diff --git a/.gitattributes b/.gitattributes index c816750670d0dc675986755fe8026927ead57c97..250c692e90754f85251ec406700accaea8bbb908 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5384,3 +5384,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/EasyHOI_[[:space:]]Unleashing[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Large[[:space:]]Models[[:space:]]for[[:space:]]Reconstructing[[:space:]]Hand-Object[[:space:]]Interactions[[:space:]]in[[:space:]]the[[:space:]]Wild/8c622e8a-33b6-400d-9aee-8eb774d315a9_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/EchoMatch_[[:space:]]Partial-to-Partial[[:space:]]Shape[[:space:]]Matching[[:space:]]via[[:space:]]Correspondence[[:space:]]Reflection/7a3c48f7-95c7-446b-9823-5172d80222c7_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/EchoMimicV2_[[:space:]]Towards[[:space:]]Striking,[[:space:]]Simplified,[[:space:]]and[[:space:]]Semi-Body[[:space:]]Human[[:space:]]Animation/288a5554-9d81-4d12-bbd8-a666ffaecfb1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Few-Shot[[:space:]]Class-Incremental[[:space:]]Learning[[:space:]]via[[:space:]]Training-Free[[:space:]]Bi-Level[[:space:]]Modality[[:space:]]Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Online[[:space:]]Continual[[:space:]]Learning[[:space:]]with[[:space:]]Plug-and-Play[[:space:]]State[[:space:]]Space[[:space:]]Model[[:space:]]and[[:space:]]Class-Conditional[[:space:]]Mixture[[:space:]]of[[:space:]]Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Privacy-Utility[[:space:]]Trade-offs[[:space:]]to[[:space:]]Mitigate[[:space:]]Memorization[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]SAM[[:space:]]with[[:space:]]Efficient[[:space:]]Prompting[[:space:]]and[[:space:]]Preference[[:space:]]Optimization[[:space:]]for[[:space:]]Semi-supervised[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Testing-Time[[:space:]]Robustness[[:space:]]for[[:space:]]Trusted[[:space:]]Multi-View[[:space:]]Classification[[:space:]]in[[:space:]]the[[:space:]]Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Video-LLM[[:space:]]Reasoning[[:space:]]via[[:space:]]Agent-of-Thoughts[[:space:]]Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Virtual[[:space:]]Try-On[[:space:]]with[[:space:]]Synthetic[[:space:]]Pairs[[:space:]]and[[:space:]]Error-Aware[[:space:]]Noise[[:space:]]Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Vision-Language[[:space:]]Compositional[[:space:]]Understanding[[:space:]]with[[:space:]]Multimodal[[:space:]]Synthetic[[:space:]]Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EnliveningGS_[[:space:]]Active[[:space:]]Locomotion[[:space:]]of[[:space:]]3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EntityErasure_[[:space:]]Erasing[[:space:]]Entity[[:space:]]Cleanly[[:space:]]via[[:space:]]Amodal[[:space:]]Entity[[:space:]]Segmentation[[:space:]]and[[:space:]]Completion/841507d5-5b60-4a74-9926-0461296e0a84_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EntitySAM_[[:space:]]Segment[[:space:]]Everything[[:space:]]in[[:space:]]Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EntropyMark_[[:space:]]Towards[[:space:]]More[[:space:]]Harmless[[:space:]]Backdoor[[:space:]]Watermark[[:space:]]via[[:space:]]Entropy-based[[:space:]]Constraint[[:space:]]for[[:space:]]Open-source[[:space:]]Dataset[[:space:]]Copyright[[:space:]]Protection/79eea005-2d9c-4999-9dbc-01e224e70254_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EnvGS_[[:space:]]Modeling[[:space:]]View-Dependent[[:space:]]Appearance[[:space:]]with[[:space:]]Environment[[:space:]]Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EnvPoser_[[:space:]]Environment-aware[[:space:]]Realistic[[:space:]]Human[[:space:]]Motion[[:space:]]Estimation[[:space:]]from[[:space:]]Sparse[[:space:]]Observations[[:space:]]with[[:space:]]Uncertainty[[:space:]]Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EquiPose_[[:space:]]Exploiting[[:space:]]Permutation[[:space:]]Equivariance[[:space:]]for[[:space:]]Relative[[:space:]]Camera[[:space:]]Pose[[:space:]]Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Erase[[:space:]]Diffusion_[[:space:]]Empowering[[:space:]]Object[[:space:]]Removal[[:space:]]Through[[:space:]]Calibrating[[:space:]]Diffusion[[:space:]]Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Erasing[[:space:]]Undesirable[[:space:]]Influence[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/cfcabb09-71fe-461a-a521-c3d740140c91_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Escaping[[:space:]]Plato's[[:space:]]Cave_[[:space:]]Towards[[:space:]]the[[:space:]]Alignment[[:space:]]of[[:space:]]3D[[:space:]]and[[:space:]]Text[[:space:]]Latent[[:space:]]Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Estimating[[:space:]]Body[[:space:]]and[[:space:]]Hand[[:space:]]Motion[[:space:]]in[[:space:]]an[[:space:]]Ego-sensed[[:space:]]World/71e1babb-034c-40be-8709-39f13bf4244a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Ev-3DOD_[[:space:]]Pushing[[:space:]]the[[:space:]]Temporal[[:space:]]Boundaries[[:space:]]of[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Event[[:space:]]Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EvEnhancer_[[:space:]]Empowering[[:space:]]Effectiveness,[[:space:]]Efficiency[[:space:]]and[[:space:]]Generalizability[[:space:]]for[[:space:]]Continuous[[:space:]]Space-Time[[:space:]]Video[[:space:]]Super-Resolution[[:space:]]with[[:space:]]Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EvOcc_[[:space:]]Accurate[[:space:]]Semantic[[:space:]]Occupancy[[:space:]]for[[:space:]]Automated[[:space:]]Driving[[:space:]]Using[[:space:]]Evidence[[:space:]]Theory/7c502132-611b-4140-a7e5-11680d62e1b1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Eval3D_[[:space:]]Interpretable[[:space:]]and[[:space:]]Fine-grained[[:space:]]Evaluation[[:space:]]for[[:space:]]3D[[:space:]]Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Evaluating[[:space:]]Model[[:space:]]Perception[[:space:]]of[[:space:]]Color[[:space:]]Illusions[[:space:]]in[[:space:]]Photorealistic[[:space:]]Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Evaluating[[:space:]]Vision-Language[[:space:]]Models[[:space:]]as[[:space:]]Evaluators[[:space:]]in[[:space:]]Path[[:space:]]Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event[[:space:]]Ellipsometer_[[:space:]]Event-based[[:space:]]Mueller-Matrix[[:space:]]Video[[:space:]]Imaging/00131384-06b2-4638-a55f-16332464dd84_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event[[:space:]]Fields_[[:space:]]Capturing[[:space:]]Light[[:space:]]Fields[[:space:]]at[[:space:]]High[[:space:]]Speed,[[:space:]]Resolution,[[:space:]]and[[:space:]]Dynamic[[:space:]]Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-Equalized[[:space:]]Dense[[:space:]]Video[[:space:]]Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-based[[:space:]]Video[[:space:]]Super-Resolution[[:space:]]via[[:space:]]State[[:space:]]Space[[:space:]]Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EventFly_[[:space:]]Event[[:space:]]Camera[[:space:]]Perception[[:space:]]from[[:space:]]Ground[[:space:]]to[[:space:]]the[[:space:]]Sky/ecab4823-da56-483f-8321-47b8a1d3b698_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EventGPT_[[:space:]]Event[[:space:]]Stream[[:space:]]Understanding[[:space:]]with[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/bb8d5233-3517-4797-9367-5fe773bd874a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EventPSR_[[:space:]]Surface[[:space:]]Normal[[:space:]]and[[:space:]]Reflectance[[:space:]]Estimation[[:space:]]from[[:space:]]Photometric[[:space:]]Stereo[[:space:]]Using[[:space:]]an[[:space:]]Event[[:space:]]Camera/133da766-89f1-4c63-bf52-44b0090134e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EventSplat_[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]from[[:space:]]Moving[[:space:]]Event[[:space:]]Cameras[[:space:]]for[[:space:]]Real-time[[:space:]]Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Every[[:space:]]SAM[[:space:]]Drop[[:space:]]Counts_[[:space:]]Embracing[[:space:]]Semantic[[:space:]]Priors[[:space:]]for[[:space:]]Multi-Modality[[:space:]]Image[[:space:]]Fusion[[:space:]]and[[:space:]]Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Everything[[:space:]]to[[:space:]]the[[:space:]]Synthetic_[[:space:]]Diffusion-driven[[:space:]]Test-time[[:space:]]Adaptation[[:space:]]via[[:space:]]Synthetic-Domain[[:space:]]Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Evolving[[:space:]]High-Quality[[:space:]]Rendering[[:space:]]and[[:space:]]Reconstruction[[:space:]]in[[:space:]]a[[:space:]]Unified[[:space:]]Framework[[:space:]]with[[:space:]]Contribution-Adaptive[[:space:]]Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exact_[[:space:]]Exploring[[:space:]]Space-Time[[:space:]]Perceptive[[:space:]]Clues[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Satellite[[:space:]]Image[[:space:]]Time[[:space:]]Series[[:space:]]Semantic[[:space:]]Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ExpertAF_[[:space:]]Expert[[:space:]]Actionable[[:space:]]Feedback[[:space:]]from[[:space:]]Video/ccac2a98-a341-4697-825f-e838745145a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Explainable[[:space:]]Saliency_[[:space:]]Articulating[[:space:]]Reasoning[[:space:]]with[[:space:]]Contextual[[:space:]]Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Explaining[[:space:]]Domain[[:space:]]Shifts[[:space:]]in[[:space:]]Language_[[:space:]]Concept[[:space:]]Erasing[[:space:]]for[[:space:]]Interpretable[[:space:]]Image[[:space:]]Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Explaining[[:space:]]in[[:space:]]Diffusion_[[:space:]]Explaining[[:space:]]a[[:space:]]Classifier[[:space:]]with[[:space:]]Diffusion[[:space:]]Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Explicit[[:space:]]Depth-Aware[[:space:]]Blurry[[:space:]]Video[[:space:]]Frame[[:space:]]Interpolation[[:space:]]Guided[[:space:]]by[[:space:]]Differential[[:space:]]Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploiting[[:space:]]Deblurring[[:space:]]Networks[[:space:]]for[[:space:]]Radiance[[:space:]]Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploiting[[:space:]]Temporal[[:space:]]State[[:space:]]Space[[:space:]]Sharing[[:space:]]for[[:space:]]Video[[:space:]]Semantic[[:space:]]Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploration-Driven[[:space:]]Generative[[:space:]]Interactive[[:space:]]Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]CLIP's[[:space:]]Dense[[:space:]]Knowledge[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Semantic[[:space:]]Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Contextual[[:space:]]Attribute[[:space:]]Density[[:space:]]in[[:space:]]Referring[[:space:]]Expression[[:space:]]Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Historical[[:space:]]Information[[:space:]]for[[:space:]]RGBE[[:space:]]Visual[[:space:]]Tracking[[:space:]]with[[:space:]]Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Intrinsic[[:space:]]Normal[[:space:]]Prototypes[[:space:]]within[[:space:]]a[[:space:]]Single[[:space:]]Image[[:space:]]for[[:space:]]Universal[[:space:]]Anomaly[[:space:]]Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Scene[[:space:]]Affinity[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]LiDAR[[:space:]]Semantic[[:space:]]Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Semantic[[:space:]]Feature[[:space:]]Discrimination[[:space:]]for[[:space:]]Perceptual[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]and[[:space:]]Opinion-Unaware[[:space:]]No-Reference[[:space:]]Image[[:space:]]Quality[[:space:]]Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Simple[[:space:]]Open-Vocabulary[[:space:]]Semantic[[:space:]]Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Sparse[[:space:]]MoE[[:space:]]in[[:space:]]GANs[[:space:]]for[[:space:]]Text-conditioned[[:space:]]Image[[:space:]]Synthesis/7078b157-7879-47cb-991a-09f81047db7a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Temporally-Aware[[:space:]]Features[[:space:]]for[[:space:]]Point[[:space:]]Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Timeline[[:space:]]Control[[:space:]]for[[:space:]]Facial[[:space:]]Motion[[:space:]]Generation/65b84757-8067-4189-a894-056e4be1d124_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Visual[[:space:]]Vulnerabilities[[:space:]]via[[:space:]]Multi-Loss[[:space:]]Adversarial[[:space:]]Search[[:space:]]for[[:space:]]Jailbreaking[[:space:]]Vision-Language[[:space:]]Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]the[[:space:]]Deep[[:space:]]Fusion[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Diffusion[[:space:]]Transformers[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exposure-slot_[[:space:]]Exposure-centric[[:space:]]Representations[[:space:]]Learning[[:space:]]with[[:space:]]Slot-in-Slot[[:space:]]Attention[[:space:]]for[[:space:]]Region-aware[[:space:]]Exposure[[:space:]]Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Extrapolating[[:space:]]and[[:space:]]Decoupling[[:space:]]Image-to-Video[[:space:]]Generation[[:space:]]Models_[[:space:]]Motion[[:space:]]Modeling[[:space:]]is[[:space:]]Easier[[:space:]]Than[[:space:]]You[[:space:]]Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Extreme[[:space:]]Rotation[[:space:]]Estimation[[:space:]]in[[:space:]]the[[:space:]]Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/F-LMM_[[:space:]]Grounding[[:space:]]Frozen[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FADA_[[:space:]]Fast[[:space:]]Diffusion[[:space:]]Avatar[[:space:]]Synthesis[[:space:]]with[[:space:]]Mixed-Supervised[[:space:]]Multi-CFG[[:space:]]Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FADE_[[:space:]]Frequency-Aware[[:space:]]Diffusion[[:space:]]Model[[:space:]]Factorization[[:space:]]for[[:space:]]Video[[:space:]]Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FALCON_[[:space:]]Fairness[[:space:]]Learning[[:space:]]via[[:space:]]Contrastive[[:space:]]Attention[[:space:]]Approach[[:space:]]to[[:space:]]Continual[[:space:]]Semantic[[:space:]]Scene[[:space:]]Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_content_list.json b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e72b9ca3f1d6e2223b97d99c2dea1b5f47a2a74c --- /dev/null +++ b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration", + "text_level": 1, + "bbox": [ + 248, + 130, + 750, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiyang Chen $^{1}$ , Tianyu Ding $^{2}$ , Lei Wang $^{3}$ , Jing Huo $^{1}$ , Yang Gao $^{1}$ , Wenbin Li $^{1,4*}$ $^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China \n $^{2}$ Applied Sciences Group, Microsoft, USA $^{3}$ University of Wollongong, Australia \n $^{4}$ Shenzhen Research Institute of Nanjing University, Shenzhen, China", + "bbox": [ + 158, + 202, + 836, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot Class-Incremental Learning (FSCIL) challenges models to adapt to new classes with limited samples, presenting greater difficulties than traditional class-incremental learning. While existing approaches rely heavily on visual models and require additional training during base or incremental phases, we propose a training-free framework that leverages pre-trained visual-language models like CLIP. At the core of our approach is a novel Bi-level Modality Calibration (BiMC) strategy. Our framework initially performs intra-modal calibration, combining LLM-generated fine-grained category descriptions with visual prototypes from the base session to achieve precise classifier estimation. This is further complemented by inter-modal calibration that fuses pre-trained linguistic knowledge with task-specific visual priors to mitigate modality-specific biases. To enhance prediction robustness, we introduce additional metrics and strategies that maximize the utilization of limited data. Extensive experimental results demonstrate that our approach significantly outperforms existing methods. Code is available at: https://github.com/yychen016/BiMC.", + "bbox": [ + 88, + 340, + 483, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 669, + 223, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of artificial intelligence has propelled deep neural networks to achieve remarkable progress. However, these models often struggle with catastrophic forgetting when faced with dynamically changing environments, impeding their ability to learn sequential tasks effectively. Class-Incremental Learning (CIL) can be a potential solution to this challenge. Yet, in practical settings, researchers frequently encounter a scarcity of training data during the incremental phase, leading to sample sparsity issues. Consequently, Few-Shot Class-Incremental Learning (FSCIL) [31] has recently emerged as a more realistic scenario, better reflecting real-world constraints.", + "bbox": [ + 89, + 695, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/030fcc11f45b67201aee0601c2407f7098dec105eb21477a20c26411ba68f89f.jpg", + "image_caption": [ + "Figure 1. Model updating paradigms in FSCIL. (a) Continuous model updates during incremental phases. (b) Leveraging the base task to learn a generalizable model. (c) Our proposed framework: training-free in both base and incremental sessions." + ], + "image_footnote": [], + "bbox": [ + 516, + 306, + 903, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current FSCIL methods can be broadly categorized into two paradigms: the continual fine-tuning paradigm and the incremental-session freezing paradigm, as shown in Figure 1 (a) and (b). The former [4, 21, 27, 43] utilizes limited samples to continuously update the model, enhancing its plasticity. Conversely, the latter [22, 35, 44, 48] focuses on learning robust representations during the base task, which are then generalized to incremental tasks, prioritizing stability by avoiding subsequent model updates. Despite their distinct approaches, both paradigms necessitate additional training to adapt to new tasks, presenting ongoing challenges in resource efficiency and model adaptability.", + "bbox": [ + 509, + 595, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of large-scale models has shifted the landscape of representation learning. Vision-language models like CLIP [24], trained on a large number of image-text pairs, excel in zero-shot tasks. However, fine-tuning is computationally expensive and may compromise generalization, especially with limited data [14, 39]. While recent approaches like prompt tuning [46, 47] and low-rank adaptation [9] offer more efficient alternatives, they still demand", + "bbox": [ + 509, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9881", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "additional training. This raises a crucial question: How can we effectively leverage pre-trained knowledge and domain-specific visual priors to adapt the large-scale model using few samples without additional training?", + "bbox": [ + 88, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While CLIP is widely used for cross-modal retrieval in visual categorization, it overlooks the potential of intramodal recognition, akin to traditional purely visual models, such as CNN [30, 41]. We believe that CLIP's visual prototypes can effectively complement zero-shot textual classifiers, serving as domain-specific priors, particularly in addressing two major challenges in FSCIL tasks: catastrophic forgetting and overfitting. Building on this insight, we introduce Bi-level Modality Calibration (BiMC), a simple yet effective framework that treats CLIP as a black-box model and operates without additional training, thereby avoiding the parameter updates that typically cause forgetting. Our framework incorporates both intra-modal and inter-modal calibration strategies to enhance CLIP's classification accuracy. For intra-modal calibration, we leverage LLM-generated fine-grained category descriptions in the textual domain to improve zero-shot classifier discrimination, while using base session prototypes in the visual domain to refine new class prototypes affected by sample scarcity. Inter-modal calibration combines pre-trained linguistic knowledge with task-aware visual priors to reduce modal biases and overfitting. The framework is further enhanced with an anisotropic covariance metric and a cross-modal category-conditioned nearest-neighbor metric, implementing the final classifier through masked ensemble inference. Extensive experiments demonstrate that our method achieves competitive performance without additional training.", + "bbox": [ + 91, + 152, + 485, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 575, + 413, + 588 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel training-free FSCIL framework that achieves continual model adaptation by treating vision-language models as black boxes, significantly enhancing practicality for real-world applications.", + "- We develop a bi-level modality calibration approach combining intra-modal and inter-modal strategies to enhance classifier accuracy across both modalities. This approach is strengthened by an innovative visual covariance metric and a category-conditioned nearest-neighbor metric, culminating in a robust masked ensemble inference strategy.", + "- We demonstrate the effectiveness of our framework through extensive experiments on standard benchmarks, where it not only achieves competitive performance but also outperforms supervised training methods in several scenarios, despite requiring no additional training." + ], + "bbox": [ + 89, + 590, + 482, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 844, + 228, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-Shot Class-Incremental Learning. Few-Shot Class-Incremental Learning (FSCIL) challenges models to adapt", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "continuously using minimal samples. Existing approaches can be categorized into three main groups: representation-based methods [1, 2, 12, 18, 22, 27, 30, 38, 44], which focus on optimizing feature representations; dynamic-architecture-based methods [31, 41], which adapt model structures to accommodate new classes; and knowledge-distillation-based methods [4, 7, 15], which transfer knowledge from previous tasks to new ones. Notable examples include CEC [41], which employs an evolving graph model for classifier optimization, and FACT [44] and SAVC [30], which introduce virtual categories as placeholders for forward compatibility. In contrast, our approach leverages a pre-trained vision-language model, eliminating the need for additional learning on the base session.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Class-Incremental Learning via Pre-Trained Models. Recent trends in incremental learning focus on effectively utilizing pre-trained models as backbones for downstream tasks. Early works primarily used prompt-based methods [11, 28, 34, 36, 37], introducing learnable prompt parameters for task adaptation. Other approaches have employed adapters [45] or LoRA [9] to facilitate adaptation [17, 40, 45]. While these methods use efficient parameter-tuning techniques, they still require further training. Our approach, however, leverages both vision and language modalities to enable ongoing model adaptation without gradient updates.", + "bbox": [ + 511, + 303, + 903, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-Assisted Classification. Linguistic information has been increasingly utilized in various visual tasks. Some studies have used linguistic cues to guide visual classification [33, 42], while others have leveraged Large Language Models (LLMs) to generate rich descriptive text [20, 23, 26]. In continual learning, several works have exploited linguistic guidance [13, 19]. Within FSCIL, methods like [2, 4, 21] also utilize linguistic information. However, these approaches have limitations: some [2] use distillation techniques on smaller visual models with linguistic knowledge as a regularization term, potentially limiting scalability and flexibility. Others [4, 21] rely on complex and often redundant architectural designs, making implementation challenging and inefficient. In contrast, our approach addresses these limitations by proposing a simpler, more efficient framework that effectively integrates linguistic and visual information.", + "bbox": [ + 511, + 484, + 903, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 758, + 604, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 784, + 640, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In FSCIL, a model continuously adapts to a sequence of sessions $\\mathcal{S} = \\{\\mathcal{D}_0, \\mathcal{D}_1, \\dots, \\mathcal{D}_T\\}$ . Each session $\\mathcal{D}_t = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{N_t}$ consists of training data with a category space $\\mathcal{C}_t^{\\mathrm{train}} = Y_t$ , where $N_t$ represents the number of samples in the $t$ -th session. The base session $\\mathcal{D}_0$ contains a substantial number of categories and samples, while subsequent incre", + "bbox": [ + 511, + 808, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9882", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b4ad33fa5bebd94ccc23844662c36caebe6b538983bd185a292c7176b0aff18.jpg", + "image_caption": [ + "(a) Bi-Level Modality Calibration Framework" + ], + "image_footnote": [], + "bbox": [ + 107, + 97, + 614, + 388 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c99d754171076c28918dddfcfc68ec896af88138cf8b42f7d23eee3226a43551.jpg", + "image_caption": [ + "(b) Semantic and Covariance-enhanced Metric", + "Figure 2. Overview of our framework. In each task adaptation phase, we construct a calibrated classifier through our proposed bi-level calibration framework, which involves intra-modal calibration and inter-modal calibration. To enhance performance, we introduce a globally shared covariance metric for visual feature modeling, complemented by a category-conditioned nearest-neighbor scoring strategy." + ], + "image_footnote": [], + "bbox": [ + 645, + 97, + 890, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mental sessions $\\mathcal{D}_t$ for $t > 0$ contain only a few training samples per category, formally expressed as $N_0\\gg N_t$", + "bbox": [ + 89, + 459, + 482, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The categories across different sessions are mutually exclusive, meaning that for any different $i,j\\in [0,T]$ $Y_{i}\\cap Y_{j} = \\varnothing$ . Furthermore, during the training of the current session, data from previous sessions is not accessible. At the testing phase of the $t$ -th session, the model evaluates against all previously encountered categories, with the cumulative category set defined as $\\mathcal{V}_t = \\bigcup_{j = 0}^t Y_j$ . Consequently, the test category space is given by $\\mathcal{C}_t^{\\mathrm{test}} = \\mathcal{V}_t$", + "bbox": [ + 89, + 491, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Bi-Level calibration framework", + "text_level": 1, + "bbox": [ + 89, + 622, + 370, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed training-free framework builds upon CLIP [24] and incorporates two essential components: intra-modal and inter-modal calibration (see Figure 2a).", + "bbox": [ + 89, + 646, + 482, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Intra-modal classifier calibration", + "text_level": 1, + "bbox": [ + 89, + 710, + 375, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In its standard usage, CLIP relies on simple prompt templates, (e.g., \"a photo of a [CLS].\") to construct zero-shot classifiers for visual tasks. However, this category-agnostic approach has inherent limitations. The generic nature of these templates fails to capture category-specific nuances, making them particularly inadequate for fine-grained classification tasks where detailed feature discrimination is crucial. A single, static template cannot effectively characterize the diverse visual characteristics within each category, especially when subtle distinctions between categories are essential.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome these limitations, we leverage Large Language Models (LLMs) to generate dynamic, category-specific descriptions. These LLM-generated descriptions provide richer, more discriminative semantic representations for each category, enhancing the model's capacity to capture fine-grained features. We integrate this enhanced semantic information through an intra-modal calibration strategy within the textual modality:", + "bbox": [ + 511, + 460, + 906, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {\\mu}} _ {c} ^ {T} = \\left(1 - \\lambda_ {T}\\right) \\boldsymbol {w} _ {c} + \\lambda_ {T} \\left(\\frac {1}{n _ {c}} \\sum_ {j = 1} ^ {n _ {c}} \\frac {g \\left(\\mathbf {t} _ {c , j}\\right)}{\\| g \\left(\\mathbf {t} _ {c , j}\\right) \\| _ {2}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 588, + 906, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this formulation, $\\lambda_{T}$ controls the intensity of intra-modal calibration within the text modality, $n_c$ denotes the number of descriptions for category $c$ , $g(\\cdot)$ represents the text encoder, $\\mathbf{t}_{c,j}$ is the $j$ -th LLM-generated description for category $c$ , $\\tilde{\\boldsymbol{\\mu}}_c^T$ is the calibrated textual prototype, and $\\boldsymbol{w}_c$ denotes the original CLIP Zero-Shot classifier weight for class $c$ . This calibration mechanism preserves the robust generalization capabilities of the CLIP Zero-Shot classifier while incorporating fine-grained, category-specific semantic information. The latter term in Eq. (1) serves as a semantically enriched category description center, enhancing the model's discriminative power.", + "bbox": [ + 511, + 643, + 906, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the FSCIL context, prototype-based classifiers are commonly employed, where the accuracy of prototype estimation directly impacts performance. During the base session, the abundance of samples allows for precise estimation of class prototypes. However, in incremental sessions", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9883", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that introduce new classes, the limited number of samples inevitably leads to biased estimates of new class prototypes. Drawing inspiration from TEEN [35], we leverage the accurately estimated prototype weights of base classes to facilitate intra-modality calibration on the prototypes of incoming new classes within the visual modality.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given class $c$ , we first utilize the encoded training image data to obtain the naive visual prototype:", + "bbox": [ + 89, + 181, + 482, + 213 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {c} ^ {I} = \\frac {1}{m _ {c}} \\sum_ {j = 1} ^ {m _ {c}} \\frac {f (\\mathbf {x} _ {c , j})}{\\| f (\\mathbf {x} _ {c , j}) \\| _ {2}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 220, + 482, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f(\\cdot)$ is the visual encoder, $m_{c}$ is the number of images in class $c$ , and $\\mathbf{x}_{c,j}$ is the $j$ -th image belonging to category $c$ . Due to the inaccuracy in visual prototype estimation during the incremental sessions, the prototypes from the base session are used to recalibrate the new class prototypes by leveraging the similarity relationship between them. This process can be formulated as:", + "bbox": [ + 89, + 271, + 483, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I} = \\left\\{ \\begin{array}{l l} \\boldsymbol {\\mu} _ {c} ^ {I} & , t = 0 \\\\ (1 - \\lambda_ {I}) \\boldsymbol {\\mu} _ {c} ^ {I} + \\lambda_ {I} \\sum_ {b = 1} ^ {| \\mathcal {Y} _ {0} |} s _ {b, c} \\boldsymbol {\\mu} _ {b} ^ {I} & , t > 0, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 385, + 482, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, $t$ represents the task identifier, $\\lambda_I$ controls the strength of the visual intra-modal calibration, and $s_{b,c}$ denotes the normalized cosine similarity between the visual prototypes of class $b$ (from the base classes) and class $c$ . The similarity $s_{b,c}$ is defined as: $s_{b,c} = \\frac{e^{\\tau \\cdot \\langle \\mu_b, \\mu_c \\rangle}}{\\sum_{i=1}^{|\\mathcal{Y}_0|} e^{\\tau \\cdot \\langle \\mu_i, \\mu_c \\rangle}}$ . Here, $\\langle \\cdot, \\cdot \\rangle$ represents the cosine similarity, and $\\tau$ is the temperature scaling parameter.", + "bbox": [ + 89, + 458, + 482, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through this visual intra-modality calibration process, the prototypes of the new classes are adjusted to align more closely with the well-calibrated base class prototypes, allowing the new class prototypes to partially inherit the discriminative properties of the base class prototypes.", + "bbox": [ + 89, + 571, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Inter-modal classifier calibration", + "text_level": 1, + "bbox": [ + 89, + 665, + 374, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following intra-modal calibration, two classifier achieves enhanced accuracy within its respective modality. However, we posit that their full potential remains unrealized when operating in isolation. This is due to the fact that classifiers operating exclusively on linguistic data derive their knowledge from pre-trained models and lack explicit downstream knowledge for visual tasks. Conversely, classifiers functioning solely within the visual domain serve as metrics within the visual sub-feature space but are deficient in semantic understanding. Recognizing the complementary nature of these two modalities, we propose a training-free inter-modal calibration strategy, allowing domain-relevant downstream visual prior to guide domain-agnostic pretrained linguistic knowledge. Formally, this process can be", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "described as:", + "bbox": [ + 513, + 90, + 602, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {c} = \\beta \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {T} + (1 - \\beta) \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 114, + 906, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\pmb{\\mu}_{c}$ is the inter-modal calibrated classifier, and $\\beta$ is the calibration coefficient. During inference, we use cosine similarity to calculate the similarity between the sample features $f(\\mathbf{x})$ and the mixed category center:", + "bbox": [ + 511, + 142, + 906, + 203 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {c} ^ {\\text {c a l i b}} = \\frac {f (\\mathbf {x}) ^ {\\top} \\boldsymbol {\\mu} _ {c}}{\\| f (\\mathbf {x}) \\| _ {2} \\cdot \\| \\boldsymbol {\\mu} _ {c} \\| _ {2}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 212, + 906, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This inter-modal calibration strategy effectively combines the strengths of both linguistic and visual modalities. It helps to enhance the overall classification performance by creating a more robust and comprehensive classifier that can better capture the nuances of visual tasks while maintaining a strong semantic understanding.", + "bbox": [ + 511, + 265, + 906, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic and covariance-enhanced metric", + "text_level": 1, + "bbox": [ + 513, + 364, + 874, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While our previous metrics effectively integrate linguistic and visual information and mitigate modal bias, they primarily measure sample distances from distribution centers. This isotropic approach, despite its simplicity, struggles to capture higher-order information and complex data relationships. To address these limitations, we propose leveraging statistical information from existing data for more comprehensive and efficient measurements, as shown in Figure 2b.", + "bbox": [ + 511, + 387, + 905, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Modeling global covariance", + "text_level": 1, + "bbox": [ + 513, + 526, + 759, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by FeCAM [8], we introduce an anisotropic covariance metric to fully utilize the visual modality's statistical information. For a session $t$ , we compute its covariance matrix $\\boldsymbol{\\Sigma}^t$ using image embeddings $f(\\mathbf{x})$ of all image data relevant to the session. To avoid singularity, we apply regularization with an identity matrix $I_d$ :", + "bbox": [ + 511, + 549, + 905, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {\\Sigma}} ^ {t} = \\boldsymbol {\\Sigma} ^ {t} + \\frac {\\gamma}{d} \\operatorname {t r} \\left(\\boldsymbol {\\Sigma} ^ {t}\\right) \\boldsymbol {I} _ {d}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 648, + 906, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a hyperparameter controlling the strength of regularization, $d$ represents the feature dimension. Here, $\\mathrm{tr}(\\cdot)$ denotes the trace operator of a matrix, and both $d$ and $\\mathrm{tr}(\\pmb{\\Sigma}^t)$ provide a prior estimation for the regularization strength. Subsequently, we construct a continuously evolving shared covariance matrix as:", + "bbox": [ + 511, + 683, + 906, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {t} = \\frac {| \\mathcal {Y} _ {t - 1} |}{| \\mathcal {Y} _ {t} |} \\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {t - 1} + \\left(1 - \\frac {| \\mathcal {Y} _ {t - 1} |}{| \\mathcal {Y} _ {t} |}\\right) \\tilde {\\boldsymbol {\\Sigma}} ^ {t}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 782, + 906, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The evolved covariance matrix captures the statistical information of all categories. Given a test sample $\\mathbf{x}$ , we calculate its score as:", + "bbox": [ + 511, + 824, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {c} ^ {\\mathrm {c o v}} (\\mathbf {x}) = - \\frac {1}{d} \\left(f (\\mathbf {x}) - \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}\\right) ^ {T} \\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {- 1} \\left(f (\\mathbf {x}) - \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 875, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9884", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a994cc2e8cb606d02da4817343093fbf085c29173d024f0f17b4a66f014e3e48.jpg", + "image_caption": [ + "Figure 3. In the inference phase, an ensemble strategy with category masking is utilized." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 475, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The division by feature dimension $d$ serves as a normalization factor, preventing the softmax transformation into probabilities from collapsing into a single category. The negative sign maintains metric consistency, ensuring that a larger output score indicates a higher probability of the sample belonging to category $c$ . Here, $\\hat{\\Sigma}_G$ denotes the global covariance, with the superscript $t$ omitted for clarity.", + "bbox": [ + 89, + 255, + 483, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Cross-modal category nearest neighbor metric", + "text_level": 1, + "bbox": [ + 89, + 382, + 467, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global covariance modeling effectively utilizes statistical information from the visual modality. To better leverage LLM-generated category descriptions, we propose a category-conditioned nearest-neighbor-based metric. Given a description text $\\mathbf{t}_{c,j}$ encoded by $g(\\cdot)$ , we obtain its normalized feature representation $\\mathbf{z}_{c,j} = \\frac{g(\\mathbf{t}_{c,j})}{\\|g(\\mathbf{t}_{c,j})\\|_2}$ . For a visual query sample, we compute its normalized representation $\\mathbf{v} = \\frac{f(\\mathbf{x})}{\\|f(\\mathbf{x})\\|_2}$ . The likelihood of this sample belonging to category $c$ is determined by the maximum dot product between $\\mathbf{v}$ and all description features in class $c$ :", + "bbox": [ + 89, + 406, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {c} ^ {\\mathrm {n n}} = \\max _ {j} \\left\\{\\boldsymbol {z} _ {c, j} ^ {\\top} \\cdot \\boldsymbol {v} \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 575, + 482, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{s}_c^{\\mathrm{nn}}$ indicates the likelihood that the query sample belongs to category $c$ . This metric leverages maximum cross-modal similarity, effectively utilizing the diverse textual descriptions generated by LLM.", + "bbox": [ + 89, + 613, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Inference score reorganization strategy", + "text_level": 1, + "bbox": [ + 89, + 695, + 416, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a globally shared covariance matrix to model distribution shape information in the visual feature space. While the base session contains substantial data for accurate covariance matrix estimation, experiments reveal that these base-derived covariance matrices perform poorly in distinguishing new class data. This limitation stems from two factors: (1) the covariance from extensive base data fails to align with new classes, and (2) the limited data available for new classes prevents accurate covariance estimation. To address this, we use the cross-modal category nearest neighbor metric for novel classes to compensate for the limitations of the second-order covariance metric, as shown in Figure 3.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first normalize the scores of three metrics using the softmax function: $\\mathbf{p}^{\\mathrm{calib}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{calib}}\\right)$ , $\\mathbf{p}^{\\mathrm{cov}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{cov}}\\right)$ , and $\\mathbf{p}^{\\mathrm{nn}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{nn}}\\right)$ . The final probability score $\\mathbf{p}_c$ is computed differently for base and novel categories:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {c} = \\left\\{ \\begin{array}{l l} \\alpha \\mathbf {p} _ {c} ^ {\\text {c a l i b}} + (1 - \\alpha) \\mathbf {p} _ {c} ^ {\\text {c o v}}, & c \\in \\mathcal {Y} _ {0} \\\\ \\alpha \\mathbf {p} _ {c} ^ {\\text {c a l i b}} + (1 - \\alpha) \\mathbf {p} _ {c} ^ {\\text {n n}}, & c \\notin \\mathcal {Y} _ {0}. \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 176, + 905, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For base categories $(c\\in \\mathcal{V}_0)$ , we use a weighted sum of mixed prototype scores and visual mahalanobis metric scores. For novel categories $(c\\notin \\mathcal{V}_0)$ , we combine mixed prototype scores with cross-modal nearest neighbor metric scores. The final category prediction for a sample is determined by $\\arg \\max_c\\mathbf{p}_c$", + "bbox": [ + 511, + 229, + 905, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 334, + 645, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 511, + 359, + 718, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three benchmarks following [22, 30, 41]: CIFAR100 [3], CUB200-2011 [32], and miniImageNet [25]. For CIFAR100 and miniImageNet, we partition each dataset into 60 base classes and 40 novel classes. The novel classes are further divided into eight incremental tasks, with each task structured as a 5-way 5-shot incremental session. The CUB200 dataset is split into 100 base classes and 100 novel classes, with each incremental task organized in a 10-way 5-shot format.", + "bbox": [ + 511, + 383, + 905, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We adopt CLIP's ViT-B/16 architecture as the backbone for fair comparison across all methods. For category descriptions, we utilize fine-grained text generated by CuPL [23] and [26]. The hyperparameters are configured with intra-modal calibration coefficients $\\lambda_{T} = 0.5$ (textual), and $\\lambda_{I} = 0.1$ (visual), and a temperature parameter $\\tau = 16$ . The inter-modal calibration coefficient $\\beta$ is optimized through a validation process. Specifically, we divide the base session into training and validation subsets. Applying our bi-level calibration framework on the training subset, we evaluate performance on the validation subset. By incrementing $\\beta$ by 0.05, we identify the value that maximizes validation accuracy. This optimal $\\beta$ remains fixed throughout subsequent incremental sessions. The shared covariance parameter $\\gamma$ is set to 1 for the base session and adjusted to 5 for incremental sessions. Finally, for score reorganization, we set $\\alpha = 0.6$ for CIFAR100 and miniImageNet, and $\\alpha = 0.8$ for CUB200.", + "bbox": [ + 511, + 518, + 906, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison results", + "text_level": 1, + "bbox": [ + 511, + 801, + 700, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with training-free approaches in the FSCIL scenario: CLIP Zero-Shot [24], Visual Prototype [29], TEEN [35] and FeCAM [8]. We present the performance curve over all benchmarks in Figure 4, while detailed session-wise performance on miniImageNet is shown", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9885", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b76da2e39c5363fe7f1fb99d4cd85d6fbb9ee2357f955fa576f15f7306b53f1a.jpg", + "table_caption": [ + "Table 1. Detailed session-wise accuracy, average accuracy (Avg) and performance degradation (PD) comparison on miniImageNet dataset. V and L represent the visual and language modalities, respectively. BiMC refers to the results obtained solely through the bi-level calibration framework, whereas $\\mathrm{BiMC}^{\\dagger}$ incorporates the ensemble classifier strategy. The best results on each each sessions are indicated in bold, while the second-best results are underlined. $\\uparrow$ means higher is better, while $\\downarrow$ means lower is better." + ], + "table_footnote": [], + "table_body": "
MethodModalityAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
CLIP Zero-Shot [24]L91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
Visual Prototype [29]V92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
TEEN [35]V92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
FeCAM [8]V94.4293.7191.6190.5990.6290.2288.4388.1788.2090.666.22
BiMCV-L94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
BiMC†V-L95.4795.3493.8093.6393.5193.1892.5192.5692.4093.603.07
", + "bbox": [ + 101, + 146, + 893, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57436e0131a346d47d8c328e86056be0ec678b944fad1e04a7d10a4ca767dc9f.jpg", + "image_caption": [ + "(a) miniImageNet" + ], + "image_footnote": [ + "CLIP Zero-Shot Visual Prototype FeCAM BiMC BiMC" + ], + "bbox": [ + 116, + 295, + 370, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f23fa610cb82228685bcc854c7c2817a4182c2de560d7d661f47e5d6dfdb565b.jpg", + "image_caption": [ + "(b) CIFAR100" + ], + "image_footnote": [], + "bbox": [ + 372, + 295, + 625, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0e46788044a63a361463168d4f9be39d998a627aa93130e235db4d4d64bb120.jpg", + "image_caption": [ + "(c) CUB200" + ], + "image_footnote": [], + "bbox": [ + 629, + 296, + 885, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Table 1. Detailed results for CIFAR100 and CUB200 are provided in the supplementary material.", + "bbox": [ + 88, + 503, + 480, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach demonstrates competitive performance across all three datasets and significantly outperforms unimodal baselines without requiring additional training, highlighting the importance of leveraging information from both modalities. On the miniImageNet dataset, while visual and textual prototype classifiers achieve comparable results, and although TEEN [35] and FeCAM [8] better utilize visual modality information, their improvement remains limited. Our framework employs a simple calibration strategy to integrate domain visual knowledge with pretrained linguistic knowledge, yielding up to $6\\%$ performance improvement over uni-modal models.", + "bbox": [ + 88, + 534, + 482, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For CIFAR100, the language prototype shows better resistance to forgetting compared to the visual one, which is more susceptible to sequential task learning. This can be attributed to the dataset's low-resolution characteristics, resulting in insufficient visual information during incremental phases. Our calibration framework achieves approximately $3.5\\%$ improvement over a language-only classifier. In contrast, for the fine-grained CUB200 dataset, the visual prototype surpasses the language prototype due to limited trans", + "bbox": [ + 89, + 715, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08124e23b9f475316d421fc74513f9a8c45c39641d4dd10d2127ed4718cebc19.jpg", + "image_caption": [ + "Figure 4. Performance curve of each incremental sessions on (a) miniImageNet, (b) CIFAR100 and (c) CUB200 datasets.", + "(a) Old Classes Entropy", + "Figure 5. Comparison of output confidence with different models. ferability of pre-trained language knowledge. By combining both modalities, our approach achieves approximately $3.5\\%$ performance improvement. Notably, our framework demonstrates robust resistance to forgetting, achieving the lowest performance drop across all metrics." + ], + "image_footnote": [], + "bbox": [ + 521, + 505, + 709, + 635 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7dbfd7fc27e2554dffc90f6f9df37d8566f965101ec383ca61d203ddffc5342.jpg", + "image_caption": [ + "(b) Novel Classes Entropy" + ], + "image_footnote": [], + "bbox": [ + 710, + 505, + 898, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One can observe that each modality's contribution to classification varies across datasets. Our proposed bi-level calibration framework effectively leverages the strengths of each modality and mitigates modality bias. Moreover, our ensemble inference strategy, denoted as $\\mathrm{BiMC}^{\\dagger}$ , provides additional performance improvements.", + "bbox": [ + 511, + 744, + 905, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Investigation of bi-level calibrated classifiers", + "text_level": 1, + "bbox": [ + 511, + 845, + 888, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we analyze two distinct attributes of the calibrated classifier to demonstrate the effectiveness of our pro", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1“textual prototype” refers to the classifier generated by CLIP in a zero-shot setting, which has the same meaning as “textual classifier” mentioned later in the paper.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9886", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e8359e8b473eaffeb161dec830755d702c86a703c5a7eec21bdccad9bb942864.jpg", + "table_caption": [ + "Table 2. Ablation studies of bi-level calibration framework on miniImageNet. Vis. and Lang. individually represent classifiers derived from the visual and language modalities. Intra-C and Inter-C refer to the strategies of intra-modal and inter-modal calibration, respectively." + ], + "table_footnote": [], + "table_body": "
VisLangIntra-CInter-CAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
✓92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
✓✓92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
✓91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
✓✓93.2793.3291.9191.5191.2190.7389.2889.2389.0191.054.26
✓✓✓94.6894.5892.8792.4992.5692.3191.5991.5891.5492.693.14
✓✓✓✓94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
", + "bbox": [ + 94, + 119, + 898, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b14a28402bd443d3d3244c8b664fc1b2abab6f0fbdeb0e21f588cb696eb32624.jpg", + "table_caption": [ + "Table 3. The left side of the table lists classification result combinations from three classifiers, while the right side shows the percentage $(\\%)$ of test samples meeting each combination. Test sets from both the base and the final session were used for analysis." + ], + "table_footnote": [], + "table_body": "
ModalityCIFAR100CUB200
VLV-LD0testD8testD0testD10test
✓✓✓65.3351.8159.0044.85
X✓✓6.9513.072.064.00
✓X✓6.105.5619.8321.40
XX✓1.202.101.081.98
", + "bbox": [ + 99, + 316, + 470, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "posed framework.", + "bbox": [ + 89, + 426, + 212, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The bi-level calibrated classifier enhances prediction confidence. We evaluate prediction confidence by calculating the entropy of predictions for each classifier in the final session across all three datasets. The entropy serves as an uncertainty indicator for model outputs, and we compute it separately for both base and novel categories, as detailed in Figure 5. The findings indicate that the bi-level calibrated classifier significantly reduces the entropy of predictions compared to uni-model classifiers, thereby enhancing prediction accuracy.", + "bbox": [ + 88, + 441, + 482, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The bi-level calibrated classifier mitigates modality bias. To elucidate the superior performance of our method, we analyze the joint prediction outcomes across different classifiers in comparison to uni-modal approaches. Using $\\checkmark$ to denote correct classifications and $x$ for classification errors, we observe that the calibrated classifier maintains accurate predictions even in cases where uni-modal classifiers fail, as demonstrated in the $2^{\\mathrm{nd}}$ and $3^{\\mathrm{rd}}$ rows of Table 3. This evidence suggests that the bi-level calibrated classifier produces more accurate and unbiased predictions, offering enhanced robustness compared to uni-modality classifiers. Remarkably, in certain rare instances, the modality-calibrated classifier successfully classifies samples that both uni-modal classifiers misclassify.", + "bbox": [ + 89, + 594, + 482, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 816, + 241, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation analyses on the miniImageNet dataset to evaluate the significance of each component within our proposed framework.", + "bbox": [ + 89, + 839, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bi-level calibration framework. Our framework com-", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/83a91a36c76ca892894c516fca47b7218f434752cc848ab2db686a2e740c3e50.jpg", + "table_caption": [ + "Table 4. Ablation study of Semantic and Covariance-enhanced Metric on the miniImageNet dataset. To verify their effectiveness, we measured the performance of the base task $\\mathcal{A}_{\\mathrm{base}}$ , the performance of last task $\\mathcal{A}_{\\mathrm{last}}$ and the average performance $\\mathcal{A}_{\\mathrm{avg.}}$ across all tasks. Additionally, we report the accuracy of base $\\mathcal{A}_{\\mathrm{last}}^b$ and novel $\\mathcal{A}_{\\mathrm{last}}^n$ class in last task." + ], + "table_footnote": [], + "table_body": "
Ablation\\( A_{base} \\)\\( A_{last} \\)\\( A_{last}^{b} \\)\\( A_{last}^{n} \\)\\( A_{avg} \\)
BiMC94.9091.8193.4389.3892.97
+ MGC95.4791.5194.3787.2292.91
+ CMNN94.8592.1593.3090.4293.18
\\( BiMC^† (w/o mask.) \\)95.4792.2094.1289.3293.37
\\( BiMC^† \\)95.4792.4093.2091.2093.60
", + "bbox": [ + 516, + 344, + 901, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prises four components: visual modality information, linguistic modality information, intra-modal calibration strategy, and inter-modal calibration strategy. As shown in Table 2, single-modality classifiers enhanced with intra-modal calibration demonstrate improved performance, attributable to more accurate estimation of the uni-modal classifier. When employing inter-modal calibration, we observe significant performance gains compared to uni-modal approaches, owing to the fusion strategy's effectiveness in addressing modality biases. The combination of both intramodal and inter-modal calibration strategies yields the best performance, with comprehensive knowledge integration leading to minimal forgetting rates.", + "bbox": [ + 509, + 465, + 906, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic and covariance-enhanced metric. We examine the effectiveness of Modeling Global Covariance (MGC) and Cross-modal category Nearest Neighbor metric (CMNN) in Table 4. The global covariance modeling strategy enhances base category performance in both the base session and the final incremental session, though at the cost of reduced recognition performance for novel classes, a consequence of the mismatch between global covariance (dominated by base data) and new classes. The semantic nearest neighbor strategy, conversely, improves novel class recognition performance. Implementing both strategies simultaneously without masking further enhances overall performance. Finally, the ensemble inference mechanism with masks fully leverages the advantages of both metrics, achieving superior performance.", + "bbox": [ + 509, + 674, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9887", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/abcb726911d4542274275d41792cf9dc8dad4acfeee6d0ffd7c8cb560e55de39.jpg", + "table_caption": [ + "Table 5. Comparison with three trainable methods on the miniImageNet dataset. $N_{p}$ is the number of parameters which require training. $\\Delta A_{\\mathrm{last}}$ reflects the last task's performance gap between our method and the comparative one." + ], + "table_footnote": [], + "table_body": "
MethodNpAbaseAlastAavg.ΔAlast
CPE-CLIP [5]400k90.2382.7786.13+9.63
CLIP-M3[6]46k96.0092.5094.10-0.10
LP-DiF [10]8.1k96.3491.6893.76+0.72
BiMC†095.4792.4093.600.00
", + "bbox": [ + 94, + 146, + 480, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5e5c32a2306938d04c274f5f4b1f6d572633587d94cedd59476b55792fccafc.jpg", + "image_caption": [ + "Figure 6. CLIP Feature Space Visualization. Squares, circles, and pentagrams represent textual, visual features, and calibrated classifiers respectively. Different colors denote different categories." + ], + "image_footnote": [], + "bbox": [ + 102, + 247, + 472, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Further analysis", + "text_level": 1, + "bbox": [ + 89, + 463, + 254, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with training-required methods. To comprehensively evaluate our approach, we conduct comparisons with three training-based methods [5, 6, 10]. As demonstrated in Table 5, our parameter-free adaptation approach to FSCIL tasks achieves comparable or superior performance to methods that require additional training.", + "bbox": [ + 89, + 488, + 483, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Feature Space. We visualized the image and text feature distributions on the miniImageNet dataset, as depicted in Figure 6. Our analysis showed that in the unified feature space encoded by CLIP, visual and textual features are located at diametrically opposed ends, exemplifying a modality gap phenomenon [16]. Textual features showed more consistent convergence, in contrast to the visual features of the same category, which were more dispersed within the feature space. Notably, despite being distinct, these features share similar representations in the visual subspace (e.g., the two purple dashed ellipses in Figure 6) but are clearly distinguishable in the textual subspace. Furthermore, after modality calibration, the classifiers, influenced by visual priors, demonstrate a shift from the textual modality to the visual modality, thereby alleviating modality bias and enhancing classification capabilities.", + "bbox": [ + 89, + 580, + 483, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of hyper-parameters. Our framework involves two critical hyper-parameters: the inter-modal calibration parameter $\\beta$ and the ensemble classifier weighting coefficient $\\alpha$ , both constrained to [0, 1]. Figure 7a illustrates how performance varies sessions for different values of $\\beta$ .", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35c68906a9f825959cb60b12f63cef715f2e5fae892fb8df54a56ca338c1af74.jpg", + "image_caption": [ + "(a) Influence of $\\beta$ and session-wise accuracy on miniImageNet" + ], + "image_footnote": [], + "bbox": [ + 527, + 90, + 890, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd4e19dacae4c2f23db4f1b41b3841d6a1154a411644a9d4ec0c955313180a82.jpg", + "image_caption": [ + "(b) Influence of $\\alpha$ on miniImageNet" + ], + "image_footnote": [], + "bbox": [ + 522, + 252, + 697, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfe36ef6505c16c1ec2265ce9357aa6f988328cced9a8110867173d92e2427d7.jpg", + "image_caption": [ + "(c) Influence of $\\alpha$ on CIFAR100", + "Figure 7. Analysis on hyper-parameters $\\beta$ and $\\alpha$ . For $\\beta$ , we present heatmaps illustrating accuracy under varying $\\beta$ across different sessions. For $\\alpha$ , we report the average task accuracy $A_{avg}$ and the final task accuracy $A_{last}$ under different settings." + ], + "image_footnote": [], + "bbox": [ + 723, + 251, + 897, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that the $\\beta$ values that perform well in the base session maintain their effectiveness throughout incremental tasks. The optimal solution identified through the validation search in the base session (indicated by the blue rectangular box) closely approaches the global optimum. For the parameter $\\alpha$ , results are presented in Figures 7b and 7c. When $\\alpha = 1$ , the method reduces to a bi-level calibrated classifier. The performance reaches its peak at $\\alpha = 0.6$ , optimizing both the average and final accuracy metrics.", + "bbox": [ + 511, + 435, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 585, + 633, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explore CLIP models in the context of Few-shot Class-Incremental Learning. To mitigate forgetting, we treat the pre-trained model as a black box and employ a simple yet effective training-free bi-level modality calibration strategy. The intra-modal calibration achieves accurate classifier estimates within each modality, while the inter-modal calibration combines knowledge from both visual and textual modalities. An ensemble inference strategy integrating covariance and nearest-neighbor metrics enhances the accuracy of prediction. Extensive experiments validate the effectiveness of our approach.", + "bbox": [ + 511, + 609, + 906, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), National Natural Science Foundation of China (62192783, 62276128), Guangdong Basic and Applied Basic Research Foundation (2024A1515011340) Jiangsu Natural Science Foundation (BK20221441), and the Australian Research Council's Discovery Project (DP220101784).", + "bbox": [ + 511, + 777, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9888", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2", + "[2] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 2", + "[3] Krizhevsky Alex. Learning multiple layers of features from tiny images. https://www.cs.toronto.edu/kriz/learning-features-2009-TR.pdf, 2009.5", + "[4] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1, 2", + "[5] Marco D'Alessandro, Alberto Alonso, Enrique Calabrés, and Mikel Galar. Multimodal parameter-efficient few-shot class incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3393-3403, 2023. 8", + "[6] Thang Doan, Sima Behpour, Xin Li, Wenbin He, Liang Gou, and Liu Ren. A streamlined approach to multimodal few-shot class incremental learning for fine-grained datasets. arXiv preprint arXiv:2403.06295, 2024. 8", + "[7] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2", + "[8] Dipam Goswami, Yuyang Liu, Bartlomiej Twardowski, and Joost van de Weijer. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 6", + "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2", + "[10] Zitong Huang, Ze Chen, Zhixing Chen, Erjin Zhou, Xinxing Xu, Rick Siow Mong Goh, Yong Liu, Chunmei Feng, and Wangmeng Zuo. Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint arXiv:2401.01598, 2024. 8", + "[11] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2", + "[12] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In European Conference on Computer Vision, pages 432-448. Springer, 2022. 2" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11463-11473, 2023. 2", + "[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 1", + "[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9029, 2021. 2", + "[16] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In Advances in Neural Information Processing Systems, 2022. 8", + "[17] Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23638-23647, 2024. 2", + "[18] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2", + "[19] Bolin Ni, Hongbo Zhao, Chenghao Zhang, Ke Hu, Gaofeng Meng, Zhaoxiang Zhang, and Shiming Xiang. Enhancing visual continual learning with language-guided supervision. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 24068-24077, 2024. 2", + "[20] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023. 2", + "[21] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 1, 2", + "[22] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 1, 2, 5", + "[23] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 2, 5", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 3, 5, 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9889", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5", + "[26] Oindrila Saha, Grant Van Horn, and Subhransu Maji. Improved zero-shot classification by adapting vlms with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17552, 2024. 2, 5", + "[27] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 1, 2", + "[28] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2", + "[29] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 5, 6", + "[30] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24183-24192, 2023. 2, 5", + "[31] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 1, 2", + "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[33] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11899-11909, 2023. 2", + "[34] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[35] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 5, 6", + "[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2", + "[38] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2", + "[39] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1", + "[40] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 2", + "[41] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2021. 2, 5", + "[42] Hai Zhang, Junzhe Xu, Shanlin Jiang, and Zhenan He. Simple semantic-aided few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28588–28597, 2024. 2", + "[43] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1", + "[44] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 9046–9056, 2022. 1, 2", + "[45] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2", + "[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1", + "[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1", + "[48] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1" + ], + "bbox": [ + 516, + 90, + 906, + 895 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9890", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_model.json b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5886f4b783abf7289fe6c5d900d9af17a098dd1c --- /dev/null +++ b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.131, + 0.751, + 0.177 + ], + "angle": 0, + "content": "Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.203, + 0.837, + 0.277 + ], + "angle": 0, + "content": "Yiyang Chen\\(^{1}\\), Tianyu Ding\\(^{2}\\), Lei Wang\\(^{3}\\), Jing Huo\\(^{1}\\), Yang Gao\\(^{1}\\), Wenbin Li\\(^{1,4*}\\) \n\\(^{1}\\)State Key Laboratory for Novel Software Technology, Nanjing University, China \n\\(^{2}\\)Applied Sciences Group, Microsoft, USA \\(^{3}\\)University of Wollongong, Australia \n\\(^{4}\\)Shenzhen Research Institute of Nanjing University, Shenzhen, China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.485, + 0.659 + ], + "angle": 0, + "content": "Few-shot Class-Incremental Learning (FSCIL) challenges models to adapt to new classes with limited samples, presenting greater difficulties than traditional class-incremental learning. While existing approaches rely heavily on visual models and require additional training during base or incremental phases, we propose a training-free framework that leverages pre-trained visual-language models like CLIP. At the core of our approach is a novel Bi-level Modality Calibration (BiMC) strategy. Our framework initially performs intra-modal calibration, combining LLM-generated fine-grained category descriptions with visual prototypes from the base session to achieve precise classifier estimation. This is further complemented by inter-modal calibration that fuses pre-trained linguistic knowledge with task-specific visual priors to mitigate modality-specific biases. To enhance prediction robustness, we introduce additional metrics and strategies that maximize the utilization of limited data. Extensive experimental results demonstrate that our approach significantly outperforms existing methods. Code is available at: https://github.com/yychen016/BiMC." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.224, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.484, + 0.879 + ], + "angle": 0, + "content": "The rapid advancement of artificial intelligence has propelled deep neural networks to achieve remarkable progress. However, these models often struggle with catastrophic forgetting when faced with dynamically changing environments, impeding their ability to learn sequential tasks effectively. Class-Incremental Learning (CIL) can be a potential solution to this challenge. Yet, in practical settings, researchers frequently encounter a scarcity of training data during the incremental phase, leading to sample sparsity issues. Consequently, Few-Shot Class-Incremental Learning (FSCIL) [31] has recently emerged as a more realistic scenario, better reflecting real-world constraints." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.528, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Model updating paradigms in FSCIL. (a) Continuous model updates during incremental phases. (b) Leveraging the base task to learn a generalizable model. (c) Our proposed framework: training-free in both base and incremental sessions." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.597, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Current FSCIL methods can be broadly categorized into two paradigms: the continual fine-tuning paradigm and the incremental-session freezing paradigm, as shown in Figure 1 (a) and (b). The former [4, 21, 27, 43] utilizes limited samples to continuously update the model, enhancing its plasticity. Conversely, the latter [22, 35, 44, 48] focuses on learning robust representations during the base task, which are then generalized to incremental tasks, prioritizing stability by avoiding subsequent model updates. Despite their distinct approaches, both paradigms necessitate additional training to adapt to new tasks, presenting ongoing challenges in resource efficiency and model adaptability." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The emergence of large-scale models has shifted the landscape of representation learning. Vision-language models like CLIP [24], trained on a large number of image-text pairs, excel in zero-shot tasks. However, fine-tuning is computationally expensive and may compromise generalization, especially with limited data [14, 39]. While recent approaches like prompt tuning [46, 47] and low-rank adaptation [9] offer more efficient alternatives, they still demand" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "additional training. This raises a crucial question: How can we effectively leverage pre-trained knowledge and domain-specific visual priors to adapt the large-scale model using few samples without additional training?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.153, + 0.486, + 0.575 + ], + "angle": 0, + "content": "While CLIP is widely used for cross-modal retrieval in visual categorization, it overlooks the potential of intramodal recognition, akin to traditional purely visual models, such as CNN [30, 41]. We believe that CLIP's visual prototypes can effectively complement zero-shot textual classifiers, serving as domain-specific priors, particularly in addressing two major challenges in FSCIL tasks: catastrophic forgetting and overfitting. Building on this insight, we introduce Bi-level Modality Calibration (BiMC), a simple yet effective framework that treats CLIP as a black-box model and operates without additional training, thereby avoiding the parameter updates that typically cause forgetting. Our framework incorporates both intra-modal and inter-modal calibration strategies to enhance CLIP's classification accuracy. For intra-modal calibration, we leverage LLM-generated fine-grained category descriptions in the textual domain to improve zero-shot classifier discrimination, while using base session prototypes in the visual domain to refine new class prototypes affected by sample scarcity. Inter-modal calibration combines pre-trained linguistic knowledge with task-aware visual priors to reduce modal biases and overfitting. The framework is further enhanced with an anisotropic covariance metric and a cross-modal category-conditioned nearest-neighbor metric, implementing the final classifier through masked ensemble inference. Extensive experiments demonstrate that our method achieves competitive performance without additional training." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.576, + 0.414, + 0.589 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.591, + 0.483, + 0.651 + ], + "angle": 0, + "content": "- We propose a novel training-free FSCIL framework that achieves continual model adaptation by treating vision-language models as black boxes, significantly enhancing practicality for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.658, + 0.483, + 0.749 + ], + "angle": 0, + "content": "- We develop a bi-level modality calibration approach combining intra-modal and inter-modal strategies to enhance classifier accuracy across both modalities. This approach is strengthened by an innovative visual covariance metric and a category-conditioned nearest-neighbor metric, culminating in a robust masked ensemble inference strategy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.755, + 0.483, + 0.83 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of our framework through extensive experiments on standard benchmarks, where it not only achieves competitive performance but also outperforms supervised training methods in several scenarios, despite requiring no additional training." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.591, + 0.483, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.229, + 0.86 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Few-Shot Class-Incremental Learning. Few-Shot Class-Incremental Learning (FSCIL) challenges models to adapt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "continuously using minimal samples. Existing approaches can be categorized into three main groups: representation-based methods [1, 2, 12, 18, 22, 27, 30, 38, 44], which focus on optimizing feature representations; dynamic-architecture-based methods [31, 41], which adapt model structures to accommodate new classes; and knowledge-distillation-based methods [4, 7, 15], which transfer knowledge from previous tasks to new ones. Notable examples include CEC [41], which employs an evolving graph model for classifier optimization, and FACT [44] and SAVC [30], which introduce virtual categories as placeholders for forward compatibility. In contrast, our approach leverages a pre-trained vision-language model, eliminating the need for additional learning on the base session." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.485 + ], + "angle": 0, + "content": "Class-Incremental Learning via Pre-Trained Models. Recent trends in incremental learning focus on effectively utilizing pre-trained models as backbones for downstream tasks. Early works primarily used prompt-based methods [11, 28, 34, 36, 37], introducing learnable prompt parameters for task adaptation. Other approaches have employed adapters [45] or LoRA [9] to facilitate adaptation [17, 40, 45]. While these methods use efficient parameter-tuning techniques, they still require further training. Our approach, however, leverages both vision and language modalities to enable ongoing model adaptation without gradient updates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.905, + 0.742 + ], + "angle": 0, + "content": "Language-Assisted Classification. Linguistic information has been increasingly utilized in various visual tasks. Some studies have used linguistic cues to guide visual classification [33, 42], while others have leveraged Large Language Models (LLMs) to generate rich descriptive text [20, 23, 26]. In continual learning, several works have exploited linguistic guidance [13, 19]. Within FSCIL, methods like [2, 4, 21] also utilize linguistic information. However, these approaches have limitations: some [2] use distillation techniques on smaller visual models with linguistic knowledge as a regularization term, potentially limiting scalability and flexibility. Others [4, 21] rely on complex and often redundant architectural designs, making implementation challenging and inefficient. In contrast, our approach addresses these limitations by proposing a simpler, more efficient framework that effectively integrates linguistic and visual information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.759, + 0.605, + 0.774 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.642, + 0.801 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.809, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In FSCIL, a model continuously adapts to a sequence of sessions \\( \\mathcal{S} = \\{\\mathcal{D}_0, \\mathcal{D}_1, \\dots, \\mathcal{D}_T\\} \\). Each session \\( \\mathcal{D}_t = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{N_t} \\) consists of training data with a category space \\( \\mathcal{C}_t^{\\mathrm{train}} = Y_t \\), where \\( N_t \\) represents the number of samples in the \\( t \\)-th session. The base session \\( \\mathcal{D}_0 \\) contains a substantial number of categories and samples, while subsequent incre" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.098, + 0.615, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.396, + 0.483, + 0.408 + ], + "angle": 0, + "content": "(a) Bi-Level Modality Calibration Framework" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.098, + 0.891, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.396, + 0.892, + 0.408 + ], + "angle": 0, + "content": "(b) Semantic and Covariance-enhanced Metric" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.411, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Figure 2. Overview of our framework. In each task adaptation phase, we construct a calibrated classifier through our proposed bi-level calibration framework, which involves intra-modal calibration and inter-modal calibration. To enhance performance, we introduce a globally shared covariance metric for visual feature modeling, complemented by a category-conditioned nearest-neighbor scoring strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.492 + ], + "angle": 0, + "content": "mental sessions \\(\\mathcal{D}_t\\) for \\(t > 0\\) contain only a few training samples per category, formally expressed as \\(N_0\\gg N_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.614 + ], + "angle": 0, + "content": "The categories across different sessions are mutually exclusive, meaning that for any different \\(i,j\\in [0,T]\\) \\(Y_{i}\\cap Y_{j} = \\varnothing\\) . Furthermore, during the training of the current session, data from previous sessions is not accessible. At the testing phase of the \\(t\\) -th session, the model evaluates against all previously encountered categories, with the cumulative category set defined as \\(\\mathcal{V}_t = \\bigcup_{j = 0}^t Y_j\\) . Consequently, the test category space is given by \\(\\mathcal{C}_t^{\\mathrm{test}} = \\mathcal{V}_t\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.371, + 0.637 + ], + "angle": 0, + "content": "3.2. Bi-Level calibration framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Our proposed training-free framework builds upon CLIP [24] and incorporates two essential components: intra-modal and inter-modal calibration (see Figure 2a)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.376, + 0.725 + ], + "angle": 0, + "content": "3.2.1 Intra-modal classifier calibration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In its standard usage, CLIP relies on simple prompt templates, (e.g., \"a photo of a [CLS].\") to construct zero-shot classifiers for visual tasks. However, this category-agnostic approach has inherent limitations. The generic nature of these templates fails to capture category-specific nuances, making them particularly inadequate for fine-grained classification tasks where detailed feature discrimination is crucial. A single, static template cannot effectively characterize the diverse visual characteristics within each category, especially when subtle distinctions between categories are essential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.582 + ], + "angle": 0, + "content": "To overcome these limitations, we leverage Large Language Models (LLMs) to generate dynamic, category-specific descriptions. These LLM-generated descriptions provide richer, more discriminative semantic representations for each category, enhancing the model's capacity to capture fine-grained features. We integrate this enhanced semantic information through an intra-modal calibration strategy within the textual modality:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.589, + 0.907, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {\\mu}} _ {c} ^ {T} = \\left(1 - \\lambda_ {T}\\right) \\boldsymbol {w} _ {c} + \\lambda_ {T} \\left(\\frac {1}{n _ {c}} \\sum_ {j = 1} ^ {n _ {c}} \\frac {g \\left(\\mathbf {t} _ {c , j}\\right)}{\\| g \\left(\\mathbf {t} _ {c , j}\\right) \\| _ {2}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.825 + ], + "angle": 0, + "content": "In this formulation, \\(\\lambda_{T}\\) controls the intensity of intra-modal calibration within the text modality, \\(n_c\\) denotes the number of descriptions for category \\(c\\), \\(g(\\cdot)\\) represents the text encoder, \\(\\mathbf{t}_{c,j}\\) is the \\(j\\)-th LLM-generated description for category \\(c\\), \\(\\tilde{\\boldsymbol{\\mu}}_c^T\\) is the calibrated textual prototype, and \\(\\boldsymbol{w}_c\\) denotes the original CLIP Zero-Shot classifier weight for class \\(c\\). This calibration mechanism preserves the robust generalization capabilities of the CLIP Zero-Shot classifier while incorporating fine-grained, category-specific semantic information. The latter term in Eq. (1) serves as a semantically enriched category description center, enhancing the model's discriminative power." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the FSCIL context, prototype-based classifiers are commonly employed, where the accuracy of prototype estimation directly impacts performance. During the base session, the abundance of samples allows for precise estimation of class prototypes. However, in incremental sessions" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "that introduce new classes, the limited number of samples inevitably leads to biased estimates of new class prototypes. Drawing inspiration from TEEN [35], we leverage the accurately estimated prototype weights of base classes to facilitate intra-modality calibration on the prototypes of incoming new classes within the visual modality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.214 + ], + "angle": 0, + "content": "For a given class \\( c \\), we first utilize the encoded training image data to obtain the naive visual prototype:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.222, + 0.483, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {c} ^ {I} = \\frac {1}{m _ {c}} \\sum_ {j = 1} ^ {m _ {c}} \\frac {f (\\mathbf {x} _ {c , j})}{\\| f (\\mathbf {x} _ {c , j}) \\| _ {2}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.484, + 0.378 + ], + "angle": 0, + "content": "where \\( f(\\cdot) \\) is the visual encoder, \\( m_{c} \\) is the number of images in class \\( c \\), and \\( \\mathbf{x}_{c,j} \\) is the \\( j \\)-th image belonging to category \\( c \\). Due to the inaccuracy in visual prototype estimation during the incremental sessions, the prototypes from the base session are used to recalibrate the new class prototypes by leveraging the similarity relationship between them. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.386, + 0.483, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I} = \\left\\{ \\begin{array}{l l} \\boldsymbol {\\mu} _ {c} ^ {I} & , t = 0 \\\\ (1 - \\lambda_ {I}) \\boldsymbol {\\mu} _ {c} ^ {I} + \\lambda_ {I} \\sum_ {b = 1} ^ {| \\mathcal {Y} _ {0} |} s _ {b, c} \\boldsymbol {\\mu} _ {b} ^ {I} & , t > 0, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.572 + ], + "angle": 0, + "content": "where, \\(t\\) represents the task identifier, \\(\\lambda_I\\) controls the strength of the visual intra-modal calibration, and \\(s_{b,c}\\) denotes the normalized cosine similarity between the visual prototypes of class \\(b\\) (from the base classes) and class \\(c\\). The similarity \\(s_{b,c}\\) is defined as: \\(s_{b,c} = \\frac{e^{\\tau \\cdot \\langle \\mu_b, \\mu_c \\rangle}}{\\sum_{i=1}^{|\\mathcal{Y}_0|} e^{\\tau \\cdot \\langle \\mu_i, \\mu_c \\rangle}}\\). Here, \\(\\langle \\cdot, \\cdot \\rangle\\) represents the cosine similarity, and \\(\\tau\\) is the temperature scaling parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.484, + 0.649 + ], + "angle": 0, + "content": "Through this visual intra-modality calibration process, the prototypes of the new classes are adjusted to align more closely with the well-calibrated base class prototypes, allowing the new class prototypes to partially inherit the discriminative properties of the base class prototypes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.375, + 0.68 + ], + "angle": 0, + "content": "3.2.2 Inter-modal classifier calibration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Following intra-modal calibration, two classifier achieves enhanced accuracy within its respective modality. However, we posit that their full potential remains unrealized when operating in isolation. This is due to the fact that classifiers operating exclusively on linguistic data derive their knowledge from pre-trained models and lack explicit downstream knowledge for visual tasks. Conversely, classifiers functioning solely within the visual domain serve as metrics within the visual sub-feature space but are deficient in semantic understanding. Recognizing the complementary nature of these two modalities, we propose a training-free inter-modal calibration strategy, allowing domain-relevant downstream visual prior to guide domain-agnostic pretrained linguistic knowledge. Formally, this process can be" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.604, + 0.105 + ], + "angle": 0, + "content": "described as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.115, + 0.907, + 0.134 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {c} = \\beta \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {T} + (1 - \\beta) \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.143, + 0.907, + 0.204 + ], + "angle": 0, + "content": "Here, \\(\\pmb{\\mu}_{c}\\) is the inter-modal calibrated classifier, and \\(\\beta\\) is the calibration coefficient. During inference, we use cosine similarity to calculate the similarity between the sample features \\(f(\\mathbf{x})\\) and the mixed category center:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.213, + 0.907, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {c} ^ {\\text {c a l i b}} = \\frac {f (\\mathbf {x}) ^ {\\top} \\boldsymbol {\\mu} _ {c}}{\\| f (\\mathbf {x}) \\| _ {2} \\cdot \\| \\boldsymbol {\\mu} _ {c} \\| _ {2}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.266, + 0.907, + 0.356 + ], + "angle": 0, + "content": "This inter-modal calibration strategy effectively combines the strengths of both linguistic and visual modalities. It helps to enhance the overall classification performance by creating a more robust and comprehensive classifier that can better capture the nuances of visual tasks while maintaining a strong semantic understanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.365, + 0.875, + 0.38 + ], + "angle": 0, + "content": "3.3. Semantic and covariance-enhanced metric" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.906, + 0.509 + ], + "angle": 0, + "content": "While our previous metrics effectively integrate linguistic and visual information and mitigate modal bias, they primarily measure sample distances from distribution centers. This isotropic approach, despite its simplicity, struggles to capture higher-order information and complex data relationships. To address these limitations, we propose leveraging statistical information from existing data for more comprehensive and efficient measurements, as shown in Figure 2b." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.527, + 0.761, + 0.542 + ], + "angle": 0, + "content": "3.3.1 Modeling global covariance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.55, + 0.906, + 0.641 + ], + "angle": 0, + "content": "Inspired by FeCAM [8], we introduce an anisotropic covariance metric to fully utilize the visual modality's statistical information. For a session \\( t \\), we compute its covariance matrix \\( \\boldsymbol{\\Sigma}^t \\) using image embeddings \\( f(\\mathbf{x}) \\) of all image data relevant to the session. To avoid singularity, we apply regularization with an identity matrix \\( I_d \\):" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.649, + 0.907, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {\\Sigma}} ^ {t} = \\boldsymbol {\\Sigma} ^ {t} + \\frac {\\gamma}{d} \\operatorname {t r} \\left(\\boldsymbol {\\Sigma} ^ {t}\\right) \\boldsymbol {I} _ {d}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.684, + 0.907, + 0.774 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a hyperparameter controlling the strength of regularization, \\(d\\) represents the feature dimension. Here, \\(\\mathrm{tr}(\\cdot)\\) denotes the trace operator of a matrix, and both \\(d\\) and \\(\\mathrm{tr}(\\pmb{\\Sigma}^t)\\) provide a prior estimation for the regularization strength. Subsequently, we construct a continuously evolving shared covariance matrix as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.783, + 0.907, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {t} = \\frac {| \\mathcal {Y} _ {t - 1} |}{| \\mathcal {Y} _ {t} |} \\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {t - 1} + \\left(1 - \\frac {| \\mathcal {Y} _ {t - 1} |}{| \\mathcal {Y} _ {t} |}\\right) \\tilde {\\boldsymbol {\\Sigma}} ^ {t}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.906, + 0.869 + ], + "angle": 0, + "content": "The evolved covariance matrix captures the statistical information of all categories. Given a test sample \\(\\mathbf{x}\\), we calculate its score as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.876, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {c} ^ {\\mathrm {c o v}} (\\mathbf {x}) = - \\frac {1}{d} \\left(f (\\mathbf {x}) - \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}\\right) ^ {T} \\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {- 1} \\left(f (\\mathbf {x}) - \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}\\right). \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9884" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.476, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.223, + 0.483, + 0.252 + ], + "angle": 0, + "content": "Figure 3. In the inference phase, an ensemble strategy with category masking is utilized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.362 + ], + "angle": 0, + "content": "The division by feature dimension \\( d \\) serves as a normalization factor, preventing the softmax transformation into probabilities from collapsing into a single category. The negative sign maintains metric consistency, ensuring that a larger output score indicates a higher probability of the sample belonging to category \\( c \\). Here, \\( \\hat{\\Sigma}_G \\) denotes the global covariance, with the superscript \\( t \\) omitted for clarity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.383, + 0.468, + 0.399 + ], + "angle": 0, + "content": "3.3.2 Cross-modal category nearest neighbor metric" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Global covariance modeling effectively utilizes statistical information from the visual modality. To better leverage LLM-generated category descriptions, we propose a category-conditioned nearest-neighbor-based metric. Given a description text \\(\\mathbf{t}_{c,j}\\) encoded by \\(g(\\cdot)\\), we obtain its normalized feature representation \\(\\mathbf{z}_{c,j} = \\frac{g(\\mathbf{t}_{c,j})}{\\|g(\\mathbf{t}_{c,j})\\|_2}\\). For a visual query sample, we compute its normalized representation \\(\\mathbf{v} = \\frac{f(\\mathbf{x})}{\\|f(\\mathbf{x})\\|_2}\\). The likelihood of this sample belonging to category \\(c\\) is determined by the maximum dot product between \\(\\mathbf{v}\\) and all description features in class \\(c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.577, + 0.483, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {c} ^ {\\mathrm {n n}} = \\max _ {j} \\left\\{\\boldsymbol {z} _ {c, j} ^ {\\top} \\cdot \\boldsymbol {v} \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.675 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}_c^{\\mathrm{nn}}\\) indicates the likelihood that the query sample belongs to category \\(c\\). This metric leverages maximum cross-modal similarity, effectively utilizing the diverse textual descriptions generated by LLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.418, + 0.712 + ], + "angle": 0, + "content": "3.3.3 Inference score reorganization strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We employ a globally shared covariance matrix to model distribution shape information in the visual feature space. While the base session contains substantial data for accurate covariance matrix estimation, experiments reveal that these base-derived covariance matrices perform poorly in distinguishing new class data. This limitation stems from two factors: (1) the covariance from extensive base data fails to align with new classes, and (2) the limited data available for new classes prevents accurate covariance estimation. To address this, we use the cross-modal category nearest neighbor metric for novel classes to compensate for the limitations of the second-order covariance metric, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "We first normalize the scores of three metrics using the softmax function: \\(\\mathbf{p}^{\\mathrm{calib}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{calib}}\\right)\\), \\(\\mathbf{p}^{\\mathrm{cov}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{cov}}\\right)\\), and \\(\\mathbf{p}^{\\mathrm{nn}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{nn}}\\right)\\). The final probability score \\(\\mathbf{p}_c\\) is computed differently for base and novel categories:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.178, + 0.906, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {c} = \\left\\{ \\begin{array}{l l} \\alpha \\mathbf {p} _ {c} ^ {\\text {c a l i b}} + (1 - \\alpha) \\mathbf {p} _ {c} ^ {\\text {c o v}}, & c \\in \\mathcal {Y} _ {0} \\\\ \\alpha \\mathbf {p} _ {c} ^ {\\text {c a l i b}} + (1 - \\alpha) \\mathbf {p} _ {c} ^ {\\text {n n}}, & c \\notin \\mathcal {Y} _ {0}. \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.906, + 0.321 + ], + "angle": 0, + "content": "For base categories \\((c\\in \\mathcal{V}_0)\\), we use a weighted sum of mixed prototype scores and visual mahalanobis metric scores. For novel categories \\((c\\notin \\mathcal{V}_0)\\), we combine mixed prototype scores with cross-modal nearest neighbor metric scores. The final category prediction for a sample is determined by \\(\\arg \\max_c\\mathbf{p}_c\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.335, + 0.646, + 0.351 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.719, + 0.377 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.906, + 0.519 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three benchmarks following [22, 30, 41]: CIFAR100 [3], CUB200-2011 [32], and miniImageNet [25]. For CIFAR100 and miniImageNet, we partition each dataset into 60 base classes and 40 novel classes. The novel classes are further divided into eight incremental tasks, with each task structured as a 5-way 5-shot incremental session. The CUB200 dataset is split into 100 base classes and 100 novel classes, with each incremental task organized in a 10-way 5-shot format." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Implementation details. We adopt CLIP's ViT-B/16 architecture as the backbone for fair comparison across all methods. For category descriptions, we utilize fine-grained text generated by CuPL [23] and [26]. The hyperparameters are configured with intra-modal calibration coefficients \\(\\lambda_{T} = 0.5\\) (textual), and \\(\\lambda_{I} = 0.1\\) (visual), and a temperature parameter \\(\\tau = 16\\). The inter-modal calibration coefficient \\(\\beta\\) is optimized through a validation process. Specifically, we divide the base session into training and validation subsets. Applying our bi-level calibration framework on the training subset, we evaluate performance on the validation subset. By incrementing \\(\\beta\\) by 0.05, we identify the value that maximizes validation accuracy. This optimal \\(\\beta\\) remains fixed throughout subsequent incremental sessions. The shared covariance parameter \\(\\gamma\\) is set to 1 for the base session and adjusted to 5 for incremental sessions. Finally, for score reorganization, we set \\(\\alpha = 0.6\\) for CIFAR100 and miniImageNet, and \\(\\alpha = 0.8\\) for CUB200." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.702, + 0.818 + ], + "angle": 0, + "content": "4.2. Comparison results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We compare our method with training-free approaches in the FSCIL scenario: CLIP Zero-Shot [24], Visual Prototype [29], TEEN [35] and FeCAM [8]. We present the performance curve over all benchmarks in Figure 4, while detailed session-wise performance on miniImageNet is shown" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9885" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Detailed session-wise accuracy, average accuracy (Avg) and performance degradation (PD) comparison on miniImageNet dataset. V and L represent the visual and language modalities, respectively. BiMC refers to the results obtained solely through the bi-level calibration framework, whereas \\(\\mathrm{BiMC}^{\\dagger}\\) incorporates the ensemble classifier strategy. The best results on each each sessions are indicated in bold, while the second-best results are underlined. \\(\\uparrow\\) means higher is better, while \\(\\downarrow\\) means lower is better." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.147, + 0.895, + 0.284 + ], + "angle": 0, + "content": "
MethodModalityAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
CLIP Zero-Shot [24]L91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
Visual Prototype [29]V92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
TEEN [35]V92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
FeCAM [8]V94.4293.7191.6190.5990.6290.2288.4388.1788.2090.666.22
BiMCV-L94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
BiMC†V-L95.4795.3493.8093.6393.5193.1892.5192.5692.4093.603.07
" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.296, + 0.371, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.447, + 0.303, + 0.459 + ], + "angle": 0, + "content": "(a) miniImageNet" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.296, + 0.627, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.447, + 0.551, + 0.459 + ], + "angle": 0, + "content": "(b) CIFAR100" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.297, + 0.886, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.447, + 0.802, + 0.459 + ], + "angle": 0, + "content": "(c) CUB200" + }, + { + "type": "image_footnote", + "bbox": [ + 0.191, + 0.465, + 0.807, + 0.477 + ], + "angle": 0, + "content": "CLIP Zero-Shot Visual Prototype FeCAM BiMC BiMC" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.482, + 0.856, + 0.496 + ], + "angle": 0, + "content": "Figure 4. Performance curve of each incremental sessions on (a) miniImageNet, (b) CIFAR100 and (c) CUB200 datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.482, + 0.534 + ], + "angle": 0, + "content": "in Table 1. Detailed results for CIFAR100 and CUB200 are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Our approach demonstrates competitive performance across all three datasets and significantly outperforms unimodal baselines without requiring additional training, highlighting the importance of leveraging information from both modalities. On the miniImageNet dataset, while visual and textual prototype classifiers achieve comparable results, and although TEEN [35] and FeCAM [8] better utilize visual modality information, their improvement remains limited. Our framework employs a simple calibration strategy to integrate domain visual knowledge with pretrained linguistic knowledge, yielding up to \\(6\\%\\) performance improvement over uni-modal models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.483, + 0.854 + ], + "angle": 0, + "content": "For CIFAR100, the language prototype shows better resistance to forgetting compared to the visual one, which is more susceptible to sequential task learning. This can be attributed to the dataset's low-resolution characteristics, resulting in insufficient visual information during incremental phases. Our calibration framework achieves approximately \\(3.5\\%\\) improvement over a language-only classifier. In contrast, for the fine-grained CUB200 dataset, the visual prototype surpasses the language prototype due to limited trans" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.506, + 0.71, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.638, + 0.69, + 0.65 + ], + "angle": 0, + "content": "(a) Old Classes Entropy" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.506, + 0.899, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.638, + 0.885, + 0.65 + ], + "angle": 0, + "content": "(b) Novel Classes Entropy" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.654, + 0.905, + 0.746 + ], + "angle": 0, + "content": "Figure 5. Comparison of output confidence with different models. ferability of pre-trained language knowledge. By combining both modalities, our approach achieves approximately \\(3.5\\%\\) performance improvement. Notably, our framework demonstrates robust resistance to forgetting, achieving the lowest performance drop across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.837 + ], + "angle": 0, + "content": "One can observe that each modality's contribution to classification varies across datasets. Our proposed bi-level calibration framework effectively leverages the strengths of each modality and mitigates modality bias. Moreover, our ensemble inference strategy, denoted as \\(\\mathrm{BiMC}^{\\dagger}\\), provides additional performance improvements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.89, + 0.864 + ], + "angle": 0, + "content": "4.3. Investigation of bi-level calibrated classifiers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In this section, we analyze two distinct attributes of the calibrated classifier to demonstrate the effectiveness of our pro" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1“textual prototype” refers to the classifier generated by CLIP in a zero-shot setting, which has the same meaning as “textual classifier” mentioned later in the paper." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9886" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 2. Ablation studies of bi-level calibration framework on miniImageNet. Vis. and Lang. individually represent classifiers derived from the visual and language modalities. Intra-C and Inter-C refer to the strategies of intra-modal and inter-modal calibration, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.12, + 0.899, + 0.248 + ], + "angle": 0, + "content": "
VisLangIntra-CInter-CAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
✓92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
✓✓92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
✓91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
✓✓93.2793.3291.9191.5191.2190.7389.2889.2389.0191.054.26
✓✓✓94.6894.5892.8792.4992.5692.3191.5991.5891.5492.693.14
✓✓✓✓94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.316 + ], + "angle": 0, + "content": "Table 3. The left side of the table lists classification result combinations from three classifiers, while the right side shows the percentage \\((\\%)\\) of test samples meeting each combination. Test sets from both the base and the final session were used for analysis." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.318, + 0.472, + 0.419 + ], + "angle": 0, + "content": "
ModalityCIFAR100CUB200
VLV-LD0testD8testD0testD10test
✓✓✓65.3351.8159.0044.85
X✓✓6.9513.072.064.00
✓X✓6.105.5619.8321.40
XX✓1.202.101.081.98
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.214, + 0.442 + ], + "angle": 0, + "content": "posed framework." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.443, + 0.483, + 0.594 + ], + "angle": 0, + "content": "The bi-level calibrated classifier enhances prediction confidence. We evaluate prediction confidence by calculating the entropy of predictions for each classifier in the final session across all three datasets. The entropy serves as an uncertainty indicator for model outputs, and we compute it separately for both base and novel categories, as detailed in Figure 5. The findings indicate that the bi-level calibrated classifier significantly reduces the entropy of predictions compared to uni-model classifiers, thereby enhancing prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.807 + ], + "angle": 0, + "content": "The bi-level calibrated classifier mitigates modality bias. To elucidate the superior performance of our method, we analyze the joint prediction outcomes across different classifiers in comparison to uni-modal approaches. Using \\(\\checkmark\\) to denote correct classifications and \\(x\\) for classification errors, we observe that the calibrated classifier maintains accurate predictions even in cases where uni-modal classifiers fail, as demonstrated in the \\(2^{\\mathrm{nd}}\\) and \\(3^{\\mathrm{rd}}\\) rows of Table 3. This evidence suggests that the bi-level calibrated classifier produces more accurate and unbiased predictions, offering enhanced robustness compared to uni-modality classifiers. Remarkably, in certain rare instances, the modality-calibrated classifier successfully classifies samples that both uni-modal classifiers misclassify." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.242, + 0.833 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.885 + ], + "angle": 0, + "content": "We conduct comprehensive ablation analyses on the miniImageNet dataset to evaluate the significance of each component within our proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Bi-level calibration framework. Our framework com-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.259, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Table 4. Ablation study of Semantic and Covariance-enhanced Metric on the miniImageNet dataset. To verify their effectiveness, we measured the performance of the base task \\(\\mathcal{A}_{\\mathrm{base}}\\), the performance of last task \\(\\mathcal{A}_{\\mathrm{last}}\\) and the average performance \\(\\mathcal{A}_{\\mathrm{avg.}}\\) across all tasks. Additionally, we report the accuracy of base \\(\\mathcal{A}_{\\mathrm{last}}^b\\) and novel \\(\\mathcal{A}_{\\mathrm{last}}^n\\) class in last task." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.345, + 0.902, + 0.442 + ], + "angle": 0, + "content": "
Ablation\\( A_{base} \\)\\( A_{last} \\)\\( A_{last}^{b} \\)\\( A_{last}^{n} \\)\\( A_{avg} \\)
BiMC94.9091.8193.4389.3892.97
+ MGC95.4791.5194.3787.2292.91
+ CMNN94.8592.1593.3090.4293.18
\\( BiMC^† (w/o mask.) \\)95.4792.2094.1289.3293.37
\\( BiMC^† \\)95.4792.4093.2091.2093.60
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.467, + 0.907, + 0.664 + ], + "angle": 0, + "content": "prises four components: visual modality information, linguistic modality information, intra-modal calibration strategy, and inter-modal calibration strategy. As shown in Table 2, single-modality classifiers enhanced with intra-modal calibration demonstrate improved performance, attributable to more accurate estimation of the uni-modal classifier. When employing inter-modal calibration, we observe significant performance gains compared to uni-modal approaches, owing to the fusion strategy's effectiveness in addressing modality biases. The combination of both intramodal and inter-modal calibration strategies yields the best performance, with comprehensive knowledge integration leading to minimal forgetting rates." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Semantic and covariance-enhanced metric. We examine the effectiveness of Modeling Global Covariance (MGC) and Cross-modal category Nearest Neighbor metric (CMNN) in Table 4. The global covariance modeling strategy enhances base category performance in both the base session and the final incremental session, though at the cost of reduced recognition performance for novel classes, a consequence of the mismatch between global covariance (dominated by base data) and new classes. The semantic nearest neighbor strategy, conversely, improves novel class recognition performance. Implementing both strategies simultaneously without masking further enhances overall performance. Finally, the ensemble inference mechanism with masks fully leverages the advantages of both metrics, achieving superior performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9887" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.146 + ], + "angle": 0, + "content": "Table 5. Comparison with three trainable methods on the miniImageNet dataset. \\( N_{p} \\) is the number of parameters which require training. \\( \\Delta A_{\\mathrm{last}} \\) reflects the last task's performance gap between our method and the comparative one." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.147, + 0.482, + 0.232 + ], + "angle": 0, + "content": "
MethodNpAbaseAlastAavg.ΔAlast
CPE-CLIP [5]400k90.2382.7786.13+9.63
CLIP-M3[6]46k96.0092.5094.10-0.10
LP-DiF [10]8.1k96.3491.6893.76+0.72
BiMC†095.4792.4093.600.00
" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.248, + 0.473, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.416, + 0.485, + 0.457 + ], + "angle": 0, + "content": "Figure 6. CLIP Feature Space Visualization. Squares, circles, and pentagrams represent textual, visual features, and calibrated classifiers respectively. Different colors denote different categories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.255, + 0.481 + ], + "angle": 0, + "content": "4.5. Further analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Comparison with training-required methods. To comprehensively evaluate our approach, we conduct comparisons with three training-based methods [5, 6, 10]. As demonstrated in Table 5, our parameter-free adaptation approach to FSCIL tasks achieves comparable or superior performance to methods that require additional training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Visualization of Feature Space. We visualized the image and text feature distributions on the miniImageNet dataset, as depicted in Figure 6. Our analysis showed that in the unified feature space encoded by CLIP, visual and textual features are located at diametrically opposed ends, exemplifying a modality gap phenomenon [16]. Textual features showed more consistent convergence, in contrast to the visual features of the same category, which were more dispersed within the feature space. Notably, despite being distinct, these features share similar representations in the visual subspace (e.g., the two purple dashed ellipses in Figure 6) but are clearly distinguishable in the textual subspace. Furthermore, after modality calibration, the classifiers, influenced by visual priors, demonstrate a shift from the textual modality to the visual modality, thereby alleviating modality bias and enhancing classification capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Analysis of hyper-parameters. Our framework involves two critical hyper-parameters: the inter-modal calibration parameter \\(\\beta\\) and the ensemble classifier weighting coefficient \\(\\alpha\\), both constrained to [0, 1]. Figure 7a illustrates how performance varies sessions for different values of \\(\\beta\\)." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.092, + 0.891, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.232, + 0.877, + 0.246 + ], + "angle": 0, + "content": "(a) Influence of \\(\\beta\\) and session-wise accuracy on miniImageNet" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.253, + 0.699, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.357, + 0.706, + 0.369 + ], + "angle": 0, + "content": "(b) Influence of \\(\\alpha\\) on miniImageNet" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.252, + 0.898, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.356, + 0.898, + 0.369 + ], + "angle": 0, + "content": "(c) Influence of \\(\\alpha\\) on CIFAR100" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.373, + 0.906, + 0.43 + ], + "angle": 0, + "content": "Figure 7. Analysis on hyper-parameters \\(\\beta\\) and \\(\\alpha\\). For \\(\\beta\\), we present heatmaps illustrating accuracy under varying \\(\\beta\\) across different sessions. For \\(\\alpha\\), we report the average task accuracy \\(A_{avg}\\) and the final task accuracy \\(A_{last}\\) under different settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.572 + ], + "angle": 0, + "content": "We observe that the \\(\\beta\\) values that perform well in the base session maintain their effectiveness throughout incremental tasks. The optimal solution identified through the validation search in the base session (indicated by the blue rectangular box) closely approaches the global optimum. For the parameter \\(\\alpha\\), results are presented in Figures 7b and 7c. When \\(\\alpha = 1\\), the method reduces to a bi-level calibrated classifier. The performance reaches its peak at \\(\\alpha = 0.6\\), optimizing both the average and final accuracy metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.634, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.777 + ], + "angle": 0, + "content": "We explore CLIP models in the context of Few-shot Class-Incremental Learning. To mitigate forgetting, we treat the pre-trained model as a black box and employ a simple yet effective training-free bi-level modality calibration strategy. The intra-modal calibration achieves accurate classifier estimates within each modality, while the inter-modal calibration combines knowledge from both visual and textual modalities. An ensemble inference strategy integrating covariance and nearest-neighbor metrics enhances the accuracy of prediction. Extensive experiments validate the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.909, + 0.899 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), National Natural Science Foundation of China (62192783, 62276128), Guangdong Basic and Applied Basic Research Foundation (2024A1515011340) Jiangsu Natural Science Foundation (BK20221441), and the Australian Research Council's Discovery Project (DP220101784)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9888" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.185 + ], + "angle": 0, + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.187, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.245, + 0.483, + 0.286 + ], + "angle": 0, + "content": "[3] Krizhevsky Alex. Learning multiple layers of features from tiny images. https://www.cs.toronto.edu/kriz/learning-features-2009-TR.pdf, 2009.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[4] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[5] Marco D'Alessandro, Alberto Alonso, Enrique Calabrés, and Mikel Galar. Multimodal parameter-efficient few-shot class incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3393-3403, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.445, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[6] Thang Doan, Sima Behpour, Xin Li, Wenbin He, Liang Gou, and Liu Ren. A streamlined approach to multimodal few-shot class incremental learning for fine-grained datasets. arXiv preprint arXiv:2403.06295, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.502, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[7] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.574, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[8] Dipam Goswami, Yuyang Liu, Bartlomiej Twardowski, and Joost van de Weijer. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.646, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[10] Zitong Huang, Ze Chen, Zhixing Chen, Erjin Zhou, Xinxing Xu, Rick Siow Mong Goh, Yong Liu, Chunmei Feng, and Wangmeng Zuo. Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint arXiv:2401.01598, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[11] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In European Conference on Computer Vision, pages 432-448. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[13] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11463-11473, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9029, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.398 + ], + "angle": 0, + "content": "[16] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In Advances in Neural Information Processing Systems, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.908, + 0.455 + ], + "angle": 0, + "content": "[17] Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23638-23647, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.497 + ], + "angle": 0, + "content": "[18] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[19] Bolin Ni, Hongbo Zhao, Chenghao Zhang, Ke Hu, Gaofeng Meng, Zhaoxiang Zhang, and Shiming Xiang. Enhancing visual continual learning with language-guided supervision. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 24068-24077, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.906, + 0.622 + ], + "angle": 0, + "content": "[20] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[21] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[22] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.906, + 0.817 + ], + "angle": 0, + "content": "[23] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9889" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[26] Oindrila Saha, Grant Van Horn, and Subhransu Maji. Improved zero-shot classification by adapting vlms with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17552, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[27] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[28] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[29] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[30] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24183-24192, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[31] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[33] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11899-11909, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[34] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[35] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[38] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[39] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[40] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.443 + ], + "angle": 0, + "content": "[41] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.907, + 0.5 + ], + "angle": 0, + "content": "[42] Hai Zhang, Junzhe Xu, Shanlin Jiang, and Zhenan He. Simple semantic-aided few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28588–28597, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[43] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[44] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 9046–9056, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[45] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.781 + ], + "angle": 0, + "content": "[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.784, + 0.905, + 0.837 + ], + "angle": 0, + "content": "[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.84, + 0.905, + 0.896 + ], + "angle": 0, + "content": "[48] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9890" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_origin.pdf b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a8bd3cc728eeccf540db84a21a092c5af1f6510f --- /dev/null +++ b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/dbbdd019-448e-4ab4-b391-57c7f46b59cd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62e77ee0ffb43cc85c164a238d52928aa349daf5b991d132ac798e67916dc50a +size 1923523 diff --git a/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/full.md b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0dc3635aa75803c2d2423f32a2a3189a29bbc2a4 --- /dev/null +++ b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/full.md @@ -0,0 +1,323 @@ +# Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration + +Yiyang Chen $^{1}$ , Tianyu Ding $^{2}$ , Lei Wang $^{3}$ , Jing Huo $^{1}$ , Yang Gao $^{1}$ , Wenbin Li $^{1,4*}$ $^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China + $^{2}$ Applied Sciences Group, Microsoft, USA $^{3}$ University of Wollongong, Australia + $^{4}$ Shenzhen Research Institute of Nanjing University, Shenzhen, China + +# Abstract + +Few-shot Class-Incremental Learning (FSCIL) challenges models to adapt to new classes with limited samples, presenting greater difficulties than traditional class-incremental learning. While existing approaches rely heavily on visual models and require additional training during base or incremental phases, we propose a training-free framework that leverages pre-trained visual-language models like CLIP. At the core of our approach is a novel Bi-level Modality Calibration (BiMC) strategy. Our framework initially performs intra-modal calibration, combining LLM-generated fine-grained category descriptions with visual prototypes from the base session to achieve precise classifier estimation. This is further complemented by inter-modal calibration that fuses pre-trained linguistic knowledge with task-specific visual priors to mitigate modality-specific biases. To enhance prediction robustness, we introduce additional metrics and strategies that maximize the utilization of limited data. Extensive experimental results demonstrate that our approach significantly outperforms existing methods. Code is available at: https://github.com/yychen016/BiMC. + +# 1. Introduction + +The rapid advancement of artificial intelligence has propelled deep neural networks to achieve remarkable progress. However, these models often struggle with catastrophic forgetting when faced with dynamically changing environments, impeding their ability to learn sequential tasks effectively. Class-Incremental Learning (CIL) can be a potential solution to this challenge. Yet, in practical settings, researchers frequently encounter a scarcity of training data during the incremental phase, leading to sample sparsity issues. Consequently, Few-Shot Class-Incremental Learning (FSCIL) [31] has recently emerged as a more realistic scenario, better reflecting real-world constraints. + +![](images/030fcc11f45b67201aee0601c2407f7098dec105eb21477a20c26411ba68f89f.jpg) +Figure 1. Model updating paradigms in FSCIL. (a) Continuous model updates during incremental phases. (b) Leveraging the base task to learn a generalizable model. (c) Our proposed framework: training-free in both base and incremental sessions. + +Current FSCIL methods can be broadly categorized into two paradigms: the continual fine-tuning paradigm and the incremental-session freezing paradigm, as shown in Figure 1 (a) and (b). The former [4, 21, 27, 43] utilizes limited samples to continuously update the model, enhancing its plasticity. Conversely, the latter [22, 35, 44, 48] focuses on learning robust representations during the base task, which are then generalized to incremental tasks, prioritizing stability by avoiding subsequent model updates. Despite their distinct approaches, both paradigms necessitate additional training to adapt to new tasks, presenting ongoing challenges in resource efficiency and model adaptability. + +The emergence of large-scale models has shifted the landscape of representation learning. Vision-language models like CLIP [24], trained on a large number of image-text pairs, excel in zero-shot tasks. However, fine-tuning is computationally expensive and may compromise generalization, especially with limited data [14, 39]. While recent approaches like prompt tuning [46, 47] and low-rank adaptation [9] offer more efficient alternatives, they still demand + +additional training. This raises a crucial question: How can we effectively leverage pre-trained knowledge and domain-specific visual priors to adapt the large-scale model using few samples without additional training? + +While CLIP is widely used for cross-modal retrieval in visual categorization, it overlooks the potential of intramodal recognition, akin to traditional purely visual models, such as CNN [30, 41]. We believe that CLIP's visual prototypes can effectively complement zero-shot textual classifiers, serving as domain-specific priors, particularly in addressing two major challenges in FSCIL tasks: catastrophic forgetting and overfitting. Building on this insight, we introduce Bi-level Modality Calibration (BiMC), a simple yet effective framework that treats CLIP as a black-box model and operates without additional training, thereby avoiding the parameter updates that typically cause forgetting. Our framework incorporates both intra-modal and inter-modal calibration strategies to enhance CLIP's classification accuracy. For intra-modal calibration, we leverage LLM-generated fine-grained category descriptions in the textual domain to improve zero-shot classifier discrimination, while using base session prototypes in the visual domain to refine new class prototypes affected by sample scarcity. Inter-modal calibration combines pre-trained linguistic knowledge with task-aware visual priors to reduce modal biases and overfitting. The framework is further enhanced with an anisotropic covariance metric and a cross-modal category-conditioned nearest-neighbor metric, implementing the final classifier through masked ensemble inference. Extensive experiments demonstrate that our method achieves competitive performance without additional training. + +Our contributions are summarized as follows: + +- We propose a novel training-free FSCIL framework that achieves continual model adaptation by treating vision-language models as black boxes, significantly enhancing practicality for real-world applications. +- We develop a bi-level modality calibration approach combining intra-modal and inter-modal strategies to enhance classifier accuracy across both modalities. This approach is strengthened by an innovative visual covariance metric and a category-conditioned nearest-neighbor metric, culminating in a robust masked ensemble inference strategy. +- We demonstrate the effectiveness of our framework through extensive experiments on standard benchmarks, where it not only achieves competitive performance but also outperforms supervised training methods in several scenarios, despite requiring no additional training. + +# 2. Related work + +Few-Shot Class-Incremental Learning. Few-Shot Class-Incremental Learning (FSCIL) challenges models to adapt + +continuously using minimal samples. Existing approaches can be categorized into three main groups: representation-based methods [1, 2, 12, 18, 22, 27, 30, 38, 44], which focus on optimizing feature representations; dynamic-architecture-based methods [31, 41], which adapt model structures to accommodate new classes; and knowledge-distillation-based methods [4, 7, 15], which transfer knowledge from previous tasks to new ones. Notable examples include CEC [41], which employs an evolving graph model for classifier optimization, and FACT [44] and SAVC [30], which introduce virtual categories as placeholders for forward compatibility. In contrast, our approach leverages a pre-trained vision-language model, eliminating the need for additional learning on the base session. + +Class-Incremental Learning via Pre-Trained Models. Recent trends in incremental learning focus on effectively utilizing pre-trained models as backbones for downstream tasks. Early works primarily used prompt-based methods [11, 28, 34, 36, 37], introducing learnable prompt parameters for task adaptation. Other approaches have employed adapters [45] or LoRA [9] to facilitate adaptation [17, 40, 45]. While these methods use efficient parameter-tuning techniques, they still require further training. Our approach, however, leverages both vision and language modalities to enable ongoing model adaptation without gradient updates. + +Language-Assisted Classification. Linguistic information has been increasingly utilized in various visual tasks. Some studies have used linguistic cues to guide visual classification [33, 42], while others have leveraged Large Language Models (LLMs) to generate rich descriptive text [20, 23, 26]. In continual learning, several works have exploited linguistic guidance [13, 19]. Within FSCIL, methods like [2, 4, 21] also utilize linguistic information. However, these approaches have limitations: some [2] use distillation techniques on smaller visual models with linguistic knowledge as a regularization term, potentially limiting scalability and flexibility. Others [4, 21] rely on complex and often redundant architectural designs, making implementation challenging and inefficient. In contrast, our approach addresses these limitations by proposing a simpler, more efficient framework that effectively integrates linguistic and visual information. + +# 3. Method + +# 3.1. Preliminary + +In FSCIL, a model continuously adapts to a sequence of sessions $\mathcal{S} = \{\mathcal{D}_0, \mathcal{D}_1, \dots, \mathcal{D}_T\}$ . Each session $\mathcal{D}_t = \{(\mathbf{x}_i, y_i)\}_{i=1}^{N_t}$ consists of training data with a category space $\mathcal{C}_t^{\mathrm{train}} = Y_t$ , where $N_t$ represents the number of samples in the $t$ -th session. The base session $\mathcal{D}_0$ contains a substantial number of categories and samples, while subsequent incre + +![](images/5b4ad33fa5bebd94ccc23844662c36caebe6b538983bd185a292c7176b0aff18.jpg) +(a) Bi-Level Modality Calibration Framework + +![](images/c99d754171076c28918dddfcfc68ec896af88138cf8b42f7d23eee3226a43551.jpg) +(b) Semantic and Covariance-enhanced Metric +Figure 2. Overview of our framework. In each task adaptation phase, we construct a calibrated classifier through our proposed bi-level calibration framework, which involves intra-modal calibration and inter-modal calibration. To enhance performance, we introduce a globally shared covariance metric for visual feature modeling, complemented by a category-conditioned nearest-neighbor scoring strategy. + +mental sessions $\mathcal{D}_t$ for $t > 0$ contain only a few training samples per category, formally expressed as $N_0\gg N_t$ + +The categories across different sessions are mutually exclusive, meaning that for any different $i,j\in [0,T]$ $Y_{i}\cap Y_{j} = \varnothing$ . Furthermore, during the training of the current session, data from previous sessions is not accessible. At the testing phase of the $t$ -th session, the model evaluates against all previously encountered categories, with the cumulative category set defined as $\mathcal{V}_t = \bigcup_{j = 0}^t Y_j$ . Consequently, the test category space is given by $\mathcal{C}_t^{\mathrm{test}} = \mathcal{V}_t$ + +# 3.2. Bi-Level calibration framework + +Our proposed training-free framework builds upon CLIP [24] and incorporates two essential components: intra-modal and inter-modal calibration (see Figure 2a). + +# 3.2.1 Intra-modal classifier calibration + +In its standard usage, CLIP relies on simple prompt templates, (e.g., "a photo of a [CLS].") to construct zero-shot classifiers for visual tasks. However, this category-agnostic approach has inherent limitations. The generic nature of these templates fails to capture category-specific nuances, making them particularly inadequate for fine-grained classification tasks where detailed feature discrimination is crucial. A single, static template cannot effectively characterize the diverse visual characteristics within each category, especially when subtle distinctions between categories are essential. + +To overcome these limitations, we leverage Large Language Models (LLMs) to generate dynamic, category-specific descriptions. These LLM-generated descriptions provide richer, more discriminative semantic representations for each category, enhancing the model's capacity to capture fine-grained features. We integrate this enhanced semantic information through an intra-modal calibration strategy within the textual modality: + +$$ +\tilde {\boldsymbol {\mu}} _ {c} ^ {T} = \left(1 - \lambda_ {T}\right) \boldsymbol {w} _ {c} + \lambda_ {T} \left(\frac {1}{n _ {c}} \sum_ {j = 1} ^ {n _ {c}} \frac {g \left(\mathbf {t} _ {c , j}\right)}{\| g \left(\mathbf {t} _ {c , j}\right) \| _ {2}}\right). \tag {1} +$$ + +In this formulation, $\lambda_{T}$ controls the intensity of intra-modal calibration within the text modality, $n_c$ denotes the number of descriptions for category $c$ , $g(\cdot)$ represents the text encoder, $\mathbf{t}_{c,j}$ is the $j$ -th LLM-generated description for category $c$ , $\tilde{\boldsymbol{\mu}}_c^T$ is the calibrated textual prototype, and $\boldsymbol{w}_c$ denotes the original CLIP Zero-Shot classifier weight for class $c$ . This calibration mechanism preserves the robust generalization capabilities of the CLIP Zero-Shot classifier while incorporating fine-grained, category-specific semantic information. The latter term in Eq. (1) serves as a semantically enriched category description center, enhancing the model's discriminative power. + +In the FSCIL context, prototype-based classifiers are commonly employed, where the accuracy of prototype estimation directly impacts performance. During the base session, the abundance of samples allows for precise estimation of class prototypes. However, in incremental sessions + +that introduce new classes, the limited number of samples inevitably leads to biased estimates of new class prototypes. Drawing inspiration from TEEN [35], we leverage the accurately estimated prototype weights of base classes to facilitate intra-modality calibration on the prototypes of incoming new classes within the visual modality. + +For a given class $c$ , we first utilize the encoded training image data to obtain the naive visual prototype: + +$$ +\boldsymbol {\mu} _ {c} ^ {I} = \frac {1}{m _ {c}} \sum_ {j = 1} ^ {m _ {c}} \frac {f (\mathbf {x} _ {c , j})}{\| f (\mathbf {x} _ {c , j}) \| _ {2}}, \tag {2} +$$ + +where $f(\cdot)$ is the visual encoder, $m_{c}$ is the number of images in class $c$ , and $\mathbf{x}_{c,j}$ is the $j$ -th image belonging to category $c$ . Due to the inaccuracy in visual prototype estimation during the incremental sessions, the prototypes from the base session are used to recalibrate the new class prototypes by leveraging the similarity relationship between them. This process can be formulated as: + +$$ +\tilde {\boldsymbol {\mu}} _ {c} ^ {I} = \left\{ \begin{array}{l l} \boldsymbol {\mu} _ {c} ^ {I} & , t = 0 \\ (1 - \lambda_ {I}) \boldsymbol {\mu} _ {c} ^ {I} + \lambda_ {I} \sum_ {b = 1} ^ {| \mathcal {Y} _ {0} |} s _ {b, c} \boldsymbol {\mu} _ {b} ^ {I} & , t > 0, \end{array} \right. \tag {3} +$$ + +where, $t$ represents the task identifier, $\lambda_I$ controls the strength of the visual intra-modal calibration, and $s_{b,c}$ denotes the normalized cosine similarity between the visual prototypes of class $b$ (from the base classes) and class $c$ . The similarity $s_{b,c}$ is defined as: $s_{b,c} = \frac{e^{\tau \cdot \langle \mu_b, \mu_c \rangle}}{\sum_{i=1}^{|\mathcal{Y}_0|} e^{\tau \cdot \langle \mu_i, \mu_c \rangle}}$ . Here, $\langle \cdot, \cdot \rangle$ represents the cosine similarity, and $\tau$ is the temperature scaling parameter. + +Through this visual intra-modality calibration process, the prototypes of the new classes are adjusted to align more closely with the well-calibrated base class prototypes, allowing the new class prototypes to partially inherit the discriminative properties of the base class prototypes. + +# 3.2.2 Inter-modal classifier calibration + +Following intra-modal calibration, two classifier achieves enhanced accuracy within its respective modality. However, we posit that their full potential remains unrealized when operating in isolation. This is due to the fact that classifiers operating exclusively on linguistic data derive their knowledge from pre-trained models and lack explicit downstream knowledge for visual tasks. Conversely, classifiers functioning solely within the visual domain serve as metrics within the visual sub-feature space but are deficient in semantic understanding. Recognizing the complementary nature of these two modalities, we propose a training-free inter-modal calibration strategy, allowing domain-relevant downstream visual prior to guide domain-agnostic pretrained linguistic knowledge. Formally, this process can be + +described as: + +$$ +\boldsymbol {\mu} _ {c} = \beta \tilde {\boldsymbol {\mu}} _ {c} ^ {T} + (1 - \beta) \tilde {\boldsymbol {\mu}} _ {c} ^ {I}. \tag {4} +$$ + +Here, $\pmb{\mu}_{c}$ is the inter-modal calibrated classifier, and $\beta$ is the calibration coefficient. During inference, we use cosine similarity to calculate the similarity between the sample features $f(\mathbf{x})$ and the mixed category center: + +$$ +\mathbf {s} _ {c} ^ {\text {c a l i b}} = \frac {f (\mathbf {x}) ^ {\top} \boldsymbol {\mu} _ {c}}{\| f (\mathbf {x}) \| _ {2} \cdot \| \boldsymbol {\mu} _ {c} \| _ {2}}. \tag {5} +$$ + +This inter-modal calibration strategy effectively combines the strengths of both linguistic and visual modalities. It helps to enhance the overall classification performance by creating a more robust and comprehensive classifier that can better capture the nuances of visual tasks while maintaining a strong semantic understanding. + +# 3.3. Semantic and covariance-enhanced metric + +While our previous metrics effectively integrate linguistic and visual information and mitigate modal bias, they primarily measure sample distances from distribution centers. This isotropic approach, despite its simplicity, struggles to capture higher-order information and complex data relationships. To address these limitations, we propose leveraging statistical information from existing data for more comprehensive and efficient measurements, as shown in Figure 2b. + +# 3.3.1 Modeling global covariance + +Inspired by FeCAM [8], we introduce an anisotropic covariance metric to fully utilize the visual modality's statistical information. For a session $t$ , we compute its covariance matrix $\boldsymbol{\Sigma}^t$ using image embeddings $f(\mathbf{x})$ of all image data relevant to the session. To avoid singularity, we apply regularization with an identity matrix $I_d$ : + +$$ +\tilde {\boldsymbol {\Sigma}} ^ {t} = \boldsymbol {\Sigma} ^ {t} + \frac {\gamma}{d} \operatorname {t r} \left(\boldsymbol {\Sigma} ^ {t}\right) \boldsymbol {I} _ {d}, \tag {6} +$$ + +where $\gamma$ is a hyperparameter controlling the strength of regularization, $d$ represents the feature dimension. Here, $\mathrm{tr}(\cdot)$ denotes the trace operator of a matrix, and both $d$ and $\mathrm{tr}(\pmb{\Sigma}^t)$ provide a prior estimation for the regularization strength. Subsequently, we construct a continuously evolving shared covariance matrix as: + +$$ +\tilde {\boldsymbol {\Sigma}} _ {G} ^ {t} = \frac {| \mathcal {Y} _ {t - 1} |}{| \mathcal {Y} _ {t} |} \tilde {\boldsymbol {\Sigma}} _ {G} ^ {t - 1} + \left(1 - \frac {| \mathcal {Y} _ {t - 1} |}{| \mathcal {Y} _ {t} |}\right) \tilde {\boldsymbol {\Sigma}} ^ {t}. \tag {7} +$$ + +The evolved covariance matrix captures the statistical information of all categories. Given a test sample $\mathbf{x}$ , we calculate its score as: + +$$ +\mathbf {s} _ {c} ^ {\mathrm {c o v}} (\mathbf {x}) = - \frac {1}{d} \left(f (\mathbf {x}) - \tilde {\boldsymbol {\mu}} _ {c} ^ {I}\right) ^ {T} \tilde {\boldsymbol {\Sigma}} _ {G} ^ {- 1} \left(f (\mathbf {x}) - \tilde {\boldsymbol {\mu}} _ {c} ^ {I}\right). \tag {8} +$$ + +![](images/a994cc2e8cb606d02da4817343093fbf085c29173d024f0f17b4a66f014e3e48.jpg) +Figure 3. In the inference phase, an ensemble strategy with category masking is utilized. + +The division by feature dimension $d$ serves as a normalization factor, preventing the softmax transformation into probabilities from collapsing into a single category. The negative sign maintains metric consistency, ensuring that a larger output score indicates a higher probability of the sample belonging to category $c$ . Here, $\hat{\Sigma}_G$ denotes the global covariance, with the superscript $t$ omitted for clarity. + +# 3.3.2 Cross-modal category nearest neighbor metric + +Global covariance modeling effectively utilizes statistical information from the visual modality. To better leverage LLM-generated category descriptions, we propose a category-conditioned nearest-neighbor-based metric. Given a description text $\mathbf{t}_{c,j}$ encoded by $g(\cdot)$ , we obtain its normalized feature representation $\mathbf{z}_{c,j} = \frac{g(\mathbf{t}_{c,j})}{\|g(\mathbf{t}_{c,j})\|_2}$ . For a visual query sample, we compute its normalized representation $\mathbf{v} = \frac{f(\mathbf{x})}{\|f(\mathbf{x})\|_2}$ . The likelihood of this sample belonging to category $c$ is determined by the maximum dot product between $\mathbf{v}$ and all description features in class $c$ : + +$$ +\mathbf {s} _ {c} ^ {\mathrm {n n}} = \max _ {j} \left\{\boldsymbol {z} _ {c, j} ^ {\top} \cdot \boldsymbol {v} \right\}, \tag {9} +$$ + +where $\mathbf{s}_c^{\mathrm{nn}}$ indicates the likelihood that the query sample belongs to category $c$ . This metric leverages maximum cross-modal similarity, effectively utilizing the diverse textual descriptions generated by LLM. + +# 3.3.3 Inference score reorganization strategy + +We employ a globally shared covariance matrix to model distribution shape information in the visual feature space. While the base session contains substantial data for accurate covariance matrix estimation, experiments reveal that these base-derived covariance matrices perform poorly in distinguishing new class data. This limitation stems from two factors: (1) the covariance from extensive base data fails to align with new classes, and (2) the limited data available for new classes prevents accurate covariance estimation. To address this, we use the cross-modal category nearest neighbor metric for novel classes to compensate for the limitations of the second-order covariance metric, as shown in Figure 3. + +We first normalize the scores of three metrics using the softmax function: $\mathbf{p}^{\mathrm{calib}} = \mathrm{softmax}\left(\mathbf{s}^{\mathrm{calib}}\right)$ , $\mathbf{p}^{\mathrm{cov}} = \mathrm{softmax}\left(\mathbf{s}^{\mathrm{cov}}\right)$ , and $\mathbf{p}^{\mathrm{nn}} = \mathrm{softmax}\left(\mathbf{s}^{\mathrm{nn}}\right)$ . The final probability score $\mathbf{p}_c$ is computed differently for base and novel categories: + +$$ +\mathbf {p} _ {c} = \left\{ \begin{array}{l l} \alpha \mathbf {p} _ {c} ^ {\text {c a l i b}} + (1 - \alpha) \mathbf {p} _ {c} ^ {\text {c o v}}, & c \in \mathcal {Y} _ {0} \\ \alpha \mathbf {p} _ {c} ^ {\text {c a l i b}} + (1 - \alpha) \mathbf {p} _ {c} ^ {\text {n n}}, & c \notin \mathcal {Y} _ {0}. \end{array} \right. \tag {10} +$$ + +For base categories $(c\in \mathcal{V}_0)$ , we use a weighted sum of mixed prototype scores and visual mahalanobis metric scores. For novel categories $(c\notin \mathcal{V}_0)$ , we combine mixed prototype scores with cross-modal nearest neighbor metric scores. The final category prediction for a sample is determined by $\arg \max_c\mathbf{p}_c$ + +# 4. Experiments + +# 4.1. Experimental settings + +Datasets. We evaluate our method on three benchmarks following [22, 30, 41]: CIFAR100 [3], CUB200-2011 [32], and miniImageNet [25]. For CIFAR100 and miniImageNet, we partition each dataset into 60 base classes and 40 novel classes. The novel classes are further divided into eight incremental tasks, with each task structured as a 5-way 5-shot incremental session. The CUB200 dataset is split into 100 base classes and 100 novel classes, with each incremental task organized in a 10-way 5-shot format. + +Implementation details. We adopt CLIP's ViT-B/16 architecture as the backbone for fair comparison across all methods. For category descriptions, we utilize fine-grained text generated by CuPL [23] and [26]. The hyperparameters are configured with intra-modal calibration coefficients $\lambda_{T} = 0.5$ (textual), and $\lambda_{I} = 0.1$ (visual), and a temperature parameter $\tau = 16$ . The inter-modal calibration coefficient $\beta$ is optimized through a validation process. Specifically, we divide the base session into training and validation subsets. Applying our bi-level calibration framework on the training subset, we evaluate performance on the validation subset. By incrementing $\beta$ by 0.05, we identify the value that maximizes validation accuracy. This optimal $\beta$ remains fixed throughout subsequent incremental sessions. The shared covariance parameter $\gamma$ is set to 1 for the base session and adjusted to 5 for incremental sessions. Finally, for score reorganization, we set $\alpha = 0.6$ for CIFAR100 and miniImageNet, and $\alpha = 0.8$ for CUB200. + +# 4.2. Comparison results + +We compare our method with training-free approaches in the FSCIL scenario: CLIP Zero-Shot [24], Visual Prototype [29], TEEN [35] and FeCAM [8]. We present the performance curve over all benchmarks in Figure 4, while detailed session-wise performance on miniImageNet is shown + +Table 1. Detailed session-wise accuracy, average accuracy (Avg) and performance degradation (PD) comparison on miniImageNet dataset. V and L represent the visual and language modalities, respectively. BiMC refers to the results obtained solely through the bi-level calibration framework, whereas $\mathrm{BiMC}^{\dagger}$ incorporates the ensemble classifier strategy. The best results on each each sessions are indicated in bold, while the second-best results are underlined. $\uparrow$ means higher is better, while $\downarrow$ means lower is better. + +
MethodModalityAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
CLIP Zero-Shot [24]L91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
Visual Prototype [29]V92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
TEEN [35]V92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
FeCAM [8]V94.4293.7191.6190.5990.6290.2288.4388.1788.2090.666.22
BiMCV-L94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
BiMC†V-L95.4795.3493.8093.6393.5193.1892.5192.5692.4093.603.07
+ +(a) miniImageNet +![](images/57436e0131a346d47d8c328e86056be0ec678b944fad1e04a7d10a4ca767dc9f.jpg) +CLIP Zero-Shot Visual Prototype FeCAM BiMC BiMC + +![](images/f23fa610cb82228685bcc854c7c2817a4182c2de560d7d661f47e5d6dfdb565b.jpg) +(b) CIFAR100 + +![](images/e0e46788044a63a361463168d4f9be39d998a627aa93130e235db4d4d64bb120.jpg) +(c) CUB200 + +in Table 1. Detailed results for CIFAR100 and CUB200 are provided in the supplementary material. + +Our approach demonstrates competitive performance across all three datasets and significantly outperforms unimodal baselines without requiring additional training, highlighting the importance of leveraging information from both modalities. On the miniImageNet dataset, while visual and textual prototype classifiers achieve comparable results, and although TEEN [35] and FeCAM [8] better utilize visual modality information, their improvement remains limited. Our framework employs a simple calibration strategy to integrate domain visual knowledge with pretrained linguistic knowledge, yielding up to $6\%$ performance improvement over uni-modal models. + +For CIFAR100, the language prototype shows better resistance to forgetting compared to the visual one, which is more susceptible to sequential task learning. This can be attributed to the dataset's low-resolution characteristics, resulting in insufficient visual information during incremental phases. Our calibration framework achieves approximately $3.5\%$ improvement over a language-only classifier. In contrast, for the fine-grained CUB200 dataset, the visual prototype surpasses the language prototype due to limited trans + +![](images/08124e23b9f475316d421fc74513f9a8c45c39641d4dd10d2127ed4718cebc19.jpg) +Figure 4. Performance curve of each incremental sessions on (a) miniImageNet, (b) CIFAR100 and (c) CUB200 datasets. +(a) Old Classes Entropy +Figure 5. Comparison of output confidence with different models. ferability of pre-trained language knowledge. By combining both modalities, our approach achieves approximately $3.5\%$ performance improvement. Notably, our framework demonstrates robust resistance to forgetting, achieving the lowest performance drop across all metrics. + +![](images/b7dbfd7fc27e2554dffc90f6f9df37d8566f965101ec383ca61d203ddffc5342.jpg) +(b) Novel Classes Entropy + +One can observe that each modality's contribution to classification varies across datasets. Our proposed bi-level calibration framework effectively leverages the strengths of each modality and mitigates modality bias. Moreover, our ensemble inference strategy, denoted as $\mathrm{BiMC}^{\dagger}$ , provides additional performance improvements. + +# 4.3. Investigation of bi-level calibrated classifiers + +In this section, we analyze two distinct attributes of the calibrated classifier to demonstrate the effectiveness of our pro + +Table 2. Ablation studies of bi-level calibration framework on miniImageNet. Vis. and Lang. individually represent classifiers derived from the visual and language modalities. Intra-C and Inter-C refer to the strategies of intra-modal and inter-modal calibration, respectively. + +
VisLangIntra-CInter-CAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
✓92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
✓✓92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
✓91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
✓✓93.2793.3291.9191.5191.2190.7389.2889.2389.0191.054.26
✓✓✓94.6894.5892.8792.4992.5692.3191.5991.5891.5492.693.14
✓✓✓✓94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
+ +Table 3. The left side of the table lists classification result combinations from three classifiers, while the right side shows the percentage $(\%)$ of test samples meeting each combination. Test sets from both the base and the final session were used for analysis. + +
ModalityCIFAR100CUB200
VLV-LD0testD8testD0testD10test
✓✓✓65.3351.8159.0044.85
X✓✓6.9513.072.064.00
✓X✓6.105.5619.8321.40
XX✓1.202.101.081.98
+ +posed framework. + +The bi-level calibrated classifier enhances prediction confidence. We evaluate prediction confidence by calculating the entropy of predictions for each classifier in the final session across all three datasets. The entropy serves as an uncertainty indicator for model outputs, and we compute it separately for both base and novel categories, as detailed in Figure 5. The findings indicate that the bi-level calibrated classifier significantly reduces the entropy of predictions compared to uni-model classifiers, thereby enhancing prediction accuracy. + +The bi-level calibrated classifier mitigates modality bias. To elucidate the superior performance of our method, we analyze the joint prediction outcomes across different classifiers in comparison to uni-modal approaches. Using $\checkmark$ to denote correct classifications and $x$ for classification errors, we observe that the calibrated classifier maintains accurate predictions even in cases where uni-modal classifiers fail, as demonstrated in the $2^{\mathrm{nd}}$ and $3^{\mathrm{rd}}$ rows of Table 3. This evidence suggests that the bi-level calibrated classifier produces more accurate and unbiased predictions, offering enhanced robustness compared to uni-modality classifiers. Remarkably, in certain rare instances, the modality-calibrated classifier successfully classifies samples that both uni-modal classifiers misclassify. + +# 4.4. Ablation study + +We conduct comprehensive ablation analyses on the miniImageNet dataset to evaluate the significance of each component within our proposed framework. + +Bi-level calibration framework. Our framework com- + +Table 4. Ablation study of Semantic and Covariance-enhanced Metric on the miniImageNet dataset. To verify their effectiveness, we measured the performance of the base task $\mathcal{A}_{\mathrm{base}}$ , the performance of last task $\mathcal{A}_{\mathrm{last}}$ and the average performance $\mathcal{A}_{\mathrm{avg.}}$ across all tasks. Additionally, we report the accuracy of base $\mathcal{A}_{\mathrm{last}}^b$ and novel $\mathcal{A}_{\mathrm{last}}^n$ class in last task. + +
Ablation\( A_{base} \)\( A_{last} \)\( A_{last}^{b} \)\( A_{last}^{n} \)\( A_{avg} \)
BiMC94.9091.8193.4389.3892.97
+ MGC95.4791.5194.3787.2292.91
+ CMNN94.8592.1593.3090.4293.18
\( BiMC^† (w/o mask.) \)95.4792.2094.1289.3293.37
\( BiMC^† \)95.4792.4093.2091.2093.60
+ +prises four components: visual modality information, linguistic modality information, intra-modal calibration strategy, and inter-modal calibration strategy. As shown in Table 2, single-modality classifiers enhanced with intra-modal calibration demonstrate improved performance, attributable to more accurate estimation of the uni-modal classifier. When employing inter-modal calibration, we observe significant performance gains compared to uni-modal approaches, owing to the fusion strategy's effectiveness in addressing modality biases. The combination of both intramodal and inter-modal calibration strategies yields the best performance, with comprehensive knowledge integration leading to minimal forgetting rates. + +Semantic and covariance-enhanced metric. We examine the effectiveness of Modeling Global Covariance (MGC) and Cross-modal category Nearest Neighbor metric (CMNN) in Table 4. The global covariance modeling strategy enhances base category performance in both the base session and the final incremental session, though at the cost of reduced recognition performance for novel classes, a consequence of the mismatch between global covariance (dominated by base data) and new classes. The semantic nearest neighbor strategy, conversely, improves novel class recognition performance. Implementing both strategies simultaneously without masking further enhances overall performance. Finally, the ensemble inference mechanism with masks fully leverages the advantages of both metrics, achieving superior performance. + +Table 5. Comparison with three trainable methods on the miniImageNet dataset. $N_{p}$ is the number of parameters which require training. $\Delta A_{\mathrm{last}}$ reflects the last task's performance gap between our method and the comparative one. + +
MethodNpAbaseAlastAavg.ΔAlast
CPE-CLIP [5]400k90.2382.7786.13+9.63
CLIP-M3[6]46k96.0092.5094.10-0.10
LP-DiF [10]8.1k96.3491.6893.76+0.72
BiMC†095.4792.4093.600.00
+ +![](images/e5e5c32a2306938d04c274f5f4b1f6d572633587d94cedd59476b55792fccafc.jpg) +Figure 6. CLIP Feature Space Visualization. Squares, circles, and pentagrams represent textual, visual features, and calibrated classifiers respectively. Different colors denote different categories. + +# 4.5. Further analysis + +Comparison with training-required methods. To comprehensively evaluate our approach, we conduct comparisons with three training-based methods [5, 6, 10]. As demonstrated in Table 5, our parameter-free adaptation approach to FSCIL tasks achieves comparable or superior performance to methods that require additional training. + +Visualization of Feature Space. We visualized the image and text feature distributions on the miniImageNet dataset, as depicted in Figure 6. Our analysis showed that in the unified feature space encoded by CLIP, visual and textual features are located at diametrically opposed ends, exemplifying a modality gap phenomenon [16]. Textual features showed more consistent convergence, in contrast to the visual features of the same category, which were more dispersed within the feature space. Notably, despite being distinct, these features share similar representations in the visual subspace (e.g., the two purple dashed ellipses in Figure 6) but are clearly distinguishable in the textual subspace. Furthermore, after modality calibration, the classifiers, influenced by visual priors, demonstrate a shift from the textual modality to the visual modality, thereby alleviating modality bias and enhancing classification capabilities. + +Analysis of hyper-parameters. Our framework involves two critical hyper-parameters: the inter-modal calibration parameter $\beta$ and the ensemble classifier weighting coefficient $\alpha$ , both constrained to [0, 1]. Figure 7a illustrates how performance varies sessions for different values of $\beta$ . + +![](images/35c68906a9f825959cb60b12f63cef715f2e5fae892fb8df54a56ca338c1af74.jpg) +(a) Influence of $\beta$ and session-wise accuracy on miniImageNet + +![](images/dd4e19dacae4c2f23db4f1b41b3841d6a1154a411644a9d4ec0c955313180a82.jpg) +(b) Influence of $\alpha$ on miniImageNet + +![](images/cfe36ef6505c16c1ec2265ce9357aa6f988328cced9a8110867173d92e2427d7.jpg) +(c) Influence of $\alpha$ on CIFAR100 +Figure 7. Analysis on hyper-parameters $\beta$ and $\alpha$ . For $\beta$ , we present heatmaps illustrating accuracy under varying $\beta$ across different sessions. For $\alpha$ , we report the average task accuracy $A_{avg}$ and the final task accuracy $A_{last}$ under different settings. + +We observe that the $\beta$ values that perform well in the base session maintain their effectiveness throughout incremental tasks. The optimal solution identified through the validation search in the base session (indicated by the blue rectangular box) closely approaches the global optimum. For the parameter $\alpha$ , results are presented in Figures 7b and 7c. When $\alpha = 1$ , the method reduces to a bi-level calibrated classifier. The performance reaches its peak at $\alpha = 0.6$ , optimizing both the average and final accuracy metrics. + +# 5. Conclusion + +We explore CLIP models in the context of Few-shot Class-Incremental Learning. To mitigate forgetting, we treat the pre-trained model as a black box and employ a simple yet effective training-free bi-level modality calibration strategy. The intra-modal calibration achieves accurate classifier estimates within each modality, while the inter-modal calibration combines knowledge from both visual and textual modalities. An ensemble inference strategy integrating covariance and nearest-neighbor metrics enhances the accuracy of prediction. Extensive experiments validate the effectiveness of our approach. + +Acknowledgements. This work is supported in part by the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), National Natural Science Foundation of China (62192783, 62276128), Guangdong Basic and Applied Basic Research Foundation (2024A1515011340) Jiangsu Natural Science Foundation (BK20221441), and the Australian Research Council's Discovery Project (DP220101784). + +# References + +[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2 +[2] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 2 +[3] Krizhevsky Alex. Learning multiple layers of features from tiny images. https://www.cs.toronto.edu/kriz/learning-features-2009-TR.pdf, 2009.5 +[4] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1, 2 +[5] Marco D'Alessandro, Alberto Alonso, Enrique Calabrés, and Mikel Galar. Multimodal parameter-efficient few-shot class incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3393-3403, 2023. 8 +[6] Thang Doan, Sima Behpour, Xin Li, Wenbin He, Liang Gou, and Liu Ren. A streamlined approach to multimodal few-shot class incremental learning for fine-grained datasets. arXiv preprint arXiv:2403.06295, 2024. 8 +[7] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2 +[8] Dipam Goswami, Yuyang Liu, Bartlomiej Twardowski, and Joost van de Weijer. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 6 +[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2 +[10] Zitong Huang, Ze Chen, Zhixing Chen, Erjin Zhou, Xinxing Xu, Rick Siow Mong Goh, Yong Liu, Chunmei Feng, and Wangmeng Zuo. Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint arXiv:2401.01598, 2024. 8 +[11] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2 +[12] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In European Conference on Computer Vision, pages 432-448. Springer, 2022. 2 + +[13] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11463-11473, 2023. 2 +[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 1 +[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9029, 2021. 2 +[16] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In Advances in Neural Information Processing Systems, 2022. 8 +[17] Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23638-23647, 2024. 2 +[18] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2 +[19] Bolin Ni, Hongbo Zhao, Chenghao Zhang, Ke Hu, Gaofeng Meng, Zhaoxiang Zhang, and Shiming Xiang. Enhancing visual continual learning with language-guided supervision. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 24068-24077, 2024. 2 +[20] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023. 2 +[21] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 1, 2 +[22] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 1, 2, 5 +[23] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 2, 5 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 3, 5, 6 + +[25] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5 +[26] Oindrila Saha, Grant Van Horn, and Subhransu Maji. Improved zero-shot classification by adapting vlms with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17552, 2024. 2, 5 +[27] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 1, 2 +[28] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2 +[29] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 5, 6 +[30] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24183-24192, 2023. 2, 5 +[31] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 1, 2 +[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[33] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11899-11909, 2023. 2 +[34] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2 +[35] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 5, 6 +[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 2 + +[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2 +[38] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2 +[39] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1 +[40] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 2 +[41] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2021. 2, 5 +[42] Hai Zhang, Junzhe Xu, Shanlin Jiang, and Zhenan He. Simple semantic-aided few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28588–28597, 2024. 2 +[43] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1 +[44] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 9046–9056, 2022. 1, 2 +[45] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2 +[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1 +[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1 +[48] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1 \ No newline at end of file diff --git a/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/images.zip b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8c147c76cce75636870178f259dc7dadc5a8ff5 --- /dev/null +++ b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:470a5c76009e59559177fc9d72089b9ceb5b88b5d8a7b31ba85f4bf4b92d70b8 +size 543268 diff --git a/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/layout.json b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7f099c1732fb31262636dce795e7398ad1ffc8d8 --- /dev/null +++ b/2025/Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration/layout.json @@ -0,0 +1,9428 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 103, + 459, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 103, + 459, + 140 + ], + "spans": [ + { + "bbox": [ + 152, + 103, + 459, + 140 + ], + "type": "text", + "content": "Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": "Yiyang Chen" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Tianyu Ding" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Lei Wang" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Jing Huo" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Yang Gao" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Wenbin Li" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1,4*}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology, Nanjing University, China \n" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": "Applied Sciences Group, Microsoft, USA " + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": "University of Wollongong, Australia \n" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 160, + 512, + 219 + ], + "type": "text", + "content": "Shenzhen Research Institute of Nanjing University, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 521 + ], + "type": "text", + "content": "Few-shot Class-Incremental Learning (FSCIL) challenges models to adapt to new classes with limited samples, presenting greater difficulties than traditional class-incremental learning. While existing approaches rely heavily on visual models and require additional training during base or incremental phases, we propose a training-free framework that leverages pre-trained visual-language models like CLIP. At the core of our approach is a novel Bi-level Modality Calibration (BiMC) strategy. Our framework initially performs intra-modal calibration, combining LLM-generated fine-grained category descriptions with visual prototypes from the base session to achieve precise classifier estimation. This is further complemented by inter-modal calibration that fuses pre-trained linguistic knowledge with task-specific visual priors to mitigate modality-specific biases. To enhance prediction robustness, we introduce additional metrics and strategies that maximize the utilization of limited data. Extensive experimental results demonstrate that our approach significantly outperforms existing methods. Code is available at: https://github.com/yychen016/BiMC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 530, + 137, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 137, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 137, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "type": "text", + "content": "The rapid advancement of artificial intelligence has propelled deep neural networks to achieve remarkable progress. However, these models often struggle with catastrophic forgetting when faced with dynamically changing environments, impeding their ability to learn sequential tasks effectively. Class-Incremental Learning (CIL) can be a potential solution to this challenge. Yet, in practical settings, researchers frequently encounter a scarcity of training data during the incremental phase, leading to sample sparsity issues. Consequently, Few-Shot Class-Incremental Learning (FSCIL) [31] has recently emerged as a more realistic scenario, better reflecting real-world constraints." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 243, + 553, + 414 + ], + "blocks": [ + { + "bbox": [ + 316, + 243, + 553, + 414 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 414 + ], + "type": "image", + "image_path": "030fcc11f45b67201aee0601c2407f7098dec105eb21477a20c26411ba68f89f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 418, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 462 + ], + "type": "text", + "content": "Figure 1. Model updating paradigms in FSCIL. (a) Continuous model updates during incremental phases. (b) Leveraging the base task to learn a generalizable model. (c) Our proposed framework: training-free in both base and incremental sessions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 472, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 616 + ], + "type": "text", + "content": "Current FSCIL methods can be broadly categorized into two paradigms: the continual fine-tuning paradigm and the incremental-session freezing paradigm, as shown in Figure 1 (a) and (b). The former [4, 21, 27, 43] utilizes limited samples to continuously update the model, enhancing its plasticity. Conversely, the latter [22, 35, 44, 48] focuses on learning robust representations during the base task, which are then generalized to incremental tasks, prioritizing stability by avoiding subsequent model updates. Despite their distinct approaches, both paradigms necessitate additional training to adapt to new tasks, presenting ongoing challenges in resource efficiency and model adaptability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "content": "The emergence of large-scale models has shifted the landscape of representation learning. Vision-language models like CLIP [24], trained on a large number of image-text pairs, excel in zero-shot tasks. However, fine-tuning is computationally expensive and may compromise generalization, especially with limited data [14, 39]. While recent approaches like prompt tuning [46, 47] and low-rank adaptation [9] offer more efficient alternatives, they still demand" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9881" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "content": "additional training. This raises a crucial question: How can we effectively leverage pre-trained knowledge and domain-specific visual priors to adapt the large-scale model using few samples without additional training?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 297, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 297, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 297, + 455 + ], + "type": "text", + "content": "While CLIP is widely used for cross-modal retrieval in visual categorization, it overlooks the potential of intramodal recognition, akin to traditional purely visual models, such as CNN [30, 41]. We believe that CLIP's visual prototypes can effectively complement zero-shot textual classifiers, serving as domain-specific priors, particularly in addressing two major challenges in FSCIL tasks: catastrophic forgetting and overfitting. Building on this insight, we introduce Bi-level Modality Calibration (BiMC), a simple yet effective framework that treats CLIP as a black-box model and operates without additional training, thereby avoiding the parameter updates that typically cause forgetting. Our framework incorporates both intra-modal and inter-modal calibration strategies to enhance CLIP's classification accuracy. For intra-modal calibration, we leverage LLM-generated fine-grained category descriptions in the textual domain to improve zero-shot classifier discrimination, while using base session prototypes in the visual domain to refine new class prototypes affected by sample scarcity. Inter-modal calibration combines pre-trained linguistic knowledge with task-aware visual priors to reduce modal biases and overfitting. The framework is further enhanced with an anisotropic covariance metric and a cross-modal category-conditioned nearest-neighbor metric, implementing the final classifier through masked ensemble inference. Extensive experiments demonstrate that our method achieves competitive performance without additional training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 456, + 253, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 456, + 253, + 466 + ], + "spans": [ + { + "bbox": [ + 67, + 456, + 253, + 466 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 295, + 657 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 468, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 515 + ], + "type": "text", + "content": "- We propose a novel training-free FSCIL framework that achieves continual model adaptation by treating vision-language models as black boxes, significantly enhancing practicality for real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": "- We develop a bi-level modality calibration approach combining intra-modal and inter-modal strategies to enhance classifier accuracy across both modalities. This approach is strengthened by an innovative visual covariance metric and a category-conditioned nearest-neighbor metric, culminating in a robust masked ensemble inference strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 657 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of our framework through extensive experiments on standard benchmarks, where it not only achieves competitive performance but also outperforms supervised training methods in several scenarios, despite requiring no additional training." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 669, + 140, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 140, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 140, + 681 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Few-Shot Class-Incremental Learning. Few-Shot Class-Incremental Learning (FSCIL) challenges models to adapt" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "continuously using minimal samples. Existing approaches can be categorized into three main groups: representation-based methods [1, 2, 12, 18, 22, 27, 30, 38, 44], which focus on optimizing feature representations; dynamic-architecture-based methods [31, 41], which adapt model structures to accommodate new classes; and knowledge-distillation-based methods [4, 7, 15], which transfer knowledge from previous tasks to new ones. Notable examples include CEC [41], which employs an evolving graph model for classifier optimization, and FACT [44] and SAVC [30], which introduce virtual categories as placeholders for forward compatibility. In contrast, our approach leverages a pre-trained vision-language model, eliminating the need for additional learning on the base session." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "text", + "content": "Class-Incremental Learning via Pre-Trained Models. Recent trends in incremental learning focus on effectively utilizing pre-trained models as backbones for downstream tasks. Early works primarily used prompt-based methods [11, 28, 34, 36, 37], introducing learnable prompt parameters for task adaptation. Other approaches have employed adapters [45] or LoRA [9] to facilitate adaptation [17, 40, 45]. While these methods use efficient parameter-tuning techniques, they still require further training. Our approach, however, leverages both vision and language modalities to enable ongoing model adaptation without gradient updates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 553, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 587 + ], + "type": "text", + "content": "Language-Assisted Classification. Linguistic information has been increasingly utilized in various visual tasks. Some studies have used linguistic cues to guide visual classification [33, 42], while others have leveraged Large Language Models (LLMs) to generate rich descriptive text [20, 23, 26]. In continual learning, several works have exploited linguistic guidance [13, 19]. Within FSCIL, methods like [2, 4, 21] also utilize linguistic information. However, these approaches have limitations: some [2] use distillation techniques on smaller visual models with linguistic knowledge as a regularization term, potentially limiting scalability and flexibility. Others [4, 21] rely on complex and often redundant architectural designs, making implementation challenging and inefficient. In contrast, our approach addresses these limitations by proposing a simpler, more efficient framework that effectively integrates linguistic and visual information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 621, + 392, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 392, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 392, + 634 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "content": "In FSCIL, a model continuously adapts to a sequence of sessions " + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{\\mathcal{D}_0, \\mathcal{D}_1, \\dots, \\mathcal{D}_T\\}" + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "content": ". Each session " + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{N_t}" + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "content": " consists of training data with a category space " + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t^{\\mathrm{train}} = Y_t" + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "content": " represents the number of samples in the " + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "content": "-th session. The base session " + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 313, + 640, + 553, + 713 + ], + "type": "text", + "content": " contains a substantial number of categories and samples, while subsequent incre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9882" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 77, + 376, + 308 + ], + "blocks": [ + { + "bbox": [ + 66, + 77, + 376, + 308 + ], + "lines": [ + { + "bbox": [ + 66, + 77, + 376, + 308 + ], + "spans": [ + { + "bbox": [ + 66, + 77, + 376, + 308 + ], + "type": "image", + "image_path": "5b4ad33fa5bebd94ccc23844662c36caebe6b538983bd185a292c7176b0aff18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 313, + 295, + 323 + ], + "lines": [ + { + "bbox": [ + 146, + 313, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 146, + 313, + 295, + 323 + ], + "type": "text", + "content": "(a) Bi-Level Modality Calibration Framework" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 395, + 77, + 545, + 308 + ], + "blocks": [ + { + "bbox": [ + 395, + 77, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 395, + 77, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 395, + 77, + 545, + 308 + ], + "type": "image", + "image_path": "c99d754171076c28918dddfcfc68ec896af88138cf8b42f7d23eee3226a43551.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 313, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 395, + 313, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 395, + 313, + 545, + 323 + ], + "type": "text", + "content": "(b) Semantic and Covariance-enhanced Metric" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 325, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 325, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 555, + 360 + ], + "type": "text", + "content": "Figure 2. Overview of our framework. In each task adaptation phase, we construct a calibrated classifier through our proposed bi-level calibration framework, which involves intra-modal calibration and inter-modal calibration. To enhance performance, we introduce a globally shared covariance metric for visual feature modeling, complemented by a category-conditioned nearest-neighbor scoring strategy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "content": "mental sessions " + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "content": " contain only a few training samples per category, formally expressed as " + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "inline_equation", + "content": "N_0\\gg N_t" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "text", + "content": "The categories across different sessions are mutually exclusive, meaning that for any different " + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "inline_equation", + "content": "i,j\\in [0,T]" + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "inline_equation", + "content": "Y_{i}\\cap Y_{j} = \\varnothing" + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "text", + "content": " . Furthermore, during the training of the current session, data from previous sessions is not accessible. At the testing phase of the " + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "text", + "content": " -th session, the model evaluates against all previously encountered categories, with the cumulative category set defined as " + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_t = \\bigcup_{j = 0}^t Y_j" + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "text", + "content": " . Consequently, the test category space is given by " + }, + { + "bbox": [ + 55, + 389, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t^{\\mathrm{test}} = \\mathcal{V}_t" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 227, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 227, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 227, + 504 + ], + "type": "text", + "content": "3.2. Bi-Level calibration framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 512, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 548 + ], + "type": "text", + "content": "Our proposed training-free framework builds upon CLIP [24] and incorporates two essential components: intra-modal and inter-modal calibration (see Figure 2a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 563, + 230, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 230, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 230, + 574 + ], + "type": "text", + "content": "3.2.1 Intra-modal classifier calibration" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "In its standard usage, CLIP relies on simple prompt templates, (e.g., \"a photo of a [CLS].\") to construct zero-shot classifiers for visual tasks. However, this category-agnostic approach has inherent limitations. The generic nature of these templates fails to capture category-specific nuances, making them particularly inadequate for fine-grained classification tasks where detailed feature discrimination is crucial. A single, static template cannot effectively characterize the diverse visual characteristics within each category, especially when subtle distinctions between categories are essential." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 460 + ], + "type": "text", + "content": "To overcome these limitations, we leverage Large Language Models (LLMs) to generate dynamic, category-specific descriptions. These LLM-generated descriptions provide richer, more discriminative semantic representations for each category, enhancing the model's capacity to capture fine-grained features. We integrate this enhanced semantic information through an intra-modal calibration strategy within the textual modality:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 466, + 555, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 466, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 325, + 466, + 555, + 506 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {\\mu}} _ {c} ^ {T} = \\left(1 - \\lambda_ {T}\\right) \\boldsymbol {w} _ {c} + \\lambda_ {T} \\left(\\frac {1}{n _ {c}} \\sum_ {j = 1} ^ {n _ {c}} \\frac {g \\left(\\mathbf {t} _ {c , j}\\right)}{\\| g \\left(\\mathbf {t} _ {c , j}\\right) \\| _ {2}}\\right). \\tag {1}", + "image_path": "9da7f3d3b430bcf8dd3d7eba153532620296396aac219abbc92e62a6d0d82bff.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "In this formulation, " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\lambda_{T}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " controls the intensity of intra-modal calibration within the text modality, " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " denotes the number of descriptions for category " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " represents the text encoder, " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{c,j}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "-th LLM-generated description for category " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{\\mu}}_c^T" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " is the calibrated textual prototype, and " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_c" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " denotes the original CLIP Zero-Shot classifier weight for class " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ". This calibration mechanism preserves the robust generalization capabilities of the CLIP Zero-Shot classifier while incorporating fine-grained, category-specific semantic information. The latter term in Eq. (1) serves as a semantically enriched category description center, enhancing the model's discriminative power." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "In the FSCIL context, prototype-based classifiers are commonly employed, where the accuracy of prototype estimation directly impacts performance. During the base session, the abundance of samples allows for precise estimation of class prototypes. However, in incremental sessions" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9883" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "that introduce new classes, the limited number of samples inevitably leads to biased estimates of new class prototypes. Drawing inspiration from TEEN [35], we leverage the accurately estimated prototype weights of base classes to facilitate intra-modality calibration on the prototypes of incoming new classes within the visual modality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "type": "text", + "content": "For a given class " + }, + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "type": "text", + "content": ", we first utilize the encoded training image data to obtain the naive visual prototype:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 175, + 295, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 117, + 175, + 295, + 207 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {c} ^ {I} = \\frac {1}{m _ {c}} \\sum_ {j = 1} ^ {m _ {c}} \\frac {f (\\mathbf {x} _ {c , j})}{\\| f (\\mathbf {x} _ {c , j}) \\| _ {2}}, \\tag {2}", + "image_path": "b18ecc288282043945e24c181b12f9a06e05a0ba26b2a5468c67c930542e2307.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "content": " is the visual encoder, " + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "inline_equation", + "content": "m_{c}" + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "content": " is the number of images in class " + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{c,j}" + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "content": "-th image belonging to category " + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 215, + 296, + 299 + ], + "type": "text", + "content": ". Due to the inaccuracy in visual prototype estimation during the incremental sessions, the prototypes from the base session are used to recalibrate the new class prototypes by leveraging the similarity relationship between them. This process can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 305, + 295, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 305, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 79, + 305, + 295, + 356 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I} = \\left\\{ \\begin{array}{l l} \\boldsymbol {\\mu} _ {c} ^ {I} & , t = 0 \\\\ (1 - \\lambda_ {I}) \\boldsymbol {\\mu} _ {c} ^ {I} + \\lambda_ {I} \\sum_ {b = 1} ^ {| \\mathcal {Y} _ {0} |} s _ {b, c} \\boldsymbol {\\mu} _ {b} ^ {I} & , t > 0, \\end{array} \\right. \\tag {3}", + "image_path": "d177a165fe45b3958c6ca2fe06628a66c86b702fa545b0348701cdc4bb5ea572.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": " represents the task identifier, " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\lambda_I" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": " controls the strength of the visual intra-modal calibration, and " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "s_{b,c}" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": " denotes the normalized cosine similarity between the visual prototypes of class " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": " (from the base classes) and class " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": ". The similarity " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "s_{b,c}" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": " is defined as: " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "s_{b,c} = \\frac{e^{\\tau \\cdot \\langle \\mu_b, \\mu_c \\rangle}}{\\sum_{i=1}^{|\\mathcal{Y}_0|} e^{\\tau \\cdot \\langle \\mu_i, \\mu_c \\rangle}}" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": " represents the cosine similarity, and " + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 363, + 295, + 453 + ], + "type": "text", + "content": " is the temperature scaling parameter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 453, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 514 + ], + "type": "text", + "content": "Through this visual intra-modality calibration process, the prototypes of the new classes are adjusted to align more closely with the well-calibrated base class prototypes, allowing the new class prototypes to partially inherit the discriminative properties of the base class prototypes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 527, + 229, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 229, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 229, + 538 + ], + "type": "text", + "content": "3.2.2 Inter-modal classifier calibration" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Following intra-modal calibration, two classifier achieves enhanced accuracy within its respective modality. However, we posit that their full potential remains unrealized when operating in isolation. This is due to the fact that classifiers operating exclusively on linguistic data derive their knowledge from pre-trained models and lack explicit downstream knowledge for visual tasks. Conversely, classifiers functioning solely within the visual domain serve as metrics within the visual sub-feature space but are deficient in semantic understanding. Recognizing the complementary nature of these two modalities, we propose a training-free inter-modal calibration strategy, allowing domain-relevant downstream visual prior to guide domain-agnostic pretrained linguistic knowledge. Formally, this process can be" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 369, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 369, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 369, + 83 + ], + "type": "text", + "content": "described as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 381, + 91, + 555, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 555, + 106 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {c} = \\beta \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {T} + (1 - \\beta) \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}. \\tag {4}", + "image_path": "b2b096ddc1c19cc173718ac2270297f7677310a86005dcceeb4c97814c102406.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{c}" + }, + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "text", + "content": " is the inter-modal calibrated classifier, and " + }, + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "text", + "content": " is the calibration coefficient. During inference, we use cosine similarity to calculate the similarity between the sample features " + }, + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 113, + 555, + 161 + ], + "type": "text", + "content": " and the mixed category center:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 378, + 168, + 555, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 168, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 378, + 168, + 555, + 198 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {c} ^ {\\text {c a l i b}} = \\frac {f (\\mathbf {x}) ^ {\\top} \\boldsymbol {\\mu} _ {c}}{\\| f (\\mathbf {x}) \\| _ {2} \\cdot \\| \\boldsymbol {\\mu} _ {c} \\| _ {2}}. \\tag {5}", + "image_path": "f4f2a250045a8471e3ec12258a80d7f7c43f98d43a3139454975ea521318ad63.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "text", + "content": "This inter-modal calibration strategy effectively combines the strengths of both linguistic and visual modalities. It helps to enhance the overall classification performance by creating a more robust and comprehensive classifier that can better capture the nuances of visual tasks while maintaining a strong semantic understanding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 289, + 535, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 535, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 535, + 300 + ], + "type": "text", + "content": "3.3. Semantic and covariance-enhanced metric" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 307, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 403 + ], + "type": "text", + "content": "While our previous metrics effectively integrate linguistic and visual information and mitigate modal bias, they primarily measure sample distances from distribution centers. This isotropic approach, despite its simplicity, struggles to capture higher-order information and complex data relationships. To address these limitations, we propose leveraging statistical information from existing data for more comprehensive and efficient measurements, as shown in Figure 2b." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 417, + 465, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 417, + 465, + 429 + ], + "spans": [ + { + "bbox": [ + 314, + 417, + 465, + 429 + ], + "type": "text", + "content": "3.3.1 Modeling global covariance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "text", + "content": "Inspired by FeCAM [8], we introduce an anisotropic covariance metric to fully utilize the visual modality's statistical information. For a session " + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "text", + "content": ", we compute its covariance matrix " + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^t" + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "text", + "content": " using image embeddings " + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "text", + "content": " of all image data relevant to the session. To avoid singularity, we apply regularization with an identity matrix " + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 313, + 435, + 554, + 507 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 382, + 514, + 555, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 514, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 382, + 514, + 555, + 535 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {\\Sigma}} ^ {t} = \\boldsymbol {\\Sigma} ^ {t} + \\frac {\\gamma}{d} \\operatorname {t r} \\left(\\boldsymbol {\\Sigma} ^ {t}\\right) \\boldsymbol {I} _ {d}, \\tag {6}", + "image_path": "9cf8f87a2ce0a5d50fc2445b7505ca3e02643c1bd6ef0fc8225f00c43a8d4d29.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "text", + "content": " is a hyperparameter controlling the strength of regularization, " + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "text", + "content": " represents the feature dimension. Here, " + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{tr}(\\cdot)" + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "text", + "content": " denotes the trace operator of a matrix, and both " + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{tr}(\\pmb{\\Sigma}^t)" + }, + { + "bbox": [ + 313, + 541, + 555, + 613 + ], + "type": "text", + "content": " provide a prior estimation for the regularization strength. Subsequently, we construct a continuously evolving shared covariance matrix as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 348, + 620, + 555, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 620, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 348, + 620, + 555, + 647 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {t} = \\frac {| \\mathcal {Y} _ {t - 1} |}{| \\mathcal {Y} _ {t} |} \\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {t - 1} + \\left(1 - \\frac {| \\mathcal {Y} _ {t - 1} |}{| \\mathcal {Y} _ {t} |}\\right) \\tilde {\\boldsymbol {\\Sigma}} ^ {t}. \\tag {7}", + "image_path": "2a72d18a2c6b8d19e4d5347721cd25566be627b89da1980fc5e3ba16949541ed.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 653, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 688 + ], + "type": "text", + "content": "The evolved covariance matrix captures the statistical information of all categories. Given a test sample " + }, + { + "bbox": [ + 313, + 653, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 653, + 554, + 688 + ], + "type": "text", + "content": ", we calculate its score as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 325, + 693, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 693, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 325, + 693, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {c} ^ {\\mathrm {c o v}} (\\mathbf {x}) = - \\frac {1}{d} \\left(f (\\mathbf {x}) - \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}\\right) ^ {T} \\tilde {\\boldsymbol {\\Sigma}} _ {G} ^ {- 1} \\left(f (\\mathbf {x}) - \\tilde {\\boldsymbol {\\mu}} _ {c} ^ {I}\\right). \\tag {8}", + "image_path": "031a7008207f25623dcc08725fee2f5566ffb9591cb0757de5340c09afad1f94.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9884" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 291, + 171 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 291, + 171 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 291, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 291, + 171 + ], + "type": "image", + "image_path": "a994cc2e8cb606d02da4817343093fbf085c29173d024f0f17b4a66f014e3e48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 176, + 295, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 295, + 199 + ], + "type": "text", + "content": "Figure 3. In the inference phase, an ensemble strategy with category masking is utilized." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "text", + "content": "The division by feature dimension " + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "text", + "content": " serves as a normalization factor, preventing the softmax transformation into probabilities from collapsing into a single category. The negative sign maintains metric consistency, ensuring that a larger output score indicates a higher probability of the sample belonging to category " + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}_G" + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "text", + "content": " denotes the global covariance, with the superscript " + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 202, + 296, + 286 + ], + "type": "text", + "content": " omitted for clarity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 286, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 286, + 316 + ], + "type": "text", + "content": "3.3.2 Cross-modal category nearest neighbor metric" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": "Global covariance modeling effectively utilizes statistical information from the visual modality. To better leverage LLM-generated category descriptions, we propose a category-conditioned nearest-neighbor-based metric. Given a description text " + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{c,j}" + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": " encoded by " + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": ", we obtain its normalized feature representation " + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{c,j} = \\frac{g(\\mathbf{t}_{c,j})}{\\|g(\\mathbf{t}_{c,j})\\|_2}" + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": ". For a visual query sample, we compute its normalized representation " + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\frac{f(\\mathbf{x})}{\\|f(\\mathbf{x})\\|_2}" + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": ". The likelihood of this sample belonging to category " + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": " is determined by the maximum dot product between " + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": " and all description features in class " + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 322, + 296, + 448 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 456, + 295, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 456, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 127, + 456, + 295, + 476 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {c} ^ {\\mathrm {n n}} = \\max _ {j} \\left\\{\\boldsymbol {z} _ {c, j} ^ {\\top} \\cdot \\boldsymbol {v} \\right\\}, \\tag {9}", + "image_path": "06ee58e96e2f0b4c14de4509d3fcac5fe029884ca54cc46d5c473788c11b6d52.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_c^{\\mathrm{nn}}" + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "content": " indicates the likelihood that the query sample belongs to category " + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "content": ". This metric leverages maximum cross-modal similarity, effectively utilizing the diverse textual descriptions generated by LLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 255, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 255, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 255, + 563 + ], + "type": "text", + "content": "3.3.3 Inference score reorganization strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "We employ a globally shared covariance matrix to model distribution shape information in the visual feature space. While the base session contains substantial data for accurate covariance matrix estimation, experiments reveal that these base-derived covariance matrices perform poorly in distinguishing new class data. This limitation stems from two factors: (1) the covariance from extensive base data fails to align with new classes, and (2) the limited data available for new classes prevents accurate covariance estimation. To address this, we use the cross-modal category nearest neighbor metric for novel classes to compensate for the limitations of the second-order covariance metric, as shown in Figure 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "We first normalize the scores of three metrics using the softmax function: " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\mathrm{calib}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{calib}}\\right)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\mathrm{cov}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{cov}}\\right)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\mathrm{nn}} = \\mathrm{softmax}\\left(\\mathbf{s}^{\\mathrm{nn}}\\right)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". The final probability score " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " is computed differently for base and novel categories:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 336, + 140, + 554, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 140, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 336, + 140, + 554, + 174 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {c} = \\left\\{ \\begin{array}{l l} \\alpha \\mathbf {p} _ {c} ^ {\\text {c a l i b}} + (1 - \\alpha) \\mathbf {p} _ {c} ^ {\\text {c o v}}, & c \\in \\mathcal {Y} _ {0} \\\\ \\alpha \\mathbf {p} _ {c} ^ {\\text {c a l i b}} + (1 - \\alpha) \\mathbf {p} _ {c} ^ {\\text {n n}}, & c \\notin \\mathcal {Y} _ {0}. \\end{array} \\right. \\tag {10}", + "image_path": "64ca92cc29d5bdf23181a3bd2274a27c7d2d7ea661e3d1d8cc2a697cb987cf0f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "type": "text", + "content": "For base categories " + }, + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "type": "inline_equation", + "content": "(c\\in \\mathcal{V}_0)" + }, + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "type": "text", + "content": ", we use a weighted sum of mixed prototype scores and visual mahalanobis metric scores. For novel categories " + }, + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "type": "inline_equation", + "content": "(c\\notin \\mathcal{V}_0)" + }, + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "type": "text", + "content": ", we combine mixed prototype scores with cross-modal nearest neighbor metric scores. The final category prediction for a sample is determined by " + }, + { + "bbox": [ + 313, + 182, + 554, + 254 + ], + "type": "inline_equation", + "content": "\\arg \\max_c\\mathbf{p}_c" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 265, + 395, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 395, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 395, + 277 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 440, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 440, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 440, + 298 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three benchmarks following [22, 30, 41]: CIFAR100 [3], CUB200-2011 [32], and miniImageNet [25]. For CIFAR100 and miniImageNet, we partition each dataset into 60 base classes and 40 novel classes. The novel classes are further divided into eight incremental tasks, with each task structured as a 5-way 5-shot incremental session. The CUB200 dataset is split into 100 base classes and 100 novel classes, with each incremental task organized in a 10-way 5-shot format." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": "Implementation details. We adopt CLIP's ViT-B/16 architecture as the backbone for fair comparison across all methods. For category descriptions, we utilize fine-grained text generated by CuPL [23] and [26]. The hyperparameters are configured with intra-modal calibration coefficients " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\lambda_{T} = 0.5" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " (textual), and " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\lambda_{I} = 0.1" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " (visual), and a temperature parameter " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\tau = 16" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": ". The inter-modal calibration coefficient " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " is optimized through a validation process. Specifically, we divide the base session into training and validation subsets. Applying our bi-level calibration framework on the training subset, we evaluate performance on the validation subset. By incrementing " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " by 0.05, we identify the value that maximizes validation accuracy. This optimal " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " remains fixed throughout subsequent incremental sessions. The shared covariance parameter " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " is set to 1 for the base session and adjusted to 5 for incremental sessions. Finally, for score reorganization, we set " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " for CIFAR100 and miniImageNet, and " + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\alpha = 0.8" + }, + { + "bbox": [ + 313, + 411, + 555, + 626 + ], + "type": "text", + "content": " for CUB200." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 635, + 429, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 429, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 429, + 647 + ], + "type": "text", + "content": "4.2. Comparison results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "We compare our method with training-free approaches in the FSCIL scenario: CLIP Zero-Shot [24], Visual Prototype [29], TEEN [35] and FeCAM [8]. We present the performance curve over all benchmarks in Figure 4, while detailed session-wise performance on miniImageNet is shown" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9885" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 116, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Detailed session-wise accuracy, average accuracy (Avg) and performance degradation (PD) comparison on miniImageNet dataset. V and L represent the visual and language modalities, respectively. BiMC refers to the results obtained solely through the bi-level calibration framework, whereas " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\mathrm{BiMC}^{\\dagger}" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " incorporates the ensemble classifier strategy. The best results on each each sessions are indicated in bold, while the second-best results are underlined. " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " means higher is better, while " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " means lower is better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 116, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 62, + 116, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 116, + 547, + 224 + ], + "type": "table", + "html": "
MethodModalityAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
CLIP Zero-Shot [24]L91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
Visual Prototype [29]V92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
TEEN [35]V92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
FeCAM [8]V94.4293.7191.6190.5990.6290.2288.4388.1788.2090.666.22
BiMCV-L94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
BiMC†V-L95.4795.3493.8093.6393.5193.1892.5192.5692.4093.603.07
", + "image_path": "b76da2e39c5363fe7f1fb99d4cd85d6fbb9ee2357f955fa576f15f7306b53f1a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 71, + 234, + 227, + 353 + ], + "blocks": [ + { + "bbox": [ + 71, + 234, + 227, + 353 + ], + "lines": [ + { + "bbox": [ + 71, + 234, + 227, + 353 + ], + "spans": [ + { + "bbox": [ + 71, + 234, + 227, + 353 + ], + "type": "image", + "image_path": "57436e0131a346d47d8c328e86056be0ec678b944fad1e04a7d10a4ca767dc9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 354, + 185, + 363 + ], + "lines": [ + { + "bbox": [ + 127, + 354, + 185, + 363 + ], + "spans": [ + { + "bbox": [ + 127, + 354, + 185, + 363 + ], + "type": "text", + "content": "(a) miniImageNet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 368, + 493, + 377 + ], + "lines": [ + { + "bbox": [ + 116, + 368, + 493, + 377 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 493, + 377 + ], + "type": "text", + "content": "CLIP Zero-Shot Visual Prototype FeCAM BiMC BiMC" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 234, + 383, + 352 + ], + "blocks": [ + { + "bbox": [ + 228, + 234, + 383, + 352 + ], + "lines": [ + { + "bbox": [ + 228, + 234, + 383, + 352 + ], + "spans": [ + { + "bbox": [ + 228, + 234, + 383, + 352 + ], + "type": "image", + "image_path": "f23fa610cb82228685bcc854c7c2817a4182c2de560d7d661f47e5d6dfdb565b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 354, + 337, + 363 + ], + "lines": [ + { + "bbox": [ + 290, + 354, + 337, + 363 + ], + "spans": [ + { + "bbox": [ + 290, + 354, + 337, + 363 + ], + "type": "text", + "content": "(b) CIFAR100" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 385, + 235, + 542, + 352 + ], + "blocks": [ + { + "bbox": [ + 385, + 235, + 542, + 352 + ], + "lines": [ + { + "bbox": [ + 385, + 235, + 542, + 352 + ], + "spans": [ + { + "bbox": [ + 385, + 235, + 542, + 352 + ], + "type": "image", + "image_path": "e0e46788044a63a361463168d4f9be39d998a627aa93130e235db4d4d64bb120.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 354, + 490, + 363 + ], + "lines": [ + { + "bbox": [ + 451, + 354, + 490, + 363 + ], + "spans": [ + { + "bbox": [ + 451, + 354, + 490, + 363 + ], + "type": "text", + "content": "(c) CUB200" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 399, + 294, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 294, + 422 + ], + "type": "text", + "content": "in Table 1. Detailed results for CIFAR100 and CUB200 are provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 423, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 567 + ], + "type": "text", + "content": "Our approach demonstrates competitive performance across all three datasets and significantly outperforms unimodal baselines without requiring additional training, highlighting the importance of leveraging information from both modalities. On the miniImageNet dataset, while visual and textual prototype classifiers achieve comparable results, and although TEEN [35] and FeCAM [8] better utilize visual modality information, their improvement remains limited. Our framework employs a simple calibration strategy to integrate domain visual knowledge with pretrained linguistic knowledge, yielding up to " + }, + { + "bbox": [ + 54, + 423, + 295, + 567 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 54, + 423, + 295, + 567 + ], + "type": "text", + "content": " performance improvement over uni-modal models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "type": "text", + "content": "For CIFAR100, the language prototype shows better resistance to forgetting compared to the visual one, which is more susceptible to sequential task learning. This can be attributed to the dataset's low-resolution characteristics, resulting in insufficient visual information during incremental phases. Our calibration framework achieves approximately " + }, + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "type": "text", + "content": " improvement over a language-only classifier. In contrast, for the fine-grained CUB200 dataset, the visual prototype surpasses the language prototype due to limited trans" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 400, + 434, + 503 + ], + "blocks": [ + { + "bbox": [ + 85, + 381, + 523, + 392 + ], + "lines": [ + { + "bbox": [ + 85, + 381, + 523, + 392 + ], + "spans": [ + { + "bbox": [ + 85, + 381, + 523, + 392 + ], + "type": "text", + "content": "Figure 4. Performance curve of each incremental sessions on (a) miniImageNet, (b) CIFAR100 and (c) CUB200 datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 400, + 434, + 503 + ], + "lines": [ + { + "bbox": [ + 319, + 400, + 434, + 503 + ], + "spans": [ + { + "bbox": [ + 319, + 400, + 434, + 503 + ], + "type": "image", + "image_path": "08124e23b9f475316d421fc74513f9a8c45c39641d4dd10d2127ed4718cebc19.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 505, + 422, + 514 + ], + "lines": [ + { + "bbox": [ + 346, + 505, + 422, + 514 + ], + "spans": [ + { + "bbox": [ + 346, + 505, + 422, + 514 + ], + "type": "text", + "content": "(a) Old Classes Entropy" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 517, + 553, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 517, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 553, + 590 + ], + "type": "text", + "content": "Figure 5. Comparison of output confidence with different models. ferability of pre-trained language knowledge. By combining both modalities, our approach achieves approximately " + }, + { + "bbox": [ + 313, + 517, + 553, + 590 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 517, + 553, + 590 + ], + "type": "text", + "content": " performance improvement. Notably, our framework demonstrates robust resistance to forgetting, achieving the lowest performance drop across all metrics." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 435, + 400, + 550, + 503 + ], + "blocks": [ + { + "bbox": [ + 435, + 400, + 550, + 503 + ], + "lines": [ + { + "bbox": [ + 435, + 400, + 550, + 503 + ], + "spans": [ + { + "bbox": [ + 435, + 400, + 550, + 503 + ], + "type": "image", + "image_path": "b7dbfd7fc27e2554dffc90f6f9df37d8566f965101ec383ca61d203ddffc5342.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 505, + 541, + 514 + ], + "lines": [ + { + "bbox": [ + 458, + 505, + 541, + 514 + ], + "spans": [ + { + "bbox": [ + 458, + 505, + 541, + 514 + ], + "type": "text", + "content": "(b) Novel Classes Entropy" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "content": "One can observe that each modality's contribution to classification varies across datasets. Our proposed bi-level calibration framework effectively leverages the strengths of each modality and mitigates modality bias. Moreover, our ensemble inference strategy, denoted as " + }, + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{BiMC}^{\\dagger}" + }, + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "content": ", provides additional performance improvements." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 670, + 544, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 544, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 544, + 684 + ], + "type": "text", + "content": "4.3. Investigation of bi-level calibrated classifiers" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "In this section, we analyze two distinct attributes of the calibrated classifier to demonstrate the effectiveness of our pro" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1“textual prototype” refers to the classifier generated by CLIP in a zero-shot setting, which has the same meaning as “textual classifier” mentioned later in the paper." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9886" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 550, + 196 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Ablation studies of bi-level calibration framework on miniImageNet. Vis. and Lang. individually represent classifiers derived from the visual and language modalities. Intra-C and Inter-C refer to the strategies of intra-modal and inter-modal calibration, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 550, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 550, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 550, + 196 + ], + "type": "table", + "html": "
VisLangIntra-CInter-CAccuracy in each session(%) ↑Avg ↑PD ↓
012345678
✓92.5891.8889.3788.2488.4287.8686.3386.1686.0988.556.49
✓✓92.5892.0689.6988.6088.8088.3186.8486.7186.6688.925.92
✓91.2791.2589.7489.4388.9888.4286.9686.6486.1588.765.12
✓✓93.2793.3291.9191.5191.2190.7389.2889.2389.0191.054.26
✓✓✓94.6894.5892.8792.4992.5692.3191.5991.5891.5492.693.14
✓✓✓✓94.9094.8093.2192.9192.8892.6091.7891.8891.8192.973.09
", + "image_path": "e8359e8b473eaffeb161dec830755d702c86a703c5a7eec21bdccad9bb942864.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 251, + 288, + 331 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "type": "text", + "content": "Table 3. The left side of the table lists classification result combinations from three classifiers, while the right side shows the percentage " + }, + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "type": "text", + "content": " of test samples meeting each combination. Test sets from both the base and the final session were used for analysis." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 251, + 288, + 331 + ], + "lines": [ + { + "bbox": [ + 61, + 251, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 288, + 331 + ], + "type": "table", + "html": "
ModalityCIFAR100CUB200
VLV-LD0testD8testD0testD10test
✓✓✓65.3351.8159.0044.85
X✓✓6.9513.072.064.00
✓X✓6.105.5619.8321.40
XX✓1.202.101.081.98
", + "image_path": "b14a28402bd443d3d3244c8b664fc1b2abab6f0fbdeb0e21f588cb696eb32624.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 130, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 130, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 130, + 350 + ], + "type": "text", + "content": "posed framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 350, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 295, + 470 + ], + "type": "text", + "content": "The bi-level calibrated classifier enhances prediction confidence. We evaluate prediction confidence by calculating the entropy of predictions for each classifier in the final session across all three datasets. The entropy serves as an uncertainty indicator for model outputs, and we compute it separately for both base and novel categories, as detailed in Figure 5. The findings indicate that the bi-level calibrated classifier significantly reduces the entropy of predictions compared to uni-model classifiers, thereby enhancing prediction accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": "The bi-level calibrated classifier mitigates modality bias. To elucidate the superior performance of our method, we analyze the joint prediction outcomes across different classifiers in comparison to uni-modal approaches. Using " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " to denote correct classifications and " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " for classification errors, we observe that the calibrated classifier maintains accurate predictions even in cases where uni-modal classifiers fail, as demonstrated in the " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " rows of Table 3. This evidence suggests that the bi-level calibrated classifier produces more accurate and unbiased predictions, offering enhanced robustness compared to uni-modality classifiers. Remarkably, in certain rare instances, the modality-calibrated classifier successfully classifies samples that both uni-modal classifiers misclassify." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 148, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 148, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 148, + 659 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "content": "We conduct comprehensive ablation analyses on the miniImageNet dataset to evaluate the significance of each component within our proposed framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Bi-level calibration framework. Our framework com-" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 273, + 552, + 350 + ], + "blocks": [ + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "text", + "content": "Table 4. Ablation study of Semantic and Covariance-enhanced Metric on the miniImageNet dataset. To verify their effectiveness, we measured the performance of the base task " + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{base}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "text", + "content": ", the performance of last task " + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{last}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "text", + "content": " and the average performance " + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{avg.}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "text", + "content": " across all tasks. Additionally, we report the accuracy of base " + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{last}}^b" + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "text", + "content": " and novel " + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{last}}^n" + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "text", + "content": " class in last task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 273, + 552, + 350 + ], + "lines": [ + { + "bbox": [ + 316, + 273, + 552, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 552, + 350 + ], + "type": "table", + "html": "
Ablation\\( A_{base} \\)\\( A_{last} \\)\\( A_{last}^{b} \\)\\( A_{last}^{n} \\)\\( A_{avg} \\)
BiMC94.9091.8193.4389.3892.97
+ MGC95.4791.5194.3787.2292.91
+ CMNN94.8592.1593.3090.4293.18
\\( BiMC^† (w/o mask.) \\)95.4792.2094.1289.3293.37
\\( BiMC^† \\)95.4792.4093.2091.2093.60
", + "image_path": "83a91a36c76ca892894c516fca47b7218f434752cc848ab2db686a2e740c3e50.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 369, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 369, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 312, + 369, + 555, + 525 + ], + "type": "text", + "content": "prises four components: visual modality information, linguistic modality information, intra-modal calibration strategy, and inter-modal calibration strategy. As shown in Table 2, single-modality classifiers enhanced with intra-modal calibration demonstrate improved performance, attributable to more accurate estimation of the uni-modal classifier. When employing inter-modal calibration, we observe significant performance gains compared to uni-modal approaches, owing to the fusion strategy's effectiveness in addressing modality biases. The combination of both intramodal and inter-modal calibration strategies yields the best performance, with comprehensive knowledge integration leading to minimal forgetting rates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "type": "text", + "content": "Semantic and covariance-enhanced metric. We examine the effectiveness of Modeling Global Covariance (MGC) and Cross-modal category Nearest Neighbor metric (CMNN) in Table 4. The global covariance modeling strategy enhances base category performance in both the base session and the final incremental session, though at the cost of reduced recognition performance for novel classes, a consequence of the mismatch between global covariance (dominated by base data) and new classes. The semantic nearest neighbor strategy, conversely, improves novel class recognition performance. Implementing both strategies simultaneously without masking further enhances overall performance. Finally, the ensemble inference mechanism with masks fully leverages the advantages of both metrics, achieving superior performance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9887" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 116, + 294, + 183 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": "Table 5. Comparison with three trainable methods on the miniImageNet dataset. " + }, + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": " is the number of parameters which require training. " + }, + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "inline_equation", + "content": "\\Delta A_{\\mathrm{last}}" + }, + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": " reflects the last task's performance gap between our method and the comparative one." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 116, + 294, + 183 + ], + "lines": [ + { + "bbox": [ + 58, + 116, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 58, + 116, + 294, + 183 + ], + "type": "table", + "html": "
MethodNpAbaseAlastAavg.ΔAlast
CPE-CLIP [5]400k90.2382.7786.13+9.63
CLIP-M3[6]46k96.0092.5094.10-0.10
LP-DiF [10]8.1k96.3491.6893.76+0.72
BiMC†095.4792.4093.600.00
", + "image_path": "abcb726911d4542274275d41792cf9dc8dad4acfeee6d0ffd7c8cb560e55de39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 196, + 289, + 328 + ], + "blocks": [ + { + "bbox": [ + 63, + 196, + 289, + 328 + ], + "lines": [ + { + "bbox": [ + 63, + 196, + 289, + 328 + ], + "spans": [ + { + "bbox": [ + 63, + 196, + 289, + 328 + ], + "type": "image", + "image_path": "e5e5c32a2306938d04c274f5f4b1f6d572633587d94cedd59476b55792fccafc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 329, + 296, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 361 + ], + "type": "text", + "content": "Figure 6. CLIP Feature Space Visualization. Squares, circles, and pentagrams represent textual, visual features, and calibrated classifiers respectively. Different colors denote different categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 156, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 156, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 156, + 380 + ], + "type": "text", + "content": "4.5. Further analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "content": "Comparison with training-required methods. To comprehensively evaluate our approach, we conduct comparisons with three training-based methods [5, 6, 10]. As demonstrated in Table 5, our parameter-free adaptation approach to FSCIL tasks achieves comparable or superior performance to methods that require additional training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 653 + ], + "type": "text", + "content": "Visualization of Feature Space. We visualized the image and text feature distributions on the miniImageNet dataset, as depicted in Figure 6. Our analysis showed that in the unified feature space encoded by CLIP, visual and textual features are located at diametrically opposed ends, exemplifying a modality gap phenomenon [16]. Textual features showed more consistent convergence, in contrast to the visual features of the same category, which were more dispersed within the feature space. Notably, despite being distinct, these features share similar representations in the visual subspace (e.g., the two purple dashed ellipses in Figure 6) but are clearly distinguishable in the textual subspace. Furthermore, after modality calibration, the classifiers, influenced by visual priors, demonstrate a shift from the textual modality to the visual modality, thereby alleviating modality bias and enhancing classification capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Analysis of hyper-parameters. Our framework involves two critical hyper-parameters: the inter-modal calibration parameter " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and the ensemble classifier weighting coefficient " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", both constrained to [0, 1]. Figure 7a illustrates how performance varies sessions for different values of " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 545, + 182 + ], + "type": "image", + "image_path": "35c68906a9f825959cb60b12f63cef715f2e5fae892fb8df54a56ca338c1af74.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 183, + 536, + 194 + ], + "lines": [ + { + "bbox": [ + 333, + 183, + 536, + 194 + ], + "spans": [ + { + "bbox": [ + 333, + 183, + 536, + 194 + ], + "type": "text", + "content": "(a) Influence of " + }, + { + "bbox": [ + 333, + 183, + 536, + 194 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 333, + 183, + 536, + 194 + ], + "type": "text", + "content": " and session-wise accuracy on miniImageNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 200, + 427, + 281 + ], + "blocks": [ + { + "bbox": [ + 320, + 200, + 427, + 281 + ], + "lines": [ + { + "bbox": [ + 320, + 200, + 427, + 281 + ], + "spans": [ + { + "bbox": [ + 320, + 200, + 427, + 281 + ], + "type": "image", + "image_path": "dd4e19dacae4c2f23db4f1b41b3841d6a1154a411644a9d4ec0c955313180a82.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 282, + 432, + 292 + ], + "lines": [ + { + "bbox": [ + 315, + 282, + 432, + 292 + ], + "spans": [ + { + "bbox": [ + 315, + 282, + 432, + 292 + ], + "type": "text", + "content": "(b) Influence of " + }, + { + "bbox": [ + 315, + 282, + 432, + 292 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 315, + 282, + 432, + 292 + ], + "type": "text", + "content": " on miniImageNet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 443, + 199, + 549, + 280 + ], + "blocks": [ + { + "bbox": [ + 443, + 199, + 549, + 280 + ], + "lines": [ + { + "bbox": [ + 443, + 199, + 549, + 280 + ], + "spans": [ + { + "bbox": [ + 443, + 199, + 549, + 280 + ], + "type": "image", + "image_path": "cfe36ef6505c16c1ec2265ce9357aa6f988328cced9a8110867173d92e2427d7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 281, + 549, + 292 + ], + "lines": [ + { + "bbox": [ + 443, + 281, + 549, + 292 + ], + "spans": [ + { + "bbox": [ + 443, + 281, + 549, + 292 + ], + "type": "text", + "content": "(c) Influence of " + }, + { + "bbox": [ + 443, + 281, + 549, + 292 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 443, + 281, + 549, + 292 + ], + "type": "text", + "content": " on CIFAR100" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": "Figure 7. Analysis on hyper-parameters " + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": ", we present heatmaps illustrating accuracy under varying " + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": " across different sessions. For " + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": ", we report the average task accuracy " + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "inline_equation", + "content": "A_{avg}" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": " and the final task accuracy " + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "inline_equation", + "content": "A_{last}" + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": " under different settings." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": "We observe that the " + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": " values that perform well in the base session maintain their effectiveness throughout incremental tasks. The optimal solution identified through the validation search in the base session (indicated by the blue rectangular box) closely approaches the global optimum. For the parameter " + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": ", results are presented in Figures 7b and 7c. When " + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": ", the method reduces to a bi-level calibrated classifier. The performance reaches its peak at " + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": ", optimizing both the average and final accuracy metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 464, + 388, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 388, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 388, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 483, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 615 + ], + "type": "text", + "content": "We explore CLIP models in the context of Few-shot Class-Incremental Learning. To mitigate forgetting, we treat the pre-trained model as a black box and employ a simple yet effective training-free bi-level modality calibration strategy. The intra-modal calibration achieves accurate classifier estimates within each modality, while the inter-modal calibration combines knowledge from both visual and textual modalities. An ensemble inference strategy integrating covariance and nearest-neighbor metrics enhances the accuracy of prediction. Extensive experiments validate the effectiveness of our approach." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 616, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 556, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), National Natural Science Foundation of China (62192783, 62276128), Guangdong Basic and Applied Basic Research Foundation (2024A1515011340) Jiangsu Natural Science Foundation (BK20221441), and the Australian Research Council's Discovery Project (DP220101784)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9888" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "text", + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "text", + "content": "[2] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 194, + 295, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 226 + ], + "type": "text", + "content": "[3] Krizhevsky Alex. Learning multiple layers of features from tiny images. https://www.cs.toronto.edu/kriz/learning-features-2009-TR.pdf, 2009.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 293 + ], + "type": "text", + "content": "[4] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "text", + "content": "[5] Marco D'Alessandro, Alberto Alonso, Enrique Calabrés, and Mikel Galar. Multimodal parameter-efficient few-shot class incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3393-3403, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 352, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 352, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 352, + 294, + 396 + ], + "type": "text", + "content": "[6] Thang Doan, Sima Behpour, Xin Li, Wenbin He, Liang Gou, and Liu Ren. A streamlined approach to multimodal few-shot class incremental learning for fine-grained datasets. arXiv preprint arXiv:2403.06295, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 397, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 397, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 61, + 397, + 296, + 452 + ], + "type": "text", + "content": "[7] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 454, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 454, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 61, + 454, + 296, + 509 + ], + "type": "text", + "content": "[8] Dipam Goswami, Yuyang Liu, Bartlomiej Twardowski, and Joost van de Weijer. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 511, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 511, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 61, + 511, + 296, + 555 + ], + "type": "text", + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 296, + 610 + ], + "type": "text", + "content": "[10] Zitong Huang, Ze Chen, Zhixing Chen, Erjin Zhou, Xinxing Xu, Rick Siow Mong Goh, Yong Liu, Chunmei Feng, and Wangmeng Zuo. Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint arXiv:2401.01598, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 666 + ], + "type": "text", + "content": "[11] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In European Conference on Computer Vision, pages 432-448. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[13] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11463-11473, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "type": "text", + "content": "[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 259 + ], + "type": "text", + "content": "[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9029, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 554, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 315 + ], + "type": "text", + "content": "[16] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In Advances in Neural Information Processing Systems, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 555, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 360 + ], + "type": "text", + "content": "[17] Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23638-23647, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 554, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 393 + ], + "type": "text", + "content": "[18] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 449 + ], + "type": "text", + "content": "[19] Bolin Ni, Hongbo Zhao, Chenghao Zhang, Ke Hu, Gaofeng Meng, Zhaoxiang Zhang, and Shiming Xiang. Enhancing visual continual learning with language-guided supervision. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 24068-24077, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 554, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 492 + ], + "type": "text", + "content": "[20] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "type": "text", + "content": "[21] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "type": "text", + "content": "[22] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 593, + 554, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 554, + 647 + ], + "type": "text", + "content": "[23] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9889" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[25] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[26] Oindrila Saha, Grant Van Horn, and Subhransu Maji. Improved zero-shot classification by adapting vlms with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17552, 2024. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[27] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 296 + ], + "type": "text", + "content": "[28] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "type": "text", + "content": "[29] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "type": "text", + "content": "[30] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24183-24192, 2023. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[31] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "text", + "content": "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[33] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11899-11909, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "text", + "content": "[34] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "type": "text", + "content": "[35] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 709 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "text", + "content": "[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[38] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[39] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "text", + "content": "[40] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "type": "text", + "content": "[41] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2021. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "type": "text", + "content": "[42] Hai Zhang, Junzhe Xu, Shanlin Jiang, and Zhenan He. Simple semantic-aided few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28588–28597, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "text", + "content": "[43] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "text", + "content": "[44] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 9046–9056, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "type": "text", + "content": "[45] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 565, + 553, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 618 + ], + "type": "text", + "content": "[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 620, + 553, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 620, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 316, + 620, + 553, + 662 + ], + "type": "text", + "content": "[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 665, + 553, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 665, + 553, + 709 + ], + "spans": [ + { + "bbox": [ + 316, + 665, + 553, + 709 + ], + "type": "text", + "content": "[48] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9890" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_content_list.json b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e473c857c4e36bd14a1c578c7686e41d3bc64b9 --- /dev/null +++ b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sihao Liu $^{1}$ Yibo Yang $^{2\\dagger}$ Xiaojie Li $^{3}$ David A. Clifton $^{4}$ Bernard Ghanem $^{2}$", + "bbox": [ + 174, + 202, + 823, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Harbin Institute of Technology $^{2}$ King Abdullah University of Science and Technology", + "bbox": [ + 197, + 222, + 792, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Harbin Institute of Technology (Shenzhen) $^{4}$ University of Oxford", + "bbox": [ + 267, + 239, + 725, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online continual learning (OCL) seeks to learn new tasks from data streams that appear only once, while retaining knowledge of previously learned tasks. Most existing methods rely on replay, focusing on enhancing memory retention through regularization or distillation. However, they often overlook the adaptability of the model, limiting the ability to learn generalizable and discriminative features incrementally from online training data. To address this, we introduce a plug-and-play module, S6MOD, which can be integrated into most existing methods and directly improve adaptability. Specifically, S6MOD introduces an extra branch after the backbone, where a mixture of discretization selectively adjusts parameters in a selective state space model, enriching selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. We further design a class-conditional routing algorithm for dynamic, uncertainty-based adjustment and implement a contrastive discretization loss to optimize it. Extensive experiments combining our module with various models demonstrate that S6MOD significantly enhances model adaptability, leading to substantial performance gains and achieving the state-of-the-art results. The code is available at https://github.com/MyToumaKazusa/S6MOD.", + "bbox": [ + 91, + 323, + 483, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continual learning (CL) [10, 29, 31, 43] is a task that requires models to efficiently acquire new abilities from incoming data. It tests the model's ability to adapt to dynamically changing environments while retaining existing knowledge. For example, autonomous vehicles are expected to continuously learn new driving environments and traffic regulations, thereby improving detection accuracy under complex driving scenes. From the perspective of task format, CL can be divided into offline and online [10, 43].", + "bbox": [ + 89, + 739, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike offline CL [7, 33], online CL (OCL) [10, 16] aligns with real-world implementations as it requires data to arrive sequentially in mini-batches and allows only one epoch of training, posing greater challenges for efficient adaptation with data accessible for learning only once.", + "bbox": [ + 511, + 292, + 906, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the challenging OCL, existing studies widely rely on the replay technique [17] that selectively stores a subset of old-class data to strengthen the memory ability of existing knowledge and mitigate catastrophic forgetting. Building on replay, various methods have been proposed with regularization, buffer allocation, and distillation strategies [17, 18, 34, 44, 46, 50]. For example, MIR [3] aims to mitigate mutual interference among tasks through a retrieval strategy, OCM [17] seeks to reduce forgetting by maximizing mutual information between different tasks, and CCLDC [44] adopts collaborative learning and distillation to improve plasticity. However, learning generalizable and discriminative features incrementally from online training data is still intractable, and imprecise features will impede and even mislead the replay strategies [45]. Therefore, inducing accurate and efficient adaptation with limited data access is critically important for OCL and there still remains significant potential for improvement.", + "bbox": [ + 511, + 369, + 908, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, selective state space models (SSMs), also known as S6 in Mamba [13], have shown promising results in modeling long-range dependencies with computational efficiency. Selective SSMs introduce a selective scan mechanism to make the interactions among sequential states aware of input context, and have been applied to vision tasks by integrating various scan directions [26, 58]. In addition to the success in sequence modeling [5, 9, 24, 25, 35, 49], selective SSMs also exhibit stronger adaptability than static parameters in few-shot class-incremental learning due to the dynamic operation weights, as reported in a recent pioneering work [22]. Despite the increased capacity of adaptation, directly applying selective SSMs to OCL will be infeasible because it is challenging for selective SSMs to capture a precise discretization pattern with limited context from online data that appears only once for training. Moreover, a single discretization mechanism may fail to capture all the", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20502", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "essential details in some nonlinear dynamic systems or systems with multi-scale characteristics [1, 15].", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we develop a plug-and-play branch based on selective state space model and class-conditional mixture of discretization. The branch with an equiangular tight frame classifier [29, 37, 48, 51, 52, 57] can be integrated on any existing OCL methods to supervise the original classification head and improve the adaptation ability in OCL. Considering multiple discretization methods can exhibit varying sensitivity to the dynamic characteristics of a system, we introduce a mixture of discretization into SSMs to enrich selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. This flexibility allows our method to comprehensively capture the system's complex dynamic features, particularly in cases of rapid state changes and evolving systems as exemplified by OCL.", + "bbox": [ + 88, + 122, + 482, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to guide the mixture of discretization for OCL, where a delicate trade-off between maintaining the stability of old knowledge and fostering the plasticity for new ability is needed, we further introduce a class-conditional routing to aggregate the discretization patterns. Concretely, we maintain a feature prototype for each class to calculate the class uncertainty based on the margin among different classes. For classes with less uncertainty, fewer discretization patterns will be aggregated to stabilize the abilities already acquired for these classes, while for classes with large uncertainty, more discretization patterns will be included to allocate additional capacity for adapting to these undeveloped abilities. After aggregation of the mixture of discretization, we employ a contrastive discretization loss that enforces within-class consistency and between-class diversity, contributing to the learning of generalizable and discriminative features after the selective scan. In experiments, we integrate our plug-and-play branch on numerous OCL methods and significant improvements can be consistently observed on multiple datasets.", + "bbox": [ + 91, + 351, + 482, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 107, + 654, + 434, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To enhance the adaptability of existing methods for OCL, we propose S6MOD, a plug-and-play module based on selective state space model and class-conditional mixture of discretization, which can strengthen the base method.", + "- We further develop a class-conditional routing strategy that dynamically satisfy both objectives of maintaining the stability of knowledge already acquired and fostering the plasticity for undeveloped abilities. A contrastive discretization loss is employed to facilitate the learning of generalizable and discriminative features.", + "- In experiments, our method can be easily integrated on different OCL methods and is compatible with distillation techniques, to consistently achieve significant improvements on multiple datasets including CIFAR-10, CIFAR-100, and Tiny-ImageNet." + ], + "bbox": [ + 89, + 671, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Learning(CL). CL can be categorized into three types based on the implementation methods [10, 29, 31, 43]: regularization-based, parameter isolation-based, and replay-based. Regularization-based methods [2, 23, 42, 56] introduce regularization terms to restrict changes in model parameters, thereby protecting the already learned knowledge. Parameter isolation-based methods [36, 38, 47, 54] typically assign different model parameters or subnetworks to different tasks to avoid interference between tasks. Replay-based methods [6, 8, 27, 33, 34] perform joint training by either storing a portion of old data or generating pseudo-data. This allows the model to revisit old data while learning new data, thereby reducing forgetting [53].", + "bbox": [ + 511, + 114, + 906, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Online Continual Learning(OCL). OCL is a specialized form of CL designed to test the ability to continuously learn and update as data arrives in real-time streams [10, 16, 28]. This means that OCL can only train for a single epoch and typically process only individual samples or mini batches at a time. Replay-based methods are widely used in OCL [17]. ER [34] introduces the combination of cross-entropy loss with a random buffer. OCM [17] uses mutual information maximization to reduce feature bias and preserve past knowledge. GSA [18] proposes a gradient-based adaptive optimization method to address dynamic training bias. On-Pro [46] uses online prototype equilibrium to address shortcut learning problem. Some methods introduce knowledge distillation based on replay. CCL-DC [44] introduces Collaborative Continual Learning (CCL) and Distillation Chain (DC) to enhance model plasticity. MOSE [50] alleviates forgetting by integrating multi-level supervision and reverse self-distillation. However, these models carry a risk when the features used to distill are not precise. In comparison, our method S6MOD as a plug-and-play module is applicable to most OCL methods to improve the adaptability of these models, enabling them to learn generalizable and discriminative features more efficiently.", + "bbox": [ + 511, + 334, + 906, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Selective State Space Model (S6) Selective State Space Model (S6) [13] has gained increasing interest as an alternative to self-attention [41] with lower computational complexity. S6 enhances the S4 model [14] by introducing the selective scan mechanism, and its effectiveness in vision tasks has been extensively studied [26, 58]. For example, Vmamba [26] introduces SS2D, a cross-scanning mechanism for images, facilitating the extension of Mamba to process vision data. A recent study Mamba-FSCIL [22] leverages the dynamic weights and sequence modeling capability of Mamba to achieve dynamic adaptation in few-shot class-incremental learning. But their method relies on an effective selective scan learned from training data. In OCL,", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20503", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the online data that appears only once for training provides limited context for Mamba models, and thus poses challenges to capture a precise discretization pattern. Different from these studies, our method integrates S6 with class-conditional mixture of discretization and effectively helps to improve the adaptability for OCL. Some existing works have introduced mixture of experts (MoE) into Mamba [4, 24, 32], following the design of switch-transformer [12]. Our method differs from them in that each expert in our mixture of discretization is only a simple linear layer for computation efficiency, and discretization patterns are aggregated with our class-conditional gating to balance between maintaining stability of old knowledge and allocating capacity for learning new abilities.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 314, + 230, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition of OCL", + "text_level": 1, + "bbox": [ + 89, + 339, + 336, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OCL requires a model to continuously update itself from a data stream, with each mini-batch containing new data sampled from a changing distribution $D_{t}$ . At each time step $t$ , the model receives a mini-batch of data $\\{(s_{i}^{(t)}, r_{i}^{(t)})\\}_{i=1}^{n_{t}}$ where $s_{i}^{(t)}$ represents the input data, $r_{i}^{(t)}$ represents the corresponding labels, and $n_{t}$ is the number of samples. The goal of OCL is to sequentially update the model to adapt to both gradual and task-specific shifts in the data distribution. Given a pre-trained network $f_{\\theta}$ with parameters $\\theta$ , the model updates its parameters by solving an optimization problem $\\arg \\min_{\\theta} L(f_{\\theta}(s), r)$ , where $L$ is the loss function, thereby enabling it to adapt effectively to the evolving data distribution.", + "bbox": [ + 89, + 361, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. State Space Models", + "text_level": 1, + "bbox": [ + 89, + 570, + 276, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State Space Models (SSMs) can be viewed as linear time-invariant systems. They map input sequences $x(t) \\in \\mathbb{R}$ to output sequences $y(t) \\in \\mathbb{R}$ through hidden states $h(t) \\in \\mathbb{R}^N$ . Mathematically, these models can be represented as linear ordinary differential equations (ODE):", + "bbox": [ + 89, + 592, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = A h (t) + B x (t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 674, + 480, + 698 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny (t) = C h (t),\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 695, + 308, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $A \\in \\mathbb{R}^{N \\times N}$ represents the state transition matrix, while $B \\in \\mathbb{R}^{N \\times 1}$ and $C \\in \\mathbb{R}^{1 \\times N}$ denote the mapping matrices from input to latent state and from latent state to output, respectively, where $N$ indicates the size of the hidden state.", + "bbox": [ + 89, + 719, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To apply the SSM to real discrete data, zero-order hold (ZOH) [14] is employed. It discretizes the continuous parameters $A$ , $B$ , and $C$ using the time scale parameter $\\Delta$ . The discretized SSM equations can be rewritten as follows:", + "bbox": [ + 89, + 795, + 483, + 856 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\bar {A} h _ {t - 1} + \\bar {B} x _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 866, + 482, + 891 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = C h _ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 888, + 290, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, Gu [13] proposed a new parameterization method for SSM with a selective scan mechanism, which is known as S6 and serves as the core of the Mamba model. To enhance S6's capability in processing visual data, Liu et al. [26] introduced SS2D, which can serialize image data from four different directions. Given the input data $x$ , the output $\\overline{x}$ processed by SS2D can be expressed as follows:", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {x} = \\operatorname {S S 2 D} (x) = \\sum_ {i = 1} ^ {4} \\operatorname {S 6} (\\operatorname {s c a n} (x, i)). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 204, + 905, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 252, + 687, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will first introduce the overall structure of our plug-and-play module S6MOD in Sec. 4.1. After that, we will present the detailed design of the state space model with mixture of discretization in Sec. 4.2. Then, we will introduce the class-conditional routing and its corresponding contrastive discretization loss in details in Sec. 4.3. Finally, we will specify the optimization details in Sec. 4.4.", + "bbox": [ + 511, + 277, + 906, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overall structure", + "text_level": 1, + "bbox": [ + 511, + 390, + 683, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better enhance existing methods' ability in learning generalizable and discriminative features incrementally from online training data, we propose a plug-and-play module named S6MOD that can be easily applied to existing OCL methods, as shown in Fig. 1 (a). Overall, S6MOD strengthens the original method by introducing a plug-and-play branch after the backbone. It consists of a block and a fixed equiangular tight frame (ETF) classifier [51, 52]. The intermediate features $\\mathbf{F}$ generated by the backbone are duplicated and sent into the classification head of the original base method, and the selective state space model (SSM) branch introduced by our method.", + "bbox": [ + 511, + 412, + 905, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our SSM branch, $\\mathbf{F}$ is projected into two paths through an MLP. The first one $\\mathbf{X} = f_{x}(\\mathbf{F})$ is used to perform selective scan with our class-conditional mixture of discretization, as will be detailed in later subsections. The other one $\\mathbf{Z} = f_{z}(\\mathbf{F})$ performs a gating mechanism as commonly adopted in Mamba models [22, 26, 30]. The output feature of this branch can be formulated as:", + "bbox": [ + 511, + 594, + 905, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} = \\operatorname {S i L U} (\\mathbf {Z}) \\otimes \\operatorname {S 6 M O D} (\\operatorname {S i L U} (\\operatorname {C o n v} (\\mathbf {X}))). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 707, + 903, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To assist the learning of the base method for generalizable and discriminative features, we adopt a regularization, $\\mathcal{L}_{\\mathrm{Diff}}$ , on the predicted distribution of the base method, which is the KL divergence between $P$ and $Q$ and can be formulated as:", + "bbox": [ + 511, + 731, + 905, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {D i f f}} = \\sum_ {i} P (i) \\log \\left(\\frac {P (i)}{Q (i)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 811, + 903, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P$ represents the predicted distribution of the base method, and $Q$ refers to the output prediction after the ETF classifier of our SSM branch, i.e., $Q = \\mathrm{softmax}(\\mathbf{W}_{\\mathrm{ETF}}\\boldsymbol {\\mu})$", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20504", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8338a9ed7bc850ad5ab28a83ba3eee2816c554b458ff4a891119dafa92d612f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 89, + 844, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/245cf302e4e89c134e247a787c1a5df0a1d38b2d97dc794fa850c28b88415bce.jpg", + "image_caption": [ + "Figure 1. Framework of S6MOD. Our method (a) introduces a plug-and-play branch after the backbone, where features are learned through S6MOD and supervised by the ETF classifier to guide the base method. S6MOD (c) utilizes MoE to enhance the discretization of SSM and applies class-conditional routing (b) to dynamically adjust the discretization based on the uncertainty. Finally, we use a contrastive discretization loss (d) to supervise the learning of both generalizable and discriminative features." + ], + "image_footnote": [], + "bbox": [ + 153, + 224, + 841, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. S6 with Mixture of Discretization", + "text_level": 1, + "bbox": [ + 89, + 561, + 382, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We integrate the selective space state model with a mixture of discretization, borrowing the concept of mixture of experts [11, 19, 39, 55]. Its structure is shown in Fig. 1 (c). Similar to the typical SSM structure, we also use an MLP layer to produce the projection matrices $B$ and $C$ from the input features, which can be expressed as follows:", + "bbox": [ + 89, + 584, + 483, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} = f _ {B} (\\hat {\\mathbf {X}}), \\quad \\mathbf {C} = f _ {C} (\\hat {\\mathbf {X}}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 688, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}$ denotes the input of the selective scan module S6MOD, i.e., $\\hat{\\mathbf{X}} = \\mathrm{SiLU}(\\mathrm{Conv}(\\mathbf{X}))$ . To enable selective SSM to capture a precise discretization pattern with limited context from online data, we develop a sparse MoE system for the discretization transformation $\\pmb{\\Delta}$ in S6 models, utilizing specialized projection layers to enrich the discretization patterns and adaptively selecting the most sensitive discretization based on the current dynamics. Specifically, each discretization candidate is produced by a projection layer as follows:", + "bbox": [ + 89, + 720, + 483, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Delta} _ {i} = f _ {\\Delta i} (\\hat {\\mathbf {X}}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input feature $\\mathbf{X}$ is also fed into a sparse gating mechanism, which produces associated importance weights $w_{i}$ corresponding to each discretization $\\pmb{\\Delta}_{i}$ . Through the class-conditional gating that will be introduced in Sec. 4.3, we dynamically control the number of discretization patterns, $N_{k}$ , to be selected for each class $k$ based on the input features $\\hat{\\mathbf{X}}$ . Finally, the selected discretization patterns are aggregated with their importance weights as follows:", + "bbox": [ + 511, + 561, + 906, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Delta} = \\sum_ {i \\in \\Omega (\\hat {\\mathbf {X}})} w _ {i} \\cdot \\boldsymbol {\\Delta} _ {i}, \\quad | \\Omega (\\hat {\\mathbf {X}}) | = N _ {k}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 688, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega (\\hat{\\mathbf{X}})$ is the index set of the top- $N_{k}$ discretization candidates of $\\hat{\\mathbf{X}}$ , according to their importance weights.", + "bbox": [ + 511, + 732, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It's noteworthy that $\\Delta$ is important because it controls the decay rate during the state update process, enabling the model to flexibly select and retain important information when handling long sequences, while avoiding the accumulation of redundant and irrelevant information [13], especially in cases of rapid changing states and evolving systems, as exemplified by OCL. Our method with mixture of discretization facilitates the learning of a precise selective scan pattern with limited data context in OCL.", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20505", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Class-Conditional Routing", + "text_level": 1, + "bbox": [ + 89, + 90, + 331, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With more discretization patterns selected, the model is able to allocate more capacity for learning new knowledge, but may cause a significant shift of the model that impairs the existing ability. If we can calculate $N_{k}$ based on the misclassification probability of each class $k$ , we can dynamically control the capacity of mixture of discretization based on the learning conditions of all classes. Therefore, in order to guide the mixture of discretization to strike a balance between maintaining the stability of old knowledge and fostering the plasticity for new ability, we further design class-conditional routing. During training, we maintain the feature prototypes $\\mathbf{M} = \\{\\mathbf{M}_c\\}$ by moving average, where $\\mathbf{M}_c$ is the within-class feature mean for class $c$ . Then, we estimate the class uncertainty $\\sigma_{k}$ based on the average margin of class $k$ with different classes,", + "bbox": [ + 89, + 112, + 483, + 338 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {k} = \\operatorname {A v g} _ {c} \\left\\{\\exp \\left(- \\lambda_ {0} \\| \\mathbf {M} _ {k} - \\mathbf {M} _ {c} \\| _ {2}\\right) \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 349, + 480, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_0$ is a hyper-parameter and $\\sigma_{k}$ denotes the class uncertainty for class $k$ ranging from $(0,1)$ . A large $\\sigma_{k}$ indicates that the class center $k$ has narrow margins with the other class centers, and thus tends to be misclassified, while a small $\\sigma_{k}$ happens when class center $k$ is distant from the other class centers with a lower likelihood of being misclassified. In inference, we replace $\\mathbf{M}_k$ with the input feature $\\hat{\\mathbf{X}}$ and calculate its uncertainty by Eq. (9) with the feature prototypes $\\mathbf{M}_c$ of all classes.", + "bbox": [ + 89, + 378, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we multiply the uncertainty by the number of total discretization candidates $N$ , to get the discretization number to select for class $k$ or input feature $\\hat{\\mathbf{X}}$ as follows,", + "bbox": [ + 89, + 513, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nN _ {k} = \\operatorname {c e i l} \\left(N \\cdot \\sigma_ {k}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 571, + 480, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{ceil}$ refers to the operation that rounds up $N \\cdot \\sigma_k$ to the nearest integer.", + "bbox": [ + 89, + 599, + 482, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The institution is that classes that are more prone to misclassification have smaller margins with the other classes in the feature space, and thus needs to aggregate more discretization patterns to allocate additional capacity for adapting to these undeveloped abilities. Conversely, for categories that are easier to classify with low uncertainty, a smaller $N_{k}$ is favored to only include the most likely discretization patterns, which can spare optimization efforts for the uncertain classes and stabilize these already acquired abilities. The combination of class-conditional routing and our mixture of discretization achieves dynamic structures dependent on input features and classes, such that the final prediction of each class selectively optimizes the corresponding discretization patterns.", + "bbox": [ + 89, + 628, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further reduce interaction among different classes and facilitate the learning of generalizable and discriminative features, we also incorporate contrastive discretization loss function $\\mathcal{L}_{\\mathrm{Cont}}$ , as shown in Fig. 1 (d). It encourages $\\pmb{\\Delta}$ to", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "have within-class consistency and between-class diversity as follows:", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {C o n t}} = - \\frac {1}{B ^ {2}} \\sum_ {m = 1} ^ {B} \\sum_ {n = 1} ^ {B} \\left(1 _ {y _ {m} = y _ {n}} - 1 _ {y _ {m} \\neq y _ {n}}\\right) \\frac {\\boldsymbol {\\Delta} _ {m} \\cdot \\boldsymbol {\\Delta} _ {n}}{\\| \\boldsymbol {\\Delta} _ {m} \\| \\| \\boldsymbol {\\Delta} _ {n} \\|} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 130, + 903, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ is the batch size, $\\Delta_{m}$ represents the aggregated $\\Delta$ by Eq. (8) for input $m$ , $y_{m}$ denotes the class of $m$ and $1_{y_m = y_n}$ is an indicator function that takes a value of 1 when the condition $y_{m} = y_{n}$ holds, and 0 otherwise.", + "bbox": [ + 511, + 185, + 906, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Optimization", + "text_level": 1, + "bbox": [ + 511, + 255, + 651, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the previously mentioned loss functions in Eqs. (5) and (11), we also directly used the classification loss function $\\mathcal{L}_{\\mathrm{DR}}$ specifically designed for ETF classifier [51, 52] to supervise our introduced plug-and-play branch, and the loss function $\\mathcal{L}_{\\mathrm{z}}$ for maintaining load balancing among the discretization patterns [59]. In summary, when integrating S6MOD on a base method, the overall loss function can be expressed as follows:", + "bbox": [ + 511, + 277, + 906, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {S 6 M O D}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 412, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S 6 M O D}} = \\mathcal {L} _ {\\mathrm {D R}} + \\alpha \\cdot \\mathcal {L} _ {\\text {D i f f}} + \\beta \\cdot \\mathcal {L} _ {\\text {C o n t}} + \\mathcal {L} _ {\\mathrm {z}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 455, + 903, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are hyperparameters.", + "bbox": [ + 511, + 479, + 754, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 510, + 645, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 511, + 535, + 710, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We test three datasets that are widely used in OCL, i.e., CIFAR-10 (10 classes) [20], CIFAR-100 (100 classes) [20] and TinyImageNet (200 classes) [21]. The dataset settings we adopted are the same as those in CCLDC[46]. We divide CIFAR-10 into 5 tasks, with 2 classes per task; CIFAR-100 into 10 tasks, with 10 classes per task; and TinyImageNet into 100 tasks, with 2 classes per task. Further details about the datasets will be provided in the supplementary materials.", + "bbox": [ + 511, + 556, + 906, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To demonstrate the effectiveness and applicability of our module, we conduct tests on 7 typical and state-of-the-art methods, including ER [34], OCM [17], OnPro [46], OCM-CCLDC [44], OnPro-CCLDC [44], MOSE[50] and MOE-MOSE [50].", + "bbox": [ + 511, + 714, + 905, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. To ensure a fair comparison, we applied the same hyperparameter settings to each baseline and the methods combined with our module (S6MOD). All the above methods use ResNet-18 as the backbone, without pretraining. For the streaming input data, we set the batch size to 10, and for the samples drawn from the buffer,", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20506", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9e6379da3c440f0fc14e92b480273ceff9c3428ecc589a4c5aa85156070a66b6.jpg", + "table_caption": [ + "Table 1. Average Accuracy (\\%, higher is better) on three benchmark datasets with difference memory buffer size $M$ , with and without our proposed S6MOD module. All values are averages of 10 runs." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]56.68±1.8962.32±4.1324.47±0.7231.89±1.4539.41±1.8110.82±0.7919.16±1.4224.71±2.52
ER + Ours57.88±3.3065.80±2.1626.50±2.2334.55±1.6639.61±3.1610.94±1.4719.67±1.3625.62±1.73
OCM [17]68.19±1.7573.15±1.0528.02±0.7435.69±1.3642.22±1.0618.36±0.9526.74±1.0231.94±1.19
OCM + Ours70.57±1.1475.31±1.1032.44±1.3538.97±2.2845.49±1.5819.12±1.6027.20±1.8332.31±1.72
OnPro [46]70.47±2.1274.70±1.5127.22±0.7733.33±0.9341.59±1.3814.32±1.4021.13±2.1226.38±2.18
OnPro + Ours72.30±1.0875.14±0.9928.84±0.5937.57±0.8144.14±1.3617.27±0.6525.62±0.7929.20±1.00
OCM-CCLDC [44]74.14±0.8577.66±1.4635.00±1.1543.34±1.5151.43±1.3723.36±1.1833.17±0.9739.25±0.88
OCM-CCLDC + Ours74.45±1.2078.21±1.0336.02±1.7744.40±2.2652.53±0.2123.68±1.0233.59±0.7239.53±1.02
OnPro-CCLDC [44]74.49±2.1478.64±1.4234.76±1.1241.89±0.8250.01±0.8521.81±1.0232.00±0.7238.18±1.02
OnPro-CCLDC + Ours75.03±1.0379.51±0.6336.07±0.7643.92±0.9650.85±0.5222.03±0.8233.29±0.3538.51±0.91
MOSE [50]61.02±1.4770.74±1.1835.05±0.3445.06±0.3254.53±0.7818.23±0.7330.98±0.6338.71±0.44
MOSE + Ours63.74±1.5572.54±0.1435.43±0.6245.38±0.3154.75±0.5219.03±0.8332.03±0.8138.18±1.02
MOE-MOSE [50]62.54±1.5972.18±1.2937.32±0.3447.03±0.5755.62±0.7220.61±0.6932.52±0.3338.41±0.53
MOE-MOSE + Ours63.86±1.1973.16±0.5337.76±0.5147.25±0.4356.32±0.5321.03±0.7733.53±0.5140.11±0.26
", + "bbox": [ + 176, + 123, + 820, + 356 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2bd1111f36b59465c0519f3b88d72b8a9bcb9c84f958d913e1509efed4087a4d.jpg", + "table_caption": [ + "Table 2. Average Forgetting (%, lower is better) on three benchmark datasets with difference memory buffer size $M$ , with and without our proposed S6MOD module. All values are averages of 10 runs." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]33.16±3.5020.94±6.7932.65±1.7822.20±2.2613.29±1.9858.38±1.6946.87±1.6040.77±2.45
ER + Ours31.25±3.4416.71±2.8630.96±2.0719.02±2.1312.97±2.6158.03±2.1945.83±1.4937.62±1.53
OCM [17]13.68±4.2511.63±2.6214.99±1.559.16±1.753.76±1.1626.12±1.6319.74±1.3015.92±1.47
OCM + Ours13.58±3.269.68±3.4715.99±1.8211.60±1.495.46±0.9824.23±2.0719.12±1.8714.66±1.75
OnPro [46]17.94±3.6914.20±2.6016.76±2.4712.42±1.396.72±0.9428.01±1.5923.52±1.7520.32±1.70
OnPro + Ours7.13±1.445.73±1.8116.03±1.699.81±1.325.07±0.8920.92±1.1216.82±0.9718.15±1.91
OCM-CCLDC [44]11.59±2.249.18±2.0316.69±2.3610.07±1.373.99±0.7826.16±1.9019.99±1.9615.56±1.06
OCM-CCLDC + Ours15.81±2.6911.28±1.8316.64±1.859.08±2.063.48±0.3125.29±1.9514.89±0.7411.55±1.31
OnPro-CCLDC [44]19.89±4.0114.62±2.7528.93±2.1920.23±1.0310.55±1.8928.21±1.5820.85±1.1316.17±0.63
OnPro-CCLDC + Ours17.23±3.1611.51±2.3322.26±1.1812.49±1.594.88±1.3626.73±1.5116.69±0.3112.23±1.37
MOSE [50]30.36±1.6920.27±1.2737.54±0.4325.89±0.4513.60±0.5947.16±1.4124.96±0.6215.51±0.33
MOSE + Ours27.38±1.9417.92±0.3437.09±0.6325.80±0.3915.55±0.6146.60±1.0724.39±0.7220.15±0.74
MOE-MOSE [50]29.39±1.7919.24±1.4935.17±0.3023.99±0.5112.81±0.7441.98±1.4622.22±0.3813.94±0.59
MOE-MOSE + Ours28.13±1.4117.98±0.5634.76±0.6623.84±0.7714.03±0.4841.26±1.0221.39±0.5814.91±0.28
", + "bbox": [ + 171, + 398, + 816, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the batch size is set to 64. We retain each baseline's original data augmentation methods without modifying the base methods. For more details, please refer to the supplementary materials.", + "bbox": [ + 89, + 642, + 482, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 89, + 712, + 184, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We combine our method with both classical and state-of-the-art approaches on CIFAR-10, CIFAR-100, and TinyImageNet. The experimental results in Table 1 demonstrate the universality of our method, as it consistently improves accuracy across numerous baselines. On CIFAR-10 and CIFAR-100, our method generally leads to a $1\\%$ improvement, while on TinyImageNet, we achieve significant gains under OCM, OnPro and MOE-MOSE. Notably, in settings like CIFAR100 $(M = 2k)$ and TinyImageNet $(M = 5k)$ , OnPro result in an approximate $4.2\\%$ and $4.5\\%$ improvement, respectively. It is also worth mention", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing that even for state-of-the-art distillation-based methods like CCLDC and MOE-MOSE, the performance improvements are quite substantial. For instance, OnPro-CCLDC with the S6MOD integration achieve an impressive $79.51\\%$ on CIFAR10 ( $M = 1k$ ), and MOE-MOSE with S6MOD reach $56.32\\%$ on CIFAR100 ( $M = 5k$ ) and $40.1\\%$ on TinyImageNet ( $M = 10k$ ).", + "bbox": [ + 511, + 642, + 906, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, as shown in Table 2, our module is also highly effective in reducing model forgetting, achieving lower forgetting rates in most settings, with some cases showing particularly significant improvements. This can be attributed to our class-conditional routing. In a few baseline settings, the inclusion of our module led to a higher forgetting rate, but this does not imply that the model is more prone to forgetting. In these cases, we think the higher forgetting rates are often due to the stronger learning capabilities of the model. Analyzing both Table 1 and Table 2, we", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20507", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c5c2c0261d5c7df568ac30198318d265eb04025187677bc700f730815bcb22f5.jpg", + "table_caption": [ + "Table 3. Ablation studies on CIFAR-100 ( $M = 2k$ ). The \"branch\" refers to using SS2D[26] as the core of the S6MOD, without routing and $\\mathcal{L}_{\\mathrm{Cont}}$ . \"routing\" represents class-conditional routing. S6MOD refers to the composition of (branch + class-conditional routing + $\\mathcal{L}_{\\mathrm{Cont}}$ ). All values are averaged over 10 runs." + ], + "table_footnote": [], + "table_body": "
MethodAcc. ↑AF ↓
OnPro33.33±0.9312.42±1.39
OnPro + branch (with LDiff)36.61±1.0410.28±1.59
OnPro + branch + routing36.92±0.728.85±0.81
OnPro + branch + LCont36.93±0.849.70±1.42
OnPro + S6MOD37.57±0.819.81±1.32
", + "bbox": [ + 129, + 162, + 434, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "can find that even in these settings with higher forgetting rates, the accuracy performance remains quite competitive.", + "bbox": [ + 89, + 258, + 483, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 299, + 256, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1, the S6MOD module we propose is very simple and can be directly applied to existing baselines. By adding an extra branch and using the $\\mathcal{L}_{\\mathrm{Diff}}$ to supervise the original classification head, we can achieve a significant improvement. To demonstrate this, we conduct an ablation study where we only add the extra branch and $\\mathcal{L}_{\\mathrm{Diff}}$ to the baseline. The results in Table 3 confirm this, showing a $3.3\\%$ improvement with supervision from the extra branch alone.", + "bbox": [ + 89, + 321, + 483, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, to verify the effectiveness and adaptability of class-conditional routing and contrastive discretization loss $\\mathcal{L}_{\\mathrm{Cont}}$ , we conduct separate ablation experiments. As shown in Table 3, when added to the base method with the extra branch, the improvements brought by adding class-conditional routing or $\\mathcal{L}_{\\mathrm{Cont}}$ alone are not significant, with accuracy gains of only $0.31\\%$ and $0.32\\%$ , respectively. However, when both components are used together, the improvement becomes significant, with an approximate $1\\%$ increase in accuracy, and the forgetting rate still decreases. This indicates that when our module is correctly combined, it can significantly enhances the model's adaptability by learning more generalizable and discriminative features through dynamic parameters.", + "bbox": [ + 89, + 459, + 483, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Analysis", + "text_level": 1, + "bbox": [ + 89, + 681, + 194, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Feature Embeddings. To verify that S6MOD improves the adaptability of the baseline and helps it learn more discriminative features, we use t-SNE [40] visualization for analysis. We compare with the baseline under the CIFAR-100 ( $M = 2k$ ) setting. As shown in Fig. 2, we visualize the features of the samples in the buffer after the final training, which are the features $\\mathbf{F}$ output by the backbone just before they are passed to the classifier for classification. The features shown in Fig. 2 (a) and (c) are very scattered, with only a few classes showing some degree of aggregation. Most classes do not have clear separations between other classes, making it difficult to distinguish. In contrast, Fig. 2 (b) and (d) present the feature distribution", + "bbox": [ + 89, + 703, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dee5334ca91198dbfe3bf8014e387bbe9f54663822eedfd246385311b5230cce.jpg", + "image_caption": [ + "(a) OnPro" + ], + "image_footnote": [], + "bbox": [ + 553, + 90, + 705, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3242c70be735c2f67ee71d100c33f530f7965e3de772d801e7f595b8e646aaec.jpg", + "image_caption": [ + "(b) OnPro + S6MOD" + ], + "image_footnote": [], + "bbox": [ + 718, + 90, + 867, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/709d94a5ea9bf0fa553c1cb3994394fc1712f0bdbe1b9883ba92168e7fb8b260.jpg", + "image_caption": [ + "(c) OCM" + ], + "image_footnote": [], + "bbox": [ + 555, + 224, + 705, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a16046da30c40d9d32a88fdf34a4d7365e1a22ee49c631b58ec7bf23e6218f1.jpg", + "image_caption": [ + "(d) OCM + S6MOD" + ], + "image_footnote": [], + "bbox": [ + 718, + 224, + 867, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eed097c6eb67dfc27acafa55c89044fa14f762eef4a279430749d907bd6dedb3.jpg", + "image_caption": [ + "Figure 2. t-SNE visualization of memory data at the end of training on CIFAR-100 ( $M = 2k$ ), showcasing baseline in (a) and (c) and baseline combined with S6MOD in (b) and (d). Different colors represent different classes.", + "(a) New-Task Accuracy", + "Figure 3. Impact of dynamically selecting different $N_{k}$ values on the ability to learn new tasks and prevent forgetting of old tasks: New-Task accuracy represents the model's accuracy on the current task. We conduct experiments by setting $N_{k} = 1$ , $N_{k} = N$ , and calculating $N_{k}$ through class-conditional routing. The dataset used is CIFAR-100 ( $M = 2k$ )." + ], + "image_footnote": [], + "bbox": [ + 522, + 426, + 689, + 559 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af18e09dfca09cb93f91eeadd68afffb784431d74cc3427634b9938df0e4d7f1.jpg", + "image_caption": [ + "(b) Average Forgetting" + ], + "image_footnote": [], + "bbox": [ + 714, + 434, + 883, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "after adding our module. We can observe that many categories have noticeably converged, forming small clusters with consistent intra-class compactness and clear inter-class separations. These features are much easier to distinguish and more conducive to classification by the classifier, which aligns with the significant improvements we demonstrate in the experiments.", + "bbox": [ + 511, + 672, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Class-conditional Routing. We further investigate the effectiveness of class-conditional routing. Under the CIFAR-100 ( $M = 2k$ ) setting, we design experiments to compare the impact of a fixed $N_{k}$ and dynamically calculated $N_{k}$ using class-conditional routing. Specifically, we set $N_{k} = N$ , meaning all patterns are activated, and $N_{k} = 1$ , meaning only the pattern with the highest prob", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20508", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d881aab3f0270682ff6073d3b905e972b9c605e67a5cb95129438cea13ad3041.jpg", + "table_caption": [ + "Table 4. Impact of hyper-parameters $\\alpha$ and $\\beta$ on performance in the CIFAR-100 dataset ( $M = 2k$ ). $\\alpha$ and $\\beta$ are the weights for $\\mathcal{L}_{\\mathrm{Diff}}$ and $\\mathcal{L}_{\\mathrm{Cont}}$ , respectively. Acc. and AF represent the average accuracy and the average forgetting over 5 runs, respectively." + ], + "table_footnote": [], + "table_body": "
Impact of hyper-parameter α
Value0.010.10.515
Acc.36.80±1.1737.09±1.0937.55±1.0837.47±0.9437.57±0.81
AF10.49±1.019.65±0.399.05±1.259.16±2.119.81±1.32
Impact of hyper-parameter β
Value0.1151025
Acc.36.50±1.2437.30±1.0337.40±0.3337.26±0.6137.57±0.81
AF10.24±0.4310.12±1.0910.10±0.609.98±0.669.81±1.32
", + "bbox": [ + 94, + 151, + 480, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ability is activated. To facilitate analysis, we present the accuracy of the model on the latest task (the current task) in Fig. 3 (a). As shown in Fig. 3 (a), when $N_{k} = N$ , all patterns are activated, and thanks to the inclusion of more patterns with different weights and larger parameters, the model has greater capacity to adapt to new tasks, resulting in a stronger ability to learn new knowledge. When $N_{k} = 1$ , only the pattern with the highest probability is activated, significantly reducing the model's ability to learn the new task. However, as shown in Fig. 3 (b), when $N_{k} = N$ , since all patterns are updated by the current task, the model has a higher risk of forgetting. When $N_{k} = 1$ , only the pattern with the highest probability is updated, and tasks do not interfere with each other, so forgetting is less likely. Combining the analysis of the entire Fig. 3, the effect of our proposed class-conditional routing is both intuitive and significant. It not only helps the model achieve strong learning ability for new tasks, comparable to the $N_{k}$ setting, but also provides excellent anti-forgetting performance, similar to $N_{k} = 1$ . This is due to the dynamic adjustment of $N_{k}$ based on uncertainty: for samples that are prone to misclassification, class-conditional routing assigns a larger $N_{k}$ , enabling more extensive learning. For easier-to-classify samples, the model's learning ability is higher, and class-conditional routing assigns a smaller $N_{k}$ to enhance anti-forgetting.", + "bbox": [ + 89, + 306, + 483, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity Analysis of Hyper-parameters $\\alpha$ and $\\beta$ . $\\alpha$ and $\\beta$ are the weights for $\\mathcal{L}_{\\mathrm{Diff}}$ and $\\mathcal{L}_{\\mathrm{Cont}}$ , respectively. We conduct experiments on the sensitivity of them, and the results are shown in Table 4. It can be observed that when the weights are low, the optimization effect of the loss is relatively weak, both accuracy and forgetting rate do not perform well. Gradually increasing the weights, we find that after a certain point, the impact of the weights on the results becomes relatively stable, with no more drastic changes.", + "bbox": [ + 89, + 700, + 483, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity Analysis of total discretization patterns number $N$ . We conduct experiments on the total number of discretization patterns in the module. The results are shown", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f880a619373a0e7ad688d11998bc6e7c6870626c6df92de9fa2343e03a5add4.jpg", + "image_caption": [ + "(a) New-Task Average Accuracy (b) Average Forgetting w/o class-w/o class-conditional routing conditional routing" + ], + "image_footnote": [], + "bbox": [ + 522, + 99, + 692, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb54119721d252a830c3c0a1d8ad2d5d16dd42b102922c8d56f9c5fc62e1d87c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 99, + 885, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e14ddd1b7dc09cbadec1ced7e49b5f258b1ef9e7751088c28be86dc5b04a503.jpg", + "image_caption": [ + "(c) New-Task Average Accuracy" + ], + "image_footnote": [], + "bbox": [ + 522, + 250, + 694, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fbee29e319e85a8d06830280eae17ac9f6c1dfce5abbd379d95c25f1343a3af.jpg", + "image_caption": [ + "(d) Average Forgetting", + "Figure 4. Impact of pattern number $N$ . The dataset used is CIFAR-100 ( $M = 2k$ )." + ], + "image_footnote": [], + "bbox": [ + 714, + 250, + 887, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 4. As the number $N$ increases, the model gains a stronger ability to learn new tasks, but at the same time, its tendency to forget becomes more pronounced. Furthermore, by comparing (a)(b) and (c)(d) in Fig. 4, we can conclude that class-conditional routing can significantly reduce forgetting with only a slight decrease in new task accuracy, which in turn leads to an overall improvement in performance. This indirectly supports the effectiveness of class-conditional routing in our ablation studies.", + "bbox": [ + 511, + 424, + 906, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 573, + 633, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a plug-and-play module S6MOD to enhance the adaptability of existing OCL methods by introducing selective state space models (SSMs) with a class-conditional mixture of discretization. By integrating a dynamic discretization mechanism and leveraging class-conditional strategies, our method can efficiently allocate discretization patterns based on class uncertainty, improving both the model's generalization and its ability to adapt to new data. Experimental results on multiple OCL datasets demonstrate that our method consistently optimize existing techniques, contributing to more robust OCL systems.", + "bbox": [ + 511, + 599, + 906, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 779, + 692, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The research reported in this publication was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940. This research used computational resources provided by IBEX managed by the Supercomputing Core Laboratory at KAUST.", + "bbox": [ + 511, + 801, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20509", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Assyr Abdulle, E Weinan, Björn Engquist, and Eric Vanden-Eijnden. The heterogeneous multiscale method. Acta Numerica, 21:1-87, 2012. 2", + "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2", + "[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. NeurIPS, 32, 2019. 1", + "[4] Quentin Anthony, Yury Tokpanov, Paolo Glorioso, and Beren Millidge. Blackmamba: Mixture of experts for statespace models. arXiv preprint arXiv:2402.01771, 2024. 3", + "[5] Fengxiang Bie, Yibo Yang, Zhongzhu Zhou, Adam Ghanem, Minjia Zhang, Zhewei Yao, Xiaoxia Wu, Connor Holmes, Pareesa Golnari, David A Clifton, et al. Renaissance: A survey into ai text-to-image generation in the era of large model. TPAMI, 2024. 1", + "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. *NeurIPS*, 33: 15920-15930, 2020. 2", + "[7] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 1", + "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2", + "[9] Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688, 2023. 1", + "[10] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 44(7):3366-3385, 2021. 1, 2", + "[11] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 4", + "[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3", + "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3, 4", + "[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3", + "[15] John Guckenheimer and Philip Holmes. Nonlinear oscillations, dynamical systems, and bifurcations of vector fields. Springer Science & Business Media, 2013. 2" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Nuwan Gunasekara, Bernhard Pfahringer, Heitor Murilo Gomes, and Albert Bifet. Survey on online streaming continual learning. In *IJCAI*, pages 6628-6637, 2023. 1, 2", + "[17] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In ICML, 2022. 1, 2, 5, 6", + "[18] Yiduo Guo, Bing Liu, and Dongyan Zhao. Dealing with cross-task class discrimination in online continual learning. In CVPR, 2023. 1, 2", + "[19] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. Neural Computation, 3(1):79-87, 1991. 4", + "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5", + "[21] Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015. 5", + "[22] Xiaojie Li, Yibo Yang, Jianlong Wu, Bernard Ghanem, Liqiang Nie, and Min Zhang. Mamba-fscil: Dynamic adaptation with selective state space model for few-shot class-incremental learning. arXiv preprint arXiv:2407.06136, 2024. 1, 2, 3", + "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 40(12):2935-2947, 2017. 2", + "[24] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 1, 3", + "[25] Sihao Liu, Mingyi Liu, Tianyu Jiang, Shuang Yu, Hanchuan Xu, and Zhongjie Wang. GSL-mash: Enhancing mashup creation service recommendations through graph structure learning. In Service-Oriented Computing, pages 176-191, Singapore, 2025. Springer Nature Singapore. 1", + "[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. ArXiv, abs/2401.10166, 2024. 1, 2, 3, 7", + "[27] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 30, 2017. 2", + "[28] Zheda Mai, Ruiwen Li, Hyunwoo Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. In CVPR, pages 3589-3599, 2021. 2", + "[29] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 1, 2", + "[30] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3", + "[31] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 2", + "[32] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, Michał Krutul, Jakub Krajewski, Szymon Antoniak, Pi" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20510", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "otr Mioś, Marek Cygan, and Sebastian Jaszcur. Moemamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 3", + "[33] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2", + "[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. *NeurIPS*, 32, 2019. 1, 2, 5, 6", + "[35] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 1", + "[36] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2", + "[37] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In CVPR, pages 23933-23942, 2024. 2", + "[38] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, pages 4548-4557. PMLR, 2018. 2", + "[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 4", + "[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 7", + "[41] A Vaswani. Attention is all you need. NeurIPS, 2017. 2", + "[42] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. NeurIPS, 34:22379-22391, 2021. 2", + "[43] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. TPAMI, 2024. 1, 2", + "[44] Maorong Wang, Nicolas Michel, Ling Xiao, and Toshihiko Yamasaki. Improving plasticity in online continual learning via collaborative learning. In CVPR, pages 23460-23469, 2024. 1, 2, 5, 6", + "[45] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, pages 18764-18774, 2023. 1", + "[46] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, 2023. 1, 2, 5, 6", + "[47] Mitchell Wortsman, Vivek Ramanujan, Rosanne Liu, Aniruddha Kembhavi, Mohammad Rastegari, Jason Yosinski, and Ali Farhadi. Supermasks in superposition. NeurIPS, 33:15173-15184, 2020. 2", + "[48] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 1", + "[50] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In CVPR, pages 23670-23680, 2024. 1, 2, 5, 6", + "[51] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, pages 37991-38002. Curran Associates, Inc., 2022. 2, 3, 5", + "[52] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR, 2023. 2, 3, 5", + "[53] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. Corda: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In NeurlPS, 2024. 2", + "[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2", + "[55] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 4", + "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995. PMLR, 2017. 2", + "[57] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19560, 2023. 2", + "[58] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2", + "[59] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5" + ], + "bbox": [ + 516, + 92, + 903, + 739 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20511", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_model.json b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_model.json new file mode 100644 index 0000000000000000000000000000000000000000..344e2b26ebb7e59b6af77450280474e63c8d98e8 --- /dev/null +++ b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.89, + 0.175 + ], + "angle": 0, + "content": "Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.203, + 0.825, + 0.221 + ], + "angle": 0, + "content": "Sihao Liu\\(^{1}\\) Yibo Yang\\(^{2\\dagger}\\) Xiaojie Li\\(^{3}\\) David A. Clifton\\(^{4}\\) Bernard Ghanem\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.223, + 0.794, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Harbin Institute of Technology \\(^{2}\\)King Abdullah University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.24, + 0.727, + 0.257 + ], + "angle": 0, + "content": "\\(^{3}\\)Harbin Institute of Technology (Shenzhen) \\(^{4}\\)University of Oxford" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.324, + 0.485, + 0.686 + ], + "angle": 0, + "content": "Online continual learning (OCL) seeks to learn new tasks from data streams that appear only once, while retaining knowledge of previously learned tasks. Most existing methods rely on replay, focusing on enhancing memory retention through regularization or distillation. However, they often overlook the adaptability of the model, limiting the ability to learn generalizable and discriminative features incrementally from online training data. To address this, we introduce a plug-and-play module, S6MOD, which can be integrated into most existing methods and directly improve adaptability. Specifically, S6MOD introduces an extra branch after the backbone, where a mixture of discretization selectively adjusts parameters in a selective state space model, enriching selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. We further design a class-conditional routing algorithm for dynamic, uncertainty-based adjustment and implement a contrastive discretization loss to optimize it. Extensive experiments combining our module with various models demonstrate that S6MOD significantly enhances model adaptability, leading to substantial performance gains and achieving the state-of-the-art results. The code is available at https://github.com/MyToumaKazusa/S6MOD." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Continual learning (CL) [10, 29, 31, 43] is a task that requires models to efficiently acquire new abilities from incoming data. It tests the model's ability to adapt to dynamically changing environments while retaining existing knowledge. For example, autonomous vehicles are expected to continuously learn new driving environments and traffic regulations, thereby improving detection accuracy under complex driving scenes. From the perspective of task format, CL can be divided into offline and online [10, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.369 + ], + "angle": 0, + "content": "Unlike offline CL [7, 33], online CL (OCL) [10, 16] aligns with real-world implementations as it requires data to arrive sequentially in mini-batches and allows only one epoch of training, posing greater challenges for efficient adaptation with data accessible for learning only once." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.909, + 0.643 + ], + "angle": 0, + "content": "To tackle the challenging OCL, existing studies widely rely on the replay technique [17] that selectively stores a subset of old-class data to strengthen the memory ability of existing knowledge and mitigate catastrophic forgetting. Building on replay, various methods have been proposed with regularization, buffer allocation, and distillation strategies [17, 18, 34, 44, 46, 50]. For example, MIR [3] aims to mitigate mutual interference among tasks through a retrieval strategy, OCM [17] seeks to reduce forgetting by maximizing mutual information between different tasks, and CCLDC [44] adopts collaborative learning and distillation to improve plasticity. However, learning generalizable and discriminative features incrementally from online training data is still intractable, and imprecise features will impede and even mislead the replay strategies [45]. Therefore, inducing accurate and efficient adaptation with limited data access is critically important for OCL and there still remains significant potential for improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, selective state space models (SSMs), also known as S6 in Mamba [13], have shown promising results in modeling long-range dependencies with computational efficiency. Selective SSMs introduce a selective scan mechanism to make the interactions among sequential states aware of input context, and have been applied to vision tasks by integrating various scan directions [26, 58]. In addition to the success in sequence modeling [5, 9, 24, 25, 35, 49], selective SSMs also exhibit stronger adaptability than static parameters in few-shot class-incremental learning due to the dynamic operation weights, as reported in a recent pioneering work [22]. Despite the increased capacity of adaptation, directly applying selective SSMs to OCL will be infeasible because it is challenging for selective SSMs to capture a precise discretization pattern with limited context from online data that appears only once for training. Moreover, a single discretization mechanism may fail to capture all the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.238, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20502" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "essential details in some nonlinear dynamic systems or systems with multi-scale characteristics [1, 15]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.123, + 0.483, + 0.35 + ], + "angle": 0, + "content": "To this end, we develop a plug-and-play branch based on selective state space model and class-conditional mixture of discretization. The branch with an equiangular tight frame classifier [29, 37, 48, 51, 52, 57] can be integrated on any existing OCL methods to supervise the original classification head and improve the adaptation ability in OCL. Considering multiple discretization methods can exhibit varying sensitivity to the dynamic characteristics of a system, we introduce a mixture of discretization into SSMs to enrich selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. This flexibility allows our method to comprehensively capture the system's complex dynamic features, particularly in cases of rapid state changes and evolving systems as exemplified by OCL." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.352, + 0.483, + 0.653 + ], + "angle": 0, + "content": "In order to guide the mixture of discretization for OCL, where a delicate trade-off between maintaining the stability of old knowledge and fostering the plasticity for new ability is needed, we further introduce a class-conditional routing to aggregate the discretization patterns. Concretely, we maintain a feature prototype for each class to calculate the class uncertainty based on the margin among different classes. For classes with less uncertainty, fewer discretization patterns will be aggregated to stabilize the abilities already acquired for these classes, while for classes with large uncertainty, more discretization patterns will be included to allocate additional capacity for adapting to these undeveloped abilities. After aggregation of the mixture of discretization, we employ a contrastive discretization loss that enforces within-class consistency and between-class diversity, contributing to the learning of generalizable and discriminative features after the selective scan. In experiments, we integrate our plug-and-play branch on numerous OCL methods and significant improvements can be consistently observed on multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.655, + 0.435, + 0.669 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.672, + 0.483, + 0.732 + ], + "angle": 0, + "content": "- To enhance the adaptability of existing methods for OCL, we propose S6MOD, a plug-and-play module based on selective state space model and class-conditional mixture of discretization, which can strengthen the base method." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.733, + 0.483, + 0.821 + ], + "angle": 0, + "content": "- We further develop a class-conditional routing strategy that dynamically satisfy both objectives of maintaining the stability of knowledge already acquired and fostering the plasticity for undeveloped abilities. A contrastive discretization loss is employed to facilitate the learning of generalizable and discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.823, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- In experiments, our method can be easily integrated on different OCL methods and is compatible with distillation techniques, to consistently achieve significant improvements on multiple datasets including CIFAR-10, CIFAR-100, and Tiny-ImageNet." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.672, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.907, + 0.314 + ], + "angle": 0, + "content": "Continual Learning(CL). CL can be categorized into three types based on the implementation methods [10, 29, 31, 43]: regularization-based, parameter isolation-based, and replay-based. Regularization-based methods [2, 23, 42, 56] introduce regularization terms to restrict changes in model parameters, thereby protecting the already learned knowledge. Parameter isolation-based methods [36, 38, 47, 54] typically assign different model parameters or subnetworks to different tasks to avoid interference between tasks. Replay-based methods [6, 8, 27, 33, 34] perform joint training by either storing a portion of old data or generating pseudo-data. This allows the model to revisit old data while learning new data, thereby reducing forgetting [53]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.335, + 0.908, + 0.683 + ], + "angle": 0, + "content": "Online Continual Learning(OCL). OCL is a specialized form of CL designed to test the ability to continuously learn and update as data arrives in real-time streams [10, 16, 28]. This means that OCL can only train for a single epoch and typically process only individual samples or mini batches at a time. Replay-based methods are widely used in OCL [17]. ER [34] introduces the combination of cross-entropy loss with a random buffer. OCM [17] uses mutual information maximization to reduce feature bias and preserve past knowledge. GSA [18] proposes a gradient-based adaptive optimization method to address dynamic training bias. On-Pro [46] uses online prototype equilibrium to address shortcut learning problem. Some methods introduce knowledge distillation based on replay. CCL-DC [44] introduces Collaborative Continual Learning (CCL) and Distillation Chain (DC) to enhance model plasticity. MOSE [50] alleviates forgetting by integrating multi-level supervision and reverse self-distillation. However, these models carry a risk when the features used to distill are not precise. In comparison, our method S6MOD as a plug-and-play module is applicable to most OCL methods to improve the adaptability of these models, enabling them to learn generalizable and discriminative features more efficiently." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Selective State Space Model (S6) Selective State Space Model (S6) [13] has gained increasing interest as an alternative to self-attention [41] with lower computational complexity. S6 enhances the S4 model [14] by introducing the selective scan mechanism, and its effectiveness in vision tasks has been extensively studied [26, 58]. For example, Vmamba [26] introduces SS2D, a cross-scanning mechanism for images, facilitating the extension of Mamba to process vision data. A recent study Mamba-FSCIL [22] leverages the dynamic weights and sequence modeling capability of Mamba to achieve dynamic adaptation in few-shot class-incremental learning. But their method relies on an effective selective scan learned from training data. In OCL," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20503" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "the online data that appears only once for training provides limited context for Mamba models, and thus poses challenges to capture a precise discretization pattern. Different from these studies, our method integrates S6 with class-conditional mixture of discretization and effectively helps to improve the adaptability for OCL. Some existing works have introduced mixture of experts (MoE) into Mamba [4, 24, 32], following the design of switch-transformer [12]. Our method differs from them in that each expert in our mixture of discretization is only a simple linear layer for computation efficiency, and discretization patterns are aggregated with our class-conditional gating to balance between maintaining stability of old knowledge and allocating capacity for learning new abilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.315, + 0.232, + 0.33 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.34, + 0.338, + 0.355 + ], + "angle": 0, + "content": "3.1. Problem Definition of OCL" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.485, + 0.562 + ], + "angle": 0, + "content": "OCL requires a model to continuously update itself from a data stream, with each mini-batch containing new data sampled from a changing distribution \\( D_{t} \\). At each time step \\( t \\), the model receives a mini-batch of data \\( \\{(s_{i}^{(t)}, r_{i}^{(t)})\\}_{i=1}^{n_{t}} \\) where \\( s_{i}^{(t)} \\) represents the input data, \\( r_{i}^{(t)} \\) represents the corresponding labels, and \\( n_{t} \\) is the number of samples. The goal of OCL is to sequentially update the model to adapt to both gradual and task-specific shifts in the data distribution. Given a pre-trained network \\( f_{\\theta} \\) with parameters \\( \\theta \\), the model updates its parameters by solving an optimization problem \\( \\arg \\min_{\\theta} L(f_{\\theta}(s), r) \\), where \\( L \\) is the loss function, thereby enabling it to adapt effectively to the evolving data distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.571, + 0.277, + 0.587 + ], + "angle": 0, + "content": "3.2. State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.669 + ], + "angle": 0, + "content": "State Space Models (SSMs) can be viewed as linear time-invariant systems. They map input sequences \\( x(t) \\in \\mathbb{R} \\) to output sequences \\( y(t) \\in \\mathbb{R} \\) through hidden states \\( h(t) \\in \\mathbb{R}^N \\). Mathematically, these models can be represented as linear ordinary differential equations (ODE):" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.675, + 0.482, + 0.699 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = A h (t) + B x (t), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.696, + 0.309, + 0.712 + ], + "angle": 0, + "content": "\\[\ny (t) = C h (t),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "where \\(A \\in \\mathbb{R}^{N \\times N}\\) represents the state transition matrix, while \\(B \\in \\mathbb{R}^{N \\times 1}\\) and \\(C \\in \\mathbb{R}^{1 \\times N}\\) denote the mapping matrices from input to latent state and from latent state to output, respectively, where \\(N\\) indicates the size of the hidden state." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.857 + ], + "angle": 0, + "content": "To apply the SSM to real discrete data, zero-order hold (ZOH) [14] is employed. It discretizes the continuous parameters \\(A\\), \\(B\\), and \\(C\\) using the time scale parameter \\(\\Delta\\). The discretized SSM equations can be rewritten as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.867, + 0.483, + 0.892 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\bar {A} h _ {t - 1} + \\bar {B} x _ {t}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.889, + 0.292, + 0.904 + ], + "angle": 0, + "content": "\\[\ny _ {t} = C h _ {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "Recently, Gu [13] proposed a new parameterization method for SSM with a selective scan mechanism, which is known as S6 and serves as the core of the Mamba model. To enhance S6's capability in processing visual data, Liu et al. [26] introduced SS2D, which can serialize image data from four different directions. Given the input data \\( x \\), the output \\( \\overline{x} \\) processed by SS2D can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.205, + 0.906, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\bar {x} = \\operatorname {S S 2 D} (x) = \\sum_ {i = 1} ^ {4} \\operatorname {S 6} (\\operatorname {s c a n} (x, i)). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.253, + 0.688, + 0.27 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.385 + ], + "angle": 0, + "content": "In this section, we will first introduce the overall structure of our plug-and-play module S6MOD in Sec. 4.1. After that, we will present the detailed design of the state space model with mixture of discretization in Sec. 4.2. Then, we will introduce the class-conditional routing and its corresponding contrastive discretization loss in details in Sec. 4.3. Finally, we will specify the optimization details in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.391, + 0.684, + 0.405 + ], + "angle": 0, + "content": "4.1. Overall structure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.906, + 0.595 + ], + "angle": 0, + "content": "To better enhance existing methods' ability in learning generalizable and discriminative features incrementally from online training data, we propose a plug-and-play module named S6MOD that can be easily applied to existing OCL methods, as shown in Fig. 1 (a). Overall, S6MOD strengthens the original method by introducing a plug-and-play branch after the backbone. It consists of a block and a fixed equiangular tight frame (ETF) classifier [51, 52]. The intermediate features \\(\\mathbf{F}\\) generated by the backbone are duplicated and sent into the classification head of the original base method, and the selective state space model (SSM) branch introduced by our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.7 + ], + "angle": 0, + "content": "In our SSM branch, \\(\\mathbf{F}\\) is projected into two paths through an MLP. The first one \\(\\mathbf{X} = f_{x}(\\mathbf{F})\\) is used to perform selective scan with our class-conditional mixture of discretization, as will be detailed in later subsections. The other one \\(\\mathbf{Z} = f_{z}(\\mathbf{F})\\) performs a gating mechanism as commonly adopted in Mamba models [22, 26, 30]. The output feature of this branch can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.708, + 0.905, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} = \\operatorname {S i L U} (\\mathbf {Z}) \\otimes \\operatorname {S 6 M O D} (\\operatorname {S i L U} (\\operatorname {C o n v} (\\mathbf {X}))). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.807 + ], + "angle": 0, + "content": "To assist the learning of the base method for generalizable and discriminative features, we adopt a regularization, \\(\\mathcal{L}_{\\mathrm{Diff}}\\), on the predicted distribution of the base method, which is the KL divergence between \\(P\\) and \\(Q\\) and can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.812, + 0.905, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {D i f f}} = \\sum_ {i} P (i) \\log \\left(\\frac {P (i)}{Q (i)}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "where \\(P\\) represents the predicted distribution of the base method, and \\(Q\\) refers to the output prediction after the ETF classifier of our SSM branch, i.e., \\(Q = \\mathrm{softmax}(\\mathbf{W}_{\\mathrm{ETF}}\\boldsymbol {\\mu})\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20504" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.153, + 0.09, + 0.845, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.226, + 0.842, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.49, + 0.908, + 0.548 + ], + "angle": 0, + "content": "Figure 1. Framework of S6MOD. Our method (a) introduces a plug-and-play branch after the backbone, where features are learned through S6MOD and supervised by the ETF classifier to guide the base method. S6MOD (c) utilizes MoE to enhance the discretization of SSM and applies class-conditional routing (b) to dynamically adjust the discretization based on the uncertainty. Finally, we use a contrastive discretization loss (d) to supervise the learning of both generalizable and discriminative features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.383, + 0.577 + ], + "angle": 0, + "content": "4.2. S6 with Mixture of Discretization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.676 + ], + "angle": 0, + "content": "We integrate the selective space state model with a mixture of discretization, borrowing the concept of mixture of experts [11, 19, 39, 55]. Its structure is shown in Fig. 1 (c). Similar to the typical SSM structure, we also use an MLP layer to produce the projection matrices \\( B \\) and \\( C \\) from the input features, which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.689, + 0.483, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} = f _ {B} (\\hat {\\mathbf {X}}), \\quad \\mathbf {C} = f _ {C} (\\hat {\\mathbf {X}}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.873 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}\\) denotes the input of the selective scan module S6MOD, i.e., \\(\\hat{\\mathbf{X}} = \\mathrm{SiLU}(\\mathrm{Conv}(\\mathbf{X}))\\). To enable selective SSM to capture a precise discretization pattern with limited context from online data, we develop a sparse MoE system for the discretization transformation \\(\\pmb{\\Delta}\\) in S6 models, utilizing specialized projection layers to enrich the discretization patterns and adaptively selecting the most sensitive discretization based on the current dynamics. Specifically, each discretization candidate is produced by a projection layer as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Delta} _ {i} = f _ {\\Delta i} (\\hat {\\mathbf {X}}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.685 + ], + "angle": 0, + "content": "The input feature \\(\\mathbf{X}\\) is also fed into a sparse gating mechanism, which produces associated importance weights \\(w_{i}\\) corresponding to each discretization \\(\\pmb{\\Delta}_{i}\\). Through the class-conditional gating that will be introduced in Sec. 4.3, we dynamically control the number of discretization patterns, \\(N_{k}\\), to be selected for each class \\(k\\) based on the input features \\(\\hat{\\mathbf{X}}\\). Finally, the selected discretization patterns are aggregated with their importance weights as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.689, + 0.907, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Delta} = \\sum_ {i \\in \\Omega (\\hat {\\mathbf {X}})} w _ {i} \\cdot \\boldsymbol {\\Delta} _ {i}, \\quad | \\Omega (\\hat {\\mathbf {X}}) | = N _ {k}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.906, + 0.765 + ], + "angle": 0, + "content": "where \\(\\Omega (\\hat{\\mathbf{X}})\\) is the index set of the top- \\(N_{k}\\) discretization candidates of \\(\\hat{\\mathbf{X}}\\), according to their importance weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "It's noteworthy that \\(\\Delta\\) is important because it controls the decay rate during the state update process, enabling the model to flexibly select and retain important information when handling long sequences, while avoiding the accumulation of redundant and irrelevant information [13], especially in cases of rapid changing states and evolving systems, as exemplified by OCL. Our method with mixture of discretization facilitates the learning of a precise selective scan pattern with limited data context in OCL." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20505" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.332, + 0.108 + ], + "angle": 0, + "content": "4.3. Class-Conditional Routing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.339 + ], + "angle": 0, + "content": "With more discretization patterns selected, the model is able to allocate more capacity for learning new knowledge, but may cause a significant shift of the model that impairs the existing ability. If we can calculate \\( N_{k} \\) based on the misclassification probability of each class \\( k \\), we can dynamically control the capacity of mixture of discretization based on the learning conditions of all classes. Therefore, in order to guide the mixture of discretization to strike a balance between maintaining the stability of old knowledge and fostering the plasticity for new ability, we further design class-conditional routing. During training, we maintain the feature prototypes \\( \\mathbf{M} = \\{\\mathbf{M}_c\\} \\) by moving average, where \\( \\mathbf{M}_c \\) is the within-class feature mean for class \\( c \\). Then, we estimate the class uncertainty \\( \\sigma_{k} \\) based on the average margin of class \\( k \\) with different classes," + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.351, + 0.482, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {k} = \\operatorname {A v g} _ {c} \\left\\{\\exp \\left(- \\lambda_ {0} \\| \\mathbf {M} _ {k} - \\mathbf {M} _ {c} \\| _ {2}\\right) \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.514 + ], + "angle": 0, + "content": "where \\(\\lambda_0\\) is a hyper-parameter and \\(\\sigma_{k}\\) denotes the class uncertainty for class \\(k\\) ranging from \\((0,1)\\). A large \\(\\sigma_{k}\\) indicates that the class center \\(k\\) has narrow margins with the other class centers, and thus tends to be misclassified, while a small \\(\\sigma_{k}\\) happens when class center \\(k\\) is distant from the other class centers with a lower likelihood of being misclassified. In inference, we replace \\(\\mathbf{M}_k\\) with the input feature \\(\\hat{\\mathbf{X}}\\) and calculate its uncertainty by Eq. (9) with the feature prototypes \\(\\mathbf{M}_c\\) of all classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Then, we multiply the uncertainty by the number of total discretization candidates \\(N\\), to get the discretization number to select for class \\(k\\) or input feature \\(\\hat{\\mathbf{X}}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.572, + 0.482, + 0.587 + ], + "angle": 0, + "content": "\\[\nN _ {k} = \\operatorname {c e i l} \\left(N \\cdot \\sigma_ {k}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.629 + ], + "angle": 0, + "content": "where \\( \\text{ceil} \\) refers to the operation that rounds up \\( N \\cdot \\sigma_k \\) to the nearest integer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.84 + ], + "angle": 0, + "content": "The institution is that classes that are more prone to misclassification have smaller margins with the other classes in the feature space, and thus needs to aggregate more discretization patterns to allocate additional capacity for adapting to these undeveloped abilities. Conversely, for categories that are easier to classify with low uncertainty, a smaller \\( N_{k} \\) is favored to only include the most likely discretization patterns, which can spare optimization efforts for the uncertain classes and stabilize these already acquired abilities. The combination of class-conditional routing and our mixture of discretization achieves dynamic structures dependent on input features and classes, such that the final prediction of each class selectively optimizes the corresponding discretization patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To further reduce interaction among different classes and facilitate the learning of generalizable and discriminative features, we also incorporate contrastive discretization loss function \\(\\mathcal{L}_{\\mathrm{Cont}}\\), as shown in Fig. 1 (d). It encourages \\(\\pmb{\\Delta}\\) to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "have within-class consistency and between-class diversity as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.131, + 0.905, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {C o n t}} = - \\frac {1}{B ^ {2}} \\sum_ {m = 1} ^ {B} \\sum_ {n = 1} ^ {B} \\left(1 _ {y _ {m} = y _ {n}} - 1 _ {y _ {m} \\neq y _ {n}}\\right) \\frac {\\boldsymbol {\\Delta} _ {m} \\cdot \\boldsymbol {\\Delta} _ {n}}{\\| \\boldsymbol {\\Delta} _ {m} \\| \\| \\boldsymbol {\\Delta} _ {n} \\|} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.907, + 0.247 + ], + "angle": 0, + "content": "where \\(B\\) is the batch size, \\(\\Delta_{m}\\) represents the aggregated \\(\\Delta\\) by Eq. (8) for input \\(m\\), \\(y_{m}\\) denotes the class of \\(m\\) and \\(1_{y_m = y_n}\\) is an indicator function that takes a value of 1 when the condition \\(y_{m} = y_{n}\\) holds, and 0 otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.256, + 0.653, + 0.272 + ], + "angle": 0, + "content": "4.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.907, + 0.4 + ], + "angle": 0, + "content": "In addition to the previously mentioned loss functions in Eqs. (5) and (11), we also directly used the classification loss function \\(\\mathcal{L}_{\\mathrm{DR}}\\) specifically designed for ETF classifier [51, 52] to supervise our introduced plug-and-play branch, and the loss function \\(\\mathcal{L}_{\\mathrm{z}}\\) for maintaining load balancing among the discretization patterns [59]. In summary, when integrating S6MOD on a base method, the overall loss function can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.413, + 0.905, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {S 6 M O D}}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.457, + 0.905, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S 6 M O D}} = \\mathcal {L} _ {\\mathrm {D R}} + \\alpha \\cdot \\mathcal {L} _ {\\text {D i f f}} + \\beta \\cdot \\mathcal {L} _ {\\text {C o n t}} + \\mathcal {L} _ {\\mathrm {z}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.481, + 0.756, + 0.496 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.646, + 0.527 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.712, + 0.552 + ], + "angle": 0, + "content": "5.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.695 + ], + "angle": 0, + "content": "Datasets. We test three datasets that are widely used in OCL, i.e., CIFAR-10 (10 classes) [20], CIFAR-100 (100 classes) [20] and TinyImageNet (200 classes) [21]. The dataset settings we adopted are the same as those in CCLDC[46]. We divide CIFAR-10 into 5 tasks, with 2 classes per task; CIFAR-100 into 10 tasks, with 10 classes per task; and TinyImageNet into 100 tasks, with 2 classes per task. Further details about the datasets will be provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.79 + ], + "angle": 0, + "content": "Baselines. To demonstrate the effectiveness and applicability of our module, we conduct tests on 7 typical and state-of-the-art methods, including ER [34], OCM [17], OnPro [46], OCM-CCLDC [44], OnPro-CCLDC [44], MOSE[50] and MOE-MOSE [50]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Implementation details. To ensure a fair comparison, we applied the same hyperparameter settings to each baseline and the methods combined with our module (S6MOD). All the above methods use ResNet-18 as the backbone, without pretraining. For the streaming input data, we set the batch size to 10, and for the samples drawn from the buffer," + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20506" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Average Accuracy (\\%, higher is better) on three benchmark datasets with difference memory buffer size \\( M \\), with and without our proposed S6MOD module. All values are averages of 10 runs." + }, + { + "type": "table", + "bbox": [ + 0.177, + 0.124, + 0.821, + 0.357 + ], + "angle": 0, + "content": "
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]56.68±1.8962.32±4.1324.47±0.7231.89±1.4539.41±1.8110.82±0.7919.16±1.4224.71±2.52
ER + Ours57.88±3.3065.80±2.1626.50±2.2334.55±1.6639.61±3.1610.94±1.4719.67±1.3625.62±1.73
OCM [17]68.19±1.7573.15±1.0528.02±0.7435.69±1.3642.22±1.0618.36±0.9526.74±1.0231.94±1.19
OCM + Ours70.57±1.1475.31±1.1032.44±1.3538.97±2.2845.49±1.5819.12±1.6027.20±1.8332.31±1.72
OnPro [46]70.47±2.1274.70±1.5127.22±0.7733.33±0.9341.59±1.3814.32±1.4021.13±2.1226.38±2.18
OnPro + Ours72.30±1.0875.14±0.9928.84±0.5937.57±0.8144.14±1.3617.27±0.6525.62±0.7929.20±1.00
OCM-CCLDC [44]74.14±0.8577.66±1.4635.00±1.1543.34±1.5151.43±1.3723.36±1.1833.17±0.9739.25±0.88
OCM-CCLDC + Ours74.45±1.2078.21±1.0336.02±1.7744.40±2.2652.53±0.2123.68±1.0233.59±0.7239.53±1.02
OnPro-CCLDC [44]74.49±2.1478.64±1.4234.76±1.1241.89±0.8250.01±0.8521.81±1.0232.00±0.7238.18±1.02
OnPro-CCLDC + Ours75.03±1.0379.51±0.6336.07±0.7643.92±0.9650.85±0.5222.03±0.8233.29±0.3538.51±0.91
MOSE [50]61.02±1.4770.74±1.1835.05±0.3445.06±0.3254.53±0.7818.23±0.7330.98±0.6338.71±0.44
MOSE + Ours63.74±1.5572.54±0.1435.43±0.6245.38±0.3154.75±0.5219.03±0.8332.03±0.8138.18±1.02
MOE-MOSE [50]62.54±1.5972.18±1.2937.32±0.3447.03±0.5755.62±0.7220.61±0.6932.52±0.3338.41±0.53
MOE-MOSE + Ours63.86±1.1973.16±0.5337.76±0.5147.25±0.4356.32±0.5321.03±0.7733.53±0.5140.11±0.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.365, + 0.908, + 0.394 + ], + "angle": 0, + "content": "Table 2. Average Forgetting (%, lower is better) on three benchmark datasets with difference memory buffer size \\( M \\), with and without our proposed S6MOD module. All values are averages of 10 runs." + }, + { + "type": "table", + "bbox": [ + 0.173, + 0.399, + 0.818, + 0.631 + ], + "angle": 0, + "content": "
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]33.16±3.5020.94±6.7932.65±1.7822.20±2.2613.29±1.9858.38±1.6946.87±1.6040.77±2.45
ER + Ours31.25±3.4416.71±2.8630.96±2.0719.02±2.1312.97±2.6158.03±2.1945.83±1.4937.62±1.53
OCM [17]13.68±4.2511.63±2.6214.99±1.559.16±1.753.76±1.1626.12±1.6319.74±1.3015.92±1.47
OCM + Ours13.58±3.269.68±3.4715.99±1.8211.60±1.495.46±0.9824.23±2.0719.12±1.8714.66±1.75
OnPro [46]17.94±3.6914.20±2.6016.76±2.4712.42±1.396.72±0.9428.01±1.5923.52±1.7520.32±1.70
OnPro + Ours7.13±1.445.73±1.8116.03±1.699.81±1.325.07±0.8920.92±1.1216.82±0.9718.15±1.91
OCM-CCLDC [44]11.59±2.249.18±2.0316.69±2.3610.07±1.373.99±0.7826.16±1.9019.99±1.9615.56±1.06
OCM-CCLDC + Ours15.81±2.6911.28±1.8316.64±1.859.08±2.063.48±0.3125.29±1.9514.89±0.7411.55±1.31
OnPro-CCLDC [44]19.89±4.0114.62±2.7528.93±2.1920.23±1.0310.55±1.8928.21±1.5820.85±1.1316.17±0.63
OnPro-CCLDC + Ours17.23±3.1611.51±2.3322.26±1.1812.49±1.594.88±1.3626.73±1.5116.69±0.3112.23±1.37
MOSE [50]30.36±1.6920.27±1.2737.54±0.4325.89±0.4513.60±0.5947.16±1.4124.96±0.6215.51±0.33
MOSE + Ours27.38±1.9417.92±0.3437.09±0.6325.80±0.3915.55±0.6146.60±1.0724.39±0.7220.15±0.74
MOE-MOSE [50]29.39±1.7919.24±1.4935.17±0.3023.99±0.5112.81±0.7441.98±1.4622.22±0.3813.94±0.59
MOE-MOSE + Ours28.13±1.4117.98±0.5634.76±0.6623.84±0.7714.03±0.4841.26±1.0221.39±0.5814.91±0.28
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.705 + ], + "angle": 0, + "content": "the batch size is set to 64. We retain each baseline's original data augmentation methods without modifying the base methods. For more details, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.185, + 0.728 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We combine our method with both classical and state-of-the-art approaches on CIFAR-10, CIFAR-100, and TinyImageNet. The experimental results in Table 1 demonstrate the universality of our method, as it consistently improves accuracy across numerous baselines. On CIFAR-10 and CIFAR-100, our method generally leads to a \\(1\\%\\) improvement, while on TinyImageNet, we achieve significant gains under OCM, OnPro and MOE-MOSE. Notably, in settings like CIFAR100 \\((M = 2k)\\) and TinyImageNet \\((M = 5k)\\), OnPro result in an approximate \\(4.2\\%\\) and \\(4.5\\%\\) improvement, respectively. It is also worth mention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.749 + ], + "angle": 0, + "content": "ing that even for state-of-the-art distillation-based methods like CCLDC and MOE-MOSE, the performance improvements are quite substantial. For instance, OnPro-CCLDC with the S6MOD integration achieve an impressive \\(79.51\\%\\) on CIFAR10 (\\(M = 1k\\)), and MOE-MOSE with S6MOD reach \\(56.32\\%\\) on CIFAR100 (\\(M = 5k\\)) and \\(40.1\\%\\) on TinyImageNet (\\(M = 10k\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In addition, as shown in Table 2, our module is also highly effective in reducing model forgetting, achieving lower forgetting rates in most settings, with some cases showing particularly significant improvements. This can be attributed to our class-conditional routing. In a few baseline settings, the inclusion of our module led to a higher forgetting rate, but this does not imply that the model is more prone to forgetting. In these cases, we think the higher forgetting rates are often due to the stronger learning capabilities of the model. Analyzing both Table 1 and Table 2, we" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20507" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.16 + ], + "angle": 0, + "content": "Table 3. Ablation studies on CIFAR-100 (\\(M = 2k\\)). The \"branch\" refers to using SS2D[26] as the core of the S6MOD, without routing and \\(\\mathcal{L}_{\\mathrm{Cont}}\\). \"routing\" represents class-conditional routing. S6MOD refers to the composition of (branch + class-conditional routing + \\(\\mathcal{L}_{\\mathrm{Cont}}\\)). All values are averaged over 10 runs." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.164, + 0.435, + 0.254 + ], + "angle": 0, + "content": "
MethodAcc. ↑AF ↓
OnPro33.33±0.9312.42±1.39
OnPro + branch (with LDiff)36.61±1.0410.28±1.59
OnPro + branch + routing36.92±0.728.85±0.81
OnPro + branch + LCont36.93±0.849.70±1.42
OnPro + S6MOD37.57±0.819.81±1.32
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.485, + 0.292 + ], + "angle": 0, + "content": "can find that even in these settings with higher forgetting rates, the accuracy performance remains quite competitive." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.3, + 0.258, + 0.317 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.485, + 0.458 + ], + "angle": 0, + "content": "As mentioned in Sec. 1, the S6MOD module we propose is very simple and can be directly applied to existing baselines. By adding an extra branch and using the \\(\\mathcal{L}_{\\mathrm{Diff}}\\) to supervise the original classification head, we can achieve a significant improvement. To demonstrate this, we conduct an ablation study where we only add the extra branch and \\(\\mathcal{L}_{\\mathrm{Diff}}\\) to the baseline. The results in Table 3 confirm this, showing a \\(3.3\\%\\) improvement with supervision from the extra branch alone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.673 + ], + "angle": 0, + "content": "In addition, to verify the effectiveness and adaptability of class-conditional routing and contrastive discretization loss \\(\\mathcal{L}_{\\mathrm{Cont}}\\), we conduct separate ablation experiments. As shown in Table 3, when added to the base method with the extra branch, the improvements brought by adding class-conditional routing or \\(\\mathcal{L}_{\\mathrm{Cont}}\\) alone are not significant, with accuracy gains of only \\(0.31\\%\\) and \\(0.32\\%\\), respectively. However, when both components are used together, the improvement becomes significant, with an approximate \\(1\\%\\) increase in accuracy, and the forgetting rate still decreases. This indicates that when our module is correctly combined, it can significantly enhances the model's adaptability by learning more generalizable and discriminative features through dynamic parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.196, + 0.699 + ], + "angle": 0, + "content": "5.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Analysis of Feature Embeddings. To verify that S6MOD improves the adaptability of the baseline and helps it learn more discriminative features, we use t-SNE [40] visualization for analysis. We compare with the baseline under the CIFAR-100 (\\(M = 2k\\)) setting. As shown in Fig. 2, we visualize the features of the samples in the buffer after the final training, which are the features \\(\\mathbf{F}\\) output by the backbone just before they are passed to the classifier for classification. The features shown in Fig. 2 (a) and (c) are very scattered, with only a few classes showing some degree of aggregation. Most classes do not have clear separations between other classes, making it difficult to distinguish. In contrast, Fig. 2 (b) and (d) present the feature distribution" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.091, + 0.707, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.211, + 0.657, + 0.222 + ], + "angle": 0, + "content": "(a) OnPro" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.092, + 0.868, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.21, + 0.849, + 0.222 + ], + "angle": 0, + "content": "(b) OnPro + S6MOD" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.226, + 0.707, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.346, + 0.655, + 0.359 + ], + "angle": 0, + "content": "(c) OCM" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.226, + 0.869, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.346, + 0.847, + 0.358 + ], + "angle": 0, + "content": "(d) OCM + S6MOD" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.371, + 0.908, + 0.426 + ], + "angle": 0, + "content": "Figure 2. t-SNE visualization of memory data at the end of training on CIFAR-100 (\\(M = 2k\\)), showcasing baseline in (a) and (c) and baseline combined with S6MOD in (b) and (d). Different colors represent different classes." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.427, + 0.691, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.565, + 0.677, + 0.577 + ], + "angle": 0, + "content": "(a) New-Task Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.435, + 0.884, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.565, + 0.868, + 0.577 + ], + "angle": 0, + "content": "(b) Average Forgetting" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.581, + 0.908, + 0.664 + ], + "angle": 0, + "content": "Figure 3. Impact of dynamically selecting different \\( N_{k} \\) values on the ability to learn new tasks and prevent forgetting of old tasks: New-Task accuracy represents the model's accuracy on the current task. We conduct experiments by setting \\( N_{k} = 1 \\), \\( N_{k} = N \\), and calculating \\( N_{k} \\) through class-conditional routing. The dataset used is CIFAR-100 (\\( M = 2k \\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.78 + ], + "angle": 0, + "content": "after adding our module. We can observe that many categories have noticeably converged, forming small clusters with consistent intra-class compactness and clear inter-class separations. These features are much easier to distinguish and more conducive to classification by the classifier, which aligns with the significant improvements we demonstrate in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis of Class-conditional Routing. We further investigate the effectiveness of class-conditional routing. Under the CIFAR-100 (\\(M = 2k\\)) setting, we design experiments to compare the impact of a fixed \\(N_{k}\\) and dynamically calculated \\(N_{k}\\) using class-conditional routing. Specifically, we set \\(N_{k} = N\\), meaning all patterns are activated, and \\(N_{k} = 1\\), meaning only the pattern with the highest prob" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20508" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.146 + ], + "angle": 0, + "content": "Table 4. Impact of hyper-parameters \\(\\alpha\\) and \\(\\beta\\) on performance in the CIFAR-100 dataset (\\(M = 2k\\)). \\(\\alpha\\) and \\(\\beta\\) are the weights for \\(\\mathcal{L}_{\\mathrm{Diff}}\\) and \\(\\mathcal{L}_{\\mathrm{Cont}}\\), respectively. Acc. and AF represent the average accuracy and the average forgetting over 5 runs, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.152, + 0.481, + 0.297 + ], + "angle": 0, + "content": "
Impact of hyper-parameter α
Value0.010.10.515
Acc.36.80±1.1737.09±1.0937.55±1.0837.47±0.9437.57±0.81
AF10.49±1.019.65±0.399.05±1.259.16±2.119.81±1.32
Impact of hyper-parameter β
Value0.1151025
Acc.36.50±1.2437.30±1.0337.40±0.3337.26±0.6137.57±0.81
AF10.24±0.4310.12±1.0910.10±0.609.98±0.669.81±1.32
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.686 + ], + "angle": 0, + "content": "ability is activated. To facilitate analysis, we present the accuracy of the model on the latest task (the current task) in Fig. 3 (a). As shown in Fig. 3 (a), when \\( N_{k} = N \\), all patterns are activated, and thanks to the inclusion of more patterns with different weights and larger parameters, the model has greater capacity to adapt to new tasks, resulting in a stronger ability to learn new knowledge. When \\( N_{k} = 1 \\), only the pattern with the highest probability is activated, significantly reducing the model's ability to learn the new task. However, as shown in Fig. 3 (b), when \\( N_{k} = N \\), since all patterns are updated by the current task, the model has a higher risk of forgetting. When \\( N_{k} = 1 \\), only the pattern with the highest probability is updated, and tasks do not interfere with each other, so forgetting is less likely. Combining the analysis of the entire Fig. 3, the effect of our proposed class-conditional routing is both intuitive and significant. It not only helps the model achieve strong learning ability for new tasks, comparable to the \\( N_{k} \\) setting, but also provides excellent anti-forgetting performance, similar to \\( N_{k} = 1 \\). This is due to the dynamic adjustment of \\( N_{k} \\) based on uncertainty: for samples that are prone to misclassification, class-conditional routing assigns a larger \\( N_{k} \\), enabling more extensive learning. For easier-to-classify samples, the model's learning ability is higher, and class-conditional routing assigns a smaller \\( N_{k} \\) to enhance anti-forgetting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Sensitivity Analysis of Hyper-parameters \\(\\alpha\\) and \\(\\beta\\). \\(\\alpha\\) and \\(\\beta\\) are the weights for \\(\\mathcal{L}_{\\mathrm{Diff}}\\) and \\(\\mathcal{L}_{\\mathrm{Cont}}\\), respectively. We conduct experiments on the sensitivity of them, and the results are shown in Table 4. It can be observed that when the weights are low, the optimization effect of the loss is relatively weak, both accuracy and forgetting rate do not perform well. Gradually increasing the weights, we find that after a certain point, the impact of the weights on the results becomes relatively stable, with no more drastic changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Sensitivity Analysis of total discretization patterns number \\(N\\). We conduct experiments on the total number of discretization patterns in the module. The results are shown" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.101, + 0.693, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.1, + 0.887, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.217, + 0.905, + 0.243 + ], + "angle": 0, + "content": "(a) New-Task Average Accuracy (b) Average Forgetting w/o class-w/o class-conditional routing conditional routing" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.251, + 0.695, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.368, + 0.7, + 0.38 + ], + "angle": 0, + "content": "(c) New-Task Average Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.251, + 0.888, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.368, + 0.868, + 0.38 + ], + "angle": 0, + "content": "(d) Average Forgetting" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.386, + 0.905, + 0.414 + ], + "angle": 0, + "content": "Figure 4. Impact of pattern number \\( N \\). The dataset used is CIFAR-100 (\\( M = 2k \\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.562 + ], + "angle": 0, + "content": "in Fig. 4. As the number \\( N \\) increases, the model gains a stronger ability to learn new tasks, but at the same time, its tendency to forget becomes more pronounced. Furthermore, by comparing (a)(b) and (c)(d) in Fig. 4, we can conclude that class-conditional routing can significantly reduce forgetting with only a slight decrease in new task accuracy, which in turn leads to an overall improvement in performance. This indirectly supports the effectiveness of class-conditional routing in our ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.574, + 0.634, + 0.59 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.766 + ], + "angle": 0, + "content": "In this paper, we propose a plug-and-play module S6MOD to enhance the adaptability of existing OCL methods by introducing selective state space models (SSMs) with a class-conditional mixture of discretization. By integrating a dynamic discretization mechanism and leveraging class-conditional strategies, our method can efficiently allocate discretization patterns based on class uncertainty, improving both the model's generalization and its ability to adapt to new data. Experimental results on multiple OCL datasets demonstrate that our method consistently optimize existing techniques, contributing to more robust OCL systems." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.78, + 0.693, + 0.796 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.907, + 0.888 + ], + "angle": 0, + "content": "The research reported in this publication was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940. This research used computational resources provided by IBEX managed by the Supercomputing Core Laboratory at KAUST." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20509" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Assyr Abdulle, E Weinan, Björn Engquist, and Eric Vanden-Eijnden. The heterogeneous multiscale method. Acta Numerica, 21:1-87, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.16, + 0.482, + 0.214 + ], + "angle": 0, + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. NeurIPS, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[4] Quentin Anthony, Yury Tokpanov, Paolo Glorioso, and Beren Millidge. Blackmamba: Mixture of experts for statespace models. arXiv preprint arXiv:2402.01771, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[5] Fengxiang Bie, Yibo Yang, Zhongzhu Zhou, Adam Ghanem, Minjia Zhang, Zhewei Yao, Xiaoxia Wu, Connor Holmes, Pareesa Golnari, David A Clifton, et al. Renaissance: A survey into ai text-to-image generation in the era of large model. TPAMI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. *NeurIPS*, 33: 15920-15930, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.482, + 0.505 + ], + "angle": 0, + "content": "[7] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[9] Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.597, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[10] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 44(7):3366-3385, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.669, + 0.482, + 0.71 + ], + "angle": 0, + "content": "[11] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.714, + 0.482, + 0.767 + ], + "angle": 0, + "content": "[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.771, + 0.482, + 0.812 + ], + "angle": 0, + "content": "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] John Guckenheimer and Philip Holmes. Nonlinear oscillations, dynamical systems, and bifurcations of vector fields. Springer Science & Business Media, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Nuwan Gunasekara, Bernhard Pfahringer, Heitor Murilo Gomes, and Albert Bifet. Survey on online streaming continual learning. In *IJCAI*, pages 6628-6637, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[17] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In ICML, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[18] Yiduo Guo, Bing Liu, and Dongyan Zhao. Dealing with cross-task class discrimination in online continual learning. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[19] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. Neural Computation, 3(1):79-87, 1991. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[21] Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.323, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[22] Xiaojie Li, Yibo Yang, Jianlong Wu, Bernard Ghanem, Liqiang Nie, and Min Zhang. Mamba-fscil: Dynamic adaptation with selective state space model for few-shot class-incremental learning. arXiv preprint arXiv:2407.06136, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.393, + 0.904, + 0.418 + ], + "angle": 0, + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 40(12):2935-2947, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.422, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[24] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[25] Sihao Liu, Mingyi Liu, Tianyu Jiang, Shuang Yu, Hanchuan Xu, and Zhongjie Wang. GSL-mash: Enhancing mashup creation service recommendations through graph structure learning. In Service-Oriented Computing, pages 176-191, Singapore, 2025. Springer Nature Singapore. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. ArXiv, abs/2401.10166, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[27] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[28] Zheda Mai, Ruiwen Li, Hyunwoo Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. In CVPR, pages 3589-3599, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[29] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[30] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[31] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, Michał Krutul, Jakub Krajewski, Szymon Antoniak, Pi" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20510" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "otr Mioś, Marek Cygan, and Sebastian Jaszcur. Moemamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[33] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. *NeurIPS*, 32, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[35] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[36] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[37] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In CVPR, pages 23933-23942, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[38] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, pages 4548-4557. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.46, + 0.566 + ], + "angle": 0, + "content": "[41] A Vaswani. Attention is all you need. NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[42] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. NeurIPS, 34:22379-22391, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[43] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. TPAMI, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[44] Maorong Wang, Nicolas Michel, Ling Xiao, and Toshihiko Yamasaki. Improving plasticity in online continual learning via collaborative learning. In CVPR, pages 23460-23469, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[45] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, pages 18764-18774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[46] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[47] Mitchell Wortsman, Vivek Ramanujan, Rosanne Liu, Aniruddha Kembhavi, Mohammad Rastegari, Jason Yosinski, and Ali Farhadi. Supermasks in superposition. NeurIPS, 33:15173-15184, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[48] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[49] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[50] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In CVPR, pages 23670-23680, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[51] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, pages 37991-38002. Curran Associates, Inc., 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[52] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[53] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. Corda: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In NeurlPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.528 + ], + "angle": 0, + "content": "[55] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[57] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19560, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[58] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[59] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20511" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_origin.pdf b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..072d3a146cee07794c618ddf65ee516992ee7b32 --- /dev/null +++ b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/3a332709-76df-43e2-b1a6-2e585b16d645_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d6fa0ab5da711f7c5588b3c204e75732c417b50433a37a40f600c2003badc34 +size 786107 diff --git a/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/full.md b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9df0da9ef388a622f451cc9be495e267b7979e2c --- /dev/null +++ b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/full.md @@ -0,0 +1,340 @@ +# Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization + +Sihao Liu $^{1}$ Yibo Yang $^{2\dagger}$ Xiaojie Li $^{3}$ David A. Clifton $^{4}$ Bernard Ghanem $^{2}$ + +$^{1}$ Harbin Institute of Technology $^{2}$ King Abdullah University of Science and Technology + +$^{3}$ Harbin Institute of Technology (Shenzhen) $^{4}$ University of Oxford + +# Abstract + +Online continual learning (OCL) seeks to learn new tasks from data streams that appear only once, while retaining knowledge of previously learned tasks. Most existing methods rely on replay, focusing on enhancing memory retention through regularization or distillation. However, they often overlook the adaptability of the model, limiting the ability to learn generalizable and discriminative features incrementally from online training data. To address this, we introduce a plug-and-play module, S6MOD, which can be integrated into most existing methods and directly improve adaptability. Specifically, S6MOD introduces an extra branch after the backbone, where a mixture of discretization selectively adjusts parameters in a selective state space model, enriching selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. We further design a class-conditional routing algorithm for dynamic, uncertainty-based adjustment and implement a contrastive discretization loss to optimize it. Extensive experiments combining our module with various models demonstrate that S6MOD significantly enhances model adaptability, leading to substantial performance gains and achieving the state-of-the-art results. The code is available at https://github.com/MyToumaKazusa/S6MOD. + +# 1. Introduction + +Continual learning (CL) [10, 29, 31, 43] is a task that requires models to efficiently acquire new abilities from incoming data. It tests the model's ability to adapt to dynamically changing environments while retaining existing knowledge. For example, autonomous vehicles are expected to continuously learn new driving environments and traffic regulations, thereby improving detection accuracy under complex driving scenes. From the perspective of task format, CL can be divided into offline and online [10, 43]. + +Unlike offline CL [7, 33], online CL (OCL) [10, 16] aligns with real-world implementations as it requires data to arrive sequentially in mini-batches and allows only one epoch of training, posing greater challenges for efficient adaptation with data accessible for learning only once. + +To tackle the challenging OCL, existing studies widely rely on the replay technique [17] that selectively stores a subset of old-class data to strengthen the memory ability of existing knowledge and mitigate catastrophic forgetting. Building on replay, various methods have been proposed with regularization, buffer allocation, and distillation strategies [17, 18, 34, 44, 46, 50]. For example, MIR [3] aims to mitigate mutual interference among tasks through a retrieval strategy, OCM [17] seeks to reduce forgetting by maximizing mutual information between different tasks, and CCLDC [44] adopts collaborative learning and distillation to improve plasticity. However, learning generalizable and discriminative features incrementally from online training data is still intractable, and imprecise features will impede and even mislead the replay strategies [45]. Therefore, inducing accurate and efficient adaptation with limited data access is critically important for OCL and there still remains significant potential for improvement. + +Recently, selective state space models (SSMs), also known as S6 in Mamba [13], have shown promising results in modeling long-range dependencies with computational efficiency. Selective SSMs introduce a selective scan mechanism to make the interactions among sequential states aware of input context, and have been applied to vision tasks by integrating various scan directions [26, 58]. In addition to the success in sequence modeling [5, 9, 24, 25, 35, 49], selective SSMs also exhibit stronger adaptability than static parameters in few-shot class-incremental learning due to the dynamic operation weights, as reported in a recent pioneering work [22]. Despite the increased capacity of adaptation, directly applying selective SSMs to OCL will be infeasible because it is challenging for selective SSMs to capture a precise discretization pattern with limited context from online data that appears only once for training. Moreover, a single discretization mechanism may fail to capture all the + +essential details in some nonlinear dynamic systems or systems with multi-scale characteristics [1, 15]. + +To this end, we develop a plug-and-play branch based on selective state space model and class-conditional mixture of discretization. The branch with an equiangular tight frame classifier [29, 37, 48, 51, 52, 57] can be integrated on any existing OCL methods to supervise the original classification head and improve the adaptation ability in OCL. Considering multiple discretization methods can exhibit varying sensitivity to the dynamic characteristics of a system, we introduce a mixture of discretization into SSMs to enrich selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. This flexibility allows our method to comprehensively capture the system's complex dynamic features, particularly in cases of rapid state changes and evolving systems as exemplified by OCL. + +In order to guide the mixture of discretization for OCL, where a delicate trade-off between maintaining the stability of old knowledge and fostering the plasticity for new ability is needed, we further introduce a class-conditional routing to aggregate the discretization patterns. Concretely, we maintain a feature prototype for each class to calculate the class uncertainty based on the margin among different classes. For classes with less uncertainty, fewer discretization patterns will be aggregated to stabilize the abilities already acquired for these classes, while for classes with large uncertainty, more discretization patterns will be included to allocate additional capacity for adapting to these undeveloped abilities. After aggregation of the mixture of discretization, we employ a contrastive discretization loss that enforces within-class consistency and between-class diversity, contributing to the learning of generalizable and discriminative features after the selective scan. In experiments, we integrate our plug-and-play branch on numerous OCL methods and significant improvements can be consistently observed on multiple datasets. + +Our contributions can be summarized as follows: + +- To enhance the adaptability of existing methods for OCL, we propose S6MOD, a plug-and-play module based on selective state space model and class-conditional mixture of discretization, which can strengthen the base method. +- We further develop a class-conditional routing strategy that dynamically satisfy both objectives of maintaining the stability of knowledge already acquired and fostering the plasticity for undeveloped abilities. A contrastive discretization loss is employed to facilitate the learning of generalizable and discriminative features. +- In experiments, our method can be easily integrated on different OCL methods and is compatible with distillation techniques, to consistently achieve significant improvements on multiple datasets including CIFAR-10, CIFAR-100, and Tiny-ImageNet. + +# 2. Related Work + +Continual Learning(CL). CL can be categorized into three types based on the implementation methods [10, 29, 31, 43]: regularization-based, parameter isolation-based, and replay-based. Regularization-based methods [2, 23, 42, 56] introduce regularization terms to restrict changes in model parameters, thereby protecting the already learned knowledge. Parameter isolation-based methods [36, 38, 47, 54] typically assign different model parameters or subnetworks to different tasks to avoid interference between tasks. Replay-based methods [6, 8, 27, 33, 34] perform joint training by either storing a portion of old data or generating pseudo-data. This allows the model to revisit old data while learning new data, thereby reducing forgetting [53]. + +Online Continual Learning(OCL). OCL is a specialized form of CL designed to test the ability to continuously learn and update as data arrives in real-time streams [10, 16, 28]. This means that OCL can only train for a single epoch and typically process only individual samples or mini batches at a time. Replay-based methods are widely used in OCL [17]. ER [34] introduces the combination of cross-entropy loss with a random buffer. OCM [17] uses mutual information maximization to reduce feature bias and preserve past knowledge. GSA [18] proposes a gradient-based adaptive optimization method to address dynamic training bias. On-Pro [46] uses online prototype equilibrium to address shortcut learning problem. Some methods introduce knowledge distillation based on replay. CCL-DC [44] introduces Collaborative Continual Learning (CCL) and Distillation Chain (DC) to enhance model plasticity. MOSE [50] alleviates forgetting by integrating multi-level supervision and reverse self-distillation. However, these models carry a risk when the features used to distill are not precise. In comparison, our method S6MOD as a plug-and-play module is applicable to most OCL methods to improve the adaptability of these models, enabling them to learn generalizable and discriminative features more efficiently. + +Selective State Space Model (S6) Selective State Space Model (S6) [13] has gained increasing interest as an alternative to self-attention [41] with lower computational complexity. S6 enhances the S4 model [14] by introducing the selective scan mechanism, and its effectiveness in vision tasks has been extensively studied [26, 58]. For example, Vmamba [26] introduces SS2D, a cross-scanning mechanism for images, facilitating the extension of Mamba to process vision data. A recent study Mamba-FSCIL [22] leverages the dynamic weights and sequence modeling capability of Mamba to achieve dynamic adaptation in few-shot class-incremental learning. But their method relies on an effective selective scan learned from training data. In OCL, + +the online data that appears only once for training provides limited context for Mamba models, and thus poses challenges to capture a precise discretization pattern. Different from these studies, our method integrates S6 with class-conditional mixture of discretization and effectively helps to improve the adaptability for OCL. Some existing works have introduced mixture of experts (MoE) into Mamba [4, 24, 32], following the design of switch-transformer [12]. Our method differs from them in that each expert in our mixture of discretization is only a simple linear layer for computation efficiency, and discretization patterns are aggregated with our class-conditional gating to balance between maintaining stability of old knowledge and allocating capacity for learning new abilities. + +# 3. Preliminaries + +# 3.1. Problem Definition of OCL + +OCL requires a model to continuously update itself from a data stream, with each mini-batch containing new data sampled from a changing distribution $D_{t}$ . At each time step $t$ , the model receives a mini-batch of data $\{(s_{i}^{(t)}, r_{i}^{(t)})\}_{i=1}^{n_{t}}$ where $s_{i}^{(t)}$ represents the input data, $r_{i}^{(t)}$ represents the corresponding labels, and $n_{t}$ is the number of samples. The goal of OCL is to sequentially update the model to adapt to both gradual and task-specific shifts in the data distribution. Given a pre-trained network $f_{\theta}$ with parameters $\theta$ , the model updates its parameters by solving an optimization problem $\arg \min_{\theta} L(f_{\theta}(s), r)$ , where $L$ is the loss function, thereby enabling it to adapt effectively to the evolving data distribution. + +# 3.2. State Space Models + +State Space Models (SSMs) can be viewed as linear time-invariant systems. They map input sequences $x(t) \in \mathbb{R}$ to output sequences $y(t) \in \mathbb{R}$ through hidden states $h(t) \in \mathbb{R}^N$ . Mathematically, these models can be represented as linear ordinary differential equations (ODE): + +$$ +h ^ {\prime} (t) = A h (t) + B x (t), \tag {1} +$$ + +$$ +y (t) = C h (t), +$$ + +where $A \in \mathbb{R}^{N \times N}$ represents the state transition matrix, while $B \in \mathbb{R}^{N \times 1}$ and $C \in \mathbb{R}^{1 \times N}$ denote the mapping matrices from input to latent state and from latent state to output, respectively, where $N$ indicates the size of the hidden state. + +To apply the SSM to real discrete data, zero-order hold (ZOH) [14] is employed. It discretizes the continuous parameters $A$ , $B$ , and $C$ using the time scale parameter $\Delta$ . The discretized SSM equations can be rewritten as follows: + +$$ +h _ {t} = \bar {A} h _ {t - 1} + \bar {B} x _ {t}, \tag {2} +$$ + +$$ +y _ {t} = C h _ {t}. +$$ + +Recently, Gu [13] proposed a new parameterization method for SSM with a selective scan mechanism, which is known as S6 and serves as the core of the Mamba model. To enhance S6's capability in processing visual data, Liu et al. [26] introduced SS2D, which can serialize image data from four different directions. Given the input data $x$ , the output $\overline{x}$ processed by SS2D can be expressed as follows: + +$$ +\bar {x} = \operatorname {S S 2 D} (x) = \sum_ {i = 1} ^ {4} \operatorname {S 6} (\operatorname {s c a n} (x, i)). \tag {3} +$$ + +# 4. Proposed Method + +In this section, we will first introduce the overall structure of our plug-and-play module S6MOD in Sec. 4.1. After that, we will present the detailed design of the state space model with mixture of discretization in Sec. 4.2. Then, we will introduce the class-conditional routing and its corresponding contrastive discretization loss in details in Sec. 4.3. Finally, we will specify the optimization details in Sec. 4.4. + +# 4.1. Overall structure + +To better enhance existing methods' ability in learning generalizable and discriminative features incrementally from online training data, we propose a plug-and-play module named S6MOD that can be easily applied to existing OCL methods, as shown in Fig. 1 (a). Overall, S6MOD strengthens the original method by introducing a plug-and-play branch after the backbone. It consists of a block and a fixed equiangular tight frame (ETF) classifier [51, 52]. The intermediate features $\mathbf{F}$ generated by the backbone are duplicated and sent into the classification head of the original base method, and the selective state space model (SSM) branch introduced by our method. + +In our SSM branch, $\mathbf{F}$ is projected into two paths through an MLP. The first one $\mathbf{X} = f_{x}(\mathbf{F})$ is used to perform selective scan with our class-conditional mixture of discretization, as will be detailed in later subsections. The other one $\mathbf{Z} = f_{z}(\mathbf{F})$ performs a gating mechanism as commonly adopted in Mamba models [22, 26, 30]. The output feature of this branch can be formulated as: + +$$ +\boldsymbol {\mu} = \operatorname {S i L U} (\mathbf {Z}) \otimes \operatorname {S 6 M O D} (\operatorname {S i L U} (\operatorname {C o n v} (\mathbf {X}))). \tag {4} +$$ + +To assist the learning of the base method for generalizable and discriminative features, we adopt a regularization, $\mathcal{L}_{\mathrm{Diff}}$ , on the predicted distribution of the base method, which is the KL divergence between $P$ and $Q$ and can be formulated as: + +$$ +\mathcal {L} _ {\text {D i f f}} = \sum_ {i} P (i) \log \left(\frac {P (i)}{Q (i)}\right), \tag {5} +$$ + +where $P$ represents the predicted distribution of the base method, and $Q$ refers to the output prediction after the ETF classifier of our SSM branch, i.e., $Q = \mathrm{softmax}(\mathbf{W}_{\mathrm{ETF}}\boldsymbol {\mu})$ + +![](images/a8338a9ed7bc850ad5ab28a83ba3eee2816c554b458ff4a891119dafa92d612f.jpg) + +![](images/245cf302e4e89c134e247a787c1a5df0a1d38b2d97dc794fa850c28b88415bce.jpg) +Figure 1. Framework of S6MOD. Our method (a) introduces a plug-and-play branch after the backbone, where features are learned through S6MOD and supervised by the ETF classifier to guide the base method. S6MOD (c) utilizes MoE to enhance the discretization of SSM and applies class-conditional routing (b) to dynamically adjust the discretization based on the uncertainty. Finally, we use a contrastive discretization loss (d) to supervise the learning of both generalizable and discriminative features. + +# 4.2. S6 with Mixture of Discretization + +We integrate the selective space state model with a mixture of discretization, borrowing the concept of mixture of experts [11, 19, 39, 55]. Its structure is shown in Fig. 1 (c). Similar to the typical SSM structure, we also use an MLP layer to produce the projection matrices $B$ and $C$ from the input features, which can be expressed as follows: + +$$ +\mathbf {B} = f _ {B} (\hat {\mathbf {X}}), \quad \mathbf {C} = f _ {C} (\hat {\mathbf {X}}), \tag {6} +$$ + +where $\hat{\mathbf{X}}$ denotes the input of the selective scan module S6MOD, i.e., $\hat{\mathbf{X}} = \mathrm{SiLU}(\mathrm{Conv}(\mathbf{X}))$ . To enable selective SSM to capture a precise discretization pattern with limited context from online data, we develop a sparse MoE system for the discretization transformation $\pmb{\Delta}$ in S6 models, utilizing specialized projection layers to enrich the discretization patterns and adaptively selecting the most sensitive discretization based on the current dynamics. Specifically, each discretization candidate is produced by a projection layer as follows: + +$$ +\boldsymbol {\Delta} _ {i} = f _ {\Delta i} (\hat {\mathbf {X}}). \tag {7} +$$ + +The input feature $\mathbf{X}$ is also fed into a sparse gating mechanism, which produces associated importance weights $w_{i}$ corresponding to each discretization $\pmb{\Delta}_{i}$ . Through the class-conditional gating that will be introduced in Sec. 4.3, we dynamically control the number of discretization patterns, $N_{k}$ , to be selected for each class $k$ based on the input features $\hat{\mathbf{X}}$ . Finally, the selected discretization patterns are aggregated with their importance weights as follows: + +$$ +\boldsymbol {\Delta} = \sum_ {i \in \Omega (\hat {\mathbf {X}})} w _ {i} \cdot \boldsymbol {\Delta} _ {i}, \quad | \Omega (\hat {\mathbf {X}}) | = N _ {k}, \tag {8} +$$ + +where $\Omega (\hat{\mathbf{X}})$ is the index set of the top- $N_{k}$ discretization candidates of $\hat{\mathbf{X}}$ , according to their importance weights. + +It's noteworthy that $\Delta$ is important because it controls the decay rate during the state update process, enabling the model to flexibly select and retain important information when handling long sequences, while avoiding the accumulation of redundant and irrelevant information [13], especially in cases of rapid changing states and evolving systems, as exemplified by OCL. Our method with mixture of discretization facilitates the learning of a precise selective scan pattern with limited data context in OCL. + +# 4.3. Class-Conditional Routing + +With more discretization patterns selected, the model is able to allocate more capacity for learning new knowledge, but may cause a significant shift of the model that impairs the existing ability. If we can calculate $N_{k}$ based on the misclassification probability of each class $k$ , we can dynamically control the capacity of mixture of discretization based on the learning conditions of all classes. Therefore, in order to guide the mixture of discretization to strike a balance between maintaining the stability of old knowledge and fostering the plasticity for new ability, we further design class-conditional routing. During training, we maintain the feature prototypes $\mathbf{M} = \{\mathbf{M}_c\}$ by moving average, where $\mathbf{M}_c$ is the within-class feature mean for class $c$ . Then, we estimate the class uncertainty $\sigma_{k}$ based on the average margin of class $k$ with different classes, + +$$ +\sigma_ {k} = \operatorname {A v g} _ {c} \left\{\exp \left(- \lambda_ {0} \| \mathbf {M} _ {k} - \mathbf {M} _ {c} \| _ {2}\right) \right\}, \tag {9} +$$ + +where $\lambda_0$ is a hyper-parameter and $\sigma_{k}$ denotes the class uncertainty for class $k$ ranging from $(0,1)$ . A large $\sigma_{k}$ indicates that the class center $k$ has narrow margins with the other class centers, and thus tends to be misclassified, while a small $\sigma_{k}$ happens when class center $k$ is distant from the other class centers with a lower likelihood of being misclassified. In inference, we replace $\mathbf{M}_k$ with the input feature $\hat{\mathbf{X}}$ and calculate its uncertainty by Eq. (9) with the feature prototypes $\mathbf{M}_c$ of all classes. + +Then, we multiply the uncertainty by the number of total discretization candidates $N$ , to get the discretization number to select for class $k$ or input feature $\hat{\mathbf{X}}$ as follows, + +$$ +N _ {k} = \operatorname {c e i l} \left(N \cdot \sigma_ {k}\right), \tag {10} +$$ + +where $\text{ceil}$ refers to the operation that rounds up $N \cdot \sigma_k$ to the nearest integer. + +The institution is that classes that are more prone to misclassification have smaller margins with the other classes in the feature space, and thus needs to aggregate more discretization patterns to allocate additional capacity for adapting to these undeveloped abilities. Conversely, for categories that are easier to classify with low uncertainty, a smaller $N_{k}$ is favored to only include the most likely discretization patterns, which can spare optimization efforts for the uncertain classes and stabilize these already acquired abilities. The combination of class-conditional routing and our mixture of discretization achieves dynamic structures dependent on input features and classes, such that the final prediction of each class selectively optimizes the corresponding discretization patterns. + +To further reduce interaction among different classes and facilitate the learning of generalizable and discriminative features, we also incorporate contrastive discretization loss function $\mathcal{L}_{\mathrm{Cont}}$ , as shown in Fig. 1 (d). It encourages $\pmb{\Delta}$ to + +have within-class consistency and between-class diversity as follows: + +$$ +\mathcal {L} _ {\text {C o n t}} = - \frac {1}{B ^ {2}} \sum_ {m = 1} ^ {B} \sum_ {n = 1} ^ {B} \left(1 _ {y _ {m} = y _ {n}} - 1 _ {y _ {m} \neq y _ {n}}\right) \frac {\boldsymbol {\Delta} _ {m} \cdot \boldsymbol {\Delta} _ {n}}{\| \boldsymbol {\Delta} _ {m} \| \| \boldsymbol {\Delta} _ {n} \|} \tag {11} +$$ + +where $B$ is the batch size, $\Delta_{m}$ represents the aggregated $\Delta$ by Eq. (8) for input $m$ , $y_{m}$ denotes the class of $m$ and $1_{y_m = y_n}$ is an indicator function that takes a value of 1 when the condition $y_{m} = y_{n}$ holds, and 0 otherwise. + +# 4.4. Optimization + +In addition to the previously mentioned loss functions in Eqs. (5) and (11), we also directly used the classification loss function $\mathcal{L}_{\mathrm{DR}}$ specifically designed for ETF classifier [51, 52] to supervise our introduced plug-and-play branch, and the loss function $\mathcal{L}_{\mathrm{z}}$ for maintaining load balancing among the discretization patterns [59]. In summary, when integrating S6MOD on a base method, the overall loss function can be expressed as follows: + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {b a s e}} + \mathcal {L} _ {\text {S 6 M O D}}, \tag {12} +$$ + +$$ +\mathcal {L} _ {\mathrm {S 6 M O D}} = \mathcal {L} _ {\mathrm {D R}} + \alpha \cdot \mathcal {L} _ {\text {D i f f}} + \beta \cdot \mathcal {L} _ {\text {C o n t}} + \mathcal {L} _ {\mathrm {z}}, \tag {13} +$$ + +where $\alpha$ and $\beta$ are hyperparameters. + +# 5. Experiments + +# 5.1. Experimental Setups + +Datasets. We test three datasets that are widely used in OCL, i.e., CIFAR-10 (10 classes) [20], CIFAR-100 (100 classes) [20] and TinyImageNet (200 classes) [21]. The dataset settings we adopted are the same as those in CCLDC[46]. We divide CIFAR-10 into 5 tasks, with 2 classes per task; CIFAR-100 into 10 tasks, with 10 classes per task; and TinyImageNet into 100 tasks, with 2 classes per task. Further details about the datasets will be provided in the supplementary materials. + +Baselines. To demonstrate the effectiveness and applicability of our module, we conduct tests on 7 typical and state-of-the-art methods, including ER [34], OCM [17], OnPro [46], OCM-CCLDC [44], OnPro-CCLDC [44], MOSE[50] and MOE-MOSE [50]. + +Implementation details. To ensure a fair comparison, we applied the same hyperparameter settings to each baseline and the methods combined with our module (S6MOD). All the above methods use ResNet-18 as the backbone, without pretraining. For the streaming input data, we set the batch size to 10, and for the samples drawn from the buffer, + +Table 1. Average Accuracy (\%, higher is better) on three benchmark datasets with difference memory buffer size $M$ , with and without our proposed S6MOD module. All values are averages of 10 runs. + +
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]56.68±1.8962.32±4.1324.47±0.7231.89±1.4539.41±1.8110.82±0.7919.16±1.4224.71±2.52
ER + Ours57.88±3.3065.80±2.1626.50±2.2334.55±1.6639.61±3.1610.94±1.4719.67±1.3625.62±1.73
OCM [17]68.19±1.7573.15±1.0528.02±0.7435.69±1.3642.22±1.0618.36±0.9526.74±1.0231.94±1.19
OCM + Ours70.57±1.1475.31±1.1032.44±1.3538.97±2.2845.49±1.5819.12±1.6027.20±1.8332.31±1.72
OnPro [46]70.47±2.1274.70±1.5127.22±0.7733.33±0.9341.59±1.3814.32±1.4021.13±2.1226.38±2.18
OnPro + Ours72.30±1.0875.14±0.9928.84±0.5937.57±0.8144.14±1.3617.27±0.6525.62±0.7929.20±1.00
OCM-CCLDC [44]74.14±0.8577.66±1.4635.00±1.1543.34±1.5151.43±1.3723.36±1.1833.17±0.9739.25±0.88
OCM-CCLDC + Ours74.45±1.2078.21±1.0336.02±1.7744.40±2.2652.53±0.2123.68±1.0233.59±0.7239.53±1.02
OnPro-CCLDC [44]74.49±2.1478.64±1.4234.76±1.1241.89±0.8250.01±0.8521.81±1.0232.00±0.7238.18±1.02
OnPro-CCLDC + Ours75.03±1.0379.51±0.6336.07±0.7643.92±0.9650.85±0.5222.03±0.8233.29±0.3538.51±0.91
MOSE [50]61.02±1.4770.74±1.1835.05±0.3445.06±0.3254.53±0.7818.23±0.7330.98±0.6338.71±0.44
MOSE + Ours63.74±1.5572.54±0.1435.43±0.6245.38±0.3154.75±0.5219.03±0.8332.03±0.8138.18±1.02
MOE-MOSE [50]62.54±1.5972.18±1.2937.32±0.3447.03±0.5755.62±0.7220.61±0.6932.52±0.3338.41±0.53
MOE-MOSE + Ours63.86±1.1973.16±0.5337.76±0.5147.25±0.4356.32±0.5321.03±0.7733.53±0.5140.11±0.26
+ +Table 2. Average Forgetting (%, lower is better) on three benchmark datasets with difference memory buffer size $M$ , with and without our proposed S6MOD module. All values are averages of 10 runs. + +
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]33.16±3.5020.94±6.7932.65±1.7822.20±2.2613.29±1.9858.38±1.6946.87±1.6040.77±2.45
ER + Ours31.25±3.4416.71±2.8630.96±2.0719.02±2.1312.97±2.6158.03±2.1945.83±1.4937.62±1.53
OCM [17]13.68±4.2511.63±2.6214.99±1.559.16±1.753.76±1.1626.12±1.6319.74±1.3015.92±1.47
OCM + Ours13.58±3.269.68±3.4715.99±1.8211.60±1.495.46±0.9824.23±2.0719.12±1.8714.66±1.75
OnPro [46]17.94±3.6914.20±2.6016.76±2.4712.42±1.396.72±0.9428.01±1.5923.52±1.7520.32±1.70
OnPro + Ours7.13±1.445.73±1.8116.03±1.699.81±1.325.07±0.8920.92±1.1216.82±0.9718.15±1.91
OCM-CCLDC [44]11.59±2.249.18±2.0316.69±2.3610.07±1.373.99±0.7826.16±1.9019.99±1.9615.56±1.06
OCM-CCLDC + Ours15.81±2.6911.28±1.8316.64±1.859.08±2.063.48±0.3125.29±1.9514.89±0.7411.55±1.31
OnPro-CCLDC [44]19.89±4.0114.62±2.7528.93±2.1920.23±1.0310.55±1.8928.21±1.5820.85±1.1316.17±0.63
OnPro-CCLDC + Ours17.23±3.1611.51±2.3322.26±1.1812.49±1.594.88±1.3626.73±1.5116.69±0.3112.23±1.37
MOSE [50]30.36±1.6920.27±1.2737.54±0.4325.89±0.4513.60±0.5947.16±1.4124.96±0.6215.51±0.33
MOSE + Ours27.38±1.9417.92±0.3437.09±0.6325.80±0.3915.55±0.6146.60±1.0724.39±0.7220.15±0.74
MOE-MOSE [50]29.39±1.7919.24±1.4935.17±0.3023.99±0.5112.81±0.7441.98±1.4622.22±0.3813.94±0.59
MOE-MOSE + Ours28.13±1.4117.98±0.5634.76±0.6623.84±0.7714.03±0.4841.26±1.0221.39±0.5814.91±0.28
+ +the batch size is set to 64. We retain each baseline's original data augmentation methods without modifying the base methods. For more details, please refer to the supplementary materials. + +# 5.2. Results + +We combine our method with both classical and state-of-the-art approaches on CIFAR-10, CIFAR-100, and TinyImageNet. The experimental results in Table 1 demonstrate the universality of our method, as it consistently improves accuracy across numerous baselines. On CIFAR-10 and CIFAR-100, our method generally leads to a $1\%$ improvement, while on TinyImageNet, we achieve significant gains under OCM, OnPro and MOE-MOSE. Notably, in settings like CIFAR100 $(M = 2k)$ and TinyImageNet $(M = 5k)$ , OnPro result in an approximate $4.2\%$ and $4.5\%$ improvement, respectively. It is also worth mention + +ing that even for state-of-the-art distillation-based methods like CCLDC and MOE-MOSE, the performance improvements are quite substantial. For instance, OnPro-CCLDC with the S6MOD integration achieve an impressive $79.51\%$ on CIFAR10 ( $M = 1k$ ), and MOE-MOSE with S6MOD reach $56.32\%$ on CIFAR100 ( $M = 5k$ ) and $40.1\%$ on TinyImageNet ( $M = 10k$ ). + +In addition, as shown in Table 2, our module is also highly effective in reducing model forgetting, achieving lower forgetting rates in most settings, with some cases showing particularly significant improvements. This can be attributed to our class-conditional routing. In a few baseline settings, the inclusion of our module led to a higher forgetting rate, but this does not imply that the model is more prone to forgetting. In these cases, we think the higher forgetting rates are often due to the stronger learning capabilities of the model. Analyzing both Table 1 and Table 2, we + +Table 3. Ablation studies on CIFAR-100 ( $M = 2k$ ). The "branch" refers to using SS2D[26] as the core of the S6MOD, without routing and $\mathcal{L}_{\mathrm{Cont}}$ . "routing" represents class-conditional routing. S6MOD refers to the composition of (branch + class-conditional routing + $\mathcal{L}_{\mathrm{Cont}}$ ). All values are averaged over 10 runs. + +
MethodAcc. ↑AF ↓
OnPro33.33±0.9312.42±1.39
OnPro + branch (with LDiff)36.61±1.0410.28±1.59
OnPro + branch + routing36.92±0.728.85±0.81
OnPro + branch + LCont36.93±0.849.70±1.42
OnPro + S6MOD37.57±0.819.81±1.32
+ +can find that even in these settings with higher forgetting rates, the accuracy performance remains quite competitive. + +# 5.3. Ablation Studies + +As mentioned in Sec. 1, the S6MOD module we propose is very simple and can be directly applied to existing baselines. By adding an extra branch and using the $\mathcal{L}_{\mathrm{Diff}}$ to supervise the original classification head, we can achieve a significant improvement. To demonstrate this, we conduct an ablation study where we only add the extra branch and $\mathcal{L}_{\mathrm{Diff}}$ to the baseline. The results in Table 3 confirm this, showing a $3.3\%$ improvement with supervision from the extra branch alone. + +In addition, to verify the effectiveness and adaptability of class-conditional routing and contrastive discretization loss $\mathcal{L}_{\mathrm{Cont}}$ , we conduct separate ablation experiments. As shown in Table 3, when added to the base method with the extra branch, the improvements brought by adding class-conditional routing or $\mathcal{L}_{\mathrm{Cont}}$ alone are not significant, with accuracy gains of only $0.31\%$ and $0.32\%$ , respectively. However, when both components are used together, the improvement becomes significant, with an approximate $1\%$ increase in accuracy, and the forgetting rate still decreases. This indicates that when our module is correctly combined, it can significantly enhances the model's adaptability by learning more generalizable and discriminative features through dynamic parameters. + +# 5.4. Analysis + +Analysis of Feature Embeddings. To verify that S6MOD improves the adaptability of the baseline and helps it learn more discriminative features, we use t-SNE [40] visualization for analysis. We compare with the baseline under the CIFAR-100 ( $M = 2k$ ) setting. As shown in Fig. 2, we visualize the features of the samples in the buffer after the final training, which are the features $\mathbf{F}$ output by the backbone just before they are passed to the classifier for classification. The features shown in Fig. 2 (a) and (c) are very scattered, with only a few classes showing some degree of aggregation. Most classes do not have clear separations between other classes, making it difficult to distinguish. In contrast, Fig. 2 (b) and (d) present the feature distribution + +![](images/dee5334ca91198dbfe3bf8014e387bbe9f54663822eedfd246385311b5230cce.jpg) +(a) OnPro + +![](images/3242c70be735c2f67ee71d100c33f530f7965e3de772d801e7f595b8e646aaec.jpg) +(b) OnPro + S6MOD + +![](images/709d94a5ea9bf0fa553c1cb3994394fc1712f0bdbe1b9883ba92168e7fb8b260.jpg) +(c) OCM + +![](images/1a16046da30c40d9d32a88fdf34a4d7365e1a22ee49c631b58ec7bf23e6218f1.jpg) +(d) OCM + S6MOD + +![](images/eed097c6eb67dfc27acafa55c89044fa14f762eef4a279430749d907bd6dedb3.jpg) +Figure 2. t-SNE visualization of memory data at the end of training on CIFAR-100 ( $M = 2k$ ), showcasing baseline in (a) and (c) and baseline combined with S6MOD in (b) and (d). Different colors represent different classes. +(a) New-Task Accuracy +Figure 3. Impact of dynamically selecting different $N_{k}$ values on the ability to learn new tasks and prevent forgetting of old tasks: New-Task accuracy represents the model's accuracy on the current task. We conduct experiments by setting $N_{k} = 1$ , $N_{k} = N$ , and calculating $N_{k}$ through class-conditional routing. The dataset used is CIFAR-100 ( $M = 2k$ ). + +![](images/af18e09dfca09cb93f91eeadd68afffb784431d74cc3427634b9938df0e4d7f1.jpg) +(b) Average Forgetting + +after adding our module. We can observe that many categories have noticeably converged, forming small clusters with consistent intra-class compactness and clear inter-class separations. These features are much easier to distinguish and more conducive to classification by the classifier, which aligns with the significant improvements we demonstrate in the experiments. + +Analysis of Class-conditional Routing. We further investigate the effectiveness of class-conditional routing. Under the CIFAR-100 ( $M = 2k$ ) setting, we design experiments to compare the impact of a fixed $N_{k}$ and dynamically calculated $N_{k}$ using class-conditional routing. Specifically, we set $N_{k} = N$ , meaning all patterns are activated, and $N_{k} = 1$ , meaning only the pattern with the highest prob + +Table 4. Impact of hyper-parameters $\alpha$ and $\beta$ on performance in the CIFAR-100 dataset ( $M = 2k$ ). $\alpha$ and $\beta$ are the weights for $\mathcal{L}_{\mathrm{Diff}}$ and $\mathcal{L}_{\mathrm{Cont}}$ , respectively. Acc. and AF represent the average accuracy and the average forgetting over 5 runs, respectively. + +
Impact of hyper-parameter α
Value0.010.10.515
Acc.36.80±1.1737.09±1.0937.55±1.0837.47±0.9437.57±0.81
AF10.49±1.019.65±0.399.05±1.259.16±2.119.81±1.32
Impact of hyper-parameter β
Value0.1151025
Acc.36.50±1.2437.30±1.0337.40±0.3337.26±0.6137.57±0.81
AF10.24±0.4310.12±1.0910.10±0.609.98±0.669.81±1.32
+ +ability is activated. To facilitate analysis, we present the accuracy of the model on the latest task (the current task) in Fig. 3 (a). As shown in Fig. 3 (a), when $N_{k} = N$ , all patterns are activated, and thanks to the inclusion of more patterns with different weights and larger parameters, the model has greater capacity to adapt to new tasks, resulting in a stronger ability to learn new knowledge. When $N_{k} = 1$ , only the pattern with the highest probability is activated, significantly reducing the model's ability to learn the new task. However, as shown in Fig. 3 (b), when $N_{k} = N$ , since all patterns are updated by the current task, the model has a higher risk of forgetting. When $N_{k} = 1$ , only the pattern with the highest probability is updated, and tasks do not interfere with each other, so forgetting is less likely. Combining the analysis of the entire Fig. 3, the effect of our proposed class-conditional routing is both intuitive and significant. It not only helps the model achieve strong learning ability for new tasks, comparable to the $N_{k}$ setting, but also provides excellent anti-forgetting performance, similar to $N_{k} = 1$ . This is due to the dynamic adjustment of $N_{k}$ based on uncertainty: for samples that are prone to misclassification, class-conditional routing assigns a larger $N_{k}$ , enabling more extensive learning. For easier-to-classify samples, the model's learning ability is higher, and class-conditional routing assigns a smaller $N_{k}$ to enhance anti-forgetting. + +Sensitivity Analysis of Hyper-parameters $\alpha$ and $\beta$ . $\alpha$ and $\beta$ are the weights for $\mathcal{L}_{\mathrm{Diff}}$ and $\mathcal{L}_{\mathrm{Cont}}$ , respectively. We conduct experiments on the sensitivity of them, and the results are shown in Table 4. It can be observed that when the weights are low, the optimization effect of the loss is relatively weak, both accuracy and forgetting rate do not perform well. Gradually increasing the weights, we find that after a certain point, the impact of the weights on the results becomes relatively stable, with no more drastic changes. + +Sensitivity Analysis of total discretization patterns number $N$ . We conduct experiments on the total number of discretization patterns in the module. The results are shown + +![](images/2f880a619373a0e7ad688d11998bc6e7c6870626c6df92de9fa2343e03a5add4.jpg) +(a) New-Task Average Accuracy (b) Average Forgetting w/o class-w/o class-conditional routing conditional routing + +![](images/bb54119721d252a830c3c0a1d8ad2d5d16dd42b102922c8d56f9c5fc62e1d87c.jpg) + +![](images/1e14ddd1b7dc09cbadec1ced7e49b5f258b1ef9e7751088c28be86dc5b04a503.jpg) +(c) New-Task Average Accuracy + +![](images/8fbee29e319e85a8d06830280eae17ac9f6c1dfce5abbd379d95c25f1343a3af.jpg) +(d) Average Forgetting +Figure 4. Impact of pattern number $N$ . The dataset used is CIFAR-100 ( $M = 2k$ ). + +in Fig. 4. As the number $N$ increases, the model gains a stronger ability to learn new tasks, but at the same time, its tendency to forget becomes more pronounced. Furthermore, by comparing (a)(b) and (c)(d) in Fig. 4, we can conclude that class-conditional routing can significantly reduce forgetting with only a slight decrease in new task accuracy, which in turn leads to an overall improvement in performance. This indirectly supports the effectiveness of class-conditional routing in our ablation studies. + +# 6. Conclusion + +In this paper, we propose a plug-and-play module S6MOD to enhance the adaptability of existing OCL methods by introducing selective state space models (SSMs) with a class-conditional mixture of discretization. By integrating a dynamic discretization mechanism and leveraging class-conditional strategies, our method can efficiently allocate discretization patterns based on class uncertainty, improving both the model's generalization and its ability to adapt to new data. Experimental results on multiple OCL datasets demonstrate that our method consistently optimize existing techniques, contributing to more robust OCL systems. + +# 7. Acknowledgments + +The research reported in this publication was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940. This research used computational resources provided by IBEX managed by the Supercomputing Core Laboratory at KAUST. + +# References + +[1] Assyr Abdulle, E Weinan, Björn Engquist, and Eric Vanden-Eijnden. The heterogeneous multiscale method. Acta Numerica, 21:1-87, 2012. 2 +[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2 +[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. NeurIPS, 32, 2019. 1 +[4] Quentin Anthony, Yury Tokpanov, Paolo Glorioso, and Beren Millidge. Blackmamba: Mixture of experts for statespace models. arXiv preprint arXiv:2402.01771, 2024. 3 +[5] Fengxiang Bie, Yibo Yang, Zhongzhu Zhou, Adam Ghanem, Minjia Zhang, Zhewei Yao, Xiaoxia Wu, Connor Holmes, Pareesa Golnari, David A Clifton, et al. Renaissance: A survey into ai text-to-image generation in the era of large model. TPAMI, 2024. 1 +[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. *NeurIPS*, 33: 15920-15930, 2020. 2 +[7] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 1 +[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2 +[9] Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688, 2023. 1 +[10] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 44(7):3366-3385, 2021. 1, 2 +[11] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 4 +[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3 +[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3, 4 +[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3 +[15] John Guckenheimer and Philip Holmes. Nonlinear oscillations, dynamical systems, and bifurcations of vector fields. Springer Science & Business Media, 2013. 2 + +[16] Nuwan Gunasekara, Bernhard Pfahringer, Heitor Murilo Gomes, and Albert Bifet. Survey on online streaming continual learning. In *IJCAI*, pages 6628-6637, 2023. 1, 2 +[17] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In ICML, 2022. 1, 2, 5, 6 +[18] Yiduo Guo, Bing Liu, and Dongyan Zhao. Dealing with cross-task class discrimination in online continual learning. In CVPR, 2023. 1, 2 +[19] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. Neural Computation, 3(1):79-87, 1991. 4 +[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5 +[21] Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015. 5 +[22] Xiaojie Li, Yibo Yang, Jianlong Wu, Bernard Ghanem, Liqiang Nie, and Min Zhang. Mamba-fscil: Dynamic adaptation with selective state space model for few-shot class-incremental learning. arXiv preprint arXiv:2407.06136, 2024. 1, 2, 3 +[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 40(12):2935-2947, 2017. 2 +[24] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 1, 3 +[25] Sihao Liu, Mingyi Liu, Tianyu Jiang, Shuang Yu, Hanchuan Xu, and Zhongjie Wang. GSL-mash: Enhancing mashup creation service recommendations through graph structure learning. In Service-Oriented Computing, pages 176-191, Singapore, 2025. Springer Nature Singapore. 1 +[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. ArXiv, abs/2401.10166, 2024. 1, 2, 3, 7 +[27] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 30, 2017. 2 +[28] Zheda Mai, Ruiwen Li, Hyunwoo Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. In CVPR, pages 3589-3599, 2021. 2 +[29] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 1, 2 +[30] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3 +[31] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 2 +[32] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, Michał Krutul, Jakub Krajewski, Szymon Antoniak, Pi + +otr Mioś, Marek Cygan, and Sebastian Jaszcur. Moemamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 3 +[33] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2 +[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. *NeurIPS*, 32, 2019. 1, 2, 5, 6 +[35] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 1 +[36] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2 +[37] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In CVPR, pages 23933-23942, 2024. 2 +[38] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, pages 4548-4557. PMLR, 2018. 2 +[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 4 +[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 7 +[41] A Vaswani. Attention is all you need. NeurIPS, 2017. 2 +[42] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. NeurIPS, 34:22379-22391, 2021. 2 +[43] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. TPAMI, 2024. 1, 2 +[44] Maorong Wang, Nicolas Michel, Ling Xiao, and Toshihiko Yamasaki. Improving plasticity in online continual learning via collaborative learning. In CVPR, pages 23460-23469, 2024. 1, 2, 5, 6 +[45] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, pages 18764-18774, 2023. 1 +[46] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, 2023. 1, 2, 5, 6 +[47] Mitchell Wortsman, Vivek Ramanujan, Rosanne Liu, Aniruddha Kembhavi, Mohammad Rastegari, Jason Yosinski, and Ali Farhadi. Supermasks in superposition. NeurIPS, 33:15173-15184, 2020. 2 +[48] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2 + +[49] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 1 +[50] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In CVPR, pages 23670-23680, 2024. 1, 2, 5, 6 +[51] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, pages 37991-38002. Curran Associates, Inc., 2022. 2, 3, 5 +[52] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR, 2023. 2, 3, 5 +[53] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. Corda: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In NeurlPS, 2024. 2 +[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2 +[55] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 4 +[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995. PMLR, 2017. 2 +[57] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19560, 2023. 2 +[58] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2 +[59] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5 \ No newline at end of file diff --git a/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/images.zip b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1047d51ce2bd035b5e65272a7a85c83f5cb83e1 --- /dev/null +++ b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5bd3aba8688e7995b30376476dc93d36d421a79a88383c010c9bb4768c57f59 +size 568827 diff --git a/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/layout.json b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7795c5dd1955716a17e16ec67324265ddef89bbf --- /dev/null +++ b/2025/Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization/layout.json @@ -0,0 +1,10249 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 138 + ], + "type": "text", + "content": "Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "text", + "content": "Sihao Liu" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "text", + "content": " Yibo Yang" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "text", + "content": " Xiaojie Li" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "text", + "content": " David A. Clifton" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "text", + "content": " Bernard Ghanem" + }, + { + "bbox": [ + 107, + 160, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 176, + 485, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 176, + 485, + 190 + ], + "spans": [ + { + "bbox": [ + 121, + 176, + 485, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 176, + 485, + 190 + ], + "type": "text", + "content": "Harbin Institute of Technology " + }, + { + "bbox": [ + 121, + 176, + 485, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 176, + 485, + 190 + ], + "type": "text", + "content": "King Abdullah University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 190, + 444, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 190, + 444, + 203 + ], + "spans": [ + { + "bbox": [ + 164, + 190, + 444, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 164, + 190, + 444, + 203 + ], + "type": "text", + "content": "Harbin Institute of Technology (Shenzhen) " + }, + { + "bbox": [ + 164, + 190, + 444, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 164, + 190, + 444, + 203 + ], + "type": "text", + "content": "University of Oxford" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 256, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 543 + ], + "type": "text", + "content": "Online continual learning (OCL) seeks to learn new tasks from data streams that appear only once, while retaining knowledge of previously learned tasks. Most existing methods rely on replay, focusing on enhancing memory retention through regularization or distillation. However, they often overlook the adaptability of the model, limiting the ability to learn generalizable and discriminative features incrementally from online training data. To address this, we introduce a plug-and-play module, S6MOD, which can be integrated into most existing methods and directly improve adaptability. Specifically, S6MOD introduces an extra branch after the backbone, where a mixture of discretization selectively adjusts parameters in a selective state space model, enriching selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. We further design a class-conditional routing algorithm for dynamic, uncertainty-based adjustment and implement a contrastive discretization loss to optimize it. Extensive experiments combining our module with various models demonstrate that S6MOD significantly enhances model adaptability, leading to substantial performance gains and achieving the state-of-the-art results. The code is available at https://github.com/MyToumaKazusa/S6MOD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "type": "text", + "content": "Continual learning (CL) [10, 29, 31, 43] is a task that requires models to efficiently acquire new abilities from incoming data. It tests the model's ability to adapt to dynamically changing environments while retaining existing knowledge. For example, autonomous vehicles are expected to continuously learn new driving environments and traffic regulations, thereby improving detection accuracy under complex driving scenes. From the perspective of task format, CL can be divided into offline and online [10, 43]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 292 + ], + "type": "text", + "content": "Unlike offline CL [7, 33], online CL (OCL) [10, 16] aligns with real-world implementations as it requires data to arrive sequentially in mini-batches and allows only one epoch of training, posing greater challenges for efficient adaptation with data accessible for learning only once." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 293, + 556, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 509 + ], + "type": "text", + "content": "To tackle the challenging OCL, existing studies widely rely on the replay technique [17] that selectively stores a subset of old-class data to strengthen the memory ability of existing knowledge and mitigate catastrophic forgetting. Building on replay, various methods have been proposed with regularization, buffer allocation, and distillation strategies [17, 18, 34, 44, 46, 50]. For example, MIR [3] aims to mitigate mutual interference among tasks through a retrieval strategy, OCM [17] seeks to reduce forgetting by maximizing mutual information between different tasks, and CCLDC [44] adopts collaborative learning and distillation to improve plasticity. However, learning generalizable and discriminative features incrementally from online training data is still intractable, and imprecise features will impede and even mislead the replay strategies [45]. Therefore, inducing accurate and efficient adaptation with limited data access is critically important for OCL and there still remains significant potential for improvement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Recently, selective state space models (SSMs), also known as S6 in Mamba [13], have shown promising results in modeling long-range dependencies with computational efficiency. Selective SSMs introduce a selective scan mechanism to make the interactions among sequential states aware of input context, and have been applied to vision tasks by integrating various scan directions [26, 58]. In addition to the success in sequence modeling [5, 9, 24, 25, 35, 49], selective SSMs also exhibit stronger adaptability than static parameters in few-shot class-incremental learning due to the dynamic operation weights, as reported in a recent pioneering work [22]. Despite the increased capacity of adaptation, directly applying selective SSMs to OCL will be infeasible because it is challenging for selective SSMs to capture a precise discretization pattern with limited context from online data that appears only once for training. Moreover, a single discretization mechanism may fail to capture all the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20502" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "essential details in some nonlinear dynamic systems or systems with multi-scale characteristics [1, 15]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 97, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 97, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 97, + 295, + 277 + ], + "type": "text", + "content": "To this end, we develop a plug-and-play branch based on selective state space model and class-conditional mixture of discretization. The branch with an equiangular tight frame classifier [29, 37, 48, 51, 52, 57] can be integrated on any existing OCL methods to supervise the original classification head and improve the adaptation ability in OCL. Considering multiple discretization methods can exhibit varying sensitivity to the dynamic characteristics of a system, we introduce a mixture of discretization into SSMs to enrich selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. This flexibility allows our method to comprehensively capture the system's complex dynamic features, particularly in cases of rapid state changes and evolving systems as exemplified by OCL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 278, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 295, + 517 + ], + "type": "text", + "content": "In order to guide the mixture of discretization for OCL, where a delicate trade-off between maintaining the stability of old knowledge and fostering the plasticity for new ability is needed, we further introduce a class-conditional routing to aggregate the discretization patterns. Concretely, we maintain a feature prototype for each class to calculate the class uncertainty based on the margin among different classes. For classes with less uncertainty, fewer discretization patterns will be aggregated to stabilize the abilities already acquired for these classes, while for classes with large uncertainty, more discretization patterns will be included to allocate additional capacity for adapting to these undeveloped abilities. After aggregation of the mixture of discretization, we employ a contrastive discretization loss that enforces within-class consistency and between-class diversity, contributing to the learning of generalizable and discriminative features after the selective scan. In experiments, we integrate our plug-and-play branch on numerous OCL methods and significant improvements can be consistently observed on multiple datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 518, + 266, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 518, + 266, + 529 + ], + "spans": [ + { + "bbox": [ + 66, + 518, + 266, + 529 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 532, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "text", + "content": "- To enhance the adaptability of existing methods for OCL, we propose S6MOD, a plug-and-play module based on selective state space model and class-conditional mixture of discretization, which can strengthen the base method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 580, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 650 + ], + "type": "text", + "content": "- We further develop a class-conditional routing strategy that dynamically satisfy both objectives of maintaining the stability of knowledge already acquired and fostering the plasticity for undeveloped abilities. A contrastive discretization loss is employed to facilitate the learning of generalizable and discriminative features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 651, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 711 + ], + "type": "text", + "content": "- In experiments, our method can be easily integrated on different OCL methods and is compatible with distillation techniques, to consistently achieve significant improvements on multiple datasets including CIFAR-10, CIFAR-100, and Tiny-ImageNet." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "type": "text", + "content": "Continual Learning(CL). CL can be categorized into three types based on the implementation methods [10, 29, 31, 43]: regularization-based, parameter isolation-based, and replay-based. Regularization-based methods [2, 23, 42, 56] introduce regularization terms to restrict changes in model parameters, thereby protecting the already learned knowledge. Parameter isolation-based methods [36, 38, 47, 54] typically assign different model parameters or subnetworks to different tasks to avoid interference between tasks. Replay-based methods [6, 8, 27, 33, 34] perform joint training by either storing a portion of old data or generating pseudo-data. This allows the model to revisit old data while learning new data, thereby reducing forgetting [53]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 265, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 540 + ], + "type": "text", + "content": "Online Continual Learning(OCL). OCL is a specialized form of CL designed to test the ability to continuously learn and update as data arrives in real-time streams [10, 16, 28]. This means that OCL can only train for a single epoch and typically process only individual samples or mini batches at a time. Replay-based methods are widely used in OCL [17]. ER [34] introduces the combination of cross-entropy loss with a random buffer. OCM [17] uses mutual information maximization to reduce feature bias and preserve past knowledge. GSA [18] proposes a gradient-based adaptive optimization method to address dynamic training bias. On-Pro [46] uses online prototype equilibrium to address shortcut learning problem. Some methods introduce knowledge distillation based on replay. CCL-DC [44] introduces Collaborative Continual Learning (CCL) and Distillation Chain (DC) to enhance model plasticity. MOSE [50] alleviates forgetting by integrating multi-level supervision and reverse self-distillation. However, these models carry a risk when the features used to distill are not precise. In comparison, our method S6MOD as a plug-and-play module is applicable to most OCL methods to improve the adaptability of these models, enabling them to learn generalizable and discriminative features more efficiently." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "Selective State Space Model (S6) Selective State Space Model (S6) [13] has gained increasing interest as an alternative to self-attention [41] with lower computational complexity. S6 enhances the S4 model [14] by introducing the selective scan mechanism, and its effectiveness in vision tasks has been extensively studied [26, 58]. For example, Vmamba [26] introduces SS2D, a cross-scanning mechanism for images, facilitating the extension of Mamba to process vision data. A recent study Mamba-FSCIL [22] leverages the dynamic weights and sequence modeling capability of Mamba to achieve dynamic adaptation in few-shot class-incremental learning. But their method relies on an effective selective scan learned from training data. In OCL," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20503" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "the online data that appears only once for training provides limited context for Mamba models, and thus poses challenges to capture a precise discretization pattern. Different from these studies, our method integrates S6 with class-conditional mixture of discretization and effectively helps to improve the adaptability for OCL. Some existing works have introduced mixture of experts (MoE) into Mamba [4, 24, 32], following the design of switch-transformer [12]. Our method differs from them in that each expert in our mixture of discretization is only a simple linear layer for computation efficiency, and discretization patterns are aggregated with our class-conditional gating to balance between maintaining stability of old knowledge and allocating capacity for learning new abilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 141, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 141, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 141, + 261 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 206, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 206, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 206, + 281 + ], + "type": "text", + "content": "3.1. Problem Definition of OCL" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": "OCL requires a model to continuously update itself from a data stream, with each mini-batch containing new data sampled from a changing distribution " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": ". At each time step " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": ", the model receives a mini-batch of data " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\{(s_{i}^{(t)}, r_{i}^{(t)})\\}_{i=1}^{n_{t}}" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "s_{i}^{(t)}" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": " represents the input data, " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "r_{i}^{(t)}" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": " represents the corresponding labels, and " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "n_{t}" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": " is the number of samples. The goal of OCL is to sequentially update the model to adapt to both gradual and task-specific shifts in the data distribution. Given a pre-trained network " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": ", the model updates its parameters by solving an optimization problem " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\theta} L(f_{\\theta}(s), r)" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 286, + 296, + 445 + ], + "type": "text", + "content": " is the loss function, thereby enabling it to adapt effectively to the evolving data distribution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 452, + 169, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 169, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 169, + 464 + ], + "type": "text", + "content": "3.2. State Space Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": "State Space Models (SSMs) can be viewed as linear time-invariant systems. They map input sequences " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": " to output sequences " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": " through hidden states " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": ". Mathematically, these models can be represented as linear ordinary differential equations (ODE):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 534, + 294, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 534, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 124, + 534, + 294, + 553 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = A h (t) + B x (t), \\tag {1}", + "image_path": "09da237761be8f673e9055b2c07eb64bfbc155584c050346f4ab8c471072a139.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 551, + 189, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 551, + 189, + 563 + ], + "spans": [ + { + "bbox": [ + 129, + 551, + 189, + 563 + ], + "type": "interline_equation", + "content": "y (t) = C h (t),", + "image_path": "3d3081ecef440bf25b05959080b9396930827f71efa42e924187dd2a7a9ddd90.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " represents the state transition matrix, while " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{1 \\times N}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " denote the mapping matrices from input to latent state and from latent state to output, respectively, where " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " indicates the size of the hidden state." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "text", + "content": "To apply the SSM to real discrete data, zero-order hold (ZOH) [14] is employed. It discretizes the continuous parameters " + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "text", + "content": " using the time scale parameter " + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 630, + 296, + 678 + ], + "type": "text", + "content": ". The discretized SSM equations can be rewritten as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 686, + 295, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 686, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 132, + 686, + 295, + 706 + ], + "type": "interline_equation", + "content": "h _ {t} = \\bar {A} h _ {t - 1} + \\bar {B} x _ {t}, \\tag {2}", + "image_path": "99d7cec3117da369c9e6d41fa7e5bd09c180aa44b04186c8e85053a35bcf7629.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 704, + 178, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 704, + 178, + 715 + ], + "spans": [ + { + "bbox": [ + 135, + 704, + 178, + 715 + ], + "type": "interline_equation", + "content": "y _ {t} = C h _ {t}.", + "image_path": "d51b073beeaabdfb28c7afe1c23ff649289280d8954e3edffee5c5883444edb1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "Recently, Gu [13] proposed a new parameterization method for SSM with a selective scan mechanism, which is known as S6 and serves as the core of the Mamba model. To enhance S6's capability in processing visual data, Liu et al. [26] introduced SS2D, which can serialize image data from four different directions. Given the input data " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\overline{x}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " processed by SS2D can be expressed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 162, + 554, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 162, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 358, + 162, + 554, + 194 + ], + "type": "interline_equation", + "content": "\\bar {x} = \\operatorname {S S 2 D} (x) = \\sum_ {i = 1} ^ {4} \\operatorname {S 6} (\\operatorname {s c a n} (x, i)). \\tag {3}", + "image_path": "3d4837c0831b7e4cf528d1cc97e7c27b78743f6bec0879e4c3b054f42e4a9a23.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 200, + 421, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 421, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 421, + 213 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "text", + "content": "In this section, we will first introduce the overall structure of our plug-and-play module S6MOD in Sec. 4.1. After that, we will present the detailed design of the state space model with mixture of discretization in Sec. 4.2. Then, we will introduce the class-conditional routing and its corresponding contrastive discretization loss in details in Sec. 4.3. Finally, we will specify the optimization details in Sec. 4.4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 309, + 418, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 418, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 418, + 320 + ], + "type": "text", + "content": "4.1. Overall structure" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 327, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 471 + ], + "type": "text", + "content": "To better enhance existing methods' ability in learning generalizable and discriminative features incrementally from online training data, we propose a plug-and-play module named S6MOD that can be easily applied to existing OCL methods, as shown in Fig. 1 (a). Overall, S6MOD strengthens the original method by introducing a plug-and-play branch after the backbone. It consists of a block and a fixed equiangular tight frame (ETF) classifier [51, 52]. The intermediate features " + }, + { + "bbox": [ + 313, + 327, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 327, + 554, + 471 + ], + "type": "text", + "content": " generated by the backbone are duplicated and sent into the classification head of the original base method, and the selective state space model (SSM) branch introduced by our method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "text", + "content": "In our SSM branch, " + }, + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "text", + "content": " is projected into two paths through an MLP. The first one " + }, + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = f_{x}(\\mathbf{F})" + }, + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "text", + "content": " is used to perform selective scan with our class-conditional mixture of discretization, as will be detailed in later subsections. The other one " + }, + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = f_{z}(\\mathbf{F})" + }, + { + "bbox": [ + 313, + 471, + 554, + 554 + ], + "type": "text", + "content": " performs a gating mechanism as commonly adopted in Mamba models [22, 26, 30]. The output feature of this branch can be formulated as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 341, + 560, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 560, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 341, + 560, + 553, + 574 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} = \\operatorname {S i L U} (\\mathbf {Z}) \\otimes \\operatorname {S 6 M O D} (\\operatorname {S i L U} (\\operatorname {C o n v} (\\mathbf {X}))). \\tag {4}", + "image_path": "196b63ccbe771e75211e05828a2c26c3f5af3f1aa753bf3b0faafb965f08ca23.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "text", + "content": "To assist the learning of the base method for generalizable and discriminative features, we adopt a regularization, " + }, + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}" + }, + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "text", + "content": ", on the predicted distribution of the base method, which is the KL divergence between " + }, + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 579, + 554, + 639 + ], + "type": "text", + "content": " and can be formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 370, + 643, + 553, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 643, + 553, + 672 + ], + "spans": [ + { + "bbox": [ + 370, + 643, + 553, + 672 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {D i f f}} = \\sum_ {i} P (i) \\log \\left(\\frac {P (i)}{Q (i)}\\right), \\tag {5}", + "image_path": "3664b665041f1f45707573dab3c757c31ffaf5eaf0c76cb8bddb18494cb8e57c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " represents the predicted distribution of the base method, and " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " refers to the output prediction after the ETF classifier of our SSM branch, i.e., " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "Q = \\mathrm{softmax}(\\mathbf{W}_{\\mathrm{ETF}}\\boldsymbol {\\mu})" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20504" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 71, + 517, + 178 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 517, + 178 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 517, + 178 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 517, + 178 + ], + "type": "image", + "image_path": "a8338a9ed7bc850ad5ab28a83ba3eee2816c554b458ff4a891119dafa92d612f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 94, + 178, + 515, + 381 + ], + "blocks": [ + { + "bbox": [ + 94, + 178, + 515, + 381 + ], + "lines": [ + { + "bbox": [ + 94, + 178, + 515, + 381 + ], + "spans": [ + { + "bbox": [ + 94, + 178, + 515, + 381 + ], + "type": "image", + "image_path": "245cf302e4e89c134e247a787c1a5df0a1d38b2d97dc794fa850c28b88415bce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 388, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 54, + 388, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 555, + 434 + ], + "type": "text", + "content": "Figure 1. Framework of S6MOD. Our method (a) introduces a plug-and-play branch after the backbone, where features are learned through S6MOD and supervised by the ETF classifier to guide the base method. S6MOD (c) utilizes MoE to enhance the discretization of SSM and applies class-conditional routing (b) to dynamically adjust the discretization based on the uncertainty. Finally, we use a contrastive discretization loss (d) to supervise the learning of both generalizable and discriminative features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 445, + 234, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 234, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 234, + 456 + ], + "type": "text", + "content": "4.2. S6 with Mixture of Discretization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 535 + ], + "type": "text", + "content": "We integrate the selective space state model with a mixture of discretization, borrowing the concept of mixture of experts [11, 19, 39, 55]. Its structure is shown in Fig. 1 (c). Similar to the typical SSM structure, we also use an MLP layer to produce the projection matrices " + }, + { + "bbox": [ + 55, + 463, + 296, + 535 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 463, + 296, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 463, + 296, + 535 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 463, + 296, + 535 + ], + "type": "text", + "content": " from the input features, which can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 545, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 545, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 295, + 559 + ], + "type": "interline_equation", + "content": "\\mathbf {B} = f _ {B} (\\hat {\\mathbf {X}}), \\quad \\mathbf {C} = f _ {C} (\\hat {\\mathbf {X}}), \\tag {6}", + "image_path": "bd694b5ca42bc18125d7a951814a7a7ae302fb1cd107557db6a6eca942bbc9d1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " denotes the input of the selective scan module S6MOD, i.e., " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} = \\mathrm{SiLU}(\\mathrm{Conv}(\\mathbf{X}))" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": ". To enable selective SSM to capture a precise discretization pattern with limited context from online data, we develop a sparse MoE system for the discretization transformation " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " in S6 models, utilizing specialized projection layers to enrich the discretization patterns and adaptively selecting the most sensitive discretization based on the current dynamics. Specifically, each discretization candidate is produced by a projection layer as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 143, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Delta} _ {i} = f _ {\\Delta i} (\\hat {\\mathbf {X}}). \\tag {7}", + "image_path": "1d53bf55b3d6212f73d58ca6fa8d3f31737a5e1959ab903dab597c63df152021.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": "The input feature " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": " is also fed into a sparse gating mechanism, which produces associated importance weights " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": " corresponding to each discretization " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_{i}" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": ". Through the class-conditional gating that will be introduced in Sec. 4.3, we dynamically control the number of discretization patterns, " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": ", to be selected for each class " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": " based on the input features " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": ". Finally, the selected discretization patterns are aggregated with their importance weights as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 354, + 545, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 545, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 354, + 545, + 555, + 574 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Delta} = \\sum_ {i \\in \\Omega (\\hat {\\mathbf {X}})} w _ {i} \\cdot \\boldsymbol {\\Delta} _ {i}, \\quad | \\Omega (\\hat {\\mathbf {X}}) | = N _ {k}, \\tag {8}", + "image_path": "683e8b6f89231dd8fd07d15c207a86cdbdcd2a5c501fc33d993d65eb5f7ab327.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\Omega (\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "text", + "content": " is the index set of the top- " + }, + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "text", + "content": " discretization candidates of " + }, + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 580, + 554, + 605 + ], + "type": "text", + "content": ", according to their importance weights." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "It's noteworthy that " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": " is important because it controls the decay rate during the state update process, enabling the model to flexibly select and retain important information when handling long sequences, while avoiding the accumulation of redundant and irrelevant information [13], especially in cases of rapid changing states and evolving systems, as exemplified by OCL. Our method with mixture of discretization facilitates the learning of a precise selective scan pattern with limited data context in OCL." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20505" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 203, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 203, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 203, + 85 + ], + "type": "text", + "content": "4.3. Class-Conditional Routing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": "With more discretization patterns selected, the model is able to allocate more capacity for learning new knowledge, but may cause a significant shift of the model that impairs the existing ability. If we can calculate " + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": " based on the misclassification probability of each class " + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": ", we can dynamically control the capacity of mixture of discretization based on the learning conditions of all classes. Therefore, in order to guide the mixture of discretization to strike a balance between maintaining the stability of old knowledge and fostering the plasticity for new ability, we further design class-conditional routing. During training, we maintain the feature prototypes " + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{\\mathbf{M}_c\\}" + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": " by moving average, where " + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_c" + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": " is the within-class feature mean for class " + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": ". Then, we estimate the class uncertainty " + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": " based on the average margin of class " + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": " with different classes," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 277, + 294, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 277, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 94, + 277, + 294, + 291 + ], + "type": "interline_equation", + "content": "\\sigma_ {k} = \\operatorname {A v g} _ {c} \\left\\{\\exp \\left(- \\lambda_ {0} \\| \\mathbf {M} _ {k} - \\mathbf {M} _ {c} \\| _ {2}\\right) \\right\\}, \\tag {9}", + "image_path": "727432099ab379d4b6f946a4fc0f940b00de40bcdcd2475d5ff6e1b176b333ff.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\lambda_0" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " is a hyper-parameter and " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " denotes the class uncertainty for class " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " ranging from " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": ". A large " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " indicates that the class center " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " has narrow margins with the other class centers, and thus tends to be misclassified, while a small " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " happens when class center " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " is distant from the other class centers with a lower likelihood of being misclassified. In inference, we replace " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_k" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " with the input feature " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " and calculate its uncertainty by Eq. (9) with the feature prototypes " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_c" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " of all classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": "Then, we multiply the uncertainty by the number of total discretization candidates " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": ", to get the discretization number to select for class " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": " or input feature " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 453, + 294, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 294, + 464 + ], + "type": "interline_equation", + "content": "N _ {k} = \\operatorname {c e i l} \\left(N \\cdot \\sigma_ {k}\\right), \\tag {10}", + "image_path": "f2c3e55147ec12625b7896e49357acc2251351f8bc7facee532822a0685f95cf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\text{ceil}" + }, + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "text", + "content": " refers to the operation that rounds up " + }, + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "inline_equation", + "content": "N \\cdot \\sigma_k" + }, + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "text", + "content": " to the nearest integer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": "The institution is that classes that are more prone to misclassification have smaller margins with the other classes in the feature space, and thus needs to aggregate more discretization patterns to allocate additional capacity for adapting to these undeveloped abilities. Conversely, for categories that are easier to classify with low uncertainty, a smaller " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": " is favored to only include the most likely discretization patterns, which can spare optimization efforts for the uncertain classes and stabilize these already acquired abilities. The combination of class-conditional routing and our mixture of discretization achieves dynamic structures dependent on input features and classes, such that the final prediction of each class selectively optimizes the corresponding discretization patterns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "To further reduce interaction among different classes and facilitate the learning of generalizable and discriminative features, we also incorporate contrastive discretization loss function " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cont}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", as shown in Fig. 1 (d). It encourages " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "have within-class consistency and between-class diversity as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 103, + 553, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 103, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 315, + 103, + 553, + 146 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {C o n t}} = - \\frac {1}{B ^ {2}} \\sum_ {m = 1} ^ {B} \\sum_ {n = 1} ^ {B} \\left(1 _ {y _ {m} = y _ {n}} - 1 _ {y _ {m} \\neq y _ {n}}\\right) \\frac {\\boldsymbol {\\Delta} _ {m} \\cdot \\boldsymbol {\\Delta} _ {n}}{\\| \\boldsymbol {\\Delta} _ {m} \\| \\| \\boldsymbol {\\Delta} _ {n} \\|} \\tag {11}", + "image_path": "541eabd244d1541e709d09e4739b055a3c09a1908bf9c9e269de1f4c18459b33.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\Delta_{m}" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": " represents the aggregated " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": " by Eq. (8) for input " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": " denotes the class of " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "1_{y_m = y_n}" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": " is an indicator function that takes a value of 1 when the condition " + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "inline_equation", + "content": "y_{m} = y_{n}" + }, + { + "bbox": [ + 313, + 147, + 555, + 195 + ], + "type": "text", + "content": " holds, and 0 otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 202, + 399, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 399, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 399, + 215 + ], + "type": "text", + "content": "4.4. Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": "In addition to the previously mentioned loss functions in Eqs. (5) and (11), we also directly used the classification loss function " + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DR}}" + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": " specifically designed for ETF classifier [51, 52] to supervise our introduced plug-and-play branch, and the loss function " + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{z}}" + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": " for maintaining load balancing among the discretization patterns [59]. In summary, when integrating S6MOD on a base method, the overall loss function can be expressed as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 327, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 327, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 386, + 327, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {S 6 M O D}}, \\tag {12}", + "image_path": "0e122d49e49737e465a9430c0fbd9a15ce11eb3b88e3c00a19da90cbfcceea31.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 361, + 553, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 361, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 335, + 361, + 553, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S 6 M O D}} = \\mathcal {L} _ {\\mathrm {D R}} + \\alpha \\cdot \\mathcal {L} _ {\\text {D i f f}} + \\beta \\cdot \\mathcal {L} _ {\\text {C o n t}} + \\mathcal {L} _ {\\mathrm {z}}, \\tag {13}", + "image_path": "efc907751119e5f6dcab0d6f3ce74bacf316e0f113e0ce51405580ecad5a6c68.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 380, + 462, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 462, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 462, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 380, + 462, + 392 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 380, + 462, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 380, + 462, + 392 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 380, + 462, + 392 + ], + "type": "text", + "content": " are hyperparameters." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 404, + 395, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 395, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 395, + 417 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 424, + 435, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 435, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 435, + 437 + ], + "type": "text", + "content": "5.1. Experimental Setups" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": "Datasets. We test three datasets that are widely used in OCL, i.e., CIFAR-10 (10 classes) [20], CIFAR-100 (100 classes) [20] and TinyImageNet (200 classes) [21]. The dataset settings we adopted are the same as those in CCLDC[46]. We divide CIFAR-10 into 5 tasks, with 2 classes per task; CIFAR-100 into 10 tasks, with 10 classes per task; and TinyImageNet into 100 tasks, with 2 classes per task. Further details about the datasets will be provided in the supplementary materials." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 566, + 554, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 625 + ], + "type": "text", + "content": "Baselines. To demonstrate the effectiveness and applicability of our module, we conduct tests on 7 typical and state-of-the-art methods, including ER [34], OCM [17], OnPro [46], OCM-CCLDC [44], OnPro-CCLDC [44], MOSE[50] and MOE-MOSE [50]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Implementation details. To ensure a fair comparison, we applied the same hyperparameter settings to each baseline and the methods combined with our module (S6MOD). All the above methods use ResNet-18 as the backbone, without pretraining. For the streaming input data, we set the batch size to 10, and for the samples drawn from the buffer," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20506" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 98, + 502, + 282 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Average Accuracy (\\%, higher is better) on three benchmark datasets with difference memory buffer size " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", with and without our proposed S6MOD module. All values are averages of 10 runs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 98, + 502, + 282 + ], + "lines": [ + { + "bbox": [ + 108, + 98, + 502, + 282 + ], + "spans": [ + { + "bbox": [ + 108, + 98, + 502, + 282 + ], + "type": "table", + "html": "
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]56.68±1.8962.32±4.1324.47±0.7231.89±1.4539.41±1.8110.82±0.7919.16±1.4224.71±2.52
ER + Ours57.88±3.3065.80±2.1626.50±2.2334.55±1.6639.61±3.1610.94±1.4719.67±1.3625.62±1.73
OCM [17]68.19±1.7573.15±1.0528.02±0.7435.69±1.3642.22±1.0618.36±0.9526.74±1.0231.94±1.19
OCM + Ours70.57±1.1475.31±1.1032.44±1.3538.97±2.2845.49±1.5819.12±1.6027.20±1.8332.31±1.72
OnPro [46]70.47±2.1274.70±1.5127.22±0.7733.33±0.9341.59±1.3814.32±1.4021.13±2.1226.38±2.18
OnPro + Ours72.30±1.0875.14±0.9928.84±0.5937.57±0.8144.14±1.3617.27±0.6525.62±0.7929.20±1.00
OCM-CCLDC [44]74.14±0.8577.66±1.4635.00±1.1543.34±1.5151.43±1.3723.36±1.1833.17±0.9739.25±0.88
OCM-CCLDC + Ours74.45±1.2078.21±1.0336.02±1.7744.40±2.2652.53±0.2123.68±1.0233.59±0.7239.53±1.02
OnPro-CCLDC [44]74.49±2.1478.64±1.4234.76±1.1241.89±0.8250.01±0.8521.81±1.0232.00±0.7238.18±1.02
OnPro-CCLDC + Ours75.03±1.0379.51±0.6336.07±0.7643.92±0.9650.85±0.5222.03±0.8233.29±0.3538.51±0.91
MOSE [50]61.02±1.4770.74±1.1835.05±0.3445.06±0.3254.53±0.7818.23±0.7330.98±0.6338.71±0.44
MOSE + Ours63.74±1.5572.54±0.1435.43±0.6245.38±0.3154.75±0.5219.03±0.8332.03±0.8138.18±1.02
MOE-MOSE [50]62.54±1.5972.18±1.2937.32±0.3447.03±0.5755.62±0.7220.61±0.6932.52±0.3338.41±0.53
MOE-MOSE + Ours63.86±1.1973.16±0.5337.76±0.5147.25±0.4356.32±0.5321.03±0.7733.53±0.5140.11±0.26
", + "image_path": "9e6379da3c440f0fc14e92b480273ceff9c3428ecc589a4c5aa85156070a66b6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 105, + 316, + 500, + 499 + ], + "blocks": [ + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "text", + "content": "Table 2. Average Forgetting (%, lower is better) on three benchmark datasets with difference memory buffer size " + }, + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "text", + "content": ", with and without our proposed S6MOD module. All values are averages of 10 runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 316, + 500, + 499 + ], + "lines": [ + { + "bbox": [ + 105, + 316, + 500, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 500, + 499 + ], + "type": "table", + "html": "
DatasetCIFAR10CIFAR100Tiny-ImageNet
Memory Size M50010001000200050002000500010000
ER [34]33.16±3.5020.94±6.7932.65±1.7822.20±2.2613.29±1.9858.38±1.6946.87±1.6040.77±2.45
ER + Ours31.25±3.4416.71±2.8630.96±2.0719.02±2.1312.97±2.6158.03±2.1945.83±1.4937.62±1.53
OCM [17]13.68±4.2511.63±2.6214.99±1.559.16±1.753.76±1.1626.12±1.6319.74±1.3015.92±1.47
OCM + Ours13.58±3.269.68±3.4715.99±1.8211.60±1.495.46±0.9824.23±2.0719.12±1.8714.66±1.75
OnPro [46]17.94±3.6914.20±2.6016.76±2.4712.42±1.396.72±0.9428.01±1.5923.52±1.7520.32±1.70
OnPro + Ours7.13±1.445.73±1.8116.03±1.699.81±1.325.07±0.8920.92±1.1216.82±0.9718.15±1.91
OCM-CCLDC [44]11.59±2.249.18±2.0316.69±2.3610.07±1.373.99±0.7826.16±1.9019.99±1.9615.56±1.06
OCM-CCLDC + Ours15.81±2.6911.28±1.8316.64±1.859.08±2.063.48±0.3125.29±1.9514.89±0.7411.55±1.31
OnPro-CCLDC [44]19.89±4.0114.62±2.7528.93±2.1920.23±1.0310.55±1.8928.21±1.5820.85±1.1316.17±0.63
OnPro-CCLDC + Ours17.23±3.1611.51±2.3322.26±1.1812.49±1.594.88±1.3626.73±1.5116.69±0.3112.23±1.37
MOSE [50]30.36±1.6920.27±1.2737.54±0.4325.89±0.4513.60±0.5947.16±1.4124.96±0.6215.51±0.33
MOSE + Ours27.38±1.9417.92±0.3437.09±0.6325.80±0.3915.55±0.6146.60±1.0724.39±0.7220.15±0.74
MOE-MOSE [50]29.39±1.7919.24±1.4935.17±0.3023.99±0.5112.81±0.7441.98±1.4622.22±0.3813.94±0.59
MOE-MOSE + Ours28.13±1.4117.98±0.5634.76±0.6623.84±0.7714.03±0.4841.26±1.0221.39±0.5814.91±0.28
", + "image_path": "2bd1111f36b59465c0519f3b88d72b8a9bcb9c84f958d913e1509efed4087a4d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 558 + ], + "type": "text", + "content": "the batch size is set to 64. We retain each baseline's original data augmentation methods without modifying the base methods. For more details, please refer to the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 113, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 113, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 113, + 576 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "We combine our method with both classical and state-of-the-art approaches on CIFAR-10, CIFAR-100, and TinyImageNet. The experimental results in Table 1 demonstrate the universality of our method, as it consistently improves accuracy across numerous baselines. On CIFAR-10 and CIFAR-100, our method generally leads to a " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " improvement, while on TinyImageNet, we achieve significant gains under OCM, OnPro and MOE-MOSE. Notably, in settings like CIFAR100 " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "(M = 2k)" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " and TinyImageNet " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "(M = 5k)" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", OnPro result in an approximate " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " improvement, respectively. It is also worth mention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "content": "ing that even for state-of-the-art distillation-based methods like CCLDC and MOE-MOSE, the performance improvements are quite substantial. For instance, OnPro-CCLDC with the S6MOD integration achieve an impressive " + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "inline_equation", + "content": "79.51\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "content": " on CIFAR10 (" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "inline_equation", + "content": "M = 1k" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "content": "), and MOE-MOSE with S6MOD reach " + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "inline_equation", + "content": "56.32\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "content": " on CIFAR100 (" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "inline_equation", + "content": "M = 5k" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "inline_equation", + "content": "40.1\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "content": " on TinyImageNet (" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "inline_equation", + "content": "M = 10k" + }, + { + "bbox": [ + 313, + 509, + 555, + 593 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In addition, as shown in Table 2, our module is also highly effective in reducing model forgetting, achieving lower forgetting rates in most settings, with some cases showing particularly significant improvements. This can be attributed to our class-conditional routing. In a few baseline settings, the inclusion of our module led to a higher forgetting rate, but this does not imply that the model is more prone to forgetting. In these cases, we think the higher forgetting rates are often due to the stronger learning capabilities of the model. Analyzing both Table 1 and Table 2, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20507" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 129, + 266, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "Table 3. Ablation studies on CIFAR-100 (" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "M = 2k" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "). The \"branch\" refers to using SS2D[26] as the core of the S6MOD, without routing and " + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cont}}" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": ". \"routing\" represents class-conditional routing. S6MOD refers to the composition of (branch + class-conditional routing + " + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cont}}" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "). All values are averaged over 10 runs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 129, + 266, + 201 + ], + "lines": [ + { + "bbox": [ + 79, + 129, + 266, + 201 + ], + "spans": [ + { + "bbox": [ + 79, + 129, + 266, + 201 + ], + "type": "table", + "html": "
MethodAcc. ↑AF ↓
OnPro33.33±0.9312.42±1.39
OnPro + branch (with LDiff)36.61±1.0410.28±1.59
OnPro + branch + routing36.92±0.728.85±0.81
OnPro + branch + LCont36.93±0.849.70±1.42
OnPro + S6MOD37.57±0.819.81±1.32
", + "image_path": "c5c2c0261d5c7df568ac30198318d265eb04025187677bc700f730815bcb22f5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 231 + ], + "type": "text", + "content": "can find that even in these settings with higher forgetting rates, the accuracy performance remains quite competitive." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 157, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 157, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 157, + 251 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "text", + "content": "As mentioned in Sec. 1, the S6MOD module we propose is very simple and can be directly applied to existing baselines. By adding an extra branch and using the " + }, + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}" + }, + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "text", + "content": " to supervise the original classification head, we can achieve a significant improvement. To demonstrate this, we conduct an ablation study where we only add the extra branch and " + }, + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}" + }, + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "text", + "content": " to the baseline. The results in Table 3 confirm this, showing a " + }, + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 55, + 255, + 296, + 362 + ], + "type": "text", + "content": " improvement with supervision from the extra branch alone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "text", + "content": "In addition, to verify the effectiveness and adaptability of class-conditional routing and contrastive discretization loss " + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cont}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "text", + "content": ", we conduct separate ablation experiments. As shown in Table 3, when added to the base method with the extra branch, the improvements brought by adding class-conditional routing or " + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cont}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "text", + "content": " alone are not significant, with accuracy gains of only " + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "inline_equation", + "content": "0.31\\%" + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "inline_equation", + "content": "0.32\\%" + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "text", + "content": ", respectively. However, when both components are used together, the improvement becomes significant, with an approximate " + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 364, + 296, + 533 + ], + "type": "text", + "content": " increase in accuracy, and the forgetting rate still decreases. This indicates that when our module is correctly combined, it can significantly enhances the model's adaptability by learning more generalizable and discriminative features through dynamic parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 119, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 119, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 119, + 553 + ], + "type": "text", + "content": "5.4. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 715 + ], + "type": "text", + "content": "Analysis of Feature Embeddings. To verify that S6MOD improves the adaptability of the baseline and helps it learn more discriminative features, we use t-SNE [40] visualization for analysis. We compare with the baseline under the CIFAR-100 (" + }, + { + "bbox": [ + 55, + 557, + 297, + 715 + ], + "type": "inline_equation", + "content": "M = 2k" + }, + { + "bbox": [ + 55, + 557, + 297, + 715 + ], + "type": "text", + "content": ") setting. As shown in Fig. 2, we visualize the features of the samples in the buffer after the final training, which are the features " + }, + { + "bbox": [ + 55, + 557, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 557, + 297, + 715 + ], + "type": "text", + "content": " output by the backbone just before they are passed to the classifier for classification. The features shown in Fig. 2 (a) and (c) are very scattered, with only a few classes showing some degree of aggregation. Most classes do not have clear separations between other classes, making it difficult to distinguish. In contrast, Fig. 2 (b) and (d) present the feature distribution" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 339, + 72, + 432, + 163 + ], + "blocks": [ + { + "bbox": [ + 339, + 72, + 432, + 163 + ], + "lines": [ + { + "bbox": [ + 339, + 72, + 432, + 163 + ], + "spans": [ + { + "bbox": [ + 339, + 72, + 432, + 163 + ], + "type": "image", + "image_path": "dee5334ca91198dbfe3bf8014e387bbe9f54663822eedfd246385311b5230cce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 167, + 402, + 175 + ], + "lines": [ + { + "bbox": [ + 368, + 167, + 402, + 175 + ], + "spans": [ + { + "bbox": [ + 368, + 167, + 402, + 175 + ], + "type": "text", + "content": "(a) OnPro" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 531, + 163 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 531, + 163 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 531, + 163 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 531, + 163 + ], + "type": "image", + "image_path": "3242c70be735c2f67ee71d100c33f530f7965e3de772d801e7f595b8e646aaec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 166, + 519, + 175 + ], + "lines": [ + { + "bbox": [ + 449, + 166, + 519, + 175 + ], + "spans": [ + { + "bbox": [ + 449, + 166, + 519, + 175 + ], + "type": "text", + "content": "(b) OnPro + S6MOD" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 340, + 178, + 432, + 270 + ], + "blocks": [ + { + "bbox": [ + 340, + 178, + 432, + 270 + ], + "lines": [ + { + "bbox": [ + 340, + 178, + 432, + 270 + ], + "spans": [ + { + "bbox": [ + 340, + 178, + 432, + 270 + ], + "type": "image", + "image_path": "709d94a5ea9bf0fa553c1cb3994394fc1712f0bdbe1b9883ba92168e7fb8b260.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 274, + 400, + 284 + ], + "lines": [ + { + "bbox": [ + 369, + 274, + 400, + 284 + ], + "spans": [ + { + "bbox": [ + 369, + 274, + 400, + 284 + ], + "type": "text", + "content": "(c) OCM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 440, + 178, + 531, + 270 + ], + "blocks": [ + { + "bbox": [ + 440, + 178, + 531, + 270 + ], + "lines": [ + { + "bbox": [ + 440, + 178, + 531, + 270 + ], + "spans": [ + { + "bbox": [ + 440, + 178, + 531, + 270 + ], + "type": "image", + "image_path": "1a16046da30c40d9d32a88fdf34a4d7365e1a22ee49c631b58ec7bf23e6218f1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 274, + 518, + 283 + ], + "lines": [ + { + "bbox": [ + 451, + 274, + 518, + 283 + ], + "spans": [ + { + "bbox": [ + 451, + 274, + 518, + 283 + ], + "type": "text", + "content": "(d) OCM + S6MOD" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 320, + 338, + 422, + 443 + ], + "blocks": [ + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "type": "text", + "content": "Figure 2. t-SNE visualization of memory data at the end of training on CIFAR-100 (" + }, + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "type": "inline_equation", + "content": "M = 2k" + }, + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "type": "text", + "content": "), showcasing baseline in (a) and (c) and baseline combined with S6MOD in (b) and (d). Different colors represent different classes." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 338, + 422, + 443 + ], + "lines": [ + { + "bbox": [ + 320, + 338, + 422, + 443 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 422, + 443 + ], + "type": "image", + "image_path": "eed097c6eb67dfc27acafa55c89044fa14f762eef4a279430749d907bd6dedb3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 447, + 414, + 456 + ], + "lines": [ + { + "bbox": [ + 336, + 447, + 414, + 456 + ], + "spans": [ + { + "bbox": [ + 336, + 447, + 414, + 456 + ], + "type": "text", + "content": "(a) New-Task Accuracy" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "text", + "content": "Figure 3. Impact of dynamically selecting different " + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "text", + "content": " values on the ability to learn new tasks and prevent forgetting of old tasks: New-Task accuracy represents the model's accuracy on the current task. We conduct experiments by setting " + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "inline_equation", + "content": "N_{k} = 1" + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "inline_equation", + "content": "N_{k} = N" + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "text", + "content": ", and calculating " + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "text", + "content": " through class-conditional routing. The dataset used is CIFAR-100 (" + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "inline_equation", + "content": "M = 2k" + }, + { + "bbox": [ + 313, + 460, + 555, + 525 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 437, + 344, + 541, + 443 + ], + "blocks": [ + { + "bbox": [ + 437, + 344, + 541, + 443 + ], + "lines": [ + { + "bbox": [ + 437, + 344, + 541, + 443 + ], + "spans": [ + { + "bbox": [ + 437, + 344, + 541, + 443 + ], + "type": "image", + "image_path": "af18e09dfca09cb93f91eeadd68afffb784431d74cc3427634b9938df0e4d7f1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 447, + 531, + 456 + ], + "lines": [ + { + "bbox": [ + 456, + 447, + 531, + 456 + ], + "spans": [ + { + "bbox": [ + 456, + 447, + 531, + 456 + ], + "type": "text", + "content": "(b) Average Forgetting" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": "after adding our module. We can observe that many categories have noticeably converged, forming small clusters with consistent intra-class compactness and clear inter-class separations. These features are much easier to distinguish and more conducive to classification by the classifier, which aligns with the significant improvements we demonstrate in the experiments." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Analysis of Class-conditional Routing. We further investigate the effectiveness of class-conditional routing. Under the CIFAR-100 (" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "M = 2k" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ") setting, we design experiments to compare the impact of a fixed " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " and dynamically calculated " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " using class-conditional routing. Specifically, we set " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{k} = N" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", meaning all patterns are activated, and " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{k} = 1" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", meaning only the pattern with the highest prob" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20508" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 120, + 294, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": "Table 4. Impact of hyper-parameters " + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": " on performance in the CIFAR-100 dataset (" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "inline_equation", + "content": "M = 2k" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": " are the weights for " + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cont}}" + }, + { + "bbox": [ + 55, + 71, + 295, + 115 + ], + "type": "text", + "content": ", respectively. Acc. and AF represent the average accuracy and the average forgetting over 5 runs, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 120, + 294, + 235 + ], + "lines": [ + { + "bbox": [ + 58, + 120, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 58, + 120, + 294, + 235 + ], + "type": "table", + "html": "
Impact of hyper-parameter α
Value0.010.10.515
Acc.36.80±1.1737.09±1.0937.55±1.0837.47±0.9437.57±0.81
AF10.49±1.019.65±0.399.05±1.259.16±2.119.81±1.32
Impact of hyper-parameter β
Value0.1151025
Acc.36.50±1.2437.30±1.0337.40±0.3337.26±0.6137.57±0.81
AF10.24±0.4310.12±1.0910.10±0.609.98±0.669.81±1.32
", + "image_path": "d881aab3f0270682ff6073d3b905e972b9c605e67a5cb95129438cea13ad3041.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": "ability is activated. To facilitate analysis, we present the accuracy of the model on the latest task (the current task) in Fig. 3 (a). As shown in Fig. 3 (a), when " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k} = N" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": ", all patterns are activated, and thanks to the inclusion of more patterns with different weights and larger parameters, the model has greater capacity to adapt to new tasks, resulting in a stronger ability to learn new knowledge. When " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k} = 1" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": ", only the pattern with the highest probability is activated, significantly reducing the model's ability to learn the new task. However, as shown in Fig. 3 (b), when " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k} = N" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": ", since all patterns are updated by the current task, the model has a higher risk of forgetting. When " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k} = 1" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": ", only the pattern with the highest probability is updated, and tasks do not interfere with each other, so forgetting is less likely. Combining the analysis of the entire Fig. 3, the effect of our proposed class-conditional routing is both intuitive and significant. It not only helps the model achieve strong learning ability for new tasks, comparable to the " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": " setting, but also provides excellent anti-forgetting performance, similar to " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k} = 1" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": ". This is due to the dynamic adjustment of " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": " based on uncertainty: for samples that are prone to misclassification, class-conditional routing assigns a larger " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": ", enabling more extensive learning. For easier-to-classify samples, the model's learning ability is higher, and class-conditional routing assigns a smaller " + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 243, + 296, + 543 + ], + "type": "text", + "content": " to enhance anti-forgetting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "content": "Sensitivity Analysis of Hyper-parameters " + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "content": " are the weights for " + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}" + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cont}}" + }, + { + "bbox": [ + 55, + 555, + 296, + 665 + ], + "type": "text", + "content": ", respectively. We conduct experiments on the sensitivity of them, and the results are shown in Table 4. It can be observed that when the weights are low, the optimization effect of the loss is relatively weak, both accuracy and forgetting rate do not perform well. Gradually increasing the weights, we find that after a certain point, the impact of the weights on the results becomes relatively stable, with no more drastic changes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Sensitivity Analysis of total discretization patterns number " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". We conduct experiments on the total number of discretization patterns in the module. The results are shown" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 79, + 424, + 169 + ], + "blocks": [ + { + "bbox": [ + 320, + 79, + 424, + 169 + ], + "lines": [ + { + "bbox": [ + 320, + 79, + 424, + 169 + ], + "spans": [ + { + "bbox": [ + 320, + 79, + 424, + 169 + ], + "type": "image", + "image_path": "2f880a619373a0e7ad688d11998bc6e7c6870626c6df92de9fa2343e03a5add4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 171, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 315, + 171, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 315, + 171, + 553, + 192 + ], + "type": "text", + "content": "(a) New-Task Average Accuracy (b) Average Forgetting w/o class-w/o class-conditional routing conditional routing" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 437, + 79, + 542, + 170 + ], + "blocks": [ + { + "bbox": [ + 437, + 79, + 542, + 170 + ], + "lines": [ + { + "bbox": [ + 437, + 79, + 542, + 170 + ], + "spans": [ + { + "bbox": [ + 437, + 79, + 542, + 170 + ], + "type": "image", + "image_path": "bb54119721d252a830c3c0a1d8ad2d5d16dd42b102922c8d56f9c5fc62e1d87c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 198, + 425, + 289 + ], + "blocks": [ + { + "bbox": [ + 320, + 198, + 425, + 289 + ], + "lines": [ + { + "bbox": [ + 320, + 198, + 425, + 289 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 425, + 289 + ], + "type": "image", + "image_path": "1e14ddd1b7dc09cbadec1ced7e49b5f258b1ef9e7751088c28be86dc5b04a503.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 291, + 428, + 300 + ], + "lines": [ + { + "bbox": [ + 321, + 291, + 428, + 300 + ], + "spans": [ + { + "bbox": [ + 321, + 291, + 428, + 300 + ], + "type": "text", + "content": "(c) New-Task Average Accuracy" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 437, + 198, + 543, + 289 + ], + "blocks": [ + { + "bbox": [ + 437, + 198, + 543, + 289 + ], + "lines": [ + { + "bbox": [ + 437, + 198, + 543, + 289 + ], + "spans": [ + { + "bbox": [ + 437, + 198, + 543, + 289 + ], + "type": "image", + "image_path": "8fbee29e319e85a8d06830280eae17ac9f6c1dfce5abbd379d95c25f1343a3af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 291, + 531, + 300 + ], + "lines": [ + { + "bbox": [ + 457, + 291, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 457, + 291, + 531, + 300 + ], + "type": "text", + "content": "(d) Average Forgetting" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 305, + 553, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 305, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 553, + 327 + ], + "type": "text", + "content": "Figure 4. Impact of pattern number " + }, + { + "bbox": [ + 313, + 305, + 553, + 327 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 305, + 553, + 327 + ], + "type": "text", + "content": ". The dataset used is CIFAR-100 (" + }, + { + "bbox": [ + 313, + 305, + 553, + 327 + ], + "type": "inline_equation", + "content": "M = 2k" + }, + { + "bbox": [ + 313, + 305, + 553, + 327 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 336, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 445 + ], + "type": "text", + "content": "in Fig. 4. As the number " + }, + { + "bbox": [ + 313, + 336, + 555, + 445 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 336, + 555, + 445 + ], + "type": "text", + "content": " increases, the model gains a stronger ability to learn new tasks, but at the same time, its tendency to forget becomes more pronounced. Furthermore, by comparing (a)(b) and (c)(d) in Fig. 4, we can conclude that class-conditional routing can significantly reduce forgetting with only a slight decrease in new task accuracy, which in turn leads to an overall improvement in performance. This indirectly supports the effectiveness of class-conditional routing in our ablation studies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 454, + 388, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 388, + 467 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 388, + 467 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 475, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 606 + ], + "type": "text", + "content": "In this paper, we propose a plug-and-play module S6MOD to enhance the adaptability of existing OCL methods by introducing selective state space models (SSMs) with a class-conditional mixture of discretization. By integrating a dynamic discretization mechanism and leveraging class-conditional strategies, our method can efficiently allocate discretization patterns based on class uncertainty, improving both the model's generalization and its ability to adapt to new data. Experimental results on multiple OCL datasets demonstrate that our method consistently optimize existing techniques, contributing to more robust OCL systems." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 424, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 424, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 424, + 630 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 703 + ], + "type": "text", + "content": "The research reported in this publication was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940. This research used computational resources provided by IBEX managed by the Supercomputing Core Laboratory at KAUST." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20509" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Assyr Abdulle, E Weinan, Björn Engquist, and Eric Vanden-Eijnden. The heterogeneous multiscale method. Acta Numerica, 21:1-87, 2012. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 126, + 294, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 126, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 294, + 169 + ], + "type": "text", + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 172, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 295, + 215 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. NeurIPS, 32, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 218, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 294, + 251 + ], + "type": "text", + "content": "[4] Quentin Anthony, Yury Tokpanov, Paolo Glorioso, and Beren Millidge. Blackmamba: Mixture of experts for statespace models. arXiv preprint arXiv:2402.01771, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 254, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 294, + 306 + ], + "type": "text", + "content": "[5] Fengxiang Bie, Yibo Yang, Zhongzhu Zhou, Adam Ghanem, Minjia Zhang, Zhewei Yao, Xiaoxia Wu, Connor Holmes, Pareesa Golnari, David A Clifton, et al. Renaissance: A survey into ai text-to-image generation in the era of large model. TPAMI, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "type": "text", + "content": "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. *NeurIPS*, 33: 15920-15930, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 294, + 399 + ], + "type": "text", + "content": "[7] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "text", + "content": "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 294, + 470 + ], + "type": "text", + "content": "[9] Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 472, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 472, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 57, + 472, + 294, + 525 + ], + "type": "text", + "content": "[10] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 44(7):3366-3385, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 529, + 294, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 529, + 294, + 562 + ], + "spans": [ + { + "bbox": [ + 57, + 529, + 294, + 562 + ], + "type": "text", + "content": "[11] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 565, + 294, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 565, + 294, + 607 + ], + "spans": [ + { + "bbox": [ + 57, + 565, + 294, + 607 + ], + "type": "text", + "content": "[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 610, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 610, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 610, + 294, + 643 + ], + "type": "text", + "content": "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "text", + "content": "[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[15] John Guckenheimer and Philip Holmes. Nonlinear oscillations, dynamical systems, and bifurcations of vector fields. Springer Science & Business Media, 2013. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Nuwan Gunasekara, Bernhard Pfahringer, Heitor Murilo Gomes, and Albert Bifet. Survey on online streaming continual learning. In *IJCAI*, pages 6628-6637, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "text", + "content": "[17] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In ICML, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[18] Yiduo Guo, Bing Liu, and Dongyan Zhao. Dealing with cross-task class discrimination in online continual learning. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "text", + "content": "[19] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. Neural Computation, 3(1):79-87, 1991. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 209, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 229 + ], + "type": "text", + "content": "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 232, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 253 + ], + "type": "text", + "content": "[21] Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "type": "text", + "content": "[22] Xiaojie Li, Yibo Yang, Jianlong Wu, Bernard Ghanem, Liqiang Nie, and Min Zhang. Mamba-fscil: Dynamic adaptation with selective state space model for few-shot class-incremental learning. arXiv preprint arXiv:2407.06136, 2024. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 311, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 553, + 331 + ], + "type": "text", + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 40(12):2935-2947, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 334, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 334, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 334, + 553, + 387 + ], + "type": "text", + "content": "[24] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 389, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 553, + 443 + ], + "type": "text", + "content": "[25] Sihao Liu, Mingyi Liu, Tianyu Jiang, Shuang Yu, Hanchuan Xu, and Zhongjie Wang. GSL-mash: Enhancing mashup creation service recommendations through graph structure learning. In Service-Oriented Computing, pages 176-191, Singapore, 2025. Springer Nature Singapore. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. ArXiv, abs/2401.10166, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 490, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 520 + ], + "type": "text", + "content": "[27] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "text", + "content": "[28] Zheda Mai, Ruiwen Li, Hyunwoo Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. In CVPR, pages 3589-3599, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 569, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 610 + ], + "type": "text", + "content": "[29] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "type": "text", + "content": "[30] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[31] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[32] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, Michał Krutul, Jakub Krajewski, Szymon Antoniak, Pi" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20510" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "otr Mioś, Marek Cygan, and Sebastian Jaszcur. Moemamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 148 + ], + "type": "text", + "content": "[33] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. *NeurIPS*, 32, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "text", + "content": "[35] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[36] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "text", + "content": "[37] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In CVPR, pages 23933-23942, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[38] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, pages 4548-4557. PMLR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "text", + "content": "[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 281, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 281, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 281, + 448 + ], + "type": "text", + "content": "[41] A Vaswani. Attention is all you need. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[42] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. NeurIPS, 34:22379-22391, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "text", + "content": "[43] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. TPAMI, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "text", + "content": "[44] Maorong Wang, Nicolas Michel, Ling Xiao, and Toshihiko Yamasaki. Improving plasticity in online continual learning via collaborative learning. In CVPR, pages 23460-23469, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[45] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, pages 18764-18774, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 604, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 635 + ], + "type": "text", + "content": "[46] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In ICCV, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 637, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 678 + ], + "type": "text", + "content": "[47] Mitchell Wortsman, Vivek Ramanujan, Rosanne Liu, Aniruddha Kembhavi, Mohammad Rastegari, Jason Yosinski, and Ali Farhadi. Supermasks in superposition. NeurIPS, 33:15173-15184, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[48] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 586 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[49] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[50] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In CVPR, pages 23670-23680, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[51] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, pages 37991-38002. Curran Associates, Inc., 2022. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "text", + "content": "[52] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR, 2023. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "text", + "content": "[53] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. Corda: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In NeurlPS, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "text", + "content": "[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 375, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 418 + ], + "type": "text", + "content": "[55] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 420, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 452 + ], + "type": "text", + "content": "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995. PMLR, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "text", + "content": "[57] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19560, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "text", + "content": "[58] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "type": "text", + "content": "[59] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20511" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_content_list.json b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69b17ccdc6200931801b579bf0f34bf496e89053 --- /dev/null +++ b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_content_list.json @@ -0,0 +1,1550 @@ +[ + { + "type": "text", + "text": "Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models", + "text_level": 1, + "bbox": [ + 116, + 128, + 880, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Chen $^{1}$ Daochang Liu $^{2}$ Mubarak Shah $^{3}$ Chang Xu $^{1}$", + "bbox": [ + 245, + 202, + 746, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Faculty of Engineering, The University of Sydney, Australia $^{2}$ School of Physics, Mathematics and Computing, The University of Western Australia, Australia $^{3}$ Center for Research in Computer Vision, University of Central Florida, USA", + "bbox": [ + 114, + 220, + 880, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cche0711@uni., c.xu@}sydney.edu.au daochang.liu@uwa.edu.au shah@crcv.ucf.edu", + "bbox": [ + 153, + 276, + 836, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models have demonstrated remarkable capabilities in creating images highly aligned with user prompts, yet their proclivity for memorizing training set images has sparked concerns about the originality of the generated images and privacy issues, potentially leading to legal complications for both model owners and users, particularly when the memorized images contain proprietary content. Although methods to mitigate these issues have been suggested, enhancing privacy often results in a significant decrease in the utility of the outputs, as indicated by text-alignment scores. To bridge the research gap, we introduce a novel method, PRSS, which refines the classifier-free guidance approach in diffusion models by integrating prompt re-anchoring (PR) to improve privacy and incorporating semantic prompt search (SS) to enhance utility. Extensive experiments across various privacy levels demonstrate that our approach consistently improves the privacy-utility trade-off, establishing a new state-of-the-art.", + "bbox": [ + 88, + 358, + 485, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 636, + 220, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leveraging classifier-free guidance (CFG) [12], text-to-image diffusion models like Stable Diffusion [28] and Midjourney [1] are now capable of generating highly realistic images that closely align with user-provided text prompts. This capability has propelled their popularity, leading to widespread use and distribution of their generated images. However, recent research [6, 8, 31, 34] has revealed a critical issue: these models can memorize training data, leading them to reproduce parts of images, such as foregrounds or backgrounds (local memorization, see Fig. 1), or even entire images (global memorization, see Fig. 2) during inference, instead of generating genuinely novel content. When the training data includes sensitive or copyrighted material, these memorization issues can infringe on copyright laws without notifying either the model's owners or users or the copyright holders of the replicated content. The risk is es", + "bbox": [ + 89, + 657, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ppecially substantial given the extensive use of these models and their reliance on massive datasets, such as LAION [30], which contain billions of web-scale images and are impractical to thoroughly review or filter manually. This risk is further highlighted by real-world cases, where several artists have filed lawsuits, arguing that models like Stable Diffusion act as “21st-century collage tools” that remix their copyrighted works, implicating Stability AI, DeviantArt, Midjourney, and Runway AI. Recognizing the potential for unauthorized reproductions, Midjourney has even banned prompts containing the term “Afghan” to prevent the generation of the copyrighted Afghan Girl photograph. Yet, as [35] demonstrates, such restrictions alone are insufficient to fully prevent the reproduction of copyrighted images. This underscores the urgent need for timely effective mitigation strategies to address these concerns.", + "bbox": [ + 511, + 327, + 906, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to these legal challenges, recent initiatives [7, 8, 27, 32, 36] have focused on developing strategies to minimize memorization, achieving notable success. These approaches vary in scope, with some targeting the model's training phase and others making adjustments during inference. While training-phase strategies are theoretically viable, they lack practical significance as they predominantly experiment on a small $10\\mathrm{k}$ subset of the entire LAION5B dataset due to the impracticality of fine-tuning on the full dataset — a constraint stemming from computational limitations. In contrast, inference-phase strategies, which involve modifications only during inference, offer computational efficiency and adaptability, as they require no additional training or fine-tuning. This makes them highly applicable across pre-trained diffusion models, regardless of specific training configurations. Despite their promise, inference-time strategies can also suffer from computational constraints. For example, the approach proposed by [7] involves searching the entire LAION dataset for the nearest neighbor to optimize the generated latent representation during each denoising step. Other inference-time strategies [8, 27, 32, 36] that avoid nearest neighbor search", + "bbox": [ + 511, + 569, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8182", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2733b0e4ece0630b5c245b132c36f86a5ccb6afcbf18f36a0b53bae8f1da8e28.jpg", + "image_caption": [ + "Figure 1. Examples of Local Memorization: Stable Diffusion generations (columns 2-4) can replicate local regions of training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7)." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6fcaa2b5f9d426a1352a61170f978de7addc41e82e072566a3906e17f2493652.jpg", + "image_caption": [ + "Figure 2. Examples of Global Memorization: Stable Diffusion generations (columns 2-4) can replicate the entire training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7)." + ], + "image_footnote": [], + "bbox": [ + 94, + 347, + 480, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrate high efficiency but only moderate effectiveness, leaving a substantial gap in optimizing the privacy-utility trade-off. In particular, these methods often require significant compromises in text-image alignment to reduce memorization and conversely, struggle to sufficiently suppress memorization when trying to preserve alignment.", + "bbox": [ + 89, + 611, + 482, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we carefully analyze why existing strategies struggle to achieve an optimal privacy-utility trade-off, uncovering critical insights that the two inputs (text-conditional and unconditional predictions) used in the CFG framework are both suboptimal for mitigating memorization, even though they are standard for most other tasks in Stable Diffusion models. Specifically, we find that the text-conditional term in CFG, when optimized for privacy by prompt engineering, can significantly reduce text alignment, indicating the need for alternative methods to find privacy-preserving yet intent-aligned text-conditioning. Additionally, we observe that the unconditional term in CFG is less effective as an anchor for privacy", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "preservation, suggesting that alternative anchoring points that enhance privacy are necessary. Building on these insights, we introduce PRSS, a novel method that incorporates two complementary strategies to refine the widely used CFG approach for improved memorization mitigation. Our method includes (1) prompt re-anchoring, designed to preserve high privacy with a smaller utility cost, which re-anchors the unconditional component in CFG using the memorized prompt to consistently guide generations away from memorization during the inference process; (2) semantic prompt search, designed to preserve high utility with minimal privacy compromise, which optimizes the prompt away from the memorized one in the language space while keeping similar semantics. Our experimental results reveal that employing both strategies synergistically can consistently outperform baseline methods in achieving superior privacy-utility trade-offs at different privacy levels. In particular, this results in higher text alignment (utility) and substantially fewer memorized generations (improved privacy). Furthermore, ablation studies validate the efficacy of each component within our strategy. From an implementation standpoint, our strategies are simple, efficient, and effective, necessitating only adjustments to the existing CFG equation for updating the noise prediction during inference. This integration is seamless, avoiding disruption to the inference workflow and eliminating the need for additional re-training, fine-tuning, or searching over the training data.", + "bbox": [ + 511, + 90, + 906, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 511, + 651, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Classifier-Free Guidance in Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 536, + 895, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [13, 23] are popular generative models that consist of a forward diffusion process that consistently adds Gaussian noise to real images $x_0$ for $T$ steps based on the following conditional function, resulting in $x_T \\sim \\mathcal{N}(0, \\mathbf{I})$ :", + "bbox": [ + 511, + 558, + 905, + 633 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 643, + 903, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $x_{t}$ represents the noised version of $x_{0}$ after adding $t$ noise, and $\\{\\beta_t\\}_{t = 1}^T$ is the schedule that controls the magnitude of added noise at time $t$ .", + "bbox": [ + 511, + 671, + 905, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This is followed by the reverse process, in which a de-noiser model is trained to predict the distribution (i.e., mean $\\mu_{\\theta}(x_t)$ and variance $\\sigma_{\\theta}^{2}(x_{t})\\mathbf{I}$ ) of a cleaner version of $x_{t}$ , denoted as $x_{t - 1}$ , at any given timestep $t$ :", + "bbox": [ + 511, + 717, + 905, + 777 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}\\right), \\sigma_ {\\theta} ^ {2} \\left(x _ {t}\\right) \\mathbf {I}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 787, + 903, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During training, the objective function is equivalent to have the denoiser $\\epsilon_{\\theta}(x_t,y,t)$ predict the noise $\\epsilon_{t}$ to be removed, rather than the image $x_{t - 1}$ at timestep $t$ conditioned on $y$ which may represent a user-provided text prompt:", + "bbox": [ + 511, + 814, + 905, + 875 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {t \\in [ 1, T ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (x _ {t}, y, t) \\| _ {2} ^ {2} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 883, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8183", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classifier-free guidance (CFG) [12] is a widely used technique for balancing fidelity and diversity in the generated outputs of diffusion models, and it can be easily applied during inference. CFG operates by linearly guiding the noise prediction from the unconditional prediction, $\\epsilon_{\\theta}(x_t,t)$ towards the conditional prediction, $\\epsilon_{\\theta}(x_t,y,t)$ , with a scaling factor $s > 1$ to control the degree of guidance. For simplicity, removing the timestep $t$ from the notation, the adjusted noise prediction can be defined as:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}\\right) + s \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, y\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 234, + 482, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Memorization in Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 257, + 390, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition and Metrics. A widely adopted practical definition of memorization refers to generated images exhibiting high object-level similarity to training data [31]. This includes the concept of reconstructive memory in diffusion models, where foreground and background objects from training images are combined, potentially with modifications such as shifting, scaling, or cropping. These reconstructions may not match any training image pixel-for-pixel but still closely resemble specific training images at the object level. The standard metric for evaluating memorization [7, 8, 27, 31, 32, 36] is designed base on a pre-trained Self-Supervised Copy Detection (SSCD) model [25]:", + "bbox": [ + 89, + 272, + 483, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S S C D} (\\hat {x}) = \\max _ {x _ {i} \\in \\mathcal {D}} \\operatorname {C o s i n e} \\left(\\mathbf {E} _ {S S C D} (\\hat {x}), \\mathbf {E} _ {S S C D} (x _ {i})\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 457, + 482, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where for each generated image $\\hat{x}$ , cosine similarity is computed between its SSCD embedding $\\mathbf{E}_{SSCD}(\\hat{x})$ and the embeddings $\\mathbf{E}_{SSCD}(x_i)$ of all training images $x_i$ in the dataset $\\mathcal{D}$ . The maximum similarity serves as the final SSCD score.", + "bbox": [ + 89, + 479, + 482, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most recently, [8] proposed a method to localize memorized regions in generated images, arguing that only memorized regions pose privacy risks, making local memorization a more meaningful concept. They introduced an efficient approach to synthesize a mask that highlights these regions without requiring access to training data. Using this mask, they proposed a new localized metric based on both object-level SSCD and pixel-level Euclidean distance:", + "bbox": [ + 89, + 540, + 483, + 660 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL S (\\hat {x}, x) = - \\mathbb {1} _ {\\mathrm {S S C D} > 0. 5} \\cdot \\| (\\hat {x} - x) \\circ \\mathbf {m} \\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 667, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{m}$ is the local memorization mask and $\\mathbb{1}_{\\mathrm{SSCD} > 0.5}$ is an indicator function that signifies whether the SSCD score exceeds the 0.5 threshold, which is commonly used to distinguish memorized from non-memorized content.", + "bbox": [ + 89, + 691, + 482, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Memorization Detection. A recent work [36] introduces an innovative approach for detecting potential memorized generations during inference by leveraging the magnitude of the difference between text-conditional $\\epsilon_{\\theta}(x_t, e_p)$ and unconditional $\\epsilon_{\\theta}(x_t, e_\\phi)$ noise predictions:", + "bbox": [ + 89, + 752, + 483, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm _ {t} = \\left\\| \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) \\right\\| _ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 835, + 482, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where memorized generations consistently exhibit larger magnitudes than non-memorized ones, making this magnitude a strong signal for detection. Furthermore, accurate", + "bbox": [ + 89, + 856, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a3778257d9c64f428098939618dc4033c07215cc89161ab6972db3949f644b30.jpg", + "image_caption": [ + "Figure 3. Illustration of the logic behind how $\\lambda$ in Eq. (10) governs the privacy-utility trade-off. Currently, the privacy-utility trade-off is governed solely by how extensively the prompt is modified." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 905, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "detection can be achieved efficiently by evaluating the magnitude $m_{T-1}$ at the first denoising step (timestep $T - 1$ , highlighting the method's computational efficiency.", + "bbox": [ + 511, + 189, + 903, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contemporary work [8] extends this detection method by incorporating their proposed local memorization mask $\\mathbf{m}$ into Eq. (7) to compute a masked magnitude, achieving state-of-the-art performance:", + "bbox": [ + 511, + 234, + 903, + 295 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm _ {t} ^ {\\prime} = \\left\\| \\left(\\varepsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\varepsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\circ \\mathbf {m} \\right\\| _ {2} / \\left(\\frac {1}{N} \\sum_ {i = 1} ^ {N} m _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 304, + 903, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{m}$ is the local memorization mask and $\\frac{1}{N} \\sum_{i=1}^{N} m_i$ is used to normalize the result by the mean of $\\mathbf{m}$ .", + "bbox": [ + 511, + 357, + 903, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 400, + 604, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin with a detailed analysis in Sec. 3.1 to uncover the key challenges that hinder existing strategies from achieving an optimal privacy-utility trade-off. Building on these insights, we evolve the baseline step by step with our proposed prompt re-anchoring (PR) strategy in Sec 3.2, which focuses on improving privacy, and semantic prompt search (SS) in Sec 3.3, which enhances utility. Finally, in Sec. 3.4, we analyze how the combination of PR and SS achieves a synergistic effect, offering the best privacy-utility trade-off.", + "bbox": [ + 511, + 426, + 905, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Trade-off Analysis of Existing Strategies", + "text_level": 1, + "bbox": [ + 511, + 571, + 859, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is built upon the widely-used classifier-free guidance (CFG) by treating the input text prompt from user $e_p$ as the condition $y$ in Eq. (4) and rewriting the unconditional noise as a null condition $e_\\phi$ ,", + "bbox": [ + 511, + 593, + 905, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 664, + 903, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The one-step prompt engineering (PE) based memorization mitigation method proposed by [36] then counteract memorization by reformulating Eq. (9) as:", + "bbox": [ + 511, + 691, + 905, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\{m _ {T - 1} < \\lambda \\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e ^ {*}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} > \\lambda \\right\\}}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 744, + 903, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $m_{T - 1} = \\| \\epsilon_{\\theta}(x_{T - 1},e_p) - \\epsilon_{\\theta}(x_{T - 1},e_\\phi)\\| _2$ represents the magnitude at the initial step $T - 1$ as per Eq. (7), $\\epsilon_{\\theta}(x_t,e_p)$ and $\\epsilon_{\\theta}(x_{t},e^{*})$ are the noise predictions conditioned on the user prompt embedding $e_p$ and engineered prompt embedding $e^*$ (optimized based on the gradient computed from the first-step magnitude $m_{T - 1}$ ), and", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8184", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0212fde5e1de8ac4ebb3f459c6cbafb2045d8f900b37c2573da45a72a39af262.jpg", + "image_caption": [ + "a) Privacy-Utility Trade-Off in Baseline" + ], + "image_footnote": [], + "bbox": [ + 94, + 112, + 338, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d36f7bb43ddcb6b429d029441996088fcf5d52e0f72545ec8f86abda62de1e9d.jpg", + "image_caption": [ + "b) Enhancing Privacy with Prompt Re-Anchoring (PR)", + "$\\mathbf{SIM}(\\mathbb{O},\\mathbb{B}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{C}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{D})$" + ], + "image_footnote": [], + "bbox": [ + 364, + 114, + 609, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69ef1818c0a04ced470b7bb6b8cd221b10dfd10a64c16d03cb3dad59a6c838c1.jpg", + "image_caption": [ + "c) Enhancing Utility with Semantic Search (SS)", + "$\\mathrm{SIM}(\\mathbb{O},\\mathbb{C}) < \\mathrm{SIM}(\\mathbb{O},\\mathbb{C}^{\\prime})$", + "$\\mathbf{SIM}(\\mathbb{O},\\mathbb{D}) < \\mathbf{SIM}(\\mathbb{O},\\mathbb{D}^{\\prime})$", + "SIM CLIP Similarity (Utility)" + ], + "image_footnote": [], + "bbox": [ + 630, + 114, + 903, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96271cda170299881d18afaffb0f27f525f9fa0b02cffc9cfd0d9563eb17ebc9.jpg", + "image_caption": [ + "Contour Line of Equal Privacy" + ], + "image_footnote": [], + "bbox": [ + 107, + 277, + 127, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2dc232078a2a1003ca08f48f3726f87b2632c37018b5000d69d8ab2879e998c7.jpg", + "image_caption": [ + "$\\mathbf{SIM}(\\mathbb{O},\\mathbb{B}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{C}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{D})$", + "$\\rightarrow$ Baseline Guidance" + ], + "image_footnote": [], + "bbox": [ + 300, + 281, + 316, + 291 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab9957472abcddd87421923a6345bf50c88af6af601cdbb38883cef9d3ba5c91.jpg", + "image_caption": [ + "Our PR Guidance" + ], + "image_footnote": [], + "bbox": [ + 460, + 281, + 468, + 291 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/177949909573a9f6dd18342974a45c2068fb5e2b034d16c2b039580587515fd5.jpg", + "image_caption": [ + "Our PRSS Guidance" + ], + "image_footnote": [], + "bbox": [ + 622, + 281, + 630, + 291 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d743170ccf694ad07aa47bf9ba3e381adfb09ed86a99bf87f302f9d3f2b8cddc.jpg", + "image_caption": [ + "Unconditional Prediction" + ], + "image_footnote": [], + "bbox": [ + 109, + 308, + 122, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb41483b64ef842a229262c827e6eea42cad7c15d5a22a22637b115eb02aae4e.jpg", + "image_caption": [ + "$e_p$ -conditional prediction" + ], + "image_footnote": [], + "bbox": [ + 292, + 309, + 303, + 318 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09116e76c0a1a6aaf5a0c28da69766ee7c6a01ac6602a11296178d6a38e2e370.jpg", + "image_caption": [ + "$e^*$ -conditional prediction with different privacy level" + ], + "image_footnote": [], + "bbox": [ + 478, + 305, + 503, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1ddcd0dc93036430c7ca620bcf16651170f2806ab61fc35b92f4abd6a63e939.jpg", + "image_caption": [ + "$e_p^{ss}$ -conditional prediction with different privacy level", + "Figure 4. Illustration of our proposed strategy (PRSS) compared to the baseline prompt engineering (PE) method. a) The privacy-utility trade-off in the baseline method, where the privacy increases at the cost of utility from B to C to D. b) Our proposed PR guidance, i.e., the green arrow, drives the image out of the low-privacy region more efficiently than the baseline guidance, i.e., the black arrows. c) Our proposed SS can enhance utility, by finding points $\\mathrm{B}^{\\prime}\\mathrm{C}^{\\prime}\\mathrm{D}^{\\prime}$ that are at the same level of privacy as BCD but with better utility. In summary, using the proposed PR and SS jointly can enhance the privacy-utility trade-off to mitigate memorization in diffusion models." + ], + "image_footnote": [], + "bbox": [ + 691, + 305, + 736, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\epsilon_{\\theta}(x_t, e_{\\phi})$ is the unconditional noise prediction. This strategy detects potential memorization for a given input prompt based on the magnitude signal $m_{T-1}$ with a threshold $\\lambda$ , and undertakes soft prompt engineering (i.e., optimizes the prompt embeddings) if potential memorization is detected. In this context, if $m_{T-1}$ is under $\\lambda$ , the prompt is considered safe from memorization, and the generation proceeds with the original user prompt $e_p$ , mirroring Stable Diffusion's output. Conversely, if $m_{T-1}$ is above $\\lambda$ , the prompt is then identified as likely to produce memorized images, thus $e_p$ is optimized to and replaced with an engineered prompt $e^*$ in the conditional prediction while keeping the unconditional prediction $\\epsilon_{\\theta}(x_t, e_{\\phi})$ as is following the classical CFG equation. Specifically, $e^*$ is optimized to reduce the first-step magnitude $m_{T-1}$ using gradient descent and the optimization halts once the magnitude $m_{T-1}$ falls below $\\lambda$ , marking the engineered prompt as less prone to memorization. Both detection and prompt optimization occur only during the initial inference step $T-1$ , rather than throughout the entire inference process, enhancing efficiency.", + "bbox": [ + 91, + 407, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Privacy-utility trade-off. Eq. (10) suffers from a severe privacy-utility trade-off as in Fig. 3. Opting for a lower $\\lambda$ improves the privacy (risk of memorization) of outputs at the expense of utility (text-alignment). It requires more substantial optimization of the input prompt, aiming to achieve a magnitude $m_{T - 1}$ lower than $\\lambda$ , thereby enhancing privacy. However, such extensive modifications to the prompt may lead to a loss of its original meaning, compromising utility significantly. Conversely, a higher $\\lambda$ upholds the utility of generated outputs but at the cost of privacy.", + "bbox": [ + 89, + 714, + 483, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric analysis. Fig. 4 (a) visualizes the privacy-utility trade-off in Eq. (10) as a function of $\\lambda$ , where vary", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing levels of $\\lambda$ define distinct contours corresponding to target magnitudes and privacy levels during prompt engineering (PE). Outer contours represent greater prompt modifications, which enhance privacy. This optimization shifts the prompt embedding from the original memorized prompt, $e_{p}$ , located at point $\\mathsf{O}$ , to the engineered prompt, $e^{*}$ , at points B, C, or D. The number of PE iterations determines the distance of the prompt from point $\\mathsf{O}$ : more iterations lead to greater divergence, achieving higher privacy levels. The process halts once the desired privacy level is attained.", + "bbox": [ + 511, + 407, + 906, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview of our improvements. Our trade-off analysis reveals that existing strategies rely solely on prompt modification methods such as prompt engineering to enhance privacy, which is suboptimal and significantly compromises utility. This highlights an urgent need to propose alternative methods to improve privacy and preserve utility. To address this, we introduce two meticulously crafted mitigation strategies that only require refining the guidance components in CFG to achieve a superior privacy-utility tradeoff: (1) prompt re-anchoring (PR) and (2) semantic prompt search (SS). Specifically, SS enhances utility with minimal privacy cost, while PR serves as an effective alternative to improve privacy, which has the potential of costing less utility when used in combination with SS.", + "bbox": [ + 511, + 559, + 908, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Enhancing Privacy Through Prompt Re-Anchoring", + "text_level": 1, + "bbox": [ + 511, + 776, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One of our objectives is to safeguard privacy while preserving as much utility as possible. Recall that in Eq. (10) enhancing privacy requires reducing the magnitude, which correlates with a significant decrease in utility. It is important to note that the utility loss stems not from the reduction in magnitude itself, but from the baseline's inefficient", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8185", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "approach to achieving that reduction. Visually, in Fig. 4, predictions that lie on the same magnitude contour may exhibit notably different levels of utility. This observation motivates us to ponder: Is it possible to reduce magnitude through a more efficient approach (incurring a lower utility cost) compared to the baseline method of prompt engineering? This insight has spurred the creation of our prompt re-anchoring (PR) strategy, which aims to reduce magnitude by steering predictions away from memorized images.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We argue that $e_p$ is discarded too hastily in Eq. (10) when the memorization is detected ( $m_{T-1} > \\lambda$ ). Instead, recognizing $e_p$ as a valuable anchor point — its conditional generation can trace the memorization path — we use $e_p$ to re-anchor the generation process, by replacing the unconditional prediction in CFG with the $e_p$ -conditioned prediction to divert the generation away from memorization. Thereby, Eq. (10) is re-written to the following form,", + "bbox": [ + 89, + 227, + 480, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} < \\lambda \\right\\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e ^ {*}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right)\\right) \\right] \\mathbb {1} _ {\\{m _ {T - 1} > \\lambda \\}}. \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 357, + 480, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis: Enhancing Privacy with Less Utility Cost. The baseline approach in Eq. (10) to enhancing privacy necessitates significant utility sacrifices, primarily due to the need for more extensive gradient updates during prompt optimization to obtain $e^{*}$ . This leads to substantial modifications of the original user prompt and a large gap between $e^{*}$ and $e_{p}$ , causing the prompt to diverge significantly from the user's initial intent and thus undermine the text alignment. In contrast, our prompt re-anchoring strategy employs the intuition from the original CFG approach to guide generations toward a preferred output while distancing them from an undesirable one, effectively achieved through linear interpolation between two generations. Uniquely, our strategy designates the undesired generation as the memorized one that enforces conditioning on the original user prompt $e_{p}$ , rather than relying on unconditional conditioning $e_{\\phi}$ . This approach steers generations away from memorized content more efficiently with the contrastive direction, which therefore requires fewer gradient updates to obtain $e^{*}$ and thus less alternation on the prompt for the same privacy enhancement level. Our previous question is hence answered.", + "bbox": [ + 89, + 417, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 represents a geometric interpretation of our prompt re-anchoring strategy. Classical CFG directs $\\hat{\\epsilon}$ away from unconditional (point A) towards conditional predictions (O for $e_p$ , and B, C, D for $e^*$ ) as in the left figure. PR, however, guides $\\hat{\\epsilon}$ from the memorized prediction (O) towards less or unmemorized predictions with lower magnitude (B, C, D), enhancing privacy, as in the middle figure.", + "bbox": [ + 89, + 734, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis: Continuous Memorization Diversion without Loss of Efficiency. Although [36] found that minimizing the initial magnitude $m_{T-1}$ can indirectly reduce magnitudes in later steps, we observe that the magnitude", + "bbox": [ + 89, + 840, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can still spike and memorization may still occur later in the inference process. For example, as illustrated in Fig. 8 (in Supplementary Material), while the engineered prompt achieves a low magnitude initially, it fails to ensure that this low level is maintained throughout subsequent inference steps, potentially leading back to high magnitudes indicative of Stable Diffusion's memorized outputs. To counteract this, the PR strategy can continuously divert generations away from initially identified memorized conditions (point $\\bigcirc$ ) across the entirety of the inference process as shown in Fig. 8 (in Supplementary Material). The effectiveness is also evident from the qualitative appearance of the final generated image, and its quantitative similarity score as in Fig. 8 (in Supplementary Material).", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Enhancing Utility Through Semantic Prompt Search", + "text_level": 1, + "bbox": [ + 511, + 310, + 903, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we aim to propose a strategy to enhance utility with minimal privacy cost. Recall that low utility is primarily due to the increased difference between the user prompt and the engineered prompt. We suggest conducting semantic search to find a prompt $e_p^{ss}$ semantically similar to the original user prompt $e_p$ but with a lower risk of memorization, aiming to replace the engineered prompt $e^*$ that results in the unsatisfactory trade-off. Eq. (11) is thus further re-formulated to the following form after incorporating semantic prompt search strategy in addition to the prompt re-anchoring strategy,", + "bbox": [ + 511, + 348, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} < \\lambda \\right\\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p} ^ {s s}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} > \\lambda \\right\\}}. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 523, + 903, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. (12), the engineered prompt $e^{*}$ from Eq. (10) and Eq. (11) is replaced by $e_{p}^{ss}$ . The search is conducted in the linguistic space to inherently preserve semantics. We use a large language model (LLM) to find up to $n_{s}$ alternatives to the user prompt. This search early stops once an alternative prompt $e_{p}^{ss}$ 's magnitude falls below the $\\lambda$ threshold. Visually in Fig. 4 (c), this is equivalent to reaching the target privacy contour at points $\\mathsf{B}'$ , $\\mathsf{C}'$ , or $\\mathsf{D}'$ . Notably, although $\\mathsf{B}$ and $\\mathsf{B}'$ lie on the same privacy contour, the utility of $\\mathsf{B}'$ is often higher than that of $\\mathsf{B}$ . Implementationally, if all $n_{s}$ alternatives result in magnitudes above $\\lambda$ , the one with the lowest magnitude will be chosen.", + "bbox": [ + 511, + 583, + 903, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis: Enhancing Utility with No Privacy Cost. This semantic search differs from soft prompt engineering in that the optimization happens in the language space instead of the embedding space. The baseline strategy seeks to boost utility by minimizing the number of prompt optimization steps, which diminishes the extent of prompt modifications, significantly undermining privacy as ostensibly secure prompts may inadvertently produce memorized outputs. Conversely, our approach employs a semantic prompt", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8186", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dfd5be107f4f622f002b851ebad88e9c5521de549c1a9823d7254b225f813db4.jpg", + "image_caption": [ + "Figure 5. Both the baseline approach and our method effectively inhibit Stable Diffusion from producing memorized images by minimizing the magnitude in the initial inference step for prompt \"The No Limits Business Woman Podcast\". Nonetheless, our semantic prompt search (SS) strategy, which results in \"The Empowered Business Woman's Podcast\", retains considerably more utility, as demonstrated by the substantially higher CLIP similarities between the original user text prompt and both our (1) modified prompt, and (2) ultimately generated image. Qualitative results further corroborate this finding. $m_{T-1}$ is the first-step magnitude, $CLIP_{txt}$ and $CLIP_{img}$ refer to the original prompt's CLIP similarity with the modified prompt and the generated image." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 468, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "search to inherently preserve utility, leveraging the capabilities of Large Language Models (LLM) to navigate the linguistic space for suitable alternatives. By pinpointing a substitute prompt in the language space, we keep the semantic similarity and obviate the need to trade privacy for enhanced utility.", + "bbox": [ + 89, + 417, + 483, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis: Enhancing Privacy with Minimum Utility Cost. This is also guaranteed by design. As discussed in Sec. 3.2, the baseline requires significant utility sacrifices to gain extra privacy. In contrast, the utility impact associated with our semantic prompt search is limited to the semantic discrepancy between the original user prompt and the LLM-derived alternative, which is much smaller. Furthermore, when the search yields an alternative prompt devoid of memorization risks, it effectively represents a significant boost in privacy for a minimal cost of utility. Geometrically as in Fig. 4 (c), this means shifting the prediction from $\\bigcirc$ to $\\mathbb{D}^{\\prime}$ with minimum utility cost.", + "bbox": [ + 89, + 507, + 483, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 demonstrate SS's superiority over the baseline using a widely-investigated memorized prompt: \"The No Limits Business Woman Podcast\". While the baseline's prompt engineering (PE) strategy successfully reduces the magnitude to 1.15 and prevents memorization, it results in a significant semantic deviation from the original prompt, as indicated by low CLIP similarity scores for both the engineered text prompt and the generated image. In contrast, our SS strategy not only reduces the magnitude further to 0.78, effectively preventing memorization, but also significantly improves text alignment. By navigating the linguistic space, our method finds \"The Empowered Business Woman's Podcast\" as a privacy-preserving yet semantically aligned alternative.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Synergy Effect of the Two Strategies", + "text_level": 1, + "bbox": [ + 511, + 90, + 830, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis: Tackling the Limitation of Semantic Search (SS). In an extreme scenario, SS might encounter a limitation where all $n_s$ searched alternatives result in magnitudes exceeding the set threshold, failing to meet the required privacy standards. This situation can be geometrically represented by a movement from point $\\mathsf{O}$ to $\\mathsf{B}'$ in Fig. 4 (c). However, it's important to note that even in such cases, the transition to a lower magnitude contour, where $\\mathsf{B}'$ is located, still offers a certain degree of privacy enhancement without additional cost. Further advancements in privacy (e.g., moving from $\\mathsf{B}'$ to $\\mathsf{D}'$ ) can be achieved through our synergistic application of the PR strategy, which incurs a smaller utility cost compared to the baseline method. The right side of Fig. 8 (in Supplementary Material) showcases the synergistic benefit of utilizing the PR strategy to compensate for the SS strategy's limitations. In this example, although SS may not secure an alternative prompt with a sufficiently reduced magnitude, lowering it from the original 7.48 to 6.02 has proved to be adequate for PR to leverage the original prompt as the new anchor point for steering the generation away from memorized content effectively.", + "bbox": [ + 511, + 112, + 906, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis: Addressing the Privacy-Utility Dilemma. For the baseline strategy [36], the dilemma in balancing the privacy-utility trade-off through adjusting the $\\lambda$ hyperparameter manifests in two ways: opting for a higher $\\lambda$ to bolster utility inadvertently raises the False Negative (FN) rate, leading to potentially unsafe prompts being misclassified as safe and thus compromising privacy; conversely, selecting a lower $\\lambda$ to enhance privacy increases the False Positive (FP) rate, resulting in unnecessary modifications to already safe prompts, thereby reducing utility. Prompt re-anchoring is tailored to mitigate the privacy repercussions of False Negatives (FN) encountered when users aim to augment utility with a higher $\\lambda$ . It achieves this by actively directing generation away from memorized content, as elucidated in Sec. 3.2 and depicted in Fig. 4 as a divergence from $\\bigcirc$ . This approach thus optimizes the privacy-utility trade-off under such conditions. Semantic prompt search, on the other hand, aims to alleviate the utility decline tied to False Positives (FP) when enhancing privacy with a lower $\\lambda$ . It conducts a linguistically sensitive search that avoids semantic degradation, consequently refining the privacy-utility balance in these scenarios. Employing both strategies synergistically ensures an enhanced privacy-utility trade-off, regardless of whether the user's priority lies with maximizing privacy or utility.", + "bbox": [ + 511, + 431, + 908, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 821, + 645, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation settings", + "text_level": 1, + "bbox": [ + 511, + 847, + 696, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although our method is designed to enhance the approach in [36], it can be easily adapted to any recent advance", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8187", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ment of the detection strategy by simply replacing $m_{T-1}$ in Eq. (12) with alternative detection signals, making PRSS highly practical and adaptable for future improvements. For example, a contemporary work [8] introduced the masked magnitude $m_t'$ (see Eq. (8)), achieving new state-of-the-art detection performance. The mitigation strategy proposed by [8] also uses prompt engineering based on the upgraded detection signal $m_t'$ , offering two key advantages: (1) $m_t'$ provides a more accurate detection trigger for initiating the mitigation strategy, (2) it serves as a more effective loss function than unmasked magnitude, yielding better gradients for optimizing prompts. These improvements make it the current state-of-the-art in mitigation. PRSS can seamlessly integrate this enhanced detection signal by replacing $m_{T-1}$ with $m_{T-1}'$ in Eq. (12), producing our best results to date. For a fair comparison, we re-implement the mitigation strategy from [8] and evaluate its performance alongside PRSS. This comparison not only highlights PRSS's capability to surpass the most recent state-of-the-art but also serves as an ablation study, assessing PRSS's impact since both methods use the same detection strategy to trigger mitigation. Additionally, we produce a version of PRSS that uses $m_t$ as the detection signal, aligning with [36], to further demonstrate PRSS's effectiveness in comparison to [36].", + "bbox": [ + 89, + 90, + 483, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow the baselines [8, 36] by experimenting on Stable Diffusion v1-4, using its corresponding prompt dataset from [34], which contains 500 prompts, over 300 of which are classified as memorized and organized by different memorization types. To provide a comprehensive view, we adopt the settings from [8] and report results for both global and local memorization prompts separately. We use the default hyperparameters from [36] to re-implement the detection and prompt engineering mitigation strategies from [36] and [8]. For semantic prompt search with LLMs, we use the GPT-4 model via the OpenAI API, instructing it to generate up to $n_s = 25$ semantically similar prompts for each user prompt. The associated inference time and cost are minimal, with each alternative prompt generated in approximately 0.9 seconds, costing around $0.02.", + "bbox": [ + 89, + 457, + 483, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following standard practice in memorization literature, we use the CLIP score to measure the alignment between the generated output and the input user prompt, serving as our utility metric. For privacy metrics, which quantify the degree of memorization, most studies rely on SSCD-based scores (see Eq. (5)), reporting metrics such as the average score, the 95th percentile score, and the percentage of memorized generations, with a threshold of 0.5 commonly indicating memorization. Additionally, [8] proposes a localized similarity metric (LS, see Eq. (6)) that measures similarity between the generated image and corresponding training image only within the localized memorization region. We present comparative results using the widely adopted SSCD metric in Fig. 6 and Fig. 7, while other metrics, including", + "bbox": [ + 89, + 688, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f701e94eadb009f9e20942b349979d6e265558bb5229dc535eff97b8966947b5.jpg", + "image_caption": [ + "Figure 6. Comparison with baselines. Across different detection signals ( $m$ and $m'$ ), PRSS consistently achieves superior privacy-utility trade-offs. The improvement is particularly pronounced in global memorization compared to local memorization." + ], + "image_footnote": [], + "bbox": [ + 524, + 94, + 715, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5d280afac8367d66c49dd9e8a475d6b9dc35683989d4387f1e06bef24a6ca95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 95, + 895, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LS, are detailed in the Supplementary Material for a more comprehensive evaluation.", + "bbox": [ + 511, + 335, + 906, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 511, + 377, + 759, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 6, PRSS consistently outperforms prompt engineering (PE) in achieving better privacy-utility trade-offs, regardless of the detection signal used (whether $m$ or $m'$ ). This is visually evident as the red and green lines (PRSS) lie further in the bottom-right direction compared to the yellow and blue lines (PE) for both global and local memorization scenarios. Notably, PRSS demonstrates astonishingly large improvements in global memorization (from the blue line to the green line and from the yellow line to the red line), while [8] achieves only marginal gains over [36] in this case (from the blue line to the yellow line and from the green line to the red line). For local memorization, although the margin of improvement is smaller, PRSS still achieves a significant improvement over PE. Specifically, even using the inferior detection signal $m$ , PRSS (green line) achieves comparable performance with the method proposed by [8] (yellow line), which leverages the superior signal $m'$ . Equipped with $m'$ , PRSS can achieve the state-of-the-art results (red line). This highlights PRSS's robust capability to leverage advancements in detection signals.", + "bbox": [ + 511, + 398, + 908, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 710, + 679, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 highlights the necessity of both components in PRSS, as evidenced by the red line lying further in the bottom-right direction compared to the yellow and green lines for both global and local memorization scenarios. The synergistic effect of combining PR and SS as discussed in Sec. 3.4, is particularly evident in the local memorization scenario, which aligns with [8]'s observation that local memorization poses greater challenges for existing mitigation strategies. Using PR alone effectively enhances privacy but results in suboptimal utility, while SS alone improves utility but achieves limited privacy protection. Together, they", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8188", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/963b6e8d9d17ff6018a78a1202f1a9aec28734ebc378b21b33a836242b1a16a1.jpg", + "image_caption": [ + "Figure 7. Ablation studies of PRSS components. Using PR alone results in low utility, while SS alone cannot ensure privacy. Combining PR and SS achieves a synergistic effect, delivering the best privacy-utility trade-off." + ], + "image_footnote": [], + "bbox": [ + 101, + 94, + 470, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "leverage each other's strengths, achieving the best privacy-utility trade-off. For global memorization, either PR or SS alone significantly outperforms the baseline. However, their combination further enhances the trade-off, underscoring the robustness of PRSS in tackling both global and local memorization challenges.", + "bbox": [ + 89, + 321, + 483, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 89, + 414, + 233, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Memorization in Generative Models The phenomenon of generative models memorizing training data has been extensively studied across domains, including language models [3, 4, 15, 16, 18, 20], image generation [2, 10, 11, 21, 22, 33], and video generation [9, 26], due to the significant privacy and legal risks it poses [5, 14, 29, 37]. Specifically, in diffusion models, [6, 31, 34] first validated the memorization issue on various models including DDPMs [13, 23] and Stable Diffusion [28] on datasets include Oxford Flowers [24], CelebA [19], CIFAR-10 [17] and LAION [30]. Further analysis by [32] revealed that Stable Diffusion's text-conditioning exacerbates overfitting, making its memorization of LAION particularly problematic. To better understand memorization, [34] categorized the phenomenon as Matching Verbatim, Retrieval Verbatim, and Template Verbatim based on the cause, distinguishing between one-to-one and many-to-many mappings in text-image pairs, while [8] classified memorization into global and local cases, demonstrating the practical significance of this distinction. In this work, we focus on Stable Diffusion due to its widespread use, open-source nature, and real-world legal challenges [29]. We adopt the framework of [8] to analyze both global and local memorization for a more nuanced evaluation of our proposed strategies.", + "bbox": [ + 89, + 433, + 482, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection Strategies for Diffusion Memorization. Early detection methods required generating large image sets to estimate density [6] or comparing generations with the entire training dataset [31], which were computationally intensive. Recent approaches [27, 36] introduced more efficient one-step detection strategies. [27] detected memorization by observing less concentrated cross-attention scores", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for memorized generations than non-memorized ones. [36] proposed using the abnormally high magnitude of text-conditional noise predictions as a detection signal. Building on this, [8] incorporated localized memorization masks into the magnitude computation, offering a more accurate detection signal and achieving state-of-the-art performance. Our proposed PRSS approach leverages these detection mechanisms to trigger mitigation strategies. Notably, PRSS can adapt to any detection method by substituting the indicator variable condition in its framework. We employ the state-of-the-art method from [8] for optimal results.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mitigation Strategies for Diffusion Memorization. Mitigation strategies have evolved rapidly. Early methods, such as retraining models on de-duplicated data [6], were computationally expensive and showed limited effectiveness. [32] introduced randomness into text prompts or embeddings but indiscriminately affected non-memorized prompts, reducing overall utility. To improve on it, [7] proposed to compute the generation's similarity with the closest training image to engineer the image latent using the classifier guidance framework. However, this relies on computationally expensive nearest-neighbor searches across the training dataset. [36] proposed to conduct prompt engineering using the magnitude of text-conditional prediction as the loss, and [27] proposed to re-weight cross-attention scores for all tokens in the prompt. They offer improved utility-privacy trade-offs and achieve comparable performance. Most recently, [8] extended these efforts by leveraging localized magnitudes as loss functions for prompt engineering, achieving the best results to date. Our investigation reveals the limitations in the prevailing method's privacy/utility balance. In response, we introduce new mitigation strategies that achieve more favorable trade-offs.", + "bbox": [ + 511, + 257, + 906, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 595, + 784, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a novel PRSS method that consists of prompt re-anchoring and semantic prompt search as deliberately designed strategies to mitigate memorization in diffusion models. Empirical results show that PRSS demonstrates superior privacy-utility trade-offs over existing methods. Beyond its effectiveness, PRSS is simple and efficient to implement. It requires only minor modifications to the CFG equation to update noise predictions during inference, seamlessly integrating into the inference pipeline without additional training, fine-tuning, or reliance on training data. The efficacy of each component is validated through comprehensive analysis and ablation studies. A current limitation is that the performance of PRSS, as well as baseline mitigation strategies, relies heavily on the accuracy of the detection mechanism. When it fails to trigger accurately, they all revert to the standard Stable Diffusion process, compromising privacy preservation. Thus, improving the detection mechanism's accuracy in future works would orthogonally bolster the effectiveness of PRSS.", + "bbox": [ + 511, + 613, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8189", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Australian Research Council under Projects DP210101859 and FT230100549.", + "bbox": [ + 89, + 114, + 483, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Midjourney. https://www.midjourney.com/. 1", + "[2] Sanjeev Arora, Andrej Risteski, and Yi Zhang. Do gans learn the distribution? some theory and empirics. In International Conference on Learning Representations, 2018. 8", + "[3] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In USENIX Security Symposium, 2021. 8", + "[4] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022.8", + "[5] Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang, Nicolas Papernot, Andreas Terzis, and Florian Tramér. The privacy onion effect: Memorization is relative. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 8", + "[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In USENIX Security Symposium, pages 5253-5270, 2023. 1, 8", + "[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In CVPR, 2024. 1, 3, 8", + "[8] Chen Chen, Daochang Liu, Mubarak Shah, and Chang Xu. Exploring local memorization in diffusion models via bright ending attention. *ICLR*, 2025. 1, 3, 7, 8", + "[9] Chen Chen, Enhuai Liu, Daochang Liu, Mubarak Shah, and Chang Xu. Investigating memorization in video diffusion models. ICLR 2nd Workshop on Navigating and Addressing Data Problems for Foundation Models (DATA-FM), 2025. 8", + "[10] Yunhao Chen, Xingjun Ma, Difan Zou, and Yu-Gang Jiang. Towards a theoretical understanding of memorization in diffusion models. arXiv preprint arXiv:2410.02467, 2024. 8", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 8", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 1, 3", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 8", + "[14] H. H. Jiang, L. Brown, J. Cheng, M. Khan, A. Gupta, D. Workman, A. Hanna, J. Flowers, and T. Gebru. AI Art and Its Impact on Artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374. ACM, 2023. 8" + ], + "bbox": [ + 93, + 198, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. Large language models struggle to learn long-tail knowledge. arXiv preprint arXiv:2211.08411, 2022. 8", + "[16] Nikhil Kandpal, Eric Wallace, and Colin Raffel. Deduplicating training data mitigates privacy risks in language models. In International Conference on Machine Learning, 2022. 8", + "[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 8", + "[18] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the Association for Computational Linguistics, 2022. 8", + "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 8", + "[20] Fatemehsadat Mireshghallah, Anish Uniyal, Tianhao Wang, David Evans, and Taylor Berg-Kirkpatrick. Memorization in nlp fine-tuning methods. arXiv preprint arXiv:2205.12506, 2022.8", + "[21] Amir Naseh, Jaemin Roh, and Amir Houmansadr. Memory triggers: Unveiling memorization in text-to-image generative models through word-level duplication. arXiv preprint arXiv:2312.03692, 2023. 8", + "[22] Ali Naseh, Jaechul Roh, and Amir Houmansadr. Understanding (un)intended memorization in text-to-image generative models. arXiv preprint arXiv:2312.07550, 2023. 8", + "[23] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 8", + "[24] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 8", + "[25] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, pages 14532-14542, 2022. 3", + "[26] Aimon Rahman, Malsha V. Perera, and Vishal M. Patel. Frame by familiar frame: Understanding replication in video diffusion models. arXiv preprint arXiv:2403.19593, 2024. 8", + "[27] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. ECCV, 2024. 1, 3, 8", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 8", + "[29] Joseph Saveri and Butterick Matthew. Stable diffusion litigation, 2023. 2023. 8", + "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8190", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 1, 8", + "[31] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In CVPR, pages 6048-6058, 2023. 1, 3, 8", + "[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. In NeurIPS, 2023. 1, 3, 8", + "[33] Patrick Tinsley, Adam Czajka, and Patrick Flynn. This face does not exist... but it might be yours! identity leakage in generative models. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 8", + "[34] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 1, 7, 8", + "[35] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Advances in Neural Information Processing Systems, pages 51008-51025. Curran Associates, Inc., 2023. 1", + "[36] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In ICLR, 2024. 1, 3, 5, 6, 7, 8", + "[37] Chiyuan Zhang, Daphne Ippolito, Katherine Lee, Matthew Jagielski, Florian Tramér, and Nicholas Carlini. Counterfactual memorization in neural language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023." + ], + "bbox": [ + 91, + 90, + 482, + 515 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 125, + 517, + 133, + 526 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8191", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_model.json b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa3ff178ffaee8768d49ef1faba68dc6f068c865 --- /dev/null +++ b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.13, + 0.882, + 0.175 + ], + "angle": 0, + "content": "Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.203, + 0.747, + 0.221 + ], + "angle": 0, + "content": "Chen Chen\\(^{1}\\) Daochang Liu\\(^{2}\\) Mubarak Shah\\(^{3}\\) Chang Xu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.222, + 0.882, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Faculty of Engineering, The University of Sydney, Australia \\(^{2}\\)School of Physics, Mathematics and Computing, The University of Western Australia, Australia \\(^{3}\\)Center for Research in Computer Vision, University of Central Florida, USA" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.277, + 0.838, + 0.293 + ], + "angle": 0, + "content": "{cche0711@uni., c.xu@}sydney.edu.au daochang.liu@uwa.edu.au shah@crcv.ucf.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.631 + ], + "angle": 0, + "content": "Text-to-image diffusion models have demonstrated remarkable capabilities in creating images highly aligned with user prompts, yet their proclivity for memorizing training set images has sparked concerns about the originality of the generated images and privacy issues, potentially leading to legal complications for both model owners and users, particularly when the memorized images contain proprietary content. Although methods to mitigate these issues have been suggested, enhancing privacy often results in a significant decrease in the utility of the outputs, as indicated by text-alignment scores. To bridge the research gap, we introduce a novel method, PRSS, which refines the classifier-free guidance approach in diffusion models by integrating prompt re-anchoring (PR) to improve privacy and incorporating semantic prompt search (SS) to enhance utility. Extensive experiments across various privacy levels demonstrate that our approach consistently improves the privacy-utility trade-off, establishing a new state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.222, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Leveraging classifier-free guidance (CFG) [12], text-to-image diffusion models like Stable Diffusion [28] and Midjourney [1] are now capable of generating highly realistic images that closely align with user-provided text prompts. This capability has propelled their popularity, leading to widespread use and distribution of their generated images. However, recent research [6, 8, 31, 34] has revealed a critical issue: these models can memorize training data, leading them to reproduce parts of images, such as foregrounds or backgrounds (local memorization, see Fig. 1), or even entire images (global memorization, see Fig. 2) during inference, instead of generating genuinely novel content. When the training data includes sensitive or copyrighted material, these memorization issues can infringe on copyright laws without notifying either the model's owners or users or the copyright holders of the replicated content. The risk is es" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.569 + ], + "angle": 0, + "content": "ppecially substantial given the extensive use of these models and their reliance on massive datasets, such as LAION [30], which contain billions of web-scale images and are impractical to thoroughly review or filter manually. This risk is further highlighted by real-world cases, where several artists have filed lawsuits, arguing that models like Stable Diffusion act as “21st-century collage tools” that remix their copyrighted works, implicating Stability AI, DeviantArt, Midjourney, and Runway AI. Recognizing the potential for unauthorized reproductions, Midjourney has even banned prompts containing the term “Afghan” to prevent the generation of the copyrighted Afghan Girl photograph. Yet, as [35] demonstrates, such restrictions alone are insufficient to fully prevent the reproduction of copyrighted images. This underscores the urgent need for timely effective mitigation strategies to address these concerns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In response to these legal challenges, recent initiatives [7, 8, 27, 32, 36] have focused on developing strategies to minimize memorization, achieving notable success. These approaches vary in scope, with some targeting the model's training phase and others making adjustments during inference. While training-phase strategies are theoretically viable, they lack practical significance as they predominantly experiment on a small \\(10\\mathrm{k}\\) subset of the entire LAION5B dataset due to the impracticality of fine-tuning on the full dataset — a constraint stemming from computational limitations. In contrast, inference-phase strategies, which involve modifications only during inference, offer computational efficiency and adaptability, as they require no additional training or fine-tuning. This makes them highly applicable across pre-trained diffusion models, regardless of specific training configurations. Despite their promise, inference-time strategies can also suffer from computational constraints. For example, the approach proposed by [7] involves searching the entire LAION dataset for the nearest neighbor to optimize the generated latent representation during each denoising step. Other inference-time strategies [8, 27, 32, 36] that avoid nearest neighbor search" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.344 + ], + "angle": 0, + "content": "Figure 1. Examples of Local Memorization: Stable Diffusion generations (columns 2-4) can replicate local regions of training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.348, + 0.482, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Figure 2. Examples of Global Memorization: Stable Diffusion generations (columns 2-4) can replicate the entire training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.703 + ], + "angle": 0, + "content": "demonstrate high efficiency but only moderate effectiveness, leaving a substantial gap in optimizing the privacy-utility trade-off. In particular, these methods often require significant compromises in text-image alignment to reduce memorization and conversely, struggle to sufficiently suppress memorization when trying to preserve alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we carefully analyze why existing strategies struggle to achieve an optimal privacy-utility trade-off, uncovering critical insights that the two inputs (text-conditional and unconditional predictions) used in the CFG framework are both suboptimal for mitigating memorization, even though they are standard for most other tasks in Stable Diffusion models. Specifically, we find that the text-conditional term in CFG, when optimized for privacy by prompt engineering, can significantly reduce text alignment, indicating the need for alternative methods to find privacy-preserving yet intent-aligned text-conditioning. Additionally, we observe that the unconditional term in CFG is less effective as an anchor for privacy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.501 + ], + "angle": 0, + "content": "preservation, suggesting that alternative anchoring points that enhance privacy are necessary. Building on these insights, we introduce PRSS, a novel method that incorporates two complementary strategies to refine the widely used CFG approach for improved memorization mitigation. Our method includes (1) prompt re-anchoring, designed to preserve high privacy with a smaller utility cost, which re-anchors the unconditional component in CFG using the memorized prompt to consistently guide generations away from memorization during the inference process; (2) semantic prompt search, designed to preserve high utility with minimal privacy compromise, which optimizes the prompt away from the memorized one in the language space while keeping similar semantics. Our experimental results reveal that employing both strategies synergistically can consistently outperform baseline methods in achieving superior privacy-utility trade-offs at different privacy levels. In particular, this results in higher text alignment (utility) and substantially fewer memorized generations (improved privacy). Furthermore, ablation studies validate the efficacy of each component within our strategy. From an implementation standpoint, our strategies are simple, efficient, and effective, necessitating only adjustments to the existing CFG equation for updating the noise prediction during inference. This integration is seamless, avoiding disruption to the inference workflow and eliminating the need for additional re-training, fine-tuning, or searching over the training data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.512, + 0.652, + 0.527 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.537, + 0.896, + 0.552 + ], + "angle": 0, + "content": "2.1. Classifier-Free Guidance in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.906, + 0.634 + ], + "angle": 0, + "content": "Diffusion models [13, 23] are popular generative models that consist of a forward diffusion process that consistently adds Gaussian noise to real images \\( x_0 \\) for \\( T \\) steps based on the following conditional function, resulting in \\( x_T \\sim \\mathcal{N}(0, \\mathbf{I}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.645, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.717 + ], + "angle": 0, + "content": "where \\( x_{t} \\) represents the noised version of \\( x_{0} \\) after adding \\( t \\) noise, and \\( \\{\\beta_t\\}_{t = 1}^T \\) is the schedule that controls the magnitude of added noise at time \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.779 + ], + "angle": 0, + "content": "This is followed by the reverse process, in which a de-noiser model is trained to predict the distribution (i.e., mean \\(\\mu_{\\theta}(x_t)\\) and variance \\(\\sigma_{\\theta}^{2}(x_{t})\\mathbf{I}\\)) of a cleaner version of \\(x_{t}\\), denoted as \\(x_{t - 1}\\), at any given timestep \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.788, + 0.905, + 0.806 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}\\right), \\sigma_ {\\theta} ^ {2} \\left(x _ {t}\\right) \\mathbf {I}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.906, + 0.876 + ], + "angle": 0, + "content": "During training, the objective function is equivalent to have the denoiser \\(\\epsilon_{\\theta}(x_t,y,t)\\) predict the noise \\(\\epsilon_{t}\\) to be removed, rather than the image \\(x_{t - 1}\\) at timestep \\(t\\) conditioned on \\(y\\) which may represent a user-provided text prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {t \\in [ 1, T ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (x _ {t}, y, t) \\| _ {2} ^ {2} \\right]. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8183" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "Classifier-free guidance (CFG) [12] is a widely used technique for balancing fidelity and diversity in the generated outputs of diffusion models, and it can be easily applied during inference. CFG operates by linearly guiding the noise prediction from the unconditional prediction, \\(\\epsilon_{\\theta}(x_t,t)\\) towards the conditional prediction, \\(\\epsilon_{\\theta}(x_t,y,t)\\), with a scaling factor \\(s > 1\\) to control the degree of guidance. For simplicity, removing the timestep \\(t\\) from the notation, the adjusted noise prediction can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.235, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}\\right) + s \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, y\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.258, + 0.392, + 0.272 + ], + "angle": 0, + "content": "2.2. Memorization in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.455 + ], + "angle": 0, + "content": "Definition and Metrics. A widely adopted practical definition of memorization refers to generated images exhibiting high object-level similarity to training data [31]. This includes the concept of reconstructive memory in diffusion models, where foreground and background objects from training images are combined, potentially with modifications such as shifting, scaling, or cropping. These reconstructions may not match any training image pixel-for-pixel but still closely resemble specific training images at the object level. The standard metric for evaluating memorization [7, 8, 27, 31, 32, 36] is designed base on a pre-trained Self-Supervised Copy Detection (SSCD) model [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.458, + 0.483, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S S C D} (\\hat {x}) = \\max _ {x _ {i} \\in \\mathcal {D}} \\operatorname {C o s i n e} \\left(\\mathbf {E} _ {S S C D} (\\hat {x}), \\mathbf {E} _ {S S C D} (x _ {i})\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.483, + 0.54 + ], + "angle": 0, + "content": "where for each generated image \\(\\hat{x}\\), cosine similarity is computed between its SSCD embedding \\(\\mathbf{E}_{SSCD}(\\hat{x})\\) and the embeddings \\(\\mathbf{E}_{SSCD}(x_i)\\) of all training images \\(x_i\\) in the dataset \\(\\mathcal{D}\\). The maximum similarity serves as the final SSCD score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.484, + 0.661 + ], + "angle": 0, + "content": "Most recently, [8] proposed a method to localize memorized regions in generated images, arguing that only memorized regions pose privacy risks, making local memorization a more meaningful concept. They introduced an efficient approach to synthesize a mask that highlights these regions without requiring access to training data. Using this mask, they proposed a new localized metric based on both object-level SSCD and pixel-level Euclidean distance:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.669, + 0.483, + 0.687 + ], + "angle": 0, + "content": "\\[\nL S (\\hat {x}, x) = - \\mathbb {1} _ {\\mathrm {S S C D} > 0. 5} \\cdot \\| (\\hat {x} - x) \\circ \\mathbf {m} \\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.752 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}\\) is the local memorization mask and \\(\\mathbb{1}_{\\mathrm{SSCD} > 0.5}\\) is an indicator function that signifies whether the SSCD score exceeds the 0.5 threshold, which is commonly used to distinguish memorized from non-memorized content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.484, + 0.829 + ], + "angle": 0, + "content": "Memorization Detection. A recent work [36] introduces an innovative approach for detecting potential memorized generations during inference by leveraging the magnitude of the difference between text-conditional \\(\\epsilon_{\\theta}(x_t, e_p)\\) and unconditional \\(\\epsilon_{\\theta}(x_t, e_\\phi)\\) noise predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.836, + 0.483, + 0.853 + ], + "angle": 0, + "content": "\\[\nm _ {t} = \\left\\| \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) \\right\\| _ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where memorized generations consistently exhibit larger magnitudes than non-memorized ones, making this magnitude a strong signal for detection. Furthermore, accurate" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.906, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.148, + 0.907, + 0.189 + ], + "angle": 0, + "content": "Figure 3. Illustration of the logic behind how \\(\\lambda\\) in Eq. (10) governs the privacy-utility trade-off. Currently, the privacy-utility trade-off is governed solely by how extensively the prompt is modified." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.905, + 0.235 + ], + "angle": 0, + "content": "detection can be achieved efficiently by evaluating the magnitude \\( m_{T-1} \\) at the first denoising step (timestep \\( T - 1 \\), highlighting the method's computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.905, + 0.296 + ], + "angle": 0, + "content": "The contemporary work [8] extends this detection method by incorporating their proposed local memorization mask \\(\\mathbf{m}\\) into Eq. (7) to compute a masked magnitude, achieving state-of-the-art performance:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.305, + 0.905, + 0.358 + ], + "angle": 0, + "content": "\\[\nm _ {t} ^ {\\prime} = \\left\\| \\left(\\varepsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\varepsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\circ \\mathbf {m} \\right\\| _ {2} / \\left(\\frac {1}{N} \\sum_ {i = 1} ^ {N} m _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.905, + 0.389 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}\\) is the local memorization mask and \\(\\frac{1}{N} \\sum_{i=1}^{N} m_i\\) is used to normalize the result by the mean of \\(\\mathbf{m}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.401, + 0.605, + 0.417 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.565 + ], + "angle": 0, + "content": "We begin with a detailed analysis in Sec. 3.1 to uncover the key challenges that hinder existing strategies from achieving an optimal privacy-utility trade-off. Building on these insights, we evolve the baseline step by step with our proposed prompt re-anchoring (PR) strategy in Sec 3.2, which focuses on improving privacy, and semantic prompt search (SS) in Sec 3.3, which enhances utility. Finally, in Sec. 3.4, we analyze how the combination of PR and SS achieves a synergistic effect, offering the best privacy-utility trade-off." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.86, + 0.589 + ], + "angle": 0, + "content": "3.1. Trade-off Analysis of Existing Strategies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Our method is built upon the widely-used classifier-free guidance (CFG) by treating the input text prompt from user \\( e_p \\) as the condition \\( y \\) in Eq. (4) and rewriting the unconditional noise as a null condition \\( e_\\phi \\)," + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.665, + 0.905, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.737 + ], + "angle": 0, + "content": "The one-step prompt engineering (PE) based memorization mitigation method proposed by [36] then counteract memorization by reformulating Eq. (9) as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.746, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\{m _ {T - 1} < \\lambda \\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e ^ {*}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} > \\lambda \\right\\}}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(m_{T - 1} = \\| \\epsilon_{\\theta}(x_{T - 1},e_p) - \\epsilon_{\\theta}(x_{T - 1},e_\\phi)\\| _2\\) represents the magnitude at the initial step \\(T - 1\\) as per Eq. (7), \\(\\epsilon_{\\theta}(x_t,e_p)\\) and \\(\\epsilon_{\\theta}(x_{t},e^{*})\\) are the noise predictions conditioned on the user prompt embedding \\(e_p\\) and engineered prompt embedding \\(e^*\\) (optimized based on the gradient computed from the first-step magnitude \\(m_{T - 1}\\)), and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8184" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.09, + 0.299, + 0.103 + ], + "angle": 0, + "content": "a) Privacy-Utility Trade-Off in Baseline" + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.09, + 0.636, + 0.103 + ], + "angle": 0, + "content": "b) Enhancing Privacy with Prompt Re-Anchoring (PR)" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.091, + 0.904, + 0.103 + ], + "angle": 0, + "content": "c) Enhancing Utility with Semantic Search (SS)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.113, + 0.339, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.251, + 0.325, + 0.263 + ], + "angle": 0, + "content": "\\(\\mathbf{SIM}(\\mathbb{O},\\mathbb{B}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{C}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{D})\\)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.116, + 0.61, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.252, + 0.592, + 0.263 + ], + "angle": 0, + "content": "\\(\\mathbf{SIM}(\\mathbb{O},\\mathbb{B}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{C}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{D})\\)" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.116, + 0.904, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.256, + 0.902, + 0.263 + ], + "angle": 0, + "content": "\\(\\mathrm{SIM}(\\mathbb{O},\\mathbb{C}) < \\mathrm{SIM}(\\mathbb{O},\\mathbb{C}^{\\prime})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.263, + 0.902, + 0.272 + ], + "angle": 0, + "content": "\\(\\mathbf{SIM}(\\mathbb{O},\\mathbb{D}) < \\mathbf{SIM}(\\mathbb{O},\\mathbb{D}^{\\prime})\\)" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.279, + 0.128, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.282, + 0.271, + 0.293 + ], + "angle": 0, + "content": "Contour Line of Equal Privacy" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.282, + 0.317, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.282, + 0.42, + 0.292 + ], + "angle": 0, + "content": "\\(\\rightarrow\\) Baseline Guidance" + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.282, + 0.47, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.282, + 0.576, + 0.291 + ], + "angle": 0, + "content": "Our PR Guidance" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.282, + 0.631, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.282, + 0.749, + 0.291 + ], + "angle": 0, + "content": "Our PRSS Guidance" + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.281, + 0.905, + 0.291 + ], + "angle": 0, + "content": "SIM CLIP Similarity (Utility)" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.309, + 0.124, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.309, + 0.245, + 0.319 + ], + "angle": 0, + "content": "Unconditional Prediction" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.31, + 0.305, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.311, + 0.425, + 0.321 + ], + "angle": 0, + "content": "\\(e_p\\) -conditional prediction" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.306, + 0.504, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.305, + 0.635, + 0.324 + ], + "angle": 0, + "content": "\\(e^*\\) -conditional prediction with different privacy level" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.306, + 0.738, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.304, + 0.866, + 0.324 + ], + "angle": 0, + "content": "\\(e_p^{ss}\\) -conditional prediction with different privacy level" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.33, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 4. Illustration of our proposed strategy (PRSS) compared to the baseline prompt engineering (PE) method. a) The privacy-utility trade-off in the baseline method, where the privacy increases at the cost of utility from B to C to D. b) Our proposed PR guidance, i.e., the green arrow, drives the image out of the low-privacy region more efficiently than the baseline guidance, i.e., the black arrows. c) Our proposed SS can enhance utility, by finding points \\( \\mathrm{B}^{\\prime}\\mathrm{C}^{\\prime}\\mathrm{D}^{\\prime} \\) that are at the same level of privacy as BCD but with better utility. In summary, using the proposed PR and SS jointly can enhance the privacy-utility trade-off to mitigate memorization in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.408, + 0.485, + 0.711 + ], + "angle": 0, + "content": "\\(\\epsilon_{\\theta}(x_t, e_{\\phi})\\) is the unconditional noise prediction. This strategy detects potential memorization for a given input prompt based on the magnitude signal \\(m_{T-1}\\) with a threshold \\(\\lambda\\), and undertakes soft prompt engineering (i.e., optimizes the prompt embeddings) if potential memorization is detected. In this context, if \\(m_{T-1}\\) is under \\(\\lambda\\), the prompt is considered safe from memorization, and the generation proceeds with the original user prompt \\(e_p\\), mirroring Stable Diffusion's output. Conversely, if \\(m_{T-1}\\) is above \\(\\lambda\\), the prompt is then identified as likely to produce memorized images, thus \\(e_p\\) is optimized to and replaced with an engineered prompt \\(e^*\\) in the conditional prediction while keeping the unconditional prediction \\(\\epsilon_{\\theta}(x_t, e_{\\phi})\\) as is following the classical CFG equation. Specifically, \\(e^*\\) is optimized to reduce the first-step magnitude \\(m_{T-1}\\) using gradient descent and the optimization halts once the magnitude \\(m_{T-1}\\) falls below \\(\\lambda\\), marking the engineered prompt as less prone to memorization. Both detection and prompt optimization occur only during the initial inference step \\(T-1\\), rather than throughout the entire inference process, enhancing efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Privacy-utility trade-off. Eq. (10) suffers from a severe privacy-utility trade-off as in Fig. 3. Opting for a lower \\(\\lambda\\) improves the privacy (risk of memorization) of outputs at the expense of utility (text-alignment). It requires more substantial optimization of the input prompt, aiming to achieve a magnitude \\(m_{T - 1}\\) lower than \\(\\lambda\\), thereby enhancing privacy. However, such extensive modifications to the prompt may lead to a loss of its original meaning, compromising utility significantly. Conversely, a higher \\(\\lambda\\) upholds the utility of generated outputs but at the cost of privacy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Geometric analysis. Fig. 4 (a) visualizes the privacy-utility trade-off in Eq. (10) as a function of \\(\\lambda\\), where vary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.559 + ], + "angle": 0, + "content": "ing levels of \\(\\lambda\\) define distinct contours corresponding to target magnitudes and privacy levels during prompt engineering (PE). Outer contours represent greater prompt modifications, which enhance privacy. This optimization shifts the prompt embedding from the original memorized prompt, \\(e_{p}\\), located at point \\(\\mathsf{O}\\), to the engineered prompt, \\(e^{*}\\), at points B, C, or D. The number of PE iterations determines the distance of the prompt from point \\(\\mathsf{O}\\): more iterations lead to greater divergence, achieving higher privacy levels. The process halts once the desired privacy level is attained." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.771 + ], + "angle": 0, + "content": "Overview of our improvements. Our trade-off analysis reveals that existing strategies rely solely on prompt modification methods such as prompt engineering to enhance privacy, which is suboptimal and significantly compromises utility. This highlights an urgent need to propose alternative methods to improve privacy and preserve utility. To address this, we introduce two meticulously crafted mitigation strategies that only require refining the guidance components in CFG to achieve a superior privacy-utility tradeoff: (1) prompt re-anchoring (PR) and (2) semantic prompt search (SS). Specifically, SS enhances utility with minimal privacy cost, while PR serves as an effective alternative to improve privacy, which has the potential of costing less utility when used in combination with SS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.908, + 0.809 + ], + "angle": 0, + "content": "3.2. Enhancing Privacy Through Prompt Re-Anchoring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "One of our objectives is to safeguard privacy while preserving as much utility as possible. Recall that in Eq. (10) enhancing privacy requires reducing the magnitude, which correlates with a significant decrease in utility. It is important to note that the utility loss stems not from the reduction in magnitude itself, but from the baseline's inefficient" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8185" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "approach to achieving that reduction. Visually, in Fig. 4, predictions that lie on the same magnitude contour may exhibit notably different levels of utility. This observation motivates us to ponder: Is it possible to reduce magnitude through a more efficient approach (incurring a lower utility cost) compared to the baseline method of prompt engineering? This insight has spurred the creation of our prompt re-anchoring (PR) strategy, which aims to reduce magnitude by steering predictions away from memorized images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.349 + ], + "angle": 0, + "content": "We argue that \\( e_p \\) is discarded too hastily in Eq. (10) when the memorization is detected (\\( m_{T-1} > \\lambda \\)). Instead, recognizing \\( e_p \\) as a valuable anchor point — its conditional generation can trace the memorization path — we use \\( e_p \\) to re-anchor the generation process, by replacing the unconditional prediction in CFG with the \\( e_p \\)-conditioned prediction to divert the generation away from memorization. Thereby, Eq. (10) is re-written to the following form," + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.358, + 0.482, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} < \\lambda \\right\\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e ^ {*}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right)\\right) \\right] \\mathbb {1} _ {\\{m _ {T - 1} > \\lambda \\}}. \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Analysis: Enhancing Privacy with Less Utility Cost. The baseline approach in Eq. (10) to enhancing privacy necessitates significant utility sacrifices, primarily due to the need for more extensive gradient updates during prompt optimization to obtain \\( e^{*} \\). This leads to substantial modifications of the original user prompt and a large gap between \\( e^{*} \\) and \\( e_{p} \\), causing the prompt to diverge significantly from the user's initial intent and thus undermine the text alignment. In contrast, our prompt re-anchoring strategy employs the intuition from the original CFG approach to guide generations toward a preferred output while distancing them from an undesirable one, effectively achieved through linear interpolation between two generations. Uniquely, our strategy designates the undesired generation as the memorized one that enforces conditioning on the original user prompt \\( e_{p} \\), rather than relying on unconditional conditioning \\( e_{\\phi} \\). This approach steers generations away from memorized content more efficiently with the contrastive direction, which therefore requires fewer gradient updates to obtain \\( e^{*} \\) and thus less alternation on the prompt for the same privacy enhancement level. Our previous question is hence answered." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Fig. 4 represents a geometric interpretation of our prompt re-anchoring strategy. Classical CFG directs \\(\\hat{\\epsilon}\\) away from unconditional (point A) towards conditional predictions (O for \\(e_p\\), and B, C, D for \\(e^*\\)) as in the left figure. PR, however, guides \\(\\hat{\\epsilon}\\) from the memorized prediction (O) towards less or unmemorized predictions with lower magnitude (B, C, D), enhancing privacy, as in the middle figure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis: Continuous Memorization Diversion without Loss of Efficiency. Although [36] found that minimizing the initial magnitude \\( m_{T-1} \\) can indirectly reduce magnitudes in later steps, we observe that the magnitude" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "can still spike and memorization may still occur later in the inference process. For example, as illustrated in Fig. 8 (in Supplementary Material), while the engineered prompt achieves a low magnitude initially, it fails to ensure that this low level is maintained throughout subsequent inference steps, potentially leading back to high magnitudes indicative of Stable Diffusion's memorized outputs. To counteract this, the PR strategy can continuously divert generations away from initially identified memorized conditions (point \\(\\bigcirc\\)) across the entirety of the inference process as shown in Fig. 8 (in Supplementary Material). The effectiveness is also evident from the qualitative appearance of the final generated image, and its quantitative similarity score as in Fig. 8 (in Supplementary Material)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.311, + 0.905, + 0.341 + ], + "angle": 0, + "content": "3.3. Enhancing Utility Through Semantic Prompt Search" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.515 + ], + "angle": 0, + "content": "In this section, we aim to propose a strategy to enhance utility with minimal privacy cost. Recall that low utility is primarily due to the increased difference between the user prompt and the engineered prompt. We suggest conducting semantic search to find a prompt \\( e_p^{ss} \\) semantically similar to the original user prompt \\( e_p \\) but with a lower risk of memorization, aiming to replace the engineered prompt \\( e^* \\) that results in the unsatisfactory trade-off. Eq. (11) is thus further re-formulated to the following form after incorporating semantic prompt search strategy in addition to the prompt re-anchoring strategy," + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.524, + 0.905, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} < \\lambda \\right\\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p} ^ {s s}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} > \\lambda \\right\\}}. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.764 + ], + "angle": 0, + "content": "In Eq. (12), the engineered prompt \\( e^{*} \\) from Eq. (10) and Eq. (11) is replaced by \\( e_{p}^{ss} \\). The search is conducted in the linguistic space to inherently preserve semantics. We use a large language model (LLM) to find up to \\( n_{s} \\) alternatives to the user prompt. This search early stops once an alternative prompt \\( e_{p}^{ss} \\)'s magnitude falls below the \\( \\lambda \\) threshold. Visually in Fig. 4 (c), this is equivalent to reaching the target privacy contour at points \\( \\mathsf{B}' \\), \\( \\mathsf{C}' \\), or \\( \\mathsf{D}' \\). Notably, although \\( \\mathsf{B} \\) and \\( \\mathsf{B}' \\) lie on the same privacy contour, the utility of \\( \\mathsf{B}' \\) is often higher than that of \\( \\mathsf{B} \\). Implementationally, if all \\( n_{s} \\) alternatives result in magnitudes above \\( \\lambda \\), the one with the lowest magnitude will be chosen." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Analysis: Enhancing Utility with No Privacy Cost. This semantic search differs from soft prompt engineering in that the optimization happens in the language space instead of the embedding space. The baseline strategy seeks to boost utility by minimizing the number of prompt optimization steps, which diminishes the extent of prompt modifications, significantly undermining privacy as ostensibly secure prompts may inadvertently produce memorized outputs. Conversely, our approach employs a semantic prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.47, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.485, + 0.414 + ], + "angle": 0, + "content": "Figure 5. Both the baseline approach and our method effectively inhibit Stable Diffusion from producing memorized images by minimizing the magnitude in the initial inference step for prompt \"The No Limits Business Woman Podcast\". Nonetheless, our semantic prompt search (SS) strategy, which results in \"The Empowered Business Woman's Podcast\", retains considerably more utility, as demonstrated by the substantially higher CLIP similarities between the original user text prompt and both our (1) modified prompt, and (2) ultimately generated image. Qualitative results further corroborate this finding. \\( m_{T-1} \\) is the first-step magnitude, \\( CLIP_{txt} \\) and \\( CLIP_{img} \\) refer to the original prompt's CLIP similarity with the modified prompt and the generated image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.508 + ], + "angle": 0, + "content": "search to inherently preserve utility, leveraging the capabilities of Large Language Models (LLM) to navigate the linguistic space for suitable alternatives. By pinpointing a substitute prompt in the language space, we keep the semantic similarity and obviate the need to trade privacy for enhanced utility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.689 + ], + "angle": 0, + "content": "Analysis: Enhancing Privacy with Minimum Utility Cost. This is also guaranteed by design. As discussed in Sec. 3.2, the baseline requires significant utility sacrifices to gain extra privacy. In contrast, the utility impact associated with our semantic prompt search is limited to the semantic discrepancy between the original user prompt and the LLM-derived alternative, which is much smaller. Furthermore, when the search yields an alternative prompt devoid of memorization risks, it effectively represents a significant boost in privacy for a minimal cost of utility. Geometrically as in Fig. 4 (c), this means shifting the prediction from \\(\\bigcirc\\) to \\(\\mathbb{D}^{\\prime}\\) with minimum utility cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 demonstrate SS's superiority over the baseline using a widely-investigated memorized prompt: \"The No Limits Business Woman Podcast\". While the baseline's prompt engineering (PE) strategy successfully reduces the magnitude to 1.15 and prevents memorization, it results in a significant semantic deviation from the original prompt, as indicated by low CLIP similarity scores for both the engineered text prompt and the generated image. In contrast, our SS strategy not only reduces the magnitude further to 0.78, effectively preventing memorization, but also significantly improves text alignment. By navigating the linguistic space, our method finds \"The Empowered Business Woman's Podcast\" as a privacy-preserving yet semantically aligned alternative." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.831, + 0.108 + ], + "angle": 0, + "content": "3.4. Synergy Effect of the Two Strategies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.431 + ], + "angle": 0, + "content": "Analysis: Tackling the Limitation of Semantic Search (SS). In an extreme scenario, SS might encounter a limitation where all \\( n_s \\) searched alternatives result in magnitudes exceeding the set threshold, failing to meet the required privacy standards. This situation can be geometrically represented by a movement from point \\( \\mathsf{O} \\) to \\( \\mathsf{B}' \\) in Fig. 4 (c). However, it's important to note that even in such cases, the transition to a lower magnitude contour, where \\( \\mathsf{B}' \\) is located, still offers a certain degree of privacy enhancement without additional cost. Further advancements in privacy (e.g., moving from \\( \\mathsf{B}' \\) to \\( \\mathsf{D}' \\)) can be achieved through our synergistic application of the PR strategy, which incurs a smaller utility cost compared to the baseline method. The right side of Fig. 8 (in Supplementary Material) showcases the synergistic benefit of utilizing the PR strategy to compensate for the SS strategy's limitations. In this example, although SS may not secure an alternative prompt with a sufficiently reduced magnitude, lowering it from the original 7.48 to 6.02 has proved to be adequate for PR to leverage the original prompt as the new anchor point for steering the generation away from memorized content effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.432, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Analysis: Addressing the Privacy-Utility Dilemma. For the baseline strategy [36], the dilemma in balancing the privacy-utility trade-off through adjusting the \\(\\lambda\\) hyperparameter manifests in two ways: opting for a higher \\(\\lambda\\) to bolster utility inadvertently raises the False Negative (FN) rate, leading to potentially unsafe prompts being misclassified as safe and thus compromising privacy; conversely, selecting a lower \\(\\lambda\\) to enhance privacy increases the False Positive (FP) rate, resulting in unnecessary modifications to already safe prompts, thereby reducing utility. Prompt re-anchoring is tailored to mitigate the privacy repercussions of False Negatives (FN) encountered when users aim to augment utility with a higher \\(\\lambda\\). It achieves this by actively directing generation away from memorized content, as elucidated in Sec. 3.2 and depicted in Fig. 4 as a divergence from \\(\\bigcirc\\). This approach thus optimizes the privacy-utility trade-off under such conditions. Semantic prompt search, on the other hand, aims to alleviate the utility decline tied to False Positives (FP) when enhancing privacy with a lower \\(\\lambda\\). It conducts a linguistically sensitive search that avoids semantic degradation, consequently refining the privacy-utility balance in these scenarios. Employing both strategies synergistically ensures an enhanced privacy-utility trade-off, regardless of whether the user's priority lies with maximizing privacy or utility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.823, + 0.646, + 0.84 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.697, + 0.866 + ], + "angle": 0, + "content": "4.1. Evaluation settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Although our method is designed to enhance the approach in [36], it can be easily adapted to any recent advance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8187" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.455 + ], + "angle": 0, + "content": "ment of the detection strategy by simply replacing \\( m_{T-1} \\) in Eq. (12) with alternative detection signals, making PRSS highly practical and adaptable for future improvements. For example, a contemporary work [8] introduced the masked magnitude \\( m_t' \\) (see Eq. (8)), achieving new state-of-the-art detection performance. The mitigation strategy proposed by [8] also uses prompt engineering based on the upgraded detection signal \\( m_t' \\), offering two key advantages: (1) \\( m_t' \\) provides a more accurate detection trigger for initiating the mitigation strategy, (2) it serves as a more effective loss function than unmasked magnitude, yielding better gradients for optimizing prompts. These improvements make it the current state-of-the-art in mitigation. PRSS can seamlessly integrate this enhanced detection signal by replacing \\( m_{T-1} \\) with \\( m_{T-1}' \\) in Eq. (12), producing our best results to date. For a fair comparison, we re-implement the mitigation strategy from [8] and evaluate its performance alongside PRSS. This comparison not only highlights PRSS's capability to surpass the most recent state-of-the-art but also serves as an ablation study, assessing PRSS's impact since both methods use the same detection strategy to trigger mitigation. Additionally, we produce a version of PRSS that uses \\( m_t \\) as the detection signal, aligning with [36], to further demonstrate PRSS's effectiveness in comparison to [36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.686 + ], + "angle": 0, + "content": "We follow the baselines [8, 36] by experimenting on Stable Diffusion v1-4, using its corresponding prompt dataset from [34], which contains 500 prompts, over 300 of which are classified as memorized and organized by different memorization types. To provide a comprehensive view, we adopt the settings from [8] and report results for both global and local memorization prompts separately. We use the default hyperparameters from [36] to re-implement the detection and prompt engineering mitigation strategies from [36] and [8]. For semantic prompt search with LLMs, we use the GPT-4 model via the OpenAI API, instructing it to generate up to \\( n_s = 25 \\) semantically similar prompts for each user prompt. The associated inference time and cost are minimal, with each alternative prompt generated in approximately 0.9 seconds, costing around $0.02." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Following standard practice in memorization literature, we use the CLIP score to measure the alignment between the generated output and the input user prompt, serving as our utility metric. For privacy metrics, which quantify the degree of memorization, most studies rely on SSCD-based scores (see Eq. (5)), reporting metrics such as the average score, the 95th percentile score, and the percentage of memorized generations, with a threshold of 0.5 commonly indicating memorization. Additionally, [8] proposes a localized similarity metric (LS, see Eq. (6)) that measures similarity between the generated image and corresponding training image only within the localized memorization region. We present comparative results using the widely adopted SSCD metric in Fig. 6 and Fig. 7, while other metrics, including" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.095, + 0.716, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.096, + 0.897, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.33 + ], + "angle": 0, + "content": "Figure 6. Comparison with baselines. Across different detection signals (\\(m\\) and \\(m'\\)), PRSS consistently achieves superior privacy-utility trade-offs. The improvement is particularly pronounced in global memorization compared to local memorization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.907, + 0.368 + ], + "angle": 0, + "content": "LS, are detailed in the Supplementary Material for a more comprehensive evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.378, + 0.761, + 0.394 + ], + "angle": 0, + "content": "4.2. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.909, + 0.704 + ], + "angle": 0, + "content": "As shown in Fig. 6, PRSS consistently outperforms prompt engineering (PE) in achieving better privacy-utility trade-offs, regardless of the detection signal used (whether \\( m \\) or \\( m' \\)). This is visually evident as the red and green lines (PRSS) lie further in the bottom-right direction compared to the yellow and blue lines (PE) for both global and local memorization scenarios. Notably, PRSS demonstrates astonishingly large improvements in global memorization (from the blue line to the green line and from the yellow line to the red line), while [8] achieves only marginal gains over [36] in this case (from the blue line to the yellow line and from the green line to the red line). For local memorization, although the margin of improvement is smaller, PRSS still achieves a significant improvement over PE. Specifically, even using the inferior detection signal \\( m \\), PRSS (green line) achieves comparable performance with the method proposed by [8] (yellow line), which leverages the superior signal \\( m' \\). Equipped with \\( m' \\), PRSS can achieve the state-of-the-art results (red line). This highlights PRSS's robust capability to leverage advancements in detection signals." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.681, + 0.727 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fig. 7 highlights the necessity of both components in PRSS, as evidenced by the red line lying further in the bottom-right direction compared to the yellow and green lines for both global and local memorization scenarios. The synergistic effect of combining PR and SS as discussed in Sec. 3.4, is particularly evident in the local memorization scenario, which aligns with [8]'s observation that local memorization poses greater challenges for existing mitigation strategies. Using PR alone effectively enhances privacy but results in suboptimal utility, while SS alone improves utility but achieves limited privacy protection. Together, they" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8188" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.095, + 0.472, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 7. Ablation studies of PRSS components. Using PR alone results in low utility, while SS alone cannot ensure privacy. Combining PR and SS achieves a synergistic effect, delivering the best privacy-utility trade-off." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.414 + ], + "angle": 0, + "content": "leverage each other's strengths, achieving the best privacy-utility trade-off. For global memorization, either PR or SS alone significantly outperforms the baseline. However, their combination further enhances the trade-off, underscoring the robustness of PRSS in tackling both global and local memorization challenges." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.415, + 0.235, + 0.43 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Memorization in Generative Models The phenomenon of generative models memorizing training data has been extensively studied across domains, including language models [3, 4, 15, 16, 18, 20], image generation [2, 10, 11, 21, 22, 33], and video generation [9, 26], due to the significant privacy and legal risks it poses [5, 14, 29, 37]. Specifically, in diffusion models, [6, 31, 34] first validated the memorization issue on various models including DDPMs [13, 23] and Stable Diffusion [28] on datasets include Oxford Flowers [24], CelebA [19], CIFAR-10 [17] and LAION [30]. Further analysis by [32] revealed that Stable Diffusion's text-conditioning exacerbates overfitting, making its memorization of LAION particularly problematic. To better understand memorization, [34] categorized the phenomenon as Matching Verbatim, Retrieval Verbatim, and Template Verbatim based on the cause, distinguishing between one-to-one and many-to-many mappings in text-image pairs, while [8] classified memorization into global and local cases, demonstrating the practical significance of this distinction. In this work, we focus on Stable Diffusion due to its widespread use, open-source nature, and real-world legal challenges [29]. We adopt the framework of [8] to analyze both global and local memorization for a more nuanced evaluation of our proposed strategies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Detection Strategies for Diffusion Memorization. Early detection methods required generating large image sets to estimate density [6] or comparing generations with the entire training dataset [31], which were computationally intensive. Recent approaches [27, 36] introduced more efficient one-step detection strategies. [27] detected memorization by observing less concentrated cross-attention scores" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "for memorized generations than non-memorized ones. [36] proposed using the abnormally high magnitude of text-conditional noise predictions as a detection signal. Building on this, [8] incorporated localized memorization masks into the magnitude computation, offering a more accurate detection signal and achieving state-of-the-art performance. Our proposed PRSS approach leverages these detection mechanisms to trigger mitigation strategies. Notably, PRSS can adapt to any detection method by substituting the indicator variable condition in its framework. We employ the state-of-the-art method from [8] for optimal results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Mitigation Strategies for Diffusion Memorization. Mitigation strategies have evolved rapidly. Early methods, such as retraining models on de-duplicated data [6], were computationally expensive and showed limited effectiveness. [32] introduced randomness into text prompts or embeddings but indiscriminately affected non-memorized prompts, reducing overall utility. To improve on it, [7] proposed to compute the generation's similarity with the closest training image to engineer the image latent using the classifier guidance framework. However, this relies on computationally expensive nearest-neighbor searches across the training dataset. [36] proposed to conduct prompt engineering using the magnitude of text-conditional prediction as the loss, and [27] proposed to re-weight cross-attention scores for all tokens in the prompt. They offer improved utility-privacy trade-offs and achieve comparable performance. Most recently, [8] extended these efforts by leveraging localized magnitudes as loss functions for prompt engineering, achieving the best results to date. Our investigation reveals the limitations in the prevailing method's privacy/utility balance. In response, we introduce new mitigation strategies that achieve more favorable trade-offs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.596, + 0.785, + 0.611 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This paper introduces a novel PRSS method that consists of prompt re-anchoring and semantic prompt search as deliberately designed strategies to mitigate memorization in diffusion models. Empirical results show that PRSS demonstrates superior privacy-utility trade-offs over existing methods. Beyond its effectiveness, PRSS is simple and efficient to implement. It requires only minor modifications to the CFG equation to update noise predictions during inference, seamlessly integrating into the inference pipeline without additional training, fine-tuning, or reliance on training data. The efficacy of each component is validated through comprehensive analysis and ablation studies. A current limitation is that the performance of PRSS, as well as baseline mitigation strategies, relies heavily on the accuracy of the detection mechanism. When it fails to trigger accurately, they all revert to the standard Stable Diffusion process, compromising privacy preservation. Thus, improving the detection mechanism's accuracy in future works would orthogonally bolster the effectiveness of PRSS." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8189" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.144 + ], + "angle": 0, + "content": "This work was supported in part by the Australian Research Council under Projects DP210101859 and FT230100549." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.46, + 0.213 + ], + "angle": 0, + "content": "[1] Midjourney. https://www.midjourney.com/. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[2] Sanjeev Arora, Andrej Risteski, and Yi Zhang. Do gans learn the distribution? some theory and empirics. In International Conference on Learning Representations, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In USENIX Security Symposium, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[4] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang, Nicolas Papernot, Andreas Terzis, and Florian Tramér. The privacy onion effect: Memorization is relative. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.512 + ], + "angle": 0, + "content": "[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In USENIX Security Symposium, pages 5253-5270, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.515, + 0.483, + 0.542 + ], + "angle": 0, + "content": "[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In CVPR, 2024. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[8] Chen Chen, Daochang Liu, Mubarak Shah, and Chang Xu. Exploring local memorization in diffusion models via bright ending attention. *ICLR*, 2025. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[9] Chen Chen, Enhuai Liu, Daochang Liu, Mubarak Shah, and Chang Xu. Investigating memorization in video diffusion models. ICLR 2nd Workshop on Navigating and Addressing Data Problems for Foundation Models (DATA-FM), 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] Yunhao Chen, Xingjun Ma, Difan Zou, and Yu-Gang Jiang. Towards a theoretical understanding of memorization in diffusion models. arXiv preprint arXiv:2410.02467, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] H. H. Jiang, L. Brown, J. Cheng, M. Khan, A. Gupta, D. Workman, A. Hanna, J. Flowers, and T. Gebru. AI Art and Its Impact on Artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374. ACM, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[15] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. Large language models struggle to learn long-tail knowledge. arXiv preprint arXiv:2211.08411, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[16] Nikhil Kandpal, Eric Wallace, and Colin Raffel. Deduplicating training data mitigates privacy risks in language models. In International Conference on Machine Learning, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[18] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the Association for Computational Linguistics, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.908, + 0.358 + ], + "angle": 0, + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[20] Fatemehsadat Mireshghallah, Anish Uniyal, Tianhao Wang, David Evans, and Taylor Berg-Kirkpatrick. Memorization in nlp fine-tuning methods. arXiv preprint arXiv:2205.12506, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[21] Amir Naseh, Jaemin Roh, and Amir Houmansadr. Memory triggers: Unveiling memorization in text-to-image generative models through word-level duplication. arXiv preprint arXiv:2312.03692, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[22] Ali Naseh, Jaechul Roh, and Amir Houmansadr. Understanding (un)intended memorization in text-to-image generative models. arXiv preprint arXiv:2312.07550, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.904, + 0.546 + ], + "angle": 0, + "content": "[23] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[24] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[25] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, pages 14532-14542, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[26] Aimon Rahman, Malsha V. Perera, and Vishal M. Patel. Frame by familiar frame: Understanding replication in video diffusion models. arXiv preprint arXiv:2403.19593, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[27] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. ECCV, 2024. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[29] Joseph Saveri and Butterick Matthew. Stable diffusion litigation, 2023. 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8190" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[31] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In CVPR, pages 6048-6058, 2023. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. In NeurIPS, 2023. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[33] Patrick Tinsley, Adam Czajka, and Patrick Flynn. This face does not exist... but it might be yours! identity leakage in generative models. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[34] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[35] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Advances in Neural Information Processing Systems, pages 51008-51025. Curran Associates, Inc., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[36] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In ICLR, 2024. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[37] Chiyuan Zhang, Daphne Ippolito, Katherine Lee, Matthew Jagielski, Florian Tramér, and Nicholas Carlini. Counterfactual memorization in neural language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.127, + 0.518, + 0.135, + 0.527 + ], + "angle": 0, + "content": "8" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8191" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_origin.pdf b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..87d529ad71ced22f87931bfba81adb1f8beff267 --- /dev/null +++ b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/0d123934-7ce7-45d9-ad3f-707ece8265c2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:711b54aa789e8995c0d90b819de6b92ec30276cba2cafceb2dc0771137735bd8 +size 2029430 diff --git a/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/full.md b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2be82df1a0b62ff7301144de865ee9ca1ec73a63 --- /dev/null +++ b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/full.md @@ -0,0 +1,309 @@ +# Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models + +Chen Chen $^{1}$ Daochang Liu $^{2}$ Mubarak Shah $^{3}$ Chang Xu $^{1}$ + +$^{1}$ School of Computer Science, Faculty of Engineering, The University of Sydney, Australia $^{2}$ School of Physics, Mathematics and Computing, The University of Western Australia, Australia $^{3}$ Center for Research in Computer Vision, University of Central Florida, USA + +{cche0711@uni., c.xu@}sydney.edu.au daochang.liu@uwa.edu.au shah@crcv.ucf.edu + +# Abstract + +Text-to-image diffusion models have demonstrated remarkable capabilities in creating images highly aligned with user prompts, yet their proclivity for memorizing training set images has sparked concerns about the originality of the generated images and privacy issues, potentially leading to legal complications for both model owners and users, particularly when the memorized images contain proprietary content. Although methods to mitigate these issues have been suggested, enhancing privacy often results in a significant decrease in the utility of the outputs, as indicated by text-alignment scores. To bridge the research gap, we introduce a novel method, PRSS, which refines the classifier-free guidance approach in diffusion models by integrating prompt re-anchoring (PR) to improve privacy and incorporating semantic prompt search (SS) to enhance utility. Extensive experiments across various privacy levels demonstrate that our approach consistently improves the privacy-utility trade-off, establishing a new state-of-the-art. + +# 1. Introduction + +Leveraging classifier-free guidance (CFG) [12], text-to-image diffusion models like Stable Diffusion [28] and Midjourney [1] are now capable of generating highly realistic images that closely align with user-provided text prompts. This capability has propelled their popularity, leading to widespread use and distribution of their generated images. However, recent research [6, 8, 31, 34] has revealed a critical issue: these models can memorize training data, leading them to reproduce parts of images, such as foregrounds or backgrounds (local memorization, see Fig. 1), or even entire images (global memorization, see Fig. 2) during inference, instead of generating genuinely novel content. When the training data includes sensitive or copyrighted material, these memorization issues can infringe on copyright laws without notifying either the model's owners or users or the copyright holders of the replicated content. The risk is es + +ppecially substantial given the extensive use of these models and their reliance on massive datasets, such as LAION [30], which contain billions of web-scale images and are impractical to thoroughly review or filter manually. This risk is further highlighted by real-world cases, where several artists have filed lawsuits, arguing that models like Stable Diffusion act as “21st-century collage tools” that remix their copyrighted works, implicating Stability AI, DeviantArt, Midjourney, and Runway AI. Recognizing the potential for unauthorized reproductions, Midjourney has even banned prompts containing the term “Afghan” to prevent the generation of the copyrighted Afghan Girl photograph. Yet, as [35] demonstrates, such restrictions alone are insufficient to fully prevent the reproduction of copyrighted images. This underscores the urgent need for timely effective mitigation strategies to address these concerns. + +In response to these legal challenges, recent initiatives [7, 8, 27, 32, 36] have focused on developing strategies to minimize memorization, achieving notable success. These approaches vary in scope, with some targeting the model's training phase and others making adjustments during inference. While training-phase strategies are theoretically viable, they lack practical significance as they predominantly experiment on a small $10\mathrm{k}$ subset of the entire LAION5B dataset due to the impracticality of fine-tuning on the full dataset — a constraint stemming from computational limitations. In contrast, inference-phase strategies, which involve modifications only during inference, offer computational efficiency and adaptability, as they require no additional training or fine-tuning. This makes them highly applicable across pre-trained diffusion models, regardless of specific training configurations. Despite their promise, inference-time strategies can also suffer from computational constraints. For example, the approach proposed by [7] involves searching the entire LAION dataset for the nearest neighbor to optimize the generated latent representation during each denoising step. Other inference-time strategies [8, 27, 32, 36] that avoid nearest neighbor search + +![](images/2733b0e4ece0630b5c245b132c36f86a5ccb6afcbf18f36a0b53bae8f1da8e28.jpg) +Figure 1. Examples of Local Memorization: Stable Diffusion generations (columns 2-4) can replicate local regions of training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7). + +![](images/6fcaa2b5f9d426a1352a61170f978de7addc41e82e072566a3906e17f2493652.jpg) +Figure 2. Examples of Global Memorization: Stable Diffusion generations (columns 2-4) can replicate the entire training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7). + +demonstrate high efficiency but only moderate effectiveness, leaving a substantial gap in optimizing the privacy-utility trade-off. In particular, these methods often require significant compromises in text-image alignment to reduce memorization and conversely, struggle to sufficiently suppress memorization when trying to preserve alignment. + +To address these limitations, we carefully analyze why existing strategies struggle to achieve an optimal privacy-utility trade-off, uncovering critical insights that the two inputs (text-conditional and unconditional predictions) used in the CFG framework are both suboptimal for mitigating memorization, even though they are standard for most other tasks in Stable Diffusion models. Specifically, we find that the text-conditional term in CFG, when optimized for privacy by prompt engineering, can significantly reduce text alignment, indicating the need for alternative methods to find privacy-preserving yet intent-aligned text-conditioning. Additionally, we observe that the unconditional term in CFG is less effective as an anchor for privacy + +preservation, suggesting that alternative anchoring points that enhance privacy are necessary. Building on these insights, we introduce PRSS, a novel method that incorporates two complementary strategies to refine the widely used CFG approach for improved memorization mitigation. Our method includes (1) prompt re-anchoring, designed to preserve high privacy with a smaller utility cost, which re-anchors the unconditional component in CFG using the memorized prompt to consistently guide generations away from memorization during the inference process; (2) semantic prompt search, designed to preserve high utility with minimal privacy compromise, which optimizes the prompt away from the memorized one in the language space while keeping similar semantics. Our experimental results reveal that employing both strategies synergistically can consistently outperform baseline methods in achieving superior privacy-utility trade-offs at different privacy levels. In particular, this results in higher text alignment (utility) and substantially fewer memorized generations (improved privacy). Furthermore, ablation studies validate the efficacy of each component within our strategy. From an implementation standpoint, our strategies are simple, efficient, and effective, necessitating only adjustments to the existing CFG equation for updating the noise prediction during inference. This integration is seamless, avoiding disruption to the inference workflow and eliminating the need for additional re-training, fine-tuning, or searching over the training data. + +# 2. Preliminaries + +# 2.1. Classifier-Free Guidance in Diffusion Models + +Diffusion models [13, 23] are popular generative models that consist of a forward diffusion process that consistently adds Gaussian noise to real images $x_0$ for $T$ steps based on the following conditional function, resulting in $x_T \sim \mathcal{N}(0, \mathbf{I})$ : + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $x_{t}$ represents the noised version of $x_{0}$ after adding $t$ noise, and $\{\beta_t\}_{t = 1}^T$ is the schedule that controls the magnitude of added noise at time $t$ . + +This is followed by the reverse process, in which a de-noiser model is trained to predict the distribution (i.e., mean $\mu_{\theta}(x_t)$ and variance $\sigma_{\theta}^{2}(x_{t})\mathbf{I}$ ) of a cleaner version of $x_{t}$ , denoted as $x_{t - 1}$ , at any given timestep $t$ : + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}\right), \sigma_ {\theta} ^ {2} \left(x _ {t}\right) \mathbf {I}\right). \tag {2} +$$ + +During training, the objective function is equivalent to have the denoiser $\epsilon_{\theta}(x_t,y,t)$ predict the noise $\epsilon_{t}$ to be removed, rather than the image $x_{t - 1}$ at timestep $t$ conditioned on $y$ which may represent a user-provided text prompt: + +$$ +\mathcal {L} = \mathbb {E} _ {t \in [ 1, T ], \epsilon \sim \mathcal {N} (0, \mathbf {I})} \left[ \| \epsilon_ {t} - \epsilon_ {\theta} (x _ {t}, y, t) \| _ {2} ^ {2} \right]. \tag {3} +$$ + +Classifier-free guidance (CFG) [12] is a widely used technique for balancing fidelity and diversity in the generated outputs of diffusion models, and it can be easily applied during inference. CFG operates by linearly guiding the noise prediction from the unconditional prediction, $\epsilon_{\theta}(x_t,t)$ towards the conditional prediction, $\epsilon_{\theta}(x_t,y,t)$ , with a scaling factor $s > 1$ to control the degree of guidance. For simplicity, removing the timestep $t$ from the notation, the adjusted noise prediction can be defined as: + +$$ +\hat {\epsilon} \leftarrow \epsilon_ {\theta} \left(x _ {t}\right) + s \cdot \left(\epsilon_ {\theta} \left(x _ {t}, y\right) - \epsilon_ {\theta} \left(x _ {t}\right)\right). \tag {4} +$$ + +# 2.2. Memorization in Diffusion Models + +Definition and Metrics. A widely adopted practical definition of memorization refers to generated images exhibiting high object-level similarity to training data [31]. This includes the concept of reconstructive memory in diffusion models, where foreground and background objects from training images are combined, potentially with modifications such as shifting, scaling, or cropping. These reconstructions may not match any training image pixel-for-pixel but still closely resemble specific training images at the object level. The standard metric for evaluating memorization [7, 8, 27, 31, 32, 36] is designed base on a pre-trained Self-Supervised Copy Detection (SSCD) model [25]: + +$$ +\operatorname {S S C D} (\hat {x}) = \max _ {x _ {i} \in \mathcal {D}} \operatorname {C o s i n e} \left(\mathbf {E} _ {S S C D} (\hat {x}), \mathbf {E} _ {S S C D} (x _ {i})\right), \tag {5} +$$ + +where for each generated image $\hat{x}$ , cosine similarity is computed between its SSCD embedding $\mathbf{E}_{SSCD}(\hat{x})$ and the embeddings $\mathbf{E}_{SSCD}(x_i)$ of all training images $x_i$ in the dataset $\mathcal{D}$ . The maximum similarity serves as the final SSCD score. + +Most recently, [8] proposed a method to localize memorized regions in generated images, arguing that only memorized regions pose privacy risks, making local memorization a more meaningful concept. They introduced an efficient approach to synthesize a mask that highlights these regions without requiring access to training data. Using this mask, they proposed a new localized metric based on both object-level SSCD and pixel-level Euclidean distance: + +$$ +L S (\hat {x}, x) = - \mathbb {1} _ {\mathrm {S S C D} > 0. 5} \cdot \| (\hat {x} - x) \circ \mathbf {m} \| _ {2}, \tag {6} +$$ + +where $\mathbf{m}$ is the local memorization mask and $\mathbb{1}_{\mathrm{SSCD} > 0.5}$ is an indicator function that signifies whether the SSCD score exceeds the 0.5 threshold, which is commonly used to distinguish memorized from non-memorized content. + +Memorization Detection. A recent work [36] introduces an innovative approach for detecting potential memorized generations during inference by leveraging the magnitude of the difference between text-conditional $\epsilon_{\theta}(x_t, e_p)$ and unconditional $\epsilon_{\theta}(x_t, e_\phi)$ noise predictions: + +$$ +m _ {t} = \left\| \epsilon_ {\theta} \left(x _ {t}, e _ {p}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right) \right\| _ {2}, \tag {7} +$$ + +where memorized generations consistently exhibit larger magnitudes than non-memorized ones, making this magnitude a strong signal for detection. Furthermore, accurate + +![](images/a3778257d9c64f428098939618dc4033c07215cc89161ab6972db3949f644b30.jpg) +Figure 3. Illustration of the logic behind how $\lambda$ in Eq. (10) governs the privacy-utility trade-off. Currently, the privacy-utility trade-off is governed solely by how extensively the prompt is modified. + +detection can be achieved efficiently by evaluating the magnitude $m_{T-1}$ at the first denoising step (timestep $T - 1$ , highlighting the method's computational efficiency. + +The contemporary work [8] extends this detection method by incorporating their proposed local memorization mask $\mathbf{m}$ into Eq. (7) to compute a masked magnitude, achieving state-of-the-art performance: + +$$ +m _ {t} ^ {\prime} = \left\| \left(\varepsilon_ {\theta} \left(x _ {t}, e _ {p}\right) - \varepsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right)\right) \circ \mathbf {m} \right\| _ {2} / \left(\frac {1}{N} \sum_ {i = 1} ^ {N} m _ {i}\right), \tag {8} +$$ + +where $\mathbf{m}$ is the local memorization mask and $\frac{1}{N} \sum_{i=1}^{N} m_i$ is used to normalize the result by the mean of $\mathbf{m}$ . + +# 3. Method + +We begin with a detailed analysis in Sec. 3.1 to uncover the key challenges that hinder existing strategies from achieving an optimal privacy-utility trade-off. Building on these insights, we evolve the baseline step by step with our proposed prompt re-anchoring (PR) strategy in Sec 3.2, which focuses on improving privacy, and semantic prompt search (SS) in Sec 3.3, which enhances utility. Finally, in Sec. 3.4, we analyze how the combination of PR and SS achieves a synergistic effect, offering the best privacy-utility trade-off. + +# 3.1. Trade-off Analysis of Existing Strategies + +Our method is built upon the widely-used classifier-free guidance (CFG) by treating the input text prompt from user $e_p$ as the condition $y$ in Eq. (4) and rewriting the unconditional noise as a null condition $e_\phi$ , + +$$ +\hat {\epsilon} \leftarrow \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right) + s \left(\epsilon_ {\theta} \left(x _ {t}, e _ {p}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right)\right). \tag {9} +$$ + +The one-step prompt engineering (PE) based memorization mitigation method proposed by [36] then counteract memorization by reformulating Eq. (9) as: + +$$ +\begin{array}{l} \hat {\epsilon} \leftarrow \left[ \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right) + s \left(\epsilon_ {\theta} \left(x _ {t}, e _ {p}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right)\right) \right] \mathbb {1} _ {\{m _ {T - 1} < \lambda \}} \\ + \left[ \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right) + s \left(\epsilon_ {\theta} \left(x _ {t}, e ^ {*}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right)\right) \right] \mathbb {1} _ {\left\{m _ {T - 1} > \lambda \right\}}, \tag {10} \\ \end{array} +$$ + +where $m_{T - 1} = \| \epsilon_{\theta}(x_{T - 1},e_p) - \epsilon_{\theta}(x_{T - 1},e_\phi)\| _2$ represents the magnitude at the initial step $T - 1$ as per Eq. (7), $\epsilon_{\theta}(x_t,e_p)$ and $\epsilon_{\theta}(x_{t},e^{*})$ are the noise predictions conditioned on the user prompt embedding $e_p$ and engineered prompt embedding $e^*$ (optimized based on the gradient computed from the first-step magnitude $m_{T - 1}$ ), and + +![](images/0212fde5e1de8ac4ebb3f459c6cbafb2045d8f900b37c2573da45a72a39af262.jpg) +a) Privacy-Utility Trade-Off in Baseline + +![](images/d36f7bb43ddcb6b429d029441996088fcf5d52e0f72545ec8f86abda62de1e9d.jpg) +b) Enhancing Privacy with Prompt Re-Anchoring (PR) +$\mathbf{SIM}(\mathbb{O},\mathbb{B}) > \mathbf{SIM}(\mathbb{O},\mathbb{C}) > \mathbf{SIM}(\mathbb{O},\mathbb{D})$ + +![](images/69ef1818c0a04ced470b7bb6b8cd221b10dfd10a64c16d03cb3dad59a6c838c1.jpg) +c) Enhancing Utility with Semantic Search (SS) +$\mathrm{SIM}(\mathbb{O},\mathbb{C}) < \mathrm{SIM}(\mathbb{O},\mathbb{C}^{\prime})$ +$\mathbf{SIM}(\mathbb{O},\mathbb{D}) < \mathbf{SIM}(\mathbb{O},\mathbb{D}^{\prime})$ +SIM CLIP Similarity (Utility) + +![](images/96271cda170299881d18afaffb0f27f525f9fa0b02cffc9cfd0d9563eb17ebc9.jpg) +Contour Line of Equal Privacy + +![](images/2dc232078a2a1003ca08f48f3726f87b2632c37018b5000d69d8ab2879e998c7.jpg) +$\mathbf{SIM}(\mathbb{O},\mathbb{B}) > \mathbf{SIM}(\mathbb{O},\mathbb{C}) > \mathbf{SIM}(\mathbb{O},\mathbb{D})$ +$\rightarrow$ Baseline Guidance + +![](images/ab9957472abcddd87421923a6345bf50c88af6af601cdbb38883cef9d3ba5c91.jpg) +Our PR Guidance + +![](images/177949909573a9f6dd18342974a45c2068fb5e2b034d16c2b039580587515fd5.jpg) +Our PRSS Guidance + +![](images/d743170ccf694ad07aa47bf9ba3e381adfb09ed86a99bf87f302f9d3f2b8cddc.jpg) +Unconditional Prediction + +![](images/eb41483b64ef842a229262c827e6eea42cad7c15d5a22a22637b115eb02aae4e.jpg) +$e_p$ -conditional prediction + +![](images/09116e76c0a1a6aaf5a0c28da69766ee7c6a01ac6602a11296178d6a38e2e370.jpg) +$e^*$ -conditional prediction with different privacy level + +![](images/e1ddcd0dc93036430c7ca620bcf16651170f2806ab61fc35b92f4abd6a63e939.jpg) +$e_p^{ss}$ -conditional prediction with different privacy level +Figure 4. Illustration of our proposed strategy (PRSS) compared to the baseline prompt engineering (PE) method. a) The privacy-utility trade-off in the baseline method, where the privacy increases at the cost of utility from B to C to D. b) Our proposed PR guidance, i.e., the green arrow, drives the image out of the low-privacy region more efficiently than the baseline guidance, i.e., the black arrows. c) Our proposed SS can enhance utility, by finding points $\mathrm{B}^{\prime}\mathrm{C}^{\prime}\mathrm{D}^{\prime}$ that are at the same level of privacy as BCD but with better utility. In summary, using the proposed PR and SS jointly can enhance the privacy-utility trade-off to mitigate memorization in diffusion models. + +$\epsilon_{\theta}(x_t, e_{\phi})$ is the unconditional noise prediction. This strategy detects potential memorization for a given input prompt based on the magnitude signal $m_{T-1}$ with a threshold $\lambda$ , and undertakes soft prompt engineering (i.e., optimizes the prompt embeddings) if potential memorization is detected. In this context, if $m_{T-1}$ is under $\lambda$ , the prompt is considered safe from memorization, and the generation proceeds with the original user prompt $e_p$ , mirroring Stable Diffusion's output. Conversely, if $m_{T-1}$ is above $\lambda$ , the prompt is then identified as likely to produce memorized images, thus $e_p$ is optimized to and replaced with an engineered prompt $e^*$ in the conditional prediction while keeping the unconditional prediction $\epsilon_{\theta}(x_t, e_{\phi})$ as is following the classical CFG equation. Specifically, $e^*$ is optimized to reduce the first-step magnitude $m_{T-1}$ using gradient descent and the optimization halts once the magnitude $m_{T-1}$ falls below $\lambda$ , marking the engineered prompt as less prone to memorization. Both detection and prompt optimization occur only during the initial inference step $T-1$ , rather than throughout the entire inference process, enhancing efficiency. + +Privacy-utility trade-off. Eq. (10) suffers from a severe privacy-utility trade-off as in Fig. 3. Opting for a lower $\lambda$ improves the privacy (risk of memorization) of outputs at the expense of utility (text-alignment). It requires more substantial optimization of the input prompt, aiming to achieve a magnitude $m_{T - 1}$ lower than $\lambda$ , thereby enhancing privacy. However, such extensive modifications to the prompt may lead to a loss of its original meaning, compromising utility significantly. Conversely, a higher $\lambda$ upholds the utility of generated outputs but at the cost of privacy. + +Geometric analysis. Fig. 4 (a) visualizes the privacy-utility trade-off in Eq. (10) as a function of $\lambda$ , where vary + +ing levels of $\lambda$ define distinct contours corresponding to target magnitudes and privacy levels during prompt engineering (PE). Outer contours represent greater prompt modifications, which enhance privacy. This optimization shifts the prompt embedding from the original memorized prompt, $e_{p}$ , located at point $\mathsf{O}$ , to the engineered prompt, $e^{*}$ , at points B, C, or D. The number of PE iterations determines the distance of the prompt from point $\mathsf{O}$ : more iterations lead to greater divergence, achieving higher privacy levels. The process halts once the desired privacy level is attained. + +Overview of our improvements. Our trade-off analysis reveals that existing strategies rely solely on prompt modification methods such as prompt engineering to enhance privacy, which is suboptimal and significantly compromises utility. This highlights an urgent need to propose alternative methods to improve privacy and preserve utility. To address this, we introduce two meticulously crafted mitigation strategies that only require refining the guidance components in CFG to achieve a superior privacy-utility tradeoff: (1) prompt re-anchoring (PR) and (2) semantic prompt search (SS). Specifically, SS enhances utility with minimal privacy cost, while PR serves as an effective alternative to improve privacy, which has the potential of costing less utility when used in combination with SS. + +# 3.2. Enhancing Privacy Through Prompt Re-Anchoring + +One of our objectives is to safeguard privacy while preserving as much utility as possible. Recall that in Eq. (10) enhancing privacy requires reducing the magnitude, which correlates with a significant decrease in utility. It is important to note that the utility loss stems not from the reduction in magnitude itself, but from the baseline's inefficient + +approach to achieving that reduction. Visually, in Fig. 4, predictions that lie on the same magnitude contour may exhibit notably different levels of utility. This observation motivates us to ponder: Is it possible to reduce magnitude through a more efficient approach (incurring a lower utility cost) compared to the baseline method of prompt engineering? This insight has spurred the creation of our prompt re-anchoring (PR) strategy, which aims to reduce magnitude by steering predictions away from memorized images. + +We argue that $e_p$ is discarded too hastily in Eq. (10) when the memorization is detected ( $m_{T-1} > \lambda$ ). Instead, recognizing $e_p$ as a valuable anchor point — its conditional generation can trace the memorization path — we use $e_p$ to re-anchor the generation process, by replacing the unconditional prediction in CFG with the $e_p$ -conditioned prediction to divert the generation away from memorization. Thereby, Eq. (10) is re-written to the following form, + +$$ +\begin{array}{l} \hat {\epsilon} \leftarrow \left[ \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right) + s \left(\epsilon_ {\theta} \left(x _ {t}, e _ {p}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right)\right) \right] \mathbb {1} _ {\left\{m _ {T - 1} < \lambda \right\}} \\ + \left[ \epsilon_ {\theta} \left(x _ {t}, e _ {p}\right) + s \left(\epsilon_ {\theta} \left(x _ {t}, e ^ {*}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {p}\right)\right) \right] \mathbb {1} _ {\{m _ {T - 1} > \lambda \}}. \tag {11} \\ \end{array} +$$ + +Analysis: Enhancing Privacy with Less Utility Cost. The baseline approach in Eq. (10) to enhancing privacy necessitates significant utility sacrifices, primarily due to the need for more extensive gradient updates during prompt optimization to obtain $e^{*}$ . This leads to substantial modifications of the original user prompt and a large gap between $e^{*}$ and $e_{p}$ , causing the prompt to diverge significantly from the user's initial intent and thus undermine the text alignment. In contrast, our prompt re-anchoring strategy employs the intuition from the original CFG approach to guide generations toward a preferred output while distancing them from an undesirable one, effectively achieved through linear interpolation between two generations. Uniquely, our strategy designates the undesired generation as the memorized one that enforces conditioning on the original user prompt $e_{p}$ , rather than relying on unconditional conditioning $e_{\phi}$ . This approach steers generations away from memorized content more efficiently with the contrastive direction, which therefore requires fewer gradient updates to obtain $e^{*}$ and thus less alternation on the prompt for the same privacy enhancement level. Our previous question is hence answered. + +Fig. 4 represents a geometric interpretation of our prompt re-anchoring strategy. Classical CFG directs $\hat{\epsilon}$ away from unconditional (point A) towards conditional predictions (O for $e_p$ , and B, C, D for $e^*$ ) as in the left figure. PR, however, guides $\hat{\epsilon}$ from the memorized prediction (O) towards less or unmemorized predictions with lower magnitude (B, C, D), enhancing privacy, as in the middle figure. + +Analysis: Continuous Memorization Diversion without Loss of Efficiency. Although [36] found that minimizing the initial magnitude $m_{T-1}$ can indirectly reduce magnitudes in later steps, we observe that the magnitude + +can still spike and memorization may still occur later in the inference process. For example, as illustrated in Fig. 8 (in Supplementary Material), while the engineered prompt achieves a low magnitude initially, it fails to ensure that this low level is maintained throughout subsequent inference steps, potentially leading back to high magnitudes indicative of Stable Diffusion's memorized outputs. To counteract this, the PR strategy can continuously divert generations away from initially identified memorized conditions (point $\bigcirc$ ) across the entirety of the inference process as shown in Fig. 8 (in Supplementary Material). The effectiveness is also evident from the qualitative appearance of the final generated image, and its quantitative similarity score as in Fig. 8 (in Supplementary Material). + +# 3.3. Enhancing Utility Through Semantic Prompt Search + +In this section, we aim to propose a strategy to enhance utility with minimal privacy cost. Recall that low utility is primarily due to the increased difference between the user prompt and the engineered prompt. We suggest conducting semantic search to find a prompt $e_p^{ss}$ semantically similar to the original user prompt $e_p$ but with a lower risk of memorization, aiming to replace the engineered prompt $e^*$ that results in the unsatisfactory trade-off. Eq. (11) is thus further re-formulated to the following form after incorporating semantic prompt search strategy in addition to the prompt re-anchoring strategy, + +$$ +\begin{array}{l} \hat {\epsilon} \leftarrow \left[ \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right) + s \left(\epsilon_ {\theta} \left(x _ {t}, e _ {p}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {\phi}\right)\right) \right] \mathbb {1} _ {\left\{m _ {T - 1} < \lambda \right\}} \\ + \left[ \epsilon_ {\theta} \left(x _ {t}, e _ {p}\right) + s \left(\epsilon_ {\theta} \left(x _ {t}, e _ {p} ^ {s s}\right) - \epsilon_ {\theta} \left(x _ {t}, e _ {p}\right)\right) \right] \mathbb {1} _ {\left\{m _ {T - 1} > \lambda \right\}}. \tag {12} \\ \end{array} +$$ + +In Eq. (12), the engineered prompt $e^{*}$ from Eq. (10) and Eq. (11) is replaced by $e_{p}^{ss}$ . The search is conducted in the linguistic space to inherently preserve semantics. We use a large language model (LLM) to find up to $n_{s}$ alternatives to the user prompt. This search early stops once an alternative prompt $e_{p}^{ss}$ 's magnitude falls below the $\lambda$ threshold. Visually in Fig. 4 (c), this is equivalent to reaching the target privacy contour at points $\mathsf{B}'$ , $\mathsf{C}'$ , or $\mathsf{D}'$ . Notably, although $\mathsf{B}$ and $\mathsf{B}'$ lie on the same privacy contour, the utility of $\mathsf{B}'$ is often higher than that of $\mathsf{B}$ . Implementationally, if all $n_{s}$ alternatives result in magnitudes above $\lambda$ , the one with the lowest magnitude will be chosen. + +Analysis: Enhancing Utility with No Privacy Cost. This semantic search differs from soft prompt engineering in that the optimization happens in the language space instead of the embedding space. The baseline strategy seeks to boost utility by minimizing the number of prompt optimization steps, which diminishes the extent of prompt modifications, significantly undermining privacy as ostensibly secure prompts may inadvertently produce memorized outputs. Conversely, our approach employs a semantic prompt + +![](images/dfd5be107f4f622f002b851ebad88e9c5521de549c1a9823d7254b225f813db4.jpg) +Figure 5. Both the baseline approach and our method effectively inhibit Stable Diffusion from producing memorized images by minimizing the magnitude in the initial inference step for prompt "The No Limits Business Woman Podcast". Nonetheless, our semantic prompt search (SS) strategy, which results in "The Empowered Business Woman's Podcast", retains considerably more utility, as demonstrated by the substantially higher CLIP similarities between the original user text prompt and both our (1) modified prompt, and (2) ultimately generated image. Qualitative results further corroborate this finding. $m_{T-1}$ is the first-step magnitude, $CLIP_{txt}$ and $CLIP_{img}$ refer to the original prompt's CLIP similarity with the modified prompt and the generated image. + +search to inherently preserve utility, leveraging the capabilities of Large Language Models (LLM) to navigate the linguistic space for suitable alternatives. By pinpointing a substitute prompt in the language space, we keep the semantic similarity and obviate the need to trade privacy for enhanced utility. + +Analysis: Enhancing Privacy with Minimum Utility Cost. This is also guaranteed by design. As discussed in Sec. 3.2, the baseline requires significant utility sacrifices to gain extra privacy. In contrast, the utility impact associated with our semantic prompt search is limited to the semantic discrepancy between the original user prompt and the LLM-derived alternative, which is much smaller. Furthermore, when the search yields an alternative prompt devoid of memorization risks, it effectively represents a significant boost in privacy for a minimal cost of utility. Geometrically as in Fig. 4 (c), this means shifting the prediction from $\bigcirc$ to $\mathbb{D}^{\prime}$ with minimum utility cost. + +Fig. 5 demonstrate SS's superiority over the baseline using a widely-investigated memorized prompt: "The No Limits Business Woman Podcast". While the baseline's prompt engineering (PE) strategy successfully reduces the magnitude to 1.15 and prevents memorization, it results in a significant semantic deviation from the original prompt, as indicated by low CLIP similarity scores for both the engineered text prompt and the generated image. In contrast, our SS strategy not only reduces the magnitude further to 0.78, effectively preventing memorization, but also significantly improves text alignment. By navigating the linguistic space, our method finds "The Empowered Business Woman's Podcast" as a privacy-preserving yet semantically aligned alternative. + +# 3.4. Synergy Effect of the Two Strategies + +Analysis: Tackling the Limitation of Semantic Search (SS). In an extreme scenario, SS might encounter a limitation where all $n_s$ searched alternatives result in magnitudes exceeding the set threshold, failing to meet the required privacy standards. This situation can be geometrically represented by a movement from point $\mathsf{O}$ to $\mathsf{B}'$ in Fig. 4 (c). However, it's important to note that even in such cases, the transition to a lower magnitude contour, where $\mathsf{B}'$ is located, still offers a certain degree of privacy enhancement without additional cost. Further advancements in privacy (e.g., moving from $\mathsf{B}'$ to $\mathsf{D}'$ ) can be achieved through our synergistic application of the PR strategy, which incurs a smaller utility cost compared to the baseline method. The right side of Fig. 8 (in Supplementary Material) showcases the synergistic benefit of utilizing the PR strategy to compensate for the SS strategy's limitations. In this example, although SS may not secure an alternative prompt with a sufficiently reduced magnitude, lowering it from the original 7.48 to 6.02 has proved to be adequate for PR to leverage the original prompt as the new anchor point for steering the generation away from memorized content effectively. + +Analysis: Addressing the Privacy-Utility Dilemma. For the baseline strategy [36], the dilemma in balancing the privacy-utility trade-off through adjusting the $\lambda$ hyperparameter manifests in two ways: opting for a higher $\lambda$ to bolster utility inadvertently raises the False Negative (FN) rate, leading to potentially unsafe prompts being misclassified as safe and thus compromising privacy; conversely, selecting a lower $\lambda$ to enhance privacy increases the False Positive (FP) rate, resulting in unnecessary modifications to already safe prompts, thereby reducing utility. Prompt re-anchoring is tailored to mitigate the privacy repercussions of False Negatives (FN) encountered when users aim to augment utility with a higher $\lambda$ . It achieves this by actively directing generation away from memorized content, as elucidated in Sec. 3.2 and depicted in Fig. 4 as a divergence from $\bigcirc$ . This approach thus optimizes the privacy-utility trade-off under such conditions. Semantic prompt search, on the other hand, aims to alleviate the utility decline tied to False Positives (FP) when enhancing privacy with a lower $\lambda$ . It conducts a linguistically sensitive search that avoids semantic degradation, consequently refining the privacy-utility balance in these scenarios. Employing both strategies synergistically ensures an enhanced privacy-utility trade-off, regardless of whether the user's priority lies with maximizing privacy or utility. + +# 4. Experiments + +# 4.1. Evaluation settings + +Although our method is designed to enhance the approach in [36], it can be easily adapted to any recent advance + +ment of the detection strategy by simply replacing $m_{T-1}$ in Eq. (12) with alternative detection signals, making PRSS highly practical and adaptable for future improvements. For example, a contemporary work [8] introduced the masked magnitude $m_t'$ (see Eq. (8)), achieving new state-of-the-art detection performance. The mitigation strategy proposed by [8] also uses prompt engineering based on the upgraded detection signal $m_t'$ , offering two key advantages: (1) $m_t'$ provides a more accurate detection trigger for initiating the mitigation strategy, (2) it serves as a more effective loss function than unmasked magnitude, yielding better gradients for optimizing prompts. These improvements make it the current state-of-the-art in mitigation. PRSS can seamlessly integrate this enhanced detection signal by replacing $m_{T-1}$ with $m_{T-1}'$ in Eq. (12), producing our best results to date. For a fair comparison, we re-implement the mitigation strategy from [8] and evaluate its performance alongside PRSS. This comparison not only highlights PRSS's capability to surpass the most recent state-of-the-art but also serves as an ablation study, assessing PRSS's impact since both methods use the same detection strategy to trigger mitigation. Additionally, we produce a version of PRSS that uses $m_t$ as the detection signal, aligning with [36], to further demonstrate PRSS's effectiveness in comparison to [36]. + +We follow the baselines [8, 36] by experimenting on Stable Diffusion v1-4, using its corresponding prompt dataset from [34], which contains 500 prompts, over 300 of which are classified as memorized and organized by different memorization types. To provide a comprehensive view, we adopt the settings from [8] and report results for both global and local memorization prompts separately. We use the default hyperparameters from [36] to re-implement the detection and prompt engineering mitigation strategies from [36] and [8]. For semantic prompt search with LLMs, we use the GPT-4 model via the OpenAI API, instructing it to generate up to $n_s = 25$ semantically similar prompts for each user prompt. The associated inference time and cost are minimal, with each alternative prompt generated in approximately 0.9 seconds, costing around $0.02. + +Following standard practice in memorization literature, we use the CLIP score to measure the alignment between the generated output and the input user prompt, serving as our utility metric. For privacy metrics, which quantify the degree of memorization, most studies rely on SSCD-based scores (see Eq. (5)), reporting metrics such as the average score, the 95th percentile score, and the percentage of memorized generations, with a threshold of 0.5 commonly indicating memorization. Additionally, [8] proposes a localized similarity metric (LS, see Eq. (6)) that measures similarity between the generated image and corresponding training image only within the localized memorization region. We present comparative results using the widely adopted SSCD metric in Fig. 6 and Fig. 7, while other metrics, including + +![](images/f701e94eadb009f9e20942b349979d6e265558bb5229dc535eff97b8966947b5.jpg) +Figure 6. Comparison with baselines. Across different detection signals ( $m$ and $m'$ ), PRSS consistently achieves superior privacy-utility trade-offs. The improvement is particularly pronounced in global memorization compared to local memorization. + +![](images/d5d280afac8367d66c49dd9e8a475d6b9dc35683989d4387f1e06bef24a6ca95.jpg) + +LS, are detailed in the Supplementary Material for a more comprehensive evaluation. + +# 4.2. Comparison with Baselines + +As shown in Fig. 6, PRSS consistently outperforms prompt engineering (PE) in achieving better privacy-utility trade-offs, regardless of the detection signal used (whether $m$ or $m'$ ). This is visually evident as the red and green lines (PRSS) lie further in the bottom-right direction compared to the yellow and blue lines (PE) for both global and local memorization scenarios. Notably, PRSS demonstrates astonishingly large improvements in global memorization (from the blue line to the green line and from the yellow line to the red line), while [8] achieves only marginal gains over [36] in this case (from the blue line to the yellow line and from the green line to the red line). For local memorization, although the margin of improvement is smaller, PRSS still achieves a significant improvement over PE. Specifically, even using the inferior detection signal $m$ , PRSS (green line) achieves comparable performance with the method proposed by [8] (yellow line), which leverages the superior signal $m'$ . Equipped with $m'$ , PRSS can achieve the state-of-the-art results (red line). This highlights PRSS's robust capability to leverage advancements in detection signals. + +# 4.3. Ablation Studies + +Fig. 7 highlights the necessity of both components in PRSS, as evidenced by the red line lying further in the bottom-right direction compared to the yellow and green lines for both global and local memorization scenarios. The synergistic effect of combining PR and SS as discussed in Sec. 3.4, is particularly evident in the local memorization scenario, which aligns with [8]'s observation that local memorization poses greater challenges for existing mitigation strategies. Using PR alone effectively enhances privacy but results in suboptimal utility, while SS alone improves utility but achieves limited privacy protection. Together, they + +![](images/963b6e8d9d17ff6018a78a1202f1a9aec28734ebc378b21b33a836242b1a16a1.jpg) +Figure 7. Ablation studies of PRSS components. Using PR alone results in low utility, while SS alone cannot ensure privacy. Combining PR and SS achieves a synergistic effect, delivering the best privacy-utility trade-off. + +leverage each other's strengths, achieving the best privacy-utility trade-off. For global memorization, either PR or SS alone significantly outperforms the baseline. However, their combination further enhances the trade-off, underscoring the robustness of PRSS in tackling both global and local memorization challenges. + +# 5. Related Work + +Memorization in Generative Models The phenomenon of generative models memorizing training data has been extensively studied across domains, including language models [3, 4, 15, 16, 18, 20], image generation [2, 10, 11, 21, 22, 33], and video generation [9, 26], due to the significant privacy and legal risks it poses [5, 14, 29, 37]. Specifically, in diffusion models, [6, 31, 34] first validated the memorization issue on various models including DDPMs [13, 23] and Stable Diffusion [28] on datasets include Oxford Flowers [24], CelebA [19], CIFAR-10 [17] and LAION [30]. Further analysis by [32] revealed that Stable Diffusion's text-conditioning exacerbates overfitting, making its memorization of LAION particularly problematic. To better understand memorization, [34] categorized the phenomenon as Matching Verbatim, Retrieval Verbatim, and Template Verbatim based on the cause, distinguishing between one-to-one and many-to-many mappings in text-image pairs, while [8] classified memorization into global and local cases, demonstrating the practical significance of this distinction. In this work, we focus on Stable Diffusion due to its widespread use, open-source nature, and real-world legal challenges [29]. We adopt the framework of [8] to analyze both global and local memorization for a more nuanced evaluation of our proposed strategies. + +Detection Strategies for Diffusion Memorization. Early detection methods required generating large image sets to estimate density [6] or comparing generations with the entire training dataset [31], which were computationally intensive. Recent approaches [27, 36] introduced more efficient one-step detection strategies. [27] detected memorization by observing less concentrated cross-attention scores + +for memorized generations than non-memorized ones. [36] proposed using the abnormally high magnitude of text-conditional noise predictions as a detection signal. Building on this, [8] incorporated localized memorization masks into the magnitude computation, offering a more accurate detection signal and achieving state-of-the-art performance. Our proposed PRSS approach leverages these detection mechanisms to trigger mitigation strategies. Notably, PRSS can adapt to any detection method by substituting the indicator variable condition in its framework. We employ the state-of-the-art method from [8] for optimal results. + +Mitigation Strategies for Diffusion Memorization. Mitigation strategies have evolved rapidly. Early methods, such as retraining models on de-duplicated data [6], were computationally expensive and showed limited effectiveness. [32] introduced randomness into text prompts or embeddings but indiscriminately affected non-memorized prompts, reducing overall utility. To improve on it, [7] proposed to compute the generation's similarity with the closest training image to engineer the image latent using the classifier guidance framework. However, this relies on computationally expensive nearest-neighbor searches across the training dataset. [36] proposed to conduct prompt engineering using the magnitude of text-conditional prediction as the loss, and [27] proposed to re-weight cross-attention scores for all tokens in the prompt. They offer improved utility-privacy trade-offs and achieve comparable performance. Most recently, [8] extended these efforts by leveraging localized magnitudes as loss functions for prompt engineering, achieving the best results to date. Our investigation reveals the limitations in the prevailing method's privacy/utility balance. In response, we introduce new mitigation strategies that achieve more favorable trade-offs. + +# 6. Conclusion and Future Work + +This paper introduces a novel PRSS method that consists of prompt re-anchoring and semantic prompt search as deliberately designed strategies to mitigate memorization in diffusion models. Empirical results show that PRSS demonstrates superior privacy-utility trade-offs over existing methods. Beyond its effectiveness, PRSS is simple and efficient to implement. It requires only minor modifications to the CFG equation to update noise predictions during inference, seamlessly integrating into the inference pipeline without additional training, fine-tuning, or reliance on training data. The efficacy of each component is validated through comprehensive analysis and ablation studies. A current limitation is that the performance of PRSS, as well as baseline mitigation strategies, relies heavily on the accuracy of the detection mechanism. When it fails to trigger accurately, they all revert to the standard Stable Diffusion process, compromising privacy preservation. Thus, improving the detection mechanism's accuracy in future works would orthogonally bolster the effectiveness of PRSS. + +# 7. Acknowledgement + +This work was supported in part by the Australian Research Council under Projects DP210101859 and FT230100549. + +# References + +[1] Midjourney. https://www.midjourney.com/. 1 +[2] Sanjeev Arora, Andrej Risteski, and Yi Zhang. Do gans learn the distribution? some theory and empirics. In International Conference on Learning Representations, 2018. 8 +[3] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In USENIX Security Symposium, 2021. 8 +[4] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022.8 +[5] Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang, Nicolas Papernot, Andreas Terzis, and Florian Tramér. The privacy onion effect: Memorization is relative. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 8 +[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In USENIX Security Symposium, pages 5253-5270, 2023. 1, 8 +[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In CVPR, 2024. 1, 3, 8 +[8] Chen Chen, Daochang Liu, Mubarak Shah, and Chang Xu. Exploring local memorization in diffusion models via bright ending attention. *ICLR*, 2025. 1, 3, 7, 8 +[9] Chen Chen, Enhuai Liu, Daochang Liu, Mubarak Shah, and Chang Xu. Investigating memorization in video diffusion models. ICLR 2nd Workshop on Navigating and Addressing Data Problems for Foundation Models (DATA-FM), 2025. 8 +[10] Yunhao Chen, Xingjun Ma, Difan Zou, and Yu-Gang Jiang. Towards a theoretical understanding of memorization in diffusion models. arXiv preprint arXiv:2410.02467, 2024. 8 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 8 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 1, 3 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 8 +[14] H. H. Jiang, L. Brown, J. Cheng, M. Khan, A. Gupta, D. Workman, A. Hanna, J. Flowers, and T. Gebru. AI Art and Its Impact on Artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374. ACM, 2023. 8 + +[15] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. Large language models struggle to learn long-tail knowledge. arXiv preprint arXiv:2211.08411, 2022. 8 +[16] Nikhil Kandpal, Eric Wallace, and Colin Raffel. Deduplicating training data mitigates privacy risks in language models. In International Conference on Machine Learning, 2022. 8 +[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 8 +[18] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the Association for Computational Linguistics, 2022. 8 +[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 8 +[20] Fatemehsadat Mireshghallah, Anish Uniyal, Tianhao Wang, David Evans, and Taylor Berg-Kirkpatrick. Memorization in nlp fine-tuning methods. arXiv preprint arXiv:2205.12506, 2022.8 +[21] Amir Naseh, Jaemin Roh, and Amir Houmansadr. Memory triggers: Unveiling memorization in text-to-image generative models through word-level duplication. arXiv preprint arXiv:2312.03692, 2023. 8 +[22] Ali Naseh, Jaechul Roh, and Amir Houmansadr. Understanding (un)intended memorization in text-to-image generative models. arXiv preprint arXiv:2312.07550, 2023. 8 +[23] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 8 +[24] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 8 +[25] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, pages 14532-14542, 2022. 3 +[26] Aimon Rahman, Malsha V. Perera, and Vishal M. Patel. Frame by familiar frame: Understanding replication in video diffusion models. arXiv preprint arXiv:2403.19593, 2024. 8 +[27] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. ECCV, 2024. 1, 3, 8 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 8 +[29] Joseph Saveri and Butterick Matthew. Stable diffusion litigation, 2023. 2023. 8 +[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine + +Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 1, 8 +[31] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In CVPR, pages 6048-6058, 2023. 1, 3, 8 +[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. In NeurIPS, 2023. 1, 3, 8 +[33] Patrick Tinsley, Adam Czajka, and Patrick Flynn. This face does not exist... but it might be yours! identity leakage in generative models. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 8 +[34] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 1, 7, 8 +[35] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Advances in Neural Information Processing Systems, pages 51008-51025. Curran Associates, Inc., 2023. 1 +[36] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In ICLR, 2024. 1, 3, 5, 6, 7, 8 +[37] Chiyuan Zhang, Daphne Ippolito, Katherine Lee, Matthew Jagielski, Florian Tramér, and Nicholas Carlini. Counterfactual memorization in neural language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. \ No newline at end of file diff --git a/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/images.zip b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5506b4a80cdec3cbd48ec46e72211ff9a3c54390 --- /dev/null +++ b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9c49aaca04586f0354d789f83023883f324af687984c1de60b0921897961fba +size 379203 diff --git a/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/layout.json b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0aa53b9e19c3151260e87e739706808085a46d74 --- /dev/null +++ b/2025/Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models/layout.json @@ -0,0 +1,9481 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 102, + 539, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 102, + 539, + 138 + ], + "spans": [ + { + "bbox": [ + 71, + 102, + 539, + 138 + ], + "type": "text", + "content": "Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "text", + "content": "Chen Chen" + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "text", + "content": " Daochang Liu" + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "text", + "content": " Mubarak Shah" + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "text", + "content": " Chang Xu" + }, + { + "bbox": [ + 150, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "spans": [ + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "type": "text", + "content": "School of Computer Science, Faculty of Engineering, The University of Sydney, Australia " + }, + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "type": "text", + "content": "School of Physics, Mathematics and Computing, The University of Western Australia, Australia " + }, + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 175, + 539, + 217 + ], + "type": "text", + "content": "Center for Research in Computer Vision, University of Central Florida, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 219, + 512, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 219, + 512, + 232 + ], + "spans": [ + { + "bbox": [ + 94, + 219, + 512, + 232 + ], + "type": "text", + "content": "{cche0711@uni., c.xu@}sydney.edu.au daochang.liu@uwa.edu.au shah@crcv.ucf.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "text", + "content": "Text-to-image diffusion models have demonstrated remarkable capabilities in creating images highly aligned with user prompts, yet their proclivity for memorizing training set images has sparked concerns about the originality of the generated images and privacy issues, potentially leading to legal complications for both model owners and users, particularly when the memorized images contain proprietary content. Although methods to mitigate these issues have been suggested, enhancing privacy often results in a significant decrease in the utility of the outputs, as indicated by text-alignment scores. To bridge the research gap, we introduce a novel method, PRSS, which refines the classifier-free guidance approach in diffusion models by integrating prompt re-anchoring (PR) to improve privacy and incorporating semantic prompt search (SS) to enhance utility. Extensive experiments across various privacy levels demonstrate that our approach consistently improves the privacy-utility trade-off, establishing a new state-of-the-art." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 504, + 135, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 135, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 135, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 297, + 714 + ], + "type": "text", + "content": "Leveraging classifier-free guidance (CFG) [12], text-to-image diffusion models like Stable Diffusion [28] and Midjourney [1] are now capable of generating highly realistic images that closely align with user-provided text prompts. This capability has propelled their popularity, leading to widespread use and distribution of their generated images. However, recent research [6, 8, 31, 34] has revealed a critical issue: these models can memorize training data, leading them to reproduce parts of images, such as foregrounds or backgrounds (local memorization, see Fig. 1), or even entire images (global memorization, see Fig. 2) during inference, instead of generating genuinely novel content. When the training data includes sensitive or copyrighted material, these memorization issues can infringe on copyright laws without notifying either the model's owners or users or the copyright holders of the replicated content. The risk is es" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 259, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 450 + ], + "type": "text", + "content": "ppecially substantial given the extensive use of these models and their reliance on massive datasets, such as LAION [30], which contain billions of web-scale images and are impractical to thoroughly review or filter manually. This risk is further highlighted by real-world cases, where several artists have filed lawsuits, arguing that models like Stable Diffusion act as “21st-century collage tools” that remix their copyrighted works, implicating Stability AI, DeviantArt, Midjourney, and Runway AI. Recognizing the potential for unauthorized reproductions, Midjourney has even banned prompts containing the term “Afghan” to prevent the generation of the copyrighted Afghan Girl photograph. Yet, as [35] demonstrates, such restrictions alone are insufficient to fully prevent the reproduction of copyrighted images. This underscores the urgent need for timely effective mitigation strategies to address these concerns." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": "In response to these legal challenges, recent initiatives [7, 8, 27, 32, 36] have focused on developing strategies to minimize memorization, achieving notable success. These approaches vary in scope, with some targeting the model's training phase and others making adjustments during inference. While training-phase strategies are theoretically viable, they lack practical significance as they predominantly experiment on a small " + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": " subset of the entire LAION5B dataset due to the impracticality of fine-tuning on the full dataset — a constraint stemming from computational limitations. In contrast, inference-phase strategies, which involve modifications only during inference, offer computational efficiency and adaptability, as they require no additional training or fine-tuning. This makes them highly applicable across pre-trained diffusion models, regardless of specific training configurations. Despite their promise, inference-time strategies can also suffer from computational constraints. For example, the approach proposed by [7] involves searching the entire LAION dataset for the nearest neighbor to optimize the generated latent representation during each denoising step. Other inference-time strategies [8, 27, 32, 36] that avoid nearest neighbor search" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8182" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 208 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 208 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 208 + ], + "type": "image", + "image_path": "2733b0e4ece0630b5c245b132c36f86a5ccb6afcbf18f36a0b53bae8f1da8e28.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 216, + 295, + 272 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 272 + ], + "type": "text", + "content": "Figure 1. Examples of Local Memorization: Stable Diffusion generations (columns 2-4) can replicate local regions of training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 275, + 294, + 411 + ], + "blocks": [ + { + "bbox": [ + 58, + 275, + 294, + 411 + ], + "lines": [ + { + "bbox": [ + 58, + 275, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 58, + 275, + 294, + 411 + ], + "type": "image", + "image_path": "6fcaa2b5f9d426a1352a61170f978de7addc41e82e072566a3906e17f2493652.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 422, + 295, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 477 + ], + "type": "text", + "content": "Figure 2. Examples of Global Memorization: Stable Diffusion generations (columns 2-4) can replicate the entire training images (column 1). Our PRSS method effectively mitigates memorization while maintaining strong alignment with the input prompts (columns 5-7)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 556 + ], + "type": "text", + "content": "demonstrate high efficiency but only moderate effectiveness, leaving a substantial gap in optimizing the privacy-utility trade-off. In particular, these methods often require significant compromises in text-image alignment to reduce memorization and conversely, struggle to sufficiently suppress memorization when trying to preserve alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "To address these limitations, we carefully analyze why existing strategies struggle to achieve an optimal privacy-utility trade-off, uncovering critical insights that the two inputs (text-conditional and unconditional predictions) used in the CFG framework are both suboptimal for mitigating memorization, even though they are standard for most other tasks in Stable Diffusion models. Specifically, we find that the text-conditional term in CFG, when optimized for privacy by prompt engineering, can significantly reduce text alignment, indicating the need for alternative methods to find privacy-preserving yet intent-aligned text-conditioning. Additionally, we observe that the unconditional term in CFG is less effective as an anchor for privacy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "text", + "content": "preservation, suggesting that alternative anchoring points that enhance privacy are necessary. Building on these insights, we introduce PRSS, a novel method that incorporates two complementary strategies to refine the widely used CFG approach for improved memorization mitigation. Our method includes (1) prompt re-anchoring, designed to preserve high privacy with a smaller utility cost, which re-anchors the unconditional component in CFG using the memorized prompt to consistently guide generations away from memorization during the inference process; (2) semantic prompt search, designed to preserve high utility with minimal privacy compromise, which optimizes the prompt away from the memorized one in the language space while keeping similar semantics. Our experimental results reveal that employing both strategies synergistically can consistently outperform baseline methods in achieving superior privacy-utility trade-offs at different privacy levels. In particular, this results in higher text alignment (utility) and substantially fewer memorized generations (improved privacy). Furthermore, ablation studies validate the efficacy of each component within our strategy. From an implementation standpoint, our strategies are simple, efficient, and effective, necessitating only adjustments to the existing CFG equation for updating the noise prediction during inference. This integration is seamless, avoiding disruption to the inference workflow and eliminating the need for additional re-training, fine-tuning, or searching over the training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 405, + 399, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 405, + 399, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 405, + 399, + 417 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 425, + 548, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 548, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 548, + 437 + ], + "type": "text", + "content": "2.1. Classifier-Free Guidance in Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "text", + "content": "Diffusion models [13, 23] are popular generative models that consist of a forward diffusion process that consistently adds Gaussian noise to real images " + }, + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "text", + "content": " steps based on the following conditional function, resulting in " + }, + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "inline_equation", + "content": "x_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 442, + 554, + 502 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 510, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 510, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 350, + 510, + 553, + 525 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "20e59606c96b1ceaa6f903b0bcc99c3c5e7e350ccd760d62244015e9a76bf626.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": " represents the noised version of " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": " after adding " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": " noise, and " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": " is the schedule that controls the magnitude of added noise at time " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": "This is followed by the reverse process, in which a de-noiser model is trained to predict the distribution (i.e., mean " + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t)" + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta}^{2}(x_{t})\\mathbf{I}" + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": ") of a cleaner version of " + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": ", at any given timestep " + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 624, + 553, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 624, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 347, + 624, + 553, + 638 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}\\right), \\sigma_ {\\theta} ^ {2} \\left(x _ {t}\\right) \\mathbf {I}\\right). \\tag {2}", + "image_path": "007cfd24a84cb5aac6606cd1c38616905ed89d7c505119f9a1a0bb8cffe662cf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": "During training, the objective function is equivalent to have the denoiser " + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,y,t)" + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": " predict the noise " + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": " to be removed, rather than the image " + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": " which may represent a user-provided text prompt:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 345, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {t \\in [ 1, T ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (x _ {t}, y, t) \\| _ {2} ^ {2} \\right]. \\tag {3}", + "image_path": "214f23b26b7fd2443b131c91753e396373d73326198e5ffb3fb91ff9c7dfa5f1.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8183" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "Classifier-free guidance (CFG) [12] is a widely used technique for balancing fidelity and diversity in the generated outputs of diffusion models, and it can be easily applied during inference. CFG operates by linearly guiding the noise prediction from the unconditional prediction, " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " towards the conditional prediction, " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,y,t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", with a scaling factor " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "s > 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " to control the degree of guidance. For simplicity, removing the timestep " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " from the notation, the adjusted noise prediction can be defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 186, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 186, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 97, + 186, + 295, + 199 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}\\right) + s \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, y\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right). \\tag {4}", + "image_path": "213064de07a2ac8188d44b6f546d44521e6e3837bed0df39d694434ec00dd432.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 239, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 239, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 239, + 215 + ], + "type": "text", + "content": "2.2. Memorization in Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": "Definition and Metrics. A widely adopted practical definition of memorization refers to generated images exhibiting high object-level similarity to training data [31]. This includes the concept of reconstructive memory in diffusion models, where foreground and background objects from training images are combined, potentially with modifications such as shifting, scaling, or cropping. These reconstructions may not match any training image pixel-for-pixel but still closely resemble specific training images at the object level. The standard metric for evaluating memorization [7, 8, 27, 31, 32, 36] is designed base on a pre-trained Self-Supervised Copy Detection (SSCD) model [25]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 362, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 362, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 63, + 362, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\operatorname {S S C D} (\\hat {x}) = \\max _ {x _ {i} \\in \\mathcal {D}} \\operatorname {C o s i n e} \\left(\\mathbf {E} _ {S S C D} (\\hat {x}), \\mathbf {E} _ {S S C D} (x _ {i})\\right), \\tag {5}", + "image_path": "468c0fa2ae762603340ac8d499c7b39e5823b16ef9f28842d7e3c03e6b236058.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "content": "where for each generated image " + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "content": ", cosine similarity is computed between its SSCD embedding " + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{SSCD}(\\hat{x})" + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "content": " and the embeddings " + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{SSCD}(x_i)" + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "content": " of all training images " + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "content": " in the dataset " + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "content": ". The maximum similarity serves as the final SSCD score." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 428, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 296, + 523 + ], + "type": "text", + "content": "Most recently, [8] proposed a method to localize memorized regions in generated images, arguing that only memorized regions pose privacy risks, making local memorization a more meaningful concept. They introduced an efficient approach to synthesize a mask that highlights these regions without requiring access to training data. Using this mask, they proposed a new localized metric based on both object-level SSCD and pixel-level Euclidean distance:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 529, + 295, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 529, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 87, + 529, + 295, + 544 + ], + "type": "interline_equation", + "content": "L S (\\hat {x}, x) = - \\mathbb {1} _ {\\mathrm {S S C D} > 0. 5} \\cdot \\| (\\hat {x} - x) \\circ \\mathbf {m} \\| _ {2}, \\tag {6}", + "image_path": "93d4750be4600f3e3d342ac8cd0ec680f8e1f29f5a94e2fbb47e7f719f236732.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 548, + 295, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 548, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 548, + 295, + 595 + ], + "type": "text", + "content": " is the local memorization mask and " + }, + { + "bbox": [ + 55, + 548, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\mathrm{SSCD} > 0.5}" + }, + { + "bbox": [ + 55, + 548, + 295, + 595 + ], + "type": "text", + "content": " is an indicator function that signifies whether the SSCD score exceeds the 0.5 threshold, which is commonly used to distinguish memorized from non-memorized content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 596, + 296, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 656 + ], + "type": "text", + "content": "Memorization Detection. A recent work [36] introduces an innovative approach for detecting potential memorized generations during inference by leveraging the magnitude of the difference between text-conditional " + }, + { + "bbox": [ + 55, + 596, + 296, + 656 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, e_p)" + }, + { + "bbox": [ + 55, + 596, + 296, + 656 + ], + "type": "text", + "content": " and unconditional " + }, + { + "bbox": [ + 55, + 596, + 296, + 656 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, e_\\phi)" + }, + { + "bbox": [ + 55, + 596, + 296, + 656 + ], + "type": "text", + "content": " noise predictions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 662, + 295, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 662, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 107, + 662, + 295, + 675 + ], + "type": "interline_equation", + "content": "m _ {t} = \\left\\| \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) \\right\\| _ {2}, \\tag {7}", + "image_path": "9f3b3ec6d12cde555df6f1744519936efa71c514c25f79ba88caa346347a7fa1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "content": "where memorized generations consistently exhibit larger magnitudes than non-memorized ones, making this magnitude a strong signal for detection. Furthermore, accurate" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 554, + 107 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 554, + 107 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 554, + 107 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 554, + 107 + ], + "type": "image", + "image_path": "a3778257d9c64f428098939618dc4033c07215cc89161ab6972db3949f644b30.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 117, + 555, + 149 + ], + "lines": [ + { + "bbox": [ + 313, + 117, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 555, + 149 + ], + "type": "text", + "content": "Figure 3. Illustration of the logic behind how " + }, + { + "bbox": [ + 313, + 117, + 555, + 149 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 117, + 555, + 149 + ], + "type": "text", + "content": " in Eq. (10) governs the privacy-utility trade-off. Currently, the privacy-utility trade-off is governed solely by how extensively the prompt is modified." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 150, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 553, + 186 + ], + "type": "text", + "content": "detection can be achieved efficiently by evaluating the magnitude " + }, + { + "bbox": [ + 313, + 150, + 553, + 186 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 313, + 150, + 553, + 186 + ], + "type": "text", + "content": " at the first denoising step (timestep " + }, + { + "bbox": [ + 313, + 150, + 553, + 186 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 313, + 150, + 553, + 186 + ], + "type": "text", + "content": ", highlighting the method's computational efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 186, + 553, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 234 + ], + "type": "text", + "content": "The contemporary work [8] extends this detection method by incorporating their proposed local memorization mask " + }, + { + "bbox": [ + 313, + 186, + 553, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 186, + 553, + 234 + ], + "type": "text", + "content": " into Eq. (7) to compute a masked magnitude, achieving state-of-the-art performance:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 241, + 553, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 553, + 283 + ], + "type": "interline_equation", + "content": "m _ {t} ^ {\\prime} = \\left\\| \\left(\\varepsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\varepsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\circ \\mathbf {m} \\right\\| _ {2} / \\left(\\frac {1}{N} \\sum_ {i = 1} ^ {N} m _ {i}\\right), \\tag {8}", + "image_path": "4d80e25df674d69b35b7a5cc1d7f88703803e846418a98bf7add467d3e41cf81.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "text", + "content": " is the local memorization mask and " + }, + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "inline_equation", + "content": "\\frac{1}{N} \\sum_{i=1}^{N} m_i" + }, + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "text", + "content": " is used to normalize the result by the mean of " + }, + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 283, + 553, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 317, + 370, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 317, + 370, + 330 + ], + "spans": [ + { + "bbox": [ + 314, + 317, + 370, + 330 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 338, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 447 + ], + "type": "text", + "content": "We begin with a detailed analysis in Sec. 3.1 to uncover the key challenges that hinder existing strategies from achieving an optimal privacy-utility trade-off. Building on these insights, we evolve the baseline step by step with our proposed prompt re-anchoring (PR) strategy in Sec 3.2, which focuses on improving privacy, and semantic prompt search (SS) in Sec 3.3, which enhances utility. Finally, in Sec. 3.4, we analyze how the combination of PR and SS achieves a synergistic effect, offering the best privacy-utility trade-off." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 453, + 526, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 526, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 526, + 466 + ], + "type": "text", + "content": "3.1. Trade-off Analysis of Existing Strategies" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": "Our method is built upon the widely-used classifier-free guidance (CFG) by treating the input text prompt from user " + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": " as the condition " + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": " in Eq. (4) and rewriting the unconditional noise as a null condition " + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "inline_equation", + "content": "e_\\phi" + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 343, + 526, + 553, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 526, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 343, + 526, + 553, + 540 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right). \\tag {9}", + "image_path": "423b424bdb918ae8434e4afd147497436ffb586ac23c2175d902dfc821c05131.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "content": "The one-step prompt engineering (PE) based memorization mitigation method proposed by [36] then counteract memorization by reformulating Eq. (9) as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 590, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 590, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 318, + 590, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\{m _ {T - 1} < \\lambda \\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e ^ {*}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} > \\lambda \\right\\}}, \\tag {10} \\\\ \\end{array}", + "image_path": "ce085f9abc267f0b7c06741361a561eede2c90d95e94aa44418ed34dd0624e9a.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "m_{T - 1} = \\| \\epsilon_{\\theta}(x_{T - 1},e_p) - \\epsilon_{\\theta}(x_{T - 1},e_\\phi)\\| _2" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " represents the magnitude at the initial step " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " as per Eq. (7), " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_p)" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_{t},e^{*})" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " are the noise predictions conditioned on the user prompt embedding " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " and engineered prompt embedding " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "e^*" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " (optimized based on the gradient computed from the first-step magnitude " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "m_{T - 1}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "), and" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8184" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 89, + 207, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 182, + 81 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 182, + 81 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 182, + 81 + ], + "type": "text", + "content": "a) Privacy-Utility Trade-Off in Baseline" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 89, + 207, + 190 + ], + "lines": [ + { + "bbox": [ + 58, + 89, + 207, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 89, + 207, + 190 + ], + "type": "image", + "image_path": "0212fde5e1de8ac4ebb3f459c6cbafb2045d8f900b37c2573da45a72a39af262.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 91, + 373, + 191 + ], + "blocks": [ + { + "bbox": [ + 213, + 71, + 389, + 81 + ], + "lines": [ + { + "bbox": [ + 213, + 71, + 389, + 81 + ], + "spans": [ + { + "bbox": [ + 213, + 71, + 389, + 81 + ], + "type": "text", + "content": "b) Enhancing Privacy with Prompt Re-Anchoring (PR)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 91, + 373, + 191 + ], + "lines": [ + { + "bbox": [ + 223, + 91, + 373, + 191 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 373, + 191 + ], + "type": "image", + "image_path": "d36f7bb43ddcb6b429d029441996088fcf5d52e0f72545ec8f86abda62de1e9d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 199, + 362, + 208 + ], + "lines": [ + { + "bbox": [ + 275, + 199, + 362, + 208 + ], + "spans": [ + { + "bbox": [ + 275, + 199, + 362, + 208 + ], + "type": "inline_equation", + "content": "\\mathbf{SIM}(\\mathbb{O},\\mathbb{B}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{C}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{D})" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 386, + 91, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 400, + 72, + 553, + 81 + ], + "lines": [ + { + "bbox": [ + 400, + 72, + 553, + 81 + ], + "spans": [ + { + "bbox": [ + 400, + 72, + 553, + 81 + ], + "type": "text", + "content": "c) Enhancing Utility with Semantic Search (SS)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 91, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 386, + 91, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 553, + 201 + ], + "type": "image", + "image_path": "69ef1818c0a04ced470b7bb6b8cd221b10dfd10a64c16d03cb3dad59a6c838c1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 202, + 552, + 208 + ], + "lines": [ + { + "bbox": [ + 493, + 202, + 552, + 208 + ], + "spans": [ + { + "bbox": [ + 493, + 202, + 552, + 208 + ], + "type": "inline_equation", + "content": "\\mathrm{SIM}(\\mathbb{O},\\mathbb{C}) < \\mathrm{SIM}(\\mathbb{O},\\mathbb{C}^{\\prime})" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 208, + 552, + 215 + ], + "lines": [ + { + "bbox": [ + 493, + 208, + 552, + 215 + ], + "spans": [ + { + "bbox": [ + 493, + 208, + 552, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{SIM}(\\mathbb{O},\\mathbb{D}) < \\mathbf{SIM}(\\mathbb{O},\\mathbb{D}^{\\prime})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 477, + 222, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 477, + 222, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 477, + 222, + 553, + 230 + ], + "type": "text", + "content": "SIM CLIP Similarity (Utility)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 66, + 220, + 78, + 232 + ], + "blocks": [ + { + "bbox": [ + 66, + 220, + 78, + 232 + ], + "lines": [ + { + "bbox": [ + 66, + 220, + 78, + 232 + ], + "spans": [ + { + "bbox": [ + 66, + 220, + 78, + 232 + ], + "type": "image", + "image_path": "96271cda170299881d18afaffb0f27f525f9fa0b02cffc9cfd0d9563eb17ebc9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 223, + 165, + 232 + ], + "lines": [ + { + "bbox": [ + 82, + 223, + 165, + 232 + ], + "spans": [ + { + "bbox": [ + 82, + 223, + 165, + 232 + ], + "type": "text", + "content": "Contour Line of Equal Privacy" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 184, + 223, + 194, + 231 + ], + "blocks": [ + { + "bbox": [ + 111, + 198, + 198, + 208 + ], + "lines": [ + { + "bbox": [ + 111, + 198, + 198, + 208 + ], + "spans": [ + { + "bbox": [ + 111, + 198, + 198, + 208 + ], + "type": "inline_equation", + "content": "\\mathbf{SIM}(\\mathbb{O},\\mathbb{B}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{C}) > \\mathbf{SIM}(\\mathbb{O},\\mathbb{D})" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 223, + 194, + 231 + ], + "lines": [ + { + "bbox": [ + 184, + 223, + 194, + 231 + ], + "spans": [ + { + "bbox": [ + 184, + 223, + 194, + 231 + ], + "type": "image", + "image_path": "2dc232078a2a1003ca08f48f3726f87b2632c37018b5000d69d8ab2879e998c7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 223, + 257, + 231 + ], + "lines": [ + { + "bbox": [ + 196, + 223, + 257, + 231 + ], + "spans": [ + { + "bbox": [ + 196, + 223, + 257, + 231 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 196, + 223, + 257, + 231 + ], + "type": "text", + "content": " Baseline Guidance" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 282, + 223, + 287, + 231 + ], + "blocks": [ + { + "bbox": [ + 282, + 223, + 287, + 231 + ], + "lines": [ + { + "bbox": [ + 282, + 223, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 282, + 223, + 287, + 231 + ], + "type": "image", + "image_path": "ab9957472abcddd87421923a6345bf50c88af6af601cdbb38883cef9d3ba5c91.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 223, + 352, + 230 + ], + "lines": [ + { + "bbox": [ + 290, + 223, + 352, + 230 + ], + "spans": [ + { + "bbox": [ + 290, + 223, + 352, + 230 + ], + "type": "text", + "content": "Our PR Guidance" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 381, + 223, + 386, + 231 + ], + "blocks": [ + { + "bbox": [ + 381, + 223, + 386, + 231 + ], + "lines": [ + { + "bbox": [ + 381, + 223, + 386, + 231 + ], + "spans": [ + { + "bbox": [ + 381, + 223, + 386, + 231 + ], + "type": "image", + "image_path": "177949909573a9f6dd18342974a45c2068fb5e2b034d16c2b039580587515fd5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 223, + 458, + 230 + ], + "lines": [ + { + "bbox": [ + 389, + 223, + 458, + 230 + ], + "spans": [ + { + "bbox": [ + 389, + 223, + 458, + 230 + ], + "type": "text", + "content": "Our PRSS Guidance" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 67, + 244, + 75, + 251 + ], + "blocks": [ + { + "bbox": [ + 67, + 244, + 75, + 251 + ], + "lines": [ + { + "bbox": [ + 67, + 244, + 75, + 251 + ], + "spans": [ + { + "bbox": [ + 67, + 244, + 75, + 251 + ], + "type": "image", + "image_path": "d743170ccf694ad07aa47bf9ba3e381adfb09ed86a99bf87f302f9d3f2b8cddc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 244, + 149, + 252 + ], + "lines": [ + { + "bbox": [ + 80, + 244, + 149, + 252 + ], + "spans": [ + { + "bbox": [ + 80, + 244, + 149, + 252 + ], + "type": "text", + "content": "Unconditional Prediction" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 179, + 245, + 186, + 252 + ], + "blocks": [ + { + "bbox": [ + 179, + 245, + 186, + 252 + ], + "lines": [ + { + "bbox": [ + 179, + 245, + 186, + 252 + ], + "spans": [ + { + "bbox": [ + 179, + 245, + 186, + 252 + ], + "type": "image", + "image_path": "eb41483b64ef842a229262c827e6eea42cad7c15d5a22a22637b115eb02aae4e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 246, + 260, + 254 + ], + "lines": [ + { + "bbox": [ + 190, + 246, + 260, + 254 + ], + "spans": [ + { + "bbox": [ + 190, + 246, + 260, + 254 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 190, + 246, + 260, + 254 + ], + "type": "text", + "content": " -conditional prediction" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 293, + 242, + 308, + 251 + ], + "blocks": [ + { + "bbox": [ + 293, + 242, + 308, + 251 + ], + "lines": [ + { + "bbox": [ + 293, + 242, + 308, + 251 + ], + "spans": [ + { + "bbox": [ + 293, + 242, + 308, + 251 + ], + "type": "image", + "image_path": "09116e76c0a1a6aaf5a0c28da69766ee7c6a01ac6602a11296178d6a38e2e370.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 241, + 388, + 256 + ], + "lines": [ + { + "bbox": [ + 315, + 241, + 388, + 256 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 388, + 256 + ], + "type": "inline_equation", + "content": "e^*" + }, + { + "bbox": [ + 315, + 241, + 388, + 256 + ], + "type": "text", + "content": " -conditional prediction with different privacy level" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 423, + 242, + 451, + 254 + ], + "blocks": [ + { + "bbox": [ + 423, + 242, + 451, + 254 + ], + "lines": [ + { + "bbox": [ + 423, + 242, + 451, + 254 + ], + "spans": [ + { + "bbox": [ + 423, + 242, + 451, + 254 + ], + "type": "image", + "image_path": "e1ddcd0dc93036430c7ca620bcf16651170f2806ab61fc35b92f4abd6a63e939.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 240, + 529, + 256 + ], + "lines": [ + { + "bbox": [ + 457, + 240, + 529, + 256 + ], + "spans": [ + { + "bbox": [ + 457, + 240, + 529, + 256 + ], + "type": "inline_equation", + "content": "e_p^{ss}" + }, + { + "bbox": [ + 457, + 240, + 529, + 256 + ], + "type": "text", + "content": " -conditional prediction with different privacy level" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 261, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 555, + 316 + ], + "type": "text", + "content": "Figure 4. Illustration of our proposed strategy (PRSS) compared to the baseline prompt engineering (PE) method. a) The privacy-utility trade-off in the baseline method, where the privacy increases at the cost of utility from B to C to D. b) Our proposed PR guidance, i.e., the green arrow, drives the image out of the low-privacy region more efficiently than the baseline guidance, i.e., the black arrows. c) Our proposed SS can enhance utility, by finding points " + }, + { + "bbox": [ + 55, + 261, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\mathrm{B}^{\\prime}\\mathrm{C}^{\\prime}\\mathrm{D}^{\\prime}" + }, + { + "bbox": [ + 55, + 261, + 555, + 316 + ], + "type": "text", + "content": " that are at the same level of privacy as BCD but with better utility. In summary, using the proposed PR and SS jointly can enhance the privacy-utility trade-off to mitigate memorization in diffusion models." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, e_{\\phi})" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " is the unconditional noise prediction. This strategy detects potential memorization for a given input prompt based on the magnitude signal " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " with a threshold " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": ", and undertakes soft prompt engineering (i.e., optimizes the prompt embeddings) if potential memorization is detected. In this context, if " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " is under " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": ", the prompt is considered safe from memorization, and the generation proceeds with the original user prompt " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": ", mirroring Stable Diffusion's output. Conversely, if " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " is above " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": ", the prompt is then identified as likely to produce memorized images, thus " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " is optimized to and replaced with an engineered prompt " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "e^*" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " in the conditional prediction while keeping the unconditional prediction " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, e_{\\phi})" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " as is following the classical CFG equation. Specifically, " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "e^*" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " is optimized to reduce the first-step magnitude " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " using gradient descent and the optimization halts once the magnitude " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": " falls below " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": ", marking the engineered prompt as less prone to memorization. Both detection and prompt optimization occur only during the initial inference step " + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "inline_equation", + "content": "T-1" + }, + { + "bbox": [ + 56, + 323, + 296, + 563 + ], + "type": "text", + "content": ", rather than throughout the entire inference process, enhancing efficiency." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "text", + "content": "Privacy-utility trade-off. Eq. (10) suffers from a severe privacy-utility trade-off as in Fig. 3. Opting for a lower " + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "text", + "content": " improves the privacy (risk of memorization) of outputs at the expense of utility (text-alignment). It requires more substantial optimization of the input prompt, aiming to achieve a magnitude " + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "inline_equation", + "content": "m_{T - 1}" + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "text", + "content": " lower than " + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "text", + "content": ", thereby enhancing privacy. However, such extensive modifications to the prompt may lead to a loss of its original meaning, compromising utility significantly. Conversely, a higher " + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 566, + 296, + 686 + ], + "type": "text", + "content": " upholds the utility of generated outputs but at the cost of privacy." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Geometric analysis. Fig. 4 (a) visualizes the privacy-utility trade-off in Eq. (10) as a function of " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", where vary" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "text", + "content": "ing levels of " + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "text", + "content": " define distinct contours corresponding to target magnitudes and privacy levels during prompt engineering (PE). Outer contours represent greater prompt modifications, which enhance privacy. This optimization shifts the prompt embedding from the original memorized prompt, " + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "inline_equation", + "content": "e_{p}" + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "text", + "content": ", located at point " + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "inline_equation", + "content": "\\mathsf{O}" + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "text", + "content": ", to the engineered prompt, " + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "inline_equation", + "content": "e^{*}" + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "text", + "content": ", at points B, C, or D. The number of PE iterations determines the distance of the prompt from point " + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "inline_equation", + "content": "\\mathsf{O}" + }, + { + "bbox": [ + 313, + 323, + 555, + 442 + ], + "type": "text", + "content": ": more iterations lead to greater divergence, achieving higher privacy levels. The process halts once the desired privacy level is attained." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 443, + 556, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 610 + ], + "type": "text", + "content": "Overview of our improvements. Our trade-off analysis reveals that existing strategies rely solely on prompt modification methods such as prompt engineering to enhance privacy, which is suboptimal and significantly compromises utility. This highlights an urgent need to propose alternative methods to improve privacy and preserve utility. To address this, we introduce two meticulously crafted mitigation strategies that only require refining the guidance components in CFG to achieve a superior privacy-utility tradeoff: (1) prompt re-anchoring (PR) and (2) semantic prompt search (SS). Specifically, SS enhances utility with minimal privacy cost, while PR serves as an effective alternative to improve privacy, which has the potential of costing less utility when used in combination with SS." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "type": "text", + "content": "3.2. Enhancing Privacy Through Prompt Re-Anchoring" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "One of our objectives is to safeguard privacy while preserving as much utility as possible. Recall that in Eq. (10) enhancing privacy requires reducing the magnitude, which correlates with a significant decrease in utility. It is important to note that the utility loss stems not from the reduction in magnitude itself, but from the baseline's inefficient" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8185" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "approach to achieving that reduction. Visually, in Fig. 4, predictions that lie on the same magnitude contour may exhibit notably different levels of utility. This observation motivates us to ponder: Is it possible to reduce magnitude through a more efficient approach (incurring a lower utility cost) compared to the baseline method of prompt engineering? This insight has spurred the creation of our prompt re-anchoring (PR) strategy, which aims to reduce magnitude by steering predictions away from memorized images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "content": "We argue that " + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "content": " is discarded too hastily in Eq. (10) when the memorization is detected (" + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "inline_equation", + "content": "m_{T-1} > \\lambda" + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "content": "). Instead, recognizing " + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "content": " as a valuable anchor point — its conditional generation can trace the memorization path — we use " + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "content": " to re-anchor the generation process, by replacing the unconditional prediction in CFG with the " + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "content": "-conditioned prediction to divert the generation away from memorization. Thereby, Eq. (10) is re-written to the following form," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 283, + 294, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 283, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 283, + 294, + 324 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} < \\lambda \\right\\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e ^ {*}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right)\\right) \\right] \\mathbb {1} _ {\\{m _ {T - 1} > \\lambda \\}}. \\tag {11} \\\\ \\end{array}", + "image_path": "704988a3925378a60cd56ea71dabe35b24840dd02e776d7bc060e0417fae8b2a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "content": "Analysis: Enhancing Privacy with Less Utility Cost. The baseline approach in Eq. (10) to enhancing privacy necessitates significant utility sacrifices, primarily due to the need for more extensive gradient updates during prompt optimization to obtain " + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "inline_equation", + "content": "e^{*}" + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "content": ". This leads to substantial modifications of the original user prompt and a large gap between " + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "inline_equation", + "content": "e^{*}" + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "inline_equation", + "content": "e_{p}" + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "content": ", causing the prompt to diverge significantly from the user's initial intent and thus undermine the text alignment. In contrast, our prompt re-anchoring strategy employs the intuition from the original CFG approach to guide generations toward a preferred output while distancing them from an undesirable one, effectively achieved through linear interpolation between two generations. Uniquely, our strategy designates the undesired generation as the memorized one that enforces conditioning on the original user prompt " + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "inline_equation", + "content": "e_{p}" + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "content": ", rather than relying on unconditional conditioning " + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "inline_equation", + "content": "e_{\\phi}" + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "content": ". This approach steers generations away from memorized content more efficiently with the contrastive direction, which therefore requires fewer gradient updates to obtain " + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "inline_equation", + "content": "e^{*}" + }, + { + "bbox": [ + 55, + 331, + 295, + 582 + ], + "type": "text", + "content": " and thus less alternation on the prompt for the same privacy enhancement level. Our previous question is hence answered." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": "Fig. 4 represents a geometric interpretation of our prompt re-anchoring strategy. Classical CFG directs " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": " away from unconditional (point A) towards conditional predictions (O for " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": ", and B, C, D for " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "e^*" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": ") as in the left figure. PR, however, guides " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": " from the memorized prediction (O) towards less or unmemorized predictions with lower magnitude (B, C, D), enhancing privacy, as in the middle figure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": "Analysis: Continuous Memorization Diversion without Loss of Efficiency. Although [36] found that minimizing the initial magnitude " + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": " can indirectly reduce magnitudes in later steps, we observe that the magnitude" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "can still spike and memorization may still occur later in the inference process. For example, as illustrated in Fig. 8 (in Supplementary Material), while the engineered prompt achieves a low magnitude initially, it fails to ensure that this low level is maintained throughout subsequent inference steps, potentially leading back to high magnitudes indicative of Stable Diffusion's memorized outputs. To counteract this, the PR strategy can continuously divert generations away from initially identified memorized conditions (point " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": ") across the entirety of the inference process as shown in Fig. 8 (in Supplementary Material). The effectiveness is also evident from the qualitative appearance of the final generated image, and its quantitative similarity score as in Fig. 8 (in Supplementary Material)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 246, + 553, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 270 + ], + "type": "text", + "content": "3.3. Enhancing Utility Through Semantic Prompt Search" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "text", + "content": "In this section, we aim to propose a strategy to enhance utility with minimal privacy cost. Recall that low utility is primarily due to the increased difference between the user prompt and the engineered prompt. We suggest conducting semantic search to find a prompt " + }, + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "inline_equation", + "content": "e_p^{ss}" + }, + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "text", + "content": " semantically similar to the original user prompt " + }, + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "text", + "content": " but with a lower risk of memorization, aiming to replace the engineered prompt " + }, + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "inline_equation", + "content": "e^*" + }, + { + "bbox": [ + 313, + 276, + 555, + 407 + ], + "type": "text", + "content": " that results in the unsatisfactory trade-off. Eq. (11) is thus further re-formulated to the following form after incorporating semantic prompt search strategy in addition to the prompt re-anchoring strategy," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 415, + 553, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 454 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\epsilon} \\leftarrow \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\phi}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} < \\lambda \\right\\}} \\\\ + \\left[ \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right) + s \\left(\\epsilon_ {\\theta} \\left(x _ {t}, e _ {p} ^ {s s}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, e _ {p}\\right)\\right) \\right] \\mathbb {1} _ {\\left\\{m _ {T - 1} > \\lambda \\right\\}}. \\tag {12} \\\\ \\end{array}", + "image_path": "5c5b37740487628812a3f538388fb4e89770006b5c08d90ab79c57d19e7bea82.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": "In Eq. (12), the engineered prompt " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "e^{*}" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": " from Eq. (10) and Eq. (11) is replaced by " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "e_{p}^{ss}" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": ". The search is conducted in the linguistic space to inherently preserve semantics. We use a large language model (LLM) to find up to " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "n_{s}" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": " alternatives to the user prompt. This search early stops once an alternative prompt " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "e_{p}^{ss}" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": "'s magnitude falls below the " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": " threshold. Visually in Fig. 4 (c), this is equivalent to reaching the target privacy contour at points " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{B}'" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{C}'" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{D}'" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": ". Notably, although " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{B}" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{B}'" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": " lie on the same privacy contour, the utility of " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{B}'" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": " is often higher than that of " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{B}" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": ". Implementationally, if all " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "n_{s}" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": " alternatives result in magnitudes above " + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 462, + 553, + 605 + ], + "type": "text", + "content": ", the one with the lowest magnitude will be chosen." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Analysis: Enhancing Utility with No Privacy Cost. This semantic search differs from soft prompt engineering in that the optimization happens in the language space instead of the embedding space. The baseline strategy seeks to boost utility by minimizing the number of prompt optimization steps, which diminishes the extent of prompt modifications, significantly undermining privacy as ostensibly secure prompts may inadvertently produce memorized outputs. Conversely, our approach employs a semantic prompt" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8186" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 287, + 185 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 287, + 185 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 287, + 185 + ], + "type": "image", + "image_path": "dfd5be107f4f622f002b851ebad88e9c5521de549c1a9823d7254b225f813db4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "type": "text", + "content": "Figure 5. Both the baseline approach and our method effectively inhibit Stable Diffusion from producing memorized images by minimizing the magnitude in the initial inference step for prompt \"The No Limits Business Woman Podcast\". Nonetheless, our semantic prompt search (SS) strategy, which results in \"The Empowered Business Woman's Podcast\", retains considerably more utility, as demonstrated by the substantially higher CLIP similarities between the original user text prompt and both our (1) modified prompt, and (2) ultimately generated image. Qualitative results further corroborate this finding. " + }, + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "type": "text", + "content": " is the first-step magnitude, " + }, + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "type": "inline_equation", + "content": "CLIP_{txt}" + }, + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "type": "inline_equation", + "content": "CLIP_{img}" + }, + { + "bbox": [ + 55, + 194, + 296, + 327 + ], + "type": "text", + "content": " refer to the original prompt's CLIP similarity with the modified prompt and the generated image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 402 + ], + "type": "text", + "content": "search to inherently preserve utility, leveraging the capabilities of Large Language Models (LLM) to navigate the linguistic space for suitable alternatives. By pinpointing a substitute prompt in the language space, we keep the semantic similarity and obviate the need to trade privacy for enhanced utility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "content": "Analysis: Enhancing Privacy with Minimum Utility Cost. This is also guaranteed by design. As discussed in Sec. 3.2, the baseline requires significant utility sacrifices to gain extra privacy. In contrast, the utility impact associated with our semantic prompt search is limited to the semantic discrepancy between the original user prompt and the LLM-derived alternative, which is much smaller. Furthermore, when the search yields an alternative prompt devoid of memorization risks, it effectively represents a significant boost in privacy for a minimal cost of utility. Geometrically as in Fig. 4 (c), this means shifting the prediction from " + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^{\\prime}" + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "content": " with minimum utility cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Fig. 5 demonstrate SS's superiority over the baseline using a widely-investigated memorized prompt: \"The No Limits Business Woman Podcast\". While the baseline's prompt engineering (PE) strategy successfully reduces the magnitude to 1.15 and prevents memorization, it results in a significant semantic deviation from the original prompt, as indicated by low CLIP similarity scores for both the engineered text prompt and the generated image. In contrast, our SS strategy not only reduces the magnitude further to 0.78, effectively preventing memorization, but also significantly improves text alignment. By navigating the linguistic space, our method finds \"The Empowered Business Woman's Podcast\" as a privacy-preserving yet semantically aligned alternative." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 508, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 508, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 508, + 85 + ], + "type": "text", + "content": "3.4. Synergy Effect of the Two Strategies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": "Analysis: Tackling the Limitation of Semantic Search (SS). In an extreme scenario, SS might encounter a limitation where all " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " searched alternatives result in magnitudes exceeding the set threshold, failing to meet the required privacy standards. This situation can be geometrically represented by a movement from point " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathsf{O}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathsf{B}'" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " in Fig. 4 (c). However, it's important to note that even in such cases, the transition to a lower magnitude contour, where " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathsf{B}'" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " is located, still offers a certain degree of privacy enhancement without additional cost. Further advancements in privacy (e.g., moving from " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathsf{B}'" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathsf{D}'" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ") can be achieved through our synergistic application of the PR strategy, which incurs a smaller utility cost compared to the baseline method. The right side of Fig. 8 (in Supplementary Material) showcases the synergistic benefit of utilizing the PR strategy to compensate for the SS strategy's limitations. In this example, although SS may not secure an alternative prompt with a sufficiently reduced magnitude, lowering it from the original 7.48 to 6.02 has proved to be adequate for PR to leverage the original prompt as the new anchor point for steering the generation away from memorized content effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "content": "Analysis: Addressing the Privacy-Utility Dilemma. For the baseline strategy [36], the dilemma in balancing the privacy-utility trade-off through adjusting the " + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "content": " hyperparameter manifests in two ways: opting for a higher " + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "content": " to bolster utility inadvertently raises the False Negative (FN) rate, leading to potentially unsafe prompts being misclassified as safe and thus compromising privacy; conversely, selecting a lower " + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "content": " to enhance privacy increases the False Positive (FP) rate, resulting in unnecessary modifications to already safe prompts, thereby reducing utility. Prompt re-anchoring is tailored to mitigate the privacy repercussions of False Negatives (FN) encountered when users aim to augment utility with a higher " + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "content": ". It achieves this by actively directing generation away from memorized content, as elucidated in Sec. 3.2 and depicted in Fig. 4 as a divergence from " + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "content": ". This approach thus optimizes the privacy-utility trade-off under such conditions. Semantic prompt search, on the other hand, aims to alleviate the utility decline tied to False Positives (FP) when enhancing privacy with a lower " + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 342, + 556, + 641 + ], + "type": "text", + "content": ". It conducts a linguistically sensitive search that avoids semantic degradation, consequently refining the privacy-utility balance in these scenarios. Employing both strategies synergistically ensures an enhanced privacy-utility trade-off, regardless of whether the user's priority lies with maximizing privacy or utility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 651, + 395, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 395, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 395, + 665 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 671, + 426, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 426, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 426, + 685 + ], + "type": "text", + "content": "4.1. Evaluation settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Although our method is designed to enhance the approach in [36], it can be easily adapted to any recent advance" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8187" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": "ment of the detection strategy by simply replacing " + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": " in Eq. (12) with alternative detection signals, making PRSS highly practical and adaptable for future improvements. For example, a contemporary work [8] introduced the masked magnitude " + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "inline_equation", + "content": "m_t'" + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": " (see Eq. (8)), achieving new state-of-the-art detection performance. The mitigation strategy proposed by [8] also uses prompt engineering based on the upgraded detection signal " + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "inline_equation", + "content": "m_t'" + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": ", offering two key advantages: (1) " + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "inline_equation", + "content": "m_t'" + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": " provides a more accurate detection trigger for initiating the mitigation strategy, (2) it serves as a more effective loss function than unmasked magnitude, yielding better gradients for optimizing prompts. These improvements make it the current state-of-the-art in mitigation. PRSS can seamlessly integrate this enhanced detection signal by replacing " + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "inline_equation", + "content": "m_{T-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "inline_equation", + "content": "m_{T-1}'" + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": " in Eq. (12), producing our best results to date. For a fair comparison, we re-implement the mitigation strategy from [8] and evaluate its performance alongside PRSS. This comparison not only highlights PRSS's capability to surpass the most recent state-of-the-art but also serves as an ablation study, assessing PRSS's impact since both methods use the same detection strategy to trigger mitigation. Additionally, we produce a version of PRSS that uses " + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": " as the detection signal, aligning with [36], to further demonstrate PRSS's effectiveness in comparison to [36]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 362, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 543 + ], + "type": "text", + "content": "We follow the baselines [8, 36] by experimenting on Stable Diffusion v1-4, using its corresponding prompt dataset from [34], which contains 500 prompts, over 300 of which are classified as memorized and organized by different memorization types. To provide a comprehensive view, we adopt the settings from [8] and report results for both global and local memorization prompts separately. We use the default hyperparameters from [36] to re-implement the detection and prompt engineering mitigation strategies from [36] and [8]. For semantic prompt search with LLMs, we use the GPT-4 model via the OpenAI API, instructing it to generate up to " + }, + { + "bbox": [ + 55, + 362, + 296, + 543 + ], + "type": "inline_equation", + "content": "n_s = 25" + }, + { + "bbox": [ + 55, + 362, + 296, + 543 + ], + "type": "text", + "content": " semantically similar prompts for each user prompt. The associated inference time and cost are minimal, with each alternative prompt generated in approximately 0.9 seconds, costing around $0.02." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "Following standard practice in memorization literature, we use the CLIP score to measure the alignment between the generated output and the input user prompt, serving as our utility metric. For privacy metrics, which quantify the degree of memorization, most studies rely on SSCD-based scores (see Eq. (5)), reporting metrics such as the average score, the 95th percentile score, and the percentage of memorized generations, with a threshold of 0.5 commonly indicating memorization. Additionally, [8] proposes a localized similarity metric (LS, see Eq. (6)) that measures similarity between the generated image and corresponding training image only within the localized memorization region. We present comparative results using the widely adopted SSCD metric in Fig. 6 and Fig. 7, while other metrics, including" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 321, + 75, + 438, + 200 + ], + "blocks": [ + { + "bbox": [ + 321, + 75, + 438, + 200 + ], + "lines": [ + { + "bbox": [ + 321, + 75, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 321, + 75, + 438, + 200 + ], + "type": "image", + "image_path": "f701e94eadb009f9e20942b349979d6e265558bb5229dc535eff97b8966947b5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 216, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 261 + ], + "type": "text", + "content": "Figure 6. Comparison with baselines. Across different detection signals (" + }, + { + "bbox": [ + 313, + 216, + 555, + 261 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 216, + 555, + 261 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 555, + 261 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 313, + 216, + 555, + 261 + ], + "type": "text", + "content": "), PRSS consistently achieves superior privacy-utility trade-offs. The improvement is particularly pronounced in global memorization compared to local memorization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 440, + 76, + 548, + 200 + ], + "blocks": [ + { + "bbox": [ + 440, + 76, + 548, + 200 + ], + "lines": [ + { + "bbox": [ + 440, + 76, + 548, + 200 + ], + "spans": [ + { + "bbox": [ + 440, + 76, + 548, + 200 + ], + "type": "image", + "image_path": "d5d280afac8367d66c49dd9e8a475d6b9dc35683989d4387f1e06bef24a6ca95.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 266, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 291 + ], + "type": "text", + "content": "LS, are detailed in the Supplementary Material for a more comprehensive evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 299, + 465, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 465, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 465, + 312 + ], + "type": "text", + "content": "4.2. Comparison with Baselines" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "text", + "content": "As shown in Fig. 6, PRSS consistently outperforms prompt engineering (PE) in achieving better privacy-utility trade-offs, regardless of the detection signal used (whether " + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "text", + "content": "). This is visually evident as the red and green lines (PRSS) lie further in the bottom-right direction compared to the yellow and blue lines (PE) for both global and local memorization scenarios. Notably, PRSS demonstrates astonishingly large improvements in global memorization (from the blue line to the green line and from the yellow line to the red line), while [8] achieves only marginal gains over [36] in this case (from the blue line to the yellow line and from the green line to the red line). For local memorization, although the margin of improvement is smaller, PRSS still achieves a significant improvement over PE. Specifically, even using the inferior detection signal " + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "text", + "content": ", PRSS (green line) achieves comparable performance with the method proposed by [8] (yellow line), which leverages the superior signal " + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "text", + "content": ". Equipped with " + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 313, + 316, + 556, + 557 + ], + "type": "text", + "content": ", PRSS can achieve the state-of-the-art results (red line). This highlights PRSS's robust capability to leverage advancements in detection signals." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Fig. 7 highlights the necessity of both components in PRSS, as evidenced by the red line lying further in the bottom-right direction compared to the yellow and green lines for both global and local memorization scenarios. The synergistic effect of combining PR and SS as discussed in Sec. 3.4, is particularly evident in the local memorization scenario, which aligns with [8]'s observation that local memorization poses greater challenges for existing mitigation strategies. Using PR alone effectively enhances privacy but results in suboptimal utility, while SS alone improves utility but achieves limited privacy protection. Together, they" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8188" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 75, + 288, + 201 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 288, + 201 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 288, + 201 + ], + "type": "image", + "image_path": "963b6e8d9d17ff6018a78a1202f1a9aec28734ebc378b21b33a836242b1a16a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 251 + ], + "type": "text", + "content": "Figure 7. Ablation studies of PRSS components. Using PR alone results in low utility, while SS alone cannot ensure privacy. Combining PR and SS achieves a synergistic effect, delivering the best privacy-utility trade-off." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": "leverage each other's strengths, achieving the best privacy-utility trade-off. For global memorization, either PR or SS alone significantly outperforms the baseline. However, their combination further enhances the trade-off, underscoring the robustness of PRSS in tackling both global and local memorization challenges." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 328, + 143, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 143, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 143, + 340 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 295, + 629 + ], + "type": "text", + "content": "Memorization in Generative Models The phenomenon of generative models memorizing training data has been extensively studied across domains, including language models [3, 4, 15, 16, 18, 20], image generation [2, 10, 11, 21, 22, 33], and video generation [9, 26], due to the significant privacy and legal risks it poses [5, 14, 29, 37]. Specifically, in diffusion models, [6, 31, 34] first validated the memorization issue on various models including DDPMs [13, 23] and Stable Diffusion [28] on datasets include Oxford Flowers [24], CelebA [19], CIFAR-10 [17] and LAION [30]. Further analysis by [32] revealed that Stable Diffusion's text-conditioning exacerbates overfitting, making its memorization of LAION particularly problematic. To better understand memorization, [34] categorized the phenomenon as Matching Verbatim, Retrieval Verbatim, and Template Verbatim based on the cause, distinguishing between one-to-one and many-to-many mappings in text-image pairs, while [8] classified memorization into global and local cases, demonstrating the practical significance of this distinction. In this work, we focus on Stable Diffusion due to its widespread use, open-source nature, and real-world legal challenges [29]. We adopt the framework of [8] to analyze both global and local memorization for a more nuanced evaluation of our proposed strategies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Detection Strategies for Diffusion Memorization. Early detection methods required generating large image sets to estimate density [6] or comparing generations with the entire training dataset [31], which were computationally intensive. Recent approaches [27, 36] introduced more efficient one-step detection strategies. [27] detected memorization by observing less concentrated cross-attention scores" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "for memorized generations than non-memorized ones. [36] proposed using the abnormally high magnitude of text-conditional noise predictions as a detection signal. Building on this, [8] incorporated localized memorization masks into the magnitude computation, offering a more accurate detection signal and achieving state-of-the-art performance. Our proposed PRSS approach leverages these detection mechanisms to trigger mitigation strategies. Notably, PRSS can adapt to any detection method by substituting the indicator variable condition in its framework. We employ the state-of-the-art method from [8] for optimal results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 204, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 467 + ], + "type": "text", + "content": "Mitigation Strategies for Diffusion Memorization. Mitigation strategies have evolved rapidly. Early methods, such as retraining models on de-duplicated data [6], were computationally expensive and showed limited effectiveness. [32] introduced randomness into text prompts or embeddings but indiscriminately affected non-memorized prompts, reducing overall utility. To improve on it, [7] proposed to compute the generation's similarity with the closest training image to engineer the image latent using the classifier guidance framework. However, this relies on computationally expensive nearest-neighbor searches across the training dataset. [36] proposed to conduct prompt engineering using the magnitude of text-conditional prediction as the loss, and [27] proposed to re-weight cross-attention scores for all tokens in the prompt. They offer improved utility-privacy trade-offs and achieve comparable performance. Most recently, [8] extended these efforts by leveraging localized magnitudes as loss functions for prompt engineering, achieving the best results to date. Our investigation reveals the limitations in the prevailing method's privacy/utility balance. In response, we introduce new mitigation strategies that achieve more favorable trade-offs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 472, + 480, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 480, + 483 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "content": "This paper introduces a novel PRSS method that consists of prompt re-anchoring and semantic prompt search as deliberately designed strategies to mitigate memorization in diffusion models. Empirical results show that PRSS demonstrates superior privacy-utility trade-offs over existing methods. Beyond its effectiveness, PRSS is simple and efficient to implement. It requires only minor modifications to the CFG equation to update noise predictions during inference, seamlessly integrating into the inference pipeline without additional training, fine-tuning, or reliance on training data. The efficacy of each component is validated through comprehensive analysis and ablation studies. A current limitation is that the performance of PRSS, as well as baseline mitigation strategies, relies heavily on the accuracy of the detection mechanism. When it fails to trigger accurately, they all revert to the standard Stable Diffusion process, compromising privacy preservation. Thus, improving the detection mechanism's accuracy in future works would orthogonally bolster the effectiveness of PRSS." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8189" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 114 + ], + "type": "text", + "content": "This work was supported in part by the Australian Research Council under Projects DP210101859 and FT230100549." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 157, + 281, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 281, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 281, + 168 + ], + "type": "text", + "content": "[1] Midjourney. https://www.midjourney.com/. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 170, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 295, + 203 + ], + "type": "text", + "content": "[2] Sanjeev Arora, Andrej Risteski, and Yi Zhang. Do gans learn the distribution? some theory and empirics. In International Conference on Learning Representations, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In USENIX Security Symposium, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 260, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 295, + 304 + ], + "type": "text", + "content": "[4] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022.8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 305, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 295, + 350 + ], + "type": "text", + "content": "[5] Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang, Nicolas Papernot, Andreas Terzis, and Florian Tramér. The privacy onion effect: Memorization is relative. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 351, + 295, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 405 + ], + "type": "text", + "content": "[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In USENIX Security Symposium, pages 5253-5270, 2023. 1, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 407, + 295, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 295, + 429 + ], + "type": "text", + "content": "[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In CVPR, 2024. 1, 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 430, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 463 + ], + "type": "text", + "content": "[8] Chen Chen, Daochang Liu, Mubarak Shah, and Chang Xu. Exploring local memorization in diffusion models via bright ending attention. *ICLR*, 2025. 1, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "type": "text", + "content": "[9] Chen Chen, Enhuai Liu, Daochang Liu, Mubarak Shah, and Chang Xu. Investigating memorization in video diffusion models. ICLR 2nd Workshop on Navigating and Addressing Data Problems for Foundation Models (DATA-FM), 2025. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "text", + "content": "[10] Yunhao Chen, Xingjun Ma, Difan Zou, and Yu-Gang Jiang. Towards a theoretical understanding of memorization in diffusion models. arXiv preprint arXiv:2410.02467, 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 544, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 598 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 601, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 622 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[14] H. H. Jiang, L. Brown, J. Cheng, M. Khan, A. Gupta, D. Workman, A. Hanna, J. Flowers, and T. Gebru. AI Art and Its Impact on Artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374. ACM, 2023. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[15] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. Large language models struggle to learn long-tail knowledge. arXiv preprint arXiv:2211.08411, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[16] Nikhil Kandpal, Eric Wallace, and Colin Raffel. Deduplicating training data mitigates privacy risks in language models. In International Conference on Machine Learning, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 184 + ], + "type": "text", + "content": "[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 186, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 239 + ], + "type": "text", + "content": "[18] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the Association for Computational Linguistics, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 242, + 555, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 555, + 283 + ], + "type": "text", + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[20] Fatemehsadat Mireshghallah, Anish Uniyal, Tianhao Wang, David Evans, and Taylor Berg-Kirkpatrick. Memorization in nlp fine-tuning methods. arXiv preprint arXiv:2205.12506, 2022.8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[21] Amir Naseh, Jaemin Roh, and Amir Houmansadr. Memory triggers: Unveiling memorization in text-to-image generative models through word-level duplication. arXiv preprint arXiv:2312.03692, 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[22] Ali Naseh, Jaechul Roh, and Amir Houmansadr. Understanding (un)intended memorization in text-to-image generative models. arXiv preprint arXiv:2312.07550, 2023. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 432 + ], + "type": "text", + "content": "[23] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "text", + "content": "[24] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 479, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 520 + ], + "type": "text", + "content": "[25] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, pages 14532-14542, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[26] Aimon Rahman, Malsha V. Perera, and Vishal M. Patel. Frame by familiar frame: Understanding replication in video diffusion models. arXiv preprint arXiv:2403.19593, 2024. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "text", + "content": "[27] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. ECCV, 2024. 1, 3, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "type": "text", + "content": "[29] Joseph Saveri and Butterick Matthew. Stable diffusion litigation, 2023. 2023. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8190" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 408 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 1, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[31] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In CVPR, pages 6048-6058, 2023. 1, 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. In NeurIPS, 2023. 1, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[33] Patrick Tinsley, Adam Czajka, and Patrick Flynn. This face does not exist... but it might be yours! identity leakage in generative models. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "text", + "content": "[34] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 1, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 329 + ], + "type": "text", + "content": "[35] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Advances in Neural Information Processing Systems, pages 51008-51025. Curran Associates, Inc., 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "type": "text", + "content": "[36] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In ICLR, 2024. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[37] Chiyuan Zhang, Daphne Ippolito, Katherine Lee, Matthew Jagielski, Florian Tramér, and Nicholas Carlini. Counterfactual memorization in neural language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 410, + 82, + 417 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 410, + 82, + 417 + ], + "spans": [ + { + "bbox": [ + 77, + 410, + 82, + 417 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8191" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_content_list.json b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9122086470058071a90e36745f94b52301f75402 --- /dev/null +++ b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_content_list.json @@ -0,0 +1,1263 @@ +[ + { + "type": "text", + "text": "Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 120, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aishik Konwer $^{1,2*}$ , Zhijian Yang $^{1}$ , Erhan Bas $^{1}$ , Cao Xiao $^{1}$ , Prateek Prasanna $^{3}$ , Parminder Bhatia $^{1}$ , Taha Kass-Hout $^{1}$ $^{1}$ GE Healthcare", + "bbox": [ + 163, + 208, + 833, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, Stony Brook University", + "bbox": [ + 263, + 263, + 730, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Biomedical Informatics, Stony Brook University", + "bbox": [ + 241, + 282, + 750, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "akonwer@cs.stonybrook.edu, prateek.prasanna@stonybrook.edu", + "bbox": [ + 241, + 301, + 751, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhijian.yang, erhan.bas, cao.xiao, paraminder.bhatia, taha.kass-hout}@gehealthcare.com", + "bbox": [ + 116, + 319, + 875, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 369, + 326, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundational models such as the Segment Anything Model (SAM) are gaining traction in medical imaging segmentation, supporting multiple downstream tasks. However, such models are supervised in nature, still relying on large annotated datasets or prompts supplied by experts. Conventional techniques such as active learning to alleviate such limitations are limited in scope and still necessitate continuous human involvement and complex domain knowledge for label refinement or establishing reward ground truth. To address these challenges, we propose an enhanced Segment Anything Model (SAM) framework that utilizes annotation-efficient prompts generated in a fully unsupervised fashion, while still capturing essential semantic, location, and shape information through contrastive language-image pretraining and visual question answering. We adopt the direct preference optimization technique to design an optimal policy that enables the model to generate high-fidelity segmentations with simple ratings or rankings provided by a virtual annotator simulating the human annotation process. State-of-the-art performance of our framework in tasks such as lung segmentation, breast tumor segmentation, and organ segmentation across various modalities, including X-ray, ultrasound, and abdominal CT, justifies its effectiveness in low-annotation data scenarios.", + "bbox": [ + 89, + 401, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 792, + 220, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With advancements in medical image analysis, there is an increasing need for sophisticated methods [20, 21, 29] to leverage the vast availability of radiology datasets (such as", + "bbox": [ + 89, + 818, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "X-ray, CT, and MRI) for accurate organ and tumor segmentation, as well as disease classification. The results of these tasks are crucial for physicians in designing effective treatment plans and surgical procedures. Several state-of-the-art deep learning-based foundational models, such as Vision-Language Models (VLMs), are now available for these purposes, relying on custom prompting to generate relevant predictions. However, they face two significant challenges: (1) despite being trained with only sparse prompts such as points or bounding boxes, these models still require human supervision for the prompt generation, leading to inefficiencies; and (2) many datasets lack comprehensive annotations, resulting in under-utilization during training of complex, data-hungry foundational architectures. Additionally, the high cost of human annotation efforts to create ground truth data can significantly escalate model development expenses.", + "bbox": [ + 511, + 369, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Popular architectures such as SAM and CLIP, have been extended to medical data, and led to innovations such as BiomedCLIP, Merlin, SAM-Med2D, and SAM-Med3D [6, 11, 47, 55]. Recently, numerous studies have focused on enhancing SAM by replacing geometric prompts and integrating semantic and spatial knowledge in an unsupervised manner through techniques such as (1) self-prompting, (2) class activation maps from CLIP, and (3) object localization models [19, 24, 57]. As shown in Fig. 1, self-prompting (b) does not require expert-provided prompts during inference. On the other hand, unsupervised prompts (c) in [19] are only used for training SAM to generate pseudo-labels for weakly supervised downstream tasks. They lack sufficient location and shape information, which could be provided by text-based prompts. Hence, the question arises: Can we come up with more refined and efficient prompts that can deliver stronger signals to foundational models without requiring intervention from annotators in both training and testing", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at GE Healthcare", + "bbox": [ + 107, + 875, + 362, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20990", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb07502a342acdf16cb10ed1cf4f3b814277888c18e400f1078248d11ecce69a.jpg", + "image_caption": [ + "Figure 1. Overview of our model: (a) SAM and SAM-based approaches rely on expert prompts during both training and inference. (b) SAM-SP [57] introduces a self-prompting module, eliminating the need for expert prompts during inference. (c) MedCLIP-SAM [19] uses unsupervised semantic prompts to generate pseudo-labels through SAM. (d) Our approach not only combines semantic, location, and generic information via unsupervised prompts but also introduces a preference-based alignment module to reward or penalize the model." + ], + "image_footnote": [], + "bbox": [ + 173, + 89, + 823, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stages? State-of-the-art models also lack comprehensive integration of semantic, locational, and generic class information in their prompting strategies. In our approach, we leverage CLIP, VQA, and LLM models [1, 55, 56] to extract this discriminative information, improving segmentation performance in unsupervised settings.", + "bbox": [ + 89, + 455, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenges posed by limited annotated datasets, several weakly supervised semantic segmentation algorithms [2, 19] have been proposed. Single-stage methods [2] leverage coarse image labels to perform segmentation in an end-to-end fashion, while many two-stage approaches [19] utilize foundational models to generate pseudo labels that can be used to train downstream segmentation architectures. Additionally, human-in-the-loop frameworks [23] have proven beneficial for tasks such as image generation, segmentation, and prognosis, involving annotators to refine segmentation labels or assess the plausibility of synthetic images. This feedback can either be incorporated back into the model or used as ground truth to train a reward function via reinforcement learning (RL), significantly reducing the need for annotated training data. However, these methods still rely on explicit reward modeling, which prevents end-to-end training. In some cases, annotators must provide complex ratings or refinements, making it difficult to train a reward function efficiently. We ask another question: Can we skip training a reward function and develop a straightforward, end-to-end pipeline that relies on simple annotator preferences? To address this, we draw inspiration from direct preference optimization tech", + "bbox": [ + 89, + 551, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "niques in the LLM preference alignment literature. We propose a preference-based reward model within our framework, using a novel loss function for end-to-end training. Our approach thus aligns with human preferences without the need for explicit reward modeling and is both easy to implement and train.", + "bbox": [ + 511, + 455, + 906, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our first goal is to develop refined and efficient prompts to fine-tune the SAM-Med2D encoder on diverse datasets, including 2D chest X-rays, breast ultrasound, and 3D abdominal CT scans. Initially, we input an image into the encoder alongside the BiomedCLIP [55] and MedViInT (VQA) [56] models. Corresponding texts, examples such as \"Chest X-ray\" and \"Describe the condition of the lungs and location of pathologies,\" are fed into BiomedCLIP and MedViInT, respectively. We gather generic information for the disease class from GPT-4, which is combined with the answers from the VQA model. Saliency maps generated by the CLIP model undergo dense CRF postprocessing to obtain bounding boxes. These bounding boxes and textual prompts are then inputted into the prompt encoder. Subsequently, the mask decoder receives both the image encoder embeddings and prompt embeddings to produce the segmentation maps.", + "bbox": [ + 511, + 549, + 908, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After fine-tuning the prompting module on a small proportion of annotated data, we introduce our second major idea: simulating human feedback through a AI preference alignment module. We generate multiple segmentation candidates for a given image by thresholding the SAM output probabilities at various levels. These candidates are rated", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20991", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on a scale of 1 to 4 based on the overlap between the candidates and ground truth, mimicking the evaluation process of a human annotator. This approach does not require explicit access to ground truth data for training, thus making our task a form of semi-supervised segmentation. We propose a DPO-inspired [38] loss function that encourages the model to prioritize desirable segmentation outputs by rewarding higher-rated candidates and penalizing lower-rated ones. The model is thus trained on the remaining portion of the dataset, without annotations, to perform relevant medical image segmentation tasks.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall our contributions can be summarized as follows:", + "bbox": [ + 109, + 258, + 480, + 272 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose refined and efficient unsupervised prompting strategies that deliver comprehensive information about location, semantics, and general disease/organ characteristics to our SAM-Med2D-based framework. Such an approach enhances segmentation performance while reducing reliance on human input for geometric prompts.", + "- We introduce a novel DPO-inspired loss function that facilitates semi-supervised model improvement using simulated human-in-the-loop feedback, eliminating the need for an explicit reward function. The framework generates multiple segmentation maps and rates them based on segmentation overlap, mimicking the evaluation of a human annotator. The model learns to distinguish between favorable and unfavorable candidates effectively." + ], + "bbox": [ + 91, + 273, + 482, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 502, + 232, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Vision-language models for medical domain", + "text_level": 1, + "bbox": [ + 89, + 527, + 462, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP [37] has gained much popularity in medical image analysis. [14] fine-tuned CLIP on various PubMed articles to create PubMedCLIP. MedCLIP [48] leverages unpaired image and text datasets along with a semantic-matching loss to align visual and textual information. Windsor et al. [49] employ unimodal self-supervision, local-global contrastive losses, and data augmentation to enhance zero-shot performance and retrieval task efficiency in low-resource data settings. Some modality-specific CLIP variants [15, 54] have been developed for Chest X-ray and Mammograms due to the readily available image-text data in these areas. However, BiomedCLIP [55] stands out as the most recent model, excelling in scalability and performance across diverse multi-organ cross-modal retrieval tasks. Therefore, we have integrated BiomedCLIP into the CLIP-driven bounding box generation module of our framework.", + "bbox": [ + 89, + 551, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on the success of large language models (LLMs) such as LLaMa [44] and GPT [35], researchers have explored merging visual features with textual representations using techniques such as cross-attention, Qformer, instruction tuning, and projection layers. This effort has resulted in vision-language models (VLMs) [4, 26, 30, 58], including Flamingo, BLIP-2, LLaVA, and", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MiniGPT, which were further adapted for the medical domain through pretraining on multimodal medical datasets. Med-Flamingo [33] is the first medical visual question-answering (VQA) model with few-shot generation capabilities. RadFM [50] serves as a foundational model that also incorporates 3D volume data. Pretraining on extensive datasets, PMC-15M and PMC-VQA [55, 56], has facilitated the development of LLaVA-Med [25] and MedViInT [56], respectively. Our segmentation pipeline leverages MedViInT's capabilities for enhanced localization and shape-based answer generation related to tumors, organs, and disease manifestations.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. SAM for multimodal biomedical data", + "text_level": 1, + "bbox": [ + 511, + 282, + 836, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MedSAM and SAM-Med2D [11, 31] focused on adapting SAM [18] for medical applications by fine-tuning it on 2D medical datasets, while SAM-Med3D [47] introduced alternative modules to accommodate 3D volumes. Efficient approaches, for example, AutoSAM [41] utilize trainable prompt encoders, whereas FastSAM3D [42] employs flash attention to accelerate inference. MedLSAM [24] proposed a localization framework to generate 3D bounding boxes as prompt input. However, most of these methods require ground truth data (bounding boxes or points) for training the prompt encoder, whereas we propose an unsupervised route for the same. We leverage ad-hoc VLM models such as CLIP, VQA, and GPT-4 together to propagate comprehensive information—encompassing semantics, location, and generic disease/organ information—that significantly enhances segmentation performance.", + "bbox": [ + 511, + 306, + 906, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Human-in-the-loop feedback", + "text_level": 1, + "bbox": [ + 511, + 560, + 769, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human-in-the-loop training paradigm in medical imaging has primarily focused on two areas: active learning [34], which identifies optimal data points for labeling to maximize performance, and interactive feedback [8] on model predictions to calibrate parameters. Examples include UI-Net, DeepIGeoS, and BIFseg [5, 45, 46], which utilize expert-provided scribbles or bounding boxes alongside geodesic transforms and graph-cut techniques to refine segmentation labels. Rao et al. [39] introduced IMIL, the first framework that assigns clinicians to guide data augmentations on mispredictions, emphasizing disease-relevant regions while eliminating irrelevant ones. Recently, the success of human feedback in instruction-tuning, and aligning large language model outputs through RL objective [52, 53], has led researchers to utilize human preferences to evaluate synthetic natural and histopathology images, thereby improving image-to-text models [23, 43] by training a reward function. Training a reward function requires the curation of dedicated human preference datasets and also prohibits the framework from operating in an end-to-end manner. Additionally, it often demands advanced do", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20992", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ff8d147200b26c99039395c2a0755a065f7f6ac071c264c6594a6c972ef87cf5.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed framework for semi-supervised segmentation: Unsupervised geometric and text prompts, obtained from pretrained BiomedCLIP, MedVInT, and GPT-4 models, are fed into the prompt encoder for finetuning the framework on a small fraction of annotated data. In the next stage, we simulate a virtual annotation process that assigns ratings to the generated segmentation candidates, which are used to fine-tune the decoder. This stage handles unannotated data, as the model does not rely on ground truth for direct supervision but only for rating while simulating a human annotator's feedback." + ], + "image_footnote": [], + "bbox": [ + 200, + 103, + 799, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "main knowledge from annotators, which can be costly. Our approach addresses this challenge by using direct preference optimization [38] to fulfill the RL objective, enabling the framework to serve as its own reward model. The model aligns generated proposals with appropriate preferences or ratings, leading to performance improvements, even with limited annotated data. These preferences are generated through a simulation mimicking human-in-the-loop feedback. Crucially, our method is easier to train and implement than traditional reward function-based RL pipelines.", + "bbox": [ + 88, + 588, + 482, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 753, + 225, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. Given an input image, our aim is to generate prompts in the form of bounding boxes and textual queries, which will be used to produce segmentation masks. The image is processed by three pretrained components: the SAM-Med2D encoder, BiomedCLIP, and MedVInT. Concurrently, GPT-4 curates generic information about the concerned disease or organ. This information, combined with the BiomedCLIP-generated box prompt and the MedVInT", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "provided textual answer prompt, serves as input for the prompt encoder. The prompt generation process is detailed in Sec. 3.1. Next, both the encoded image and prompt embeddings are fed into the decoder to create segmentation maps. Following this initial fine-tuning on annotated data, we introduce a novel loss to mimic human-in-the-loop feedback. By thresholding the output probabilities, our model produces four different segmentation candidates, and we incorporate insights about the quality of these candidates into the framework through the proposed loss function. This part, trained with unannotated data, is elaborated in Sec. 3.2.", + "bbox": [ + 511, + 588, + 906, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Visual and textual prompt generation", + "text_level": 1, + "bbox": [ + 511, + 768, + 836, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual prompt. To generate visual prompts, we first input the image along with corresponding text prompts (e.g. \"chest x-ray\", \"benign breast tumor\", \"left kidney\", \"liver\", etc.) into BiomedCLIP, a foundational model pretrained on millions of medical image-caption pairs. Next, we leverage gScoreCAM [10] to create a saliency map highlighting targeted regions (organs or tumors) in the image corre", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20993", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sponding to the supplied text. These saliency maps are then post-processed with a conditional random field filter [22] to produce coarse segmentation masks. We apply an area constraint, retaining the largest connected component or up to two, depending on the dataset (the lung dataset may yield two components). Finally, we identify the bounding box coordinates within these closed components for the box prompt, while randomly sampling several points from the designated area for the point prompt.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual prompt. We extract visual embeddings from a given image by processing it through a vision encoder derived from the PMC-CLIP architecture [28]. This encoder features a pretrained ResNet50 [16] and a trainable projection layer constructed with stacks of transformer decoder blocks. Next, we create a prompt template that incorporates the question for the image, formatted as \"Question: $\\{\\}$ , Answer is:\". This prompt is then passed through a tokenization embedding layer initialized with the weights of PMC-LLaMA [51] to generate the text embedding. Finally, we concatenate the visual and text embeddings to form the input space for a pre-initialized multimodal transformer decoder. The answer generated from this VQA setup delivers essential information regarding the shape and location of anatomical structures and pathologies. Sample questions included in the prompt are: \"What is the shape of the liver and where is it located?\", and \"What is the shape of breast tumor and where is it located?\" Additionally, we prompt GPT-4 with the relevant organ or disease label to obtain a generic description of its characteristics. Finally, both types of textual prompts are concatenated and provided as input to the prompt encoder, described in the following paragraph.", + "bbox": [ + 91, + 228, + 482, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt encoder and mask decoder. The prompt encoder, same as the one in SAM, supports three types of prompts: point, box, and text. Point and box prompts are represented by their positional encodings—specific coordinates for points, and the top-left and bottom-right corners for boxes—along with learnable feature representations. Text prompts are processed through a pretrained BiomedCLIP encoder to generate corresponding text embeddings. All prompt embeddings are then projected into 256-dimensional vectors. The feature map from the first iteration of the model, is downsampled through multiple convolutional layers followed by GELU activation to match the 256-channel dimension. Finally, these downsampled masks are combined element-wise with visual encoder embeddings. The mask decoder takes both the prompt embeddings and the visual embeddings to produce a segmentation map. The architecture is illustrated in Fig. 2.", + "bbox": [ + 89, + 561, + 482, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Segmentation proposal generation and Preference Alignment", + "text_level": 1, + "bbox": [ + 89, + 830, + 480, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposal generation. After fine-tuning our framework, comprising the prompt encoder and mask decoder, using", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ground truth for a fraction (10%) of the dataset, we propose integrating AI-based preferences for the next training episode. First, we generate multiple segmentation proposals by sampling different thresholds from the pixel probability scores output by the model. We then simulate an expert annotator to evaluate the quality of these proposals, assigning ratings based on the overlap between each generated proposal and the corresponding ground truth. Although ground truth is not explicitly used in this training phase, it plays a passive role by enabling the simulation of preferences in the absence of an actual human annotator in our experiments. We also experiment with two alternative rating mechanisms: one inspired by SAM, which backpropagates loss solely for the best candidate, and another that ranks all proposals rather than merely rating them. After obtaining the ratings or rankings, we fine-tune our decoder using the direct preference optimization technique to better align the segmentation outputs with the preferences of the virtual annotator.", + "bbox": [ + 511, + 90, + 903, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RLHF Preliminaries. Language models typically utilize a reward function to align their generated responses with user preferences. This preference distribution is often modeled using the Bradley-Terry [7] model, which operates on pairwise comparisons. For a given prompt $X$ , when the language model produces two responses, one more favorable, $Y_{m}$ , and the other less favorable, $Y_{l}$ , the distribution can be expressed as:", + "bbox": [ + 511, + 377, + 903, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(Y _ {m} > Y _ {l} \\mid X\\right) = \\frac {e ^ {r ^ {*} \\left(X , Y _ {m}\\right)}}{e ^ {r ^ {*} \\left(X , Y _ {m}\\right)} + e ^ {r ^ {*} \\left(X , Y _ {l}\\right)}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 508, + 903, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameters of the reward model can be estimated via Maximum Likelihood Estimation (MLE) and the goal is to minimize the below loss function:", + "bbox": [ + 511, + 553, + 903, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {R}} = - \\mathbb {E} _ {(X, Y _ {m}, Y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r _ {\\phi} (X, Y _ {m}) - r _ {\\phi} (X, Y _ {l})\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 609, + 903, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is the logistic function and $\\mathcal{D}$ denotes the dataset of preferences.", + "bbox": [ + 511, + 642, + 903, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reinforcement Learning from Human Feedback (RLHF) is a widely used method that involves training a reward model based on user ratings. The primary goal is to discover an optimal policy with parameter $\\pi_{\\theta}$ that maximizes this reward function while incorporating a KL divergence term to ensure the model's outputs do not deviate significantly from the original policy $\\pi_{ref}$ . The optimization problem can be formulated as follows:", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {X \\sim \\mathcal {D}, Y \\sim \\pi_ {\\theta} (Y | X)} \\left[ r _ {\\phi} (X, Y) \\right] \\tag {3} \\\\ - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ \\pi_ {\\theta} (Y \\mid X) | | \\pi_ {r e f} (Y \\mid X) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 805, + 903, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Direct Preference optimization. Due to the discrete nature of language generation, this objective is not differentiable, which usually necessitates optimization through", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20994", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reinforcement learning. The primary language model then leverages this framework to align its outputs with user ratings, ultimately generating high-scoring responses. In contrast, Direct Preference Optimization (DPO) is a more recent and streamlined approach that eliminates the need for a separate reward model. Instead, it fine-tunes the main language model directly using user preferences. This is achievable because DPO focuses on optimizing the policy itself rather than the reward function. The maximum likelihood objective for a parameterized policy $\\pi_{\\theta}$ in such a case can be formulated as:", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\pi_ {\\theta}; \\pi_ {r e f}\\right) = - \\mathbb {E} _ {\\left(X, Y _ {m}, Y _ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(Y _ {m} \\mid X\\right)}{\\pi_ {r e f} \\left(Y _ {m} \\mid X\\right)} \\right. \\right. \\tag {4} \\\\ \\left. \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(Y _ {l} \\mid X\\right)}{\\pi_ {r e f} \\left(Y _ {l} \\mid X\\right)}\\right) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 276, + 483, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ is the weight applied to reward or penalize the responses. In DPO, models are fine-tuned using pairs of outputs, explicitly comparing a preferred response $Y_{m}$ with less preferred ones $Y_{l}$ . We adapt this concept for our application in the imaging domain, where multiple segmentation candidates $Y_{1},\\ldots ,Y_{4}$ are generated for the same {image, prompt} pair $I$ . With ratings available for each candidate, we assign weights $\\beta_{1},\\beta_{2}$ to these candidates (from best to worst) to reward the higher-quality segmentations $Y_{1},Y_{2}$ and penalize the less desirable ones $Y_{3},Y_{4}$ . Consequently, Equation 4 is modified to create Equation 5. This approach allows us to incorporate real-world annotator preferences into our method through a straightforward yet effective loss function. The parameters of the initially fine-tuned model from Sec. 3.1 are denoted as $\\pi_{\\mathrm{fine}}$ . Our objective is to find the optimal parameters $\\pi_{\\psi}$ for the final architecture without deviating much from $\\pi_{\\mathrm{fine}}$ .", + "bbox": [ + 89, + 335, + 483, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} (\\pi_ {\\psi}; \\pi_ {f i n e}) = - \\mathbb {E} _ {(I, Y _ {1}, Y _ {2}, Y _ {3}, Y _ {4}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta_ {1} \\log \\frac {\\pi_ {\\psi} (Y _ {1} \\mid I)}{\\pi_ {f i n e} (Y _ {1} \\mid I)} \\right. \\right. \\\\ \\left. + \\beta_ {2} \\log \\frac {\\pi_ {\\psi} (Y _ {2} \\mid I)}{\\pi_ {f i n e} (Y _ {2} \\mid I)} - \\beta_ {2} \\log \\frac {\\pi_ {\\psi} (Y _ {3} \\mid I)}{\\pi_ {f i n e} (Y _ {3} \\mid I)} - \\beta_ {1} \\log \\frac {\\pi_ {\\psi} (Y _ {4} \\mid I)}{\\pi_ {f i n e} (Y _ {4} \\mid I)}\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 604, + 488, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Design and Results", + "text_level": 1, + "bbox": [ + 89, + 679, + 393, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluated our framework for semi-supervised segmentation using three public datasets covering lung, breast tumor, and abdominal organ segmentation across X-ray, ultrasound, and CT. For Ultrasound Breast Tumor Segmentation, we used 810 images from BUSI [3] and UDIAT [9], with 600 for training and 210 for testing. For Chest X-ray Lung Segmentation, we used 27,132 images from COVID-QU-Ex [12], spanning normal, lung opacity, viral pneumonia, and COVID-19 cases. For Abdominal CT Organ Segmentation, 200 annotated CT scans from the AMOS-CT training set were used for training. Full dataset details, including train-test splits, are in the supplementary Sec. 6.", + "bbox": [ + 89, + 704, + 485, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our method is implemented in PyTorch [36] on an EC2 instance (with 64 GB NVIDIA T4 Tensor Core GPUs). For feature extraction, we utilize the SAM-Med2D pretrained encoder. First we use annotations for only $10\\%$ of the training dataset, during which all components (visual encoder, prompt encoder, and mask decoder) are fine-tuned. In this step, only the unsupervised prompting strategy is employed. Bounding box and point prompts are used together for all experiments. The remaining portion of the dataset was used in an unannotated form to train the DPO-driven alignment strategy. We optimize the model using the Adam optimizer [13], training for 15 epochs for prompt module fine-tuning and 30 epochs for alignment. In both stages, the initial learning rate is set to 1e-4 and is halved every 10 epochs. All images are resized to a resolution of $256 \\times 256$ using the same resizing strategy as SAM-Med2D. Similar to our framework, we have also fine-tuned SOTA SAM-based methods on medical imaging datasets using LoRA by adding low-rank matrices to transformer layers. This enables efficient adaptation with fewer parameters, enhancing domain-specific performance while preserving pretrained features. The loss function used during the initial fine-tuning is a 20:1 weighted combination of focal loss [27] and Dice loss [32]. Details on the preference alignment module and proposal generation are provided in the supplementary material Sec. 6. Dice Similarity Coefficient, Intersection over Union (IoU), and Surface Dice Similarity (SDC) scores are used to evaluate segmentation performance.", + "bbox": [ + 511, + 90, + 906, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 511, + 542, + 803, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Fig. 3 compares the performance of our framework with relevant methods (U-Net [40], nnU-Net [17], SAM [18], SAM-Med2D [11], Self-prompt). The self-prompt method is designed as a variant of [57], excluding the knowledge distillation module. We also design a prompt-only baseline of our framework which is trained on different splits of fully annotated data. Our framework is initially trained with only the prompting module using $10\\%$ of data (annotated). As a result, both the prompt-only and final versions exhibit identical performance on this $10\\%$ subset. For the final model, the additional data used for training are considered unannotated, as training the alignment mechanism does not use supervision from ground truth. SOTA methods directly use ground truth for the entire data subset. Nevertheless, our method consistently outperforms them in limited data settings $(10 - 50\\%)$ due to the preference alignment mechanism. Our architecture thus reduces the reliance on large datasets, highlighting its lower annotation requirements, which makes it highly effective in low-data regimes. On the Chest-Xray dataset, for instance, with just $20\\%$ of the data, our method achieves a Dice score of 78.87, compared to 58.66 (U-Net), 60.97 (nnU-Net), 61.64 (SAM),", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20995", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/050dfb3d8688e5f902c8fdcb41ded97f761872112a1c636967ced906c1fba6b7.jpg", + "image_caption": [ + "Figure 3. Quantitative comparison with SOTA. Dice score (for Chest Xray, Breast USD) and mean Dice score (for AMOS CT) have been shown to measure the model segmentation performance on different proportions of training data $(10\\%, 20\\%, 50\\%,$ and $100\\%)$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 357, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/491cb3ff6193712fc86b0dc8a6ee47923f533c5f8c0cbc52a182c46747ba49f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 88, + 645, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adf7f09c9307423f79d8a845ac3a4a6fa65799c80aa5665f412048bb55fa6ec1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 88, + 905, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab375ac6a1e9a03b9d80d0a0c9216ea07abf42b6ebe17e04345333489888f38d.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons were made between the segmentation results of nnUnet, SAM-Med2D, and our framework on 2D datasets. BiomedCLIP-based saliency maps are also depicted. Experiments were conducted in $50\\%$ data settings." + ], + "image_footnote": [], + "bbox": [ + 133, + 268, + 867, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "67.81 (SAM-Med2D), and 68.41 (Self-prompt). Both U-Net and nnU-Net are known to be data-hungry models, struggling with smaller data subsets. Our usage of visual and textual prompts, derived through BiomedCLIP and VLM models, offers superior signals to those used in SAM and SAM-Med2D. Unlike these SAM variants, Self-Prompt SAM does not require expert-supplied prompts during inference; it instead generates prompts from the output masks in each iteration. This method offers a slight improvement $(+1\\%)$ over SAM-Med2D. Our method shows more stable performance gains across different data subsets, while other SOTA methods exhibit steeper improvements. In the $50\\%$ data setting, our semi-supervised framework achieves a Dice score of 89.68, compared to 91.42 for the supervised prompt-only version. Similar result trends are observed in the breast ultrasound dataset. Notably, the performance jump of our method from $20\\%$ to $50\\%$ data is much more pronounced than from $10\\%$ to $20\\%$ . This can be due to the nature of the dataset, as the model requires more data to achieve precise and accurate segmentation of breast tumors. We also evaluated our method on a 3D abdominal organ segmentation dataset, including organs such as the liver, kidneys, spleen, pancreas, aorta, bladder, etc (see Fig. 3). Our method outperforms the SOTA across all data proportions except for the full dataset. With $20\\%$ of the data, our method achieves a mean Dice score of 77.69, surpassing U", + "bbox": [ + 91, + 494, + 483, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Net (59.35), nnU-Net (65.21), SAM (64.93), SAM-Med2D (66.57), SAMMed3D (72.54), and Self-prompt (71.83). At the $50\\%$ data setting, it reaches a Dice score of 85.70, comparable to the 86.36 achieved by the prompt-only version.", + "bbox": [ + 511, + 494, + 908, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/769e6a3b7bc6d232524c11ee4463ef95231d2e5eeded689edb65501c78e7e839.jpg", + "image_caption": [ + "Figure 5. Segmentation maps of different anatomical structures (liver and spleen) for SAM-Med3D and our method" + ], + "image_footnote": [], + "bbox": [ + 535, + 570, + 885, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. We present segmentation maps for 3 datasets in Fig. 4 and Fig. 5, generated by our model trained at $50\\%$ data settings. The saliency maps could correctly highlight the target regions and proved to be an excellent source of supervision. It can be noted from Fig. 4 that our segmentation quality around the boundaries of tumor or lung is much superior compared to both nnUnet and SAM-Med2D. In Fig. 5, SAM-Med3D tended to under-segment the edges of both the abdominal organs, the liver,", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20996", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/631aac687fca7e61ed569ba628d958d8da741c501dcba7d13e092bf33879b0cf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SupervisionMethodsChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedOurs74.3078.8764.3575.8877.6978.34
20%- Alignment75.0279.1367.5181.3879.2080.66
10%- Alignment68.4375.6061.4473.6274.7776.06
- Alignment - VQA66.9073.3557.6070.5374.0875.41
- Alignment - VQA - GPT463.1672.7654.2668.8973.1674.89
- Alignment - VQA - CAM50.3857.0245.1659.0569.9771.45
", + "bbox": [ + 176, + 87, + 823, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/60c0a3f133cb85d8c15c4423613565fa7baed9e7fa12441b6c92e906756ad6c4.jpg", + "table_caption": [ + "Table 1. Ablation results demonstrating effectiveness of major components." + ], + "table_footnote": [], + "table_body": "
SupervisionPreference AlignmentChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedLoss for best candidate70.3777.0961.7673.8175.0176.36
Rating73.9978.4163.9775.5277.2377.98
Ranking74.3078.8764.3575.8877.6978.34
10% + 20% unannotatedRanking79.7485.1573.5684.2380.5481.95
10% + 40% unannotatedRanking88.9689.6885.9288.1584.3085.70
", + "bbox": [ + 176, + 236, + 823, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation results on different proportions of training data for 3 types of preference scoring strategies.", + "bbox": [ + 176, + 348, + 818, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and spleen. Additional results are in supplementary Sec. 1.", + "bbox": [ + 89, + 388, + 480, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 420, + 251, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of major components. We conduct several ablation experiments (see Tab. 1) to evaluate the contribution of each module in our architecture. For simplicity, we focus on the Chest-Xray dataset for analysis. Before incorporating the DPO-driven alignment strategy, all baselines were trained with $10\\%$ labeled data. First, we developed a naive baseline (last row), which relies solely on textual answers generated from GPT-4. As expected, it performed poorly, achieving a 57.02 Dice score. A second baseline (second-last row) was designed to use only visual prompts from BiomedCLIP, which provides semantic information. This method significantly benefited from the coarse segmentation masks derived from CLIP saliency maps, improving the Dice score to 72.76. Next, we combined both textual and visual prompts to form a third baseline (-Alignment-VQA), which resulted in a slight improvement $(+0.59\\%)$ over the CLIP-only baseline. Finally, we integrated VQA into the prompting strategy to obtain answers related to the shape and location of the target regions, completing our fully empowered unsupervised prompting strategy (-Alignment). This model achieved a Dice score of 75.60 with $10\\%$ data and improved to 79.13 with $20\\%$ data. In comparison, with the preference alignment mechanism, fine-tuning the prompting module with $10\\%$ annotated data and training the alignment module on an additional $10\\%$ unannotated data achieved a Dice score of 78.87. This is on par with the fully supervised prompt-only model, underscoring the effectiveness of our alignment module.", + "bbox": [ + 89, + 445, + 482, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Preference Scoring strategies. We conducted several ablation studies (detailed in Tab. 2) to evaluate the effec", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tiveness of the preference-scoring strategy. One baseline approach involved backpropagating the loss based only on the best candidate, while another compared ranking the candidates rather than simply rating them. The results are presented for different proportions of unannotated data, on top of the foundational fine-tuning of both the prompt encoder and the decoder using $10\\%$ annotated data. At $10\\%$ unannotated data, both the ranking and rating approaches performed similarly (mean Dice scores of 77.69 and 77.23 for AMOS CT, respectively), significantly outperforming the best-candidate-only method (75.01). Additionally, we observed that as the proportion of unannotated data increased, the model's performance improved substantially.", + "bbox": [ + 511, + 388, + 906, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to noisy rating. We experimented with introducing noise into the rating mechanism by flipping the ratings of closely ranked candidates to enhance the framework's robustness. Results are in supplementary material Sec. 2.", + "bbox": [ + 511, + 585, + 906, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a novel training strategy to enhance SAM representations for semi-supervised medical image segmentation. We extract integrated semantic, location, and shape information from pretrained vision-language models in an unsupervised manner. This information is used as refined prompts for our model. We also implement an optimal policy, inspired by direct preference optimization in language models. This enables human-in-the-loop feedback simulation within a streamlined training framework, without the need for a separate reward function or extensive knowledge from annotators. These modules ensure that our framework outperforms state-of-the-art methods across datasets spanning multiple modalities in low-annotation data scenarios.", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20997", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank everyone from the GE Healthcare Science & Technology Organization for their support and contributions to the project. The primary model development and evaluations were conducted during a summer internship at GEHC. After the internship, Konwer and Prasanna were supported by the following grants during the paper writing and rebuttal process: NIH 1R01CA297843-01, NIH 3R21CA258493-02S1, and 1R21CA258493-01A1.", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 251, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] Peri Akiva and Kristin Dana. Towards single stage weakly supervised semantic segmentation. arXiv preprint arXiv:2106.10309, 2021. 2", + "[3] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 2", + "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3", + "[5] Mario Amrehn, Sven Gaube, Mathias Unberath, Frank Schebesch, Tim Horz, Maddalena Strumia, Stefan Steidl, Markus Kowarschik, and Andreas Maier. Ui-net: Interactive artificial neural networks for iterative image segmentation based on a user model. arXiv preprint arXiv:1709.03450, 2017. 3", + "[6] Louis Blankemeier, Joseph Paul Cohen, Ashwin Kumar, Dave Van Veen, Syed Jamal Safdar Gardezi, Magdalini Paschali, Zhihong Chen, Jean-Benoit Delbrouck, Eduardo Reis, Cesar Truyts, et al. Merlin: A vision language foundation model for 3d computed tomography. arXiv preprint arXiv:2406.06512, 2024. 1", + "[7] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5", + "[8] Samuel Budd, Emma C Robinson, and Bernhard Kainz. A survey on active learning and human-in-the-loop deep learning for medical image analysis. Medical image analysis, 71: 102062, 2021. 3", + "[9] Michal Byra, Piotr Jarosik, Aleksandra Szubert, Michael Galperin, Haydee Ojeda-Fournier, Linda Olson, Mary O'Boyle, Christopher Comstock, and Michael Andre. Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control, 61:102027, 2020. 6, 2", + "[10] Peijie Chen, Qi Li, Saad Biaz, Trung Bui, and Anh Nguyen. gscorecam: What objects is clip looking at? In Proceedings" + ], + "bbox": [ + 93, + 276, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the Asian Conference on Computer Vision (ACCV), pages 1959-1975, 2022. 4", + "[11] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 3, 6", + "[12] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and coid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 6, 2", + "[13] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6", + "[14] Sedigheh Eslami, Christoph Meinel, and Gerard De Melo. Pubmedclip: How much does clip benefit visual question answering in the medical domain? In Findings of the Association for Computational Linguistics: EACL 2023, pages 1181-1193, 2023. 3", + "[15] Shantanu Ghosh, Clare B Poynton, Shyam Visweswaran, and Kayhan Batmanghelich. Mammo-clip: A vision language foundation model to enhance data efficiency and robustness in mammography. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 632–642. Springer, 2024. 3", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[17] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6", + "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3, 6", + "[19] Taha Koleilat, Hojat Asgariandehkordi, Hassan Rivaz, and Yiming Xiao. Medclip-sam: Bridging text and image towards universal medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 643-653. Springer, 2024. 1, 2", + "[20] Aishik Konwer, Xuan Xu, Joseph Bae, Chao Chen, and Pratek Prasanna. Temporal context matters: Enhancing single image prediction with disease progression representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18824-18835, 2022. 1", + "[21] Aishik Konwer, Xiaoling Hu, Joseph Bae, Xuan Xu, Chao Chen, and Prateek Prasanna. Enhancing modality-agnostic representations via meta-learning for brain tumor segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21415–21425, 2023. 1", + "[22] Philipp Krahenbuhl and Vladlen Koltun. Parameter learning and convergent inference for dense random fields. In Inter-" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20998", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national conference on machine learning, pages 513-521. \nPMLR, 2013. 5", + "[23] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3", + "[24] Wenhui Lei, Xu Wei, Xiaofan Zhang, Kang Li, and Shaoting Zhang. Medlsam: Localize and segment anything model for 3d medical images. arXiv preprint arXiv:2306.14752, 2023. 1, 3", + "[25] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3", + "[27] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 6", + "[28] Weixiong Lin, Ziheng Zhao, Xiaoman Zhang, Chaoyi Wu, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-clip: Contrastive language-image pre-training using biomedical documents. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 525–536. Springer, 2023. 5", + "[29] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3", + "[31] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 3", + "[32] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6", + "[33] Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Eduardo Pontes Reis, and Pranav Rajpurkar. Med-flamingo: a multimodal medical few-shot learner. In Machine Learning for Health (ML4H), pages 353-367. PMLR, 2023. 3", + "[34] Vishwesh Nath, Dong Yang, Bennett A. Landman, Daguang Xu, and Holger R. Roth. Diminishing uncertainty within the training pool: Active learning for medical image segmentation. IEEE Transactions on Medical Imaging, 40(10): 2534-2547, 2021. 3", + "[35] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3", + "[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[38] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4", + "[39] Adrit Rao, Andrea Fisher, Ken Chang, John Christopher Panagides, Katherine McNamara, Joon-Young Lee, and Oliver Aalami. Imil: Interactive medical image learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5241–5250, 2024. 3", + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6", + "[41] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3", + "[42] Yiqing Shen, Jingxing Li, Xinyuan Shao, Blanca Inigo Romillo, Ankush Jindal, David Dreizin, and Mathias Unberath. Fastsam3d: An efficient segment anything model for 3d volumetric medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 542-552. Springer, 2024. 3", + "[43] Shenghuan Sun, Greg Goldgof, Atul Butte, and Ahmed M Alaa. Aligning synthetic medical images with clinical knowledge using human feedback. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[44] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[45] Guotai Wang, Wenqi Li, Maria A Zuluaga, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sebastien Ourselin, et al. Interactive medical image segmentation using deep learning with image-specific fine tuning. IEEE transactions on medical imaging, 37(7): 1562-1573, 2018. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20999", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Guotai Wang, Maria A Zuluaga, Wenqi Li, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sébastien Ourselin, et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 41(7):1559-1572, 2018. 3", + "[47] H Wang et al. Sam-med3d: towards general-purpose segmentation models for volumetric medical images. Preprint at https://arxiv.org/abs/2310.15161, 2024. 1, 3", + "[48] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163, 2022. 3", + "[49] Rhydian Windsor, Amir Jamaludin, Timor Kadir, and Andrew Zisserman. Vision-language modelling for radiological imaging and reports in the low data regime. arXiv preprint arXiv:2303.17644, 2023. 3", + "[50] Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology. arXiv preprint arXiv:2308.02463, 2023. 3", + "[51] Chaoyi Wu, Weixiong Lin, Xiaoman Zhang, Ya Zhang, Weidi Xie, and Yanfeng Wang. Pmc-llama: toward building open-source language models for medicine. Journal of the American Medical Informatics Association, page ocae045, 2024. 5", + "[52] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3", + "[53] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems, 36:59008-59033, 2023. 3", + "[54] Kihyun You, Jawook Gu, Jiyeon Ham, Beomhee Park, Jiho Kim, Eun K Hong, Woonhyuk Baek, and Byungseok Roh. Cxr-clip: Toward large scale chest x-ray language-image pre-training. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 101-111. Springer, 2023. 3", + "[55] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 1, 2, 3", + "[56] Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023. 2, 3", + "[57] Chunpeng Zhou, Kangjie Ning, Qianqian Shen, Sheng Zhou, Zhi Yu, and Haishuai Wang. Sam-sp: Self-prompting makes sam great again. arXiv preprint arXiv:2408.12364, 2024. 1, 2, 6", + "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + ], + "bbox": [ + 89, + 92, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3", + "bbox": [ + 545, + 92, + 903, + 119 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21000", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_model.json b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_model.json new file mode 100644 index 0000000000000000000000000000000000000000..38f7bca683db53e3183a0b4e8492d6580a402cb0 --- /dev/null +++ b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_model.json @@ -0,0 +1,1916 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.209, + 0.834, + 0.263 + ], + "angle": 0, + "content": "Aishik Konwer\\(^{1,2*}\\), Zhijian Yang\\(^{1}\\), Erhan Bas\\(^{1}\\), Cao Xiao\\(^{1}\\), Prateek Prasanna\\(^{3}\\), Parminder Bhatia\\(^{1}\\), Taha Kass-Hout\\(^{1}\\) \n\\(^{1}\\)GE Healthcare" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.264, + 0.732, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science, Stony Brook University" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.283, + 0.751, + 0.301 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Biomedical Informatics, Stony Brook University" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.303, + 0.753, + 0.318 + ], + "angle": 0, + "content": "akonwer@cs.stonybrook.edu, prateek.prasanna@stonybrook.edu" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.32, + 0.876, + 0.335 + ], + "angle": 0, + "content": "{zhijian.yang, erhan.bas, cao.xiao, paraminder.bhatia, taha.kass-hout}@gehealthcare.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.37, + 0.327, + 0.387 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.486, + 0.765 + ], + "angle": 0, + "content": "Foundational models such as the Segment Anything Model (SAM) are gaining traction in medical imaging segmentation, supporting multiple downstream tasks. However, such models are supervised in nature, still relying on large annotated datasets or prompts supplied by experts. Conventional techniques such as active learning to alleviate such limitations are limited in scope and still necessitate continuous human involvement and complex domain knowledge for label refinement or establishing reward ground truth. To address these challenges, we propose an enhanced Segment Anything Model (SAM) framework that utilizes annotation-efficient prompts generated in a fully unsupervised fashion, while still capturing essential semantic, location, and shape information through contrastive language-image pretraining and visual question answering. We adopt the direct preference optimization technique to design an optimal policy that enables the model to generate high-fidelity segmentations with simple ratings or rankings provided by a virtual annotator simulating the human annotation process. State-of-the-art performance of our framework in tasks such as lung segmentation, breast tumor segmentation, and organ segmentation across various modalities, including X-ray, ultrasound, and abdominal CT, justifies its effectiveness in low-annotation data scenarios." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.794, + 0.222, + 0.809 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.484, + 0.866 + ], + "angle": 0, + "content": "With advancements in medical image analysis, there is an increasing need for sophisticated methods [20, 21, 29] to leverage the vast availability of radiology datasets (such as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.908, + 0.629 + ], + "angle": 0, + "content": "X-ray, CT, and MRI) for accurate organ and tumor segmentation, as well as disease classification. The results of these tasks are crucial for physicians in designing effective treatment plans and surgical procedures. Several state-of-the-art deep learning-based foundational models, such as Vision-Language Models (VLMs), are now available for these purposes, relying on custom prompting to generate relevant predictions. However, they face two significant challenges: (1) despite being trained with only sparse prompts such as points or bounding boxes, these models still require human supervision for the prompt generation, leading to inefficiencies; and (2) many datasets lack comprehensive annotations, resulting in under-utilization during training of complex, data-hungry foundational architectures. Additionally, the high cost of human annotation efforts to create ground truth data can significantly escalate model development expenses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Popular architectures such as SAM and CLIP, have been extended to medical data, and led to innovations such as BiomedCLIP, Merlin, SAM-Med2D, and SAM-Med3D [6, 11, 47, 55]. Recently, numerous studies have focused on enhancing SAM by replacing geometric prompts and integrating semantic and spatial knowledge in an unsupervised manner through techniques such as (1) self-prompting, (2) class activation maps from CLIP, and (3) object localization models [19, 24, 57]. As shown in Fig. 1, self-prompting (b) does not require expert-provided prompts during inference. On the other hand, unsupervised prompts (c) in [19] are only used for training SAM to generate pseudo-labels for weakly supervised downstream tasks. They lack sufficient location and shape information, which could be provided by text-based prompts. Hence, the question arises: Can we come up with more refined and efficient prompts that can deliver stronger signals to foundational models without requiring intervention from annotators in both training and testing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.363, + 0.888 + ], + "angle": 0, + "content": "*Work done during internship at GE Healthcare" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20990" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.174, + 0.09, + 0.825, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.374, + 0.908, + 0.432 + ], + "angle": 0, + "content": "Figure 1. Overview of our model: (a) SAM and SAM-based approaches rely on expert prompts during both training and inference. (b) SAM-SP [57] introduces a self-prompting module, eliminating the need for expert prompts during inference. (c) MedCLIP-SAM [19] uses unsupervised semantic prompts to generate pseudo-labels through SAM. (d) Our approach not only combines semantic, location, and generic information via unsupervised prompts but also introduces a preference-based alignment module to reward or penalize the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.483, + 0.548 + ], + "angle": 0, + "content": "stages? State-of-the-art models also lack comprehensive integration of semantic, locational, and generic class information in their prompting strategies. In our approach, we leverage CLIP, VQA, and LLM models [1, 55, 56] to extract this discriminative information, improving segmentation performance in unsupervised settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To tackle the challenges posed by limited annotated datasets, several weakly supervised semantic segmentation algorithms [2, 19] have been proposed. Single-stage methods [2] leverage coarse image labels to perform segmentation in an end-to-end fashion, while many two-stage approaches [19] utilize foundational models to generate pseudo labels that can be used to train downstream segmentation architectures. Additionally, human-in-the-loop frameworks [23] have proven beneficial for tasks such as image generation, segmentation, and prognosis, involving annotators to refine segmentation labels or assess the plausibility of synthetic images. This feedback can either be incorporated back into the model or used as ground truth to train a reward function via reinforcement learning (RL), significantly reducing the need for annotated training data. However, these methods still rely on explicit reward modeling, which prevents end-to-end training. In some cases, annotators must provide complex ratings or refinements, making it difficult to train a reward function efficiently. We ask another question: Can we skip training a reward function and develop a straightforward, end-to-end pipeline that relies on simple annotator preferences? To address this, we draw inspiration from direct preference optimization tech" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.547 + ], + "angle": 0, + "content": "niques in the LLM preference alignment literature. We propose a preference-based reward model within our framework, using a novel loss function for end-to-end training. Our approach thus aligns with human preferences without the need for explicit reward modeling and is both easy to implement and train." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Our first goal is to develop refined and efficient prompts to fine-tune the SAM-Med2D encoder on diverse datasets, including 2D chest X-rays, breast ultrasound, and 3D abdominal CT scans. Initially, we input an image into the encoder alongside the BiomedCLIP [55] and MedViInT (VQA) [56] models. Corresponding texts, examples such as \"Chest X-ray\" and \"Describe the condition of the lungs and location of pathologies,\" are fed into BiomedCLIP and MedViInT, respectively. We gather generic information for the disease class from GPT-4, which is combined with the answers from the VQA model. Saliency maps generated by the CLIP model undergo dense CRF postprocessing to obtain bounding boxes. These bounding boxes and textual prompts are then inputted into the prompt encoder. Subsequently, the mask decoder receives both the image encoder embeddings and prompt embeddings to produce the segmentation maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "After fine-tuning the prompting module on a small proportion of annotated data, we introduce our second major idea: simulating human feedback through a AI preference alignment module. We generate multiple segmentation candidates for a given image by thresholding the SAM output probabilities at various levels. These candidates are rated" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20991" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "on a scale of 1 to 4 based on the overlap between the candidates and ground truth, mimicking the evaluation process of a human annotator. This approach does not require explicit access to ground truth data for training, thus making our task a form of semi-supervised segmentation. We propose a DPO-inspired [38] loss function that encourages the model to prioritize desirable segmentation outputs by rewarding higher-rated candidates and penalizing lower-rated ones. The model is thus trained on the remaining portion of the dataset, without annotations, to perform relevant medical image segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.259, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Overall our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.275, + 0.483, + 0.365 + ], + "angle": 0, + "content": "- We propose refined and efficient unsupervised prompting strategies that deliver comprehensive information about location, semantics, and general disease/organ characteristics to our SAM-Med2D-based framework. Such an approach enhances segmentation performance while reducing reliance on human input for geometric prompts." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.366, + 0.483, + 0.486 + ], + "angle": 0, + "content": "- We introduce a novel DPO-inspired loss function that facilitates semi-supervised model improvement using simulated human-in-the-loop feedback, eliminating the need for an explicit reward function. The framework generates multiple segmentation maps and rates them based on segmentation overlap, mimicking the evaluation of a human annotator. The model learns to distinguish between favorable and unfavorable candidates effectively." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.275, + 0.483, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.503, + 0.233, + 0.518 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.463, + 0.544 + ], + "angle": 0, + "content": "2.1. Vision-language models for medical domain" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.794 + ], + "angle": 0, + "content": "CLIP [37] has gained much popularity in medical image analysis. [14] fine-tuned CLIP on various PubMed articles to create PubMedCLIP. MedCLIP [48] leverages unpaired image and text datasets along with a semantic-matching loss to align visual and textual information. Windsor et al. [49] employ unimodal self-supervision, local-global contrastive losses, and data augmentation to enhance zero-shot performance and retrieval task efficiency in low-resource data settings. Some modality-specific CLIP variants [15, 54] have been developed for Chest X-ray and Mammograms due to the readily available image-text data in these areas. However, BiomedCLIP [55] stands out as the most recent model, excelling in scalability and performance across diverse multi-organ cross-modal retrieval tasks. Therefore, we have integrated BiomedCLIP into the CLIP-driven bounding box generation module of our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Building on the success of large language models (LLMs) such as LLaMa [44] and GPT [35], researchers have explored merging visual features with textual representations using techniques such as cross-attention, Qformer, instruction tuning, and projection layers. This effort has resulted in vision-language models (VLMs) [4, 26, 30, 58], including Flamingo, BLIP-2, LLaVA, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "MiniGPT, which were further adapted for the medical domain through pretraining on multimodal medical datasets. Med-Flamingo [33] is the first medical visual question-answering (VQA) model with few-shot generation capabilities. RadFM [50] serves as a foundational model that also incorporates 3D volume data. Pretraining on extensive datasets, PMC-15M and PMC-VQA [55, 56], has facilitated the development of LLaVA-Med [25] and MedViInT [56], respectively. Our segmentation pipeline leverages MedViInT's capabilities for enhanced localization and shape-based answer generation related to tumors, organs, and disease manifestations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.837, + 0.299 + ], + "angle": 0, + "content": "2.2. SAM for multimodal biomedical data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.55 + ], + "angle": 0, + "content": "MedSAM and SAM-Med2D [11, 31] focused on adapting SAM [18] for medical applications by fine-tuning it on 2D medical datasets, while SAM-Med3D [47] introduced alternative modules to accommodate 3D volumes. Efficient approaches, for example, AutoSAM [41] utilize trainable prompt encoders, whereas FastSAM3D [42] employs flash attention to accelerate inference. MedLSAM [24] proposed a localization framework to generate 3D bounding boxes as prompt input. However, most of these methods require ground truth data (bounding boxes or points) for training the prompt encoder, whereas we propose an unsupervised route for the same. We leverage ad-hoc VLM models such as CLIP, VQA, and GPT-4 together to propagate comprehensive information—encompassing semantics, location, and generic disease/organ information—that significantly enhances segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.771, + 0.576 + ], + "angle": 0, + "content": "2.3. Human-in-the-loop feedback" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Human-in-the-loop training paradigm in medical imaging has primarily focused on two areas: active learning [34], which identifies optimal data points for labeling to maximize performance, and interactive feedback [8] on model predictions to calibrate parameters. Examples include UI-Net, DeepIGeoS, and BIFseg [5, 45, 46], which utilize expert-provided scribbles or bounding boxes alongside geodesic transforms and graph-cut techniques to refine segmentation labels. Rao et al. [39] introduced IMIL, the first framework that assigns clinicians to guide data augmentations on mispredictions, emphasizing disease-relevant regions while eliminating irrelevant ones. Recently, the success of human feedback in instruction-tuning, and aligning large language model outputs through RL objective [52, 53], has led researchers to utilize human preferences to evaluate synthetic natural and histopathology images, thereby improving image-to-text models [23, 43] by training a reward function. Training a reward function requires the curation of dedicated human preference datasets and also prohibits the framework from operating in an end-to-end manner. Additionally, it often demands advanced do" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20992" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.202, + 0.104, + 0.8, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.492, + 0.908, + 0.565 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed framework for semi-supervised segmentation: Unsupervised geometric and text prompts, obtained from pretrained BiomedCLIP, MedVInT, and GPT-4 models, are fed into the prompt encoder for finetuning the framework on a small fraction of annotated data. In the next stage, we simulate a virtual annotation process that assigns ratings to the generated segmentation candidates, which are used to fine-tune the decoder. This stage handles unannotated data, as the model does not rely on ground truth for direct supervision but only for rating while simulating a human annotator's feedback." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.589, + 0.483, + 0.742 + ], + "angle": 0, + "content": "main knowledge from annotators, which can be costly. Our approach addresses this challenge by using direct preference optimization [38] to fulfill the RL objective, enabling the framework to serve as its own reward model. The model aligns generated proposals with appropriate preferences or ratings, leading to performance improvements, even with limited annotated data. These preferences are generated through a simulation mimicking human-in-the-loop feedback. Crucially, our method is easier to train and implement than traditional reward function-based RL pipelines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.226, + 0.772 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Overview. Given an input image, our aim is to generate prompts in the form of bounding boxes and textual queries, which will be used to produce segmentation masks. The image is processed by three pretrained components: the SAM-Med2D encoder, BiomedCLIP, and MedVInT. Concurrently, GPT-4 curates generic information about the concerned disease or organ. This information, combined with the BiomedCLIP-generated box prompt and the MedVInT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.907, + 0.756 + ], + "angle": 0, + "content": "provided textual answer prompt, serves as input for the prompt encoder. The prompt generation process is detailed in Sec. 3.1. Next, both the encoded image and prompt embeddings are fed into the decoder to create segmentation maps. Following this initial fine-tuning on annotated data, we introduce a novel loss to mimic human-in-the-loop feedback. By thresholding the output probabilities, our model produces four different segmentation candidates, and we incorporate insights about the quality of these candidates into the framework through the proposed loss function. This part, trained with unannotated data, is elaborated in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.838, + 0.788 + ], + "angle": 0, + "content": "3.1. Visual and textual prompt generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Visual prompt. To generate visual prompts, we first input the image along with corresponding text prompts (e.g. \"chest x-ray\", \"benign breast tumor\", \"left kidney\", \"liver\", etc.) into BiomedCLIP, a foundational model pretrained on millions of medical image-caption pairs. Next, we leverage gScoreCAM [10] to create a saliency map highlighting targeted regions (organs or tumors) in the image corre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20993" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "sponding to the supplied text. These saliency maps are then post-processed with a conditional random field filter [22] to produce coarse segmentation masks. We apply an area constraint, retaining the largest connected component or up to two, depending on the dataset (the lung dataset may yield two components). Finally, we identify the bounding box coordinates within these closed components for the box prompt, while randomly sampling several points from the designated area for the point prompt." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.229, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Textual prompt. We extract visual embeddings from a given image by processing it through a vision encoder derived from the PMC-CLIP architecture [28]. This encoder features a pretrained ResNet50 [16] and a trainable projection layer constructed with stacks of transformer decoder blocks. Next, we create a prompt template that incorporates the question for the image, formatted as \"Question: \\(\\{\\}\\), Answer is:\". This prompt is then passed through a tokenization embedding layer initialized with the weights of PMC-LLaMA [51] to generate the text embedding. Finally, we concatenate the visual and text embeddings to form the input space for a pre-initialized multimodal transformer decoder. The answer generated from this VQA setup delivers essential information regarding the shape and location of anatomical structures and pathologies. Sample questions included in the prompt are: \"What is the shape of the liver and where is it located?\", and \"What is the shape of breast tumor and where is it located?\" Additionally, we prompt GPT-4 with the relevant organ or disease label to obtain a generic description of its characteristics. Finally, both types of textual prompts are concatenated and provided as input to the prompt encoder, described in the following paragraph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Prompt encoder and mask decoder. The prompt encoder, same as the one in SAM, supports three types of prompts: point, box, and text. Point and box prompts are represented by their positional encodings—specific coordinates for points, and the top-left and bottom-right corners for boxes—along with learnable feature representations. Text prompts are processed through a pretrained BiomedCLIP encoder to generate corresponding text embeddings. All prompt embeddings are then projected into 256-dimensional vectors. The feature map from the first iteration of the model, is downsampled through multiple convolutional layers followed by GELU activation to match the 256-channel dimension. Finally, these downsampled masks are combined element-wise with visual encoder embeddings. The mask decoder takes both the prompt embeddings and the visual embeddings to produce a segmentation map. The architecture is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.482, + 0.863 + ], + "angle": 0, + "content": "3.2. Segmentation proposal generation and Preference Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Proposal generation. After fine-tuning our framework, comprising the prompt encoder and mask decoder, using" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.377 + ], + "angle": 0, + "content": "ground truth for a fraction (10%) of the dataset, we propose integrating AI-based preferences for the next training episode. First, we generate multiple segmentation proposals by sampling different thresholds from the pixel probability scores output by the model. We then simulate an expert annotator to evaluate the quality of these proposals, assigning ratings based on the overlap between each generated proposal and the corresponding ground truth. Although ground truth is not explicitly used in this training phase, it plays a passive role by enabling the simulation of preferences in the absence of an actual human annotator in our experiments. We also experiment with two alternative rating mechanisms: one inspired by SAM, which backpropagates loss solely for the best candidate, and another that ranks all proposals rather than merely rating them. After obtaining the ratings or rankings, we fine-tune our decoder using the direct preference optimization technique to better align the segmentation outputs with the preferences of the virtual annotator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.905, + 0.5 + ], + "angle": 0, + "content": "RLHF Preliminaries. Language models typically utilize a reward function to align their generated responses with user preferences. This preference distribution is often modeled using the Bradley-Terry [7] model, which operates on pairwise comparisons. For a given prompt \\( X \\), when the language model produces two responses, one more favorable, \\( Y_{m} \\), and the other less favorable, \\( Y_{l} \\), the distribution can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.51, + 0.905, + 0.545 + ], + "angle": 0, + "content": "\\[\nP \\left(Y _ {m} > Y _ {l} \\mid X\\right) = \\frac {e ^ {r ^ {*} \\left(X , Y _ {m}\\right)}}{e ^ {r ^ {*} \\left(X , Y _ {m}\\right)} + e ^ {r ^ {*} \\left(X , Y _ {l}\\right)}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.905, + 0.599 + ], + "angle": 0, + "content": "Parameters of the reward model can be estimated via Maximum Likelihood Estimation (MLE) and the goal is to minimize the below loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.611, + 0.905, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {R}} = - \\mathbb {E} _ {(X, Y _ {m}, Y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r _ {\\phi} (X, Y _ {m}) - r _ {\\phi} (X, Y _ {l})\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.905, + 0.672 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the logistic function and \\(\\mathcal{D}\\) denotes the dataset of preferences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Reinforcement Learning from Human Feedback (RLHF) is a widely used method that involves training a reward model based on user ratings. The primary goal is to discover an optimal policy with parameter \\(\\pi_{\\theta}\\) that maximizes this reward function while incorporating a KL divergence term to ensure the model's outputs do not deviate significantly from the original policy \\(\\pi_{ref}\\). The optimization problem can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.806, + 0.905, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {X \\sim \\mathcal {D}, Y \\sim \\pi_ {\\theta} (Y | X)} \\left[ r _ {\\phi} (X, Y) \\right] \\tag {3} \\\\ - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ \\pi_ {\\theta} (Y \\mid X) | | \\pi_ {r e f} (Y \\mid X) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Direct Preference optimization. Due to the discrete nature of language generation, this objective is not differentiable, which usually necessitates optimization through" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20994" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "reinforcement learning. The primary language model then leverages this framework to align its outputs with user ratings, ultimately generating high-scoring responses. In contrast, Direct Preference Optimization (DPO) is a more recent and streamlined approach that eliminates the need for a separate reward model. Instead, it fine-tunes the main language model directly using user preferences. This is achievable because DPO focuses on optimizing the policy itself rather than the reward function. The maximum likelihood objective for a parameterized policy \\(\\pi_{\\theta}\\) in such a case can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.277, + 0.484, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\pi_ {\\theta}; \\pi_ {r e f}\\right) = - \\mathbb {E} _ {\\left(X, Y _ {m}, Y _ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(Y _ {m} \\mid X\\right)}{\\pi_ {r e f} \\left(Y _ {m} \\mid X\\right)} \\right. \\right. \\tag {4} \\\\ \\left. \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(Y _ {l} \\mid X\\right)}{\\pi_ {r e f} \\left(Y _ {l} \\mid X\\right)}\\right) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.595 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the weight applied to reward or penalize the responses. In DPO, models are fine-tuned using pairs of outputs, explicitly comparing a preferred response \\(Y_{m}\\) with less preferred ones \\(Y_{l}\\). We adapt this concept for our application in the imaging domain, where multiple segmentation candidates \\(Y_{1},\\ldots ,Y_{4}\\) are generated for the same {image, prompt} pair \\(I\\). With ratings available for each candidate, we assign weights \\(\\beta_{1},\\beta_{2}\\) to these candidates (from best to worst) to reward the higher-quality segmentations \\(Y_{1},Y_{2}\\) and penalize the less desirable ones \\(Y_{3},Y_{4}\\). Consequently, Equation 4 is modified to create Equation 5. This approach allows us to incorporate real-world annotator preferences into our method through a straightforward yet effective loss function. The parameters of the initially fine-tuned model from Sec. 3.1 are denoted as \\(\\pi_{\\mathrm{fine}}\\). Our objective is to find the optimal parameters \\(\\pi_{\\psi}\\) for the final architecture without deviating much from \\(\\pi_{\\mathrm{fine}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.605, + 0.489, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} (\\pi_ {\\psi}; \\pi_ {f i n e}) = - \\mathbb {E} _ {(I, Y _ {1}, Y _ {2}, Y _ {3}, Y _ {4}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta_ {1} \\log \\frac {\\pi_ {\\psi} (Y _ {1} \\mid I)}{\\pi_ {f i n e} (Y _ {1} \\mid I)} \\right. \\right. \\\\ \\left. + \\beta_ {2} \\log \\frac {\\pi_ {\\psi} (Y _ {2} \\mid I)}{\\pi_ {f i n e} (Y _ {2} \\mid I)} - \\beta_ {2} \\log \\frac {\\pi_ {\\psi} (Y _ {3} \\mid I)}{\\pi_ {f i n e} (Y _ {3} \\mid I)} - \\beta_ {1} \\log \\frac {\\pi_ {\\psi} (Y _ {4} \\mid I)}{\\pi_ {f i n e} (Y _ {4} \\mid I)}\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.395, + 0.697 + ], + "angle": 0, + "content": "4. Experimental Design and Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.9 + ], + "angle": 0, + "content": "Datasets. We evaluated our framework for semi-supervised segmentation using three public datasets covering lung, breast tumor, and abdominal organ segmentation across X-ray, ultrasound, and CT. For Ultrasound Breast Tumor Segmentation, we used 810 images from BUSI [3] and UDIAT [9], with 600 for training and 210 for testing. For Chest X-ray Lung Segmentation, we used 27,132 images from COVID-QU-Ex [12], spanning normal, lung opacity, viral pneumonia, and COVID-19 cases. For Abdominal CT Organ Segmentation, 200 annotated CT scans from the AMOS-CT training set were used for training. Full dataset details, including train-test splits, are in the supplementary Sec. 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.53 + ], + "angle": 0, + "content": "Implementation Details. Our method is implemented in PyTorch [36] on an EC2 instance (with 64 GB NVIDIA T4 Tensor Core GPUs). For feature extraction, we utilize the SAM-Med2D pretrained encoder. First we use annotations for only \\(10\\%\\) of the training dataset, during which all components (visual encoder, prompt encoder, and mask decoder) are fine-tuned. In this step, only the unsupervised prompting strategy is employed. Bounding box and point prompts are used together for all experiments. The remaining portion of the dataset was used in an unannotated form to train the DPO-driven alignment strategy. We optimize the model using the Adam optimizer [13], training for 15 epochs for prompt module fine-tuning and 30 epochs for alignment. In both stages, the initial learning rate is set to 1e-4 and is halved every 10 epochs. All images are resized to a resolution of \\(256 \\times 256\\) using the same resizing strategy as SAM-Med2D. Similar to our framework, we have also fine-tuned SOTA SAM-based methods on medical imaging datasets using LoRA by adding low-rank matrices to transformer layers. This enables efficient adaptation with fewer parameters, enhancing domain-specific performance while preserving pretrained features. The loss function used during the initial fine-tuning is a 20:1 weighted combination of focal loss [27] and Dice loss [32]. Details on the preference alignment module and proposal generation are provided in the supplementary material Sec. 6. Dice Similarity Coefficient, Intersection over Union (IoU), and Surface Dice Similarity (SDC) scores are used to evaluate segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.805, + 0.56 + ], + "angle": 0, + "content": "4.1. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative results. Fig. 3 compares the performance of our framework with relevant methods (U-Net [40], nnU-Net [17], SAM [18], SAM-Med2D [11], Self-prompt). The self-prompt method is designed as a variant of [57], excluding the knowledge distillation module. We also design a prompt-only baseline of our framework which is trained on different splits of fully annotated data. Our framework is initially trained with only the prompting module using \\(10\\%\\) of data (annotated). As a result, both the prompt-only and final versions exhibit identical performance on this \\(10\\%\\) subset. For the final model, the additional data used for training are considered unannotated, as training the alignment mechanism does not use supervision from ground truth. SOTA methods directly use ground truth for the entire data subset. Nevertheless, our method consistently outperforms them in limited data settings \\((10 - 50\\%)\\) due to the preference alignment mechanism. Our architecture thus reduces the reliance on large datasets, highlighting its lower annotation requirements, which makes it highly effective in low-data regimes. On the Chest-Xray dataset, for instance, with just \\(20\\%\\) of the data, our method achieves a Dice score of 78.87, compared to 58.66 (U-Net), 60.97 (nnU-Net), 61.64 (SAM)," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20995" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.359, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.089, + 0.647, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.089, + 0.906, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.226, + 0.908, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Quantitative comparison with SOTA. Dice score (for Chest Xray, Breast USD) and mean Dice score (for AMOS CT) have been shown to measure the model segmentation performance on different proportions of training data \\((10\\%, 20\\%, 50\\%,\\) and \\(100\\%)\\)." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.269, + 0.868, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.442, + 0.908, + 0.471 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons were made between the segmentation results of nnUnet, SAM-Med2D, and our framework on 2D datasets. BiomedCLIP-based saliency maps are also depicted. Experiments were conducted in \\(50\\%\\) data settings." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.496, + 0.485, + 0.89 + ], + "angle": 0, + "content": "67.81 (SAM-Med2D), and 68.41 (Self-prompt). Both U-Net and nnU-Net are known to be data-hungry models, struggling with smaller data subsets. Our usage of visual and textual prompts, derived through BiomedCLIP and VLM models, offers superior signals to those used in SAM and SAM-Med2D. Unlike these SAM variants, Self-Prompt SAM does not require expert-supplied prompts during inference; it instead generates prompts from the output masks in each iteration. This method offers a slight improvement \\((+1\\%)\\) over SAM-Med2D. Our method shows more stable performance gains across different data subsets, while other SOTA methods exhibit steeper improvements. In the \\(50\\%\\) data setting, our semi-supervised framework achieves a Dice score of 89.68, compared to 91.42 for the supervised prompt-only version. Similar result trends are observed in the breast ultrasound dataset. Notably, the performance jump of our method from \\(20\\%\\) to \\(50\\%\\) data is much more pronounced than from \\(10\\%\\) to \\(20\\%\\). This can be due to the nature of the dataset, as the model requires more data to achieve precise and accurate segmentation of breast tumors. We also evaluated our method on a 3D abdominal organ segmentation dataset, including organs such as the liver, kidneys, spleen, pancreas, aorta, bladder, etc (see Fig. 3). Our method outperforms the SOTA across all data proportions except for the full dataset. With \\(20\\%\\) of the data, our method achieves a mean Dice score of 77.69, surpassing U" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.496, + 0.909, + 0.558 + ], + "angle": 0, + "content": "Net (59.35), nnU-Net (65.21), SAM (64.93), SAM-Med2D (66.57), SAMMed3D (72.54), and Self-prompt (71.83). At the \\(50\\%\\) data setting, it reaches a Dice score of 85.70, comparable to the 86.36 achieved by the prompt-only version." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.571, + 0.887, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.72, + 0.908, + 0.748 + ], + "angle": 0, + "content": "Figure 5. Segmentation maps of different anatomical structures (liver and spleen) for SAM-Med3D and our method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative results. We present segmentation maps for 3 datasets in Fig. 4 and Fig. 5, generated by our model trained at \\(50\\%\\) data settings. The saliency maps could correctly highlight the target regions and proved to be an excellent source of supervision. It can be noted from Fig. 4 that our segmentation quality around the boundaries of tumor or lung is much superior compared to both nnUnet and SAM-Med2D. In Fig. 5, SAM-Med3D tended to under-segment the edges of both the abdominal organs, the liver," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20996" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.178, + 0.088, + 0.825, + 0.201 + ], + "angle": 0, + "content": "
SupervisionMethodsChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedOurs74.3078.8764.3575.8877.6978.34
20%- Alignment75.0279.1367.5181.3879.2080.66
10%- Alignment68.4375.6061.4473.6274.7776.06
- Alignment - VQA66.9073.3557.6070.5374.0875.41
- Alignment - VQA - GPT463.1672.7654.2668.8973.1674.89
- Alignment - VQA - CAM50.3857.0245.1659.0569.9771.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.274, + 0.21, + 0.723, + 0.224 + ], + "angle": 0, + "content": "Table 1. Ablation results demonstrating effectiveness of major components." + }, + { + "type": "table", + "bbox": [ + 0.177, + 0.237, + 0.825, + 0.34 + ], + "angle": 0, + "content": "
SupervisionPreference AlignmentChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedLoss for best candidate70.3777.0961.7673.8175.0176.36
Rating73.9978.4163.9775.5277.2377.98
Ranking74.3078.8764.3575.8877.6978.34
10% + 20% unannotatedRanking79.7485.1573.5684.2380.5481.95
10% + 40% unannotatedRanking88.9689.6885.9288.1584.3085.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.349, + 0.82, + 0.364 + ], + "angle": 0, + "content": "Table 2. Ablation results on different proportions of training data for 3 types of preference scoring strategies." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.39, + 0.481, + 0.407 + ], + "angle": 0, + "content": "and spleen. Additional results are in supplementary Sec. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.253, + 0.435 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Effectiveness of major components. We conduct several ablation experiments (see Tab. 1) to evaluate the contribution of each module in our architecture. For simplicity, we focus on the Chest-Xray dataset for analysis. Before incorporating the DPO-driven alignment strategy, all baselines were trained with \\(10\\%\\) labeled data. First, we developed a naive baseline (last row), which relies solely on textual answers generated from GPT-4. As expected, it performed poorly, achieving a 57.02 Dice score. A second baseline (second-last row) was designed to use only visual prompts from BiomedCLIP, which provides semantic information. This method significantly benefited from the coarse segmentation masks derived from CLIP saliency maps, improving the Dice score to 72.76. Next, we combined both textual and visual prompts to form a third baseline (-Alignment-VQA), which resulted in a slight improvement \\((+0.59\\%)\\) over the CLIP-only baseline. Finally, we integrated VQA into the prompting strategy to obtain answers related to the shape and location of the target regions, completing our fully empowered unsupervised prompting strategy (-Alignment). This model achieved a Dice score of 75.60 with \\(10\\%\\) data and improved to 79.13 with \\(20\\%\\) data. In comparison, with the preference alignment mechanism, fine-tuning the prompting module with \\(10\\%\\) annotated data and training the alignment module on an additional \\(10\\%\\) unannotated data achieved a Dice score of 78.87. This is on par with the fully supervised prompt-only model, underscoring the effectiveness of our alignment module." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Preference Scoring strategies. We conducted several ablation studies (detailed in Tab. 2) to evaluate the effec" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.587 + ], + "angle": 0, + "content": "tiveness of the preference-scoring strategy. One baseline approach involved backpropagating the loss based only on the best candidate, while another compared ranking the candidates rather than simply rating them. The results are presented for different proportions of unannotated data, on top of the foundational fine-tuning of both the prompt encoder and the decoder using \\(10\\%\\) annotated data. At \\(10\\%\\) unannotated data, both the ranking and rating approaches performed similarly (mean Dice scores of 77.69 and 77.23 for AMOS CT, respectively), significantly outperforming the best-candidate-only method (75.01). Additionally, we observed that as the proportion of unannotated data increased, the model's performance improved substantially." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.907, + 0.662 + ], + "angle": 0, + "content": "Robustness to noisy rating. We experimented with introducing noise into the rating mechanism by flipping the ratings of closely ranked candidates to enhance the framework's robustness. Results are in supplementary material Sec. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduce a novel training strategy to enhance SAM representations for semi-supervised medical image segmentation. We extract integrated semantic, location, and shape information from pretrained vision-language models in an unsupervised manner. This information is used as refined prompts for our model. We also implement an optimal policy, inspired by direct preference optimization in language models. This enables human-in-the-loop feedback simulation within a streamlined training framework, without the need for a separate reward function or extensive knowledge from annotators. These modules ensure that our framework outperforms state-of-the-art methods across datasets spanning multiple modalities in low-annotation data scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20997" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.237 + ], + "angle": 0, + "content": "We sincerely thank everyone from the GE Healthcare Science & Technology Organization for their support and contributions to the project. The primary model development and evaluations were conducted during a summer internship at GEHC. After the internship, Konwer and Prasanna were supported by the following grants during the paper writing and rebuttal process: NIH 1R01CA297843-01, NIH 3R21CA258493-02S1, and 1R21CA258493-01A1." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.252, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.347, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[2] Peri Akiva and Kristin Dana. Towards single stage weakly supervised semantic segmentation. arXiv preprint arXiv:2106.10309, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.391, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[3] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.434, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.485, + 0.601 + ], + "angle": 0, + "content": "[5] Mario Amrehn, Sven Gaube, Mathias Unberath, Frank Schebesch, Tim Horz, Maddalena Strumia, Stefan Steidl, Markus Kowarschik, and Andreas Maier. Ui-net: Interactive artificial neural networks for iterative image segmentation based on a user model. arXiv preprint arXiv:1709.03450, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[6] Louis Blankemeier, Joseph Paul Cohen, Ashwin Kumar, Dave Van Veen, Syed Jamal Safdar Gardezi, Magdalini Paschali, Zhihong Chen, Jean-Benoit Delbrouck, Eduardo Reis, Cesar Truyts, et al. Merlin: A vision language foundation model for 3d computed tomography. arXiv preprint arXiv:2406.06512, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[7] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[8] Samuel Budd, Emma C Robinson, and Bernhard Kainz. A survey on active learning and human-in-the-loop deep learning for medical image analysis. Medical image analysis, 71: 102062, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[9] Michal Byra, Piotr Jarosik, Aleksandra Szubert, Michael Galperin, Haydee Ojeda-Fournier, Linda Olson, Mary O'Boyle, Christopher Comstock, and Michael Andre. Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control, 61:102027, 2020. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Peijie Chen, Qi Li, Saad Biaz, Trung Bui, and Anh Nguyen. gscorecam: What objects is clip looking at? In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.277, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "of the Asian Conference on Computer Vision (ACCV), pages 1959-1975, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[11] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[12] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and coid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[13] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.356 + ], + "angle": 0, + "content": "[14] Sedigheh Eslami, Christoph Meinel, and Gerard De Melo. Pubmedclip: How much does clip benefit visual question answering in the medical domain? In Findings of the Association for Computational Linguistics: EACL 2023, pages 1181-1193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[15] Shantanu Ghosh, Clare B Poynton, Shyam Visweswaran, and Kayhan Batmanghelich. Mammo-clip: A vision language foundation model to enhance data efficiency and robustness in mammography. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 632–642. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[17] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[19] Taha Koleilat, Hojat Asgariandehkordi, Hassan Rivaz, and Yiming Xiao. Medclip-sam: Bridging text and image towards universal medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 643-653. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[20] Aishik Konwer, Xuan Xu, Joseph Bae, Chao Chen, and Pratek Prasanna. Temporal context matters: Enhancing single image prediction with disease progression representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18824-18835, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[21] Aishik Konwer, Xiaoling Hu, Joseph Bae, Xuan Xu, Chao Chen, and Prateek Prasanna. Enhancing modality-agnostic representations via meta-learning for brain tumor segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21415–21425, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Philipp Krahenbuhl and Vladlen Koltun. Parameter learning and convergent inference for dense random fields. In Inter-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20998" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "national conference on machine learning, pages 513-521. \nPMLR, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[23] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[24] Wenhui Lei, Xu Wei, Xiaofan Zhang, Kang Li, and Shaoting Zhang. Medlsam: Localize and segment anything model for 3d medical images. arXiv preprint arXiv:2306.14752, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[25] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[27] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[28] Weixiong Lin, Ziheng Zhao, Xiaoman Zhang, Chaoyi Wu, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-clip: Contrastive language-image pre-training using biomedical documents. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 525–536. Springer, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[29] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[31] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[32] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[33] Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Eduardo Pontes Reis, and Pranav Rajpurkar. Med-flamingo: a multimodal medical few-shot learner. In Machine Learning for Health (ML4H), pages 353-367. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Vishwesh Nath, Dong Yang, Bennett A. Landman, Daguang Xu, and Holger R. Roth. Diminishing uncertainty within the training pool: Active learning for medical image segmentation. IEEE Transactions on Medical Imaging, 40(10): 2534-2547, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[38] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[39] Adrit Rao, Andrea Fisher, Ken Chang, John Christopher Panagides, Katherine McNamara, Joon-Young Lee, and Oliver Aalami. Imil: Interactive medical image learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5241–5250, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.606 + ], + "angle": 0, + "content": "[41] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[42] Yiqing Shen, Jingxing Li, Xinyuan Shao, Blanca Inigo Romillo, Ankush Jindal, David Dreizin, and Mathias Unberath. Fastsam3d: An efficient segment anything model for 3d volumetric medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 542-552. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[43] Shenghuan Sun, Greg Goldgof, Atul Butte, and Ahmed M Alaa. Aligning synthetic medical images with clinical knowledge using human feedback. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[44] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[45] Guotai Wang, Wenqi Li, Maria A Zuluaga, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sebastien Ourselin, et al. Interactive medical image segmentation using deep learning with image-specific fine tuning. IEEE transactions on medical imaging, 37(7): 1562-1573, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20999" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[46] Guotai Wang, Maria A Zuluaga, Wenqi Li, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sébastien Ourselin, et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 41(7):1559-1572, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.178, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[47] H Wang et al. Sam-med3d: towards general-purpose segmentation models for volumetric medical images. Preprint at https://arxiv.org/abs/2310.15161, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.221, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[48] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[49] Rhydian Windsor, Amir Jamaludin, Timor Kadir, and Andrew Zisserman. Vision-language modelling for radiological imaging and reports in the low data regime. arXiv preprint arXiv:2303.17644, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[50] Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology. arXiv preprint arXiv:2308.02463, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.364, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[51] Chaoyi Wu, Weixiong Lin, Xiaoman Zhang, Ya Zhang, Weidi Xie, and Yanfeng Wang. Pmc-llama: toward building open-source language models for medicine. Journal of the American Medical Informatics Association, page ocae045, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.435, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[52] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.506, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[53] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems, 36:59008-59033, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.591, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[54] Kihyun You, Jawook Gu, Jiyeon Ham, Beomhee Park, Jiho Kim, Eun K Hong, Woonhyuk Baek, and Byungseok Roh. Cxr-clip: Toward large scale chest x-ray language-image pre-training. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 101-111. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.676, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[55] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[56] Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[57] Chunpeng Zhou, Kangjie Ning, Qianqian Shen, Sheng Zhou, Zhi Yu, and Haishuai Wang. Sam-sp: Self-prompting makes sam great again. arXiv preprint arXiv:2408.12364, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.546, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21000" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_origin.pdf b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9199477adce7a64567f633a494dfd54cdc3e27e5 --- /dev/null +++ b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/246f03d7-fa39-4c5a-a7fe-2aa5a7ce4b40_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:899a756aac3f20ed3d1bbc0884c9ee88ad6cd7ba5c8a6426aec46737ee1dab97 +size 1073460 diff --git a/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/full.md b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6180d767d3e124a4feb0c5323f6621ab3d457d03 --- /dev/null +++ b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/full.md @@ -0,0 +1,250 @@ +# Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation + +Aishik Konwer $^{1,2*}$ , Zhijian Yang $^{1}$ , Erhan Bas $^{1}$ , Cao Xiao $^{1}$ , Prateek Prasanna $^{3}$ , Parminder Bhatia $^{1}$ , Taha Kass-Hout $^{1}$ $^{1}$ GE Healthcare + +$^{2}$ Department of Computer Science, Stony Brook University + +$^{3}$ Department of Biomedical Informatics, Stony Brook University + +akonwer@cs.stonybrook.edu, prateek.prasanna@stonybrook.edu + +{zhijian.yang, erhan.bas, cao.xiao, paraminder.bhatia, taha.kass-hout}@gehealthcare.com + +# Abstract + +Foundational models such as the Segment Anything Model (SAM) are gaining traction in medical imaging segmentation, supporting multiple downstream tasks. However, such models are supervised in nature, still relying on large annotated datasets or prompts supplied by experts. Conventional techniques such as active learning to alleviate such limitations are limited in scope and still necessitate continuous human involvement and complex domain knowledge for label refinement or establishing reward ground truth. To address these challenges, we propose an enhanced Segment Anything Model (SAM) framework that utilizes annotation-efficient prompts generated in a fully unsupervised fashion, while still capturing essential semantic, location, and shape information through contrastive language-image pretraining and visual question answering. We adopt the direct preference optimization technique to design an optimal policy that enables the model to generate high-fidelity segmentations with simple ratings or rankings provided by a virtual annotator simulating the human annotation process. State-of-the-art performance of our framework in tasks such as lung segmentation, breast tumor segmentation, and organ segmentation across various modalities, including X-ray, ultrasound, and abdominal CT, justifies its effectiveness in low-annotation data scenarios. + +# 1. Introduction + +With advancements in medical image analysis, there is an increasing need for sophisticated methods [20, 21, 29] to leverage the vast availability of radiology datasets (such as + +X-ray, CT, and MRI) for accurate organ and tumor segmentation, as well as disease classification. The results of these tasks are crucial for physicians in designing effective treatment plans and surgical procedures. Several state-of-the-art deep learning-based foundational models, such as Vision-Language Models (VLMs), are now available for these purposes, relying on custom prompting to generate relevant predictions. However, they face two significant challenges: (1) despite being trained with only sparse prompts such as points or bounding boxes, these models still require human supervision for the prompt generation, leading to inefficiencies; and (2) many datasets lack comprehensive annotations, resulting in under-utilization during training of complex, data-hungry foundational architectures. Additionally, the high cost of human annotation efforts to create ground truth data can significantly escalate model development expenses. + +Popular architectures such as SAM and CLIP, have been extended to medical data, and led to innovations such as BiomedCLIP, Merlin, SAM-Med2D, and SAM-Med3D [6, 11, 47, 55]. Recently, numerous studies have focused on enhancing SAM by replacing geometric prompts and integrating semantic and spatial knowledge in an unsupervised manner through techniques such as (1) self-prompting, (2) class activation maps from CLIP, and (3) object localization models [19, 24, 57]. As shown in Fig. 1, self-prompting (b) does not require expert-provided prompts during inference. On the other hand, unsupervised prompts (c) in [19] are only used for training SAM to generate pseudo-labels for weakly supervised downstream tasks. They lack sufficient location and shape information, which could be provided by text-based prompts. Hence, the question arises: Can we come up with more refined and efficient prompts that can deliver stronger signals to foundational models without requiring intervention from annotators in both training and testing + +![](images/bb07502a342acdf16cb10ed1cf4f3b814277888c18e400f1078248d11ecce69a.jpg) +Figure 1. Overview of our model: (a) SAM and SAM-based approaches rely on expert prompts during both training and inference. (b) SAM-SP [57] introduces a self-prompting module, eliminating the need for expert prompts during inference. (c) MedCLIP-SAM [19] uses unsupervised semantic prompts to generate pseudo-labels through SAM. (d) Our approach not only combines semantic, location, and generic information via unsupervised prompts but also introduces a preference-based alignment module to reward or penalize the model. + +stages? State-of-the-art models also lack comprehensive integration of semantic, locational, and generic class information in their prompting strategies. In our approach, we leverage CLIP, VQA, and LLM models [1, 55, 56] to extract this discriminative information, improving segmentation performance in unsupervised settings. + +To tackle the challenges posed by limited annotated datasets, several weakly supervised semantic segmentation algorithms [2, 19] have been proposed. Single-stage methods [2] leverage coarse image labels to perform segmentation in an end-to-end fashion, while many two-stage approaches [19] utilize foundational models to generate pseudo labels that can be used to train downstream segmentation architectures. Additionally, human-in-the-loop frameworks [23] have proven beneficial for tasks such as image generation, segmentation, and prognosis, involving annotators to refine segmentation labels or assess the plausibility of synthetic images. This feedback can either be incorporated back into the model or used as ground truth to train a reward function via reinforcement learning (RL), significantly reducing the need for annotated training data. However, these methods still rely on explicit reward modeling, which prevents end-to-end training. In some cases, annotators must provide complex ratings or refinements, making it difficult to train a reward function efficiently. We ask another question: Can we skip training a reward function and develop a straightforward, end-to-end pipeline that relies on simple annotator preferences? To address this, we draw inspiration from direct preference optimization tech + +niques in the LLM preference alignment literature. We propose a preference-based reward model within our framework, using a novel loss function for end-to-end training. Our approach thus aligns with human preferences without the need for explicit reward modeling and is both easy to implement and train. + +Our first goal is to develop refined and efficient prompts to fine-tune the SAM-Med2D encoder on diverse datasets, including 2D chest X-rays, breast ultrasound, and 3D abdominal CT scans. Initially, we input an image into the encoder alongside the BiomedCLIP [55] and MedViInT (VQA) [56] models. Corresponding texts, examples such as "Chest X-ray" and "Describe the condition of the lungs and location of pathologies," are fed into BiomedCLIP and MedViInT, respectively. We gather generic information for the disease class from GPT-4, which is combined with the answers from the VQA model. Saliency maps generated by the CLIP model undergo dense CRF postprocessing to obtain bounding boxes. These bounding boxes and textual prompts are then inputted into the prompt encoder. Subsequently, the mask decoder receives both the image encoder embeddings and prompt embeddings to produce the segmentation maps. + +After fine-tuning the prompting module on a small proportion of annotated data, we introduce our second major idea: simulating human feedback through a AI preference alignment module. We generate multiple segmentation candidates for a given image by thresholding the SAM output probabilities at various levels. These candidates are rated + +on a scale of 1 to 4 based on the overlap between the candidates and ground truth, mimicking the evaluation process of a human annotator. This approach does not require explicit access to ground truth data for training, thus making our task a form of semi-supervised segmentation. We propose a DPO-inspired [38] loss function that encourages the model to prioritize desirable segmentation outputs by rewarding higher-rated candidates and penalizing lower-rated ones. The model is thus trained on the remaining portion of the dataset, without annotations, to perform relevant medical image segmentation tasks. + +Overall our contributions can be summarized as follows: + +- We propose refined and efficient unsupervised prompting strategies that deliver comprehensive information about location, semantics, and general disease/organ characteristics to our SAM-Med2D-based framework. Such an approach enhances segmentation performance while reducing reliance on human input for geometric prompts. +- We introduce a novel DPO-inspired loss function that facilitates semi-supervised model improvement using simulated human-in-the-loop feedback, eliminating the need for an explicit reward function. The framework generates multiple segmentation maps and rates them based on segmentation overlap, mimicking the evaluation of a human annotator. The model learns to distinguish between favorable and unfavorable candidates effectively. + +# 2. Related Work + +# 2.1. Vision-language models for medical domain + +CLIP [37] has gained much popularity in medical image analysis. [14] fine-tuned CLIP on various PubMed articles to create PubMedCLIP. MedCLIP [48] leverages unpaired image and text datasets along with a semantic-matching loss to align visual and textual information. Windsor et al. [49] employ unimodal self-supervision, local-global contrastive losses, and data augmentation to enhance zero-shot performance and retrieval task efficiency in low-resource data settings. Some modality-specific CLIP variants [15, 54] have been developed for Chest X-ray and Mammograms due to the readily available image-text data in these areas. However, BiomedCLIP [55] stands out as the most recent model, excelling in scalability and performance across diverse multi-organ cross-modal retrieval tasks. Therefore, we have integrated BiomedCLIP into the CLIP-driven bounding box generation module of our framework. + +Building on the success of large language models (LLMs) such as LLaMa [44] and GPT [35], researchers have explored merging visual features with textual representations using techniques such as cross-attention, Qformer, instruction tuning, and projection layers. This effort has resulted in vision-language models (VLMs) [4, 26, 30, 58], including Flamingo, BLIP-2, LLaVA, and + +MiniGPT, which were further adapted for the medical domain through pretraining on multimodal medical datasets. Med-Flamingo [33] is the first medical visual question-answering (VQA) model with few-shot generation capabilities. RadFM [50] serves as a foundational model that also incorporates 3D volume data. Pretraining on extensive datasets, PMC-15M and PMC-VQA [55, 56], has facilitated the development of LLaVA-Med [25] and MedViInT [56], respectively. Our segmentation pipeline leverages MedViInT's capabilities for enhanced localization and shape-based answer generation related to tumors, organs, and disease manifestations. + +# 2.2. SAM for multimodal biomedical data + +MedSAM and SAM-Med2D [11, 31] focused on adapting SAM [18] for medical applications by fine-tuning it on 2D medical datasets, while SAM-Med3D [47] introduced alternative modules to accommodate 3D volumes. Efficient approaches, for example, AutoSAM [41] utilize trainable prompt encoders, whereas FastSAM3D [42] employs flash attention to accelerate inference. MedLSAM [24] proposed a localization framework to generate 3D bounding boxes as prompt input. However, most of these methods require ground truth data (bounding boxes or points) for training the prompt encoder, whereas we propose an unsupervised route for the same. We leverage ad-hoc VLM models such as CLIP, VQA, and GPT-4 together to propagate comprehensive information—encompassing semantics, location, and generic disease/organ information—that significantly enhances segmentation performance. + +# 2.3. Human-in-the-loop feedback + +Human-in-the-loop training paradigm in medical imaging has primarily focused on two areas: active learning [34], which identifies optimal data points for labeling to maximize performance, and interactive feedback [8] on model predictions to calibrate parameters. Examples include UI-Net, DeepIGeoS, and BIFseg [5, 45, 46], which utilize expert-provided scribbles or bounding boxes alongside geodesic transforms and graph-cut techniques to refine segmentation labels. Rao et al. [39] introduced IMIL, the first framework that assigns clinicians to guide data augmentations on mispredictions, emphasizing disease-relevant regions while eliminating irrelevant ones. Recently, the success of human feedback in instruction-tuning, and aligning large language model outputs through RL objective [52, 53], has led researchers to utilize human preferences to evaluate synthetic natural and histopathology images, thereby improving image-to-text models [23, 43] by training a reward function. Training a reward function requires the curation of dedicated human preference datasets and also prohibits the framework from operating in an end-to-end manner. Additionally, it often demands advanced do + +![](images/ff8d147200b26c99039395c2a0755a065f7f6ac071c264c6594a6c972ef87cf5.jpg) +Figure 2. Illustration of the proposed framework for semi-supervised segmentation: Unsupervised geometric and text prompts, obtained from pretrained BiomedCLIP, MedVInT, and GPT-4 models, are fed into the prompt encoder for finetuning the framework on a small fraction of annotated data. In the next stage, we simulate a virtual annotation process that assigns ratings to the generated segmentation candidates, which are used to fine-tune the decoder. This stage handles unannotated data, as the model does not rely on ground truth for direct supervision but only for rating while simulating a human annotator's feedback. + +main knowledge from annotators, which can be costly. Our approach addresses this challenge by using direct preference optimization [38] to fulfill the RL objective, enabling the framework to serve as its own reward model. The model aligns generated proposals with appropriate preferences or ratings, leading to performance improvements, even with limited annotated data. These preferences are generated through a simulation mimicking human-in-the-loop feedback. Crucially, our method is easier to train and implement than traditional reward function-based RL pipelines. + +# 3. Methodology + +Overview. Given an input image, our aim is to generate prompts in the form of bounding boxes and textual queries, which will be used to produce segmentation masks. The image is processed by three pretrained components: the SAM-Med2D encoder, BiomedCLIP, and MedVInT. Concurrently, GPT-4 curates generic information about the concerned disease or organ. This information, combined with the BiomedCLIP-generated box prompt and the MedVInT + +provided textual answer prompt, serves as input for the prompt encoder. The prompt generation process is detailed in Sec. 3.1. Next, both the encoded image and prompt embeddings are fed into the decoder to create segmentation maps. Following this initial fine-tuning on annotated data, we introduce a novel loss to mimic human-in-the-loop feedback. By thresholding the output probabilities, our model produces four different segmentation candidates, and we incorporate insights about the quality of these candidates into the framework through the proposed loss function. This part, trained with unannotated data, is elaborated in Sec. 3.2. + +# 3.1. Visual and textual prompt generation + +Visual prompt. To generate visual prompts, we first input the image along with corresponding text prompts (e.g. "chest x-ray", "benign breast tumor", "left kidney", "liver", etc.) into BiomedCLIP, a foundational model pretrained on millions of medical image-caption pairs. Next, we leverage gScoreCAM [10] to create a saliency map highlighting targeted regions (organs or tumors) in the image corre + +sponding to the supplied text. These saliency maps are then post-processed with a conditional random field filter [22] to produce coarse segmentation masks. We apply an area constraint, retaining the largest connected component or up to two, depending on the dataset (the lung dataset may yield two components). Finally, we identify the bounding box coordinates within these closed components for the box prompt, while randomly sampling several points from the designated area for the point prompt. + +Textual prompt. We extract visual embeddings from a given image by processing it through a vision encoder derived from the PMC-CLIP architecture [28]. This encoder features a pretrained ResNet50 [16] and a trainable projection layer constructed with stacks of transformer decoder blocks. Next, we create a prompt template that incorporates the question for the image, formatted as "Question: $\{\}$ , Answer is:". This prompt is then passed through a tokenization embedding layer initialized with the weights of PMC-LLaMA [51] to generate the text embedding. Finally, we concatenate the visual and text embeddings to form the input space for a pre-initialized multimodal transformer decoder. The answer generated from this VQA setup delivers essential information regarding the shape and location of anatomical structures and pathologies. Sample questions included in the prompt are: "What is the shape of the liver and where is it located?", and "What is the shape of breast tumor and where is it located?" Additionally, we prompt GPT-4 with the relevant organ or disease label to obtain a generic description of its characteristics. Finally, both types of textual prompts are concatenated and provided as input to the prompt encoder, described in the following paragraph. + +Prompt encoder and mask decoder. The prompt encoder, same as the one in SAM, supports three types of prompts: point, box, and text. Point and box prompts are represented by their positional encodings—specific coordinates for points, and the top-left and bottom-right corners for boxes—along with learnable feature representations. Text prompts are processed through a pretrained BiomedCLIP encoder to generate corresponding text embeddings. All prompt embeddings are then projected into 256-dimensional vectors. The feature map from the first iteration of the model, is downsampled through multiple convolutional layers followed by GELU activation to match the 256-channel dimension. Finally, these downsampled masks are combined element-wise with visual encoder embeddings. The mask decoder takes both the prompt embeddings and the visual embeddings to produce a segmentation map. The architecture is illustrated in Fig. 2. + +# 3.2. Segmentation proposal generation and Preference Alignment + +Proposal generation. After fine-tuning our framework, comprising the prompt encoder and mask decoder, using + +ground truth for a fraction (10%) of the dataset, we propose integrating AI-based preferences for the next training episode. First, we generate multiple segmentation proposals by sampling different thresholds from the pixel probability scores output by the model. We then simulate an expert annotator to evaluate the quality of these proposals, assigning ratings based on the overlap between each generated proposal and the corresponding ground truth. Although ground truth is not explicitly used in this training phase, it plays a passive role by enabling the simulation of preferences in the absence of an actual human annotator in our experiments. We also experiment with two alternative rating mechanisms: one inspired by SAM, which backpropagates loss solely for the best candidate, and another that ranks all proposals rather than merely rating them. After obtaining the ratings or rankings, we fine-tune our decoder using the direct preference optimization technique to better align the segmentation outputs with the preferences of the virtual annotator. + +RLHF Preliminaries. Language models typically utilize a reward function to align their generated responses with user preferences. This preference distribution is often modeled using the Bradley-Terry [7] model, which operates on pairwise comparisons. For a given prompt $X$ , when the language model produces two responses, one more favorable, $Y_{m}$ , and the other less favorable, $Y_{l}$ , the distribution can be expressed as: + +$$ +P \left(Y _ {m} > Y _ {l} \mid X\right) = \frac {e ^ {r ^ {*} \left(X , Y _ {m}\right)}}{e ^ {r ^ {*} \left(X , Y _ {m}\right)} + e ^ {r ^ {*} \left(X , Y _ {l}\right)}}, \tag {1} +$$ + +Parameters of the reward model can be estimated via Maximum Likelihood Estimation (MLE) and the goal is to minimize the below loss function: + +$$ +\mathcal {L} _ {\mathcal {R}} = - \mathbb {E} _ {(X, Y _ {m}, Y _ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(r _ {\phi} (X, Y _ {m}) - r _ {\phi} (X, Y _ {l})\right) \right], \tag {2} +$$ + +where $\sigma$ is the logistic function and $\mathcal{D}$ denotes the dataset of preferences. + +Reinforcement Learning from Human Feedback (RLHF) is a widely used method that involves training a reward model based on user ratings. The primary goal is to discover an optimal policy with parameter $\pi_{\theta}$ that maximizes this reward function while incorporating a KL divergence term to ensure the model's outputs do not deviate significantly from the original policy $\pi_{ref}$ . The optimization problem can be formulated as follows: + +$$ +\begin{array}{l} \max _ {\pi_ {\theta}} \mathbb {E} _ {X \sim \mathcal {D}, Y \sim \pi_ {\theta} (Y | X)} \left[ r _ {\phi} (X, Y) \right] \tag {3} \\ - \beta \mathbb {D} _ {\mathrm {K L}} \left[ \pi_ {\theta} (Y \mid X) | | \pi_ {r e f} (Y \mid X) \right], \\ \end{array} +$$ + +Direct Preference optimization. Due to the discrete nature of language generation, this objective is not differentiable, which usually necessitates optimization through + +reinforcement learning. The primary language model then leverages this framework to align its outputs with user ratings, ultimately generating high-scoring responses. In contrast, Direct Preference Optimization (DPO) is a more recent and streamlined approach that eliminates the need for a separate reward model. Instead, it fine-tunes the main language model directly using user preferences. This is achievable because DPO focuses on optimizing the policy itself rather than the reward function. The maximum likelihood objective for a parameterized policy $\pi_{\theta}$ in such a case can be formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {D P O}} \left(\pi_ {\theta}; \pi_ {r e f}\right) = - \mathbb {E} _ {\left(X, Y _ {m}, Y _ {l}\right) \sim \mathcal {D}} \left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta} \left(Y _ {m} \mid X\right)}{\pi_ {r e f} \left(Y _ {m} \mid X\right)} \right. \right. \tag {4} \\ \left. \left. - \beta \log \frac {\pi_ {\theta} \left(Y _ {l} \mid X\right)}{\pi_ {r e f} \left(Y _ {l} \mid X\right)}\right) \right], \\ \end{array} +$$ + +where $\beta$ is the weight applied to reward or penalize the responses. In DPO, models are fine-tuned using pairs of outputs, explicitly comparing a preferred response $Y_{m}$ with less preferred ones $Y_{l}$ . We adapt this concept for our application in the imaging domain, where multiple segmentation candidates $Y_{1},\ldots ,Y_{4}$ are generated for the same {image, prompt} pair $I$ . With ratings available for each candidate, we assign weights $\beta_{1},\beta_{2}$ to these candidates (from best to worst) to reward the higher-quality segmentations $Y_{1},Y_{2}$ and penalize the less desirable ones $Y_{3},Y_{4}$ . Consequently, Equation 4 is modified to create Equation 5. This approach allows us to incorporate real-world annotator preferences into our method through a straightforward yet effective loss function. The parameters of the initially fine-tuned model from Sec. 3.1 are denoted as $\pi_{\mathrm{fine}}$ . Our objective is to find the optimal parameters $\pi_{\psi}$ for the final architecture without deviating much from $\pi_{\mathrm{fine}}$ . + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {D P O}} (\pi_ {\psi}; \pi_ {f i n e}) = - \mathbb {E} _ {(I, Y _ {1}, Y _ {2}, Y _ {3}, Y _ {4}) \sim \mathcal {D}} \left[ \log \sigma \left(\beta_ {1} \log \frac {\pi_ {\psi} (Y _ {1} \mid I)}{\pi_ {f i n e} (Y _ {1} \mid I)} \right. \right. \\ \left. + \beta_ {2} \log \frac {\pi_ {\psi} (Y _ {2} \mid I)}{\pi_ {f i n e} (Y _ {2} \mid I)} - \beta_ {2} \log \frac {\pi_ {\psi} (Y _ {3} \mid I)}{\pi_ {f i n e} (Y _ {3} \mid I)} - \beta_ {1} \log \frac {\pi_ {\psi} (Y _ {4} \mid I)}{\pi_ {f i n e} (Y _ {4} \mid I)}\right) \tag {5} \\ \end{array} +$$ + +# 4. Experimental Design and Results + +Datasets. We evaluated our framework for semi-supervised segmentation using three public datasets covering lung, breast tumor, and abdominal organ segmentation across X-ray, ultrasound, and CT. For Ultrasound Breast Tumor Segmentation, we used 810 images from BUSI [3] and UDIAT [9], with 600 for training and 210 for testing. For Chest X-ray Lung Segmentation, we used 27,132 images from COVID-QU-Ex [12], spanning normal, lung opacity, viral pneumonia, and COVID-19 cases. For Abdominal CT Organ Segmentation, 200 annotated CT scans from the AMOS-CT training set were used for training. Full dataset details, including train-test splits, are in the supplementary Sec. 6. + +Implementation Details. Our method is implemented in PyTorch [36] on an EC2 instance (with 64 GB NVIDIA T4 Tensor Core GPUs). For feature extraction, we utilize the SAM-Med2D pretrained encoder. First we use annotations for only $10\%$ of the training dataset, during which all components (visual encoder, prompt encoder, and mask decoder) are fine-tuned. In this step, only the unsupervised prompting strategy is employed. Bounding box and point prompts are used together for all experiments. The remaining portion of the dataset was used in an unannotated form to train the DPO-driven alignment strategy. We optimize the model using the Adam optimizer [13], training for 15 epochs for prompt module fine-tuning and 30 epochs for alignment. In both stages, the initial learning rate is set to 1e-4 and is halved every 10 epochs. All images are resized to a resolution of $256 \times 256$ using the same resizing strategy as SAM-Med2D. Similar to our framework, we have also fine-tuned SOTA SAM-based methods on medical imaging datasets using LoRA by adding low-rank matrices to transformer layers. This enables efficient adaptation with fewer parameters, enhancing domain-specific performance while preserving pretrained features. The loss function used during the initial fine-tuning is a 20:1 weighted combination of focal loss [27] and Dice loss [32]. Details on the preference alignment module and proposal generation are provided in the supplementary material Sec. 6. Dice Similarity Coefficient, Intersection over Union (IoU), and Surface Dice Similarity (SDC) scores are used to evaluate segmentation performance. + +# 4.1. Comparison with state-of-the-art + +Quantitative results. Fig. 3 compares the performance of our framework with relevant methods (U-Net [40], nnU-Net [17], SAM [18], SAM-Med2D [11], Self-prompt). The self-prompt method is designed as a variant of [57], excluding the knowledge distillation module. We also design a prompt-only baseline of our framework which is trained on different splits of fully annotated data. Our framework is initially trained with only the prompting module using $10\%$ of data (annotated). As a result, both the prompt-only and final versions exhibit identical performance on this $10\%$ subset. For the final model, the additional data used for training are considered unannotated, as training the alignment mechanism does not use supervision from ground truth. SOTA methods directly use ground truth for the entire data subset. Nevertheless, our method consistently outperforms them in limited data settings $(10 - 50\%)$ due to the preference alignment mechanism. Our architecture thus reduces the reliance on large datasets, highlighting its lower annotation requirements, which makes it highly effective in low-data regimes. On the Chest-Xray dataset, for instance, with just $20\%$ of the data, our method achieves a Dice score of 78.87, compared to 58.66 (U-Net), 60.97 (nnU-Net), 61.64 (SAM), + +![](images/050dfb3d8688e5f902c8fdcb41ded97f761872112a1c636967ced906c1fba6b7.jpg) +Figure 3. Quantitative comparison with SOTA. Dice score (for Chest Xray, Breast USD) and mean Dice score (for AMOS CT) have been shown to measure the model segmentation performance on different proportions of training data $(10\%, 20\%, 50\%,$ and $100\%)$ . + +![](images/491cb3ff6193712fc86b0dc8a6ee47923f533c5f8c0cbc52a182c46747ba49f0.jpg) + +![](images/adf7f09c9307423f79d8a845ac3a4a6fa65799c80aa5665f412048bb55fa6ec1.jpg) + +![](images/ab375ac6a1e9a03b9d80d0a0c9216ea07abf42b6ebe17e04345333489888f38d.jpg) +Figure 4. Qualitative comparisons were made between the segmentation results of nnUnet, SAM-Med2D, and our framework on 2D datasets. BiomedCLIP-based saliency maps are also depicted. Experiments were conducted in $50\%$ data settings. + +67.81 (SAM-Med2D), and 68.41 (Self-prompt). Both U-Net and nnU-Net are known to be data-hungry models, struggling with smaller data subsets. Our usage of visual and textual prompts, derived through BiomedCLIP and VLM models, offers superior signals to those used in SAM and SAM-Med2D. Unlike these SAM variants, Self-Prompt SAM does not require expert-supplied prompts during inference; it instead generates prompts from the output masks in each iteration. This method offers a slight improvement $(+1\%)$ over SAM-Med2D. Our method shows more stable performance gains across different data subsets, while other SOTA methods exhibit steeper improvements. In the $50\%$ data setting, our semi-supervised framework achieves a Dice score of 89.68, compared to 91.42 for the supervised prompt-only version. Similar result trends are observed in the breast ultrasound dataset. Notably, the performance jump of our method from $20\%$ to $50\%$ data is much more pronounced than from $10\%$ to $20\%$ . This can be due to the nature of the dataset, as the model requires more data to achieve precise and accurate segmentation of breast tumors. We also evaluated our method on a 3D abdominal organ segmentation dataset, including organs such as the liver, kidneys, spleen, pancreas, aorta, bladder, etc (see Fig. 3). Our method outperforms the SOTA across all data proportions except for the full dataset. With $20\%$ of the data, our method achieves a mean Dice score of 77.69, surpassing U + +Net (59.35), nnU-Net (65.21), SAM (64.93), SAM-Med2D (66.57), SAMMed3D (72.54), and Self-prompt (71.83). At the $50\%$ data setting, it reaches a Dice score of 85.70, comparable to the 86.36 achieved by the prompt-only version. + +![](images/769e6a3b7bc6d232524c11ee4463ef95231d2e5eeded689edb65501c78e7e839.jpg) +Figure 5. Segmentation maps of different anatomical structures (liver and spleen) for SAM-Med3D and our method + +Qualitative results. We present segmentation maps for 3 datasets in Fig. 4 and Fig. 5, generated by our model trained at $50\%$ data settings. The saliency maps could correctly highlight the target regions and proved to be an excellent source of supervision. It can be noted from Fig. 4 that our segmentation quality around the boundaries of tumor or lung is much superior compared to both nnUnet and SAM-Med2D. In Fig. 5, SAM-Med3D tended to under-segment the edges of both the abdominal organs, the liver, + +
SupervisionMethodsChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedOurs74.3078.8764.3575.8877.6978.34
20%- Alignment75.0279.1367.5181.3879.2080.66
10%- Alignment68.4375.6061.4473.6274.7776.06
- Alignment - VQA66.9073.3557.6070.5374.0875.41
- Alignment - VQA - GPT463.1672.7654.2668.8973.1674.89
- Alignment - VQA - CAM50.3857.0245.1659.0569.9771.45
+ +Table 1. Ablation results demonstrating effectiveness of major components. + +
SupervisionPreference AlignmentChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedLoss for best candidate70.3777.0961.7673.8175.0176.36
Rating73.9978.4163.9775.5277.2377.98
Ranking74.3078.8764.3575.8877.6978.34
10% + 20% unannotatedRanking79.7485.1573.5684.2380.5481.95
10% + 40% unannotatedRanking88.9689.6885.9288.1584.3085.70
+ +Table 2. Ablation results on different proportions of training data for 3 types of preference scoring strategies. + +and spleen. Additional results are in supplementary Sec. 1. + +# 4.2. Ablation studies + +Effectiveness of major components. We conduct several ablation experiments (see Tab. 1) to evaluate the contribution of each module in our architecture. For simplicity, we focus on the Chest-Xray dataset for analysis. Before incorporating the DPO-driven alignment strategy, all baselines were trained with $10\%$ labeled data. First, we developed a naive baseline (last row), which relies solely on textual answers generated from GPT-4. As expected, it performed poorly, achieving a 57.02 Dice score. A second baseline (second-last row) was designed to use only visual prompts from BiomedCLIP, which provides semantic information. This method significantly benefited from the coarse segmentation masks derived from CLIP saliency maps, improving the Dice score to 72.76. Next, we combined both textual and visual prompts to form a third baseline (-Alignment-VQA), which resulted in a slight improvement $(+0.59\%)$ over the CLIP-only baseline. Finally, we integrated VQA into the prompting strategy to obtain answers related to the shape and location of the target regions, completing our fully empowered unsupervised prompting strategy (-Alignment). This model achieved a Dice score of 75.60 with $10\%$ data and improved to 79.13 with $20\%$ data. In comparison, with the preference alignment mechanism, fine-tuning the prompting module with $10\%$ annotated data and training the alignment module on an additional $10\%$ unannotated data achieved a Dice score of 78.87. This is on par with the fully supervised prompt-only model, underscoring the effectiveness of our alignment module. + +Preference Scoring strategies. We conducted several ablation studies (detailed in Tab. 2) to evaluate the effec + +tiveness of the preference-scoring strategy. One baseline approach involved backpropagating the loss based only on the best candidate, while another compared ranking the candidates rather than simply rating them. The results are presented for different proportions of unannotated data, on top of the foundational fine-tuning of both the prompt encoder and the decoder using $10\%$ annotated data. At $10\%$ unannotated data, both the ranking and rating approaches performed similarly (mean Dice scores of 77.69 and 77.23 for AMOS CT, respectively), significantly outperforming the best-candidate-only method (75.01). Additionally, we observed that as the proportion of unannotated data increased, the model's performance improved substantially. + +Robustness to noisy rating. We experimented with introducing noise into the rating mechanism by flipping the ratings of closely ranked candidates to enhance the framework's robustness. Results are in supplementary material Sec. 2. + +# 5. Conclusion + +We introduce a novel training strategy to enhance SAM representations for semi-supervised medical image segmentation. We extract integrated semantic, location, and shape information from pretrained vision-language models in an unsupervised manner. This information is used as refined prompts for our model. We also implement an optimal policy, inspired by direct preference optimization in language models. This enables human-in-the-loop feedback simulation within a streamlined training framework, without the need for a separate reward function or extensive knowledge from annotators. These modules ensure that our framework outperforms state-of-the-art methods across datasets spanning multiple modalities in low-annotation data scenarios. + +# 6. Acknowledgments + +We sincerely thank everyone from the GE Healthcare Science & Technology Organization for their support and contributions to the project. The primary model development and evaluations were conducted during a summer internship at GEHC. After the internship, Konwer and Prasanna were supported by the following grants during the paper writing and rebuttal process: NIH 1R01CA297843-01, NIH 3R21CA258493-02S1, and 1R21CA258493-01A1. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Peri Akiva and Kristin Dana. Towards single stage weakly supervised semantic segmentation. arXiv preprint arXiv:2106.10309, 2021. 2 +[3] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 2 +[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3 +[5] Mario Amrehn, Sven Gaube, Mathias Unberath, Frank Schebesch, Tim Horz, Maddalena Strumia, Stefan Steidl, Markus Kowarschik, and Andreas Maier. Ui-net: Interactive artificial neural networks for iterative image segmentation based on a user model. arXiv preprint arXiv:1709.03450, 2017. 3 +[6] Louis Blankemeier, Joseph Paul Cohen, Ashwin Kumar, Dave Van Veen, Syed Jamal Safdar Gardezi, Magdalini Paschali, Zhihong Chen, Jean-Benoit Delbrouck, Eduardo Reis, Cesar Truyts, et al. Merlin: A vision language foundation model for 3d computed tomography. arXiv preprint arXiv:2406.06512, 2024. 1 +[7] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5 +[8] Samuel Budd, Emma C Robinson, and Bernhard Kainz. A survey on active learning and human-in-the-loop deep learning for medical image analysis. Medical image analysis, 71: 102062, 2021. 3 +[9] Michal Byra, Piotr Jarosik, Aleksandra Szubert, Michael Galperin, Haydee Ojeda-Fournier, Linda Olson, Mary O'Boyle, Christopher Comstock, and Michael Andre. Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control, 61:102027, 2020. 6, 2 +[10] Peijie Chen, Qi Li, Saad Biaz, Trung Bui, and Anh Nguyen. gscorecam: What objects is clip looking at? In Proceedings + +of the Asian Conference on Computer Vision (ACCV), pages 1959-1975, 2022. 4 +[11] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 3, 6 +[12] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and coid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 6, 2 +[13] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6 +[14] Sedigheh Eslami, Christoph Meinel, and Gerard De Melo. Pubmedclip: How much does clip benefit visual question answering in the medical domain? In Findings of the Association for Computational Linguistics: EACL 2023, pages 1181-1193, 2023. 3 +[15] Shantanu Ghosh, Clare B Poynton, Shyam Visweswaran, and Kayhan Batmanghelich. Mammo-clip: A vision language foundation model to enhance data efficiency and robustness in mammography. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 632–642. Springer, 2024. 3 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[17] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6 +[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3, 6 +[19] Taha Koleilat, Hojat Asgariandehkordi, Hassan Rivaz, and Yiming Xiao. Medclip-sam: Bridging text and image towards universal medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 643-653. Springer, 2024. 1, 2 +[20] Aishik Konwer, Xuan Xu, Joseph Bae, Chao Chen, and Pratek Prasanna. Temporal context matters: Enhancing single image prediction with disease progression representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18824-18835, 2022. 1 +[21] Aishik Konwer, Xiaoling Hu, Joseph Bae, Xuan Xu, Chao Chen, and Prateek Prasanna. Enhancing modality-agnostic representations via meta-learning for brain tumor segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21415–21425, 2023. 1 +[22] Philipp Krahenbuhl and Vladlen Koltun. Parameter learning and convergent inference for dense random fields. In Inter- + +national conference on machine learning, pages 513-521. +PMLR, 2013. 5 +[23] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3 +[24] Wenhui Lei, Xu Wei, Xiaofan Zhang, Kang Li, and Shaoting Zhang. Medlsam: Localize and segment anything model for 3d medical images. arXiv preprint arXiv:2306.14752, 2023. 1, 3 +[25] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 3 +[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3 +[27] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 6 +[28] Weixiong Lin, Ziheng Zhao, Xiaoman Zhang, Chaoyi Wu, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-clip: Contrastive language-image pre-training using biomedical documents. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 525–536. Springer, 2023. 5 +[29] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3 +[31] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 3 +[32] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6 +[33] Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Eduardo Pontes Reis, and Pranav Rajpurkar. Med-flamingo: a multimodal medical few-shot learner. In Machine Learning for Health (ML4H), pages 353-367. PMLR, 2023. 3 +[34] Vishwesh Nath, Dong Yang, Bennett A. Landman, Daguang Xu, and Holger R. Roth. Diminishing uncertainty within the training pool: Active learning for medical image segmentation. IEEE Transactions on Medical Imaging, 40(10): 2534-2547, 2021. 3 +[35] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini + +Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3 +[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[38] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4 +[39] Adrit Rao, Andrea Fisher, Ken Chang, John Christopher Panagides, Katherine McNamara, Joon-Young Lee, and Oliver Aalami. Imil: Interactive medical image learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5241–5250, 2024. 3 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6 +[41] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3 +[42] Yiqing Shen, Jingxing Li, Xinyuan Shao, Blanca Inigo Romillo, Ankush Jindal, David Dreizin, and Mathias Unberath. Fastsam3d: An efficient segment anything model for 3d volumetric medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 542-552. Springer, 2024. 3 +[43] Shenghuan Sun, Greg Goldgof, Atul Butte, and Ahmed M Alaa. Aligning synthetic medical images with clinical knowledge using human feedback. Advances in Neural Information Processing Systems, 36, 2024. 3 +[44] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[45] Guotai Wang, Wenqi Li, Maria A Zuluaga, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sebastien Ourselin, et al. Interactive medical image segmentation using deep learning with image-specific fine tuning. IEEE transactions on medical imaging, 37(7): 1562-1573, 2018. 3 + +[46] Guotai Wang, Maria A Zuluaga, Wenqi Li, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sébastien Ourselin, et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 41(7):1559-1572, 2018. 3 +[47] H Wang et al. Sam-med3d: towards general-purpose segmentation models for volumetric medical images. Preprint at https://arxiv.org/abs/2310.15161, 2024. 1, 3 +[48] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163, 2022. 3 +[49] Rhydian Windsor, Amir Jamaludin, Timor Kadir, and Andrew Zisserman. Vision-language modelling for radiological imaging and reports in the low data regime. arXiv preprint arXiv:2303.17644, 2023. 3 +[50] Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology. arXiv preprint arXiv:2308.02463, 2023. 3 +[51] Chaoyi Wu, Weixiong Lin, Xiaoman Zhang, Ya Zhang, Weidi Xie, and Yanfeng Wang. Pmc-llama: toward building open-source language models for medicine. Journal of the American Medical Informatics Association, page ocae045, 2024. 5 +[52] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3 +[53] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems, 36:59008-59033, 2023. 3 +[54] Kihyun You, Jawook Gu, Jiyeon Ham, Beomhee Park, Jiho Kim, Eun K Hong, Woonhyuk Baek, and Byungseok Roh. Cxr-clip: Toward large scale chest x-ray language-image pre-training. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 101-111. Springer, 2023. 3 +[55] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 1, 2, 3 +[56] Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023. 2, 3 +[57] Chunpeng Zhou, Kangjie Ning, Qianqian Shen, Sheng Zhou, Zhi Yu, and Haishuai Wang. Sam-sp: Self-prompting makes sam great again. arXiv preprint arXiv:2408.12364, 2024. 1, 2, 6 +[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language + +understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 \ No newline at end of file diff --git a/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/images.zip b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..76797051a179672271841bcba8558b059552db7c --- /dev/null +++ b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f2d580ecbc888d2dd9c61827367141ea65d177e16af5134ebbf73288f1dd36e +size 425972 diff --git a/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/layout.json b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e66aec486ea670749b3b7dcae454c96cc70b13f --- /dev/null +++ b/2025/Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation/layout.json @@ -0,0 +1,6894 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 536, + 140 + ], + "type": "text", + "content": "Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "spans": [ + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": "Aishik Konwer" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": ", Zhijian Yang" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": ", Erhan Bas" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": ", Cao Xiao" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": ", Prateek Prasanna" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": ", Parminder Bhatia" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": ", Taha Kass-Hout" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 165, + 510, + 208 + ], + "type": "text", + "content": "GE Healthcare" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 209, + 447, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 209, + 447, + 224 + ], + "spans": [ + { + "bbox": [ + 161, + 209, + 447, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 209, + 447, + 224 + ], + "type": "text", + "content": "Department of Computer Science, Stony Brook University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 224, + 459, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 224, + 459, + 238 + ], + "spans": [ + { + "bbox": [ + 148, + 224, + 459, + 238 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 224, + 459, + 238 + ], + "type": "text", + "content": "Department of Biomedical Informatics, Stony Brook University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 148, + 239, + 460, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 239, + 460, + 251 + ], + "spans": [ + { + "bbox": [ + 148, + 239, + 460, + 251 + ], + "type": "text", + "content": "akonwer@cs.stonybrook.edu, prateek.prasanna@stonybrook.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 253, + 536, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 253, + 536, + 265 + ], + "spans": [ + { + "bbox": [ + 71, + 253, + 536, + 265 + ], + "type": "text", + "content": "{zhijian.yang, erhan.bas, cao.xiao, paraminder.bhatia, taha.kass-hout}@gehealthcare.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 293, + 200, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 293, + 200, + 306 + ], + "spans": [ + { + "bbox": [ + 151, + 293, + 200, + 306 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 318, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 297, + 605 + ], + "type": "text", + "content": "Foundational models such as the Segment Anything Model (SAM) are gaining traction in medical imaging segmentation, supporting multiple downstream tasks. However, such models are supervised in nature, still relying on large annotated datasets or prompts supplied by experts. Conventional techniques such as active learning to alleviate such limitations are limited in scope and still necessitate continuous human involvement and complex domain knowledge for label refinement or establishing reward ground truth. To address these challenges, we propose an enhanced Segment Anything Model (SAM) framework that utilizes annotation-efficient prompts generated in a fully unsupervised fashion, while still capturing essential semantic, location, and shape information through contrastive language-image pretraining and visual question answering. We adopt the direct preference optimization technique to design an optimal policy that enables the model to generate high-fidelity segmentations with simple ratings or rankings provided by a virtual annotator simulating the human annotation process. State-of-the-art performance of our framework in tasks such as lung segmentation, breast tumor segmentation, and organ segmentation across various modalities, including X-ray, ultrasound, and abdominal CT, justifies its effectiveness in low-annotation data scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "type": "text", + "content": "With advancements in medical image analysis, there is an increasing need for sophisticated methods [20, 21, 29] to leverage the vast availability of radiology datasets (such as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 293, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 498 + ], + "type": "text", + "content": "X-ray, CT, and MRI) for accurate organ and tumor segmentation, as well as disease classification. The results of these tasks are crucial for physicians in designing effective treatment plans and surgical procedures. Several state-of-the-art deep learning-based foundational models, such as Vision-Language Models (VLMs), are now available for these purposes, relying on custom prompting to generate relevant predictions. However, they face two significant challenges: (1) despite being trained with only sparse prompts such as points or bounding boxes, these models still require human supervision for the prompt generation, leading to inefficiencies; and (2) many datasets lack comprehensive annotations, resulting in under-utilization during training of complex, data-hungry foundational architectures. Additionally, the high cost of human annotation efforts to create ground truth data can significantly escalate model development expenses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Popular architectures such as SAM and CLIP, have been extended to medical data, and led to innovations such as BiomedCLIP, Merlin, SAM-Med2D, and SAM-Med3D [6, 11, 47, 55]. Recently, numerous studies have focused on enhancing SAM by replacing geometric prompts and integrating semantic and spatial knowledge in an unsupervised manner through techniques such as (1) self-prompting, (2) class activation maps from CLIP, and (3) object localization models [19, 24, 57]. As shown in Fig. 1, self-prompting (b) does not require expert-provided prompts during inference. On the other hand, unsupervised prompts (c) in [19] are only used for training SAM to generate pseudo-labels for weakly supervised downstream tasks. They lack sufficient location and shape information, which could be provided by text-based prompts. Hence, the question arises: Can we come up with more refined and efficient prompts that can deliver stronger signals to foundational models without requiring intervention from annotators in both training and testing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 222, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 222, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 222, + 703 + ], + "type": "text", + "content": "*Work done during internship at GE Healthcare" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20990" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 71, + 504, + 289 + ], + "blocks": [ + { + "bbox": [ + 106, + 71, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 71, + 504, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 71, + 504, + 289 + ], + "type": "image", + "image_path": "bb07502a342acdf16cb10ed1cf4f3b814277888c18e400f1078248d11ecce69a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 555, + 342 + ], + "type": "text", + "content": "Figure 1. Overview of our model: (a) SAM and SAM-based approaches rely on expert prompts during both training and inference. (b) SAM-SP [57] introduces a self-prompting module, eliminating the need for expert prompts during inference. (c) MedCLIP-SAM [19] uses unsupervised semantic prompts to generate pseudo-labels through SAM. (d) Our approach not only combines semantic, location, and generic information via unsupervised prompts but also introduces a preference-based alignment module to reward or penalize the model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 361, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 434 + ], + "type": "text", + "content": "stages? State-of-the-art models also lack comprehensive integration of semantic, locational, and generic class information in their prompting strategies. In our approach, we leverage CLIP, VQA, and LLM models [1, 55, 56] to extract this discriminative information, improving segmentation performance in unsupervised settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 715 + ], + "type": "text", + "content": "To tackle the challenges posed by limited annotated datasets, several weakly supervised semantic segmentation algorithms [2, 19] have been proposed. Single-stage methods [2] leverage coarse image labels to perform segmentation in an end-to-end fashion, while many two-stage approaches [19] utilize foundational models to generate pseudo labels that can be used to train downstream segmentation architectures. Additionally, human-in-the-loop frameworks [23] have proven beneficial for tasks such as image generation, segmentation, and prognosis, involving annotators to refine segmentation labels or assess the plausibility of synthetic images. This feedback can either be incorporated back into the model or used as ground truth to train a reward function via reinforcement learning (RL), significantly reducing the need for annotated training data. However, these methods still rely on explicit reward modeling, which prevents end-to-end training. In some cases, annotators must provide complex ratings or refinements, making it difficult to train a reward function efficiently. We ask another question: Can we skip training a reward function and develop a straightforward, end-to-end pipeline that relies on simple annotator preferences? To address this, we draw inspiration from direct preference optimization tech" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 361, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 433 + ], + "type": "text", + "content": "niques in the LLM preference alignment literature. We propose a preference-based reward model within our framework, using a novel loss function for end-to-end training. Our approach thus aligns with human preferences without the need for explicit reward modeling and is both easy to implement and train." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 435, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 639 + ], + "type": "text", + "content": "Our first goal is to develop refined and efficient prompts to fine-tune the SAM-Med2D encoder on diverse datasets, including 2D chest X-rays, breast ultrasound, and 3D abdominal CT scans. Initially, we input an image into the encoder alongside the BiomedCLIP [55] and MedViInT (VQA) [56] models. Corresponding texts, examples such as \"Chest X-ray\" and \"Describe the condition of the lungs and location of pathologies,\" are fed into BiomedCLIP and MedViInT, respectively. We gather generic information for the disease class from GPT-4, which is combined with the answers from the VQA model. Saliency maps generated by the CLIP model undergo dense CRF postprocessing to obtain bounding boxes. These bounding boxes and textual prompts are then inputted into the prompt encoder. Subsequently, the mask decoder receives both the image encoder embeddings and prompt embeddings to produce the segmentation maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "After fine-tuning the prompting module on a small proportion of annotated data, we introduce our second major idea: simulating human feedback through a AI preference alignment module. We generate multiple segmentation candidates for a given image by thresholding the SAM output probabilities at various levels. These candidates are rated" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20991" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "on a scale of 1 to 4 based on the overlap between the candidates and ground truth, mimicking the evaluation process of a human annotator. This approach does not require explicit access to ground truth data for training, thus making our task a form of semi-supervised segmentation. We propose a DPO-inspired [38] loss function that encourages the model to prioritize desirable segmentation outputs by rewarding higher-rated candidates and penalizing lower-rated ones. The model is thus trained on the remaining portion of the dataset, without annotations, to perform relevant medical image segmentation tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 205, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 205, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 67, + 205, + 294, + 216 + ], + "type": "text", + "content": "Overall our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 217, + 295, + 384 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 217, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 289 + ], + "type": "text", + "content": "- We propose refined and efficient unsupervised prompting strategies that deliver comprehensive information about location, semantics, and general disease/organ characteristics to our SAM-Med2D-based framework. Such an approach enhances segmentation performance while reducing reliance on human input for geometric prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 289, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 384 + ], + "type": "text", + "content": "- We introduce a novel DPO-inspired loss function that facilitates semi-supervised model improvement using simulated human-in-the-loop feedback, eliminating the need for an explicit reward function. The framework generates multiple segmentation maps and rates them based on segmentation overlap, mimicking the evaluation of a human annotator. The model learns to distinguish between favorable and unfavorable candidates effectively." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 398, + 142, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 142, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 142, + 410 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 283, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 283, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 283, + 430 + ], + "type": "text", + "content": "2.1. Vision-language models for medical domain" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 628 + ], + "type": "text", + "content": "CLIP [37] has gained much popularity in medical image analysis. [14] fine-tuned CLIP on various PubMed articles to create PubMedCLIP. MedCLIP [48] leverages unpaired image and text datasets along with a semantic-matching loss to align visual and textual information. Windsor et al. [49] employ unimodal self-supervision, local-global contrastive losses, and data augmentation to enhance zero-shot performance and retrieval task efficiency in low-resource data settings. Some modality-specific CLIP variants [15, 54] have been developed for Chest X-ray and Mammograms due to the readily available image-text data in these areas. However, BiomedCLIP [55] stands out as the most recent model, excelling in scalability and performance across diverse multi-organ cross-modal retrieval tasks. Therefore, we have integrated BiomedCLIP into the CLIP-driven bounding box generation module of our framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Building on the success of large language models (LLMs) such as LLaMa [44] and GPT [35], researchers have explored merging visual features with textual representations using techniques such as cross-attention, Qformer, instruction tuning, and projection layers. This effort has resulted in vision-language models (VLMs) [4, 26, 30, 58], including Flamingo, BLIP-2, LLaVA, and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "MiniGPT, which were further adapted for the medical domain through pretraining on multimodal medical datasets. Med-Flamingo [33] is the first medical visual question-answering (VQA) model with few-shot generation capabilities. RadFM [50] serves as a foundational model that also incorporates 3D volume data. Pretraining on extensive datasets, PMC-15M and PMC-VQA [55, 56], has facilitated the development of LLaVA-Med [25] and MedViInT [56], respectively. Our segmentation pipeline leverages MedViInT's capabilities for enhanced localization and shape-based answer generation related to tumors, organs, and disease manifestations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 224, + 512, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 512, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 512, + 236 + ], + "type": "text", + "content": "2.2. SAM for multimodal biomedical data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 435 + ], + "type": "text", + "content": "MedSAM and SAM-Med2D [11, 31] focused on adapting SAM [18] for medical applications by fine-tuning it on 2D medical datasets, while SAM-Med3D [47] introduced alternative modules to accommodate 3D volumes. Efficient approaches, for example, AutoSAM [41] utilize trainable prompt encoders, whereas FastSAM3D [42] employs flash attention to accelerate inference. MedLSAM [24] proposed a localization framework to generate 3D bounding boxes as prompt input. However, most of these methods require ground truth data (bounding boxes or points) for training the prompt encoder, whereas we propose an unsupervised route for the same. We leverage ad-hoc VLM models such as CLIP, VQA, and GPT-4 together to propagate comprehensive information—encompassing semantics, location, and generic disease/organ information—that significantly enhances segmentation performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 444, + 471, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 471, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 471, + 456 + ], + "type": "text", + "content": "2.3. Human-in-the-loop feedback" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "Human-in-the-loop training paradigm in medical imaging has primarily focused on two areas: active learning [34], which identifies optimal data points for labeling to maximize performance, and interactive feedback [8] on model predictions to calibrate parameters. Examples include UI-Net, DeepIGeoS, and BIFseg [5, 45, 46], which utilize expert-provided scribbles or bounding boxes alongside geodesic transforms and graph-cut techniques to refine segmentation labels. Rao et al. [39] introduced IMIL, the first framework that assigns clinicians to guide data augmentations on mispredictions, emphasizing disease-relevant regions while eliminating irrelevant ones. Recently, the success of human feedback in instruction-tuning, and aligning large language model outputs through RL objective [52, 53], has led researchers to utilize human preferences to evaluate synthetic natural and histopathology images, thereby improving image-to-text models [23, 43] by training a reward function. Training a reward function requires the curation of dedicated human preference datasets and also prohibits the framework from operating in an end-to-end manner. Additionally, it often demands advanced do" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20992" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 82, + 489, + 370 + ], + "blocks": [ + { + "bbox": [ + 123, + 82, + 489, + 370 + ], + "lines": [ + { + "bbox": [ + 123, + 82, + 489, + 370 + ], + "spans": [ + { + "bbox": [ + 123, + 82, + 489, + 370 + ], + "type": "image", + "image_path": "ff8d147200b26c99039395c2a0755a065f7f6ac071c264c6594a6c972ef87cf5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 389, + 555, + 447 + ], + "lines": [ + { + "bbox": [ + 54, + 389, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 555, + 447 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed framework for semi-supervised segmentation: Unsupervised geometric and text prompts, obtained from pretrained BiomedCLIP, MedVInT, and GPT-4 models, are fed into the prompt encoder for finetuning the framework on a small fraction of annotated data. In the next stage, we simulate a virtual annotation process that assigns ratings to the generated segmentation candidates, which are used to fine-tune the decoder. This stage handles unannotated data, as the model does not rely on ground truth for direct supervision but only for rating while simulating a human annotator's feedback." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 466, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 466, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 54, + 466, + 295, + 587 + ], + "type": "text", + "content": "main knowledge from annotators, which can be costly. Our approach addresses this challenge by using direct preference optimization [38] to fulfill the RL objective, enabling the framework to serve as its own reward model. The model aligns generated proposals with appropriate preferences or ratings, leading to performance improvements, even with limited annotated data. These preferences are generated through a simulation mimicking human-in-the-loop feedback. Crucially, our method is easier to train and implement than traditional reward function-based RL pipelines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 597, + 138, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 138, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 138, + 611 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Overview. Given an input image, our aim is to generate prompts in the form of bounding boxes and textual queries, which will be used to produce segmentation masks. The image is processed by three pretrained components: the SAM-Med2D encoder, BiomedCLIP, and MedVInT. Concurrently, GPT-4 curates generic information about the concerned disease or organ. This information, combined with the BiomedCLIP-generated box prompt and the MedVInT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 466, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 598 + ], + "type": "text", + "content": "provided textual answer prompt, serves as input for the prompt encoder. The prompt generation process is detailed in Sec. 3.1. Next, both the encoded image and prompt embeddings are fed into the decoder to create segmentation maps. Following this initial fine-tuning on annotated data, we introduce a novel loss to mimic human-in-the-loop feedback. By thresholding the output probabilities, our model produces four different segmentation candidates, and we incorporate insights about the quality of these candidates into the framework through the proposed loss function. This part, trained with unannotated data, is elaborated in Sec. 3.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 609, + 512, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 512, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 512, + 624 + ], + "type": "text", + "content": "3.1. Visual and textual prompt generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Visual prompt. To generate visual prompts, we first input the image along with corresponding text prompts (e.g. \"chest x-ray\", \"benign breast tumor\", \"left kidney\", \"liver\", etc.) into BiomedCLIP, a foundational model pretrained on millions of medical image-caption pairs. Next, we leverage gScoreCAM [10] to create a saliency map highlighting targeted regions (organs or tumors) in the image corre" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20993" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "sponding to the supplied text. These saliency maps are then post-processed with a conditional random field filter [22] to produce coarse segmentation masks. We apply an area constraint, retaining the largest connected component or up to two, depending on the dataset (the lung dataset may yield two components). Finally, we identify the bounding box coordinates within these closed components for the box prompt, while randomly sampling several points from the designated area for the point prompt." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 181, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 181, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 181, + 295, + 443 + ], + "type": "text", + "content": "Textual prompt. We extract visual embeddings from a given image by processing it through a vision encoder derived from the PMC-CLIP architecture [28]. This encoder features a pretrained ResNet50 [16] and a trainable projection layer constructed with stacks of transformer decoder blocks. Next, we create a prompt template that incorporates the question for the image, formatted as \"Question: " + }, + { + "bbox": [ + 56, + 181, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 56, + 181, + 295, + 443 + ], + "type": "text", + "content": ", Answer is:\". This prompt is then passed through a tokenization embedding layer initialized with the weights of PMC-LLaMA [51] to generate the text embedding. Finally, we concatenate the visual and text embeddings to form the input space for a pre-initialized multimodal transformer decoder. The answer generated from this VQA setup delivers essential information regarding the shape and location of anatomical structures and pathologies. Sample questions included in the prompt are: \"What is the shape of the liver and where is it located?\", and \"What is the shape of breast tumor and where is it located?\" Additionally, we prompt GPT-4 with the relevant organ or disease label to obtain a generic description of its characteristics. Finally, both types of textual prompts are concatenated and provided as input to the prompt encoder, described in the following paragraph." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "type": "text", + "content": "Prompt encoder and mask decoder. The prompt encoder, same as the one in SAM, supports three types of prompts: point, box, and text. Point and box prompts are represented by their positional encodings—specific coordinates for points, and the top-left and bottom-right corners for boxes—along with learnable feature representations. Text prompts are processed through a pretrained BiomedCLIP encoder to generate corresponding text embeddings. All prompt embeddings are then projected into 256-dimensional vectors. The feature map from the first iteration of the model, is downsampled through multiple convolutional layers followed by GELU activation to match the 256-channel dimension. Finally, these downsampled masks are combined element-wise with visual encoder embeddings. The mask decoder takes both the prompt embeddings and the visual embeddings to produce a segmentation map. The architecture is illustrated in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 658, + 294, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 294, + 683 + ], + "type": "text", + "content": "3.2. Segmentation proposal generation and Preference Alignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Proposal generation. After fine-tuning our framework, comprising the prompt encoder and mask decoder, using" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 298 + ], + "type": "text", + "content": "ground truth for a fraction (10%) of the dataset, we propose integrating AI-based preferences for the next training episode. First, we generate multiple segmentation proposals by sampling different thresholds from the pixel probability scores output by the model. We then simulate an expert annotator to evaluate the quality of these proposals, assigning ratings based on the overlap between each generated proposal and the corresponding ground truth. Although ground truth is not explicitly used in this training phase, it plays a passive role by enabling the simulation of preferences in the absence of an actual human annotator in our experiments. We also experiment with two alternative rating mechanisms: one inspired by SAM, which backpropagates loss solely for the best candidate, and another that ranks all proposals rather than merely rating them. After obtaining the ratings or rankings, we fine-tune our decoder using the direct preference optimization technique to better align the segmentation outputs with the preferences of the virtual annotator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "text", + "content": "RLHF Preliminaries. Language models typically utilize a reward function to align their generated responses with user preferences. This preference distribution is often modeled using the Bradley-Terry [7] model, which operates on pairwise comparisons. For a given prompt " + }, + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "text", + "content": ", when the language model produces two responses, one more favorable, " + }, + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "inline_equation", + "content": "Y_{m}" + }, + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "text", + "content": ", and the other less favorable, " + }, + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "inline_equation", + "content": "Y_{l}" + }, + { + "bbox": [ + 313, + 299, + 553, + 396 + ], + "type": "text", + "content": ", the distribution can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 348, + 403, + 553, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 403, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 348, + 403, + 553, + 431 + ], + "type": "interline_equation", + "content": "P \\left(Y _ {m} > Y _ {l} \\mid X\\right) = \\frac {e ^ {r ^ {*} \\left(X , Y _ {m}\\right)}}{e ^ {r ^ {*} \\left(X , Y _ {m}\\right)} + e ^ {r ^ {*} \\left(X , Y _ {l}\\right)}}, \\tag {1}", + "image_path": "817d84a1b8bbb4ce88c08a374c8f4a19348fc0422d8e38f320150b57f3f9f283.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 438, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 553, + 474 + ], + "type": "text", + "content": "Parameters of the reward model can be estimated via Maximum Likelihood Estimation (MLE) and the goal is to minimize the below loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 483, + 553, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 483, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 319, + 483, + 553, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {R}} = - \\mathbb {E} _ {(X, Y _ {m}, Y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r _ {\\phi} (X, Y _ {m}) - r _ {\\phi} (X, Y _ {l})\\right) \\right], \\tag {2}", + "image_path": "d1ed023513ffd57fd05d03ffb5e0cad7d6e944e0cf64237d4543609ef63958a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 509, + 553, + 532 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 509, + 553, + 532 + ], + "type": "text", + "content": " is the logistic function and " + }, + { + "bbox": [ + 313, + 509, + 553, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 509, + 553, + 532 + ], + "type": "text", + "content": " denotes the dataset of preferences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "Reinforcement Learning from Human Feedback (RLHF) is a widely used method that involves training a reward model based on user ratings. The primary goal is to discover an optimal policy with parameter " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " that maximizes this reward function while incorporating a KL divergence term to ensure the model's outputs do not deviate significantly from the original policy " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\pi_{ref}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": ". The optimization problem can be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 360, + 638, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 638, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 360, + 638, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {X \\sim \\mathcal {D}, Y \\sim \\pi_ {\\theta} (Y | X)} \\left[ r _ {\\phi} (X, Y) \\right] \\tag {3} \\\\ - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ \\pi_ {\\theta} (Y \\mid X) | | \\pi_ {r e f} (Y \\mid X) \\right], \\\\ \\end{array}", + "image_path": "ba262a2f3d3a3e19e48a5db6b577cfd0652a7480ef3a6502dbbc7f6d1c371227.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Direct Preference optimization. Due to the discrete nature of language generation, this objective is not differentiable, which usually necessitates optimization through" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20994" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "reinforcement learning. The primary language model then leverages this framework to align its outputs with user ratings, ultimately generating high-scoring responses. In contrast, Direct Preference Optimization (DPO) is a more recent and streamlined approach that eliminates the need for a separate reward model. Instead, it fine-tunes the main language model directly using user preferences. This is achievable because DPO focuses on optimizing the policy itself rather than the reward function. The maximum likelihood objective for a parameterized policy " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " in such a case can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 219, + 296, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 219, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 63, + 219, + 296, + 261 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\pi_ {\\theta}; \\pi_ {r e f}\\right) = - \\mathbb {E} _ {\\left(X, Y _ {m}, Y _ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(Y _ {m} \\mid X\\right)}{\\pi_ {r e f} \\left(Y _ {m} \\mid X\\right)} \\right. \\right. \\tag {4} \\\\ \\left. \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(Y _ {l} \\mid X\\right)}{\\pi_ {r e f} \\left(Y _ {l} \\mid X\\right)}\\right) \\right], \\\\ \\end{array}", + "image_path": "60e4628cea4da76af6674461940b354ac0667afad4afa951b5a63cf16b3669ae.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": " is the weight applied to reward or penalize the responses. In DPO, models are fine-tuned using pairs of outputs, explicitly comparing a preferred response " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "Y_{m}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": " with less preferred ones " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "Y_{l}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": ". We adapt this concept for our application in the imaging domain, where multiple segmentation candidates " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "Y_{1},\\ldots ,Y_{4}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": " are generated for the same {image, prompt} pair " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": ". With ratings available for each candidate, we assign weights " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\beta_{2}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": " to these candidates (from best to worst) to reward the higher-quality segmentations " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "Y_{1},Y_{2}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": " and penalize the less desirable ones " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "Y_{3},Y_{4}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": ". Consequently, Equation 4 is modified to create Equation 5. This approach allows us to incorporate real-world annotator preferences into our method through a straightforward yet effective loss function. The parameters of the initially fine-tuned model from Sec. 3.1 are denoted as " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{fine}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": ". Our objective is to find the optimal parameters " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\pi_{\\psi}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": " for the final architecture without deviating much from " + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{fine}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 479, + 299, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 299, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 299, + 528 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} (\\pi_ {\\psi}; \\pi_ {f i n e}) = - \\mathbb {E} _ {(I, Y _ {1}, Y _ {2}, Y _ {3}, Y _ {4}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta_ {1} \\log \\frac {\\pi_ {\\psi} (Y _ {1} \\mid I)}{\\pi_ {f i n e} (Y _ {1} \\mid I)} \\right. \\right. \\\\ \\left. + \\beta_ {2} \\log \\frac {\\pi_ {\\psi} (Y _ {2} \\mid I)}{\\pi_ {f i n e} (Y _ {2} \\mid I)} - \\beta_ {2} \\log \\frac {\\pi_ {\\psi} (Y _ {3} \\mid I)}{\\pi_ {f i n e} (Y _ {3} \\mid I)} - \\beta_ {1} \\log \\frac {\\pi_ {\\psi} (Y _ {4} \\mid I)}{\\pi_ {f i n e} (Y _ {4} \\mid I)}\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "26d8b50cd6a85eb288502675e458d3e2f7b5db4d18f43c80156ac5befc8293d0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 538, + 241, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 241, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 241, + 552 + ], + "type": "text", + "content": "4. Experimental Design and Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 712 + ], + "type": "text", + "content": "Datasets. We evaluated our framework for semi-supervised segmentation using three public datasets covering lung, breast tumor, and abdominal organ segmentation across X-ray, ultrasound, and CT. For Ultrasound Breast Tumor Segmentation, we used 810 images from BUSI [3] and UDIAT [9], with 600 for training and 210 for testing. For Chest X-ray Lung Segmentation, we used 27,132 images from COVID-QU-Ex [12], spanning normal, lung opacity, viral pneumonia, and COVID-19 cases. For Abdominal CT Organ Segmentation, 200 annotated CT scans from the AMOS-CT training set were used for training. Full dataset details, including train-test splits, are in the supplementary Sec. 6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "text", + "content": "Implementation Details. Our method is implemented in PyTorch [36] on an EC2 instance (with 64 GB NVIDIA T4 Tensor Core GPUs). For feature extraction, we utilize the SAM-Med2D pretrained encoder. First we use annotations for only " + }, + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "text", + "content": " of the training dataset, during which all components (visual encoder, prompt encoder, and mask decoder) are fine-tuned. In this step, only the unsupervised prompting strategy is employed. Bounding box and point prompts are used together for all experiments. The remaining portion of the dataset was used in an unannotated form to train the DPO-driven alignment strategy. We optimize the model using the Adam optimizer [13], training for 15 epochs for prompt module fine-tuning and 30 epochs for alignment. In both stages, the initial learning rate is set to 1e-4 and is halved every 10 epochs. All images are resized to a resolution of " + }, + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "text", + "content": " using the same resizing strategy as SAM-Med2D. Similar to our framework, we have also fine-tuned SOTA SAM-based methods on medical imaging datasets using LoRA by adding low-rank matrices to transformer layers. This enables efficient adaptation with fewer parameters, enhancing domain-specific performance while preserving pretrained features. The loss function used during the initial fine-tuning is a 20:1 weighted combination of focal loss [27] and Dice loss [32]. Details on the preference alignment module and proposal generation are provided in the supplementary material Sec. 6. Dice Similarity Coefficient, Intersection over Union (IoU), and Surface Dice Similarity (SDC) scores are used to evaluate segmentation performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 430, + 492, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 492, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 492, + 443 + ], + "type": "text", + "content": "4.1. Comparison with state-of-the-art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "Quantitative results. Fig. 3 compares the performance of our framework with relevant methods (U-Net [40], nnU-Net [17], SAM [18], SAM-Med2D [11], Self-prompt). The self-prompt method is designed as a variant of [57], excluding the knowledge distillation module. We also design a prompt-only baseline of our framework which is trained on different splits of fully annotated data. Our framework is initially trained with only the prompting module using " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " of data (annotated). As a result, both the prompt-only and final versions exhibit identical performance on this " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " subset. For the final model, the additional data used for training are considered unannotated, as training the alignment mechanism does not use supervision from ground truth. SOTA methods directly use ground truth for the entire data subset. Nevertheless, our method consistently outperforms them in limited data settings " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "(10 - 50\\%)" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " due to the preference alignment mechanism. Our architecture thus reduces the reliance on large datasets, highlighting its lower annotation requirements, which makes it highly effective in low-data regimes. On the Chest-Xray dataset, for instance, with just " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " of the data, our method achieves a Dice score of 78.87, compared to 58.66 (U-Net), 60.97 (nnU-Net), 61.64 (SAM)," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20995" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 219, + 171 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 219, + 171 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 219, + 171 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 219, + 171 + ], + "type": "image", + "image_path": "050dfb3d8688e5f902c8fdcb41ded97f761872112a1c636967ced906c1fba6b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "text", + "content": "Figure 3. Quantitative comparison with SOTA. Dice score (for Chest Xray, Breast USD) and mean Dice score (for AMOS CT) have been shown to measure the model segmentation performance on different proportions of training data " + }, + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "inline_equation", + "content": "(10\\%, 20\\%, 50\\%," + }, + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "inline_equation", + "content": "100\\%)" + }, + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 70, + 395, + 171 + ], + "blocks": [ + { + "bbox": [ + 219, + 70, + 395, + 171 + ], + "lines": [ + { + "bbox": [ + 219, + 70, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 219, + 70, + 395, + 171 + ], + "type": "image", + "image_path": "491cb3ff6193712fc86b0dc8a6ee47923f533c5f8c0cbc52a182c46747ba49f0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 395, + 70, + 554, + 171 + ], + "blocks": [ + { + "bbox": [ + 395, + 70, + 554, + 171 + ], + "lines": [ + { + "bbox": [ + 395, + 70, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 395, + 70, + 554, + 171 + ], + "type": "image", + "image_path": "adf7f09c9307423f79d8a845ac3a4a6fa65799c80aa5665f412048bb55fa6ec1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 213, + 531, + 340 + ], + "blocks": [ + { + "bbox": [ + 82, + 213, + 531, + 340 + ], + "lines": [ + { + "bbox": [ + 82, + 213, + 531, + 340 + ], + "spans": [ + { + "bbox": [ + 82, + 213, + 531, + 340 + ], + "type": "image", + "image_path": "ab375ac6a1e9a03b9d80d0a0c9216ea07abf42b6ebe17e04345333489888f38d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons were made between the segmentation results of nnUnet, SAM-Med2D, and our framework on 2D datasets. BiomedCLIP-based saliency maps are also depicted. Experiments were conducted in " + }, + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "type": "text", + "content": " data settings." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": "67.81 (SAM-Med2D), and 68.41 (Self-prompt). Both U-Net and nnU-Net are known to be data-hungry models, struggling with smaller data subsets. Our usage of visual and textual prompts, derived through BiomedCLIP and VLM models, offers superior signals to those used in SAM and SAM-Med2D. Unlike these SAM variants, Self-Prompt SAM does not require expert-supplied prompts during inference; it instead generates prompts from the output masks in each iteration. This method offers a slight improvement " + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "inline_equation", + "content": "(+1\\%)" + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": " over SAM-Med2D. Our method shows more stable performance gains across different data subsets, while other SOTA methods exhibit steeper improvements. In the " + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": " data setting, our semi-supervised framework achieves a Dice score of 89.68, compared to 91.42 for the supervised prompt-only version. Similar result trends are observed in the breast ultrasound dataset. Notably, the performance jump of our method from " + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": " data is much more pronounced than from " + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": ". This can be due to the nature of the dataset, as the model requires more data to achieve precise and accurate segmentation of breast tumors. We also evaluated our method on a 3D abdominal organ segmentation dataset, including organs such as the liver, kidneys, spleen, pancreas, aorta, bladder, etc (see Fig. 3). Our method outperforms the SOTA across all data proportions except for the full dataset. With " + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 392, + 296, + 704 + ], + "type": "text", + "content": " of the data, our method achieves a mean Dice score of 77.69, surpassing U" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 392, + 556, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 441 + ], + "type": "text", + "content": "Net (59.35), nnU-Net (65.21), SAM (64.93), SAM-Med2D (66.57), SAMMed3D (72.54), and Self-prompt (71.83). At the " + }, + { + "bbox": [ + 313, + 392, + 556, + 441 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 392, + 556, + 441 + ], + "type": "text", + "content": " data setting, it reaches a Dice score of 85.70, comparable to the 86.36 achieved by the prompt-only version." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 452, + 542, + 563 + ], + "blocks": [ + { + "bbox": [ + 328, + 452, + 542, + 563 + ], + "lines": [ + { + "bbox": [ + 328, + 452, + 542, + 563 + ], + "spans": [ + { + "bbox": [ + 328, + 452, + 542, + 563 + ], + "type": "image", + "image_path": "769e6a3b7bc6d232524c11ee4463ef95231d2e5eeded689edb65501c78e7e839.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 570, + 555, + 592 + ], + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 592 + ], + "type": "text", + "content": "Figure 5. Segmentation maps of different anatomical structures (liver and spleen) for SAM-Med3D and our method" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Qualitative results. We present segmentation maps for 3 datasets in Fig. 4 and Fig. 5, generated by our model trained at " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " data settings. The saliency maps could correctly highlight the target regions and proved to be an excellent source of supervision. It can be noted from Fig. 4 that our segmentation quality around the boundaries of tumor or lung is much superior compared to both nnUnet and SAM-Med2D. In Fig. 5, SAM-Med3D tended to under-segment the edges of both the abdominal organs, the liver," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20996" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 69, + 504, + 159 + ], + "blocks": [ + { + "bbox": [ + 108, + 69, + 504, + 159 + ], + "lines": [ + { + "bbox": [ + 108, + 69, + 504, + 159 + ], + "spans": [ + { + "bbox": [ + 108, + 69, + 504, + 159 + ], + "type": "table", + "html": "
SupervisionMethodsChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedOurs74.3078.8764.3575.8877.6978.34
20%- Alignment75.0279.1367.5181.3879.2080.66
10%- Alignment68.4375.6061.4473.6274.7776.06
- Alignment - VQA66.9073.3557.6070.5374.0875.41
- Alignment - VQA - GPT463.1672.7654.2668.8973.1674.89
- Alignment - VQA - CAM50.3857.0245.1659.0569.9771.45
", + "image_path": "631aac687fca7e61ed569ba628d958d8da741c501dcba7d13e092bf33879b0cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 108, + 187, + 504, + 269 + ], + "blocks": [ + { + "bbox": [ + 167, + 166, + 442, + 177 + ], + "lines": [ + { + "bbox": [ + 167, + 166, + 442, + 177 + ], + "spans": [ + { + "bbox": [ + 167, + 166, + 442, + 177 + ], + "type": "text", + "content": "Table 1. Ablation results demonstrating effectiveness of major components." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 187, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 108, + 187, + 504, + 269 + ], + "spans": [ + { + "bbox": [ + 108, + 187, + 504, + 269 + ], + "type": "table", + "html": "
SupervisionPreference AlignmentChest-XrayBreast-USDAMOS-CT
IoUDiceIoUDicemDicemSDC
10% + 10% unannotatedLoss for best candidate70.3777.0961.7673.8175.0176.36
Rating73.9978.4163.9775.5277.2377.98
Ranking74.3078.8764.3575.8877.6978.34
10% + 20% unannotatedRanking79.7485.1573.5684.2380.5481.95
10% + 40% unannotatedRanking88.9689.6885.9288.1584.3085.70
", + "image_path": "60c0a3f133cb85d8c15c4423613565fa7baed9e7fa12441b6c92e906756ad6c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 276, + 501, + 288 + ], + "lines": [ + { + "bbox": [ + 108, + 276, + 501, + 288 + ], + "spans": [ + { + "bbox": [ + 108, + 276, + 501, + 288 + ], + "type": "text", + "content": "Table 2. Ablation results on different proportions of training data for 3 types of preference scoring strategies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 308, + 294, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 294, + 322 + ], + "type": "text", + "content": "and spleen. Additional results are in supplementary Sec. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 333, + 154, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 154, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 154, + 344 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "content": "Effectiveness of major components. We conduct several ablation experiments (see Tab. 1) to evaluate the contribution of each module in our architecture. For simplicity, we focus on the Chest-Xray dataset for analysis. Before incorporating the DPO-driven alignment strategy, all baselines were trained with " + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "content": " labeled data. First, we developed a naive baseline (last row), which relies solely on textual answers generated from GPT-4. As expected, it performed poorly, achieving a 57.02 Dice score. A second baseline (second-last row) was designed to use only visual prompts from BiomedCLIP, which provides semantic information. This method significantly benefited from the coarse segmentation masks derived from CLIP saliency maps, improving the Dice score to 72.76. Next, we combined both textual and visual prompts to form a third baseline (-Alignment-VQA), which resulted in a slight improvement " + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "inline_equation", + "content": "(+0.59\\%)" + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "content": " over the CLIP-only baseline. Finally, we integrated VQA into the prompting strategy to obtain answers related to the shape and location of the target regions, completing our fully empowered unsupervised prompting strategy (-Alignment). This model achieved a Dice score of 75.60 with " + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "content": " data and improved to 79.13 with " + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "content": " data. In comparison, with the preference alignment mechanism, fine-tuning the prompting module with " + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "content": " annotated data and training the alignment module on an additional " + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 353, + 295, + 687 + ], + "type": "text", + "content": " unannotated data achieved a Dice score of 78.87. This is on par with the fully supervised prompt-only model, underscoring the effectiveness of our alignment module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Preference Scoring strategies. We conducted several ablation studies (detailed in Tab. 2) to evaluate the effec" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 308, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 464 + ], + "type": "text", + "content": "tiveness of the preference-scoring strategy. One baseline approach involved backpropagating the loss based only on the best candidate, while another compared ranking the candidates rather than simply rating them. The results are presented for different proportions of unannotated data, on top of the foundational fine-tuning of both the prompt encoder and the decoder using " + }, + { + "bbox": [ + 313, + 308, + 555, + 464 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 464 + ], + "type": "text", + "content": " annotated data. At " + }, + { + "bbox": [ + 313, + 308, + 555, + 464 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 464 + ], + "type": "text", + "content": " unannotated data, both the ranking and rating approaches performed similarly (mean Dice scores of 77.69 and 77.23 for AMOS CT, respectively), significantly outperforming the best-candidate-only method (75.01). Additionally, we observed that as the proportion of unannotated data increased, the model's performance improved substantially." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 464, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 524 + ], + "type": "text", + "content": "Robustness to noisy rating. We experimented with introducing noise into the rating mechanism by flipping the ratings of closely ranked candidates to enhance the framework's robustness. Results are in supplementary material Sec. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "We introduce a novel training strategy to enhance SAM representations for semi-supervised medical image segmentation. We extract integrated semantic, location, and shape information from pretrained vision-language models in an unsupervised manner. This information is used as refined prompts for our model. We also implement an optimal policy, inspired by direct preference optimization in language models. This enables human-in-the-loop feedback simulation within a streamlined training framework, without the need for a separate reward function or extensive knowledge from annotators. These modules ensure that our framework outperforms state-of-the-art methods across datasets spanning multiple modalities in low-annotation data scenarios." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20997" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "We sincerely thank everyone from the GE Healthcare Science & Technology Organization for their support and contributions to the project. The primary model development and evaluations were conducted during a summer internship at GEHC. After the internship, Konwer and Prasanna were supported by the following grants during the paper writing and rebuttal process: NIH 1R01CA297843-01, NIH 3R21CA258493-02S1, and 1R21CA258493-01A1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 308 + ], + "type": "text", + "content": "[2] Peri Akiva and Kristin Dana. Towards single stage weakly supervised semantic segmentation. arXiv preprint arXiv:2106.10309, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 309, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 309, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 61, + 309, + 295, + 342 + ], + "type": "text", + "content": "[3] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 343, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 343, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 61, + 343, + 296, + 408 + ], + "type": "text", + "content": "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "type": "text", + "content": "[5] Mario Amrehn, Sven Gaube, Mathias Unberath, Frank Schebesch, Tim Horz, Maddalena Strumia, Stefan Steidl, Markus Kowarschik, and Andreas Maier. Ui-net: Interactive artificial neural networks for iterative image segmentation based on a user model. arXiv preprint arXiv:1709.03450, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 544 + ], + "type": "text", + "content": "[6] Louis Blankemeier, Joseph Paul Cohen, Ashwin Kumar, Dave Van Veen, Syed Jamal Safdar Gardezi, Magdalini Paschali, Zhihong Chen, Jean-Benoit Delbrouck, Eduardo Reis, Cesar Truyts, et al. Merlin: A vision language foundation model for 3d computed tomography. arXiv preprint arXiv:2406.06512, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 545, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 578 + ], + "type": "text", + "content": "[7] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 295, + 622 + ], + "type": "text", + "content": "[8] Samuel Budd, Emma C Robinson, and Bernhard Kainz. A survey on active learning and human-in-the-loop deep learning for medical image analysis. Medical image analysis, 71: 102062, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 689 + ], + "type": "text", + "content": "[9] Michal Byra, Piotr Jarosik, Aleksandra Szubert, Michael Galperin, Haydee Ojeda-Fournier, Linda Olson, Mary O'Boyle, Christopher Comstock, and Michael Andre. Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control, 61:102027, 2020. 6, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Peijie Chen, Qi Li, Saad Biaz, Trung Bui, and Anh Nguyen. gscorecam: What objects is clip looking at? In Proceedings" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "of the Asian Conference on Computer Vision (ACCV), pages 1959-1975, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "text", + "content": "[11] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "type": "text", + "content": "[12] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and coid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 6, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "text", + "content": "[13] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "type": "text", + "content": "[14] Sedigheh Eslami, Christoph Meinel, and Gerard De Melo. Pubmedclip: How much does clip benefit visual question answering in the medical domain? In Findings of the Association for Computational Linguistics: EACL 2023, pages 1181-1193, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 348 + ], + "type": "text", + "content": "[15] Shantanu Ghosh, Clare B Poynton, Shyam Visweswaran, and Kayhan Batmanghelich. Mammo-clip: A vision language foundation model to enhance data efficiency and robustness in mammography. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 632–642. Springer, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[17] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 438, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 492 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 494, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 554, + 557 + ], + "type": "text", + "content": "[19] Taha Koleilat, Hojat Asgariandehkordi, Hassan Rivaz, and Yiming Xiao. Medclip-sam: Bridging text and image towards universal medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 643-653. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 554, + 623 + ], + "type": "text", + "content": "[20] Aishik Konwer, Xuan Xu, Joseph Bae, Chao Chen, and Pratek Prasanna. Temporal context matters: Enhancing single image prediction with disease progression representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18824-18835, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 689 + ], + "type": "text", + "content": "[21] Aishik Konwer, Xiaoling Hu, Joseph Bae, Xuan Xu, Chao Chen, and Prateek Prasanna. Enhancing modality-agnostic representations via meta-learning for brain tumor segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21415–21425, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[22] Philipp Krahenbuhl and Vladlen Koltun. Parameter learning and convergent inference for dense random fields. In Inter-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20998" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "national conference on machine learning, pages 513-521. \nPMLR, 2013. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[23] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[24] Wenhui Lei, Xu Wei, Xiaofan Zhang, Kang Li, and Shaoting Zhang. Medlsam: Localize and segment anything model for 3d medical images. arXiv preprint arXiv:2306.14752, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "text", + "content": "[25] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 326 + ], + "type": "text", + "content": "[27] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[28] Weixiong Lin, Ziheng Zhao, Xiaoman Zhang, Chaoyi Wu, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-clip: Contrastive language-image pre-training using biomedical documents. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 525–536. Springer, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "type": "text", + "content": "[29] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "text", + "content": "[31] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "text", + "content": "[32] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "text", + "content": "[33] Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Eduardo Pontes Reis, and Pranav Rajpurkar. Med-flamingo: a multimodal medical few-shot learner. In Machine Learning for Health (ML4H), pages 353-367. PMLR, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[34] Vishwesh Nath, Dong Yang, Bennett A. Landman, Daguang Xu, and Holger R. Roth. Diminishing uncertainty within the training pool: Active learning for medical image segmentation. IEEE Transactions on Medical Imaging, 40(10): 2534-2547, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "text", + "content": "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "text", + "content": "[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 183, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 249 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[38] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 370 + ], + "type": "text", + "content": "[39] Adrit Rao, Andrea Fisher, Ken Chang, John Christopher Panagides, Katherine McNamara, Joon-Young Lee, and Oliver Aalami. Imil: Interactive medical image learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5241–5250, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 437 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 479 + ], + "type": "text", + "content": "[41] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 547 + ], + "type": "text", + "content": "[42] Yiqing Shen, Jingxing Li, Xinyuan Shao, Blanca Inigo Romillo, Ankush Jindal, David Dreizin, and Mathias Unberath. Fastsam3d: An efficient segment anything model for 3d volumetric medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 542-552. Springer, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "text", + "content": "[43] Shenghuan Sun, Greg Goldgof, Atul Butte, and Ahmed M Alaa. Aligning synthetic medical images with clinical knowledge using human feedback. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 592, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 647 + ], + "type": "text", + "content": "[44] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[45] Guotai Wang, Wenqi Li, Maria A Zuluaga, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sebastien Ourselin, et al. Interactive medical image segmentation using deep learning with image-specific fine tuning. IEEE transactions on medical imaging, 37(7): 1562-1573, 2018. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20999" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[46] Guotai Wang, Maria A Zuluaga, Wenqi Li, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, Sébastien Ourselin, et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 41(7):1559-1572, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 173 + ], + "type": "text", + "content": "[47] H Wang et al. Sam-med3d: towards general-purpose segmentation models for volumetric medical images. Preprint at https://arxiv.org/abs/2310.15161, 2024. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "type": "text", + "content": "[48] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 294, + 251 + ], + "type": "text", + "content": "[49] Rhydian Windsor, Amir Jamaludin, Timor Kadir, and Andrew Zisserman. Vision-language modelling for radiological imaging and reports in the low data regime. arXiv preprint arXiv:2303.17644, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 254, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 294, + 286 + ], + "type": "text", + "content": "[50] Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology. arXiv preprint arXiv:2308.02463, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 288, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 288, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 57, + 288, + 294, + 342 + ], + "type": "text", + "content": "[51] Chaoyi Wu, Weixiong Lin, Xiaoman Zhang, Ya Zhang, Weidi Xie, and Yanfeng Wang. Pmc-llama: toward building open-source language models for medicine. Journal of the American Medical Informatics Association, page ocae045, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 344, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 344, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 344, + 294, + 399 + ], + "type": "text", + "content": "[52] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 400, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 400, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 57, + 400, + 294, + 464 + ], + "type": "text", + "content": "[53] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems, 36:59008-59033, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 468, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 468, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 468, + 294, + 533 + ], + "type": "text", + "content": "[54] Kihyun You, Jawook Gu, Jiyeon Ham, Beomhee Park, Jiho Kim, Eun K Hong, Woonhyuk Baek, and Byungseok Roh. Cxr-clip: Toward large scale chest x-ray language-image pre-training. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 101-111. Springer, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 535, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 535, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 294, + 599 + ], + "type": "text", + "content": "[55] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 645 + ], + "type": "text", + "content": "[56] Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[57] Chunpeng Zhou, Kangjie Ning, Qianqian Shen, Sheng Zhou, Zhi Yu, and Haishuai Wang. Sam-sp: Self-prompting makes sam great again. arXiv preprint arXiv:2408.12364, 2024. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "type": "text", + "content": "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "type": "text", + "content": "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21000" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_content_list.json b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6576de3ee4b2bbf11122beeb1794c0b3e778c5be --- /dev/null +++ b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_content_list.json @@ -0,0 +1,1642 @@ +[ + { + "type": "text", + "text": "Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild", + "text_level": 1, + "bbox": [ + 250, + 130, + 746, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Liu $^{1}$ , Yufei Chen $^{1*}$ , Xiaodong Yue $^{2,3}$ $^{1}$ School of Computer Science and Technology, Tongji University, Shanghai, China \n $^{2}$ Artificial Intelligence Institute, Shanghai University, Shanghai, China \n $^{3}$ VLN Lab, Shanghai, China \nldachuan@outlook.com, yufeichen@tongji.edu.cn, yswantfly@shu.edu.cn", + "bbox": [ + 174, + 202, + 823, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trusted multi-view classification (TMVC) addresses variations in data quality by evaluating the reliability of each view based on prediction uncertainty at the evidence level, reducing the impact of low-quality views commonly encountered in real-world scenarios. However, existing TMVC methods often struggle to maintain robustness during testing, particularly when integrating noisy or corrupted views. This limitation arises because the evidence collected by TMVC may be unreliable, frequently providing incorrect information due to complex view distributions and optimization challenges, ultimately leading to classification performance degradation. To enhance the robustness of TMVC methods in real-world conditions, we propose a generalized evidence filtering mechanism that is compatible with various fusion strategies commonly used in TMVC, including Belief Constraint Fusion, Aleatory Cumulative Belief Fusion, and Averaging Belief Fusion. Specifically, we frame the identification of unreliable evidence as a multiple testing problem and introduce $p$ -values to control the risk of false identification. By selectively down-weighting unreliable evidence during testing, our mechanism ensures robust fusion and mitigates performance degradation. Both theoretical guarantees and empirical results demonstrate significant improvements in the classification performance of TMVC methods, supporting their reliable application in challenging, real-world environments.", + "bbox": [ + 86, + 357, + 485, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 777, + 220, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-view classification (MVC) enhances the accuracy of deep learning models by integrating information from multiple data views. By combining insights from diverse views, MVC improves decision-making in complex domains [7, 10, 11, 14, 15, 25, 31]. However, traditional", + "bbox": [ + 89, + 803, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f7ebd7586ae13d0d54a2da1a914d155d44851fd2267499c93d4a7a00c9eb0057.jpg", + "image_caption": [ + "Figure 1. Classification accuracy degradation with noise (w/) compared to no noise (w/o) across three fusion strategies (BCF, A-CBF, and ABF) on six datasets. Noise is introduced by replacing half of the views to simulate data collected from malfunctioning sensors." + ], + "image_footnote": [], + "bbox": [ + 516, + 325, + 906, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MVC approaches often assume that each view consistently provides high-quality, reliable information—an assumption that is frequently unrealistic in real-world applications. In practical scenarios, data quality and reliability across views can vary significantly, leading to contradictory or misleading information that compromises prediction accuracy and reliability. As a result, conventional MVC models may struggle to reliably identify the ground truth, particularly when dealing with noisy, corrupted, or otherwise low-quality data views [38].", + "bbox": [ + 509, + 623, + 906, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, trusted multi-view classification (TMVC) [8, 9, 16, 17, 32] has emerged as an effective solution. Unlike traditional MVC, TMVC evaluates the trustworthiness of each view by leveraging prediction uncertainty derived from view-specific evidence, enabling it to identify unreliable views and prioritize informative ones in the final predictions. This approach has proven highly effective in critical applications such as medical imaging diagno", + "bbox": [ + 509, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15508", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sis [29] and autonomous driving [21], where distinguishing between reliable and unreliable predictions is essential.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In TMVC, evidence from each sample for each view contributes to its prediction and is also used to calculate a view-specific uncertainty score $u$ , which reflects the confidence in the quality of that view for the given sample. A high uncertainty score (i.e., $u = 1$ ) indicates a substantial deviation from the training data, potentially due to noise or corruption. In such cases, evidence supporting all classes is minimized and fused using one of three strategies: Belief Constraint Fusion (BCF) [8, 9], Aleatory Cumulative Belief Fusion (A-CBF) [16-18, 37], and Averaging Belief Fusion (ABF) [32]. However, despite their success in controlled settings, these methods often struggle to remain robust in the wild. Complex and diverse data view distributions, combined with varying levels of optimization, can distort uncertainty estimates, resulting in unreliable evidence that misleads the model and biases predictions toward incorrect classes. As shown in Figure 1, introducing random noise (e.g., replacing half of the views in a dataset) to simulate sensor failure conditions leads to a significant drop in classification accuracy across all fusion strategies on six real-world datasets, underscoring the risks of decision-making in unpredictable and noisy environments.", + "bbox": [ + 91, + 123, + 483, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose a generalized and robust testing-time evidence filtering mechanism to enhance the robustness of TMVC methods in real-world, unpredictable environments. Specifically, we frame the identification of unreliable evidence as a multiple testing problem, introducing $p$ -values to control the risk of false identification within a small margin. This approach enables effective post-hoc adjustments during testing. By selectively down-weighting evidence from unreliable views, our mechanism enhances fusion robustness and mitigates performance degradation caused by low-quality data sources. In summary, our contributions are as follows:", + "bbox": [ + 89, + 455, + 482, + 636 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a testing-time evidence filtering mechanism that is compatible with multiple fusion strategies, enhancing the robustness of existing TMVC methods.", + "2. Our approach frames the identification of unreliable evidence as a multiple testing problem and selectively adjusts evidence weights during testing, mitigating the negative impact of unreliable views on fusion.", + "3. Theoretical and empirical analyses validate the effectiveness of the proposed method in improving both classification accuracy and robustness." + ], + "bbox": [ + 89, + 638, + 482, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 806, + 240, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-view Classification", + "text_level": 1, + "bbox": [ + 89, + 832, + 316, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Over the past few years, numerous multi-view classification (MVC) approaches [1, 2, 16, 22, 25, 26, 30, 34, 36] have been developed, particularly those leveraging deep neural", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "networks (DNNs). These methods aim to learn shared representations by integrating information from different types of views, resulting in improved performance. By combining insights from diverse views, MVC has achieved remarkable gains and has been widely applied across various tasks [23, 35]. However, in real-world scenarios, the quality of data across views can vary significantly, introducing noise and inconsistencies that undermine both classification accuracy and reliability—factors that are especially critical in areas like medical diagnosis and autonomous driving.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Trusted Multi-view Classification", + "text_level": 1, + "bbox": [ + 511, + 255, + 805, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, Trusted Multi-View Classification (TMVC) has emerged as a framework that assesses the reliability of each view through prediction uncertainty, allowing models to prioritize reliable views and down-weight unreliable ones [8, 9, 16, 17, 32, 33]. TMVC methods often rely on Subjective Logic (SL) [12] to model view-specific predictions as multinomial opinions, which are then fused using belief combination rules tailored to the context [12]. For instance, extensions of Dempster-Shafer theory [24] were firstly introduced by [8, 9] to facilitate evidence fusion in multi-view settings. Other fusion approaches, such as Cumulative Belief Fusion [16, 17] and Averaging Belief Fusion [32], have also been proposed to further enhance the effectiveness of TMVC. Despite their success, all these methods fall short of achieving robust performance during testing, as the magnitude of estimated uncertainty is highly sensitive to the diverse distributions of data views and varying levels of model optimization, which significantly degrades classification accuracy in real-world scenarios. In contrast, our method improves testing-time robustness through an evidence filtering mechanism, thereby reducing the risks associated with final decision-making.", + "bbox": [ + 511, + 279, + 906, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Overview and Analysis of TMVC", + "text_level": 1, + "bbox": [ + 511, + 628, + 816, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Generalized Framework of TMVC", + "text_level": 1, + "bbox": [ + 511, + 655, + 816, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the training data $D_{train} = \\left\\{\\left\\{\\boldsymbol{x}_i^v\\right\\}_{v=1}^V, \\boldsymbol{y}_i\\right\\}_{i=1}^N$ , where $N$ is the number of samples, each sample has $V$ views with a one-hot label $\\boldsymbol{y}_i = \\left\\{y_{ik}\\right\\}_{k=1}^K$ for a $K$ -class classification task. TMVC uses subjective logic [12] to learn a function $f$ that maps $\\left\\{\\boldsymbol{x}_i^v\\right\\}_{v=1}^V$ to $\\boldsymbol{y}_i$ along with a corresponding overall uncertainty score $u_i$ , which quantifies prediction confidence at the evidence level.", + "bbox": [ + 511, + 674, + 906, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, let $f_{v}(\\boldsymbol{x}_{i}^{v})$ denote the output of the penultimate (logits) layer of the view-specific neural network for sample $\\boldsymbol{x}_{i}^{v}$ in a single view. Instead of the standard softmax function for a single categorical prediction, TMVC employs an alternative activation function $a(\\cdot)$ (e.g., ReLU) to capture non-negative evidence $e_{i}^{v} = a(f_{v}(\\boldsymbol{x}_{i}^{v}))$ , where $e_{i}^{v} = \\{e_{ik}^{v}\\}_{k=1}^{K}$ . The view-specific prediction uncertainty is", + "bbox": [ + 511, + 792, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15509", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "given by", + "bbox": [ + 89, + 92, + 151, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu _ {i} ^ {v} = K / S _ {i} ^ {v}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 117, + 482, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S_{i}^{v} = \\sum_{k=1}^{K} e_{ik}^{v} + K$ . This reflects the lack of total evidence across all classes in a single view.", + "bbox": [ + 89, + 146, + 482, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TMVC then applies evidence fusion strategies, such as BCF, A-CBF, and ABF, to integrate the evidence from each view into a unified multi-view evidence representation $e_i = \\{e_{ik}\\}_{k=1}^K$ . These three fusion rules for two views are commonly expressed as follows [20]:", + "bbox": [ + 89, + 193, + 483, + 268 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} A - C B F \\left(\\boldsymbol {e} _ {i} ^ {1}, \\boldsymbol {e} _ {i} ^ {2}\\right) = \\boldsymbol {e} _ {i} ^ {1} + \\boldsymbol {e} _ {i} ^ {2}. \\\\ A B F \\left(e _ {i} ^ {1}, e _ {i} ^ {2}\\right) = \\left(e _ {i} ^ {1} + e _ {i} ^ {2}\\right) / 2. \\\\ B C F \\left(\\boldsymbol {e} _ {i} ^ {1}, \\boldsymbol {e} _ {i} ^ {2}\\right) = \\boldsymbol {e} _ {i} ^ {1} + \\boldsymbol {e} _ {i} ^ {2} + \\boldsymbol {e} _ {i} ^ {1} \\boldsymbol {e} _ {i} ^ {2} / K. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 277, + 480, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These fusion rules can also be extended to handle multiple views with ease.", + "bbox": [ + 89, + 347, + 482, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the training phase, TMVC links the generated evidence to the concentration parameter of the Dirichlet distribution and optimizes model parameters using a task-specific loss function. This loss primarily consists of a cross-entropy term, which maximizes the evidence for the correct label of each sample, and a Kullback-Leibler (KL)-divergence term, which drives the evidence for incorrect labels toward zero, as described in [8]. The final multi-view prediction $\\pmb{p}_i = \\{p_{ik}\\}_{k=1}^K$ is obtained as the expectation of the learned Dirichlet distribution:", + "bbox": [ + 89, + 378, + 483, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i k} = \\left(e _ {i k} + 1\\right) / S _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 540, + 482, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S_{i} = \\sum_{k = 1}^{K}e_{ik} + K$ , and the overall uncertainty $u_{i}$ is given by $u_{i} = K / S_{i}$ . Finally, the model's best performance on the validation dataset $D_{val}$ is used to select the optimal parameters for testing.", + "bbox": [ + 89, + 568, + 483, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Limitations of TMVC at Testing Time", + "text_level": 1, + "bbox": [ + 89, + 655, + 418, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite its success across various domains [21, 29], TMVC methods face significant limitations at testing time, particularly in real-world scenarios with diverse data view distributions and varying levels of model optimization. These factors often degrade TMVC's robustness, especially when handling low-quality data views.", + "bbox": [ + 89, + 676, + 482, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given a multi-view test dataset $D_{test} = \\left\\{\\left\\{\\boldsymbol{x}_j^v\\right\\}_{v=1}^V\\right\\}_{j=1}^M$ , a trained TMVC model with $V$ view-specific networks provides uncertainty estimates for each view of a testing sample $\\boldsymbol{x}_j^v$ . Ideally, a high uncertainty score (e.g., $u_j^v = 1$ ) for a low-quality view would reduce its evidence contribution across all classes to zero ( $e_j^v = 0$ ), minimizing its impact during fusion. However, these uncertainty estimates are sensitive to data distributions and", + "bbox": [ + 89, + 768, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e8f495d20d6f49ae6ff828e7a9737257e9e39f68115c8c8f25415bd8c9ebbef1.jpg", + "image_caption": [ + "Figure 2. Density distribution of uncertainty for noisy view results under noise corruption and the corresponding density distribution of remaining unreliable evidence across three fusion strategies on the CUB dataset." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 714, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3ac0fa0a6cbd09095427c090e1b652a17dfd50ffe197cbe7e7ba6a32a69a0d0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 90, + 903, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model optimization levels, making it challenging to reliably achieve $u_{j}^{v} = 1$ for extremely noisy or corrupted views.", + "bbox": [ + 511, + 329, + 903, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For example, in the CUB dataset, even when one view is entirely corrupted by noise (i.e., features of one view are replaced by random noise), the uncertainty score for that view typically falls within the range [0.7, 1.0], leaving residual evidence mostly between [0.0, 4.0] (as illustrated in Figure 2) that contributes to fusion and can potentially bias predictions. Since this view lacks useful information, its residual evidence is unreliable and negatively impacts multi-view classification accuracy, as shown in Figure 1. Moreover, the overall uncertainty score derived from fusion strategies like BCF and A-CBF is often lower as additional views are fused (as demonstrated in [9, 16]). Even with ABF fusion, the overall uncertainty remains below the highest view-specific uncertainty [32]. As a result, decision-makers may struggle to detect potentially misclassified samples caused by low-quality views. These limitations underscore the challenges TMVC faces in real-world conditions, where testing-time robustness is crucial for reliable decision-making.", + "bbox": [ + 511, + 359, + 906, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 662, + 687, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present a strategy to enhance the robustness of TMVC methods at testing time. Performance degradation in TMVC methods often arises from the negative impact of such low-quality views, which typically exhibit high prediction uncertainty during testing, compared to the reliable views seen during training. An intuitive solution to this problem is to reduce the contribution of unreliable evidence from low-quality views during multi-view fusion. To address this, we propose a generalized evidence filtering mechanism that is compatible with various TMVC fusion rules. The mechanism consists of two key components: (1) identifying unreliable evidence collected from low-quality views, and (2) assigning lower weights to minimize their impact on the final decision.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15510", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Reliable Evidence Discovery in the Wild", + "text_level": 1, + "bbox": [ + 89, + 90, + 434, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Identifying unreliable evidence is equivalent to discovering reliable evidence. The primary challenge in this process is controlling the false discovery rate (FDR) [3], since the true label is unknown. We address this challenge by framing reliable evidence discovery as a multiple hypothesis testing problem from a statistical learning perspective.", + "bbox": [ + 89, + 112, + 483, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a multi-view test dataset $D_{test}$ , the objective is to identify as many reliable pieces of evidence as possible while ensuring that the FDR is maintained below a user-defined threshold $\\alpha \\in (0,1)$ . Let us define a discovery function $\\Lambda(\\kappa^v)$ , which selects a subset $S^v$ of view-specific evidence for the test points of interest. Evidences with smaller values of $u_j^v$ , i.e., $u_j^v \\leq \\kappa^v$ , is likely to be included in the subset, where $\\kappa^v$ is a threshold determined by the discovery function. The FDR for this function is then defined as:", + "bbox": [ + 89, + 203, + 483, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F D R} \\left(\\Lambda \\left(\\kappa^ {v}\\right)\\right) := \\mathbb {E} \\left[ \\frac {\\sum_ {j \\in D _ {t e s t}} \\mathbb {1} \\left\\{j \\in S ^ {v} , u _ {j} ^ {v} > \\kappa^ {v} \\right\\}}{\\max \\left\\{1 , | S ^ {v} | \\right\\}} \\right]. (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 345, + 483, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to ensure that the subset $S^v$ identified by $\\Lambda(\\kappa^v)$ satisfies the following error-control property:", + "bbox": [ + 89, + 396, + 483, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 (Risk-controlling discovery). Let $\\kappa^v$ be a random variable. We define $\\Lambda(\\kappa^v)$ as an $(\\alpha, \\delta)$ -risk-controlling discovery (RCD) function if", + "bbox": [ + 89, + 435, + 483, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left(\\operatorname {F D R} \\left(\\Lambda \\left(\\kappa^ {v}\\right)\\right) \\leq \\alpha\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 491, + 480, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "holds with probability at least $1 - \\delta$", + "bbox": [ + 89, + 518, + 318, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\delta \\in (0,1)$ is the confidence level. To meet this risk control criterion, we need to identify a suitable $\\Lambda (\\kappa^v)$ that adapts to diverse data view distributions encountered in the wild. For this purpose, we gather statistical evidence via hypothesis testing, where we consider the view-specific evidence of a test sample to be discovered (i.e, $u_{j}^{v}\\leq \\kappa^{v}$ ) by $\\Lambda (\\kappa^v)$ as the null hypothesis. Rejecting the null hypothesis reflects that $u_{j}^{v} > \\kappa^{v}$ , and therefore the evidence is unreliable.", + "bbox": [ + 89, + 542, + 483, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under this framework, discovering reliable evidence boils down to simultaneously testing the $M$ hypotheses. To do this, for any arbitrary view-specific evidence $\\pmb{e}_j^v$ , we compute the $p$ -value:", + "bbox": [ + 89, + 679, + 483, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\text {- v a l u e} \\left(\\boldsymbol {e} _ {j} ^ {v}\\right) = \\frac {1 + \\sum_ {l \\in D _ {v a l}} \\mathbb {1} \\left\\{u _ {j} ^ {v} \\geq u _ {l} ^ {v} \\right\\}}{\\left| D _ {v a l} \\right| + 1}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 750, + 483, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is clear that the $p$ -value quantifies how representative the evidence $e_j^v$ is of reliable evidence. If the $p$ -value is less than $\\alpha$ , the null hypothesis of exchangeability with the validation data is rejected, deeming the evidence unreliable. We obtain marginally valid $p$ -values, which satisfy:", + "bbox": [ + 89, + 792, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left(p - \\text {v a l u e} \\left(\\boldsymbol {e} _ {j} ^ {v}\\right) \\leq \\alpha\\right) \\leq \\alpha , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 878, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "indicating that the probability of falsely identifying unreliable evidence is controlled by the small margin $\\alpha$ . To achieve this, we set a view-specific, distribution-free threshold using the Benjamini-Hochberg (BH) procedure [3]:", + "bbox": [ + 511, + 90, + 905, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa^ {v} = Q \\left(\\left\\{u _ {l} ^ {v} \\right\\} _ {l \\in D _ {v a l}}, \\alpha\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 162, + 906, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q(\\cdot, \\alpha)$ denotes the $\\lceil (1 - \\alpha)(|D_{val}| + 1) \\rceil$ -th smallest value in $\\{u_l^v\\}_{l \\in D_{val}}$ . Then, we construct the reliable evidence set using $\\Lambda(\\kappa^v)$ :", + "bbox": [ + 511, + 190, + 905, + 237 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {v} = \\left\\{j \\in D _ {\\text {t e s t}}: u _ {j} ^ {v} \\leq \\kappa^ {v} \\right\\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 247, + 906, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we have the following proposition:", + "bbox": [ + 511, + 276, + 799, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1. $\\Lambda (\\kappa^v)$ is an $(\\alpha ,\\delta)$ -risk-controlling discovery (RCD) function.", + "bbox": [ + 511, + 300, + 905, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Mitigating the Influence of Unreliable Evidence", + "text_level": 1, + "bbox": [ + 511, + 340, + 905, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After identifying the reliability of view-specific evidence, we aim to mitigate the impact of unreliable evidence during multi-view fusion. To achieve this, we define the filtered evidence for each testing sample as follows:", + "bbox": [ + 511, + 363, + 905, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\boldsymbol {e}} _ {j} ^ {v} = \\left\\{ \\begin{array}{c c} \\boldsymbol {e} _ {j} ^ {v}, & j \\in S ^ {v}. \\\\ \\alpha \\cdot \\boldsymbol {e} _ {j} ^ {v}, & j \\notin S ^ {v}. \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 434, + 905, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\alpha$ is a user-defined parameter that controls the weight assigned to unreliable evidence, directly influencing the FDR control. Smaller values of $\\alpha$ effectively reduce the influence of low-quality views, minimizing the risk of unreliable evidence negatively impacting the final prediction. This tight control is especially valuable in high-stakes applications where high confidence and accuracy in predictions is crucial. On the other hand, if $\\alpha$ is set to a larger value, the mechanism retains more influence from low-quality views, which may be beneficial when such views contain partially valuable information despite high uncertainty.", + "bbox": [ + 511, + 477, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, the choice of $\\alpha$ balances FDR control and information retention, allowing flexibility in managing the contribution of uncertain views to the fusion process. A lower $\\alpha$ tightens FDR control, reducing the likelihood of unreliable evidence being included and enhancing prediction reliability. Importantly, we do not fully exclude unreliable evidence, as even noisy views may contain useful information for classification. Instead, we adopt a weighted filtering approach that retains the potential benefits of low-quality views while mitigating their negative effects. This soft filtering strategy allows the model to remain adaptive and robust under real-world conditions.", + "bbox": [ + 511, + 643, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After filtering, the view-specific evidence is fused using the selected fusion rules (e.g., BCF, A-CBF, or ABF), achieving an adaptable and robust multi-view fusion process that is compatible with various TMVC fusion strategies. This leads to the following propositions:", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 2. Applying evidence filtering improves the classification performance of TMVC fusion strategies (e.g., BCF, A-CBF, or ABF), under the assumption that unreliable evidence provides incorrect information in the wild.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 3. The overall uncertainty of multi-view results generated by TMVC fusion strategies with evidence filtering for unreliable evidence will exhibit greater uncertainty than fusion without evidence filtering.", + "bbox": [ + 86, + 157, + 482, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both Propositions 2 and 3 guarantee the robustness of our method when confronted with unreliable evidence (i.e., evidence from low-quality views) in the wild. Note that unreliable evidence providing correct predictions by chance (a counterintuitive situation) may hurt classification performance in our method, particularly when evidence from other views is incorrect. To address this, our method exhibits greater overall uncertainty for these multi-view results, as shown in Proposition 3, which can alert humans to be cautious with these results.", + "bbox": [ + 89, + 226, + 483, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, our evidence filtering mechanism improves testing-time robustness by adjusting the weights of view-specific evidence based on reliability. This enables TMVC methods to perform reliably in environments with data inconsistencies. This testing-time robustness is crucial for dependable deployment in high-stakes, real-world applications.", + "bbox": [ + 89, + 377, + 483, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Computational Complexity Analysis", + "text_level": 1, + "bbox": [ + 89, + 489, + 405, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method introduces a generalized framework that can be applied to various DNNs for multi-view testing. Given a single test pass with complexity $O(T_{test})$ on an arbitrary TMVC model, the overall computational complexity is:", + "bbox": [ + 89, + 513, + 483, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO (L \\log L) + O (M \\cdot T _ {t e s t}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 595, + 480, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L$ and $M$ represent the sizes of the validation and test datasets, respectively. The term $O(L\\log L)$ reflects the complexity involved in threshold calculation, primarily due to sorting the uncertainty scores within the validation set to determine an appropriate threshold for filtering unreliable evidence. The term $O(M\\cdot T_{test})$ represents the complexity of running one test pass on $M$ samples in the testing set. In most practical scenarios, the term $O(M\\cdot T_{test})$ will typically dominate, especially for large-scale DNNs, unless the validation set is extremely large or the neural networks are shallow and computationally lightweight.", + "bbox": [ + 89, + 621, + 483, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 800, + 223, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we validate the effectiveness of our proposed method on six real-world datasets. Each experiment is run 10 times, and we report the mean and standard deviation of accuracy (ACC) to ensure reliability. The extensive results demonstrate the clear advantage of our proposed method.", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 90, + 702, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 113, + 619, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on six real-world multi-view datasets as follows: CUB [28]: This dataset contains 11,788 bird images from 200 distinct categories, each paired with text descriptions. For our experiments, we utilize the first 10 categories. Image features are extracted using GoogleNet, while text features are extracted using doc2vec. SCENE [5]: This dataset comprises 4,485 images from 15 indoor and outdoor scene categories. Multiple views are represented using three feature types: GIST, PHOG, and LBP. CAL [6]: The CAL dataset consists of 8,677 images from 101 categories. We focus on the first 20 categories and extract six types of features: GABOR, WAVELET-MOMENTS, CENHIST, HOG, GIST, and LBP. HAND [27]: This dataset contains 2,000 samples across 10 digit classes (from '0' to '9'), with 200 samples per class. Six feature descriptors (PIX, FOU, FAC, ZER, KAR, MOR) are extracted as multiple views. PIE: This dataset consists of 680 facial images from 68 individuals. Three types of features are extracted: intensity, LBP, and Gabor. NUS: A subset of the NUS-WIDE dataset [4], it contains 30,000 images spanning 31 object categories, represented in five views, including a 64-D color histogram, 225-D blockwise color moments, 144-D color correlogram, 73-D edge direction histogram, and 128-D wavelet texture.", + "bbox": [ + 511, + 132, + 906, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.2. Compared Methods", + "text_level": 1, + "bbox": [ + 513, + 503, + 697, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our method is a generalized framework applicable to various trusted multi-view classification methods, we primarily evaluate its effectiveness by comparing baseline methods with and without our approach in these experiments. We apply our method to three fusion rules: BCF, A-CBF, and ABF, using the general loss function proposed in the work of [8]. We then integrate our method with state-of-the-art multi-view deep classification methods, including TMC [8], TMDOA [16], ETMC [9], SMDC [17], ECML [32], and CCML [19]. These methods use the same neural network architecture as ours, ensuring a consistent basis for comparison.", + "bbox": [ + 511, + 523, + 906, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.3. Implementations", + "text_level": 1, + "bbox": [ + 513, + 714, + 678, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For all datasets, we split the data into training (70%), validation (10%), and testing (20%) sets. Fully connected networks were used as view-specific backbones. To obtain a trained TMVC model, we employed the Adam optimizer [13] with an $l_{2}$ -norm regularization of $1e^{-5}$ and a learning rate of $1e^{-3}$ . During testing, we set $\\alpha = 0.05$ by default, ensuring an FDR below 5% unless otherwise specified. The model was implemented in PyTorch and run on a GeForce RTX 4090 GPU with 24GB of memory.", + "bbox": [ + 511, + 733, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note: Optimization during training is not a primary focus of this work; hence, finding the optimal model parameters", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15512", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3af809e3e41194dce520df021a0df0816f9d3fcd08ee5e205c8e748b497f59a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Fusion with Clean Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB98.75±0.4298.75±0.4297.50±1.6798.33±1.6797.08±0.4297.92±0.42
SCENE70.94±0.0670.69±0.3171.63±0.7571.31±1.0670.50±0.1370.25±0.25
CAL98.50±0.5098.38±0.3897.63±0.3897.25±0.2598.75±0.5098.50±0.25
HAND97.88±0.1398.25±0.2599.00±0.0599.00±0.7598.88±0.3899.25±0.50
PIE98.08±1.1598.46±0.7798.85±0.3898.85±0.3897.69±1.5498.08±1.15
NUS53.85±1.0054.05±0.8055.35±0.3055.43±0.0857.75±0.7557.88±0.78
Fusion with Corrupted Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB64.17±5.0080.00±0.8356.67±6.6779.17±1.6765.33±3.3383.75±1.25
SCENE46.00±3.3853.38±2.3150.75±2.0051.38±1.9449.25±0.6352.75±0.75
CAL56.75±1.8876.75±5.0070.25±0.7571.50±2.0073.25±0.1375.50±2.00
HAND64.50±1.6370.00±1.8870.00±0.7570.75±0.6375.75±1.7578.25±1.75
PIE89.77±0.0196.54±0.3888.46±3.8590.77±2.3188.08±0.3891.54±0.01
NUS27.70±1.6037.70±0.8536.73±1.2238.95±1.2037.00±1.6040.63±1.38
", + "bbox": [ + 101, + 88, + 893, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Classification performance (ACC (%)) for fusion with clean views and corrupted views, based on three fusion strategies, with or without our method.", + "bbox": [ + 89, + 357, + 906, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "was not the primary goal. Instead, we provide comparative results with and without our method under identical settings (same random seed, dataset splits, optimization parameters, etc.) to evaluate the effectiveness of our method.", + "bbox": [ + 89, + 412, + 480, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 481, + 294, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1. Effect of Our Method on Basic Fusion Strategies", + "text_level": 1, + "bbox": [ + 89, + 503, + 475, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this evaluation, we assess the effectiveness of our proposed method across three fusion strategies (ABF, A-CBF, and BCF) under both clean and corrupted view conditions. Initially, we applied our method to these fusion strategies in the context of clean views, where no noise was introduced. As shown in Table 1, our method generally improves classification accuracy across most datasets, even in noise-free environments. This demonstrates that our approach enhances model stability in both noisy and normal conditions. Minor performance decreases observed in the SCENE and CAL datasets with clean views may be attributed to our method's down-weighting of higher uncertainty evidence for some samples in certain views. These higher uncertainty samples could otherwise positively impact classification in some cases.", + "bbox": [ + 88, + 523, + 482, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Under corrupted view conditions, where random noise in the range of $[-5, 5]$ is added to half of the views, our method significantly boosts accuracy across all datasets. For example, on the CUB dataset with A-CBF fusion, our method achieves nearly a $23\\%$ improvement. For datasets like HAND, the modest performance improvement under corrupted conditions can be explained by the high uncertainty values (close to 1) associated with low total evidence (close to 0) in the corrupted HAND views. This naturally limits the impact of unreliable evidence during fusion.", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, these results underscore our method's robustness in enhancing existing TMVC fusion strategies, particularly in challenging, noisy environments, while also maintaining stable performance in clean data conditions.", + "bbox": [ + 511, + 412, + 906, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2. Effect on Popular TMVC Methods", + "text_level": 1, + "bbox": [ + 511, + 479, + 802, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also demonstrate the effectiveness of our method when applied to various popular TMVC methods under corrupted view conditions (similar to Table 1) across six datasets. As shown in Figure 3, each radar chart represents one popular TMVC method, displaying classification accuracy under two conditions: with our method (w/ Ours, shown in red) and without our method (w/o Ours, shown in blue). Across all six TMVC methods, our method consistently improves performance under corrupted view conditions, as indicated by the expanded area covered by the red line compared to the blue line. The radar charts clearly show that our method effectively boosts classification accuracy on each dataset when faced with noisy and corrupted views. This improvement underscores the robustness of our approach in handling unreliable or degraded views across various TMVC frameworks. Particularly in challenging datasets like CUB, where noise can severely degrade model accuracy, our method provides a noticeable performance enhancement, indicating that it successfully mitigates the impact of corrupted views on multi-view classification accuracy. Overall, these results confirm the applicability and robustness of our method when integrated into diverse TMVC methods.", + "bbox": [ + 511, + 498, + 906, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.3. Sensitivity Analysis on Noisy Ratio", + "text_level": 1, + "bbox": [ + 511, + 835, + 802, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this experiment, we demonstrate the effectiveness of our method under varying noisy ratios in corrupted view conditions. We randomly select samples from half of", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de8b0859a823f81376e60849e79864afd769a43cf65dfb0f5ce4398470e40b84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 88, + 380, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c0f01b2ac4e5aea2576ccae5ba8e155aee3b56edfd08d00ce4a019b2eda08c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 88, + 604, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec9392ed813ce6d285679c47f117acb818d793c13282c49237d52e122bd0f752.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 88, + 830, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e68622b40361e16bfc07039ed007e0dd24922d600fdd02812c17f6a5a936a6c.jpg", + "image_caption": [ + "Figure 3. Comparison of classification accuracy between TMVC methods with and without our method." + ], + "image_footnote": [], + "bbox": [ + 168, + 263, + 379, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e04ec4757e5f8704007ad6dd906c4305d2c12289ec6033ef514cc716d7483e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 263, + 604, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/443332ab02c5a57058db11cb8c6f9c6b62440267a846ab09bcff5ad896ef6c82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 262, + 826, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e783faeb1e085466394b96283e4b7bae25f5c4b3e1a0e0265ff22f0013547701.jpg", + "image_caption": [ + "Figure 4. Sensitivity analysis of accuracy with varying noisy ratios under corrupted view conditions (where half of the views contain noise) on the CUB dataset." + ], + "image_footnote": [], + "bbox": [ + 112, + 452, + 372, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b775281c623353fb681beb21ec9dbf847e8eeed30566d7fa47a22e37fd0bff38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 452, + 630, + 587 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b68fbfdd59b4df89a0f34e4d8851d2ce29d9830274b356138a676c8b687e5fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 452, + 885, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the views in the CUB dataset and apply different noisy ratios to create polluted samples. As shown in Figure 4, the experimental results reveal that the proposed method significantly enhances the robustness of all three fusion strategies (ABF, A-CBF, and BCF) against varying noisy ratios. As the noisy ratio increases, both the baseline and proposed methods experience a decline in accuracy. However, the proposed method consistently outperforms the baseline across all fusion rules. Notably, the accuracy decline is slower for the method with the proposed fusion approach, and it is accompanied by reduced variability, indicating a greater resilience to noise. These findings highlight the effectiveness of the proposed method in maintaining both accuracy and stability, especially in high-noise environments. This makes it a promising solution for noise-robust multi-view fusion tasks.", + "bbox": [ + 88, + 652, + 485, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.4. Robustness Analysis under Varying Noise Intensity", + "text_level": 1, + "bbox": [ + 511, + 652, + 906, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the robustness of our method under varying noise intensities. In this part, noise vectors (denoted by $\\epsilon$ ) are sampled from a Gaussian distribution $\\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ . These noise vectors are then multiplied by an intensity factor $\\eta$ and added to half of the views in the test samples from the CUB dataset, i.e., $\\tilde{\\boldsymbol{x}}^v = \\boldsymbol{x}^v + \\eta \\epsilon^v$ . Figure 5 presents the robustness analysis under varying noise intensities. The results indicate that the proposed method significantly enhances the robustness of the three fusion strategies (ABF, A-CBF, and BCF) against different noise levels. Across all noise intensities, our method consistently outperforms the baseline, showing more stable accuracy and reduced variability, particularly at higher noise levels. These findings highlight the effectiveness of our method in improving performance and stability in noisy environments.", + "bbox": [ + 509, + 672, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15514", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1042192b852cb7f0395c7ee05a814731e2728323264ba4ba0b942daa0394054.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 89, + 374, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12237ecb883beb199a255e761e85ff45d52db53705b2472f38cec5bd92dda42e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 89, + 629, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/388b04a61267d04f5e800150f1a041205f923b7fbc7f43d56d98ebc6da9db48e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 89, + 885, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8329d23d6584ed5903c4cdedb1d6c1fb4db0449b96199a9ea0a8d94a9d36c81e.jpg", + "table_caption": [ + "Figure 5. Robustness analysis under varying noise intensities on the CUB dataset." + ], + "table_footnote": [], + "table_body": "
A-CBFABFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
α =0.00569.60 ± 2.7069.60 ± 2.7053.75 ± 1.7553.75 ± 1.7570.00 ± 1.4070.10 ± 1.40
α =0.00672.50 ± 0.3072.60 ± 0.3049.88 ± 0.6350.63 ± 0.1369.50 ± 1.0069.50 ± 1.00
α =0.00769.70 ± 1.4069.70 ± 1.4058.63 ± 6.3858.88 ± 6.6371.30 ± 1.6071.40 ± 1.60
α =0.00871.70 ± 1.0071.70 ± 1.0051.63 ± 0.6351.75 ± 0.7570.70 ± 0.7071.00 ± 0.70
α =0.00970.00 ± 1.3070.10 ± 1.3049.63 ± 1.3850.63 ± 0.3869.50 ± 1.0069.80 ± 1.00
α =0.01070.50 ± 1.8070.60 ± 1.8054.18 ± 4.6354.25 ± 2.5070.00 ± 0.8070.20 ± 0.80
α =0.02072.00 ± 1.3072.20 ± 1.3052.25 ± 0.0052.75 ± 0.2571.70 ± 1.3071.90 ± 1.30
α =0.03070.30 ± 1.3070.40 ± 1.3054.75 ± 1.7558.63 ± 1.3869.50 ± 0.8069.80 ± 0.80
α =0.04070.70 ± 1.8070.90 ± 1.8055.13 ± 2.3867.00 ± 1.0072.20 ± 1.0072.50 ± 1.00
α =0.05071.80 ± 1.1071.90 ± 1.1055.13 ± 0.3878.63 ± 1.3873.00 ± 0.9075.30 ± 0.90
α =0.06069.70 ± 1.2069.80 ± 1.2057.50 ± 2.5068.13 ± 1.3871.50 ± 1.2071.70 ± 1.20
α =0.07070.30 ± 1.3070.40 ± 1.3055.88 ± 2.6376.50 ± 1.0069.80 ± 1.0070.00 ± 1.00
α =0.08070.20 ± 1.3070.30 ± 1.3056.50 ± 1.2578.63 ± 0.1370.50 ± 1.3070.70 ± 1.30
α =0.09072.20 ± 1.7072.40 ± 1.7056.25 ± 3.5076.88 ± 2.3871.70 ± 1.0072.00 ± 1.00
α =0.10069.50 ± 1.4069.70 ± 1.4055.13 ± 1.6378.00 ± 1.2570.70 ± 1.2070.90 ± 1.20
", + "bbox": [ + 114, + 266, + 883, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Sensitivity analysis of user-defined $\\alpha$ under corrupted view conditions (with noise in half of the views) on the CAL dataset.", + "bbox": [ + 107, + 501, + 885, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.5. Sensitivity Analysis on User-defined $\\alpha$", + "text_level": 1, + "bbox": [ + 89, + 542, + 398, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we analyze the impact of different user-defined $\\alpha$ values under corrupted view conditions (similar to Table 1) on the CAL dataset. Table 2 shows the effectiveness of our method across various fusion strategies (A-CBF, ABF, and BCF) as $\\alpha$ changes. Generally, incorporating our method (\"w/Ours\") improves classification accuracy in most cases, particularly at higher $\\alpha$ values. Our method consistently achieves higher accuracy with relatively low variance across different $\\alpha$ values, especially in the ABF fusion, indicating strong performance in mitigating noise-induced fluctuations. Notably, when $\\alpha = 0.05$ , the accuracy across all three fusion strategies reaches a relatively high level of performance.", + "bbox": [ + 88, + 571, + 482, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is important to note that because noise was randomly sampled with different $\\alpha$ settings in our experiment, the performance of the \"w/o Ours\" condition fluctuates with changes in $\\alpha$ . However, despite these fluctuations, our method (\"w/ Ours\") consistently improves accuracy with same $\\alpha$ under identical noisy conditions (i.e., with the same sampled noise), demonstrating its capability to effectively reduce the impact of noise.", + "bbox": [ + 88, + 780, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed a generalized evidence filtering mechanism to improve the robustness of Trusted Multi-View Classification (TMVC) methods in the wild, particularly under challenging conditions with noisy and corrupted views. We framed the identification of unreliable evidence as a multiple testing problem, using $p$ -values to control the risk of false identification within a minimal range. This approach enables effective post-hoc adjustments during testing. By selectively adjusting evidence weights, our mechanism strengthens fusion robustness and mitigates performance degradation from low-quality data sources. Both theoretical analysis and extensive experiments on multiple real-world datasets demonstrate that our method significantly enhances classification accuracy and stability across various fusion strategies (ABF, A-CBF, and BCF), even in environments with substantial noise.", + "bbox": [ + 509, + 569, + 906, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 830, + 671, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62472315, 62476165).", + "bbox": [ + 511, + 857, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15515", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis. In International conference on machine learning, pages 1247-1255. PMLR, 2013. 2", + "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 2", + "[3] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1):289-300, 1995. 4", + "[4] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 5", + "[5] Li Fei-Fei and Pietro Perona. A bayesian hierarchical model for learning natural scene categories. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 524-531. IEEE, 2005. 5", + "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5", + "[7] Pinhan Fu, Xinyan Liang, Yuhua Qian, Qian Guo, Yayu Zhang, Qin Huang, and Ke Tang. Multi-scale features are effective for multi-modal classification: An architecture search viewpoint. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1", + "[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In International Conference on Learning Representations, 2021. 1, 2, 3, 5", + "[9] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1, 2, 3, 5", + "[10] Bingbing Jiang, Chenglong Zhang, Yan Zhong, Yi Liu, Yingwei Zhang, Xingyu Wu, and Weiguo Sheng. Adaptive collaborative fusion for multi-view semi-supervised classification. Information Fusion, 96:37-50, 2023. 1", + "[11] Bingbing Jiang, Xingyu Wu, Xiren Zhou, Anthony G Cohn, Yi Liu, Weiguo Sheng, and Huanhuan Chen. Semi-supervised multi-view feature selection with adaptive graph learning. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3615-3629, 2024. 1", + "[12] AUDUN. Jøsang. Subjective Logic: A formalism for reasoning under uncertainty. Springer, 2018. 2", + "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[14] Xinyan Liang, Yuhua Qian, Qian Guo, Honghong Cheng, and Jiye Liang. Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9236-9254, 2021. 1" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Xinyan Liang, Pinhan Fu, Qian Guo, Keyin Zheng, and Yuhua Qian. Dc-nas: Divide-and-conquer neural architecture search for multi-modal classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13754-13762, 2024. 1", + "[16] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7585-7593, 2022. 1, 2, 3, 5", + "[17] Wei Liu, Yufei Chen, Xiaodong Yue, Changqing Zhang, and Shaorong Xie. Safe multi-view deep classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8870-8878, 2023. 1, 2, 5", + "[18] Wei Liu, Yufei Chen, and Xiaodong Yue. Building trust in decision with conformalized multi-view deep classification. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7278-7287, 2024. 2", + "[19] Ying Liu, Lihong Liu, Cai Xu, Xiangyu Song, Ziyu Guan, and Wei Zhao. Dynamic evidence decoupling for trusted multi-view learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7269-7277, 2024. 5", + "[20] Jueqing Lu, Lan Du, Wray Buntine, Myong Chol Jung, Joanna Dipnall, and Belinda Gabbe. Navigating conflicting views: Harnessing trust for learning. arXiv preprint arXiv:2406.00958, 2024.3", + "[21] Jay Patrikar, Sushant Veer, Apoorva Sharma, Marco Pavone, and Sebastian Scherer. Rulefuser: Injecting rules in evidential networks for robust out-of-distribution trajectory prediction. arXiv preprint arXiv:2405.11139, 2024. 2, 3", + "[22] Zhenwen Ren and Quansen Sun. Simultaneous global and local graph structure preserving for multiple kernel clustering. IEEE transactions on neural networks and learning systems, 32(5):1839-1851, 2020. 2", + "[23] Yaniv Shachor, Hayit Greenspan, and Jacob Goldberger. A mixture of views network with applications to multi-view medical imaging. Neurocomputing, 374:1-9, 2020. 2", + "[24] Glenn Shafer. A mathematical theory of evidence. Princeton university press, 1976. 2", + "[25] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2", + "[26] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2", + "[27] Martijn van Breukelen, Robert PW Duin, David MJ Tax, and JE Den Hartog. Handwritten digit recognition by combined classifiers. Kybernetika, 34(4):381-386, 1998. 5", + "[28] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. California Institute of Technology, 2011. 5", + "[29] Meng Wang, Tian Lin, Lianyu Wang, Aidi Lin, Ke Zou, Xinxing Xu, Yi Zhou, Yuanyuan Peng, Qingquan Meng, Yiming Qian, et al. Uncertainty-inspired open set learning for retinal anomaly identification. Nature Communications, 14(1):6757, 2023. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15516", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Weiran Wang, Raman Arora, Karen Livescu, and Jeff Bilmes. On deep multi-view representation learning. In International conference on machine learning, pages 1083-1092. PMLR, 2015. 2", + "[31] Chang Xu, Dacheng Tao, and Chao Xu. A survey on multiview learning. arXiv preprint arXiv:1304.5634, 2013. 1", + "[32] Cai Xu, Jiajun Si, Ziyu Guan, Wei Zhao, Yue Wu, and Xiyue Gao. Reliable conflictive multi-view learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16129-16137, 2024. 1, 2, 3, 5", + "[33] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 2", + "[34] Jinglin Xu, Wenbin Li, Xinwang Liu, Dingwen Zhang, Ji Liu, and Junwei Han. Deep embedded complementary and interactive information for multi-view classification. In Proceedings of the AAAI conference on artificial intelligence, pages 6494-6501, 2020. 2", + "[35] Shaoxun Xu, Yufei Chen, Chao Ma, and Xiaodong Yue. Deep evidential fusion network for medical image classification. International Journal of Approximate Reasoning, 150: 188-198, 2022. 2", + "[36] Xiaoyu Yang, Jie Lu, and En Yu. Adapting multi-modal large language model to concept drift from pre-training onwards. In The Thirteenth International Conference on Learning Representations, 2025. 2", + "[37] Liyuan Zhang, Wei Liu, Yufei Chen, and Xiaodong Yue. Reliable multi-view deep patent classification. Mathematics, 10 (23):4545, 2022. 2", + "[38] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1" + ], + "bbox": [ + 89, + 92, + 483, + 558 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15517", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_model.json b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f99751d4db4e7dbae7b37acecf7e9824f77108cf --- /dev/null +++ b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.131, + 0.748, + 0.175 + ], + "angle": 0, + "content": "Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.203, + 0.825, + 0.294 + ], + "angle": 0, + "content": "Wei Liu\\(^{1}\\), Yufei Chen\\(^{1*}\\), Xiaodong Yue\\(^{2,3}\\) \n\\(^{1}\\)School of Computer Science and Technology, Tongji University, Shanghai, China \n\\(^{2}\\)Artificial Intelligence Institute, Shanghai University, Shanghai, China \n\\(^{3}\\)VLN Lab, Shanghai, China \nldachuan@outlook.com, yufeichen@tongji.edu.cn, yswantfly@shu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.358, + 0.486, + 0.752 + ], + "angle": 0, + "content": "Trusted multi-view classification (TMVC) addresses variations in data quality by evaluating the reliability of each view based on prediction uncertainty at the evidence level, reducing the impact of low-quality views commonly encountered in real-world scenarios. However, existing TMVC methods often struggle to maintain robustness during testing, particularly when integrating noisy or corrupted views. This limitation arises because the evidence collected by TMVC may be unreliable, frequently providing incorrect information due to complex view distributions and optimization challenges, ultimately leading to classification performance degradation. To enhance the robustness of TMVC methods in real-world conditions, we propose a generalized evidence filtering mechanism that is compatible with various fusion strategies commonly used in TMVC, including Belief Constraint Fusion, Aleatory Cumulative Belief Fusion, and Averaging Belief Fusion. Specifically, we frame the identification of unreliable evidence as a multiple testing problem and introduce \\( p \\)-values to control the risk of false identification. By selectively down-weighting unreliable evidence during testing, our mechanism ensures robust fusion and mitigates performance degradation. Both theoretical guarantees and empirical results demonstrate significant improvements in the classification performance of TMVC methods, supporting their reliable application in challenging, real-world environments." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.779, + 0.222, + 0.795 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Multi-view classification (MVC) enhances the accuracy of deep learning models by integrating information from multiple data views. By combining insights from diverse views, MVC improves decision-making in complex domains [7, 10, 11, 14, 15, 25, 31]. However, traditional" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.327, + 0.907, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.518, + 0.908, + 0.588 + ], + "angle": 0, + "content": "Figure 1. Classification accuracy degradation with noise (w/) compared to no noise (w/o) across three fusion strategies (BCF, A-CBF, and ABF) on six datasets. Noise is introduced by replacing half of the views to simulate data collected from malfunctioning sensors." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.624, + 0.907, + 0.776 + ], + "angle": 0, + "content": "MVC approaches often assume that each view consistently provides high-quality, reliable information—an assumption that is frequently unrealistic in real-world applications. In practical scenarios, data quality and reliability across views can vary significantly, leading to contradictory or misleading information that compromises prediction accuracy and reliability. As a result, conventional MVC models may struggle to reliably identify the ground truth, particularly when dealing with noisy, corrupted, or otherwise low-quality data views [38]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address these limitations, trusted multi-view classification (TMVC) [8, 9, 16, 17, 32] has emerged as an effective solution. Unlike traditional MVC, TMVC evaluates the trustworthiness of each view by leveraging prediction uncertainty derived from view-specific evidence, enabling it to identify unreliable views and prioritize informative ones in the final predictions. This approach has proven highly effective in critical applications such as medical imaging diagno" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15508" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "sis [29] and autonomous driving [21], where distinguishing between reliable and unreliable predictions is essential." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.124, + 0.484, + 0.455 + ], + "angle": 0, + "content": "In TMVC, evidence from each sample for each view contributes to its prediction and is also used to calculate a view-specific uncertainty score \\( u \\), which reflects the confidence in the quality of that view for the given sample. A high uncertainty score (i.e., \\( u = 1 \\)) indicates a substantial deviation from the training data, potentially due to noise or corruption. In such cases, evidence supporting all classes is minimized and fused using one of three strategies: Belief Constraint Fusion (BCF) [8, 9], Aleatory Cumulative Belief Fusion (A-CBF) [16-18, 37], and Averaging Belief Fusion (ABF) [32]. However, despite their success in controlled settings, these methods often struggle to remain robust in the wild. Complex and diverse data view distributions, combined with varying levels of optimization, can distort uncertainty estimates, resulting in unreliable evidence that misleads the model and biases predictions toward incorrect classes. As shown in Figure 1, introducing random noise (e.g., replacing half of the views in a dataset) to simulate sensor failure conditions leads to a significant drop in classification accuracy across all fusion strategies on six real-world datasets, underscoring the risks of decision-making in unpredictable and noisy environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.637 + ], + "angle": 0, + "content": "To address this challenge, we propose a generalized and robust testing-time evidence filtering mechanism to enhance the robustness of TMVC methods in real-world, unpredictable environments. Specifically, we frame the identification of unreliable evidence as a multiple testing problem, introducing \\( p \\)-values to control the risk of false identification within a small margin. This approach enables effective post-hoc adjustments during testing. By selectively down-weighting evidence from unreliable views, our mechanism enhances fusion robustness and mitigates performance degradation caused by low-quality data sources. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.639, + 0.483, + 0.684 + ], + "angle": 0, + "content": "1. We propose a testing-time evidence filtering mechanism that is compatible with multiple fusion strategies, enhancing the robustness of existing TMVC methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.684, + 0.483, + 0.744 + ], + "angle": 0, + "content": "2. Our approach frames the identification of unreliable evidence as a multiple testing problem and selectively adjusts evidence weights during testing, mitigating the negative impact of unreliable views on fusion." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.745, + 0.483, + 0.789 + ], + "angle": 0, + "content": "3. Theoretical and empirical analyses validate the effectiveness of the proposed method in improving both classification accuracy and robustness." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.639, + 0.483, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.241, + 0.822 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.318, + 0.847 + ], + "angle": 0, + "content": "2.1. Multi-view Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Over the past few years, numerous multi-view classification (MVC) approaches [1, 2, 16, 22, 25, 26, 30, 34, 36] have been developed, particularly those leveraging deep neural" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "networks (DNNs). These methods aim to learn shared representations by integrating information from different types of views, resulting in improved performance. By combining insights from diverse views, MVC has achieved remarkable gains and has been widely applied across various tasks [23, 35]. However, in real-world scenarios, the quality of data across views can vary significantly, introducing noise and inconsistencies that undermine both classification accuracy and reliability—factors that are especially critical in areas like medical diagnosis and autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.256, + 0.806, + 0.271 + ], + "angle": 0, + "content": "2.2. Trusted Multi-view Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.908, + 0.612 + ], + "angle": 0, + "content": "To address these challenges, Trusted Multi-View Classification (TMVC) has emerged as a framework that assesses the reliability of each view through prediction uncertainty, allowing models to prioritize reliable views and down-weight unreliable ones [8, 9, 16, 17, 32, 33]. TMVC methods often rely on Subjective Logic (SL) [12] to model view-specific predictions as multinomial opinions, which are then fused using belief combination rules tailored to the context [12]. For instance, extensions of Dempster-Shafer theory [24] were firstly introduced by [8, 9] to facilitate evidence fusion in multi-view settings. Other fusion approaches, such as Cumulative Belief Fusion [16, 17] and Averaging Belief Fusion [32], have also been proposed to further enhance the effectiveness of TMVC. Despite their success, all these methods fall short of achieving robust performance during testing, as the magnitude of estimated uncertainty is highly sensitive to the diverse distributions of data views and varying levels of model optimization, which significantly degrades classification accuracy in real-world scenarios. In contrast, our method improves testing-time robustness through an evidence filtering mechanism, thereby reducing the risks associated with final decision-making." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.629, + 0.818, + 0.646 + ], + "angle": 0, + "content": "3. Overview and Analysis of TMVC" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.656, + 0.818, + 0.67 + ], + "angle": 0, + "content": "3.1. Generalized Framework of TMVC" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.792 + ], + "angle": 0, + "content": "Given the training data \\( D_{train} = \\left\\{\\left\\{\\boldsymbol{x}_i^v\\right\\}_{v=1}^V, \\boldsymbol{y}_i\\right\\}_{i=1}^N \\), where \\( N \\) is the number of samples, each sample has \\( V \\) views with a one-hot label \\( \\boldsymbol{y}_i = \\left\\{y_{ik}\\right\\}_{k=1}^K \\) for a \\( K \\)-class classification task. TMVC uses subjective logic [12] to learn a function \\( f \\) that maps \\( \\left\\{\\boldsymbol{x}_i^v\\right\\}_{v=1}^V \\) to \\( \\boldsymbol{y}_i \\) along with a corresponding overall uncertainty score \\( u_i \\), which quantifies prediction confidence at the evidence level." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Specifically, let \\( f_{v}(\\boldsymbol{x}_{i}^{v}) \\) denote the output of the penultimate (logits) layer of the view-specific neural network for sample \\( \\boldsymbol{x}_{i}^{v} \\) in a single view. Instead of the standard softmax function for a single categorical prediction, TMVC employs an alternative activation function \\( a(\\cdot) \\) (e.g., ReLU) to capture non-negative evidence \\( e_{i}^{v} = a(f_{v}(\\boldsymbol{x}_{i}^{v})) \\), where \\( e_{i}^{v} = \\{e_{ik}^{v}\\}_{k=1}^{K} \\). The view-specific prediction uncertainty is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15509" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.093, + 0.152, + 0.107 + ], + "angle": 0, + "content": "given by" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.118, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\nu _ {i} ^ {v} = K / S _ {i} ^ {v}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.147, + 0.483, + 0.193 + ], + "angle": 0, + "content": "where \\( S_{i}^{v} = \\sum_{k=1}^{K} e_{ik}^{v} + K \\). This reflects the lack of total evidence across all classes in a single view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.194, + 0.484, + 0.269 + ], + "angle": 0, + "content": "TMVC then applies evidence fusion strategies, such as BCF, A-CBF, and ABF, to integrate the evidence from each view into a unified multi-view evidence representation \\( e_i = \\{e_{ik}\\}_{k=1}^K \\). These three fusion rules for two views are commonly expressed as follows [20]:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.278, + 0.482, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} A - C B F \\left(\\boldsymbol {e} _ {i} ^ {1}, \\boldsymbol {e} _ {i} ^ {2}\\right) = \\boldsymbol {e} _ {i} ^ {1} + \\boldsymbol {e} _ {i} ^ {2}. \\\\ A B F \\left(e _ {i} ^ {1}, e _ {i} ^ {2}\\right) = \\left(e _ {i} ^ {1} + e _ {i} ^ {2}\\right) / 2. \\\\ B C F \\left(\\boldsymbol {e} _ {i} ^ {1}, \\boldsymbol {e} _ {i} ^ {2}\\right) = \\boldsymbol {e} _ {i} ^ {1} + \\boldsymbol {e} _ {i} ^ {2} + \\boldsymbol {e} _ {i} ^ {1} \\boldsymbol {e} _ {i} ^ {2} / K. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.377 + ], + "angle": 0, + "content": "These fusion rules can also be extended to handle multiple views with ease." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.528 + ], + "angle": 0, + "content": "In the training phase, TMVC links the generated evidence to the concentration parameter of the Dirichlet distribution and optimizes model parameters using a task-specific loss function. This loss primarily consists of a cross-entropy term, which maximizes the evidence for the correct label of each sample, and a Kullback-Leibler (KL)-divergence term, which drives the evidence for incorrect labels toward zero, as described in [8]. The final multi-view prediction \\( \\pmb{p}_i = \\{p_{ik}\\}_{k=1}^K \\) is obtained as the expectation of the learned Dirichlet distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.541, + 0.483, + 0.559 + ], + "angle": 0, + "content": "\\[\np _ {i k} = \\left(e _ {i k} + 1\\right) / S _ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.648 + ], + "angle": 0, + "content": "where \\( S_{i} = \\sum_{k = 1}^{K}e_{ik} + K \\), and the overall uncertainty \\( u_{i} \\) is given by \\( u_{i} = K / S_{i} \\). Finally, the model's best performance on the validation dataset \\( D_{val} \\) is used to select the optimal parameters for testing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.419, + 0.673 + ], + "angle": 0, + "content": "3.2. Limitations of TMVC at Testing Time" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.768 + ], + "angle": 0, + "content": "Despite its success across various domains [21, 29], TMVC methods face significant limitations at testing time, particularly in real-world scenarios with diverse data view distributions and varying levels of model optimization. These factors often degrade TMVC's robustness, especially when handling low-quality data views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Specifically, given a multi-view test dataset \\( D_{test} = \\left\\{\\left\\{\\boldsymbol{x}_j^v\\right\\}_{v=1}^V\\right\\}_{j=1}^M \\), a trained TMVC model with \\( V \\) view-specific networks provides uncertainty estimates for each view of a testing sample \\( \\boldsymbol{x}_j^v \\). Ideally, a high uncertainty score (e.g., \\( u_j^v = 1 \\)) for a low-quality view would reduce its evidence contribution across all classes to zero (\\( e_j^v = 0 \\)), minimizing its impact during fusion. However, these uncertainty estimates are sensitive to data distributions and" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.715, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.091, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.245, + 0.906, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Density distribution of uncertainty for noisy view results under noise corruption and the corresponding density distribution of remaining unreliable evidence across three fusion strategies on the CUB dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.905, + 0.361 + ], + "angle": 0, + "content": "model optimization levels, making it challenging to reliably achieve \\( u_{j}^{v} = 1 \\) for extremely noisy or corrupted views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.648 + ], + "angle": 0, + "content": "For example, in the CUB dataset, even when one view is entirely corrupted by noise (i.e., features of one view are replaced by random noise), the uncertainty score for that view typically falls within the range [0.7, 1.0], leaving residual evidence mostly between [0.0, 4.0] (as illustrated in Figure 2) that contributes to fusion and can potentially bias predictions. Since this view lacks useful information, its residual evidence is unreliable and negatively impacts multi-view classification accuracy, as shown in Figure 1. Moreover, the overall uncertainty score derived from fusion strategies like BCF and A-CBF is often lower as additional views are fused (as demonstrated in [9, 16]). Even with ABF fusion, the overall uncertainty remains below the highest view-specific uncertainty [32]. As a result, decision-makers may struggle to detect potentially misclassified samples caused by low-quality views. These limitations underscore the challenges TMVC faces in real-world conditions, where testing-time robustness is crucial for reliable decision-making." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.688, + 0.681 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we present a strategy to enhance the robustness of TMVC methods at testing time. Performance degradation in TMVC methods often arises from the negative impact of such low-quality views, which typically exhibit high prediction uncertainty during testing, compared to the reliable views seen during training. An intuitive solution to this problem is to reduce the contribution of unreliable evidence from low-quality views during multi-view fusion. To address this, we propose a generalized evidence filtering mechanism that is compatible with various TMVC fusion rules. The mechanism consists of two key components: (1) identifying unreliable evidence collected from low-quality views, and (2) assigning lower weights to minimize their impact on the final decision." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15510" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.436, + 0.108 + ], + "angle": 0, + "content": "4.1. Reliable Evidence Discovery in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.204 + ], + "angle": 0, + "content": "Identifying unreliable evidence is equivalent to discovering reliable evidence. The primary challenge in this process is controlling the false discovery rate (FDR) [3], since the true label is unknown. We address this challenge by framing reliable evidence discovery as a multiple hypothesis testing problem from a statistical learning perspective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.339 + ], + "angle": 0, + "content": "Given a multi-view test dataset \\( D_{test} \\), the objective is to identify as many reliable pieces of evidence as possible while ensuring that the FDR is maintained below a user-defined threshold \\( \\alpha \\in (0,1) \\). Let us define a discovery function \\( \\Lambda(\\kappa^v) \\), which selects a subset \\( S^v \\) of view-specific evidence for the test points of interest. Evidences with smaller values of \\( u_j^v \\), i.e., \\( u_j^v \\leq \\kappa^v \\), is likely to be included in the subset, where \\( \\kappa^v \\) is a threshold determined by the discovery function. The FDR for this function is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.346, + 0.484, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F D R} \\left(\\Lambda \\left(\\kappa^ {v}\\right)\\right) := \\mathbb {E} \\left[ \\frac {\\sum_ {j \\in D _ {t e s t}} \\mathbb {1} \\left\\{j \\in S ^ {v} , u _ {j} ^ {v} > \\kappa^ {v} \\right\\}}{\\max \\left\\{1 , | S ^ {v} | \\right\\}} \\right]. (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Our goal is to ensure that the subset \\( S^v \\) identified by \\( \\Lambda(\\kappa^v) \\) satisfies the following error-control property:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.482 + ], + "angle": 0, + "content": "Definition 1 (Risk-controlling discovery). Let \\(\\kappa^v\\) be a random variable. We define \\(\\Lambda(\\kappa^v)\\) as an \\((\\alpha, \\delta)\\)-risk-controlling discovery (RCD) function if" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.492, + 0.482, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left(\\operatorname {F D R} \\left(\\Lambda \\left(\\kappa^ {v}\\right)\\right) \\leq \\alpha\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.52, + 0.32, + 0.535 + ], + "angle": 0, + "content": "holds with probability at least \\(1 - \\delta\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.679 + ], + "angle": 0, + "content": "Here, \\(\\delta \\in (0,1)\\) is the confidence level. To meet this risk control criterion, we need to identify a suitable \\(\\Lambda (\\kappa^v)\\) that adapts to diverse data view distributions encountered in the wild. For this purpose, we gather statistical evidence via hypothesis testing, where we consider the view-specific evidence of a test sample to be discovered (i.e, \\(u_{j}^{v}\\leq \\kappa^{v}\\)) by \\(\\Lambda (\\kappa^v)\\) as the null hypothesis. Rejecting the null hypothesis reflects that \\(u_{j}^{v} > \\kappa^{v}\\), and therefore the evidence is unreliable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.484, + 0.741 + ], + "angle": 0, + "content": "Under this framework, discovering reliable evidence boils down to simultaneously testing the \\(M\\) hypotheses. To do this, for any arbitrary view-specific evidence \\(\\pmb{e}_j^v\\), we compute the \\(p\\)-value:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.75, + 0.484, + 0.786 + ], + "angle": 0, + "content": "\\[\np \\text {- v a l u e} \\left(\\boldsymbol {e} _ {j} ^ {v}\\right) = \\frac {1 + \\sum_ {l \\in D _ {v a l}} \\mathbb {1} \\left\\{u _ {j} ^ {v} \\geq u _ {l} ^ {v} \\right\\}}{\\left| D _ {v a l} \\right| + 1}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.484, + 0.87 + ], + "angle": 0, + "content": "It is clear that the \\( p \\)-value quantifies how representative the evidence \\( e_j^v \\) is of reliable evidence. If the \\( p \\)-value is less than \\( \\alpha \\), the null hypothesis of exchangeability with the validation data is rejected, deeming the evidence unreliable. We obtain marginally valid \\( p \\)-values, which satisfy:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.88, + 0.484, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left(p - \\text {v a l u e} \\left(\\boldsymbol {e} _ {j} ^ {v}\\right) \\leq \\alpha\\right) \\leq \\alpha , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.153 + ], + "angle": 0, + "content": "indicating that the probability of falsely identifying unreliable evidence is controlled by the small margin \\(\\alpha\\). To achieve this, we set a view-specific, distribution-free threshold using the Benjamini-Hochberg (BH) procedure [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.163, + 0.907, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\kappa^ {v} = Q \\left(\\left\\{u _ {l} ^ {v} \\right\\} _ {l \\in D _ {v a l}}, \\alpha\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.191, + 0.906, + 0.238 + ], + "angle": 0, + "content": "where \\(Q(\\cdot, \\alpha)\\) denotes the \\(\\lceil (1 - \\alpha)(|D_{val}| + 1) \\rceil\\)-th smallest value in \\(\\{u_l^v\\}_{l \\in D_{val}}\\). Then, we construct the reliable evidence set using \\(\\Lambda(\\kappa^v)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.248, + 0.907, + 0.268 + ], + "angle": 0, + "content": "\\[\nS ^ {v} = \\left\\{j \\in D _ {\\text {t e s t}}: u _ {j} ^ {v} \\leq \\kappa^ {v} \\right\\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.277, + 0.8, + 0.292 + ], + "angle": 0, + "content": "Finally, we have the following proposition:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.906, + 0.333 + ], + "angle": 0, + "content": "Proposition 1. \\(\\Lambda (\\kappa^v)\\) is an \\((\\alpha ,\\delta)\\)-risk-controlling discovery (RCD) function." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.341, + 0.906, + 0.358 + ], + "angle": 0, + "content": "4.2. Mitigating the Influence of Unreliable Evidence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.424 + ], + "angle": 0, + "content": "After identifying the reliability of view-specific evidence, we aim to mitigate the impact of unreliable evidence during multi-view fusion. To achieve this, we define the filtered evidence for each testing sample as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.435, + 0.906, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\boldsymbol {e}} _ {j} ^ {v} = \\left\\{ \\begin{array}{c c} \\boldsymbol {e} _ {j} ^ {v}, & j \\in S ^ {v}. \\\\ \\alpha \\cdot \\boldsymbol {e} _ {j} ^ {v}, & j \\notin S ^ {v}. \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Here, \\(\\alpha\\) is a user-defined parameter that controls the weight assigned to unreliable evidence, directly influencing the FDR control. Smaller values of \\(\\alpha\\) effectively reduce the influence of low-quality views, minimizing the risk of unreliable evidence negatively impacting the final prediction. This tight control is especially valuable in high-stakes applications where high confidence and accuracy in predictions is crucial. On the other hand, if \\(\\alpha\\) is set to a larger value, the mechanism retains more influence from low-quality views, which may be beneficial when such views contain partially valuable information despite high uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Moreover, the choice of \\(\\alpha\\) balances FDR control and information retention, allowing flexibility in managing the contribution of uncertain views to the fusion process. A lower \\(\\alpha\\) tightens FDR control, reducing the likelihood of unreliable evidence being included and enhancing prediction reliability. Importantly, we do not fully exclude unreliable evidence, as even noisy views may contain useful information for classification. Instead, we adopt a weighted filtering approach that retains the potential benefits of low-quality views while mitigating their negative effects. This soft filtering strategy allows the model to remain adaptive and robust under real-world conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "After filtering, the view-specific evidence is fused using the selected fusion rules (e.g., BCF, A-CBF, or ABF), achieving an adaptable and robust multi-view fusion process that is compatible with various TMVC fusion strategies. This leads to the following propositions:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15511" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "Proposition 2. Applying evidence filtering improves the classification performance of TMVC fusion strategies (e.g., BCF, A-CBF, or ABF), under the assumption that unreliable evidence provides incorrect information in the wild." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.159, + 0.483, + 0.22 + ], + "angle": 0, + "content": "Proposition 3. The overall uncertainty of multi-view results generated by TMVC fusion strategies with evidence filtering for unreliable evidence will exhibit greater uncertainty than fusion without evidence filtering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.484, + 0.377 + ], + "angle": 0, + "content": "Both Propositions 2 and 3 guarantee the robustness of our method when confronted with unreliable evidence (i.e., evidence from low-quality views) in the wild. Note that unreliable evidence providing correct predictions by chance (a counterintuitive situation) may hurt classification performance in our method, particularly when evidence from other views is incorrect. To address this, our method exhibits greater overall uncertainty for these multi-view results, as shown in Proposition 3, which can alert humans to be cautious with these results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.483 + ], + "angle": 0, + "content": "In summary, our evidence filtering mechanism improves testing-time robustness by adjusting the weights of view-specific evidence based on reliability. This enables TMVC methods to perform reliably in environments with data inconsistencies. This testing-time robustness is crucial for dependable deployment in high-stakes, real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.491, + 0.406, + 0.508 + ], + "angle": 0, + "content": "4.3. Computational Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Our method introduces a generalized framework that can be applied to various DNNs for multi-view testing. Given a single test pass with complexity \\( O(T_{test}) \\) on an arbitrary TMVC model, the overall computational complexity is:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.597, + 0.482, + 0.614 + ], + "angle": 0, + "content": "\\[\nO (L \\log L) + O (M \\cdot T _ {t e s t}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.789 + ], + "angle": 0, + "content": "where \\( L \\) and \\( M \\) represent the sizes of the validation and test datasets, respectively. The term \\( O(L\\log L) \\) reflects the complexity involved in threshold calculation, primarily due to sorting the uncertainty scores within the validation set to determine an appropriate threshold for filtering unreliable evidence. The term \\( O(M\\cdot T_{test}) \\) represents the complexity of running one test pass on \\( M \\) samples in the testing set. In most practical scenarios, the term \\( O(M\\cdot T_{test}) \\) will typically dominate, especially for large-scale DNNs, unless the validation set is extremely large or the neural networks are shallow and computationally lightweight." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.224, + 0.817 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we validate the effectiveness of our proposed method on six real-world datasets. Each experiment is run 10 times, and we report the mean and standard deviation of accuracy (ACC) to ensure reliability. The extensive results demonstrate the clear advantage of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.704, + 0.107 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.114, + 0.62, + 0.127 + ], + "angle": 0, + "content": "5.1.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.133, + 0.907, + 0.496 + ], + "angle": 0, + "content": "We conduct experiments on six real-world multi-view datasets as follows: CUB [28]: This dataset contains 11,788 bird images from 200 distinct categories, each paired with text descriptions. For our experiments, we utilize the first 10 categories. Image features are extracted using GoogleNet, while text features are extracted using doc2vec. SCENE [5]: This dataset comprises 4,485 images from 15 indoor and outdoor scene categories. Multiple views are represented using three feature types: GIST, PHOG, and LBP. CAL [6]: The CAL dataset consists of 8,677 images from 101 categories. We focus on the first 20 categories and extract six types of features: GABOR, WAVELET-MOMENTS, CENHIST, HOG, GIST, and LBP. HAND [27]: This dataset contains 2,000 samples across 10 digit classes (from '0' to '9'), with 200 samples per class. Six feature descriptors (PIX, FOU, FAC, ZER, KAR, MOR) are extracted as multiple views. PIE: This dataset consists of 680 facial images from 68 individuals. Three types of features are extracted: intensity, LBP, and Gabor. NUS: A subset of the NUS-WIDE dataset [4], it contains 30,000 images spanning 31 object categories, represented in five views, including a 64-D color histogram, 225-D blockwise color moments, 144-D color correlogram, 73-D edge direction histogram, and 128-D wavelet texture." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.505, + 0.699, + 0.52 + ], + "angle": 0, + "content": "5.1.2. Compared Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.907, + 0.706 + ], + "angle": 0, + "content": "Since our method is a generalized framework applicable to various trusted multi-view classification methods, we primarily evaluate its effectiveness by comparing baseline methods with and without our approach in these experiments. We apply our method to three fusion rules: BCF, A-CBF, and ABF, using the general loss function proposed in the work of [8]. We then integrate our method with state-of-the-art multi-view deep classification methods, including TMC [8], TMDOA [16], ETMC [9], SMDC [17], ECML [32], and CCML [19]. These methods use the same neural network architecture as ours, ensuring a consistent basis for comparison." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.715, + 0.679, + 0.73 + ], + "angle": 0, + "content": "5.1.3. Implementations" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.734, + 0.906, + 0.87 + ], + "angle": 0, + "content": "For all datasets, we split the data into training (70%), validation (10%), and testing (20%) sets. Fully connected networks were used as view-specific backbones. To obtain a trained TMVC model, we employed the Adam optimizer [13] with an \\( l_{2} \\)-norm regularization of \\( 1e^{-5} \\) and a learning rate of \\( 1e^{-3} \\). During testing, we set \\( \\alpha = 0.05 \\) by default, ensuring an FDR below 5% unless otherwise specified. The model was implemented in PyTorch and run on a GeForce RTX 4090 GPU with 24GB of memory." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Note: Optimization during training is not a primary focus of this work; hence, finding the optimal model parameters" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15512" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.895, + 0.348 + ], + "angle": 0, + "content": "
Fusion with Clean Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB98.75±0.4298.75±0.4297.50±1.6798.33±1.6797.08±0.4297.92±0.42
SCENE70.94±0.0670.69±0.3171.63±0.7571.31±1.0670.50±0.1370.25±0.25
CAL98.50±0.5098.38±0.3897.63±0.3897.25±0.2598.75±0.5098.50±0.25
HAND97.88±0.1398.25±0.2599.00±0.0599.00±0.7598.88±0.3899.25±0.50
PIE98.08±1.1598.46±0.7798.85±0.3898.85±0.3897.69±1.5498.08±1.15
NUS53.85±1.0054.05±0.8055.35±0.3055.43±0.0857.75±0.7557.88±0.78
Fusion with Corrupted Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB64.17±5.0080.00±0.8356.67±6.6779.17±1.6765.33±3.3383.75±1.25
SCENE46.00±3.3853.38±2.3150.75±2.0051.38±1.9449.25±0.6352.75±0.75
CAL56.75±1.8876.75±5.0070.25±0.7571.50±2.0073.25±0.1375.50±2.00
HAND64.50±1.6370.00±1.8870.00±0.7570.75±0.6375.75±1.7578.25±1.75
PIE89.77±0.0196.54±0.3888.46±3.8590.77±2.3188.08±0.3891.54±0.01
NUS27.70±1.6037.70±0.8536.73±1.2238.95±1.2037.00±1.6040.63±1.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.358, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Table 1. Classification performance (ACC (%)) for fusion with clean views and corrupted views, based on three fusion strategies, with or without our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.482, + 0.473 + ], + "angle": 0, + "content": "was not the primary goal. Instead, we provide comparative results with and without our method under identical settings (same random seed, dataset splits, optimization parameters, etc.) to evaluate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.295, + 0.499 + ], + "angle": 0, + "content": "5.2. Quantitative Analysis" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.505, + 0.476, + 0.52 + ], + "angle": 0, + "content": "5.2.1. Effect of Our Method on Basic Fusion Strategies" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.749 + ], + "angle": 0, + "content": "In this evaluation, we assess the effectiveness of our proposed method across three fusion strategies (ABF, A-CBF, and BCF) under both clean and corrupted view conditions. Initially, we applied our method to these fusion strategies in the context of clean views, where no noise was introduced. As shown in Table 1, our method generally improves classification accuracy across most datasets, even in noise-free environments. This demonstrates that our approach enhances model stability in both noisy and normal conditions. Minor performance decreases observed in the SCENE and CAL datasets with clean views may be attributed to our method's down-weighting of higher uncertainty evidence for some samples in certain views. These higher uncertainty samples could otherwise positively impact classification in some cases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Under corrupted view conditions, where random noise in the range of \\([-5, 5]\\) is added to half of the views, our method significantly boosts accuracy across all datasets. For example, on the CUB dataset with A-CBF fusion, our method achieves nearly a \\(23\\%\\) improvement. For datasets like HAND, the modest performance improvement under corrupted conditions can be explained by the high uncertainty values (close to 1) associated with low total evidence (close to 0) in the corrupted HAND views. This naturally limits the impact of unreliable evidence during fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Overall, these results underscore our method's robustness in enhancing existing TMVC fusion strategies, particularly in challenging, noisy environments, while also maintaining stable performance in clean data conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.803, + 0.495 + ], + "angle": 0, + "content": "5.2.2. Effect on Popular TMVC Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.908, + 0.831 + ], + "angle": 0, + "content": "We also demonstrate the effectiveness of our method when applied to various popular TMVC methods under corrupted view conditions (similar to Table 1) across six datasets. As shown in Figure 3, each radar chart represents one popular TMVC method, displaying classification accuracy under two conditions: with our method (w/ Ours, shown in red) and without our method (w/o Ours, shown in blue). Across all six TMVC methods, our method consistently improves performance under corrupted view conditions, as indicated by the expanded area covered by the red line compared to the blue line. The radar charts clearly show that our method effectively boosts classification accuracy on each dataset when faced with noisy and corrupted views. This improvement underscores the robustness of our approach in handling unreliable or degraded views across various TMVC frameworks. Particularly in challenging datasets like CUB, where noise can severely degrade model accuracy, our method provides a noticeable performance enhancement, indicating that it successfully mitigates the impact of corrupted views on multi-view classification accuracy. Overall, these results confirm the applicability and robustness of our method when integrated into diverse TMVC methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.837, + 0.803, + 0.852 + ], + "angle": 0, + "content": "5.2.3. Sensitivity Analysis on Noisy Ratio" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this experiment, we demonstrate the effectiveness of our method under varying noisy ratios in corrupted view conditions. We randomly select samples from half of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15513" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.089, + 0.382, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.089, + 0.605, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.089, + 0.831, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.264, + 0.38, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.264, + 0.605, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.263, + 0.828, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.424, + 0.807, + 0.438 + ], + "angle": 0, + "content": "Figure 3. Comparison of classification accuracy between TMVC methods with and without our method." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.453, + 0.373, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.453, + 0.631, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.453, + 0.886, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.599, + 0.908, + 0.627 + ], + "angle": 0, + "content": "Figure 4. Sensitivity analysis of accuracy with varying noisy ratios under corrupted view conditions (where half of the views contain noise) on the CUB dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.654, + 0.486, + 0.895 + ], + "angle": 0, + "content": "the views in the CUB dataset and apply different noisy ratios to create polluted samples. As shown in Figure 4, the experimental results reveal that the proposed method significantly enhances the robustness of all three fusion strategies (ABF, A-CBF, and BCF) against varying noisy ratios. As the noisy ratio increases, both the baseline and proposed methods experience a decline in accuracy. However, the proposed method consistently outperforms the baseline across all fusion rules. Notably, the accuracy decline is slower for the method with the proposed fusion approach, and it is accompanied by reduced variability, indicating a greater resilience to noise. These findings highlight the effectiveness of the proposed method in maintaining both accuracy and stability, especially in high-noise environments. This makes it a promising solution for noise-robust multi-view fusion tasks." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.67 + ], + "angle": 0, + "content": "5.2.4. Robustness Analysis under Varying Noise Intensity" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We analyze the robustness of our method under varying noise intensities. In this part, noise vectors (denoted by \\(\\epsilon\\)) are sampled from a Gaussian distribution \\(\\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\). These noise vectors are then multiplied by an intensity factor \\(\\eta\\) and added to half of the views in the test samples from the CUB dataset, i.e., \\(\\tilde{\\boldsymbol{x}}^v = \\boldsymbol{x}^v + \\eta \\epsilon^v\\). Figure 5 presents the robustness analysis under varying noise intensities. The results indicate that the proposed method significantly enhances the robustness of the three fusion strategies (ABF, A-CBF, and BCF) against different noise levels. Across all noise intensities, our method consistently outperforms the baseline, showing more stable accuracy and reduced variability, particularly at higher noise levels. These findings highlight the effectiveness of our method in improving performance and stability in noisy environments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15514" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.09, + 0.375, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.09, + 0.63, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.09, + 0.887, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.239, + 0.741, + 0.254 + ], + "angle": 0, + "content": "Figure 5. Robustness analysis under varying noise intensities on the CUB dataset." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.267, + 0.884, + 0.491 + ], + "angle": 0, + "content": "
A-CBFABFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
α =0.00569.60 ± 2.7069.60 ± 2.7053.75 ± 1.7553.75 ± 1.7570.00 ± 1.4070.10 ± 1.40
α =0.00672.50 ± 0.3072.60 ± 0.3049.88 ± 0.6350.63 ± 0.1369.50 ± 1.0069.50 ± 1.00
α =0.00769.70 ± 1.4069.70 ± 1.4058.63 ± 6.3858.88 ± 6.6371.30 ± 1.6071.40 ± 1.60
α =0.00871.70 ± 1.0071.70 ± 1.0051.63 ± 0.6351.75 ± 0.7570.70 ± 0.7071.00 ± 0.70
α =0.00970.00 ± 1.3070.10 ± 1.3049.63 ± 1.3850.63 ± 0.3869.50 ± 1.0069.80 ± 1.00
α =0.01070.50 ± 1.8070.60 ± 1.8054.18 ± 4.6354.25 ± 2.5070.00 ± 0.8070.20 ± 0.80
α =0.02072.00 ± 1.3072.20 ± 1.3052.25 ± 0.0052.75 ± 0.2571.70 ± 1.3071.90 ± 1.30
α =0.03070.30 ± 1.3070.40 ± 1.3054.75 ± 1.7558.63 ± 1.3869.50 ± 0.8069.80 ± 0.80
α =0.04070.70 ± 1.8070.90 ± 1.8055.13 ± 2.3867.00 ± 1.0072.20 ± 1.0072.50 ± 1.00
α =0.05071.80 ± 1.1071.90 ± 1.1055.13 ± 0.3878.63 ± 1.3873.00 ± 0.9075.30 ± 0.90
α =0.06069.70 ± 1.2069.80 ± 1.2057.50 ± 2.5068.13 ± 1.3871.50 ± 1.2071.70 ± 1.20
α =0.07070.30 ± 1.3070.40 ± 1.3055.88 ± 2.6376.50 ± 1.0069.80 ± 1.0070.00 ± 1.00
α =0.08070.20 ± 1.3070.30 ± 1.3056.50 ± 1.2578.63 ± 0.1370.50 ± 1.3070.70 ± 1.30
α =0.09072.20 ± 1.7072.40 ± 1.7056.25 ± 3.5076.88 ± 2.3871.70 ± 1.0072.00 ± 1.00
α =0.10069.50 ± 1.4069.70 ± 1.4055.13 ± 1.6378.00 ± 1.2570.70 ± 1.2070.90 ± 1.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.502, + 0.887, + 0.516 + ], + "angle": 0, + "content": "Table 2. Sensitivity analysis of user-defined \\( \\alpha \\) under corrupted view conditions (with noise in half of the views) on the CAL dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.4, + 0.558 + ], + "angle": 0, + "content": "5.2.5. Sensitivity Analysis on User-defined \\(\\alpha\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.573, + 0.483, + 0.771 + ], + "angle": 0, + "content": "In this section, we analyze the impact of different user-defined \\(\\alpha\\) values under corrupted view conditions (similar to Table 1) on the CAL dataset. Table 2 shows the effectiveness of our method across various fusion strategies (A-CBF, ABF, and BCF) as \\(\\alpha\\) changes. Generally, incorporating our method (\"w/Ours\") improves classification accuracy in most cases, particularly at higher \\(\\alpha\\) values. Our method consistently achieves higher accuracy with relatively low variance across different \\(\\alpha\\) values, especially in the ABF fusion, indicating strong performance in mitigating noise-induced fluctuations. Notably, when \\(\\alpha = 0.05\\), the accuracy across all three fusion strategies reaches a relatively high level of performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "It is important to note that because noise was randomly sampled with different \\(\\alpha\\) settings in our experiment, the performance of the \"w/o Ours\" condition fluctuates with changes in \\(\\alpha\\). However, despite these fluctuations, our method (\"w/ Ours\") consistently improves accuracy with same \\(\\alpha\\) under identical noisy conditions (i.e., with the same sampled noise), demonstrating its capability to effectively reduce the impact of noise." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.542, + 0.634, + 0.558 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.57, + 0.907, + 0.812 + ], + "angle": 0, + "content": "In this work, we proposed a generalized evidence filtering mechanism to improve the robustness of Trusted Multi-View Classification (TMVC) methods in the wild, particularly under challenging conditions with noisy and corrupted views. We framed the identification of unreliable evidence as a multiple testing problem, using \\( p \\)-values to control the risk of false identification within a minimal range. This approach enables effective post-hoc adjustments during testing. By selectively adjusting evidence weights, our mechanism strengthens fusion robustness and mitigates performance degradation from low-quality data sources. Both theoretical analysis and extensive experiments on multiple real-world datasets demonstrate that our method significantly enhances classification accuracy and stability across various fusion strategies (ABF, A-CBF, and BCF), even in environments with substantial noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.673, + 0.849 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.858, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62472315, 62476165)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15515" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis. In International conference on machine learning, pages 1247-1255. PMLR, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1):289-300, 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.339 + ], + "angle": 0, + "content": "[4] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[5] Li Fei-Fei and Pietro Perona. A bayesian hierarchical model for learning natural scene categories. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 524-531. IEEE, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Pinhan Fu, Xinyan Liang, Yuhua Qian, Qian Guo, Yayu Zhang, Qin Huang, and Ke Tang. Multi-scale features are effective for multi-modal classification: An architecture search viewpoint. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In International Conference on Learning Representations, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[9] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[10] Bingbing Jiang, Chenglong Zhang, Yan Zhong, Yi Liu, Yingwei Zhang, Xingyu Wu, and Weiguo Sheng. Adaptive collaborative fusion for multi-view semi-supervised classification. Information Fusion, 96:37-50, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[11] Bingbing Jiang, Xingyu Wu, Xiren Zhou, Anthony G Cohn, Yi Liu, Weiguo Sheng, and Huanhuan Chen. Semi-supervised multi-view feature selection with adaptive graph learning. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3615-3629, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[12] AUDUN. Jøsang. Subjective Logic: A formalism for reasoning under uncertainty. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[14] Xinyan Liang, Yuhua Qian, Qian Guo, Honghong Cheng, and Jiye Liang. Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9236-9254, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[15] Xinyan Liang, Pinhan Fu, Qian Guo, Keyin Zheng, and Yuhua Qian. Dc-nas: Divide-and-conquer neural architecture search for multi-modal classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13754-13762, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[16] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7585-7593, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[17] Wei Liu, Yufei Chen, Xiaodong Yue, Changqing Zhang, and Shaorong Xie. Safe multi-view deep classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8870-8878, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[18] Wei Liu, Yufei Chen, and Xiaodong Yue. Building trust in decision with conformalized multi-view deep classification. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7278-7287, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.401 + ], + "angle": 0, + "content": "[19] Ying Liu, Lihong Liu, Cai Xu, Xiangyu Song, Ziyu Guan, and Wei Zhao. Dynamic evidence decoupling for trusted multi-view learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7269-7277, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[20] Jueqing Lu, Lan Du, Wray Buntine, Myong Chol Jung, Joanna Dipnall, and Belinda Gabbe. Navigating conflicting views: Harnessing trust for learning. arXiv preprint arXiv:2406.00958, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[21] Jay Patrikar, Sushant Veer, Apoorva Sharma, Marco Pavone, and Sebastian Scherer. Rulefuser: Injecting rules in evidential networks for robust out-of-distribution trajectory prediction. arXiv preprint arXiv:2405.11139, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[22] Zhenwen Ren and Quansen Sun. Simultaneous global and local graph structure preserving for multiple kernel clustering. IEEE transactions on neural networks and learning systems, 32(5):1839-1851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[23] Yaniv Shachor, Hayit Greenspan, and Jacob Goldberger. A mixture of views network with applications to multi-view medical imaging. Neurocomputing, 374:1-9, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[24] Glenn Shafer. A mathematical theory of evidence. Princeton university press, 1976. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.7 + ], + "angle": 0, + "content": "[25] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.906, + 0.745 + ], + "angle": 0, + "content": "[26] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[27] Martijn van Breukelen, Robert PW Duin, David MJ Tax, and JE Den Hartog. Handwritten digit recognition by combined classifiers. Kybernetika, 34(4):381-386, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[28] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. California Institute of Technology, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[29] Meng Wang, Tian Lin, Lianyu Wang, Aidi Lin, Ke Zou, Xinxing Xu, Yi Zhou, Yuanyuan Peng, Qingquan Meng, Yiming Qian, et al. Uncertainty-inspired open set learning for retinal anomaly identification. Nature Communications, 14(1):6757, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15516" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.484, + 0.146 + ], + "angle": 0, + "content": "[30] Weiran Wang, Raman Arora, Karen Livescu, and Jeff Bilmes. On deep multi-view representation learning. In International conference on machine learning, pages 1083-1092. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[31] Chang Xu, Dacheng Tao, and Chao Xu. A survey on multiview learning. arXiv preprint arXiv:1304.5634, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[32] Cai Xu, Jiajun Si, Ziyu Guan, Wei Zhao, Yue Wu, and Xiyue Gao. Reliable conflictive multi-view learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16129-16137, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.235, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[33] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[34] Jinglin Xu, Wenbin Li, Xinwang Liu, Dingwen Zhang, Ji Liu, and Junwei Han. Deep embedded complementary and interactive information for multi-view classification. In Proceedings of the AAAI conference on artificial intelligence, pages 6494-6501, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.349, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[35] Shaoxun Xu, Yufei Chen, Chao Ma, and Xiaodong Yue. Deep evidential fusion network for medical image classification. International Journal of Approximate Reasoning, 150: 188-198, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[36] Xiaoyu Yang, Jie Lu, and En Yu. Adapting multi-modal large language model to concept drift from pre-training onwards. In The Thirteenth International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.462, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[37] Liyuan Zhang, Wei Liu, Yufei Chen, and Xiaodong Yue. Reliable multi-view deep patent classification. Mathematics, 10 (23):4545, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.505, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[38] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15517" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_origin.pdf b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2147927b8b1fdb709c72d12ca4b01da925df3d75 --- /dev/null +++ b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/5a9084f0-d242-4d6e-a2eb-6959dbb6b0a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68ec4cb687fc196f18d5c2651b616092dbd241fedd22237f8c46a128c6cccdfd +size 8508045 diff --git a/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/full.md b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b2d97deb1051a2dde835341f8fd317f724c8c4e6 --- /dev/null +++ b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/full.md @@ -0,0 +1,315 @@ +# Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild + +Wei Liu $^{1}$ , Yufei Chen $^{1*}$ , Xiaodong Yue $^{2,3}$ $^{1}$ School of Computer Science and Technology, Tongji University, Shanghai, China + $^{2}$ Artificial Intelligence Institute, Shanghai University, Shanghai, China + $^{3}$ VLN Lab, Shanghai, China +ldachuan@outlook.com, yufeichen@tongji.edu.cn, yswantfly@shu.edu.cn + +# Abstract + +Trusted multi-view classification (TMVC) addresses variations in data quality by evaluating the reliability of each view based on prediction uncertainty at the evidence level, reducing the impact of low-quality views commonly encountered in real-world scenarios. However, existing TMVC methods often struggle to maintain robustness during testing, particularly when integrating noisy or corrupted views. This limitation arises because the evidence collected by TMVC may be unreliable, frequently providing incorrect information due to complex view distributions and optimization challenges, ultimately leading to classification performance degradation. To enhance the robustness of TMVC methods in real-world conditions, we propose a generalized evidence filtering mechanism that is compatible with various fusion strategies commonly used in TMVC, including Belief Constraint Fusion, Aleatory Cumulative Belief Fusion, and Averaging Belief Fusion. Specifically, we frame the identification of unreliable evidence as a multiple testing problem and introduce $p$ -values to control the risk of false identification. By selectively down-weighting unreliable evidence during testing, our mechanism ensures robust fusion and mitigates performance degradation. Both theoretical guarantees and empirical results demonstrate significant improvements in the classification performance of TMVC methods, supporting their reliable application in challenging, real-world environments. + +# 1. Introduction + +Multi-view classification (MVC) enhances the accuracy of deep learning models by integrating information from multiple data views. By combining insights from diverse views, MVC improves decision-making in complex domains [7, 10, 11, 14, 15, 25, 31]. However, traditional + +![](images/f7ebd7586ae13d0d54a2da1a914d155d44851fd2267499c93d4a7a00c9eb0057.jpg) +Figure 1. Classification accuracy degradation with noise (w/) compared to no noise (w/o) across three fusion strategies (BCF, A-CBF, and ABF) on six datasets. Noise is introduced by replacing half of the views to simulate data collected from malfunctioning sensors. + +MVC approaches often assume that each view consistently provides high-quality, reliable information—an assumption that is frequently unrealistic in real-world applications. In practical scenarios, data quality and reliability across views can vary significantly, leading to contradictory or misleading information that compromises prediction accuracy and reliability. As a result, conventional MVC models may struggle to reliably identify the ground truth, particularly when dealing with noisy, corrupted, or otherwise low-quality data views [38]. + +To address these limitations, trusted multi-view classification (TMVC) [8, 9, 16, 17, 32] has emerged as an effective solution. Unlike traditional MVC, TMVC evaluates the trustworthiness of each view by leveraging prediction uncertainty derived from view-specific evidence, enabling it to identify unreliable views and prioritize informative ones in the final predictions. This approach has proven highly effective in critical applications such as medical imaging diagno + +sis [29] and autonomous driving [21], where distinguishing between reliable and unreliable predictions is essential. + +In TMVC, evidence from each sample for each view contributes to its prediction and is also used to calculate a view-specific uncertainty score $u$ , which reflects the confidence in the quality of that view for the given sample. A high uncertainty score (i.e., $u = 1$ ) indicates a substantial deviation from the training data, potentially due to noise or corruption. In such cases, evidence supporting all classes is minimized and fused using one of three strategies: Belief Constraint Fusion (BCF) [8, 9], Aleatory Cumulative Belief Fusion (A-CBF) [16-18, 37], and Averaging Belief Fusion (ABF) [32]. However, despite their success in controlled settings, these methods often struggle to remain robust in the wild. Complex and diverse data view distributions, combined with varying levels of optimization, can distort uncertainty estimates, resulting in unreliable evidence that misleads the model and biases predictions toward incorrect classes. As shown in Figure 1, introducing random noise (e.g., replacing half of the views in a dataset) to simulate sensor failure conditions leads to a significant drop in classification accuracy across all fusion strategies on six real-world datasets, underscoring the risks of decision-making in unpredictable and noisy environments. + +To address this challenge, we propose a generalized and robust testing-time evidence filtering mechanism to enhance the robustness of TMVC methods in real-world, unpredictable environments. Specifically, we frame the identification of unreliable evidence as a multiple testing problem, introducing $p$ -values to control the risk of false identification within a small margin. This approach enables effective post-hoc adjustments during testing. By selectively down-weighting evidence from unreliable views, our mechanism enhances fusion robustness and mitigates performance degradation caused by low-quality data sources. In summary, our contributions are as follows: + +1. We propose a testing-time evidence filtering mechanism that is compatible with multiple fusion strategies, enhancing the robustness of existing TMVC methods. +2. Our approach frames the identification of unreliable evidence as a multiple testing problem and selectively adjusts evidence weights during testing, mitigating the negative impact of unreliable views on fusion. +3. Theoretical and empirical analyses validate the effectiveness of the proposed method in improving both classification accuracy and robustness. + +# 2. Related Works + +# 2.1. Multi-view Classification + +Over the past few years, numerous multi-view classification (MVC) approaches [1, 2, 16, 22, 25, 26, 30, 34, 36] have been developed, particularly those leveraging deep neural + +networks (DNNs). These methods aim to learn shared representations by integrating information from different types of views, resulting in improved performance. By combining insights from diverse views, MVC has achieved remarkable gains and has been widely applied across various tasks [23, 35]. However, in real-world scenarios, the quality of data across views can vary significantly, introducing noise and inconsistencies that undermine both classification accuracy and reliability—factors that are especially critical in areas like medical diagnosis and autonomous driving. + +# 2.2. Trusted Multi-view Classification + +To address these challenges, Trusted Multi-View Classification (TMVC) has emerged as a framework that assesses the reliability of each view through prediction uncertainty, allowing models to prioritize reliable views and down-weight unreliable ones [8, 9, 16, 17, 32, 33]. TMVC methods often rely on Subjective Logic (SL) [12] to model view-specific predictions as multinomial opinions, which are then fused using belief combination rules tailored to the context [12]. For instance, extensions of Dempster-Shafer theory [24] were firstly introduced by [8, 9] to facilitate evidence fusion in multi-view settings. Other fusion approaches, such as Cumulative Belief Fusion [16, 17] and Averaging Belief Fusion [32], have also been proposed to further enhance the effectiveness of TMVC. Despite their success, all these methods fall short of achieving robust performance during testing, as the magnitude of estimated uncertainty is highly sensitive to the diverse distributions of data views and varying levels of model optimization, which significantly degrades classification accuracy in real-world scenarios. In contrast, our method improves testing-time robustness through an evidence filtering mechanism, thereby reducing the risks associated with final decision-making. + +# 3. Overview and Analysis of TMVC + +# 3.1. Generalized Framework of TMVC + +Given the training data $D_{train} = \left\{\left\{\boldsymbol{x}_i^v\right\}_{v=1}^V, \boldsymbol{y}_i\right\}_{i=1}^N$ , where $N$ is the number of samples, each sample has $V$ views with a one-hot label $\boldsymbol{y}_i = \left\{y_{ik}\right\}_{k=1}^K$ for a $K$ -class classification task. TMVC uses subjective logic [12] to learn a function $f$ that maps $\left\{\boldsymbol{x}_i^v\right\}_{v=1}^V$ to $\boldsymbol{y}_i$ along with a corresponding overall uncertainty score $u_i$ , which quantifies prediction confidence at the evidence level. + +Specifically, let $f_{v}(\boldsymbol{x}_{i}^{v})$ denote the output of the penultimate (logits) layer of the view-specific neural network for sample $\boldsymbol{x}_{i}^{v}$ in a single view. Instead of the standard softmax function for a single categorical prediction, TMVC employs an alternative activation function $a(\cdot)$ (e.g., ReLU) to capture non-negative evidence $e_{i}^{v} = a(f_{v}(\boldsymbol{x}_{i}^{v}))$ , where $e_{i}^{v} = \{e_{ik}^{v}\}_{k=1}^{K}$ . The view-specific prediction uncertainty is + +given by + +$$ +u _ {i} ^ {v} = K / S _ {i} ^ {v}, \tag {1} +$$ + +where $S_{i}^{v} = \sum_{k=1}^{K} e_{ik}^{v} + K$ . This reflects the lack of total evidence across all classes in a single view. + +TMVC then applies evidence fusion strategies, such as BCF, A-CBF, and ABF, to integrate the evidence from each view into a unified multi-view evidence representation $e_i = \{e_{ik}\}_{k=1}^K$ . These three fusion rules for two views are commonly expressed as follows [20]: + +$$ +\begin{array}{l} A - C B F \left(\boldsymbol {e} _ {i} ^ {1}, \boldsymbol {e} _ {i} ^ {2}\right) = \boldsymbol {e} _ {i} ^ {1} + \boldsymbol {e} _ {i} ^ {2}. \\ A B F \left(e _ {i} ^ {1}, e _ {i} ^ {2}\right) = \left(e _ {i} ^ {1} + e _ {i} ^ {2}\right) / 2. \\ B C F \left(\boldsymbol {e} _ {i} ^ {1}, \boldsymbol {e} _ {i} ^ {2}\right) = \boldsymbol {e} _ {i} ^ {1} + \boldsymbol {e} _ {i} ^ {2} + \boldsymbol {e} _ {i} ^ {1} \boldsymbol {e} _ {i} ^ {2} / K. \tag {2} \\ \end{array} +$$ + +These fusion rules can also be extended to handle multiple views with ease. + +In the training phase, TMVC links the generated evidence to the concentration parameter of the Dirichlet distribution and optimizes model parameters using a task-specific loss function. This loss primarily consists of a cross-entropy term, which maximizes the evidence for the correct label of each sample, and a Kullback-Leibler (KL)-divergence term, which drives the evidence for incorrect labels toward zero, as described in [8]. The final multi-view prediction $\pmb{p}_i = \{p_{ik}\}_{k=1}^K$ is obtained as the expectation of the learned Dirichlet distribution: + +$$ +p _ {i k} = \left(e _ {i k} + 1\right) / S _ {i}, \tag {3} +$$ + +where $S_{i} = \sum_{k = 1}^{K}e_{ik} + K$ , and the overall uncertainty $u_{i}$ is given by $u_{i} = K / S_{i}$ . Finally, the model's best performance on the validation dataset $D_{val}$ is used to select the optimal parameters for testing. + +# 3.2. Limitations of TMVC at Testing Time + +Despite its success across various domains [21, 29], TMVC methods face significant limitations at testing time, particularly in real-world scenarios with diverse data view distributions and varying levels of model optimization. These factors often degrade TMVC's robustness, especially when handling low-quality data views. + +Specifically, given a multi-view test dataset $D_{test} = \left\{\left\{\boldsymbol{x}_j^v\right\}_{v=1}^V\right\}_{j=1}^M$ , a trained TMVC model with $V$ view-specific networks provides uncertainty estimates for each view of a testing sample $\boldsymbol{x}_j^v$ . Ideally, a high uncertainty score (e.g., $u_j^v = 1$ ) for a low-quality view would reduce its evidence contribution across all classes to zero ( $e_j^v = 0$ ), minimizing its impact during fusion. However, these uncertainty estimates are sensitive to data distributions and + +![](images/e8f495d20d6f49ae6ff828e7a9737257e9e39f68115c8c8f25415bd8c9ebbef1.jpg) +Figure 2. Density distribution of uncertainty for noisy view results under noise corruption and the corresponding density distribution of remaining unreliable evidence across three fusion strategies on the CUB dataset. + +![](images/3ac0fa0a6cbd09095427c090e1b652a17dfd50ffe197cbe7e7ba6a32a69a0d0f.jpg) + +model optimization levels, making it challenging to reliably achieve $u_{j}^{v} = 1$ for extremely noisy or corrupted views. + +For example, in the CUB dataset, even when one view is entirely corrupted by noise (i.e., features of one view are replaced by random noise), the uncertainty score for that view typically falls within the range [0.7, 1.0], leaving residual evidence mostly between [0.0, 4.0] (as illustrated in Figure 2) that contributes to fusion and can potentially bias predictions. Since this view lacks useful information, its residual evidence is unreliable and negatively impacts multi-view classification accuracy, as shown in Figure 1. Moreover, the overall uncertainty score derived from fusion strategies like BCF and A-CBF is often lower as additional views are fused (as demonstrated in [9, 16]). Even with ABF fusion, the overall uncertainty remains below the highest view-specific uncertainty [32]. As a result, decision-makers may struggle to detect potentially misclassified samples caused by low-quality views. These limitations underscore the challenges TMVC faces in real-world conditions, where testing-time robustness is crucial for reliable decision-making. + +# 4. Proposed Method + +In this section, we present a strategy to enhance the robustness of TMVC methods at testing time. Performance degradation in TMVC methods often arises from the negative impact of such low-quality views, which typically exhibit high prediction uncertainty during testing, compared to the reliable views seen during training. An intuitive solution to this problem is to reduce the contribution of unreliable evidence from low-quality views during multi-view fusion. To address this, we propose a generalized evidence filtering mechanism that is compatible with various TMVC fusion rules. The mechanism consists of two key components: (1) identifying unreliable evidence collected from low-quality views, and (2) assigning lower weights to minimize their impact on the final decision. + +# 4.1. Reliable Evidence Discovery in the Wild + +Identifying unreliable evidence is equivalent to discovering reliable evidence. The primary challenge in this process is controlling the false discovery rate (FDR) [3], since the true label is unknown. We address this challenge by framing reliable evidence discovery as a multiple hypothesis testing problem from a statistical learning perspective. + +Given a multi-view test dataset $D_{test}$ , the objective is to identify as many reliable pieces of evidence as possible while ensuring that the FDR is maintained below a user-defined threshold $\alpha \in (0,1)$ . Let us define a discovery function $\Lambda(\kappa^v)$ , which selects a subset $S^v$ of view-specific evidence for the test points of interest. Evidences with smaller values of $u_j^v$ , i.e., $u_j^v \leq \kappa^v$ , is likely to be included in the subset, where $\kappa^v$ is a threshold determined by the discovery function. The FDR for this function is then defined as: + +$$ +\mathrm {F D R} \left(\Lambda \left(\kappa^ {v}\right)\right) := \mathbb {E} \left[ \frac {\sum_ {j \in D _ {t e s t}} \mathbb {1} \left\{j \in S ^ {v} , u _ {j} ^ {v} > \kappa^ {v} \right\}}{\max \left\{1 , | S ^ {v} | \right\}} \right]. (4) +$$ + +Our goal is to ensure that the subset $S^v$ identified by $\Lambda(\kappa^v)$ satisfies the following error-control property: + +Definition 1 (Risk-controlling discovery). Let $\kappa^v$ be a random variable. We define $\Lambda(\kappa^v)$ as an $(\alpha, \delta)$ -risk-controlling discovery (RCD) function if + +$$ +\mathbb {P} \left(\operatorname {F D R} \left(\Lambda \left(\kappa^ {v}\right)\right) \leq \alpha\right) \tag {5} +$$ + +holds with probability at least $1 - \delta$ + +Here, $\delta \in (0,1)$ is the confidence level. To meet this risk control criterion, we need to identify a suitable $\Lambda (\kappa^v)$ that adapts to diverse data view distributions encountered in the wild. For this purpose, we gather statistical evidence via hypothesis testing, where we consider the view-specific evidence of a test sample to be discovered (i.e, $u_{j}^{v}\leq \kappa^{v}$ ) by $\Lambda (\kappa^v)$ as the null hypothesis. Rejecting the null hypothesis reflects that $u_{j}^{v} > \kappa^{v}$ , and therefore the evidence is unreliable. + +Under this framework, discovering reliable evidence boils down to simultaneously testing the $M$ hypotheses. To do this, for any arbitrary view-specific evidence $\pmb{e}_j^v$ , we compute the $p$ -value: + +$$ +p \text {- v a l u e} \left(\boldsymbol {e} _ {j} ^ {v}\right) = \frac {1 + \sum_ {l \in D _ {v a l}} \mathbb {1} \left\{u _ {j} ^ {v} \geq u _ {l} ^ {v} \right\}}{\left| D _ {v a l} \right| + 1}. \tag {6} +$$ + +It is clear that the $p$ -value quantifies how representative the evidence $e_j^v$ is of reliable evidence. If the $p$ -value is less than $\alpha$ , the null hypothesis of exchangeability with the validation data is rejected, deeming the evidence unreliable. We obtain marginally valid $p$ -values, which satisfy: + +$$ +\mathbb {P} \left(p - \text {v a l u e} \left(\boldsymbol {e} _ {j} ^ {v}\right) \leq \alpha\right) \leq \alpha , \tag {7} +$$ + +indicating that the probability of falsely identifying unreliable evidence is controlled by the small margin $\alpha$ . To achieve this, we set a view-specific, distribution-free threshold using the Benjamini-Hochberg (BH) procedure [3]: + +$$ +\kappa^ {v} = Q \left(\left\{u _ {l} ^ {v} \right\} _ {l \in D _ {v a l}}, \alpha\right), \tag {8} +$$ + +where $Q(\cdot, \alpha)$ denotes the $\lceil (1 - \alpha)(|D_{val}| + 1) \rceil$ -th smallest value in $\{u_l^v\}_{l \in D_{val}}$ . Then, we construct the reliable evidence set using $\Lambda(\kappa^v)$ : + +$$ +S ^ {v} = \left\{j \in D _ {\text {t e s t}}: u _ {j} ^ {v} \leq \kappa^ {v} \right\}. \tag {9} +$$ + +Finally, we have the following proposition: + +Proposition 1. $\Lambda (\kappa^v)$ is an $(\alpha ,\delta)$ -risk-controlling discovery (RCD) function. + +# 4.2. Mitigating the Influence of Unreliable Evidence + +After identifying the reliability of view-specific evidence, we aim to mitigate the impact of unreliable evidence during multi-view fusion. To achieve this, we define the filtered evidence for each testing sample as follows: + +$$ +\widehat {\boldsymbol {e}} _ {j} ^ {v} = \left\{ \begin{array}{c c} \boldsymbol {e} _ {j} ^ {v}, & j \in S ^ {v}. \\ \alpha \cdot \boldsymbol {e} _ {j} ^ {v}, & j \notin S ^ {v}. \end{array} \right. \tag {10} +$$ + +Here, $\alpha$ is a user-defined parameter that controls the weight assigned to unreliable evidence, directly influencing the FDR control. Smaller values of $\alpha$ effectively reduce the influence of low-quality views, minimizing the risk of unreliable evidence negatively impacting the final prediction. This tight control is especially valuable in high-stakes applications where high confidence and accuracy in predictions is crucial. On the other hand, if $\alpha$ is set to a larger value, the mechanism retains more influence from low-quality views, which may be beneficial when such views contain partially valuable information despite high uncertainty. + +Moreover, the choice of $\alpha$ balances FDR control and information retention, allowing flexibility in managing the contribution of uncertain views to the fusion process. A lower $\alpha$ tightens FDR control, reducing the likelihood of unreliable evidence being included and enhancing prediction reliability. Importantly, we do not fully exclude unreliable evidence, as even noisy views may contain useful information for classification. Instead, we adopt a weighted filtering approach that retains the potential benefits of low-quality views while mitigating their negative effects. This soft filtering strategy allows the model to remain adaptive and robust under real-world conditions. + +After filtering, the view-specific evidence is fused using the selected fusion rules (e.g., BCF, A-CBF, or ABF), achieving an adaptable and robust multi-view fusion process that is compatible with various TMVC fusion strategies. This leads to the following propositions: + +Proposition 2. Applying evidence filtering improves the classification performance of TMVC fusion strategies (e.g., BCF, A-CBF, or ABF), under the assumption that unreliable evidence provides incorrect information in the wild. + +Proposition 3. The overall uncertainty of multi-view results generated by TMVC fusion strategies with evidence filtering for unreliable evidence will exhibit greater uncertainty than fusion without evidence filtering. + +Both Propositions 2 and 3 guarantee the robustness of our method when confronted with unreliable evidence (i.e., evidence from low-quality views) in the wild. Note that unreliable evidence providing correct predictions by chance (a counterintuitive situation) may hurt classification performance in our method, particularly when evidence from other views is incorrect. To address this, our method exhibits greater overall uncertainty for these multi-view results, as shown in Proposition 3, which can alert humans to be cautious with these results. + +In summary, our evidence filtering mechanism improves testing-time robustness by adjusting the weights of view-specific evidence based on reliability. This enables TMVC methods to perform reliably in environments with data inconsistencies. This testing-time robustness is crucial for dependable deployment in high-stakes, real-world applications. + +# 4.3. Computational Complexity Analysis + +Our method introduces a generalized framework that can be applied to various DNNs for multi-view testing. Given a single test pass with complexity $O(T_{test})$ on an arbitrary TMVC model, the overall computational complexity is: + +$$ +O (L \log L) + O (M \cdot T _ {t e s t}), \tag {11} +$$ + +where $L$ and $M$ represent the sizes of the validation and test datasets, respectively. The term $O(L\log L)$ reflects the complexity involved in threshold calculation, primarily due to sorting the uncertainty scores within the validation set to determine an appropriate threshold for filtering unreliable evidence. The term $O(M\cdot T_{test})$ represents the complexity of running one test pass on $M$ samples in the testing set. In most practical scenarios, the term $O(M\cdot T_{test})$ will typically dominate, especially for large-scale DNNs, unless the validation set is extremely large or the neural networks are shallow and computationally lightweight. + +# 5. Experiments + +In this section, we validate the effectiveness of our proposed method on six real-world datasets. Each experiment is run 10 times, and we report the mean and standard deviation of accuracy (ACC) to ensure reliability. The extensive results demonstrate the clear advantage of our proposed method. + +# 5.1. Experimental Setup + +# 5.1.1. Datasets + +We conduct experiments on six real-world multi-view datasets as follows: CUB [28]: This dataset contains 11,788 bird images from 200 distinct categories, each paired with text descriptions. For our experiments, we utilize the first 10 categories. Image features are extracted using GoogleNet, while text features are extracted using doc2vec. SCENE [5]: This dataset comprises 4,485 images from 15 indoor and outdoor scene categories. Multiple views are represented using three feature types: GIST, PHOG, and LBP. CAL [6]: The CAL dataset consists of 8,677 images from 101 categories. We focus on the first 20 categories and extract six types of features: GABOR, WAVELET-MOMENTS, CENHIST, HOG, GIST, and LBP. HAND [27]: This dataset contains 2,000 samples across 10 digit classes (from '0' to '9'), with 200 samples per class. Six feature descriptors (PIX, FOU, FAC, ZER, KAR, MOR) are extracted as multiple views. PIE: This dataset consists of 680 facial images from 68 individuals. Three types of features are extracted: intensity, LBP, and Gabor. NUS: A subset of the NUS-WIDE dataset [4], it contains 30,000 images spanning 31 object categories, represented in five views, including a 64-D color histogram, 225-D blockwise color moments, 144-D color correlogram, 73-D edge direction histogram, and 128-D wavelet texture. + +# 5.1.2. Compared Methods + +Since our method is a generalized framework applicable to various trusted multi-view classification methods, we primarily evaluate its effectiveness by comparing baseline methods with and without our approach in these experiments. We apply our method to three fusion rules: BCF, A-CBF, and ABF, using the general loss function proposed in the work of [8]. We then integrate our method with state-of-the-art multi-view deep classification methods, including TMC [8], TMDOA [16], ETMC [9], SMDC [17], ECML [32], and CCML [19]. These methods use the same neural network architecture as ours, ensuring a consistent basis for comparison. + +# 5.1.3. Implementations + +For all datasets, we split the data into training (70%), validation (10%), and testing (20%) sets. Fully connected networks were used as view-specific backbones. To obtain a trained TMVC model, we employed the Adam optimizer [13] with an $l_{2}$ -norm regularization of $1e^{-5}$ and a learning rate of $1e^{-3}$ . During testing, we set $\alpha = 0.05$ by default, ensuring an FDR below 5% unless otherwise specified. The model was implemented in PyTorch and run on a GeForce RTX 4090 GPU with 24GB of memory. + +Note: Optimization during training is not a primary focus of this work; hence, finding the optimal model parameters + +
Fusion with Clean Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB98.75±0.4298.75±0.4297.50±1.6798.33±1.6797.08±0.4297.92±0.42
SCENE70.94±0.0670.69±0.3171.63±0.7571.31±1.0670.50±0.1370.25±0.25
CAL98.50±0.5098.38±0.3897.63±0.3897.25±0.2598.75±0.5098.50±0.25
HAND97.88±0.1398.25±0.2599.00±0.0599.00±0.7598.88±0.3899.25±0.50
PIE98.08±1.1598.46±0.7798.85±0.3898.85±0.3897.69±1.5498.08±1.15
NUS53.85±1.0054.05±0.8055.35±0.3055.43±0.0857.75±0.7557.88±0.78
Fusion with Corrupted Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB64.17±5.0080.00±0.8356.67±6.6779.17±1.6765.33±3.3383.75±1.25
SCENE46.00±3.3853.38±2.3150.75±2.0051.38±1.9449.25±0.6352.75±0.75
CAL56.75±1.8876.75±5.0070.25±0.7571.50±2.0073.25±0.1375.50±2.00
HAND64.50±1.6370.00±1.8870.00±0.7570.75±0.6375.75±1.7578.25±1.75
PIE89.77±0.0196.54±0.3888.46±3.8590.77±2.3188.08±0.3891.54±0.01
NUS27.70±1.6037.70±0.8536.73±1.2238.95±1.2037.00±1.6040.63±1.38
+ +Table 1. Classification performance (ACC (%)) for fusion with clean views and corrupted views, based on three fusion strategies, with or without our method. + +was not the primary goal. Instead, we provide comparative results with and without our method under identical settings (same random seed, dataset splits, optimization parameters, etc.) to evaluate the effectiveness of our method. + +# 5.2. Quantitative Analysis + +# 5.2.1. Effect of Our Method on Basic Fusion Strategies + +In this evaluation, we assess the effectiveness of our proposed method across three fusion strategies (ABF, A-CBF, and BCF) under both clean and corrupted view conditions. Initially, we applied our method to these fusion strategies in the context of clean views, where no noise was introduced. As shown in Table 1, our method generally improves classification accuracy across most datasets, even in noise-free environments. This demonstrates that our approach enhances model stability in both noisy and normal conditions. Minor performance decreases observed in the SCENE and CAL datasets with clean views may be attributed to our method's down-weighting of higher uncertainty evidence for some samples in certain views. These higher uncertainty samples could otherwise positively impact classification in some cases. + +Under corrupted view conditions, where random noise in the range of $[-5, 5]$ is added to half of the views, our method significantly boosts accuracy across all datasets. For example, on the CUB dataset with A-CBF fusion, our method achieves nearly a $23\%$ improvement. For datasets like HAND, the modest performance improvement under corrupted conditions can be explained by the high uncertainty values (close to 1) associated with low total evidence (close to 0) in the corrupted HAND views. This naturally limits the impact of unreliable evidence during fusion. + +Overall, these results underscore our method's robustness in enhancing existing TMVC fusion strategies, particularly in challenging, noisy environments, while also maintaining stable performance in clean data conditions. + +# 5.2.2. Effect on Popular TMVC Methods + +We also demonstrate the effectiveness of our method when applied to various popular TMVC methods under corrupted view conditions (similar to Table 1) across six datasets. As shown in Figure 3, each radar chart represents one popular TMVC method, displaying classification accuracy under two conditions: with our method (w/ Ours, shown in red) and without our method (w/o Ours, shown in blue). Across all six TMVC methods, our method consistently improves performance under corrupted view conditions, as indicated by the expanded area covered by the red line compared to the blue line. The radar charts clearly show that our method effectively boosts classification accuracy on each dataset when faced with noisy and corrupted views. This improvement underscores the robustness of our approach in handling unreliable or degraded views across various TMVC frameworks. Particularly in challenging datasets like CUB, where noise can severely degrade model accuracy, our method provides a noticeable performance enhancement, indicating that it successfully mitigates the impact of corrupted views on multi-view classification accuracy. Overall, these results confirm the applicability and robustness of our method when integrated into diverse TMVC methods. + +# 5.2.3. Sensitivity Analysis on Noisy Ratio + +In this experiment, we demonstrate the effectiveness of our method under varying noisy ratios in corrupted view conditions. We randomly select samples from half of + +![](images/de8b0859a823f81376e60849e79864afd769a43cf65dfb0f5ce4398470e40b84.jpg) + +![](images/7c0f01b2ac4e5aea2576ccae5ba8e155aee3b56edfd08d00ce4a019b2eda08c6.jpg) + +![](images/ec9392ed813ce6d285679c47f117acb818d793c13282c49237d52e122bd0f752.jpg) + +![](images/7e68622b40361e16bfc07039ed007e0dd24922d600fdd02812c17f6a5a936a6c.jpg) +Figure 3. Comparison of classification accuracy between TMVC methods with and without our method. + +![](images/0e04ec4757e5f8704007ad6dd906c4305d2c12289ec6033ef514cc716d7483e5.jpg) + +![](images/443332ab02c5a57058db11cb8c6f9c6b62440267a846ab09bcff5ad896ef6c82.jpg) + +![](images/e783faeb1e085466394b96283e4b7bae25f5c4b3e1a0e0265ff22f0013547701.jpg) +Figure 4. Sensitivity analysis of accuracy with varying noisy ratios under corrupted view conditions (where half of the views contain noise) on the CUB dataset. + +![](images/b775281c623353fb681beb21ec9dbf847e8eeed30566d7fa47a22e37fd0bff38.jpg) + +![](images/4b68fbfdd59b4df89a0f34e4d8851d2ce29d9830274b356138a676c8b687e5fa.jpg) + +the views in the CUB dataset and apply different noisy ratios to create polluted samples. As shown in Figure 4, the experimental results reveal that the proposed method significantly enhances the robustness of all three fusion strategies (ABF, A-CBF, and BCF) against varying noisy ratios. As the noisy ratio increases, both the baseline and proposed methods experience a decline in accuracy. However, the proposed method consistently outperforms the baseline across all fusion rules. Notably, the accuracy decline is slower for the method with the proposed fusion approach, and it is accompanied by reduced variability, indicating a greater resilience to noise. These findings highlight the effectiveness of the proposed method in maintaining both accuracy and stability, especially in high-noise environments. This makes it a promising solution for noise-robust multi-view fusion tasks. + +# 5.2.4. Robustness Analysis under Varying Noise Intensity + +We analyze the robustness of our method under varying noise intensities. In this part, noise vectors (denoted by $\epsilon$ ) are sampled from a Gaussian distribution $\mathcal{N}(\mathbf{0}, \mathbf{I})$ . These noise vectors are then multiplied by an intensity factor $\eta$ and added to half of the views in the test samples from the CUB dataset, i.e., $\tilde{\boldsymbol{x}}^v = \boldsymbol{x}^v + \eta \epsilon^v$ . Figure 5 presents the robustness analysis under varying noise intensities. The results indicate that the proposed method significantly enhances the robustness of the three fusion strategies (ABF, A-CBF, and BCF) against different noise levels. Across all noise intensities, our method consistently outperforms the baseline, showing more stable accuracy and reduced variability, particularly at higher noise levels. These findings highlight the effectiveness of our method in improving performance and stability in noisy environments. + +![](images/c1042192b852cb7f0395c7ee05a814731e2728323264ba4ba0b942daa0394054.jpg) + +![](images/12237ecb883beb199a255e761e85ff45d52db53705b2472f38cec5bd92dda42e.jpg) + +![](images/388b04a61267d04f5e800150f1a041205f923b7fbc7f43d56d98ebc6da9db48e.jpg) + +Figure 5. Robustness analysis under varying noise intensities on the CUB dataset. + +
A-CBFABFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
α =0.00569.60 ± 2.7069.60 ± 2.7053.75 ± 1.7553.75 ± 1.7570.00 ± 1.4070.10 ± 1.40
α =0.00672.50 ± 0.3072.60 ± 0.3049.88 ± 0.6350.63 ± 0.1369.50 ± 1.0069.50 ± 1.00
α =0.00769.70 ± 1.4069.70 ± 1.4058.63 ± 6.3858.88 ± 6.6371.30 ± 1.6071.40 ± 1.60
α =0.00871.70 ± 1.0071.70 ± 1.0051.63 ± 0.6351.75 ± 0.7570.70 ± 0.7071.00 ± 0.70
α =0.00970.00 ± 1.3070.10 ± 1.3049.63 ± 1.3850.63 ± 0.3869.50 ± 1.0069.80 ± 1.00
α =0.01070.50 ± 1.8070.60 ± 1.8054.18 ± 4.6354.25 ± 2.5070.00 ± 0.8070.20 ± 0.80
α =0.02072.00 ± 1.3072.20 ± 1.3052.25 ± 0.0052.75 ± 0.2571.70 ± 1.3071.90 ± 1.30
α =0.03070.30 ± 1.3070.40 ± 1.3054.75 ± 1.7558.63 ± 1.3869.50 ± 0.8069.80 ± 0.80
α =0.04070.70 ± 1.8070.90 ± 1.8055.13 ± 2.3867.00 ± 1.0072.20 ± 1.0072.50 ± 1.00
α =0.05071.80 ± 1.1071.90 ± 1.1055.13 ± 0.3878.63 ± 1.3873.00 ± 0.9075.30 ± 0.90
α =0.06069.70 ± 1.2069.80 ± 1.2057.50 ± 2.5068.13 ± 1.3871.50 ± 1.2071.70 ± 1.20
α =0.07070.30 ± 1.3070.40 ± 1.3055.88 ± 2.6376.50 ± 1.0069.80 ± 1.0070.00 ± 1.00
α =0.08070.20 ± 1.3070.30 ± 1.3056.50 ± 1.2578.63 ± 0.1370.50 ± 1.3070.70 ± 1.30
α =0.09072.20 ± 1.7072.40 ± 1.7056.25 ± 3.5076.88 ± 2.3871.70 ± 1.0072.00 ± 1.00
α =0.10069.50 ± 1.4069.70 ± 1.4055.13 ± 1.6378.00 ± 1.2570.70 ± 1.2070.90 ± 1.20
+ +Table 2. Sensitivity analysis of user-defined $\alpha$ under corrupted view conditions (with noise in half of the views) on the CAL dataset. + +# 5.2.5. Sensitivity Analysis on User-defined $\alpha$ + +In this section, we analyze the impact of different user-defined $\alpha$ values under corrupted view conditions (similar to Table 1) on the CAL dataset. Table 2 shows the effectiveness of our method across various fusion strategies (A-CBF, ABF, and BCF) as $\alpha$ changes. Generally, incorporating our method ("w/Ours") improves classification accuracy in most cases, particularly at higher $\alpha$ values. Our method consistently achieves higher accuracy with relatively low variance across different $\alpha$ values, especially in the ABF fusion, indicating strong performance in mitigating noise-induced fluctuations. Notably, when $\alpha = 0.05$ , the accuracy across all three fusion strategies reaches a relatively high level of performance. + +It is important to note that because noise was randomly sampled with different $\alpha$ settings in our experiment, the performance of the "w/o Ours" condition fluctuates with changes in $\alpha$ . However, despite these fluctuations, our method ("w/ Ours") consistently improves accuracy with same $\alpha$ under identical noisy conditions (i.e., with the same sampled noise), demonstrating its capability to effectively reduce the impact of noise. + +# 6. Conclusion + +In this work, we proposed a generalized evidence filtering mechanism to improve the robustness of Trusted Multi-View Classification (TMVC) methods in the wild, particularly under challenging conditions with noisy and corrupted views. We framed the identification of unreliable evidence as a multiple testing problem, using $p$ -values to control the risk of false identification within a minimal range. This approach enables effective post-hoc adjustments during testing. By selectively adjusting evidence weights, our mechanism strengthens fusion robustness and mitigates performance degradation from low-quality data sources. Both theoretical analysis and extensive experiments on multiple real-world datasets demonstrate that our method significantly enhances classification accuracy and stability across various fusion strategies (ABF, A-CBF, and BCF), even in environments with substantial noise. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (No. 62472315, 62476165). + +# References + +[1] Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis. In International conference on machine learning, pages 1247-1255. PMLR, 2013. 2 +[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 2 +[3] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1):289-300, 1995. 4 +[4] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 5 +[5] Li Fei-Fei and Pietro Perona. A bayesian hierarchical model for learning natural scene categories. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 524-531. IEEE, 2005. 5 +[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5 +[7] Pinhan Fu, Xinyan Liang, Yuhua Qian, Qian Guo, Yayu Zhang, Qin Huang, and Ke Tang. Multi-scale features are effective for multi-modal classification: An architecture search viewpoint. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1 +[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In International Conference on Learning Representations, 2021. 1, 2, 3, 5 +[9] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1, 2, 3, 5 +[10] Bingbing Jiang, Chenglong Zhang, Yan Zhong, Yi Liu, Yingwei Zhang, Xingyu Wu, and Weiguo Sheng. Adaptive collaborative fusion for multi-view semi-supervised classification. Information Fusion, 96:37-50, 2023. 1 +[11] Bingbing Jiang, Xingyu Wu, Xiren Zhou, Anthony G Cohn, Yi Liu, Weiguo Sheng, and Huanhuan Chen. Semi-supervised multi-view feature selection with adaptive graph learning. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3615-3629, 2024. 1 +[12] AUDUN. Jøsang. Subjective Logic: A formalism for reasoning under uncertainty. Springer, 2018. 2 +[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[14] Xinyan Liang, Yuhua Qian, Qian Guo, Honghong Cheng, and Jiye Liang. Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9236-9254, 2021. 1 + +[15] Xinyan Liang, Pinhan Fu, Qian Guo, Keyin Zheng, and Yuhua Qian. Dc-nas: Divide-and-conquer neural architecture search for multi-modal classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13754-13762, 2024. 1 +[16] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7585-7593, 2022. 1, 2, 3, 5 +[17] Wei Liu, Yufei Chen, Xiaodong Yue, Changqing Zhang, and Shaorong Xie. Safe multi-view deep classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8870-8878, 2023. 1, 2, 5 +[18] Wei Liu, Yufei Chen, and Xiaodong Yue. Building trust in decision with conformalized multi-view deep classification. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7278-7287, 2024. 2 +[19] Ying Liu, Lihong Liu, Cai Xu, Xiangyu Song, Ziyu Guan, and Wei Zhao. Dynamic evidence decoupling for trusted multi-view learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7269-7277, 2024. 5 +[20] Jueqing Lu, Lan Du, Wray Buntine, Myong Chol Jung, Joanna Dipnall, and Belinda Gabbe. Navigating conflicting views: Harnessing trust for learning. arXiv preprint arXiv:2406.00958, 2024.3 +[21] Jay Patrikar, Sushant Veer, Apoorva Sharma, Marco Pavone, and Sebastian Scherer. Rulefuser: Injecting rules in evidential networks for robust out-of-distribution trajectory prediction. arXiv preprint arXiv:2405.11139, 2024. 2, 3 +[22] Zhenwen Ren and Quansen Sun. Simultaneous global and local graph structure preserving for multiple kernel clustering. IEEE transactions on neural networks and learning systems, 32(5):1839-1851, 2020. 2 +[23] Yaniv Shachor, Hayit Greenspan, and Jacob Goldberger. A mixture of views network with applications to multi-view medical imaging. Neurocomputing, 374:1-9, 2020. 2 +[24] Glenn Shafer. A mathematical theory of evidence. Princeton university press, 1976. 2 +[25] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2 +[26] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2 +[27] Martijn van Breukelen, Robert PW Duin, David MJ Tax, and JE Den Hartog. Handwritten digit recognition by combined classifiers. Kybernetika, 34(4):381-386, 1998. 5 +[28] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. California Institute of Technology, 2011. 5 +[29] Meng Wang, Tian Lin, Lianyu Wang, Aidi Lin, Ke Zou, Xinxing Xu, Yi Zhou, Yuanyuan Peng, Qingquan Meng, Yiming Qian, et al. Uncertainty-inspired open set learning for retinal anomaly identification. Nature Communications, 14(1):6757, 2023. 2, 3 + +[30] Weiran Wang, Raman Arora, Karen Livescu, and Jeff Bilmes. On deep multi-view representation learning. In International conference on machine learning, pages 1083-1092. PMLR, 2015. 2 +[31] Chang Xu, Dacheng Tao, and Chao Xu. A survey on multiview learning. arXiv preprint arXiv:1304.5634, 2013. 1 +[32] Cai Xu, Jiajun Si, Ziyu Guan, Wei Zhao, Yue Wu, and Xiyue Gao. Reliable conflictive multi-view learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16129-16137, 2024. 1, 2, 3, 5 +[33] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 2 +[34] Jinglin Xu, Wenbin Li, Xinwang Liu, Dingwen Zhang, Ji Liu, and Junwei Han. Deep embedded complementary and interactive information for multi-view classification. In Proceedings of the AAAI conference on artificial intelligence, pages 6494-6501, 2020. 2 +[35] Shaoxun Xu, Yufei Chen, Chao Ma, and Xiaodong Yue. Deep evidential fusion network for medical image classification. International Journal of Approximate Reasoning, 150: 188-198, 2022. 2 +[36] Xiaoyu Yang, Jie Lu, and En Yu. Adapting multi-modal large language model to concept drift from pre-training onwards. In The Thirteenth International Conference on Learning Representations, 2025. 2 +[37] Liyuan Zhang, Wei Liu, Yufei Chen, and Xiaodong Yue. Reliable multi-view deep patent classification. Mathematics, 10 (23):4545, 2022. 2 +[38] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1 \ No newline at end of file diff --git a/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/images.zip b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4f5fb6135183e4e64cdebb0487c0b9560b3899b4 --- /dev/null +++ b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66856b10a2beb90a9d687973c87ef13bed12d3ccaa33351021d9a6e72e2e7718 +size 630688 diff --git a/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/layout.json b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d32a4370da9d0c038185e4545b59515160b6f7ed --- /dev/null +++ b/2025/Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild/layout.json @@ -0,0 +1,8326 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 103, + 457, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 103, + 457, + 138 + ], + "spans": [ + { + "bbox": [ + 153, + 103, + 457, + 138 + ], + "type": "text", + "content": "Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "text", + "content": "Wei Liu" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "text", + "content": ", Yufei Chen" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "text", + "content": ", Xiaodong Yue" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "text", + "content": "School of Computer Science and Technology, Tongji University, Shanghai, China \n" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "text", + "content": "Artificial Intelligence Institute, Shanghai University, Shanghai, China \n" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 160, + 504, + 232 + ], + "type": "text", + "content": "VLN Lab, Shanghai, China \nldachuan@outlook.com, yufeichen@tongji.edu.cn, yswantfly@shu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 297, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 297, + 595 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 297, + 595 + ], + "type": "text", + "content": "Trusted multi-view classification (TMVC) addresses variations in data quality by evaluating the reliability of each view based on prediction uncertainty at the evidence level, reducing the impact of low-quality views commonly encountered in real-world scenarios. However, existing TMVC methods often struggle to maintain robustness during testing, particularly when integrating noisy or corrupted views. This limitation arises because the evidence collected by TMVC may be unreliable, frequently providing incorrect information due to complex view distributions and optimization challenges, ultimately leading to classification performance degradation. To enhance the robustness of TMVC methods in real-world conditions, we propose a generalized evidence filtering mechanism that is compatible with various fusion strategies commonly used in TMVC, including Belief Constraint Fusion, Aleatory Cumulative Belief Fusion, and Averaging Belief Fusion. Specifically, we frame the identification of unreliable evidence as a multiple testing problem and introduce " + }, + { + "bbox": [ + 53, + 283, + 297, + 595 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 53, + 283, + 297, + 595 + ], + "type": "text", + "content": "-values to control the risk of false identification. By selectively down-weighting unreliable evidence during testing, our mechanism ensures robust fusion and mitigates performance degradation. Both theoretical guarantees and empirical results demonstrate significant improvements in the classification performance of TMVC methods, supporting their reliable application in challenging, real-world environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 616, + 135, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 135, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 135, + 629 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "content": "Multi-view classification (MVC) enhances the accuracy of deep learning models by integrating information from multiple data views. By combining insights from diverse views, MVC improves decision-making in complex domains [7, 10, 11, 14, 15, 25, 31]. However, traditional" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 258, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 316, + 258, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 555, + 401 + ], + "type": "image", + "image_path": "f7ebd7586ae13d0d54a2da1a914d155d44851fd2267499c93d4a7a00c9eb0057.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 410, + 555, + 465 + ], + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 465 + ], + "type": "text", + "content": "Figure 1. Classification accuracy degradation with noise (w/) compared to no noise (w/o) across three fusion strategies (BCF, A-CBF, and ABF) on six datasets. Noise is introduced by replacing half of the views to simulate data collected from malfunctioning sensors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 494, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 494, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 312, + 494, + 555, + 614 + ], + "type": "text", + "content": "MVC approaches often assume that each view consistently provides high-quality, reliable information—an assumption that is frequently unrealistic in real-world applications. In practical scenarios, data quality and reliability across views can vary significantly, leading to contradictory or misleading information that compromises prediction accuracy and reliability. As a result, conventional MVC models may struggle to reliably identify the ground truth, particularly when dealing with noisy, corrupted, or otherwise low-quality data views [38]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 556, + 715 + ], + "type": "text", + "content": "To address these limitations, trusted multi-view classification (TMVC) [8, 9, 16, 17, 32] has emerged as an effective solution. Unlike traditional MVC, TMVC evaluates the trustworthiness of each view by leveraging prediction uncertainty derived from view-specific evidence, enabling it to identify unreliable views and prioritize informative ones in the final predictions. This approach has proven highly effective in critical applications such as medical imaging diagno" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15508" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "sis [29] and autonomous driving [21], where distinguishing between reliable and unreliable predictions is essential." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 296, + 360 + ], + "type": "text", + "content": "In TMVC, evidence from each sample for each view contributes to its prediction and is also used to calculate a view-specific uncertainty score " + }, + { + "bbox": [ + 56, + 98, + 296, + 360 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 56, + 98, + 296, + 360 + ], + "type": "text", + "content": ", which reflects the confidence in the quality of that view for the given sample. A high uncertainty score (i.e., " + }, + { + "bbox": [ + 56, + 98, + 296, + 360 + ], + "type": "inline_equation", + "content": "u = 1" + }, + { + "bbox": [ + 56, + 98, + 296, + 360 + ], + "type": "text", + "content": ") indicates a substantial deviation from the training data, potentially due to noise or corruption. In such cases, evidence supporting all classes is minimized and fused using one of three strategies: Belief Constraint Fusion (BCF) [8, 9], Aleatory Cumulative Belief Fusion (A-CBF) [16-18, 37], and Averaging Belief Fusion (ABF) [32]. However, despite their success in controlled settings, these methods often struggle to remain robust in the wild. Complex and diverse data view distributions, combined with varying levels of optimization, can distort uncertainty estimates, resulting in unreliable evidence that misleads the model and biases predictions toward incorrect classes. As shown in Figure 1, introducing random noise (e.g., replacing half of the views in a dataset) to simulate sensor failure conditions leads to a significant drop in classification accuracy across all fusion strategies on six real-world datasets, underscoring the risks of decision-making in unpredictable and noisy environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 504 + ], + "type": "text", + "content": "To address this challenge, we propose a generalized and robust testing-time evidence filtering mechanism to enhance the robustness of TMVC methods in real-world, unpredictable environments. Specifically, we frame the identification of unreliable evidence as a multiple testing problem, introducing " + }, + { + "bbox": [ + 55, + 361, + 295, + 504 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 361, + 295, + 504 + ], + "type": "text", + "content": "-values to control the risk of false identification within a small margin. This approach enables effective post-hoc adjustments during testing. By selectively down-weighting evidence from unreliable views, our mechanism enhances fusion robustness and mitigates performance degradation caused by low-quality data sources. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 506, + 295, + 624 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 506, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 506, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 295, + 541 + ], + "type": "text", + "content": "1. We propose a testing-time evidence filtering mechanism that is compatible with multiple fusion strategies, enhancing the robustness of existing TMVC methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 541, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 589 + ], + "type": "text", + "content": "2. Our approach frames the identification of unreliable evidence as a multiple testing problem and selectively adjusts evidence weights during testing, mitigating the negative impact of unreliable views on fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 590, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 624 + ], + "type": "text", + "content": "3. Theoretical and empirical analyses validate the effectiveness of the proposed method in improving both classification accuracy and robustness." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 639, + 147, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 147, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 147, + 651 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 194, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 194, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 194, + 670 + ], + "type": "text", + "content": "2.1. Multi-view Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Over the past few years, numerous multi-view classification (MVC) approaches [1, 2, 16, 22, 25, 26, 30, 34, 36] have been developed, particularly those leveraging deep neural" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "networks (DNNs). These methods aim to learn shared representations by integrating information from different types of views, resulting in improved performance. By combining insights from diverse views, MVC has achieved remarkable gains and has been widely applied across various tasks [23, 35]. However, in real-world scenarios, the quality of data across views can vary significantly, introducing noise and inconsistencies that undermine both classification accuracy and reliability—factors that are especially critical in areas like medical diagnosis and autonomous driving." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 202, + 493, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 493, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 493, + 214 + ], + "type": "text", + "content": "2.2. Trusted Multi-view Classification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 221, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 484 + ], + "type": "text", + "content": "To address these challenges, Trusted Multi-View Classification (TMVC) has emerged as a framework that assesses the reliability of each view through prediction uncertainty, allowing models to prioritize reliable views and down-weight unreliable ones [8, 9, 16, 17, 32, 33]. TMVC methods often rely on Subjective Logic (SL) [12] to model view-specific predictions as multinomial opinions, which are then fused using belief combination rules tailored to the context [12]. For instance, extensions of Dempster-Shafer theory [24] were firstly introduced by [8, 9] to facilitate evidence fusion in multi-view settings. Other fusion approaches, such as Cumulative Belief Fusion [16, 17] and Averaging Belief Fusion [32], have also been proposed to further enhance the effectiveness of TMVC. Despite their success, all these methods fall short of achieving robust performance during testing, as the magnitude of estimated uncertainty is highly sensitive to the diverse distributions of data views and varying levels of model optimization, which significantly degrades classification accuracy in real-world scenarios. In contrast, our method improves testing-time robustness through an evidence filtering mechanism, thereby reducing the risks associated with final decision-making." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 500, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 500, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 500, + 511 + ], + "type": "text", + "content": "3. Overview and Analysis of TMVC" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 519, + 500, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 500, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 500, + 530 + ], + "type": "text", + "content": "3.1. Generalized Framework of TMVC" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": "Given the training data " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "D_{train} = \\left\\{\\left\\{\\boldsymbol{x}_i^v\\right\\}_{v=1}^V, \\boldsymbol{y}_i\\right\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": " is the number of samples, each sample has " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": " views with a one-hot label " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}_i = \\left\\{y_{ik}\\right\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": " for a " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": "-class classification task. TMVC uses subjective logic [12] to learn a function " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": " that maps " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "\\left\\{\\boldsymbol{x}_i^v\\right\\}_{v=1}^V" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}_i" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": " along with a corresponding overall uncertainty score " + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 313, + 534, + 555, + 627 + ], + "type": "text", + "content": ", which quantifies prediction confidence at the evidence level." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": "Specifically, let " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "f_{v}(\\boldsymbol{x}_{i}^{v})" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": " denote the output of the penultimate (logits) layer of the view-specific neural network for sample " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{i}^{v}" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": " in a single view. Instead of the standard softmax function for a single categorical prediction, TMVC employs an alternative activation function " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "a(\\cdot)" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": " (e.g., ReLU) to capture non-negative evidence " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "e_{i}^{v} = a(f_{v}(\\boldsymbol{x}_{i}^{v}))" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "e_{i}^{v} = \\{e_{ik}^{v}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": ". The view-specific prediction uncertainty is" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15509" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 93, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 93, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 93, + 84 + ], + "type": "text", + "content": "given by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 93, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 93, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 148, + 93, + 295, + 107 + ], + "type": "interline_equation", + "content": "u _ {i} ^ {v} = K / S _ {i} ^ {v}, \\tag {1}", + "image_path": "d11f49344ff7c3dadd9a0bd1c030a85ebb11af21dd75f6f299fa644942944e1f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 116, + 295, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 295, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 116, + 295, + 152 + ], + "type": "inline_equation", + "content": "S_{i}^{v} = \\sum_{k=1}^{K} e_{ik}^{v} + K" + }, + { + "bbox": [ + 55, + 116, + 295, + 152 + ], + "type": "text", + "content": ". This reflects the lack of total evidence across all classes in a single view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 153, + 296, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 296, + 213 + ], + "type": "text", + "content": "TMVC then applies evidence fusion strategies, such as BCF, A-CBF, and ABF, to integrate the evidence from each view into a unified multi-view evidence representation " + }, + { + "bbox": [ + 55, + 153, + 296, + 213 + ], + "type": "inline_equation", + "content": "e_i = \\{e_{ik}\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 153, + 296, + 213 + ], + "type": "text", + "content": ". These three fusion rules for two views are commonly expressed as follows [20]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 220, + 294, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 220, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 92, + 220, + 294, + 268 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} A - C B F \\left(\\boldsymbol {e} _ {i} ^ {1}, \\boldsymbol {e} _ {i} ^ {2}\\right) = \\boldsymbol {e} _ {i} ^ {1} + \\boldsymbol {e} _ {i} ^ {2}. \\\\ A B F \\left(e _ {i} ^ {1}, e _ {i} ^ {2}\\right) = \\left(e _ {i} ^ {1} + e _ {i} ^ {2}\\right) / 2. \\\\ B C F \\left(\\boldsymbol {e} _ {i} ^ {1}, \\boldsymbol {e} _ {i} ^ {2}\\right) = \\boldsymbol {e} _ {i} ^ {1} + \\boldsymbol {e} _ {i} ^ {2} + \\boldsymbol {e} _ {i} ^ {1} \\boldsymbol {e} _ {i} ^ {2} / K. \\tag {2} \\\\ \\end{array}", + "image_path": "4291c55161150638743542b2a1f85e510a3ad079c8333995cd48bed3b6df48a5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 275, + 295, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 298 + ], + "type": "text", + "content": "These fusion rules can also be extended to handle multiple views with ease." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "content": "In the training phase, TMVC links the generated evidence to the concentration parameter of the Dirichlet distribution and optimizes model parameters using a task-specific loss function. This loss primarily consists of a cross-entropy term, which maximizes the evidence for the correct label of each sample, and a Kullback-Leibler (KL)-divergence term, which drives the evidence for incorrect labels toward zero, as described in [8]. The final multi-view prediction " + }, + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i = \\{p_{ik}\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "content": " is obtained as the expectation of the learned Dirichlet distribution:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 428, + 295, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 428, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 133, + 428, + 295, + 442 + ], + "type": "interline_equation", + "content": "p _ {i k} = \\left(e _ {i k} + 1\\right) / S _ {i}, \\tag {3}", + "image_path": "7ee2251f80b9972d3b0edef43f379302bbdc87ee4228632697442a43aeb09b73.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "inline_equation", + "content": "S_{i} = \\sum_{k = 1}^{K}e_{ik} + K" + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "text", + "content": ", and the overall uncertainty " + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "inline_equation", + "content": "u_{i} = K / S_{i}" + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "text", + "content": ". Finally, the model's best performance on the validation dataset " + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "inline_equation", + "content": "D_{val}" + }, + { + "bbox": [ + 55, + 450, + 296, + 513 + ], + "type": "text", + "content": " is used to select the optimal parameters for testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 519, + 256, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 256, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 256, + 533 + ], + "type": "text", + "content": "3.2. Limitations of TMVC at Testing Time" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 536, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 608 + ], + "type": "text", + "content": "Despite its success across various domains [21, 29], TMVC methods face significant limitations at testing time, particularly in real-world scenarios with diverse data view distributions and varying levels of model optimization. These factors often degrade TMVC's robustness, especially when handling low-quality data views." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "text", + "content": "Specifically, given a multi-view test dataset " + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "inline_equation", + "content": "D_{test} = \\left\\{\\left\\{\\boldsymbol{x}_j^v\\right\\}_{v=1}^V\\right\\}_{j=1}^M" + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "text", + "content": ", a trained TMVC model with " + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "text", + "content": " view-specific networks provides uncertainty estimates for each view of a testing sample " + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_j^v" + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "text", + "content": ". Ideally, a high uncertainty score (e.g., " + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "inline_equation", + "content": "u_j^v = 1" + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "text", + "content": ") for a low-quality view would reduce its evidence contribution across all classes to zero (" + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "inline_equation", + "content": "e_j^v = 0" + }, + { + "bbox": [ + 55, + 609, + 296, + 713 + ], + "type": "text", + "content": "), minimizing its impact during fusion. However, these uncertainty estimates are sensitive to data distributions and" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 437, + 185 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 437, + 185 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 437, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 437, + 185 + ], + "type": "image", + "image_path": "e8f495d20d6f49ae6ff828e7a9737257e9e39f68115c8c8f25415bd8c9ebbef1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 554, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 554, + 237 + ], + "type": "text", + "content": "Figure 2. Density distribution of uncertainty for noisy view results under noise corruption and the corresponding density distribution of remaining unreliable evidence across three fusion strategies on the CUB dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 553, + 185 + ], + "type": "image", + "image_path": "3ac0fa0a6cbd09095427c090e1b652a17dfd50ffe197cbe7e7ba6a32a69a0d0f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": "model optimization levels, making it challenging to reliably achieve " + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "inline_equation", + "content": "u_{j}^{v} = 1" + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": " for extremely noisy or corrupted views." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 285, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 513 + ], + "type": "text", + "content": "For example, in the CUB dataset, even when one view is entirely corrupted by noise (i.e., features of one view are replaced by random noise), the uncertainty score for that view typically falls within the range [0.7, 1.0], leaving residual evidence mostly between [0.0, 4.0] (as illustrated in Figure 2) that contributes to fusion and can potentially bias predictions. Since this view lacks useful information, its residual evidence is unreliable and negatively impacts multi-view classification accuracy, as shown in Figure 1. Moreover, the overall uncertainty score derived from fusion strategies like BCF and A-CBF is often lower as additional views are fused (as demonstrated in [9, 16]). Even with ABF fusion, the overall uncertainty remains below the highest view-specific uncertainty [32]. As a result, decision-makers may struggle to detect potentially misclassified samples caused by low-quality views. These limitations underscore the challenges TMVC faces in real-world conditions, where testing-time robustness is crucial for reliable decision-making." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 525, + 421, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 421, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 421, + 539 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "In this section, we present a strategy to enhance the robustness of TMVC methods at testing time. Performance degradation in TMVC methods often arises from the negative impact of such low-quality views, which typically exhibit high prediction uncertainty during testing, compared to the reliable views seen during training. An intuitive solution to this problem is to reduce the contribution of unreliable evidence from low-quality views during multi-view fusion. To address this, we propose a generalized evidence filtering mechanism that is compatible with various TMVC fusion rules. The mechanism consists of two key components: (1) identifying unreliable evidence collected from low-quality views, and (2) assigning lower weights to minimize their impact on the final decision." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15510" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "type": "text", + "content": "4.1. Reliable Evidence Discovery in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 161 + ], + "type": "text", + "content": "Identifying unreliable evidence is equivalent to discovering reliable evidence. The primary challenge in this process is controlling the false discovery rate (FDR) [3], since the true label is unknown. We address this challenge by framing reliable evidence discovery as a multiple hypothesis testing problem from a statistical learning perspective." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": "Given a multi-view test dataset " + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "inline_equation", + "content": "D_{test}" + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": ", the objective is to identify as many reliable pieces of evidence as possible while ensuring that the FDR is maintained below a user-defined threshold " + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1)" + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": ". Let us define a discovery function " + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\Lambda(\\kappa^v)" + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": ", which selects a subset " + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "inline_equation", + "content": "S^v" + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": " of view-specific evidence for the test points of interest. Evidences with smaller values of " + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "inline_equation", + "content": "u_j^v" + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "inline_equation", + "content": "u_j^v \\leq \\kappa^v" + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": ", is likely to be included in the subset, where " + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\kappa^v" + }, + { + "bbox": [ + 55, + 161, + 296, + 268 + ], + "type": "text", + "content": " is a threshold determined by the discovery function. The FDR for this function is then defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 274, + 296, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 308 + ], + "type": "interline_equation", + "content": "\\mathrm {F D R} \\left(\\Lambda \\left(\\kappa^ {v}\\right)\\right) := \\mathbb {E} \\left[ \\frac {\\sum_ {j \\in D _ {t e s t}} \\mathbb {1} \\left\\{j \\in S ^ {v} , u _ {j} ^ {v} > \\kappa^ {v} \\right\\}}{\\max \\left\\{1 , | S ^ {v} | \\right\\}} \\right]. (4)", + "image_path": "345705b260897492e738cd1906c7842a9c6b1d8b727f82627dcb801b00765d66.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 314, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 339 + ], + "type": "text", + "content": "Our goal is to ensure that the subset " + }, + { + "bbox": [ + 55, + 314, + 296, + 339 + ], + "type": "inline_equation", + "content": "S^v" + }, + { + "bbox": [ + 55, + 314, + 296, + 339 + ], + "type": "text", + "content": " identified by " + }, + { + "bbox": [ + 55, + 314, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\Lambda(\\kappa^v)" + }, + { + "bbox": [ + 55, + 314, + 296, + 339 + ], + "type": "text", + "content": " satisfies the following error-control property:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": "Definition 1 (Risk-controlling discovery). Let " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\kappa^v" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": " be a random variable. We define " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\Lambda(\\kappa^v)" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": " as an " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "(\\alpha, \\delta)" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": "-risk-controlling discovery (RCD) function if" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 389, + 294, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 389, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 127, + 389, + 294, + 403 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left(\\operatorname {F D R} \\left(\\Lambda \\left(\\kappa^ {v}\\right)\\right) \\leq \\alpha\\right) \\tag {5}", + "image_path": "73f9932b8659224c73062f95c91ce32e424fcc95e5a5b484e68ab14fbeea81a9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 195, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 195, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 195, + 423 + ], + "type": "text", + "content": "holds with probability at least " + }, + { + "bbox": [ + 55, + 411, + 195, + 423 + ], + "type": "inline_equation", + "content": "1 - \\delta" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\delta \\in (0,1)" + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "text", + "content": " is the confidence level. To meet this risk control criterion, we need to identify a suitable " + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\Lambda (\\kappa^v)" + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "text", + "content": " that adapts to diverse data view distributions encountered in the wild. For this purpose, we gather statistical evidence via hypothesis testing, where we consider the view-specific evidence of a test sample to be discovered (i.e, " + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "inline_equation", + "content": "u_{j}^{v}\\leq \\kappa^{v}" + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "text", + "content": ") by " + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\Lambda (\\kappa^v)" + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "text", + "content": " as the null hypothesis. Rejecting the null hypothesis reflects that " + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "inline_equation", + "content": "u_{j}^{v} > \\kappa^{v}" + }, + { + "bbox": [ + 55, + 430, + 296, + 537 + ], + "type": "text", + "content": ", and therefore the evidence is unreliable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": "Under this framework, discovering reliable evidence boils down to simultaneously testing the " + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": " hypotheses. To do this, for any arbitrary view-specific evidence " + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{e}_j^v" + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": ", we compute the " + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": "-value:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 84, + 594, + 296, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 594, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 84, + 594, + 296, + 622 + ], + "type": "interline_equation", + "content": "p \\text {- v a l u e} \\left(\\boldsymbol {e} _ {j} ^ {v}\\right) = \\frac {1 + \\sum_ {l \\in D _ {v a l}} \\mathbb {1} \\left\\{u _ {j} ^ {v} \\geq u _ {l} ^ {v} \\right\\}}{\\left| D _ {v a l} \\right| + 1}. \\tag {6}", + "image_path": "2e5230acd351636194c9c052aa70c6dc5a81d437ab25607bf31f67dcc059683f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "text", + "content": "It is clear that the " + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "text", + "content": "-value quantifies how representative the evidence " + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "inline_equation", + "content": "e_j^v" + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "text", + "content": " is of reliable evidence. If the " + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "text", + "content": "-value is less than " + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "text", + "content": ", the null hypothesis of exchangeability with the validation data is rejected, deeming the evidence unreliable. We obtain marginally valid " + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 628, + 296, + 689 + ], + "type": "text", + "content": "-values, which satisfy:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 118, + 696, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 696, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 118, + 696, + 296, + 712 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left(p - \\text {v a l u e} \\left(\\boldsymbol {e} _ {j} ^ {v}\\right) \\leq \\alpha\\right) \\leq \\alpha , \\tag {7}", + "image_path": "2119682c3b016ca090fe245dfd734c7d75fca2fdea1d4707e8efa3815888caeb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": "indicating that the probability of falsely identifying unreliable evidence is controlled by the small margin " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": ". To achieve this, we set a view-specific, distribution-free threshold using the Benjamini-Hochberg (BH) procedure [3]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 129, + 555, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 129, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 380, + 129, + 555, + 144 + ], + "type": "interline_equation", + "content": "\\kappa^ {v} = Q \\left(\\left\\{u _ {l} ^ {v} \\right\\} _ {l \\in D _ {v a l}}, \\alpha\\right), \\tag {8}", + "image_path": "3528a2672630fb29c476ae42c2f101ba064f79f1f988c3c7bcc06c848b89d8f2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "inline_equation", + "content": "Q(\\cdot, \\alpha)" + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "inline_equation", + "content": "\\lceil (1 - \\alpha)(|D_{val}| + 1) \\rceil" + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "text", + "content": "-th smallest value in " + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "inline_equation", + "content": "\\{u_l^v\\}_{l \\in D_{val}}" + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "text", + "content": ". Then, we construct the reliable evidence set using " + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "inline_equation", + "content": "\\Lambda(\\kappa^v)" + }, + { + "bbox": [ + 313, + 151, + 554, + 188 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 371, + 196, + 555, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 196, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 371, + 196, + 555, + 212 + ], + "type": "interline_equation", + "content": "S ^ {v} = \\left\\{j \\in D _ {\\text {t e s t}}: u _ {j} ^ {v} \\leq \\kappa^ {v} \\right\\}. \\tag {9}", + "image_path": "a952fe39806a44f6cc7b106ba201dc64ba97cfc4ef403ef84d3e69b0e7e486c8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 219, + 489, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 489, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 489, + 231 + ], + "type": "text", + "content": "Finally, we have the following proposition:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 238, + 554, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 263 + ], + "type": "text", + "content": "Proposition 1. " + }, + { + "bbox": [ + 313, + 238, + 554, + 263 + ], + "type": "inline_equation", + "content": "\\Lambda (\\kappa^v)" + }, + { + "bbox": [ + 313, + 238, + 554, + 263 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 313, + 238, + 554, + 263 + ], + "type": "inline_equation", + "content": "(\\alpha ,\\delta)" + }, + { + "bbox": [ + 313, + 238, + 554, + 263 + ], + "type": "text", + "content": "-risk-controlling discovery (RCD) function." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 270, + 554, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 283 + ], + "type": "text", + "content": "4.2. Mitigating the Influence of Unreliable Evidence" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 288, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 335 + ], + "type": "text", + "content": "After identifying the reliability of view-specific evidence, we aim to mitigate the impact of unreliable evidence during multi-view fusion. To achieve this, we define the filtered evidence for each testing sample as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 376, + 344, + 554, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 344, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 376, + 344, + 554, + 371 + ], + "type": "interline_equation", + "content": "\\widehat {\\boldsymbol {e}} _ {j} ^ {v} = \\left\\{ \\begin{array}{c c} \\boldsymbol {e} _ {j} ^ {v}, & j \\in S ^ {v}. \\\\ \\alpha \\cdot \\boldsymbol {e} _ {j} ^ {v}, & j \\notin S ^ {v}. \\end{array} \\right. \\tag {10}", + "image_path": "ef1f782ddee7c8ebea0b7db7fc745919bff360364a084e7bcf84b9894095c535.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "text", + "content": " is a user-defined parameter that controls the weight assigned to unreliable evidence, directly influencing the FDR control. Smaller values of " + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "text", + "content": " effectively reduce the influence of low-quality views, minimizing the risk of unreliable evidence negatively impacting the final prediction. This tight control is especially valuable in high-stakes applications where high confidence and accuracy in predictions is crucial. On the other hand, if " + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "text", + "content": " is set to a larger value, the mechanism retains more influence from low-quality views, which may be beneficial when such views contain partially valuable information despite high uncertainty." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "Moreover, the choice of " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " balances FDR control and information retention, allowing flexibility in managing the contribution of uncertain views to the fusion process. A lower " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " tightens FDR control, reducing the likelihood of unreliable evidence being included and enhancing prediction reliability. Importantly, we do not fully exclude unreliable evidence, as even noisy views may contain useful information for classification. Instead, we adopt a weighted filtering approach that retains the potential benefits of low-quality views while mitigating their negative effects. This soft filtering strategy allows the model to remain adaptive and robust under real-world conditions." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "After filtering, the view-specific evidence is fused using the selected fusion rules (e.g., BCF, A-CBF, or ABF), achieving an adaptable and robust multi-view fusion process that is compatible with various TMVC fusion strategies. This leads to the following propositions:" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15511" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "Proposition 2. Applying evidence filtering improves the classification performance of TMVC fusion strategies (e.g., BCF, A-CBF, or ABF), under the assumption that unreliable evidence provides incorrect information in the wild." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 125, + 295, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 295, + 174 + ], + "type": "text", + "content": "Proposition 3. The overall uncertainty of multi-view results generated by TMVC fusion strategies with evidence filtering for unreliable evidence will exhibit greater uncertainty than fusion without evidence filtering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 179, + 296, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 296, + 298 + ], + "type": "text", + "content": "Both Propositions 2 and 3 guarantee the robustness of our method when confronted with unreliable evidence (i.e., evidence from low-quality views) in the wild. Note that unreliable evidence providing correct predictions by chance (a counterintuitive situation) may hurt classification performance in our method, particularly when evidence from other views is incorrect. To address this, our method exhibits greater overall uncertainty for these multi-view results, as shown in Proposition 3, which can alert humans to be cautious with these results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 382 + ], + "type": "text", + "content": "In summary, our evidence filtering mechanism improves testing-time robustness by adjusting the weights of view-specific evidence based on reliability. This enables TMVC methods to perform reliably in environments with data inconsistencies. This testing-time robustness is crucial for dependable deployment in high-stakes, real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 248, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 248, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 248, + 402 + ], + "type": "text", + "content": "4.3. Computational Complexity Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 455 + ], + "type": "text", + "content": "Our method introduces a generalized framework that can be applied to various DNNs for multi-view testing. Given a single test pass with complexity " + }, + { + "bbox": [ + 55, + 407, + 296, + 455 + ], + "type": "inline_equation", + "content": "O(T_{test})" + }, + { + "bbox": [ + 55, + 407, + 296, + 455 + ], + "type": "text", + "content": " on an arbitrary TMVC model, the overall computational complexity is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 472, + 294, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 472, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 114, + 472, + 294, + 486 + ], + "type": "interline_equation", + "content": "O (L \\log L) + O (M \\cdot T _ {t e s t}), \\tag {11}", + "image_path": "79436750673e40bcd8d39838b9c500cd411171966923aa03a70ddbe83aad490f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " represent the sizes of the validation and test datasets, respectively. The term " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "O(L\\log L)" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " reflects the complexity involved in threshold calculation, primarily due to sorting the uncertainty scores within the validation set to determine an appropriate threshold for filtering unreliable evidence. The term " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "O(M\\cdot T_{test})" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " represents the complexity of running one test pass on " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " samples in the testing set. In most practical scenarios, the term " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "O(M\\cdot T_{test})" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " will typically dominate, especially for large-scale DNNs, unless the validation set is extremely large or the neural networks are shallow and computationally lightweight." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 634, + 137, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 137, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 137, + 647 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "In this section, we validate the effectiveness of our proposed method on six real-world datasets. Each experiment is run 10 times, and we report the mean and standard deviation of accuracy (ACC) to ensure reliability. The extensive results demonstrate the clear advantage of our proposed method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 430, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 430, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 430, + 84 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 90, + 379, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 90, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 314, + 90, + 379, + 100 + ], + "type": "text", + "content": "5.1.1. Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 105, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 555, + 392 + ], + "type": "text", + "content": "We conduct experiments on six real-world multi-view datasets as follows: CUB [28]: This dataset contains 11,788 bird images from 200 distinct categories, each paired with text descriptions. For our experiments, we utilize the first 10 categories. Image features are extracted using GoogleNet, while text features are extracted using doc2vec. SCENE [5]: This dataset comprises 4,485 images from 15 indoor and outdoor scene categories. Multiple views are represented using three feature types: GIST, PHOG, and LBP. CAL [6]: The CAL dataset consists of 8,677 images from 101 categories. We focus on the first 20 categories and extract six types of features: GABOR, WAVELET-MOMENTS, CENHIST, HOG, GIST, and LBP. HAND [27]: This dataset contains 2,000 samples across 10 digit classes (from '0' to '9'), with 200 samples per class. Six feature descriptors (PIX, FOU, FAC, ZER, KAR, MOR) are extracted as multiple views. PIE: This dataset consists of 680 facial images from 68 individuals. Three types of features are extracted: intensity, LBP, and Gabor. NUS: A subset of the NUS-WIDE dataset [4], it contains 30,000 images spanning 31 object categories, represented in five views, including a 64-D color histogram, 225-D blockwise color moments, 144-D color correlogram, 73-D edge direction histogram, and 128-D wavelet texture." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 399, + 427, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 427, + 411 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 427, + 411 + ], + "type": "text", + "content": "5.1.2. Compared Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 415, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 559 + ], + "type": "text", + "content": "Since our method is a generalized framework applicable to various trusted multi-view classification methods, we primarily evaluate its effectiveness by comparing baseline methods with and without our approach in these experiments. We apply our method to three fusion rules: BCF, A-CBF, and ABF, using the general loss function proposed in the work of [8]. We then integrate our method with state-of-the-art multi-view deep classification methods, including TMC [8], TMDOA [16], ETMC [9], SMDC [17], ECML [32], and CCML [19]. These methods use the same neural network architecture as ours, ensuring a consistent basis for comparison." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 566, + 415, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 566, + 415, + 578 + ], + "spans": [ + { + "bbox": [ + 314, + 566, + 415, + 578 + ], + "type": "text", + "content": "5.1.3. Implementations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "content": "For all datasets, we split the data into training (70%), validation (10%), and testing (20%) sets. Fully connected networks were used as view-specific backbones. To obtain a trained TMVC model, we employed the Adam optimizer [13] with an " + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "content": "-norm regularization of " + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "content": " and a learning rate of " + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "content": ". During testing, we set " + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "content": " by default, ensuring an FDR below 5% unless otherwise specified. The model was implemented in PyTorch and run on a GeForce RTX 4090 GPU with 24GB of memory." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Note: Optimization during training is not a primary focus of this work; hence, finding the optimal model parameters" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15512" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 547, + 275 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 547, + 275 + ], + "type": "table", + "html": "
Fusion with Clean Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB98.75±0.4298.75±0.4297.50±1.6798.33±1.6797.08±0.4297.92±0.42
SCENE70.94±0.0670.69±0.3171.63±0.7571.31±1.0670.50±0.1370.25±0.25
CAL98.50±0.5098.38±0.3897.63±0.3897.25±0.2598.75±0.5098.50±0.25
HAND97.88±0.1398.25±0.2599.00±0.0599.00±0.7598.88±0.3899.25±0.50
PIE98.08±1.1598.46±0.7798.85±0.3898.85±0.3897.69±1.5498.08±1.15
NUS53.85±1.0054.05±0.8055.35±0.3055.43±0.0857.75±0.7557.88±0.78
Fusion with Corrupted Views
ABFA-CBFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
CUB64.17±5.0080.00±0.8356.67±6.6779.17±1.6765.33±3.3383.75±1.25
SCENE46.00±3.3853.38±2.3150.75±2.0051.38±1.9449.25±0.6352.75±0.75
CAL56.75±1.8876.75±5.0070.25±0.7571.50±2.0073.25±0.1375.50±2.00
HAND64.50±1.6370.00±1.8870.00±0.7570.75±0.6375.75±1.7578.25±1.75
PIE89.77±0.0196.54±0.3888.46±3.8590.77±2.3188.08±0.3891.54±0.01
NUS27.70±1.6037.70±0.8536.73±1.2238.95±1.2037.00±1.6040.63±1.38
", + "image_path": "3af809e3e41194dce520df021a0df0816f9d3fcd08ee5e205c8e748b497f59a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 555, + 304 + ], + "type": "text", + "content": "Table 1. Classification performance (ACC (%)) for fusion with clean views and corrupted views, based on three fusion strategies, with or without our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 327, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 294, + 374 + ], + "type": "text", + "content": "was not the primary goal. Instead, we provide comparative results with and without our method under identical settings (same random seed, dataset splits, optimization parameters, etc.) to evaluate the effectiveness of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 180, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 180, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 180, + 395 + ], + "type": "text", + "content": "5.2. Quantitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 291, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 291, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 291, + 411 + ], + "type": "text", + "content": "5.2.1. Effect of Our Method on Basic Fusion Strategies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 415, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 593 + ], + "type": "text", + "content": "In this evaluation, we assess the effectiveness of our proposed method across three fusion strategies (ABF, A-CBF, and BCF) under both clean and corrupted view conditions. Initially, we applied our method to these fusion strategies in the context of clean views, where no noise was introduced. As shown in Table 1, our method generally improves classification accuracy across most datasets, even in noise-free environments. This demonstrates that our approach enhances model stability in both noisy and normal conditions. Minor performance decreases observed in the SCENE and CAL datasets with clean views may be attributed to our method's down-weighting of higher uncertainty evidence for some samples in certain views. These higher uncertainty samples could otherwise positively impact classification in some cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Under corrupted view conditions, where random noise in the range of " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "[-5, 5]" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " is added to half of the views, our method significantly boosts accuracy across all datasets. For example, on the CUB dataset with A-CBF fusion, our method achieves nearly a " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " improvement. For datasets like HAND, the modest performance improvement under corrupted conditions can be explained by the high uncertainty values (close to 1) associated with low total evidence (close to 0) in the corrupted HAND views. This naturally limits the impact of unreliable evidence during fusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 327, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 374 + ], + "type": "text", + "content": "Overall, these results underscore our method's robustness in enhancing existing TMVC fusion strategies, particularly in challenging, noisy environments, while also maintaining stable performance in clean data conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 491, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 491, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 491, + 392 + ], + "type": "text", + "content": "5.2.2. Effect on Popular TMVC Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 395, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 658 + ], + "type": "text", + "content": "We also demonstrate the effectiveness of our method when applied to various popular TMVC methods under corrupted view conditions (similar to Table 1) across six datasets. As shown in Figure 3, each radar chart represents one popular TMVC method, displaying classification accuracy under two conditions: with our method (w/ Ours, shown in red) and without our method (w/o Ours, shown in blue). Across all six TMVC methods, our method consistently improves performance under corrupted view conditions, as indicated by the expanded area covered by the red line compared to the blue line. The radar charts clearly show that our method effectively boosts classification accuracy on each dataset when faced with noisy and corrupted views. This improvement underscores the robustness of our approach in handling unreliable or degraded views across various TMVC frameworks. Particularly in challenging datasets like CUB, where noise can severely degrade model accuracy, our method provides a noticeable performance enhancement, indicating that it successfully mitigates the impact of corrupted views on multi-view classification accuracy. Overall, these results confirm the applicability and robustness of our method when integrated into diverse TMVC methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 662, + 491, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 491, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 491, + 674 + ], + "type": "text", + "content": "5.2.3. Sensitivity Analysis on Noisy Ratio" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In this experiment, we demonstrate the effectiveness of our method under varying noisy ratios in corrupted view conditions. We randomly select samples from half of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15513" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 70, + 233, + 193 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 233, + 193 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 233, + 193 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 233, + 193 + ], + "type": "image", + "image_path": "de8b0859a823f81376e60849e79864afd769a43cf65dfb0f5ce4398470e40b84.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 241, + 70, + 370, + 204 + ], + "blocks": [ + { + "bbox": [ + 241, + 70, + 370, + 204 + ], + "lines": [ + { + "bbox": [ + 241, + 70, + 370, + 204 + ], + "spans": [ + { + "bbox": [ + 241, + 70, + 370, + 204 + ], + "type": "image", + "image_path": "7c0f01b2ac4e5aea2576ccae5ba8e155aee3b56edfd08d00ce4a019b2eda08c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 508, + 191 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 508, + 191 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 508, + 191 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 508, + 191 + ], + "type": "image", + "image_path": "ec9392ed813ce6d285679c47f117acb818d793c13282c49237d52e122bd0f752.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 103, + 209, + 232, + 327 + ], + "blocks": [ + { + "bbox": [ + 103, + 209, + 232, + 327 + ], + "lines": [ + { + "bbox": [ + 103, + 209, + 232, + 327 + ], + "spans": [ + { + "bbox": [ + 103, + 209, + 232, + 327 + ], + "type": "image", + "image_path": "7e68622b40361e16bfc07039ed007e0dd24922d600fdd02812c17f6a5a936a6c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 335, + 493, + 346 + ], + "lines": [ + { + "bbox": [ + 116, + 335, + 493, + 346 + ], + "spans": [ + { + "bbox": [ + 116, + 335, + 493, + 346 + ], + "type": "text", + "content": "Figure 3. Comparison of classification accuracy between TMVC methods with and without our method." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 242, + 209, + 370, + 327 + ], + "blocks": [ + { + "bbox": [ + 242, + 209, + 370, + 327 + ], + "lines": [ + { + "bbox": [ + 242, + 209, + 370, + 327 + ], + "spans": [ + { + "bbox": [ + 242, + 209, + 370, + 327 + ], + "type": "image", + "image_path": "0e04ec4757e5f8704007ad6dd906c4305d2c12289ec6033ef514cc716d7483e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 378, + 208, + 506, + 327 + ], + "blocks": [ + { + "bbox": [ + 378, + 208, + 506, + 327 + ], + "lines": [ + { + "bbox": [ + 378, + 208, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 378, + 208, + 506, + 327 + ], + "type": "image", + "image_path": "443332ab02c5a57058db11cb8c6f9c6b62440267a846ab09bcff5ad896ef6c82.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 69, + 358, + 228, + 466 + ], + "blocks": [ + { + "bbox": [ + 69, + 358, + 228, + 466 + ], + "lines": [ + { + "bbox": [ + 69, + 358, + 228, + 466 + ], + "spans": [ + { + "bbox": [ + 69, + 358, + 228, + 466 + ], + "type": "image", + "image_path": "e783faeb1e085466394b96283e4b7bae25f5c4b3e1a0e0265ff22f0013547701.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 474, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 55, + 474, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 555, + 496 + ], + "type": "text", + "content": "Figure 4. Sensitivity analysis of accuracy with varying noisy ratios under corrupted view conditions (where half of the views contain noise) on the CUB dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 228, + 358, + 386, + 465 + ], + "blocks": [ + { + "bbox": [ + 228, + 358, + 386, + 465 + ], + "lines": [ + { + "bbox": [ + 228, + 358, + 386, + 465 + ], + "spans": [ + { + "bbox": [ + 228, + 358, + 386, + 465 + ], + "type": "image", + "image_path": "b775281c623353fb681beb21ec9dbf847e8eeed30566d7fa47a22e37fd0bff38.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 386, + 358, + 542, + 466 + ], + "blocks": [ + { + "bbox": [ + 386, + 358, + 542, + 466 + ], + "lines": [ + { + "bbox": [ + 386, + 358, + 542, + 466 + ], + "spans": [ + { + "bbox": [ + 386, + 358, + 542, + 466 + ], + "type": "image", + "image_path": "4b68fbfdd59b4df89a0f34e4d8851d2ce29d9830274b356138a676c8b687e5fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 517, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 517, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 54, + 517, + 297, + 708 + ], + "type": "text", + "content": "the views in the CUB dataset and apply different noisy ratios to create polluted samples. As shown in Figure 4, the experimental results reveal that the proposed method significantly enhances the robustness of all three fusion strategies (ABF, A-CBF, and BCF) against varying noisy ratios. As the noisy ratio increases, both the baseline and proposed methods experience a decline in accuracy. However, the proposed method consistently outperforms the baseline across all fusion rules. Notably, the accuracy decline is slower for the method with the proposed fusion approach, and it is accompanied by reduced variability, indicating a greater resilience to noise. These findings highlight the effectiveness of the proposed method in maintaining both accuracy and stability, especially in high-noise environments. This makes it a promising solution for noise-robust multi-view fusion tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 517, + 555, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 530 + ], + "type": "text", + "content": "5.2.4. Robustness Analysis under Varying Noise Intensity" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": "We analyze the robustness of our method under varying noise intensities. In this part, noise vectors (denoted by " + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": ") are sampled from a Gaussian distribution " + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": ". These noise vectors are then multiplied by an intensity factor " + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": " and added to half of the views in the test samples from the CUB dataset, i.e., " + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{x}}^v = \\boldsymbol{x}^v + \\eta \\epsilon^v" + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": ". Figure 5 presents the robustness analysis under varying noise intensities. The results indicate that the proposed method significantly enhances the robustness of the three fusion strategies (ABF, A-CBF, and BCF) against different noise levels. Across all noise intensities, our method consistently outperforms the baseline, showing more stable accuracy and reduced variability, particularly at higher noise levels. These findings highlight the effectiveness of our method in improving performance and stability in noisy environments." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15514" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 229, + 180 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 229, + 180 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 229, + 180 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 229, + 180 + ], + "type": "image", + "image_path": "c1042192b852cb7f0395c7ee05a814731e2728323264ba4ba0b942daa0394054.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 226, + 71, + 385, + 180 + ], + "blocks": [ + { + "bbox": [ + 226, + 71, + 385, + 180 + ], + "lines": [ + { + "bbox": [ + 226, + 71, + 385, + 180 + ], + "spans": [ + { + "bbox": [ + 226, + 71, + 385, + 180 + ], + "type": "image", + "image_path": "12237ecb883beb199a255e761e85ff45d52db53705b2472f38cec5bd92dda42e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 383, + 71, + 542, + 180 + ], + "blocks": [ + { + "bbox": [ + 383, + 71, + 542, + 180 + ], + "lines": [ + { + "bbox": [ + 383, + 71, + 542, + 180 + ], + "spans": [ + { + "bbox": [ + 383, + 71, + 542, + 180 + ], + "type": "image", + "image_path": "388b04a61267d04f5e800150f1a041205f923b7fbc7f43d56d98ebc6da9db48e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 211, + 541, + 388 + ], + "blocks": [ + { + "bbox": [ + 156, + 189, + 453, + 201 + ], + "lines": [ + { + "bbox": [ + 156, + 189, + 453, + 201 + ], + "spans": [ + { + "bbox": [ + 156, + 189, + 453, + 201 + ], + "type": "text", + "content": "Figure 5. Robustness analysis under varying noise intensities on the CUB dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 211, + 541, + 388 + ], + "lines": [ + { + "bbox": [ + 70, + 211, + 541, + 388 + ], + "spans": [ + { + "bbox": [ + 70, + 211, + 541, + 388 + ], + "type": "table", + "html": "
A-CBFABFBCF
w/o Oursw/ Oursw/o Oursw/ Oursw/o Oursw/ Ours
α =0.00569.60 ± 2.7069.60 ± 2.7053.75 ± 1.7553.75 ± 1.7570.00 ± 1.4070.10 ± 1.40
α =0.00672.50 ± 0.3072.60 ± 0.3049.88 ± 0.6350.63 ± 0.1369.50 ± 1.0069.50 ± 1.00
α =0.00769.70 ± 1.4069.70 ± 1.4058.63 ± 6.3858.88 ± 6.6371.30 ± 1.6071.40 ± 1.60
α =0.00871.70 ± 1.0071.70 ± 1.0051.63 ± 0.6351.75 ± 0.7570.70 ± 0.7071.00 ± 0.70
α =0.00970.00 ± 1.3070.10 ± 1.3049.63 ± 1.3850.63 ± 0.3869.50 ± 1.0069.80 ± 1.00
α =0.01070.50 ± 1.8070.60 ± 1.8054.18 ± 4.6354.25 ± 2.5070.00 ± 0.8070.20 ± 0.80
α =0.02072.00 ± 1.3072.20 ± 1.3052.25 ± 0.0052.75 ± 0.2571.70 ± 1.3071.90 ± 1.30
α =0.03070.30 ± 1.3070.40 ± 1.3054.75 ± 1.7558.63 ± 1.3869.50 ± 0.8069.80 ± 0.80
α =0.04070.70 ± 1.8070.90 ± 1.8055.13 ± 2.3867.00 ± 1.0072.20 ± 1.0072.50 ± 1.00
α =0.05071.80 ± 1.1071.90 ± 1.1055.13 ± 0.3878.63 ± 1.3873.00 ± 0.9075.30 ± 0.90
α =0.06069.70 ± 1.2069.80 ± 1.2057.50 ± 2.5068.13 ± 1.3871.50 ± 1.2071.70 ± 1.20
α =0.07070.30 ± 1.3070.40 ± 1.3055.88 ± 2.6376.50 ± 1.0069.80 ± 1.0070.00 ± 1.00
α =0.08070.20 ± 1.3070.30 ± 1.3056.50 ± 1.2578.63 ± 0.1370.50 ± 1.3070.70 ± 1.30
α =0.09072.20 ± 1.7072.40 ± 1.7056.25 ± 3.5076.88 ± 2.3871.70 ± 1.0072.00 ± 1.00
α =0.10069.50 ± 1.4069.70 ± 1.4055.13 ± 1.6378.00 ± 1.2570.70 ± 1.2070.90 ± 1.20
", + "image_path": "8329d23d6584ed5903c4cdedb1d6c1fb4db0449b96199a9ea0a8d94a9d36c81e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 397, + 542, + 408 + ], + "lines": [ + { + "bbox": [ + 66, + 397, + 542, + 408 + ], + "spans": [ + { + "bbox": [ + 66, + 397, + 542, + 408 + ], + "type": "text", + "content": "Table 2. Sensitivity analysis of user-defined " + }, + { + "bbox": [ + 66, + 397, + 542, + 408 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 66, + 397, + 542, + 408 + ], + "type": "text", + "content": " under corrupted view conditions (with noise in half of the views) on the CAL dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 430, + 244, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 244, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 244, + 441 + ], + "type": "text", + "content": "5.2.5. Sensitivity Analysis on User-defined " + }, + { + "bbox": [ + 55, + 430, + 244, + 441 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "text", + "content": "In this section, we analyze the impact of different user-defined " + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "text", + "content": " values under corrupted view conditions (similar to Table 1) on the CAL dataset. Table 2 shows the effectiveness of our method across various fusion strategies (A-CBF, ABF, and BCF) as " + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "text", + "content": " changes. Generally, incorporating our method (\"w/Ours\") improves classification accuracy in most cases, particularly at higher " + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "text", + "content": " values. Our method consistently achieves higher accuracy with relatively low variance across different " + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "text", + "content": " values, especially in the ABF fusion, indicating strong performance in mitigating noise-induced fluctuations. Notably, when " + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 54, + 453, + 295, + 610 + ], + "type": "text", + "content": ", the accuracy across all three fusion strategies reaches a relatively high level of performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "content": "It is important to note that because noise was randomly sampled with different " + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "content": " settings in our experiment, the performance of the \"w/o Ours\" condition fluctuates with changes in " + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "content": ". However, despite these fluctuations, our method (\"w/ Ours\") consistently improves accuracy with same " + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "content": " under identical noisy conditions (i.e., with the same sampled noise), demonstrating its capability to effectively reduce the impact of noise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 451, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 451, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 312, + 451, + 555, + 643 + ], + "type": "text", + "content": "In this work, we proposed a generalized evidence filtering mechanism to improve the robustness of Trusted Multi-View Classification (TMVC) methods in the wild, particularly under challenging conditions with noisy and corrupted views. We framed the identification of unreliable evidence as a multiple testing problem, using " + }, + { + "bbox": [ + 312, + 451, + 555, + 643 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 312, + 451, + 555, + 643 + ], + "type": "text", + "content": "-values to control the risk of false identification within a minimal range. This approach enables effective post-hoc adjustments during testing. By selectively adjusting evidence weights, our mechanism strengthens fusion robustness and mitigates performance degradation from low-quality data sources. Both theoretical analysis and extensive experiments on multiple real-world datasets demonstrate that our method significantly enhances classification accuracy and stability across various fusion strategies (ABF, A-CBF, and BCF), even in environments with substantial noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 411, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 411, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 411, + 672 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 679, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62472315, 62476165)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15515" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis. In International conference on machine learning, pages 1247-1255. PMLR, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "text", + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "text", + "content": "[3] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1):289-300, 1995. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "type": "text", + "content": "[4] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "text", + "content": "[5] Li Fei-Fei and Pietro Perona. A bayesian hierarchical model for learning natural scene categories. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 524-531. IEEE, 2005. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 368 + ], + "type": "text", + "content": "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "text", + "content": "[7] Pinhan Fu, Xinyan Liang, Yuhua Qian, Qian Guo, Yayu Zhang, Qin Huang, and Ke Tang. Multi-scale features are effective for multi-modal classification: An architecture search viewpoint. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 456 + ], + "type": "text", + "content": "[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In International Conference on Learning Representations, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "text", + "content": "[9] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "text", + "content": "[10] Bingbing Jiang, Chenglong Zhang, Yan Zhong, Yi Liu, Yingwei Zhang, Xingyu Wu, and Weiguo Sheng. Adaptive collaborative fusion for multi-view semi-supervised classification. Information Fusion, 96:37-50, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "type": "text", + "content": "[11] Bingbing Jiang, Xingyu Wu, Xiren Zhou, Anthony G Cohn, Yi Liu, Weiguo Sheng, and Huanhuan Chen. Semi-supervised multi-view feature selection with adaptive graph learning. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3615-3629, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "type": "text", + "content": "[12] AUDUN. Jøsang. Subjective Logic: A formalism for reasoning under uncertainty. Springer, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 656 + ], + "type": "text", + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[14] Xinyan Liang, Yuhua Qian, Qian Guo, Honghong Cheng, and Jiye Liang. Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9236-9254, 2021. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[15] Xinyan Liang, Pinhan Fu, Qian Guo, Keyin Zheng, and Yuhua Qian. Dc-nas: Divide-and-conquer neural architecture search for multi-modal classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13754-13762, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "text", + "content": "[16] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7585-7593, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 217 + ], + "type": "text", + "content": "[17] Wei Liu, Yufei Chen, Xiaodong Yue, Changqing Zhang, and Shaorong Xie. Safe multi-view deep classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8870-8878, 2023. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "text", + "content": "[18] Wei Liu, Yufei Chen, and Xiaodong Yue. Building trust in decision with conformalized multi-view deep classification. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7278-7287, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "type": "text", + "content": "[19] Ying Liu, Lihong Liu, Cai Xu, Xiangyu Song, Ziyu Guan, and Wei Zhao. Dynamic evidence decoupling for trusted multi-view learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7269-7277, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 320, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 554, + 363 + ], + "type": "text", + "content": "[20] Jueqing Lu, Lan Du, Wray Buntine, Myong Chol Jung, Joanna Dipnall, and Belinda Gabbe. Navigating conflicting views: Harnessing trust for learning. arXiv preprint arXiv:2406.00958, 2024.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "text", + "content": "[21] Jay Patrikar, Sushant Veer, Apoorva Sharma, Marco Pavone, and Sebastian Scherer. Rulefuser: Injecting rules in evidential networks for robust out-of-distribution trajectory prediction. arXiv preprint arXiv:2405.11139, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "text", + "content": "[22] Zhenwen Ren and Quansen Sun. Simultaneous global and local graph structure preserving for multiple kernel clustering. IEEE transactions on neural networks and learning systems, 32(5):1839-1851, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 487 + ], + "type": "text", + "content": "[23] Yaniv Shachor, Hayit Greenspan, and Jacob Goldberger. A mixture of views network with applications to multi-view medical imaging. Neurocomputing, 374:1-9, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 489, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 554, + 510 + ], + "type": "text", + "content": "[24] Glenn Shafer. A mathematical theory of evidence. Princeton university press, 1976. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 512, + 554, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 554 + ], + "type": "text", + "content": "[25] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 556, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 554, + 590 + ], + "type": "text", + "content": "[26] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 591, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 554, + 623 + ], + "type": "text", + "content": "[27] Martijn van Breukelen, Robert PW Duin, David MJ Tax, and JE Den Hartog. Handwritten digit recognition by combined classifiers. Kybernetika, 34(4):381-386, 1998. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "type": "text", + "content": "[28] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. California Institute of Technology, 2011. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[29] Meng Wang, Tian Lin, Lianyu Wang, Aidi Lin, Ke Zou, Xinxing Xu, Yi Zhou, Yuanyuan Peng, Qingquan Meng, Yiming Qian, et al. Uncertainty-inspired open set learning for retinal anomaly identification. Nature Communications, 14(1):6757, 2023. 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15516" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 442 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "text", + "content": "[30] Weiran Wang, Raman Arora, Karen Livescu, and Jeff Bilmes. On deep multi-view representation learning. In International conference on machine learning, pages 1083-1092. PMLR, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 138 + ], + "type": "text", + "content": "[31] Chang Xu, Dacheng Tao, and Chao Xu. A survey on multiview learning. arXiv preprint arXiv:1304.5634, 2013. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 295, + 184 + ], + "type": "text", + "content": "[32] Cai Xu, Jiajun Si, Ziyu Guan, Wei Zhao, Yue Wu, and Xiyue Gao. Reliable conflictive multi-view learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16129-16137, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 295, + 217 + ], + "type": "text", + "content": "[33] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 274 + ], + "type": "text", + "content": "[34] Jinglin Xu, Wenbin Li, Xinwang Liu, Dingwen Zhang, Ji Liu, and Junwei Han. Deep embedded complementary and interactive information for multi-view classification. In Proceedings of the AAAI conference on artificial intelligence, pages 6494-6501, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 276, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 276, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 57, + 276, + 294, + 318 + ], + "type": "text", + "content": "[35] Shaoxun Xu, Yufei Chen, Chao Ma, and Xiaodong Yue. Deep evidential fusion network for medical image classification. International Journal of Approximate Reasoning, 150: 188-198, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "type": "text", + "content": "[36] Xiaoyu Yang, Jie Lu, and En Yu. Adapting multi-modal large language model to concept drift from pre-training onwards. In The Thirteenth International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 365, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 294, + 396 + ], + "type": "text", + "content": "[37] Liyuan Zhang, Wei Liu, Yufei Chen, and Xiaodong Yue. Reliable multi-view deep patent classification. Mathematics, 10 (23):4545, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 57, + 399, + 294, + 442 + ], + "type": "text", + "content": "[38] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15517" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_content_list.json b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2596c1d49cb8094a4cd344adb6d95c127cbe2030 --- /dev/null +++ b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_content_list.json @@ -0,0 +1,1610 @@ +[ + { + "type": "text", + "text": "Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation", + "text_level": 1, + "bbox": [ + 150, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yudi Shi $^{1,2}$ , Shangzhe Di $^{1,2}$ , Qirui Chen $^{1,2}$ , Weidi Xie $^{1,\\dagger}$", + "bbox": [ + 274, + 179, + 718, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Artificial Intelligence, Shanghai Jiao Tong University, China", + "bbox": [ + 215, + 200, + 781, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Coop. Medianet Innovation Center, Shanghai Jiao Tong University, China", + "bbox": [ + 204, + 222, + 792, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://zhengrongz.github.io/AoTD/", + "bbox": [ + 292, + 244, + 697, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 311, + 326, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper tackles the problem of video question answering (VideoQA), a task that often requires multi-step reasoning and a profound understanding of spatial-temporal dynamics. While large video-language models perform well on benchmarks, they often lack explainability and spatial-temporal grounding. In this paper, we propose Agent-of-Thoughts Distillation (AoTD), a method that enhances models by incorporating automatically generated Chain-of-Thoughts (CoTs) into the instruction-tuning process. Specifically, we leverage an agent-based system to decompose complex questions into sub-tasks, and address them with specialized vision models, the intermediate results are then treated as reasoning chains. We also introduce a verification mechanism using a large language model (LLM) to ensure the reliability of generated CoTs. Extensive experiments demonstrate that AoTD improves the performance on multiple-choice and open-ended benchmarks.", + "bbox": [ + 88, + 344, + 485, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 628, + 220, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Question Answering (VideoQA) refers to a critical task that offers a natural interface for human-machine interaction through language [33, 42, 43, 49]. This synergy of visual content and language enhances the accessibility of AI systems for the general public, allowing users to query complex visual content with natural language. By encompassing tasks such as action recognition, object detection, and scene understanding, VideoQA serves as a comprehensive benchmark for evaluating AI's ability to interpret videos, addressing the fundamental questions of 'who', 'what', 'when', and 'where' that are crucial to understand daily life activities, pushing the boundaries of what AI systems can interpret from dynamic visual content.", + "bbox": [ + 89, + 652, + 482, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent literature has primarily explored two avenues in VideoQA. The first involves training large video lan", + "bbox": [ + 89, + 849, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guage models (Video-LLMs) through direct instruction-tuning, using videos paired solely with corresponding questions and answers [1, 22, 24, 53]. While these models excel on public benchmarks, they often lack explainability and struggle with spatial-temporal grounding. This limitation hinders their ability to provide clear reasoning, which is essential for real-world applications where transparency and interpretability are critical [29].", + "bbox": [ + 511, + 299, + 906, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conversely, an emerging approach utilizes agent-based systems that decompose complex questions into manageable sub-tasks, each addressed by specialized tools [15, 17, 37]. The results are then aggregated to form a coherent answer. Theoretically, such approach naturally offers greater interpretability, as the reasoning process is divided into explainable steps that can be independently assessed. However, our experiments indicate that current video understanding tools are not strong enough for building reliable agent-based systems. In addition, the high memory demands and time-consuming nature of these systems present significant challenges for their practical use.", + "bbox": [ + 511, + 422, + 908, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to leverage the advantage of both research lines, enhancing Video-LLM by integrating Chain-of-Thoughts (CoTs) into instruction-tuning, with the CoTs being constructed from the outputs of specialized agent models, capturing the step-by-step reasoning procedure, as illustrated in Figure 1.", + "bbox": [ + 511, + 607, + 906, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In specific, we start by systematically evaluating the off-the-shelf models tailored for atomic video understanding tasks, such as action recognition [39, 41] and language grounding [23], using well-annotated datasets. This comprehensive evaluation allows us to pinpoint the most effective tools for each sub-task, thus laying a robust foundation for constructing reliable chains. Moreover, this process also provides a critical assessment of the broader capabilities of visual models across general and complex scenes, offering valuable insights for future research within the community.", + "bbox": [ + 511, + 700, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In addition, we introduce a verification mechanism using a large language model (LLM), designed to assess if the generated CoTs adhere to a clear, step-by-step reason-", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†: Corresponding author.", + "bbox": [ + 91, + 887, + 227, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8523", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17bb1544992c14d17c78007bd09f29b16c38ac5375d015f13f76caae5c65d36f.jpg", + "image_caption": [ + "Figure 1. Our method, AoTD, distills multi-step reasoning and spatial-temporal understanding into a single generative video-language model. When addressing complex VideoQA tasks, the model trained with AoTD (as shown in (b)) enables to generate a step-by-step reasoning to get the correct answer. In contrast, previous models trained solely on question-answer pairs (as in (a)) generate only a final answer, often without intermediate reasoning, which can lead to incorrect conclusions." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 834, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing process and incorporate essential information for answering the queries effectively. This mechanism filters out low-quality or logically inconsistent reasoning paths. The remaining CoTs that pass this verification are then distilled into large generative video-language models, significantly enhancing both their performance and interpretability, ultimately leading to the development of more robust, accurate, and interpretable VideoQA systems.", + "bbox": [ + 88, + 470, + 482, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold: (i) we propose a novel approach for enhancing Video-LLMs by distilling high-quality CoTs into their instruction-tuning process. These CoTs capture step-by-step reasoning paths, improving both the model's performance and its interpretability; (ii) to automatically construct the CoTs for any dataset, we employ an agent-based system to decompose complex VideoQA questions into simpler sub-tasks, leveraging off-the-shelf vision models to handle each sub-task. The intermediate outputs from these models can therefore be collected as CoTs for addressing the corresponding visual question; (iii) through extensive experiments, we demonstrate that our distilled model outperforms existing methods across both multiple-choice and open-ended VideoQA benchmarks, enabling to deliver not only accurate answers but also comprehensive reasoning explanations.", + "bbox": [ + 89, + 592, + 482, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 844, + 232, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-language models (Video-LLMs). Recent works such as VideoLLaMA2 [4], LLaVA-NeXT-Video [54] and", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VideoChat2 [19], with their excellent architecture design and reasonable instruction-tuning data collection, have achieved a new level of zero-shot results in VideoQA task. However, current end-to-end models still lack interpretability for questions, as well as the ability to think and visually process complex problems in multiple steps, which is an important part for embodied learning and autonomous driving.", + "bbox": [ + 511, + 470, + 906, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Programming and Agents. With the progress of LLMs, some recent works [5, 15, 37, 45] begin to try to use LLM as planner to solve the complex reasoning task in real scenarios. They attempt to decompose the question into some easier sub-questions, and use different specialist models as agents to solve these sub-questions, and finally gather them to get the answer of the raw question. These models demonstrate a strong ability to obtain trustworthy answers based on the intermediate evidence they get, but they lag far behind the end-to-end model in terms of inference speed.", + "bbox": [ + 511, + 595, + 908, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Chain-of-Thoughts (CoTs). The potential of Chain-of-Thought (CoT) reasoning [40, 46] extends from NLP to the visual domain, highlighting a growing interest in applying this approach across various fields. Numerous studies have incorporated CoTs into visual understanding tasks [11, 30, 36, 55], utilizing powerful Multi-Modal Large Language Models (MLLMs) for generating CoTs or adopting tool-based architectures for sequential problem-solving.", + "bbox": [ + 511, + 748, + 908, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent innovations, for example, Visual Program Distillation (VPD) [16] and Fact [10] attempt to address these", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8524", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cc83821021dd8b9d6f71f7c08b74b07c06a7143507b7df54fa6523cd2d235fdb.jpg", + "image_caption": [ + "Figure 2. Overview on Agent-of-Thoughts Distillation (AoTD). Step 1: Selecting best-performing agents for each sub-task to construct an agent-based system. Step 2: Decomposing question into executable program and leveraging chosen models to solve it sequentially to generate execution trace. Step 3: The execution trace is converted and filtered by LLM to produce high quality natural language CoTs. Step 4: Distilling CoTs into Video-LLM with two forms of prompt, allowing it to achieve a balance between concise answers and comprehensive rationales. The final model is Video-LLM-AoTD." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "issues by maintaining the accuracy and diversity of CoTs, while leveraging MLLMs to generate them directly. These approaches decompose complex tasks into code programs, call upon expert models to handle sub-tasks, and utilize the resulting CoTs as training data to fine-tune visual-language models. This process significantly improves the models' ability to generate detailed rationales. Despite the progress in image understanding, there remains a notable oversight in video domains, where reasoning chains can be particularly effective due to the complex spatial-temporal dynamics of video understanding tasks. This is the focus of our paper.", + "bbox": [ + 88, + 520, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent Work. In the recent literature, we notice two work that share similar idea with ours, specifically, Video-STaR [56] construct CoTs using videos and existing labels for instruction-tuning, yet they do not develop an agent-based system. Meanwhile, MotionEpic [8] introduces a Video-of-Thought reasoning framework that integrates video spatial-temporal scene graphs, marking a significant stride towards more nuanced video reasoning.", + "bbox": [ + 89, + 690, + 482, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Agent-of-Thoughts Distillation", + "text_level": 1, + "bbox": [ + 89, + 828, + 372, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach, termed Agent-of-Thoughts Distillation (AoTD), to enhance the Video-LLMs by training them with multi-step Chain-of-Thoughts.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we start by developing an agent-based video understanding system, to generate multi-step reasoning chains that address complex video questions. These reasoning chains are then distilled into one Video-LLM through instruction-tuning. By combining the strengths of agent-based systems and large generative models, our proposed AoTD enables to build more reliable and interpretable VideoQA systems. Figure 2 illustrates the entire process of our method.", + "bbox": [ + 511, + 520, + 906, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 664, + 718, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video clip with $t$ frames, $\\mathcal{V} = \\{x_1, \\ldots, x_t\\}$ , and a set of $n$ questions $\\mathcal{Q} = \\{q_1, q_2, \\ldots, q_n\\}$ , our goal is to train a Video-LLM capable of producing both concise answers and comprehensive rationales. Depending on the suffix prompt $p_s$ , the model can generate different types of outputs. The process can be formulated as:", + "bbox": [ + 511, + 686, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{a _ {i}, \\mathcal {S} _ {i} \\right\\} = \\Phi (\\mathcal {V}, q _ {i}, p _ {s}), \\mathcal {S} _ {i} = \\{\\emptyset \\} \\text {o r} \\left\\{s _ {i, 1}, \\dots , s _ {i, k} \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 785, + 883, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q_{i}$ denotes the $i$ -th question, $a_{i}$ is the answer in freeform text, and $S_{i}$ represents the rationale, consisting of the reasoning process. If the prompt specifies to only generate the answer, $S_{i} = \\{\\emptyset\\}$ . Otherwise, if the prompt requires the generation of rationales, $S_{i} = \\{s_{i,1}, \\ldots, s_{i,k}\\}$ , where each $s_{i,j}$ corresponds to a reasoning step.", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8525", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/68ec4f03f282cec6fecd90b25251552a7502eceaf5be1c6ac64f55f3fadc66db.jpg", + "image_caption": [ + "Video input", + "Question: Which object did the person put down before they held the book? Possible answers: (A) The broom. (B) The paper. (C) The door. (D) The cup." + ], + "image_footnote": [ + "def execute_command(video, query, possible_answers): hold Clip $=$ Filter Frames_with_act(video, 'person held the boo before_hold Clip $=$ trim(video,end=hold clip.sta put clip $=$ Filter Frames_with_act(before_hold_cl person put down somethi person clip $=$ Find(place,'person') put_objs $=$ Query_Objects(person clip, 'object put down by the perso info $=$ {'object put down by the person': put_obj answer $=$ select_answer(query,info,possible的答案 return answer", + "Ground Truth Hold a book: Frame[6.64,7.45,...,16.32]" + ], + "bbox": [ + 93, + 87, + 205, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33473034039d4be3540613bf8c2f0129853df88be5023d9374259f6dcc822f4a.jpg", + "image_caption": [ + "frame 2.88", + "Ground Truth Put something: Frame[2.88,3.41,4.78,6.12,6.64]", + "Ground Truth Person:[128,64,244,371] in frame 2.88" + ], + "image_footnote": [ + "Object Detection Eval Person:[119,0,241,408] in frame 2.88" + ], + "bbox": [ + 555, + 147, + 620, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28053488798311b6240ae2401679e65af7099e28bef2b290410a2c33a74d9364.jpg", + "image_caption": [ + "frame 6.64", + "Program Output The broom.", + "Ground Truth Answer The broom.", + "Ground Truth Put a broom somewhere" + ], + "image_footnote": [ + "Evaluation IoU 65.8%" + ], + "bbox": [ + 622, + 147, + 712, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f5814326e3b4cbfed30c218ec9301e3a608aec00f0c573f60075b8813f43a901.jpg", + "image_caption": [ + "frame 7", + "Figure 3. Program execution process in an agent-based system. We uniformly sample 32 frames from the video, and to ensure scale consistency, the frame ids of key frames are normalized into these 32 frames. The blue boxes represent the program execution steps, the red boxes denote the ground truth for each step. The combination of red and yellow boxes represents one example process of evaluating Object detection model candidates." + ], + "image_footnote": [ + "Filter_frames_with actu hold Clip: Frame 7 to 15" + ], + "bbox": [ + 99, + 311, + 163, + 397 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/064d0c33aa2d1448fea26394963b6b091ec7e32e3ac5c2b5551a0df957bb733b.jpg", + "image_caption": [ + "frame 15" + ], + "image_footnote": [], + "bbox": [ + 176, + 311, + 240, + 397 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b2bef87da79297f89f8430cee1e8a8925bfcdc034f9d33183a7c663f407b30e9.jpg", + "image_caption": [ + "Ground Truth \nFrame[2.88,3.41,4.78, 6.12,6.64]", + "trim\nbefore_hold Clip:\nFrame 0 to 7", + "frame 0" + ], + "image_footnote": [], + "bbox": [ + 250, + 311, + 312, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d593a95de6c5bae5dc1613aebc8019086b2472167d1aae2446cbb690be4fa80f.jpg", + "image_caption": [ + "frame 7" + ], + "image_footnote": [], + "bbox": [ + 320, + 311, + 383, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7ee3066790d72441e1c60cd4265410b280fe81a97b442e8115d29c956db4c94.jpg", + "image_caption": [ + "Filter Frames with act put Clip: Frame 2 to 7", + "frame 2" + ], + "image_footnote": [], + "bbox": [ + 395, + 311, + 457, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7b8e75d154e539de4bde52c453f260139df2b6283158532a65bdacdbcbf5ff0d.jpg", + "image_caption": [ + "frame 7" + ], + "image_footnote": [], + "bbox": [ + 460, + 311, + 544, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa4388d34ac4a7bb7dce5958c6f8dbdd68dbc98f96024c08f1c6f7fb1a5437f6.jpg", + "image_caption": [ + "Find person Clip:[120,0,239, 413] in frame 2 ...", + "frame 2" + ], + "image_footnote": [], + "bbox": [ + 555, + 311, + 620, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/23a55f3e19616eaa13cd27195ab0d75a545a2ceee26f47814bdef5a57f5b4c2a.jpg", + "image_caption": [ + "frame 7" + ], + "image_footnote": [], + "bbox": [ + 645, + 311, + 707, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/745f32b8e47f4bd4e024ad8eb436aed743fa1fe27d1b25abdf9c5a793e137487.jpg", + "image_caption": [ + "Query_Objects\nput_objects: broom", + "frame 2" + ], + "image_footnote": [], + "bbox": [ + 720, + 311, + 784, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d20ccc585c3a81054b4e64d2582800c4a1aa792b891431597e221db85c6225ab.jpg", + "image_caption": [ + "frame 7" + ], + "image_footnote": [], + "bbox": [ + 805, + 311, + 867, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Unlike existing models that are instruction-tuned on VideoQA datasets using simple question-answer pairs, which bypass the intermediate thought process, our approach emphasizes the importance of training with Chain-of-Thoughts (CoTs). In the following section, we outline the process for generating high-quality CoTs from existing VideoQA datasets.", + "bbox": [ + 89, + 489, + 483, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CoT Construction with Agent-based System", + "text_level": 1, + "bbox": [ + 89, + 603, + 464, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent work, such as STAR [42], has introduced the idea of employing executable symbolic programs, to directly decompose questions into sub-tasks. When combined with scene graphs that contain comprehensive video information from key frames—such as object locations, interactions, and actions—these programs facilitate the generation of concise Chain-of-Thoughts (CoTs) through direct execution of symbolic operations. However, datasets of this nature are limited in scale, we therefore propose to first build an agent-based system, capable of breaking down complex questions into simpler sub-tasks, and the intermediate outputs from this system can then be employed to construct CoTs for any existing VideoQA dataset.", + "bbox": [ + 89, + 625, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Agent-based VideoQA. Assuming we are given a video input $(\\mathcal{V})$ , questions $(\\mathcal{Q})$ , and a set of visual models $(\\mathcal{M} = \\{\\phi_{\\mathrm{act}}, \\phi_{\\mathrm{det}}, \\dots, \\phi_{\\mathrm{qa}}\\})$ , an LLM-based agent core $(\\pi(\\cdot))$ processes the question along with the documentation of the visual models $(\\mathcal{T})$ , which includes variables and function-", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ba208850aa8a27cae03d9cd702c94f5f5d8efce19508089e06e0e467aa20830f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sub-taskModel nameMetricNumber (%)
question decompositionCodeQwen1.5-Chat (7B) [2]52.7
GPT-3.5-Turbo [31]Acc73.1
DeepSeek-Coder-Instruct (6.7B) [6]85.7
object detectionOWL-ViT v1 [26]47.3
GLIP [20]IoU58.9
OWL-ViT v2 [28]63.0
temporal groundingLITA (13B) [18]11.7 / 20.2
TimeChat (7B) [35]IoU / Recall13.9 / 23.1
UniVTG [23]24.7 / 35.3
action recognitionInternVideo2 (1B) [39]7.6
Open-VCLIP [41]Top1-Acc8.9
LLaVA-NeXT-Video-DPO (7B) [54]18.2
question answeringLLaMA-VID (7B) [21]43.5
SeViLA [48]Acc46.5
LLaVA-NeXT-Video-DPO (7B) [54]53.4
", + "bbox": [ + 517, + 487, + 903, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Sub-tasks definition and evaluation results. We choose 3 model candidates for each sub-task and evaluate them in STAR training set with the corresponding metrics. Models with best performance are placed at the bottom of each column.", + "bbox": [ + 511, + 724, + 906, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "alities. The agent subsequently decomposes the question into sub-tasks formatted as Python code, and resolves them by invoking the appropriate visual models through function calls. It is important to note that the visual models can be arranged in various orders depending on the specific question, ensuring flexibility in problem-solving.", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as illustrated by the example in Figure 3, the", + "bbox": [ + 532, + 885, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8526", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "question is first decomposed into a series of sub-tasks, including temporal grounding, object detection, and question answering. The corresponding specialized models are then executed sequentially to address these sub-tasks, ultimately yielding the final answer $y_{i}$ :", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\{\\phi_ {\\text {g r o u n d}}, \\phi_ {\\det }, \\phi_ {\\mathrm {q a}} \\right\\} := \\pi \\left(q _ {i}, \\mathcal {T}\\right), \\\\ y _ {i} = \\phi_ {\\text {g r o u n d}} (\\mathcal {V}) \\rightarrow \\phi_ {\\det} (\\mathcal {V}) \\rightarrow \\phi_ {\\mathrm {q a}} (\\mathcal {V}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 176, + 411, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CoT Construction. To ensure the correctness of outputs at all the intermediate steps, we leverage the training set from STAR for hyperparameter tuning, enabling us to identify the most effective model for each sub-task within the agent-based system. By following the provided programs, we evaluate the performance of the corresponding vision models on tasks such as object detection and action recognition. Given the availability of complete reasoning chains, we independently assess each sub-task using ground truth data for all preceding steps.", + "bbox": [ + 89, + 220, + 483, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1, we present the evaluation results for the various sub-tasks. Specifically, for question decomposition, we compare several code LLMs, with DeepSeekCoder-Instruct achieving the highest performance, outperforming even GPT-3.5-Turbo. In object detection, OWLViT v2 records the highest Intersection over Union (IoU) score, showcasing its superior open-vocabulary detection capability. The results for temporal grounding indicate that while UniVTG leads in performance, there remains a need for further advancements in this area. In action recognition, our evaluations show that generative models outperformed discriminative models, likely due to the fine-grained action list provided by the STAR dataset. However, the performance of both model types reveals significant room for improvement. Finally, in the one-hop question answering sub-task, all models perform admirably, with LLaVA-NeXT-Video-DPO standing out as a top performer, consistent with its strong results on other benchmarks.", + "bbox": [ + 89, + 372, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With these high-performing models, we implement the agent-based approach on VideoQA datasets that consist solely of QA pairs. During the execution of the programs, we record all intermediate outputs to construct the CoTs. Since the outputs from these vision models vary in format—such as bounding boxes and free-form text—we employ another LLM to translate the execution trace into natural language for better use in the distillation process. Detailed examples are provided in Appendix C.", + "bbox": [ + 89, + 643, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. CoT Verification", + "text_level": 1, + "bbox": [ + 89, + 787, + 254, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To refine the quality of reasoning chains for VideoQA samples, we implement a two-step verification: (i) we filter execution traces to retain only those, where the program can reach correct output. For multiple-choice datasets, the output must match the correct answer exactly, while for open-ended datasets, we prompt the LLM to verify correctness,", + "bbox": [ + 89, + 809, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f4a343deca7b142e248d2934e3e4ef9a109335e927428e83f33a8eff24bbf846.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDescription# Labels# CoTs
AGQACompositional25.0K5.4K
ANetQACompositional25.0K3.6K
STARCompositional45.7K11.2K
NExT-QATemporal & Causal34.1K12.1K
CLEVRERSpatial & Temporal21.0K-
EgoQAEgo-centric7.8K-
Total158.6K32.3K
", + "bbox": [ + 542, + 89, + 874, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Dataset statistics. The column \"#Labels\" indicates the number of VideoQA pairs, which include the video, query, possible answers (multiple-choice), and the correct answer. \"#CoTs\" refers to the number of CoTs generated using our agent-based system for each dataset.", + "bbox": [ + 511, + 218, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "accounting for format differences; (ii) we prompt the LLM to evaluate the logical coherence and usefulness of the reasoning chains in solving the problem. The model assesses whether the CoTs follow a clear, step-by-step reasoning process and provides a binary evaluation ('Yes' or 'No') to indicate their quality (detailed prompts are included in Appendix D). This two-step approach ensures that only high-quality CoTs are retained for further distillation.", + "bbox": [ + 511, + 308, + 903, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 2, we provide the statistics for the remaining generated CoTs for different datasets. We primarily select compositional QA datasets, as these require the model to process spatial-temporal information from different events comprehensively.", + "bbox": [ + 511, + 436, + 903, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Step-by-step Distillation", + "text_level": 1, + "bbox": [ + 511, + 541, + 733, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe the process of distilling the generated CoTs into a Video-LLM. This distillation enhances the model's ability for spatial-temporal video understanding and multi-step reasoning, thereby improving its performance on complex VideoQA tasks.", + "bbox": [ + 511, + 569, + 903, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In specific, using the generated CoTs, we can build the dataset $D = \\{(\\mathcal{V}_j,q_j,\\hat{y}_j,c_j,p_s)\\}_{j = 1}^N$ , where $N$ is the total number of samples in the distilling dataset, $\\mathcal{V}_j$ is the video input, $q_{j}$ is the question, $\\hat{y}_j$ is the ground-truth answer, $c_{j}$ is the generated CoT, $p_{s}$ is the task-specific suffix prompt, to distinguish different tasks, for example, for multiple-choice VQA, the prompt can be: \"Answer with the option's letter from the given choices directly and only give the best option\", and for open-ended VQA, the prompt can be: \"Answer in one word or phrase\". Please refer to detailed prompts in Appendix D.", + "bbox": [ + 511, + 652, + 905, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At distillation stage, we minimize the cross-entropy loss of predicting both the answer and the CoTs, we replace the suffix prompt $p_{s}$ with \"Explain the rationale to answer the question\", to indicate whether we want a question answer or a rationale to explain the thinking steps. Following [10]", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8527", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cddd742c8d91e51cb002c8205855594521b7d735f28b8dd74af340e6e7ffb96b.jpg", + "image_caption": [ + "Figure 4. Visualization of rationales. LLaVA-NeXT-Video-AoTD can output rationales containing both spatial-temporal grounding of key information and step-by-step thinking process to solve the question." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 486, + 333 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56007bbdf03faa4e89fc8b0794b261f650a83e927808135b2cc00ea59b927197.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 87, + 903, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and [16], our optimization objective is:", + "bbox": [ + 89, + 383, + 349, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\lambda \\mathcal {L} _ {\\text {r a t i o n a l e}} \\\\ = \\sum_ {j = 1} ^ {N} \\ell (\\Phi (\\mathcal {V} _ {j}, q _ {j}, p _ {s}), \\hat {y} _ {j}) + \\lambda \\ell (\\Phi (\\mathcal {V} _ {j}, q _ {j}, p _ {s}), c _ {j}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 411, + 460, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we set $\\lambda$ to 1 to ensure the importance of answer and rationale are equally considered. Notice that, not all the QA pairs can generate qualified CoT. In that case, the $\\mathcal{L}_{\\mathrm{rationale}}$ will be set to 0.", + "bbox": [ + 89, + 484, + 483, + 546 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24465b0e0f20c366b1f2226a485cf8200148735254a8e7e780c05e43a68c8928.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSizeTypeTrainEval
traineval
MC-VQA
STAR [42]45.7K7.1KCompositional✓✓
NExT-QA [43]34.1K5.0KTemporal & Causal✓✓
CLEVRER [47]21.0K-Spatial-temporal✓✗
Perception-Test [33]-11.5KGeneral✗✓
MVBench [19]-4.0KGeneral✗✓
VideoMME [9]-2.7KGeneral✗✓
VSIBench [44]-5.0Kspatial-temporal✗✓
OE-VQA
AGQA [14]25.0K2.0KCompositional✓✓
ANetQA [50]25.0K2.0KCompositional✓✓
EgoQA [13]7.8K-Ego-centric✓✗
Activitynet-QA [49]-8.0KGeneral✗✓
Video-ChatGPT [24]-3.0KGeneral✗✓
", + "bbox": [ + 91, + 556, + 478, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Training and evaluation datasets statics.", + "bbox": [ + 140, + 784, + 429, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 814, + 220, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present the experimental setup (Sec. 4.1) and results on various VideoQA benchmarks (Sec. 4.2). Extensive ablation studies have also been conducted to further examine the contributions of our approach in Sec. 4.3, and", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "an evaluation on the quality of rationales generated by the distilled model is made in Sec. 4.4.", + "bbox": [ + 511, + 383, + 906, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 428, + 702, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Base model. We use LLaVA-NeXT-Video (7B) [54] (LNV for short) as base Video-LLM, which has shown remarkable performance on image-centric tasks, for example image question answering [51]. We present comparison on naive instruction-tuning with video question answering dataset or with additional CoT distillation. For CoT conversion and verification, we prompt LLaMA-3.1-8B with the manually-designed instruction and some in-context examples. Detailed prompts are provided in Appendix D.", + "bbox": [ + 511, + 450, + 906, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction tuning. We utilize both multiple-choice and open-ended QA data, along with the generated CoTs, to fine-tune the base video question answering model, as summarised in Table 2. The resulting distilled model is named LLaVA-NeXT-Video-AoTD (LNV-AoTD for short). Additionally, as baseline, we also train another version of the model using only the basic QA data, which we refer to as LLaVA-NeXT-Video-Instruct (LNV-Instruct for short).", + "bbox": [ + 511, + 593, + 905, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation benchmarks. We conduct extensive evaluations on Multiple-Choice Video QA (MC-VQA) and Open-Ended Video QA (OE-VQA). We report the top-1 accuracy for all MC benchmarks, which means the proportion of the output equal to the answer. We report a GPT-assessed Acc. and Score with the help of GPT-3.5-turbo-0613 for all OE benchmarks. For each question, GPT delivers a binary decision indicating whether the output is correct or incorrect, along with a similarity score reflecting the degree of alignment between the output and the correct answer. The term 'Acc.' refers to the percentage of correct outputs, while 'Score' represents the average similarity scores. For the", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8528", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b5693ec54e54aabfa2d623aed0481be0e214ab27dfe26246f9d31d822886deb4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMVBench (Acc.)VideoMME (Acc.)STAR (Acc.)NExT-QA (Acc.)Perception-Test (Acc.)VSIBench (Acc.)
Proprietary Models
Gemini 1.5 Pro [12]-75.0---45.4
GPT4-V [32]43.759.9----
Open-source Models
LLaMA-VID (7B) [21]41.925.9--44.6-
Video-LLaVA (7B) [22]41.039.9--44.3-
VideoChat2 (7B) [19]51.133.759.0*68.6*47.3-
VideoLLaMA2 (7B) [4]53.445.158.5*62.3*49.6-
LLaVA-NeXT-Video (7B) [54]46.5*41.0*52.4*61.6*47.5*19.7*
LLaVA-NeXT-Video-Instruct (7B)53.443.272.277.150.326.7
LLaVA-NeXT-Video-AoTD (7B)55.645.074.377.650.628.8
", + "bbox": [ + 125, + 88, + 870, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison with Video-LLMs on MC-VQA benchmarks. LLaVA-NeXT-Video-AoTD outperforms all other baselines the and the version without CoT distillation. * means results reproduced by ourselves. Results without signs are retrieved from [4] and [44].", + "bbox": [ + 89, + 301, + 906, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluation on AGQA and ANetQA, due to the large volume of test set, we test on a subset of samples. We evenly select the benchmark in-domain and out-of-domain for testing to ensure a comprehensive and reasonable evaluation of the model capability. Noted that though VSIBench has both MC and OE questions, it doesn't need GPT for score, so we classify it into MC benchmarks for convenience. Detailed statistics for evaluation benchmarks are shown in Table 3.", + "bbox": [ + 89, + 335, + 485, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 467, + 284, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We divide the comparison into two parts: the first focuses on comparing the distilled model with other baselines, while the second examines the difference between the instruct version and the AoTD version. Note that, the latter part will be mainly compared and discussed, to demonstrate the model's improvement relative to its previous performance, as well as establishing the transferability of the method across models.", + "bbox": [ + 89, + 488, + 483, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MC-VQA performance. As shown in Table 4, our LLaVA-NeXT-Video-AoTD achieves superior performance across all benchmarks. Several key observations can be made: (i) comparing to the base model, even a simple instruction-tuning on certain VideoQA datasets significantly enhances the model's question-answering performance. This improvement is notable, as the base model was primarily trained on static images and struggled with video understanding; (ii) our model, instruction-tuned with CoT distillation, demonstrates further performance enhancements across all benchmarks, particularly on the compositional VideoQA benchmark (STAR) and comprehensive benchmarks (VideoMME, MVBench). This suggests that our AoTD method effectively improves the model's ability to address complex problems and interpret spatial-temporal scenes; (iii) the distilled model consistently outperforms all other baselines across almost all benchmarks, even when compared to more powerful models. This finding shows that our method effectively bridges performance gaps created by varying model components.", + "bbox": [ + 89, + 598, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OE-VQA performance. As shown in Table 5, LLaVA-NeXT-Video-AoTD outperforms the Instruct variant across all open-ended VideoQA benchmarks. Notably, it achieves a greater percentage increase compared to the MC-VQA benchmarks, suggesting that CoT distillation may be more effective for open-ended generation than for multiple-choice selection. While the distilled model scores higher than most models listed in the table, it does not surpass LLaVA-NeXT-Video on certain benchmarks. We conjecture this is due to the model's extensive training on images, that can also benefit the question answering without requiring complex reasonings, as also suggested by the findings in VideoLLaMA2 [4]. Additionally, the inherent challenges of evaluating open-ended VQA may influence the results. Assessments conducted by GPT can be biased or inaccurate, and the metrics we employ primarily indicate general trends rather than providing absolute accuracy.", + "bbox": [ + 511, + 335, + 906, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 603, + 666, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on CoT filtering. To prove the effectiveness of our filtering mechanism, we trained an alternative model without CoT filtering while maintaining all other settings, i.e., using 36.3K verified CoTs for distillation. As shown in Table 9, the model's performance declines significantly on both the Multiple-Choice VQA and Open-Ended VQA benchmarks when the CoT filtering mechanism is not utilized. This confirms that employing large language models (LLMs) to filter CoTs is crucial for enhancing data quality.", + "bbox": [ + 511, + 625, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on model transferability. As AoTD is a distillation method that leverages Chain-of-Thoughts (CoTs), it can theoretically be applied to any Video-LLMs. To assess the transferability of our method, we conduct experiments on another very recent model, LLaVA-OneVision (7B) [3]. As shown in Table 9, our method also demonstrates significant improvements on the benchmarks, showing the transferability and robustness of the approach. Due to the rapid advancements in the computer vision field, evaluating all", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8529", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/79c8196bcb3e5bb302a71ba3b64537fdac3677ed8b8d94b09128420b55ff4f0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelANetQA (Acc./Score)AGQA (Acc./Score)Corr.Video-ChatGPT (Score)ActivityNet (Acc./Score)
Deta.Cont.Temp.Cons.
Proprietary Models
Gemini 1.5 Pro [12]-------56.7/-
GPT4-V [32]--4.093.884.373.944.0259.5/-
Open-Source Models
VideoLLaMA (7B) [53]--1.962.182.161.821.7912.4/1.1
Video-ChatGPT (7B) [24]--2.502.572.692.162.2035.2/2.7
LLaMA-VID (7B) [21]--2.963.003.532.462.5147.4/3.3
Video-LLaVA (7B) [22]--2.872.943.442.452.5145.3/3.3
VideoChat2 (7B) [19]--3.022.883.512.662.8149.1/3.3
VideoLLaMA2 (7B) [4]--3.093.093.682.633.2549.9/3.3
LLaVA-NeXT-Video (7B) [54]46.4/3.3*27.4/2.2*3.26*3.22*3.77*2.47*2.99*54.3/3.2*
LLaVA-NeXT-Video-Instruct (7B)47.1/3.159.3/3.42.962.813.352.422.8250.0/3.3
LLaVA-NeXT-Video-AoTD (7B)53.9/3.460.9/3.63.113.003.602.412.9153.2/3.4
", + "bbox": [ + 155, + 88, + 841, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models and benchmarks is prohibitively infeasible. Thus, we focus on assessing some representative models against selected benchmarks to provide a representative evaluation.", + "bbox": [ + 89, + 371, + 482, + 416 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f7476e825bd72f9d7530ff5fb097ca67a55bfbde0ae0416fcb964229440c15a8.jpg", + "table_caption": [ + "Table 5. Comparison with Video-LLMs on OE-VQA benchmarks. LLaVA-NeXT-Video-AoTD improves performance in all open-ended benchmarks compared with the Instruct version. * means results reproduced by ourselves. Results without signs are retrieved from [4]." + ], + "table_footnote": [], + "table_body": "
ModelFilteringMVBench (Acc.)STAR (Acc.)AGQA (Acc. / Score)
LNV-AoTDX53.773.359.5/3.5
LNV-AoTD✓55.674.360.9/3.6
Onevision-Instruct-59.275.865.6/3.7
Onevision-AoTD✓60.576.665.7/3.7
", + "bbox": [ + 102, + 419, + 468, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Evaluation on Rationales", + "text_level": 1, + "bbox": [ + 89, + 539, + 318, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify whether the model has effectively learned multistep reasoning through CoTs distillation, we analyze the rationales generated by the model. Specifically, we extract and evaluate the temporal and spatial information embedded within these rationales. This approach extends beyond merely assessing the correctness of the final answer, which could be influenced by biases or other external factors. By examining the reasoning process in detail, it enables a more accurate understanding of the model's ability to perceive and reason about spatial and temporal relationships.", + "bbox": [ + 88, + 560, + 482, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation protocols. We randomly select 200 samples from the STAR validation set and run inference on them using the suffix prompt, recording the generated rationales. From these rationales, we extract the predicted temporal windows and bounding boxes, comparing them to the ground truth. For the spatial part, we calculate the IoU between the predicted and ground truth bounding boxes. For the temporal part, we compute IoU and Recall, leveraging the frame-level annotations provided in the dataset.", + "bbox": [ + 88, + 715, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation results. Table 7 presents the evaluation results. For comparison, we also test UniVTG for temporal reasoning and OWL-ViT v2 for spatial reasoning. The re", + "bbox": [ + 88, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed677f5ba042f6499b9267ef48730d5abf5f72f63689e51714eba38728ae60ca.jpg", + "table_caption": [ + "Table 6. Ablation results of CoT filtering and transferability." + ], + "table_footnote": [], + "table_body": "
ModelTemporal GroundingSpatial Grounding
IoU (%)Recall (%)IoU (%)
UniVTG22.831.0-
OWL-ViT v2--64.7
LNV-InstructXXX
LNV-AoTD21.734.045.2
", + "bbox": [ + 514, + 368, + 903, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Temporal and spatial abilities evaluation results.", + "bbox": [ + 537, + 462, + 877, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sults show that LNV-Instruct struggles to generate valid rationales, even when using the suffix prompt. In contrast, LNV-AoTD demonstrates comparable performance to specialized models in both spatial and temporal reasoning, indicating that the model successfully acquired these abilities through the distillation process.", + "bbox": [ + 511, + 486, + 906, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 589, + 633, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Agent-of-Thoughts Distillation (AoTD), that aims to distill multi-step reasoning and spatial-temporal understanding into a large video-language model (Video-LLM). Our method introduces an agent-based system that automates the generation of Chain-of-Thoughts (CoTs) from various VideoQA datasets, by breaking down complex questions into manageable sub-tasks that can be addressed by specialized vision models. Extensive experiments validate that the distilled model significantly enhances performance on both MC-VQA and OE-VQA benchmarks, underscoring the effectiveness of our approach. We believe AoTD represents a promising future direction for advancing the reasoning abilities in Video-LLMs.", + "bbox": [ + 511, + 618, + 906, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 840, + 671, + 858 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is funded by National Key R&D Program of China (No.2022ZD0161400).", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8530", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1", + "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 4", + "[3] Li Bo, Zhang Yuanhan, Guo Dong, Zhang Renrui, Li Feng, Zhang Hao, Zhang Kaichen, Li Yanwei, Liu Ziwei, and Li Chunyuan. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7", + "[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. arXiv preprint arXiv:2406.07476, 2024. 2, 7, 8", + "[5] Rohan Choudhury, Koichiro Niinuma, Kris M. Kitani, and Laszlo A. Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 2", + "[6] Guo Daya, Zhu Qihao, Yang Dejian, Dong Zhenda Xie, Kai, Zhang Wentao, Chen Guanting, Bi Xiao, Y. Wu, Y.K. Li, Luo Fuli, and Liang Yingfei, Xiongand Wenfeng. Deepseek-coder: When the large language model meets programming - the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 4", + "[7] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 1", + "[8] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In Proceedings of the International Conference on Machine Learning, 2024. 3", + "[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 6", + "[10] Minghe Gao, Shuang Chen, Liang Pang, Yuan Yao, Jisheng Dang, Wenqiao Zhang, Juncheng Li, Siliang Tang, Yueting Zhuang, and Tat-Seng Chua. Fact: Teaching mllms with faithful, concise and transferable rationales. In ACM Multimedia, 2024. 2, 5, 1" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Timin Gao, Peixian Chen, Mengdan Zhang, Chaoyou Fu, Yunhang Shen, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Xing Sun, Liujuan Cao, et al. Cantor: Inspiring multimodal chain-of-thought of mllm. In ACM Multimedia, 2024. 2", + "[12] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 8", + "[13] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6", + "[14] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 6", + "[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2", + "[16] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 1", + "[17] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent. In Advances in Neural Information Processing Systems, 2024. 1", + "[18] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In Proceedings of the European Conference on Computer Vision, 2024. 4", + "[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7, 8", + "[20] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 4", + "[21] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 4, 7, 8", + "[22] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8531", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Empirical Methods in Natural Language Processing, 2024. 1, 7, 8", + "[23] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In Proceedings of the International Conference on Computer Vision, 2023. 1, 4", + "[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 1, 6, 8", + "[25] Ahmad Mahmood, Ashmal Vayani, Muzammal Naseer, Salman Khan, and Fahad Khan. Vurf: A general-purpose reasoning and self-refinement framework for video understanding. arXiv preprint arXiv:2403.14743, 2024. 8", + "[26] Minderer Matthias, Gritsenko Alexey, Stone Austin, Neumann Maxim, Weissenborn Dirk, Dosovitskiy Alexey, Mahendran Aravindh, Arnab Anurag, Dehghani Mostafa, Shen Zhuoran, Wang Xiao, Zhai Xiaohua, Kipf Thomas, and Houlsby Neil. Simple open-vocabulary object detection with vision transformers. In Proceedings of the European Conference on Computer Vision, 2022. 4", + "[27] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 8", + "[28] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. In Advances in Neural Information Processing Systems, 2024. 4", + "[29] Arindam Mitra, Luciano Del Corro, Shweti Mahajan, Andres Codas, Clarisse Simoes, Sahaj Agarwal, Xuxi Chen, Anastasia Razdaibiedina, Erik Jones, Kriti Aggarwal, et al. Orca 2: Teaching small language models how to reason. arXiv preprint arXiv:2311.11045, 2023. 1", + "[30] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[31] OpenAI. Gpt-3.5-turbo system card. https:// platform.openuai.com/docs/models/gpt-3- 5-turbo,2023.4", + "[32] OpenAI. Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card, 2023.7,8", + "[33] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models. In Advances in Neural Information Processing Systems, 2023. 1, 6", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn-" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 8", + "[35] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 4", + "[36] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. In Advances in Neural Information Processing Systems, 2024. 2", + "[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2, 3", + "[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8", + "[39] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In Proceedings of the European Conference on Computer Vision, 2024. 1, 4", + "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, 2022. 2", + "[41] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In Proceedings of the International Conference on Machine Learning, 2023. 1, 4", + "[42] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Advances in Neural Information Processing Systems, 2021. 1, 4, 6", + "[43] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 6, 2", + "[44] Jihan Yang, Shusheng Yang, Anjali Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces. arXiv preprint arXiv:2412.14171, 2024. 6, 7", + "[45] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In Proceedings of the International Conference on Machine Learning, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8532", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, 2024. 2", + "[47] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: collision events for video representation and reasoning. In Proceedings of the International Conference on Learning Representations, 2020. 6", + "[48] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In Advances in Neural Information Processing Systems, 2023. 4", + "[49] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 1, 6", + "[50] Zhou Yu, Lixiang Zheng, Zhou Zhao, Fei Wu, Jianping Fan, Kui Ren, and Jun Yu. Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 3", + "[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 6", + "[52] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, 2023. 8", + "[53] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 1, 8", + "[54] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024. 2, 4, 6, 7, 8", + "[55] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2", + "[56] Orr Zohar, Xiaohan Wang, Yonatan Bitton, Idan Szpektor, and Serena Yeung-Levy. Video-star: Self-training enables video instruction tuning with any supervision. arXiv preprint arXiv:2407.06189, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 810 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8533", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_model.json b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9187296b10ba4f9a551966f5a9c7685ef162a65f --- /dev/null +++ b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.18, + 0.719, + 0.199 + ], + "angle": 0, + "content": "Yudi Shi\\(^{1,2}\\), Shangzhe Di\\(^{1,2}\\), Qirui Chen\\(^{1,2}\\), Weidi Xie\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.202, + 0.782, + 0.221 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Artificial Intelligence, Shanghai Jiao Tong University, China" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.223, + 0.794, + 0.241 + ], + "angle": 0, + "content": "\\(^{2}\\)Coop. Medianet Innovation Center, Shanghai Jiao Tong University, China" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.246, + 0.698, + 0.264 + ], + "angle": 0, + "content": "https://zhengrongz.github.io/AoTD/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.313, + 0.327, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.486, + 0.603 + ], + "angle": 0, + "content": "This paper tackles the problem of video question answering (VideoQA), a task that often requires multi-step reasoning and a profound understanding of spatial-temporal dynamics. While large video-language models perform well on benchmarks, they often lack explainability and spatial-temporal grounding. In this paper, we propose Agent-of-Thoughts Distillation (AoTD), a method that enhances models by incorporating automatically generated Chain-of-Thoughts (CoTs) into the instruction-tuning process. Specifically, we leverage an agent-based system to decompose complex questions into sub-tasks, and address them with specialized vision models, the intermediate results are then treated as reasoning chains. We also introduce a verification mechanism using a large language model (LLM) to ensure the reliability of generated CoTs. Extensive experiments demonstrate that AoTD improves the performance on multiple-choice and open-ended benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.629, + 0.222, + 0.644 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.85 + ], + "angle": 0, + "content": "Video Question Answering (VideoQA) refers to a critical task that offers a natural interface for human-machine interaction through language [33, 42, 43, 49]. This synergy of visual content and language enhances the accessibility of AI systems for the general public, allowing users to query complex visual content with natural language. By encompassing tasks such as action recognition, object detection, and scene understanding, VideoQA serves as a comprehensive benchmark for evaluating AI's ability to interpret videos, addressing the fundamental questions of 'who', 'what', 'when', and 'where' that are crucial to understand daily life activities, pushing the boundaries of what AI systems can interpret from dynamic visual content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.85, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Recent literature has primarily explored two avenues in VideoQA. The first involves training large video lan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.42 + ], + "angle": 0, + "content": "guage models (Video-LLMs) through direct instruction-tuning, using videos paired solely with corresponding questions and answers [1, 22, 24, 53]. While these models excel on public benchmarks, they often lack explainability and struggle with spatial-temporal grounding. This limitation hinders their ability to provide clear reasoning, which is essential for real-world applications where transparency and interpretability are critical [29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.606 + ], + "angle": 0, + "content": "Conversely, an emerging approach utilizes agent-based systems that decompose complex questions into manageable sub-tasks, each addressed by specialized tools [15, 17, 37]. The results are then aggregated to form a coherent answer. Theoretically, such approach naturally offers greater interpretability, as the reasoning process is divided into explainable steps that can be independently assessed. However, our experiments indicate that current video understanding tools are not strong enough for building reliable agent-based systems. In addition, the high memory demands and time-consuming nature of these systems present significant challenges for their practical use." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.907, + 0.699 + ], + "angle": 0, + "content": "In this paper, we aim to leverage the advantage of both research lines, enhancing Video-LLM by integrating Chain-of-Thoughts (CoTs) into instruction-tuning, with the CoTs being constructed from the outputs of specialized agent models, capturing the step-by-step reasoning procedure, as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.909, + 0.853 + ], + "angle": 0, + "content": "In specific, we start by systematically evaluating the off-the-shelf models tailored for atomic video understanding tasks, such as action recognition [39, 41] and language grounding [23], using well-annotated datasets. This comprehensive evaluation allows us to pinpoint the most effective tools for each sub-task, thus laying a robust foundation for constructing reliable chains. Moreover, this process also provides a critical assessment of the broader capabilities of visual models across general and complex scenes, offering valuable insights for future research within the community." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In addition, we introduce a verification mechanism using a large language model (LLM), designed to assess if the generated CoTs adhere to a clear, step-by-step reason-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.888, + 0.228, + 0.901 + ], + "angle": 0, + "content": "†: Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8523" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.089, + 0.836, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.402, + 0.908, + 0.459 + ], + "angle": 0, + "content": "Figure 1. Our method, AoTD, distills multi-step reasoning and spatial-temporal understanding into a single generative video-language model. When addressing complex VideoQA tasks, the model trained with AoTD (as shown in (b)) enables to generate a step-by-step reasoning to get the correct answer. In contrast, previous models trained solely on question-answer pairs (as in (a)) generate only a final answer, often without intermediate reasoning, which can lead to incorrect conclusions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.483, + 0.592 + ], + "angle": 0, + "content": "ing process and incorporate essential information for answering the queries effectively. This mechanism filters out low-quality or logically inconsistent reasoning paths. The remaining CoTs that pass this verification are then distilled into large generative video-language models, significantly enhancing both their performance and interpretability, ultimately leading to the development of more robust, accurate, and interpretable VideoQA systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.835 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold: (i) we propose a novel approach for enhancing Video-LLMs by distilling high-quality CoTs into their instruction-tuning process. These CoTs capture step-by-step reasoning paths, improving both the model's performance and its interpretability; (ii) to automatically construct the CoTs for any dataset, we employ an agent-based system to decompose complex VideoQA questions into simpler sub-tasks, leveraging off-the-shelf vision models to handle each sub-task. The intermediate outputs from these models can therefore be collected as CoTs for addressing the corresponding visual question; (iii) through extensive experiments, we demonstrate that our distilled model outperforms existing methods across both multiple-choice and open-ended VideoQA benchmarks, enabling to deliver not only accurate answers but also comprehensive reasoning explanations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.233, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Video-language models (Video-LLMs). Recent works such as VideoLLaMA2 [4], LLaVA-NeXT-Video [54] and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.593 + ], + "angle": 0, + "content": "VideoChat2 [19], with their excellent architecture design and reasonable instruction-tuning data collection, have achieved a new level of zero-shot results in VideoQA task. However, current end-to-end models still lack interpretability for questions, as well as the ability to think and visually process complex problems in multiple steps, which is an important part for embodied learning and autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.746 + ], + "angle": 0, + "content": "Visual Programming and Agents. With the progress of LLMs, some recent works [5, 15, 37, 45] begin to try to use LLM as planner to solve the complex reasoning task in real scenarios. They attempt to decompose the question into some easier sub-questions, and use different specialist models as agents to solve these sub-questions, and finally gather them to get the answer of the raw question. These models demonstrate a strong ability to obtain trustworthy answers based on the intermediate evidence they get, but they lag far behind the end-to-end model in terms of inference speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Visual Chain-of-Thoughts (CoTs). The potential of Chain-of-Thought (CoT) reasoning [40, 46] extends from NLP to the visual domain, highlighting a growing interest in applying this approach across various fields. Numerous studies have incorporated CoTs into visual understanding tasks [11, 30, 36, 55], utilizing powerful Multi-Modal Large Language Models (MLLMs) for generating CoTs or adopting tool-based architectures for sequential problem-solving." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Recent innovations, for example, Visual Program Distillation (VPD) [16] and Fact [10] attempt to address these" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8524" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.891, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.438, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Figure 2. Overview on Agent-of-Thoughts Distillation (AoTD). Step 1: Selecting best-performing agents for each sub-task to construct an agent-based system. Step 2: Decomposing question into executable program and leveraging chosen models to solve it sequentially to generate execution trace. Step 3: The execution trace is converted and filtered by LLM to produce high quality natural language CoTs. Step 4: Distilling CoTs into Video-LLM with two forms of prompt, allowing it to achieve a balance between concise answers and comprehensive rationales. The final model is Video-LLM-AoTD." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.521, + 0.483, + 0.688 + ], + "angle": 0, + "content": "issues by maintaining the accuracy and diversity of CoTs, while leveraging MLLMs to generate them directly. These approaches decompose complex tasks into code programs, call upon expert models to handle sub-tasks, and utilize the resulting CoTs as training data to fine-tune visual-language models. This process significantly improves the models' ability to generate detailed rationales. Despite the progress in image understanding, there remains a notable oversight in video domains, where reasoning chains can be particularly effective due to the complex spatial-temporal dynamics of video understanding tasks. This is the focus of our paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.812 + ], + "angle": 0, + "content": "Concurrent Work. In the recent literature, we notice two work that share similar idea with ours, specifically, Video-STaR [56] construct CoTs using videos and existing labels for instruction-tuning, yet they do not develop an agent-based system. Meanwhile, MotionEpic [8] introduces a Video-of-Thought reasoning framework that integrates video spatial-temporal scene graphs, marking a significant stride towards more nuanced video reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.373, + 0.846 + ], + "angle": 0, + "content": "3. Agent-of-Thoughts Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach, termed Agent-of-Thoughts Distillation (AoTD), to enhance the Video-LLMs by training them with multi-step Chain-of-Thoughts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Specifically, we start by developing an agent-based video understanding system, to generate multi-step reasoning chains that address complex video questions. These reasoning chains are then distilled into one Video-LLM through instruction-tuning. By combining the strengths of agent-based systems and large generative models, our proposed AoTD enables to build more reliable and interpretable VideoQA systems. Figure 2 illustrates the entire process of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.719, + 0.68 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Given a video clip with \\( t \\) frames, \\( \\mathcal{V} = \\{x_1, \\ldots, x_t\\} \\), and a set of \\( n \\) questions \\( \\mathcal{Q} = \\{q_1, q_2, \\ldots, q_n\\} \\), our goal is to train a Video-LLM capable of producing both concise answers and comprehensive rationales. Depending on the suffix prompt \\( p_s \\), the model can generate different types of outputs. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.786, + 0.885, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\left\\{a _ {i}, \\mathcal {S} _ {i} \\right\\} = \\Phi (\\mathcal {V}, q _ {i}, p _ {s}), \\mathcal {S} _ {i} = \\{\\emptyset \\} \\text {o r} \\left\\{s _ {i, 1}, \\dots , s _ {i, k} \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\( q_{i} \\) denotes the \\( i \\)-th question, \\( a_{i} \\) is the answer in freeform text, and \\( S_{i} \\) represents the rationale, consisting of the reasoning process. If the prompt specifies to only generate the answer, \\( S_{i} = \\{\\emptyset\\} \\). Otherwise, if the prompt requires the generation of rationales, \\( S_{i} = \\{s_{i,1}, \\ldots, s_{i,k}\\} \\), where each \\( s_{i,j} \\) corresponds to a reasoning step." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8525" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.207, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.235, + 0.182, + 0.245 + ], + "angle": 0, + "content": "Video input" + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.092, + 0.627, + 0.115 + ], + "angle": 0, + "content": "Question: Which object did the person put down before they held the book? Possible answers: (A) The broom. (B) The paper. (C) The door. (D) The cup." + }, + { + "type": "image_footnote", + "bbox": [ + 0.207, + 0.12, + 0.523, + 0.231 + ], + "angle": 0, + "content": "def execute_command(video, query, possible_answers): hold Clip \\(=\\) Filter Frames_with_act(video, 'person held the boo before_hold Clip \\(=\\) trim(video,end=hold clip.sta put clip \\(=\\) Filter Frames_with_act(before_hold_cl person put down somethi person clip \\(=\\) Find(place,'person') put_objs \\(=\\) Query_Objects(person clip, 'object put down by the perso info \\(=\\) {'object put down by the person': put_obj answer \\(=\\) select_answer(query,info,possible的答案 return answer" + }, + { + "type": "image_footnote", + "bbox": [ + 0.556, + 0.118, + 0.692, + 0.147 + ], + "angle": 0, + "content": "Object Detection Eval Person:[119,0,241,408] in frame 2.88" + }, + { + "type": "image_footnote", + "bbox": [ + 0.765, + 0.116, + 0.867, + 0.141 + ], + "angle": 0, + "content": "Evaluation IoU 65.8%" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.148, + 0.622, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.233, + 0.614, + 0.242 + ], + "angle": 0, + "content": "frame 2.88" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.148, + 0.714, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.233, + 0.707, + 0.242 + ], + "angle": 0, + "content": "frame 6.64" + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.171, + 0.85, + 0.195 + ], + "angle": 0, + "content": "Program Output The broom." + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.21, + 0.867, + 0.234 + ], + "angle": 0, + "content": "Ground Truth Answer The broom." + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.251, + 0.242, + 0.28 + ], + "angle": 0, + "content": "Ground Truth Hold a book: Frame[6.64,7.45,...,16.32]" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.252, + 0.377, + 0.28 + ], + "angle": 0, + "content": "Ground Truth \nFrame[2.88,3.41,4.78, 6.12,6.64]" + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.252, + 0.547, + 0.28 + ], + "angle": 0, + "content": "Ground Truth Put something: Frame[2.88,3.41,4.78,6.12,6.64]" + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.252, + 0.696, + 0.277 + ], + "angle": 0, + "content": "Ground Truth Person:[128,64,244,371] in frame 2.88" + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.251, + 0.85, + 0.268 + ], + "angle": 0, + "content": "Ground Truth Put a broom somewhere" + }, + { + "type": "image_footnote", + "bbox": [ + 0.101, + 0.284, + 0.238, + 0.313 + ], + "angle": 0, + "content": "Filter_frames_with actu hold Clip: Frame 7 to 15" + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.284, + 0.355, + 0.312 + ], + "angle": 0, + "content": "trim\nbefore_hold Clip:\nFrame 0 to 7" + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.284, + 0.53, + 0.312 + ], + "angle": 0, + "content": "Filter Frames with act put Clip: Frame 2 to 7" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.284, + 0.697, + 0.309 + ], + "angle": 0, + "content": "Find person Clip:[120,0,239, 413] in frame 2 ..." + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.284, + 0.816, + 0.302 + ], + "angle": 0, + "content": "Query_Objects\nput_objects: broom" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.313, + 0.165, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.4, + 0.152, + 0.408 + ], + "angle": 0, + "content": "frame 7" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.313, + 0.241, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.4, + 0.231, + 0.408 + ], + "angle": 0, + "content": "frame 15" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.313, + 0.313, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.4, + 0.298, + 0.408 + ], + "angle": 0, + "content": "frame 0" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.313, + 0.385, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.4, + 0.372, + 0.408 + ], + "angle": 0, + "content": "frame 7" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.313, + 0.458, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.4, + 0.445, + 0.408 + ], + "angle": 0, + "content": "frame 2" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.313, + 0.545, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.4, + 0.531, + 0.408 + ], + "angle": 0, + "content": "frame 7" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.313, + 0.621, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.4, + 0.608, + 0.408 + ], + "angle": 0, + "content": "frame 2" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.313, + 0.709, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.4, + 0.695, + 0.408 + ], + "angle": 0, + "content": "frame 7" + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.313, + 0.785, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.4, + 0.771, + 0.408 + ], + "angle": 0, + "content": "frame 2" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.313, + 0.869, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.4, + 0.855, + 0.408 + ], + "angle": 0, + "content": "frame 7" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.907, + 0.477 + ], + "angle": 0, + "content": "Figure 3. Program execution process in an agent-based system. We uniformly sample 32 frames from the video, and to ensure scale consistency, the frame ids of key frames are normalized into these 32 frames. The blue boxes represent the program execution steps, the red boxes denote the ground truth for each step. The combination of red and yellow boxes represents one example process of evaluating Object detection model candidates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.596 + ], + "angle": 0, + "content": "Discussion. Unlike existing models that are instruction-tuned on VideoQA datasets using simple question-answer pairs, which bypass the intermediate thought process, our approach emphasizes the importance of training with Chain-of-Thoughts (CoTs). In the following section, we outline the process for generating high-quality CoTs from existing VideoQA datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.465, + 0.62 + ], + "angle": 0, + "content": "3.2. CoT Construction with Agent-based System" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Recent work, such as STAR [42], has introduced the idea of employing executable symbolic programs, to directly decompose questions into sub-tasks. When combined with scene graphs that contain comprehensive video information from key frames—such as object locations, interactions, and actions—these programs facilitate the generation of concise Chain-of-Thoughts (CoTs) through direct execution of symbolic operations. However, datasets of this nature are limited in scale, we therefore propose to first build an agent-based system, capable of breaking down complex questions into simpler sub-tasks, and the intermediate outputs from this system can then be employed to construct CoTs for any existing VideoQA dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Agent-based VideoQA. Assuming we are given a video input \\((\\mathcal{V})\\), questions \\((\\mathcal{Q})\\), and a set of visual models \\((\\mathcal{M} = \\{\\phi_{\\mathrm{act}}, \\phi_{\\mathrm{det}}, \\dots, \\phi_{\\mathrm{qa}}\\})\\), an LLM-based agent core \\((\\pi(\\cdot))\\) processes the question along with the documentation of the visual models \\((\\mathcal{T})\\), which includes variables and function-" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.488, + 0.905, + 0.72 + ], + "angle": 0, + "content": "
Sub-taskModel nameMetricNumber (%)
question decompositionCodeQwen1.5-Chat (7B) [2]52.7
GPT-3.5-Turbo [31]Acc73.1
DeepSeek-Coder-Instruct (6.7B) [6]85.7
object detectionOWL-ViT v1 [26]47.3
GLIP [20]IoU58.9
OWL-ViT v2 [28]63.0
temporal groundingLITA (13B) [18]11.7 / 20.2
TimeChat (7B) [35]IoU / Recall13.9 / 23.1
UniVTG [23]24.7 / 35.3
action recognitionInternVideo2 (1B) [39]7.6
Open-VCLIP [41]Top1-Acc8.9
LLaVA-NeXT-Video-DPO (7B) [54]18.2
question answeringLLaMA-VID (7B) [21]43.5
SeViLA [48]Acc46.5
LLaVA-NeXT-Video-DPO (7B) [54]53.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.725, + 0.907, + 0.782 + ], + "angle": 0, + "content": "Table 1. Sub-tasks definition and evaluation results. We choose 3 model candidates for each sub-task and evaluate them in STAR training set with the corresponding metrics. Models with best performance are placed at the bottom of each column." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.885 + ], + "angle": 0, + "content": "alities. The agent subsequently decomposes the question into sub-tasks formatted as Python code, and resolves them by invoking the appropriate visual models through function calls. It is important to note that the visual models can be arranged in various orders depending on the specific question, ensuring flexibility in problem-solving." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Specifically, as illustrated by the example in Figure 3, the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8526" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "question is first decomposed into a series of sub-tasks, including temporal grounding, object detection, and question answering. The corresponding specialized models are then executed sequentially to address these sub-tasks, ultimately yielding the final answer \\( y_{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.177, + 0.412, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\{\\phi_ {\\text {g r o u n d}}, \\phi_ {\\det }, \\phi_ {\\mathrm {q a}} \\right\\} := \\pi \\left(q _ {i}, \\mathcal {T}\\right), \\\\ y _ {i} = \\phi_ {\\text {g r o u n d}} (\\mathcal {V}) \\rightarrow \\phi_ {\\det} (\\mathcal {V}) \\rightarrow \\phi_ {\\mathrm {q a}} (\\mathcal {V}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.484, + 0.372 + ], + "angle": 0, + "content": "CoT Construction. To ensure the correctness of outputs at all the intermediate steps, we leverage the training set from STAR for hyperparameter tuning, enabling us to identify the most effective model for each sub-task within the agent-based system. By following the provided programs, we evaluate the performance of the corresponding vision models on tasks such as object detection and action recognition. Given the availability of complete reasoning chains, we independently assess each sub-task using ground truth data for all preceding steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.644 + ], + "angle": 0, + "content": "As shown in Table 1, we present the evaluation results for the various sub-tasks. Specifically, for question decomposition, we compare several code LLMs, with DeepSeekCoder-Instruct achieving the highest performance, outperforming even GPT-3.5-Turbo. In object detection, OWLViT v2 records the highest Intersection over Union (IoU) score, showcasing its superior open-vocabulary detection capability. The results for temporal grounding indicate that while UniVTG leads in performance, there remains a need for further advancements in this area. In action recognition, our evaluations show that generative models outperformed discriminative models, likely due to the fine-grained action list provided by the STAR dataset. However, the performance of both model types reveals significant room for improvement. Finally, in the one-hop question answering sub-task, all models perform admirably, with LLaVA-NeXT-Video-DPO standing out as a top performer, consistent with its strong results on other benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.781 + ], + "angle": 0, + "content": "With these high-performing models, we implement the agent-based approach on VideoQA datasets that consist solely of QA pairs. During the execution of the programs, we record all intermediate outputs to construct the CoTs. Since the outputs from these vision models vary in format—such as bounding boxes and free-form text—we employ another LLM to translate the execution trace into natural language for better use in the distillation process. Detailed examples are provided in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.255, + 0.803 + ], + "angle": 0, + "content": "3.3. CoT Verification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To refine the quality of reasoning chains for VideoQA samples, we implement a two-step verification: (i) we filter execution traces to retain only those, where the program can reach correct output. For multiple-choice datasets, the output must match the correct answer exactly, while for open-ended datasets, we prompt the LLM to verify correctness," + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.09, + 0.875, + 0.215 + ], + "angle": 0, + "content": "
DatasetDescription# Labels# CoTs
AGQACompositional25.0K5.4K
ANetQACompositional25.0K3.6K
STARCompositional45.7K11.2K
NExT-QATemporal & Causal34.1K12.1K
CLEVRERSpatial & Temporal21.0K-
EgoQAEgo-centric7.8K-
Total158.6K32.3K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.219, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Table 2. Dataset statistics. The column \"#Labels\" indicates the number of VideoQA pairs, which include the video, query, possible answers (multiple-choice), and the correct answer. \"#CoTs\" refers to the number of CoTs generated using our agent-based system for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.905, + 0.43 + ], + "angle": 0, + "content": "accounting for format differences; (ii) we prompt the LLM to evaluate the logical coherence and usefulness of the reasoning chains in solving the problem. The model assesses whether the CoTs follow a clear, step-by-step reasoning process and provides a binary evaluation ('Yes' or 'No') to indicate their quality (detailed prompts are included in Appendix D). This two-step approach ensures that only high-quality CoTs are retained for further distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.905, + 0.512 + ], + "angle": 0, + "content": "In Table 2, we provide the statistics for the remaining generated CoTs for different datasets. We primarily select compositional QA datasets, as these require the model to process spatial-temporal information from different events comprehensively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.735, + 0.558 + ], + "angle": 0, + "content": "3.4. Step-by-step Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.905, + 0.646 + ], + "angle": 0, + "content": "In this section, we describe the process of distilling the generated CoTs into a Video-LLM. This distillation enhances the model's ability for spatial-temporal video understanding and multi-step reasoning, thereby improving its performance on complex VideoQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.906, + 0.82 + ], + "angle": 0, + "content": "In specific, using the generated CoTs, we can build the dataset \\( D = \\{(\\mathcal{V}_j,q_j,\\hat{y}_j,c_j,p_s)\\}_{j = 1}^N \\), where \\( N \\) is the total number of samples in the distilling dataset, \\( \\mathcal{V}_j \\) is the video input, \\( q_{j} \\) is the question, \\( \\hat{y}_j \\) is the ground-truth answer, \\( c_{j} \\) is the generated CoT, \\( p_{s} \\) is the task-specific suffix prompt, to distinguish different tasks, for example, for multiple-choice VQA, the prompt can be: \"Answer with the option's letter from the given choices directly and only give the best option\", and for open-ended VQA, the prompt can be: \"Answer in one word or phrase\". Please refer to detailed prompts in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "At distillation stage, we minimize the cross-entropy loss of predicting both the answer and the CoTs, we replace the suffix prompt \\( p_{s} \\) with \"Explain the rationale to answer the question\", to indicate whether we want a question answer or a rationale to explain the thinking steps. Following [10]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8527" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.487, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.088, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.342, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 4. Visualization of rationales. LLaVA-NeXT-Video-AoTD can output rationales containing both spatial-temporal grounding of key information and step-by-step thinking process to solve the question." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.385, + 0.351, + 0.4 + ], + "angle": 0, + "content": "and [16], our optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.412, + 0.462, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\lambda \\mathcal {L} _ {\\text {r a t i o n a l e}} \\\\ = \\sum_ {j = 1} ^ {N} \\ell (\\Phi (\\mathcal {V} _ {j}, q _ {j}, p _ {s}), \\hat {y} _ {j}) + \\lambda \\ell (\\Phi (\\mathcal {V} _ {j}, q _ {j}, p _ {s}), c _ {j}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Here, we set \\(\\lambda\\) to 1 to ensure the importance of answer and rationale are equally considered. Notice that, not all the QA pairs can generate qualified CoT. In that case, the \\(\\mathcal{L}_{\\mathrm{rationale}}\\) will be set to 0." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.558, + 0.48, + 0.78 + ], + "angle": 0, + "content": "
DatasetSizeTypeTrainEval
traineval
MC-VQA
STAR [42]45.7K7.1KCompositional✓✓
NExT-QA [43]34.1K5.0KTemporal & Causal✓✓
CLEVRER [47]21.0K-Spatial-temporal✓✗
Perception-Test [33]-11.5KGeneral✗✓
MVBench [19]-4.0KGeneral✗✓
VideoMME [9]-2.7KGeneral✗✓
VSIBench [44]-5.0Kspatial-temporal✗✓
OE-VQA
AGQA [14]25.0K2.0KCompositional✓✓
ANetQA [50]25.0K2.0KCompositional✓✓
EgoQA [13]7.8K-Ego-centric✓✗
Activitynet-QA [49]-8.0KGeneral✗✓
Video-ChatGPT [24]-3.0KGeneral✗✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.142, + 0.785, + 0.431, + 0.798 + ], + "angle": 0, + "content": "Table 3. Training and evaluation datasets statics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.222, + 0.832 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we present the experimental setup (Sec. 4.1) and results on various VideoQA benchmarks (Sec. 4.2). Extensive ablation studies have also been conducted to further examine the contributions of our approach in Sec. 4.3, and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.907, + 0.414 + ], + "angle": 0, + "content": "an evaluation on the quality of rationales generated by the distilled model is made in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.429, + 0.704, + 0.446 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Base model. We use LLaVA-NeXT-Video (7B) [54] (LNV for short) as base Video-LLM, which has shown remarkable performance on image-centric tasks, for example image question answering [51]. We present comparison on naive instruction-tuning with video question answering dataset or with additional CoT distillation. For CoT conversion and verification, we prompt LLaMA-3.1-8B with the manually-designed instruction and some in-context examples. Detailed prompts are provided in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.714 + ], + "angle": 0, + "content": "Instruction tuning. We utilize both multiple-choice and open-ended QA data, along with the generated CoTs, to fine-tune the base video question answering model, as summarised in Table 2. The resulting distilled model is named LLaVA-NeXT-Video-AoTD (LNV-AoTD for short). Additionally, as baseline, we also train another version of the model using only the basic QA data, which we refer to as LLaVA-NeXT-Video-Instruct (LNV-Instruct for short)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Evaluation benchmarks. We conduct extensive evaluations on Multiple-Choice Video QA (MC-VQA) and Open-Ended Video QA (OE-VQA). We report the top-1 accuracy for all MC benchmarks, which means the proportion of the output equal to the answer. We report a GPT-assessed Acc. and Score with the help of GPT-3.5-turbo-0613 for all OE benchmarks. For each question, GPT delivers a binary decision indicating whether the output is correct or incorrect, along with a similarity score reflecting the degree of alignment between the output and the correct answer. The term 'Acc.' refers to the percentage of correct outputs, while 'Score' represents the average similarity scores. For the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8528" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.872, + 0.299 + ], + "angle": 0, + "content": "
ModelMVBench (Acc.)VideoMME (Acc.)STAR (Acc.)NExT-QA (Acc.)Perception-Test (Acc.)VSIBench (Acc.)
Proprietary Models
Gemini 1.5 Pro [12]-75.0---45.4
GPT4-V [32]43.759.9----
Open-source Models
LLaMA-VID (7B) [21]41.925.9--44.6-
Video-LLaVA (7B) [22]41.039.9--44.3-
VideoChat2 (7B) [19]51.133.759.0*68.6*47.3-
VideoLLaMA2 (7B) [4]53.445.158.5*62.3*49.6-
LLaVA-NeXT-Video (7B) [54]46.5*41.0*52.4*61.6*47.5*19.7*
LLaVA-NeXT-Video-Instruct (7B)53.443.272.277.150.326.7
LLaVA-NeXT-Video-AoTD (7B)55.645.074.377.650.628.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.303, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 4. Comparison with Video-LLMs on MC-VQA benchmarks. LLaVA-NeXT-Video-AoTD outperforms all other baselines the and the version without CoT distillation. * means results reproduced by ourselves. Results without signs are retrieved from [4] and [44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.459 + ], + "angle": 0, + "content": "evaluation on AGQA and ANetQA, due to the large volume of test set, we test on a subset of samples. We evenly select the benchmark in-domain and out-of-domain for testing to ensure a comprehensive and reasonable evaluation of the model capability. Noted that though VSIBench has both MC and OE questions, it doesn't need GPT for score, so we classify it into MC benchmarks for convenience. Detailed statistics for evaluation benchmarks are shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.285, + 0.484 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.597 + ], + "angle": 0, + "content": "We divide the comparison into two parts: the first focuses on comparing the distilled model with other baselines, while the second examines the difference between the instruct version and the AoTD version. Note that, the latter part will be mainly compared and discussed, to demonstrate the model's improvement relative to its previous performance, as well as establishing the transferability of the method across models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.902 + ], + "angle": 0, + "content": "MC-VQA performance. As shown in Table 4, our LLaVA-NeXT-Video-AoTD achieves superior performance across all benchmarks. Several key observations can be made: (i) comparing to the base model, even a simple instruction-tuning on certain VideoQA datasets significantly enhances the model's question-answering performance. This improvement is notable, as the base model was primarily trained on static images and struggled with video understanding; (ii) our model, instruction-tuned with CoT distillation, demonstrates further performance enhancements across all benchmarks, particularly on the compositional VideoQA benchmark (STAR) and comprehensive benchmarks (VideoMME, MVBench). This suggests that our AoTD method effectively improves the model's ability to address complex problems and interpret spatial-temporal scenes; (iii) the distilled model consistently outperforms all other baselines across almost all benchmarks, even when compared to more powerful models. This finding shows that our method effectively bridges performance gaps created by varying model components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.908, + 0.596 + ], + "angle": 0, + "content": "OE-VQA performance. As shown in Table 5, LLaVA-NeXT-Video-AoTD outperforms the Instruct variant across all open-ended VideoQA benchmarks. Notably, it achieves a greater percentage increase compared to the MC-VQA benchmarks, suggesting that CoT distillation may be more effective for open-ended generation than for multiple-choice selection. While the distilled model scores higher than most models listed in the table, it does not surpass LLaVA-NeXT-Video on certain benchmarks. We conjecture this is due to the model's extensive training on images, that can also benefit the question answering without requiring complex reasonings, as also suggested by the findings in VideoLLaMA2 [4]. Additionally, the inherent challenges of evaluating open-ended VQA may influence the results. Assessments conducted by GPT can be biased or inaccurate, and the metrics we employ primarily indicate general trends rather than providing absolute accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.667, + 0.62 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Analysis on CoT filtering. To prove the effectiveness of our filtering mechanism, we trained an alternative model without CoT filtering while maintaining all other settings, i.e., using 36.3K verified CoTs for distillation. As shown in Table 9, the model's performance declines significantly on both the Multiple-Choice VQA and Open-Ended VQA benchmarks when the CoT filtering mechanism is not utilized. This confirms that employing large language models (LLMs) to filter CoTs is crucial for enhancing data quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis on model transferability. As AoTD is a distillation method that leverages Chain-of-Thoughts (CoTs), it can theoretically be applied to any Video-LLMs. To assess the transferability of our method, we conduct experiments on another very recent model, LLaVA-OneVision (7B) [3]. As shown in Table 9, our method also demonstrates significant improvements on the benchmarks, showing the transferability and robustness of the approach. Due to the rapid advancements in the computer vision field, evaluating all" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8529" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.842, + 0.327 + ], + "angle": 0, + "content": "
ModelANetQA (Acc./Score)AGQA (Acc./Score)Corr.Video-ChatGPT (Score)ActivityNet (Acc./Score)
Deta.Cont.Temp.Cons.
Proprietary Models
Gemini 1.5 Pro [12]-------56.7/-
GPT4-V [32]--4.093.884.373.944.0259.5/-
Open-Source Models
VideoLLaMA (7B) [53]--1.962.182.161.821.7912.4/1.1
Video-ChatGPT (7B) [24]--2.502.572.692.162.2035.2/2.7
LLaMA-VID (7B) [21]--2.963.003.532.462.5147.4/3.3
Video-LLaVA (7B) [22]--2.872.943.442.452.5145.3/3.3
VideoChat2 (7B) [19]--3.022.883.512.662.8149.1/3.3
VideoLLaMA2 (7B) [4]--3.093.093.682.633.2549.9/3.3
LLaVA-NeXT-Video (7B) [54]46.4/3.3*27.4/2.2*3.26*3.22*3.77*2.47*2.99*54.3/3.2*
LLaVA-NeXT-Video-Instruct (7B)47.1/3.159.3/3.42.962.813.352.422.8250.0/3.3
LLaVA-NeXT-Video-AoTD (7B)53.9/3.460.9/3.63.113.003.602.412.9153.2/3.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.332, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 5. Comparison with Video-LLMs on OE-VQA benchmarks. LLaVA-NeXT-Video-AoTD improves performance in all open-ended benchmarks compared with the Instruct version. * means results reproduced by ourselves. Results without signs are retrieved from [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.417 + ], + "angle": 0, + "content": "models and benchmarks is prohibitively infeasible. Thus, we focus on assessing some representative models against selected benchmarks to provide a representative evaluation." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.42, + 0.469, + 0.515 + ], + "angle": 0, + "content": "
ModelFilteringMVBench (Acc.)STAR (Acc.)AGQA (Acc. / Score)
LNV-AoTDX53.773.359.5/3.5
LNV-AoTD✓55.674.360.9/3.6
Onevision-Instruct-59.275.865.6/3.7
Onevision-AoTD✓60.576.665.7/3.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.52, + 0.466, + 0.533 + ], + "angle": 0, + "content": "Table 6. Ablation results of CoT filtering and transferability." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.54, + 0.32, + 0.554 + ], + "angle": 0, + "content": "4.4. Evaluation on Rationales" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.561, + 0.483, + 0.713 + ], + "angle": 0, + "content": "To verify whether the model has effectively learned multistep reasoning through CoTs distillation, we analyze the rationales generated by the model. Specifically, we extract and evaluate the temporal and spatial information embedded within these rationales. This approach extends beyond merely assessing the correctness of the final answer, which could be influenced by biases or other external factors. By examining the reasoning process in detail, it enables a more accurate understanding of the model's ability to perceive and reason about spatial and temporal relationships." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.716, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Evaluation protocols. We randomly select 200 samples from the STAR validation set and run inference on them using the suffix prompt, recording the generated rationales. From these rationales, we extract the predicted temporal windows and bounding boxes, comparing them to the ground truth. For the spatial part, we calculate the IoU between the predicted and ground truth bounding boxes. For the temporal part, we compute IoU and Recall, leveraging the frame-level annotations provided in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation results. Table 7 presents the evaluation results. For comparison, we also test UniVTG for temporal reasoning and OWL-ViT v2 for spatial reasoning. The re" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.369, + 0.905, + 0.458 + ], + "angle": 0, + "content": "
ModelTemporal GroundingSpatial Grounding
IoU (%)Recall (%)IoU (%)
UniVTG22.831.0-
OWL-ViT v2--64.7
LNV-InstructXXX
LNV-AoTD21.734.045.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.539, + 0.463, + 0.878, + 0.477 + ], + "angle": 0, + "content": "Table 7. Temporal and spatial abilities evaluation results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.579 + ], + "angle": 0, + "content": "sults show that LNV-Instruct struggles to generate valid rationales, even when using the suffix prompt. In contrast, LNV-AoTD demonstrates comparable performance to specialized models in both spatial and temporal reasoning, indicating that the model successfully acquired these abilities through the distillation process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.634, + 0.607 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.816 + ], + "angle": 0, + "content": "We present Agent-of-Thoughts Distillation (AoTD), that aims to distill multi-step reasoning and spatial-temporal understanding into a large video-language model (Video-LLM). Our method introduces an agent-based system that automates the generation of Chain-of-Thoughts (CoTs) from various VideoQA datasets, by breaking down complex questions into manageable sub-tasks that can be addressed by specialized vision models. Extensive experiments validate that the distilled model significantly enhances performance on both MC-VQA and OE-VQA benchmarks, underscoring the effectiveness of our approach. We believe AoTD represents a promising future direction for advancing the reasoning abilities in Video-LLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.842, + 0.673, + 0.859 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work is funded by National Key R&D Program of China (No.2022ZD0161400)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8530" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.185 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[3] Li Bo, Zhang Yuanhan, Guo Dong, Zhang Renrui, Li Feng, Zhang Hao, Zhang Kaichen, Li Yanwei, Liu Ziwei, and Li Chunyuan. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. arXiv preprint arXiv:2406.07476, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[5] Rohan Choudhury, Koichiro Niinuma, Kris M. Kitani, and Laszlo A. Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.626 + ], + "angle": 0, + "content": "[6] Guo Daya, Zhu Qihao, Yang Dejian, Dong Zhenda Xie, Kai, Zhang Wentao, Chen Guanting, Bi Xiao, Y. Wu, Y.K. Li, Luo Fuli, and Liang Yingfei, Xiongand Wenfeng. Deepseek-coder: When the large language model meets programming - the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[7] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[8] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In Proceedings of the International Conference on Machine Learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Minghe Gao, Shuang Chen, Liang Pang, Yuan Yao, Jisheng Dang, Wenqiao Zhang, Juncheng Li, Siliang Tang, Yueting Zhuang, and Tat-Seng Chua. Fact: Teaching mllms with faithful, concise and transferable rationales. In ACM Multimedia, 2024. 2, 5, 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[11] Timin Gao, Peixian Chen, Mengdan Zhang, Chaoyou Fu, Yunhang Shen, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Xing Sun, Liujuan Cao, et al. Cantor: Inspiring multimodal chain-of-thought of mllm. In ACM Multimedia, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[12] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[13] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[14] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[16] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[17] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent. In Advances in Neural Information Processing Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[18] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In Proceedings of the European Conference on Computer Vision, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[20] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[21] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[22] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8531" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Conference on Empirical Methods in Natural Language Processing, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[23] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In Proceedings of the International Conference on Computer Vision, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[25] Ahmad Mahmood, Ashmal Vayani, Muzammal Naseer, Salman Khan, and Fahad Khan. Vurf: A general-purpose reasoning and self-refinement framework for video understanding. arXiv preprint arXiv:2403.14743, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[26] Minderer Matthias, Gritsenko Alexey, Stone Austin, Neumann Maxim, Weissenborn Dirk, Dosovitskiy Alexey, Mahendran Aravindh, Arnab Anurag, Dehghani Mostafa, Shen Zhuoran, Wang Xiao, Zhai Xiaohua, Kipf Thomas, and Houlsby Neil. Simple open-vocabulary object detection with vision transformers. In Proceedings of the European Conference on Computer Vision, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[27] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[28] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. In Advances in Neural Information Processing Systems, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[29] Arindam Mitra, Luciano Del Corro, Shweti Mahajan, Andres Codas, Clarisse Simoes, Sahaj Agarwal, Xuxi Chen, Anastasia Razdaibiedina, Erik Jones, Kriti Aggarwal, et al. Orca 2: Teaching small language models how to reason. arXiv preprint arXiv:2311.11045, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[30] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[31] OpenAI. Gpt-3.5-turbo system card. https:// platform.openuai.com/docs/models/gpt-3- 5-turbo,2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[32] OpenAI. Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card, 2023.7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[33] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models. In Advances in Neural Information Processing Systems, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "ing transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[35] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[36] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. In Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[39] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In Proceedings of the European Conference on Computer Vision, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[41] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In Proceedings of the International Conference on Machine Learning, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[42] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Advances in Neural Information Processing Systems, 2021. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[43] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[44] Jihan Yang, Shusheng Yang, Anjali Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces. arXiv preprint arXiv:2412.14171, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[45] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In Proceedings of the International Conference on Machine Learning, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8532" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[47] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: collision events for video representation and reasoning. In Proceedings of the International Conference on Learning Representations, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[48] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In Advances in Neural Information Processing Systems, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[49] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.36, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[50] Zhou Yu, Lixiang Zheng, Zhou Zhao, Fei Wu, Jianping Fan, Kui Ren, and Jun Yu. Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.515, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[52] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.571, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[53] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[54] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.699, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[55] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.755, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[56] Orr Zohar, Xiaohan Wang, Yonatan Bitton, Idan Szpektor, and Serena Yeung-Levy. Video-star: Self-training enables video instruction tuning with any supervision. arXiv preprint arXiv:2407.06189, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8533" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_origin.pdf b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..69e0f2571378dae68e4fb9b8efe432d9fbff17d4 --- /dev/null +++ b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/9af6fd66-0efe-4a13-87b2-86d45a972105_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68e01efbabd120b33b4e0fe46530032714fc11ef5a6be1630336aa25529a5b69 +size 7037405 diff --git a/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/full.md b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..84eb145e09b47dd5ef5cf1ffcfd57a49e304ae20 --- /dev/null +++ b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/full.md @@ -0,0 +1,327 @@ +# Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation + +Yudi Shi $^{1,2}$ , Shangzhe Di $^{1,2}$ , Qirui Chen $^{1,2}$ , Weidi Xie $^{1,\dagger}$ + +$^{1}$ School of Artificial Intelligence, Shanghai Jiao Tong University, China + +$^{2}$ Coop. Medianet Innovation Center, Shanghai Jiao Tong University, China + +https://zhengrongz.github.io/AoTD/ + +# Abstract + +This paper tackles the problem of video question answering (VideoQA), a task that often requires multi-step reasoning and a profound understanding of spatial-temporal dynamics. While large video-language models perform well on benchmarks, they often lack explainability and spatial-temporal grounding. In this paper, we propose Agent-of-Thoughts Distillation (AoTD), a method that enhances models by incorporating automatically generated Chain-of-Thoughts (CoTs) into the instruction-tuning process. Specifically, we leverage an agent-based system to decompose complex questions into sub-tasks, and address them with specialized vision models, the intermediate results are then treated as reasoning chains. We also introduce a verification mechanism using a large language model (LLM) to ensure the reliability of generated CoTs. Extensive experiments demonstrate that AoTD improves the performance on multiple-choice and open-ended benchmarks. + +# 1. Introduction + +Video Question Answering (VideoQA) refers to a critical task that offers a natural interface for human-machine interaction through language [33, 42, 43, 49]. This synergy of visual content and language enhances the accessibility of AI systems for the general public, allowing users to query complex visual content with natural language. By encompassing tasks such as action recognition, object detection, and scene understanding, VideoQA serves as a comprehensive benchmark for evaluating AI's ability to interpret videos, addressing the fundamental questions of 'who', 'what', 'when', and 'where' that are crucial to understand daily life activities, pushing the boundaries of what AI systems can interpret from dynamic visual content. + +Recent literature has primarily explored two avenues in VideoQA. The first involves training large video lan + +guage models (Video-LLMs) through direct instruction-tuning, using videos paired solely with corresponding questions and answers [1, 22, 24, 53]. While these models excel on public benchmarks, they often lack explainability and struggle with spatial-temporal grounding. This limitation hinders their ability to provide clear reasoning, which is essential for real-world applications where transparency and interpretability are critical [29]. + +Conversely, an emerging approach utilizes agent-based systems that decompose complex questions into manageable sub-tasks, each addressed by specialized tools [15, 17, 37]. The results are then aggregated to form a coherent answer. Theoretically, such approach naturally offers greater interpretability, as the reasoning process is divided into explainable steps that can be independently assessed. However, our experiments indicate that current video understanding tools are not strong enough for building reliable agent-based systems. In addition, the high memory demands and time-consuming nature of these systems present significant challenges for their practical use. + +In this paper, we aim to leverage the advantage of both research lines, enhancing Video-LLM by integrating Chain-of-Thoughts (CoTs) into instruction-tuning, with the CoTs being constructed from the outputs of specialized agent models, capturing the step-by-step reasoning procedure, as illustrated in Figure 1. + +In specific, we start by systematically evaluating the off-the-shelf models tailored for atomic video understanding tasks, such as action recognition [39, 41] and language grounding [23], using well-annotated datasets. This comprehensive evaluation allows us to pinpoint the most effective tools for each sub-task, thus laying a robust foundation for constructing reliable chains. Moreover, this process also provides a critical assessment of the broader capabilities of visual models across general and complex scenes, offering valuable insights for future research within the community. + +In addition, we introduce a verification mechanism using a large language model (LLM), designed to assess if the generated CoTs adhere to a clear, step-by-step reason- + +![](images/17bb1544992c14d17c78007bd09f29b16c38ac5375d015f13f76caae5c65d36f.jpg) +Figure 1. Our method, AoTD, distills multi-step reasoning and spatial-temporal understanding into a single generative video-language model. When addressing complex VideoQA tasks, the model trained with AoTD (as shown in (b)) enables to generate a step-by-step reasoning to get the correct answer. In contrast, previous models trained solely on question-answer pairs (as in (a)) generate only a final answer, often without intermediate reasoning, which can lead to incorrect conclusions. + +ing process and incorporate essential information for answering the queries effectively. This mechanism filters out low-quality or logically inconsistent reasoning paths. The remaining CoTs that pass this verification are then distilled into large generative video-language models, significantly enhancing both their performance and interpretability, ultimately leading to the development of more robust, accurate, and interpretable VideoQA systems. + +In summary, our contributions are three-fold: (i) we propose a novel approach for enhancing Video-LLMs by distilling high-quality CoTs into their instruction-tuning process. These CoTs capture step-by-step reasoning paths, improving both the model's performance and its interpretability; (ii) to automatically construct the CoTs for any dataset, we employ an agent-based system to decompose complex VideoQA questions into simpler sub-tasks, leveraging off-the-shelf vision models to handle each sub-task. The intermediate outputs from these models can therefore be collected as CoTs for addressing the corresponding visual question; (iii) through extensive experiments, we demonstrate that our distilled model outperforms existing methods across both multiple-choice and open-ended VideoQA benchmarks, enabling to deliver not only accurate answers but also comprehensive reasoning explanations. + +# 2. Related Work + +Video-language models (Video-LLMs). Recent works such as VideoLLaMA2 [4], LLaVA-NeXT-Video [54] and + +VideoChat2 [19], with their excellent architecture design and reasonable instruction-tuning data collection, have achieved a new level of zero-shot results in VideoQA task. However, current end-to-end models still lack interpretability for questions, as well as the ability to think and visually process complex problems in multiple steps, which is an important part for embodied learning and autonomous driving. + +Visual Programming and Agents. With the progress of LLMs, some recent works [5, 15, 37, 45] begin to try to use LLM as planner to solve the complex reasoning task in real scenarios. They attempt to decompose the question into some easier sub-questions, and use different specialist models as agents to solve these sub-questions, and finally gather them to get the answer of the raw question. These models demonstrate a strong ability to obtain trustworthy answers based on the intermediate evidence they get, but they lag far behind the end-to-end model in terms of inference speed. + +Visual Chain-of-Thoughts (CoTs). The potential of Chain-of-Thought (CoT) reasoning [40, 46] extends from NLP to the visual domain, highlighting a growing interest in applying this approach across various fields. Numerous studies have incorporated CoTs into visual understanding tasks [11, 30, 36, 55], utilizing powerful Multi-Modal Large Language Models (MLLMs) for generating CoTs or adopting tool-based architectures for sequential problem-solving. + +Recent innovations, for example, Visual Program Distillation (VPD) [16] and Fact [10] attempt to address these + +![](images/cc83821021dd8b9d6f71f7c08b74b07c06a7143507b7df54fa6523cd2d235fdb.jpg) +Figure 2. Overview on Agent-of-Thoughts Distillation (AoTD). Step 1: Selecting best-performing agents for each sub-task to construct an agent-based system. Step 2: Decomposing question into executable program and leveraging chosen models to solve it sequentially to generate execution trace. Step 3: The execution trace is converted and filtered by LLM to produce high quality natural language CoTs. Step 4: Distilling CoTs into Video-LLM with two forms of prompt, allowing it to achieve a balance between concise answers and comprehensive rationales. The final model is Video-LLM-AoTD. + +issues by maintaining the accuracy and diversity of CoTs, while leveraging MLLMs to generate them directly. These approaches decompose complex tasks into code programs, call upon expert models to handle sub-tasks, and utilize the resulting CoTs as training data to fine-tune visual-language models. This process significantly improves the models' ability to generate detailed rationales. Despite the progress in image understanding, there remains a notable oversight in video domains, where reasoning chains can be particularly effective due to the complex spatial-temporal dynamics of video understanding tasks. This is the focus of our paper. + +Concurrent Work. In the recent literature, we notice two work that share similar idea with ours, specifically, Video-STaR [56] construct CoTs using videos and existing labels for instruction-tuning, yet they do not develop an agent-based system. Meanwhile, MotionEpic [8] introduces a Video-of-Thought reasoning framework that integrates video spatial-temporal scene graphs, marking a significant stride towards more nuanced video reasoning. + +# 3. Agent-of-Thoughts Distillation + +In this paper, we propose a novel approach, termed Agent-of-Thoughts Distillation (AoTD), to enhance the Video-LLMs by training them with multi-step Chain-of-Thoughts. + +Specifically, we start by developing an agent-based video understanding system, to generate multi-step reasoning chains that address complex video questions. These reasoning chains are then distilled into one Video-LLM through instruction-tuning. By combining the strengths of agent-based systems and large generative models, our proposed AoTD enables to build more reliable and interpretable VideoQA systems. Figure 2 illustrates the entire process of our method. + +# 3.1. Problem Formulation + +Given a video clip with $t$ frames, $\mathcal{V} = \{x_1, \ldots, x_t\}$ , and a set of $n$ questions $\mathcal{Q} = \{q_1, q_2, \ldots, q_n\}$ , our goal is to train a Video-LLM capable of producing both concise answers and comprehensive rationales. Depending on the suffix prompt $p_s$ , the model can generate different types of outputs. The process can be formulated as: + +$$ +\left\{a _ {i}, \mathcal {S} _ {i} \right\} = \Phi (\mathcal {V}, q _ {i}, p _ {s}), \mathcal {S} _ {i} = \{\emptyset \} \text {o r} \left\{s _ {i, 1}, \dots , s _ {i, k} \right\} +$$ + +where $q_{i}$ denotes the $i$ -th question, $a_{i}$ is the answer in freeform text, and $S_{i}$ represents the rationale, consisting of the reasoning process. If the prompt specifies to only generate the answer, $S_{i} = \{\emptyset\}$ . Otherwise, if the prompt requires the generation of rationales, $S_{i} = \{s_{i,1}, \ldots, s_{i,k}\}$ , where each $s_{i,j}$ corresponds to a reasoning step. + +Video input +Question: Which object did the person put down before they held the book? Possible answers: (A) The broom. (B) The paper. (C) The door. (D) The cup. +![](images/68ec4f03f282cec6fecd90b25251552a7502eceaf5be1c6ac64f55f3fadc66db.jpg) +def execute_command(video, query, possible_answers): hold Clip $=$ Filter Frames_with_act(video, 'person held the boo before_hold Clip $=$ trim(video,end=hold clip.sta put clip $=$ Filter Frames_with_act(before_hold_cl person put down somethi person clip $=$ Find(place,'person') put_objs $=$ Query_Objects(person clip, 'object put down by the perso info $=$ {'object put down by the person': put_obj answer $=$ select_answer(query,info,possible的答案 return answer +Ground Truth Hold a book: Frame[6.64,7.45,...,16.32] + +frame 2.88 +Ground Truth Put something: Frame[2.88,3.41,4.78,6.12,6.64] +Ground Truth Person:[128,64,244,371] in frame 2.88 +![](images/33473034039d4be3540613bf8c2f0129853df88be5023d9374259f6dcc822f4a.jpg) +Object Detection Eval Person:[119,0,241,408] in frame 2.88 + +frame 6.64 +Program Output The broom. +Ground Truth Answer The broom. +Ground Truth Put a broom somewhere +![](images/28053488798311b6240ae2401679e65af7099e28bef2b290410a2c33a74d9364.jpg) +Evaluation IoU 65.8% + +frame 7 +Figure 3. Program execution process in an agent-based system. We uniformly sample 32 frames from the video, and to ensure scale consistency, the frame ids of key frames are normalized into these 32 frames. The blue boxes represent the program execution steps, the red boxes denote the ground truth for each step. The combination of red and yellow boxes represents one example process of evaluating Object detection model candidates. +![](images/f5814326e3b4cbfed30c218ec9301e3a608aec00f0c573f60075b8813f43a901.jpg) +Filter_frames_with actu hold Clip: Frame 7 to 15 + +![](images/064d0c33aa2d1448fea26394963b6b091ec7e32e3ac5c2b5551a0df957bb733b.jpg) +frame 15 + +![](images/b2bef87da79297f89f8430cee1e8a8925bfcdc034f9d33183a7c663f407b30e9.jpg) +Ground Truth +Frame[2.88,3.41,4.78, 6.12,6.64] +trim +before_hold Clip: +Frame 0 to 7 +frame 0 + +![](images/d593a95de6c5bae5dc1613aebc8019086b2472167d1aae2446cbb690be4fa80f.jpg) +frame 7 + +![](images/f7ee3066790d72441e1c60cd4265410b280fe81a97b442e8115d29c956db4c94.jpg) +Filter Frames with act put Clip: Frame 2 to 7 +frame 2 + +![](images/7b8e75d154e539de4bde52c453f260139df2b6283158532a65bdacdbcbf5ff0d.jpg) +frame 7 + +![](images/fa4388d34ac4a7bb7dce5958c6f8dbdd68dbc98f96024c08f1c6f7fb1a5437f6.jpg) +Find person Clip:[120,0,239, 413] in frame 2 ... +frame 2 + +![](images/23a55f3e19616eaa13cd27195ab0d75a545a2ceee26f47814bdef5a57f5b4c2a.jpg) +frame 7 + +![](images/745f32b8e47f4bd4e024ad8eb436aed743fa1fe27d1b25abdf9c5a793e137487.jpg) +Query_Objects +put_objects: broom +frame 2 + +![](images/d20ccc585c3a81054b4e64d2582800c4a1aa792b891431597e221db85c6225ab.jpg) +frame 7 + +Discussion. Unlike existing models that are instruction-tuned on VideoQA datasets using simple question-answer pairs, which bypass the intermediate thought process, our approach emphasizes the importance of training with Chain-of-Thoughts (CoTs). In the following section, we outline the process for generating high-quality CoTs from existing VideoQA datasets. + +# 3.2. CoT Construction with Agent-based System + +Recent work, such as STAR [42], has introduced the idea of employing executable symbolic programs, to directly decompose questions into sub-tasks. When combined with scene graphs that contain comprehensive video information from key frames—such as object locations, interactions, and actions—these programs facilitate the generation of concise Chain-of-Thoughts (CoTs) through direct execution of symbolic operations. However, datasets of this nature are limited in scale, we therefore propose to first build an agent-based system, capable of breaking down complex questions into simpler sub-tasks, and the intermediate outputs from this system can then be employed to construct CoTs for any existing VideoQA dataset. + +Agent-based VideoQA. Assuming we are given a video input $(\mathcal{V})$ , questions $(\mathcal{Q})$ , and a set of visual models $(\mathcal{M} = \{\phi_{\mathrm{act}}, \phi_{\mathrm{det}}, \dots, \phi_{\mathrm{qa}}\})$ , an LLM-based agent core $(\pi(\cdot))$ processes the question along with the documentation of the visual models $(\mathcal{T})$ , which includes variables and function- + +
Sub-taskModel nameMetricNumber (%)
question decompositionCodeQwen1.5-Chat (7B) [2]52.7
GPT-3.5-Turbo [31]Acc73.1
DeepSeek-Coder-Instruct (6.7B) [6]85.7
object detectionOWL-ViT v1 [26]47.3
GLIP [20]IoU58.9
OWL-ViT v2 [28]63.0
temporal groundingLITA (13B) [18]11.7 / 20.2
TimeChat (7B) [35]IoU / Recall13.9 / 23.1
UniVTG [23]24.7 / 35.3
action recognitionInternVideo2 (1B) [39]7.6
Open-VCLIP [41]Top1-Acc8.9
LLaVA-NeXT-Video-DPO (7B) [54]18.2
question answeringLLaMA-VID (7B) [21]43.5
SeViLA [48]Acc46.5
LLaVA-NeXT-Video-DPO (7B) [54]53.4
+ +Table 1. Sub-tasks definition and evaluation results. We choose 3 model candidates for each sub-task and evaluate them in STAR training set with the corresponding metrics. Models with best performance are placed at the bottom of each column. + +alities. The agent subsequently decomposes the question into sub-tasks formatted as Python code, and resolves them by invoking the appropriate visual models through function calls. It is important to note that the visual models can be arranged in various orders depending on the specific question, ensuring flexibility in problem-solving. + +Specifically, as illustrated by the example in Figure 3, the + +question is first decomposed into a series of sub-tasks, including temporal grounding, object detection, and question answering. The corresponding specialized models are then executed sequentially to address these sub-tasks, ultimately yielding the final answer $y_{i}$ : + +$$ +\begin{array}{l} \left\{\phi_ {\text {g r o u n d}}, \phi_ {\det }, \phi_ {\mathrm {q a}} \right\} := \pi \left(q _ {i}, \mathcal {T}\right), \\ y _ {i} = \phi_ {\text {g r o u n d}} (\mathcal {V}) \rightarrow \phi_ {\det} (\mathcal {V}) \rightarrow \phi_ {\mathrm {q a}} (\mathcal {V}) \\ \end{array} +$$ + +CoT Construction. To ensure the correctness of outputs at all the intermediate steps, we leverage the training set from STAR for hyperparameter tuning, enabling us to identify the most effective model for each sub-task within the agent-based system. By following the provided programs, we evaluate the performance of the corresponding vision models on tasks such as object detection and action recognition. Given the availability of complete reasoning chains, we independently assess each sub-task using ground truth data for all preceding steps. + +As shown in Table 1, we present the evaluation results for the various sub-tasks. Specifically, for question decomposition, we compare several code LLMs, with DeepSeekCoder-Instruct achieving the highest performance, outperforming even GPT-3.5-Turbo. In object detection, OWLViT v2 records the highest Intersection over Union (IoU) score, showcasing its superior open-vocabulary detection capability. The results for temporal grounding indicate that while UniVTG leads in performance, there remains a need for further advancements in this area. In action recognition, our evaluations show that generative models outperformed discriminative models, likely due to the fine-grained action list provided by the STAR dataset. However, the performance of both model types reveals significant room for improvement. Finally, in the one-hop question answering sub-task, all models perform admirably, with LLaVA-NeXT-Video-DPO standing out as a top performer, consistent with its strong results on other benchmarks. + +With these high-performing models, we implement the agent-based approach on VideoQA datasets that consist solely of QA pairs. During the execution of the programs, we record all intermediate outputs to construct the CoTs. Since the outputs from these vision models vary in format—such as bounding boxes and free-form text—we employ another LLM to translate the execution trace into natural language for better use in the distillation process. Detailed examples are provided in Appendix C. + +# 3.3. CoT Verification + +To refine the quality of reasoning chains for VideoQA samples, we implement a two-step verification: (i) we filter execution traces to retain only those, where the program can reach correct output. For multiple-choice datasets, the output must match the correct answer exactly, while for open-ended datasets, we prompt the LLM to verify correctness, + +
DatasetDescription# Labels# CoTs
AGQACompositional25.0K5.4K
ANetQACompositional25.0K3.6K
STARCompositional45.7K11.2K
NExT-QATemporal & Causal34.1K12.1K
CLEVRERSpatial & Temporal21.0K-
EgoQAEgo-centric7.8K-
Total158.6K32.3K
+ +Table 2. Dataset statistics. The column "#Labels" indicates the number of VideoQA pairs, which include the video, query, possible answers (multiple-choice), and the correct answer. "#CoTs" refers to the number of CoTs generated using our agent-based system for each dataset. + +accounting for format differences; (ii) we prompt the LLM to evaluate the logical coherence and usefulness of the reasoning chains in solving the problem. The model assesses whether the CoTs follow a clear, step-by-step reasoning process and provides a binary evaluation ('Yes' or 'No') to indicate their quality (detailed prompts are included in Appendix D). This two-step approach ensures that only high-quality CoTs are retained for further distillation. + +In Table 2, we provide the statistics for the remaining generated CoTs for different datasets. We primarily select compositional QA datasets, as these require the model to process spatial-temporal information from different events comprehensively. + +# 3.4. Step-by-step Distillation + +In this section, we describe the process of distilling the generated CoTs into a Video-LLM. This distillation enhances the model's ability for spatial-temporal video understanding and multi-step reasoning, thereby improving its performance on complex VideoQA tasks. + +In specific, using the generated CoTs, we can build the dataset $D = \{(\mathcal{V}_j,q_j,\hat{y}_j,c_j,p_s)\}_{j = 1}^N$ , where $N$ is the total number of samples in the distilling dataset, $\mathcal{V}_j$ is the video input, $q_{j}$ is the question, $\hat{y}_j$ is the ground-truth answer, $c_{j}$ is the generated CoT, $p_{s}$ is the task-specific suffix prompt, to distinguish different tasks, for example, for multiple-choice VQA, the prompt can be: "Answer with the option's letter from the given choices directly and only give the best option", and for open-ended VQA, the prompt can be: "Answer in one word or phrase". Please refer to detailed prompts in Appendix D. + +At distillation stage, we minimize the cross-entropy loss of predicting both the answer and the CoTs, we replace the suffix prompt $p_{s}$ with "Explain the rationale to answer the question", to indicate whether we want a question answer or a rationale to explain the thinking steps. Following [10] + +![](images/cddd742c8d91e51cb002c8205855594521b7d735f28b8dd74af340e6e7ffb96b.jpg) +Figure 4. Visualization of rationales. LLaVA-NeXT-Video-AoTD can output rationales containing both spatial-temporal grounding of key information and step-by-step thinking process to solve the question. + +![](images/56007bbdf03faa4e89fc8b0794b261f650a83e927808135b2cc00ea59b927197.jpg) + +and [16], our optimization objective is: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\text {l a b e l}} + \lambda \mathcal {L} _ {\text {r a t i o n a l e}} \\ = \sum_ {j = 1} ^ {N} \ell (\Phi (\mathcal {V} _ {j}, q _ {j}, p _ {s}), \hat {y} _ {j}) + \lambda \ell (\Phi (\mathcal {V} _ {j}, q _ {j}, p _ {s}), c _ {j}) \\ \end{array} +$$ + +Here, we set $\lambda$ to 1 to ensure the importance of answer and rationale are equally considered. Notice that, not all the QA pairs can generate qualified CoT. In that case, the $\mathcal{L}_{\mathrm{rationale}}$ will be set to 0. + +
DatasetSizeTypeTrainEval
traineval
MC-VQA
STAR [42]45.7K7.1KCompositional✓✓
NExT-QA [43]34.1K5.0KTemporal & Causal✓✓
CLEVRER [47]21.0K-Spatial-temporal✓✗
Perception-Test [33]-11.5KGeneral✗✓
MVBench [19]-4.0KGeneral✗✓
VideoMME [9]-2.7KGeneral✗✓
VSIBench [44]-5.0Kspatial-temporal✗✓
OE-VQA
AGQA [14]25.0K2.0KCompositional✓✓
ANetQA [50]25.0K2.0KCompositional✓✓
EgoQA [13]7.8K-Ego-centric✓✗
Activitynet-QA [49]-8.0KGeneral✗✓
Video-ChatGPT [24]-3.0KGeneral✗✓
+ +Table 3. Training and evaluation datasets statics. + +# 4. Experiments + +In this section, we present the experimental setup (Sec. 4.1) and results on various VideoQA benchmarks (Sec. 4.2). Extensive ablation studies have also been conducted to further examine the contributions of our approach in Sec. 4.3, and + +an evaluation on the quality of rationales generated by the distilled model is made in Sec. 4.4. + +# 4.1. Experimental Setup + +Base model. We use LLaVA-NeXT-Video (7B) [54] (LNV for short) as base Video-LLM, which has shown remarkable performance on image-centric tasks, for example image question answering [51]. We present comparison on naive instruction-tuning with video question answering dataset or with additional CoT distillation. For CoT conversion and verification, we prompt LLaMA-3.1-8B with the manually-designed instruction and some in-context examples. Detailed prompts are provided in Appendix D. + +Instruction tuning. We utilize both multiple-choice and open-ended QA data, along with the generated CoTs, to fine-tune the base video question answering model, as summarised in Table 2. The resulting distilled model is named LLaVA-NeXT-Video-AoTD (LNV-AoTD for short). Additionally, as baseline, we also train another version of the model using only the basic QA data, which we refer to as LLaVA-NeXT-Video-Instruct (LNV-Instruct for short). + +Evaluation benchmarks. We conduct extensive evaluations on Multiple-Choice Video QA (MC-VQA) and Open-Ended Video QA (OE-VQA). We report the top-1 accuracy for all MC benchmarks, which means the proportion of the output equal to the answer. We report a GPT-assessed Acc. and Score with the help of GPT-3.5-turbo-0613 for all OE benchmarks. For each question, GPT delivers a binary decision indicating whether the output is correct or incorrect, along with a similarity score reflecting the degree of alignment between the output and the correct answer. The term 'Acc.' refers to the percentage of correct outputs, while 'Score' represents the average similarity scores. For the + +
ModelMVBench (Acc.)VideoMME (Acc.)STAR (Acc.)NExT-QA (Acc.)Perception-Test (Acc.)VSIBench (Acc.)
Proprietary Models
Gemini 1.5 Pro [12]-75.0---45.4
GPT4-V [32]43.759.9----
Open-source Models
LLaMA-VID (7B) [21]41.925.9--44.6-
Video-LLaVA (7B) [22]41.039.9--44.3-
VideoChat2 (7B) [19]51.133.759.0*68.6*47.3-
VideoLLaMA2 (7B) [4]53.445.158.5*62.3*49.6-
LLaVA-NeXT-Video (7B) [54]46.5*41.0*52.4*61.6*47.5*19.7*
LLaVA-NeXT-Video-Instruct (7B)53.443.272.277.150.326.7
LLaVA-NeXT-Video-AoTD (7B)55.645.074.377.650.628.8
+ +Table 4. Comparison with Video-LLMs on MC-VQA benchmarks. LLaVA-NeXT-Video-AoTD outperforms all other baselines the and the version without CoT distillation. * means results reproduced by ourselves. Results without signs are retrieved from [4] and [44]. + +evaluation on AGQA and ANetQA, due to the large volume of test set, we test on a subset of samples. We evenly select the benchmark in-domain and out-of-domain for testing to ensure a comprehensive and reasonable evaluation of the model capability. Noted that though VSIBench has both MC and OE questions, it doesn't need GPT for score, so we classify it into MC benchmarks for convenience. Detailed statistics for evaluation benchmarks are shown in Table 3. + +# 4.2. Quantitative Results + +We divide the comparison into two parts: the first focuses on comparing the distilled model with other baselines, while the second examines the difference between the instruct version and the AoTD version. Note that, the latter part will be mainly compared and discussed, to demonstrate the model's improvement relative to its previous performance, as well as establishing the transferability of the method across models. + +MC-VQA performance. As shown in Table 4, our LLaVA-NeXT-Video-AoTD achieves superior performance across all benchmarks. Several key observations can be made: (i) comparing to the base model, even a simple instruction-tuning on certain VideoQA datasets significantly enhances the model's question-answering performance. This improvement is notable, as the base model was primarily trained on static images and struggled with video understanding; (ii) our model, instruction-tuned with CoT distillation, demonstrates further performance enhancements across all benchmarks, particularly on the compositional VideoQA benchmark (STAR) and comprehensive benchmarks (VideoMME, MVBench). This suggests that our AoTD method effectively improves the model's ability to address complex problems and interpret spatial-temporal scenes; (iii) the distilled model consistently outperforms all other baselines across almost all benchmarks, even when compared to more powerful models. This finding shows that our method effectively bridges performance gaps created by varying model components. + +OE-VQA performance. As shown in Table 5, LLaVA-NeXT-Video-AoTD outperforms the Instruct variant across all open-ended VideoQA benchmarks. Notably, it achieves a greater percentage increase compared to the MC-VQA benchmarks, suggesting that CoT distillation may be more effective for open-ended generation than for multiple-choice selection. While the distilled model scores higher than most models listed in the table, it does not surpass LLaVA-NeXT-Video on certain benchmarks. We conjecture this is due to the model's extensive training on images, that can also benefit the question answering without requiring complex reasonings, as also suggested by the findings in VideoLLaMA2 [4]. Additionally, the inherent challenges of evaluating open-ended VQA may influence the results. Assessments conducted by GPT can be biased or inaccurate, and the metrics we employ primarily indicate general trends rather than providing absolute accuracy. + +# 4.3. Ablation Study + +Analysis on CoT filtering. To prove the effectiveness of our filtering mechanism, we trained an alternative model without CoT filtering while maintaining all other settings, i.e., using 36.3K verified CoTs for distillation. As shown in Table 9, the model's performance declines significantly on both the Multiple-Choice VQA and Open-Ended VQA benchmarks when the CoT filtering mechanism is not utilized. This confirms that employing large language models (LLMs) to filter CoTs is crucial for enhancing data quality. + +Analysis on model transferability. As AoTD is a distillation method that leverages Chain-of-Thoughts (CoTs), it can theoretically be applied to any Video-LLMs. To assess the transferability of our method, we conduct experiments on another very recent model, LLaVA-OneVision (7B) [3]. As shown in Table 9, our method also demonstrates significant improvements on the benchmarks, showing the transferability and robustness of the approach. Due to the rapid advancements in the computer vision field, evaluating all + +
ModelANetQA (Acc./Score)AGQA (Acc./Score)Corr.Video-ChatGPT (Score)ActivityNet (Acc./Score)
Deta.Cont.Temp.Cons.
Proprietary Models
Gemini 1.5 Pro [12]-------56.7/-
GPT4-V [32]--4.093.884.373.944.0259.5/-
Open-Source Models
VideoLLaMA (7B) [53]--1.962.182.161.821.7912.4/1.1
Video-ChatGPT (7B) [24]--2.502.572.692.162.2035.2/2.7
LLaMA-VID (7B) [21]--2.963.003.532.462.5147.4/3.3
Video-LLaVA (7B) [22]--2.872.943.442.452.5145.3/3.3
VideoChat2 (7B) [19]--3.022.883.512.662.8149.1/3.3
VideoLLaMA2 (7B) [4]--3.093.093.682.633.2549.9/3.3
LLaVA-NeXT-Video (7B) [54]46.4/3.3*27.4/2.2*3.26*3.22*3.77*2.47*2.99*54.3/3.2*
LLaVA-NeXT-Video-Instruct (7B)47.1/3.159.3/3.42.962.813.352.422.8250.0/3.3
LLaVA-NeXT-Video-AoTD (7B)53.9/3.460.9/3.63.113.003.602.412.9153.2/3.4
+ +models and benchmarks is prohibitively infeasible. Thus, we focus on assessing some representative models against selected benchmarks to provide a representative evaluation. + +Table 5. Comparison with Video-LLMs on OE-VQA benchmarks. LLaVA-NeXT-Video-AoTD improves performance in all open-ended benchmarks compared with the Instruct version. * means results reproduced by ourselves. Results without signs are retrieved from [4]. + +
ModelFilteringMVBench (Acc.)STAR (Acc.)AGQA (Acc. / Score)
LNV-AoTDX53.773.359.5/3.5
LNV-AoTD✓55.674.360.9/3.6
Onevision-Instruct-59.275.865.6/3.7
Onevision-AoTD✓60.576.665.7/3.7
+ +# 4.4. Evaluation on Rationales + +To verify whether the model has effectively learned multistep reasoning through CoTs distillation, we analyze the rationales generated by the model. Specifically, we extract and evaluate the temporal and spatial information embedded within these rationales. This approach extends beyond merely assessing the correctness of the final answer, which could be influenced by biases or other external factors. By examining the reasoning process in detail, it enables a more accurate understanding of the model's ability to perceive and reason about spatial and temporal relationships. + +Evaluation protocols. We randomly select 200 samples from the STAR validation set and run inference on them using the suffix prompt, recording the generated rationales. From these rationales, we extract the predicted temporal windows and bounding boxes, comparing them to the ground truth. For the spatial part, we calculate the IoU between the predicted and ground truth bounding boxes. For the temporal part, we compute IoU and Recall, leveraging the frame-level annotations provided in the dataset. + +Evaluation results. Table 7 presents the evaluation results. For comparison, we also test UniVTG for temporal reasoning and OWL-ViT v2 for spatial reasoning. The re + +Table 6. Ablation results of CoT filtering and transferability. + +
ModelTemporal GroundingSpatial Grounding
IoU (%)Recall (%)IoU (%)
UniVTG22.831.0-
OWL-ViT v2--64.7
LNV-InstructXXX
LNV-AoTD21.734.045.2
+ +Table 7. Temporal and spatial abilities evaluation results. + +sults show that LNV-Instruct struggles to generate valid rationales, even when using the suffix prompt. In contrast, LNV-AoTD demonstrates comparable performance to specialized models in both spatial and temporal reasoning, indicating that the model successfully acquired these abilities through the distillation process. + +# 5. Conclusion + +We present Agent-of-Thoughts Distillation (AoTD), that aims to distill multi-step reasoning and spatial-temporal understanding into a large video-language model (Video-LLM). Our method introduces an agent-based system that automates the generation of Chain-of-Thoughts (CoTs) from various VideoQA datasets, by breaking down complex questions into manageable sub-tasks that can be addressed by specialized vision models. Extensive experiments validate that the distilled model significantly enhances performance on both MC-VQA and OE-VQA benchmarks, underscoring the effectiveness of our approach. We believe AoTD represents a promising future direction for advancing the reasoning abilities in Video-LLMs. + +# Acknowledgments + +This work is funded by National Key R&D Program of China (No.2022ZD0161400). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1 +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 4 +[3] Li Bo, Zhang Yuanhan, Guo Dong, Zhang Renrui, Li Feng, Zhang Hao, Zhang Kaichen, Li Yanwei, Liu Ziwei, and Li Chunyuan. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7 +[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. arXiv preprint arXiv:2406.07476, 2024. 2, 7, 8 +[5] Rohan Choudhury, Koichiro Niinuma, Kris M. Kitani, and Laszlo A. Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 2 +[6] Guo Daya, Zhu Qihao, Yang Dejian, Dong Zhenda Xie, Kai, Zhang Wentao, Chen Guanting, Bi Xiao, Y. Wu, Y.K. Li, Luo Fuli, and Liang Yingfei, Xiongand Wenfeng. Deepseek-coder: When the large language model meets programming - the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 4 +[7] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 1 +[8] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In Proceedings of the International Conference on Machine Learning, 2024. 3 +[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 6 +[10] Minghe Gao, Shuang Chen, Liang Pang, Yuan Yao, Jisheng Dang, Wenqiao Zhang, Juncheng Li, Siliang Tang, Yueting Zhuang, and Tat-Seng Chua. Fact: Teaching mllms with faithful, concise and transferable rationales. In ACM Multimedia, 2024. 2, 5, 1 + +[11] Timin Gao, Peixian Chen, Mengdan Zhang, Chaoyou Fu, Yunhang Shen, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Xing Sun, Liujuan Cao, et al. Cantor: Inspiring multimodal chain-of-thought of mllm. In ACM Multimedia, 2024. 2 +[12] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 8 +[13] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6 +[14] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 6 +[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2 +[16] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 1 +[17] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent. In Advances in Neural Information Processing Systems, 2024. 1 +[18] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In Proceedings of the European Conference on Computer Vision, 2024. 4 +[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7, 8 +[20] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 4 +[21] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 4, 7, 8 +[22] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the + +Conference on Empirical Methods in Natural Language Processing, 2024. 1, 7, 8 +[23] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In Proceedings of the International Conference on Computer Vision, 2023. 1, 4 +[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 1, 6, 8 +[25] Ahmad Mahmood, Ashmal Vayani, Muzammal Naseer, Salman Khan, and Fahad Khan. Vurf: A general-purpose reasoning and self-refinement framework for video understanding. arXiv preprint arXiv:2403.14743, 2024. 8 +[26] Minderer Matthias, Gritsenko Alexey, Stone Austin, Neumann Maxim, Weissenborn Dirk, Dosovitskiy Alexey, Mahendran Aravindh, Arnab Anurag, Dehghani Mostafa, Shen Zhuoran, Wang Xiao, Zhai Xiaohua, Kipf Thomas, and Houlsby Neil. Simple open-vocabulary object detection with vision transformers. In Proceedings of the European Conference on Computer Vision, 2022. 4 +[27] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 8 +[28] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. In Advances in Neural Information Processing Systems, 2024. 4 +[29] Arindam Mitra, Luciano Del Corro, Shweti Mahajan, Andres Codas, Clarisse Simoes, Sahaj Agarwal, Xuxi Chen, Anastasia Razdaibiedina, Erik Jones, Kriti Aggarwal, et al. Orca 2: Teaching small language models how to reason. arXiv preprint arXiv:2311.11045, 2023. 1 +[30] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[31] OpenAI. Gpt-3.5-turbo system card. https:// platform.openuai.com/docs/models/gpt-3- 5-turbo,2023.4 +[32] OpenAI. Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card, 2023.7,8 +[33] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models. In Advances in Neural Information Processing Systems, 2023. 1, 6 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- + +ing transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 8 +[35] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 4 +[36] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. In Advances in Neural Information Processing Systems, 2024. 2 +[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2, 3 +[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8 +[39] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In Proceedings of the European Conference on Computer Vision, 2024. 1, 4 +[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, 2022. 2 +[41] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In Proceedings of the International Conference on Machine Learning, 2023. 1, 4 +[42] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Advances in Neural Information Processing Systems, 2021. 1, 4, 6 +[43] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 6, 2 +[44] Jihan Yang, Shusheng Yang, Anjali Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces. arXiv preprint arXiv:2412.14171, 2024. 6, 7 +[45] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In Proceedings of the International Conference on Machine Learning, 2024. 2 + +[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, 2024. 2 +[47] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: collision events for video representation and reasoning. In Proceedings of the International Conference on Learning Representations, 2020. 6 +[48] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In Advances in Neural Information Processing Systems, 2023. 4 +[49] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 1, 6 +[50] Zhou Yu, Lixiang Zheng, Zhou Zhao, Fei Wu, Jianping Fan, Kui Ren, and Jun Yu. Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 3 +[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 6 +[52] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, 2023. 8 +[53] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 1, 8 +[54] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024. 2, 4, 6, 7, 8 +[55] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2 +[56] Orr Zohar, Xiaohan Wang, Yonatan Bitton, Idan Szpektor, and Serena Yeung-Levy. Video-star: Self-training enables video instruction tuning with any supervision. arXiv preprint arXiv:2407.06189, 2024. 3 \ No newline at end of file diff --git a/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/images.zip b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f6908631c390e99c4cca6bc8a8e4e1bce73b79f --- /dev/null +++ b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef95d9e11465a4e958f2e4d4d4d4808d1f56e37d73ff5dd6ecb5dd8f70a68af9 +size 801623 diff --git a/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/layout.json b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b4470b1ed2b6f27fdf1cadb006855fa86f31de8 --- /dev/null +++ b/2025/Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation/layout.json @@ -0,0 +1,8413 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "text", + "content": "Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "spans": [ + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "text", + "content": "Yudi Shi" + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "text", + "content": ", Shangzhe Di" + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "text", + "content": ", Qirui Chen" + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "text", + "content": ", Weidi Xie" + }, + { + "bbox": [ + 168, + 142, + 440, + 157 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 159, + 478, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 159, + 478, + 175 + ], + "type": "text", + "content": "School of Artificial Intelligence, Shanghai Jiao Tong University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 176, + 485, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 176, + 485, + 190 + ], + "spans": [ + { + "bbox": [ + 125, + 176, + 485, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 176, + 485, + 190 + ], + "type": "text", + "content": "Coop. Medianet Innovation Center, Shanghai Jiao Tong University, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 179, + 194, + 427, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 194, + 427, + 209 + ], + "spans": [ + { + "bbox": [ + 179, + 194, + 427, + 209 + ], + "type": "text", + "content": "https://zhengrongz.github.io/AoTD/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "spans": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 273, + 297, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 297, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 297, + 477 + ], + "type": "text", + "content": "This paper tackles the problem of video question answering (VideoQA), a task that often requires multi-step reasoning and a profound understanding of spatial-temporal dynamics. While large video-language models perform well on benchmarks, they often lack explainability and spatial-temporal grounding. In this paper, we propose Agent-of-Thoughts Distillation (AoTD), a method that enhances models by incorporating automatically generated Chain-of-Thoughts (CoTs) into the instruction-tuning process. Specifically, we leverage an agent-based system to decompose complex questions into sub-tasks, and address them with specialized vision models, the intermediate results are then treated as reasoning chains. We also introduce a verification mechanism using a large language model (LLM) to ensure the reliability of generated CoTs. Extensive experiments demonstrate that AoTD improves the performance on multiple-choice and open-ended benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 673 + ], + "type": "text", + "content": "Video Question Answering (VideoQA) refers to a critical task that offers a natural interface for human-machine interaction through language [33, 42, 43, 49]. This synergy of visual content and language enhances the accessibility of AI systems for the general public, allowing users to query complex visual content with natural language. By encompassing tasks such as action recognition, object detection, and scene understanding, VideoQA serves as a comprehensive benchmark for evaluating AI's ability to interpret videos, addressing the fundamental questions of 'who', 'what', 'when', and 'where' that are crucial to understand daily life activities, pushing the boundaries of what AI systems can interpret from dynamic visual content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "type": "text", + "content": "Recent literature has primarily explored two avenues in VideoQA. The first involves training large video lan" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 237, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 332 + ], + "type": "text", + "content": "guage models (Video-LLMs) through direct instruction-tuning, using videos paired solely with corresponding questions and answers [1, 22, 24, 53]. While these models excel on public benchmarks, they often lack explainability and struggle with spatial-temporal grounding. This limitation hinders their ability to provide clear reasoning, which is essential for real-world applications where transparency and interpretability are critical [29]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 335, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 479 + ], + "type": "text", + "content": "Conversely, an emerging approach utilizes agent-based systems that decompose complex questions into manageable sub-tasks, each addressed by specialized tools [15, 17, 37]. The results are then aggregated to form a coherent answer. Theoretically, such approach naturally offers greater interpretability, as the reasoning process is divided into explainable steps that can be independently assessed. However, our experiments indicate that current video understanding tools are not strong enough for building reliable agent-based systems. In addition, the high memory demands and time-consuming nature of these systems present significant challenges for their practical use." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 481, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 555, + 553 + ], + "type": "text", + "content": "In this paper, we aim to leverage the advantage of both research lines, enhancing Video-LLM by integrating Chain-of-Thoughts (CoTs) into instruction-tuning, with the CoTs being constructed from the outputs of specialized agent models, capturing the step-by-step reasoning procedure, as illustrated in Figure 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "type": "text", + "content": "In specific, we start by systematically evaluating the off-the-shelf models tailored for atomic video understanding tasks, such as action recognition [39, 41] and language grounding [23], using well-annotated datasets. This comprehensive evaluation allows us to pinpoint the most effective tools for each sub-task, thus laying a robust foundation for constructing reliable chains. Moreover, this process also provides a critical assessment of the broader capabilities of visual models across general and complex scenes, offering valuable insights for future research within the community." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "In addition, we introduce a verification mechanism using a large language model (LLM), designed to assess if the generated CoTs adhere to a clear, step-by-step reason-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 703, + 139, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 139, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 139, + 713 + ], + "type": "text", + "content": "†: Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8523" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 511, + 312 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 511, + 312 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 511, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 511, + 312 + ], + "type": "image", + "image_path": "17bb1544992c14d17c78007bd09f29b16c38ac5375d015f13f76caae5c65d36f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 318, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 54, + 318, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 555, + 363 + ], + "type": "text", + "content": "Figure 1. Our method, AoTD, distills multi-step reasoning and spatial-temporal understanding into a single generative video-language model. When addressing complex VideoQA tasks, the model trained with AoTD (as shown in (b)) enables to generate a step-by-step reasoning to get the correct answer. In contrast, previous models trained solely on question-answer pairs (as in (a)) generate only a final answer, often without intermediate reasoning, which can lead to incorrect conclusions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 468 + ], + "type": "text", + "content": "ing process and incorporate essential information for answering the queries effectively. This mechanism filters out low-quality or logically inconsistent reasoning paths. The remaining CoTs that pass this verification are then distilled into large generative video-language models, significantly enhancing both their performance and interpretability, ultimately leading to the development of more robust, accurate, and interpretable VideoQA systems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 469, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 661 + ], + "type": "text", + "content": "In summary, our contributions are three-fold: (i) we propose a novel approach for enhancing Video-LLMs by distilling high-quality CoTs into their instruction-tuning process. These CoTs capture step-by-step reasoning paths, improving both the model's performance and its interpretability; (ii) to automatically construct the CoTs for any dataset, we employ an agent-based system to decompose complex VideoQA questions into simpler sub-tasks, leveraging off-the-shelf vision models to handle each sub-task. The intermediate outputs from these models can therefore be collected as CoTs for addressing the corresponding visual question; (iii) through extensive experiments, we demonstrate that our distilled model outperforms existing methods across both multiple-choice and open-ended VideoQA benchmarks, enabling to deliver not only accurate answers but also comprehensive reasoning explanations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Video-language models (Video-LLMs). Recent works such as VideoLLaMA2 [4], LLaVA-NeXT-Video [54] and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 373, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 469 + ], + "type": "text", + "content": "VideoChat2 [19], with their excellent architecture design and reasonable instruction-tuning data collection, have achieved a new level of zero-shot results in VideoQA task. However, current end-to-end models still lack interpretability for questions, as well as the ability to think and visually process complex problems in multiple steps, which is an important part for embodied learning and autonomous driving." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 472, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 590 + ], + "type": "text", + "content": "Visual Programming and Agents. With the progress of LLMs, some recent works [5, 15, 37, 45] begin to try to use LLM as planner to solve the complex reasoning task in real scenarios. They attempt to decompose the question into some easier sub-questions, and use different specialist models as agents to solve these sub-questions, and finally gather them to get the answer of the raw question. These models demonstrate a strong ability to obtain trustworthy answers based on the intermediate evidence they get, but they lag far behind the end-to-end model in terms of inference speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "type": "text", + "content": "Visual Chain-of-Thoughts (CoTs). The potential of Chain-of-Thought (CoT) reasoning [40, 46] extends from NLP to the visual domain, highlighting a growing interest in applying this approach across various fields. Numerous studies have incorporated CoTs into visual understanding tasks [11, 30, 36, 55], utilizing powerful Multi-Modal Large Language Models (MLLMs) for generating CoTs or adopting tool-based architectures for sequential problem-solving." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Recent innovations, for example, Visual Program Distillation (VPD) [16] and Fact [10] attempt to address these" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8524" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 545, + 339 + ], + "type": "image", + "image_path": "cc83821021dd8b9d6f71f7c08b74b07c06a7143507b7df54fa6523cd2d235fdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 346, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 555, + 401 + ], + "type": "text", + "content": "Figure 2. Overview on Agent-of-Thoughts Distillation (AoTD). Step 1: Selecting best-performing agents for each sub-task to construct an agent-based system. Step 2: Decomposing question into executable program and leveraging chosen models to solve it sequentially to generate execution trace. Step 3: The execution trace is converted and filtered by LLM to produce high quality natural language CoTs. Step 4: Distilling CoTs into Video-LLM with two forms of prompt, allowing it to achieve a balance between concise answers and comprehensive rationales. The final model is Video-LLM-AoTD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 412, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 412, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 412, + 295, + 544 + ], + "type": "text", + "content": "issues by maintaining the accuracy and diversity of CoTs, while leveraging MLLMs to generate them directly. These approaches decompose complex tasks into code programs, call upon expert models to handle sub-tasks, and utilize the resulting CoTs as training data to fine-tune visual-language models. This process significantly improves the models' ability to generate detailed rationales. Despite the progress in image understanding, there remains a notable oversight in video domains, where reasoning chains can be particularly effective due to the complex spatial-temporal dynamics of video understanding tasks. This is the focus of our paper." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 547, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 643 + ], + "type": "text", + "content": "Concurrent Work. In the recent literature, we notice two work that share similar idea with ours, specifically, Video-STaR [56] construct CoTs using videos and existing labels for instruction-tuning, yet they do not develop an agent-based system. Meanwhile, MotionEpic [8] introduces a Video-of-Thought reasoning framework that integrates video spatial-temporal scene graphs, marking a significant stride towards more nuanced video reasoning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 656, + 228, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 228, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 228, + 670 + ], + "type": "text", + "content": "3. Agent-of-Thoughts Distillation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "In this paper, we propose a novel approach, termed Agent-of-Thoughts Distillation (AoTD), to enhance the Video-LLMs by training them with multi-step Chain-of-Thoughts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "type": "text", + "content": "Specifically, we start by developing an agent-based video understanding system, to generate multi-step reasoning chains that address complex video questions. These reasoning chains are then distilled into one Video-LLM through instruction-tuning. By combining the strengths of agent-based systems and large generative models, our proposed AoTD enables to build more reliable and interpretable VideoQA systems. Figure 2 illustrates the entire process of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 526, + 440, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 440, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 440, + 538 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "text", + "content": "Given a video clip with " + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "text", + "content": " frames, " + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{x_1, \\ldots, x_t\\}" + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "text", + "content": ", and a set of " + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "text", + "content": " questions " + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{q_1, q_2, \\ldots, q_n\\}" + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "text", + "content": ", our goal is to train a Video-LLM capable of producing both concise answers and comprehensive rationales. Depending on the suffix prompt " + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 313, + 544, + 555, + 616 + ], + "type": "text", + "content": ", the model can generate different types of outputs. The process can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 622, + 541, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 622, + 541, + 635 + ], + "spans": [ + { + "bbox": [ + 326, + 622, + 541, + 635 + ], + "type": "interline_equation", + "content": "\\left\\{a _ {i}, \\mathcal {S} _ {i} \\right\\} = \\Phi (\\mathcal {V}, q _ {i}, p _ {s}), \\mathcal {S} _ {i} = \\{\\emptyset \\} \\text {o r} \\left\\{s _ {i, 1}, \\dots , s _ {i, k} \\right\\}", + "image_path": "0ffbb977e032a1590134a73c77aad74e8c4ba8be6f5d80d907ebdfcb4f349662.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "-th question, " + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": " is the answer in freeform text, and " + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": " represents the rationale, consisting of the reasoning process. If the prompt specifies to only generate the answer, " + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "inline_equation", + "content": "S_{i} = \\{\\emptyset\\}" + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": ". Otherwise, if the prompt requires the generation of rationales, " + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "inline_equation", + "content": "S_{i} = \\{s_{i,1}, \\ldots, s_{i,k}\\}" + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "inline_equation", + "content": "s_{i,j}" + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": " corresponds to a reasoning step." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8525" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 126, + 183 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 126, + 183 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 126, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 126, + 183 + ], + "type": "image", + "image_path": "68ec4f03f282cec6fecd90b25251552a7502eceaf5be1c6ac64f55f3fadc66db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 186, + 111, + 194 + ], + "lines": [ + { + "bbox": [ + 73, + 186, + 111, + 194 + ], + "spans": [ + { + "bbox": [ + 73, + 186, + 111, + 194 + ], + "type": "text", + "content": "Video input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 72, + 383, + 91 + ], + "lines": [ + { + "bbox": [ + 126, + 72, + 383, + 91 + ], + "spans": [ + { + "bbox": [ + 126, + 72, + 383, + 91 + ], + "type": "text", + "content": "Question: Which object did the person put down before they held the book? Possible answers: (A) The broom. (B) The paper. (C) The door. (D) The cup." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "lines": [ + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "spans": [ + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": "def execute_command(video, query, possible_answers): hold Clip " + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": " Filter Frames_with_act(video, 'person held the boo before_hold Clip " + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": " trim(video,end=hold clip.sta put clip " + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": " Filter Frames_with_act(before_hold_cl person put down somethi person clip " + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": " Find(place,'person') put_objs " + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": " Query_Objects(person clip, 'object put down by the perso info " + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": " {'object put down by the person': put_obj answer " + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 126, + 95, + 320, + 182 + ], + "type": "text", + "content": " select_answer(query,info,possible的答案 return answer" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 59, + 198, + 148, + 221 + ], + "lines": [ + { + "bbox": [ + 59, + 198, + 148, + 221 + ], + "spans": [ + { + "bbox": [ + 59, + 198, + 148, + 221 + ], + "type": "text", + "content": "Ground Truth Hold a book: Frame[6.64,7.45,...,16.32]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 340, + 117, + 380, + 182 + ], + "blocks": [ + { + "bbox": [ + 340, + 93, + 423, + 116 + ], + "lines": [ + { + "bbox": [ + 340, + 93, + 423, + 116 + ], + "spans": [ + { + "bbox": [ + 340, + 93, + 423, + 116 + ], + "type": "text", + "content": "Object Detection Eval Person:[119,0,241,408] in frame 2.88" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 340, + 117, + 380, + 182 + ], + "lines": [ + { + "bbox": [ + 340, + 117, + 380, + 182 + ], + "spans": [ + { + "bbox": [ + 340, + 117, + 380, + 182 + ], + "type": "image", + "image_path": "33473034039d4be3540613bf8c2f0129853df88be5023d9374259f6dcc822f4a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 184, + 375, + 191 + ], + "lines": [ + { + "bbox": [ + 343, + 184, + 375, + 191 + ], + "spans": [ + { + "bbox": [ + 343, + 184, + 375, + 191 + ], + "type": "text", + "content": "frame 2.88" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 240, + 199, + 334, + 221 + ], + "lines": [ + { + "bbox": [ + 240, + 199, + 334, + 221 + ], + "spans": [ + { + "bbox": [ + 240, + 199, + 334, + 221 + ], + "type": "text", + "content": "Ground Truth Put something: Frame[2.88,3.41,4.78,6.12,6.64]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 338, + 199, + 425, + 219 + ], + "lines": [ + { + "bbox": [ + 338, + 199, + 425, + 219 + ], + "spans": [ + { + "bbox": [ + 338, + 199, + 425, + 219 + ], + "type": "text", + "content": "Ground Truth Person:[128,64,244,371] in frame 2.88" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 381, + 117, + 436, + 182 + ], + "blocks": [ + { + "bbox": [ + 468, + 91, + 530, + 111 + ], + "lines": [ + { + "bbox": [ + 468, + 91, + 530, + 111 + ], + "spans": [ + { + "bbox": [ + 468, + 91, + 530, + 111 + ], + "type": "text", + "content": "Evaluation IoU 65.8%" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 381, + 117, + 436, + 182 + ], + "lines": [ + { + "bbox": [ + 381, + 117, + 436, + 182 + ], + "spans": [ + { + "bbox": [ + 381, + 117, + 436, + 182 + ], + "type": "image", + "image_path": "28053488798311b6240ae2401679e65af7099e28bef2b290410a2c33a74d9364.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 184, + 432, + 191 + ], + "lines": [ + { + "bbox": [ + 383, + 184, + 432, + 191 + ], + "spans": [ + { + "bbox": [ + 383, + 184, + 432, + 191 + ], + "type": "text", + "content": "frame 6.64" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 135, + 520, + 154 + ], + "lines": [ + { + "bbox": [ + 459, + 135, + 520, + 154 + ], + "spans": [ + { + "bbox": [ + 459, + 135, + 520, + 154 + ], + "type": "text", + "content": "Program Output The broom." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 166, + 530, + 185 + ], + "lines": [ + { + "bbox": [ + 447, + 166, + 530, + 185 + ], + "spans": [ + { + "bbox": [ + 447, + 166, + 530, + 185 + ], + "type": "text", + "content": "Ground Truth Answer The broom." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 439, + 198, + 520, + 212 + ], + "lines": [ + { + "bbox": [ + 439, + 198, + 520, + 212 + ], + "spans": [ + { + "bbox": [ + 439, + 198, + 520, + 212 + ], + "type": "text", + "content": "Ground Truth Put a broom somewhere" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 61, + 247, + 100, + 315 + ], + "blocks": [ + { + "bbox": [ + 61, + 224, + 145, + 247 + ], + "lines": [ + { + "bbox": [ + 61, + 224, + 145, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 145, + 247 + ], + "type": "text", + "content": "Filter_frames_with actu hold Clip: Frame 7 to 15" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 61, + 247, + 100, + 315 + ], + "lines": [ + { + "bbox": [ + 61, + 247, + 100, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 100, + 315 + ], + "type": "image", + "image_path": "f5814326e3b4cbfed30c218ec9301e3a608aec00f0c573f60075b8813f43a901.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 316, + 93, + 323 + ], + "lines": [ + { + "bbox": [ + 70, + 316, + 93, + 323 + ], + "spans": [ + { + "bbox": [ + 70, + 316, + 93, + 323 + ], + "type": "text", + "content": "frame 7" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 334, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 555, + 377 + ], + "type": "text", + "content": "Figure 3. Program execution process in an agent-based system. We uniformly sample 32 frames from the video, and to ensure scale consistency, the frame ids of key frames are normalized into these 32 frames. The blue boxes represent the program execution steps, the red boxes denote the ground truth for each step. The combination of red and yellow boxes represents one example process of evaluating Object detection model candidates." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 108, + 247, + 147, + 315 + ], + "blocks": [ + { + "bbox": [ + 108, + 247, + 147, + 315 + ], + "lines": [ + { + "bbox": [ + 108, + 247, + 147, + 315 + ], + "spans": [ + { + "bbox": [ + 108, + 247, + 147, + 315 + ], + "type": "image", + "image_path": "064d0c33aa2d1448fea26394963b6b091ec7e32e3ac5c2b5551a0df957bb733b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 316, + 141, + 323 + ], + "lines": [ + { + "bbox": [ + 115, + 316, + 141, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 141, + 323 + ], + "type": "text", + "content": "frame 15" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 153, + 247, + 191, + 314 + ], + "blocks": [ + { + "bbox": [ + 151, + 199, + 230, + 221 + ], + "lines": [ + { + "bbox": [ + 151, + 199, + 230, + 221 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 230, + 221 + ], + "type": "text", + "content": "Ground Truth \nFrame[2.88,3.41,4.78, 6.12,6.64]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 152, + 224, + 217, + 247 + ], + "lines": [ + { + "bbox": [ + 152, + 224, + 217, + 247 + ], + "spans": [ + { + "bbox": [ + 152, + 224, + 217, + 247 + ], + "type": "text", + "content": "trim\nbefore_hold Clip:\nFrame 0 to 7" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 247, + 191, + 314 + ], + "lines": [ + { + "bbox": [ + 153, + 247, + 191, + 314 + ], + "spans": [ + { + "bbox": [ + 153, + 247, + 191, + 314 + ], + "type": "image", + "image_path": "b2bef87da79297f89f8430cee1e8a8925bfcdc034f9d33183a7c663f407b30e9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 316, + 182, + 323 + ], + "lines": [ + { + "bbox": [ + 160, + 316, + 182, + 323 + ], + "spans": [ + { + "bbox": [ + 160, + 316, + 182, + 323 + ], + "type": "text", + "content": "frame 0" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 196, + 247, + 235, + 314 + ], + "blocks": [ + { + "bbox": [ + 196, + 247, + 235, + 314 + ], + "lines": [ + { + "bbox": [ + 196, + 247, + 235, + 314 + ], + "spans": [ + { + "bbox": [ + 196, + 247, + 235, + 314 + ], + "type": "image", + "image_path": "d593a95de6c5bae5dc1613aebc8019086b2472167d1aae2446cbb690be4fa80f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 316, + 227, + 323 + ], + "lines": [ + { + "bbox": [ + 205, + 316, + 227, + 323 + ], + "spans": [ + { + "bbox": [ + 205, + 316, + 227, + 323 + ], + "type": "text", + "content": "frame 7" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 242, + 247, + 280, + 314 + ], + "blocks": [ + { + "bbox": [ + 241, + 224, + 324, + 247 + ], + "lines": [ + { + "bbox": [ + 241, + 224, + 324, + 247 + ], + "spans": [ + { + "bbox": [ + 241, + 224, + 324, + 247 + ], + "type": "text", + "content": "Filter Frames with act put Clip: Frame 2 to 7" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 247, + 280, + 314 + ], + "lines": [ + { + "bbox": [ + 242, + 247, + 280, + 314 + ], + "spans": [ + { + "bbox": [ + 242, + 247, + 280, + 314 + ], + "type": "image", + "image_path": "f7ee3066790d72441e1c60cd4265410b280fe81a97b442e8115d29c956db4c94.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 316, + 272, + 323 + ], + "lines": [ + { + "bbox": [ + 249, + 316, + 272, + 323 + ], + "spans": [ + { + "bbox": [ + 249, + 316, + 272, + 323 + ], + "type": "text", + "content": "frame 2" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 282, + 247, + 333, + 314 + ], + "blocks": [ + { + "bbox": [ + 282, + 247, + 333, + 314 + ], + "lines": [ + { + "bbox": [ + 282, + 247, + 333, + 314 + ], + "spans": [ + { + "bbox": [ + 282, + 247, + 333, + 314 + ], + "type": "image", + "image_path": "7b8e75d154e539de4bde52c453f260139df2b6283158532a65bdacdbcbf5ff0d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 316, + 324, + 323 + ], + "lines": [ + { + "bbox": [ + 302, + 316, + 324, + 323 + ], + "spans": [ + { + "bbox": [ + 302, + 316, + 324, + 323 + ], + "type": "text", + "content": "frame 7" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 340, + 247, + 380, + 314 + ], + "blocks": [ + { + "bbox": [ + 340, + 224, + 426, + 244 + ], + "lines": [ + { + "bbox": [ + 340, + 224, + 426, + 244 + ], + "spans": [ + { + "bbox": [ + 340, + 224, + 426, + 244 + ], + "type": "text", + "content": "Find person Clip:[120,0,239, 413] in frame 2 ..." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 247, + 380, + 314 + ], + "lines": [ + { + "bbox": [ + 340, + 247, + 380, + 314 + ], + "spans": [ + { + "bbox": [ + 340, + 247, + 380, + 314 + ], + "type": "image", + "image_path": "fa4388d34ac4a7bb7dce5958c6f8dbdd68dbc98f96024c08f1c6f7fb1a5437f6.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 316, + 372, + 323 + ], + "lines": [ + { + "bbox": [ + 348, + 316, + 372, + 323 + ], + "spans": [ + { + "bbox": [ + 348, + 316, + 372, + 323 + ], + "type": "text", + "content": "frame 2" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 395, + 247, + 433, + 314 + ], + "blocks": [ + { + "bbox": [ + 395, + 247, + 433, + 314 + ], + "lines": [ + { + "bbox": [ + 395, + 247, + 433, + 314 + ], + "spans": [ + { + "bbox": [ + 395, + 247, + 433, + 314 + ], + "type": "image", + "image_path": "23a55f3e19616eaa13cd27195ab0d75a545a2ceee26f47814bdef5a57f5b4c2a.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 316, + 425, + 323 + ], + "lines": [ + { + "bbox": [ + 402, + 316, + 425, + 323 + ], + "spans": [ + { + "bbox": [ + 402, + 316, + 425, + 323 + ], + "type": "text", + "content": "frame 7" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 441, + 247, + 480, + 314 + ], + "blocks": [ + { + "bbox": [ + 441, + 224, + 499, + 239 + ], + "lines": [ + { + "bbox": [ + 441, + 224, + 499, + 239 + ], + "spans": [ + { + "bbox": [ + 441, + 224, + 499, + 239 + ], + "type": "text", + "content": "Query_Objects\nput_objects: broom" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 247, + 480, + 314 + ], + "lines": [ + { + "bbox": [ + 441, + 247, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 441, + 247, + 480, + 314 + ], + "type": "image", + "image_path": "745f32b8e47f4bd4e024ad8eb436aed743fa1fe27d1b25abdf9c5a793e137487.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 316, + 471, + 323 + ], + "lines": [ + { + "bbox": [ + 449, + 316, + 471, + 323 + ], + "spans": [ + { + "bbox": [ + 449, + 316, + 471, + 323 + ], + "type": "text", + "content": "frame 2" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 493, + 247, + 531, + 314 + ], + "blocks": [ + { + "bbox": [ + 493, + 247, + 531, + 314 + ], + "lines": [ + { + "bbox": [ + 493, + 247, + 531, + 314 + ], + "spans": [ + { + "bbox": [ + 493, + 247, + 531, + 314 + ], + "type": "image", + "image_path": "d20ccc585c3a81054b4e64d2582800c4a1aa792b891431597e221db85c6225ab.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 316, + 523, + 323 + ], + "lines": [ + { + "bbox": [ + 500, + 316, + 523, + 323 + ], + "spans": [ + { + "bbox": [ + 500, + 316, + 523, + 323 + ], + "type": "text", + "content": "frame 7" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "content": "Discussion. Unlike existing models that are instruction-tuned on VideoQA datasets using simple question-answer pairs, which bypass the intermediate thought process, our approach emphasizes the importance of training with Chain-of-Thoughts (CoTs). In the following section, we outline the process for generating high-quality CoTs from existing VideoQA datasets." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 478, + 284, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 284, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 284, + 491 + ], + "type": "text", + "content": "3.2. CoT Construction with Agent-based System" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": "Recent work, such as STAR [42], has introduced the idea of employing executable symbolic programs, to directly decompose questions into sub-tasks. When combined with scene graphs that contain comprehensive video information from key frames—such as object locations, interactions, and actions—these programs facilitate the generation of concise Chain-of-Thoughts (CoTs) through direct execution of symbolic operations. However, datasets of this nature are limited in scale, we therefore propose to first build an agent-based system, capable of breaking down complex questions into simpler sub-tasks, and the intermediate outputs from this system can then be employed to construct CoTs for any existing VideoQA dataset." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Agent-based VideoQA. Assuming we are given a video input " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{V})" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", questions " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q})" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", and a set of visual models " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{M} = \\{\\phi_{\\mathrm{act}}, \\phi_{\\mathrm{det}}, \\dots, \\phi_{\\mathrm{qa}}\\})" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", an LLM-based agent core " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\pi(\\cdot))" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " processes the question along with the documentation of the visual models " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{T})" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", which includes variables and function-" + } + ] + } + ], + "index": 46 + }, + { + "type": "table", + "bbox": [ + 317, + 386, + 553, + 570 + ], + "blocks": [ + { + "bbox": [ + 317, + 386, + 553, + 570 + ], + "lines": [ + { + "bbox": [ + 317, + 386, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 386, + 553, + 570 + ], + "type": "table", + "html": "
Sub-taskModel nameMetricNumber (%)
question decompositionCodeQwen1.5-Chat (7B) [2]52.7
GPT-3.5-Turbo [31]Acc73.1
DeepSeek-Coder-Instruct (6.7B) [6]85.7
object detectionOWL-ViT v1 [26]47.3
GLIP [20]IoU58.9
OWL-ViT v2 [28]63.0
temporal groundingLITA (13B) [18]11.7 / 20.2
TimeChat (7B) [35]IoU / Recall13.9 / 23.1
UniVTG [23]24.7 / 35.3
action recognitionInternVideo2 (1B) [39]7.6
Open-VCLIP [41]Top1-Acc8.9
LLaVA-NeXT-Video-DPO (7B) [54]18.2
question answeringLLaMA-VID (7B) [21]43.5
SeViLA [48]Acc46.5
LLaVA-NeXT-Video-DPO (7B) [54]53.4
", + "image_path": "ba208850aa8a27cae03d9cd702c94f5f5d8efce19508089e06e0e467aa20830f.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "table_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "text", + "content": "Table 1. Sub-tasks definition and evaluation results. We choose 3 model candidates for each sub-task and evaluate them in STAR training set with the corresponding metrics. Models with best performance are placed at the bottom of each column." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "alities. The agent subsequently decomposes the question into sub-tasks formatted as Python code, and resolves them by invoking the appropriate visual models through function calls. It is important to note that the visual models can be arranged in various orders depending on the specific question, ensuring flexibility in problem-solving." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "content": "Specifically, as illustrated by the example in Figure 3, the" + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8526" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "question is first decomposed into a series of sub-tasks, including temporal grounding, object detection, and question answering. The corresponding specialized models are then executed sequentially to address these sub-tasks, ultimately yielding the final answer " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 140, + 252, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 140, + 252, + 168 + ], + "spans": [ + { + "bbox": [ + 97, + 140, + 252, + 168 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\{\\phi_ {\\text {g r o u n d}}, \\phi_ {\\det }, \\phi_ {\\mathrm {q a}} \\right\\} := \\pi \\left(q _ {i}, \\mathcal {T}\\right), \\\\ y _ {i} = \\phi_ {\\text {g r o u n d}} (\\mathcal {V}) \\rightarrow \\phi_ {\\det} (\\mathcal {V}) \\rightarrow \\phi_ {\\mathrm {q a}} (\\mathcal {V}) \\\\ \\end{array}", + "image_path": "2f0246cb2a20eb4c36d9034faec616cd95f9ddc84564f0f43492286cdfddf115.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "content": "CoT Construction. To ensure the correctness of outputs at all the intermediate steps, we leverage the training set from STAR for hyperparameter tuning, enabling us to identify the most effective model for each sub-task within the agent-based system. By following the provided programs, we evaluate the performance of the corresponding vision models on tasks such as object detection and action recognition. Given the availability of complete reasoning chains, we independently assess each sub-task using ground truth data for all preceding steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 510 + ], + "type": "text", + "content": "As shown in Table 1, we present the evaluation results for the various sub-tasks. Specifically, for question decomposition, we compare several code LLMs, with DeepSeekCoder-Instruct achieving the highest performance, outperforming even GPT-3.5-Turbo. In object detection, OWLViT v2 records the highest Intersection over Union (IoU) score, showcasing its superior open-vocabulary detection capability. The results for temporal grounding indicate that while UniVTG leads in performance, there remains a need for further advancements in this area. In action recognition, our evaluations show that generative models outperformed discriminative models, likely due to the fine-grained action list provided by the STAR dataset. However, the performance of both model types reveals significant room for improvement. Finally, in the one-hop question answering sub-task, all models perform admirably, with LLaVA-NeXT-Video-DPO standing out as a top performer, consistent with its strong results on other benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": "With these high-performing models, we implement the agent-based approach on VideoQA datasets that consist solely of QA pairs. During the execution of the programs, we record all intermediate outputs to construct the CoTs. Since the outputs from these vision models vary in format—such as bounding boxes and free-form text—we employ another LLM to translate the execution trace into natural language for better use in the distillation process. Detailed examples are provided in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 156, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 156, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 156, + 635 + ], + "type": "text", + "content": "3.3. CoT Verification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "content": "To refine the quality of reasoning chains for VideoQA samples, we implement a two-step verification: (i) we filter execution traces to retain only those, where the program can reach correct output. For multiple-choice datasets, the output must match the correct answer exactly, while for open-ended datasets, we prompt the LLM to verify correctness," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 332, + 71, + 535, + 170 + ], + "blocks": [ + { + "bbox": [ + 332, + 71, + 535, + 170 + ], + "lines": [ + { + "bbox": [ + 332, + 71, + 535, + 170 + ], + "spans": [ + { + "bbox": [ + 332, + 71, + 535, + 170 + ], + "type": "table", + "html": "
DatasetDescription# Labels# CoTs
AGQACompositional25.0K5.4K
ANetQACompositional25.0K3.6K
STARCompositional45.7K11.2K
NExT-QATemporal & Causal34.1K12.1K
CLEVRERSpatial & Temporal21.0K-
EgoQAEgo-centric7.8K-
Total158.6K32.3K
", + "image_path": "f4a343deca7b142e248d2934e3e4ef9a109335e927428e83f33a8eff24bbf846.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 173, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 553, + 228 + ], + "type": "text", + "content": "Table 2. Dataset statistics. The column \"#Labels\" indicates the number of VideoQA pairs, which include the video, query, possible answers (multiple-choice), and the correct answer. \"#CoTs\" refers to the number of CoTs generated using our agent-based system for each dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 244, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 340 + ], + "type": "text", + "content": "accounting for format differences; (ii) we prompt the LLM to evaluate the logical coherence and usefulness of the reasoning chains in solving the problem. The model assesses whether the CoTs follow a clear, step-by-step reasoning process and provides a binary evaluation ('Yes' or 'No') to indicate their quality (detailed prompts are included in Appendix D). This two-step approach ensures that only high-quality CoTs are retained for further distillation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 553, + 405 + ], + "type": "text", + "content": "In Table 2, we provide the statistics for the remaining generated CoTs for different datasets. We primarily select compositional QA datasets, as these require the model to process spatial-temporal information from different events comprehensively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 429, + 449, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 449, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 449, + 441 + ], + "type": "text", + "content": "3.4. Step-by-step Distillation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 451, + 553, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 553, + 511 + ], + "type": "text", + "content": "In this section, we describe the process of distilling the generated CoTs into a Video-LLM. This distillation enhances the model's ability for spatial-temporal video understanding and multi-step reasoning, thereby improving its performance on complex VideoQA tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": "In specific, using the generated CoTs, we can build the dataset " + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "inline_equation", + "content": "D = \\{(\\mathcal{V}_j,q_j,\\hat{y}_j,c_j,p_s)\\}_{j = 1}^N" + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": " is the total number of samples in the distilling dataset, " + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_j" + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": " is the video input, " + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": " is the question, " + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "inline_equation", + "content": "\\hat{y}_j" + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": " is the ground-truth answer, " + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": " is the generated CoT, " + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "inline_equation", + "content": "p_{s}" + }, + { + "bbox": [ + 313, + 517, + 554, + 649 + ], + "type": "text", + "content": " is the task-specific suffix prompt, to distinguish different tasks, for example, for multiple-choice VQA, the prompt can be: \"Answer with the option's letter from the given choices directly and only give the best option\", and for open-ended VQA, the prompt can be: \"Answer in one word or phrase\". Please refer to detailed prompts in Appendix D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "At distillation stage, we minimize the cross-entropy loss of predicting both the answer and the CoTs, we replace the suffix prompt " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "p_{s}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " with \"Explain the rationale to answer the question\", to indicate whether we want a question answer or a rationale to explain the thinking steps. Following [10]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8527" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 298, + 264 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 298, + 264 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 298, + 264 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 298, + 264 + ], + "type": "image", + "image_path": "cddd742c8d91e51cb002c8205855594521b7d735f28b8dd74af340e6e7ffb96b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "type": "text", + "content": "Figure 4. Visualization of rationales. LLaVA-NeXT-Video-AoTD can output rationales containing both spatial-temporal grounding of key information and step-by-step thinking process to solve the question." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 308, + 69, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 308, + 69, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 308, + 69, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 308, + 69, + 553, + 264 + ], + "type": "image", + "image_path": "56007bbdf03faa4e89fc8b0794b261f650a83e927808135b2cc00ea59b927197.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 214, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 214, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 214, + 316 + ], + "type": "text", + "content": "and [16], our optimization objective is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 326, + 282, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 326, + 282, + 376 + ], + "spans": [ + { + "bbox": [ + 69, + 326, + 282, + 376 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\lambda \\mathcal {L} _ {\\text {r a t i o n a l e}} \\\\ = \\sum_ {j = 1} ^ {N} \\ell (\\Phi (\\mathcal {V} _ {j}, q _ {j}, p _ {s}), \\hat {y} _ {j}) + \\lambda \\ell (\\Phi (\\mathcal {V} _ {j}, q _ {j}, p _ {s}), c _ {j}) \\\\ \\end{array}", + "image_path": "1e3021c9eb5457a9b6be921720fd26ff97511f46028eb8ea8dd2055955ba27d8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 433 + ], + "type": "text", + "content": "Here, we set " + }, + { + "bbox": [ + 55, + 384, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 384, + 296, + 433 + ], + "type": "text", + "content": " to 1 to ensure the importance of answer and rationale are equally considered. Notice that, not all the QA pairs can generate qualified CoT. In that case, the " + }, + { + "bbox": [ + 55, + 384, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rationale}}" + }, + { + "bbox": [ + 55, + 384, + 296, + 433 + ], + "type": "text", + "content": " will be set to 0." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 441, + 293, + 617 + ], + "blocks": [ + { + "bbox": [ + 56, + 441, + 293, + 617 + ], + "lines": [ + { + "bbox": [ + 56, + 441, + 293, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 293, + 617 + ], + "type": "table", + "html": "
DatasetSizeTypeTrainEval
traineval
MC-VQA
STAR [42]45.7K7.1KCompositional✓✓
NExT-QA [43]34.1K5.0KTemporal & Causal✓✓
CLEVRER [47]21.0K-Spatial-temporal✓✗
Perception-Test [33]-11.5KGeneral✗✓
MVBench [19]-4.0KGeneral✗✓
VideoMME [9]-2.7KGeneral✗✓
VSIBench [44]-5.0Kspatial-temporal✗✓
OE-VQA
AGQA [14]25.0K2.0KCompositional✓✓
ANetQA [50]25.0K2.0KCompositional✓✓
EgoQA [13]7.8K-Ego-centric✓✗
Activitynet-QA [49]-8.0KGeneral✗✓
Video-ChatGPT [24]-3.0KGeneral✗✓
", + "image_path": "24465b0e0f20c366b1f2226a485cf8200148735254a8e7e780c05e43a68c8928.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 621, + 263, + 632 + ], + "lines": [ + { + "bbox": [ + 86, + 621, + 263, + 632 + ], + "spans": [ + { + "bbox": [ + 86, + 621, + 263, + 632 + ], + "type": "text", + "content": "Table 3. Training and evaluation datasets statics." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "In this section, we present the experimental setup (Sec. 4.1) and results on various VideoQA benchmarks (Sec. 4.2). Extensive ablation studies have also been conducted to further examine the contributions of our approach in Sec. 4.3, and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "type": "text", + "content": "an evaluation on the quality of rationales generated by the distilled model is made in Sec. 4.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 339, + 430, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 430, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 430, + 353 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "type": "text", + "content": "Base model. We use LLaVA-NeXT-Video (7B) [54] (LNV for short) as base Video-LLM, which has shown remarkable performance on image-centric tasks, for example image question answering [51]. We present comparison on naive instruction-tuning with video question answering dataset or with additional CoT distillation. For CoT conversion and verification, we prompt LLaMA-3.1-8B with the manually-designed instruction and some in-context examples. Detailed prompts are provided in Appendix D." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 470, + 554, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 565 + ], + "type": "text", + "content": "Instruction tuning. We utilize both multiple-choice and open-ended QA data, along with the generated CoTs, to fine-tune the base video question answering model, as summarised in Table 2. The resulting distilled model is named LLaVA-NeXT-Video-AoTD (LNV-AoTD for short). Additionally, as baseline, we also train another version of the model using only the basic QA data, which we refer to as LLaVA-NeXT-Video-Instruct (LNV-Instruct for short)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Evaluation benchmarks. We conduct extensive evaluations on Multiple-Choice Video QA (MC-VQA) and Open-Ended Video QA (OE-VQA). We report the top-1 accuracy for all MC benchmarks, which means the proportion of the output equal to the answer. We report a GPT-assessed Acc. and Score with the help of GPT-3.5-turbo-0613 for all OE benchmarks. For each question, GPT delivers a binary decision indicating whether the output is correct or incorrect, along with a similarity score reflecting the degree of alignment between the output and the correct answer. The term 'Acc.' refers to the percentage of correct outputs, while 'Score' represents the average similarity scores. For the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8528" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 533, + 236 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 533, + 236 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 533, + 236 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 533, + 236 + ], + "type": "table", + "html": "
ModelMVBench (Acc.)VideoMME (Acc.)STAR (Acc.)NExT-QA (Acc.)Perception-Test (Acc.)VSIBench (Acc.)
Proprietary Models
Gemini 1.5 Pro [12]-75.0---45.4
GPT4-V [32]43.759.9----
Open-source Models
LLaMA-VID (7B) [21]41.925.9--44.6-
Video-LLaVA (7B) [22]41.039.9--44.3-
VideoChat2 (7B) [19]51.133.759.0*68.6*47.3-
VideoLLaMA2 (7B) [4]53.445.158.5*62.3*49.6-
LLaVA-NeXT-Video (7B) [54]46.5*41.0*52.4*61.6*47.5*19.7*
LLaVA-NeXT-Video-Instruct (7B)53.443.272.277.150.326.7
LLaVA-NeXT-Video-AoTD (7B)55.645.074.377.650.628.8
", + "image_path": "b5693ec54e54aabfa2d623aed0481be0e214ab27dfe26246f9d31d822886deb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "type": "text", + "content": "Table 4. Comparison with Video-LLMs on MC-VQA benchmarks. LLaVA-NeXT-Video-AoTD outperforms all other baselines the and the version without CoT distillation. * means results reproduced by ourselves. Results without signs are retrieved from [4] and [44]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 266, + 297, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 363 + ], + "type": "text", + "content": "evaluation on AGQA and ANetQA, due to the large volume of test set, we test on a subset of samples. We evenly select the benchmark in-domain and out-of-domain for testing to ensure a comprehensive and reasonable evaluation of the model capability. Noted that though VSIBench has both MC and OE questions, it doesn't need GPT for score, so we classify it into MC benchmarks for convenience. Detailed statistics for evaluation benchmarks are shown in Table 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 174, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 174, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 174, + 383 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 472 + ], + "type": "text", + "content": "We divide the comparison into two parts: the first focuses on comparing the distilled model with other baselines, while the second examines the difference between the instruct version and the AoTD version. Note that, the latter part will be mainly compared and discussed, to demonstrate the model's improvement relative to its previous performance, as well as establishing the transferability of the method across models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": "MC-VQA performance. As shown in Table 4, our LLaVA-NeXT-Video-AoTD achieves superior performance across all benchmarks. Several key observations can be made: (i) comparing to the base model, even a simple instruction-tuning on certain VideoQA datasets significantly enhances the model's question-answering performance. This improvement is notable, as the base model was primarily trained on static images and struggled with video understanding; (ii) our model, instruction-tuned with CoT distillation, demonstrates further performance enhancements across all benchmarks, particularly on the compositional VideoQA benchmark (STAR) and comprehensive benchmarks (VideoMME, MVBench). This suggests that our AoTD method effectively improves the model's ability to address complex problems and interpret spatial-temporal scenes; (iii) the distilled model consistently outperforms all other baselines across almost all benchmarks, even when compared to more powerful models. This finding shows that our method effectively bridges performance gaps created by varying model components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 266, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 472 + ], + "type": "text", + "content": "OE-VQA performance. As shown in Table 5, LLaVA-NeXT-Video-AoTD outperforms the Instruct variant across all open-ended VideoQA benchmarks. Notably, it achieves a greater percentage increase compared to the MC-VQA benchmarks, suggesting that CoT distillation may be more effective for open-ended generation than for multiple-choice selection. While the distilled model scores higher than most models listed in the table, it does not surpass LLaVA-NeXT-Video on certain benchmarks. We conjecture this is due to the model's extensive training on images, that can also benefit the question answering without requiring complex reasonings, as also suggested by the findings in VideoLLaMA2 [4]. Additionally, the inherent challenges of evaluating open-ended VQA may influence the results. Assessments conducted by GPT can be biased or inaccurate, and the metrics we employ primarily indicate general trends rather than providing absolute accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 478, + 408, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 408, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 408, + 491 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "type": "text", + "content": "Analysis on CoT filtering. To prove the effectiveness of our filtering mechanism, we trained an alternative model without CoT filtering while maintaining all other settings, i.e., using 36.3K verified CoTs for distillation. As shown in Table 9, the model's performance declines significantly on both the Multiple-Choice VQA and Open-Ended VQA benchmarks when the CoT filtering mechanism is not utilized. This confirms that employing large language models (LLMs) to filter CoTs is crucial for enhancing data quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Analysis on model transferability. As AoTD is a distillation method that leverages Chain-of-Thoughts (CoTs), it can theoretically be applied to any Video-LLMs. To assess the transferability of our method, we conduct experiments on another very recent model, LLaVA-OneVision (7B) [3]. As shown in Table 9, our method also demonstrates significant improvements on the benchmarks, showing the transferability and robustness of the approach. Due to the rapid advancements in the computer vision field, evaluating all" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8529" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 515, + 258 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 515, + 258 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 515, + 258 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 515, + 258 + ], + "type": "table", + "html": "
ModelANetQA (Acc./Score)AGQA (Acc./Score)Corr.Video-ChatGPT (Score)ActivityNet (Acc./Score)
Deta.Cont.Temp.Cons.
Proprietary Models
Gemini 1.5 Pro [12]-------56.7/-
GPT4-V [32]--4.093.884.373.944.0259.5/-
Open-Source Models
VideoLLaMA (7B) [53]--1.962.182.161.821.7912.4/1.1
Video-ChatGPT (7B) [24]--2.502.572.692.162.2035.2/2.7
LLaMA-VID (7B) [21]--2.963.003.532.462.5147.4/3.3
Video-LLaVA (7B) [22]--2.872.943.442.452.5145.3/3.3
VideoChat2 (7B) [19]--3.022.883.512.662.8149.1/3.3
VideoLLaMA2 (7B) [4]--3.093.093.682.633.2549.9/3.3
LLaVA-NeXT-Video (7B) [54]46.4/3.3*27.4/2.2*3.26*3.22*3.77*2.47*2.99*54.3/3.2*
LLaVA-NeXT-Video-Instruct (7B)47.1/3.159.3/3.42.962.813.352.422.8250.0/3.3
LLaVA-NeXT-Video-AoTD (7B)53.9/3.460.9/3.63.113.003.602.412.9153.2/3.4
", + "image_path": "79c8196bcb3e5bb302a71ba3b64537fdac3677ed8b8d94b09128420b55ff4f0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 294, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 330 + ], + "type": "text", + "content": "models and benchmarks is prohibitively infeasible. Thus, we focus on assessing some representative models against selected benchmarks to provide a representative evaluation." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 332, + 287, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Table 5. Comparison with Video-LLMs on OE-VQA benchmarks. LLaVA-NeXT-Video-AoTD improves performance in all open-ended benchmarks compared with the Instruct version. * means results reproduced by ourselves. Results without signs are retrieved from [4]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 332, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 63, + 332, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 63, + 332, + 287, + 407 + ], + "type": "table", + "html": "
ModelFilteringMVBench (Acc.)STAR (Acc.)AGQA (Acc. / Score)
LNV-AoTDX53.773.359.5/3.5
LNV-AoTD✓55.674.360.9/3.6
Onevision-Instruct-59.275.865.6/3.7
Onevision-AoTD✓60.576.665.7/3.7
", + "image_path": "f7476e825bd72f9d7530ff5fb097ca67a55bfbde0ae0416fcb964229440c15a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 427, + 195, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 195, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 195, + 438 + ], + "type": "text", + "content": "4.4. Evaluation on Rationales" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "type": "text", + "content": "To verify whether the model has effectively learned multistep reasoning through CoTs distillation, we analyze the rationales generated by the model. Specifically, we extract and evaluate the temporal and spatial information embedded within these rationales. This approach extends beyond merely assessing the correctness of the final answer, which could be influenced by biases or other external factors. By examining the reasoning process in detail, it enables a more accurate understanding of the model's ability to perceive and reason about spatial and temporal relationships." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 567, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 567, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 567, + 295, + 674 + ], + "type": "text", + "content": "Evaluation protocols. We randomly select 200 samples from the STAR validation set and run inference on them using the suffix prompt, recording the generated rationales. From these rationales, we extract the predicted temporal windows and bounding boxes, comparing them to the ground truth. For the spatial part, we calculate the IoU between the predicted and ground truth bounding boxes. For the temporal part, we compute IoU and Recall, leveraging the frame-level annotations provided in the dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "type": "text", + "content": "Evaluation results. Table 7 presents the evaluation results. For comparison, we also test UniVTG for temporal reasoning and OWL-ViT v2 for spatial reasoning. The re" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 292, + 553, + 362 + ], + "blocks": [ + { + "bbox": [ + 65, + 411, + 285, + 422 + ], + "lines": [ + { + "bbox": [ + 65, + 411, + 285, + 422 + ], + "spans": [ + { + "bbox": [ + 65, + 411, + 285, + 422 + ], + "type": "text", + "content": "Table 6. Ablation results of CoT filtering and transferability." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 292, + 553, + 362 + ], + "lines": [ + { + "bbox": [ + 315, + 292, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 315, + 292, + 553, + 362 + ], + "type": "table", + "html": "
ModelTemporal GroundingSpatial Grounding
IoU (%)Recall (%)IoU (%)
UniVTG22.831.0-
OWL-ViT v2--64.7
LNV-InstructXXX
LNV-AoTD21.734.045.2
", + "image_path": "ed677f5ba042f6499b9267ef48730d5abf5f72f63689e51714eba38728ae60ca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 329, + 366, + 537, + 377 + ], + "lines": [ + { + "bbox": [ + 329, + 366, + 537, + 377 + ], + "spans": [ + { + "bbox": [ + 329, + 366, + 537, + 377 + ], + "type": "text", + "content": "Table 7. Temporal and spatial abilities evaluation results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 385, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 458 + ], + "type": "text", + "content": "sults show that LNV-Instruct struggles to generate valid rationales, even when using the suffix prompt. In contrast, LNV-AoTD demonstrates comparable performance to specialized models in both spatial and temporal reasoning, indicating that the model successfully acquired these abilities through the distillation process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 467, + 388, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 388, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 388, + 480 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 490, + 555, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 646 + ], + "type": "text", + "content": "We present Agent-of-Thoughts Distillation (AoTD), that aims to distill multi-step reasoning and spatial-temporal understanding into a large video-language model (Video-LLM). Our method introduces an agent-based system that automates the generation of Chain-of-Thoughts (CoTs) from various VideoQA datasets, by breaking down complex questions into manageable sub-tasks that can be addressed by specialized vision models. Extensive experiments validate that the distilled model significantly enhances performance on both MC-VQA and OE-VQA benchmarks, underscoring the effectiveness of our approach. We believe AoTD represents a promising future direction for advancing the reasoning abilities in Video-LLMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 411, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 411, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 411, + 680 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "This work is funded by National Key R&D Program of China (No.2022ZD0161400)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8530" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 280 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 282, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 295, + 326 + ], + "type": "text", + "content": "[3] Li Bo, Zhang Yuanhan, Guo Dong, Zhang Renrui, Li Feng, Zhang Hao, Zhang Kaichen, Li Yanwei, Liu Ziwei, and Li Chunyuan. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 327, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 295, + 382 + ], + "type": "text", + "content": "[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. arXiv preprint arXiv:2406.07476, 2024. 2, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 384, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 295, + 427 + ], + "type": "text", + "content": "[5] Rohan Choudhury, Koichiro Niinuma, Kris M. Kitani, and Laszlo A. Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 431, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 495 + ], + "type": "text", + "content": "[6] Guo Daya, Zhu Qihao, Yang Dejian, Dong Zhenda Xie, Kai, Zhang Wentao, Chen Guanting, Bi Xiao, Y. Wu, Y.K. Li, Luo Fuli, and Liang Yingfei, Xiongand Wenfeng. Deepseek-coder: When the large language model meets programming - the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "type": "text", + "content": "[7] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 295, + 599 + ], + "type": "text", + "content": "[8] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In Proceedings of the International Conference on Machine Learning, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 601, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 295, + 656 + ], + "type": "text", + "content": "[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[10] Minghe Gao, Shuang Chen, Liang Pang, Yuan Yao, Jisheng Dang, Wenqiao Zhang, Juncheng Li, Siliang Tang, Yueting Zhuang, and Tat-Seng Chua. Fact: Teaching mllms with faithful, concise and transferable rationales. In ACM Multimedia, 2024. 2, 5, 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[11] Timin Gao, Peixian Chen, Mengdan Zhang, Chaoyou Fu, Yunhang Shen, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Xing Sun, Liujuan Cao, et al. Cantor: Inspiring multimodal chain-of-thought of mllm. In ACM Multimedia, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 129, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 163 + ], + "type": "text", + "content": "[12] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "text", + "content": "[13] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 231, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 284 + ], + "type": "text", + "content": "[14] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "text", + "content": "[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 332, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 398 + ], + "type": "text", + "content": "[16] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "text", + "content": "[17] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent. In Advances in Neural Information Processing Systems, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "text", + "content": "[18] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In Proceedings of the European Conference on Computer Vision, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "text", + "content": "[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "type": "text", + "content": "[20] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[21] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 4, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[22] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8531" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "Conference on Empirical Methods in Natural Language Processing, 2024. 1, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[23] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In Proceedings of the International Conference on Computer Vision, 2023. 1, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 1, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[25] Ahmad Mahmood, Ashmal Vayani, Muzammal Naseer, Salman Khan, and Fahad Khan. Vurf: A general-purpose reasoning and self-refinement framework for video understanding. arXiv preprint arXiv:2403.14743, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 319 + ], + "type": "text", + "content": "[26] Minderer Matthias, Gritsenko Alexey, Stone Austin, Neumann Maxim, Weissenborn Dirk, Dosovitskiy Alexey, Mahendran Aravindh, Arnab Anurag, Dehghani Mostafa, Shen Zhuoran, Wang Xiao, Zhai Xiaohua, Kipf Thomas, and Houlsby Neil. Simple open-vocabulary object detection with vision transformers. In Proceedings of the European Conference on Computer Vision, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[27] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[28] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. In Advances in Neural Information Processing Systems, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[29] Arindam Mitra, Luciano Del Corro, Shweti Mahajan, Andres Codas, Clarisse Simoes, Sahaj Agarwal, Xuxi Chen, Anastasia Razdaibiedina, Erik Jones, Kriti Aggarwal, et al. Orca 2: Teaching small language models how to reason. arXiv preprint arXiv:2311.11045, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[30] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[31] OpenAI. Gpt-3.5-turbo system card. https:// platform.openuai.com/docs/models/gpt-3- 5-turbo,2023.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 568 + ], + "type": "text", + "content": "[32] OpenAI. Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card, 2023.7,8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 678 + ], + "type": "text", + "content": "[33] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models. In Advances in Neural Information Processing Systems, 2023. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "ing transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[35] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "type": "text", + "content": "[36] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. In Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "type": "text", + "content": "[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 413 + ], + "type": "text", + "content": "[39] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In Proceedings of the European Conference on Computer Vision, 2024. 1, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[41] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In Proceedings of the International Conference on Machine Learning, 2023. 1, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "text", + "content": "[42] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Advances in Neural Information Processing Systems, 2021. 1, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "text", + "content": "[43] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 6, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "text", + "content": "[44] Jihan Yang, Shusheng Yang, Anjali Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces. arXiv preprint arXiv:2412.14171, 2024. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[45] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In Proceedings of the International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8532" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 642 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[47] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: collision events for video representation and reasoning. In Proceedings of the International Conference on Learning Representations, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[48] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In Advances in Neural Information Processing Systems, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "text", + "content": "[49] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "type": "text", + "content": "[50] Zhou Yu, Lixiang Zheng, Zhou Zhao, Fei Wu, Jianping Fan, Kui Ren, and Jun Yu. Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 407 + ], + "type": "text", + "content": "[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 407, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 451 + ], + "type": "text", + "content": "[52] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 452, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 452, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 294, + 496 + ], + "type": "text", + "content": "[53] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 1, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "text", + "content": "[54] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 553, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 294, + 597 + ], + "type": "text", + "content": "[55] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 597, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 597, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 597, + 294, + 642 + ], + "type": "text", + "content": "[56] Orr Zohar, Xiaohan Wang, Yonatan Bitton, Idan Szpektor, and Serena Yeung-Levy. Video-star: Self-training enables video instruction tuning with any supervision. arXiv preprint arXiv:2407.06189, 2024. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8533" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_content_list.json b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..07f0f987c8299e5baddeb1899e51b6405af933c9 --- /dev/null +++ b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_content_list.json @@ -0,0 +1,1539 @@ +[ + { + "type": "text", + "text": "Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling", + "text_level": 1, + "bbox": [ + 138, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nannan Li \nBoston University \nnnli@bu.edu", + "bbox": [ + 236, + 203, + 383, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kevin J. Shih NVIDIA", + "bbox": [ + 441, + 203, + 550, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bryan A. Plummer \nBoston University \nbplum@bu.edu", + "bbox": [ + 607, + 203, + 761, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given an isolated garment image in a canonical product view and a separate image of a person, the virtual try-on task aims to generate a new image of the person wearing the target garment. Prior virtual try-on works face two major challenges in achieving this goal: a) the paired (human, garment) training data has limited availability; b) generating textures on the human that perfectly match that of the prompted garment is difficult, often resulting in distorted text and faded textures. Our work explores ways to tackle these issues through both synthetic data as well as model refinement. We introduce a garment extraction model that generates (human, synthetic garment) pairs from a single image of a clothed individual. The synthetic pairs can then be used to augment the training of virtual try-on. We also propose an Error-Aware Refinement-based Schrödinger Bridge (EARSB) that surgically targets localized generation errors for correcting the output of a base virtual try-on model. To identify likely errors, we propose a weakly-supervised error classifier that localizes regions for refinement, subsequently augmenting the Schrödinger Bridge's noise schedule with its confidence heatmap. Experiments on VITON-HD and DressCode-Upper demonstrate that our synthetic data augmentation enhances the performance of prior work, while EARSB improves the overall image quality. In user studies, our model is preferred by the users in an average of $59\\%$ of cases. Code is available at this link.", + "bbox": [ + 86, + 325, + 485, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 767, + 220, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virtual try-on aims to generate a photorealistic image of a target person wearing a prompted product-view garment [22, 36, 39]. It allows users to visualize how garments would fit and appear on their bodies without the need for physical trials. While recent methods have made significant strides in this field [18, 29, 35, 36], noticeable artifacts such as text distortion and faded textures persist in gener", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce5ff823a899f3e784948fde1ca27fdd30f51870986ddaa5e941c699096ee03a.jpg", + "image_caption": [ + "Figure 1. Example of our proposed Error-Aware Refinement Schrödinger Bridge (EARSB). EARSB can refine the artifacts (marked by bounding boxes) in an initial image generated by an existing try-on model. The initial image is generated by [18] in the top row and by [29] in the bottom row. + Syn. Data in the last column strengthens the refinement with the proposed synthetic data augmentation in training." + ], + "image_footnote": [], + "bbox": [ + 516, + 287, + 903, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ated images. For example, as illustrated in the second row of Fig. 1, the logo and the text on the t-shirt noticeably fade away in the initial image generated by a prior try-on model [29]. These imperfections stem from two primary challenges in virtual try-on: limited data availability and the complexity of accurate garment texture deformation. To address these issues, we propose a two-pronged approach: augmenting training data through cost-effective synthetic data generation, and surgically targeting known generation artifacts using our proposed Error-Aware Refinement-based Schrödinger Bridge (EARSB).", + "bbox": [ + 511, + 638, + 908, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At a minimum, the training data of virtual try-on requires paired (human, product-view garment) images. The product-view garment image is a canonical, front-facing view of the clothing with a clean background. A substantial amount of data is needed to capture the combinatorial space comprising all possible human poses, skin tones, viewing", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21238", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "angles, and their respective physical interactions with fabric textures, shapes, letterings, and other material properties. Unfortunately, these images are generally available only on copyright-protected product webpages and, therefore, are not readily available for use. To mitigate this issue, we propose to augment training with synthetic data generated from the easier symmetric human-to-garment task, wherein we train a garment-extraction model to extract a canonical product-view garment image from an image of a clothed person. This will allow us to create synthetic paired training data from unpaired datasets [11, 24, 34]. Our results demonstrate that incorporating the more readily available synthetic training pairs can improve image generation quality in the virtual try-on task.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to addressing the data scarcity issue, we aim to construct a refinement model that can make localized adjustments to a weaker model's generation results. Our approach draws inspiration from classical boosting approaches where every model in a cascade of models targets the shortcomings of the preceding models. We are interested in a targeted refinement approach for two main reasons: it allows a training objective that is focused solely on fixing specific errors, and potentially saves computation when initial predictions are sufficiently good.", + "bbox": [ + 89, + 304, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Two components are necessary to achieve such a pipeline: a classifier for identifying localized generation errors, and a refinement model that can re-synthesize content specifically in these localized regions. We found that an effective Weakly-Supervised error Classifier (WSC) can be constructed with just a few hours of manual labeling of generation errors. Another benefit of this approach is that it can be easily tailored for the errors of a specific model that produces images with artifacts. The resulting WSC will produce an error map highlighting low-quality regions. Subsequently, we adopt an Image-to-Image Schrödinger Bridge $(\\mathrm{I}^2\\mathrm{SB})$ [23] to learn the refinement of these regions in the generated images. While typical diffusion models map from noise to data, $\\mathrm{I}^2\\mathrm{SB}$ constructs a Schrödinger Bridge (SB) that allows us to map from data to data, or in our setup, generations with artifacts to ground truth images. In addition, we introduce an adaptive noise schedule to direct the SB process to focus on the localized errors by incorporating the classifier's prediction error into the noise schedule, which we describe in more detail in Sec. 4.1. As shown in the first row of Fig. 1, our refinement SB model (i.e., EARSB) corrects the distorted text in the initially generated image.", + "bbox": [ + 89, + 458, + 482, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our paper are:", + "bbox": [ + 109, + 792, + 344, + 808 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce (human, synthetic garment) pairs as an augmentation in the training of virtual try-on task. The synthetic garment is obtained from our human-to-garment model, which can generate product-view garment images from human images.", + "- We introduce a spatially adaptive Schrödinger Bridge" + ], + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model (EARSB) to refine the outputs of a base virtual try-on model. Our formulation incorporates a spatially varying diffusion noise schedule, with noise proportional to the degree of refinement we wish to perform locally. We find this to yield better results than the baseline Schrödinger Bridge framework.", + "bbox": [ + 526, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments on two datasets (VITON-HD [20] and DressCode-Upper [25]) show that EARSB enhances the quality of the images generated by prior work, and is preferred by the users in $59\\%$ cases on average.", + "bbox": [ + 513, + 181, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 256, + 653, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training with Synthetic Data. The addition of synthetic data is often an effective means of improving downstream task performance when it is difficult to amass real data at the necessary scale. This has been demonstrated in the domains of image generation [17, 30] and image editing [5, 33]. Careful applications can also be used to ameliorate dataset imbalance issues, as shown in [9]. Other works, such as [1], use self-synthesized data to provide negative guidance for the diffusion model. Our incorporation of synthetic data in the virtual try-on task tackles a specific sub-problem in the broader image editing domain and is similar in spirit to [5, 33]. Specifically, we aim to synthesize paired training data that satisfies the stringent requirements of virtual try-on paired training data – a canonical product-view garment image paired with an example of it being worn. Images of people in clothing are readily available, but it is difficult to obtain a product-view image of the exact clothing they are wearing. To address this, our work tackles the human-to-garment problem, which aims to extract the clothing from a person's photo and project it to the canonical product view, making it roughly symmetric to the virtual try-on task.", + "bbox": [ + 511, + 281, + 903, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Virtual Try-On. There has been a shift from earlier GAN-based framework [14, 20, 22, 29, 35] to diffusion-based methods [7, 18] in the virtual try-on literature. Diffusion models fit an SDE process mapping from the image distribution to the noise distribution, and tend to be easier to train than GAN-based approaches due to the simplicity of the L2 denoising loss [10, 12, 31]. At inference, the diffusion model denoises a random Gaussian noise distribution to a human-readable image via multiple sampling steps. [7, 36] propose parameter-efficient approaches that concatenate the human image and the garment images along the spatial dimension such that the self-attention layer in the denoising UNet can achieve texture transfer without extra parameters. In [18], the authors introduce additional cross-attention layers to learn the semantic correspondences between the garment and the human image. The methods in [3, 21, 27] align different embedding spaces in the attention module to achieve flexible clothing editing after try-on, such as style change or graphics insertion. In contrast to prior work that samples from random noise, we build upon", + "bbox": [ + 511, + 599, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21239", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc735dfcbe2cb0fed7de7d183d8f99ba89111d1b3a2ebaaf015381c33912280f.jpg", + "image_caption": [ + "(a) Human-to-Garment model." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 496, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a866f0be82048b99dcd5de44ca60bff340b1cfbd5b09d33289accdc2096578d8.jpg", + "image_caption": [ + "(b) Examples in H2G-UH and H2G-FH.", + "Figure 2. (a) Our human-to-garment model, which is explained in Sec. 3.1 (b) Examples of the constructed (human, synthetic garment) pairs in Sec. 3.1." + ], + "image_footnote": [], + "bbox": [ + 93, + 270, + 500, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "recent advances in Schrödinger bridges, notably [23], to directly sample from an initial image generated by prior try-on models. Our work is similar in spirit to [37], which initializes the noisy image with a GAN-generated image and small amounts of random noise. However, our work explores varying the local noise schedule based on the error level at a given location.", + "bbox": [ + 89, + 507, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Augmented Training with Synthetic Data", + "text_level": 1, + "bbox": [ + 89, + 630, + 459, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Virtual Try-On Task Definition. Let $(x_0, C)$ be the (human image, product-view of worn garment) pair in virtual try-on training. We will refer to $(x_0, C)$ as paired data. Let $\\bar{x}_0$ be a masked version of $x_0$ in which the worn garment corresponding to $C$ is masked out. We can then set up a learning task in which we aim to fit the following function: $F(\\bar{x}_0, C, \\phi; \\theta) \\to x_0$ , where $\\phi$ corresponds to other conditionals such as pose representations from DensePose [13].", + "bbox": [ + 89, + 657, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Acquiring high-quality pairs $(x_0, C)$ at scale is challenging due to copyright and brand protection, but acquiring images of just humans $(x_0)$ at scale is considerably more feasible [11, 21, 24, 34]. This observation motivates proposed human-to-garment process to extract a synthetic canonical view image $\\hat{C}$ from $x_0$ . We can then augment our virtual try-on training with $(x_0, \\hat{C})$ pair, requiring only single human images. In the following, Sec. 3.1 explains the archi", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tecture of our human-to-garment model, Sec. 3.2 discusses how we use this model to construct the synthetic dataset, and Sec. 3.3 describes how the synthetic data is used to augment the virtual try-on training.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Human-to-Garment Model", + "text_level": 1, + "bbox": [ + 511, + 162, + 758, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While virtual try-on requires generating skin and deforming the product-view garment to accommodate diverse postures, the human-to-garment task simply aims to map the clothing item to its canonical view. To achieve this, we use existing paired (human, garment) data (e.g., VITON-HD [20]) to train our human-to-garment model. As illustrated in Fig. 2a, we first segment and extract the clothing on the person map and then feed the clothing item to a generator that synthesizes its canonical view. The generator is based on the UNet model proposed in [14], which uses a flow-like mechanism for warping latent features in an optical-flow-like manner. The generator was trained using a combined L1 reconstruction and adversarial loss.", + "bbox": [ + 511, + 185, + 906, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Constructing Synthetic H2G-UH and H2G-FH", + "text_level": 1, + "bbox": [ + 511, + 392, + 905, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic images $\\hat{C}$ produced from our models necessarily contain generation errors. We use the following criteria to filter for high-quality synthetic data: a) The single human image $x_0$ has a clean background (low pixel variance in the non-human region); b) $x_0$ is frontal view (classified by its DensePose representation [13]); c) the reconstruction error (LPIPS distance) is small when reconstructing the human image $x_0$ in a try-on model using the $(x_0, \\hat{C})$ pair (e.g., [20, 29]). Under these criteria, we select human images from DeepFashion2 [24] and UPT [34], eventually creating 12,730 synthetic pairs of upper-body human images (referred to as H2G-UH) and 8,939 pairs of full-body human images (referred to as H2G-FH). Examples of the synthetic pairs are shown in Fig. 2b.", + "bbox": [ + 511, + 414, + 905, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Augmented Virtual Try-on Training", + "text_level": 1, + "bbox": [ + 511, + 638, + 826, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further prevent distribution leakage of incorporating synthetic data, we explore two means of limiting the effect of the real-synthetic domain gap: (a) two training stages involving pretraining the try-on model using synthetic pairs, and then finetuning on real pairs [19]; (b) training simultaneously on real and synthetic data, but conditioning the try-on model on a real/synthetic flag, similar to [16]. We found empirically that the second augmentation performs slightly better than the first (See Sec. 5.1).", + "bbox": [ + 511, + 660, + 905, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Error-Aware Refinement Schrödinger Bridge", + "text_level": 1, + "bbox": [ + 511, + 811, + 905, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Apart from the synthetic data augmentation from Sec. 3, our second approach to enhancing existing try-on methods is a refinement pipeline with two steps. First, given some", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21240", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1a97d47f3f18a1c439164979b9d506dff344fd07bb800734353798f309804de5.jpg", + "image_caption": [ + "Figure 3. The diffusion process in our refinement-based EARSB. We first preprocess the input image, then use a base try-on model that takes the masked human image $\\bar{x}_0$ , its pose representation $P$ , and its garment $C$ as input to generate an initial human image $x_1$ . $x_1$ is fed to our weakly-supervised classifier (WSC) to obtain the error map $M$ (see Sec. 4.1). This map reweights the noise distribution $\\epsilon$ to $\\epsilon^r$ in $\\mathrm{I}^2\\mathrm{SB}$ diffusion and refines $x_1$ that has generation errors to the ground truth image $x_0$ (see Sec. 4.2)." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 906, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "base model $F_{base}(\\bar{x}_0, C, \\phi) \\to x_1$ where $F_{base}(.)$ can be any pretrained GAN or diffusion-based approach to virtual try-on, $x_1$ closely approximates the true real human image $x_0$ with some generation artifacts. To automatically identify the artifacts in $x_1$ , we construct a weakly-supervised error classifier $\\mathrm{WSC}(x_1, C) \\to M$ as in Fig. 3, where $M$ is a confidence map predicting a heatmap for likely generation errors. Then, our second step performs the final refinement by fitting a Schrödinger bridge based on $\\mathrm{I}^2\\mathrm{SB}$ [23] via the following mapping: $F_{EARSB}(x_1, C, M, \\phi; \\theta) \\to x_0$ .", + "bbox": [ + 88, + 457, + 480, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The approach is weakly inspired by boosting methods in that we wish to fit a targeted refinement model that is trained specifically on the generation errors of an existing model. The refinement goal applies to the general setting where we want to refine a flawed image output, though we focus on virtual try-on for this work. Thus, as illustrated in Fig. 3, the training of EARSB includes three steps:", + "bbox": [ + 89, + 609, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Pre-process the images in the training set and feed them to existing try-on models to get the initial images $x_{1}$ .", + "2. Obtain the error maps $M$ on the initial images $x_{1}$ using our WSC (Sec. 4.1).", + "3. Use $M$ to adjust the noise schedule in $\\mathrm{I}^2\\mathrm{SB}$ [23] and train the noise prediction model in EARSB following Eq. (9) (Sec. 4.2)." + ], + "bbox": [ + 109, + 717, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the first step simply requires running an off-the-shelf virtual try-on model to obtain an initial image, our discussion in the next section will begin by describing the second step- obtaining the error map.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Obtaining the Error Map", + "text_level": 1, + "bbox": [ + 511, + 457, + 746, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start by obtaining the error map $M$ that highlights the corrupted or incorrect area of the initial image $x_{1}$ using our proposed Weakly-Supervised error Classifier (WSC).", + "bbox": [ + 511, + 479, + 906, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classifier Architecture. As shown in the green dotted box of Fig. 3, our WSC has two encoders to match the image features of $x_{1}$ and $C$ with cross attention to predict a sigmoid-activated error map.", + "bbox": [ + 511, + 525, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Data Annotation. In practice, it is labor-intensive to fully annotate all the initial images for where the generated artifacts are located. To mitigate this issue, we used a few hours to hand-label a small portion of the initial try-on images in the training set at the patch level, using bounding boxes for poorly generated regions.", + "bbox": [ + 511, + 585, + 905, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weakly-Supervised Training. Let $x_0, x_1^u, x_1^l$ be the real human image, the unlabeled initial image, and the labeled initial image with bounding boxes annotating artifacts. Our WSC loss terms are defined as:", + "bbox": [ + 511, + 676, + 905, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m g} = - \\log \\left(\\mathrm {W S C} \\left(x _ {1} ^ {u}, C\\right) _ {\\max }\\right) + \\log \\left(1 - \\mathrm {W S C} \\left(x _ {0}, C\\right) _ {\\max }\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 737, + 924, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {p a t} = - \\log \\left(\\operatorname {W S C} \\left(x _ {1} ^ {l}, C\\right) \\odot B _ {b o x}\\right) \\\\ - \\log \\left(1 - \\operatorname {W S C} \\left(x _ {1} ^ {l}, C\\right) \\odot \\left(1 - B _ {\\text {b o x}}\\right)\\right) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 757, + 903, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{img}$ is the image-level loss and $\\mathcal{L}_{pat}$ is the patch-level loss. In $\\mathcal{L}_{img}$ , $\\mathrm{WSC}(\\cdot)$ is the output error map and $\\mathrm{WSC}(\\cdot)_{\\mathrm{max}}$ denotes the spatially max-pooled score in the error map. In $\\mathcal{L}_{pat}$ , $B_{box}$ is the spatial binary mask for the annotated regions, thereby maximizing and minimizing the scores for regions within and outside of the annotated boxes", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21241", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "respectively. Our final loss is: $\\mathcal{L}_{\\mathrm{WSC}} = \\mathcal{L}_{ins} + \\mathcal{L}_{pat}$", + "bbox": [ + 89, + 90, + 437, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The trained WSC will predict an error map $M$ for the initial image $x_{1}$ , which is then used to adjust the noise schedule in the diffusion process described in the next section.", + "bbox": [ + 89, + 108, + 482, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Error-Map-Reweighted SB Formulation", + "text_level": 1, + "bbox": [ + 89, + 169, + 434, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve the refinement goal, our diffusion process extends Schrödinger bridges as formulated in $\\mathrm{I}^2\\mathrm{SB}$ [23], where we incrementally add noise to the initial image $x_{1}$ , and then remove the noise to approximate the refined image $x_0$ . However, without additional information, a naively trained $\\mathrm{I}^2\\mathrm{SB}$ model must implicitly learn what to refine and what to retain. Our formulation aims to explicitly incorporate prior knowledge of localized generation errors via the error map $M$ into the Schrödinger process by using $M$ to locally scale the noise schedule for the Schrödinger process.", + "bbox": [ + 89, + 193, + 483, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our choice of locally scaling the noise schedule is based on several observations. We want the model to directly copy pixels over to $x_0$ for correctly generated regions in $x_1$ - it would be nice to avoid training the model to add and remove noise from these regions. In contrast, erroneous regions in the initial try-on images, especially more noticeable ones, include generation errors that share little to no structural similarity to the target. These errors include examples such as deformed limbs, and distorted textures/fabrics, which may need more added noise to prevent the model from conditioning too strongly on the original pattern.", + "bbox": [ + 89, + 345, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As such, we construct our refinement model $F_{EARSB}(x_1,C,M,P;\\theta) \\to x_0$ , where the model is conditioned on the canonical view garment $C$ , the error map $M$ , and the pose representation $P$ . Fig. 3 shows our Weakly-Supervised Classifier (WSC) first locates the errors in the error map $M$ , then $M$ reweights the noise schedule of the I²SB stochastic process to assign a higher volume of noise to the low-quality region \"rebel\" so the model can focus on refining it.", + "bbox": [ + 89, + 515, + 483, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Error-Map-Reweighted Diffusion Process. Following I²SB [23], our diffusion Schrödinger bridge maps from the initial image $x_{1}$ to the ground truth image $x_{0}$ . It fits to the following stochastic process:", + "bbox": [ + 89, + 652, + 483, + 714 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t} = \\mu_ {t} \\left(x _ {0}, x _ {1}\\right) + \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon \\\\ \\mu_ {t} = \\frac {\\bar {\\sigma} _ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} x _ {0} + \\frac {\\sigma_ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} x _ {1}, \\Sigma_ {t} = \\frac {\\bar {\\sigma} _ {t} ^ {2} \\sigma_ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} \\cdot I, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 717, + 482, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_t^2 = \\int_0^t\\beta_\\tau d\\tau$ , $\\bar{\\sigma}_t^2 = \\int_t^1\\beta_\\tau d\\tau$ and $\\beta_{\\tau}$ is a symmetrical noise schedule. $\\epsilon \\sim \\mathcal{N}(0,I)$ is random Gaussian noise. The above equation stochastically adds noise and then removes it between $x_{1}$ and $x_0$ .", + "bbox": [ + 89, + 775, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We extend $\\mathrm{I}^2\\mathrm{SB}$ such that the noise schedule can vary spatially based on the error map $M$ (obtained from WSC in Sec. 4.1). Good regions will be assigned less noise (i.e., smaller variance) in the diffusion process, while poor qual", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ity regions will be assigned more:", + "bbox": [ + 511, + 90, + 740, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\mu_ {t} \\left(x _ {0}, x _ {1}\\right) + \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon^ {r}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 109, + 903, + 127 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {r} = M \\cdot \\epsilon , M = \\operatorname {W S C} \\left(x _ {1}, C\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 130, + 903, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_t$ is the same as Eq. (2) and $\\epsilon^r$ is the adaptive noise.", + "bbox": [ + 511, + 148, + 903, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling Process. The initial image $x_{1}$ is iteratively refined to $x_{0}$ via a denoising/sampling process, where a model predicts the noise distribution at each time step. In contrast to prior soft-attention-based UNets [18, 26, 36, 37], our denoising model uses cloth-flow-learning UNet for more precise garment deformation [14]. It accepts the garment $C$ , the error map $M$ , the pose representation $P$ , and the noisy image $x_{t}$ as inputs and predicts the error-adapted noise $\\epsilon_{\\theta}^{r}(\\cdot ;t)$ , where $(\\cdot ;t)$ omits the inputs $M, P, x_{t}, C$ . See Supp. B for the detailed model architecture. With the predicted noise $\\epsilon_{\\theta}^{r}(\\cdot ;t)$ , we define our sampling process:", + "bbox": [ + 511, + 172, + 906, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {0} = x _ {t} - \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon_ {\\theta} ^ {r} (M, P, x _ {t}, C; t) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 343, + 903, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - \\Delta t} = \\hat {\\mu} _ {t - \\Delta t} \\left(\\hat {x} _ {0}, x _ {t}\\right) + M \\cdot \\sqrt {\\hat {\\Sigma} _ {t}} \\cdot \\epsilon \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 364, + 903, + 390 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mu} _ {t - \\Delta t} = \\frac {\\sigma_ {t - \\Delta t} ^ {2}}{\\sigma_ {t} ^ {2}} \\hat {x} _ {0} + \\frac {\\sigma_ {t} ^ {2} - \\sigma_ {t - \\Delta t} ^ {2}}{\\sigma_ {t} ^ {2}} x _ {t} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 392, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Sigma} _ {t} = \\frac {\\sigma_ {t - \\Delta t} ^ {2} \\left(\\sigma_ {t} ^ {2} - \\sigma_ {t - \\Delta t} ^ {2}\\right)}{\\sigma_ {t} ^ {2}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 428, + 903, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta t > 0$ and it is the sampling interval. Starting from $t = 1$ , the process iteratively refines the initial human image $x_{1}$ based on the error map $M$ . When $M$ is all ones in Eq. (5), our model reverts to the $\\mathrm{I}^2\\mathrm{SB}$ formulation. When $M$ is all zeros (i.e., no error), $x_{1}$ is believed to be perfect $x_{1}$ does not need to be refined in the sampling process.", + "bbox": [ + 511, + 464, + 905, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training objective of our model is the mean squared error between the predicted noise $\\epsilon_{\\theta}^{r}$ and the reweighted Gaussian noise $\\epsilon^{r}$", + "bbox": [ + 511, + 561, + 905, + 607 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E A R S B}} = \\mathbb {E} _ {t \\sim U (0, 1)} | | \\epsilon_ {\\theta} ^ {r} (M, P, x _ {t}, C; t) - \\epsilon^ {r} | | ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 611, + 903, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Further Improvements via Classifier Guidance and Expert Denoisers", + "text_level": 1, + "bbox": [ + 511, + 659, + 903, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whereas prior work used an object category classifier to guide the sampling process [10], our WSC guidance gives a direction toward the real data distribution. Chung et al. [8] shows we can estimate the guidance score $\\nabla_{x_t}\\log p(y|x_t)$ using the denoised clean image $\\hat{x}_0$ : $\\nabla_{x_t}\\log p(y|x_t)\\simeq \\nabla_{x_t}\\log p(y|\\hat{x}_0)$ , where $y$ is the fake/real label. Since the label for real data is 0 in WSC, the classifier guidance is:", + "bbox": [ + 511, + 702, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mu} _ {t - \\Delta t} \\leftarrow \\hat {\\mu} _ {t - \\Delta t} + M \\cdot \\hat {\\Sigma} _ {t} \\cdot \\nabla_ {x _ {t}} \\log p (\\mathbf {0} | \\hat {x} _ {0}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 811, + 903, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(\\mathbf{0}|\\hat{x}_0) = 1 - \\mathrm{WSC}(\\hat{x}_0,C)$ .", + "bbox": [ + 511, + 830, + 751, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [2], a trained EARSB model is split into two models, each having an expert denoiser that is fine-tuned on denoising ranges $t \\in [0, 0.5]$ and $t \\in [0.5, 1]$ , respectively.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21242", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/21610391de687195a77c7934e86820f65779137986bcd97e443d5a16e8c7066d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UnpairedVITON-HDDressCode-Upper
PairedUnpairedPaired
FID↓KID↓FID↓KID↓SSIM↑LPIPS↓FID↓KID↓FID↓KID↓SSIM↑
(a) GAN-Based
HR-VTON [20]10.750.288.460.260.9010.07515.260.3911.760.320.947
SD-VTON [29]9.050.126.470.090.9070.07014.730.3210.990.240.947
GP-VTON [35]8.610.865.530.070.9130.06426.191.7123.661.590.816
EARSB (ours)8.420.075.250.050.9180.05910.890.137.150.130.961
EARSB +H2G-UH/FH (ours)8.260.065.140.040.9190.05810.700.117.050.110.965
(b) SD-Based
LaDI-VTON [26]8.950.126.050.080.9020.07114.880.3911.610.320.939
CatVTON [7]8.870.085.490.070.9150.05911.910.217.660.100.950
CAT-DM [37]8.550.105.980.070.9080.06712.910.298.580.160.948
IDM-VTON [6]8.590.115.510.090.9020.06111.090.166.790.120.956
TPD [36]8.230.064.860.040.9170.057-
StableVITON [18]8.200.075.160.050.9170.057-
EARSB(SD) +H2G-UH/FH (ours)8.040.064.900.030.9250.05310.410.096.760.080.968
", + "bbox": [ + 106, + 88, + 885, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 425, + 223, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use VITON-HD [20], DressCode-Upper [25], and our synthetic H2G-UH and H2G-FH for training. They include 11,647, 13,564, 12,730, 8,939 training images, respectively. For synthetic data augmentation, we combine VITON-HD with our H2G-UH since they both include mostly upper-body human images. DressCode-Upper is combined with H2G-FH as both consist of full-body human photos. For evaluation, VITON-HD contains 2,032 (human, garment) test pairs and DressCode-Upper has 1,800 test pairs and include both paired and unpaired settings. In the paired setting, the input garment image and the garment in the human image are the same item. Conversely, the unpaired setting uses a different garment image.", + "bbox": [ + 88, + 452, + 482, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We use Structural Similarity Index Measure (SSIM) [32], Frechet Inception Distance (FID) [15], Kernel Inception Distance (KID) [4], and Learned Perceptual Image Patch Similarity (LPIPS) [38] to evaluate image quality. All the compared methods use the same image size $512 \\times 512$ and padding when computing the above metrics.", + "bbox": [ + 89, + 652, + 482, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental setup. We compare EARSB with GAN-based methods HR-VTON [20], SD-VTON [29] and GPVTON [35], as well as Stable Diffusion (SD) based methods including CAT-DM [37], StableVITON [18], TPD [36], IDM-VTON [6] and CatVTON [7]. Unless otherwise specified, all diffusion models use 25 sampling steps.", + "bbox": [ + 89, + 746, + 482, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report results of multiple variants of our approach. EARSB uses GAN-based GP-VTON [35] to generate the initial image that was trained without synthetic data augmentation. EARSB+H2G-UH/FH trains with either H2G-", + "bbox": [ + 89, + 839, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3860d04fbf53713fb67a5e86e8d15a85dde7f91ee347191df8e7e87f7ca4461d.jpg", + "table_caption": [ + "Table 1. Virtual try-on results on VITON-HD [20] and DressCode-Upper [25] for (a) GAN-based and (b) diffusion-based models using 25 sampling steps. KID is multiplied by 100. We find our EARSB approach outperforms prior work on average. See Sec. 5.1 for discussion." + ], + "table_footnote": [], + "table_body": "
MethodsGP-VTON EARSBStableVITON EARSB
Consistency42%58%38%62%
Fidelity39%61%45%55%
", + "bbox": [ + 522, + 424, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. User studies on VITON-HD. Our EARSB method is preferred in an average of $59\\%$ cases.", + "bbox": [ + 511, + 489, + 903, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UH or H2G-FH. We add the upper-body synthetic subset H2G-UH for the upper-body-human dataset VITON-HD, and the full-body synthetic H2G-FH when on DressCodeUpper. EARSB(SD)+H2G-UH/FH uses the diffusion-based CatVTON [7] to generate the initial image.", + "bbox": [ + 511, + 520, + 903, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Results", + "text_level": 1, + "bbox": [ + 513, + 606, + 606, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 compares our approach with those from prior work. We find our full model variants EARSB+H2G-UH/FH and EARSB(SD)+H2G-UH/FH boosts performance over the GAN and SD-based methods, respectively. The last two rows of Tab. 1(a) show that incorporating our synthetic training pairs provide a consistent boost in performance on both datasets. Comparing the last rows of Tab. 1(a) and Tab. 1(b) we observe that using the diffusion model to generate the initial image gives better performance in EARSB(SD)+H2G-UH/FH, but is more costly.", + "bbox": [ + 509, + 628, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study. We asked Amazon MTurk workers to evaluate the texture consistency and image fidelity of synthesized images, comparing our model against GP-VTON and StableVITON. We randomly selected 100 pairs from VITON-HD to evaluate on, assigning at least 3 workers per image. Study results in Tab. 2 report our method is preferred at least $10\\%$ more than the GAN-based GP-VTON and the SD-based StableVITON (59% overall).", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21243", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e65258728eb019c37b994906975e059994c4ee72d62edc40b57937f329a02678.jpg", + "image_caption": [ + "Unpaired FID", + "Figure 4. Results on VITON-HD at 5, 10, 25, 50, and 100 sampling steps. Our method consistently improves our baseline starting model GP-VTON (black, dotted line), making it competitive with StableVITON (especially at under 50 sampling steps). Legend is shared for all." + ], + "image_footnote": [], + "bbox": [ + 96, + 114, + 423, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83f869cd286951474efcd491f9ea0d87dbd916af9470854750c7e64028f3938e.jpg", + "image_caption": [ + "Paired FID" + ], + "image_footnote": [], + "bbox": [ + 423, + 114, + 648, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d476eb027df389ed00b1301f4df47f7193d704f7adfa5eef0fbd85bd07908158.jpg", + "image_caption": [ + "Paired SSIM" + ], + "image_footnote": [], + "bbox": [ + 653, + 114, + 898, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57fde6822d1365bbef2bd0a29506996da1076e8ab9d9a559ef2e97bc6f88ef45.jpg", + "image_caption": [ + "Figure 5. Visualizations on VITON-HD (top row) and DressCode (bottom row). Our EARSB+H2G-UH and EARSB(SD)+H2G-UH better recover the intricate textures in the garment." + ], + "image_footnote": [], + "bbox": [ + 93, + 280, + 903, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trade-Off Between Sampling Efficiency and Image Quality. Diffusion-generated images often show degraded quality with fewer sampling steps. In EARSB, the spatially adaptive noise schedule can preserve correct clothing textures in the initial image with a low noise level and only fix the erroneous parts, potentially resulting in less image quality degradation with fewer steps. In Fig. 4, while other SD-based methods have a sharp performance drop with decreasing sampling steps, EARSB and EARSB+H2G-UH show consistent performance across different sampling steps, demonstrating a better trade-off between image quality and computational efficiency.", + "bbox": [ + 88, + 532, + 482, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 5 gives examples of generated images from different approaches. The top row is from VITON-HD dataset and the bottom row is from DressCodeUpper. The third images in the two rows are GAN-generated results. We see that our EARSB+H2G-UH/FH in the last column improves the low-quality textures from the GAN-generated images, which are the distorted graphics in the center. Additional visualizations are in Supp. G.", + "bbox": [ + 89, + 715, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablations", + "text_level": 1, + "bbox": [ + 89, + 847, + 200, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic Pairs Augmentation. Tab. 3 incorporates H2GUH into the training of StableVITON [18] and CAT-DM", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c23edea79ed0df040465b5474751b33b6968b60f0efe4c1ec839249b897cc5d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Data Aug.Unpaired FID↓Paired FID↓SSIM↑LPIPS↓
CAT-DM [37]None8.565.900.9110.067
CAT-DMH2G-UH8.365.670.9130.063
StableVITON [18]None8.255.150.9170.056
StableVITONH2G-UH8.175.040.9190.054
EARSB(SD)H2G-UH8.044.900.9250.053
", + "bbox": [ + 513, + 531, + 915, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparing the effect of our H2G-UH data augmentation approach on VITON-HD. We bold the best overall results and underline the best results for a single model with and without H2G-UH. Each method uses the number of sampling steps from their paper: 2 for CAT-DM, 50 for StableVITON, and 25 for our own EARSB. We find H2G-UH consistently boosts performance.", + "bbox": [ + 511, + 664, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[37] on the VITON-HD dataset to validate the effectiveness of synthetic pairs on enhancing existing diffusion methods. We observe that training with our synthetic H2G-UH is effective in improving most metrics.", + "bbox": [ + 511, + 777, + 903, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 explores different ways of incorporating the synthetic data H2G-UH during the training of EARSB. We find that using a randomly warped version of the clothing $(W(\\mathrm{H}2\\mathrm{G}-\\mathrm{U}\\mathrm{H}))$ hinders performance, demonstrating the", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21244", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3462a3cbfbd624b48fe01a109cf5958c83939e776755b70e74c393c199587c4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
None8.425.250.9180.059
W(H2G-UH)8.685.440.9090.063
plain H2G-UH9.646.520.9020.073
pre. H2G-UH8.355.180.9180.059
H2G-UH8.265.140.9190.058
", + "bbox": [ + 125, + 88, + 444, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c012b101f541b93b1fc390eba36a999b70f43a247fe4df2ccedccbd0bf678f10.jpg", + "table_caption": [ + "Table 4. Ablations of our H2G-UH augmentation on VITON-HD. Specifically, pre. H2G-UH is pretrained using the synthetic pairs and finetuned on real data, $W(\\mathbf{H2G - UH})$ replaces the synthetic garment in each pair with a randomly warped version of the clothing cropped from the real human image, plain H2G-UH is trained using the mixed distribution of the real and synthetic pairs without the augmentation label identifying them, and H2G-UH uses the mixed data with the identifying label. See Sec. 5.2 for discussion." + ], + "table_footnote": [], + "table_body": "
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
Inpainting9.266.330.9090.068
EARSB (w.o. M)9.216.270.9120.061
EARSB (rand(M))9.136.550.9020.071
EARSB (w.o. CG)8.485.320.9180.059
EARSB (full)8.425.250.9180.059
", + "bbox": [ + 109, + 348, + 457, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparing noise scheduling strategies on VITON-HD. We include a simple inpainting baseline, not using the error map during training (w.o. $M$ ), a random error map $(\\mathrm{rand}(M))$ , and removing classifier guidance (w.o. CG). These results demonstrate the importance of using a meaningful error map.", + "bbox": [ + 89, + 481, + 482, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "importance of the synthetic product-view. Additionally, the poor plain H2G-UH results indicate the presence of a synthetic-real domain gap when using H2G-UH. Inspired by [28], one way to address this issue is by pretraining on the synthetic data and finetuning on real samples (see pre. H2G-UH). However, we find it most effective to condition on a synthetic data indicator while training on mixed data.", + "bbox": [ + 89, + 580, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Error-Aware Noise Schedule. Tab. 5 explores the importance of the error-aware noise schedule described in Sec. 4.2, where the error map adapts the noise distribution according to the quality of the image patches in the initial image $x_{1}$ . This adaptive approach contrasts with a uniform Gaussian noise application across all locations, which would reduce our model to $\\mathrm{I}^2\\mathrm{SB}$ . We consider three baselines: Inpainting regenerates rather than refines erroneous regions (where the mean confidence $M$ of an image patch containing an error is greater than 0.5), w.o. $M$ removes the error map during training, and $\\mathbf{rand}(M)$ indicates a random error map. As shown in Tab. 5, EARSB outperforms all these baselines, underscoring the importance of a meaningful error map in precisely locating and enhancing", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3449d3c1a0689566e3e505a7aeb64ce903b67d6f61e178ddb0a9e52b1303762.jpg", + "image_caption": [ + "Figure 6. The precision-recall curve for retrieving annotated errors at the pixel level, comparing our WSC to two fully unsupervised baselines (UC, CC). WSC performs best at retrieving generation artifacts at a nominal labeling cost." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 887, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing targeted regions. Additionally, the slight performance degradation observed when removing the classifier guidance (EARSB(w.o. CG)) suggests that the error map employed in our classifier guidance also contributes to overall image quality improvement. Collectively, these findings highlight the crucial role of our adaptive noise schedule in achieving superior results.", + "bbox": [ + 511, + 309, + 903, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weakly-Supervised error Classifier (WSC). Our weakly supervised classifier from Sec. 4.1 highlights low-quality regions in the initial image $x_{1}$ with only a few hours of labeling. To validate its effectiveness, we train two ablations of our WSC: the Unsupervised Classifier (UC) that only uses image labels (i.e., fake or real), and the Fake/Real Composite Classifier (CC). CC uses both image fake/real labels as well as fake region-level labels which are created by compositing real image patches and fake image patches. The compositing is a fully automatic alternative to manual labeling that provides patch-level labels. We annotated 100 images in the test set to validate their effectiveness. Fig. 6 shows the pixel-level precision-recall curve for retrieving annotated artifact pixels within the bounding boxes using the classifiers' confidence maps. It is clear that weak supervision remains an incredibly cost-effective approach.", + "bbox": [ + 511, + 416, + 906, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper addresses two shortcomings of prior work on virtual try-on. First, we address the limited data availability by introducing a human-to-garment model that generates (human, synthetic garment) pairs from a single image of a clothed individual. Second, we propose a refinement model EARSB that surgically targets localized generation errors from the output of a prior model. EARSB improves the low-quality region of an initially generated image based on a spatially-varying noise schedule that targets known artifacts. Experiments on two benchmark datasets demonstrate that our synthetic data augmentation improves the performance of prior work and that EARSB enhances the overall image quality.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21245", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This material is based upon work supported, in part, by DARPA under agreement number HR00112020054 and the National Science Foundation under Grant No. DBI-2134696. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the supporting agencies.", + "bbox": [ + 89, + 90, + 485, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 213, + 187, + 229 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sina Alemohammad, Ahmed Imtiaz Humayun, Shruti Agarwal, John Collomosse, and Richard Baraniuk. Self-improving diffusion models with synthetic data. arXiv preprint arXiv:2408.16333, 2024. 2", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5", + "[3] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In CVPR, 2023. 2", + "[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 6", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2", + "[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. In ECCV, 2024. 6", + "[7] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 2, 6", + "[8] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 5", + "[9] Damien Dablain, Bartosz Krawczyk, and Nitesh V Chawla. Deepsmote: Fusing deep learning and smote for imbalanced data. IEEE Transactions on Neural Networks and Learning Systems, 34(9):6390-6404, 2022. 2", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2, 5", + "[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, 2022. 2, 3", + "[12] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 2" + ], + "bbox": [ + 93, + 238, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In CVPR, 2018. 3", + "[14] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In CVPR, 2019. 2, 3, 5", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6", + "[16] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In ICML, 2020. 3", + "[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2", + "[18] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. StableVITON: Learning semantic correspondence with latent diffusion model for virtual try-on. In CVPR, 2024. 1, 2, 5, 6, 7", + "[19] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 3", + "[20] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In ECCV, 2022. 2, 3, 6", + "[21] Nannan Li, Qing Liu, Krishna Kumar Singh, Yilin Wang, Jianming Zhang, Bryan A Plummer, and Zhe Lin. UniHuman: A unified model for editing human images in the wild. In CVPR, 2024. 2, 3", + "[22] Zhi Li, Pengfei Wei, Xiang Yin, Zejun Ma, and Alex C Kot. Virtual try-on with pose-garment keypoints guided inpainting. In ICCV, 2023. 1, 2", + "[23] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I2sb: image-to-image schrodinger bridge. In ICML, 2023. 2, 3, 4, 5", + "[24] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3", + "[25] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In CVPR, 2022. 2, 6", + "[26] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. LaDIVTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On. In ACM Multimedia, 2023. 5, 6", + "[27] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In CVPR, 2024. 2", + "[28] Maan Qraitem, Kate Saenko, and Bryan A. Plummer. From fake to real: Pretraining on balanced synthetic images to pre" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21246", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vent spurious correlations in image recognition. In The European Conference on Computer Vision (ECCV), 2024. 8", + "[29] Sang-Heon Shim, Jiwoo Chung, and Jae-Pil Heo. Towards squeezing-averse virtual try-on via sequential deformation. In AAAI, 2024. 1, 2, 3, 6", + "[30] C Shivashankar and Shane Miller. Semantic data augmentation with generative models. In CVPR, 2023. 2", + "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2", + "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6", + "[33] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024. 2", + "[34] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive gan. NeurIPS, 2021. 2, 3", + "[35] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. GPVTON: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In CVPR, 2023. 1, 2, 6", + "[36] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In CVPR, 2024. 1, 2, 5, 6", + "[37] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In CVPR, 2024. 3, 5, 6, 7", + "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[39] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. TryOnDiffusion: A tale of two unets. In CVPR, 2023. 1" + ], + "bbox": [ + 91, + 92, + 482, + 672 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21247", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_model.json b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1da5fbf3f8bb45f68392f5fa63f0beced405fb64 --- /dev/null +++ b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.13, + 0.861, + 0.178 + ], + "angle": 0, + "content": "Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.204, + 0.384, + 0.257 + ], + "angle": 0, + "content": "Nannan Li \nBoston University \nnnli@bu.edu" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.204, + 0.551, + 0.238 + ], + "angle": 0, + "content": "Kevin J. Shih NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.204, + 0.762, + 0.257 + ], + "angle": 0, + "content": "Bryan A. Plummer \nBoston University \nbplum@bu.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.326, + 0.486, + 0.734 + ], + "angle": 0, + "content": "Given an isolated garment image in a canonical product view and a separate image of a person, the virtual try-on task aims to generate a new image of the person wearing the target garment. Prior virtual try-on works face two major challenges in achieving this goal: a) the paired (human, garment) training data has limited availability; b) generating textures on the human that perfectly match that of the prompted garment is difficult, often resulting in distorted text and faded textures. Our work explores ways to tackle these issues through both synthetic data as well as model refinement. We introduce a garment extraction model that generates (human, synthetic garment) pairs from a single image of a clothed individual. The synthetic pairs can then be used to augment the training of virtual try-on. We also propose an Error-Aware Refinement-based Schrödinger Bridge (EARSB) that surgically targets localized generation errors for correcting the output of a base virtual try-on model. To identify likely errors, we propose a weakly-supervised error classifier that localizes regions for refinement, subsequently augmenting the Schrödinger Bridge's noise schedule with its confidence heatmap. Experiments on VITON-HD and DressCode-Upper demonstrate that our synthetic data augmentation enhances the performance of prior work, while EARSB improves the overall image quality. In user studies, our model is preferred by the users in an average of \\(59\\%\\) of cases. Code is available at this link." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.768, + 0.222, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Virtual try-on aims to generate a photorealistic image of a target person wearing a prompted product-view garment [22, 36, 39]. It allows users to visualize how garments would fit and appear on their bodies without the need for physical trials. While recent methods have made significant strides in this field [18, 29, 35, 36], noticeable artifacts such as text distortion and faded textures persist in gener" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.52, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Example of our proposed Error-Aware Refinement Schrödinger Bridge (EARSB). EARSB can refine the artifacts (marked by bounding boxes) in an initial image generated by an existing try-on model. The initial image is generated by [18] in the top row and by [29] in the bottom row. + Syn. Data in the last column strengthens the refinement with the proposed synthetic data augmentation in training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.806 + ], + "angle": 0, + "content": "ated images. For example, as illustrated in the second row of Fig. 1, the logo and the text on the t-shirt noticeably fade away in the initial image generated by a prior try-on model [29]. These imperfections stem from two primary challenges in virtual try-on: limited data availability and the complexity of accurate garment texture deformation. To address these issues, we propose a two-pronged approach: augmenting training data through cost-effective synthetic data generation, and surgically targeting known generation artifacts using our proposed Error-Aware Refinement-based Schrödinger Bridge (EARSB)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "At a minimum, the training data of virtual try-on requires paired (human, product-view garment) images. The product-view garment image is a canonical, front-facing view of the clothing with a clean background. A substantial amount of data is needed to capture the combinatorial space comprising all possible human poses, skin tones, viewing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21238" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "angles, and their respective physical interactions with fabric textures, shapes, letterings, and other material properties. Unfortunately, these images are generally available only on copyright-protected product webpages and, therefore, are not readily available for use. To mitigate this issue, we propose to augment training with synthetic data generated from the easier symmetric human-to-garment task, wherein we train a garment-extraction model to extract a canonical product-view garment image from an image of a clothed person. This will allow us to create synthetic paired training data from unpaired datasets [11, 24, 34]. Our results demonstrate that incorporating the more readily available synthetic training pairs can improve image generation quality in the virtual try-on task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.482, + 0.457 + ], + "angle": 0, + "content": "In addition to addressing the data scarcity issue, we aim to construct a refinement model that can make localized adjustments to a weaker model's generation results. Our approach draws inspiration from classical boosting approaches where every model in a cascade of models targets the shortcomings of the preceding models. We are interested in a targeted refinement approach for two main reasons: it allows a training objective that is focused solely on fixing specific errors, and potentially saves computation when initial predictions are sufficiently good." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Two components are necessary to achieve such a pipeline: a classifier for identifying localized generation errors, and a refinement model that can re-synthesize content specifically in these localized regions. We found that an effective Weakly-Supervised error Classifier (WSC) can be constructed with just a few hours of manual labeling of generation errors. Another benefit of this approach is that it can be easily tailored for the errors of a specific model that produces images with artifacts. The resulting WSC will produce an error map highlighting low-quality regions. Subsequently, we adopt an Image-to-Image Schrödinger Bridge \\((\\mathrm{I}^2\\mathrm{SB})\\) [23] to learn the refinement of these regions in the generated images. While typical diffusion models map from noise to data, \\(\\mathrm{I}^2\\mathrm{SB}\\) constructs a Schrödinger Bridge (SB) that allows us to map from data to data, or in our setup, generations with artifacts to ground truth images. In addition, we introduce an adaptive noise schedule to direct the SB process to focus on the localized errors by incorporating the classifier's prediction error into the noise schedule, which we describe in more detail in Sec. 4.1. As shown in the first row of Fig. 1, our refinement SB model (i.e., EARSB) corrects the distorted text in the initially generated image." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.794, + 0.345, + 0.809 + ], + "angle": 0, + "content": "The contributions of our paper are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We introduce (human, synthetic garment) pairs as an augmentation in the training of virtual try-on task. The synthetic garment is obtained from our human-to-garment model, which can generate product-view garment images from human images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We introduce a spatially adaptive Schrödinger Bridge" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "model (EARSB) to refine the outputs of a base virtual try-on model. Our formulation incorporates a spatially varying diffusion noise schedule, with noise proportional to the degree of refinement we wish to perform locally. We find this to yield better results than the baseline Schrödinger Bridge framework." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.243 + ], + "angle": 0, + "content": "- Extensive experiments on two datasets (VITON-HD [20] and DressCode-Upper [25]) show that EARSB enhances the quality of the images generated by prior work, and is preferred by the users in \\(59\\%\\) cases on average." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.257, + 0.655, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.905, + 0.599 + ], + "angle": 0, + "content": "Training with Synthetic Data. The addition of synthetic data is often an effective means of improving downstream task performance when it is difficult to amass real data at the necessary scale. This has been demonstrated in the domains of image generation [17, 30] and image editing [5, 33]. Careful applications can also be used to ameliorate dataset imbalance issues, as shown in [9]. Other works, such as [1], use self-synthesized data to provide negative guidance for the diffusion model. Our incorporation of synthetic data in the virtual try-on task tackles a specific sub-problem in the broader image editing domain and is similar in spirit to [5, 33]. Specifically, we aim to synthesize paired training data that satisfies the stringent requirements of virtual try-on paired training data – a canonical product-view garment image paired with an example of it being worn. Images of people in clothing are readily available, but it is difficult to obtain a product-view image of the exact clothing they are wearing. To address this, our work tackles the human-to-garment problem, which aims to extract the clothing from a person's photo and project it to the canonical product view, making it roughly symmetric to the virtual try-on task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.6, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Virtual Try-On. There has been a shift from earlier GAN-based framework [14, 20, 22, 29, 35] to diffusion-based methods [7, 18] in the virtual try-on literature. Diffusion models fit an SDE process mapping from the image distribution to the noise distribution, and tend to be easier to train than GAN-based approaches due to the simplicity of the L2 denoising loss [10, 12, 31]. At inference, the diffusion model denoises a random Gaussian noise distribution to a human-readable image via multiple sampling steps. [7, 36] propose parameter-efficient approaches that concatenate the human image and the garment images along the spatial dimension such that the self-attention layer in the denoising UNet can achieve texture transfer without extra parameters. In [18], the authors introduce additional cross-attention layers to learn the semantic correspondences between the garment and the human image. The methods in [3, 21, 27] align different embedding spaces in the attention module to achieve flexible clothing editing after try-on, such as style change or graphics insertion. In contrast to prior work that samples from random noise, we build upon" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21239" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.498, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.256, + 0.378, + 0.268 + ], + "angle": 0, + "content": "(a) Human-to-Garment model." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.271, + 0.5, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.425, + 0.404, + 0.437 + ], + "angle": 0, + "content": "(b) Examples in H2G-UH and H2G-FH." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Figure 2. (a) Our human-to-garment model, which is explained in Sec. 3.1 (b) Examples of the constructed (human, synthetic garment) pairs in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.614 + ], + "angle": 0, + "content": "recent advances in Schrödinger bridges, notably [23], to directly sample from an initial image generated by prior try-on models. Our work is similar in spirit to [37], which initializes the noisy image with a GAN-generated image and small amounts of random noise. However, our work explores varying the local noise schedule based on the error level at a given location." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.46, + 0.649 + ], + "angle": 0, + "content": "3. Augmented Training with Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Virtual Try-On Task Definition. Let \\((x_0, C)\\) be the (human image, product-view of worn garment) pair in virtual try-on training. We will refer to \\((x_0, C)\\) as paired data. Let \\(\\bar{x}_0\\) be a masked version of \\(x_0\\) in which the worn garment corresponding to \\(C\\) is masked out. We can then set up a learning task in which we aim to fit the following function: \\(F(\\bar{x}_0, C, \\phi; \\theta) \\to x_0\\), where \\(\\phi\\) corresponds to other conditionals such as pose representations from DensePose [13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Acquiring high-quality pairs \\((x_0, C)\\) at scale is challenging due to copyright and brand protection, but acquiring images of just humans \\((x_0)\\) at scale is considerably more feasible [11, 21, 24, 34]. This observation motivates proposed human-to-garment process to extract a synthetic canonical view image \\(\\hat{C}\\) from \\(x_0\\). We can then augment our virtual try-on training with \\((x_0, \\hat{C})\\) pair, requiring only single human images. In the following, Sec. 3.1 explains the archi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "tecture of our human-to-garment model, Sec. 3.2 discusses how we use this model to construct the synthetic dataset, and Sec. 3.3 describes how the synthetic data is used to augment the virtual try-on training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.163, + 0.759, + 0.178 + ], + "angle": 0, + "content": "3.1. Human-to-Garment Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.907, + 0.381 + ], + "angle": 0, + "content": "While virtual try-on requires generating skin and deforming the product-view garment to accommodate diverse postures, the human-to-garment task simply aims to map the clothing item to its canonical view. To achieve this, we use existing paired (human, garment) data (e.g., VITON-HD [20]) to train our human-to-garment model. As illustrated in Fig. 2a, we first segment and extract the clothing on the person map and then feed the clothing item to a generator that synthesizes its canonical view. The generator is based on the UNet model proposed in [14], which uses a flow-like mechanism for warping latent features in an optical-flow-like manner. The generator was trained using a combined L1 reconstruction and adversarial loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.906, + 0.41 + ], + "angle": 0, + "content": "3.2. Constructing Synthetic H2G-UH and H2G-FH" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.906, + 0.628 + ], + "angle": 0, + "content": "Synthetic images \\(\\hat{C}\\) produced from our models necessarily contain generation errors. We use the following criteria to filter for high-quality synthetic data: a) The single human image \\(x_0\\) has a clean background (low pixel variance in the non-human region); b) \\(x_0\\) is frontal view (classified by its DensePose representation [13]); c) the reconstruction error (LPIPS distance) is small when reconstructing the human image \\(x_0\\) in a try-on model using the \\((x_0, \\hat{C})\\) pair (e.g., [20, 29]). Under these criteria, we select human images from DeepFashion2 [24] and UPT [34], eventually creating 12,730 synthetic pairs of upper-body human images (referred to as H2G-UH) and 8,939 pairs of full-body human images (referred to as H2G-FH). Examples of the synthetic pairs are shown in Fig. 2b." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.827, + 0.656 + ], + "angle": 0, + "content": "3.3. Augmented Virtual Try-on Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.797 + ], + "angle": 0, + "content": "To further prevent distribution leakage of incorporating synthetic data, we explore two means of limiting the effect of the real-synthetic domain gap: (a) two training stages involving pretraining the try-on model using synthetic pairs, and then finetuning on real pairs [19]; (b) training simultaneously on real and synthetic data, but conditioning the try-on model on a real/synthetic flag, similar to [16]. We found empirically that the second augmentation performs slightly better than the first (See Sec. 5.1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.812, + 0.906, + 0.848 + ], + "angle": 0, + "content": "4. Error-Aware Refinement Schrödinger Bridge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Apart from the synthetic data augmentation from Sec. 3, our second approach to enhancing existing try-on methods is a refinement pipeline with two steps. First, given some" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21240" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.907, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.389, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Figure 3. The diffusion process in our refinement-based EARSB. We first preprocess the input image, then use a base try-on model that takes the masked human image \\(\\bar{x}_0\\), its pose representation \\(P\\), and its garment \\(C\\) as input to generate an initial human image \\(x_1\\). \\(x_1\\) is fed to our weakly-supervised classifier (WSC) to obtain the error map \\(M\\) (see Sec. 4.1). This map reweights the noise distribution \\(\\epsilon\\) to \\(\\epsilon^r\\) in \\(\\mathrm{I}^2\\mathrm{SB}\\) diffusion and refines \\(x_1\\) that has generation errors to the ground truth image \\(x_0\\) (see Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.458, + 0.482, + 0.61 + ], + "angle": 0, + "content": "base model \\( F_{base}(\\bar{x}_0, C, \\phi) \\to x_1 \\) where \\( F_{base}(.) \\) can be any pretrained GAN or diffusion-based approach to virtual try-on, \\( x_1 \\) closely approximates the true real human image \\( x_0 \\) with some generation artifacts. To automatically identify the artifacts in \\( x_1 \\), we construct a weakly-supervised error classifier \\( \\mathrm{WSC}(x_1, C) \\to M \\) as in Fig. 3, where \\( M \\) is a confidence map predicting a heatmap for likely generation errors. Then, our second step performs the final refinement by fitting a Schrödinger bridge based on \\( \\mathrm{I}^2\\mathrm{SB} \\) [23] via the following mapping: \\( F_{EARSB}(x_1, C, M, \\phi; \\theta) \\to x_0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.717 + ], + "angle": 0, + "content": "The approach is weakly inspired by boosting methods in that we wish to fit a targeted refinement model that is trained specifically on the generation errors of an existing model. The refinement goal applies to the general setting where we want to refine a flawed image output, though we focus on virtual try-on for this work. Thus, as illustrated in Fig. 3, the training of EARSB includes three steps:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.718, + 0.483, + 0.763 + ], + "angle": 0, + "content": "1. Pre-process the images in the training set and feed them to existing try-on models to get the initial images \\( x_{1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.763, + 0.483, + 0.793 + ], + "angle": 0, + "content": "2. Obtain the error maps \\( M \\) on the initial images \\( x_{1} \\) using our WSC (Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.794, + 0.483, + 0.839 + ], + "angle": 0, + "content": "3. Use \\(M\\) to adjust the noise schedule in \\(\\mathrm{I}^2\\mathrm{SB}\\) [23] and train the noise prediction model in EARSB following Eq. (9) (Sec. 4.2)." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.718, + 0.483, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As the first step simply requires running an off-the-shelf virtual try-on model to obtain an initial image, our discussion in the next section will begin by describing the second step- obtaining the error map." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.458, + 0.747, + 0.475 + ], + "angle": 0, + "content": "4.1. Obtaining the Error Map" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.526 + ], + "angle": 0, + "content": "We start by obtaining the error map \\( M \\) that highlights the corrupted or incorrect area of the initial image \\( x_{1} \\) using our proposed Weakly-Supervised error Classifier (WSC)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.906, + 0.586 + ], + "angle": 0, + "content": "Classifier Architecture. As shown in the green dotted box of Fig. 3, our WSC has two encoders to match the image features of \\( x_{1} \\) and \\( C \\) with cross attention to predict a sigmoid-activated error map." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.677 + ], + "angle": 0, + "content": "Training Data Annotation. In practice, it is labor-intensive to fully annotate all the initial images for where the generated artifacts are located. To mitigate this issue, we used a few hours to hand-label a small portion of the initial try-on images in the training set at the patch level, using bounding boxes for poorly generated regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Weakly-Supervised Training. Let \\( x_0, x_1^u, x_1^l \\) be the real human image, the unlabeled initial image, and the labeled initial image with bounding boxes annotating artifacts. Our WSC loss terms are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.738, + 0.926, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m g} = - \\log \\left(\\mathrm {W S C} \\left(x _ {1} ^ {u}, C\\right) _ {\\max }\\right) + \\log \\left(1 - \\mathrm {W S C} \\left(x _ {0}, C\\right) _ {\\max }\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.758, + 0.905, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {p a t} = - \\log \\left(\\operatorname {W S C} \\left(x _ {1} ^ {l}, C\\right) \\odot B _ {b o x}\\right) \\\\ - \\log \\left(1 - \\operatorname {W S C} \\left(x _ {1} ^ {l}, C\\right) \\odot \\left(1 - B _ {\\text {b o x}}\\right)\\right) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{img}\\) is the image-level loss and \\(\\mathcal{L}_{pat}\\) is the patch-level loss. In \\(\\mathcal{L}_{img}\\), \\(\\mathrm{WSC}(\\cdot)\\) is the output error map and \\(\\mathrm{WSC}(\\cdot)_{\\mathrm{max}}\\) denotes the spatially max-pooled score in the error map. In \\(\\mathcal{L}_{pat}\\), \\(B_{box}\\) is the spatial binary mask for the annotated regions, thereby maximizing and minimizing the scores for regions within and outside of the annotated boxes" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21241" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.438, + 0.108 + ], + "angle": 0, + "content": "respectively. Our final loss is: \\(\\mathcal{L}_{\\mathrm{WSC}} = \\mathcal{L}_{ins} + \\mathcal{L}_{pat}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.483, + 0.155 + ], + "angle": 0, + "content": "The trained WSC will predict an error map \\( M \\) for the initial image \\( x_{1} \\), which is then used to adjust the noise schedule in the diffusion process described in the next section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.17, + 0.436, + 0.187 + ], + "angle": 0, + "content": "4.2. Error-Map-Reweighted SB Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.194, + 0.484, + 0.346 + ], + "angle": 0, + "content": "To achieve the refinement goal, our diffusion process extends Schrödinger bridges as formulated in \\(\\mathrm{I}^2\\mathrm{SB}\\) [23], where we incrementally add noise to the initial image \\(x_{1}\\), and then remove the noise to approximate the refined image \\(x_0\\). However, without additional information, a naively trained \\(\\mathrm{I}^2\\mathrm{SB}\\) model must implicitly learn what to refine and what to retain. Our formulation aims to explicitly incorporate prior knowledge of localized generation errors via the error map \\(M\\) into the Schrödinger process by using \\(M\\) to locally scale the noise schedule for the Schrödinger process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.484, + 0.514 + ], + "angle": 0, + "content": "Our choice of locally scaling the noise schedule is based on several observations. We want the model to directly copy pixels over to \\( x_0 \\) for correctly generated regions in \\( x_1 \\) - it would be nice to avoid training the model to add and remove noise from these regions. In contrast, erroneous regions in the initial try-on images, especially more noticeable ones, include generation errors that share little to no structural similarity to the target. These errors include examples such as deformed limbs, and distorted textures/fabrics, which may need more added noise to prevent the model from conditioning too strongly on the original pattern." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.652 + ], + "angle": 0, + "content": "As such, we construct our refinement model \\( F_{EARSB}(x_1,C,M,P;\\theta) \\to x_0 \\), where the model is conditioned on the canonical view garment \\( C \\), the error map \\( M \\), and the pose representation \\( P \\). Fig. 3 shows our Weakly-Supervised Classifier (WSC) first locates the errors in the error map \\( M \\), then \\( M \\) reweights the noise schedule of the I²SB stochastic process to assign a higher volume of noise to the low-quality region \"rebel\" so the model can focus on refining it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.715 + ], + "angle": 0, + "content": "Error-Map-Reweighted Diffusion Process. Following I²SB [23], our diffusion Schrödinger bridge maps from the initial image \\( x_{1} \\) to the ground truth image \\( x_{0} \\). It fits to the following stochastic process:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.718, + 0.483, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t} = \\mu_ {t} \\left(x _ {0}, x _ {1}\\right) + \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon \\\\ \\mu_ {t} = \\frac {\\bar {\\sigma} _ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} x _ {0} + \\frac {\\sigma_ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} x _ {1}, \\Sigma_ {t} = \\frac {\\bar {\\sigma} _ {t} ^ {2} \\sigma_ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} \\cdot I, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.839 + ], + "angle": 0, + "content": "where \\(\\sigma_t^2 = \\int_0^t\\beta_\\tau d\\tau\\), \\(\\bar{\\sigma}_t^2 = \\int_t^1\\beta_\\tau d\\tau\\) and \\(\\beta_{\\tau}\\) is a symmetrical noise schedule. \\(\\epsilon \\sim \\mathcal{N}(0,I)\\) is random Gaussian noise. The above equation stochastically adds noise and then removes it between \\(x_{1}\\) and \\(x_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We extend \\(\\mathrm{I}^2\\mathrm{SB}\\) such that the noise schedule can vary spatially based on the error map \\(M\\) (obtained from WSC in Sec. 4.1). Good regions will be assigned less noise (i.e., smaller variance) in the diffusion process, while poor qual" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.741, + 0.107 + ], + "angle": 0, + "content": "ity regions will be assigned more:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.111, + 0.905, + 0.128 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\mu_ {t} \\left(x _ {0}, x _ {1}\\right) + \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon^ {r}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.131, + 0.905, + 0.147 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {r} = M \\cdot \\epsilon , M = \\operatorname {W S C} \\left(x _ {1}, C\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.15, + 0.905, + 0.165 + ], + "angle": 0, + "content": "where \\(\\mu_t\\) is the same as Eq. (2) and \\(\\epsilon^r\\) is the adaptive noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.173, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Sampling Process. The initial image \\( x_{1} \\) is iteratively refined to \\( x_{0} \\) via a denoising/sampling process, where a model predicts the noise distribution at each time step. In contrast to prior soft-attention-based UNets [18, 26, 36, 37], our denoising model uses cloth-flow-learning UNet for more precise garment deformation [14]. It accepts the garment \\( C \\), the error map \\( M \\), the pose representation \\( P \\), and the noisy image \\( x_{t} \\) as inputs and predicts the error-adapted noise \\( \\epsilon_{\\theta}^{r}(\\cdot ;t) \\), where \\( (\\cdot ;t) \\) omits the inputs \\( M, P, x_{t}, C \\). See Supp. B for the detailed model architecture. With the predicted noise \\( \\epsilon_{\\theta}^{r}(\\cdot ;t) \\), we define our sampling process:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.344, + 0.905, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {0} = x _ {t} - \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon_ {\\theta} ^ {r} (M, P, x _ {t}, C; t) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.366, + 0.905, + 0.391 + ], + "angle": 0, + "content": "\\[\nx _ {t - \\Delta t} = \\hat {\\mu} _ {t - \\Delta t} \\left(\\hat {x} _ {0}, x _ {t}\\right) + M \\cdot \\sqrt {\\hat {\\Sigma} _ {t}} \\cdot \\epsilon \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.393, + 0.905, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mu} _ {t - \\Delta t} = \\frac {\\sigma_ {t - \\Delta t} ^ {2}}{\\sigma_ {t} ^ {2}} \\hat {x} _ {0} + \\frac {\\sigma_ {t} ^ {2} - \\sigma_ {t - \\Delta t} ^ {2}}{\\sigma_ {t} ^ {2}} x _ {t} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.429, + 0.905, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Sigma} _ {t} = \\frac {\\sigma_ {t - \\Delta t} ^ {2} \\left(\\sigma_ {t} ^ {2} - \\sigma_ {t - \\Delta t} ^ {2}\\right)}{\\sigma_ {t} ^ {2}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.465, + 0.906, + 0.556 + ], + "angle": 0, + "content": "where \\(\\Delta t > 0\\) and it is the sampling interval. Starting from \\(t = 1\\), the process iteratively refines the initial human image \\(x_{1}\\) based on the error map \\(M\\). When \\(M\\) is all ones in Eq. (5), our model reverts to the \\(\\mathrm{I}^2\\mathrm{SB}\\) formulation. When \\(M\\) is all zeros (i.e., no error), \\(x_{1}\\) is believed to be perfect \\(x_{1}\\) does not need to be refined in the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.906, + 0.608 + ], + "angle": 0, + "content": "The training objective of our model is the mean squared error between the predicted noise \\(\\epsilon_{\\theta}^{r}\\) and the reweighted Gaussian noise \\(\\epsilon^{r}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.612, + 0.905, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E A R S B}} = \\mathbb {E} _ {t \\sim U (0, 1)} | | \\epsilon_ {\\theta} ^ {r} (M, P, x _ {t}, C; t) - \\epsilon^ {r} | | ^ {2} \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.691 + ], + "angle": 0, + "content": "4.2.1. Further Improvements via Classifier Guidance and Expert Denoisers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Whereas prior work used an object category classifier to guide the sampling process [10], our WSC guidance gives a direction toward the real data distribution. Chung et al. [8] shows we can estimate the guidance score \\(\\nabla_{x_t}\\log p(y|x_t)\\) using the denoised clean image \\(\\hat{x}_0\\): \\(\\nabla_{x_t}\\log p(y|x_t)\\simeq \\nabla_{x_t}\\log p(y|\\hat{x}_0)\\), where \\(y\\) is the fake/real label. Since the label for real data is 0 in WSC, the classifier guidance is:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.812, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mu} _ {t - \\Delta t} \\leftarrow \\hat {\\mu} _ {t - \\Delta t} + M \\cdot \\hat {\\Sigma} _ {t} \\cdot \\nabla_ {x _ {t}} \\log p (\\mathbf {0} | \\hat {x} _ {0}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.832, + 0.753, + 0.849 + ], + "angle": 0, + "content": "where \\(p(\\mathbf{0}|\\hat{x}_0) = 1 - \\mathrm{WSC}(\\hat{x}_0,C)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Following [2], a trained EARSB model is split into two models, each having an expert denoiser that is fine-tuned on denoising ranges \\( t \\in [0, 0.5] \\) and \\( t \\in [0.5, 1] \\), respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21242" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.887, + 0.379 + ], + "angle": 0, + "content": "
UnpairedVITON-HDDressCode-Upper
PairedUnpairedPaired
FID↓KID↓FID↓KID↓SSIM↑LPIPS↓FID↓KID↓FID↓KID↓SSIM↑
(a) GAN-Based
HR-VTON [20]10.750.288.460.260.9010.07515.260.3911.760.320.947
SD-VTON [29]9.050.126.470.090.9070.07014.730.3210.990.240.947
GP-VTON [35]8.610.865.530.070.9130.06426.191.7123.661.590.816
EARSB (ours)8.420.075.250.050.9180.05910.890.137.150.130.961
EARSB +H2G-UH/FH (ours)8.260.065.140.040.9190.05810.700.117.050.110.965
(b) SD-Based
LaDI-VTON [26]8.950.126.050.080.9020.07114.880.3911.610.320.939
CatVTON [7]8.870.085.490.070.9150.05911.910.217.660.100.950
CAT-DM [37]8.550.105.980.070.9080.06712.910.298.580.160.948
IDM-VTON [6]8.590.115.510.090.9020.06111.090.166.790.120.956
TPD [36]8.230.064.860.040.9170.057-
StableVITON [18]8.200.075.160.050.9170.057-
EARSB(SD) +H2G-UH/FH (ours)8.040.064.900.030.9250.05310.410.096.760.080.968
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.385, + 0.904, + 0.414 + ], + "angle": 0, + "content": "Table 1. Virtual try-on results on VITON-HD [20] and DressCode-Upper [25] for (a) GAN-based and (b) diffusion-based models using 25 sampling steps. KID is multiplied by 100. We find our EARSB approach outperforms prior work on average. See Sec. 5.1 for discussion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.224, + 0.442 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.453, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Datasets. We use VITON-HD [20], DressCode-Upper [25], and our synthetic H2G-UH and H2G-FH for training. They include 11,647, 13,564, 12,730, 8,939 training images, respectively. For synthetic data augmentation, we combine VITON-HD with our H2G-UH since they both include mostly upper-body human images. DressCode-Upper is combined with H2G-FH as both consist of full-body human photos. For evaluation, VITON-HD contains 2,032 (human, garment) test pairs and DressCode-Upper has 1,800 test pairs and include both paired and unpaired settings. In the paired setting, the input garment image and the garment in the human image are the same item. Conversely, the unpaired setting uses a different garment image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Metrics. We use Structural Similarity Index Measure (SSIM) [32], Frechet Inception Distance (FID) [15], Kernel Inception Distance (KID) [4], and Learned Perceptual Image Patch Similarity (LPIPS) [38] to evaluate image quality. All the compared methods use the same image size \\( 512 \\times 512 \\) and padding when computing the above metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Experimental setup. We compare EARSB with GAN-based methods HR-VTON [20], SD-VTON [29] and GPVTON [35], as well as Stable Diffusion (SD) based methods including CAT-DM [37], StableVITON [18], TPD [36], IDM-VTON [6] and CatVTON [7]. Unless otherwise specified, all diffusion models use 25 sampling steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We report results of multiple variants of our approach. EARSB uses GAN-based GP-VTON [35] to generate the initial image that was trained without synthetic data augmentation. EARSB+H2G-UH/FH trains with either H2G-" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.425, + 0.892, + 0.486 + ], + "angle": 0, + "content": "
MethodsGP-VTON EARSBStableVITON EARSB
Consistency42%58%38%62%
Fidelity39%61%45%55%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.49, + 0.904, + 0.518 + ], + "angle": 0, + "content": "Table 2. User studies on VITON-HD. Our EARSB method is preferred in an average of \\(59\\%\\) cases." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.905, + 0.597 + ], + "angle": 0, + "content": "UH or H2G-FH. We add the upper-body synthetic subset H2G-UH for the upper-body-human dataset VITON-HD, and the full-body synthetic H2G-FH when on DressCodeUpper. EARSB(SD)+H2G-UH/FH uses the diffusion-based CatVTON [7] to generate the initial image." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.607, + 0.607, + 0.621 + ], + "angle": 0, + "content": "5.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Tab. 1 compares our approach with those from prior work. We find our full model variants EARSB+H2G-UH/FH and EARSB(SD)+H2G-UH/FH boosts performance over the GAN and SD-based methods, respectively. The last two rows of Tab. 1(a) show that incorporating our synthetic training pairs provide a consistent boost in performance on both datasets. Comparing the last rows of Tab. 1(a) and Tab. 1(b) we observe that using the diffusion model to generate the initial image gives better performance in EARSB(SD)+H2G-UH/FH, but is more costly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "User Study. We asked Amazon MTurk workers to evaluate the texture consistency and image fidelity of synthesized images, comparing our model against GP-VTON and StableVITON. We randomly selected 100 pairs from VITON-HD to evaluate on, assigning at least 3 workers per image. Study results in Tab. 2 report our method is preferred at least \\(10\\%\\) more than the GAN-based GP-VTON and the SD-based StableVITON (59% overall)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21243" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.092, + 0.282, + 0.107 + ], + "angle": 0, + "content": "Unpaired FID" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.115, + 0.424, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.095, + 0.592, + 0.108 + ], + "angle": 0, + "content": "Paired FID" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.115, + 0.649, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.095, + 0.85, + 0.108 + ], + "angle": 0, + "content": "Paired SSIM" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.115, + 0.9, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 4. Results on VITON-HD at 5, 10, 25, 50, and 100 sampling steps. Our method consistently improves our baseline starting model GP-VTON (black, dotted line), making it competitive with StableVITON (especially at under 50 sampling steps). Legend is shared for all." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.281, + 0.905, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.498, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Figure 5. Visualizations on VITON-HD (top row) and DressCode (bottom row). Our EARSB+H2G-UH and EARSB(SD)+H2G-UH better recover the intricate textures in the garment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.534, + 0.483, + 0.715 + ], + "angle": 0, + "content": "Trade-Off Between Sampling Efficiency and Image Quality. Diffusion-generated images often show degraded quality with fewer sampling steps. In EARSB, the spatially adaptive noise schedule can preserve correct clothing textures in the initial image with a low noise level and only fix the erroneous parts, potentially resulting in less image quality degradation with fewer steps. In Fig. 4, while other SD-based methods have a sharp performance drop with decreasing sampling steps, EARSB and EARSB+H2G-UH show consistent performance across different sampling steps, demonstrating a better trade-off between image quality and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 5 gives examples of generated images from different approaches. The top row is from VITON-HD dataset and the bottom row is from DressCodeUpper. The third images in the two rows are GAN-generated results. We see that our EARSB+H2G-UH/FH in the last column improves the low-quality textures from the GAN-generated images, which are the distorted graphics in the center. Additional visualizations are in Supp. G." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.202, + 0.863 + ], + "angle": 0, + "content": "5.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Synthetic Pairs Augmentation. Tab. 3 incorporates H2GUH into the training of StableVITON [18] and CAT-DM" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.532, + 0.916, + 0.661 + ], + "angle": 0, + "content": "
Data Aug.Unpaired FID↓Paired FID↓SSIM↑LPIPS↓
CAT-DM [37]None8.565.900.9110.067
CAT-DMH2G-UH8.365.670.9130.063
StableVITON [18]None8.255.150.9170.056
StableVITONH2G-UH8.175.040.9190.054
EARSB(SD)H2G-UH8.044.900.9250.053
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Table 3. Comparing the effect of our H2G-UH data augmentation approach on VITON-HD. We bold the best overall results and underline the best results for a single model with and without H2G-UH. Each method uses the number of sampling steps from their paper: 2 for CAT-DM, 50 for StableVITON, and 25 for our own EARSB. We find H2G-UH consistently boosts performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.905, + 0.839 + ], + "angle": 0, + "content": "[37] on the VITON-HD dataset to validate the effectiveness of synthetic pairs on enhancing existing diffusion methods. We observe that training with our synthetic H2G-UH is effective in improving most metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Tab. 4 explores different ways of incorporating the synthetic data H2G-UH during the training of EARSB. We find that using a randomly warped version of the clothing \\((W(\\mathrm{H}2\\mathrm{G}-\\mathrm{U}\\mathrm{H}))\\) hinders performance, demonstrating the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21244" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.445, + 0.217 + ], + "angle": 0, + "content": "
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
None8.425.250.9180.059
W(H2G-UH)8.685.440.9090.063
plain H2G-UH9.646.520.9020.073
pre. H2G-UH8.355.180.9180.059
H2G-UH8.265.140.9190.058
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.221, + 0.482, + 0.332 + ], + "angle": 0, + "content": "Table 4. Ablations of our H2G-UH augmentation on VITON-HD. Specifically, pre. H2G-UH is pretrained using the synthetic pairs and finetuned on real data, \\( W(\\mathbf{H2G - UH}) \\) replaces the synthetic garment in each pair with a randomly warped version of the clothing cropped from the real human image, plain H2G-UH is trained using the mixed distribution of the real and synthetic pairs without the augmentation label identifying them, and H2G-UH uses the mixed data with the identifying label. See Sec. 5.2 for discussion." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.349, + 0.459, + 0.478 + ], + "angle": 0, + "content": "
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
Inpainting9.266.330.9090.068
EARSB (w.o. M)9.216.270.9120.061
EARSB (rand(M))9.136.550.9020.071
EARSB (w.o. CG)8.485.320.9180.059
EARSB (full)8.425.250.9180.059
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.482, + 0.483, + 0.553 + ], + "angle": 0, + "content": "Table 5. Comparing noise scheduling strategies on VITON-HD. We include a simple inpainting baseline, not using the error map during training (w.o. \\(M\\)), a random error map \\((\\mathrm{rand}(M))\\), and removing classifier guidance (w.o. CG). These results demonstrate the importance of using a meaningful error map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.687 + ], + "angle": 0, + "content": "importance of the synthetic product-view. Additionally, the poor plain H2G-UH results indicate the presence of a synthetic-real domain gap when using H2G-UH. Inspired by [28], one way to address this issue is by pretraining on the synthetic data and finetuning on real samples (see pre. H2G-UH). However, we find it most effective to condition on a synthetic data indicator while training on mixed data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Error-Aware Noise Schedule. Tab. 5 explores the importance of the error-aware noise schedule described in Sec. 4.2, where the error map adapts the noise distribution according to the quality of the image patches in the initial image \\( x_{1} \\). This adaptive approach contrasts with a uniform Gaussian noise application across all locations, which would reduce our model to \\( \\mathrm{I}^2\\mathrm{SB} \\). We consider three baselines: Inpainting regenerates rather than refines erroneous regions (where the mean confidence \\( M \\) of an image patch containing an error is greater than 0.5), w.o. \\( M \\) removes the error map during training, and \\( \\mathbf{rand}(M) \\) indicates a random error map. As shown in Tab. 5, EARSB outperforms all these baselines, underscoring the importance of a meaningful error map in precisely locating and enhancing" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.888, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.224, + 0.905, + 0.28 + ], + "angle": 0, + "content": "Figure 6. The precision-recall curve for retrieving annotated errors at the pixel level, comparing our WSC to two fully unsupervised baselines (UC, CC). WSC performs best at retrieving generation artifacts at a nominal labeling cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.415 + ], + "angle": 0, + "content": "ing targeted regions. Additionally, the slight performance degradation observed when removing the classifier guidance (EARSB(w.o. CG)) suggests that the error map employed in our classifier guidance also contributes to overall image quality improvement. Collectively, these findings highlight the crucial role of our adaptive noise schedule in achieving superior results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Weakly-Supervised error Classifier (WSC). Our weakly supervised classifier from Sec. 4.1 highlights low-quality regions in the initial image \\( x_{1} \\) with only a few hours of labeling. To validate its effectiveness, we train two ablations of our WSC: the Unsupervised Classifier (UC) that only uses image labels (i.e., fake or real), and the Fake/Real Composite Classifier (CC). CC uses both image fake/real labels as well as fake region-level labels which are created by compositing real image patches and fake image patches. The compositing is a fully automatic alternative to manual labeling that provides patch-level labels. We annotated 100 images in the test set to validate their effectiveness. Fig. 6 shows the pixel-level precision-recall curve for retrieving annotated artifact pixels within the bounding boxes using the classifiers' confidence maps. It is clear that weak supervision remains an incredibly cost-effective approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.693 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper addresses two shortcomings of prior work on virtual try-on. First, we address the limited data availability by introducing a human-to-garment model that generates (human, synthetic garment) pairs from a single image of a clothed individual. Second, we propose a refinement model EARSB that surgically targets localized generation errors from the output of a prior model. EARSB improves the low-quality region of an initially generated image based on a spatially-varying noise schedule that targets known artifacts. Experiments on two benchmark datasets demonstrate that our synthetic data augmentation improves the performance of prior work and that EARSB enhances the overall image quality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21245" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.199 + ], + "angle": 0, + "content": "Acknowledgments This material is based upon work supported, in part, by DARPA under agreement number HR00112020054 and the National Science Foundation under Grant No. DBI-2134696. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the supporting agencies." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.214, + 0.188, + 0.23 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.239, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[1] Sina Alemohammad, Ahmed Imtiaz Humayun, Shruti Agarwal, John Collomosse, and Richard Baraniuk. Self-improving diffusion models with synthetic data. arXiv preprint arXiv:2408.16333, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.297, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.382, + 0.484, + 0.438 + ], + "angle": 0, + "content": "[3] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.44, + 0.484, + 0.48 + ], + "angle": 0, + "content": "[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.483, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.527, + 0.484, + 0.569 + ], + "angle": 0, + "content": "[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. In ECCV, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.571, + 0.484, + 0.64 + ], + "angle": 0, + "content": "[7] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.642, + 0.484, + 0.684 + ], + "angle": 0, + "content": "[8] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.686, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[9] Damien Dablain, Bartosz Krawczyk, and Nitesh V Chawla. Deepsmote: Fusing deep learning and smote for imbalanced data. IEEE Transactions on Neural Networks and Learning Systems, 34(9):6390-6404, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.744, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.239, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[13] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[14] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In CVPR, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[16] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. StableVITON: Learning semantic correspondence with latent diffusion model for virtual try-on. In CVPR, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[19] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[20] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[21] Nannan Li, Qing Liu, Krishna Kumar Singh, Yilin Wang, Jianming Zhang, Bryan A Plummer, and Zhe Lin. UniHuman: A unified model for editing human images in the wild. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[22] Zhi Li, Pengfei Wei, Xiang Yin, Zejun Ma, and Alex C Kot. Virtual try-on with pose-garment keypoints guided inpainting. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[23] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I2sb: image-to-image schrodinger bridge. In ICML, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[24] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[25] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[26] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. LaDIVTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On. In ACM Multimedia, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[27] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[28] Maan Qraitem, Kate Saenko, and Bryan A. Plummer. From fake to real: Pretraining on balanced synthetic images to pre" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21246" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "vent spurious correlations in image recognition. In The European Conference on Computer Vision (ECCV), 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[29] Sang-Heon Shim, Jiwoo Chung, and Jae-Pil Heo. Towards squeezing-averse virtual try-on via sequential deformation. In AAAI, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[30] C Shivashankar and Shane Miller. Semantic data augmentation with generative models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[33] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[34] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive gan. NeurIPS, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[35] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. GPVTON: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In CVPR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[36] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In CVPR, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[37] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In CVPR, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[39] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. TryOnDiffusion: A tale of two unets. In CVPR, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21247" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_origin.pdf b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4ad4c6cc570b612e5a6b6c6ef91796aca46ccd22 --- /dev/null +++ b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/05159407-58d0-4826-9db8-7d0aa541ed98_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83613c4c6b5b414cc809d7db921422a7a3d2bc77d6876d7bd6b2cafe3a0d34e7 +size 3512655 diff --git a/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/full.md b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc7178e7bd1944b6d95a2e6d838ec59f382378dd --- /dev/null +++ b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/full.md @@ -0,0 +1,310 @@ +# Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling + +Nannan Li +Boston University +nnli@bu.edu + +Kevin J. Shih NVIDIA + +Bryan A. Plummer +Boston University +bplum@bu.edu + +# Abstract + +Given an isolated garment image in a canonical product view and a separate image of a person, the virtual try-on task aims to generate a new image of the person wearing the target garment. Prior virtual try-on works face two major challenges in achieving this goal: a) the paired (human, garment) training data has limited availability; b) generating textures on the human that perfectly match that of the prompted garment is difficult, often resulting in distorted text and faded textures. Our work explores ways to tackle these issues through both synthetic data as well as model refinement. We introduce a garment extraction model that generates (human, synthetic garment) pairs from a single image of a clothed individual. The synthetic pairs can then be used to augment the training of virtual try-on. We also propose an Error-Aware Refinement-based Schrödinger Bridge (EARSB) that surgically targets localized generation errors for correcting the output of a base virtual try-on model. To identify likely errors, we propose a weakly-supervised error classifier that localizes regions for refinement, subsequently augmenting the Schrödinger Bridge's noise schedule with its confidence heatmap. Experiments on VITON-HD and DressCode-Upper demonstrate that our synthetic data augmentation enhances the performance of prior work, while EARSB improves the overall image quality. In user studies, our model is preferred by the users in an average of $59\%$ of cases. Code is available at this link. + +# 1. Introduction + +Virtual try-on aims to generate a photorealistic image of a target person wearing a prompted product-view garment [22, 36, 39]. It allows users to visualize how garments would fit and appear on their bodies without the need for physical trials. While recent methods have made significant strides in this field [18, 29, 35, 36], noticeable artifacts such as text distortion and faded textures persist in gener + +![](images/ce5ff823a899f3e784948fde1ca27fdd30f51870986ddaa5e941c699096ee03a.jpg) +Figure 1. Example of our proposed Error-Aware Refinement Schrödinger Bridge (EARSB). EARSB can refine the artifacts (marked by bounding boxes) in an initial image generated by an existing try-on model. The initial image is generated by [18] in the top row and by [29] in the bottom row. + Syn. Data in the last column strengthens the refinement with the proposed synthetic data augmentation in training. + +ated images. For example, as illustrated in the second row of Fig. 1, the logo and the text on the t-shirt noticeably fade away in the initial image generated by a prior try-on model [29]. These imperfections stem from two primary challenges in virtual try-on: limited data availability and the complexity of accurate garment texture deformation. To address these issues, we propose a two-pronged approach: augmenting training data through cost-effective synthetic data generation, and surgically targeting known generation artifacts using our proposed Error-Aware Refinement-based Schrödinger Bridge (EARSB). + +At a minimum, the training data of virtual try-on requires paired (human, product-view garment) images. The product-view garment image is a canonical, front-facing view of the clothing with a clean background. A substantial amount of data is needed to capture the combinatorial space comprising all possible human poses, skin tones, viewing + +angles, and their respective physical interactions with fabric textures, shapes, letterings, and other material properties. Unfortunately, these images are generally available only on copyright-protected product webpages and, therefore, are not readily available for use. To mitigate this issue, we propose to augment training with synthetic data generated from the easier symmetric human-to-garment task, wherein we train a garment-extraction model to extract a canonical product-view garment image from an image of a clothed person. This will allow us to create synthetic paired training data from unpaired datasets [11, 24, 34]. Our results demonstrate that incorporating the more readily available synthetic training pairs can improve image generation quality in the virtual try-on task. + +In addition to addressing the data scarcity issue, we aim to construct a refinement model that can make localized adjustments to a weaker model's generation results. Our approach draws inspiration from classical boosting approaches where every model in a cascade of models targets the shortcomings of the preceding models. We are interested in a targeted refinement approach for two main reasons: it allows a training objective that is focused solely on fixing specific errors, and potentially saves computation when initial predictions are sufficiently good. + +Two components are necessary to achieve such a pipeline: a classifier for identifying localized generation errors, and a refinement model that can re-synthesize content specifically in these localized regions. We found that an effective Weakly-Supervised error Classifier (WSC) can be constructed with just a few hours of manual labeling of generation errors. Another benefit of this approach is that it can be easily tailored for the errors of a specific model that produces images with artifacts. The resulting WSC will produce an error map highlighting low-quality regions. Subsequently, we adopt an Image-to-Image Schrödinger Bridge $(\mathrm{I}^2\mathrm{SB})$ [23] to learn the refinement of these regions in the generated images. While typical diffusion models map from noise to data, $\mathrm{I}^2\mathrm{SB}$ constructs a Schrödinger Bridge (SB) that allows us to map from data to data, or in our setup, generations with artifacts to ground truth images. In addition, we introduce an adaptive noise schedule to direct the SB process to focus on the localized errors by incorporating the classifier's prediction error into the noise schedule, which we describe in more detail in Sec. 4.1. As shown in the first row of Fig. 1, our refinement SB model (i.e., EARSB) corrects the distorted text in the initially generated image. + +The contributions of our paper are: + +- We introduce (human, synthetic garment) pairs as an augmentation in the training of virtual try-on task. The synthetic garment is obtained from our human-to-garment model, which can generate product-view garment images from human images. +- We introduce a spatially adaptive Schrödinger Bridge + +model (EARSB) to refine the outputs of a base virtual try-on model. Our formulation incorporates a spatially varying diffusion noise schedule, with noise proportional to the degree of refinement we wish to perform locally. We find this to yield better results than the baseline Schrödinger Bridge framework. + +- Extensive experiments on two datasets (VITON-HD [20] and DressCode-Upper [25]) show that EARSB enhances the quality of the images generated by prior work, and is preferred by the users in $59\%$ cases on average. + +# 2. Related Work + +Training with Synthetic Data. The addition of synthetic data is often an effective means of improving downstream task performance when it is difficult to amass real data at the necessary scale. This has been demonstrated in the domains of image generation [17, 30] and image editing [5, 33]. Careful applications can also be used to ameliorate dataset imbalance issues, as shown in [9]. Other works, such as [1], use self-synthesized data to provide negative guidance for the diffusion model. Our incorporation of synthetic data in the virtual try-on task tackles a specific sub-problem in the broader image editing domain and is similar in spirit to [5, 33]. Specifically, we aim to synthesize paired training data that satisfies the stringent requirements of virtual try-on paired training data – a canonical product-view garment image paired with an example of it being worn. Images of people in clothing are readily available, but it is difficult to obtain a product-view image of the exact clothing they are wearing. To address this, our work tackles the human-to-garment problem, which aims to extract the clothing from a person's photo and project it to the canonical product view, making it roughly symmetric to the virtual try-on task. + +Virtual Try-On. There has been a shift from earlier GAN-based framework [14, 20, 22, 29, 35] to diffusion-based methods [7, 18] in the virtual try-on literature. Diffusion models fit an SDE process mapping from the image distribution to the noise distribution, and tend to be easier to train than GAN-based approaches due to the simplicity of the L2 denoising loss [10, 12, 31]. At inference, the diffusion model denoises a random Gaussian noise distribution to a human-readable image via multiple sampling steps. [7, 36] propose parameter-efficient approaches that concatenate the human image and the garment images along the spatial dimension such that the self-attention layer in the denoising UNet can achieve texture transfer without extra parameters. In [18], the authors introduce additional cross-attention layers to learn the semantic correspondences between the garment and the human image. The methods in [3, 21, 27] align different embedding spaces in the attention module to achieve flexible clothing editing after try-on, such as style change or graphics insertion. In contrast to prior work that samples from random noise, we build upon + +![](images/bc735dfcbe2cb0fed7de7d183d8f99ba89111d1b3a2ebaaf015381c33912280f.jpg) +(a) Human-to-Garment model. + +![](images/a866f0be82048b99dcd5de44ca60bff340b1cfbd5b09d33289accdc2096578d8.jpg) +(b) Examples in H2G-UH and H2G-FH. +Figure 2. (a) Our human-to-garment model, which is explained in Sec. 3.1 (b) Examples of the constructed (human, synthetic garment) pairs in Sec. 3.1. + +recent advances in Schrödinger bridges, notably [23], to directly sample from an initial image generated by prior try-on models. Our work is similar in spirit to [37], which initializes the noisy image with a GAN-generated image and small amounts of random noise. However, our work explores varying the local noise schedule based on the error level at a given location. + +# 3. Augmented Training with Synthetic Data + +Virtual Try-On Task Definition. Let $(x_0, C)$ be the (human image, product-view of worn garment) pair in virtual try-on training. We will refer to $(x_0, C)$ as paired data. Let $\bar{x}_0$ be a masked version of $x_0$ in which the worn garment corresponding to $C$ is masked out. We can then set up a learning task in which we aim to fit the following function: $F(\bar{x}_0, C, \phi; \theta) \to x_0$ , where $\phi$ corresponds to other conditionals such as pose representations from DensePose [13]. + +Acquiring high-quality pairs $(x_0, C)$ at scale is challenging due to copyright and brand protection, but acquiring images of just humans $(x_0)$ at scale is considerably more feasible [11, 21, 24, 34]. This observation motivates proposed human-to-garment process to extract a synthetic canonical view image $\hat{C}$ from $x_0$ . We can then augment our virtual try-on training with $(x_0, \hat{C})$ pair, requiring only single human images. In the following, Sec. 3.1 explains the archi + +tecture of our human-to-garment model, Sec. 3.2 discusses how we use this model to construct the synthetic dataset, and Sec. 3.3 describes how the synthetic data is used to augment the virtual try-on training. + +# 3.1. Human-to-Garment Model + +While virtual try-on requires generating skin and deforming the product-view garment to accommodate diverse postures, the human-to-garment task simply aims to map the clothing item to its canonical view. To achieve this, we use existing paired (human, garment) data (e.g., VITON-HD [20]) to train our human-to-garment model. As illustrated in Fig. 2a, we first segment and extract the clothing on the person map and then feed the clothing item to a generator that synthesizes its canonical view. The generator is based on the UNet model proposed in [14], which uses a flow-like mechanism for warping latent features in an optical-flow-like manner. The generator was trained using a combined L1 reconstruction and adversarial loss. + +# 3.2. Constructing Synthetic H2G-UH and H2G-FH + +Synthetic images $\hat{C}$ produced from our models necessarily contain generation errors. We use the following criteria to filter for high-quality synthetic data: a) The single human image $x_0$ has a clean background (low pixel variance in the non-human region); b) $x_0$ is frontal view (classified by its DensePose representation [13]); c) the reconstruction error (LPIPS distance) is small when reconstructing the human image $x_0$ in a try-on model using the $(x_0, \hat{C})$ pair (e.g., [20, 29]). Under these criteria, we select human images from DeepFashion2 [24] and UPT [34], eventually creating 12,730 synthetic pairs of upper-body human images (referred to as H2G-UH) and 8,939 pairs of full-body human images (referred to as H2G-FH). Examples of the synthetic pairs are shown in Fig. 2b. + +# 3.3. Augmented Virtual Try-on Training + +To further prevent distribution leakage of incorporating synthetic data, we explore two means of limiting the effect of the real-synthetic domain gap: (a) two training stages involving pretraining the try-on model using synthetic pairs, and then finetuning on real pairs [19]; (b) training simultaneously on real and synthetic data, but conditioning the try-on model on a real/synthetic flag, similar to [16]. We found empirically that the second augmentation performs slightly better than the first (See Sec. 5.1). + +# 4. Error-Aware Refinement Schrödinger Bridge + +Apart from the synthetic data augmentation from Sec. 3, our second approach to enhancing existing try-on methods is a refinement pipeline with two steps. First, given some + +![](images/1a97d47f3f18a1c439164979b9d506dff344fd07bb800734353798f309804de5.jpg) +Figure 3. The diffusion process in our refinement-based EARSB. We first preprocess the input image, then use a base try-on model that takes the masked human image $\bar{x}_0$ , its pose representation $P$ , and its garment $C$ as input to generate an initial human image $x_1$ . $x_1$ is fed to our weakly-supervised classifier (WSC) to obtain the error map $M$ (see Sec. 4.1). This map reweights the noise distribution $\epsilon$ to $\epsilon^r$ in $\mathrm{I}^2\mathrm{SB}$ diffusion and refines $x_1$ that has generation errors to the ground truth image $x_0$ (see Sec. 4.2). + +base model $F_{base}(\bar{x}_0, C, \phi) \to x_1$ where $F_{base}(.)$ can be any pretrained GAN or diffusion-based approach to virtual try-on, $x_1$ closely approximates the true real human image $x_0$ with some generation artifacts. To automatically identify the artifacts in $x_1$ , we construct a weakly-supervised error classifier $\mathrm{WSC}(x_1, C) \to M$ as in Fig. 3, where $M$ is a confidence map predicting a heatmap for likely generation errors. Then, our second step performs the final refinement by fitting a Schrödinger bridge based on $\mathrm{I}^2\mathrm{SB}$ [23] via the following mapping: $F_{EARSB}(x_1, C, M, \phi; \theta) \to x_0$ . + +The approach is weakly inspired by boosting methods in that we wish to fit a targeted refinement model that is trained specifically on the generation errors of an existing model. The refinement goal applies to the general setting where we want to refine a flawed image output, though we focus on virtual try-on for this work. Thus, as illustrated in Fig. 3, the training of EARSB includes three steps: + +1. Pre-process the images in the training set and feed them to existing try-on models to get the initial images $x_{1}$ . +2. Obtain the error maps $M$ on the initial images $x_{1}$ using our WSC (Sec. 4.1). +3. Use $M$ to adjust the noise schedule in $\mathrm{I}^2\mathrm{SB}$ [23] and train the noise prediction model in EARSB following Eq. (9) (Sec. 4.2). + +As the first step simply requires running an off-the-shelf virtual try-on model to obtain an initial image, our discussion in the next section will begin by describing the second step- obtaining the error map. + +# 4.1. Obtaining the Error Map + +We start by obtaining the error map $M$ that highlights the corrupted or incorrect area of the initial image $x_{1}$ using our proposed Weakly-Supervised error Classifier (WSC). + +Classifier Architecture. As shown in the green dotted box of Fig. 3, our WSC has two encoders to match the image features of $x_{1}$ and $C$ with cross attention to predict a sigmoid-activated error map. + +Training Data Annotation. In practice, it is labor-intensive to fully annotate all the initial images for where the generated artifacts are located. To mitigate this issue, we used a few hours to hand-label a small portion of the initial try-on images in the training set at the patch level, using bounding boxes for poorly generated regions. + +Weakly-Supervised Training. Let $x_0, x_1^u, x_1^l$ be the real human image, the unlabeled initial image, and the labeled initial image with bounding boxes annotating artifacts. Our WSC loss terms are defined as: + +$$ +\mathcal {L} _ {i m g} = - \log \left(\mathrm {W S C} \left(x _ {1} ^ {u}, C\right) _ {\max }\right) + \log \left(1 - \mathrm {W S C} \left(x _ {0}, C\right) _ {\max }\right) +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {p a t} = - \log \left(\operatorname {W S C} \left(x _ {1} ^ {l}, C\right) \odot B _ {b o x}\right) \\ - \log \left(1 - \operatorname {W S C} \left(x _ {1} ^ {l}, C\right) \odot \left(1 - B _ {\text {b o x}}\right)\right) \tag {1} \\ \end{array} +$$ + +where $\mathcal{L}_{img}$ is the image-level loss and $\mathcal{L}_{pat}$ is the patch-level loss. In $\mathcal{L}_{img}$ , $\mathrm{WSC}(\cdot)$ is the output error map and $\mathrm{WSC}(\cdot)_{\mathrm{max}}$ denotes the spatially max-pooled score in the error map. In $\mathcal{L}_{pat}$ , $B_{box}$ is the spatial binary mask for the annotated regions, thereby maximizing and minimizing the scores for regions within and outside of the annotated boxes + +respectively. Our final loss is: $\mathcal{L}_{\mathrm{WSC}} = \mathcal{L}_{ins} + \mathcal{L}_{pat}$ + +The trained WSC will predict an error map $M$ for the initial image $x_{1}$ , which is then used to adjust the noise schedule in the diffusion process described in the next section. + +# 4.2. Error-Map-Reweighted SB Formulation + +To achieve the refinement goal, our diffusion process extends Schrödinger bridges as formulated in $\mathrm{I}^2\mathrm{SB}$ [23], where we incrementally add noise to the initial image $x_{1}$ , and then remove the noise to approximate the refined image $x_0$ . However, without additional information, a naively trained $\mathrm{I}^2\mathrm{SB}$ model must implicitly learn what to refine and what to retain. Our formulation aims to explicitly incorporate prior knowledge of localized generation errors via the error map $M$ into the Schrödinger process by using $M$ to locally scale the noise schedule for the Schrödinger process. + +Our choice of locally scaling the noise schedule is based on several observations. We want the model to directly copy pixels over to $x_0$ for correctly generated regions in $x_1$ - it would be nice to avoid training the model to add and remove noise from these regions. In contrast, erroneous regions in the initial try-on images, especially more noticeable ones, include generation errors that share little to no structural similarity to the target. These errors include examples such as deformed limbs, and distorted textures/fabrics, which may need more added noise to prevent the model from conditioning too strongly on the original pattern. + +As such, we construct our refinement model $F_{EARSB}(x_1,C,M,P;\theta) \to x_0$ , where the model is conditioned on the canonical view garment $C$ , the error map $M$ , and the pose representation $P$ . Fig. 3 shows our Weakly-Supervised Classifier (WSC) first locates the errors in the error map $M$ , then $M$ reweights the noise schedule of the I²SB stochastic process to assign a higher volume of noise to the low-quality region "rebel" so the model can focus on refining it. + +Error-Map-Reweighted Diffusion Process. Following I²SB [23], our diffusion Schrödinger bridge maps from the initial image $x_{1}$ to the ground truth image $x_{0}$ . It fits to the following stochastic process: + +$$ +\begin{array}{l} x _ {t} = \mu_ {t} \left(x _ {0}, x _ {1}\right) + \sqrt {\Sigma_ {t}} \cdot \epsilon \\ \mu_ {t} = \frac {\bar {\sigma} _ {t} ^ {2}}{\bar {\sigma} _ {t} ^ {2} + \sigma_ {t} ^ {2}} x _ {0} + \frac {\sigma_ {t} ^ {2}}{\bar {\sigma} _ {t} ^ {2} + \sigma_ {t} ^ {2}} x _ {1}, \Sigma_ {t} = \frac {\bar {\sigma} _ {t} ^ {2} \sigma_ {t} ^ {2}}{\bar {\sigma} _ {t} ^ {2} + \sigma_ {t} ^ {2}} \cdot I, \tag {2} \\ \end{array} +$$ + +where $\sigma_t^2 = \int_0^t\beta_\tau d\tau$ , $\bar{\sigma}_t^2 = \int_t^1\beta_\tau d\tau$ and $\beta_{\tau}$ is a symmetrical noise schedule. $\epsilon \sim \mathcal{N}(0,I)$ is random Gaussian noise. The above equation stochastically adds noise and then removes it between $x_{1}$ and $x_0$ . + +We extend $\mathrm{I}^2\mathrm{SB}$ such that the noise schedule can vary spatially based on the error map $M$ (obtained from WSC in Sec. 4.1). Good regions will be assigned less noise (i.e., smaller variance) in the diffusion process, while poor qual + +ity regions will be assigned more: + +$$ +x _ {t} = \mu_ {t} \left(x _ {0}, x _ {1}\right) + \sqrt {\Sigma_ {t}} \cdot \epsilon^ {r}, \tag {3} +$$ + +$$ +\epsilon^ {r} = M \cdot \epsilon , M = \operatorname {W S C} \left(x _ {1}, C\right) \tag {4} +$$ + +where $\mu_t$ is the same as Eq. (2) and $\epsilon^r$ is the adaptive noise. + +Sampling Process. The initial image $x_{1}$ is iteratively refined to $x_{0}$ via a denoising/sampling process, where a model predicts the noise distribution at each time step. In contrast to prior soft-attention-based UNets [18, 26, 36, 37], our denoising model uses cloth-flow-learning UNet for more precise garment deformation [14]. It accepts the garment $C$ , the error map $M$ , the pose representation $P$ , and the noisy image $x_{t}$ as inputs and predicts the error-adapted noise $\epsilon_{\theta}^{r}(\cdot ;t)$ , where $(\cdot ;t)$ omits the inputs $M, P, x_{t}, C$ . See Supp. B for the detailed model architecture. With the predicted noise $\epsilon_{\theta}^{r}(\cdot ;t)$ , we define our sampling process: + +$$ +\hat {x} _ {0} = x _ {t} - \sqrt {\Sigma_ {t}} \cdot \epsilon_ {\theta} ^ {r} (M, P, x _ {t}, C; t) \tag {5} +$$ + +$$ +x _ {t - \Delta t} = \hat {\mu} _ {t - \Delta t} \left(\hat {x} _ {0}, x _ {t}\right) + M \cdot \sqrt {\hat {\Sigma} _ {t}} \cdot \epsilon \tag {6} +$$ + +$$ +\hat {\mu} _ {t - \Delta t} = \frac {\sigma_ {t - \Delta t} ^ {2}}{\sigma_ {t} ^ {2}} \hat {x} _ {0} + \frac {\sigma_ {t} ^ {2} - \sigma_ {t - \Delta t} ^ {2}}{\sigma_ {t} ^ {2}} x _ {t} \tag {7} +$$ + +$$ +\hat {\Sigma} _ {t} = \frac {\sigma_ {t - \Delta t} ^ {2} \left(\sigma_ {t} ^ {2} - \sigma_ {t - \Delta t} ^ {2}\right)}{\sigma_ {t} ^ {2}} \tag {8} +$$ + +where $\Delta t > 0$ and it is the sampling interval. Starting from $t = 1$ , the process iteratively refines the initial human image $x_{1}$ based on the error map $M$ . When $M$ is all ones in Eq. (5), our model reverts to the $\mathrm{I}^2\mathrm{SB}$ formulation. When $M$ is all zeros (i.e., no error), $x_{1}$ is believed to be perfect $x_{1}$ does not need to be refined in the sampling process. + +The training objective of our model is the mean squared error between the predicted noise $\epsilon_{\theta}^{r}$ and the reweighted Gaussian noise $\epsilon^{r}$ + +$$ +\mathcal {L} _ {\mathrm {E A R S B}} = \mathbb {E} _ {t \sim U (0, 1)} | | \epsilon_ {\theta} ^ {r} (M, P, x _ {t}, C; t) - \epsilon^ {r} | | ^ {2} \tag {9} +$$ + +# 4.2.1. Further Improvements via Classifier Guidance and Expert Denoisers + +Whereas prior work used an object category classifier to guide the sampling process [10], our WSC guidance gives a direction toward the real data distribution. Chung et al. [8] shows we can estimate the guidance score $\nabla_{x_t}\log p(y|x_t)$ using the denoised clean image $\hat{x}_0$ : $\nabla_{x_t}\log p(y|x_t)\simeq \nabla_{x_t}\log p(y|\hat{x}_0)$ , where $y$ is the fake/real label. Since the label for real data is 0 in WSC, the classifier guidance is: + +$$ +\hat {\mu} _ {t - \Delta t} \leftarrow \hat {\mu} _ {t - \Delta t} + M \cdot \hat {\Sigma} _ {t} \cdot \nabla_ {x _ {t}} \log p (\mathbf {0} | \hat {x} _ {0}) \tag {10} +$$ + +where $p(\mathbf{0}|\hat{x}_0) = 1 - \mathrm{WSC}(\hat{x}_0,C)$ . + +Following [2], a trained EARSB model is split into two models, each having an expert denoiser that is fine-tuned on denoising ranges $t \in [0, 0.5]$ and $t \in [0.5, 1]$ , respectively. + +
UnpairedVITON-HDDressCode-Upper
PairedUnpairedPaired
FID↓KID↓FID↓KID↓SSIM↑LPIPS↓FID↓KID↓FID↓KID↓SSIM↑
(a) GAN-Based
HR-VTON [20]10.750.288.460.260.9010.07515.260.3911.760.320.947
SD-VTON [29]9.050.126.470.090.9070.07014.730.3210.990.240.947
GP-VTON [35]8.610.865.530.070.9130.06426.191.7123.661.590.816
EARSB (ours)8.420.075.250.050.9180.05910.890.137.150.130.961
EARSB +H2G-UH/FH (ours)8.260.065.140.040.9190.05810.700.117.050.110.965
(b) SD-Based
LaDI-VTON [26]8.950.126.050.080.9020.07114.880.3911.610.320.939
CatVTON [7]8.870.085.490.070.9150.05911.910.217.660.100.950
CAT-DM [37]8.550.105.980.070.9080.06712.910.298.580.160.948
IDM-VTON [6]8.590.115.510.090.9020.06111.090.166.790.120.956
TPD [36]8.230.064.860.040.9170.057-
StableVITON [18]8.200.075.160.050.9170.057-
EARSB(SD) +H2G-UH/FH (ours)8.040.064.900.030.9250.05310.410.096.760.080.968
+ +# 5. Experiments + +Datasets. We use VITON-HD [20], DressCode-Upper [25], and our synthetic H2G-UH and H2G-FH for training. They include 11,647, 13,564, 12,730, 8,939 training images, respectively. For synthetic data augmentation, we combine VITON-HD with our H2G-UH since they both include mostly upper-body human images. DressCode-Upper is combined with H2G-FH as both consist of full-body human photos. For evaluation, VITON-HD contains 2,032 (human, garment) test pairs and DressCode-Upper has 1,800 test pairs and include both paired and unpaired settings. In the paired setting, the input garment image and the garment in the human image are the same item. Conversely, the unpaired setting uses a different garment image. + +Metrics. We use Structural Similarity Index Measure (SSIM) [32], Frechet Inception Distance (FID) [15], Kernel Inception Distance (KID) [4], and Learned Perceptual Image Patch Similarity (LPIPS) [38] to evaluate image quality. All the compared methods use the same image size $512 \times 512$ and padding when computing the above metrics. + +Experimental setup. We compare EARSB with GAN-based methods HR-VTON [20], SD-VTON [29] and GPVTON [35], as well as Stable Diffusion (SD) based methods including CAT-DM [37], StableVITON [18], TPD [36], IDM-VTON [6] and CatVTON [7]. Unless otherwise specified, all diffusion models use 25 sampling steps. + +We report results of multiple variants of our approach. EARSB uses GAN-based GP-VTON [35] to generate the initial image that was trained without synthetic data augmentation. EARSB+H2G-UH/FH trains with either H2G- + +Table 1. Virtual try-on results on VITON-HD [20] and DressCode-Upper [25] for (a) GAN-based and (b) diffusion-based models using 25 sampling steps. KID is multiplied by 100. We find our EARSB approach outperforms prior work on average. See Sec. 5.1 for discussion. + +
MethodsGP-VTON EARSBStableVITON EARSB
Consistency42%58%38%62%
Fidelity39%61%45%55%
+ +Table 2. User studies on VITON-HD. Our EARSB method is preferred in an average of $59\%$ cases. + +UH or H2G-FH. We add the upper-body synthetic subset H2G-UH for the upper-body-human dataset VITON-HD, and the full-body synthetic H2G-FH when on DressCodeUpper. EARSB(SD)+H2G-UH/FH uses the diffusion-based CatVTON [7] to generate the initial image. + +# 5.1. Results + +Tab. 1 compares our approach with those from prior work. We find our full model variants EARSB+H2G-UH/FH and EARSB(SD)+H2G-UH/FH boosts performance over the GAN and SD-based methods, respectively. The last two rows of Tab. 1(a) show that incorporating our synthetic training pairs provide a consistent boost in performance on both datasets. Comparing the last rows of Tab. 1(a) and Tab. 1(b) we observe that using the diffusion model to generate the initial image gives better performance in EARSB(SD)+H2G-UH/FH, but is more costly. + +User Study. We asked Amazon MTurk workers to evaluate the texture consistency and image fidelity of synthesized images, comparing our model against GP-VTON and StableVITON. We randomly selected 100 pairs from VITON-HD to evaluate on, assigning at least 3 workers per image. Study results in Tab. 2 report our method is preferred at least $10\%$ more than the GAN-based GP-VTON and the SD-based StableVITON (59% overall). + +![](images/e65258728eb019c37b994906975e059994c4ee72d62edc40b57937f329a02678.jpg) +Unpaired FID +Figure 4. Results on VITON-HD at 5, 10, 25, 50, and 100 sampling steps. Our method consistently improves our baseline starting model GP-VTON (black, dotted line), making it competitive with StableVITON (especially at under 50 sampling steps). Legend is shared for all. + +![](images/83f869cd286951474efcd491f9ea0d87dbd916af9470854750c7e64028f3938e.jpg) +Paired FID + +![](images/d476eb027df389ed00b1301f4df47f7193d704f7adfa5eef0fbd85bd07908158.jpg) +Paired SSIM + +![](images/57fde6822d1365bbef2bd0a29506996da1076e8ab9d9a559ef2e97bc6f88ef45.jpg) +Figure 5. Visualizations on VITON-HD (top row) and DressCode (bottom row). Our EARSB+H2G-UH and EARSB(SD)+H2G-UH better recover the intricate textures in the garment. + +Trade-Off Between Sampling Efficiency and Image Quality. Diffusion-generated images often show degraded quality with fewer sampling steps. In EARSB, the spatially adaptive noise schedule can preserve correct clothing textures in the initial image with a low noise level and only fix the erroneous parts, potentially resulting in less image quality degradation with fewer steps. In Fig. 4, while other SD-based methods have a sharp performance drop with decreasing sampling steps, EARSB and EARSB+H2G-UH show consistent performance across different sampling steps, demonstrating a better trade-off between image quality and computational efficiency. + +Qualitative Results. Fig. 5 gives examples of generated images from different approaches. The top row is from VITON-HD dataset and the bottom row is from DressCodeUpper. The third images in the two rows are GAN-generated results. We see that our EARSB+H2G-UH/FH in the last column improves the low-quality textures from the GAN-generated images, which are the distorted graphics in the center. Additional visualizations are in Supp. G. + +# 5.2. Ablations + +Synthetic Pairs Augmentation. Tab. 3 incorporates H2GUH into the training of StableVITON [18] and CAT-DM + +
Data Aug.Unpaired FID↓Paired FID↓SSIM↑LPIPS↓
CAT-DM [37]None8.565.900.9110.067
CAT-DMH2G-UH8.365.670.9130.063
StableVITON [18]None8.255.150.9170.056
StableVITONH2G-UH8.175.040.9190.054
EARSB(SD)H2G-UH8.044.900.9250.053
+ +Table 3. Comparing the effect of our H2G-UH data augmentation approach on VITON-HD. We bold the best overall results and underline the best results for a single model with and without H2G-UH. Each method uses the number of sampling steps from their paper: 2 for CAT-DM, 50 for StableVITON, and 25 for our own EARSB. We find H2G-UH consistently boosts performance. + +[37] on the VITON-HD dataset to validate the effectiveness of synthetic pairs on enhancing existing diffusion methods. We observe that training with our synthetic H2G-UH is effective in improving most metrics. + +Tab. 4 explores different ways of incorporating the synthetic data H2G-UH during the training of EARSB. We find that using a randomly warped version of the clothing $(W(\mathrm{H}2\mathrm{G}-\mathrm{U}\mathrm{H}))$ hinders performance, demonstrating the + +
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
None8.425.250.9180.059
W(H2G-UH)8.685.440.9090.063
plain H2G-UH9.646.520.9020.073
pre. H2G-UH8.355.180.9180.059
H2G-UH8.265.140.9190.058
+ +Table 4. Ablations of our H2G-UH augmentation on VITON-HD. Specifically, pre. H2G-UH is pretrained using the synthetic pairs and finetuned on real data, $W(\mathbf{H2G - UH})$ replaces the synthetic garment in each pair with a randomly warped version of the clothing cropped from the real human image, plain H2G-UH is trained using the mixed distribution of the real and synthetic pairs without the augmentation label identifying them, and H2G-UH uses the mixed data with the identifying label. See Sec. 5.2 for discussion. + +
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
Inpainting9.266.330.9090.068
EARSB (w.o. M)9.216.270.9120.061
EARSB (rand(M))9.136.550.9020.071
EARSB (w.o. CG)8.485.320.9180.059
EARSB (full)8.425.250.9180.059
+ +Table 5. Comparing noise scheduling strategies on VITON-HD. We include a simple inpainting baseline, not using the error map during training (w.o. $M$ ), a random error map $(\mathrm{rand}(M))$ , and removing classifier guidance (w.o. CG). These results demonstrate the importance of using a meaningful error map. + +importance of the synthetic product-view. Additionally, the poor plain H2G-UH results indicate the presence of a synthetic-real domain gap when using H2G-UH. Inspired by [28], one way to address this issue is by pretraining on the synthetic data and finetuning on real samples (see pre. H2G-UH). However, we find it most effective to condition on a synthetic data indicator while training on mixed data. + +Error-Aware Noise Schedule. Tab. 5 explores the importance of the error-aware noise schedule described in Sec. 4.2, where the error map adapts the noise distribution according to the quality of the image patches in the initial image $x_{1}$ . This adaptive approach contrasts with a uniform Gaussian noise application across all locations, which would reduce our model to $\mathrm{I}^2\mathrm{SB}$ . We consider three baselines: Inpainting regenerates rather than refines erroneous regions (where the mean confidence $M$ of an image patch containing an error is greater than 0.5), w.o. $M$ removes the error map during training, and $\mathbf{rand}(M)$ indicates a random error map. As shown in Tab. 5, EARSB outperforms all these baselines, underscoring the importance of a meaningful error map in precisely locating and enhancing + +![](images/e3449d3c1a0689566e3e505a7aeb64ce903b67d6f61e178ddb0a9e52b1303762.jpg) +Figure 6. The precision-recall curve for retrieving annotated errors at the pixel level, comparing our WSC to two fully unsupervised baselines (UC, CC). WSC performs best at retrieving generation artifacts at a nominal labeling cost. + +ing targeted regions. Additionally, the slight performance degradation observed when removing the classifier guidance (EARSB(w.o. CG)) suggests that the error map employed in our classifier guidance also contributes to overall image quality improvement. Collectively, these findings highlight the crucial role of our adaptive noise schedule in achieving superior results. + +Weakly-Supervised error Classifier (WSC). Our weakly supervised classifier from Sec. 4.1 highlights low-quality regions in the initial image $x_{1}$ with only a few hours of labeling. To validate its effectiveness, we train two ablations of our WSC: the Unsupervised Classifier (UC) that only uses image labels (i.e., fake or real), and the Fake/Real Composite Classifier (CC). CC uses both image fake/real labels as well as fake region-level labels which are created by compositing real image patches and fake image patches. The compositing is a fully automatic alternative to manual labeling that provides patch-level labels. We annotated 100 images in the test set to validate their effectiveness. Fig. 6 shows the pixel-level precision-recall curve for retrieving annotated artifact pixels within the bounding boxes using the classifiers' confidence maps. It is clear that weak supervision remains an incredibly cost-effective approach. + +# 6. Conclusion + +This paper addresses two shortcomings of prior work on virtual try-on. First, we address the limited data availability by introducing a human-to-garment model that generates (human, synthetic garment) pairs from a single image of a clothed individual. Second, we propose a refinement model EARSB that surgically targets localized generation errors from the output of a prior model. EARSB improves the low-quality region of an initially generated image based on a spatially-varying noise schedule that targets known artifacts. Experiments on two benchmark datasets demonstrate that our synthetic data augmentation improves the performance of prior work and that EARSB enhances the overall image quality. + +Acknowledgments This material is based upon work supported, in part, by DARPA under agreement number HR00112020054 and the National Science Foundation under Grant No. DBI-2134696. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the supporting agencies. + +# References + +[1] Sina Alemohammad, Ahmed Imtiaz Humayun, Shruti Agarwal, John Collomosse, and Richard Baraniuk. Self-improving diffusion models with synthetic data. arXiv preprint arXiv:2408.16333, 2024. 2 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5 +[3] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In CVPR, 2023. 2 +[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 6 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2 +[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. In ECCV, 2024. 6 +[7] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 2, 6 +[8] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 5 +[9] Damien Dablain, Bartosz Krawczyk, and Nitesh V Chawla. Deepsmote: Fusing deep learning and smote for imbalanced data. IEEE Transactions on Neural Networks and Learning Systems, 34(9):6390-6404, 2022. 2 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2, 5 +[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, 2022. 2, 3 +[12] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 2 + +[13] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In CVPR, 2018. 3 +[14] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In CVPR, 2019. 2, 3, 5 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6 +[16] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In ICML, 2020. 3 +[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2 +[18] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. StableVITON: Learning semantic correspondence with latent diffusion model for virtual try-on. In CVPR, 2024. 1, 2, 5, 6, 7 +[19] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 3 +[20] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In ECCV, 2022. 2, 3, 6 +[21] Nannan Li, Qing Liu, Krishna Kumar Singh, Yilin Wang, Jianming Zhang, Bryan A Plummer, and Zhe Lin. UniHuman: A unified model for editing human images in the wild. In CVPR, 2024. 2, 3 +[22] Zhi Li, Pengfei Wei, Xiang Yin, Zejun Ma, and Alex C Kot. Virtual try-on with pose-garment keypoints guided inpainting. In ICCV, 2023. 1, 2 +[23] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I2sb: image-to-image schrodinger bridge. In ICML, 2023. 2, 3, 4, 5 +[24] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3 +[25] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In CVPR, 2022. 2, 6 +[26] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. LaDIVTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On. In ACM Multimedia, 2023. 5, 6 +[27] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In CVPR, 2024. 2 +[28] Maan Qraitem, Kate Saenko, and Bryan A. Plummer. From fake to real: Pretraining on balanced synthetic images to pre + +vent spurious correlations in image recognition. In The European Conference on Computer Vision (ECCV), 2024. 8 +[29] Sang-Heon Shim, Jiwoo Chung, and Jae-Pil Heo. Towards squeezing-averse virtual try-on via sequential deformation. In AAAI, 2024. 1, 2, 3, 6 +[30] C Shivashankar and Shane Miller. Semantic data augmentation with generative models. In CVPR, 2023. 2 +[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2 +[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6 +[33] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024. 2 +[34] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive gan. NeurIPS, 2021. 2, 3 +[35] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. GPVTON: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In CVPR, 2023. 1, 2, 6 +[36] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In CVPR, 2024. 1, 2, 5, 6 +[37] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In CVPR, 2024. 3, 5, 6, 7 +[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[39] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. TryOnDiffusion: A tale of two unets. In CVPR, 2023. 1 \ No newline at end of file diff --git a/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/images.zip b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f0fb06cd0650e5b0afdbbd23ea7fe27e136265e --- /dev/null +++ b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33ce0131004325d4d4af1f2ccfc2b8df63ab94c7aab53670a1a4d9537bd82c9e +size 620906 diff --git a/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/layout.json b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..179b9dfaaf2a23c6890613556e8f94d1027f337b --- /dev/null +++ b/2025/Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling/layout.json @@ -0,0 +1,8850 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "text", + "content": "Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 161, + 235, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 161, + 235, + 203 + ], + "spans": [ + { + "bbox": [ + 145, + 161, + 235, + 203 + ], + "type": "text", + "content": "Nannan Li \nBoston University \nnnli@bu.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 270, + 161, + 337, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 161, + 337, + 188 + ], + "spans": [ + { + "bbox": [ + 270, + 161, + 337, + 188 + ], + "type": "text", + "content": "Kevin J. Shih NVIDIA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 372, + 161, + 466, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 161, + 466, + 203 + ], + "spans": [ + { + "bbox": [ + 372, + 161, + 466, + 203 + ], + "type": "text", + "content": "Bryan A. Plummer \nBoston University \nbplum@bu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 258, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 297, + 581 + ], + "type": "text", + "content": "Given an isolated garment image in a canonical product view and a separate image of a person, the virtual try-on task aims to generate a new image of the person wearing the target garment. Prior virtual try-on works face two major challenges in achieving this goal: a) the paired (human, garment) training data has limited availability; b) generating textures on the human that perfectly match that of the prompted garment is difficult, often resulting in distorted text and faded textures. Our work explores ways to tackle these issues through both synthetic data as well as model refinement. We introduce a garment extraction model that generates (human, synthetic garment) pairs from a single image of a clothed individual. The synthetic pairs can then be used to augment the training of virtual try-on. We also propose an Error-Aware Refinement-based Schrödinger Bridge (EARSB) that surgically targets localized generation errors for correcting the output of a base virtual try-on model. To identify likely errors, we propose a weakly-supervised error classifier that localizes regions for refinement, subsequently augmenting the Schrödinger Bridge's noise schedule with its confidence heatmap. Experiments on VITON-HD and DressCode-Upper demonstrate that our synthetic data augmentation enhances the performance of prior work, while EARSB improves the overall image quality. In user studies, our model is preferred by the users in an average of " + }, + { + "bbox": [ + 53, + 258, + 297, + 581 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 53, + 258, + 297, + 581 + ], + "type": "text", + "content": " of cases. Code is available at this link." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Virtual try-on aims to generate a photorealistic image of a target person wearing a prompted product-view garment [22, 36, 39]. It allows users to visualize how garments would fit and appear on their bodies without the need for physical trials. While recent methods have made significant strides in this field [18, 29, 35, 36], noticeable artifacts such as text distortion and faded textures persist in gener" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 553, + 403 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 403 + ], + "type": "image", + "image_path": "ce5ff823a899f3e784948fde1ca27fdd30f51870986ddaa5e941c699096ee03a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 489 + ], + "type": "text", + "content": "Figure 1. Example of our proposed Error-Aware Refinement Schrödinger Bridge (EARSB). EARSB can refine the artifacts (marked by bounding boxes) in an initial image generated by an existing try-on model. The initial image is generated by [18] in the top row and by [29] in the bottom row. + Syn. Data in the last column strengthens the refinement with the proposed synthetic data augmentation in training." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 638 + ], + "type": "text", + "content": "ated images. For example, as illustrated in the second row of Fig. 1, the logo and the text on the t-shirt noticeably fade away in the initial image generated by a prior try-on model [29]. These imperfections stem from two primary challenges in virtual try-on: limited data availability and the complexity of accurate garment texture deformation. To address these issues, we propose a two-pronged approach: augmenting training data through cost-effective synthetic data generation, and surgically targeting known generation artifacts using our proposed Error-Aware Refinement-based Schrödinger Bridge (EARSB)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "At a minimum, the training data of virtual try-on requires paired (human, product-view garment) images. The product-view garment image is a canonical, front-facing view of the clothing with a clean background. A substantial amount of data is needed to capture the combinatorial space comprising all possible human poses, skin tones, viewing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21238" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "angles, and their respective physical interactions with fabric textures, shapes, letterings, and other material properties. Unfortunately, these images are generally available only on copyright-protected product webpages and, therefore, are not readily available for use. To mitigate this issue, we propose to augment training with synthetic data generated from the easier symmetric human-to-garment task, wherein we train a garment-extraction model to extract a canonical product-view garment image from an image of a clothed person. This will allow us to create synthetic paired training data from unpaired datasets [11, 24, 34]. Our results demonstrate that incorporating the more readily available synthetic training pairs can improve image generation quality in the virtual try-on task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 241, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 294, + 361 + ], + "type": "text", + "content": "In addition to addressing the data scarcity issue, we aim to construct a refinement model that can make localized adjustments to a weaker model's generation results. Our approach draws inspiration from classical boosting approaches where every model in a cascade of models targets the shortcomings of the preceding models. We are interested in a targeted refinement approach for two main reasons: it allows a training objective that is focused solely on fixing specific errors, and potentially saves computation when initial predictions are sufficiently good." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 626 + ], + "type": "text", + "content": "Two components are necessary to achieve such a pipeline: a classifier for identifying localized generation errors, and a refinement model that can re-synthesize content specifically in these localized regions. We found that an effective Weakly-Supervised error Classifier (WSC) can be constructed with just a few hours of manual labeling of generation errors. Another benefit of this approach is that it can be easily tailored for the errors of a specific model that produces images with artifacts. The resulting WSC will produce an error map highlighting low-quality regions. Subsequently, we adopt an Image-to-Image Schrödinger Bridge " + }, + { + "bbox": [ + 55, + 363, + 295, + 626 + ], + "type": "inline_equation", + "content": "(\\mathrm{I}^2\\mathrm{SB})" + }, + { + "bbox": [ + 55, + 363, + 295, + 626 + ], + "type": "text", + "content": " [23] to learn the refinement of these regions in the generated images. While typical diffusion models map from noise to data, " + }, + { + "bbox": [ + 55, + 363, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 55, + 363, + 295, + 626 + ], + "type": "text", + "content": " constructs a Schrödinger Bridge (SB) that allows us to map from data to data, or in our setup, generations with artifacts to ground truth images. In addition, we introduce an adaptive noise schedule to direct the SB process to focus on the localized errors by incorporating the classifier's prediction error into the noise schedule, which we describe in more detail in Sec. 4.1. As shown in the first row of Fig. 1, our refinement SB model (i.e., EARSB) corrects the distorted text in the initially generated image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 628, + 211, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 628, + 211, + 640 + ], + "spans": [ + { + "bbox": [ + 67, + 628, + 211, + 640 + ], + "type": "text", + "content": "The contributions of our paper are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "content": "- We introduce (human, synthetic garment) pairs as an augmentation in the training of virtual try-on task. The synthetic garment is obtained from our human-to-garment model, which can generate product-view garment images from human images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- We introduce a spatially adaptive Schrödinger Bridge" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 144 + ], + "type": "text", + "content": "model (EARSB) to refine the outputs of a base virtual try-on model. Our formulation incorporates a spatially varying diffusion noise schedule, with noise proportional to the degree of refinement we wish to perform locally. We find this to yield better results than the baseline Schrödinger Bridge framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": "- Extensive experiments on two datasets (VITON-HD [20] and DressCode-Upper [25]) show that EARSB enhances the quality of the images generated by prior work, and is preferred by the users in " + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": " cases on average." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 203, + 400, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 203, + 400, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 203, + 400, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 223, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 474 + ], + "type": "text", + "content": "Training with Synthetic Data. The addition of synthetic data is often an effective means of improving downstream task performance when it is difficult to amass real data at the necessary scale. This has been demonstrated in the domains of image generation [17, 30] and image editing [5, 33]. Careful applications can also be used to ameliorate dataset imbalance issues, as shown in [9]. Other works, such as [1], use self-synthesized data to provide negative guidance for the diffusion model. Our incorporation of synthetic data in the virtual try-on task tackles a specific sub-problem in the broader image editing domain and is similar in spirit to [5, 33]. Specifically, we aim to synthesize paired training data that satisfies the stringent requirements of virtual try-on paired training data – a canonical product-view garment image paired with an example of it being worn. Images of people in clothing are readily available, but it is difficult to obtain a product-view image of the exact clothing they are wearing. To address this, our work tackles the human-to-garment problem, which aims to extract the clothing from a person's photo and project it to the canonical product view, making it roughly symmetric to the virtual try-on task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "content": "Virtual Try-On. There has been a shift from earlier GAN-based framework [14, 20, 22, 29, 35] to diffusion-based methods [7, 18] in the virtual try-on literature. Diffusion models fit an SDE process mapping from the image distribution to the noise distribution, and tend to be easier to train than GAN-based approaches due to the simplicity of the L2 denoising loss [10, 12, 31]. At inference, the diffusion model denoises a random Gaussian noise distribution to a human-readable image via multiple sampling steps. [7, 36] propose parameter-efficient approaches that concatenate the human image and the garment images along the spatial dimension such that the self-attention layer in the denoising UNet can achieve texture transfer without extra parameters. In [18], the authors introduce additional cross-attention layers to learn the semantic correspondences between the garment and the human image. The methods in [3, 21, 27] align different embedding spaces in the attention module to achieve flexible clothing editing after try-on, such as style change or graphics insertion. In contrast to prior work that samples from random noise, we build upon" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21239" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 304, + 201 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 304, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 304, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 304, + 201 + ], + "type": "image", + "image_path": "bc735dfcbe2cb0fed7de7d183d8f99ba89111d1b3a2ebaaf015381c33912280f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 202, + 231, + 212 + ], + "lines": [ + { + "bbox": [ + 132, + 202, + 231, + 212 + ], + "spans": [ + { + "bbox": [ + 132, + 202, + 231, + 212 + ], + "type": "text", + "content": "(a) Human-to-Garment model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 214, + 306, + 335 + ], + "blocks": [ + { + "bbox": [ + 57, + 214, + 306, + 335 + ], + "lines": [ + { + "bbox": [ + 57, + 214, + 306, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 214, + 306, + 335 + ], + "type": "image", + "image_path": "a866f0be82048b99dcd5de44ca60bff340b1cfbd5b09d33289accdc2096578d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 336, + 247, + 346 + ], + "lines": [ + { + "bbox": [ + 116, + 336, + 247, + 346 + ], + "spans": [ + { + "bbox": [ + 116, + 336, + 247, + 346 + ], + "type": "text", + "content": "(b) Examples in H2G-UH and H2G-FH." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 356, + 295, + 389 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 389 + ], + "type": "text", + "content": "Figure 2. (a) Our human-to-garment model, which is explained in Sec. 3.1 (b) Examples of the constructed (human, synthetic garment) pairs in Sec. 3.1." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 486 + ], + "type": "text", + "content": "recent advances in Schrödinger bridges, notably [23], to directly sample from an initial image generated by prior try-on models. Our work is similar in spirit to [37], which initializes the noisy image with a GAN-generated image and small amounts of random noise. However, our work explores varying the local noise schedule based on the error level at a given location." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 499, + 281, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 281, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 281, + 514 + ], + "type": "text", + "content": "3. Augmented Training with Synthetic Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": "Virtual Try-On Task Definition. Let " + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "(x_0, C)" + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": " be the (human image, product-view of worn garment) pair in virtual try-on training. We will refer to " + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "(x_0, C)" + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": " as paired data. Let " + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\bar{x}_0" + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": " be a masked version of " + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": " in which the worn garment corresponding to " + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": " is masked out. We can then set up a learning task in which we aim to fit the following function: " + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "F(\\bar{x}_0, C, \\phi; \\theta) \\to x_0" + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": " corresponds to other conditionals such as pose representations from DensePose [13]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Acquiring high-quality pairs " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x_0, C)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " at scale is challenging due to copyright and brand protection, but acquiring images of just humans " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x_0)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " at scale is considerably more feasible [11, 21, 24, 34]. This observation motivates proposed human-to-garment process to extract a synthetic canonical view image " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ". We can then augment our virtual try-on training with " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x_0, \\hat{C})" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " pair, requiring only single human images. In the following, Sec. 3.1 explains the archi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "tecture of our human-to-garment model, Sec. 3.2 discusses how we use this model to construct the synthetic dataset, and Sec. 3.3 describes how the synthetic data is used to augment the virtual try-on training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 129, + 464, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 464, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 464, + 140 + ], + "type": "text", + "content": "3.1. Human-to-Garment Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 147, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 301 + ], + "type": "text", + "content": "While virtual try-on requires generating skin and deforming the product-view garment to accommodate diverse postures, the human-to-garment task simply aims to map the clothing item to its canonical view. To achieve this, we use existing paired (human, garment) data (e.g., VITON-HD [20]) to train our human-to-garment model. As illustrated in Fig. 2a, we first segment and extract the clothing on the person map and then feed the clothing item to a generator that synthesizes its canonical view. The generator is based on the UNet model proposed in [14], which uses a flow-like mechanism for warping latent features in an optical-flow-like manner. The generator was trained using a combined L1 reconstruction and adversarial loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 311, + 554, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 324 + ], + "type": "text", + "content": "3.2. Constructing Synthetic H2G-UH and H2G-FH" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "text", + "content": "Synthetic images " + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "text", + "content": " produced from our models necessarily contain generation errors. We use the following criteria to filter for high-quality synthetic data: a) The single human image " + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "text", + "content": " has a clean background (low pixel variance in the non-human region); b) " + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "text", + "content": " is frontal view (classified by its DensePose representation [13]); c) the reconstruction error (LPIPS distance) is small when reconstructing the human image " + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "text", + "content": " in a try-on model using the " + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "inline_equation", + "content": "(x_0, \\hat{C})" + }, + { + "bbox": [ + 313, + 328, + 554, + 497 + ], + "type": "text", + "content": " pair (e.g., [20, 29]). Under these criteria, we select human images from DeepFashion2 [24] and UPT [34], eventually creating 12,730 synthetic pairs of upper-body human images (referred to as H2G-UH) and 8,939 pairs of full-body human images (referred to as H2G-FH). Examples of the synthetic pairs are shown in Fig. 2b." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 506, + 506, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 506, + 519 + ], + "type": "text", + "content": "3.3. Augmented Virtual Try-on Training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 523, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 631 + ], + "type": "text", + "content": "To further prevent distribution leakage of incorporating synthetic data, we explore two means of limiting the effect of the real-synthetic domain gap: (a) two training stages involving pretraining the try-on model using synthetic pairs, and then finetuning on real pairs [19]; (b) training simultaneously on real and synthetic data, but conditioning the try-on model on a real/synthetic flag, similar to [16]. We found empirically that the second augmentation performs slightly better than the first (See Sec. 5.1)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 643, + 554, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 671 + ], + "type": "text", + "content": "4. Error-Aware Refinement Schrödinger Bridge" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Apart from the synthetic data augmentation from Sec. 3, our second approach to enhancing existing try-on methods is a refinement pipeline with two steps. First, given some" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21240" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 555, + 300 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 300 + ], + "type": "image", + "image_path": "1a97d47f3f18a1c439164979b9d506dff344fd07bb800734353798f309804de5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": "Figure 3. The diffusion process in our refinement-based EARSB. We first preprocess the input image, then use a base try-on model that takes the masked human image " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\bar{x}_0" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": ", its pose representation " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": ", and its garment " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " as input to generate an initial human image " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " is fed to our weakly-supervised classifier (WSC) to obtain the error map " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " (see Sec. 4.1). This map reweights the noise distribution " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\epsilon^r" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " diffusion and refines " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " that has generation errors to the ground truth image " + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 54, + 308, + 555, + 353 + ], + "type": "text", + "content": " (see Sec. 4.2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": "base model " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "F_{base}(\\bar{x}_0, C, \\phi) \\to x_1" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "F_{base}(.)" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": " can be any pretrained GAN or diffusion-based approach to virtual try-on, " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": " closely approximates the true real human image " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": " with some generation artifacts. To automatically identify the artifacts in " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": ", we construct a weakly-supervised error classifier " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{WSC}(x_1, C) \\to M" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": " as in Fig. 3, where " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": " is a confidence map predicting a heatmap for likely generation errors. Then, our second step performs the final refinement by fitting a Schrödinger bridge based on " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": " [23] via the following mapping: " + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "inline_equation", + "content": "F_{EARSB}(x_1, C, M, \\phi; \\theta) \\to x_0" + }, + { + "bbox": [ + 54, + 362, + 294, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "content": "The approach is weakly inspired by boosting methods in that we wish to fit a targeted refinement model that is trained specifically on the generation errors of an existing model. The refinement goal applies to the general setting where we want to refine a flawed image output, though we focus on virtual try-on for this work. Thus, as illustrated in Fig. 3, the training of EARSB includes three steps:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 568, + 295, + 664 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 568, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 568, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 67, + 568, + 295, + 604 + ], + "type": "text", + "content": "1. Pre-process the images in the training set and feed them to existing try-on models to get the initial images " + }, + { + "bbox": [ + 67, + 568, + 295, + 604 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 67, + 568, + 295, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 604, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 604, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 67, + 604, + 295, + 628 + ], + "type": "text", + "content": "2. Obtain the error maps " + }, + { + "bbox": [ + 67, + 604, + 295, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 604, + 295, + 628 + ], + "type": "text", + "content": " on the initial images " + }, + { + "bbox": [ + 67, + 604, + 295, + 628 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 67, + 604, + 295, + 628 + ], + "type": "text", + "content": " using our WSC (Sec. 4.1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 628, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 628, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 628, + 295, + 664 + ], + "type": "text", + "content": "3. Use " + }, + { + "bbox": [ + 67, + 628, + 295, + 664 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 628, + 295, + 664 + ], + "type": "text", + "content": " to adjust the noise schedule in " + }, + { + "bbox": [ + 67, + 628, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 67, + 628, + 295, + 664 + ], + "type": "text", + "content": " [23] and train the noise prediction model in EARSB following Eq. (9) (Sec. 4.2)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "As the first step simply requires running an off-the-shelf virtual try-on model to obtain an initial image, our discussion in the next section will begin by describing the second step- obtaining the error map." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 362, + 457, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 457, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 457, + 376 + ], + "type": "text", + "content": "4.1. Obtaining the Error Map" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 416 + ], + "type": "text", + "content": "We start by obtaining the error map " + }, + { + "bbox": [ + 313, + 380, + 555, + 416 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 380, + 555, + 416 + ], + "type": "text", + "content": " that highlights the corrupted or incorrect area of the initial image " + }, + { + "bbox": [ + 313, + 380, + 555, + 416 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 380, + 555, + 416 + ], + "type": "text", + "content": " using our proposed Weakly-Supervised error Classifier (WSC)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 416, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 464 + ], + "type": "text", + "content": "Classifier Architecture. As shown in the green dotted box of Fig. 3, our WSC has two encoders to match the image features of " + }, + { + "bbox": [ + 313, + 416, + 554, + 464 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 416, + 554, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 416, + 554, + 464 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 416, + 554, + 464 + ], + "type": "text", + "content": " with cross attention to predict a sigmoid-activated error map." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "content": "Training Data Annotation. In practice, it is labor-intensive to fully annotate all the initial images for where the generated artifacts are located. To mitigate this issue, we used a few hours to hand-label a small portion of the initial try-on images in the training set at the patch level, using bounding boxes for poorly generated regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 536, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 583 + ], + "type": "text", + "content": "Weakly-Supervised Training. Let " + }, + { + "bbox": [ + 313, + 536, + 554, + 583 + ], + "type": "inline_equation", + "content": "x_0, x_1^u, x_1^l" + }, + { + "bbox": [ + 313, + 536, + 554, + 583 + ], + "type": "text", + "content": " be the real human image, the unlabeled initial image, and the labeled initial image with bounding boxes annotating artifacts. Our WSC loss terms are defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 584, + 566, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 584, + 566, + 598 + ], + "spans": [ + { + "bbox": [ + 314, + 584, + 566, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m g} = - \\log \\left(\\mathrm {W S C} \\left(x _ {1} ^ {u}, C\\right) _ {\\max }\\right) + \\log \\left(1 - \\mathrm {W S C} \\left(x _ {0}, C\\right) _ {\\max }\\right)", + "image_path": "0298dfa2f512671cc647958815e9b932ddccf8fe447adc2bd029aecf39f43b89.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 600, + 553, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 553, + 641 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {p a t} = - \\log \\left(\\operatorname {W S C} \\left(x _ {1} ^ {l}, C\\right) \\odot B _ {b o x}\\right) \\\\ - \\log \\left(1 - \\operatorname {W S C} \\left(x _ {1} ^ {l}, C\\right) \\odot \\left(1 - B _ {\\text {b o x}}\\right)\\right) \\tag {1} \\\\ \\end{array}", + "image_path": "94dda0b6deb5a3c7655707cd2d03ffe8424a44a737ab188519ff4e8cedea8a8a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " is the image-level loss and " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pat}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " is the patch-level loss. In " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{WSC}(\\cdot)" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " is the output error map and " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{WSC}(\\cdot)_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " denotes the spatially max-pooled score in the error map. In " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pat}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "B_{box}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " is the spatial binary mask for the annotated regions, thereby maximizing and minimizing the scores for regions within and outside of the annotated boxes" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21241" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 268, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 268, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 268, + 85 + ], + "type": "text", + "content": "respectively. Our final loss is: " + }, + { + "bbox": [ + 55, + 72, + 268, + 85 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{WSC}} = \\mathcal{L}_{ins} + \\mathcal{L}_{pat}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 295, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 295, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 295, + 122 + ], + "type": "text", + "content": "The trained WSC will predict an error map " + }, + { + "bbox": [ + 55, + 86, + 295, + 122 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 86, + 295, + 122 + ], + "type": "text", + "content": " for the initial image " + }, + { + "bbox": [ + 55, + 86, + 295, + 122 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 86, + 295, + 122 + ], + "type": "text", + "content": ", which is then used to adjust the noise schedule in the diffusion process described in the next section." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 134, + 266, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 266, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 266, + 148 + ], + "type": "text", + "content": "4.2. Error-Map-Reweighted SB Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": "To achieve the refinement goal, our diffusion process extends Schrödinger bridges as formulated in " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " [23], where we incrementally add noise to the initial image " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": ", and then remove the noise to approximate the refined image " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": ". However, without additional information, a naively trained " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " model must implicitly learn what to refine and what to retain. Our formulation aims to explicitly incorporate prior knowledge of localized generation errors via the error map " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " into the Schrödinger process by using " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " to locally scale the noise schedule for the Schrödinger process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 407 + ], + "type": "text", + "content": "Our choice of locally scaling the noise schedule is based on several observations. We want the model to directly copy pixels over to " + }, + { + "bbox": [ + 55, + 274, + 296, + 407 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 274, + 296, + 407 + ], + "type": "text", + "content": " for correctly generated regions in " + }, + { + "bbox": [ + 55, + 274, + 296, + 407 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 55, + 274, + 296, + 407 + ], + "type": "text", + "content": " - it would be nice to avoid training the model to add and remove noise from these regions. In contrast, erroneous regions in the initial try-on images, especially more noticeable ones, include generation errors that share little to no structural similarity to the target. These errors include examples such as deformed limbs, and distorted textures/fabrics, which may need more added noise to prevent the model from conditioning too strongly on the original pattern." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": "As such, we construct our refinement model " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "F_{EARSB}(x_1,C,M,P;\\theta) \\to x_0" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": ", where the model is conditioned on the canonical view garment " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": ", the error map " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": ", and the pose representation " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": ". Fig. 3 shows our Weakly-Supervised Classifier (WSC) first locates the errors in the error map " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " reweights the noise schedule of the I²SB stochastic process to assign a higher volume of noise to the low-quality region \"rebel\" so the model can focus on refining it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 566 + ], + "type": "text", + "content": "Error-Map-Reweighted Diffusion Process. Following I²SB [23], our diffusion Schrödinger bridge maps from the initial image " + }, + { + "bbox": [ + 55, + 517, + 296, + 566 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 517, + 296, + 566 + ], + "type": "text", + "content": " to the ground truth image " + }, + { + "bbox": [ + 55, + 517, + 296, + 566 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 55, + 517, + 296, + 566 + ], + "type": "text", + "content": ". It fits to the following stochastic process:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 568, + 295, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 568, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 66, + 568, + 295, + 611 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t} = \\mu_ {t} \\left(x _ {0}, x _ {1}\\right) + \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon \\\\ \\mu_ {t} = \\frac {\\bar {\\sigma} _ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} x _ {0} + \\frac {\\sigma_ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} x _ {1}, \\Sigma_ {t} = \\frac {\\bar {\\sigma} _ {t} ^ {2} \\sigma_ {t} ^ {2}}{\\bar {\\sigma} _ {t} ^ {2} + \\sigma_ {t} ^ {2}} \\cdot I, \\tag {2} \\\\ \\end{array}", + "image_path": "093c2431d845225be4f8912dd13329cafa0a0eb0257e018113dd0adf3c57e9de.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\sigma_t^2 = \\int_0^t\\beta_\\tau d\\tau" + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\bar{\\sigma}_t^2 = \\int_t^1\\beta_\\tau d\\tau" + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\beta_{\\tau}" + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "content": " is a symmetrical noise schedule. " + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "content": " is random Gaussian noise. The above equation stochastically adds noise and then removes it between " + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 614, + 295, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We extend " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " such that the noise schedule can vary spatially based on the error map " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " (obtained from WSC in Sec. 4.1). Good regions will be assigned less noise (i.e., smaller variance) in the diffusion process, while poor qual" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 453, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 453, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 453, + 84 + ], + "type": "text", + "content": "ity regions will be assigned more:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 87, + 553, + 101 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 87, + 553, + 101 + ], + "spans": [ + { + "bbox": [ + 376, + 87, + 553, + 101 + ], + "type": "interline_equation", + "content": "x _ {t} = \\mu_ {t} \\left(x _ {0}, x _ {1}\\right) + \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon^ {r}, \\tag {3}", + "image_path": "c06dfa5dc74fdcf62234befcd48d339c5c7707e42c7f19681150a34a1817c6ce.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 103, + 553, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 103, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 372, + 103, + 553, + 116 + ], + "type": "interline_equation", + "content": "\\epsilon^ {r} = M \\cdot \\epsilon , M = \\operatorname {W S C} \\left(x _ {1}, C\\right) \\tag {4}", + "image_path": "dc4c908ac978288c6a3b0bf6612a7a921549948dc1cd1b2b675a7d59ec8af662.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 118, + 553, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 553, + 130 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 118, + 553, + 130 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 313, + 118, + 553, + 130 + ], + "type": "text", + "content": " is the same as Eq. (2) and " + }, + { + "bbox": [ + 313, + 118, + 553, + 130 + ], + "type": "inline_equation", + "content": "\\epsilon^r" + }, + { + "bbox": [ + 313, + 118, + 553, + 130 + ], + "type": "text", + "content": " is the adaptive noise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": "Sampling Process. The initial image " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": " is iteratively refined to " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": " via a denoising/sampling process, where a model predicts the noise distribution at each time step. In contrast to prior soft-attention-based UNets [18, 26, 36, 37], our denoising model uses cloth-flow-learning UNet for more precise garment deformation [14]. It accepts the garment " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": ", the error map " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": ", the pose representation " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": ", and the noisy image " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": " as inputs and predicts the error-adapted noise " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{r}(\\cdot ;t)" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "(\\cdot ;t)" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": " omits the inputs " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "M, P, x_{t}, C" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": ". See Supp. B for the detailed model architecture. With the predicted noise " + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{r}(\\cdot ;t)" + }, + { + "bbox": [ + 313, + 137, + 555, + 269 + ], + "type": "text", + "content": ", we define our sampling process:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 272, + 553, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 272, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 361, + 272, + 553, + 285 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {0} = x _ {t} - \\sqrt {\\Sigma_ {t}} \\cdot \\epsilon_ {\\theta} ^ {r} (M, P, x _ {t}, C; t) \\tag {5}", + "image_path": "ec21d2647648c3db55c8d2f25124e4ae0020c046054e0ebfd6497a069dbb9cbd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 289, + 553, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 289, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 354, + 289, + 553, + 309 + ], + "type": "interline_equation", + "content": "x _ {t - \\Delta t} = \\hat {\\mu} _ {t - \\Delta t} \\left(\\hat {x} _ {0}, x _ {t}\\right) + M \\cdot \\sqrt {\\hat {\\Sigma} _ {t}} \\cdot \\epsilon \\tag {6}", + "image_path": "2ea4fc7cd3857a9fe07ade45d6c767a97490780fb4b4ec67d4da60de578b7506.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 311, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 311, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 311, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\hat {\\mu} _ {t - \\Delta t} = \\frac {\\sigma_ {t - \\Delta t} ^ {2}}{\\sigma_ {t} ^ {2}} \\hat {x} _ {0} + \\frac {\\sigma_ {t} ^ {2} - \\sigma_ {t - \\Delta t} ^ {2}}{\\sigma_ {t} ^ {2}} x _ {t} \\tag {7}", + "image_path": "3854e7bda64e29f4ee6c2b4a89ca702c2884a3f62159ac0b1369de8ebe9ed230.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 380, + 339, + 553, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 339, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 380, + 339, + 553, + 366 + ], + "type": "interline_equation", + "content": "\\hat {\\Sigma} _ {t} = \\frac {\\sigma_ {t - \\Delta t} ^ {2} \\left(\\sigma_ {t} ^ {2} - \\sigma_ {t - \\Delta t} ^ {2}\\right)}{\\sigma_ {t} ^ {2}} \\tag {8}", + "image_path": "e37bbc8ef644786d37b5d99078cfa5228da71aefdd67649575b79c8a96510f57.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "\\Delta t > 0" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " and it is the sampling interval. Starting from " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": ", the process iteratively refines the initial human image " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " based on the error map " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " is all ones in Eq. (5), our model reverts to the " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " formulation. When " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " is all zeros (i.e., no error), " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " is believed to be perfect " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " does not need to be refined in the sampling process." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "type": "text", + "content": "The training objective of our model is the mean squared error between the predicted noise " + }, + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{r}" + }, + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "type": "text", + "content": " and the reweighted Gaussian noise " + }, + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "type": "inline_equation", + "content": "\\epsilon^{r}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 329, + 484, + 553, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 484, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 329, + 484, + 553, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E A R S B}} = \\mathbb {E} _ {t \\sim U (0, 1)} | | \\epsilon_ {\\theta} ^ {r} (M, P, x _ {t}, C; t) - \\epsilon^ {r} | | ^ {2} \\tag {9}", + "image_path": "9a19b5d232266331bfb3ca2289ec9459c34edcdd95a2775d9a87234014733804.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 522, + 553, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 547 + ], + "type": "text", + "content": "4.2.1. Further Improvements via Classifier Guidance and Expert Denoisers" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "text", + "content": "Whereas prior work used an object category classifier to guide the sampling process [10], our WSC guidance gives a direction toward the real data distribution. Chung et al. [8] shows we can estimate the guidance score " + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p(y|x_t)" + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "text", + "content": " using the denoised clean image " + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p(y|x_t)\\simeq \\nabla_{x_t}\\log p(y|\\hat{x}_0)" + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 556, + 554, + 640 + ], + "type": "text", + "content": " is the fake/real label. Since the label for real data is 0 in WSC, the classifier guidance is:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 334, + 643, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 643, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 334, + 643, + 553, + 656 + ], + "type": "interline_equation", + "content": "\\hat {\\mu} _ {t - \\Delta t} \\leftarrow \\hat {\\mu} _ {t - \\Delta t} + M \\cdot \\hat {\\Sigma} _ {t} \\cdot \\nabla_ {x _ {t}} \\log p (\\mathbf {0} | \\hat {x} _ {0}) \\tag {10}", + "image_path": "dd08bbb3e1054d5f6a53c672fc375cf7bf27fd4b12eeac4f80757d0684f915c4.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 658, + 460, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 460, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 460, + 672 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 658, + 460, + 672 + ], + "type": "inline_equation", + "content": "p(\\mathbf{0}|\\hat{x}_0) = 1 - \\mathrm{WSC}(\\hat{x}_0,C)" + }, + { + "bbox": [ + 313, + 658, + 460, + 672 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Following [2], a trained EARSB model is split into two models, each having an expert denoiser that is fine-tuned on denoising ranges " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "t \\in [0, 0.5]" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "t \\in [0.5, 1]" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21242" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 542, + 300 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 542, + 300 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 542, + 300 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 542, + 300 + ], + "type": "table", + "html": "
UnpairedVITON-HDDressCode-Upper
PairedUnpairedPaired
FID↓KID↓FID↓KID↓SSIM↑LPIPS↓FID↓KID↓FID↓KID↓SSIM↑
(a) GAN-Based
HR-VTON [20]10.750.288.460.260.9010.07515.260.3911.760.320.947
SD-VTON [29]9.050.126.470.090.9070.07014.730.3210.990.240.947
GP-VTON [35]8.610.865.530.070.9130.06426.191.7123.661.590.816
EARSB (ours)8.420.075.250.050.9180.05910.890.137.150.130.961
EARSB +H2G-UH/FH (ours)8.260.065.140.040.9190.05810.700.117.050.110.965
(b) SD-Based
LaDI-VTON [26]8.950.126.050.080.9020.07114.880.3911.610.320.939
CatVTON [7]8.870.085.490.070.9150.05911.910.217.660.100.950
CAT-DM [37]8.550.105.980.070.9080.06712.910.298.580.160.948
IDM-VTON [6]8.590.115.510.090.9020.06111.090.166.790.120.956
TPD [36]8.230.064.860.040.9170.057-
StableVITON [18]8.200.075.160.050.9170.057-
EARSB(SD) +H2G-UH/FH (ours)8.040.064.900.030.9250.05310.410.096.760.080.968
", + "image_path": "21610391de687195a77c7934e86820f65779137986bcd97e443d5a16e8c7066d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 337, + 137, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 137, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 137, + 350 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 358, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 358, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 295, + 515 + ], + "type": "text", + "content": "Datasets. We use VITON-HD [20], DressCode-Upper [25], and our synthetic H2G-UH and H2G-FH for training. They include 11,647, 13,564, 12,730, 8,939 training images, respectively. For synthetic data augmentation, we combine VITON-HD with our H2G-UH since they both include mostly upper-body human images. DressCode-Upper is combined with H2G-FH as both consist of full-body human photos. For evaluation, VITON-HD contains 2,032 (human, garment) test pairs and DressCode-Upper has 1,800 test pairs and include both paired and unpaired settings. In the paired setting, the input garment image and the garment in the human image are the same item. Conversely, the unpaired setting uses a different garment image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "content": "Metrics. We use Structural Similarity Index Measure (SSIM) [32], Frechet Inception Distance (FID) [15], Kernel Inception Distance (KID) [4], and Learned Perceptual Image Patch Similarity (LPIPS) [38] to evaluate image quality. All the compared methods use the same image size " + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "content": " and padding when computing the above metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 591, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 663 + ], + "type": "text", + "content": "Experimental setup. We compare EARSB with GAN-based methods HR-VTON [20], SD-VTON [29] and GPVTON [35], as well as Stable Diffusion (SD) based methods including CAT-DM [37], StableVITON [18], TPD [36], IDM-VTON [6] and CatVTON [7]. Unless otherwise specified, all diffusion models use 25 sampling steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 712 + ], + "type": "text", + "content": "We report results of multiple variants of our approach. EARSB uses GAN-based GP-VTON [35] to generate the initial image that was trained without synthetic data augmentation. EARSB+H2G-UH/FH trains with either H2G-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 336, + 545, + 384 + ], + "blocks": [ + { + "bbox": [ + 55, + 304, + 553, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 553, + 327 + ], + "type": "text", + "content": "Table 1. Virtual try-on results on VITON-HD [20] and DressCode-Upper [25] for (a) GAN-based and (b) diffusion-based models using 25 sampling steps. KID is multiplied by 100. We find our EARSB approach outperforms prior work on average. See Sec. 5.1 for discussion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 336, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 320, + 336, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 320, + 336, + 545, + 384 + ], + "type": "table", + "html": "
MethodsGP-VTON EARSBStableVITON EARSB
Consistency42%58%38%62%
Fidelity39%61%45%55%
", + "image_path": "3860d04fbf53713fb67a5e86e8d15a85dde7f91ee347191df8e7e87f7ca4461d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 388, + 553, + 410 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 410 + ], + "type": "text", + "content": "Table 2. User studies on VITON-HD. Our EARSB method is preferred in an average of " + }, + { + "bbox": [ + 313, + 388, + 553, + 410 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 313, + 388, + 553, + 410 + ], + "type": "text", + "content": " cases." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 412, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 553, + 472 + ], + "type": "text", + "content": "UH or H2G-FH. We add the upper-body synthetic subset H2G-UH for the upper-body-human dataset VITON-HD, and the full-body synthetic H2G-FH when on DressCodeUpper. EARSB(SD)+H2G-UH/FH uses the diffusion-based CatVTON [7] to generate the initial image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 480, + 371, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 480, + 371, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 480, + 371, + 491 + ], + "type": "text", + "content": "5.1. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 498, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 617 + ], + "type": "text", + "content": "Tab. 1 compares our approach with those from prior work. We find our full model variants EARSB+H2G-UH/FH and EARSB(SD)+H2G-UH/FH boosts performance over the GAN and SD-based methods, respectively. The last two rows of Tab. 1(a) show that incorporating our synthetic training pairs provide a consistent boost in performance on both datasets. Comparing the last rows of Tab. 1(a) and Tab. 1(b) we observe that using the diffusion model to generate the initial image gives better performance in EARSB(SD)+H2G-UH/FH, but is more costly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "User Study. We asked Amazon MTurk workers to evaluate the texture consistency and image fidelity of synthesized images, comparing our model against GP-VTON and StableVITON. We randomly selected 100 pairs from VITON-HD to evaluate on, assigning at least 3 workers per image. Study results in Tab. 2 report our method is preferred at least " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " more than the GAN-based GP-VTON and the SD-based StableVITON (59% overall)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21243" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 91, + 259, + 194 + ], + "blocks": [ + { + "bbox": [ + 115, + 72, + 172, + 84 + ], + "lines": [ + { + "bbox": [ + 115, + 72, + 172, + 84 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 172, + 84 + ], + "type": "text", + "content": "Unpaired FID" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 91, + 259, + 194 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 259, + 194 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 259, + 194 + ], + "type": "image", + "image_path": "e65258728eb019c37b994906975e059994c4ee72d62edc40b57937f329a02678.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 220 + ], + "type": "text", + "content": "Figure 4. Results on VITON-HD at 5, 10, 25, 50, and 100 sampling steps. Our method consistently improves our baseline starting model GP-VTON (black, dotted line), making it competitive with StableVITON (especially at under 50 sampling steps). Legend is shared for all." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 259, + 91, + 397, + 194 + ], + "blocks": [ + { + "bbox": [ + 316, + 75, + 362, + 85 + ], + "lines": [ + { + "bbox": [ + 316, + 75, + 362, + 85 + ], + "spans": [ + { + "bbox": [ + 316, + 75, + 362, + 85 + ], + "type": "text", + "content": "Paired FID" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 259, + 91, + 397, + 194 + ], + "lines": [ + { + "bbox": [ + 259, + 91, + 397, + 194 + ], + "spans": [ + { + "bbox": [ + 259, + 91, + 397, + 194 + ], + "type": "image", + "image_path": "83f869cd286951474efcd491f9ea0d87dbd916af9470854750c7e64028f3938e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 400, + 91, + 550, + 194 + ], + "blocks": [ + { + "bbox": [ + 466, + 75, + 520, + 85 + ], + "lines": [ + { + "bbox": [ + 466, + 75, + 520, + 85 + ], + "spans": [ + { + "bbox": [ + 466, + 75, + 520, + 85 + ], + "type": "text", + "content": "Paired SSIM" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 91, + 550, + 194 + ], + "lines": [ + { + "bbox": [ + 400, + 91, + 550, + 194 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 550, + 194 + ], + "type": "image", + "image_path": "d476eb027df389ed00b1301f4df47f7193d704f7adfa5eef0fbd85bd07908158.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 222, + 553, + 387 + ], + "blocks": [ + { + "bbox": [ + 57, + 222, + 553, + 387 + ], + "lines": [ + { + "bbox": [ + 57, + 222, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 57, + 222, + 553, + 387 + ], + "type": "image", + "image_path": "57fde6822d1365bbef2bd0a29506996da1076e8ab9d9a559ef2e97bc6f88ef45.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 394, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 555, + 417 + ], + "type": "text", + "content": "Figure 5. Visualizations on VITON-HD (top row) and DressCode (bottom row). Our EARSB+H2G-UH and EARSB(SD)+H2G-UH better recover the intricate textures in the garment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 422, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 295, + 566 + ], + "type": "text", + "content": "Trade-Off Between Sampling Efficiency and Image Quality. Diffusion-generated images often show degraded quality with fewer sampling steps. In EARSB, the spatially adaptive noise schedule can preserve correct clothing textures in the initial image with a low noise level and only fix the erroneous parts, potentially resulting in less image quality degradation with fewer steps. In Fig. 4, while other SD-based methods have a sharp performance drop with decreasing sampling steps, EARSB and EARSB+H2G-UH show consistent performance across different sampling steps, demonstrating a better trade-off between image quality and computational efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 567, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 662 + ], + "type": "text", + "content": "Qualitative Results. Fig. 5 gives examples of generated images from different approaches. The top row is from VITON-HD dataset and the bottom row is from DressCodeUpper. The third images in the two rows are GAN-generated results. We see that our EARSB+H2G-UH/FH in the last column improves the low-quality textures from the GAN-generated images, which are the distorted graphics in the center. Additional visualizations are in Supp. G." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "type": "text", + "content": "5.2. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Synthetic Pairs Augmentation. Tab. 3 incorporates H2GUH into the training of StableVITON [18] and CAT-DM" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 314, + 421, + 560, + 523 + ], + "blocks": [ + { + "bbox": [ + 314, + 421, + 560, + 523 + ], + "lines": [ + { + "bbox": [ + 314, + 421, + 560, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 560, + 523 + ], + "type": "table", + "html": "
Data Aug.Unpaired FID↓Paired FID↓SSIM↑LPIPS↓
CAT-DM [37]None8.565.900.9110.067
CAT-DMH2G-UH8.365.670.9130.063
StableVITON [18]None8.255.150.9170.056
StableVITONH2G-UH8.175.040.9190.054
EARSB(SD)H2G-UH8.044.900.9250.053
", + "image_path": "8c23edea79ed0df040465b5474751b33b6968b60f0efe4c1ec839249b897cc5d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 555, + 594 + ], + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 594 + ], + "type": "text", + "content": "Table 3. Comparing the effect of our H2G-UH data augmentation approach on VITON-HD. We bold the best overall results and underline the best results for a single model with and without H2G-UH. Each method uses the number of sampling steps from their paper: 2 for CAT-DM, 50 for StableVITON, and 25 for our own EARSB. We find H2G-UH consistently boosts performance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "type": "text", + "content": "[37] on the VITON-HD dataset to validate the effectiveness of synthetic pairs on enhancing existing diffusion methods. We observe that training with our synthetic H2G-UH is effective in improving most metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Tab. 4 explores different ways of incorporating the synthetic data H2G-UH during the training of EARSB. We find that using a randomly warped version of the clothing " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "(W(\\mathrm{H}2\\mathrm{G}-\\mathrm{U}\\mathrm{H}))" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " hinders performance, demonstrating the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21244" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 272, + 171 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 272, + 171 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 272, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 272, + 171 + ], + "type": "table", + "html": "
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
None8.425.250.9180.059
W(H2G-UH)8.685.440.9090.063
plain H2G-UH9.646.520.9020.073
pre. H2G-UH8.355.180.9180.059
H2G-UH8.265.140.9190.058
", + "image_path": "3462a3cbfbd624b48fe01a109cf5958c83939e776755b70e74c393c199587c4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 276, + 280, + 378 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 294, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 294, + 262 + ], + "type": "text", + "content": "Table 4. Ablations of our H2G-UH augmentation on VITON-HD. Specifically, pre. H2G-UH is pretrained using the synthetic pairs and finetuned on real data, " + }, + { + "bbox": [ + 55, + 175, + 294, + 262 + ], + "type": "inline_equation", + "content": "W(\\mathbf{H2G - UH})" + }, + { + "bbox": [ + 55, + 175, + 294, + 262 + ], + "type": "text", + "content": " replaces the synthetic garment in each pair with a randomly warped version of the clothing cropped from the real human image, plain H2G-UH is trained using the mixed distribution of the real and synthetic pairs without the augmentation label identifying them, and H2G-UH uses the mixed data with the identifying label. See Sec. 5.2 for discussion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 276, + 280, + 378 + ], + "lines": [ + { + "bbox": [ + 67, + 276, + 280, + 378 + ], + "spans": [ + { + "bbox": [ + 67, + 276, + 280, + 378 + ], + "type": "table", + "html": "
UnpairedPaired
FID↓FID↓SSIM↑LPIPS↓
Inpainting9.266.330.9090.068
EARSB (w.o. M)9.216.270.9120.061
EARSB (rand(M))9.136.550.9020.071
EARSB (w.o. CG)8.485.320.9180.059
EARSB (full)8.425.250.9180.059
", + "image_path": "c012b101f541b93b1fc390eba36a999b70f43a247fe4df2ccedccbd0bf678f10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 295, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 381, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 295, + 437 + ], + "type": "text", + "content": "Table 5. Comparing noise scheduling strategies on VITON-HD. We include a simple inpainting baseline, not using the error map during training (w.o. " + }, + { + "bbox": [ + 55, + 381, + 295, + 437 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 381, + 295, + 437 + ], + "type": "text", + "content": "), a random error map " + }, + { + "bbox": [ + 55, + 381, + 295, + 437 + ], + "type": "inline_equation", + "content": "(\\mathrm{rand}(M))" + }, + { + "bbox": [ + 55, + 381, + 295, + 437 + ], + "type": "text", + "content": ", and removing classifier guidance (w.o. CG). These results demonstrate the importance of using a meaningful error map." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": "importance of the synthetic product-view. Additionally, the poor plain H2G-UH results indicate the presence of a synthetic-real domain gap when using H2G-UH. Inspired by [28], one way to address this issue is by pretraining on the synthetic data and finetuning on real samples (see pre. H2G-UH). However, we find it most effective to condition on a synthetic data indicator while training on mixed data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Error-Aware Noise Schedule. Tab. 5 explores the importance of the error-aware noise schedule described in Sec. 4.2, where the error map adapts the noise distribution according to the quality of the image patches in the initial image " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". This adaptive approach contrasts with a uniform Gaussian noise application across all locations, which would reduce our model to " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2\\mathrm{SB}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". We consider three baselines: Inpainting regenerates rather than refines erroneous regions (where the mean confidence " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " of an image patch containing an error is greater than 0.5), w.o. " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " removes the error map during training, and " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{rand}(M)" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " indicates a random error map. As shown in Tab. 5, EARSB outperforms all these baselines, underscoring the importance of a meaningful error map in precisely locating and enhancing" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 543, + 171 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 543, + 171 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 543, + 171 + ], + "type": "image", + "image_path": "e3449d3c1a0689566e3e505a7aeb64ce903b67d6f61e178ddb0a9e52b1303762.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 177, + 553, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 221 + ], + "type": "text", + "content": "Figure 6. The precision-recall curve for retrieving annotated errors at the pixel level, comparing our WSC to two fully unsupervised baselines (UC, CC). WSC performs best at retrieving generation artifacts at a nominal labeling cost." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "text", + "content": "ing targeted regions. Additionally, the slight performance degradation observed when removing the classifier guidance (EARSB(w.o. CG)) suggests that the error map employed in our classifier guidance also contributes to overall image quality improvement. Collectively, these findings highlight the crucial role of our adaptive noise schedule in achieving superior results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "type": "text", + "content": "Weakly-Supervised error Classifier (WSC). Our weakly supervised classifier from Sec. 4.1 highlights low-quality regions in the initial image " + }, + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "type": "text", + "content": " with only a few hours of labeling. To validate its effectiveness, we train two ablations of our WSC: the Unsupervised Classifier (UC) that only uses image labels (i.e., fake or real), and the Fake/Real Composite Classifier (CC). CC uses both image fake/real labels as well as fake region-level labels which are created by compositing real image patches and fake image patches. The compositing is a fully automatic alternative to manual labeling that provides patch-level labels. We annotated 100 images in the test set to validate their effectiveness. Fig. 6 shows the pixel-level precision-recall curve for retrieving annotated artifact pixels within the bounding boxes using the classifiers' confidence maps. It is clear that weak supervision remains an incredibly cost-effective approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "This paper addresses two shortcomings of prior work on virtual try-on. First, we address the limited data availability by introducing a human-to-garment model that generates (human, synthetic garment) pairs from a single image of a clothed individual. Second, we propose a refinement model EARSB that surgically targets localized generation errors from the output of a prior model. EARSB improves the low-quality region of an initially generated image based on a spatially-varying noise schedule that targets known artifacts. Experiments on two benchmark datasets demonstrate that our synthetic data augmentation improves the performance of prior work and that EARSB enhances the overall image quality." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21245" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 157 + ], + "type": "text", + "content": "Acknowledgments This material is based upon work supported, in part, by DARPA under agreement number HR00112020054 and the National Science Foundation under Grant No. DBI-2134696. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the supporting agencies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 169, + 115, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 169, + 115, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 169, + 115, + 182 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 189, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 189, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 189, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 189, + 296, + 233 + ], + "type": "text", + "content": "[1] Sina Alemohammad, Ahmed Imtiaz Humayun, Shruti Agarwal, John Collomosse, and Richard Baraniuk. Self-improving diffusion models with synthetic data. arXiv preprint arXiv:2408.16333, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 235, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 235, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 235, + 296, + 300 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 302, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 302, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 302, + 296, + 346 + ], + "type": "text", + "content": "[3] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 348, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 348, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 348, + 296, + 380 + ], + "type": "text", + "content": "[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 382, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 382, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 382, + 296, + 415 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 417, + 296, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 417, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 62, + 417, + 296, + 450 + ], + "type": "text", + "content": "[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. In ECCV, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 452, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 452, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 452, + 296, + 506 + ], + "type": "text", + "content": "[7] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 508, + 296, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 508, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 508, + 296, + 541 + ], + "type": "text", + "content": "[8] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 543, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 543, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 543, + 296, + 586 + ], + "type": "text", + "content": "[9] Damien Dablain, Bartosz Krawczyk, and Nitesh V Chawla. Deepsmote: Fusing deep learning and smote for imbalanced data. IEEE Transactions on Neural Networks and Learning Systems, 34(9):6390-6404, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 589, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 589, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 589, + 295, + 611 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "type": "text", + "content": "[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[13] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In CVPR, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "text", + "content": "[14] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In CVPR, 2019. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "text", + "content": "[16] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In ICML, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "type": "text", + "content": "[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[18] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. StableVITON: Learning semantic correspondence with latent diffusion model for virtual try-on. In CVPR, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[19] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[20] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[21] Nannan Li, Qing Liu, Krishna Kumar Singh, Yilin Wang, Jianming Zhang, Bryan A Plummer, and Zhe Lin. UniHuman: A unified model for editing human images in the wild. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "text", + "content": "[22] Zhi Li, Pengfei Wei, Xiang Yin, Zejun Ma, and Alex C Kot. Virtual try-on with pose-garment keypoints guided inpainting. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "type": "text", + "content": "[23] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I2sb: image-to-image schrodinger bridge. In ICML, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[24] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 556, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 600 + ], + "type": "text", + "content": "[25] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[26] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. LaDIVTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On. In ACM Multimedia, 2023. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[27] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In CVPR, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[28] Maan Qraitem, Kate Saenko, and Bryan A. Plummer. From fake to real: Pretraining on balanced synthetic images to pre" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21246" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 533 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "vent spurious correlations in image recognition. In The European Conference on Computer Vision (ECCV), 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "text", + "content": "[29] Sang-Heon Shim, Jiwoo Chung, and Jae-Pil Heo. Towards squeezing-averse virtual try-on via sequential deformation. In AAAI, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "text", + "content": "[30] C Shivashankar and Shane Miller. Semantic data augmentation with generative models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 153, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 184 + ], + "type": "text", + "content": "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "text", + "content": "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[33] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[34] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive gan. NeurIPS, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 322, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 375 + ], + "type": "text", + "content": "[35] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. GPVTON: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In CVPR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[36] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In CVPR, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[37] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In CVPR, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[39] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. TryOnDiffusion: A tale of two unets. In CVPR, 2023. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21247" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_content_list.json b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f5b9bbb1699b08ff8472590c45c5c75a396bce3c --- /dev/null +++ b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_content_list.json @@ -0,0 +1,1787 @@ +[ + { + "type": "text", + "text": "Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data", + "text_level": 1, + "bbox": [ + 112, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoxin Li and Boyang Li", + "bbox": [ + 393, + 203, + 602, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nanyang Technological University", + "bbox": [ + 359, + 222, + 635, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{haoxin003,boyang.li}@ntu.edu.sg", + "bbox": [ + 351, + 241, + 640, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paired image-text data with subtle variations in-between (e.g., people holding surfboards vs. people holding shovels) hold the promise of producing Vision-Language Models with proper compositional understanding. Synthesizing such training data from generative models is a highly coveted prize due to the reduced cost of data collection. However, synthesizing training images for compositional learning presents three challenges: (1) efficiency in generating large quantities of images, (2) text alignment between the generated image and the caption in the exact place of the subtle change, and (3) image fidelity in ensuring sufficient similarity with the original real images in all other places. We propose SPARCL (Synthetic Perturbations for Advancing Robust Compositional Learning), which integrates image feature injection into a fast text-to-image generative model, followed by an image style transfer step, to meet the three challenges. Further, to cope with any residual issues of text alignment, we propose an adaptive margin loss to filter out potentially incorrect synthetic samples and focus the learning on informative hard samples. Evaluation on four compositional understanding benchmarks demonstrates that SPARCL significantly improves the compositionality of CLIP, boosting the average accuracy of the CLIP base model by over $8\\%$ across all benchmarks and outperforming state-of-the-art methods by $2\\%$ on three benchmarks.", + "bbox": [ + 88, + 325, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 752, + 220, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current Vision-Language Models (VLMs) still face limitation in accurately interpreting compositional relationships between objects and attributes, as demonstrated by numerous evaluations [32, 62, 88, 106, 113]. This limitation primarily stems from the absence of subtle variations in the training data [42] (e.g., the subtle variations between the two captions in Figure 1 (a)). As a result, it becomes possible to maximize empirical image-caption alignment us", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Source caption: Two people carrying surfboards on a beach.", + "bbox": [ + 519, + 289, + 841, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Target caption: Two people carrying shovels on a beach.", + "bbox": [ + 521, + 301, + 820, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8364424f094b3ea50439a4f9064f100a2481797dd7a8b5cd896d37e2ebd8d57.jpg", + "image_caption": [ + "Source image" + ], + "image_footnote": [], + "bbox": [ + 521, + 313, + 635, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8dc5e82fe8dfcae27e4391674bb6a2c65ccbc04cc89d6d54e04638c132aac93b.jpg", + "image_caption": [ + "Edited image" + ], + "image_footnote": [], + "bbox": [ + 653, + 311, + 767, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/782b2f58781d01e515a527e3f48d9ef639c6c2d869b49b7e80e5ccace4429e55.jpg", + "image_caption": [ + "Generated image" + ], + "image_footnote": [], + "bbox": [ + 785, + 311, + 900, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a) Difficulty in accurately creating precise variations.", + "bbox": [ + 566, + 411, + 852, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/179fbc60fa9f1bcd7a064bc872ed467a961e3dd099e5b33c20c917d1e3dee137.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 433, + 633, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "+Real positive caption: A man is stirring a silver pot filled with food.", + "bbox": [ + 517, + 523, + 630, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- Synthetic negative caption: A man is stirring a copper pot filled with food.", + "bbox": [ + 519, + 551, + 630, + 587 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3273596626e139739b29826575bfa9910cdcb4118cc8bb36b163d658d3ad7c47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 431, + 767, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "+Real positive caption: Random people sitting in a public transport bus.", + "bbox": [ + 651, + 523, + 763, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- Synthetic negative caption: Random animals riding a purple elephant.", + "bbox": [ + 653, + 551, + 750, + 587 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9bf9a3ebee93449cdb354a66b8a537523b9f487963f1d0a693caa97c09dc5b9.jpg", + "image_caption": [ + "Easy Sample" + ], + "image_footnote": [], + "bbox": [ + 651, + 597, + 676, + 612 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/254e148fec8e8f637db25798d94eddb4b471ba89022de0918338303651080fa1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 431, + 901, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "+Real positive caption: A double decker bus parked at a station.", + "bbox": [ + 784, + 523, + 895, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "+ Synthetic positive caption: A double-decker bus parked at a bus station with a few passengers getting on.", + "bbox": [ + 784, + 551, + 898, + 597 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f5f758e2e76f32567ea946c823a4658c96d37f424f9b23f6b426fc32568da5a6.jpg", + "image_caption": [ + "Wrong Generation" + ], + "image_footnote": [], + "bbox": [ + 785, + 597, + 810, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(b) Inconsistency in cross-modal alignment quality of synthetic samples.", + "bbox": [ + 517, + 614, + 900, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Challenges in generating and training on synthetic data: (a) When generating an image with subtle variations based on a real image and a target caption specifying the variations, an image editing model [6] struggles with text alignment (middle), while an image generation model [75] fails to maintain image fidelity (right). (b) Synthetic positive and negative image_caption pairs show different levels of alignment quality. The subtle variations in the synthetic negative caption (left) make it difficult to distinguish from the positive; the over-modified negative caption (middle) is easy to distinguish; and the hallucinated content in the synthetic positive caption (right) results in an incorrect positive.", + "bbox": [ + 511, + 640, + 906, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing shortcut features [23] rather than genuinely learning nuanced distinctions. While collecting training samples with subtle variations could enhance compositionality, this approach is time-consuming and labor-intensive, rendering it impractical at scale.", + "bbox": [ + 511, + 809, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advances in generative models [7, 12, 20, 43, 61, 75,", + "bbox": [ + 532, + 885, + 905, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 808, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24849", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "78, 91, 105, 110] now facilitate the synthetic generation of training samples with subtle variations between paired samples [11, 18, 49, 68, 71, 79, 83, 86, 106, 108, 109]. By starting with real image-caption pairs, generative models can create subtle edits to both captions and images, providing valuable training data with minimal manual effort. These generated variations enable VLMs to enhance their compositionality by learning from the nuanced differences.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, generating and training on synthetic data presents two key challenges. The first is the difficulty of efficiently and accurately creating precise variations in synthetic images. To generate large-scale training data with precise variations, the image generation process must meet three criteria: efficiency in producing large quantities of images, text alignment of the generated image to the corresponding caption, which is subtly changed from the original caption, and image fidelity in ensuring the generated image otherwise closely matches their real counterpart. Nevertheless, current image generation methods struggle to meet all three criteria simultaneously. Per-sample optimization methods [44, 82, 94, 112] lack efficiency, while zero-shot image editing methods [6, 22, 63] and text-to-image (T2I) models [75, 78] often fail to achieve proper text alignment and image fidelity, respectively, as illustrated in Figure 1 (a). The inaccurate synthetic variations produced by these models can mislead the learning of VLMs. To alleviate this issue, we take a fast T2I model, which excels in efficiency and text alignment, and inject real image features into the text prompt features in order to enhance image fidelity. By combining this approach with AdaIN [35], we manage to substantially improve the image generation process per the three criteria.", + "bbox": [ + 91, + 215, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Still, the generated data do not perfectly meet the text alignment criterion, and we propose to deal with the remaining problems with an innovative approach to model training. As Figure 1 (b) shows, the similarity between synthetic positive and negative pairs may vary, resulting in a mixture of hard text-image pairs, easy pairs, and incorrect pairs. Instead of treating all positive and negative samples uniformly, we propose a novel loss that differentiates between positive, hard negative, and easy negative samples. Further, we propose an adaptive margin that helps to filter out potentially incorrect generations and focus the learning on informative hard samples.", + "bbox": [ + 89, + 580, + 482, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose SPARCL, which integrates image feature injection into a fast T2I model to improve the quality of synthetic variations in images and employs an adaptive margin loss to leverage the varying alignment quality in synthetic samples for training VLMs. Evaluations on four compositional understanding benchmarks demonstrate that SPARCL significantly enhances VLM compositionality, improving CLIP by over $5\\%$ on VL-CheckList [113] and $7\\%$ on SugarCrepe [32], while surpassing state-of-the", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "art methods by $1\\%$ and $2\\%$ on the two benchmarks. The main contributions of this paper are as follows: (1) we propose image feature injection to enhance the quality of synthetic variations in images, which provide valuable training data to improve the compositionality of VLMs; (2) we introduce an adaptive margin loss to leverage varying levels of cross-modal alignment in synthetic samples to effectively differentiate positive and negative samples; (3) experimental results validate that SPARCL significantly improves the compositional understanding capabilities of CLIP models.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 258, + 653, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Limitations in Compositionality of VLMs", + "text_level": 1, + "bbox": [ + 511, + 282, + 867, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While VLMs excel in many multi-modal tasks [40, 52, 53, 74, 84, 101, 115], they still struggle with compositional understanding—the ability to interpret novel combinations of known visual and textual components. Benchmarks like What'sUp [42] reveal difficulties in understanding spatial relationships, SPEC [70] highlights issues with object size, position, and count, and ARO [106] uncovers limitations in understanding attributes, relations, and word order. Winoground [88], SNARE [96], and VL-CheckList [113] also expose these shortcomings. SugarCrepe [32] addresses hackable biases in prior benchmarks, where text-only models achieve artificially high performance, by introducing fluent and meaningful hard negatives. Their findings suggest that previous benchmarks overestimated compositional understanding. Building on this, SugarCrepe++ [19] further introduces semantically equivalent but lexically varied captions as hard positives and shows the difficulties of VLMs in distinguishing between lexical variations.", + "bbox": [ + 511, + 306, + 906, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Improving Compositionality of VLMs", + "text_level": 1, + "bbox": [ + 511, + 590, + 841, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior approaches to improving the compositionality of VLMs can be broadly classified into the following categories: (1) Leveraging detailed image captions: Detailed captions from dense captioning models [17, 54], simulation platforms [9], and video annotations [45] are collected to train VLMs. However, these samples often lack pairs with subtle variations, limiting their contribution to compositionality. (2) Distilling from pretrained models: SDS-CLIP [3], SF-CLIP [80] and IL-CLIP [114] distill knowledge from pretrained image generation models and visual-language foundation models. IL-CLIP [114] refines representations through iterative learning with pretrained vision and language agents. However, pretrained models also face limitations in compositionality [34, 90]. (3) Incorporating structural knowledge: MosaiCLIP [85] and Structure-CLIP [36] incorporate scene graph knowledge in text features. CLIP-SGVL [28] and 3VL [103] train VLMs to predict scene graphs. [66] utilizes scene graphs as prompts to elicit compositional knowledge from VLMs without fur", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24850", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "other training. However, these methods rely on models trained with expensive dense structure annotations (e.g., scene graphs). (4) Utilizing synthetic negative samples: Rule-based tools [30] or large language models (LLMs) [15, 73] are used to generate negative captions by editing real captions [10, 11, 18, 68, 83, 86, 106, 109]. Image generation models [75] are also used to create or edit images for training [49, 71, 79, 108]. Despite their utility, efficiently generating large amounts of images with precise variations remains challenging due to the inherent limitations of generative models. (5) Applying fine-grained alignment constraints: MCD [46] enforces multi-scale alignment across images with varying augmentations and the corresponding text captions. SPARC [5] learns local alignment by associating each text token with a group of local image patches. CE-CLIP [109] applies intra-modal contrastive loss and cross-modal ranking loss to improve alignment. However, uniform supervision applied to synthetic samples with varying alignment quality limits their effectiveness. In this paper, we propose SPARCL to address two key challenges in learning from synthetic data: the difficulty in accurately creating precise variations and the inconsistency in cross-modal alignment quality in synthetic data.", + "bbox": [ + 89, + 90, + 485, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Training with Synthetic Data", + "text_level": 1, + "bbox": [ + 89, + 446, + 351, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic data for training machine learning models have been studied in various fields [14, 26, 48, 64, 67, 76, 77, 93, 99]. Synthetic data generated through simulations and graphics engines supports a wide range of tasks [16, 72, 95]. However, these synthetic datasets often diverge significantly from real-world data. Recent advances in generative models [7, 12, 43, 61, 75, 78, 91, 105, 110] have made it possible to synthesize data that more closely resembles real-world scenarios. Synthetic data are widely used in both language tasks [21, 31, 60, 64, 65, 87, 97, 98, 100, 116] and vision tasks [1, 2, 4, 25, 29, 39, 57, 81, 89, 92, 111]. Synthetic data are often noisy, necessitating noise-resistant training methods [13, 37, 41, 50, 51, 102], especially those designed for contrastive learning [38, 56]. In this paper, we generate multimodal samples with subtle variations and filter out potentially incorrect ones during training to improve the compositionality of VLMs.", + "bbox": [ + 89, + 469, + 483, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SPARCL", + "text_level": 1, + "bbox": [ + 89, + 739, + 192, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve the compositional understanding abilities of VLMs, we propose Synthetic Perturbations for Advancing Robust Compositional Learning, or SPARCL, which generates multimodal samples with subtle variations from real samples and trains VLMs to learn nuanced differences through synthetic data. In the generation phase, SPARCL creates positive and negative captions with slight variations from real captions using an LLM, then generates images based on the real image and these modified captions through", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a fast T2I model. To enhance the quality of subtle variations in synthetic images, we introduce image feature injection, which integrates real image features into the text prompt features of the T2I model to improve fidelity to the real image, as detailed in Sec. 3.2. In the training phase, SPARCL employs an adaptive margin loss that leverages varying levels of multimodal alignment in the synthetic samples to effectively learn informative nuanced distinctions, as described in Sec. 3.3. The framework of SPARCL is shown in Figure 2.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generating Negative and Positive Captions", + "text_level": 1, + "bbox": [ + 511, + 251, + 879, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Captions with subtle variations are crucial for learning compositional knowledge, as shown by previous work that generates captions by randomly swapping or replacing nouns and adjectives [106, 109]. However, manually designed generation rules often introduce nonsensical or grammatically incorrect artifacts, creating shortcut features [23] that obstruct true compositional understanding [32]. To address this issue, we use an LLM to generate natural synthetic captions. To further mitigate the impact of generative artifacts, we generate both negative and positive captions, ensuring that VLMs cannot easily differentiate negative captions from positive ones based solely on artifacts. We denote the $i^{th}$ real image-caption pair in the training set as $(\\boldsymbol{I}_i^r,\\boldsymbol{T}_i^r)$ . Given a real caption $\\boldsymbol{T}_i^r$ , we prompt the LLM to generate a synthetic negative caption $\\boldsymbol{T}_i^{sn}$ and a synthetic positive caption $\\boldsymbol{T}_i^{sp}$ , as specified by the prompts in Figure A1 in the Appendix.", + "bbox": [ + 511, + 273, + 908, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Generating Images via Image Feature Injection", + "text_level": 1, + "bbox": [ + 511, + 539, + 906, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic images that exhibit subtle variations from real images while aligning with the captions are valuable but challenging to generate, as discussed in Sec. 1. Although previous works [49, 71, 79, 108] have utilized object segmentation or filtered dissimilar generations to improve fidelity to real images, they struggle with manipulating relationships or lack efficiency. We aim to enhance fidelity by injecting image features into a fast T2I model [61], which already achieves high efficiency and text alignment.", + "bbox": [ + 511, + 561, + 908, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Feature Injection. The images generated by T2I models lack image fidelity to real images, as no real image information is input to the models. To enhance the fidelity of synthetic images, we inject real image features into the text prompt features, enabling the model to incorporate information from the real images.", + "bbox": [ + 511, + 704, + 908, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In T2I models, content and style are separated in the semantic and padding embeddings. The semantic embeddings (before the embedding of the [EOS] token) usually capture most of the image content in the text prompts, while the padding embeddings (after the [EOS] token) usually represent the image style [104]. Therefore, we can inject real image features into the padding embeddings to guide", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24851", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/497a83f463f5bee0abcf4b37147e88b27e9e7c59b43f1c10da506ae2c143c20a.jpg", + "image_caption": [ + "(a) Generating synthetic captions and images" + ], + "image_footnote": [], + "bbox": [ + 98, + 113, + 566, + 286 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3847209be1ab50fffa2408ebd0dc56d68078672bc9d1c6f9837fba18b4a02182.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 94, + 900, + 285 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8ebc6cae65f82dcfdfdafda7c6c32aa8337d353637752417e23f6a58cfe5ca5.jpg", + "image_caption": [ + "(b) Training with real and synthetic samples" + ], + "image_footnote": [], + "bbox": [ + 94, + 311, + 388, + 448 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ec229f2694a8b91cd1a306779aa55a2ea6ef0916c20f5d7649715a48dfb61a1.jpg", + "image_caption": [ + "Adaptive Margin Loss", + "Figure 2. An overview of SPARCL. (a) Starting with a real image-caption pair, we generate synthetic positive and negative pairs with subtle variations using an LLM and a fast T2I model. To improve the quality of subtle variations in synthetic images, we introduce image feature injection to reduce unintended variations from a standard T2I model (see Sec. 3.2). (b) We train the VLM using both real and synthetic samples. In addition to a sigmoid loss for distinguishing positive and negative pairs, we apply an adaptive margin loss that leverages varying alignment levels across training samples to learn informative nuanced distinctions (see Sec. 3.3)." + ], + "image_footnote": [], + "bbox": [ + 397, + 311, + 743, + 450 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c88d2ee2497dc57fe24ef84887af0bf723486a9c5ae33532b4e969e775179112.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 297, + 898, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the model in generating images with a similar style to real images, without affecting the alignment with the captions, thanks to the decoupling of content and style in the prompt embeddings.", + "bbox": [ + 89, + 550, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We extract features for a real image $\\pmb{I}_i^r$ and the corresponding synthetic caption $\\pmb{T}_i^s$ ( $s \\in \\{sn, sp\\}$ ) using two aligned feature encoders (e.g., CLIP image and text encoders). The image embedding $\\pmb{f}_i^r$ is the encoder output at the position of the [CLS] token. The text embedding is the text encoder output $e_i^s = \\langle e_{i,j}^s \\rangle_{j=1}^L$ , where $L$ is the maximum sequence length. Let the index of the [EOS] token in $e_i^s$ be $k_i^s$ , $k_i^s \\leq L$ . That is, the padding token [PAD] is used as input to the text encoder at positions between $k_i^s$ and $L$ . The key step of SPARCL is to replace the text embeddings at those positions with the image embedding, producing a new sequence of text embeddings, $\\hat{e}_i^s$ ,", + "bbox": [ + 89, + 609, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {e} _ {i} ^ {s} = \\left\\langle e _ {i, 1} ^ {s}, e _ {i, 2} ^ {s}, \\dots , e _ {i, k _ {i}} ^ {s}, \\underbrace {\\boldsymbol {f} _ {i} ^ {r} , \\dots , \\boldsymbol {f} _ {i} ^ {r}} _ {L - k _ {i} ^ {s} \\text {t i m e s}} \\right\\rangle . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 782, + 482, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The embeddings $\\hat{e}_i^s$ is then used as input to the T2I model to generate a synthetic image $\\tilde{I}_i^s$ . This process, depicted in Figure 2, reduces unintended variations in synthetic images generated by the standard T2I model and enhances image fidelity to real images, thereby improving the quality of", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "synthetic subtle variations.", + "bbox": [ + 513, + 550, + 691, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style Transfer. To further reduce the domain gap between synthetic and real images, we use AdaIN [35] to transfer the style of the synthetic image to that of the real image. We use a pretrained AdaIN encoder to extract content features from $\\tilde{I}_i^s$ and style features from $I_i^r$ . We normalize the content features using instance normalization and then scale and shift them with the mean and variance of the style features. The transformed features are fed into a pretrained AdaIN decoder to generate $I_i^s$ , which is subsequently used for model training.", + "bbox": [ + 511, + 573, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training with Real and Synthetic Samples", + "text_level": 1, + "bbox": [ + 511, + 739, + 872, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After generating captions and images, each real image-caption pair $(I_i^r, T_i^r)$ is extended to $(I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})$ by adding one synthetic negative and one synthetic positive pair. We then train the VLM using these extended samples. Given a batch of $n$ sample groups, $\\{(I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})\\}_{i=1}^n$ , we organize the images by concatenating all real, synthetic negative, and synthetic positive images into an image batch $I^B$ of $3n$ images, and similar all captions into a caption", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24852", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "batch $\\pmb{T}^B$ of $3n$ captions.", + "bbox": [ + 89, + 90, + 261, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We calculate the similarity between each image-caption pair in the batch. The similarity between the $i^{\\mathrm{th}}$ image $\\pmb{I}_i^B$ and the $j^{\\mathrm{th}}$ caption $\\pmb{T}_j^B$ is given by $S_{i,j} = c(\\mathcal{E}_I(\\pmb{I}_i^B),\\mathcal{E}_T(\\pmb{T}_j^B))$ , where $c(\\cdot ,\\cdot)$ denotes cosine similarity, and $\\mathcal{E}_I$ and $\\mathcal{E}_T$ represent the image and text encoders of the VLM, respectively. We then define the ground-truth alignment variable $M_{i,j}$ , which takes value 1 if the image matches with the caption, and -1 otherwise. Clearly, the only positive image-caption pairs are $(\\pmb{I}_i^r,\\pmb{T}_i^r),(I_i^{sn},T_i^{sn}),(I_i^{sp},T_i^{sp}),(I_i^r,T_i^{sp})$ and $(I_i^{sp},T_i^r)$ . To account for multiple positive associations of an image or a caption, we apply a sigmoid-based contrastive loss [8, 107] to encourage higher similarity for positive pairs and lower similarity for negative pairs,", + "bbox": [ + 89, + 104, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o n} = - \\frac {1}{3 n} \\sum_ {i = 1} ^ {3 n} \\sum_ {j = 1} ^ {3 n} \\log \\frac {1}{1 + \\exp \\left(- M _ {i , j} \\left(S _ {i , j} / \\tau + b\\right)\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 311, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter and $b$ is a bias term.", + "bbox": [ + 89, + 357, + 472, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Margin Loss. Despite best efforts toward controlling generative models, synthetic samples may still have variable quality in terms of the alignment between text and imagery. That is, purported positive (resp. negative) pairs may not be semantically similar (resp. dissimilar). To account for variation in synthetic data quality, we propose to differentiate between real data and synthetic data and between easy negatives and hard negatives. Further, we propose an adaptive margin that filters out potential incorrect samples and prioritizes learning from hard samples.", + "bbox": [ + 89, + 378, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each image in a batch, we define four sets of captions, a positive set $\\mathbb{P}$ , a hard negative set $\\mathbb{N}_h$ , an easy negative set $\\mathbb{N}_e$ , and a real negative set $\\mathbb{N}_r$ , which represent different levels of alignment within the batch. If the image is a real or synthetic positive image,", + "bbox": [ + 89, + 530, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\mathbb{P}$ contains the real positive captions $T_{i}^{r}$ and the synthetic positive captions $T_{i}^{sp}$ of the current, $i$ -th image.", + "- $\\mathbb{N}_h$ contains the synthetic negative captions $T_i^{sn}$ of the current image.", + "- $\\mathbb{N}_e = \\{\\pmb{T}_j^r,\\pmb{T}_j^{sp},\\pmb{T}_j^{sn}|i\\neq j\\}$ contains all real and synthetic captions from all other images.", + "- $\\mathbb{N}_r = \\{\\pmb{T}_j^r | i \\neq j\\}$ contains all real captions belonging to the other images." + ], + "bbox": [ + 89, + 604, + 482, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If the image is a synthetic negative image, the sets $\\mathbb{N}_e$ and $\\mathbb{N}_r$ are unchanged but $\\mathbb{P}$ and $\\mathbb{N}_h$ differ:", + "bbox": [ + 89, + 726, + 482, + 755 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\mathbb{P}$ contains the synthetic negative captions $T_{i}^{sn}$ .", + "- $\\mathbb{N}_h$ contains $T_i^r$ and $T_i^{sp}$ ." + ], + "bbox": [ + 89, + 756, + 416, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define the margin loss for the $i^{th}$ image as:", + "bbox": [ + 89, + 786, + 400, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {m a r, i} ^ {I} = \\frac {1}{| \\mathbb {P} | \\cdot | \\mathbb {N} _ {h} |} \\sum_ {j _ {1} \\in \\mathbb {P}, j _ {2} \\in \\mathbb {N} _ {h}} \\max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}) \\\\ + \\frac {1}{\\left| \\mathbb {N} _ {h} \\right| \\cdot \\left| \\mathbb {N} _ {e} \\right|} \\sum_ {j _ {1} \\in \\mathbb {N} _ {h}, j _ {2} \\in \\mathbb {N} _ {e}} \\max \\left(0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}\\right) \\tag {3} \\\\ + \\frac {\\alpha}{| \\mathbb {P} | \\cdot | \\mathbb {N} _ {r} |} \\sum_ {j _ {1} \\in \\mathbb {P}, j _ {2} \\in \\mathbb {N} _ {r}} \\max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 805, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This loss encourages positive pairs to have higher similarity scores than hard negative pairs, which should have higher similarity than easy negative pairs. Additionally, a weight $\\alpha > 1$ is applied to comparisons between positive pairs and real negative pairs to emphasize these comparisons, as real samples are generally correct, making comparisons involving them more reliable.", + "bbox": [ + 511, + 90, + 903, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we propose an adaptive margin in the above loss. Let $d = S_{i,j_1} - S_{i,j_2}$ denote the difference between two similarity scores in Eq. (3). The adaptive margin $m$ is then computed as follows:", + "bbox": [ + 511, + 196, + 903, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm = \\left\\{ \\begin{array}{l l} d, & d < \\beta \\\\ (\\frac {m _ {0} - d}{m _ {0} - \\beta} \\gamma + 1) m _ {0}, & \\beta \\leq d \\leq m _ {0} \\\\ m _ {0} & d > m _ {0} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 262, + 903, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m_0$ is a base margin, $\\beta < 0$ is a cutoff threshold, $\\gamma$ is a scaling factor. The adaptive margin is designed with the following rationale: when $d < \\beta$ , the purported positive pair is much less similar than the purported negative pair, suggesting the presence of incorrect or mislabeled samples, so we zero out the loss by setting the margin to $d$ . For $\\beta \\leq d \\leq m_0$ , the margin is scaled, with smaller differences receiving larger margins to emphasize learning from harder samples. When $d > m_0$ , the samples are already well-separated with the margin $m_0$ , so the margin is capped at $m_0$ , again leading to zero loss. We visualize both the adaptive margin and the corresponding loss as functions of $d$ in Figure 2 (b).", + "bbox": [ + 511, + 316, + 905, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss for all images is given by $L_{mar}^{I} = \\frac{1}{3n}\\sum_{i}L_{mar,i}^{I}$ . Analogously, we compute $L_{mar}^{T}$ for all captions. The total adaptive margin loss is $L_{mar} = L_{mar}^{I} + L_{mar}^{T}$ . Finally, the overall training loss is a weighted combination of the contrastive loss and the adaptive margin loss, with weight $\\lambda$ :", + "bbox": [ + 511, + 512, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {c o n} + \\lambda L _ {m a r}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 609, + 903, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 650, + 643, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 676, + 614, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We use the COCO-2014 dataset [55] as the training data source. The training set consists of 82,783 images, each paired with five captions. For each real image-caption pair, we generate one positive and one negative synthetic pair. In line with previous approaches [18, 106, 109], we train the VLMs using both the original COCO-2014 training data and the synthetic samples.", + "bbox": [ + 511, + 696, + 905, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. We evaluate our model on four vision-language compositional understanding benchmarks: (1) ARO [106], which consists of 23,937 cases for relation understanding and 28,748 for attribute understanding. We exclude the subsets for order understanding, as they contain significant nonsensical and non-fluent artifacts [32].", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24853", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3524514efdf5cc4c7fb32200eda28faef94a453bff77696fc34250f0e8499f6a.jpg", + "table_caption": [ + "Table 1. Comparison of accuracy (%) between SPARCL and baselines on four benchmarks. \"img.\" represents images, \"cap.\" represents captions, \"syn.\" represents synthetic data." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataAROVL-CheckListSugarCrepeSugarCrepe++
Source# real img.# real cap.# syn. img.# syn. cap.
CLIP [74] (Zero-Shot)-----61.173.273.459.8
CLIP [74] (Finetune)COCO82K410K0064.172.879.962.3
SDS-CLIP [3]COCO82K410K0057.5---
[79]COCO0082K82K65.069.9--
AMR-NegCLIP [83]COCO100K100K0500K79.4-85.2-
NegCLIP [106]COCO100K100K0500K76.074.682.564.9
MosaiCLIP [85]COCO109K109K0981K80.376.8--
FSC-CLIP [68]COCO100K100K01.5M-77.285.1-
CE-CLIP [109]COCO82K410K02M79.776.385.2-
COMO [49]COCO113K567K567K567K-76.9--
SPARCL (our method)COCO82K410K820K820K77.279.287.166.1
SPEC [70]LAION20K20K20K20K70.1---
CounterCurate [108]Flickr30K30K150K150K--82.8-
[18]CC3M3M3M09M-75.3-55.3
CE-CLIP+ [109]COCO+CC3M3M3M015M80.479.387.5-
CLOVE [11]LAION-COCO>1B>1B0>1B73.2-85.1-
IL-CLIP [114]CC12M12M12M00--70.3-
SF-CLIP [80]YFCC15M15M15M00--71.2-
syn-CLIP [9]SyViC00>1M>1M69.274.8--
FiGCLIP [45]VidSitu20K videos0067.0-74.6-
", + "bbox": [ + 99, + 127, + 895, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(2) VL-CheckList [113], a large-scale benchmark with over 100,000 samples, evaluates compositionality across subsets of objects, attributes, and relationships, which are further divided into various fine-grained categories. (3) SugarCrepe [32] includes 7,000 test cases across seven subsets. In the above three benchmarks, each test case containing one image, one positive caption, and one negative caption. (4) SugarCrepe++ [19] includes 4,757 test samples across five subsets, where each test case consists of one image, two positive captions, and one negative caption. All benchmarks involve classifying captions as positive or negative for the given images. We report the average accuracy across all subsets of each benchmark and include the accuracy for each subset in Appendix A3.", + "bbox": [ + 88, + 469, + 486, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 690, + 308, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We use ViT-B/32 and ViT-L/14 architectures from OpenAI's CLIP model [74] as our base models, initialized with pretrained checkpoints. Following syn-CLIP [9], we integrate LoRA adapters [33] into both the image and text encoders of CLIP to improve training efficiency and mitigate knowledge forgetting. Only the LoRA adapters are fine-tuned during training.", + "bbox": [ + 89, + 712, + 485, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Setups. We use the AdamW optimizer [59] with a cosine learning rate schedule [58]. Training is conducted on two Tesla V100 GPUs, with a batch size of 128 sample groups for ViT-B/32 and 16 for ViT-L/14. The base learning rate is set to 0.01 for a total batch size of 256, and scaled", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "linearly [24] based on the actual batch size. Training is performed for 3,000 steps for ViT-B/32 and 15,000 steps for ViT-L/14, corresponding to fewer than 5 epochs in previous studies [18, 106, 109]. More details and hyperparameters settings are in Appendix A2.", + "bbox": [ + 511, + 469, + 906, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 559, + 653, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SPARCL with several baseline methods: (1) methods that utilize synthetic samples for training [49, 68, 70, 79, 83, 106, 108, 109], (2) methods that distill knowledge from pretrained models [3, 11, 80, 114], (3) methods that incorporate knowledge from scene graphs [18, 85], and (4) methods that leverage detailed image captions [9, 45]. Both SPARCL and the baseline methods use ViT-B/32 as the base model. The results are presented in Table 1.", + "bbox": [ + 511, + 580, + 906, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that SPARCL achieves the best performance on VL-CheckList, SugarCrepe, and SugarCrepe++ compared to baselines trained on the same data source, namely COCO. Specifically, SPARCL surpasses the strongest baseline by $1.4\\%$ on VL-CheckList and $2.5\\%$ on SugarCrepe, underscoring its effectiveness in enhancing compositional understanding. However, SPARCL performs worse on ARO than some baselines. We hypothesize that this may be due to nonsensical or grammatically incorrect artifacts in the ARO test samples [32], which could favor methods that utilize rule-based synthetic training samples (e.g., [85]) containing similar artifacts. Notably, even with only $50\\%$ of the training data, SPARCL outperforms most base", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24854", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d61ed83d7a9b3ace1242042cee4693c9d8eafcf67b645cccad677fb4376ae6f.jpg", + "table_caption": [ + "Table 2. Ablated performance (\\%) of SPARCL. \"SynCap\" refers to synthetic captions, \"SynImg\" refers to synthetic images, \"FeatInj\" denotes image feature injection, and \"CompSet\" indicates the comparison sets used in Eq. (3)." + ], + "table_footnote": [], + "table_body": "
ModelVariantSynCapSynImgFeatInjAdaINAdaptive Margin LossAROVL-CheckListSugarCrepeSugarCrepe++Average
CompSetMargin
ViT-B/32#1XXXXX60.4972.6179.3664.8569.33
#2✓XXXX71.7773.5286.3564.3273.99
#3X✓XXX62.6271.7079.9764.8469.79
#4✓✓XXX71.8675.5485.4365.2274.51
#5✓✓✓XX73.4076.5685.5465.7875.32
#6✓✓X✓X73.7975.7285.7964.4974.95
#7✓✓✓✓X74.1276.3585.4066.4475.58
#8✓✓✓✓AllFixed76.7978.5987.0865.4276.97
#9✓✓✓✓AllAdaptive77.1579.1687.1166.1277.38
#10✓✓✓✓AllAdaptive Inversed76.6779.2686.6865.3076.97
#11✓✓✓✓Only (P, Nh)Adaptive77.4880.4886.1564.7077.20
ViT-L/14#1XXXXX59.8073.2681.4964.9069.86
#2✓XXXX72.1675.4487.3864.4474.85
#3X✓XXX60.7072.2382.5866.6370.54
#4✓✓XXX75.2078.2987.2964.6176.35
#5✓✓✓XX75.1178.5887.5464.5776.45
#6✓✓X✓X74.8879.1087.5064.6276.52
#7✓✓✓✓X74.9380.0487.7965.3077.01
#8✓✓✓✓AllFixed75.2180.7588.0266.4177.60
#9✓✓✓✓AllAdaptive75.8380.8188.2366.8377.93
#10✓✓✓✓AllAdaptive Inversed75.1580.7187.9366.6177.60
#11✓✓✓✓Only (P, Nh)Adaptive76.2680.5787.4266.2877.63
", + "bbox": [ + 106, + 127, + 893, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "line methods (see Table 3 for SPARCL performance with reduced training data). Compared to baselines using additional data sources, SPARCL surpasses or matches their performance. For example, SPARCL outperforms methods in [11, 18] despite their use of more training data. SPARCL performs comparably to CE-CLIP+ [109] on VL-CheckList and SugarCrepe, despite CE-CLIP+ leveraging CC3M as an additional data source and utilizing significantly more samples.", + "bbox": [ + 88, + 460, + 480, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 89, + 611, + 346, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation studies to assess the impact of each component and design choice in SPARCL. The results, presented in Table 2, show the overall performance.", + "bbox": [ + 89, + 635, + 482, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic captions vs. synthetic images. To analyze the impact of synthetic captions and images, we compare Variant #1 (using only real samples) with Variant #2 (using real samples and synthetic captions) and #3 (using real samples and synthetic images). Variant #2 shows a significant improvement over Variant #1, with the average accuracy improving from $69.33\\%$ to $73.99\\%$ for ViT-B/32. However, Variant #3 obtains only a marginal improvements of about $0.5\\%$ . These results suggest that synthetic captions substantially enhance compositional understanding, while synthetic images alone provide limited benefit, which aligns with findings in [69]. We hypothesize that generative artifacts in synthetic images are more pronounced than in captions, which negatively affects the effective learning of nu", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "anced distinctions by VLMs. When both synthetic captions and images are used (Variant #4), performance improves further, indicating a synergistic effect.", + "bbox": [ + 511, + 460, + 906, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image feature injection vs. AdaIN. To analyze the impact of image feature injection and AdaIN, we compare Variant #4 (without image feature injection or AdaIN) with Variant #5 (with image feature injection only) and #6 (with AdaIN only). Using ViT-B/32, we observe that Variant #5 outperforms Variant #4 across all four benchmarks, with an average gap of $0.8\\%$ . While Variant #6 shows a smaller improvement of $0.4\\%$ , it outperforms Variant #5 on ARO and SugarCrepe. These results suggest that image feature injection is more effective for improving compositionality than AdaIN, although the two methods are somewhat complementary. Through the improvements obtained using ViT-L/14 are a little bit different, both techniques improve performance over Variant #4. Combining both methods in Variant #7 leads to further improvements, with a more than $1\\%$ increase in average accuracy for ViT-B/32 and a $0.65\\%$ increase for ViT-L/14 over Variant #4. This highlights the importance of reducing unintended changes in synthetic images, as such changes could cause the model to rely on them for distinguishing positive and negative samples rather than the semantic differences specified in the corresponding captions. Examples of synthetic images in Figure A3 and A4 in the Appendix show how image feature injection helps mitigate these unintended changes.", + "bbox": [ + 511, + 513, + 908, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of the margin loss and adaptive margin strate-", + "bbox": [ + 513, + 885, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24855", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a94cd500998d001a314457782a3a5ed2706114dd3ca8c958abe1f9fcd01b87f1.jpg", + "table_caption": [ + "Table 3. Performance (\\%) of SPARCL using different subsets of training samples. \"Neg.\" represents synthetic negative samples, \"Pos.\" represents synthetic positive samples, and \"Prop.\" represents the proportion of training samples." + ], + "table_footnote": [], + "table_body": "
VariantNeg.Pos.Prop.AROVL-CheckListSugarCrepeSugarCrepe++Average
#7✓×100%72.1274.6586.4157.9972.79
#7✓✓100%74.1276.3585.4066.4475.58
#9✓✓20%75.9077.5685.7864.4475.92
#9✓✓50%77.9079.3886.7164.2177.05
#9✓✓100%77.1579.1687.1166.1277.38
", + "bbox": [ + 91, + 157, + 488, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gies. We compare three variants with distinct margin approaches: (1) Fixed (Variant #8): a fixed margin without adaptive adjustments; (2) Adaptive (Variant #9): the margin is calculated according to Eq. (4); and (3) Adaptive Inversed (Variant #10): a larger margin is applied to samples with higher similarity differences, representing an inverse version of Eq. (4) that prioritizes learning from easier samples. Comparing Variant #8 with Variant #7, we observe that #8 improves by over $1\\%$ for ViT-B/32 and $0.6\\%$ for ViT-L/14 in average accuracy, suggesting that margin loss effectively aids in compositional understanding. Between Variants #8, #9, and #10, we find that Variant #9 outperforms Variant #8 across all four benchmarks, with an average boost of over $0.3\\%$ for both ViT-B/32 and ViT-L/14. In contrast, Variant #10 results in a performance decline on three benchmarks and no improvement in average accuracy. These findings highlight the superior effectiveness of adaptive margins.", + "bbox": [ + 88, + 267, + 485, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is applying adaptive margin loss only to hard samples sufficient? To explore whether applying margin loss solely to hard samples is adequate, we construct Variant #11, which uses only the positive set $\\mathbb{P}$ and the hard negative set $\\mathbb{N}_h$ for margin loss calculation in Eq. (3), which is similar to learning strategies in [49, 109]. When comparing Variant #11 with Variant #8 and #9, we observe that Variant #11 performs better on ARO and VL-CheckList, but worse on SugarCrepe and SugarCrepe++. We hypothesize that by focusing exclusively on the positive and hard negative sets, the model may learn nonsensical artifacts present in the hard negatives. These artifacts improve performance on ARO and VL-CheckList, where the test samples exhibit similar patterns, but hinder performance on SugarCrepe and SugarCrepe++, which are designed to avoid such patterns [32]. In contrast, Variant #9, which incorporates all sets in Eq. (3) for margin loss calculation, achieves better results on SugarCrepe and SugarCrepe++ with only a minor drop in performance on ARO and VL-CheckList. This suggests that easy negative samples serve as regularization, preventing overfitting to the artifacts in hard negative samples.", + "bbox": [ + 91, + 531, + 485, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of synthetic positive samples. We compare the performance of training with only synthetic negative samples versus using both synthetic negative and positive samples", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Table 3. Based on Variant #7, we train ViT-B/32 with these two combinations of training data. We observe nearly $3\\%$ decrease in average accuracy when using only synthetic negative samples, compared to using both synthetic negative and positive samples. This decrease is particularly noticeable on SugarCrepe++, ARO, and VL-CheckList. These results underscore the importance of incorporating synthetic positive samples in training, as they provide essential information about variations that maintain semantic consistency, which is critical for compositional understanding.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of model size and training data size. To evaluate the impact of model size on compositional understanding, we compare the performance of ViT-B/32 and ViT-L/14 in Table 2. We find that ViT-L/14 only provides a modest improvement of about $0.5\\%$ of average accuracy over ViT-B/32. This suggests that increasing model size does not significantly enhance compositional understanding. To investigate the effect of training data size, we train ViT-B/32 using $20\\%$ and $50\\%$ of randomly sampled training data and report the results in Table 3. We observe that performance improves as the proportion of training data increases, indicating that a larger training set helps the model better capture compositional knowledge. However, the performance gain diminishes when the training data size increases from $50\\%$ to $100\\%$ . Perhaps generating or selecting high-quality data would be more effective than simplistic data scaling.", + "bbox": [ + 511, + 251, + 908, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional experiments on hyperparameter analysis can be found in Appendix A3.", + "bbox": [ + 511, + 496, + 908, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 549, + 633, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SPARCL to enhance the compositional understanding capabilities of VLMs by generating and training with synthetic data. To tackle two key challenges in using synthetic data—namely, the difficulty of generating accurate variations and the inconsistency in cross-modal alignment quality—SPARCL integrates image feature injection into a T2I model to improve the quality of synthetic variations and introduces an adaptive margin loss to account for varying levels of cross-modal alignment for effectively learning nuanced distinctions. Experiments on four visual-language compositional understanding benchmarks demonstrate the effectiveness of SPARCL.", + "bbox": [ + 511, + 577, + 908, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 781, + 692, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF-NRFF13-2021-0006), Singapore. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24856", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 3", + "[2] Manel Baradad Jurjo, Jonas Wulff, Tongzhou Wang, Phillip Isola, and Antonio Torralba. Learning to see by looking at noise. Advances in Neural Information Processing Systems, 34:2556-2569, 2021. 3", + "[3] Samyadeep Basu, Maziar Sanjabi, Daniela Massiceti, Shell Xu Hu, and Soheil Feizi. Augmenting clip with improved visio-linguistic reasoning. arXiv preprint arXiv:2307.09233, 2023. 2, 6", + "[4] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This dataset does not exist: training models from generated images. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2020. 3", + "[5] Ioana Bica, Anastasija Ilic, Matthias Bauer, Goker Erdoğan, Matko Bošnjak, Christos Kaplanis, Alexey A Gritsenko, Matthias Minderer, Charles Blundell, Razvan Pascanu, et al. Improving fine-grained understanding in image-text pre-training. arXiv preprint arXiv:2401.09865, 2024. 3", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3", + "[8] Adrian Bulat, Yassine Ouali, and Georgios Tzimiropoulos. Fff: Fixing flawed foundations in contrastive pre-training results in very strong vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14172-14182, 2024. 5", + "[9] Paola Cascante-Bonilla, Khaled Shehada, James Seale Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gul Varol, Aude Oliva, Vicente Ordonez, Rogerio Feris, et al. Going beyond nouns with vision & language models using synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20155-20165, 2023. 2, 6", + "[10] Paola Cascante-Bonilla, Yu Hou, Yang Trista Cao, Hal Daumé III, and Rachel Rudinger. Natural language inference improves compositionality in vision-language models. arXiv preprint arXiv:2410.22315, 2024. 3", + "[11] Santiago Castro, Amir Ziai, Avneesh Saluja, Zhuoning Yuan, and Rada Mihalcea. Clove: Encoding compositional language in contrastive vision-language models. arXiv preprint arXiv:2402.15021, 2024. 2, 3, 6, 7", + "[12] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy," + ], + "bbox": [ + 101, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3", + "[13] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 3", + "[14] Yabo Dan, Yong Zhao, Xiang Li, Shaobo Li, Ming Hu, and Jianjun Hu. Generative adversarial networks (gan) based efficient sampling of chemical composition space for inverse design of inorganic materials. npj Computational Materials, 6(1):84, 2020. 3", + "[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[16] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3", + "[17] Sivan Doveh, Assaf Arbelle, Sivan Harary, Amit Alfassy, Roei Herzig, Donghyun Kim, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, et al. Dense and aligned captions (dac) promote compositional reasoning in v1 models. arXiv preprint arXiv:2305.19595, 2023. 2", + "[18] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2, 3, 5, 6, 7", + "[19] Sri Harsha Dumpala, Aman Jaiswal, Chandramouli Sastry, Evangelos Milios, Sageev Oore, and Hassan Sajjad. Sugarcrepe++ dataset: Vision-language model sensitivity to semantic and lexical alterations. arXiv preprint arXiv:2406.11171, 2024. 2, 6", + "[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1", + "[21] Jiahui Gao, Renjie Pi, Lin Yong, Hang Xu, Jiacheng Ye, Zhiyong Wu, Weizhong Zhang, Xiaodan Liang, Zhenguo Li, and Lingpeng Kong. Self-guided noise-free data generation for efficient zero-shot learning. In International Conference on Learning Representations (ICLR 2023), 2023. 3", + "[22] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. arXiv preprint arXiv:2403.14602, 2024. 2, 1", + "[23] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural net" + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24857", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works.Nature Machine Intelligence,2(11):665-673,2020. 1,3", + "[24] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6", + "[25] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 3", + "[26] Xuanli He, Islam Nassar, Jamie Kiros, Gholamreza Haffari, and Mohammad Norouzi. Generate, annotate, and learn: Nlp with synthetic text. Transactions of the Association for Computational Linguistics, 10:826-842, 2022. 3", + "[27] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 3", + "[28] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arbelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2", + "[29] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3", + "[30] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1):411-420, 2017. 3", + "[31] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language models with (almost) no human labor. arXiv preprint arXiv:2212.09689, 2022. 3", + "[32] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 1, 2, 3, 5, 6, 8", + "[33] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[34] Hang Hua, Yunlong Tang, Ziyun Zeng, Liangliang Cao, Zhengyuan Yang, Hangfeng He, Chenliang Xu, and Jiebo Luo. Mmcomposition: Revisiting the compositionality of pre-trained vision-language models. arXiv preprint arXiv:2410.09733, 2024. 2", + "[35] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4", + "[36] Yufeng Huang, Jiji Tang, Zhuo Chen, Rongsheng Zhang, Xinfeng Zhang, Weijie Chen, Zeng Zhao, Zhou Zhao," + ], + "bbox": [ + 99, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tangjie Lv, Zhipeng Hu, et al. Structure-clip: Towards scene graph knowledge to enhance multi-modal structured representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2417-2425, 2024. 2", + "[37] Zhizhong Huang, Junping Zhang, and Hongming Shan. Twin contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2023. 3", + "[38] Sarah Ibrahimi, Arnaud Sors, Rafael Sampaio de Rezende, and Stéphane Clinchant. Learning with label noise for image retrieval by selecting interactions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2181-2190, 2022. 3", + "[39] Ali Jahanian, Xavier Puig, Yonglong Tian, and Phillip Isola. Generative models as a data source for multiview representation learning. arXiv preprint arXiv:2106.05258, 2021. 3", + "[40] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2", + "[41] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 3", + "[42] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's\" up\" with vision-language models? investigating their struggle with spatial reasoning. arXiv preprint arXiv:2310.19785, 2023. 1, 2", + "[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 3", + "[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2", + "[45] Zeeshan Khan, Makarand Tapaswi, et al. Figclip: Fine-grained clip adaptation via densely annotated videos. arXiv preprint arXiv:2401.07669, 2024. 2, 6", + "[46] Bumsoo Kim, Yeonsik Jo, Jinhyung Kim, and Seunghwan Kim. Misalign, contrast then distill: Rethinking misalignments in language-image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2023. 3", + "[47] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1", + "[48] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 3", + "[49] Chengen Lai, Shengli Song, Sitong Yan, and Guangneng Hu. Improving vision and language concepts understanding with multimodal counterfactual samples. In European" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24858", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 174-191. Springer, 2024. 2, 3, 6, 8", + "[50] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5051-5059, 2019. 3", + "[51] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9485-9494, 2021. 3", + "[52] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2", + "[53] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2", + "[54] Liunian Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[56] Chang Liu, Han Yu, Boyang Li, Zhiqi Shen, Zhanning Gao, Peiran Ren, Xuansong Xie, Lizhen Cui, and Chunyan Miao. Noise-resistant deep metric learning with ranking-based instance selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6811-6820, 2021. 3", + "[57] Hao Liu, Tom Zahavy, Volodymyr Mnih, and Satinder Singh. Palm up: Playing in the latent manifold for unsupervised pretraining. Advances in Neural Information Processing Systems, 35:35880-35893, 2022. 3", + "[58] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[60] Yujie Lu, Wanrong Zhu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Imagination-augmented natural language understanding. arXiv preprint arXiv:2204.08535, 2022. 3", + "[61] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3", + "[62] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 1" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 1", + "[64] Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. Generating training data with language models: Towards zero-shot language understanding. Advances in Neural Information Processing Systems, 35:462-477, 2022. 3", + "[65] Yu Meng, Martin Michalski, Jiaxin Huang, Yu Zhang, Tarek Abdelzaher, and Jiawei Han. Tuning language models as training data generators for augmentation-enhanced few-shot learning. In International Conference on Machine Learning, pages 24457-24477. PMLR, 2023. 3", + "[66] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076, 2023. 2", + "[67] Sergey I Nikolenko. Synthetic data for deep learning. Springer, 2021. 3", + "[68] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pre-trained vlms for improving vision-linguistic compositionality. arXiv preprint arXiv:2410.05210, 2024. 2, 3, 6", + "[69] Maitreya Patel, Abhiram Kusumba, Sheng Cheng, Changhoon Kim, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Tripletclip: Improving compositional reasoning of clip via synthetic vision-language negatives. arXiv preprint arXiv:2411.02545, 2024. 7", + "[70] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. arXiv preprint arXiv:2312.00081, 2023. 2, 6", + "[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize diagnose and optimize: Towards fine-grained vision-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13279-13288, 2024. 2, 3", + "[72] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.3", + "[73] Alec Radford. Improving language understanding by generative pre-training. 2018. 3", + "[74] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 1, 3", + "[75] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3", + "[76] Andrew Rosenberg, Yu Zhang, Bhuvana Ramabhadran, Ye Jia, Pedro Moreno, Yonghui Wu, and Zelin Wu. Speech recognition with augmented synthesized speech. In 2019" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24859", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE automatic speech recognition and understanding workshop (ASRU), pages 996-1002. IEEE, 2019. 3", + "[77] Nick Rossenbach, Albert Zeyer, Ralf Schlüter, and Hermann Ney. Generating synthetic audio data for attention-based speech recognition systems. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7069-7073. IEEE, 2020. 3", + "[78] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3", + "[79] Ugur Sahin, Hang Li, Qadeer Khan, Daniel Cremers, and Volker Tresp. Enhancing multimodal compositional reasoning of visual language models with generative negative mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5563-5573, 2024. 2, 3, 6", + "[80] Sepehr Sameni, Kushal Kafle, Hao Tan, and Simon Jenni. Building vision-language models on solid foundations with masked distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14216-14226, 2024. 2, 6", + "[81] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[82] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2", + "[83] Ziyi Shou and Fangzhen Lin. Enhancing semantic understanding in vision language models using meaning representation negative generation. In Fourth Workshop on Knowledge-infused Learning, 2024. 2, 3, 6", + "[84] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2", + "[85] Harman Singh, Pengchuan Zhang, Qifan Wang, Mengjiao Wang, Wenhan Xiong, Jingfei Du, and Yu Chen. Coarse-to-fine contrastive learning in image-text-graph space for improved vision-language compositionality. arXiv preprint arXiv:2305.13812, 2023. 2, 6", + "[86] Jaisidh Singh, Ishaan Shrivastava, Mayank Vatsa, Richa Singh, and Aparna Bharati. Learn\" no\" to say\" yes\" better: Improving vision-language models via negations. arXiv preprint arXiv:2403.20312, 2024. 2, 3", + "[87] Tianyi Tang, Yushuo Chen, Yifan Du, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen. Learning to imagine: Visually-" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "augmented natural language generation. arXiv preprint arXiv:2305.16944, 2023. 3", + "[88] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2", + "[89] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[90] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2", + "[91] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3", + "[92] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 3", + "[93] Allan Tucker, Zhenchen Wang, Ylenia Rotalinti, and Puja Myles. Generating high-fidelity synthetic patient data for assessing machine learning healthcare software. NPJ digital medicine, 3(1):1-13, 2020. 3", + "[94] Dani Valevski, Matan Kalman, Eyal Molad, Eyal Segalis, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning a diffusion model on a single image. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2", + "[95] Gül Varol, Ivan Laptev, Cordelia Schmid, and Andrew Zisserman. Synthetic humans for action recognition from unseen viewpoints. International Journal of Computer Vision, 129(7):2264-2287, 2021. 3", + "[96] Fei Wang, Liang Ding, Jun Rao, Ye Liu, Li Shen, and Changxing Ding. Can linguistic knowledge improve multimodal alignment in vision-language pretraining? arXiv preprint arXiv:2308.12898, 2023. 2", + "[97] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022.3", + "[98] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 3" + ], + "bbox": [ + 522, + 93, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24860", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[99] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. arXiv preprint arXiv:2004.11546, 2020. 3", + "[100] Yue Yang, Wenlin Yao, Hongming Zhang, Xiaoyang Wang, Dong Yu, and Jianshu Chen. Z-lavi: Zero-shot language solver fueled by visual imagination. arXiv preprint arXiv:2210.12261, 2022. 3", + "[101] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2", + "[102] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 3", + "[103] Nir Yellinek, Leonid Karlinsky, and Raja Giryes. 3vl: using trees to teach vision & language models compositional concepts. arXiv preprint arXiv:2312.17345, 2023. 2", + "[104] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3", + "[105] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3", + "[106] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 5, 6", + "[107] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 5", + "[108] Jianrui Zhang, Mu Cai, Tengyang Xie, and Yong Jae Lee. Countercurate: Enhancing physical and semantic visio-linguistic compositional reasoning via counterfactual examples. arXiv preprint arXiv:2402.13254, 2024. 2, 3, 6", + "[109] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic fine-grained understanding. arXiv preprint arXiv:2306.08832, 2023. 2, 3, 5, 6, 7, 8", + "[110] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 3", + "[111] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[112] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2", + "[113] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6", + "[114] Chenhao Zheng, Jieyu Zhang, Aniruddha Kembhavi, and Ranjay Krishna. Iterated learning improves compositionality in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13785-13795, 2024. 2, 6", + "[115] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2", + "[116] Wanrong Zhu, An Yan, Yujie Lu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Visualize before you write: Imagination-guided open-ended text generation. arXiv preprint arXiv:2210.03765, 2022. 3" + ], + "bbox": [ + 516, + 90, + 906, + 415 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "24861", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_model.json b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..053a68fa16d37480dc169d7ba35708aab24d4a7f --- /dev/null +++ b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_model.json @@ -0,0 +1,3119 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.204, + 0.603, + 0.222 + ], + "angle": 0, + "content": "Haoxin Li and Boyang Li" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.223, + 0.637, + 0.241 + ], + "angle": 0, + "content": "Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.242, + 0.642, + 0.257 + ], + "angle": 0, + "content": "{haoxin003,boyang.li}@ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Paired image-text data with subtle variations in-between (e.g., people holding surfboards vs. people holding shovels) hold the promise of producing Vision-Language Models with proper compositional understanding. Synthesizing such training data from generative models is a highly coveted prize due to the reduced cost of data collection. However, synthesizing training images for compositional learning presents three challenges: (1) efficiency in generating large quantities of images, (2) text alignment between the generated image and the caption in the exact place of the subtle change, and (3) image fidelity in ensuring sufficient similarity with the original real images in all other places. We propose SPARCL (Synthetic Perturbations for Advancing Robust Compositional Learning), which integrates image feature injection into a fast text-to-image generative model, followed by an image style transfer step, to meet the three challenges. Further, to cope with any residual issues of text alignment, we propose an adaptive margin loss to filter out potentially incorrect synthetic samples and focus the learning on informative hard samples. Evaluation on four compositional understanding benchmarks demonstrates that SPARCL significantly improves the compositionality of CLIP, boosting the average accuracy of the CLIP base model by over \\(8\\%\\) across all benchmarks and outperforming state-of-the-art methods by \\(2\\%\\) on three benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.753, + 0.222, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Current Vision-Language Models (VLMs) still face limitation in accurately interpreting compositional relationships between objects and attributes, as demonstrated by numerous evaluations [32, 62, 88, 106, 113]. This limitation primarily stems from the absence of subtle variations in the training data [42] (e.g., the subtle variations between the two captions in Figure 1 (a)). As a result, it becomes possible to maximize empirical image-caption alignment us" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.29, + 0.842, + 0.302 + ], + "angle": 0, + "content": "Source caption: Two people carrying surfboards on a beach." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.302, + 0.821, + 0.313 + ], + "angle": 0, + "content": "Target caption: Two people carrying shovels on a beach." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.314, + 0.636, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.401, + 0.61, + 0.41 + ], + "angle": 0, + "content": "Source image" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.313, + 0.768, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.401, + 0.741, + 0.41 + ], + "angle": 0, + "content": "Edited image" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.313, + 0.901, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.401, + 0.882, + 0.41 + ], + "angle": 0, + "content": "Generated image" + }, + { + "type": "text", + "bbox": [ + 0.568, + 0.412, + 0.853, + 0.425 + ], + "angle": 0, + "content": "(a) Difficulty in accurately creating precise variations." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.434, + 0.635, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.525, + 0.632, + 0.552 + ], + "angle": 0, + "content": "+Real positive caption: A man is stirring a silver pot filled with food." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.552, + 0.632, + 0.588 + ], + "angle": 0, + "content": "- Synthetic negative caption: A man is stirring a copper pot filled with food." + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.433, + 0.769, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.525, + 0.764, + 0.552 + ], + "angle": 0, + "content": "+Real positive caption: Random people sitting in a public transport bus." + }, + { + "type": "text", + "bbox": [ + 0.654, + 0.552, + 0.75, + 0.588 + ], + "angle": 0, + "content": "- Synthetic negative caption: Random animals riding a purple elephant." + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.598, + 0.678, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.6, + 0.751, + 0.611 + ], + "angle": 0, + "content": "Easy Sample" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.433, + 0.902, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.785, + 0.525, + 0.897, + 0.552 + ], + "angle": 0, + "content": "+Real positive caption: A double decker bus parked at a station." + }, + { + "type": "text", + "bbox": [ + 0.785, + 0.552, + 0.9, + 0.598 + ], + "angle": 0, + "content": "+ Synthetic positive caption: A double-decker bus parked at a bus station with a few passengers getting on." + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.598, + 0.812, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.6, + 0.897, + 0.611 + ], + "angle": 0, + "content": "Wrong Generation" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.616, + 0.901, + 0.631 + ], + "angle": 0, + "content": "(b) Inconsistency in cross-modal alignment quality of synthetic samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Figure 1. Challenges in generating and training on synthetic data: (a) When generating an image with subtle variations based on a real image and a target caption specifying the variations, an image editing model [6] struggles with text alignment (middle), while an image generation model [75] fails to maintain image fidelity (right). (b) Synthetic positive and negative image_caption pairs show different levels of alignment quality. The subtle variations in the synthetic negative caption (left) make it difficult to distinguish from the positive; the over-modified negative caption (middle) is easy to distinguish; and the hallucinated content in the synthetic positive caption (right) results in an incorrect positive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.885 + ], + "angle": 0, + "content": "ing shortcut features [23] rather than genuinely learning nuanced distinctions. While collecting training samples with subtle variations could enhance compositionality, this approach is time-consuming and labor-intensive, rendering it impractical at scale." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Advances in generative models [7, 12, 20, 43, 61, 75," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24849" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "78, 91, 105, 110] now facilitate the synthetic generation of training samples with subtle variations between paired samples [11, 18, 49, 68, 71, 79, 83, 86, 106, 108, 109]. By starting with real image-caption pairs, generative models can create subtle edits to both captions and images, providing valuable training data with minimal manual effort. These generated variations enable VLMs to enhance their compositionality by learning from the nuanced differences." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.216, + 0.483, + 0.576 + ], + "angle": 0, + "content": "However, generating and training on synthetic data presents two key challenges. The first is the difficulty of efficiently and accurately creating precise variations in synthetic images. To generate large-scale training data with precise variations, the image generation process must meet three criteria: efficiency in producing large quantities of images, text alignment of the generated image to the corresponding caption, which is subtly changed from the original caption, and image fidelity in ensuring the generated image otherwise closely matches their real counterpart. Nevertheless, current image generation methods struggle to meet all three criteria simultaneously. Per-sample optimization methods [44, 82, 94, 112] lack efficiency, while zero-shot image editing methods [6, 22, 63] and text-to-image (T2I) models [75, 78] often fail to achieve proper text alignment and image fidelity, respectively, as illustrated in Figure 1 (a). The inaccurate synthetic variations produced by these models can mislead the learning of VLMs. To alleviate this issue, we take a fast T2I model, which excels in efficiency and text alignment, and inject real image features into the text prompt features in order to enhance image fidelity. By combining this approach with AdaIN [35], we manage to substantially improve the image generation process per the three criteria." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.762 + ], + "angle": 0, + "content": "Still, the generated data do not perfectly meet the text alignment criterion, and we propose to deal with the remaining problems with an innovative approach to model training. As Figure 1 (b) shows, the similarity between synthetic positive and negative pairs may vary, resulting in a mixture of hard text-image pairs, easy pairs, and incorrect pairs. Instead of treating all positive and negative samples uniformly, we propose a novel loss that differentiates between positive, hard negative, and easy negative samples. Further, we propose an adaptive margin that helps to filter out potentially incorrect generations and focus the learning on informative hard samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In summary, we propose SPARCL, which integrates image feature injection into a fast T2I model to improve the quality of synthetic variations in images and employs an adaptive margin loss to leverage the varying alignment quality in synthetic samples for training VLMs. Evaluations on four compositional understanding benchmarks demonstrate that SPARCL significantly enhances VLM compositionality, improving CLIP by over \\(5\\%\\) on VL-CheckList [113] and \\(7\\%\\) on SugarCrepe [32], while surpassing state-of-the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "art methods by \\(1\\%\\) and \\(2\\%\\) on the two benchmarks. The main contributions of this paper are as follows: (1) we propose image feature injection to enhance the quality of synthetic variations in images, which provide valuable training data to improve the compositionality of VLMs; (2) we introduce an adaptive margin loss to leverage varying levels of cross-modal alignment in synthetic samples to effectively differentiate positive and negative samples; (3) experimental results validate that SPARCL significantly improves the compositional understanding capabilities of CLIP models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.259, + 0.655, + 0.274 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.868, + 0.3 + ], + "angle": 0, + "content": "2.1. Limitations in Compositionality of VLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.58 + ], + "angle": 0, + "content": "While VLMs excel in many multi-modal tasks [40, 52, 53, 74, 84, 101, 115], they still struggle with compositional understanding—the ability to interpret novel combinations of known visual and textual components. Benchmarks like What'sUp [42] reveal difficulties in understanding spatial relationships, SPEC [70] highlights issues with object size, position, and count, and ARO [106] uncovers limitations in understanding attributes, relations, and word order. Winoground [88], SNARE [96], and VL-CheckList [113] also expose these shortcomings. SugarCrepe [32] addresses hackable biases in prior benchmarks, where text-only models achieve artificially high performance, by introducing fluent and meaningful hard negatives. Their findings suggest that previous benchmarks overestimated compositional understanding. Building on this, SugarCrepe++ [19] further introduces semantically equivalent but lexically varied captions as hard positives and shows the difficulties of VLMs in distinguishing between lexical variations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.842, + 0.607 + ], + "angle": 0, + "content": "2.2. Improving Compositionality of VLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Prior approaches to improving the compositionality of VLMs can be broadly classified into the following categories: (1) Leveraging detailed image captions: Detailed captions from dense captioning models [17, 54], simulation platforms [9], and video annotations [45] are collected to train VLMs. However, these samples often lack pairs with subtle variations, limiting their contribution to compositionality. (2) Distilling from pretrained models: SDS-CLIP [3], SF-CLIP [80] and IL-CLIP [114] distill knowledge from pretrained image generation models and visual-language foundation models. IL-CLIP [114] refines representations through iterative learning with pretrained vision and language agents. However, pretrained models also face limitations in compositionality [34, 90]. (3) Incorporating structural knowledge: MosaiCLIP [85] and Structure-CLIP [36] incorporate scene graph knowledge in text features. CLIP-SGVL [28] and 3VL [103] train VLMs to predict scene graphs. [66] utilizes scene graphs as prompts to elicit compositional knowledge from VLMs without fur" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24850" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.44 + ], + "angle": 0, + "content": "other training. However, these methods rely on models trained with expensive dense structure annotations (e.g., scene graphs). (4) Utilizing synthetic negative samples: Rule-based tools [30] or large language models (LLMs) [15, 73] are used to generate negative captions by editing real captions [10, 11, 18, 68, 83, 86, 106, 109]. Image generation models [75] are also used to create or edit images for training [49, 71, 79, 108]. Despite their utility, efficiently generating large amounts of images with precise variations remains challenging due to the inherent limitations of generative models. (5) Applying fine-grained alignment constraints: MCD [46] enforces multi-scale alignment across images with varying augmentations and the corresponding text captions. SPARC [5] learns local alignment by associating each text token with a group of local image patches. CE-CLIP [109] applies intra-modal contrastive loss and cross-modal ranking loss to improve alignment. However, uniform supervision applied to synthetic samples with varying alignment quality limits their effectiveness. In this paper, we propose SPARCL to address two key challenges in learning from synthetic data: the difficulty in accurately creating precise variations and the inconsistency in cross-modal alignment quality in synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.448, + 0.352, + 0.465 + ], + "angle": 0, + "content": "2.3. Training with Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.485, + 0.727 + ], + "angle": 0, + "content": "Synthetic data for training machine learning models have been studied in various fields [14, 26, 48, 64, 67, 76, 77, 93, 99]. Synthetic data generated through simulations and graphics engines supports a wide range of tasks [16, 72, 95]. However, these synthetic datasets often diverge significantly from real-world data. Recent advances in generative models [7, 12, 43, 61, 75, 78, 91, 105, 110] have made it possible to synthesize data that more closely resembles real-world scenarios. Synthetic data are widely used in both language tasks [21, 31, 60, 64, 65, 87, 97, 98, 100, 116] and vision tasks [1, 2, 4, 25, 29, 39, 57, 81, 89, 92, 111]. Synthetic data are often noisy, necessitating noise-resistant training methods [13, 37, 41, 50, 51, 102], especially those designed for contrastive learning [38, 56]. In this paper, we generate multimodal samples with subtle variations and filter out potentially incorrect ones during training to improve the compositionality of VLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.194, + 0.755 + ], + "angle": 0, + "content": "3. SPARCL" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To improve the compositional understanding abilities of VLMs, we propose Synthetic Perturbations for Advancing Robust Compositional Learning, or SPARCL, which generates multimodal samples with subtle variations from real samples and trains VLMs to learn nuanced differences through synthetic data. In the generation phase, SPARCL creates positive and negative captions with slight variations from real captions using an LLM, then generates images based on the real image and these modified captions through" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.244 + ], + "angle": 0, + "content": "a fast T2I model. To enhance the quality of subtle variations in synthetic images, we introduce image feature injection, which integrates real image features into the text prompt features of the T2I model to improve fidelity to the real image, as detailed in Sec. 3.2. In the training phase, SPARCL employs an adaptive margin loss that leverages varying levels of multimodal alignment in the synthetic samples to effectively learn informative nuanced distinctions, as described in Sec. 3.3. The framework of SPARCL is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.88, + 0.269 + ], + "angle": 0, + "content": "3.1. Generating Negative and Positive Captions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.909, + 0.532 + ], + "angle": 0, + "content": "Captions with subtle variations are crucial for learning compositional knowledge, as shown by previous work that generates captions by randomly swapping or replacing nouns and adjectives [106, 109]. However, manually designed generation rules often introduce nonsensical or grammatically incorrect artifacts, creating shortcut features [23] that obstruct true compositional understanding [32]. To address this issue, we use an LLM to generate natural synthetic captions. To further mitigate the impact of generative artifacts, we generate both negative and positive captions, ensuring that VLMs cannot easily differentiate negative captions from positive ones based solely on artifacts. We denote the \\(i^{th}\\) real image-caption pair in the training set as \\((\\boldsymbol{I}_i^r,\\boldsymbol{T}_i^r)\\). Given a real caption \\(\\boldsymbol{T}_i^r\\), we prompt the LLM to generate a synthetic negative caption \\(\\boldsymbol{T}_i^{sn}\\) and a synthetic positive caption \\(\\boldsymbol{T}_i^{sp}\\), as specified by the prompts in Figure A1 in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.54, + 0.907, + 0.557 + ], + "angle": 0, + "content": "3.2. Generating Images via Image Feature Injection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.7 + ], + "angle": 0, + "content": "Synthetic images that exhibit subtle variations from real images while aligning with the captions are valuable but challenging to generate, as discussed in Sec. 1. Although previous works [49, 71, 79, 108] have utilized object segmentation or filtered dissimilar generations to improve fidelity to real images, they struggle with manipulating relationships or lack efficiency. We aim to enhance fidelity by injecting image features into a fast T2I model [61], which already achieves high efficiency and text alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.796 + ], + "angle": 0, + "content": "Image Feature Injection. The images generated by T2I models lack image fidelity to real images, as no real image information is input to the models. To enhance the fidelity of synthetic images, we inject real image features into the text prompt features, enabling the model to incorporate information from the real images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In T2I models, content and style are separated in the semantic and padding embeddings. The semantic embeddings (before the embedding of the [EOS] token) usually capture most of the image content in the text prompts, while the padding embeddings (after the [EOS] token) usually represent the image style [104]. Therefore, we can inject real image features into the padding embeddings to guide" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24851" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.095, + 0.394, + 0.11 + ], + "angle": 0, + "content": "(a) Generating synthetic captions and images" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.114, + 0.568, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.095, + 0.901, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.295, + 0.384, + 0.312 + ], + "angle": 0, + "content": "(b) Training with real and synthetic samples" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.313, + 0.39, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.298, + 0.543, + 0.312 + ], + "angle": 0, + "content": "Adaptive Margin Loss" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.313, + 0.744, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.298, + 0.899, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 2. An overview of SPARCL. (a) Starting with a real image-caption pair, we generate synthetic positive and negative pairs with subtle variations using an LLM and a fast T2I model. To improve the quality of subtle variations in synthetic images, we introduce image feature injection to reduce unintended variations from a standard T2I model (see Sec. 3.2). (b) We train the VLM using both real and synthetic samples. In addition to a sigmoid loss for distinguishing positive and negative pairs, we apply an adaptive margin loss that leverages varying alignment levels across training samples to learn informative nuanced distinctions (see Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.61 + ], + "angle": 0, + "content": "the model in generating images with a similar style to real images, without affecting the alignment with the captions, thanks to the decoupling of content and style in the prompt embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.778 + ], + "angle": 0, + "content": "We extract features for a real image \\( \\pmb{I}_i^r \\) and the corresponding synthetic caption \\( \\pmb{T}_i^s \\) (\\( s \\in \\{sn, sp\\} \\)) using two aligned feature encoders (e.g., CLIP image and text encoders). The image embedding \\( \\pmb{f}_i^r \\) is the encoder output at the position of the [CLS] token. The text embedding is the text encoder output \\( e_i^s = \\langle e_{i,j}^s \\rangle_{j=1}^L \\), where \\( L \\) is the maximum sequence length. Let the index of the [EOS] token in \\( e_i^s \\) be \\( k_i^s \\), \\( k_i^s \\leq L \\). That is, the padding token [PAD] is used as input to the text encoder at positions between \\( k_i^s \\) and \\( L \\). The key step of SPARCL is to replace the text embeddings at those positions with the image embedding, producing a new sequence of text embeddings, \\( \\hat{e}_i^s \\)," + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.784, + 0.483, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\hat {e} _ {i} ^ {s} = \\left\\langle e _ {i, 1} ^ {s}, e _ {i, 2} ^ {s}, \\dots , e _ {i, k _ {i}} ^ {s}, \\underbrace {\\boldsymbol {f} _ {i} ^ {r} , \\dots , \\boldsymbol {f} _ {i} ^ {r}} _ {L - k _ {i} ^ {s} \\text {t i m e s}} \\right\\rangle . \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The embeddings \\(\\hat{e}_i^s\\) is then used as input to the T2I model to generate a synthetic image \\(\\tilde{I}_i^s\\). This process, depicted in Figure 2, reduces unintended variations in synthetic images generated by the standard T2I model and enhances image fidelity to real images, thereby improving the quality of" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.551, + 0.692, + 0.565 + ], + "angle": 0, + "content": "synthetic subtle variations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.726 + ], + "angle": 0, + "content": "Style Transfer. To further reduce the domain gap between synthetic and real images, we use AdaIN [35] to transfer the style of the synthetic image to that of the real image. We use a pretrained AdaIN encoder to extract content features from \\(\\tilde{I}_i^s\\) and style features from \\(I_i^r\\). We normalize the content features using instance normalization and then scale and shift them with the mean and variance of the style features. The transformed features are fed into a pretrained AdaIN decoder to generate \\(I_i^s\\), which is subsequently used for model training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.873, + 0.758 + ], + "angle": 0, + "content": "3.3. Training with Real and Synthetic Samples" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "After generating captions and images, each real image-caption pair \\((I_i^r, T_i^r)\\) is extended to \\((I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})\\) by adding one synthetic negative and one synthetic positive pair. We then train the VLM using these extended samples. Given a batch of \\(n\\) sample groups, \\(\\{(I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})\\}_{i=1}^n\\), we organize the images by concatenating all real, synthetic negative, and synthetic positive images into an image batch \\(I^B\\) of \\(3n\\) images, and similar all captions into a caption" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24852" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.262, + 0.106 + ], + "angle": 0, + "content": "batch \\(\\pmb{T}^B\\) of \\(3n\\) captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.106, + 0.484, + 0.304 + ], + "angle": 0, + "content": "We calculate the similarity between each image-caption pair in the batch. The similarity between the \\(i^{\\mathrm{th}}\\) image \\(\\pmb{I}_i^B\\) and the \\(j^{\\mathrm{th}}\\) caption \\(\\pmb{T}_j^B\\) is given by \\(S_{i,j} = c(\\mathcal{E}_I(\\pmb{I}_i^B),\\mathcal{E}_T(\\pmb{T}_j^B))\\), where \\(c(\\cdot ,\\cdot)\\) denotes cosine similarity, and \\(\\mathcal{E}_I\\) and \\(\\mathcal{E}_T\\) represent the image and text encoders of the VLM, respectively. We then define the ground-truth alignment variable \\(M_{i,j}\\), which takes value 1 if the image matches with the caption, and -1 otherwise. Clearly, the only positive image-caption pairs are \\((\\pmb{I}_i^r,\\pmb{T}_i^r),(I_i^{sn},T_i^{sn}),(I_i^{sp},T_i^{sp}),(I_i^r,T_i^{sp})\\) and \\((I_i^{sp},T_i^r)\\). To account for multiple positive associations of an image or a caption, we apply a sigmoid-based contrastive loss [8, 107] to encourage higher similarity for positive pairs and lower similarity for negative pairs," + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.312, + 0.483, + 0.35 + ], + "angle": 0, + "content": "\\[\nL _ {c o n} = - \\frac {1}{3 n} \\sum_ {i = 1} ^ {3 n} \\sum_ {j = 1} ^ {3 n} \\log \\frac {1}{1 + \\exp \\left(- M _ {i , j} \\left(S _ {i , j} / \\tau + b\\right)\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.358, + 0.473, + 0.373 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter and \\(b\\) is a bias term." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.529 + ], + "angle": 0, + "content": "Adaptive Margin Loss. Despite best efforts toward controlling generative models, synthetic samples may still have variable quality in terms of the alignment between text and imagery. That is, purported positive (resp. negative) pairs may not be semantically similar (resp. dissimilar). To account for variation in synthetic data quality, we propose to differentiate between real data and synthetic data and between easy negatives and hard negatives. Further, we propose an adaptive margin that filters out potential incorrect samples and prioritizes learning from hard samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.605 + ], + "angle": 0, + "content": "For each image in a batch, we define four sets of captions, a positive set \\(\\mathbb{P}\\), a hard negative set \\(\\mathbb{N}_h\\), an easy negative set \\(\\mathbb{N}_e\\), and a real negative set \\(\\mathbb{N}_r\\), which represent different levels of alignment within the batch. If the image is a real or synthetic positive image," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.606, + 0.483, + 0.636 + ], + "angle": 0, + "content": "- \\(\\mathbb{P}\\) contains the real positive captions \\(T_{i}^{r}\\) and the synthetic positive captions \\(T_{i}^{sp}\\) of the current, \\(i\\)-th image." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.636, + 0.483, + 0.665 + ], + "angle": 0, + "content": "- \\(\\mathbb{N}_h\\) contains the synthetic negative captions \\(T_i^{sn}\\) of the current image." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.665, + 0.483, + 0.696 + ], + "angle": 0, + "content": "- \\(\\mathbb{N}_e = \\{\\pmb{T}_j^r,\\pmb{T}_j^{sp},\\pmb{T}_j^{sn}|i\\neq j\\}\\) contains all real and synthetic captions from all other images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.696, + 0.483, + 0.726 + ], + "angle": 0, + "content": "- \\(\\mathbb{N}_r = \\{\\pmb{T}_j^r | i \\neq j\\}\\) contains all real captions belonging to the other images." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.606, + 0.483, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.756 + ], + "angle": 0, + "content": "If the image is a synthetic negative image, the sets \\(\\mathbb{N}_e\\) and \\(\\mathbb{N}_r\\) are unchanged but \\(\\mathbb{P}\\) and \\(\\mathbb{N}_h\\) differ:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.757, + 0.418, + 0.772 + ], + "angle": 0, + "content": "- \\(\\mathbb{P}\\) contains the synthetic negative captions \\(T_{i}^{sn}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.772, + 0.275, + 0.787 + ], + "angle": 0, + "content": "- \\(\\mathbb{N}_h\\) contains \\(T_i^r\\) and \\(T_i^{sp}\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.757, + 0.418, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.787, + 0.401, + 0.802 + ], + "angle": 0, + "content": "We define the margin loss for the \\(i^{th}\\) image as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.806, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {m a r, i} ^ {I} = \\frac {1}{| \\mathbb {P} | \\cdot | \\mathbb {N} _ {h} |} \\sum_ {j _ {1} \\in \\mathbb {P}, j _ {2} \\in \\mathbb {N} _ {h}} \\max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}) \\\\ + \\frac {1}{\\left| \\mathbb {N} _ {h} \\right| \\cdot \\left| \\mathbb {N} _ {e} \\right|} \\sum_ {j _ {1} \\in \\mathbb {N} _ {h}, j _ {2} \\in \\mathbb {N} _ {e}} \\max \\left(0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}\\right) \\tag {3} \\\\ + \\frac {\\alpha}{| \\mathbb {P} | \\cdot | \\mathbb {N} _ {r} |} \\sum_ {j _ {1} \\in \\mathbb {P}, j _ {2} \\in \\mathbb {N} _ {r}} \\max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "This loss encourages positive pairs to have higher similarity scores than hard negative pairs, which should have higher similarity than easy negative pairs. Additionally, a weight \\(\\alpha > 1\\) is applied to comparisons between positive pairs and real negative pairs to emphasize these comparisons, as real samples are generally correct, making comparisons involving them more reliable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.905, + 0.253 + ], + "angle": 0, + "content": "Lastly, we propose an adaptive margin in the above loss. Let \\( d = S_{i,j_1} - S_{i,j_2} \\) denote the difference between two similarity scores in Eq. (3). The adaptive margin \\( m \\) is then computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.263, + 0.905, + 0.308 + ], + "angle": 0, + "content": "\\[\nm = \\left\\{ \\begin{array}{l l} d, & d < \\beta \\\\ (\\frac {m _ {0} - d}{m _ {0} - \\beta} \\gamma + 1) m _ {0}, & \\beta \\leq d \\leq m _ {0} \\\\ m _ {0} & d > m _ {0} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.906, + 0.513 + ], + "angle": 0, + "content": "where \\(m_0\\) is a base margin, \\(\\beta < 0\\) is a cutoff threshold, \\(\\gamma\\) is a scaling factor. The adaptive margin is designed with the following rationale: when \\(d < \\beta\\), the purported positive pair is much less similar than the purported negative pair, suggesting the presence of incorrect or mislabeled samples, so we zero out the loss by setting the margin to \\(d\\). For \\(\\beta \\leq d \\leq m_0\\), the margin is scaled, with smaller differences receiving larger margins to emphasize learning from harder samples. When \\(d > m_0\\), the samples are already well-separated with the margin \\(m_0\\), so the margin is capped at \\(m_0\\), again leading to zero loss. We visualize both the adaptive margin and the corresponding loss as functions of \\(d\\) in Figure 2 (b)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.513, + 0.906, + 0.6 + ], + "angle": 0, + "content": "The loss for all images is given by \\( L_{mar}^{I} = \\frac{1}{3n}\\sum_{i}L_{mar,i}^{I} \\). Analogously, we compute \\( L_{mar}^{T} \\) for all captions. The total adaptive margin loss is \\( L_{mar} = L_{mar}^{I} + L_{mar}^{T} \\). Finally, the overall training loss is a weighted combination of the contrastive loss and the adaptive margin loss, with weight \\( \\lambda \\):" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.611, + 0.905, + 0.626 + ], + "angle": 0, + "content": "\\[\nL = L _ {c o n} + \\lambda L _ {m a r}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.645, + 0.668 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.615, + 0.69 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.906, + 0.804 + ], + "angle": 0, + "content": "Training. We use the COCO-2014 dataset [55] as the training data source. The training set consists of 82,783 images, each paired with five captions. For each real image-caption pair, we generate one positive and one negative synthetic pair. In line with previous approaches [18, 106, 109], we train the VLMs using both the original COCO-2014 training data and the synthetic samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation. We evaluate our model on four vision-language compositional understanding benchmarks: (1) ARO [106], which consists of 23,937 cases for relation understanding and 28,748 for attribute understanding. We exclude the subsets for order understanding, as they contain significant nonsensical and non-fluent artifacts [32]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24853" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of accuracy (%) between SPARCL and baselines on four benchmarks. \"img.\" represents images, \"cap.\" represents captions, \"syn.\" represents synthetic data." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.128, + 0.897, + 0.446 + ], + "angle": 0, + "content": "
MethodTraining DataAROVL-CheckListSugarCrepeSugarCrepe++
Source# real img.# real cap.# syn. img.# syn. cap.
CLIP [74] (Zero-Shot)-----61.173.273.459.8
CLIP [74] (Finetune)COCO82K410K0064.172.879.962.3
SDS-CLIP [3]COCO82K410K0057.5---
[79]COCO0082K82K65.069.9--
AMR-NegCLIP [83]COCO100K100K0500K79.4-85.2-
NegCLIP [106]COCO100K100K0500K76.074.682.564.9
MosaiCLIP [85]COCO109K109K0981K80.376.8--
FSC-CLIP [68]COCO100K100K01.5M-77.285.1-
CE-CLIP [109]COCO82K410K02M79.776.385.2-
COMO [49]COCO113K567K567K567K-76.9--
SPARCL (our method)COCO82K410K820K820K77.279.287.166.1
SPEC [70]LAION20K20K20K20K70.1---
CounterCurate [108]Flickr30K30K150K150K--82.8-
[18]CC3M3M3M09M-75.3-55.3
CE-CLIP+ [109]COCO+CC3M3M3M015M80.479.387.5-
CLOVE [11]LAION-COCO>1B>1B0>1B73.2-85.1-
IL-CLIP [114]CC12M12M12M00--70.3-
SF-CLIP [80]YFCC15M15M15M00--71.2-
syn-CLIP [9]SyViC00>1M>1M69.274.8--
FiGCLIP [45]VidSitu20K videos0067.0-74.6-
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.487, + 0.685 + ], + "angle": 0, + "content": "(2) VL-CheckList [113], a large-scale benchmark with over 100,000 samples, evaluates compositionality across subsets of objects, attributes, and relationships, which are further divided into various fine-grained categories. (3) SugarCrepe [32] includes 7,000 test cases across seven subsets. In the above three benchmarks, each test case containing one image, one positive caption, and one negative caption. (4) SugarCrepe++ [19] includes 4,757 test samples across five subsets, where each test case consists of one image, two positive captions, and one negative caption. All benchmarks involve classifying captions as positive or negative for the given images. We report the average accuracy across all subsets of each benchmark and include the accuracy for each subset in Appendix A3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.691, + 0.31, + 0.708 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.486, + 0.821 + ], + "angle": 0, + "content": "Models. We use ViT-B/32 and ViT-L/14 architectures from OpenAI's CLIP model [74] as our base models, initialized with pretrained checkpoints. Following syn-CLIP [9], we integrate LoRA adapters [33] into both the image and text encoders of CLIP to improve training efficiency and mitigate knowledge forgetting. Only the LoRA adapters are fine-tuned during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Training Setups. We use the AdamW optimizer [59] with a cosine learning rate schedule [58]. Training is conducted on two Tesla V100 GPUs, with a batch size of 128 sample groups for ViT-B/32 and 16 for ViT-L/14. The base learning rate is set to 0.01 for a total batch size of 256, and scaled" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.908, + 0.548 + ], + "angle": 0, + "content": "linearly [24] based on the actual batch size. Training is performed for 3,000 steps for ViT-B/32 and 15,000 steps for ViT-L/14, corresponding to fewer than 5 epochs in previous studies [18, 106, 109]. More details and hyperparameters settings are in Appendix A2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.655, + 0.575 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.703 + ], + "angle": 0, + "content": "We compare SPARCL with several baseline methods: (1) methods that utilize synthetic samples for training [49, 68, 70, 79, 83, 106, 108, 109], (2) methods that distill knowledge from pretrained models [3, 11, 80, 114], (3) methods that incorporate knowledge from scene graphs [18, 85], and (4) methods that leverage detailed image captions [9, 45]. Both SPARCL and the baseline methods use ViT-B/32 as the base model. The results are presented in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We observe that SPARCL achieves the best performance on VL-CheckList, SugarCrepe, and SugarCrepe++ compared to baselines trained on the same data source, namely COCO. Specifically, SPARCL surpasses the strongest baseline by \\(1.4\\%\\) on VL-CheckList and \\(2.5\\%\\) on SugarCrepe, underscoring its effectiveness in enhancing compositional understanding. However, SPARCL performs worse on ARO than some baselines. We hypothesize that this may be due to nonsensical or grammatically incorrect artifacts in the ARO test samples [32], which could favor methods that utilize rule-based synthetic training samples (e.g., [85]) containing similar artifacts. Notably, even with only \\(50\\%\\) of the training data, SPARCL outperforms most base" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24854" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 2. Ablated performance (\\%) of SPARCL. \"SynCap\" refers to synthetic captions, \"SynImg\" refers to synthetic images, \"FeatInj\" denotes image feature injection, and \"CompSet\" indicates the comparison sets used in Eq. (3)." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.128, + 0.894, + 0.438 + ], + "angle": 0, + "content": "
ModelVariantSynCapSynImgFeatInjAdaINAdaptive Margin LossAROVL-CheckListSugarCrepeSugarCrepe++Average
CompSetMargin
ViT-B/32#1XXXXX60.4972.6179.3664.8569.33
#2✓XXXX71.7773.5286.3564.3273.99
#3X✓XXX62.6271.7079.9764.8469.79
#4✓✓XXX71.8675.5485.4365.2274.51
#5✓✓✓XX73.4076.5685.5465.7875.32
#6✓✓X✓X73.7975.7285.7964.4974.95
#7✓✓✓✓X74.1276.3585.4066.4475.58
#8✓✓✓✓AllFixed76.7978.5987.0865.4276.97
#9✓✓✓✓AllAdaptive77.1579.1687.1166.1277.38
#10✓✓✓✓AllAdaptive Inversed76.6779.2686.6865.3076.97
#11✓✓✓✓Only (P, Nh)Adaptive77.4880.4886.1564.7077.20
ViT-L/14#1XXXXX59.8073.2681.4964.9069.86
#2✓XXXX72.1675.4487.3864.4474.85
#3X✓XXX60.7072.2382.5866.6370.54
#4✓✓XXX75.2078.2987.2964.6176.35
#5✓✓✓XX75.1178.5887.5464.5776.45
#6✓✓X✓X74.8879.1087.5064.6276.52
#7✓✓✓✓X74.9380.0487.7965.3077.01
#8✓✓✓✓AllFixed75.2180.7588.0266.4177.60
#9✓✓✓✓AllAdaptive75.8380.8188.2366.8377.93
#10✓✓✓✓AllAdaptive Inversed75.1580.7187.9366.6177.60
#11✓✓✓✓Only (P, Nh)Adaptive76.2680.5787.4266.2877.63
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.482, + 0.599 + ], + "angle": 0, + "content": "line methods (see Table 3 for SPARCL performance with reduced training data). Compared to baselines using additional data sources, SPARCL surpasses or matches their performance. For example, SPARCL outperforms methods in [11, 18] despite their use of more training data. SPARCL performs comparably to CE-CLIP+ [109] on VL-CheckList and SugarCrepe, despite CE-CLIP+ leveraging CC3M as an additional data source and utilizing significantly more samples." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.612, + 0.348, + 0.629 + ], + "angle": 0, + "content": "4.4. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.682 + ], + "angle": 0, + "content": "We perform ablation studies to assess the impact of each component and design choice in SPARCL. The results, presented in Table 2, show the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Synthetic captions vs. synthetic images. To analyze the impact of synthetic captions and images, we compare Variant #1 (using only real samples) with Variant #2 (using real samples and synthetic captions) and #3 (using real samples and synthetic images). Variant #2 shows a significant improvement over Variant #1, with the average accuracy improving from \\(69.33\\%\\) to \\(73.99\\%\\) for ViT-B/32. However, Variant #3 obtains only a marginal improvements of about \\(0.5\\%\\). These results suggest that synthetic captions substantially enhance compositional understanding, while synthetic images alone provide limited benefit, which aligns with findings in [69]. We hypothesize that generative artifacts in synthetic images are more pronounced than in captions, which negatively affects the effective learning of nu" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.461, + 0.907, + 0.507 + ], + "angle": 0, + "content": "anced distinctions by VLMs. When both synthetic captions and images are used (Variant #4), performance improves further, indicating a synergistic effect." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.909, + 0.879 + ], + "angle": 0, + "content": "Image feature injection vs. AdaIN. To analyze the impact of image feature injection and AdaIN, we compare Variant #4 (without image feature injection or AdaIN) with Variant #5 (with image feature injection only) and #6 (with AdaIN only). Using ViT-B/32, we observe that Variant #5 outperforms Variant #4 across all four benchmarks, with an average gap of \\(0.8\\%\\). While Variant #6 shows a smaller improvement of \\(0.4\\%\\), it outperforms Variant #5 on ARO and SugarCrepe. These results suggest that image feature injection is more effective for improving compositionality than AdaIN, although the two methods are somewhat complementary. Through the improvements obtained using ViT-L/14 are a little bit different, both techniques improve performance over Variant #4. Combining both methods in Variant #7 leads to further improvements, with a more than \\(1\\%\\) increase in average accuracy for ViT-B/32 and a \\(0.65\\%\\) increase for ViT-L/14 over Variant #4. This highlights the importance of reducing unintended changes in synthetic images, as such changes could cause the model to rely on them for distinguishing positive and negative samples rather than the semantic differences specified in the corresponding captions. Examples of synthetic images in Figure A3 and A4 in the Appendix show how image feature injection helps mitigate these unintended changes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Effects of the margin loss and adaptive margin strate-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24855" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 3. Performance (\\%) of SPARCL using different subsets of training samples. \"Neg.\" represents synthetic negative samples, \"Pos.\" represents synthetic positive samples, and \"Prop.\" represents the proportion of training samples." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.158, + 0.49, + 0.243 + ], + "angle": 0, + "content": "
VariantNeg.Pos.Prop.AROVL-CheckListSugarCrepeSugarCrepe++Average
#7✓×100%72.1274.6586.4157.9972.79
#7✓✓100%74.1276.3585.4066.4475.58
#9✓✓20%75.9077.5685.7864.4475.92
#9✓✓50%77.9079.3886.7164.2177.05
#9✓✓100%77.1579.1687.1166.1277.38
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.268, + 0.486, + 0.526 + ], + "angle": 0, + "content": "gies. We compare three variants with distinct margin approaches: (1) Fixed (Variant #8): a fixed margin without adaptive adjustments; (2) Adaptive (Variant #9): the margin is calculated according to Eq. (4); and (3) Adaptive Inversed (Variant #10): a larger margin is applied to samples with higher similarity differences, representing an inverse version of Eq. (4) that prioritizes learning from easier samples. Comparing Variant #8 with Variant #7, we observe that #8 improves by over \\(1\\%\\) for ViT-B/32 and \\(0.6\\%\\) for ViT-L/14 in average accuracy, suggesting that margin loss effectively aids in compositional understanding. Between Variants #8, #9, and #10, we find that Variant #9 outperforms Variant #8 across all four benchmarks, with an average boost of over \\(0.3\\%\\) for both ViT-B/32 and ViT-L/14. In contrast, Variant #10 results in a performance decline on three benchmarks and no improvement in average accuracy. These findings highlight the superior effectiveness of adaptive margins." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.532, + 0.486, + 0.851 + ], + "angle": 0, + "content": "Is applying adaptive margin loss only to hard samples sufficient? To explore whether applying margin loss solely to hard samples is adequate, we construct Variant #11, which uses only the positive set \\(\\mathbb{P}\\) and the hard negative set \\(\\mathbb{N}_h\\) for margin loss calculation in Eq. (3), which is similar to learning strategies in [49, 109]. When comparing Variant #11 with Variant #8 and #9, we observe that Variant #11 performs better on ARO and VL-CheckList, but worse on SugarCrepe and SugarCrepe++. We hypothesize that by focusing exclusively on the positive and hard negative sets, the model may learn nonsensical artifacts present in the hard negatives. These artifacts improve performance on ARO and VL-CheckList, where the test samples exhibit similar patterns, but hinder performance on SugarCrepe and SugarCrepe++, which are designed to avoid such patterns [32]. In contrast, Variant #9, which incorporates all sets in Eq. (3) for margin loss calculation, achieves better results on SugarCrepe and SugarCrepe++ with only a minor drop in performance on ARO and VL-CheckList. This suggests that easy negative samples serve as regularization, preventing overfitting to the artifacts in hard negative samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Effects of synthetic positive samples. We compare the performance of training with only synthetic negative samples versus using both synthetic negative and positive samples" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.244 + ], + "angle": 0, + "content": "in Table 3. Based on Variant #7, we train ViT-B/32 with these two combinations of training data. We observe nearly \\(3\\%\\) decrease in average accuracy when using only synthetic negative samples, compared to using both synthetic negative and positive samples. This decrease is particularly noticeable on SugarCrepe++, ARO, and VL-CheckList. These results underscore the importance of incorporating synthetic positive samples in training, as they provide essential information about variations that maintain semantic consistency, which is critical for compositional understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.909, + 0.494 + ], + "angle": 0, + "content": "Influence of model size and training data size. To evaluate the impact of model size on compositional understanding, we compare the performance of ViT-B/32 and ViT-L/14 in Table 2. We find that ViT-L/14 only provides a modest improvement of about \\(0.5\\%\\) of average accuracy over ViT-B/32. This suggests that increasing model size does not significantly enhance compositional understanding. To investigate the effect of training data size, we train ViT-B/32 using \\(20\\%\\) and \\(50\\%\\) of randomly sampled training data and report the results in Table 3. We observe that performance improves as the proportion of training data increases, indicating that a larger training set helps the model better capture compositional knowledge. However, the performance gain diminishes when the training data size increases from \\(50\\%\\) to \\(100\\%\\). Perhaps generating or selecting high-quality data would be more effective than simplistic data scaling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.529 + ], + "angle": 0, + "content": "Additional experiments on hyperparameter analysis can be found in Appendix A3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.634, + 0.567 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.909, + 0.76 + ], + "angle": 0, + "content": "In this paper, we introduce SPARCL to enhance the compositional understanding capabilities of VLMs by generating and training with synthetic data. To tackle two key challenges in using synthetic data—namely, the difficulty of generating accurate variations and the inconsistency in cross-modal alignment quality—SPARCL integrates image feature injection into a T2I model to improve the quality of synthetic variations and introduces an adaptive margin loss to account for varying levels of cross-modal alignment for effectively learning nuanced distinctions. Experiments on four visual-language compositional understanding benchmarks demonstrate the effectiveness of SPARCL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.782, + 0.693, + 0.8 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF-NRFF13-2021-0006), Singapore. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24856" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.172, + 0.483, + 0.226 + ], + "angle": 0, + "content": "[2] Manel Baradad Jurjo, Jonas Wulff, Tongzhou Wang, Phillip Isola, and Antonio Torralba. Learning to see by looking at noise. Advances in Neural Information Processing Systems, 34:2556-2569, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Samyadeep Basu, Maziar Sanjabi, Daniela Massiceti, Shell Xu Hu, and Soheil Feizi. Augmenting clip with improved visio-linguistic reasoning. arXiv preprint arXiv:2307.09233, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.284, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This dataset does not exist: training models from generated images. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.355, + 0.483, + 0.435 + ], + "angle": 0, + "content": "[5] Ioana Bica, Anastasija Ilic, Matthias Bauer, Goker Erdoğan, Matko Bošnjak, Christos Kaplanis, Alexey A Gritsenko, Matthias Minderer, Charles Blundell, Razvan Pascanu, et al. Improving fine-grained understanding in image-text pre-training. arXiv preprint arXiv:2401.09865, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.439, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.509, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.593, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Adrian Bulat, Yassine Ouali, and Georgios Tzimiropoulos. Fff: Fixing flawed foundations in contrastive pre-training results in very strong vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14172-14182, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.664, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[9] Paola Cascante-Bonilla, Khaled Shehada, James Seale Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gul Varol, Aude Oliva, Vicente Ordonez, Rogerio Feris, et al. Going beyond nouns with vision & language models using synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20155-20165, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[10] Paola Cascante-Bonilla, Yu Hou, Yang Trista Cao, Hal Daumé III, and Rachel Rudinger. Natural language inference improves compositionality in vision-language models. arXiv preprint arXiv:2410.22315, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Santiago Castro, Amir Ziai, Avneesh Saluja, Zhuoning Yuan, and Rada Mihalcea. Clove: Encoding compositional language in contrastive vision-language models. arXiv preprint arXiv:2402.15021, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy," + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[13] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Yabo Dan, Yong Zhao, Xiang Li, Shaobo Li, Ming Hu, and Jianjun Hu. Generative adversarial networks (gan) based efficient sampling of chemical composition space for inverse design of inorganic materials. npj Computational Materials, 6(1):84, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.278, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.321, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.406, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[17] Sivan Doveh, Assaf Arbelle, Sivan Harary, Amit Alfassy, Roei Herzig, Donghyun Kim, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, et al. Dense and aligned captions (dac) promote compositional reasoning in v1 models. arXiv preprint arXiv:2305.19595, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.477, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[18] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[19] Sri Harsha Dumpala, Aman Jaiswal, Chandramouli Sastry, Evangelos Milios, Sageev Oore, and Hassan Sajjad. Sugarcrepe++ dataset: Vision-language model sensitivity to semantic and lexical alterations. arXiv preprint arXiv:2406.11171, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[21] Jiahui Gao, Renjie Pi, Lin Yong, Hang Xu, Jiacheng Ye, Zhiyong Wu, Weizhong Zhang, Xiaodan Liang, Zhenguo Li, and Lingpeng Kong. Self-guided noise-free data generation for efficient zero-shot learning. In International Conference on Learning Representations (ICLR 2023), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[22] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. arXiv preprint arXiv:2403.14602, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural net" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24857" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "works.Nature Machine Intelligence,2(11):665-673,2020. 1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[24] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[25] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[26] Xuanli He, Islam Nassar, Jamie Kiros, Gholamreza Haffari, and Mohammad Norouzi. Generate, annotate, and learn: Nlp with synthetic text. Transactions of the Association for Computational Linguistics, 10:826-842, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[27] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.377, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[28] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arbelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[29] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[30] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1):411-420, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[31] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language models with (almost) no human labor. arXiv preprint arXiv:2212.09689, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[32] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[33] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[34] Hang Hua, Yunlong Tang, Ziyun Zeng, Liangliang Cao, Zhengyuan Yang, Hangfeng He, Chenliang Xu, and Jiebo Luo. Mmcomposition: Revisiting the compositionality of pre-trained vision-language models. arXiv preprint arXiv:2410.09733, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Yufeng Huang, Jiji Tang, Zhuo Chen, Rongsheng Zhang, Xinfeng Zhang, Weijie Chen, Zeng Zhao, Zhou Zhao," + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Tangjie Lv, Zhipeng Hu, et al. Structure-clip: Towards scene graph knowledge to enhance multi-modal structured representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2417-2425, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[37] Zhizhong Huang, Junping Zhang, and Hongming Shan. Twin contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.205, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[38] Sarah Ibrahimi, Arnaud Sors, Rafael Sampaio de Rezende, and Stéphane Clinchant. Learning with label noise for image retrieval by selecting interactions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2181-2190, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.274, + 0.904, + 0.315 + ], + "angle": 0, + "content": "[39] Ali Jahanian, Xavier Puig, Yonglong Tian, and Phillip Isola. Generative models as a data source for multiview representation learning. arXiv preprint arXiv:2106.05258, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.316, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[40] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.4, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[41] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.469, + 0.904, + 0.523 + ], + "angle": 0, + "content": "[42] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's\" up\" with vision-language models? investigating their struggle with spatial reasoning. arXiv preprint arXiv:2310.19785, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.595, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.678, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[45] Zeeshan Khan, Makarand Tapaswi, et al. Figclip: Fine-grained clip adaptation via densely annotated videos. arXiv preprint arXiv:2401.07669, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.72, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[46] Bumsoo Kim, Yeonsik Jo, Jinhyung Kim, and Seunghwan Kim. Misalign, contrast then distill: Rethinking misalignments in language-image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[47] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[48] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Chengen Lai, Shengli Song, Sitong Yan, and Guangneng Hu. Improving vision and language concepts understanding with multimodal counterfactual samples. In European" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24858" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 174-191. Springer, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[50] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5051-5059, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[51] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9485-9494, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[52] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[53] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[54] Liunian Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[56] Chang Liu, Han Yu, Boyang Li, Zhiqi Shen, Zhanning Gao, Peiran Ren, Xuansong Xie, Lizhen Cui, and Chunyan Miao. Noise-resistant deep metric learning with ranking-based instance selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6811-6820, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[57] Hao Liu, Tom Zahavy, Volodymyr Mnih, and Satinder Singh. Palm up: Playing in the latent manifold for unsupervised pretraining. Advances in Neural Information Processing Systems, 35:35880-35893, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[58] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[60] Yujie Lu, Wanrong Zhu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Imagination-augmented natural language understanding. arXiv preprint arXiv:2204.08535, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[61] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[62] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[63] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[64] Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. Generating training data with language models: Towards zero-shot language understanding. Advances in Neural Information Processing Systems, 35:462-477, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[65] Yu Meng, Martin Michalski, Jiaxin Huang, Yu Zhang, Tarek Abdelzaher, and Jiawei Han. Tuning language models as training data generators for augmentation-enhanced few-shot learning. In International Conference on Machine Learning, pages 24457-24477. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[66] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.334, + 0.904, + 0.362 + ], + "angle": 0, + "content": "[67] Sergey I Nikolenko. Synthetic data for deep learning. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.364, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[68] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pre-trained vlms for improving vision-linguistic compositionality. arXiv preprint arXiv:2410.05210, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[69] Maitreya Patel, Abhiram Kusumba, Sheng Cheng, Changhoon Kim, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Tripletclip: Improving compositional reasoning of clip via synthetic vision-language negatives. arXiv preprint arXiv:2411.02545, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[70] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. arXiv preprint arXiv:2312.00081, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.548, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize diagnose and optimize: Towards fine-grained vision-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13279-13288, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[72] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.676, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[73] Alec Radford. Improving language understanding by generative pre-training. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.705, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[74] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.789, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[75] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[76] Andrew Rosenberg, Yu Zhang, Bhuvana Ramabhadran, Ye Jia, Pedro Moreno, Yonghui Wu, and Zelin Wu. Speech recognition with augmented synthesized speech. In 2019" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24859" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE automatic speech recognition and understanding workshop (ASRU), pages 996-1002. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[77] Nick Rossenbach, Albert Zeyer, Ralf Schlüter, and Hermann Ney. Generating synthetic audio data for attention-based speech recognition systems. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7069-7073. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.205, + 0.482, + 0.299 + ], + "angle": 0, + "content": "[78] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[79] Ugur Sahin, Hang Li, Qadeer Khan, Daniel Cremers, and Volker Tresp. Enhancing multimodal compositional reasoning of visual language models with generative negative mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5563-5573, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[80] Sepehr Sameni, Kushal Kafle, Hao Tan, and Simon Jenni. Building vision-language models on solid foundations with masked distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14216-14226, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.456, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[81] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[82] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[83] Ziyi Shou and Fangzhen Lin. Enhancing semantic understanding in vision language models using meaning representation negative generation. In Fourth Workshop on Knowledge-infused Learning, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[84] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[85] Harman Singh, Pengchuan Zhang, Qifan Wang, Mengjiao Wang, Wenhan Xiong, Jingfei Du, and Yu Chen. Coarse-to-fine contrastive learning in image-text-graph space for improved vision-language compositionality. arXiv preprint arXiv:2305.13812, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[86] Jaisidh Singh, Ishaan Shrivastava, Mayank Vatsa, Richa Singh, and Aparna Bharati. Learn\" no\" to say\" yes\" better: Improving vision-language models via negations. arXiv preprint arXiv:2403.20312, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[87] Tianyi Tang, Yushuo Chen, Yifan Du, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen. Learning to imagine: Visually-" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "augmented natural language generation. arXiv preprint arXiv:2305.16944, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[88] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[89] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[90] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.904, + 0.417 + ], + "angle": 0, + "content": "[91] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.42, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[92] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.519, + 0.904, + 0.572 + ], + "angle": 0, + "content": "[93] Allan Tucker, Zhenchen Wang, Ylenia Rotalinti, and Puja Myles. Generating high-fidelity synthetic patient data for assessing machine learning healthcare software. NPJ digital medicine, 3(1):1-13, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.576, + 0.904, + 0.643 + ], + "angle": 0, + "content": "[94] Dani Valevski, Matan Kalman, Eyal Molad, Eyal Segalis, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning a diffusion model on a single image. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.647, + 0.904, + 0.701 + ], + "angle": 0, + "content": "[95] Gül Varol, Ivan Laptev, Cordelia Schmid, and Andrew Zisserman. Synthetic humans for action recognition from unseen viewpoints. International Journal of Computer Vision, 129(7):2264-2287, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[96] Fei Wang, Liang Ding, Jun Rao, Ye Liu, Li Shen, and Changxing Ding. Can linguistic knowledge improve multimodal alignment in vision-language pretraining? arXiv preprint arXiv:2308.12898, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.761, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[97] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[98] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24860" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[99] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. arXiv preprint arXiv:2004.11546, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[100] Yue Yang, Wenlin Yao, Hongming Zhang, Xiaoyang Wang, Dong Yu, and Jianshu Chen. Z-lavi: Zero-shot language solver fueled by visual imagination. arXiv preprint arXiv:2210.12261, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[101] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[102] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[103] Nir Yellinek, Leonid Karlinsky, and Raja Giryes. 3vl: using trees to teach vision & language models compositional concepts. arXiv preprint arXiv:2312.17345, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[104] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[105] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[106] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[107] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[108] Jianrui Zhang, Mu Cai, Tengyang Xie, and Yong Jae Lee. Countercurate: Enhancing physical and semantic visio-linguistic compositional reasoning via counterfactual examples. arXiv preprint arXiv:2402.13254, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[109] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic fine-grained understanding. arXiv preprint arXiv:2306.08832, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[110] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[111] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[112] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[113] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[114] Chenhao Zheng, Jieyu Zhang, Aniruddha Kembhavi, and Ranjay Krishna. Iterated learning improves compositionality in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13785-13795, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[115] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[116] Wanrong Zhu, An Yan, Yujie Lu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Visualize before you write: Imagination-guided open-ended text generation. arXiv preprint arXiv:2210.03765, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24861" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_origin.pdf b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3daa5da47cba380f7bcc2a28bac3c39b63fb1df7 --- /dev/null +++ b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/653d9e05-cace-4615-bcbc-a402fd9e9c3e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e37525e6aaaddf9de74fcce642bdf9cb3ae484d8be55ab571f9208a5eb61988d +size 791720 diff --git a/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/full.md b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..651e59cae12b50c6ceeba4e62e7f840ef091523c --- /dev/null +++ b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/full.md @@ -0,0 +1,388 @@ +# Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data + +Haoxin Li and Boyang Li + +Nanyang Technological University + +{haoxin003,boyang.li}@ntu.edu.sg + +# Abstract + +Paired image-text data with subtle variations in-between (e.g., people holding surfboards vs. people holding shovels) hold the promise of producing Vision-Language Models with proper compositional understanding. Synthesizing such training data from generative models is a highly coveted prize due to the reduced cost of data collection. However, synthesizing training images for compositional learning presents three challenges: (1) efficiency in generating large quantities of images, (2) text alignment between the generated image and the caption in the exact place of the subtle change, and (3) image fidelity in ensuring sufficient similarity with the original real images in all other places. We propose SPARCL (Synthetic Perturbations for Advancing Robust Compositional Learning), which integrates image feature injection into a fast text-to-image generative model, followed by an image style transfer step, to meet the three challenges. Further, to cope with any residual issues of text alignment, we propose an adaptive margin loss to filter out potentially incorrect synthetic samples and focus the learning on informative hard samples. Evaluation on four compositional understanding benchmarks demonstrates that SPARCL significantly improves the compositionality of CLIP, boosting the average accuracy of the CLIP base model by over $8\%$ across all benchmarks and outperforming state-of-the-art methods by $2\%$ on three benchmarks. + +# 1. Introduction + +Current Vision-Language Models (VLMs) still face limitation in accurately interpreting compositional relationships between objects and attributes, as demonstrated by numerous evaluations [32, 62, 88, 106, 113]. This limitation primarily stems from the absence of subtle variations in the training data [42] (e.g., the subtle variations between the two captions in Figure 1 (a)). As a result, it becomes possible to maximize empirical image-caption alignment us + +Source caption: Two people carrying surfboards on a beach. + +Target caption: Two people carrying shovels on a beach. + +![](images/f8364424f094b3ea50439a4f9064f100a2481797dd7a8b5cd896d37e2ebd8d57.jpg) +Source image + +![](images/8dc5e82fe8dfcae27e4391674bb6a2c65ccbc04cc89d6d54e04638c132aac93b.jpg) +Edited image + +![](images/782b2f58781d01e515a527e3f48d9ef639c6c2d869b49b7e80e5ccace4429e55.jpg) +Generated image + +(a) Difficulty in accurately creating precise variations. + +![](images/179fbc60fa9f1bcd7a064bc872ed467a961e3dd099e5b33c20c917d1e3dee137.jpg) + ++Real positive caption: A man is stirring a silver pot filled with food. + +- Synthetic negative caption: A man is stirring a copper pot filled with food. + +![](images/3273596626e139739b29826575bfa9910cdcb4118cc8bb36b163d658d3ad7c47.jpg) + ++Real positive caption: Random people sitting in a public transport bus. + +- Synthetic negative caption: Random animals riding a purple elephant. + +![](images/b9bf9a3ebee93449cdb354a66b8a537523b9f487963f1d0a693caa97c09dc5b9.jpg) +Easy Sample + +![](images/254e148fec8e8f637db25798d94eddb4b471ba89022de0918338303651080fa1.jpg) + ++Real positive caption: A double decker bus parked at a station. + ++ Synthetic positive caption: A double-decker bus parked at a bus station with a few passengers getting on. + +![](images/f5f758e2e76f32567ea946c823a4658c96d37f424f9b23f6b426fc32568da5a6.jpg) +Wrong Generation + +(b) Inconsistency in cross-modal alignment quality of synthetic samples. + +Figure 1. Challenges in generating and training on synthetic data: (a) When generating an image with subtle variations based on a real image and a target caption specifying the variations, an image editing model [6] struggles with text alignment (middle), while an image generation model [75] fails to maintain image fidelity (right). (b) Synthetic positive and negative image_caption pairs show different levels of alignment quality. The subtle variations in the synthetic negative caption (left) make it difficult to distinguish from the positive; the over-modified negative caption (middle) is easy to distinguish; and the hallucinated content in the synthetic positive caption (right) results in an incorrect positive. + +ing shortcut features [23] rather than genuinely learning nuanced distinctions. While collecting training samples with subtle variations could enhance compositionality, this approach is time-consuming and labor-intensive, rendering it impractical at scale. + +Advances in generative models [7, 12, 20, 43, 61, 75, + +78, 91, 105, 110] now facilitate the synthetic generation of training samples with subtle variations between paired samples [11, 18, 49, 68, 71, 79, 83, 86, 106, 108, 109]. By starting with real image-caption pairs, generative models can create subtle edits to both captions and images, providing valuable training data with minimal manual effort. These generated variations enable VLMs to enhance their compositionality by learning from the nuanced differences. + +However, generating and training on synthetic data presents two key challenges. The first is the difficulty of efficiently and accurately creating precise variations in synthetic images. To generate large-scale training data with precise variations, the image generation process must meet three criteria: efficiency in producing large quantities of images, text alignment of the generated image to the corresponding caption, which is subtly changed from the original caption, and image fidelity in ensuring the generated image otherwise closely matches their real counterpart. Nevertheless, current image generation methods struggle to meet all three criteria simultaneously. Per-sample optimization methods [44, 82, 94, 112] lack efficiency, while zero-shot image editing methods [6, 22, 63] and text-to-image (T2I) models [75, 78] often fail to achieve proper text alignment and image fidelity, respectively, as illustrated in Figure 1 (a). The inaccurate synthetic variations produced by these models can mislead the learning of VLMs. To alleviate this issue, we take a fast T2I model, which excels in efficiency and text alignment, and inject real image features into the text prompt features in order to enhance image fidelity. By combining this approach with AdaIN [35], we manage to substantially improve the image generation process per the three criteria. + +Still, the generated data do not perfectly meet the text alignment criterion, and we propose to deal with the remaining problems with an innovative approach to model training. As Figure 1 (b) shows, the similarity between synthetic positive and negative pairs may vary, resulting in a mixture of hard text-image pairs, easy pairs, and incorrect pairs. Instead of treating all positive and negative samples uniformly, we propose a novel loss that differentiates between positive, hard negative, and easy negative samples. Further, we propose an adaptive margin that helps to filter out potentially incorrect generations and focus the learning on informative hard samples. + +In summary, we propose SPARCL, which integrates image feature injection into a fast T2I model to improve the quality of synthetic variations in images and employs an adaptive margin loss to leverage the varying alignment quality in synthetic samples for training VLMs. Evaluations on four compositional understanding benchmarks demonstrate that SPARCL significantly enhances VLM compositionality, improving CLIP by over $5\%$ on VL-CheckList [113] and $7\%$ on SugarCrepe [32], while surpassing state-of-the + +art methods by $1\%$ and $2\%$ on the two benchmarks. The main contributions of this paper are as follows: (1) we propose image feature injection to enhance the quality of synthetic variations in images, which provide valuable training data to improve the compositionality of VLMs; (2) we introduce an adaptive margin loss to leverage varying levels of cross-modal alignment in synthetic samples to effectively differentiate positive and negative samples; (3) experimental results validate that SPARCL significantly improves the compositional understanding capabilities of CLIP models. + +# 2. Related Work + +# 2.1. Limitations in Compositionality of VLMs + +While VLMs excel in many multi-modal tasks [40, 52, 53, 74, 84, 101, 115], they still struggle with compositional understanding—the ability to interpret novel combinations of known visual and textual components. Benchmarks like What'sUp [42] reveal difficulties in understanding spatial relationships, SPEC [70] highlights issues with object size, position, and count, and ARO [106] uncovers limitations in understanding attributes, relations, and word order. Winoground [88], SNARE [96], and VL-CheckList [113] also expose these shortcomings. SugarCrepe [32] addresses hackable biases in prior benchmarks, where text-only models achieve artificially high performance, by introducing fluent and meaningful hard negatives. Their findings suggest that previous benchmarks overestimated compositional understanding. Building on this, SugarCrepe++ [19] further introduces semantically equivalent but lexically varied captions as hard positives and shows the difficulties of VLMs in distinguishing between lexical variations. + +# 2.2. Improving Compositionality of VLMs + +Prior approaches to improving the compositionality of VLMs can be broadly classified into the following categories: (1) Leveraging detailed image captions: Detailed captions from dense captioning models [17, 54], simulation platforms [9], and video annotations [45] are collected to train VLMs. However, these samples often lack pairs with subtle variations, limiting their contribution to compositionality. (2) Distilling from pretrained models: SDS-CLIP [3], SF-CLIP [80] and IL-CLIP [114] distill knowledge from pretrained image generation models and visual-language foundation models. IL-CLIP [114] refines representations through iterative learning with pretrained vision and language agents. However, pretrained models also face limitations in compositionality [34, 90]. (3) Incorporating structural knowledge: MosaiCLIP [85] and Structure-CLIP [36] incorporate scene graph knowledge in text features. CLIP-SGVL [28] and 3VL [103] train VLMs to predict scene graphs. [66] utilizes scene graphs as prompts to elicit compositional knowledge from VLMs without fur + +other training. However, these methods rely on models trained with expensive dense structure annotations (e.g., scene graphs). (4) Utilizing synthetic negative samples: Rule-based tools [30] or large language models (LLMs) [15, 73] are used to generate negative captions by editing real captions [10, 11, 18, 68, 83, 86, 106, 109]. Image generation models [75] are also used to create or edit images for training [49, 71, 79, 108]. Despite their utility, efficiently generating large amounts of images with precise variations remains challenging due to the inherent limitations of generative models. (5) Applying fine-grained alignment constraints: MCD [46] enforces multi-scale alignment across images with varying augmentations and the corresponding text captions. SPARC [5] learns local alignment by associating each text token with a group of local image patches. CE-CLIP [109] applies intra-modal contrastive loss and cross-modal ranking loss to improve alignment. However, uniform supervision applied to synthetic samples with varying alignment quality limits their effectiveness. In this paper, we propose SPARCL to address two key challenges in learning from synthetic data: the difficulty in accurately creating precise variations and the inconsistency in cross-modal alignment quality in synthetic data. + +# 2.3. Training with Synthetic Data + +Synthetic data for training machine learning models have been studied in various fields [14, 26, 48, 64, 67, 76, 77, 93, 99]. Synthetic data generated through simulations and graphics engines supports a wide range of tasks [16, 72, 95]. However, these synthetic datasets often diverge significantly from real-world data. Recent advances in generative models [7, 12, 43, 61, 75, 78, 91, 105, 110] have made it possible to synthesize data that more closely resembles real-world scenarios. Synthetic data are widely used in both language tasks [21, 31, 60, 64, 65, 87, 97, 98, 100, 116] and vision tasks [1, 2, 4, 25, 29, 39, 57, 81, 89, 92, 111]. Synthetic data are often noisy, necessitating noise-resistant training methods [13, 37, 41, 50, 51, 102], especially those designed for contrastive learning [38, 56]. In this paper, we generate multimodal samples with subtle variations and filter out potentially incorrect ones during training to improve the compositionality of VLMs. + +# 3. SPARCL + +To improve the compositional understanding abilities of VLMs, we propose Synthetic Perturbations for Advancing Robust Compositional Learning, or SPARCL, which generates multimodal samples with subtle variations from real samples and trains VLMs to learn nuanced differences through synthetic data. In the generation phase, SPARCL creates positive and negative captions with slight variations from real captions using an LLM, then generates images based on the real image and these modified captions through + +a fast T2I model. To enhance the quality of subtle variations in synthetic images, we introduce image feature injection, which integrates real image features into the text prompt features of the T2I model to improve fidelity to the real image, as detailed in Sec. 3.2. In the training phase, SPARCL employs an adaptive margin loss that leverages varying levels of multimodal alignment in the synthetic samples to effectively learn informative nuanced distinctions, as described in Sec. 3.3. The framework of SPARCL is shown in Figure 2. + +# 3.1. Generating Negative and Positive Captions + +Captions with subtle variations are crucial for learning compositional knowledge, as shown by previous work that generates captions by randomly swapping or replacing nouns and adjectives [106, 109]. However, manually designed generation rules often introduce nonsensical or grammatically incorrect artifacts, creating shortcut features [23] that obstruct true compositional understanding [32]. To address this issue, we use an LLM to generate natural synthetic captions. To further mitigate the impact of generative artifacts, we generate both negative and positive captions, ensuring that VLMs cannot easily differentiate negative captions from positive ones based solely on artifacts. We denote the $i^{th}$ real image-caption pair in the training set as $(\boldsymbol{I}_i^r,\boldsymbol{T}_i^r)$ . Given a real caption $\boldsymbol{T}_i^r$ , we prompt the LLM to generate a synthetic negative caption $\boldsymbol{T}_i^{sn}$ and a synthetic positive caption $\boldsymbol{T}_i^{sp}$ , as specified by the prompts in Figure A1 in the Appendix. + +# 3.2. Generating Images via Image Feature Injection + +Synthetic images that exhibit subtle variations from real images while aligning with the captions are valuable but challenging to generate, as discussed in Sec. 1. Although previous works [49, 71, 79, 108] have utilized object segmentation or filtered dissimilar generations to improve fidelity to real images, they struggle with manipulating relationships or lack efficiency. We aim to enhance fidelity by injecting image features into a fast T2I model [61], which already achieves high efficiency and text alignment. + +Image Feature Injection. The images generated by T2I models lack image fidelity to real images, as no real image information is input to the models. To enhance the fidelity of synthetic images, we inject real image features into the text prompt features, enabling the model to incorporate information from the real images. + +In T2I models, content and style are separated in the semantic and padding embeddings. The semantic embeddings (before the embedding of the [EOS] token) usually capture most of the image content in the text prompts, while the padding embeddings (after the [EOS] token) usually represent the image style [104]. Therefore, we can inject real image features into the padding embeddings to guide + +![](images/497a83f463f5bee0abcf4b37147e88b27e9e7c59b43f1c10da506ae2c143c20a.jpg) +(a) Generating synthetic captions and images + +![](images/3847209be1ab50fffa2408ebd0dc56d68078672bc9d1c6f9837fba18b4a02182.jpg) + +![](images/f8ebc6cae65f82dcfdfdafda7c6c32aa8337d353637752417e23f6a58cfe5ca5.jpg) +(b) Training with real and synthetic samples + +![](images/0ec229f2694a8b91cd1a306779aa55a2ea6ef0916c20f5d7649715a48dfb61a1.jpg) +Adaptive Margin Loss +Figure 2. An overview of SPARCL. (a) Starting with a real image-caption pair, we generate synthetic positive and negative pairs with subtle variations using an LLM and a fast T2I model. To improve the quality of subtle variations in synthetic images, we introduce image feature injection to reduce unintended variations from a standard T2I model (see Sec. 3.2). (b) We train the VLM using both real and synthetic samples. In addition to a sigmoid loss for distinguishing positive and negative pairs, we apply an adaptive margin loss that leverages varying alignment levels across training samples to learn informative nuanced distinctions (see Sec. 3.3). + +![](images/c88d2ee2497dc57fe24ef84887af0bf723486a9c5ae33532b4e969e775179112.jpg) + +the model in generating images with a similar style to real images, without affecting the alignment with the captions, thanks to the decoupling of content and style in the prompt embeddings. + +We extract features for a real image $\pmb{I}_i^r$ and the corresponding synthetic caption $\pmb{T}_i^s$ ( $s \in \{sn, sp\}$ ) using two aligned feature encoders (e.g., CLIP image and text encoders). The image embedding $\pmb{f}_i^r$ is the encoder output at the position of the [CLS] token. The text embedding is the text encoder output $e_i^s = \langle e_{i,j}^s \rangle_{j=1}^L$ , where $L$ is the maximum sequence length. Let the index of the [EOS] token in $e_i^s$ be $k_i^s$ , $k_i^s \leq L$ . That is, the padding token [PAD] is used as input to the text encoder at positions between $k_i^s$ and $L$ . The key step of SPARCL is to replace the text embeddings at those positions with the image embedding, producing a new sequence of text embeddings, $\hat{e}_i^s$ , + +$$ +\hat {e} _ {i} ^ {s} = \left\langle e _ {i, 1} ^ {s}, e _ {i, 2} ^ {s}, \dots , e _ {i, k _ {i}} ^ {s}, \underbrace {\boldsymbol {f} _ {i} ^ {r} , \dots , \boldsymbol {f} _ {i} ^ {r}} _ {L - k _ {i} ^ {s} \text {t i m e s}} \right\rangle . \tag {1} +$$ + +The embeddings $\hat{e}_i^s$ is then used as input to the T2I model to generate a synthetic image $\tilde{I}_i^s$ . This process, depicted in Figure 2, reduces unintended variations in synthetic images generated by the standard T2I model and enhances image fidelity to real images, thereby improving the quality of + +synthetic subtle variations. + +Style Transfer. To further reduce the domain gap between synthetic and real images, we use AdaIN [35] to transfer the style of the synthetic image to that of the real image. We use a pretrained AdaIN encoder to extract content features from $\tilde{I}_i^s$ and style features from $I_i^r$ . We normalize the content features using instance normalization and then scale and shift them with the mean and variance of the style features. The transformed features are fed into a pretrained AdaIN decoder to generate $I_i^s$ , which is subsequently used for model training. + +# 3.3. Training with Real and Synthetic Samples + +After generating captions and images, each real image-caption pair $(I_i^r, T_i^r)$ is extended to $(I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})$ by adding one synthetic negative and one synthetic positive pair. We then train the VLM using these extended samples. Given a batch of $n$ sample groups, $\{(I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})\}_{i=1}^n$ , we organize the images by concatenating all real, synthetic negative, and synthetic positive images into an image batch $I^B$ of $3n$ images, and similar all captions into a caption + +batch $\pmb{T}^B$ of $3n$ captions. + +We calculate the similarity between each image-caption pair in the batch. The similarity between the $i^{\mathrm{th}}$ image $\pmb{I}_i^B$ and the $j^{\mathrm{th}}$ caption $\pmb{T}_j^B$ is given by $S_{i,j} = c(\mathcal{E}_I(\pmb{I}_i^B),\mathcal{E}_T(\pmb{T}_j^B))$ , where $c(\cdot ,\cdot)$ denotes cosine similarity, and $\mathcal{E}_I$ and $\mathcal{E}_T$ represent the image and text encoders of the VLM, respectively. We then define the ground-truth alignment variable $M_{i,j}$ , which takes value 1 if the image matches with the caption, and -1 otherwise. Clearly, the only positive image-caption pairs are $(\pmb{I}_i^r,\pmb{T}_i^r),(I_i^{sn},T_i^{sn}),(I_i^{sp},T_i^{sp}),(I_i^r,T_i^{sp})$ and $(I_i^{sp},T_i^r)$ . To account for multiple positive associations of an image or a caption, we apply a sigmoid-based contrastive loss [8, 107] to encourage higher similarity for positive pairs and lower similarity for negative pairs, + +$$ +L _ {c o n} = - \frac {1}{3 n} \sum_ {i = 1} ^ {3 n} \sum_ {j = 1} ^ {3 n} \log \frac {1}{1 + \exp \left(- M _ {i , j} \left(S _ {i , j} / \tau + b\right)\right)}, \tag {2} +$$ + +where $\tau$ is the temperature parameter and $b$ is a bias term. + +Adaptive Margin Loss. Despite best efforts toward controlling generative models, synthetic samples may still have variable quality in terms of the alignment between text and imagery. That is, purported positive (resp. negative) pairs may not be semantically similar (resp. dissimilar). To account for variation in synthetic data quality, we propose to differentiate between real data and synthetic data and between easy negatives and hard negatives. Further, we propose an adaptive margin that filters out potential incorrect samples and prioritizes learning from hard samples. + +For each image in a batch, we define four sets of captions, a positive set $\mathbb{P}$ , a hard negative set $\mathbb{N}_h$ , an easy negative set $\mathbb{N}_e$ , and a real negative set $\mathbb{N}_r$ , which represent different levels of alignment within the batch. If the image is a real or synthetic positive image, + +- $\mathbb{P}$ contains the real positive captions $T_{i}^{r}$ and the synthetic positive captions $T_{i}^{sp}$ of the current, $i$ -th image. +- $\mathbb{N}_h$ contains the synthetic negative captions $T_i^{sn}$ of the current image. +- $\mathbb{N}_e = \{\pmb{T}_j^r,\pmb{T}_j^{sp},\pmb{T}_j^{sn}|i\neq j\}$ contains all real and synthetic captions from all other images. +- $\mathbb{N}_r = \{\pmb{T}_j^r | i \neq j\}$ contains all real captions belonging to the other images. + +If the image is a synthetic negative image, the sets $\mathbb{N}_e$ and $\mathbb{N}_r$ are unchanged but $\mathbb{P}$ and $\mathbb{N}_h$ differ: + +- $\mathbb{P}$ contains the synthetic negative captions $T_{i}^{sn}$ . +- $\mathbb{N}_h$ contains $T_i^r$ and $T_i^{sp}$ . + +We define the margin loss for the $i^{th}$ image as: + +$$ +\begin{array}{l} L _ {m a r, i} ^ {I} = \frac {1}{| \mathbb {P} | \cdot | \mathbb {N} _ {h} |} \sum_ {j _ {1} \in \mathbb {P}, j _ {2} \in \mathbb {N} _ {h}} \max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}) \\ + \frac {1}{\left| \mathbb {N} _ {h} \right| \cdot \left| \mathbb {N} _ {e} \right|} \sum_ {j _ {1} \in \mathbb {N} _ {h}, j _ {2} \in \mathbb {N} _ {e}} \max \left(0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}\right) \tag {3} \\ + \frac {\alpha}{| \mathbb {P} | \cdot | \mathbb {N} _ {r} |} \sum_ {j _ {1} \in \mathbb {P}, j _ {2} \in \mathbb {N} _ {r}} \max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}). \\ \end{array} +$$ + +This loss encourages positive pairs to have higher similarity scores than hard negative pairs, which should have higher similarity than easy negative pairs. Additionally, a weight $\alpha > 1$ is applied to comparisons between positive pairs and real negative pairs to emphasize these comparisons, as real samples are generally correct, making comparisons involving them more reliable. + +Lastly, we propose an adaptive margin in the above loss. Let $d = S_{i,j_1} - S_{i,j_2}$ denote the difference between two similarity scores in Eq. (3). The adaptive margin $m$ is then computed as follows: + +$$ +m = \left\{ \begin{array}{l l} d, & d < \beta \\ (\frac {m _ {0} - d}{m _ {0} - \beta} \gamma + 1) m _ {0}, & \beta \leq d \leq m _ {0} \\ m _ {0} & d > m _ {0} \end{array} \right. \tag {4} +$$ + +where $m_0$ is a base margin, $\beta < 0$ is a cutoff threshold, $\gamma$ is a scaling factor. The adaptive margin is designed with the following rationale: when $d < \beta$ , the purported positive pair is much less similar than the purported negative pair, suggesting the presence of incorrect or mislabeled samples, so we zero out the loss by setting the margin to $d$ . For $\beta \leq d \leq m_0$ , the margin is scaled, with smaller differences receiving larger margins to emphasize learning from harder samples. When $d > m_0$ , the samples are already well-separated with the margin $m_0$ , so the margin is capped at $m_0$ , again leading to zero loss. We visualize both the adaptive margin and the corresponding loss as functions of $d$ in Figure 2 (b). + +The loss for all images is given by $L_{mar}^{I} = \frac{1}{3n}\sum_{i}L_{mar,i}^{I}$ . Analogously, we compute $L_{mar}^{T}$ for all captions. The total adaptive margin loss is $L_{mar} = L_{mar}^{I} + L_{mar}^{T}$ . Finally, the overall training loss is a weighted combination of the contrastive loss and the adaptive margin loss, with weight $\lambda$ : + +$$ +L = L _ {c o n} + \lambda L _ {m a r}. \tag {5} +$$ + +# 4. Experiments + +# 4.1. Datasets + +Training. We use the COCO-2014 dataset [55] as the training data source. The training set consists of 82,783 images, each paired with five captions. For each real image-caption pair, we generate one positive and one negative synthetic pair. In line with previous approaches [18, 106, 109], we train the VLMs using both the original COCO-2014 training data and the synthetic samples. + +Evaluation. We evaluate our model on four vision-language compositional understanding benchmarks: (1) ARO [106], which consists of 23,937 cases for relation understanding and 28,748 for attribute understanding. We exclude the subsets for order understanding, as they contain significant nonsensical and non-fluent artifacts [32]. + +Table 1. Comparison of accuracy (%) between SPARCL and baselines on four benchmarks. "img." represents images, "cap." represents captions, "syn." represents synthetic data. + +
MethodTraining DataAROVL-CheckListSugarCrepeSugarCrepe++
Source# real img.# real cap.# syn. img.# syn. cap.
CLIP [74] (Zero-Shot)-----61.173.273.459.8
CLIP [74] (Finetune)COCO82K410K0064.172.879.962.3
SDS-CLIP [3]COCO82K410K0057.5---
[79]COCO0082K82K65.069.9--
AMR-NegCLIP [83]COCO100K100K0500K79.4-85.2-
NegCLIP [106]COCO100K100K0500K76.074.682.564.9
MosaiCLIP [85]COCO109K109K0981K80.376.8--
FSC-CLIP [68]COCO100K100K01.5M-77.285.1-
CE-CLIP [109]COCO82K410K02M79.776.385.2-
COMO [49]COCO113K567K567K567K-76.9--
SPARCL (our method)COCO82K410K820K820K77.279.287.166.1
SPEC [70]LAION20K20K20K20K70.1---
CounterCurate [108]Flickr30K30K150K150K--82.8-
[18]CC3M3M3M09M-75.3-55.3
CE-CLIP+ [109]COCO+CC3M3M3M015M80.479.387.5-
CLOVE [11]LAION-COCO>1B>1B0>1B73.2-85.1-
IL-CLIP [114]CC12M12M12M00--70.3-
SF-CLIP [80]YFCC15M15M15M00--71.2-
syn-CLIP [9]SyViC00>1M>1M69.274.8--
FiGCLIP [45]VidSitu20K videos0067.0-74.6-
+ +(2) VL-CheckList [113], a large-scale benchmark with over 100,000 samples, evaluates compositionality across subsets of objects, attributes, and relationships, which are further divided into various fine-grained categories. (3) SugarCrepe [32] includes 7,000 test cases across seven subsets. In the above three benchmarks, each test case containing one image, one positive caption, and one negative caption. (4) SugarCrepe++ [19] includes 4,757 test samples across five subsets, where each test case consists of one image, two positive captions, and one negative caption. All benchmarks involve classifying captions as positive or negative for the given images. We report the average accuracy across all subsets of each benchmark and include the accuracy for each subset in Appendix A3. + +# 4.2. Implementation Details + +Models. We use ViT-B/32 and ViT-L/14 architectures from OpenAI's CLIP model [74] as our base models, initialized with pretrained checkpoints. Following syn-CLIP [9], we integrate LoRA adapters [33] into both the image and text encoders of CLIP to improve training efficiency and mitigate knowledge forgetting. Only the LoRA adapters are fine-tuned during training. + +Training Setups. We use the AdamW optimizer [59] with a cosine learning rate schedule [58]. Training is conducted on two Tesla V100 GPUs, with a batch size of 128 sample groups for ViT-B/32 and 16 for ViT-L/14. The base learning rate is set to 0.01 for a total batch size of 256, and scaled + +linearly [24] based on the actual batch size. Training is performed for 3,000 steps for ViT-B/32 and 15,000 steps for ViT-L/14, corresponding to fewer than 5 epochs in previous studies [18, 106, 109]. More details and hyperparameters settings are in Appendix A2. + +# 4.3. Main Results + +We compare SPARCL with several baseline methods: (1) methods that utilize synthetic samples for training [49, 68, 70, 79, 83, 106, 108, 109], (2) methods that distill knowledge from pretrained models [3, 11, 80, 114], (3) methods that incorporate knowledge from scene graphs [18, 85], and (4) methods that leverage detailed image captions [9, 45]. Both SPARCL and the baseline methods use ViT-B/32 as the base model. The results are presented in Table 1. + +We observe that SPARCL achieves the best performance on VL-CheckList, SugarCrepe, and SugarCrepe++ compared to baselines trained on the same data source, namely COCO. Specifically, SPARCL surpasses the strongest baseline by $1.4\%$ on VL-CheckList and $2.5\%$ on SugarCrepe, underscoring its effectiveness in enhancing compositional understanding. However, SPARCL performs worse on ARO than some baselines. We hypothesize that this may be due to nonsensical or grammatically incorrect artifacts in the ARO test samples [32], which could favor methods that utilize rule-based synthetic training samples (e.g., [85]) containing similar artifacts. Notably, even with only $50\%$ of the training data, SPARCL outperforms most base + +Table 2. Ablated performance (\%) of SPARCL. "SynCap" refers to synthetic captions, "SynImg" refers to synthetic images, "FeatInj" denotes image feature injection, and "CompSet" indicates the comparison sets used in Eq. (3). + +
ModelVariantSynCapSynImgFeatInjAdaINAdaptive Margin LossAROVL-CheckListSugarCrepeSugarCrepe++Average
CompSetMargin
ViT-B/32#1XXXXX60.4972.6179.3664.8569.33
#2✓XXXX71.7773.5286.3564.3273.99
#3X✓XXX62.6271.7079.9764.8469.79
#4✓✓XXX71.8675.5485.4365.2274.51
#5✓✓✓XX73.4076.5685.5465.7875.32
#6✓✓X✓X73.7975.7285.7964.4974.95
#7✓✓✓✓X74.1276.3585.4066.4475.58
#8✓✓✓✓AllFixed76.7978.5987.0865.4276.97
#9✓✓✓✓AllAdaptive77.1579.1687.1166.1277.38
#10✓✓✓✓AllAdaptive Inversed76.6779.2686.6865.3076.97
#11✓✓✓✓Only (P, Nh)Adaptive77.4880.4886.1564.7077.20
ViT-L/14#1XXXXX59.8073.2681.4964.9069.86
#2✓XXXX72.1675.4487.3864.4474.85
#3X✓XXX60.7072.2382.5866.6370.54
#4✓✓XXX75.2078.2987.2964.6176.35
#5✓✓✓XX75.1178.5887.5464.5776.45
#6✓✓X✓X74.8879.1087.5064.6276.52
#7✓✓✓✓X74.9380.0487.7965.3077.01
#8✓✓✓✓AllFixed75.2180.7588.0266.4177.60
#9✓✓✓✓AllAdaptive75.8380.8188.2366.8377.93
#10✓✓✓✓AllAdaptive Inversed75.1580.7187.9366.6177.60
#11✓✓✓✓Only (P, Nh)Adaptive76.2680.5787.4266.2877.63
+ +line methods (see Table 3 for SPARCL performance with reduced training data). Compared to baselines using additional data sources, SPARCL surpasses or matches their performance. For example, SPARCL outperforms methods in [11, 18] despite their use of more training data. SPARCL performs comparably to CE-CLIP+ [109] on VL-CheckList and SugarCrepe, despite CE-CLIP+ leveraging CC3M as an additional data source and utilizing significantly more samples. + +# 4.4. Ablation Study and Analysis + +We perform ablation studies to assess the impact of each component and design choice in SPARCL. The results, presented in Table 2, show the overall performance. + +Synthetic captions vs. synthetic images. To analyze the impact of synthetic captions and images, we compare Variant #1 (using only real samples) with Variant #2 (using real samples and synthetic captions) and #3 (using real samples and synthetic images). Variant #2 shows a significant improvement over Variant #1, with the average accuracy improving from $69.33\%$ to $73.99\%$ for ViT-B/32. However, Variant #3 obtains only a marginal improvements of about $0.5\%$ . These results suggest that synthetic captions substantially enhance compositional understanding, while synthetic images alone provide limited benefit, which aligns with findings in [69]. We hypothesize that generative artifacts in synthetic images are more pronounced than in captions, which negatively affects the effective learning of nu + +anced distinctions by VLMs. When both synthetic captions and images are used (Variant #4), performance improves further, indicating a synergistic effect. + +Image feature injection vs. AdaIN. To analyze the impact of image feature injection and AdaIN, we compare Variant #4 (without image feature injection or AdaIN) with Variant #5 (with image feature injection only) and #6 (with AdaIN only). Using ViT-B/32, we observe that Variant #5 outperforms Variant #4 across all four benchmarks, with an average gap of $0.8\%$ . While Variant #6 shows a smaller improvement of $0.4\%$ , it outperforms Variant #5 on ARO and SugarCrepe. These results suggest that image feature injection is more effective for improving compositionality than AdaIN, although the two methods are somewhat complementary. Through the improvements obtained using ViT-L/14 are a little bit different, both techniques improve performance over Variant #4. Combining both methods in Variant #7 leads to further improvements, with a more than $1\%$ increase in average accuracy for ViT-B/32 and a $0.65\%$ increase for ViT-L/14 over Variant #4. This highlights the importance of reducing unintended changes in synthetic images, as such changes could cause the model to rely on them for distinguishing positive and negative samples rather than the semantic differences specified in the corresponding captions. Examples of synthetic images in Figure A3 and A4 in the Appendix show how image feature injection helps mitigate these unintended changes. + +Effects of the margin loss and adaptive margin strate- + +Table 3. Performance (\%) of SPARCL using different subsets of training samples. "Neg." represents synthetic negative samples, "Pos." represents synthetic positive samples, and "Prop." represents the proportion of training samples. + +
VariantNeg.Pos.Prop.AROVL-CheckListSugarCrepeSugarCrepe++Average
#7✓×100%72.1274.6586.4157.9972.79
#7✓✓100%74.1276.3585.4066.4475.58
#9✓✓20%75.9077.5685.7864.4475.92
#9✓✓50%77.9079.3886.7164.2177.05
#9✓✓100%77.1579.1687.1166.1277.38
+ +gies. We compare three variants with distinct margin approaches: (1) Fixed (Variant #8): a fixed margin without adaptive adjustments; (2) Adaptive (Variant #9): the margin is calculated according to Eq. (4); and (3) Adaptive Inversed (Variant #10): a larger margin is applied to samples with higher similarity differences, representing an inverse version of Eq. (4) that prioritizes learning from easier samples. Comparing Variant #8 with Variant #7, we observe that #8 improves by over $1\%$ for ViT-B/32 and $0.6\%$ for ViT-L/14 in average accuracy, suggesting that margin loss effectively aids in compositional understanding. Between Variants #8, #9, and #10, we find that Variant #9 outperforms Variant #8 across all four benchmarks, with an average boost of over $0.3\%$ for both ViT-B/32 and ViT-L/14. In contrast, Variant #10 results in a performance decline on three benchmarks and no improvement in average accuracy. These findings highlight the superior effectiveness of adaptive margins. + +Is applying adaptive margin loss only to hard samples sufficient? To explore whether applying margin loss solely to hard samples is adequate, we construct Variant #11, which uses only the positive set $\mathbb{P}$ and the hard negative set $\mathbb{N}_h$ for margin loss calculation in Eq. (3), which is similar to learning strategies in [49, 109]. When comparing Variant #11 with Variant #8 and #9, we observe that Variant #11 performs better on ARO and VL-CheckList, but worse on SugarCrepe and SugarCrepe++. We hypothesize that by focusing exclusively on the positive and hard negative sets, the model may learn nonsensical artifacts present in the hard negatives. These artifacts improve performance on ARO and VL-CheckList, where the test samples exhibit similar patterns, but hinder performance on SugarCrepe and SugarCrepe++, which are designed to avoid such patterns [32]. In contrast, Variant #9, which incorporates all sets in Eq. (3) for margin loss calculation, achieves better results on SugarCrepe and SugarCrepe++ with only a minor drop in performance on ARO and VL-CheckList. This suggests that easy negative samples serve as regularization, preventing overfitting to the artifacts in hard negative samples. + +Effects of synthetic positive samples. We compare the performance of training with only synthetic negative samples versus using both synthetic negative and positive samples + +in Table 3. Based on Variant #7, we train ViT-B/32 with these two combinations of training data. We observe nearly $3\%$ decrease in average accuracy when using only synthetic negative samples, compared to using both synthetic negative and positive samples. This decrease is particularly noticeable on SugarCrepe++, ARO, and VL-CheckList. These results underscore the importance of incorporating synthetic positive samples in training, as they provide essential information about variations that maintain semantic consistency, which is critical for compositional understanding. + +Influence of model size and training data size. To evaluate the impact of model size on compositional understanding, we compare the performance of ViT-B/32 and ViT-L/14 in Table 2. We find that ViT-L/14 only provides a modest improvement of about $0.5\%$ of average accuracy over ViT-B/32. This suggests that increasing model size does not significantly enhance compositional understanding. To investigate the effect of training data size, we train ViT-B/32 using $20\%$ and $50\%$ of randomly sampled training data and report the results in Table 3. We observe that performance improves as the proportion of training data increases, indicating that a larger training set helps the model better capture compositional knowledge. However, the performance gain diminishes when the training data size increases from $50\%$ to $100\%$ . Perhaps generating or selecting high-quality data would be more effective than simplistic data scaling. + +Additional experiments on hyperparameter analysis can be found in Appendix A3. + +# 5. Conclusion + +In this paper, we introduce SPARCL to enhance the compositional understanding capabilities of VLMs by generating and training with synthetic data. To tackle two key challenges in using synthetic data—namely, the difficulty of generating accurate variations and the inconsistency in cross-modal alignment quality—SPARCL integrates image feature injection into a T2I model to improve the quality of synthetic variations and introduces an adaptive margin loss to account for varying levels of cross-modal alignment for effectively learning nuanced distinctions. Experiments on four visual-language compositional understanding benchmarks demonstrate the effectiveness of SPARCL. + +# 6. Acknowledgments + +This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF-NRFF13-2021-0006), Singapore. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies. + +# References + +[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 3 +[2] Manel Baradad Jurjo, Jonas Wulff, Tongzhou Wang, Phillip Isola, and Antonio Torralba. Learning to see by looking at noise. Advances in Neural Information Processing Systems, 34:2556-2569, 2021. 3 +[3] Samyadeep Basu, Maziar Sanjabi, Daniela Massiceti, Shell Xu Hu, and Soheil Feizi. Augmenting clip with improved visio-linguistic reasoning. arXiv preprint arXiv:2307.09233, 2023. 2, 6 +[4] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This dataset does not exist: training models from generated images. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2020. 3 +[5] Ioana Bica, Anastasija Ilic, Matthias Bauer, Goker Erdoğan, Matko Bošnjak, Christos Kaplanis, Alexey A Gritsenko, Matthias Minderer, Charles Blundell, Razvan Pascanu, et al. Improving fine-grained understanding in image-text pre-training. arXiv preprint arXiv:2401.09865, 2024. 3 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3 +[8] Adrian Bulat, Yassine Ouali, and Georgios Tzimiropoulos. Fff: Fixing flawed foundations in contrastive pre-training results in very strong vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14172-14182, 2024. 5 +[9] Paola Cascante-Bonilla, Khaled Shehada, James Seale Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gul Varol, Aude Oliva, Vicente Ordonez, Rogerio Feris, et al. Going beyond nouns with vision & language models using synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20155-20165, 2023. 2, 6 +[10] Paola Cascante-Bonilla, Yu Hou, Yang Trista Cao, Hal Daumé III, and Rachel Rudinger. Natural language inference improves compositionality in vision-language models. arXiv preprint arXiv:2410.22315, 2024. 3 +[11] Santiago Castro, Amir Ziai, Avneesh Saluja, Zhuoning Yuan, and Rada Mihalcea. Clove: Encoding compositional language in contrastive vision-language models. arXiv preprint arXiv:2402.15021, 2024. 2, 3, 6, 7 +[12] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, + +William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3 +[13] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 3 +[14] Yabo Dan, Yong Zhao, Xiang Li, Shaobo Li, Ming Hu, and Jianjun Hu. Generative adversarial networks (gan) based efficient sampling of chemical composition space for inverse design of inorganic materials. npj Computational Materials, 6(1):84, 2020. 3 +[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[16] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3 +[17] Sivan Doveh, Assaf Arbelle, Sivan Harary, Amit Alfassy, Roei Herzig, Donghyun Kim, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, et al. Dense and aligned captions (dac) promote compositional reasoning in v1 models. arXiv preprint arXiv:2305.19595, 2023. 2 +[18] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2, 3, 5, 6, 7 +[19] Sri Harsha Dumpala, Aman Jaiswal, Chandramouli Sastry, Evangelos Milios, Sageev Oore, and Hassan Sajjad. Sugarcrepe++ dataset: Vision-language model sensitivity to semantic and lexical alterations. arXiv preprint arXiv:2406.11171, 2024. 2, 6 +[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1 +[21] Jiahui Gao, Renjie Pi, Lin Yong, Hang Xu, Jiacheng Ye, Zhiyong Wu, Weizhong Zhang, Xiaodan Liang, Zhenguo Li, and Lingpeng Kong. Self-guided noise-free data generation for efficient zero-shot learning. In International Conference on Learning Representations (ICLR 2023), 2023. 3 +[22] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. arXiv preprint arXiv:2403.14602, 2024. 2, 1 +[23] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural net + +works.Nature Machine Intelligence,2(11):665-673,2020. 1,3 +[24] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6 +[25] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 3 +[26] Xuanli He, Islam Nassar, Jamie Kiros, Gholamreza Haffari, and Mohammad Norouzi. Generate, annotate, and learn: Nlp with synthetic text. Transactions of the Association for Computational Linguistics, 10:826-842, 2022. 3 +[27] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 3 +[28] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arbelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2 +[29] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3 +[30] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1):411-420, 2017. 3 +[31] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language models with (almost) no human labor. arXiv preprint arXiv:2212.09689, 2022. 3 +[32] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 1, 2, 3, 5, 6, 8 +[33] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[34] Hang Hua, Yunlong Tang, Ziyun Zeng, Liangliang Cao, Zhengyuan Yang, Hangfeng He, Chenliang Xu, and Jiebo Luo. Mmcomposition: Revisiting the compositionality of pre-trained vision-language models. arXiv preprint arXiv:2410.09733, 2024. 2 +[35] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4 +[36] Yufeng Huang, Jiji Tang, Zhuo Chen, Rongsheng Zhang, Xinfeng Zhang, Weijie Chen, Zeng Zhao, Zhou Zhao, + +Tangjie Lv, Zhipeng Hu, et al. Structure-clip: Towards scene graph knowledge to enhance multi-modal structured representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2417-2425, 2024. 2 +[37] Zhizhong Huang, Junping Zhang, and Hongming Shan. Twin contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2023. 3 +[38] Sarah Ibrahimi, Arnaud Sors, Rafael Sampaio de Rezende, and Stéphane Clinchant. Learning with label noise for image retrieval by selecting interactions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2181-2190, 2022. 3 +[39] Ali Jahanian, Xavier Puig, Yonglong Tian, and Phillip Isola. Generative models as a data source for multiview representation learning. arXiv preprint arXiv:2106.05258, 2021. 3 +[40] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[41] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 3 +[42] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's" up" with vision-language models? investigating their struggle with spatial reasoning. arXiv preprint arXiv:2310.19785, 2023. 1, 2 +[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 3 +[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2 +[45] Zeeshan Khan, Makarand Tapaswi, et al. Figclip: Fine-grained clip adaptation via densely annotated videos. arXiv preprint arXiv:2401.07669, 2024. 2, 6 +[46] Bumsoo Kim, Yeonsik Jo, Jinhyung Kim, and Seunghwan Kim. Misalign, contrast then distill: Rethinking misalignments in language-image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2023. 3 +[47] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1 +[48] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 3 +[49] Chengen Lai, Shengli Song, Sitong Yan, and Guangneng Hu. Improving vision and language concepts understanding with multimodal counterfactual samples. In European + +Conference on Computer Vision, pages 174-191. Springer, 2024. 2, 3, 6, 8 +[50] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5051-5059, 2019. 3 +[51] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9485-9494, 2021. 3 +[52] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2 +[53] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2 +[54] Liunian Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. Advances in Neural Information Processing Systems, 36, 2024. 2 +[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[56] Chang Liu, Han Yu, Boyang Li, Zhiqi Shen, Zhanning Gao, Peiran Ren, Xuansong Xie, Lizhen Cui, and Chunyan Miao. Noise-resistant deep metric learning with ranking-based instance selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6811-6820, 2021. 3 +[57] Hao Liu, Tom Zahavy, Volodymyr Mnih, and Satinder Singh. Palm up: Playing in the latent manifold for unsupervised pretraining. Advances in Neural Information Processing Systems, 35:35880-35893, 2022. 3 +[58] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[60] Yujie Lu, Wanrong Zhu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Imagination-augmented natural language understanding. arXiv preprint arXiv:2204.08535, 2022. 3 +[61] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3 +[62] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 1 + +[63] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 1 +[64] Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. Generating training data with language models: Towards zero-shot language understanding. Advances in Neural Information Processing Systems, 35:462-477, 2022. 3 +[65] Yu Meng, Martin Michalski, Jiaxin Huang, Yu Zhang, Tarek Abdelzaher, and Jiawei Han. Tuning language models as training data generators for augmentation-enhanced few-shot learning. In International Conference on Machine Learning, pages 24457-24477. PMLR, 2023. 3 +[66] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076, 2023. 2 +[67] Sergey I Nikolenko. Synthetic data for deep learning. Springer, 2021. 3 +[68] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pre-trained vlms for improving vision-linguistic compositionality. arXiv preprint arXiv:2410.05210, 2024. 2, 3, 6 +[69] Maitreya Patel, Abhiram Kusumba, Sheng Cheng, Changhoon Kim, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Tripletclip: Improving compositional reasoning of clip via synthetic vision-language negatives. arXiv preprint arXiv:2411.02545, 2024. 7 +[70] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. arXiv preprint arXiv:2312.00081, 2023. 2, 6 +[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize diagnose and optimize: Towards fine-grained vision-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13279-13288, 2024. 2, 3 +[72] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.3 +[73] Alec Radford. Improving language understanding by generative pre-training. 2018. 3 +[74] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 1, 3 +[75] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3 +[76] Andrew Rosenberg, Yu Zhang, Bhuvana Ramabhadran, Ye Jia, Pedro Moreno, Yonghui Wu, and Zelin Wu. Speech recognition with augmented synthesized speech. In 2019 + +IEEE automatic speech recognition and understanding workshop (ASRU), pages 996-1002. IEEE, 2019. 3 +[77] Nick Rossenbach, Albert Zeyer, Ralf Schlüter, and Hermann Ney. Generating synthetic audio data for attention-based speech recognition systems. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7069-7073. IEEE, 2020. 3 +[78] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3 +[79] Ugur Sahin, Hang Li, Qadeer Khan, Daniel Cremers, and Volker Tresp. Enhancing multimodal compositional reasoning of visual language models with generative negative mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5563-5573, 2024. 2, 3, 6 +[80] Sepehr Sameni, Kushal Kafle, Hao Tan, and Simon Jenni. Building vision-language models on solid foundations with masked distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14216-14226, 2024. 2, 6 +[81] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3 +[82] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2 +[83] Ziyi Shou and Fangzhen Lin. Enhancing semantic understanding in vision language models using meaning representation negative generation. In Fourth Workshop on Knowledge-infused Learning, 2024. 2, 3, 6 +[84] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2 +[85] Harman Singh, Pengchuan Zhang, Qifan Wang, Mengjiao Wang, Wenhan Xiong, Jingfei Du, and Yu Chen. Coarse-to-fine contrastive learning in image-text-graph space for improved vision-language compositionality. arXiv preprint arXiv:2305.13812, 2023. 2, 6 +[86] Jaisidh Singh, Ishaan Shrivastava, Mayank Vatsa, Richa Singh, and Aparna Bharati. Learn" no" to say" yes" better: Improving vision-language models via negations. arXiv preprint arXiv:2403.20312, 2024. 2, 3 +[87] Tianyi Tang, Yushuo Chen, Yifan Du, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen. Learning to imagine: Visually- + +augmented natural language generation. arXiv preprint arXiv:2305.16944, 2023. 3 +[88] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2 +[89] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36, 2024. 3 +[90] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2 +[91] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3 +[92] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 3 +[93] Allan Tucker, Zhenchen Wang, Ylenia Rotalinti, and Puja Myles. Generating high-fidelity synthetic patient data for assessing machine learning healthcare software. NPJ digital medicine, 3(1):1-13, 2020. 3 +[94] Dani Valevski, Matan Kalman, Eyal Molad, Eyal Segalis, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning a diffusion model on a single image. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2 +[95] Gül Varol, Ivan Laptev, Cordelia Schmid, and Andrew Zisserman. Synthetic humans for action recognition from unseen viewpoints. International Journal of Computer Vision, 129(7):2264-2287, 2021. 3 +[96] Fei Wang, Liang Ding, Jun Rao, Ye Liu, Li Shen, and Changxing Ding. Can linguistic knowledge improve multimodal alignment in vision-language pretraining? arXiv preprint arXiv:2308.12898, 2023. 2 +[97] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022.3 +[98] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 3 + +[99] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. arXiv preprint arXiv:2004.11546, 2020. 3 +[100] Yue Yang, Wenlin Yao, Hongming Zhang, Xiaoyang Wang, Dong Yu, and Jianshu Chen. Z-lavi: Zero-shot language solver fueled by visual imagination. arXiv preprint arXiv:2210.12261, 2022. 3 +[101] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2 +[102] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 3 +[103] Nir Yellinek, Leonid Karlinsky, and Raja Giryes. 3vl: using trees to teach vision & language models compositional concepts. arXiv preprint arXiv:2312.17345, 2023. 2 +[104] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3 +[105] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3 +[106] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 5, 6 +[107] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 5 +[108] Jianrui Zhang, Mu Cai, Tengyang Xie, and Yong Jae Lee. Countercurate: Enhancing physical and semantic visio-linguistic compositional reasoning via counterfactual examples. arXiv preprint arXiv:2402.13254, 2024. 2, 3, 6 +[109] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic fine-grained understanding. arXiv preprint arXiv:2306.08832, 2023. 2, 3, 5, 6, 7, 8 +[110] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 3 +[111] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3 + +[112] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2 +[113] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6 +[114] Chenhao Zheng, Jieyu Zhang, Aniruddha Kembhavi, and Ranjay Krishna. Iterated learning improves compositionality in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13785-13795, 2024. 2, 6 +[115] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 +[116] Wanrong Zhu, An Yan, Yujie Lu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Visualize before you write: Imagination-guided open-ended text generation. arXiv preprint arXiv:2210.03765, 2022. 3 \ No newline at end of file diff --git a/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/images.zip b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..732d2d2b7906b04250d5738f467e5e35e34afa01 --- /dev/null +++ b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2f4b5f7186365db7c6ac3ad1f25924566f74b682d4cf4eb288a29996d8fac5e +size 450610 diff --git a/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/layout.json b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e846166921825e84a5543bcd91fda08583131c82 --- /dev/null +++ b/2025/Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data/layout.json @@ -0,0 +1,11574 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "type": "text", + "content": "Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 241, + 161, + 369, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 161, + 369, + 175 + ], + "spans": [ + { + "bbox": [ + 241, + 161, + 369, + 175 + ], + "type": "text", + "content": "Haoxin Li and Boyang Li" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 176, + 389, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 176, + 389, + 190 + ], + "spans": [ + { + "bbox": [ + 220, + 176, + 389, + 190 + ], + "type": "text", + "content": "Nanyang Technological University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 191, + 392, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 191, + 392, + 203 + ], + "spans": [ + { + "bbox": [ + 215, + 191, + 392, + 203 + ], + "type": "text", + "content": "{haoxin003,boyang.li}@ntu.edu.sg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 258, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 568 + ], + "type": "text", + "content": "Paired image-text data with subtle variations in-between (e.g., people holding surfboards vs. people holding shovels) hold the promise of producing Vision-Language Models with proper compositional understanding. Synthesizing such training data from generative models is a highly coveted prize due to the reduced cost of data collection. However, synthesizing training images for compositional learning presents three challenges: (1) efficiency in generating large quantities of images, (2) text alignment between the generated image and the caption in the exact place of the subtle change, and (3) image fidelity in ensuring sufficient similarity with the original real images in all other places. We propose SPARCL (Synthetic Perturbations for Advancing Robust Compositional Learning), which integrates image feature injection into a fast text-to-image generative model, followed by an image style transfer step, to meet the three challenges. Further, to cope with any residual issues of text alignment, we propose an adaptive margin loss to filter out potentially incorrect synthetic samples and focus the learning on informative hard samples. Evaluation on four compositional understanding benchmarks demonstrates that SPARCL significantly improves the compositionality of CLIP, boosting the average accuracy of the CLIP base model by over " + }, + { + "bbox": [ + 54, + 258, + 297, + 568 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 568 + ], + "type": "text", + "content": " across all benchmarks and outperforming state-of-the-art methods by " + }, + { + "bbox": [ + 54, + 258, + 297, + 568 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 568 + ], + "type": "text", + "content": " on three benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 596, + 135, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 596, + 135, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 596, + 135, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Current Vision-Language Models (VLMs) still face limitation in accurately interpreting compositional relationships between objects and attributes, as demonstrated by numerous evaluations [32, 62, 88, 106, 113]. This limitation primarily stems from the absence of subtle variations in the training data [42] (e.g., the subtle variations between the two captions in Figure 1 (a)). As a result, it becomes possible to maximize empirical image-caption alignment us" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 229, + 515, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 229, + 515, + 239 + ], + "spans": [ + { + "bbox": [ + 318, + 229, + 515, + 239 + ], + "type": "text", + "content": "Source caption: Two people carrying surfboards on a beach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 239, + 502, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 239, + 502, + 247 + ], + "spans": [ + { + "bbox": [ + 319, + 239, + 502, + 247 + ], + "type": "text", + "content": "Target caption: Two people carrying shovels on a beach." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 248, + 389, + 317 + ], + "blocks": [ + { + "bbox": [ + 319, + 248, + 389, + 317 + ], + "lines": [ + { + "bbox": [ + 319, + 248, + 389, + 317 + ], + "spans": [ + { + "bbox": [ + 319, + 248, + 389, + 317 + ], + "type": "image", + "image_path": "f8364424f094b3ea50439a4f9064f100a2481797dd7a8b5cd896d37e2ebd8d57.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 317, + 373, + 324 + ], + "lines": [ + { + "bbox": [ + 334, + 317, + 373, + 324 + ], + "spans": [ + { + "bbox": [ + 334, + 317, + 373, + 324 + ], + "type": "text", + "content": "Source image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 400, + 247, + 470, + 317 + ], + "blocks": [ + { + "bbox": [ + 400, + 247, + 470, + 317 + ], + "lines": [ + { + "bbox": [ + 400, + 247, + 470, + 317 + ], + "spans": [ + { + "bbox": [ + 400, + 247, + 470, + 317 + ], + "type": "image", + "image_path": "8dc5e82fe8dfcae27e4391674bb6a2c65ccbc04cc89d6d54e04638c132aac93b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 317, + 453, + 324 + ], + "lines": [ + { + "bbox": [ + 416, + 317, + 453, + 324 + ], + "spans": [ + { + "bbox": [ + 416, + 317, + 453, + 324 + ], + "type": "text", + "content": "Edited image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 481, + 247, + 551, + 317 + ], + "blocks": [ + { + "bbox": [ + 481, + 247, + 551, + 317 + ], + "lines": [ + { + "bbox": [ + 481, + 247, + 551, + 317 + ], + "spans": [ + { + "bbox": [ + 481, + 247, + 551, + 317 + ], + "type": "image", + "image_path": "782b2f58781d01e515a527e3f48d9ef639c6c2d869b49b7e80e5ccace4429e55.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 317, + 539, + 324 + ], + "lines": [ + { + "bbox": [ + 492, + 317, + 539, + 324 + ], + "spans": [ + { + "bbox": [ + 492, + 317, + 539, + 324 + ], + "type": "text", + "content": "Generated image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 326, + 522, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 326, + 522, + 336 + ], + "spans": [ + { + "bbox": [ + 347, + 326, + 522, + 336 + ], + "type": "text", + "content": "(a) Difficulty in accurately creating precise variations." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 318, + 343, + 388, + 413 + ], + "blocks": [ + { + "bbox": [ + 318, + 343, + 388, + 413 + ], + "lines": [ + { + "bbox": [ + 318, + 343, + 388, + 413 + ], + "spans": [ + { + "bbox": [ + 318, + 343, + 388, + 413 + ], + "type": "image", + "image_path": "179fbc60fa9f1bcd7a064bc872ed467a961e3dd099e5b33c20c917d1e3dee137.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 415, + 386, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 386, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 386, + 437 + ], + "type": "text", + "content": "+Real positive caption: A man is stirring a silver pot filled with food." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 437, + 386, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 437, + 386, + 465 + ], + "spans": [ + { + "bbox": [ + 318, + 437, + 386, + 465 + ], + "type": "text", + "content": "- Synthetic negative caption: A man is stirring a copper pot filled with food." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 399, + 342, + 470, + 413 + ], + "blocks": [ + { + "bbox": [ + 399, + 342, + 470, + 413 + ], + "lines": [ + { + "bbox": [ + 399, + 342, + 470, + 413 + ], + "spans": [ + { + "bbox": [ + 399, + 342, + 470, + 413 + ], + "type": "image", + "image_path": "3273596626e139739b29826575bfa9910cdcb4118cc8bb36b163d658d3ad7c47.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 399, + 415, + 467, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 415, + 467, + 437 + ], + "spans": [ + { + "bbox": [ + 399, + 415, + 467, + 437 + ], + "type": "text", + "content": "+Real positive caption: Random people sitting in a public transport bus." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 400, + 437, + 459, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 437, + 459, + 465 + ], + "spans": [ + { + "bbox": [ + 400, + 437, + 459, + 465 + ], + "type": "text", + "content": "- Synthetic negative caption: Random animals riding a purple elephant." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 399, + 473, + 414, + 485 + ], + "blocks": [ + { + "bbox": [ + 399, + 473, + 414, + 485 + ], + "lines": [ + { + "bbox": [ + 399, + 473, + 414, + 485 + ], + "spans": [ + { + "bbox": [ + 399, + 473, + 414, + 485 + ], + "type": "image", + "image_path": "b9bf9a3ebee93449cdb354a66b8a537523b9f487963f1d0a693caa97c09dc5b9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 475, + 459, + 483 + ], + "lines": [ + { + "bbox": [ + 421, + 475, + 459, + 483 + ], + "spans": [ + { + "bbox": [ + 421, + 475, + 459, + 483 + ], + "type": "text", + "content": "Easy Sample" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 481, + 342, + 552, + 413 + ], + "blocks": [ + { + "bbox": [ + 481, + 342, + 552, + 413 + ], + "lines": [ + { + "bbox": [ + 481, + 342, + 552, + 413 + ], + "spans": [ + { + "bbox": [ + 481, + 342, + 552, + 413 + ], + "type": "image", + "image_path": "254e148fec8e8f637db25798d94eddb4b471ba89022de0918338303651080fa1.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 480, + 415, + 548, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 480, + 415, + 548, + 437 + ], + "spans": [ + { + "bbox": [ + 480, + 415, + 548, + 437 + ], + "type": "text", + "content": "+Real positive caption: A double decker bus parked at a station." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 480, + 437, + 550, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 480, + 437, + 550, + 473 + ], + "spans": [ + { + "bbox": [ + 480, + 437, + 550, + 473 + ], + "type": "text", + "content": "+ Synthetic positive caption: A double-decker bus parked at a bus station with a few passengers getting on." + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 481, + 473, + 496, + 485 + ], + "blocks": [ + { + "bbox": [ + 481, + 473, + 496, + 485 + ], + "lines": [ + { + "bbox": [ + 481, + 473, + 496, + 485 + ], + "spans": [ + { + "bbox": [ + 481, + 473, + 496, + 485 + ], + "type": "image", + "image_path": "f5f758e2e76f32567ea946c823a4658c96d37f424f9b23f6b426fc32568da5a6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 475, + 548, + 483 + ], + "lines": [ + { + "bbox": [ + 496, + 475, + 548, + 483 + ], + "spans": [ + { + "bbox": [ + 496, + 475, + 548, + 483 + ], + "type": "text", + "content": "Wrong Generation" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 487, + 551, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 551, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 551, + 499 + ], + "type": "text", + "content": "(b) Inconsistency in cross-modal alignment quality of synthetic samples." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 507, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 629 + ], + "type": "text", + "content": "Figure 1. Challenges in generating and training on synthetic data: (a) When generating an image with subtle variations based on a real image and a target caption specifying the variations, an image editing model [6] struggles with text alignment (middle), while an image generation model [75] fails to maintain image fidelity (right). (b) Synthetic positive and negative image_caption pairs show different levels of alignment quality. The subtle variations in the synthetic negative caption (left) make it difficult to distinguish from the positive; the over-modified negative caption (middle) is easy to distinguish; and the hallucinated content in the synthetic positive caption (right) results in an incorrect positive." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "type": "text", + "content": "ing shortcut features [23] rather than genuinely learning nuanced distinctions. While collecting training samples with subtle variations could enhance compositionality, this approach is time-consuming and labor-intensive, rendering it impractical at scale." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "content": "Advances in generative models [7, 12, 20, 43, 61, 75," + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24849" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "78, 91, 105, 110] now facilitate the synthetic generation of training samples with subtle variations between paired samples [11, 18, 49, 68, 71, 79, 83, 86, 106, 108, 109]. By starting with real image-caption pairs, generative models can create subtle edits to both captions and images, providing valuable training data with minimal manual effort. These generated variations enable VLMs to enhance their compositionality by learning from the nuanced differences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 295, + 456 + ], + "type": "text", + "content": "However, generating and training on synthetic data presents two key challenges. The first is the difficulty of efficiently and accurately creating precise variations in synthetic images. To generate large-scale training data with precise variations, the image generation process must meet three criteria: efficiency in producing large quantities of images, text alignment of the generated image to the corresponding caption, which is subtly changed from the original caption, and image fidelity in ensuring the generated image otherwise closely matches their real counterpart. Nevertheless, current image generation methods struggle to meet all three criteria simultaneously. Per-sample optimization methods [44, 82, 94, 112] lack efficiency, while zero-shot image editing methods [6, 22, 63] and text-to-image (T2I) models [75, 78] often fail to achieve proper text alignment and image fidelity, respectively, as illustrated in Figure 1 (a). The inaccurate synthetic variations produced by these models can mislead the learning of VLMs. To alleviate this issue, we take a fast T2I model, which excels in efficiency and text alignment, and inject real image features into the text prompt features in order to enhance image fidelity. By combining this approach with AdaIN [35], we manage to substantially improve the image generation process per the three criteria." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 603 + ], + "type": "text", + "content": "Still, the generated data do not perfectly meet the text alignment criterion, and we propose to deal with the remaining problems with an innovative approach to model training. As Figure 1 (b) shows, the similarity between synthetic positive and negative pairs may vary, resulting in a mixture of hard text-image pairs, easy pairs, and incorrect pairs. Instead of treating all positive and negative samples uniformly, we propose a novel loss that differentiates between positive, hard negative, and easy negative samples. Further, we propose an adaptive margin that helps to filter out potentially incorrect generations and focus the learning on informative hard samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "In summary, we propose SPARCL, which integrates image feature injection into a fast T2I model to improve the quality of synthetic variations in images and employs an adaptive margin loss to leverage the varying alignment quality in synthetic samples for training VLMs. Evaluations on four compositional understanding benchmarks demonstrate that SPARCL significantly enhances VLM compositionality, improving CLIP by over " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " on VL-CheckList [113] and " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " on SugarCrepe [32], while surpassing state-of-the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "art methods by " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " on the two benchmarks. The main contributions of this paper are as follows: (1) we propose image feature injection to enhance the quality of synthetic variations in images, which provide valuable training data to improve the compositionality of VLMs; (2) we introduce an adaptive margin loss to leverage varying levels of cross-modal alignment in synthetic samples to effectively differentiate positive and negative samples; (3) experimental results validate that SPARCL significantly improves the compositional understanding capabilities of CLIP models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 205, + 400, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 400, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 400, + 217 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 224, + 531, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 531, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 531, + 237 + ], + "type": "text", + "content": "2.1. Limitations in Compositionality of VLMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": "While VLMs excel in many multi-modal tasks [40, 52, 53, 74, 84, 101, 115], they still struggle with compositional understanding—the ability to interpret novel combinations of known visual and textual components. Benchmarks like What'sUp [42] reveal difficulties in understanding spatial relationships, SPEC [70] highlights issues with object size, position, and count, and ARO [106] uncovers limitations in understanding attributes, relations, and word order. Winoground [88], SNARE [96], and VL-CheckList [113] also expose these shortcomings. SugarCrepe [32] addresses hackable biases in prior benchmarks, where text-only models achieve artificially high performance, by introducing fluent and meaningful hard negatives. Their findings suggest that previous benchmarks overestimated compositional understanding. Building on this, SugarCrepe++ [19] further introduces semantically equivalent but lexically varied captions as hard positives and shows the difficulties of VLMs in distinguishing between lexical variations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 468, + 515, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 515, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 515, + 480 + ], + "type": "text", + "content": "2.2. Improving Compositionality of VLMs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Prior approaches to improving the compositionality of VLMs can be broadly classified into the following categories: (1) Leveraging detailed image captions: Detailed captions from dense captioning models [17, 54], simulation platforms [9], and video annotations [45] are collected to train VLMs. However, these samples often lack pairs with subtle variations, limiting their contribution to compositionality. (2) Distilling from pretrained models: SDS-CLIP [3], SF-CLIP [80] and IL-CLIP [114] distill knowledge from pretrained image generation models and visual-language foundation models. IL-CLIP [114] refines representations through iterative learning with pretrained vision and language agents. However, pretrained models also face limitations in compositionality [34, 90]. (3) Incorporating structural knowledge: MosaiCLIP [85] and Structure-CLIP [36] incorporate scene graph knowledge in text features. CLIP-SGVL [28] and 3VL [103] train VLMs to predict scene graphs. [66] utilizes scene graphs as prompts to elicit compositional knowledge from VLMs without fur" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24850" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 348 + ], + "type": "text", + "content": "other training. However, these methods rely on models trained with expensive dense structure annotations (e.g., scene graphs). (4) Utilizing synthetic negative samples: Rule-based tools [30] or large language models (LLMs) [15, 73] are used to generate negative captions by editing real captions [10, 11, 18, 68, 83, 86, 106, 109]. Image generation models [75] are also used to create or edit images for training [49, 71, 79, 108]. Despite their utility, efficiently generating large amounts of images with precise variations remains challenging due to the inherent limitations of generative models. (5) Applying fine-grained alignment constraints: MCD [46] enforces multi-scale alignment across images with varying augmentations and the corresponding text captions. SPARC [5] learns local alignment by associating each text token with a group of local image patches. CE-CLIP [109] applies intra-modal contrastive loss and cross-modal ranking loss to improve alignment. However, uniform supervision applied to synthetic samples with varying alignment quality limits their effectiveness. In this paper, we propose SPARCL to address two key challenges in learning from synthetic data: the difficulty in accurately creating precise variations and the inconsistency in cross-modal alignment quality in synthetic data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 354, + 215, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 215, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 215, + 368 + ], + "type": "text", + "content": "2.3. Training with Synthetic Data" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 372, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 575 + ], + "type": "text", + "content": "Synthetic data for training machine learning models have been studied in various fields [14, 26, 48, 64, 67, 76, 77, 93, 99]. Synthetic data generated through simulations and graphics engines supports a wide range of tasks [16, 72, 95]. However, these synthetic datasets often diverge significantly from real-world data. Recent advances in generative models [7, 12, 43, 61, 75, 78, 91, 105, 110] have made it possible to synthesize data that more closely resembles real-world scenarios. Synthetic data are widely used in both language tasks [21, 31, 60, 64, 65, 87, 97, 98, 100, 116] and vision tasks [1, 2, 4, 25, 29, 39, 57, 81, 89, 92, 111]. Synthetic data are often noisy, necessitating noise-resistant training methods [13, 37, 41, 50, 51, 102], especially those designed for contrastive learning [38, 56]. In this paper, we generate multimodal samples with subtle variations and filter out potentially incorrect ones during training to improve the compositionality of VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 586, + 118, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 118, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 118, + 597 + ], + "type": "text", + "content": "3. SPARCL" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "To improve the compositional understanding abilities of VLMs, we propose Synthetic Perturbations for Advancing Robust Compositional Learning, or SPARCL, which generates multimodal samples with subtle variations from real samples and trains VLMs to learn nuanced differences through synthetic data. In the generation phase, SPARCL creates positive and negative captions with slight variations from real captions using an LLM, then generates images based on the real image and these modified captions through" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "a fast T2I model. To enhance the quality of subtle variations in synthetic images, we introduce image feature injection, which integrates real image features into the text prompt features of the T2I model to improve fidelity to the real image, as detailed in Sec. 3.2. In the training phase, SPARCL employs an adaptive margin loss that leverages varying levels of multimodal alignment in the synthetic samples to effectively learn informative nuanced distinctions, as described in Sec. 3.3. The framework of SPARCL is shown in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 199, + 538, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 538, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 538, + 213 + ], + "type": "text", + "content": "3.1. Generating Negative and Positive Captions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "text", + "content": "Captions with subtle variations are crucial for learning compositional knowledge, as shown by previous work that generates captions by randomly swapping or replacing nouns and adjectives [106, 109]. However, manually designed generation rules often introduce nonsensical or grammatically incorrect artifacts, creating shortcut features [23] that obstruct true compositional understanding [32]. To address this issue, we use an LLM to generate natural synthetic captions. To further mitigate the impact of generative artifacts, we generate both negative and positive captions, ensuring that VLMs cannot easily differentiate negative captions from positive ones based solely on artifacts. We denote the " + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "text", + "content": " real image-caption pair in the training set as " + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{I}_i^r,\\boldsymbol{T}_i^r)" + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "text", + "content": ". Given a real caption " + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "inline_equation", + "content": "\\boldsymbol{T}_i^r" + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "text", + "content": ", we prompt the LLM to generate a synthetic negative caption " + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "inline_equation", + "content": "\\boldsymbol{T}_i^{sn}" + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "text", + "content": " and a synthetic positive caption " + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "inline_equation", + "content": "\\boldsymbol{T}_i^{sp}" + }, + { + "bbox": [ + 313, + 217, + 556, + 421 + ], + "type": "text", + "content": ", as specified by the prompts in Figure A1 in the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 427, + 555, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 441 + ], + "type": "text", + "content": "3.2. Generating Images via Image Feature Injection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 445, + 556, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 554 + ], + "type": "text", + "content": "Synthetic images that exhibit subtle variations from real images while aligning with the captions are valuable but challenging to generate, as discussed in Sec. 1. Although previous works [49, 71, 79, 108] have utilized object segmentation or filtered dissimilar generations to improve fidelity to real images, they struggle with manipulating relationships or lack efficiency. We aim to enhance fidelity by injecting image features into a fast T2I model [61], which already achieves high efficiency and text alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 630 + ], + "type": "text", + "content": "Image Feature Injection. The images generated by T2I models lack image fidelity to real images, as no real image information is input to the models. To enhance the fidelity of synthetic images, we inject real image features into the text prompt features, enabling the model to incorporate information from the real images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "In T2I models, content and style are separated in the semantic and padding embeddings. The semantic embeddings (before the embedding of the [EOS] token) usually capture most of the image content in the text prompts, while the padding embeddings (after the [EOS] token) usually represent the image style [104]. Therefore, we can inject real image features into the padding embeddings to guide" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24851" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 90, + 347, + 227 + ], + "blocks": [ + { + "bbox": [ + 60, + 75, + 241, + 87 + ], + "lines": [ + { + "bbox": [ + 60, + 75, + 241, + 87 + ], + "spans": [ + { + "bbox": [ + 60, + 75, + 241, + 87 + ], + "type": "text", + "content": "(a) Generating synthetic captions and images" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 90, + 347, + 227 + ], + "lines": [ + { + "bbox": [ + 60, + 90, + 347, + 227 + ], + "spans": [ + { + "bbox": [ + 60, + 90, + 347, + 227 + ], + "type": "image", + "image_path": "497a83f463f5bee0abcf4b37147e88b27e9e7c59b43f1c10da506ae2c143c20a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 352, + 75, + 551, + 226 + ], + "blocks": [ + { + "bbox": [ + 352, + 75, + 551, + 226 + ], + "lines": [ + { + "bbox": [ + 352, + 75, + 551, + 226 + ], + "spans": [ + { + "bbox": [ + 352, + 75, + 551, + 226 + ], + "type": "image", + "image_path": "3847209be1ab50fffa2408ebd0dc56d68078672bc9d1c6f9837fba18b4a02182.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 247, + 238, + 355 + ], + "blocks": [ + { + "bbox": [ + 59, + 233, + 235, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 233, + 235, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 233, + 235, + 247 + ], + "type": "text", + "content": "(b) Training with real and synthetic samples" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 247, + 238, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 238, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 238, + 355 + ], + "type": "image", + "image_path": "f8ebc6cae65f82dcfdfdafda7c6c32aa8337d353637752417e23f6a58cfe5ca5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 243, + 247, + 455, + 357 + ], + "blocks": [ + { + "bbox": [ + 243, + 236, + 332, + 247 + ], + "lines": [ + { + "bbox": [ + 243, + 236, + 332, + 247 + ], + "spans": [ + { + "bbox": [ + 243, + 236, + 332, + 247 + ], + "type": "text", + "content": "Adaptive Margin Loss" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 243, + 247, + 455, + 357 + ], + "lines": [ + { + "bbox": [ + 243, + 247, + 455, + 357 + ], + "spans": [ + { + "bbox": [ + 243, + 247, + 455, + 357 + ], + "type": "image", + "image_path": "0ec229f2694a8b91cd1a306779aa55a2ea6ef0916c20f5d7649715a48dfb61a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 425 + ], + "type": "text", + "content": "Figure 2. An overview of SPARCL. (a) Starting with a real image-caption pair, we generate synthetic positive and negative pairs with subtle variations using an LLM and a fast T2I model. To improve the quality of subtle variations in synthetic images, we introduce image feature injection to reduce unintended variations from a standard T2I model (see Sec. 3.2). (b) We train the VLM using both real and synthetic samples. In addition to a sigmoid loss for distinguishing positive and negative pairs, we apply an adaptive margin loss that leverages varying alignment levels across training samples to learn informative nuanced distinctions (see Sec. 3.3)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 455, + 236, + 550, + 356 + ], + "blocks": [ + { + "bbox": [ + 455, + 236, + 550, + 356 + ], + "lines": [ + { + "bbox": [ + 455, + 236, + 550, + 356 + ], + "spans": [ + { + "bbox": [ + 455, + 236, + 550, + 356 + ], + "type": "image", + "image_path": "c88d2ee2497dc57fe24ef84887af0bf723486a9c5ae33532b4e969e775179112.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": "the model in generating images with a similar style to real images, without affecting the alignment with the captions, thanks to the decoupling of content and style in the prompt embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": "We extract features for a real image " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{I}_i^r" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": " and the corresponding synthetic caption " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i^s" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "s \\in \\{sn, sp\\}" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": ") using two aligned feature encoders (e.g., CLIP image and text encoders). The image embedding " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^r" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": " is the encoder output at the position of the [CLS] token. The text embedding is the text encoder output " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "e_i^s = \\langle e_{i,j}^s \\rangle_{j=1}^L" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": " is the maximum sequence length. Let the index of the [EOS] token in " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "e_i^s" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "k_i^s" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "k_i^s \\leq L" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": ". That is, the padding token [PAD] is used as input to the text encoder at positions between " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "k_i^s" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": ". The key step of SPARCL is to replace the text embeddings at those positions with the image embedding, producing a new sequence of text embeddings, " + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\hat{e}_i^s" + }, + { + "bbox": [ + 55, + 483, + 295, + 616 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 620, + 295, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 620, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 99, + 620, + 295, + 648 + ], + "type": "interline_equation", + "content": "\\hat {e} _ {i} ^ {s} = \\left\\langle e _ {i, 1} ^ {s}, e _ {i, 2} ^ {s}, \\dots , e _ {i, k _ {i}} ^ {s}, \\underbrace {\\boldsymbol {f} _ {i} ^ {r} , \\dots , \\boldsymbol {f} _ {i} ^ {r}} _ {L - k _ {i} ^ {s} \\text {t i m e s}} \\right\\rangle . \\tag {1}", + "image_path": "b18468beed28ac0c92db7549ecb1f9b2afbe20ed4e2d9023238e72ff13a881cd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "The embeddings " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{e}_i^s" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is then used as input to the T2I model to generate a synthetic image " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{I}_i^s" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ". This process, depicted in Figure 2, reduces unintended variations in synthetic images generated by the standard T2I model and enhances image fidelity to real images, thereby improving the quality of" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 436, + 423, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 423, + 447 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 423, + 447 + ], + "type": "text", + "content": "synthetic subtle variations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "text", + "content": "Style Transfer. To further reduce the domain gap between synthetic and real images, we use AdaIN [35] to transfer the style of the synthetic image to that of the real image. We use a pretrained AdaIN encoder to extract content features from " + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{I}_i^s" + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "text", + "content": " and style features from " + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "inline_equation", + "content": "I_i^r" + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "text", + "content": ". We normalize the content features using instance normalization and then scale and shift them with the mean and variance of the style features. The transformed features are fed into a pretrained AdaIN decoder to generate " + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "inline_equation", + "content": "I_i^s" + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "text", + "content": ", which is subsequently used for model training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 586, + 534, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 534, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 534, + 600 + ], + "type": "text", + "content": "3.3. Training with Real and Synthetic Samples" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "After generating captions and images, each real image-caption pair " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "(I_i^r, T_i^r)" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " is extended to " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "(I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " by adding one synthetic negative and one synthetic positive pair. We then train the VLM using these extended samples. Given a batch of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " sample groups, " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{(I_i^r, T_i^r, I_i^{sn}, T_i^{sn}, I_i^{sp}, T_i^{sp})\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", we organize the images by concatenating all real, synthetic negative, and synthetic positive images into an image batch " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "I^B" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "3n" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " images, and similar all captions into a caption" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24852" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 160, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 160, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 160, + 83 + ], + "type": "text", + "content": "batch " + }, + { + "bbox": [ + 55, + 72, + 160, + 83 + ], + "type": "inline_equation", + "content": "\\pmb{T}^B" + }, + { + "bbox": [ + 55, + 72, + 160, + 83 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 72, + 160, + 83 + ], + "type": "inline_equation", + "content": "3n" + }, + { + "bbox": [ + 55, + 72, + 160, + 83 + ], + "type": "text", + "content": " captions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": "We calculate the similarity between each image-caption pair in the batch. The similarity between the " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " image " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{I}_i^B" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " caption " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{T}_j^B" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "S_{i,j} = c(\\mathcal{E}_I(\\pmb{I}_i^B),\\mathcal{E}_T(\\pmb{T}_j^B))" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "c(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " denotes cosine similarity, and " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_I" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_T" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " represent the image and text encoders of the VLM, respectively. We then define the ground-truth alignment variable " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "M_{i,j}" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": ", which takes value 1 if the image matches with the caption, and -1 otherwise. Clearly, the only positive image-caption pairs are " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "(\\pmb{I}_i^r,\\pmb{T}_i^r),(I_i^{sn},T_i^{sn}),(I_i^{sp},T_i^{sp}),(I_i^r,T_i^{sp})" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "inline_equation", + "content": "(I_i^{sp},T_i^r)" + }, + { + "bbox": [ + 55, + 83, + 296, + 240 + ], + "type": "text", + "content": ". To account for multiple positive associations of an image or a caption, we apply a sigmoid-based contrastive loss [8, 107] to encourage higher similarity for positive pairs and lower similarity for negative pairs," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 247, + 295, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 247, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 66, + 247, + 295, + 277 + ], + "type": "interline_equation", + "content": "L _ {c o n} = - \\frac {1}{3 n} \\sum_ {i = 1} ^ {3 n} \\sum_ {j = 1} ^ {3 n} \\log \\frac {1}{1 + \\exp \\left(- M _ {i , j} \\left(S _ {i , j} / \\tau + b\\right)\\right)}, \\tag {2}", + "image_path": "165c0f22a0088031d9037e2b35a45c2df0982761d74b91e18319fdadb43ef4ae.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 289, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 289, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 289, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 283, + 289, + 295 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 283, + 289, + 295 + ], + "type": "text", + "content": " is the temperature parameter and " + }, + { + "bbox": [ + 55, + 283, + 289, + 295 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 283, + 289, + 295 + ], + "type": "text", + "content": " is a bias term." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "content": "Adaptive Margin Loss. Despite best efforts toward controlling generative models, synthetic samples may still have variable quality in terms of the alignment between text and imagery. That is, purported positive (resp. negative) pairs may not be semantically similar (resp. dissimilar). To account for variation in synthetic data quality, we propose to differentiate between real data and synthetic data and between easy negatives and hard negatives. Further, we propose an adaptive margin that filters out potential incorrect samples and prioritizes learning from hard samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "content": "For each image in a batch, we define four sets of captions, a positive set " + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "content": ", a hard negative set " + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_h" + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "content": ", an easy negative set " + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_e" + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "content": ", and a real negative set " + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_r" + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "content": ", which represent different levels of alignment within the batch. If the image is a real or synthetic positive image," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 479, + 295, + 574 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "content": " contains the real positive captions " + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "inline_equation", + "content": "T_{i}^{r}" + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "content": " and the synthetic positive captions " + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "inline_equation", + "content": "T_{i}^{sp}" + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "content": " of the current, " + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "content": "-th image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 503, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 526 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 503, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_h" + }, + { + "bbox": [ + 55, + 503, + 295, + 526 + ], + "type": "text", + "content": " contains the synthetic negative captions " + }, + { + "bbox": [ + 55, + 503, + 295, + 526 + ], + "type": "inline_equation", + "content": "T_i^{sn}" + }, + { + "bbox": [ + 55, + 503, + 295, + 526 + ], + "type": "text", + "content": " of the current image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 526, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 551 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 526, + 295, + 551 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_e = \\{\\pmb{T}_j^r,\\pmb{T}_j^{sp},\\pmb{T}_j^{sn}|i\\neq j\\}" + }, + { + "bbox": [ + 55, + 526, + 295, + 551 + ], + "type": "text", + "content": " contains all real and synthetic captions from all other images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 551, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 295, + 574 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 551, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_r = \\{\\pmb{T}_j^r | i \\neq j\\}" + }, + { + "bbox": [ + 55, + 551, + 295, + 574 + ], + "type": "text", + "content": " contains all real captions belonging to the other images." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "text", + "content": "If the image is a synthetic negative image, the sets " + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_e" + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_r" + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "text", + "content": " are unchanged but " + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_h" + }, + { + "bbox": [ + 55, + 575, + 295, + 598 + ], + "type": "text", + "content": " differ:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 599, + 255, + 623 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 55, + 599, + 255, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 255, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 255, + 611 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 599, + 255, + 611 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 599, + 255, + 611 + ], + "type": "text", + "content": " contains the synthetic negative captions " + }, + { + "bbox": [ + 55, + 599, + 255, + 611 + ], + "type": "inline_equation", + "content": "T_{i}^{sn}" + }, + { + "bbox": [ + 55, + 599, + 255, + 611 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_h" + }, + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "inline_equation", + "content": "T_i^r" + }, + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "inline_equation", + "content": "T_i^{sp}" + }, + { + "bbox": [ + 55, + 611, + 168, + 623 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 623, + 245, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 245, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 245, + 635 + ], + "type": "text", + "content": "We define the margin loss for the " + }, + { + "bbox": [ + 55, + 623, + 245, + 635 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 623, + 245, + 635 + ], + "type": "text", + "content": " image as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 638, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 638, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 63, + 638, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {m a r, i} ^ {I} = \\frac {1}{| \\mathbb {P} | \\cdot | \\mathbb {N} _ {h} |} \\sum_ {j _ {1} \\in \\mathbb {P}, j _ {2} \\in \\mathbb {N} _ {h}} \\max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}) \\\\ + \\frac {1}{\\left| \\mathbb {N} _ {h} \\right| \\cdot \\left| \\mathbb {N} _ {e} \\right|} \\sum_ {j _ {1} \\in \\mathbb {N} _ {h}, j _ {2} \\in \\mathbb {N} _ {e}} \\max \\left(0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}\\right) \\tag {3} \\\\ + \\frac {\\alpha}{| \\mathbb {P} | \\cdot | \\mathbb {N} _ {r} |} \\sum_ {j _ {1} \\in \\mathbb {P}, j _ {2} \\in \\mathbb {N} _ {r}} \\max (0, m + S _ {i, j _ {2}} - S _ {i, j _ {1}}). \\\\ \\end{array}", + "image_path": "c4cfc394e66845dbb7ed2b2122e06ebb62cc9a91a4f213cc014546022aab4141.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": "This loss encourages positive pairs to have higher similarity scores than hard negative pairs, which should have higher similarity than easy negative pairs. Additionally, a weight " + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\alpha > 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": " is applied to comparisons between positive pairs and real negative pairs to emphasize these comparisons, as real samples are generally correct, making comparisons involving them more reliable." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 156, + 553, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 200 + ], + "type": "text", + "content": "Lastly, we propose an adaptive margin in the above loss. Let " + }, + { + "bbox": [ + 313, + 156, + 553, + 200 + ], + "type": "inline_equation", + "content": "d = S_{i,j_1} - S_{i,j_2}" + }, + { + "bbox": [ + 313, + 156, + 553, + 200 + ], + "type": "text", + "content": " denote the difference between two similarity scores in Eq. (3). The adaptive margin " + }, + { + "bbox": [ + 313, + 156, + 553, + 200 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 156, + 553, + 200 + ], + "type": "text", + "content": " is then computed as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 208, + 553, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 208, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 350, + 208, + 553, + 243 + ], + "type": "interline_equation", + "content": "m = \\left\\{ \\begin{array}{l l} d, & d < \\beta \\\\ (\\frac {m _ {0} - d}{m _ {0} - \\beta} \\gamma + 1) m _ {0}, & \\beta \\leq d \\leq m _ {0} \\\\ m _ {0} & d > m _ {0} \\end{array} \\right. \\tag {4}", + "image_path": "09db38467b329939643d0c83ea108d54c03b20b8241b591ba2630aae1f538e76.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "m_0" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": " is a base margin, " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "\\beta < 0" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": " is a cutoff threshold, " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": " is a scaling factor. The adaptive margin is designed with the following rationale: when " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "d < \\beta" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": ", the purported positive pair is much less similar than the purported negative pair, suggesting the presence of incorrect or mislabeled samples, so we zero out the loss by setting the margin to " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "\\beta \\leq d \\leq m_0" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": ", the margin is scaled, with smaller differences receiving larger margins to emphasize learning from harder samples. When " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "d > m_0" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": ", the samples are already well-separated with the margin " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "m_0" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": ", so the margin is capped at " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "m_0" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": ", again leading to zero loss. We visualize both the adaptive margin and the corresponding loss as functions of " + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 251, + 554, + 406 + ], + "type": "text", + "content": " in Figure 2 (b)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "text", + "content": "The loss for all images is given by " + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "inline_equation", + "content": "L_{mar}^{I} = \\frac{1}{3n}\\sum_{i}L_{mar,i}^{I}" + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "text", + "content": ". Analogously, we compute " + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "inline_equation", + "content": "L_{mar}^{T}" + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "text", + "content": " for all captions. The total adaptive margin loss is " + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "inline_equation", + "content": "L_{mar} = L_{mar}^{I} + L_{mar}^{T}" + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "text", + "content": ". Finally, the overall training loss is a weighted combination of the contrastive loss and the adaptive margin loss, with weight " + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 406, + 554, + 475 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 394, + 483, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 483, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 394, + 483, + 553, + 495 + ], + "type": "interline_equation", + "content": "L = L _ {c o n} + \\lambda L _ {m a r}. \\tag {5}", + "image_path": "153fa8641bfdd8ffad6bda17340e38151f16c084085bbebd41d1d27f24bb8b7f.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 515, + 394, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 394, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 394, + 529 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 536, + 376, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 376, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 376, + 546 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 552, + 554, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 636 + ], + "type": "text", + "content": "Training. We use the COCO-2014 dataset [55] as the training data source. The training set consists of 82,783 images, each paired with five captions. For each real image-caption pair, we generate one positive and one negative synthetic pair. In line with previous approaches [18, 106, 109], we train the VLMs using both the original COCO-2014 training data and the synthetic samples." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "Evaluation. We evaluate our model on four vision-language compositional understanding benchmarks: (1) ARO [106], which consists of 23,937 cases for relation understanding and 28,748 for attribute understanding. We exclude the subsets for order understanding, as they contain significant nonsensical and non-fluent artifacts [32]." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24853" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 101, + 548, + 353 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of accuracy (%) between SPARCL and baselines on four benchmarks. \"img.\" represents images, \"cap.\" represents captions, \"syn.\" represents synthetic data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 101, + 548, + 353 + ], + "lines": [ + { + "bbox": [ + 61, + 101, + 548, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 101, + 548, + 353 + ], + "type": "table", + "html": "
MethodTraining DataAROVL-CheckListSugarCrepeSugarCrepe++
Source# real img.# real cap.# syn. img.# syn. cap.
CLIP [74] (Zero-Shot)-----61.173.273.459.8
CLIP [74] (Finetune)COCO82K410K0064.172.879.962.3
SDS-CLIP [3]COCO82K410K0057.5---
[79]COCO0082K82K65.069.9--
AMR-NegCLIP [83]COCO100K100K0500K79.4-85.2-
NegCLIP [106]COCO100K100K0500K76.074.682.564.9
MosaiCLIP [85]COCO109K109K0981K80.376.8--
FSC-CLIP [68]COCO100K100K01.5M-77.285.1-
CE-CLIP [109]COCO82K410K02M79.776.385.2-
COMO [49]COCO113K567K567K567K-76.9--
SPARCL (our method)COCO82K410K820K820K77.279.287.166.1
SPEC [70]LAION20K20K20K20K70.1---
CounterCurate [108]Flickr30K30K150K150K--82.8-
[18]CC3M3M3M09M-75.3-55.3
CE-CLIP+ [109]COCO+CC3M3M3M015M80.479.387.5-
CLOVE [11]LAION-COCO>1B>1B0>1B73.2-85.1-
IL-CLIP [114]CC12M12M12M00--70.3-
SF-CLIP [80]YFCC15M15M15M00--71.2-
syn-CLIP [9]SyViC00>1M>1M69.274.8--
FiGCLIP [45]VidSitu20K videos0067.0-74.6-
", + "image_path": "3524514efdf5cc4c7fb32200eda28faef94a453bff77696fc34250f0e8499f6a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 372, + 298, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 298, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 298, + 542 + ], + "type": "text", + "content": "(2) VL-CheckList [113], a large-scale benchmark with over 100,000 samples, evaluates compositionality across subsets of objects, attributes, and relationships, which are further divided into various fine-grained categories. (3) SugarCrepe [32] includes 7,000 test cases across seven subsets. In the above three benchmarks, each test case containing one image, one positive caption, and one negative caption. (4) SugarCrepe++ [19] includes 4,757 test samples across five subsets, where each test case consists of one image, two positive captions, and one negative caption. All benchmarks involve classifying captions as positive or negative for the given images. We report the average accuracy across all subsets of each benchmark and include the accuracy for each subset in Appendix A3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 547, + 189, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 189, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 189, + 560 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 297, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 297, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 297, + 650 + ], + "type": "text", + "content": "Models. We use ViT-B/32 and ViT-L/14 architectures from OpenAI's CLIP model [74] as our base models, initialized with pretrained checkpoints. Following syn-CLIP [9], we integrate LoRA adapters [33] into both the image and text encoders of CLIP to improve training efficiency and mitigate knowledge forgetting. Only the LoRA adapters are fine-tuned during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Training Setups. We use the AdamW optimizer [59] with a cosine learning rate schedule [58]. Training is conducted on two Tesla V100 GPUs, with a batch size of 128 sample groups for ViT-B/32 and 16 for ViT-L/14. The base learning rate is set to 0.01 for a total batch size of 256, and scaled" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 372, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 434 + ], + "type": "text", + "content": "linearly [24] based on the actual batch size. Training is performed for 3,000 steps for ViT-B/32 and 15,000 steps for ViT-L/14, corresponding to fewer than 5 epochs in previous studies [18, 106, 109]. More details and hyperparameters settings are in Appendix A2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 443, + 400, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 400, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 400, + 455 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 556 + ], + "type": "text", + "content": "We compare SPARCL with several baseline methods: (1) methods that utilize synthetic samples for training [49, 68, 70, 79, 83, 106, 108, 109], (2) methods that distill knowledge from pretrained models [3, 11, 80, 114], (3) methods that incorporate knowledge from scene graphs [18, 85], and (4) methods that leverage detailed image captions [9, 45]. Both SPARCL and the baseline methods use ViT-B/32 as the base model. The results are presented in Table 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "We observe that SPARCL achieves the best performance on VL-CheckList, SugarCrepe, and SugarCrepe++ compared to baselines trained on the same data source, namely COCO. Specifically, SPARCL surpasses the strongest baseline by " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " on VL-CheckList and " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " on SugarCrepe, underscoring its effectiveness in enhancing compositional understanding. However, SPARCL performs worse on ARO than some baselines. We hypothesize that this may be due to nonsensical or grammatically incorrect artifacts in the ARO test samples [32], which could favor methods that utilize rule-based synthetic training samples (e.g., [85]) containing similar artifacts. Notably, even with only " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " of the training data, SPARCL outperforms most base" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24854" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 101, + 547, + 346 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Ablated performance (\\%) of SPARCL. \"SynCap\" refers to synthetic captions, \"SynImg\" refers to synthetic images, \"FeatInj\" denotes image feature injection, and \"CompSet\" indicates the comparison sets used in Eq. (3)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 101, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 65, + 101, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 65, + 101, + 547, + 346 + ], + "type": "table", + "html": "
ModelVariantSynCapSynImgFeatInjAdaINAdaptive Margin LossAROVL-CheckListSugarCrepeSugarCrepe++Average
CompSetMargin
ViT-B/32#1XXXXX60.4972.6179.3664.8569.33
#2✓XXXX71.7773.5286.3564.3273.99
#3X✓XXX62.6271.7079.9764.8469.79
#4✓✓XXX71.8675.5485.4365.2274.51
#5✓✓✓XX73.4076.5685.5465.7875.32
#6✓✓X✓X73.7975.7285.7964.4974.95
#7✓✓✓✓X74.1276.3585.4066.4475.58
#8✓✓✓✓AllFixed76.7978.5987.0865.4276.97
#9✓✓✓✓AllAdaptive77.1579.1687.1166.1277.38
#10✓✓✓✓AllAdaptive Inversed76.6779.2686.6865.3076.97
#11✓✓✓✓Only (P, Nh)Adaptive77.4880.4886.1564.7077.20
ViT-L/14#1XXXXX59.8073.2681.4964.9069.86
#2✓XXXX72.1675.4487.3864.4474.85
#3X✓XXX60.7072.2382.5866.6370.54
#4✓✓XXX75.2078.2987.2964.6176.35
#5✓✓✓XX75.1178.5887.5464.5776.45
#6✓✓X✓X74.8879.1087.5064.6276.52
#7✓✓✓✓X74.9380.0487.7965.3077.01
#8✓✓✓✓AllFixed75.2180.7588.0266.4177.60
#9✓✓✓✓AllAdaptive75.8380.8188.2366.8377.93
#10✓✓✓✓AllAdaptive Inversed75.1580.7187.9366.6177.60
#11✓✓✓✓Only (P, Nh)Adaptive76.2680.5787.4266.2877.63
", + "image_path": "9d61ed83d7a9b3ace1242042cee4693c9d8eafcf67b645cccad677fb4376ae6f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 365, + 294, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 294, + 474 + ], + "type": "text", + "content": "line methods (see Table 3 for SPARCL performance with reduced training data). Compared to baselines using additional data sources, SPARCL surpasses or matches their performance. For example, SPARCL outperforms methods in [11, 18] despite their use of more training data. SPARCL performs comparably to CE-CLIP+ [109] on VL-CheckList and SugarCrepe, despite CE-CLIP+ leveraging CC3M as an additional data source and utilizing significantly more samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 212, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 212, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 212, + 498 + ], + "type": "text", + "content": "4.4. Ablation Study and Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "type": "text", + "content": "We perform ablation studies to assess the impact of each component and design choice in SPARCL. The results, presented in Table 2, show the overall performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Synthetic captions vs. synthetic images. To analyze the impact of synthetic captions and images, we compare Variant #1 (using only real samples) with Variant #2 (using real samples and synthetic captions) and #3 (using real samples and synthetic images). Variant #2 shows a significant improvement over Variant #1, with the average accuracy improving from " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "69.33\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "73.99\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " for ViT-B/32. However, Variant #3 obtains only a marginal improvements of about " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". These results suggest that synthetic captions substantially enhance compositional understanding, while synthetic images alone provide limited benefit, which aligns with findings in [69]. We hypothesize that generative artifacts in synthetic images are more pronounced than in captions, which negatively affects the effective learning of nu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "text", + "content": "anced distinctions by VLMs. When both synthetic captions and images are used (Variant #4), performance improves further, indicating a synergistic effect." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "text", + "content": "Image feature injection vs. AdaIN. To analyze the impact of image feature injection and AdaIN, we compare Variant #4 (without image feature injection or AdaIN) with Variant #5 (with image feature injection only) and #6 (with AdaIN only). Using ViT-B/32, we observe that Variant #5 outperforms Variant #4 across all four benchmarks, with an average gap of " + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "text", + "content": ". While Variant #6 shows a smaller improvement of " + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "text", + "content": ", it outperforms Variant #5 on ARO and SugarCrepe. These results suggest that image feature injection is more effective for improving compositionality than AdaIN, although the two methods are somewhat complementary. Through the improvements obtained using ViT-L/14 are a little bit different, both techniques improve performance over Variant #4. Combining both methods in Variant #7 leads to further improvements, with a more than " + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "text", + "content": " increase in average accuracy for ViT-B/32 and a " + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "inline_equation", + "content": "0.65\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 696 + ], + "type": "text", + "content": " increase for ViT-L/14 over Variant #4. This highlights the importance of reducing unintended changes in synthetic images, as such changes could cause the model to rely on them for distinguishing positive and negative samples rather than the semantic differences specified in the corresponding captions. Examples of synthetic images in Figure A3 and A4 in the Appendix show how image feature injection helps mitigate these unintended changes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": "Effects of the margin loss and adaptive margin strate-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24855" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 125, + 299, + 192 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 3. Performance (\\%) of SPARCL using different subsets of training samples. \"Neg.\" represents synthetic negative samples, \"Pos.\" represents synthetic positive samples, and \"Prop.\" represents the proportion of training samples." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 125, + 299, + 192 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 299, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 299, + 192 + ], + "type": "table", + "html": "
VariantNeg.Pos.Prop.AROVL-CheckListSugarCrepeSugarCrepe++Average
#7✓×100%72.1274.6586.4157.9972.79
#7✓✓100%74.1276.3585.4066.4475.58
#9✓✓20%75.9077.5685.7864.4475.92
#9✓✓50%77.9079.3886.7164.2177.05
#9✓✓100%77.1579.1687.1166.1277.38
", + "image_path": "a94cd500998d001a314457782a3a5ed2706114dd3ca8c958abe1f9fcd01b87f1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "text", + "content": "gies. We compare three variants with distinct margin approaches: (1) Fixed (Variant #8): a fixed margin without adaptive adjustments; (2) Adaptive (Variant #9): the margin is calculated according to Eq. (4); and (3) Adaptive Inversed (Variant #10): a larger margin is applied to samples with higher similarity differences, representing an inverse version of Eq. (4) that prioritizes learning from easier samples. Comparing Variant #8 with Variant #7, we observe that #8 improves by over " + }, + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "text", + "content": " for ViT-B/32 and " + }, + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "text", + "content": " for ViT-L/14 in average accuracy, suggesting that margin loss effectively aids in compositional understanding. Between Variants #8, #9, and #10, we find that Variant #9 outperforms Variant #8 across all four benchmarks, with an average boost of over " + }, + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 54, + 212, + 297, + 416 + ], + "type": "text", + "content": " for both ViT-B/32 and ViT-L/14. In contrast, Variant #10 results in a performance decline on three benchmarks and no improvement in average accuracy. These findings highlight the superior effectiveness of adaptive margins." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 421, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 297, + 673 + ], + "type": "text", + "content": "Is applying adaptive margin loss only to hard samples sufficient? To explore whether applying margin loss solely to hard samples is adequate, we construct Variant #11, which uses only the positive set " + }, + { + "bbox": [ + 56, + 421, + 297, + 673 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 56, + 421, + 297, + 673 + ], + "type": "text", + "content": " and the hard negative set " + }, + { + "bbox": [ + 56, + 421, + 297, + 673 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_h" + }, + { + "bbox": [ + 56, + 421, + 297, + 673 + ], + "type": "text", + "content": " for margin loss calculation in Eq. (3), which is similar to learning strategies in [49, 109]. When comparing Variant #11 with Variant #8 and #9, we observe that Variant #11 performs better on ARO and VL-CheckList, but worse on SugarCrepe and SugarCrepe++. We hypothesize that by focusing exclusively on the positive and hard negative sets, the model may learn nonsensical artifacts present in the hard negatives. These artifacts improve performance on ARO and VL-CheckList, where the test samples exhibit similar patterns, but hinder performance on SugarCrepe and SugarCrepe++, which are designed to avoid such patterns [32]. In contrast, Variant #9, which incorporates all sets in Eq. (3) for margin loss calculation, achieves better results on SugarCrepe and SugarCrepe++ with only a minor drop in performance on ARO and VL-CheckList. This suggests that easy negative samples serve as regularization, preventing overfitting to the artifacts in hard negative samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Effects of synthetic positive samples. We compare the performance of training with only synthetic negative samples versus using both synthetic negative and positive samples" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "in Table 3. Based on Variant #7, we train ViT-B/32 with these two combinations of training data. We observe nearly " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " decrease in average accuracy when using only synthetic negative samples, compared to using both synthetic negative and positive samples. This decrease is particularly noticeable on SugarCrepe++, ARO, and VL-CheckList. These results underscore the importance of incorporating synthetic positive samples in training, as they provide essential information about variations that maintain semantic consistency, which is critical for compositional understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "text", + "content": "Influence of model size and training data size. To evaluate the impact of model size on compositional understanding, we compare the performance of ViT-B/32 and ViT-L/14 in Table 2. We find that ViT-L/14 only provides a modest improvement of about " + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "text", + "content": " of average accuracy over ViT-B/32. This suggests that increasing model size does not significantly enhance compositional understanding. To investigate the effect of training data size, we train ViT-B/32 using " + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "text", + "content": " of randomly sampled training data and report the results in Table 3. We observe that performance improves as the proportion of training data increases, indicating that a larger training set helps the model better capture compositional knowledge. However, the performance gain diminishes when the training data size increases from " + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 199, + 556, + 391 + ], + "type": "text", + "content": ". Perhaps generating or selecting high-quality data would be more effective than simplistic data scaling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 393, + 556, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 418 + ], + "type": "text", + "content": "Additional experiments on hyperparameter analysis can be found in Appendix A3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 435, + 388, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 388, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 388, + 449 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 457, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 556, + 601 + ], + "type": "text", + "content": "In this paper, we introduce SPARCL to enhance the compositional understanding capabilities of VLMs by generating and training with synthetic data. To tackle two key challenges in using synthetic data—namely, the difficulty of generating accurate variations and the inconsistency in cross-modal alignment quality—SPARCL integrates image feature injection into a T2I model to improve the quality of synthetic variations and introduces an adaptive margin loss to account for varying levels of cross-modal alignment for effectively learning nuanced distinctions. Experiments on four visual-language compositional understanding benchmarks demonstrate the effectiveness of SPARCL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 619, + 424, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 424, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 424, + 633 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF-NRFF13-2021-0006), Singapore. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24856" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 136, + 295, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 136, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 66, + 136, + 295, + 178 + ], + "type": "text", + "content": "[2] Manel Baradad Jurjo, Jonas Wulff, Tongzhou Wang, Phillip Isola, and Antonio Torralba. Learning to see by looking at noise. Advances in Neural Information Processing Systems, 34:2556-2569, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 66, + 180, + 295, + 223 + ], + "type": "text", + "content": "[3] Samyadeep Basu, Maziar Sanjabi, Daniela Massiceti, Shell Xu Hu, and Soheil Feizi. Augmenting clip with improved visio-linguistic reasoning. arXiv preprint arXiv:2307.09233, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 224, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 224, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 66, + 224, + 295, + 280 + ], + "type": "text", + "content": "[4] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This dataset does not exist: training models from generated images. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 281, + 295, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 281, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 295, + 344 + ], + "type": "text", + "content": "[5] Ioana Bica, Anastasija Ilic, Matthias Bauer, Goker Erdoğan, Matko Bošnjak, Christos Kaplanis, Alexey A Gritsenko, Matthias Minderer, Charles Blundell, Razvan Pascanu, et al. Improving fine-grained understanding in image-text pre-training. arXiv preprint arXiv:2401.09865, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 347, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 347, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 66, + 347, + 295, + 401 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 403, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 403, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 66, + 403, + 295, + 467 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 469, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 469, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 66, + 469, + 295, + 523 + ], + "type": "text", + "content": "[8] Adrian Bulat, Yassine Ouali, and Georgios Tzimiropoulos. Fff: Fixing flawed foundations in contrastive pre-training results in very strong vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14172-14182, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 525, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 525, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 66, + 525, + 295, + 600 + ], + "type": "text", + "content": "[9] Paola Cascante-Bonilla, Khaled Shehada, James Seale Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gul Varol, Aude Oliva, Vicente Ordonez, Rogerio Feris, et al. Going beyond nouns with vision & language models using synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20155-20165, 2023. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 295, + 645 + ], + "type": "text", + "content": "[10] Paola Cascante-Bonilla, Yu Hou, Yang Trista Cao, Hal Daumé III, and Rachel Rudinger. Natural language inference improves compositionality in vision-language models. arXiv preprint arXiv:2410.22315, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "text", + "content": "[11] Santiago Castro, Amir Ziai, Avneesh Saluja, Zhuoning Yuan, and Rada Mihalcea. Clove: Encoding compositional language in contrastive vision-language models. arXiv preprint arXiv:2402.15021, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "text", + "content": "William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 553, + 162 + ], + "type": "text", + "content": "[13] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 164, + 553, + 217 + ], + "type": "text", + "content": "[14] Yabo Dan, Yong Zhao, Xiang Li, Shaobo Li, Ming Hu, and Jianjun Hu. Generative adversarial networks (gan) based efficient sampling of chemical composition space for inverse design of inorganic materials. npj Computational Materials, 6(1):84, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 220, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 220, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 553, + 251 + ], + "type": "text", + "content": "[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 254, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 254, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 254, + 553, + 319 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 321, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 553, + 376 + ], + "type": "text", + "content": "[17] Sivan Doveh, Assaf Arbelle, Sivan Harary, Amit Alfassy, Roei Herzig, Donghyun Kim, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, et al. Dense and aligned captions (dac) promote compositional reasoning in v1 models. arXiv preprint arXiv:2305.19595, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 377, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 377, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 377, + 553, + 453 + ], + "type": "text", + "content": "[18] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "type": "text", + "content": "[19] Sri Harsha Dumpala, Aman Jaiswal, Chandramouli Sastry, Evangelos Milios, Sageev Oore, and Hassan Sajjad. Sugarcrepe++ dataset: Vision-language model sensitivity to semantic and lexical alterations. arXiv preprint arXiv:2406.11171, 2024. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 512, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 577 + ], + "type": "text", + "content": "[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 321, + 579, + 553, + 633 + ], + "type": "text", + "content": "[21] Jiahui Gao, Renjie Pi, Lin Yong, Hang Xu, Jiacheng Ye, Zhiyong Wu, Weizhong Zhang, Xiaodan Liang, Zhenguo Li, and Lingpeng Kong. Self-guided noise-free data generation for efficient zero-shot learning. In International Conference on Learning Representations (ICLR 2023), 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 553, + 678 + ], + "type": "text", + "content": "[22] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. arXiv preprint arXiv:2403.14602, 2024. 2, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "text", + "content": "[23] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural net" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24857" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "text", + "content": "works.Nature Machine Intelligence,2(11):665-673,2020. 1,3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 150 + ], + "type": "text", + "content": "[24] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "type": "text", + "content": "[25] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 241 + ], + "type": "text", + "content": "[26] Xuanli He, Islam Nassar, Jamie Kiros, Gholamreza Haffari, and Mohammad Norouzi. Generate, annotate, and learn: Nlp with synthetic text. Transactions of the Association for Computational Linguistics, 10:826-842, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "type": "text", + "content": "[27] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 298, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 294, + 353 + ], + "type": "text", + "content": "[28] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arbelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 408 + ], + "type": "text", + "content": "[29] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "type": "text", + "content": "[30] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1):411-420, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "type": "text", + "content": "[31] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language models with (almost) no human labor. arXiv preprint arXiv:2212.09689, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 544 + ], + "type": "text", + "content": "[32] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 294, + 589 + ], + "type": "text", + "content": "[33] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 644 + ], + "type": "text", + "content": "[34] Hang Hua, Yunlong Tang, Ziyun Zeng, Liangliang Cao, Zhengyuan Yang, Hangfeng He, Chenliang Xu, and Jiebo Luo. Mmcomposition: Revisiting the compositionality of pre-trained vision-language models. arXiv preprint arXiv:2410.09733, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "text", + "content": "[35] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Yufeng Huang, Jiji Tang, Zhuo Chen, Rongsheng Zhang, Xinfeng Zhang, Weijie Chen, Zeng Zhao, Zhou Zhao," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 117 + ], + "type": "text", + "content": "Tangjie Lv, Zhipeng Hu, et al. Structure-clip: Towards scene graph knowledge to enhance multi-modal structured representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2417-2425, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "type": "text", + "content": "[37] Zhizhong Huang, Junping Zhang, and Hongming Shan. Twin contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 162, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 553, + 216 + ], + "type": "text", + "content": "[38] Sarah Ibrahimi, Arnaud Sors, Rafael Sampaio de Rezende, and Stéphane Clinchant. Learning with label noise for image retrieval by selecting interactions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2181-2190, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 553, + 249 + ], + "type": "text", + "content": "[39] Ali Jahanian, Xavier Puig, Yonglong Tian, and Phillip Isola. Generative models as a data source for multiview representation learning. arXiv preprint arXiv:2106.05258, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 250, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 250, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 320, + 250, + 553, + 315 + ], + "type": "text", + "content": "[40] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 316, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 553, + 369 + ], + "type": "text", + "content": "[41] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 320, + 371, + 553, + 414 + ], + "type": "text", + "content": "[42] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's\" up\" with vision-language models? investigating their struggle with spatial reasoning. arXiv preprint arXiv:2310.19785, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 469 + ], + "type": "text", + "content": "[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 471, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 471, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 320, + 471, + 553, + 534 + ], + "type": "text", + "content": "[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 536, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 536, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 320, + 536, + 553, + 569 + ], + "type": "text", + "content": "[45] Zeeshan Khan, Makarand Tapaswi, et al. Figclip: Fine-grained clip adaptation via densely annotated videos. arXiv preprint arXiv:2401.07669, 2024. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 570, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 570, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 570, + 555, + 624 + ], + "type": "text", + "content": "[46] Bumsoo Kim, Yeonsik Jo, Jinhyung Kim, and Seunghwan Kim. Misalign, contrast then distill: Rethinking misalignments in language-image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 625, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 553, + 646 + ], + "type": "text", + "content": "[47] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 680 + ], + "type": "text", + "content": "[48] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "text", + "content": "[49] Chengen Lai, Shengli Song, Sitong Yan, and Guangneng Hu. Improving vision and language concepts understanding with multimodal counterfactual samples. In European" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24858" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 174-191. Springer, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 95, + 295, + 139 + ], + "type": "text", + "content": "[50] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5051-5059, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "type": "text", + "content": "[51] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9485-9494, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 237 + ], + "type": "text", + "content": "[52] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "text", + "content": "[53] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "type": "text", + "content": "[54] Liunian Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 393 + ], + "type": "text", + "content": "[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 394, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 294, + 458 + ], + "type": "text", + "content": "[56] Chang Liu, Han Yu, Boyang Li, Zhiqi Shen, Zhanning Gao, Peiran Ren, Xuansong Xie, Lizhen Cui, and Chunyan Miao. Noise-resistant deep metric learning with ranking-based instance selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6811-6820, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 502 + ], + "type": "text", + "content": "[57] Hao Liu, Tom Zahavy, Volodymyr Mnih, and Satinder Singh. Palm up: Playing in the latent manifold for unsupervised pretraining. Advances in Neural Information Processing Systems, 35:35880-35893, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 504, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 294, + 536 + ], + "type": "text", + "content": "[58] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 537, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 294, + 559 + ], + "type": "text", + "content": "[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 601 + ], + "type": "text", + "content": "[60] Yujie Lu, Wanrong Zhu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Imagination-augmented natural language understanding. arXiv preprint arXiv:2204.08535, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 646 + ], + "type": "text", + "content": "[61] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "text", + "content": "[62] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 117 + ], + "type": "text", + "content": "[63] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 162 + ], + "type": "text", + "content": "[64] Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. Generating training data with language models: Towards zero-shot language understanding. Advances in Neural Information Processing Systems, 35:462-477, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 553, + 217 + ], + "type": "text", + "content": "[65] Yu Meng, Martin Michalski, Jiaxin Huang, Yu Zhang, Tarek Abdelzaher, and Jiawei Han. Tuning language models as training data generators for augmentation-enhanced few-shot learning. In International Conference on Machine Learning, pages 24457-24477. PMLR, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 262 + ], + "type": "text", + "content": "[66] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 264, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 264, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 320, + 264, + 553, + 286 + ], + "type": "text", + "content": "[67] Sergey I Nikolenko. Synthetic data for deep learning. Springer, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 288, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 288, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 320, + 288, + 553, + 331 + ], + "type": "text", + "content": "[68] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pre-trained vlms for improving vision-linguistic compositionality. arXiv preprint arXiv:2410.05210, 2024. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 320, + 332, + 553, + 386 + ], + "type": "text", + "content": "[69] Maitreya Patel, Abhiram Kusumba, Sheng Cheng, Changhoon Kim, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Tripletclip: Improving compositional reasoning of clip via synthetic vision-language negatives. arXiv preprint arXiv:2411.02545, 2024. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 553, + 431 + ], + "type": "text", + "content": "[70] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. arXiv preprint arXiv:2312.00081, 2023. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 434, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 555, + 488 + ], + "type": "text", + "content": "[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize diagnose and optimize: Towards fine-grained vision-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13279-13288, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "type": "text", + "content": "[72] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 535, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 535, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 320, + 535, + 553, + 556 + ], + "type": "text", + "content": "[73] Alec Radford. Improving language understanding by generative pre-training. 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 558, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 558, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 320, + 558, + 553, + 623 + ], + "type": "text", + "content": "[74] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 624, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 555, + 679 + ], + "type": "text", + "content": "[75] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "text", + "content": "[76] Andrew Rosenberg, Yu Zhang, Bhuvana Ramabhadran, Ye Jia, Pedro Moreno, Yonghui Wu, and Zelin Wu. Speech recognition with augmented synthesized speech. In 2019" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24859" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "IEEE automatic speech recognition and understanding workshop (ASRU), pages 996-1002. IEEE, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 160 + ], + "type": "text", + "content": "[77] Nick Rossenbach, Albert Zeyer, Ralf Schlüter, and Hermann Ney. Generating synthetic audio data for attention-based speech recognition systems. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7069-7073. IEEE, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 162, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 294, + 236 + ], + "type": "text", + "content": "[78] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 239, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 303 + ], + "type": "text", + "content": "[79] Ugur Sahin, Hang Li, Qadeer Khan, Daniel Cremers, and Volker Tresp. Enhancing multimodal compositional reasoning of visual language models with generative negative mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5563-5573, 2024. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 304, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 358 + ], + "type": "text", + "content": "[80] Sepehr Sameni, Kushal Kafle, Hao Tan, and Simon Jenni. Building vision-language models on solid foundations with masked distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14216-14226, 2024. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 361, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 414 + ], + "type": "text", + "content": "[81] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 415, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 480 + ], + "type": "text", + "content": "[82] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 294, + 525 + ], + "type": "text", + "content": "[83] Ziyi Shou and Fangzhen Lin. Enhancing semantic understanding in vision language models using meaning representation negative generation. In Fourth Workshop on Knowledge-infused Learning, 2024. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 526, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 590 + ], + "type": "text", + "content": "[84] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 592, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 294, + 645 + ], + "type": "text", + "content": "[85] Harman Singh, Pengchuan Zhang, Qifan Wang, Mengjiao Wang, Wenhan Xiong, Jingfei Du, and Yu Chen. Coarse-to-fine contrastive learning in image-text-graph space for improved vision-language compositionality. arXiv preprint arXiv:2305.13812, 2023. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 690 + ], + "type": "text", + "content": "[86] Jaisidh Singh, Ishaan Shrivastava, Mayank Vatsa, Richa Singh, and Aparna Bharati. Learn\" no\" to say\" yes\" better: Improving vision-language models via negations. arXiv preprint arXiv:2403.20312, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[87] Tianyi Tang, Yushuo Chen, Yifan Du, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen. Learning to imagine: Visually-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 339, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 74, + 553, + 95 + ], + "type": "text", + "content": "augmented natural language generation. arXiv preprint arXiv:2305.16944, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "type": "text", + "content": "[88] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 553, + 217 + ], + "type": "text", + "content": "[89] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 274 + ], + "type": "text", + "content": "[90] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 553, + 330 + ], + "type": "text", + "content": "[91] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 332, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 332, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 332, + 553, + 409 + ], + "type": "text", + "content": "[92] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 553, + 453 + ], + "type": "text", + "content": "[93] Allan Tucker, Zhenchen Wang, Ylenia Rotalinti, and Puja Myles. Generating high-fidelity synthetic patient data for assessing machine learning healthcare software. NPJ digital medicine, 3(1):1-13, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 456, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 456, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 320, + 456, + 553, + 509 + ], + "type": "text", + "content": "[94] Dani Valevski, Matan Kalman, Eyal Molad, Eyal Segalis, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning a diffusion model on a single image. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 512, + 553, + 555 + ], + "type": "text", + "content": "[95] Gül Varol, Ivan Laptev, Cordelia Schmid, and Andrew Zisserman. Synthetic humans for action recognition from unseen viewpoints. International Journal of Computer Vision, 129(7):2264-2287, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "type": "text", + "content": "[96] Fei Wang, Liang Ding, Jun Rao, Ye Liu, Li Shen, and Changxing Ding. Can linguistic knowledge improve multimodal alignment in vision-language pretraining? arXiv preprint arXiv:2308.12898, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 602, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 553, + 655 + ], + "type": "text", + "content": "[97] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "type": "text", + "content": "[98] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24860" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 294, + 127 + ], + "type": "text", + "content": "[99] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. arXiv preprint arXiv:2004.11546, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[100] Yue Yang, Wenlin Yao, Hongming Zhang, Xiaoyang Wang, Dong Yu, and Jianshu Chen. Z-lavi: Zero-shot language solver fueled by visual imagination. arXiv preprint arXiv:2210.12261, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[101] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 275 + ], + "type": "text", + "content": "[102] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[103] Nir Yellinek, Leonid Karlinsky, and Raja Giryes. 3vl: using trees to teach vision & language models compositional concepts. arXiv preprint arXiv:2312.17345, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "type": "text", + "content": "[104] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "text", + "content": "[105] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 455 + ], + "type": "text", + "content": "[106] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "text", + "content": "[107] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "type": "text", + "content": "[108] Jianrui Zhang, Mu Cai, Tengyang Xie, and Yong Jae Lee. Countercurate: Enhancing physical and semantic visio-linguistic compositional reasoning via counterfactual examples. arXiv preprint arXiv:2402.13254, 2024. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 590 + ], + "type": "text", + "content": "[109] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic fine-grained understanding. arXiv preprint arXiv:2306.08832, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 591, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 646 + ], + "type": "text", + "content": "[110] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[111] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 329 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[112] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[113] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "text", + "content": "[114] Chenhao Zheng, Jieyu Zhang, Aniruddha Kembhavi, and Ranjay Krishna. Iterated learning improves compositionality in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13785-13795, 2024. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "text", + "content": "[115] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "text", + "content": "[116] Wanrong Zhu, An Yan, Yujie Lu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Visualize before you write: Imagination-guided open-ended text generation. arXiv preprint arXiv:2210.03765, 2022. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24861" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_content_list.json b/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b89aa14f0be6c738b5e3379ff655e362f9c39191 --- /dev/null +++ b/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_content_list.json @@ -0,0 +1,1977 @@ +[ + { + "type": "text", + "text": "EnliveningGS: Active Locomotion of 3DGS", + "text_level": 1, + "bbox": [ + 279, + 130, + 718, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siyuan Shen $^{1}$ Tianjia Shao $^{1*}$ Kun Zhou $^{1}$ Chenfanfu Jiang $^{2}$ Yin Yang $^{3}$ $^{1}$ State Key Laboratory of CAD&CG, Zhejiang University, $^{2}$ UCLA, $^{3}$ University of Utah", + "bbox": [ + 143, + 179, + 846, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 250, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents a novel pipeline named EnliveningGS, which enables active locomotion of 3D models represented with 3D Gaussian splatting (3DGS). We are inspired by the fact that real-world lives pose their bodies in a natural and physically meaningful manner by compressing or elongating muscle fibers embedded in the body. EnliveningGS aims to replicate the similar functionality of 3DGS models so that the object within a 3DGS scene acts like a living creature rather than a static shape — they walk, jump, and twist in the scene under provided motion trajectories driven by muscle activations. While the concept is straightforward, many challenging technical difficulties need to be taken care of. Synthesizing realistic locomotion of a 3DGS model embodies an inverse physics problem of very high dimensions. The core challenge is how to efficiently and robustly model frictional contacts between an “enlivened model” and the environment, as it is the composition of contact/collision/friction forces triggered by muscle activation that generates the final movement of the object. We propose a hybrid numerical method mixing LCP and penalty method to tackle this NP-hard problem robustly. Our pipeline also addresses the limitation of existing 3DGS deformation algorithms and inpainting the missing information when models move around.", + "bbox": [ + 88, + 282, + 485, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 672, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of learning-based radiance field rendering techniques like NeRF [34] and 3D Gaussian splitting (3DGS) [26] have rebranded the pipeline of novel view synthesis. From synthesized RGB information, users can further alter and edit the illumination, appearance, and geometry. It is also possible to integrate those neural scene representations with physics to synthesize plausible physics-based dynamics. For instance, PIE-NeRF uses a quadratic moving least square interpolation to approximate the nonlinear elastodynamics of a NeRF model [11]. PhysGaussian [45] integrates 3DGS with the material point method (MPM) [39]. GaussianSplashing leverages a universal", + "bbox": [ + 89, + 698, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0a6ac79fe35d4d681cd010c4b9b97c7468258d5cd0c6169c59fc58bd80b25f8.jpg", + "image_caption": [ + "Figure 1. Jumping Chess. We use static 3DGS as input and generates physics-aware 4D animations. The chess piece contracts and expands its body, creating jumping motions through collisions and friction with the chessboard, fitting the user's input trajectory." + ], + "image_footnote": [], + "bbox": [ + 517, + 252, + 901, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "particle-based discretization to animate 3DGS scenes with fluids [10]. Those frameworks allow deeper integrations of 3DGS with relevant downstream graphics/vision tasks, enabling a wide range of exciting applications.", + "bbox": [ + 511, + 531, + 906, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is a common paradigm, when combining physics and 3DGS, that the physical effects of the neural scene are passively triggered by external forces, which are normally specified by the user or pre-scripted (e.g., the gravity force). This makes sense for inanimate objects — they are not changing their kinematic states unless forced by external excitements. However, the locomotion of a living being follows a distinct modality. Rather than relying on external forces passively, a living entity actively adjusts its poses to achieve a desired movement such as walking, bending, or jumping. The enabler of such locomotion is muscle activations that contract or elongate different parts of the body, which interact with the environment e.g., the floor. The resultant contact and friction forces, in turn, provide the necessary momentum of the motion. This procedure is beyond the capability of existing physics-based 3DGS techniques as the muscle activations can hardly be measured or synthesized with a passive physics simulator. It is unfortunate that an interesting 3DGS scene can only house lifeless objects.", + "bbox": [ + 511, + 595, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we show a novel framework to further en", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 114, + 887, + 245, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "896", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hance the 3DGS expressivity. We name our framework EnliveningGS, which inversely computes muscle activations embedded in a 3DGS model that lead to user-specific motion. As shown in Figure 1, the user only needs to provide a desired trajectory of the object (or other high-level descriptions of the kinematics). Given a set of embedded muscle fibers (approximated by piece-wise linear springs), EnliveningGS segments the object from the scene, estimates the possible contact/friction forces between the object and its surrounding environment, and calculates the optimal muscle activations to drive the corresponding body motion so that the object behaves like a living creature and moves following the prescribed trajectories as much as possible. As contact modeling is known to be an NP-hard problem and quickly becomes unsolvable when there are a moderate number of contact incidents, EnliveningGS introduces a two-stage locomotion procedure to reduce the solution space and efficiently handle frictional contacts between the 3DGS model and the scene. To summarize, our contributions include:", + "bbox": [ + 89, + 90, + 480, + 391 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Active Animatable 3DGS Pipeline: We present En-livingGS pipeline that brings any household object to life. When they walk or jump, occluded areas become visible again. To overcome such visual artifacts, we propose an improved inpainting algorithm to fill colors and textures back to those exposed areas based on the known prior of the environment. We also present a self-splitting mechanism for Gaussian kernels to capture and reduce spiky artifacts under significant deformation.", + "- Hybrid Contact Modeling Framework for 3DGS: We design a novel hybrid numerical procedure combing the penalty method, which softens the hard contact constraint, and the classic Linear Complementarity Programming (LCP) [38] so that the combinatorial space of the active set can be significantly reduced, and the two-way coupling with the underlying nonlinear physics simulation becomes feasible." + ], + "bbox": [ + 89, + 398, + 480, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 676, + 232, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian editing. Recently, 3D Gaussian splatting (GS) [25] utilizes a set of 3D Gaussian kernels to represent static scenes explicitly and achieve state-of-the-art results. Subsequently, editing on GS has become a popular topic[7, 13, 30, 31, 47, 48, 50]. To make static 3D scenes deformable, some works extract explicit high-quality meshes from 3DGS representations and align GS kernels with mesh surfaces for guidance when deformed [12, 14, 44]. Their reliance on strong topological information and high-quality meshes makes it difficult to balance quality and interactivity. SC-GS [20] learns sparse control points for 3D scene dynamics but faces challenges with large movements and complex deformation.", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physics-aware dynamics on 3D GS. To combine a static GS scene with a physics engine to generate dynamics on GS, PhysGaussian[45] leverages the first-order information from the displacement map (deformation gradient) to assist dynamic simulations. VR-GS [23] constructed a tetrahedral cage for each segmented Gaussian kernel group and embedded these kernel groups into corresponding meshes, allowing the deformed mesh driven by eXtended Position-based Dynamics (XPBD) [32] to guide the deformation of the Gaussian kernels. GASP[6] treats the simulation process as a black box and integrates Newtonian dynamics into the GaMeS[44] framework. PhysDreamer [49] distills dynamics priors learned by video generation models. It is worth mentioning that existing methods consider the simulation process as a forward process, leaving the inverse solving and control of physical parameters of objects on 3D GS as an unexplored area.", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contacts in locomotion control Friction plays a crucial role in soft body locomotion control. Stewart and Trinkle linearize the 3D friction model using a polyhedral cone and propose an LCP formulation to approximate Coulomb's friction cone conditions [38] and many improve upon it [2, 3, 24, 35, 43]. LCP-based methods offer exact solutions and are efficient for small to medium-sized problems but can be computationally expensive for optimization problems. Due to this drawback, many previous methods explicitly assumed that the contacts remain static [21, 27] while optimizing control forces subject to equations of motion. A few studies [41] exploit the physical meaning of complementarity constraints as heuristics to improve the solution and performance of the solver but still often fall back into exhaustive searching due to a lack of good initial state. The penalty method models contacts and frictions as spring forces and evaluates them at the instant [16-18, 42, 46] but can introduce numerical challenges such as stiffness and instability [1, 36]. The combination of the penalty methods and LCP-based methods has been less explored.", + "bbox": [ + 511, + 368, + 903, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 686, + 604, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 710, + 624, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A 3DGS scene consists of a collection of 3D Gaussians $\\mathcal{G}_{\\text {scene }}$ . Each Gaussian kernel is defined by a center position $\\mu \\in \\mathbb{R}^3$ , a covariance matrix $\\Sigma$ , the opacity $\\sigma \\in [0,1)$ and a set of spherical harmonics coordinates $c \\in \\mathbb{R}^k$ describing the colors emitted by the Gaussian in all directions. The covariance matrix $\\Sigma$ can be factorized into a rotation matrix $\\mathbf{R}$ expressed as a quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ and a scaling matrix $\\mathbf{S}$ of its diagonal vector $\\mathbf{s} \\in \\mathbb{R}^3$ : $\\boldsymbol{\\Sigma} = \\mathbf{RSS}^T\\mathbf{R}^T$ .", + "bbox": [ + 511, + 733, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vanilla 3DGS represents a static environment where users can switch their viewpoints within the scene freely. It is also possible to include dynamics to the scene by adding", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "897", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2de87bca3b56f490f5be11449fd51a68385a6ffa02669947cb6192d6dd88bb4.jpg", + "image_caption": [ + "Figure 2. Bidirectional local embedding. Our bidirectional local embedding (a) serves as guidance for (b) self-splitting GS deformation, (c) accurate collision handling, and (d) non-penetration regularization." + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 478, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "external forces e.g., as in PhysGaussian [45] so that 3DGS models move or deform under user-specified forces passively. Nevertheless, real-world living creatures are self-driven — their movements are enabled by embedded muscles given a desired kinematic motion target. Such an active movement mechanism is much more difficult to obtain as it inevitably involves frictional contacts between the model and the environment. To this end, we incorporate additional physical expressions into the Gaussian elements, enabling them to be self-driven in a physics-based way. These elements can change their appearance at each time step and approximate user-specified motion targets interacting with the environment.", + "bbox": [ + 89, + 281, + 483, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hybrid Mesh-Gaussian Representation", + "text_level": 1, + "bbox": [ + 89, + 488, + 426, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{G}_o$ denote the Gaussian elements of an object of interest, which is segmented from the scene $\\mathcal{G}_{\\text{scene}}$ . We first create an encapsulating tetrahedral mesh $\\mathcal{T} = (\\mathcal{V}, \\mathcal{F})$ to approximate the object's overall volume, mass, and deformation. Here, $\\mathcal{V}$ and $\\mathcal{F}$ refer to the sets of nodal points and faces in $\\mathcal{T}$ . To achieve this segmentation, we approximate the nearby environment with planes (more in Secs. 3.5 and 3.6). The plane $P$ expressed as a unit normal vector $\\mathbf{n}$ and its distance $d$ from the origin:", + "bbox": [ + 89, + 508, + 483, + 645 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathbf {n}, d) = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} \\mid \\mathbf {N} ^ {T} \\mathbf {x} + d = 0 \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 656, + 482, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is used to \"cut out\" irrelevant Gaussians, reduce the blurred regions at the cut, and serve as a prior for inpainting the missing parts after segmentation. Additionally, it is utilized for collision detection to determine the collision forces acting on $\\mathcal{G}_o$ .", + "bbox": [ + 89, + 684, + 482, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{T}$ determines the states of Gaussian kernels in $\\mathcal{G}_o$ after deformation. The deformed position of a point $\\mathbf{x}'$ originally at $\\mathbf{x}$ within a tetrahedron is represented by the barycentric coordinates of the four vertices of the tetrahedron $\\mathbf{v}'$ :", + "bbox": [ + 89, + 760, + 483, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = \\sum_ {i} \\lambda_ {i} (\\mathbf {x}) \\mathbf {v} _ {i} ^ {\\prime}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 828, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After deformation, $\\mathcal{T}$ provides a global displacement map for all Gaussians in $\\mathcal{G}_o$ . However, the internal rotation and", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scaling of the Gaussians cannot be accurately described by position-based interpolation, resulting in spiky artifacts. Additionally, the tetrahedral mesh is a coarse mesh. Its faces have significant gaps from the boundaries of the object, and it does not accurately describe the collision between $\\mathcal{G}_o$ and the plane. To address these issues, we establish a bidirectional local embedding within the Gaussian around the mean $\\mathbf{m}$ to precisely describe its influence range. This is determined by the endpoints of the ellipsoid corresponding to the Gaussian $[\\mathbf{v}_1,\\mathbf{v}_2,\\mathbf{v}_3]$ :", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {T} ^ {\\pm} = [ \\mathbf {m}, \\mathbf {v} _ {1}, \\mathbf {v} _ {2}, \\mathbf {v} _ {3} ] \\\\ = [ \\mathbf {m}, \\mathbf {m} \\pm \\gamma s _ {1} \\mathbf {r} _ {1}, \\mathbf {m} \\pm \\gamma s _ {2} \\mathbf {r} _ {2}, \\mathbf {m} \\pm \\gamma s _ {3} \\mathbf {r} _ {3} ], \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 253, + 903, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the relevant parameters can be obtained from the scaling matrix $\\mathbf{S} = \\mathrm{diag}(s_1, s_2, s_3)$ and the rotation Matrix $\\mathbf{R} = [\\mathbf{r}_1, \\mathbf{r}_2, \\mathbf{r}_3]$ . $\\gamma$ is a parameter that describes the range of influence, and we set its default value as 3.0. Each Gaussian has two directional embeddings, $\\mathbf{T}^{+}$ and $\\mathbf{T}^{-}$ , and each contains four points.", + "bbox": [ + 511, + 300, + 905, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design bidirectional local embeddings for several purposes. For collision processing, it is used to detect the extent to which an object collides with the environmental plane; and for segmentation, it detects whether the Gaussian is cut by the plane and establishes a loss term to penalize Gaussian kernels that intersect the plane under initial setting; For deformation, it is used to accurately describe the scaling and rotation of the deformed Gaussian kernel. Finally, based on the angle between two directions, it determines whether the deformed Gaussian kernel is sufficient to cover the radiance field.", + "bbox": [ + 511, + 390, + 905, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Soft Body Locomotion on Plane", + "text_level": 1, + "bbox": [ + 511, + 566, + 794, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core of our pipeline is an active locomotion solver for $\\mathcal{T}$ and $\\mathcal{G}_o$ . This is in the form of an inverse problem involving high-dimension inequalities constraints for processing frictional contacts between a 3DGS object and the environment.", + "bbox": [ + 511, + 589, + 905, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Muscle-Driven Dynamics", + "text_level": 1, + "bbox": [ + 511, + 672, + 736, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We embed muscle fibers within $\\mathcal{T}$ to serve as the primary drivers of motion. At each time step, the nodal positions of $N$ nodes in $\\mathcal{V}$ , denoted as $\\mathbf{p}$ , is determined through numerical integration of the dynamic equations of motion:", + "bbox": [ + 511, + 691, + 905, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} \\ddot {\\mathbf {p}} = \\mathbf {f} _ {e x t} + \\mathbf {f} _ {i n t} + \\mathbf {f} _ {d} + \\mathbf {f} _ {m} + \\mathbf {f} _ {c}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 765, + 903, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ is the mass matrix, $\\mathbf{f}_{ext},\\mathbf{f}_{int},\\mathbf{f}_d,\\mathbf{f}_c,\\mathbf{f}_m$ represent external, internal, damping, muscle, and contact forces, respectively. We use $(\\cdot)$ for velocity and $(\\cdot)$ for acceleration. The external force $\\mathbf{f}_{ext}$ includes e.g., the gravity force. The internal force $\\mathbf{f}_{int}$ , also known as the elastic force, is the resultant force generated by the material's elastic properties that resist deformation. In our implementation, this force", + "bbox": [ + 511, + 792, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "898", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cf42db0c67c03a6c9d9f2550c30d291b96a5824a96ffd9352abf68fae674a06a.jpg", + "image_caption": [ + "Figure 3. An overview of EnliveningGS pipeline. Our hybrid mesh-Gaussian representation plays an important role in all tasks within the pipeline. Our locomotion solver inversely computes the muscle activation embedded in the segmented 3DGS object that leads to user-specific motion." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 903, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is determined as the negative gradient of the Neo-Hookean energy [5]:", + "bbox": [ + 89, + 325, + 482, + 358 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi = \\frac {\\mu}{2} (I _ {C} - 3) - \\mu \\ln J + \\frac {\\lambda}{2} (\\ln J) ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 368, + 482, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mu$ and $\\lambda$ are the Lamé parameters. $I_{C} = \\operatorname{tr}(\\mathbf{F}^{T}\\mathbf{F})$ and $J = \\det(\\mathbf{F})$ are computed based on the deformation gradient $\\mathbf{F}$ . We use the Rayleigh damping for the damping force:", + "bbox": [ + 89, + 409, + 483, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {d} = \\left(\\alpha \\mathbf {M} + \\beta \\mathbf {K} _ {0}\\right) \\dot {\\mathbf {p}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 472, + 482, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{K}_0$ is the rest-shape stiffness matrix.", + "bbox": [ + 89, + 496, + 380, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider muscle fibers as polygonal curves with $M$ segments. The spring force caused by each segment is computed as: $f = ka$ where $k$ is the stiffness of the muscle fiber and $a$ denotes the change in the segment's length. When a muscle segment activates, the spring force will shorten or bend the body around it. Muscle force on the whole soft body is encoded in a pose-dependent activation matrix $\\mathbf{A} \\in \\mathbb{R}^{3N \\times M}$ :", + "bbox": [ + 89, + 511, + 483, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {m} = \\mathbf {A} (\\mathbf {p}) \\mathbf {a}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 633, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{a} \\in \\mathbb{R}^{M}$ is the vector of activations of all the muscle segments. Please refer to the supplementary materials for a complete description of the calculations involving activation matrix $\\mathbf{A}$ .", + "bbox": [ + 89, + 657, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide an automatic muscle generation tool given $\\mathcal{T}$ , which uses the medial axis transform [8, 29] of the model. Four longitudinal muscles are placed parallel near the medial axis at a given radius, and radial muscles are placed at given intervals. Users can also customize the muscle setup with preferred arrangements inside the object to achieve different movements. For instance, by implanting several longitudinal muscle fibers that run the length of the body, the body can be shortened when these muscles contract. Bending the body can be achieved by contracting these muscles unevenly on one side. Additionally, circular muscles wrap around the body, enabling it to twist, while", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "radial muscles extend across the body's cross-section, allowing the body to elongate while preserving its volume.", + "bbox": [ + 511, + 325, + 903, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Contacting 3DGS", + "text_level": 1, + "bbox": [ + 511, + 362, + 683, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section we present a novel contact model for 3DGS. For Gaussian kernels in $\\mathcal{G}_o$ , we first define $\\mathbf{u}$ as the position of points in bidirectional local embedding $\\mathbf{T}^{\\pm}$ below the plane as $\\mathbf{N}^T\\mathbf{u} + d\\leq 0$ . Non-interpenetration constraints on 3DGS is thus defined as inequalities:", + "bbox": [ + 511, + 381, + 905, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {n} (\\mathbf {u}) \\equiv \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} \\geq \\mathbf {0}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 464, + 903, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We decompose contact forces on the Gaussian as $\\mathbf{f}_g = \\mathbf{N}f_{\\perp} + \\mathbf{D}\\mathbf{f}_{\\parallel}$ , where $\\mathbf{D}$ is a set of directions at the contact point approximating the friction cone, and they are perpendicular to the plane normal $\\mathbf{N}$ . $f_{\\perp}$ and $\\mathbf{f}_{\\parallel}$ are the magnitudes of normal and tangent forces.", + "bbox": [ + 511, + 489, + 905, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LCP formulation [2] that regulates the contact velocity and contact force is:", + "bbox": [ + 511, + 566, + 905, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {0} \\leq \\left[ \\begin{array}{l} f _ {\\perp} \\\\ \\mathbf {f} _ {\\parallel} \\\\ \\lambda \\end{array} \\right] \\perp \\left[ \\begin{array}{c} \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} \\\\ \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} + \\mathbf {E} \\lambda \\\\ \\mu f _ {\\perp} - \\mathbf {E} ^ {T} \\mathbf {f} _ {\\parallel} \\end{array} \\right] \\geq \\mathbf {0}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 599, + 903, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ is the friction coefficient, and $\\lambda$ is an auxiliary variable (e.g., the Lagrange multiplier) that relates the tangent velocity to a sliding contact. $\\mathbf{E}$ is a block-diagonal matrix of $\\mathbf{e}$ . LCP condition notation $\\mathbf{x} \\perp \\mathbf{y}$ indicates $\\mathbf{x}^T\\mathbf{y} = 0$ . The first line of Equation (9) is a complementarity formulation of contact that enforces the non-penetration condition, and the following two lines encode Coulomb's friction law.", + "bbox": [ + 511, + 656, + 905, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Like [33], contact can be viewed as a regularization of the complementarity form of the non-penetration constraint, which gives penalty forms of contact as spring force $\\mathbf{f}_g^{spring}$ :", + "bbox": [ + 511, + 763, + 905, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\perp} ^ {s p r i n g} \\equiv - k _ {n} \\min \\left(0, C _ {n} (\\mathbf {u})\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 835, + 903, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\parallel} ^ {s p r i n g} \\equiv - \\min \\left(k _ {f}, \\mu \\frac {\\| \\mathbf {f} ^ {s p r i n g} \\|}{\\| \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} \\|}\\right) \\mathbf {D} ^ {T} \\dot {\\mathbf {u}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 869, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "899", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k_{n}$ and $k_{f}$ control the stiffness of the contact. Forces on local embeddings can be converted back onto points on $\\mathcal{T}$ using the same weights in Equation (2), giving us $\\mathbf{f}_c$ or $\\mathbf{f}_c^{spring}$ .", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Two-stage Locomotion", + "text_level": 1, + "bbox": [ + 89, + 156, + 294, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike force-based passive dynamics, we achieve the desired locomotion mainly by intuitively prescribing the motion trajectory of the object. For the position of a target point $\\mathbf{p}^{\\star}$ , the objective function that constrains the position can be written as:", + "bbox": [ + 89, + 175, + 483, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\text {p o s i t i o n}} = \\left\\| f (\\mathbf {p}) - \\mathbf {p} ^ {\\star} \\right\\| ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 255, + 482, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f$ can be functions that compute the positional target of the object. For instance, a commonly used setup is $f_{COM}$ , which calculates the COM (center of mass) $\\mathbf{c} = f_{COM}(\\mathbf{p})$ of the body, of local COM coordinates (foot, head, base, etc.). Similar objective functions can be used to track the velocity, acceleration, angular momentum, base area, etc.", + "bbox": [ + 89, + 280, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first solve Equation (4) by substituting $\\mathbf{f}_c$ with $\\mathbf{f}_c^{spring}$ and temporarily keeping muscle activation a same as the last time step. Based on the resulting $\\dot{\\mathbf{p}}$ , we convert the state of contact points back into the inequalities in Equation (9), and solve a Quadratic Programming (QP) problem to find the best $\\mathbf{a}$ . For example, if a contact point is in static contact, the corresponding QP is:", + "bbox": [ + 89, + 386, + 483, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {a}, f _ {\\perp}, \\mathbf {f} _ {| |}, \\lambda} G \\left(\\mathbf {a}, f _ {\\perp}, \\mathbf {f} _ {| |}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 497, + 334, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "subject to", + "bbox": [ + 183, + 526, + 253, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {0} \\leq f _ {\\perp}, \\quad \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} ^ {n + 1} = \\mathbf {0} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 544, + 482, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {0} \\leq \\mathbf {f} _ {\\parallel}, \\quad \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} ^ {n + 1} + \\mathbf {E} \\lambda = \\mathbf {0}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 565, + 388, + 584 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {0} = \\lambda , \\quad \\mu f _ {\\perp} - \\mathbf {E} ^ {T} \\mathbf {f} _ {\\parallel} \\geq \\mathbf {0}.\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 587, + 372, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that all contact points may have their own states. If the results are not satisfactory, we will use the pivoting method as in [41] to find the appropriate contact states.", + "bbox": [ + 89, + 609, + 483, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Gaussian Deformation in Tetrahedral Mesh", + "text_level": 1, + "bbox": [ + 89, + 662, + 460, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The barycentric coordinates of the points in $\\mathbf{T}^{+}$ are calculated in the reference configuration, and the deformed coordinates in the current configuration $\\mathbf{T}_s^+ = [\\mathbf{m}',\\mathbf{v}_1',\\mathbf{v}_2',\\mathbf{v}_3']$ are then calculated at each timestep to describe the deformation.", + "bbox": [ + 89, + 684, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After that, we construct two matrices $\\mathbf{D}_m$ and $\\mathbf{D}_s$ , representing the differences in vertex coordinates in the reference and current configurations, respectively:", + "bbox": [ + 89, + 760, + 483, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r c l} \\mathbf {D} _ {m} & = & \\left[ \\mathbf {v} _ {1} ^ {0} - \\mathbf {m}, \\mathbf {v} _ {2} ^ {0} - \\mathbf {m}, \\mathbf {v} _ {3} ^ {0} - \\mathbf {m} \\right] \\\\ \\mathbf {D} _ {m} & = & \\left[ \\mathbf {v} _ {1} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\mathbf {v} _ {2} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\mathbf {v} _ {3} ^ {\\prime} - \\mathbf {m} ^ {\\prime} \\right] \\end{array} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 810, + 482, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {s} = \\left[ \\begin{array}{c} \\mathbf {v} _ {1} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\bar {\\mathbf {v}} _ {2} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\bar {\\mathbf {v}} _ {3} ^ {\\prime} - \\mathbf {m} ^ {\\prime} \\end{array} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 825, + 411, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The linear transformation $\\mathbf{F}$ that maps the reference configuration to the current configuration can be calculated as:", + "bbox": [ + 89, + 848, + 482, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\mathbf {D} _ {s} \\cdot \\mathbf {D} _ {m} ^ {- 1}.\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 883, + 339, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The deformed covariance becomes: $\\Sigma^{\\prime} = \\mathbf{F}\\Sigma \\mathbf{F}^{T}$ and is converted into scaling and rotation by performing singular value decomposition (SVD) on $\\Sigma^{\\prime}$ :", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {+} = \\operatorname {d i a g} \\left(\\sqrt {\\operatorname {S V D} _ {\\Sigma} \\left(\\Sigma^ {\\prime}\\right)}\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 147, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} ^ {+} = \\mathsf {S V D} _ {U} (\\Sigma^ {\\prime}).\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 162, + 756, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that if $\\operatorname{det}(\\mathbf{R}^{+}) = -1$ , one needs to negate its last column to ensure it is a valid rotation matrix in $SO(3)$ . It is also necessary to adjust the spherical harmonics' orientation by applying rotation $R_{SH} = \\mathsf{SVD}_U(\\Sigma')\\mathsf{SVD}_V(\\Sigma')^T$ to display the correct colors from various angles.", + "bbox": [ + 511, + 191, + 905, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The points in local embeddings represent the shape of the deformed Gaussian kernel but they do not necessarily belong to the same tetrahedron. For some larger Gaussian kernels, significant rotations may cause them to extend beyond the original radiance field's boundaries. To avoid this situation, we use the reverse embedding $\\mathbf{T}^{-}$ and use the aforementioned method to obtain its corresponding rotation matrix $\\mathbf{R}^{-}$ . The angle between $\\mathbf{R}^{+}$ and $\\mathbf{R}^{-}$ can be expressed as:", + "bbox": [ + 511, + 267, + 906, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\arccos \\left(\\frac {\\operatorname {t r} \\left(\\mathbf {R} ^ {+ T} \\mathbf {R} ^ {-}\\right) - 1}{2}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 412, + 903, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If $\\theta$ exceeds the threshold $\\theta >\\eta$ , it indicates that the region covered by the Gaussian has transformed from a plane to a curved surface. A single kernel can no longer represent the deformed area accurately; thus, this Gaussian will be split into two from the center. Figure 2 (b) provides a schematic of our splitting process, while Figure 8 presents the corresponding experimental results.", + "bbox": [ + 511, + 463, + 905, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [9, 15], the splitting of a Gaussian $(\\alpha_0,\\mu_0,\\Sigma_0)$ into two new Gaussian kernels $(\\alpha_{l},\\mu_{l},\\Sigma_{l})$ and $(\\alpha_{r},\\mu_{r},\\Sigma_{r})$ at the center is calculated by:", + "bbox": [ + 513, + 569, + 905, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {l} = \\alpha_ {r} = \\frac {\\alpha_ {0}}{2},\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 625, + 655, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {l} = \\mu_ {0} - \\frac {\\mathbf {v} _ {k} - \\mu_ {0}}{\\kappa}, \\quad \\mu_ {r} = \\mu_ {0} + \\frac {\\mathbf {v} _ {k} - \\mu_ {0}}{\\kappa}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 652, + 903, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {l} = \\boldsymbol {\\Sigma} _ {r} = \\boldsymbol {\\Sigma} _ {0} - \\frac {(\\mathbf {v} _ {k} - \\mu_ {0}) (\\mathbf {v} _ {k} - \\mu_ {0}) ^ {T}}{\\kappa^ {2}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 681, + 825, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the above, $\\mathbf{v}_k \\in \\mathbf{T}^+$ stands for the endpoint on main principle axis with the largest scaling factor $s_k$ . Factor $\\kappa = \\frac{2\\gamma}{\\sqrt{2\\pi}}$ is derived by maximizing the similarity between each new Gaussian and their own part of the original. We refer readers to supplementary material for the complete derivation.", + "bbox": [ + 511, + 723, + 905, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Plane Estimation", + "text_level": 1, + "bbox": [ + 511, + 809, + 681, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To find the parameters $\\mathbf{N}$ and $d$ of a plane in a 3DGS scene, we first construct a depth map on the plane:", + "bbox": [ + 511, + 830, + 903, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {p} (u, v) = - \\frac {\\mathbf {N} \\cdot \\mathbf {t} _ {c} + d}{\\mathbf {n} ^ {T} \\mathbf {R} _ {c} ^ {T} \\mathbf {K} ^ {- 1} \\mathbf {p} (u , v)}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 871, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "900", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}(u,v)$ represents the pixel coordinates in homogeneous form $[u,v,1]^T$ , $\\mathbf{R}_c$ and $\\mathbf{t}_c$ are the rotation and translation extracted from the camera's extrinsic matrix, and $\\mathbf{K}$ denotes the camera's intrinsic matrix. Initially, $\\mathbf{N}$ is set to camera's look-at direction and $d$ is set to zero to ensure solvability.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By minimizing the error between the depth map within the region and the GS depth map $\\mathcal{D}_g$ , the plane parameters can be optimized by:", + "bbox": [ + 89, + 181, + 480, + 227 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\mathbf {N}, d} \\sum_ {(u, v) \\in \\mathcal {M}} \\| \\mathcal {D} _ {p} (u, v) - \\mathcal {D} _ {g} (u, v) \\|, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 233, + 482, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is a mask of the planar region which can be obtained through image-based segmentation algorithms [28] or roughly drawn by the user.", + "bbox": [ + 89, + 268, + 480, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The user can repeat this step on multiple images and optimize simultaneously to obtain fine-tuned plane parameters. In practice, a single viewpoint is usually sufficient to estimate an accurate plane.", + "bbox": [ + 89, + 315, + 480, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Object Segmentation with Plane Priors", + "text_level": 1, + "bbox": [ + 89, + 383, + 426, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We categorize the Gaussian scene into two types: dynamic splitting and static splitting. Equation (17) showcases dynamic online splitting to enhance the quality during the deformation process and ensure the speed of splitting on animating objects. For a Gaussian model, since the segmentation plane of the object and environment is already determined, it is more suitable to incorporate the plane information during training to regulate the Gaussian parameters at the cutting plane.", + "bbox": [ + 89, + 405, + 482, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To perform static splitting on the training stage, we first determine whether the Gaussian intersects with the plane. This requires checking if all the endpoints in $\\mathbf{T}^{\\pm}$ are on the same side of the plane. We project center-to-endpoints distance onto the plane normal:", + "bbox": [ + 89, + 541, + 482, + 614 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\eta_ {s e g} = \\max _ {\\mathbf {v} _ {i} \\in \\mathbf {T} ^ {\\pm}} | \\mathbf {N} ^ {T} (\\mathbf {v} _ {i} - \\mathbf {m}) |. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 623, + 480, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If the distance from Gaussian kernel to the plane $|d_0| = |P(\\mathbf{n}, d)(\\mathbf{m})| < \\eta_{seg}$ , we will constrain the Gaussian's parameter or split the Gaussian with the plane.", + "bbox": [ + 89, + 651, + 482, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We fine-tune the trained Gaussian model by introducing the following non-penetration regularization term to ensure the model's influence range does not cross the plane:", + "bbox": [ + 89, + 696, + 482, + 742 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e g} = \\left| \\left| \\eta_ {s e g} - \\left| d _ {0} \\right| \\right| \\right| _ {2} \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 750, + 480, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.7. Inpainting 3D Object on GS", + "text_level": 1, + "bbox": [ + 89, + 772, + 341, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After extracting the object $\\mathcal{G}_o$ , a hole is left in the background $\\mathcal{G}_{\\text {scene }}$ . To address this issue, we introduce the depth map of the plane in Equation (18) into the depth-based inpainting method [30] and use a 2D inpainting tool [40] to obtain the color of the missing parts from the current viewpoint. Subsequently, we fine-tune the filled patch using the inpainted single-view image.", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 643, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implemented our pipeline on a workstation with a single NVIDIA RTX4090 GPU and Intel 14900K CPU. We used PyTorch for storing and editing 3DGS on GPU, and OSQP [37] for solving QP in Equation (13). Our method supports the original 3DGS implementation [25]. For better visual quality on the planar surface, we set the minimum scaling factor on Gaussian kernels to $\\epsilon = e^{-10}$ , similar to 2DGS[19]. More benchmarks for the locomotion examples are reported in the supplementary document.", + "bbox": [ + 511, + 114, + 903, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Locomotion Design", + "text_level": 1, + "bbox": [ + 511, + 260, + 697, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Walking. We achieve walking by controlling the trajectories of all four legs. The diagonal support legs simultaneously lift the body, while the remaining two legs swing forward with the body's movement. As shown in Figure 4, the wooden stool moves one position to the right by taking four steps.", + "bbox": [ + 511, + 281, + 903, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Jumping. Jumping soft-bodied characters are more commonly seen in cartoons and animations. Jumping is a complex action because the contact state between the object and the ground changes rapidly and significantly. To achieve continuous jumping, it is necessary to precisely control the position of the COM and the relative velocity of the center of the base to jump the appropriate distance and maintain balance. We have designed an animation of a game of chess, where the soft-bodied chess pieces move across the board by continuously jumping, as shown in Figure 1.", + "bbox": [ + 511, + 378, + 905, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Twisting. Figure 5 illustrates an example of the Peashooter doll twisting, jumping, and rotating 180 degrees. Based on the jumping motion, we additionally control the change in angular momentum along the central axis, allowing the Peashooter to rotate its body while airborne and face the zombie upon landing.", + "bbox": [ + 511, + 536, + 905, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison I. Deformation on GS", + "text_level": 1, + "bbox": [ + 511, + 635, + 813, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the current state-of-the-art physico-aware GS methods. Figure 6 shows a qualitative comparison, where our method demonstrates better results under large deformation. Table 1 presents the related quantitative evaluation: we apply bending, twisting, and falling on the ground deformations to the ficus, chair, and mic examples from the NeRF synthetic dataset [34], and compare them with the rendered images of the source 3D mesh subjected to the same deformations as the ground truth.", + "bbox": [ + 511, + 657, + 905, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison II. Two-stage Locomotion", + "text_level": 1, + "bbox": [ + 511, + 803, + 844, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our two-stage locomotion solver, we conducted frame-by-frame tests on an animation sequence with given muscle activation. This animation sequence includes 50 frames, 60 variables, and 50 pairs of linear complementarity constraints. We compared our solution with a QP", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "901", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be7f40502de4eb5fd1d0cf95dfa904018e5ff900e922910a6d7994978e6acc9f.jpg", + "image_caption": [ + "Figure 4. Walking stool. In the bedroom scene, a wooden stool walks like a quadruped." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4de9d7892211b4138d355b7a33f08eaf5c94581ddee23317007830d4ec57e737.jpg", + "image_caption": [ + "Figure 5. Peashooter vs. Zombie. On a cloth-covered coffee table, a Peashooter figurine contracts, jumps, and twists its body, landing after a 180-degree spin to face the zombie." + ], + "image_footnote": [], + "bbox": [ + 91, + 207, + 903, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/526b1ceff8a22c32a277fca95150967414931650572624af579c3b89cc489d4d.jpg", + "image_caption": [ + "Figure 6. Visual quality comparison on deforming objects. Compared to PhysGaussian [45] and VR-GS [23], our method can still suppress spiky artifacts under large deformations." + ], + "image_footnote": [], + "bbox": [ + 96, + 347, + 477, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "solver based directly on the static contact assumption and the QPCC solver proposed by [41], and reported the average best and worst results across all frames in Table 2. Our locomotion solver uses the results predicted by the penalty force as guidance, reducing the number of iterations and lowering the objective value.", + "bbox": [ + 89, + 625, + 483, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison III. Inpainting", + "text_level": 1, + "bbox": [ + 89, + 727, + 338, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 7, we demonstrate the inpainting effect on a plane using the garden example from Mip-NeRF360 dataset [4] and the truck example from DTU dataset [22], comparing it with the existing 3D object inpainting method on GS. In both examples, the removed objects occupy a significant portion of the photo, leaving a large hole on the ground after removal. Existing methods, due to the lack of relevant depth information, struggle to fill such a large missing area even after extensive finetuning of the inpainted 2D image (approximately 10,000 iterations). In contrast,", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b8bf8c7662b9b530f14806fe8a7a93253f5c213cfa2eaa3352c52825e63b165.jpg", + "image_caption": [ + "Figure 7. Large 3D object inpainting. Comparison on large 3D object inpainting cases on the ground, where Gaussian Grouping [48] requires 20 minutes fine-tuning but still fails to fill the large hole on the ground while our method with better inpainting quality only needs 20 seconds fine-tuning. We show the render result as well as the depth." + ], + "image_footnote": [], + "bbox": [ + 514, + 347, + 594, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee8c7d348c3552bc7c94b808a2e882413ffcf21a131b20b3e2d78ce1ff39ead0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 347, + 750, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f68b70a34949f5b6dd57f03f444a35da0bb0abdf2483d18eb41cb27ac41d958.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 347, + 905, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method leverages the plane as a prior, placing a dense and reasonably positioned point cloud in the missing area. This approach requires only about 100 iterations of finetuning, speeding up the inpainting process by several orders of magnitude and remaining unaffected by the size of the missing region.", + "bbox": [ + 511, + 564, + 906, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 662, + 679, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bidirectional local embedding We first use a checkered bar to showcase the effect of bidirectional local embedding as in Figure 8. The checkered bar twists increasingly from right to left, and the sparky artifacts become progressively more pronounced. Without the need to detect whether the Gaussian kernel exceeds the object's surface, our bidirectional local embedding automatically identifies the regions causing sparky artifacts based on the information within the Gaussian kernel. Table 1 also reports quantitative results.", + "bbox": [ + 511, + 685, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Non-penetration regularization Since the segmentation plane under $\\mathcal{G}_o$ is occluded from any viewpoint, directly masking out the object from the environment results in significant blurring and penetration at the segmentation plane.", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "902", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/601f37ed30a617aa560b2e8a37a0a45bb3d50141d48f85ad464406f2d8f5c2a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BendingTwistingFalling
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
SuGaR [14]28.040.8900.04626.370.8230.10125.010.9130.053
PhysGaussian [45]28.140.9070.04326.530.8560.08925.100.9230.048
VR-GS [23]28.630.9170.03826.530.8930.06625.310.9380.044
Ours (w/o split)28.620.9180.03826.470.8910.06225.340.9390.042
Ours28.790.9230.03126.890.9100.04125.670.9510.032
", + "bbox": [ + 129, + 88, + 861, + 207 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1c8f94ea0c96bbb69305ec05d38adaec5832d5478e969412959beb7dc9e10a0.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of rendering quality after deformation on the NeRF synthetic dataset." + ], + "table_footnote": [], + "table_body": "
Iterations ↓Object value ↓
Avg.BestWorstAvg.BestWorst
QP---9.430.1623.86
QPCC [41]46.1915101.290.1412.11
Ours17.0113820.830.137.89
", + "bbox": [ + 96, + 243, + 470, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/270088c2abe18f4b67809c5d0b707df7d4df7998780e0056c50e942458b20af4.jpg", + "image_caption": [ + "Rest shape" + ], + "image_footnote": [], + "bbox": [ + 96, + 385, + 272, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c683898487a24bf74613ecd0e6983c757d8ac416af434b0c507523b8a3d3cae2.jpg", + "image_caption": [ + "w/o split" + ], + "image_footnote": [], + "bbox": [ + 300, + 377, + 477, + 434 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b13452fc052c610c093f662b7204b7ef682ac79110c38e6e4607fd66bc15e2c.jpg", + "image_caption": [ + "Detected artifacts", + "Figure 8. Ablation study on bidirectional local embedding. Our method captures the regions that produce spiky artifacts and removes them by splitting the Gaussian kernels to fit the radiance field." + ], + "image_footnote": [], + "bbox": [ + 96, + 446, + 272, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ebd8957b8e9b6b23abec4c0da4d908c03474664a8252488c40cbe48fead5fa7.jpg", + "image_caption": [ + "w/split" + ], + "image_footnote": [], + "bbox": [ + 299, + 449, + 475, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d9ddf6deff199238fea0af31ab12d22e1370ff0d66196f1586900581c5e3b28c.jpg", + "table_caption": [ + "Table 2. Numerical experiments of locomotion solvers." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓Residual ↓
Mask (w/o Lseg)---0.058
Remove25.9170.8270.211-
w/ Lseg35.2910.9450.0340.002
", + "bbox": [ + 91, + 584, + 480, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on non-penetration regularization: quantitative results.", + "bbox": [ + 89, + 667, + 482, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Gaussian kernel that penetrates the plane contributes to the object's appearance, so a direct Remove operation would create holes in the object. Figure 9 shows the relevant qualitative comparison results. Our non-penetration regularization restricts the Gaussian from penetrating the plane, maintaining the integrity of the object's surface. In the quantitative experiments shown in Table 3, we placed the camera on the segmentation plane, compared the image quality of the object's upper half with full mask-out, and evaluated the residual of the lower half. The residuals are expressed as the ratio of the sum of pixel values in the lower half to those in the upper half.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/474bef880366a2ecc61aec97330eaefc99bd5a880e08a0e5e64680f38fda7c2f.jpg", + "image_caption": [ + "Figure 9. Ablation study on non-penetration regularization: qualitative results. With $L_{seg}$ applied to Gaussian kernels penetrating the plane, the object retains its complete appearance near the cutting plane while removing the parts below the cutting plane." + ], + "image_footnote": [], + "bbox": [ + 517, + 243, + 901, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 510, + 632, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present EnliveningGS, a novel pipeline that enables active locomotion for 3D models represented by 3DGS. We use a hybrid mesh embedding of 3DGS to better estimate the contact between a 3DGS model and the environment. Given a high-level motion goal, we find the optimal muscle activations driving the movement of the object's body. An LCP-penalty mixed optimizer is used to efficiently calculate the contact and friction force.", + "bbox": [ + 511, + 529, + 905, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation Currently, our method relies on plane-based simplification of the environment. When the surroundings consist of more complicated geometries, users may need to use more sophisticated approximation strategies to estimate the contact plane. It is also possible to use triangulated surfaces. Doint so could impose more computational overhead for the contact solver.", + "bbox": [ + 511, + 662, + 905, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements The authors would like to thank the reviewers for their insightful comments. Yin Yang is partially supported by NSF 2301040. This work is also supported by NSF China (No. 62322209 and No. 62421003), the gift from Adobe Research, the XPLORER PRIZE, and the 100 Talents Program of Zhejiang University. The source code and data are available at https://gapszju.github.io/EnliveningGS.", + "bbox": [ + 511, + 773, + 905, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "903", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sheldon Andrews, Kenny Erleben, and Zachary Ferguson. Contact and friction simulation for computer graphics. In ACM SIGGRAPH 2022 Courses, pages 1-172, Vancouver British Columbia Canada, 2022. ACM. 2", + "[2] Mihai Anitescu and Florian A Potra. Formulating dynamic multi-rigid-body contact problems with friction as solvable linear complementarity problems. Nonlinear Dynamics, 14: 231-247, 1997. 2, 4", + "[3] Lijie Bai, John E Mitchell, and Jong-Shi Pang. On convex quadratic programs with linear complementarity constraints. Computational Optimization and Applications, 54(3):517-554, 2013. 2", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7", + "[5] Javier Bonet and Richard D Wood. Nonlinear continuum mechanics for finite element analysis. Cambridge university press, 1997. 4", + "[6] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physico-based simulations. arXiv preprint arXiv:2409.05819, 2024. 2", + "[7] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2", + "[8] Hyeong In Choi, Sung Woo Choi, and Hwan Pyo Moon. Mathematical theory of medial axis transform. *pacific journal of mathematics*, 181(1):57-88, 1997. 4", + "[9] Qiyuan Feng, Gengchen Cao, Haoxiang Chen, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. A new split algorithm for 3d gaussian splatting. arXiv preprint arXiv:2403.09143, 2024.5", + "[10] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Dynamic fluid synthesis with gaussian splatting. arXiv preprint arXiv:2401.15318, 2024. 1", + "[11] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 1", + "[12] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based Gaussian Splatting for Real-time Large-scale Deformation. arXiv preprint arXiv:2402.04796, 2024. 2", + "[13] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. arXiv preprint arXiv:2403.14554, 2024. 2", + "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 8", + "[15] Peter Hall, David Marshall, and Ralph Martin. Adding and subtracting eigenspaces with eigenvalue decomposition and singular value decomposition. Image and vision computing, 20(13-14):1009-1016, 2002. 5", + "[16] Fursan Hamad, Shreyas Giridharan, and Christian Moormann. A penalty function method for modelling frictional contact in mpm. Procedia Engineering, 175:116-123, 2017. 2", + "[17] David Harmon, Etienne Vouga, Breannan Smith, Rasmus Tamstorf, and Eitan Grinspun. Asynchronous contact mechanics. In ACM SIGGRAPH 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery.", + "[18] Shoichi Hasegawa and Nobuaki Fujii. Real-time rigid body simulation based on volumetric penalty method. In 11th Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems, 2003. HAPTICS 2003. Proceedings., pages 326-332. IEEE, 2003. 2", + "[19] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers* '24, pages 1-11, 2024. 6", + "[20] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4220-4230, 2024. 2", + "[21] Sumit Jain, Yuting Ye, and C Karen Liu. Optimization-based interactive motion synthesis. ACM Transactions on Graphics (TOG), 28(1):1-12, 2009. 2", + "[22] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanaes. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7", + "[23] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, and Chenfanfu Jiang. VR-GS: A Physical Dynamics-Aware Interactive Gaussian Splitting System in Virtual Reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 2, 7, 8", + "[24] Joaquim J. Judice and Fernanda M Pires. A block principal pivoting algorithm for large-scale strictly monotone linear complementarity problems. Computers & operations research, 21(5):587-596, 1994. 2", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 6", + "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1", + "[27] Junggon Kim and Nancy S Pollard. Direct control of simulated nonhuman characters. IEEE Computer Graphics and Applications, 31(4):56-65, 2011. 2" + ], + "bbox": [ + 517, + 93, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "904", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6", + "[29] Lei Lan, Ran Luo, Marco Fratarcangeli, Weiwei Xu, Huamin Wang, Xiaohu Guo, Junfeng Yao, and Yin Yang. Medial Elastics: Efficient and Collision-Ready Deformation via Medial Axis Transform. ACM Transactions on Graphics, 39(3): 1-17, 2020. 4", + "[30] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. InFusion: Inpainting 3D Gaussians via Learning Depth Completion from Diffusion Prior. arXiv preprint arXiv:2404.11613, 2024. 2, 6", + "[31] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2", + "[32] Miles Macklin, Matthias Müller, and Nuttapong Chentanez. XPBD: Position-based simulation of compliant constrained dynamics. In Proceedings of the 9th International Conference on Motion in Games, pages 49-54, New York, NY, USA, 2016. Association for Computing Machinery. 2", + "[33] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Tae-Yong Kim. Primal/dual descent methods for dynamics. In Computer Graphics Forum, pages 89-100. Wiley Online Library, 2020. 4", + "[34] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1, 6", + "[35] Todd S Munson, Francisco Facchinei, Michael C Ferris, Andreas Fischer, and Christian Kanzow. The semismooth algorithm for large scale complementarity problems. INFORMS Journal on Computing, 13(4):294-311, 2001. 2", + "[36] Andrei-Ionut Stefancu, Silviu-Cristian Melenciuc, and Mihai Budescu. Penalty based algorithms for frictional contact problems. Buletinul Institutului Politehnic din lasi. Sectia Constructii, Architectura, 57(3):119, 2011. 2", + "[37] B. Stellato, G. Banjac, P. Goulart, A. Temporad, and S. Boyd. OSQP: an operator splitting solver for quadratic programs. Mathematical Programming Computation, 12(4): 637-672, 2020. 6", + "[38] D. E. Stewart and J. C. Trinkle. An Implicit Time-Stepping Scheme for Rigid Body Dynamics with Inelastic Collisions and Coulomb Friction. International Journal for Numerical Methods in Engineering, 39(15):2673-2691, 1996. 2", + "[39] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 1", + "[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6", + "[41] Jie Tan, Greg Turk, and C. Karen Liu. Soft body locomotion. ACM Transactions on Graphics, 31(4):26:1-26:11, 2012. 2, 5, 7, 8", + "[42] Min Tang, Dinesh Manocha, Miguel A Otaduy, and Ruofeng Tong. Continuous penalty forces. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 2", + "[43] Alessandro Tasora, Dario Mangoni, Simone Benatti, and Rinaldo Garziera. Solving variational inequalities and cone complementarity problems in nonsmooth dynamics using the alternating direction method of multipliers. International Journal for Numerical Methods in Engineering, 122(16): 4093-4113, 2021. 2", + "[44] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. GaMeS: Mesh-Based Adapting and Modification of Gaussian Splatting. arXiv preprint arXiv:2402.01459, 2024. 2", + "[45] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4389-4398, 2024. 1, 2, 3, 7, 8", + "[46] Hongyi Xu, Yili Zhao, and Jernej Barbič. Implicit multibody penalty-based distributed contact. IEEE transactions on visualization and computer graphics, 20(9):1266-1279, 2014. 2", + "[47] Shuojue Yang, Qian Li, Daiyun Shen, Bingchen Gong, Qi Dou, and Yueming Jin. Deform3dgs: Flexible deformation for fast surgical scene reconstruction with gaussian splattering. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 132-142. Springer, 2024. 2", + "[48] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. arXiv preprint arXiv:2312.00732, 2024. 2, 7", + "[49] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2024. 2", + "[50] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 726 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "905", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_model.json b/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3da2adcf439f35a3986ca2e6510b8cd18ce1d76e --- /dev/null +++ b/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_model.json @@ -0,0 +1,2574 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.28, + 0.131, + 0.72, + 0.152 + ], + "angle": 0, + "content": "EnliveningGS: Active Locomotion of 3DGS" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.18, + 0.848, + 0.217 + ], + "angle": 0, + "content": "Siyuan Shen\\(^{1}\\) Tianjia Shao\\(^{1*}\\) Kun Zhou\\(^{1}\\) Chenfanfu Jiang\\(^{2}\\) Yin Yang\\(^{3}\\) \n\\(^{1}\\) State Key Laboratory of CAD&CG, Zhejiang University, \\(^{2}\\) UCLA, \\(^{3}\\) University of Utah" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.251, + 0.327, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.283, + 0.486, + 0.645 + ], + "angle": 0, + "content": "This paper presents a novel pipeline named EnliveningGS, which enables active locomotion of 3D models represented with 3D Gaussian splatting (3DGS). We are inspired by the fact that real-world lives pose their bodies in a natural and physically meaningful manner by compressing or elongating muscle fibers embedded in the body. EnliveningGS aims to replicate the similar functionality of 3DGS models so that the object within a 3DGS scene acts like a living creature rather than a static shape — they walk, jump, and twist in the scene under provided motion trajectories driven by muscle activations. While the concept is straightforward, many challenging technical difficulties need to be taken care of. Synthesizing realistic locomotion of a 3DGS model embodies an inverse physics problem of very high dimensions. The core challenge is how to efficiently and robustly model frictional contacts between an “enlivened model” and the environment, as it is the composition of contact/collision/friction forces triggered by muscle activation that generates the final movement of the object. We propose a hybrid numerical method mixing LCP and penalty method to tackle this NP-hard problem robustly. Our pipeline also addresses the limitation of existing 3DGS deformation algorithms and inpainting the missing information when models move around." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.222, + 0.688 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.484, + 0.88 + ], + "angle": 0, + "content": "The advent of learning-based radiance field rendering techniques like NeRF [34] and 3D Gaussian splitting (3DGS) [26] have rebranded the pipeline of novel view synthesis. From synthesized RGB information, users can further alter and edit the illumination, appearance, and geometry. It is also possible to integrate those neural scene representations with physics to synthesize plausible physics-based dynamics. For instance, PIE-NeRF uses a quadratic moving least square interpolation to approximate the nonlinear elastodynamics of a NeRF model [11]. PhysGaussian [45] integrates 3DGS with the material point method (MPM) [39]. GaussianSplashing leverages a universal" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.253, + 0.903, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.45, + 0.907, + 0.506 + ], + "angle": 0, + "content": "Figure 1. Jumping Chess. We use static 3DGS as input and generates physics-aware 4D animations. The chess piece contracts and expands its body, creating jumping motions through collisions and friction with the chessboard, fitting the user's input trajectory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.594 + ], + "angle": 0, + "content": "particle-based discretization to animate 3DGS scenes with fluids [10]. Those frameworks allow deeper integrations of 3DGS with relevant downstream graphics/vision tasks, enabling a wide range of exciting applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.883 + ], + "angle": 0, + "content": "It is a common paradigm, when combining physics and 3DGS, that the physical effects of the neural scene are passively triggered by external forces, which are normally specified by the user or pre-scripted (e.g., the gravity force). This makes sense for inanimate objects — they are not changing their kinematic states unless forced by external excitements. However, the locomotion of a living being follows a distinct modality. Rather than relying on external forces passively, a living entity actively adjusts its poses to achieve a desired movement such as walking, bending, or jumping. The enabler of such locomotion is muscle activations that contract or elongate different parts of the body, which interact with the environment e.g., the floor. The resultant contact and friction forces, in turn, provide the necessary momentum of the motion. This procedure is beyond the capability of existing physics-based 3DGS techniques as the muscle activations can hardly be measured or synthesized with a passive physics simulator. It is unfortunate that an interesting 3DGS scene can only house lifeless objects." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In this paper, we show a novel framework to further en" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.246, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "896" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.392 + ], + "angle": 0, + "content": "hance the 3DGS expressivity. We name our framework EnliveningGS, which inversely computes muscle activations embedded in a 3DGS model that lead to user-specific motion. As shown in Figure 1, the user only needs to provide a desired trajectory of the object (or other high-level descriptions of the kinematics). Given a set of embedded muscle fibers (approximated by piece-wise linear springs), EnliveningGS segments the object from the scene, estimates the possible contact/friction forces between the object and its surrounding environment, and calculates the optimal muscle activations to drive the corresponding body motion so that the object behaves like a living creature and moves following the prescribed trajectories as much as possible. As contact modeling is known to be an NP-hard problem and quickly becomes unsolvable when there are a moderate number of contact incidents, EnliveningGS introduces a two-stage locomotion procedure to reduce the solution space and efficiently handle frictional contacts between the 3DGS model and the scene. To summarize, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.4, + 0.482, + 0.536 + ], + "angle": 0, + "content": "- Active Animatable 3DGS Pipeline: We present En-livingGS pipeline that brings any household object to life. When they walk or jump, occluded areas become visible again. To overcome such visual artifacts, we propose an improved inpainting algorithm to fill colors and textures back to those exposed areas based on the known prior of the environment. We also present a self-splitting mechanism for Gaussian kernels to capture and reduce spiky artifacts under significant deformation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.537, + 0.482, + 0.656 + ], + "angle": 0, + "content": "- Hybrid Contact Modeling Framework for 3DGS: We design a novel hybrid numerical procedure combing the penalty method, which softens the hard contact constraint, and the classic Linear Complementarity Programming (LCP) [38] so that the combinatorial space of the active set can be significantly reduced, and the two-way coupling with the underlying nonlinear physics simulation becomes feasible." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.4, + 0.482, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.677, + 0.233, + 0.693 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3D Gaussian editing. Recently, 3D Gaussian splatting (GS) [25] utilizes a set of 3D Gaussian kernels to represent static scenes explicitly and achieve state-of-the-art results. Subsequently, editing on GS has become a popular topic[7, 13, 30, 31, 47, 48, 50]. To make static 3D scenes deformable, some works extract explicit high-quality meshes from 3DGS representations and align GS kernels with mesh surfaces for guidance when deformed [12, 14, 44]. Their reliance on strong topological information and high-quality meshes makes it difficult to balance quality and interactivity. SC-GS [20] learns sparse control points for 3D scene dynamics but faces challenges with large movements and complex deformation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Physics-aware dynamics on 3D GS. To combine a static GS scene with a physics engine to generate dynamics on GS, PhysGaussian[45] leverages the first-order information from the displacement map (deformation gradient) to assist dynamic simulations. VR-GS [23] constructed a tetrahedral cage for each segmented Gaussian kernel group and embedded these kernel groups into corresponding meshes, allowing the deformed mesh driven by eXtended Position-based Dynamics (XPBD) [32] to guide the deformation of the Gaussian kernels. GASP[6] treats the simulation process as a black box and integrates Newtonian dynamics into the GaMeS[44] framework. PhysDreamer [49] distills dynamics priors learned by video generation models. It is worth mentioning that existing methods consider the simulation process as a forward process, leaving the inverse solving and control of physical parameters of objects on 3D GS as an unexplored area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.905, + 0.672 + ], + "angle": 0, + "content": "Contacts in locomotion control Friction plays a crucial role in soft body locomotion control. Stewart and Trinkle linearize the 3D friction model using a polyhedral cone and propose an LCP formulation to approximate Coulomb's friction cone conditions [38] and many improve upon it [2, 3, 24, 35, 43]. LCP-based methods offer exact solutions and are efficient for small to medium-sized problems but can be computationally expensive for optimization problems. Due to this drawback, many previous methods explicitly assumed that the contacts remain static [21, 27] while optimizing control forces subject to equations of motion. A few studies [41] exploit the physical meaning of complementarity constraints as heuristics to improve the solution and performance of the solver but still often fall back into exhaustive searching due to a lack of good initial state. The penalty method models contacts and frictions as spring forces and evaluates them at the instant [16-18, 42, 46] but can introduce numerical challenges such as stiffness and instability [1, 36]. The combination of the penalty methods and LCP-based methods has been less explored." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.687, + 0.605, + 0.701 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.625, + 0.727 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.905, + 0.854 + ], + "angle": 0, + "content": "A 3DGS scene consists of a collection of 3D Gaussians \\(\\mathcal{G}_{\\text {scene }}\\). Each Gaussian kernel is defined by a center position \\(\\mu \\in \\mathbb{R}^3\\), a covariance matrix \\(\\Sigma\\), the opacity \\(\\sigma \\in [0,1)\\) and a set of spherical harmonics coordinates \\(c \\in \\mathbb{R}^k\\) describing the colors emitted by the Gaussian in all directions. The covariance matrix \\(\\Sigma\\) can be factorized into a rotation matrix \\(\\mathbf{R}\\) expressed as a quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\) and a scaling matrix \\(\\mathbf{S}\\) of its diagonal vector \\(\\mathbf{s} \\in \\mathbb{R}^3\\): \\(\\boldsymbol{\\Sigma} = \\mathbf{RSS}^T\\mathbf{R}^T\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Vanilla 3DGS represents a static environment where users can switch their viewpoints within the scene freely. It is also possible to include dynamics to the scene by adding" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.48, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Figure 2. Bidirectional local embedding. Our bidirectional local embedding (a) serves as guidance for (b) self-splitting GS deformation, (c) accurate collision handling, and (d) non-penetration regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.478 + ], + "angle": 0, + "content": "external forces e.g., as in PhysGaussian [45] so that 3DGS models move or deform under user-specified forces passively. Nevertheless, real-world living creatures are self-driven — their movements are enabled by embedded muscles given a desired kinematic motion target. Such an active movement mechanism is much more difficult to obtain as it inevitably involves frictional contacts between the model and the environment. To this end, we incorporate additional physical expressions into the Gaussian elements, enabling them to be self-driven in a physics-based way. These elements can change their appearance at each time step and approximate user-specified motion targets interacting with the environment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.489, + 0.428, + 0.505 + ], + "angle": 0, + "content": "3.2. Hybrid Mesh-Gaussian Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.646 + ], + "angle": 0, + "content": "Let \\(\\mathcal{G}_o\\) denote the Gaussian elements of an object of interest, which is segmented from the scene \\(\\mathcal{G}_{\\text{scene}}\\). We first create an encapsulating tetrahedral mesh \\(\\mathcal{T} = (\\mathcal{V}, \\mathcal{F})\\) to approximate the object's overall volume, mass, and deformation. Here, \\(\\mathcal{V}\\) and \\(\\mathcal{F}\\) refer to the sets of nodal points and faces in \\(\\mathcal{T}\\). To achieve this segmentation, we approximate the nearby environment with planes (more in Secs. 3.5 and 3.6). The plane \\(P\\) expressed as a unit normal vector \\(\\mathbf{n}\\) and its distance \\(d\\) from the origin:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.657, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\nP (\\mathbf {n}, d) = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} \\mid \\mathbf {N} ^ {T} \\mathbf {x} + d = 0 \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.76 + ], + "angle": 0, + "content": "is used to \"cut out\" irrelevant Gaussians, reduce the blurred regions at the cut, and serve as a prior for inpainting the missing parts after segmentation. Additionally, it is utilized for collision detection to determine the collision forces acting on \\(\\mathcal{G}_o\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.821 + ], + "angle": 0, + "content": "\\(\\mathcal{T}\\) determines the states of Gaussian kernels in \\(\\mathcal{G}_o\\) after deformation. The deformed position of a point \\(\\mathbf{x}'\\) originally at \\(\\mathbf{x}\\) within a tetrahedron is represented by the barycentric coordinates of the four vertices of the tetrahedron \\(\\mathbf{v}'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.829, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = \\sum_ {i} \\lambda_ {i} (\\mathbf {x}) \\mathbf {v} _ {i} ^ {\\prime}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "After deformation, \\(\\mathcal{T}\\) provides a global displacement map for all Gaussians in \\(\\mathcal{G}_o\\). However, the internal rotation and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "scaling of the Gaussians cannot be accurately described by position-based interpolation, resulting in spiky artifacts. Additionally, the tetrahedral mesh is a coarse mesh. Its faces have significant gaps from the boundaries of the object, and it does not accurately describe the collision between \\(\\mathcal{G}_o\\) and the plane. To address these issues, we establish a bidirectional local embedding within the Gaussian around the mean \\(\\mathbf{m}\\) to precisely describe its influence range. This is determined by the endpoints of the ellipsoid corresponding to the Gaussian \\([\\mathbf{v}_1,\\mathbf{v}_2,\\mathbf{v}_3]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.255, + 0.905, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {T} ^ {\\pm} = [ \\mathbf {m}, \\mathbf {v} _ {1}, \\mathbf {v} _ {2}, \\mathbf {v} _ {3} ] \\\\ = [ \\mathbf {m}, \\mathbf {m} \\pm \\gamma s _ {1} \\mathbf {r} _ {1}, \\mathbf {m} \\pm \\gamma s _ {2} \\mathbf {r} _ {2}, \\mathbf {m} \\pm \\gamma s _ {3} \\mathbf {r} _ {3} ], \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.906, + 0.39 + ], + "angle": 0, + "content": "where the relevant parameters can be obtained from the scaling matrix \\(\\mathbf{S} = \\mathrm{diag}(s_1, s_2, s_3)\\) and the rotation Matrix \\(\\mathbf{R} = [\\mathbf{r}_1, \\mathbf{r}_2, \\mathbf{r}_3]\\). \\(\\gamma\\) is a parameter that describes the range of influence, and we set its default value as 3.0. Each Gaussian has two directional embeddings, \\(\\mathbf{T}^{+}\\) and \\(\\mathbf{T}^{-}\\), and each contains four points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.906, + 0.556 + ], + "angle": 0, + "content": "We design bidirectional local embeddings for several purposes. For collision processing, it is used to detect the extent to which an object collides with the environmental plane; and for segmentation, it detects whether the Gaussian is cut by the plane and establishes a loss term to penalize Gaussian kernels that intersect the plane under initial setting; For deformation, it is used to accurately describe the scaling and rotation of the deformed Gaussian kernel. Finally, based on the angle between two directions, it determines whether the deformed Gaussian kernel is sufficient to cover the radiance field." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.567, + 0.795, + 0.583 + ], + "angle": 0, + "content": "3.3. Soft Body Locomotion on Plane" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.664 + ], + "angle": 0, + "content": "The core of our pipeline is an active locomotion solver for \\(\\mathcal{T}\\) and \\(\\mathcal{G}_o\\). This is in the form of an inverse problem involving high-dimension inequalities constraints for processing frictional contacts between a 3DGS object and the environment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.673, + 0.738, + 0.688 + ], + "angle": 0, + "content": "3.3.1. Muscle-Driven Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.753 + ], + "angle": 0, + "content": "We embed muscle fibers within \\(\\mathcal{T}\\) to serve as the primary drivers of motion. At each time step, the nodal positions of \\(N\\) nodes in \\(\\mathcal{V}\\), denoted as \\(\\mathbf{p}\\), is determined through numerical integration of the dynamic equations of motion:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.766, + 0.905, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} \\ddot {\\mathbf {p}} = \\mathbf {f} _ {e x t} + \\mathbf {f} _ {i n t} + \\mathbf {f} _ {d} + \\mathbf {f} _ {m} + \\mathbf {f} _ {c}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}\\) is the mass matrix, \\(\\mathbf{f}_{ext},\\mathbf{f}_{int},\\mathbf{f}_d,\\mathbf{f}_c,\\mathbf{f}_m\\) represent external, internal, damping, muscle, and contact forces, respectively. We use \\((\\cdot)\\) for velocity and \\((\\cdot)\\) for acceleration. The external force \\(\\mathbf{f}_{ext}\\) includes e.g., the gravity force. The internal force \\(\\mathbf{f}_{int}\\), also known as the elastic force, is the resultant force generated by the material's elastic properties that resist deformation. In our implementation, this force" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.904, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.27, + 0.908, + 0.315 + ], + "angle": 0, + "content": "Figure 3. An overview of EnliveningGS pipeline. Our hybrid mesh-Gaussian representation plays an important role in all tasks within the pipeline. Our locomotion solver inversely computes the muscle activation embedded in the segmented 3DGS object that leads to user-specific motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.483, + 0.359 + ], + "angle": 0, + "content": "is determined as the negative gradient of the Neo-Hookean energy [5]:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.369, + 0.483, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\Psi = \\frac {\\mu}{2} (I _ {C} - 3) - \\mu \\ln J + \\frac {\\lambda}{2} (\\ln J) ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.469 + ], + "angle": 0, + "content": "Here, \\(\\mu\\) and \\(\\lambda\\) are the Lamé parameters. \\(I_{C} = \\operatorname{tr}(\\mathbf{F}^{T}\\mathbf{F})\\) and \\(J = \\det(\\mathbf{F})\\) are computed based on the deformation gradient \\(\\mathbf{F}\\). We use the Rayleigh damping for the damping force:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.473, + 0.483, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {d} = \\left(\\alpha \\mathbf {M} + \\beta \\mathbf {K} _ {0}\\right) \\dot {\\mathbf {p}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.497, + 0.381, + 0.511 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}_0\\) is the rest-shape stiffness matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.631 + ], + "angle": 0, + "content": "We consider muscle fibers as polygonal curves with \\(M\\) segments. The spring force caused by each segment is computed as: \\(f = ka\\) where \\(k\\) is the stiffness of the muscle fiber and \\(a\\) denotes the change in the segment's length. When a muscle segment activates, the spring force will shorten or bend the body around it. Muscle force on the whole soft body is encoded in a pose-dependent activation matrix \\(\\mathbf{A} \\in \\mathbb{R}^{3N \\times M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.635, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {m} = \\mathbf {A} (\\mathbf {p}) \\mathbf {a}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.718 + ], + "angle": 0, + "content": "where \\(\\mathbf{a} \\in \\mathbb{R}^{M}\\) is the vector of activations of all the muscle segments. Please refer to the supplementary materials for a complete description of the calculations involving activation matrix \\(\\mathbf{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We provide an automatic muscle generation tool given \\(\\mathcal{T}\\), which uses the medial axis transform [8, 29] of the model. Four longitudinal muscles are placed parallel near the medial axis at a given radius, and radial muscles are placed at given intervals. Users can also customize the muscle setup with preferred arrangements inside the object to achieve different movements. For instance, by implanting several longitudinal muscle fibers that run the length of the body, the body can be shortened when these muscles contract. Bending the body can be achieved by contracting these muscles unevenly on one side. Additionally, circular muscles wrap around the body, enabling it to twist, while" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.905, + 0.358 + ], + "angle": 0, + "content": "radial muscles extend across the body's cross-section, allowing the body to elongate while preserving its volume." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.363, + 0.684, + 0.378 + ], + "angle": 0, + "content": "3.3.2. Contacting 3DGS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.906, + 0.457 + ], + "angle": 0, + "content": "In this section we present a novel contact model for 3DGS. For Gaussian kernels in \\(\\mathcal{G}_o\\), we first define \\(\\mathbf{u}\\) as the position of points in bidirectional local embedding \\(\\mathbf{T}^{\\pm}\\) below the plane as \\(\\mathbf{N}^T\\mathbf{u} + d\\leq 0\\). Non-interpenetration constraints on 3DGS is thus defined as inequalities:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.465, + 0.905, + 0.483 + ], + "angle": 0, + "content": "\\[\nC _ {n} (\\mathbf {u}) \\equiv \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} \\geq \\mathbf {0}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.567 + ], + "angle": 0, + "content": "We decompose contact forces on the Gaussian as \\(\\mathbf{f}_g = \\mathbf{N}f_{\\perp} + \\mathbf{D}\\mathbf{f}_{\\parallel}\\), where \\(\\mathbf{D}\\) is a set of directions at the contact point approximating the friction cone, and they are perpendicular to the plane normal \\(\\mathbf{N}\\). \\(f_{\\perp}\\) and \\(\\mathbf{f}_{\\parallel}\\) are the magnitudes of normal and tangent forces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.595 + ], + "angle": 0, + "content": "LCP formulation [2] that regulates the contact velocity and contact force is:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.601, + 0.905, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathbf {0} \\leq \\left[ \\begin{array}{l} f _ {\\perp} \\\\ \\mathbf {f} _ {\\parallel} \\\\ \\lambda \\end{array} \\right] \\perp \\left[ \\begin{array}{c} \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} \\\\ \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} + \\mathbf {E} \\lambda \\\\ \\mu f _ {\\perp} - \\mathbf {E} ^ {T} \\mathbf {f} _ {\\parallel} \\end{array} \\right] \\geq \\mathbf {0}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.763 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the friction coefficient, and \\(\\lambda\\) is an auxiliary variable (e.g., the Lagrange multiplier) that relates the tangent velocity to a sliding contact. \\(\\mathbf{E}\\) is a block-diagonal matrix of \\(\\mathbf{e}\\). LCP condition notation \\(\\mathbf{x} \\perp \\mathbf{y}\\) indicates \\(\\mathbf{x}^T\\mathbf{y} = 0\\). The first line of Equation (9) is a complementarity formulation of contact that enforces the non-penetration condition, and the following two lines encode Coulomb's friction law." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.827 + ], + "angle": 0, + "content": "Like [33], contact can be viewed as a regularization of the complementarity form of the non-penetration constraint, which gives penalty forms of contact as spring force \\( \\mathbf{f}_g^{spring} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.837, + 0.905, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\perp} ^ {s p r i n g} \\equiv - k _ {n} \\min \\left(0, C _ {n} (\\mathbf {u})\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.87, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\parallel} ^ {s p r i n g} \\equiv - \\min \\left(k _ {f}, \\mu \\frac {\\| \\mathbf {f} ^ {s p r i n g} \\|}{\\| \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} \\|}\\right) \\mathbf {D} ^ {T} \\dot {\\mathbf {u}}, \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "899" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "where \\(k_{n}\\) and \\(k_{f}\\) control the stiffness of the contact. Forces on local embeddings can be converted back onto points on \\(\\mathcal{T}\\) using the same weights in Equation (2), giving us \\(\\mathbf{f}_c\\) or \\(\\mathbf{f}_c^{spring}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.157, + 0.295, + 0.172 + ], + "angle": 0, + "content": "3.3.3. Two-stage Locomotion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.176, + 0.484, + 0.25 + ], + "angle": 0, + "content": "Unlike force-based passive dynamics, we achieve the desired locomotion mainly by intuitively prescribing the motion trajectory of the object. For the position of a target point \\(\\mathbf{p}^{\\star}\\), the objective function that constrains the position can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.256, + 0.483, + 0.275 + ], + "angle": 0, + "content": "\\[\nG _ {\\text {p o s i t i o n}} = \\left\\| f (\\mathbf {p}) - \\mathbf {p} ^ {\\star} \\right\\| ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.385 + ], + "angle": 0, + "content": "where \\( f \\) can be functions that compute the positional target of the object. For instance, a commonly used setup is \\( f_{COM} \\), which calculates the COM (center of mass) \\( \\mathbf{c} = f_{COM}(\\mathbf{p}) \\) of the body, of local COM coordinates (foot, head, base, etc.). Similar objective functions can be used to track the velocity, acceleration, angular momentum, base area, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.492 + ], + "angle": 0, + "content": "We first solve Equation (4) by substituting \\(\\mathbf{f}_c\\) with \\(\\mathbf{f}_c^{spring}\\) and temporarily keeping muscle activation a same as the last time step. Based on the resulting \\(\\dot{\\mathbf{p}}\\), we convert the state of contact points back into the inequalities in Equation (9), and solve a Quadratic Programming (QP) problem to find the best \\(\\mathbf{a}\\). For example, if a contact point is in static contact, the corresponding QP is:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.498, + 0.335, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {a}, f _ {\\perp}, \\mathbf {f} _ {| |}, \\lambda} G \\left(\\mathbf {a}, f _ {\\perp}, \\mathbf {f} _ {| |}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.527, + 0.254, + 0.542 + ], + "angle": 0, + "content": "subject to" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.545, + 0.483, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathbf {0} \\leq f _ {\\perp}, \\quad \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} ^ {n + 1} = \\mathbf {0} \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.566, + 0.39, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathbf {0} \\leq \\mathbf {f} _ {\\parallel}, \\quad \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} ^ {n + 1} + \\mathbf {E} \\lambda = \\mathbf {0}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.588, + 0.373, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathbf {0} = \\lambda , \\quad \\mu f _ {\\perp} - \\mathbf {E} ^ {T} \\mathbf {f} _ {\\parallel} \\geq \\mathbf {0}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.657 + ], + "angle": 0, + "content": "Note that all contact points may have their own states. If the results are not satisfactory, we will use the pivoting method as in [41] to find the appropriate contact states." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.462, + 0.678 + ], + "angle": 0, + "content": "3.4. Gaussian Deformation in Tetrahedral Mesh" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.759 + ], + "angle": 0, + "content": "The barycentric coordinates of the points in \\(\\mathbf{T}^{+}\\) are calculated in the reference configuration, and the deformed coordinates in the current configuration \\(\\mathbf{T}_s^+ = [\\mathbf{m}',\\mathbf{v}_1',\\mathbf{v}_2',\\mathbf{v}_3']\\) are then calculated at each timestep to describe the deformation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.806 + ], + "angle": 0, + "content": "After that, we construct two matrices \\(\\mathbf{D}_m\\) and \\(\\mathbf{D}_s\\), representing the differences in vertex coordinates in the reference and current configurations, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.811, + 0.483, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r c l} \\mathbf {D} _ {m} & = & \\left[ \\mathbf {v} _ {1} ^ {0} - \\mathbf {m}, \\mathbf {v} _ {2} ^ {0} - \\mathbf {m}, \\mathbf {v} _ {3} ^ {0} - \\mathbf {m} \\right] \\\\ \\mathbf {D} _ {m} & = & \\left[ \\mathbf {v} _ {1} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\mathbf {v} _ {2} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\mathbf {v} _ {3} ^ {\\prime} - \\mathbf {m} ^ {\\prime} \\right] \\end{array} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.827, + 0.412, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {s} = \\left[ \\begin{array}{c} \\mathbf {v} _ {1} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\bar {\\mathbf {v}} _ {2} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\bar {\\mathbf {v}} _ {3} ^ {\\prime} - \\mathbf {m} ^ {\\prime} \\end{array} \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.878 + ], + "angle": 0, + "content": "The linear transformation \\(\\mathbf{F}\\) that maps the reference configuration to the current configuration can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.884, + 0.341, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\mathbf {D} _ {s} \\cdot \\mathbf {D} _ {m} ^ {- 1}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.906, + 0.137 + ], + "angle": 0, + "content": "The deformed covariance becomes: \\(\\Sigma^{\\prime} = \\mathbf{F}\\Sigma \\mathbf{F}^{T}\\) and is converted into scaling and rotation by performing singular value decomposition (SVD) on \\(\\Sigma^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.148, + 0.905, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {+} = \\operatorname {d i a g} \\left(\\sqrt {\\operatorname {S V D} _ {\\Sigma} \\left(\\Sigma^ {\\prime}\\right)}\\right) \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.164, + 0.758, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} ^ {+} = \\mathsf {S V D} _ {U} (\\Sigma^ {\\prime}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.906, + 0.268 + ], + "angle": 0, + "content": "Note that if \\(\\operatorname{det}(\\mathbf{R}^{+}) = -1\\), one needs to negate its last column to ensure it is a valid rotation matrix in \\(SO(3)\\). It is also necessary to adjust the spherical harmonics' orientation by applying rotation \\(R_{SH} = \\mathsf{SVD}_U(\\Sigma')\\mathsf{SVD}_V(\\Sigma')^T\\) to display the correct colors from various angles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.404 + ], + "angle": 0, + "content": "The points in local embeddings represent the shape of the deformed Gaussian kernel but they do not necessarily belong to the same tetrahedron. For some larger Gaussian kernels, significant rotations may cause them to extend beyond the original radiance field's boundaries. To avoid this situation, we use the reverse embedding \\(\\mathbf{T}^{-}\\) and use the aforementioned method to obtain its corresponding rotation matrix \\(\\mathbf{R}^{-}\\). The angle between \\(\\mathbf{R}^{+}\\) and \\(\\mathbf{R}^{-}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\theta = \\arccos \\left(\\frac {\\operatorname {t r} \\left(\\mathbf {R} ^ {+ T} \\mathbf {R} ^ {-}\\right) - 1}{2}\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.464, + 0.906, + 0.569 + ], + "angle": 0, + "content": "If \\(\\theta\\) exceeds the threshold \\(\\theta >\\eta\\), it indicates that the region covered by the Gaussian has transformed from a plane to a curved surface. A single kernel can no longer represent the deformed area accurately; thus, this Gaussian will be split into two from the center. Figure 2 (b) provides a schematic of our splitting process, while Figure 8 presents the corresponding experimental results." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.57, + 0.906, + 0.616 + ], + "angle": 0, + "content": "Inspired by [9, 15], the splitting of a Gaussian \\((\\alpha_0,\\mu_0,\\Sigma_0)\\) into two new Gaussian kernels \\((\\alpha_{l},\\mu_{l},\\Sigma_{l})\\) and \\((\\alpha_{r},\\mu_{r},\\Sigma_{r})\\) at the center is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.625, + 0.656, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {l} = \\alpha_ {r} = \\frac {\\alpha_ {0}}{2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.654, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mu_ {l} = \\mu_ {0} - \\frac {\\mathbf {v} _ {k} - \\mu_ {0}}{\\kappa}, \\quad \\mu_ {r} = \\mu_ {0} + \\frac {\\mathbf {v} _ {k} - \\mu_ {0}}{\\kappa}, \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.682, + 0.826, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {l} = \\boldsymbol {\\Sigma} _ {r} = \\boldsymbol {\\Sigma} _ {0} - \\frac {(\\mathbf {v} _ {k} - \\mu_ {0}) (\\mathbf {v} _ {k} - \\mu_ {0}) ^ {T}}{\\kappa^ {2}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.724, + 0.906, + 0.802 + ], + "angle": 0, + "content": "In the above, \\(\\mathbf{v}_k \\in \\mathbf{T}^+\\) stands for the endpoint on main principle axis with the largest scaling factor \\(s_k\\). Factor \\(\\kappa = \\frac{2\\gamma}{\\sqrt{2\\pi}}\\) is derived by maximizing the similarity between each new Gaussian and their own part of the original. We refer readers to supplementary material for the complete derivation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.81, + 0.682, + 0.825 + ], + "angle": 0, + "content": "3.5. Plane Estimation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.832, + 0.905, + 0.863 + ], + "angle": 0, + "content": "To find the parameters \\(\\mathbf{N}\\) and \\(d\\) of a plane in a 3DGS scene, we first construct a depth map on the plane:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.872, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {p} (u, v) = - \\frac {\\mathbf {N} \\cdot \\mathbf {t} _ {c} + d}{\\mathbf {n} ^ {T} \\mathbf {R} _ {c} ^ {T} \\mathbf {K} ^ {- 1} \\mathbf {p} (u , v)}, \\tag {18}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "900" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}(u,v)\\) represents the pixel coordinates in homogeneous form \\([u,v,1]^T\\), \\(\\mathbf{R}_c\\) and \\(\\mathbf{t}_c\\) are the rotation and translation extracted from the camera's extrinsic matrix, and \\(\\mathbf{K}\\) denotes the camera's intrinsic matrix. Initially, \\(\\mathbf{N}\\) is set to camera's look-at direction and \\(d\\) is set to zero to ensure solvability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.228 + ], + "angle": 0, + "content": "By minimizing the error between the depth map within the region and the GS depth map \\(\\mathcal{D}_g\\), the plane parameters can be optimized by:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.234, + 0.483, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\mathbf {N}, d} \\sum_ {(u, v) \\in \\mathcal {M}} \\| \\mathcal {D} _ {p} (u, v) - \\mathcal {D} _ {g} (u, v) \\|, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.482, + 0.316 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is a mask of the planar region which can be obtained through image-based segmentation algorithms [28] or roughly drawn by the user." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.482, + 0.376 + ], + "angle": 0, + "content": "The user can repeat this step on multiple images and optimize simultaneously to obtain fine-tuned plane parameters. In practice, a single viewpoint is usually sufficient to estimate an accurate plane." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.384, + 0.427, + 0.399 + ], + "angle": 0, + "content": "3.6. Object Segmentation with Plane Priors" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.483, + 0.541 + ], + "angle": 0, + "content": "We categorize the Gaussian scene into two types: dynamic splitting and static splitting. Equation (17) showcases dynamic online splitting to enhance the quality during the deformation process and ensure the speed of splitting on animating objects. For a Gaussian model, since the segmentation plane of the object and environment is already determined, it is more suitable to incorporate the plane information during training to regulate the Gaussian parameters at the cutting plane." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.616 + ], + "angle": 0, + "content": "To perform static splitting on the training stage, we first determine whether the Gaussian intersects with the plane. This requires checking if all the endpoints in \\(\\mathbf{T}^{\\pm}\\) are on the same side of the plane. We project center-to-endpoints distance onto the plane normal:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.624, + 0.482, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\eta_ {s e g} = \\max _ {\\mathbf {v} _ {i} \\in \\mathbf {T} ^ {\\pm}} | \\mathbf {N} ^ {T} (\\mathbf {v} _ {i} - \\mathbf {m}) |. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.697 + ], + "angle": 0, + "content": "If the distance from Gaussian kernel to the plane \\(|d_0| = |P(\\mathbf{n}, d)(\\mathbf{m})| < \\eta_{seg}\\), we will constrain the Gaussian's parameter or split the Gaussian with the plane." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.743 + ], + "angle": 0, + "content": "We fine-tune the trained Gaussian model by introducing the following non-penetration regularization term to ensure the model's influence range does not cross the plane:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.751, + 0.482, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e g} = \\left| \\left| \\eta_ {s e g} - \\left| d _ {0} \\right| \\right| \\right| _ {2} \\tag {21}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.343, + 0.79 + ], + "angle": 0, + "content": "3.7. Inpainting 3D Object on GS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "After extracting the object \\(\\mathcal{G}_o\\), a hole is left in the background \\(\\mathcal{G}_{\\text {scene }}\\). To address this issue, we introduce the depth map of the plane in Equation (18) into the depth-based inpainting method [30] and use a 2D inpainting tool [40] to obtain the color of the missing parts from the current viewpoint. Subsequently, we fine-tune the filled patch using the inpainted single-view image." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.645, + 0.107 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.252 + ], + "angle": 0, + "content": "We implemented our pipeline on a workstation with a single NVIDIA RTX4090 GPU and Intel 14900K CPU. We used PyTorch for storing and editing 3DGS on GPU, and OSQP [37] for solving QP in Equation (13). Our method supports the original 3DGS implementation [25]. For better visual quality on the planar surface, we set the minimum scaling factor on Gaussian kernels to \\(\\epsilon = e^{-10}\\), similar to 2DGS[19]. More benchmarks for the locomotion examples are reported in the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.261, + 0.698, + 0.277 + ], + "angle": 0, + "content": "4.1. Locomotion Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.905, + 0.373 + ], + "angle": 0, + "content": "Walking. We achieve walking by controlling the trajectories of all four legs. The diagonal support legs simultaneously lift the body, while the remaining two legs swing forward with the body's movement. As shown in Figure 4, the wooden stool moves one position to the right by taking four steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.531 + ], + "angle": 0, + "content": "Jumping. Jumping soft-bodied characters are more commonly seen in cartoons and animations. Jumping is a complex action because the contact state between the object and the ground changes rapidly and significantly. To achieve continuous jumping, it is necessary to precisely control the position of the COM and the relative velocity of the center of the base to jump the appropriate distance and maintain balance. We have designed an animation of a game of chess, where the soft-bodied chess pieces move across the board by continuously jumping, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Twisting. Figure 5 illustrates an example of the Peashooter doll twisting, jumping, and rotating 180 degrees. Based on the jumping motion, we additionally control the change in angular momentum along the central axis, allowing the Peashooter to rotate its body while airborne and face the zombie upon landing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.815, + 0.652 + ], + "angle": 0, + "content": "4.2. Comparison I. Deformation on GS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.795 + ], + "angle": 0, + "content": "We compare our method with the current state-of-the-art physico-aware GS methods. Figure 6 shows a qualitative comparison, where our method demonstrates better results under large deformation. Table 1 presents the related quantitative evaluation: we apply bending, twisting, and falling on the ground deformations to the ficus, chair, and mic examples from the NeRF synthetic dataset [34], and compare them with the rendered images of the source 3D mesh subjected to the same deformations as the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.845, + 0.819 + ], + "angle": 0, + "content": "4.3. Comparison II. Two-stage Locomotion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To evaluate our two-stage locomotion solver, we conducted frame-by-frame tests on an animation sequence with given muscle activation. This animation sequence includes 50 frames, 60 variables, and 50 pairs of linear complementarity constraints. We compared our solution with a QP" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "901" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.189, + 0.759, + 0.204 + ], + "angle": 0, + "content": "Figure 4. Walking stool. In the bedroom scene, a wooden stool walks like a quadruped." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.208, + 0.905, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Peashooter vs. Zombie. On a cloth-covered coffee table, a Peashooter figurine contracts, jumps, and twists its body, landing after a 180-degree spin to face the zombie." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.348, + 0.478, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Figure 6. Visual quality comparison on deforming objects. Compared to PhysGaussian [45] and VR-GS [23], our method can still suppress spiky artifacts under large deformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.716 + ], + "angle": 0, + "content": "solver based directly on the static contact assumption and the QPCC solver proposed by [41], and reported the average best and worst results across all frames in Table 2. Our locomotion solver uses the results predicted by the penalty force as guidance, reducing the number of iterations and lowering the objective value." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.339, + 0.745 + ], + "angle": 0, + "content": "4.4. Comparison III. Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 7, we demonstrate the inpainting effect on a plane using the garden example from Mip-NeRF360 dataset [4] and the truck example from DTU dataset [22], comparing it with the existing 3D object inpainting method on GS. In both examples, the removed objects occupy a significant portion of the photo, leaving a large hole on the ground after removal. Existing methods, due to the lack of relevant depth information, struggle to fill such a large missing area even after extensive finetuning of the inpainted 2D image (approximately 10,000 iterations). In contrast," + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.348, + 0.596, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.348, + 0.75, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.348, + 0.906, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Figure 7. Large 3D object inpainting. Comparison on large 3D object inpainting cases on the ground, where Gaussian Grouping [48] requires 20 minutes fine-tuning but still fails to fill the large hole on the ground while our method with better inpainting quality only needs 20 seconds fine-tuning. We show the render result as well as the depth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.656 + ], + "angle": 0, + "content": "our method leverages the plane as a prior, placing a dense and reasonably positioned point cloud in the missing area. This approach requires only about 100 iterations of finetuning, speeding up the inpainting process by several orders of magnitude and remaining unaffected by the size of the missing region." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.68, + 0.68 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.823 + ], + "angle": 0, + "content": "Bidirectional local embedding We first use a checkered bar to showcase the effect of bidirectional local embedding as in Figure 8. The checkered bar twists increasingly from right to left, and the sparky artifacts become progressively more pronounced. Without the need to detect whether the Gaussian kernel exceeds the object's surface, our bidirectional local embedding automatically identifies the regions causing sparky artifacts based on the information within the Gaussian kernel. Table 1 also reports quantitative results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Non-penetration regularization Since the segmentation plane under \\(\\mathcal{G}_o\\) is occluded from any viewpoint, directly masking out the object from the environment results in significant blurring and penetration at the segmentation plane." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "902" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.13, + 0.089, + 0.862, + 0.208 + ], + "angle": 0, + "content": "
BendingTwistingFalling
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
SuGaR [14]28.040.8900.04626.370.8230.10125.010.9130.053
PhysGaussian [45]28.140.9070.04326.530.8560.08925.100.9230.048
VR-GS [23]28.630.9170.03826.530.8930.06625.310.9380.044
Ours (w/o split)28.620.9180.03826.470.8910.06225.340.9390.042
Ours28.790.9230.03126.890.9100.04125.670.9510.032
" + }, + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.219, + 0.819, + 0.233 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of rendering quality after deformation on the NeRF synthetic dataset." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.244, + 0.471, + 0.33 + ], + "angle": 0, + "content": "
Iterations ↓Object value ↓
Avg.BestWorstAvg.BestWorst
QP---9.430.1623.86
QPCC [41]46.1915101.290.1412.11
Ours17.0113820.830.137.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.341, + 0.456, + 0.355 + ], + "angle": 0, + "content": "Table 2. Numerical experiments of locomotion solvers." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.386, + 0.273, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.432, + 0.215, + 0.444 + ], + "angle": 0, + "content": "Rest shape" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.378, + 0.478, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.437, + 0.415, + 0.444 + ], + "angle": 0, + "content": "w/o split" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.447, + 0.274, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.501, + 0.236, + 0.511 + ], + "angle": 0, + "content": "Detected artifacts" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.45, + 0.476, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.501, + 0.412, + 0.512 + ], + "angle": 0, + "content": "w/split" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Figure 8. Ablation study on bidirectional local embedding. Our method captures the regions that produce spiky artifacts and removes them by splitting the Gaussian kernels to fit the radiance field." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.585, + 0.481, + 0.657 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓Residual ↓
Mask (w/o Lseg)---0.058
Remove25.9170.8270.211-
w/ Lseg35.2910.9450.0340.002
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Table 3. Ablation study on non-penetration regularization: quantitative results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The Gaussian kernel that penetrates the plane contributes to the object's appearance, so a direct Remove operation would create holes in the object. Figure 9 shows the relevant qualitative comparison results. Our non-penetration regularization restricts the Gaussian from penetrating the plane, maintaining the integrity of the object's surface. In the quantitative experiments shown in Table 3, we placed the camera on the segmentation plane, compared the image quality of the object's upper half with full mask-out, and evaluated the residual of the lower half. The residuals are expressed as the ratio of the sum of pixel values in the lower half to those in the upper half." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.244, + 0.903, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.493 + ], + "angle": 0, + "content": "Figure 9. Ablation study on non-penetration regularization: qualitative results. With \\( L_{seg} \\) applied to Gaussian kernels penetrating the plane, the object retains its complete appearance near the cutting plane while removing the parts below the cutting plane." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.633, + 0.525 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.651 + ], + "angle": 0, + "content": "In this paper, we present EnliveningGS, a novel pipeline that enables active locomotion for 3D models represented by 3DGS. We use a hybrid mesh embedding of 3DGS to better estimate the contact between a 3DGS model and the environment. Given a high-level motion goal, we find the optimal muscle activations driving the movement of the object's body. An LCP-penalty mixed optimizer is used to efficiently calculate the contact and friction force." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.906, + 0.769 + ], + "angle": 0, + "content": "Limitation Currently, our method relies on plane-based simplification of the environment. When the surroundings consist of more complicated geometries, users may need to use more sophisticated approximation strategies to estimate the contact plane. It is also possible to use triangulated surfaces. Doint so could impose more computational overhead for the contact solver." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.906, + 0.888 + ], + "angle": 0, + "content": "Acknowledgements The authors would like to thank the reviewers for their insightful comments. Yin Yang is partially supported by NSF 2301040. This work is also supported by NSF China (No. 62322209 and No. 62421003), the gift from Adobe Research, the XPLORER PRIZE, and the 100 Talents Program of Zhejiang University. The source code and data are available at https://gapszju.github.io/EnliveningGS." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "903" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Sheldon Andrews, Kenny Erleben, and Zachary Ferguson. Contact and friction simulation for computer graphics. In ACM SIGGRAPH 2022 Courses, pages 1-172, Vancouver British Columbia Canada, 2022. ACM. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Mihai Anitescu and Florian A Potra. Formulating dynamic multi-rigid-body contact problems with friction as solvable linear complementarity problems. Nonlinear Dynamics, 14: 231-247, 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Lijie Bai, John E Mitchell, and Jong-Shi Pang. On convex quadratic programs with linear complementarity constraints. Computational Optimization and Applications, 54(3):517-554, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[5] Javier Bonet and Richard D Wood. Nonlinear continuum mechanics for finite element analysis. Cambridge university press, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[6] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physico-based simulations. arXiv preprint arXiv:2409.05819, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.485, + 0.514 + ], + "angle": 0, + "content": "[7] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[8] Hyeong In Choi, Sung Woo Choi, and Hwan Pyo Moon. Mathematical theory of medial axis transform. *pacific journal of mathematics*, 181(1):57-88, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[9] Qiyuan Feng, Gengchen Cao, Haoxiang Chen, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. A new split algorithm for 3d gaussian splatting. arXiv preprint arXiv:2403.09143, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Dynamic fluid synthesis with gaussian splatting. arXiv preprint arXiv:2401.15318, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[11] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[12] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based Gaussian Splatting for Real-time Large-scale Deformation. arXiv preprint arXiv:2402.04796, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[13] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. arXiv preprint arXiv:2403.14554, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[15] Peter Hall, David Marshall, and Ralph Martin. Adding and subtracting eigenspaces with eigenvalue decomposition and singular value decomposition. Image and vision computing, 20(13-14):1009-1016, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[16] Fursan Hamad, Shreyas Giridharan, and Christian Moormann. A penalty function method for modelling frictional contact in mpm. Procedia Engineering, 175:116-123, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[17] David Harmon, Etienne Vouga, Breannan Smith, Rasmus Tamstorf, and Eitan Grinspun. Asynchronous contact mechanics. In ACM SIGGRAPH 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[18] Shoichi Hasegawa and Nobuaki Fujii. Real-time rigid body simulation based on volumetric penalty method. In 11th Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems, 2003. HAPTICS 2003. Proceedings., pages 326-332. IEEE, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.364, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[19] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers* '24, pages 1-11, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[20] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4220-4230, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[21] Sumit Jain, Yuting Ye, and C Karen Liu. Optimization-based interactive motion synthesis. ACM Transactions on Graphics (TOG), 28(1):1-12, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[22] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanaes. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[23] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, and Chenfanfu Jiang. VR-GS: A Physical Dynamics-Aware Interactive Gaussian Splitting System in Virtual Reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[24] Joaquim J. Judice and Fernanda M Pires. A block principal pivoting algorithm for large-scale strictly monotone linear complementarity problems. Computers & operations research, 21(5):587-596, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Junggon Kim and Nancy S Pollard. Direct control of simulated nonhuman characters. IEEE Computer Graphics and Applications, 31(4):56-65, 2011. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "904" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Lei Lan, Ran Luo, Marco Fratarcangeli, Weiwei Xu, Huamin Wang, Xiaohu Guo, Junfeng Yao, and Yin Yang. Medial Elastics: Efficient and Collision-Ready Deformation via Medial Axis Transform. ACM Transactions on Graphics, 39(3): 1-17, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[30] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. InFusion: Inpainting 3D Gaussians via Learning Depth Completion from Diffusion Prior. arXiv preprint arXiv:2404.11613, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[31] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[32] Miles Macklin, Matthias Müller, and Nuttapong Chentanez. XPBD: Position-based simulation of compliant constrained dynamics. In Proceedings of the 9th International Conference on Motion in Games, pages 49-54, New York, NY, USA, 2016. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[33] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Tae-Yong Kim. Primal/dual descent methods for dynamics. In Computer Graphics Forum, pages 89-100. Wiley Online Library, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[34] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[35] Todd S Munson, Francisco Facchinei, Michael C Ferris, Andreas Fischer, and Christian Kanzow. The semismooth algorithm for large scale complementarity problems. INFORMS Journal on Computing, 13(4):294-311, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[36] Andrei-Ionut Stefancu, Silviu-Cristian Melenciuc, and Mihai Budescu. Penalty based algorithms for frictional contact problems. Buletinul Institutului Politehnic din lasi. Sectia Constructii, Architectura, 57(3):119, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[37] B. Stellato, G. Banjac, P. Goulart, A. Temporad, and S. Boyd. OSQP: an operator splitting solver for quadratic programs. Mathematical Programming Computation, 12(4): 637-672, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[38] D. E. Stewart and J. C. Trinkle. An Implicit Time-Stepping Scheme for Rigid Body Dynamics with Inelastic Collisions and Coulomb Friction. International Journal for Numerical Methods in Engineering, 39(15):2673-2691, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[39] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Jie Tan, Greg Turk, and C. Karen Liu. Soft body locomotion. ACM Transactions on Graphics, 31(4):26:1-26:11, 2012. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[42] Min Tang, Dinesh Manocha, Miguel A Otaduy, and Ruofeng Tong. Continuous penalty forces. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[43] Alessandro Tasora, Dario Mangoni, Simone Benatti, and Rinaldo Garziera. Solving variational inequalities and cone complementarity problems in nonsmooth dynamics using the alternating direction method of multipliers. International Journal for Numerical Methods in Engineering, 122(16): 4093-4113, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[44] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. GaMeS: Mesh-Based Adapting and Modification of Gaussian Splatting. arXiv preprint arXiv:2402.01459, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[45] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4389-4398, 2024. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[46] Hongyi Xu, Yili Zhao, and Jernej Barbič. Implicit multibody penalty-based distributed contact. IEEE transactions on visualization and computer graphics, 20(9):1266-1279, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[47] Shuojue Yang, Qian Li, Daiyun Shen, Bingchen Gong, Qi Dou, and Yueming Jin. Deform3dgs: Flexible deformation for fast surgical scene reconstruction with gaussian splattering. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 132-142. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[48] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. arXiv preprint arXiv:2312.00732, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[49] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[50] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "905" + } + ] +] \ No newline at end of file diff --git a/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_origin.pdf b/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..156d1b37ecba3705c9c9c5f157c54d4010b1464e --- /dev/null +++ b/2025/EnliveningGS_ Active Locomotion of 3DGS/678bd279-c773-4d3f-9e3f-a7dfe7dbc9fa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a1b6a449d56a50881d2b9912dae95bb8d10d6455f5b10e06cad9aa68acca0cf +size 8004939 diff --git a/2025/EnliveningGS_ Active Locomotion of 3DGS/full.md b/2025/EnliveningGS_ Active Locomotion of 3DGS/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae86d8f80f0c6c059d2000ee7849e9d0f69d85d4 --- /dev/null +++ b/2025/EnliveningGS_ Active Locomotion of 3DGS/full.md @@ -0,0 +1,423 @@ +# EnliveningGS: Active Locomotion of 3DGS + +Siyuan Shen $^{1}$ Tianjia Shao $^{1*}$ Kun Zhou $^{1}$ Chenfanfu Jiang $^{2}$ Yin Yang $^{3}$ $^{1}$ State Key Laboratory of CAD&CG, Zhejiang University, $^{2}$ UCLA, $^{3}$ University of Utah + +# Abstract + +This paper presents a novel pipeline named EnliveningGS, which enables active locomotion of 3D models represented with 3D Gaussian splatting (3DGS). We are inspired by the fact that real-world lives pose their bodies in a natural and physically meaningful manner by compressing or elongating muscle fibers embedded in the body. EnliveningGS aims to replicate the similar functionality of 3DGS models so that the object within a 3DGS scene acts like a living creature rather than a static shape — they walk, jump, and twist in the scene under provided motion trajectories driven by muscle activations. While the concept is straightforward, many challenging technical difficulties need to be taken care of. Synthesizing realistic locomotion of a 3DGS model embodies an inverse physics problem of very high dimensions. The core challenge is how to efficiently and robustly model frictional contacts between an “enlivened model” and the environment, as it is the composition of contact/collision/friction forces triggered by muscle activation that generates the final movement of the object. We propose a hybrid numerical method mixing LCP and penalty method to tackle this NP-hard problem robustly. Our pipeline also addresses the limitation of existing 3DGS deformation algorithms and inpainting the missing information when models move around. + +# 1. Introduction + +The advent of learning-based radiance field rendering techniques like NeRF [34] and 3D Gaussian splitting (3DGS) [26] have rebranded the pipeline of novel view synthesis. From synthesized RGB information, users can further alter and edit the illumination, appearance, and geometry. It is also possible to integrate those neural scene representations with physics to synthesize plausible physics-based dynamics. For instance, PIE-NeRF uses a quadratic moving least square interpolation to approximate the nonlinear elastodynamics of a NeRF model [11]. PhysGaussian [45] integrates 3DGS with the material point method (MPM) [39]. GaussianSplashing leverages a universal + +![](images/a0a6ac79fe35d4d681cd010c4b9b97c7468258d5cd0c6169c59fc58bd80b25f8.jpg) +Figure 1. Jumping Chess. We use static 3DGS as input and generates physics-aware 4D animations. The chess piece contracts and expands its body, creating jumping motions through collisions and friction with the chessboard, fitting the user's input trajectory. + +particle-based discretization to animate 3DGS scenes with fluids [10]. Those frameworks allow deeper integrations of 3DGS with relevant downstream graphics/vision tasks, enabling a wide range of exciting applications. + +It is a common paradigm, when combining physics and 3DGS, that the physical effects of the neural scene are passively triggered by external forces, which are normally specified by the user or pre-scripted (e.g., the gravity force). This makes sense for inanimate objects — they are not changing their kinematic states unless forced by external excitements. However, the locomotion of a living being follows a distinct modality. Rather than relying on external forces passively, a living entity actively adjusts its poses to achieve a desired movement such as walking, bending, or jumping. The enabler of such locomotion is muscle activations that contract or elongate different parts of the body, which interact with the environment e.g., the floor. The resultant contact and friction forces, in turn, provide the necessary momentum of the motion. This procedure is beyond the capability of existing physics-based 3DGS techniques as the muscle activations can hardly be measured or synthesized with a passive physics simulator. It is unfortunate that an interesting 3DGS scene can only house lifeless objects. + +In this paper, we show a novel framework to further en + +hance the 3DGS expressivity. We name our framework EnliveningGS, which inversely computes muscle activations embedded in a 3DGS model that lead to user-specific motion. As shown in Figure 1, the user only needs to provide a desired trajectory of the object (or other high-level descriptions of the kinematics). Given a set of embedded muscle fibers (approximated by piece-wise linear springs), EnliveningGS segments the object from the scene, estimates the possible contact/friction forces between the object and its surrounding environment, and calculates the optimal muscle activations to drive the corresponding body motion so that the object behaves like a living creature and moves following the prescribed trajectories as much as possible. As contact modeling is known to be an NP-hard problem and quickly becomes unsolvable when there are a moderate number of contact incidents, EnliveningGS introduces a two-stage locomotion procedure to reduce the solution space and efficiently handle frictional contacts between the 3DGS model and the scene. To summarize, our contributions include: + +- Active Animatable 3DGS Pipeline: We present En-livingGS pipeline that brings any household object to life. When they walk or jump, occluded areas become visible again. To overcome such visual artifacts, we propose an improved inpainting algorithm to fill colors and textures back to those exposed areas based on the known prior of the environment. We also present a self-splitting mechanism for Gaussian kernels to capture and reduce spiky artifacts under significant deformation. +- Hybrid Contact Modeling Framework for 3DGS: We design a novel hybrid numerical procedure combing the penalty method, which softens the hard contact constraint, and the classic Linear Complementarity Programming (LCP) [38] so that the combinatorial space of the active set can be significantly reduced, and the two-way coupling with the underlying nonlinear physics simulation becomes feasible. + +# 2. Related Work + +3D Gaussian editing. Recently, 3D Gaussian splatting (GS) [25] utilizes a set of 3D Gaussian kernels to represent static scenes explicitly and achieve state-of-the-art results. Subsequently, editing on GS has become a popular topic[7, 13, 30, 31, 47, 48, 50]. To make static 3D scenes deformable, some works extract explicit high-quality meshes from 3DGS representations and align GS kernels with mesh surfaces for guidance when deformed [12, 14, 44]. Their reliance on strong topological information and high-quality meshes makes it difficult to balance quality and interactivity. SC-GS [20] learns sparse control points for 3D scene dynamics but faces challenges with large movements and complex deformation. + +Physics-aware dynamics on 3D GS. To combine a static GS scene with a physics engine to generate dynamics on GS, PhysGaussian[45] leverages the first-order information from the displacement map (deformation gradient) to assist dynamic simulations. VR-GS [23] constructed a tetrahedral cage for each segmented Gaussian kernel group and embedded these kernel groups into corresponding meshes, allowing the deformed mesh driven by eXtended Position-based Dynamics (XPBD) [32] to guide the deformation of the Gaussian kernels. GASP[6] treats the simulation process as a black box and integrates Newtonian dynamics into the GaMeS[44] framework. PhysDreamer [49] distills dynamics priors learned by video generation models. It is worth mentioning that existing methods consider the simulation process as a forward process, leaving the inverse solving and control of physical parameters of objects on 3D GS as an unexplored area. + +Contacts in locomotion control Friction plays a crucial role in soft body locomotion control. Stewart and Trinkle linearize the 3D friction model using a polyhedral cone and propose an LCP formulation to approximate Coulomb's friction cone conditions [38] and many improve upon it [2, 3, 24, 35, 43]. LCP-based methods offer exact solutions and are efficient for small to medium-sized problems but can be computationally expensive for optimization problems. Due to this drawback, many previous methods explicitly assumed that the contacts remain static [21, 27] while optimizing control forces subject to equations of motion. A few studies [41] exploit the physical meaning of complementarity constraints as heuristics to improve the solution and performance of the solver but still often fall back into exhaustive searching due to a lack of good initial state. The penalty method models contacts and frictions as spring forces and evaluates them at the instant [16-18, 42, 46] but can introduce numerical challenges such as stiffness and instability [1, 36]. The combination of the penalty methods and LCP-based methods has been less explored. + +# 3. Method + +# 3.1. Overview + +A 3DGS scene consists of a collection of 3D Gaussians $\mathcal{G}_{\text {scene }}$ . Each Gaussian kernel is defined by a center position $\mu \in \mathbb{R}^3$ , a covariance matrix $\Sigma$ , the opacity $\sigma \in [0,1)$ and a set of spherical harmonics coordinates $c \in \mathbb{R}^k$ describing the colors emitted by the Gaussian in all directions. The covariance matrix $\Sigma$ can be factorized into a rotation matrix $\mathbf{R}$ expressed as a quaternion $\mathbf{q} \in \mathbb{R}^4$ and a scaling matrix $\mathbf{S}$ of its diagonal vector $\mathbf{s} \in \mathbb{R}^3$ : $\boldsymbol{\Sigma} = \mathbf{RSS}^T\mathbf{R}^T$ . + +Vanilla 3DGS represents a static environment where users can switch their viewpoints within the scene freely. It is also possible to include dynamics to the scene by adding + +![](images/b2de87bca3b56f490f5be11449fd51a68385a6ffa02669947cb6192d6dd88bb4.jpg) +Figure 2. Bidirectional local embedding. Our bidirectional local embedding (a) serves as guidance for (b) self-splitting GS deformation, (c) accurate collision handling, and (d) non-penetration regularization. + +external forces e.g., as in PhysGaussian [45] so that 3DGS models move or deform under user-specified forces passively. Nevertheless, real-world living creatures are self-driven — their movements are enabled by embedded muscles given a desired kinematic motion target. Such an active movement mechanism is much more difficult to obtain as it inevitably involves frictional contacts between the model and the environment. To this end, we incorporate additional physical expressions into the Gaussian elements, enabling them to be self-driven in a physics-based way. These elements can change their appearance at each time step and approximate user-specified motion targets interacting with the environment. + +# 3.2. Hybrid Mesh-Gaussian Representation + +Let $\mathcal{G}_o$ denote the Gaussian elements of an object of interest, which is segmented from the scene $\mathcal{G}_{\text{scene}}$ . We first create an encapsulating tetrahedral mesh $\mathcal{T} = (\mathcal{V}, \mathcal{F})$ to approximate the object's overall volume, mass, and deformation. Here, $\mathcal{V}$ and $\mathcal{F}$ refer to the sets of nodal points and faces in $\mathcal{T}$ . To achieve this segmentation, we approximate the nearby environment with planes (more in Secs. 3.5 and 3.6). The plane $P$ expressed as a unit normal vector $\mathbf{n}$ and its distance $d$ from the origin: + +$$ +P (\mathbf {n}, d) = \left\{\mathbf {x} \in \mathbb {R} ^ {3} \mid \mathbf {N} ^ {T} \mathbf {x} + d = 0 \right\}, \tag {1} +$$ + +is used to "cut out" irrelevant Gaussians, reduce the blurred regions at the cut, and serve as a prior for inpainting the missing parts after segmentation. Additionally, it is utilized for collision detection to determine the collision forces acting on $\mathcal{G}_o$ . + +$\mathcal{T}$ determines the states of Gaussian kernels in $\mathcal{G}_o$ after deformation. The deformed position of a point $\mathbf{x}'$ originally at $\mathbf{x}$ within a tetrahedron is represented by the barycentric coordinates of the four vertices of the tetrahedron $\mathbf{v}'$ : + +$$ +\mathbf {x} ^ {\prime} = \sum_ {i} \lambda_ {i} (\mathbf {x}) \mathbf {v} _ {i} ^ {\prime}. \tag {2} +$$ + +After deformation, $\mathcal{T}$ provides a global displacement map for all Gaussians in $\mathcal{G}_o$ . However, the internal rotation and + +scaling of the Gaussians cannot be accurately described by position-based interpolation, resulting in spiky artifacts. Additionally, the tetrahedral mesh is a coarse mesh. Its faces have significant gaps from the boundaries of the object, and it does not accurately describe the collision between $\mathcal{G}_o$ and the plane. To address these issues, we establish a bidirectional local embedding within the Gaussian around the mean $\mathbf{m}$ to precisely describe its influence range. This is determined by the endpoints of the ellipsoid corresponding to the Gaussian $[\mathbf{v}_1,\mathbf{v}_2,\mathbf{v}_3]$ : + +$$ +\begin{array}{l} \mathbf {T} ^ {\pm} = [ \mathbf {m}, \mathbf {v} _ {1}, \mathbf {v} _ {2}, \mathbf {v} _ {3} ] \\ = [ \mathbf {m}, \mathbf {m} \pm \gamma s _ {1} \mathbf {r} _ {1}, \mathbf {m} \pm \gamma s _ {2} \mathbf {r} _ {2}, \mathbf {m} \pm \gamma s _ {3} \mathbf {r} _ {3} ], \tag {3} \\ \end{array} +$$ + +where the relevant parameters can be obtained from the scaling matrix $\mathbf{S} = \mathrm{diag}(s_1, s_2, s_3)$ and the rotation Matrix $\mathbf{R} = [\mathbf{r}_1, \mathbf{r}_2, \mathbf{r}_3]$ . $\gamma$ is a parameter that describes the range of influence, and we set its default value as 3.0. Each Gaussian has two directional embeddings, $\mathbf{T}^{+}$ and $\mathbf{T}^{-}$ , and each contains four points. + +We design bidirectional local embeddings for several purposes. For collision processing, it is used to detect the extent to which an object collides with the environmental plane; and for segmentation, it detects whether the Gaussian is cut by the plane and establishes a loss term to penalize Gaussian kernels that intersect the plane under initial setting; For deformation, it is used to accurately describe the scaling and rotation of the deformed Gaussian kernel. Finally, based on the angle between two directions, it determines whether the deformed Gaussian kernel is sufficient to cover the radiance field. + +# 3.3. Soft Body Locomotion on Plane + +The core of our pipeline is an active locomotion solver for $\mathcal{T}$ and $\mathcal{G}_o$ . This is in the form of an inverse problem involving high-dimension inequalities constraints for processing frictional contacts between a 3DGS object and the environment. + +# 3.3.1. Muscle-Driven Dynamics + +We embed muscle fibers within $\mathcal{T}$ to serve as the primary drivers of motion. At each time step, the nodal positions of $N$ nodes in $\mathcal{V}$ , denoted as $\mathbf{p}$ , is determined through numerical integration of the dynamic equations of motion: + +$$ +\mathbf {M} \ddot {\mathbf {p}} = \mathbf {f} _ {e x t} + \mathbf {f} _ {i n t} + \mathbf {f} _ {d} + \mathbf {f} _ {m} + \mathbf {f} _ {c}, \tag {4} +$$ + +where $\mathbf{M}$ is the mass matrix, $\mathbf{f}_{ext},\mathbf{f}_{int},\mathbf{f}_d,\mathbf{f}_c,\mathbf{f}_m$ represent external, internal, damping, muscle, and contact forces, respectively. We use $(\cdot)$ for velocity and $(\cdot)$ for acceleration. The external force $\mathbf{f}_{ext}$ includes e.g., the gravity force. The internal force $\mathbf{f}_{int}$ , also known as the elastic force, is the resultant force generated by the material's elastic properties that resist deformation. In our implementation, this force + +![](images/cf42db0c67c03a6c9d9f2550c30d291b96a5824a96ffd9352abf68fae674a06a.jpg) +Figure 3. An overview of EnliveningGS pipeline. Our hybrid mesh-Gaussian representation plays an important role in all tasks within the pipeline. Our locomotion solver inversely computes the muscle activation embedded in the segmented 3DGS object that leads to user-specific motion. + +is determined as the negative gradient of the Neo-Hookean energy [5]: + +$$ +\Psi = \frac {\mu}{2} (I _ {C} - 3) - \mu \ln J + \frac {\lambda}{2} (\ln J) ^ {2}. \tag {5} +$$ + +Here, $\mu$ and $\lambda$ are the Lamé parameters. $I_{C} = \operatorname{tr}(\mathbf{F}^{T}\mathbf{F})$ and $J = \det(\mathbf{F})$ are computed based on the deformation gradient $\mathbf{F}$ . We use the Rayleigh damping for the damping force: + +$$ +\mathbf {f} _ {d} = \left(\alpha \mathbf {M} + \beta \mathbf {K} _ {0}\right) \dot {\mathbf {p}}, \tag {6} +$$ + +where $\mathbf{K}_0$ is the rest-shape stiffness matrix. + +We consider muscle fibers as polygonal curves with $M$ segments. The spring force caused by each segment is computed as: $f = ka$ where $k$ is the stiffness of the muscle fiber and $a$ denotes the change in the segment's length. When a muscle segment activates, the spring force will shorten or bend the body around it. Muscle force on the whole soft body is encoded in a pose-dependent activation matrix $\mathbf{A} \in \mathbb{R}^{3N \times M}$ : + +$$ +\mathbf {f} _ {m} = \mathbf {A} (\mathbf {p}) \mathbf {a}, \tag {7} +$$ + +where $\mathbf{a} \in \mathbb{R}^{M}$ is the vector of activations of all the muscle segments. Please refer to the supplementary materials for a complete description of the calculations involving activation matrix $\mathbf{A}$ . + +We provide an automatic muscle generation tool given $\mathcal{T}$ , which uses the medial axis transform [8, 29] of the model. Four longitudinal muscles are placed parallel near the medial axis at a given radius, and radial muscles are placed at given intervals. Users can also customize the muscle setup with preferred arrangements inside the object to achieve different movements. For instance, by implanting several longitudinal muscle fibers that run the length of the body, the body can be shortened when these muscles contract. Bending the body can be achieved by contracting these muscles unevenly on one side. Additionally, circular muscles wrap around the body, enabling it to twist, while + +radial muscles extend across the body's cross-section, allowing the body to elongate while preserving its volume. + +# 3.3.2. Contacting 3DGS + +In this section we present a novel contact model for 3DGS. For Gaussian kernels in $\mathcal{G}_o$ , we first define $\mathbf{u}$ as the position of points in bidirectional local embedding $\mathbf{T}^{\pm}$ below the plane as $\mathbf{N}^T\mathbf{u} + d\leq 0$ . Non-interpenetration constraints on 3DGS is thus defined as inequalities: + +$$ +C _ {n} (\mathbf {u}) \equiv \mathbf {N} ^ {T} \dot {\mathbf {u}} \geq \mathbf {0}. \tag {8} +$$ + +We decompose contact forces on the Gaussian as $\mathbf{f}_g = \mathbf{N}f_{\perp} + \mathbf{D}\mathbf{f}_{\parallel}$ , where $\mathbf{D}$ is a set of directions at the contact point approximating the friction cone, and they are perpendicular to the plane normal $\mathbf{N}$ . $f_{\perp}$ and $\mathbf{f}_{\parallel}$ are the magnitudes of normal and tangent forces. + +LCP formulation [2] that regulates the contact velocity and contact force is: + +$$ +\mathbf {0} \leq \left[ \begin{array}{l} f _ {\perp} \\ \mathbf {f} _ {\parallel} \\ \lambda \end{array} \right] \perp \left[ \begin{array}{c} \mathbf {N} ^ {T} \dot {\mathbf {u}} \\ \mathbf {D} ^ {T} \dot {\mathbf {u}} + \mathbf {E} \lambda \\ \mu f _ {\perp} - \mathbf {E} ^ {T} \mathbf {f} _ {\parallel} \end{array} \right] \geq \mathbf {0}, \tag {9} +$$ + +where $\mu$ is the friction coefficient, and $\lambda$ is an auxiliary variable (e.g., the Lagrange multiplier) that relates the tangent velocity to a sliding contact. $\mathbf{E}$ is a block-diagonal matrix of $\mathbf{e}$ . LCP condition notation $\mathbf{x} \perp \mathbf{y}$ indicates $\mathbf{x}^T\mathbf{y} = 0$ . The first line of Equation (9) is a complementarity formulation of contact that enforces the non-penetration condition, and the following two lines encode Coulomb's friction law. + +Like [33], contact can be viewed as a regularization of the complementarity form of the non-penetration constraint, which gives penalty forms of contact as spring force $\mathbf{f}_g^{spring}$ : + +$$ +\mathbf {f} _ {\perp} ^ {s p r i n g} \equiv - k _ {n} \min \left(0, C _ {n} (\mathbf {u})\right), \tag {10} +$$ + +$$ +\mathbf {f} _ {\parallel} ^ {s p r i n g} \equiv - \min \left(k _ {f}, \mu \frac {\| \mathbf {f} ^ {s p r i n g} \|}{\| \mathbf {D} ^ {T} \dot {\mathbf {u}} \|}\right) \mathbf {D} ^ {T} \dot {\mathbf {u}}, \tag {11} +$$ + +where $k_{n}$ and $k_{f}$ control the stiffness of the contact. Forces on local embeddings can be converted back onto points on $\mathcal{T}$ using the same weights in Equation (2), giving us $\mathbf{f}_c$ or $\mathbf{f}_c^{spring}$ . + +# 3.3.3. Two-stage Locomotion + +Unlike force-based passive dynamics, we achieve the desired locomotion mainly by intuitively prescribing the motion trajectory of the object. For the position of a target point $\mathbf{p}^{\star}$ , the objective function that constrains the position can be written as: + +$$ +G _ {\text {p o s i t i o n}} = \left\| f (\mathbf {p}) - \mathbf {p} ^ {\star} \right\| ^ {2}, \tag {12} +$$ + +where $f$ can be functions that compute the positional target of the object. For instance, a commonly used setup is $f_{COM}$ , which calculates the COM (center of mass) $\mathbf{c} = f_{COM}(\mathbf{p})$ of the body, of local COM coordinates (foot, head, base, etc.). Similar objective functions can be used to track the velocity, acceleration, angular momentum, base area, etc. + +We first solve Equation (4) by substituting $\mathbf{f}_c$ with $\mathbf{f}_c^{spring}$ and temporarily keeping muscle activation a same as the last time step. Based on the resulting $\dot{\mathbf{p}}$ , we convert the state of contact points back into the inequalities in Equation (9), and solve a Quadratic Programming (QP) problem to find the best $\mathbf{a}$ . For example, if a contact point is in static contact, the corresponding QP is: + +$$ +\min _ {\mathbf {a}, f _ {\perp}, \mathbf {f} _ {| |}, \lambda} G \left(\mathbf {a}, f _ {\perp}, \mathbf {f} _ {| |}\right) +$$ + +subject to + +$$ +\mathbf {0} \leq f _ {\perp}, \quad \mathbf {N} ^ {T} \dot {\mathbf {u}} ^ {n + 1} = \mathbf {0} \tag {13} +$$ + +$$ +\mathbf {0} \leq \mathbf {f} _ {\parallel}, \quad \mathbf {D} ^ {T} \dot {\mathbf {u}} ^ {n + 1} + \mathbf {E} \lambda = \mathbf {0} +$$ + +$$ +\mathbf {0} = \lambda , \quad \mu f _ {\perp} - \mathbf {E} ^ {T} \mathbf {f} _ {\parallel} \geq \mathbf {0}. +$$ + +Note that all contact points may have their own states. If the results are not satisfactory, we will use the pivoting method as in [41] to find the appropriate contact states. + +# 3.4. Gaussian Deformation in Tetrahedral Mesh + +The barycentric coordinates of the points in $\mathbf{T}^{+}$ are calculated in the reference configuration, and the deformed coordinates in the current configuration $\mathbf{T}_s^+ = [\mathbf{m}',\mathbf{v}_1',\mathbf{v}_2',\mathbf{v}_3']$ are then calculated at each timestep to describe the deformation. + +After that, we construct two matrices $\mathbf{D}_m$ and $\mathbf{D}_s$ , representing the differences in vertex coordinates in the reference and current configurations, respectively: + +$$ +\begin{array}{r c l} \mathbf {D} _ {m} & = & \left[ \mathbf {v} _ {1} ^ {0} - \mathbf {m}, \mathbf {v} _ {2} ^ {0} - \mathbf {m}, \mathbf {v} _ {3} ^ {0} - \mathbf {m} \right] \\ \mathbf {D} _ {m} & = & \left[ \mathbf {v} _ {1} ^ {\prime} - \mathbf {m} ^ {\prime}, \mathbf {v} _ {2} ^ {\prime} - \mathbf {m} ^ {\prime}, \mathbf {v} _ {3} ^ {\prime} - \mathbf {m} ^ {\prime} \right] \end{array} \tag {14} +$$ + +$$ +\mathbf {D} _ {s} = \left[ \begin{array}{c} \mathbf {v} _ {1} ^ {\prime} - \mathbf {m} ^ {\prime}, \bar {\mathbf {v}} _ {2} ^ {\prime} - \mathbf {m} ^ {\prime}, \bar {\mathbf {v}} _ {3} ^ {\prime} - \mathbf {m} ^ {\prime} \end{array} \right]. +$$ + +The linear transformation $\mathbf{F}$ that maps the reference configuration to the current configuration can be calculated as: + +$$ +\mathbf {F} = \mathbf {D} _ {s} \cdot \mathbf {D} _ {m} ^ {- 1}. +$$ + +The deformed covariance becomes: $\Sigma^{\prime} = \mathbf{F}\Sigma \mathbf{F}^{T}$ and is converted into scaling and rotation by performing singular value decomposition (SVD) on $\Sigma^{\prime}$ : + +$$ +\mathbf {S} ^ {+} = \operatorname {d i a g} \left(\sqrt {\operatorname {S V D} _ {\Sigma} \left(\Sigma^ {\prime}\right)}\right) \tag {15} +$$ + +$$ +\mathbf {R} ^ {+} = \mathsf {S V D} _ {U} (\Sigma^ {\prime}). +$$ + +Note that if $\operatorname{det}(\mathbf{R}^{+}) = -1$ , one needs to negate its last column to ensure it is a valid rotation matrix in $SO(3)$ . It is also necessary to adjust the spherical harmonics' orientation by applying rotation $R_{SH} = \mathsf{SVD}_U(\Sigma')\mathsf{SVD}_V(\Sigma')^T$ to display the correct colors from various angles. + +The points in local embeddings represent the shape of the deformed Gaussian kernel but they do not necessarily belong to the same tetrahedron. For some larger Gaussian kernels, significant rotations may cause them to extend beyond the original radiance field's boundaries. To avoid this situation, we use the reverse embedding $\mathbf{T}^{-}$ and use the aforementioned method to obtain its corresponding rotation matrix $\mathbf{R}^{-}$ . The angle between $\mathbf{R}^{+}$ and $\mathbf{R}^{-}$ can be expressed as: + +$$ +\theta = \arccos \left(\frac {\operatorname {t r} \left(\mathbf {R} ^ {+ T} \mathbf {R} ^ {-}\right) - 1}{2}\right). \tag {16} +$$ + +If $\theta$ exceeds the threshold $\theta >\eta$ , it indicates that the region covered by the Gaussian has transformed from a plane to a curved surface. A single kernel can no longer represent the deformed area accurately; thus, this Gaussian will be split into two from the center. Figure 2 (b) provides a schematic of our splitting process, while Figure 8 presents the corresponding experimental results. + +Inspired by [9, 15], the splitting of a Gaussian $(\alpha_0,\mu_0,\Sigma_0)$ into two new Gaussian kernels $(\alpha_{l},\mu_{l},\Sigma_{l})$ and $(\alpha_{r},\mu_{r},\Sigma_{r})$ at the center is calculated by: + +$$ +\alpha_ {l} = \alpha_ {r} = \frac {\alpha_ {0}}{2}, +$$ + +$$ +\mu_ {l} = \mu_ {0} - \frac {\mathbf {v} _ {k} - \mu_ {0}}{\kappa}, \quad \mu_ {r} = \mu_ {0} + \frac {\mathbf {v} _ {k} - \mu_ {0}}{\kappa}, \tag {17} +$$ + +$$ +\boldsymbol {\Sigma} _ {l} = \boldsymbol {\Sigma} _ {r} = \boldsymbol {\Sigma} _ {0} - \frac {(\mathbf {v} _ {k} - \mu_ {0}) (\mathbf {v} _ {k} - \mu_ {0}) ^ {T}}{\kappa^ {2}}. +$$ + +In the above, $\mathbf{v}_k \in \mathbf{T}^+$ stands for the endpoint on main principle axis with the largest scaling factor $s_k$ . Factor $\kappa = \frac{2\gamma}{\sqrt{2\pi}}$ is derived by maximizing the similarity between each new Gaussian and their own part of the original. We refer readers to supplementary material for the complete derivation. + +# 3.5. Plane Estimation + +To find the parameters $\mathbf{N}$ and $d$ of a plane in a 3DGS scene, we first construct a depth map on the plane: + +$$ +\mathcal {D} _ {p} (u, v) = - \frac {\mathbf {N} \cdot \mathbf {t} _ {c} + d}{\mathbf {n} ^ {T} \mathbf {R} _ {c} ^ {T} \mathbf {K} ^ {- 1} \mathbf {p} (u , v)}, \tag {18} +$$ + +where $\mathbf{p}(u,v)$ represents the pixel coordinates in homogeneous form $[u,v,1]^T$ , $\mathbf{R}_c$ and $\mathbf{t}_c$ are the rotation and translation extracted from the camera's extrinsic matrix, and $\mathbf{K}$ denotes the camera's intrinsic matrix. Initially, $\mathbf{N}$ is set to camera's look-at direction and $d$ is set to zero to ensure solvability. + +By minimizing the error between the depth map within the region and the GS depth map $\mathcal{D}_g$ , the plane parameters can be optimized by: + +$$ +\arg \min _ {\mathbf {N}, d} \sum_ {(u, v) \in \mathcal {M}} \| \mathcal {D} _ {p} (u, v) - \mathcal {D} _ {g} (u, v) \|, \tag {19} +$$ + +where $\mathcal{M}$ is a mask of the planar region which can be obtained through image-based segmentation algorithms [28] or roughly drawn by the user. + +The user can repeat this step on multiple images and optimize simultaneously to obtain fine-tuned plane parameters. In practice, a single viewpoint is usually sufficient to estimate an accurate plane. + +# 3.6. Object Segmentation with Plane Priors + +We categorize the Gaussian scene into two types: dynamic splitting and static splitting. Equation (17) showcases dynamic online splitting to enhance the quality during the deformation process and ensure the speed of splitting on animating objects. For a Gaussian model, since the segmentation plane of the object and environment is already determined, it is more suitable to incorporate the plane information during training to regulate the Gaussian parameters at the cutting plane. + +To perform static splitting on the training stage, we first determine whether the Gaussian intersects with the plane. This requires checking if all the endpoints in $\mathbf{T}^{\pm}$ are on the same side of the plane. We project center-to-endpoints distance onto the plane normal: + +$$ +\eta_ {s e g} = \max _ {\mathbf {v} _ {i} \in \mathbf {T} ^ {\pm}} | \mathbf {N} ^ {T} (\mathbf {v} _ {i} - \mathbf {m}) |. \tag {20} +$$ + +If the distance from Gaussian kernel to the plane $|d_0| = |P(\mathbf{n}, d)(\mathbf{m})| < \eta_{seg}$ , we will constrain the Gaussian's parameter or split the Gaussian with the plane. + +We fine-tune the trained Gaussian model by introducing the following non-penetration regularization term to ensure the model's influence range does not cross the plane: + +$$ +\mathcal {L} _ {s e g} = \left| \left| \eta_ {s e g} - \left| d _ {0} \right| \right| \right| _ {2} \tag {21} +$$ + +# 3.7. Inpainting 3D Object on GS + +After extracting the object $\mathcal{G}_o$ , a hole is left in the background $\mathcal{G}_{\text {scene }}$ . To address this issue, we introduce the depth map of the plane in Equation (18) into the depth-based inpainting method [30] and use a 2D inpainting tool [40] to obtain the color of the missing parts from the current viewpoint. Subsequently, we fine-tune the filled patch using the inpainted single-view image. + +# 4. Experiments + +We implemented our pipeline on a workstation with a single NVIDIA RTX4090 GPU and Intel 14900K CPU. We used PyTorch for storing and editing 3DGS on GPU, and OSQP [37] for solving QP in Equation (13). Our method supports the original 3DGS implementation [25]. For better visual quality on the planar surface, we set the minimum scaling factor on Gaussian kernels to $\epsilon = e^{-10}$ , similar to 2DGS[19]. More benchmarks for the locomotion examples are reported in the supplementary document. + +# 4.1. Locomotion Design + +Walking. We achieve walking by controlling the trajectories of all four legs. The diagonal support legs simultaneously lift the body, while the remaining two legs swing forward with the body's movement. As shown in Figure 4, the wooden stool moves one position to the right by taking four steps. + +Jumping. Jumping soft-bodied characters are more commonly seen in cartoons and animations. Jumping is a complex action because the contact state between the object and the ground changes rapidly and significantly. To achieve continuous jumping, it is necessary to precisely control the position of the COM and the relative velocity of the center of the base to jump the appropriate distance and maintain balance. We have designed an animation of a game of chess, where the soft-bodied chess pieces move across the board by continuously jumping, as shown in Figure 1. + +Twisting. Figure 5 illustrates an example of the Peashooter doll twisting, jumping, and rotating 180 degrees. Based on the jumping motion, we additionally control the change in angular momentum along the central axis, allowing the Peashooter to rotate its body while airborne and face the zombie upon landing. + +# 4.2. Comparison I. Deformation on GS + +We compare our method with the current state-of-the-art physico-aware GS methods. Figure 6 shows a qualitative comparison, where our method demonstrates better results under large deformation. Table 1 presents the related quantitative evaluation: we apply bending, twisting, and falling on the ground deformations to the ficus, chair, and mic examples from the NeRF synthetic dataset [34], and compare them with the rendered images of the source 3D mesh subjected to the same deformations as the ground truth. + +# 4.3. Comparison II. Two-stage Locomotion + +To evaluate our two-stage locomotion solver, we conducted frame-by-frame tests on an animation sequence with given muscle activation. This animation sequence includes 50 frames, 60 variables, and 50 pairs of linear complementarity constraints. We compared our solution with a QP + +![](images/be7f40502de4eb5fd1d0cf95dfa904018e5ff900e922910a6d7994978e6acc9f.jpg) +Figure 4. Walking stool. In the bedroom scene, a wooden stool walks like a quadruped. + +![](images/4de9d7892211b4138d355b7a33f08eaf5c94581ddee23317007830d4ec57e737.jpg) +Figure 5. Peashooter vs. Zombie. On a cloth-covered coffee table, a Peashooter figurine contracts, jumps, and twists its body, landing after a 180-degree spin to face the zombie. + +![](images/526b1ceff8a22c32a277fca95150967414931650572624af579c3b89cc489d4d.jpg) +Figure 6. Visual quality comparison on deforming objects. Compared to PhysGaussian [45] and VR-GS [23], our method can still suppress spiky artifacts under large deformations. + +solver based directly on the static contact assumption and the QPCC solver proposed by [41], and reported the average best and worst results across all frames in Table 2. Our locomotion solver uses the results predicted by the penalty force as guidance, reducing the number of iterations and lowering the objective value. + +# 4.4. Comparison III. Inpainting + +As shown in Figure 7, we demonstrate the inpainting effect on a plane using the garden example from Mip-NeRF360 dataset [4] and the truck example from DTU dataset [22], comparing it with the existing 3D object inpainting method on GS. In both examples, the removed objects occupy a significant portion of the photo, leaving a large hole on the ground after removal. Existing methods, due to the lack of relevant depth information, struggle to fill such a large missing area even after extensive finetuning of the inpainted 2D image (approximately 10,000 iterations). In contrast, + +![](images/3b8bf8c7662b9b530f14806fe8a7a93253f5c213cfa2eaa3352c52825e63b165.jpg) +Figure 7. Large 3D object inpainting. Comparison on large 3D object inpainting cases on the ground, where Gaussian Grouping [48] requires 20 minutes fine-tuning but still fails to fill the large hole on the ground while our method with better inpainting quality only needs 20 seconds fine-tuning. We show the render result as well as the depth. + +![](images/ee8c7d348c3552bc7c94b808a2e882413ffcf21a131b20b3e2d78ce1ff39ead0.jpg) + +![](images/5f68b70a34949f5b6dd57f03f444a35da0bb0abdf2483d18eb41cb27ac41d958.jpg) + +our method leverages the plane as a prior, placing a dense and reasonably positioned point cloud in the missing area. This approach requires only about 100 iterations of finetuning, speeding up the inpainting process by several orders of magnitude and remaining unaffected by the size of the missing region. + +# 4.5. Ablation Studies + +Bidirectional local embedding We first use a checkered bar to showcase the effect of bidirectional local embedding as in Figure 8. The checkered bar twists increasingly from right to left, and the sparky artifacts become progressively more pronounced. Without the need to detect whether the Gaussian kernel exceeds the object's surface, our bidirectional local embedding automatically identifies the regions causing sparky artifacts based on the information within the Gaussian kernel. Table 1 also reports quantitative results. + +Non-penetration regularization Since the segmentation plane under $\mathcal{G}_o$ is occluded from any viewpoint, directly masking out the object from the environment results in significant blurring and penetration at the segmentation plane. + +
BendingTwistingFalling
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
SuGaR [14]28.040.8900.04626.370.8230.10125.010.9130.053
PhysGaussian [45]28.140.9070.04326.530.8560.08925.100.9230.048
VR-GS [23]28.630.9170.03826.530.8930.06625.310.9380.044
Ours (w/o split)28.620.9180.03826.470.8910.06225.340.9390.042
Ours28.790.9230.03126.890.9100.04125.670.9510.032
+ +Table 1. Quantitative evaluation of rendering quality after deformation on the NeRF synthetic dataset. + +
Iterations ↓Object value ↓
Avg.BestWorstAvg.BestWorst
QP---9.430.1623.86
QPCC [41]46.1915101.290.1412.11
Ours17.0113820.830.137.89
+ +![](images/270088c2abe18f4b67809c5d0b707df7d4df7998780e0056c50e942458b20af4.jpg) +Rest shape + +![](images/c683898487a24bf74613ecd0e6983c757d8ac416af434b0c507523b8a3d3cae2.jpg) +w/o split + +![](images/1b13452fc052c610c093f662b7204b7ef682ac79110c38e6e4607fd66bc15e2c.jpg) +Detected artifacts +Figure 8. Ablation study on bidirectional local embedding. Our method captures the regions that produce spiky artifacts and removes them by splitting the Gaussian kernels to fit the radiance field. + +![](images/5ebd8957b8e9b6b23abec4c0da4d908c03474664a8252488c40cbe48fead5fa7.jpg) +w/split + +Table 2. Numerical experiments of locomotion solvers. + +
PSNR ↑SSIM ↑LPIPS ↓Residual ↓
Mask (w/o Lseg)---0.058
Remove25.9170.8270.211-
w/ Lseg35.2910.9450.0340.002
+ +Table 3. Ablation study on non-penetration regularization: quantitative results. + +The Gaussian kernel that penetrates the plane contributes to the object's appearance, so a direct Remove operation would create holes in the object. Figure 9 shows the relevant qualitative comparison results. Our non-penetration regularization restricts the Gaussian from penetrating the plane, maintaining the integrity of the object's surface. In the quantitative experiments shown in Table 3, we placed the camera on the segmentation plane, compared the image quality of the object's upper half with full mask-out, and evaluated the residual of the lower half. The residuals are expressed as the ratio of the sum of pixel values in the lower half to those in the upper half. + +![](images/474bef880366a2ecc61aec97330eaefc99bd5a880e08a0e5e64680f38fda7c2f.jpg) +Figure 9. Ablation study on non-penetration regularization: qualitative results. With $L_{seg}$ applied to Gaussian kernels penetrating the plane, the object retains its complete appearance near the cutting plane while removing the parts below the cutting plane. + +# 5. Conclusion + +In this paper, we present EnliveningGS, a novel pipeline that enables active locomotion for 3D models represented by 3DGS. We use a hybrid mesh embedding of 3DGS to better estimate the contact between a 3DGS model and the environment. Given a high-level motion goal, we find the optimal muscle activations driving the movement of the object's body. An LCP-penalty mixed optimizer is used to efficiently calculate the contact and friction force. + +Limitation Currently, our method relies on plane-based simplification of the environment. When the surroundings consist of more complicated geometries, users may need to use more sophisticated approximation strategies to estimate the contact plane. It is also possible to use triangulated surfaces. Doint so could impose more computational overhead for the contact solver. + +Acknowledgements The authors would like to thank the reviewers for their insightful comments. Yin Yang is partially supported by NSF 2301040. This work is also supported by NSF China (No. 62322209 and No. 62421003), the gift from Adobe Research, the XPLORER PRIZE, and the 100 Talents Program of Zhejiang University. The source code and data are available at https://gapszju.github.io/EnliveningGS. + +# References + +[1] Sheldon Andrews, Kenny Erleben, and Zachary Ferguson. Contact and friction simulation for computer graphics. In ACM SIGGRAPH 2022 Courses, pages 1-172, Vancouver British Columbia Canada, 2022. ACM. 2 +[2] Mihai Anitescu and Florian A Potra. Formulating dynamic multi-rigid-body contact problems with friction as solvable linear complementarity problems. Nonlinear Dynamics, 14: 231-247, 1997. 2, 4 +[3] Lijie Bai, John E Mitchell, and Jong-Shi Pang. On convex quadratic programs with linear complementarity constraints. Computational Optimization and Applications, 54(3):517-554, 2013. 2 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7 +[5] Javier Bonet and Richard D Wood. Nonlinear continuum mechanics for finite element analysis. Cambridge university press, 1997. 4 +[6] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physico-based simulations. arXiv preprint arXiv:2409.05819, 2024. 2 +[7] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2 +[8] Hyeong In Choi, Sung Woo Choi, and Hwan Pyo Moon. Mathematical theory of medial axis transform. *pacific journal of mathematics*, 181(1):57-88, 1997. 4 +[9] Qiyuan Feng, Gengchen Cao, Haoxiang Chen, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. A new split algorithm for 3d gaussian splatting. arXiv preprint arXiv:2403.09143, 2024.5 +[10] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Dynamic fluid synthesis with gaussian splatting. arXiv preprint arXiv:2401.15318, 2024. 1 +[11] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 1 +[12] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based Gaussian Splatting for Real-time Large-scale Deformation. arXiv preprint arXiv:2402.04796, 2024. 2 +[13] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. arXiv preprint arXiv:2403.14554, 2024. 2 +[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 8 +[15] Peter Hall, David Marshall, and Ralph Martin. Adding and subtracting eigenspaces with eigenvalue decomposition and singular value decomposition. Image and vision computing, 20(13-14):1009-1016, 2002. 5 +[16] Fursan Hamad, Shreyas Giridharan, and Christian Moormann. A penalty function method for modelling frictional contact in mpm. Procedia Engineering, 175:116-123, 2017. 2 +[17] David Harmon, Etienne Vouga, Breannan Smith, Rasmus Tamstorf, and Eitan Grinspun. Asynchronous contact mechanics. In ACM SIGGRAPH 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery. +[18] Shoichi Hasegawa and Nobuaki Fujii. Real-time rigid body simulation based on volumetric penalty method. In 11th Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems, 2003. HAPTICS 2003. Proceedings., pages 326-332. IEEE, 2003. 2 +[19] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers* '24, pages 1-11, 2024. 6 +[20] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4220-4230, 2024. 2 +[21] Sumit Jain, Yuting Ye, and C Karen Liu. Optimization-based interactive motion synthesis. ACM Transactions on Graphics (TOG), 28(1):1-12, 2009. 2 +[22] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanaes. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7 +[23] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, and Chenfanfu Jiang. VR-GS: A Physical Dynamics-Aware Interactive Gaussian Splitting System in Virtual Reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 2, 7, 8 +[24] Joaquim J. Judice and Fernanda M Pires. A block principal pivoting algorithm for large-scale strictly monotone linear complementarity problems. Computers & operations research, 21(5):587-596, 1994. 2 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 6 +[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1 +[27] Junggon Kim and Nancy S Pollard. Direct control of simulated nonhuman characters. IEEE Computer Graphics and Applications, 31(4):56-65, 2011. 2 + +[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6 +[29] Lei Lan, Ran Luo, Marco Fratarcangeli, Weiwei Xu, Huamin Wang, Xiaohu Guo, Junfeng Yao, and Yin Yang. Medial Elastics: Efficient and Collision-Ready Deformation via Medial Axis Transform. ACM Transactions on Graphics, 39(3): 1-17, 2020. 4 +[30] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. InFusion: Inpainting 3D Gaussians via Learning Depth Completion from Diffusion Prior. arXiv preprint arXiv:2404.11613, 2024. 2, 6 +[31] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2 +[32] Miles Macklin, Matthias Müller, and Nuttapong Chentanez. XPBD: Position-based simulation of compliant constrained dynamics. In Proceedings of the 9th International Conference on Motion in Games, pages 49-54, New York, NY, USA, 2016. Association for Computing Machinery. 2 +[33] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Tae-Yong Kim. Primal/dual descent methods for dynamics. In Computer Graphics Forum, pages 89-100. Wiley Online Library, 2020. 4 +[34] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1, 6 +[35] Todd S Munson, Francisco Facchinei, Michael C Ferris, Andreas Fischer, and Christian Kanzow. The semismooth algorithm for large scale complementarity problems. INFORMS Journal on Computing, 13(4):294-311, 2001. 2 +[36] Andrei-Ionut Stefancu, Silviu-Cristian Melenciuc, and Mihai Budescu. Penalty based algorithms for frictional contact problems. Buletinul Institutului Politehnic din lasi. Sectia Constructii, Architectura, 57(3):119, 2011. 2 +[37] B. Stellato, G. Banjac, P. Goulart, A. Temporad, and S. Boyd. OSQP: an operator splitting solver for quadratic programs. Mathematical Programming Computation, 12(4): 637-672, 2020. 6 +[38] D. E. Stewart and J. C. Trinkle. An Implicit Time-Stepping Scheme for Rigid Body Dynamics with Inelastic Collisions and Coulomb Friction. International Journal for Numerical Methods in Engineering, 39(15):2673-2691, 1996. 2 +[39] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 1 +[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with + +fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6 +[41] Jie Tan, Greg Turk, and C. Karen Liu. Soft body locomotion. ACM Transactions on Graphics, 31(4):26:1-26:11, 2012. 2, 5, 7, 8 +[42] Min Tang, Dinesh Manocha, Miguel A Otaduy, and Ruofeng Tong. Continuous penalty forces. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 2 +[43] Alessandro Tasora, Dario Mangoni, Simone Benatti, and Rinaldo Garziera. Solving variational inequalities and cone complementarity problems in nonsmooth dynamics using the alternating direction method of multipliers. International Journal for Numerical Methods in Engineering, 122(16): 4093-4113, 2021. 2 +[44] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. GaMeS: Mesh-Based Adapting and Modification of Gaussian Splatting. arXiv preprint arXiv:2402.01459, 2024. 2 +[45] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4389-4398, 2024. 1, 2, 3, 7, 8 +[46] Hongyi Xu, Yili Zhao, and Jernej Barbič. Implicit multibody penalty-based distributed contact. IEEE transactions on visualization and computer graphics, 20(9):1266-1279, 2014. 2 +[47] Shuojue Yang, Qian Li, Daiyun Shen, Bingchen Gong, Qi Dou, and Yueming Jin. Deform3dgs: Flexible deformation for fast surgical scene reconstruction with gaussian splattering. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 132-142. Springer, 2024. 2 +[48] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. arXiv preprint arXiv:2312.00732, 2024. 2, 7 +[49] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2024. 2 +[50] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2 \ No newline at end of file diff --git a/2025/EnliveningGS_ Active Locomotion of 3DGS/images.zip b/2025/EnliveningGS_ Active Locomotion of 3DGS/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..93e096063cca9abc047b0fb95b32dd6672dadb10 --- /dev/null +++ b/2025/EnliveningGS_ Active Locomotion of 3DGS/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dafbb0ab817bd2219805c09662de28ab047569e42e04fc6cd67279c59465aaea +size 592335 diff --git a/2025/EnliveningGS_ Active Locomotion of 3DGS/layout.json b/2025/EnliveningGS_ Active Locomotion of 3DGS/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e31bb787fda4f304380e59bc4612840d59946408 --- /dev/null +++ b/2025/EnliveningGS_ Active Locomotion of 3DGS/layout.json @@ -0,0 +1,10673 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 171, + 103, + 440, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 103, + 440, + 120 + ], + "spans": [ + { + "bbox": [ + 171, + 103, + 440, + 120 + ], + "type": "text", + "content": "EnliveningGS: Active Locomotion of 3DGS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "spans": [ + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": "Siyuan Shen" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": " Tianjia Shao" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": " Kun Zhou" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": " Chenfanfu Jiang" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": " Yin Yang" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": " State Key Laboratory of CAD&CG, Zhejiang University, " + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": " UCLA, " + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 142, + 518, + 171 + ], + "type": "text", + "content": " University of Utah" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 198, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 198, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 198, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 224, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 297, + 510 + ], + "type": "text", + "content": "This paper presents a novel pipeline named EnliveningGS, which enables active locomotion of 3D models represented with 3D Gaussian splatting (3DGS). We are inspired by the fact that real-world lives pose their bodies in a natural and physically meaningful manner by compressing or elongating muscle fibers embedded in the body. EnliveningGS aims to replicate the similar functionality of 3DGS models so that the object within a 3DGS scene acts like a living creature rather than a static shape — they walk, jump, and twist in the scene under provided motion trajectories driven by muscle activations. While the concept is straightforward, many challenging technical difficulties need to be taken care of. Synthesizing realistic locomotion of a 3DGS model embodies an inverse physics problem of very high dimensions. The core challenge is how to efficiently and robustly model frictional contacts between an “enlivened model” and the environment, as it is the composition of contact/collision/friction forces triggered by muscle activation that generates the final movement of the object. We propose a hybrid numerical method mixing LCP and penalty method to tackle this NP-hard problem robustly. Our pipeline also addresses the limitation of existing 3DGS deformation algorithms and inpainting the missing information when models move around." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 553, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 296, + 696 + ], + "type": "text", + "content": "The advent of learning-based radiance field rendering techniques like NeRF [34] and 3D Gaussian splitting (3DGS) [26] have rebranded the pipeline of novel view synthesis. From synthesized RGB information, users can further alter and edit the illumination, appearance, and geometry. It is also possible to integrate those neural scene representations with physics to synthesize plausible physics-based dynamics. For instance, PIE-NeRF uses a quadratic moving least square interpolation to approximate the nonlinear elastodynamics of a NeRF model [11]. PhysGaussian [45] integrates 3DGS with the material point method (MPM) [39]. GaussianSplashing leverages a universal" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 200, + 552, + 345 + ], + "blocks": [ + { + "bbox": [ + 317, + 200, + 552, + 345 + ], + "lines": [ + { + "bbox": [ + 317, + 200, + 552, + 345 + ], + "spans": [ + { + "bbox": [ + 317, + 200, + 552, + 345 + ], + "type": "image", + "image_path": "a0a6ac79fe35d4d681cd010c4b9b97c7468258d5cd0c6169c59fc58bd80b25f8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 356, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 400 + ], + "type": "text", + "content": "Figure 1. Jumping Chess. We use static 3DGS as input and generates physics-aware 4D animations. The chess piece contracts and expands its body, creating jumping motions through collisions and friction with the chessboard, fitting the user's input trajectory." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 421, + 555, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 470 + ], + "type": "text", + "content": "particle-based discretization to animate 3DGS scenes with fluids [10]. Those frameworks allow deeper integrations of 3DGS with relevant downstream graphics/vision tasks, enabling a wide range of exciting applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "content": "It is a common paradigm, when combining physics and 3DGS, that the physical effects of the neural scene are passively triggered by external forces, which are normally specified by the user or pre-scripted (e.g., the gravity force). This makes sense for inanimate objects — they are not changing their kinematic states unless forced by external excitements. However, the locomotion of a living being follows a distinct modality. Rather than relying on external forces passively, a living entity actively adjusts its poses to achieve a desired movement such as walking, bending, or jumping. The enabler of such locomotion is muscle activations that contract or elongate different parts of the body, which interact with the environment e.g., the floor. The resultant contact and friction forces, in turn, provide the necessary momentum of the motion. This procedure is beyond the capability of existing physics-based 3DGS techniques as the muscle activations can hardly be measured or synthesized with a passive physics simulator. It is unfortunate that an interesting 3DGS scene can only house lifeless objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "In this paper, we show a novel framework to further en" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 150, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 150, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "896" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 310 + ], + "type": "text", + "content": "hance the 3DGS expressivity. We name our framework EnliveningGS, which inversely computes muscle activations embedded in a 3DGS model that lead to user-specific motion. As shown in Figure 1, the user only needs to provide a desired trajectory of the object (or other high-level descriptions of the kinematics). Given a set of embedded muscle fibers (approximated by piece-wise linear springs), EnliveningGS segments the object from the scene, estimates the possible contact/friction forces between the object and its surrounding environment, and calculates the optimal muscle activations to drive the corresponding body motion so that the object behaves like a living creature and moves following the prescribed trajectories as much as possible. As contact modeling is known to be an NP-hard problem and quickly becomes unsolvable when there are a moderate number of contact incidents, EnliveningGS introduces a two-stage locomotion procedure to reduce the solution space and efficiently handle frictional contacts between the 3DGS model and the scene. To summarize, our contributions include:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 316, + 294, + 519 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 316, + 294, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 294, + 424 + ], + "type": "text", + "content": "- Active Animatable 3DGS Pipeline: We present En-livingGS pipeline that brings any household object to life. When they walk or jump, occluded areas become visible again. To overcome such visual artifacts, we propose an improved inpainting algorithm to fill colors and textures back to those exposed areas based on the known prior of the environment. We also present a self-splitting mechanism for Gaussian kernels to capture and reduce spiky artifacts under significant deformation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 425, + 294, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 294, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 294, + 519 + ], + "type": "text", + "content": "- Hybrid Contact Modeling Framework for 3DGS: We design a novel hybrid numerical procedure combing the penalty method, which softens the hard contact constraint, and the classic Linear Complementarity Programming (LCP) [38] so that the combinatorial space of the active set can be significantly reduced, and the two-way coupling with the underlying nonlinear physics simulation becomes feasible." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 536, + 142, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 142, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 142, + 548 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "3D Gaussian editing. Recently, 3D Gaussian splatting (GS) [25] utilizes a set of 3D Gaussian kernels to represent static scenes explicitly and achieve state-of-the-art results. Subsequently, editing on GS has become a popular topic[7, 13, 30, 31, 47, 48, 50]. To make static 3D scenes deformable, some works extract explicit high-quality meshes from 3DGS representations and align GS kernels with mesh surfaces for guidance when deformed [12, 14, 44]. Their reliance on strong topological information and high-quality meshes makes it difficult to balance quality and interactivity. SC-GS [20] learns sparse control points for 3D scene dynamics but faces challenges with large movements and complex deformation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "Physics-aware dynamics on 3D GS. To combine a static GS scene with a physics engine to generate dynamics on GS, PhysGaussian[45] leverages the first-order information from the displacement map (deformation gradient) to assist dynamic simulations. VR-GS [23] constructed a tetrahedral cage for each segmented Gaussian kernel group and embedded these kernel groups into corresponding meshes, allowing the deformed mesh driven by eXtended Position-based Dynamics (XPBD) [32] to guide the deformation of the Gaussian kernels. GASP[6] treats the simulation process as a black box and integrates Newtonian dynamics into the GaMeS[44] framework. PhysDreamer [49] distills dynamics priors learned by video generation models. It is worth mentioning that existing methods consider the simulation process as a forward process, leaving the inverse solving and control of physical parameters of objects on 3D GS as an unexplored area." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 292, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 532 + ], + "type": "text", + "content": "Contacts in locomotion control Friction plays a crucial role in soft body locomotion control. Stewart and Trinkle linearize the 3D friction model using a polyhedral cone and propose an LCP formulation to approximate Coulomb's friction cone conditions [38] and many improve upon it [2, 3, 24, 35, 43]. LCP-based methods offer exact solutions and are efficient for small to medium-sized problems but can be computationally expensive for optimization problems. Due to this drawback, many previous methods explicitly assumed that the contacts remain static [21, 27] while optimizing control forces subject to equations of motion. A few studies [41] exploit the physical meaning of complementarity constraints as heuristics to improve the solution and performance of the solver but still often fall back into exhaustive searching due to a lack of good initial state. The penalty method models contacts and frictions as spring forces and evaluates them at the instant [16-18, 42, 46] but can introduce numerical challenges such as stiffness and instability [1, 36]. The combination of the penalty methods and LCP-based methods has been less explored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 544, + 370, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 370, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 370, + 555 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 382, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 382, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 382, + 575 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": "A 3DGS scene consists of a collection of 3D Gaussians " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text {scene }}" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": ". Each Gaussian kernel is defined by a center position " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": ", a covariance matrix " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": ", the opacity " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\sigma \\in [0,1)" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": " and a set of spherical harmonics coordinates " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": " describing the colors emitted by the Gaussian in all directions. The covariance matrix " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": " can be factorized into a rotation matrix " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": " expressed as a quaternion " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": " and a scaling matrix " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": " of its diagonal vector " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} = \\mathbf{RSS}^T\\mathbf{R}^T" + }, + { + "bbox": [ + 313, + 581, + 553, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Vanilla 3DGS represents a static environment where users can switch their viewpoints within the scene freely. It is also possible to include dynamics to the scene by adding" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "897" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 293, + 149 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 293, + 149 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 293, + 149 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 293, + 149 + ], + "type": "image", + "image_path": "b2de87bca3b56f490f5be11449fd51a68385a6ffa02669947cb6192d6dd88bb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 158, + 295, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 204 + ], + "type": "text", + "content": "Figure 2. Bidirectional local embedding. Our bidirectional local embedding (a) serves as guidance for (b) self-splitting GS deformation, (c) accurate collision handling, and (d) non-penetration regularization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 378 + ], + "type": "text", + "content": "external forces e.g., as in PhysGaussian [45] so that 3DGS models move or deform under user-specified forces passively. Nevertheless, real-world living creatures are self-driven — their movements are enabled by embedded muscles given a desired kinematic motion target. Such an active movement mechanism is much more difficult to obtain as it inevitably involves frictional contacts between the model and the environment. To this end, we incorporate additional physical expressions into the Gaussian elements, enabling them to be self-driven in a physics-based way. These elements can change their appearance at each time step and approximate user-specified motion targets interacting with the environment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 261, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 261, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 261, + 399 + ], + "type": "text", + "content": "3.2. Hybrid Mesh-Gaussian Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": " denote the Gaussian elements of an object of interest, which is segmented from the scene " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{scene}}" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": ". We first create an encapsulating tetrahedral mesh " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = (\\mathcal{V}, \\mathcal{F})" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": " to approximate the object's overall volume, mass, and deformation. Here, " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": " refer to the sets of nodal points and faces in " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": ". To achieve this segmentation, we approximate the nearby environment with planes (more in Secs. 3.5 and 3.6). The plane " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": " expressed as a unit normal vector " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": " and its distance " + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 403, + 296, + 511 + ], + "type": "text", + "content": " from the origin:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 520, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 520, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 97, + 520, + 295, + 534 + ], + "type": "interline_equation", + "content": "P (\\mathbf {n}, d) = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} \\mid \\mathbf {N} ^ {T} \\mathbf {x} + d = 0 \\right\\}, \\tag {1}", + "image_path": "0dd1472696f16e229c7cabb01b7727e32ef66e9ed3b008e8eb03e548dd178254.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 542, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 601 + ], + "type": "text", + "content": "is used to \"cut out\" irrelevant Gaussians, reduce the blurred regions at the cut, and serve as a prior for inpainting the missing parts after segmentation. Additionally, it is utilized for collision detection to determine the collision forces acting on " + }, + { + "bbox": [ + 55, + 542, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 55, + 542, + 295, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "text", + "content": " determines the states of Gaussian kernels in " + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "text", + "content": " after deformation. The deformed position of a point " + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "text", + "content": " originally at " + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "text", + "content": " within a tetrahedron is represented by the barycentric coordinates of the four vertices of the tetrahedron " + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{v}'" + }, + { + "bbox": [ + 55, + 602, + 296, + 650 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 656, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 656, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 136, + 656, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = \\sum_ {i} \\lambda_ {i} (\\mathbf {x}) \\mathbf {v} _ {i} ^ {\\prime}. \\tag {2}", + "image_path": "1d2df5467945360b7f4b116fe7f5669d3cae9f26c64ce70481f53e0c4a25e39a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "After deformation, " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " provides a global displacement map for all Gaussians in " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ". However, the internal rotation and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "scaling of the Gaussians cannot be accurately described by position-based interpolation, resulting in spiky artifacts. Additionally, the tetrahedral mesh is a coarse mesh. Its faces have significant gaps from the boundaries of the object, and it does not accurately describe the collision between " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " and the plane. To address these issues, we establish a bidirectional local embedding within the Gaussian around the mean " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " to precisely describe its influence range. This is determined by the endpoints of the ellipsoid corresponding to the Gaussian " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "[\\mathbf{v}_1,\\mathbf{v}_2,\\mathbf{v}_3]" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 201, + 553, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 201, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 326, + 201, + 553, + 237 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {T} ^ {\\pm} = [ \\mathbf {m}, \\mathbf {v} _ {1}, \\mathbf {v} _ {2}, \\mathbf {v} _ {3} ] \\\\ = [ \\mathbf {m}, \\mathbf {m} \\pm \\gamma s _ {1} \\mathbf {r} _ {1}, \\mathbf {m} \\pm \\gamma s _ {2} \\mathbf {r} _ {2}, \\mathbf {m} \\pm \\gamma s _ {3} \\mathbf {r} _ {3} ], \\tag {3} \\\\ \\end{array}", + "image_path": "3b706c9ec9d5f3e00e1dbdea189e3d1a3f53a6e6980e96f82c24e461f2637f47.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "text", + "content": "where the relevant parameters can be obtained from the scaling matrix " + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathrm{diag}(s_1, s_2, s_3)" + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "text", + "content": " and the rotation Matrix " + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = [\\mathbf{r}_1, \\mathbf{r}_2, \\mathbf{r}_3]" + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "text", + "content": " is a parameter that describes the range of influence, and we set its default value as 3.0. Each Gaussian has two directional embeddings, " + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{+}" + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{-}" + }, + { + "bbox": [ + 313, + 238, + 554, + 308 + ], + "type": "text", + "content": ", and each contains four points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 309, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 554, + 440 + ], + "type": "text", + "content": "We design bidirectional local embeddings for several purposes. For collision processing, it is used to detect the extent to which an object collides with the environmental plane; and for segmentation, it detects whether the Gaussian is cut by the plane and establishes a loss term to penalize Gaussian kernels that intersect the plane under initial setting; For deformation, it is used to accurately describe the scaling and rotation of the deformed Gaussian kernel. Finally, based on the angle between two directions, it determines whether the deformed Gaussian kernel is sufficient to cover the radiance field." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 486, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 486, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 486, + 461 + ], + "type": "text", + "content": "3.3. Soft Body Locomotion on Plane" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 467, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 525 + ], + "type": "text", + "content": "The core of our pipeline is an active locomotion solver for " + }, + { + "bbox": [ + 313, + 467, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 467, + 554, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 467, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 313, + 467, + 554, + 525 + ], + "type": "text", + "content": ". This is in the form of an inverse problem involving high-dimension inequalities constraints for processing frictional contacts between a 3DGS object and the environment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 533, + 451, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 451, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 451, + 544 + ], + "type": "text", + "content": "3.3.1. Muscle-Driven Dynamics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "text", + "content": "We embed muscle fibers within " + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "text", + "content": " to serve as the primary drivers of motion. At each time step, the nodal positions of " + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "text", + "content": " nodes in " + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 548, + 554, + 596 + ], + "type": "text", + "content": ", is determined through numerical integration of the dynamic equations of motion:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 362, + 606, + 553, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 606, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 362, + 606, + 553, + 619 + ], + "type": "interline_equation", + "content": "\\mathbf {M} \\ddot {\\mathbf {p}} = \\mathbf {f} _ {e x t} + \\mathbf {f} _ {i n t} + \\mathbf {f} _ {d} + \\mathbf {f} _ {m} + \\mathbf {f} _ {c}, \\tag {4}", + "image_path": "9798bbf0ce8afc762d21c1d4567fdfe7f290594baebc91737c6acdaedd5316ad.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " is the mass matrix, " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ext},\\mathbf{f}_{int},\\mathbf{f}_d,\\mathbf{f}_c,\\mathbf{f}_m" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " represent external, internal, damping, muscle, and contact forces, respectively. We use " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " for velocity and " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " for acceleration. The external force " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ext}" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " includes e.g., the gravity force. The internal force " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{int}" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", also known as the elastic force, is the resultant force generated by the material's elastic properties that resist deformation. In our implementation, this force" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "898" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 553, + 205 + ], + "type": "image", + "image_path": "cf42db0c67c03a6c9d9f2550c30d291b96a5824a96ffd9352abf68fae674a06a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 249 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 249 + ], + "type": "text", + "content": "Figure 3. An overview of EnliveningGS pipeline. Our hybrid mesh-Gaussian representation plays an important role in all tasks within the pipeline. Our locomotion solver inversely computes the muscle activation embedded in the segmented 3DGS object that leads to user-specific motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 295, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 284 + ], + "type": "text", + "content": "is determined as the negative gradient of the Neo-Hookean energy [5]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 292, + 295, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 292, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 94, + 292, + 295, + 316 + ], + "type": "interline_equation", + "content": "\\Psi = \\frac {\\mu}{2} (I _ {C} - 3) - \\mu \\ln J + \\frac {\\lambda}{2} (\\ln J) ^ {2}. \\tag {5}", + "image_path": "76f1c1da6572bb3ae7ec8ed339a41c97a99df2d422ebf356b61e245043586779.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "text", + "content": " are the Lamé parameters. " + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "inline_equation", + "content": "I_{C} = \\operatorname{tr}(\\mathbf{F}^{T}\\mathbf{F})" + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "inline_equation", + "content": "J = \\det(\\mathbf{F})" + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "text", + "content": " are computed based on the deformation gradient " + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 324, + 296, + 371 + ], + "type": "text", + "content": ". We use the Rayleigh damping for the damping force:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 374, + 295, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 374, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 129, + 374, + 295, + 388 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {d} = \\left(\\alpha \\mathbf {M} + \\beta \\mathbf {K} _ {0}\\right) \\dot {\\mathbf {p}}, \\tag {6}", + "image_path": "942898bc8a5f7fdff7d3a6b60557888883b6413204d69cf2420ba8b721aa1fae.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 393, + 233, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 233, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 233, + 404 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 393, + 233, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_0" + }, + { + "bbox": [ + 55, + 393, + 233, + 404 + ], + "type": "text", + "content": " is the rest-shape stiffness matrix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "text", + "content": "We consider muscle fibers as polygonal curves with " + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "text", + "content": " segments. The spring force caused by each segment is computed as: " + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "inline_equation", + "content": "f = ka" + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "text", + "content": " is the stiffness of the muscle fiber and " + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "text", + "content": " denotes the change in the segment's length. When a muscle segment activates, the spring force will shorten or bend the body around it. Muscle force on the whole soft body is encoded in a pose-dependent activation matrix " + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{3N \\times M}" + }, + { + "bbox": [ + 55, + 405, + 296, + 499 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 145, + 502, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 502, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 145, + 502, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {m} = \\mathbf {A} (\\mathbf {p}) \\mathbf {a}, \\tag {7}", + "image_path": "ee49123938e1f3a4b41016460614d8ae3c4377d87bf82d97c335aea7195a6a31.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": " is the vector of activations of all the muscle segments. Please refer to the supplementary materials for a complete description of the calculations involving activation matrix " + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "We provide an automatic muscle generation tool given " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", which uses the medial axis transform [8, 29] of the model. Four longitudinal muscles are placed parallel near the medial axis at a given radius, and radial muscles are placed at given intervals. Users can also customize the muscle setup with preferred arrangements inside the object to achieve different movements. For instance, by implanting several longitudinal muscle fibers that run the length of the body, the body can be shortened when these muscles contract. Bending the body can be achieved by contracting these muscles unevenly on one side. Additionally, circular muscles wrap around the body, enabling it to twist, while" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 258, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 283 + ], + "type": "text", + "content": "radial muscles extend across the body's cross-section, allowing the body to elongate while preserving its volume." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 287, + 418, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 418, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 418, + 299 + ], + "type": "text", + "content": "3.3.2. Contacting 3DGS" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "text", + "content": "In this section we present a novel contact model for 3DGS. For Gaussian kernels in " + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "text", + "content": ", we first define " + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "text", + "content": " as the position of points in bidirectional local embedding " + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\pm}" + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "text", + "content": " below the plane as " + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^T\\mathbf{u} + d\\leq 0" + }, + { + "bbox": [ + 313, + 302, + 554, + 361 + ], + "type": "text", + "content": ". Non-interpenetration constraints on 3DGS is thus defined as inequalities:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 391, + 368, + 553, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 368, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 391, + 368, + 553, + 382 + ], + "type": "interline_equation", + "content": "C _ {n} (\\mathbf {u}) \\equiv \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} \\geq \\mathbf {0}. \\tag {8}", + "image_path": "0627023d9abcf0ec28f6e2131fd522566f1e7f966f0b946a92fd78b88b3ee557.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "text", + "content": "We decompose contact forces on the Gaussian as " + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_g = \\mathbf{N}f_{\\perp} + \\mathbf{D}\\mathbf{f}_{\\parallel}" + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "text", + "content": " is a set of directions at the contact point approximating the friction cone, and they are perpendicular to the plane normal " + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "inline_equation", + "content": "f_{\\perp}" + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\parallel}" + }, + { + "bbox": [ + 313, + 388, + 554, + 449 + ], + "type": "text", + "content": " are the magnitudes of normal and tangent forces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 449, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 471 + ], + "type": "text", + "content": "LCP formulation [2] that regulates the contact velocity and contact force is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 475, + 553, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 475, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 350, + 475, + 553, + 516 + ], + "type": "interline_equation", + "content": "\\mathbf {0} \\leq \\left[ \\begin{array}{l} f _ {\\perp} \\\\ \\mathbf {f} _ {\\parallel} \\\\ \\lambda \\end{array} \\right] \\perp \\left[ \\begin{array}{c} \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} \\\\ \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} + \\mathbf {E} \\lambda \\\\ \\mu f _ {\\perp} - \\mathbf {E} ^ {T} \\mathbf {f} _ {\\parallel} \\end{array} \\right] \\geq \\mathbf {0}, \\tag {9}", + "image_path": "a27113ca7ab095324327f195bf8b4be71ef629b470877b08cc9d9b9d729b4280.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": " is the friction coefficient, and " + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": " is an auxiliary variable (e.g., the Lagrange multiplier) that relates the tangent velocity to a sliding contact. " + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": " is a block-diagonal matrix of " + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": ". LCP condition notation " + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\perp \\mathbf{y}" + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^T\\mathbf{y} = 0" + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": ". The first line of Equation (9) is a complementarity formulation of contact that enforces the non-penetration condition, and the following two lines encode Coulomb's friction law." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 654 + ], + "type": "text", + "content": "Like [33], contact can be viewed as a regularization of the complementarity form of the non-penetration constraint, which gives penalty forms of contact as spring force " + }, + { + "bbox": [ + 313, + 605, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_g^{spring}" + }, + { + "bbox": [ + 313, + 605, + 554, + 654 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 662, + 553, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 662, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 369, + 662, + 553, + 677 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\perp} ^ {s p r i n g} \\equiv - k _ {n} \\min \\left(0, C _ {n} (\\mathbf {u})\\right), \\tag {10}", + "image_path": "69787158e76a21b12baab7711a2db15586f0d2f6cca6fdd213541542e0d08117.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 338, + 689, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 689, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 338, + 689, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\parallel} ^ {s p r i n g} \\equiv - \\min \\left(k _ {f}, \\mu \\frac {\\| \\mathbf {f} ^ {s p r i n g} \\|}{\\| \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} \\|}\\right) \\mathbf {D} ^ {T} \\dot {\\mathbf {u}}, \\tag {11}", + "image_path": "f0467dd790cac88addb73785c273e16fff955de0b6b510e03fdb5400640ae130.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "899" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "k_{n}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "k_{f}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " control the stiffness of the contact. Forces on local embeddings can be converted back onto points on " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " using the same weights in Equation (2), giving us " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c^{spring}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 124, + 180, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 180, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 180, + 136 + ], + "type": "text", + "content": "3.3.3. Two-stage Locomotion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 139, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 296, + 198 + ], + "type": "text", + "content": "Unlike force-based passive dynamics, we achieve the desired locomotion mainly by intuitively prescribing the motion trajectory of the object. For the position of a target point " + }, + { + "bbox": [ + 55, + 139, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\star}" + }, + { + "bbox": [ + 55, + 139, + 296, + 198 + ], + "type": "text", + "content": ", the objective function that constrains the position can be written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 202, + 295, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 202, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 117, + 202, + 295, + 217 + ], + "type": "interline_equation", + "content": "G _ {\\text {p o s i t i o n}} = \\left\\| f (\\mathbf {p}) - \\mathbf {p} ^ {\\star} \\right\\| ^ {2}, \\tag {12}", + "image_path": "49248d4d12d16d88c1e272a30e723cd1e3bd2d97c0bab09662d97cd5424438f8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "text", + "content": " can be functions that compute the positional target of the object. For instance, a commonly used setup is " + }, + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "inline_equation", + "content": "f_{COM}" + }, + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "text", + "content": ", which calculates the COM (center of mass) " + }, + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = f_{COM}(\\mathbf{p})" + }, + { + "bbox": [ + 55, + 222, + 296, + 304 + ], + "type": "text", + "content": " of the body, of local COM coordinates (foot, head, base, etc.). Similar objective functions can be used to track the velocity, acceleration, angular momentum, base area, etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "text", + "content": "We first solve Equation (4) by substituting " + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c" + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c^{spring}" + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "text", + "content": " and temporarily keeping muscle activation a same as the last time step. Based on the resulting " + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{p}}" + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "text", + "content": ", we convert the state of contact points back into the inequalities in Equation (9), and solve a Quadratic Programming (QP) problem to find the best " + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 306, + 296, + 389 + ], + "type": "text", + "content": ". For example, if a contact point is in static contact, the corresponding QP is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 394, + 205, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 394, + 205, + 413 + ], + "spans": [ + { + "bbox": [ + 113, + 394, + 205, + 413 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {a}, f _ {\\perp}, \\mathbf {f} _ {| |}, \\lambda} G \\left(\\mathbf {a}, f _ {\\perp}, \\mathbf {f} _ {| |}\\right)", + "image_path": "a3c60038b3e1420a8623eec07ef806e9077070849ecce8770d4036c8c5b7207b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 417, + 155, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 417, + 155, + 429 + ], + "spans": [ + { + "bbox": [ + 112, + 417, + 155, + 429 + ], + "type": "text", + "content": "subject to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 112, + 431, + 295, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 431, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 112, + 431, + 295, + 445 + ], + "type": "interline_equation", + "content": "\\mathbf {0} \\leq f _ {\\perp}, \\quad \\mathbf {N} ^ {T} \\dot {\\mathbf {u}} ^ {n + 1} = \\mathbf {0} \\tag {13}", + "image_path": "b3bc3de52be470f4fa03563f6417de40113d58ceb2598614e10a17e5989b8594.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 448, + 238, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 448, + 238, + 463 + ], + "spans": [ + { + "bbox": [ + 113, + 448, + 238, + 463 + ], + "type": "interline_equation", + "content": "\\mathbf {0} \\leq \\mathbf {f} _ {\\parallel}, \\quad \\mathbf {D} ^ {T} \\dot {\\mathbf {u}} ^ {n + 1} + \\mathbf {E} \\lambda = \\mathbf {0}", + "image_path": "2280fdfa08b83d35f3f74a3064408676a94d37245dc76a651b45c772cfa01ae1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 465, + 228, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 465, + 228, + 479 + ], + "spans": [ + { + "bbox": [ + 113, + 465, + 228, + 479 + ], + "type": "interline_equation", + "content": "\\mathbf {0} = \\lambda , \\quad \\mu f _ {\\perp} - \\mathbf {E} ^ {T} \\mathbf {f} _ {\\parallel} \\geq \\mathbf {0}.", + "image_path": "40925333c440390ac7f793790406d693f8b13bc4e107de92221b52653b64696d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 483, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 520 + ], + "type": "text", + "content": "Note that all contact points may have their own states. If the results are not satisfactory, we will use the pivoting method as in [41] to find the appropriate contact states." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 525, + 282, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 282, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 282, + 536 + ], + "type": "text", + "content": "3.4. Gaussian Deformation in Tetrahedral Mesh" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": "The barycentric coordinates of the points in " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{+}" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " are calculated in the reference configuration, and the deformed coordinates in the current configuration " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^+ = [\\mathbf{m}',\\mathbf{v}_1',\\mathbf{v}_2',\\mathbf{v}_3']" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " are then calculated at each timestep to describe the deformation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "text", + "content": "After that, we construct two matrices " + }, + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_m" + }, + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_s" + }, + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "text", + "content": ", representing the differences in vertex coordinates in the reference and current configurations, respectively:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 82, + 642, + 295, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 642, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 82, + 642, + 295, + 662 + ], + "type": "interline_equation", + "content": "\\begin{array}{r c l} \\mathbf {D} _ {m} & = & \\left[ \\mathbf {v} _ {1} ^ {0} - \\mathbf {m}, \\mathbf {v} _ {2} ^ {0} - \\mathbf {m}, \\mathbf {v} _ {3} ^ {0} - \\mathbf {m} \\right] \\\\ \\mathbf {D} _ {m} & = & \\left[ \\mathbf {v} _ {1} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\mathbf {v} _ {2} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\mathbf {v} _ {3} ^ {\\prime} - \\mathbf {m} ^ {\\prime} \\right] \\end{array} \\tag {14}", + "image_path": "c577cdd286fb73b34fea8287a3db63fd866f1da426e78eb1888b366503ab2f98.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 85, + 654, + 252, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 654, + 252, + 668 + ], + "spans": [ + { + "bbox": [ + 85, + 654, + 252, + 668 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {s} = \\left[ \\begin{array}{c} \\mathbf {v} _ {1} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\bar {\\mathbf {v}} _ {2} ^ {\\prime} - \\mathbf {m} ^ {\\prime}, \\bar {\\mathbf {v}} _ {3} ^ {\\prime} - \\mathbf {m} ^ {\\prime} \\end{array} \\right].", + "image_path": "2d4d4244b5a52dbd08a7fedea330324e9049c731b75c7d2d88aff2ce7bf16e41.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "text", + "content": "The linear transformation " + }, + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "text", + "content": " that maps the reference configuration to the current configuration can be calculated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 700, + 208, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 700, + 208, + 714 + ], + "spans": [ + { + "bbox": [ + 141, + 700, + 208, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\mathbf {D} _ {s} \\cdot \\mathbf {D} _ {m} ^ {- 1}.", + "image_path": "5f9bf831f586d661ec7d2c359feb8146fa144b13a25d65de7417d68a48b14aa2.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "The deformed covariance becomes: " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime} = \\mathbf{F}\\Sigma \\mathbf{F}^{T}" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " and is converted into scaling and rotation by performing singular value decomposition (SVD) on " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 372, + 117, + 553, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 117, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 372, + 117, + 553, + 133 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {+} = \\operatorname {d i a g} \\left(\\sqrt {\\operatorname {S V D} _ {\\Sigma} \\left(\\Sigma^ {\\prime}\\right)}\\right) \\tag {15}", + "image_path": "6eff081f5887395014e3824eb868e69da0c5390e033d6617576b046d93a6e1db.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 372, + 129, + 463, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 129, + 463, + 142 + ], + "spans": [ + { + "bbox": [ + 372, + 129, + 463, + 142 + ], + "type": "interline_equation", + "content": "\\mathbf {R} ^ {+} = \\mathsf {S V D} _ {U} (\\Sigma^ {\\prime}).", + "image_path": "e44628567664347bf1c39d880d0a98f8c9692fd6efe0fca0b97d88d780fd9e37.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "text", + "content": "Note that if " + }, + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "inline_equation", + "content": "\\operatorname{det}(\\mathbf{R}^{+}) = -1" + }, + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "text", + "content": ", one needs to negate its last column to ensure it is a valid rotation matrix in " + }, + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "text", + "content": ". It is also necessary to adjust the spherical harmonics' orientation by applying rotation " + }, + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "inline_equation", + "content": "R_{SH} = \\mathsf{SVD}_U(\\Sigma')\\mathsf{SVD}_V(\\Sigma')^T" + }, + { + "bbox": [ + 313, + 152, + 554, + 212 + ], + "type": "text", + "content": " to display the correct colors from various angles." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": "The points in local embeddings represent the shape of the deformed Gaussian kernel but they do not necessarily belong to the same tetrahedron. For some larger Gaussian kernels, significant rotations may cause them to extend beyond the original radiance field's boundaries. To avoid this situation, we use the reverse embedding " + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{-}" + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": " and use the aforementioned method to obtain its corresponding rotation matrix " + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{-}" + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": ". The angle between " + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{+}" + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{-}" + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 363, + 327, + 553, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 363, + 327, + 553, + 360 + ], + "type": "interline_equation", + "content": "\\theta = \\arccos \\left(\\frac {\\operatorname {t r} \\left(\\mathbf {R} ^ {+ T} \\mathbf {R} ^ {-}\\right) - 1}{2}\\right). \\tag {16}", + "image_path": "741fa28086cb9fb1d8ce9d00d38514c226611fc3258cd4fbcb83072045362360.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 367, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 450 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 313, + 367, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 367, + 554, + 450 + ], + "type": "text", + "content": " exceeds the threshold " + }, + { + "bbox": [ + 313, + 367, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\theta >\\eta" + }, + { + "bbox": [ + 313, + 367, + 554, + 450 + ], + "type": "text", + "content": ", it indicates that the region covered by the Gaussian has transformed from a plane to a curved surface. A single kernel can no longer represent the deformed area accurately; thus, this Gaussian will be split into two from the center. Figure 2 (b) provides a schematic of our splitting process, while Figure 8 presents the corresponding experimental results." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "text", + "content": "Inspired by [9, 15], the splitting of a Gaussian " + }, + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "inline_equation", + "content": "(\\alpha_0,\\mu_0,\\Sigma_0)" + }, + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "text", + "content": " into two new Gaussian kernels " + }, + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "inline_equation", + "content": "(\\alpha_{l},\\mu_{l},\\Sigma_{l})" + }, + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "inline_equation", + "content": "(\\alpha_{r},\\mu_{r},\\Sigma_{r})" + }, + { + "bbox": [ + 314, + 451, + 554, + 487 + ], + "type": "text", + "content": " at the center is calculated by:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 334, + 495, + 401, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 495, + 401, + 515 + ], + "spans": [ + { + "bbox": [ + 334, + 495, + 401, + 515 + ], + "type": "interline_equation", + "content": "\\alpha_ {l} = \\alpha_ {r} = \\frac {\\alpha_ {0}}{2},", + "image_path": "2da44ba23aeb81351dbd34efb3ec8810cc2679bd1b957e10e584ee87375d33e5.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 334, + 517, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 517, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 334, + 517, + 553, + 539 + ], + "type": "interline_equation", + "content": "\\mu_ {l} = \\mu_ {0} - \\frac {\\mathbf {v} _ {k} - \\mu_ {0}}{\\kappa}, \\quad \\mu_ {r} = \\mu_ {0} + \\frac {\\mathbf {v} _ {k} - \\mu_ {0}}{\\kappa}, \\tag {17}", + "image_path": "1f99812d121f9d0e8c57b18158dfdad7aeb35794f62ca76b333bcbc568c3f0ae.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 333, + 540, + 505, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 540, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 333, + 540, + 505, + 565 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {l} = \\boldsymbol {\\Sigma} _ {r} = \\boldsymbol {\\Sigma} _ {0} - \\frac {(\\mathbf {v} _ {k} - \\mu_ {0}) (\\mathbf {v} _ {k} - \\mu_ {0}) ^ {T}}{\\kappa^ {2}}.", + "image_path": "c7e54aef371f5fd084b7b4bc5b1feecb2c7093e546a53f3c6ebbfd78c041af64.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "text", + "content": "In the above, " + }, + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_k \\in \\mathbf{T}^+" + }, + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "text", + "content": " stands for the endpoint on main principle axis with the largest scaling factor " + }, + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "text", + "content": ". Factor " + }, + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\kappa = \\frac{2\\gamma}{\\sqrt{2\\pi}}" + }, + { + "bbox": [ + 313, + 573, + 554, + 635 + ], + "type": "text", + "content": " is derived by maximizing the similarity between each new Gaussian and their own part of the original. We refer readers to supplementary material for the complete derivation." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 641, + 417, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 417, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 417, + 653 + ], + "type": "text", + "content": "3.5. Plane Estimation" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 658, + 553, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 553, + 683 + ], + "type": "text", + "content": "To find the parameters " + }, + { + "bbox": [ + 313, + 658, + 553, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 313, + 658, + 553, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 658, + 553, + 683 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 658, + 553, + 683 + ], + "type": "text", + "content": " of a plane in a 3DGS scene, we first construct a depth map on the plane:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 363, + 690, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 690, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 363, + 690, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {p} (u, v) = - \\frac {\\mathbf {N} \\cdot \\mathbf {t} _ {c} + d}{\\mathbf {n} ^ {T} \\mathbf {R} _ {c} ^ {T} \\mathbf {K} ^ {- 1} \\mathbf {p} (u , v)}, \\tag {18}", + "image_path": "f2e70a545ebdec2a39001365a21249351bdd73159b375858dba8ab0cd8c26800.jpg" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "900" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{p}(u,v)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " represents the pixel coordinates in homogeneous form " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "[u,v,1]^T" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_c" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " are the rotation and translation extracted from the camera's extrinsic matrix, and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " denotes the camera's intrinsic matrix. Initially, " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is set to camera's look-at direction and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is set to zero to ensure solvability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 180 + ], + "type": "text", + "content": "By minimizing the error between the depth map within the region and the GS depth map " + }, + { + "bbox": [ + 55, + 144, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 55, + 144, + 294, + 180 + ], + "type": "text", + "content": ", the plane parameters can be optimized by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 185, + 295, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 185, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 91, + 185, + 295, + 211 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\mathbf {N}, d} \\sum_ {(u, v) \\in \\mathcal {M}} \\| \\mathcal {D} _ {p} (u, v) - \\mathcal {D} _ {g} (u, v) \\|, \\tag {19}", + "image_path": "fcdda4ecdf8c0d65f5661848e62024281e841817a5e37f13ae6af5cc867bc679.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 294, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 294, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 213, + 294, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 213, + 294, + 250 + ], + "type": "text", + "content": " is a mask of the planar region which can be obtained through image-based segmentation algorithms [28] or roughly drawn by the user." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 250, + 294, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 294, + 297 + ], + "type": "text", + "content": "The user can repeat this step on multiple images and optimize simultaneously to obtain fine-tuned plane parameters. In practice, a single viewpoint is usually sufficient to estimate an accurate plane." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 261, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 261, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 261, + 316 + ], + "type": "text", + "content": "3.6. Object Segmentation with Plane Priors" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 321, + 295, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 295, + 428 + ], + "type": "text", + "content": "We categorize the Gaussian scene into two types: dynamic splitting and static splitting. Equation (17) showcases dynamic online splitting to enhance the quality during the deformation process and ensure the speed of splitting on animating objects. For a Gaussian model, since the segmentation plane of the object and environment is already determined, it is more suitable to incorporate the plane information during training to regulate the Gaussian parameters at the cutting plane." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 429, + 295, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 487 + ], + "type": "text", + "content": "To perform static splitting on the training stage, we first determine whether the Gaussian intersects with the plane. This requires checking if all the endpoints in " + }, + { + "bbox": [ + 55, + 429, + 295, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\pm}" + }, + { + "bbox": [ + 55, + 429, + 295, + 487 + ], + "type": "text", + "content": " are on the same side of the plane. We project center-to-endpoints distance onto the plane normal:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 494, + 294, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 494, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 113, + 494, + 294, + 513 + ], + "type": "interline_equation", + "content": "\\eta_ {s e g} = \\max _ {\\mathbf {v} _ {i} \\in \\mathbf {T} ^ {\\pm}} | \\mathbf {N} ^ {T} (\\mathbf {v} _ {i} - \\mathbf {m}) |. \\tag {20}", + "image_path": "82d499709228ac884e2f4b63aaec8554b7ad0f9173569799df26ba9d39bb5855.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "type": "text", + "content": "If the distance from Gaussian kernel to the plane " + }, + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "type": "inline_equation", + "content": "|d_0| = |P(\\mathbf{n}, d)(\\mathbf{m})| < \\eta_{seg}" + }, + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "type": "text", + "content": ", we will constrain the Gaussian's parameter or split the Gaussian with the plane." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "type": "text", + "content": "We fine-tune the trained Gaussian model by introducing the following non-penetration regularization term to ensure the model's influence range does not cross the plane:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 594, + 294, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 594, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 127, + 594, + 294, + 608 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e g} = \\left| \\left| \\eta_ {s e g} - \\left| d _ {0} \\right| \\right| \\right| _ {2} \\tag {21}", + "image_path": "e0e549d328736bceae0a3231595b62115b54b6fe77c2b992bba770868f3786ad.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 612, + 209, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 209, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 209, + 625 + ], + "type": "text", + "content": "3.7. Inpainting 3D Object on GS" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "After extracting the object " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": ", a hole is left in the background " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text {scene }}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": ". To address this issue, we introduce the depth map of the plane in Equation (18) into the depth-based inpainting method [30] and use a 2D inpainting tool [40] to obtain the color of the missing parts from the current viewpoint. Subsequently, we fine-tune the filled patch using the inpainted single-view image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 394, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 394, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 394, + 84 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 91, + 553, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 199 + ], + "type": "text", + "content": "We implemented our pipeline on a workstation with a single NVIDIA RTX4090 GPU and Intel 14900K CPU. We used PyTorch for storing and editing 3DGS on GPU, and OSQP [37] for solving QP in Equation (13). Our method supports the original 3DGS implementation [25]. For better visual quality on the planar surface, we set the minimum scaling factor on Gaussian kernels to " + }, + { + "bbox": [ + 313, + 91, + 553, + 199 + ], + "type": "inline_equation", + "content": "\\epsilon = e^{-10}" + }, + { + "bbox": [ + 313, + 91, + 553, + 199 + ], + "type": "text", + "content": ", similar to 2DGS[19]. More benchmarks for the locomotion examples are reported in the supplementary document." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 206, + 427, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 427, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 427, + 219 + ], + "type": "text", + "content": "4.1. Locomotion Design" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 223, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 295 + ], + "type": "text", + "content": "Walking. We achieve walking by controlling the trajectories of all four legs. The diagonal support legs simultaneously lift the body, while the remaining two legs swing forward with the body's movement. As shown in Figure 4, the wooden stool moves one position to the right by taking four steps." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 300, + 554, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 420 + ], + "type": "text", + "content": "Jumping. Jumping soft-bodied characters are more commonly seen in cartoons and animations. Jumping is a complex action because the contact state between the object and the ground changes rapidly and significantly. To achieve continuous jumping, it is necessary to precisely control the position of the COM and the relative velocity of the center of the base to jump the appropriate distance and maintain balance. We have designed an animation of a game of chess, where the soft-bodied chess pieces move across the board by continuously jumping, as shown in Figure 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 425, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 496 + ], + "type": "text", + "content": "Twisting. Figure 5 illustrates an example of the Peashooter doll twisting, jumping, and rotating 180 degrees. Based on the jumping motion, we additionally control the change in angular momentum along the central axis, allowing the Peashooter to rotate its body while airborne and face the zombie upon landing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 503, + 498, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 498, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 498, + 516 + ], + "type": "text", + "content": "4.2. Comparison I. Deformation on GS" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "type": "text", + "content": "We compare our method with the current state-of-the-art physico-aware GS methods. Figure 6 shows a qualitative comparison, where our method demonstrates better results under large deformation. Table 1 presents the related quantitative evaluation: we apply bending, twisting, and falling on the ground deformations to the ficus, chair, and mic examples from the NeRF synthetic dataset [34], and compare them with the rendered images of the source 3D mesh subjected to the same deformations as the ground truth." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 636, + 517, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 517, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 517, + 648 + ], + "type": "text", + "content": "4.3. Comparison II. Two-stage Locomotion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "To evaluate our two-stage locomotion solver, we conducted frame-by-frame tests on an animation sequence with given muscle activation. This animation sequence includes 50 frames, 60 variables, and 50 pairs of linear complementarity constraints. We compared our solution with a QP" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "901" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 142 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 142 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 142 + ], + "type": "image", + "image_path": "be7f40502de4eb5fd1d0cf95dfa904018e5ff900e922910a6d7994978e6acc9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 149, + 464, + 161 + ], + "lines": [ + { + "bbox": [ + 145, + 149, + 464, + 161 + ], + "spans": [ + { + "bbox": [ + 145, + 149, + 464, + 161 + ], + "type": "text", + "content": "Figure 4. Walking stool. In the bedroom scene, a wooden stool walks like a quadruped." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 164, + 553, + 236 + ], + "blocks": [ + { + "bbox": [ + 56, + 164, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 164, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 553, + 236 + ], + "type": "image", + "image_path": "4de9d7892211b4138d355b7a33f08eaf5c94581ddee23317007830d4ec57e737.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "text", + "content": "Figure 5. Peashooter vs. Zombie. On a cloth-covered coffee table, a Peashooter figurine contracts, jumps, and twists its body, landing after a 180-degree spin to face the zombie." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 275, + 292, + 439 + ], + "blocks": [ + { + "bbox": [ + 59, + 275, + 292, + 439 + ], + "lines": [ + { + "bbox": [ + 59, + 275, + 292, + 439 + ], + "spans": [ + { + "bbox": [ + 59, + 275, + 292, + 439 + ], + "type": "image", + "image_path": "526b1ceff8a22c32a277fca95150967414931650572624af579c3b89cc489d4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 449, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 483 + ], + "type": "text", + "content": "Figure 6. Visual quality comparison on deforming objects. Compared to PhysGaussian [45] and VR-GS [23], our method can still suppress spiky artifacts under large deformations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": "solver based directly on the static contact assumption and the QPCC solver proposed by [41], and reported the average best and worst results across all frames in Table 2. Our locomotion solver uses the results predicted by the penalty force as guidance, reducing the number of iterations and lowering the objective value." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 207, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 207, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 207, + 590 + ], + "type": "text", + "content": "4.4. Comparison III. Inpainting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "As shown in Figure 7, we demonstrate the inpainting effect on a plane using the garden example from Mip-NeRF360 dataset [4] and the truck example from DTU dataset [22], comparing it with the existing 3D object inpainting method on GS. In both examples, the removed objects occupy a significant portion of the photo, leaving a large hole on the ground after removal. Existing methods, due to the lack of relevant depth information, struggle to fill such a large missing area even after extensive finetuning of the inpainted 2D image (approximately 10,000 iterations). In contrast," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 275, + 364, + 361 + ], + "blocks": [ + { + "bbox": [ + 315, + 275, + 364, + 361 + ], + "lines": [ + { + "bbox": [ + 315, + 275, + 364, + 361 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 364, + 361 + ], + "type": "image", + "image_path": "3b8bf8c7662b9b530f14806fe8a7a93253f5c213cfa2eaa3352c52825e63b165.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 437 + ], + "type": "text", + "content": "Figure 7. Large 3D object inpainting. Comparison on large 3D object inpainting cases on the ground, where Gaussian Grouping [48] requires 20 minutes fine-tuning but still fails to fill the large hole on the ground while our method with better inpainting quality only needs 20 seconds fine-tuning. We show the render result as well as the depth." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 367, + 275, + 459, + 361 + ], + "blocks": [ + { + "bbox": [ + 367, + 275, + 459, + 361 + ], + "lines": [ + { + "bbox": [ + 367, + 275, + 459, + 361 + ], + "spans": [ + { + "bbox": [ + 367, + 275, + 459, + 361 + ], + "type": "image", + "image_path": "ee8c7d348c3552bc7c94b808a2e882413ffcf21a131b20b3e2d78ce1ff39ead0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 461, + 275, + 554, + 361 + ], + "blocks": [ + { + "bbox": [ + 461, + 275, + 554, + 361 + ], + "lines": [ + { + "bbox": [ + 461, + 275, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 461, + 275, + 554, + 361 + ], + "type": "image", + "image_path": "5f68b70a34949f5b6dd57f03f444a35da0bb0abdf2483d18eb41cb27ac41d958.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "type": "text", + "content": "our method leverages the plane as a prior, placing a dense and reasonably positioned point cloud in the missing area. This approach requires only about 100 iterations of finetuning, speeding up the inpainting process by several orders of magnitude and remaining unaffected by the size of the missing region." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 416, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 416, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 416, + 538 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 543, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 651 + ], + "type": "text", + "content": "Bidirectional local embedding We first use a checkered bar to showcase the effect of bidirectional local embedding as in Figure 8. The checkered bar twists increasingly from right to left, and the sparky artifacts become progressively more pronounced. Without the need to detect whether the Gaussian kernel exceeds the object's surface, our bidirectional local embedding automatically identifies the regions causing sparky artifacts based on the information within the Gaussian kernel. Table 1 also reports quantitative results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Non-penetration regularization Since the segmentation plane under " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " is occluded from any viewpoint, directly masking out the object from the environment results in significant blurring and penetration at the segmentation plane." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "902" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 70, + 527, + 164 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 527, + 164 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 527, + 164 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 527, + 164 + ], + "type": "table", + "html": "
BendingTwistingFalling
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
SuGaR [14]28.040.8900.04626.370.8230.10125.010.9130.053
PhysGaussian [45]28.140.9070.04326.530.8560.08925.100.9230.048
VR-GS [23]28.630.9170.03826.530.8930.06625.310.9380.044
Ours (w/o split)28.620.9180.03826.470.8910.06225.340.9390.042
Ours28.790.9230.03126.890.9100.04125.670.9510.032
", + "image_path": "601f37ed30a617aa560b2e8a37a0a45bb3d50141d48f85ad464406f2d8f5c2a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 193, + 288, + 261 + ], + "blocks": [ + { + "bbox": [ + 108, + 173, + 501, + 184 + ], + "lines": [ + { + "bbox": [ + 108, + 173, + 501, + 184 + ], + "spans": [ + { + "bbox": [ + 108, + 173, + 501, + 184 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of rendering quality after deformation on the NeRF synthetic dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 193, + 288, + 261 + ], + "lines": [ + { + "bbox": [ + 59, + 193, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 288, + 261 + ], + "type": "table", + "html": "
Iterations ↓Object value ↓
Avg.BestWorstAvg.BestWorst
QP---9.430.1623.86
QPCC [41]46.1915101.290.1412.11
Ours17.0113820.830.137.89
", + "image_path": "f1c8f94ea0c96bbb69305ec05d38adaec5832d5478e969412959beb7dc9e10a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 305, + 167, + 338 + ], + "blocks": [ + { + "bbox": [ + 59, + 305, + 167, + 338 + ], + "lines": [ + { + "bbox": [ + 59, + 305, + 167, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 305, + 167, + 338 + ], + "type": "image", + "image_path": "270088c2abe18f4b67809c5d0b707df7d4df7998780e0056c50e942458b20af4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 342, + 131, + 351 + ], + "lines": [ + { + "bbox": [ + 93, + 342, + 131, + 351 + ], + "spans": [ + { + "bbox": [ + 93, + 342, + 131, + 351 + ], + "type": "text", + "content": "Rest shape" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 184, + 299, + 292, + 344 + ], + "blocks": [ + { + "bbox": [ + 184, + 299, + 292, + 344 + ], + "lines": [ + { + "bbox": [ + 184, + 299, + 292, + 344 + ], + "spans": [ + { + "bbox": [ + 184, + 299, + 292, + 344 + ], + "type": "image", + "image_path": "c683898487a24bf74613ecd0e6983c757d8ac416af434b0c507523b8a3d3cae2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 346, + 253, + 351 + ], + "lines": [ + { + "bbox": [ + 222, + 346, + 253, + 351 + ], + "spans": [ + { + "bbox": [ + 222, + 346, + 253, + 351 + ], + "type": "text", + "content": "w/o split" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 354, + 167, + 396 + ], + "blocks": [ + { + "bbox": [ + 59, + 354, + 167, + 396 + ], + "lines": [ + { + "bbox": [ + 59, + 354, + 167, + 396 + ], + "spans": [ + { + "bbox": [ + 59, + 354, + 167, + 396 + ], + "type": "image", + "image_path": "1b13452fc052c610c093f662b7204b7ef682ac79110c38e6e4607fd66bc15e2c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 396, + 144, + 404 + ], + "lines": [ + { + "bbox": [ + 83, + 396, + 144, + 404 + ], + "spans": [ + { + "bbox": [ + 83, + 396, + 144, + 404 + ], + "type": "text", + "content": "Detected artifacts" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 415, + 295, + 458 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 458 + ], + "type": "text", + "content": "Figure 8. Ablation study on bidirectional local embedding. Our method captures the regions that produce spiky artifacts and removes them by splitting the Gaussian kernels to fit the radiance field." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 183, + 356, + 291, + 396 + ], + "blocks": [ + { + "bbox": [ + 183, + 356, + 291, + 396 + ], + "lines": [ + { + "bbox": [ + 183, + 356, + 291, + 396 + ], + "spans": [ + { + "bbox": [ + 183, + 356, + 291, + 396 + ], + "type": "image", + "image_path": "5ebd8957b8e9b6b23abec4c0da4d908c03474664a8252488c40cbe48fead5fa7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 396, + 252, + 405 + ], + "lines": [ + { + "bbox": [ + 225, + 396, + 252, + 405 + ], + "spans": [ + { + "bbox": [ + 225, + 396, + 252, + 405 + ], + "type": "text", + "content": "w/split" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 56, + 463, + 294, + 520 + ], + "blocks": [ + { + "bbox": [ + 70, + 270, + 279, + 281 + ], + "lines": [ + { + "bbox": [ + 70, + 270, + 279, + 281 + ], + "spans": [ + { + "bbox": [ + 70, + 270, + 279, + 281 + ], + "type": "text", + "content": "Table 2. Numerical experiments of locomotion solvers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 463, + 294, + 520 + ], + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 520 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓Residual ↓
Mask (w/o Lseg)---0.058
Remove25.9170.8270.211-
w/ Lseg35.2910.9450.0340.002
", + "image_path": "d9ddf6deff199238fea0af31ab12d22e1370ff0d66196f1586900581c5e3b28c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "text", + "content": "Table 3. Ablation study on non-penetration regularization: quantitative results." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "The Gaussian kernel that penetrates the plane contributes to the object's appearance, so a direct Remove operation would create holes in the object. Figure 9 shows the relevant qualitative comparison results. Our non-penetration regularization restricts the Gaussian from penetrating the plane, maintaining the integrity of the object's surface. In the quantitative experiments shown in Table 3, we placed the camera on the segmentation plane, compared the image quality of the object's upper half with full mask-out, and evaluated the residual of the lower half. The residuals are expressed as the ratio of the sum of pixel values in the lower half to those in the upper half." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 193, + 552, + 338 + ], + "blocks": [ + { + "bbox": [ + 317, + 193, + 552, + 338 + ], + "lines": [ + { + "bbox": [ + 317, + 193, + 552, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 552, + 338 + ], + "type": "image", + "image_path": "474bef880366a2ecc61aec97330eaefc99bd5a880e08a0e5e64680f38fda7c2f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "type": "text", + "content": "Figure 9. Ablation study on non-penetration regularization: qualitative results. With " + }, + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "type": "inline_equation", + "content": "L_{seg}" + }, + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "type": "text", + "content": " applied to Gaussian kernels penetrating the plane, the object retains its complete appearance near the cutting plane while removing the parts below the cutting plane." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 404, + 387, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 387, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 387, + 415 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 419, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 515 + ], + "type": "text", + "content": "In this paper, we present EnliveningGS, a novel pipeline that enables active locomotion for 3D models represented by 3DGS. We use a hybrid mesh embedding of 3DGS to better estimate the contact between a 3DGS model and the environment. Given a high-level motion goal, we find the optimal muscle activations driving the movement of the object's body. An LCP-penalty mixed optimizer is used to efficiently calculate the contact and friction force." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 525, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 609 + ], + "type": "text", + "content": "Limitation Currently, our method relies on plane-based simplification of the environment. When the surroundings consist of more complicated geometries, users may need to use more sophisticated approximation strategies to estimate the contact plane. It is also possible to use triangulated surfaces. Doint so could impose more computational overhead for the contact solver." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 613, + 554, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 703 + ], + "type": "text", + "content": "Acknowledgements The authors would like to thank the reviewers for their insightful comments. Yin Yang is partially supported by NSF 2301040. This work is also supported by NSF China (No. 62322209 and No. 62421003), the gift from Adobe Research, the XPLORER PRIZE, and the 100 Talents Program of Zhejiang University. The source code and data are available at https://gapszju.github.io/EnliveningGS." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "903" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Sheldon Andrews, Kenny Erleben, and Zachary Ferguson. Contact and friction simulation for computer graphics. In ACM SIGGRAPH 2022 Courses, pages 1-172, Vancouver British Columbia Canada, 2022. ACM. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "text", + "content": "[2] Mihai Anitescu and Florian A Potra. Formulating dynamic multi-rigid-body contact problems with friction as solvable linear complementarity problems. Nonlinear Dynamics, 14: 231-247, 1997. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Lijie Bai, John E Mitchell, and Jong-Shi Pang. On convex quadratic programs with linear complementarity constraints. Computational Optimization and Applications, 54(3):517-554, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 295, + 258 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 294 + ], + "type": "text", + "content": "[5] Javier Bonet and Richard D Wood. Nonlinear continuum mechanics for finite element analysis. Cambridge university press, 1997. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 339 + ], + "type": "text", + "content": "[6] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physico-based simulations. arXiv preprint arXiv:2409.05819, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 407 + ], + "type": "text", + "content": "[7] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "text", + "content": "[8] Hyeong In Choi, Sung Woo Choi, and Hwan Pyo Moon. Mathematical theory of medial axis transform. *pacific journal of mathematics*, 181(1):57-88, 1997. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 294, + 485 + ], + "type": "text", + "content": "[9] Qiyuan Feng, Gengchen Cao, Haoxiang Chen, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. A new split algorithm for 3d gaussian splatting. arXiv preprint arXiv:2403.09143, 2024.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "text", + "content": "[10] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Dynamic fluid synthesis with gaussian splatting. arXiv preprint arXiv:2401.15318, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "text", + "content": "[11] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 643 + ], + "type": "text", + "content": "[12] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based Gaussian Splatting for Real-time Large-scale Deformation. arXiv preprint arXiv:2402.04796, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[13] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. arXiv preprint arXiv:2403.14554, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 714 + ], + "type": "text", + "content": "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[15] Peter Hall, David Marshall, and Ralph Martin. Adding and subtracting eigenspaces with eigenvalue decomposition and singular value decomposition. Image and vision computing, 20(13-14):1009-1016, 2002. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[16] Fursan Hamad, Shreyas Giridharan, and Christian Moormann. A penalty function method for modelling frictional contact in mpm. Procedia Engineering, 175:116-123, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "text", + "content": "[17] David Harmon, Etienne Vouga, Breannan Smith, Rasmus Tamstorf, and Eitan Grinspun. Asynchronous contact mechanics. In ACM SIGGRAPH 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "text", + "content": "[18] Shoichi Hasegawa and Nobuaki Fujii. Real-time rigid body simulation based on volumetric penalty method. In 11th Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems, 2003. HAPTICS 2003. Proceedings., pages 326-332. IEEE, 2003. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 288, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 553, + 342 + ], + "type": "text", + "content": "[19] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers* '24, pages 1-11, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "type": "text", + "content": "[20] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4220-4230, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "type": "text", + "content": "[21] Sumit Jain, Yuting Ye, and C Karen Liu. Optimization-based interactive motion synthesis. ACM Transactions on Graphics (TOG), 28(1):1-12, 2009. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "text", + "content": "[22] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanaes. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "text", + "content": "[23] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, and Chenfanfu Jiang. VR-GS: A Physical Dynamics-Aware Interactive Gaussian Splitting System in Virtual Reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 2, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "text", + "content": "[24] Joaquim J. Judice and Fernanda M Pires. A block principal pivoting algorithm for large-scale strictly monotone linear complementarity problems. Computers & operations research, 21(5):587-596, 1994. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "text", + "content": "[27] Junggon Kim and Nancy S Pollard. Direct control of simulated nonhuman characters. IEEE Computer Graphics and Applications, 31(4):56-65, 2011. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "904" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[29] Lei Lan, Ran Luo, Marco Fratarcangeli, Weiwei Xu, Huamin Wang, Xiaohu Guo, Junfeng Yao, and Yin Yang. Medial Elastics: Efficient and Collision-Ready Deformation via Medial Axis Transform. ACM Transactions on Graphics, 39(3): 1-17, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[30] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. InFusion: Inpainting 3D Gaussians via Learning Depth Completion from Diffusion Prior. arXiv preprint arXiv:2404.11613, 2024. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "type": "text", + "content": "[31] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "text", + "content": "[32] Miles Macklin, Matthias Müller, and Nuttapong Chentanez. XPBD: Position-based simulation of compliant constrained dynamics. In Proceedings of the 9th International Conference on Motion in Games, pages 49-54, New York, NY, USA, 2016. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "text", + "content": "[33] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Tae-Yong Kim. Primal/dual descent methods for dynamics. In Computer Graphics Forum, pages 89-100. Wiley Online Library, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[34] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[35] Todd S Munson, Francisco Facchinei, Michael C Ferris, Andreas Fischer, and Christian Kanzow. The semismooth algorithm for large scale complementarity problems. INFORMS Journal on Computing, 13(4):294-311, 2001. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "text", + "content": "[36] Andrei-Ionut Stefancu, Silviu-Cristian Melenciuc, and Mihai Budescu. Penalty based algorithms for frictional contact problems. Buletinul Institutului Politehnic din lasi. Sectia Constructii, Architectura, 57(3):119, 2011. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 577 + ], + "type": "text", + "content": "[37] B. Stellato, G. Banjac, P. Goulart, A. Temporad, and S. Boyd. OSQP: an operator splitting solver for quadratic programs. Mathematical Programming Computation, 12(4): 637-672, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[38] D. E. Stewart and J. C. Trinkle. An Implicit Time-Stepping Scheme for Rigid Body Dynamics with Inelastic Collisions and Coulomb Friction. International Journal for Numerical Methods in Engineering, 39(15):2673-2691, 1996. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[39] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 575 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[41] Jie Tan, Greg Turk, and C. Karen Liu. Soft body locomotion. ACM Transactions on Graphics, 31(4):26:1-26:11, 2012. 2, 5, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "text", + "content": "[42] Min Tang, Dinesh Manocha, Miguel A Otaduy, and Ruofeng Tong. Continuous penalty forces. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 164, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 228 + ], + "type": "text", + "content": "[43] Alessandro Tasora, Dario Mangoni, Simone Benatti, and Rinaldo Garziera. Solving variational inequalities and cone complementarity problems in nonsmooth dynamics using the alternating direction method of multipliers. International Journal for Numerical Methods in Engineering, 122(16): 4093-4113, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[44] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. GaMeS: Mesh-Based Adapting and Modification of Gaussian Splatting. arXiv preprint arXiv:2402.01459, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "text", + "content": "[45] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4389-4398, 2024. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "text", + "content": "[46] Hongyi Xu, Yili Zhao, and Jernej Barbič. Implicit multibody penalty-based distributed contact. IEEE transactions on visualization and computer graphics, 20(9):1266-1279, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "type": "text", + "content": "[47] Shuojue Yang, Qian Li, Daiyun Shen, Bingchen Gong, Qi Dou, and Yueming Jin. Deform3dgs: Flexible deformation for fast surgical scene reconstruction with gaussian splattering. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 132-142. Springer, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "type": "text", + "content": "[48] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. arXiv preprint arXiv:2312.00732, 2024. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 477, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 531 + ], + "type": "text", + "content": "[49] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 533, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 575 + ], + "type": "text", + "content": "[50] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "905" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_content_list.json b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1dcaeff0ef1ffebdec58c274bd5b955ee5e22697 --- /dev/null +++ b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_content_list.json @@ -0,0 +1,2504 @@ +[ + { + "type": "text", + "text": "EntityErasure: Erasing Entity Cleanly via Amodal Entity Segmentation and Completion", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixing Zhu $^{1}$ Qing Zhang $^{1,4*}$ Yitong Wang $^{2}$ Yongwei Nie $^{3}$ Wei-Shi Zheng $^{1,4}$", + "bbox": [ + 169, + 202, + 826, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China $^{2}$ ByteDance Inc $^{3}$ South China University of Technology, China", + "bbox": [ + 191, + 220, + 805, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 107, + 256, + 883, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95d68646c3ec89202c98c78e765196351db5d2edb0a8499ca1e52f188f1ae3ec.jpg", + "image_caption": [ + "Input", + "Figure 1. Comparison with previous methods on object erasure. As shown, our method can cleanly erase objects without generating unwanted sundries that appear in the results of other methods." + ], + "image_footnote": [], + "bbox": [ + 93, + 292, + 205, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/131cb6b292d5f197c6a2d155c2d628b2b5c6d55b2c4161eb05f11fe8a80d13ab.jpg", + "image_caption": [ + "Inpainting mask" + ], + "image_footnote": [], + "bbox": [ + 207, + 292, + 321, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfef5d6930f374224a8feaa56a09d4f842594a5420ad71952219f09db9d36daa.jpg", + "image_caption": [ + "Our amodal seg." + ], + "image_footnote": [], + "bbox": [ + 323, + 292, + 437, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f4ff6b74efe85b6fd6a7531d95293baa720a4ad925069fb509ad7cb7071e939.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 439, + 292, + 552, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26c31fc6a41ab27fb110304641abe2cc6823374ba1c8f9a67dca2e1a1a0fa3cf.jpg", + "image_caption": [ + "PowerPaint [52]" + ], + "image_footnote": [], + "bbox": [ + 555, + 292, + 669, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3f21267127e00dc2955e926d00b8b0a26c92e23fb63d28e1b2072fed5a45498.jpg", + "image_caption": [ + "CLIPAway [8]" + ], + "image_footnote": [], + "bbox": [ + 671, + 292, + 785, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40a1884e85a79121dd26a553b53de8a43c8e68a911cc0ddc563b488979fdfa4b.jpg", + "image_caption": [ + "MAT [20]" + ], + "image_footnote": [], + "bbox": [ + 787, + 292, + 901, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 521, + 326, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents EntityErasure, a novel diffusion-based inpainting method that can effectively erase entities without inducing unwanted sundries. To this end, we propose to address this problem by dividing it into amodal entity segmentation and completion, such that the region to inpaint takes only entities in the non-inpainting area as reference, avoiding the possibility to generate unpredictable sundries. Moreover, we develop two entity segmentation based metrics for quantitatively assessing the performance of object erasure, which are shown be more effective than existing metrics. Experimental results demonstrate that our approach outperforms other state-of-the-art object erasure methods. Our code and data are available at https://zyxunh.github.io/EntityErasure-ProjectPage/.", + "bbox": [ + 88, + 551, + 485, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 792, + 220, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object erasure is a widely used technique in the field of image editing, which aims to remove the specified object and fill the area with coherent image content without generating new sundries. Early methods [4, 20, 27] are primarily", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "built upon Generative Adversarial Networks (GANs) [9], but they often do not perform well in open-world scenarios. Thanks to the powerful generation capabilities of diffusion models [32], diffusion-based inpainting [8, 52] has emerged to be a preferable choice for object erasure. However, these methods may generate unwanted sundries (see Fig. 1).", + "bbox": [ + 511, + 522, + 906, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to address the above problem, PowerPaint [52] proposes to learn task-related embedding for object generation/erasure, and introduces classifier-free-guidance to help avoid generating sundries. CLIPAway [8] leverages CLIP embeddings to make diffusion model focus on background regions while excluding foreground elements. Inst-Paint [44] adopts text prompts to guide diffusion model remove objects, while MagicEraser [17] improves the erasure effect by increasing the background weight and reducing the foreground weight in self attention. However, these methods are less effective in avoiding sundries in complex scenes, which motivates us to tackle object erasure from a different perspective.", + "bbox": [ + 511, + 613, + 908, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present EntityErasure, a diffusion-based inpainting method that can generate high-quality object erasure results free of sundries. Particularly, we propose to treat object erasure as a multi-entity amodal completion problem. To do so, we first perform amodal entity segmentation, and then utilize the segmentation output to guide", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (zhangq93@mail.sysu.edu.cn).", + "bbox": [ + 114, + 886, + 406, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28274", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc345f8f405e3dcb25c3b48f705f46c8b20bd49c434c1865078c0547a984b951.jpg", + "image_caption": [ + "Figure 2. Overview of EntityErasure. It contains a amodal entity segmentation model (AESM) and a amodal entity completion model (AECM). The output of AESM serves as input for AECM, AECM then injects content information into the denoising UNet with entity attention (EA). Particularly, EA is designed to constrain the cross-attention mask to focus only on the corresponding entities, ensuring that each region only looks for content information of its corresponding entity, so that the final result is aligned with the amodal entity segmentation." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the completion. Note, this is different from single-entity amodal completion that aims to generate a complete entity for each one, as object erasure requires to repair the missing entities and ensure the sequential order of the objects and the result to be coherent and contextually appropriate. By this means, we are able to predict the corresponding entity of each region, and constrain that each region only utilizes feature information from its corresponding entity during amodal entity completion, thereby preventing the generation of sundries. Considering that existing metrics (such as FID [11] and CLIP Dist. [44]) fail to reflect the effectiveness of object erasure in avoiding sundries, we further develop two metrics for performance evaluation.", + "bbox": [ + 89, + 412, + 483, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 609, + 410, + 625 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a diffusion-based inpainting method that allows to erase entities without generating sundries.", + "- We propose to tackle object erasure from a perspective of amodal entity segmentation and completion, and define two new metrics for performance evaluation.", + "- We show that our method outperforms state-of-the-art methods on object erasure." + ], + "bbox": [ + 109, + 628, + 483, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 762, + 233, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Inpainting", + "text_level": 1, + "bbox": [ + 89, + 787, + 261, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since GANs [9] have demonstrated powerful generative capabilities across various domains, including image generation and editing, numerous GAN-based inpainting methods [23, 27, 45-47] are developed in the past decade. Additionally, LaMa [37] introduces Fourier Convolutions [3] into the inpainting field, enabling better restoration of repetitive tex", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ture information. In recent years, diffusion models have become a preferable choice for generation. Sohl-Dickstein et al. [35] introduced the foundation theory of diffusion, while Dhariwal et al. [5] later demonstrated that diffusion can outperform GANs. To reduce the computational complexity, Stable Diffusion (SD) [32] proposes to scale up the model size and employs a Variational Autoencoder (VAE) [6] to map the original image into latent space, offering a powerful text-to-image model trained on LAION dataset [33] with billions of image-text pairs. Some inpainting methods [32, 38, 50] add new channels or modules to trained diffusion models for better injecting the original image information. For example, SD-inpainting [32] directly combines the inpainting mask and the original image as input. ControlNet [50] replicates a trainable U-Net architecture to facilitate the injection of additional conditional information.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering that the text prompt may not align with the semantic of the inpainting area, SmartBrush [38] uses text prompt and object shape to guide object inpainting by training with paired object-text data. However, this method is tailored for object generation, not object erasure that we target in this paper. PowerPaint [52] integrates text-guided generation and object erasure within a single model, using different text prompts to represent various tasks. For the erasure task, it introduces classifier-free guidance by using the generation prompt as the negative prompt to reduce sundries. CLIPAway [8] proposes a plug-and-play solution that extracts the background area features as prompts to guide the inpainting model to generate the background and remove the foreground. However, these methods may also generate sundries in some complex scenes.", + "bbox": [ + 511, + 334, + 906, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Amodal Segmentation and Completion", + "text_level": 1, + "bbox": [ + 511, + 574, + 848, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Amodal completion [11, 25, 49] aims to generate a complete object when only a partial view of that object is provided as input, while the goal of amodal segmentation [14, 19, 28, 51] is to produce a segmentation map that represents the entire object. As directly completing the entire object is a very challenging task, so various methods primarily focus on amodal segmentation. PCNet [49] proposes to complete the segmentation and then use the completed segmentation to guide subsequent amodal completion. To remedy the difficulty of collecting large-scale amodal data, SeGAN [7] attempts to train on synthetic datasets. However, these approaches still have respective limitations, and only work well for a limited set of closed-world objects.", + "bbox": [ + 511, + 597, + 905, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the emergence of diffusion methods exemplified by Stable Diffusion [32], many amodal completion approaches [26, 40, 48] are also trying to utilize SD. Benefiting from the superior generation capability of SD, these methods are no longer limited to specific scene types or fixed categories. pix2gestalt [26] incorporates visible amodal masks as additional conditional inputs into SD to in", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28275", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b4e9f173f94334ded1560b09b5c5875df730d43c63df65750c68a0337f75b4d.jpg", + "image_caption": [ + "Figure 3. Illustration of entity attention (EA). It leverages amodal entity segmentation to restrain attention map calculated between the entity reference feature and input feature. By limiting each entity's influence solely to pixels within regions predicted by the amodal entity segmentation, it completes corresponding entities as segmentation guidance. Here, $Q$ , $K$ , and $V$ represent input queries, entity reference keys, and entity reference values, respectively." + ], + "image_footnote": [], + "bbox": [ + 174, + 90, + 823, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "form the model which object needs to be completed, then it predicts the shape and appearance of the entire objects. Xu et al. [40] proposes a training-free method that directly uses the pre-trained model of SD-inpaint [32]. It directly blends the visible portions of the object that need to be completed with the background image, which is then fed to the SD-inpainting model to reconstruct the remaining parts. This also highlights the strong correlation between amodal completion and object inpainting, as both tasks involve reconstructing missing information based on available contextual cues. SDAmodal [48] enhances the generalization capabilities of amodal segmentation by incorporating a pre-trained SD model into their pipeline.", + "bbox": [ + 88, + 376, + 485, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 585, + 183, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 presents the pipeline of our method. As shown, we firstly feeds the noisy latent $z_{T}$ , masked image latent $c_{i}$ , and inpainting mask to Amodal Entity Segmentation Model (AESM) to predict amodal entity segmentation. Next, the amodal entity segmentation, masked image latent $c_{i}$ , inpainting mask, and the noisy latent $z_{t}$ will be fed to Amodal Entity Completion Model (AECM) to predict the final amodal entity completion result. The noisy latent $z_{T}$ and $z_{t}$ can be set differently. Below we describe our method in detail.", + "bbox": [ + 89, + 612, + 483, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Amodal Entity Segmentation", + "text_level": 1, + "bbox": [ + 89, + 772, + 349, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are different levels of granularity for object segmentation. A segmentation map of a person can be the entire person, or specific parts such as limbs, the head, or even finer details like the eyes and nose. Segment Anything Model (SAM) [15]'s levels of granularity are various, and it's not easy to find an appropriate level of granularity. Considering this, we choose to use entity segmenta", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion [29, 30] as our entity unit which defines a complete object or stuff as a single entity. Well-trained diffusion models such as stable diffusion can serve as pretrained models for other vision tasks and have shown strong performance compared to other self-supervised algorithms such as MAE [10]. Inspired by ODISE [39], we use pretrained stable diffusion model [32] as backbone of segmentation. Same with ODISE, we use Mask2Former [2] as our decoder. During testing, it makes only a single prediction. However, with different random noise inputs, it can produce varying amodal entity segmentation results.", + "bbox": [ + 511, + 376, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use Feature pyramid networks (FPN) [22] to fuse the features of different layers. We set $N$ segmentation queries in the Mask2Former decoder head, then use masked cross attention between queries and FPN layers to exact things and stuff features into queries. Finally, each query generates a category score and a segmentation mask for each thing and stuff. For more details, please refer to [2].", + "bbox": [ + 511, + 544, + 908, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Amodal Entity Completion", + "text_level": 1, + "bbox": [ + 511, + 665, + 758, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initial diffusion-based inpainting method directly adds the inpainting conditional input on text-to-image model and then fine-tunes this model. BrushNet [13] proposes to use a decomposed dual-branch to inject non-inpainting area information into main branch. Inspired by BrushNet, we also use decomposed dual-branch to inject non-inpainting area information to inpainting area of the corresponding entity. However, unlike the depth, edge, and other conditions of ControlNet [50], the generated area and the condition area of inpainting task are not element-wise corresponding, so we choose to inject entity information in the attention layer. Ip-Adapter [43] uses CLIP [31] and Magicanimate [41] uses copied Unet as appearance feature encoder. We choose Unet as our Entity Reference Encoder (ERE) because Unet", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28276", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can provide more low-level detailed information than CLIP.", + "bbox": [ + 89, + 90, + 482, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We copy the main branch diffusion model $\\mathcal{F}_a(\\cdot;\\Theta_a)$ as Entity Reference Encoder $\\mathcal{F}_b(\\cdot;\\Theta_b)$ . $\\mathcal{F}_b(\\cdot;\\Theta_b)$ is used to extract the features of non-inpainting area, $\\Theta_b$ is its trainable parameters. We incorporate the features from the Entity Reference Encoder (ERE) into the main branch by injecting them into attention layers, where $h_b$ represents the hidden state feature of the ERE:", + "bbox": [ + 88, + 106, + 482, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {b} = \\mathcal {F} _ {b} \\left(c _ {i}; \\Theta_ {b}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 226, + 482, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{i}$ is the condition latent feature, $h_{b}$ contains the information of condition images and processes it to the high-level semantic feature. So we inject $h_{b}$ to cross attention of main branch Unet. The hidden state feature of main branch is represented as $h_{a}$ . In order to ensure that each entity area only injects the semantic information of the corresponding entity, inspired by Freestyle [42], we use the amodal entity mask obtained in the previous stage to restrict the attention mask. As shown in Fig. 3, the image query $(Q)$ and entity reference key $(K)$ , value $(V)$ are obtained with three projection layers. Then standard attention score map $\\mathcal{M}$ is calculated as:", + "bbox": [ + 89, + 250, + 483, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = \\frac {Q K ^ {T}}{\\sqrt {d}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 426, + 482, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ is the scaling factor. Next we utilize the amodal entity mask to constrain the distribution of the attention maps:", + "bbox": [ + 89, + 464, + 483, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {M}} _ {i, j} = \\left\\{ \\begin{array}{l l} \\mathcal {M} _ {i, j} & \\text {i f} I _ {i} = = I _ {j}, \\\\ - \\inf & e l s e. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 505, + 482, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{i}$ is the $i$ -th element's ID of amodal entity segmentation predicted in the first stage. Final attention output can be calculated as:", + "bbox": [ + 89, + 550, + 483, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} (\\widehat {\\mathcal {M}}) V, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 604, + 482, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, only information of the same entity will be injected into the corresponding inpainting area.", + "bbox": [ + 88, + 633, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the non-inpainting areas are element-wise corresponding, its generation is straightforward, requiring only a copy of the corresponding pixel information. Therefore, we only apply entity attention in the inpainting area. This process is expressed as follows:", + "bbox": [ + 89, + 665, + 483, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {o} = I _ {i n} \\cdot W ^ {o 1} \\mathrm {A t t n} (Q, K, V) + (1 - I _ {i n}) \\cdot W ^ {o 2} h _ {b}, (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 750, + 482, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h_{o}$ is the output hidden state, $I_{in}$ is the inpainting mask, $W^{o1}$ and $W^{o2}$ are two fully connected layers's weight.", + "bbox": [ + 89, + 777, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Classifier-free Guidance", + "text_level": 1, + "bbox": [ + 89, + 833, + 310, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classifier-free guidance (CFG) [12] is widely used in diffusion methods to enhance the strength of the condition guidance. A positive prompt and a negative prompt are", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "required. In PowerPoint [52], a special token is used to represent the erasure task, while another token is used to represent the generation task. Then the erasure token is set to positive prompt, and the generation token is set to negative prompt. This helps enhance the erasure instruction, thereby reducing the possibility of generating sundries. We also use classifier-free guidance to enhance the erasure task. However, unlike PowerPoint, we do not train an additional generation task. Instead, we choose to enhance the guidance of the input condition image. Similar to the IP-Adapter [43], during training, we randomly drop the input condition image with a probability of $p_d$ by replacing it with an all-zero image, $p_d$ is set to 0.05 as default. During testing, the positive input is the condition image containing the non-inpainting regions, while the negative input is an all-zero image. We then use the following linear combination of the conditional and unconditional result:", + "bbox": [ + 511, + 90, + 906, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c _ {i}, c _ {t}\\right) = (1 + w) \\epsilon_ {\\theta} \\left(z _ {t}, c _ {i}, c _ {t}\\right) - w \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing_ {i}, \\varnothing_ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 363, + 906, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{i}$ and $c_{t}$ are latent features of the condition image and text conditions, respectively, while $\\varnothing_{i}$ is a blank, all-zero latent feature, and $\\varnothing_{t}$ is an empty string. We find that significantly improves the model's erasure performance and the quality of the generated images.", + "bbox": [ + 511, + 393, + 906, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 482, + 730, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use about 1750,000 images from Open Images [16] V7 to train our model. Entity segmentation pseudo labels are generated by a pretrained entity segmentation model released by [30], and text prompt pseudo labels are generated with blip [18]. Given a source image, we randomly generate an inpainting mask to occlude some parts of the image. The image with mask occlusion is then treated as the input image, while the entity segmentation of the source image is now the ground truth for amodal entity segmentation, and the source image is the ground truth for amodal completion.", + "bbox": [ + 511, + 506, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use LaMa's strategy [37] and entity mask to generate the inpainting mask. In addition, since we are addressing an amodal completion task, we constrain that at least $40\\%$ of each entity is not masked. AESM and AECM are trained separately. Both of them are initialized by pretrained RealisticVisionV60B1 [34]. For AESM, the number of mask queries $N$ is set as 100. All parameters of AESM are trainable. We set the batch size to 64, the number of iterations to 80,000, and the learning rate to 1e-5, respectively. For AECM, the main branch Unet is frozen. We use the AdamW [24] optimizer. We set the batch size to 128, the number of iterations to 20,000 iterations, and the learning rate to 5e-5. Input images are resized to maintain their aspect ratio, with a fixed area of around $512 \\times 512$ . During inference, we use DDIM [36] sampling with a number of 20 sampling steps.", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28277", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/80e271e934f55ea4ddbbc95231d64d6617031bc66a738e0b495c8d1628697020.jpg", + "image_caption": [ + "Figure 4. Pipeline of our object erasure metric. The objective of this case is to erase the dog within the inpainting mask. To determine if any new sundries have been generated in the result, we first obtain an entity segmentation of the erased output using an open-world entity segmentation model. We then calculate the Intersection over Self (IoS) for each entity relative to the inpainting mask. Entities with IoS greater than threshold $\\lambda$ are considered newly generated sundries. In this case, the tiger is newly generated." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 898, + 224 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7e496f1f40c4ffd9635487fab289b2c041b2193abad210b8ff9d84597a8c6ddb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEntitySeg [30]coco [21]
FID↓MSN↓MARS↓ReMOVE↑FID↓MSN↓MARS↓ReMOVE↑
MAT [20]50.00.4330.04486.6725.181.8760.13379.79
SDInpaint [32]57.10.7460.15184.6323.272.9670.28877.14
CLIPAway [8]59.00.6790.13084.2824.172.6860.24078.17
PowerPaint [52]56.90.5540.11584.8025.682.6540.23978.0
Ours58.00.1220.03786.2422.530.5060.08679.56
CLIPAway* [8]60.40.8360.13484.2925.332.7520.22777.11
PowerPaint* [52]58.20.6970.10884.8126.952.7140.22576.93
Ours*59.10.1580.03286.1023.350.5520.08078.05
", + "bbox": [ + 148, + 301, + 849, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with state-of-the-art methods. * indicates results produced with loose masks.", + "bbox": [ + 156, + 486, + 836, + 501 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66a0b663ee869ca819ab1edbe7e5eb7502de074bcd18ffe5895b413f1909009f.jpg", + "image_caption": [ + "Figure 5. Comparison between our proposed metrics (MSN↓, MARS↓) and CLIP accuracy [44] (CLIP@1,3,5↑). CLIP accuracy only focuses on the difference between the generated result and the original input, so toys with larger differences in the generated image actually lead to an improvement in the CLIP@1 and CLIP@3 metrics." + ], + "image_footnote": [], + "bbox": [ + 96, + 515, + 181, + 609 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/90ec4b2fd28fa710a30f8957a81f8610297d3fc41ecd9c0348655fa81e989e9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 515, + 333, + 609 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5a9bac4596dcee38e8013e38e34050d2c5bf0292e62dbcc938b37b10f4e99381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 515, + 421, + 609 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30b5ea547f78dfd30e476e2dd14638d387af5e41b6f9aec665d44f3b1f94113e.jpg", + "image_caption": [ + "Figure 6. Comparison between our proposed metrics (MSN↓, MARS↓) and ReMOVE score [1] (ReMOVE↑). ReMOVE focuses only on whether the generated result is consistent with the surrounding context. The newly generated blue chair is consistent with the surrounding environment, so it received a higher Remove score." + ], + "image_footnote": [], + "bbox": [ + 421, + 515, + 480, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 729, + 223, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 756, + 274, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "How to evaluate the model's ability of erasing objects has always been a difficult problem. Other inpainting tasks such as text-guide inpainting are generally evaluated by using pre-trained CLIP model to calculate the correlation between the guided text prompt features and the generated image features. The higher the correlation, the more the generated results are in line with expectations. However, for erasing tasks, it is necessary to ensure that no new entities are gener-", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d578b70186f94bd3086fe8b379766c6164ae4b49ec27f86fa99106b442619f17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 515, + 602, + 604 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c505d8ff9871d5484ae85515cf6612fef4b380828e7e47e761155c20427beb71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 515, + 689, + 604 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/75b74904cb51644ad26135748d562ffec6284924418c825af9fcb76e2358e7fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 515, + 751, + 606 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/478a1f70c6dafde6d4cc288961b65ff1b6451e8ba99c4b01979a46b823abdd3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 515, + 843, + 606 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/90dd6b9e56583e5004ea6c59840422df6bf3ea200249486362aa29a43ca9be55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 844, + 515, + 903, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ated and the content of the generated result is part of original entity. Considering this, we directly determine whether any new entities are generated within the inpainting mask. As shown Fig. 4, we use the pre-trained state-of-the-art openworld segmentation model EntitySeg [30] to predict the entity segmentation of the generated result, and then calculate the Intersection over Self (IoS) between each entity with inpainting mask by:", + "bbox": [ + 511, + 724, + 906, + 847 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I o S} (A, B) = \\frac {| A \\cap B |}{| A |}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 854, + 906, + 887 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28278", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfddbc5192a6c050cdc7b10e19145b6b8e188bead0f98757d475f1de96939de9.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 102, + 205, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/503d41e9a14b0142fc1f28e346284f96156f12f2b020cc6b3b2231e57b89ab35.jpg", + "image_caption": [ + "Inpainting mask" + ], + "image_footnote": [], + "bbox": [ + 207, + 102, + 321, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ecf565fc4b9d4e2a3c0042659ac12a272f99c05fdd53bb578ebd9e4bd1ed3c6.jpg", + "image_caption": [ + "Our amodal seg." + ], + "image_footnote": [], + "bbox": [ + 325, + 102, + 437, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29f2c9794fcf195b5474a75db5ad150c85a4519132a27c5e548cf7c44fe415d7.jpg", + "image_caption": [ + "Ours", + "PowerPaint [52]" + ], + "image_footnote": [], + "bbox": [ + 442, + 102, + 553, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/420c230bc84ec3701523eb95f4c5f85f95fd74ef302dee4c7e07308985dc5157.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 161, + 204, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f2b14d98689993b3ff77d4d38e43fc9a0d83fb886e643dd652b63fb6c9bdbb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 161, + 320, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b2c05a09a2d42dc456cea9181bc0a29207e753a8c8b9677e073200e8bcbda4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 161, + 437, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3fe885ed4a4c35dd60e1b461ca4dc32bb4f0bc4c77116a30433cd009d47d0271.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 161, + 553, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/821b9930147fe0860fceaa12030efad941ed7071a6f7e2dfb89de06ffafd2e8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 161, + 669, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c168f87aa8cae347ac3075e7e17d0884dd8b82eb3e96ce3f0aed52a35505628e.jpg", + "image_caption": [ + "CLIPAway [8]" + ], + "image_footnote": [], + "bbox": [ + 674, + 161, + 785, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/244df2dae948974448fb1d1c43c6d3139dfe2db042746499b7e52772a8ecc2fd.jpg", + "image_caption": [ + "MAT [20]" + ], + "image_footnote": [], + "bbox": [ + 790, + 161, + 901, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b33196469bb7bad83aaa453dc570c0dde2650c1369163d3b6a2fea6a9d0a8b2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 224, + 204, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00e54a27e0f28e5962baad0240bb6d931a7501427112fce45a8c11fa5f504fec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 224, + 320, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ece6eead487f878dcfa11f9c05023c2e73ef0328d58389a369287bc0b3cffe6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 224, + 437, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/607ca004c99a37075c00d810382430cab655d043d218f5768efa768166d7eeb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 224, + 553, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7be25138b0c60bc8e4a40a4ae02ea471488e3282079cc3d4691f384c45ebf08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 224, + 669, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5fc8ad3c7a0d33380ea55ffd71219a62ae5a2e12fedc216081887171c7854818.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 224, + 785, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b4f44ef2aa7bc0092c38a802bd12e84611bdcf7206b301f02d313c7d9b8eb41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 224, + 900, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a07f3fc7b0b55c793961bf30678fc3ccbbb678c7e77e3b0312e714850b9771b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 291, + 204, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e9f3a4ba94da554f22161c462b5cbc18b1bd7028711b129b9d99acafa1ad57b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 291, + 320, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d872abbaab29b3766b0af1385e885fffa61f03f77128e8078f25f3f1f6d7468.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 291, + 437, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e07a5a659da4041911111cdb7a69d7b2617912fe0e8e2d0b87942bda0cb23b03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 291, + 553, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1139b27b4cd9fc48c7a41387b852ef066939d4654898dbcbd6b6894da8cbb598.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 291, + 669, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5970e85c01d66b196072f3e680f3744a73c3981bd3bbb769b3b2002adad172f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 291, + 785, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/784b21c6b154c0aefffa48ea7afdf14f5845404a98f9bca9068d706c22760d8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 291, + 900, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c32db83d5dbb778ddf3c48f744dac297bc50c0f5c41e609c518f81cfc284802.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 352, + 204, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f23fc56921060d77673eaf644301da08ec766844001208aff3b710a68da3ab0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 352, + 320, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c570319cc88043640bf6696e37a17a9b166dc5b4386530f7a80430b9c05e5e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 352, + 437, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6813dd7a55f627650a5217ec289190ddbbb03f0982b28e3c90512f3fbd37c5d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 352, + 553, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/11f228a7ffd52f14fe2e395e6a7f3f6ae965ee4215f68577e75f4cde79754bc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 352, + 669, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54c3d9bb194c7a80cc7b395e45430c1854515df480bc7c9846e942b9c12297bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 352, + 785, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7cb8337df6863f3882f3820179085a6abedf9b61600797006b9391aea50d087.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 352, + 900, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/80bc1a3d5e1c8b7d61b2abf6f85f73627b33b938deecb0f3754a330af92abcb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 420, + 204, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8531769c74a7afa17953508a8604a2b45b160a6cc3361be849ee6fae81169aa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 420, + 320, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2b3c4aef2d166c200ae4d89db70f18c8924ab52d0ee34c603556e18404a581b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 420, + 437, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b27b8735aa964fe2463500be9b15c7d25a0d2acb72f10438062ec2e918565ac9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 420, + 553, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17b65febfd662f7db2a494f3eb870ae247f51a8406c9eddf63b0d72a53c35500.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 420, + 669, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/903588f52c5235edf9c0f8105e4a919c6492f9d222f2b3a4d3590f70693e6cad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 420, + 785, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab1cb0c4abb2c9c265aedc9bba2b2fdcb907c4051bb6a0d9c2d1b44fd23f4cf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 420, + 900, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/940862c7ca1f5d2b7653d9b11b0ed1519ea638e3031a58ad5b95a811ef56213f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 484, + 204, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/697b3b982cb243fbbc9d8479f28ae6bbac7b03140a38247a23c39900c717710b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 484, + 320, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/651e9193db581e5fea2be522fbc57d104dc5cecc17c827954d6430071abc259a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 484, + 437, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0de4e5c79f125e6003575cdc6d41456401c7975659cb85e2fd81e64f937c4a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 484, + 553, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75e13bdadcce2aa28797a8bfa4ad573282a3d45eb459d16f0ddfd8752607e8bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 484, + 669, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1fa7abd658bedc6cb099a77400efb4013dafcfc57218199d07ad356560e0a3e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 484, + 785, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea0cc1002a0555877001172bb5c4aacddb94836c15be7049fc4e86952dfde631.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 484, + 900, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1032b9177d616fda0b3d78baa3d609873d5996cdf06f9dee7f788a291f19820.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with state-of-the-art methods. As shown, our method effectively erases the target objects while avoiding generating unwanted sundries like the compared methods." + ], + "image_footnote": [], + "bbox": [ + 93, + 551, + 204, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dcf0071bf0608a84ccaaad113640db89a9cc30100366270a512782b104450717.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 551, + 320, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8032237434a1cbe42d18c4d38c856e5e8935d2febc1bd3ccec1b51f4d320239f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 551, + 437, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/558d1ac2aacdebcb047abdbb82d6b23df2ed4c0bba9f1ae760823b203f8eca20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 551, + 553, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c9b8c512f9c2987c42080d6e8bcacbcf6e9807dc43a08702d5aa0568a88c92e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 551, + 669, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79de545d0cabd71e8e420a81ef6e0aa61bb8212ceca6b920b2259b1d0c2e379c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 551, + 785, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef807835733c81372c38eb27bc9bf78332fa6933b3bf83a668254b950e717947.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 551, + 900, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $A$ is the entity segmentation mask and $|A|$ refers to the area of $A$ . $B$ is the inpainting mask. If the IoS is greater than the threshold $\\lambda$ , these entities are considered as newly generated sundries (SDR). $\\lambda$ is set to 0.95 by default. We traverse all entities and calculate the number of all newly generated sundries (referred to as $|SDR|$ ). By setting $N$ as the number of test samples, the Mean Sundries Num (MSN) is defined by:", + "bbox": [ + 89, + 672, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M S N} = \\frac {| S D R |}{N}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 792, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Mean Area Ratio of Sundries (MARS) to the inpainting area can be computed by:", + "bbox": [ + 89, + 825, + 482, + 856 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M A R S} = \\sum_ {A \\in S D R} \\frac {| A \\cap B |}{| B |} \\times \\frac {1}{N}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 864, + 482, + 904 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MSN reflects the quantity of sundries generated by the model, while MARS indicates the ratio of the sundries area relative to the inpainting mask area. A higher MARS value signifies that the generated sundries are larger and more prominent. Both of these metrics should be minimized for optimal performance. In Fig. 5, we present a comparison between our metric and CLIP accuracy [44]. It is evident that CLIP accuracy only focuses on the magnitude of the change between the generated result and the original image. We also compare our method with ReMOVE [1] in Fig. 6. ReMOVE focuses solely on whether the generated image is consistent with the surrounding context. It fails to provide a reasonable score when the newly generated object (blue chair) is consistent with the surrounding context. In contrast, our metric is more reasonable and reliable.", + "bbox": [ + 511, + 672, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28279", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51b05103348b269308f9b3dea1d27894befa6774c448c5948283df0bdec8bf8d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 93, + 98, + 168, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44eb1bb973a22f2011546a8eeaf001df50774afb85df8d63fc5f8b0a67999cb5.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 169, + 98, + 245, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83772e6c4613e244c45a53dcf65de447bab0ae71bcf9029fefa2a9c1a4dcfdf8.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 246, + 98, + 323, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50e74cbddf0052dab63aae9c97f1e4a5f09da4fb657778c0b09c2cd9cee5cc9a.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 325, + 98, + 400, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21ddf1320cc8ba197025e24cfa98b0bbadf6012e8eb2c7866dc4f8f4479146fe.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 401, + 98, + 477, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02cf79cfe422caae5aaab0d27257ebc9b985c01c8200273658494c8b102b7d37.jpg", + "image_caption": [ + "Figure 8. Effect of entity attention. From left to right: (a) input, (b) inpainting mask, (c) amodal entity segmentation, (d) entity attention, (e) cross attention. As can be seen, with our entity attention, our method faithfully repairs the missing parts of entities according to the amodal entity segmentation, effectively reducing the generation of sundries.", + "(a)", + "Figure 9. Impact of applying entity attention on non-inpainting area: (a) input + inpainting mask, (b) not applying entity attention on non-inpainting area, (c) applying entity attention on non-inpainting area. As shown, applying entity attention on non-inpainting areas may lead to unwanted changes to these areas compared to the original image (see the red box)." + ], + "image_footnote": [], + "bbox": [ + 93, + 441, + 215, + 539 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41d1d2616f9ec7f9d265ccb12686877434c828aad90b6760c600779d15f2d6cd.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 220, + 441, + 344, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68f1136591874d8648bc205429838f09ee8fb4aac2f7da07f1a06c6c802e88ac.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 349, + 441, + 473, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since we evaluate our method on non-synthetic real-world scenarios, where paired data is unavailable, we use non-paired metrics for evaluation. Besides MSN and MARS, we also report ReMOVE score [1] and Frechet Inception Distance (FID) [11]. We compare our method with other state-of-the-art inpainting approaches on entity segmentation test dataset [30] and coco val2017 dataset [21]. Excluding stuff objects, the segmentation mask for thing object is selected from each image as the inpainting mask, then we dilate the mask. Meanwhile, we find that CLIP-Away [8] and PowerPaint [52] would generate additional sundries from floating handheld objects due to precise inpainting mask. Therefore, we also conduct comparisons using looser inpainting masks. We obtain looser inpainting masks by dilating masks with larger sizes and filling the", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "holes. A total of around 700 test samples are generated from the entity segmentation dataset, and 4000 test samples are generated from the COCO dataset.", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 145, + 880, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our method with various state-of-the-art inpainting methods, including MAT [20], SD-Inpainting [32], CLIPAway [8], and PowerPaint [52], where MAT is a GAN-based method while the other three are stable diffusion based methods.", + "bbox": [ + 511, + 167, + 905, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. As shown in Table 1, our method demonstrates significant advantages over CLIP-Away [8] and PowerPaint [52] in both the MSN and MARS metrics on entity segmentation and coco dataset, using either precise or loose inpainting masks. And among all the diffusion-based methods, our approach also achieves the highest score on the ReMOVE metric. This demonstrates that the proposed model significantly reduces the probability of sundries generation and the area of generated sundries compared with other methods. We also observe that the GAN-based method MAT has relatively low MSN, MARS score and high ReMOVE score compared with SD-based baselines. This may be due to the powerful generative capability of the SD model.", + "bbox": [ + 511, + 250, + 906, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. In Fig. 7, we show some prediction results of EntityErasure alongside other state-of-the-art methods. It is evident that MAT generates fewer sundries but their generated textures lack realism. Meanwhile, CLIPAway and PowerPaint produce higher quality results, but still produce sundries in some cases. In contrast, our proposed EntityErasure demonstrates a lower probability of sundries generation and produces higher-quality images. Please see the supplementary material for more results.", + "bbox": [ + 511, + 465, + 905, + 603 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/42a2b16e53836d60fba77180792a20ee2da8d033212cacad75592e15cdff6e61.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMATCLIPAwaySDInpaintPowerPaintOurs
Time(s)↓0.084.042.215.002.98
", + "bbox": [ + 517, + 616, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison on inference time.", + "bbox": [ + 584, + 667, + 833, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on inference time. In Table 2, we conduct comparison of inference speed on a single RTX 4090. The number of our sampling steps is set as 20. As shown, the inference speed of our method is comparable to those of other diffusion-based methods. Note, we experimentally find that larger number of sampling steps (e.g., 50 steps) does not lead to clear performance improvement, while the inference time cost will be increased.", + "bbox": [ + 511, + 703, + 905, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 833, + 666, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of the modules proposed in EntityErasure, we conduct ablation experiments on entity segmentation dataset [30].", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28280", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f13073755b2d7b267b57a6ef366f82712e95dac502c5870fcf04207746e8bc18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓MSN↓MARS↓ReMOVE↑
cross attn57.10.2120.07785.82
entity attn58.00.1220.03786.24
", + "bbox": [ + 106, + 88, + 470, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of amodal entity segmentation guidance. We use amodal entity segmentation guidance in entity attention to inject content into the corresponding entity. To evaluate its impact, we replace entity attention with standard cross attention for comparison. As shown in Table 3, FID shows a slight decrease, while MSN and MARS increase notably, indicating a rise in sundries generation. This highlights the effectiveness of our amodal entity segmentation guidance in reducing sundries generation. It is further supported by the qualitative ablation results presented in Fig. 8, where model without amodal entity segmentation guidance tends to generate sundries easily.", + "bbox": [ + 89, + 200, + 483, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of applying entity attention on non-inpainting area. We observe that applying entity attention for injecting information into the non-inpainting area does not incur sundries, but it may cause alterations to the non-inpainting region. As shown in Fig. 9, if entity attention is applied to the non-inpainting area, changes can be observed. This implies that, the non-inpainting area only requires feature injection via element-wise addition, without the need for entity attention.", + "bbox": [ + 89, + 388, + 483, + 523 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/137ae5ca71384cfd7f3e6e7b9c796f0a784e714666b60051cda423f810400ffa.jpg", + "table_caption": [ + "Table 3. Effect of entity attention." + ], + "table_footnote": [], + "table_body": "
TIFID↓MSN↓MARS↓ReMOVE↑
1159.80.4490.078384.86
1358.00.1220.036986.24
3159.90.4250.075684.88
3358.00.1220.036686.24
", + "bbox": [ + 114, + 535, + 460, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Effect of classifier-free guidance. T: Text-CFG weight, I: Image-CFG weight.", + "bbox": [ + 89, + 638, + 482, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of classifier-free guidance. We apply classifier-free guidance to amplify the influence of condition image and amodal entity segmentation guidance. To assess its effectiveness, we perform experiments with varying classifier-free guidance weights. As shown in Table 4, when the weight of Image-CFG is set to 1, effectively turning off Image-CFG, both MSN and MARS increase significantly, which signals an increase in sundries generation. Compared to the impact of Text-CFG, the influence of Image-CFG on the results is more significant. It is further supported by the qualitative ablation results presented in Fig. 4. In the first butterfly image, without CFG, the model would generate excessive textures. This underscores the role of Image-CFG in reducing the occurrence of sundries.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89b2e00ee18584c6f805ad1e551cf7f88545473b8f0b28895637be1d52f228d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 101, + 612, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/607a1a117cd0a8da134e1f3e601f548928cbaba0c7660dacc27475ffed1089b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 101, + 709, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa05338e5d91422fad3a92143773165207beb87fc23cc63f77f62307bf8eb9e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 101, + 805, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d2bdc1818bb1946e85eaadb2b215460d9fdce54db13d9c409155d12d7fdfa0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 101, + 903, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e62895cb0ae35471ffd1a96be84b5018b3f9b2bf8d7972c8e7ea54ffd8d2a8fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 145, + 609, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e8a86f809f1258c07ec97084883695bef4ab0045833947e2ff487daa61ff6a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 145, + 709, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a55a6f9f04cd77eae6dc26fb55e97321c6fae6d63d8fff890647aa5d93fc938.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 145, + 805, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32b15670a7dcdb48723cd7e244ca814eb50c3a61e85d40e7d2ef07e28cf40e17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 145, + 903, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34a6b55b8831a96591dc5a3fc4ed00f8ed84dbe3c6126d2ee883d12335c5c967.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 516, + 258, + 611, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85c419f09dc6c7c4598325eb7ea9e247caf3030c950deae04af63a0cc70ca5b9.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 614, + 258, + 709, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3697e753b11bc3d3f69a6e4866f199c2da02ddb1cb30ccd35389e6a402c156c1.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 710, + 258, + 805, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3bba271f19a65a45172d4499f88ce862ebe243d4c38858c6c64de9b982b6b352.jpg", + "image_caption": [ + "(d)", + "Figure 10. Effect of classifier-free guidance. From left to right: (a) input, (b) inpainting mask, (c) Imagecfg=3, (d) Imagecfg=1. As shown, utilizing classifier-free guidance to strengthen the guidance from the condition image benefits obtaining cleaner images without sundries." + ], + "image_footnote": [], + "bbox": [ + 808, + 258, + 903, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 449, + 633, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel diffusion-based inpainting method named EntityErasure. We approach the erasure task from a new perspective, framing it as a partial amodal entity completion task. Our method decomposes the erasure process into two steps: partial amodal entity segmentation and partial amodal entity completion, making the erasure more controllable. The proposed entity attention injects entity information with amodal entity segmentation guidance. We ensure that all generated content is completed from the entities in the non-inpainting area, effectively preventing the creation of new sundries. EntityErasure demonstrates state-of-the-art performance in terms of object erasure.", + "bbox": [ + 511, + 481, + 906, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Although our method can produce high-quality object erasure results without unwanted sundries, it also has several limitations. First of all, the success of our method may be affected by the quality of amodal segmentation, though our method has some tolerance to poor amodal segmentation (see the supplementary material for validation). Besides, it currently relies on manual mask extension to remove shadows associated with the erased objects. In the future, we will try to address this limitation by incorporating instance shadow detection.", + "bbox": [ + 511, + 676, + 908, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002).", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28281", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In CVPRW, 2024. 5, 6, 7", + "[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 3", + "[3] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. In NeurIPS, 2020. 2", + "[4] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1", + "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2", + "[6] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2", + "[7] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2", + "[8] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models. In NeurIPS, 2024. 1, 2, 5, 6, 7", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2", + "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 7", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 4", + "[13] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024. 3", + "[14] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2", + "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3", + "[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 4" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Fan Li, Zixiao Zhang, Yi Huang, Jianzhuang Liu, Renjing Pei, Bin Shao, and Songcen Xu. Magiceraser: Erasing any objects via semantics-aware control. In ECCV, 2024. 1", + "[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 4", + "[19] Ke Li and Jitendra Malik. Amodal instance segmentation. In ECCV, 2016. 2", + "[20] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In CVPR, 2022. 1, 5, 6, 7", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7", + "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3", + "[23] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4", + "[25] Khoi Nguyen and Sinisa Todorovic. A weakly supervised amodal segmenter with boundary uncertainty estimation. In ICCV, 2021. 2", + "[26] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In CVPR, 2024. 2", + "[27] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1, 2", + "[28] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2", + "[29] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Philip Torr, Zhe Lin, and Jiaya Jia. Open world entity segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8743-8756, 2022. 3", + "[30] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High-quality entity segmentation. In ICCV, 2023. 3, 4, 5, 7", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 7", + "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28282", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy1, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2", + "[34] SG161222. Realistic vision v6.0 b1, 2024. 4", + "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2", + "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4", + "[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 2, 4", + "[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2", + "[39] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3", + "[40] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. A modal completion via progressive mixed context diffusion. In CVPR, 2024. 2, 3", + "[41] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanate: Temporally consistent human image animation using diffusion model. In CVPR, 2024. 3", + "[42] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In CVPR, 2023. 4", + "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 3, 4", + "[44] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 2, 5, 6", + "[45] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2", + "[46] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In CVPR, 2019.", + "[47] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2", + "[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, 2024. 2, 3", + "[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, 2020. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3", + "[51] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2", + "[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In ECCV, 2024. 1, 2, 4, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 219 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28283", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_model.json b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c76a39eb202e82c102978d83782612e26d352fd4 --- /dev/null +++ b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_model.json @@ -0,0 +1,3190 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.885, + 0.178 + ], + "angle": 0, + "content": "EntityErasure: Erasing Entity Cleanly via Amodal Entity Segmentation and Completion" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.203, + 0.828, + 0.223 + ], + "angle": 0, + "content": "Yixing Zhu\\(^{1}\\) Qing Zhang\\(^{1,4*}\\) Yitong Wang\\(^{2}\\) Yongwei Nie\\(^{3}\\) Wei-Shi Zheng\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.222, + 0.807, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-sen University, China \\(^{2}\\)ByteDance Inc \\(^{3}\\)South China University of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.257, + 0.885, + 0.275 + ], + "angle": 0, + "content": "4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.28, + 0.168, + 0.293 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.293, + 0.206, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.28, + 0.321, + 0.294 + ], + "angle": 0, + "content": "Inpainting mask" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.293, + 0.322, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.28, + 0.436, + 0.294 + ], + "angle": 0, + "content": "Our amodal seg." + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.293, + 0.438, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.281, + 0.514, + 0.294 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.293, + 0.553, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.28, + 0.667, + 0.294 + ], + "angle": 0, + "content": "PowerPaint [52]" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.293, + 0.67, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.28, + 0.778, + 0.294 + ], + "angle": 0, + "content": "CLIPAway [8]" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.293, + 0.786, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.28, + 0.879, + 0.294 + ], + "angle": 0, + "content": "MAT [20]" + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.293, + 0.902, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.481, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Figure 1. Comparison with previous methods on object erasure. As shown, our method can cleanly erase objects without generating unwanted sundries that appear in the results of other methods." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.522, + 0.327, + 0.537 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.486, + 0.766 + ], + "angle": 0, + "content": "This paper presents EntityErasure, a novel diffusion-based inpainting method that can effectively erase entities without inducing unwanted sundries. To this end, we propose to address this problem by dividing it into amodal entity segmentation and completion, such that the region to inpaint takes only entities in the non-inpainting area as reference, avoiding the possibility to generate unpredictable sundries. Moreover, we develop two entity segmentation based metrics for quantitatively assessing the performance of object erasure, which are shown be more effective than existing metrics. Experimental results demonstrate that our approach outperforms other state-of-the-art object erasure methods. Our code and data are available at https://zyxunh.github.io/EntityErasure-ProjectPage/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.793, + 0.222, + 0.809 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Object erasure is a widely used technique in the field of image editing, which aims to remove the specified object and fill the area with coherent image content without generating new sundries. Early methods [4, 20, 27] are primarily" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.614 + ], + "angle": 0, + "content": "built upon Generative Adversarial Networks (GANs) [9], but they often do not perform well in open-world scenarios. Thanks to the powerful generation capabilities of diffusion models [32], diffusion-based inpainting [8, 52] has emerged to be a preferable choice for object erasure. However, these methods may generate unwanted sundries (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.81 + ], + "angle": 0, + "content": "In order to address the above problem, PowerPaint [52] proposes to learn task-related embedding for object generation/erasure, and introduces classifier-free-guidance to help avoid generating sundries. CLIPAway [8] leverages CLIP embeddings to make diffusion model focus on background regions while excluding foreground elements. Inst-Paint [44] adopts text prompts to guide diffusion model remove objects, while MagicEraser [17] improves the erasure effect by increasing the background weight and reducing the foreground weight in self attention. However, these methods are less effective in avoiding sundries in complex scenes, which motivates us to tackle object erasure from a different perspective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present EntityErasure, a diffusion-based inpainting method that can generate high-quality object erasure results free of sundries. Particularly, we propose to treat object erasure as a multi-entity amodal completion problem. To do so, we first perform amodal entity segmentation, and then utilize the segmentation output to guide" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.407, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (zhangq93@mail.sysu.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28274" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.28, + 0.485, + 0.406 + ], + "angle": 0, + "content": "Figure 2. Overview of EntityErasure. It contains a amodal entity segmentation model (AESM) and a amodal entity completion model (AECM). The output of AESM serves as input for AECM, AECM then injects content information into the denoising UNet with entity attention (EA). Particularly, EA is designed to constrain the cross-attention mask to focus only on the corresponding entities, ensuring that each region only looks for content information of its corresponding entity, so that the final result is aligned with the amodal entity segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.61 + ], + "angle": 0, + "content": "the completion. Note, this is different from single-entity amodal completion that aims to generate a complete entity for each one, as object erasure requires to repair the missing entities and ensure the sequential order of the objects and the result to be coherent and contextually appropriate. By this means, we are able to predict the corresponding entity of each region, and constrain that each region only utilizes feature information from its corresponding entity during amodal entity completion, thereby preventing the generation of sundries. Considering that existing metrics (such as FID [11] and CLIP Dist. [44]) fail to reflect the effectiveness of object erasure in avoiding sundries, we further develop two metrics for performance evaluation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.611, + 0.411, + 0.626 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.63, + 0.483, + 0.66 + ], + "angle": 0, + "content": "- We present a diffusion-based inpainting method that allows to erase entities without generating sundries." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.666, + 0.484, + 0.711 + ], + "angle": 0, + "content": "- We propose to tackle object erasure from a perspective of amodal entity segmentation and completion, and define two new metrics for performance evaluation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.718, + 0.484, + 0.749 + ], + "angle": 0, + "content": "- We show that our method outperforms state-of-the-art methods on object erasure." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.63, + 0.484, + 0.749 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.234, + 0.779 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.262, + 0.805 + ], + "angle": 0, + "content": "2.1. Image Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Since GANs [9] have demonstrated powerful generative capabilities across various domains, including image generation and editing, numerous GAN-based inpainting methods [23, 27, 45-47] are developed in the past decade. Additionally, LaMa [37] introduces Fourier Convolutions [3] into the inpainting field, enabling better restoration of repetitive tex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "ture information. In recent years, diffusion models have become a preferable choice for generation. Sohl-Dickstein et al. [35] introduced the foundation theory of diffusion, while Dhariwal et al. [5] later demonstrated that diffusion can outperform GANs. To reduce the computational complexity, Stable Diffusion (SD) [32] proposes to scale up the model size and employs a Variational Autoencoder (VAE) [6] to map the original image into latent space, offering a powerful text-to-image model trained on LAION dataset [33] with billions of image-text pairs. Some inpainting methods [32, 38, 50] add new channels or modules to trained diffusion models for better injecting the original image information. For example, SD-inpainting [32] directly combines the inpainting mask and the original image as input. ControlNet [50] replicates a trainable U-Net architecture to facilitate the injection of additional conditional information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Considering that the text prompt may not align with the semantic of the inpainting area, SmartBrush [38] uses text prompt and object shape to guide object inpainting by training with paired object-text data. However, this method is tailored for object generation, not object erasure that we target in this paper. PowerPaint [52] integrates text-guided generation and object erasure within a single model, using different text prompts to represent various tasks. For the erasure task, it introduces classifier-free guidance by using the generation prompt as the negative prompt to reduce sundries. CLIPAway [8] proposes a plug-and-play solution that extracts the background area features as prompts to guide the inpainting model to generate the background and remove the foreground. However, these methods may also generate sundries in some complex scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.849, + 0.59 + ], + "angle": 0, + "content": "2.2. Amodal Segmentation and Completion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Amodal completion [11, 25, 49] aims to generate a complete object when only a partial view of that object is provided as input, while the goal of amodal segmentation [14, 19, 28, 51] is to produce a segmentation map that represents the entire object. As directly completing the entire object is a very challenging task, so various methods primarily focus on amodal segmentation. PCNet [49] proposes to complete the segmentation and then use the completed segmentation to guide subsequent amodal completion. To remedy the difficulty of collecting large-scale amodal data, SeGAN [7] attempts to train on synthetic datasets. However, these approaches still have respective limitations, and only work well for a limited set of closed-world objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "With the emergence of diffusion methods exemplified by Stable Diffusion [32], many amodal completion approaches [26, 40, 48] are also trying to utilize SD. Benefiting from the superior generation capability of SD, these methods are no longer limited to specific scene types or fixed categories. pix2gestalt [26] incorporates visible amodal masks as additional conditional inputs into SD to in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28275" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.092, + 0.825, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.312, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Illustration of entity attention (EA). It leverages amodal entity segmentation to restrain attention map calculated between the entity reference feature and input feature. By limiting each entity's influence solely to pixels within regions predicted by the amodal entity segmentation, it completes corresponding entities as segmentation guidance. Here, \\( Q \\), \\( K \\), and \\( V \\) represent input queries, entity reference keys, and entity reference values, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.486, + 0.575 + ], + "angle": 0, + "content": "form the model which object needs to be completed, then it predicts the shape and appearance of the entire objects. Xu et al. [40] proposes a training-free method that directly uses the pre-trained model of SD-inpaint [32]. It directly blends the visible portions of the object that need to be completed with the background image, which is then fed to the SD-inpainting model to reconstruct the remaining parts. This also highlights the strong correlation between amodal completion and object inpainting, as both tasks involve reconstructing missing information based on available contextual cues. SDAmodal [48] enhances the generalization capabilities of amodal segmentation by incorporating a pre-trained SD model into their pipeline." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.184, + 0.603 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Fig. 2 presents the pipeline of our method. As shown, we firstly feeds the noisy latent \\( z_{T} \\), masked image latent \\( c_{i} \\), and inpainting mask to Amodal Entity Segmentation Model (AESM) to predict amodal entity segmentation. Next, the amodal entity segmentation, masked image latent \\( c_{i} \\), inpainting mask, and the noisy latent \\( z_{t} \\) will be fed to Amodal Entity Completion Model (AECM) to predict the final amodal entity completion result. The noisy latent \\( z_{T} \\) and \\( z_{t} \\) can be set differently. Below we describe our method in detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.35, + 0.789 + ], + "angle": 0, + "content": "3.1. Amodal Entity Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "There are different levels of granularity for object segmentation. A segmentation map of a person can be the entire person, or specific parts such as limbs, the head, or even finer details like the eyes and nose. Segment Anything Model (SAM) [15]'s levels of granularity are various, and it's not easy to find an appropriate level of granularity. Considering this, we choose to use entity segmenta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.543 + ], + "angle": 0, + "content": "tion [29, 30] as our entity unit which defines a complete object or stuff as a single entity. Well-trained diffusion models such as stable diffusion can serve as pretrained models for other vision tasks and have shown strong performance compared to other self-supervised algorithms such as MAE [10]. Inspired by ODISE [39], we use pretrained stable diffusion model [32] as backbone of segmentation. Same with ODISE, we use Mask2Former [2] as our decoder. During testing, it makes only a single prediction. However, with different random noise inputs, it can produce varying amodal entity segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.909, + 0.651 + ], + "angle": 0, + "content": "We use Feature pyramid networks (FPN) [22] to fuse the features of different layers. We set \\( N \\) segmentation queries in the Mask2Former decoder head, then use masked cross attention between queries and FPN layers to exact things and stuff features into queries. Finally, each query generates a category score and a segmentation mask for each thing and stuff. For more details, please refer to [2]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.759, + 0.683 + ], + "angle": 0, + "content": "3.2. Amodal Entity Completion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The initial diffusion-based inpainting method directly adds the inpainting conditional input on text-to-image model and then fine-tunes this model. BrushNet [13] proposes to use a decomposed dual-branch to inject non-inpainting area information into main branch. Inspired by BrushNet, we also use decomposed dual-branch to inject non-inpainting area information to inpainting area of the corresponding entity. However, unlike the depth, edge, and other conditions of ControlNet [50], the generated area and the condition area of inpainting task are not element-wise corresponding, so we choose to inject entity information in the attention layer. Ip-Adapter [43] uses CLIP [31] and Magicanimate [41] uses copied Unet as appearance feature encoder. We choose Unet as our Entity Reference Encoder (ERE) because Unet" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28276" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.107 + ], + "angle": 0, + "content": "can provide more low-level detailed information than CLIP." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.483, + 0.212 + ], + "angle": 0, + "content": "We copy the main branch diffusion model \\(\\mathcal{F}_a(\\cdot;\\Theta_a)\\) as Entity Reference Encoder \\(\\mathcal{F}_b(\\cdot;\\Theta_b)\\). \\(\\mathcal{F}_b(\\cdot;\\Theta_b)\\) is used to extract the features of non-inpainting area, \\(\\Theta_b\\) is its trainable parameters. We incorporate the features from the Entity Reference Encoder (ERE) into the main branch by injecting them into attention layers, where \\(h_b\\) represents the hidden state feature of the ERE:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.227, + 0.483, + 0.243 + ], + "angle": 0, + "content": "\\[\nh _ {b} = \\mathcal {F} _ {b} \\left(c _ {i}; \\Theta_ {b}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.484, + 0.429 + ], + "angle": 0, + "content": "where \\( c_{i} \\) is the condition latent feature, \\( h_{b} \\) contains the information of condition images and processes it to the high-level semantic feature. So we inject \\( h_{b} \\) to cross attention of main branch Unet. The hidden state feature of main branch is represented as \\( h_{a} \\). In order to ensure that each entity area only injects the semantic information of the corresponding entity, inspired by Freestyle [42], we use the amodal entity mask obtained in the previous stage to restrict the attention mask. As shown in Fig. 3, the image query \\( (Q) \\) and entity reference key \\( (K) \\), value \\( (V) \\) are obtained with three projection layers. Then standard attention score map \\( \\mathcal{M} \\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.428, + 0.483, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = \\frac {Q K ^ {T}}{\\sqrt {d}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.497 + ], + "angle": 0, + "content": "where \\(d\\) is the scaling factor. Next we utilize the amodal entity mask to constrain the distribution of the attention maps:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.506, + 0.483, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {M}} _ {i, j} = \\left\\{ \\begin{array}{l l} \\mathcal {M} _ {i, j} & \\text {i f} I _ {i} = = I _ {j}, \\\\ - \\inf & e l s e. \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.596 + ], + "angle": 0, + "content": "where \\( I_{i} \\) is the \\( i \\)-th element's ID of amodal entity segmentation predicted in the first stage. Final attention output can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.606, + 0.483, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} (\\widehat {\\mathcal {M}}) V, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.635, + 0.483, + 0.665 + ], + "angle": 0, + "content": "In this way, only information of the same entity will be injected into the corresponding inpainting area." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.484, + 0.743 + ], + "angle": 0, + "content": "Since the non-inpainting areas are element-wise corresponding, its generation is straightforward, requiring only a copy of the corresponding pixel information. Therefore, we only apply entity attention in the inpainting area. This process is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.751, + 0.483, + 0.77 + ], + "angle": 0, + "content": "\\[\nh _ {o} = I _ {i n} \\cdot W ^ {o 1} \\mathrm {A t t n} (Q, K, V) + (1 - I _ {i n}) \\cdot W ^ {o 2} h _ {b}, (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.826 + ], + "angle": 0, + "content": "where \\( h_{o} \\) is the output hidden state, \\( I_{in} \\) is the inpainting mask, \\( W^{o1} \\) and \\( W^{o2} \\) are two fully connected layers's weight." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.311, + 0.849 + ], + "angle": 0, + "content": "3.3. Classifier-free Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Classifier-free guidance (CFG) [12] is widely used in diffusion methods to enhance the strength of the condition guidance. A positive prompt and a negative prompt are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.349 + ], + "angle": 0, + "content": "required. In PowerPoint [52], a special token is used to represent the erasure task, while another token is used to represent the generation task. Then the erasure token is set to positive prompt, and the generation token is set to negative prompt. This helps enhance the erasure instruction, thereby reducing the possibility of generating sundries. We also use classifier-free guidance to enhance the erasure task. However, unlike PowerPoint, we do not train an additional generation task. Instead, we choose to enhance the guidance of the input condition image. Similar to the IP-Adapter [43], during training, we randomly drop the input condition image with a probability of \\( p_d \\) by replacing it with an all-zero image, \\( p_d \\) is set to 0.05 as default. During testing, the positive input is the condition image containing the non-inpainting regions, while the negative input is an all-zero image. We then use the following linear combination of the conditional and unconditional result:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.364, + 0.907, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c _ {i}, c _ {t}\\right) = (1 + w) \\epsilon_ {\\theta} \\left(z _ {t}, c _ {i}, c _ {t}\\right) - w \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing_ {i}, \\varnothing_ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.907, + 0.47 + ], + "angle": 0, + "content": "where \\( c_{i} \\) and \\( c_{t} \\) are latent features of the condition image and text conditions, respectively, while \\( \\varnothing_{i} \\) is a blank, all-zero latent feature, and \\( \\varnothing_{t} \\) is an empty string. We find that significantly improves the model's erasure performance and the quality of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.731, + 0.499 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.658 + ], + "angle": 0, + "content": "We use about 1750,000 images from Open Images [16] V7 to train our model. Entity segmentation pseudo labels are generated by a pretrained entity segmentation model released by [30], and text prompt pseudo labels are generated with blip [18]. Given a source image, we randomly generate an inpainting mask to occlude some parts of the image. The image with mask occlusion is then treated as the input image, while the entity segmentation of the source image is now the ground truth for amodal entity segmentation, and the source image is the ground truth for amodal completion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We use LaMa's strategy [37] and entity mask to generate the inpainting mask. In addition, since we are addressing an amodal completion task, we constrain that at least \\(40\\%\\) of each entity is not masked. AESM and AECM are trained separately. Both of them are initialized by pretrained RealisticVisionV60B1 [34]. For AESM, the number of mask queries \\(N\\) is set as 100. All parameters of AESM are trainable. We set the batch size to 64, the number of iterations to 80,000, and the learning rate to 1e-5, respectively. For AECM, the main branch Unet is frozen. We use the AdamW [24] optimizer. We set the batch size to 128, the number of iterations to 20,000 iterations, and the learning rate to 5e-5. Input images are resized to maintain their aspect ratio, with a fixed area of around \\(512 \\times 512\\). During inference, we use DDIM [36] sampling with a number of 20 sampling steps." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28277" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.9, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Figure 4. Pipeline of our object erasure metric. The objective of this case is to erase the dog within the inpainting mask. To determine if any new sundries have been generated in the result, we first obtain an entity segmentation of the erased output using an open-world entity segmentation model. We then calculate the Intersection over Self (IoS) for each entity relative to the inpainting mask. Entities with IoS greater than threshold \\(\\lambda\\) are considered newly generated sundries. In this case, the tiger is newly generated." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.302, + 0.85, + 0.478 + ], + "angle": 0, + "content": "
MethodEntitySeg [30]coco [21]
FID↓MSN↓MARS↓ReMOVE↑FID↓MSN↓MARS↓ReMOVE↑
MAT [20]50.00.4330.04486.6725.181.8760.13379.79
SDInpaint [32]57.10.7460.15184.6323.272.9670.28877.14
CLIPAway [8]59.00.6790.13084.2824.172.6860.24078.17
PowerPaint [52]56.90.5540.11584.8025.682.6540.23978.0
Ours58.00.1220.03786.2422.530.5060.08679.56
CLIPAway* [8]60.40.8360.13484.2925.332.7520.22777.11
PowerPaint* [52]58.20.6970.10884.8126.952.7140.22576.93
Ours*59.10.1580.03286.1023.350.5520.08078.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.487, + 0.837, + 0.502 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with state-of-the-art methods. * indicates results produced with loose masks." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.516, + 0.182, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.516, + 0.334, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.516, + 0.422, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.516, + 0.482, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Figure 5. Comparison between our proposed metrics (MSN↓, MARS↓) and CLIP accuracy [44] (CLIP@1,3,5↑). CLIP accuracy only focuses on the difference between the generated result and the original input, so toys with larger differences in the generated image actually lead to an improvement in the CLIP@1 and CLIP@3 metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.224, + 0.748 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.276, + 0.772 + ], + "angle": 0, + "content": "4.1. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "How to evaluate the model's ability of erasing objects has always been a difficult problem. Other inpainting tasks such as text-guide inpainting are generally evaluated by using pre-trained CLIP model to calculate the correlation between the guided text prompt features and the generated image features. The higher the correlation, the more the generated results are in line with expectations. However, for erasing tasks, it is necessary to ensure that no new entities are gener-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.516, + 0.604, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.516, + 0.691, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.516, + 0.753, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.516, + 0.844, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.846, + 0.516, + 0.905, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.618, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Figure 6. Comparison between our proposed metrics (MSN↓, MARS↓) and ReMOVE score [1] (ReMOVE↑). ReMOVE focuses only on whether the generated result is consistent with the surrounding context. The newly generated blue chair is consistent with the surrounding environment, so it received a higher Remove score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.907, + 0.848 + ], + "angle": 0, + "content": "ated and the content of the generated result is part of original entity. Considering this, we directly determine whether any new entities are generated within the inpainting mask. As shown Fig. 4, we use the pre-trained state-of-the-art openworld segmentation model EntitySeg [30] to predict the entity segmentation of the generated result, and then calculate the Intersection over Self (IoS) between each entity with inpainting mask by:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.855, + 0.907, + 0.888 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I o S} (A, B) = \\frac {| A \\cap B |}{| A |}, \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28278" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.089, + 0.168, + 0.101 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.089, + 0.32, + 0.102 + ], + "angle": 0, + "content": "Inpainting mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.089, + 0.437, + 0.102 + ], + "angle": 0, + "content": "Our amodal seg." + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.091, + 0.516, + 0.102 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.103, + 0.206, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.103, + 0.322, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.103, + 0.439, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.103, + 0.555, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.089, + 0.67, + 0.103 + ], + "angle": 0, + "content": "PowerPaint [52]" + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.089, + 0.78, + 0.102 + ], + "angle": 0, + "content": "CLIPAway [8]" + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.089, + 0.882, + 0.102 + ], + "angle": 0, + "content": "MAT [20]" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.162, + 0.205, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.162, + 0.321, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.162, + 0.438, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.162, + 0.554, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.162, + 0.67, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.162, + 0.787, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.162, + 0.902, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.226, + 0.205, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.226, + 0.321, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.225, + 0.438, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.226, + 0.554, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.226, + 0.67, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.226, + 0.786, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.226, + 0.901, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.292, + 0.205, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.292, + 0.321, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.292, + 0.438, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.292, + 0.554, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.292, + 0.67, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.292, + 0.786, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.292, + 0.901, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.353, + 0.205, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.353, + 0.321, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.353, + 0.438, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.353, + 0.554, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.353, + 0.67, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.353, + 0.786, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.353, + 0.901, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.421, + 0.205, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.421, + 0.321, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.421, + 0.438, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.421, + 0.554, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.421, + 0.67, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.421, + 0.786, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.421, + 0.901, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.485, + 0.205, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.485, + 0.321, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.485, + 0.438, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.485, + 0.554, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.485, + 0.67, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.485, + 0.786, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.485, + 0.901, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.553, + 0.205, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.553, + 0.321, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.553, + 0.438, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.553, + 0.554, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.553, + 0.67, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.553, + 0.786, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.553, + 0.901, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.619, + 0.907, + 0.647 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with state-of-the-art methods. As shown, our method effectively erases the target objects while avoiding generating unwanted sundries like the compared methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.793 + ], + "angle": 0, + "content": "where \\( A \\) is the entity segmentation mask and \\( |A| \\) refers to the area of \\( A \\). \\( B \\) is the inpainting mask. If the IoS is greater than the threshold \\( \\lambda \\), these entities are considered as newly generated sundries (SDR). \\( \\lambda \\) is set to 0.95 by default. We traverse all entities and calculate the number of all newly generated sundries (referred to as \\( |SDR| \\)). By setting \\( N \\) as the number of test samples, the Mean Sundries Num (MSN) is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.793, + 0.483, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M S N} = \\frac {| S D R |}{N}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.857 + ], + "angle": 0, + "content": "The Mean Area Ratio of Sundries (MARS) to the inpainting area can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.866, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M A R S} = \\sum_ {A \\in S D R} \\frac {| A \\cap B |}{| B |} \\times \\frac {1}{N}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.9 + ], + "angle": 0, + "content": "MSN reflects the quantity of sundries generated by the model, while MARS indicates the ratio of the sundries area relative to the inpainting mask area. A higher MARS value signifies that the generated sundries are larger and more prominent. Both of these metrics should be minimized for optimal performance. In Fig. 5, we present a comparison between our metric and CLIP accuracy [44]. It is evident that CLIP accuracy only focuses on the magnitude of the change between the generated result and the original image. We also compare our method with ReMOVE [1] in Fig. 6. ReMOVE focuses solely on whether the generated image is consistent with the surrounding context. It fails to provide a reasonable score when the newly generated object (blue chair) is consistent with the surrounding context. In contrast, our metric is more reasonable and reliable." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28279" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.099, + 0.169, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.321, + 0.141, + 0.334 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.099, + 0.246, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.321, + 0.219, + 0.334 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.099, + 0.324, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.321, + 0.296, + 0.334 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.099, + 0.401, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.321, + 0.374, + 0.334 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.099, + 0.478, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.321, + 0.45, + 0.334 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.425 + ], + "angle": 0, + "content": "Figure 8. Effect of entity attention. From left to right: (a) input, (b) inpainting mask, (c) amodal entity segmentation, (d) entity attention, (e) cross attention. As can be seen, with our entity attention, our method faithfully repairs the missing parts of entities according to the amodal entity segmentation, effectively reducing the generation of sundries." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.443, + 0.217, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.544, + 0.166, + 0.556 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.443, + 0.346, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.544, + 0.294, + 0.557 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.443, + 0.475, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.544, + 0.423, + 0.557 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.648 + ], + "angle": 0, + "content": "Figure 9. Impact of applying entity attention on non-inpainting area: (a) input + inpainting mask, (b) not applying entity attention on non-inpainting area, (c) applying entity attention on non-inpainting area. As shown, applying entity attention on non-inpainting areas may lead to unwanted changes to these areas compared to the original image (see the red box)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Since we evaluate our method on non-synthetic real-world scenarios, where paired data is unavailable, we use non-paired metrics for evaluation. Besides MSN and MARS, we also report ReMOVE score [1] and Frechet Inception Distance (FID) [11]. We compare our method with other state-of-the-art inpainting approaches on entity segmentation test dataset [30] and coco val2017 dataset [21]. Excluding stuff objects, the segmentation mask for thing object is selected from each image as the inpainting mask, then we dilate the mask. Meanwhile, we find that CLIP-Away [8] and PowerPaint [52] would generate additional sundries from floating handheld objects due to precise inpainting mask. Therefore, we also conduct comparisons using looser inpainting masks. We obtain looser inpainting masks by dilating masks with larger sizes and filling the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "holes. A total of around 700 test samples are generated from the entity segmentation dataset, and 4000 test samples are generated from the COCO dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.146, + 0.882, + 0.162 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.243 + ], + "angle": 0, + "content": "Baselines. We compare our method with various state-of-the-art inpainting methods, including MAT [20], SD-Inpainting [32], CLIPAway [8], and PowerPaint [52], where MAT is a GAN-based method while the other three are stable diffusion based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Quantitative comparison. As shown in Table 1, our method demonstrates significant advantages over CLIP-Away [8] and PowerPaint [52] in both the MSN and MARS metrics on entity segmentation and coco dataset, using either precise or loose inpainting masks. And among all the diffusion-based methods, our approach also achieves the highest score on the ReMOVE metric. This demonstrates that the proposed model significantly reduces the probability of sundries generation and the area of generated sundries compared with other methods. We also observe that the GAN-based method MAT has relatively low MSN, MARS score and high ReMOVE score compared with SD-based baselines. This may be due to the powerful generative capability of the SD model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.906, + 0.604 + ], + "angle": 0, + "content": "Qualitative comparison. In Fig. 7, we show some prediction results of EntityErasure alongside other state-of-the-art methods. It is evident that MAT generates fewer sundries but their generated textures lack realism. Meanwhile, CLIPAway and PowerPaint produce higher quality results, but still produce sundries in some cases. In contrast, our proposed EntityErasure demonstrates a lower probability of sundries generation and produces higher-quality images. Please see the supplementary material for more results." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.617, + 0.906, + 0.658 + ], + "angle": 0, + "content": "
MethodMATCLIPAwaySDInpaintPowerPaintOurs
Time(s)↓0.084.042.215.002.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.585, + 0.668, + 0.834, + 0.682 + ], + "angle": 0, + "content": "Table 2. Comparison on inference time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Comparison on inference time. In Table 2, we conduct comparison of inference speed on a single RTX 4090. The number of our sampling steps is set as 20. As shown, the inference speed of our method is comparable to those of other diffusion-based methods. Note, we experimentally find that larger number of sampling steps (e.g., 50 steps) does not lead to clear performance improvement, while the inference time cost will be increased." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.667, + 0.85 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To validate the effectiveness of the modules proposed in EntityErasure, we conduct ablation experiments on entity segmentation dataset [30]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28280" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.472, + 0.151 + ], + "angle": 0, + "content": "
MethodFID↓MSN↓MARS↓ReMOVE↑
cross attn57.10.2120.07785.82
entity attn58.00.1220.03786.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.161, + 0.393, + 0.175 + ], + "angle": 0, + "content": "Table 3. Effect of entity attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.484, + 0.382 + ], + "angle": 0, + "content": "Effect of amodal entity segmentation guidance. We use amodal entity segmentation guidance in entity attention to inject content into the corresponding entity. To evaluate its impact, we replace entity attention with standard cross attention for comparison. As shown in Table 3, FID shows a slight decrease, while MSN and MARS increase notably, indicating a rise in sundries generation. This highlights the effectiveness of our amodal entity segmentation guidance in reducing sundries generation. It is further supported by the qualitative ablation results presented in Fig. 8, where model without amodal entity segmentation guidance tends to generate sundries easily." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.525 + ], + "angle": 0, + "content": "Impact of applying entity attention on non-inpainting area. We observe that applying entity attention for injecting information into the non-inpainting area does not incur sundries, but it may cause alterations to the non-inpainting region. As shown in Fig. 9, if entity attention is applied to the non-inpainting area, changes can be observed. This implies that, the non-inpainting area only requires feature injection via element-wise addition, without the need for entity attention." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.536, + 0.462, + 0.63 + ], + "angle": 0, + "content": "
TIFID↓MSN↓MARS↓ReMOVE↑
1159.80.4490.078384.86
1358.00.1220.036986.24
3159.90.4250.075684.88
3358.00.1220.036686.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Table 4. Effect of classifier-free guidance. T: Text-CFG weight, I: Image-CFG weight." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effect of classifier-free guidance. We apply classifier-free guidance to amplify the influence of condition image and amodal entity segmentation guidance. To assess its effectiveness, we perform experiments with varying classifier-free guidance weights. As shown in Table 4, when the weight of Image-CFG is set to 1, effectively turning off Image-CFG, both MSN and MARS increase significantly, which signals an increase in sundries generation. Compared to the impact of Text-CFG, the influence of Image-CFG on the results is more significant. It is further supported by the qualitative ablation results presented in Fig. 4. In the first butterfly image, without CFG, the model would generate excessive textures. This underscores the role of Image-CFG in reducing the occurrence of sundries." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.102, + 0.613, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.102, + 0.71, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.102, + 0.807, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.102, + 0.905, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.146, + 0.611, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.146, + 0.71, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.146, + 0.807, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.146, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.259, + 0.612, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.259, + 0.71, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.259, + 0.807, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.259, + 0.905, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.313, + 0.575, + 0.326 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.313, + 0.672, + 0.326 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.313, + 0.77, + 0.326 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.312, + 0.868, + 0.325 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 10. Effect of classifier-free guidance. From left to right: (a) input, (b) inpainting mask, (c) Imagecfg=3, (d) Imagecfg=1. As shown, utilizing classifier-free guidance to strengthen the guidance from the condition image benefits obtaining cleaner images without sundries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.45, + 0.634, + 0.466 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.664 + ], + "angle": 0, + "content": "In this paper, we propose a novel diffusion-based inpainting method named EntityErasure. We approach the erasure task from a new perspective, framing it as a partial amodal entity completion task. Our method decomposes the erasure process into two steps: partial amodal entity segmentation and partial amodal entity completion, making the erasure more controllable. The proposed entity attention injects entity information with amodal entity segmentation guidance. We ensure that all generated content is completed from the entities in the non-inpainting area, effectively preventing the creation of new sundries. EntityErasure demonstrates state-of-the-art performance in terms of object erasure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.909, + 0.829 + ], + "angle": 0, + "content": "Limitations and future work. Although our method can produce high-quality object erasure results without unwanted sundries, it also has several limitations. First of all, the success of our method may be affected by the quality of amodal segmentation, though our method has some tolerance to poor amodal segmentation (see the supplementary material for validation). Besides, it currently relies on manual mask extension to remove shadows associated with the erased objects. In the future, we will try to address this limitation by incorporating instance shadow detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28281" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In CVPRW, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.229, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[4] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[6] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[7] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[8] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models. In NeurIPS, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.518, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.561, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[13] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[14] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[17] Fan Li, Zixiao Zhang, Yi Huang, Jianzhuang Liu, Renjing Pei, Bin Shao, and Songcen Xu. Magiceraser: Erasing any objects via semantics-aware control. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[19] Ke Li and Jitendra Malik. Amodal instance segmentation. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.905, + 0.265 + ], + "angle": 0, + "content": "[20] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In CVPR, 2022. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.268, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.325, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[23] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[25] Khoi Nguyen and Sinisa Todorovic. A weakly supervised amodal segmenter with boundary uncertainty estimation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[26] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.584 + ], + "angle": 0, + "content": "[27] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[28] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[29] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Philip Torr, Zhe Lin, and Jiaya Jia. Open world entity segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8743-8756, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[30] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High-quality entity segmentation. In ICCV, 2023. 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28282" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.162 + ], + "angle": 0, + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy1, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.164, + 0.392, + 0.176 + ], + "angle": 0, + "content": "[34] SG161222. Realistic vision v6.0 b1, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[39] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[40] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. A modal completion via progressive mixed context diffusion. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[41] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanate: Temporally consistent human image animation using diffusion model. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[42] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[44] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[45] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[46] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[47] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[51] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In ECCV, 2024. 1, 2, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28283" + } + ] +] \ No newline at end of file diff --git a/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_origin.pdf b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b481002817deec54d3e68af0493f87e7a02204d0 --- /dev/null +++ b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/841507d5-5b60-4a74-9926-0461296e0a84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3835ddbeaee215b1d00037258cc969e3fb1082610940b537c0ab6d2963ba2cc7 +size 9487367 diff --git a/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/full.md b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..21474d6288acd57aff8afa31182d608623e2f401 --- /dev/null +++ b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/full.md @@ -0,0 +1,470 @@ +# EntityErasure: Erasing Entity Cleanly via Amodal Entity Segmentation and Completion + +Yixing Zhu $^{1}$ Qing Zhang $^{1,4*}$ Yitong Wang $^{2}$ Yongwei Nie $^{3}$ Wei-Shi Zheng $^{1,4}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China $^{2}$ ByteDance Inc $^{3}$ South China University of Technology, China + +4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +![](images/95d68646c3ec89202c98c78e765196351db5d2edb0a8499ca1e52f188f1ae3ec.jpg) +Input +Figure 1. Comparison with previous methods on object erasure. As shown, our method can cleanly erase objects without generating unwanted sundries that appear in the results of other methods. + +![](images/131cb6b292d5f197c6a2d155c2d628b2b5c6d55b2c4161eb05f11fe8a80d13ab.jpg) +Inpainting mask + +![](images/bfef5d6930f374224a8feaa56a09d4f842594a5420ad71952219f09db9d36daa.jpg) +Our amodal seg. + +![](images/7f4ff6b74efe85b6fd6a7531d95293baa720a4ad925069fb509ad7cb7071e939.jpg) +Ours + +![](images/26c31fc6a41ab27fb110304641abe2cc6823374ba1c8f9a67dca2e1a1a0fa3cf.jpg) +PowerPaint [52] + +![](images/e3f21267127e00dc2955e926d00b8b0a26c92e23fb63d28e1b2072fed5a45498.jpg) +CLIPAway [8] + +![](images/40a1884e85a79121dd26a553b53de8a43c8e68a911cc0ddc563b488979fdfa4b.jpg) +MAT [20] + +# Abstract + +This paper presents EntityErasure, a novel diffusion-based inpainting method that can effectively erase entities without inducing unwanted sundries. To this end, we propose to address this problem by dividing it into amodal entity segmentation and completion, such that the region to inpaint takes only entities in the non-inpainting area as reference, avoiding the possibility to generate unpredictable sundries. Moreover, we develop two entity segmentation based metrics for quantitatively assessing the performance of object erasure, which are shown be more effective than existing metrics. Experimental results demonstrate that our approach outperforms other state-of-the-art object erasure methods. Our code and data are available at https://zyxunh.github.io/EntityErasure-ProjectPage/. + +# 1. Introduction + +Object erasure is a widely used technique in the field of image editing, which aims to remove the specified object and fill the area with coherent image content without generating new sundries. Early methods [4, 20, 27] are primarily + +built upon Generative Adversarial Networks (GANs) [9], but they often do not perform well in open-world scenarios. Thanks to the powerful generation capabilities of diffusion models [32], diffusion-based inpainting [8, 52] has emerged to be a preferable choice for object erasure. However, these methods may generate unwanted sundries (see Fig. 1). + +In order to address the above problem, PowerPaint [52] proposes to learn task-related embedding for object generation/erasure, and introduces classifier-free-guidance to help avoid generating sundries. CLIPAway [8] leverages CLIP embeddings to make diffusion model focus on background regions while excluding foreground elements. Inst-Paint [44] adopts text prompts to guide diffusion model remove objects, while MagicEraser [17] improves the erasure effect by increasing the background weight and reducing the foreground weight in self attention. However, these methods are less effective in avoiding sundries in complex scenes, which motivates us to tackle object erasure from a different perspective. + +In this paper, we present EntityErasure, a diffusion-based inpainting method that can generate high-quality object erasure results free of sundries. Particularly, we propose to treat object erasure as a multi-entity amodal completion problem. To do so, we first perform amodal entity segmentation, and then utilize the segmentation output to guide + +![](images/bc345f8f405e3dcb25c3b48f705f46c8b20bd49c434c1865078c0547a984b951.jpg) +Figure 2. Overview of EntityErasure. It contains a amodal entity segmentation model (AESM) and a amodal entity completion model (AECM). The output of AESM serves as input for AECM, AECM then injects content information into the denoising UNet with entity attention (EA). Particularly, EA is designed to constrain the cross-attention mask to focus only on the corresponding entities, ensuring that each region only looks for content information of its corresponding entity, so that the final result is aligned with the amodal entity segmentation. + +the completion. Note, this is different from single-entity amodal completion that aims to generate a complete entity for each one, as object erasure requires to repair the missing entities and ensure the sequential order of the objects and the result to be coherent and contextually appropriate. By this means, we are able to predict the corresponding entity of each region, and constrain that each region only utilizes feature information from its corresponding entity during amodal entity completion, thereby preventing the generation of sundries. Considering that existing metrics (such as FID [11] and CLIP Dist. [44]) fail to reflect the effectiveness of object erasure in avoiding sundries, we further develop two metrics for performance evaluation. + +In summary, our contributions are as follows: + +- We present a diffusion-based inpainting method that allows to erase entities without generating sundries. +- We propose to tackle object erasure from a perspective of amodal entity segmentation and completion, and define two new metrics for performance evaluation. +- We show that our method outperforms state-of-the-art methods on object erasure. + +# 2. Related Work + +# 2.1. Image Inpainting + +Since GANs [9] have demonstrated powerful generative capabilities across various domains, including image generation and editing, numerous GAN-based inpainting methods [23, 27, 45-47] are developed in the past decade. Additionally, LaMa [37] introduces Fourier Convolutions [3] into the inpainting field, enabling better restoration of repetitive tex + +ture information. In recent years, diffusion models have become a preferable choice for generation. Sohl-Dickstein et al. [35] introduced the foundation theory of diffusion, while Dhariwal et al. [5] later demonstrated that diffusion can outperform GANs. To reduce the computational complexity, Stable Diffusion (SD) [32] proposes to scale up the model size and employs a Variational Autoencoder (VAE) [6] to map the original image into latent space, offering a powerful text-to-image model trained on LAION dataset [33] with billions of image-text pairs. Some inpainting methods [32, 38, 50] add new channels or modules to trained diffusion models for better injecting the original image information. For example, SD-inpainting [32] directly combines the inpainting mask and the original image as input. ControlNet [50] replicates a trainable U-Net architecture to facilitate the injection of additional conditional information. + +Considering that the text prompt may not align with the semantic of the inpainting area, SmartBrush [38] uses text prompt and object shape to guide object inpainting by training with paired object-text data. However, this method is tailored for object generation, not object erasure that we target in this paper. PowerPaint [52] integrates text-guided generation and object erasure within a single model, using different text prompts to represent various tasks. For the erasure task, it introduces classifier-free guidance by using the generation prompt as the negative prompt to reduce sundries. CLIPAway [8] proposes a plug-and-play solution that extracts the background area features as prompts to guide the inpainting model to generate the background and remove the foreground. However, these methods may also generate sundries in some complex scenes. + +# 2.2. Amodal Segmentation and Completion + +Amodal completion [11, 25, 49] aims to generate a complete object when only a partial view of that object is provided as input, while the goal of amodal segmentation [14, 19, 28, 51] is to produce a segmentation map that represents the entire object. As directly completing the entire object is a very challenging task, so various methods primarily focus on amodal segmentation. PCNet [49] proposes to complete the segmentation and then use the completed segmentation to guide subsequent amodal completion. To remedy the difficulty of collecting large-scale amodal data, SeGAN [7] attempts to train on synthetic datasets. However, these approaches still have respective limitations, and only work well for a limited set of closed-world objects. + +With the emergence of diffusion methods exemplified by Stable Diffusion [32], many amodal completion approaches [26, 40, 48] are also trying to utilize SD. Benefiting from the superior generation capability of SD, these methods are no longer limited to specific scene types or fixed categories. pix2gestalt [26] incorporates visible amodal masks as additional conditional inputs into SD to in + +![](images/5b4e9f173f94334ded1560b09b5c5875df730d43c63df65750c68a0337f75b4d.jpg) +Figure 3. Illustration of entity attention (EA). It leverages amodal entity segmentation to restrain attention map calculated between the entity reference feature and input feature. By limiting each entity's influence solely to pixels within regions predicted by the amodal entity segmentation, it completes corresponding entities as segmentation guidance. Here, $Q$ , $K$ , and $V$ represent input queries, entity reference keys, and entity reference values, respectively. + +form the model which object needs to be completed, then it predicts the shape and appearance of the entire objects. Xu et al. [40] proposes a training-free method that directly uses the pre-trained model of SD-inpaint [32]. It directly blends the visible portions of the object that need to be completed with the background image, which is then fed to the SD-inpainting model to reconstruct the remaining parts. This also highlights the strong correlation between amodal completion and object inpainting, as both tasks involve reconstructing missing information based on available contextual cues. SDAmodal [48] enhances the generalization capabilities of amodal segmentation by incorporating a pre-trained SD model into their pipeline. + +# 3. Method + +Fig. 2 presents the pipeline of our method. As shown, we firstly feeds the noisy latent $z_{T}$ , masked image latent $c_{i}$ , and inpainting mask to Amodal Entity Segmentation Model (AESM) to predict amodal entity segmentation. Next, the amodal entity segmentation, masked image latent $c_{i}$ , inpainting mask, and the noisy latent $z_{t}$ will be fed to Amodal Entity Completion Model (AECM) to predict the final amodal entity completion result. The noisy latent $z_{T}$ and $z_{t}$ can be set differently. Below we describe our method in detail. + +# 3.1. Amodal Entity Segmentation + +There are different levels of granularity for object segmentation. A segmentation map of a person can be the entire person, or specific parts such as limbs, the head, or even finer details like the eyes and nose. Segment Anything Model (SAM) [15]'s levels of granularity are various, and it's not easy to find an appropriate level of granularity. Considering this, we choose to use entity segmenta + +tion [29, 30] as our entity unit which defines a complete object or stuff as a single entity. Well-trained diffusion models such as stable diffusion can serve as pretrained models for other vision tasks and have shown strong performance compared to other self-supervised algorithms such as MAE [10]. Inspired by ODISE [39], we use pretrained stable diffusion model [32] as backbone of segmentation. Same with ODISE, we use Mask2Former [2] as our decoder. During testing, it makes only a single prediction. However, with different random noise inputs, it can produce varying amodal entity segmentation results. + +We use Feature pyramid networks (FPN) [22] to fuse the features of different layers. We set $N$ segmentation queries in the Mask2Former decoder head, then use masked cross attention between queries and FPN layers to exact things and stuff features into queries. Finally, each query generates a category score and a segmentation mask for each thing and stuff. For more details, please refer to [2]. + +# 3.2. Amodal Entity Completion + +The initial diffusion-based inpainting method directly adds the inpainting conditional input on text-to-image model and then fine-tunes this model. BrushNet [13] proposes to use a decomposed dual-branch to inject non-inpainting area information into main branch. Inspired by BrushNet, we also use decomposed dual-branch to inject non-inpainting area information to inpainting area of the corresponding entity. However, unlike the depth, edge, and other conditions of ControlNet [50], the generated area and the condition area of inpainting task are not element-wise corresponding, so we choose to inject entity information in the attention layer. Ip-Adapter [43] uses CLIP [31] and Magicanimate [41] uses copied Unet as appearance feature encoder. We choose Unet as our Entity Reference Encoder (ERE) because Unet + +can provide more low-level detailed information than CLIP. + +We copy the main branch diffusion model $\mathcal{F}_a(\cdot;\Theta_a)$ as Entity Reference Encoder $\mathcal{F}_b(\cdot;\Theta_b)$ . $\mathcal{F}_b(\cdot;\Theta_b)$ is used to extract the features of non-inpainting area, $\Theta_b$ is its trainable parameters. We incorporate the features from the Entity Reference Encoder (ERE) into the main branch by injecting them into attention layers, where $h_b$ represents the hidden state feature of the ERE: + +$$ +h _ {b} = \mathcal {F} _ {b} \left(c _ {i}; \Theta_ {b}\right), \tag {1} +$$ + +where $c_{i}$ is the condition latent feature, $h_{b}$ contains the information of condition images and processes it to the high-level semantic feature. So we inject $h_{b}$ to cross attention of main branch Unet. The hidden state feature of main branch is represented as $h_{a}$ . In order to ensure that each entity area only injects the semantic information of the corresponding entity, inspired by Freestyle [42], we use the amodal entity mask obtained in the previous stage to restrict the attention mask. As shown in Fig. 3, the image query $(Q)$ and entity reference key $(K)$ , value $(V)$ are obtained with three projection layers. Then standard attention score map $\mathcal{M}$ is calculated as: + +$$ +\mathcal {M} = \frac {Q K ^ {T}}{\sqrt {d}}, \tag {2} +$$ + +where $d$ is the scaling factor. Next we utilize the amodal entity mask to constrain the distribution of the attention maps: + +$$ +\widehat {\mathcal {M}} _ {i, j} = \left\{ \begin{array}{l l} \mathcal {M} _ {i, j} & \text {i f} I _ {i} = = I _ {j}, \\ - \inf & e l s e. \end{array} \right. \tag {3} +$$ + +where $I_{i}$ is the $i$ -th element's ID of amodal entity segmentation predicted in the first stage. Final attention output can be calculated as: + +$$ +\operatorname {A t t n} (Q, K, V) = \operatorname {s o f t m a x} (\widehat {\mathcal {M}}) V, \tag {4} +$$ + +In this way, only information of the same entity will be injected into the corresponding inpainting area. + +Since the non-inpainting areas are element-wise corresponding, its generation is straightforward, requiring only a copy of the corresponding pixel information. Therefore, we only apply entity attention in the inpainting area. This process is expressed as follows: + +$$ +h _ {o} = I _ {i n} \cdot W ^ {o 1} \mathrm {A t t n} (Q, K, V) + (1 - I _ {i n}) \cdot W ^ {o 2} h _ {b}, (5) +$$ + +where $h_{o}$ is the output hidden state, $I_{in}$ is the inpainting mask, $W^{o1}$ and $W^{o2}$ are two fully connected layers's weight. + +# 3.3. Classifier-free Guidance + +Classifier-free guidance (CFG) [12] is widely used in diffusion methods to enhance the strength of the condition guidance. A positive prompt and a negative prompt are + +required. In PowerPoint [52], a special token is used to represent the erasure task, while another token is used to represent the generation task. Then the erasure token is set to positive prompt, and the generation token is set to negative prompt. This helps enhance the erasure instruction, thereby reducing the possibility of generating sundries. We also use classifier-free guidance to enhance the erasure task. However, unlike PowerPoint, we do not train an additional generation task. Instead, we choose to enhance the guidance of the input condition image. Similar to the IP-Adapter [43], during training, we randomly drop the input condition image with a probability of $p_d$ by replacing it with an all-zero image, $p_d$ is set to 0.05 as default. During testing, the positive input is the condition image containing the non-inpainting regions, while the negative input is an all-zero image. We then use the following linear combination of the conditional and unconditional result: + +$$ +\tilde {\epsilon} _ {\theta} \left(z _ {t}, c _ {i}, c _ {t}\right) = (1 + w) \epsilon_ {\theta} \left(z _ {t}, c _ {i}, c _ {t}\right) - w \epsilon_ {\theta} \left(z _ {t}, \varnothing_ {i}, \varnothing_ {t}\right), \tag {6} +$$ + +where $c_{i}$ and $c_{t}$ are latent features of the condition image and text conditions, respectively, while $\varnothing_{i}$ is a blank, all-zero latent feature, and $\varnothing_{t}$ is an empty string. We find that significantly improves the model's erasure performance and the quality of the generated images. + +# 3.4. Implementation Details + +We use about 1750,000 images from Open Images [16] V7 to train our model. Entity segmentation pseudo labels are generated by a pretrained entity segmentation model released by [30], and text prompt pseudo labels are generated with blip [18]. Given a source image, we randomly generate an inpainting mask to occlude some parts of the image. The image with mask occlusion is then treated as the input image, while the entity segmentation of the source image is now the ground truth for amodal entity segmentation, and the source image is the ground truth for amodal completion. + +We use LaMa's strategy [37] and entity mask to generate the inpainting mask. In addition, since we are addressing an amodal completion task, we constrain that at least $40\%$ of each entity is not masked. AESM and AECM are trained separately. Both of them are initialized by pretrained RealisticVisionV60B1 [34]. For AESM, the number of mask queries $N$ is set as 100. All parameters of AESM are trainable. We set the batch size to 64, the number of iterations to 80,000, and the learning rate to 1e-5, respectively. For AECM, the main branch Unet is frozen. We use the AdamW [24] optimizer. We set the batch size to 128, the number of iterations to 20,000 iterations, and the learning rate to 5e-5. Input images are resized to maintain their aspect ratio, with a fixed area of around $512 \times 512$ . During inference, we use DDIM [36] sampling with a number of 20 sampling steps. + +![](images/80e271e934f55ea4ddbbc95231d64d6617031bc66a738e0b495c8d1628697020.jpg) +Figure 4. Pipeline of our object erasure metric. The objective of this case is to erase the dog within the inpainting mask. To determine if any new sundries have been generated in the result, we first obtain an entity segmentation of the erased output using an open-world entity segmentation model. We then calculate the Intersection over Self (IoS) for each entity relative to the inpainting mask. Entities with IoS greater than threshold $\lambda$ are considered newly generated sundries. In this case, the tiger is newly generated. + +
MethodEntitySeg [30]coco [21]
FID↓MSN↓MARS↓ReMOVE↑FID↓MSN↓MARS↓ReMOVE↑
MAT [20]50.00.4330.04486.6725.181.8760.13379.79
SDInpaint [32]57.10.7460.15184.6323.272.9670.28877.14
CLIPAway [8]59.00.6790.13084.2824.172.6860.24078.17
PowerPaint [52]56.90.5540.11584.8025.682.6540.23978.0
Ours58.00.1220.03786.2422.530.5060.08679.56
CLIPAway* [8]60.40.8360.13484.2925.332.7520.22777.11
PowerPaint* [52]58.20.6970.10884.8126.952.7140.22576.93
Ours*59.10.1580.03286.1023.350.5520.08078.05
+ +Table 1. Quantitative comparison with state-of-the-art methods. * indicates results produced with loose masks. + +![](images/66a0b663ee869ca819ab1edbe7e5eb7502de074bcd18ffe5895b413f1909009f.jpg) +Figure 5. Comparison between our proposed metrics (MSN↓, MARS↓) and CLIP accuracy [44] (CLIP@1,3,5↑). CLIP accuracy only focuses on the difference between the generated result and the original input, so toys with larger differences in the generated image actually lead to an improvement in the CLIP@1 and CLIP@3 metrics. + +![](images/90ec4b2fd28fa710a30f8957a81f8610297d3fc41ecd9c0348655fa81e989e9e.jpg) + +![](images/5a9bac4596dcee38e8013e38e34050d2c5bf0292e62dbcc938b37b10f4e99381.jpg) + +![](images/30b5ea547f78dfd30e476e2dd14638d387af5e41b6f9aec665d44f3b1f94113e.jpg) +Figure 6. Comparison between our proposed metrics (MSN↓, MARS↓) and ReMOVE score [1] (ReMOVE↑). ReMOVE focuses only on whether the generated result is consistent with the surrounding context. The newly generated blue chair is consistent with the surrounding environment, so it received a higher Remove score. + +# 4. Experiments + +# 4.1. Evaluation Metrics + +How to evaluate the model's ability of erasing objects has always been a difficult problem. Other inpainting tasks such as text-guide inpainting are generally evaluated by using pre-trained CLIP model to calculate the correlation between the guided text prompt features and the generated image features. The higher the correlation, the more the generated results are in line with expectations. However, for erasing tasks, it is necessary to ensure that no new entities are gener- + +![](images/d578b70186f94bd3086fe8b379766c6164ae4b49ec27f86fa99106b442619f17.jpg) + +![](images/c505d8ff9871d5484ae85515cf6612fef4b380828e7e47e761155c20427beb71.jpg) + +![](images/75b74904cb51644ad26135748d562ffec6284924418c825af9fcb76e2358e7fb.jpg) + +![](images/478a1f70c6dafde6d4cc288961b65ff1b6451e8ba99c4b01979a46b823abdd3f.jpg) + +![](images/90dd6b9e56583e5004ea6c59840422df6bf3ea200249486362aa29a43ca9be55.jpg) + +ated and the content of the generated result is part of original entity. Considering this, we directly determine whether any new entities are generated within the inpainting mask. As shown Fig. 4, we use the pre-trained state-of-the-art openworld segmentation model EntitySeg [30] to predict the entity segmentation of the generated result, and then calculate the Intersection over Self (IoS) between each entity with inpainting mask by: + +$$ +\operatorname {I o S} (A, B) = \frac {| A \cap B |}{| A |}, \tag {7} +$$ + +![](images/bfddbc5192a6c050cdc7b10e19145b6b8e188bead0f98757d475f1de96939de9.jpg) +Input + +![](images/503d41e9a14b0142fc1f28e346284f96156f12f2b020cc6b3b2231e57b89ab35.jpg) +Inpainting mask + +![](images/8ecf565fc4b9d4e2a3c0042659ac12a272f99c05fdd53bb578ebd9e4bd1ed3c6.jpg) +Our amodal seg. + +![](images/29f2c9794fcf195b5474a75db5ad150c85a4519132a27c5e548cf7c44fe415d7.jpg) +Ours +PowerPaint [52] + +![](images/420c230bc84ec3701523eb95f4c5f85f95fd74ef302dee4c7e07308985dc5157.jpg) + +![](images/7f2b14d98689993b3ff77d4d38e43fc9a0d83fb886e643dd652b63fb6c9bdbb4.jpg) + +![](images/8b2c05a09a2d42dc456cea9181bc0a29207e753a8c8b9677e073200e8bcbda4e.jpg) + +![](images/3fe885ed4a4c35dd60e1b461ca4dc32bb4f0bc4c77116a30433cd009d47d0271.jpg) + +![](images/821b9930147fe0860fceaa12030efad941ed7071a6f7e2dfb89de06ffafd2e8d.jpg) + +![](images/c168f87aa8cae347ac3075e7e17d0884dd8b82eb3e96ce3f0aed52a35505628e.jpg) +CLIPAway [8] + +![](images/244df2dae948974448fb1d1c43c6d3139dfe2db042746499b7e52772a8ecc2fd.jpg) +MAT [20] + +![](images/b33196469bb7bad83aaa453dc570c0dde2650c1369163d3b6a2fea6a9d0a8b2e.jpg) + +![](images/00e54a27e0f28e5962baad0240bb6d931a7501427112fce45a8c11fa5f504fec.jpg) + +![](images/5ece6eead487f878dcfa11f9c05023c2e73ef0328d58389a369287bc0b3cffe6.jpg) + +![](images/607ca004c99a37075c00d810382430cab655d043d218f5768efa768166d7eeb5.jpg) + +![](images/d7be25138b0c60bc8e4a40a4ae02ea471488e3282079cc3d4691f384c45ebf08.jpg) + +![](images/5fc8ad3c7a0d33380ea55ffd71219a62ae5a2e12fedc216081887171c7854818.jpg) + +![](images/5b4f44ef2aa7bc0092c38a802bd12e84611bdcf7206b301f02d313c7d9b8eb41.jpg) + +![](images/5a07f3fc7b0b55c793961bf30678fc3ccbbb678c7e77e3b0312e714850b9771b.jpg) + +![](images/2e9f3a4ba94da554f22161c462b5cbc18b1bd7028711b129b9d99acafa1ad57b.jpg) + +![](images/7d872abbaab29b3766b0af1385e885fffa61f03f77128e8078f25f3f1f6d7468.jpg) + +![](images/e07a5a659da4041911111cdb7a69d7b2617912fe0e8e2d0b87942bda0cb23b03.jpg) + +![](images/1139b27b4cd9fc48c7a41387b852ef066939d4654898dbcbd6b6894da8cbb598.jpg) + +![](images/5970e85c01d66b196072f3e680f3744a73c3981bd3bbb769b3b2002adad172f7.jpg) + +![](images/784b21c6b154c0aefffa48ea7afdf14f5845404a98f9bca9068d706c22760d8b.jpg) + +![](images/7c32db83d5dbb778ddf3c48f744dac297bc50c0f5c41e609c518f81cfc284802.jpg) + +![](images/f23fc56921060d77673eaf644301da08ec766844001208aff3b710a68da3ab0d.jpg) + +![](images/2c570319cc88043640bf6696e37a17a9b166dc5b4386530f7a80430b9c05e5e9.jpg) + +![](images/6813dd7a55f627650a5217ec289190ddbbb03f0982b28e3c90512f3fbd37c5d0.jpg) + +![](images/11f228a7ffd52f14fe2e395e6a7f3f6ae965ee4215f68577e75f4cde79754bc5.jpg) + +![](images/54c3d9bb194c7a80cc7b395e45430c1854515df480bc7c9846e942b9c12297bf.jpg) + +![](images/e7cb8337df6863f3882f3820179085a6abedf9b61600797006b9391aea50d087.jpg) + +![](images/80bc1a3d5e1c8b7d61b2abf6f85f73627b33b938deecb0f3754a330af92abcb2.jpg) + +![](images/8531769c74a7afa17953508a8604a2b45b160a6cc3361be849ee6fae81169aa8.jpg) + +![](images/e2b3c4aef2d166c200ae4d89db70f18c8924ab52d0ee34c603556e18404a581b.jpg) + +![](images/b27b8735aa964fe2463500be9b15c7d25a0d2acb72f10438062ec2e918565ac9.jpg) + +![](images/17b65febfd662f7db2a494f3eb870ae247f51a8406c9eddf63b0d72a53c35500.jpg) + +![](images/903588f52c5235edf9c0f8105e4a919c6492f9d222f2b3a4d3590f70693e6cad.jpg) + +![](images/ab1cb0c4abb2c9c265aedc9bba2b2fdcb907c4051bb6a0d9c2d1b44fd23f4cf6.jpg) + +![](images/940862c7ca1f5d2b7653d9b11b0ed1519ea638e3031a58ad5b95a811ef56213f.jpg) + +![](images/697b3b982cb243fbbc9d8479f28ae6bbac7b03140a38247a23c39900c717710b.jpg) + +![](images/651e9193db581e5fea2be522fbc57d104dc5cecc17c827954d6430071abc259a.jpg) + +![](images/a0de4e5c79f125e6003575cdc6d41456401c7975659cb85e2fd81e64f937c4a4.jpg) + +![](images/75e13bdadcce2aa28797a8bfa4ad573282a3d45eb459d16f0ddfd8752607e8bd.jpg) + +![](images/1fa7abd658bedc6cb099a77400efb4013dafcfc57218199d07ad356560e0a3e7.jpg) + +![](images/ea0cc1002a0555877001172bb5c4aacddb94836c15be7049fc4e86952dfde631.jpg) + +![](images/b1032b9177d616fda0b3d78baa3d609873d5996cdf06f9dee7f788a291f19820.jpg) +Figure 7. Qualitative comparison with state-of-the-art methods. As shown, our method effectively erases the target objects while avoiding generating unwanted sundries like the compared methods. + +![](images/dcf0071bf0608a84ccaaad113640db89a9cc30100366270a512782b104450717.jpg) + +![](images/8032237434a1cbe42d18c4d38c856e5e8935d2febc1bd3ccec1b51f4d320239f.jpg) + +![](images/558d1ac2aacdebcb047abdbb82d6b23df2ed4c0bba9f1ae760823b203f8eca20.jpg) + +![](images/8c9b8c512f9c2987c42080d6e8bcacbcf6e9807dc43a08702d5aa0568a88c92e.jpg) + +![](images/79de545d0cabd71e8e420a81ef6e0aa61bb8212ceca6b920b2259b1d0c2e379c.jpg) + +![](images/ef807835733c81372c38eb27bc9bf78332fa6933b3bf83a668254b950e717947.jpg) + +where $A$ is the entity segmentation mask and $|A|$ refers to the area of $A$ . $B$ is the inpainting mask. If the IoS is greater than the threshold $\lambda$ , these entities are considered as newly generated sundries (SDR). $\lambda$ is set to 0.95 by default. We traverse all entities and calculate the number of all newly generated sundries (referred to as $|SDR|$ ). By setting $N$ as the number of test samples, the Mean Sundries Num (MSN) is defined by: + +$$ +\mathrm {M S N} = \frac {| S D R |}{N}. \tag {8} +$$ + +The Mean Area Ratio of Sundries (MARS) to the inpainting area can be computed by: + +$$ +\mathrm {M A R S} = \sum_ {A \in S D R} \frac {| A \cap B |}{| B |} \times \frac {1}{N}, \tag {9} +$$ + +MSN reflects the quantity of sundries generated by the model, while MARS indicates the ratio of the sundries area relative to the inpainting mask area. A higher MARS value signifies that the generated sundries are larger and more prominent. Both of these metrics should be minimized for optimal performance. In Fig. 5, we present a comparison between our metric and CLIP accuracy [44]. It is evident that CLIP accuracy only focuses on the magnitude of the change between the generated result and the original image. We also compare our method with ReMOVE [1] in Fig. 6. ReMOVE focuses solely on whether the generated image is consistent with the surrounding context. It fails to provide a reasonable score when the newly generated object (blue chair) is consistent with the surrounding context. In contrast, our metric is more reasonable and reliable. + +![](images/51b05103348b269308f9b3dea1d27894befa6774c448c5948283df0bdec8bf8d.jpg) +(a) + +![](images/44eb1bb973a22f2011546a8eeaf001df50774afb85df8d63fc5f8b0a67999cb5.jpg) +(b) + +![](images/83772e6c4613e244c45a53dcf65de447bab0ae71bcf9029fefa2a9c1a4dcfdf8.jpg) +(c) + +![](images/50e74cbddf0052dab63aae9c97f1e4a5f09da4fb657778c0b09c2cd9cee5cc9a.jpg) +(d) + +![](images/21ddf1320cc8ba197025e24cfa98b0bbadf6012e8eb2c7866dc4f8f4479146fe.jpg) +(e) + +![](images/02cf79cfe422caae5aaab0d27257ebc9b985c01c8200273658494c8b102b7d37.jpg) +Figure 8. Effect of entity attention. From left to right: (a) input, (b) inpainting mask, (c) amodal entity segmentation, (d) entity attention, (e) cross attention. As can be seen, with our entity attention, our method faithfully repairs the missing parts of entities according to the amodal entity segmentation, effectively reducing the generation of sundries. +(a) +Figure 9. Impact of applying entity attention on non-inpainting area: (a) input + inpainting mask, (b) not applying entity attention on non-inpainting area, (c) applying entity attention on non-inpainting area. As shown, applying entity attention on non-inpainting areas may lead to unwanted changes to these areas compared to the original image (see the red box). + +![](images/41d1d2616f9ec7f9d265ccb12686877434c828aad90b6760c600779d15f2d6cd.jpg) +(b) + +![](images/68f1136591874d8648bc205429838f09ee8fb4aac2f7da07f1a06c6c802e88ac.jpg) +(c) + +Since we evaluate our method on non-synthetic real-world scenarios, where paired data is unavailable, we use non-paired metrics for evaluation. Besides MSN and MARS, we also report ReMOVE score [1] and Frechet Inception Distance (FID) [11]. We compare our method with other state-of-the-art inpainting approaches on entity segmentation test dataset [30] and coco val2017 dataset [21]. Excluding stuff objects, the segmentation mask for thing object is selected from each image as the inpainting mask, then we dilate the mask. Meanwhile, we find that CLIP-Away [8] and PowerPaint [52] would generate additional sundries from floating handheld objects due to precise inpainting mask. Therefore, we also conduct comparisons using looser inpainting masks. We obtain looser inpainting masks by dilating masks with larger sizes and filling the + +holes. A total of around 700 test samples are generated from the entity segmentation dataset, and 4000 test samples are generated from the COCO dataset. + +# 4.2. Comparison with State-of-the-Art Methods + +Baselines. We compare our method with various state-of-the-art inpainting methods, including MAT [20], SD-Inpainting [32], CLIPAway [8], and PowerPaint [52], where MAT is a GAN-based method while the other three are stable diffusion based methods. + +Quantitative comparison. As shown in Table 1, our method demonstrates significant advantages over CLIP-Away [8] and PowerPaint [52] in both the MSN and MARS metrics on entity segmentation and coco dataset, using either precise or loose inpainting masks. And among all the diffusion-based methods, our approach also achieves the highest score on the ReMOVE metric. This demonstrates that the proposed model significantly reduces the probability of sundries generation and the area of generated sundries compared with other methods. We also observe that the GAN-based method MAT has relatively low MSN, MARS score and high ReMOVE score compared with SD-based baselines. This may be due to the powerful generative capability of the SD model. + +Qualitative comparison. In Fig. 7, we show some prediction results of EntityErasure alongside other state-of-the-art methods. It is evident that MAT generates fewer sundries but their generated textures lack realism. Meanwhile, CLIPAway and PowerPaint produce higher quality results, but still produce sundries in some cases. In contrast, our proposed EntityErasure demonstrates a lower probability of sundries generation and produces higher-quality images. Please see the supplementary material for more results. + +
MethodMATCLIPAwaySDInpaintPowerPaintOurs
Time(s)↓0.084.042.215.002.98
+ +Table 2. Comparison on inference time. + +Comparison on inference time. In Table 2, we conduct comparison of inference speed on a single RTX 4090. The number of our sampling steps is set as 20. As shown, the inference speed of our method is comparable to those of other diffusion-based methods. Note, we experimentally find that larger number of sampling steps (e.g., 50 steps) does not lead to clear performance improvement, while the inference time cost will be increased. + +# 4.3. Ablation Study + +To validate the effectiveness of the modules proposed in EntityErasure, we conduct ablation experiments on entity segmentation dataset [30]. + +
MethodFID↓MSN↓MARS↓ReMOVE↑
cross attn57.10.2120.07785.82
entity attn58.00.1220.03786.24
+ +Effect of amodal entity segmentation guidance. We use amodal entity segmentation guidance in entity attention to inject content into the corresponding entity. To evaluate its impact, we replace entity attention with standard cross attention for comparison. As shown in Table 3, FID shows a slight decrease, while MSN and MARS increase notably, indicating a rise in sundries generation. This highlights the effectiveness of our amodal entity segmentation guidance in reducing sundries generation. It is further supported by the qualitative ablation results presented in Fig. 8, where model without amodal entity segmentation guidance tends to generate sundries easily. + +Impact of applying entity attention on non-inpainting area. We observe that applying entity attention for injecting information into the non-inpainting area does not incur sundries, but it may cause alterations to the non-inpainting region. As shown in Fig. 9, if entity attention is applied to the non-inpainting area, changes can be observed. This implies that, the non-inpainting area only requires feature injection via element-wise addition, without the need for entity attention. + +Table 3. Effect of entity attention. + +
TIFID↓MSN↓MARS↓ReMOVE↑
1159.80.4490.078384.86
1358.00.1220.036986.24
3159.90.4250.075684.88
3358.00.1220.036686.24
+ +Table 4. Effect of classifier-free guidance. T: Text-CFG weight, I: Image-CFG weight. + +Effect of classifier-free guidance. We apply classifier-free guidance to amplify the influence of condition image and amodal entity segmentation guidance. To assess its effectiveness, we perform experiments with varying classifier-free guidance weights. As shown in Table 4, when the weight of Image-CFG is set to 1, effectively turning off Image-CFG, both MSN and MARS increase significantly, which signals an increase in sundries generation. Compared to the impact of Text-CFG, the influence of Image-CFG on the results is more significant. It is further supported by the qualitative ablation results presented in Fig. 4. In the first butterfly image, without CFG, the model would generate excessive textures. This underscores the role of Image-CFG in reducing the occurrence of sundries. + +![](images/89b2e00ee18584c6f805ad1e551cf7f88545473b8f0b28895637be1d52f228d3.jpg) + +![](images/607a1a117cd0a8da134e1f3e601f548928cbaba0c7660dacc27475ffed1089b4.jpg) + +![](images/fa05338e5d91422fad3a92143773165207beb87fc23cc63f77f62307bf8eb9e3.jpg) + +![](images/3d2bdc1818bb1946e85eaadb2b215460d9fdce54db13d9c409155d12d7fdfa0a.jpg) + +![](images/e62895cb0ae35471ffd1a96be84b5018b3f9b2bf8d7972c8e7ea54ffd8d2a8fe.jpg) + +![](images/9e8a86f809f1258c07ec97084883695bef4ab0045833947e2ff487daa61ff6a9.jpg) + +![](images/2a55a6f9f04cd77eae6dc26fb55e97321c6fae6d63d8fff890647aa5d93fc938.jpg) + +![](images/32b15670a7dcdb48723cd7e244ca814eb50c3a61e85d40e7d2ef07e28cf40e17.jpg) + +![](images/34a6b55b8831a96591dc5a3fc4ed00f8ed84dbe3c6126d2ee883d12335c5c967.jpg) +(a) + +![](images/85c419f09dc6c7c4598325eb7ea9e247caf3030c950deae04af63a0cc70ca5b9.jpg) +(b) + +![](images/3697e753b11bc3d3f69a6e4866f199c2da02ddb1cb30ccd35389e6a402c156c1.jpg) +(c) + +![](images/3bba271f19a65a45172d4499f88ce862ebe243d4c38858c6c64de9b982b6b352.jpg) +(d) +Figure 10. Effect of classifier-free guidance. From left to right: (a) input, (b) inpainting mask, (c) Imagecfg=3, (d) Imagecfg=1. As shown, utilizing classifier-free guidance to strengthen the guidance from the condition image benefits obtaining cleaner images without sundries. + +# 5. Conclusion + +In this paper, we propose a novel diffusion-based inpainting method named EntityErasure. We approach the erasure task from a new perspective, framing it as a partial amodal entity completion task. Our method decomposes the erasure process into two steps: partial amodal entity segmentation and partial amodal entity completion, making the erasure more controllable. The proposed entity attention injects entity information with amodal entity segmentation guidance. We ensure that all generated content is completed from the entities in the non-inpainting area, effectively preventing the creation of new sundries. EntityErasure demonstrates state-of-the-art performance in terms of object erasure. + +Limitations and future work. Although our method can produce high-quality object erasure results without unwanted sundries, it also has several limitations. First of all, the success of our method may be affected by the quality of amodal segmentation, though our method has some tolerance to poor amodal segmentation (see the supplementary material for validation). Besides, it currently relies on manual mask extension to remove shadows associated with the erased objects. In the future, we will try to address this limitation by incorporating instance shadow detection. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002). + +# References + +[1] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In CVPRW, 2024. 5, 6, 7 +[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 3 +[3] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. In NeurIPS, 2020. 2 +[4] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1 +[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2 +[6] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2 +[7] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2 +[8] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models. In NeurIPS, 2024. 1, 2, 5, 6, 7 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2 +[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 7 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 4 +[13] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024. 3 +[14] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2 +[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3 +[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 4 + +[17] Fan Li, Zixiao Zhang, Yi Huang, Jianzhuang Liu, Renjing Pei, Bin Shao, and Songcen Xu. Magiceraser: Erasing any objects via semantics-aware control. In ECCV, 2024. 1 +[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 4 +[19] Ke Li and Jitendra Malik. Amodal instance segmentation. In ECCV, 2016. 2 +[20] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In CVPR, 2022. 1, 5, 6, 7 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7 +[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3 +[23] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4 +[25] Khoi Nguyen and Sinisa Todorovic. A weakly supervised amodal segmenter with boundary uncertainty estimation. In ICCV, 2021. 2 +[26] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In CVPR, 2024. 2 +[27] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1, 2 +[28] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2 +[29] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Philip Torr, Zhe Lin, and Jiaya Jia. Open world entity segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8743-8756, 2022. 3 +[30] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High-quality entity segmentation. In ICCV, 2023. 3, 4, 5, 7 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 7 +[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo + +Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy1, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2 +[34] SG161222. Realistic vision v6.0 b1, 2024. 4 +[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2 +[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4 +[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 2, 4 +[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2 +[39] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3 +[40] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. A modal completion via progressive mixed context diffusion. In CVPR, 2024. 2, 3 +[41] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanate: Temporally consistent human image animation using diffusion model. In CVPR, 2024. 3 +[42] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In CVPR, 2023. 4 +[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 3, 4 +[44] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 2, 5, 6 +[45] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2 +[46] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In CVPR, 2019. +[47] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2 +[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, 2024. 2, 3 +[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, 2020. 2 + +[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3 +[51] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2 +[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In ECCV, 2024. 1, 2, 4, 5, 6, 7 \ No newline at end of file diff --git a/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/images.zip b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..705be286771e22ad86cea3a0a6bf63a5ecb3e1a2 --- /dev/null +++ b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a8bf0ec4a3a59d0b5e8033665120d26c54536e82746ae5a3be7a1fa421dd2aa +size 1016103 diff --git a/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/layout.json b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64c7df79bad2e909b0193ab160e675b1ab97fc44 --- /dev/null +++ b/2025/EntityErasure_ Erasing Entity Cleanly via Amodal Entity Segmentation and Completion/layout.json @@ -0,0 +1,11790 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "text", + "content": "EntityErasure: Erasing Entity Cleanly via Amodal Entity Segmentation and Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "text", + "content": "Yixing Zhu" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "text", + "content": " Qing Zhang" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "inline_equation", + "content": "^{1,4*}" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "text", + "content": " Yitong Wang" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "text", + "content": " Yongwei Nie" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "text", + "content": " Wei-Shi Zheng" + }, + { + "bbox": [ + 104, + 160, + 506, + 176 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "spans": [ + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University, China " + }, + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "type": "text", + "content": "ByteDance Inc " + }, + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 117, + 175, + 493, + 204 + ], + "type": "text", + "content": "South China University of Technology, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 203, + 541, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 203, + 541, + 217 + ], + "spans": [ + { + "bbox": [ + 66, + 203, + 541, + 217 + ], + "type": "text", + "content": "4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 232, + 126, + 381 + ], + "blocks": [ + { + "bbox": [ + 79, + 221, + 102, + 232 + ], + "lines": [ + { + "bbox": [ + 79, + 221, + 102, + 232 + ], + "spans": [ + { + "bbox": [ + 79, + 221, + 102, + 232 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 232, + 126, + 381 + ], + "lines": [ + { + "bbox": [ + 57, + 232, + 126, + 381 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 126, + 381 + ], + "type": "image", + "image_path": "95d68646c3ec89202c98c78e765196351db5d2edb0a8499ca1e52f188f1ae3ec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Comparison with previous methods on object erasure. As shown, our method can cleanly erase objects without generating unwanted sundries that appear in the results of other methods." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 127, + 232, + 197, + 381 + ], + "blocks": [ + { + "bbox": [ + 129, + 221, + 196, + 232 + ], + "lines": [ + { + "bbox": [ + 129, + 221, + 196, + 232 + ], + "spans": [ + { + "bbox": [ + 129, + 221, + 196, + 232 + ], + "type": "text", + "content": "Inpainting mask" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 232, + 197, + 381 + ], + "lines": [ + { + "bbox": [ + 127, + 232, + 197, + 381 + ], + "spans": [ + { + "bbox": [ + 127, + 232, + 197, + 381 + ], + "type": "image", + "image_path": "131cb6b292d5f197c6a2d155c2d628b2b5c6d55b2c4161eb05f11fe8a80d13ab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 198, + 232, + 268, + 381 + ], + "blocks": [ + { + "bbox": [ + 199, + 221, + 266, + 232 + ], + "lines": [ + { + "bbox": [ + 199, + 221, + 266, + 232 + ], + "spans": [ + { + "bbox": [ + 199, + 221, + 266, + 232 + ], + "type": "text", + "content": "Our amodal seg." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 198, + 232, + 268, + 381 + ], + "lines": [ + { + "bbox": [ + 198, + 232, + 268, + 381 + ], + "spans": [ + { + "bbox": [ + 198, + 232, + 268, + 381 + ], + "type": "image", + "image_path": "bfef5d6930f374224a8feaa56a09d4f842594a5420ad71952219f09db9d36daa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 269, + 232, + 338, + 381 + ], + "blocks": [ + { + "bbox": [ + 293, + 222, + 314, + 232 + ], + "lines": [ + { + "bbox": [ + 293, + 222, + 314, + 232 + ], + "spans": [ + { + "bbox": [ + 293, + 222, + 314, + 232 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 269, + 232, + 338, + 381 + ], + "lines": [ + { + "bbox": [ + 269, + 232, + 338, + 381 + ], + "spans": [ + { + "bbox": [ + 269, + 232, + 338, + 381 + ], + "type": "image", + "image_path": "7f4ff6b74efe85b6fd6a7531d95293baa720a4ad925069fb509ad7cb7071e939.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 340, + 232, + 410, + 381 + ], + "blocks": [ + { + "bbox": [ + 341, + 221, + 408, + 232 + ], + "lines": [ + { + "bbox": [ + 341, + 221, + 408, + 232 + ], + "spans": [ + { + "bbox": [ + 341, + 221, + 408, + 232 + ], + "type": "text", + "content": "PowerPaint [52]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 232, + 410, + 381 + ], + "lines": [ + { + "bbox": [ + 340, + 232, + 410, + 381 + ], + "spans": [ + { + "bbox": [ + 340, + 232, + 410, + 381 + ], + "type": "image", + "image_path": "26c31fc6a41ab27fb110304641abe2cc6823374ba1c8f9a67dca2e1a1a0fa3cf.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 411, + 232, + 481, + 381 + ], + "blocks": [ + { + "bbox": [ + 416, + 221, + 476, + 232 + ], + "lines": [ + { + "bbox": [ + 416, + 221, + 476, + 232 + ], + "spans": [ + { + "bbox": [ + 416, + 221, + 476, + 232 + ], + "type": "text", + "content": "CLIPAway [8]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 232, + 481, + 381 + ], + "lines": [ + { + "bbox": [ + 411, + 232, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 411, + 232, + 481, + 381 + ], + "type": "image", + "image_path": "e3f21267127e00dc2955e926d00b8b0a26c92e23fb63d28e1b2072fed5a45498.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 482, + 232, + 552, + 381 + ], + "blocks": [ + { + "bbox": [ + 496, + 221, + 537, + 232 + ], + "lines": [ + { + "bbox": [ + 496, + 221, + 537, + 232 + ], + "spans": [ + { + "bbox": [ + 496, + 221, + 537, + 232 + ], + "type": "text", + "content": "MAT [20]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 482, + 232, + 552, + 381 + ], + "lines": [ + { + "bbox": [ + 482, + 232, + 552, + 381 + ], + "spans": [ + { + "bbox": [ + 482, + 232, + 552, + 381 + ], + "type": "image", + "image_path": "40a1884e85a79121dd26a553b53de8a43c8e68a911cc0ddc563b488979fdfa4b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 151, + 413, + 200, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 413, + 200, + 425 + ], + "spans": [ + { + "bbox": [ + 151, + 413, + 200, + 425 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 437, + 297, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 297, + 606 + ], + "type": "text", + "content": "This paper presents EntityErasure, a novel diffusion-based inpainting method that can effectively erase entities without inducing unwanted sundries. To this end, we propose to address this problem by dividing it into amodal entity segmentation and completion, such that the region to inpaint takes only entities in the non-inpainting area as reference, avoiding the possibility to generate unpredictable sundries. Moreover, we develop two entity segmentation based metrics for quantitatively assessing the performance of object erasure, which are shown be more effective than existing metrics. Experimental results demonstrate that our approach outperforms other state-of-the-art object erasure methods. Our code and data are available at https://zyxunh.github.io/EntityErasure-ProjectPage/." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 628, + 135, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 135, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 135, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "Object erasure is a widely used technique in the field of image editing, which aims to remove the specified object and fill the area with coherent image content without generating new sundries. Early methods [4, 20, 27] are primarily" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": "built upon Generative Adversarial Networks (GANs) [9], but they often do not perform well in open-world scenarios. Thanks to the powerful generation capabilities of diffusion models [32], diffusion-based inpainting [8, 52] has emerged to be a preferable choice for object erasure. However, these methods may generate unwanted sundries (see Fig. 1)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 486, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 641 + ], + "type": "text", + "content": "In order to address the above problem, PowerPaint [52] proposes to learn task-related embedding for object generation/erasure, and introduces classifier-free-guidance to help avoid generating sundries. CLIPAway [8] leverages CLIP embeddings to make diffusion model focus on background regions while excluding foreground elements. Inst-Paint [44] adopts text prompts to guide diffusion model remove objects, while MagicEraser [17] improves the erasure effect by increasing the background weight and reducing the foreground weight in self attention. However, these methods are less effective in avoiding sundries in complex scenes, which motivates us to tackle object erasure from a different perspective." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "In this paper, we present EntityErasure, a diffusion-based inpainting method that can generate high-quality object erasure results free of sundries. Particularly, we propose to treat object erasure as a multi-entity amodal completion problem. To do so, we first perform amodal entity segmentation, and then utilize the segmentation output to guide" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 249, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 249, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 249, + 713 + ], + "type": "text", + "content": "*Corresponding author (zhangq93@mail.sysu.edu.cn)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28274" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 217 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 217 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 217 + ], + "type": "image", + "image_path": "bc345f8f405e3dcb25c3b48f705f46c8b20bd49c434c1865078c0547a984b951.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 221, + 296, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 321 + ], + "type": "text", + "content": "Figure 2. Overview of EntityErasure. It contains a amodal entity segmentation model (AESM) and a amodal entity completion model (AECM). The output of AESM serves as input for AECM, AECM then injects content information into the denoising UNet with entity attention (EA). Particularly, EA is designed to constrain the cross-attention mask to focus only on the corresponding entities, ensuring that each region only looks for content information of its corresponding entity, so that the final result is aligned with the amodal entity segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": "the completion. Note, this is different from single-entity amodal completion that aims to generate a complete entity for each one, as object erasure requires to repair the missing entities and ensure the sequential order of the objects and the result to be coherent and contextually appropriate. By this means, we are able to predict the corresponding entity of each region, and constrain that each region only utilizes feature information from its corresponding entity during amodal entity completion, thereby preventing the generation of sundries. Considering that existing metrics (such as FID [11] and CLIP Dist. [44]) fail to reflect the effectiveness of object erasure in avoiding sundries, we further develop two metrics for performance evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 483, + 251, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 483, + 251, + 495 + ], + "spans": [ + { + "bbox": [ + 67, + 483, + 251, + 495 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 498, + 296, + 593 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 498, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 498, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 67, + 498, + 295, + 522 + ], + "type": "text", + "content": "- We present a diffusion-based inpainting method that allows to erase entities without generating sundries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 527, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 527, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 67, + 527, + 296, + 563 + ], + "type": "text", + "content": "- We propose to tackle object erasure from a perspective of amodal entity segmentation and completion, and define two new metrics for performance evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 568, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 568, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 67, + 568, + 296, + 593 + ], + "type": "text", + "content": "- We show that our method outperforms state-of-the-art methods on object erasure." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 604, + 143, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 143, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 143, + 616 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 160, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 160, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 160, + 637 + ], + "type": "text", + "content": "2.1. Image Inpainting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Since GANs [9] have demonstrated powerful generative capabilities across various domains, including image generation and editing, numerous GAN-based inpainting methods [23, 27, 45-47] are developed in the past decade. Additionally, LaMa [37] introduces Fourier Convolutions [3] into the inpainting field, enabling better restoration of repetitive tex" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "ture information. In recent years, diffusion models have become a preferable choice for generation. Sohl-Dickstein et al. [35] introduced the foundation theory of diffusion, while Dhariwal et al. [5] later demonstrated that diffusion can outperform GANs. To reduce the computational complexity, Stable Diffusion (SD) [32] proposes to scale up the model size and employs a Variational Autoencoder (VAE) [6] to map the original image into latent space, offering a powerful text-to-image model trained on LAION dataset [33] with billions of image-text pairs. Some inpainting methods [32, 38, 50] add new channels or modules to trained diffusion models for better injecting the original image information. For example, SD-inpainting [32] directly combines the inpainting mask and the original image as input. ControlNet [50] replicates a trainable U-Net architecture to facilitate the injection of additional conditional information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 265, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 445 + ], + "type": "text", + "content": "Considering that the text prompt may not align with the semantic of the inpainting area, SmartBrush [38] uses text prompt and object shape to guide object inpainting by training with paired object-text data. However, this method is tailored for object generation, not object erasure that we target in this paper. PowerPaint [52] integrates text-guided generation and object erasure within a single model, using different text prompts to represent various tasks. For the erasure task, it introduces classifier-free guidance by using the generation prompt as the negative prompt to reduce sundries. CLIPAway [8] proposes a plug-and-play solution that extracts the background area features as prompts to guide the inpainting model to generate the background and remove the foreground. However, these methods may also generate sundries in some complex scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 455, + 519, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 519, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 519, + 467 + ], + "type": "text", + "content": "2.2. Amodal Segmentation and Completion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "text", + "content": "Amodal completion [11, 25, 49] aims to generate a complete object when only a partial view of that object is provided as input, while the goal of amodal segmentation [14, 19, 28, 51] is to produce a segmentation map that represents the entire object. As directly completing the entire object is a very challenging task, so various methods primarily focus on amodal segmentation. PCNet [49] proposes to complete the segmentation and then use the completed segmentation to guide subsequent amodal completion. To remedy the difficulty of collecting large-scale amodal data, SeGAN [7] attempts to train on synthetic datasets. However, these approaches still have respective limitations, and only work well for a limited set of closed-world objects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "With the emergence of diffusion methods exemplified by Stable Diffusion [32], many amodal completion approaches [26, 40, 48] are also trying to utilize SD. Benefiting from the superior generation capability of SD, these methods are no longer limited to specific scene types or fixed categories. pix2gestalt [26] incorporates visible amodal masks as additional conditional inputs into SD to in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28275" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 72, + 504, + 243 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 504, + 243 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 504, + 243 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 504, + 243 + ], + "type": "image", + "image_path": "5b4e9f173f94334ded1560b09b5c5875df730d43c63df65750c68a0337f75b4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "type": "text", + "content": "Figure 3. Illustration of entity attention (EA). It leverages amodal entity segmentation to restrain attention map calculated between the entity reference feature and input feature. By limiting each entity's influence solely to pixels within regions predicted by the amodal entity segmentation, it completes corresponding entities as segmentation guidance. Here, " + }, + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 247, + 555, + 291 + ], + "type": "text", + "content": " represent input queries, entity reference keys, and entity reference values, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 298, + 297, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 297, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 297, + 455 + ], + "type": "text", + "content": "form the model which object needs to be completed, then it predicts the shape and appearance of the entire objects. Xu et al. [40] proposes a training-free method that directly uses the pre-trained model of SD-inpaint [32]. It directly blends the visible portions of the object that need to be completed with the background image, which is then fed to the SD-inpainting model to reconstruct the remaining parts. This also highlights the strong correlation between amodal completion and object inpainting, as both tasks involve reconstructing missing information based on available contextual cues. SDAmodal [48] enhances the generalization capabilities of amodal segmentation by incorporating a pre-trained SD model into their pipeline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 464, + 112, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 112, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 112, + 477 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": "Fig. 2 presents the pipeline of our method. As shown, we firstly feeds the noisy latent " + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": ", masked image latent " + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": ", and inpainting mask to Amodal Entity Segmentation Model (AESM) to predict amodal entity segmentation. Next, the amodal entity segmentation, masked image latent " + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": ", inpainting mask, and the noisy latent " + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": " will be fed to Amodal Entity Completion Model (AECM) to predict the final amodal entity completion result. The noisy latent " + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": " can be set differently. Below we describe our method in detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 214, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 214, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 214, + 624 + ], + "type": "text", + "content": "3.1. Amodal Entity Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "There are different levels of granularity for object segmentation. A segmentation map of a person can be the entire person, or specific parts such as limbs, the head, or even finer details like the eyes and nose. Segment Anything Model (SAM) [15]'s levels of granularity are various, and it's not easy to find an appropriate level of granularity. Considering this, we choose to use entity segmenta" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 298, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 430 + ], + "type": "text", + "content": "tion [29, 30] as our entity unit which defines a complete object or stuff as a single entity. Well-trained diffusion models such as stable diffusion can serve as pretrained models for other vision tasks and have shown strong performance compared to other self-supervised algorithms such as MAE [10]. Inspired by ODISE [39], we use pretrained stable diffusion model [32] as backbone of segmentation. Same with ODISE, we use Mask2Former [2] as our decoder. During testing, it makes only a single prediction. However, with different random noise inputs, it can produce varying amodal entity segmentation results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 431, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 556, + 515 + ], + "type": "text", + "content": "We use Feature pyramid networks (FPN) [22] to fuse the features of different layers. We set " + }, + { + "bbox": [ + 313, + 431, + 556, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 431, + 556, + 515 + ], + "type": "text", + "content": " segmentation queries in the Mask2Former decoder head, then use masked cross attention between queries and FPN layers to exact things and stuff features into queries. Finally, each query generates a category score and a segmentation mask for each thing and stuff. For more details, please refer to [2]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 527, + 464, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 464, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 464, + 540 + ], + "type": "text", + "content": "3.2. Amodal Entity Completion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "The initial diffusion-based inpainting method directly adds the inpainting conditional input on text-to-image model and then fine-tunes this model. BrushNet [13] proposes to use a decomposed dual-branch to inject non-inpainting area information into main branch. Inspired by BrushNet, we also use decomposed dual-branch to inject non-inpainting area information to inpainting area of the corresponding entity. However, unlike the depth, edge, and other conditions of ControlNet [50], the generated area and the condition area of inpainting task are not element-wise corresponding, so we choose to inject entity information in the attention layer. Ip-Adapter [43] uses CLIP [31] and Magicanimate [41] uses copied Unet as appearance feature encoder. We choose Unet as our Entity Reference Encoder (ERE) because Unet" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28276" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 84 + ], + "type": "text", + "content": "can provide more low-level detailed information than CLIP." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "text", + "content": "We copy the main branch diffusion model " + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_a(\\cdot;\\Theta_a)" + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "text", + "content": " as Entity Reference Encoder " + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_b(\\cdot;\\Theta_b)" + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_b(\\cdot;\\Theta_b)" + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "text", + "content": " is used to extract the features of non-inpainting area, " + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "inline_equation", + "content": "\\Theta_b" + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "text", + "content": " is its trainable parameters. We incorporate the features from the Entity Reference Encoder (ERE) into the main branch by injecting them into attention layers, where " + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "inline_equation", + "content": "h_b" + }, + { + "bbox": [ + 54, + 84, + 295, + 167 + ], + "type": "text", + "content": " represents the hidden state feature of the ERE:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 179, + 295, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 179, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 139, + 179, + 295, + 192 + ], + "type": "interline_equation", + "content": "h _ {b} = \\mathcal {F} _ {b} \\left(c _ {i}; \\Theta_ {b}\\right), \\tag {1}", + "image_path": "655424ed034d0638be968faa8ddfee744d8c16c37897733c3115f3b1b4cffea0.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": " is the condition latent feature, " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "h_{b}" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": " contains the information of condition images and processes it to the high-level semantic feature. So we inject " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "h_{b}" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": " to cross attention of main branch Unet. The hidden state feature of main branch is represented as " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "h_{a}" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": ". In order to ensure that each entity area only injects the semantic information of the corresponding entity, inspired by Freestyle [42], we use the amodal entity mask obtained in the previous stage to restrict the attention mask. As shown in Fig. 3, the image query " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": " and entity reference key " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": ", value " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": " are obtained with three projection layers. Then standard attention score map " + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 198, + 296, + 339 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 338, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 338, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 146, + 338, + 295, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = \\frac {Q K ^ {T}}{\\sqrt {d}}, \\tag {2}", + "image_path": "c34c31f32393f6ce5e149000074c4c0079140ecf5f63aa2a79ca47fc1af06e8b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 368, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 368, + 296, + 393 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 368, + 296, + 393 + ], + "type": "text", + "content": " is the scaling factor. Next we utilize the amodal entity mask to constrain the distribution of the attention maps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 400, + 295, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 400, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 111, + 400, + 295, + 433 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {M}} _ {i, j} = \\left\\{ \\begin{array}{l l} \\mathcal {M} _ {i, j} & \\text {i f} I _ {i} = = I _ {j}, \\\\ - \\inf & e l s e. \\end{array} \\right. \\tag {3}", + "image_path": "c0cd9f110f2991610d93e915250c3ab724670b0fdd215c4dd2b994ab5edc1d10.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "text", + "content": "-th element's ID of amodal entity segmentation predicted in the first stage. Final attention output can be calculated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 479, + 295, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 479, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 107, + 479, + 295, + 495 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} (\\widehat {\\mathcal {M}}) V, \\tag {4}", + "image_path": "4ee260acc477386cfd25b5eb6d29f9d198eef6cf36fe20c1bb60c883b3a9562c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 502, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 502, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 502, + 295, + 526 + ], + "type": "text", + "content": "In this way, only information of the same entity will be injected into the corresponding inpainting area." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 527, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 296, + 588 + ], + "type": "text", + "content": "Since the non-inpainting areas are element-wise corresponding, its generation is straightforward, requiring only a copy of the corresponding pixel information. Therefore, we only apply entity attention in the inpainting area. This process is expressed as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 594, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 594, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 594, + 295, + 609 + ], + "type": "interline_equation", + "content": "h _ {o} = I _ {i n} \\cdot W ^ {o 1} \\mathrm {A t t n} (Q, K, V) + (1 - I _ {i n}) \\cdot W ^ {o 2} h _ {b}, (5)", + "image_path": "1fd0271c70b660bdcd0a8e1c1bbf61f85659404ee15dedbf98cc4989a0e3130e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "h_{o}" + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": " is the output hidden state, " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": " is the inpainting mask, " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "W^{o1}" + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "W^{o2}" + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": " are two fully connected layers's weight." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 660, + 190, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 190, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 190, + 672 + ], + "type": "text", + "content": "3.3. Classifier-free Guidance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Classifier-free guidance (CFG) [12] is widely used in diffusion methods to enhance the strength of the condition guidance. A positive prompt and a negative prompt are" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "content": "required. In PowerPoint [52], a special token is used to represent the erasure task, while another token is used to represent the generation task. Then the erasure token is set to positive prompt, and the generation token is set to negative prompt. This helps enhance the erasure instruction, thereby reducing the possibility of generating sundries. We also use classifier-free guidance to enhance the erasure task. However, unlike PowerPoint, we do not train an additional generation task. Instead, we choose to enhance the guidance of the input condition image. Similar to the IP-Adapter [43], during training, we randomly drop the input condition image with a probability of " + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "content": " by replacing it with an all-zero image, " + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "content": " is set to 0.05 as default. During testing, the positive input is the condition image containing the non-inpainting regions, while the negative input is an all-zero image. We then use the following linear combination of the conditional and unconditional result:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 288, + 555, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 288, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 318, + 288, + 555, + 300 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c _ {i}, c _ {t}\\right) = (1 + w) \\epsilon_ {\\theta} \\left(z _ {t}, c _ {i}, c _ {t}\\right) - w \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing_ {i}, \\varnothing_ {t}\\right), \\tag {6}", + "image_path": "31af88880b702b7cc650895d631fd245a559efbe2faf65e5afe3b5687bb94492.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "text", + "content": " are latent features of the condition image and text conditions, respectively, while " + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\varnothing_{i}" + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "text", + "content": " is a blank, all-zero latent feature, and " + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\varnothing_{t}" + }, + { + "bbox": [ + 313, + 312, + 555, + 372 + ], + "type": "text", + "content": " is an empty string. We find that significantly improves the model's erasure performance and the quality of the generated images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 382, + 447, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 447, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 447, + 395 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 401, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 521 + ], + "type": "text", + "content": "We use about 1750,000 images from Open Images [16] V7 to train our model. Entity segmentation pseudo labels are generated by a pretrained entity segmentation model released by [30], and text prompt pseudo labels are generated with blip [18]. Given a source image, we randomly generate an inpainting mask to occlude some parts of the image. The image with mask occlusion is then treated as the input image, while the entity segmentation of the source image is now the ground truth for amodal entity segmentation, and the source image is the ground truth for amodal completion." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "We use LaMa's strategy [37] and entity mask to generate the inpainting mask. In addition, since we are addressing an amodal completion task, we constrain that at least " + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": " of each entity is not masked. AESM and AECM are trained separately. Both of them are initialized by pretrained RealisticVisionV60B1 [34]. For AESM, the number of mask queries " + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": " is set as 100. All parameters of AESM are trainable. We set the batch size to 64, the number of iterations to 80,000, and the learning rate to 1e-5, respectively. For AECM, the main branch Unet is frozen. We use the AdamW [24] optimizer. We set the batch size to 128, the number of iterations to 20,000 iterations, and the learning rate to 5e-5. Input images are resized to maintain their aspect ratio, with a fixed area of around " + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": ". During inference, we use DDIM [36] sampling with a number of 20 sampling steps." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28277" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 550, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 178 + ], + "type": "image", + "image_path": "80e271e934f55ea4ddbbc95231d64d6617031bc66a738e0b495c8d1628697020.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 184, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 555, + 231 + ], + "type": "text", + "content": "Figure 4. Pipeline of our object erasure metric. The objective of this case is to erase the dog within the inpainting mask. To determine if any new sundries have been generated in the result, we first obtain an entity segmentation of the erased output using an open-world entity segmentation model. We then calculate the Intersection over Self (IoS) for each entity relative to the inpainting mask. Entities with IoS greater than threshold " + }, + { + "bbox": [ + 55, + 184, + 555, + 231 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 184, + 555, + 231 + ], + "type": "text", + "content": " are considered newly generated sundries. In this case, the tiger is newly generated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 91, + 239, + 520, + 378 + ], + "blocks": [ + { + "bbox": [ + 91, + 239, + 520, + 378 + ], + "lines": [ + { + "bbox": [ + 91, + 239, + 520, + 378 + ], + "spans": [ + { + "bbox": [ + 91, + 239, + 520, + 378 + ], + "type": "table", + "html": "
MethodEntitySeg [30]coco [21]
FID↓MSN↓MARS↓ReMOVE↑FID↓MSN↓MARS↓ReMOVE↑
MAT [20]50.00.4330.04486.6725.181.8760.13379.79
SDInpaint [32]57.10.7460.15184.6323.272.9670.28877.14
CLIPAway [8]59.00.6790.13084.2824.172.6860.24078.17
PowerPaint [52]56.90.5540.11584.8025.682.6540.23978.0
Ours58.00.1220.03786.2422.530.5060.08679.56
CLIPAway* [8]60.40.8360.13484.2925.332.7520.22777.11
PowerPaint* [52]58.20.6970.10884.8126.952.7140.22576.93
Ours*59.10.1580.03286.1023.350.5520.08078.05
", + "image_path": "7e496f1f40c4ffd9635487fab289b2c041b2193abad210b8ff9d84597a8c6ddb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 385, + 512, + 397 + ], + "lines": [ + { + "bbox": [ + 96, + 385, + 512, + 397 + ], + "spans": [ + { + "bbox": [ + 96, + 385, + 512, + 397 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with state-of-the-art methods. * indicates results produced with loose masks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 408, + 111, + 483 + ], + "blocks": [ + { + "bbox": [ + 59, + 408, + 111, + 483 + ], + "lines": [ + { + "bbox": [ + 59, + 408, + 111, + 483 + ], + "spans": [ + { + "bbox": [ + 59, + 408, + 111, + 483 + ], + "type": "image", + "image_path": "66a0b663ee869ca819ab1edbe7e5eb7502de074bcd18ffe5895b413f1909009f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 490, + 296, + 555 + ], + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 555 + ], + "type": "text", + "content": "Figure 5. Comparison between our proposed metrics (MSN↓, MARS↓) and CLIP accuracy [44] (CLIP@1,3,5↑). CLIP accuracy only focuses on the difference between the generated result and the original input, so toys with larger differences in the generated image actually lead to an improvement in the CLIP@1 and CLIP@3 metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 117, + 408, + 204, + 483 + ], + "blocks": [ + { + "bbox": [ + 117, + 408, + 204, + 483 + ], + "lines": [ + { + "bbox": [ + 117, + 408, + 204, + 483 + ], + "spans": [ + { + "bbox": [ + 117, + 408, + 204, + 483 + ], + "type": "image", + "image_path": "90ec4b2fd28fa710a30f8957a81f8610297d3fc41ecd9c0348655fa81e989e9e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 206, + 408, + 258, + 483 + ], + "blocks": [ + { + "bbox": [ + 206, + 408, + 258, + 483 + ], + "lines": [ + { + "bbox": [ + 206, + 408, + 258, + 483 + ], + "spans": [ + { + "bbox": [ + 206, + 408, + 258, + 483 + ], + "type": "image", + "image_path": "5a9bac4596dcee38e8013e38e34050d2c5bf0292e62dbcc938b37b10f4e99381.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 258, + 408, + 294, + 483 + ], + "blocks": [ + { + "bbox": [ + 258, + 408, + 294, + 483 + ], + "lines": [ + { + "bbox": [ + 258, + 408, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 258, + 408, + 294, + 483 + ], + "type": "image", + "image_path": "30b5ea547f78dfd30e476e2dd14638d387af5e41b6f9aec665d44f3b1f94113e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 489, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 489, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 555, + 555 + ], + "type": "text", + "content": "Figure 6. Comparison between our proposed metrics (MSN↓, MARS↓) and ReMOVE score [1] (ReMOVE↑). ReMOVE focuses only on whether the generated result is consistent with the surrounding context. The newly generated blue chair is consistent with the surrounding environment, so it received a higher Remove score." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 578, + 137, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 137, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 137, + 592 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 168, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 168, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 168, + 611 + ], + "type": "text", + "content": "4.1. Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "How to evaluate the model's ability of erasing objects has always been a difficult problem. Other inpainting tasks such as text-guide inpainting are generally evaluated by using pre-trained CLIP model to calculate the correlation between the guided text prompt features and the generated image features. The higher the correlation, the more the generated results are in line with expectations. However, for erasing tasks, it is necessary to ensure that no new entities are gener-" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 408, + 369, + 479 + ], + "blocks": [ + { + "bbox": [ + 317, + 408, + 369, + 479 + ], + "lines": [ + { + "bbox": [ + 317, + 408, + 369, + 479 + ], + "spans": [ + { + "bbox": [ + 317, + 408, + 369, + 479 + ], + "type": "image", + "image_path": "d578b70186f94bd3086fe8b379766c6164ae4b49ec27f86fa99106b442619f17.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 370, + 408, + 422, + 479 + ], + "blocks": [ + { + "bbox": [ + 370, + 408, + 422, + 479 + ], + "lines": [ + { + "bbox": [ + 370, + 408, + 422, + 479 + ], + "spans": [ + { + "bbox": [ + 370, + 408, + 422, + 479 + ], + "type": "image", + "image_path": "c505d8ff9871d5484ae85515cf6612fef4b380828e7e47e761155c20427beb71.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 423, + 408, + 460, + 480 + ], + "blocks": [ + { + "bbox": [ + 423, + 408, + 460, + 480 + ], + "lines": [ + { + "bbox": [ + 423, + 408, + 460, + 480 + ], + "spans": [ + { + "bbox": [ + 423, + 408, + 460, + 480 + ], + "type": "image", + "image_path": "75b74904cb51644ad26135748d562ffec6284924418c825af9fcb76e2358e7fb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 462, + 408, + 516, + 480 + ], + "blocks": [ + { + "bbox": [ + 462, + 408, + 516, + 480 + ], + "lines": [ + { + "bbox": [ + 462, + 408, + 516, + 480 + ], + "spans": [ + { + "bbox": [ + 462, + 408, + 516, + 480 + ], + "type": "image", + "image_path": "478a1f70c6dafde6d4cc288961b65ff1b6451e8ba99c4b01979a46b823abdd3f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 517, + 408, + 553, + 480 + ], + "blocks": [ + { + "bbox": [ + 517, + 408, + 553, + 480 + ], + "lines": [ + { + "bbox": [ + 517, + 408, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 517, + 408, + 553, + 480 + ], + "type": "image", + "image_path": "90dd6b9e56583e5004ea6c59840422df6bf3ea200249486362aa29a43ca9be55.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 574, + 555, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 671 + ], + "type": "text", + "content": "ated and the content of the generated result is part of original entity. Considering this, we directly determine whether any new entities are generated within the inpainting mask. As shown Fig. 4, we use the pre-trained state-of-the-art openworld segmentation model EntitySeg [30] to predict the entity segmentation of the generated result, and then calculate the Intersection over Self (IoS) between each entity with inpainting mask by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 386, + 677, + 555, + 703 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 677, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 386, + 677, + 555, + 703 + ], + "type": "interline_equation", + "content": "\\operatorname {I o S} (A, B) = \\frac {| A \\cap B |}{| A |}, \\tag {7}", + "image_path": "85e3524f1e13663472f4fd7588f834968a5adcf736d5c4862ce9ea0b9c38865f.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28278" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 81, + 126, + 125 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 102, + 79 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 102, + 79 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 102, + 79 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 81, + 126, + 125 + ], + "lines": [ + { + "bbox": [ + 57, + 81, + 126, + 125 + ], + "spans": [ + { + "bbox": [ + 57, + 81, + 126, + 125 + ], + "type": "image", + "image_path": "bfddbc5192a6c050cdc7b10e19145b6b8e188bead0f98757d475f1de96939de9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 127, + 81, + 197, + 125 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 195, + 80 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 195, + 80 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 195, + 80 + ], + "type": "text", + "content": "Inpainting mask" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 81, + 197, + 125 + ], + "lines": [ + { + "bbox": [ + 127, + 81, + 197, + 125 + ], + "spans": [ + { + "bbox": [ + 127, + 81, + 197, + 125 + ], + "type": "image", + "image_path": "503d41e9a14b0142fc1f28e346284f96156f12f2b020cc6b3b2231e57b89ab35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 199, + 81, + 268, + 125 + ], + "blocks": [ + { + "bbox": [ + 200, + 70, + 267, + 80 + ], + "lines": [ + { + "bbox": [ + 200, + 70, + 267, + 80 + ], + "spans": [ + { + "bbox": [ + 200, + 70, + 267, + 80 + ], + "type": "text", + "content": "Our amodal seg." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 199, + 81, + 268, + 125 + ], + "lines": [ + { + "bbox": [ + 199, + 81, + 268, + 125 + ], + "spans": [ + { + "bbox": [ + 199, + 81, + 268, + 125 + ], + "type": "image", + "image_path": "8ecf565fc4b9d4e2a3c0042659ac12a272f99c05fdd53bb578ebd9e4bd1ed3c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 271, + 81, + 339, + 125 + ], + "blocks": [ + { + "bbox": [ + 295, + 72, + 315, + 80 + ], + "lines": [ + { + "bbox": [ + 295, + 72, + 315, + 80 + ], + "spans": [ + { + "bbox": [ + 295, + 72, + 315, + 80 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 271, + 81, + 339, + 125 + ], + "lines": [ + { + "bbox": [ + 271, + 81, + 339, + 125 + ], + "spans": [ + { + "bbox": [ + 271, + 81, + 339, + 125 + ], + "type": "image", + "image_path": "29f2c9794fcf195b5474a75db5ad150c85a4519132a27c5e548cf7c44fe415d7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 70, + 410, + 81 + ], + "lines": [ + { + "bbox": [ + 343, + 70, + 410, + 81 + ], + "spans": [ + { + "bbox": [ + 343, + 70, + 410, + 81 + ], + "type": "text", + "content": "PowerPaint [52]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 128, + 125, + 175 + ], + "blocks": [ + { + "bbox": [ + 57, + 128, + 125, + 175 + ], + "lines": [ + { + "bbox": [ + 57, + 128, + 125, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 128, + 125, + 175 + ], + "type": "image", + "image_path": "420c230bc84ec3701523eb95f4c5f85f95fd74ef302dee4c7e07308985dc5157.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 127, + 128, + 196, + 175 + ], + "blocks": [ + { + "bbox": [ + 127, + 128, + 196, + 175 + ], + "lines": [ + { + "bbox": [ + 127, + 128, + 196, + 175 + ], + "spans": [ + { + "bbox": [ + 127, + 128, + 196, + 175 + ], + "type": "image", + "image_path": "7f2b14d98689993b3ff77d4d38e43fc9a0d83fb886e643dd652b63fb6c9bdbb4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 198, + 128, + 268, + 175 + ], + "blocks": [ + { + "bbox": [ + 198, + 128, + 268, + 175 + ], + "lines": [ + { + "bbox": [ + 198, + 128, + 268, + 175 + ], + "spans": [ + { + "bbox": [ + 198, + 128, + 268, + 175 + ], + "type": "image", + "image_path": "8b2c05a09a2d42dc456cea9181bc0a29207e753a8c8b9677e073200e8bcbda4e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 270, + 128, + 339, + 175 + ], + "blocks": [ + { + "bbox": [ + 270, + 128, + 339, + 175 + ], + "lines": [ + { + "bbox": [ + 270, + 128, + 339, + 175 + ], + "spans": [ + { + "bbox": [ + 270, + 128, + 339, + 175 + ], + "type": "image", + "image_path": "3fe885ed4a4c35dd60e1b461ca4dc32bb4f0bc4c77116a30433cd009d47d0271.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 341, + 128, + 410, + 175 + ], + "blocks": [ + { + "bbox": [ + 341, + 128, + 410, + 175 + ], + "lines": [ + { + "bbox": [ + 341, + 128, + 410, + 175 + ], + "spans": [ + { + "bbox": [ + 341, + 128, + 410, + 175 + ], + "type": "image", + "image_path": "821b9930147fe0860fceaa12030efad941ed7071a6f7e2dfb89de06ffafd2e8d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 413, + 128, + 481, + 175 + ], + "blocks": [ + { + "bbox": [ + 419, + 70, + 477, + 80 + ], + "lines": [ + { + "bbox": [ + 419, + 70, + 477, + 80 + ], + "spans": [ + { + "bbox": [ + 419, + 70, + 477, + 80 + ], + "type": "text", + "content": "CLIPAway [8]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 128, + 481, + 175 + ], + "lines": [ + { + "bbox": [ + 413, + 128, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 413, + 128, + 481, + 175 + ], + "type": "image", + "image_path": "c168f87aa8cae347ac3075e7e17d0884dd8b82eb3e96ce3f0aed52a35505628e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 484, + 128, + 552, + 175 + ], + "blocks": [ + { + "bbox": [ + 499, + 70, + 539, + 80 + ], + "lines": [ + { + "bbox": [ + 499, + 70, + 539, + 80 + ], + "spans": [ + { + "bbox": [ + 499, + 70, + 539, + 80 + ], + "type": "text", + "content": "MAT [20]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 484, + 128, + 552, + 175 + ], + "lines": [ + { + "bbox": [ + 484, + 128, + 552, + 175 + ], + "spans": [ + { + "bbox": [ + 484, + 128, + 552, + 175 + ], + "type": "image", + "image_path": "244df2dae948974448fb1d1c43c6d3139dfe2db042746499b7e52772a8ecc2fd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 57, + 178, + 125, + 228 + ], + "blocks": [ + { + "bbox": [ + 57, + 178, + 125, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 178, + 125, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 178, + 125, + 228 + ], + "type": "image", + "image_path": "b33196469bb7bad83aaa453dc570c0dde2650c1369163d3b6a2fea6a9d0a8b2e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 127, + 178, + 196, + 228 + ], + "blocks": [ + { + "bbox": [ + 127, + 178, + 196, + 228 + ], + "lines": [ + { + "bbox": [ + 127, + 178, + 196, + 228 + ], + "spans": [ + { + "bbox": [ + 127, + 178, + 196, + 228 + ], + "type": "image", + "image_path": "00e54a27e0f28e5962baad0240bb6d931a7501427112fce45a8c11fa5f504fec.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 198, + 178, + 268, + 228 + ], + "blocks": [ + { + "bbox": [ + 198, + 178, + 268, + 228 + ], + "lines": [ + { + "bbox": [ + 198, + 178, + 268, + 228 + ], + "spans": [ + { + "bbox": [ + 198, + 178, + 268, + 228 + ], + "type": "image", + "image_path": "5ece6eead487f878dcfa11f9c05023c2e73ef0328d58389a369287bc0b3cffe6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 270, + 178, + 339, + 228 + ], + "blocks": [ + { + "bbox": [ + 270, + 178, + 339, + 228 + ], + "lines": [ + { + "bbox": [ + 270, + 178, + 339, + 228 + ], + "spans": [ + { + "bbox": [ + 270, + 178, + 339, + 228 + ], + "type": "image", + "image_path": "607ca004c99a37075c00d810382430cab655d043d218f5768efa768166d7eeb5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 341, + 178, + 410, + 228 + ], + "blocks": [ + { + "bbox": [ + 341, + 178, + 410, + 228 + ], + "lines": [ + { + "bbox": [ + 341, + 178, + 410, + 228 + ], + "spans": [ + { + "bbox": [ + 341, + 178, + 410, + 228 + ], + "type": "image", + "image_path": "d7be25138b0c60bc8e4a40a4ae02ea471488e3282079cc3d4691f384c45ebf08.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 412, + 178, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 412, + 178, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 412, + 178, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 412, + 178, + 481, + 228 + ], + "type": "image", + "image_path": "5fc8ad3c7a0d33380ea55ffd71219a62ae5a2e12fedc216081887171c7854818.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 483, + 178, + 551, + 228 + ], + "blocks": [ + { + "bbox": [ + 483, + 178, + 551, + 228 + ], + "lines": [ + { + "bbox": [ + 483, + 178, + 551, + 228 + ], + "spans": [ + { + "bbox": [ + 483, + 178, + 551, + 228 + ], + "type": "image", + "image_path": "5b4f44ef2aa7bc0092c38a802bd12e84611bdcf7206b301f02d313c7d9b8eb41.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 57, + 231, + 125, + 277 + ], + "blocks": [ + { + "bbox": [ + 57, + 231, + 125, + 277 + ], + "lines": [ + { + "bbox": [ + 57, + 231, + 125, + 277 + ], + "spans": [ + { + "bbox": [ + 57, + 231, + 125, + 277 + ], + "type": "image", + "image_path": "5a07f3fc7b0b55c793961bf30678fc3ccbbb678c7e77e3b0312e714850b9771b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 127, + 231, + 196, + 277 + ], + "blocks": [ + { + "bbox": [ + 127, + 231, + 196, + 277 + ], + "lines": [ + { + "bbox": [ + 127, + 231, + 196, + 277 + ], + "spans": [ + { + "bbox": [ + 127, + 231, + 196, + 277 + ], + "type": "image", + "image_path": "2e9f3a4ba94da554f22161c462b5cbc18b1bd7028711b129b9d99acafa1ad57b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 198, + 231, + 268, + 277 + ], + "blocks": [ + { + "bbox": [ + 198, + 231, + 268, + 277 + ], + "lines": [ + { + "bbox": [ + 198, + 231, + 268, + 277 + ], + "spans": [ + { + "bbox": [ + 198, + 231, + 268, + 277 + ], + "type": "image", + "image_path": "7d872abbaab29b3766b0af1385e885fffa61f03f77128e8078f25f3f1f6d7468.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 270, + 231, + 339, + 277 + ], + "blocks": [ + { + "bbox": [ + 270, + 231, + 339, + 277 + ], + "lines": [ + { + "bbox": [ + 270, + 231, + 339, + 277 + ], + "spans": [ + { + "bbox": [ + 270, + 231, + 339, + 277 + ], + "type": "image", + "image_path": "e07a5a659da4041911111cdb7a69d7b2617912fe0e8e2d0b87942bda0cb23b03.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 341, + 231, + 410, + 277 + ], + "blocks": [ + { + "bbox": [ + 341, + 231, + 410, + 277 + ], + "lines": [ + { + "bbox": [ + 341, + 231, + 410, + 277 + ], + "spans": [ + { + "bbox": [ + 341, + 231, + 410, + 277 + ], + "type": "image", + "image_path": "1139b27b4cd9fc48c7a41387b852ef066939d4654898dbcbd6b6894da8cbb598.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 412, + 231, + 481, + 277 + ], + "blocks": [ + { + "bbox": [ + 412, + 231, + 481, + 277 + ], + "lines": [ + { + "bbox": [ + 412, + 231, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 412, + 231, + 481, + 277 + ], + "type": "image", + "image_path": "5970e85c01d66b196072f3e680f3744a73c3981bd3bbb769b3b2002adad172f7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 483, + 231, + 551, + 277 + ], + "blocks": [ + { + "bbox": [ + 483, + 231, + 551, + 277 + ], + "lines": [ + { + "bbox": [ + 483, + 231, + 551, + 277 + ], + "spans": [ + { + "bbox": [ + 483, + 231, + 551, + 277 + ], + "type": "image", + "image_path": "784b21c6b154c0aefffa48ea7afdf14f5845404a98f9bca9068d706c22760d8b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 57, + 279, + 125, + 331 + ], + "blocks": [ + { + "bbox": [ + 57, + 279, + 125, + 331 + ], + "lines": [ + { + "bbox": [ + 57, + 279, + 125, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 279, + 125, + 331 + ], + "type": "image", + "image_path": "7c32db83d5dbb778ddf3c48f744dac297bc50c0f5c41e609c518f81cfc284802.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 127, + 279, + 196, + 331 + ], + "blocks": [ + { + "bbox": [ + 127, + 279, + 196, + 331 + ], + "lines": [ + { + "bbox": [ + 127, + 279, + 196, + 331 + ], + "spans": [ + { + "bbox": [ + 127, + 279, + 196, + 331 + ], + "type": "image", + "image_path": "f23fc56921060d77673eaf644301da08ec766844001208aff3b710a68da3ab0d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 198, + 279, + 268, + 331 + ], + "blocks": [ + { + "bbox": [ + 198, + 279, + 268, + 331 + ], + "lines": [ + { + "bbox": [ + 198, + 279, + 268, + 331 + ], + "spans": [ + { + "bbox": [ + 198, + 279, + 268, + 331 + ], + "type": "image", + "image_path": "2c570319cc88043640bf6696e37a17a9b166dc5b4386530f7a80430b9c05e5e9.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 270, + 279, + 339, + 331 + ], + "blocks": [ + { + "bbox": [ + 270, + 279, + 339, + 331 + ], + "lines": [ + { + "bbox": [ + 270, + 279, + 339, + 331 + ], + "spans": [ + { + "bbox": [ + 270, + 279, + 339, + 331 + ], + "type": "image", + "image_path": "6813dd7a55f627650a5217ec289190ddbbb03f0982b28e3c90512f3fbd37c5d0.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 341, + 279, + 410, + 331 + ], + "blocks": [ + { + "bbox": [ + 341, + 279, + 410, + 331 + ], + "lines": [ + { + "bbox": [ + 341, + 279, + 410, + 331 + ], + "spans": [ + { + "bbox": [ + 341, + 279, + 410, + 331 + ], + "type": "image", + "image_path": "11f228a7ffd52f14fe2e395e6a7f3f6ae965ee4215f68577e75f4cde79754bc5.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 412, + 279, + 481, + 331 + ], + "blocks": [ + { + "bbox": [ + 412, + 279, + 481, + 331 + ], + "lines": [ + { + "bbox": [ + 412, + 279, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 412, + 279, + 481, + 331 + ], + "type": "image", + "image_path": "54c3d9bb194c7a80cc7b395e45430c1854515df480bc7c9846e942b9c12297bf.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 483, + 279, + 551, + 331 + ], + "blocks": [ + { + "bbox": [ + 483, + 279, + 551, + 331 + ], + "lines": [ + { + "bbox": [ + 483, + 279, + 551, + 331 + ], + "spans": [ + { + "bbox": [ + 483, + 279, + 551, + 331 + ], + "type": "image", + "image_path": "e7cb8337df6863f3882f3820179085a6abedf9b61600797006b9391aea50d087.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 57, + 333, + 125, + 381 + ], + "blocks": [ + { + "bbox": [ + 57, + 333, + 125, + 381 + ], + "lines": [ + { + "bbox": [ + 57, + 333, + 125, + 381 + ], + "spans": [ + { + "bbox": [ + 57, + 333, + 125, + 381 + ], + "type": "image", + "image_path": "80bc1a3d5e1c8b7d61b2abf6f85f73627b33b938deecb0f3754a330af92abcb2.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 127, + 333, + 196, + 381 + ], + "blocks": [ + { + "bbox": [ + 127, + 333, + 196, + 381 + ], + "lines": [ + { + "bbox": [ + 127, + 333, + 196, + 381 + ], + "spans": [ + { + "bbox": [ + 127, + 333, + 196, + 381 + ], + "type": "image", + "image_path": "8531769c74a7afa17953508a8604a2b45b160a6cc3361be849ee6fae81169aa8.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 198, + 333, + 268, + 381 + ], + "blocks": [ + { + "bbox": [ + 198, + 333, + 268, + 381 + ], + "lines": [ + { + "bbox": [ + 198, + 333, + 268, + 381 + ], + "spans": [ + { + "bbox": [ + 198, + 333, + 268, + 381 + ], + "type": "image", + "image_path": "e2b3c4aef2d166c200ae4d89db70f18c8924ab52d0ee34c603556e18404a581b.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 270, + 333, + 339, + 381 + ], + "blocks": [ + { + "bbox": [ + 270, + 333, + 339, + 381 + ], + "lines": [ + { + "bbox": [ + 270, + 333, + 339, + 381 + ], + "spans": [ + { + "bbox": [ + 270, + 333, + 339, + 381 + ], + "type": "image", + "image_path": "b27b8735aa964fe2463500be9b15c7d25a0d2acb72f10438062ec2e918565ac9.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 341, + 333, + 410, + 381 + ], + "blocks": [ + { + "bbox": [ + 341, + 333, + 410, + 381 + ], + "lines": [ + { + "bbox": [ + 341, + 333, + 410, + 381 + ], + "spans": [ + { + "bbox": [ + 341, + 333, + 410, + 381 + ], + "type": "image", + "image_path": "17b65febfd662f7db2a494f3eb870ae247f51a8406c9eddf63b0d72a53c35500.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 412, + 333, + 481, + 381 + ], + "blocks": [ + { + "bbox": [ + 412, + 333, + 481, + 381 + ], + "lines": [ + { + "bbox": [ + 412, + 333, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 412, + 333, + 481, + 381 + ], + "type": "image", + "image_path": "903588f52c5235edf9c0f8105e4a919c6492f9d222f2b3a4d3590f70693e6cad.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 483, + 333, + 551, + 381 + ], + "blocks": [ + { + "bbox": [ + 483, + 333, + 551, + 381 + ], + "lines": [ + { + "bbox": [ + 483, + 333, + 551, + 381 + ], + "spans": [ + { + "bbox": [ + 483, + 333, + 551, + 381 + ], + "type": "image", + "image_path": "ab1cb0c4abb2c9c265aedc9bba2b2fdcb907c4051bb6a0d9c2d1b44fd23f4cf6.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 57, + 384, + 125, + 436 + ], + "blocks": [ + { + "bbox": [ + 57, + 384, + 125, + 436 + ], + "lines": [ + { + "bbox": [ + 57, + 384, + 125, + 436 + ], + "spans": [ + { + "bbox": [ + 57, + 384, + 125, + 436 + ], + "type": "image", + "image_path": "940862c7ca1f5d2b7653d9b11b0ed1519ea638e3031a58ad5b95a811ef56213f.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 127, + 384, + 196, + 436 + ], + "blocks": [ + { + "bbox": [ + 127, + 384, + 196, + 436 + ], + "lines": [ + { + "bbox": [ + 127, + 384, + 196, + 436 + ], + "spans": [ + { + "bbox": [ + 127, + 384, + 196, + 436 + ], + "type": "image", + "image_path": "697b3b982cb243fbbc9d8479f28ae6bbac7b03140a38247a23c39900c717710b.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 198, + 384, + 268, + 436 + ], + "blocks": [ + { + "bbox": [ + 198, + 384, + 268, + 436 + ], + "lines": [ + { + "bbox": [ + 198, + 384, + 268, + 436 + ], + "spans": [ + { + "bbox": [ + 198, + 384, + 268, + 436 + ], + "type": "image", + "image_path": "651e9193db581e5fea2be522fbc57d104dc5cecc17c827954d6430071abc259a.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 270, + 384, + 339, + 436 + ], + "blocks": [ + { + "bbox": [ + 270, + 384, + 339, + 436 + ], + "lines": [ + { + "bbox": [ + 270, + 384, + 339, + 436 + ], + "spans": [ + { + "bbox": [ + 270, + 384, + 339, + 436 + ], + "type": "image", + "image_path": "a0de4e5c79f125e6003575cdc6d41456401c7975659cb85e2fd81e64f937c4a4.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 341, + 384, + 410, + 436 + ], + "blocks": [ + { + "bbox": [ + 341, + 384, + 410, + 436 + ], + "lines": [ + { + "bbox": [ + 341, + 384, + 410, + 436 + ], + "spans": [ + { + "bbox": [ + 341, + 384, + 410, + 436 + ], + "type": "image", + "image_path": "75e13bdadcce2aa28797a8bfa4ad573282a3d45eb459d16f0ddfd8752607e8bd.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 412, + 384, + 481, + 436 + ], + "blocks": [ + { + "bbox": [ + 412, + 384, + 481, + 436 + ], + "lines": [ + { + "bbox": [ + 412, + 384, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 412, + 384, + 481, + 436 + ], + "type": "image", + "image_path": "1fa7abd658bedc6cb099a77400efb4013dafcfc57218199d07ad356560e0a3e7.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 483, + 384, + 551, + 436 + ], + "blocks": [ + { + "bbox": [ + 483, + 384, + 551, + 436 + ], + "lines": [ + { + "bbox": [ + 483, + 384, + 551, + 436 + ], + "spans": [ + { + "bbox": [ + 483, + 384, + 551, + 436 + ], + "type": "image", + "image_path": "ea0cc1002a0555877001172bb5c4aacddb94836c15be7049fc4e86952dfde631.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 57, + 437, + 125, + 483 + ], + "blocks": [ + { + "bbox": [ + 57, + 437, + 125, + 483 + ], + "lines": [ + { + "bbox": [ + 57, + 437, + 125, + 483 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 125, + 483 + ], + "type": "image", + "image_path": "b1032b9177d616fda0b3d78baa3d609873d5996cdf06f9dee7f788a291f19820.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 490, + 555, + 512 + ], + "lines": [ + { + "bbox": [ + 55, + 490, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 555, + 512 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with state-of-the-art methods. As shown, our method effectively erases the target objects while avoiding generating unwanted sundries like the compared methods." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 127, + 437, + 196, + 483 + ], + "blocks": [ + { + "bbox": [ + 127, + 437, + 196, + 483 + ], + "lines": [ + { + "bbox": [ + 127, + 437, + 196, + 483 + ], + "spans": [ + { + "bbox": [ + 127, + 437, + 196, + 483 + ], + "type": "image", + "image_path": "dcf0071bf0608a84ccaaad113640db89a9cc30100366270a512782b104450717.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 198, + 437, + 268, + 483 + ], + "blocks": [ + { + "bbox": [ + 198, + 437, + 268, + 483 + ], + "lines": [ + { + "bbox": [ + 198, + 437, + 268, + 483 + ], + "spans": [ + { + "bbox": [ + 198, + 437, + 268, + 483 + ], + "type": "image", + "image_path": "8032237434a1cbe42d18c4d38c856e5e8935d2febc1bd3ccec1b51f4d320239f.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 270, + 437, + 339, + 483 + ], + "blocks": [ + { + "bbox": [ + 270, + 437, + 339, + 483 + ], + "lines": [ + { + "bbox": [ + 270, + 437, + 339, + 483 + ], + "spans": [ + { + "bbox": [ + 270, + 437, + 339, + 483 + ], + "type": "image", + "image_path": "558d1ac2aacdebcb047abdbb82d6b23df2ed4c0bba9f1ae760823b203f8eca20.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 341, + 437, + 410, + 483 + ], + "blocks": [ + { + "bbox": [ + 341, + 437, + 410, + 483 + ], + "lines": [ + { + "bbox": [ + 341, + 437, + 410, + 483 + ], + "spans": [ + { + "bbox": [ + 341, + 437, + 410, + 483 + ], + "type": "image", + "image_path": "8c9b8c512f9c2987c42080d6e8bcacbcf6e9807dc43a08702d5aa0568a88c92e.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 412, + 437, + 481, + 483 + ], + "blocks": [ + { + "bbox": [ + 412, + 437, + 481, + 483 + ], + "lines": [ + { + "bbox": [ + 412, + 437, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 412, + 437, + 481, + 483 + ], + "type": "image", + "image_path": "79de545d0cabd71e8e420a81ef6e0aa61bb8212ceca6b920b2259b1d0c2e379c.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 483, + 437, + 551, + 483 + ], + "blocks": [ + { + "bbox": [ + 483, + 437, + 551, + 483 + ], + "lines": [ + { + "bbox": [ + 483, + 437, + 551, + 483 + ], + "spans": [ + { + "bbox": [ + 483, + 437, + 551, + 483 + ], + "type": "image", + "image_path": "ef807835733c81372c38eb27bc9bf78332fa6933b3bf83a668254b950e717947.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": " is the entity segmentation mask and " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "|A|" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": " refers to the area of " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": " is the inpainting mask. If the IoS is greater than the threshold " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": ", these entities are considered as newly generated sundries (SDR). " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": " is set to 0.95 by default. We traverse all entities and calculate the number of all newly generated sundries (referred to as " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "|SDR|" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": "). By setting " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": " as the number of test samples, the Mean Sundries Num (MSN) is defined by:" + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 139, + 628, + 295, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 628, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 139, + 628, + 295, + 651 + ], + "type": "interline_equation", + "content": "\\mathrm {M S N} = \\frac {| S D R |}{N}. \\tag {8}", + "image_path": "fe310f3e62cbabc4a109da60ac9a122be9a236ddfede69ec7873e0a7281474c9.jpg" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 55, + 654, + 295, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 678 + ], + "type": "text", + "content": "The Mean Area Ratio of Sundries (MARS) to the inpainting area can be computed by:" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 106, + 685, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 685, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\mathrm {M A R S} = \\sum_ {A \\in S D R} \\frac {| A \\cap B |}{| B |} \\times \\frac {1}{N}, \\tag {9}", + "image_path": "e9272d37010ecb2f704592ff21de69773aa8a85bdf3aca810ca0ba559505eafe.jpg" + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": "MSN reflects the quantity of sundries generated by the model, while MARS indicates the ratio of the sundries area relative to the inpainting mask area. A higher MARS value signifies that the generated sundries are larger and more prominent. Both of these metrics should be minimized for optimal performance. In Fig. 5, we present a comparison between our metric and CLIP accuracy [44]. It is evident that CLIP accuracy only focuses on the magnitude of the change between the generated result and the original image. We also compare our method with ReMOVE [1] in Fig. 6. ReMOVE focuses solely on whether the generated image is consistent with the surrounding context. It fails to provide a reasonable score when the newly generated object (blue chair) is consistent with the surrounding context. In contrast, our metric is more reasonable and reliable." + } + ] + } + ], + "index": 65 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28279" + } + ] + } + ], + "index": 66 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 78, + 103, + 251 + ], + "blocks": [ + { + "bbox": [ + 57, + 78, + 103, + 251 + ], + "lines": [ + { + "bbox": [ + 57, + 78, + 103, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 78, + 103, + 251 + ], + "type": "image", + "image_path": "51b05103348b269308f9b3dea1d27894befa6774c448c5948283df0bdec8bf8d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 254, + 86, + 264 + ], + "lines": [ + { + "bbox": [ + 74, + 254, + 86, + 264 + ], + "spans": [ + { + "bbox": [ + 74, + 254, + 86, + 264 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 104, + 78, + 150, + 251 + ], + "blocks": [ + { + "bbox": [ + 104, + 78, + 150, + 251 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 150, + 251 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 150, + 251 + ], + "type": "image", + "image_path": "44eb1bb973a22f2011546a8eeaf001df50774afb85df8d63fc5f8b0a67999cb5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 254, + 134, + 264 + ], + "lines": [ + { + "bbox": [ + 121, + 254, + 134, + 264 + ], + "spans": [ + { + "bbox": [ + 121, + 254, + 134, + 264 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 78, + 198, + 251 + ], + "blocks": [ + { + "bbox": [ + 151, + 78, + 198, + 251 + ], + "lines": [ + { + "bbox": [ + 151, + 78, + 198, + 251 + ], + "spans": [ + { + "bbox": [ + 151, + 78, + 198, + 251 + ], + "type": "image", + "image_path": "83772e6c4613e244c45a53dcf65de447bab0ae71bcf9029fefa2a9c1a4dcfdf8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 254, + 181, + 264 + ], + "lines": [ + { + "bbox": [ + 168, + 254, + 181, + 264 + ], + "spans": [ + { + "bbox": [ + 168, + 254, + 181, + 264 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 199, + 78, + 245, + 251 + ], + "blocks": [ + { + "bbox": [ + 199, + 78, + 245, + 251 + ], + "lines": [ + { + "bbox": [ + 199, + 78, + 245, + 251 + ], + "spans": [ + { + "bbox": [ + 199, + 78, + 245, + 251 + ], + "type": "image", + "image_path": "50e74cbddf0052dab63aae9c97f1e4a5f09da4fb657778c0b09c2cd9cee5cc9a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 254, + 228, + 264 + ], + "lines": [ + { + "bbox": [ + 216, + 254, + 228, + 264 + ], + "spans": [ + { + "bbox": [ + 216, + 254, + 228, + 264 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 246, + 78, + 292, + 251 + ], + "blocks": [ + { + "bbox": [ + 246, + 78, + 292, + 251 + ], + "lines": [ + { + "bbox": [ + 246, + 78, + 292, + 251 + ], + "spans": [ + { + "bbox": [ + 246, + 78, + 292, + 251 + ], + "type": "image", + "image_path": "21ddf1320cc8ba197025e24cfa98b0bbadf6012e8eb2c7866dc4f8f4479146fe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 254, + 275, + 264 + ], + "lines": [ + { + "bbox": [ + 263, + 254, + 275, + 264 + ], + "spans": [ + { + "bbox": [ + 263, + 254, + 275, + 264 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 350, + 132, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 270, + 295, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 336 + ], + "type": "text", + "content": "Figure 8. Effect of entity attention. From left to right: (a) input, (b) inpainting mask, (c) amodal entity segmentation, (d) entity attention, (e) cross attention. As can be seen, with our entity attention, our method faithfully repairs the missing parts of entities according to the amodal entity segmentation, effectively reducing the generation of sundries." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 350, + 132, + 427 + ], + "lines": [ + { + "bbox": [ + 57, + 350, + 132, + 427 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 132, + 427 + ], + "type": "image", + "image_path": "02cf79cfe422caae5aaab0d27257ebc9b985c01c8200273658494c8b102b7d37.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 430, + 101, + 440 + ], + "lines": [ + { + "bbox": [ + 89, + 430, + 101, + 440 + ], + "spans": [ + { + "bbox": [ + 89, + 430, + 101, + 440 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 447, + 295, + 513 + ], + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 513 + ], + "type": "text", + "content": "Figure 9. Impact of applying entity attention on non-inpainting area: (a) input + inpainting mask, (b) not applying entity attention on non-inpainting area, (c) applying entity attention on non-inpainting area. As shown, applying entity attention on non-inpainting areas may lead to unwanted changes to these areas compared to the original image (see the red box)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 350, + 211, + 426 + ], + "blocks": [ + { + "bbox": [ + 135, + 350, + 211, + 426 + ], + "lines": [ + { + "bbox": [ + 135, + 350, + 211, + 426 + ], + "spans": [ + { + "bbox": [ + 135, + 350, + 211, + 426 + ], + "type": "image", + "image_path": "41d1d2616f9ec7f9d265ccb12686877434c828aad90b6760c600779d15f2d6cd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 430, + 179, + 441 + ], + "lines": [ + { + "bbox": [ + 167, + 430, + 179, + 441 + ], + "spans": [ + { + "bbox": [ + 167, + 430, + 179, + 441 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 214, + 350, + 290, + 426 + ], + "blocks": [ + { + "bbox": [ + 214, + 350, + 290, + 426 + ], + "lines": [ + { + "bbox": [ + 214, + 350, + 290, + 426 + ], + "spans": [ + { + "bbox": [ + 214, + 350, + 290, + 426 + ], + "type": "image", + "image_path": "68f1136591874d8648bc205429838f09ee8fb4aac2f7da07f1a06c6c802e88ac.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 430, + 258, + 441 + ], + "lines": [ + { + "bbox": [ + 246, + 430, + 258, + 441 + ], + "spans": [ + { + "bbox": [ + 246, + 430, + 258, + 441 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Since we evaluate our method on non-synthetic real-world scenarios, where paired data is unavailable, we use non-paired metrics for evaluation. Besides MSN and MARS, we also report ReMOVE score [1] and Frechet Inception Distance (FID) [11]. We compare our method with other state-of-the-art inpainting approaches on entity segmentation test dataset [30] and coco val2017 dataset [21]. Excluding stuff objects, the segmentation mask for thing object is selected from each image as the inpainting mask, then we dilate the mask. Meanwhile, we find that CLIP-Away [8] and PowerPaint [52] would generate additional sundries from floating handheld objects due to precise inpainting mask. Therefore, we also conduct comparisons using looser inpainting masks. We obtain looser inpainting masks by dilating masks with larger sizes and filling the" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "holes. A total of around 700 test samples are generated from the entity segmentation dataset, and 4000 test samples are generated from the COCO dataset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 115, + 539, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 539, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 539, + 128 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 133, + 554, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 192 + ], + "type": "text", + "content": "Baselines. We compare our method with various state-of-the-art inpainting methods, including MAT [20], SD-Inpainting [32], CLIPAway [8], and PowerPaint [52], where MAT is a GAN-based method while the other three are stable diffusion based methods." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": "Quantitative comparison. As shown in Table 1, our method demonstrates significant advantages over CLIP-Away [8] and PowerPaint [52] in both the MSN and MARS metrics on entity segmentation and coco dataset, using either precise or loose inpainting masks. And among all the diffusion-based methods, our approach also achieves the highest score on the ReMOVE metric. This demonstrates that the proposed model significantly reduces the probability of sundries generation and the area of generated sundries compared with other methods. We also observe that the GAN-based method MAT has relatively low MSN, MARS score and high ReMOVE score compared with SD-based baselines. This may be due to the powerful generative capability of the SD model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 369, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 478 + ], + "type": "text", + "content": "Qualitative comparison. In Fig. 7, we show some prediction results of EntityErasure alongside other state-of-the-art methods. It is evident that MAT generates fewer sundries but their generated textures lack realism. Meanwhile, CLIPAway and PowerPaint produce higher quality results, but still produce sundries in some cases. In contrast, our proposed EntityErasure demonstrates a lower probability of sundries generation and produces higher-quality images. Please see the supplementary material for more results." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 317, + 488, + 554, + 521 + ], + "blocks": [ + { + "bbox": [ + 317, + 488, + 554, + 521 + ], + "lines": [ + { + "bbox": [ + 317, + 488, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 554, + 521 + ], + "type": "table", + "html": "
MethodMATCLIPAwaySDInpaintPowerPaintOurs
Time(s)↓0.084.042.215.002.98
", + "image_path": "42a2b16e53836d60fba77180792a20ee2da8d033212cacad75592e15cdff6e61.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 358, + 529, + 510, + 540 + ], + "lines": [ + { + "bbox": [ + 358, + 529, + 510, + 540 + ], + "spans": [ + { + "bbox": [ + 358, + 529, + 510, + 540 + ], + "type": "text", + "content": "Table 2. Comparison on inference time." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 557, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 653 + ], + "type": "text", + "content": "Comparison on inference time. In Table 2, we conduct comparison of inference speed on a single RTX 4090. The number of our sampling steps is set as 20. As shown, the inference speed of our method is comparable to those of other diffusion-based methods. Note, we experimentally find that larger number of sampling steps (e.g., 50 steps) does not lead to clear performance improvement, while the inference time cost will be increased." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 660, + 408, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 408, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 408, + 673 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "To validate the effectiveness of the modules proposed in EntityErasure, we conduct ablation experiments on entity segmentation dataset [30]." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28280" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 288, + 119 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 288, + 119 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 288, + 119 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 288, + 119 + ], + "type": "table", + "html": "
MethodFID↓MSN↓MARS↓ReMOVE↑
cross attn57.10.2120.07785.82
entity attn58.00.1220.03786.24
", + "image_path": "f13073755b2d7b267b57a6ef366f82712e95dac502c5870fcf04207746e8bc18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 302 + ], + "type": "text", + "content": "Effect of amodal entity segmentation guidance. We use amodal entity segmentation guidance in entity attention to inject content into the corresponding entity. To evaluate its impact, we replace entity attention with standard cross attention for comparison. As shown in Table 3, FID shows a slight decrease, while MSN and MARS increase notably, indicating a rise in sundries generation. This highlights the effectiveness of our amodal entity segmentation guidance in reducing sundries generation. It is further supported by the qualitative ablation results presented in Fig. 8, where model without amodal entity segmentation guidance tends to generate sundries easily." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 415 + ], + "type": "text", + "content": "Impact of applying entity attention on non-inpainting area. We observe that applying entity attention for injecting information into the non-inpainting area does not incur sundries, but it may cause alterations to the non-inpainting region. As shown in Fig. 9, if entity attention is applied to the non-inpainting area, changes can be observed. This implies that, the non-inpainting area only requires feature injection via element-wise addition, without the need for entity attention." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 70, + 424, + 282, + 498 + ], + "blocks": [ + { + "bbox": [ + 110, + 127, + 240, + 138 + ], + "lines": [ + { + "bbox": [ + 110, + 127, + 240, + 138 + ], + "spans": [ + { + "bbox": [ + 110, + 127, + 240, + 138 + ], + "type": "text", + "content": "Table 3. Effect of entity attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 424, + 282, + 498 + ], + "lines": [ + { + "bbox": [ + 70, + 424, + 282, + 498 + ], + "spans": [ + { + "bbox": [ + 70, + 424, + 282, + 498 + ], + "type": "table", + "html": "
TIFID↓MSN↓MARS↓ReMOVE↑
1159.80.4490.078384.86
1358.00.1220.036986.24
3159.90.4250.075684.88
3358.00.1220.036686.24
", + "image_path": "137ae5ca71384cfd7f3e6e7b9c796f0a784e714666b60051cda423f810400ffa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "type": "text", + "content": "Table 4. Effect of classifier-free guidance. T: Text-CFG weight, I: Image-CFG weight." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Effect of classifier-free guidance. We apply classifier-free guidance to amplify the influence of condition image and amodal entity segmentation guidance. To assess its effectiveness, we perform experiments with varying classifier-free guidance weights. As shown in Table 4, when the weight of Image-CFG is set to 1, effectively turning off Image-CFG, both MSN and MARS increase significantly, which signals an increase in sundries generation. Compared to the impact of Text-CFG, the influence of Image-CFG on the results is more significant. It is further supported by the qualitative ablation results presented in Fig. 4. In the first butterfly image, without CFG, the model would generate excessive textures. This underscores the role of Image-CFG in reducing the occurrence of sundries." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 80, + 375, + 114 + ], + "blocks": [ + { + "bbox": [ + 316, + 80, + 375, + 114 + ], + "lines": [ + { + "bbox": [ + 316, + 80, + 375, + 114 + ], + "spans": [ + { + "bbox": [ + 316, + 80, + 375, + 114 + ], + "type": "image", + "image_path": "89b2e00ee18584c6f805ad1e551cf7f88545473b8f0b28895637be1d52f228d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 376, + 80, + 434, + 114 + ], + "blocks": [ + { + "bbox": [ + 376, + 80, + 434, + 114 + ], + "lines": [ + { + "bbox": [ + 376, + 80, + 434, + 114 + ], + "spans": [ + { + "bbox": [ + 376, + 80, + 434, + 114 + ], + "type": "image", + "image_path": "607a1a117cd0a8da134e1f3e601f548928cbaba0c7660dacc27475ffed1089b4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 435, + 80, + 493, + 114 + ], + "blocks": [ + { + "bbox": [ + 435, + 80, + 493, + 114 + ], + "lines": [ + { + "bbox": [ + 435, + 80, + 493, + 114 + ], + "spans": [ + { + "bbox": [ + 435, + 80, + 493, + 114 + ], + "type": "image", + "image_path": "fa05338e5d91422fad3a92143773165207beb87fc23cc63f77f62307bf8eb9e3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 495, + 80, + 553, + 114 + ], + "blocks": [ + { + "bbox": [ + 495, + 80, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 495, + 80, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 495, + 80, + 553, + 114 + ], + "type": "image", + "image_path": "3d2bdc1818bb1946e85eaadb2b215460d9fdce54db13d9c409155d12d7fdfa0a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 115, + 373, + 203 + ], + "blocks": [ + { + "bbox": [ + 316, + 115, + 373, + 203 + ], + "lines": [ + { + "bbox": [ + 316, + 115, + 373, + 203 + ], + "spans": [ + { + "bbox": [ + 316, + 115, + 373, + 203 + ], + "type": "image", + "image_path": "e62895cb0ae35471ffd1a96be84b5018b3f9b2bf8d7972c8e7ea54ffd8d2a8fe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 115, + 434, + 203 + ], + "blocks": [ + { + "bbox": [ + 376, + 115, + 434, + 203 + ], + "lines": [ + { + "bbox": [ + 376, + 115, + 434, + 203 + ], + "spans": [ + { + "bbox": [ + 376, + 115, + 434, + 203 + ], + "type": "image", + "image_path": "9e8a86f809f1258c07ec97084883695bef4ab0045833947e2ff487daa61ff6a9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 115, + 493, + 203 + ], + "blocks": [ + { + "bbox": [ + 435, + 115, + 493, + 203 + ], + "lines": [ + { + "bbox": [ + 435, + 115, + 493, + 203 + ], + "spans": [ + { + "bbox": [ + 435, + 115, + 493, + 203 + ], + "type": "image", + "image_path": "2a55a6f9f04cd77eae6dc26fb55e97321c6fae6d63d8fff890647aa5d93fc938.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 495, + 115, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 495, + 115, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 495, + 115, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 495, + 115, + 553, + 203 + ], + "type": "image", + "image_path": "32b15670a7dcdb48723cd7e244ca814eb50c3a61e85d40e7d2ef07e28cf40e17.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 205, + 374, + 243 + ], + "blocks": [ + { + "bbox": [ + 316, + 205, + 374, + 243 + ], + "lines": [ + { + "bbox": [ + 316, + 205, + 374, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 374, + 243 + ], + "type": "image", + "image_path": "34a6b55b8831a96591dc5a3fc4ed00f8ed84dbe3c6126d2ee883d12335c5c967.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 247, + 351, + 258 + ], + "lines": [ + { + "bbox": [ + 339, + 247, + 351, + 258 + ], + "spans": [ + { + "bbox": [ + 339, + 247, + 351, + 258 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 376, + 205, + 434, + 243 + ], + "blocks": [ + { + "bbox": [ + 376, + 205, + 434, + 243 + ], + "lines": [ + { + "bbox": [ + 376, + 205, + 434, + 243 + ], + "spans": [ + { + "bbox": [ + 376, + 205, + 434, + 243 + ], + "type": "image", + "image_path": "85c419f09dc6c7c4598325eb7ea9e247caf3030c950deae04af63a0cc70ca5b9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 247, + 411, + 258 + ], + "lines": [ + { + "bbox": [ + 399, + 247, + 411, + 258 + ], + "spans": [ + { + "bbox": [ + 399, + 247, + 411, + 258 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 435, + 205, + 493, + 243 + ], + "blocks": [ + { + "bbox": [ + 435, + 205, + 493, + 243 + ], + "lines": [ + { + "bbox": [ + 435, + 205, + 493, + 243 + ], + "spans": [ + { + "bbox": [ + 435, + 205, + 493, + 243 + ], + "type": "image", + "image_path": "3697e753b11bc3d3f69a6e4866f199c2da02ddb1cb30ccd35389e6a402c156c1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 247, + 471, + 258 + ], + "lines": [ + { + "bbox": [ + 459, + 247, + 471, + 258 + ], + "spans": [ + { + "bbox": [ + 459, + 247, + 471, + 258 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 495, + 205, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 495, + 205, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 495, + 205, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 495, + 205, + 553, + 243 + ], + "type": "image", + "image_path": "3bba271f19a65a45172d4499f88ce862ebe243d4c38858c6c64de9b982b6b352.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 247, + 531, + 257 + ], + "lines": [ + { + "bbox": [ + 518, + 247, + 531, + 257 + ], + "spans": [ + { + "bbox": [ + 518, + 247, + 531, + 257 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 270, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 326 + ], + "type": "text", + "content": "Figure 10. Effect of classifier-free guidance. From left to right: (a) input, (b) inpainting mask, (c) Imagecfg=3, (d) Imagecfg=1. As shown, utilizing classifier-free guidance to strengthen the guidance from the condition image benefits obtaining cleaner images without sundries." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 356, + 388, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 388, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 388, + 369 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": "In this paper, we propose a novel diffusion-based inpainting method named EntityErasure. We approach the erasure task from a new perspective, framing it as a partial amodal entity completion task. Our method decomposes the erasure process into two steps: partial amodal entity segmentation and partial amodal entity completion, making the erasure more controllable. The proposed entity attention injects entity information with amodal entity segmentation guidance. We ensure that all generated content is completed from the entities in the non-inpainting area, effectively preventing the creation of new sundries. EntityErasure demonstrates state-of-the-art performance in terms of object erasure." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 536, + 556, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 556, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 556, + 656 + ], + "type": "text", + "content": "Limitations and future work. Although our method can produce high-quality object erasure results without unwanted sundries, it also has several limitations. First of all, the success of our method may be affected by the quality of amodal segmentation, though our method has some tolerance to poor amodal segmentation (see the supplementary material for validation). Besides, it currently relies on manual mask extension to remove shadows associated with the erased objects. In the future, we will try to address this limitation by incorporating instance shadow detection." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002)." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28281" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In CVPRW, 2024. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "type": "text", + "content": "[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 181, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 181, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 181, + 295, + 203 + ], + "type": "text", + "content": "[3] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "type": "text", + "content": "[4] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 250, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 294, + 271 + ], + "type": "text", + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 273, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 295 + ], + "type": "text", + "content": "[6] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 297, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 294, + 318 + ], + "type": "text", + "content": "[7] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 363 + ], + "type": "text", + "content": "[8] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models. In NeurIPS, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 365, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 407 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 410, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 410, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 57, + 410, + 295, + 442 + ], + "type": "text", + "content": "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 488, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 294, + 511 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 555 + ], + "type": "text", + "content": "[13] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 557, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 589 + ], + "type": "text", + "content": "[14] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "text", + "content": "[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Fan Li, Zixiao Zhang, Yi Huang, Jianzhuang Liu, Renjing Pei, Bin Shao, and Songcen Xu. Magiceraser: Erasing any objects via semantics-aware control. In ECCV, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 153, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 175 + ], + "type": "text", + "content": "[19] Ke Li and Jitendra Malik. Amodal instance segmentation. In ECCV, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 177, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 553, + 209 + ], + "type": "text", + "content": "[20] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In CVPR, 2022. 1, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 212, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 212, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 553, + 255 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "type": "text", + "content": "[23] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 326, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 347 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[25] Khoi Nguyen and Sinisa Todorovic. A weakly supervised amodal segmenter with boundary uncertainty estimation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "type": "text", + "content": "[26] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In CVPR, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 430, + 553, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 462 + ], + "type": "text", + "content": "[27] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "type": "text", + "content": "[28] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "text", + "content": "[29] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Philip Torr, Zhe Lin, and Jiaya Jia. Open world entity segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8743-8756, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "text", + "content": "[30] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High-quality entity segmentation. In ICCV, 2023. 3, 4, 5, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 578, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 643 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28282" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 128 + ], + "type": "text", + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy1, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 129, + 239, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 129, + 239, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 129, + 239, + 139 + ], + "type": "text", + "content": "[34] SG161222. Realistic vision v6.0 b1, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 295, + 174 + ], + "type": "text", + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 196 + ], + "type": "text", + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "text", + "content": "[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "text", + "content": "[39] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "text", + "content": "[40] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. A modal completion via progressive mixed context diffusion. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "type": "text", + "content": "[41] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanate: Temporally consistent human image animation using diffusion model. In CVPR, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 441 + ], + "type": "text", + "content": "[42] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In CVPR, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "text", + "content": "[44] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 566 + ], + "type": "text", + "content": "[45] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "text", + "content": "[46] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In CVPR, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "text", + "content": "[47] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "text", + "content": "[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 174 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[51] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In ECCV, 2024. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28283" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_content_list.json b/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..250d1ea04457c9d69a767a2059ff05aaedfaa72d --- /dev/null +++ b/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_content_list.json @@ -0,0 +1,1521 @@ +[ + { + "type": "text", + "text": "EntitySAM: Segment Everything in Video", + "text_level": 1, + "bbox": [ + 285, + 130, + 712, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingqiao Ye $^{1,*}$ Seoung Wug Oh $^{2}$ Lei Ke $^{3}$ Joon-Young Lee $^{2}$ $^{1}$ EPFL 2Adobe Research 3Carnegie Mellon University", + "bbox": [ + 217, + 179, + 759, + 218 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72cda20c7485f39168bcc495c67ad7fc72aa5b31e7be012a612f4209f8f75d12.jpg", + "image_caption": [ + "Figure 1. Zero-shot video entity segmentation performance comparison on VIPEntySeg dataset using models trained on COCO, showing: 1) SAM 2 [39] using Mask2Former [7] mask prompts for the initial frame, 2) Mask2Former with DEVA [11] association, and 3) our proposed EntitySAM. Our EntitySAM enhances SAM 2 by automatically segmenting and tracking novel entities without requiring user-specified prompts, achieving superior performance compared to existing state-of-the-art zero-shot tracking methods." + ], + "image_footnote": [], + "bbox": [ + 96, + 227, + 906, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 525, + 326, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automatically tracking and segmenting every video entity remains a significant challenge. Despite rapid advancements in video segmentation, even state-of-the-art models like SAM 2 struggle to consistently track all entities across a video—a task we refer to as Video Entity Segmentation. We propose EntitySAM, a framework for zero-shot video entity segmentation. EntitySAM extends SAM 2 by removing the need for explicit prompts, allowing automatic discovery and tracking of all entities, including those appearing in later frames. We incorporate query-based entity discovery and association into SAM 2, inspired by transformer-based object detectors. Specifically, we introduce an entity decoder to facilitate inter-object communication and an automatic prompt generator using learnable object queries. Additionally, we add a semantic encoder to enhance SAM 2's semantic awareness, improving segmentation quality. Trained on image-level mask annotations without category information from the COCO dataset, EntitySAM demonstrates strong generalization on four zero-shot video segmentation tasks: Video Entity, Panoptic, Instance, and Semantic Segmentation. Results on six popular benchmarks show that", + "bbox": [ + 86, + 556, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "EntitySAM outperforms previous unified video segmentation methods and strong baselines, setting new standards for zero-shot video segmentation. Our code and models are at github.com/ymq2017-entitysam.", + "bbox": [ + 511, + 526, + 906, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 617, + 643, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video segmentation is an essential task for various applications in video scene understanding, including video editing, AR/VR, and robotic perception. There are roughly two lines of video segmentation approaches. One line of approaches focuses on selecting semantic thing/stuff from predefined sets [27, 28, 48]. While these methods show strong performance in the benchmark, the predictions are limited to a pre-defined class set and the training data domain.", + "bbox": [ + 509, + 642, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another line of approaches works by initialing the information about the target object. The most common approach in this group is semi-supervised video object segmentation (VOS) which tracks a given mask in the first frame. The recent SAM 2 [39] also falls into this group. SAM 2 generalizes the VOS task by taking multiple prompt types, including masks, boxes, and clicks similar to image segmentation methods [19, 22], and enables interactive video segmentation. Trained on large-scale training data, SAM 2 shows", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during an internship at Adobe Research.", + "bbox": [ + 107, + 886, + 437, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24234", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strong zero-shot broad-domain selection capability.", + "bbox": [ + 89, + 90, + 429, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite recent progress, a significant challenge remains in selecting and tracking all entities from a video—a task we term Video Entity Segmentation, which aims to segment any distinct group with coherent semantic meaning. While SAM 2 is an effective model for tracking each individual entity, we identified two limitations in selecting all entities: (1) It initializes tracking from user prompts so it cannot track unmarked entities and new entities appearing in different frames, as shown in Figure 1. (2) It tracks each entity independently which is not only inefficient and can lead to less optimal results (e.g., mask overlapping).", + "bbox": [ + 88, + 106, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose EntitySAM, an extension of SAM 2 for zero-shot video entity segmentation. EntitySAM retains the zero-shot potential of SAM 2 while eliminating the need for explicit prompts for each entity and enabling seamless integration of new entities during tracking. Inspired by transformer-based object detectors [5, 6, 42], we sophisticatedly integrated query-based object discovery and association into SAM 2 making minimal changes to the original structure. To be specific, we replace the original mask decoder with a new entity decoder and introduce an automatic prompt generator that generates SAM 2 consumable prompts from learnable object queries. In this way, we can preserve most of the pre-trained SAM 2 weights. The proposed entity decoder differs from the original SAM 2 mask decoder in that it enables inter-entity communication to predict multiple entity masks without overlaps. Another notable addition to SAM 2 is a semantic encoder. We empirically find that SAM 2 encoder has a weak semantic awareness and supplementing semantic information using an external feature encoder greatly boosts the performance.", + "bbox": [ + 91, + 273, + 483, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train EntitySAM on the COCO dataset [23] without using any category annotation and evaluate it on various zero-shot video segmentation tasks. First we introduce a video entity segmentation task. Video entity segmentation is class-agnostic, focusing on the segmentation and tracking of \"entities\" rather than predefined object categories. For this, we re-purpose the video panoptic segmentation dataset [28] by removing category annotation. In addition to this, we evaluate our method on multiple video instance and semantic segmentation benchmarks in zero-shot setting. Experimental results show that EntitySAM greatly outperforms previous unified video segmentation models [52] and strong baselines that combine an image detector [7] and a mask propagator [11]. Moreover, our method can be easily adapted to class-specific video panoptic segmentation by integrating state-of-the-art vision-language models [1].", + "bbox": [ + 89, + 574, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 829, + 241, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taxonomy of Video Segmentation Tasks Video segmentation tasks can be broadly categorized into class-specific and class-agnostic paradigms. Class-specific tasks,", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such as Video Instance Segmentation [36, 42, 48], Video Semantic Segmentation [30, 40], and Video Panoptic Segmentation [20, 28], operate within a closed vocabulary, segmenting \"thing\" or \"stuff\" categories based on predefined labels throughout video sequences.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, class-agnostic tasks, such as Video Object Segmentation (VOS) [34, 46], rely on prompt-based guidance. Unlike image-level prompts, video prompts typically require initial mask input in the first frame for temporal propagation. Recent work, such as SAM 2 [39], demonstrates that large-scale pre-training can generalize VOS to zero-shot applications. However, its reliance on explicit prompts becomes inefficient when tracking multiple dense objects with frequent occlusions or reappearance.", + "bbox": [ + 511, + 169, + 906, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances have expanded video segmentation to open-world and open-vocabulary paradigms. The BURST dataset [3] introduces annotations across 482 diverse categories with a focus on thing-level segmentation, while UVO [13] takes a class-agnostic open-world approach, treating all video objects as a single category without predefined labels. Entity Segmentation [26, 37] has shown promising generalization to novel categories at the image level, but extending this approach to video introduces significant challenges for zero-shot evaluation. To bridge this gap, we propose Video Entity Segmentation, a new task aimed at segmenting entities across video sequences without relying on predefined categories or explicit prompts. This approach extends image-level entity understanding to maintain temporal consistency across frames, enabling generalized handling of novel categories in a zero-shot setting. By avoiding rigid class definitions, our framework provides greater flexibility, particularly beneficial for real-world applications with unknown or fluid category boundaries.", + "bbox": [ + 511, + 306, + 906, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Query-based Video Segmentation Query-based video segmentation models are widely used for dense prediction tasks, including video instance [6, 14, 17, 18, 42-44] and semantic segmentation [2, 21]. Inspired by detection transformers [5], these methods use queries to represent object information and apply cross-attention with image features to generate segmentation masks, demonstrating effectiveness in video instance segmentation. Temporal association is achieved through either direct query matching [17, 45] or additional tracking modules [15]. While effective in controlled scenarios, these query-based methods without explicit memory propagation struggle with long or complex video sequences due to challenges in modeling long-term temporal dependencies [36].", + "bbox": [ + 511, + 595, + 908, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Memory-based Video Segmentation Memory-based methods [9, 24, 31, 33, 47, 49, 50] are effective in Video Object Segmentation [4, 12, 46, 46], using initial ground truth masks as prompts. Memory features are stored in memory banks for segmentation in subsequent frames, excelling in long-term video tracking [8, 10]. SAM 2 [39] validates the", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24235", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generalizability of memory-based designs but processes objects separately, limiting efficiency and lacking inter-object communication. Recent works have attempted to integrate query-based and memory mechanisms, such as Cutie [10] for object-level VOS and UniVS [29] for multi-task inference. However, joint training of memory and query modules is computationally challenging for long videos. Our EntitySAM addresses these issues by leveraging SAM 2's frozen memory encoder and introducing a novel query-based entity decoder for efficient joint training, combining the strengths of both approaches.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 263, + 181, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present EntitySAM for Video Entity Segmentation, enabling automatically segmenting and tracking all entities in the video sequence. By integrating query and memory-based video segmentation designs, EntitySAM handles multi-object tracking with inter-object communication and object-level contextual information. In Section 3.1, we first introduce the Video Entity Segmentation task, which offers improved generalization and reduced annotation complexity compared to other video segmentation tasks, such as Video Instance / Panoptic / Object Segmentation. In Section 3.2, we provide an overview of SAM 2's architecture and key characteristics. Section 3.3 details the architectural design of our EntitySAM based on SAM 2, highlighting our enhanced feature fusion and decoding mechanisms, along with comprehensive training and inference protocols.", + "bbox": [ + 89, + 287, + 483, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Video Entity Segmentation", + "text_level": 1, + "bbox": [ + 89, + 523, + 333, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Entity Segmentation aims to segment and track all entities within video sequences using a class-agnostic approach. In this context, an \"entity\" encompasses both thing and stuff categories, representing any distinct group with coherent semantic meaning. The task requires continuous tracking and segmentation of these entities through temporal sequences, with annotations represented as tube masks that capture the spatial-temporal evolution of each entity.", + "bbox": [ + 89, + 546, + 483, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given an input video sequence $I \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ , both annotations and predictions consist of $N$ non-overlapping masklets $M_{i_{i=1}}^{N}$ , where each masklet $M_{i} \\in [0,1]^T \\times [0,1]$ . For predictions, the task additionally requires a confidence score $s_{i} \\in [0,1]$ associated with each masklet. Under the non-overlapping constraint, these $N$ masklets can be merged into a single video sequence $M \\in \\mathbb{N}^{T \\times H \\times W}$ , where each pixel value represents the corresponding prediction ID.", + "bbox": [ + 89, + 667, + 483, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For evaluation, we first review traditional Video Panoptic Quality (VPQ), which computes metrics per category:", + "bbox": [ + 89, + 803, + 482, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV P Q = \\frac {1}{| C |} \\sum_ {c \\in C} \\frac {\\sum_ {t} I o U _ {c , t}}{T P _ {c} + 0 . 5 \\times F P _ {c} + 0 . 5 \\times F N _ {c}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 840, + 482, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better evaluate video entity segmentation, we propose", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/5bbe247da584689b00b2543bc33fc42fad4e06ceb739e54d4452a671dfdb475d.jpg", + "table_caption": [ + "Table 1. Comparison of Different Video Segmentation Tasks" + ], + "table_footnote": [], + "table_body": "
TaskCategoryInference
ThingsStuffAutomatic PromptCross-DatasetZero-shot
Video Instance Seg.✓✓✓
Video Semantic Seg.✓✓✓
Video Object Seg.✓✓✓
Video Panoptic Seg.✓✓✓✓
Video Entity Seg.✓✓✓✓✓
", + "bbox": [ + 514, + 104, + 903, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Entity Quality (VEQ), a class-agnostic metric that assesses performance across temporal sequences:", + "bbox": [ + 511, + 191, + 903, + 223 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V E Q} = \\frac {\\sum_ {t} I o U _ {t}}{T P + 0 . 5 \\times F P + 0 . 5 \\times F N} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 232, + 903, + 263 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V E Q} ^ {S Q} = \\quad \\frac {\\sum_ {t} I o U _ {t}}{T P} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 267, + 903, + 296 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V E Q} ^ {R Q} = \\frac {\\sum_ {t} T P}{T P + 0 . 5 \\times F P + 0 . 5 \\times F N} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 297, + 903, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $IoU_{t}$ represents the temporal intersection over union across frames, measuring both spatial accuracy and temporal consistency. Here, $TP$ , $FP$ , and $FN$ are computed globally across all entities without category differentiation. For consistency with Video Panoptic Segmentation evaluation, we also implement a class-agnostic variant of the STQ (Segmentation and Tracking Quality) metric, denoted as $\\mathrm{STQ}^{EN}$ , which emphasizes tracking consistency across frames. The formulation of VEQ provides a more direct assessment of video-level segmentation and tracking performance while eliminating the complexity of category-specific evaluation. Moreover, this class-agnostic approach reduces annotation requirements, enables large-scale training, and enhances model generalizability across datasets, particularly for open-vocabulary and zero-shot scenarios.", + "bbox": [ + 511, + 339, + 906, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Table 1, Video Entity Segmentation presents distinct advantages compared to existing video segmentation paradigms: Unlike Video Object Segmentation, it operates automatically and accommodates multiple dynamic objects. In contrast to Video Instance Segmentation, it extends beyond foreground objectness to include background entities. Compared to Video Semantic Segmentation, it provides instance-level annotations. Unlike Video Panoptic Segmentation, it adopts a class-agnostic approach that eliminates the constraint of pre-defined category vocabularies. This unified entity representation allows both thing and stuff categories to be learned jointly and subsequently classified using external Vision Language Models, enabling advanced open-vocabulary evaluation and reasoning-based segmentation tasks.", + "bbox": [ + 511, + 566, + 908, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Method Background: SAM 2", + "text_level": 1, + "bbox": [ + 511, + 801, + 774, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM 2 [39] extends the Segment Anything model [22] to encompass both image and video domains. Through an iterative process of data annotation and model training, SAM 2 leverages both manual and automated datasets generated by its data engine. For video segmentation functionality,", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24236", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb6ed7ec0b4a7475fdc8f8a5abb4e21daa7ad49e64373544165aed1ea7bb37e7.jpg", + "image_caption": [ + "Figure 2. (a) Overview of the EntitySAM framework. EntitySAM utilizes the frozen encoder and memory parameters from SAM 2, incorporating a dual encoder design for enhanced semantic features (Section 3.3.1). The PromptGenerator (Section 3.3.2) automatically generates prompts from Prompt Queries. The enhanced features and distinct query groups are processed by the EntityDecoder (Section 3.3.3) to produce video mask outputs. (b) Self-attention and cross-attention mechanisms in EntityDecoder layers." + ], + "image_footnote": [], + "bbox": [ + 91, + 71, + 903, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAM 2 incorporates a classical memory mechanism to facilitate object tracking.", + "bbox": [ + 89, + 306, + 480, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture processes video input through multiple stages: initially, a multi-scale Image encoder extracts frame-level features, which then undergo cross-attention with previous video-level memory features to derive per-mask temporal information. Concurrently, input prompts are processed through a dedicated prompt encoder to generate prompt encodings. These various features and prompts are integrated through cross-attention mechanisms and dynamic convolution operations within the mask decoder to generate mask outputs. The model maintains temporal coherence by encoding and storing mask sequences in a memory bank, which updates frame-by-frame and feeds into memory attention for processing subsequent frames.", + "bbox": [ + 88, + 338, + 482, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through extensive training on 35M manual and automated mask annotations, SAM 2 shows exceptional zero-shot performance across various video object segmentation benchmarks, including DAVIS [35], YTVOS [46], LVOS [16], and MOSE [12]. However, SAM 2 exhibits notable limitations: it requires initial mask/prompt specification at the sequence start and cannot effectively accommodate new objects introduced during tracking. Moreover, the model's efficiency deteriorates when handling large numbers of concurrent objects, indicating challenges in scaling to more complex scenarios with multiple dynamic objects.", + "bbox": [ + 88, + 536, + 482, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. EntitySAM", + "text_level": 1, + "bbox": [ + 89, + 715, + 215, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coupled with the proposed Video Entity Segmentation task, we present EntitySAM, a novel model that achieves strong zero-shot video entity segmentation performance while requiring only image-level class-agnostic entity annotations. As illustrated in Figure 2, our model preserves the memory-based architecture of the original SAM by freezing both the pretrained image encoder and memory-related modules, thus retaining its advantages in long-term tracking of objects. Meanwhile, EntitySAM comprises three learnable components: Dual Visual Encoder, EntityPromptGenerator,", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and EntityDecoder. The following sections will detail these architectural components.", + "bbox": [ + 511, + 306, + 906, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Dual Visual Encoder", + "text_level": 1, + "bbox": [ + 511, + 344, + 712, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our EntitySAM leverages the frozen Segment Anything Model 2 (SAM 2) image encoder as its backbone, as well as its frozen memory encoder and memory attention layers. The frozen design preserves SAM 2's zero-shot feature extraction capabilities, which were developed through training on extensive video object tracking datasets. Additionally, freezing the memory encoder substantially reduces GPU memory requirements during the training of multiple dense entity tracking in video sequences.", + "bbox": [ + 509, + 369, + 906, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although SAM 2 exhibits impressive performance across multi-granularity class-agnostic segmentation tasks, we reveal its limitations in maintaining consistent entity-level representations as in Table 9. Specifically, SAM 2 struggles to generate reliable confidence scores for entity classification, where an 'entity' refers to a semantically coherent group. Thus in addition to the original SAM 2's backbone feature, we integrate the backbone features from DINOv2 [32], which produces consistent feature representations for semantically similar regions by large-scale self-supervised training.", + "bbox": [ + 511, + 506, + 905, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, EntitySAM incorporates DINOv2 through a linear projection layer, aligning its feature dimensionality with that of SAM 2. The aligned features are then concatenated and fed into the decoder to form a fused representation. This enhances the encoded visual representation by preserving the zero-shot capabilities for video entity segmentation while effectively mitigating SAM 2's limitations in entity-level understanding.", + "bbox": [ + 511, + 671, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Automatic Prompt Generator", + "text_level": 1, + "bbox": [ + 511, + 800, + 774, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In SAM 2's architecture, the prompt encoder plays a crucial role in image-level segmentation, where input prompts, such as points and bounding boxes, are processed to generate prompt embeddings for decoder segmentation. However, in video-level operations, SAM 2 bypasses the prompt", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24237", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "encoder, focusing solely on object tracking using historical memory features. To extend the model's capability for new entity tracking in video sequences, EntitySAM incorporates prompt embeddings during video training.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our EntitySAM implements a prompt generator utilizing learnable queries that perform cross-attention with enhanced visual features to predict input prompts for the prompt encoder. The resulting prompt embeddings are subsequently fed into the decoder at each frame to localize both existing and novel entities within the video sequence. Concurrently, this module predicts an entity confidence score, denoted as $S(\\text{entity})$ , which performs binary classification between foreground entities and background regions.", + "bbox": [ + 89, + 154, + 482, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More formally, let $\\mathcal{Q} \\in \\mathbb{R}^{N \\times D}$ represent the learnable queries and $\\mathcal{F} \\in \\mathbb{R}^{HW \\times D}$ denote the enhanced visual features, where $N$ is the number of queries, $D$ is the feature dimension, and $H, W$ are spatial dimensions. The prompt generation and scoring process can be expressed as: $\\mathcal{A} = \\text{CrossAttention}(\\mathcal{Q}, \\mathcal{F})$ , $\\mathcal{P} = \\text{FC}_{\\text{prompt}}(A)$ , and $\\mathcal{S} = \\text{FC}_{\\text{score}}(A)$ , where $A \\in \\mathbb{R}^{N \\times D}$ is the cross-attention output, $\\mathcal{P} \\in \\mathbb{R}^{N \\times D'}$ represents the generated prompts with $D'$ matching the expected prompt dimension required by the prompt encoder, and $\\mathcal{S} \\in \\mathbb{R}^{N \\times 1}$ denotes the entity confidence scores.", + "bbox": [ + 89, + 292, + 483, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Entity Mask Decoder", + "text_level": 1, + "bbox": [ + 89, + 478, + 294, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Entity Mask Decoder predicts output masks based on enhanced features in Section 3.3.1 and the learnable prompts through a multi-stage processing pipeline in Section 3.3.2. It employs two groups of batched queries: mask queries $\\mathcal{Q}_m\\in \\mathbb{R}^{N\\times 1\\times C}$ and IoU queries $\\mathcal{Q}_{iou}\\in \\mathbb{R}^{N\\times 1\\times C}$ , where $N$ denotes the number of entities and $C$ represents the feature dimension. These queries are concatenated with input prompts for self-attention to facilitate inter-entity communication. These queries are subsequently reshaped for cross-attention operations between token and image representations. In the final layer, mask queries undergo dynamic convolution with upsampled mask features to generate output masks, while IoU queries are processed through a dedicated prediction head to estimate conditioned IoU scores. The computational flow can be formalized as:", + "bbox": [ + 89, + 506, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} [ \\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime} ] = \\text {S e l f A t t e n t i o n} (\\text {c o n c a t} (\\mathcal {Q} _ {m}; \\mathcal {Q} _ {i o u})) \\\\ w h e r e \\operatorname {c o n c a t} \\left(\\mathcal {Q} _ {m}; \\mathcal {Q} _ {i o u}\\right) \\in \\mathbb {R} ^ {1 \\times 2 N \\times C} \\\\ F _ {o u t} = \\operatorname {C r o s s A t t e n t i o n} (\\operatorname {r e s h a p e} \\left(\\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime}\\right), F) \\\\ \\operatorname {w h e r e} \\operatorname {r e s h a p e} \\left(\\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime}\\right) \\in \\mathbb {R} ^ {N \\times 2 \\times C}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 748, + 480, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The supervision mechanism integrates outputs from both the prompt generator (entity scores) and mask decoder (masks and conditioned IoU predictions) through Hungarian matching with ground truth annotations. The expected", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "IoU for query $q$ can be expressed as:", + "bbox": [ + 513, + 90, + 756, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E (\\mathrm {I o U} _ {q}) = E (\\mathrm {I o U} _ {q} | \\mathrm {E n t} _ {q}) S (\\mathrm {E n t} _ {q}) + E (\\mathrm {I o U} _ {q} | \\neg \\mathrm {E n t} _ {q}) S (\\neg \\mathrm {E n t} _ {q}) \\\\ = E \\left(\\operatorname {I o U} _ {q} \\mid \\operatorname {E n t} _ {q}\\right) S \\left(\\operatorname {E n t} _ {q}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 114, + 903, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S(\\mathrm{Ent}_q)$ represents the entity confidence score from PromptGenerator in Section 3.3.2 and $E(\\mathrm{IoU}_q|\\neg \\mathrm{Ent}_q) = 0$ .", + "bbox": [ + 511, + 161, + 903, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The resulting output masks are processed through the frozen memory encoder, memory bank, and memory attention mechanisms, with gradient computation detached for memory-related components. This enables efficient training of the queries-based tracking component while maintaining frozen memory modules. Consequently, our approach achieves scalable training for over 50 entities in 8-frame video clips, significantly surpassing SAM 2's original capacity of handling only 3 video objects during training.", + "bbox": [ + 511, + 191, + 906, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4 Training Loss and Open-Vocabulary Inference", + "text_level": 1, + "bbox": [ + 511, + 344, + 890, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ an end-to-end training objective for EntitySAM, formulated as:", + "bbox": [ + 511, + 368, + 903, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {M a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\text {E n t i t y}} + \\lambda_ {2} \\mathcal {L} _ {\\text {I o U}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 409, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ supervises both mask prediction and binary entity classification, following the Mask2former framework [7]. The hyperparameters $\\lambda_{1}$ and $\\lambda_{2}$ balance the respective loss terms and are empirically set to 1.0 and 1.0.", + "bbox": [ + 511, + 436, + 903, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, EntitySAM accommodates both class-agnostic and class-specific evaluation protocols. For class-agnostic scenarios, entity masks are directly generated and merged based on confidence scores and mask overlap metrics. In class-specific evaluation, we select the frame with the highest entity confidence score for each masklet prediction. The predicted entity masks are utilized to crop corresponding regions of interest, which, together with carefully designed prompts containing candidate category names, are processed through GPT-4o [1] for final classification. We demonstrate in the supplementary materials that this mask-guided cropping approach achieves superior classification performance compared to alternative methods.", + "bbox": [ + 511, + 497, + 905, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 700, + 645, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 726, + 702, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets Video entity segmentation typically requires extensive class-agnostic dense entity video segmentation annotations. Although it eliminates the need for class-specific annotations, creating frame-by-frame annotations remains labor-intensive. Inspired by [51, 52], our training pipeline leverages COCO [23] panoptic segmentation data without category label annotation (denoted as $\\mathrm{COCO}^*$ ) and extends it to pseudo-video sequences. Our approach utilizes only image-level annotations, eschewing video-level labels, to achieve effective video entity segmentation. The pipeline", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24238", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "also enables zero-shot cross-dataset evaluation across multiple benchmarks. We demonstrate our model's versatility through comprehensive zero-shot evaluations on various tasks: video panoptic segmentation on VIPSeg [28], video semantic segmentation on VSPW [27], video instance segmentation on Youtube-VIS 2019 [48], Youtube-VIS 2021 [48], and OVIS [36], and video entity segmentation on our proposed VIPEntitySeg dataset—a class-agnostic variant of VIPSeg. Notably, all zero-shot evaluations are performed using a single trained model, demonstrating its robust zero-shot generalization capabilities across different video segmentation tasks and datasets.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details In our experimental setup, we employ two backbone architectures: ViT-Small and ViT-Large. The models are trained using the AdamW optimizer with a learning rate of $5 \\times 10^{-5}$ and weight decay of 0.05. Input images are cropped to $1024 \\times 1024$ pixels for compatibility with SAM 2 ViT/16 and $896 \\times 896$ pixels for DINOv2 ViT/14. Training is conducted with a batch size of 8. The loss function combines multiple components: cross-entropy and dice losses for mask prediction, L1 loss for IoU prediction, and focal loss for binary entity classification. Our training strategy consists of two phases: initial training on static COCO images for 40K iterations, followed by training on COCO pseudo-video sequences with clip length 8 for 10K iterations. Both stages are class-agnostic. During inference, we utilize a single entity category output mode with confidence scoring. Video entity quality (VEQ) is evaluated in a class-agnostic manner. For class-specific evaluation tasks, we implement a multi-step process: first selecting the highest-confidence mask from the masklet for each prediction, then extracting the corresponding frame region within the mask, and finally employing GPT-4o for classification.", + "bbox": [ + 91, + 273, + 486, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Zero-shot Comparison", + "text_level": 1, + "bbox": [ + 89, + 604, + 302, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot Video Entity Segmentation In Table 2 (left), we evaluate our approach on VIPEntitySeg using class-agnostic metrics against several baseline methods. For SAM 2, we adopt two prompting manners, a grid point prompt sampling initialization and a COCO-trained Mask2Former [7] mask prompts initialization. We additionally compare against state-of-the-art zero-shot segmentation models, including: 1) FC-CLIP [51] integrated with MinVIS [17] association, 2) DEVA [11] coupled with COCO-trained Mask2Former [7] for per-frame segmentation, and 3) OV-DVIS++ [52] in both online and offline modes. All these SOTA models were trained on COCO with category annotations, with DEVA utilizing additional external datasets for its association model. We compare with these methods on our class-agnostic VIPEntitySeg benchmark. Our EntitySAM achieves significant performance improvements over all baselines and SOTA models with both ViT-Small and ViT-Large backbones, despite training", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "without category annotations. This superior performance demonstrates the effectiveness of our class-agnostic training approach and proposed modules. Notably, we observe substantial improvements in $\\mathrm{VEQ}^{RQ}$ , indicating enhanced cross-frame entity identification and reduced false positives. Figure 3 further illustrates our model's robustness by evaluating VEQ across varying frame intervals, demonstrating that EntitySAM maintains consistent performance over extended temporal sequences.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding the class-agnostic $\\mathrm{STQ}^{EN}$ metric, which emphasizes tracking performance, DEVA achieves higher scores through its tracking association model trained on external video object segmentation datasets. However, DEVA relies on a two-stage pipeline of detection followed by association, lacking end-to-end integration. While our approach achieves lower scores, it delivers comparable performance through a streamlined process that requires only COCO training data and single-pass inference.", + "bbox": [ + 511, + 227, + 903, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot Video Panoptic Segmentation In Table 2 (right), we extend our evaluation to VIPSeg for class-specific video panoptic segmentation. Several baseline models, including Mask2Former and DEVA, are constrained to VIPSeg evaluation due to their closed-vocabulary nature. When compared to open-vocabulary models such as FC-CLIP and OV-DVIS++, our approach demonstrates substantial improvements, particularly with the ViT-Small backbone. These results validate our approach of \"post-classification with multimodal models\", indicating that it surpasses traditional open-vocabulary segmentation training pipelines despite operating without category annotations. Our EntitySAM achieves competitive performance on both class-agnostic and class-specific tasks, suggesting that category annotations may not be essential for effective open-vocabulary video segmentation.", + "bbox": [ + 511, + 367, + 903, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot Video Semantic Segmentation Table 3 presents our results on video semantic segmentation, obtained by merging per-pixel mask predictions. Our method achieves improvements in both mean Intersection over Union (mIoU) and Video Consistency (VC) metrics. The VSPW evaluation dataset, which contains the same video content but with extended frame sequences, validates our model's capability for effective long-term tracking.", + "bbox": [ + 511, + 611, + 903, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot Video Instance Segmentation As shown in Table 4, we evaluate EntitySAM on three popular video instance segmentation datasets. Our approach on extending video entity segmentation to video instance segmentation requires no model retraining—we simply add a background/stuff class during inference and filter out predictions with that class to obtain instance results. Our model demonstrates significant performance improvements over state-of-the-art open-vocabulary video instance segmentation models. In particular, on the challenging OVIS dataset, which features frequent occlusions, our model maintains robust", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24239", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5da507146307bb83555f58559ecf5d2ca8f1c1bb07e5eceab35918957d1679ff.jpg", + "table_caption": [ + "Table 2. Comparison of Zero-shot Video Entity Segmentation (left) and Video Panoptic Segmentation (right) on COCO $\\rightarrow$ VIPSeg evaluation. EntitySAM is trained on the COCO dataset without category annotations. In VIPSeg, - means the results of open-vocabulary evaluation for closed-vocabulary models are not available. * denotes that DEVA is trained with multiple external video object segmentation datasets for the temporal association model." + ], + "table_footnote": [], + "table_body": "
MethodBackboneClass Agnostic TrainingVIPEntitySegVIPSeg
VEQVEQSQVEQRQSTQENVPQVPQThVPQStSTQ
SAM 2.1 [39] (Everything Init)ViT-S✓26.982.532.638.7-
SAM 2.1 [39] (Mask2former Init)ViT-S + ResNet5046.283.655.234.5-
DEVA* [11]ResNet-5049.083.658.642.4-
FC-CLIP [51]ResNet-5041.584.049.429.722.325.519.119.7
OV-DVIS++(Online) [52]ResNet-5045.983.754.833.624.426.822.422.0
OV-DVIS++(Offline) [52]ResNet-5046.782.856.335.623.826.421.424.4
OursViT-S✓51.884.761.141.028.732.925.131.4
SAM 2.1 [39] (Everything Init)ViT-L✓30.583.136.741.7-
SAM 2.1 [39] (Mask2former Init)ViT-L + Swin-L50.684.460.039.1-
DEVA* [11]Swin-L50.883.860.645.3-
FC-CLIP [51]ConvNext-L43.484.851.232.427.930.925.024.2
OV-DVIS++(Online) [52]ConvNext-L49.483.559.237.928.931.326.828.4
OV-DVIS++(Offline) [52]ConvNext-L49.083.159.040.030.431.929.132.2
OursViT-L✓54.684.764.543.331.437.526.033.5
", + "bbox": [ + 91, + 128, + 906, + 348 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/510c47af21cd23436b03f2987b081f2c5ac8840a264e858d5619f36250961e53.jpg", + "table_caption": [ + "Table 3. Zero-Shot Open-vocabulary Video Semantic Segmentation on COCO $\\rightarrow$ VSPW. EntitySAM is trained on the COCO dataset without category annotations." + ], + "table_footnote": [], + "table_body": "
MethodBackboneClass Agnostic TrainingmVC8VSPW mVC16mIOU
FC-CLIP [51]ResNet-5084.982.724.3
OV-DVIS++(Online) [52]ResNet-5092.791.527.6
OV-DVIS++(Offline) [52]ResNet-5092.491.328.4
OursViT-S✓94.093.034.6
FC-CLIP [51]ConvNext-L89.988.428.9
OV-DVIS++(Online) [52]ConvNext-L94.293.334.3
OV-DVIS++(Offline) [52]ConvNext-L93.993.034.1
OursViT-L✓94.693.735.5
", + "bbox": [ + 91, + 397, + 482, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance, benefiting from the entity learning strategy employed during training.", + "bbox": [ + 89, + 525, + 482, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Experiments", + "text_level": 1, + "bbox": [ + 89, + 560, + 295, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation studies using the ViT-Small backbone on the COCO $\\rightarrow$ VIPEntitySeg video entity segmentation task with zero-shot protocol to evaluate the effectiveness of our proposed modules.", + "bbox": [ + 89, + 583, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the Proposed Modules Table 5 presents ablation studies of our proposed modules. Our baseline, SAM 2 with memory-based tracking modules, cannot generate predictions automatically without appropriate prompting. Using grid-point sampling leads to numerous false positive predictions, achieving a VEQ of only 26.9. Removing the memory design and implementing solely a query-based EntityDecoder with query association for tracking improves performance to 41.3, though this remains suboptimal. Combining memory-based and query-based approaches significantly enhances performance to 50.2, surpassing both DEVA and OV-DVIS++. Finally, incorporating a complementary DINOv2 dual feature encoder achieves the best result (51.8) for the ViT-Small backbone.", + "bbox": [ + 89, + 643, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Query Design As illustrated in Figure 2, our design incorporates three query types: two groups in EntityDecoder and one group in PromptGenerator. Our abla", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faff9aa980542777667efb97d103d049495e6848ff1377b97d5ecac62b7451a9.jpg", + "image_caption": [ + "Figure 3. VEQ comparison with different evaluation time intervals. EntitySAM improves over long intervals." + ], + "image_footnote": [], + "bbox": [ + 516, + 353, + 694, + 464 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4298e2160ec64bc276eea6e068907ffaf70a6180032fc2487f0d8276ec0248ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 353, + 903, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion experiment of query design in Table 6 reveals that using a single query group to predict all outputs (mask, IoU, and confidence score) yields suboptimal performance with a VEQ of 49.2. Adding a second group of prompt queries improves performance to 49.7. Further separation into a third query group for distinct mask and IoU prediction achieves the best performance of 51.8. These results suggest that different query groups learn specialized distributions for their respective prediction tasks, departing from traditional detection transformers' shared-query architecture.", + "bbox": [ + 511, + 516, + 906, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on EntityDecoder Depth Table 7 studies the influence of the entity decoder's depth layer. For the ViT-Small backbone, a decoder depth of 4 achieves optimal performance with a VEQ of 51.8 on COCO $\\rightarrow$ VIPSeg setting. Further increasing decoder depth shows no additional performance benefits.", + "bbox": [ + 511, + 670, + 905, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Training Stages As detailed in Section 4.1, our training process consists of two stages: initial training on single images (treated as zero-length videos) followed by training on pseudo video clips of length 8. During the first stage, the memory component remains unused while the EntityDecoder learns to identify new entities in each image. Table 8 shows that this image-only training achieves limited video entity segmentation performance with a VEQ of 31.4. However, this stage exposes the model to diverse en", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24240", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c29258108f6ef968b52fb15b9c62d94b3ded3902c57c2b7310e6421fe97b7b82.jpg", + "table_caption": [ + "Table 4. Zero-Shot Open-vocabulary Video Instance Segmentation on COCO $\\rightarrow$ Validation sets of YouTube-VIS 2019, 2021, and OVIS. EntitySAM is trained on the COCO dataset without category annotations." + ], + "table_footnote": [], + "table_body": "
MethodBackboneClass Agnostic TrainingYoutube-VIS 2019Youtube-VIS 2021OVIS
AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)
Detric[53]-OWTB[25]ResNet-5017.9----16.7----9.0----
MindVLT[41]ResNet-5023.1----20.9----11.4----
FC-CLIP [51]ResNet-5028.943.929.932.441.225.540.026.828.037.011.825.110.58.516.4
OV-DVIS++(Online) [52]ResNet-5034.550.639.239.549.530.946.734.834.445.814.831.213.110.524.7
OV-DVIS++(Offline) [52]ResNet-5034.451.839.240.050.631.048.534.334.646.813.028.410.99.825.1
OursViT-S✓44.960.749.543.662.438.852.543.136.256.521.035.820.812.336.0
", + "bbox": [ + 91, + 111, + 906, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b3503f943d827de1dead3578b761f94ace03c5a033c51fe383287289bfb56977.jpg", + "table_caption": [ + "Table 5. Ablation Study on the Proposed Modules." + ], + "table_footnote": [], + "table_body": "
ModelMemoryEntityDecoderDual Visual EncoderVEQ
SAM 2 [39]✓26.9
Ours✓✓41.3
✓✓50.2
✓✓✓51.8
", + "bbox": [ + 91, + 227, + 480, + 323 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7847958b38175703587365d3369a2e006363c3d22ed2090e0c2353702def00e4.jpg", + "table_caption": [ + "Table 6. Ablation Study on Query Design." + ], + "table_footnote": [], + "table_body": "
EntityDecoderPromptGeneratorCOCO → VIPSeq
VEQVEQSQVEQRQ
Shared Queries49.284.058.6
Shared Queries✓49.784.259.0
Separate Queries✓51.884.761.1
", + "bbox": [ + 91, + 343, + 480, + 421 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66e6605ad3899d12f0d6a4ea3389ac552ad0e4b7daf74ae50d1daac951b37e3f.jpg", + "table_caption": [ + "Table 7. Ablation Study on EntityDecoder Depth." + ], + "table_footnote": [], + "table_body": "
DecoderDecoder depthCOCO → VIPSeg
VEQVEQSQVEQRQ
EntityDecoder251.384.760.5
451.884.761.1
651.684.661.0
851.184.460.6
", + "bbox": [ + 91, + 450, + 480, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tities, and when combined with the subsequent video training stage—which aligns EntityDecoder with memory encoders—performance improves substantially to 51.8 VEQ. This two-stage approach outperforms training exclusively on pseudo video sequences.", + "bbox": [ + 89, + 564, + 482, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Dual Encoder In Section 3.3.1, we introduced encoder feature fusion designs for dual encoders. As shown in Table 9, our ablation studies reveal that incorporating features from either CLIP [38] or SAM [22] alongside the SAM 2 encoder features yields no improvement in zero-shot performance. This can be attributed to CLIP's vision encoder being optimized for classification rather than segmentation tasks, resulting in representations less suitable for segmentation. Similarly, SAM, being trained with an approach analogous to SAM 2, provides features that largely overlap with SAM 2's capabilities rather than offering complementary information. In contrast, DINOv2, trained through self-supervision, demonstrates significant benefits when combined with SAM 2. Its self-supervised training enables DINOv2 to learn consistent image representations that effectively complement SAM 2's supervised features, ultimately enhancing the overall performance.", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ee8aa851fed3ac9f1bb0a01fbe378e7e697659ecf1992179902157d4e079a77f.jpg", + "table_caption": [ + "Table 8. Ablation Study on Training stages." + ], + "table_footnote": [], + "table_body": "
ModelImagePseudo VideoCOCO → VIPSeg
VEQVEQSQVEQRQ
Ours✓31.479.439.5
✓49.284.258.4
✓✓51.884.761.1
", + "bbox": [ + 514, + 227, + 903, + 311 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2263415b5bbfe8948e49f88974c7dd80a5846d977bb235645c22a6a6393c6dbe.jpg", + "table_caption": [ + "Table 9. Ablation Study on Dual Encoder Feature." + ], + "table_footnote": [], + "table_body": "
Encoder FeatureDual EncoderCOCO → VIPSeg
VEQVEQSQVEQRQ
SAM 2-50.284.559.4
SAM [22]47.684.256.5
CLIP [38]49.284.058.6
DINOv2 [32]51.884.761.1
", + "bbox": [ + 514, + 333, + 903, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative experiments and visualization", + "text_level": 1, + "bbox": [ + 511, + 446, + 867, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 demonstrates our model's performance on zero-shot video sequences, showcasing consistent entity tracking and segmentation across frames. Additional qualitative results are provided in the supplementary materials.", + "bbox": [ + 511, + 470, + 905, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/971a321b7254a0fef2a56a5217755b09d596e0f91ed9a1790494d156c4a0ff8e.jpg", + "image_caption": [ + "Figure 4. Visualization of zero-shot video entity segmentation." + ], + "image_footnote": [], + "bbox": [ + 514, + 537, + 645, + 651 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b224e69f03d2ef284977ad493ef40659edb97820e6316c8007ed3d070f783760.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 537, + 774, + 651 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55082acb74841d8e6cd604a3be21c25b6e52b401b5ba35cb27a125feb2afdd45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 537, + 903, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present EntitySAM, a framework that integrates query and memory mechanisms to enhance SAM 2 for zero-shot video entity segmentation. Our approach eliminates the need for user-defined prompt inputs and enables automatic tracking of new entities across frames. Trained exclusively on COCO image-level mask annotations without category information, EntitySAM achieves superior performance on the Video Entity Segmentation task. The model demonstrates strong generalization capabilities across various zero-shot evaluations, including Video Panoptic / Semantic / Instance Segmentation. These results provide valuable insights into class-agnostic training approaches and their effectiveness in zero-shot scenarios.", + "bbox": [ + 511, + 703, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24241", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5", + "[2] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2", + "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2", + "[4] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. In arXiv preprint arXiv:1905.00737, 2019. 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[6] Bowen Cheng, Anwesa Choudhuri andd Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. In https://arxiv.org/abs/2112.10764, 2021. 2", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 5, 6", + "[8] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 2", + "[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 2", + "[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In arXiv, 2023. 2, 3", + "[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 1, 2, 6, 7", + "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, 2023. 2, 4", + "[13] Yuming Du, Wen Guo, Yang Xiao, and Vincent Lepetit. Uvo challenge on video-based open-world segmentation 2021: 1st place solution. ICCV Workshop, 2021. 2", + "[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. Advances in Neural Information Processing Systems, 35:23109–23120, 2022. 2", + "[15] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14623-14632, 2023. 2", + "[16] Lingyi Hong, Wenchao Chen, Zhongying Liu, Wei Zhang, Pinxue Guo, Zhaoyu Chen, and Wenqiang Zhang. Lvos: A benchmark for long-term video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13480-13492, 2023. 4", + "[17] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6", + "[18] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 2", + "[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, 2023. 1", + "[20] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2", + "[21] Dahun Kim, Jun Xie, Huiyu Wang, Siyuan Qiao, Qihang Yu, Hong-Seok Kim, Hartwig Adam, In So Kweon, and Liang-Chieh Chen. Tubeformer-deeplab: Video mask transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13914-13924, 2022. 2", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 3, 8", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5", + "[24] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, 2022. 2", + "[25] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 8", + "[26] Qi Lu, Jason Kuen, Shen Tiancheng, Gu Jiumiang, Guo Weidong, Jia Jiaya, Lin Zhe, and Yang Ming-Hsuan. High-quality entity segmentation. In ICCV, 2023. 2", + "[27] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. Vspw: A large-scale dataset for video scene parsing in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4133-4143, 2021. 1, 6", + "[28] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 1, 2, 6", + "[29] Xindong Zhang Minghan Li, Shuai Li and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries, 2024. 3", + "[30] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6819-6828, 2018. 2", + "[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24242", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4, 8", + "[33] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In CVPR, 2022. 2", + "[34] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 2", + "[35] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 4", + "[36] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation. IJCV, 2022. 2, 6", + "[37] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. In arXiv preprint arXiv:2107.14228, 2021. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8", + "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3, 7, 8", + "[40] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3126-3137, 2022. 2", + "[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In proceedings of the IEEE/CVF international conference on computer vision, pages 4057-4066, 2023. 8", + "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2", + "[43] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In European Conference on Computer Vision, pages 553-569. Springer, 2022." + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, 2022. 2", + "[45] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In European Conference on Computer Vision, pages 588-605. Springer, 2022. 2", + "[46] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. In ECCV, 2018. 2, 4", + "[47] Xiaohao Xu, Jinglu Wang, Xiao Li, and Yan Lu. Reliable propagation-correction modulation for video object segmentation. In AAAI, 2022. 2", + "[48] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 6", + "[49] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 2", + "[50] Zongxin Yang, Jiaxu Miao, Yunchao Wei, Wenguan Wang, Xiaohan Wang, and Yi Yang. Scalable video object segmentation with identification mechanism. TPAMI, 2024. 2", + "[51] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 5, 6, 7, 8", + "[52] Tao Zhang, Xingye Tian, Yikang Zhou, Shunping Ji, Xuebo Wang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, and Yu Wu. Dvis++: Improved decoupled framework for universal video segmentation. arXiv preprint arXiv:2312.13305, 2023. 2, 5, 6, 7, 8", + "[53] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8" + ], + "bbox": [ + 516, + 92, + 903, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24243", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_model.json b/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0bf6d1ff7861738a487b5b9e21fe76c56b7bcd1e --- /dev/null +++ b/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.287, + 0.131, + 0.713, + 0.154 + ], + "angle": 0, + "content": "EntitySAM: Segment Everything in Video" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.18, + 0.76, + 0.219 + ], + "angle": 0, + "content": "Mingqiao Ye\\(^{1,*}\\) Seoung Wug Oh\\(^{2}\\) Lei Ke\\(^{3}\\) Joon-Young Lee\\(^{2}\\) \n\\(^{1}\\)EPFL 2Adobe Research 3Carnegie Mellon University" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.228, + 0.907, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.45, + 0.908, + 0.507 + ], + "angle": 0, + "content": "Figure 1. Zero-shot video entity segmentation performance comparison on VIPEntySeg dataset using models trained on COCO, showing: 1) SAM 2 [39] using Mask2Former [7] mask prompts for the initial frame, 2) Mask2Former with DEVA [11] association, and 3) our proposed EntitySAM. Our EntitySAM enhances SAM 2 by automatically segmenting and tracking novel entities without requiring user-specified prompts, achieving superior performance compared to existing state-of-the-art zero-shot tracking methods." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.526, + 0.327, + 0.541 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.558, + 0.486, + 0.876 + ], + "angle": 0, + "content": "Automatically tracking and segmenting every video entity remains a significant challenge. Despite rapid advancements in video segmentation, even state-of-the-art models like SAM 2 struggle to consistently track all entities across a video—a task we refer to as Video Entity Segmentation. We propose EntitySAM, a framework for zero-shot video entity segmentation. EntitySAM extends SAM 2 by removing the need for explicit prompts, allowing automatic discovery and tracking of all entities, including those appearing in later frames. We incorporate query-based entity discovery and association into SAM 2, inspired by transformer-based object detectors. Specifically, we introduce an entity decoder to facilitate inter-object communication and an automatic prompt generator using learnable object queries. Additionally, we add a semantic encoder to enhance SAM 2's semantic awareness, improving segmentation quality. Trained on image-level mask annotations without category information from the COCO dataset, EntitySAM demonstrates strong generalization on four zero-shot video segmentation tasks: Video Entity, Panoptic, Instance, and Semantic Segmentation. Results on six popular benchmarks show that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.588 + ], + "angle": 0, + "content": "EntitySAM outperforms previous unified video segmentation methods and strong baselines, setting new standards for zero-shot video segmentation. Our code and models are at github.com/ymq2017-entitysam." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.645, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Video segmentation is an essential task for various applications in video scene understanding, including video editing, AR/VR, and robotic perception. There are roughly two lines of video segmentation approaches. One line of approaches focuses on selecting semantic thing/stuff from predefined sets [27, 28, 48]. While these methods show strong performance in the benchmark, the predictions are limited to a pre-defined class set and the training data domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Another line of approaches works by initialing the information about the target object. The most common approach in this group is semi-supervised video object segmentation (VOS) which tracks a given mask in the first frame. The recent SAM 2 [39] also falls into this group. SAM 2 generalizes the VOS task by taking multiple prompt types, including masks, boxes, and clicks similar to image segmentation methods [19, 22], and enables interactive video segmentation. Trained on large-scale training data, SAM 2 shows" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.438, + 0.901 + ], + "angle": 0, + "content": "*This work was done during an internship at Adobe Research." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24234" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.431, + 0.106 + ], + "angle": 0, + "content": "strong zero-shot broad-domain selection capability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.482, + 0.272 + ], + "angle": 0, + "content": "Despite recent progress, a significant challenge remains in selecting and tracking all entities from a video—a task we term Video Entity Segmentation, which aims to segment any distinct group with coherent semantic meaning. While SAM 2 is an effective model for tracking each individual entity, we identified two limitations in selecting all entities: (1) It initializes tracking from user prompts so it cannot track unmarked entities and new entities appearing in different frames, as shown in Figure 1. (2) It tracks each entity independently which is not only inefficient and can lead to less optimal results (e.g., mask overlapping)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.274, + 0.485, + 0.575 + ], + "angle": 0, + "content": "In this paper, we propose EntitySAM, an extension of SAM 2 for zero-shot video entity segmentation. EntitySAM retains the zero-shot potential of SAM 2 while eliminating the need for explicit prompts for each entity and enabling seamless integration of new entities during tracking. Inspired by transformer-based object detectors [5, 6, 42], we sophisticatedly integrated query-based object discovery and association into SAM 2 making minimal changes to the original structure. To be specific, we replace the original mask decoder with a new entity decoder and introduce an automatic prompt generator that generates SAM 2 consumable prompts from learnable object queries. In this way, we can preserve most of the pre-trained SAM 2 weights. The proposed entity decoder differs from the original SAM 2 mask decoder in that it enables inter-entity communication to predict multiple entity masks without overlaps. Another notable addition to SAM 2 is a semantic encoder. We empirically find that SAM 2 encoder has a weak semantic awareness and supplementing semantic information using an external feature encoder greatly boosts the performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.818 + ], + "angle": 0, + "content": "We train EntitySAM on the COCO dataset [23] without using any category annotation and evaluate it on various zero-shot video segmentation tasks. First we introduce a video entity segmentation task. Video entity segmentation is class-agnostic, focusing on the segmentation and tracking of \"entities\" rather than predefined object categories. For this, we re-purpose the video panoptic segmentation dataset [28] by removing category annotation. In addition to this, we evaluate our method on multiple video instance and semantic segmentation benchmarks in zero-shot setting. Experimental results show that EntitySAM greatly outperforms previous unified video segmentation models [52] and strong baselines that combine an image detector [7] and a mask propagator [11]. Moreover, our method can be easily adapted to class-specific video panoptic segmentation by integrating state-of-the-art vision-language models [1]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.242, + 0.847 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Taxonomy of Video Segmentation Tasks Video segmentation tasks can be broadly categorized into class-specific and class-agnostic paradigms. Class-specific tasks," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "such as Video Instance Segmentation [36, 42, 48], Video Semantic Segmentation [30, 40], and Video Panoptic Segmentation [20, 28], operate within a closed vocabulary, segmenting \"thing\" or \"stuff\" categories based on predefined labels throughout video sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.907, + 0.306 + ], + "angle": 0, + "content": "In contrast, class-agnostic tasks, such as Video Object Segmentation (VOS) [34, 46], rely on prompt-based guidance. Unlike image-level prompts, video prompts typically require initial mask input in the first frame for temporal propagation. Recent work, such as SAM 2 [39], demonstrates that large-scale pre-training can generalize VOS to zero-shot applications. However, its reliance on explicit prompts becomes inefficient when tracking multiple dense objects with frequent occlusions or reappearance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.908, + 0.595 + ], + "angle": 0, + "content": "Recent advances have expanded video segmentation to open-world and open-vocabulary paradigms. The BURST dataset [3] introduces annotations across 482 diverse categories with a focus on thing-level segmentation, while UVO [13] takes a class-agnostic open-world approach, treating all video objects as a single category without predefined labels. Entity Segmentation [26, 37] has shown promising generalization to novel categories at the image level, but extending this approach to video introduces significant challenges for zero-shot evaluation. To bridge this gap, we propose Video Entity Segmentation, a new task aimed at segmenting entities across video sequences without relying on predefined categories or explicit prompts. This approach extends image-level entity understanding to maintain temporal consistency across frames, enabling generalized handling of novel categories in a zero-shot setting. By avoiding rigid class definitions, our framework provides greater flexibility, particularly beneficial for real-world applications with unknown or fluid category boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Query-based Video Segmentation Query-based video segmentation models are widely used for dense prediction tasks, including video instance [6, 14, 17, 18, 42-44] and semantic segmentation [2, 21]. Inspired by detection transformers [5], these methods use queries to represent object information and apply cross-attention with image features to generate segmentation masks, demonstrating effectiveness in video instance segmentation. Temporal association is achieved through either direct query matching [17, 45] or additional tracking modules [15]. While effective in controlled scenarios, these query-based methods without explicit memory propagation struggle with long or complex video sequences due to challenges in modeling long-term temporal dependencies [36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Memory-based Video Segmentation Memory-based methods [9, 24, 31, 33, 47, 49, 50] are effective in Video Object Segmentation [4, 12, 46, 46], using initial ground truth masks as prompts. Memory features are stored in memory banks for segmentation in subsequent frames, excelling in long-term video tracking [8, 10]. SAM 2 [39] validates the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24235" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "generalizability of memory-based designs but processes objects separately, limiting efficiency and lacking inter-object communication. Recent works have attempted to integrate query-based and memory mechanisms, such as Cutie [10] for object-level VOS and UniVS [29] for multi-task inference. However, joint training of memory and query modules is computationally challenging for long videos. Our EntitySAM addresses these issues by leveraging SAM 2's frozen memory encoder and introducing a novel query-based entity decoder for efficient joint training, combining the strengths of both approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.182, + 0.279 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.485, + 0.516 + ], + "angle": 0, + "content": "We present EntitySAM for Video Entity Segmentation, enabling automatically segmenting and tracking all entities in the video sequence. By integrating query and memory-based video segmentation designs, EntitySAM handles multi-object tracking with inter-object communication and object-level contextual information. In Section 3.1, we first introduce the Video Entity Segmentation task, which offers improved generalization and reduced annotation complexity compared to other video segmentation tasks, such as Video Instance / Panoptic / Object Segmentation. In Section 3.2, we provide an overview of SAM 2's architecture and key characteristics. Section 3.3 details the architectural design of our EntitySAM based on SAM 2, highlighting our enhanced feature fusion and decoding mechanisms, along with comprehensive training and inference protocols." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.334, + 0.54 + ], + "angle": 0, + "content": "3.1. Video Entity Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.667 + ], + "angle": 0, + "content": "Video Entity Segmentation aims to segment and track all entities within video sequences using a class-agnostic approach. In this context, an \"entity\" encompasses both thing and stuff categories, representing any distinct group with coherent semantic meaning. The task requires continuous tracking and segmentation of these entities through temporal sequences, with annotations represented as tube masks that capture the spatial-temporal evolution of each entity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.803 + ], + "angle": 0, + "content": "Specifically, given an input video sequence \\(I \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\), both annotations and predictions consist of \\(N\\) non-overlapping masklets \\(M_{i_{i=1}}^{N}\\), where each masklet \\(M_{i} \\in [0,1]^T \\times [0,1]\\). For predictions, the task additionally requires a confidence score \\(s_{i} \\in [0,1]\\) associated with each masklet. Under the non-overlapping constraint, these \\(N\\) masklets can be merged into a single video sequence \\(M \\in \\mathbb{N}^{T \\times H \\times W}\\), where each pixel value represents the corresponding prediction ID." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.483, + 0.835 + ], + "angle": 0, + "content": "For evaluation, we first review traditional Video Panoptic Quality (VPQ), which computes metrics per category:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.841, + 0.483, + 0.879 + ], + "angle": 0, + "content": "\\[\nV P Q = \\frac {1}{| C |} \\sum_ {c \\in C} \\frac {\\sum_ {t} I o U _ {c , t}}{T P _ {c} + 0 . 5 \\times F P _ {c} + 0 . 5 \\times F N _ {c}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To better evaluate video entity segmentation, we propose" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.09, + 0.89, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of Different Video Segmentation Tasks" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.106, + 0.905, + 0.189 + ], + "angle": 0, + "content": "
TaskCategoryInference
ThingsStuffAutomatic PromptCross-DatasetZero-shot
Video Instance Seg.✓✓✓
Video Semantic Seg.✓✓✓
Video Object Seg.✓✓✓
Video Panoptic Seg.✓✓✓✓
Video Entity Seg.✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.193, + 0.905, + 0.224 + ], + "angle": 0, + "content": "Video Entity Quality (VEQ), a class-agnostic metric that assesses performance across temporal sequences:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.233, + 0.905, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V E Q} = \\frac {\\sum_ {t} I o U _ {t}}{T P + 0 . 5 \\times F P + 0 . 5 \\times F N} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.268, + 0.905, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V E Q} ^ {S Q} = \\quad \\frac {\\sum_ {t} I o U _ {t}}{T P} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.299, + 0.905, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V E Q} ^ {R Q} = \\frac {\\sum_ {t} T P}{T P + 0 . 5 \\times F P + 0 . 5 \\times F N} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.566 + ], + "angle": 0, + "content": "where \\( IoU_{t} \\) represents the temporal intersection over union across frames, measuring both spatial accuracy and temporal consistency. Here, \\( TP \\), \\( FP \\), and \\( FN \\) are computed globally across all entities without category differentiation. For consistency with Video Panoptic Segmentation evaluation, we also implement a class-agnostic variant of the STQ (Segmentation and Tracking Quality) metric, denoted as \\( \\mathrm{STQ}^{EN} \\), which emphasizes tracking consistency across frames. The formulation of VEQ provides a more direct assessment of video-level segmentation and tracking performance while eliminating the complexity of category-specific evaluation. Moreover, this class-agnostic approach reduces annotation requirements, enables large-scale training, and enhances model generalizability across datasets, particularly for open-vocabulary and zero-shot scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.793 + ], + "angle": 0, + "content": "As illustrated in Table 1, Video Entity Segmentation presents distinct advantages compared to existing video segmentation paradigms: Unlike Video Object Segmentation, it operates automatically and accommodates multiple dynamic objects. In contrast to Video Instance Segmentation, it extends beyond foreground objectness to include background entities. Compared to Video Semantic Segmentation, it provides instance-level annotations. Unlike Video Panoptic Segmentation, it adopts a class-agnostic approach that eliminates the constraint of pre-defined category vocabularies. This unified entity representation allows both thing and stuff categories to be learned jointly and subsequently classified using external Vision Language Models, enabling advanced open-vocabulary evaluation and reasoning-based segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.776, + 0.818 + ], + "angle": 0, + "content": "3.2. Method Background: SAM 2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "SAM 2 [39] extends the Segment Anything model [22] to encompass both image and video domains. Through an iterative process of data annotation and model training, SAM 2 leverages both manual and automated datasets generated by its data engine. For video segmentation functionality," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24236" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.072, + 0.905, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.242, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 2. (a) Overview of the EntitySAM framework. EntitySAM utilizes the frozen encoder and memory parameters from SAM 2, incorporating a dual encoder design for enhanced semantic features (Section 3.3.1). The PromptGenerator (Section 3.3.2) automatically generates prompts from Prompt Queries. The enhanced features and distinct query groups are processed by the EntityDecoder (Section 3.3.3) to produce video mask outputs. (b) Self-attention and cross-attention mechanisms in EntityDecoder layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.482, + 0.337 + ], + "angle": 0, + "content": "SAM 2 incorporates a classical memory mechanism to facilitate object tracking." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.483, + 0.534 + ], + "angle": 0, + "content": "The architecture processes video input through multiple stages: initially, a multi-scale Image encoder extracts frame-level features, which then undergo cross-attention with previous video-level memory features to derive per-mask temporal information. Concurrently, input prompts are processed through a dedicated prompt encoder to generate prompt encodings. These various features and prompts are integrated through cross-attention mechanisms and dynamic convolution operations within the mask decoder to generate mask outputs. The model maintains temporal coherence by encoding and storing mask sequences in a memory bank, which updates frame-by-frame and feeds into memory attention for processing subsequent frames." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Through extensive training on 35M manual and automated mask annotations, SAM 2 shows exceptional zero-shot performance across various video object segmentation benchmarks, including DAVIS [35], YTVOS [46], LVOS [16], and MOSE [12]. However, SAM 2 exhibits notable limitations: it requires initial mask/prompt specification at the sequence start and cannot effectively accommodate new objects introduced during tracking. Moreover, the model's efficiency deteriorates when handling large numbers of concurrent objects, indicating challenges in scaling to more complex scenarios with multiple dynamic objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.216, + 0.732 + ], + "angle": 0, + "content": "3.3. EntitySAM" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Coupled with the proposed Video Entity Segmentation task, we present EntitySAM, a novel model that achieves strong zero-shot video entity segmentation performance while requiring only image-level class-agnostic entity annotations. As illustrated in Figure 2, our model preserves the memory-based architecture of the original SAM by freezing both the pretrained image encoder and memory-related modules, thus retaining its advantages in long-term tracking of objects. Meanwhile, EntitySAM comprises three learnable components: Dual Visual Encoder, EntityPromptGenerator," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.337 + ], + "angle": 0, + "content": "and EntityDecoder. The following sections will detail these architectural components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.714, + 0.358 + ], + "angle": 0, + "content": "3.3.1 Dual Visual Encoder" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.37, + 0.907, + 0.505 + ], + "angle": 0, + "content": "Our EntitySAM leverages the frozen Segment Anything Model 2 (SAM 2) image encoder as its backbone, as well as its frozen memory encoder and memory attention layers. The frozen design preserves SAM 2's zero-shot feature extraction capabilities, which were developed through training on extensive video object tracking datasets. Additionally, freezing the memory encoder substantially reduces GPU memory requirements during the training of multiple dense entity tracking in video sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.672 + ], + "angle": 0, + "content": "Although SAM 2 exhibits impressive performance across multi-granularity class-agnostic segmentation tasks, we reveal its limitations in maintaining consistent entity-level representations as in Table 9. Specifically, SAM 2 struggles to generate reliable confidence scores for entity classification, where an 'entity' refers to a semantically coherent group. Thus in addition to the original SAM 2's backbone feature, we integrate the backbone features from DINOv2 [32], which produces consistent feature representations for semantically similar regions by large-scale self-supervised training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.793 + ], + "angle": 0, + "content": "As illustrated in Figure 2, EntitySAM incorporates DINOv2 through a linear projection layer, aligning its feature dimensionality with that of SAM 2. The aligned features are then concatenated and fed into the decoder to form a fused representation. This enhances the encoded visual representation by preserving the zero-shot capabilities for video entity segmentation while effectively mitigating SAM 2's limitations in entity-level understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.776, + 0.816 + ], + "angle": 0, + "content": "3.3.2 Automatic Prompt Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "In SAM 2's architecture, the prompt encoder plays a crucial role in image-level segmentation, where input prompts, such as points and bounding boxes, are processed to generate prompt embeddings for decoder segmentation. However, in video-level operations, SAM 2 bypasses the prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24237" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "encoder, focusing solely on object tracking using historical memory features. To extend the model's capability for new entity tracking in video sequences, EntitySAM incorporates prompt embeddings during video training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.291 + ], + "angle": 0, + "content": "Our EntitySAM implements a prompt generator utilizing learnable queries that perform cross-attention with enhanced visual features to predict input prompts for the prompt encoder. The resulting prompt embeddings are subsequently fed into the decoder at each frame to localize both existing and novel entities within the video sequence. Concurrently, this module predicts an entity confidence score, denoted as \\( S(\\text{entity}) \\), which performs binary classification between foreground entities and background regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.459 + ], + "angle": 0, + "content": "More formally, let \\(\\mathcal{Q} \\in \\mathbb{R}^{N \\times D}\\) represent the learnable queries and \\(\\mathcal{F} \\in \\mathbb{R}^{HW \\times D}\\) denote the enhanced visual features, where \\(N\\) is the number of queries, \\(D\\) is the feature dimension, and \\(H, W\\) are spatial dimensions. The prompt generation and scoring process can be expressed as: \\(\\mathcal{A} = \\text{CrossAttention}(\\mathcal{Q}, \\mathcal{F})\\), \\(\\mathcal{P} = \\text{FC}_{\\text{prompt}}(A)\\), and \\(\\mathcal{S} = \\text{FC}_{\\text{score}}(A)\\), where \\(A \\in \\mathbb{R}^{N \\times D}\\) is the cross-attention output, \\(\\mathcal{P} \\in \\mathbb{R}^{N \\times D'}\\) represents the generated prompts with \\(D'\\) matching the expected prompt dimension required by the prompt encoder, and \\(\\mathcal{S} \\in \\mathbb{R}^{N \\times 1}\\) denotes the entity confidence scores." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.479, + 0.295, + 0.493 + ], + "angle": 0, + "content": "3.3.3 Entity Mask Decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.734 + ], + "angle": 0, + "content": "The Entity Mask Decoder predicts output masks based on enhanced features in Section 3.3.1 and the learnable prompts through a multi-stage processing pipeline in Section 3.3.2. It employs two groups of batched queries: mask queries \\(\\mathcal{Q}_m\\in \\mathbb{R}^{N\\times 1\\times C}\\) and IoU queries \\(\\mathcal{Q}_{iou}\\in \\mathbb{R}^{N\\times 1\\times C}\\), where \\(N\\) denotes the number of entities and \\(C\\) represents the feature dimension. These queries are concatenated with input prompts for self-attention to facilitate inter-entity communication. These queries are subsequently reshaped for cross-attention operations between token and image representations. In the final layer, mask queries undergo dynamic convolution with upsampled mask features to generate output masks, while IoU queries are processed through a dedicated prediction head to estimate conditioned IoU scores. The computational flow can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.749, + 0.482, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} [ \\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime} ] = \\text {S e l f A t t e n t i o n} (\\text {c o n c a t} (\\mathcal {Q} _ {m}; \\mathcal {Q} _ {i o u})) \\\\ w h e r e \\operatorname {c o n c a t} \\left(\\mathcal {Q} _ {m}; \\mathcal {Q} _ {i o u}\\right) \\in \\mathbb {R} ^ {1 \\times 2 N \\times C} \\\\ F _ {o u t} = \\operatorname {C r o s s A t t e n t i o n} (\\operatorname {r e s h a p e} \\left(\\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime}\\right), F) \\\\ \\operatorname {w h e r e} \\operatorname {r e s h a p e} \\left(\\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime}\\right) \\in \\mathbb {R} ^ {N \\times 2 \\times C}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The supervision mechanism integrates outputs from both the prompt generator (entity scores) and mask decoder (masks and conditioned IoU predictions) through Hungarian matching with ground truth annotations. The expected" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.757, + 0.107 + ], + "angle": 0, + "content": "IoU for query \\( q \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.115, + 0.905, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E (\\mathrm {I o U} _ {q}) = E (\\mathrm {I o U} _ {q} | \\mathrm {E n t} _ {q}) S (\\mathrm {E n t} _ {q}) + E (\\mathrm {I o U} _ {q} | \\neg \\mathrm {E n t} _ {q}) S (\\neg \\mathrm {E n t} _ {q}) \\\\ = E \\left(\\operatorname {I o U} _ {q} \\mid \\operatorname {E n t} _ {q}\\right) S \\left(\\operatorname {E n t} _ {q}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.162, + 0.905, + 0.191 + ], + "angle": 0, + "content": "where \\( S(\\mathrm{Ent}_q) \\) represents the entity confidence score from PromptGenerator in Section 3.3.2 and \\( E(\\mathrm{IoU}_q|\\neg \\mathrm{Ent}_q) = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.327 + ], + "angle": 0, + "content": "The resulting output masks are processed through the frozen memory encoder, memory bank, and memory attention mechanisms, with gradient computation detached for memory-related components. This enables efficient training of the queries-based tracking component while maintaining frozen memory modules. Consequently, our approach achieves scalable training for over 50 entities in 8-frame video clips, significantly surpassing SAM 2's original capacity of handling only 3 video objects during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.892, + 0.361 + ], + "angle": 0, + "content": "3.3.4 Training Loss and Open-Vocabulary Inference" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.369, + 0.905, + 0.397 + ], + "angle": 0, + "content": "We employ an end-to-end training objective for EntitySAM, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.41, + 0.905, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {M a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\text {E n t i t y}} + \\lambda_ {2} \\mathcal {L} _ {\\text {I o U}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.905, + 0.497 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) supervises both mask prediction and binary entity classification, following the Mask2former framework [7]. The hyperparameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) balance the respective loss terms and are empirically set to 1.0 and 1.0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.694 + ], + "angle": 0, + "content": "During inference, EntitySAM accommodates both class-agnostic and class-specific evaluation protocols. For class-agnostic scenarios, entity masks are directly generated and merged based on confidence scores and mask overlap metrics. In class-specific evaluation, we select the frame with the highest entity confidence score for each masklet prediction. The predicted entity masks are utilized to crop corresponding regions of interest, which, together with carefully designed prompts containing candidate category names, are processed through GPT-4o [1] for final classification. We demonstrate in the supplementary materials that this mask-guided cropping approach achieves superior classification performance compared to alternative methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.646, + 0.719 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.704, + 0.743 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Datasets Video entity segmentation typically requires extensive class-agnostic dense entity video segmentation annotations. Although it eliminates the need for class-specific annotations, creating frame-by-frame annotations remains labor-intensive. Inspired by [51, 52], our training pipeline leverages COCO [23] panoptic segmentation data without category label annotation (denoted as \\(\\mathrm{COCO}^*\\)) and extends it to pseudo-video sequences. Our approach utilizes only image-level annotations, eschewing video-level labels, to achieve effective video entity segmentation. The pipeline" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24238" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "also enables zero-shot cross-dataset evaluation across multiple benchmarks. We demonstrate our model's versatility through comprehensive zero-shot evaluations on various tasks: video panoptic segmentation on VIPSeg [28], video semantic segmentation on VSPW [27], video instance segmentation on Youtube-VIS 2019 [48], Youtube-VIS 2021 [48], and OVIS [36], and video entity segmentation on our proposed VIPEntitySeg dataset—a class-agnostic variant of VIPSeg. Notably, all zero-shot evaluations are performed using a single trained model, demonstrating its robust zero-shot generalization capabilities across different video segmentation tasks and datasets." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.274, + 0.487, + 0.593 + ], + "angle": 0, + "content": "Implementation details In our experimental setup, we employ two backbone architectures: ViT-Small and ViT-Large. The models are trained using the AdamW optimizer with a learning rate of \\(5 \\times 10^{-5}\\) and weight decay of 0.05. Input images are cropped to \\(1024 \\times 1024\\) pixels for compatibility with SAM 2 ViT/16 and \\(896 \\times 896\\) pixels for DINOv2 ViT/14. Training is conducted with a batch size of 8. The loss function combines multiple components: cross-entropy and dice losses for mask prediction, L1 loss for IoU prediction, and focal loss for binary entity classification. Our training strategy consists of two phases: initial training on static COCO images for 40K iterations, followed by training on COCO pseudo-video sequences with clip length 8 for 10K iterations. Both stages are class-agnostic. During inference, we utilize a single entity category output mode with confidence scoring. Video entity quality (VEQ) is evaluated in a class-agnostic manner. For class-specific evaluation tasks, we implement a multi-step process: first selecting the highest-confidence mask from the masklet for each prediction, then extracting the corresponding frame region within the mask, and finally employing GPT-4o for classification." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.303, + 0.621 + ], + "angle": 0, + "content": "4.2. Zero-shot Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Zero-shot Video Entity Segmentation In Table 2 (left), we evaluate our approach on VIPEntitySeg using class-agnostic metrics against several baseline methods. For SAM 2, we adopt two prompting manners, a grid point prompt sampling initialization and a COCO-trained Mask2Former [7] mask prompts initialization. We additionally compare against state-of-the-art zero-shot segmentation models, including: 1) FC-CLIP [51] integrated with MinVIS [17] association, 2) DEVA [11] coupled with COCO-trained Mask2Former [7] for per-frame segmentation, and 3) OV-DVIS++ [52] in both online and offline modes. All these SOTA models were trained on COCO with category annotations, with DEVA utilizing additional external datasets for its association model. We compare with these methods on our class-agnostic VIPEntitySeg benchmark. Our EntitySAM achieves significant performance improvements over all baselines and SOTA models with both ViT-Small and ViT-Large backbones, despite training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "without category annotations. This superior performance demonstrates the effectiveness of our class-agnostic training approach and proposed modules. Notably, we observe substantial improvements in \\(\\mathrm{VEQ}^{RQ}\\), indicating enhanced cross-frame entity identification and reduced false positives. Figure 3 further illustrates our model's robustness by evaluating VEQ across varying frame intervals, demonstrating that EntitySAM maintains consistent performance over extended temporal sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.366 + ], + "angle": 0, + "content": "Regarding the class-agnostic \\(\\mathrm{STQ}^{EN}\\) metric, which emphasizes tracking performance, DEVA achieves higher scores through its tracking association model trained on external video object segmentation datasets. However, DEVA relies on a two-stage pipeline of detection followed by association, lacking end-to-end integration. While our approach achieves lower scores, it delivers comparable performance through a streamlined process that requires only COCO training data and single-pass inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.61 + ], + "angle": 0, + "content": "Zero-shot Video Panoptic Segmentation In Table 2 (right), we extend our evaluation to VIPSeg for class-specific video panoptic segmentation. Several baseline models, including Mask2Former and DEVA, are constrained to VIPSeg evaluation due to their closed-vocabulary nature. When compared to open-vocabulary models such as FC-CLIP and OV-DVIS++, our approach demonstrates substantial improvements, particularly with the ViT-Small backbone. These results validate our approach of \"post-classification with multimodal models\", indicating that it surpasses traditional open-vocabulary segmentation training pipelines despite operating without category annotations. Our EntitySAM achieves competitive performance on both class-agnostic and class-specific tasks, suggesting that category annotations may not be essential for effective open-vocabulary video segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.905, + 0.733 + ], + "angle": 0, + "content": "Zero-shot Video Semantic Segmentation Table 3 presents our results on video semantic segmentation, obtained by merging per-pixel mask predictions. Our method achieves improvements in both mean Intersection over Union (mIoU) and Video Consistency (VC) metrics. The VSPW evaluation dataset, which contains the same video content but with extended frame sequences, validates our model's capability for effective long-term tracking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Zero-shot Video Instance Segmentation As shown in Table 4, we evaluate EntitySAM on three popular video instance segmentation datasets. Our approach on extending video entity segmentation to video instance segmentation requires no model retraining—we simply add a background/stuff class during inference and filter out predictions with that class to obtain instance results. Our model demonstrates significant performance improvements over state-of-the-art open-vocabulary video instance segmentation models. In particular, on the challenging OVIS dataset, which features frequent occlusions, our model maintains robust" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24239" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.071, + 0.907, + 0.128 + ], + "angle": 0, + "content": "Table 2. Comparison of Zero-shot Video Entity Segmentation (left) and Video Panoptic Segmentation (right) on COCO \\(\\rightarrow\\) VIPSeg evaluation. EntitySAM is trained on the COCO dataset without category annotations. In VIPSeg, - means the results of open-vocabulary evaluation for closed-vocabulary models are not available. * denotes that DEVA is trained with multiple external video object segmentation datasets for the temporal association model." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.13, + 0.907, + 0.349 + ], + "angle": 0, + "content": "
MethodBackboneClass Agnostic TrainingVIPEntitySegVIPSeg
VEQVEQSQVEQRQSTQENVPQVPQThVPQStSTQ
SAM 2.1 [39] (Everything Init)ViT-S✓26.982.532.638.7-
SAM 2.1 [39] (Mask2former Init)ViT-S + ResNet5046.283.655.234.5-
DEVA* [11]ResNet-5049.083.658.642.4-
FC-CLIP [51]ResNet-5041.584.049.429.722.325.519.119.7
OV-DVIS++(Online) [52]ResNet-5045.983.754.833.624.426.822.422.0
OV-DVIS++(Offline) [52]ResNet-5046.782.856.335.623.826.421.424.4
OursViT-S✓51.884.761.141.028.732.925.131.4
SAM 2.1 [39] (Everything Init)ViT-L✓30.583.136.741.7-
SAM 2.1 [39] (Mask2former Init)ViT-L + Swin-L50.684.460.039.1-
DEVA* [11]Swin-L50.883.860.645.3-
FC-CLIP [51]ConvNext-L43.484.851.232.427.930.925.024.2
OV-DVIS++(Online) [52]ConvNext-L49.483.559.237.928.931.326.828.4
OV-DVIS++(Offline) [52]ConvNext-L49.083.159.040.030.431.929.132.2
OursViT-L✓54.684.764.543.331.437.526.033.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.354, + 0.483, + 0.396 + ], + "angle": 0, + "content": "Table 3. Zero-Shot Open-vocabulary Video Semantic Segmentation on COCO \\(\\rightarrow\\) VSPW. EntitySAM is trained on the COCO dataset without category annotations." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.398, + 0.483, + 0.511 + ], + "angle": 0, + "content": "
MethodBackboneClass Agnostic TrainingmVC8VSPW mVC16mIOU
FC-CLIP [51]ResNet-5084.982.724.3
OV-DVIS++(Online) [52]ResNet-5092.791.527.6
OV-DVIS++(Offline) [52]ResNet-5092.491.328.4
OursViT-S✓94.093.034.6
FC-CLIP [51]ConvNext-L89.988.428.9
OV-DVIS++(Online) [52]ConvNext-L94.293.334.3
OV-DVIS++(Offline) [52]ConvNext-L93.993.034.1
OursViT-L✓94.693.735.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.483, + 0.557 + ], + "angle": 0, + "content": "performance, benefiting from the entity learning strategy employed during training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.297, + 0.577 + ], + "angle": 0, + "content": "4.3. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.644 + ], + "angle": 0, + "content": "We perform ablation studies using the ViT-Small backbone on the COCO \\(\\rightarrow\\) VIPEntitySeg video entity segmentation task with zero-shot protocol to evaluate the effectiveness of our proposed modules." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Ablation on the Proposed Modules Table 5 presents ablation studies of our proposed modules. Our baseline, SAM 2 with memory-based tracking modules, cannot generate predictions automatically without appropriate prompting. Using grid-point sampling leads to numerous false positive predictions, achieving a VEQ of only 26.9. Removing the memory design and implementing solely a query-based EntityDecoder with query association for tracking improves performance to 41.3, though this remains suboptimal. Combining memory-based and query-based approaches significantly enhances performance to 50.2, surpassing both DEVA and OV-DVIS++. Finally, incorporating a complementary DINOv2 dual feature encoder achieves the best result (51.8) for the ViT-Small backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation on Query Design As illustrated in Figure 2, our design incorporates three query types: two groups in EntityDecoder and one group in PromptGenerator. Our abla" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.354, + 0.695, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.354, + 0.904, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.473, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Figure 3. VEQ comparison with different evaluation time intervals. EntitySAM improves over long intervals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.669 + ], + "angle": 0, + "content": "tion experiment of query design in Table 6 reveals that using a single query group to predict all outputs (mask, IoU, and confidence score) yields suboptimal performance with a VEQ of 49.2. Adding a second group of prompt queries improves performance to 49.7. Further separation into a third query group for distinct mask and IoU prediction achieves the best performance of 51.8. These results suggest that different query groups learn specialized distributions for their respective prediction tasks, departing from traditional detection transformers' shared-query architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.761 + ], + "angle": 0, + "content": "Ablation on EntityDecoder Depth Table 7 studies the influence of the entity decoder's depth layer. For the ViT-Small backbone, a decoder depth of 4 achieves optimal performance with a VEQ of 51.8 on COCO \\(\\rightarrow\\) VIPSeg setting. Further increasing decoder depth shows no additional performance benefits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ablation on Training Stages As detailed in Section 4.1, our training process consists of two stages: initial training on single images (treated as zero-length videos) followed by training on pseudo video clips of length 8. During the first stage, the memory component remains unused while the EntityDecoder learns to identify new entities in each image. Table 8 shows that this image-only training achieves limited video entity segmentation performance with a VEQ of 31.4. However, this stage exposes the model to diverse en" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24240" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.08, + 0.908, + 0.109 + ], + "angle": 0, + "content": "Table 4. Zero-Shot Open-vocabulary Video Instance Segmentation on COCO \\(\\rightarrow\\) Validation sets of YouTube-VIS 2019, 2021, and OVIS. EntitySAM is trained on the COCO dataset without category annotations." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.112, + 0.907, + 0.206 + ], + "angle": 0, + "content": "
MethodBackboneClass Agnostic TrainingYoutube-VIS 2019Youtube-VIS 2021OVIS
AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)
Detric[53]-OWTB[25]ResNet-5017.9----16.7----9.0----
MindVLT[41]ResNet-5023.1----20.9----11.4----
FC-CLIP [51]ResNet-5028.943.929.932.441.225.540.026.828.037.011.825.110.58.516.4
OV-DVIS++(Online) [52]ResNet-5034.550.639.239.549.530.946.734.834.445.814.831.213.110.524.7
OV-DVIS++(Offline) [52]ResNet-5034.451.839.240.050.631.048.534.334.646.813.028.410.99.825.1
OursViT-S✓44.960.749.543.662.438.852.543.136.256.521.035.820.812.336.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.211, + 0.438, + 0.225 + ], + "angle": 0, + "content": "Table 5. Ablation Study on the Proposed Modules." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.228, + 0.482, + 0.324 + ], + "angle": 0, + "content": "
ModelMemoryEntityDecoderDual Visual EncoderVEQ
SAM 2 [39]✓26.9
Ours✓✓41.3
✓✓50.2
✓✓✓51.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.327, + 0.413, + 0.342 + ], + "angle": 0, + "content": "Table 6. Ablation Study on Query Design." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.422 + ], + "angle": 0, + "content": "
EntityDecoderPromptGeneratorCOCO → VIPSeq
VEQVEQSQVEQRQ
Shared Queries49.284.058.6
Shared Queries✓49.784.259.0
Separate Queries✓51.884.761.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.435, + 0.435, + 0.449 + ], + "angle": 0, + "content": "Table 7. Ablation Study on EntityDecoder Depth." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.451, + 0.482, + 0.549 + ], + "angle": 0, + "content": "
DecoderDecoder depthCOCO → VIPSeg
VEQVEQSQVEQRQ
EntityDecoder251.384.760.5
451.884.761.1
651.684.661.0
851.184.460.6
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.642 + ], + "angle": 0, + "content": "tities, and when combined with the subsequent video training stage—which aligns EntityDecoder with memory encoders—performance improves substantially to 51.8 VEQ. This two-stage approach outperforms training exclusively on pseudo video sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation on Dual Encoder In Section 3.3.1, we introduced encoder feature fusion designs for dual encoders. As shown in Table 9, our ablation studies reveal that incorporating features from either CLIP [38] or SAM [22] alongside the SAM 2 encoder features yields no improvement in zero-shot performance. This can be attributed to CLIP's vision encoder being optimized for classification rather than segmentation tasks, resulting in representations less suitable for segmentation. Similarly, SAM, being trained with an approach analogous to SAM 2, provides features that largely overlap with SAM 2's capabilities rather than offering complementary information. In contrast, DINOv2, trained through self-supervision, demonstrates significant benefits when combined with SAM 2. Its self-supervised training enables DINOv2 to learn consistent image representations that effectively complement SAM 2's supervised features, ultimately enhancing the overall performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.211, + 0.84, + 0.225 + ], + "angle": 0, + "content": "Table 8. Ablation Study on Training stages." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.228, + 0.905, + 0.312 + ], + "angle": 0, + "content": "
ModelImagePseudo VideoCOCO → VIPSeg
VEQVEQSQVEQRQ
Ours✓31.479.439.5
✓49.284.258.4
✓✓51.884.761.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.317, + 0.859, + 0.332 + ], + "angle": 0, + "content": "Table 9. Ablation Study on Dual Encoder Feature." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.334, + 0.905, + 0.431 + ], + "angle": 0, + "content": "
Encoder FeatureDual EncoderCOCO → VIPSeg
VEQVEQSQVEQRQ
SAM 2-50.284.559.4
SAM [22]47.684.256.5
CLIP [38]49.284.058.6
DINOv2 [32]51.884.761.1
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.447, + 0.869, + 0.464 + ], + "angle": 0, + "content": "4.4. Qualitative experiments and visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.532 + ], + "angle": 0, + "content": "Figure 4 demonstrates our model's performance on zero-shot video sequences, showcasing consistent entity tracking and segmentation across frames. Additional qualitative results are provided in the supplementary materials." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.538, + 0.646, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.538, + 0.776, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.538, + 0.905, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.654, + 0.895, + 0.668 + ], + "angle": 0, + "content": "Figure 4. Visualization of zero-shot video entity segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.9 + ], + "angle": 0, + "content": "We present EntitySAM, a framework that integrates query and memory mechanisms to enhance SAM 2 for zero-shot video entity segmentation. Our approach eliminates the need for user-defined prompt inputs and enables automatic tracking of new entities across frames. Trained exclusively on COCO image-level mask annotations without category information, EntitySAM achieves superior performance on the Video Entity Segmentation task. The model demonstrates strong generalization capabilities across various zero-shot evaluations, including Video Panoptic / Semantic / Instance Segmentation. These results provide valuable insights into class-agnostic training approaches and their effectiveness in zero-shot scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24241" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. In arXiv preprint arXiv:1905.00737, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Anwesa Choudhuri andd Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. In https://arxiv.org/abs/2112.10764, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[8] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.615, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In arXiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[13] Yuming Du, Wen Guo, Yang Xiao, and Vincent Lepetit. Uvo challenge on video-based open-world segmentation 2021: 1st place solution. ICCV Workshop, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. Advances in Neural Information Processing Systems, 35:23109–23120, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[15] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14623-14632, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[16] Lingyi Hong, Wenchao Chen, Zhongying Liu, Wei Zhang, Pinxue Guo, Zhaoyu Chen, and Wenqiang Zhang. Lvos: A benchmark for long-term video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13480-13492, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[17] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[18] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[20] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[21] Dahun Kim, Jun Xie, Huiyu Wang, Siyuan Qiao, Qihang Yu, Hong-Seok Kim, Hartwig Adam, In So Kweon, and Liang-Chieh Chen. Tubeformer-deeplab: Video mask transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13914-13924, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[24] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.609 + ], + "angle": 0, + "content": "[25] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[26] Qi Lu, Jason Kuen, Shen Tiancheng, Gu Jiumiang, Guo Weidong, Jia Jiaya, Lin Zhe, and Yang Ming-Hsuan. High-quality entity segmentation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[27] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. Vspw: A large-scale dataset for video scene parsing in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4133-4143, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[28] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[29] Xindong Zhang Minghan Li, Shuai Li and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[30] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6819-6828, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24242" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[32] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[33] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[34] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[35] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[36] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation. IJCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[37] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. In arXiv preprint arXiv:2107.14228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[40] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3126-3137, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In proceedings of the IEEE/CVF international conference on computer vision, pages 4057-4066, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In European Conference on Computer Vision, pages 553-569. Springer, 2022." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[44] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[45] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In European Conference on Computer Vision, pages 588-605. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[46] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. In ECCV, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[47] Xiaohao Xu, Jinglu Wang, Xiao Li, and Yan Lu. Reliable propagation-correction modulation for video object segmentation. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[48] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[49] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[50] Zongxin Yang, Jiaxu Miao, Yunchao Wei, Wenguan Wang, Xiaohan Wang, and Yi Yang. Scalable video object segmentation with identification mechanism. TPAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[51] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[52] Tao Zhang, Xingye Tian, Yikang Zhou, Shunping Ji, Xuebo Wang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, and Yu Wu. Dvis++: Improved decoupled framework for universal video segmentation. arXiv preprint arXiv:2312.13305, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[53] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24243" + } + ] +] \ No newline at end of file diff --git a/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_origin.pdf b/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f021bd4e0c2a5414082e02470c6fa3fb01cc8721 --- /dev/null +++ b/2025/EntitySAM_ Segment Everything in Video/cb4f400f-a69b-43ea-b233-f7f126e62f4a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:258a2123fc6a0b6cd5b23d52b24391d836f8d7aec873d416031ba9b7d3165e68 +size 5855710 diff --git a/2025/EntitySAM_ Segment Everything in Video/full.md b/2025/EntitySAM_ Segment Everything in Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fe4ebbadcc84b7b51218040d53d81f9d5ddc8f0b --- /dev/null +++ b/2025/EntitySAM_ Segment Everything in Video/full.md @@ -0,0 +1,307 @@ +# EntitySAM: Segment Everything in Video + +Mingqiao Ye $^{1,*}$ Seoung Wug Oh $^{2}$ Lei Ke $^{3}$ Joon-Young Lee $^{2}$ $^{1}$ EPFL 2Adobe Research 3Carnegie Mellon University + +![](images/72cda20c7485f39168bcc495c67ad7fc72aa5b31e7be012a612f4209f8f75d12.jpg) +Figure 1. Zero-shot video entity segmentation performance comparison on VIPEntySeg dataset using models trained on COCO, showing: 1) SAM 2 [39] using Mask2Former [7] mask prompts for the initial frame, 2) Mask2Former with DEVA [11] association, and 3) our proposed EntitySAM. Our EntitySAM enhances SAM 2 by automatically segmenting and tracking novel entities without requiring user-specified prompts, achieving superior performance compared to existing state-of-the-art zero-shot tracking methods. + +# Abstract + +Automatically tracking and segmenting every video entity remains a significant challenge. Despite rapid advancements in video segmentation, even state-of-the-art models like SAM 2 struggle to consistently track all entities across a video—a task we refer to as Video Entity Segmentation. We propose EntitySAM, a framework for zero-shot video entity segmentation. EntitySAM extends SAM 2 by removing the need for explicit prompts, allowing automatic discovery and tracking of all entities, including those appearing in later frames. We incorporate query-based entity discovery and association into SAM 2, inspired by transformer-based object detectors. Specifically, we introduce an entity decoder to facilitate inter-object communication and an automatic prompt generator using learnable object queries. Additionally, we add a semantic encoder to enhance SAM 2's semantic awareness, improving segmentation quality. Trained on image-level mask annotations without category information from the COCO dataset, EntitySAM demonstrates strong generalization on four zero-shot video segmentation tasks: Video Entity, Panoptic, Instance, and Semantic Segmentation. Results on six popular benchmarks show that + +EntitySAM outperforms previous unified video segmentation methods and strong baselines, setting new standards for zero-shot video segmentation. Our code and models are at github.com/ymq2017-entitysam. + +# 1. Introduction + +Video segmentation is an essential task for various applications in video scene understanding, including video editing, AR/VR, and robotic perception. There are roughly two lines of video segmentation approaches. One line of approaches focuses on selecting semantic thing/stuff from predefined sets [27, 28, 48]. While these methods show strong performance in the benchmark, the predictions are limited to a pre-defined class set and the training data domain. + +Another line of approaches works by initialing the information about the target object. The most common approach in this group is semi-supervised video object segmentation (VOS) which tracks a given mask in the first frame. The recent SAM 2 [39] also falls into this group. SAM 2 generalizes the VOS task by taking multiple prompt types, including masks, boxes, and clicks similar to image segmentation methods [19, 22], and enables interactive video segmentation. Trained on large-scale training data, SAM 2 shows + +strong zero-shot broad-domain selection capability. + +Despite recent progress, a significant challenge remains in selecting and tracking all entities from a video—a task we term Video Entity Segmentation, which aims to segment any distinct group with coherent semantic meaning. While SAM 2 is an effective model for tracking each individual entity, we identified two limitations in selecting all entities: (1) It initializes tracking from user prompts so it cannot track unmarked entities and new entities appearing in different frames, as shown in Figure 1. (2) It tracks each entity independently which is not only inefficient and can lead to less optimal results (e.g., mask overlapping). + +In this paper, we propose EntitySAM, an extension of SAM 2 for zero-shot video entity segmentation. EntitySAM retains the zero-shot potential of SAM 2 while eliminating the need for explicit prompts for each entity and enabling seamless integration of new entities during tracking. Inspired by transformer-based object detectors [5, 6, 42], we sophisticatedly integrated query-based object discovery and association into SAM 2 making minimal changes to the original structure. To be specific, we replace the original mask decoder with a new entity decoder and introduce an automatic prompt generator that generates SAM 2 consumable prompts from learnable object queries. In this way, we can preserve most of the pre-trained SAM 2 weights. The proposed entity decoder differs from the original SAM 2 mask decoder in that it enables inter-entity communication to predict multiple entity masks without overlaps. Another notable addition to SAM 2 is a semantic encoder. We empirically find that SAM 2 encoder has a weak semantic awareness and supplementing semantic information using an external feature encoder greatly boosts the performance. + +We train EntitySAM on the COCO dataset [23] without using any category annotation and evaluate it on various zero-shot video segmentation tasks. First we introduce a video entity segmentation task. Video entity segmentation is class-agnostic, focusing on the segmentation and tracking of "entities" rather than predefined object categories. For this, we re-purpose the video panoptic segmentation dataset [28] by removing category annotation. In addition to this, we evaluate our method on multiple video instance and semantic segmentation benchmarks in zero-shot setting. Experimental results show that EntitySAM greatly outperforms previous unified video segmentation models [52] and strong baselines that combine an image detector [7] and a mask propagator [11]. Moreover, our method can be easily adapted to class-specific video panoptic segmentation by integrating state-of-the-art vision-language models [1]. + +# 2. Related Works + +Taxonomy of Video Segmentation Tasks Video segmentation tasks can be broadly categorized into class-specific and class-agnostic paradigms. Class-specific tasks, + +such as Video Instance Segmentation [36, 42, 48], Video Semantic Segmentation [30, 40], and Video Panoptic Segmentation [20, 28], operate within a closed vocabulary, segmenting "thing" or "stuff" categories based on predefined labels throughout video sequences. + +In contrast, class-agnostic tasks, such as Video Object Segmentation (VOS) [34, 46], rely on prompt-based guidance. Unlike image-level prompts, video prompts typically require initial mask input in the first frame for temporal propagation. Recent work, such as SAM 2 [39], demonstrates that large-scale pre-training can generalize VOS to zero-shot applications. However, its reliance on explicit prompts becomes inefficient when tracking multiple dense objects with frequent occlusions or reappearance. + +Recent advances have expanded video segmentation to open-world and open-vocabulary paradigms. The BURST dataset [3] introduces annotations across 482 diverse categories with a focus on thing-level segmentation, while UVO [13] takes a class-agnostic open-world approach, treating all video objects as a single category without predefined labels. Entity Segmentation [26, 37] has shown promising generalization to novel categories at the image level, but extending this approach to video introduces significant challenges for zero-shot evaluation. To bridge this gap, we propose Video Entity Segmentation, a new task aimed at segmenting entities across video sequences without relying on predefined categories or explicit prompts. This approach extends image-level entity understanding to maintain temporal consistency across frames, enabling generalized handling of novel categories in a zero-shot setting. By avoiding rigid class definitions, our framework provides greater flexibility, particularly beneficial for real-world applications with unknown or fluid category boundaries. + +Query-based Video Segmentation Query-based video segmentation models are widely used for dense prediction tasks, including video instance [6, 14, 17, 18, 42-44] and semantic segmentation [2, 21]. Inspired by detection transformers [5], these methods use queries to represent object information and apply cross-attention with image features to generate segmentation masks, demonstrating effectiveness in video instance segmentation. Temporal association is achieved through either direct query matching [17, 45] or additional tracking modules [15]. While effective in controlled scenarios, these query-based methods without explicit memory propagation struggle with long or complex video sequences due to challenges in modeling long-term temporal dependencies [36]. + +Memory-based Video Segmentation Memory-based methods [9, 24, 31, 33, 47, 49, 50] are effective in Video Object Segmentation [4, 12, 46, 46], using initial ground truth masks as prompts. Memory features are stored in memory banks for segmentation in subsequent frames, excelling in long-term video tracking [8, 10]. SAM 2 [39] validates the + +generalizability of memory-based designs but processes objects separately, limiting efficiency and lacking inter-object communication. Recent works have attempted to integrate query-based and memory mechanisms, such as Cutie [10] for object-level VOS and UniVS [29] for multi-task inference. However, joint training of memory and query modules is computationally challenging for long videos. Our EntitySAM addresses these issues by leveraging SAM 2's frozen memory encoder and introducing a novel query-based entity decoder for efficient joint training, combining the strengths of both approaches. + +# 3. Method + +We present EntitySAM for Video Entity Segmentation, enabling automatically segmenting and tracking all entities in the video sequence. By integrating query and memory-based video segmentation designs, EntitySAM handles multi-object tracking with inter-object communication and object-level contextual information. In Section 3.1, we first introduce the Video Entity Segmentation task, which offers improved generalization and reduced annotation complexity compared to other video segmentation tasks, such as Video Instance / Panoptic / Object Segmentation. In Section 3.2, we provide an overview of SAM 2's architecture and key characteristics. Section 3.3 details the architectural design of our EntitySAM based on SAM 2, highlighting our enhanced feature fusion and decoding mechanisms, along with comprehensive training and inference protocols. + +# 3.1. Video Entity Segmentation + +Video Entity Segmentation aims to segment and track all entities within video sequences using a class-agnostic approach. In this context, an "entity" encompasses both thing and stuff categories, representing any distinct group with coherent semantic meaning. The task requires continuous tracking and segmentation of these entities through temporal sequences, with annotations represented as tube masks that capture the spatial-temporal evolution of each entity. + +Specifically, given an input video sequence $I \in \mathbb{R}^{T \times H \times W \times 3}$ , both annotations and predictions consist of $N$ non-overlapping masklets $M_{i_{i=1}}^{N}$ , where each masklet $M_{i} \in [0,1]^T \times [0,1]$ . For predictions, the task additionally requires a confidence score $s_{i} \in [0,1]$ associated with each masklet. Under the non-overlapping constraint, these $N$ masklets can be merged into a single video sequence $M \in \mathbb{N}^{T \times H \times W}$ , where each pixel value represents the corresponding prediction ID. + +For evaluation, we first review traditional Video Panoptic Quality (VPQ), which computes metrics per category: + +$$ +V P Q = \frac {1}{| C |} \sum_ {c \in C} \frac {\sum_ {t} I o U _ {c , t}}{T P _ {c} + 0 . 5 \times F P _ {c} + 0 . 5 \times F N _ {c}} \tag {1} +$$ + +To better evaluate video entity segmentation, we propose + +Table 1. Comparison of Different Video Segmentation Tasks + +
TaskCategoryInference
ThingsStuffAutomatic PromptCross-DatasetZero-shot
Video Instance Seg.✓✓✓
Video Semantic Seg.✓✓✓
Video Object Seg.✓✓✓
Video Panoptic Seg.✓✓✓✓
Video Entity Seg.✓✓✓✓✓
+ +Video Entity Quality (VEQ), a class-agnostic metric that assesses performance across temporal sequences: + +$$ +\mathrm {V E Q} = \frac {\sum_ {t} I o U _ {t}}{T P + 0 . 5 \times F P + 0 . 5 \times F N} \tag {2} +$$ + +$$ +\mathrm {V E Q} ^ {S Q} = \quad \frac {\sum_ {t} I o U _ {t}}{T P} \tag {3} +$$ + +$$ +\mathrm {V E Q} ^ {R Q} = \frac {\sum_ {t} T P}{T P + 0 . 5 \times F P + 0 . 5 \times F N} \tag {4} +$$ + +where $IoU_{t}$ represents the temporal intersection over union across frames, measuring both spatial accuracy and temporal consistency. Here, $TP$ , $FP$ , and $FN$ are computed globally across all entities without category differentiation. For consistency with Video Panoptic Segmentation evaluation, we also implement a class-agnostic variant of the STQ (Segmentation and Tracking Quality) metric, denoted as $\mathrm{STQ}^{EN}$ , which emphasizes tracking consistency across frames. The formulation of VEQ provides a more direct assessment of video-level segmentation and tracking performance while eliminating the complexity of category-specific evaluation. Moreover, this class-agnostic approach reduces annotation requirements, enables large-scale training, and enhances model generalizability across datasets, particularly for open-vocabulary and zero-shot scenarios. + +As illustrated in Table 1, Video Entity Segmentation presents distinct advantages compared to existing video segmentation paradigms: Unlike Video Object Segmentation, it operates automatically and accommodates multiple dynamic objects. In contrast to Video Instance Segmentation, it extends beyond foreground objectness to include background entities. Compared to Video Semantic Segmentation, it provides instance-level annotations. Unlike Video Panoptic Segmentation, it adopts a class-agnostic approach that eliminates the constraint of pre-defined category vocabularies. This unified entity representation allows both thing and stuff categories to be learned jointly and subsequently classified using external Vision Language Models, enabling advanced open-vocabulary evaluation and reasoning-based segmentation tasks. + +# 3.2. Method Background: SAM 2 + +SAM 2 [39] extends the Segment Anything model [22] to encompass both image and video domains. Through an iterative process of data annotation and model training, SAM 2 leverages both manual and automated datasets generated by its data engine. For video segmentation functionality, + +![](images/fb6ed7ec0b4a7475fdc8f8a5abb4e21daa7ad49e64373544165aed1ea7bb37e7.jpg) +Figure 2. (a) Overview of the EntitySAM framework. EntitySAM utilizes the frozen encoder and memory parameters from SAM 2, incorporating a dual encoder design for enhanced semantic features (Section 3.3.1). The PromptGenerator (Section 3.3.2) automatically generates prompts from Prompt Queries. The enhanced features and distinct query groups are processed by the EntityDecoder (Section 3.3.3) to produce video mask outputs. (b) Self-attention and cross-attention mechanisms in EntityDecoder layers. + +SAM 2 incorporates a classical memory mechanism to facilitate object tracking. + +The architecture processes video input through multiple stages: initially, a multi-scale Image encoder extracts frame-level features, which then undergo cross-attention with previous video-level memory features to derive per-mask temporal information. Concurrently, input prompts are processed through a dedicated prompt encoder to generate prompt encodings. These various features and prompts are integrated through cross-attention mechanisms and dynamic convolution operations within the mask decoder to generate mask outputs. The model maintains temporal coherence by encoding and storing mask sequences in a memory bank, which updates frame-by-frame and feeds into memory attention for processing subsequent frames. + +Through extensive training on 35M manual and automated mask annotations, SAM 2 shows exceptional zero-shot performance across various video object segmentation benchmarks, including DAVIS [35], YTVOS [46], LVOS [16], and MOSE [12]. However, SAM 2 exhibits notable limitations: it requires initial mask/prompt specification at the sequence start and cannot effectively accommodate new objects introduced during tracking. Moreover, the model's efficiency deteriorates when handling large numbers of concurrent objects, indicating challenges in scaling to more complex scenarios with multiple dynamic objects. + +# 3.3. EntitySAM + +Coupled with the proposed Video Entity Segmentation task, we present EntitySAM, a novel model that achieves strong zero-shot video entity segmentation performance while requiring only image-level class-agnostic entity annotations. As illustrated in Figure 2, our model preserves the memory-based architecture of the original SAM by freezing both the pretrained image encoder and memory-related modules, thus retaining its advantages in long-term tracking of objects. Meanwhile, EntitySAM comprises three learnable components: Dual Visual Encoder, EntityPromptGenerator, + +and EntityDecoder. The following sections will detail these architectural components. + +# 3.3.1 Dual Visual Encoder + +Our EntitySAM leverages the frozen Segment Anything Model 2 (SAM 2) image encoder as its backbone, as well as its frozen memory encoder and memory attention layers. The frozen design preserves SAM 2's zero-shot feature extraction capabilities, which were developed through training on extensive video object tracking datasets. Additionally, freezing the memory encoder substantially reduces GPU memory requirements during the training of multiple dense entity tracking in video sequences. + +Although SAM 2 exhibits impressive performance across multi-granularity class-agnostic segmentation tasks, we reveal its limitations in maintaining consistent entity-level representations as in Table 9. Specifically, SAM 2 struggles to generate reliable confidence scores for entity classification, where an 'entity' refers to a semantically coherent group. Thus in addition to the original SAM 2's backbone feature, we integrate the backbone features from DINOv2 [32], which produces consistent feature representations for semantically similar regions by large-scale self-supervised training. + +As illustrated in Figure 2, EntitySAM incorporates DINOv2 through a linear projection layer, aligning its feature dimensionality with that of SAM 2. The aligned features are then concatenated and fed into the decoder to form a fused representation. This enhances the encoded visual representation by preserving the zero-shot capabilities for video entity segmentation while effectively mitigating SAM 2's limitations in entity-level understanding. + +# 3.3.2 Automatic Prompt Generator + +In SAM 2's architecture, the prompt encoder plays a crucial role in image-level segmentation, where input prompts, such as points and bounding boxes, are processed to generate prompt embeddings for decoder segmentation. However, in video-level operations, SAM 2 bypasses the prompt + +encoder, focusing solely on object tracking using historical memory features. To extend the model's capability for new entity tracking in video sequences, EntitySAM incorporates prompt embeddings during video training. + +Our EntitySAM implements a prompt generator utilizing learnable queries that perform cross-attention with enhanced visual features to predict input prompts for the prompt encoder. The resulting prompt embeddings are subsequently fed into the decoder at each frame to localize both existing and novel entities within the video sequence. Concurrently, this module predicts an entity confidence score, denoted as $S(\text{entity})$ , which performs binary classification between foreground entities and background regions. + +More formally, let $\mathcal{Q} \in \mathbb{R}^{N \times D}$ represent the learnable queries and $\mathcal{F} \in \mathbb{R}^{HW \times D}$ denote the enhanced visual features, where $N$ is the number of queries, $D$ is the feature dimension, and $H, W$ are spatial dimensions. The prompt generation and scoring process can be expressed as: $\mathcal{A} = \text{CrossAttention}(\mathcal{Q}, \mathcal{F})$ , $\mathcal{P} = \text{FC}_{\text{prompt}}(A)$ , and $\mathcal{S} = \text{FC}_{\text{score}}(A)$ , where $A \in \mathbb{R}^{N \times D}$ is the cross-attention output, $\mathcal{P} \in \mathbb{R}^{N \times D'}$ represents the generated prompts with $D'$ matching the expected prompt dimension required by the prompt encoder, and $\mathcal{S} \in \mathbb{R}^{N \times 1}$ denotes the entity confidence scores. + +# 3.3.3 Entity Mask Decoder + +The Entity Mask Decoder predicts output masks based on enhanced features in Section 3.3.1 and the learnable prompts through a multi-stage processing pipeline in Section 3.3.2. It employs two groups of batched queries: mask queries $\mathcal{Q}_m\in \mathbb{R}^{N\times 1\times C}$ and IoU queries $\mathcal{Q}_{iou}\in \mathbb{R}^{N\times 1\times C}$ , where $N$ denotes the number of entities and $C$ represents the feature dimension. These queries are concatenated with input prompts for self-attention to facilitate inter-entity communication. These queries are subsequently reshaped for cross-attention operations between token and image representations. In the final layer, mask queries undergo dynamic convolution with upsampled mask features to generate output masks, while IoU queries are processed through a dedicated prediction head to estimate conditioned IoU scores. The computational flow can be formalized as: + +$$ +\begin{array}{l} [ \mathcal {Q} _ {m} ^ {\prime}, \mathcal {Q} _ {i o u} ^ {\prime} ] = \text {S e l f A t t e n t i o n} (\text {c o n c a t} (\mathcal {Q} _ {m}; \mathcal {Q} _ {i o u})) \\ w h e r e \operatorname {c o n c a t} \left(\mathcal {Q} _ {m}; \mathcal {Q} _ {i o u}\right) \in \mathbb {R} ^ {1 \times 2 N \times C} \\ F _ {o u t} = \operatorname {C r o s s A t t e n t i o n} (\operatorname {r e s h a p e} \left(\mathcal {Q} _ {m} ^ {\prime}, \mathcal {Q} _ {i o u} ^ {\prime}\right), F) \\ \operatorname {w h e r e} \operatorname {r e s h a p e} \left(\mathcal {Q} _ {m} ^ {\prime}, \mathcal {Q} _ {i o u} ^ {\prime}\right) \in \mathbb {R} ^ {N \times 2 \times C}. \tag {5} \\ \end{array} +$$ + +The supervision mechanism integrates outputs from both the prompt generator (entity scores) and mask decoder (masks and conditioned IoU predictions) through Hungarian matching with ground truth annotations. The expected + +IoU for query $q$ can be expressed as: + +$$ +\begin{array}{l} E (\mathrm {I o U} _ {q}) = E (\mathrm {I o U} _ {q} | \mathrm {E n t} _ {q}) S (\mathrm {E n t} _ {q}) + E (\mathrm {I o U} _ {q} | \neg \mathrm {E n t} _ {q}) S (\neg \mathrm {E n t} _ {q}) \\ = E \left(\operatorname {I o U} _ {q} \mid \operatorname {E n t} _ {q}\right) S \left(\operatorname {E n t} _ {q}\right), \tag {6} \\ \end{array} +$$ + +where $S(\mathrm{Ent}_q)$ represents the entity confidence score from PromptGenerator in Section 3.3.2 and $E(\mathrm{IoU}_q|\neg \mathrm{Ent}_q) = 0$ . + +The resulting output masks are processed through the frozen memory encoder, memory bank, and memory attention mechanisms, with gradient computation detached for memory-related components. This enables efficient training of the queries-based tracking component while maintaining frozen memory modules. Consequently, our approach achieves scalable training for over 50 entities in 8-frame video clips, significantly surpassing SAM 2's original capacity of handling only 3 video objects during training. + +# 3.3.4 Training Loss and Open-Vocabulary Inference + +We employ an end-to-end training objective for EntitySAM, formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {M a s k}} + \lambda_ {1} \mathcal {L} _ {\text {E n t i t y}} + \lambda_ {2} \mathcal {L} _ {\text {I o U}}, \tag {7} +$$ + +where $\mathcal{L}$ supervises both mask prediction and binary entity classification, following the Mask2former framework [7]. The hyperparameters $\lambda_{1}$ and $\lambda_{2}$ balance the respective loss terms and are empirically set to 1.0 and 1.0. + +During inference, EntitySAM accommodates both class-agnostic and class-specific evaluation protocols. For class-agnostic scenarios, entity masks are directly generated and merged based on confidence scores and mask overlap metrics. In class-specific evaluation, we select the frame with the highest entity confidence score for each masklet prediction. The predicted entity masks are utilized to crop corresponding regions of interest, which, together with carefully designed prompts containing candidate category names, are processed through GPT-4o [1] for final classification. We demonstrate in the supplementary materials that this mask-guided cropping approach achieves superior classification performance compared to alternative methods. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets Video entity segmentation typically requires extensive class-agnostic dense entity video segmentation annotations. Although it eliminates the need for class-specific annotations, creating frame-by-frame annotations remains labor-intensive. Inspired by [51, 52], our training pipeline leverages COCO [23] panoptic segmentation data without category label annotation (denoted as $\mathrm{COCO}^*$ ) and extends it to pseudo-video sequences. Our approach utilizes only image-level annotations, eschewing video-level labels, to achieve effective video entity segmentation. The pipeline + +also enables zero-shot cross-dataset evaluation across multiple benchmarks. We demonstrate our model's versatility through comprehensive zero-shot evaluations on various tasks: video panoptic segmentation on VIPSeg [28], video semantic segmentation on VSPW [27], video instance segmentation on Youtube-VIS 2019 [48], Youtube-VIS 2021 [48], and OVIS [36], and video entity segmentation on our proposed VIPEntitySeg dataset—a class-agnostic variant of VIPSeg. Notably, all zero-shot evaluations are performed using a single trained model, demonstrating its robust zero-shot generalization capabilities across different video segmentation tasks and datasets. + +Implementation details In our experimental setup, we employ two backbone architectures: ViT-Small and ViT-Large. The models are trained using the AdamW optimizer with a learning rate of $5 \times 10^{-5}$ and weight decay of 0.05. Input images are cropped to $1024 \times 1024$ pixels for compatibility with SAM 2 ViT/16 and $896 \times 896$ pixels for DINOv2 ViT/14. Training is conducted with a batch size of 8. The loss function combines multiple components: cross-entropy and dice losses for mask prediction, L1 loss for IoU prediction, and focal loss for binary entity classification. Our training strategy consists of two phases: initial training on static COCO images for 40K iterations, followed by training on COCO pseudo-video sequences with clip length 8 for 10K iterations. Both stages are class-agnostic. During inference, we utilize a single entity category output mode with confidence scoring. Video entity quality (VEQ) is evaluated in a class-agnostic manner. For class-specific evaluation tasks, we implement a multi-step process: first selecting the highest-confidence mask from the masklet for each prediction, then extracting the corresponding frame region within the mask, and finally employing GPT-4o for classification. + +# 4.2. Zero-shot Comparison + +Zero-shot Video Entity Segmentation In Table 2 (left), we evaluate our approach on VIPEntitySeg using class-agnostic metrics against several baseline methods. For SAM 2, we adopt two prompting manners, a grid point prompt sampling initialization and a COCO-trained Mask2Former [7] mask prompts initialization. We additionally compare against state-of-the-art zero-shot segmentation models, including: 1) FC-CLIP [51] integrated with MinVIS [17] association, 2) DEVA [11] coupled with COCO-trained Mask2Former [7] for per-frame segmentation, and 3) OV-DVIS++ [52] in both online and offline modes. All these SOTA models were trained on COCO with category annotations, with DEVA utilizing additional external datasets for its association model. We compare with these methods on our class-agnostic VIPEntitySeg benchmark. Our EntitySAM achieves significant performance improvements over all baselines and SOTA models with both ViT-Small and ViT-Large backbones, despite training + +without category annotations. This superior performance demonstrates the effectiveness of our class-agnostic training approach and proposed modules. Notably, we observe substantial improvements in $\mathrm{VEQ}^{RQ}$ , indicating enhanced cross-frame entity identification and reduced false positives. Figure 3 further illustrates our model's robustness by evaluating VEQ across varying frame intervals, demonstrating that EntitySAM maintains consistent performance over extended temporal sequences. + +Regarding the class-agnostic $\mathrm{STQ}^{EN}$ metric, which emphasizes tracking performance, DEVA achieves higher scores through its tracking association model trained on external video object segmentation datasets. However, DEVA relies on a two-stage pipeline of detection followed by association, lacking end-to-end integration. While our approach achieves lower scores, it delivers comparable performance through a streamlined process that requires only COCO training data and single-pass inference. + +Zero-shot Video Panoptic Segmentation In Table 2 (right), we extend our evaluation to VIPSeg for class-specific video panoptic segmentation. Several baseline models, including Mask2Former and DEVA, are constrained to VIPSeg evaluation due to their closed-vocabulary nature. When compared to open-vocabulary models such as FC-CLIP and OV-DVIS++, our approach demonstrates substantial improvements, particularly with the ViT-Small backbone. These results validate our approach of "post-classification with multimodal models", indicating that it surpasses traditional open-vocabulary segmentation training pipelines despite operating without category annotations. Our EntitySAM achieves competitive performance on both class-agnostic and class-specific tasks, suggesting that category annotations may not be essential for effective open-vocabulary video segmentation. + +Zero-shot Video Semantic Segmentation Table 3 presents our results on video semantic segmentation, obtained by merging per-pixel mask predictions. Our method achieves improvements in both mean Intersection over Union (mIoU) and Video Consistency (VC) metrics. The VSPW evaluation dataset, which contains the same video content but with extended frame sequences, validates our model's capability for effective long-term tracking. + +Zero-shot Video Instance Segmentation As shown in Table 4, we evaluate EntitySAM on three popular video instance segmentation datasets. Our approach on extending video entity segmentation to video instance segmentation requires no model retraining—we simply add a background/stuff class during inference and filter out predictions with that class to obtain instance results. Our model demonstrates significant performance improvements over state-of-the-art open-vocabulary video instance segmentation models. In particular, on the challenging OVIS dataset, which features frequent occlusions, our model maintains robust + +Table 2. Comparison of Zero-shot Video Entity Segmentation (left) and Video Panoptic Segmentation (right) on COCO $\rightarrow$ VIPSeg evaluation. EntitySAM is trained on the COCO dataset without category annotations. In VIPSeg, - means the results of open-vocabulary evaluation for closed-vocabulary models are not available. * denotes that DEVA is trained with multiple external video object segmentation datasets for the temporal association model. + +
MethodBackboneClass Agnostic TrainingVIPEntitySegVIPSeg
VEQVEQSQVEQRQSTQENVPQVPQThVPQStSTQ
SAM 2.1 [39] (Everything Init)ViT-S✓26.982.532.638.7-
SAM 2.1 [39] (Mask2former Init)ViT-S + ResNet5046.283.655.234.5-
DEVA* [11]ResNet-5049.083.658.642.4-
FC-CLIP [51]ResNet-5041.584.049.429.722.325.519.119.7
OV-DVIS++(Online) [52]ResNet-5045.983.754.833.624.426.822.422.0
OV-DVIS++(Offline) [52]ResNet-5046.782.856.335.623.826.421.424.4
OursViT-S✓51.884.761.141.028.732.925.131.4
SAM 2.1 [39] (Everything Init)ViT-L✓30.583.136.741.7-
SAM 2.1 [39] (Mask2former Init)ViT-L + Swin-L50.684.460.039.1-
DEVA* [11]Swin-L50.883.860.645.3-
FC-CLIP [51]ConvNext-L43.484.851.232.427.930.925.024.2
OV-DVIS++(Online) [52]ConvNext-L49.483.559.237.928.931.326.828.4
OV-DVIS++(Offline) [52]ConvNext-L49.083.159.040.030.431.929.132.2
OursViT-L✓54.684.764.543.331.437.526.033.5
+ +Table 3. Zero-Shot Open-vocabulary Video Semantic Segmentation on COCO $\rightarrow$ VSPW. EntitySAM is trained on the COCO dataset without category annotations. + +
MethodBackboneClass Agnostic TrainingmVC8VSPW mVC16mIOU
FC-CLIP [51]ResNet-5084.982.724.3
OV-DVIS++(Online) [52]ResNet-5092.791.527.6
OV-DVIS++(Offline) [52]ResNet-5092.491.328.4
OursViT-S✓94.093.034.6
FC-CLIP [51]ConvNext-L89.988.428.9
OV-DVIS++(Online) [52]ConvNext-L94.293.334.3
OV-DVIS++(Offline) [52]ConvNext-L93.993.034.1
OursViT-L✓94.693.735.5
+ +performance, benefiting from the entity learning strategy employed during training. + +# 4.3. Ablation Experiments + +We perform ablation studies using the ViT-Small backbone on the COCO $\rightarrow$ VIPEntitySeg video entity segmentation task with zero-shot protocol to evaluate the effectiveness of our proposed modules. + +Ablation on the Proposed Modules Table 5 presents ablation studies of our proposed modules. Our baseline, SAM 2 with memory-based tracking modules, cannot generate predictions automatically without appropriate prompting. Using grid-point sampling leads to numerous false positive predictions, achieving a VEQ of only 26.9. Removing the memory design and implementing solely a query-based EntityDecoder with query association for tracking improves performance to 41.3, though this remains suboptimal. Combining memory-based and query-based approaches significantly enhances performance to 50.2, surpassing both DEVA and OV-DVIS++. Finally, incorporating a complementary DINOv2 dual feature encoder achieves the best result (51.8) for the ViT-Small backbone. + +Ablation on Query Design As illustrated in Figure 2, our design incorporates three query types: two groups in EntityDecoder and one group in PromptGenerator. Our abla + +![](images/faff9aa980542777667efb97d103d049495e6848ff1377b97d5ecac62b7451a9.jpg) +Figure 3. VEQ comparison with different evaluation time intervals. EntitySAM improves over long intervals. + +![](images/4298e2160ec64bc276eea6e068907ffaf70a6180032fc2487f0d8276ec0248ba.jpg) + +tion experiment of query design in Table 6 reveals that using a single query group to predict all outputs (mask, IoU, and confidence score) yields suboptimal performance with a VEQ of 49.2. Adding a second group of prompt queries improves performance to 49.7. Further separation into a third query group for distinct mask and IoU prediction achieves the best performance of 51.8. These results suggest that different query groups learn specialized distributions for their respective prediction tasks, departing from traditional detection transformers' shared-query architecture. + +Ablation on EntityDecoder Depth Table 7 studies the influence of the entity decoder's depth layer. For the ViT-Small backbone, a decoder depth of 4 achieves optimal performance with a VEQ of 51.8 on COCO $\rightarrow$ VIPSeg setting. Further increasing decoder depth shows no additional performance benefits. + +Ablation on Training Stages As detailed in Section 4.1, our training process consists of two stages: initial training on single images (treated as zero-length videos) followed by training on pseudo video clips of length 8. During the first stage, the memory component remains unused while the EntityDecoder learns to identify new entities in each image. Table 8 shows that this image-only training achieves limited video entity segmentation performance with a VEQ of 31.4. However, this stage exposes the model to diverse en + +Table 4. Zero-Shot Open-vocabulary Video Instance Segmentation on COCO $\rightarrow$ Validation sets of YouTube-VIS 2019, 2021, and OVIS. EntitySAM is trained on the COCO dataset without category annotations. + +
MethodBackboneClass Agnostic TrainingYoutube-VIS 2019Youtube-VIS 2021OVIS
AP\( AP_{50} \)\( AP_{75} \)\( AR_{1} \)\( AR_{10} \)AP\( AP_{50} \)\( AP_{75} \)\( AR_{1} \)\( AR_{10} \)AP\( AP_{50} \)\( AP_{75} \)\( AR_{1} \)\( AR_{10} \)
Detric[53]-OWTB[25]ResNet-5017.9----16.7----9.0----
MindVLT[41]ResNet-5023.1----20.9----11.4----
FC-CLIP [51]ResNet-5028.943.929.932.441.225.540.026.828.037.011.825.110.58.516.4
OV-DVIS++(Online) [52]ResNet-5034.550.639.239.549.530.946.734.834.445.814.831.213.110.524.7
OV-DVIS++(Offline) [52]ResNet-5034.451.839.240.050.631.048.534.334.646.813.028.410.99.825.1
OursViT-S✓44.960.749.543.662.438.852.543.136.256.521.035.820.812.336.0
+ +Table 5. Ablation Study on the Proposed Modules. + +
ModelMemoryEntityDecoderDual Visual EncoderVEQ
SAM 2 [39]✓26.9
Ours✓✓41.3
✓✓50.2
✓✓✓51.8
+ +Table 6. Ablation Study on Query Design. + +
EntityDecoderPromptGeneratorCOCO → VIPSeq
VEQVEQSQVEQRQ
Shared Queries49.284.058.6
Shared Queries✓49.784.259.0
Separate Queries✓51.884.761.1
+ +Table 7. Ablation Study on EntityDecoder Depth. + +
DecoderDecoder depthCOCO → VIPSeg
VEQVEQSQVEQRQ
EntityDecoder251.384.760.5
451.884.761.1
651.684.661.0
851.184.460.6
+ +tities, and when combined with the subsequent video training stage—which aligns EntityDecoder with memory encoders—performance improves substantially to 51.8 VEQ. This two-stage approach outperforms training exclusively on pseudo video sequences. + +Ablation on Dual Encoder In Section 3.3.1, we introduced encoder feature fusion designs for dual encoders. As shown in Table 9, our ablation studies reveal that incorporating features from either CLIP [38] or SAM [22] alongside the SAM 2 encoder features yields no improvement in zero-shot performance. This can be attributed to CLIP's vision encoder being optimized for classification rather than segmentation tasks, resulting in representations less suitable for segmentation. Similarly, SAM, being trained with an approach analogous to SAM 2, provides features that largely overlap with SAM 2's capabilities rather than offering complementary information. In contrast, DINOv2, trained through self-supervision, demonstrates significant benefits when combined with SAM 2. Its self-supervised training enables DINOv2 to learn consistent image representations that effectively complement SAM 2's supervised features, ultimately enhancing the overall performance. + +Table 8. Ablation Study on Training stages. + +
ModelImagePseudo VideoCOCO → VIPSeg
VEQVEQSQVEQRQ
Ours✓31.479.439.5
✓49.284.258.4
✓✓51.884.761.1
+ +Table 9. Ablation Study on Dual Encoder Feature. + +
Encoder FeatureDual EncoderCOCO → VIPSeg
VEQVEQSQVEQRQ
SAM 2-50.284.559.4
SAM [22]47.684.256.5
CLIP [38]49.284.058.6
DINOv2 [32]51.884.761.1
+ +# 4.4. Qualitative experiments and visualization + +Figure 4 demonstrates our model's performance on zero-shot video sequences, showcasing consistent entity tracking and segmentation across frames. Additional qualitative results are provided in the supplementary materials. + +![](images/971a321b7254a0fef2a56a5217755b09d596e0f91ed9a1790494d156c4a0ff8e.jpg) +Figure 4. Visualization of zero-shot video entity segmentation. + +![](images/b224e69f03d2ef284977ad493ef40659edb97820e6316c8007ed3d070f783760.jpg) + +![](images/55082acb74841d8e6cd604a3be21c25b6e52b401b5ba35cb27a125feb2afdd45.jpg) + +# 5. Conclusion + +We present EntitySAM, a framework that integrates query and memory mechanisms to enhance SAM 2 for zero-shot video entity segmentation. Our approach eliminates the need for user-defined prompt inputs and enables automatic tracking of new entities across frames. Trained exclusively on COCO image-level mask annotations without category information, EntitySAM achieves superior performance on the Video Entity Segmentation task. The model demonstrates strong generalization capabilities across various zero-shot evaluations, including Video Panoptic / Semantic / Instance Segmentation. These results provide valuable insights into class-agnostic training approaches and their effectiveness in zero-shot scenarios. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5 +[2] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2 +[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2 +[4] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. In arXiv preprint arXiv:1905.00737, 2019. 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[6] Bowen Cheng, Anwesa Choudhuri andd Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. In https://arxiv.org/abs/2112.10764, 2021. 2 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 5, 6 +[8] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 2 +[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 2 +[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In arXiv, 2023. 2, 3 +[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 1, 2, 6, 7 +[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, 2023. 2, 4 +[13] Yuming Du, Wen Guo, Yang Xiao, and Vincent Lepetit. Uvo challenge on video-based open-world segmentation 2021: 1st place solution. ICCV Workshop, 2021. 2 +[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. Advances in Neural Information Processing Systems, 35:23109–23120, 2022. 2 +[15] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14623-14632, 2023. 2 +[16] Lingyi Hong, Wenchao Chen, Zhongying Liu, Wei Zhang, Pinxue Guo, Zhaoyu Chen, and Wenqiang Zhang. Lvos: A benchmark for long-term video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13480-13492, 2023. 4 +[17] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6 +[18] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 2 +[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, 2023. 1 +[20] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2 +[21] Dahun Kim, Jun Xie, Huiyu Wang, Siyuan Qiao, Qihang Yu, Hong-Seok Kim, Hartwig Adam, In So Kweon, and Liang-Chieh Chen. Tubeformer-deeplab: Video mask transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13914-13924, 2022. 2 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 3, 8 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5 +[24] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, 2022. 2 +[25] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 8 +[26] Qi Lu, Jason Kuen, Shen Tiancheng, Gu Jiumiang, Guo Weidong, Jia Jiaya, Lin Zhe, and Yang Ming-Hsuan. High-quality entity segmentation. In ICCV, 2023. 2 +[27] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. Vspw: A large-scale dataset for video scene parsing in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4133-4143, 2021. 1, 6 +[28] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 1, 2, 6 +[29] Xindong Zhang Minghan Li, Shuai Li and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries, 2024. 3 +[30] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6819-6828, 2018. 2 +[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2 + +[32] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4, 8 +[33] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In CVPR, 2022. 2 +[34] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 2 +[35] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 4 +[36] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation. IJCV, 2022. 2, 6 +[37] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. In arXiv preprint arXiv:2107.14228, 2021. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8 +[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3, 7, 8 +[40] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3126-3137, 2022. 2 +[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In proceedings of the IEEE/CVF international conference on computer vision, pages 4057-4066, 2023. 8 +[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2 +[43] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In European Conference on Computer Vision, pages 553-569. Springer, 2022. + +[44] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, 2022. 2 +[45] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In European Conference on Computer Vision, pages 588-605. Springer, 2022. 2 +[46] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. In ECCV, 2018. 2, 4 +[47] Xiaohao Xu, Jinglu Wang, Xiao Li, and Yan Lu. Reliable propagation-correction modulation for video object segmentation. In AAAI, 2022. 2 +[48] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 6 +[49] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 2 +[50] Zongxin Yang, Jiaxu Miao, Yunchao Wei, Wenguan Wang, Xiaohan Wang, and Yi Yang. Scalable video object segmentation with identification mechanism. TPAMI, 2024. 2 +[51] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 5, 6, 7, 8 +[52] Tao Zhang, Xingye Tian, Yikang Zhou, Shunping Ji, Xuebo Wang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, and Yu Wu. Dvis++: Improved decoupled framework for universal video segmentation. arXiv preprint arXiv:2312.13305, 2023. 2, 5, 6, 7, 8 +[53] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8 \ No newline at end of file diff --git a/2025/EntitySAM_ Segment Everything in Video/images.zip b/2025/EntitySAM_ Segment Everything in Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c317c851c062f40aa0c544ce8037a10413a73c4 --- /dev/null +++ b/2025/EntitySAM_ Segment Everything in Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3c45c90d62cbcbb77b9da9791ff952d1351bcc0a921ae8b143bd52397c96ab8 +size 514385 diff --git a/2025/EntitySAM_ Segment Everything in Video/layout.json b/2025/EntitySAM_ Segment Everything in Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5ea0f8e526fde11b7cd024bb09157b0452bc53e --- /dev/null +++ b/2025/EntitySAM_ Segment Everything in Video/layout.json @@ -0,0 +1,7561 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 175, + 103, + 436, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 103, + 436, + 121 + ], + "spans": [ + { + "bbox": [ + 175, + 103, + 436, + 121 + ], + "type": "text", + "content": "EntitySAM: Segment Everything in Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "text", + "content": "Mingqiao Ye" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "text", + "content": " Seoung Wug Oh" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "text", + "content": " Lei Ke" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "text", + "content": " Joon-Young Lee" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 142, + 465, + 173 + ], + "type": "text", + "content": "EPFL 2Adobe Research 3Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 180, + 555, + 350 + ], + "blocks": [ + { + "bbox": [ + 59, + 180, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 59, + 180, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 59, + 180, + 555, + 350 + ], + "type": "image", + "image_path": "72cda20c7485f39168bcc495c67ad7fc72aa5b31e7be012a612f4209f8f75d12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 356, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 555, + 401 + ], + "type": "text", + "content": "Figure 1. Zero-shot video entity segmentation performance comparison on VIPEntySeg dataset using models trained on COCO, showing: 1) SAM 2 [39] using Mask2Former [7] mask prompts for the initial frame, 2) Mask2Former with DEVA [11] association, and 3) our proposed EntitySAM. Our EntitySAM enhances SAM 2 by automatically segmenting and tracking novel entities without requiring user-specified prompts, achieving superior performance compared to existing state-of-the-art zero-shot tracking methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 416, + 200, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 416, + 200, + 428 + ], + "spans": [ + { + "bbox": [ + 151, + 416, + 200, + 428 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 441, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 297, + 693 + ], + "type": "text", + "content": "Automatically tracking and segmenting every video entity remains a significant challenge. Despite rapid advancements in video segmentation, even state-of-the-art models like SAM 2 struggle to consistently track all entities across a video—a task we refer to as Video Entity Segmentation. We propose EntitySAM, a framework for zero-shot video entity segmentation. EntitySAM extends SAM 2 by removing the need for explicit prompts, allowing automatic discovery and tracking of all entities, including those appearing in later frames. We incorporate query-based entity discovery and association into SAM 2, inspired by transformer-based object detectors. Specifically, we introduce an entity decoder to facilitate inter-object communication and an automatic prompt generator using learnable object queries. Additionally, we add a semantic encoder to enhance SAM 2's semantic awareness, improving segmentation quality. Trained on image-level mask annotations without category information from the COCO dataset, EntitySAM demonstrates strong generalization on four zero-shot video segmentation tasks: Video Entity, Panoptic, Instance, and Semantic Segmentation. Results on six popular benchmarks show that" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 417, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 465 + ], + "type": "text", + "content": "EntitySAM outperforms previous unified video segmentation methods and strong baselines, setting new standards for zero-shot video segmentation. Our code and models are at github.com/ymq2017-entitysam." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 489, + 394, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 394, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 394, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 509, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 605 + ], + "type": "text", + "content": "Video segmentation is an essential task for various applications in video scene understanding, including video editing, AR/VR, and robotic perception. There are roughly two lines of video segmentation approaches. One line of approaches focuses on selecting semantic thing/stuff from predefined sets [27, 28, 48]. While these methods show strong performance in the benchmark, the predictions are limited to a pre-defined class set and the training data domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "Another line of approaches works by initialing the information about the target object. The most common approach in this group is semi-supervised video object segmentation (VOS) which tracks a given mask in the first frame. The recent SAM 2 [39] also falls into this group. SAM 2 generalizes the VOS task by taking multiple prompt types, including masks, boxes, and clicks similar to image segmentation methods [19, 22], and enables interactive video segmentation. Trained on large-scale training data, SAM 2 shows" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "text", + "content": "*This work was done during an internship at Adobe Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24234" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 263, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 263, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 263, + 83 + ], + "type": "text", + "content": "strong zero-shot broad-domain selection capability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 294, + 215 + ], + "type": "text", + "content": "Despite recent progress, a significant challenge remains in selecting and tracking all entities from a video—a task we term Video Entity Segmentation, which aims to segment any distinct group with coherent semantic meaning. While SAM 2 is an effective model for tracking each individual entity, we identified two limitations in selecting all entities: (1) It initializes tracking from user prompts so it cannot track unmarked entities and new entities appearing in different frames, as shown in Figure 1. (2) It tracks each entity independently which is not only inefficient and can lead to less optimal results (e.g., mask overlapping)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 217, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 455 + ], + "type": "text", + "content": "In this paper, we propose EntitySAM, an extension of SAM 2 for zero-shot video entity segmentation. EntitySAM retains the zero-shot potential of SAM 2 while eliminating the need for explicit prompts for each entity and enabling seamless integration of new entities during tracking. Inspired by transformer-based object detectors [5, 6, 42], we sophisticatedly integrated query-based object discovery and association into SAM 2 making minimal changes to the original structure. To be specific, we replace the original mask decoder with a new entity decoder and introduce an automatic prompt generator that generates SAM 2 consumable prompts from learnable object queries. In this way, we can preserve most of the pre-trained SAM 2 weights. The proposed entity decoder differs from the original SAM 2 mask decoder in that it enables inter-entity communication to predict multiple entity masks without overlaps. Another notable addition to SAM 2 is a semantic encoder. We empirically find that SAM 2 encoder has a weak semantic awareness and supplementing semantic information using an external feature encoder greatly boosts the performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 647 + ], + "type": "text", + "content": "We train EntitySAM on the COCO dataset [23] without using any category annotation and evaluate it on various zero-shot video segmentation tasks. First we introduce a video entity segmentation task. Video entity segmentation is class-agnostic, focusing on the segmentation and tracking of \"entities\" rather than predefined object categories. For this, we re-purpose the video panoptic segmentation dataset [28] by removing category annotation. In addition to this, we evaluate our method on multiple video instance and semantic segmentation benchmarks in zero-shot setting. Experimental results show that EntitySAM greatly outperforms previous unified video segmentation models [52] and strong baselines that combine an image detector [7] and a mask propagator [11]. Moreover, our method can be easily adapted to class-specific video panoptic segmentation by integrating state-of-the-art vision-language models [1]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 657, + 148, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 148, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 148, + 670 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Taxonomy of Video Segmentation Tasks Video segmentation tasks can be broadly categorized into class-specific and class-agnostic paradigms. Class-specific tasks," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "such as Video Instance Segmentation [36, 42, 48], Video Semantic Segmentation [30, 40], and Video Panoptic Segmentation [20, 28], operate within a closed vocabulary, segmenting \"thing\" or \"stuff\" categories based on predefined labels throughout video sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 134, + 555, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 555, + 242 + ], + "type": "text", + "content": "In contrast, class-agnostic tasks, such as Video Object Segmentation (VOS) [34, 46], rely on prompt-based guidance. Unlike image-level prompts, video prompts typically require initial mask input in the first frame for temporal propagation. Recent work, such as SAM 2 [39], demonstrates that large-scale pre-training can generalize VOS to zero-shot applications. However, its reliance on explicit prompts becomes inefficient when tracking multiple dense objects with frequent occlusions or reappearance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 471 + ], + "type": "text", + "content": "Recent advances have expanded video segmentation to open-world and open-vocabulary paradigms. The BURST dataset [3] introduces annotations across 482 diverse categories with a focus on thing-level segmentation, while UVO [13] takes a class-agnostic open-world approach, treating all video objects as a single category without predefined labels. Entity Segmentation [26, 37] has shown promising generalization to novel categories at the image level, but extending this approach to video introduces significant challenges for zero-shot evaluation. To bridge this gap, we propose Video Entity Segmentation, a new task aimed at segmenting entities across video sequences without relying on predefined categories or explicit prompts. This approach extends image-level entity understanding to maintain temporal consistency across frames, enabling generalized handling of novel categories in a zero-shot setting. By avoiding rigid class definitions, our framework provides greater flexibility, particularly beneficial for real-world applications with unknown or fluid category boundaries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 640 + ], + "type": "text", + "content": "Query-based Video Segmentation Query-based video segmentation models are widely used for dense prediction tasks, including video instance [6, 14, 17, 18, 42-44] and semantic segmentation [2, 21]. Inspired by detection transformers [5], these methods use queries to represent object information and apply cross-attention with image features to generate segmentation masks, demonstrating effectiveness in video instance segmentation. Temporal association is achieved through either direct query matching [17, 45] or additional tracking modules [15]. While effective in controlled scenarios, these query-based methods without explicit memory propagation struggle with long or complex video sequences due to challenges in modeling long-term temporal dependencies [36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Memory-based Video Segmentation Memory-based methods [9, 24, 31, 33, 47, 49, 50] are effective in Video Object Segmentation [4, 12, 46, 46], using initial ground truth masks as prompts. Memory features are stored in memory banks for segmentation in subsequent frames, excelling in long-term video tracking [8, 10]. SAM 2 [39] validates the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24235" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "generalizability of memory-based designs but processes objects separately, limiting efficiency and lacking inter-object communication. Recent works have attempted to integrate query-based and memory mechanisms, such as Cutie [10] for object-level VOS and UniVS [29] for multi-task inference. However, joint training of memory and query modules is computationally challenging for long videos. Our EntitySAM addresses these issues by leveraging SAM 2's frozen memory encoder and introducing a novel query-based entity decoder for efficient joint training, combining the strengths of both approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 111, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 111, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 111, + 220 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 408 + ], + "type": "text", + "content": "We present EntitySAM for Video Entity Segmentation, enabling automatically segmenting and tracking all entities in the video sequence. By integrating query and memory-based video segmentation designs, EntitySAM handles multi-object tracking with inter-object communication and object-level contextual information. In Section 3.1, we first introduce the Video Entity Segmentation task, which offers improved generalization and reduced annotation complexity compared to other video segmentation tasks, such as Video Instance / Panoptic / Object Segmentation. In Section 3.2, we provide an overview of SAM 2's architecture and key characteristics. Section 3.3 details the architectural design of our EntitySAM based on SAM 2, highlighting our enhanced feature fusion and decoding mechanisms, along with comprehensive training and inference protocols." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 204, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 204, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 204, + 427 + ], + "type": "text", + "content": "3.1. Video Entity Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 528 + ], + "type": "text", + "content": "Video Entity Segmentation aims to segment and track all entities within video sequences using a class-agnostic approach. In this context, an \"entity\" encompasses both thing and stuff categories, representing any distinct group with coherent semantic meaning. The task requires continuous tracking and segmentation of these entities through temporal sequences, with annotations represented as tube masks that capture the spatial-temporal evolution of each entity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": "Specifically, given an input video sequence " + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": ", both annotations and predictions consist of " + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": " non-overlapping masklets " + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "inline_equation", + "content": "M_{i_{i=1}}^{N}" + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": ", where each masklet " + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "inline_equation", + "content": "M_{i} \\in [0,1]^T \\times [0,1]" + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": ". For predictions, the task additionally requires a confidence score " + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "inline_equation", + "content": "s_{i} \\in [0,1]" + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": " associated with each masklet. Under the non-overlapping constraint, these " + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": " masklets can be merged into a single video sequence " + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{N}^{T \\times H \\times W}" + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": ", where each pixel value represents the corresponding prediction ID." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 295, + 661 + ], + "type": "text", + "content": "For evaluation, we first review traditional Video Panoptic Quality (VPQ), which computes metrics per category:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 666, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 666, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 66, + 666, + 295, + 696 + ], + "type": "interline_equation", + "content": "V P Q = \\frac {1}{| C |} \\sum_ {c \\in C} \\frac {\\sum_ {t} I o U _ {c , t}}{T P _ {c} + 0 . 5 \\times F P _ {c} + 0 . 5 \\times F N _ {c}} \\tag {1}", + "image_path": "52fc205303b8193035d8c6f90134b841bb1f7b7e729a0ccead8098ceaa46c23c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "To better evaluate video entity segmentation, we propose" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 83, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 544, + 82 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 544, + 82 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 544, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of Different Video Segmentation Tasks" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 83, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 315, + 83, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 83, + 553, + 149 + ], + "type": "table", + "html": "
TaskCategoryInference
ThingsStuffAutomatic PromptCross-DatasetZero-shot
Video Instance Seg.✓✓✓
Video Semantic Seg.✓✓✓
Video Object Seg.✓✓✓
Video Panoptic Seg.✓✓✓✓
Video Entity Seg.✓✓✓✓✓
", + "image_path": "5bbe247da584689b00b2543bc33fc42fad4e06ceb739e54d4452a671dfdb475d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 152, + 553, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 177 + ], + "type": "text", + "content": "Video Entity Quality (VEQ), a class-agnostic metric that assesses performance across temporal sequences:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 184, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 184, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 359, + 184, + 553, + 209 + ], + "type": "interline_equation", + "content": "\\mathrm {V E Q} = \\frac {\\sum_ {t} I o U _ {t}}{T P + 0 . 5 \\times F P + 0 . 5 \\times F N} \\tag {2}", + "image_path": "36ffc0c77232a288461b62bbdf00d56eb07fe9d3986f28c517d6b04fd5eb463b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 212, + 553, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 212, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 348, + 212, + 553, + 235 + ], + "type": "interline_equation", + "content": "\\mathrm {V E Q} ^ {S Q} = \\quad \\frac {\\sum_ {t} I o U _ {t}}{T P} \\tag {3}", + "image_path": "902ed7de634e2e92dd8b034b48c940e11f2a59d517e5699d369ece19b67241ca.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 236, + 553, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 236, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 347, + 236, + 553, + 261 + ], + "type": "interline_equation", + "content": "\\mathrm {V E Q} ^ {R Q} = \\frac {\\sum_ {t} T P}{T P + 0 . 5 \\times F P + 0 . 5 \\times F N} \\tag {4}", + "image_path": "d7fa72fa08fecff278f4c70eed4bad02a76b02bbcf4f01a5c05bd74e3aa7b160.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "inline_equation", + "content": "IoU_{t}" + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "text", + "content": " represents the temporal intersection over union across frames, measuring both spatial accuracy and temporal consistency. Here, " + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "text", + "content": " are computed globally across all entities without category differentiation. For consistency with Video Panoptic Segmentation evaluation, we also implement a class-agnostic variant of the STQ (Segmentation and Tracking Quality) metric, denoted as " + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathrm{STQ}^{EN}" + }, + { + "bbox": [ + 313, + 269, + 555, + 448 + ], + "type": "text", + "content": ", which emphasizes tracking consistency across frames. The formulation of VEQ provides a more direct assessment of video-level segmentation and tracking performance while eliminating the complexity of category-specific evaluation. Moreover, this class-agnostic approach reduces annotation requirements, enables large-scale training, and enhances model generalizability across datasets, particularly for open-vocabulary and zero-shot scenarios." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 449, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 628 + ], + "type": "text", + "content": "As illustrated in Table 1, Video Entity Segmentation presents distinct advantages compared to existing video segmentation paradigms: Unlike Video Object Segmentation, it operates automatically and accommodates multiple dynamic objects. In contrast to Video Instance Segmentation, it extends beyond foreground objectness to include background entities. Compared to Video Semantic Segmentation, it provides instance-level annotations. Unlike Video Panoptic Segmentation, it adopts a class-agnostic approach that eliminates the constraint of pre-defined category vocabularies. This unified entity representation allows both thing and stuff categories to be learned jointly and subsequently classified using external Vision Language Models, enabling advanced open-vocabulary evaluation and reasoning-based segmentation tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 474, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 474, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 474, + 647 + ], + "type": "text", + "content": "3.2. Method Background: SAM 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "SAM 2 [39] extends the Segment Anything model [22] to encompass both image and video domains. Through an iterative process of data annotation and model training, SAM 2 leverages both manual and automated datasets generated by its data engine. For video segmentation functionality," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24236" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 57, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 56, + 57, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 56, + 57, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 57, + 553, + 186 + ], + "type": "image", + "image_path": "fb6ed7ec0b4a7475fdc8f8a5abb4e21daa7ad49e64373544165aed1ea7bb37e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 191, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 54, + 191, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 191, + 555, + 237 + ], + "type": "text", + "content": "Figure 2. (a) Overview of the EntitySAM framework. EntitySAM utilizes the frozen encoder and memory parameters from SAM 2, incorporating a dual encoder design for enhanced semantic features (Section 3.3.1). The PromptGenerator (Section 3.3.2) automatically generates prompts from Prompt Queries. The enhanced features and distinct query groups are processed by the EntityDecoder (Section 3.3.3) to produce video mask outputs. (b) Self-attention and cross-attention mechanisms in EntityDecoder layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "type": "text", + "content": "SAM 2 incorporates a classical memory mechanism to facilitate object tracking." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 268, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 295, + 422 + ], + "type": "text", + "content": "The architecture processes video input through multiple stages: initially, a multi-scale Image encoder extracts frame-level features, which then undergo cross-attention with previous video-level memory features to derive per-mask temporal information. Concurrently, input prompts are processed through a dedicated prompt encoder to generate prompt encodings. These various features and prompts are integrated through cross-attention mechanisms and dynamic convolution operations within the mask decoder to generate mask outputs. The model maintains temporal coherence by encoding and storing mask sequences in a memory bank, which updates frame-by-frame and feeds into memory attention for processing subsequent frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 425, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 557 + ], + "type": "text", + "content": "Through extensive training on 35M manual and automated mask annotations, SAM 2 shows exceptional zero-shot performance across various video object segmentation benchmarks, including DAVIS [35], YTVOS [46], LVOS [16], and MOSE [12]. However, SAM 2 exhibits notable limitations: it requires initial mask/prompt specification at the sequence start and cannot effectively accommodate new objects introduced during tracking. Moreover, the model's efficiency deteriorates when handling large numbers of concurrent objects, indicating challenges in scaling to more complex scenarios with multiple dynamic objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 567, + 132, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 132, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 132, + 579 + ], + "type": "text", + "content": "3.3. EntitySAM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "Coupled with the proposed Video Entity Segmentation task, we present EntitySAM, a novel model that achieves strong zero-shot video entity segmentation performance while requiring only image-level class-agnostic entity annotations. As illustrated in Figure 2, our model preserves the memory-based architecture of the original SAM by freezing both the pretrained image encoder and memory-related modules, thus retaining its advantages in long-term tracking of objects. Meanwhile, EntitySAM comprises three learnable components: Dual Visual Encoder, EntityPromptGenerator," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 266 + ], + "type": "text", + "content": "and EntityDecoder. The following sections will detail these architectural components." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 273, + 436, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 436, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 436, + 283 + ], + "type": "text", + "content": "3.3.1 Dual Visual Encoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 293, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 293, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 555, + 399 + ], + "type": "text", + "content": "Our EntitySAM leverages the frozen Segment Anything Model 2 (SAM 2) image encoder as its backbone, as well as its frozen memory encoder and memory attention layers. The frozen design preserves SAM 2's zero-shot feature extraction capabilities, which were developed through training on extensive video object tracking datasets. Additionally, freezing the memory encoder substantially reduces GPU memory requirements during the training of multiple dense entity tracking in video sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 532 + ], + "type": "text", + "content": "Although SAM 2 exhibits impressive performance across multi-granularity class-agnostic segmentation tasks, we reveal its limitations in maintaining consistent entity-level representations as in Table 9. Specifically, SAM 2 struggles to generate reliable confidence scores for entity classification, where an 'entity' refers to a semantically coherent group. Thus in addition to the original SAM 2's backbone feature, we integrate the backbone features from DINOv2 [32], which produces consistent feature representations for semantically similar regions by large-scale self-supervised training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 532, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 628 + ], + "type": "text", + "content": "As illustrated in Figure 2, EntitySAM incorporates DINOv2 through a linear projection layer, aligning its feature dimensionality with that of SAM 2. The aligned features are then concatenated and fed into the decoder to form a fused representation. This enhances the encoded visual representation by preserving the zero-shot capabilities for video entity segmentation while effectively mitigating SAM 2's limitations in entity-level understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 634, + 474, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 474, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 474, + 646 + ], + "type": "text", + "content": "3.3.2 Automatic Prompt Generator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "In SAM 2's architecture, the prompt encoder plays a crucial role in image-level segmentation, where input prompts, such as points and bounding boxes, are processed to generate prompt embeddings for decoder segmentation. However, in video-level operations, SAM 2 bypasses the prompt" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24237" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "encoder, focusing solely on object tracking using historical memory features. To extend the model's capability for new entity tracking in video sequences, EntitySAM incorporates prompt embeddings during video training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 230 + ], + "type": "text", + "content": "Our EntitySAM implements a prompt generator utilizing learnable queries that perform cross-attention with enhanced visual features to predict input prompts for the prompt encoder. The resulting prompt embeddings are subsequently fed into the decoder at each frame to localize both existing and novel entities within the video sequence. Concurrently, this module predicts an entity confidence score, denoted as " + }, + { + "bbox": [ + 55, + 122, + 295, + 230 + ], + "type": "inline_equation", + "content": "S(\\text{entity})" + }, + { + "bbox": [ + 55, + 122, + 295, + 230 + ], + "type": "text", + "content": ", which performs binary classification between foreground entities and background regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": "More formally, let " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " represent the learnable queries and " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\in \\mathbb{R}^{HW \\times D}" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " denote the enhanced visual features, where " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " is the number of queries, " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " is the feature dimension, and " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " are spatial dimensions. The prompt generation and scoring process can be expressed as: " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\text{CrossAttention}(\\mathcal{Q}, \\mathcal{F})" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\text{FC}_{\\text{prompt}}(A)" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\text{FC}_{\\text{score}}(A)" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " is the cross-attention output, " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N \\times D'}" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " represents the generated prompts with " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " matching the expected prompt dimension required by the prompt encoder, and " + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{S} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 232, + 296, + 363 + ], + "type": "text", + "content": " denotes the entity confidence scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 180, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 180, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 180, + 390 + ], + "type": "text", + "content": "3.3.3 Entity Mask Decoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "text", + "content": "The Entity Mask Decoder predicts output masks based on enhanced features in Section 3.3.1 and the learnable prompts through a multi-stage processing pipeline in Section 3.3.2. It employs two groups of batched queries: mask queries " + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_m\\in \\mathbb{R}^{N\\times 1\\times C}" + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "text", + "content": " and IoU queries " + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{iou}\\in \\mathbb{R}^{N\\times 1\\times C}" + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "text", + "content": " denotes the number of entities and " + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 401, + 296, + 581 + ], + "type": "text", + "content": " represents the feature dimension. These queries are concatenated with input prompts for self-attention to facilitate inter-entity communication. These queries are subsequently reshaped for cross-attention operations between token and image representations. In the final layer, mask queries undergo dynamic convolution with upsampled mask features to generate output masks, while IoU queries are processed through a dedicated prediction head to estimate conditioned IoU scores. The computational flow can be formalized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 593, + 294, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 593, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 64, + 593, + 294, + 665 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} [ \\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime} ] = \\text {S e l f A t t e n t i o n} (\\text {c o n c a t} (\\mathcal {Q} _ {m}; \\mathcal {Q} _ {i o u})) \\\\ w h e r e \\operatorname {c o n c a t} \\left(\\mathcal {Q} _ {m}; \\mathcal {Q} _ {i o u}\\right) \\in \\mathbb {R} ^ {1 \\times 2 N \\times C} \\\\ F _ {o u t} = \\operatorname {C r o s s A t t e n t i o n} (\\operatorname {r e s h a p e} \\left(\\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime}\\right), F) \\\\ \\operatorname {w h e r e} \\operatorname {r e s h a p e} \\left(\\mathcal {Q} _ {m} ^ {\\prime}, \\mathcal {Q} _ {i o u} ^ {\\prime}\\right) \\in \\mathbb {R} ^ {N \\times 2 \\times C}. \\tag {5} \\\\ \\end{array}", + "image_path": "a8dc6461fb7bf5bdd994ffb706ed1d83233a10c789b317e89c980e2f46769a57.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "The supervision mechanism integrates outputs from both the prompt generator (entity scores) and mask decoder (masks and conditioned IoU predictions) through Hungarian matching with ground truth annotations. The expected" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "type": "text", + "content": "IoU for query " + }, + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 91, + 553, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 553, + 126 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E (\\mathrm {I o U} _ {q}) = E (\\mathrm {I o U} _ {q} | \\mathrm {E n t} _ {q}) S (\\mathrm {E n t} _ {q}) + E (\\mathrm {I o U} _ {q} | \\neg \\mathrm {E n t} _ {q}) S (\\neg \\mathrm {E n t} _ {q}) \\\\ = E \\left(\\operatorname {I o U} _ {q} \\mid \\operatorname {E n t} _ {q}\\right) S \\left(\\operatorname {E n t} _ {q}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "196a316585ce84b30e5316b9d0006e8e849a1317396aec9accc2e2cba4fd7909.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 128, + 553, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 553, + 151 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 128, + 553, + 151 + ], + "type": "inline_equation", + "content": "S(\\mathrm{Ent}_q)" + }, + { + "bbox": [ + 313, + 128, + 553, + 151 + ], + "type": "text", + "content": " represents the entity confidence score from PromptGenerator in Section 3.3.2 and " + }, + { + "bbox": [ + 313, + 128, + 553, + 151 + ], + "type": "inline_equation", + "content": "E(\\mathrm{IoU}_q|\\neg \\mathrm{Ent}_q) = 0" + }, + { + "bbox": [ + 313, + 128, + 553, + 151 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 152, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 258 + ], + "type": "text", + "content": "The resulting output masks are processed through the frozen memory encoder, memory bank, and memory attention mechanisms, with gradient computation detached for memory-related components. This enables efficient training of the queries-based tracking component while maintaining frozen memory modules. Consequently, our approach achieves scalable training for over 50 entities in 8-frame video clips, significantly surpassing SAM 2's original capacity of handling only 3 video objects during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 273, + 545, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 545, + 285 + ], + "type": "text", + "content": "3.3.4 Training Loss and Open-Vocabulary Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 292, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 314 + ], + "type": "text", + "content": "We employ an end-to-end training objective for EntitySAM, formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 324, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 324, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 365, + 324, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {M a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\text {E n t i t y}} + \\lambda_ {2} \\mathcal {L} _ {\\text {I o U}}, \\tag {7}", + "image_path": "2196b94c3f62142fcd0abfd92169b5ef5d42fb87c789dd993b03234b2c3a5899.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "text", + "content": " supervises both mask prediction and binary entity classification, following the Mask2former framework [7]. The hyperparameters " + }, + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 346, + 553, + 393 + ], + "type": "text", + "content": " balance the respective loss terms and are empirically set to 1.0 and 1.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 394, + 554, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 549 + ], + "type": "text", + "content": "During inference, EntitySAM accommodates both class-agnostic and class-specific evaluation protocols. For class-agnostic scenarios, entity masks are directly generated and merged based on confidence scores and mask overlap metrics. In class-specific evaluation, we select the frame with the highest entity confidence score for each masklet prediction. The predicted entity masks are utilized to crop corresponding regions of interest, which, together with carefully designed prompts containing candidate category names, are processed through GPT-4o [1] for final classification. We demonstrate in the supplementary materials that this mask-guided cropping approach achieves superior classification performance compared to alternative methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 555, + 395, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 395, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 395, + 569 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 575, + 430, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 430, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 430, + 588 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Datasets Video entity segmentation typically requires extensive class-agnostic dense entity video segmentation annotations. Although it eliminates the need for class-specific annotations, creating frame-by-frame annotations remains labor-intensive. Inspired by [51, 52], our training pipeline leverages COCO [23] panoptic segmentation data without category label annotation (denoted as " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO}^*" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ") and extends it to pseudo-video sequences. Our approach utilizes only image-level annotations, eschewing video-level labels, to achieve effective video entity segmentation. The pipeline" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24238" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "also enables zero-shot cross-dataset evaluation across multiple benchmarks. We demonstrate our model's versatility through comprehensive zero-shot evaluations on various tasks: video panoptic segmentation on VIPSeg [28], video semantic segmentation on VSPW [27], video instance segmentation on Youtube-VIS 2019 [48], Youtube-VIS 2021 [48], and OVIS [36], and video entity segmentation on our proposed VIPEntitySeg dataset—a class-agnostic variant of VIPSeg. Notably, all zero-shot evaluations are performed using a single trained model, demonstrating its robust zero-shot generalization capabilities across different video segmentation tasks and datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "text", + "content": "Implementation details In our experimental setup, we employ two backbone architectures: ViT-Small and ViT-Large. The models are trained using the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "text", + "content": " and weight decay of 0.05. Input images are cropped to " + }, + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "text", + "content": " pixels for compatibility with SAM 2 ViT/16 and " + }, + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "inline_equation", + "content": "896 \\times 896" + }, + { + "bbox": [ + 56, + 217, + 298, + 469 + ], + "type": "text", + "content": " pixels for DINOv2 ViT/14. Training is conducted with a batch size of 8. The loss function combines multiple components: cross-entropy and dice losses for mask prediction, L1 loss for IoU prediction, and focal loss for binary entity classification. Our training strategy consists of two phases: initial training on static COCO images for 40K iterations, followed by training on COCO pseudo-video sequences with clip length 8 for 10K iterations. Both stages are class-agnostic. During inference, we utilize a single entity category output mode with confidence scoring. Video entity quality (VEQ) is evaluated in a class-agnostic manner. For class-specific evaluation tasks, we implement a multi-step process: first selecting the highest-confidence mask from the masklet for each prediction, then extracting the corresponding frame region within the mask, and finally employing GPT-4o for classification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 479, + 185, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 185, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 185, + 491 + ], + "type": "text", + "content": "4.2. Zero-shot Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "Zero-shot Video Entity Segmentation In Table 2 (left), we evaluate our approach on VIPEntitySeg using class-agnostic metrics against several baseline methods. For SAM 2, we adopt two prompting manners, a grid point prompt sampling initialization and a COCO-trained Mask2Former [7] mask prompts initialization. We additionally compare against state-of-the-art zero-shot segmentation models, including: 1) FC-CLIP [51] integrated with MinVIS [17] association, 2) DEVA [11] coupled with COCO-trained Mask2Former [7] for per-frame segmentation, and 3) OV-DVIS++ [52] in both online and offline modes. All these SOTA models were trained on COCO with category annotations, with DEVA utilizing additional external datasets for its association model. We compare with these methods on our class-agnostic VIPEntitySeg benchmark. Our EntitySAM achieves significant performance improvements over all baselines and SOTA models with both ViT-Small and ViT-Large backbones, despite training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "without category annotations. This superior performance demonstrates the effectiveness of our class-agnostic training approach and proposed modules. Notably, we observe substantial improvements in " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{VEQ}^{RQ}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", indicating enhanced cross-frame entity identification and reduced false positives. Figure 3 further illustrates our model's robustness by evaluating VEQ across varying frame intervals, demonstrating that EntitySAM maintains consistent performance over extended temporal sequences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 289 + ], + "type": "text", + "content": "Regarding the class-agnostic " + }, + { + "bbox": [ + 313, + 180, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{STQ}^{EN}" + }, + { + "bbox": [ + 313, + 180, + 553, + 289 + ], + "type": "text", + "content": " metric, which emphasizes tracking performance, DEVA achieves higher scores through its tracking association model trained on external video object segmentation datasets. However, DEVA relies on a two-stage pipeline of detection followed by association, lacking end-to-end integration. While our approach achieves lower scores, it delivers comparable performance through a streamlined process that requires only COCO training data and single-pass inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 291, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 483 + ], + "type": "text", + "content": "Zero-shot Video Panoptic Segmentation In Table 2 (right), we extend our evaluation to VIPSeg for class-specific video panoptic segmentation. Several baseline models, including Mask2Former and DEVA, are constrained to VIPSeg evaluation due to their closed-vocabulary nature. When compared to open-vocabulary models such as FC-CLIP and OV-DVIS++, our approach demonstrates substantial improvements, particularly with the ViT-Small backbone. These results validate our approach of \"post-classification with multimodal models\", indicating that it surpasses traditional open-vocabulary segmentation training pipelines despite operating without category annotations. Our EntitySAM achieves competitive performance on both class-agnostic and class-specific tasks, suggesting that category annotations may not be essential for effective open-vocabulary video segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 484, + 553, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 553, + 580 + ], + "type": "text", + "content": "Zero-shot Video Semantic Segmentation Table 3 presents our results on video semantic segmentation, obtained by merging per-pixel mask predictions. Our method achieves improvements in both mean Intersection over Union (mIoU) and Video Consistency (VC) metrics. The VSPW evaluation dataset, which contains the same video content but with extended frame sequences, validates our model's capability for effective long-term tracking." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Zero-shot Video Instance Segmentation As shown in Table 4, we evaluate EntitySAM on three popular video instance segmentation datasets. Our approach on extending video entity segmentation to video instance segmentation requires no model retraining—we simply add a background/stuff class during inference and filter out predictions with that class to obtain instance results. Our model demonstrates significant performance improvements over state-of-the-art open-vocabulary video instance segmentation models. In particular, on the challenging OVIS dataset, which features frequent occlusions, our model maintains robust" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24239" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 555, + 276 + ], + "blocks": [ + { + "bbox": [ + 55, + 56, + 555, + 101 + ], + "lines": [ + { + "bbox": [ + 55, + 56, + 555, + 101 + ], + "spans": [ + { + "bbox": [ + 55, + 56, + 555, + 101 + ], + "type": "text", + "content": "Table 2. Comparison of Zero-shot Video Entity Segmentation (left) and Video Panoptic Segmentation (right) on COCO " + }, + { + "bbox": [ + 55, + 56, + 555, + 101 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 56, + 555, + 101 + ], + "type": "text", + "content": " VIPSeg evaluation. EntitySAM is trained on the COCO dataset without category annotations. In VIPSeg, - means the results of open-vocabulary evaluation for closed-vocabulary models are not available. * denotes that DEVA is trained with multiple external video object segmentation datasets for the temporal association model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 276 + ], + "type": "table", + "html": "
MethodBackboneClass Agnostic TrainingVIPEntitySegVIPSeg
VEQVEQSQVEQRQSTQENVPQVPQThVPQStSTQ
SAM 2.1 [39] (Everything Init)ViT-S✓26.982.532.638.7-
SAM 2.1 [39] (Mask2former Init)ViT-S + ResNet5046.283.655.234.5-
DEVA* [11]ResNet-5049.083.658.642.4-
FC-CLIP [51]ResNet-5041.584.049.429.722.325.519.119.7
OV-DVIS++(Online) [52]ResNet-5045.983.754.833.624.426.822.422.0
OV-DVIS++(Offline) [52]ResNet-5046.782.856.335.623.826.421.424.4
OursViT-S✓51.884.761.141.028.732.925.131.4
SAM 2.1 [39] (Everything Init)ViT-L✓30.583.136.741.7-
SAM 2.1 [39] (Mask2former Init)ViT-L + Swin-L50.684.460.039.1-
DEVA* [11]Swin-L50.883.860.645.3-
FC-CLIP [51]ConvNext-L43.484.851.232.427.930.925.024.2
OV-DVIS++(Online) [52]ConvNext-L49.483.559.237.928.931.326.828.4
OV-DVIS++(Offline) [52]ConvNext-L49.083.159.040.030.431.929.132.2
OursViT-L✓54.684.764.543.331.437.526.033.5
", + "image_path": "5da507146307bb83555f58559ecf5d2ca8f1c1bb07e5eceab35918957d1679ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 315, + 295, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 280, + 295, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 280, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 295, + 313 + ], + "type": "text", + "content": "Table 3. Zero-Shot Open-vocabulary Video Semantic Segmentation on COCO " + }, + { + "bbox": [ + 55, + 280, + 295, + 313 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 280, + 295, + 313 + ], + "type": "text", + "content": " VSPW. EntitySAM is trained on the COCO dataset without category annotations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 315, + 295, + 404 + ], + "lines": [ + { + "bbox": [ + 56, + 315, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 295, + 404 + ], + "type": "table", + "html": "
MethodBackboneClass Agnostic TrainingmVC8VSPW mVC16mIOU
FC-CLIP [51]ResNet-5084.982.724.3
OV-DVIS++(Online) [52]ResNet-5092.791.527.6
OV-DVIS++(Offline) [52]ResNet-5092.491.328.4
OursViT-S✓94.093.034.6
FC-CLIP [51]ConvNext-L89.988.428.9
OV-DVIS++(Online) [52]ConvNext-L94.293.334.3
OV-DVIS++(Offline) [52]ConvNext-L93.993.034.1
OursViT-L✓94.693.735.5
", + "image_path": "510c47af21cd23436b03f2987b081f2c5ac8840a264e858d5619f36250961e53.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 441 + ], + "type": "text", + "content": "performance, benefiting from the entity learning strategy employed during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 181, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 181, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 181, + 456 + ], + "type": "text", + "content": "4.3. Ablation Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "type": "text", + "content": "We perform ablation studies using the ViT-Small backbone on the COCO " + }, + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "type": "text", + "content": " VIPEntitySeg video entity segmentation task with zero-shot protocol to evaluate the effectiveness of our proposed modules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": "Ablation on the Proposed Modules Table 5 presents ablation studies of our proposed modules. Our baseline, SAM 2 with memory-based tracking modules, cannot generate predictions automatically without appropriate prompting. Using grid-point sampling leads to numerous false positive predictions, achieving a VEQ of only 26.9. Removing the memory design and implementing solely a query-based EntityDecoder with query association for tracking improves performance to 41.3, though this remains suboptimal. Combining memory-based and query-based approaches significantly enhances performance to 50.2, surpassing both DEVA and OV-DVIS++. Finally, incorporating a complementary DINOv2 dual feature encoder achieves the best result (51.8) for the ViT-Small backbone." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Ablation on Query Design As illustrated in Figure 2, our design incorporates three query types: two groups in EntityDecoder and one group in PromptGenerator. Our abla" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 280, + 425, + 368 + ], + "blocks": [ + { + "bbox": [ + 316, + 280, + 425, + 368 + ], + "lines": [ + { + "bbox": [ + 316, + 280, + 425, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 425, + 368 + ], + "type": "image", + "image_path": "faff9aa980542777667efb97d103d049495e6848ff1377b97d5ecac62b7451a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 374, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 397 + ], + "type": "text", + "content": "Figure 3. VEQ comparison with different evaluation time intervals. EntitySAM improves over long intervals." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 428, + 280, + 553, + 368 + ], + "blocks": [ + { + "bbox": [ + 428, + 280, + 553, + 368 + ], + "lines": [ + { + "bbox": [ + 428, + 280, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 428, + 280, + 553, + 368 + ], + "type": "image", + "image_path": "4298e2160ec64bc276eea6e068907ffaf70a6180032fc2487f0d8276ec0248ba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 409, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 529 + ], + "type": "text", + "content": "tion experiment of query design in Table 6 reveals that using a single query group to predict all outputs (mask, IoU, and confidence score) yields suboptimal performance with a VEQ of 49.2. Adding a second group of prompt queries improves performance to 49.7. Further separation into a third query group for distinct mask and IoU prediction achieves the best performance of 51.8. These results suggest that different query groups learn specialized distributions for their respective prediction tasks, departing from traditional detection transformers' shared-query architecture." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 531, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 602 + ], + "type": "text", + "content": "Ablation on EntityDecoder Depth Table 7 studies the influence of the entity decoder's depth layer. For the ViT-Small backbone, a decoder depth of 4 achieves optimal performance with a VEQ of 51.8 on COCO " + }, + { + "bbox": [ + 313, + 531, + 554, + 602 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 531, + 554, + 602 + ], + "type": "text", + "content": " VIPSeg setting. Further increasing decoder depth shows no additional performance benefits." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Ablation on Training Stages As detailed in Section 4.1, our training process consists of two stages: initial training on single images (treated as zero-length videos) followed by training on pseudo video clips of length 8. During the first stage, the memory component remains unused while the EntityDecoder learns to identify new entities in each image. Table 8 shows that this image-only training achieves limited video entity segmentation performance with a VEQ of 31.4. However, this stage exposes the model to diverse en" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24240" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 88, + 555, + 163 + ], + "blocks": [ + { + "bbox": [ + 55, + 63, + 555, + 86 + ], + "lines": [ + { + "bbox": [ + 55, + 63, + 555, + 86 + ], + "spans": [ + { + "bbox": [ + 55, + 63, + 555, + 86 + ], + "type": "text", + "content": "Table 4. Zero-Shot Open-vocabulary Video Instance Segmentation on COCO " + }, + { + "bbox": [ + 55, + 63, + 555, + 86 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 63, + 555, + 86 + ], + "type": "text", + "content": " Validation sets of YouTube-VIS 2019, 2021, and OVIS. EntitySAM is trained on the COCO dataset without category annotations." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 88, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 56, + 88, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 555, + 163 + ], + "type": "table", + "html": "
MethodBackboneClass Agnostic TrainingYoutube-VIS 2019Youtube-VIS 2021OVIS
AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)
Detric[53]-OWTB[25]ResNet-5017.9----16.7----9.0----
MindVLT[41]ResNet-5023.1----20.9----11.4----
FC-CLIP [51]ResNet-5028.943.929.932.441.225.540.026.828.037.011.825.110.58.516.4
OV-DVIS++(Online) [52]ResNet-5034.550.639.239.549.530.946.734.834.445.814.831.213.110.524.7
OV-DVIS++(Offline) [52]ResNet-5034.451.839.240.050.631.048.534.334.646.813.028.410.99.825.1
OursViT-S✓44.960.749.543.662.438.852.543.136.256.521.035.820.812.336.0
", + "image_path": "c29258108f6ef968b52fb15b9c62d94b3ded3902c57c2b7310e6421fe97b7b82.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 180, + 294, + 256 + ], + "blocks": [ + { + "bbox": [ + 83, + 167, + 268, + 178 + ], + "lines": [ + { + "bbox": [ + 83, + 167, + 268, + 178 + ], + "spans": [ + { + "bbox": [ + 83, + 167, + 268, + 178 + ], + "type": "text", + "content": "Table 5. Ablation Study on the Proposed Modules." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 180, + 294, + 256 + ], + "lines": [ + { + "bbox": [ + 56, + 180, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 294, + 256 + ], + "type": "table", + "html": "
ModelMemoryEntityDecoderDual Visual EncoderVEQ
SAM 2 [39]✓26.9
Ours✓✓41.3
✓✓50.2
✓✓✓51.8
", + "image_path": "b3503f943d827de1dead3578b761f94ace03c5a033c51fe383287289bfb56977.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 272, + 294, + 334 + ], + "blocks": [ + { + "bbox": [ + 97, + 258, + 252, + 270 + ], + "lines": [ + { + "bbox": [ + 97, + 258, + 252, + 270 + ], + "spans": [ + { + "bbox": [ + 97, + 258, + 252, + 270 + ], + "type": "text", + "content": "Table 6. Ablation Study on Query Design." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 272, + 294, + 334 + ], + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 334 + ], + "type": "table", + "html": "
EntityDecoderPromptGeneratorCOCO → VIPSeq
VEQVEQSQVEQRQ
Shared Queries49.284.058.6
Shared Queries✓49.784.259.0
Separate Queries✓51.884.761.1
", + "image_path": "7847958b38175703587365d3369a2e006363c3d22ed2090e0c2353702def00e4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 357, + 294, + 434 + ], + "blocks": [ + { + "bbox": [ + 85, + 344, + 266, + 355 + ], + "lines": [ + { + "bbox": [ + 85, + 344, + 266, + 355 + ], + "spans": [ + { + "bbox": [ + 85, + 344, + 266, + 355 + ], + "type": "text", + "content": "Table 7. Ablation Study on EntityDecoder Depth." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 357, + 294, + 434 + ], + "lines": [ + { + "bbox": [ + 56, + 357, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 294, + 434 + ], + "type": "table", + "html": "
DecoderDecoder depthCOCO → VIPSeg
VEQVEQSQVEQRQ
EntityDecoder251.384.760.5
451.884.761.1
651.684.661.0
851.184.460.6
", + "image_path": "66e6605ad3899d12f0d6a4ea3389ac552ad0e4b7daf74ae50d1daac951b37e3f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 447, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 508 + ], + "type": "text", + "content": "tities, and when combined with the subsequent video training stage—which aligns EntityDecoder with memory encoders—performance improves substantially to 51.8 VEQ. This two-stage approach outperforms training exclusively on pseudo video sequences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Ablation on Dual Encoder In Section 3.3.1, we introduced encoder feature fusion designs for dual encoders. As shown in Table 9, our ablation studies reveal that incorporating features from either CLIP [38] or SAM [22] alongside the SAM 2 encoder features yields no improvement in zero-shot performance. This can be attributed to CLIP's vision encoder being optimized for classification rather than segmentation tasks, resulting in representations less suitable for segmentation. Similarly, SAM, being trained with an approach analogous to SAM 2, provides features that largely overlap with SAM 2's capabilities rather than offering complementary information. In contrast, DINOv2, trained through self-supervision, demonstrates significant benefits when combined with SAM 2. Its self-supervised training enables DINOv2 to learn consistent image representations that effectively complement SAM 2's supervised features, ultimately enhancing the overall performance." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 180, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 354, + 167, + 514, + 178 + ], + "lines": [ + { + "bbox": [ + 354, + 167, + 514, + 178 + ], + "spans": [ + { + "bbox": [ + 354, + 167, + 514, + 178 + ], + "type": "text", + "content": "Table 8. Ablation Study on Training stages." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 180, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 315, + 180, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 315, + 180, + 553, + 247 + ], + "type": "table", + "html": "
ModelImagePseudo VideoCOCO → VIPSeg
VEQVEQSQVEQRQ
Ours✓31.479.439.5
✓49.284.258.4
✓✓51.884.761.1
", + "image_path": "ee8aa851fed3ac9f1bb0a01fbe378e7e697659ecf1992179902157d4e079a77f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 264, + 553, + 341 + ], + "blocks": [ + { + "bbox": [ + 343, + 251, + 525, + 262 + ], + "lines": [ + { + "bbox": [ + 343, + 251, + 525, + 262 + ], + "spans": [ + { + "bbox": [ + 343, + 251, + 525, + 262 + ], + "type": "text", + "content": "Table 9. Ablation Study on Dual Encoder Feature." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 264, + 553, + 341 + ], + "lines": [ + { + "bbox": [ + 315, + 264, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 553, + 341 + ], + "type": "table", + "html": "
Encoder FeatureDual EncoderCOCO → VIPSeg
VEQVEQSQVEQRQ
SAM 2-50.284.559.4
SAM [22]47.684.256.5
CLIP [38]49.284.058.6
DINOv2 [32]51.884.761.1
", + "image_path": "2263415b5bbfe8948e49f88974c7dd80a5846d977bb235645c22a6a6393c6dbe.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 531, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 531, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 531, + 367 + ], + "type": "text", + "content": "4.4. Qualitative experiments and visualization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 373, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 421 + ], + "type": "text", + "content": "Figure 4 demonstrates our model's performance on zero-shot video sequences, showcasing consistent entity tracking and segmentation across frames. Additional qualitative results are provided in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 426, + 395, + 516 + ], + "blocks": [ + { + "bbox": [ + 315, + 426, + 395, + 516 + ], + "lines": [ + { + "bbox": [ + 315, + 426, + 395, + 516 + ], + "spans": [ + { + "bbox": [ + 315, + 426, + 395, + 516 + ], + "type": "image", + "image_path": "971a321b7254a0fef2a56a5217755b09d596e0f91ed9a1790494d156c4a0ff8e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 517, + 547, + 529 + ], + "lines": [ + { + "bbox": [ + 320, + 517, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 320, + 517, + 547, + 529 + ], + "type": "text", + "content": "Figure 4. Visualization of zero-shot video entity segmentation." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 395, + 426, + 474, + 516 + ], + "blocks": [ + { + "bbox": [ + 395, + 426, + 474, + 516 + ], + "lines": [ + { + "bbox": [ + 395, + 426, + 474, + 516 + ], + "spans": [ + { + "bbox": [ + 395, + 426, + 474, + 516 + ], + "type": "image", + "image_path": "b224e69f03d2ef284977ad493ef40659edb97820e6316c8007ed3d070f783760.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 474, + 426, + 553, + 516 + ], + "blocks": [ + { + "bbox": [ + 474, + 426, + 553, + 516 + ], + "lines": [ + { + "bbox": [ + 474, + 426, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 474, + 426, + 553, + 516 + ], + "type": "image", + "image_path": "55082acb74841d8e6cd604a3be21c25b6e52b401b5ba35cb27a125feb2afdd45.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "content": "We present EntitySAM, a framework that integrates query and memory mechanisms to enhance SAM 2 for zero-shot video entity segmentation. Our approach eliminates the need for user-defined prompt inputs and enables automatic tracking of new entities across frames. Trained exclusively on COCO image-level mask annotations without category information, EntitySAM achieves superior performance on the Video Entity Segmentation task. The model demonstrates strong generalization capabilities across various zero-shot evaluations, including Video Panoptic / Semantic / Instance Segmentation. These results provide valuable insights into class-agnostic training approaches and their effectiveness in zero-shot scenarios." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24241" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "text", + "content": "[2] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 236 + ], + "type": "text", + "content": "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "type": "text", + "content": "[4] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. In arXiv preprint arXiv:1905.00737, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 283, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 295, + 327 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 328, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 372 + ], + "type": "text", + "content": "[6] Bowen Cheng, Anwesa Choudhuri andd Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. In https://arxiv.org/abs/2112.10764, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 295, + 417 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 418, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 295, + 452 + ], + "type": "text", + "content": "[8] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 453, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 486 + ], + "type": "text", + "content": "[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 487, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 295, + 520 + ], + "type": "text", + "content": "[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In arXiv, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 521, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 554 + ], + "type": "text", + "content": "[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "text", + "content": "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, 2023. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "type": "text", + "content": "[13] Yuming Du, Wen Guo, Yang Xiao, and Vincent Lepetit. Uvo challenge on video-based open-world segmentation 2021: 1st place solution. ICCV Workshop, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "type": "text", + "content": "[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. Advances in Neural Information Processing Systems, 35:23109–23120, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[15] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14623-14632, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[16] Lingyi Hong, Wenchao Chen, Zhongying Liu, Wei Zhang, Pinxue Guo, Zhaoyu Chen, and Wenqiang Zhang. Lvos: A benchmark for long-term video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13480-13492, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "text", + "content": "[17] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "text", + "content": "[18] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "text", + "content": "[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 251, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 272 + ], + "type": "text", + "content": "[20] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "type": "text", + "content": "[21] Dahun Kim, Jun Xie, Huiyu Wang, Siyuan Qiao, Qihang Yu, Hong-Seok Kim, Hartwig Adam, In So Kweon, and Liang-Chieh Chen. Tubeformer-deeplab: Video mask transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13914-13924, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 3, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "text", + "content": "[24] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 449, + 553, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 482 + ], + "type": "text", + "content": "[25] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 483, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 553, + 514 + ], + "type": "text", + "content": "[26] Qi Lu, Jason Kuen, Shen Tiancheng, Gu Jiumiang, Guo Weidong, Jia Jiaya, Lin Zhe, and Yang Ming-Hsuan. High-quality entity segmentation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "text", + "content": "[27] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. Vspw: A large-scale dataset for video scene parsing in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4133-4143, 2021. 1, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "text", + "content": "[28] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "text", + "content": "[29] Xindong Zhang Minghan Li, Shuai Li and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[30] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6819-6828, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24242" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 171 + ], + "type": "text", + "content": "[32] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "text", + "content": "[33] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[34] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[35] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "type": "text", + "content": "[36] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation. IJCV, 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[37] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. In arXiv preprint arXiv:2107.14228, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 436 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 514 + ], + "type": "text", + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[40] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3126-3137, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In proceedings of the IEEE/CVF international conference on computer vision, pages 4057-4066, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "text", + "content": "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[43] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In European Conference on Computer Vision, pages 553-569. Springer, 2022." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 475 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[44] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[45] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In European Conference on Computer Vision, pages 588-605. Springer, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[46] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. In ECCV, 2018. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[47] Xiaohao Xu, Jinglu Wang, Xiao Li, and Yan Lu. Reliable propagation-correction modulation for video object segmentation. In AAAI, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 251 + ], + "type": "text", + "content": "[48] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[49] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[50] Zongxin Yang, Jiaxu Miao, Yunchao Wei, Wenguan Wang, Xiaohan Wang, and Yi Yang. Scalable video object segmentation with identification mechanism. TPAMI, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "text", + "content": "[51] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "text", + "content": "[52] Tao Zhang, Xingye Tian, Yikang Zhou, Shunping Ji, Xuebo Wang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, and Yu Wu. Dvis++: Improved decoupled framework for universal video segmentation. arXiv preprint arXiv:2312.13305, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "text", + "content": "[53] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24243" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_content_list.json b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1cb0732bc93895093034097573015c0ee4af1693 --- /dev/null +++ b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_content_list.json @@ -0,0 +1,1950 @@ +[ + { + "type": "text", + "text": "EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Sun $^{1,2}$ , Rui Wang $^{1,2*}$ , Zixuan Zhu $^{1,2}$ , Lihua Jing $^{1,2}$ , Yuanfang Guo $^{3}$", + "bbox": [ + 212, + 202, + 779, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences", + "bbox": [ + 218, + 220, + 772, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Cyber Security, University of Chinese Academy of Sciences", + "bbox": [ + 214, + 238, + 782, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 School of Computer Science and Engineering, Beihang University", + "bbox": [ + 228, + 256, + 766, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nming, wangrui, zhuzixuan, jinglihua}@iie.ac.cn, andyguo@buaa.edu.cn", + "bbox": [ + 209, + 277, + 807, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality open-source datasets are essential for advancing deep neural networks. However, the unauthorized commercial use of these datasets has raised significant concerns about copyright protection. One promising approach is backdoor watermark-based dataset ownership verification (BW-DOV), in which dataset protectors implant specific backdoors into illicit models through dataset watermarking, enabling the tracing of these models through abnormal prediction behaviors. Unfortunately, the targeted nature of these BW-DOV methods can be maliciously exploited, potentially leading to harmful side effects. While existing harmless methods attempt to mitigate these risks, watermarked datasets can still negatively affect prediction results, partially compromising dataset functionality. In this paper, we propose a more harmless backdoor watermark, called EntropyMark, which improves prediction confidence without altering the final prediction results. For this purpose, an entropy-based constraint is introduced to regulate the probability distribution. Specifically, we design an iterative clean-label dataset watermarking framework. Our framework employs gradient matching and adaptive data selection to optimize backdoor injection. In parallel, we introduce a hypothesis test method grounded in entropy inconsistency to verify dataset ownership. Extensive experiments on benchmark datasets demonstrate the effectiveness, transferability, and defense resistance of our approach. Our code is available at https://github.com/AaronSun2000/EntropyMark.", + "bbox": [ + 86, + 358, + 483, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 808, + 220, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, deep neural networks (DNNs) have rapidly advanced and been widely deployed in numerous real-world applications, such as autonomous driving [8] and face", + "bbox": [ + 89, + 834, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c94024f018bb9f1eced0dac4a858983c4e562b75b100696ecf1becf9600b5a0.jpg", + "image_caption": [ + "I. Training", + "II. Inference", + "Figure 1. Comparison of EntropyMark with existing back-door watermarks: Existing watermarks alter the prediction results of backdoored models and produce conspicuous modifications on datasets. In contrast, EntropyMark preserves the original results and is significantly more stealthy." + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 903, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "recognition [39]. The prosperity of DNNs largely depends on the availability of high-quality datasets (e.g., CIFAR [18] and ImageNet [5]). Researchers use open-source datasets to evaluate and iteratively optimize their proposed methods. However, the value of these datasets has attracted dataset stealers who illegally copy, distribute and use them without permission. Currently, most open-source datasets are intended solely for educational or research purposes and explicitly prohibit unauthorized commercial use1. Therefore, copyright protection for open-source datasets has become an urgent and necessary topic.", + "bbox": [ + 509, + 657, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, classic data protection methods, including differential privacy [1, 44], encryption [4, 40], and digital watermarking [11, 15], are not well-suited for open-", + "bbox": [ + 511, + 825, + 903, + 871 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1See IMDb data policy for details.", + "bbox": [ + 531, + 886, + 715, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30692", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "source datasets, as they are primarily designed for private data protection. Safeguarding open-source datasets presents more challenges because their protectors lack control over the training process and have no access to training components. Instead, they can only query the outputs of suspicious models to determine whether those models were trained on the protected datasets. In addition, due to the open-source nature of these datasets, any protection strategy must avoid compromising dataset functionality or altering the integrity of information representation.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor watermark-based dataset ownership verification (BW-DOV) [26] offers a feasible solution for protecting open-source dataset copyrights. This approach involves modifying samples in the original dataset and releasing the watermarked version for open-source use. When a model is trained on this watermarked dataset, a specific backdoor is implanted. To verify whether a suspicious model was trained on the protected dataset, protectors can simply detect the presence of the predetermined backdoor.", + "bbox": [ + 89, + 244, + 480, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, due to the targeted nature (i.e., specified erroneous prediction results) of existing BW-DOV methods, watermarked models are vulnerable to manipulation and may maliciously output predetermined results. To address this issue, UBW [24] proposes an untargeted backdoor watermark, which mitigates the targeted nature of the backdoor by increasing prediction randomness on watermarked samples. Domain-Watermark [10] ensures that watermarked models correctly classify certain 'hard' samples that benign models may misclassify. However, these methods still alter the original prediction results, inevitably impacting dataset functionality. To address the limitations of existing BW-DOV methods, we argue that harmless backdoor watermarks should not only avoid generating specified erroneous prediction results but also preserve the original prediction outputs, thereby maintaining dataset functionality.", + "bbox": [ + 89, + 383, + 482, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose EntropyMark, a more harmless backdoor watermark designed to improve prediction confidence for verification samples without affecting overall prediction results. To ensure that the prediction probabilities for these verification samples are more concentrated on the ground-truth label, we introduce an entropy-based constraint to regulate the probability distribution. For implementation, we design an iterative clean-label dataset watermarking framework. Within this framework, we employ gradient matching to optimize training convergence, ensuring effective backdoor injection into models trained on the watermarked dataset. We also utilize gradient-based adaptive selection to identify more \"decisive\" samples for watermarking. Fig. 1 compares EntropyMark with existing BW-DOV methods, emphasizing the harmlessness and stealthiness of our method. Unlike radioactive data [31], the classic data tracing method through watermarking, EntropyMark exhibits greater robustness in black-box settings and offers", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enhanced resilience against various defenses. Finally, to further improve verification accuracy, we incorporate a pairwise T-test based on entropy inconsistency.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions can be summarized as follows:", + "bbox": [ + 532, + 137, + 893, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose EntropyMark, a backdoor watermark under an entropy-based constraint. Based on our backdoor watermark, we design an iterative clean-label dataset watermarking framework. EntropyMark achieves enhanced stealthiness and minimal harmfulness while retaining watermark effectiveness.", + "(2) Our framework adaptively selects \"decisive\" samples to be watermarked and incorporates gradient matching to enhance training convergence.", + "(3) Building on the properties of EntropyMark, we develop an entropy inconsistency hypothesis test to quantitatively verify dataset ownership.", + "(4) Extensive experiments on benchmark datasets validate the effectiveness, transferability, and defense resistance of our method." + ], + "bbox": [ + 501, + 152, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 395, + 653, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Backdoor Attacks", + "text_level": 1, + "bbox": [ + 513, + 421, + 687, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.1. Dirty-label Attacks", + "text_level": 1, + "bbox": [ + 513, + 444, + 692, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dirty-label attacks forcibly modify the labels of selected malicious samples. BadNets [9], the first dirty-label attack, uses a static patch as the trigger. To enhance stealthiness, Blend [3] generates malicious samples by blending the trigger with original images. Subsequently, more stealthy and effective methods [17, 21, 30] have been proposed, varying the forms of triggers. However, dirty-label attacks are still detectable due to label-semantic inconsistencies.", + "bbox": [ + 511, + 463, + 906, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.2. Clean-label Attacks", + "text_level": 1, + "bbox": [ + 513, + 594, + 696, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Clean-label attacks, which retain original sample labels, are well-suited for scenarios where attackers lack control over label settings. Label-Consistent [38] introduces the first patch-based attack under this paradigm. However, many early clean-label attacks [2, 29] require access to a pre-trained model and often perform poorly on higher-resolution datasets or those with numerous categories. Sleeper Agent [33] addresses these limitations by using gradient matching in a bi-level optimization process to generate poisoned samples.", + "bbox": [ + 511, + 614, + 906, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3. Positive Application", + "text_level": 1, + "bbox": [ + 513, + 775, + 697, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Positive applications of backdoor attacks are also being exploited, such as membership reasoning [16], algorithm evaluation [27, 42], and adversarial defense [32]. Backdoor watermark [26] has also emerged as an effective solution for protecting open-source datasets. However, existing methods still interfere with the original prediction results of watermarked models, thus compromising dataset functionality.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30693", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Data Protection", + "text_level": 1, + "bbox": [ + 89, + 90, + 248, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.1. Classic Data Protection", + "text_level": 1, + "bbox": [ + 89, + 112, + 299, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data protection aims to prevent unauthorized use and safeguard privacy. Common methods include (1) Digital Watermarking: Embedding owner-specific tags to assert ownership [11, 15]. (2) Encryption: Encrypting data, which can only be accessible with a secret key, securing sensitive information [4, 20, 22, 40, 41]. (3) Differential Privacy: Adding randomness (e.g., Laplace noise) to ensure indistinguishable outputs from neighboring datasets samples, preserving membership privacy [1, 44].", + "bbox": [ + 89, + 131, + 483, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By combining digital watermarking with poisoning-based backdoor attacks, radioactive data [31] trace protected data by increasing prediction confidence for certain categories. Specifically, it uses carriers to shift the feature representation of samples within a particular category toward a predetermined direction, thereby aiding classification. However, this technique requires a high poisoning rate (over $20\\%$ ) and lacks rigorous hypothesis testing in the black-box setting. Moreover, the isotropic carriers used as watermarks make radioactive data susceptible to detection through statistical defenses (e.g., subspace analysis).", + "bbox": [ + 89, + 268, + 483, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.2. Dataset Copyright Protection", + "text_level": 1, + "bbox": [ + 89, + 441, + 341, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset copyright protection is an emerging topic within the data protection field, aiming to verify whether a suspicious model has been trained on a protected dataset. Conventional data protection methods are not feasible for open-source datasets, as they either require access to training components or compromise dataset functionality. Backdoor watermark-based dataset ownership verification (BW-DOV) [26, 36] presents a feasible solution for protecting open-source datasets. DVBW [26] is the first to propose employing backdoor watermarks in original datasets and then detecting abnormal backdoors in suspicious models. To mitigate the harmfulness of backdoor watermarks, UBW [24] introduces an untargeted backdoor watermark under both dirty-label and clean-label settings. DomainWatermark [10] proposes a more harmless clean-label backdoor, where watermarked models correctly classify \"hard\" samples that benign models may misclassify. However, these methods still cause results inconsistency predicted by watermarked and benign models.", + "bbox": [ + 89, + 460, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. BW-DOV Formulation", + "text_level": 1, + "bbox": [ + 89, + 762, + 307, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Black-box Setting", + "text_level": 1, + "bbox": [ + 89, + 787, + 264, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main pipeline of existing BW-DOV methods consists of two phases: dataset watermarking and dataset ownership verification. During the dataset watermarking phase, they can only watermark a portion of the benign dataset and have no access to training components (e.g., model architecture, loss function). During the dataset ownership verification", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "phase, they can only obtain the prediction probabilities for the verification samples.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Main Pipeline of BW-DOV", + "text_level": 1, + "bbox": [ + 511, + 128, + 756, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Dataset Watermarking", + "text_level": 1, + "bbox": [ + 511, + 151, + 720, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^N$ represent the original dataset, where image $x_i \\in \\mathcal{X} = [0,1]^{C \\times H \\times W}$ , and label $y_i \\in \\mathcal{Y} = \\{1, \\ldots, K\\}$ . A small selected subset $\\mathcal{D}_s$ ( $\\mathcal{D}_s \\subset \\mathcal{D}$ ) is watermarked to produce a modified version $\\mathcal{D}_m = \\{G_\\psi(x_t), \\eta(y_t) | (x_t, y_t) \\in \\mathcal{D}_s\\}$ , where $G_\\psi: \\mathcal{X} \\to \\mathcal{X}$ indicates the watermark generator $G$ with parameters $\\psi$ and $\\eta$ indicates target label transformation. The modified samples $\\mathcal{D}_m$ are combined with the remaining benign samples $\\mathcal{D}_b$ ( $\\mathcal{D}_b = \\mathcal{D} \\setminus \\mathcal{D}_s$ ) to create the watermarked dataset $\\mathcal{D}_w$ ( $\\mathcal{D}_w = \\mathcal{D}_m + \\mathcal{D}_b$ ). The watermarking rate $\\gamma = |\\mathcal{D}_s| / |\\mathcal{D}|$ quantifies the proportion of modified samples. $\\mathcal{D}_w$ is finally released in place of the original dataset $\\mathcal{D}$ for protection.", + "bbox": [ + 511, + 169, + 906, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2. Dataset Ownership Verification", + "text_level": 1, + "bbox": [ + 511, + 357, + 777, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset protectors aim to verify whether a suspicious model $F$ has been trained on the protected dataset $\\mathcal{D}_w$ by identifying the specific implanted backdoors. These backdoors can be detected using hypothesis-test-based methods (e.g., pairwise T-test, Wilcoxon-test [14]) on verification sets $\\mathcal{D}_v$ .", + "bbox": [ + 511, + 376, + 905, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Dataset Watermarking", + "text_level": 1, + "bbox": [ + 511, + 464, + 730, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a more harmless BW-DOV method, EntropyMark implants a result-preserving backdoor under an entropy-based constraint. Specifically, it employs an iterative clean-label dataset watermarking framework.", + "bbox": [ + 511, + 488, + 905, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Result-preserving Backdoor", + "text_level": 1, + "bbox": [ + 511, + 556, + 766, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike conventional backdoors, EntropyMark improves prediction confidence while preserving original prediction results. This introduces a distributional bias in prediction probabilities, thereby disrupting the assumption of distribution consistency across samples from the same dataset [16].", + "bbox": [ + 511, + 579, + 905, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1.1. Entropy-based Constraint", + "text_level": 1, + "bbox": [ + 511, + 661, + 741, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Entropy is a measure of uncertainty in prediction probabilities, which can be formulated as follows:", + "bbox": [ + 511, + 680, + 905, + 709 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} (x) = - \\sum_ {i = 1} ^ {N} p _ {i} \\log p _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 715, + 906, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_i$ represents the probability of sample $x$ being classified as the $i^{th}$ label, and $N$ is the number of possible labels.", + "bbox": [ + 511, + 763, + 905, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The entropy-based constraint is introduced to ensure that watermarked samples are more confidently predicted by the backdoored models, thus introducing a distributional bias between the watermarked and benign samples. As a result, dataset protectors can verify whether the model was trained on the protected dataset by checking for entropy inconsistency between benign and watermarked samples.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30694", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/daea1abcdd2f163a47dd289b478e26cf6609b080c09cfa6d37e296554938aa55.jpg", + "image_caption": [ + "Figure 2. Main Pipeline of EntropyMark: In the dataset watermarking phase, we fine-tune a watermark generator with the aid of a surrogate model. This fine-tuned generator is then used to produce the watermarked dataset with backdoor capabilities. In the dataset ownership verification phase, we verify whether the suspicious model has been trained on our protected dataset by checking for entropy inconsistency in its predictions on benign and patched samples." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. Objective Formulation", + "text_level": 1, + "bbox": [ + 89, + 417, + 294, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, as the constraint may force watermarked samples to aggregate the feature representation toward one independent cluster, relying solely on the entropy-based constraint could lead to misclassification. To address this, we incorporate the cross-entropy loss to mitigate the negative impact on prediction accuracy. Specifically, we design the dispersion loss $\\mathcal{L}_d$ , which serves as the objective of our backdoor:", + "bbox": [ + 89, + 436, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} \\left(S _ {\\theta} (\\boldsymbol {x}, y)\\right) = \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right) + \\beta \\cdot \\mathcal {H} \\left(S _ {\\theta} (\\boldsymbol {x})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 569, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyper-parameter that balances the prediction accuracy and confidence, and $\\mathcal{H}$ denotes the entropy of the prediction probabilities.", + "bbox": [ + 89, + 597, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Iterative Dataset Watermarking Framework", + "text_level": 1, + "bbox": [ + 89, + 651, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the dataset watermarking phase, we modify a portion of samples to produce the watermarked datasets with backdoor capabilities. We adopt an iterative training paradigm to gradually fine-tune the watermark generator $G_{\\psi}$ , thereby enhancing the \"toxicity\" of the generated watermarked samples and then more effectively implanting the specific backdoor into models.", + "bbox": [ + 89, + 674, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the iterative process of our method. Initially, we follow the standard image steganography training process [35] to pre-train the watermark generator on the original dataset and randomly select samples to form a subset. We then utilize the pre-trained generator (with fixed parameters) to watermark the selected subset and combine the modified samples with the remaining benign samples to train the surrogate model. Subsequently, we fine-tune", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the watermark generator on the modified subset guided by the trained surrogate model (with fixed parameters). The surrogate model is then employed to adaptively select samples for watermarking in a new iteration round. This iterative process, involving surrogate model training, generator fine-tuning, and adaptive data selection, is repeated to fine-tune the watermark generator and finally produce the watermarked dataset with backdoor capabilities.", + "bbox": [ + 511, + 417, + 906, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Surrogate Model Training", + "text_level": 1, + "bbox": [ + 511, + 551, + 741, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We periodically train a surrogate model $S_{\\theta}$ to simulate the real-time performance of the suspicious model $F$ on the current watermarked dataset $\\mathcal{D}_w$ , aiding in the fine-tuning of the watermark generator $G_{\\psi}$ . Operating under a clean-label framework, we watermark the samples without altering their labels:", + "bbox": [ + 511, + 571, + 905, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(\\boldsymbol {x}, y) \\Longrightarrow (G _ {\\psi} (\\boldsymbol {x}), y), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 679, + 903, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the original sample $(x,y)\\in \\mathcal{D}_s$ , and $\\mathcal{D}_s$ is a selected subset from the original dataset.", + "bbox": [ + 511, + 710, + 905, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under the black-box setting, we cannot control the training components (e.g., model architecture, hyperparameters, optimizer) during model training. Therefore, we follow the vallina training process to train $S_{\\theta}$ on $\\mathcal{D}_w$ using the cross-entropy loss, formulated as:", + "bbox": [ + 511, + 743, + 905, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {w}} \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 834, + 903, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{ce}$ denotes the cross-entropy loss, $S_{\\theta}$ represents the surrogate model with parameters $\\theta$", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30695", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2. Generator Fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 90, + 295, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Directly fine-tuning the generator using the dispersion loss $\\mathcal{L}_d$ is challenging for training convergence, as it would require differentiating through the surrogate model's training procedure [7]. To address this, we employ a heuristic method known as gradient matching for training optimization. Let $\\mathcal{D}_v$ denote the patched verification set. We aim for our backdoor to be triggered on $\\mathcal{D}_v$ . Therefore, we define the verification loss $\\mathcal{L}_{ver}$ using the dispersion loss $\\mathcal{L}_d$ :", + "bbox": [ + 89, + 109, + 483, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v e r} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {v}} \\mathcal {L} _ {d} \\left(S _ {\\theta} (\\boldsymbol {x} + \\boldsymbol {p}, y)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 241, + 482, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}$ indicates the patch added to the original verification samples. Since the verification samples are private to dataset protectors, strong stealthiness is unnecessary.", + "bbox": [ + 89, + 267, + 482, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we simply employ the cross-entropy loss as the fine-tuning loss $\\mathcal{L}_{train}$ on the modified subset $\\mathcal{D}_s$ :", + "bbox": [ + 89, + 313, + 483, + 344 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a i n}} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {s}} \\mathcal {L} _ {\\text {c e}} \\left(S _ {\\theta} \\left(G _ {\\psi} (\\boldsymbol {x}), y\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 353, + 482, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we establish the following objective to align the verification loss $\\mathcal{L}_{ver}$ with the fine-tuning loss $\\mathcal{L}_{train}$ :", + "bbox": [ + 89, + 378, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\psi} 1 - \\frac {\\nabla_ {\\theta} \\mathcal {L} _ {t r a i n} \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {v e r}}{| | \\nabla_ {\\theta} \\mathcal {L} _ {t r a i n} | | \\cdot | | \\nabla_ {\\theta} \\mathcal {L} _ {v e r} | |}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 417, + 482, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this mode, we fine-tune the watermark generator $G_{\\psi}$ using the cross-entropy loss to generate the watermarked dataset with backdoor capabilities. The watermarked models will then exhibit higher prediction confidence on the patched verification samples. This mode replaces direct solvers on watermark generator $G_{\\psi}$ using the dispersion loss $\\mathcal{L}_d$ and triggers the backdoor solely on a specific subset, reducing the difficulty of backdoor implantation.", + "bbox": [ + 89, + 458, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3. Adaptive Data Selection", + "text_level": 1, + "bbox": [ + 89, + 585, + 303, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adaptively select more \"decisive\" samples to be watermarked for better backdoor implantation. Specifically, we choose the preferred subset $\\mathcal{D}_s$ with the smallest gradient norm of the cross-entropy loss:", + "bbox": [ + 89, + 604, + 483, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {s} = \\underset {(\\boldsymbol {x}, y) \\in \\mathcal {D}} {\\arg \\min } \\left| \\left| \\nabla_ {\\theta} \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right) \\right| \\right|. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 675, + 482, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive watermarked data selection greatly improves the stability of EntropyMark. By focusing on samples with smaller gradient norms, we can more effectively implant the desired backdoor into models.", + "bbox": [ + 89, + 710, + 483, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Dataset Ownership Verification", + "text_level": 1, + "bbox": [ + 89, + 784, + 377, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EntropyMark ensures that watermarked models $\\mathcal{M}_w$ show higher prediction confidence for patched verification samples compared to benign samples. To verify ownership, we use the entropy of the prediction probabilities as the verification metric and design a corresponding hypothesis-test-guided method:", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Suppose $F(\\cdot)$ is the probability vector predicted by the suspicious model $F$ . Let $\\pmb{x}$ denote a benign sample, and $\\pmb{x}' = \\pmb{x} + \\pmb{p}$ denote its patched version. Denote $E_{b} = \\mathcal{H}(F(\\pmb{x}))$ and $E_{w} = \\mathcal{H}(F(\\pmb{x}'))$ as the entropy of the prediction probabilities for $\\pmb{x}$ and $\\pmb{x}'$ , respectively. Given the null hypothesis $H_{0}: E_{b} + \\tau = E_{w}$ ( $H_{1}: E_{b} + \\tau > E_{w}$ ) where the hyper-parameter $\\tau \\in [-1, 1]$ , it is claimed that the suspicious model is trained on the protected dataset if and only if $H_{0}$ is rejected.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, in practice, we randomly select $m$ benign samples from the original dataset as the verification set. We then use these selected verification samples and their corresponding patched versions to conduct an entropy-based pairwise T-test [19] and compute the $p$ -value. A lower $p$ -value increases the likelihood of rejecting the null hypothesis $H_0$ , thereby indicating a higher probability that the suspicious model was trained on our protected dataset. In addition, we compute the confidence score $\\Delta E = E_b - E_w$ to represent the verification confidence.", + "bbox": [ + 511, + 227, + 908, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 513, + 391, + 645, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate EntropyMark using two popular benchmark datasets, including CIFAR-10 [18] and a 12-class subset of ImageNet [23]. Please refer to Appendix A for more details on the datasets.", + "bbox": [ + 511, + 416, + 906, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Evaluation on Backdoor Watermark", + "text_level": 1, + "bbox": [ + 511, + 486, + 830, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1.1. Settings", + "text_level": 1, + "bbox": [ + 511, + 508, + 614, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare EntropyMark against seven representative BW-DOV methods, including three dirty-label backdoor watermarks (i.e., BadNets [9], Blend [3], and UBW-P [24]) and four clean-label backdoor watermarks (i.e., Label-Consistent [38], Sleeper Agent [33], UBW-C [24], and Domain-Watermark [10]). We uniformly set the watermarking rate $\\gamma$ to $10\\%$ . More information on baseline settings is available in Appendix C.", + "bbox": [ + 511, + 527, + 906, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In EntropyMark, ResNet-18 [12] is used as the network architecture for both surrogate and watermarked models, each trained from scratch with randomly initialized weights. As demonstrated in Sec. 7.2, EntropyMark can ensure transferability under the black-box setting. We set the source label $y_{s}$ to 0, and the dispersion loss hyper-parameter $\\beta$ to 2.0. Further training configurations are provided in Appendix B.", + "bbox": [ + 511, + 648, + 908, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1.2. Metrics", + "text_level": 1, + "bbox": [ + 511, + 761, + 612, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of the backdoor watermarks using two key metrics: watermark success rate (WSR) and benign accuracy (BA). WSR represents the ratio of correctly predicted watermarked samples, while BA represents the ratio of correctly predicted benign samples.", + "bbox": [ + 511, + 779, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To quantitatively assess the harmfulness of different backdoor watermarks, we design the prediction consistency indices $\\mathcal{C}_b$ and $\\mathcal{C}_w$ . The indices measure the ratio of", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30696", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/98bab555976ba6fe151220c6866c36403c1d1b94519745e13fbfda8cb484af66.jpg", + "image_caption": [ + "Figure 3. Examples in Various BW-DOV Methods: In all BW-DOV methods, the target label is set to 'cat', while the source label is 'shelf' for Sleeper Agent, UBW-C, and ours. Our method stands out by not altering the ground-truth label of any sample while ensuring more stealthy modifications." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 906, + 334 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/29e1cc9dd0b08b7ab20d417ef2fb38c3bd21009e1601798a0b419401e8ba9e1d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodCIFAR-10Sub-ImageNet-12
WSR (%)BA (%)CbCwWSR (%)BA (%)CbCw
-Clean Model-92.43---79.84--
Dirty-labelBadNets [9]10091.670.93130.107610080.160.82500.0872
Blend [3]10091.620.92780.036110079.040.81960.0888
UBW-P [24]82.2090.700.91770.116763.4675.770.81440.2022
Clean-labelLabel-Consistent [38]94.4592.360.95180.067599.5173.300.74710.0065
Sleeper Agent [33]86.3291.540.91120.23018.2180.040.81250.7821
UBW-C [24]86.5087.870.87600.763948.8570.640.70770.7090
Domain-Watermark [10]99.4090.270.91630.143051.15*47.69*0.4750*0.4756*
EntropyMark (ours)87.0087.790.89360.869569.6271.350.75100.7471
", + "bbox": [ + 112, + 398, + 885, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Watermark performance on CIFAR-10 and Sub-ImageNet-12. Harmful $(\\mathcal{C}_w < 0.25$ or $c_{b} < 0.25)$ or abnormal results (WSR $<$ $10\\%$ or $BA_{drop} > 30\\%$ ) are marked in red. Results marked with $\\ast$ are reproduced on our own.", + "bbox": [ + 89, + 584, + 906, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "samples with consistent prediction results between watermarked models and clean models:", + "bbox": [ + 89, + 638, + 482, + 670 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\frac {1}{| \\mathcal {D} |} \\sum_ {x \\in \\mathcal {D}} \\mathbb {I} \\left\\{F _ {w} (\\boldsymbol {x}) = F _ {b} (\\boldsymbol {x}) \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 676, + 482, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F_{w}$ and $F_{b}$ denote the watermarked model and the benign model, respectively. $\\mathbb{I}$ represents the indicator function. This formula applies to $\\mathcal{C}_w$ and $\\mathcal{C}_b$ on the fullwatermarked $\\mathcal{D}_f$ and benign datasets $\\mathcal{D}$ , respectively.", + "bbox": [ + 89, + 723, + 483, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1.3. Results", + "text_level": 1, + "bbox": [ + 89, + 791, + 189, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 presents examples generated by different methods. Modified samples in BadNets, Blend, UBW-P, and Label Consistent all adopt static patches or patterns, making their modifications relatively easy to detect. Although Sleeper Agent and UBW-C apply more subtle perturbations, the modified samples still exhibit observable noise. In contrast,", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EntropyMark achieves a nearly imperceptible modification, demonstrating superior stealthiness.", + "bbox": [ + 511, + 638, + 903, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, the WSR and BA of EntropyMark are comparable to those of other clean-label backdoor watermarks. Due to the clean-label setting and specific nature of our backdoor, the WSR of our watermarked model is restricted by BA. However, this limitation does not hinder the effectiveness of the backdoor implantation, as we can still trace the unauthorized model through significant entropy inconsistency. Through extensive trials with different methods, we find that EntropyMark is more stable, with each metric value exhibiting minimal fluctuations.", + "bbox": [ + 509, + 671, + 906, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EntropyMark achieves prediction consistency indices $\\mathcal{C}_b$ and $\\mathcal{C}_w$ greater than 0.7, demonstrating optimal harmlessness. For most backdoor watermarks, watermarked models tend to exhibit flaws on malicious samples, resulting in extremely low $C_w$ . In contrast, EntropyMark preserves", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30697", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b22e663cbb05f2839172b87341861b516b7fd879ba1f43ab63c57ed7bbd9ef3d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CIFAR-10Sub-ImageNet-12
WatermarkBenignOtherWatermarkBenignOther
p-value0.00171.00001.00000.00181.00000.8375
ΔE0.00370.00820.0126-0.00310.0002-0.0003
", + "bbox": [ + 274, + 88, + 722, + 154 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/95eaa2d73fa9cec471f14a82afa3007e14af9de7554b2ce35545088093848e8a.jpg", + "table_caption": [ + "Table 2. The effectiveness of dataset ownership verification via our EntropyMark." + ], + "table_footnote": [], + "table_body": "
Iterative Fine-tuningGradient MatchingAdaptive SelectionWSR (%)BA (%)WatermarkBenignOther
XXX88.8288.690.07640.99980.0001
✓XX88.9689.030.81570.98370.5794
✓X✓86.8389.000.00390.98370.8655
✓✓X73.1887.280.00521.00001.0000
✓✓✓82.4787.790.00171.00001.0000
", + "bbox": [ + 133, + 193, + 867, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on CIFAR-10 with ResNet-18. Verification failures ( $p$ -value in 'Watermark' greater than those in 'Benign' or 'Other') are highlighted in red, and verification successes ( $p$ -value in 'Watermark' smaller than those in 'Benign' and 'Other') are highlighted in green. The same highlighting convention is applied in the following tables.", + "bbox": [ + 88, + 308, + 906, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset functionality, and corresponding watermarked models do not show obvious backdoor vulnerabilities, thus avoiding malicious prediction manipulation.", + "bbox": [ + 88, + 377, + 480, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Evaluation on Dataset Ownership Verification", + "text_level": 1, + "bbox": [ + 89, + 435, + 478, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 458, + 192, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate EntropyMark in three representative scenarios, including 1) Watermark: Detect watermarked models using our patched verification samples; 2) Benign: Detect benign models using the same patched verification samples; 3) Other: Detect watermarked models using patched samples with different patterns.", + "bbox": [ + 89, + 477, + 483, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.2. Metrics", + "text_level": 1, + "bbox": [ + 89, + 578, + 191, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our evaluation employs $p$ -value and $\\Delta E$ , where smaller values indicate higher accuracy in the 'Watermark' scenario, while higher values are anticipated in the 'Benign' and 'Other' scenarios. This approach effectively distinguishes models trained on our watermarked dataset from clean models and those trained on other protected datasets.", + "bbox": [ + 89, + 598, + 483, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.3. Results", + "text_level": 1, + "bbox": [ + 89, + 698, + 187, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 2, our verification provides consistently correct and deterministic results across all three scenarios. In the 'Watermark' scenario, we achieve a very low $p$ -value ( $p$ -value $\\ll 0.002$ ), indicating a strong likelihood of rejecting the null hypothesis. Conversely, the 'Benign' and 'Other' scenarios yield significantly higher $p$ -values ( $p$ -value $\\gg 0.5$ ). This differentiation aligns with our verification goals. In addition, $\\Delta E$ in the 'Watermark' scenario is smaller than in the 'Benign' and 'Other' scenarios, demonstrating high verification confidence.", + "bbox": [ + 89, + 718, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is noteworthy that watermarked datasets may slightly affect dataset functionality, leading to a subtle decline in", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prediction confidence on patched verification samples, especially in the CIFAR-10 dataset. Fortunately, this influence is effectively suppressed in the 'Watermark' scenario, enabling successful and reliable verification.", + "bbox": [ + 511, + 377, + 906, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 513, + 453, + 627, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 479, + 678, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our ablation studies, we evaluate the contribution of each technique in EntropyMark, including 1) iterative fine-tuning, 2) gradient matching, and 3) adaptive selection.", + "bbox": [ + 511, + 502, + 906, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 presents the results of various combinations of these techniques. Without fine-tuning, the pre-trained model cannot produce watermarked datasets with effective backdoor capabilities. Adding gradient matching refines the fine-tuning, significantly boosting verification accuracy. The adaptive selection mechanism further enhances verification, making it effective even without gradient matching. When all three techniques are combined, the watermarked model yields a high WSR and BA, creating the largest $p$ -value gap between the 'Watermark' scenario and the other two scenarios, indicating optimal verification performance.", + "bbox": [ + 511, + 547, + 908, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.2. Black-box Evaluation", + "text_level": 1, + "bbox": [ + 511, + 726, + 718, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the transferability of EntropyMark in the black-box setting. Specifically, we fix the training components of our surrogate model while varying the corresponding components of the suspicious models to be verified.", + "bbox": [ + 511, + 750, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 4, EntropyMark achieves a large $p$ -value gap between the 'Watermark' scenario and the other two scenarios, effectively distinguishing among the three scenarios. These results demonstrate that our method can verify suspicious models in the black-box setting with relatively stable WSR and BA, indicating strong transferability", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30698", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/945e174b6b57d513b13f356c6d1416ffecc2a32a8aa9b72eb45ce86a36d16927.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureBatch SizeOptimizerLearning RateWSR (%)BA (%)WatermarkBenignOther
ResNet-18128SGD0.0585.0688.200.00711.00000.5772
ResNet-1864SGD0.0584.6788.490.15390.86981.0000
ResNet-18128AdamW0.1074.8680.720.32661.00000.8113
ResNet-34128SGD0.1085.2888.440.36160.94671.0000
VGG-19128SGD0.1082.8487.090.46611.00000.7026
ShuffleNet128SGD0.1083.1788.230.39961.00000.5911
GoogleNet128SGD0.1088.0692.260.22530.96020.8639
", + "bbox": [ + 153, + 88, + 843, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6e8da1e98ff113c976e3493c9625ae67d7768e45fec77938a58cc9a8e59eb407.jpg", + "table_caption": [ + "Table 4. The results under the black-box setting for model architectures and hyper-parameters in the training process. For the surrogate model, we fix its training settings to use ResNet-18 as the model, 128 as the batch size, SGD as the optimizer, and 0.1 as the learning rate." + ], + "table_footnote": [], + "table_body": "
RWSR (%)BA (%)WatermarkBenignOther
184.8388.190.49250.99980.6289
282.4787.790.00171.00001.0000
380.4388.040.00790.98700.9977
469.1787.080.44391.00000.7764
572.4787.040.88850.61290.3223
", + "bbox": [ + 112, + 273, + 460, + 364 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6dc34110e7e1750d3259d76b3a0a899a911241681778816af72ef3abc27e18a3.jpg", + "table_caption": [ + "Table 5. The effects of iteration rounds $R$ ." + ], + "table_footnote": [], + "table_body": "
γWSR (%)BA (%)WatermarkBenignOther
0.0185.0788.860.05040.98700.9998
0.0581.1188.500.00181.00000.9999
0.1082.4787.790.00171.00001.0000
0.2079.7187.100.65831.00000.2552
", + "bbox": [ + 102, + 404, + 470, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. The effects of watermarking rate $\\gamma$ .", + "bbox": [ + 156, + 494, + 416, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and applicability to real-world scenarios.", + "bbox": [ + 94, + 534, + 359, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.3. Parameters Analysis", + "text_level": 1, + "bbox": [ + 94, + 556, + 282, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the impact of key hyper-parameters on the effectiveness of our method. Further analysis on additional hyper-parameters is provided in Appendix D.", + "bbox": [ + 94, + 579, + 480, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.3.1. Iteration Rounds", + "text_level": 1, + "bbox": [ + 94, + 631, + 254, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results in Tab. 5 indicate that: (1) As $R$ increases, WSR tends to decrease, while BA remains relatively stable; (2) With too few iterations, $p$ -values in the 'Watermark' and 'Other' scenarios are not significantly different, potentially causing confusion between different trigger patterns; (3) With too many iterations, WSR drops sharply, weakening backdoor capabilities. This effect may result from the entropy-based constraint becoming stronger with each iteration. As a consequence, the feature representations of watermarked samples start to cluster, which eventually leads to model misclassification. These results suggest that setting $R = 2$ enables the watermarked dataset to retain the strongest backdoor capabilities.", + "bbox": [ + 93, + 648, + 480, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.3.2. Watermarking Rate", + "text_level": 1, + "bbox": [ + 94, + 852, + 274, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 6, at lower watermarking rates (1%-10%), both WSR and BA remain stable. However, when $\\gamma$ is", + "bbox": [ + 94, + 871, + 480, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56226b4270987dfdf45270a5aa7db66768d449fdfc5d24b0498f145387d954da.jpg", + "image_caption": [ + "Figure 4. Resistance to defenses on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 516, + 273, + 901, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "too high $(20\\%)$ , the $p$ -value in the 'Other' scenario drops sharply, leading to verification failures. This likely happens because a higher watermarking rate increases the susceptibility of watermarked models to being triggered by other watermark patterns. These results indicate that a watermarking rate between $5\\%-10\\%$ achieves optimal verification performance.", + "bbox": [ + 516, + 433, + 903, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.4. Resistance to Defenses", + "text_level": 1, + "bbox": [ + 516, + 549, + 718, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset stealers may employ defenses to bypass backdoor watermarks during model training. We evaluate two practical backdoor removal methods: fine-tuning and model pruning [28].", + "bbox": [ + 516, + 571, + 903, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 4, model pruning fails to remove our backdoor without sacrificing BA while fine-tuning only has minor effects on both WSR and BA. Since EntropyMark is sample-specific, it naturally resists most conventional sample-filtering defenses. For further details on additional defenses, please refer to Appendix E.", + "bbox": [ + 516, + 631, + 903, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 516, + 736, + 630, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a more harmless backdoor watermark that influences prediction confidence without altering prediction results. In addition, we design an iterative clean-label dataset watermarking framework under an entropy-based constraint. For better backdoor implantation, we utilize gradient matching and adaptive data selection to optimize training convergence. Extensive experiments on benchmark datasets validate the effectiveness, transferability, and resistance to defenses of our method.", + "bbox": [ + 516, + 761, + 903, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30699", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.62272020.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 148, + 187, + 165 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiawang Bai, Yiming Li, Jiawei Li, Xue Yang, Yong Jiang, and Shu-Tao Xia. Multinomial random forest. Pattern Recognition, 122:108331, 2022. 1, 3", + "[2] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 2", + "[3] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 5, 6", + "[4] Hua Deng, Zheng Qin, Qianhong Wu, Zhenyu Guan, Robert H Deng, Yujue Wang, and Yunya Zhou. Identity-based encryption transformation for flexible sharing of encrypted data in public cloud. IEEE Transactions on Information Forensics and Security, 15:3168-3180, 2020. 1, 3", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1", + "[6] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference, pages 113-125, 2019. 3", + "[7] Jonas Geiping, Liam Fowl, W Ronny Huang, Wojciech Czaja, Gavin Taylor, Michael Moeller, and Tom Goldstein. Witches' brew: Industrial scale data poisoning via gradient matching. arXiv preprint arXiv:2009.02276, 2020. 5", + "[8] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020. 1", + "[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 2, 5, 6", + "[10] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: effective and harmless dataset copyright protection is closed at hand. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 54421-54450, 2023. 2, 3, 5, 6", + "[11] Yuanfang Guo, Oscar C Au, Rui Wang, Lu Fang, and Xiaochun Cao. Halftone image watermarking by content aware double-sided embedding error diffusion. IEEE Transactions on Image Processing, 27(7):3387-3402, 2018. 1, 3", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + ], + "bbox": [ + 93, + 172, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4", + "[14] Robert V Hogg, Joseph W McKean, Allen T Craig, et al. Introduction to mathematical statistics. Pearson Education India, 2013. 3", + "[15] Chiou-Ting Hsu and Ja-Ling Wu. Hidden digital watermarks in images. IEEE Transactions on image processing, 8(1):58-68, 1999. 1, 3", + "[16] Hongsheng Hu, Zoran Salcic, Gillian Dobbie, Jinjun Chen, Lichao Sun, and Xuyun Zhang. Membership inference via backdoorsing. arXiv preprint arXiv:2206.04823, 2022. 2, 3", + "[17] Wenbo Jiang, Hongwei Li, Guowen Xu, and Tianwei Zhang. Color backdoor: A robust poisoning attack in color space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8133-8142, 2023. 2", + "[18] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 1, 5", + "[19] Richard J Larsen and Morris L Marx. An introduction to mathematical statistics. Prentice Hall Hoboken, NJ, 2005. 5", + "[20] Jiguo Li, Qihong Yu, and Yichen Zhang. Hierarchical attribute based encryption with continuous leakage-resilience. Information Sciences, 484:113-134, 2019. 3", + "[21] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16463-16472, 2021. 2", + "[22] Yiming Li, Peidong Liu, Yong Jiang, and Shu-Tao Xia. Visual privacy protection via mapping distortion. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3740-3744. IEEE, 2021. 3", + "[23] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 5, 1", + "[24] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 5, 6", + "[25] Yiming Li, Mengxi Ya, Yang Bai, Yong Jiang, and Shu-Tao Xia. Backdoorbox: A python toolbox for backdoor learning. arXiv preprint arXiv:2302.01762, 2023. 2", + "[26] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 2, 3", + "[27] Yi-Shan Lin, Wen-Chuan Lee, and Z Berkay Celik. What do you see? evaluation of explainable artificial intelligence (xai) interpretability through neural backdoors. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 1027-1035, 2021. 2", + "[28] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30700", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 8", + "[29] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 182-199. Springer, 2020. 2", + "[30] Anh Nguyen and Anh Tran. Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369, 2021. 2", + "[31] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 3", + "[32] Shawn Shan, Emily Wenger, Bolun Wang, Bo Li, Haitao Zheng, and Ben Y Zhao. Gotta catch'em all: Using honeypots to catch adversarial attacks on neural networks. In Proceedings of the 2020 ACM SIGSAC conference on computer and communications security, pages 67-83, 2020. 2", + "[33] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. Advances in Neural Information Processing Systems, 35: 19165-19178, 2022. 2, 5, 6", + "[34] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 4", + "[35] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4, 1", + "[36] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 3", + "[37] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 3", + "[38] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 5, 6", + "[39] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1", + "[40] Shulan Wang, Junwei Zhou, Joseph K Liu, Jianping Yu, Jianyong Chen, and Weixin Xie. An efficient file hierarchy attribute-based encryption scheme in cloud computing. IEEE Transactions on Information Forensics and Security, 11(6): 1265-1277, 2016. 1, 3", + "[41] Zuobin Xiong, Zhipeng Cai, Qilong Han, Arwa Alrawais, and Wei Li. Adgan: Protect your location privacy in camera data of auto-driving vehicles. IEEE Transactions on Industrial Informatics, 17(9):6200-6210, 2020. 3", + "[42] Mengxi Ya, Yiming Li, Tao Dai, Bin Wang, Yong Jiang, and Shu-Tao Xia. Towards faithful xai evaluation via generalization-limited backdoor watermark. In The" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Twelfth International Conference on Learning Representations, 2023. 2", + "[43] Runkai Zheng, Rongjun Tang, Jianze Li, and Li Liu. Data-free backdoor removal based on channel lipschitzness. arXiv preprint arXiv:2208.03111, 2022. 4", + "[44] Linghui Zhu, Xinyi Liu, Yiming Li, Xue Yang, Shu-Tao Xia, and Rongxing Lu. A fine-grained differentially private federated learning against leakage from gradients. IEEE Internet of Things Journal, 9(13):11500-11512, 2021. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 219 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_model.json b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_model.json new file mode 100644 index 0000000000000000000000000000000000000000..41fe04abc6615fb3fe09cd45880bdb121e42c0c1 --- /dev/null +++ b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_model.json @@ -0,0 +1,2475 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.868, + 0.177 + ], + "angle": 0, + "content": "EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.78, + 0.222 + ], + "angle": 0, + "content": "Ming Sun\\(^{1,2}\\), Rui Wang\\(^{1,2*}\\), Zixuan Zhu\\(^{1,2}\\), Lihua Jing\\(^{1,2}\\), Yuanfang Guo\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.222, + 0.773, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Information Engineering, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.239, + 0.783, + 0.256 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Cyber Security, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.257, + 0.767, + 0.275 + ], + "angle": 0, + "content": "3 School of Computer Science and Engineering, Beihang University" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.278, + 0.808, + 0.292 + ], + "angle": 0, + "content": "nming, wangrui, zhuzixuan, jinglihua}@iie.ac.cn, andyguo@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.359, + 0.485, + 0.782 + ], + "angle": 0, + "content": "High-quality open-source datasets are essential for advancing deep neural networks. However, the unauthorized commercial use of these datasets has raised significant concerns about copyright protection. One promising approach is backdoor watermark-based dataset ownership verification (BW-DOV), in which dataset protectors implant specific backdoors into illicit models through dataset watermarking, enabling the tracing of these models through abnormal prediction behaviors. Unfortunately, the targeted nature of these BW-DOV methods can be maliciously exploited, potentially leading to harmful side effects. While existing harmless methods attempt to mitigate these risks, watermarked datasets can still negatively affect prediction results, partially compromising dataset functionality. In this paper, we propose a more harmless backdoor watermark, called EntropyMark, which improves prediction confidence without altering the final prediction results. For this purpose, an entropy-based constraint is introduced to regulate the probability distribution. Specifically, we design an iterative clean-label dataset watermarking framework. Our framework employs gradient matching and adaptive data selection to optimize backdoor injection. In parallel, we introduce a hypothesis test method grounded in entropy inconsistency to verify dataset ownership. Extensive experiments on benchmark datasets demonstrate the effectiveness, transferability, and defense resistance of our approach. Our code is available at https://github.com/AaronSun2000/EntropyMark." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.809, + 0.222, + 0.824 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.483, + 0.88 + ], + "angle": 0, + "content": "In recent years, deep neural networks (DNNs) have rapidly advanced and been widely deployed in numerous real-world applications, such as autonomous driving [8] and face" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.326, + 0.905, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.537, + 0.625, + 0.548 + ], + "angle": 0, + "content": "I. Training" + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.537, + 0.825, + 0.547 + ], + "angle": 0, + "content": "II. Inference" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.559, + 0.908, + 0.63 + ], + "angle": 0, + "content": "Figure 1. Comparison of EntropyMark with existing back-door watermarks: Existing watermarks alter the prediction results of backdoored models and produce conspicuous modifications on datasets. In contrast, EntropyMark preserves the original results and is significantly more stealthy." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.825 + ], + "angle": 0, + "content": "recognition [39]. The prosperity of DNNs largely depends on the availability of high-quality datasets (e.g., CIFAR [18] and ImageNet [5]). Researchers use open-source datasets to evaluate and iteratively optimize their proposed methods. However, the value of these datasets has attracted dataset stealers who illegally copy, distribute and use them without permission. Currently, most open-source datasets are intended solely for educational or research purposes and explicitly prohibit unauthorized commercial use1. Therefore, copyright protection for open-source datasets has become an urgent and necessary topic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.905, + 0.872 + ], + "angle": 0, + "content": "Unfortunately, classic data protection methods, including differential privacy [1, 44], encryption [4, 40], and digital watermarking [11, 15], are not well-suited for open-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.716, + 0.9 + ], + "angle": 0, + "content": "1See IMDb data policy for details." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30692" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "source datasets, as they are primarily designed for private data protection. Safeguarding open-source datasets presents more challenges because their protectors lack control over the training process and have no access to training components. Instead, they can only query the outputs of suspicious models to determine whether those models were trained on the protected datasets. In addition, due to the open-source nature of these datasets, any protection strategy must avoid compromising dataset functionality or altering the integrity of information representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.482, + 0.381 + ], + "angle": 0, + "content": "Backdoor watermark-based dataset ownership verification (BW-DOV) [26] offers a feasible solution for protecting open-source dataset copyrights. This approach involves modifying samples in the original dataset and releasing the watermarked version for open-source use. When a model is trained on this watermarked dataset, a specific backdoor is implanted. To verify whether a suspicious model was trained on the protected dataset, protectors can simply detect the presence of the predetermined backdoor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.627 + ], + "angle": 0, + "content": "However, due to the targeted nature (i.e., specified erroneous prediction results) of existing BW-DOV methods, watermarked models are vulnerable to manipulation and may maliciously output predetermined results. To address this issue, UBW [24] proposes an untargeted backdoor watermark, which mitigates the targeted nature of the backdoor by increasing prediction randomness on watermarked samples. Domain-Watermark [10] ensures that watermarked models correctly classify certain 'hard' samples that benign models may misclassify. However, these methods still alter the original prediction results, inevitably impacting dataset functionality. To address the limitations of existing BW-DOV methods, we argue that harmless backdoor watermarks should not only avoid generating specified erroneous prediction results but also preserve the original prediction outputs, thereby maintaining dataset functionality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose EntropyMark, a more harmless backdoor watermark designed to improve prediction confidence for verification samples without affecting overall prediction results. To ensure that the prediction probabilities for these verification samples are more concentrated on the ground-truth label, we introduce an entropy-based constraint to regulate the probability distribution. For implementation, we design an iterative clean-label dataset watermarking framework. Within this framework, we employ gradient matching to optimize training convergence, ensuring effective backdoor injection into models trained on the watermarked dataset. We also utilize gradient-based adaptive selection to identify more \"decisive\" samples for watermarking. Fig. 1 compares EntropyMark with existing BW-DOV methods, emphasizing the harmlessness and stealthiness of our method. Unlike radioactive data [31], the classic data tracing method through watermarking, EntropyMark exhibits greater robustness in black-box settings and offers" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "enhanced resilience against various defenses. Finally, to further improve verification accuracy, we incorporate a pairwise T-test based on entropy inconsistency." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.138, + 0.895, + 0.152 + ], + "angle": 0, + "content": "The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.153, + 0.905, + 0.242 + ], + "angle": 0, + "content": "(1) We propose EntropyMark, a backdoor watermark under an entropy-based constraint. Based on our backdoor watermark, we design an iterative clean-label dataset watermarking framework. EntropyMark achieves enhanced stealthiness and minimal harmfulness while retaining watermark effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.244, + 0.905, + 0.289 + ], + "angle": 0, + "content": "(2) Our framework adaptively selects \"decisive\" samples to be watermarked and incorporates gradient matching to enhance training convergence." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.29, + 0.905, + 0.334 + ], + "angle": 0, + "content": "(3) Building on the properties of EntropyMark, we develop an entropy inconsistency hypothesis test to quantitatively verify dataset ownership." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.335, + 0.905, + 0.379 + ], + "angle": 0, + "content": "(4) Extensive experiments on benchmark datasets validate the effectiveness, transferability, and defense resistance of our method." + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.153, + 0.905, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.396, + 0.655, + 0.412 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.422, + 0.689, + 0.436 + ], + "angle": 0, + "content": "2.1. Backdoor Attacks" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.445, + 0.694, + 0.459 + ], + "angle": 0, + "content": "2.1.1. Dirty-label Attacks" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.464, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Dirty-label attacks forcibly modify the labels of selected malicious samples. BadNets [9], the first dirty-label attack, uses a static patch as the trigger. To enhance stealthiness, Blend [3] generates malicious samples by blending the trigger with original images. Subsequently, more stealthy and effective methods [17, 21, 30] have been proposed, varying the forms of triggers. However, dirty-label attacks are still detectable due to label-semantic inconsistencies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.595, + 0.697, + 0.609 + ], + "angle": 0, + "content": "2.1.2. Clean-label Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.766 + ], + "angle": 0, + "content": "Clean-label attacks, which retain original sample labels, are well-suited for scenarios where attackers lack control over label settings. Label-Consistent [38] introduces the first patch-based attack under this paradigm. However, many early clean-label attacks [2, 29] require access to a pre-trained model and often perform poorly on higher-resolution datasets or those with numerous categories. Sleeper Agent [33] addresses these limitations by using gradient matching in a bi-level optimization process to generate poisoned samples." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.776, + 0.699, + 0.79 + ], + "angle": 0, + "content": "2.1.3. Positive Application" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Positive applications of backdoor attacks are also being exploited, such as membership reasoning [16], algorithm evaluation [27, 42], and adversarial defense [32]. Backdoor watermark [26] has also emerged as an effective solution for protecting open-source datasets. However, existing methods still interfere with the original prediction results of watermarked models, thus compromising dataset functionality." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30693" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.249, + 0.105 + ], + "angle": 0, + "content": "2.2. Data Protection" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.113, + 0.3, + 0.127 + ], + "angle": 0, + "content": "2.2.1. Classic Data Protection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.132, + 0.484, + 0.268 + ], + "angle": 0, + "content": "Data protection aims to prevent unauthorized use and safeguard privacy. Common methods include (1) Digital Watermarking: Embedding owner-specific tags to assert ownership [11, 15]. (2) Encryption: Encrypting data, which can only be accessible with a secret key, securing sensitive information [4, 20, 22, 40, 41]. (3) Differential Privacy: Adding randomness (e.g., Laplace noise) to ensure indistinguishable outputs from neighboring datasets samples, preserving membership privacy [1, 44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.484, + 0.435 + ], + "angle": 0, + "content": "By combining digital watermarking with poisoning-based backdoor attacks, radioactive data [31] trace protected data by increasing prediction confidence for certain categories. Specifically, it uses carriers to shift the feature representation of samples within a particular category toward a predetermined direction, thereby aiding classification. However, this technique requires a high poisoning rate (over \\(20\\%\\) ) and lacks rigorous hypothesis testing in the black-box setting. Moreover, the isotropic carriers used as watermarks make radioactive data susceptible to detection through statistical defenses (e.g., subspace analysis)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.443, + 0.342, + 0.457 + ], + "angle": 0, + "content": "2.2.2. Dataset Copyright Protection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Dataset copyright protection is an emerging topic within the data protection field, aiming to verify whether a suspicious model has been trained on a protected dataset. Conventional data protection methods are not feasible for open-source datasets, as they either require access to training components or compromise dataset functionality. Backdoor watermark-based dataset ownership verification (BW-DOV) [26, 36] presents a feasible solution for protecting open-source datasets. DVBW [26] is the first to propose employing backdoor watermarks in original datasets and then detecting abnormal backdoors in suspicious models. To mitigate the harmfulness of backdoor watermarks, UBW [24] introduces an untargeted backdoor watermark under both dirty-label and clean-label settings. DomainWatermark [10] proposes a more harmless clean-label backdoor, where watermarked models correctly classify \"hard\" samples that benign models may misclassify. However, these methods still cause results inconsistency predicted by watermarked and benign models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.308, + 0.779 + ], + "angle": 0, + "content": "3. BW-DOV Formulation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.265, + 0.804 + ], + "angle": 0, + "content": "3.1. Black-box Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The main pipeline of existing BW-DOV methods consists of two phases: dataset watermarking and dataset ownership verification. During the dataset watermarking phase, they can only watermark a portion of the benign dataset and have no access to training components (e.g., model architecture, loss function). During the dataset ownership verification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "phase, they can only obtain the prediction probabilities for the verification samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.129, + 0.757, + 0.145 + ], + "angle": 0, + "content": "3.2. Main Pipeline of BW-DOV" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.152, + 0.722, + 0.167 + ], + "angle": 0, + "content": "3.2.1. Dataset Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Let \\(\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^N\\) represent the original dataset, where image \\(x_i \\in \\mathcal{X} = [0,1]^{C \\times H \\times W}\\), and label \\(y_i \\in \\mathcal{Y} = \\{1, \\ldots, K\\}\\). A small selected subset \\(\\mathcal{D}_s\\) (\\(\\mathcal{D}_s \\subset \\mathcal{D}\\)) is watermarked to produce a modified version \\(\\mathcal{D}_m = \\{G_\\psi(x_t), \\eta(y_t) | (x_t, y_t) \\in \\mathcal{D}_s\\}\\), where \\(G_\\psi: \\mathcal{X} \\to \\mathcal{X}\\) indicates the watermark generator \\(G\\) with parameters \\(\\psi\\) and \\(\\eta\\) indicates target label transformation. The modified samples \\(\\mathcal{D}_m\\) are combined with the remaining benign samples \\(\\mathcal{D}_b\\) (\\(\\mathcal{D}_b = \\mathcal{D} \\setminus \\mathcal{D}_s\\)) to create the watermarked dataset \\(\\mathcal{D}_w\\) (\\(\\mathcal{D}_w = \\mathcal{D}_m + \\mathcal{D}_b\\)). The watermarking rate \\(\\gamma = |\\mathcal{D}_s| / |\\mathcal{D}|\\) quantifies the proportion of modified samples. \\(\\mathcal{D}_w\\) is finally released in place of the original dataset \\(\\mathcal{D}\\) for protection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.358, + 0.779, + 0.373 + ], + "angle": 0, + "content": "3.2.2. Dataset Ownership Verification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.906, + 0.453 + ], + "angle": 0, + "content": "Dataset protectors aim to verify whether a suspicious model \\(F\\) has been trained on the protected dataset \\(\\mathcal{D}_w\\) by identifying the specific implanted backdoors. These backdoors can be detected using hypothesis-test-based methods (e.g., pairwise T-test, Wilcoxon-test [14]) on verification sets \\(\\mathcal{D}_v\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.731, + 0.481 + ], + "angle": 0, + "content": "4. Dataset Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.551 + ], + "angle": 0, + "content": "As a more harmless BW-DOV method, EntropyMark implants a result-preserving backdoor under an entropy-based constraint. Specifically, it employs an iterative clean-label dataset watermarking framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.767, + 0.574 + ], + "angle": 0, + "content": "4.1. Result-preserving Backdoor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Unlike conventional backdoors, EntropyMark improves prediction confidence while preserving original prediction results. This introduces a distributional bias in prediction probabilities, thereby disrupting the assumption of distribution consistency across samples from the same dataset [16]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.662, + 0.743, + 0.676 + ], + "angle": 0, + "content": "4.1.1. Entropy-based Constraint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.71 + ], + "angle": 0, + "content": "Entropy is a measure of uncertainty in prediction probabilities, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.717, + 0.907, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} (x) = - \\sum_ {i = 1} ^ {N} p _ {i} \\log p _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.795 + ], + "angle": 0, + "content": "where \\( p_i \\) represents the probability of sample \\( x \\) being classified as the \\( i^{th} \\) label, and \\( N \\) is the number of possible labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The entropy-based constraint is introduced to ensure that watermarked samples are more confidently predicted by the backdoored models, thus introducing a distributional bias between the watermarked and benign samples. As a result, dataset protectors can verify whether the model was trained on the protected dataset by checking for entropy inconsistency between benign and watermarked samples." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.336, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Main Pipeline of EntropyMark: In the dataset watermarking phase, we fine-tune a watermark generator with the aid of a surrogate model. This fine-tuned generator is then used to produce the watermarked dataset with backdoor capabilities. In the dataset ownership verification phase, we verify whether the suspicious model has been trained on our protected dataset by checking for entropy inconsistency in its predictions on benign and patched samples." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.419, + 0.295, + 0.433 + ], + "angle": 0, + "content": "4.1.2. Objective Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.543 + ], + "angle": 0, + "content": "However, as the constraint may force watermarked samples to aggregate the feature representation toward one independent cluster, relying solely on the entropy-based constraint could lead to misclassification. To address this, we incorporate the cross-entropy loss to mitigate the negative impact on prediction accuracy. Specifically, we design the dispersion loss \\(\\mathcal{L}_d\\), which serves as the objective of our backdoor:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.57, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} \\left(S _ {\\theta} (\\boldsymbol {x}, y)\\right) = \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right) + \\beta \\cdot \\mathcal {H} \\left(S _ {\\theta} (\\boldsymbol {x})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.645 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyper-parameter that balances the prediction accuracy and confidence, and \\(\\mathcal{H}\\) denotes the entropy of the prediction probabilities." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.652, + 0.469, + 0.669 + ], + "angle": 0, + "content": "4.2. Iterative Dataset Watermarking Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.779 + ], + "angle": 0, + "content": "During the dataset watermarking phase, we modify a portion of samples to produce the watermarked datasets with backdoor capabilities. We adopt an iterative training paradigm to gradually fine-tune the watermark generator \\( G_{\\psi} \\), thereby enhancing the \"toxicity\" of the generated watermarked samples and then more effectively implanting the specific backdoor into models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fig. 2 illustrates the iterative process of our method. Initially, we follow the standard image steganography training process [35] to pre-train the watermark generator on the original dataset and randomly select samples to form a subset. We then utilize the pre-trained generator (with fixed parameters) to watermark the selected subset and combine the modified samples with the remaining benign samples to train the surrogate model. Subsequently, we fine-tune" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.54 + ], + "angle": 0, + "content": "the watermark generator on the modified subset guided by the trained surrogate model (with fixed parameters). The surrogate model is then employed to adaptively select samples for watermarking in a new iteration round. This iterative process, involving surrogate model training, generator fine-tuning, and adaptive data selection, is repeated to fine-tune the watermark generator and finally produce the watermarked dataset with backdoor capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.552, + 0.743, + 0.568 + ], + "angle": 0, + "content": "4.2.1. Surrogate Model Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.906, + 0.663 + ], + "angle": 0, + "content": "We periodically train a surrogate model \\( S_{\\theta} \\) to simulate the real-time performance of the suspicious model \\( F \\) on the current watermarked dataset \\( \\mathcal{D}_w \\), aiding in the fine-tuning of the watermark generator \\( G_{\\psi} \\). Operating under a clean-label framework, we watermark the samples without altering their labels:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.68, + 0.905, + 0.697 + ], + "angle": 0, + "content": "\\[\n(\\boldsymbol {x}, y) \\Longrightarrow (G _ {\\psi} (\\boldsymbol {x}), y), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.906, + 0.742 + ], + "angle": 0, + "content": "where the original sample \\((x,y)\\in \\mathcal{D}_s\\) , and \\(\\mathcal{D}_s\\) is a selected subset from the original dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.906, + 0.819 + ], + "angle": 0, + "content": "Under the black-box setting, we cannot control the training components (e.g., model architecture, hyperparameters, optimizer) during model training. Therefore, we follow the vallina training process to train \\( S_{\\theta} \\) on \\( \\mathcal{D}_w \\) using the cross-entropy loss, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.835, + 0.905, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {w}} \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{ce}\\) denotes the cross-entropy loss, \\(S_{\\theta}\\) represents the surrogate model with parameters \\(\\theta\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30695" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.296, + 0.108 + ], + "angle": 0, + "content": "4.2.2. Generator Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.484, + 0.232 + ], + "angle": 0, + "content": "Directly fine-tuning the generator using the dispersion loss \\(\\mathcal{L}_d\\) is challenging for training convergence, as it would require differentiating through the surrogate model's training procedure [7]. To address this, we employ a heuristic method known as gradient matching for training optimization. Let \\(\\mathcal{D}_v\\) denote the patched verification set. We aim for our backdoor to be triggered on \\(\\mathcal{D}_v\\). Therefore, we define the verification loss \\(\\mathcal{L}_{ver}\\) using the dispersion loss \\(\\mathcal{L}_d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.242, + 0.483, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v e r} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {v}} \\mathcal {L} _ {d} \\left(S _ {\\theta} (\\boldsymbol {x} + \\boldsymbol {p}, y)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.314 + ], + "angle": 0, + "content": "where \\( \\pmb{p} \\) indicates the patch added to the original verification samples. Since the verification samples are private to dataset protectors, strong stealthiness is unnecessary." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.314, + 0.484, + 0.345 + ], + "angle": 0, + "content": "Then we simply employ the cross-entropy loss as the fine-tuning loss \\(\\mathcal{L}_{train}\\) on the modified subset \\(\\mathcal{D}_s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.354, + 0.483, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a i n}} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {s}} \\mathcal {L} _ {\\text {c e}} \\left(S _ {\\theta} \\left(G _ {\\psi} (\\boldsymbol {x}), y\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.38, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Finally, we establish the following objective to align the verification loss \\(\\mathcal{L}_{ver}\\) with the fine-tuning loss \\(\\mathcal{L}_{train}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.418, + 0.483, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\psi} 1 - \\frac {\\nabla_ {\\theta} \\mathcal {L} _ {t r a i n} \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {v e r}}{| | \\nabla_ {\\theta} \\mathcal {L} _ {t r a i n} | | \\cdot | | \\nabla_ {\\theta} \\mathcal {L} _ {v e r} | |}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.581 + ], + "angle": 0, + "content": "In this mode, we fine-tune the watermark generator \\( G_{\\psi} \\) using the cross-entropy loss to generate the watermarked dataset with backdoor capabilities. The watermarked models will then exhibit higher prediction confidence on the patched verification samples. This mode replaces direct solvers on watermark generator \\( G_{\\psi} \\) using the dispersion loss \\( \\mathcal{L}_d \\) and triggers the backdoor solely on a specific subset, reducing the difficulty of backdoor implantation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.305, + 0.602 + ], + "angle": 0, + "content": "4.2.3. Adaptive Data Selection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.667 + ], + "angle": 0, + "content": "We adaptively select more \"decisive\" samples to be watermarked for better backdoor implantation. Specifically, we choose the preferred subset \\( \\mathcal{D}_s \\) with the smallest gradient norm of the cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.676, + 0.483, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {s} = \\underset {(\\boldsymbol {x}, y) \\in \\mathcal {D}} {\\arg \\min } \\left| \\left| \\nabla_ {\\theta} \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right) \\right| \\right|. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.772 + ], + "angle": 0, + "content": "Adaptive watermarked data selection greatly improves the stability of EntropyMark. By focusing on samples with smaller gradient norms, we can more effectively implant the desired backdoor into models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.379, + 0.803 + ], + "angle": 0, + "content": "5. Dataset Ownership Verification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "EntropyMark ensures that watermarked models \\(\\mathcal{M}_w\\) show higher prediction confidence for patched verification samples compared to benign samples. To verify ownership, we use the entropy of the prediction probabilities as the verification metric and design a corresponding hypothesis-test-guided method:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Suppose \\( F(\\cdot) \\) is the probability vector predicted by the suspicious model \\( F \\). Let \\( \\pmb{x} \\) denote a benign sample, and \\( \\pmb{x}' = \\pmb{x} + \\pmb{p} \\) denote its patched version. Denote \\( E_{b} = \\mathcal{H}(F(\\pmb{x})) \\) and \\( E_{w} = \\mathcal{H}(F(\\pmb{x}')) \\) as the entropy of the prediction probabilities for \\( \\pmb{x} \\) and \\( \\pmb{x}' \\), respectively. Given the null hypothesis \\( H_{0}: E_{b} + \\tau = E_{w} \\) (\\( H_{1}: E_{b} + \\tau > E_{w} \\)) where the hyper-parameter \\( \\tau \\in [-1, 1] \\), it is claimed that the suspicious model is trained on the protected dataset if and only if \\( H_{0} \\) is rejected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.909, + 0.379 + ], + "angle": 0, + "content": "As shown in Fig. 2, in practice, we randomly select \\( m \\) benign samples from the original dataset as the verification set. We then use these selected verification samples and their corresponding patched versions to conduct an entropy-based pairwise T-test [19] and compute the \\( p \\)-value. A lower \\( p \\)-value increases the likelihood of rejecting the null hypothesis \\( H_0 \\), thereby indicating a higher probability that the suspicious model was trained on our protected dataset. In addition, we compute the confidence score \\( \\Delta E = E_b - E_w \\) to represent the verification confidence." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.392, + 0.646, + 0.409 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.478 + ], + "angle": 0, + "content": "We evaluate EntropyMark using two popular benchmark datasets, including CIFAR-10 [18] and a 12-class subset of ImageNet [23]. Please refer to Appendix A for more details on the datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.487, + 0.831, + 0.502 + ], + "angle": 0, + "content": "6.1. Evaluation on Backdoor Watermark" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.509, + 0.616, + 0.524 + ], + "angle": 0, + "content": "6.1.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.649 + ], + "angle": 0, + "content": "We compare EntropyMark against seven representative BW-DOV methods, including three dirty-label backdoor watermarks (i.e., BadNets [9], Blend [3], and UBW-P [24]) and four clean-label backdoor watermarks (i.e., Label-Consistent [38], Sleeper Agent [33], UBW-C [24], and Domain-Watermark [10]). We uniformly set the watermarking rate \\(\\gamma\\) to \\(10\\%\\). More information on baseline settings is available in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.909, + 0.756 + ], + "angle": 0, + "content": "In EntropyMark, ResNet-18 [12] is used as the network architecture for both surrogate and watermarked models, each trained from scratch with randomly initialized weights. As demonstrated in Sec. 7.2, EntropyMark can ensure transferability under the black-box setting. We set the source label \\( y_{s} \\) to 0, and the dispersion loss hyper-parameter \\( \\beta \\) to 2.0. Further training configurations are provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.614, + 0.775 + ], + "angle": 0, + "content": "6.1.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.856 + ], + "angle": 0, + "content": "We evaluate the effectiveness of the backdoor watermarks using two key metrics: watermark success rate (WSR) and benign accuracy (BA). WSR represents the ratio of correctly predicted watermarked samples, while BA represents the ratio of correctly predicted benign samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To quantitatively assess the harmfulness of different backdoor watermarks, we design the prediction consistency indices \\( \\mathcal{C}_b \\) and \\( \\mathcal{C}_w \\). The indices measure the ratio of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "30696" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.907, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Figure 3. Examples in Various BW-DOV Methods: In all BW-DOV methods, the target label is set to 'cat', while the source label is 'shelf' for Sleeper Agent, UBW-C, and ours. Our method stands out by not altering the ground-truth label of any sample while ensuring more stealthy modifications." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.4, + 0.887, + 0.575 + ], + "angle": 0, + "content": "
TypeMethodCIFAR-10Sub-ImageNet-12
WSR (%)BA (%)CbCwWSR (%)BA (%)CbCw
-Clean Model-92.43---79.84--
Dirty-labelBadNets [9]10091.670.93130.107610080.160.82500.0872
Blend [3]10091.620.92780.036110079.040.81960.0888
UBW-P [24]82.2090.700.91770.116763.4675.770.81440.2022
Clean-labelLabel-Consistent [38]94.4592.360.95180.067599.5173.300.74710.0065
Sleeper Agent [33]86.3291.540.91120.23018.2180.040.81250.7821
UBW-C [24]86.5087.870.87600.763948.8570.640.70770.7090
Domain-Watermark [10]99.4090.270.91630.143051.15*47.69*0.4750*0.4756*
EntropyMark (ours)87.0087.790.89360.869569.6271.350.75100.7471
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.585, + 0.908, + 0.614 + ], + "angle": 0, + "content": "Table 1. Watermark performance on CIFAR-10 and Sub-ImageNet-12. Harmful \\((\\mathcal{C}_w < 0.25\\) or \\(c_{b} < 0.25)\\) or abnormal results (WSR \\(< \\) \\(10\\%\\) or \\(BA_{drop} > 30\\%\\) ) are marked in red. Results marked with \\(\\ast\\) are reproduced on our own." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.671 + ], + "angle": 0, + "content": "samples with consistent prediction results between watermarked models and clean models:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.678, + 0.483, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\frac {1}{| \\mathcal {D} |} \\sum_ {x \\in \\mathcal {D}} \\mathbb {I} \\left\\{F _ {w} (\\boldsymbol {x}) = F _ {b} (\\boldsymbol {x}) \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.484, + 0.786 + ], + "angle": 0, + "content": "where \\(F_{w}\\) and \\(F_{b}\\) denote the watermarked model and the benign model, respectively. \\(\\mathbb{I}\\) represents the indicator function. This formula applies to \\(\\mathcal{C}_w\\) and \\(\\mathcal{C}_b\\) on the fullwatermarked \\(\\mathcal{D}_f\\) and benign datasets \\(\\mathcal{D}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.19, + 0.805 + ], + "angle": 0, + "content": "6.1.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 presents examples generated by different methods. Modified samples in BadNets, Blend, UBW-P, and Label Consistent all adopt static patches or patterns, making their modifications relatively easy to detect. Although Sleeper Agent and UBW-C apply more subtle perturbations, the modified samples still exhibit observable noise. In contrast," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.905, + 0.671 + ], + "angle": 0, + "content": "EntropyMark achieves a nearly imperceptible modification, demonstrating superior stealthiness." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.672, + 0.907, + 0.824 + ], + "angle": 0, + "content": "As shown in Tab. 1, the WSR and BA of EntropyMark are comparable to those of other clean-label backdoor watermarks. Due to the clean-label setting and specific nature of our backdoor, the WSR of our watermarked model is restricted by BA. However, this limitation does not hinder the effectiveness of the backdoor implantation, as we can still trace the unauthorized model through significant entropy inconsistency. Through extensive trials with different methods, we find that EntropyMark is more stable, with each metric value exhibiting minimal fluctuations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "EntropyMark achieves prediction consistency indices \\( \\mathcal{C}_b \\) and \\( \\mathcal{C}_w \\) greater than 0.7, demonstrating optimal harmlessness. For most backdoor watermarks, watermarked models tend to exhibit flaws on malicious samples, resulting in extremely low \\( C_w \\). In contrast, EntropyMark preserves" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30697" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.275, + 0.089, + 0.723, + 0.155 + ], + "angle": 0, + "content": "
CIFAR-10Sub-ImageNet-12
WatermarkBenignOtherWatermarkBenignOther
p-value0.00171.00001.00000.00181.00000.8375
ΔE0.00370.00820.0126-0.00310.0002-0.0003
" + }, + { + "type": "table_caption", + "bbox": [ + 0.256, + 0.166, + 0.74, + 0.18 + ], + "angle": 0, + "content": "Table 2. The effectiveness of dataset ownership verification via our EntropyMark." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.194, + 0.868, + 0.3 + ], + "angle": 0, + "content": "
Iterative Fine-tuningGradient MatchingAdaptive SelectionWSR (%)BA (%)WatermarkBenignOther
XXX88.8288.690.07640.99980.0001
✓XX88.9689.030.81570.98370.5794
✓X✓86.8389.000.00390.98370.8655
✓✓X73.1887.280.00521.00001.0000
✓✓✓82.4787.790.00171.00001.0000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.309, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Table 3. Ablation studies on CIFAR-10 with ResNet-18. Verification failures (\\(p\\)-value in 'Watermark' greater than those in 'Benign' or 'Other') are highlighted in red, and verification successes (\\(p\\)-value in 'Watermark' smaller than those in 'Benign' and 'Other') are highlighted in green. The same highlighting convention is applied in the following tables." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.378, + 0.482, + 0.425 + ], + "angle": 0, + "content": "dataset functionality, and corresponding watermarked models do not show obvious backdoor vulnerabilities, thus avoiding malicious prediction manipulation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.436, + 0.48, + 0.452 + ], + "angle": 0, + "content": "6.2. Evaluation on Dataset Ownership Verification" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.459, + 0.194, + 0.473 + ], + "angle": 0, + "content": "6.2.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.569 + ], + "angle": 0, + "content": "We validate EntropyMark in three representative scenarios, including 1) Watermark: Detect watermarked models using our patched verification samples; 2) Benign: Detect benign models using the same patched verification samples; 3) Other: Detect watermarked models using patched samples with different patterns." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.579, + 0.192, + 0.593 + ], + "angle": 0, + "content": "6.2.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.69 + ], + "angle": 0, + "content": "Our evaluation employs \\( p \\)-value and \\( \\Delta E \\), where smaller values indicate higher accuracy in the 'Watermark' scenario, while higher values are anticipated in the 'Benign' and 'Other' scenarios. This approach effectively distinguishes models trained on our watermarked dataset from clean models and those trained on other protected datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.699, + 0.189, + 0.713 + ], + "angle": 0, + "content": "6.2.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.869 + ], + "angle": 0, + "content": "As shown in Tab. 2, our verification provides consistently correct and deterministic results across all three scenarios. In the 'Watermark' scenario, we achieve a very low \\( p \\)-value (\\( p \\)-value \\( \\ll 0.002 \\)), indicating a strong likelihood of rejecting the null hypothesis. Conversely, the 'Benign' and 'Other' scenarios yield significantly higher \\( p \\)-values (\\( p \\)-value \\( \\gg 0.5 \\)). This differentiation aligns with our verification goals. In addition, \\( \\Delta E \\) in the 'Watermark' scenario is smaller than in the 'Benign' and 'Other' scenarios, demonstrating high verification confidence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "It is noteworthy that watermarked datasets may slightly affect dataset functionality, leading to a subtle decline in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.44 + ], + "angle": 0, + "content": "prediction confidence on patched verification samples, especially in the CIFAR-10 dataset. Fortunately, this influence is effectively suppressed in the 'Watermark' scenario, enabling successful and reliable verification." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.454, + 0.628, + 0.47 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.679, + 0.495 + ], + "angle": 0, + "content": "7.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.548 + ], + "angle": 0, + "content": "In our ablation studies, we evaluate the contribution of each technique in EntropyMark, including 1) iterative fine-tuning, 2) gradient matching, and 3) adaptive selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.909, + 0.716 + ], + "angle": 0, + "content": "Tab. 3 presents the results of various combinations of these techniques. Without fine-tuning, the pre-trained model cannot produce watermarked datasets with effective backdoor capabilities. Adding gradient matching refines the fine-tuning, significantly boosting verification accuracy. The adaptive selection mechanism further enhances verification, making it effective even without gradient matching. When all three techniques are combined, the watermarked model yields a high WSR and BA, creating the largest \\( p \\)-value gap between the 'Watermark' scenario and the other two scenarios, indicating optimal verification performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.719, + 0.741 + ], + "angle": 0, + "content": "7.2. Black-box Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We evaluate the transferability of EntropyMark in the black-box setting. Specifically, we fix the training components of our surrogate model while varying the corresponding components of the suspicious models to be verified." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "As shown in Tab. 4, EntropyMark achieves a large \\( p \\)-value gap between the 'Watermark' scenario and the other two scenarios, effectively distinguishing among the three scenarios. These results demonstrate that our method can verify suspicious models in the black-box setting with relatively stable WSR and BA, indicating strong transferability" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30698" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.154, + 0.089, + 0.844, + 0.212 + ], + "angle": 0, + "content": "
ArchitectureBatch SizeOptimizerLearning RateWSR (%)BA (%)WatermarkBenignOther
ResNet-18128SGD0.0585.0688.200.00711.00000.5772
ResNet-1864SGD0.0584.6788.490.15390.86981.0000
ResNet-18128AdamW0.1074.8680.720.32661.00000.8113
ResNet-34128SGD0.1085.2888.440.36160.94671.0000
VGG-19128SGD0.1082.8487.090.46611.00000.7026
ShuffleNet128SGD0.1083.1788.230.39961.00000.5911
GoogleNet128SGD0.1088.0692.260.22530.96020.8639
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.224, + 0.905, + 0.251 + ], + "angle": 0, + "content": "Table 4. The results under the black-box setting for model architectures and hyper-parameters in the training process. For the surrogate model, we fix its training settings to use ResNet-18 as the model, 128 as the batch size, SGD as the optimizer, and 0.1 as the learning rate." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.275, + 0.462, + 0.366 + ], + "angle": 0, + "content": "
RWSR (%)BA (%)WatermarkBenignOther
184.8388.190.49250.99980.6289
282.4787.790.00171.00001.0000
380.4388.040.00790.98700.9977
469.1787.080.44391.00000.7764
572.4787.040.88850.61290.3223
" + }, + { + "type": "table_caption", + "bbox": [ + 0.163, + 0.378, + 0.411, + 0.391 + ], + "angle": 0, + "content": "Table 5. The effects of iteration rounds \\( R \\) ." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.405, + 0.472, + 0.484 + ], + "angle": 0, + "content": "
γWSR (%)BA (%)WatermarkBenignOther
0.0185.0788.860.05040.98700.9998
0.0581.1188.500.00181.00000.9999
0.1082.4787.790.00171.00001.0000
0.2079.7187.100.65831.00000.2552
" + }, + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.496, + 0.417, + 0.508 + ], + "angle": 0, + "content": "Table 6. The effects of watermarking rate \\( \\gamma \\) ." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.535, + 0.36, + 0.548 + ], + "angle": 0, + "content": "and applicability to real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.558, + 0.284, + 0.572 + ], + "angle": 0, + "content": "7.3. Parameters Analysis" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.58, + 0.481, + 0.624 + ], + "angle": 0, + "content": "We analyze the impact of key hyper-parameters on the effectiveness of our method. Further analysis on additional hyper-parameters is provided in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.632, + 0.255, + 0.644 + ], + "angle": 0, + "content": "7.3.1. Iteration Rounds" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.65, + 0.482, + 0.845 + ], + "angle": 0, + "content": "Results in Tab. 5 indicate that: (1) As \\( R \\) increases, WSR tends to decrease, while BA remains relatively stable; (2) With too few iterations, \\( p \\)-values in the 'Watermark' and 'Other' scenarios are not significantly different, potentially causing confusion between different trigger patterns; (3) With too many iterations, WSR drops sharply, weakening backdoor capabilities. This effect may result from the entropy-based constraint becoming stronger with each iteration. As a consequence, the feature representations of watermarked samples start to cluster, which eventually leads to model misclassification. These results suggest that setting \\( R = 2 \\) enables the watermarked dataset to retain the strongest backdoor capabilities." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.853, + 0.276, + 0.867 + ], + "angle": 0, + "content": "7.3.2. Watermarking Rate" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.872, + 0.481, + 0.9 + ], + "angle": 0, + "content": "As shown in Tab. 6, at lower watermarking rates (1%-10%), both WSR and BA remain stable. However, when \\(\\gamma\\) is" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.275, + 0.903, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.394, + 0.849, + 0.406 + ], + "angle": 0, + "content": "Figure 4. Resistance to defenses on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.434, + 0.904, + 0.539 + ], + "angle": 0, + "content": "too high \\((20\\%)\\), the \\(p\\)-value in the 'Other' scenario drops sharply, leading to verification failures. This likely happens because a higher watermarking rate increases the susceptibility of watermarked models to being triggered by other watermark patterns. These results indicate that a watermarking rate between \\(5\\%-10\\%\\) achieves optimal verification performance." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.55, + 0.72, + 0.563 + ], + "angle": 0, + "content": "7.4. Resistance to Defenses" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.572, + 0.904, + 0.631 + ], + "angle": 0, + "content": "Dataset stealers may employ defenses to bypass backdoor watermarks during model training. We evaluate two practical backdoor removal methods: fine-tuning and model pruning [28]." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.632, + 0.904, + 0.721 + ], + "angle": 0, + "content": "As shown in Fig. 4, model pruning fails to remove our backdoor without sacrificing BA while fine-tuning only has minor effects on both WSR and BA. Since EntropyMark is sample-specific, it naturally resists most conventional sample-filtering defenses. For further details on additional defenses, please refer to Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.737, + 0.631, + 0.751 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.762, + 0.904, + 0.896 + ], + "angle": 0, + "content": "In this work, we propose a more harmless backdoor watermark that influences prediction confidence without altering prediction results. In addition, we design an iterative clean-label dataset watermarking framework under an entropy-based constraint. For better backdoor implantation, we utilize gradient matching and adaptive data selection to optimize training convergence. Extensive experiments on benchmark datasets validate the effectiveness, transferability, and resistance to defenses of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30699" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.62272020." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.149, + 0.188, + 0.166 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.174, + 0.482, + 0.216 + ], + "angle": 0, + "content": "[1] Jiawang Bai, Yiming Li, Jiawei Li, Xue Yang, Yong Jiang, and Shu-Tao Xia. Multinomial random forest. Pattern Recognition, 122:108331, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.217, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[2] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.273, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[4] Hua Deng, Zheng Qin, Qianhong Wu, Zhenyu Guan, Robert H Deng, Yujue Wang, and Yunya Zhou. Identity-based encryption transformation for flexible sharing of encrypted data in public cloud. IEEE Transactions on Information Forensics and Security, 15:3168-3180, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[6] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference, pages 113-125, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[7] Jonas Geiping, Liam Fowl, W Ronny Huang, Wojciech Czaja, Gavin Taylor, Michael Moeller, and Tom Goldstein. Witches' brew: Industrial scale data poisoning via gradient matching. arXiv preprint arXiv:2009.02276, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[8] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.679, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[10] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: effective and harmless dataset copyright protection is closed at hand. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 54421-54450, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[11] Yuanfang Guo, Oscar C Au, Rui Wang, Lu Fang, and Xiaochun Cao. Halftone image watermarking by content aware double-sided embedding error diffusion. IEEE Transactions on Image Processing, 27(7):3387-3402, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.174, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Robert V Hogg, Joseph W McKean, Allen T Craig, et al. Introduction to mathematical statistics. Pearson Education India, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] Chiou-Ting Hsu and Ja-Ling Wu. Hidden digital watermarks in images. IEEE Transactions on image processing, 8(1):58-68, 1999. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[16] Hongsheng Hu, Zoran Salcic, Gillian Dobbie, Jinjun Chen, Lichao Sun, and Xuyun Zhang. Membership inference via backdoorsing. arXiv preprint arXiv:2206.04823, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[17] Wenbo Jiang, Hongwei Li, Guowen Xu, and Tianwei Zhang. Color backdoor: A robust poisoning attack in color space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8133-8142, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[18] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.377 + ], + "angle": 0, + "content": "[19] Richard J Larsen and Morris L Marx. An introduction to mathematical statistics. Prentice Hall Hoboken, NJ, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[20] Jiguo Li, Qihong Yu, and Yichen Zhang. Hierarchical attribute based encryption with continuous leakage-resilience. Information Sciences, 484:113-134, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[21] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16463-16472, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[22] Yiming Li, Peidong Liu, Yong Jiang, and Shu-Tao Xia. Visual privacy protection via mapping distortion. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3740-3744. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[23] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[24] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[25] Yiming Li, Mengxi Ya, Yang Bai, Yong Jiang, and Shu-Tao Xia. Backdoorbox: A python toolbox for backdoor learning. arXiv preprint arXiv:2302.01762, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[26] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[27] Yi-Shan Lin, Wen-Chuan Lee, and Z Berkay Celik. What do you see? evaluation of explainable artificial intelligence (xai) interpretability through neural backdoors. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 1027-1035, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30700" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[29] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 182-199. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[30] Anh Nguyen and Anh Tran. Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[31] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[32] Shawn Shan, Emily Wenger, Bolun Wang, Bo Li, Haitao Zheng, and Ben Y Zhao. Gotta catch'em all: Using honeypots to catch adversarial attacks on neural networks. In Proceedings of the 2020 ACM SIGSAC conference on computer and communications security, pages 67-83, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[33] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. Advances in Neural Information Processing Systems, 35: 19165-19178, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[34] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[35] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[36] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[37] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[38] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[39] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[40] Shulan Wang, Junwei Zhou, Joseph K Liu, Jianping Yu, Jianyong Chen, and Weixin Xie. An efficient file hierarchy attribute-based encryption scheme in cloud computing. IEEE Transactions on Information Forensics and Security, 11(6): 1265-1277, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[41] Zuobin Xiong, Zhipeng Cai, Qilong Han, Arwa Alrawais, and Wei Li. Adgan: Protect your location privacy in camera data of auto-driving vehicles. IEEE Transactions on Industrial Informatics, 17(9):6200-6210, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Mengxi Ya, Yiming Li, Tao Dai, Bin Wang, Yong Jiang, and Shu-Tao Xia. Towards faithful xai evaluation via generalization-limited backdoor watermark. In The" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[43] Runkai Zheng, Rongjun Tang, Jianze Li, and Li Liu. Data-free backdoor removal based on channel lipschitzness. arXiv preprint arXiv:2208.03111, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[44] Linghui Zhu, Xinyi Liu, Yiming Li, Xue Yang, Shu-Tao Xia, and Rongxing Lu. A fine-grained differentially private federated learning against leakage from gradients. IEEE Internet of Things Journal, 9(13):11500-11512, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30701" + } + ] +] \ No newline at end of file diff --git a/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_origin.pdf b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd3c0f2d3a5a05ae3b96ae5b6ae22e60ce433af6 --- /dev/null +++ b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/79eea005-2d9c-4999-9dbc-01e224e70254_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed0b3d24b3e57da4a40ee4c75bdcd6292d050beaedbdf3f6a8759582b13528c1 +size 1236958 diff --git a/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/full.md b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e19e8d78a4a802ebfb201031c3b9702619eebc75 --- /dev/null +++ b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/full.md @@ -0,0 +1,373 @@ +# EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection + +Ming Sun $^{1,2}$ , Rui Wang $^{1,2*}$ , Zixuan Zhu $^{1,2}$ , Lihua Jing $^{1,2}$ , Yuanfang Guo $^{3}$ + +$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences + +$^{2}$ School of Cyber Security, University of Chinese Academy of Sciences + +3 School of Computer Science and Engineering, Beihang University + +nming, wangrui, zhuzixuan, jinglihua}@iie.ac.cn, andyguo@buaa.edu.cn + +# Abstract + +High-quality open-source datasets are essential for advancing deep neural networks. However, the unauthorized commercial use of these datasets has raised significant concerns about copyright protection. One promising approach is backdoor watermark-based dataset ownership verification (BW-DOV), in which dataset protectors implant specific backdoors into illicit models through dataset watermarking, enabling the tracing of these models through abnormal prediction behaviors. Unfortunately, the targeted nature of these BW-DOV methods can be maliciously exploited, potentially leading to harmful side effects. While existing harmless methods attempt to mitigate these risks, watermarked datasets can still negatively affect prediction results, partially compromising dataset functionality. In this paper, we propose a more harmless backdoor watermark, called EntropyMark, which improves prediction confidence without altering the final prediction results. For this purpose, an entropy-based constraint is introduced to regulate the probability distribution. Specifically, we design an iterative clean-label dataset watermarking framework. Our framework employs gradient matching and adaptive data selection to optimize backdoor injection. In parallel, we introduce a hypothesis test method grounded in entropy inconsistency to verify dataset ownership. Extensive experiments on benchmark datasets demonstrate the effectiveness, transferability, and defense resistance of our approach. Our code is available at https://github.com/AaronSun2000/EntropyMark. + +# 1. Introduction + +In recent years, deep neural networks (DNNs) have rapidly advanced and been widely deployed in numerous real-world applications, such as autonomous driving [8] and face + +![](images/0c94024f018bb9f1eced0dac4a858983c4e562b75b100696ecf1becf9600b5a0.jpg) +I. Training +II. Inference +Figure 1. Comparison of EntropyMark with existing back-door watermarks: Existing watermarks alter the prediction results of backdoored models and produce conspicuous modifications on datasets. In contrast, EntropyMark preserves the original results and is significantly more stealthy. + +recognition [39]. The prosperity of DNNs largely depends on the availability of high-quality datasets (e.g., CIFAR [18] and ImageNet [5]). Researchers use open-source datasets to evaluate and iteratively optimize their proposed methods. However, the value of these datasets has attracted dataset stealers who illegally copy, distribute and use them without permission. Currently, most open-source datasets are intended solely for educational or research purposes and explicitly prohibit unauthorized commercial use1. Therefore, copyright protection for open-source datasets has become an urgent and necessary topic. + +Unfortunately, classic data protection methods, including differential privacy [1, 44], encryption [4, 40], and digital watermarking [11, 15], are not well-suited for open- + +source datasets, as they are primarily designed for private data protection. Safeguarding open-source datasets presents more challenges because their protectors lack control over the training process and have no access to training components. Instead, they can only query the outputs of suspicious models to determine whether those models were trained on the protected datasets. In addition, due to the open-source nature of these datasets, any protection strategy must avoid compromising dataset functionality or altering the integrity of information representation. + +Backdoor watermark-based dataset ownership verification (BW-DOV) [26] offers a feasible solution for protecting open-source dataset copyrights. This approach involves modifying samples in the original dataset and releasing the watermarked version for open-source use. When a model is trained on this watermarked dataset, a specific backdoor is implanted. To verify whether a suspicious model was trained on the protected dataset, protectors can simply detect the presence of the predetermined backdoor. + +However, due to the targeted nature (i.e., specified erroneous prediction results) of existing BW-DOV methods, watermarked models are vulnerable to manipulation and may maliciously output predetermined results. To address this issue, UBW [24] proposes an untargeted backdoor watermark, which mitigates the targeted nature of the backdoor by increasing prediction randomness on watermarked samples. Domain-Watermark [10] ensures that watermarked models correctly classify certain 'hard' samples that benign models may misclassify. However, these methods still alter the original prediction results, inevitably impacting dataset functionality. To address the limitations of existing BW-DOV methods, we argue that harmless backdoor watermarks should not only avoid generating specified erroneous prediction results but also preserve the original prediction outputs, thereby maintaining dataset functionality. + +In this paper, we propose EntropyMark, a more harmless backdoor watermark designed to improve prediction confidence for verification samples without affecting overall prediction results. To ensure that the prediction probabilities for these verification samples are more concentrated on the ground-truth label, we introduce an entropy-based constraint to regulate the probability distribution. For implementation, we design an iterative clean-label dataset watermarking framework. Within this framework, we employ gradient matching to optimize training convergence, ensuring effective backdoor injection into models trained on the watermarked dataset. We also utilize gradient-based adaptive selection to identify more "decisive" samples for watermarking. Fig. 1 compares EntropyMark with existing BW-DOV methods, emphasizing the harmlessness and stealthiness of our method. Unlike radioactive data [31], the classic data tracing method through watermarking, EntropyMark exhibits greater robustness in black-box settings and offers + +enhanced resilience against various defenses. Finally, to further improve verification accuracy, we incorporate a pairwise T-test based on entropy inconsistency. + +The main contributions can be summarized as follows: + +(1) We propose EntropyMark, a backdoor watermark under an entropy-based constraint. Based on our backdoor watermark, we design an iterative clean-label dataset watermarking framework. EntropyMark achieves enhanced stealthiness and minimal harmfulness while retaining watermark effectiveness. +(2) Our framework adaptively selects "decisive" samples to be watermarked and incorporates gradient matching to enhance training convergence. +(3) Building on the properties of EntropyMark, we develop an entropy inconsistency hypothesis test to quantitatively verify dataset ownership. +(4) Extensive experiments on benchmark datasets validate the effectiveness, transferability, and defense resistance of our method. + +# 2. Related Work + +# 2.1. Backdoor Attacks + +# 2.1.1. Dirty-label Attacks + +Dirty-label attacks forcibly modify the labels of selected malicious samples. BadNets [9], the first dirty-label attack, uses a static patch as the trigger. To enhance stealthiness, Blend [3] generates malicious samples by blending the trigger with original images. Subsequently, more stealthy and effective methods [17, 21, 30] have been proposed, varying the forms of triggers. However, dirty-label attacks are still detectable due to label-semantic inconsistencies. + +# 2.1.2. Clean-label Attacks + +Clean-label attacks, which retain original sample labels, are well-suited for scenarios where attackers lack control over label settings. Label-Consistent [38] introduces the first patch-based attack under this paradigm. However, many early clean-label attacks [2, 29] require access to a pre-trained model and often perform poorly on higher-resolution datasets or those with numerous categories. Sleeper Agent [33] addresses these limitations by using gradient matching in a bi-level optimization process to generate poisoned samples. + +# 2.1.3. Positive Application + +Positive applications of backdoor attacks are also being exploited, such as membership reasoning [16], algorithm evaluation [27, 42], and adversarial defense [32]. Backdoor watermark [26] has also emerged as an effective solution for protecting open-source datasets. However, existing methods still interfere with the original prediction results of watermarked models, thus compromising dataset functionality. + +# 2.2. Data Protection + +# 2.2.1. Classic Data Protection + +Data protection aims to prevent unauthorized use and safeguard privacy. Common methods include (1) Digital Watermarking: Embedding owner-specific tags to assert ownership [11, 15]. (2) Encryption: Encrypting data, which can only be accessible with a secret key, securing sensitive information [4, 20, 22, 40, 41]. (3) Differential Privacy: Adding randomness (e.g., Laplace noise) to ensure indistinguishable outputs from neighboring datasets samples, preserving membership privacy [1, 44]. + +By combining digital watermarking with poisoning-based backdoor attacks, radioactive data [31] trace protected data by increasing prediction confidence for certain categories. Specifically, it uses carriers to shift the feature representation of samples within a particular category toward a predetermined direction, thereby aiding classification. However, this technique requires a high poisoning rate (over $20\%$ ) and lacks rigorous hypothesis testing in the black-box setting. Moreover, the isotropic carriers used as watermarks make radioactive data susceptible to detection through statistical defenses (e.g., subspace analysis). + +# 2.2.2. Dataset Copyright Protection + +Dataset copyright protection is an emerging topic within the data protection field, aiming to verify whether a suspicious model has been trained on a protected dataset. Conventional data protection methods are not feasible for open-source datasets, as they either require access to training components or compromise dataset functionality. Backdoor watermark-based dataset ownership verification (BW-DOV) [26, 36] presents a feasible solution for protecting open-source datasets. DVBW [26] is the first to propose employing backdoor watermarks in original datasets and then detecting abnormal backdoors in suspicious models. To mitigate the harmfulness of backdoor watermarks, UBW [24] introduces an untargeted backdoor watermark under both dirty-label and clean-label settings. DomainWatermark [10] proposes a more harmless clean-label backdoor, where watermarked models correctly classify "hard" samples that benign models may misclassify. However, these methods still cause results inconsistency predicted by watermarked and benign models. + +# 3. BW-DOV Formulation + +# 3.1. Black-box Setting + +The main pipeline of existing BW-DOV methods consists of two phases: dataset watermarking and dataset ownership verification. During the dataset watermarking phase, they can only watermark a portion of the benign dataset and have no access to training components (e.g., model architecture, loss function). During the dataset ownership verification + +phase, they can only obtain the prediction probabilities for the verification samples. + +# 3.2. Main Pipeline of BW-DOV + +# 3.2.1. Dataset Watermarking + +Let $\mathcal{D} = \{(x_i, y_i)\}_{i=1}^N$ represent the original dataset, where image $x_i \in \mathcal{X} = [0,1]^{C \times H \times W}$ , and label $y_i \in \mathcal{Y} = \{1, \ldots, K\}$ . A small selected subset $\mathcal{D}_s$ ( $\mathcal{D}_s \subset \mathcal{D}$ ) is watermarked to produce a modified version $\mathcal{D}_m = \{G_\psi(x_t), \eta(y_t) | (x_t, y_t) \in \mathcal{D}_s\}$ , where $G_\psi: \mathcal{X} \to \mathcal{X}$ indicates the watermark generator $G$ with parameters $\psi$ and $\eta$ indicates target label transformation. The modified samples $\mathcal{D}_m$ are combined with the remaining benign samples $\mathcal{D}_b$ ( $\mathcal{D}_b = \mathcal{D} \setminus \mathcal{D}_s$ ) to create the watermarked dataset $\mathcal{D}_w$ ( $\mathcal{D}_w = \mathcal{D}_m + \mathcal{D}_b$ ). The watermarking rate $\gamma = |\mathcal{D}_s| / |\mathcal{D}|$ quantifies the proportion of modified samples. $\mathcal{D}_w$ is finally released in place of the original dataset $\mathcal{D}$ for protection. + +# 3.2.2. Dataset Ownership Verification + +Dataset protectors aim to verify whether a suspicious model $F$ has been trained on the protected dataset $\mathcal{D}_w$ by identifying the specific implanted backdoors. These backdoors can be detected using hypothesis-test-based methods (e.g., pairwise T-test, Wilcoxon-test [14]) on verification sets $\mathcal{D}_v$ . + +# 4. Dataset Watermarking + +As a more harmless BW-DOV method, EntropyMark implants a result-preserving backdoor under an entropy-based constraint. Specifically, it employs an iterative clean-label dataset watermarking framework. + +# 4.1. Result-preserving Backdoor + +Unlike conventional backdoors, EntropyMark improves prediction confidence while preserving original prediction results. This introduces a distributional bias in prediction probabilities, thereby disrupting the assumption of distribution consistency across samples from the same dataset [16]. + +# 4.1.1. Entropy-based Constraint + +Entropy is a measure of uncertainty in prediction probabilities, which can be formulated as follows: + +$$ +\mathcal {H} (x) = - \sum_ {i = 1} ^ {N} p _ {i} \log p _ {i}, \tag {1} +$$ + +where $p_i$ represents the probability of sample $x$ being classified as the $i^{th}$ label, and $N$ is the number of possible labels. + +The entropy-based constraint is introduced to ensure that watermarked samples are more confidently predicted by the backdoored models, thus introducing a distributional bias between the watermarked and benign samples. As a result, dataset protectors can verify whether the model was trained on the protected dataset by checking for entropy inconsistency between benign and watermarked samples. + +![](images/daea1abcdd2f163a47dd289b478e26cf6609b080c09cfa6d37e296554938aa55.jpg) +Figure 2. Main Pipeline of EntropyMark: In the dataset watermarking phase, we fine-tune a watermark generator with the aid of a surrogate model. This fine-tuned generator is then used to produce the watermarked dataset with backdoor capabilities. In the dataset ownership verification phase, we verify whether the suspicious model has been trained on our protected dataset by checking for entropy inconsistency in its predictions on benign and patched samples. + +# 4.1.2. Objective Formulation + +However, as the constraint may force watermarked samples to aggregate the feature representation toward one independent cluster, relying solely on the entropy-based constraint could lead to misclassification. To address this, we incorporate the cross-entropy loss to mitigate the negative impact on prediction accuracy. Specifically, we design the dispersion loss $\mathcal{L}_d$ , which serves as the objective of our backdoor: + +$$ +\mathcal {L} _ {d} \left(S _ {\theta} (\boldsymbol {x}, y)\right) = \mathcal {L} _ {c e} \left(S _ {\theta} (\boldsymbol {x}), y\right) + \beta \cdot \mathcal {H} \left(S _ {\theta} (\boldsymbol {x})\right), \tag {2} +$$ + +where $\beta$ is a hyper-parameter that balances the prediction accuracy and confidence, and $\mathcal{H}$ denotes the entropy of the prediction probabilities. + +# 4.2. Iterative Dataset Watermarking Framework + +During the dataset watermarking phase, we modify a portion of samples to produce the watermarked datasets with backdoor capabilities. We adopt an iterative training paradigm to gradually fine-tune the watermark generator $G_{\psi}$ , thereby enhancing the "toxicity" of the generated watermarked samples and then more effectively implanting the specific backdoor into models. + +Fig. 2 illustrates the iterative process of our method. Initially, we follow the standard image steganography training process [35] to pre-train the watermark generator on the original dataset and randomly select samples to form a subset. We then utilize the pre-trained generator (with fixed parameters) to watermark the selected subset and combine the modified samples with the remaining benign samples to train the surrogate model. Subsequently, we fine-tune + +the watermark generator on the modified subset guided by the trained surrogate model (with fixed parameters). The surrogate model is then employed to adaptively select samples for watermarking in a new iteration round. This iterative process, involving surrogate model training, generator fine-tuning, and adaptive data selection, is repeated to fine-tune the watermark generator and finally produce the watermarked dataset with backdoor capabilities. + +# 4.2.1. Surrogate Model Training + +We periodically train a surrogate model $S_{\theta}$ to simulate the real-time performance of the suspicious model $F$ on the current watermarked dataset $\mathcal{D}_w$ , aiding in the fine-tuning of the watermark generator $G_{\psi}$ . Operating under a clean-label framework, we watermark the samples without altering their labels: + +$$ +(\boldsymbol {x}, y) \Longrightarrow (G _ {\psi} (\boldsymbol {x}), y), \tag {3} +$$ + +where the original sample $(x,y)\in \mathcal{D}_s$ , and $\mathcal{D}_s$ is a selected subset from the original dataset. + +Under the black-box setting, we cannot control the training components (e.g., model architecture, hyperparameters, optimizer) during model training. Therefore, we follow the vallina training process to train $S_{\theta}$ on $\mathcal{D}_w$ using the cross-entropy loss, formulated as: + +$$ +\min _ {\theta} \mathbb {E} _ {(\boldsymbol {x}, y) \in \mathcal {D} _ {w}} \mathcal {L} _ {c e} \left(S _ {\theta} (\boldsymbol {x}), y\right), \tag {4} +$$ + +where $\mathcal{L}_{ce}$ denotes the cross-entropy loss, $S_{\theta}$ represents the surrogate model with parameters $\theta$ + +# 4.2.2. Generator Fine-tuning + +Directly fine-tuning the generator using the dispersion loss $\mathcal{L}_d$ is challenging for training convergence, as it would require differentiating through the surrogate model's training procedure [7]. To address this, we employ a heuristic method known as gradient matching for training optimization. Let $\mathcal{D}_v$ denote the patched verification set. We aim for our backdoor to be triggered on $\mathcal{D}_v$ . Therefore, we define the verification loss $\mathcal{L}_{ver}$ using the dispersion loss $\mathcal{L}_d$ : + +$$ +\mathcal {L} _ {v e r} = \mathbb {E} _ {(\boldsymbol {x}, y) \in \mathcal {D} _ {v}} \mathcal {L} _ {d} \left(S _ {\theta} (\boldsymbol {x} + \boldsymbol {p}, y)\right), \tag {5} +$$ + +where $\pmb{p}$ indicates the patch added to the original verification samples. Since the verification samples are private to dataset protectors, strong stealthiness is unnecessary. + +Then we simply employ the cross-entropy loss as the fine-tuning loss $\mathcal{L}_{train}$ on the modified subset $\mathcal{D}_s$ : + +$$ +\mathcal {L} _ {\text {t r a i n}} = \mathbb {E} _ {(\boldsymbol {x}, y) \in \mathcal {D} _ {s}} \mathcal {L} _ {\text {c e}} \left(S _ {\theta} \left(G _ {\psi} (\boldsymbol {x}), y\right)\right). \tag {6} +$$ + +Finally, we establish the following objective to align the verification loss $\mathcal{L}_{ver}$ with the fine-tuning loss $\mathcal{L}_{train}$ : + +$$ +\min _ {\psi} 1 - \frac {\nabla_ {\theta} \mathcal {L} _ {t r a i n} \cdot \nabla_ {\theta} \mathcal {L} _ {v e r}}{| | \nabla_ {\theta} \mathcal {L} _ {t r a i n} | | \cdot | | \nabla_ {\theta} \mathcal {L} _ {v e r} | |}. \tag {7} +$$ + +In this mode, we fine-tune the watermark generator $G_{\psi}$ using the cross-entropy loss to generate the watermarked dataset with backdoor capabilities. The watermarked models will then exhibit higher prediction confidence on the patched verification samples. This mode replaces direct solvers on watermark generator $G_{\psi}$ using the dispersion loss $\mathcal{L}_d$ and triggers the backdoor solely on a specific subset, reducing the difficulty of backdoor implantation. + +# 4.2.3. Adaptive Data Selection + +We adaptively select more "decisive" samples to be watermarked for better backdoor implantation. Specifically, we choose the preferred subset $\mathcal{D}_s$ with the smallest gradient norm of the cross-entropy loss: + +$$ +\mathcal {D} _ {s} = \underset {(\boldsymbol {x}, y) \in \mathcal {D}} {\arg \min } \left| \left| \nabla_ {\theta} \mathcal {L} _ {c e} \left(S _ {\theta} (\boldsymbol {x}), y\right) \right| \right|. \tag {8} +$$ + +Adaptive watermarked data selection greatly improves the stability of EntropyMark. By focusing on samples with smaller gradient norms, we can more effectively implant the desired backdoor into models. + +# 5. Dataset Ownership Verification + +EntropyMark ensures that watermarked models $\mathcal{M}_w$ show higher prediction confidence for patched verification samples compared to benign samples. To verify ownership, we use the entropy of the prediction probabilities as the verification metric and design a corresponding hypothesis-test-guided method: + +Suppose $F(\cdot)$ is the probability vector predicted by the suspicious model $F$ . Let $\pmb{x}$ denote a benign sample, and $\pmb{x}' = \pmb{x} + \pmb{p}$ denote its patched version. Denote $E_{b} = \mathcal{H}(F(\pmb{x}))$ and $E_{w} = \mathcal{H}(F(\pmb{x}'))$ as the entropy of the prediction probabilities for $\pmb{x}$ and $\pmb{x}'$ , respectively. Given the null hypothesis $H_{0}: E_{b} + \tau = E_{w}$ ( $H_{1}: E_{b} + \tau > E_{w}$ ) where the hyper-parameter $\tau \in [-1, 1]$ , it is claimed that the suspicious model is trained on the protected dataset if and only if $H_{0}$ is rejected. + +As shown in Fig. 2, in practice, we randomly select $m$ benign samples from the original dataset as the verification set. We then use these selected verification samples and their corresponding patched versions to conduct an entropy-based pairwise T-test [19] and compute the $p$ -value. A lower $p$ -value increases the likelihood of rejecting the null hypothesis $H_0$ , thereby indicating a higher probability that the suspicious model was trained on our protected dataset. In addition, we compute the confidence score $\Delta E = E_b - E_w$ to represent the verification confidence. + +# 6. Experiments + +We evaluate EntropyMark using two popular benchmark datasets, including CIFAR-10 [18] and a 12-class subset of ImageNet [23]. Please refer to Appendix A for more details on the datasets. + +# 6.1. Evaluation on Backdoor Watermark + +# 6.1.1. Settings + +We compare EntropyMark against seven representative BW-DOV methods, including three dirty-label backdoor watermarks (i.e., BadNets [9], Blend [3], and UBW-P [24]) and four clean-label backdoor watermarks (i.e., Label-Consistent [38], Sleeper Agent [33], UBW-C [24], and Domain-Watermark [10]). We uniformly set the watermarking rate $\gamma$ to $10\%$ . More information on baseline settings is available in Appendix C. + +In EntropyMark, ResNet-18 [12] is used as the network architecture for both surrogate and watermarked models, each trained from scratch with randomly initialized weights. As demonstrated in Sec. 7.2, EntropyMark can ensure transferability under the black-box setting. We set the source label $y_{s}$ to 0, and the dispersion loss hyper-parameter $\beta$ to 2.0. Further training configurations are provided in Appendix B. + +# 6.1.2. Metrics + +We evaluate the effectiveness of the backdoor watermarks using two key metrics: watermark success rate (WSR) and benign accuracy (BA). WSR represents the ratio of correctly predicted watermarked samples, while BA represents the ratio of correctly predicted benign samples. + +To quantitatively assess the harmfulness of different backdoor watermarks, we design the prediction consistency indices $\mathcal{C}_b$ and $\mathcal{C}_w$ . The indices measure the ratio of + +![](images/98bab555976ba6fe151220c6866c36403c1d1b94519745e13fbfda8cb484af66.jpg) +Figure 3. Examples in Various BW-DOV Methods: In all BW-DOV methods, the target label is set to 'cat', while the source label is 'shelf' for Sleeper Agent, UBW-C, and ours. Our method stands out by not altering the ground-truth label of any sample while ensuring more stealthy modifications. + +
TypeMethodCIFAR-10Sub-ImageNet-12
WSR (%)BA (%)CbCwWSR (%)BA (%)CbCw
-Clean Model-92.43---79.84--
Dirty-labelBadNets [9]10091.670.93130.107610080.160.82500.0872
Blend [3]10091.620.92780.036110079.040.81960.0888
UBW-P [24]82.2090.700.91770.116763.4675.770.81440.2022
Clean-labelLabel-Consistent [38]94.4592.360.95180.067599.5173.300.74710.0065
Sleeper Agent [33]86.3291.540.91120.23018.2180.040.81250.7821
UBW-C [24]86.5087.870.87600.763948.8570.640.70770.7090
Domain-Watermark [10]99.4090.270.91630.143051.15*47.69*0.4750*0.4756*
EntropyMark (ours)87.0087.790.89360.869569.6271.350.75100.7471
+ +Table 1. Watermark performance on CIFAR-10 and Sub-ImageNet-12. Harmful $(\mathcal{C}_w < 0.25$ or $c_{b} < 0.25)$ or abnormal results (WSR $<$ $10\%$ or $BA_{drop} > 30\%$ ) are marked in red. Results marked with $\ast$ are reproduced on our own. + +samples with consistent prediction results between watermarked models and clean models: + +$$ +\mathcal {C} = \frac {1}{| \mathcal {D} |} \sum_ {x \in \mathcal {D}} \mathbb {I} \left\{F _ {w} (\boldsymbol {x}) = F _ {b} (\boldsymbol {x}) \right\}, \tag {9} +$$ + +where $F_{w}$ and $F_{b}$ denote the watermarked model and the benign model, respectively. $\mathbb{I}$ represents the indicator function. This formula applies to $\mathcal{C}_w$ and $\mathcal{C}_b$ on the fullwatermarked $\mathcal{D}_f$ and benign datasets $\mathcal{D}$ , respectively. + +# 6.1.3. Results + +Fig. 3 presents examples generated by different methods. Modified samples in BadNets, Blend, UBW-P, and Label Consistent all adopt static patches or patterns, making their modifications relatively easy to detect. Although Sleeper Agent and UBW-C apply more subtle perturbations, the modified samples still exhibit observable noise. In contrast, + +EntropyMark achieves a nearly imperceptible modification, demonstrating superior stealthiness. + +As shown in Tab. 1, the WSR and BA of EntropyMark are comparable to those of other clean-label backdoor watermarks. Due to the clean-label setting and specific nature of our backdoor, the WSR of our watermarked model is restricted by BA. However, this limitation does not hinder the effectiveness of the backdoor implantation, as we can still trace the unauthorized model through significant entropy inconsistency. Through extensive trials with different methods, we find that EntropyMark is more stable, with each metric value exhibiting minimal fluctuations. + +EntropyMark achieves prediction consistency indices $\mathcal{C}_b$ and $\mathcal{C}_w$ greater than 0.7, demonstrating optimal harmlessness. For most backdoor watermarks, watermarked models tend to exhibit flaws on malicious samples, resulting in extremely low $C_w$ . In contrast, EntropyMark preserves + +
CIFAR-10Sub-ImageNet-12
WatermarkBenignOtherWatermarkBenignOther
p-value0.00171.00001.00000.00181.00000.8375
ΔE0.00370.00820.0126-0.00310.0002-0.0003
+ +Table 2. The effectiveness of dataset ownership verification via our EntropyMark. + +
Iterative Fine-tuningGradient MatchingAdaptive SelectionWSR (%)BA (%)WatermarkBenignOther
XXX88.8288.690.07640.99980.0001
✓XX88.9689.030.81570.98370.5794
✓X✓86.8389.000.00390.98370.8655
✓✓X73.1887.280.00521.00001.0000
✓✓✓82.4787.790.00171.00001.0000
+ +Table 3. Ablation studies on CIFAR-10 with ResNet-18. Verification failures ( $p$ -value in 'Watermark' greater than those in 'Benign' or 'Other') are highlighted in red, and verification successes ( $p$ -value in 'Watermark' smaller than those in 'Benign' and 'Other') are highlighted in green. The same highlighting convention is applied in the following tables. + +dataset functionality, and corresponding watermarked models do not show obvious backdoor vulnerabilities, thus avoiding malicious prediction manipulation. + +# 6.2. Evaluation on Dataset Ownership Verification + +# 6.2.1. Settings + +We validate EntropyMark in three representative scenarios, including 1) Watermark: Detect watermarked models using our patched verification samples; 2) Benign: Detect benign models using the same patched verification samples; 3) Other: Detect watermarked models using patched samples with different patterns. + +# 6.2.2. Metrics + +Our evaluation employs $p$ -value and $\Delta E$ , where smaller values indicate higher accuracy in the 'Watermark' scenario, while higher values are anticipated in the 'Benign' and 'Other' scenarios. This approach effectively distinguishes models trained on our watermarked dataset from clean models and those trained on other protected datasets. + +# 6.2.3. Results + +As shown in Tab. 2, our verification provides consistently correct and deterministic results across all three scenarios. In the 'Watermark' scenario, we achieve a very low $p$ -value ( $p$ -value $\ll 0.002$ ), indicating a strong likelihood of rejecting the null hypothesis. Conversely, the 'Benign' and 'Other' scenarios yield significantly higher $p$ -values ( $p$ -value $\gg 0.5$ ). This differentiation aligns with our verification goals. In addition, $\Delta E$ in the 'Watermark' scenario is smaller than in the 'Benign' and 'Other' scenarios, demonstrating high verification confidence. + +It is noteworthy that watermarked datasets may slightly affect dataset functionality, leading to a subtle decline in + +prediction confidence on patched verification samples, especially in the CIFAR-10 dataset. Fortunately, this influence is effectively suppressed in the 'Watermark' scenario, enabling successful and reliable verification. + +# 7. Discussion + +# 7.1. Ablation Studies + +In our ablation studies, we evaluate the contribution of each technique in EntropyMark, including 1) iterative fine-tuning, 2) gradient matching, and 3) adaptive selection. + +Tab. 3 presents the results of various combinations of these techniques. Without fine-tuning, the pre-trained model cannot produce watermarked datasets with effective backdoor capabilities. Adding gradient matching refines the fine-tuning, significantly boosting verification accuracy. The adaptive selection mechanism further enhances verification, making it effective even without gradient matching. When all three techniques are combined, the watermarked model yields a high WSR and BA, creating the largest $p$ -value gap between the 'Watermark' scenario and the other two scenarios, indicating optimal verification performance. + +# 7.2. Black-box Evaluation + +We evaluate the transferability of EntropyMark in the black-box setting. Specifically, we fix the training components of our surrogate model while varying the corresponding components of the suspicious models to be verified. + +As shown in Tab. 4, EntropyMark achieves a large $p$ -value gap between the 'Watermark' scenario and the other two scenarios, effectively distinguishing among the three scenarios. These results demonstrate that our method can verify suspicious models in the black-box setting with relatively stable WSR and BA, indicating strong transferability + +
ArchitectureBatch SizeOptimizerLearning RateWSR (%)BA (%)WatermarkBenignOther
ResNet-18128SGD0.0585.0688.200.00711.00000.5772
ResNet-1864SGD0.0584.6788.490.15390.86981.0000
ResNet-18128AdamW0.1074.8680.720.32661.00000.8113
ResNet-34128SGD0.1085.2888.440.36160.94671.0000
VGG-19128SGD0.1082.8487.090.46611.00000.7026
ShuffleNet128SGD0.1083.1788.230.39961.00000.5911
GoogleNet128SGD0.1088.0692.260.22530.96020.8639
+ +Table 4. The results under the black-box setting for model architectures and hyper-parameters in the training process. For the surrogate model, we fix its training settings to use ResNet-18 as the model, 128 as the batch size, SGD as the optimizer, and 0.1 as the learning rate. + +
RWSR (%)BA (%)WatermarkBenignOther
184.8388.190.49250.99980.6289
282.4787.790.00171.00001.0000
380.4388.040.00790.98700.9977
469.1787.080.44391.00000.7764
572.4787.040.88850.61290.3223
+ +Table 5. The effects of iteration rounds $R$ . + +
γWSR (%)BA (%)WatermarkBenignOther
0.0185.0788.860.05040.98700.9998
0.0581.1188.500.00181.00000.9999
0.1082.4787.790.00171.00001.0000
0.2079.7187.100.65831.00000.2552
+ +Table 6. The effects of watermarking rate $\gamma$ . + +and applicability to real-world scenarios. + +# 7.3. Parameters Analysis + +We analyze the impact of key hyper-parameters on the effectiveness of our method. Further analysis on additional hyper-parameters is provided in Appendix D. + +# 7.3.1. Iteration Rounds + +Results in Tab. 5 indicate that: (1) As $R$ increases, WSR tends to decrease, while BA remains relatively stable; (2) With too few iterations, $p$ -values in the 'Watermark' and 'Other' scenarios are not significantly different, potentially causing confusion between different trigger patterns; (3) With too many iterations, WSR drops sharply, weakening backdoor capabilities. This effect may result from the entropy-based constraint becoming stronger with each iteration. As a consequence, the feature representations of watermarked samples start to cluster, which eventually leads to model misclassification. These results suggest that setting $R = 2$ enables the watermarked dataset to retain the strongest backdoor capabilities. + +# 7.3.2. Watermarking Rate + +As shown in Tab. 6, at lower watermarking rates (1%-10%), both WSR and BA remain stable. However, when $\gamma$ is + +![](images/56226b4270987dfdf45270a5aa7db66768d449fdfc5d24b0498f145387d954da.jpg) +Figure 4. Resistance to defenses on CIFAR-10. + +too high $(20\%)$ , the $p$ -value in the 'Other' scenario drops sharply, leading to verification failures. This likely happens because a higher watermarking rate increases the susceptibility of watermarked models to being triggered by other watermark patterns. These results indicate that a watermarking rate between $5\%-10\%$ achieves optimal verification performance. + +# 7.4. Resistance to Defenses + +Dataset stealers may employ defenses to bypass backdoor watermarks during model training. We evaluate two practical backdoor removal methods: fine-tuning and model pruning [28]. + +As shown in Fig. 4, model pruning fails to remove our backdoor without sacrificing BA while fine-tuning only has minor effects on both WSR and BA. Since EntropyMark is sample-specific, it naturally resists most conventional sample-filtering defenses. For further details on additional defenses, please refer to Appendix E. + +# 8. Conclusion + +In this work, we propose a more harmless backdoor watermark that influences prediction confidence without altering prediction results. In addition, we design an iterative clean-label dataset watermarking framework under an entropy-based constraint. For better backdoor implantation, we utilize gradient matching and adaptive data selection to optimize training convergence. Extensive experiments on benchmark datasets validate the effectiveness, transferability, and resistance to defenses of our method. + +Acknowledgments. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.62272020. + +# References + +[1] Jiawang Bai, Yiming Li, Jiawei Li, Xue Yang, Yong Jiang, and Shu-Tao Xia. Multinomial random forest. Pattern Recognition, 122:108331, 2022. 1, 3 +[2] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 2 +[3] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 5, 6 +[4] Hua Deng, Zheng Qin, Qianhong Wu, Zhenyu Guan, Robert H Deng, Yujue Wang, and Yunya Zhou. Identity-based encryption transformation for flexible sharing of encrypted data in public cloud. IEEE Transactions on Information Forensics and Security, 15:3168-3180, 2020. 1, 3 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1 +[6] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference, pages 113-125, 2019. 3 +[7] Jonas Geiping, Liam Fowl, W Ronny Huang, Wojciech Czaja, Gavin Taylor, Michael Moeller, and Tom Goldstein. Witches' brew: Industrial scale data poisoning via gradient matching. arXiv preprint arXiv:2009.02276, 2020. 5 +[8] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020. 1 +[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 2, 5, 6 +[10] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: effective and harmless dataset copyright protection is closed at hand. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 54421-54450, 2023. 2, 3, 5, 6 +[11] Yuanfang Guo, Oscar C Au, Rui Wang, Lu Fang, and Xiaochun Cao. Halftone image watermarking by content aware double-sided embedding error diffusion. IEEE Transactions on Image Processing, 27(7):3387-3402, 2018. 1, 3 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a + +two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4 +[14] Robert V Hogg, Joseph W McKean, Allen T Craig, et al. Introduction to mathematical statistics. Pearson Education India, 2013. 3 +[15] Chiou-Ting Hsu and Ja-Ling Wu. Hidden digital watermarks in images. IEEE Transactions on image processing, 8(1):58-68, 1999. 1, 3 +[16] Hongsheng Hu, Zoran Salcic, Gillian Dobbie, Jinjun Chen, Lichao Sun, and Xuyun Zhang. Membership inference via backdoorsing. arXiv preprint arXiv:2206.04823, 2022. 2, 3 +[17] Wenbo Jiang, Hongwei Li, Guowen Xu, and Tianwei Zhang. Color backdoor: A robust poisoning attack in color space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8133-8142, 2023. 2 +[18] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 1, 5 +[19] Richard J Larsen and Morris L Marx. An introduction to mathematical statistics. Prentice Hall Hoboken, NJ, 2005. 5 +[20] Jiguo Li, Qihong Yu, and Yichen Zhang. Hierarchical attribute based encryption with continuous leakage-resilience. Information Sciences, 484:113-134, 2019. 3 +[21] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16463-16472, 2021. 2 +[22] Yiming Li, Peidong Liu, Yong Jiang, and Shu-Tao Xia. Visual privacy protection via mapping distortion. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3740-3744. IEEE, 2021. 3 +[23] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 5, 1 +[24] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 5, 6 +[25] Yiming Li, Mengxi Ya, Yang Bai, Yong Jiang, and Shu-Tao Xia. Backdoorbox: A python toolbox for backdoor learning. arXiv preprint arXiv:2302.01762, 2023. 2 +[26] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 2, 3 +[27] Yi-Shan Lin, Wen-Chuan Lee, and Z Berkay Celik. What do you see? evaluation of explainable artificial intelligence (xai) interpretability through neural backdoors. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 1027-1035, 2021. 2 +[28] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in + +attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 8 +[29] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 182-199. Springer, 2020. 2 +[30] Anh Nguyen and Anh Tran. Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369, 2021. 2 +[31] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 3 +[32] Shawn Shan, Emily Wenger, Bolun Wang, Bo Li, Haitao Zheng, and Ben Y Zhao. Gotta catch'em all: Using honeypots to catch adversarial attacks on neural networks. In Proceedings of the 2020 ACM SIGSAC conference on computer and communications security, pages 67-83, 2020. 2 +[33] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. Advances in Neural Information Processing Systems, 35: 19165-19178, 2022. 2, 5, 6 +[34] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 4 +[35] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4, 1 +[36] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 3 +[37] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 3 +[38] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 5, 6 +[39] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1 +[40] Shulan Wang, Junwei Zhou, Joseph K Liu, Jianping Yu, Jianyong Chen, and Weixin Xie. An efficient file hierarchy attribute-based encryption scheme in cloud computing. IEEE Transactions on Information Forensics and Security, 11(6): 1265-1277, 2016. 1, 3 +[41] Zuobin Xiong, Zhipeng Cai, Qilong Han, Arwa Alrawais, and Wei Li. Adgan: Protect your location privacy in camera data of auto-driving vehicles. IEEE Transactions on Industrial Informatics, 17(9):6200-6210, 2020. 3 +[42] Mengxi Ya, Yiming Li, Tao Dai, Bin Wang, Yong Jiang, and Shu-Tao Xia. Towards faithful xai evaluation via generalization-limited backdoor watermark. In The + +Twelfth International Conference on Learning Representations, 2023. 2 +[43] Runkai Zheng, Rongjun Tang, Jianze Li, and Li Liu. Data-free backdoor removal based on channel lipschitzness. arXiv preprint arXiv:2208.03111, 2022. 4 +[44] Linghui Zhu, Xinyi Liu, Yiming Li, Xue Yang, Shu-Tao Xia, and Rongxing Lu. A fine-grained differentially private federated learning against leakage from gradients. IEEE Internet of Things Journal, 9(13):11500-11512, 2021. 1, 3 \ No newline at end of file diff --git a/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/images.zip b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c5a66a2d1d071beacfc07b03d1e332ea8694ee0 --- /dev/null +++ b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f0827893d8024a9a4b6c63847453d9068de053bd603b3f018d83acdb42ee09f +size 599583 diff --git a/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/layout.json b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3468828f39318faad58df634b38216844964157c --- /dev/null +++ b/2025/EntropyMark_ Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection/layout.json @@ -0,0 +1,9945 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 140 + ], + "type": "text", + "content": "EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "text", + "content": "Ming Sun" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "text", + "content": ", Rui Wang" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "text", + "content": ", Zixuan Zhu" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "text", + "content": ", Lihua Jing" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "text", + "content": ", Yuanfang Guo" + }, + { + "bbox": [ + 130, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 175, + 473, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 175, + 473, + 189 + ], + "spans": [ + { + "bbox": [ + 134, + 175, + 473, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 175, + 473, + 189 + ], + "type": "text", + "content": " Institute of Information Engineering, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 189, + 479, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 189, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 131, + 189, + 479, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 131, + 189, + 479, + 202 + ], + "type": "text", + "content": " School of Cyber Security, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 203, + 469, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 203, + 469, + 217 + ], + "spans": [ + { + "bbox": [ + 140, + 203, + 469, + 217 + ], + "type": "text", + "content": "3 School of Computer Science and Engineering, Beihang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 220, + 494, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 220, + 494, + 231 + ], + "spans": [ + { + "bbox": [ + 128, + 220, + 494, + 231 + ], + "type": "text", + "content": "nming, wangrui, zhuzixuan, jinglihua}@iie.ac.cn, andyguo@buaa.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 152, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 284, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 296, + 619 + ], + "type": "text", + "content": "High-quality open-source datasets are essential for advancing deep neural networks. However, the unauthorized commercial use of these datasets has raised significant concerns about copyright protection. One promising approach is backdoor watermark-based dataset ownership verification (BW-DOV), in which dataset protectors implant specific backdoors into illicit models through dataset watermarking, enabling the tracing of these models through abnormal prediction behaviors. Unfortunately, the targeted nature of these BW-DOV methods can be maliciously exploited, potentially leading to harmful side effects. While existing harmless methods attempt to mitigate these risks, watermarked datasets can still negatively affect prediction results, partially compromising dataset functionality. In this paper, we propose a more harmless backdoor watermark, called EntropyMark, which improves prediction confidence without altering the final prediction results. For this purpose, an entropy-based constraint is introduced to regulate the probability distribution. Specifically, we design an iterative clean-label dataset watermarking framework. Our framework employs gradient matching and adaptive data selection to optimize backdoor injection. In parallel, we introduce a hypothesis test method grounded in entropy inconsistency to verify dataset ownership. Extensive experiments on benchmark datasets demonstrate the effectiveness, transferability, and defense resistance of our approach. Our code is available at https://github.com/AaronSun2000/EntropyMark." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "spans": [ + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 661, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 295, + 696 + ], + "type": "text", + "content": "In recent years, deep neural networks (DNNs) have rapidly advanced and been widely deployed in numerous real-world applications, such as autonomous driving [8] and face" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 553, + 424 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 553, + 424 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 424 + ], + "type": "image", + "image_path": "0c94024f018bb9f1eced0dac4a858983c4e562b75b100696ecf1becf9600b5a0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 425, + 382, + 434 + ], + "lines": [ + { + "bbox": [ + 348, + 425, + 382, + 434 + ], + "spans": [ + { + "bbox": [ + 348, + 425, + 382, + 434 + ], + "type": "text", + "content": "I. Training" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 425, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 466, + 425, + 504, + 433 + ], + "spans": [ + { + "bbox": [ + 466, + 425, + 504, + 433 + ], + "type": "text", + "content": "II. Inference" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 442, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. Comparison of EntropyMark with existing back-door watermarks: Existing watermarks alter the prediction results of backdoored models and produce conspicuous modifications on datasets. In contrast, EntropyMark preserves the original results and is significantly more stealthy." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 521, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 653 + ], + "type": "text", + "content": "recognition [39]. The prosperity of DNNs largely depends on the availability of high-quality datasets (e.g., CIFAR [18] and ImageNet [5]). Researchers use open-source datasets to evaluate and iteratively optimize their proposed methods. However, the value of these datasets has attracted dataset stealers who illegally copy, distribute and use them without permission. Currently, most open-source datasets are intended solely for educational or research purposes and explicitly prohibit unauthorized commercial use1. Therefore, copyright protection for open-source datasets has become an urgent and necessary topic." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "text", + "content": "Unfortunately, classic data protection methods, including differential privacy [1, 44], encryption [4, 40], and digital watermarking [11, 15], are not well-suited for open-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 702, + 438, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 438, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 438, + 712 + ], + "type": "text", + "content": "1See IMDb data policy for details." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30692" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "source datasets, as they are primarily designed for private data protection. Safeguarding open-source datasets presents more challenges because their protectors lack control over the training process and have no access to training components. Instead, they can only query the outputs of suspicious models to determine whether those models were trained on the protected datasets. In addition, due to the open-source nature of these datasets, any protection strategy must avoid compromising dataset functionality or altering the integrity of information representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 294, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 301 + ], + "type": "text", + "content": "Backdoor watermark-based dataset ownership verification (BW-DOV) [26] offers a feasible solution for protecting open-source dataset copyrights. This approach involves modifying samples in the original dataset and releasing the watermarked version for open-source use. When a model is trained on this watermarked dataset, a specific backdoor is implanted. To verify whether a suspicious model was trained on the protected dataset, protectors can simply detect the presence of the predetermined backdoor." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 496 + ], + "type": "text", + "content": "However, due to the targeted nature (i.e., specified erroneous prediction results) of existing BW-DOV methods, watermarked models are vulnerable to manipulation and may maliciously output predetermined results. To address this issue, UBW [24] proposes an untargeted backdoor watermark, which mitigates the targeted nature of the backdoor by increasing prediction randomness on watermarked samples. Domain-Watermark [10] ensures that watermarked models correctly classify certain 'hard' samples that benign models may misclassify. However, these methods still alter the original prediction results, inevitably impacting dataset functionality. To address the limitations of existing BW-DOV methods, we argue that harmless backdoor watermarks should not only avoid generating specified erroneous prediction results but also preserve the original prediction outputs, thereby maintaining dataset functionality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose EntropyMark, a more harmless backdoor watermark designed to improve prediction confidence for verification samples without affecting overall prediction results. To ensure that the prediction probabilities for these verification samples are more concentrated on the ground-truth label, we introduce an entropy-based constraint to regulate the probability distribution. For implementation, we design an iterative clean-label dataset watermarking framework. Within this framework, we employ gradient matching to optimize training convergence, ensuring effective backdoor injection into models trained on the watermarked dataset. We also utilize gradient-based adaptive selection to identify more \"decisive\" samples for watermarking. Fig. 1 compares EntropyMark with existing BW-DOV methods, emphasizing the harmlessness and stealthiness of our method. Unlike radioactive data [31], the classic data tracing method through watermarking, EntropyMark exhibits greater robustness in black-box settings and offers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "enhanced resilience against various defenses. Finally, to further improve verification accuracy, we incorporate a pairwise T-test based on entropy inconsistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 109, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 109, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 109, + 547, + 120 + ], + "type": "text", + "content": "The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 121, + 553, + 300 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 307, + 121, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 121, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 121, + 553, + 191 + ], + "type": "text", + "content": "(1) We propose EntropyMark, a backdoor watermark under an entropy-based constraint. Based on our backdoor watermark, we design an iterative clean-label dataset watermarking framework. EntropyMark achieves enhanced stealthiness and minimal harmfulness while retaining watermark effectiveness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 193, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 193, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 193, + 553, + 228 + ], + "type": "text", + "content": "(2) Our framework adaptively selects \"decisive\" samples to be watermarked and incorporates gradient matching to enhance training convergence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 229, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 553, + 264 + ], + "type": "text", + "content": "(3) Building on the properties of EntropyMark, we develop an entropy inconsistency hypothesis test to quantitatively verify dataset ownership." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 265, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 553, + 300 + ], + "type": "text", + "content": "(4) Extensive experiments on benchmark datasets validate the effectiveness, transferability, and defense resistance of our method." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 313, + 400, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 313, + 400, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 313, + 400, + 326 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 334, + 421, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 334, + 421, + 345 + ], + "spans": [ + { + "bbox": [ + 314, + 334, + 421, + 345 + ], + "type": "text", + "content": "2.1. Backdoor Attacks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 352, + 424, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 352, + 424, + 363 + ], + "spans": [ + { + "bbox": [ + 314, + 352, + 424, + 363 + ], + "type": "text", + "content": "2.1.1. Dirty-label Attacks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 367, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 462 + ], + "type": "text", + "content": "Dirty-label attacks forcibly modify the labels of selected malicious samples. BadNets [9], the first dirty-label attack, uses a static patch as the trigger. To enhance stealthiness, Blend [3] generates malicious samples by blending the trigger with original images. Subsequently, more stealthy and effective methods [17, 21, 30] have been proposed, varying the forms of triggers. However, dirty-label attacks are still detectable due to label-semantic inconsistencies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 471, + 426, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 471, + 426, + 482 + ], + "spans": [ + { + "bbox": [ + 314, + 471, + 426, + 482 + ], + "type": "text", + "content": "2.1.2. Clean-label Attacks" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 487, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 606 + ], + "type": "text", + "content": "Clean-label attacks, which retain original sample labels, are well-suited for scenarios where attackers lack control over label settings. Label-Consistent [38] introduces the first patch-based attack under this paradigm. However, many early clean-label attacks [2, 29] require access to a pre-trained model and often perform poorly on higher-resolution datasets or those with numerous categories. Sleeper Agent [33] addresses these limitations by using gradient matching in a bi-level optimization process to generate poisoned samples." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 614, + 427, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 614, + 427, + 625 + ], + "spans": [ + { + "bbox": [ + 314, + 614, + 427, + 625 + ], + "type": "text", + "content": "2.1.3. Positive Application" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Positive applications of backdoor attacks are also being exploited, such as membership reasoning [16], algorithm evaluation [27, 42], and adversarial defense [32]. Backdoor watermark [26] has also emerged as an effective solution for protecting open-source datasets. However, existing methods still interfere with the original prediction results of watermarked models, thus compromising dataset functionality." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30693" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 152, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 152, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 152, + 83 + ], + "type": "text", + "content": "2.2. Data Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 183, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 183, + 100 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 183, + 100 + ], + "type": "text", + "content": "2.2.1. Classic Data Protection" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 104, + 296, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 296, + 212 + ], + "type": "text", + "content": "Data protection aims to prevent unauthorized use and safeguard privacy. Common methods include (1) Digital Watermarking: Embedding owner-specific tags to assert ownership [11, 15]. (2) Encryption: Encrypting data, which can only be accessible with a secret key, securing sensitive information [4, 20, 22, 40, 41]. (3) Differential Privacy: Adding randomness (e.g., Laplace noise) to ensure indistinguishable outputs from neighboring datasets samples, preserving membership privacy [1, 44]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 344 + ], + "type": "text", + "content": "By combining digital watermarking with poisoning-based backdoor attacks, radioactive data [31] trace protected data by increasing prediction confidence for certain categories. Specifically, it uses carriers to shift the feature representation of samples within a particular category toward a predetermined direction, thereby aiding classification. However, this technique requires a high poisoning rate (over " + }, + { + "bbox": [ + 55, + 213, + 296, + 344 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 213, + 296, + 344 + ], + "type": "text", + "content": " ) and lacks rigorous hypothesis testing in the black-box setting. Moreover, the isotropic carriers used as watermarks make radioactive data susceptible to detection through statistical defenses (e.g., subspace analysis)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 209, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 209, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 209, + 361 + ], + "type": "text", + "content": "2.2.2. Dataset Copyright Protection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 594 + ], + "type": "text", + "content": "Dataset copyright protection is an emerging topic within the data protection field, aiming to verify whether a suspicious model has been trained on a protected dataset. Conventional data protection methods are not feasible for open-source datasets, as they either require access to training components or compromise dataset functionality. Backdoor watermark-based dataset ownership verification (BW-DOV) [26, 36] presents a feasible solution for protecting open-source datasets. DVBW [26] is the first to propose employing backdoor watermarks in original datasets and then detecting abnormal backdoors in suspicious models. To mitigate the harmfulness of backdoor watermarks, UBW [24] introduces an untargeted backdoor watermark under both dirty-label and clean-label settings. DomainWatermark [10] proposes a more harmless clean-label backdoor, where watermarked models correctly classify \"hard\" samples that benign models may misclassify. However, these methods still cause results inconsistency predicted by watermarked and benign models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 188, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 188, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 188, + 616 + ], + "type": "text", + "content": "3. BW-DOV Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 162, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 162, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 162, + 636 + ], + "type": "text", + "content": "3.1. Black-box Setting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "The main pipeline of existing BW-DOV methods consists of two phases: dataset watermarking and dataset ownership verification. During the dataset watermarking phase, they can only watermark a portion of the benign dataset and have no access to training components (e.g., model architecture, loss function). During the dataset ownership verification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "phase, they can only obtain the prediction probabilities for the verification samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 102, + 463, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 463, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 463, + 114 + ], + "type": "text", + "content": "3.2. Main Pipeline of BW-DOV" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 120, + 441, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 441, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 441, + 132 + ], + "type": "text", + "content": "3.2.1. Dataset Watermarking" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " represent the original dataset, where image " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X} = [0,1]^{C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": ", and label " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{Y} = \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": ". A small selected subset " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s \\subset \\mathcal{D}" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": ") is watermarked to produce a modified version " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m = \\{G_\\psi(x_t), \\eta(y_t) | (x_t, y_t) \\in \\mathcal{D}_s\\}" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "G_\\psi: \\mathcal{X} \\to \\mathcal{X}" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " indicates the watermark generator " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " indicates target label transformation. The modified samples " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " are combined with the remaining benign samples " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_b" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_b = \\mathcal{D} \\setminus \\mathcal{D}_s" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": ") to create the watermarked dataset " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_w" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_w = \\mathcal{D}_m + \\mathcal{D}_b" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": "). The watermarking rate " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\gamma = |\\mathcal{D}_s| / |\\mathcal{D}|" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " quantifies the proportion of modified samples. " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_w" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " is finally released in place of the original dataset " + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 134, + 555, + 278 + ], + "type": "text", + "content": " for protection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 283, + 476, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 476, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 476, + 295 + ], + "type": "text", + "content": "3.2.2. Dataset Ownership Verification" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "content": "Dataset protectors aim to verify whether a suspicious model " + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "content": " has been trained on the protected dataset " + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_w" + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "content": " by identifying the specific implanted backdoors. These backdoors can be detected using hypothesis-test-based methods (e.g., pairwise T-test, Wilcoxon-test [14]) on verification sets " + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 368, + 447, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 447, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 447, + 380 + ], + "type": "text", + "content": "4. Dataset Watermarking" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 387, + 554, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 436 + ], + "type": "text", + "content": "As a more harmless BW-DOV method, EntropyMark implants a result-preserving backdoor under an entropy-based constraint. Specifically, it employs an iterative clean-label dataset watermarking framework." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 441, + 469, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 469, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 469, + 454 + ], + "type": "text", + "content": "4.1. Result-preserving Backdoor" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": "Unlike conventional backdoors, EntropyMark improves prediction confidence while preserving original prediction results. This introduces a distributional bias in prediction probabilities, thereby disrupting the assumption of distribution consistency across samples from the same dataset [16]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 524, + 454, + 535 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 454, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 454, + 535 + ], + "type": "text", + "content": "4.1.1. Entropy-based Constraint" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 539, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 562 + ], + "type": "text", + "content": "Entropy is a measure of uncertainty in prediction probabilities, which can be formulated as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 383, + 567, + 555, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 567, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 383, + 567, + 555, + 600 + ], + "type": "interline_equation", + "content": "\\mathcal {H} (x) = - \\sum_ {i = 1} ^ {N} p _ {i} \\log p _ {i}, \\tag {1}", + "image_path": "06a7cc9183f989d4ae5a011aec3e9d6ebdee2af6da0a2119d54d1ce9acd07c25.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "content": " represents the probability of sample " + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "content": " being classified as the " + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "content": " label, and " + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "content": " is the number of possible labels." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "The entropy-based constraint is introduced to ensure that watermarked samples are more confidently predicted by the backdoored models, thus introducing a distributional bias between the watermarked and benign samples. As a result, dataset protectors can verify whether the model was trained on the protected dataset by checking for entropy inconsistency between benign and watermarked samples." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30694" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 257 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 257 + ], + "type": "image", + "image_path": "daea1abcdd2f163a47dd289b478e26cf6609b080c09cfa6d37e296554938aa55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. Main Pipeline of EntropyMark: In the dataset watermarking phase, we fine-tune a watermark generator with the aid of a surrogate model. This fine-tuned generator is then used to produce the watermarked dataset with backdoor capabilities. In the dataset ownership verification phase, we verify whether the suspicious model has been trained on our protected dataset by checking for entropy inconsistency in its predictions on benign and patched samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 180, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 180, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 180, + 342 + ], + "type": "text", + "content": "4.1.2. Objective Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 430 + ], + "type": "text", + "content": "However, as the constraint may force watermarked samples to aggregate the feature representation toward one independent cluster, relying solely on the entropy-based constraint could lead to misclassification. To address this, we incorporate the cross-entropy loss to mitigate the negative impact on prediction accuracy. Specifically, we design the dispersion loss " + }, + { + "bbox": [ + 55, + 346, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 346, + 296, + 430 + ], + "type": "text", + "content": ", which serves as the objective of our backdoor:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 451, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 451, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 70, + 451, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} \\left(S _ {\\theta} (\\boldsymbol {x}, y)\\right) = \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right) + \\beta \\cdot \\mathcal {H} \\left(S _ {\\theta} (\\boldsymbol {x})\\right), \\tag {2}", + "image_path": "1048671678de7b91e241919b67bfac5ba4cd1bea6e4caf2b795f1aad86692a20.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": " is a hyper-parameter that balances the prediction accuracy and confidence, and " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": " denotes the entropy of the prediction probabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 287, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 287, + 529 + ], + "type": "text", + "content": "4.2. Iterative Dataset Watermarking Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "type": "text", + "content": "During the dataset watermarking phase, we modify a portion of samples to produce the watermarked datasets with backdoor capabilities. We adopt an iterative training paradigm to gradually fine-tune the watermark generator " + }, + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "type": "inline_equation", + "content": "G_{\\psi}" + }, + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "type": "text", + "content": ", thereby enhancing the \"toxicity\" of the generated watermarked samples and then more effectively implanting the specific backdoor into models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Fig. 2 illustrates the iterative process of our method. Initially, we follow the standard image steganography training process [35] to pre-train the watermark generator on the original dataset and randomly select samples to form a subset. We then utilize the pre-trained generator (with fixed parameters) to watermark the selected subset and combine the modified samples with the remaining benign samples to train the surrogate model. Subsequently, we fine-tune" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 427 + ], + "type": "text", + "content": "the watermark generator on the modified subset guided by the trained surrogate model (with fixed parameters). The surrogate model is then employed to adaptively select samples for watermarking in a new iteration round. This iterative process, involving surrogate model training, generator fine-tuning, and adaptive data selection, is repeated to fine-tune the watermark generator and finally produce the watermarked dataset with backdoor capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 454, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 454, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 454, + 449 + ], + "type": "text", + "content": "4.2.1. Surrogate Model Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "content": "We periodically train a surrogate model " + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "content": " to simulate the real-time performance of the suspicious model " + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "content": " on the current watermarked dataset " + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_w" + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "content": ", aiding in the fine-tuning of the watermark generator " + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "inline_equation", + "content": "G_{\\psi}" + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "content": ". Operating under a clean-label framework, we watermark the samples without altering their labels:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 386, + 538, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 538, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 386, + 538, + 553, + 552 + ], + "type": "interline_equation", + "content": "(\\boldsymbol {x}, y) \\Longrightarrow (G _ {\\psi} (\\boldsymbol {x}), y), \\tag {3}", + "image_path": "43626043df6f3744e1882ed0ccb2ce70980e11870ff75ce849e50823b895d972.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 554, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 587 + ], + "type": "text", + "content": "where the original sample " + }, + { + "bbox": [ + 313, + 563, + 554, + 587 + ], + "type": "inline_equation", + "content": "(x,y)\\in \\mathcal{D}_s" + }, + { + "bbox": [ + 313, + 563, + 554, + 587 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 563, + 554, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 313, + 563, + 554, + 587 + ], + "type": "text", + "content": " is a selected subset from the original dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 589, + 554, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 554, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 554, + 648 + ], + "type": "text", + "content": "Under the black-box setting, we cannot control the training components (e.g., model architecture, hyperparameters, optimizer) during model training. Therefore, we follow the vallina training process to train " + }, + { + "bbox": [ + 313, + 589, + 554, + 648 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 313, + 589, + 554, + 648 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 589, + 554, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_w" + }, + { + "bbox": [ + 313, + 589, + 554, + 648 + ], + "type": "text", + "content": " using the cross-entropy loss, formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 372, + 661, + 553, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 661, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 372, + 661, + 553, + 678 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {w}} \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right), \\tag {4}", + "image_path": "6e84e3b7e6f9e2d473a71c29d81e806b09cb4920b0812645d3e63ac3ea452ee1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " denotes the cross-entropy loss, " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " represents the surrogate model with parameters " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30695" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 181, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 181, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 181, + 85 + ], + "type": "text", + "content": "4.2.2. Generator Fine-tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "text", + "content": "Directly fine-tuning the generator using the dispersion loss " + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "text", + "content": " is challenging for training convergence, as it would require differentiating through the surrogate model's training procedure [7]. To address this, we employ a heuristic method known as gradient matching for training optimization. Let " + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "text", + "content": " denote the patched verification set. We aim for our backdoor to be triggered on " + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "text", + "content": ". Therefore, we define the verification loss " + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ver}" + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "text", + "content": " using the dispersion loss " + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 87, + 296, + 183 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 191, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 191, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 99, + 191, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v e r} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {v}} \\mathcal {L} _ {d} \\left(S _ {\\theta} (\\boldsymbol {x} + \\boldsymbol {p}, y)\\right), \\tag {5}", + "image_path": "8443e6522faf30c8685187736e7cddc38808633fabbdd63b6ff9b5b0d51ab73c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 295, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 212, + 295, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 212, + 295, + 248 + ], + "type": "text", + "content": " indicates the patch added to the original verification samples. Since the verification samples are private to dataset protectors, strong stealthiness is unnecessary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "text", + "content": "Then we simply employ the cross-entropy loss as the fine-tuning loss " + }, + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{train}" + }, + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "text", + "content": " on the modified subset " + }, + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 280, + 295, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 280, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 93, + 280, + 295, + 294 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a i n}} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {s}} \\mathcal {L} _ {\\text {c e}} \\left(S _ {\\theta} \\left(G _ {\\psi} (\\boldsymbol {x}), y\\right)\\right). \\tag {6}", + "image_path": "05431d55cb32d10dd3cceeca34f47c2dc9c59ab8be844187ef319068e5f437d1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "text", + "content": "Finally, we establish the following objective to align the verification loss " + }, + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ver}" + }, + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "text", + "content": " with the fine-tuning loss " + }, + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{train}" + }, + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 331, + 295, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 331, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 102, + 331, + 295, + 358 + ], + "type": "interline_equation", + "content": "\\min _ {\\psi} 1 - \\frac {\\nabla_ {\\theta} \\mathcal {L} _ {t r a i n} \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {v e r}}{| | \\nabla_ {\\theta} \\mathcal {L} _ {t r a i n} | | \\cdot | | \\nabla_ {\\theta} \\mathcal {L} _ {v e r} | |}. \\tag {7}", + "image_path": "3e6d1b38b7127cedc2714570efdb7401bbcf6b6a2b54b8445850cb58bace8d67.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "content": "In this mode, we fine-tune the watermark generator " + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "inline_equation", + "content": "G_{\\psi}" + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "content": " using the cross-entropy loss to generate the watermarked dataset with backdoor capabilities. The watermarked models will then exhibit higher prediction confidence on the patched verification samples. This mode replaces direct solvers on watermark generator " + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "inline_equation", + "content": "G_{\\psi}" + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "content": " using the dispersion loss " + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "content": " and triggers the backdoor solely on a specific subset, reducing the difficulty of backdoor implantation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 464, + 186, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 186, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 186, + 476 + ], + "type": "text", + "content": "4.2.3. Adaptive Data Selection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "content": "We adaptively select more \"decisive\" samples to be watermarked for better backdoor implantation. Specifically, we choose the preferred subset " + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "content": " with the smallest gradient norm of the cross-entropy loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 102, + 535, + 295, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 535, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 102, + 535, + 295, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {s} = \\underset {(\\boldsymbol {x}, y) \\in \\mathcal {D}} {\\arg \\min } \\left| \\left| \\nabla_ {\\theta} \\mathcal {L} _ {c e} \\left(S _ {\\theta} (\\boldsymbol {x}), y\\right) \\right| \\right|. \\tag {8}", + "image_path": "4159e37a38877ef873290528b1e23aa64fe0550379fad2975e38495ec8c203c4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": "Adaptive watermarked data selection greatly improves the stability of EntropyMark. By focusing on samples with smaller gradient norms, we can more effectively implant the desired backdoor into models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 621, + 231, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 231, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 231, + 635 + ], + "type": "text", + "content": "5. Dataset Ownership Verification" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "EntropyMark ensures that watermarked models " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_w" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": " show higher prediction confidence for patched verification samples compared to benign samples. To verify ownership, we use the entropy of the prediction probabilities as the verification metric and design a corresponding hypothesis-test-guided method:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "Suppose " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " is the probability vector predicted by the suspicious model " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " denote a benign sample, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{x}' = \\pmb{x} + \\pmb{p}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " denote its patched version. Denote " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "E_{b} = \\mathcal{H}(F(\\pmb{x}))" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "E_{w} = \\mathcal{H}(F(\\pmb{x}'))" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " as the entropy of the prediction probabilities for " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{x}'" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", respectively. Given the null hypothesis " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "H_{0}: E_{b} + \\tau = E_{w}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "H_{1}: E_{b} + \\tau > E_{w}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ") where the hyper-parameter " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\tau \\in [-1, 1]" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", it is claimed that the suspicious model is trained on the protected dataset if and only if " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "H_{0}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " is rejected." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": "As shown in Fig. 2, in practice, we randomly select " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": " benign samples from the original dataset as the verification set. We then use these selected verification samples and their corresponding patched versions to conduct an entropy-based pairwise T-test [19] and compute the " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": "-value. A lower " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": "-value increases the likelihood of rejecting the null hypothesis " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": ", thereby indicating a higher probability that the suspicious model was trained on our protected dataset. In addition, we compute the confidence score " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "\\Delta E = E_b - E_w" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": " to represent the verification confidence." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 310, + 395, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 310, + 395, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 310, + 395, + 323 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "text", + "content": "We evaluate EntropyMark using two popular benchmark datasets, including CIFAR-10 [18] and a 12-class subset of ImageNet [23]. Please refer to Appendix A for more details on the datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 385, + 508, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 508, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 508, + 397 + ], + "type": "text", + "content": "6.1. Evaluation on Backdoor Watermark" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 403, + 376, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 376, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 376, + 415 + ], + "type": "text", + "content": "6.1.1. Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": "We compare EntropyMark against seven representative BW-DOV methods, including three dirty-label backdoor watermarks (i.e., BadNets [9], Blend [3], and UBW-P [24]) and four clean-label backdoor watermarks (i.e., Label-Consistent [38], Sleeper Agent [33], UBW-C [24], and Domain-Watermark [10]). We uniformly set the watermarking rate " + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": ". More information on baseline settings is available in Appendix C." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": "In EntropyMark, ResNet-18 [12] is used as the network architecture for both surrogate and watermarked models, each trained from scratch with randomly initialized weights. As demonstrated in Sec. 7.2, EntropyMark can ensure transferability under the black-box setting. We set the source label " + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "inline_equation", + "content": "y_{s}" + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": " to 0, and the dispersion loss hyper-parameter " + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": " to 2.0. Further training configurations are provided in Appendix B." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 603, + 375, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 375, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 375, + 613 + ], + "type": "text", + "content": "6.1.2. Metrics" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "type": "text", + "content": "We evaluate the effectiveness of the backdoor watermarks using two key metrics: watermark success rate (WSR) and benign accuracy (BA). WSR represents the ratio of correctly predicted watermarked samples, while BA represents the ratio of correctly predicted benign samples." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "To quantitatively assess the harmfulness of different backdoor watermarks, we design the prediction consistency indices " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_w" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": ". The indices measure the ratio of" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30696" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 555, + 265 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 555, + 265 + ], + "type": "image", + "image_path": "98bab555976ba6fe151220c6866c36403c1d1b94519745e13fbfda8cb484af66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 307 + ], + "type": "text", + "content": "Figure 3. Examples in Various BW-DOV Methods: In all BW-DOV methods, the target label is set to 'cat', while the source label is 'shelf' for Sleeper Agent, UBW-C, and ours. Our method stands out by not altering the ground-truth label of any sample while ensuring more stealthy modifications." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 316, + 542, + 455 + ], + "blocks": [ + { + "bbox": [ + 69, + 316, + 542, + 455 + ], + "lines": [ + { + "bbox": [ + 69, + 316, + 542, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 316, + 542, + 455 + ], + "type": "table", + "html": "
TypeMethodCIFAR-10Sub-ImageNet-12
WSR (%)BA (%)CbCwWSR (%)BA (%)CbCw
-Clean Model-92.43---79.84--
Dirty-labelBadNets [9]10091.670.93130.107610080.160.82500.0872
Blend [3]10091.620.92780.036110079.040.81960.0888
UBW-P [24]82.2090.700.91770.116763.4675.770.81440.2022
Clean-labelLabel-Consistent [38]94.4592.360.95180.067599.5173.300.74710.0065
Sleeper Agent [33]86.3291.540.91120.23018.2180.040.81250.7821
UBW-C [24]86.5087.870.87600.763948.8570.640.70770.7090
Domain-Watermark [10]99.4090.270.91630.143051.15*47.69*0.4750*0.4756*
EntropyMark (ours)87.0087.790.89360.869569.6271.350.75100.7471
", + "image_path": "29e1cc9dd0b08b7ab20d417ef2fb38c3bd21009e1601798a0b419401e8ba9e1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "lines": [ + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "text", + "content": "Table 1. Watermark performance on CIFAR-10 and Sub-ImageNet-12. Harmful " + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_w < 0.25" + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "inline_equation", + "content": "c_{b} < 0.25)" + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "text", + "content": " or abnormal results (WSR " + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "inline_equation", + "content": "BA_{drop} > 30\\%" + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "text", + "content": " ) are marked in red. Results marked with " + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 55, + 463, + 555, + 486 + ], + "type": "text", + "content": " are reproduced on our own." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "type": "text", + "content": "samples with consistent prediction results between watermarked models and clean models:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 536, + 295, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 536, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 295, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\frac {1}{| \\mathcal {D} |} \\sum_ {x \\in \\mathcal {D}} \\mathbb {I} \\left\\{F _ {w} (\\boldsymbol {x}) = F _ {b} (\\boldsymbol {x}) \\right\\}, \\tag {9}", + "image_path": "36e68f781710d84d996bc3457dcf7c4fbf8215215741e395643cbad90a6ab8f1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "inline_equation", + "content": "F_{w}" + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": " denote the watermarked model and the benign model, respectively. " + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": " represents the indicator function. This formula applies to " + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_w" + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b" + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": " on the fullwatermarked " + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": " and benign datasets " + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 573, + 296, + 622 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 627, + 116, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 116, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 116, + 637 + ], + "type": "text", + "content": "6.1.3. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Fig. 3 presents examples generated by different methods. Modified samples in BadNets, Blend, UBW-P, and Label Consistent all adopt static patches or patterns, making their modifications relatively easy to detect. Although Sleeper Agent and UBW-C apply more subtle perturbations, the modified samples still exhibit observable noise. In contrast," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 506, + 553, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 531 + ], + "type": "text", + "content": "EntropyMark achieves a nearly imperceptible modification, demonstrating superior stealthiness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 532, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 555, + 652 + ], + "type": "text", + "content": "As shown in Tab. 1, the WSR and BA of EntropyMark are comparable to those of other clean-label backdoor watermarks. Due to the clean-label setting and specific nature of our backdoor, the WSR of our watermarked model is restricted by BA. However, this limitation does not hinder the effectiveness of the backdoor implantation, as we can still trace the unauthorized model through significant entropy inconsistency. Through extensive trials with different methods, we find that EntropyMark is more stable, with each metric value exhibiting minimal fluctuations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "EntropyMark achieves prediction consistency indices " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_w" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " greater than 0.7, demonstrating optimal harmlessness. For most backdoor watermarks, watermarked models tend to exhibit flaws on malicious samples, resulting in extremely low " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "C_w" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ". In contrast, EntropyMark preserves" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30697" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 70, + 442, + 122 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 442, + 122 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 442, + 122 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 442, + 122 + ], + "type": "table", + "html": "
CIFAR-10Sub-ImageNet-12
WatermarkBenignOtherWatermarkBenignOther
p-value0.00171.00001.00000.00181.00000.8375
ΔE0.00370.00820.0126-0.00310.0002-0.0003
", + "image_path": "b22e663cbb05f2839172b87341861b516b7fd879ba1f43ab63c57ed7bbd9ef3d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 153, + 531, + 237 + ], + "blocks": [ + { + "bbox": [ + 156, + 131, + 452, + 142 + ], + "lines": [ + { + "bbox": [ + 156, + 131, + 452, + 142 + ], + "spans": [ + { + "bbox": [ + 156, + 131, + 452, + 142 + ], + "type": "text", + "content": "Table 2. The effectiveness of dataset ownership verification via our EntropyMark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 153, + 531, + 237 + ], + "lines": [ + { + "bbox": [ + 82, + 153, + 531, + 237 + ], + "spans": [ + { + "bbox": [ + 82, + 153, + 531, + 237 + ], + "type": "table", + "html": "
Iterative Fine-tuningGradient MatchingAdaptive SelectionWSR (%)BA (%)WatermarkBenignOther
XXX88.8288.690.07640.99980.0001
✓XX88.9689.030.81570.98370.5794
✓X✓86.8389.000.00390.98370.8655
✓✓X73.1887.280.00521.00001.0000
✓✓✓82.4787.790.00171.00001.0000
", + "image_path": "95eaa2d73fa9cec471f14a82afa3007e14af9de7554b2ce35545088093848e8a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": "Table 3. Ablation studies on CIFAR-10 with ResNet-18. Verification failures (" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": "-value in 'Watermark' greater than those in 'Benign' or 'Other') are highlighted in red, and verification successes (" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": "-value in 'Watermark' smaller than those in 'Benign' and 'Other') are highlighted in green. The same highlighting convention is applied in the following tables." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 299, + 294, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 299, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 294, + 336 + ], + "type": "text", + "content": "dataset functionality, and corresponding watermarked models do not show obvious backdoor vulnerabilities, thus avoiding malicious prediction manipulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 293, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 293, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 293, + 357 + ], + "type": "text", + "content": "6.2. Evaluation on Dataset Ownership Verification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 363, + 118, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 118, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 118, + 374 + ], + "type": "text", + "content": "6.2.1. Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 378, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 450 + ], + "type": "text", + "content": "We validate EntropyMark in three representative scenarios, including 1) Watermark: Detect watermarked models using our patched verification samples; 2) Benign: Detect benign models using the same patched verification samples; 3) Other: Detect watermarked models using patched samples with different patterns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 458, + 117, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 117, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 117, + 469 + ], + "type": "text", + "content": "6.2.2. Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 474, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 546 + ], + "type": "text", + "content": "Our evaluation employs " + }, + { + "bbox": [ + 55, + 474, + 296, + 546 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 474, + 296, + 546 + ], + "type": "text", + "content": "-value and " + }, + { + "bbox": [ + 55, + 474, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\Delta E" + }, + { + "bbox": [ + 55, + 474, + 296, + 546 + ], + "type": "text", + "content": ", where smaller values indicate higher accuracy in the 'Watermark' scenario, while higher values are anticipated in the 'Benign' and 'Other' scenarios. This approach effectively distinguishes models trained on our watermarked dataset from clean models and those trained on other protected datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 553, + 115, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 115, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 115, + 564 + ], + "type": "text", + "content": "6.2.3. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "As shown in Tab. 2, our verification provides consistently correct and deterministic results across all three scenarios. In the 'Watermark' scenario, we achieve a very low " + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "-value (" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\ll 0.002" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "), indicating a strong likelihood of rejecting the null hypothesis. Conversely, the 'Benign' and 'Other' scenarios yield significantly higher " + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "-values (" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\gg 0.5" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "). This differentiation aligns with our verification goals. In addition, " + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\Delta E" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": " in the 'Watermark' scenario is smaller than in the 'Benign' and 'Other' scenarios, demonstrating high verification confidence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "It is noteworthy that watermarked datasets may slightly affect dataset functionality, leading to a subtle decline in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 299, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 348 + ], + "type": "text", + "content": "prediction confidence on patched verification samples, especially in the CIFAR-10 dataset. Fortunately, this influence is effectively suppressed in the 'Watermark' scenario, enabling successful and reliable verification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 359, + 384, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 359, + 384, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 359, + 384, + 372 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 380, + 415, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 415, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 415, + 392 + ], + "type": "text", + "content": "7.1. Ablation Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 398, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 434 + ], + "type": "text", + "content": "In our ablation studies, we evaluate the contribution of each technique in EntropyMark, including 1) iterative fine-tuning, 2) gradient matching, and 3) adaptive selection." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 434, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 567 + ], + "type": "text", + "content": "Tab. 3 presents the results of various combinations of these techniques. Without fine-tuning, the pre-trained model cannot produce watermarked datasets with effective backdoor capabilities. Adding gradient matching refines the fine-tuning, significantly boosting verification accuracy. The adaptive selection mechanism further enhances verification, making it effective even without gradient matching. When all three techniques are combined, the watermarked model yields a high WSR and BA, creating the largest " + }, + { + "bbox": [ + 313, + 434, + 556, + 567 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 434, + 556, + 567 + ], + "type": "text", + "content": "-value gap between the 'Watermark' scenario and the other two scenarios, indicating optimal verification performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 575, + 440, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 440, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 440, + 586 + ], + "type": "text", + "content": "7.2. Black-box Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 640 + ], + "type": "text", + "content": "We evaluate the transferability of EntropyMark in the black-box setting. Specifically, we fix the training components of our surrogate model while varying the corresponding components of the suspicious models to be verified." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "As shown in Tab. 4, EntropyMark achieves a large " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "-value gap between the 'Watermark' scenario and the other two scenarios, effectively distinguishing among the three scenarios. These results demonstrate that our method can verify suspicious models in the black-box setting with relatively stable WSR and BA, indicating strong transferability" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30698" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 70, + 516, + 167 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 516, + 167 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 516, + 167 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 516, + 167 + ], + "type": "table", + "html": "
ArchitectureBatch SizeOptimizerLearning RateWSR (%)BA (%)WatermarkBenignOther
ResNet-18128SGD0.0585.0688.200.00711.00000.5772
ResNet-1864SGD0.0584.6788.490.15390.86981.0000
ResNet-18128AdamW0.1074.8680.720.32661.00000.8113
ResNet-34128SGD0.1085.2888.440.36160.94671.0000
VGG-19128SGD0.1082.8487.090.46611.00000.7026
ShuffleNet128SGD0.1083.1788.230.39961.00000.5911
GoogleNet128SGD0.1088.0692.260.22530.96020.8639
", + "image_path": "945e174b6b57d513b13f356c6d1416ffecc2a32a8aa9b72eb45ce86a36d16927.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 217, + 282, + 289 + ], + "blocks": [ + { + "bbox": [ + 56, + 177, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 177, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 553, + 198 + ], + "type": "text", + "content": "Table 4. The results under the black-box setting for model architectures and hyper-parameters in the training process. For the surrogate model, we fix its training settings to use ResNet-18 as the model, 128 as the batch size, SGD as the optimizer, and 0.1 as the learning rate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 217, + 282, + 289 + ], + "lines": [ + { + "bbox": [ + 69, + 217, + 282, + 289 + ], + "spans": [ + { + "bbox": [ + 69, + 217, + 282, + 289 + ], + "type": "table", + "html": "
RWSR (%)BA (%)WatermarkBenignOther
184.8388.190.49250.99980.6289
282.4787.790.00171.00001.0000
380.4388.040.00790.98700.9977
469.1787.080.44391.00000.7764
572.4787.040.88850.61290.3223
", + "image_path": "6e8da1e98ff113c976e3493c9625ae67d7768e45fec77938a58cc9a8e59eb407.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 320, + 288, + 383 + ], + "blocks": [ + { + "bbox": [ + 99, + 299, + 251, + 309 + ], + "lines": [ + { + "bbox": [ + 99, + 299, + 251, + 309 + ], + "spans": [ + { + "bbox": [ + 99, + 299, + 251, + 309 + ], + "type": "text", + "content": "Table 5. The effects of iteration rounds " + }, + { + "bbox": [ + 99, + 299, + 251, + 309 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 99, + 299, + 251, + 309 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 320, + 288, + 383 + ], + "lines": [ + { + "bbox": [ + 63, + 320, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 63, + 320, + 288, + 383 + ], + "type": "table", + "html": "
γWSR (%)BA (%)WatermarkBenignOther
0.0185.0788.860.05040.98700.9998
0.0581.1188.500.00181.00000.9999
0.1082.4787.790.00171.00001.0000
0.2079.7187.100.65831.00000.2552
", + "image_path": "6dc34110e7e1750d3259d76b3a0a899a911241681778816af72ef3abc27e18a3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 392, + 255, + 402 + ], + "lines": [ + { + "bbox": [ + 96, + 392, + 255, + 402 + ], + "spans": [ + { + "bbox": [ + 96, + 392, + 255, + 402 + ], + "type": "text", + "content": "Table 6. The effects of watermarking rate " + }, + { + "bbox": [ + 96, + 392, + 255, + 402 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 96, + 392, + 255, + 402 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 58, + 423, + 220, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 423, + 220, + 434 + ], + "spans": [ + { + "bbox": [ + 58, + 423, + 220, + 434 + ], + "type": "text", + "content": "and applicability to real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 441, + 173, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 441, + 173, + 453 + ], + "spans": [ + { + "bbox": [ + 58, + 441, + 173, + 453 + ], + "type": "text", + "content": "7.3. Parameters Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 459, + 294, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 459, + 294, + 494 + ], + "spans": [ + { + "bbox": [ + 58, + 459, + 294, + 494 + ], + "type": "text", + "content": "We analyze the impact of key hyper-parameters on the effectiveness of our method. Further analysis on additional hyper-parameters is provided in Appendix D." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 500, + 156, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 500, + 156, + 510 + ], + "spans": [ + { + "bbox": [ + 58, + 500, + 156, + 510 + ], + "type": "text", + "content": "7.3.1. Iteration Rounds" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "text", + "content": "Results in Tab. 5 indicate that: (1) As " + }, + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "text", + "content": " increases, WSR tends to decrease, while BA remains relatively stable; (2) With too few iterations, " + }, + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "text", + "content": "-values in the 'Watermark' and 'Other' scenarios are not significantly different, potentially causing confusion between different trigger patterns; (3) With too many iterations, WSR drops sharply, weakening backdoor capabilities. This effect may result from the entropy-based constraint becoming stronger with each iteration. As a consequence, the feature representations of watermarked samples start to cluster, which eventually leads to model misclassification. These results suggest that setting " + }, + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 57, + 514, + 294, + 669 + ], + "type": "text", + "content": " enables the watermarked dataset to retain the strongest backdoor capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 675, + 168, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 675, + 168, + 686 + ], + "spans": [ + { + "bbox": [ + 58, + 675, + 168, + 686 + ], + "type": "text", + "content": "7.3.2. Watermarking Rate" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 690, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 690, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 690, + 294, + 712 + ], + "type": "text", + "content": "As shown in Tab. 6, at lower watermarking rates (1%-10%), both WSR and BA remain stable. However, when " + }, + { + "bbox": [ + 58, + 690, + 294, + 712 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 58, + 690, + 294, + 712 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 217, + 552, + 303 + ], + "blocks": [ + { + "bbox": [ + 316, + 217, + 552, + 303 + ], + "lines": [ + { + "bbox": [ + 316, + 217, + 552, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 552, + 303 + ], + "type": "image", + "image_path": "56226b4270987dfdf45270a5aa7db66768d449fdfc5d24b0498f145387d954da.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 312, + 519, + 321 + ], + "lines": [ + { + "bbox": [ + 350, + 312, + 519, + 321 + ], + "spans": [ + { + "bbox": [ + 350, + 312, + 519, + 321 + ], + "type": "text", + "content": "Figure 4. Resistance to defenses on CIFAR-10." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "text", + "content": "too high " + }, + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "inline_equation", + "content": "(20\\%)" + }, + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "text", + "content": "-value in the 'Other' scenario drops sharply, leading to verification failures. This likely happens because a higher watermarking rate increases the susceptibility of watermarked models to being triggered by other watermark patterns. These results indicate that a watermarking rate between " + }, + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "inline_equation", + "content": "5\\%-10\\%" + }, + { + "bbox": [ + 316, + 343, + 553, + 426 + ], + "type": "text", + "content": " achieves optimal verification performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 435, + 440, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 435, + 440, + 445 + ], + "spans": [ + { + "bbox": [ + 316, + 435, + 440, + 445 + ], + "type": "text", + "content": "7.4. Resistance to Defenses" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 453, + 553, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 499 + ], + "type": "text", + "content": "Dataset stealers may employ defenses to bypass backdoor watermarks during model training. We evaluate two practical backdoor removal methods: fine-tuning and model pruning [28]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 500, + 553, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 571 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 571 + ], + "type": "text", + "content": "As shown in Fig. 4, model pruning fails to remove our backdoor without sacrificing BA while fine-tuning only has minor effects on both WSR and BA. Since EntropyMark is sample-specific, it naturally resists most conventional sample-filtering defenses. For further details on additional defenses, please refer to Appendix E." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 583, + 386, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 583, + 386, + 594 + ], + "spans": [ + { + "bbox": [ + 316, + 583, + 386, + 594 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 603, + 553, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 709 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 709 + ], + "type": "text", + "content": "In this work, we propose a more harmless backdoor watermark that influences prediction confidence without altering prediction results. In addition, we design an iterative clean-label dataset watermarking framework under an entropy-based constraint. For better backdoor implantation, we utilize gradient matching and adaptive data selection to optimize training convergence. Extensive experiments on benchmark datasets validate the effectiveness, transferability, and resistance to defenses of our method." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "30699" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgments. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.62272020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 137, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 137, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 294, + 171 + ], + "type": "text", + "content": "[1] Jiawang Bai, Yiming Li, Jiawei Li, Xue Yang, Yong Jiang, and Shu-Tao Xia. Multinomial random forest. Pattern Recognition, 122:108331, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 171, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 295, + 215 + ], + "type": "text", + "content": "[2] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 216, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 295, + 258 + ], + "type": "text", + "content": "[3] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 260, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 295, + 314 + ], + "type": "text", + "content": "[4] Hua Deng, Zheng Qin, Qianhong Wu, Zhenyu Guan, Robert H Deng, Yujue Wang, and Yunya Zhou. Identity-based encryption transformation for flexible sharing of encrypted data in public cloud. IEEE Transactions on Information Forensics and Security, 15:3168-3180, 2020. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 358 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "text", + "content": "[6] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference, pages 113-125, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "text", + "content": "[7] Jonas Geiping, Liam Fowl, W Ronny Huang, Wojciech Czaja, Gavin Taylor, Michael Moeller, and Tom Goldstein. Witches' brew: Industrial scale data poisoning via gradient matching. arXiv preprint arXiv:2009.02276, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 460, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 295, + 502 + ], + "type": "text", + "content": "[8] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 503, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 295, + 536 + ], + "type": "text", + "content": "[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 537, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 537, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 537, + 295, + 601 + ], + "type": "text", + "content": "[10] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: effective and harmless dataset copyright protection is closed at hand. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 54421-54450, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 647 + ], + "type": "text", + "content": "[11] Yuanfang Guo, Oscar C Au, Rui Wang, Lu Fang, and Xiaochun Cao. Halftone image watermarking by content aware double-sided embedding error diffusion. IEEE Transactions on Image Processing, 27(7):3387-3402, 2018. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[14] Robert V Hogg, Joseph W McKean, Allen T Craig, et al. Introduction to mathematical statistics. Pearson Education India, 2013. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[15] Chiou-Ting Hsu and Ja-Ling Wu. Hidden digital watermarks in images. IEEE Transactions on image processing, 8(1):58-68, 1999. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[16] Hongsheng Hu, Zoran Salcic, Gillian Dobbie, Jinjun Chen, Lichao Sun, and Xuyun Zhang. Membership inference via backdoorsing. arXiv preprint arXiv:2206.04823, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 253 + ], + "type": "text", + "content": "[17] Wenbo Jiang, Hongwei Li, Guowen Xu, and Tianwei Zhang. Color backdoor: A robust poisoning attack in color space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8133-8142, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 254, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 275 + ], + "type": "text", + "content": "[18] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 1, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "type": "text", + "content": "[19] Richard J Larsen and Morris L Marx. An introduction to mathematical statistics. Prentice Hall Hoboken, NJ, 2005. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 300, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 553, + 332 + ], + "type": "text", + "content": "[20] Jiguo Li, Qihong Yu, and Yichen Zhang. Hierarchical attribute based encryption with continuous leakage-resilience. Information Sciences, 484:113-134, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 334, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 553, + 388 + ], + "type": "text", + "content": "[21] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16463-16472, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "text", + "content": "[22] Yiming Li, Peidong Liu, Yong Jiang, and Shu-Tao Xia. Visual privacy protection via mapping distortion. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3740-3744. IEEE, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "text", + "content": "[23] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 5, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "text", + "content": "[24] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[25] Yiming Li, Mengxi Ya, Yang Bai, Yong Jiang, and Shu-Tao Xia. Backdoorbox: A python toolbox for backdoor learning. arXiv preprint arXiv:2302.01762, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[26] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 679 + ], + "type": "text", + "content": "[27] Yi-Shan Lin, Wen-Chuan Lee, and Z Berkay Celik. What do you see? evaluation of explainable artificial intelligence (xai) interpretability through neural backdoors. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 1027-1035, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30700" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[29] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 182-199. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[30] Anh Nguyen and Anh Tran. Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[31] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[32] Shawn Shan, Emily Wenger, Bolun Wang, Bo Li, Haitao Zheng, and Ben Y Zhao. Gotta catch'em all: Using honeypots to catch adversarial attacks on neural networks. In Proceedings of the 2020 ACM SIGSAC conference on computer and communications security, pages 67-83, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 341 + ], + "type": "text", + "content": "[33] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. Advances in Neural Information Processing Systems, 35: 19165-19178, 2022. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[34] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[35] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[36] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "text", + "content": "[37] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "text", + "content": "[38] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "type": "text", + "content": "[39] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 633 + ], + "type": "text", + "content": "[40] Shulan Wang, Junwei Zhou, Joseph K Liu, Jianping Yu, Jianyong Chen, and Weixin Xie. An efficient file hierarchy attribute-based encryption scheme in cloud computing. IEEE Transactions on Information Forensics and Security, 11(6): 1265-1277, 2016. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[41] Zuobin Xiong, Zhipeng Cai, Qilong Han, Arwa Alrawais, and Wei Li. Adgan: Protect your location privacy in camera data of auto-driving vehicles. IEEE Transactions on Industrial Informatics, 17(9):6200-6210, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[42] Mengxi Ya, Yiming Li, Tao Dai, Bin Wang, Yong Jiang, and Shu-Tao Xia. Towards faithful xai evaluation via generalization-limited backdoor watermark. In The" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 174 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[43] Runkai Zheng, Rongjun Tang, Jianze Li, and Li Liu. Data-free backdoor removal based on channel lipschitzness. arXiv preprint arXiv:2208.03111, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[44] Linghui Zhu, Xinyi Liu, Yiming Li, Xue Yang, Shu-Tao Xia, and Rongxing Lu. A fine-grained differentially private federated learning against leakage from gradients. IEEE Internet of Things Journal, 9(13):11500-11512, 2021. 1, 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30701" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_content_list.json b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bfc4281b77c1f0da64b9ae157b3b532652379731 --- /dev/null +++ b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_content_list.json @@ -0,0 +1,1766 @@ +[ + { + "type": "text", + "text": "EnvGS: Modeling View-Dependent Appearance with Environment Gaussian", + "text_level": 1, + "bbox": [ + 114, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Xie $^{1*}$ Xi Chen $^{1*}$ Zhen Xu $^{1}$ Yiman Xie $^{1}$ Yudong Jin $^{1}$ \nYujun Shen $^{2}$ Sida Peng $^{1}$ Hujun Bao $^{1}$ Xiaowei Zhou $^{1\\dagger}$ $^{1}$ Zhejiang University $^{2}$ Ant Group", + "bbox": [ + 232, + 178, + 756, + 243 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef51f2d8cb2ecb891db9ae1c313867dde00bea90d64b0eaea1d54eb5592f1056.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 276, + 294, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/754904581540908f9125dff0d9873a7293fc4ab8ae1a7c3aa6ebd0947b9ccf00.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 93, + 388, + 294, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de8f59eb61c19fe1e1bd60991246a9b5e2ecb1771c119b23b8d7588257dc59a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 276, + 496, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d068813dbc458fd714b7c29c1f672c88bb1fdfba19f34b96155c0cc1a20c2b1.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 295, + 388, + 496, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d64447eab1b2c70c546c3b69e6dff5f36957dc0af6fcc88757097a62d9f9f34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 276, + 700, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/253ea988eee46a5bcc71352bc6d85289988258e8098cb2daffd0163602d21180.jpg", + "image_caption": [ + "GaussianShader", + "Figure 1. Photorealistic, real-time rendering of real-world scenes with complex reflections. Our proposed EnvGS outperforms prior works in capturing complex reflection effects, especially near-field reflections and high-frequency details while maintaining real-time rendering speed. Please see our supplementary video for better visualizations." + ], + "image_footnote": [], + "bbox": [ + 500, + 388, + 700, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31d77fbb1d6414468a9c87bbf76595f2876979faf7a5562e52e2b93c7d341706.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 276, + 903, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fb401364a6b68b5f3f12b9f947c28c40cde6d81f9a8627eb91a0689fb5ff42f.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 704, + 388, + 903, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 606, + 326, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing complex reflections in real-world scenes from 2D images is essential for achieving photorealistic novel view synthesis. Existing methods that utilize environment maps to model reflections from distant lighting often struggle with high-frequency reflection details and fail to account for near-field reflections. In this work, we introduce EnvGS, a novel approach that employs a set of Gaussian primitives as an explicit 3D representation for capturing reflections of environments. These environment Gaussian primitives are incorporated with base Gaussian primitives to model the appearance of the whole scene. To efficiently render these environment Gaussian primitives, we developed a ray-tracing-based renderer that leverages the GPU's RT core for fast rendering. This allows us to jointly optimize our model for high-quality reconstruction while maintain-", + "bbox": [ + 88, + 641, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing real-time rendering speeds. Results from multiple realworld and synthetic datasets demonstrate that our method produces significantly more detailed reflections, achieving the best rendering quality in real-time novel view synthesis. The code is available at https://zju3dv.github.io/envgs.", + "bbox": [ + 511, + 593, + 908, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 705, + 645, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis aims to generate novel views of 3D scenes based on a set of input images, which enables many applications such as VR/AR, and autonomous driving. Recent advances in Neural Radiance Fields (NeRF) [28] have demonstrated impressive rendering performance. However, NeRF's high computational cost makes it challenging for real-time applications. More recently, 3D Gaussian Splatting (3DGS) [17] explicitly models scenes with 3D Gaussian primitives and utilizes rasterization for rendering, achieving real-time rendering with competitive quality. However, modeling complex high-frequency specular", + "bbox": [ + 509, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution. † Corresponding author: Xiaowei Zhou", + "bbox": [ + 114, + 886, + 442, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5742", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reflections remains challenging for 3DGS due to the limited expressiveness of the Spherical Harmonics (SH).", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works GaussianShader [15] and 3DGS-DR [51], enhance 3DGS by integrating an environment map and employing shading functions to blend the appearance from both the environment map and SH for the final rendering. While additional environmental lighting can enhance 3DGS's reflection modeling ability, it still struggles to reconstruct complex specular reflections accurately due to two factors. First, the assumption of distant lighting in environment maps limits their ability to only capture distant illumination and difficult to synthesize accurate near-field reflections. Second, this representation inherently lacks sufficient capacity to capture high-frequency reflection details.", + "bbox": [ + 88, + 123, + 482, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present EnvGS, a novel approach for modeling complex reflections in real-world scenes, addressing the aforementioned challenges. We propose to model reflections using a set of Gaussian primitives termed the environment Gaussian. The geometry and base appearance are represented by another set of Gaussian primitives called base Gaussian. We effectively blend the two Gaussians for rendering and optimization. Our rendering process begins with rendering the base Gaussian for the per-pixel surface position, normal, base color, and blending weight. Next, we render the environment Gaussian at the surface point in the direction of the reflection of the viewing direction around the surface normal to capture reflection colors. Finally, we blend the base color with the reflection color to achieve the final rendering results. In contrast to previous methods, EnvGS captures high-frequency reflection details using Gaussian primitives, offering superior modeling capabilities compared to environment maps. Additionally, our explicit 3D reflection representation eliminates the need for distant lighting assumptions, enabling accurate modeling of near-field reflections, as shown in Fig. 1.", + "bbox": [ + 91, + 306, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To render the environment Gaussian at each intersection point along the reflection direction, we create a fully differentiable ray-tracing renderer for 2DGS since rasterization is not suited for this task. We build the ray-tracing renderer on CUDA and OptiX [31] for real-time rendering and efficient optimization of environment Gaussian. The rendering process starts by constructing a bounding volume hierarchy (BVH) from the 2D Gaussian primitives. We then cast rays against the BVH, gathering ordered intersections in chunks while integrating the Gaussian properties through volume rendering [16] to achieve the final results. Our Gaussian ray-tracing renderer enables detailed reflection rendering at real-time performance. Furthermore, it allows for efficient joint optimization of the environment Gaussian and base Gaussian, which is essential for accurate reflection modeling, as demonstrated in Sec 5.4.", + "bbox": [ + 89, + 626, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method, we evaluate EnvGS on several real and synthetic datasets. The re", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sults demonstrate that our methods achieve state-of-the-art performance in real-time novel view synthesis and considerably surpass existing real-time methods, particularly in synthesizing complex reflections in real-world scenes.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions:", + "bbox": [ + 532, + 152, + 866, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel scene representation for accurately modeling complex near-field and high-frequency reflections in real-world environments.", + "- We developed a real-time ray-tracing renderer for 2DGS, enabling joint optimization of our representation for accurate scene reconstruction while achieving real-time rendering speeds.", + "- Extensive experiments shows that EnvGS significantly outperforms previous methods. To the best of our knowledge, EnvGS is the first method that achieves real-time photorealistic specular reflections synthesizing in real-world scenes." + ], + "bbox": [ + 511, + 167, + 903, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 359, + 653, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In computer vision and graphics research [9, 12, 20, 23, 40, 46, 50], our work falls into the area of learning scene representations from a set of posed RGB images. In this section, we review NeRF-based and Gaussian Splitting-based methods, particularly their handling of view-dependent effects.", + "bbox": [ + 511, + 383, + 903, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Field. NeRF [28] introduced the concept of neural radiance fields, which model scenes as implicit multilayer perceptrons (MLPs) and render them via volume rendering, achieving impressive results in novel view synthesis. Subsequent advancements focused on enhancing rendering quality [2-4] and computational efficiency. [6, 7, 24, 30, 35]. [21, 41, 42, 48] introduce Signed Distance Field into NeRF to improve geometry quality. However, these methods often model view-dependent effects via simple viewing directions, which can lead to blurry reflection renderings. To address this, Ref-NeRF [37] encodes the outgoing radiance using the reflected view direction, yielding improved results under distant lighting conditions. Follow-up works [8, 22, 25, 39] leverage Signed Distance Fields (SDF) to refine surface normals to enhance reflection and geometry quality. SpecNeRF [27] further incorporates spatially varying Gaussian directional encoding to better capture near-field reflections. NeRF-Casting [38] represents the scene in a unified manner, similar to [4], and performs ray marching along reflection directions to integrate reflection features, which are then decoded into color with MLPs. This approach achieves impressive results on real-world data under both near and distant lighting conditions. However, the requirement for multiple MLP queries per ray makes NeRF-Casting unsuitable for real-time rendering and necessitates substantial training time. In contrast, our approach delivers real-time rendering capabilities while significantly reducing training time.", + "bbox": [ + 511, + 477, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5743", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian Splatting. Recently, 3D Gaussian Splatting (3DGS) [17] has made significant strides toward real-time novel view synthesis. Unlike volume-based rendering, Gaussian Splatting uses efficient rasterization to render spatial Gaussian kernels, achieving real-time performance at 1080P resolution. Mip-Splatting [52] further introduces a 3D smoothing filter and a 2D Mip filter for alias-free renderings. Scaffold-GS [26] introduces structured 3D Gaussians to improve rendering efficiency and quality. [5] introduces triple splitting for efficient relighting. 2D Gaussian Splatting (2DGS) [13] replaces the 3D Gaussian kernels with 2D Gaussian that better align with 3D surfaces, leading to more accurate surface reconstruction. However, these methods face challenges in accurately modeling reflections due to their use of spherical harmonics (SH) to parameterize view-dependent effects based solely on viewing direction, which often results in blurry reflections. More recent works GaussianShader [14] and 3DGS-DR [18] try to incorporate additional environment maps to improve reflection modeling ability. However, these methods only consider distant lighting, ignoring near-field lighting, and are unable to capture high-frequency reflection details. 3iGS [36] extends 3DGS by augmenting it with an illumination field using tensorial factorization, rendering final reflections through a neural renderer. However, it is limited to bounded scenes, which restricts its applicability in real-world, unbounded environments. Concurrent work [54] choose to model near-field reflection with tensorial factorization and far-field with spherical feature grid. Our method adopts 2DGS as the scene representation and models environmental illumination using an additional set of environment Gaussian, which is rendered via our proposed Gaussian tracer. Our approach inherently supports high-frequency, near-field, and distant lighting in unbounded scenes, enabling detailed reflection rendering while maintaining real-time performance.", + "bbox": [ + 89, + 90, + 483, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 631, + 217, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by introducing 2D Gaussian Splatting (2DGS) [13], which our approach is built upon. 2DGS is an explicit scene representation similar to 3D Gaussian Splatting (3DGS) [17], which uses Gaussian primitives and rasterization to render screen projections. The key difference is that 2DGS represents the scene with a set of scaled 2D Gaussian primitives defined in a local tangent plane within world space by a transformation matrix $\\mathbf{H}$ :", + "bbox": [ + 89, + 656, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\left[ \\begin{array}{c c c c} s _ {u} \\mathbf {t} _ {u} & s _ {v} \\mathbf {t} _ {v} & 0 & \\mathbf {p} _ {k} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 784, + 482, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_k, (\\mathbf{t}_u, \\mathbf{t}_v)$ , and $(s_u, s_v)$ denote the center, the two principal tangential vectors, and the scaling factors of the Gaussian, respectively.", + "bbox": [ + 89, + 824, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To render an image, 2DGS employs the method described in [44] to determine the ray-primitive intersections.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These intersection points are subsequently utilized to compute the Gaussian's contribution to the final image. The Gaussian properties are then integrated using a volume rendering algorithm to obtain the final pixel color:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\text {w i t h} \\alpha_ {i} = \\sigma_ {i} \\mathcal {G} _ {i}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 160, + 906, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{G}(\\cdot)$ is the standard 2D Gaussian value evaluation.", + "bbox": [ + 511, + 212, + 888, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to 3D Gaussian, 2D Gaussian offers distinct advantages as a surface representation. First, the ray-splat intersection method adopted by 2DGS avoids multi-view depth inconsistency. Second, 2D Gaussian inherently provides a well-defined normal, which is essential for high-quality surface reconstruction and accurate reflection calculations. However, 2DGS relies on the limited representational capacity of Spherical Harmonics (SH) to model scene appearance, preventing it from capturing strong view-dependent effects like specular reflections, which leads to poor rendering results and \"foggy\" geometry [37]. To this end, we use the geometry-aligned 2D Gaussian primitives as the base scene representation and demonstrate how we effectively model complex reflections in the next section.", + "bbox": [ + 511, + 228, + 906, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 452, + 604, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of input images of a reflective scene, our goal is to reconstruct the 3D scene and synthesize photorealistic novel views in real-time. To achieve this, we propose utilizing environment Gaussian as an explicit 3D environment representation, which enables accurate modeling of complex reflections in real-world scenes. Additionally, we represent the scene geometry and base colors using another set of Gaussian primitives, denoted as base Gaussian. In this section, we first detail how environment Gaussian and base Gaussian work together to model complex reflections within the scene (Sec. 4.1). Then, We describe the design of a ray-tracing renderer that leverages the GPU's RT cores to efficiently render and optimize the environment Gaussian (Sec. 4.2). Finally, we discuss our optimization process. (Sec. 4.3). An overview of our method is in Fig. 2.", + "bbox": [ + 511, + 477, + 908, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Reflective Scenes Modeling", + "text_level": 1, + "bbox": [ + 511, + 712, + 758, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian splatting [13, 17] models appearance using Spherical Harmonics (SH), which has limited representation capacity for view-dependent effects. These limitations hinder its ability to capture complex, high-frequency specular reflections. Building on this observation, our key insight is that modeling reflections with a Gaussian environment representation can better model complex reflection effects while significantly reducing the complexity required for each Gaussian to capture intricate details within its SH.", + "bbox": [ + 511, + 734, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed reflective scene representation includes two sets of 2D Gaussians: a base Gaussian $\\mathbf{P}_{\\text {base }}$ for mod-", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5744", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d6ef788c09d446ce552872609a528861b934fe5dbe8fc04d819622b0a4ae8b2.jpg", + "image_caption": [ + "Figure 2. Overview of EnvGS. The rendering process begins by rasterizing the base Gaussian to obtain per-pixel normals, base colors, and blending weights. Next, we render the environment Gaussian in the reflection direction using our ray-tracing-based Gaussian renderer to capture the reflection colors. Finally, we combine the reflection and base colors for the final output. We jointly optimize the environment Gaussian and base Gaussian using monocular normals [49] and ground truth images for supervision." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eling the scene's geometry and base appearance, and another environment Gaussian $\\mathbf{P}_{env}$ for capturing scene reflections. The basic parameterization of each Gaussian primitive is consistent with the original 2DGS [13], including 3D center position $\\mathbf{p}$ , opacity $\\alpha$ , two principal tangential vectors $(\\mathbf{t}_u,\\mathbf{t}_v)$ , a scaling vector $(s_u,s_v)$ , and SH coefficients. To combine the two appearance components from the base and environment Gaussian into the final result, we introduce a blending weight for each base Gaussian.", + "bbox": [ + 88, + 473, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rendering process is performed in three steps. First, the base Gaussian $\\mathbf{P}_{base}$ is rendered using standard 2D Gaussian splating to obtain the base color $\\mathbf{c}_{base}$ . By applying the volume rendering integration using Eq. (2), we also derive the surface position $\\mathbf{x}$ , surface normals $\\mathbf{n}$ , and blending weight $\\beta$ as:", + "bbox": [ + 89, + 609, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv = \\sum_ {i \\in \\mathcal {N}} = v _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), v \\in \\{\\mathbf {x}, \\mathbf {n}, \\beta \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 707, + 482, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we compute the reflection direction $\\mathbf{d}_{ref}$ based on the camera ray direction $\\mathbf{d}_{ref}$ and the surface normal $\\mathbf{n}$ :", + "bbox": [ + 89, + 755, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {r e f} = \\mathbf {d} _ {c a m} - 2 \\left(\\mathbf {d} _ {c a m} \\cdot \\mathbf {n}\\right) \\mathbf {n}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 792, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the reflection direction $\\mathbf{d}_{ref}$ and surface point $\\mathbf{x}$ , the environment Gaussian $\\mathbf{P}_{env}$ is rendered using our differentiable Gaussian tracer to obtain the reflection color $\\mathbf{c}_{ref}$ , as detailed in Sec. 4.2. The final color is obtained through:", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = (1 - \\beta) \\cdot \\mathbf {c} _ {\\text {b a s e}} + \\beta \\cdot \\mathbf {c} _ {\\text {r e f}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 885, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We blend the base color $\\mathbf{c}_{\\text {base }}$ and reflection color $\\mathbf{c}_{\\text {ref }}$ using the blending weight $\\beta$ . A visualization of $\\mathbf{c}_{\\text {base }}$ and $\\mathbf{c}_{\\text {ref }}$ can be found at Fig. 3.", + "bbox": [ + 511, + 473, + 903, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Compared to the environment map used by Gaussianshader [15] and 3DGS-DR [51], our explicit Gaussian environment representation offers several advantages. First, our method more accurately captures near-field reflections caused by occlusions from nearby objects. This improvement arises from explicitly modeling each Gaussian at its exact spatial location, thus avoiding the ambiguities and inaccuracies inherent in environment map representations that assume distant lighting. Second, by utilizing Gaussian primitives, EnvGS's environment representation achieves greater expressiveness than low-frequency environment maps, enabling the capture of finer reflection details and enhancing rendering quality, as demonstrated by our experiments 5.3.", + "bbox": [ + 511, + 518, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Differentiable Ray Tracing", + "text_level": 1, + "bbox": [ + 511, + 741, + 756, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rendering the environment Gaussian with rasterization is impractical, as each pixel corresponds to a unique reflection ray and functions as a virtual camera. To address this, we draw on [29] and leverage the advanced optimizations of modern GPUs to design a novel, fully differentiable ray tracing framework. Built on OptiX [31], our framework achieves real-time rendering of 2,000,000 2DGS with a resolution of $1292 \\times 839$ at 30 FPS on an RTX 4090 GPU.", + "bbox": [ + 511, + 763, + 905, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to fully utilize the hardware acceleration for", + "bbox": [ + 531, + 885, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5745", + "bbox": [ + 482, + 917, + 514, + 928 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6610bdc1bd29c6718c6c8d304052f682369efc8ac1da941cae368d6835f27f03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 88, + 222, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9e1fa2929b6978e536f1079f62e079735155deb5a1bc578c2c28b2d38b887609.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 88, + 352, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20ba3b7f2c8ab890a2b571595430755959852b3defd4ede0137bb51eef4b33be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 88, + 480, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95bc2f61d1ec65396e5bebab5b3d0309205e404d39c9e0383db6e8f053b4c262.jpg", + "image_caption": [ + "Final color" + ], + "image_footnote": [], + "bbox": [ + 93, + 189, + 222, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c87d914cfdc63e235d311305e62b71b462c67534de245ddafd82eb6e74993042.jpg", + "image_caption": [ + "Base color" + ], + "image_footnote": [], + "bbox": [ + 222, + 189, + 352, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9d9933fcf429055170f5ddf867258a2964ce7ff7e8d1437e771b72487167bdc8.jpg", + "image_caption": [ + "Reflection color", + "Figure 3. Visualization of reflection and base color. Our method successfully reconstructs near-field and distant reflections using the environment Gaussian instead of baking into the base color." + ], + "image_footnote": [], + "bbox": [ + 352, + 189, + 480, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ray-primitive intersections, we need to convert each 2D Gaussian into a geometric primitive compatible with GPU processing and insert it into a bounding volume hierarchy (BVH). In light of this, we propose to represent each 2D Gaussian with two triangles. Specifically, we first define the four Gaussian bounding vertices $\\mathbf{V}_{local} = \\{(\\mathrm{sgn}(r),\\mathrm{sgn}(r))\\}$ in the local tangent plane, where $\\mathrm{sgn}(\\cdot)$ is the sign function and $r$ is set to 3 representing three times the sigma range. Then, the four local bounding vertices $\\mathbf{V}_{local}$ are transformed to world space as the vertices of the two triangles covering the Gaussian $\\mathbf{V}_{world}$ using Eq. 1. After the transformation, the triangles are organized into a BVH, which serves as the input for the ray tracing process.", + "bbox": [ + 89, + 369, + 482, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We develop a custom CUDA kernel using the raygen and anyhit programmable entry points of OptiX. Inspired by [29], the rendering is done in a chunk-by-chunk manner. The anyhit kernel traces the input ray to obtain a chunk of size $\\mathrm{k}$ , while raygen integrates this chunk and invokes anyhit to retrieve the next chunk along the ray. Specifically, consider an input ray with origin $\\mathbf{o}$ and direction $\\mathbf{d}$ . The raygen program first initiates a traversal against the BVH to identify all possible intersections along the ray. During the traversal, the anyhit program sorts each intersected Gaussian by the depth and maintains a sorted k-buffer for the closest $\\mathrm{k}$ intersections. We empirically found that $\\mathrm{k}$ with 16 is the best trade-off between traversal counts and the number of Gaussians sorted per traversal. After traversal, the raygen program integrates properties of the sorted Gaussians in the buffer following Eq. 2. The Gaussian response is calculated by applying the inverse of the transformation matrix $\\mathbf{H}$ to the ray's intersection point $\\mathbf{x}_i$ and evaluating the Gaussian value at the transformed point as:", + "bbox": [ + 89, + 566, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {i} \\left(\\mathbf {u} _ {i}\\right) = \\mathcal {G} _ {i} \\left(\\mathbf {H} ^ {- 1} \\mathbf {x} _ {i}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 859, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This process repeats until no further intersections are found", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "along the ray or the accumulated transmittance drops below a specified threshold. More details can be found in supplementary materials.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our ray tracing framework is fully differentiable, allowing for end-to-end optimization of both the base and the environment Gaussian primitives. However, storing all intersections during the forward pass and performing the backward pass in back-to-front order, as done in 3DGS [17], is impractical due to high memory consumption. To address this, we implement the backward pass in the same front-to-back order as the forward pass by re-casting rays and computing gradients for each integration step. A key aspect is calculating the gradient with respect to the input ray origin $\\frac{\\partial\\mathcal{L}}{\\partial\\mathbf{o}}$ and direction $\\frac{\\partial\\mathcal{L}}{\\partial\\mathbf{d}}$ , which is crucial for the joint optimization of our model (see supplementary for more details). Our ablation in Sec 5.4 demonstrates that this joint optimization of geometry and appearance is essential for recovering geometrically accurate surfaces.", + "bbox": [ + 511, + 136, + 906, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Optimization", + "text_level": 1, + "bbox": [ + 511, + 369, + 653, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance training stability, we initiate optimization by first training the base Gaussian $\\mathbf{P}_{\\text {base }}$ , which is initialized using the sparse point cloud obtained from Structure-from-Motion (SfM) [11, 34]. After the bootstrapping phase, we initialize the environment Gaussian $\\mathbf{P}_{env}$ by partitioning the scene's bounding box $\\mathbf{B}_{\\text {scene }}$ into $N^3$ sub-grids and randomly sampling $K$ primitives within each grid, and then optimize the base Gaussian and environment Gaussian jointly. The scene's bounding box $\\mathbf{B}_{\\text {scene }}$ is determined by the $99.5\\%$ quantile of the sparse point cloud's bounding box $\\mathbf{B}_{\\text {sfm }}$ obtained from SfM, the sub-grids resolution $N$ is set to 32, and $K$ is set to 5 for each grid.", + "bbox": [ + 511, + 392, + 906, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow 2DGS [13] to add a normal consistency constraint between the rendered normal map $\\mathbf{n}$ and the gradients of the depth map $\\mathbf{N}_d$ :", + "bbox": [ + 511, + 573, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\mathbf {n} _ {i} ^ {\\top} \\mathbf {N} _ {d}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 626, + 903, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{p}$ is the number of pixels in the image and $\\mathbf{N}_d$ is calculated as finite differences of neighboring pixels in the depth map:", + "bbox": [ + 511, + 674, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {N} _ {d} (\\mathbf {u}) = \\frac {\\nabla_ {u} \\mathbf {p} _ {d} \\times \\nabla_ {v} \\mathbf {p} _ {d}}{\\| \\nabla_ {u} \\mathbf {p} _ {d} \\times \\nabla_ {v} \\mathbf {p} _ {d} \\|}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 726, + 903, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the normal consistency constraint alone is insufficient for accurately modeling ambiguous surfaces involving both reflection and refraction. Inspired by recent advances in monocular normal estimation [1, 10, 49], we propose to supervise the rendered normal map $\\mathbf{n}$ using monocular normal estimates $\\mathbf{N}_m$ :", + "bbox": [ + 511, + 763, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m o n o}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\mathbf {n} _ {i} ^ {\\top} \\mathbf {N} _ {m}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 861, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 482, + 917, + 514, + 928 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f50060daa754012bb7e24a1d4f011b8219ba9e6a428731fd662082ea646a4023.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on real scenes. Our method significantly improves rendering quality over previous approaches, particularly in producing more detailed reflections. Zoom in for more details." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 903, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the $\\mathcal{L}_1$ image loss and D-SSIM loss $\\mathcal{L}$ employed by 3DGS [17], we also use perceptual loss [53] to enhance the perceived quality of the rendered image:", + "bbox": [ + 89, + 641, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p e r c}} = \\left\\| \\Phi (\\mathbf {I}) - \\Phi \\left(\\mathbf {I} _ {g t}\\right) \\right\\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 704, + 482, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Phi$ is the pre-trained VGG-16 network [33] and $\\mathbf{I},\\mathbf{I}_{gt}$ are the rendered and ground truth images, respectively.", + "bbox": [ + 89, + 739, + 482, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final loss function is defined as:", + "bbox": [ + 109, + 773, + 351, + 786 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {\\text {n o r m}} + \\lambda_ {2} \\mathcal {L} _ {\\text {m o n o}} + \\lambda_ {3} \\mathcal {L} _ {\\text {p e r c}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 805, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{rgb}$ is the photometric reconstruction loss combining the $\\mathcal{L}_1$ loss and a D-SSIM [43] term with a ratio of 0.8 and 0.2 respectively. We set $\\lambda_1 = 0.04$ , $\\lambda_2 = 0.01$ , and $\\lambda_3 = 0.01$ across our experiments.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 638, + 645, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 666, + 730, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement EnvGS with custom OptiX kernels and optimize our model using the PyTorch framework [32, 47] with the Adam optimizer [19]. Specifically, we set the learning rates for the parameters of each base and environment Gaussian to match those used in 2DGS [13]. The learning rate for the blending weight is set to $1e^{-2}$ . During training, we apply the adaptive Gaussian control strategy of 3DGS [17] with the normal propagation and color sabotage introduced in 3DGS-DR [18]. Since our Gaussian tracer integrates the Gaussian properties directly in 3D space, there is no valid gradient for the projected 2D center, which is used as the densification criterion in 3DGS. Following [29], we accumulate the 3D spatial gradients of the Gaussian position to achieve a similar effect. Note that each accumulated gra", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 482, + 917, + 491, + 928 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ead50a1a984a9ecc7fa6b28f6b20db3cced035415e22a55dfd6a1bab624df529.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRef-NeRF Real Scenes [37]NeRF-Casting Scenes [38]FPS↑Training Time↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Non real-timeRef-NeRF* [37]23.0870.6250.26130.5830.8900.124<0.147h
UniSDF [39]23.7000.6350.26630.8380.8890.130<0.1>47h
ZipNeRF [4]23.6770.6350.24731.7400.9040.105<0.1>47h
NeRF-Casting [38]24.6700.6590.24631.0230.8890.128<0.1>47h
Real-time3DGS [17]23.7000.6410.26228.8600.8770.159182.3070.6h
2DGS [13]23.8040.6540.28128.2760.8620.193159.1880.7h
GaussianShader [14]22.8750.6220.31426.4120.8350.21627.9451.6h
3DGS-DR [51]23.5220.6400.27428.4870.8580.197133.5931.0h
Ours24.6170.6710.24130.4440.8860.14826.2212.5h
", + "bbox": [ + 147, + 88, + 849, + 248 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4fad18c74660ece645922244c9eac995494ad9bdd46b40991f1d21602fd6b212.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on Ref-Real [37] and NeRF-Casting Scenes [38] datasets. Our method delivers the highest rendering quality among real-time techniques and outperforms several non-real-time methods, achieving competitive results with the state-of-the-art non-real-time method NeRF-Casting [38], while being 100 times faster. Note that Ref-NeRF* is an improved version of Ref-NeRF [37] that uses Zip-NeRF's geometry model. All metrics are evaluated at 1/4 resolution as prior works [38]." + ], + "table_footnote": [], + "table_body": "
MethodsForegroundNear-Field
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [17]31.6810.9550.04644.1610.9940.008
2DGS [13]31.3280.9530.05343.6240.9940.010
GaussianShader [14]30.6940.9470.05842.3910.9930.010
3DGS-DR31.8140.9530.05044.0070.9940.009
Ours33.2950.9620.04046.3920.9960.007
", + "bbox": [ + 112, + 328, + 460, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dient is scaled by half of the intersection depth to prevent under-densification in distant regions. All experiments are conducted on a single NVIDIA RTX 4090 GPU.", + "bbox": [ + 89, + 483, + 483, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 540, + 377, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train and evaluate EnvGS on a range of datasets with a focus on real-world scenes characterized by complex view-dependent effects. We evaluate the Ref-Real [37] and NeRF-Casting Shiny Scenes [38] to demonstrate our method's ability for complex specular reflections in real-world scenes. We additionally captured two more real-world scenes for a more comprehensive evaluation. We demonstrate that our methods can reconstruct detailed reflections on complex real-world scenes with real-time rendering speed. Additionally, we evaluate our method on the synthetic Shiny Blender dataset [37], which is rendered using environment maps. More evaluation results can be found in the supplementary materials.", + "bbox": [ + 89, + 563, + 482, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We maintain consistent training and testing splits and image resolution across all datasets, following prior works [3, 17, 37]. We use three commonly used metrics for evaluation: PSNR, SSIM [43], and LPIPS [53].", + "bbox": [ + 89, + 760, + 483, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Baseline Comparisons", + "text_level": 1, + "bbox": [ + 89, + 832, + 297, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with both implicit and explicit prior works, including Zip-NeRF [13], 3DGS [17], 2DGS [13], which are designed for general scenes with pri", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e94dfbc6fb837772da6f1d4cc86e38ffe2ef454a9e6b4f10928123923436f263.jpg", + "table_caption": [ + "Table 2. Quantitative results of foreground and near-field region on Ref-Real [37] and NeRF-Casting Scenes [38]. See supplementary Appendix A.1 for more details and qualitative results." + ], + "table_footnote": [], + "table_body": "
MethodsAudiDog
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Ref-NeRF [37]24.5290.7130.39723.6200.6730.389
3DGS [17]26.1710.8250.18125.3000.8820.165
2DGS [13]25.8690.8190.19824.9330.8760.184
GaussianShader [14]24.7680.7950.22523.0610.8400.226
3DGS-DR [51]26.0930.8200.19223.9430.8500.210
Ours28.6290.8760.12125.5990.8850.151
", + "bbox": [ + 535, + 328, + 883, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results on our self-captured scenes.", + "bbox": [ + 529, + 441, + 887, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "marily diffuse appearances, as well as with methods specifically tailored for scenes with strong specular reflections, including Ref-NeRF [37], GaussianShader [14], 3DGS-DR [18], and NeRF-Casting [38]. We also compare with ENVIDR [22], NDE [45] these two object-level methods.", + "bbox": [ + 511, + 469, + 906, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present quantitative and qualitative results to illustrate the advantages of our methods. Since accurate reflection reconstruction and rendering are the essential advantages of our method, we encourage readers to refer to the various rendered continuous video results in the supplementary material for a more comprehensive evaluation of our method.", + "bbox": [ + 511, + 547, + 905, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first evaluate our method on nine real-world scenes from the Ref-Real dataset [37], NeRF-Casting Shiny Scenes [38] and our self-captured real-world scenes, which feature complex geometry and specular reflections. The quantitative results, shown in Tab. 1 and Tab. 3, demonstrate that our method outperforms all explicit methods by a large margin and achieves comparable results to the current state-of-the-art implicit method, NeRF-Casting, while being significantly faster. The qualitative results in Fig.4 further highlight the superior performance of our method in capturing complex view-dependent effects, especially nearfield reflections and high-frequency details.", + "bbox": [ + 511, + 656, + 905, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide additional comparisons on the Mip-NeRF 360 dataset [3], the Shiny Blender dataset [37] and with ENVIDR [22], NDE [45] in supplementary Appendix A, which also includes per-scene metric breakdowns from Tab. 1.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5748", + "bbox": [ + 482, + 917, + 514, + 928 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c760e364bbce1deff1cd01d9d5e019dc22fc49f5b30caaa27c86c59ef5a62e4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 88, + 225, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2791f0a2d6bee68f3e7acba85d3a66896169bc981aa2de7bc27b6de9dec373d4.jpg", + "image_caption": [ + "(a) w/o joint optimization" + ], + "image_footnote": [], + "bbox": [ + 93, + 193, + 225, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe0fde4400901a472eb394fcbcca998095ac1c6ab6f290cf0d2cb3bb93a71a5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 88, + 362, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb20b86d0890f29745968e343dd29d123ed0550111ab051385579c9b7bc19a8b.jpg", + "image_caption": [ + "(b) w/o mono normal loss" + ], + "image_footnote": [], + "bbox": [ + 228, + 193, + 361, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31296fec1e6c85513c6d78699d8ee602da67f339fab9eba9c7efe7a97ee941c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 496, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1481b2a5b7bebafef9ac43e119dc91ba90c4c478f6636edcc4bc339525cc4d5.jpg", + "image_caption": [ + "(c) w/ environment map", + "Figure 5. Ablation study of proposed components on the Ref-Real dataset [37]. Removing either the monocular normal constraint or the joint optimization of base and environment Gaussians results in noisy geometry and inaccurate reflection reconstruction. The \"w/ environment map\" variant fails to capture near-field reflections." + ], + "image_footnote": [], + "bbox": [ + 364, + 193, + 496, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29ed4b41adca4e473701762271b7f73583c3b7ccc6ee4e45ed67145333d0bb44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 633, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/766b3bc215f9be33c0ba73681bd6d72d0ba4c51cc66a9ea7835fd2f94636aecf.jpg", + "image_caption": [ + "(d) w/o normal propagation" + ], + "image_footnote": [], + "bbox": [ + 500, + 193, + 633, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b627fda375fbb21841a9d8a7dc1502d3c9427fea7ba2a43c039ec39c201900c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 767, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e2577d09bd87237be35a1e685f52d1ebd981e27f3e83c787d16419672e887a7.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 635, + 193, + 767, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5ecd83fb96bc13565b89964d1524297ab984bf776cdec2c07136025e200b7de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 88, + 905, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d03ef3e2689e95fb639a8a29b94bf0e01b71f3418f6506a1eefdf09650ee86dc.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 769, + 193, + 903, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method achieves competitive results on both datasets, demonstrating its versatility and effectiveness in handling diverse scenes with complex view-dependent effects.", + "bbox": [ + 89, + 378, + 482, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 433, + 254, + 448 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/793a7d2400e158b3f9370fb0c0ef83083b52dc506ca58a5613c238d7e1733e87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
w/o joint optimization24.0340.6440.287
w/o mono normal loss24.1070.6480.270
w/ environment map24.1450.6460.285
w/o color sabotage24.2680.6500.269
w/o normal propagation24.1920.6520.271
w/o lpips loss24.5670.6710.280
Ours24.6170.6710.241
", + "bbox": [ + 112, + 457, + 460, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation studies.", + "bbox": [ + 205, + 563, + 367, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies of our key components on the Ref-Real dataset. Quantitative and qualitative results are shown in Tab. 4 and Fig. 5, respectively.", + "bbox": [ + 89, + 583, + 482, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Joint optimization. The \"w/o joint optimization\" variant detaches the joint optimization of the base Gaussian and environment Gaussian from the reflection rendering step. As shown in Tab. 4 and Fig. 5, This variant fails to recover accurate geometry, leading to inferior reflection reconstruction and rendering quality.", + "bbox": [ + 89, + 628, + 482, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Monocular normal constraint The \"w/o monocular normal\" variant removes the monocular normal constraint as described in Sec. 4.3. Training may become trapped in largely incorrect geometry, resulting in inaccurate reflection reconstruction, as demonstrated in Tab.4 and Fig.5.", + "bbox": [ + 89, + 719, + 482, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Environment map representation The \"w/ environment map\" variant replaces our core Gaussian environment representation with an environment map representation while keeping all other components unchanged. Figure 5 illustrates that while it effectively captures smooth distant reflections, it has difficulty modeling near-field and high-frequency reflections, and produces more bumpy geometry.", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Color sabotage and normal propagation. The “ $w/o$ color sabotage” and “ $w/o$ normal propagation” variants omit the color sabotage and normal propagation steps from our method, respectively. As demonstrated in Tab. 4, both variants result in reduced rendering quality.", + "bbox": [ + 511, + 378, + 905, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perceptual loss. The \"w/o lipps loss\" variants removes the perceptual loss. The results in Tab. 4 demonstrate that our method still produces detailed and accurate reflections without perceptual loss, which offers only marginal improvements.", + "bbox": [ + 511, + 455, + 905, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 544, + 761, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduced EnvGS, a novel reflective scene representation for high-quality complex reflection capturing and real-time rendering. Our method explicitly models reflections with a set of environment Gaussian primitives. The environment Gaussian primitives are used together with a set of base Gaussian primitives that model basic scene properties (geometry, base color, and blending weight) to model the appearance of the whole scene. Furthermore, we develop a differentiable Gaussian ray tracer utilizing GPU's RT core to effectively optimize and render the environment Gaussian. The proposed method demonstrates superior performance in capturing complex reflections across various datasets.", + "bbox": [ + 511, + 570, + 906, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A limitation of EnvGS is its difficulty with transparent and refractive materials, as it only addresses reflection direction. Future work could explore extending our method to accommodate these materials.", + "bbox": [ + 511, + 767, + 905, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was partially supported by NSFC (No. U24B20154, 62402427), Ant Group, and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5749", + "bbox": [ + 482, + 917, + 514, + 928 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5", + "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2", + "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7, 1, 8", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2, 7, 1, 6, 8", + "[5] Zoubin Bi, Yixin Zeng, Chong Zeng, Fan Pei, Xiang Feng, Kun Zhou, and Hongzhi Wu. Gs3: Efficient relighting with triple gaussian splatting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 3", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2", + "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[8] Chen Guangcheng, He Yicheng, He Li, and Zhang Hong. Pisr: Polarimetric neural implicit surface reconstruction for textureless and specular objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2", + "[9] Yang Hang, Chen Rui, An Shipeng, Wei Hao, and Zhang Heng. The growth of image-related three dimensional reconstruction techniques in deep learning-driven era: a critical summary, 2023. 2", + "[10] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 5", + "[11] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 5", + "[12] Fuyuan Hu, Chenlu Li, Tao Zhou, Hongfu Cheng, and Minming Gu. A survey on point cloud completion algorithms for deep learning, 2025. 2", + "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 4, 5, 6, 7, 8", + "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 3, 7, 6, 8", + "[15] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for re" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "flective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 2, 4, 1", + "[16] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 3, 5, 6, 7, 2, 8", + "[18] Ye Keyang, Hou Qiming, and Zhou Kun. 3d gaussian splatting with deferred reflection. 2024. 3, 6, 7, 1", + "[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[20] Yingqun Li, Xiao Hu, Xiang Xu, Yanning Xu, and Lu Wang. Deep learning-based foveated rendering in 3d space: a review, 2024. 2", + "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[22] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 2, 7", + "[23] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. 2025. 2", + "[24] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2", + "[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 2", + "[26] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3", + "[27] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections, 2023. 2", + "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[29] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian Ray Tracing: Fast tracing of particle scenes. ACM Transactions on Graphics and SIGGRAPH Asia, 2024. 4, 5, 6", + "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5750", + "bbox": [ + 482, + 917, + 514, + 928 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[31] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. Acm transactions on graphics (tog), 29(4):1-13, 2010. 2, 4", + "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6", + "[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 5", + "[35] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2", + "[36] Zhe Jun Tang and Tat-Jen Cham. 3igs: Factorised tensorial illumination for 3d gaussian splatting. arXiv preprint arXiv:2408.03753, 2024. 3, 7", + "[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3, 7, 8, 1, 6", + "[38] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2, 7, 1, 6, 8", + "[39] Fangjinhua Wang, Marie-Julie Rakotosaona, Michael Niemeyer, Richard Szeliski, Marc Pollefeys, and Federico Tombari. Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In NeurIPS, 2024. 2, 7, 6, 8", + "[40] Jinke Wang, Xingxing Zuo, Xiangrui Zhao, Jiajun Lyu, and Yong Liu. Review of multi-source fusion slam: current status and challenges, 2022. 2", + "[41] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2", + "[42] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7", + "[44] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall," + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kaspar Rohrer, Norbert Felber, et al. A hardware architecture for surface splattering. ACM Transactions on Graphics (TOG), 26(3):90-es, 2007. 3", + "[45] Liwen Wu, Sai Bi, Zexiang Xu, Fujun Luan, Kai Zhang, Iliyan Georgiev, Kalyan Sunkavalli, and Ravi Ramamoorthi. Neural directional encoding for efficient and accurate view-dependent appearance modeling. In CVPR, 2024. 7, 2", + "[46] Jiankai Xing and Kun Xu. Physically based differentiable rendering: a survey, 2024. 2", + "[47] Zhen Xu, Tao Xie, Sida Peng, Haotong Lin, Qing Shuai, Zhiyuan Yu, Guangzhao He, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Easyvolcap: Accelerating neural volumetric video research. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6", + "[48] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2", + "[49] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics, 2024. 4, 5", + "[50] Hanqiao Ye, Yangdong Liu, and Shuhan Shen. Lightweight visual-based localization technology, 2024. 2", + "[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2, 4, 7, 1, 6, 8", + "[52] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7", + "[54] Youjia Zhang, Anpei Chen, Yumin Wan,zikai Song, Junqing Yu, Yawei Luo, and Wei Yang. Ref-gs: Directional factorization for 2d gaussian splatting. arXiv preprint arXiv:2412.00905, 2024.3" + ], + "bbox": [ + 516, + 92, + 903, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5751", + "bbox": [ + 482, + 917, + 513, + 928 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_model.json b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..421bb5ca59331f428c850a82840ee3d751b34d1f --- /dev/null +++ b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "EnvGS: Modeling View-Dependent Appearance with Environment Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.179, + 0.758, + 0.244 + ], + "angle": 0, + "content": "Tao Xie\\(^{1*}\\) Xi Chen\\(^{1*}\\) Zhen Xu\\(^{1}\\) Yiman Xie\\(^{1}\\) Yudong Jin\\(^{1}\\) \nYujun Shen\\(^{2}\\) Sida Peng\\(^{1}\\) Hujun Bao\\(^{1}\\) Xiaowei Zhou\\(^{1\\dagger}\\) \n\\(^{1}\\) Zhejiang University \\(^{2}\\) Ant Group" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.277, + 0.295, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.389, + 0.295, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.501, + 0.242, + 0.515 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.277, + 0.498, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.389, + 0.498, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.501, + 0.416, + 0.515 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.277, + 0.702, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.389, + 0.702, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.501, + 0.658, + 0.515 + ], + "angle": 0, + "content": "GaussianShader" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.277, + 0.905, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.389, + 0.905, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.501, + 0.825, + 0.515 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.535, + 0.908, + 0.579 + ], + "angle": 0, + "content": "Figure 1. Photorealistic, real-time rendering of real-world scenes with complex reflections. Our proposed EnvGS outperforms prior works in capturing complex reflection effects, especially near-field reflections and high-frequency details while maintaining real-time rendering speed. Please see our supplementary video for better visualizations." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.607, + 0.327, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Reconstructing complex reflections in real-world scenes from 2D images is essential for achieving photorealistic novel view synthesis. Existing methods that utilize environment maps to model reflections from distant lighting often struggle with high-frequency reflection details and fail to account for near-field reflections. In this work, we introduce EnvGS, a novel approach that employs a set of Gaussian primitives as an explicit 3D representation for capturing reflections of environments. These environment Gaussian primitives are incorporated with base Gaussian primitives to model the appearance of the whole scene. To efficiently render these environment Gaussian primitives, we developed a ray-tracing-based renderer that leverages the GPU's RT core for fast rendering. This allows us to jointly optimize our model for high-quality reconstruction while maintain-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.909, + 0.671 + ], + "angle": 0, + "content": "ing real-time rendering speeds. Results from multiple realworld and synthetic datasets demonstrate that our method produces significantly more detailed reflections, achieving the best rendering quality in real-time novel view synthesis. The code is available at https://zju3dv.github.io/envgs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.707, + 0.646, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Novel view synthesis aims to generate novel views of 3D scenes based on a set of input images, which enables many applications such as VR/AR, and autonomous driving. Recent advances in Neural Radiance Fields (NeRF) [28] have demonstrated impressive rendering performance. However, NeRF's high computational cost makes it challenging for real-time applications. More recently, 3D Gaussian Splatting (3DGS) [17] explicitly models scenes with 3D Gaussian primitives and utilizes rasterization for rendering, achieving real-time rendering with competitive quality. However, modeling complex high-frequency specular" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.444, + 0.901 + ], + "angle": 0, + "content": "* Equal Contribution. † Corresponding author: Xiaowei Zhou" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "reflections remains challenging for 3DGS due to the limited expressiveness of the Spherical Harmonics (SH)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Recent works GaussianShader [15] and 3DGS-DR [51], enhance 3DGS by integrating an environment map and employing shading functions to blend the appearance from both the environment map and SH for the final rendering. While additional environmental lighting can enhance 3DGS's reflection modeling ability, it still struggles to reconstruct complex specular reflections accurately due to two factors. First, the assumption of distant lighting in environment maps limits their ability to only capture distant illumination and difficult to synthesize accurate near-field reflections. Second, this representation inherently lacks sufficient capacity to capture high-frequency reflection details." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.485, + 0.625 + ], + "angle": 0, + "content": "In this paper, we present EnvGS, a novel approach for modeling complex reflections in real-world scenes, addressing the aforementioned challenges. We propose to model reflections using a set of Gaussian primitives termed the environment Gaussian. The geometry and base appearance are represented by another set of Gaussian primitives called base Gaussian. We effectively blend the two Gaussians for rendering and optimization. Our rendering process begins with rendering the base Gaussian for the per-pixel surface position, normal, base color, and blending weight. Next, we render the environment Gaussian at the surface point in the direction of the reflection of the viewing direction around the surface normal to capture reflection colors. Finally, we blend the base color with the reflection color to achieve the final rendering results. In contrast to previous methods, EnvGS captures high-frequency reflection details using Gaussian primitives, offering superior modeling capabilities compared to environment maps. Additionally, our explicit 3D reflection representation eliminates the need for distant lighting assumptions, enabling accurate modeling of near-field reflections, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.868 + ], + "angle": 0, + "content": "To render the environment Gaussian at each intersection point along the reflection direction, we create a fully differentiable ray-tracing renderer for 2DGS since rasterization is not suited for this task. We build the ray-tracing renderer on CUDA and OptiX [31] for real-time rendering and efficient optimization of environment Gaussian. The rendering process starts by constructing a bounding volume hierarchy (BVH) from the 2D Gaussian primitives. We then cast rays against the BVH, gathering ordered intersections in chunks while integrating the Gaussian properties through volume rendering [16] to achieve the final results. Our Gaussian ray-tracing renderer enables detailed reflection rendering at real-time performance. Furthermore, it allows for efficient joint optimization of the environment Gaussian and base Gaussian, which is essential for accurate reflection modeling, as demonstrated in Sec 5.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To validate the effectiveness of our method, we evaluate EnvGS on several real and synthetic datasets. The re" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "sults demonstrate that our methods achieve state-of-the-art performance in real-time novel view synthesis and considerably surpass existing real-time methods, particularly in synthesizing complex reflections in real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.153, + 0.867, + 0.167 + ], + "angle": 0, + "content": "In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.211 + ], + "angle": 0, + "content": "- We propose a novel scene representation for accurately modeling complex near-field and high-frequency reflections in real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- We developed a real-time ray-tracing renderer for 2DGS, enabling joint optimization of our representation for accurate scene reconstruction while achieving real-time rendering speeds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.347 + ], + "angle": 0, + "content": "- Extensive experiments shows that EnvGS significantly outperforms previous methods. To the best of our knowledge, EnvGS is the first method that achieves real-time photorealistic specular reflections synthesizing in real-world scenes." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.36, + 0.655, + 0.375 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.905, + 0.462 + ], + "angle": 0, + "content": "In computer vision and graphics research [9, 12, 20, 23, 40, 46, 50], our work falls into the area of learning scene representations from a set of posed RGB images. In this section, we review NeRF-based and Gaussian Splitting-based methods, particularly their handling of view-dependent effects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Neural Radiance Field. NeRF [28] introduced the concept of neural radiance fields, which model scenes as implicit multilayer perceptrons (MLPs) and render them via volume rendering, achieving impressive results in novel view synthesis. Subsequent advancements focused on enhancing rendering quality [2-4] and computational efficiency. [6, 7, 24, 30, 35]. [21, 41, 42, 48] introduce Signed Distance Field into NeRF to improve geometry quality. However, these methods often model view-dependent effects via simple viewing directions, which can lead to blurry reflection renderings. To address this, Ref-NeRF [37] encodes the outgoing radiance using the reflected view direction, yielding improved results under distant lighting conditions. Follow-up works [8, 22, 25, 39] leverage Signed Distance Fields (SDF) to refine surface normals to enhance reflection and geometry quality. SpecNeRF [27] further incorporates spatially varying Gaussian directional encoding to better capture near-field reflections. NeRF-Casting [38] represents the scene in a unified manner, similar to [4], and performs ray marching along reflection directions to integrate reflection features, which are then decoded into color with MLPs. This approach achieves impressive results on real-world data under both near and distant lighting conditions. However, the requirement for multiple MLP queries per ray makes NeRF-Casting unsuitable for real-time rendering and necessitates substantial training time. In contrast, our approach delivers real-time rendering capabilities while significantly reducing training time." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5743" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.622 + ], + "angle": 0, + "content": "Gaussian Splatting. Recently, 3D Gaussian Splatting (3DGS) [17] has made significant strides toward real-time novel view synthesis. Unlike volume-based rendering, Gaussian Splatting uses efficient rasterization to render spatial Gaussian kernels, achieving real-time performance at 1080P resolution. Mip-Splatting [52] further introduces a 3D smoothing filter and a 2D Mip filter for alias-free renderings. Scaffold-GS [26] introduces structured 3D Gaussians to improve rendering efficiency and quality. [5] introduces triple splitting for efficient relighting. 2D Gaussian Splatting (2DGS) [13] replaces the 3D Gaussian kernels with 2D Gaussian that better align with 3D surfaces, leading to more accurate surface reconstruction. However, these methods face challenges in accurately modeling reflections due to their use of spherical harmonics (SH) to parameterize view-dependent effects based solely on viewing direction, which often results in blurry reflections. More recent works GaussianShader [14] and 3DGS-DR [18] try to incorporate additional environment maps to improve reflection modeling ability. However, these methods only consider distant lighting, ignoring near-field lighting, and are unable to capture high-frequency reflection details. 3iGS [36] extends 3DGS by augmenting it with an illumination field using tensorial factorization, rendering final reflections through a neural renderer. However, it is limited to bounded scenes, which restricts its applicability in real-world, unbounded environments. Concurrent work [54] choose to model near-field reflection with tensorial factorization and far-field with spherical feature grid. Our method adopts 2DGS as the scene representation and models environmental illumination using an additional set of environment Gaussian, which is rendered via our proposed Gaussian tracer. Our approach inherently supports high-frequency, near-field, and distant lighting in unbounded scenes, enabling detailed reflection rendering while maintaining real-time performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.218, + 0.649 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.779 + ], + "angle": 0, + "content": "We begin by introducing 2D Gaussian Splatting (2DGS) [13], which our approach is built upon. 2DGS is an explicit scene representation similar to 3D Gaussian Splatting (3DGS) [17], which uses Gaussian primitives and rasterization to render screen projections. The key difference is that 2DGS represents the scene with a set of scaled 2D Gaussian primitives defined in a local tangent plane within world space by a transformation matrix \\(\\mathbf{H}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.785, + 0.483, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\left[ \\begin{array}{c c c c} s _ {u} \\mathbf {t} _ {u} & s _ {v} \\mathbf {t} _ {v} & 0 & \\mathbf {p} _ {k} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.871 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_k, (\\mathbf{t}_u, \\mathbf{t}_v)\\), and \\((s_u, s_v)\\) denote the center, the two principal tangential vectors, and the scaling factors of the Gaussian, respectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To render an image, 2DGS employs the method described in [44] to determine the ray-primitive intersections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "These intersection points are subsequently utilized to compute the Gaussian's contribution to the final image. The Gaussian properties are then integrated using a volume rendering algorithm to obtain the final pixel color:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.161, + 0.907, + 0.204 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\text {w i t h} \\alpha_ {i} = \\sigma_ {i} \\mathcal {G} _ {i}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.89, + 0.227 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}(\\cdot)\\) is the standard 2D Gaussian value evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Compared to 3D Gaussian, 2D Gaussian offers distinct advantages as a surface representation. First, the ray-splat intersection method adopted by 2DGS avoids multi-view depth inconsistency. Second, 2D Gaussian inherently provides a well-defined normal, which is essential for high-quality surface reconstruction and accurate reflection calculations. However, 2DGS relies on the limited representational capacity of Spherical Harmonics (SH) to model scene appearance, preventing it from capturing strong view-dependent effects like specular reflections, which leads to poor rendering results and \"foggy\" geometry [37]. To this end, we use the geometry-aligned 2D Gaussian primitives as the base scene representation and demonstrate how we effectively model complex reflections in the next section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.605, + 0.468 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.909, + 0.706 + ], + "angle": 0, + "content": "Given a set of input images of a reflective scene, our goal is to reconstruct the 3D scene and synthesize photorealistic novel views in real-time. To achieve this, we propose utilizing environment Gaussian as an explicit 3D environment representation, which enables accurate modeling of complex reflections in real-world scenes. Additionally, we represent the scene geometry and base colors using another set of Gaussian primitives, denoted as base Gaussian. In this section, we first detail how environment Gaussian and base Gaussian work together to model complex reflections within the scene (Sec. 4.1). Then, We describe the design of a ray-tracing renderer that leverages the GPU's RT cores to efficiently render and optimize the environment Gaussian (Sec. 4.2). Finally, we discuss our optimization process. (Sec. 4.3). An overview of our method is in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.759, + 0.729 + ], + "angle": 0, + "content": "4.1. Reflective Scenes Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Gaussian splatting [13, 17] models appearance using Spherical Harmonics (SH), which has limited representation capacity for view-dependent effects. These limitations hinder its ability to capture complex, high-frequency specular reflections. Building on this observation, our key insight is that modeling reflections with a Gaussian environment representation can better model complex reflection effects while significantly reducing the complexity required for each Gaussian to capture intricate details within its SH." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our proposed reflective scene representation includes two sets of 2D Gaussians: a base Gaussian \\(\\mathbf{P}_{\\text {base }}\\) for mod-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5744" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.404, + 0.908, + 0.463 + ], + "angle": 0, + "content": "Figure 2. Overview of EnvGS. The rendering process begins by rasterizing the base Gaussian to obtain per-pixel normals, base colors, and blending weights. Next, we render the environment Gaussian in the reflection direction using our ray-tracing-based Gaussian renderer to capture the reflection colors. Finally, we combine the reflection and base colors for the final output. We jointly optimize the environment Gaussian and base Gaussian using monocular normals [49] and ground truth images for supervision." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.483, + 0.61 + ], + "angle": 0, + "content": "eling the scene's geometry and base appearance, and another environment Gaussian \\(\\mathbf{P}_{env}\\) for capturing scene reflections. The basic parameterization of each Gaussian primitive is consistent with the original 2DGS [13], including 3D center position \\(\\mathbf{p}\\), opacity \\(\\alpha\\), two principal tangential vectors \\((\\mathbf{t}_u,\\mathbf{t}_v)\\), a scaling vector \\((s_u,s_v)\\), and SH coefficients. To combine the two appearance components from the base and environment Gaussian into the final result, we introduce a blending weight for each base Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.701 + ], + "angle": 0, + "content": "The rendering process is performed in three steps. First, the base Gaussian \\(\\mathbf{P}_{base}\\) is rendered using standard 2D Gaussian splating to obtain the base color \\(\\mathbf{c}_{base}\\). By applying the volume rendering integration using Eq. (2), we also derive the surface position \\(\\mathbf{x}\\), surface normals \\(\\mathbf{n}\\), and blending weight \\(\\beta\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.708, + 0.483, + 0.75 + ], + "angle": 0, + "content": "\\[\nv = \\sum_ {i \\in \\mathcal {N}} = v _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), v \\in \\{\\mathbf {x}, \\mathbf {n}, \\beta \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Then, we compute the reflection direction \\(\\mathbf{d}_{ref}\\) based on the camera ray direction \\(\\mathbf{d}_{ref}\\) and the surface normal \\(\\mathbf{n}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.794, + 0.483, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {r e f} = \\mathbf {d} _ {c a m} - 2 \\left(\\mathbf {d} _ {c a m} \\cdot \\mathbf {n}\\right) \\mathbf {n}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.484, + 0.879 + ], + "angle": 0, + "content": "With the reflection direction \\(\\mathbf{d}_{ref}\\) and surface point \\(\\mathbf{x}\\), the environment Gaussian \\(\\mathbf{P}_{env}\\) is rendered using our differentiable Gaussian tracer to obtain the reflection color \\(\\mathbf{c}_{ref}\\), as detailed in Sec. 4.2. The final color is obtained through:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.886, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = (1 - \\beta) \\cdot \\mathbf {c} _ {\\text {b a s e}} + \\beta \\cdot \\mathbf {c} _ {\\text {r e f}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.905, + 0.519 + ], + "angle": 0, + "content": "We blend the base color \\(\\mathbf{c}_{\\text {base }}\\) and reflection color \\(\\mathbf{c}_{\\text {ref }}\\) using the blending weight \\(\\beta\\). A visualization of \\(\\mathbf{c}_{\\text {base }}\\) and \\(\\mathbf{c}_{\\text {ref }}\\) can be found at Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Discussion. Compared to the environment map used by Gaussianshader [15] and 3DGS-DR [51], our explicit Gaussian environment representation offers several advantages. First, our method more accurately captures near-field reflections caused by occlusions from nearby objects. This improvement arises from explicitly modeling each Gaussian at its exact spatial location, thus avoiding the ambiguities and inaccuracies inherent in environment map representations that assume distant lighting. Second, by utilizing Gaussian primitives, EnvGS's environment representation achieves greater expressiveness than low-frequency environment maps, enabling the capture of finer reflection details and enhancing rendering quality, as demonstrated by our experiments 5.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.757, + 0.759 + ], + "angle": 0, + "content": "4.2. Differentiable Ray Tracing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.884 + ], + "angle": 0, + "content": "Rendering the environment Gaussian with rasterization is impractical, as each pixel corresponds to a unique reflection ray and functions as a virtual camera. To address this, we draw on [29] and leverage the advanced optimizations of modern GPUs to design a novel, fully differentiable ray tracing framework. Built on OptiX [31], our framework achieves real-time rendering of 2,000,000 2DGS with a resolution of \\(1292 \\times 839\\) at 30 FPS on an RTX 4090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In order to fully utilize the hardware acceleration for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.918, + 0.516, + 0.929 + ], + "angle": 0, + "content": "5745" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.223, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.089, + 0.353, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.089, + 0.482, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.19, + 0.223, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.29, + 0.183, + 0.299 + ], + "angle": 0, + "content": "Final color" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.19, + 0.353, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.29, + 0.313, + 0.299 + ], + "angle": 0, + "content": "Base color" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.19, + 0.482, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.29, + 0.453, + 0.299 + ], + "angle": 0, + "content": "Reflection color" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.483, + 0.358 + ], + "angle": 0, + "content": "Figure 3. Visualization of reflection and base color. Our method successfully reconstructs near-field and distant reflections using the environment Gaussian instead of baking into the base color." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.566 + ], + "angle": 0, + "content": "ray-primitive intersections, we need to convert each 2D Gaussian into a geometric primitive compatible with GPU processing and insert it into a bounding volume hierarchy (BVH). In light of this, we propose to represent each 2D Gaussian with two triangles. Specifically, we first define the four Gaussian bounding vertices \\(\\mathbf{V}_{local} = \\{(\\mathrm{sgn}(r),\\mathrm{sgn}(r))\\}\\) in the local tangent plane, where \\(\\mathrm{sgn}(\\cdot)\\) is the sign function and \\(r\\) is set to 3 representing three times the sigma range. Then, the four local bounding vertices \\(\\mathbf{V}_{local}\\) are transformed to world space as the vertices of the two triangles covering the Gaussian \\(\\mathbf{V}_{world}\\) using Eq. 1. After the transformation, the triangles are organized into a BVH, which serves as the input for the ray tracing process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We develop a custom CUDA kernel using the raygen and anyhit programmable entry points of OptiX. Inspired by [29], the rendering is done in a chunk-by-chunk manner. The anyhit kernel traces the input ray to obtain a chunk of size \\( \\mathrm{k} \\), while raygen integrates this chunk and invokes anyhit to retrieve the next chunk along the ray. Specifically, consider an input ray with origin \\( \\mathbf{o} \\) and direction \\( \\mathbf{d} \\). The raygen program first initiates a traversal against the BVH to identify all possible intersections along the ray. During the traversal, the anyhit program sorts each intersected Gaussian by the depth and maintains a sorted k-buffer for the closest \\( \\mathrm{k} \\) intersections. We empirically found that \\( \\mathrm{k} \\) with 16 is the best trade-off between traversal counts and the number of Gaussians sorted per traversal. After traversal, the raygen program integrates properties of the sorted Gaussians in the buffer following Eq. 2. The Gaussian response is calculated by applying the inverse of the transformation matrix \\( \\mathbf{H} \\) to the ray's intersection point \\( \\mathbf{x}_i \\) and evaluating the Gaussian value at the transformed point as:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.861, + 0.483, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {i} \\left(\\mathbf {u} _ {i}\\right) = \\mathcal {G} _ {i} \\left(\\mathbf {H} ^ {- 1} \\mathbf {x} _ {i}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This process repeats until no further intersections are found" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "along the ray or the accumulated transmittance drops below a specified threshold. More details can be found in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Our ray tracing framework is fully differentiable, allowing for end-to-end optimization of both the base and the environment Gaussian primitives. However, storing all intersections during the forward pass and performing the backward pass in back-to-front order, as done in 3DGS [17], is impractical due to high memory consumption. To address this, we implement the backward pass in the same front-to-back order as the forward pass by re-casting rays and computing gradients for each integration step. A key aspect is calculating the gradient with respect to the input ray origin \\(\\frac{\\partial\\mathcal{L}}{\\partial\\mathbf{o}}\\) and direction \\(\\frac{\\partial\\mathcal{L}}{\\partial\\mathbf{d}}\\), which is crucial for the joint optimization of our model (see supplementary for more details). Our ablation in Sec 5.4 demonstrates that this joint optimization of geometry and appearance is essential for recovering geometrically accurate surfaces." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.655, + 0.387 + ], + "angle": 0, + "content": "4.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.574 + ], + "angle": 0, + "content": "To enhance training stability, we initiate optimization by first training the base Gaussian \\(\\mathbf{P}_{\\text {base }}\\), which is initialized using the sparse point cloud obtained from Structure-from-Motion (SfM) [11, 34]. After the bootstrapping phase, we initialize the environment Gaussian \\(\\mathbf{P}_{env}\\) by partitioning the scene's bounding box \\(\\mathbf{B}_{\\text {scene }}\\) into \\(N^3\\) sub-grids and randomly sampling \\(K\\) primitives within each grid, and then optimize the base Gaussian and environment Gaussian jointly. The scene's bounding box \\(\\mathbf{B}_{\\text {scene }}\\) is determined by the \\(99.5\\%\\) quantile of the sparse point cloud's bounding box \\(\\mathbf{B}_{\\text {sfm }}\\) obtained from SfM, the sub-grids resolution \\(N\\) is set to 32, and \\(K\\) is set to 5 for each grid." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.62 + ], + "angle": 0, + "content": "We follow 2DGS [13] to add a normal consistency constraint between the rendered normal map \\(\\mathbf{n}\\) and the gradients of the depth map \\(\\mathbf{N}_d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.627, + 0.905, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\mathbf {n} _ {i} ^ {\\top} \\mathbf {N} _ {d}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.72 + ], + "angle": 0, + "content": "where \\(N_{p}\\) is the number of pixels in the image and \\(\\mathbf{N}_d\\) is calculated as finite differences of neighboring pixels in the depth map:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.727, + 0.905, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathbf {N} _ {d} (\\mathbf {u}) = \\frac {\\nabla_ {u} \\mathbf {p} _ {d} \\times \\nabla_ {v} \\mathbf {p} _ {d}}{\\| \\nabla_ {u} \\mathbf {p} _ {d} \\times \\nabla_ {v} \\mathbf {p} _ {d} \\|}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.855 + ], + "angle": 0, + "content": "However, the normal consistency constraint alone is insufficient for accurately modeling ambiguous surfaces involving both reflection and refraction. Inspired by recent advances in monocular normal estimation [1, 10, 49], we propose to supervise the rendered normal map \\(\\mathbf{n}\\) using monocular normal estimates \\(\\mathbf{N}_m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.862, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m o n o}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\mathbf {n} _ {i} ^ {\\top} \\mathbf {N} _ {m}\\right). \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.918, + 0.516, + 0.929 + ], + "angle": 0, + "content": "5" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.904, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.602, + 0.908, + 0.631 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on real scenes. Our method significantly improves rendering quality over previous approaches, particularly in producing more detailed reflections. Zoom in for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.688 + ], + "angle": 0, + "content": "In addition to the \\(\\mathcal{L}_1\\) image loss and D-SSIM loss \\(\\mathcal{L}\\) employed by 3DGS [17], we also use perceptual loss [53] to enhance the perceived quality of the rendered image:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.705, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p e r c}} = \\left\\| \\Phi (\\mathbf {I}) - \\Phi \\left(\\mathbf {I} _ {g t}\\right) \\right\\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.483, + 0.771 + ], + "angle": 0, + "content": "where \\(\\Phi\\) is the pre-trained VGG-16 network [33] and \\(\\mathbf{I},\\mathbf{I}_{gt}\\) are the rendered and ground truth images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.774, + 0.352, + 0.787 + ], + "angle": 0, + "content": "The final loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.806, + 0.483, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {\\text {n o r m}} + \\lambda_ {2} \\mathcal {L} _ {\\text {m o n o}} + \\lambda_ {3} \\mathcal {L} _ {\\text {p e r c}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{rgb}\\) is the photometric reconstruction loss combining the \\(\\mathcal{L}_1\\) loss and a D-SSIM [43] term with a ratio of 0.8 and 0.2 respectively. We set \\(\\lambda_1 = 0.04\\), \\(\\lambda_2 = 0.01\\), and \\(\\lambda_3 = 0.01\\) across our experiments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.64, + 0.646, + 0.657 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.731, + 0.683 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We implement EnvGS with custom OptiX kernels and optimize our model using the PyTorch framework [32, 47] with the Adam optimizer [19]. Specifically, we set the learning rates for the parameters of each base and environment Gaussian to match those used in 2DGS [13]. The learning rate for the blending weight is set to \\(1e^{-2}\\). During training, we apply the adaptive Gaussian control strategy of 3DGS [17] with the normal propagation and color sabotage introduced in 3DGS-DR [18]. Since our Gaussian tracer integrates the Gaussian properties directly in 3D space, there is no valid gradient for the projected 2D center, which is used as the densification criterion in 3DGS. Following [29], we accumulate the 3D spatial gradients of the Gaussian position to achieve a similar effect. Note that each accumulated gra" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.918, + 0.493, + 0.929 + ], + "angle": 0, + "content": "5" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.089, + 0.851, + 0.249 + ], + "angle": 0, + "content": "
MethodsRef-NeRF Real Scenes [37]NeRF-Casting Scenes [38]FPS↑Training Time↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Non real-timeRef-NeRF* [37]23.0870.6250.26130.5830.8900.124<0.147h
UniSDF [39]23.7000.6350.26630.8380.8890.130<0.1>47h
ZipNeRF [4]23.6770.6350.24731.7400.9040.105<0.1>47h
NeRF-Casting [38]24.6700.6590.24631.0230.8890.128<0.1>47h
Real-time3DGS [17]23.7000.6410.26228.8600.8770.159182.3070.6h
2DGS [13]23.8040.6540.28128.2760.8620.193159.1880.7h
GaussianShader [14]22.8750.6220.31426.4120.8350.21627.9451.6h
3DGS-DR [51]23.5220.6400.27428.4870.8580.197133.5931.0h
Ours24.6170.6710.24130.4440.8860.14826.2212.5h
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.26, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on Ref-Real [37] and NeRF-Casting Scenes [38] datasets. Our method delivers the highest rendering quality among real-time techniques and outperforms several non-real-time methods, achieving competitive results with the state-of-the-art non-real-time method NeRF-Casting [38], while being 100 times faster. Note that Ref-NeRF* is an improved version of Ref-NeRF [37] that uses Zip-NeRF's geometry model. All metrics are evaluated at 1/4 resolution as prior works [38]." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.329, + 0.462, + 0.424 + ], + "angle": 0, + "content": "
MethodsForegroundNear-Field
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [17]31.6810.9550.04644.1610.9940.008
2DGS [13]31.3280.9530.05343.6240.9940.010
GaussianShader [14]30.6940.9470.05842.3910.9930.010
3DGS-DR31.8140.9530.05044.0070.9940.009
Ours33.2950.9620.04046.3920.9960.007
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Table 2. Quantitative results of foreground and near-field region on Ref-Real [37] and NeRF-Casting Scenes [38]. See supplementary Appendix A.1 for more details and qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.53 + ], + "angle": 0, + "content": "dient is scaled by half of the intersection depth to prevent under-densification in distant regions. All experiments are conducted on a single NVIDIA RTX 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.541, + 0.378, + 0.556 + ], + "angle": 0, + "content": "5.2. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.76 + ], + "angle": 0, + "content": "We train and evaluate EnvGS on a range of datasets with a focus on real-world scenes characterized by complex view-dependent effects. We evaluate the Ref-Real [37] and NeRF-Casting Shiny Scenes [38] to demonstrate our method's ability for complex specular reflections in real-world scenes. We additionally captured two more real-world scenes for a more comprehensive evaluation. We demonstrate that our methods can reconstruct detailed reflections on complex real-world scenes with real-time rendering speed. Additionally, we evaluate our method on the synthetic Shiny Blender dataset [37], which is rendered using environment maps. More evaluation results can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.822 + ], + "angle": 0, + "content": "We maintain consistent training and testing splits and image resolution across all datasets, following prior works [3, 17, 37]. We use three commonly used metrics for evaluation: PSNR, SSIM [43], and LPIPS [53]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.299, + 0.849 + ], + "angle": 0, + "content": "5.3. Baseline Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare our method with both implicit and explicit prior works, including Zip-NeRF [13], 3DGS [17], 2DGS [13], which are designed for general scenes with pri" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.329, + 0.885, + 0.433 + ], + "angle": 0, + "content": "
MethodsAudiDog
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Ref-NeRF [37]24.5290.7130.39723.6200.6730.389
3DGS [17]26.1710.8250.18125.3000.8820.165
2DGS [13]25.8690.8190.19824.9330.8760.184
GaussianShader [14]24.7680.7950.22523.0610.8400.226
3DGS-DR [51]26.0930.8200.19223.9430.8500.210
Ours28.6290.8760.12125.5990.8850.151
" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.443, + 0.888, + 0.457 + ], + "angle": 0, + "content": "Table 3. Quantitative results on our self-captured scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.547 + ], + "angle": 0, + "content": "marily diffuse appearances, as well as with methods specifically tailored for scenes with strong specular reflections, including Ref-NeRF [37], GaussianShader [14], 3DGS-DR [18], and NeRF-Casting [38]. We also compare with ENVIDR [22], NDE [45] these two object-level methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.653 + ], + "angle": 0, + "content": "In this section, we present quantitative and qualitative results to illustrate the advantages of our methods. Since accurate reflection reconstruction and rendering are the essential advantages of our method, we encourage readers to refer to the various rendered continuous video results in the supplementary material for a more comprehensive evaluation of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.838 + ], + "angle": 0, + "content": "We first evaluate our method on nine real-world scenes from the Ref-Real dataset [37], NeRF-Casting Shiny Scenes [38] and our self-captured real-world scenes, which feature complex geometry and specular reflections. The quantitative results, shown in Tab. 1 and Tab. 3, demonstrate that our method outperforms all explicit methods by a large margin and achieves comparable results to the current state-of-the-art implicit method, NeRF-Casting, while being significantly faster. The qualitative results in Fig.4 further highlight the superior performance of our method in capturing complex view-dependent effects, especially nearfield reflections and high-frequency details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We provide additional comparisons on the Mip-NeRF 360 dataset [3], the Shiny Blender dataset [37] and with ENVIDR [22], NDE [45] in supplementary Appendix A, which also includes per-scene metric breakdowns from Tab. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.918, + 0.516, + 0.929 + ], + "angle": 0, + "content": "5748" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.227, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.194, + 0.226, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.3, + 0.219, + 0.31 + ], + "angle": 0, + "content": "(a) w/o joint optimization" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.089, + 0.363, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.194, + 0.362, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.301, + 0.355, + 0.31 + ], + "angle": 0, + "content": "(b) w/o mono normal loss" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.089, + 0.498, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.194, + 0.498, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.301, + 0.486, + 0.311 + ], + "angle": 0, + "content": "(c) w/ environment map" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.634, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.194, + 0.634, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.301, + 0.63, + 0.311 + ], + "angle": 0, + "content": "(d) w/o normal propagation" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.769, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.194, + 0.769, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.301, + 0.721, + 0.311 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.089, + 0.906, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.194, + 0.905, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.301, + 0.867, + 0.31 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.323, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Figure 5. Ablation study of proposed components on the Ref-Real dataset [37]. Removing either the monocular normal constraint or the joint optimization of base and environment Gaussians results in noisy geometry and inaccurate reflection reconstruction. The \"w/ environment map\" variant fails to capture near-field reflections." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.425 + ], + "angle": 0, + "content": "Our method achieves competitive results on both datasets, demonstrating its versatility and effectiveness in handling diverse scenes with complex view-dependent effects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.255, + 0.449 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.458, + 0.462, + 0.554 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
w/o joint optimization24.0340.6440.287
w/o mono normal loss24.1070.6480.270
w/ environment map24.1450.6460.285
w/o color sabotage24.2680.6500.269
w/o normal propagation24.1920.6520.271
w/o lpips loss24.5670.6710.280
Ours24.6170.6710.241
" + }, + { + "type": "table_caption", + "bbox": [ + 0.206, + 0.564, + 0.368, + 0.577 + ], + "angle": 0, + "content": "Table 4. Ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.629 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies of our key components on the Ref-Real dataset. Quantitative and qualitative results are shown in Tab. 4 and Fig. 5, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Joint optimization. The \"w/o joint optimization\" variant detaches the joint optimization of the base Gaussian and environment Gaussian from the reflection rendering step. As shown in Tab. 4 and Fig. 5, This variant fails to recover accurate geometry, leading to inferior reflection reconstruction and rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Monocular normal constraint The \"w/o monocular normal\" variant removes the monocular normal constraint as described in Sec. 4.3. Training may become trapped in largely incorrect geometry, resulting in inaccurate reflection reconstruction, as demonstrated in Tab.4 and Fig.5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Environment map representation The \"w/ environment map\" variant replaces our core Gaussian environment representation with an environment map representation while keeping all other components unchanged. Figure 5 illustrates that while it effectively captures smooth distant reflections, it has difficulty modeling near-field and high-frequency reflections, and produces more bumpy geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Color sabotage and normal propagation. The “\\(w/o\\) color sabotage” and “\\(w/o\\) normal propagation” variants omit the color sabotage and normal propagation steps from our method, respectively. As demonstrated in Tab. 4, both variants result in reduced rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.53 + ], + "angle": 0, + "content": "Perceptual loss. The \"w/o lipps loss\" variants removes the perceptual loss. The results in Tab. 4 demonstrate that our method still produces detailed and accurate reflections without perceptual loss, which offers only marginal improvements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.762, + 0.56 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.907, + 0.766 + ], + "angle": 0, + "content": "This paper introduced EnvGS, a novel reflective scene representation for high-quality complex reflection capturing and real-time rendering. Our method explicitly models reflections with a set of environment Gaussian primitives. The environment Gaussian primitives are used together with a set of base Gaussian primitives that model basic scene properties (geometry, base color, and blending weight) to model the appearance of the whole scene. Furthermore, we develop a differentiable Gaussian ray tracer utilizing GPU's RT core to effectively optimize and render the environment Gaussian. The proposed method demonstrates superior performance in capturing complex reflections across various datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.768, + 0.906, + 0.827 + ], + "angle": 0, + "content": "A limitation of EnvGS is its difficulty with transparent and refractive materials, as it only addresses reflection direction. Future work could explore extending our method to accommodate these materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This work was partially supported by NSFC (No. U24B20154, 62402427), Ant Group, and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.918, + 0.516, + 0.929 + ], + "angle": 0, + "content": "5749" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7, 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2, 7, 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[5] Zoubin Bi, Yixin Zeng, Chong Zeng, Fan Pei, Xiang Feng, Kun Zhou, and Hongzhi Wu. Gs3: Efficient relighting with triple gaussian splatting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[8] Chen Guangcheng, He Yicheng, He Li, and Zhang Hong. Pisr: Polarimetric neural implicit surface reconstruction for textureless and specular objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[9] Yang Hang, Chen Rui, An Shipeng, Wei Hao, and Zhang Heng. The growth of image-related three dimensional reconstruction techniques in deep learning-driven era: a critical summary, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.566, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[10] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[11] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[12] Fuyuan Hu, Chenlu Li, Tao Zhou, Hongfu Cheng, and Minming Gu. A survey on point cloud completion algorithms for deep learning, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 3, 7, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for re" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.134 + ], + "angle": 0, + "content": "flective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 2, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[16] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 3, 5, 6, 7, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.904, + 0.262 + ], + "angle": 0, + "content": "[18] Ye Keyang, Hou Qiming, and Zhou Kun. 3d gaussian splatting with deferred reflection. 2024. 3, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.904, + 0.291 + ], + "angle": 0, + "content": "[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.904, + 0.333 + ], + "angle": 0, + "content": "[20] Yingqun Li, Xiao Hu, Xiang Xu, Yanning Xu, and Lu Wang. Deep learning-based foveated rendering in 3d space: a review, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.336, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[22] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.464, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[23] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[24] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[26] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[27] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[29] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian Ray Tracing: Fast tracing of particle scenes. ACM Transactions on Graphics and SIGGRAPH Asia, 2024. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.918, + 0.516, + 0.929 + ], + "angle": 0, + "content": "5750" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[31] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. Acm transactions on graphics (tog), 29(4):1-13, 2010. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[35] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[36] Zhe Jun Tang and Tat-Jen Cham. 3igs: Factorised tensorial illumination for 3d gaussian splatting. arXiv preprint arXiv:2408.03753, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3, 7, 8, 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[38] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2, 7, 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[39] Fangjinhua Wang, Marie-Julie Rakotosaona, Michael Niemeyer, Richard Szeliski, Marc Pollefeys, and Federico Tombari. Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In NeurIPS, 2024. 2, 7, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[40] Jinke Wang, Xingxing Zuo, Xiangrui Zhao, Jiajun Lyu, and Yong Liu. Review of multi-source fusion slam: current status and challenges, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[41] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[42] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Kaspar Rohrer, Norbert Felber, et al. A hardware architecture for surface splattering. ACM Transactions on Graphics (TOG), 26(3):90-es, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[45] Liwen Wu, Sai Bi, Zexiang Xu, Fujun Luan, Kai Zhang, Iliyan Georgiev, Kalyan Sunkavalli, and Ravi Ramamoorthi. Neural directional encoding for efficient and accurate view-dependent appearance modeling. In CVPR, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[46] Jiankai Xing and Kun Xu. Physically based differentiable rendering: a survey, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[47] Zhen Xu, Tao Xie, Sida Peng, Haotong Lin, Qing Shuai, Zhiyuan Yu, Guangzhao He, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Easyvolcap: Accelerating neural volumetric video research. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[48] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[49] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[50] Hanqiao Ye, Yangdong Liu, and Shuhan Shen. Lightweight visual-based localization technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2, 4, 7, 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[52] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[54] Youjia Zhang, Anpei Chen, Yumin Wan,zikai Song, Junqing Yu, Yawei Luo, and Wei Yang. Ref-gs: Directional factorization for 2d gaussian splatting. arXiv preprint arXiv:2412.00905, 2024.3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.918, + 0.514, + 0.929 + ], + "angle": 0, + "content": "5751" + } + ] +] \ No newline at end of file diff --git a/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_origin.pdf b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a6f6855d4b092c915cfd7dc8153bd89a547e79c --- /dev/null +++ b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/9aaa52c4-5f11-4ae5-bb15-8f551966ee84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0ac2c7d8e9098e0be139fde4eeecba7d4cd00eae3946dacc86b8e494ccbf6ff +size 7305925 diff --git a/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/full.md b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dfb44d68e8acf05a230f697824d2be5d76614baa --- /dev/null +++ b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/full.md @@ -0,0 +1,357 @@ +# EnvGS: Modeling View-Dependent Appearance with Environment Gaussian + +Tao Xie $^{1*}$ Xi Chen $^{1*}$ Zhen Xu $^{1}$ Yiman Xie $^{1}$ Yudong Jin $^{1}$ +Yujun Shen $^{2}$ Sida Peng $^{1}$ Hujun Bao $^{1}$ Xiaowei Zhou $^{1\dagger}$ $^{1}$ Zhejiang University $^{2}$ Ant Group + +![](images/ef51f2d8cb2ecb891db9ae1c313867dde00bea90d64b0eaea1d54eb5592f1056.jpg) + +![](images/754904581540908f9125dff0d9873a7293fc4ab8ae1a7c3aa6ebd0947b9ccf00.jpg) +Ground Truth + +![](images/de8f59eb61c19fe1e1bd60991246a9b5e2ecb1771c119b23b8d7588257dc59a3.jpg) + +![](images/0d068813dbc458fd714b7c29c1f672c88bb1fdfba19f34b96155c0cc1a20c2b1.jpg) +Ours + +![](images/5d64447eab1b2c70c546c3b69e6dff5f36957dc0af6fcc88757097a62d9f9f34.jpg) + +![](images/253ea988eee46a5bcc71352bc6d85289988258e8098cb2daffd0163602d21180.jpg) +GaussianShader +Figure 1. Photorealistic, real-time rendering of real-world scenes with complex reflections. Our proposed EnvGS outperforms prior works in capturing complex reflection effects, especially near-field reflections and high-frequency details while maintaining real-time rendering speed. Please see our supplementary video for better visualizations. + +![](images/31d77fbb1d6414468a9c87bbf76595f2876979faf7a5562e52e2b93c7d341706.jpg) + +![](images/0fb401364a6b68b5f3f12b9f947c28c40cde6d81f9a8627eb91a0689fb5ff42f.jpg) +3DGS + +# Abstract + +Reconstructing complex reflections in real-world scenes from 2D images is essential for achieving photorealistic novel view synthesis. Existing methods that utilize environment maps to model reflections from distant lighting often struggle with high-frequency reflection details and fail to account for near-field reflections. In this work, we introduce EnvGS, a novel approach that employs a set of Gaussian primitives as an explicit 3D representation for capturing reflections of environments. These environment Gaussian primitives are incorporated with base Gaussian primitives to model the appearance of the whole scene. To efficiently render these environment Gaussian primitives, we developed a ray-tracing-based renderer that leverages the GPU's RT core for fast rendering. This allows us to jointly optimize our model for high-quality reconstruction while maintain- + +ing real-time rendering speeds. Results from multiple realworld and synthetic datasets demonstrate that our method produces significantly more detailed reflections, achieving the best rendering quality in real-time novel view synthesis. The code is available at https://zju3dv.github.io/envgs. + +# 1. Introduction + +Novel view synthesis aims to generate novel views of 3D scenes based on a set of input images, which enables many applications such as VR/AR, and autonomous driving. Recent advances in Neural Radiance Fields (NeRF) [28] have demonstrated impressive rendering performance. However, NeRF's high computational cost makes it challenging for real-time applications. More recently, 3D Gaussian Splatting (3DGS) [17] explicitly models scenes with 3D Gaussian primitives and utilizes rasterization for rendering, achieving real-time rendering with competitive quality. However, modeling complex high-frequency specular + +reflections remains challenging for 3DGS due to the limited expressiveness of the Spherical Harmonics (SH). + +Recent works GaussianShader [15] and 3DGS-DR [51], enhance 3DGS by integrating an environment map and employing shading functions to blend the appearance from both the environment map and SH for the final rendering. While additional environmental lighting can enhance 3DGS's reflection modeling ability, it still struggles to reconstruct complex specular reflections accurately due to two factors. First, the assumption of distant lighting in environment maps limits their ability to only capture distant illumination and difficult to synthesize accurate near-field reflections. Second, this representation inherently lacks sufficient capacity to capture high-frequency reflection details. + +In this paper, we present EnvGS, a novel approach for modeling complex reflections in real-world scenes, addressing the aforementioned challenges. We propose to model reflections using a set of Gaussian primitives termed the environment Gaussian. The geometry and base appearance are represented by another set of Gaussian primitives called base Gaussian. We effectively blend the two Gaussians for rendering and optimization. Our rendering process begins with rendering the base Gaussian for the per-pixel surface position, normal, base color, and blending weight. Next, we render the environment Gaussian at the surface point in the direction of the reflection of the viewing direction around the surface normal to capture reflection colors. Finally, we blend the base color with the reflection color to achieve the final rendering results. In contrast to previous methods, EnvGS captures high-frequency reflection details using Gaussian primitives, offering superior modeling capabilities compared to environment maps. Additionally, our explicit 3D reflection representation eliminates the need for distant lighting assumptions, enabling accurate modeling of near-field reflections, as shown in Fig. 1. + +To render the environment Gaussian at each intersection point along the reflection direction, we create a fully differentiable ray-tracing renderer for 2DGS since rasterization is not suited for this task. We build the ray-tracing renderer on CUDA and OptiX [31] for real-time rendering and efficient optimization of environment Gaussian. The rendering process starts by constructing a bounding volume hierarchy (BVH) from the 2D Gaussian primitives. We then cast rays against the BVH, gathering ordered intersections in chunks while integrating the Gaussian properties through volume rendering [16] to achieve the final results. Our Gaussian ray-tracing renderer enables detailed reflection rendering at real-time performance. Furthermore, it allows for efficient joint optimization of the environment Gaussian and base Gaussian, which is essential for accurate reflection modeling, as demonstrated in Sec 5.4. + +To validate the effectiveness of our method, we evaluate EnvGS on several real and synthetic datasets. The re + +sults demonstrate that our methods achieve state-of-the-art performance in real-time novel view synthesis and considerably surpass existing real-time methods, particularly in synthesizing complex reflections in real-world scenes. + +In summary, we make the following contributions: + +- We propose a novel scene representation for accurately modeling complex near-field and high-frequency reflections in real-world environments. +- We developed a real-time ray-tracing renderer for 2DGS, enabling joint optimization of our representation for accurate scene reconstruction while achieving real-time rendering speeds. +- Extensive experiments shows that EnvGS significantly outperforms previous methods. To the best of our knowledge, EnvGS is the first method that achieves real-time photorealistic specular reflections synthesizing in real-world scenes. + +# 2. Related Work + +In computer vision and graphics research [9, 12, 20, 23, 40, 46, 50], our work falls into the area of learning scene representations from a set of posed RGB images. In this section, we review NeRF-based and Gaussian Splitting-based methods, particularly their handling of view-dependent effects. + +Neural Radiance Field. NeRF [28] introduced the concept of neural radiance fields, which model scenes as implicit multilayer perceptrons (MLPs) and render them via volume rendering, achieving impressive results in novel view synthesis. Subsequent advancements focused on enhancing rendering quality [2-4] and computational efficiency. [6, 7, 24, 30, 35]. [21, 41, 42, 48] introduce Signed Distance Field into NeRF to improve geometry quality. However, these methods often model view-dependent effects via simple viewing directions, which can lead to blurry reflection renderings. To address this, Ref-NeRF [37] encodes the outgoing radiance using the reflected view direction, yielding improved results under distant lighting conditions. Follow-up works [8, 22, 25, 39] leverage Signed Distance Fields (SDF) to refine surface normals to enhance reflection and geometry quality. SpecNeRF [27] further incorporates spatially varying Gaussian directional encoding to better capture near-field reflections. NeRF-Casting [38] represents the scene in a unified manner, similar to [4], and performs ray marching along reflection directions to integrate reflection features, which are then decoded into color with MLPs. This approach achieves impressive results on real-world data under both near and distant lighting conditions. However, the requirement for multiple MLP queries per ray makes NeRF-Casting unsuitable for real-time rendering and necessitates substantial training time. In contrast, our approach delivers real-time rendering capabilities while significantly reducing training time. + +Gaussian Splatting. Recently, 3D Gaussian Splatting (3DGS) [17] has made significant strides toward real-time novel view synthesis. Unlike volume-based rendering, Gaussian Splatting uses efficient rasterization to render spatial Gaussian kernels, achieving real-time performance at 1080P resolution. Mip-Splatting [52] further introduces a 3D smoothing filter and a 2D Mip filter for alias-free renderings. Scaffold-GS [26] introduces structured 3D Gaussians to improve rendering efficiency and quality. [5] introduces triple splitting for efficient relighting. 2D Gaussian Splatting (2DGS) [13] replaces the 3D Gaussian kernels with 2D Gaussian that better align with 3D surfaces, leading to more accurate surface reconstruction. However, these methods face challenges in accurately modeling reflections due to their use of spherical harmonics (SH) to parameterize view-dependent effects based solely on viewing direction, which often results in blurry reflections. More recent works GaussianShader [14] and 3DGS-DR [18] try to incorporate additional environment maps to improve reflection modeling ability. However, these methods only consider distant lighting, ignoring near-field lighting, and are unable to capture high-frequency reflection details. 3iGS [36] extends 3DGS by augmenting it with an illumination field using tensorial factorization, rendering final reflections through a neural renderer. However, it is limited to bounded scenes, which restricts its applicability in real-world, unbounded environments. Concurrent work [54] choose to model near-field reflection with tensorial factorization and far-field with spherical feature grid. Our method adopts 2DGS as the scene representation and models environmental illumination using an additional set of environment Gaussian, which is rendered via our proposed Gaussian tracer. Our approach inherently supports high-frequency, near-field, and distant lighting in unbounded scenes, enabling detailed reflection rendering while maintaining real-time performance. + +# 3. Preliminary + +We begin by introducing 2D Gaussian Splatting (2DGS) [13], which our approach is built upon. 2DGS is an explicit scene representation similar to 3D Gaussian Splatting (3DGS) [17], which uses Gaussian primitives and rasterization to render screen projections. The key difference is that 2DGS represents the scene with a set of scaled 2D Gaussian primitives defined in a local tangent plane within world space by a transformation matrix $\mathbf{H}$ : + +$$ +\mathbf {H} = \left[ \begin{array}{c c c c} s _ {u} \mathbf {t} _ {u} & s _ {v} \mathbf {t} _ {v} & 0 & \mathbf {p} _ {k} \\ 0 & 0 & 0 & 1 \end{array} \right], \tag {1} +$$ + +where $\mathbf{p}_k, (\mathbf{t}_u, \mathbf{t}_v)$ , and $(s_u, s_v)$ denote the center, the two principal tangential vectors, and the scaling factors of the Gaussian, respectively. + +To render an image, 2DGS employs the method described in [44] to determine the ray-primitive intersections. + +These intersection points are subsequently utilized to compute the Gaussian's contribution to the final image. The Gaussian properties are then integrated using a volume rendering algorithm to obtain the final pixel color: + +$$ +\mathbf {c} = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \mathbf {c} _ {i}, \text {w i t h} \alpha_ {i} = \sigma_ {i} \mathcal {G} _ {i}, T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {2} +$$ + +where $\mathcal{G}(\cdot)$ is the standard 2D Gaussian value evaluation. + +Compared to 3D Gaussian, 2D Gaussian offers distinct advantages as a surface representation. First, the ray-splat intersection method adopted by 2DGS avoids multi-view depth inconsistency. Second, 2D Gaussian inherently provides a well-defined normal, which is essential for high-quality surface reconstruction and accurate reflection calculations. However, 2DGS relies on the limited representational capacity of Spherical Harmonics (SH) to model scene appearance, preventing it from capturing strong view-dependent effects like specular reflections, which leads to poor rendering results and "foggy" geometry [37]. To this end, we use the geometry-aligned 2D Gaussian primitives as the base scene representation and demonstrate how we effectively model complex reflections in the next section. + +# 4. Method + +Given a set of input images of a reflective scene, our goal is to reconstruct the 3D scene and synthesize photorealistic novel views in real-time. To achieve this, we propose utilizing environment Gaussian as an explicit 3D environment representation, which enables accurate modeling of complex reflections in real-world scenes. Additionally, we represent the scene geometry and base colors using another set of Gaussian primitives, denoted as base Gaussian. In this section, we first detail how environment Gaussian and base Gaussian work together to model complex reflections within the scene (Sec. 4.1). Then, We describe the design of a ray-tracing renderer that leverages the GPU's RT cores to efficiently render and optimize the environment Gaussian (Sec. 4.2). Finally, we discuss our optimization process. (Sec. 4.3). An overview of our method is in Fig. 2. + +# 4.1. Reflective Scenes Modeling + +Gaussian splatting [13, 17] models appearance using Spherical Harmonics (SH), which has limited representation capacity for view-dependent effects. These limitations hinder its ability to capture complex, high-frequency specular reflections. Building on this observation, our key insight is that modeling reflections with a Gaussian environment representation can better model complex reflection effects while significantly reducing the complexity required for each Gaussian to capture intricate details within its SH. + +Our proposed reflective scene representation includes two sets of 2D Gaussians: a base Gaussian $\mathbf{P}_{\text {base }}$ for mod- + +![](images/7d6ef788c09d446ce552872609a528861b934fe5dbe8fc04d819622b0a4ae8b2.jpg) +Figure 2. Overview of EnvGS. The rendering process begins by rasterizing the base Gaussian to obtain per-pixel normals, base colors, and blending weights. Next, we render the environment Gaussian in the reflection direction using our ray-tracing-based Gaussian renderer to capture the reflection colors. Finally, we combine the reflection and base colors for the final output. We jointly optimize the environment Gaussian and base Gaussian using monocular normals [49] and ground truth images for supervision. + +eling the scene's geometry and base appearance, and another environment Gaussian $\mathbf{P}_{env}$ for capturing scene reflections. The basic parameterization of each Gaussian primitive is consistent with the original 2DGS [13], including 3D center position $\mathbf{p}$ , opacity $\alpha$ , two principal tangential vectors $(\mathbf{t}_u,\mathbf{t}_v)$ , a scaling vector $(s_u,s_v)$ , and SH coefficients. To combine the two appearance components from the base and environment Gaussian into the final result, we introduce a blending weight for each base Gaussian. + +The rendering process is performed in three steps. First, the base Gaussian $\mathbf{P}_{base}$ is rendered using standard 2D Gaussian splating to obtain the base color $\mathbf{c}_{base}$ . By applying the volume rendering integration using Eq. (2), we also derive the surface position $\mathbf{x}$ , surface normals $\mathbf{n}$ , and blending weight $\beta$ as: + +$$ +v = \sum_ {i \in \mathcal {N}} = v _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), v \in \{\mathbf {x}, \mathbf {n}, \beta \}. \tag {3} +$$ + +Then, we compute the reflection direction $\mathbf{d}_{ref}$ based on the camera ray direction $\mathbf{d}_{ref}$ and the surface normal $\mathbf{n}$ : + +$$ +\mathbf {d} _ {r e f} = \mathbf {d} _ {c a m} - 2 \left(\mathbf {d} _ {c a m} \cdot \mathbf {n}\right) \mathbf {n}. \tag {4} +$$ + +With the reflection direction $\mathbf{d}_{ref}$ and surface point $\mathbf{x}$ , the environment Gaussian $\mathbf{P}_{env}$ is rendered using our differentiable Gaussian tracer to obtain the reflection color $\mathbf{c}_{ref}$ , as detailed in Sec. 4.2. The final color is obtained through: + +$$ +\mathbf {c} = (1 - \beta) \cdot \mathbf {c} _ {\text {b a s e}} + \beta \cdot \mathbf {c} _ {\text {r e f}}. \tag {5} +$$ + +We blend the base color $\mathbf{c}_{\text {base }}$ and reflection color $\mathbf{c}_{\text {ref }}$ using the blending weight $\beta$ . A visualization of $\mathbf{c}_{\text {base }}$ and $\mathbf{c}_{\text {ref }}$ can be found at Fig. 3. + +Discussion. Compared to the environment map used by Gaussianshader [15] and 3DGS-DR [51], our explicit Gaussian environment representation offers several advantages. First, our method more accurately captures near-field reflections caused by occlusions from nearby objects. This improvement arises from explicitly modeling each Gaussian at its exact spatial location, thus avoiding the ambiguities and inaccuracies inherent in environment map representations that assume distant lighting. Second, by utilizing Gaussian primitives, EnvGS's environment representation achieves greater expressiveness than low-frequency environment maps, enabling the capture of finer reflection details and enhancing rendering quality, as demonstrated by our experiments 5.3. + +# 4.2. Differentiable Ray Tracing + +Rendering the environment Gaussian with rasterization is impractical, as each pixel corresponds to a unique reflection ray and functions as a virtual camera. To address this, we draw on [29] and leverage the advanced optimizations of modern GPUs to design a novel, fully differentiable ray tracing framework. Built on OptiX [31], our framework achieves real-time rendering of 2,000,000 2DGS with a resolution of $1292 \times 839$ at 30 FPS on an RTX 4090 GPU. + +In order to fully utilize the hardware acceleration for + +![](images/6610bdc1bd29c6718c6c8d304052f682369efc8ac1da941cae368d6835f27f03.jpg) + +![](images/9e1fa2929b6978e536f1079f62e079735155deb5a1bc578c2c28b2d38b887609.jpg) + +![](images/20ba3b7f2c8ab890a2b571595430755959852b3defd4ede0137bb51eef4b33be.jpg) + +![](images/95bc2f61d1ec65396e5bebab5b3d0309205e404d39c9e0383db6e8f053b4c262.jpg) +Final color + +![](images/c87d914cfdc63e235d311305e62b71b462c67534de245ddafd82eb6e74993042.jpg) +Base color + +![](images/9d9933fcf429055170f5ddf867258a2964ce7ff7e8d1437e771b72487167bdc8.jpg) +Reflection color +Figure 3. Visualization of reflection and base color. Our method successfully reconstructs near-field and distant reflections using the environment Gaussian instead of baking into the base color. + +ray-primitive intersections, we need to convert each 2D Gaussian into a geometric primitive compatible with GPU processing and insert it into a bounding volume hierarchy (BVH). In light of this, we propose to represent each 2D Gaussian with two triangles. Specifically, we first define the four Gaussian bounding vertices $\mathbf{V}_{local} = \{(\mathrm{sgn}(r),\mathrm{sgn}(r))\}$ in the local tangent plane, where $\mathrm{sgn}(\cdot)$ is the sign function and $r$ is set to 3 representing three times the sigma range. Then, the four local bounding vertices $\mathbf{V}_{local}$ are transformed to world space as the vertices of the two triangles covering the Gaussian $\mathbf{V}_{world}$ using Eq. 1. After the transformation, the triangles are organized into a BVH, which serves as the input for the ray tracing process. + +We develop a custom CUDA kernel using the raygen and anyhit programmable entry points of OptiX. Inspired by [29], the rendering is done in a chunk-by-chunk manner. The anyhit kernel traces the input ray to obtain a chunk of size $\mathrm{k}$ , while raygen integrates this chunk and invokes anyhit to retrieve the next chunk along the ray. Specifically, consider an input ray with origin $\mathbf{o}$ and direction $\mathbf{d}$ . The raygen program first initiates a traversal against the BVH to identify all possible intersections along the ray. During the traversal, the anyhit program sorts each intersected Gaussian by the depth and maintains a sorted k-buffer for the closest $\mathrm{k}$ intersections. We empirically found that $\mathrm{k}$ with 16 is the best trade-off between traversal counts and the number of Gaussians sorted per traversal. After traversal, the raygen program integrates properties of the sorted Gaussians in the buffer following Eq. 2. The Gaussian response is calculated by applying the inverse of the transformation matrix $\mathbf{H}$ to the ray's intersection point $\mathbf{x}_i$ and evaluating the Gaussian value at the transformed point as: + +$$ +\mathcal {G} _ {i} \left(\mathbf {u} _ {i}\right) = \mathcal {G} _ {i} \left(\mathbf {H} ^ {- 1} \mathbf {x} _ {i}\right). \tag {6} +$$ + +This process repeats until no further intersections are found + +along the ray or the accumulated transmittance drops below a specified threshold. More details can be found in supplementary materials. + +Our ray tracing framework is fully differentiable, allowing for end-to-end optimization of both the base and the environment Gaussian primitives. However, storing all intersections during the forward pass and performing the backward pass in back-to-front order, as done in 3DGS [17], is impractical due to high memory consumption. To address this, we implement the backward pass in the same front-to-back order as the forward pass by re-casting rays and computing gradients for each integration step. A key aspect is calculating the gradient with respect to the input ray origin $\frac{\partial\mathcal{L}}{\partial\mathbf{o}}$ and direction $\frac{\partial\mathcal{L}}{\partial\mathbf{d}}$ , which is crucial for the joint optimization of our model (see supplementary for more details). Our ablation in Sec 5.4 demonstrates that this joint optimization of geometry and appearance is essential for recovering geometrically accurate surfaces. + +# 4.3. Optimization + +To enhance training stability, we initiate optimization by first training the base Gaussian $\mathbf{P}_{\text {base }}$ , which is initialized using the sparse point cloud obtained from Structure-from-Motion (SfM) [11, 34]. After the bootstrapping phase, we initialize the environment Gaussian $\mathbf{P}_{env}$ by partitioning the scene's bounding box $\mathbf{B}_{\text {scene }}$ into $N^3$ sub-grids and randomly sampling $K$ primitives within each grid, and then optimize the base Gaussian and environment Gaussian jointly. The scene's bounding box $\mathbf{B}_{\text {scene }}$ is determined by the $99.5\%$ quantile of the sparse point cloud's bounding box $\mathbf{B}_{\text {sfm }}$ obtained from SfM, the sub-grids resolution $N$ is set to 32, and $K$ is set to 5 for each grid. + +We follow 2DGS [13] to add a normal consistency constraint between the rendered normal map $\mathbf{n}$ and the gradients of the depth map $\mathbf{N}_d$ : + +$$ +\mathcal {L} _ {\text {n o r m}} = \frac {1}{N _ {p}} \sum_ {i = 1} ^ {N _ {p}} \left(1 - \mathbf {n} _ {i} ^ {\top} \mathbf {N} _ {d}\right), \tag {7} +$$ + +where $N_{p}$ is the number of pixels in the image and $\mathbf{N}_d$ is calculated as finite differences of neighboring pixels in the depth map: + +$$ +\mathbf {N} _ {d} (\mathbf {u}) = \frac {\nabla_ {u} \mathbf {p} _ {d} \times \nabla_ {v} \mathbf {p} _ {d}}{\| \nabla_ {u} \mathbf {p} _ {d} \times \nabla_ {v} \mathbf {p} _ {d} \|}. \tag {8} +$$ + +However, the normal consistency constraint alone is insufficient for accurately modeling ambiguous surfaces involving both reflection and refraction. Inspired by recent advances in monocular normal estimation [1, 10, 49], we propose to supervise the rendered normal map $\mathbf{n}$ using monocular normal estimates $\mathbf{N}_m$ : + +$$ +\mathcal {L} _ {\text {m o n o}} = \frac {1}{N _ {p}} \sum_ {i = 1} ^ {N _ {p}} \left(1 - \mathbf {n} _ {i} ^ {\top} \mathbf {N} _ {m}\right). \tag {9} +$$ + +![](images/f50060daa754012bb7e24a1d4f011b8219ba9e6a428731fd662082ea646a4023.jpg) +Figure 4. Qualitative comparison on real scenes. Our method significantly improves rendering quality over previous approaches, particularly in producing more detailed reflections. Zoom in for more details. + +In addition to the $\mathcal{L}_1$ image loss and D-SSIM loss $\mathcal{L}$ employed by 3DGS [17], we also use perceptual loss [53] to enhance the perceived quality of the rendered image: + +$$ +\mathcal {L} _ {\text {p e r c}} = \left\| \Phi (\mathbf {I}) - \Phi \left(\mathbf {I} _ {g t}\right) \right\| _ {1}, \tag {10} +$$ + +where $\Phi$ is the pre-trained VGG-16 network [33] and $\mathbf{I},\mathbf{I}_{gt}$ are the rendered and ground truth images, respectively. + +The final loss function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {r g b} + \lambda_ {1} \mathcal {L} _ {\text {n o r m}} + \lambda_ {2} \mathcal {L} _ {\text {m o n o}} + \lambda_ {3} \mathcal {L} _ {\text {p e r c}}, \tag {11} +$$ + +where $\mathcal{L}_{rgb}$ is the photometric reconstruction loss combining the $\mathcal{L}_1$ loss and a D-SSIM [43] term with a ratio of 0.8 and 0.2 respectively. We set $\lambda_1 = 0.04$ , $\lambda_2 = 0.01$ , and $\lambda_3 = 0.01$ across our experiments. + +# 5. Experiments + +# 5.1. Implementation Details + +We implement EnvGS with custom OptiX kernels and optimize our model using the PyTorch framework [32, 47] with the Adam optimizer [19]. Specifically, we set the learning rates for the parameters of each base and environment Gaussian to match those used in 2DGS [13]. The learning rate for the blending weight is set to $1e^{-2}$ . During training, we apply the adaptive Gaussian control strategy of 3DGS [17] with the normal propagation and color sabotage introduced in 3DGS-DR [18]. Since our Gaussian tracer integrates the Gaussian properties directly in 3D space, there is no valid gradient for the projected 2D center, which is used as the densification criterion in 3DGS. Following [29], we accumulate the 3D spatial gradients of the Gaussian position to achieve a similar effect. Note that each accumulated gra + +
MethodsRef-NeRF Real Scenes [37]NeRF-Casting Scenes [38]FPS↑Training Time↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Non real-timeRef-NeRF* [37]23.0870.6250.26130.5830.8900.124<0.147h
UniSDF [39]23.7000.6350.26630.8380.8890.130<0.1>47h
ZipNeRF [4]23.6770.6350.24731.7400.9040.105<0.1>47h
NeRF-Casting [38]24.6700.6590.24631.0230.8890.128<0.1>47h
Real-time3DGS [17]23.7000.6410.26228.8600.8770.159182.3070.6h
2DGS [13]23.8040.6540.28128.2760.8620.193159.1880.7h
GaussianShader [14]22.8750.6220.31426.4120.8350.21627.9451.6h
3DGS-DR [51]23.5220.6400.27428.4870.8580.197133.5931.0h
Ours24.6170.6710.24130.4440.8860.14826.2212.5h
+ +Table 1. Quantitative comparison on Ref-Real [37] and NeRF-Casting Scenes [38] datasets. Our method delivers the highest rendering quality among real-time techniques and outperforms several non-real-time methods, achieving competitive results with the state-of-the-art non-real-time method NeRF-Casting [38], while being 100 times faster. Note that Ref-NeRF* is an improved version of Ref-NeRF [37] that uses Zip-NeRF's geometry model. All metrics are evaluated at 1/4 resolution as prior works [38]. + +
MethodsForegroundNear-Field
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [17]31.6810.9550.04644.1610.9940.008
2DGS [13]31.3280.9530.05343.6240.9940.010
GaussianShader [14]30.6940.9470.05842.3910.9930.010
3DGS-DR31.8140.9530.05044.0070.9940.009
Ours33.2950.9620.04046.3920.9960.007
+ +dient is scaled by half of the intersection depth to prevent under-densification in distant regions. All experiments are conducted on a single NVIDIA RTX 4090 GPU. + +# 5.2. Datasets and Evaluation Metrics + +We train and evaluate EnvGS on a range of datasets with a focus on real-world scenes characterized by complex view-dependent effects. We evaluate the Ref-Real [37] and NeRF-Casting Shiny Scenes [38] to demonstrate our method's ability for complex specular reflections in real-world scenes. We additionally captured two more real-world scenes for a more comprehensive evaluation. We demonstrate that our methods can reconstruct detailed reflections on complex real-world scenes with real-time rendering speed. Additionally, we evaluate our method on the synthetic Shiny Blender dataset [37], which is rendered using environment maps. More evaluation results can be found in the supplementary materials. + +We maintain consistent training and testing splits and image resolution across all datasets, following prior works [3, 17, 37]. We use three commonly used metrics for evaluation: PSNR, SSIM [43], and LPIPS [53]. + +# 5.3. Baseline Comparisons + +We compare our method with both implicit and explicit prior works, including Zip-NeRF [13], 3DGS [17], 2DGS [13], which are designed for general scenes with pri + +Table 2. Quantitative results of foreground and near-field region on Ref-Real [37] and NeRF-Casting Scenes [38]. See supplementary Appendix A.1 for more details and qualitative results. + +
MethodsAudiDog
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Ref-NeRF [37]24.5290.7130.39723.6200.6730.389
3DGS [17]26.1710.8250.18125.3000.8820.165
2DGS [13]25.8690.8190.19824.9330.8760.184
GaussianShader [14]24.7680.7950.22523.0610.8400.226
3DGS-DR [51]26.0930.8200.19223.9430.8500.210
Ours28.6290.8760.12125.5990.8850.151
+ +Table 3. Quantitative results on our self-captured scenes. + +marily diffuse appearances, as well as with methods specifically tailored for scenes with strong specular reflections, including Ref-NeRF [37], GaussianShader [14], 3DGS-DR [18], and NeRF-Casting [38]. We also compare with ENVIDR [22], NDE [45] these two object-level methods. + +In this section, we present quantitative and qualitative results to illustrate the advantages of our methods. Since accurate reflection reconstruction and rendering are the essential advantages of our method, we encourage readers to refer to the various rendered continuous video results in the supplementary material for a more comprehensive evaluation of our method. + +We first evaluate our method on nine real-world scenes from the Ref-Real dataset [37], NeRF-Casting Shiny Scenes [38] and our self-captured real-world scenes, which feature complex geometry and specular reflections. The quantitative results, shown in Tab. 1 and Tab. 3, demonstrate that our method outperforms all explicit methods by a large margin and achieves comparable results to the current state-of-the-art implicit method, NeRF-Casting, while being significantly faster. The qualitative results in Fig.4 further highlight the superior performance of our method in capturing complex view-dependent effects, especially nearfield reflections and high-frequency details. + +We provide additional comparisons on the Mip-NeRF 360 dataset [3], the Shiny Blender dataset [37] and with ENVIDR [22], NDE [45] in supplementary Appendix A, which also includes per-scene metric breakdowns from Tab. 1. + +![](images/c760e364bbce1deff1cd01d9d5e019dc22fc49f5b30caaa27c86c59ef5a62e4d.jpg) + +![](images/2791f0a2d6bee68f3e7acba85d3a66896169bc981aa2de7bc27b6de9dec373d4.jpg) +(a) w/o joint optimization + +![](images/fe0fde4400901a472eb394fcbcca998095ac1c6ab6f290cf0d2cb3bb93a71a5f.jpg) + +![](images/fb20b86d0890f29745968e343dd29d123ed0550111ab051385579c9b7bc19a8b.jpg) +(b) w/o mono normal loss + +![](images/31296fec1e6c85513c6d78699d8ee602da67f339fab9eba9c7efe7a97ee941c8.jpg) + +![](images/a1481b2a5b7bebafef9ac43e119dc91ba90c4c478f6636edcc4bc339525cc4d5.jpg) +(c) w/ environment map +Figure 5. Ablation study of proposed components on the Ref-Real dataset [37]. Removing either the monocular normal constraint or the joint optimization of base and environment Gaussians results in noisy geometry and inaccurate reflection reconstruction. The "w/ environment map" variant fails to capture near-field reflections. + +![](images/29ed4b41adca4e473701762271b7f73583c3b7ccc6ee4e45ed67145333d0bb44.jpg) + +![](images/766b3bc215f9be33c0ba73681bd6d72d0ba4c51cc66a9ea7835fd2f94636aecf.jpg) +(d) w/o normal propagation + +![](images/0b627fda375fbb21841a9d8a7dc1502d3c9427fea7ba2a43c039ec39c201900c.jpg) + +![](images/4e2577d09bd87237be35a1e685f52d1ebd981e27f3e83c787d16419672e887a7.jpg) +(e) Ours + +![](images/c5ecd83fb96bc13565b89964d1524297ab984bf776cdec2c07136025e200b7de.jpg) + +![](images/d03ef3e2689e95fb639a8a29b94bf0e01b71f3418f6506a1eefdf09650ee86dc.jpg) +Ground Truth + +Our method achieves competitive results on both datasets, demonstrating its versatility and effectiveness in handling diverse scenes with complex view-dependent effects. + +# 5.4. Ablation Studies + +
PSNR ↑SSIM ↑LPIPS ↓
w/o joint optimization24.0340.6440.287
w/o mono normal loss24.1070.6480.270
w/ environment map24.1450.6460.285
w/o color sabotage24.2680.6500.269
w/o normal propagation24.1920.6520.271
w/o lpips loss24.5670.6710.280
Ours24.6170.6710.241
+ +Table 4. Ablation studies. + +In this section, we conduct ablation studies of our key components on the Ref-Real dataset. Quantitative and qualitative results are shown in Tab. 4 and Fig. 5, respectively. + +Joint optimization. The "w/o joint optimization" variant detaches the joint optimization of the base Gaussian and environment Gaussian from the reflection rendering step. As shown in Tab. 4 and Fig. 5, This variant fails to recover accurate geometry, leading to inferior reflection reconstruction and rendering quality. + +Monocular normal constraint The "w/o monocular normal" variant removes the monocular normal constraint as described in Sec. 4.3. Training may become trapped in largely incorrect geometry, resulting in inaccurate reflection reconstruction, as demonstrated in Tab.4 and Fig.5. + +Environment map representation The "w/ environment map" variant replaces our core Gaussian environment representation with an environment map representation while keeping all other components unchanged. Figure 5 illustrates that while it effectively captures smooth distant reflections, it has difficulty modeling near-field and high-frequency reflections, and produces more bumpy geometry. + +Color sabotage and normal propagation. The “ $w/o$ color sabotage” and “ $w/o$ normal propagation” variants omit the color sabotage and normal propagation steps from our method, respectively. As demonstrated in Tab. 4, both variants result in reduced rendering quality. + +Perceptual loss. The "w/o lipps loss" variants removes the perceptual loss. The results in Tab. 4 demonstrate that our method still produces detailed and accurate reflections without perceptual loss, which offers only marginal improvements. + +# 6. Conclusion and Discussion + +This paper introduced EnvGS, a novel reflective scene representation for high-quality complex reflection capturing and real-time rendering. Our method explicitly models reflections with a set of environment Gaussian primitives. The environment Gaussian primitives are used together with a set of base Gaussian primitives that model basic scene properties (geometry, base color, and blending weight) to model the appearance of the whole scene. Furthermore, we develop a differentiable Gaussian ray tracer utilizing GPU's RT core to effectively optimize and render the environment Gaussian. The proposed method demonstrates superior performance in capturing complex reflections across various datasets. + +A limitation of EnvGS is its difficulty with transparent and refractive materials, as it only addresses reflection direction. Future work could explore extending our method to accommodate these materials. + +Acknowledgements. This work was partially supported by NSFC (No. U24B20154, 62402427), Ant Group, and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5 +[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7, 1, 8 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2, 7, 1, 6, 8 +[5] Zoubin Bi, Yixin Zeng, Chong Zeng, Fan Pei, Xiang Feng, Kun Zhou, and Hongzhi Wu. Gs3: Efficient relighting with triple gaussian splatting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 3 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2 +[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[8] Chen Guangcheng, He Yicheng, He Li, and Zhang Hong. Pisr: Polarimetric neural implicit surface reconstruction for textureless and specular objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2 +[9] Yang Hang, Chen Rui, An Shipeng, Wei Hao, and Zhang Heng. The growth of image-related three dimensional reconstruction techniques in deep learning-driven era: a critical summary, 2023. 2 +[10] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 5 +[11] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 5 +[12] Fuyuan Hu, Chenlu Li, Tao Zhou, Hongfu Cheng, and Minming Gu. A survey on point cloud completion algorithms for deep learning, 2025. 2 +[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 4, 5, 6, 7, 8 +[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 3, 7, 6, 8 +[15] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for re + +flective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 2, 4, 1 +[16] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 3, 5, 6, 7, 2, 8 +[18] Ye Keyang, Hou Qiming, and Zhou Kun. 3d gaussian splatting with deferred reflection. 2024. 3, 6, 7, 1 +[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[20] Yingqun Li, Xiao Hu, Xiang Xu, Yanning Xu, and Lu Wang. Deep learning-based foveated rendering in 3d space: a review, 2024. 2 +[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[22] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 2, 7 +[23] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. 2025. 2 +[24] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2 +[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 2 +[26] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3 +[27] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections, 2023. 2 +[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[29] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian Ray Tracing: Fast tracing of particle scenes. ACM Transactions on Graphics and SIGGRAPH Asia, 2024. 4, 5, 6 +[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires + +olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[31] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. Acm transactions on graphics (tog), 29(4):1-13, 2010. 2, 4 +[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 5 +[35] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2 +[36] Zhe Jun Tang and Tat-Jen Cham. 3igs: Factorised tensorial illumination for 3d gaussian splatting. arXiv preprint arXiv:2408.03753, 2024. 3, 7 +[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3, 7, 8, 1, 6 +[38] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2, 7, 1, 6, 8 +[39] Fangjinhua Wang, Marie-Julie Rakotosaona, Michael Niemeyer, Richard Szeliski, Marc Pollefeys, and Federico Tombari. Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In NeurIPS, 2024. 2, 7, 6, 8 +[40] Jinke Wang, Xingxing Zuo, Xiangrui Zhao, Jiajun Lyu, and Yong Liu. Review of multi-source fusion slam: current status and challenges, 2022. 2 +[41] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2 +[42] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7 +[44] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall, + +Kaspar Rohrer, Norbert Felber, et al. A hardware architecture for surface splattering. ACM Transactions on Graphics (TOG), 26(3):90-es, 2007. 3 +[45] Liwen Wu, Sai Bi, Zexiang Xu, Fujun Luan, Kai Zhang, Iliyan Georgiev, Kalyan Sunkavalli, and Ravi Ramamoorthi. Neural directional encoding for efficient and accurate view-dependent appearance modeling. In CVPR, 2024. 7, 2 +[46] Jiankai Xing and Kun Xu. Physically based differentiable rendering: a survey, 2024. 2 +[47] Zhen Xu, Tao Xie, Sida Peng, Haotong Lin, Qing Shuai, Zhiyuan Yu, Guangzhao He, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Easyvolcap: Accelerating neural volumetric video research. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6 +[48] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2 +[49] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics, 2024. 4, 5 +[50] Hanqiao Ye, Yangdong Liu, and Shuhan Shen. Lightweight visual-based localization technology, 2024. 2 +[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2, 4, 7, 1, 6, 8 +[52] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7 +[54] Youjia Zhang, Anpei Chen, Yumin Wan,zikai Song, Junqing Yu, Yawei Luo, and Wei Yang. Ref-gs: Directional factorization for 2d gaussian splatting. arXiv preprint arXiv:2412.00905, 2024.3 \ No newline at end of file diff --git a/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/images.zip b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e6fd881913c60cf92caa2fe65117d4c1746f77f8 --- /dev/null +++ b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2cde854ef4e72fb2431b452612640d854bef96c715858f842f7d85b1c52511a +size 964450 diff --git a/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/layout.json b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a79bdc5ef3780c1aef0cbe2882ad567606b0d50c --- /dev/null +++ b/2025/EnvGS_ Modeling View-Dependent Appearance with Environment Gaussian/layout.json @@ -0,0 +1,9356 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 541, + 121 + ], + "type": "text", + "content": "EnvGS: Modeling View-Dependent Appearance with Environment Gaussian" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "spans": [ + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": "Tao Xie" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Xi Chen" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Zhen Xu" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Yiman Xie" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Yudong Jin" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " \nYujun Shen" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Sida Peng" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Hujun Bao" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Xiaowei Zhou" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Zhejiang University " + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 141, + 463, + 193 + ], + "type": "text", + "content": " Ant Group" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 219, + 180, + 306 + ], + "blocks": [ + { + "bbox": [ + 57, + 219, + 180, + 306 + ], + "lines": [ + { + "bbox": [ + 57, + 219, + 180, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 219, + 180, + 306 + ], + "type": "image", + "image_path": "ef51f2d8cb2ecb891db9ae1c313867dde00bea90d64b0eaea1d54eb5592f1056.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 308, + 180, + 395 + ], + "blocks": [ + { + "bbox": [ + 57, + 308, + 180, + 395 + ], + "lines": [ + { + "bbox": [ + 57, + 308, + 180, + 395 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 180, + 395 + ], + "type": "image", + "image_path": "754904581540908f9125dff0d9873a7293fc4ab8ae1a7c3aa6ebd0947b9ccf00.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 396, + 148, + 407 + ], + "lines": [ + { + "bbox": [ + 88, + 396, + 148, + 407 + ], + "spans": [ + { + "bbox": [ + 88, + 396, + 148, + 407 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 181, + 219, + 304, + 306 + ], + "blocks": [ + { + "bbox": [ + 181, + 219, + 304, + 306 + ], + "lines": [ + { + "bbox": [ + 181, + 219, + 304, + 306 + ], + "spans": [ + { + "bbox": [ + 181, + 219, + 304, + 306 + ], + "type": "image", + "image_path": "de8f59eb61c19fe1e1bd60991246a9b5e2ecb1771c119b23b8d7588257dc59a3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 181, + 308, + 304, + 395 + ], + "blocks": [ + { + "bbox": [ + 181, + 308, + 304, + 395 + ], + "lines": [ + { + "bbox": [ + 181, + 308, + 304, + 395 + ], + "spans": [ + { + "bbox": [ + 181, + 308, + 304, + 395 + ], + "type": "image", + "image_path": "0d068813dbc458fd714b7c29c1f672c88bb1fdfba19f34b96155c0cc1a20c2b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 396, + 254, + 407 + ], + "lines": [ + { + "bbox": [ + 231, + 396, + 254, + 407 + ], + "spans": [ + { + "bbox": [ + 231, + 396, + 254, + 407 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 219, + 429, + 306 + ], + "blocks": [ + { + "bbox": [ + 307, + 219, + 429, + 306 + ], + "lines": [ + { + "bbox": [ + 307, + 219, + 429, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 429, + 306 + ], + "type": "image", + "image_path": "5d64447eab1b2c70c546c3b69e6dff5f36957dc0af6fcc88757097a62d9f9f34.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 308, + 429, + 395 + ], + "blocks": [ + { + "bbox": [ + 306, + 308, + 429, + 395 + ], + "lines": [ + { + "bbox": [ + 306, + 308, + 429, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 429, + 395 + ], + "type": "image", + "image_path": "253ea988eee46a5bcc71352bc6d85289988258e8098cb2daffd0163602d21180.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 396, + 402, + 407 + ], + "lines": [ + { + "bbox": [ + 333, + 396, + 402, + 407 + ], + "spans": [ + { + "bbox": [ + 333, + 396, + 402, + 407 + ], + "type": "text", + "content": "GaussianShader" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 423, + 555, + 458 + ], + "lines": [ + { + "bbox": [ + 54, + 423, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 555, + 458 + ], + "type": "text", + "content": "Figure 1. Photorealistic, real-time rendering of real-world scenes with complex reflections. Our proposed EnvGS outperforms prior works in capturing complex reflection effects, especially near-field reflections and high-frequency details while maintaining real-time rendering speed. Please see our supplementary video for better visualizations." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 431, + 219, + 553, + 306 + ], + "blocks": [ + { + "bbox": [ + 431, + 219, + 553, + 306 + ], + "lines": [ + { + "bbox": [ + 431, + 219, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 431, + 219, + 553, + 306 + ], + "type": "image", + "image_path": "31d77fbb1d6414468a9c87bbf76595f2876979faf7a5562e52e2b93c7d341706.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 431, + 308, + 553, + 395 + ], + "blocks": [ + { + "bbox": [ + 431, + 308, + 553, + 395 + ], + "lines": [ + { + "bbox": [ + 431, + 308, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 431, + 308, + 553, + 395 + ], + "type": "image", + "image_path": "0fb401364a6b68b5f3f12b9f947c28c40cde6d81f9a8627eb91a0689fb5ff42f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 396, + 504, + 407 + ], + "lines": [ + { + "bbox": [ + 480, + 396, + 504, + 407 + ], + "spans": [ + { + "bbox": [ + 480, + 396, + 504, + 407 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 152, + 480, + 200, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 480, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 152, + 480, + 200, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 508, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 297, + 687 + ], + "type": "text", + "content": "Reconstructing complex reflections in real-world scenes from 2D images is essential for achieving photorealistic novel view synthesis. Existing methods that utilize environment maps to model reflections from distant lighting often struggle with high-frequency reflection details and fail to account for near-field reflections. In this work, we introduce EnvGS, a novel approach that employs a set of Gaussian primitives as an explicit 3D representation for capturing reflections of environments. These environment Gaussian primitives are incorporated with base Gaussian primitives to model the appearance of the whole scene. To efficiently render these environment Gaussian primitives, we developed a ray-tracing-based renderer that leverages the GPU's RT core for fast rendering. This allows us to jointly optimize our model for high-quality reconstruction while maintain-" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 470, + 556, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 556, + 531 + ], + "type": "text", + "content": "ing real-time rendering speeds. Results from multiple realworld and synthetic datasets demonstrate that our method produces significantly more detailed reflections, achieving the best rendering quality in real-time novel view synthesis. The code is available at https://zju3dv.github.io/envgs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 559, + 395, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 559, + 395, + 572 + ], + "spans": [ + { + "bbox": [ + 314, + 559, + 395, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "content": "Novel view synthesis aims to generate novel views of 3D scenes based on a set of input images, which enables many applications such as VR/AR, and autonomous driving. Recent advances in Neural Radiance Fields (NeRF) [28] have demonstrated impressive rendering performance. However, NeRF's high computational cost makes it challenging for real-time applications. More recently, 3D Gaussian Splatting (3DGS) [17] explicitly models scenes with 3D Gaussian primitives and utilizes rasterization for rendering, achieving real-time rendering with competitive quality. However, modeling complex high-frequency specular" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 271, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 271, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 271, + 713 + ], + "type": "text", + "content": "* Equal Contribution. † Corresponding author: Xiaowei Zhou" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5742" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "reflections remains challenging for 3DGS due to the limited expressiveness of the Spherical Harmonics (SH)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 295, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 295, + 242 + ], + "type": "text", + "content": "Recent works GaussianShader [15] and 3DGS-DR [51], enhance 3DGS by integrating an environment map and employing shading functions to blend the appearance from both the environment map and SH for the final rendering. While additional environmental lighting can enhance 3DGS's reflection modeling ability, it still struggles to reconstruct complex specular reflections accurately due to two factors. First, the assumption of distant lighting in environment maps limits their ability to only capture distant illumination and difficult to synthesize accurate near-field reflections. Second, this representation inherently lacks sufficient capacity to capture high-frequency reflection details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "content": "In this paper, we present EnvGS, a novel approach for modeling complex reflections in real-world scenes, addressing the aforementioned challenges. We propose to model reflections using a set of Gaussian primitives termed the environment Gaussian. The geometry and base appearance are represented by another set of Gaussian primitives called base Gaussian. We effectively blend the two Gaussians for rendering and optimization. Our rendering process begins with rendering the base Gaussian for the per-pixel surface position, normal, base color, and blending weight. Next, we render the environment Gaussian at the surface point in the direction of the reflection of the viewing direction around the surface normal to capture reflection colors. Finally, we blend the base color with the reflection color to achieve the final rendering results. In contrast to previous methods, EnvGS captures high-frequency reflection details using Gaussian primitives, offering superior modeling capabilities compared to environment maps. Additionally, our explicit 3D reflection representation eliminates the need for distant lighting assumptions, enabling accurate modeling of near-field reflections, as shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 687 + ], + "type": "text", + "content": "To render the environment Gaussian at each intersection point along the reflection direction, we create a fully differentiable ray-tracing renderer for 2DGS since rasterization is not suited for this task. We build the ray-tracing renderer on CUDA and OptiX [31] for real-time rendering and efficient optimization of environment Gaussian. The rendering process starts by constructing a bounding volume hierarchy (BVH) from the 2D Gaussian primitives. We then cast rays against the BVH, gathering ordered intersections in chunks while integrating the Gaussian properties through volume rendering [16] to achieve the final results. Our Gaussian ray-tracing renderer enables detailed reflection rendering at real-time performance. Furthermore, it allows for efficient joint optimization of the environment Gaussian and base Gaussian, which is essential for accurate reflection modeling, as demonstrated in Sec 5.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "To validate the effectiveness of our method, we evaluate EnvGS on several real and synthetic datasets. The re" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "sults demonstrate that our methods achieve state-of-the-art performance in real-time novel view synthesis and considerably surpass existing real-time methods, particularly in synthesizing complex reflections in real-world scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 121, + 530, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 121, + 530, + 132 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 530, + 132 + ], + "type": "text", + "content": "In summary, we make the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 553, + 274 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 133, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 167 + ], + "type": "text", + "content": "- We propose a novel scene representation for accurately modeling complex near-field and high-frequency reflections in real-world environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 168, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 215 + ], + "type": "text", + "content": "- We developed a real-time ray-tracing renderer for 2DGS, enabling joint optimization of our representation for accurate scene reconstruction while achieving real-time rendering speeds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 274 + ], + "type": "text", + "content": "- Extensive experiments shows that EnvGS significantly outperforms previous methods. To the best of our knowledge, EnvGS is the first method that achieves real-time photorealistic specular reflections synthesizing in real-world scenes." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 285, + 400, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 400, + 297 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 400, + 297 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "text", + "content": "In computer vision and graphics research [9, 12, 20, 23, 40, 46, 50], our work falls into the area of learning scene representations from a set of posed RGB images. In this section, we review NeRF-based and Gaussian Splitting-based methods, particularly their handling of view-dependent effects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 714 + ], + "type": "text", + "content": "Neural Radiance Field. NeRF [28] introduced the concept of neural radiance fields, which model scenes as implicit multilayer perceptrons (MLPs) and render them via volume rendering, achieving impressive results in novel view synthesis. Subsequent advancements focused on enhancing rendering quality [2-4] and computational efficiency. [6, 7, 24, 30, 35]. [21, 41, 42, 48] introduce Signed Distance Field into NeRF to improve geometry quality. However, these methods often model view-dependent effects via simple viewing directions, which can lead to blurry reflection renderings. To address this, Ref-NeRF [37] encodes the outgoing radiance using the reflected view direction, yielding improved results under distant lighting conditions. Follow-up works [8, 22, 25, 39] leverage Signed Distance Fields (SDF) to refine surface normals to enhance reflection and geometry quality. SpecNeRF [27] further incorporates spatially varying Gaussian directional encoding to better capture near-field reflections. NeRF-Casting [38] represents the scene in a unified manner, similar to [4], and performs ray marching along reflection directions to integrate reflection features, which are then decoded into color with MLPs. This approach achieves impressive results on real-world data under both near and distant lighting conditions. However, the requirement for multiple MLP queries per ray makes NeRF-Casting unsuitable for real-time rendering and necessitates substantial training time. In contrast, our approach delivers real-time rendering capabilities while significantly reducing training time." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5743" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 492 + ], + "type": "text", + "content": "Gaussian Splatting. Recently, 3D Gaussian Splatting (3DGS) [17] has made significant strides toward real-time novel view synthesis. Unlike volume-based rendering, Gaussian Splatting uses efficient rasterization to render spatial Gaussian kernels, achieving real-time performance at 1080P resolution. Mip-Splatting [52] further introduces a 3D smoothing filter and a 2D Mip filter for alias-free renderings. Scaffold-GS [26] introduces structured 3D Gaussians to improve rendering efficiency and quality. [5] introduces triple splitting for efficient relighting. 2D Gaussian Splatting (2DGS) [13] replaces the 3D Gaussian kernels with 2D Gaussian that better align with 3D surfaces, leading to more accurate surface reconstruction. However, these methods face challenges in accurately modeling reflections due to their use of spherical harmonics (SH) to parameterize view-dependent effects based solely on viewing direction, which often results in blurry reflections. More recent works GaussianShader [14] and 3DGS-DR [18] try to incorporate additional environment maps to improve reflection modeling ability. However, these methods only consider distant lighting, ignoring near-field lighting, and are unable to capture high-frequency reflection details. 3iGS [36] extends 3DGS by augmenting it with an illumination field using tensorial factorization, rendering final reflections through a neural renderer. However, it is limited to bounded scenes, which restricts its applicability in real-world, unbounded environments. Concurrent work [54] choose to model near-field reflection with tensorial factorization and far-field with spherical feature grid. Our method adopts 2DGS as the scene representation and models environmental illumination using an additional set of environment Gaussian, which is rendered via our proposed Gaussian tracer. Our approach inherently supports high-frequency, near-field, and distant lighting in unbounded scenes, enabling detailed reflection rendering while maintaining real-time performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 500, + 133, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 133, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 133, + 514 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 520, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 616 + ], + "type": "text", + "content": "We begin by introducing 2D Gaussian Splatting (2DGS) [13], which our approach is built upon. 2DGS is an explicit scene representation similar to 3D Gaussian Splatting (3DGS) [17], which uses Gaussian primitives and rasterization to render screen projections. The key difference is that 2DGS represents the scene with a set of scaled 2D Gaussian primitives defined in a local tangent plane within world space by a transformation matrix " + }, + { + "bbox": [ + 55, + 520, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 520, + 295, + 616 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 621, + 295, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 621, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 107, + 621, + 295, + 648 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\left[ \\begin{array}{c c c c} s _ {u} \\mathbf {t} _ {u} & s _ {v} \\mathbf {t} _ {v} & 0 & \\mathbf {p} _ {k} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right], \\tag {1}", + "image_path": "27aca62f1a7227f2acea0e21ee0c6a9108754e49b4f51e5dbfdde69405a173dd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k, (\\mathbf{t}_u, \\mathbf{t}_v)" + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "inline_equation", + "content": "(s_u, s_v)" + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "content": " denote the center, the two principal tangential vectors, and the scaling factors of the Gaussian, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "To render an image, 2DGS employs the method described in [44] to determine the ray-primitive intersections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "These intersection points are subsequently utilized to compute the Gaussian's contribution to the final image. The Gaussian properties are then integrated using a volume rendering algorithm to obtain the final pixel color:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 319, + 127, + 555, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 127, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 319, + 127, + 555, + 161 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\text {w i t h} \\alpha_ {i} = \\sigma_ {i} \\mathcal {G} _ {i}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}", + "image_path": "25cc445a9fe58f38d551e62dfdfcc25c8191e6d83147c8151b5b489017993d20.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 168, + 544, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 544, + 179 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 168, + 544, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 313, + 168, + 544, + 179 + ], + "type": "text", + "content": " is the standard 2D Gaussian value evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 181, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 348 + ], + "type": "text", + "content": "Compared to 3D Gaussian, 2D Gaussian offers distinct advantages as a surface representation. First, the ray-splat intersection method adopted by 2DGS avoids multi-view depth inconsistency. Second, 2D Gaussian inherently provides a well-defined normal, which is essential for high-quality surface reconstruction and accurate reflection calculations. However, 2DGS relies on the limited representational capacity of Spherical Harmonics (SH) to model scene appearance, preventing it from capturing strong view-dependent effects like specular reflections, which leads to poor rendering results and \"foggy\" geometry [37]. To this end, we use the geometry-aligned 2D Gaussian primitives as the base scene representation and demonstrate how we effectively model complex reflections in the next section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 358, + 370, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 370, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 370, + 370 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 378, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 556, + 559 + ], + "type": "text", + "content": "Given a set of input images of a reflective scene, our goal is to reconstruct the 3D scene and synthesize photorealistic novel views in real-time. To achieve this, we propose utilizing environment Gaussian as an explicit 3D environment representation, which enables accurate modeling of complex reflections in real-world scenes. Additionally, we represent the scene geometry and base colors using another set of Gaussian primitives, denoted as base Gaussian. In this section, we first detail how environment Gaussian and base Gaussian work together to model complex reflections within the scene (Sec. 4.1). Then, We describe the design of a ray-tracing renderer that leverages the GPU's RT cores to efficiently render and optimize the environment Gaussian (Sec. 4.2). Finally, we discuss our optimization process. (Sec. 4.3). An overview of our method is in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 564, + 464, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 464, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 464, + 577 + ], + "type": "text", + "content": "4.1. Reflective Scenes Modeling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "Gaussian splatting [13, 17] models appearance using Spherical Harmonics (SH), which has limited representation capacity for view-dependent effects. These limitations hinder its ability to capture complex, high-frequency specular reflections. Building on this observation, our key insight is that modeling reflections with a Gaussian environment representation can better model complex reflection effects while significantly reducing the complexity required for each Gaussian to capture intricate details within its SH." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Our proposed reflective scene representation includes two sets of 2D Gaussians: a base Gaussian " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\text {base }}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " for mod-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5744" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 311 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 311 + ], + "type": "image", + "image_path": "7d6ef788c09d446ce552872609a528861b934fe5dbe8fc04d819622b0a4ae8b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": "Figure 2. Overview of EnvGS. The rendering process begins by rasterizing the base Gaussian to obtain per-pixel normals, base colors, and blending weights. Next, we render the environment Gaussian in the reflection direction using our ray-tracing-based Gaussian renderer to capture the reflection colors. Finally, we combine the reflection and base colors for the final output. We jointly optimize the environment Gaussian and base Gaussian using monocular normals [49] and ground truth images for supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "text", + "content": "eling the scene's geometry and base appearance, and another environment Gaussian " + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{env}" + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "text", + "content": " for capturing scene reflections. The basic parameterization of each Gaussian primitive is consistent with the original 2DGS [13], including 3D center position " + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "text", + "content": ", two principal tangential vectors " + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "inline_equation", + "content": "(\\mathbf{t}_u,\\mathbf{t}_v)" + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "text", + "content": ", a scaling vector " + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "inline_equation", + "content": "(s_u,s_v)" + }, + { + "bbox": [ + 54, + 375, + 295, + 483 + ], + "type": "text", + "content": ", and SH coefficients. To combine the two appearance components from the base and environment Gaussian into the final result, we introduce a blending weight for each base Gaussian." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "content": "The rendering process is performed in three steps. First, the base Gaussian " + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{base}" + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "content": " is rendered using standard 2D Gaussian splating to obtain the base color " + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{base}" + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "content": ". By applying the volume rendering integration using Eq. (2), we also derive the surface position " + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "content": ", surface normals " + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "content": ", and blending weight " + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 560, + 295, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 560, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 81, + 560, + 295, + 594 + ], + "type": "interline_equation", + "content": "v = \\sum_ {i \\in \\mathcal {N}} = v _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), v \\in \\{\\mathbf {x}, \\mathbf {n}, \\beta \\}. \\tag {3}", + "image_path": "ca4759368ee42a8e903ba936fa18142b6fe87c17bf03db6d5a396e3ec0e0d206.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": "Then, we compute the reflection direction " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ref}" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": " based on the camera ray direction " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ref}" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": " and the surface normal " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 628, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 628, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 111, + 628, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {r e f} = \\mathbf {d} _ {c a m} - 2 \\left(\\mathbf {d} _ {c a m} \\cdot \\mathbf {n}\\right) \\mathbf {n}. \\tag {4}", + "image_path": "8cef30b3fa2490e18859665d8b67803c80576317f41dcb8fdf8fc565510fa5ec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "With the reflection direction " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ref}" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": " and surface point " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": ", the environment Gaussian " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{env}" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": " is rendered using our differentiable Gaussian tracer to obtain the reflection color " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{ref}" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": ", as detailed in Sec. 4.2. The final color is obtained through:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 701, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 111, + 701, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = (1 - \\beta) \\cdot \\mathbf {c} _ {\\text {b a s e}} + \\beta \\cdot \\mathbf {c} _ {\\text {r e f}}. \\tag {5}", + "image_path": "7e7c2bf6eedf50b47e923b742fc36397c5bfad8f0bda9a297a5b522e4d6d42aa.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "text", + "content": "We blend the base color " + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\text {base }}" + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "text", + "content": " and reflection color " + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\text {ref }}" + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "text", + "content": " using the blending weight " + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "text", + "content": ". A visualization of " + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\text {base }}" + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\text {ref }}" + }, + { + "bbox": [ + 313, + 375, + 553, + 411 + ], + "type": "text", + "content": " can be found at Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "content": "Discussion. Compared to the environment map used by Gaussianshader [15] and 3DGS-DR [51], our explicit Gaussian environment representation offers several advantages. First, our method more accurately captures near-field reflections caused by occlusions from nearby objects. This improvement arises from explicitly modeling each Gaussian at its exact spatial location, thus avoiding the ambiguities and inaccuracies inherent in environment map representations that assume distant lighting. Second, by utilizing Gaussian primitives, EnvGS's environment representation achieves greater expressiveness than low-frequency environment maps, enabling the capture of finer reflection details and enhancing rendering quality, as demonstrated by our experiments 5.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 587, + 463, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 463, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 463, + 601 + ], + "type": "text", + "content": "4.2. Differentiable Ray Tracing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "type": "text", + "content": "Rendering the environment Gaussian with rasterization is impractical, as each pixel corresponds to a unique reflection ray and functions as a virtual camera. To address this, we draw on [29] and leverage the advanced optimizations of modern GPUs to design a novel, fully differentiable ray tracing framework. Built on OptiX [31], our framework achieves real-time rendering of 2,000,000 2DGS with a resolution of " + }, + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "type": "inline_equation", + "content": "1292 \\times 839" + }, + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "type": "text", + "content": " at 30 FPS on an RTX 4090 GPU." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "content": "In order to fully utilize the hardware acceleration for" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "spans": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "text", + "content": "5745" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 136, + 149 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 136, + 149 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 136, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 136, + 149 + ], + "type": "image", + "image_path": "6610bdc1bd29c6718c6c8d304052f682369efc8ac1da941cae368d6835f27f03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 136, + 70, + 216, + 149 + ], + "blocks": [ + { + "bbox": [ + 136, + 70, + 216, + 149 + ], + "lines": [ + { + "bbox": [ + 136, + 70, + 216, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 70, + 216, + 149 + ], + "type": "image", + "image_path": "9e1fa2929b6978e536f1079f62e079735155deb5a1bc578c2c28b2d38b887609.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 216, + 70, + 294, + 149 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 294, + 149 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 294, + 149 + ], + "type": "image", + "image_path": "20ba3b7f2c8ab890a2b571595430755959852b3defd4ede0137bb51eef4b33be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 150, + 136, + 228 + ], + "blocks": [ + { + "bbox": [ + 57, + 150, + 136, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 150, + 136, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 150, + 136, + 228 + ], + "type": "image", + "image_path": "95bc2f61d1ec65396e5bebab5b3d0309205e404d39c9e0383db6e8f053b4c262.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 229, + 111, + 236 + ], + "lines": [ + { + "bbox": [ + 82, + 229, + 111, + 236 + ], + "spans": [ + { + "bbox": [ + 82, + 229, + 111, + 236 + ], + "type": "text", + "content": "Final color" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 150, + 216, + 228 + ], + "blocks": [ + { + "bbox": [ + 136, + 150, + 216, + 228 + ], + "lines": [ + { + "bbox": [ + 136, + 150, + 216, + 228 + ], + "spans": [ + { + "bbox": [ + 136, + 150, + 216, + 228 + ], + "type": "image", + "image_path": "c87d914cfdc63e235d311305e62b71b462c67534de245ddafd82eb6e74993042.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 229, + 191, + 236 + ], + "lines": [ + { + "bbox": [ + 162, + 229, + 191, + 236 + ], + "spans": [ + { + "bbox": [ + 162, + 229, + 191, + 236 + ], + "type": "text", + "content": "Base color" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 216, + 150, + 294, + 228 + ], + "blocks": [ + { + "bbox": [ + 216, + 150, + 294, + 228 + ], + "lines": [ + { + "bbox": [ + 216, + 150, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 216, + 150, + 294, + 228 + ], + "type": "image", + "image_path": "9d9933fcf429055170f5ddf867258a2964ce7ff7e8d1437e771b72487167bdc8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 229, + 277, + 236 + ], + "lines": [ + { + "bbox": [ + 233, + 229, + 277, + 236 + ], + "spans": [ + { + "bbox": [ + 233, + 229, + 277, + 236 + ], + "type": "text", + "content": "Reflection color" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 250, + 295, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 295, + 283 + ], + "type": "text", + "content": "Figure 3. Visualization of reflection and base color. Our method successfully reconstructs near-field and distant reflections using the environment Gaussian instead of baking into the base color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "text", + "content": "ray-primitive intersections, we need to convert each 2D Gaussian into a geometric primitive compatible with GPU processing and insert it into a bounding volume hierarchy (BVH). In light of this, we propose to represent each 2D Gaussian with two triangles. Specifically, we first define the four Gaussian bounding vertices " + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{local} = \\{(\\mathrm{sgn}(r),\\mathrm{sgn}(r))\\}" + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "text", + "content": " in the local tangent plane, where " + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\mathrm{sgn}(\\cdot)" + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "text", + "content": " is the sign function and " + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "text", + "content": " is set to 3 representing three times the sigma range. Then, the four local bounding vertices " + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{local}" + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "text", + "content": " are transformed to world space as the vertices of the two triangles covering the Gaussian " + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{world}" + }, + { + "bbox": [ + 55, + 293, + 295, + 448 + ], + "type": "text", + "content": " using Eq. 1. After the transformation, the triangles are organized into a BVH, which serves as the input for the ray tracing process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": "We develop a custom CUDA kernel using the raygen and anyhit programmable entry points of OptiX. Inspired by [29], the rendering is done in a chunk-by-chunk manner. The anyhit kernel traces the input ray to obtain a chunk of size " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{k}" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": ", while raygen integrates this chunk and invokes anyhit to retrieve the next chunk along the ray. Specifically, consider an input ray with origin " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": ". The raygen program first initiates a traversal against the BVH to identify all possible intersections along the ray. During the traversal, the anyhit program sorts each intersected Gaussian by the depth and maintains a sorted k-buffer for the closest " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{k}" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " intersections. We empirically found that " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{k}" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " with 16 is the best trade-off between traversal counts and the number of Gaussians sorted per traversal. After traversal, the raygen program integrates properties of the sorted Gaussians in the buffer following Eq. 2. The Gaussian response is calculated by applying the inverse of the transformation matrix " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " to the ray's intersection point " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " and evaluating the Gaussian value at the transformed point as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 681, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 681, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 129, + 681, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {i} \\left(\\mathbf {u} _ {i}\\right) = \\mathcal {G} _ {i} \\left(\\mathbf {H} ^ {- 1} \\mathbf {x} _ {i}\\right). \\tag {6}", + "image_path": "bdee35dec8704d31060e2f40bf11d4b308021e4181f9468df9241143596d75b2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "This process repeats until no further intersections are found" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "along the ray or the accumulated transmittance drops below a specified threshold. More details can be found in supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 108, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 288 + ], + "type": "text", + "content": "Our ray tracing framework is fully differentiable, allowing for end-to-end optimization of both the base and the environment Gaussian primitives. However, storing all intersections during the forward pass and performing the backward pass in back-to-front order, as done in 3DGS [17], is impractical due to high memory consumption. To address this, we implement the backward pass in the same front-to-back order as the forward pass by re-casting rays and computing gradients for each integration step. A key aspect is calculating the gradient with respect to the input ray origin " + }, + { + "bbox": [ + 313, + 108, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\mathbf{o}}" + }, + { + "bbox": [ + 313, + 108, + 555, + 288 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 313, + 108, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\mathbf{d}}" + }, + { + "bbox": [ + 313, + 108, + 555, + 288 + ], + "type": "text", + "content": ", which is crucial for the joint optimization of our model (see supplementary for more details). Our ablation in Sec 5.4 demonstrates that this joint optimization of geometry and appearance is essential for recovering geometrically accurate surfaces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 293, + 400, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 400, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 400, + 306 + ], + "type": "text", + "content": "4.3. Optimization" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": "To enhance training stability, we initiate optimization by first training the base Gaussian " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\text {base }}" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": ", which is initialized using the sparse point cloud obtained from Structure-from-Motion (SfM) [11, 34]. After the bootstrapping phase, we initialize the environment Gaussian " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{env}" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " by partitioning the scene's bounding box " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\text {scene }}" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " sub-grids and randomly sampling " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " primitives within each grid, and then optimize the base Gaussian and environment Gaussian jointly. The scene's bounding box " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\text {scene }}" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " is determined by the " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "99.5\\%" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " quantile of the sparse point cloud's bounding box " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\text {sfm }}" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " obtained from SfM, the sub-grids resolution " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " is set to 32, and " + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 311, + 555, + 454 + ], + "type": "text", + "content": " is set to 5 for each grid." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "text", + "content": "We follow 2DGS [13] to add a normal consistency constraint between the rendered normal map " + }, + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "text", + "content": " and the gradients of the depth map " + }, + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_d" + }, + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 369, + 496, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 496, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 369, + 496, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\mathbf {n} _ {i} ^ {\\top} \\mathbf {N} _ {d}\\right), \\tag {7}", + "image_path": "25351c5897b44c2bf409f55ce34c7b9f5446004463008214937fdb1ecc018e0f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 534, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 534, + 555, + 570 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 313, + 534, + 555, + 570 + ], + "type": "text", + "content": " is the number of pixels in the image and " + }, + { + "bbox": [ + 313, + 534, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_d" + }, + { + "bbox": [ + 313, + 534, + 555, + 570 + ], + "type": "text", + "content": " is calculated as finite differences of neighboring pixels in the depth map:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 373, + 575, + 553, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 575, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 373, + 575, + 553, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {N} _ {d} (\\mathbf {u}) = \\frac {\\nabla_ {u} \\mathbf {p} _ {d} \\times \\nabla_ {v} \\mathbf {p} _ {d}}{\\| \\nabla_ {u} \\mathbf {p} _ {d} \\times \\nabla_ {v} \\mathbf {p} _ {d} \\|}. \\tag {8}", + "image_path": "486e72196d7f8c17c6a46cd9727f3acd77470f46d5c3defe08e122cc18906655.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 677 + ], + "type": "text", + "content": "However, the normal consistency constraint alone is insufficient for accurately modeling ambiguous surfaces involving both reflection and refraction. Inspired by recent advances in monocular normal estimation [1, 10, 49], we propose to supervise the rendered normal map " + }, + { + "bbox": [ + 313, + 605, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 605, + 555, + 677 + ], + "type": "text", + "content": " using monocular normal estimates " + }, + { + "bbox": [ + 313, + 605, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_m" + }, + { + "bbox": [ + 313, + 605, + 555, + 677 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 367, + 682, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 682, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 367, + 682, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m o n o}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\mathbf {n} _ {i} ^ {\\top} \\mathbf {N} _ {m}\\right). \\tag {9}", + "image_path": "e92fb965b58be504cfdf6a3f32a270c76394b34bf5604d942852193e28f81d13.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "spans": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 553, + 469 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 553, + 469 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 553, + 469 + ], + "type": "image", + "image_path": "f50060daa754012bb7e24a1d4f011b8219ba9e6a428731fd662082ea646a4023.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 476, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 476, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 555, + 499 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on real scenes. Our method significantly improves rendering quality over previous approaches, particularly in producing more detailed reflections. Zoom in for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "content": "In addition to the " + }, + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "content": " image loss and D-SSIM loss " + }, + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "content": " employed by 3DGS [17], we also use perceptual loss [53] to enhance the perceived quality of the rendered image:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 558, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 558, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 117, + 558, + 295, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p e r c}} = \\left\\| \\Phi (\\mathbf {I}) - \\Phi \\left(\\mathbf {I} _ {g t}\\right) \\right\\| _ {1}, \\tag {10}", + "image_path": "f5fef218e13ae6e80e7cb71fb0d4c5add2afef98ec17f8956309d96c316331a3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "text", + "content": " is the pre-trained VGG-16 network [33] and " + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{I},\\mathbf{I}_{gt}" + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "text", + "content": " are the rendered and ground truth images, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 613, + 215, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 613, + 215, + 623 + ], + "spans": [ + { + "bbox": [ + 67, + 613, + 215, + 623 + ], + "type": "text", + "content": "The final loss function is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 638, + 295, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 638, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 72, + 638, + 295, + 652 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {\\text {n o r m}} + \\lambda_ {2} \\mathcal {L} _ {\\text {m o n o}} + \\lambda_ {3} \\mathcal {L} _ {\\text {p e r c}}, \\tag {11}", + "image_path": "871603b8e763256a282c1c90322fb4418f039a86e65915a8e868eb9731df3f4d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the photometric reconstruction loss combining the " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " loss and a D-SSIM [43] term with a ratio of 0.8 and 0.2 respectively. We set " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.04" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 0.01" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_3 = 0.01" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " across our experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 506, + 395, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 506, + 395, + 520 + ], + "spans": [ + { + "bbox": [ + 314, + 506, + 395, + 520 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 528, + 447, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 447, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 447, + 540 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "We implement EnvGS with custom OptiX kernels and optimize our model using the PyTorch framework [32, 47] with the Adam optimizer [19]. Specifically, we set the learning rates for the parameters of each base and environment Gaussian to match those used in 2DGS [13]. The learning rate for the blending weight is set to " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "1e^{-2}" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": ". During training, we apply the adaptive Gaussian control strategy of 3DGS [17] with the normal propagation and color sabotage introduced in 3DGS-DR [18]. Since our Gaussian tracer integrates the Gaussian properties directly in 3D space, there is no valid gradient for the projected 2D center, which is used as the densification criterion in 3DGS. Following [29], we accumulate the 3D spatial gradients of the Gaussian position to achieve a similar effect. Note that each accumulated gra" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 727, + 301, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 727, + 301, + 735 + ], + "spans": [ + { + "bbox": [ + 295, + 727, + 301, + 735 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 70, + 520, + 197 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 520, + 197 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 520, + 197 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 520, + 197 + ], + "type": "table", + "html": "
MethodsRef-NeRF Real Scenes [37]NeRF-Casting Scenes [38]FPS↑Training Time↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Non real-timeRef-NeRF* [37]23.0870.6250.26130.5830.8900.124<0.147h
UniSDF [39]23.7000.6350.26630.8380.8890.130<0.1>47h
ZipNeRF [4]23.6770.6350.24731.7400.9040.105<0.1>47h
NeRF-Casting [38]24.6700.6590.24631.0230.8890.128<0.1>47h
Real-time3DGS [17]23.7000.6410.26228.8600.8770.159182.3070.6h
2DGS [13]23.8040.6540.28128.2760.8620.193159.1880.7h
GaussianShader [14]22.8750.6220.31426.4120.8350.21627.9451.6h
3DGS-DR [51]23.5220.6400.27428.4870.8580.197133.5931.0h
Ours24.6170.6710.24130.4440.8860.14826.2212.5h
", + "image_path": "ead50a1a984a9ecc7fa6b28f6b20db3cced035415e22a55dfd6a1bab624df529.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 260, + 282, + 335 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 251 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on Ref-Real [37] and NeRF-Casting Scenes [38] datasets. Our method delivers the highest rendering quality among real-time techniques and outperforms several non-real-time methods, achieving competitive results with the state-of-the-art non-real-time method NeRF-Casting [38], while being 100 times faster. Note that Ref-NeRF* is an improved version of Ref-NeRF [37] that uses Zip-NeRF's geometry model. All metrics are evaluated at 1/4 resolution as prior works [38]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 260, + 282, + 335 + ], + "lines": [ + { + "bbox": [ + 69, + 260, + 282, + 335 + ], + "spans": [ + { + "bbox": [ + 69, + 260, + 282, + 335 + ], + "type": "table", + "html": "
MethodsForegroundNear-Field
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [17]31.6810.9550.04644.1610.9940.008
2DGS [13]31.3280.9530.05343.6240.9940.010
GaussianShader [14]30.6940.9470.05842.3910.9930.010
3DGS-DR31.8140.9530.05044.0070.9940.009
Ours33.2950.9620.04046.3920.9960.007
", + "image_path": "4fad18c74660ece645922244c9eac995494ad9bdd46b40991f1d21602fd6b212.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 419 + ], + "type": "text", + "content": "dient is scaled by half of the intersection depth to prevent under-densification in distant regions. All experiments are conducted on a single NVIDIA RTX 4090 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 428, + 231, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 231, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 231, + 440 + ], + "type": "text", + "content": "5.2. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 446, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 601 + ], + "type": "text", + "content": "We train and evaluate EnvGS on a range of datasets with a focus on real-world scenes characterized by complex view-dependent effects. We evaluate the Ref-Real [37] and NeRF-Casting Shiny Scenes [38] to demonstrate our method's ability for complex specular reflections in real-world scenes. We additionally captured two more real-world scenes for a more comprehensive evaluation. We demonstrate that our methods can reconstruct detailed reflections on complex real-world scenes with real-time rendering speed. Additionally, we evaluate our method on the synthetic Shiny Blender dataset [37], which is rendered using environment maps. More evaluation results can be found in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 602, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 651 + ], + "type": "text", + "content": "We maintain consistent training and testing splits and image resolution across all datasets, following prior works [3, 17, 37]. We use three commonly used metrics for evaluation: PSNR, SSIM [43], and LPIPS [53]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 182, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 182, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 182, + 672 + ], + "type": "text", + "content": "5.3. Baseline Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We compare our method with both implicit and explicit prior works, including Zip-NeRF [13], 3DGS [17], 2DGS [13], which are designed for general scenes with pri" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 260, + 541, + 342 + ], + "blocks": [ + { + "bbox": [ + 55, + 342, + 295, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 376 + ], + "type": "text", + "content": "Table 2. Quantitative results of foreground and near-field region on Ref-Real [37] and NeRF-Casting Scenes [38]. See supplementary Appendix A.1 for more details and qualitative results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 260, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 328, + 260, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 328, + 260, + 541, + 342 + ], + "type": "table", + "html": "
MethodsAudiDog
PSNR↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Ref-NeRF [37]24.5290.7130.39723.6200.6730.389
3DGS [17]26.1710.8250.18125.3000.8820.165
2DGS [13]25.8690.8190.19824.9330.8760.184
GaussianShader [14]24.7680.7950.22523.0610.8400.226
3DGS-DR [51]26.0930.8200.19223.9430.8500.210
Ours28.6290.8760.12125.5990.8850.151
", + "image_path": "e94dfbc6fb837772da6f1d4cc86e38ffe2ef454a9e6b4f10928123923436f263.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 350, + 543, + 361 + ], + "lines": [ + { + "bbox": [ + 324, + 350, + 543, + 361 + ], + "spans": [ + { + "bbox": [ + 324, + 350, + 543, + 361 + ], + "type": "text", + "content": "Table 3. Quantitative results on our self-captured scenes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 372, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 433 + ], + "type": "text", + "content": "marily diffuse appearances, as well as with methods specifically tailored for scenes with strong specular reflections, including Ref-NeRF [37], GaussianShader [14], 3DGS-DR [18], and NeRF-Casting [38]. We also compare with ENVIDR [22], NDE [45] these two object-level methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 517 + ], + "type": "text", + "content": "In this section, we present quantitative and qualitative results to illustrate the advantages of our methods. Since accurate reflection reconstruction and rendering are the essential advantages of our method, we encourage readers to refer to the various rendered continuous video results in the supplementary material for a more comprehensive evaluation of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 520, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 663 + ], + "type": "text", + "content": "We first evaluate our method on nine real-world scenes from the Ref-Real dataset [37], NeRF-Casting Shiny Scenes [38] and our self-captured real-world scenes, which feature complex geometry and specular reflections. The quantitative results, shown in Tab. 1 and Tab. 3, demonstrate that our method outperforms all explicit methods by a large margin and achieves comparable results to the current state-of-the-art implicit method, NeRF-Casting, while being significantly faster. The qualitative results in Fig.4 further highlight the superior performance of our method in capturing complex view-dependent effects, especially nearfield reflections and high-frequency details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We provide additional comparisons on the Mip-NeRF 360 dataset [3], the Shiny Blender dataset [37] and with ENVIDR [22], NDE [45] in supplementary Appendix A, which also includes per-scene metric breakdowns from Tab. 1." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "spans": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "text", + "content": "5748" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 138, + 153 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 138, + 153 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 138, + 153 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 138, + 153 + ], + "type": "image", + "image_path": "c760e364bbce1deff1cd01d9d5e019dc22fc49f5b30caaa27c86c59ef5a62e4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 153, + 138, + 236 + ], + "blocks": [ + { + "bbox": [ + 57, + 153, + 138, + 236 + ], + "lines": [ + { + "bbox": [ + 57, + 153, + 138, + 236 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 138, + 236 + ], + "type": "image", + "image_path": "2791f0a2d6bee68f3e7acba85d3a66896169bc981aa2de7bc27b6de9dec373d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 237, + 134, + 245 + ], + "lines": [ + { + "bbox": [ + 62, + 237, + 134, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 237, + 134, + 245 + ], + "type": "text", + "content": "(a) w/o joint optimization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 140, + 70, + 222, + 153 + ], + "blocks": [ + { + "bbox": [ + 140, + 70, + 222, + 153 + ], + "lines": [ + { + "bbox": [ + 140, + 70, + 222, + 153 + ], + "spans": [ + { + "bbox": [ + 140, + 70, + 222, + 153 + ], + "type": "image", + "image_path": "fe0fde4400901a472eb394fcbcca998095ac1c6ab6f290cf0d2cb3bb93a71a5f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 153, + 221, + 236 + ], + "blocks": [ + { + "bbox": [ + 140, + 153, + 221, + 236 + ], + "lines": [ + { + "bbox": [ + 140, + 153, + 221, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 153, + 221, + 236 + ], + "type": "image", + "image_path": "fb20b86d0890f29745968e343dd29d123ed0550111ab051385579c9b7bc19a8b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 238, + 217, + 245 + ], + "lines": [ + { + "bbox": [ + 145, + 238, + 217, + 245 + ], + "spans": [ + { + "bbox": [ + 145, + 238, + 217, + 245 + ], + "type": "text", + "content": "(b) w/o mono normal loss" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 304, + 153 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 304, + 153 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 304, + 153 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 304, + 153 + ], + "type": "image", + "image_path": "31296fec1e6c85513c6d78699d8ee602da67f339fab9eba9c7efe7a97ee941c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 153, + 304, + 236 + ], + "blocks": [ + { + "bbox": [ + 223, + 153, + 304, + 236 + ], + "lines": [ + { + "bbox": [ + 223, + 153, + 304, + 236 + ], + "spans": [ + { + "bbox": [ + 223, + 153, + 304, + 236 + ], + "type": "image", + "image_path": "a1481b2a5b7bebafef9ac43e119dc91ba90c4c478f6636edcc4bc339525cc4d5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 238, + 297, + 246 + ], + "lines": [ + { + "bbox": [ + 229, + 238, + 297, + 246 + ], + "spans": [ + { + "bbox": [ + 229, + 238, + 297, + 246 + ], + "type": "text", + "content": "(c) w/ environment map" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "type": "text", + "content": "Figure 5. Ablation study of proposed components on the Ref-Real dataset [37]. Removing either the monocular normal constraint or the joint optimization of base and environment Gaussians results in noisy geometry and inaccurate reflection reconstruction. The \"w/ environment map\" variant fails to capture near-field reflections." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 388, + 153 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 388, + 153 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 388, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 388, + 153 + ], + "type": "image", + "image_path": "29ed4b41adca4e473701762271b7f73583c3b7ccc6ee4e45ed67145333d0bb44.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 153, + 388, + 236 + ], + "blocks": [ + { + "bbox": [ + 306, + 153, + 388, + 236 + ], + "lines": [ + { + "bbox": [ + 306, + 153, + 388, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 388, + 236 + ], + "type": "image", + "image_path": "766b3bc215f9be33c0ba73681bd6d72d0ba4c51cc66a9ea7835fd2f94636aecf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 238, + 385, + 246 + ], + "lines": [ + { + "bbox": [ + 308, + 238, + 385, + 246 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 385, + 246 + ], + "type": "text", + "content": "(d) w/o normal propagation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 470, + 153 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 470, + 153 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 470, + 153 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 470, + 153 + ], + "type": "image", + "image_path": "0b627fda375fbb21841a9d8a7dc1502d3c9427fea7ba2a43c039ec39c201900c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 153, + 470, + 236 + ], + "blocks": [ + { + "bbox": [ + 389, + 153, + 470, + 236 + ], + "lines": [ + { + "bbox": [ + 389, + 153, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 389, + 153, + 470, + 236 + ], + "type": "image", + "image_path": "4e2577d09bd87237be35a1e685f52d1ebd981e27f3e83c787d16419672e887a7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 238, + 441, + 246 + ], + "lines": [ + { + "bbox": [ + 417, + 238, + 441, + 246 + ], + "spans": [ + { + "bbox": [ + 417, + 238, + 441, + 246 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 471, + 70, + 554, + 153 + ], + "blocks": [ + { + "bbox": [ + 471, + 70, + 554, + 153 + ], + "lines": [ + { + "bbox": [ + 471, + 70, + 554, + 153 + ], + "spans": [ + { + "bbox": [ + 471, + 70, + 554, + 153 + ], + "type": "image", + "image_path": "c5ecd83fb96bc13565b89964d1524297ab984bf776cdec2c07136025e200b7de.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 471, + 153, + 553, + 236 + ], + "blocks": [ + { + "bbox": [ + 471, + 153, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 471, + 153, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 471, + 153, + 553, + 236 + ], + "type": "image", + "image_path": "d03ef3e2689e95fb639a8a29b94bf0e01b71f3418f6506a1eefdf09650ee86dc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 238, + 530, + 245 + ], + "lines": [ + { + "bbox": [ + 492, + 238, + 530, + 245 + ], + "spans": [ + { + "bbox": [ + 492, + 238, + 530, + 245 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 300, + 295, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 336 + ], + "type": "text", + "content": "Our method achieves competitive results on both datasets, demonstrating its versatility and effectiveness in handling diverse scenes with complex view-dependent effects." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 343, + 156, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 156, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 156, + 355 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 69, + 362, + 282, + 438 + ], + "blocks": [ + { + "bbox": [ + 69, + 362, + 282, + 438 + ], + "lines": [ + { + "bbox": [ + 69, + 362, + 282, + 438 + ], + "spans": [ + { + "bbox": [ + 69, + 362, + 282, + 438 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
w/o joint optimization24.0340.6440.287
w/o mono normal loss24.1070.6480.270
w/ environment map24.1450.6460.285
w/o color sabotage24.2680.6500.269
w/o normal propagation24.1920.6520.271
w/o lpips loss24.5670.6710.280
Ours24.6170.6710.241
", + "image_path": "793a7d2400e158b3f9370fb0c0ef83083b52dc506ca58a5613c238d7e1733e87.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 446, + 225, + 456 + ], + "lines": [ + { + "bbox": [ + 126, + 446, + 225, + 456 + ], + "spans": [ + { + "bbox": [ + 126, + 446, + 225, + 456 + ], + "type": "text", + "content": "Table 4. Ablation studies." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "text", + "content": "In this section, we conduct ablation studies of our key components on the Ref-Real dataset. Quantitative and qualitative results are shown in Tab. 4 and Fig. 5, respectively." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "content": "Joint optimization. The \"w/o joint optimization\" variant detaches the joint optimization of the base Gaussian and environment Gaussian from the reflection rendering step. As shown in Tab. 4 and Fig. 5, This variant fails to recover accurate geometry, leading to inferior reflection reconstruction and rendering quality." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": "Monocular normal constraint The \"w/o monocular normal\" variant removes the monocular normal constraint as described in Sec. 4.3. Training may become trapped in largely incorrect geometry, resulting in inaccurate reflection reconstruction, as demonstrated in Tab.4 and Fig.5." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Environment map representation The \"w/ environment map\" variant replaces our core Gaussian environment representation with an environment map representation while keeping all other components unchanged. Figure 5 illustrates that while it effectively captures smooth distant reflections, it has difficulty modeling near-field and high-frequency reflections, and produces more bumpy geometry." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "text", + "content": "Color sabotage and normal propagation. The “" + }, + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "text", + "content": " color sabotage” and “" + }, + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "text", + "content": " normal propagation” variants omit the color sabotage and normal propagation steps from our method, respectively. As demonstrated in Tab. 4, both variants result in reduced rendering quality." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 361, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 419 + ], + "type": "text", + "content": "Perceptual loss. The \"w/o lipps loss\" variants removes the perceptual loss. The results in Tab. 4 demonstrate that our method still produces detailed and accurate reflections without perceptual loss, which offers only marginal improvements." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 431, + 466, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 466, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 466, + 443 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 452, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 606 + ], + "type": "text", + "content": "This paper introduced EnvGS, a novel reflective scene representation for high-quality complex reflection capturing and real-time rendering. Our method explicitly models reflections with a set of environment Gaussian primitives. The environment Gaussian primitives are used together with a set of base Gaussian primitives that model basic scene properties (geometry, base color, and blending weight) to model the appearance of the whole scene. Furthermore, we develop a differentiable Gaussian ray tracer utilizing GPU's RT core to effectively optimize and render the environment Gaussian. The proposed method demonstrates superior performance in capturing complex reflections across various datasets." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 608, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 554, + 654 + ], + "type": "text", + "content": "A limitation of EnvGS is its difficulty with transparent and refractive materials, as it only addresses reflection direction. Future work could explore extending our method to accommodate these materials." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Acknowledgements. This work was partially supported by NSFC (No. U24B20154, 62402427), Ant Group, and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "spans": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "text", + "content": "5749" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7, 1, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2, 7, 1, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "type": "text", + "content": "[5] Zoubin Bi, Yixin Zeng, Chong Zeng, Fan Pei, Xiang Feng, Kun Zhou, and Hongzhi Wu. Gs3: Efficient relighting with triple gaussian splatting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 324 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 326, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 294, + 357 + ], + "type": "text", + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 359, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 294, + 402 + ], + "type": "text", + "content": "[8] Chen Guangcheng, He Yicheng, He Li, and Zhang Hong. Pisr: Polarimetric neural implicit surface reconstruction for textureless and specular objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "type": "text", + "content": "[9] Yang Hang, Chen Rui, An Shipeng, Wei Hao, and Zhang Heng. The growth of image-related three dimensional reconstruction techniques in deep learning-driven era: a critical summary, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 448, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 448, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 57, + 448, + 294, + 501 + ], + "type": "text", + "content": "[10] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 502, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 294, + 555 + ], + "type": "text", + "content": "[11] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 590 + ], + "type": "text", + "content": "[12] Fuyuan Hu, Chenlu Li, Tao Zhou, Hongfu Cheng, and Minming Gu. A survey on point cloud completion algorithms for deep learning, 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "text", + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 637, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 294, + 679 + ], + "type": "text", + "content": "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 3, 7, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[15] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for re" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "text", + "content": "flective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 2, 4, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "text", + "content": "[16] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 3, 5, 6, 7, 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "text", + "content": "[18] Ye Keyang, Hou Qiming, and Zhou Kun. 3d gaussian splatting with deferred reflection. 2024. 3, 6, 7, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "type": "text", + "content": "[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "text", + "content": "[20] Yingqun Li, Xiao Hu, Xiang Xu, Yanning Xu, and Lu Wang. Deep learning-based foveated rendering in 3d space: a review, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "type": "text", + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 365 + ], + "type": "text", + "content": "[22] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 367, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 367, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 367, + 553, + 409 + ], + "type": "text", + "content": "[23] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 411, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 442 + ], + "type": "text", + "content": "[24] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "text", + "content": "[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 491, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 555, + 544 + ], + "type": "text", + "content": "[26] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "type": "text", + "content": "[27] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[29] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian Ray Tracing: Fast tracing of particle scenes. ACM Transactions on Graphics and SIGGRAPH Asia, 2024. 4, 5, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "spans": [ + { + "bbox": [ + 295, + 727, + 315, + 735 + ], + "type": "text", + "content": "5750" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "text", + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[31] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. Acm transactions on graphics (tog), 29(4):1-13, 2010. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[35] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "type": "text", + "content": "[36] Zhe Jun Tang and Tat-Jen Cham. 3igs: Factorised tensorial illumination for 3d gaussian splatting. arXiv preprint arXiv:2408.03753, 2024. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 398 + ], + "type": "text", + "content": "[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3, 7, 8, 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "text", + "content": "[38] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2, 7, 1, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[39] Fangjinhua Wang, Marie-Julie Rakotosaona, Michael Niemeyer, Richard Szeliski, Marc Pollefeys, and Federico Tombari. Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In NeurIPS, 2024. 2, 7, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "text", + "content": "[40] Jinke Wang, Xingxing Zuo, Xiangrui Zhao, Jiajun Lyu, and Yong Liu. Review of multi-source fusion slam: current status and challenges, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[41] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[42] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[44] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 542 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Kaspar Rohrer, Norbert Felber, et al. A hardware architecture for surface splattering. ACM Transactions on Graphics (TOG), 26(3):90-es, 2007. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[45] Liwen Wu, Sai Bi, Zexiang Xu, Fujun Luan, Kai Zhang, Iliyan Georgiev, Kalyan Sunkavalli, and Ravi Ramamoorthi. Neural directional encoding for efficient and accurate view-dependent appearance modeling. In CVPR, 2024. 7, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 174 + ], + "type": "text", + "content": "[46] Jiankai Xing and Kun Xu. Physically based differentiable rendering: a survey, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[47] Zhen Xu, Tao Xie, Sida Peng, Haotong Lin, Qing Shuai, Zhiyuan Yu, Guangzhao He, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Easyvolcap: Accelerating neural volumetric video research. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "text", + "content": "[48] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[49] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics, 2024. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 363 + ], + "type": "text", + "content": "[50] Hanqiao Ye, Yangdong Liu, and Shuhan Shen. Lightweight visual-based localization technology, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2, 4, 7, 1, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "text", + "content": "[52] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 542 + ], + "type": "text", + "content": "[54] Youjia Zhang, Anpei Chen, Yumin Wan,zikai Song, Junqing Yu, Yawei Luo, and Wei Yang. Ref-gs: Directional factorization for 2d gaussian splatting. arXiv preprint arXiv:2412.00905, 2024.3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 727, + 314, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 727, + 314, + 735 + ], + "spans": [ + { + "bbox": [ + 295, + 727, + 314, + 735 + ], + "type": "text", + "content": "5751" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_content_list.json b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1398bbb3754ca6eb6a00ce89037dc73dfb79456 --- /dev/null +++ b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_content_list.json @@ -0,0 +1,1874 @@ +[ + { + "type": "text", + "text": "EnvPoser: Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling", + "text_level": 1, + "bbox": [ + 80, + 130, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Songpengcheng Xia $^{1}$ Yu Zhang $^{1}$ Zhuo Su $^{2*}$ Xiaozheng Zheng $^{2}$ Zheng Lv $^{2}$ Guidong Wang $^{2}$ Yongjie Zhang $^{2}$ Qi Wu $^{1,2}$ Lei Chu $^{3}$ Ling Pei $^{1\\dagger}$ Shanghai Jiao Tong University ${}^{2}\\mathrm{ByteDance}$ ${}^{3}\\mathrm{University}$ of Southern California", + "bbox": [ + 84, + 200, + 906, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating full-body motion using the tracking signals of head and hands from VR devices holds great potential for various applications. However, the sparsity and unique distribution of observations present a significant challenge, resulting in an ill-posed problem with multiple feasible solutions (i.e., hypotheses). This amplifies uncertainty and ambiguity in full-body motion estimation, especially for the lower-body joints. Therefore, we propose a new method, EnvPoser, that employs a two-stage framework to perform full-body motion estimation using sparse tracking signals and pre-scanned environment from VR devices. EnvPoser models the multi-hypothesis nature of human motion through an uncertainty-aware estimation module in the first stage. In the second stage, we refine these multi-hypothesis estimates by integrating semantic and geometric environmental constraints, ensuring that the final motion estimation aligns realistically with both the environmental context and physical interactions. Qualitative and quantitative experiments on two public datasets demonstrate that our method achieves state-of-the-art performance, highlighting significant improvements in human motion estimation within motion-environment interaction scenarios. Project page: https://xspc.github.io/EnvPoser/.", + "bbox": [ + 75, + 323, + 473, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, motion capture algorithms using sparse tracking signals have achieved significant progress across various research fields, establishing themselves as essential technologies for applications such as motion-sensing games and immersive VR/AR experiences [13, 18, 42, 49]. With the rapid development of VR devices, full-body motion estimation algorithms based on head-mounted display (HMD) have shown great application potential. However, devices", + "bbox": [ + 75, + 722, + 468, + 844 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2fc829f832043b95837aaa864a6bf54e4c8e93fb0ba447c596d072c32f298c89.jpg", + "image_caption": [ + "Figure 1. EnvPoser can estimate the full-body motion using three tracking signals (HMD and hand controllers) and a pre-scanned environment mesh." + ], + "image_footnote": [], + "bbox": [ + 503, + 292, + 890, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "like PICO and Quest typically provide sparse tracking signals of the head and hands, presenting challenges in estimating the full-body motion from these limited observations.", + "bbox": [ + 498, + 652, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current sparse-observation-based human motion estimation methods typically rely on large-scale motion capture datasets, such as AMASS [28], to capture full-body motion through regression [14, 55] or generative models [2, 6, 8]. Alternatively, physics engines and reinforcement learning have been used [39] to synthesize full-body motion, bypassing dataset dependency. Utilizing various advanced architectures or large-scale datasets, these methods [2, 6, 8, 39, 55] attempt to model the mapping from sparse tracking signals to the most probable full-body motion. However, the lack of direct joint observations results in the same sparse input often corresponding to multiple plausible motions, making it a challenge for sparse-observation-", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Project Lead", + "bbox": [ + 94, + 851, + 171, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 96, + 863, + 223, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by the National Nature Science Foundation of China (NSFC) under Grant 62273229.", + "bbox": [ + 78, + 876, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1839", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based motion estimation algorithms to converge on the most probable full-body motion. To address this one-to-many mapping ambiguity, existing methods [4, 30, 31] attempt to add sensors to regions such as the pelvis and legs to narrow the distribution of possible motion outcomes. Although additional sensors can enhance the accuracy and robustness of full-body motion estimation, this approach may negatively impact user experience due to increased hardware requirements [13, 17, 41, 49, 54].", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, an alternative approach to addressing the inherent ambiguity in sparse observations is to incorporate additional environmental constraints, as shown in Fig. 1. Human motion is highly correlated with the surrounding environment, which contains rich context that can provide valuable cues for motion estimation. However, effectively utilizing environmental constraints presents its own challenges, as existing methods [35, 55] often simplify human-environment interactions to basic foot-to-ground contact, overlooking the complexities of interactions. In reality, individuals frequently engage in more complex interactions with their surroundings.", + "bbox": [ + 75, + 228, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight is that joint uncertainty estimation can explicitly model the multi-hypothesis nature of full-body motion estimation, and by incorporating environmental information, we can reduce estimation uncertainty, guiding the motion estimation to converge toward the most plausible outcome that aligns with both the environment and sparse input observations. Therefore, we design a two-stage deep learning framework to handle the one-to-many mapping ambiguity challenge. We first explicitly model the multi-hypothesis human motion by the joint uncertainty estimation and its sampling strategy. Then, by integrating environmental information, we guide the uncertainty-based multi-hypothesis motion estimations to converge on a full-body motion that best aligns with both the environment and sparse observations. To address the challenge of effectively utilizing environmental information, we introduce semantic constraints for non-contact motions, providing contextual cues for realistic estimation, and geometric constraints for contact scenarios, ensuring consistency and preventing collisions.", + "bbox": [ + 75, + 410, + 472, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key contributions of this paper are as follows:", + "bbox": [ + 96, + 714, + 428, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EnvPoser, a novel framework that significantly advances full-body motion estimation from sparse tracking signals by incorporating pre-scanned environmental context. This framework addresses the complexities of estimating human motion from sparse observations, ensuring robustness across diverse interactive scenarios.", + "- This work presents a two-stage framework for human motion estimation: an uncertainty-aware module provides initial estimates that account for multi-hypothesis motion, while an environment-aware refinement module in" + ], + "bbox": [ + 76, + 734, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "tegrates semantic and geometric constraints to ensure realistic outcomes in diverse scenarios, setting a new benchmark in motion estimation with environmental interactions.", + "- Comprehensive quantitative and qualitative experiments demonstrate that our proposed method achieves state-of-the-art performance across two public datasets and excels in scenarios with frequent environmental interactions." + ], + "bbox": [ + 500, + 92, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 227, + 640, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Motion estimation with sparse observations", + "text_level": 1, + "bbox": [ + 500, + 252, + 867, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current Wearable-based human motion estimation approaches mainly adopt two configurations: one uses multiple inertial measurement units (IMUs) placed on body extremities [13, 17, 36, 40, 48, 54], while the other relies on an HMD with two controllers [2, 4, 7, 14, 15, 55].", + "bbox": [ + 496, + 276, + 890, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While methods utilizing six IMUs have shown impressive results, HMD-based techniques leveraging sparse three-point observations show significant potential for practical applications. For instance, AvatarPoser [14] employs a transformer architecture to achieve accurate full-body motion. While Zheng et al. [55] developed a two-stage framework to capture joint dependencies, improving motion accuracy and fluidity. Some studies [1, 5, 8] address the ill-posed problem of sparse-to-dense mapping using generative models. Dittadi et al. [5] and Aliakbarian et al. [1] proposed using Variational Autoencoders (VAE) and normalizing flows, respectively, to estimate full-body poses from three tracking points. Additionally, the diffusion model [6, 8, 35] has also shown notable performance in motion estimation with sparse data. Despite these advances, most methods rely solely on sparse tracking signals, leading to high uncertainty and multiple potential poses for the same input. Therefore, this paper introduces joint uncertainty estimation and constrains physically implausible motions using environmental information.", + "bbox": [ + 496, + 352, + 892, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Pose regression with uncertainty estimation", + "text_level": 1, + "bbox": [ + 500, + 665, + 870, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The sparse nature of observations in 3D human pose estimation introduces significant uncertainty for unobserved joints, as identical inputs can correspond to multiple pose hypotheses [8, 54, 55]. This challenge is particularly evident in image-based pose estimation [3, 9, 22, 24, 43], where monocular methods face depth ambiguity. For instance, Wehrbein et al. [38] modeled the posterior distribution of 3D pose hypotheses using normalizing flows, while MHFormer [24] employed a multi-hypothesis transformer to capture diverse pose hypotheses through a one-to-many mapping. Recently, Shan et al. [33] proposed diffusion-based methods to aggregate multi-hypothesis predictions with joint-wise re-projection.", + "bbox": [ + 496, + 688, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our paper, we aim to use uncertainty estimation", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1840", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to model the multi-hypothesis nature of motion from sparse observations. Uncertainty in deep learning can be categorized into two types [19]: aleatoric (data-related) and epistemic (model-related). Li et al. [23] refined 3D poses by incorporating joint uncertainties through uncertainty guided-sampling and uncertainty-guided self-attention, while Zhang et al. [53] introduced a probabilistic framework that encodes aleatoric uncertainty using a robust negative log-likelihood loss, with epistemic uncertainty guiding model refinement. In wearable motion estimation, Yang et al. [46] leveraged uncertainty to weight IMU-derived features, enhancing pose corrections with additional text annotations.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Scene-aware motion estimation and generation", + "text_level": 1, + "bbox": [ + 76, + 297, + 468, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements in human motion estimation and generation have expanded from solely analyzing body motion to integrating interactions with the surrounding environment, especially vision-based approaches [11, 12, 16, 34, 47, 52]. For instance, PROX [10] incorporated scene constraints into monocular human pose estimation to address depth ambiguity, while Shen et al. [34] used sparse 3D CNNs to estimate absolute positions and dense scene contacts, refining human mesh recovery via cross-attention with 3D scene cues. EgoHMR [52] leveraged ego-camera images and diffusion models within pre-scanned environments to recover full-body motion. Similarly, S2Fusion [35] utilized sparse observations by learning human motion priors on synthetic datasets and conditioned motion generation on scene encoding and VR inputs. However, S2Fusion primarily focuses on lower body and environment interactions. QuestEnvsim [21] employed reinforcement learning to generate realistic poses in highly constrained environments using HMDs. Additionally, recent studies further emphasize the integration of scene context and historical motion to produce plausible human motion [27, 44, 45].", + "bbox": [ + 75, + 321, + 472, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These scene-aware methods underscore the importance of environmental context for enhancing motion estimation. Our approach builds on these foundations by comprehensively leveraging human-environment interactions, addressing both contact and non-contact scenarios to improve motion estimation with sparse tracking signals.", + "bbox": [ + 75, + 640, + 470, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 744, + 169, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 76, + 770, + 261, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to estimate the full-body motion sequence $\\theta = \\{\\theta_t\\}_{t=1}^T \\in \\mathcal{R}^{T \\times 132}$ (the first 22 joints of SMPL [26] model with the 6D representation of rotations) using the sparse tracking signals $\\mathbf{X} = \\{\\mathbf{x}_t\\}_{t=1}^T \\in \\mathcal{R}^{T \\times N_c}$ from the HMD and hand controllers, along with a pre-scanned 3D environment point cloud $\\mathbf{S}$ , where $T$ denotes the length of input, $N_c = 36$ means the channel of input. Each sparse tracking", + "bbox": [ + 76, + 794, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "signal at time step $t$ consists of the head and hand positions $\\pmb{p}_t$ , rotations $\\pmb{R}_t$ , and linear velocities $\\pmb{v}_t$ , represented as $\\pmb{x}_t = [\\pmb{p}_t, \\pmb{R}_t, \\pmb{v}_t]$ . Moreover, we utilize a cropped 3D environment point cloud $V_S \\in \\mathcal{R}^{N_S \\times 3}$ , where $N_S$ , the number of sampled points, is set to 1000. With the human translation, this point cloud refines pose estimation by incorporating environmental context, leveraging both contact and noncontact interactions between the body and the environment to reduce the uncertainty inherent in sparse observations.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our approach is a two-stage model comprising two core technical modules: 1) an uncertainty-aware initial motion estimation module; and 2) an environment-aware motion refinement module that leverages semantic and geometric information from the environment. We first train the uncertainty-aware initial motion estimation module (Stage I) on the AMASS dataset, with a focus on explicitly capturing the multi-hypothesis nature of motion estimation through uncertainty quantification. Subsequently, we jointly train both modules (Stage II) on the motion-environment interaction datasets, applying environmental semantic and geometric constraints to refine the multi-hypothesis estimates from Stage I into the most plausible motion estimation.", + "bbox": [ + 496, + 228, + 892, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Uncertainty-aware Initial Human Motion Estimation Module", + "text_level": 1, + "bbox": [ + 498, + 450, + 890, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain the joint uncertainty quantification and explicitly model the multi-hypothesis nature of motion estimation, we propose an uncertainty-aware initial human motion estimation module. This module is designed with three primary objectives: (1) to integrate historical motion states and sparse observations for the current initial motion estimation; (2) to effectively extract features from sparse signals; and (3) to quantify uncertainty in motion estimation, explicitly modeling the multi-hypothesis nature of motion under identical input conditions. Accordingly, this module comprises three sub-modules: sparse input signal and historical motion embedding, transformer-based motion feature extraction, and human motion and joint uncertainty regression.", + "bbox": [ + 496, + 489, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse Input Signal and Historical Human Motion Embedding. To integrate historical motion states and sparse observations for the current initial motion estimation, we adopt an auto-regressive structure in the uncertainty-aware initial human motion estimation module. Our model takes sparse observations $\\mathbf{X} \\in \\mathcal{R}^{T \\times 36}$ and historical motion $\\mathbf{X}_{hm} \\in \\mathcal{R}^{T \\times 132}$ as inputs, embedding these inputs through linear layers to obtain the observation representation $Z_{X}$ and motion historical state representation $Z_{hm}$ . Then, we concatenate the sparse observation embedding and historical motion embedding to obtain the shallow motion representation $Z_{S} = [Z_{X}, Z_{hm}]$ , which can be used for subsequent transformer-based feature extraction.", + "bbox": [ + 496, + 688, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer-based Motion Feature Extraction.", + "bbox": [ + 519, + 885, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1841", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7868b37cd6581f6594f5d4bf64e84d642a1ad73fc868415e5640199603cefbc6.jpg", + "image_caption": [ + "Figure 2. Overview of EnvPoser: A Two-Stage Motion Estimation Model. Stage I involves training the uncertainty-aware initial estimation module on the AMASS dataset to produce initial motion estimates with uncertainty quantification. Stage II refines these estimates by training on motion-environment datasets, incorporating semantic and geometric environmental constraints." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the shallow motion representation $Z_{S}$ , we employ a transformer-based feature extraction module [17] for motion features learning. To retain sequence order information, temporal positional encoding is applied to the shallow motion representation [37]. Using multi-head self-attention and a position-wise feed-forward network within the transformer encoder $F_{t}$ , this module efficiently captures intricate relationships in the sparse tracking signals and historical motion, resulting in high-dimensional motion features $Z_{H} = F_{t}(Z_{S})$ . While the transformer architecture is effective for extracting motion features, the sparsity of the input data—limited to tracking signals from the head and hands—introduces high uncertainty in estimating unobserved joints, particularly in the lower body. This uncertainty arises from the lack of direct observations for these joints, making their estimation inherently ambiguous. To address this challenge, we incorporate a mechanism to quantify the uncertainty in joint estimations.", + "bbox": [ + 75, + 411, + 470, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Motion and Joint Uncertainty Regression. With the high-dimensional motion features $Z_{H}$ , we employ a pose regression network $F_{ph}(\\cdot)$ to predict the human motion $\\widetilde{\\theta} = F_{ph}(Z_H)$ . Given the sparse nature of the input signals, capturing joint uncertainty is crucial in mapping the limited observations to a full-body motion estimation. To achieve this, we utilize a heteroscedastic neural network [19, 23] to estimate the joint uncertainty. Specifically, an uncertainty regression head $F_{uh}(\\cdot)$ is employed on the high-dimensional motion feature to estimate the joint uncertainty $\\delta \\in \\mathcal{R}^{T\\times 132}$ of the predicted motion.", + "bbox": [ + 75, + 684, + 468, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the predicted motion and its associated uncertainty, we resample the full-body motion $\\overline{\\theta}$ around the predicted mean motion $\\widetilde{\\theta}$ according to a Gaussian distribution", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{N}(\\widetilde{\\boldsymbol{\\theta}},\\delta)$ with the uncertainty $\\delta$ . This resampling technique introduces a multi-hypothesis representation, characterizing the inherent ambiguity in the unobserved joints. To ensure differentiability throughout the model training process, we apply the reparameterization trick on this sampling operation, drawing a parameter $\\epsilon$ from a standard Gaussian distribution $\\mathcal{N}(0,1)$ , represented by $\\overline{\\theta} = \\widetilde{\\theta} +\\delta \\cdot \\epsilon$ .", + "bbox": [ + 496, + 410, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through the uncertainty regression head, we can obtain the joint uncertainty estimation for predicted human motion, enabling a multi-hypothesis motion representation with the sampling strategy. We pre-train the first stage of our method on the large synthetic dataset AMASS [28]. Based on the predicted mean human motion $\\widetilde{\\theta}$ and uncertainty $\\delta$ , we train the first stage module with the following loss function:", + "bbox": [ + 496, + 518, + 892, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M} = \\| \\widetilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta} \\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 638, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ represents the ground-truth of human motion. Additionally, to ensure accurate uncertainty prediction, we also incorporate an uncertainty estimation loss as [19] stated:", + "bbox": [ + 496, + 662, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\delta} = \\left\\| \\frac {\\widetilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta}}{\\delta} \\right\\| _ {2} + \\log (\\left\\| \\boldsymbol {\\delta} \\right\\| _ {2}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 715, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, based on the uncertainty-aware initial human motion module, the objective function for Stage I is defined as follows:", + "bbox": [ + 496, + 758, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S - I} = \\lambda_ {M} L _ {M} + \\lambda_ {\\delta} L _ {\\delta}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 803, + 890, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\lambda_{M}$ and $\\lambda_{\\delta}$ are the hyper-parameters for Stage I training.", + "bbox": [ + 496, + 825, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Stage I on the AMASS dataset, along with the design for motion uncertainty quantification, allows our network to effectively capture the multi-hypothesis nature of", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1842", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "motion estimation arising from sparse observations. To address the inherent one-to-many ambiguity, we introduce environmental information as a refinement stage. By incorporating both semantic and geometric environmental information, our method guides the multi-hypothesis motion estimates to converge toward the solution that best aligns with the sparse observations and environmental constraints.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Environment-aware Human Motion Refinement with Semantic and Geometry Constraint", + "text_level": 1, + "bbox": [ + 76, + 205, + 470, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the initial motion estimation and its uncertainty produced by the first-stage uncertainty-aware module, we further refine the multi-hypothesis human motion estimates using the pre-scanned environment point clouds collected by real-world devices. To effectively utilize the environmental information, our design is guided by a key insight: human motion within an environment involves both object interactions (contact scenarios) and collision avoidance (non-contact scenarios). In this module, the environment point clouds are processed through cropping, sampling, and encoding. We then apply constraints based on the environment's semantic and geometric properties to ensure that the full-body motion estimation realistically aligns with the surrounding context.", + "bbox": [ + 75, + 243, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Environmental Point Cloud Embedding. To enhance computational efficiency and focus on relevant interaction areas, we first clip the environmental point clouds based on the human's global position. A circular sampling approach is used to minimize the impact of body orientation, selecting a 1-meter radius around the body and uniformly distributing $N_{S}$ points within this area, denoted as $\\mathbf{V}_{\\mathbf{S}} \\in \\mathcal{R}^{N_{S} \\times 3}$ . The cropped environmental point cloud $\\mathbf{V}_{\\mathbf{S}}$ is then processed by the environment encoder $F_{env}(\\cdot)$ to obtain the environmental point cloud embedding $Z_{env}$ , represented by:", + "bbox": [ + 75, + 455, + 470, + 608 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} _ {\\text {e n v}} = F _ {\\text {e n v}} \\left(\\boldsymbol {V} _ {\\boldsymbol {S}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 617, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the environment encoder $F_{env}(\\cdot)$ is implemented using the vanilla PointNet++ [32]. The extracted environmental embedding would be used in the environment semantic-aware motion refinement.", + "bbox": [ + 75, + 643, + 468, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Environment Semantic-Aware Constraint for Motion Refinement. Capturing semantic information from the environmental point cloud provides crucial guidance for motion generation [27, 44, 47], as human motion is highly correlated with environmental context. Therefore, leveraging this motion-environment correlation as a valuable supplement has the potential to enhance the sparse-observation-based human motion estimation. Inspired by [27, 34], we employ a cross-attention module between the initial motion estimates and the cropped environmental point clouds to refine the human motion estimation. This module comprises an environment-embedding network, a motion information embedding network and an environment-motion", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cross-attention network.", + "bbox": [ + 498, + 92, + 661, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The motion information embedding network $F_{motion}(\\cdot)$ takes the sampled motion estimation $\\widetilde{\\pmb{\\theta}}$ , head translation $\\pmb{p}^{head}$ and the extended sparse observations $X_{new} = \\{X, h, \\pmb{\\theta}^{up}\\} \\in \\mathcal{R}^{T \\times 40}$ as input. The extended sparse observations $X_{new}$ include the relative head height $h$ and head up-vector $\\pmb{\\theta}^{up}$ , obtained via a relative head height calculation module. Following [17, 20], this module calculates terrain height by using the height of the feet when stationary and nearby environmental points, and then determines $h$ as the difference between the absolute head height from the HMD and the estimated terrain height. These inputs are concatenated and processed through linear layers to produce the motion information embedding $Z_M = F_{motion}(\\text{concat}(\\overline{\\theta}, \\pmb{p}^{head}, X_{new}))$ .", + "bbox": [ + 496, + 106, + 890, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the environment-motion cross-attention network, the Key $K$ and Value $V$ vectors can be obtained by multiplying the weight matrix $W_{k}, W_{v}$ with the environmental embedding $Z_{env}$ , while the motion information embedding $Z_{M}$ multiplies the weight matrices $W_{q}$ to get the Query $Q$ vectors. Inspired by [25, 27], we incorporate a spatial salience $s_{\\text{spatial}} = MLP(V_{S})$ that expects our cross-attention model to be able to perceive the distance between environmental points clouds and human bodies. Using the spatial salience $s_{\\text{spatial}}$ in conjunction with attention scores $\\text{Attn}(Q, K)$ , we derive a motion-environment representation $Z_{ME}$ , which could be represented by", + "bbox": [ + 496, + 319, + 892, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} _ {M E} = \\left(\\operatorname {A t t n} (\\boldsymbol {Q}, \\boldsymbol {K}) + \\boldsymbol {s} _ {\\text {s p a t i a l}}\\right) \\cdot \\boldsymbol {V}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 512, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, by concatenating the motion-environment representation $Z_{ME}$ with the motion information embedding $Z_{M}$ and passing them through an MLP network, we obtain the environment-refined motion representation $Z_{RM} = MLP(\\text{concat}(Z_{ME}, Z_{M}))$ .", + "bbox": [ + 496, + 540, + 890, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the environment-refined motion representation $Z_{RM}$ , which integrates environmental semantics and spatial priors between the human and environment, the model can then regress the probability of human-environment contact and generate the final motion estimation.", + "bbox": [ + 496, + 614, + 890, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Regression with Contact Estimation. With the environment-refined motion representation $Z_{RM}$ , we estimate the contact probability $C = \\{c\\}_{t=1}^{T} \\in \\mathcal{R}^{T \\times 22}$ and regress the final human motion $\\hat{\\theta}_{RM}$ . We take the environment-refined motion representation $Z_{RM}$ and the extended sparse observations $X_{new}$ as input, and apply two-layers MLP layers on the concatenated input to get the contact probability prediction $\\hat{C} = MLP(concat(X_{new}, Z_{RM}))$ .", + "bbox": [ + 496, + 690, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contact probability prediction is typically treated as a binary classification problem, where Binary Cross Entropy (BCE) loss is used for training:", + "bbox": [ + 496, + 828, + 890, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o n t a c t}} = B C E L o s s (\\hat {\\boldsymbol {C}}, \\boldsymbol {C}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1843", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the contact probability $\\hat{C}$ and environment-refined motion representation $Z_{RM}$ are concatenated to regress the final motion with a motion decoder network $F_{out}(\\cdot)$ :", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\theta}} _ {R M} = F _ {o u t} (\\operatorname {c o n c a t} (\\mathbf {Z} _ {R M}, \\hat {\\mathbf {C}})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 148, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F_{out}(\\cdot)$ consists of two linear layers with a ReLU activation function. We then obtain the final motion loss:", + "bbox": [ + 76, + 176, + 468, + 207 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M ^ {\\prime}} = \\left\\| \\hat {\\boldsymbol {\\theta}} _ {R M} - \\boldsymbol {\\theta} \\right\\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 220, + 468, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Geometry-Aware Constraint for Motion Refinement. To complement semantic-level constraints, we leverage the geometric structure of the environment to enhance motion estimation accuracy, especially in human-environment interactions (contact scenarios). A straightforward approach is to detect collisions between the estimated human motion and the environment point cloud, thereby refining unrealistic motions. While existing methods [10, 50] often convert 3D environments into Signed Distance Fields (SDF) for collision detection, this approach can be computationally intensive. Inspired by recent advancements [29, 52], we instead employ the COAP model [29] to efficiently compute collisions between the human body and environmental points. With this novel model and the environment point clouds, the COAP-based collision loss [52] can be formulated as:", + "bbox": [ + 75, + 250, + 468, + 489 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o a p} = \\frac {1}{N _ {S}} \\sum_ {i = 1} ^ {N _ {S}} \\sigma \\left(f _ {\\Theta} \\left(V _ {S _ {i}} \\mid \\mathcal {G}\\right)\\right) \\mathbb {I} _ {f _ {\\Theta} \\left(V _ {S _ {i}} \\mid \\mathcal {G}\\right) > 0}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 500, + 468, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $V_{S}$ denotes the cropped environment point clouds, with $N_{S}$ representing the number of points. $\\Theta$ represents the transformation parameters of the human body, encompassing both the motion pose $\\pmb{\\theta}$ and shape $\\beta$ . $f_{\\Theta}(V_{S_i}|\\mathcal{G})$ represents the spatial relationship between the human body model and environmental points. The function $\\sigma(\\cdot)$ is the Sigmoid function, which maps the value of $f_{\\Theta}(V_{S_i}|\\mathcal{G})$ to a range between 0 and 1.", + "bbox": [ + 76, + 551, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since most motions occur on the ground, we remove ground points from the scene to ensure that the COAP-based collision loss specifically targets interactions with surrounding objects. To account for interactions with the ground, we introduce additional constraints, including foot contact, foot height, and ground penetration losses [55]:", + "bbox": [ + 75, + 674, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f c} = \\left\\| \\left(\\hat {\\boldsymbol {P}} _ {R M} ^ {f e e t} - \\boldsymbol {P} ^ {f e e t}\\right) \\cdot \\boldsymbol {C} \\right\\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 775, + 380, + 797 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g f h} = \\left\\| \\hat {\\boldsymbol {z}} _ {P _ {R M}} ^ {f e e t} - \\boldsymbol {z} ^ {g r o u n d} \\right\\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 800, + 468, + 820 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g p} = \\left\\| \\left(\\hat {z} _ {P _ {R M}} ^ {\\text {m i n}} - z ^ {\\text {g r o u n d}}\\right) \\cdot \\boldsymbol {l} \\right\\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 821, + 383, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the $\\hat{z}_{P_{RM}}^{feet}$ and $z_{ground}$ means the height of feet joint positions and the floor height, and $l$ denotes whether the joint is lower than the ground.", + "bbox": [ + 75, + 853, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Final Loss Function. In addition to the first stage loss (Eq.2, Eq.1), COAP-based collision loss (Eq.9), contact loss (Eq.6), foot-ground loss (Eq.10) and final motion loss (Eq.8), we also apply some additional losses to improve model training. With the forward kinematic chain, the joint position $P$ , $\\hat{P}_{RM}$ , and the global hands positions $P_{hand}^{\\prime}$ , $\\hat{P}_{RM}^{\\text {hand }}$ could be calculated by the motion pose $\\theta$ and $\\hat{\\theta}$ . To penalize the errors accumulating along the kinematic chain, we apply a joint position loss $(L_{\\text {posi}})$ and a hand alignment loss $(L_{hAL})$ :", + "bbox": [ + 498, + 90, + 890, + 244 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p o s i}} = \\left\\| \\hat {\\boldsymbol {P}} _ {R M} - \\boldsymbol {P} \\right\\| _ {2}, L _ {h A L} = \\left\\| \\hat {\\boldsymbol {P}} _ {R M} ^ {\\text {h a n d}} - \\boldsymbol {P} ^ {\\text {h a n d}} \\right\\| _ {1}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 258, + 890, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, with the technical analysis provided above, we can formulate the second stage objective function as follows:", + "bbox": [ + 498, + 295, + 890, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {S - I I} = L _ {S - I} + L _ {M ^ {\\prime}} + \\lambda_ {1} L _ {p o s i} + \\lambda_ {2} L _ {h A L} + \\lambda_ {3} L _ {f c} \\\\ + \\lambda_ {4} L _ {\\text {c o n t a c t}} + \\lambda_ {5} L _ {g f h} + \\lambda_ {6} L _ {g p} + \\lambda_ {7} L _ {\\text {c o a p}}. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 352, + 890, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\{\\lambda_i\\}_{i = 1,\\dots ,7}$ represents the hyperparameters. Our model is first trained on the AMASS using the Eq. 3 and then fine-tuned on a motion-environment interaction dataset with Eq. 12. The training process and selection of hyperparameters can be found in the supplementary material.", + "bbox": [ + 498, + 401, + 890, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 494, + 632, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section introduces the datasets, metrics, and competing algorithms used in our experiments. We then compare our method EnvPoser with state-of-the-art algorithms, highlighting its advantages in motion estimation. Finally, ablation studies validate the effectiveness of each module. Additional case studies and discussions are provided in the supplementary materials.", + "bbox": [ + 498, + 520, + 890, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To comprehensively evaluate the efficacy of our proposed method, we conduct experiments on two challenging public motion-scene interactions datasets: Egobody [51] and GIMO [56] datasets. Detailed information about these datasets can be found in the supplementary document.", + "bbox": [ + 498, + 627, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competing methods. We compare the motion estimation performance of the proposed model with the state-of-the-art competing methods, including AvatarPoser [14], AGRoL [6], AvatarJLM [55], and S2Fusion [35] on two public datasets. We re-train all competing methods on the EgoBody and GIMO datasets until convergence. For the results on the GIMO dataset, we reused the values reported in previous studies [35]. For the EgoBody dataset, we tested our trained model under real-time settings and presented the results as follows. It is worth noting that S2Fusion [35] is not fully open-source; therefore, we reproduced S2Fusion based on the details provided in its paper and partial open-source code, marking it as S2Fusion* in this experiment.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1844", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fcc6b04db8a1b9fa5b5710fd5efd781ae7d071850727bb8ddaddb7d39f212ca6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 88, + 210, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eeccb6f85d0353a3add391a9684a75b6bc9a5befb4817815ee321c0b52d97148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 152, + 210, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8df69a351ba0e5095e4051cacdffc2c14a4bc15fbeea277438ed3e7ba150f3f.jpg", + "image_caption": [ + "AvatarPoser" + ], + "image_footnote": [], + "bbox": [ + 86, + 220, + 210, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a9d94105bfb1daa52173b9419d47cb45d3a2eaeb5f5b7dd92b5d1ad0479ddbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 88, + 344, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69f5bda580c60f6382a354848cba29ab4d04c233aaf6827b71b4c4f138a15b7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 152, + 344, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/013ef28dda0aa6516e2c7b7b2041a96e02a5bc50db6bf09f2d58c27cd5967f55.jpg", + "image_caption": [ + "AGRoL" + ], + "image_footnote": [], + "bbox": [ + 220, + 220, + 344, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e32435c167ea98cd5bede7271e9673202fd37f468b5aaa59d4cdcd679a0831a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 88, + 478, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0b30636be4608df50ef55d4ebe1e69b7a348c909fe51f336c82246e82a86991.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 152, + 477, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e3bd1f9f176c5b124822d96098b04117c686403d55ea71857aebcd030a7f80a.jpg", + "image_caption": [ + "AvatarJLM" + ], + "image_footnote": [], + "bbox": [ + 352, + 220, + 477, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e01b1e12b18e5edaee2dfe4088cfaf58ffca40a127134ecbe8baf44e0696574.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 611, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ef11262d323e7f1ed946b4260bd898972639512d281c2a11b284f8f21805dd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 152, + 609, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef03fa587ae612799b7066941ae87a89a87c879bb9384ee38d1685d560e59d7e.jpg", + "image_caption": [ + "S2Fusion*" + ], + "image_footnote": [], + "bbox": [ + 486, + 220, + 609, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4dd84c13ca8b0d4abc7cbfee2af35b369e456fce4a58343fb1e6670bd8999906.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 88, + 743, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d1ab4d43dfb3ee93fe6b801c8ca4073a9ce5ab4c0845312dfad667cb85fcfe6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 152, + 741, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53265b806f4c1dc67114f4ef9e89bb600d8ab26c67820249454b5e3f0957f04b.jpg", + "image_caption": [ + "EnvPoser" + ], + "image_footnote": [], + "bbox": [ + 620, + 220, + 741, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7548ddbc7ee007cc2b5d476cfc9907142e3039681ee606d1b5cfc18d862aabc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 88, + 879, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4406b8fd50c52b55be0afe8818029e09c08eb0701641b609339d5d6e07ef1ccb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 152, + 879, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e07d7e0960ee82bf32007e269e50eb65d143d2604ebc91b567b06f5eabcbf5fa.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 751, + 220, + 879, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8085ac68422a0f1baa5c45d2a261d6315c0d298657bc1b993e3781d680abc2da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EgoBody
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]6.8097.8237.311.0
AGRoL[6]7.21100.9370.321.1
AvatarJLM[55]6.4291.7177.47.2
S2Fusion*[35]6.6589.2219.412.5
EnvPoser6.0074.7174.06.6
GIMO
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]7.0291.3324.016.4
AGRoL[6]6.5888.6269.412.5
AvatarJLM[55]4.9570.7258.110.7
S2Fusion[35]4.6557.8235.710.1
EnvPoser4.3857.6234.68.9
", + "bbox": [ + 81, + 358, + 486, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. The performance comparison with SOTAs.", + "bbox": [ + 117, + 561, + 426, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. Following [35, 55], we use these metrics for evaluation: 1) MPJRE $[^\\circ]$ : The mean joint rotational error across all body joints, measured in degrees; 2) MPJPE $[cm]$ : The mean positional error per joint, measured in centimeters. 3) MPJVE $[cm/s]$ : Represents the average error in joint velocities, expressed in centimeters per second; 4) Jitter: Measures motion smoothness by calculating the average jerk (rate of change of acceleration) across all joints, with lower values indicating smoother motion.", + "bbox": [ + 75, + 588, + 468, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparison with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 76, + 739, + 401, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Experimental Analysis. We compare our proposed method with the competing methods on the two public motion-environment interaction datasets. Among these, AvatarPoser [14], AGRoL [6], and AvatarJLM [55] are the SOTAs that rely solely on sparse tracking signals. As shown in Tab. 1, our model, which incorporates pre-scanned environmental point cloud information, effectively constrains the uncertainty in motion reconstruction caused by sparse tracking signals. Compared to Avatar-", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5754a7bc84d97da9af2379af66d01e2a4bc34b3b1e82254311779b2bf66f2682.jpg", + "image_caption": [ + "Figure 3. Visualization of motion estimation on three test sequences from EgoBody Dataset [51]." + ], + "image_footnote": [], + "bbox": [ + 506, + 358, + 594, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26d7eb7f316c4f4b88e773443cf961a2d1eaae70ff1db42660224557dd3c6da8.jpg", + "image_caption": [ + "AvatarJLM", + "Figure 4. Qualitative Comparison of Interaction Details." + ], + "image_footnote": [], + "bbox": [ + 506, + 460, + 594, + 559 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d8a775c966be2865842459aa367303da0d91332de7e016827a4dc4de5938bf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 359, + 689, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb3e45feca9a6e2ee927513996c620342eac609868b39f138637098e084e152f.jpg", + "image_caption": [ + "S2Fusion*" + ], + "image_footnote": [], + "bbox": [ + 602, + 460, + 689, + 559 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c9684560dbce0e51a442427cc50e998eb743463601f6cd2716a9d8e586e1ff1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 359, + 785, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1ef914f738e78ec581bea0abe9b36200cf678089f222935027308051aae96fd.jpg", + "image_caption": [ + "EnvPoser" + ], + "image_footnote": [], + "bbox": [ + 697, + 460, + 785, + 559 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc6eb79f0d5c5d93ca80447fd4a702ac42e27f32f05b77e22031cc04e1a7d530.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 359, + 880, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/601230f5242a6002a29dd0b1dadf58280575e953e679a1f66834a22079482781.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 794, + 460, + 880, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "JLM [55], which does not utilize environmental information, our method achieves significant improvements on two key metrics—MPJRE and MPJPE—by 0.42 (6.5%)/17.0 (18.5%) and 0.57 (11.5%)/13.1 (18.5%) on the EgoBody and GIMO datasets, respectively. These results underscore the increased accuracy in motion estimation achieved through the integration of environmental information.", + "bbox": [ + 496, + 609, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we benchmark our model against the latest algorithm S2Fusion [35], which introduces environmental information as a constraint for human motion estimation. On the GIMO dataset, our method reduces MPJRE errors by $5.8\\%$ compared to S2Fusion [35], and it outperforms S2Fusion across all evaluation metrics, demonstrating the superior effectiveness of our approach in leveraging environmental information. Moreover, the motion estimation performance of our method also outperforms S2Fusion* on the EgoBody dataset, which reduces MPJRE and MPJPE errors by $9.8\\%$ and $16.3\\%$ . Furthermore, our method shows a marked improvement in motion smoothness, with an $11.9\\%$", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1845", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df34b9791708b799670fe9353dbbc3d54f3e46c1dd79adb9bfe8ef2067701f5a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EgoBodyGIMO
MPJREMPJPEMPJREMPJPE
①Baseline7.4587.15.7274.6
②w/o env-semantic6.0478.04.5562.5
③w/o env-geometry6.0377.54.6363.2
EnvPoser6.0074.74.3857.6
", + "bbox": [ + 81, + 88, + 444, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reduction in Jitter over S2Fusion and a $16.8\\%$ improvement over AvatarJLM on the GIMO dataset.", + "bbox": [ + 78, + 212, + 467, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Experimental Analysis. Fig. 3 showcases the motion estimation results on the EgoBody dataset, where we present three representative sequences and visualize the human motion estimation results at key moments. For state-of-the-art algorithms that do not incorporate environmental information, the visualization results in Fig. 3 show multiple instances of environment penetration during interactions. For example, AvatarJLM exhibits errors such as hand intersections with walls (second row) and leg penetrations through boxes (third row). Furthermore, in scenarios without chairs, squatting motions are misinterpreted as sitting (first row) in AvatarJLM due to the absence of environment understanding. In contrast, EnvPoser effectively leverages environmental information to reduce penetration errors and produce more realistic motion estimates.", + "bbox": [ + 78, + 243, + 467, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In more challenging interaction scenarios, EnvPoser shows certain advantages over S2Fusion, which also incorporates environmental information. Notably, S2Fusion fails to prevent interpenetration between the lower body and surrounding objects (third row in Fig. 3). Moreover, Fig. 4 provides a comparison between EnvPoser and S2Fusion, highlighting the advantages of our approach in handling upper-body interactions with objects and complex lower-body motions through the integration of uncertainty modeling and environmental semantic/geometric constraints. Additionally, compared to AvatarJLM, both EnvPoser and S2Fusion demonstrate the effectiveness of incorporating environmental information to improve human motion estimation.", + "bbox": [ + 78, + 470, + 467, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 78, + 679, + 227, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the environment-aware refinement module. We conducted ablation studies comparing our method with three variants: ① Baseline: an auto-regressive transformer-based network with joint uncertainty estimation; ② w/o env-semantic: incorporating only geometric constraints; ③ w/o env-geometry: incorporating only semantic constraints. Tab. 2 presents the motion estimation performance for each variant. Compared to variant ①, our method achieves substantial improvements by integrating both semantic and geometric environmental information. Further analysis shows that each environment-aware refinement module independently enhances performance, with the combination yielding the best overall results.", + "bbox": [ + 78, + 704, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/73d006bac0133b7fef43565c27ca54687413e801eed78ef6b15bfe187d579d36.jpg", + "table_caption": [ + "Table 2. The ablation study on environment refinement module." + ], + "table_footnote": [], + "table_body": "
EgoBodyGIMO
MPJREMPJPEMPJVEMPJREMPJPEMPJVE
1)Baseline w/o UNC7.6390.8333.05.9077.5270.2
2)Baseline7.4587.1342.95.7274.6266.7
3)EnvPoser w/o UNC6.5680.1172.74.4560.2278.6
EnvPoser6.0074.7174.04.3857.6234.6
", + "bbox": [ + 504, + 88, + 903, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. The effectiveness of the uncertainty estimation.", + "bbox": [ + 529, + 176, + 861, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the uncertainty estimation. As shown in Tab. 3, we compare two variants: 1) Baseline w/o UNC: a baseline without uncertainty estimation and 2) Baseline: a baseline with uncertainty estimation. Results show that integrating the uncertainty estimation module not only quantifies joint uncertainty across motions but also improves full-body motion estimation through additional supervision. We further incorporated both baseline methods into the environment-aware refinement module. As shown in Tab. 3, our method achieves significant improvements over variant 3) EnvPoser w/o UNC: our method without uncertainty estimation, reducing MPJRE and MPJPE errors by $8.11\\%$ and $5.68\\%$ , respectively, on the EgoBody dataset. These results indicate that the uncertainty estimation and resampling methods effectively capture the multi-hypothesis nature of initial motion estimates and, when combined with environmental refinement, adaptively align motion estimation with sparse tracking signals and environmental context.", + "bbox": [ + 501, + 198, + 890, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 501, + 482, + 624, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel two-stage framework for reliable human motion estimation leveraging sparse tracking signals and environmental context. In the first stage, we model the multi-hypothesis nature of motion through uncertainty estimation, followed by a refinement phase that applies environment-aware constraints from both semantic and geometric perspectives. Extensive experiments show that EnvPoser outperforms existing methods, particularly in complex interaction scenarios, highlighting the significance of environmental context in wearable motion estimation and paving the way for future approaches to harness this valuable contextual information. This work has the potential to impact a wide range of interactive applications that benefit from more accurate and context-aware motion estimation.", + "bbox": [ + 501, + 508, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future works. The model assumes a static environment, lacking consideration for dynamic multi-user interactions or object motion, which may limit its effectiveness. This limitation could be addressed by incorporating a third-person perspective to estimate the movement of other people and objects, along with additional semantic and geometric constraints. Furthermore, real-time mesh quality in complex environments is often limited, which may affect performance. Future work will explore using raw images to infer contacts in complex scenes via 2D semantic understanding and introducing additional constraints to improve overall reconstruction quality.", + "bbox": [ + 501, + 719, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1846", + "bbox": [ + 485, + 946, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022. 2", + "[2] Sadegh Aliakbarian, Fatemeh Saleh, David Collier, Pashmina Cameron, and Darren Cosker. Hmd-nemo: Online 3d avatar motion generation from sparse observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9622-9631, 2023. 1, 2", + "[3] Rongyu Chen, Linlin Yang, and Angela Yao. Mhentropy: Entropy meets multiple hypotheses for pose and shape recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14840-14849, 2023. 2", + "[4] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 2", + "[5] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating smpl poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021. 2", + "[6] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 1, 2, 6, 7", + "[7] Zhen Fan, Peng Dai, Zhuo Su, Xu Gao, Zheng Lv, Jiarui Zhang, Tianyuan Du, Guidong Wang, and Yang Zhang. Emhi: A multimodal egocentric human motion dataset with hmd and body-worn imus. arXiv preprint arXiv:2408.17168, 2024. 2", + "[8] Han Feng, Wenchao Ma, Quankai Gao, Xianwei Zheng, Nan Xue, and Huijuan Xu. Stratified avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 153-163, 2024. 1, 2", + "[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 2", + "[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 6", + "[11] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 3", + "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3", + "[13] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: learning to reconstruct human pose from sparse inertial measurements in real time. ACM Trans. Graph. (TOG), 37(6):1-15, 2018. 1, 2", + "[14] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1, 2, 6, 7", + "[15] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time egocentric pose estimation from sparse and intermittent observations everywhere. In European Conference on Computer Vision, pages 277-294. Springer, 2025. 2", + "[16] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 3", + "[17] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers (SA' 22), pages 1-9, 2022. 2, 4, 5", + "[18] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision(ICCV), pages 11510-11520, 2021. 1", + "[19] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3, 4", + "[20] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 5", + "[21] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3", + "[22] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1847", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Han Li, Bowen Shi, Wenrui Dai, Hongwei Zheng, Botao Wang, Yu Sun, Min Guo, Chenglin Li, Junni Zou, and Hongkai Xiong. Pose-oriented transformer with uncertainty-guided refinement for 2d-to-3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1296–1304, 2023. 3, 4", + "[24] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 2", + "[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5", + "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 3", + "[27] Zhenyu Lou, Qiongjie Cui, Haofan Wang, Xu Tang, and Hong Zhou. Multimodal sense-informed forecasting of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2144-2154, 2024. 3, 5", + "[28] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 1, 4", + "[29] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. Coap: Compositional articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13201-13210, 2022. 6", + "[30] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2", + "[31] Jose Luis Ponton, Eduard Pujol, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Dragposer: Motion reconstruction from variable sparse tracking signals via latent space optimization. arXiv preprint arXiv:2406.14567, 2024. 2", + "[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5", + "[33] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2", + "[34] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3, 5", + "[35] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2, 3, 6, 7", + "[36] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusionposer: Real-time human motion reconstruction from arbitrary sparse sensors using autoregressive diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2513-2523, 2024. 2", + "[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4", + "[38] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11199-11208, 2021. 2", + "[39] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 1", + "[40] Yinghao Wu, chaoran wang, Lu Yin, Shihui Guo, and Yipeng Qin. Accurate and steady inertial pose estimation through sequence structure learning and modulation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[41] Songpengcheng Xia, Lei Chu, Ling Pei, Zixuan Zhang, Wenxian Yu, and Robert C Qiu. Learning disentangled representation for mixed-reality human activity recognition with a single imu sensor. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2", + "[42] Songpengcheng Xia, Lei Chu, Ling Pei, Jiarui Yang, Wenxian Yu, and Robert C Qiu. Timestamp-supervised wearable-based activity segmentation and recognition with contrastive learning and order-preserving optimal transport. IEEE Transactions on Mobile Computing, 2024. 1", + "[43] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–989, 2024. 2", + "[44] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22268-22278, 2023. 3, 5", + "[45] Hongdi Yang, Chengyang Li, Zhenxuan Wu, Gaozheng Li, Jingya Wang, Jingyi Yu, Zhuo Su, and Lan Xu. Smgdiff: Soccer motion generation using diffusion probabilistic models. arXiv preprint arXiv:2411.16216, 2024. 3", + "[46] Xueyuan Yang, Chao Yao, and Xiaojuan Ban. Spatial-related sensors matters: 3d human motion reconstruction assisted with textual semantics. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10225-10233, 2024. 3" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1848", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3, 5", + "[48] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 2", + "[49] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), pages 13167-13178, 2022. 1, 2", + "[50] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 6", + "[51] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision, pages 180-200. Springer, 2022. 6, 7", + "[52] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 3, 6", + "[53] Yufei Zhang, Hanjing Wang, Jeffrey O Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9032, 2023. 3", + "[54] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 2", + "[55] Xiaozheng Zheng, Zhuo Su, Chao Wen, Zhou Xue, and Xiaojie Jin. Realistic full-body tracking from sparse observations via joint-level modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14678-14688, 2023. 1, 2, 6, 7", + "[56] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision, pages 676-694. Springer, 2022. 6" + ], + "bbox": [ + 78, + 90, + 470, + 809 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1849", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_model.json b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b4f5f0cec8a7d5171bd81710224df586c33ade5 --- /dev/null +++ b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.131, + 0.892, + 0.177 + ], + "angle": 0, + "content": "EnvPoser: Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.202, + 0.908, + 0.259 + ], + "angle": 0, + "content": "Songpengcheng Xia\\(^{1}\\) Yu Zhang\\(^{1}\\) Zhuo Su\\(^{2*}\\) Xiaozheng Zheng\\(^{2}\\) Zheng Lv\\(^{2}\\) Guidong Wang\\(^{2}\\) Yongjie Zhang\\(^{2}\\) Qi Wu\\(^{1,2}\\) Lei Chu\\(^{3}\\) Ling Pei\\(^{1\\dagger}\\) Shanghai Jiao Tong University \\({}^{2}\\mathrm{ByteDance}\\) \\({}^{3}\\mathrm{University}\\) of Southern California" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.673 + ], + "angle": 0, + "content": "Estimating full-body motion using the tracking signals of head and hands from VR devices holds great potential for various applications. However, the sparsity and unique distribution of observations present a significant challenge, resulting in an ill-posed problem with multiple feasible solutions (i.e., hypotheses). This amplifies uncertainty and ambiguity in full-body motion estimation, especially for the lower-body joints. Therefore, we propose a new method, EnvPoser, that employs a two-stage framework to perform full-body motion estimation using sparse tracking signals and pre-scanned environment from VR devices. EnvPoser models the multi-hypothesis nature of human motion through an uncertainty-aware estimation module in the first stage. In the second stage, we refine these multi-hypothesis estimates by integrating semantic and geometric environmental constraints, ensuring that the final motion estimation aligns realistically with both the environmental context and physical interactions. Qualitative and quantitative experiments on two public datasets demonstrate that our method achieves state-of-the-art performance, highlighting significant improvements in human motion estimation within motion-environment interaction scenarios. Project page: https://xspc.github.io/EnvPoser/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.21, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.845 + ], + "angle": 0, + "content": "In recent years, motion capture algorithms using sparse tracking signals have achieved significant progress across various research fields, establishing themselves as essential technologies for applications such as motion-sensing games and immersive VR/AR experiences [13, 18, 42, 49]. With the rapid development of VR devices, full-body motion estimation algorithms based on head-mounted display (HMD) have shown great application potential. However, devices" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.293, + 0.891, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.594, + 0.894, + 0.636 + ], + "angle": 0, + "content": "Figure 1. EnvPoser can estimate the full-body motion using three tracking signals (HMD and hand controllers) and a pre-scanned environment mesh." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.654, + 0.893, + 0.701 + ], + "angle": 0, + "content": "like PICO and Quest typically provide sparse tracking signals of the head and hands, presenting challenges in estimating the full-body motion from these limited observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Current sparse-observation-based human motion estimation methods typically rely on large-scale motion capture datasets, such as AMASS [28], to capture full-body motion through regression [14, 55] or generative models [2, 6, 8]. Alternatively, physics engines and reinforcement learning have been used [39] to synthesize full-body motion, bypassing dataset dependency. Utilizing various advanced architectures or large-scale datasets, these methods [2, 6, 8, 39, 55] attempt to model the mapping from sparse tracking signals to the most probable full-body motion. However, the lack of direct joint observations results in the same sparse input often corresponding to multiple plausible motions, making it a challenge for sparse-observation-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.852, + 0.173, + 0.864 + ], + "angle": 0, + "content": "*Project Lead" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.864, + 0.224, + 0.876 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.877, + 0.47, + 0.899 + ], + "angle": 0, + "content": "This work was supported by the National Nature Science Foundation of China (NSFC) under Grant 62273229." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.852, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1839" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "based motion estimation algorithms to converge on the most probable full-body motion. To address this one-to-many mapping ambiguity, existing methods [4, 30, 31] attempt to add sensors to regions such as the pelvis and legs to narrow the distribution of possible motion outcomes. Although additional sensors can enhance the accuracy and robustness of full-body motion estimation, this approach may negatively impact user experience due to increased hardware requirements [13, 17, 41, 49, 54]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Therefore, an alternative approach to addressing the inherent ambiguity in sparse observations is to incorporate additional environmental constraints, as shown in Fig. 1. Human motion is highly correlated with the surrounding environment, which contains rich context that can provide valuable cues for motion estimation. However, effectively utilizing environmental constraints presents its own challenges, as existing methods [35, 55] often simplify human-environment interactions to basic foot-to-ground contact, overlooking the complexities of interactions. In reality, individuals frequently engage in more complex interactions with their surroundings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.473, + 0.713 + ], + "angle": 0, + "content": "Our key insight is that joint uncertainty estimation can explicitly model the multi-hypothesis nature of full-body motion estimation, and by incorporating environmental information, we can reduce estimation uncertainty, guiding the motion estimation to converge toward the most plausible outcome that aligns with both the environment and sparse input observations. Therefore, we design a two-stage deep learning framework to handle the one-to-many mapping ambiguity challenge. We first explicitly model the multi-hypothesis human motion by the joint uncertainty estimation and its sampling strategy. Then, by integrating environmental information, we guide the uncertainty-based multi-hypothesis motion estimations to converge on a full-body motion that best aligns with both the environment and sparse observations. To address the challenge of effectively utilizing environmental information, we introduce semantic constraints for non-contact motions, providing contextual cues for realistic estimation, and geometric constraints for contact scenarios, ensuring consistency and preventing collisions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.715, + 0.429, + 0.73 + ], + "angle": 0, + "content": "The key contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.839 + ], + "angle": 0, + "content": "- We propose EnvPoser, a novel framework that significantly advances full-body motion estimation from sparse tracking signals by incorporating pre-scanned environmental context. This framework addresses the complexities of estimating human motion from sparse observations, ensuring robustness across diverse interactive scenarios." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.47, + 0.9 + ], + "angle": 0, + "content": "- This work presents a two-stage framework for human motion estimation: an uncertainty-aware module provides initial estimates that account for multi-hypothesis motion, while an environment-aware refinement module in" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.093, + 0.892, + 0.151 + ], + "angle": 0, + "content": "tegrates semantic and geometric constraints to ensure realistic outcomes in diverse scenarios, setting a new benchmark in motion estimation with environmental interactions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.894, + 0.213 + ], + "angle": 0, + "content": "- Comprehensive quantitative and qualitative experiments demonstrate that our proposed method achieves state-of-the-art performance across two public datasets and excels in scenarios with frequent environmental interactions." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.093, + 0.894, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.642, + 0.243 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.253, + 0.869, + 0.269 + ], + "angle": 0, + "content": "2.1. Motion estimation with sparse observations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Current Wearable-based human motion estimation approaches mainly adopt two configurations: one uses multiple inertial measurement units (IMUs) placed on body extremities [13, 17, 36, 40, 48, 54], while the other relies on an HMD with two controllers [2, 4, 7, 14, 15, 55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.893, + 0.655 + ], + "angle": 0, + "content": "While methods utilizing six IMUs have shown impressive results, HMD-based techniques leveraging sparse three-point observations show significant potential for practical applications. For instance, AvatarPoser [14] employs a transformer architecture to achieve accurate full-body motion. While Zheng et al. [55] developed a two-stage framework to capture joint dependencies, improving motion accuracy and fluidity. Some studies [1, 5, 8] address the ill-posed problem of sparse-to-dense mapping using generative models. Dittadi et al. [5] and Aliakbarian et al. [1] proposed using Variational Autoencoders (VAE) and normalizing flows, respectively, to estimate full-body poses from three tracking points. Additionally, the diffusion model [6, 8, 35] has also shown notable performance in motion estimation with sparse data. Despite these advances, most methods rely solely on sparse tracking signals, leading to high uncertainty and multiple potential poses for the same input. Therefore, this paper introduces joint uncertainty estimation and constrains physically implausible motions using environmental information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.871, + 0.682 + ], + "angle": 0, + "content": "2.2. Pose regression with uncertainty estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.885 + ], + "angle": 0, + "content": "The sparse nature of observations in 3D human pose estimation introduces significant uncertainty for unobserved joints, as identical inputs can correspond to multiple pose hypotheses [8, 54, 55]. This challenge is particularly evident in image-based pose estimation [3, 9, 22, 24, 43], where monocular methods face depth ambiguity. For instance, Wehrbein et al. [38] modeled the posterior distribution of 3D pose hypotheses using normalizing flows, while MHFormer [24] employed a multi-hypothesis transformer to capture diverse pose hypotheses through a one-to-many mapping. Recently, Shan et al. [33] proposed diffusion-based methods to aggregate multi-hypothesis predictions with joint-wise re-projection." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our paper, we aim to use uncertainty estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1840" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "to model the multi-hypothesis nature of motion from sparse observations. Uncertainty in deep learning can be categorized into two types [19]: aleatoric (data-related) and epistemic (model-related). Li et al. [23] refined 3D poses by incorporating joint uncertainties through uncertainty guided-sampling and uncertainty-guided self-attention, while Zhang et al. [53] introduced a probabilistic framework that encodes aleatoric uncertainty using a robust negative log-likelihood loss, with epistemic uncertainty guiding model refinement. In wearable motion estimation, Yang et al. [46] leveraged uncertainty to weight IMU-derived features, enhancing pose corrections with additional text annotations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.299, + 0.47, + 0.315 + ], + "angle": 0, + "content": "2.3. Scene-aware motion estimation and generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.64 + ], + "angle": 0, + "content": "Recent advancements in human motion estimation and generation have expanded from solely analyzing body motion to integrating interactions with the surrounding environment, especially vision-based approaches [11, 12, 16, 34, 47, 52]. For instance, PROX [10] incorporated scene constraints into monocular human pose estimation to address depth ambiguity, while Shen et al. [34] used sparse 3D CNNs to estimate absolute positions and dense scene contacts, refining human mesh recovery via cross-attention with 3D scene cues. EgoHMR [52] leveraged ego-camera images and diffusion models within pre-scanned environments to recover full-body motion. Similarly, S2Fusion [35] utilized sparse observations by learning human motion priors on synthetic datasets and conditioned motion generation on scene encoding and VR inputs. However, S2Fusion primarily focuses on lower body and environment interactions. QuestEnvsim [21] employed reinforcement learning to generate realistic poses in highly constrained environments using HMDs. Additionally, recent studies further emphasize the integration of scene context and historical motion to produce plausible human motion [27, 44, 45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.472, + 0.732 + ], + "angle": 0, + "content": "These scene-aware methods underscore the importance of environmental context for enhancing motion estimation. Our approach builds on these foundations by comprehensively leveraging human-environment interactions, addressing both contact and non-contact scenarios to improve motion estimation with sparse tracking signals." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.746, + 0.17, + 0.761 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.262, + 0.787 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our goal is to estimate the full-body motion sequence \\(\\theta = \\{\\theta_t\\}_{t=1}^T \\in \\mathcal{R}^{T \\times 132}\\) (the first 22 joints of SMPL [26] model with the 6D representation of rotations) using the sparse tracking signals \\(\\mathbf{X} = \\{\\mathbf{x}_t\\}_{t=1}^T \\in \\mathcal{R}^{T \\times N_c}\\) from the HMD and hand controllers, along with a pre-scanned 3D environment point cloud \\(\\mathbf{S}\\), where \\(T\\) denotes the length of input, \\(N_c = 36\\) means the channel of input. Each sparse tracking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "signal at time step \\( t \\) consists of the head and hand positions \\( \\pmb{p}_t \\), rotations \\( \\pmb{R}_t \\), and linear velocities \\( \\pmb{v}_t \\), represented as \\( \\pmb{x}_t = [\\pmb{p}_t, \\pmb{R}_t, \\pmb{v}_t] \\). Moreover, we utilize a cropped 3D environment point cloud \\( V_S \\in \\mathcal{R}^{N_S \\times 3} \\), where \\( N_S \\), the number of sampled points, is set to 1000. With the human translation, this point cloud refines pose estimation by incorporating environmental context, leveraging both contact and noncontact interactions between the body and the environment to reduce the uncertainty inherent in sparse observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.893, + 0.44 + ], + "angle": 0, + "content": "As shown in Fig. 2, our approach is a two-stage model comprising two core technical modules: 1) an uncertainty-aware initial motion estimation module; and 2) an environment-aware motion refinement module that leverages semantic and geometric information from the environment. We first train the uncertainty-aware initial motion estimation module (Stage I) on the AMASS dataset, with a focus on explicitly capturing the multi-hypothesis nature of motion estimation through uncertainty quantification. Subsequently, we jointly train both modules (Stage II) on the motion-environment interaction datasets, applying environmental semantic and geometric constraints to refine the multi-hypothesis estimates from Stage I into the most plausible motion estimation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.452, + 0.892, + 0.482 + ], + "angle": 0, + "content": "3.2. Uncertainty-aware Initial Human Motion Estimation Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.688 + ], + "angle": 0, + "content": "To obtain the joint uncertainty quantification and explicitly model the multi-hypothesis nature of motion estimation, we propose an uncertainty-aware initial human motion estimation module. This module is designed with three primary objectives: (1) to integrate historical motion states and sparse observations for the current initial motion estimation; (2) to effectively extract features from sparse signals; and (3) to quantify uncertainty in motion estimation, explicitly modeling the multi-hypothesis nature of motion under identical input conditions. Accordingly, this module comprises three sub-modules: sparse input signal and historical motion embedding, transformer-based motion feature extraction, and human motion and joint uncertainty regression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Sparse Input Signal and Historical Human Motion Embedding. To integrate historical motion states and sparse observations for the current initial motion estimation, we adopt an auto-regressive structure in the uncertainty-aware initial human motion estimation module. Our model takes sparse observations \\( \\mathbf{X} \\in \\mathcal{R}^{T \\times 36} \\) and historical motion \\( \\mathbf{X}_{hm} \\in \\mathcal{R}^{T \\times 132} \\) as inputs, embedding these inputs through linear layers to obtain the observation representation \\( Z_{X} \\) and motion historical state representation \\( Z_{hm} \\). Then, we concatenate the sparse observation embedding and historical motion embedding to obtain the shallow motion representation \\( Z_{S} = [Z_{X}, Z_{hm}] \\), which can be used for subsequent transformer-based feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Transformer-based Motion Feature Extraction." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1841" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.893, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Overview of EnvPoser: A Two-Stage Motion Estimation Model. Stage I involves training the uncertainty-aware initial estimation module on the AMASS dataset to produce initial motion estimates with uncertainty quantification. Stage II refines these estimates by training on motion-environment datasets, incorporating semantic and geometric environmental constraints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.471, + 0.684 + ], + "angle": 0, + "content": "Based on the shallow motion representation \\( Z_{S} \\), we employ a transformer-based feature extraction module [17] for motion features learning. To retain sequence order information, temporal positional encoding is applied to the shallow motion representation [37]. Using multi-head self-attention and a position-wise feed-forward network within the transformer encoder \\( F_{t} \\), this module efficiently captures intricate relationships in the sparse tracking signals and historical motion, resulting in high-dimensional motion features \\( Z_{H} = F_{t}(Z_{S}) \\). While the transformer architecture is effective for extracting motion features, the sparsity of the input data—limited to tracking signals from the head and hands—introduces high uncertainty in estimating unobserved joints, particularly in the lower body. This uncertainty arises from the lack of direct observations for these joints, making their estimation inherently ambiguous. To address this challenge, we incorporate a mechanism to quantify the uncertainty in joint estimations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Human Motion and Joint Uncertainty Regression. With the high-dimensional motion features \\( Z_{H} \\), we employ a pose regression network \\( F_{ph}(\\cdot) \\) to predict the human motion \\( \\widetilde{\\theta} = F_{ph}(Z_H) \\). Given the sparse nature of the input signals, capturing joint uncertainty is crucial in mapping the limited observations to a full-body motion estimation. To achieve this, we utilize a heteroscedastic neural network [19, 23] to estimate the joint uncertainty. Specifically, an uncertainty regression head \\( F_{uh}(\\cdot) \\) is employed on the high-dimensional motion feature to estimate the joint uncertainty \\( \\delta \\in \\mathcal{R}^{T\\times 132} \\) of the predicted motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Using the predicted motion and its associated uncertainty, we resample the full-body motion \\(\\overline{\\theta}\\) around the predicted mean motion \\(\\widetilde{\\theta}\\) according to a Gaussian distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.519 + ], + "angle": 0, + "content": "\\(\\mathcal{N}(\\widetilde{\\boldsymbol{\\theta}},\\delta)\\) with the uncertainty \\(\\delta\\). This resampling technique introduces a multi-hypothesis representation, characterizing the inherent ambiguity in the unobserved joints. To ensure differentiability throughout the model training process, we apply the reparameterization trick on this sampling operation, drawing a parameter \\(\\epsilon\\) from a standard Gaussian distribution \\(\\mathcal{N}(0,1)\\), represented by \\(\\overline{\\theta} = \\widetilde{\\theta} +\\delta \\cdot \\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.893, + 0.639 + ], + "angle": 0, + "content": "Through the uncertainty regression head, we can obtain the joint uncertainty estimation for predicted human motion, enabling a multi-hypothesis motion representation with the sampling strategy. We pre-train the first stage of our method on the large synthetic dataset AMASS [28]. Based on the predicted mean human motion \\(\\widetilde{\\theta}\\) and uncertainty \\(\\delta\\), we train the first stage module with the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.64, + 0.891, + 0.658 + ], + "angle": 0, + "content": "\\[\nL _ {M} = \\| \\widetilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta} \\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.709 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents the ground-truth of human motion. Additionally, to ensure accurate uncertainty prediction, we also incorporate an uncertainty estimation loss as [19] stated:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.717, + 0.891, + 0.751 + ], + "angle": 0, + "content": "\\[\nL _ {\\delta} = \\left\\| \\frac {\\widetilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta}}{\\delta} \\right\\| _ {2} + \\log (\\left\\| \\boldsymbol {\\delta} \\right\\| _ {2}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.891, + 0.802 + ], + "angle": 0, + "content": "Thus, based on the uncertainty-aware initial human motion module, the objective function for Stage I is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.804, + 0.891, + 0.82 + ], + "angle": 0, + "content": "\\[\nL _ {S - I} = \\lambda_ {M} L _ {M} + \\lambda_ {\\delta} L _ {\\delta}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.891, + 0.855 + ], + "angle": 0, + "content": "where the \\(\\lambda_{M}\\) and \\(\\lambda_{\\delta}\\) are the hyper-parameters for Stage I training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training Stage I on the AMASS dataset, along with the design for motion uncertainty quantification, allows our network to effectively capture the multi-hypothesis nature of" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1842" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "motion estimation arising from sparse observations. To address the inherent one-to-many ambiguity, we introduce environmental information as a refinement stage. By incorporating both semantic and geometric environmental information, our method guides the multi-hypothesis motion estimates to converge toward the solution that best aligns with the sparse observations and environmental constraints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.206, + 0.471, + 0.238 + ], + "angle": 0, + "content": "3.3. Environment-aware Human Motion Refinement with Semantic and Geometry Constraint" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Based on the initial motion estimation and its uncertainty produced by the first-stage uncertainty-aware module, we further refine the multi-hypothesis human motion estimates using the pre-scanned environment point clouds collected by real-world devices. To effectively utilize the environmental information, our design is guided by a key insight: human motion within an environment involves both object interactions (contact scenarios) and collision avoidance (non-contact scenarios). In this module, the environment point clouds are processed through cropping, sampling, and encoding. We then apply constraints based on the environment's semantic and geometric properties to ensure that the full-body motion estimation realistically aligns with the surrounding context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.471, + 0.609 + ], + "angle": 0, + "content": "Environmental Point Cloud Embedding. To enhance computational efficiency and focus on relevant interaction areas, we first clip the environmental point clouds based on the human's global position. A circular sampling approach is used to minimize the impact of body orientation, selecting a 1-meter radius around the body and uniformly distributing \\( N_{S} \\) points within this area, denoted as \\( \\mathbf{V}_{\\mathbf{S}} \\in \\mathcal{R}^{N_{S} \\times 3} \\). The cropped environmental point cloud \\( \\mathbf{V}_{\\mathbf{S}} \\) is then processed by the environment encoder \\( F_{env}(\\cdot) \\) to obtain the environmental point cloud embedding \\( Z_{env} \\), represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.618, + 0.47, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} _ {\\text {e n v}} = F _ {\\text {e n v}} \\left(\\boldsymbol {V} _ {\\boldsymbol {S}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.705 + ], + "angle": 0, + "content": "where the environment encoder \\( F_{env}(\\cdot) \\) is implemented using the vanilla PointNet++ [32]. The extracted environmental embedding would be used in the environment semantic-aware motion refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Environment Semantic-Aware Constraint for Motion Refinement. Capturing semantic information from the environmental point cloud provides crucial guidance for motion generation [27, 44, 47], as human motion is highly correlated with environmental context. Therefore, leveraging this motion-environment correlation as a valuable supplement has the potential to enhance the sparse-observation-based human motion estimation. Inspired by [27, 34], we employ a cross-attention module between the initial motion estimates and the cropped environmental point clouds to refine the human motion estimation. This module comprises an environment-embedding network, a motion information embedding network and an environment-motion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.093, + 0.663, + 0.106 + ], + "angle": 0, + "content": "cross-attention network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.32 + ], + "angle": 0, + "content": "The motion information embedding network \\( F_{motion}(\\cdot) \\) takes the sampled motion estimation \\( \\widetilde{\\pmb{\\theta}} \\), head translation \\( \\pmb{p}^{head} \\) and the extended sparse observations \\( X_{new} = \\{X, h, \\pmb{\\theta}^{up}\\} \\in \\mathcal{R}^{T \\times 40} \\) as input. The extended sparse observations \\( X_{new} \\) include the relative head height \\( h \\) and head up-vector \\( \\pmb{\\theta}^{up} \\), obtained via a relative head height calculation module. Following [17, 20], this module calculates terrain height by using the height of the feet when stationary and nearby environmental points, and then determines \\( h \\) as the difference between the absolute head height from the HMD and the estimated terrain height. These inputs are concatenated and processed through linear layers to produce the motion information embedding \\( Z_M = F_{motion}(\\text{concat}(\\overline{\\theta}, \\pmb{p}^{head}, X_{new})) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.893, + 0.502 + ], + "angle": 0, + "content": "For the environment-motion cross-attention network, the Key \\( K \\) and Value \\( V \\) vectors can be obtained by multiplying the weight matrix \\( W_{k}, W_{v} \\) with the environmental embedding \\( Z_{env} \\), while the motion information embedding \\( Z_{M} \\) multiplies the weight matrices \\( W_{q} \\) to get the Query \\( Q \\) vectors. Inspired by [25, 27], we incorporate a spatial salience \\( s_{\\text{spatial}} = MLP(V_{S}) \\) that expects our cross-attention model to be able to perceive the distance between environmental points clouds and human bodies. Using the spatial salience \\( s_{\\text{spatial}} \\) in conjunction with attention scores \\( \\text{Attn}(Q, K) \\), we derive a motion-environment representation \\( Z_{ME} \\), which could be represented by" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.513, + 0.892, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} _ {M E} = \\left(\\operatorname {A t t n} (\\boldsymbol {Q}, \\boldsymbol {K}) + \\boldsymbol {s} _ {\\text {s p a t i a l}}\\right) \\cdot \\boldsymbol {V}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.616 + ], + "angle": 0, + "content": "Then, by concatenating the motion-environment representation \\( Z_{ME} \\) with the motion information embedding \\( Z_{M} \\) and passing them through an MLP network, we obtain the environment-refined motion representation \\( Z_{RM} = MLP(\\text{concat}(Z_{ME}, Z_{M})) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.691 + ], + "angle": 0, + "content": "Using the environment-refined motion representation \\( Z_{RM} \\), which integrates environmental semantics and spatial priors between the human and environment, the model can then regress the probability of human-environment contact and generate the final motion estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Motion Regression with Contact Estimation. With the environment-refined motion representation \\( Z_{RM} \\), we estimate the contact probability \\( C = \\{c\\}_{t=1}^{T} \\in \\mathcal{R}^{T \\times 22} \\) and regress the final human motion \\( \\hat{\\theta}_{RM} \\). We take the environment-refined motion representation \\( Z_{RM} \\) and the extended sparse observations \\( X_{new} \\) as input, and apply two-layers MLP layers on the concatenated input to get the contact probability prediction \\( \\hat{C} = MLP(concat(X_{new}, Z_{RM})) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.829, + 0.892, + 0.874 + ], + "angle": 0, + "content": "Contact probability prediction is typically treated as a binary classification problem, where Binary Cross Entropy (BCE) loss is used for training:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o n t a c t}} = B C E L o s s (\\hat {\\boldsymbol {C}}, \\boldsymbol {C}). \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1843" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Finally, the contact probability \\(\\hat{C}\\) and environment-refined motion representation \\(Z_{RM}\\) are concatenated to regress the final motion with a motion decoder network \\(F_{out}(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.149, + 0.469, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\theta}} _ {R M} = F _ {o u t} (\\operatorname {c o n c a t} (\\mathbf {Z} _ {R M}, \\hat {\\mathbf {C}})), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.178, + 0.469, + 0.208 + ], + "angle": 0, + "content": "where \\(F_{out}(\\cdot)\\) consists of two linear layers with a ReLU activation function. We then obtain the final motion loss:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.221, + 0.47, + 0.239 + ], + "angle": 0, + "content": "\\[\nL _ {M ^ {\\prime}} = \\left\\| \\hat {\\boldsymbol {\\theta}} _ {R M} - \\boldsymbol {\\theta} \\right\\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.491 + ], + "angle": 0, + "content": "Scene Geometry-Aware Constraint for Motion Refinement. To complement semantic-level constraints, we leverage the geometric structure of the environment to enhance motion estimation accuracy, especially in human-environment interactions (contact scenarios). A straightforward approach is to detect collisions between the estimated human motion and the environment point cloud, thereby refining unrealistic motions. While existing methods [10, 50] often convert 3D environments into Signed Distance Fields (SDF) for collision detection, this approach can be computationally intensive. Inspired by recent advancements [29, 52], we instead employ the COAP model [29] to efficiently compute collisions between the human body and environmental points. With this novel model and the environment point clouds, the COAP-based collision loss [52] can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.501, + 0.47, + 0.542 + ], + "angle": 0, + "content": "\\[\nL _ {c o a p} = \\frac {1}{N _ {S}} \\sum_ {i = 1} ^ {N _ {S}} \\sigma \\left(f _ {\\Theta} \\left(V _ {S _ {i}} \\mid \\mathcal {G}\\right)\\right) \\mathbb {I} _ {f _ {\\Theta} \\left(V _ {S _ {i}} \\mid \\mathcal {G}\\right) > 0}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.469, + 0.673 + ], + "angle": 0, + "content": "where \\( V_{S} \\) denotes the cropped environment point clouds, with \\( N_{S} \\) representing the number of points. \\( \\Theta \\) represents the transformation parameters of the human body, encompassing both the motion pose \\( \\pmb{\\theta} \\) and shape \\( \\beta \\). \\( f_{\\Theta}(V_{S_i}|\\mathcal{G}) \\) represents the spatial relationship between the human body model and environmental points. The function \\( \\sigma(\\cdot) \\) is the Sigmoid function, which maps the value of \\( f_{\\Theta}(V_{S_i}|\\mathcal{G}) \\) to a range between 0 and 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.765 + ], + "angle": 0, + "content": "Since most motions occur on the ground, we remove ground points from the scene to ensure that the COAP-based collision loss specifically targets interactions with surrounding objects. To account for interactions with the ground, we introduce additional constraints, including foot contact, foot height, and ground penetration losses [55]:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.776, + 0.382, + 0.798 + ], + "angle": 0, + "content": "\\[\nL _ {f c} = \\left\\| \\left(\\hat {\\boldsymbol {P}} _ {R M} ^ {f e e t} - \\boldsymbol {P} ^ {f e e t}\\right) \\cdot \\boldsymbol {C} \\right\\| _ {1},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.801, + 0.469, + 0.821 + ], + "angle": 0, + "content": "\\[\nL _ {g f h} = \\left\\| \\hat {\\boldsymbol {z}} _ {P _ {R M}} ^ {f e e t} - \\boldsymbol {z} ^ {g r o u n d} \\right\\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.823, + 0.384, + 0.842 + ], + "angle": 0, + "content": "\\[\nL _ {g p} = \\left\\| \\left(\\hat {z} _ {P _ {R M}} ^ {\\text {m i n}} - z ^ {\\text {g r o u n d}}\\right) \\cdot \\boldsymbol {l} \\right\\| _ {1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.854, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where the \\(\\hat{z}_{P_{RM}}^{feet}\\) and \\(z_{ground}\\) means the height of feet joint positions and the floor height, and \\(l\\) denotes whether the joint is lower than the ground." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.246 + ], + "angle": 0, + "content": "Final Loss Function. In addition to the first stage loss (Eq.2, Eq.1), COAP-based collision loss (Eq.9), contact loss (Eq.6), foot-ground loss (Eq.10) and final motion loss (Eq.8), we also apply some additional losses to improve model training. With the forward kinematic chain, the joint position \\( P \\), \\( \\hat{P}_{RM} \\), and the global hands positions \\( P_{hand}^{\\prime} \\), \\( \\hat{P}_{RM}^{\\text {hand }} \\) could be calculated by the motion pose \\( \\theta \\) and \\( \\hat{\\theta} \\). To penalize the errors accumulating along the kinematic chain, we apply a joint position loss \\( (L_{\\text {posi}}) \\) and a hand alignment loss \\( (L_{hAL}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.259, + 0.891, + 0.295 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p o s i}} = \\left\\| \\hat {\\boldsymbol {P}} _ {R M} - \\boldsymbol {P} \\right\\| _ {2}, L _ {h A L} = \\left\\| \\hat {\\boldsymbol {P}} _ {R M} ^ {\\text {h a n d}} - \\boldsymbol {P} ^ {\\text {h a n d}} \\right\\| _ {1}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.296, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Therefore, with the technical analysis provided above, we can formulate the second stage objective function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.353, + 0.891, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {S - I I} = L _ {S - I} + L _ {M ^ {\\prime}} + \\lambda_ {1} L _ {p o s i} + \\lambda_ {2} L _ {h A L} + \\lambda_ {3} L _ {f c} \\\\ + \\lambda_ {4} L _ {\\text {c o n t a c t}} + \\lambda_ {5} L _ {g f h} + \\lambda_ {6} L _ {g p} + \\lambda_ {7} L _ {\\text {c o a p}}. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.402, + 0.892, + 0.477 + ], + "angle": 0, + "content": "Here, \\(\\{\\lambda_i\\}_{i = 1,\\dots ,7}\\) represents the hyperparameters. Our model is first trained on the AMASS using the Eq. 3 and then fine-tuned on a motion-environment interaction dataset with Eq. 12. The training process and selection of hyperparameters can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.633, + 0.511 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.891, + 0.627 + ], + "angle": 0, + "content": "This section introduces the datasets, metrics, and competing algorithms used in our experiments. We then compare our method EnvPoser with state-of-the-art algorithms, highlighting its advantages in motion estimation. Finally, ablation studies validate the effectiveness of each module. Additional case studies and discussions are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Datasets. To comprehensively evaluate the efficacy of our proposed method, we conduct experiments on two challenging public motion-scene interactions datasets: Egobody [51] and GIMO [56] datasets. Detailed information about these datasets can be found in the supplementary document." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Competing methods. We compare the motion estimation performance of the proposed model with the state-of-the-art competing methods, including AvatarPoser [14], AGRoL [6], AvatarJLM [55], and S2Fusion [35] on two public datasets. We re-train all competing methods on the EgoBody and GIMO datasets until convergence. For the results on the GIMO dataset, we reused the values reported in previous studies [35]. For the EgoBody dataset, we tested our trained model under real-time settings and presented the results as follows. It is worth noting that S2Fusion [35] is not fully open-source; therefore, we reproduced S2Fusion based on the details provided in its paper and partial open-source code, marking it as S2Fusion* in this experiment." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.212, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.154, + 0.212, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.222, + 0.212, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.323, + 0.178, + 0.332 + ], + "angle": 0, + "content": "AvatarPoser" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.089, + 0.346, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.153, + 0.346, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.222, + 0.346, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.323, + 0.302, + 0.332 + ], + "angle": 0, + "content": "AGRoL" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.089, + 0.479, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.153, + 0.478, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.222, + 0.478, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.323, + 0.45, + 0.332 + ], + "angle": 0, + "content": "AvatarJLM" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.089, + 0.612, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.153, + 0.611, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.222, + 0.61, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.323, + 0.58, + 0.332 + ], + "angle": 0, + "content": "S2Fusion*" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.089, + 0.744, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.153, + 0.743, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.222, + 0.743, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.323, + 0.711, + 0.332 + ], + "angle": 0, + "content": "EnvPoser" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.089, + 0.88, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.153, + 0.88, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.222, + 0.88, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.323, + 0.827, + 0.333 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.34, + 0.773, + 0.355 + ], + "angle": 0, + "content": "Figure 3. Visualization of motion estimation on three test sequences from EgoBody Dataset [51]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.359, + 0.487, + 0.56 + ], + "angle": 0, + "content": "
EgoBody
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]6.8097.8237.311.0
AGRoL[6]7.21100.9370.321.1
AvatarJLM[55]6.4291.7177.47.2
S2Fusion*[35]6.6589.2219.412.5
EnvPoser6.0074.7174.06.6
GIMO
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]7.0291.3324.016.4
AGRoL[6]6.5888.6269.412.5
AvatarJLM[55]4.9570.7258.110.7
S2Fusion[35]4.6557.8235.710.1
EnvPoser4.3857.6234.68.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.562, + 0.427, + 0.575 + ], + "angle": 0, + "content": "Table 1. The performance comparison with SOTAs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Metrics. Following [35, 55], we use these metrics for evaluation: 1) MPJRE \\([^\\circ]\\): The mean joint rotational error across all body joints, measured in degrees; 2) MPJPE \\([cm]\\): The mean positional error per joint, measured in centimeters. 3) MPJVE \\([cm/s]\\): Represents the average error in joint velocities, expressed in centimeters per second; 4) Jitter: Measures motion smoothness by calculating the average jerk (rate of change of acceleration) across all joints, with lower values indicating smoother motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.403, + 0.755 + ], + "angle": 0, + "content": "4.1. Comparison with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Quantitative Experimental Analysis. We compare our proposed method with the competing methods on the two public motion-environment interaction datasets. Among these, AvatarPoser [14], AGRoL [6], and AvatarJLM [55] are the SOTAs that rely solely on sparse tracking signals. As shown in Tab. 1, our model, which incorporates pre-scanned environmental point cloud information, effectively constrains the uncertainty in motion reconstruction caused by sparse tracking signals. Compared to Avatar-" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.359, + 0.596, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.461, + 0.596, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.562, + 0.583, + 0.572 + ], + "angle": 0, + "content": "AvatarJLM" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.36, + 0.691, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.461, + 0.691, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.562, + 0.674, + 0.572 + ], + "angle": 0, + "content": "S2Fusion*" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.36, + 0.787, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.461, + 0.787, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.562, + 0.767, + 0.572 + ], + "angle": 0, + "content": "EnvPoser" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.36, + 0.882, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.461, + 0.882, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.562, + 0.847, + 0.571 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.579, + 0.862, + 0.593 + ], + "angle": 0, + "content": "Figure 4. Qualitative Comparison of Interaction Details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.716 + ], + "angle": 0, + "content": "JLM [55], which does not utilize environmental information, our method achieves significant improvements on two key metrics—MPJRE and MPJPE—by 0.42 (6.5%)/17.0 (18.5%) and 0.57 (11.5%)/13.1 (18.5%) on the EgoBody and GIMO datasets, respectively. These results underscore the increased accuracy in motion estimation achieved through the integration of environmental information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Additionally, we benchmark our model against the latest algorithm S2Fusion [35], which introduces environmental information as a constraint for human motion estimation. On the GIMO dataset, our method reduces MPJRE errors by \\(5.8\\%\\) compared to S2Fusion [35], and it outperforms S2Fusion across all evaluation metrics, demonstrating the superior effectiveness of our approach in leveraging environmental information. Moreover, the motion estimation performance of our method also outperforms S2Fusion* on the EgoBody dataset, which reduces MPJRE and MPJPE errors by \\(9.8\\%\\) and \\(16.3\\%\\). Furthermore, our method shows a marked improvement in motion smoothness, with an \\(11.9\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1845" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.445, + 0.184 + ], + "angle": 0, + "content": "
EgoBodyGIMO
MPJREMPJPEMPJREMPJPE
①Baseline7.4587.15.7274.6
②w/o env-semantic6.0478.04.5562.5
③w/o env-geometry6.0377.54.6363.2
EnvPoser6.0074.74.3857.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.188, + 0.461, + 0.201 + ], + "angle": 0, + "content": "Table 2. The ablation study on environment refinement module." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.213, + 0.468, + 0.241 + ], + "angle": 0, + "content": "reduction in Jitter over S2Fusion and a \\(16.8\\%\\) improvement over AvatarJLM on the GIMO dataset." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.244, + 0.468, + 0.469 + ], + "angle": 0, + "content": "Qualitative Experimental Analysis. Fig. 3 showcases the motion estimation results on the EgoBody dataset, where we present three representative sequences and visualize the human motion estimation results at key moments. For state-of-the-art algorithms that do not incorporate environmental information, the visualization results in Fig. 3 show multiple instances of environment penetration during interactions. For example, AvatarJLM exhibits errors such as hand intersections with walls (second row) and leg penetrations through boxes (third row). Furthermore, in scenarios without chairs, squatting motions are misinterpreted as sitting (first row) in AvatarJLM due to the absence of environment understanding. In contrast, EnvPoser effectively leverages environmental information to reduce penetration errors and produce more realistic motion estimates." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.471, + 0.468, + 0.668 + ], + "angle": 0, + "content": "In more challenging interaction scenarios, EnvPoser shows certain advantages over S2Fusion, which also incorporates environmental information. Notably, S2Fusion fails to prevent interpenetration between the lower body and surrounding objects (third row in Fig. 3). Moreover, Fig. 4 provides a comparison between EnvPoser and S2Fusion, highlighting the advantages of our approach in handling upper-body interactions with objects and complex lower-body motions through the integration of uncertainty modeling and environmental semantic/geometric constraints. Additionally, compared to AvatarJLM, both EnvPoser and S2Fusion demonstrate the effectiveness of incorporating environmental information to improve human motion estimation." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.68, + 0.228, + 0.696 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Effectiveness of the environment-aware refinement module. We conducted ablation studies comparing our method with three variants: ① Baseline: an auto-regressive transformer-based network with joint uncertainty estimation; ② w/o env-semantic: incorporating only geometric constraints; ③ w/o env-geometry: incorporating only semantic constraints. Tab. 2 presents the motion estimation performance for each variant. Compared to variant ①, our method achieves substantial improvements by integrating both semantic and geometric environmental information. Further analysis shows that each environment-aware refinement module independently enhances performance, with the combination yielding the best overall results." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.905, + 0.171 + ], + "angle": 0, + "content": "
EgoBodyGIMO
MPJREMPJPEMPJVEMPJREMPJPEMPJVE
1)Baseline w/o UNC7.6390.8333.05.9077.5270.2
2)Baseline7.4587.1342.95.7274.6266.7
3)EnvPoser w/o UNC6.5680.1172.74.4560.2278.6
EnvPoser6.0074.7174.04.3857.6234.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.177, + 0.862, + 0.19 + ], + "angle": 0, + "content": "Table 3. The effectiveness of the uncertainty estimation." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.199, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Effectiveness of the uncertainty estimation. As shown in Tab. 3, we compare two variants: 1) Baseline w/o UNC: a baseline without uncertainty estimation and 2) Baseline: a baseline with uncertainty estimation. Results show that integrating the uncertainty estimation module not only quantifies joint uncertainty across motions but also improves full-body motion estimation through additional supervision. We further incorporated both baseline methods into the environment-aware refinement module. As shown in Tab. 3, our method achieves significant improvements over variant 3) EnvPoser w/o UNC: our method without uncertainty estimation, reducing MPJRE and MPJPE errors by \\(8.11\\%\\) and \\(5.68\\%\\), respectively, on the EgoBody dataset. These results indicate that the uncertainty estimation and resampling methods effectively capture the multi-hypothesis nature of initial motion estimates and, when combined with environmental refinement, adaptively align motion estimation with sparse tracking signals and environmental context." + }, + { + "type": "title", + "bbox": [ + 0.503, + 0.483, + 0.625, + 0.498 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.509, + 0.892, + 0.718 + ], + "angle": 0, + "content": "This paper presents a novel two-stage framework for reliable human motion estimation leveraging sparse tracking signals and environmental context. In the first stage, we model the multi-hypothesis nature of motion through uncertainty estimation, followed by a refinement phase that applies environment-aware constraints from both semantic and geometric perspectives. Extensive experiments show that EnvPoser outperforms existing methods, particularly in complex interaction scenarios, highlighting the significance of environmental context in wearable motion estimation and paving the way for future approaches to harness this valuable contextual information. This work has the potential to impact a wide range of interactive applications that benefit from more accurate and context-aware motion estimation." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Limitations and future works. The model assumes a static environment, lacking consideration for dynamic multi-user interactions or object motion, which may limit its effectiveness. This limitation could be addressed by incorporating a third-person perspective to estimate the movement of other people and objects, along with additional semantic and geometric constraints. Furthermore, real-time mesh quality in complex environments is often limited, which may affect performance. Future work will explore using raw images to infer contacts in complex scenes via 2D semantic understanding and introducing additional constraints to improve overall reconstruction quality." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.947, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1846" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Sadegh Aliakbarian, Fatemeh Saleh, David Collier, Pashmina Cameron, and Darren Cosker. Hmd-nemo: Online 3d avatar motion generation from sparse observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9622-9631, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[3] Rongyu Chen, Linlin Yang, and Angela Yao. Mhentropy: Entropy meets multiple hypotheses for pose and shape recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14840-14849, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.48 + ], + "angle": 0, + "content": "[5] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating smpl poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[6] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.632 + ], + "angle": 0, + "content": "[7] Zhen Fan, Peng Dai, Zhuo Su, Xu Gao, Zheng Lv, Jiarui Zhang, Tianyuan Du, Guidong Wang, and Yang Zhang. Emhi: A multimodal egocentric human motion dataset with hmd and body-worn imus. arXiv preprint arXiv:2408.17168, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[8] Han Feng, Wenchao Ma, Quankai Gao, Xianwei Zheng, Nan Xue, and Huijuan Xu. Stratified avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 153-163, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[11] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[13] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: learning to reconstruct human pose from sparse inertial measurements in real time. ACM Trans. Graph. (TOG), 37(6):1-15, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.894, + 0.348 + ], + "angle": 0, + "content": "[14] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[15] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time egocentric pose estimation from sparse and intermittent observations everywhere. In European Conference on Computer Vision, pages 277-294. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[16] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.894, + 0.586 + ], + "angle": 0, + "content": "[17] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers (SA' 22), pages 1-9, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[18] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision(ICCV), pages 11510-11520, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[19] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.894, + 0.786 + ], + "angle": 0, + "content": "[20] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[21] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[22] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1847" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[23] Han Li, Bowen Shi, Wenrui Dai, Hongwei Zheng, Botao Wang, Yu Sun, Min Guo, Chenglin Li, Junni Zou, and Hongkai Xiong. Pose-oriented transformer with uncertainty-guided refinement for 2d-to-3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1296–1304, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[24] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[27] Zhenyu Lou, Qiongjie Cui, Haofan Wang, Xu Tang, and Hong Zhou. Multimodal sense-informed forecasting of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2144-2154, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.472, + 0.517 + ], + "angle": 0, + "content": "[28] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[29] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. Coap: Compositional articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13201-13210, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[30] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[31] Jose Luis Ponton, Eduard Pujol, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Dragposer: Motion reconstruction from variable sparse tracking signals via latent space optimization. arXiv preprint arXiv:2406.14567, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[33] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[34] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[35] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[36] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusionposer: Real-time human motion reconstruction from arbitrary sparse sensors using autoregressive diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2513-2523, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[38] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11199-11208, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[39] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[40] Yinghao Wu, chaoran wang, Lu Yin, Shihui Guo, and Yipeng Qin. Accurate and steady inertial pose estimation through sequence structure learning and modulation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[41] Songpengcheng Xia, Lei Chu, Ling Pei, Zixuan Zhang, Wenxian Yu, and Robert C Qiu. Learning disentangled representation for mixed-reality human activity recognition with a single imu sensor. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[42] Songpengcheng Xia, Lei Chu, Ling Pei, Jiarui Yang, Wenxian Yu, and Robert C Qiu. Timestamp-supervised wearable-based activity segmentation and recognition with contrastive learning and order-preserving optimal transport. IEEE Transactions on Mobile Computing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[43] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–989, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[44] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22268-22278, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[45] Hongdi Yang, Chengyang Li, Zhenxuan Wu, Gaozheng Li, Jingya Wang, Jingyi Yu, Zhuo Su, and Lan Xu. Smgdiff: Soccer motion generation using diffusion probabilistic models. arXiv preprint arXiv:2411.16216, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Xueyuan Yang, Chao Yao, and Xiaojuan Ban. Spatial-related sensors matters: 3d human motion reconstruction assisted with textual semantics. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10225-10233, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1848" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[47] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[48] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[49] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), pages 13167-13178, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[50] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[51] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision, pages 180-200. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[52] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[53] Yufei Zhang, Hanjing Wang, Jeffrey O Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9032, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.586, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[54] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.739 + ], + "angle": 0, + "content": "[55] Xiaozheng Zheng, Zhuo Su, Chao Wen, Zhou Xue, and Xiaojie Jin. Realistic full-body tracking from sparse observations via joint-level modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14678-14688, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.741, + 0.469, + 0.81 + ], + "angle": 0, + "content": "[56] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision, pages 676-694. Springer, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1849" + } + ] +] \ No newline at end of file diff --git a/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_origin.pdf b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a99839a95c6e00958d85dd5ee65ce4626151f9a --- /dev/null +++ b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/ca9b40de-cd3a-44e8-85b9-dec7c06efbee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16c8f98fdcdcece7e0f4bc3a7e1e4f31f8dc2c272a4cbcc38f9ae565377fa426 +size 4849564 diff --git a/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/full.md b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2d9b7ef3f56bdd7994ca2960b20098ac0869a7a2 --- /dev/null +++ b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/full.md @@ -0,0 +1,370 @@ +# EnvPoser: Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling + +Songpengcheng Xia $^{1}$ Yu Zhang $^{1}$ Zhuo Su $^{2*}$ Xiaozheng Zheng $^{2}$ Zheng Lv $^{2}$ Guidong Wang $^{2}$ Yongjie Zhang $^{2}$ Qi Wu $^{1,2}$ Lei Chu $^{3}$ Ling Pei $^{1\dagger}$ Shanghai Jiao Tong University ${}^{2}\mathrm{ByteDance}$ ${}^{3}\mathrm{University}$ of Southern California + +# Abstract + +Estimating full-body motion using the tracking signals of head and hands from VR devices holds great potential for various applications. However, the sparsity and unique distribution of observations present a significant challenge, resulting in an ill-posed problem with multiple feasible solutions (i.e., hypotheses). This amplifies uncertainty and ambiguity in full-body motion estimation, especially for the lower-body joints. Therefore, we propose a new method, EnvPoser, that employs a two-stage framework to perform full-body motion estimation using sparse tracking signals and pre-scanned environment from VR devices. EnvPoser models the multi-hypothesis nature of human motion through an uncertainty-aware estimation module in the first stage. In the second stage, we refine these multi-hypothesis estimates by integrating semantic and geometric environmental constraints, ensuring that the final motion estimation aligns realistically with both the environmental context and physical interactions. Qualitative and quantitative experiments on two public datasets demonstrate that our method achieves state-of-the-art performance, highlighting significant improvements in human motion estimation within motion-environment interaction scenarios. Project page: https://xspc.github.io/EnvPoser/. + +# 1. Introduction + +In recent years, motion capture algorithms using sparse tracking signals have achieved significant progress across various research fields, establishing themselves as essential technologies for applications such as motion-sensing games and immersive VR/AR experiences [13, 18, 42, 49]. With the rapid development of VR devices, full-body motion estimation algorithms based on head-mounted display (HMD) have shown great application potential. However, devices + +![](images/2fc829f832043b95837aaa864a6bf54e4c8e93fb0ba447c596d072c32f298c89.jpg) +Figure 1. EnvPoser can estimate the full-body motion using three tracking signals (HMD and hand controllers) and a pre-scanned environment mesh. + +like PICO and Quest typically provide sparse tracking signals of the head and hands, presenting challenges in estimating the full-body motion from these limited observations. + +Current sparse-observation-based human motion estimation methods typically rely on large-scale motion capture datasets, such as AMASS [28], to capture full-body motion through regression [14, 55] or generative models [2, 6, 8]. Alternatively, physics engines and reinforcement learning have been used [39] to synthesize full-body motion, bypassing dataset dependency. Utilizing various advanced architectures or large-scale datasets, these methods [2, 6, 8, 39, 55] attempt to model the mapping from sparse tracking signals to the most probable full-body motion. However, the lack of direct joint observations results in the same sparse input often corresponding to multiple plausible motions, making it a challenge for sparse-observation- + +based motion estimation algorithms to converge on the most probable full-body motion. To address this one-to-many mapping ambiguity, existing methods [4, 30, 31] attempt to add sensors to regions such as the pelvis and legs to narrow the distribution of possible motion outcomes. Although additional sensors can enhance the accuracy and robustness of full-body motion estimation, this approach may negatively impact user experience due to increased hardware requirements [13, 17, 41, 49, 54]. + +Therefore, an alternative approach to addressing the inherent ambiguity in sparse observations is to incorporate additional environmental constraints, as shown in Fig. 1. Human motion is highly correlated with the surrounding environment, which contains rich context that can provide valuable cues for motion estimation. However, effectively utilizing environmental constraints presents its own challenges, as existing methods [35, 55] often simplify human-environment interactions to basic foot-to-ground contact, overlooking the complexities of interactions. In reality, individuals frequently engage in more complex interactions with their surroundings. + +Our key insight is that joint uncertainty estimation can explicitly model the multi-hypothesis nature of full-body motion estimation, and by incorporating environmental information, we can reduce estimation uncertainty, guiding the motion estimation to converge toward the most plausible outcome that aligns with both the environment and sparse input observations. Therefore, we design a two-stage deep learning framework to handle the one-to-many mapping ambiguity challenge. We first explicitly model the multi-hypothesis human motion by the joint uncertainty estimation and its sampling strategy. Then, by integrating environmental information, we guide the uncertainty-based multi-hypothesis motion estimations to converge on a full-body motion that best aligns with both the environment and sparse observations. To address the challenge of effectively utilizing environmental information, we introduce semantic constraints for non-contact motions, providing contextual cues for realistic estimation, and geometric constraints for contact scenarios, ensuring consistency and preventing collisions. + +The key contributions of this paper are as follows: + +- We propose EnvPoser, a novel framework that significantly advances full-body motion estimation from sparse tracking signals by incorporating pre-scanned environmental context. This framework addresses the complexities of estimating human motion from sparse observations, ensuring robustness across diverse interactive scenarios. +- This work presents a two-stage framework for human motion estimation: an uncertainty-aware module provides initial estimates that account for multi-hypothesis motion, while an environment-aware refinement module in + +tegrates semantic and geometric constraints to ensure realistic outcomes in diverse scenarios, setting a new benchmark in motion estimation with environmental interactions. +- Comprehensive quantitative and qualitative experiments demonstrate that our proposed method achieves state-of-the-art performance across two public datasets and excels in scenarios with frequent environmental interactions. + +# 2. Related Work + +# 2.1. Motion estimation with sparse observations + +Current Wearable-based human motion estimation approaches mainly adopt two configurations: one uses multiple inertial measurement units (IMUs) placed on body extremities [13, 17, 36, 40, 48, 54], while the other relies on an HMD with two controllers [2, 4, 7, 14, 15, 55]. + +While methods utilizing six IMUs have shown impressive results, HMD-based techniques leveraging sparse three-point observations show significant potential for practical applications. For instance, AvatarPoser [14] employs a transformer architecture to achieve accurate full-body motion. While Zheng et al. [55] developed a two-stage framework to capture joint dependencies, improving motion accuracy and fluidity. Some studies [1, 5, 8] address the ill-posed problem of sparse-to-dense mapping using generative models. Dittadi et al. [5] and Aliakbarian et al. [1] proposed using Variational Autoencoders (VAE) and normalizing flows, respectively, to estimate full-body poses from three tracking points. Additionally, the diffusion model [6, 8, 35] has also shown notable performance in motion estimation with sparse data. Despite these advances, most methods rely solely on sparse tracking signals, leading to high uncertainty and multiple potential poses for the same input. Therefore, this paper introduces joint uncertainty estimation and constrains physically implausible motions using environmental information. + +# 2.2. Pose regression with uncertainty estimation + +The sparse nature of observations in 3D human pose estimation introduces significant uncertainty for unobserved joints, as identical inputs can correspond to multiple pose hypotheses [8, 54, 55]. This challenge is particularly evident in image-based pose estimation [3, 9, 22, 24, 43], where monocular methods face depth ambiguity. For instance, Wehrbein et al. [38] modeled the posterior distribution of 3D pose hypotheses using normalizing flows, while MHFormer [24] employed a multi-hypothesis transformer to capture diverse pose hypotheses through a one-to-many mapping. Recently, Shan et al. [33] proposed diffusion-based methods to aggregate multi-hypothesis predictions with joint-wise re-projection. + +In our paper, we aim to use uncertainty estimation + +to model the multi-hypothesis nature of motion from sparse observations. Uncertainty in deep learning can be categorized into two types [19]: aleatoric (data-related) and epistemic (model-related). Li et al. [23] refined 3D poses by incorporating joint uncertainties through uncertainty guided-sampling and uncertainty-guided self-attention, while Zhang et al. [53] introduced a probabilistic framework that encodes aleatoric uncertainty using a robust negative log-likelihood loss, with epistemic uncertainty guiding model refinement. In wearable motion estimation, Yang et al. [46] leveraged uncertainty to weight IMU-derived features, enhancing pose corrections with additional text annotations. + +# 2.3. Scene-aware motion estimation and generation + +Recent advancements in human motion estimation and generation have expanded from solely analyzing body motion to integrating interactions with the surrounding environment, especially vision-based approaches [11, 12, 16, 34, 47, 52]. For instance, PROX [10] incorporated scene constraints into monocular human pose estimation to address depth ambiguity, while Shen et al. [34] used sparse 3D CNNs to estimate absolute positions and dense scene contacts, refining human mesh recovery via cross-attention with 3D scene cues. EgoHMR [52] leveraged ego-camera images and diffusion models within pre-scanned environments to recover full-body motion. Similarly, S2Fusion [35] utilized sparse observations by learning human motion priors on synthetic datasets and conditioned motion generation on scene encoding and VR inputs. However, S2Fusion primarily focuses on lower body and environment interactions. QuestEnvsim [21] employed reinforcement learning to generate realistic poses in highly constrained environments using HMDs. Additionally, recent studies further emphasize the integration of scene context and historical motion to produce plausible human motion [27, 44, 45]. + +These scene-aware methods underscore the importance of environmental context for enhancing motion estimation. Our approach builds on these foundations by comprehensively leveraging human-environment interactions, addressing both contact and non-contact scenarios to improve motion estimation with sparse tracking signals. + +# 3. Method + +# 3.1. Problem Statement + +Our goal is to estimate the full-body motion sequence $\theta = \{\theta_t\}_{t=1}^T \in \mathcal{R}^{T \times 132}$ (the first 22 joints of SMPL [26] model with the 6D representation of rotations) using the sparse tracking signals $\mathbf{X} = \{\mathbf{x}_t\}_{t=1}^T \in \mathcal{R}^{T \times N_c}$ from the HMD and hand controllers, along with a pre-scanned 3D environment point cloud $\mathbf{S}$ , where $T$ denotes the length of input, $N_c = 36$ means the channel of input. Each sparse tracking + +signal at time step $t$ consists of the head and hand positions $\pmb{p}_t$ , rotations $\pmb{R}_t$ , and linear velocities $\pmb{v}_t$ , represented as $\pmb{x}_t = [\pmb{p}_t, \pmb{R}_t, \pmb{v}_t]$ . Moreover, we utilize a cropped 3D environment point cloud $V_S \in \mathcal{R}^{N_S \times 3}$ , where $N_S$ , the number of sampled points, is set to 1000. With the human translation, this point cloud refines pose estimation by incorporating environmental context, leveraging both contact and noncontact interactions between the body and the environment to reduce the uncertainty inherent in sparse observations. + +As shown in Fig. 2, our approach is a two-stage model comprising two core technical modules: 1) an uncertainty-aware initial motion estimation module; and 2) an environment-aware motion refinement module that leverages semantic and geometric information from the environment. We first train the uncertainty-aware initial motion estimation module (Stage I) on the AMASS dataset, with a focus on explicitly capturing the multi-hypothesis nature of motion estimation through uncertainty quantification. Subsequently, we jointly train both modules (Stage II) on the motion-environment interaction datasets, applying environmental semantic and geometric constraints to refine the multi-hypothesis estimates from Stage I into the most plausible motion estimation. + +# 3.2. Uncertainty-aware Initial Human Motion Estimation Module + +To obtain the joint uncertainty quantification and explicitly model the multi-hypothesis nature of motion estimation, we propose an uncertainty-aware initial human motion estimation module. This module is designed with three primary objectives: (1) to integrate historical motion states and sparse observations for the current initial motion estimation; (2) to effectively extract features from sparse signals; and (3) to quantify uncertainty in motion estimation, explicitly modeling the multi-hypothesis nature of motion under identical input conditions. Accordingly, this module comprises three sub-modules: sparse input signal and historical motion embedding, transformer-based motion feature extraction, and human motion and joint uncertainty regression. + +Sparse Input Signal and Historical Human Motion Embedding. To integrate historical motion states and sparse observations for the current initial motion estimation, we adopt an auto-regressive structure in the uncertainty-aware initial human motion estimation module. Our model takes sparse observations $\mathbf{X} \in \mathcal{R}^{T \times 36}$ and historical motion $\mathbf{X}_{hm} \in \mathcal{R}^{T \times 132}$ as inputs, embedding these inputs through linear layers to obtain the observation representation $Z_{X}$ and motion historical state representation $Z_{hm}$ . Then, we concatenate the sparse observation embedding and historical motion embedding to obtain the shallow motion representation $Z_{S} = [Z_{X}, Z_{hm}]$ , which can be used for subsequent transformer-based feature extraction. + +Transformer-based Motion Feature Extraction. + +![](images/7868b37cd6581f6594f5d4bf64e84d642a1ad73fc868415e5640199603cefbc6.jpg) +Figure 2. Overview of EnvPoser: A Two-Stage Motion Estimation Model. Stage I involves training the uncertainty-aware initial estimation module on the AMASS dataset to produce initial motion estimates with uncertainty quantification. Stage II refines these estimates by training on motion-environment datasets, incorporating semantic and geometric environmental constraints. + +Based on the shallow motion representation $Z_{S}$ , we employ a transformer-based feature extraction module [17] for motion features learning. To retain sequence order information, temporal positional encoding is applied to the shallow motion representation [37]. Using multi-head self-attention and a position-wise feed-forward network within the transformer encoder $F_{t}$ , this module efficiently captures intricate relationships in the sparse tracking signals and historical motion, resulting in high-dimensional motion features $Z_{H} = F_{t}(Z_{S})$ . While the transformer architecture is effective for extracting motion features, the sparsity of the input data—limited to tracking signals from the head and hands—introduces high uncertainty in estimating unobserved joints, particularly in the lower body. This uncertainty arises from the lack of direct observations for these joints, making their estimation inherently ambiguous. To address this challenge, we incorporate a mechanism to quantify the uncertainty in joint estimations. + +Human Motion and Joint Uncertainty Regression. With the high-dimensional motion features $Z_{H}$ , we employ a pose regression network $F_{ph}(\cdot)$ to predict the human motion $\widetilde{\theta} = F_{ph}(Z_H)$ . Given the sparse nature of the input signals, capturing joint uncertainty is crucial in mapping the limited observations to a full-body motion estimation. To achieve this, we utilize a heteroscedastic neural network [19, 23] to estimate the joint uncertainty. Specifically, an uncertainty regression head $F_{uh}(\cdot)$ is employed on the high-dimensional motion feature to estimate the joint uncertainty $\delta \in \mathcal{R}^{T\times 132}$ of the predicted motion. + +Using the predicted motion and its associated uncertainty, we resample the full-body motion $\overline{\theta}$ around the predicted mean motion $\widetilde{\theta}$ according to a Gaussian distribution + +$\mathcal{N}(\widetilde{\boldsymbol{\theta}},\delta)$ with the uncertainty $\delta$ . This resampling technique introduces a multi-hypothesis representation, characterizing the inherent ambiguity in the unobserved joints. To ensure differentiability throughout the model training process, we apply the reparameterization trick on this sampling operation, drawing a parameter $\epsilon$ from a standard Gaussian distribution $\mathcal{N}(0,1)$ , represented by $\overline{\theta} = \widetilde{\theta} +\delta \cdot \epsilon$ . + +Through the uncertainty regression head, we can obtain the joint uncertainty estimation for predicted human motion, enabling a multi-hypothesis motion representation with the sampling strategy. We pre-train the first stage of our method on the large synthetic dataset AMASS [28]. Based on the predicted mean human motion $\widetilde{\theta}$ and uncertainty $\delta$ , we train the first stage module with the following loss function: + +$$ +L _ {M} = \| \widetilde {\boldsymbol {\theta}} - \boldsymbol {\theta} \| _ {2}, \tag {1} +$$ + +where $\theta$ represents the ground-truth of human motion. Additionally, to ensure accurate uncertainty prediction, we also incorporate an uncertainty estimation loss as [19] stated: + +$$ +L _ {\delta} = \left\| \frac {\widetilde {\boldsymbol {\theta}} - \boldsymbol {\theta}}{\delta} \right\| _ {2} + \log (\left\| \boldsymbol {\delta} \right\| _ {2}). \tag {2} +$$ + +Thus, based on the uncertainty-aware initial human motion module, the objective function for Stage I is defined as follows: + +$$ +L _ {S - I} = \lambda_ {M} L _ {M} + \lambda_ {\delta} L _ {\delta}, \tag {3} +$$ + +where the $\lambda_{M}$ and $\lambda_{\delta}$ are the hyper-parameters for Stage I training. + +Training Stage I on the AMASS dataset, along with the design for motion uncertainty quantification, allows our network to effectively capture the multi-hypothesis nature of + +motion estimation arising from sparse observations. To address the inherent one-to-many ambiguity, we introduce environmental information as a refinement stage. By incorporating both semantic and geometric environmental information, our method guides the multi-hypothesis motion estimates to converge toward the solution that best aligns with the sparse observations and environmental constraints. + +# 3.3. Environment-aware Human Motion Refinement with Semantic and Geometry Constraint + +Based on the initial motion estimation and its uncertainty produced by the first-stage uncertainty-aware module, we further refine the multi-hypothesis human motion estimates using the pre-scanned environment point clouds collected by real-world devices. To effectively utilize the environmental information, our design is guided by a key insight: human motion within an environment involves both object interactions (contact scenarios) and collision avoidance (non-contact scenarios). In this module, the environment point clouds are processed through cropping, sampling, and encoding. We then apply constraints based on the environment's semantic and geometric properties to ensure that the full-body motion estimation realistically aligns with the surrounding context. + +Environmental Point Cloud Embedding. To enhance computational efficiency and focus on relevant interaction areas, we first clip the environmental point clouds based on the human's global position. A circular sampling approach is used to minimize the impact of body orientation, selecting a 1-meter radius around the body and uniformly distributing $N_{S}$ points within this area, denoted as $\mathbf{V}_{\mathbf{S}} \in \mathcal{R}^{N_{S} \times 3}$ . The cropped environmental point cloud $\mathbf{V}_{\mathbf{S}}$ is then processed by the environment encoder $F_{env}(\cdot)$ to obtain the environmental point cloud embedding $Z_{env}$ , represented by: + +$$ +\boldsymbol {Z} _ {\text {e n v}} = F _ {\text {e n v}} \left(\boldsymbol {V} _ {\boldsymbol {S}}\right), \tag {4} +$$ + +where the environment encoder $F_{env}(\cdot)$ is implemented using the vanilla PointNet++ [32]. The extracted environmental embedding would be used in the environment semantic-aware motion refinement. + +Environment Semantic-Aware Constraint for Motion Refinement. Capturing semantic information from the environmental point cloud provides crucial guidance for motion generation [27, 44, 47], as human motion is highly correlated with environmental context. Therefore, leveraging this motion-environment correlation as a valuable supplement has the potential to enhance the sparse-observation-based human motion estimation. Inspired by [27, 34], we employ a cross-attention module between the initial motion estimates and the cropped environmental point clouds to refine the human motion estimation. This module comprises an environment-embedding network, a motion information embedding network and an environment-motion + +cross-attention network. + +The motion information embedding network $F_{motion}(\cdot)$ takes the sampled motion estimation $\widetilde{\pmb{\theta}}$ , head translation $\pmb{p}^{head}$ and the extended sparse observations $X_{new} = \{X, h, \pmb{\theta}^{up}\} \in \mathcal{R}^{T \times 40}$ as input. The extended sparse observations $X_{new}$ include the relative head height $h$ and head up-vector $\pmb{\theta}^{up}$ , obtained via a relative head height calculation module. Following [17, 20], this module calculates terrain height by using the height of the feet when stationary and nearby environmental points, and then determines $h$ as the difference between the absolute head height from the HMD and the estimated terrain height. These inputs are concatenated and processed through linear layers to produce the motion information embedding $Z_M = F_{motion}(\text{concat}(\overline{\theta}, \pmb{p}^{head}, X_{new}))$ . + +For the environment-motion cross-attention network, the Key $K$ and Value $V$ vectors can be obtained by multiplying the weight matrix $W_{k}, W_{v}$ with the environmental embedding $Z_{env}$ , while the motion information embedding $Z_{M}$ multiplies the weight matrices $W_{q}$ to get the Query $Q$ vectors. Inspired by [25, 27], we incorporate a spatial salience $s_{\text{spatial}} = MLP(V_{S})$ that expects our cross-attention model to be able to perceive the distance between environmental points clouds and human bodies. Using the spatial salience $s_{\text{spatial}}$ in conjunction with attention scores $\text{Attn}(Q, K)$ , we derive a motion-environment representation $Z_{ME}$ , which could be represented by + +$$ +\boldsymbol {Z} _ {M E} = \left(\operatorname {A t t n} (\boldsymbol {Q}, \boldsymbol {K}) + \boldsymbol {s} _ {\text {s p a t i a l}}\right) \cdot \boldsymbol {V}. \tag {5} +$$ + +Then, by concatenating the motion-environment representation $Z_{ME}$ with the motion information embedding $Z_{M}$ and passing them through an MLP network, we obtain the environment-refined motion representation $Z_{RM} = MLP(\text{concat}(Z_{ME}, Z_{M}))$ . + +Using the environment-refined motion representation $Z_{RM}$ , which integrates environmental semantics and spatial priors between the human and environment, the model can then regress the probability of human-environment contact and generate the final motion estimation. + +Motion Regression with Contact Estimation. With the environment-refined motion representation $Z_{RM}$ , we estimate the contact probability $C = \{c\}_{t=1}^{T} \in \mathcal{R}^{T \times 22}$ and regress the final human motion $\hat{\theta}_{RM}$ . We take the environment-refined motion representation $Z_{RM}$ and the extended sparse observations $X_{new}$ as input, and apply two-layers MLP layers on the concatenated input to get the contact probability prediction $\hat{C} = MLP(concat(X_{new}, Z_{RM}))$ . + +Contact probability prediction is typically treated as a binary classification problem, where Binary Cross Entropy (BCE) loss is used for training: + +$$ +L _ {\text {c o n t a c t}} = B C E L o s s (\hat {\boldsymbol {C}}, \boldsymbol {C}). \tag {6} +$$ + +Finally, the contact probability $\hat{C}$ and environment-refined motion representation $Z_{RM}$ are concatenated to regress the final motion with a motion decoder network $F_{out}(\cdot)$ : + +$$ +\hat {\boldsymbol {\theta}} _ {R M} = F _ {o u t} (\operatorname {c o n c a t} (\mathbf {Z} _ {R M}, \hat {\mathbf {C}})), \tag {7} +$$ + +where $F_{out}(\cdot)$ consists of two linear layers with a ReLU activation function. We then obtain the final motion loss: + +$$ +L _ {M ^ {\prime}} = \left\| \hat {\boldsymbol {\theta}} _ {R M} - \boldsymbol {\theta} \right\| _ {2}. \tag {8} +$$ + +Scene Geometry-Aware Constraint for Motion Refinement. To complement semantic-level constraints, we leverage the geometric structure of the environment to enhance motion estimation accuracy, especially in human-environment interactions (contact scenarios). A straightforward approach is to detect collisions between the estimated human motion and the environment point cloud, thereby refining unrealistic motions. While existing methods [10, 50] often convert 3D environments into Signed Distance Fields (SDF) for collision detection, this approach can be computationally intensive. Inspired by recent advancements [29, 52], we instead employ the COAP model [29] to efficiently compute collisions between the human body and environmental points. With this novel model and the environment point clouds, the COAP-based collision loss [52] can be formulated as: + +$$ +L _ {c o a p} = \frac {1}{N _ {S}} \sum_ {i = 1} ^ {N _ {S}} \sigma \left(f _ {\Theta} \left(V _ {S _ {i}} \mid \mathcal {G}\right)\right) \mathbb {I} _ {f _ {\Theta} \left(V _ {S _ {i}} \mid \mathcal {G}\right) > 0}, \tag {9} +$$ + +where $V_{S}$ denotes the cropped environment point clouds, with $N_{S}$ representing the number of points. $\Theta$ represents the transformation parameters of the human body, encompassing both the motion pose $\pmb{\theta}$ and shape $\beta$ . $f_{\Theta}(V_{S_i}|\mathcal{G})$ represents the spatial relationship between the human body model and environmental points. The function $\sigma(\cdot)$ is the Sigmoid function, which maps the value of $f_{\Theta}(V_{S_i}|\mathcal{G})$ to a range between 0 and 1. + +Since most motions occur on the ground, we remove ground points from the scene to ensure that the COAP-based collision loss specifically targets interactions with surrounding objects. To account for interactions with the ground, we introduce additional constraints, including foot contact, foot height, and ground penetration losses [55]: + +$$ +L _ {f c} = \left\| \left(\hat {\boldsymbol {P}} _ {R M} ^ {f e e t} - \boldsymbol {P} ^ {f e e t}\right) \cdot \boldsymbol {C} \right\| _ {1}, +$$ + +$$ +L _ {g f h} = \left\| \hat {\boldsymbol {z}} _ {P _ {R M}} ^ {f e e t} - \boldsymbol {z} ^ {g r o u n d} \right\| _ {1}, \tag {10} +$$ + +$$ +L _ {g p} = \left\| \left(\hat {z} _ {P _ {R M}} ^ {\text {m i n}} - z ^ {\text {g r o u n d}}\right) \cdot \boldsymbol {l} \right\| _ {1}, +$$ + +where the $\hat{z}_{P_{RM}}^{feet}$ and $z_{ground}$ means the height of feet joint positions and the floor height, and $l$ denotes whether the joint is lower than the ground. + +Final Loss Function. In addition to the first stage loss (Eq.2, Eq.1), COAP-based collision loss (Eq.9), contact loss (Eq.6), foot-ground loss (Eq.10) and final motion loss (Eq.8), we also apply some additional losses to improve model training. With the forward kinematic chain, the joint position $P$ , $\hat{P}_{RM}$ , and the global hands positions $P_{hand}^{\prime}$ , $\hat{P}_{RM}^{\text {hand }}$ could be calculated by the motion pose $\theta$ and $\hat{\theta}$ . To penalize the errors accumulating along the kinematic chain, we apply a joint position loss $(L_{\text {posi}})$ and a hand alignment loss $(L_{hAL})$ : + +$$ +L _ {\text {p o s i}} = \left\| \hat {\boldsymbol {P}} _ {R M} - \boldsymbol {P} \right\| _ {2}, L _ {h A L} = \left\| \hat {\boldsymbol {P}} _ {R M} ^ {\text {h a n d}} - \boldsymbol {P} ^ {\text {h a n d}} \right\| _ {1}. \tag {11} +$$ + +Therefore, with the technical analysis provided above, we can formulate the second stage objective function as follows: + +$$ +\begin{array}{l} L _ {S - I I} = L _ {S - I} + L _ {M ^ {\prime}} + \lambda_ {1} L _ {p o s i} + \lambda_ {2} L _ {h A L} + \lambda_ {3} L _ {f c} \\ + \lambda_ {4} L _ {\text {c o n t a c t}} + \lambda_ {5} L _ {g f h} + \lambda_ {6} L _ {g p} + \lambda_ {7} L _ {\text {c o a p}}. \tag {12} \\ \end{array} +$$ + +Here, $\{\lambda_i\}_{i = 1,\dots ,7}$ represents the hyperparameters. Our model is first trained on the AMASS using the Eq. 3 and then fine-tuned on a motion-environment interaction dataset with Eq. 12. The training process and selection of hyperparameters can be found in the supplementary material. + +# 4. Experiments + +This section introduces the datasets, metrics, and competing algorithms used in our experiments. We then compare our method EnvPoser with state-of-the-art algorithms, highlighting its advantages in motion estimation. Finally, ablation studies validate the effectiveness of each module. Additional case studies and discussions are provided in the supplementary materials. + +Datasets. To comprehensively evaluate the efficacy of our proposed method, we conduct experiments on two challenging public motion-scene interactions datasets: Egobody [51] and GIMO [56] datasets. Detailed information about these datasets can be found in the supplementary document. + +Competing methods. We compare the motion estimation performance of the proposed model with the state-of-the-art competing methods, including AvatarPoser [14], AGRoL [6], AvatarJLM [55], and S2Fusion [35] on two public datasets. We re-train all competing methods on the EgoBody and GIMO datasets until convergence. For the results on the GIMO dataset, we reused the values reported in previous studies [35]. For the EgoBody dataset, we tested our trained model under real-time settings and presented the results as follows. It is worth noting that S2Fusion [35] is not fully open-source; therefore, we reproduced S2Fusion based on the details provided in its paper and partial open-source code, marking it as S2Fusion* in this experiment. + +![](images/fcc6b04db8a1b9fa5b5710fd5efd781ae7d071850727bb8ddaddb7d39f212ca6.jpg) + +![](images/eeccb6f85d0353a3add391a9684a75b6bc9a5befb4817815ee321c0b52d97148.jpg) + +![](images/a8df69a351ba0e5095e4051cacdffc2c14a4bc15fbeea277438ed3e7ba150f3f.jpg) +AvatarPoser + +![](images/1a9d94105bfb1daa52173b9419d47cb45d3a2eaeb5f5b7dd92b5d1ad0479ddbd.jpg) + +![](images/69f5bda580c60f6382a354848cba29ab4d04c233aaf6827b71b4c4f138a15b7a.jpg) + +![](images/013ef28dda0aa6516e2c7b7b2041a96e02a5bc50db6bf09f2d58c27cd5967f55.jpg) +AGRoL + +![](images/8e32435c167ea98cd5bede7271e9673202fd37f468b5aaa59d4cdcd679a0831a.jpg) + +![](images/c0b30636be4608df50ef55d4ebe1e69b7a348c909fe51f336c82246e82a86991.jpg) + +![](images/6e3bd1f9f176c5b124822d96098b04117c686403d55ea71857aebcd030a7f80a.jpg) +AvatarJLM + +![](images/2e01b1e12b18e5edaee2dfe4088cfaf58ffca40a127134ecbe8baf44e0696574.jpg) + +![](images/7ef11262d323e7f1ed946b4260bd898972639512d281c2a11b284f8f21805dd1.jpg) + +![](images/ef03fa587ae612799b7066941ae87a89a87c879bb9384ee38d1685d560e59d7e.jpg) +S2Fusion* + +![](images/4dd84c13ca8b0d4abc7cbfee2af35b369e456fce4a58343fb1e6670bd8999906.jpg) + +![](images/6d1ab4d43dfb3ee93fe6b801c8ca4073a9ce5ab4c0845312dfad667cb85fcfe6.jpg) + +![](images/53265b806f4c1dc67114f4ef9e89bb600d8ab26c67820249454b5e3f0957f04b.jpg) +EnvPoser + +![](images/c7548ddbc7ee007cc2b5d476cfc9907142e3039681ee606d1b5cfc18d862aabc.jpg) + +![](images/4406b8fd50c52b55be0afe8818029e09c08eb0701641b609339d5d6e07ef1ccb.jpg) + +![](images/e07d7e0960ee82bf32007e269e50eb65d143d2604ebc91b567b06f5eabcbf5fa.jpg) +GT + +
EgoBody
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]6.8097.8237.311.0
AGRoL[6]7.21100.9370.321.1
AvatarJLM[55]6.4291.7177.47.2
S2Fusion*[35]6.6589.2219.412.5
EnvPoser6.0074.7174.06.6
GIMO
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]7.0291.3324.016.4
AGRoL[6]6.5888.6269.412.5
AvatarJLM[55]4.9570.7258.110.7
S2Fusion[35]4.6557.8235.710.1
EnvPoser4.3857.6234.68.9
+ +Table 1. The performance comparison with SOTAs. + +Metrics. Following [35, 55], we use these metrics for evaluation: 1) MPJRE $[^\circ]$ : The mean joint rotational error across all body joints, measured in degrees; 2) MPJPE $[cm]$ : The mean positional error per joint, measured in centimeters. 3) MPJVE $[cm/s]$ : Represents the average error in joint velocities, expressed in centimeters per second; 4) Jitter: Measures motion smoothness by calculating the average jerk (rate of change of acceleration) across all joints, with lower values indicating smoother motion. + +# 4.1. Comparison with the State-of-the-Art + +Quantitative Experimental Analysis. We compare our proposed method with the competing methods on the two public motion-environment interaction datasets. Among these, AvatarPoser [14], AGRoL [6], and AvatarJLM [55] are the SOTAs that rely solely on sparse tracking signals. As shown in Tab. 1, our model, which incorporates pre-scanned environmental point cloud information, effectively constrains the uncertainty in motion reconstruction caused by sparse tracking signals. Compared to Avatar- + +![](images/5754a7bc84d97da9af2379af66d01e2a4bc34b3b1e82254311779b2bf66f2682.jpg) +Figure 3. Visualization of motion estimation on three test sequences from EgoBody Dataset [51]. + +![](images/26d7eb7f316c4f4b88e773443cf961a2d1eaae70ff1db42660224557dd3c6da8.jpg) +AvatarJLM +Figure 4. Qualitative Comparison of Interaction Details. + +![](images/1d8a775c966be2865842459aa367303da0d91332de7e016827a4dc4de5938bf5.jpg) + +![](images/eb3e45feca9a6e2ee927513996c620342eac609868b39f138637098e084e152f.jpg) +S2Fusion* + +![](images/6c9684560dbce0e51a442427cc50e998eb743463601f6cd2716a9d8e586e1ff1.jpg) + +![](images/c1ef914f738e78ec581bea0abe9b36200cf678089f222935027308051aae96fd.jpg) +EnvPoser + +![](images/fc6eb79f0d5c5d93ca80447fd4a702ac42e27f32f05b77e22031cc04e1a7d530.jpg) + +![](images/601230f5242a6002a29dd0b1dadf58280575e953e679a1f66834a22079482781.jpg) +GT + +JLM [55], which does not utilize environmental information, our method achieves significant improvements on two key metrics—MPJRE and MPJPE—by 0.42 (6.5%)/17.0 (18.5%) and 0.57 (11.5%)/13.1 (18.5%) on the EgoBody and GIMO datasets, respectively. These results underscore the increased accuracy in motion estimation achieved through the integration of environmental information. + +Additionally, we benchmark our model against the latest algorithm S2Fusion [35], which introduces environmental information as a constraint for human motion estimation. On the GIMO dataset, our method reduces MPJRE errors by $5.8\%$ compared to S2Fusion [35], and it outperforms S2Fusion across all evaluation metrics, demonstrating the superior effectiveness of our approach in leveraging environmental information. Moreover, the motion estimation performance of our method also outperforms S2Fusion* on the EgoBody dataset, which reduces MPJRE and MPJPE errors by $9.8\%$ and $16.3\%$ . Furthermore, our method shows a marked improvement in motion smoothness, with an $11.9\%$ + +
EgoBodyGIMO
MPJREMPJPEMPJREMPJPE
①Baseline7.4587.15.7274.6
②w/o env-semantic6.0478.04.5562.5
③w/o env-geometry6.0377.54.6363.2
EnvPoser6.0074.74.3857.6
+ +reduction in Jitter over S2Fusion and a $16.8\%$ improvement over AvatarJLM on the GIMO dataset. + +Qualitative Experimental Analysis. Fig. 3 showcases the motion estimation results on the EgoBody dataset, where we present three representative sequences and visualize the human motion estimation results at key moments. For state-of-the-art algorithms that do not incorporate environmental information, the visualization results in Fig. 3 show multiple instances of environment penetration during interactions. For example, AvatarJLM exhibits errors such as hand intersections with walls (second row) and leg penetrations through boxes (third row). Furthermore, in scenarios without chairs, squatting motions are misinterpreted as sitting (first row) in AvatarJLM due to the absence of environment understanding. In contrast, EnvPoser effectively leverages environmental information to reduce penetration errors and produce more realistic motion estimates. + +In more challenging interaction scenarios, EnvPoser shows certain advantages over S2Fusion, which also incorporates environmental information. Notably, S2Fusion fails to prevent interpenetration between the lower body and surrounding objects (third row in Fig. 3). Moreover, Fig. 4 provides a comparison between EnvPoser and S2Fusion, highlighting the advantages of our approach in handling upper-body interactions with objects and complex lower-body motions through the integration of uncertainty modeling and environmental semantic/geometric constraints. Additionally, compared to AvatarJLM, both EnvPoser and S2Fusion demonstrate the effectiveness of incorporating environmental information to improve human motion estimation. + +# 4.2. Ablation Study + +Effectiveness of the environment-aware refinement module. We conducted ablation studies comparing our method with three variants: ① Baseline: an auto-regressive transformer-based network with joint uncertainty estimation; ② w/o env-semantic: incorporating only geometric constraints; ③ w/o env-geometry: incorporating only semantic constraints. Tab. 2 presents the motion estimation performance for each variant. Compared to variant ①, our method achieves substantial improvements by integrating both semantic and geometric environmental information. Further analysis shows that each environment-aware refinement module independently enhances performance, with the combination yielding the best overall results. + +Table 2. The ablation study on environment refinement module. + +
EgoBodyGIMO
MPJREMPJPEMPJVEMPJREMPJPEMPJVE
1)Baseline w/o UNC7.6390.8333.05.9077.5270.2
2)Baseline7.4587.1342.95.7274.6266.7
3)EnvPoser w/o UNC6.5680.1172.74.4560.2278.6
EnvPoser6.0074.7174.04.3857.6234.6
+ +Table 3. The effectiveness of the uncertainty estimation. + +Effectiveness of the uncertainty estimation. As shown in Tab. 3, we compare two variants: 1) Baseline w/o UNC: a baseline without uncertainty estimation and 2) Baseline: a baseline with uncertainty estimation. Results show that integrating the uncertainty estimation module not only quantifies joint uncertainty across motions but also improves full-body motion estimation through additional supervision. We further incorporated both baseline methods into the environment-aware refinement module. As shown in Tab. 3, our method achieves significant improvements over variant 3) EnvPoser w/o UNC: our method without uncertainty estimation, reducing MPJRE and MPJPE errors by $8.11\%$ and $5.68\%$ , respectively, on the EgoBody dataset. These results indicate that the uncertainty estimation and resampling methods effectively capture the multi-hypothesis nature of initial motion estimates and, when combined with environmental refinement, adaptively align motion estimation with sparse tracking signals and environmental context. + +# 5. Conclusions + +This paper presents a novel two-stage framework for reliable human motion estimation leveraging sparse tracking signals and environmental context. In the first stage, we model the multi-hypothesis nature of motion through uncertainty estimation, followed by a refinement phase that applies environment-aware constraints from both semantic and geometric perspectives. Extensive experiments show that EnvPoser outperforms existing methods, particularly in complex interaction scenarios, highlighting the significance of environmental context in wearable motion estimation and paving the way for future approaches to harness this valuable contextual information. This work has the potential to impact a wide range of interactive applications that benefit from more accurate and context-aware motion estimation. + +Limitations and future works. The model assumes a static environment, lacking consideration for dynamic multi-user interactions or object motion, which may limit its effectiveness. This limitation could be addressed by incorporating a third-person perspective to estimate the movement of other people and objects, along with additional semantic and geometric constraints. Furthermore, real-time mesh quality in complex environments is often limited, which may affect performance. Future work will explore using raw images to infer contacts in complex scenes via 2D semantic understanding and introducing additional constraints to improve overall reconstruction quality. + +# References + +[1] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022. 2 +[2] Sadegh Aliakbarian, Fatemeh Saleh, David Collier, Pashmina Cameron, and Darren Cosker. Hmd-nemo: Online 3d avatar motion generation from sparse observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9622-9631, 2023. 1, 2 +[3] Rongyu Chen, Linlin Yang, and Angela Yao. Mhentropy: Entropy meets multiple hypotheses for pose and shape recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14840-14849, 2023. 2 +[4] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 2 +[5] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating smpl poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021. 2 +[6] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 1, 2, 6, 7 +[7] Zhen Fan, Peng Dai, Zhuo Su, Xu Gao, Zheng Lv, Jiarui Zhang, Tianyuan Du, Guidong Wang, and Yang Zhang. Emhi: A multimodal egocentric human motion dataset with hmd and body-worn imus. arXiv preprint arXiv:2408.17168, 2024. 2 +[8] Han Feng, Wenchao Ma, Quankai Gao, Xianwei Zheng, Nan Xue, and Huijuan Xu. Stratified avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 153-163, 2024. 1, 2 +[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 2 +[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 6 +[11] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 3 +[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3 +[13] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: learning to reconstruct human pose from sparse inertial measurements in real time. ACM Trans. Graph. (TOG), 37(6):1-15, 2018. 1, 2 +[14] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1, 2, 6, 7 +[15] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time egocentric pose estimation from sparse and intermittent observations everywhere. In European Conference on Computer Vision, pages 277-294. Springer, 2025. 2 +[16] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 3 +[17] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers (SA' 22), pages 1-9, 2022. 2, 4, 5 +[18] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision(ICCV), pages 11510-11520, 2021. 1 +[19] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3, 4 +[20] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 5 +[21] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3 +[22] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 2 + +[23] Han Li, Bowen Shi, Wenrui Dai, Hongwei Zheng, Botao Wang, Yu Sun, Min Guo, Chenglin Li, Junni Zou, and Hongkai Xiong. Pose-oriented transformer with uncertainty-guided refinement for 2d-to-3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1296–1304, 2023. 3, 4 +[24] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 2 +[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5 +[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 3 +[27] Zhenyu Lou, Qiongjie Cui, Haofan Wang, Xu Tang, and Hong Zhou. Multimodal sense-informed forecasting of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2144-2154, 2024. 3, 5 +[28] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 1, 4 +[29] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. Coap: Compositional articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13201-13210, 2022. 6 +[30] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2 +[31] Jose Luis Ponton, Eduard Pujol, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Dragposer: Motion reconstruction from variable sparse tracking signals via latent space optimization. arXiv preprint arXiv:2406.14567, 2024. 2 +[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5 +[33] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2 +[34] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3, 5 +[35] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2, 3, 6, 7 +[36] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusionposer: Real-time human motion reconstruction from arbitrary sparse sensors using autoregressive diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2513-2523, 2024. 2 +[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4 +[38] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11199-11208, 2021. 2 +[39] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 1 +[40] Yinghao Wu, chaoran wang, Lu Yin, Shihui Guo, and Yipeng Qin. Accurate and steady inertial pose estimation through sequence structure learning and modulation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[41] Songpengcheng Xia, Lei Chu, Ling Pei, Zixuan Zhang, Wenxian Yu, and Robert C Qiu. Learning disentangled representation for mixed-reality human activity recognition with a single imu sensor. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2 +[42] Songpengcheng Xia, Lei Chu, Ling Pei, Jiarui Yang, Wenxian Yu, and Robert C Qiu. Timestamp-supervised wearable-based activity segmentation and recognition with contrastive learning and order-preserving optimal transport. IEEE Transactions on Mobile Computing, 2024. 1 +[43] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–989, 2024. 2 +[44] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22268-22278, 2023. 3, 5 +[45] Hongdi Yang, Chengyang Li, Zhenxuan Wu, Gaozheng Li, Jingya Wang, Jingyi Yu, Zhuo Su, and Lan Xu. Smgdiff: Soccer motion generation using diffusion probabilistic models. arXiv preprint arXiv:2411.16216, 2024. 3 +[46] Xueyuan Yang, Chao Yao, and Xiaojuan Ban. Spatial-related sensors matters: 3d human motion reconstruction assisted with textual semantics. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10225-10233, 2024. 3 + +[47] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3, 5 +[48] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 2 +[49] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), pages 13167-13178, 2022. 1, 2 +[50] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 6 +[51] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision, pages 180-200. Springer, 2022. 6, 7 +[52] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 3, 6 +[53] Yufei Zhang, Hanjing Wang, Jeffrey O Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9032, 2023. 3 +[54] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 2 +[55] Xiaozheng Zheng, Zhuo Su, Chao Wen, Zhou Xue, and Xiaojie Jin. Realistic full-body tracking from sparse observations via joint-level modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14678-14688, 2023. 1, 2, 6, 7 +[56] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision, pages 676-694. Springer, 2022. 6 \ No newline at end of file diff --git a/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/images.zip b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9d7116f244dd7d1815d823188e8b6425489170e9 --- /dev/null +++ b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:471fed876c924c60838b77331b554c8a52f1c72364020a58098de3546f879118 +size 466228 diff --git a/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/layout.json b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..edb18fa349054353e494a2551a297769e725323e --- /dev/null +++ b/2025/EnvPoser_ Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling/layout.json @@ -0,0 +1,10290 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 49, + 103, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 103, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 545, + 140 + ], + "type": "text", + "content": "EnvPoser: Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": "Songpengcheng Xia" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Yu Zhang" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Zhuo Su" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Xiaozheng Zheng" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Zheng Lv" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Guidong Wang" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Yongjie Zhang" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Qi Wu" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Lei Chu" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Ling Pei" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " Shanghai Jiao Tong University " + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "{}^{2}\\mathrm{ByteDance}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "{}^{3}\\mathrm{University}" + }, + { + "bbox": [ + 52, + 159, + 555, + 205 + ], + "type": "text", + "content": " of Southern California" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 290, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 533 + ], + "type": "text", + "content": "Estimating full-body motion using the tracking signals of head and hands from VR devices holds great potential for various applications. However, the sparsity and unique distribution of observations present a significant challenge, resulting in an ill-posed problem with multiple feasible solutions (i.e., hypotheses). This amplifies uncertainty and ambiguity in full-body motion estimation, especially for the lower-body joints. Therefore, we propose a new method, EnvPoser, that employs a two-stage framework to perform full-body motion estimation using sparse tracking signals and pre-scanned environment from VR devices. EnvPoser models the multi-hypothesis nature of human motion through an uncertainty-aware estimation module in the first stage. In the second stage, we refine these multi-hypothesis estimates by integrating semantic and geometric environmental constraints, ensuring that the final motion estimation aligns realistically with both the environmental context and physical interactions. Qualitative and quantitative experiments on two public datasets demonstrate that our method achieves state-of-the-art performance, highlighting significant improvements in human motion estimation within motion-environment interaction scenarios. Project page: https://xspc.github.io/EnvPoser/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 572, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 669 + ], + "type": "text", + "content": "In recent years, motion capture algorithms using sparse tracking signals have achieved significant progress across various research fields, establishing themselves as essential technologies for applications such as motion-sensing games and immersive VR/AR experiences [13, 18, 42, 49]. With the rapid development of VR devices, full-body motion estimation algorithms based on head-mounted display (HMD) have shown great application potential. However, devices" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 232, + 545, + 462 + ], + "blocks": [ + { + "bbox": [ + 308, + 232, + 545, + 462 + ], + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 462 + ], + "type": "image", + "image_path": "2fc829f832043b95837aaa864a6bf54e4c8e93fb0ba447c596d072c32f298c89.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 470, + 547, + 503 + ], + "lines": [ + { + "bbox": [ + 305, + 470, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 547, + 503 + ], + "type": "text", + "content": "Figure 1. EnvPoser can estimate the full-body motion using three tracking signals (HMD and hand controllers) and a pre-scanned environment mesh." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 517, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 546, + 555 + ], + "type": "text", + "content": "like PICO and Quest typically provide sparse tracking signals of the head and hands, presenting challenges in estimating the full-body motion from these limited observations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Current sparse-observation-based human motion estimation methods typically rely on large-scale motion capture datasets, such as AMASS [28], to capture full-body motion through regression [14, 55] or generative models [2, 6, 8]. Alternatively, physics engines and reinforcement learning have been used [39] to synthesize full-body motion, bypassing dataset dependency. Utilizing various advanced architectures or large-scale datasets, these methods [2, 6, 8, 39, 55] attempt to model the mapping from sparse tracking signals to the most probable full-body motion. However, the lack of direct joint observations results in the same sparse input often corresponding to multiple plausible motions, making it a challenge for sparse-observation-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 674, + 105, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 105, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 105, + 684 + ], + "type": "text", + "content": "*Project Lead" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 684, + 137, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 684, + 137, + 693 + ], + "spans": [ + { + "bbox": [ + 59, + 684, + 137, + 693 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "text", + "content": "This work was supported by the National Nature Science Foundation of China (NSFC) under Grant 62273229." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1839" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "based motion estimation algorithms to converge on the most probable full-body motion. To address this one-to-many mapping ambiguity, existing methods [4, 30, 31] attempt to add sensors to regions such as the pelvis and legs to narrow the distribution of possible motion outcomes. Although additional sensors can enhance the accuracy and robustness of full-body motion estimation, this approach may negatively impact user experience due to increased hardware requirements [13, 17, 41, 49, 54]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 181, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 289, + 324 + ], + "type": "text", + "content": "Therefore, an alternative approach to addressing the inherent ambiguity in sparse observations is to incorporate additional environmental constraints, as shown in Fig. 1. Human motion is highly correlated with the surrounding environment, which contains rich context that can provide valuable cues for motion estimation. However, effectively utilizing environmental constraints presents its own challenges, as existing methods [35, 55] often simplify human-environment interactions to basic foot-to-ground contact, overlooking the complexities of interactions. In reality, individuals frequently engage in more complex interactions with their surroundings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 325, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 289, + 564 + ], + "type": "text", + "content": "Our key insight is that joint uncertainty estimation can explicitly model the multi-hypothesis nature of full-body motion estimation, and by incorporating environmental information, we can reduce estimation uncertainty, guiding the motion estimation to converge toward the most plausible outcome that aligns with both the environment and sparse input observations. Therefore, we design a two-stage deep learning framework to handle the one-to-many mapping ambiguity challenge. We first explicitly model the multi-hypothesis human motion by the joint uncertainty estimation and its sampling strategy. Then, by integrating environmental information, we guide the uncertainty-based multi-hypothesis motion estimations to converge on a full-body motion that best aligns with both the environment and sparse observations. To address the challenge of effectively utilizing environmental information, we introduce semantic constraints for non-contact motions, providing contextual cues for realistic estimation, and geometric constraints for contact scenarios, ensuring consistency and preventing collisions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 566, + 262, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 566, + 262, + 578 + ], + "spans": [ + { + "bbox": [ + 59, + 566, + 262, + 578 + ], + "type": "text", + "content": "The key contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 582, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 582, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 664 + ], + "type": "text", + "content": "- We propose EnvPoser, a novel framework that significantly advances full-body motion estimation from sparse tracking signals by incorporating pre-scanned environmental context. This framework addresses the complexities of estimating human motion from sparse observations, ensuring robustness across diverse interactive scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "type": "text", + "content": "- This work presents a two-stage framework for human motion estimation: an uncertainty-aware module provides initial estimates that account for multi-hypothesis motion, while an environment-aware refinement module in" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 73, + 547, + 168 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 73, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 545, + 119 + ], + "type": "text", + "content": "tegrates semantic and geometric constraints to ensure realistic outcomes in diverse scenarios, setting a new benchmark in motion estimation with environmental interactions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 121, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 547, + 168 + ], + "type": "text", + "content": "- Comprehensive quantitative and qualitative experiments demonstrate that our proposed method achieves state-of-the-art performance across two public datasets and excels in scenarios with frequent environmental interactions." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 180, + 392, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 392, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 392, + 192 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 200, + 531, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 531, + 213 + ], + "type": "text", + "content": "2.1. Motion estimation with sparse observations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "type": "text", + "content": "Current Wearable-based human motion estimation approaches mainly adopt two configurations: one uses multiple inertial measurement units (IMUs) placed on body extremities [13, 17, 36, 40, 48, 54], while the other relies on an HMD with two controllers [2, 4, 7, 14, 15, 55]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 279, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 546, + 518 + ], + "type": "text", + "content": "While methods utilizing six IMUs have shown impressive results, HMD-based techniques leveraging sparse three-point observations show significant potential for practical applications. For instance, AvatarPoser [14] employs a transformer architecture to achieve accurate full-body motion. While Zheng et al. [55] developed a two-stage framework to capture joint dependencies, improving motion accuracy and fluidity. Some studies [1, 5, 8] address the ill-posed problem of sparse-to-dense mapping using generative models. Dittadi et al. [5] and Aliakbarian et al. [1] proposed using Variational Autoencoders (VAE) and normalizing flows, respectively, to estimate full-body poses from three tracking points. Additionally, the diffusion model [6, 8, 35] has also shown notable performance in motion estimation with sparse data. Despite these advances, most methods rely solely on sparse tracking signals, leading to high uncertainty and multiple potential poses for the same input. Therefore, this paper introduces joint uncertainty estimation and constrains physically implausible motions using environmental information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 527, + 533, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 533, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 533, + 540 + ], + "type": "text", + "content": "2.2. Pose regression with uncertainty estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": "The sparse nature of observations in 3D human pose estimation introduces significant uncertainty for unobserved joints, as identical inputs can correspond to multiple pose hypotheses [8, 54, 55]. This challenge is particularly evident in image-based pose estimation [3, 9, 22, 24, 43], where monocular methods face depth ambiguity. For instance, Wehrbein et al. [38] modeled the posterior distribution of 3D pose hypotheses using normalizing flows, while MHFormer [24] employed a multi-hypothesis transformer to capture diverse pose hypotheses through a one-to-many mapping. Recently, Shan et al. [33] proposed diffusion-based methods to aggregate multi-hypothesis predictions with joint-wise re-projection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In our paper, we aim to use uncertainty estimation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1840" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "to model the multi-hypothesis nature of motion from sparse observations. Uncertainty in deep learning can be categorized into two types [19]: aleatoric (data-related) and epistemic (model-related). Li et al. [23] refined 3D poses by incorporating joint uncertainties through uncertainty guided-sampling and uncertainty-guided self-attention, while Zhang et al. [53] introduced a probabilistic framework that encodes aleatoric uncertainty using a robust negative log-likelihood loss, with epistemic uncertainty guiding model refinement. In wearable motion estimation, Yang et al. [46] leveraged uncertainty to weight IMU-derived features, enhancing pose corrections with additional text annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 236, + 287, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 287, + 249 + ], + "type": "text", + "content": "2.3. Scene-aware motion estimation and generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 506 + ], + "type": "text", + "content": "Recent advancements in human motion estimation and generation have expanded from solely analyzing body motion to integrating interactions with the surrounding environment, especially vision-based approaches [11, 12, 16, 34, 47, 52]. For instance, PROX [10] incorporated scene constraints into monocular human pose estimation to address depth ambiguity, while Shen et al. [34] used sparse 3D CNNs to estimate absolute positions and dense scene contacts, refining human mesh recovery via cross-attention with 3D scene cues. EgoHMR [52] leveraged ego-camera images and diffusion models within pre-scanned environments to recover full-body motion. Similarly, S2Fusion [35] utilized sparse observations by learning human motion priors on synthetic datasets and conditioned motion generation on scene encoding and VR inputs. However, S2Fusion primarily focuses on lower body and environment interactions. QuestEnvsim [21] employed reinforcement learning to generate realistic poses in highly constrained environments using HMDs. Additionally, recent studies further emphasize the integration of scene context and historical motion to produce plausible human motion [27, 44, 45]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 507, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 288, + 579 + ], + "type": "text", + "content": "These scene-aware methods underscore the importance of environmental context for enhancing motion estimation. Our approach builds on these foundations by comprehensively leveraging human-environment interactions, addressing both contact and non-contact scenarios to improve motion estimation with sparse tracking signals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 590, + 104, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 104, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 104, + 602 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 160, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 160, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 160, + 623 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": "Our goal is to estimate the full-body motion sequence " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\theta_t\\}_{t=1}^T \\in \\mathcal{R}^{T \\times 132}" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " (the first 22 joints of SMPL [26] model with the 6D representation of rotations) using the sparse tracking signals " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{\\mathbf{x}_t\\}_{t=1}^T \\in \\mathcal{R}^{T \\times N_c}" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " from the HMD and hand controllers, along with a pre-scanned 3D environment point cloud " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " denotes the length of input, " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_c = 36" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " means the channel of input. Each sparse tracking" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "signal at time step " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " consists of the head and hand positions " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", rotations " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{R}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", and linear velocities " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", represented as " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = [\\pmb{p}_t, \\pmb{R}_t, \\pmb{v}_t]" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ". Moreover, we utilize a cropped 3D environment point cloud " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "V_S \\in \\mathcal{R}^{N_S \\times 3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "N_S" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", the number of sampled points, is set to 1000. With the human translation, this point cloud refines pose estimation by incorporating environmental context, leveraging both contact and noncontact interactions between the body and the environment to reduce the uncertainty inherent in sparse observations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 181, + 546, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 546, + 348 + ], + "type": "text", + "content": "As shown in Fig. 2, our approach is a two-stage model comprising two core technical modules: 1) an uncertainty-aware initial motion estimation module; and 2) an environment-aware motion refinement module that leverages semantic and geometric information from the environment. We first train the uncertainty-aware initial motion estimation module (Stage I) on the AMASS dataset, with a focus on explicitly capturing the multi-hypothesis nature of motion estimation through uncertainty quantification. Subsequently, we jointly train both modules (Stage II) on the motion-environment interaction datasets, applying environmental semantic and geometric constraints to refine the multi-hypothesis estimates from Stage I into the most plausible motion estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 357, + 545, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 545, + 381 + ], + "type": "text", + "content": "3.2. Uncertainty-aware Initial Human Motion Estimation Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 544 + ], + "type": "text", + "content": "To obtain the joint uncertainty quantification and explicitly model the multi-hypothesis nature of motion estimation, we propose an uncertainty-aware initial human motion estimation module. This module is designed with three primary objectives: (1) to integrate historical motion states and sparse observations for the current initial motion estimation; (2) to effectively extract features from sparse signals; and (3) to quantify uncertainty in motion estimation, explicitly modeling the multi-hypothesis nature of motion under identical input conditions. Accordingly, this module comprises three sub-modules: sparse input signal and historical motion embedding, transformer-based motion feature extraction, and human motion and joint uncertainty regression." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": "Sparse Input Signal and Historical Human Motion Embedding. To integrate historical motion states and sparse observations for the current initial motion estimation, we adopt an auto-regressive structure in the uncertainty-aware initial human motion estimation module. Our model takes sparse observations " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathcal{R}^{T \\times 36}" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": " and historical motion " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{hm} \\in \\mathcal{R}^{T \\times 132}" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": " as inputs, embedding these inputs through linear layers to obtain the observation representation " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "Z_{X}" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": " and motion historical state representation " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "Z_{hm}" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": ". Then, we concatenate the sparse observation embedding and historical motion embedding to obtain the shallow motion representation " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "Z_{S} = [Z_{X}, Z_{hm}]" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": ", which can be used for subsequent transformer-based feature extraction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "content": "Transformer-based Motion Feature Extraction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1841" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 281 + ], + "type": "image", + "image_path": "7868b37cd6581f6594f5d4bf64e84d642a1ad73fc868415e5640199603cefbc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 546, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 546, + 319 + ], + "type": "text", + "content": "Figure 2. Overview of EnvPoser: A Two-Stage Motion Estimation Model. Stage I involves training the uncertainty-aware initial estimation module on the AMASS dataset to produce initial motion estimates with uncertainty quantification. Stage II refines these estimates by training on motion-environment datasets, incorporating semantic and geometric environmental constraints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "text", + "content": "Based on the shallow motion representation " + }, + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "inline_equation", + "content": "Z_{S}" + }, + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "text", + "content": ", we employ a transformer-based feature extraction module [17] for motion features learning. To retain sequence order information, temporal positional encoding is applied to the shallow motion representation [37]. Using multi-head self-attention and a position-wise feed-forward network within the transformer encoder " + }, + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "text", + "content": ", this module efficiently captures intricate relationships in the sparse tracking signals and historical motion, resulting in high-dimensional motion features " + }, + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "inline_equation", + "content": "Z_{H} = F_{t}(Z_{S})" + }, + { + "bbox": [ + 46, + 326, + 288, + 541 + ], + "type": "text", + "content": ". While the transformer architecture is effective for extracting motion features, the sparsity of the input data—limited to tracking signals from the head and hands—introduces high uncertainty in estimating unobserved joints, particularly in the lower body. This uncertainty arises from the lack of direct observations for these joints, making their estimation inherently ambiguous. To address this challenge, we incorporate a mechanism to quantify the uncertainty in joint estimations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "text", + "content": "Human Motion and Joint Uncertainty Regression. With the high-dimensional motion features " + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "inline_equation", + "content": "Z_{H}" + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "text", + "content": ", we employ a pose regression network " + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "inline_equation", + "content": "F_{ph}(\\cdot)" + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "text", + "content": " to predict the human motion " + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta} = F_{ph}(Z_H)" + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "text", + "content": ". Given the sparse nature of the input signals, capturing joint uncertainty is crucial in mapping the limited observations to a full-body motion estimation. To achieve this, we utilize a heteroscedastic neural network [19, 23] to estimate the joint uncertainty. Specifically, an uncertainty regression head " + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "inline_equation", + "content": "F_{uh}(\\cdot)" + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "text", + "content": " is employed on the high-dimensional motion feature to estimate the joint uncertainty " + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathcal{R}^{T\\times 132}" + }, + { + "bbox": [ + 46, + 542, + 287, + 674 + ], + "type": "text", + "content": " of the predicted motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Using the predicted motion and its associated uncertainty, we resample the full-body motion " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\overline{\\theta}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " around the predicted mean motion " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " according to a Gaussian distribution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\widetilde{\\boldsymbol{\\theta}},\\delta)" + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "text", + "content": " with the uncertainty " + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "text", + "content": ". This resampling technique introduces a multi-hypothesis representation, characterizing the inherent ambiguity in the unobserved joints. To ensure differentiability throughout the model training process, we apply the reparameterization trick on this sampling operation, drawing a parameter " + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "text", + "content": " from a standard Gaussian distribution " + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "text", + "content": ", represented by " + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\overline{\\theta} = \\widetilde{\\theta} +\\delta \\cdot \\epsilon" + }, + { + "bbox": [ + 304, + 325, + 545, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 411, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 506 + ], + "type": "text", + "content": "Through the uncertainty regression head, we can obtain the joint uncertainty estimation for predicted human motion, enabling a multi-hypothesis motion representation with the sampling strategy. We pre-train the first stage of our method on the large synthetic dataset AMASS [28]. Based on the predicted mean human motion " + }, + { + "bbox": [ + 304, + 411, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}" + }, + { + "bbox": [ + 304, + 411, + 546, + 506 + ], + "type": "text", + "content": " and uncertainty " + }, + { + "bbox": [ + 304, + 411, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 411, + 546, + 506 + ], + "type": "text", + "content": ", we train the first stage module with the following loss function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 384, + 506, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 506, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 384, + 506, + 545, + 521 + ], + "type": "interline_equation", + "content": "L _ {M} = \\| \\widetilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta} \\| _ {2}, \\tag {1}", + "image_path": "5a614c820af0bfe9d0f2a22d009a0f56a5524df96bcbf0bf94e53e63cae5b600.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": " represents the ground-truth of human motion. Additionally, to ensure accurate uncertainty prediction, we also incorporate an uncertainty estimation loss as [19] stated:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 567, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 567, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 359, + 567, + 545, + 594 + ], + "type": "interline_equation", + "content": "L _ {\\delta} = \\left\\| \\frac {\\widetilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta}}{\\delta} \\right\\| _ {2} + \\log (\\left\\| \\boldsymbol {\\delta} \\right\\| _ {2}). \\tag {2}", + "image_path": "8864b99a8144b95e311219e1f49f8e3b271274df74ae71e0c5d8726e6e301fa4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 601, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 635 + ], + "type": "text", + "content": "Thus, based on the uncertainty-aware initial human motion module, the objective function for Stage I is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 636, + 545, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 636, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 367, + 636, + 545, + 649 + ], + "type": "interline_equation", + "content": "L _ {S - I} = \\lambda_ {M} L _ {M} + \\lambda_ {\\delta} L _ {\\delta}, \\tag {3}", + "image_path": "4d46520c0862f583a8841e229ccca4a5256707423be8b2d6c827a185df5cf853.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\lambda_{M}" + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\lambda_{\\delta}" + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "content": " are the hyper-parameters for Stage I training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Training Stage I on the AMASS dataset, along with the design for motion uncertainty quantification, allows our network to effectively capture the multi-hypothesis nature of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1842" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "motion estimation arising from sparse observations. To address the inherent one-to-many ambiguity, we introduce environmental information as a refinement stage. By incorporating both semantic and geometric environmental information, our method guides the multi-hypothesis motion estimates to converge toward the solution that best aligns with the sparse observations and environmental constraints." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 288, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 288, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 288, + 188 + ], + "type": "text", + "content": "3.3. Environment-aware Human Motion Refinement with Semantic and Geometry Constraint" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 361 + ], + "type": "text", + "content": "Based on the initial motion estimation and its uncertainty produced by the first-stage uncertainty-aware module, we further refine the multi-hypothesis human motion estimates using the pre-scanned environment point clouds collected by real-world devices. To effectively utilize the environmental information, our design is guided by a key insight: human motion within an environment involves both object interactions (contact scenarios) and collision avoidance (non-contact scenarios). In this module, the environment point clouds are processed through cropping, sampling, and encoding. We then apply constraints based on the environment's semantic and geometric properties to ensure that the full-body motion estimation realistically aligns with the surrounding context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": "Environmental Point Cloud Embedding. To enhance computational efficiency and focus on relevant interaction areas, we first clip the environmental point clouds based on the human's global position. A circular sampling approach is used to minimize the impact of body orientation, selecting a 1-meter radius around the body and uniformly distributing " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": " points within this area, denoted as " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{S}} \\in \\mathcal{R}^{N_{S} \\times 3}" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": ". The cropped environmental point cloud " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{S}}" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": " is then processed by the environment encoder " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "F_{env}(\\cdot)" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": " to obtain the environmental point cloud embedding " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "Z_{env}" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": ", represented by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 489, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 489, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 121, + 489, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} _ {\\text {e n v}} = F _ {\\text {e n v}} \\left(\\boldsymbol {V} _ {\\boldsymbol {S}}\\right), \\tag {4}", + "image_path": "87282820f18cf076a141ab044040fe73c1851ee2ab577037f662c8aca0adaea2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 558 + ], + "type": "text", + "content": "where the environment encoder " + }, + { + "bbox": [ + 46, + 510, + 287, + 558 + ], + "type": "inline_equation", + "content": "F_{env}(\\cdot)" + }, + { + "bbox": [ + 46, + 510, + 287, + 558 + ], + "type": "text", + "content": " is implemented using the vanilla PointNet++ [32]. The extracted environmental embedding would be used in the environment semantic-aware motion refinement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Environment Semantic-Aware Constraint for Motion Refinement. Capturing semantic information from the environmental point cloud provides crucial guidance for motion generation [27, 44, 47], as human motion is highly correlated with environmental context. Therefore, leveraging this motion-environment correlation as a valuable supplement has the potential to enhance the sparse-observation-based human motion estimation. Inspired by [27, 34], we employ a cross-attention module between the initial motion estimates and the cropped environmental point clouds to refine the human motion estimation. This module comprises an environment-embedding network, a motion information embedding network and an environment-motion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 73, + 405, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 405, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 405, + 83 + ], + "type": "text", + "content": "cross-attention network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": "The motion information embedding network " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "F_{motion}(\\cdot)" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": " takes the sampled motion estimation " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{\\theta}}" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": ", head translation " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{head}" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": " and the extended sparse observations " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "X_{new} = \\{X, h, \\pmb{\\theta}^{up}\\} \\in \\mathcal{R}^{T \\times 40}" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": " as input. The extended sparse observations " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "X_{new}" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": " include the relative head height " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": " and head up-vector " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^{up}" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": ", obtained via a relative head height calculation module. Following [17, 20], this module calculates terrain height by using the height of the feet when stationary and nearby environmental points, and then determines " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": " as the difference between the absolute head height from the HMD and the estimated terrain height. These inputs are concatenated and processed through linear layers to produce the motion information embedding " + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "inline_equation", + "content": "Z_M = F_{motion}(\\text{concat}(\\overline{\\theta}, \\pmb{p}^{head}, X_{new}))" + }, + { + "bbox": [ + 304, + 84, + 545, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": "For the environment-motion cross-attention network, the Key " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " and Value " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " vectors can be obtained by multiplying the weight matrix " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "W_{k}, W_{v}" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " with the environmental embedding " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "Z_{env}" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": ", while the motion information embedding " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "Z_{M}" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " multiplies the weight matrices " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " to get the Query " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " vectors. Inspired by [25, 27], we incorporate a spatial salience " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "s_{\\text{spatial}} = MLP(V_{S})" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " that expects our cross-attention model to be able to perceive the distance between environmental points clouds and human bodies. Using the spatial salience " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "s_{\\text{spatial}}" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": " in conjunction with attention scores " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "\\text{Attn}(Q, K)" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": ", we derive a motion-environment representation " + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "inline_equation", + "content": "Z_{ME}" + }, + { + "bbox": [ + 304, + 253, + 546, + 397 + ], + "type": "text", + "content": ", which could be represented by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 406, + 545, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 406, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 340, + 406, + 545, + 419 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} _ {M E} = \\left(\\operatorname {A t t n} (\\boldsymbol {Q}, \\boldsymbol {K}) + \\boldsymbol {s} _ {\\text {s p a t i a l}}\\right) \\cdot \\boldsymbol {V}. \\tag {5}", + "image_path": "b0a0481a0f1976a4db86154d3f4592c5a9b9583185fb5106ab4f2a4f5cd06f1e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": "Then, by concatenating the motion-environment representation " + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "inline_equation", + "content": "Z_{ME}" + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": " with the motion information embedding " + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "inline_equation", + "content": "Z_{M}" + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": " and passing them through an MLP network, we obtain the environment-refined motion representation " + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "inline_equation", + "content": "Z_{RM} = MLP(\\text{concat}(Z_{ME}, Z_{M}))" + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "type": "text", + "content": "Using the environment-refined motion representation " + }, + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "type": "inline_equation", + "content": "Z_{RM}" + }, + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "type": "text", + "content": ", which integrates environmental semantics and spatial priors between the human and environment, the model can then regress the probability of human-environment contact and generate the final motion estimation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "content": "Motion Regression with Contact Estimation. With the environment-refined motion representation " + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "inline_equation", + "content": "Z_{RM}" + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "content": ", we estimate the contact probability " + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "inline_equation", + "content": "C = \\{c\\}_{t=1}^{T} \\in \\mathcal{R}^{T \\times 22}" + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "content": " and regress the final human motion " + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{RM}" + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "content": ". We take the environment-refined motion representation " + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "inline_equation", + "content": "Z_{RM}" + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "content": " and the extended sparse observations " + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "inline_equation", + "content": "X_{new}" + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "content": " as input, and apply two-layers MLP layers on the concatenated input to get the contact probability prediction " + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\hat{C} = MLP(concat(X_{new}, Z_{RM}))" + }, + { + "bbox": [ + 304, + 547, + 545, + 656 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 656, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 656, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 656, + 545, + 692 + ], + "type": "text", + "content": "Contact probability prediction is typically treated as a binary classification problem, where Binary Cross Entropy (BCE) loss is used for training:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 357, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 357, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o n t a c t}} = B C E L o s s (\\hat {\\boldsymbol {C}}, \\boldsymbol {C}). \\tag {6}", + "image_path": "ad6d3c58732cfc9bedf795db13399fee980db6af06ca6a2b57f1155e4e6317f1.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1843" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "Finally, the contact probability " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " and environment-refined motion representation " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "Z_{RM}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " are concatenated to regress the final motion with a motion decoder network " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "F_{out}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 118, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 118, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 91, + 118, + 287, + 133 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\theta}} _ {R M} = F _ {o u t} (\\operatorname {c o n c a t} (\\mathbf {Z} _ {R M}, \\hat {\\mathbf {C}})), \\tag {7}", + "image_path": "7cff32f311b821e1294c0392ccf53a9ea46016c827797d7f2feefd411b9642c1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "inline_equation", + "content": "F_{out}(\\cdot)" + }, + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "text", + "content": " consists of two linear layers with a ReLU activation function. We then obtain the final motion loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 175, + 287, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 175, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 287, + 189 + ], + "type": "interline_equation", + "content": "L _ {M ^ {\\prime}} = \\left\\| \\hat {\\boldsymbol {\\theta}} _ {R M} - \\boldsymbol {\\theta} \\right\\| _ {2}. \\tag {8}", + "image_path": "f6eeb235fdbcb6b216b1bf9d685b343da09c969ce585439a3dae4e1d25853bdc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 198, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 388 + ], + "type": "text", + "content": "Scene Geometry-Aware Constraint for Motion Refinement. To complement semantic-level constraints, we leverage the geometric structure of the environment to enhance motion estimation accuracy, especially in human-environment interactions (contact scenarios). A straightforward approach is to detect collisions between the estimated human motion and the environment point cloud, thereby refining unrealistic motions. While existing methods [10, 50] often convert 3D environments into Signed Distance Fields (SDF) for collision detection, this approach can be computationally intensive. Inspired by recent advancements [29, 52], we instead employ the COAP model [29] to efficiently compute collisions between the human body and environmental points. With this novel model and the environment point clouds, the COAP-based collision loss [52] can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 396, + 287, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 396, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 64, + 396, + 287, + 429 + ], + "type": "interline_equation", + "content": "L _ {c o a p} = \\frac {1}{N _ {S}} \\sum_ {i = 1} ^ {N _ {S}} \\sigma \\left(f _ {\\Theta} \\left(V _ {S _ {i}} \\mid \\mathcal {G}\\right)\\right) \\mathbb {I} _ {f _ {\\Theta} \\left(V _ {S _ {i}} \\mid \\mathcal {G}\\right) > 0}, \\tag {9}", + "image_path": "987268980ad9093ac60aae3b7eacc8eb9ab4dfe1a30fcc0f6e07e06bf0d26a8e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "V_{S}" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " denotes the cropped environment point clouds, with " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " representing the number of points. " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " represents the transformation parameters of the human body, encompassing both the motion pose " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " and shape " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "f_{\\Theta}(V_{S_i}|\\mathcal{G})" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " represents the spatial relationship between the human body model and environmental points. The function " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " is the Sigmoid function, which maps the value of " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "f_{\\Theta}(V_{S_i}|\\mathcal{G})" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " to a range between 0 and 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "content": "Since most motions occur on the ground, we remove ground points from the scene to ensure that the COAP-based collision loss specifically targets interactions with surrounding objects. To account for interactions with the ground, we introduce additional constraints, including foot contact, foot height, and ground penetration losses [55]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 614, + 233, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 614, + 233, + 632 + ], + "spans": [ + { + "bbox": [ + 98, + 614, + 233, + 632 + ], + "type": "interline_equation", + "content": "L _ {f c} = \\left\\| \\left(\\hat {\\boldsymbol {P}} _ {R M} ^ {f e e t} - \\boldsymbol {P} ^ {f e e t}\\right) \\cdot \\boldsymbol {C} \\right\\| _ {1},", + "image_path": "e9308a3192bb88d91f86aba59e055407f1a5a28fcf26efee495bc0c7244a3e56.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 634, + 287, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 634, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 100, + 634, + 287, + 650 + ], + "type": "interline_equation", + "content": "L _ {g f h} = \\left\\| \\hat {\\boldsymbol {z}} _ {P _ {R M}} ^ {f e e t} - \\boldsymbol {z} ^ {g r o u n d} \\right\\| _ {1}, \\tag {10}", + "image_path": "1e8c3531455ebd000bde1b32280a4f4feac4a7efc4c9ad004f8164a539821906.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 651, + 235, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 651, + 235, + 666 + ], + "spans": [ + { + "bbox": [ + 100, + 651, + 235, + 666 + ], + "type": "interline_equation", + "content": "L _ {g p} = \\left\\| \\left(\\hat {z} _ {P _ {R M}} ^ {\\text {m i n}} - z ^ {\\text {g r o u n d}}\\right) \\cdot \\boldsymbol {l} \\right\\| _ {1},", + "image_path": "57769cc61c79c1e5f76f9a5be1030709e93e3739e756bdf22dbe74d542d06784.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{z}_{P_{RM}}^{feet}" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "z_{ground}" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": " means the height of feet joint positions and the floor height, and " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": " denotes whether the joint is lower than the ground." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": "Final Loss Function. In addition to the first stage loss (Eq.2, Eq.1), COAP-based collision loss (Eq.9), contact loss (Eq.6), foot-ground loss (Eq.10) and final motion loss (Eq.8), we also apply some additional losses to improve model training. With the forward kinematic chain, the joint position " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\hat{P}_{RM}" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": ", and the global hands positions " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "P_{hand}^{\\prime}" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\hat{P}_{RM}^{\\text {hand }}" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": " could be calculated by the motion pose " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": ". To penalize the errors accumulating along the kinematic chain, we apply a joint position loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "(L_{\\text {posi}})" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": " and a hand alignment loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "(L_{hAL})" + }, + { + "bbox": [ + 305, + 72, + 545, + 194 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 205, + 545, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 205, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 314, + 205, + 545, + 233 + ], + "type": "interline_equation", + "content": "L _ {\\text {p o s i}} = \\left\\| \\hat {\\boldsymbol {P}} _ {R M} - \\boldsymbol {P} \\right\\| _ {2}, L _ {h A L} = \\left\\| \\hat {\\boldsymbol {P}} _ {R M} ^ {\\text {h a n d}} - \\boldsymbol {P} ^ {\\text {h a n d}} \\right\\| _ {1}. \\tag {11}", + "image_path": "4adad507e597d32dad30a622a8efcc16ce4918298e7df079214490da50d5efae.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 234, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 545, + 269 + ], + "type": "text", + "content": "Therefore, with the technical analysis provided above, we can formulate the second stage objective function as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 279, + 545, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 545, + 316 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {S - I I} = L _ {S - I} + L _ {M ^ {\\prime}} + \\lambda_ {1} L _ {p o s i} + \\lambda_ {2} L _ {h A L} + \\lambda_ {3} L _ {f c} \\\\ + \\lambda_ {4} L _ {\\text {c o n t a c t}} + \\lambda_ {5} L _ {g f h} + \\lambda_ {6} L _ {g p} + \\lambda_ {7} L _ {\\text {c o a p}}. \\tag {12} \\\\ \\end{array}", + "image_path": "0e75f6f0f196decf239575ab170186afda7b97dc37303cd36ce5b9b416a54822.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 318, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 545, + 377 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 305, + 318, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\{\\lambda_i\\}_{i = 1,\\dots ,7}" + }, + { + "bbox": [ + 305, + 318, + 545, + 377 + ], + "type": "text", + "content": " represents the hyperparameters. Our model is first trained on the AMASS using the Eq. 3 and then fine-tuned on a motion-environment interaction dataset with Eq. 12. The training process and selection of hyperparameters can be found in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 392, + 387, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 387, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 387, + 404 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 412, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 496 + ], + "type": "text", + "content": "This section introduces the datasets, metrics, and competing algorithms used in our experiments. We then compare our method EnvPoser with state-of-the-art algorithms, highlighting its advantages in motion estimation. Finally, ablation studies validate the effectiveness of each module. Additional case studies and discussions are provided in the supplementary materials." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 497, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 556 + ], + "type": "text", + "content": "Datasets. To comprehensively evaluate the efficacy of our proposed method, we conduct experiments on two challenging public motion-scene interactions datasets: Egobody [51] and GIMO [56] datasets. Detailed information about these datasets can be found in the supplementary document." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Competing methods. We compare the motion estimation performance of the proposed model with the state-of-the-art competing methods, including AvatarPoser [14], AGRoL [6], AvatarJLM [55], and S2Fusion [35] on two public datasets. We re-train all competing methods on the EgoBody and GIMO datasets until convergence. For the results on the GIMO dataset, we reused the values reported in previous studies [35]. For the EgoBody dataset, we tested our trained model under real-time settings and presented the results as follows. It is worth noting that S2Fusion [35] is not fully open-source; therefore, we reproduced S2Fusion based on the details provided in its paper and partial open-source code, marking it as S2Fusion* in this experiment." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1844" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 129, + 121 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 129, + 121 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 129, + 121 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 129, + 121 + ], + "type": "image", + "image_path": "fcc6b04db8a1b9fa5b5710fd5efd781ae7d071850727bb8ddaddb7d39f212ca6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 121, + 129, + 175 + ], + "blocks": [ + { + "bbox": [ + 53, + 121, + 129, + 175 + ], + "lines": [ + { + "bbox": [ + 53, + 121, + 129, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 121, + 129, + 175 + ], + "type": "image", + "image_path": "eeccb6f85d0353a3add391a9684a75b6bc9a5befb4817815ee321c0b52d97148.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 175, + 129, + 255 + ], + "blocks": [ + { + "bbox": [ + 53, + 175, + 129, + 255 + ], + "lines": [ + { + "bbox": [ + 53, + 175, + 129, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 129, + 255 + ], + "type": "image", + "image_path": "a8df69a351ba0e5095e4051cacdffc2c14a4bc15fbeea277438ed3e7ba150f3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 255, + 108, + 262 + ], + "lines": [ + { + "bbox": [ + 75, + 255, + 108, + 262 + ], + "spans": [ + { + "bbox": [ + 75, + 255, + 108, + 262 + ], + "type": "text", + "content": "AvatarPoser" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 70, + 211, + 121 + ], + "blocks": [ + { + "bbox": [ + 135, + 70, + 211, + 121 + ], + "lines": [ + { + "bbox": [ + 135, + 70, + 211, + 121 + ], + "spans": [ + { + "bbox": [ + 135, + 70, + 211, + 121 + ], + "type": "image", + "image_path": "1a9d94105bfb1daa52173b9419d47cb45d3a2eaeb5f5b7dd92b5d1ad0479ddbd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 121, + 211, + 175 + ], + "blocks": [ + { + "bbox": [ + 135, + 121, + 211, + 175 + ], + "lines": [ + { + "bbox": [ + 135, + 121, + 211, + 175 + ], + "spans": [ + { + "bbox": [ + 135, + 121, + 211, + 175 + ], + "type": "image", + "image_path": "69f5bda580c60f6382a354848cba29ab4d04c233aaf6827b71b4c4f138a15b7a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 135, + 175, + 211, + 255 + ], + "blocks": [ + { + "bbox": [ + 135, + 175, + 211, + 255 + ], + "lines": [ + { + "bbox": [ + 135, + 175, + 211, + 255 + ], + "spans": [ + { + "bbox": [ + 135, + 175, + 211, + 255 + ], + "type": "image", + "image_path": "013ef28dda0aa6516e2c7b7b2041a96e02a5bc50db6bf09f2d58c27cd5967f55.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 255, + 184, + 262 + ], + "lines": [ + { + "bbox": [ + 162, + 255, + 184, + 262 + ], + "spans": [ + { + "bbox": [ + 162, + 255, + 184, + 262 + ], + "type": "text", + "content": "AGRoL" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 216, + 70, + 293, + 121 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 293, + 121 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 293, + 121 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 293, + 121 + ], + "type": "image", + "image_path": "8e32435c167ea98cd5bede7271e9673202fd37f468b5aaa59d4cdcd679a0831a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 216, + 121, + 292, + 175 + ], + "blocks": [ + { + "bbox": [ + 216, + 121, + 292, + 175 + ], + "lines": [ + { + "bbox": [ + 216, + 121, + 292, + 175 + ], + "spans": [ + { + "bbox": [ + 216, + 121, + 292, + 175 + ], + "type": "image", + "image_path": "c0b30636be4608df50ef55d4ebe1e69b7a348c909fe51f336c82246e82a86991.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 216, + 175, + 292, + 255 + ], + "blocks": [ + { + "bbox": [ + 216, + 175, + 292, + 255 + ], + "lines": [ + { + "bbox": [ + 216, + 175, + 292, + 255 + ], + "spans": [ + { + "bbox": [ + 216, + 175, + 292, + 255 + ], + "type": "image", + "image_path": "6e3bd1f9f176c5b124822d96098b04117c686403d55ea71857aebcd030a7f80a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 255, + 275, + 262 + ], + "lines": [ + { + "bbox": [ + 244, + 255, + 275, + 262 + ], + "spans": [ + { + "bbox": [ + 244, + 255, + 275, + 262 + ], + "type": "text", + "content": "AvatarJLM" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 374, + 121 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 374, + 121 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 374, + 121 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 374, + 121 + ], + "type": "image", + "image_path": "2e01b1e12b18e5edaee2dfe4088cfaf58ffca40a127134ecbe8baf44e0696574.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 298, + 121, + 373, + 175 + ], + "blocks": [ + { + "bbox": [ + 298, + 121, + 373, + 175 + ], + "lines": [ + { + "bbox": [ + 298, + 121, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 298, + 121, + 373, + 175 + ], + "type": "image", + "image_path": "7ef11262d323e7f1ed946b4260bd898972639512d281c2a11b284f8f21805dd1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 298, + 175, + 373, + 255 + ], + "blocks": [ + { + "bbox": [ + 298, + 175, + 373, + 255 + ], + "lines": [ + { + "bbox": [ + 298, + 175, + 373, + 255 + ], + "spans": [ + { + "bbox": [ + 298, + 175, + 373, + 255 + ], + "type": "image", + "image_path": "ef03fa587ae612799b7066941ae87a89a87c879bb9384ee38d1685d560e59d7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 255, + 354, + 262 + ], + "lines": [ + { + "bbox": [ + 326, + 255, + 354, + 262 + ], + "spans": [ + { + "bbox": [ + 326, + 255, + 354, + 262 + ], + "type": "text", + "content": "S2Fusion*" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 380, + 70, + 455, + 121 + ], + "blocks": [ + { + "bbox": [ + 380, + 70, + 455, + 121 + ], + "lines": [ + { + "bbox": [ + 380, + 70, + 455, + 121 + ], + "spans": [ + { + "bbox": [ + 380, + 70, + 455, + 121 + ], + "type": "image", + "image_path": "4dd84c13ca8b0d4abc7cbfee2af35b369e456fce4a58343fb1e6670bd8999906.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 380, + 121, + 454, + 175 + ], + "blocks": [ + { + "bbox": [ + 380, + 121, + 454, + 175 + ], + "lines": [ + { + "bbox": [ + 380, + 121, + 454, + 175 + ], + "spans": [ + { + "bbox": [ + 380, + 121, + 454, + 175 + ], + "type": "image", + "image_path": "6d1ab4d43dfb3ee93fe6b801c8ca4073a9ce5ab4c0845312dfad667cb85fcfe6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 380, + 175, + 454, + 255 + ], + "blocks": [ + { + "bbox": [ + 380, + 175, + 454, + 255 + ], + "lines": [ + { + "bbox": [ + 380, + 175, + 454, + 255 + ], + "spans": [ + { + "bbox": [ + 380, + 175, + 454, + 255 + ], + "type": "image", + "image_path": "53265b806f4c1dc67114f4ef9e89bb600d8ab26c67820249454b5e3f0957f04b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 255, + 435, + 262 + ], + "lines": [ + { + "bbox": [ + 408, + 255, + 435, + 262 + ], + "spans": [ + { + "bbox": [ + 408, + 255, + 435, + 262 + ], + "type": "text", + "content": "EnvPoser" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 460, + 70, + 538, + 121 + ], + "blocks": [ + { + "bbox": [ + 460, + 70, + 538, + 121 + ], + "lines": [ + { + "bbox": [ + 460, + 70, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 460, + 70, + 538, + 121 + ], + "type": "image", + "image_path": "c7548ddbc7ee007cc2b5d476cfc9907142e3039681ee606d1b5cfc18d862aabc.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 460, + 121, + 538, + 175 + ], + "blocks": [ + { + "bbox": [ + 460, + 121, + 538, + 175 + ], + "lines": [ + { + "bbox": [ + 460, + 121, + 538, + 175 + ], + "spans": [ + { + "bbox": [ + 460, + 121, + 538, + 175 + ], + "type": "image", + "image_path": "4406b8fd50c52b55be0afe8818029e09c08eb0701641b609339d5d6e07ef1ccb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 460, + 175, + 538, + 255 + ], + "blocks": [ + { + "bbox": [ + 460, + 175, + 538, + 255 + ], + "lines": [ + { + "bbox": [ + 460, + 175, + 538, + 255 + ], + "spans": [ + { + "bbox": [ + 460, + 175, + 538, + 255 + ], + "type": "image", + "image_path": "e07d7e0960ee82bf32007e269e50eb65d143d2604ebc91b567b06f5eabcbf5fa.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 255, + 506, + 263 + ], + "lines": [ + { + "bbox": [ + 496, + 255, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 496, + 255, + 506, + 263 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 50, + 284, + 298, + 443 + ], + "blocks": [ + { + "bbox": [ + 50, + 284, + 298, + 443 + ], + "lines": [ + { + "bbox": [ + 50, + 284, + 298, + 443 + ], + "spans": [ + { + "bbox": [ + 50, + 284, + 298, + 443 + ], + "type": "table", + "html": "
EgoBody
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]6.8097.8237.311.0
AGRoL[6]7.21100.9370.321.1
AvatarJLM[55]6.4291.7177.47.2
S2Fusion*[35]6.6589.2219.412.5
EnvPoser6.0074.7174.06.6
GIMO
MPJRE(°)MPJPE(mm)MPJVE(mm/s)Jitter
AvatarPoser[14]7.0291.3324.016.4
AGRoL[6]6.5888.6269.412.5
AvatarJLM[55]4.9570.7258.110.7
S2Fusion[35]4.6557.8235.710.1
EnvPoser4.3857.6234.68.9
", + "image_path": "8085ac68422a0f1baa5c45d2a261d6315c0d298657bc1b993e3781d680abc2da.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 72, + 445, + 261, + 455 + ], + "lines": [ + { + "bbox": [ + 72, + 445, + 261, + 455 + ], + "spans": [ + { + "bbox": [ + 72, + 445, + 261, + 455 + ], + "type": "text", + "content": "Table 1. The performance comparison with SOTAs." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "text", + "content": "Metrics. Following [35, 55], we use these metrics for evaluation: 1) MPJRE " + }, + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "inline_equation", + "content": "[^\\circ]" + }, + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "text", + "content": ": The mean joint rotational error across all body joints, measured in degrees; 2) MPJPE " + }, + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "inline_equation", + "content": "[cm]" + }, + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "text", + "content": ": The mean positional error per joint, measured in centimeters. 3) MPJVE " + }, + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "inline_equation", + "content": "[cm/s]" + }, + { + "bbox": [ + 46, + 466, + 287, + 574 + ], + "type": "text", + "content": ": Represents the average error in joint velocities, expressed in centimeters per second; 4) Jitter: Measures motion smoothness by calculating the average jerk (rate of change of acceleration) across all joints, with lower values indicating smoother motion." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 586, + 246, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 246, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 246, + 597 + ], + "type": "text", + "content": "4.1. Comparison with the State-of-the-Art" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Quantitative Experimental Analysis. We compare our proposed method with the competing methods on the two public motion-environment interaction datasets. Among these, AvatarPoser [14], AGRoL [6], and AvatarJLM [55] are the SOTAs that rely solely on sparse tracking signals. As shown in Tab. 1, our model, which incorporates pre-scanned environmental point cloud information, effectively constrains the uncertainty in motion reconstruction caused by sparse tracking signals. Compared to Avatar-" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 310, + 284, + 364, + 364 + ], + "blocks": [ + { + "bbox": [ + 120, + 269, + 473, + 281 + ], + "lines": [ + { + "bbox": [ + 120, + 269, + 473, + 281 + ], + "spans": [ + { + "bbox": [ + 120, + 269, + 473, + 281 + ], + "type": "text", + "content": "Figure 3. Visualization of motion estimation on three test sequences from EgoBody Dataset [51]." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 284, + 364, + 364 + ], + "lines": [ + { + "bbox": [ + 310, + 284, + 364, + 364 + ], + "spans": [ + { + "bbox": [ + 310, + 284, + 364, + 364 + ], + "type": "image", + "image_path": "5754a7bc84d97da9af2379af66d01e2a4bc34b3b1e82254311779b2bf66f2682.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 310, + 365, + 364, + 443 + ], + "blocks": [ + { + "bbox": [ + 310, + 365, + 364, + 443 + ], + "lines": [ + { + "bbox": [ + 310, + 365, + 364, + 443 + ], + "spans": [ + { + "bbox": [ + 310, + 365, + 364, + 443 + ], + "type": "image", + "image_path": "26d7eb7f316c4f4b88e773443cf961a2d1eaae70ff1db42660224557dd3c6da8.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 445, + 356, + 453 + ], + "lines": [ + { + "bbox": [ + 321, + 445, + 356, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 445, + 356, + 453 + ], + "type": "text", + "content": "AvatarJLM" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 458, + 527, + 469 + ], + "lines": [ + { + "bbox": [ + 323, + 458, + 527, + 469 + ], + "spans": [ + { + "bbox": [ + 323, + 458, + 527, + 469 + ], + "type": "text", + "content": "Figure 4. Qualitative Comparison of Interaction Details." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 369, + 285, + 422, + 364 + ], + "blocks": [ + { + "bbox": [ + 369, + 285, + 422, + 364 + ], + "lines": [ + { + "bbox": [ + 369, + 285, + 422, + 364 + ], + "spans": [ + { + "bbox": [ + 369, + 285, + 422, + 364 + ], + "type": "image", + "image_path": "1d8a775c966be2865842459aa367303da0d91332de7e016827a4dc4de5938bf5.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 369, + 365, + 422, + 443 + ], + "blocks": [ + { + "bbox": [ + 369, + 365, + 422, + 443 + ], + "lines": [ + { + "bbox": [ + 369, + 365, + 422, + 443 + ], + "spans": [ + { + "bbox": [ + 369, + 365, + 422, + 443 + ], + "type": "image", + "image_path": "eb3e45feca9a6e2ee927513996c620342eac609868b39f138637098e084e152f.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 445, + 412, + 453 + ], + "lines": [ + { + "bbox": [ + 379, + 445, + 412, + 453 + ], + "spans": [ + { + "bbox": [ + 379, + 445, + 412, + 453 + ], + "type": "text", + "content": "S2Fusion*" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 427, + 285, + 481, + 364 + ], + "blocks": [ + { + "bbox": [ + 427, + 285, + 481, + 364 + ], + "lines": [ + { + "bbox": [ + 427, + 285, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 427, + 285, + 481, + 364 + ], + "type": "image", + "image_path": "6c9684560dbce0e51a442427cc50e998eb743463601f6cd2716a9d8e586e1ff1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 427, + 365, + 481, + 443 + ], + "blocks": [ + { + "bbox": [ + 427, + 365, + 481, + 443 + ], + "lines": [ + { + "bbox": [ + 427, + 365, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 427, + 365, + 481, + 443 + ], + "type": "image", + "image_path": "c1ef914f738e78ec581bea0abe9b36200cf678089f222935027308051aae96fd.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 445, + 469, + 453 + ], + "lines": [ + { + "bbox": [ + 440, + 445, + 469, + 453 + ], + "spans": [ + { + "bbox": [ + 440, + 445, + 469, + 453 + ], + "type": "text", + "content": "EnvPoser" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 486, + 285, + 539, + 364 + ], + "blocks": [ + { + "bbox": [ + 486, + 285, + 539, + 364 + ], + "lines": [ + { + "bbox": [ + 486, + 285, + 539, + 364 + ], + "spans": [ + { + "bbox": [ + 486, + 285, + 539, + 364 + ], + "type": "image", + "image_path": "fc6eb79f0d5c5d93ca80447fd4a702ac42e27f32f05b77e22031cc04e1a7d530.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 486, + 365, + 539, + 443 + ], + "blocks": [ + { + "bbox": [ + 486, + 365, + 539, + 443 + ], + "lines": [ + { + "bbox": [ + 486, + 365, + 539, + 443 + ], + "spans": [ + { + "bbox": [ + 486, + 365, + 539, + 443 + ], + "type": "image", + "image_path": "601230f5242a6002a29dd0b1dadf58280575e953e679a1f66834a22079482781.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 445, + 518, + 452 + ], + "lines": [ + { + "bbox": [ + 507, + 445, + 518, + 452 + ], + "spans": [ + { + "bbox": [ + 507, + 445, + 518, + 452 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": "JLM [55], which does not utilize environmental information, our method achieves significant improvements on two key metrics—MPJRE and MPJPE—by 0.42 (6.5%)/17.0 (18.5%) and 0.57 (11.5%)/13.1 (18.5%) on the EgoBody and GIMO datasets, respectively. These results underscore the increased accuracy in motion estimation achieved through the integration of environmental information." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Additionally, we benchmark our model against the latest algorithm S2Fusion [35], which introduces environmental information as a constraint for human motion estimation. On the GIMO dataset, our method reduces MPJRE errors by " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " compared to S2Fusion [35], and it outperforms S2Fusion across all evaluation metrics, demonstrating the superior effectiveness of our approach in leveraging environmental information. Moreover, the motion estimation performance of our method also outperforms S2Fusion* on the EgoBody dataset, which reduces MPJRE and MPJPE errors by " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". Furthermore, our method shows a marked improvement in motion smoothness, with an " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "11.9\\%" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1845" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 272, + 145 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 272, + 145 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 272, + 145 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 272, + 145 + ], + "type": "table", + "html": "
EgoBodyGIMO
MPJREMPJPEMPJREMPJPE
①Baseline7.4587.15.7274.6
②w/o env-semantic6.0478.04.5562.5
③w/o env-geometry6.0377.54.6363.2
EnvPoser6.0074.74.3857.6
", + "image_path": "df34b9791708b799670fe9353dbbc3d54f3e46c1dd79adb9bfe8ef2067701f5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 168, + 286, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 168, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 168, + 286, + 190 + ], + "type": "text", + "content": "reduction in Jitter over S2Fusion and a " + }, + { + "bbox": [ + 48, + 168, + 286, + 190 + ], + "type": "inline_equation", + "content": "16.8\\%" + }, + { + "bbox": [ + 48, + 168, + 286, + 190 + ], + "type": "text", + "content": " improvement over AvatarJLM on the GIMO dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 193, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 193, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 193, + 286, + 371 + ], + "type": "text", + "content": "Qualitative Experimental Analysis. Fig. 3 showcases the motion estimation results on the EgoBody dataset, where we present three representative sequences and visualize the human motion estimation results at key moments. For state-of-the-art algorithms that do not incorporate environmental information, the visualization results in Fig. 3 show multiple instances of environment penetration during interactions. For example, AvatarJLM exhibits errors such as hand intersections with walls (second row) and leg penetrations through boxes (third row). Furthermore, in scenarios without chairs, squatting motions are misinterpreted as sitting (first row) in AvatarJLM due to the absence of environment understanding. In contrast, EnvPoser effectively leverages environmental information to reduce penetration errors and produce more realistic motion estimates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 373, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 286, + 529 + ], + "type": "text", + "content": "In more challenging interaction scenarios, EnvPoser shows certain advantages over S2Fusion, which also incorporates environmental information. Notably, S2Fusion fails to prevent interpenetration between the lower body and surrounding objects (third row in Fig. 3). Moreover, Fig. 4 provides a comparison between EnvPoser and S2Fusion, highlighting the advantages of our approach in handling upper-body interactions with objects and complex lower-body motions through the integration of uncertainty modeling and environmental semantic/geometric constraints. Additionally, compared to AvatarJLM, both EnvPoser and S2Fusion demonstrate the effectiveness of incorporating environmental information to improve human motion estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 538, + 139, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 538, + 139, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 538, + 139, + 551 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 558, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 712 + ], + "type": "text", + "content": "Effectiveness of the environment-aware refinement module. We conducted ablation studies comparing our method with three variants: ① Baseline: an auto-regressive transformer-based network with joint uncertainty estimation; ② w/o env-semantic: incorporating only geometric constraints; ③ w/o env-geometry: incorporating only semantic constraints. Tab. 2 presents the motion estimation performance for each variant. Compared to variant ①, our method achieves substantial improvements by integrating both semantic and geometric environmental information. Further analysis shows that each environment-aware refinement module independently enhances performance, with the combination yielding the best overall results." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 553, + 135 + ], + "blocks": [ + { + "bbox": [ + 52, + 148, + 282, + 159 + ], + "lines": [ + { + "bbox": [ + 52, + 148, + 282, + 159 + ], + "spans": [ + { + "bbox": [ + 52, + 148, + 282, + 159 + ], + "type": "text", + "content": "Table 2. The ablation study on environment refinement module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 553, + 135 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 553, + 135 + ], + "type": "table", + "html": "
EgoBodyGIMO
MPJREMPJPEMPJVEMPJREMPJPEMPJVE
1)Baseline w/o UNC7.6390.8333.05.9077.5270.2
2)Baseline7.4587.1342.95.7274.6266.7
3)EnvPoser w/o UNC6.5680.1172.74.4560.2278.6
EnvPoser6.0074.7174.04.3857.6234.6
", + "image_path": "73d006bac0133b7fef43565c27ca54687413e801eed78ef6b15bfe187d579d36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 140, + 527, + 150 + ], + "lines": [ + { + "bbox": [ + 324, + 140, + 527, + 150 + ], + "spans": [ + { + "bbox": [ + 324, + 140, + 527, + 150 + ], + "type": "text", + "content": "Table 3. The effectiveness of the uncertainty estimation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 307, + 157, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 157, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 157, + 545, + 371 + ], + "type": "text", + "content": "Effectiveness of the uncertainty estimation. As shown in Tab. 3, we compare two variants: 1) Baseline w/o UNC: a baseline without uncertainty estimation and 2) Baseline: a baseline with uncertainty estimation. Results show that integrating the uncertainty estimation module not only quantifies joint uncertainty across motions but also improves full-body motion estimation through additional supervision. We further incorporated both baseline methods into the environment-aware refinement module. As shown in Tab. 3, our method achieves significant improvements over variant 3) EnvPoser w/o UNC: our method without uncertainty estimation, reducing MPJRE and MPJPE errors by " + }, + { + "bbox": [ + 307, + 157, + 545, + 371 + ], + "type": "inline_equation", + "content": "8.11\\%" + }, + { + "bbox": [ + 307, + 157, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 307, + 157, + 545, + 371 + ], + "type": "inline_equation", + "content": "5.68\\%" + }, + { + "bbox": [ + 307, + 157, + 545, + 371 + ], + "type": "text", + "content": ", respectively, on the EgoBody dataset. These results indicate that the uncertainty estimation and resampling methods effectively capture the multi-hypothesis nature of initial motion estimates and, when combined with environmental refinement, adaptively align motion estimation with sparse tracking signals and environmental context." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 382, + 382, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 382, + 394 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 382, + 394 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 403, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 545, + 568 + ], + "type": "text", + "content": "This paper presents a novel two-stage framework for reliable human motion estimation leveraging sparse tracking signals and environmental context. In the first stage, we model the multi-hypothesis nature of motion through uncertainty estimation, followed by a refinement phase that applies environment-aware constraints from both semantic and geometric perspectives. Extensive experiments show that EnvPoser outperforms existing methods, particularly in complex interaction scenarios, highlighting the significance of environmental context in wearable motion estimation and paving the way for future approaches to harness this valuable contextual information. This work has the potential to impact a wide range of interactive applications that benefit from more accurate and context-aware motion estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 712 + ], + "type": "text", + "content": "Limitations and future works. The model assumes a static environment, lacking consideration for dynamic multi-user interactions or object motion, which may limit its effectiveness. This limitation could be addressed by incorporating a third-person perspective to estimate the movement of other people and objects, along with additional semantic and geometric constraints. Furthermore, real-time mesh quality in complex environments is often limited, which may affect performance. Future work will explore using raw images to infer contacts in complex scenes via 2D semantic understanding and introducing additional constraints to improve overall reconstruction quality." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "text", + "content": "1846" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Sadegh Aliakbarian, Fatemeh Saleh, David Collier, Pashmina Cameron, and Darren Cosker. Hmd-nemo: Online 3d avatar motion generation from sparse observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9622-9631, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 247 + ], + "type": "text", + "content": "[3] Rongyu Chen, Linlin Yang, and Angela Yao. Mhentropy: Entropy meets multiple hypotheses for pose and shape recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14840-14849, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 312 + ], + "type": "text", + "content": "[4] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 380 + ], + "type": "text", + "content": "[5] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating smpl poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 446 + ], + "type": "text", + "content": "[6] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "type": "text", + "content": "[7] Zhen Fan, Peng Dai, Zhuo Su, Xu Gao, Zheng Lv, Jiarui Zhang, Tianyuan Du, Guidong Wang, and Yang Zhang. Emhi: A multimodal egocentric human motion dataset with hmd and body-worn imus. arXiv preprint arXiv:2408.17168, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "text", + "content": "[8] Han Feng, Wenchao Ma, Quankai Gao, Xianwei Zheng, Nan Xue, and Huijuan Xu. Stratified avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 153-163, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "text", + "content": "[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "text", + "content": "[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "text", + "content": "[11] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "text", + "content": "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 218 + ], + "type": "text", + "content": "[13] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: learning to reconstruct human pose from sparse inertial measurements in real time. ACM Trans. Graph. (TOG), 37(6):1-15, 2018. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 547, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 275 + ], + "type": "text", + "content": "[14] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 331 + ], + "type": "text", + "content": "[15] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time egocentric pose estimation from sparse and intermittent observations everywhere. In European Conference on Computer Vision, pages 277-294. Springer, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 547, + 397 + ], + "type": "text", + "content": "[16] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 399, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 464 + ], + "type": "text", + "content": "[17] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers (SA' 22), pages 1-9, 2022. 2, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 466, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 547, + 532 + ], + "type": "text", + "content": "[18] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision(ICCV), pages 11510-11520, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 533, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 547, + 567 + ], + "type": "text", + "content": "[19] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 568, + 547, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 547, + 622 + ], + "type": "text", + "content": "[20] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 624, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 547, + 669 + ], + "type": "text", + "content": "[21] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "type": "text", + "content": "[22] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1847" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[23] Han Li, Bowen Shi, Wenrui Dai, Hongwei Zheng, Botao Wang, Yu Sun, Min Guo, Chenglin Li, Junni Zou, and Hongkai Xiong. Pose-oriented transformer with uncertainty-guided refinement for 2d-to-3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1296–1304, 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "text", + "content": "[24] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "text", + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "type": "text", + "content": "[27] Zhenyu Lou, Qiongjie Cui, Haofan Wang, Xu Tang, and Hong Zhou. Multimodal sense-informed forecasting of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2144-2154, 2024. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 409 + ], + "type": "text", + "content": "[28] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "text", + "content": "[29] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. Coap: Compositional articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13201-13210, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "text", + "content": "[30] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "text", + "content": "[31] Jose Luis Ponton, Eduard Pujol, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Dragposer: Motion reconstruction from variable sparse tracking signals via latent space optimization. arXiv preprint arXiv:2406.14567, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 567, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 612 + ], + "type": "text", + "content": "[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[33] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[34] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "text", + "content": "[35] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 151, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 547, + 215 + ], + "type": "text", + "content": "[36] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusionposer: Real-time human motion reconstruction from arbitrary sparse sensors using autoregressive diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2513-2523, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 217, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 238 + ], + "type": "text", + "content": "[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[38] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11199-11208, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "text", + "content": "[39] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[40] Yinghao Wu, chaoran wang, Lu Yin, Shihui Guo, and Yipeng Qin. Accurate and steady inertial pose estimation through sequence structure learning and modulation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "text", + "content": "[41] Songpengcheng Xia, Lei Chu, Ling Pei, Zixuan Zhang, Wenxian Yu, and Robert C Qiu. Learning disentangled representation for mixed-reality human activity recognition with a single imu sensor. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "text", + "content": "[42] Songpengcheng Xia, Lei Chu, Ling Pei, Jiarui Yang, Wenxian Yu, and Robert C Qiu. Timestamp-supervised wearable-based activity segmentation and recognition with contrastive learning and order-preserving optimal transport. IEEE Transactions on Mobile Computing, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "text", + "content": "[43] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–989, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "text", + "content": "[44] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22268-22278, 2023. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[45] Hongdi Yang, Chengyang Li, Zhenxuan Wu, Gaozheng Li, Jingya Wang, Jingyi Yu, Zhuo Su, and Lan Xu. Smgdiff: Soccer motion generation using diffusion probabilistic models. arXiv preprint arXiv:2411.16216, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[46] Xueyuan Yang, Chao Yao, and Xiaojuan Ban. Spatial-related sensors matters: 3d human motion reconstruction assisted with textual semantics. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10225-10233, 2024. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1848" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 641 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[47] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[48] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 239 + ], + "type": "text", + "content": "[49] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), pages 13167-13178, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "type": "text", + "content": "[50] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[51] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision, pages 180-200. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "text", + "content": "[52] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 408, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 287, + 462 + ], + "type": "text", + "content": "[53] Yufei Zhang, Hanjing Wang, Jeffrey O Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9020-9032, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "text", + "content": "[54] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "type": "text", + "content": "[55] Xiaozheng Zheng, Zhuo Su, Chao Wen, Zhou Xue, and Xiaojie Jin. Realistic full-body tracking from sparse observations via joint-level modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14678-14688, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 586, + 287, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 641 + ], + "type": "text", + "content": "[56] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision, pages 676-694. Springer, 2022. 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1849" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_content_list.json b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a67a90794721161520faab3042c4a360c22d6ab7 --- /dev/null +++ b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_content_list.json @@ -0,0 +1,1894 @@ +[ + { + "type": "text", + "text": "EquiPose: Exploiting Permutation Equivariance for Relative Camera Pose Estimation", + "text_level": 1, + "bbox": [ + 254, + 130, + 743, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuzhen Liu1,2 Qialei Dong1,2,*", + "bbox": [ + 370, + 202, + 632, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, Chinese Academy of Sciences, Institute of Automation $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences liuyuzhen2022@ia.ac.cn, qldong@nlpr.ia.ac.cn", + "bbox": [ + 129, + 223, + 867, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 311, + 326, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Relative camera pose estimation between two images is a fundamental task in 3D computer vision. Recently, many relative pose estimation networks have been explored for learning a mapping from two input images to their corresponding relative pose, however, the estimated relative poses by these methods do not have the intrinsic Pose Permutation Equivariance (PPE) property: the estimated relative pose from Image A to Image B should be the inverse of that from Image B to Image A. It means that permuting the input order of two images would cause these methods to obtain inconsistent relative poses. To address this problem, we firstly introduce the concept of PPE mapping, which indicates such a mapping that captures the intrinsic PPE property of relative poses. Then by enforcing the aforementioned PPE property, we propose a general framework for relative pose estimation, called EquiPose, which could easily accommodate various relative pose estimation networks in literature as its baseline models. We further theoretically prove that the proposed EquiPose framework could guarantee that its obtained mapping is a PPE mapping. Given a pre-trained baseline model, the proposed EquiPose framework could improve its performance even without fine-tuning, and could further boost its performance with fine-tuning. Experimental results on four public datasets demonstrate that EquiPose could significantly improve the performances of various state-of-the-art models.", + "bbox": [ + 88, + 344, + 483, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 777, + 223, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating the relative camera pose between an image pair is one of the fundamental tasks in 3D computer vision [6, 46]. It has a broad range of applications, such as structure-from-motion (SfM) [14, 15, 42], visual localization [10, 40, 45], novel view synthesis [7, 53, 54], etc.", + "bbox": [ + 89, + 804, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods for relative camera pose estimation in literature could be roughly divided into two categories: matching-based methods and end-to-end methods. The matching-based methods [13, 33] generally employ a two-stage computational strategy: Firstly, they establish the point correspondences between an input pair of images by utilizing local feature extraction and matching techniques [13, 33]. Then according to the established correspondences, they solve a two-view geometric problem via a typical 5-point algorithm [24] with robust fitting [21] for pursuing the relative pose. However, since the performances of the matching-based methods depend on the established point correspondences, their performance would significantly decrease or even fail in cases of weak/repeated textures and degenerated configurations as indicated in [20].", + "bbox": [ + 511, + 313, + 906, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike the above matching-based methods, many recent methods [6, 12, 19, 28, 36, 46, 57, 61] have employed deep neural networks to directly learn a mapping from the input two images to their relative pose without explicitly establishing point correspondences. Here, it has to be pointed out that ideally, the estimated relative pose by these end-to-end methods should have the following intrinsic pose permutation equivariance (PPE) property between an image pair:", + "bbox": [ + 511, + 541, + 908, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pose Permutation Equivariance", + "text_level": 1, + "bbox": [ + 539, + 675, + 764, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Let $\\{I_1, I_2\\}$ be a pair of images, then the relative pose $P_{12}$ from $I_1$ to $I_2$ should be the inverse of the relative pose $P_{21}$ from $I_2$ to $I_1$ :", + "bbox": [ + 535, + 702, + 880, + 748 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\nP _ {2 1} = P _ {1 2} ^ {- 1}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 666, + 758, + 880, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By writing $P_{12}$ and $P_{21}$ in the matrix form:", + "bbox": [ + 532, + 796, + 821, + 811 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\nP _ {1 2} = \\left[ \\begin{array}{c c} R _ {1 2} & t _ {1 2} \\\\ 0 & 1 \\end{array} \\right], \\quad P _ {2 1} = \\left[ \\begin{array}{c c} R _ {2 1} & t _ {2 1} \\\\ 0 & 1 \\end{array} \\right], \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 824, + 906, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $R_{12}$ and $t_{12}$ are the relative rotation and translation from $I_1$ to $I_2$ , and $R_{21}$ and $t_{21}$ are the relative rotation and", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1127", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95d0c77578704945fac8104711d83f639038b1d61035d9c8a87370e9d4843937.jpg", + "image_caption": [ + "(a) MF6D" + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 312, + 202 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a1e6494f3bbda22b17b6a21ccb77bac05e18b6b496cdc5dd3e7f27567f80a1f.jpg", + "image_caption": [ + "Figure 1. Distribution of the $d_{R}$ and $d_{t}$ obtained by existing models MF6D [1] and GRelPose [28]." + ], + "image_footnote": [], + "bbox": [ + 313, + 88, + 493, + 202 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7177829fded36f4b5fde7718d37c7cf190ca49c4d3791bae9c79c12a40577087.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 679, + 202 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/948ac687236a6619f2a1ca5c04e11d47a6ca809eefdc519933841743e0ec182a.jpg", + "image_caption": [ + "(b) GRelPose" + ], + "image_footnote": [], + "bbox": [ + 681, + 89, + 862, + 203 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "translation from $I_{2}$ to $I_{1}$ , Eq. (1) could be reformulated as:", + "bbox": [ + 89, + 255, + 478, + 271 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nR _ {2 1} = R _ {1 2} ^ {\\mathrm {T}}, \\quad t _ {2 1} = - R _ {1 2} ^ {\\mathrm {T}} t _ {1 2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 277, + 482, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the PPE property is generally overlooked when designing model architectures in existing end-to-end methods [1, 28]. We validate this by conducting the following experiment: we trained two typical end-to-end models (MF6D [1] and GRelPose [28]) on the training set of ScanNet [11]. Then, we adopt the test set of ScanNet [48, 50] which contains 1500 image pairs, and feed them with two different orders to the above models respectively, leading to totally 1500 pairs of the predicted relative poses by each model, where each pair consists of the estimated pose $P_{12}$ from $I_{1}$ to $I_{2}$ and the estimated pose $P_{21}$ from $I_{2}$ to $I_{1}$ . Next, we compute the differences between the estimated relative pose $P_{21}$ and $P_{12}^{-1}$ . Specifically, we compute the angular difference $d_{R}$ between relative rotations $R_{21}$ and $R_{12}^{\\mathrm{T}}$ , and the Euclidean distance $d_{t}$ between relative translations $t_{21}$ and $-R_{12}^{\\mathrm{T}} t_{12}$ respectively via:", + "bbox": [ + 89, + 303, + 483, + 546 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nd _ {R} = \\arccos ((\\operatorname {t r} (R _ {2 1} R _ {1 2}) - 1) / 2), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 553, + 482, + 577 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nd _ {t} = \\left\\| t _ {2 1} + R _ {1 2} ^ {\\mathrm {T}} t _ {1 2} \\right\\| _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 573, + 325, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathrm{tr}(\\cdot)$ denotes matrix trace, $\\| \\cdot \\| _2$ denotes $L_{2}$ norm.", + "bbox": [ + 89, + 598, + 482, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The distributions of the rotation difference $d_{R}$ and translation difference $d_{t}$ from the above two models are shown in Fig. 1. As seen from this figure, the pose differences (both rotation and translation) estimated by these models are generally larger than zero (particularly some of them are not trivial), demonstrating that the above models do not satisfy the PPE property. Such an issue may pose risks to downstream applications that require a stable and consistent estimation of relative poses, such as robotics and visual navigation. Moreover, it is also noted that some other visual tasks involve their specific equivariances, different from the PPE property in the relative pose estimation task, and once an existing model among these tasks is equipped with a task-specific equivariance, its performance could be further boosted [4, 27, 35, 39, 43].", + "bbox": [ + 88, + 613, + 482, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above issues, we investigate whether it is possible to explore such a relative pose estimation framework that could not only bestow the PPE property on various relative pose estimation networks but also improve their", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performances further. Specifically, we first introduce the concept of PPE mapping, which refers to a mapping that captures the PPE property in relative poses. Then, we establish the foundational principle for equivariant relative pose estimation. Accordingly, we propose the EquiPose framework which holds the PPE property theoretically. The EquiPose could improve the performance of a pre-trained model even without fine-tuning, and fine-tuning could enhance their performances further.", + "bbox": [ + 511, + 255, + 903, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions include:", + "bbox": [ + 531, + 392, + 843, + 405 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Equivariance: The simple but intrinsic pose permutation equivariance (PPE) property is revealed for relative pose estimation, and the PPE mapping is defined.", + "- Generality: By imposing the PPE property, the framework EquiPose for equivariant relative pose estimation is proposed. An arbitrary end-to-end relative pose estimation model could be seamlessly embedded into this framework as a baseline, and accordingly obtains the PPE property as analyzed in theory in Sec. 3. The EquiPose framework and the theoretical analyses might give new insights into model design strategies for relative pose estimation.", + "- Flexibility: Thanks to the imposed PPE property, the performances of the derived methods from the EquiPose framework by utilizing many off-the-shelf baseline models, would be boosted even without fine-tuning by adding several lines of code. Moreover, once they are fine-tuned under EquiPose, their performances would be further improved, as demonstrated by the results in Sec. 4." + ], + "bbox": [ + 513, + 407, + 905, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 693, + 656, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we review some typical matching-based methods and end-to-end methods for relative pose estimation respectively.", + "bbox": [ + 511, + 719, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Matching-Based Relative Pose Estimation. The matching-based relative pose estimation methods typically involve the following two steps: establishing point correspondence via image matching [3, 33] and solving the two-view geometry with RANSAC [21, 24, 41]. Early approaches relied on handcrafted keypoint detectors and descriptors to establish point correspondences, such as SIFT [33] and SURF [3]. In recent years, numerous methods emerged that replace the handcrafted detectors or descrip", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1128", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tors with their learned counterparts [2, 17, 18, 32, 34, 37, 44, 51, 59]. Additionally, a variety of learning-based image-matching methods directly established point correspondences from keypoints or raw images [8, 31, 48, 50, 60].", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, since these methods depend on point correspondences, they may fail in cases of weak textures, repeated textures, large viewpoint changes, and degenerated configurations [20]. Additionally, these methods could be time-consuming due to image matching and RANSAC loops. Moreover, they could only recover the relative pose up to an unknown scale factor [24].", + "bbox": [ + 89, + 152, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "End-to-End Relative Pose Estimation. Recently, unlike the above strategies, many works have adopted deep neural networks to regress relative poses in an end-to-end manner [1, 6, 9, 12, 19, 28, 36]. These methods do not explicitly implement image matching or solve the two-view geometry problem. Instead, the networks directly learn the underlying mapping from the input image pair to the corresponding relative pose. Compared with matching-based methods, they are shown to be more robust in challenging scenes [6, 9, 19] with a significantly faster inference speed. Moreover, these end-to-end methods could be further incorporated with traditional methods for improving robustness [47].", + "bbox": [ + 89, + 257, + 483, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These end-to-end methods generally adopt Siamese architectures [5] to extract feature, which are aggregated and decoded into the relative pose subsequently [6, 12, 26, 28, 46]. To facilitate stable training, early approaches [19, 36] parameterized the rotation as quaternions. Different from them, Zhou et al. [62] suggested to represent 3D rotations in a 6D space, which benefits the training of neural networks. Cai et al. [6] and Dekel et al. [12] showed that the network could learn from implicit cues, allowing the network to estimate the relative pose even when two images have no overlap. Rockwell et al. [46] introduced the essential matrix module to emulate the 8-point algorithm [24]. Arnold et al. [1] adopted a differentiable warping module for feature aggregation. Khatib et al. [28] embedded a semidense image matcher [50] for feature extraction to pursue a better generalization ability.", + "bbox": [ + 89, + 439, + 483, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the existing end-to-end methods generally do not satisfy the PPE property, leading to inconsistent estimation when permuting the input order of an image pair as shown in Sec. 1. Unlike them, this work does not aim to propose a new model for relative pose estimation, but pursue a general framework that could bestow the PPE property on various relative pose estimation models, and improve their performances further.", + "bbox": [ + 89, + 680, + 483, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 815, + 225, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the key definitions and propositions related to the pose permutation equivariance (PPE) property of relative poses. Then, we describe the proposed EquiPose framework in detail.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. PPE Mapping and Preliminary Knowledge", + "text_level": 1, + "bbox": [ + 511, + 90, + 879, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior to the introduction of the PPE mapping, 'equivariant mapping' in mathematics is defined as [30]:", + "bbox": [ + 511, + 114, + 906, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Equivariant mapping). Let $\\mathcal{G}$ be a group that acts on spaces $\\mathcal{X}$ and $\\mathcal{Y}$ , and $\\mathbb{T}$ and $\\mathbb{T}'$ be the induced actions of $\\mathcal{G}$ on $\\mathcal{X}$ and $\\mathcal{Y}$ . A mapping $f: \\mathcal{X} \\to \\mathcal{Y}$ is equivariant with the action in $\\mathcal{G}$ if $f(\\mathbb{T}_g(x)) = \\mathbb{T}_g'(f(x)), \\forall x \\in \\mathcal{X}, g \\in \\mathcal{G}$ .", + "bbox": [ + 511, + 157, + 906, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to Definition 1 and Eq. (1), a PPE mapping with the PPE property is defined as:", + "bbox": [ + 511, + 244, + 906, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Pose permutation equivariant mapping). A mapping $f: \\mathbb{R}^{H \\times W} \\times \\mathbb{R}^{H \\times W} \\to \\mathrm{SE}(3)$ (the special Euclidean group) is pose permutation equivariant, if $f(I_2, I_1) = f(I_1, I_2)^{-1}, \\forall I_1, I_2 \\in \\mathbb{R}^{H \\times W}$ .", + "bbox": [ + 511, + 287, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 could be considered as an instance of 'equivariant mapping' from Definition 1 on relative camera poses, where the input variable $x$ is an image pair $\\{I_1, I_2\\}$ , the action $\\mathbb{T}_g$ on $\\{I_1, I_2\\}$ is an order permutation operator, and the action $\\mathbb{T}_q'$ is the inverse of a matrix.", + "bbox": [ + 511, + 361, + 905, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is noted that Definition 2 is applicable to neural networks, since any neural networks essentially aim to learn a mapping from their input to the output.", + "bbox": [ + 511, + 438, + 905, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to the above two definitions, we introduce the following definition on the geodesic mean of rotations [22, 23], as well as two theorems, which would be used to support the fundamental principle for designing a relative pose estimation framework with the PPE property in the next subsection:", + "bbox": [ + 511, + 484, + 905, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3 (Geodesic mean of rotations). Given a set of rotations $\\{R_i\\}_{i=1}^n$ , their geodesic mean $S$ is defined as:", + "bbox": [ + 511, + 587, + 906, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS = \\mathfrak {G} \\left(R _ {1}, \\dots , R _ {n}\\right) = \\underset {S \\in \\mathrm {S O} (3)} {\\arg \\min } \\sum_ {i = 1} ^ {n} \\| \\log \\left(R _ {i} ^ {\\mathrm {T}} S\\right) \\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 631, + 906, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\log (\\cdot)$ is the mapping from an element in the Lie group to its corresponding Lie algebra.", + "bbox": [ + 511, + 683, + 905, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From Definition 3, we have the two theorems as:", + "bbox": [ + 531, + 726, + 854, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. Given a set of rotations $\\{R_i\\}_{i=1}^n$ , it always holds that: $\\mathfrak{G}(R_{\\sigma(1)}, \\dots, R_{\\sigma(n)}) = \\mathfrak{G}(R_1, \\dots, R_n)$ for any permutation $\\sigma$ .", + "bbox": [ + 511, + 753, + 906, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 2. If $S \\in \\mathrm{SO}(3)$ is the geodesic mean of the set of rotations $\\{R_i\\}_{i=1}^n$ , then $S^{\\mathrm{T}}$ is the geodesic mean of the set of rotations $\\{R_i^{\\mathrm{T}}\\}_{i=1}^n$ .", + "bbox": [ + 511, + 811, + 905, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the space limitation, the proofs of the above two theorems are given in the supplementary material.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1129", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/93c2cc69918823371827c0cec9501ef7356e583cfb66fa72da7c1d62d2638db8.jpg", + "image_caption": [ + "Figure 2. Illustration of the existing methods $f$ and our method $f^{*}$ on the SE(3) manifold. Left: The existing method $f$ does not capture the PPE property in the relative pose. Right: Within EquiPose $f^{*}$ , the desired PPE property is satisfied (Theorem 3), and the results could be more reliable (Sec. 4)." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 472, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Fundamental Principle of Pose Estimation Framework with PPE", + "text_level": 1, + "bbox": [ + 89, + 275, + 482, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the aforementioned PPE mapping and related preliminary knowledge, we elaborate the fundamental principle for designing a pose estimation framework with PPE, which could use an arbitrary end-to-end relative pose estimation model as its baseline and derive the corresponding upgraded model with the PPE property as follows:", + "bbox": [ + 89, + 313, + 483, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an arbitrary relative pose estimation network corresponding to a mapping $f$ , there exists the following mapping $f^{*}$ , which is a PPE mapping:", + "bbox": [ + 89, + 404, + 483, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3. Given an arbitrary mapping $f: \\mathbb{R}^{H \\times W} \\times \\mathbb{R}^{H \\times W} \\to \\mathrm{SE}(3)$ , where $f(I_1, I_2) := (R_{12}, t_{12})$ and $f(I_2, I_1) := (R_{21}, t_{21})$ , the mapping $f^*$ defined as:", + "bbox": [ + 89, + 455, + 483, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {*} \\left(I _ {1}, I _ {2}\\right) := \\left(R _ {1 2} ^ {*}, t _ {1 2} ^ {*}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 508, + 480, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 89, + 534, + 137, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 2} ^ {*} = \\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 555, + 482, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {1 2} ^ {*} = \\left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\\right) / 2, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 575, + 482, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is a PPE mapping.", + "bbox": [ + 89, + 599, + 215, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. According to Theorem 1, we have", + "bbox": [ + 89, + 621, + 367, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 2} ^ {*} = \\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right) = \\mathfrak {G} \\left(R _ {2 1} ^ {\\mathrm {T}}, R _ {1 2}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 648, + 482, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, according to Theorem 2, we have:", + "bbox": [ + 109, + 672, + 380, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathfrak {G} \\left(R _ {2 1} ^ {\\mathrm {T}}, R _ {1 2}\\right) = \\mathfrak {G} \\left(R _ {2 1}, R _ {1 2} ^ {\\mathrm {T}}\\right) ^ {\\mathrm {T}} = R _ {2 1} ^ {* \\mathrm {T}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 691, + 482, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Eqs. (9) and (10), it holds:", + "bbox": [ + 109, + 719, + 375, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 2} ^ {*} = R _ {2 1} ^ {* \\mathrm {T}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 739, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "i.e., $f^{*}$ is permutation equivariant in terms of rotation.", + "bbox": [ + 89, + 765, + 447, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, according to (6), (7), (8) and (11), it holds:", + "bbox": [ + 109, + 780, + 431, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f ^ {*} (I _ {1}, I _ {2}) ^ {- 1} = \\left(R _ {1 2} ^ {* \\mathrm {T}}, - R _ {1 2} ^ {* \\mathrm {T}} t _ {1 2} ^ {*}\\right) \\\\ = \\left(R _ {1 2} ^ {* \\mathrm {T}}, - R _ {1 2} ^ {* \\mathrm {T}} \\left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\\right) / 2\\right) \\tag {12} \\\\ = \\left(R _ {2 1} ^ {*}, \\left(t _ {2 1} - R _ {2 1} ^ {*} t _ {1 2}\\right) / 2\\right) \\\\ = f ^ {*} \\left(I _ {2}, I _ {1}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 801, + 480, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which completes the proof.", + "bbox": [ + 89, + 885, + 272, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da3249d170907e792178917639d91d45072a5375612001bac512854000cab9e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 886, + 480, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 The EquiPose Framework", + "text_level": 1, + "bbox": [ + 514, + 90, + 777, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: A relative pose estimation network $f$ ; A pair of images $\\{I_1, I_2\\}$", + "bbox": [ + 517, + 109, + 821, + 141 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Equivariant pose estimation $R_{12}^{*}, t_{12}^{*}$", + "bbox": [ + 517, + 141, + 823, + 156 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $R_{12}, t_{12} \\gets f(I_1, I_2)$", + "2: $R_{21}, t_{21} \\gets f(I_2, I_1)$", + "3: $R_{12}^{*} \\gets R_{12}(\\exp \\left(\\frac{1}{2}\\log (R_{12}^{\\mathrm{T}}R_{21}^{\\mathrm{T}})\\right))$", + "4: $t_{12}^{*}\\gets (t_{12} - R_{12}^{*}t_{21}) / 2$" + ], + "bbox": [ + 524, + 156, + 771, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For ease of understanding, we visualize the difference between $f$ and $f^{*}$ in Fig. 2. As seen from this figure, given two input images $\\{I_1, I_2\\}$ , the existing relative pose estimation model (considered as a mapping) $f$ does not satisfy the PPE property. Once it is embedded into EquiPose, the upgrade version $f^{*}$ captures the PPE property strictly. Moreover, its estimation could be more reliable than the baseline model $f$ , as demonstrated in Sec. 4 (also in Sec. S8.3 of the supplementary material).", + "bbox": [ + 511, + 247, + 903, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is noted in Eq. (7) that the construction of $f^{*}$ involves computing the geodesic mean of two rotations, which could be computed explicitly via [38]:", + "bbox": [ + 511, + 383, + 905, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right) = R _ {1 2} \\left(R _ {1 2} ^ {\\mathrm {T}} R _ {2 1} ^ {\\mathrm {T}}\\right) ^ {1 / 2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 439, + 903, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that we work on the 3D rotation group, the square root of a rotation $R \\in \\mathrm{SO}(3)$ could be computed by:", + "bbox": [ + 511, + 470, + 905, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {1 / 2} = \\exp \\left(\\frac {1}{2} \\log (R)\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 513, + 903, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\exp (\\cdot)$ is the mapping from the Lie algebra to the Lie group.", + "bbox": [ + 511, + 549, + 903, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above analysis reveals that given an arbitrary relative pose estimation network, once its upgrade version $f^{*}$ is constructed by (6), the mapping is a pose permutation equivariant mapping, this is to say, the upgrade version has the PPE property.", + "bbox": [ + 511, + 579, + 905, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. EquiPose", + "text_level": 1, + "bbox": [ + 513, + 665, + 622, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the revealed point above as well as Eqs. (7) and (8), a simple framework for relative pose estimation, i.e., the EquiPose, could be straightforwardly built.", + "bbox": [ + 511, + 688, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the proposed EquiPose uses an existing model as its baseline, and outputs the final estimation on relative pose by calculating Eqs. (7) and (8), whose variable values are assigned by the baseline with or even without fine-tuning. Algorithm 1 shows the complete pipeline of EquiPose to obtain the relative pose.", + "bbox": [ + 511, + 734, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It has to be pointed out that the computational cost of $f^{*}$ in this implementation is twice that of $f$ , since it involves two times of forward propagation of $f$ : one for $f(I_{1},I_{2})$ , and one for $f(I_{2},I_{1})$ after permuting the input images $\\{I_1,I_2\\}$ . However, in practice, the computational", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1130", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c0b88ccccba0312ba9e2174af8dcb30830b696b3f49912ccd59909016c59abc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 88, + 759, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/53580ed6212c9472eb2b2d6d6b80de5a9714aa6f378b03440eea2dec914682eb.jpg", + "image_caption": [ + "(a)", + "(b)", + "Figure 3. The pipelines of existing methods and our method. (a) The pipeline of existing relative pose estimation networks, which consists of two stages: Stage (i) Single-Image Feature Encoding $(f_{\\mathrm{enc}})$ extracts the feature maps $\\{F_1,F_2\\}$ from $\\{I_1,I_2\\}$ . Stage (ii) Cross-Image Feature Decoding $(f_{\\mathrm{dec}})$ takes the extracted two feature maps $\\{F_1,F_2\\}$ as input, and it outputs the estimated relative pose. (b) The pipeline of EquiPose. The computational overhead from EquiPose is an additional forward propagation of $f_{\\mathrm{dec}}$ , which is slight." + ], + "image_footnote": [], + "bbox": [ + 171, + 284, + 823, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "overhead could be reduced further with the feature permutation strategy, which is elaborated in the following.", + "bbox": [ + 88, + 724, + 482, + 756 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "To begin with, it is noted that the general pipeline of existing relative pose estimation models [1, 6, 9, 12, 19, 36, 46] could be summarized in Fig. 3a. As is seen, the pipeline consists of the two stages, which we name as (i) Single-Image Feature Encoding (denoted as $f_{\\mathrm{enc}}$ ), and (ii) Cross-Image Feature Decoding (denoted as $f_{\\mathrm{dec}}$ ) respectively.", + "(i) Single-Image Feature Encoding $f_{\\mathrm{enc}}$ : This stage takes two images as inputs, and it outputs the corresponding feature maps $\\{F_1,F_2\\}$ , i.e., $f_{\\mathrm{enc}}(I_1,I_2)\\coloneqq \\{F_1,F_2\\}$ ." + ], + "bbox": [ + 88, + 758, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It contains a feature encoder, which could be any feature-extraction backbone in computer vision. It should be noted that the weights of the feature encoder for extracting the features of image $I_{1}$ and $I_{2}$ are shared (generally referred to as the Siamese architecture [5, 29]).", + "bbox": [ + 511, + 724, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(ii) Cross-Image Feature Decoding $f_{\\mathrm{dec}}$ : This stage takes the feature maps $\\{F_1, F_2\\}$ obtained from Stage (i) as input, and it outputs the estimated relative pose $P_{12}$ , i.e., $f_{\\mathrm{dec}}(F_1, F_2) := P_{12}$ . It contains a feature aggregation module and a pose decoder. The feature aggregation module maps $\\{F_1, F_2\\}$ into a single aggregated feature map (or vec", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1131", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c72533480fe92f9e20f61a204c677300acd824b939ba62134e84380c29a7e904.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison on the ScanNet dataset [11, 50]. RE: rotation errors in degrees. TE: translation errors in meters. +EquiPose: the models are implemented under EquiPose only during the inference stage. +EquiPose†: the models are further fine-tuned under EquiPose. -: the models only predict the relative camera rotation and thus could not be evaluated in translation. The best methods for each baseline are marked in bold." + ], + "table_body": "
MethodRot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m
ER[6]13.1528.2350.43---14.66-
+EquiPose15.2230.4051.51---14.11-
+EquiPose†16.0031.0152.98---12.53-
Reg6D[62]10.1622.9745.76---12.26-
+EquiPose11.7625.2847.94---11.76-
+EquiPose†16.1431.5854.20---10.81-
EightVit[46]23.6940.2561.9214.9032.9558.478.900.41
+EquiPose24.5841.3262.8215.4934.1160.038.400.35
+EquiPose†25.3042.5464.0516.3736.1463.077.970.31
MF6D[1]22.3637.1957.1014.3932.5757.349.810.44
+EquiPose25.4940.5359.9417.5135.7860.498.420.35
+EquiPose†27.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]28.9046.3666.5316.4635.3660.068.160.38
+EquiPose30.7048.0267.9017.8537.1761.807.970.32
+EquiPose†32.3249.7469.3719.6740.4065.957.610.28
", + "bbox": [ + 91, + 88, + 488, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tor), which is subsequently decoded by the pose decoder to obtain the estimated relative pose.", + "bbox": [ + 89, + 452, + 483, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the weights of the feature encoder in $f_{\\mathrm{enc}}$ are shared, we do not have to implement two complete forward propagation of $f$ . Instead, we could reuse the extracted feature maps $\\{F_1, F_2\\}$ at the first forward propagation, then permute the feature maps to $\\{F_2, F_1\\}$ , and implement one additional forward propagation of $f_{\\mathrm{dec}}$ .", + "bbox": [ + 89, + 489, + 483, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3b illustrates the EquiPose framework with the feature permutation strategy. As is seen, it first performs one forward propagation of $f_{\\mathrm{enc}}$ to extract feature maps $\\{F_1,F_2\\}$ . Then, the feature maps $\\{F_1,F_2\\}$ and its permuted counterpart $\\{F_2,F_1\\}$ are inputted to $f_{\\mathrm{dec}}$ to obtain $P_{12}$ and $P_{21}$ respectively. The final relative pose is obtained via $P_{12}$ and $P_{21}$ through Eq. (6). Compared to the original implementations of the baseline models, EquiPose only involves an additional forward propagation of $f_{\\mathrm{dec}}$ .", + "bbox": [ + 89, + 585, + 483, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Under the EquiPose framework, a pre-trained model could be directly adopted as the baseline without fine-tuning to boost its performance. Moreover, the model could also be further fine-tuned under EquiPose to achieve further improvement. We simply use an $\\mathcal{L}_1$ loss for fine-tuning:", + "bbox": [ + 89, + 728, + 483, + 806 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\| R ^ {*} - \\hat {R} \\| _ {1} + \\lambda \\| t ^ {*} - \\hat {t} \\| _ {1}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 828, + 482, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{R}$ and $\\hat{t}$ are the ground truth relative rotation and translation. $\\lambda$ is a preset constant factor.", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7de4c1b8034b8f0c163a634504aff009436fbf6c83d2c5a1d9fced443207081a.jpg", + "image_caption": [ + "Figure 4. Visualization of epipolar lines on ScanNet [11]. Five SIFT keypoints are sampled from image $I_{1}$ (column 1) and their epipolar lines are projected onto image $I_{2}$ , using the pose from: ground truth (column 2), GRelPose baseline [28] (column 3), EquiPose without (column 4) and with finetuning (column 5)." + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 893, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 367, + 638, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 392, + 802, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on both indoor and outdoor scenes. For indoor scenes, we use the ScanNet [11], the 7-Scene [49] and the 12-Scene [52] datasets, which are widely used for indoor pose estimation. For outdoor scenes, we use the recently proposed Mapfree dataset [1].", + "bbox": [ + 511, + 415, + 905, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For ScanNet, we use 2M image pairs from the training set for training, and use the test set which contains 1500 pairs for evaluation following [8, 48, 50, 55, 56]. For both 7-Scene and 12-Scene, we use 720 image pairs for evaluation using the model trained on ScanNet. For Mapfree, we follow the official setup in [1] for training and evaluation. More details are provided in the supplementary material.", + "bbox": [ + 511, + 491, + 905, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For ScanNet [11], 7-Scene [49] and 12-Scene [52], we calculate the area under the cumulative curve (AUC) as done in [31, 48, 50]. We report the AUC for rotation $(3^{\\circ} / 5^{\\circ} / 10^{\\circ})$ and translation $(0.1\\mathrm{m} / 0.2\\mathrm{m} / 0.5\\mathrm{m})$ respectively. Besides, we also report the average rotation and translation errors respectively. For the Mapfree dataset [1], we report the pose error and virtual correspondence reprojection error (VCRE) respectively following [1].", + "bbox": [ + 511, + 597, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 511, + 726, + 622, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It should be pointed out that the goal of this study is to introduce a general framework for existing end-to-end relative pose estimation models to achieve the PPE property with improved performances. Thus, we do not further evaluate and analyze other traditional matching-based methods, such as the classic SIFT [33] followed by the 5-point solver [41].", + "bbox": [ + 511, + 750, + 905, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we adopt the following five end-to-end models as baselines and evaluate the effectiveness of EquiPose in boosting their performances: (i) ExtremeRotation [6], (ii) Reg6D [6, 62], (iii) EightVit [46], (iv)", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1132", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7229146a4f9e431bbf9349cfa0c53288f322b33987069901486b563fc549daf7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method7-Scene12-Scene
Rot. AUC↑Trans. AUC↑RE↓TE↓Rot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m@3°@5°@10°@.1m@.2m@.5m
ER[6]11.1724.2846.70---9.82-14.6131.6057.15---6.94-
+EquiPose13.7226.6047.83---9.09-17.5734.3958.96---6.66-
+EquiPose†13.8227.1449.35---7.84-17.8335.3860.10---5.39-
Reg6D[62]10.0822.4445.92---7.78-13.1628.4255.27---5.67-
+EquiPose11.2023.7746.93---7.67-14.0530.0856.62---5.41-
+EquiPose†13.2627.3251.60---6.60-16.9933.5559.59---5.07-
EightVit[46]19.7735.0059.2513.9932.0561.114.920.2230.0950.3972.9538.0062.8084.032.890.08
+EquiPose19.8835.5159.8714.2132.9162.904.780.2031.1751.5373.5838.6763.7284.592.780.08
+EquiPose†20.8737.0161.4414.4733.7365.224.560.1932.7452.7274.4140.6365.3585.452.710.07
MF6D[1]18.7334.4657.6211.9030.3560.134.980.2330.4451.0773.1136.3761.5283.242.790.09
+EquiPose21.7837.4060.0315.4634.2863.714.710.1934.1754.2974.9840.8764.6785.022.560.08
+EquiPose†24.4440.8363.3217.1736.9066.444.290.1835.8055.3675.8243.0266.3685.782.450.07
GRelPose[28]26.7845.4767.9515.7034.2762.483.960.2135.6255.2075.6339.8763.8184.342.580.08
+EquiPose28.4446.8568.8715.8134.4163.483.880.2037.1356.6776.7142.2465.6585.212.410.07
+EquiPose†29.7648.2470.0815.8335.4364.913.660.1938.6358.1177.8644.1967.4386.042.240.07
", + "bbox": [ + 107, + 88, + 885, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MF6D [1], and (v) GRelPose [28]. Note that (i) ExtremeRotation and (ii) Reg6D only predict relative rotation, while the remaining ones could predict the 6 DoF relative camera pose (including both rotation and translation).", + "bbox": [ + 89, + 371, + 482, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 439, + 316, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Setup. For each of the five baseline models, including ER (ExtremeRotation) [6], Reg6D [62], EightVit [46], MF6D [1] and GRelPose [28], we compare the following variants: 1) the model in its original implementation, 2) the model implemented under EquiPose only during the inference stage (denoted with $^+$ EquiPose), 3) the model fine-tuned under EquiPose (denoted with $^+$ EquiPose $\\dagger$ ).", + "bbox": [ + 89, + 462, + 482, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Indoor Pose Evaluation. The evaluation results on ScanNet are reported in Table 1. As seen from this table, the five models under the EquiPose framework achieve consistently better performance than their corresponding baselines under the eight metrics. Moreover, these models after fine-tuning under EquiPose could achieve further enhancements in most cases by a significant margin. For example, an inference-only implementation of GRelPose under EquiPose leads to a relative decrease of $2.3\\%$ (1-7.97/8.16) and $15.8\\%$ (1-0.32/0.38) in rotation and translation errors respectively. Its fine-tuned implementation has a relative reduction of $6.7\\%$ and $26.3\\%$ in rotation and translation errors respectively. The fine-tuned MF6D could even achieve a relative reduction of $23.9\\%$ and $34.1\\%$ in rotation and translation errors respectively. These results validate the effectiveness of the proposed EquiPose framework.", + "bbox": [ + 89, + 569, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The cross-dataset evaluation results on the 7-Scene [49] and 12-Scene [52] datasets are reported in Tab. 2, where the models trained on ScanNet are used for evaluation. As seen from the table, the five models under EquiPose could also achieve consistent improvements than their corresponding baselines, and fine-tuning further boosts the performances.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/02db4de956f44e425d0ea593b847aa79e7b607b24143ac48bdd6174858961c91.jpg", + "table_caption": [ + "Table 2. Cross-dataset evaluation on the 7-Scene [49] and 12-Scene [52] datasets. All the models are trained on ScanNet." + ], + "table_footnote": [], + "table_body": "
MethodPose ErrorVCRE
Pr.↑Avg.Med.↓Med.↓Pr.@45↑Pr.@90↑
MFBin[1]9.1210cm/33.7°199.715.430.1
+EquiPose12.0185cm/32.5°196.918.231.6
MFBinTS[1]10.5183cm/23.3°166.418.135.5
+EquiPose12.9152cm/23.4°166.021.739.3
MFQuat[1]5.1203cm/27.3°160.714.333.6
+EquiPose6.3184cm/23.3°166.416.635.0
MFQuatTS[1]4.4185cm/26.3°154.814.234.7
+EquiPose6.6170cm/23.5°157.617.336.4
MF6D[1]6.1168cm/22.5°147.118.740.2
+EquiPose9.6153cm/18.6°148.222.140.9
+EquiPose†10.1148cm/18.4°145.023.042.1
EightVit[46]6.0204cm/26.6°155.217.835.8
+EquiPose6.4195cm/26.3°157.218.336.4
+EquiPose†6.8188cm/25.1°153.119.137.0
GRelPose[28]12.7168cm/22.4°137.725.843.3
+EquiPose13.0166cm/21.9°136.426.143.3
+EquiPose†13.2163cm/21.7°132.926.244.1
", + "bbox": [ + 514, + 368, + 908, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison on the Mapfree dataset [1]. As done in [1], we report Pose Error in terms of rotation error and translation error and VCRE in terms of median error and precision.", + "bbox": [ + 511, + 627, + 906, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These results demonstrate that EquiPose is also effective under the cross-dataset setup.", + "bbox": [ + 511, + 700, + 903, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the above quantitative results, we also provide visualization results from ScanNet. Considering GRelPose [28] performs relatively better than other baseline models in most cases, we adopt it as the baseline and compare it with EquiPose. The corresponding results are shown in Fig. 4. As seen from this figure, the proposed EquiPose is effective in improving the performance of relative pose estimation across different types of challenges, such as large viewpoint changes, and texture-less patterns. An inference-only implementation demonstrates obvious performance improvements, while subsequent fine-tuning could achieve", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1133", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7dba6c516837cf92c78fbd6587ad5b97834ba27f17f698b45e53c37a77cf15a9.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. EquiPose VS. Random Permutation. $^+\\mathrm{RP}$ denotes the model is trained with random permutation. $d_{R}$ and $d_{t}$ denote the rotation and translation difference computed via Eq. (4)." + ], + "table_body": "
MethoddRdtRot. AUCTrans. AUCRETE
@3°@5°@10°@.1m@.2m@.5m
EightVit[46]2.70.1223.6940.2561.9214.9032.9558.478.900.41
+RP1.20.0522.7239.9361.8314.6033.8261.238.230.33
+RP+EquiPose0022.9640.1162.0315.0234.3061.738.200.31
+EquiPose†0025.3042.5464.0516.3736.1463.077.970.31
MF6D[1]7.30.1822.3637.1957.1014.3932.5757.349.810.44
+RP5.40.1120.4236.6258.1114.5633.0160.368.170.32
+RP+EquiPose0024.6140.2660.8516.4335.6362.757.740.30
+EquiPose†0027.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]3.20.1228.9046.3666.5316.4635.3660.068.160.38
+RP2.10.0628.1145.6666.3616.9537.6563.647.740.30
+RP+EquiPose0029.0246.3566.8018.2038.8864.477.650.29
+EquiPose†0032.3249.7469.3719.6740.4065.957.610.28
", + "bbox": [ + 91, + 88, + 509, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "additional performance gains.", + "bbox": [ + 89, + 343, + 292, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Outdoor Pose Evaluation. The Mapfree dataset [1] was introduced for evaluating 6DoF relocalization, thus we only evaluate EightVit, MF6D, and GRelPose which could predict 6DoF poses. As a supplement, we further evaluate EquiPose in an inference-only manner on four additional baselines, including MFBin, MFBinTS, MFQuatTS which were released1 along with MF6D [1]. The results are reported in Tab. 3. As seen from this table, EquiPose could boost the performances of the above seven models in most cases. And the fine-tuned MF6D+EquiPose† achieves the lowest pose error among them. These results validate the effectiveness of our method under the outdoor setup.", + "bbox": [ + 88, + 359, + 482, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "EquiPose VS. Random Permutation. Intuitively, another way to leverage the PPE property for relative pose estimation is to permute the input order of image pairs randomly during training. To evaluate the effect of random permutation, we have further trained and tested EightVit [46], MF6D [1] and GRelPose [28] on the ScanNet dataset by using random permutation (RP). The corresponding results are reported in Tab. 4. In addition, we also report the average values of $d_{R}$ and $d_{t}$ computed via Eq. (4).", + "bbox": [ + 88, + 542, + 482, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As seen from this table, random permutation could improve model performances in average rotation and translation errors. However, the models trained under EquiPose perform significantly better than using random permutation. Besides, random permutation could not strictly enforce the PPE property as evidenced by both $d_{R}$ and $d_{t}$ being larger than zero, while EquiPose could achieve PPE strictly. To this end, EquiPose is not opposed to using random permutation, i.e., by enforcing the PPE property, it could also boost the performance of models trained using random permutation as observed in this table. The above results further validate the effectiveness of our method.", + "bbox": [ + 88, + 680, + 482, + 861 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7c9155e86c7c90e9ca778cd6ea41e46fcf0449e3bc15443a3b06e1b9e12971ee.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Comparison of parameter numbers, FLOPs, and inference times (ms) per image on an RTX 3090Ti GPU." + ], + "table_body": "
MethodParams.FLOPsTime (ms)
ER1.92 × 1079.86 × 1092.0
ER+EquiPose1.92 × 1071.09 × 10102.3
Reg6D2.33 × 1079.91 × 1092.6
Reg6D+EquiPose2.33 × 1079.97 × 1092.6
EightVit1.91 × 1076.88 × 1093.5
EightVit+EquiPose1.91 × 1079.64 × 1095.1
MF6D2.19 × 1074.57 × 10105.6
MF6D+EquiPose2.19 × 1074.60 × 10106.7
GRelPose1.37 × 1079.82 × 10911.8
GRelPose+EquiPose1.37 × 1071.13 × 101012.0
", + "bbox": [ + 524, + 88, + 893, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Computation Comparison and Limitation", + "text_level": 1, + "bbox": [ + 511, + 306, + 870, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we compare the parameter numbers, Floating Point Operations (FLOPs) and the inference times of the baselines and their EquiPose implementations. All the models are evaluated on an RTX 3090Ti GPU on ScanNet. The results are reported in Tab. 5. As is seen, compared with the baseline model, the EquiPose implementation does not introduce any additional parameters, since it uses the same backbone as the baseline, and thus does not incur extra storage costs. Moreover, compared with each baseline model, EquiPose results in a modest increment in computational costs generally, due to the feature permutation strategy.", + "bbox": [ + 511, + 329, + 903, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, the extra computational cost depends on the architecture of the baseline model: For models with lightweight feature aggregation modules and pose decoders, the increase in computation is slight. For models with heavy feature aggregation modules and pose decoders, the extra computational cost would be high, which is a limitation of EquiPose. Despite this, all the models under EquiPose maintain an inference time of less than 15 ms per image, which is sufficient for practical applications.", + "bbox": [ + 511, + 496, + 905, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 647, + 638, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work focuses on end-to-end relative pose estimation. Firstly, we introduce the overlooked pose permutation equivariance (PPE) property for relative pose estimation. Leveraging this property, we establish the foundational principle for designing a relative pose estimation framework that incorporates the PPE property. Building on this theoretical basis, we propose the general and flexible EquiPose framework for relative pose estimation. Experiments on four public datasets further demonstrate that the performances of pre-trained models could be improved even without re-training, and moreover, their performances could be further boosted after being fine-tuned under EquiPose.", + "bbox": [ + 511, + 672, + 905, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62376269).", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1Pretrained weights are downloaded from the website of Mapfree: https://github.com/nianticlabs/map-free-reloc.", + "bbox": [ + 89, + 875, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1134", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 2, 3, 5, 6, 7, 8, 1, 4", + "[2] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 3", + "[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2", + "[4] Georg Bokman, Johan Edstedt, Michael Felsberg, and Fredrik Kahl. Steerers: A framework for rotation equivariant keypoint descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4885-4895, 2024. 2", + "[5] Jane Bromley, Isabelle Guyon, Yann LeCun, and et al. Signature verification using a \"siamese\" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993. 3, 5", + "[6] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14566–14575, 2021. 1, 3, 5, 6, 7", + "[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 1", + "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 3, 6", + "[9] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 3, 5", + "[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1", + "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2, 6, 1, 3", + "[12] Shay Dekel and Yosi Keller. Estimating extreme 3d image rotation with transformer cross-attention. arXiv preprint arXiv:2303.02615, 2023. 1, 3, 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition workshops, pages 224-236, 2018. 1", + "[14] Qiulei Dong, Xiang Gao, Hainan Cui, and Zhanyi Hu. Robust camera translation estimation via rank enforcement. IEEE Transactions on Cybernetics, 52(2):862-872, 2020. 1", + "[15] Qulei Dong, Shuang Deng, and Yuzhen Liu. Cayley rotation averaging: Multiple camera averaging under the cayley framework. IEEE Transactions on Image Processing, 2024. 1", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3", + "[18] Johan Edstedt, Georg Bökman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. In 2024 International Conference on 3D Vision, pages 148–157. IEEE, 2024. 3", + "[19] Sovann En, Alexis Lechery, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision Workshops, 2018. 1, 3, 5", + "[20] Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and ransac's role. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8935-8943, 2022. 1, 3", + "[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2", + "[22] Richard Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In CVPR 2011, pages 3041-3048. IEEE, 2011. 3", + "[23] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hongdong Li. Rotation averaging. International Journal of Computer Vision, 103:267-305, 2013. 3, 1", + "[24] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(6):580-593, 1997. 1, 2, 3", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1", + "[26] Will Hutchcroft, Yuguang Li, Ivaylo Boyadzhiev, Zhiqiang Wan, Haiyan Wang, and Sing Bing Kang. Covispose: Covisibility pose transformer for wide-baseline relative pose estimation in 360 indoor panoramicas. In European Conference on Computer Vision, pages 615-633. Springer, 2022. 3, 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1135", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Nicolas Keriven and Gabriel Peyre. Universal invariant and equivariant graph neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[28] Fadi Khatib, Yuval Margalit, Meirav Galun, and Ronen Basri. Leveraging image matching toward end-to-end relative camera pose regression, 2024. 1, 2, 3, 6, 7, 8, 4, 5", + "[29] Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Deep Learning Workshop, 2015. 5", + "[30] Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International Conference on Machine Learning, pages 2747-2755. PMLR, 2018. 3", + "[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 3, 6", + "[32] Yuzhen Liu and Qiulei Dong. Descriptor distillation: A teacher-student-regularized framework for learning local descriptors. International Journal of Computer Vision, pages 1-19, 2024. 3", + "[33] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 6", + "[34] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3", + "[35] Clare Lyle, Mark van der Wilk, Marta Kwiatkowska, Yarin Gal, and Benjamin Bloem-Reddy. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020. 2", + "[36] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional neural networks. In International Conference on Advanced Concepts for Intelligent Vision Systems, pages 675-687. Springer, 2017. 1, 3, 5", + "[37] Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. In Advances in Neural Information Processing Systems, 2017. 3", + "[38] Maher Moakher. Means and averaging in the group of rotations. SIAM journal on matrix analysis and applications, 24 (1):1-16, 2002. 4", + "[39] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 2", + "[40] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mokolajczyk. Ood-pose: camera pose regression from out-of-distribution synthetic views. In International Conference on 3D Vision, pages 722-732. IEEE, 2022. 1", + "[41] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 6" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 1", + "[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 2", + "[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[45] Jerome Revaud, Yohann Cabon, Romain Brégier, JongMin Lee, and Philippe Weinzaepfel. Sacreg: Scene-agnostic coordinate regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 688–698, 2024. 1", + "[46] Chris Rockwell, Justin Johnson, and David F. Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision, 2022. 1, 3, 5, 6, 7, 8, 2, 4", + "[47] Chris Rockwell, Nilesh Kulkarni, Linyi Jin, Jeong Joon Park, Justin Johnson, and David F Fouhey. Far: Flexible accurate and robust 6 dof relative camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19854-19864, 2024. 3", + "[48] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 2, 3, 6, 1", + "[49] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 6, 7, 1", + "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 2, 3, 6, 1", + "[51] Yurun Tian, Axel Barroso-Laguna, Tony Ng, Vassileios Balntas, and Krystian Mikolajczyk. Hynet: Learning local descriptor with hybrid similarity measure and triplet loss. In Advances in Neural Information Processing Systems, pages 7401-7412, 2020. 3", + "[52] Julien Valentin, Angela Dai, Matthias Nießner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In International Conference on 3D Vision, pages 323-332. IEEE, 2016. 6, 7, 1", + "[53] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1", + "[54] Guangyu Wang, Jinzhi Zhang, Fan Wang, Ruqi Huang, and Lu Fang. Xscale-nvs: Cross-scale novel view synthesis with" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1136", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hash featurized manifold. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21029-21039, 2024. 1", + "[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Asian Conference on Computer Vision, 2022. 6", + "[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 6", + "[57] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 1", + "[58] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual u-net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018. 1, 2", + "[59] Xiaoming Zhao, Xingming Wu, Jinyu Miao, Weihai Chen, Peter CY Chen, and Zhengguo Li. Alike: Accurate and lightweight keypoint detection and descriptor extraction. IEEE Transactions on Multimedia, 2022. 3", + "[60] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4669-4678, 2021. 3", + "[61] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1851-1858, 2017. 1", + "[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 3, 6, 7, 1" + ], + "bbox": [ + 91, + 90, + 483, + 614 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1137", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_model.json b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b220eccad781694ac734796679e1d146f8a16143 --- /dev/null +++ b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.131, + 0.745, + 0.175 + ], + "angle": 0, + "content": "EquiPose: Exploiting Permutation Equivariance for Relative Camera Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.203, + 0.633, + 0.223 + ], + "angle": 0, + "content": "Yuzhen Liu1,2 Qialei Dong1,2,*" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.224, + 0.868, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Multimodal Artificial Intelligence Systems, Chinese Academy of Sciences, Institute of Automation \\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences liuyuzhen2022@ia.ac.cn, qldong@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.313, + 0.327, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.485, + 0.739 + ], + "angle": 0, + "content": "Relative camera pose estimation between two images is a fundamental task in 3D computer vision. Recently, many relative pose estimation networks have been explored for learning a mapping from two input images to their corresponding relative pose, however, the estimated relative poses by these methods do not have the intrinsic Pose Permutation Equivariance (PPE) property: the estimated relative pose from Image A to Image B should be the inverse of that from Image B to Image A. It means that permuting the input order of two images would cause these methods to obtain inconsistent relative poses. To address this problem, we firstly introduce the concept of PPE mapping, which indicates such a mapping that captures the intrinsic PPE property of relative poses. Then by enforcing the aforementioned PPE property, we propose a general framework for relative pose estimation, called EquiPose, which could easily accommodate various relative pose estimation networks in literature as its baseline models. We further theoretically prove that the proposed EquiPose framework could guarantee that its obtained mapping is a PPE mapping. Given a pre-trained baseline model, the proposed EquiPose framework could improve its performance even without fine-tuning, and could further boost its performance with fine-tuning. Experimental results on four public datasets demonstrate that EquiPose could significantly improve the performances of various state-of-the-art models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.779, + 0.224, + 0.795 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Estimating the relative camera pose between an image pair is one of the fundamental tasks in 3D computer vision [6, 46]. It has a broad range of applications, such as structure-from-motion (SfM) [14, 15, 42], visual localization [10, 40, 45], novel view synthesis [7, 53, 54], etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Existing methods for relative camera pose estimation in literature could be roughly divided into two categories: matching-based methods and end-to-end methods. The matching-based methods [13, 33] generally employ a two-stage computational strategy: Firstly, they establish the point correspondences between an input pair of images by utilizing local feature extraction and matching techniques [13, 33]. Then according to the established correspondences, they solve a two-view geometric problem via a typical 5-point algorithm [24] with robust fitting [21] for pursuing the relative pose. However, since the performances of the matching-based methods depend on the established point correspondences, their performance would significantly decrease or even fail in cases of weak/repeated textures and degenerated configurations as indicated in [20]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.665 + ], + "angle": 0, + "content": "Unlike the above matching-based methods, many recent methods [6, 12, 19, 28, 36, 46, 57, 61] have employed deep neural networks to directly learn a mapping from the input two images to their relative pose without explicitly establishing point correspondences. Here, it has to be pointed out that ideally, the estimated relative pose by these end-to-end methods should have the following intrinsic pose permutation equivariance (PPE) property between an image pair:" + }, + { + "type": "title", + "bbox": [ + 0.54, + 0.676, + 0.765, + 0.691 + ], + "angle": 0, + "content": "Pose Permutation Equivariance" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.703, + 0.882, + 0.749 + ], + "angle": 0, + "content": "Let \\(\\{I_1, I_2\\}\\) be a pair of images, then the relative pose \\(P_{12}\\) from \\(I_1\\) to \\(I_2\\) should be the inverse of the relative pose \\(P_{21}\\) from \\(I_2\\) to \\(I_1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.668, + 0.759, + 0.882, + 0.779 + ], + "angle": 0, + "content": "\\[\nP _ {2 1} = P _ {1 2} ^ {- 1}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.797, + 0.822, + 0.812 + ], + "angle": 0, + "content": "By writing \\(P_{12}\\) and \\(P_{21}\\) in the matrix form:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.825, + 0.907, + 0.86 + ], + "angle": 0, + "content": "\\[\nP _ {1 2} = \\left[ \\begin{array}{c c} R _ {1 2} & t _ {1 2} \\\\ 0 & 1 \\end{array} \\right], \\quad P _ {2 1} = \\left[ \\begin{array}{c c} R _ {2 1} & t _ {2 1} \\\\ 0 & 1 \\end{array} \\right], \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\( R_{12} \\) and \\( t_{12} \\) are the relative rotation and translation from \\( I_1 \\) to \\( I_2 \\), and \\( R_{21} \\) and \\( t_{21} \\) are the relative rotation and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.313, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.204, + 0.341, + 0.215 + ], + "angle": 0, + "content": "(a) MF6D" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.089, + 0.495, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.681, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.09, + 0.863, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.204, + 0.718, + 0.215 + ], + "angle": 0, + "content": "(b) GRelPose" + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.215, + 0.792, + 0.23 + ], + "angle": 0, + "content": "Figure 1. Distribution of the \\(d_{R}\\) and \\(d_{t}\\) obtained by existing models MF6D [1] and GRelPose [28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.479, + 0.272 + ], + "angle": 0, + "content": "translation from \\(I_{2}\\) to \\(I_{1}\\), Eq. (1) could be reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.278, + 0.483, + 0.297 + ], + "angle": 0, + "content": "\\[\nR _ {2 1} = R _ {1 2} ^ {\\mathrm {T}}, \\quad t _ {2 1} = - R _ {1 2} ^ {\\mathrm {T}} t _ {1 2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.547 + ], + "angle": 0, + "content": "However, the PPE property is generally overlooked when designing model architectures in existing end-to-end methods [1, 28]. We validate this by conducting the following experiment: we trained two typical end-to-end models (MF6D [1] and GRelPose [28]) on the training set of ScanNet [11]. Then, we adopt the test set of ScanNet [48, 50] which contains 1500 image pairs, and feed them with two different orders to the above models respectively, leading to totally 1500 pairs of the predicted relative poses by each model, where each pair consists of the estimated pose \\(P_{12}\\) from \\(I_{1}\\) to \\(I_{2}\\) and the estimated pose \\(P_{21}\\) from \\(I_{2}\\) to \\(I_{1}\\). Next, we compute the differences between the estimated relative pose \\(P_{21}\\) and \\(P_{12}^{-1}\\). Specifically, we compute the angular difference \\(d_{R}\\) between relative rotations \\(R_{21}\\) and \\(R_{12}^{\\mathrm{T}}\\), and the Euclidean distance \\(d_{t}\\) between relative translations \\(t_{21}\\) and \\(-R_{12}^{\\mathrm{T}} t_{12}\\) respectively via:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.554, + 0.483, + 0.578 + ], + "angle": 0, + "content": "\\[\nd _ {R} = \\arccos ((\\operatorname {t r} (R _ {2 1} R _ {1 2}) - 1) / 2), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.574, + 0.326, + 0.591 + ], + "angle": 0, + "content": "\\[\nd _ {t} = \\left\\| t _ {2 1} + R _ {1 2} ^ {\\mathrm {T}} t _ {1 2} \\right\\| _ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.614 + ], + "angle": 0, + "content": "where \\(\\mathrm{tr}(\\cdot)\\) denotes matrix trace, \\(\\| \\cdot \\| _2\\) denotes \\(L_{2}\\) norm." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.84 + ], + "angle": 0, + "content": "The distributions of the rotation difference \\( d_{R} \\) and translation difference \\( d_{t} \\) from the above two models are shown in Fig. 1. As seen from this figure, the pose differences (both rotation and translation) estimated by these models are generally larger than zero (particularly some of them are not trivial), demonstrating that the above models do not satisfy the PPE property. Such an issue may pose risks to downstream applications that require a stable and consistent estimation of relative poses, such as robotics and visual navigation. Moreover, it is also noted that some other visual tasks involve their specific equivariances, different from the PPE property in the relative pose estimation task, and once an existing model among these tasks is equipped with a task-specific equivariance, its performance could be further boosted [4, 27, 35, 39, 43]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Motivated by the above issues, we investigate whether it is possible to explore such a relative pose estimation framework that could not only bestow the PPE property on various relative pose estimation networks but also improve their" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.905, + 0.391 + ], + "angle": 0, + "content": "performances further. Specifically, we first introduce the concept of PPE mapping, which refers to a mapping that captures the PPE property in relative poses. Then, we establish the foundational principle for equivariant relative pose estimation. Accordingly, we propose the EquiPose framework which holds the PPE property theoretically. The EquiPose could improve the performance of a pre-trained model even without fine-tuning, and fine-tuning could enhance their performances further." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.393, + 0.844, + 0.406 + ], + "angle": 0, + "content": "To summarize, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.408, + 0.905, + 0.453 + ], + "angle": 0, + "content": "- Equivariance: The simple but intrinsic pose permutation equivariance (PPE) property is revealed for relative pose estimation, and the PPE mapping is defined." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.453, + 0.906, + 0.574 + ], + "angle": 0, + "content": "- Generality: By imposing the PPE property, the framework EquiPose for equivariant relative pose estimation is proposed. An arbitrary end-to-end relative pose estimation model could be seamlessly embedded into this framework as a baseline, and accordingly obtains the PPE property as analyzed in theory in Sec. 3. The EquiPose framework and the theoretical analyses might give new insights into model design strategies for relative pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.574, + 0.906, + 0.68 + ], + "angle": 0, + "content": "- Flexibility: Thanks to the imposed PPE property, the performances of the derived methods from the EquiPose framework by utilizing many off-the-shelf baseline models, would be boosted even without fine-tuning by adding several lines of code. Moreover, once they are fine-tuned under EquiPose, their performances would be further improved, as demonstrated by the results in Sec. 4." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.408, + 0.906, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.657, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.764 + ], + "angle": 0, + "content": "In this section, we review some typical matching-based methods and end-to-end methods for relative pose estimation respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Matching-Based Relative Pose Estimation. The matching-based relative pose estimation methods typically involve the following two steps: establishing point correspondence via image matching [3, 33] and solving the two-view geometry with RANSAC [21, 24, 41]. Early approaches relied on handcrafted keypoint detectors and descriptors to establish point correspondences, such as SIFT [33] and SURF [3]. In recent years, numerous methods emerged that replace the handcrafted detectors or descrip" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1128" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "tors with their learned counterparts [2, 17, 18, 32, 34, 37, 44, 51, 59]. Additionally, a variety of learning-based image-matching methods directly established point correspondences from keypoints or raw images [8, 31, 48, 50, 60]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.484, + 0.258 + ], + "angle": 0, + "content": "However, since these methods depend on point correspondences, they may fail in cases of weak textures, repeated textures, large viewpoint changes, and degenerated configurations [20]. Additionally, these methods could be time-consuming due to image matching and RANSAC loops. Moreover, they could only recover the relative pose up to an unknown scale factor [24]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.439 + ], + "angle": 0, + "content": "End-to-End Relative Pose Estimation. Recently, unlike the above strategies, many works have adopted deep neural networks to regress relative poses in an end-to-end manner [1, 6, 9, 12, 19, 28, 36]. These methods do not explicitly implement image matching or solve the two-view geometry problem. Instead, the networks directly learn the underlying mapping from the input image pair to the corresponding relative pose. Compared with matching-based methods, they are shown to be more robust in challenging scenes [6, 9, 19] with a significantly faster inference speed. Moreover, these end-to-end methods could be further incorporated with traditional methods for improving robustness [47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.68 + ], + "angle": 0, + "content": "These end-to-end methods generally adopt Siamese architectures [5] to extract feature, which are aggregated and decoded into the relative pose subsequently [6, 12, 26, 28, 46]. To facilitate stable training, early approaches [19, 36] parameterized the rotation as quaternions. Different from them, Zhou et al. [62] suggested to represent 3D rotations in a 6D space, which benefits the training of neural networks. Cai et al. [6] and Dekel et al. [12] showed that the network could learn from implicit cues, allowing the network to estimate the relative pose even when two images have no overlap. Rockwell et al. [46] introduced the essential matrix module to emulate the 8-point algorithm [24]. Arnold et al. [1] adopted a differentiable warping module for feature aggregation. Khatib et al. [28] embedded a semidense image matcher [50] for feature extraction to pursue a better generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.802 + ], + "angle": 0, + "content": "However, the existing end-to-end methods generally do not satisfy the PPE property, leading to inconsistent estimation when permuting the input order of an image pair as shown in Sec. 1. Unlike them, this work does not aim to propose a new model for relative pose estimation, but pursue a general framework that could bestow the PPE property on various relative pose estimation models, and improve their performances further." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.226, + 0.833 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this section, we first introduce the key definitions and propositions related to the pose permutation equivariance (PPE) property of relative poses. Then, we describe the proposed EquiPose framework in detail." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.88, + 0.108 + ], + "angle": 0, + "content": "3.1. PPE Mapping and Preliminary Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Prior to the introduction of the PPE mapping, 'equivariant mapping' in mathematics is defined as [30]:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.158, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Definition 1 (Equivariant mapping). Let \\(\\mathcal{G}\\) be a group that acts on spaces \\(\\mathcal{X}\\) and \\(\\mathcal{Y}\\), and \\(\\mathbb{T}\\) and \\(\\mathbb{T}'\\) be the induced actions of \\(\\mathcal{G}\\) on \\(\\mathcal{X}\\) and \\(\\mathcal{Y}\\). A mapping \\(f: \\mathcal{X} \\to \\mathcal{Y}\\) is equivariant with the action in \\(\\mathcal{G}\\) if \\(f(\\mathbb{T}_g(x)) = \\mathbb{T}_g'(f(x)), \\forall x \\in \\mathcal{X}, g \\in \\mathcal{G}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.907, + 0.277 + ], + "angle": 0, + "content": "According to Definition 1 and Eq. (1), a PPE mapping with the PPE property is defined as:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Definition 2 (Pose permutation equivariant mapping). A mapping \\( f: \\mathbb{R}^{H \\times W} \\times \\mathbb{R}^{H \\times W} \\to \\mathrm{SE}(3) \\) (the special Euclidean group) is pose permutation equivariant, if \\( f(I_2, I_1) = f(I_1, I_2)^{-1}, \\forall I_1, I_2 \\in \\mathbb{R}^{H \\times W} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.362, + 0.906, + 0.438 + ], + "angle": 0, + "content": "Definition 2 could be considered as an instance of 'equivariant mapping' from Definition 1 on relative camera poses, where the input variable \\( x \\) is an image pair \\( \\{I_1, I_2\\} \\), the action \\( \\mathbb{T}_g \\) on \\( \\{I_1, I_2\\} \\) is an order permutation operator, and the action \\( \\mathbb{T}_q' \\) is the inverse of a matrix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.906, + 0.485 + ], + "angle": 0, + "content": "It is noted that Definition 2 is applicable to neural networks, since any neural networks essentially aim to learn a mapping from their input to the output." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.485, + 0.906, + 0.575 + ], + "angle": 0, + "content": "In addition to the above two definitions, we introduce the following definition on the geodesic mean of rotations [22, 23], as well as two theorems, which would be used to support the fundamental principle for designing a relative pose estimation framework with the PPE property in the next subsection:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.588, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Definition 3 (Geodesic mean of rotations). Given a set of rotations \\(\\{R_i\\}_{i=1}^n\\), their geodesic mean \\(S\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.632, + 0.907, + 0.672 + ], + "angle": 0, + "content": "\\[\nS = \\mathfrak {G} \\left(R _ {1}, \\dots , R _ {n}\\right) = \\underset {S \\in \\mathrm {S O} (3)} {\\arg \\min } \\sum_ {i = 1} ^ {n} \\| \\log \\left(R _ {i} ^ {\\mathrm {T}} S\\right) \\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.684, + 0.906, + 0.715 + ], + "angle": 0, + "content": "where \\(\\log (\\cdot)\\) is the mapping from an element in the Lie group to its corresponding Lie algebra." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.727, + 0.855, + 0.741 + ], + "angle": 0, + "content": "From Definition 3, we have the two theorems as:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.755, + 0.907, + 0.801 + ], + "angle": 0, + "content": "Theorem 1. Given a set of rotations \\(\\{R_i\\}_{i=1}^n\\), it always holds that: \\(\\mathfrak{G}(R_{\\sigma(1)}, \\dots, R_{\\sigma(n)}) = \\mathfrak{G}(R_1, \\dots, R_n)\\) for any permutation \\(\\sigma\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.812, + 0.906, + 0.86 + ], + "angle": 0, + "content": "Theorem 2. If \\( S \\in \\mathrm{SO}(3) \\) is the geodesic mean of the set of rotations \\( \\{R_i\\}_{i=1}^n \\), then \\( S^{\\mathrm{T}} \\) is the geodesic mean of the set of rotations \\( \\{R_i^{\\mathrm{T}}\\}_{i=1}^n \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Due to the space limitation, the proofs of the above two theorems are given in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1129" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.473, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.253 + ], + "angle": 0, + "content": "Figure 2. Illustration of the existing methods \\( f \\) and our method \\( f^{*} \\) on the SE(3) manifold. Left: The existing method \\( f \\) does not capture the PPE property in the relative pose. Right: Within EquiPose \\( f^{*} \\), the desired PPE property is satisfied (Theorem 3), and the results could be more reliable (Sec. 4)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.307 + ], + "angle": 0, + "content": "3.2. Fundamental Principle of Pose Estimation Framework with PPE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.405 + ], + "angle": 0, + "content": "According to the aforementioned PPE mapping and related preliminary knowledge, we elaborate the fundamental principle for designing a pose estimation framework with PPE, which could use an arbitrary end-to-end relative pose estimation model as its baseline and derive the corresponding upgraded model with the PPE property as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.484, + 0.452 + ], + "angle": 0, + "content": "Given an arbitrary relative pose estimation network corresponding to a mapping \\( f \\), there exists the following mapping \\( f^{*} \\), which is a PPE mapping:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Theorem 3. Given an arbitrary mapping \\( f: \\mathbb{R}^{H \\times W} \\times \\mathbb{R}^{H \\times W} \\to \\mathrm{SE}(3) \\), where \\( f(I_1, I_2) := (R_{12}, t_{12}) \\) and \\( f(I_2, I_1) := (R_{21}, t_{21}) \\), the mapping \\( f^* \\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.51, + 0.482, + 0.528 + ], + "angle": 0, + "content": "\\[\nf ^ {*} \\left(I _ {1}, I _ {2}\\right) := \\left(R _ {1 2} ^ {*}, t _ {1 2} ^ {*}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.535, + 0.138, + 0.547 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.556, + 0.483, + 0.574 + ], + "angle": 0, + "content": "\\[\nR _ {1 2} ^ {*} = \\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.576, + 0.483, + 0.593 + ], + "angle": 0, + "content": "\\[\nt _ {1 2} ^ {*} = \\left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\\right) / 2, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.216, + 0.616 + ], + "angle": 0, + "content": "is a PPE mapping." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.622, + 0.368, + 0.637 + ], + "angle": 0, + "content": "Proof. According to Theorem 1, we have" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.65, + 0.483, + 0.669 + ], + "angle": 0, + "content": "\\[\nR _ {1 2} ^ {*} = \\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right) = \\mathfrak {G} \\left(R _ {2 1} ^ {\\mathrm {T}}, R _ {1 2}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.673, + 0.381, + 0.687 + ], + "angle": 0, + "content": "Then, according to Theorem 2, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.693, + 0.483, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathfrak {G} \\left(R _ {2 1} ^ {\\mathrm {T}}, R _ {1 2}\\right) = \\mathfrak {G} \\left(R _ {2 1}, R _ {1 2} ^ {\\mathrm {T}}\\right) ^ {\\mathrm {T}} = R _ {2 1} ^ {* \\mathrm {T}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.72, + 0.377, + 0.734 + ], + "angle": 0, + "content": "According to Eqs. (9) and (10), it holds:" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.74, + 0.483, + 0.759 + ], + "angle": 0, + "content": "\\[\nR _ {1 2} ^ {*} = R _ {2 1} ^ {* \\mathrm {T}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.449, + 0.781 + ], + "angle": 0, + "content": "i.e., \\( f^{*} \\) is permutation equivariant in terms of rotation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.781, + 0.433, + 0.796 + ], + "angle": 0, + "content": "Now, according to (6), (7), (8) and (11), it holds:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.803, + 0.482, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f ^ {*} (I _ {1}, I _ {2}) ^ {- 1} = \\left(R _ {1 2} ^ {* \\mathrm {T}}, - R _ {1 2} ^ {* \\mathrm {T}} t _ {1 2} ^ {*}\\right) \\\\ = \\left(R _ {1 2} ^ {* \\mathrm {T}}, - R _ {1 2} ^ {* \\mathrm {T}} \\left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\\right) / 2\\right) \\tag {12} \\\\ = \\left(R _ {2 1} ^ {*}, \\left(t _ {2 1} - R _ {2 1} ^ {*} t _ {1 2}\\right) / 2\\right) \\\\ = f ^ {*} \\left(I _ {2}, I _ {1}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.274, + 0.901 + ], + "angle": 0, + "content": "which completes the proof." + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.887, + 0.482, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.091, + 0.778, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 The EquiPose Framework" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.111, + 0.822, + 0.142 + ], + "angle": 0, + "content": "Input: A relative pose estimation network \\(f\\); A pair of images \\(\\{I_1, I_2\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.142, + 0.825, + 0.157 + ], + "angle": 0, + "content": "Output: Equivariant pose estimation \\( R_{12}^{*}, t_{12}^{*} \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.157, + 0.682, + 0.171 + ], + "angle": 0, + "content": "1: \\(R_{12}, t_{12} \\gets f(I_1, I_2)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.172, + 0.682, + 0.187 + ], + "angle": 0, + "content": "2: \\(R_{21}, t_{21} \\gets f(I_2, I_1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.186, + 0.772, + 0.203 + ], + "angle": 0, + "content": "3: \\( R_{12}^{*} \\gets R_{12}(\\exp \\left(\\frac{1}{2}\\log (R_{12}^{\\mathrm{T}}R_{21}^{\\mathrm{T}})\\right)) \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.203, + 0.706, + 0.217 + ], + "angle": 0, + "content": "4: \\(t_{12}^{*}\\gets (t_{12} - R_{12}^{*}t_{21}) / 2\\)" + }, + { + "type": "list", + "bbox": [ + 0.526, + 0.157, + 0.772, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.905, + 0.384 + ], + "angle": 0, + "content": "For ease of understanding, we visualize the difference between \\( f \\) and \\( f^{*} \\) in Fig. 2. As seen from this figure, given two input images \\( \\{I_1, I_2\\} \\), the existing relative pose estimation model (considered as a mapping) \\( f \\) does not satisfy the PPE property. Once it is embedded into EquiPose, the upgrade version \\( f^{*} \\) captures the PPE property strictly. Moreover, its estimation could be more reliable than the baseline model \\( f \\), as demonstrated in Sec. 4 (also in Sec. S8.3 of the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.906, + 0.43 + ], + "angle": 0, + "content": "It is noted in Eq. (7) that the construction of \\( f^{*} \\) involves computing the geodesic mean of two rotations, which could be computed explicitly via [38]:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.44, + 0.905, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right) = R _ {1 2} \\left(R _ {1 2} ^ {\\mathrm {T}} R _ {2 1} ^ {\\mathrm {T}}\\right) ^ {1 / 2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.515 + ], + "angle": 0, + "content": "Considering that we work on the 3D rotation group, the square root of a rotation \\( R \\in \\mathrm{SO}(3) \\) could be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.515, + 0.905, + 0.543 + ], + "angle": 0, + "content": "\\[\nR ^ {1 / 2} = \\exp \\left(\\frac {1}{2} \\log (R)\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.905, + 0.58 + ], + "angle": 0, + "content": "where \\(\\exp (\\cdot)\\) is the mapping from the Lie algebra to the Lie group." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.656 + ], + "angle": 0, + "content": "The above analysis reveals that given an arbitrary relative pose estimation network, once its upgrade version \\( f^{*} \\) is constructed by (6), the mapping is a pose permutation equivariant mapping, this is to say, the upgrade version has the PPE property." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.666, + 0.623, + 0.682 + ], + "angle": 0, + "content": "3.3. EquiPose" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.735 + ], + "angle": 0, + "content": "According to the revealed point above as well as Eqs. (7) and (8), a simple framework for relative pose estimation, i.e., the EquiPose, could be straightforwardly built." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Specifically, the proposed EquiPose uses an existing model as its baseline, and outputs the final estimation on relative pose by calculating Eqs. (7) and (8), whose variable values are assigned by the baseline with or even without fine-tuning. Algorithm 1 shows the complete pipeline of EquiPose to obtain the relative pose." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "It has to be pointed out that the computational cost of \\( f^{*} \\) in this implementation is twice that of \\( f \\), since it involves two times of forward propagation of \\( f \\): one for \\( f(I_{1},I_{2}) \\), and one for \\( f(I_{2},I_{1}) \\) after permuting the input images \\( \\{I_1,I_2\\} \\). However, in practice, the computational" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1130" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.172, + 0.089, + 0.761, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.271, + 0.508, + 0.283 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.285, + 0.825, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.626, + 0.509, + 0.638 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.643, + 0.908, + 0.701 + ], + "angle": 0, + "content": "Figure 3. The pipelines of existing methods and our method. (a) The pipeline of existing relative pose estimation networks, which consists of two stages: Stage (i) Single-Image Feature Encoding \\((f_{\\mathrm{enc}})\\) extracts the feature maps \\(\\{F_1,F_2\\}\\) from \\(\\{I_1,I_2\\}\\). Stage (ii) Cross-Image Feature Decoding \\((f_{\\mathrm{dec}})\\) takes the extracted two feature maps \\(\\{F_1,F_2\\}\\) as input, and it outputs the estimated relative pose. (b) The pipeline of EquiPose. The computational overhead from EquiPose is an additional forward propagation of \\(f_{\\mathrm{dec}}\\), which is slight." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.726, + 0.483, + 0.757 + ], + "angle": 0, + "content": "overhead could be reduced further with the feature permutation strategy, which is elaborated in the following." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.76, + 0.483, + 0.852 + ], + "angle": 0, + "content": "To begin with, it is noted that the general pipeline of existing relative pose estimation models [1, 6, 9, 12, 19, 36, 46] could be summarized in Fig. 3a. As is seen, the pipeline consists of the two stages, which we name as (i) Single-Image Feature Encoding (denoted as \\( f_{\\mathrm{enc}} \\)), and (ii) Cross-Image Feature Decoding (denoted as \\( f_{\\mathrm{dec}} \\)) respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "(i) Single-Image Feature Encoding \\( f_{\\mathrm{enc}} \\): This stage takes two images as inputs, and it outputs the corresponding feature maps \\( \\{F_1,F_2\\} \\), i.e., \\( f_{\\mathrm{enc}}(I_1,I_2)\\coloneqq \\{F_1,F_2\\} \\)." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.76, + 0.483, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.907, + 0.802 + ], + "angle": 0, + "content": "It contains a feature encoder, which could be any feature-extraction backbone in computer vision. It should be noted that the weights of the feature encoder for extracting the features of image \\( I_{1} \\) and \\( I_{2} \\) are shared (generally referred to as the Siamese architecture [5, 29])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "(ii) Cross-Image Feature Decoding \\( f_{\\mathrm{dec}} \\): This stage takes the feature maps \\( \\{F_1, F_2\\} \\) obtained from Stage (i) as input, and it outputs the estimated relative pose \\( P_{12} \\), i.e., \\( f_{\\mathrm{dec}}(F_1, F_2) := P_{12} \\). It contains a feature aggregation module and a pose decoder. The feature aggregation module maps \\( \\{F_1, F_2\\} \\) into a single aggregated feature map (or vec" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "1131" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.49, + 0.314 + ], + "angle": 0, + "content": "
MethodRot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m
ER[6]13.1528.2350.43---14.66-
+EquiPose15.2230.4051.51---14.11-
+EquiPose†16.0031.0152.98---12.53-
Reg6D[62]10.1622.9745.76---12.26-
+EquiPose11.7625.2847.94---11.76-
+EquiPose†16.1431.5854.20---10.81-
EightVit[46]23.6940.2561.9214.9032.9558.478.900.41
+EquiPose24.5841.3262.8215.4934.1160.038.400.35
+EquiPose†25.3042.5464.0516.3736.1463.077.970.31
MF6D[1]22.3637.1957.1014.3932.5757.349.810.44
+EquiPose25.4940.5359.9417.5135.7860.498.420.35
+EquiPose†27.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]28.9046.3666.5316.4635.3660.068.160.38
+EquiPose30.7048.0267.9017.8537.1761.807.970.32
+EquiPose†32.3249.7469.3719.6740.4065.957.610.28
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.414 + ], + "angle": 0, + "content": "Table 1. Comparison on the ScanNet dataset [11, 50]. RE: rotation errors in degrees. TE: translation errors in meters. +EquiPose: the models are implemented under EquiPose only during the inference stage. +EquiPose†: the models are further fine-tuned under EquiPose. -: the models only predict the relative camera rotation and thus could not be evaluated in translation. The best methods for each baseline are marked in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.485 + ], + "angle": 0, + "content": "tor), which is subsequently decoded by the pose decoder to obtain the estimated relative pose." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.583 + ], + "angle": 0, + "content": "Since the weights of the feature encoder in \\( f_{\\mathrm{enc}} \\) are shared, we do not have to implement two complete forward propagation of \\( f \\). Instead, we could reuse the extracted feature maps \\( \\{F_1, F_2\\} \\) at the first forward propagation, then permute the feature maps to \\( \\{F_2, F_1\\} \\), and implement one additional forward propagation of \\( f_{\\mathrm{dec}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.725 + ], + "angle": 0, + "content": "Fig. 3b illustrates the EquiPose framework with the feature permutation strategy. As is seen, it first performs one forward propagation of \\( f_{\\mathrm{enc}} \\) to extract feature maps \\( \\{F_1,F_2\\} \\). Then, the feature maps \\( \\{F_1,F_2\\} \\) and its permuted counterpart \\( \\{F_2,F_1\\} \\) are inputted to \\( f_{\\mathrm{dec}} \\) to obtain \\( P_{12} \\) and \\( P_{21} \\) respectively. The final relative pose is obtained via \\( P_{12} \\) and \\( P_{21} \\) through Eq. (6). Compared to the original implementations of the baseline models, EquiPose only involves an additional forward propagation of \\( f_{\\mathrm{dec}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Under the EquiPose framework, a pre-trained model could be directly adopted as the baseline without fine-tuning to boost its performance. Moreover, the model could also be further fine-tuned under EquiPose to achieve further improvement. We simply use an \\(\\mathcal{L}_1\\) loss for fine-tuning:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.829, + 0.483, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\| R ^ {*} - \\hat {R} \\| _ {1} + \\lambda \\| t ^ {*} - \\hat {t} \\| _ {1}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{R}\\) and \\(\\hat{t}\\) are the ground truth relative rotation and translation. \\(\\lambda\\) is a preset constant factor." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.089, + 0.895, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.274, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 4. Visualization of epipolar lines on ScanNet [11]. Five SIFT keypoints are sampled from image \\( I_{1} \\) (column 1) and their epipolar lines are projected onto image \\( I_{2} \\), using the pose from: ground truth (column 2), GRelPose baseline [28] (column 3), EquiPose without (column 4) and with finetuning (column 5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.368, + 0.64, + 0.387 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.803, + 0.409 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on both indoor and outdoor scenes. For indoor scenes, we use the ScanNet [11], the 7-Scene [49] and the 12-Scene [52] datasets, which are widely used for indoor pose estimation. For outdoor scenes, we use the recently proposed Mapfree dataset [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.598 + ], + "angle": 0, + "content": "For ScanNet, we use 2M image pairs from the training set for training, and use the test set which contains 1500 pairs for evaluation following [8, 48, 50, 55, 56]. For both 7-Scene and 12-Scene, we use 720 image pairs for evaluation using the model trained on ScanNet. For Mapfree, we follow the official setup in [1] for training and evaluation. More details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Evaluation Metrics. For ScanNet [11], 7-Scene [49] and 12-Scene [52], we calculate the area under the cumulative curve (AUC) as done in [31, 48, 50]. We report the AUC for rotation \\((3^{\\circ} / 5^{\\circ} / 10^{\\circ})\\) and translation \\((0.1\\mathrm{m} / 0.2\\mathrm{m} / 0.5\\mathrm{m})\\) respectively. Besides, we also report the average rotation and translation errors respectively. For the Mapfree dataset [1], we report the pose error and virtual correspondence reprojection error (VCRE) respectively following [1]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.623, + 0.743 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.841 + ], + "angle": 0, + "content": "It should be pointed out that the goal of this study is to introduce a general framework for existing end-to-end relative pose estimation models to achieve the PPE property with improved performances. Thus, we do not further evaluate and analyze other traditional matching-based methods, such as the classic SIFT [33] followed by the 5-point solver [41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Specifically, we adopt the following five end-to-end models as baselines and evaluate the effectiveness of EquiPose in boosting their performances: (i) ExtremeRotation [6], (ii) Reg6D [6, 62], (iii) EightVit [46], (iv)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1132" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.887, + 0.33 + ], + "angle": 0, + "content": "
Method7-Scene12-Scene
Rot. AUC↑Trans. AUC↑RE↓TE↓Rot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m@3°@5°@10°@.1m@.2m@.5m
ER[6]11.1724.2846.70---9.82-14.6131.6057.15---6.94-
+EquiPose13.7226.6047.83---9.09-17.5734.3958.96---6.66-
+EquiPose†13.8227.1449.35---7.84-17.8335.3860.10---5.39-
Reg6D[62]10.0822.4445.92---7.78-13.1628.4255.27---5.67-
+EquiPose11.2023.7746.93---7.67-14.0530.0856.62---5.41-
+EquiPose†13.2627.3251.60---6.60-16.9933.5559.59---5.07-
EightVit[46]19.7735.0059.2513.9932.0561.114.920.2230.0950.3972.9538.0062.8084.032.890.08
+EquiPose19.8835.5159.8714.2132.9162.904.780.2031.1751.5373.5838.6763.7284.592.780.08
+EquiPose†20.8737.0161.4414.4733.7365.224.560.1932.7452.7274.4140.6365.3585.452.710.07
MF6D[1]18.7334.4657.6211.9030.3560.134.980.2330.4451.0773.1136.3761.5283.242.790.09
+EquiPose21.7837.4060.0315.4634.2863.714.710.1934.1754.2974.9840.8764.6785.022.560.08
+EquiPose†24.4440.8363.3217.1736.9066.444.290.1835.8055.3675.8243.0266.3685.782.450.07
GRelPose[28]26.7845.4767.9515.7034.2762.483.960.2135.6255.2075.6339.8763.8184.342.580.08
+EquiPose28.4446.8568.8715.8134.4163.483.880.2037.1356.6776.7142.2465.6585.212.410.07
+EquiPose†29.7648.2470.0815.8335.4364.913.660.1938.6358.1177.8644.1967.4386.042.240.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.332, + 0.867, + 0.345 + ], + "angle": 0, + "content": "Table 2. Cross-dataset evaluation on the 7-Scene [49] and 12-Scene [52] datasets. All the models are trained on ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.435 + ], + "angle": 0, + "content": "MF6D [1], and (v) GRelPose [28]. Note that (i) ExtremeRotation and (ii) Reg6D only predict relative rotation, while the remaining ones could predict the 6 DoF relative camera pose (including both rotation and translation)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.317, + 0.456 + ], + "angle": 0, + "content": "4.3. Comparative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Evaluation Setup. For each of the five baseline models, including ER (ExtremeRotation) [6], Reg6D [62], EightVit [46], MF6D [1] and GRelPose [28], we compare the following variants: 1) the model in its original implementation, 2) the model implemented under EquiPose only during the inference stage (denoted with \\(^+\\) EquiPose), 3) the model fine-tuned under EquiPose (denoted with \\(^+\\) EquiPose\\(\\dagger\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Indoor Pose Evaluation. The evaluation results on ScanNet are reported in Table 1. As seen from this table, the five models under the EquiPose framework achieve consistently better performance than their corresponding baselines under the eight metrics. Moreover, these models after fine-tuning under EquiPose could achieve further enhancements in most cases by a significant margin. For example, an inference-only implementation of GRelPose under EquiPose leads to a relative decrease of \\(2.3\\%\\) (1-7.97/8.16) and \\(15.8\\%\\) (1-0.32/0.38) in rotation and translation errors respectively. Its fine-tuned implementation has a relative reduction of \\(6.7\\%\\) and \\(26.3\\%\\) in rotation and translation errors respectively. The fine-tuned MF6D could even achieve a relative reduction of \\(23.9\\%\\) and \\(34.1\\%\\) in rotation and translation errors respectively. These results validate the effectiveness of the proposed EquiPose framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The cross-dataset evaluation results on the 7-Scene [49] and 12-Scene [52] datasets are reported in Tab. 2, where the models trained on ScanNet are used for evaluation. As seen from the table, the five models under EquiPose could also achieve consistent improvements than their corresponding baselines, and fine-tuning further boosts the performances." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.369, + 0.909, + 0.625 + ], + "angle": 0, + "content": "
MethodPose ErrorVCRE
Pr.↑Avg.Med.↓Med.↓Pr.@45↑Pr.@90↑
MFBin[1]9.1210cm/33.7°199.715.430.1
+EquiPose12.0185cm/32.5°196.918.231.6
MFBinTS[1]10.5183cm/23.3°166.418.135.5
+EquiPose12.9152cm/23.4°166.021.739.3
MFQuat[1]5.1203cm/27.3°160.714.333.6
+EquiPose6.3184cm/23.3°166.416.635.0
MFQuatTS[1]4.4185cm/26.3°154.814.234.7
+EquiPose6.6170cm/23.5°157.617.336.4
MF6D[1]6.1168cm/22.5°147.118.740.2
+EquiPose9.6153cm/18.6°148.222.140.9
+EquiPose†10.1148cm/18.4°145.023.042.1
EightVit[46]6.0204cm/26.6°155.217.835.8
+EquiPose6.4195cm/26.3°157.218.336.4
+EquiPose†6.8188cm/25.1°153.119.137.0
GRelPose[28]12.7168cm/22.4°137.725.843.3
+EquiPose13.0166cm/21.9°136.426.143.3
+EquiPose†13.2163cm/21.7°132.926.244.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.628, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Table 3. Comparison on the Mapfree dataset [1]. As done in [1], we report Pose Error in terms of rotation error and translation error and VCRE in terms of median error and precision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.905, + 0.733 + ], + "angle": 0, + "content": "These results demonstrate that EquiPose is also effective under the cross-dataset setup." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Besides the above quantitative results, we also provide visualization results from ScanNet. Considering GRelPose [28] performs relatively better than other baseline models in most cases, we adopt it as the baseline and compare it with EquiPose. The corresponding results are shown in Fig. 4. As seen from this figure, the proposed EquiPose is effective in improving the performance of relative pose estimation across different types of challenges, such as large viewpoint changes, and texture-less patterns. An inference-only implementation demonstrates obvious performance improvements, while subsequent fine-tuning could achieve" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1133" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.511, + 0.268 + ], + "angle": 0, + "content": "
MethoddRdtRot. AUCTrans. AUCRETE
@3°@5°@10°@.1m@.2m@.5m
EightVit[46]2.70.1223.6940.2561.9214.9032.9558.478.900.41
+RP1.20.0522.7239.9361.8314.6033.8261.238.230.33
+RP+EquiPose0022.9640.1162.0315.0234.3061.738.200.31
+EquiPose†0025.3042.5464.0516.3736.1463.077.970.31
MF6D[1]7.30.1822.3637.1957.1014.3932.5757.349.810.44
+RP5.40.1120.4236.6258.1114.5633.0160.368.170.32
+RP+EquiPose0024.6140.2660.8516.4335.6362.757.740.30
+EquiPose†0027.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]3.20.1228.9046.3666.5316.4635.3660.068.160.38
+RP2.10.0628.1145.6666.3616.9537.6563.647.740.30
+RP+EquiPose0029.0246.3566.8018.2038.8864.477.650.29
+EquiPose†0032.3249.7469.3719.6740.4065.957.610.28
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.271, + 0.483, + 0.314 + ], + "angle": 0, + "content": "Table 4. EquiPose VS. Random Permutation. \\(^+\\mathrm{RP}\\) denotes the model is trained with random permutation. \\(d_{R}\\) and \\(d_{t}\\) denote the rotation and translation difference computed via Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.293, + 0.359 + ], + "angle": 0, + "content": "additional performance gains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Outdoor Pose Evaluation. The Mapfree dataset [1] was introduced for evaluating 6DoF relocalization, thus we only evaluate EightVit, MF6D, and GRelPose which could predict 6DoF poses. As a supplement, we further evaluate EquiPose in an inference-only manner on four additional baselines, including MFBin, MFBinTS, MFQuatTS which were released1 along with MF6D [1]. The results are reported in Tab. 3. As seen from this table, EquiPose could boost the performances of the above seven models in most cases. And the fine-tuned MF6D+EquiPose† achieves the lowest pose error among them. These results validate the effectiveness of our method under the outdoor setup." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.543, + 0.483, + 0.68 + ], + "angle": 0, + "content": "EquiPose VS. Random Permutation. Intuitively, another way to leverage the PPE property for relative pose estimation is to permute the input order of image pairs randomly during training. To evaluate the effect of random permutation, we have further trained and tested EightVit [46], MF6D [1] and GRelPose [28] on the ScanNet dataset by using random permutation (RP). The corresponding results are reported in Tab. 4. In addition, we also report the average values of \\( d_{R} \\) and \\( d_{t} \\) computed via Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.681, + 0.483, + 0.862 + ], + "angle": 0, + "content": "As seen from this table, random permutation could improve model performances in average rotation and translation errors. However, the models trained under EquiPose perform significantly better than using random permutation. Besides, random permutation could not strictly enforce the PPE property as evidenced by both \\( d_{R} \\) and \\( d_{t} \\) being larger than zero, while EquiPose could achieve PPE strictly. To this end, EquiPose is not opposed to using random permutation, i.e., by enforcing the PPE property, it could also boost the performance of models trained using random permutation as observed in this table. The above results further validate the effectiveness of our method." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.089, + 0.895, + 0.25 + ], + "angle": 0, + "content": "
MethodParams.FLOPsTime (ms)
ER1.92 × 1079.86 × 1092.0
ER+EquiPose1.92 × 1071.09 × 10102.3
Reg6D2.33 × 1079.91 × 1092.6
Reg6D+EquiPose2.33 × 1079.97 × 1092.6
EightVit1.91 × 1076.88 × 1093.5
EightVit+EquiPose1.91 × 1079.64 × 1095.1
MF6D2.19 × 1074.57 × 10105.6
MF6D+EquiPose2.19 × 1074.60 × 10106.7
GRelPose1.37 × 1079.82 × 10911.8
GRelPose+EquiPose1.37 × 1071.13 × 101012.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.251, + 0.905, + 0.28 + ], + "angle": 0, + "content": "Table 5. Comparison of parameter numbers, FLOPs, and inference times (ms) per image on an RTX 3090Ti GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.307, + 0.871, + 0.323 + ], + "angle": 0, + "content": "4.4. Computation Comparison and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.905, + 0.497 + ], + "angle": 0, + "content": "Here, we compare the parameter numbers, Floating Point Operations (FLOPs) and the inference times of the baselines and their EquiPose implementations. All the models are evaluated on an RTX 3090Ti GPU on ScanNet. The results are reported in Tab. 5. As is seen, compared with the baseline model, the EquiPose implementation does not introduce any additional parameters, since it uses the same backbone as the baseline, and thus does not incur extra storage costs. Moreover, compared with each baseline model, EquiPose results in a modest increment in computational costs generally, due to the feature permutation strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.633 + ], + "angle": 0, + "content": "However, the extra computational cost depends on the architecture of the baseline model: For models with lightweight feature aggregation modules and pose decoders, the increase in computation is slight. For models with heavy feature aggregation modules and pose decoders, the extra computational cost would be high, which is a limitation of EquiPose. Despite this, all the models under EquiPose maintain an inference time of less than 15 ms per image, which is sufficient for practical applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.64, + 0.664 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.854 + ], + "angle": 0, + "content": "This work focuses on end-to-end relative pose estimation. Firstly, we introduce the overlooked pose permutation equivariance (PPE) property for relative pose estimation. Leveraging this property, we establish the foundational principle for designing a relative pose estimation framework that incorporates the PPE property. Building on this theoretical basis, we propose the general and flexible EquiPose framework for relative pose estimation. Experiments on four public datasets further demonstrate that the performances of pre-trained models could be improved even without re-training, and moreover, their performances could be further boosted after being fine-tuned under EquiPose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62376269)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.903 + ], + "angle": 0, + "content": "1Pretrained weights are downloaded from the website of Mapfree: https://github.com/nianticlabs/map-free-reloc." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1134" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 2, 3, 5, 6, 7, 8, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Georg Bokman, Johan Edstedt, Michael Felsberg, and Fredrik Kahl. Steerers: A framework for rotation equivariant keypoint descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4885-4895, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[5] Jane Bromley, Isabelle Guyon, Yann LeCun, and et al. Signature verification using a \"siamese\" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14566–14575, 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[9] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Shay Dekel and Yosi Keller. Estimating extreme 3d image rotation with transformer cross-attention. arXiv preprint arXiv:2303.02615, 2023. 1, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition workshops, pages 224-236, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[14] Qiulei Dong, Xiang Gao, Hainan Cui, and Zhanyi Hu. Robust camera translation estimation via rank enforcement. IEEE Transactions on Cybernetics, 52(2):862-872, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[15] Qulei Dong, Shuang Deng, and Yuzhen Liu. Cayley rotation averaging: Multiple camera averaging under the cayley framework. IEEE Transactions on Image Processing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.415, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[18] Johan Edstedt, Georg Bökman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. In 2024 International Conference on 3D Vision, pages 148–157. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[19] Sovann En, Alexis Lechery, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision Workshops, 2018. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[20] Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and ransac's role. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8935-8943, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[22] Richard Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In CVPR 2011, pages 3041-3048. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.694, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[23] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hongdong Li. Rotation averaging. International Journal of Computer Vision, 103:267-305, 2013. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[24] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(6):580-593, 1997. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Will Hutchcroft, Yuguang Li, Ivaylo Boyadzhiev, Zhiqiang Wan, Haiyan Wang, and Sing Bing Kang. Covispose: Covisibility pose transformer for wide-baseline relative pose estimation in 360 indoor panoramicas. In European Conference on Computer Vision, pages 615-633. Springer, 2022. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1135" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Nicolas Keriven and Gabriel Peyre. Universal invariant and equivariant graph neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[28] Fadi Khatib, Yuval Margalit, Meirav Galun, and Ronen Basri. Leveraging image matching toward end-to-end relative camera pose regression, 2024. 1, 2, 3, 6, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[29] Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Deep Learning Workshop, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[30] Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International Conference on Machine Learning, pages 2747-2755. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[32] Yuzhen Liu and Qiulei Dong. Descriptor distillation: A teacher-student-regularized framework for learning local descriptors. International Journal of Computer Vision, pages 1-19, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[33] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[34] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[35] Clare Lyle, Mark van der Wilk, Marta Kwiatkowska, Yarin Gal, and Benjamin Bloem-Reddy. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[36] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional neural networks. In International Conference on Advanced Concepts for Intelligent Vision Systems, pages 675-687. Springer, 2017. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[37] Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. In Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[38] Maher Moakher. Means and averaging in the group of rotations. SIAM journal on matrix analysis and applications, 24 (1):1-16, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[39] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[40] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mokolajczyk. Ood-pose: camera pose regression from out-of-distribution synthetic views. In International Conference on 3D Vision, pages 722-732. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[42] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.907, + 0.328 + ], + "angle": 0, + "content": "[45] Jerome Revaud, Yohann Cabon, Romain Brégier, JongMin Lee, and Philippe Weinzaepfel. Sacreg: Scene-agnostic coordinate regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 688–698, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[46] Chris Rockwell, Justin Johnson, and David F. Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision, 2022. 1, 3, 5, 6, 7, 8, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[47] Chris Rockwell, Nilesh Kulkarni, Linyi Jin, Jeong Joon Park, Justin Johnson, and David F Fouhey. Far: Flexible accurate and robust 6 dof relative camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19854-19864, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[48] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[49] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[51] Yurun Tian, Axel Barroso-Laguna, Tony Ng, Vassileios Balntas, and Krystian Mikolajczyk. Hynet: Learning local descriptor with hybrid similarity measure and triplet loss. In Advances in Neural Information Processing Systems, pages 7401-7412, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[52] Julien Valentin, Angela Dai, Matthias Nießner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In International Conference on 3D Vision, pages 323-332. IEEE, 2016. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.805, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[53] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Guangyu Wang, Jinzhi Zhang, Fan Wang, Ruqi Huang, and Lu Fang. Xscale-nvs: Cross-scale novel view synthesis with" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1136" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "hash featurized manifold. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21029-21039, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Asian Conference on Computer Vision, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[57] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[58] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual u-net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[59] Xiaoming Zhao, Xingming Wu, Jinyu Miao, Weihai Chen, Peter CY Chen, and Zhengguo Li. Alike: Accurate and lightweight keypoint detection and descriptor extraction. IEEE Transactions on Multimedia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[60] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4669-4678, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[61] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1851-1858, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 3, 6, 7, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1137" + } + ] +] \ No newline at end of file diff --git a/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_origin.pdf b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..44b637a77429917eed77efba1a85c3c6145179b2 --- /dev/null +++ b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/6aabd5e5-4c6d-47d9-bb8c-4cbc4163ed65_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41c8106c20103d7d033c69cc3f13bff49b338796206ed5b8f92253a08537fde7 +size 990780 diff --git a/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/full.md b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d4e135b5170471b9db5a248d48ec1e6c1624120f --- /dev/null +++ b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/full.md @@ -0,0 +1,393 @@ +# EquiPose: Exploiting Permutation Equivariance for Relative Camera Pose Estimation + +Yuzhen Liu1,2 Qialei Dong1,2,* + +$^{1}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, Chinese Academy of Sciences, Institute of Automation $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences liuyuzhen2022@ia.ac.cn, qldong@nlpr.ia.ac.cn + +# Abstract + +Relative camera pose estimation between two images is a fundamental task in 3D computer vision. Recently, many relative pose estimation networks have been explored for learning a mapping from two input images to their corresponding relative pose, however, the estimated relative poses by these methods do not have the intrinsic Pose Permutation Equivariance (PPE) property: the estimated relative pose from Image A to Image B should be the inverse of that from Image B to Image A. It means that permuting the input order of two images would cause these methods to obtain inconsistent relative poses. To address this problem, we firstly introduce the concept of PPE mapping, which indicates such a mapping that captures the intrinsic PPE property of relative poses. Then by enforcing the aforementioned PPE property, we propose a general framework for relative pose estimation, called EquiPose, which could easily accommodate various relative pose estimation networks in literature as its baseline models. We further theoretically prove that the proposed EquiPose framework could guarantee that its obtained mapping is a PPE mapping. Given a pre-trained baseline model, the proposed EquiPose framework could improve its performance even without fine-tuning, and could further boost its performance with fine-tuning. Experimental results on four public datasets demonstrate that EquiPose could significantly improve the performances of various state-of-the-art models. + +# 1. Introduction + +Estimating the relative camera pose between an image pair is one of the fundamental tasks in 3D computer vision [6, 46]. It has a broad range of applications, such as structure-from-motion (SfM) [14, 15, 42], visual localization [10, 40, 45], novel view synthesis [7, 53, 54], etc. + +Existing methods for relative camera pose estimation in literature could be roughly divided into two categories: matching-based methods and end-to-end methods. The matching-based methods [13, 33] generally employ a two-stage computational strategy: Firstly, they establish the point correspondences between an input pair of images by utilizing local feature extraction and matching techniques [13, 33]. Then according to the established correspondences, they solve a two-view geometric problem via a typical 5-point algorithm [24] with robust fitting [21] for pursuing the relative pose. However, since the performances of the matching-based methods depend on the established point correspondences, their performance would significantly decrease or even fail in cases of weak/repeated textures and degenerated configurations as indicated in [20]. + +Unlike the above matching-based methods, many recent methods [6, 12, 19, 28, 36, 46, 57, 61] have employed deep neural networks to directly learn a mapping from the input two images to their relative pose without explicitly establishing point correspondences. Here, it has to be pointed out that ideally, the estimated relative pose by these end-to-end methods should have the following intrinsic pose permutation equivariance (PPE) property between an image pair: + +# Pose Permutation Equivariance + +Let $\{I_1, I_2\}$ be a pair of images, then the relative pose $P_{12}$ from $I_1$ to $I_2$ should be the inverse of the relative pose $P_{21}$ from $I_2$ to $I_1$ : + +$$ +P _ {2 1} = P _ {1 2} ^ {- 1}. \tag {1} +$$ + +By writing $P_{12}$ and $P_{21}$ in the matrix form: + +$$ +P _ {1 2} = \left[ \begin{array}{c c} R _ {1 2} & t _ {1 2} \\ 0 & 1 \end{array} \right], \quad P _ {2 1} = \left[ \begin{array}{c c} R _ {2 1} & t _ {2 1} \\ 0 & 1 \end{array} \right], \qquad (2) +$$ + +where $R_{12}$ and $t_{12}$ are the relative rotation and translation from $I_1$ to $I_2$ , and $R_{21}$ and $t_{21}$ are the relative rotation and + +![](images/95d0c77578704945fac8104711d83f639038b1d61035d9c8a87370e9d4843937.jpg) +(a) MF6D + +![](images/0a1e6494f3bbda22b17b6a21ccb77bac05e18b6b496cdc5dd3e7f27567f80a1f.jpg) +Figure 1. Distribution of the $d_{R}$ and $d_{t}$ obtained by existing models MF6D [1] and GRelPose [28]. + +![](images/7177829fded36f4b5fde7718d37c7cf190ca49c4d3791bae9c79c12a40577087.jpg) + +![](images/948ac687236a6619f2a1ca5c04e11d47a6ca809eefdc519933841743e0ec182a.jpg) +(b) GRelPose + +translation from $I_{2}$ to $I_{1}$ , Eq. (1) could be reformulated as: + +$$ +R _ {2 1} = R _ {1 2} ^ {\mathrm {T}}, \quad t _ {2 1} = - R _ {1 2} ^ {\mathrm {T}} t _ {1 2}. \tag {3} +$$ + +However, the PPE property is generally overlooked when designing model architectures in existing end-to-end methods [1, 28]. We validate this by conducting the following experiment: we trained two typical end-to-end models (MF6D [1] and GRelPose [28]) on the training set of ScanNet [11]. Then, we adopt the test set of ScanNet [48, 50] which contains 1500 image pairs, and feed them with two different orders to the above models respectively, leading to totally 1500 pairs of the predicted relative poses by each model, where each pair consists of the estimated pose $P_{12}$ from $I_{1}$ to $I_{2}$ and the estimated pose $P_{21}$ from $I_{2}$ to $I_{1}$ . Next, we compute the differences between the estimated relative pose $P_{21}$ and $P_{12}^{-1}$ . Specifically, we compute the angular difference $d_{R}$ between relative rotations $R_{21}$ and $R_{12}^{\mathrm{T}}$ , and the Euclidean distance $d_{t}$ between relative translations $t_{21}$ and $-R_{12}^{\mathrm{T}} t_{12}$ respectively via: + +$$ +d _ {R} = \arccos ((\operatorname {t r} (R _ {2 1} R _ {1 2}) - 1) / 2), \tag {4} +$$ + +$$ +d _ {t} = \left\| t _ {2 1} + R _ {1 2} ^ {\mathrm {T}} t _ {1 2} \right\| _ {2}, +$$ + +where $\mathrm{tr}(\cdot)$ denotes matrix trace, $\| \cdot \| _2$ denotes $L_{2}$ norm. + +The distributions of the rotation difference $d_{R}$ and translation difference $d_{t}$ from the above two models are shown in Fig. 1. As seen from this figure, the pose differences (both rotation and translation) estimated by these models are generally larger than zero (particularly some of them are not trivial), demonstrating that the above models do not satisfy the PPE property. Such an issue may pose risks to downstream applications that require a stable and consistent estimation of relative poses, such as robotics and visual navigation. Moreover, it is also noted that some other visual tasks involve their specific equivariances, different from the PPE property in the relative pose estimation task, and once an existing model among these tasks is equipped with a task-specific equivariance, its performance could be further boosted [4, 27, 35, 39, 43]. + +Motivated by the above issues, we investigate whether it is possible to explore such a relative pose estimation framework that could not only bestow the PPE property on various relative pose estimation networks but also improve their + +performances further. Specifically, we first introduce the concept of PPE mapping, which refers to a mapping that captures the PPE property in relative poses. Then, we establish the foundational principle for equivariant relative pose estimation. Accordingly, we propose the EquiPose framework which holds the PPE property theoretically. The EquiPose could improve the performance of a pre-trained model even without fine-tuning, and fine-tuning could enhance their performances further. + +To summarize, our main contributions include: + +- Equivariance: The simple but intrinsic pose permutation equivariance (PPE) property is revealed for relative pose estimation, and the PPE mapping is defined. +- Generality: By imposing the PPE property, the framework EquiPose for equivariant relative pose estimation is proposed. An arbitrary end-to-end relative pose estimation model could be seamlessly embedded into this framework as a baseline, and accordingly obtains the PPE property as analyzed in theory in Sec. 3. The EquiPose framework and the theoretical analyses might give new insights into model design strategies for relative pose estimation. +- Flexibility: Thanks to the imposed PPE property, the performances of the derived methods from the EquiPose framework by utilizing many off-the-shelf baseline models, would be boosted even without fine-tuning by adding several lines of code. Moreover, once they are fine-tuned under EquiPose, their performances would be further improved, as demonstrated by the results in Sec. 4. + +# 2. Related Work + +In this section, we review some typical matching-based methods and end-to-end methods for relative pose estimation respectively. + +Matching-Based Relative Pose Estimation. The matching-based relative pose estimation methods typically involve the following two steps: establishing point correspondence via image matching [3, 33] and solving the two-view geometry with RANSAC [21, 24, 41]. Early approaches relied on handcrafted keypoint detectors and descriptors to establish point correspondences, such as SIFT [33] and SURF [3]. In recent years, numerous methods emerged that replace the handcrafted detectors or descrip + +tors with their learned counterparts [2, 17, 18, 32, 34, 37, 44, 51, 59]. Additionally, a variety of learning-based image-matching methods directly established point correspondences from keypoints or raw images [8, 31, 48, 50, 60]. + +However, since these methods depend on point correspondences, they may fail in cases of weak textures, repeated textures, large viewpoint changes, and degenerated configurations [20]. Additionally, these methods could be time-consuming due to image matching and RANSAC loops. Moreover, they could only recover the relative pose up to an unknown scale factor [24]. + +End-to-End Relative Pose Estimation. Recently, unlike the above strategies, many works have adopted deep neural networks to regress relative poses in an end-to-end manner [1, 6, 9, 12, 19, 28, 36]. These methods do not explicitly implement image matching or solve the two-view geometry problem. Instead, the networks directly learn the underlying mapping from the input image pair to the corresponding relative pose. Compared with matching-based methods, they are shown to be more robust in challenging scenes [6, 9, 19] with a significantly faster inference speed. Moreover, these end-to-end methods could be further incorporated with traditional methods for improving robustness [47]. + +These end-to-end methods generally adopt Siamese architectures [5] to extract feature, which are aggregated and decoded into the relative pose subsequently [6, 12, 26, 28, 46]. To facilitate stable training, early approaches [19, 36] parameterized the rotation as quaternions. Different from them, Zhou et al. [62] suggested to represent 3D rotations in a 6D space, which benefits the training of neural networks. Cai et al. [6] and Dekel et al. [12] showed that the network could learn from implicit cues, allowing the network to estimate the relative pose even when two images have no overlap. Rockwell et al. [46] introduced the essential matrix module to emulate the 8-point algorithm [24]. Arnold et al. [1] adopted a differentiable warping module for feature aggregation. Khatib et al. [28] embedded a semidense image matcher [50] for feature extraction to pursue a better generalization ability. + +However, the existing end-to-end methods generally do not satisfy the PPE property, leading to inconsistent estimation when permuting the input order of an image pair as shown in Sec. 1. Unlike them, this work does not aim to propose a new model for relative pose estimation, but pursue a general framework that could bestow the PPE property on various relative pose estimation models, and improve their performances further. + +# 3. Methodology + +In this section, we first introduce the key definitions and propositions related to the pose permutation equivariance (PPE) property of relative poses. Then, we describe the proposed EquiPose framework in detail. + +# 3.1. PPE Mapping and Preliminary Knowledge + +Prior to the introduction of the PPE mapping, 'equivariant mapping' in mathematics is defined as [30]: + +Definition 1 (Equivariant mapping). Let $\mathcal{G}$ be a group that acts on spaces $\mathcal{X}$ and $\mathcal{Y}$ , and $\mathbb{T}$ and $\mathbb{T}'$ be the induced actions of $\mathcal{G}$ on $\mathcal{X}$ and $\mathcal{Y}$ . A mapping $f: \mathcal{X} \to \mathcal{Y}$ is equivariant with the action in $\mathcal{G}$ if $f(\mathbb{T}_g(x)) = \mathbb{T}_g'(f(x)), \forall x \in \mathcal{X}, g \in \mathcal{G}$ . + +According to Definition 1 and Eq. (1), a PPE mapping with the PPE property is defined as: + +Definition 2 (Pose permutation equivariant mapping). A mapping $f: \mathbb{R}^{H \times W} \times \mathbb{R}^{H \times W} \to \mathrm{SE}(3)$ (the special Euclidean group) is pose permutation equivariant, if $f(I_2, I_1) = f(I_1, I_2)^{-1}, \forall I_1, I_2 \in \mathbb{R}^{H \times W}$ . + +Definition 2 could be considered as an instance of 'equivariant mapping' from Definition 1 on relative camera poses, where the input variable $x$ is an image pair $\{I_1, I_2\}$ , the action $\mathbb{T}_g$ on $\{I_1, I_2\}$ is an order permutation operator, and the action $\mathbb{T}_q'$ is the inverse of a matrix. + +It is noted that Definition 2 is applicable to neural networks, since any neural networks essentially aim to learn a mapping from their input to the output. + +In addition to the above two definitions, we introduce the following definition on the geodesic mean of rotations [22, 23], as well as two theorems, which would be used to support the fundamental principle for designing a relative pose estimation framework with the PPE property in the next subsection: + +Definition 3 (Geodesic mean of rotations). Given a set of rotations $\{R_i\}_{i=1}^n$ , their geodesic mean $S$ is defined as: + +$$ +S = \mathfrak {G} \left(R _ {1}, \dots , R _ {n}\right) = \underset {S \in \mathrm {S O} (3)} {\arg \min } \sum_ {i = 1} ^ {n} \| \log \left(R _ {i} ^ {\mathrm {T}} S\right) \| _ {2} ^ {2}, \tag {5} +$$ + +where $\log (\cdot)$ is the mapping from an element in the Lie group to its corresponding Lie algebra. + +From Definition 3, we have the two theorems as: + +Theorem 1. Given a set of rotations $\{R_i\}_{i=1}^n$ , it always holds that: $\mathfrak{G}(R_{\sigma(1)}, \dots, R_{\sigma(n)}) = \mathfrak{G}(R_1, \dots, R_n)$ for any permutation $\sigma$ . + +Theorem 2. If $S \in \mathrm{SO}(3)$ is the geodesic mean of the set of rotations $\{R_i\}_{i=1}^n$ , then $S^{\mathrm{T}}$ is the geodesic mean of the set of rotations $\{R_i^{\mathrm{T}}\}_{i=1}^n$ . + +Due to the space limitation, the proofs of the above two theorems are given in the supplementary material. + +![](images/93c2cc69918823371827c0cec9501ef7356e583cfb66fa72da7c1d62d2638db8.jpg) +Figure 2. Illustration of the existing methods $f$ and our method $f^{*}$ on the SE(3) manifold. Left: The existing method $f$ does not capture the PPE property in the relative pose. Right: Within EquiPose $f^{*}$ , the desired PPE property is satisfied (Theorem 3), and the results could be more reliable (Sec. 4). + +# 3.2. Fundamental Principle of Pose Estimation Framework with PPE + +According to the aforementioned PPE mapping and related preliminary knowledge, we elaborate the fundamental principle for designing a pose estimation framework with PPE, which could use an arbitrary end-to-end relative pose estimation model as its baseline and derive the corresponding upgraded model with the PPE property as follows: + +Given an arbitrary relative pose estimation network corresponding to a mapping $f$ , there exists the following mapping $f^{*}$ , which is a PPE mapping: + +Theorem 3. Given an arbitrary mapping $f: \mathbb{R}^{H \times W} \times \mathbb{R}^{H \times W} \to \mathrm{SE}(3)$ , where $f(I_1, I_2) := (R_{12}, t_{12})$ and $f(I_2, I_1) := (R_{21}, t_{21})$ , the mapping $f^*$ defined as: + +$$ +f ^ {*} \left(I _ {1}, I _ {2}\right) := \left(R _ {1 2} ^ {*}, t _ {1 2} ^ {*}\right), \tag {6} +$$ + +where + +$$ +R _ {1 2} ^ {*} = \mathfrak {G} \left(R _ {1 2}, R _ {2 1} ^ {\mathrm {T}}\right), \tag {7} +$$ + +$$ +t _ {1 2} ^ {*} = \left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\right) / 2, \tag {8} +$$ + +is a PPE mapping. + +Proof. According to Theorem 1, we have + +$$ +R _ {1 2} ^ {*} = \mathfrak {G} \left(R _ {1 2}, R _ {2 1} ^ {\mathrm {T}}\right) = \mathfrak {G} \left(R _ {2 1} ^ {\mathrm {T}}, R _ {1 2}\right) \tag {9} +$$ + +Then, according to Theorem 2, we have: + +$$ +\mathfrak {G} \left(R _ {2 1} ^ {\mathrm {T}}, R _ {1 2}\right) = \mathfrak {G} \left(R _ {2 1}, R _ {1 2} ^ {\mathrm {T}}\right) ^ {\mathrm {T}} = R _ {2 1} ^ {* \mathrm {T}} \tag {10} +$$ + +According to Eqs. (9) and (10), it holds: + +$$ +R _ {1 2} ^ {*} = R _ {2 1} ^ {* \mathrm {T}}, \tag {11} +$$ + +i.e., $f^{*}$ is permutation equivariant in terms of rotation. + +Now, according to (6), (7), (8) and (11), it holds: + +$$ +\begin{array}{l} f ^ {*} (I _ {1}, I _ {2}) ^ {- 1} = \left(R _ {1 2} ^ {* \mathrm {T}}, - R _ {1 2} ^ {* \mathrm {T}} t _ {1 2} ^ {*}\right) \\ = \left(R _ {1 2} ^ {* \mathrm {T}}, - R _ {1 2} ^ {* \mathrm {T}} \left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\right) / 2\right) \tag {12} \\ = \left(R _ {2 1} ^ {*}, \left(t _ {2 1} - R _ {2 1} ^ {*} t _ {1 2}\right) / 2\right) \\ = f ^ {*} \left(I _ {2}, I _ {1}\right), \\ \end{array} +$$ + +which completes the proof. + +![](images/da3249d170907e792178917639d91d45072a5375612001bac512854000cab9e5.jpg) + +# Algorithm 1 The EquiPose Framework + +Input: A relative pose estimation network $f$ ; A pair of images $\{I_1, I_2\}$ + +Output: Equivariant pose estimation $R_{12}^{*}, t_{12}^{*}$ + +1: $R_{12}, t_{12} \gets f(I_1, I_2)$ +2: $R_{21}, t_{21} \gets f(I_2, I_1)$ +3: $R_{12}^{*} \gets R_{12}(\exp \left(\frac{1}{2}\log (R_{12}^{\mathrm{T}}R_{21}^{\mathrm{T}})\right))$ +4: $t_{12}^{*}\gets (t_{12} - R_{12}^{*}t_{21}) / 2$ + +For ease of understanding, we visualize the difference between $f$ and $f^{*}$ in Fig. 2. As seen from this figure, given two input images $\{I_1, I_2\}$ , the existing relative pose estimation model (considered as a mapping) $f$ does not satisfy the PPE property. Once it is embedded into EquiPose, the upgrade version $f^{*}$ captures the PPE property strictly. Moreover, its estimation could be more reliable than the baseline model $f$ , as demonstrated in Sec. 4 (also in Sec. S8.3 of the supplementary material). + +It is noted in Eq. (7) that the construction of $f^{*}$ involves computing the geodesic mean of two rotations, which could be computed explicitly via [38]: + +$$ +\mathfrak {G} \left(R _ {1 2}, R _ {2 1} ^ {\mathrm {T}}\right) = R _ {1 2} \left(R _ {1 2} ^ {\mathrm {T}} R _ {2 1} ^ {\mathrm {T}}\right) ^ {1 / 2}. \tag {13} +$$ + +Considering that we work on the 3D rotation group, the square root of a rotation $R \in \mathrm{SO}(3)$ could be computed by: + +$$ +R ^ {1 / 2} = \exp \left(\frac {1}{2} \log (R)\right), \tag {14} +$$ + +where $\exp (\cdot)$ is the mapping from the Lie algebra to the Lie group. + +The above analysis reveals that given an arbitrary relative pose estimation network, once its upgrade version $f^{*}$ is constructed by (6), the mapping is a pose permutation equivariant mapping, this is to say, the upgrade version has the PPE property. + +# 3.3. EquiPose + +According to the revealed point above as well as Eqs. (7) and (8), a simple framework for relative pose estimation, i.e., the EquiPose, could be straightforwardly built. + +Specifically, the proposed EquiPose uses an existing model as its baseline, and outputs the final estimation on relative pose by calculating Eqs. (7) and (8), whose variable values are assigned by the baseline with or even without fine-tuning. Algorithm 1 shows the complete pipeline of EquiPose to obtain the relative pose. + +It has to be pointed out that the computational cost of $f^{*}$ in this implementation is twice that of $f$ , since it involves two times of forward propagation of $f$ : one for $f(I_{1},I_{2})$ , and one for $f(I_{2},I_{1})$ after permuting the input images $\{I_1,I_2\}$ . However, in practice, the computational + +![](images/8c0b88ccccba0312ba9e2174af8dcb30830b696b3f49912ccd59909016c59abc.jpg) + +![](images/53580ed6212c9472eb2b2d6d6b80de5a9714aa6f378b03440eea2dec914682eb.jpg) +(a) +(b) +Figure 3. The pipelines of existing methods and our method. (a) The pipeline of existing relative pose estimation networks, which consists of two stages: Stage (i) Single-Image Feature Encoding $(f_{\mathrm{enc}})$ extracts the feature maps $\{F_1,F_2\}$ from $\{I_1,I_2\}$ . Stage (ii) Cross-Image Feature Decoding $(f_{\mathrm{dec}})$ takes the extracted two feature maps $\{F_1,F_2\}$ as input, and it outputs the estimated relative pose. (b) The pipeline of EquiPose. The computational overhead from EquiPose is an additional forward propagation of $f_{\mathrm{dec}}$ , which is slight. + +overhead could be reduced further with the feature permutation strategy, which is elaborated in the following. + +To begin with, it is noted that the general pipeline of existing relative pose estimation models [1, 6, 9, 12, 19, 36, 46] could be summarized in Fig. 3a. As is seen, the pipeline consists of the two stages, which we name as (i) Single-Image Feature Encoding (denoted as $f_{\mathrm{enc}}$ ), and (ii) Cross-Image Feature Decoding (denoted as $f_{\mathrm{dec}}$ ) respectively. +(i) Single-Image Feature Encoding $f_{\mathrm{enc}}$ : This stage takes two images as inputs, and it outputs the corresponding feature maps $\{F_1,F_2\}$ , i.e., $f_{\mathrm{enc}}(I_1,I_2)\coloneqq \{F_1,F_2\}$ . + +It contains a feature encoder, which could be any feature-extraction backbone in computer vision. It should be noted that the weights of the feature encoder for extracting the features of image $I_{1}$ and $I_{2}$ are shared (generally referred to as the Siamese architecture [5, 29]). + +(ii) Cross-Image Feature Decoding $f_{\mathrm{dec}}$ : This stage takes the feature maps $\{F_1, F_2\}$ obtained from Stage (i) as input, and it outputs the estimated relative pose $P_{12}$ , i.e., $f_{\mathrm{dec}}(F_1, F_2) := P_{12}$ . It contains a feature aggregation module and a pose decoder. The feature aggregation module maps $\{F_1, F_2\}$ into a single aggregated feature map (or vec + +
MethodRot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m
ER[6]13.1528.2350.43---14.66-
+EquiPose15.2230.4051.51---14.11-
+EquiPose†16.0031.0152.98---12.53-
Reg6D[62]10.1622.9745.76---12.26-
+EquiPose11.7625.2847.94---11.76-
+EquiPose†16.1431.5854.20---10.81-
EightVit[46]23.6940.2561.9214.9032.9558.478.900.41
+EquiPose24.5841.3262.8215.4934.1160.038.400.35
+EquiPose†25.3042.5464.0516.3736.1463.077.970.31
MF6D[1]22.3637.1957.1014.3932.5757.349.810.44
+EquiPose25.4940.5359.9417.5135.7860.498.420.35
+EquiPose†27.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]28.9046.3666.5316.4635.3660.068.160.38
+EquiPose30.7048.0267.9017.8537.1761.807.970.32
+EquiPose†32.3249.7469.3719.6740.4065.957.610.28
+ +Table 1. Comparison on the ScanNet dataset [11, 50]. RE: rotation errors in degrees. TE: translation errors in meters. +EquiPose: the models are implemented under EquiPose only during the inference stage. +EquiPose†: the models are further fine-tuned under EquiPose. -: the models only predict the relative camera rotation and thus could not be evaluated in translation. The best methods for each baseline are marked in bold. + +tor), which is subsequently decoded by the pose decoder to obtain the estimated relative pose. + +Since the weights of the feature encoder in $f_{\mathrm{enc}}$ are shared, we do not have to implement two complete forward propagation of $f$ . Instead, we could reuse the extracted feature maps $\{F_1, F_2\}$ at the first forward propagation, then permute the feature maps to $\{F_2, F_1\}$ , and implement one additional forward propagation of $f_{\mathrm{dec}}$ . + +Fig. 3b illustrates the EquiPose framework with the feature permutation strategy. As is seen, it first performs one forward propagation of $f_{\mathrm{enc}}$ to extract feature maps $\{F_1,F_2\}$ . Then, the feature maps $\{F_1,F_2\}$ and its permuted counterpart $\{F_2,F_1\}$ are inputted to $f_{\mathrm{dec}}$ to obtain $P_{12}$ and $P_{21}$ respectively. The final relative pose is obtained via $P_{12}$ and $P_{21}$ through Eq. (6). Compared to the original implementations of the baseline models, EquiPose only involves an additional forward propagation of $f_{\mathrm{dec}}$ . + +Under the EquiPose framework, a pre-trained model could be directly adopted as the baseline without fine-tuning to boost its performance. Moreover, the model could also be further fine-tuned under EquiPose to achieve further improvement. We simply use an $\mathcal{L}_1$ loss for fine-tuning: + +$$ +\mathcal {L} _ {1} = \| R ^ {*} - \hat {R} \| _ {1} + \lambda \| t ^ {*} - \hat {t} \| _ {1}, \tag {15} +$$ + +where $\hat{R}$ and $\hat{t}$ are the ground truth relative rotation and translation. $\lambda$ is a preset constant factor. + +![](images/7de4c1b8034b8f0c163a634504aff009436fbf6c83d2c5a1d9fced443207081a.jpg) +Figure 4. Visualization of epipolar lines on ScanNet [11]. Five SIFT keypoints are sampled from image $I_{1}$ (column 1) and their epipolar lines are projected onto image $I_{2}$ , using the pose from: ground truth (column 2), GRelPose baseline [28] (column 3), EquiPose without (column 4) and with finetuning (column 5). + +# 4. Experiment + +# 4.1. Datasets and Evaluation Metrics + +Datasets. We evaluate our method on both indoor and outdoor scenes. For indoor scenes, we use the ScanNet [11], the 7-Scene [49] and the 12-Scene [52] datasets, which are widely used for indoor pose estimation. For outdoor scenes, we use the recently proposed Mapfree dataset [1]. + +For ScanNet, we use 2M image pairs from the training set for training, and use the test set which contains 1500 pairs for evaluation following [8, 48, 50, 55, 56]. For both 7-Scene and 12-Scene, we use 720 image pairs for evaluation using the model trained on ScanNet. For Mapfree, we follow the official setup in [1] for training and evaluation. More details are provided in the supplementary material. + +Evaluation Metrics. For ScanNet [11], 7-Scene [49] and 12-Scene [52], we calculate the area under the cumulative curve (AUC) as done in [31, 48, 50]. We report the AUC for rotation $(3^{\circ} / 5^{\circ} / 10^{\circ})$ and translation $(0.1\mathrm{m} / 0.2\mathrm{m} / 0.5\mathrm{m})$ respectively. Besides, we also report the average rotation and translation errors respectively. For the Mapfree dataset [1], we report the pose error and virtual correspondence reprojection error (VCRE) respectively following [1]. + +# 4.2. Baselines + +It should be pointed out that the goal of this study is to introduce a general framework for existing end-to-end relative pose estimation models to achieve the PPE property with improved performances. Thus, we do not further evaluate and analyze other traditional matching-based methods, such as the classic SIFT [33] followed by the 5-point solver [41]. + +Specifically, we adopt the following five end-to-end models as baselines and evaluate the effectiveness of EquiPose in boosting their performances: (i) ExtremeRotation [6], (ii) Reg6D [6, 62], (iii) EightVit [46], (iv) + +
Method7-Scene12-Scene
Rot. AUC↑Trans. AUC↑RE↓TE↓Rot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m@3°@5°@10°@.1m@.2m@.5m
ER[6]11.1724.2846.70---9.82-14.6131.6057.15---6.94-
+EquiPose13.7226.6047.83---9.09-17.5734.3958.96---6.66-
+EquiPose†13.8227.1449.35---7.84-17.8335.3860.10---5.39-
Reg6D[62]10.0822.4445.92---7.78-13.1628.4255.27---5.67-
+EquiPose11.2023.7746.93---7.67-14.0530.0856.62---5.41-
+EquiPose†13.2627.3251.60---6.60-16.9933.5559.59---5.07-
EightVit[46]19.7735.0059.2513.9932.0561.114.920.2230.0950.3972.9538.0062.8084.032.890.08
+EquiPose19.8835.5159.8714.2132.9162.904.780.2031.1751.5373.5838.6763.7284.592.780.08
+EquiPose†20.8737.0161.4414.4733.7365.224.560.1932.7452.7274.4140.6365.3585.452.710.07
MF6D[1]18.7334.4657.6211.9030.3560.134.980.2330.4451.0773.1136.3761.5283.242.790.09
+EquiPose21.7837.4060.0315.4634.2863.714.710.1934.1754.2974.9840.8764.6785.022.560.08
+EquiPose†24.4440.8363.3217.1736.9066.444.290.1835.8055.3675.8243.0266.3685.782.450.07
GRelPose[28]26.7845.4767.9515.7034.2762.483.960.2135.6255.2075.6339.8763.8184.342.580.08
+EquiPose28.4446.8568.8715.8134.4163.483.880.2037.1356.6776.7142.2465.6585.212.410.07
+EquiPose†29.7648.2470.0815.8335.4364.913.660.1938.6358.1177.8644.1967.4386.042.240.07
+ +MF6D [1], and (v) GRelPose [28]. Note that (i) ExtremeRotation and (ii) Reg6D only predict relative rotation, while the remaining ones could predict the 6 DoF relative camera pose (including both rotation and translation). + +# 4.3. Comparative Evaluation + +Evaluation Setup. For each of the five baseline models, including ER (ExtremeRotation) [6], Reg6D [62], EightVit [46], MF6D [1] and GRelPose [28], we compare the following variants: 1) the model in its original implementation, 2) the model implemented under EquiPose only during the inference stage (denoted with $^+$ EquiPose), 3) the model fine-tuned under EquiPose (denoted with $^+$ EquiPose $\dagger$ ). + +Indoor Pose Evaluation. The evaluation results on ScanNet are reported in Table 1. As seen from this table, the five models under the EquiPose framework achieve consistently better performance than their corresponding baselines under the eight metrics. Moreover, these models after fine-tuning under EquiPose could achieve further enhancements in most cases by a significant margin. For example, an inference-only implementation of GRelPose under EquiPose leads to a relative decrease of $2.3\%$ (1-7.97/8.16) and $15.8\%$ (1-0.32/0.38) in rotation and translation errors respectively. Its fine-tuned implementation has a relative reduction of $6.7\%$ and $26.3\%$ in rotation and translation errors respectively. The fine-tuned MF6D could even achieve a relative reduction of $23.9\%$ and $34.1\%$ in rotation and translation errors respectively. These results validate the effectiveness of the proposed EquiPose framework. + +The cross-dataset evaluation results on the 7-Scene [49] and 12-Scene [52] datasets are reported in Tab. 2, where the models trained on ScanNet are used for evaluation. As seen from the table, the five models under EquiPose could also achieve consistent improvements than their corresponding baselines, and fine-tuning further boosts the performances. + +Table 2. Cross-dataset evaluation on the 7-Scene [49] and 12-Scene [52] datasets. All the models are trained on ScanNet. + +
MethodPose ErrorVCRE
Pr.↑Avg.Med.↓Med.↓Pr.@45↑Pr.@90↑
MFBin[1]9.1210cm/33.7°199.715.430.1
+EquiPose12.0185cm/32.5°196.918.231.6
MFBinTS[1]10.5183cm/23.3°166.418.135.5
+EquiPose12.9152cm/23.4°166.021.739.3
MFQuat[1]5.1203cm/27.3°160.714.333.6
+EquiPose6.3184cm/23.3°166.416.635.0
MFQuatTS[1]4.4185cm/26.3°154.814.234.7
+EquiPose6.6170cm/23.5°157.617.336.4
MF6D[1]6.1168cm/22.5°147.118.740.2
+EquiPose9.6153cm/18.6°148.222.140.9
+EquiPose†10.1148cm/18.4°145.023.042.1
EightVit[46]6.0204cm/26.6°155.217.835.8
+EquiPose6.4195cm/26.3°157.218.336.4
+EquiPose†6.8188cm/25.1°153.119.137.0
GRelPose[28]12.7168cm/22.4°137.725.843.3
+EquiPose13.0166cm/21.9°136.426.143.3
+EquiPose†13.2163cm/21.7°132.926.244.1
+ +Table 3. Comparison on the Mapfree dataset [1]. As done in [1], we report Pose Error in terms of rotation error and translation error and VCRE in terms of median error and precision. + +These results demonstrate that EquiPose is also effective under the cross-dataset setup. + +Besides the above quantitative results, we also provide visualization results from ScanNet. Considering GRelPose [28] performs relatively better than other baseline models in most cases, we adopt it as the baseline and compare it with EquiPose. The corresponding results are shown in Fig. 4. As seen from this figure, the proposed EquiPose is effective in improving the performance of relative pose estimation across different types of challenges, such as large viewpoint changes, and texture-less patterns. An inference-only implementation demonstrates obvious performance improvements, while subsequent fine-tuning could achieve + +
MethoddRdtRot. AUCTrans. AUCRETE
@3°@5°@10°@.1m@.2m@.5m
EightVit[46]2.70.1223.6940.2561.9214.9032.9558.478.900.41
+RP1.20.0522.7239.9361.8314.6033.8261.238.230.33
+RP+EquiPose0022.9640.1162.0315.0234.3061.738.200.31
+EquiPose†0025.3042.5464.0516.3736.1463.077.970.31
MF6D[1]7.30.1822.3637.1957.1014.3932.5757.349.810.44
+RP5.40.1120.4236.6258.1114.5633.0160.368.170.32
+RP+EquiPose0024.6140.2660.8516.4335.6362.757.740.30
+EquiPose†0027.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]3.20.1228.9046.3666.5316.4635.3660.068.160.38
+RP2.10.0628.1145.6666.3616.9537.6563.647.740.30
+RP+EquiPose0029.0246.3566.8018.2038.8864.477.650.29
+EquiPose†0032.3249.7469.3719.6740.4065.957.610.28
+ +Table 4. EquiPose VS. Random Permutation. $^+\mathrm{RP}$ denotes the model is trained with random permutation. $d_{R}$ and $d_{t}$ denote the rotation and translation difference computed via Eq. (4). + +additional performance gains. + +Outdoor Pose Evaluation. The Mapfree dataset [1] was introduced for evaluating 6DoF relocalization, thus we only evaluate EightVit, MF6D, and GRelPose which could predict 6DoF poses. As a supplement, we further evaluate EquiPose in an inference-only manner on four additional baselines, including MFBin, MFBinTS, MFQuatTS which were released1 along with MF6D [1]. The results are reported in Tab. 3. As seen from this table, EquiPose could boost the performances of the above seven models in most cases. And the fine-tuned MF6D+EquiPose† achieves the lowest pose error among them. These results validate the effectiveness of our method under the outdoor setup. + +EquiPose VS. Random Permutation. Intuitively, another way to leverage the PPE property for relative pose estimation is to permute the input order of image pairs randomly during training. To evaluate the effect of random permutation, we have further trained and tested EightVit [46], MF6D [1] and GRelPose [28] on the ScanNet dataset by using random permutation (RP). The corresponding results are reported in Tab. 4. In addition, we also report the average values of $d_{R}$ and $d_{t}$ computed via Eq. (4). + +As seen from this table, random permutation could improve model performances in average rotation and translation errors. However, the models trained under EquiPose perform significantly better than using random permutation. Besides, random permutation could not strictly enforce the PPE property as evidenced by both $d_{R}$ and $d_{t}$ being larger than zero, while EquiPose could achieve PPE strictly. To this end, EquiPose is not opposed to using random permutation, i.e., by enforcing the PPE property, it could also boost the performance of models trained using random permutation as observed in this table. The above results further validate the effectiveness of our method. + +
MethodParams.FLOPsTime (ms)
ER1.92 × 1079.86 × 1092.0
ER+EquiPose1.92 × 1071.09 × 10102.3
Reg6D2.33 × 1079.91 × 1092.6
Reg6D+EquiPose2.33 × 1079.97 × 1092.6
EightVit1.91 × 1076.88 × 1093.5
EightVit+EquiPose1.91 × 1079.64 × 1095.1
MF6D2.19 × 1074.57 × 10105.6
MF6D+EquiPose2.19 × 1074.60 × 10106.7
GRelPose1.37 × 1079.82 × 10911.8
GRelPose+EquiPose1.37 × 1071.13 × 101012.0
+ +Table 5. Comparison of parameter numbers, FLOPs, and inference times (ms) per image on an RTX 3090Ti GPU. + +# 4.4. Computation Comparison and Limitation + +Here, we compare the parameter numbers, Floating Point Operations (FLOPs) and the inference times of the baselines and their EquiPose implementations. All the models are evaluated on an RTX 3090Ti GPU on ScanNet. The results are reported in Tab. 5. As is seen, compared with the baseline model, the EquiPose implementation does not introduce any additional parameters, since it uses the same backbone as the baseline, and thus does not incur extra storage costs. Moreover, compared with each baseline model, EquiPose results in a modest increment in computational costs generally, due to the feature permutation strategy. + +However, the extra computational cost depends on the architecture of the baseline model: For models with lightweight feature aggregation modules and pose decoders, the increase in computation is slight. For models with heavy feature aggregation modules and pose decoders, the extra computational cost would be high, which is a limitation of EquiPose. Despite this, all the models under EquiPose maintain an inference time of less than 15 ms per image, which is sufficient for practical applications. + +# 5. Conclusions + +This work focuses on end-to-end relative pose estimation. Firstly, we introduce the overlooked pose permutation equivariance (PPE) property for relative pose estimation. Leveraging this property, we establish the foundational principle for designing a relative pose estimation framework that incorporates the PPE property. Building on this theoretical basis, we propose the general and flexible EquiPose framework for relative pose estimation. Experiments on four public datasets further demonstrate that the performances of pre-trained models could be improved even without re-training, and moreover, their performances could be further boosted after being fine-tuned under EquiPose. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62376269). + +# References + +[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 2, 3, 5, 6, 7, 8, 1, 4 +[2] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 3 +[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2 +[4] Georg Bokman, Johan Edstedt, Michael Felsberg, and Fredrik Kahl. Steerers: A framework for rotation equivariant keypoint descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4885-4895, 2024. 2 +[5] Jane Bromley, Isabelle Guyon, Yann LeCun, and et al. Signature verification using a "siamese" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993. 3, 5 +[6] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14566–14575, 2021. 1, 3, 5, 6, 7 +[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 1 +[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 3, 6 +[9] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 3, 5 +[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1 +[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2, 6, 1, 3 +[12] Shay Dekel and Yosi Keller. Estimating extreme 3d image rotation with transformer cross-attention. arXiv preprint arXiv:2303.02615, 2023. 1, 3, 5 + +[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition workshops, pages 224-236, 2018. 1 +[14] Qiulei Dong, Xiang Gao, Hainan Cui, and Zhanyi Hu. Robust camera translation estimation via rank enforcement. IEEE Transactions on Cybernetics, 52(2):862-872, 2020. 1 +[15] Qulei Dong, Shuang Deng, and Yuzhen Liu. Cayley rotation averaging: Multiple camera averaging under the cayley framework. IEEE Transactions on Image Processing, 2024. 1 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3 +[18] Johan Edstedt, Georg Bökman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. In 2024 International Conference on 3D Vision, pages 148–157. IEEE, 2024. 3 +[19] Sovann En, Alexis Lechery, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision Workshops, 2018. 1, 3, 5 +[20] Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and ransac's role. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8935-8943, 2022. 1, 3 +[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2 +[22] Richard Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In CVPR 2011, pages 3041-3048. IEEE, 2011. 3 +[23] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hongdong Li. Rotation averaging. International Journal of Computer Vision, 103:267-305, 2013. 3, 1 +[24] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(6):580-593, 1997. 1, 2, 3 +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1 +[26] Will Hutchcroft, Yuguang Li, Ivaylo Boyadzhiev, Zhiqiang Wan, Haiyan Wang, and Sing Bing Kang. Covispose: Covisibility pose transformer for wide-baseline relative pose estimation in 360 indoor panoramicas. In European Conference on Computer Vision, pages 615-633. Springer, 2022. 3, 5 + +[27] Nicolas Keriven and Gabriel Peyre. Universal invariant and equivariant graph neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2 +[28] Fadi Khatib, Yuval Margalit, Meirav Galun, and Ronen Basri. Leveraging image matching toward end-to-end relative camera pose regression, 2024. 1, 2, 3, 6, 7, 8, 4, 5 +[29] Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Deep Learning Workshop, 2015. 5 +[30] Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International Conference on Machine Learning, pages 2747-2755. PMLR, 2018. 3 +[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 3, 6 +[32] Yuzhen Liu and Qiulei Dong. Descriptor distillation: A teacher-student-regularized framework for learning local descriptors. International Journal of Computer Vision, pages 1-19, 2024. 3 +[33] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 6 +[34] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3 +[35] Clare Lyle, Mark van der Wilk, Marta Kwiatkowska, Yarin Gal, and Benjamin Bloem-Reddy. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020. 2 +[36] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional neural networks. In International Conference on Advanced Concepts for Intelligent Vision Systems, pages 675-687. Springer, 2017. 1, 3, 5 +[37] Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. In Advances in Neural Information Processing Systems, 2017. 3 +[38] Maher Moakher. Means and averaging in the group of rotations. SIAM journal on matrix analysis and applications, 24 (1):1-16, 2002. 4 +[39] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 2 +[40] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mokolajczyk. Ood-pose: camera pose regression from out-of-distribution synthetic views. In International Conference on 3D Vision, pages 722-732. IEEE, 2022. 1 +[41] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 6 + +[42] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 1 +[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 2 +[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 3 +[45] Jerome Revaud, Yohann Cabon, Romain Brégier, JongMin Lee, and Philippe Weinzaepfel. Sacreg: Scene-agnostic coordinate regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 688–698, 2024. 1 +[46] Chris Rockwell, Justin Johnson, and David F. Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision, 2022. 1, 3, 5, 6, 7, 8, 2, 4 +[47] Chris Rockwell, Nilesh Kulkarni, Linyi Jin, Jeong Joon Park, Justin Johnson, and David F Fouhey. Far: Flexible accurate and robust 6 dof relative camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19854-19864, 2024. 3 +[48] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 2, 3, 6, 1 +[49] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 6, 7, 1 +[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 2, 3, 6, 1 +[51] Yurun Tian, Axel Barroso-Laguna, Tony Ng, Vassileios Balntas, and Krystian Mikolajczyk. Hynet: Learning local descriptor with hybrid similarity measure and triplet loss. In Advances in Neural Information Processing Systems, pages 7401-7412, 2020. 3 +[52] Julien Valentin, Angela Dai, Matthias Nießner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In International Conference on 3D Vision, pages 323-332. IEEE, 2016. 6, 7, 1 +[53] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1 +[54] Guangyu Wang, Jinzhi Zhang, Fan Wang, Ruqi Huang, and Lu Fang. Xscale-nvs: Cross-scale novel view synthesis with + +hash featurized manifold. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21029-21039, 2024. 1 +[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Asian Conference on Computer Vision, 2022. 6 +[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 6 +[57] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 1 +[58] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual u-net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018. 1, 2 +[59] Xiaoming Zhao, Xingming Wu, Jinyu Miao, Weihai Chen, Peter CY Chen, and Zhengguo Li. Alike: Accurate and lightweight keypoint detection and descriptor extraction. IEEE Transactions on Multimedia, 2022. 3 +[60] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4669-4678, 2021. 3 +[61] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1851-1858, 2017. 1 +[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 3, 6, 7, 1 \ No newline at end of file diff --git a/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/images.zip b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ea2750d175940810a754e968f1a333ad3a8c039a --- /dev/null +++ b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a5bfe590314df704b026299f99904c2098257d24e9077dc5ba0acf42af733f5 +size 600947 diff --git a/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/layout.json b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..673c6af462682fc0cc934408e803fe8abad0c5b9 --- /dev/null +++ b/2025/EquiPose_ Exploiting Permutation Equivariance for Relative Camera Pose Estimation/layout.json @@ -0,0 +1,11194 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 103, + 455, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 103, + 455, + 138 + ], + "spans": [ + { + "bbox": [ + 156, + 103, + 455, + 138 + ], + "type": "text", + "content": "EquiPose: Exploiting Permutation Equivariance for Relative Camera Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 160, + 387, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 160, + 387, + 176 + ], + "spans": [ + { + "bbox": [ + 227, + 160, + 387, + 176 + ], + "type": "text", + "content": "Yuzhen Liu1,2 Qialei Dong1,2,*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 177, + 531, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 177, + 531, + 217 + ], + "spans": [ + { + "bbox": [ + 79, + 177, + 531, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 177, + 531, + 217 + ], + "type": "text", + "content": "State Key Laboratory of Multimodal Artificial Intelligence Systems, Chinese Academy of Sciences, Institute of Automation " + }, + { + "bbox": [ + 79, + 177, + 531, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 177, + 531, + 217 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences liuyuzhen2022@ia.ac.cn, qldong@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "spans": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 273, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 296, + 585 + ], + "type": "text", + "content": "Relative camera pose estimation between two images is a fundamental task in 3D computer vision. Recently, many relative pose estimation networks have been explored for learning a mapping from two input images to their corresponding relative pose, however, the estimated relative poses by these methods do not have the intrinsic Pose Permutation Equivariance (PPE) property: the estimated relative pose from Image A to Image B should be the inverse of that from Image B to Image A. It means that permuting the input order of two images would cause these methods to obtain inconsistent relative poses. To address this problem, we firstly introduce the concept of PPE mapping, which indicates such a mapping that captures the intrinsic PPE property of relative poses. Then by enforcing the aforementioned PPE property, we propose a general framework for relative pose estimation, called EquiPose, which could easily accommodate various relative pose estimation networks in literature as its baseline models. We further theoretically prove that the proposed EquiPose framework could guarantee that its obtained mapping is a PPE mapping. Given a pre-trained baseline model, the proposed EquiPose framework could improve its performance even without fine-tuning, and could further boost its performance with fine-tuning. Experimental results on four public datasets demonstrate that EquiPose could significantly improve the performances of various state-of-the-art models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 616, + 137, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 137, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 137, + 629 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 637, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 696 + ], + "type": "text", + "content": "Estimating the relative camera pose between an image pair is one of the fundamental tasks in 3D computer vision [6, 46]. It has a broad range of applications, such as structure-from-motion (SfM) [14, 15, 42], visual localization [10, 40, 45], novel view synthesis [7, 53, 54], etc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 248, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 429 + ], + "type": "text", + "content": "Existing methods for relative camera pose estimation in literature could be roughly divided into two categories: matching-based methods and end-to-end methods. The matching-based methods [13, 33] generally employ a two-stage computational strategy: Firstly, they establish the point correspondences between an input pair of images by utilizing local feature extraction and matching techniques [13, 33]. Then according to the established correspondences, they solve a two-view geometric problem via a typical 5-point algorithm [24] with robust fitting [21] for pursuing the relative pose. However, since the performances of the matching-based methods depend on the established point correspondences, their performance would significantly decrease or even fail in cases of weak/repeated textures and degenerated configurations as indicated in [20]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 429, + 556, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 526 + ], + "type": "text", + "content": "Unlike the above matching-based methods, many recent methods [6, 12, 19, 28, 36, 46, 57, 61] have employed deep neural networks to directly learn a mapping from the input two images to their relative pose without explicitly establishing point correspondences. Here, it has to be pointed out that ideally, the estimated relative pose by these end-to-end methods should have the following intrinsic pose permutation equivariance (PPE) property between an image pair:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 330, + 535, + 468, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 535, + 468, + 547 + ], + "spans": [ + { + "bbox": [ + 330, + 535, + 468, + 547 + ], + "type": "text", + "content": "Pose Permutation Equivariance" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "spans": [ + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "inline_equation", + "content": "\\{I_1, I_2\\}" + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": " be a pair of images, then the relative pose " + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": " should be the inverse of the relative pose " + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "inline_equation", + "content": "P_{21}" + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 328, + 556, + 539, + 593 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 408, + 601, + 539, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 601, + 539, + 616 + ], + "spans": [ + { + "bbox": [ + 408, + 601, + 539, + 616 + ], + "type": "interline_equation", + "content": "P _ {2 1} = P _ {1 2} ^ {- 1}. \\tag {1}", + "image_path": "8c102165de8ba9f70a8f04358fc6d85300b323bd7d60936f7aff32c257cf0264.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 631, + 503, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 631, + 503, + 643 + ], + "spans": [ + { + "bbox": [ + 326, + 631, + 503, + 643 + ], + "type": "text", + "content": "By writing " + }, + { + "bbox": [ + 326, + 631, + 503, + 643 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 326, + 631, + 503, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 326, + 631, + 503, + 643 + ], + "type": "inline_equation", + "content": "P_{21}" + }, + { + "bbox": [ + 326, + 631, + 503, + 643 + ], + "type": "text", + "content": " in the matrix form:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 653, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 653, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 345, + 653, + 555, + 681 + ], + "type": "interline_equation", + "content": "P _ {1 2} = \\left[ \\begin{array}{c c} R _ {1 2} & t _ {1 2} \\\\ 0 & 1 \\end{array} \\right], \\quad P _ {2 1} = \\left[ \\begin{array}{c c} R _ {2 1} & t _ {2 1} \\\\ 0 & 1 \\end{array} \\right], \\qquad (2)", + "image_path": "5bfe885511fb2ba55b9365af9de95e7fd02bf7eb6868b18c82c2fd98fe4be4a9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "R_{12}" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "t_{12}" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " are the relative rotation and translation from " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "R_{21}" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "t_{21}" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " are the relative rotation and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1127" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 191, + 160 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 191, + 160 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 191, + 160 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 191, + 160 + ], + "type": "image", + "image_path": "95d0c77578704945fac8104711d83f639038b1d61035d9c8a87370e9d4843937.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 161, + 208, + 170 + ], + "lines": [ + { + "bbox": [ + 175, + 161, + 208, + 170 + ], + "spans": [ + { + "bbox": [ + 175, + 161, + 208, + 170 + ], + "type": "text", + "content": "(a) MF6D" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 192, + 70, + 302, + 160 + ], + "blocks": [ + { + "bbox": [ + 192, + 70, + 302, + 160 + ], + "lines": [ + { + "bbox": [ + 192, + 70, + 302, + 160 + ], + "spans": [ + { + "bbox": [ + 192, + 70, + 302, + 160 + ], + "type": "image", + "image_path": "0a1e6494f3bbda22b17b6a21ccb77bac05e18b6b496cdc5dd3e7f27567f80a1f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 170, + 484, + 182 + ], + "lines": [ + { + "bbox": [ + 125, + 170, + 484, + 182 + ], + "spans": [ + { + "bbox": [ + 125, + 170, + 484, + 182 + ], + "type": "text", + "content": "Figure 1. Distribution of the " + }, + { + "bbox": [ + 125, + 170, + 484, + 182 + ], + "type": "inline_equation", + "content": "d_{R}" + }, + { + "bbox": [ + 125, + 170, + 484, + 182 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 125, + 170, + 484, + 182 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 125, + 170, + 484, + 182 + ], + "type": "text", + "content": " obtained by existing models MF6D [1] and GRelPose [28]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 416, + 160 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 416, + 160 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 416, + 160 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 416, + 160 + ], + "type": "image", + "image_path": "7177829fded36f4b5fde7718d37c7cf190ca49c4d3791bae9c79c12a40577087.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 417, + 71, + 528, + 161 + ], + "blocks": [ + { + "bbox": [ + 417, + 71, + 528, + 161 + ], + "lines": [ + { + "bbox": [ + 417, + 71, + 528, + 161 + ], + "spans": [ + { + "bbox": [ + 417, + 71, + 528, + 161 + ], + "type": "image", + "image_path": "948ac687236a6619f2a1ca5c04e11d47a6ca809eefdc519933841743e0ec182a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 161, + 439, + 170 + ], + "lines": [ + { + "bbox": [ + 394, + 161, + 439, + 170 + ], + "spans": [ + { + "bbox": [ + 394, + 161, + 439, + 170 + ], + "type": "text", + "content": "(b) GRelPose" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 202, + 293, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 293, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 293, + 215 + ], + "type": "text", + "content": "translation from " + }, + { + "bbox": [ + 55, + 202, + 293, + 215 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 55, + 202, + 293, + 215 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 202, + 293, + 215 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 55, + 202, + 293, + 215 + ], + "type": "text", + "content": ", Eq. (1) could be reformulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 220, + 295, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 220, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 111, + 220, + 295, + 235 + ], + "type": "interline_equation", + "content": "R _ {2 1} = R _ {1 2} ^ {\\mathrm {T}}, \\quad t _ {2 1} = - R _ {1 2} ^ {\\mathrm {T}} t _ {1 2}. \\tag {3}", + "image_path": "688e6e3f17e84fd37e2bd996b34b1eeac89739ad7fcc3b60aa3a65e36b5b6ebf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": "However, the PPE property is generally overlooked when designing model architectures in existing end-to-end methods [1, 28]. We validate this by conducting the following experiment: we trained two typical end-to-end models (MF6D [1] and GRelPose [28]) on the training set of ScanNet [11]. Then, we adopt the test set of ScanNet [48, 50] which contains 1500 image pairs, and feed them with two different orders to the above models respectively, leading to totally 1500 pairs of the predicted relative poses by each model, where each pair consists of the estimated pose " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " and the estimated pose " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "P_{21}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": ". Next, we compute the differences between the estimated relative pose " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "P_{21}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "P_{12}^{-1}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": ". Specifically, we compute the angular difference " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "d_{R}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " between relative rotations " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "R_{21}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "R_{12}^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": ", and the Euclidean distance " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " between relative translations " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "t_{21}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "inline_equation", + "content": "-R_{12}^{\\mathrm{T}} t_{12}" + }, + { + "bbox": [ + 55, + 240, + 296, + 433 + ], + "type": "text", + "content": " respectively via:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 438, + 295, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 438, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 100, + 438, + 295, + 457 + ], + "type": "interline_equation", + "content": "d _ {R} = \\arccos ((\\operatorname {t r} (R _ {2 1} R _ {1 2}) - 1) / 2), \\tag {4}", + "image_path": "845ef6d03ac76a03a3967345685862c12ab643368665aec490fce5c4e205231b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 454, + 199, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 454, + 199, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 199, + 468 + ], + "type": "interline_equation", + "content": "d _ {t} = \\left\\| t _ {2 1} + R _ {1 2} ^ {\\mathrm {T}} t _ {1 2} \\right\\| _ {2},", + "image_path": "306bedb1ac1c3e571653d9f430f7bb16b97a17472c2d31d16f151c5aec17cc0c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{tr}(\\cdot)" + }, + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "text", + "content": " denotes matrix trace, " + }, + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 474, + 295, + 486 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "text", + "content": "The distributions of the rotation difference " + }, + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "inline_equation", + "content": "d_{R}" + }, + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "text", + "content": " and translation difference " + }, + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "text", + "content": " from the above two models are shown in Fig. 1. As seen from this figure, the pose differences (both rotation and translation) estimated by these models are generally larger than zero (particularly some of them are not trivial), demonstrating that the above models do not satisfy the PPE property. Such an issue may pose risks to downstream applications that require a stable and consistent estimation of relative poses, such as robotics and visual navigation. Moreover, it is also noted that some other visual tasks involve their specific equivariances, different from the PPE property in the relative pose estimation task, and once an existing model among these tasks is equipped with a task-specific equivariance, its performance could be further boosted [4, 27, 35, 39, 43]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Motivated by the above issues, we investigate whether it is possible to explore such a relative pose estimation framework that could not only bestow the PPE property on various relative pose estimation networks but also improve their" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 202, + 553, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 309 + ], + "type": "text", + "content": "performances further. Specifically, we first introduce the concept of PPE mapping, which refers to a mapping that captures the PPE property in relative poses. Then, we establish the foundational principle for equivariant relative pose estimation. Accordingly, we propose the EquiPose framework which holds the PPE property theoretically. The EquiPose could improve the performance of a pre-trained model even without fine-tuning, and fine-tuning could enhance their performances further." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 311, + 516, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 311, + 516, + 321 + ], + "spans": [ + { + "bbox": [ + 325, + 311, + 516, + 321 + ], + "type": "text", + "content": "To summarize, our main contributions include:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 323, + 554, + 538 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "type": "text", + "content": "- Equivariance: The simple but intrinsic pose permutation equivariance (PPE) property is revealed for relative pose estimation, and the PPE mapping is defined." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 358, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 358, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 358, + 554, + 454 + ], + "type": "text", + "content": "- Generality: By imposing the PPE property, the framework EquiPose for equivariant relative pose estimation is proposed. An arbitrary end-to-end relative pose estimation model could be seamlessly embedded into this framework as a baseline, and accordingly obtains the PPE property as analyzed in theory in Sec. 3. The EquiPose framework and the theoretical analyses might give new insights into model design strategies for relative pose estimation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 454, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 554, + 538 + ], + "type": "text", + "content": "- Flexibility: Thanks to the imposed PPE property, the performances of the derived methods from the EquiPose framework by utilizing many off-the-shelf baseline models, would be boosted even without fine-tuning by adding several lines of code. Moreover, once they are fine-tuned under EquiPose, their performances would be further improved, as demonstrated by the results in Sec. 4." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 549, + 402, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 402, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 402, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "type": "text", + "content": "In this section, we review some typical matching-based methods and end-to-end methods for relative pose estimation respectively." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Matching-Based Relative Pose Estimation. The matching-based relative pose estimation methods typically involve the following two steps: establishing point correspondence via image matching [3, 33] and solving the two-view geometry with RANSAC [21, 24, 41]. Early approaches relied on handcrafted keypoint detectors and descriptors to establish point correspondences, such as SIFT [33] and SURF [3]. In recent years, numerous methods emerged that replace the handcrafted detectors or descrip" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1128" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "tors with their learned counterparts [2, 17, 18, 32, 34, 37, 44, 51, 59]. Additionally, a variety of learning-based image-matching methods directly established point correspondences from keypoints or raw images [8, 31, 48, 50, 60]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 204 + ], + "type": "text", + "content": "However, since these methods depend on point correspondences, they may fail in cases of weak textures, repeated textures, large viewpoint changes, and degenerated configurations [20]. Additionally, these methods could be time-consuming due to image matching and RANSAC loops. Moreover, they could only recover the relative pose up to an unknown scale factor [24]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 347 + ], + "type": "text", + "content": "End-to-End Relative Pose Estimation. Recently, unlike the above strategies, many works have adopted deep neural networks to regress relative poses in an end-to-end manner [1, 6, 9, 12, 19, 28, 36]. These methods do not explicitly implement image matching or solve the two-view geometry problem. Instead, the networks directly learn the underlying mapping from the input image pair to the corresponding relative pose. Compared with matching-based methods, they are shown to be more robust in challenging scenes [6, 9, 19] with a significantly faster inference speed. Moreover, these end-to-end methods could be further incorporated with traditional methods for improving robustness [47]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 538 + ], + "type": "text", + "content": "These end-to-end methods generally adopt Siamese architectures [5] to extract feature, which are aggregated and decoded into the relative pose subsequently [6, 12, 26, 28, 46]. To facilitate stable training, early approaches [19, 36] parameterized the rotation as quaternions. Different from them, Zhou et al. [62] suggested to represent 3D rotations in a 6D space, which benefits the training of neural networks. Cai et al. [6] and Dekel et al. [12] showed that the network could learn from implicit cues, allowing the network to estimate the relative pose even when two images have no overlap. Rockwell et al. [46] introduced the essential matrix module to emulate the 8-point algorithm [24]. Arnold et al. [1] adopted a differentiable warping module for feature aggregation. Khatib et al. [28] embedded a semidense image matcher [50] for feature extraction to pursue a better generalization ability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": "However, the existing end-to-end methods generally do not satisfy the PPE property, leading to inconsistent estimation when permuting the input order of an image pair as shown in Sec. 1. Unlike them, this work does not aim to propose a new model for relative pose estimation, but pursue a general framework that could bestow the PPE property on various relative pose estimation models, and improve their performances further." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 646, + 138, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 138, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 138, + 659 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In this section, we first introduce the key definitions and propositions related to the pose permutation equivariance (PPE) property of relative poses. Then, we describe the proposed EquiPose framework in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 538, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 538, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 538, + 85 + ], + "type": "text", + "content": "3.1. PPE Mapping and Preliminary Knowledge" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 116 + ], + "type": "text", + "content": "Prior to the introduction of the PPE mapping, 'equivariant mapping' in mathematics is defined as [30]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": "Definition 1 (Equivariant mapping). Let " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " be a group that acts on spaces " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathbb{T}'" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " be the induced actions of " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": ". A mapping " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " is equivariant with the action in " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "inline_equation", + "content": "f(\\mathbb{T}_g(x)) = \\mathbb{T}_g'(f(x)), \\forall x \\in \\mathcal{X}, g \\in \\mathcal{G}" + }, + { + "bbox": [ + 313, + 125, + 555, + 185 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 194, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 219 + ], + "type": "text", + "content": "According to Definition 1 and Eq. (1), a PPE mapping with the PPE property is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 277 + ], + "type": "text", + "content": "Definition 2 (Pose permutation equivariant mapping). A mapping " + }, + { + "bbox": [ + 313, + 228, + 555, + 277 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^{H \\times W} \\times \\mathbb{R}^{H \\times W} \\to \\mathrm{SE}(3)" + }, + { + "bbox": [ + 313, + 228, + 555, + 277 + ], + "type": "text", + "content": " (the special Euclidean group) is pose permutation equivariant, if " + }, + { + "bbox": [ + 313, + 228, + 555, + 277 + ], + "type": "inline_equation", + "content": "f(I_2, I_1) = f(I_1, I_2)^{-1}, \\forall I_1, I_2 \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 228, + 555, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "text", + "content": "Definition 2 could be considered as an instance of 'equivariant mapping' from Definition 1 on relative camera poses, where the input variable " + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "text", + "content": " is an image pair " + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\{I_1, I_2\\}" + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "text", + "content": ", the action " + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_g" + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\{I_1, I_2\\}" + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "text", + "content": " is an order permutation operator, and the action " + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_q'" + }, + { + "bbox": [ + 313, + 286, + 554, + 346 + ], + "type": "text", + "content": " is the inverse of a matrix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 347, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 554, + 384 + ], + "type": "text", + "content": "It is noted that Definition 2 is applicable to neural networks, since any neural networks essentially aim to learn a mapping from their input to the output." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 384, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 455 + ], + "type": "text", + "content": "In addition to the above two definitions, we introduce the following definition on the geodesic mean of rotations [22, 23], as well as two theorems, which would be used to support the fundamental principle for designing a relative pose estimation framework with the PPE property in the next subsection:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 465, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 555, + 491 + ], + "type": "text", + "content": "Definition 3 (Geodesic mean of rotations). Given a set of rotations " + }, + { + "bbox": [ + 313, + 465, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\{R_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 465, + 555, + 491 + ], + "type": "text", + "content": ", their geodesic mean " + }, + { + "bbox": [ + 313, + 465, + 555, + 491 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 465, + 555, + 491 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 500, + 555, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 500, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 322, + 500, + 555, + 532 + ], + "type": "interline_equation", + "content": "S = \\mathfrak {G} \\left(R _ {1}, \\dots , R _ {n}\\right) = \\underset {S \\in \\mathrm {S O} (3)} {\\arg \\min } \\sum_ {i = 1} ^ {n} \\| \\log \\left(R _ {i} ^ {\\mathrm {T}} S\\right) \\| _ {2} ^ {2}, \\tag {5}", + "image_path": "401a26e494b5331cbd34ee504fd93cfa697855022e54104eec0fa3284d794d47.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 541, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 541, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\log (\\cdot)" + }, + { + "bbox": [ + 313, + 541, + 554, + 566 + ], + "type": "text", + "content": " is the mapping from an element in the Lie group to its corresponding Lie algebra." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 575, + 523, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 575, + 523, + 586 + ], + "spans": [ + { + "bbox": [ + 325, + 575, + 523, + 586 + ], + "type": "text", + "content": "From Definition 3, we have the two theorems as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "text", + "content": "Theorem 1. Given a set of rotations " + }, + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "inline_equation", + "content": "\\{R_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "text", + "content": ", it always holds that: " + }, + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "inline_equation", + "content": "\\mathfrak{G}(R_{\\sigma(1)}, \\dots, R_{\\sigma(n)}) = \\mathfrak{G}(R_1, \\dots, R_n)" + }, + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "text", + "content": " for any permutation " + }, + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 597, + 555, + 634 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "text", + "content": "Theorem 2. If " + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "inline_equation", + "content": "S \\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "text", + "content": " is the geodesic mean of the set of rotations " + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\{R_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "text", + "content": " is the geodesic mean of the set of rotations " + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\{R_i^{\\mathrm{T}}\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 643, + 554, + 681 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Due to the space limitation, the proofs of the above two theorems are given in the supplementary material." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1129" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 289, + 144 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 289, + 144 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 289, + 144 + ], + "type": "image", + "image_path": "93c2cc69918823371827c0cec9501ef7356e583cfb66fa72da7c1d62d2638db8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "text", + "content": "Figure 2. Illustration of the existing methods " + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "text", + "content": " and our method " + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "text", + "content": " on the SE(3) manifold. Left: The existing method " + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "text", + "content": " does not capture the PPE property in the relative pose. Right: Within EquiPose " + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 55, + 146, + 295, + 200 + ], + "type": "text", + "content": ", the desired PPE property is satisfied (Theorem 3), and the results could be more reliable (Sec. 4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 218, + 295, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 243 + ], + "type": "text", + "content": "3.2. Fundamental Principle of Pose Estimation Framework with PPE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 320 + ], + "type": "text", + "content": "According to the aforementioned PPE mapping and related preliminary knowledge, we elaborate the fundamental principle for designing a pose estimation framework with PPE, which could use an arbitrary end-to-end relative pose estimation model as its baseline and derive the corresponding upgraded model with the PPE property as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 320, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 296, + 357 + ], + "type": "text", + "content": "Given an arbitrary relative pose estimation network corresponding to a mapping " + }, + { + "bbox": [ + 55, + 320, + 296, + 357 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 320, + 296, + 357 + ], + "type": "text", + "content": ", there exists the following mapping " + }, + { + "bbox": [ + 55, + 320, + 296, + 357 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 55, + 320, + 296, + 357 + ], + "type": "text", + "content": ", which is a PPE mapping:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "text", + "content": "Theorem 3. Given an arbitrary mapping " + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^{H \\times W} \\times \\mathbb{R}^{H \\times W} \\to \\mathrm{SE}(3)" + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "inline_equation", + "content": "f(I_1, I_2) := (R_{12}, t_{12})" + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "inline_equation", + "content": "f(I_2, I_1) := (R_{21}, t_{21})" + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "text", + "content": ", the mapping " + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "inline_equation", + "content": "f^*" + }, + { + "bbox": [ + 55, + 361, + 296, + 399 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 123, + 403, + 294, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 403, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 123, + 403, + 294, + 418 + ], + "type": "interline_equation", + "content": "f ^ {*} \\left(I _ {1}, I _ {2}\\right) := \\left(R _ {1 2} ^ {*}, t _ {1 2} ^ {*}\\right), \\tag {6}", + "image_path": "4ff0ce08510b350fc0af20c7bf6c6b41ee4026f27f79a990cbe06d61d50a520e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 423, + 84, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 84, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 84, + 433 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 440, + 295, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 440, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 123, + 440, + 295, + 454 + ], + "type": "interline_equation", + "content": "R _ {1 2} ^ {*} = \\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right), \\tag {7}", + "image_path": "65115a10df6dc1364168f398e3f5688748eb8b1cee2f5f4a8b0f748a953517b4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 456, + 295, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 456, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 127, + 456, + 295, + 469 + ], + "type": "interline_equation", + "content": "t _ {1 2} ^ {*} = \\left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\\right) / 2, \\tag {8}", + "image_path": "65eda5362dcc973fe03b0cfbb5c07c212565c9692f18c58df1e0bb600b615cbb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 475, + 132, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 132, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 132, + 487 + ], + "type": "text", + "content": "is a PPE mapping." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 492, + 225, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 225, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 225, + 504 + ], + "type": "text", + "content": "Proof. According to Theorem 1, we have" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 100, + 514, + 295, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 514, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 100, + 514, + 295, + 529 + ], + "type": "interline_equation", + "content": "R _ {1 2} ^ {*} = \\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right) = \\mathfrak {G} \\left(R _ {2 1} ^ {\\mathrm {T}}, R _ {1 2}\\right) \\tag {9}", + "image_path": "aabb7d68897e5970d03ddf98953573f7c4ff73813554f87076c8cca86385ca01.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 533, + 233, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 533, + 233, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 533, + 233, + 544 + ], + "type": "text", + "content": "Then, according to Theorem 2, we have:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 96, + 548, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 548, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 96, + 548, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\mathfrak {G} \\left(R _ {2 1} ^ {\\mathrm {T}}, R _ {1 2}\\right) = \\mathfrak {G} \\left(R _ {2 1}, R _ {1 2} ^ {\\mathrm {T}}\\right) ^ {\\mathrm {T}} = R _ {2 1} ^ {* \\mathrm {T}} \\tag {10}", + "image_path": "e8ce3d566a9af6820bd55b746b1eecb0f2351aa5cf506ad9fc061df1eb64c9bf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 570, + 230, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 230, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 230, + 581 + ], + "type": "text", + "content": "According to Eqs. (9) and (10), it holds:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 149, + 586, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 586, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 149, + 586, + 295, + 601 + ], + "type": "interline_equation", + "content": "R _ {1 2} ^ {*} = R _ {2 1} ^ {* \\mathrm {T}}, \\tag {11}", + "image_path": "288da6da691c1c35dec7a1d589e37c3a6daa306e84cbc4d70ca1be2c0f874999.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 606, + 274, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 274, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 274, + 618 + ], + "type": "text", + "content": "i.e., " + }, + { + "bbox": [ + 55, + 606, + 274, + 618 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 55, + 606, + 274, + 618 + ], + "type": "text", + "content": " is permutation equivariant in terms of rotation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 618, + 264, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 618, + 264, + 630 + ], + "spans": [ + { + "bbox": [ + 67, + 618, + 264, + 630 + ], + "type": "text", + "content": "Now, according to (6), (7), (8) and (11), it holds:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 70, + 635, + 294, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 635, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 70, + 635, + 294, + 696 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f ^ {*} (I _ {1}, I _ {2}) ^ {- 1} = \\left(R _ {1 2} ^ {* \\mathrm {T}}, - R _ {1 2} ^ {* \\mathrm {T}} t _ {1 2} ^ {*}\\right) \\\\ = \\left(R _ {1 2} ^ {* \\mathrm {T}}, - R _ {1 2} ^ {* \\mathrm {T}} \\left(t _ {1 2} - R _ {1 2} ^ {*} t _ {2 1}\\right) / 2\\right) \\tag {12} \\\\ = \\left(R _ {2 1} ^ {*}, \\left(t _ {2 1} - R _ {2 1} ^ {*} t _ {1 2}\\right) / 2\\right) \\\\ = f ^ {*} \\left(I _ {2}, I _ {1}\\right), \\\\ \\end{array}", + "image_path": "d5684866cc1666f5bd92167a6d6dabfb0834f04965080514d32c01ab9fcb1ee2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 701, + 167, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 167, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 167, + 713 + ], + "type": "text", + "content": "which completes the proof." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 285, + 702, + 294, + 711 + ], + "blocks": [ + { + "bbox": [ + 285, + 702, + 294, + 711 + ], + "lines": [ + { + "bbox": [ + 285, + 702, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 285, + 702, + 294, + 711 + ], + "type": "image", + "image_path": "da3249d170907e792178917639d91d45072a5375612001bac512854000cab9e5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 72, + 476, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 476, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 476, + 84 + ], + "type": "text", + "content": "Algorithm 1 The EquiPose Framework" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 87, + 503, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 87, + 503, + 112 + ], + "spans": [ + { + "bbox": [ + 317, + 87, + 503, + 112 + ], + "type": "text", + "content": "Input: A relative pose estimation network " + }, + { + "bbox": [ + 317, + 87, + 503, + 112 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 317, + 87, + 503, + 112 + ], + "type": "text", + "content": "; A pair of images " + }, + { + "bbox": [ + 317, + 87, + 503, + 112 + ], + "type": "inline_equation", + "content": "\\{I_1, I_2\\}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 112, + 504, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 112, + 504, + 124 + ], + "spans": [ + { + "bbox": [ + 317, + 112, + 504, + 124 + ], + "type": "text", + "content": "Output: Equivariant pose estimation " + }, + { + "bbox": [ + 317, + 112, + 504, + 124 + ], + "type": "inline_equation", + "content": "R_{12}^{*}, t_{12}^{*}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 124, + 472, + 171 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 321, + 124, + 417, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 124, + 417, + 135 + ], + "spans": [ + { + "bbox": [ + 321, + 124, + 417, + 135 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 321, + 124, + 417, + 135 + ], + "type": "inline_equation", + "content": "R_{12}, t_{12} \\gets f(I_1, I_2)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 136, + 417, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 136, + 417, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 136, + 417, + 148 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 136, + 417, + 148 + ], + "type": "inline_equation", + "content": "R_{21}, t_{21} \\gets f(I_2, I_1)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 147, + 472, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 147, + 472, + 160 + ], + "spans": [ + { + "bbox": [ + 321, + 147, + 472, + 160 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 147, + 472, + 160 + ], + "type": "inline_equation", + "content": "R_{12}^{*} \\gets R_{12}(\\exp \\left(\\frac{1}{2}\\log (R_{12}^{\\mathrm{T}}R_{21}^{\\mathrm{T}})\\right))" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 160, + 432, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 160, + 432, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 160, + 432, + 171 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 321, + 160, + 432, + 171 + ], + "type": "inline_equation", + "content": "t_{12}^{*}\\gets (t_{12} - R_{12}^{*}t_{21}) / 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "content": "For ease of understanding, we visualize the difference between " + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "content": " in Fig. 2. As seen from this figure, given two input images " + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\{I_1, I_2\\}" + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "content": ", the existing relative pose estimation model (considered as a mapping) " + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "content": " does not satisfy the PPE property. Once it is embedded into EquiPose, the upgrade version " + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "content": " captures the PPE property strictly. Moreover, its estimation could be more reliable than the baseline model " + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 196, + 553, + 304 + ], + "type": "text", + "content": ", as demonstrated in Sec. 4 (also in Sec. S8.3 of the supplementary material)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 304, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 340 + ], + "type": "text", + "content": "It is noted in Eq. (7) that the construction of " + }, + { + "bbox": [ + 313, + 304, + 554, + 340 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 313, + 304, + 554, + 340 + ], + "type": "text", + "content": " involves computing the geodesic mean of two rotations, which could be computed explicitly via [38]:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 364, + 348, + 553, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 348, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 364, + 348, + 553, + 364 + ], + "type": "interline_equation", + "content": "\\mathfrak {G} \\left(R _ {1 2}, R _ {2 1} ^ {\\mathrm {T}}\\right) = R _ {1 2} \\left(R _ {1 2} ^ {\\mathrm {T}} R _ {2 1} ^ {\\mathrm {T}}\\right) ^ {1 / 2}. \\tag {13}", + "image_path": "89600cf95e27212522ef085d6769d127a1d6726ef58f83edea3ffde77ba098d6.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 373, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 407 + ], + "type": "text", + "content": "Considering that we work on the 3D rotation group, the square root of a rotation " + }, + { + "bbox": [ + 313, + 373, + 554, + 407 + ], + "type": "inline_equation", + "content": "R \\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 313, + 373, + 554, + 407 + ], + "type": "text", + "content": " could be computed by:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 384, + 407, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 407, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 384, + 407, + 553, + 430 + ], + "type": "interline_equation", + "content": "R ^ {1 / 2} = \\exp \\left(\\frac {1}{2} \\log (R)\\right), \\tag {14}", + "image_path": "a9c739d4b0613085090a3cfb8e0737cd8f87b06a78437d95662b2cec0695f4e2.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 435, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 435, + 553, + 459 + ], + "type": "inline_equation", + "content": "\\exp (\\cdot)" + }, + { + "bbox": [ + 313, + 435, + 553, + 459 + ], + "type": "text", + "content": " is the mapping from the Lie algebra to the Lie group." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": "The above analysis reveals that given an arbitrary relative pose estimation network, once its upgrade version " + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": " is constructed by (6), the mapping is a pose permutation equivariant mapping, this is to say, the upgrade version has the PPE property." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 314, + 527, + 381, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 527, + 381, + 540 + ], + "spans": [ + { + "bbox": [ + 314, + 527, + 381, + 540 + ], + "type": "text", + "content": "3.3. EquiPose" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 545, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 582 + ], + "type": "text", + "content": "According to the revealed point above as well as Eqs. (7) and (8), a simple framework for relative pose estimation, i.e., the EquiPose, could be straightforwardly built." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "type": "text", + "content": "Specifically, the proposed EquiPose uses an existing model as its baseline, and outputs the final estimation on relative pose by calculating Eqs. (7) and (8), whose variable values are assigned by the baseline with or even without fine-tuning. Algorithm 1 shows the complete pipeline of EquiPose to obtain the relative pose." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "It has to be pointed out that the computational cost of " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " in this implementation is twice that of " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ", since it involves two times of forward propagation of " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ": one for " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "f(I_{1},I_{2})" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ", and one for " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "f(I_{2},I_{1})" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " after permuting the input images " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\{I_1,I_2\\}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ". However, in practice, the computational" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1130" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 70, + 465, + 210 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 465, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 465, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 465, + 210 + ], + "type": "image", + "image_path": "8c0b88ccccba0312ba9e2174af8dcb30830b696b3f49912ccd59909016c59abc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 105, + 225, + 504, + 494 + ], + "blocks": [ + { + "bbox": [ + 300, + 214, + 310, + 224 + ], + "lines": [ + { + "bbox": [ + 300, + 214, + 310, + 224 + ], + "spans": [ + { + "bbox": [ + 300, + 214, + 310, + 224 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 105, + 225, + 504, + 494 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 504, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 504, + 494 + ], + "type": "image", + "image_path": "53580ed6212c9472eb2b2d6d6b80de5a9714aa6f378b03440eea2dec914682eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 495, + 311, + 505 + ], + "lines": [ + { + "bbox": [ + 299, + 495, + 311, + 505 + ], + "spans": [ + { + "bbox": [ + 299, + 495, + 311, + 505 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "text", + "content": "Figure 3. The pipelines of existing methods and our method. (a) The pipeline of existing relative pose estimation networks, which consists of two stages: Stage (i) Single-Image Feature Encoding " + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{enc}})" + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "text", + "content": " extracts the feature maps " + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\{F_1,F_2\\}" + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\{I_1,I_2\\}" + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "text", + "content": ". Stage (ii) Cross-Image Feature Decoding " + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{dec}})" + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "text", + "content": " takes the extracted two feature maps " + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\{F_1,F_2\\}" + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "text", + "content": " as input, and it outputs the estimated relative pose. (b) The pipeline of EquiPose. The computational overhead from EquiPose is an additional forward propagation of " + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}" + }, + { + "bbox": [ + 54, + 509, + 555, + 555 + ], + "type": "text", + "content": ", which is slight." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 574, + 295, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 574, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 54, + 574, + 295, + 599 + ], + "type": "text", + "content": "overhead could be reduced further with the feature permutation strategy, which is elaborated in the following." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 601, + 295, + 715 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 54, + 601, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 601, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 601, + 295, + 674 + ], + "type": "text", + "content": "To begin with, it is noted that the general pipeline of existing relative pose estimation models [1, 6, 9, 12, 19, 36, 46] could be summarized in Fig. 3a. As is seen, the pipeline consists of the two stages, which we name as (i) Single-Image Feature Encoding (denoted as " + }, + { + "bbox": [ + 54, + 601, + 295, + 674 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}" + }, + { + "bbox": [ + 54, + 601, + 295, + 674 + ], + "type": "text", + "content": "), and (ii) Cross-Image Feature Decoding (denoted as " + }, + { + "bbox": [ + 54, + 601, + 295, + 674 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}" + }, + { + "bbox": [ + 54, + 601, + 295, + 674 + ], + "type": "text", + "content": ") respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "text", + "content": "(i) Single-Image Feature Encoding " + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}" + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "text", + "content": ": This stage takes two images as inputs, and it outputs the corresponding feature maps " + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{F_1,F_2\\}" + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}(I_1,I_2)\\coloneqq \\{F_1,F_2\\}" + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 574, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 635 + ], + "type": "text", + "content": "It contains a feature encoder, which could be any feature-extraction backbone in computer vision. It should be noted that the weights of the feature encoder for extracting the features of image " + }, + { + "bbox": [ + 313, + 574, + 555, + 635 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 574, + 555, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 574, + 555, + 635 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 313, + 574, + 555, + 635 + ], + "type": "text", + "content": " are shared (generally referred to as the Siamese architecture [5, 29])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "(ii) Cross-Image Feature Decoding " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ": This stage takes the feature maps " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{F_1, F_2\\}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " obtained from Stage (i) as input, and it outputs the estimated relative pose " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}(F_1, F_2) := P_{12}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ". It contains a feature aggregation module and a pose decoder. The feature aggregation module maps " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{F_1, F_2\\}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " into a single aggregated feature map (or vec" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1131" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 299, + 248 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 299, + 248 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 299, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 299, + 248 + ], + "type": "table", + "html": "
MethodRot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m
ER[6]13.1528.2350.43---14.66-
+EquiPose15.2230.4051.51---14.11-
+EquiPose†16.0031.0152.98---12.53-
Reg6D[62]10.1622.9745.76---12.26-
+EquiPose11.7625.2847.94---11.76-
+EquiPose†16.1431.5854.20---10.81-
EightVit[46]23.6940.2561.9214.9032.9558.478.900.41
+EquiPose24.5841.3262.8215.4934.1160.038.400.35
+EquiPose†25.3042.5464.0516.3736.1463.077.970.31
MF6D[1]22.3637.1957.1014.3932.5757.349.810.44
+EquiPose25.4940.5359.9417.5135.7860.498.420.35
+EquiPose†27.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]28.9046.3666.5316.4635.3660.068.160.38
+EquiPose30.7048.0267.9017.8537.1761.807.970.32
+EquiPose†32.3249.7469.3719.6740.4065.957.610.28
", + "image_path": "c72533480fe92f9e20f61a204c677300acd824b939ba62134e84380c29a7e904.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 250, + 296, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 327 + ], + "type": "text", + "content": "Table 1. Comparison on the ScanNet dataset [11, 50]. RE: rotation errors in degrees. TE: translation errors in meters. +EquiPose: the models are implemented under EquiPose only during the inference stage. +EquiPose†: the models are further fine-tuned under EquiPose. -: the models only predict the relative camera rotation and thus could not be evaluated in translation. The best methods for each baseline are marked in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 358, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 384 + ], + "type": "text", + "content": "tor), which is subsequently decoded by the pose decoder to obtain the estimated relative pose." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "text", + "content": "Since the weights of the feature encoder in " + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}" + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "text", + "content": " are shared, we do not have to implement two complete forward propagation of " + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "text", + "content": ". Instead, we could reuse the extracted feature maps " + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\{F_1, F_2\\}" + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "text", + "content": " at the first forward propagation, then permute the feature maps to " + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\{F_2, F_1\\}" + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "text", + "content": ", and implement one additional forward propagation of " + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}" + }, + { + "bbox": [ + 55, + 388, + 296, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": "Fig. 3b illustrates the EquiPose framework with the feature permutation strategy. As is seen, it first performs one forward propagation of " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " to extract feature maps " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\{F_1,F_2\\}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": ". Then, the feature maps " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\{F_1,F_2\\}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " and its permuted counterpart " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\{F_2,F_1\\}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " are inputted to " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "P_{21}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " respectively. The final relative pose is obtained via " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "P_{21}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": " through Eq. (6). Compared to the original implementations of the baseline models, EquiPose only involves an additional forward propagation of " + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 577, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 639 + ], + "type": "text", + "content": "Under the EquiPose framework, a pre-trained model could be directly adopted as the baseline without fine-tuning to boost its performance. Moreover, the model could also be further fine-tuned under EquiPose to achieve further improvement. We simply use an " + }, + { + "bbox": [ + 55, + 577, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 577, + 296, + 639 + ], + "type": "text", + "content": " loss for fine-tuning:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 656, + 295, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 656, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 108, + 656, + 295, + 671 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\| R ^ {*} - \\hat {R} \\| _ {1} + \\lambda \\| t ^ {*} - \\hat {t} \\| _ {1}, \\tag {15}", + "image_path": "48defdd97dc1b87141390d6e98f84cfface0ae05ce5c5e56c03938d1bde46282.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " are the ground truth relative rotation and translation. " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is a preset constant factor." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 547, + 215 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 547, + 215 + ], + "type": "image", + "image_path": "7de4c1b8034b8f0c163a634504aff009436fbf6c83d2c5a1d9fced443207081a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "type": "text", + "content": "Figure 4. Visualization of epipolar lines on ScanNet [11]. Five SIFT keypoints are sampled from image " + }, + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "type": "text", + "content": " (column 1) and their epipolar lines are projected onto image " + }, + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "type": "text", + "content": ", using the pose from: ground truth (column 2), GRelPose baseline [28] (column 3), EquiPose without (column 4) and with finetuning (column 5)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 391, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 391, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 391, + 306 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 311, + 491, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 491, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 491, + 323 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "text", + "content": "Datasets. We evaluate our method on both indoor and outdoor scenes. For indoor scenes, we use the ScanNet [11], the 7-Scene [49] and the 12-Scene [52] datasets, which are widely used for indoor pose estimation. For outdoor scenes, we use the recently proposed Mapfree dataset [1]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": "For ScanNet, we use 2M image pairs from the training set for training, and use the test set which contains 1500 pairs for evaluation following [8, 48, 50, 55, 56]. For both 7-Scene and 12-Scene, we use 720 image pairs for evaluation using the model trained on ScanNet. For Mapfree, we follow the official setup in [1] for training and evaluation. More details are provided in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "text", + "content": "Evaluation Metrics. For ScanNet [11], 7-Scene [49] and 12-Scene [52], we calculate the area under the cumulative curve (AUC) as done in [31, 48, 50]. We report the AUC for rotation " + }, + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "inline_equation", + "content": "(3^{\\circ} / 5^{\\circ} / 10^{\\circ})" + }, + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "inline_equation", + "content": "(0.1\\mathrm{m} / 0.2\\mathrm{m} / 0.5\\mathrm{m})" + }, + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "text", + "content": " respectively. Besides, we also report the average rotation and translation errors respectively. For the Mapfree dataset [1], we report the pose error and virtual correspondence reprojection error (VCRE) respectively following [1]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 575, + 381, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 381, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 381, + 588 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 666 + ], + "type": "text", + "content": "It should be pointed out that the goal of this study is to introduce a general framework for existing end-to-end relative pose estimation models to achieve the PPE property with improved performances. Thus, we do not further evaluate and analyze other traditional matching-based methods, such as the classic SIFT [33] followed by the 5-point solver [41]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Specifically, we adopt the following five end-to-end models as baselines and evaluate the effectiveness of EquiPose in boosting their performances: (i) ExtremeRotation [6], (ii) Reg6D [6, 62], (iii) EightVit [46], (iv)" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1132" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 542, + 261 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 542, + 261 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 542, + 261 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 542, + 261 + ], + "type": "table", + "html": "
Method7-Scene12-Scene
Rot. AUC↑Trans. AUC↑RE↓TE↓Rot. AUC↑Trans. AUC↑RE↓TE↓
@3°@5°@10°@.1m@.2m@.5m@3°@5°@10°@.1m@.2m@.5m
ER[6]11.1724.2846.70---9.82-14.6131.6057.15---6.94-
+EquiPose13.7226.6047.83---9.09-17.5734.3958.96---6.66-
+EquiPose†13.8227.1449.35---7.84-17.8335.3860.10---5.39-
Reg6D[62]10.0822.4445.92---7.78-13.1628.4255.27---5.67-
+EquiPose11.2023.7746.93---7.67-14.0530.0856.62---5.41-
+EquiPose†13.2627.3251.60---6.60-16.9933.5559.59---5.07-
EightVit[46]19.7735.0059.2513.9932.0561.114.920.2230.0950.3972.9538.0062.8084.032.890.08
+EquiPose19.8835.5159.8714.2132.9162.904.780.2031.1751.5373.5838.6763.7284.592.780.08
+EquiPose†20.8737.0161.4414.4733.7365.224.560.1932.7452.7274.4140.6365.3585.452.710.07
MF6D[1]18.7334.4657.6211.9030.3560.134.980.2330.4451.0773.1136.3761.5283.242.790.09
+EquiPose21.7837.4060.0315.4634.2863.714.710.1934.1754.2974.9840.8764.6785.022.560.08
+EquiPose†24.4440.8363.3217.1736.9066.444.290.1835.8055.3675.8243.0266.3685.782.450.07
GRelPose[28]26.7845.4767.9515.7034.2762.483.960.2135.6255.2075.6339.8763.8184.342.580.08
+EquiPose28.4446.8568.8715.8134.4163.483.880.2037.1356.6776.7142.2465.6585.212.410.07
+EquiPose†29.7648.2470.0815.8335.4364.913.660.1938.6358.1177.8644.1967.4386.042.240.07
", + "image_path": "7229146a4f9e431bbf9349cfa0c53288f322b33987069901486b563fc549daf7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 294, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 344 + ], + "type": "text", + "content": "MF6D [1], and (v) GRelPose [28]. Note that (i) ExtremeRotation and (ii) Reg6D only predict relative rotation, while the remaining ones could predict the 6 DoF relative camera pose (including both rotation and translation)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 194, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 194, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 194, + 361 + ], + "type": "text", + "content": "4.3. Comparative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "text", + "content": "Evaluation Setup. For each of the five baseline models, including ER (ExtremeRotation) [6], Reg6D [62], EightVit [46], MF6D [1] and GRelPose [28], we compare the following variants: 1) the model in its original implementation, 2) the model implemented under EquiPose only during the inference stage (denoted with " + }, + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "text", + "content": " EquiPose), 3) the model fine-tuned under EquiPose (denoted with " + }, + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "text", + "content": " EquiPose" + }, + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 366, + 295, + 450 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "content": "Indoor Pose Evaluation. The evaluation results on ScanNet are reported in Table 1. As seen from this table, the five models under the EquiPose framework achieve consistently better performance than their corresponding baselines under the eight metrics. Moreover, these models after fine-tuning under EquiPose could achieve further enhancements in most cases by a significant margin. For example, an inference-only implementation of GRelPose under EquiPose leads to a relative decrease of " + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "content": " (1-7.97/8.16) and " + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "content": " (1-0.32/0.38) in rotation and translation errors respectively. Its fine-tuned implementation has a relative reduction of " + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "inline_equation", + "content": "26.3\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "content": " in rotation and translation errors respectively. The fine-tuned MF6D could even achieve a relative reduction of " + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "inline_equation", + "content": "23.9\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "inline_equation", + "content": "34.1\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 641 + ], + "type": "text", + "content": " in rotation and translation errors respectively. These results validate the effectiveness of the proposed EquiPose framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "The cross-dataset evaluation results on the 7-Scene [49] and 12-Scene [52] datasets are reported in Tab. 2, where the models trained on ScanNet are used for evaluation. As seen from the table, the five models under EquiPose could also achieve consistent improvements than their corresponding baselines, and fine-tuning further boosts the performances." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 292, + 556, + 495 + ], + "blocks": [ + { + "bbox": [ + 79, + 262, + 530, + 273 + ], + "lines": [ + { + "bbox": [ + 79, + 262, + 530, + 273 + ], + "spans": [ + { + "bbox": [ + 79, + 262, + 530, + 273 + ], + "type": "text", + "content": "Table 2. Cross-dataset evaluation on the 7-Scene [49] and 12-Scene [52] datasets. All the models are trained on ScanNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 292, + 556, + 495 + ], + "lines": [ + { + "bbox": [ + 315, + 292, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 315, + 292, + 556, + 495 + ], + "type": "table", + "html": "
MethodPose ErrorVCRE
Pr.↑Avg.Med.↓Med.↓Pr.@45↑Pr.@90↑
MFBin[1]9.1210cm/33.7°199.715.430.1
+EquiPose12.0185cm/32.5°196.918.231.6
MFBinTS[1]10.5183cm/23.3°166.418.135.5
+EquiPose12.9152cm/23.4°166.021.739.3
MFQuat[1]5.1203cm/27.3°160.714.333.6
+EquiPose6.3184cm/23.3°166.416.635.0
MFQuatTS[1]4.4185cm/26.3°154.814.234.7
+EquiPose6.6170cm/23.5°157.617.336.4
MF6D[1]6.1168cm/22.5°147.118.740.2
+EquiPose9.6153cm/18.6°148.222.140.9
+EquiPose†10.1148cm/18.4°145.023.042.1
EightVit[46]6.0204cm/26.6°155.217.835.8
+EquiPose6.4195cm/26.3°157.218.336.4
+EquiPose†6.8188cm/25.1°153.119.137.0
GRelPose[28]12.7168cm/22.4°137.725.843.3
+EquiPose13.0166cm/21.9°136.426.143.3
+EquiPose†13.2163cm/21.7°132.926.244.1
", + "image_path": "02db4de956f44e425d0ea593b847aa79e7b607b24143ac48bdd6174858961c91.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 497, + 555, + 531 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 531 + ], + "type": "text", + "content": "Table 3. Comparison on the Mapfree dataset [1]. As done in [1], we report Pose Error in terms of rotation error and translation error and VCRE in terms of median error and precision." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 555, + 553, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 580 + ], + "type": "text", + "content": "These results demonstrate that EquiPose is also effective under the cross-dataset setup." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Besides the above quantitative results, we also provide visualization results from ScanNet. Considering GRelPose [28] performs relatively better than other baseline models in most cases, we adopt it as the baseline and compare it with EquiPose. The corresponding results are shown in Fig. 4. As seen from this figure, the proposed EquiPose is effective in improving the performance of relative pose estimation across different types of challenges, such as large viewpoint changes, and texture-less patterns. An inference-only implementation demonstrates obvious performance improvements, while subsequent fine-tuning could achieve" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1133" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 312, + 212 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 312, + 212 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 312, + 212 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 312, + 212 + ], + "type": "table", + "html": "
MethoddRdtRot. AUCTrans. AUCRETE
@3°@5°@10°@.1m@.2m@.5m
EightVit[46]2.70.1223.6940.2561.9214.9032.9558.478.900.41
+RP1.20.0522.7239.9361.8314.6033.8261.238.230.33
+RP+EquiPose0022.9640.1162.0315.0234.3061.738.200.31
+EquiPose†0025.3042.5464.0516.3736.1463.077.970.31
MF6D[1]7.30.1822.3637.1957.1014.3932.5757.349.810.44
+RP5.40.1120.4236.6258.1114.5633.0160.368.170.32
+RP+EquiPose0024.6140.2660.8516.4335.6362.757.740.30
+EquiPose†0027.1442.9962.6617.7937.7063.487.470.29
GRelPose[28]3.20.1228.9046.3666.5316.4635.3660.068.160.38
+RP2.10.0628.1145.6666.3616.9537.6563.647.740.30
+RP+EquiPose0029.0246.3566.8018.2038.8864.477.650.29
+EquiPose†0032.3249.7469.3719.6740.4065.957.610.28
", + "image_path": "7dba6c516837cf92c78fbd6587ad5b97834ba27f17f698b45e53c37a77cf15a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "type": "text", + "content": "Table 4. EquiPose VS. Random Permutation. " + }, + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "type": "inline_equation", + "content": "^+\\mathrm{RP}" + }, + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "type": "text", + "content": " denotes the model is trained with random permutation. " + }, + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "type": "inline_equation", + "content": "d_{R}" + }, + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 55, + 214, + 295, + 248 + ], + "type": "text", + "content": " denote the rotation and translation difference computed via Eq. (4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 272, + 179, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 179, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 179, + 284 + ], + "type": "text", + "content": "additional performance gains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 285, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 295, + 429 + ], + "type": "text", + "content": "Outdoor Pose Evaluation. The Mapfree dataset [1] was introduced for evaluating 6DoF relocalization, thus we only evaluate EightVit, MF6D, and GRelPose which could predict 6DoF poses. As a supplement, we further evaluate EquiPose in an inference-only manner on four additional baselines, including MFBin, MFBinTS, MFQuatTS which were released1 along with MF6D [1]. The results are reported in Tab. 3. As seen from this table, EquiPose could boost the performances of the above seven models in most cases. And the fine-tuned MF6D+EquiPose† achieves the lowest pose error among them. These results validate the effectiveness of our method under the outdoor setup." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 430, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 430, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 295, + 538 + ], + "type": "text", + "content": "EquiPose VS. Random Permutation. Intuitively, another way to leverage the PPE property for relative pose estimation is to permute the input order of image pairs randomly during training. To evaluate the effect of random permutation, we have further trained and tested EightVit [46], MF6D [1] and GRelPose [28] on the ScanNet dataset by using random permutation (RP). The corresponding results are reported in Tab. 4. In addition, we also report the average values of " + }, + { + "bbox": [ + 54, + 430, + 295, + 538 + ], + "type": "inline_equation", + "content": "d_{R}" + }, + { + "bbox": [ + 54, + 430, + 295, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 430, + 295, + 538 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 54, + 430, + 295, + 538 + ], + "type": "text", + "content": " computed via Eq. (4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 539, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 539, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 539, + 295, + 682 + ], + "type": "text", + "content": "As seen from this table, random permutation could improve model performances in average rotation and translation errors. However, the models trained under EquiPose perform significantly better than using random permutation. Besides, random permutation could not strictly enforce the PPE property as evidenced by both " + }, + { + "bbox": [ + 54, + 539, + 295, + 682 + ], + "type": "inline_equation", + "content": "d_{R}" + }, + { + "bbox": [ + 54, + 539, + 295, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 539, + 295, + 682 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 54, + 539, + 295, + 682 + ], + "type": "text", + "content": " being larger than zero, while EquiPose could achieve PPE strictly. To this end, EquiPose is not opposed to using random permutation, i.e., by enforcing the PPE property, it could also boost the performance of models trained using random permutation as observed in this table. The above results further validate the effectiveness of our method." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 547, + 198 + ], + "type": "table", + "html": "
MethodParams.FLOPsTime (ms)
ER1.92 × 1079.86 × 1092.0
ER+EquiPose1.92 × 1071.09 × 10102.3
Reg6D2.33 × 1079.91 × 1092.6
Reg6D+EquiPose2.33 × 1079.97 × 1092.6
EightVit1.91 × 1076.88 × 1093.5
EightVit+EquiPose1.91 × 1079.64 × 1095.1
MF6D2.19 × 1074.57 × 10105.6
MF6D+EquiPose2.19 × 1074.60 × 10106.7
GRelPose1.37 × 1079.82 × 10911.8
GRelPose+EquiPose1.37 × 1071.13 × 101012.0
", + "image_path": "7c9155e86c7c90e9ca778cd6ea41e46fcf0449e3bc15443a3b06e1b9e12971ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "type": "text", + "content": "Table 5. Comparison of parameter numbers, FLOPs, and inference times (ms) per image on an RTX 3090Ti GPU." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 533, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 533, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 533, + 255 + ], + "type": "text", + "content": "4.4. Computation Comparison and Limitation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 261, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 553, + 393 + ], + "type": "text", + "content": "Here, we compare the parameter numbers, Floating Point Operations (FLOPs) and the inference times of the baselines and their EquiPose implementations. All the models are evaluated on an RTX 3090Ti GPU on ScanNet. The results are reported in Tab. 5. As is seen, compared with the baseline model, the EquiPose implementation does not introduce any additional parameters, since it uses the same backbone as the baseline, and thus does not incur extra storage costs. Moreover, compared with each baseline model, EquiPose results in a modest increment in computational costs generally, due to the feature permutation strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "type": "text", + "content": "However, the extra computational cost depends on the architecture of the baseline model: For models with lightweight feature aggregation modules and pose decoders, the increase in computation is slight. For models with heavy feature aggregation modules and pose decoders, the extra computational cost would be high, which is a limitation of EquiPose. Despite this, all the models under EquiPose maintain an inference time of less than 15 ms per image, which is sufficient for practical applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 513, + 391, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 391, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 391, + 525 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 676 + ], + "type": "text", + "content": "This work focuses on end-to-end relative pose estimation. Firstly, we introduce the overlooked pose permutation equivariance (PPE) property for relative pose estimation. Leveraging this property, we establish the foundational principle for designing a relative pose estimation framework that incorporates the PPE property. Building on this theoretical basis, we propose the general and flexible EquiPose framework for relative pose estimation. Experiments on four public datasets further demonstrate that the performances of pre-trained models could be improved even without re-training, and moreover, their performances could be further boosted after being fine-tuned under EquiPose." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62376269)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "text", + "content": "1Pretrained weights are downloaded from the website of Mapfree: https://github.com/nianticlabs/map-free-reloc." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1134" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 2, 3, 5, 6, 7, 8, 1, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 246 + ], + "type": "text", + "content": "[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "text", + "content": "[4] Georg Bokman, Johan Edstedt, Michael Felsberg, and Fredrik Kahl. Steerers: A framework for rotation equivariant keypoint descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4885-4895, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "type": "text", + "content": "[5] Jane Bromley, Isabelle Guyon, Yann LeCun, and et al. Signature verification using a \"siamese\" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "type": "text", + "content": "[6] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14566–14575, 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 457 + ], + "type": "text", + "content": "[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 513 + ], + "type": "text", + "content": "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "text", + "content": "[9] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "type": "text", + "content": "[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2, 6, 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[12] Shay Dekel and Yosi Keller. Estimating extreme 3d image rotation with transformer cross-attention. arXiv preprint arXiv:2303.02615, 2023. 1, 3, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition workshops, pages 224-236, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "text", + "content": "[14] Qiulei Dong, Xiang Gao, Hainan Cui, and Zhanyi Hu. Robust camera translation estimation via rank enforcement. IEEE Transactions on Cybernetics, 52(2):862-872, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 204 + ], + "type": "text", + "content": "[15] Qulei Dong, Shuang Deng, and Yuzhen Liu. Cayley rotation averaging: Multiple camera averaging under the cayley framework. IEEE Transactions on Image Processing, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 206, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 270 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 272, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 327 + ], + "type": "text", + "content": "[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 328, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 328, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 553, + 380 + ], + "type": "text", + "content": "[18] Johan Edstedt, Georg Bökman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. In 2024 International Conference on 3D Vision, pages 148–157. IEEE, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "text", + "content": "[19] Sovann En, Alexis Lechery, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision Workshops, 2018. 1, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "text", + "content": "[20] Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and ransac's role. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8935-8943, 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 515, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 548 + ], + "type": "text", + "content": "[22] Richard Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In CVPR 2011, pages 3041-3048. IEEE, 2011. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 549, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 549, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 549, + 553, + 580 + ], + "type": "text", + "content": "[23] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hongdong Li. Rotation averaging. International Journal of Computer Vision, 103:267-305, 2013. 3, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "type": "text", + "content": "[24] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(6):580-593, 1997. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "type": "text", + "content": "[26] Will Hutchcroft, Yuguang Li, Ivaylo Boyadzhiev, Zhiqiang Wan, Haiyan Wang, and Sing Bing Kang. Covispose: Covisibility pose transformer for wide-baseline relative pose estimation in 360 indoor panoramicas. In European Conference on Computer Vision, pages 615-633. Springer, 2022. 3, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1135" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[27] Nicolas Keriven and Gabriel Peyre. Universal invariant and equivariant graph neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[28] Fadi Khatib, Yuval Margalit, Meirav Galun, and Ronen Basri. Leveraging image matching toward end-to-end relative camera pose regression, 2024. 1, 2, 3, 6, 7, 8, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[29] Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Deep Learning Workshop, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "text", + "content": "[30] Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International Conference on Machine Learning, pages 2747-2755. PMLR, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "text", + "content": "[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "text", + "content": "[32] Yuzhen Liu and Qiulei Dong. Descriptor distillation: A teacher-student-regularized framework for learning local descriptors. International Journal of Computer Vision, pages 1-19, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "type": "text", + "content": "[33] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "text", + "content": "[34] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In Proceedings of the 2019 IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "type": "text", + "content": "[35] Clare Lyle, Mark van der Wilk, Marta Kwiatkowska, Yarin Gal, and Benjamin Bloem-Reddy. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[36] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional neural networks. In International Conference on Advanced Concepts for Intelligent Vision Systems, pages 675-687. Springer, 2017. 1, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[37] Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. In Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[38] Maher Moakher. Means and averaging in the group of rotations. SIAM journal on matrix analysis and applications, 24 (1):1-16, 2002. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "text", + "content": "[39] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[40] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mokolajczyk. Ood-pose: camera pose regression from out-of-distribution synthetic views. In International Conference on 3D Vision, pages 722-732. IEEE, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[41] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[42] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 259 + ], + "type": "text", + "content": "[45] Jerome Revaud, Yohann Cabon, Romain Brégier, JongMin Lee, and Philippe Weinzaepfel. Sacreg: Scene-agnostic coordinate regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 688–698, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[46] Chris Rockwell, Justin Johnson, and David F. Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision, 2022. 1, 3, 5, 6, 7, 8, 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "text", + "content": "[47] Chris Rockwell, Nilesh Kulkarni, Linyi Jin, Jeong Joon Park, Justin Johnson, and David F Fouhey. Far: Flexible accurate and robust 6 dof relative camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19854-19864, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "type": "text", + "content": "[48] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 2, 3, 6, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 480 + ], + "type": "text", + "content": "[49] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 6, 7, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "text", + "content": "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 2, 3, 6, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "text", + "content": "[51] Yurun Tian, Axel Barroso-Laguna, Tony Ng, Vassileios Balntas, and Krystian Mikolajczyk. Hynet: Learning local descriptor with hybrid similarity measure and triplet loss. In Advances in Neural Information Processing Systems, pages 7401-7412, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "text", + "content": "[52] Julien Valentin, Angela Dai, Matthias Nießner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In International Conference on 3D Vision, pages 323-332. IEEE, 2016. 6, 7, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "type": "text", + "content": "[53] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[54] Guangyu Wang, Jinzhi Zhang, Fan Wang, Ruqi Huang, and Lu Fang. Xscale-nvs: Cross-scale novel view synthesis with" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1136" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 487 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "hash featurized manifold. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21029-21039, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Asian Conference on Computer Vision, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "text", + "content": "[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "text", + "content": "[57] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "text", + "content": "[58] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual u-net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "type": "text", + "content": "[59] Xiaoming Zhao, Xingming Wu, Jinyu Miao, Weihai Chen, Peter CY Chen, and Zhengguo Li. Alike: Accurate and lightweight keypoint detection and descriptor extraction. IEEE Transactions on Multimedia, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 296, + 386 + ], + "type": "text", + "content": "[60] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4669-4678, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 296, + 431 + ], + "type": "text", + "content": "[61] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1851-1858, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "text", + "content": "[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 3, 6, 7, 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1137" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_content_list.json b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79fa1e2a1f31d21c65c28a1992f1ac76e8fce560 --- /dev/null +++ b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_content_list.json @@ -0,0 +1,1483 @@ +[ + { + "type": "text", + "text": "Erase Diffusion: Empowering Object Removal Through Calibrating Diffusion Pathways", + "text_level": 1, + "bbox": [ + 155, + 128, + 844, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Liu*, Hao Zhou*, Benlei Cui†, Wenxiang Shang, Ran Lin Alibaba Group", + "bbox": [ + 259, + 202, + 738, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ly402089,baishu.zh,cuibenlei.cbl,shangwenxiang.swx,linran.lr09}@taobao.com", + "bbox": [ + 163, + 241, + 828, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9012fe4d2c76d6425a890207ec4dda4fca8af07c8402537114a83df786fa68e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 292, + 287, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9cf332200c98842a53ab2c8c9aac846ae491cf3249e33a069ad25cfabd9344e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 395, + 287, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a34409901c6394bf90a22338c1058e9efee906a2bb5f94268d252e96aa3936a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 292, + 495, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4df7140a17c3ba39d51bae3a8f2580ba3948b56ec3f7ae8416f02349d668b78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 292, + 700, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c4558ad4aa3d86c08071d1759c17b99380e815b154b2c34675817f1dcabe776.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 292, + 901, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b085af2e69fc4034c70413fb5df5a155ebc8e10ad8d83d0a1482e9c6eb45742.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 395, + 901, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbad07ff923cf89d006fe593ae8ae0ec30a826e844348a4dc68009846188a954.jpg", + "image_caption": [ + "Figure 1. Diverse erase inpainting results produced by our proposed EraDiff, where images before and after removal are presented in pairs, and the areas to be erased in the original images have been marked. The EraDiff can eliminate targets in various complex real-world scenes while ensuring visual coherence in the generated images." + ], + "image_footnote": [], + "bbox": [ + 94, + 501, + 290, + 652 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bcd965ade37fa1876b3bb3addedb71b91e0b63033a5ec47d5b80eb70a72c2817.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 501, + 493, + 652 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f64599edb3f1420c16da5fbf8bd07cc4dcc8e4bf50ab8b783d83a8e393c14f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 501, + 697, + 652 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50a1fe9e7a07290dac8a2e84117ed3890217db55ba42f54f9e1ab5c9401a75eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 501, + 901, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 726, + 326, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Erase inpainting, or object removal, aims to precisely remove target objects within masked regions while preserving the overall consistency of the surrounding content. Despite diffusion-based methods have made significant strides in the field of image inpainting, challenges remain regarding the emergence of unexpected objects or artifacts. We assert that the inexact diffusion pathways established by existing", + "bbox": [ + 88, + 758, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "standard optimization paradigms constrain the efficacy of object removal. To tackle these challenges, we propose a novel Erase Diffusion, termed EraDiff, aimed at unleashing the potential power of standard diffusion in the context of object removal. In contrast to standard diffusion, the EraDiff adapts both the optimization paradigm and the network to improve the coherence and elimination of the erasure results. We first introduce a Chain-Rectifying Optimization (CRO) paradigm, a sophisticated diffusion process specifically designed to align with the objectives of erasure. This paradigm establishes innovative diffusion transition", + "bbox": [ + 511, + 728, + 906, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally to this work.", + "bbox": [ + 89, + 875, + 320, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 91, + 887, + 207, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2418", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pathways that simulate the gradual elimination of objects during optimization, allowing the model to accurately capture the intent of object removal. Furthermore, to mitigate deviations caused by artifacts during the sampling pathways, we develop a simple yet effective Self-Rectifying Attention (SRA) mechanism. The SRA calibrates the sampling pathways by altering self-attention activation, allowing the model to effectively bypass artifacts while further enhancing the coherence of the generated content. With this design, our proposed EraDiff achieves state-of-the-art performance on the OpenImages V5 dataset and demonstrates significant superiority in real-world scenarios.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 301, + 222, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image inpainting is a crucial technique in computer vision, aimed at reconstructing missing or damaged regions while maintaining visual coherence and contextual integrity [15, 32]. As one specialized form of image inpainting, object removal, also called erase inpainting, has broader applications in various fields such as social media [3, 50], advertising design [36], and image processing [37, 46]. However, in comparison to general image inpainting, an ideal erase inpainting model must address two critical challenges. Coherence: The model should seamlessly inpaint the masked area, ensuring consistency in lighting, content, and other relevant aspects. Elimination: The model should accurately remove objects within the masked area while preventing the generation of extraneous elements or artifacts.", + "bbox": [ + 89, + 325, + 482, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior erase inpainting methods primarily relied on nonparametric patch sampling techniques [4, 10, 19] or Generative Adversarial Networks (GANs) [5, 8, 14, 22, 29, 33, 41, 44, 48]. These approaches often resort to filling large masked regions with repetitive patches, resulting in a lack of coherence in the generated images. In recent years, methods based on Latent Diffusion Models (LDMs) [27] have demonstrated superiority in generating more natural images. However, these methods struggle to eliminate target objects. For example, when a user seeks to erase a piece of pizza in Figure 1, the LDMs may mistakenly generate another piece of pizza instead of the expected clean plate.", + "bbox": [ + 89, + 537, + 482, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We attribute the above issues to the fact that standard diffusion pathways in most existing methods are not suitable for erase tasks. During optimization, most diffusion-based erase inpainting models take the original image with added noise and randomly generated masks as input, aiming to recover the original image in the presence of noise. However, this standard training paradigm only establishes a denoising process that transitions from random noise to clear images, without addressing the specific goal of object removal. Consequently, the model may follow a denoising pathway from noise to an image with objects. Based on this observation, we argue that an ideal erase inpainting model should", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "establish a diffusion pathway directly from objects to backgrounds to ensure the definitive removal of unwanted elements. Furthermore, in the early stages of denoising, the reconstructed regions are significantly influenced by the shape of the masks and the level of noise. This impact may lead to deviations in the early latent states, resulting in the emergence of artifacts. The standard self-attention mechanism tends to incorrectly regard the features of these artifacts as important information, which gradually amplifies the deviations during the subsequent denoising process. Ultimately, it leads to unexpected objects in the generated images.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel erase diffusion model, termed EraDiff, specifically designed for object removal, which effectively unleashes the potential erasure power of standard diffusion. First, we introduce a Chain-Rectifying Optimization (CRO) paradigm that supports the establishment of new diffusion pathways from noise to backgrounds. Specifically, we develop a dynamic image synthesis strategy that enables the generation of a variety of dynamic images at different time steps without the need for extra data. These synthesized dynamic images effectively simulate the gradual elimination of objects during the denoising process, yielding intermediate latent states corresponding to different time steps. Next, we introduce a new dedicated optimization objective for erase inpainting. This objective guides the establishment of transitions among multiple intermediate latent states. By applying the CRO paradigm, the model learns to accurately identify the intention to erase and improve content coherence. Furthermore, to address potential deviations caused by artifacts during the early denoising process, we design a Self-Rectifying Attention (SRA) mechanism that explicitly guides the model in executing object removal more effectively. By altering self-attention activation, the SRA enhances background features while rectifying its incorrect reliance on artifacts. This replacement of the standard attention mechanism results in final generated images that appear more realistic and devoid of unexpected objects. To the best of our knowledge, our proposed method achieves state-of-the-art performance in both the public OpenImages V5 [18] dataset and large-scale real-world scenarios.", + "bbox": [ + 511, + 258, + 903, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 723, + 653, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Inpainting. Image inpainting, which involves the reconstruction of large-scale missing regions, has garnered substantial attention within the field. Traditional image inpainting methods employ heuristic patch-based propagation algorithms to borrow texture and structure from neighboring regions of the corrupted areas [4, 10, 19]. However, these methods often struggle to effectively restore areas with complex structures or semantics [42]. To tackle this issue, numerous studies have harnessed the power of deep neural networks [6, 34, 49] and proposed various en", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2419", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hancements to GANs [25] aimed at improving the coherence of inpainted regions. Subsequent research has extensively examined different facets, including semantic context and texture [1, 7, 13], edges and contours [5, 8], as well as the design of hand-engineered architectures [20, 51]. Notably, Suvorov et al. [33] introduced FFCs into residual modules, providing a remedy for enhancing the perceptual quality of GAN-generated outputs. More recently, promising efforts have been made in applying LDMs to inpainting tasks, showcasing their potential to achieve impressive results [9, 23, 26, 31, 35, 38]. However, general image inpainting models may generate unexpected objects in the masked areas, making these LDM-based methods unsuitable for effective object removal.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Erase Inpainting. Erase inpainting, one specialized form of image inpainting, focuses on removing unwanted content from images. The typical setup for erase inpainting is to provide an image along with an object mask as input, ultimately generating a clean background that excludes the specified object. Many studies have concentrated on improvements in model architecture [22, 40, 43, 44] and loss functions [29, 33]. While these methods effectively improve the coherence of masked regions, they still struggle with object removal. Recent works [2, 12, 21, 30, 45] have attempted to leverage textual prompts to locate erase targets within images. However, relying on textual cues to accurately identify areas for erasure has proven to be unstable. Models fail to follow textual instructions, which can lead to unfavorable results. As a result, these approaches face challenges in large-scale practical applications. Through in-depth investigation, we find that the struggle to effectively erase objects stems from the inexact construction of the erase diffusion chain. Our proposed method attempts to directly establish diffusion pathways from the noise to the background. This approach enables effective object removal without relying on additional textual information and produces one clean and natural background.", + "bbox": [ + 91, + 303, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 661, + 230, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Diffusion Models. In this paper, we utilize latent diffusion models (LDMs) as our erase inpainting model. LDMs are probabilistic models designed to learn a data distribution $p(\\pmb{x})$ by gradually denoising a normally distributed variable, which corresponds to learning the reverse sampling process of a fixed Markov Chain of length $T$ . Specifically, the forward diffusion process progressively adds noise to the initial latent state $\\pmb{x}_0$ ,", + "bbox": [ + 89, + 686, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\boldsymbol {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 815, + 482, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $1 - \\bar{\\alpha}_t$ is the variance schedule at step $t$ and represents the level of noise, with $\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s$ .", + "bbox": [ + 89, + 840, + 482, + 871 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to the DDIM algorithm, the reverse sampling process can be defined as", + "bbox": [ + 89, + 871, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} _ {p r e v} = \\sqrt {\\bar {\\alpha} _ {p r e v}} \\left(\\frac {\\boldsymbol {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) \\tag {2} \\\\ + \\sqrt {1 - \\bar {\\alpha} _ {p r e v} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t}) + \\sigma_ {t} \\epsilon_ {t}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 108, + 903, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ is standard Gaussian noise. Here, $\\sigma_t$ are hyper-parameters, and the reverse sampling process can be considered deterministic as $\\sigma_t \\to 0$ . The term $\\epsilon_\\theta^{(t)}(\\boldsymbol{x}_t)$ is trained to predict the noise added to $\\boldsymbol{x}_t$ by minimizing the following optimization objective", + "bbox": [ + 511, + 181, + 906, + 260 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (\\mathbf {0}, I)} \\left\\| \\epsilon - \\epsilon_ {\\theta} ^ {(t)} \\left(\\boldsymbol {x} _ {t}\\right) \\right\\| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 265, + 903, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structure of Erasing Ipainting. Most advanced erasing inpainting methods are based on diffusion models. In this paper, we focus on the SD2-Inpaint model[27], which mainly consists of a Variational AutoEncoder (VAE) [17] and a U-Net network [28]. The VAE facilitates the transformation of input images into latent space. For Erasing Ipainting, the inputs typically consist of a noisy image $\\boldsymbol{x}_t^{ori} \\in \\mathbb{R}^{H \\times W \\times 3}$ , a random binary mask $\\mathbf{M} \\in \\{0,1\\}^{H \\times W}$ , and the corresponding masked image $\\boldsymbol{x}' = \\boldsymbol{x}_0^{ori} \\odot \\mathbf{M}$ . The U-Net network predicts the associated noise at timestamp $t$ based on the Equation 3. This methodology enables the model to establish a diffusion chain between the original image and Gaussian noise, allowing for seamlessly inpainting masked regions.", + "bbox": [ + 511, + 299, + 906, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the self-attention mechanism within U-Net networks plays a significant role in the erasing process. This mechanism aggregates information from the entire image, thereby controlling the generation of features in the masked regions during the reverse sampling process. Given one latent feature map $z \\in \\mathbb{R}^{h \\times w \\times c}$ , where $h$ , $w$ and $c$ are the height, width, and channel dimensions of $z$ respectively, the self-attention process can be represented as follows", + "bbox": [ + 511, + 496, + 905, + 616 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} = \\ell_ {Q} (\\boldsymbol {z}), \\ell_ {K} (\\boldsymbol {z}), \\ell_ {V} (\\boldsymbol {z}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 622, + 903, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 646, + 903, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ell_{Q},\\ell_{K},\\ell_{V}$ are learnable linear layers, and $d$ denotes the scaling factor.", + "bbox": [ + 511, + 683, + 905, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 724, + 604, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. The overall architecture of our proposed Erase Diffusion, termed EraDiff, is shown in Figure 2, consisting of a Chain-Rectifying Optimization paradigm (see Section 4.1) and a Self-Rectifying Attention mechanism (see Section 4.2). This CRO paradigm introduces the design of dynamic latent states along with a dedicated optimization objective, thereby establishing novel diffusion chains that better align with the erasure objectives. Meanwhile, the SRA mechanism effectively guides the sampling process to prevent state deviations caused by artifacts.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2420", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/daf41fa0b78b4a11615ca8cf467fbe269d34071c6fae5815d754f1db6f9ed420.jpg", + "image_caption": [ + "Figure 2. The overview of our proposed Erase Diffusion, termed EraDiff. Left: Dynamic image synthesis. Each image is initially transformed using techniques like matting, scaling, and copy-pasting. A mix-up strategy then synthesizes a series of dynamic images $\\{\\tilde{x}_t^{mix}\\}$ that simulate the gradual fading of the object. Top: Chain-Rectifying Optimization (CRO). The standard sampling pathway is prone to generating artifacts (black dashed lines). In contrast, we establish a new sampling path for erasing (red dashed lines) that better aligns the reverse sampling trajectory with a clear background. Bottom: Self-Rectifying Attention (SRA). The standard self-attention mechanism may inadvertently amplify artifacts, diverging from the expected diffusion pathway. By modifying the attention activation, we guide the model to bypass artifact regions, enhancing its focus on the background and ensuring a more accurate erase sampling path." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Chain-Rectifying Optimization", + "text_level": 1, + "bbox": [ + 89, + 604, + 366, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the optimization process, existing erase inpainting methods typically employ randomly generated or object-based masks, predicting the noise in the masked regions using Equation 3. These masked regions may cover objects, and the model's optimization objective is to reconstruct these objects rather than eliminate them. As discussed above, such standard diffusion chains may inadvertently encourage the model to generate unexpected artifacts. An effective solution is to dedicate diffusion paths for the erase inpainting task between objects and backgrounds. By simulating the gradual fading of objects, this approach can suppress their emergence during the sampling process. Even if the sampling process does not start with standard Gaussian noise (a common enhancement trick), it can still help eliminate leaked objects within the masked region. Thus, the main challenge of this approach is to develop new latent states within the erase diffusion chains and to optimize the model to manipulate transitions among these latent states.", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Latent States. To design dedicated diffusion chains for erasure between objects and backgrounds, it is essential to obtain corresponding image pairs. However, existing public datasets often lack these paired samples, highlighting the need for an innovative data synthesis strategy.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\pmb{x}_0^{ori}$ denote the original image, and a trained matting model is employed to segment the primary object within it. We then apply various image transformation techniques, including rotation and scaling, to modify the segmented objects. These transformed objects are randomly pasted onto the background region, resulting in a new synthesized image, $\\pmb{x}_0^{obj}$ . This method allows us to construct a training dataset of object-background image pairs at a relatively low cost. Furthermore, to effectively simulate the gradual fading of the objects, we input dynamic images $\\tilde{\\pmb{x}}_t^{mix}$ for each time step $t$ during the training process,", + "bbox": [ + 511, + 167, + 906, + 334 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {x}} _ {t} ^ {m i x} = (1 - \\lambda_ {t}) \\boldsymbol {x} _ {0} ^ {o r i} + \\lambda_ {t} \\boldsymbol {x} _ {0} ^ {o b j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 344, + 906, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the decreasing sequence $\\lambda_{:T} \\in [0,1]^T$ controls the mix-up level of object-background image pairs. In the case of $t = 0$ , $\\tilde{\\pmb{x}}_0^{mix}$ corresponds to the original image $\\pmb{x}_0^{ori}$ . These dynamic mix-up images assist the model in understanding the smooth transition from objects to backgrounds. Ultimately, based on Equation 1, we can derive a set of new latent states $\\pmb{x}_t^{mix}$ for the self-rectifying diffusion chains,", + "bbox": [ + 511, + 375, + 905, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} ^ {m i x} = \\sqrt {\\bar {\\alpha} _ {t}} \\tilde {\\boldsymbol {x}} _ {t} ^ {m i x} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 492, + 906, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$", + "bbox": [ + 511, + 521, + 648, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization Objective. The traditional diffusion process starts from a fixed $\\boldsymbol{x}_0^{ori}$ , and each intermediate latent state $\\boldsymbol{x}_t^{ori}$ can be obtained directly using Equation 1. However, in the new self-rectifying diffusion chains, each latent state $\\boldsymbol{x}_t^{mix}$ is derived from the dynamic image $\\tilde{\\boldsymbol{x}}_t^{mix}$ , making it infeasible to apply the standard optimization objective in Equation 3. To mitigate this challenge, we propose a novel optimization objective that enhances the alignment between the model's predicted distribution and the true distribution, as illustrated in Algorithm 1.", + "bbox": [ + 511, + 537, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the new loss function aims to minimize the distance between the model-predicted latent states $\\hat{\\pmb{x}}_t^{mix}$ and the true states $\\pmb{x}_t^{mix}$ . It allows the model to gradually alter its parameters to adapt to the new distribution shifts. Given one latent state $\\pmb{x}_t^{mix}$ , we can derive the model-predicted latent state $\\hat{\\pmb{x}}_{t - \\gamma}^{mix}$ at the previous $\\gamma$ time step,", + "bbox": [ + 511, + 688, + 905, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\theta} (\\hat {\\boldsymbol {x}} _ {t - \\gamma} ^ {m i x} | \\boldsymbol {x} _ {t} ^ {m i x}) = \\sqrt {\\bar {\\alpha} _ {t - \\gamma}} \\left(\\frac {\\boldsymbol {x} _ {t} ^ {m i x} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t} ^ {m i x})}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) \\\\ + \\sqrt {1 - \\bar {\\alpha} _ {t - \\gamma}} \\epsilon_ {\\theta} ^ {(t)} \\left(\\boldsymbol {x} _ {t} ^ {m i x}\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 792, + 903, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma \\in (0,\\gamma_{m})$ . Given the original image $\\pmb{x}_0^{ori}$ and the synthesized image $\\pmb{x}_0^{obj}$ , the corresponding true state $\\pmb{x}_{t - \\gamma}^{mix}$", + "bbox": [ + 511, + 869, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2421", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Chain-Rectifying Optimization" + ], + "code_body": "repeat \n $\\pmb{x}_0^{ori}\\sim q(x)$ $t\\sim$ Uniform $(\\{1,\\dots ,T\\})$ $\\gamma \\sim$ Uniform $(\\{1,\\ldots ,\\gamma_{m}\\})$ $\\epsilon \\sim \\mathcal{N}(0,I)$ $\\pmb{x}_0^{obj}\\gets \\pmb{x}_0^{ori}$ image transformation; \n $\\pmb{x}_{t}^{mix},\\pmb{x}_{t - \\gamma}^{mix}\\leftarrow \\pmb{x}_{0}^{ori},\\pmb{x}_{0}^{obj}$ based on Equ 6-7; \n $\\hat{\\pmb{x}}_{t - \\gamma}^{mix}\\longleftarrow \\pmb{x}_{t}^{mix},\\epsilon_{\\theta}^{(t)}$ based on Equ 8; \nTake gradient descent step on \n $\\nabla_{\\theta}\\left\\| \\pmb{x}_{t - \\gamma}^{mix} - \\hat{\\pmb{x}}_{t - \\gamma}^{mix}\\right\\| ^2$ based on Equ 9 \nuntil converged;", + "bbox": [ + 112, + 116, + 480, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can be obtained using Equation 6 and 7. Finally, we define the new optimization objective as follows", + "bbox": [ + 89, + 327, + 482, + 358 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\gamma \\sim \\operatorname {U n i f o r m} (1, \\gamma_ {m}), t} \\left\\| \\boldsymbol {x} _ {t - \\gamma} ^ {m i x} - p _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {t - \\gamma} ^ {m i x} \\mid \\boldsymbol {x} _ {t} ^ {m i x}\\right) \\right\\| _ {2} ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 366, + 483, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Self-Rectifying Attention", + "text_level": 1, + "bbox": [ + 89, + 398, + 318, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have established a self-rectifying diffusion pathway to guide the model toward object removal. However, information leakage from the mask's shape can still lead to artifacts during the early stages of denoising, resulting in latent state shifts. The self-attention mechanism tends to give the masked region stronger attention to itself rather than to the background. This phenomenon can continuously amplify artifacts along the reverse sampling path, ultimately leading to a deviation from the object removal direction. An intuitive solution is to alter the current attention layers for path calibration to mitigate the above risk of generating unexpected objects due to deviations from specific states.", + "bbox": [ + 89, + 420, + 483, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on our observations, we believe that the generation of the foreground region should rely more on backgrounds rather than focusing on itself. Additionally, the background region remains visible and thus should not be affected by the content of the foregrounds. By altering the self-attention activation, it is possible to ignore the negative effects of artifacts while emphasizing the background, thereby further enhancing the coherence and elimination of the generated content. Thus, we propose a simple yet effective Self-Rectifying Attention mechanism to replace the standard self-attention mechanism. Specifically, we first downsample and flatten the image mask $\\mathbf{M}$ to obtain the corresponding mask vector $m\\in \\{0,1\\}^{wh}$ . We then design an extended mask $m^{\\prime}\\in \\{-inf,1\\}^{wh\\times wh}$ as follows", + "bbox": [ + 89, + 603, + 483, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i, j} ^ {\\prime} = \\left\\{ \\begin{array}{l l} 1, & m _ {i} = 0 \\text {o r} m _ {j} = 0 \\\\ - i n f, & \\text {e l s e} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 821, + 482, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The extended mask is subsequently applied directly to the corresponding attention activations, effectively suppressing", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "objects within the masked regions while enhancing background features. This is represented mathematically as", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S R A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}} \\cdot m ^ {\\prime}\\right) \\mathbf {V}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 133, + 906, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This mechanism enhances the model's capacity to perceive background information without any additional computational cost, while also effectively diminishing the interference of artifacts on the final generated results. Consequently, this improvement ensures that the sampling process can calibrate itself to the target direction of object removal.", + "bbox": [ + 511, + 181, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 289, + 645, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 315, + 702, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Datasets. We conducted a thorough evaluation of our proposed pipeline utilizing the publicly available OpenImages V5 segmentation dataset [18]. Each instance within this dataset comprises the original image, a corresponding segmentation mask, segmentation bounding boxes, and associated class labels, thereby enabling a rigorous comparative analysis against several baseline models. For our experiments, we randomly selected a subset of 10,000 samples from the OpenImages V5 test set.", + "bbox": [ + 511, + 338, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Baselines. We have chosen several cutting-edge image inpainting methods as our baselines for comparison. These include two mask-guided approaches: SD2-Inpaint [27] and LaMa [33], alongside two text-guided techniques: Inst-Inpaint [45] and PowerPaint [52].", + "bbox": [ + 511, + 474, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. In accordance with LaMa [33], we employ two primary evaluation metrics: Fréchet Inception Distance (FID) [11] and Learned Perceptual Image Patch Similarity (LPIPS) [47]. These metrics are adept at assessing the overall visual coherence of the inpainted images. To provide a more nuanced evaluation of the quality of content generated within the masked regions, we introduce the Local FID metric [39], which allows for a detailed assessment of local visual fidelity. Furthermore, we augment our evaluation framework by incorporating analyses derived from GPT-4o [24] along with expert human annotations, thereby examining the effectiveness of these erasure-targeted models in eliminating objects and artifacts.", + "bbox": [ + 511, + 551, + 905, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Our proposed pipeline builds upon the foundations established by SD2-Inpaint and is implemented using PyTorch along with the Diffusers library. During the training phase, we employed the Adam [16] optimizer with a learning rate set to $3 \\times 10^{-6}$ on the OpenImages V5 training set. To simplify parameters, the increasing sequence $\\lambda_{:T}$ follows the same schedule as the sequence $1 - \\bar{\\alpha}_{:T}$ . For loss computation, we ensured that the time intervals between two timestamps did not exceed $\\gamma_{m} = 100$ . All experiments were conducted on NVIDIA A100 GPUs.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2422", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66a19e63e37ad046e2679e958ac26eac83779cd461df6111808b1824b14ebf27.jpg", + "image_caption": [ + "Figure 3. Qualitative results of OpenImages V5 dataset compared among SD2-Inpaint [27], SD2-Inpaint with prompt guidance [27], PowerPaint [52], Inst-Inpaint [45], LaMa [33], and our approach." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 903, + 554 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89892a31795a499aff256d3f7c1fe903ef07903a2f9a898906acdaffdf7a1f0d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓LPIPS↓Local FID↓
SD2-Inpaint3.8050.3018.852
SD2-Inpaint*†4.0190.3087.194
PowerPaint6.0270.28910.021
Inst-Inpaint11.4230.41043.472
LaMa7.5330.2196.091
EraDiff (ours)6.5400.1923.799
", + "bbox": [ + 91, + 609, + 480, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative assessment of various erase inpainting models on the OpenImages V5 dataset. Optimal results are highlighted in bold, with runner-up performance underlined.", + "bbox": [ + 89, + 743, + 482, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Qualitative and Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 89, + 813, + 454, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative evaluation results of our proposed method, in comparison to several established models, are summarized in Table 1. Additionally, we present the experimental findings when the prompt is used as supplementary guid", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ance for SD2-Inpaint (abbreviated as SD2-Inpaint*, consistent throughout the paper). As indicated in the table, our model is in the mid-range of all baseline models in terms of the FID score, yet it significantly surpasses the others in the Local FID metric. This observation highlighting its remarkable ability to generate visually coherent results specifically within the designated erased regions, all the while maintaining a commendable level of visual coherence throughout the entire image. Additionally, our method records the highest performance on the LPIPS metric, suggesting that the images produced exhibit enhanced visual fidelity following the removal process. Figure 3 illustrates relevant visualization results, clearly demonstrating the model's robustness in addressing challenges such as the presence of objects within the image that closely resemble the target erasure region, as well as extensive areas that necessitate removal.", + "bbox": [ + 511, + 612, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is essential to highlight that both the FID and LPIPS metrics primarily evaluate the visual coherence and aesthetic quality of the final generated images. However, these", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2423", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/575226ac7d2200ae76735e5e81f2bed82c464e5c8576796747ce2bfd9a278b16.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSuperiorComparableInferior
SD2-Inpaint1.03%18.07%80.90%
SD2-Inpaint*2.20%24.79%73.01%
LaMa13.17%35.29%51.54%
", + "bbox": [ + 93, + 85, + 480, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of OpenImages V5 dataset among SD2-Inpaint, SD2-Inpaint*, LaMa, and EraDiff. This table delineates a comparative analysis of the elimination performance results obtained by these methodologies relative to ours, highlighting whether their outcomes are superior, comparable, or inferior to those achieved by our approach.", + "bbox": [ + 89, + 172, + 482, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d19dbdd762df7979d9831daf6c2ff8d527184d3c0c4c7ade567ddd53a6494853.jpg", + "image_caption": [ + "Figure 4. Results from the user study. EraDiff demonstrates enhanced performance, as indicated by its higher mean scores in both elimination and coherence evaluations." + ], + "image_footnote": [], + "bbox": [ + 91, + 273, + 475, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "metrics do not provide a direct assessment of the effectiveness regarding object or artifact elimination in the denoising process. This limitation is clearly illustrated in Figure 3, which demonstrates that despite the SD2-Inpaint model achieving the highest FID score among all evaluated models, it frequently fails to adequately eliminate unwanted objects or artifacts from the designated masked areas.", + "bbox": [ + 89, + 566, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the baseline models and our proposed method in eliminating objects and artifacts, we systematically selected the top-performing models: SD2-Inpaint, SD2-Inpaint*, and LaMa, as indicated in Table 1. A comparative analysis was performed through pairwise comparisons against our method, utilizing GPT-4o to generate objective evaluations focused on identifying superior results among the competing models. The outcomes of this analysis are described in Table 2. Furthermore, we undertook a user study that involved 20 experts that was tasked with appraising the effectiveness of the erasure and the visual aesthetics of the generated results, using a scoring system ranging from 1 to 10; where a score of 1 reflects poor performance and a score of 10 indicates exceptional quality. The outcomes of this assessment are illustrated in Fig-", + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10383b492f5c1611b226cc85d3fbfde99ccefc8216bde19f98f8813dd27ca33c.jpg", + "image_caption": [ + "Figure 5. Visualization of EraDiff's performance across a diverse array of in-the-wild scenarios: animated imagery, e-commerce content, oil paintings, and glasses-free 3D visuals." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ure 4. The results from both experimental methodologies distinctly demonstrate that our proposed approach significantly outperforms the competitors in effectively eliminating target objects. This conclusion is further corroborated by the finding that, despite the lower FID score recorded for SD2-Inpaint, its efficacy in object removal does not surpass that of our innovative model.", + "bbox": [ + 511, + 409, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we find that EraDiff shows superior performance across a diverse array of in-the-wild scenarios, as illustrated in Figure 5. A more detailed analysis of this will be provided in the appendix.", + "bbox": [ + 511, + 515, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 590, + 666, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the impact of calibrating the sampling pathway on the performance of erase inpainting, we conducted four extensive experiments: each focusing on the removal of CRO, SRA, and the simultaneous exclusion of both CRO and SRA, all based on the EraDiff model. Furthermore, we assessed the model's performance without the incorporation of the mix-up strategy during training, wherein the value of $\\lambda_{t}$ was held constant. For our evaluation, we employed the GPT-4o to quantify the success rate of object elimination. Results are detailed in Table 3. The results indicate that the removal of either the CRO or SRA components, as well as their simultaneous exclusion, leads to a substantial deterioration in both the visual coherence of the erased images and the efficacy of object elimination. Particularly noteworthy is the fact that the omission of CRO results in a more pronounced decline in overall performance. Additionally, the absence of the mix-up strategy produced considerable fluctuations in training loss and impeded model convergence. The instability observed can be attributed to", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2424", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7edab3f66294605d48a6f8b08850865f18ab8c5410d8fa29b403b4d371581944.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLocal FID↓GPT score↑
EraDiff3.79983.43%
-w/o CRO5.71372.96%
-w/o SRA4.95078.54%
-w/o CRO ∪ SRA8.85227.80%
-w/o mix-upNaNNaN
", + "bbox": [ + 109, + 88, + 467, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Results of the ablation study highlight the individual and combined effects of CRO and SRA methodologies.", + "bbox": [ + 89, + 205, + 483, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f81a14e564eefb95cabdcf7545742082a795d608bdf07351a210edccd6753c0.jpg", + "image_caption": [ + "Figure 6. Visual examples for the ablation study comparing baseline, baseline with CRO, baseline with SRA, and baseline with both CRO and SRA, displayed left to right." + ], + "image_footnote": [], + "bbox": [ + 94, + 247, + 483, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the limited noise present at earlier time steps, which complicates the model's ability to accurately predict the entirety of the masked region as noise. Figure 6 presents the visual outcomes of these ablation experiments, clearly illustrating the effectiveness of the calibration sampling pathway in enhancing the erasure task. In this regard, SRA serves a corrective function, effectively addressing specific anomalous denoising artifacts.", + "bbox": [ + 88, + 460, + 482, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. How the CRO and SRA work?", + "text_level": 1, + "bbox": [ + 89, + 590, + 362, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further elucidate the roles of the CRO and SRA methodologies in the denoising process of erase inpainting, we established two types of extended experimental setups.", + "bbox": [ + 89, + 613, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the mechanism underlying CRO, we established an extremely challenging condition by setting the denoising strength to 0.6. This configuration allows for significant leakage of original image information, including the target objects intended for removal. This leakage can lead to considerable disruptions, resulting in artifacts within the erased regions and potentially even the restoration of the erasure-targeted objects. As illustrated in Figure 7, we observed the denoising processes of the EraDiff alongside the baseline model (i.e., SD2-Inpaint) under this condition. Notably, as the denoising process progressed, the EraDiff model effectively concealed the object artifacts in the erased region, whereas the baseline model highlighted it. These experimental findings provide further validation of our earlier discussion in Section 4, asserting that EraDiff accomplishes the erasure task by traversing a sample pathway that closely", + "bbox": [ + 89, + 659, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c45982b5cf794fe1b2c6d47b1dea7664de05df4ee2ee135528461cdc8707e15.jpg", + "image_caption": [ + "Figure 7. Comparison of EraDiff's and baseline model's denoising process with strength set to 0.6." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 905, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e984b43fc0ef25c5bec9de0c1fb1b7a723ffeb592280176c85bcf9257b47c65e.jpg", + "image_caption": [ + "Figure 8. Visualization of heatmaps representing attention block outputs in the presence and absence of the SRA mechanism." + ], + "image_footnote": [], + "bbox": [ + 516, + 253, + 903, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "resembles $x_{t}^{mix}$", + "bbox": [ + 511, + 479, + 624, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To reveal the mechanism of the SRA approach, we examined a misleading scenario where multiple similar objects exist within an image, necessitating the removal of one specific object. The heatmaps generated before and after the attention block for both the EraDiff and the baseline model are presented in Figure 8. The results indicate that the baseline model tends to focus on similar objects, while our proposed method shifts its attention to the background. This strategic focus allows the model to extract critical information from the background rather than from the foreground, thereby facilitating the elimination of the target object while maintaining coherence between the erased region and the surrounding background.", + "bbox": [ + 511, + 496, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present the EraDiff, which enhances object elimination while maintaining visual coherence in erase inpainting. By introducing a CRO paradigm, EraDiff establishes innovative diffusion pathways that facilitate a gradual removal of objects, allowing the model to better understand the erasure intent. Moreover, the SRA mechanism effectively reduces artifacts during the sampling process. With inclusive experiments, we demonstrate that these advancements significantly improve performance in challenging scenarios, positioning EraDiff as a valuable contribution to the field of erase inpainting.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fazil Altinel, Mete Ozay, and Takayuki Okatani. Deep structured energy-based image inpainting. In ICPR, pages 423-428. IEEE Computer Society, 2018. 3", + "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 3", + "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: text-driven layered image and video editing. In ECCV, pages 707-723. Springer, 2022. 2", + "[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B. Goldman. Patchmatch: a randomized correspondence algorithm for structural image editing. ACM TOG, 28(3):24, 2009. 2", + "[5] Chenjie Cao, Qiaole Dong, and Yanwei Fu. ZITS++: image inpainting by improving the incremental transformer on structural priors. IEEE TPAMI, 45(10):12667-12684, 2023. 2, 3", + "[6] Benlei Cui, Xue-Mei Dong, Qiaoqiao Zhan, Jiangtao Peng, and Weiwei Sun. LitedealthWisenet: A lightweight network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2022. 2", + "[7] Ding Ding, Sundaresh Ram, and Jeffrey J. Rodríguez. Image inpainting using nonlocal texture matching and nonlinear filtering. IEEE TIP, 28(4):1705-1719, 2019. 3", + "[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In CVPR, 2022. 2, 3", + "[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, Dong Chen, and Baining Guo. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, pages 12709-12720. IEEE, 2024. 3", + "[10] James Hays and Alexei A. Efros. Scene completion using millions of photographs. ACM TOG, 26(3):4, 2007. 2", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5", + "[12] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, and Ying Shan. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR. IEEE, 2024. 3", + "[13] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM TOG, 36(4):107:1-107:14, 2017. 3", + "[14] Jitesh Jain, Yuqian Zhou, Ning Yu, and Humphrey Shi. Keys to better image inpainting: Structure and texture go hand in hand. In WACV, pages 208-217, 2023. 2", + "[15] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T. Freeman, David Salesin, Brian Curless, and Ce Liu. SLIDE: single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12498-12507. IEEE, 2021. 2", + "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3", + "[18] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Tom Duerig, and Vittorio Ferrari. The open images dataset V4: unified image classification, object detection, and visual relationship detection at scale. CoRR, abs/1811.00982, 2018. 2, 5", + "[19] Anat Levin, Assaf Zomet, and Yair Weiss. Learning how to inpaint from global image statistics. In ICCV, pages 305-312, 2003. 2", + "[20] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, pages 7757-7765. Computer Vision Foundation / IEEE, 2020. 3", + "[21] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. ZONE: zero-shot instruction-guided local editing. In CVPR, pages 6254-6263. IEEE, 2024. 3", + "[22] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. MAT: mask-aware transformer for large hole image inpainting. In CVPR, pages 10748-10758, 2022. 2, 3", + "[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11451-11461. IEEE, 2022. 3", + "[24] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-07-10. 5", + "[25] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544. IEEE Computer Society, 2016. 3", + "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 3", + "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 2, 3, 5, 6", + "[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 3", + "[29] Andranik Sargsyan, Shant Navasardyan, Xingqian Xu, and Humphrey Shi. MI-GAN: A simple baseline for image inpainting on mobile devices. In ICCV, pages 7301-7311, 2023. 2, 3", + "[30] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879. IEEE, 2024. 3", + "[31] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3", + "[32] Wenhao Sun, Benlei Cui, Xue-Mei Dong, and Jingqun Tang. Attentive eraser: Unleashing diffusion model's object re" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2426", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "moyal potential via self-attention redirection guidance. In AAAI, 2025. 2", + "[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 3172–3182, 2022. 2, 3, 5, 6", + "[34] Jingqun Tang, Su Qiao, Benlei Cui, Yuhang Ma, Sheng Zhang, and Dimitrios Kanoulas. You can even annotate text with voice: Transcription-only-supervised text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4154-4163, New York, NY, USA, 2022. Association for Computing Machinery. 2", + "[35] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369. IEEE, 2023. 3", + "[36] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Learning 3d geometry and feature consistent gaussian splatt-ting for object removal. In ECCV, pages 1-17. Springer, 2024. 2", + "[37] Daniel Winter, Matan Cohen, Shlomi Fruchtner, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. In ECCV, pages 112-129. Springer, 2024. 2", + "[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 3", + "[39] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 5", + "[40] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin C. K. Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. CoRR, abs/2312.03771, 2023. 3", + "[41] Xingqian Xu, Shant Navasardyan, Vahram Tadevosyan, Andranik Sargsyan, Yadong Mu, and Humphrey Shi. Image completion with heterogeneously filtered spectral hints. In WACV, pages 4580-4590, 2023. 2", + "[42] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 2", + "[43] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 3", + "[44] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, pages 7505-7514, 2020. 2, 3" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. CoRR, abs/2304.03246, 2023. 3, 5, 6", + "[46] Donggeun Yoon and Donghyeon Cho. CORE-MPI: consistency object removal with embedding multiplane image. In CVPR, pages 20081-20090. IEEE, 2024. 2", + "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5", + "[48] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I-Chao Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In ICLR, 2021. 2", + "[49] Hao Zhou, Chuanping Hu, Chongyang Zhang, and Shengyang Shen. Visual relationship recognition via language and position guided attention. In ICASSP, 2019. 2", + "[50] Hao Zhou, Chongyang Zhang, Yan Luo, Yanjun Chen, and Chuanping Hu. Embracing uncertainty: Decoupling and debias for robust temporal grounding. In CVPR, pages 8445-8454, 2021. 2", + "[51] Tong Zhou, Changxing Ding, Shaowen Lin, Xinchao Wang, and Dacheng Tao. Learning oracle attention for high-fidelity face completion. In CVPR, pages 7677-7686, 2020. 3", + "[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 5, 6" + ], + "bbox": [ + 514, + 90, + 905, + 501 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2427", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_model.json b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a50c5652bd10eb8054fa9461d381fb12233231c --- /dev/null +++ b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.13, + 0.845, + 0.176 + ], + "angle": 0, + "content": "Erase Diffusion: Empowering Object Removal Through Calibrating Diffusion Pathways" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.203, + 0.74, + 0.24 + ], + "angle": 0, + "content": "Yi Liu*, Hao Zhou*, Benlei Cui†, Wenxiang Shang, Ran Lin Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.242, + 0.829, + 0.258 + ], + "angle": 0, + "content": "{ly402089,baishu.zh,cuibenlei.cbl,shangwenxiang.swx,linran.lr09}@taobao.com" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.293, + 0.289, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.396, + 0.289, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.293, + 0.496, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.293, + 0.701, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.293, + 0.902, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.396, + 0.902, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.502, + 0.292, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.502, + 0.495, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.502, + 0.698, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.502, + 0.902, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.672, + 0.908, + 0.715 + ], + "angle": 0, + "content": "Figure 1. Diverse erase inpainting results produced by our proposed EraDiff, where images before and after removal are presented in pairs, and the areas to be erased in the original images have been marked. The EraDiff can eliminate targets in various complex real-world scenes while ensuring visual coherence in the generated images." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.727, + 0.327, + 0.743 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.759, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Erase inpainting, or object removal, aims to precisely remove target objects within masked regions while preserving the overall consistency of the surrounding content. Despite diffusion-based methods have made significant strides in the field of image inpainting, challenges remain regarding the emergence of unexpected objects or artifacts. We assert that the inexact diffusion pathways established by existing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.907, + 0.895 + ], + "angle": 0, + "content": "standard optimization paradigms constrain the efficacy of object removal. To tackle these challenges, we propose a novel Erase Diffusion, termed EraDiff, aimed at unleashing the potential power of standard diffusion in the context of object removal. In contrast to standard diffusion, the EraDiff adapts both the optimization paradigm and the network to improve the coherence and elimination of the erasure results. We first introduce a Chain-Rectifying Optimization (CRO) paradigm, a sophisticated diffusion process specifically designed to align with the objectives of erasure. This paradigm establishes innovative diffusion transition" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.876, + 0.321, + 0.888 + ], + "angle": 0, + "content": "* These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.888, + 0.209, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.876, + 0.321, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2418" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "pathways that simulate the gradual elimination of objects during optimization, allowing the model to accurately capture the intent of object removal. Furthermore, to mitigate deviations caused by artifacts during the sampling pathways, we develop a simple yet effective Self-Rectifying Attention (SRA) mechanism. The SRA calibrates the sampling pathways by altering self-attention activation, allowing the model to effectively bypass artifacts while further enhancing the coherence of the generated content. With this design, our proposed EraDiff achieves state-of-the-art performance on the OpenImages V5 dataset and demonstrates significant superiority in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.302, + 0.223, + 0.318 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.538 + ], + "angle": 0, + "content": "Image inpainting is a crucial technique in computer vision, aimed at reconstructing missing or damaged regions while maintaining visual coherence and contextual integrity [15, 32]. As one specialized form of image inpainting, object removal, also called erase inpainting, has broader applications in various fields such as social media [3, 50], advertising design [36], and image processing [37, 46]. However, in comparison to general image inpainting, an ideal erase inpainting model must address two critical challenges. Coherence: The model should seamlessly inpaint the masked area, ensuring consistency in lighting, content, and other relevant aspects. Elimination: The model should accurately remove objects within the masked area while preventing the generation of extraneous elements or artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Prior erase inpainting methods primarily relied on nonparametric patch sampling techniques [4, 10, 19] or Generative Adversarial Networks (GANs) [5, 8, 14, 22, 29, 33, 41, 44, 48]. These approaches often resort to filling large masked regions with repetitive patches, resulting in a lack of coherence in the generated images. In recent years, methods based on Latent Diffusion Models (LDMs) [27] have demonstrated superiority in generating more natural images. However, these methods struggle to eliminate target objects. For example, when a user seeks to erase a piece of pizza in Figure 1, the LDMs may mistakenly generate another piece of pizza instead of the expected clean plate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We attribute the above issues to the fact that standard diffusion pathways in most existing methods are not suitable for erase tasks. During optimization, most diffusion-based erase inpainting models take the original image with added noise and randomly generated masks as input, aiming to recover the original image in the presence of noise. However, this standard training paradigm only establishes a denoising process that transitions from random noise to clear images, without addressing the specific goal of object removal. Consequently, the model may follow a denoising pathway from noise to an image with objects. Based on this observation, we argue that an ideal erase inpainting model should" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "establish a diffusion pathway directly from objects to backgrounds to ensure the definitive removal of unwanted elements. Furthermore, in the early stages of denoising, the reconstructed regions are significantly influenced by the shape of the masks and the level of noise. This impact may lead to deviations in the early latent states, resulting in the emergence of artifacts. The standard self-attention mechanism tends to incorrectly regard the features of these artifacts as important information, which gradually amplifies the deviations during the subsequent denoising process. Ultimately, it leads to unexpected objects in the generated images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.71 + ], + "angle": 0, + "content": "In this paper, we present a novel erase diffusion model, termed EraDiff, specifically designed for object removal, which effectively unleashes the potential erasure power of standard diffusion. First, we introduce a Chain-Rectifying Optimization (CRO) paradigm that supports the establishment of new diffusion pathways from noise to backgrounds. Specifically, we develop a dynamic image synthesis strategy that enables the generation of a variety of dynamic images at different time steps without the need for extra data. These synthesized dynamic images effectively simulate the gradual elimination of objects during the denoising process, yielding intermediate latent states corresponding to different time steps. Next, we introduce a new dedicated optimization objective for erase inpainting. This objective guides the establishment of transitions among multiple intermediate latent states. By applying the CRO paradigm, the model learns to accurately identify the intention to erase and improve content coherence. Furthermore, to address potential deviations caused by artifacts during the early denoising process, we design a Self-Rectifying Attention (SRA) mechanism that explicitly guides the model in executing object removal more effectively. By altering self-attention activation, the SRA enhances background features while rectifying its incorrect reliance on artifacts. This replacement of the standard attention mechanism results in final generated images that appear more realistic and devoid of unexpected objects. To the best of our knowledge, our proposed method achieves state-of-the-art performance in both the public OpenImages V5 [18] dataset and large-scale real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.724, + 0.655, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Image Inpainting. Image inpainting, which involves the reconstruction of large-scale missing regions, has garnered substantial attention within the field. Traditional image inpainting methods employ heuristic patch-based propagation algorithms to borrow texture and structure from neighboring regions of the corrupted areas [4, 10, 19]. However, these methods often struggle to effectively restore areas with complex structures or semantics [42]. To tackle this issue, numerous studies have harnessed the power of deep neural networks [6, 34, 49] and proposed various en" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "hancements to GANs [25] aimed at improving the coherence of inpainted regions. Subsequent research has extensively examined different facets, including semantic context and texture [1, 7, 13], edges and contours [5, 8], as well as the design of hand-engineered architectures [20, 51]. Notably, Suvorov et al. [33] introduced FFCs into residual modules, providing a remedy for enhancing the perceptual quality of GAN-generated outputs. More recently, promising efforts have been made in applying LDMs to inpainting tasks, showcasing their potential to achieve impressive results [9, 23, 26, 31, 35, 38]. However, general image inpainting models may generate unexpected objects in the masked areas, making these LDM-based methods unsuitable for effective object removal." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.485, + 0.651 + ], + "angle": 0, + "content": "Erase Inpainting. Erase inpainting, one specialized form of image inpainting, focuses on removing unwanted content from images. The typical setup for erase inpainting is to provide an image along with an object mask as input, ultimately generating a clean background that excludes the specified object. Many studies have concentrated on improvements in model architecture [22, 40, 43, 44] and loss functions [29, 33]. While these methods effectively improve the coherence of masked regions, they still struggle with object removal. Recent works [2, 12, 21, 30, 45] have attempted to leverage textual prompts to locate erase targets within images. However, relying on textual cues to accurately identify areas for erasure has proven to be unstable. Models fail to follow textual instructions, which can lead to unfavorable results. As a result, these approaches face challenges in large-scale practical applications. Through in-depth investigation, we find that the struggle to effectively erase objects stems from the inexact construction of the erase diffusion chain. Our proposed method attempts to directly establish diffusion pathways from the noise to the background. This approach enables effective object removal without relying on additional textual information and produces one clean and natural background." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.662, + 0.231, + 0.679 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Latent Diffusion Models. In this paper, we utilize latent diffusion models (LDMs) as our erase inpainting model. LDMs are probabilistic models designed to learn a data distribution \\( p(\\pmb{x}) \\) by gradually denoising a normally distributed variable, which corresponds to learning the reverse sampling process of a fixed Markov Chain of length \\( T \\). Specifically, the forward diffusion process progressively adds noise to the initial latent state \\( \\pmb{x}_0 \\)," + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.816, + 0.483, + 0.834 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\boldsymbol {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.872 + ], + "angle": 0, + "content": "where \\(1 - \\bar{\\alpha}_t\\) is the variance schedule at step \\(t\\) and represents the level of noise, with \\(\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "According to the DDIM algorithm, the reverse sampling process can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.109, + 0.905, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} _ {p r e v} = \\sqrt {\\bar {\\alpha} _ {p r e v}} \\left(\\frac {\\boldsymbol {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) \\tag {2} \\\\ + \\sqrt {1 - \\bar {\\alpha} _ {p r e v} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t}) + \\sigma_ {t} \\epsilon_ {t}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.261 + ], + "angle": 0, + "content": "where \\(\\epsilon_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) is standard Gaussian noise. Here, \\(\\sigma_t\\) are hyper-parameters, and the reverse sampling process can be considered deterministic as \\(\\sigma_t \\to 0\\). The term \\(\\epsilon_\\theta^{(t)}(\\boldsymbol{x}_t)\\) is trained to predict the noise added to \\(\\boldsymbol{x}_t\\) by minimizing the following optimization objective" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.266, + 0.905, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (\\mathbf {0}, I)} \\left\\| \\epsilon - \\epsilon_ {\\theta} ^ {(t)} \\left(\\boldsymbol {x} _ {t}\\right) \\right\\| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Structure of Erasing Ipainting. Most advanced erasing inpainting methods are based on diffusion models. In this paper, we focus on the SD2-Inpaint model[27], which mainly consists of a Variational AutoEncoder (VAE) [17] and a U-Net network [28]. The VAE facilitates the transformation of input images into latent space. For Erasing Ipainting, the inputs typically consist of a noisy image \\( \\boldsymbol{x}_t^{ori} \\in \\mathbb{R}^{H \\times W \\times 3} \\), a random binary mask \\( \\mathbf{M} \\in \\{0,1\\}^{H \\times W} \\), and the corresponding masked image \\( \\boldsymbol{x}' = \\boldsymbol{x}_0^{ori} \\odot \\mathbf{M} \\). The U-Net network predicts the associated noise at timestamp \\( t \\) based on the Equation 3. This methodology enables the model to establish a diffusion chain between the original image and Gaussian noise, allowing for seamlessly inpainting masked regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.617 + ], + "angle": 0, + "content": "Additionally, the self-attention mechanism within U-Net networks plays a significant role in the erasing process. This mechanism aggregates information from the entire image, thereby controlling the generation of features in the masked regions during the reverse sampling process. Given one latent feature map \\( z \\in \\mathbb{R}^{h \\times w \\times c} \\), where \\( h \\), \\( w \\) and \\( c \\) are the height, width, and channel dimensions of \\( z \\) respectively, the self-attention process can be represented as follows" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.623, + 0.905, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} = \\ell_ {Q} (\\boldsymbol {z}), \\ell_ {K} (\\boldsymbol {z}), \\ell_ {V} (\\boldsymbol {z}), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.647, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.906, + 0.713 + ], + "angle": 0, + "content": "where \\(\\ell_{Q},\\ell_{K},\\ell_{V}\\) are learnable linear layers, and \\(d\\) denotes the scaling factor." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.725, + 0.605, + 0.74 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Overview. The overall architecture of our proposed Erase Diffusion, termed EraDiff, is shown in Figure 2, consisting of a Chain-Rectifying Optimization paradigm (see Section 4.1) and a Self-Rectifying Attention mechanism (see Section 4.2). This CRO paradigm introduces the design of dynamic latent states along with a dedicated optimization objective, thereby establishing novel diffusion chains that better align with the erasure objectives. Meanwhile, the SRA mechanism effectively guides the sampling process to prevent state deviations caused by artifacts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.576 + ], + "angle": 0, + "content": "Figure 2. The overview of our proposed Erase Diffusion, termed EraDiff. Left: Dynamic image synthesis. Each image is initially transformed using techniques like matting, scaling, and copy-pasting. A mix-up strategy then synthesizes a series of dynamic images \\(\\{\\tilde{x}_t^{mix}\\}\\) that simulate the gradual fading of the object. Top: Chain-Rectifying Optimization (CRO). The standard sampling pathway is prone to generating artifacts (black dashed lines). In contrast, we establish a new sampling path for erasing (red dashed lines) that better aligns the reverse sampling trajectory with a clear background. Bottom: Self-Rectifying Attention (SRA). The standard self-attention mechanism may inadvertently amplify artifacts, diverging from the expected diffusion pathway. By modifying the attention activation, we guide the model to bypass artifact regions, enhancing its focus on the background and ensuring a more accurate erase sampling path." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.367, + 0.623 + ], + "angle": 0, + "content": "4.1. Chain-Rectifying Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In the optimization process, existing erase inpainting methods typically employ randomly generated or object-based masks, predicting the noise in the masked regions using Equation 3. These masked regions may cover objects, and the model's optimization objective is to reconstruct these objects rather than eliminate them. As discussed above, such standard diffusion chains may inadvertently encourage the model to generate unexpected artifacts. An effective solution is to dedicate diffusion paths for the erase inpainting task between objects and backgrounds. By simulating the gradual fading of objects, this approach can suppress their emergence during the sampling process. Even if the sampling process does not start with standard Gaussian noise (a common enhancement trick), it can still help eliminate leaked objects within the masked region. Thus, the main challenge of this approach is to develop new latent states within the erase diffusion chains and to optimize the model to manipulate transitions among these latent states." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Dynamic Latent States. To design dedicated diffusion chains for erasure between objects and backgrounds, it is essential to obtain corresponding image pairs. However, existing public datasets often lack these paired samples, highlighting the need for an innovative data synthesis strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.335 + ], + "angle": 0, + "content": "Let \\( \\pmb{x}_0^{ori} \\) denote the original image, and a trained matting model is employed to segment the primary object within it. We then apply various image transformation techniques, including rotation and scaling, to modify the segmented objects. These transformed objects are randomly pasted onto the background region, resulting in a new synthesized image, \\( \\pmb{x}_0^{obj} \\). This method allows us to construct a training dataset of object-background image pairs at a relatively low cost. Furthermore, to effectively simulate the gradual fading of the objects, we input dynamic images \\( \\tilde{\\pmb{x}}_t^{mix} \\) for each time step \\( t \\) during the training process," + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.345, + 0.907, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {x}} _ {t} ^ {m i x} = (1 - \\lambda_ {t}) \\boldsymbol {x} _ {0} ^ {o r i} + \\lambda_ {t} \\boldsymbol {x} _ {0} ^ {o b j}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.906, + 0.482 + ], + "angle": 0, + "content": "where the decreasing sequence \\(\\lambda_{:T} \\in [0,1]^T\\) controls the mix-up level of object-background image pairs. In the case of \\(t = 0\\), \\(\\tilde{\\pmb{x}}_0^{mix}\\) corresponds to the original image \\(\\pmb{x}_0^{ori}\\). These dynamic mix-up images assist the model in understanding the smooth transition from objects to backgrounds. Ultimately, based on Equation 1, we can derive a set of new latent states \\(\\pmb{x}_t^{mix}\\) for the self-rectifying diffusion chains," + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.493, + 0.907, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} ^ {m i x} = \\sqrt {\\bar {\\alpha} _ {t}} \\tilde {\\boldsymbol {x}} _ {t} ^ {m i x} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.65, + 0.537 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Optimization Objective. The traditional diffusion process starts from a fixed \\( \\boldsymbol{x}_0^{ori} \\), and each intermediate latent state \\( \\boldsymbol{x}_t^{ori} \\) can be obtained directly using Equation 1. However, in the new self-rectifying diffusion chains, each latent state \\( \\boldsymbol{x}_t^{mix} \\) is derived from the dynamic image \\( \\tilde{\\boldsymbol{x}}_t^{mix} \\), making it infeasible to apply the standard optimization objective in Equation 3. To mitigate this challenge, we propose a novel optimization objective that enhances the alignment between the model's predicted distribution and the true distribution, as illustrated in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.782 + ], + "angle": 0, + "content": "Specifically, the new loss function aims to minimize the distance between the model-predicted latent states \\(\\hat{\\pmb{x}}_t^{mix}\\) and the true states \\(\\pmb{x}_t^{mix}\\). It allows the model to gradually alter its parameters to adapt to the new distribution shifts. Given one latent state \\(\\pmb{x}_t^{mix}\\), we can derive the model-predicted latent state \\(\\hat{\\pmb{x}}_{t - \\gamma}^{mix}\\) at the previous \\(\\gamma\\) time step," + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.794, + 0.905, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\theta} (\\hat {\\boldsymbol {x}} _ {t - \\gamma} ^ {m i x} | \\boldsymbol {x} _ {t} ^ {m i x}) = \\sqrt {\\bar {\\alpha} _ {t - \\gamma}} \\left(\\frac {\\boldsymbol {x} _ {t} ^ {m i x} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t} ^ {m i x})}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) \\\\ + \\sqrt {1 - \\bar {\\alpha} _ {t - \\gamma}} \\epsilon_ {\\theta} ^ {(t)} \\left(\\boldsymbol {x} _ {t} ^ {m i x}\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.905, + 0.903 + ], + "angle": 0, + "content": "where \\(\\gamma \\in (0,\\gamma_{m})\\). Given the original image \\(\\pmb{x}_0^{ori}\\) and the synthesized image \\(\\pmb{x}_0^{obj}\\), the corresponding true state \\(\\pmb{x}_{t - \\gamma}^{mix}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2421" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.102, + 0.095, + 0.404, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: Chain-Rectifying Optimization" + }, + { + "type": "algorithm", + "bbox": [ + 0.113, + 0.117, + 0.482, + 0.297 + ], + "angle": 0, + "content": "repeat \n\\(\\pmb{x}_0^{ori}\\sim q(x)\\) \n\\(t\\sim\\) Uniform \\((\\{1,\\dots ,T\\})\\) \n\\(\\gamma \\sim\\) Uniform \\((\\{1,\\ldots ,\\gamma_{m}\\})\\) \n\\(\\epsilon \\sim \\mathcal{N}(0,I)\\) \n\\(\\pmb{x}_0^{obj}\\gets \\pmb{x}_0^{ori}\\) image transformation; \n\\(\\pmb{x}_{t}^{mix},\\pmb{x}_{t - \\gamma}^{mix}\\leftarrow \\pmb{x}_{0}^{ori},\\pmb{x}_{0}^{obj}\\) based on Equ 6-7; \n\\(\\hat{\\pmb{x}}_{t - \\gamma}^{mix}\\longleftarrow \\pmb{x}_{t}^{mix},\\epsilon_{\\theta}^{(t)}\\) based on Equ 8; \nTake gradient descent step on \n\\(\\nabla_{\\theta}\\left\\| \\pmb{x}_{t - \\gamma}^{mix} - \\hat{\\pmb{x}}_{t - \\gamma}^{mix}\\right\\| ^2\\) based on Equ 9 \nuntil converged;" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.483, + 0.359 + ], + "angle": 0, + "content": "can be obtained using Equation 6 and 7. Finally, we define the new optimization objective as follows" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.367, + 0.484, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\gamma \\sim \\operatorname {U n i f o r m} (1, \\gamma_ {m}), t} \\left\\| \\boldsymbol {x} _ {t - \\gamma} ^ {m i x} - p _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {t - \\gamma} ^ {m i x} \\mid \\boldsymbol {x} _ {t} ^ {m i x}\\right) \\right\\| _ {2} ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.4, + 0.32, + 0.416 + ], + "angle": 0, + "content": "4.2. Self-Rectifying Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.603 + ], + "angle": 0, + "content": "We have established a self-rectifying diffusion pathway to guide the model toward object removal. However, information leakage from the mask's shape can still lead to artifacts during the early stages of denoising, resulting in latent state shifts. The self-attention mechanism tends to give the masked region stronger attention to itself rather than to the background. This phenomenon can continuously amplify artifacts along the reverse sampling path, ultimately leading to a deviation from the object removal direction. An intuitive solution is to alter the current attention layers for path calibration to mitigate the above risk of generating unexpected objects due to deviations from specific states." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.816 + ], + "angle": 0, + "content": "Based on our observations, we believe that the generation of the foreground region should rely more on backgrounds rather than focusing on itself. Additionally, the background region remains visible and thus should not be affected by the content of the foregrounds. By altering the self-attention activation, it is possible to ignore the negative effects of artifacts while emphasizing the background, thereby further enhancing the coherence and elimination of the generated content. Thus, we propose a simple yet effective Self-Rectifying Attention mechanism to replace the standard self-attention mechanism. Specifically, we first downsample and flatten the image mask \\(\\mathbf{M}\\) to obtain the corresponding mask vector \\(m\\in \\{0,1\\}^{wh}\\). We then design an extended mask \\(m^{\\prime}\\in \\{-inf,1\\}^{wh\\times wh}\\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.823, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\nm _ {i, j} ^ {\\prime} = \\left\\{ \\begin{array}{l l} 1, & m _ {i} = 0 \\text {o r} m _ {j} = 0 \\\\ - i n f, & \\text {e l s e} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The extended mask is subsequently applied directly to the corresponding attention activations, effectively suppressing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "objects within the masked regions while enhancing background features. This is represented mathematically as" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.135, + 0.907, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S R A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}} \\cdot m ^ {\\prime}\\right) \\mathbf {V}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.907, + 0.273 + ], + "angle": 0, + "content": "This mechanism enhances the model's capacity to perceive background information without any additional computational cost, while also effectively diminishing the interference of artifacts on the final generated results. Consequently, this improvement ensures that the sampling process can calibrate itself to the target direction of object removal." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.29, + 0.646, + 0.307 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.316, + 0.704, + 0.332 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Benchmark Datasets. We conducted a thorough evaluation of our proposed pipeline utilizing the publicly available OpenImages V5 segmentation dataset [18]. Each instance within this dataset comprises the original image, a corresponding segmentation mask, segmentation bounding boxes, and associated class labels, thereby enabling a rigorous comparative analysis against several baseline models. For our experiments, we randomly selected a subset of 10,000 samples from the OpenImages V5 test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Comparison Baselines. We have chosen several cutting-edge image inpainting methods as our baselines for comparison. These include two mask-guided approaches: SD2-Inpaint [27] and LaMa [33], alongside two text-guided techniques: Inst-Inpaint [45] and PowerPaint [52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Evaluation Metrics. In accordance with LaMa [33], we employ two primary evaluation metrics: Fréchet Inception Distance (FID) [11] and Learned Perceptual Image Patch Similarity (LPIPS) [47]. These metrics are adept at assessing the overall visual coherence of the inpainted images. To provide a more nuanced evaluation of the quality of content generated within the masked regions, we introduce the Local FID metric [39], which allows for a detailed assessment of local visual fidelity. Furthermore, we augment our evaluation framework by incorporating analyses derived from GPT-4o [24] along with expert human annotations, thereby examining the effectiveness of these erasure-targeted models in eliminating objects and artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Training Details. Our proposed pipeline builds upon the foundations established by SD2-Inpaint and is implemented using PyTorch along with the Diffusers library. During the training phase, we employed the Adam [16] optimizer with a learning rate set to \\(3 \\times 10^{-6}\\) on the OpenImages V5 training set. To simplify parameters, the increasing sequence \\(\\lambda_{:T}\\) follows the same schedule as the sequence \\(1 - \\bar{\\alpha}_{:T}\\). For loss computation, we ensured that the time intervals between two timestamps did not exceed \\(\\gamma_{m} = 100\\). All experiments were conducted on NVIDIA A100 GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2422" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.905, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.566, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of OpenImages V5 dataset compared among SD2-Inpaint [27], SD2-Inpaint with prompt guidance [27], PowerPaint [52], Inst-Inpaint [45], LaMa [33], and our approach." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.611, + 0.481, + 0.734 + ], + "angle": 0, + "content": "
MethodFID↓LPIPS↓Local FID↓
SD2-Inpaint3.8050.3018.852
SD2-Inpaint*†4.0190.3087.194
PowerPaint6.0270.28910.021
Inst-Inpaint11.4230.41043.472
LaMa7.5330.2196.091
EraDiff (ours)6.5400.1923.799
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.787 + ], + "angle": 0, + "content": "Table 1. Quantitative assessment of various erase inpainting models on the OpenImages V5 dataset. Optimal results are highlighted in bold, with runner-up performance underlined." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.455, + 0.831 + ], + "angle": 0, + "content": "5.2. Qualitative and Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The quantitative evaluation results of our proposed method, in comparison to several established models, are summarized in Table 1. Additionally, we present the experimental findings when the prompt is used as supplementary guid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.854 + ], + "angle": 0, + "content": "ance for SD2-Inpaint (abbreviated as SD2-Inpaint*, consistent throughout the paper). As indicated in the table, our model is in the mid-range of all baseline models in terms of the FID score, yet it significantly surpasses the others in the Local FID metric. This observation highlighting its remarkable ability to generate visually coherent results specifically within the designated erased regions, all the while maintaining a commendable level of visual coherence throughout the entire image. Additionally, our method records the highest performance on the LPIPS metric, suggesting that the images produced exhibit enhanced visual fidelity following the removal process. Figure 3 illustrates relevant visualization results, clearly demonstrating the model's robustness in addressing challenges such as the presence of objects within the image that closely resemble the target erasure region, as well as extensive areas that necessitate removal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "It is essential to highlight that both the FID and LPIPS metrics primarily evaluate the visual coherence and aesthetic quality of the final generated images. However, these" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2423" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.087, + 0.482, + 0.164 + ], + "angle": 0, + "content": "
MethodSuperiorComparableInferior
SD2-Inpaint1.03%18.07%80.90%
SD2-Inpaint*2.20%24.79%73.01%
LaMa13.17%35.29%51.54%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Table 2. Quantitative results of OpenImages V5 dataset among SD2-Inpaint, SD2-Inpaint*, LaMa, and EraDiff. This table delineates a comparative analysis of the elimination performance results obtained by these methodologies relative to ours, highlighting whether their outcomes are superior, comparable, or inferior to those achieved by our approach." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.274, + 0.476, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.548 + ], + "angle": 0, + "content": "Figure 4. Results from the user study. EraDiff demonstrates enhanced performance, as indicated by its higher mean scores in both elimination and coherence evaluations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.673 + ], + "angle": 0, + "content": "metrics do not provide a direct assessment of the effectiveness regarding object or artifact elimination in the denoising process. This limitation is clearly illustrated in Figure 3, which demonstrates that despite the SD2-Inpaint model achieving the highest FID score among all evaluated models, it frequently fails to adequately eliminate unwanted objects or artifacts from the designated masked areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the baseline models and our proposed method in eliminating objects and artifacts, we systematically selected the top-performing models: SD2-Inpaint, SD2-Inpaint*, and LaMa, as indicated in Table 1. A comparative analysis was performed through pairwise comparisons against our method, utilizing GPT-4o to generate objective evaluations focused on identifying superior results among the competing models. The outcomes of this analysis are described in Table 2. Furthermore, we undertook a user study that involved 20 experts that was tasked with appraising the effectiveness of the erasure and the visual aesthetics of the generated results, using a scoring system ranging from 1 to 10; where a score of 1 reflects poor performance and a score of 10 indicates exceptional quality. The outcomes of this assessment are illustrated in Fig-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.35, + 0.907, + 0.393 + ], + "angle": 0, + "content": "Figure 5. Visualization of EraDiff's performance across a diverse array of in-the-wild scenarios: animated imagery, e-commerce content, oil paintings, and glasses-free 3D visuals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.514 + ], + "angle": 0, + "content": "ure 4. The results from both experimental methodologies distinctly demonstrate that our proposed approach significantly outperforms the competitors in effectively eliminating target objects. This conclusion is further corroborated by the finding that, despite the lower FID score recorded for SD2-Inpaint, its efficacy in object removal does not surpass that of our innovative model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Furthermore, we find that EraDiff shows superior performance across a diverse array of in-the-wild scenarios, as illustrated in Figure 5. A more detailed analysis of this will be provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.667, + 0.607 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To evaluate the impact of calibrating the sampling pathway on the performance of erase inpainting, we conducted four extensive experiments: each focusing on the removal of CRO, SRA, and the simultaneous exclusion of both CRO and SRA, all based on the EraDiff model. Furthermore, we assessed the model's performance without the incorporation of the mix-up strategy during training, wherein the value of \\(\\lambda_{t}\\) was held constant. For our evaluation, we employed the GPT-4o to quantify the success rate of object elimination. Results are detailed in Table 3. The results indicate that the removal of either the CRO or SRA components, as well as their simultaneous exclusion, leads to a substantial deterioration in both the visual coherence of the erased images and the efficacy of object elimination. Particularly noteworthy is the fact that the omission of CRO results in a more pronounced decline in overall performance. Additionally, the absence of the mix-up strategy produced considerable fluctuations in training loss and impeded model convergence. The instability observed can be attributed to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2424" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.468, + 0.198 + ], + "angle": 0, + "content": "
MethodLocal FID↓GPT score↑
EraDiff3.79983.43%
-w/o CRO5.71372.96%
-w/o SRA4.95078.54%
-w/o CRO ∪ SRA8.85227.80%
-w/o mix-upNaNNaN
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.206, + 0.484, + 0.237 + ], + "angle": 0, + "content": "Table 3. Results of the ablation study highlight the individual and combined effects of CRO and SRA methodologies." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.248, + 0.484, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.438 + ], + "angle": 0, + "content": "Figure 6. Visual examples for the ablation study comparing baseline, baseline with CRO, baseline with SRA, and baseline with both CRO and SRA, displayed left to right." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.583 + ], + "angle": 0, + "content": "the limited noise present at earlier time steps, which complicates the model's ability to accurately predict the entirety of the masked region as noise. Figure 6 presents the visual outcomes of these ablation experiments, clearly illustrating the effectiveness of the calibration sampling pathway in enhancing the erasure task. In this regard, SRA serves a corrective function, effectively addressing specific anomalous denoising artifacts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.363, + 0.607 + ], + "angle": 0, + "content": "5.4. How the CRO and SRA work?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.659 + ], + "angle": 0, + "content": "To further elucidate the roles of the CRO and SRA methodologies in the denoising process of erase inpainting, we established two types of extended experimental setups." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To investigate the mechanism underlying CRO, we established an extremely challenging condition by setting the denoising strength to 0.6. This configuration allows for significant leakage of original image information, including the target objects intended for removal. This leakage can lead to considerable disruptions, resulting in artifacts within the erased regions and potentially even the restoration of the erasure-targeted objects. As illustrated in Figure 7, we observed the denoising processes of the EraDiff alongside the baseline model (i.e., SD2-Inpaint) under this condition. Notably, as the denoising process progressed, the EraDiff model effectively concealed the object artifacts in the erased region, whereas the baseline model highlighted it. These experimental findings provide further validation of our earlier discussion in Section 4, asserting that EraDiff accomplishes the erasure task by traversing a sample pathway that closely" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.906, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.213, + 0.907, + 0.243 + ], + "angle": 0, + "content": "Figure 7. Comparison of EraDiff's and baseline model's denoising process with strength set to 0.6." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.255, + 0.905, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.427, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Figure 8. Visualization of heatmaps representing attention block outputs in the presence and absence of the SRA mechanism." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.481, + 0.625, + 0.496 + ], + "angle": 0, + "content": "resembles \\(x_{t}^{mix}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.693 + ], + "angle": 0, + "content": "To reveal the mechanism of the SRA approach, we examined a misleading scenario where multiple similar objects exist within an image, necessitating the removal of one specific object. The heatmaps generated before and after the attention block for both the EraDiff and the baseline model are presented in Figure 8. The results indicate that the baseline model tends to focus on similar objects, while our proposed method shifts its attention to the background. This strategic focus allows the model to extract critical information from the background rather than from the foreground, thereby facilitating the elimination of the target object while maintaining coherence between the erased region and the surrounding background." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.725 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present the EraDiff, which enhances object elimination while maintaining visual coherence in erase inpainting. By introducing a CRO paradigm, EraDiff establishes innovative diffusion pathways that facilitate a gradual removal of objects, allowing the model to better understand the erasure intent. Moreover, the SRA mechanism effectively reduces artifacts during the sampling process. With inclusive experiments, we demonstrate that these advancements significantly improve performance in challenging scenarios, positioning EraDiff as a valuable contribution to the field of erase inpainting." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2425" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Fazil Altinel, Mete Ozay, and Takayuki Okatani. Deep structured energy-based image inpainting. In ICPR, pages 423-428. IEEE Computer Society, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.484, + 0.185 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.187, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: text-driven layered image and video editing. In ECCV, pages 707-723. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B. Goldman. Patchmatch: a randomized correspondence algorithm for structural image editing. ACM TOG, 28(3):24, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[5] Chenjie Cao, Qiaole Dong, and Yanwei Fu. ZITS++: image inpainting by improving the incremental transformer on structural priors. IEEE TPAMI, 45(10):12667-12684, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[6] Benlei Cui, Xue-Mei Dong, Qiaoqiao Zhan, Jiangtao Peng, and Weiwei Sun. LitedealthWisenet: A lightweight network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[7] Ding Ding, Sundaresh Ram, and Jeffrey J. Rodríguez. Image inpainting using nonlocal texture matching and nonlinear filtering. IEEE TIP, 28(4):1705-1719, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, Dong Chen, and Baining Guo. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, pages 12709-12720. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.551, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[10] James Hays and Alexei A. Efros. Scene completion using millions of photographs. ACM TOG, 26(3):4, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.58, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[12] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, and Ying Shan. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[13] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM TOG, 36(4):107:1-107:14, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[14] Jitesh Jain, Yuqian Zhou, Ning Yu, and Humphrey Shi. Keys to better image inpainting: Structure and texture go hand in hand. In WACV, pages 208-217, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[15] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T. Freeman, David Salesin, Brian Curless, and Ce Liu. SLIDE: single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12498-12507. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[18] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Tom Duerig, and Vittorio Ferrari. The open images dataset V4: unified image classification, object detection, and visual relationship detection at scale. CoRR, abs/1811.00982, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[19] Anat Levin, Assaf Zomet, and Yair Weiss. Learning how to inpaint from global image statistics. In ICCV, pages 305-312, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[20] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, pages 7757-7765. Computer Vision Foundation / IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[21] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. ZONE: zero-shot instruction-guided local editing. In CVPR, pages 6254-6263. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[22] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. MAT: mask-aware transformer for large hole image inpainting. In CVPR, pages 10748-10758, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11451-11461. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[24] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-07-10. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[25] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544. IEEE Computer Society, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[29] Andranik Sargsyan, Shant Navasardyan, Xingqian Xu, and Humphrey Shi. MI-GAN: A simple baseline for image inpainting on mobile devices. In ICCV, pages 7301-7311, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[30] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[31] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Wenhao Sun, Benlei Cui, Xue-Mei Dong, and Jingqun Tang. Attentive eraser: Unleashing diffusion model's object re" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2426" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "moyal potential via self-attention redirection guidance. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 3172–3182, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.483, + 0.293 + ], + "angle": 0, + "content": "[34] Jingqun Tang, Su Qiao, Benlei Cui, Yuhang Ma, Sheng Zhang, and Dimitrios Kanoulas. You can even annotate text with voice: Transcription-only-supervised text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4154-4163, New York, NY, USA, 2022. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.391 + ], + "angle": 0, + "content": "[35] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.394, + 0.482, + 0.449 + ], + "angle": 0, + "content": "[36] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Learning 3d geometry and feature consistent gaussian splatt-ting for object removal. In ECCV, pages 1-17. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.452, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[37] Daniel Winter, Matan Cohen, Shlomi Fruchtner, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. In ECCV, pages 112-129. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.569, + 0.482, + 0.623 + ], + "angle": 0, + "content": "[39] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.627, + 0.482, + 0.682 + ], + "angle": 0, + "content": "[40] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin C. K. Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. CoRR, abs/2312.03771, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[41] Xingqian Xu, Shant Navasardyan, Vahram Tadevosyan, Andranik Sargsyan, Yadong Mu, and Humphrey Shi. Image completion with heterogeneously filtered spectral hints. In WACV, pages 4580-4590, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.482, + 0.797 + ], + "angle": 0, + "content": "[42] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[43] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[44] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, pages 7505-7514, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[45] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. CoRR, abs/2304.03246, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[46] Donggeun Yoon and Donghyeon Cho. CORE-MPI: consistency object removal with embedding multiplane image. In CVPR, pages 20081-20090. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[48] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I-Chao Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[49] Hao Zhou, Chuanping Hu, Chongyang Zhang, and Shengyang Shen. Visual relationship recognition via language and position guided attention. In ICASSP, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[50] Hao Zhou, Chongyang Zhang, Yan Luo, Yanjun Chen, and Chuanping Hu. Embracing uncertainty: Decoupling and debias for robust temporal grounding. In CVPR, pages 8445-8454, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.405, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[51] Tong Zhou, Changxing Ding, Shaowen Lin, Xinchao Wang, and Dacheng Tao. Learning oracle attention for high-fidelity face completion. In CVPR, pages 7677-7686, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2427" + } + ] +] \ No newline at end of file diff --git a/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_origin.pdf b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..41c51ee6307da36cac1d0da76f1de3eaa1de2a08 --- /dev/null +++ b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/eb77825c-6cbb-4c39-948d-52efc39ace4f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78abaefb03442e130996882af7833952b16faeaac88e5a708508e6462d7036e4 +size 6668425 diff --git a/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/full.md b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/full.md new file mode 100644 index 0000000000000000000000000000000000000000..58a19bdb88d14ad40a171b56019f3ade9aed5b85 --- /dev/null +++ b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/full.md @@ -0,0 +1,299 @@ +# Erase Diffusion: Empowering Object Removal Through Calibrating Diffusion Pathways + +Yi Liu*, Hao Zhou*, Benlei Cui†, Wenxiang Shang, Ran Lin Alibaba Group + +{ly402089,baishu.zh,cuibenlei.cbl,shangwenxiang.swx,linran.lr09}@taobao.com + +![](images/9012fe4d2c76d6425a890207ec4dda4fca8af07c8402537114a83df786fa68e2.jpg) + +![](images/9cf332200c98842a53ab2c8c9aac846ae491cf3249e33a069ad25cfabd9344e6.jpg) + +![](images/a34409901c6394bf90a22338c1058e9efee906a2bb5f94268d252e96aa3936a0.jpg) + +![](images/c4df7140a17c3ba39d51bae3a8f2580ba3948b56ec3f7ae8416f02349d668b78.jpg) + +![](images/7c4558ad4aa3d86c08071d1759c17b99380e815b154b2c34675817f1dcabe776.jpg) + +![](images/7b085af2e69fc4034c70413fb5df5a155ebc8e10ad8d83d0a1482e9c6eb45742.jpg) + +![](images/bbad07ff923cf89d006fe593ae8ae0ec30a826e844348a4dc68009846188a954.jpg) +Figure 1. Diverse erase inpainting results produced by our proposed EraDiff, where images before and after removal are presented in pairs, and the areas to be erased in the original images have been marked. The EraDiff can eliminate targets in various complex real-world scenes while ensuring visual coherence in the generated images. + +![](images/bcd965ade37fa1876b3bb3addedb71b91e0b63033a5ec47d5b80eb70a72c2817.jpg) + +![](images/5f64599edb3f1420c16da5fbf8bd07cc4dcc8e4bf50ab8b783d83a8e393c14f8.jpg) + +![](images/50a1fe9e7a07290dac8a2e84117ed3890217db55ba42f54f9e1ab5c9401a75eb.jpg) + +# Abstract + +Erase inpainting, or object removal, aims to precisely remove target objects within masked regions while preserving the overall consistency of the surrounding content. Despite diffusion-based methods have made significant strides in the field of image inpainting, challenges remain regarding the emergence of unexpected objects or artifacts. We assert that the inexact diffusion pathways established by existing + +standard optimization paradigms constrain the efficacy of object removal. To tackle these challenges, we propose a novel Erase Diffusion, termed EraDiff, aimed at unleashing the potential power of standard diffusion in the context of object removal. In contrast to standard diffusion, the EraDiff adapts both the optimization paradigm and the network to improve the coherence and elimination of the erasure results. We first introduce a Chain-Rectifying Optimization (CRO) paradigm, a sophisticated diffusion process specifically designed to align with the objectives of erasure. This paradigm establishes innovative diffusion transition + +pathways that simulate the gradual elimination of objects during optimization, allowing the model to accurately capture the intent of object removal. Furthermore, to mitigate deviations caused by artifacts during the sampling pathways, we develop a simple yet effective Self-Rectifying Attention (SRA) mechanism. The SRA calibrates the sampling pathways by altering self-attention activation, allowing the model to effectively bypass artifacts while further enhancing the coherence of the generated content. With this design, our proposed EraDiff achieves state-of-the-art performance on the OpenImages V5 dataset and demonstrates significant superiority in real-world scenarios. + +# 1. Introduction + +Image inpainting is a crucial technique in computer vision, aimed at reconstructing missing or damaged regions while maintaining visual coherence and contextual integrity [15, 32]. As one specialized form of image inpainting, object removal, also called erase inpainting, has broader applications in various fields such as social media [3, 50], advertising design [36], and image processing [37, 46]. However, in comparison to general image inpainting, an ideal erase inpainting model must address two critical challenges. Coherence: The model should seamlessly inpaint the masked area, ensuring consistency in lighting, content, and other relevant aspects. Elimination: The model should accurately remove objects within the masked area while preventing the generation of extraneous elements or artifacts. + +Prior erase inpainting methods primarily relied on nonparametric patch sampling techniques [4, 10, 19] or Generative Adversarial Networks (GANs) [5, 8, 14, 22, 29, 33, 41, 44, 48]. These approaches often resort to filling large masked regions with repetitive patches, resulting in a lack of coherence in the generated images. In recent years, methods based on Latent Diffusion Models (LDMs) [27] have demonstrated superiority in generating more natural images. However, these methods struggle to eliminate target objects. For example, when a user seeks to erase a piece of pizza in Figure 1, the LDMs may mistakenly generate another piece of pizza instead of the expected clean plate. + +We attribute the above issues to the fact that standard diffusion pathways in most existing methods are not suitable for erase tasks. During optimization, most diffusion-based erase inpainting models take the original image with added noise and randomly generated masks as input, aiming to recover the original image in the presence of noise. However, this standard training paradigm only establishes a denoising process that transitions from random noise to clear images, without addressing the specific goal of object removal. Consequently, the model may follow a denoising pathway from noise to an image with objects. Based on this observation, we argue that an ideal erase inpainting model should + +establish a diffusion pathway directly from objects to backgrounds to ensure the definitive removal of unwanted elements. Furthermore, in the early stages of denoising, the reconstructed regions are significantly influenced by the shape of the masks and the level of noise. This impact may lead to deviations in the early latent states, resulting in the emergence of artifacts. The standard self-attention mechanism tends to incorrectly regard the features of these artifacts as important information, which gradually amplifies the deviations during the subsequent denoising process. Ultimately, it leads to unexpected objects in the generated images. + +In this paper, we present a novel erase diffusion model, termed EraDiff, specifically designed for object removal, which effectively unleashes the potential erasure power of standard diffusion. First, we introduce a Chain-Rectifying Optimization (CRO) paradigm that supports the establishment of new diffusion pathways from noise to backgrounds. Specifically, we develop a dynamic image synthesis strategy that enables the generation of a variety of dynamic images at different time steps without the need for extra data. These synthesized dynamic images effectively simulate the gradual elimination of objects during the denoising process, yielding intermediate latent states corresponding to different time steps. Next, we introduce a new dedicated optimization objective for erase inpainting. This objective guides the establishment of transitions among multiple intermediate latent states. By applying the CRO paradigm, the model learns to accurately identify the intention to erase and improve content coherence. Furthermore, to address potential deviations caused by artifacts during the early denoising process, we design a Self-Rectifying Attention (SRA) mechanism that explicitly guides the model in executing object removal more effectively. By altering self-attention activation, the SRA enhances background features while rectifying its incorrect reliance on artifacts. This replacement of the standard attention mechanism results in final generated images that appear more realistic and devoid of unexpected objects. To the best of our knowledge, our proposed method achieves state-of-the-art performance in both the public OpenImages V5 [18] dataset and large-scale real-world scenarios. + +# 2. Related Work + +Image Inpainting. Image inpainting, which involves the reconstruction of large-scale missing regions, has garnered substantial attention within the field. Traditional image inpainting methods employ heuristic patch-based propagation algorithms to borrow texture and structure from neighboring regions of the corrupted areas [4, 10, 19]. However, these methods often struggle to effectively restore areas with complex structures or semantics [42]. To tackle this issue, numerous studies have harnessed the power of deep neural networks [6, 34, 49] and proposed various en + +hancements to GANs [25] aimed at improving the coherence of inpainted regions. Subsequent research has extensively examined different facets, including semantic context and texture [1, 7, 13], edges and contours [5, 8], as well as the design of hand-engineered architectures [20, 51]. Notably, Suvorov et al. [33] introduced FFCs into residual modules, providing a remedy for enhancing the perceptual quality of GAN-generated outputs. More recently, promising efforts have been made in applying LDMs to inpainting tasks, showcasing their potential to achieve impressive results [9, 23, 26, 31, 35, 38]. However, general image inpainting models may generate unexpected objects in the masked areas, making these LDM-based methods unsuitable for effective object removal. + +Erase Inpainting. Erase inpainting, one specialized form of image inpainting, focuses on removing unwanted content from images. The typical setup for erase inpainting is to provide an image along with an object mask as input, ultimately generating a clean background that excludes the specified object. Many studies have concentrated on improvements in model architecture [22, 40, 43, 44] and loss functions [29, 33]. While these methods effectively improve the coherence of masked regions, they still struggle with object removal. Recent works [2, 12, 21, 30, 45] have attempted to leverage textual prompts to locate erase targets within images. However, relying on textual cues to accurately identify areas for erasure has proven to be unstable. Models fail to follow textual instructions, which can lead to unfavorable results. As a result, these approaches face challenges in large-scale practical applications. Through in-depth investigation, we find that the struggle to effectively erase objects stems from the inexact construction of the erase diffusion chain. Our proposed method attempts to directly establish diffusion pathways from the noise to the background. This approach enables effective object removal without relying on additional textual information and produces one clean and natural background. + +# 3. Preliminaries + +Latent Diffusion Models. In this paper, we utilize latent diffusion models (LDMs) as our erase inpainting model. LDMs are probabilistic models designed to learn a data distribution $p(\pmb{x})$ by gradually denoising a normally distributed variable, which corresponds to learning the reverse sampling process of a fixed Markov Chain of length $T$ . Specifically, the forward diffusion process progressively adds noise to the initial latent state $\pmb{x}_0$ , + +$$ +q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0}, (1 - \bar {\alpha} _ {t}) \boldsymbol {I}\right), \tag {1} +$$ + +where $1 - \bar{\alpha}_t$ is the variance schedule at step $t$ and represents the level of noise, with $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$ . + +According to the DDIM algorithm, the reverse sampling process can be defined as + +$$ +\begin{array}{l} \boldsymbol {x} _ {p r e v} = \sqrt {\bar {\alpha} _ {p r e v}} \left(\frac {\boldsymbol {x} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} ^ {(t)} (\boldsymbol {x} _ {t})}{\sqrt {\bar {\alpha} _ {t}}}\right) \tag {2} \\ + \sqrt {1 - \bar {\alpha} _ {p r e v} - \sigma_ {t} ^ {2}} \epsilon_ {\theta} ^ {(t)} (\boldsymbol {x} _ {t}) + \sigma_ {t} \epsilon_ {t}, \\ \end{array} +$$ + +where $\epsilon_t \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ is standard Gaussian noise. Here, $\sigma_t$ are hyper-parameters, and the reverse sampling process can be considered deterministic as $\sigma_t \to 0$ . The term $\epsilon_\theta^{(t)}(\boldsymbol{x}_t)$ is trained to predict the noise added to $\boldsymbol{x}_t$ by minimizing the following optimization objective + +$$ +\min _ {\theta} \mathbb {E} _ {\epsilon \sim \mathcal {N} (\mathbf {0}, I)} \left\| \epsilon - \epsilon_ {\theta} ^ {(t)} \left(\boldsymbol {x} _ {t}\right) \right\| _ {2} ^ {2}. \tag {3} +$$ + +Structure of Erasing Ipainting. Most advanced erasing inpainting methods are based on diffusion models. In this paper, we focus on the SD2-Inpaint model[27], which mainly consists of a Variational AutoEncoder (VAE) [17] and a U-Net network [28]. The VAE facilitates the transformation of input images into latent space. For Erasing Ipainting, the inputs typically consist of a noisy image $\boldsymbol{x}_t^{ori} \in \mathbb{R}^{H \times W \times 3}$ , a random binary mask $\mathbf{M} \in \{0,1\}^{H \times W}$ , and the corresponding masked image $\boldsymbol{x}' = \boldsymbol{x}_0^{ori} \odot \mathbf{M}$ . The U-Net network predicts the associated noise at timestamp $t$ based on the Equation 3. This methodology enables the model to establish a diffusion chain between the original image and Gaussian noise, allowing for seamlessly inpainting masked regions. + +Additionally, the self-attention mechanism within U-Net networks plays a significant role in the erasing process. This mechanism aggregates information from the entire image, thereby controlling the generation of features in the masked regions during the reverse sampling process. Given one latent feature map $z \in \mathbb{R}^{h \times w \times c}$ , where $h$ , $w$ and $c$ are the height, width, and channel dimensions of $z$ respectively, the self-attention process can be represented as follows + +$$ +\mathbf {Q}, \mathbf {K}, \mathbf {V} = \ell_ {Q} (\boldsymbol {z}), \ell_ {K} (\boldsymbol {z}), \ell_ {V} (\boldsymbol {z}), \tag {4} +$$ + +$$ +\operatorname {S A} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {\top}}{\sqrt {d}}\right) \mathbf {V}, \tag {5} +$$ + +where $\ell_{Q},\ell_{K},\ell_{V}$ are learnable linear layers, and $d$ denotes the scaling factor. + +# 4. Method + +Overview. The overall architecture of our proposed Erase Diffusion, termed EraDiff, is shown in Figure 2, consisting of a Chain-Rectifying Optimization paradigm (see Section 4.1) and a Self-Rectifying Attention mechanism (see Section 4.2). This CRO paradigm introduces the design of dynamic latent states along with a dedicated optimization objective, thereby establishing novel diffusion chains that better align with the erasure objectives. Meanwhile, the SRA mechanism effectively guides the sampling process to prevent state deviations caused by artifacts. + +![](images/daf41fa0b78b4a11615ca8cf467fbe269d34071c6fae5815d754f1db6f9ed420.jpg) +Figure 2. The overview of our proposed Erase Diffusion, termed EraDiff. Left: Dynamic image synthesis. Each image is initially transformed using techniques like matting, scaling, and copy-pasting. A mix-up strategy then synthesizes a series of dynamic images $\{\tilde{x}_t^{mix}\}$ that simulate the gradual fading of the object. Top: Chain-Rectifying Optimization (CRO). The standard sampling pathway is prone to generating artifacts (black dashed lines). In contrast, we establish a new sampling path for erasing (red dashed lines) that better aligns the reverse sampling trajectory with a clear background. Bottom: Self-Rectifying Attention (SRA). The standard self-attention mechanism may inadvertently amplify artifacts, diverging from the expected diffusion pathway. By modifying the attention activation, we guide the model to bypass artifact regions, enhancing its focus on the background and ensuring a more accurate erase sampling path. + +# 4.1. Chain-Rectifying Optimization + +In the optimization process, existing erase inpainting methods typically employ randomly generated or object-based masks, predicting the noise in the masked regions using Equation 3. These masked regions may cover objects, and the model's optimization objective is to reconstruct these objects rather than eliminate them. As discussed above, such standard diffusion chains may inadvertently encourage the model to generate unexpected artifacts. An effective solution is to dedicate diffusion paths for the erase inpainting task between objects and backgrounds. By simulating the gradual fading of objects, this approach can suppress their emergence during the sampling process. Even if the sampling process does not start with standard Gaussian noise (a common enhancement trick), it can still help eliminate leaked objects within the masked region. Thus, the main challenge of this approach is to develop new latent states within the erase diffusion chains and to optimize the model to manipulate transitions among these latent states. + +Dynamic Latent States. To design dedicated diffusion chains for erasure between objects and backgrounds, it is essential to obtain corresponding image pairs. However, existing public datasets often lack these paired samples, highlighting the need for an innovative data synthesis strategy. + +Let $\pmb{x}_0^{ori}$ denote the original image, and a trained matting model is employed to segment the primary object within it. We then apply various image transformation techniques, including rotation and scaling, to modify the segmented objects. These transformed objects are randomly pasted onto the background region, resulting in a new synthesized image, $\pmb{x}_0^{obj}$ . This method allows us to construct a training dataset of object-background image pairs at a relatively low cost. Furthermore, to effectively simulate the gradual fading of the objects, we input dynamic images $\tilde{\pmb{x}}_t^{mix}$ for each time step $t$ during the training process, + +$$ +\tilde {\boldsymbol {x}} _ {t} ^ {m i x} = (1 - \lambda_ {t}) \boldsymbol {x} _ {0} ^ {o r i} + \lambda_ {t} \boldsymbol {x} _ {0} ^ {o b j}, \tag {6} +$$ + +where the decreasing sequence $\lambda_{:T} \in [0,1]^T$ controls the mix-up level of object-background image pairs. In the case of $t = 0$ , $\tilde{\pmb{x}}_0^{mix}$ corresponds to the original image $\pmb{x}_0^{ori}$ . These dynamic mix-up images assist the model in understanding the smooth transition from objects to backgrounds. Ultimately, based on Equation 1, we can derive a set of new latent states $\pmb{x}_t^{mix}$ for the self-rectifying diffusion chains, + +$$ +\boldsymbol {x} _ {t} ^ {m i x} = \sqrt {\bar {\alpha} _ {t}} \tilde {\boldsymbol {x}} _ {t} ^ {m i x} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {7} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ + +Optimization Objective. The traditional diffusion process starts from a fixed $\boldsymbol{x}_0^{ori}$ , and each intermediate latent state $\boldsymbol{x}_t^{ori}$ can be obtained directly using Equation 1. However, in the new self-rectifying diffusion chains, each latent state $\boldsymbol{x}_t^{mix}$ is derived from the dynamic image $\tilde{\boldsymbol{x}}_t^{mix}$ , making it infeasible to apply the standard optimization objective in Equation 3. To mitigate this challenge, we propose a novel optimization objective that enhances the alignment between the model's predicted distribution and the true distribution, as illustrated in Algorithm 1. + +Specifically, the new loss function aims to minimize the distance between the model-predicted latent states $\hat{\pmb{x}}_t^{mix}$ and the true states $\pmb{x}_t^{mix}$ . It allows the model to gradually alter its parameters to adapt to the new distribution shifts. Given one latent state $\pmb{x}_t^{mix}$ , we can derive the model-predicted latent state $\hat{\pmb{x}}_{t - \gamma}^{mix}$ at the previous $\gamma$ time step, + +$$ +\begin{array}{l} p _ {\theta} (\hat {\boldsymbol {x}} _ {t - \gamma} ^ {m i x} | \boldsymbol {x} _ {t} ^ {m i x}) = \sqrt {\bar {\alpha} _ {t - \gamma}} \left(\frac {\boldsymbol {x} _ {t} ^ {m i x} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} ^ {(t)} (\boldsymbol {x} _ {t} ^ {m i x})}{\sqrt {\bar {\alpha} _ {t}}}\right) \\ + \sqrt {1 - \bar {\alpha} _ {t - \gamma}} \epsilon_ {\theta} ^ {(t)} \left(\boldsymbol {x} _ {t} ^ {m i x}\right), \tag {8} \\ \end{array} +$$ + +where $\gamma \in (0,\gamma_{m})$ . Given the original image $\pmb{x}_0^{ori}$ and the synthesized image $\pmb{x}_0^{obj}$ , the corresponding true state $\pmb{x}_{t - \gamma}^{mix}$ + +Algorithm 1: Chain-Rectifying Optimization +repeat + $\pmb{x}_0^{ori}\sim q(x)$ $t\sim$ Uniform $(\{1,\dots ,T\})$ $\gamma \sim$ Uniform $(\{1,\ldots ,\gamma_{m}\})$ $\epsilon \sim \mathcal{N}(0,I)$ $\pmb{x}_0^{obj}\gets \pmb{x}_0^{ori}$ image transformation; + $\pmb{x}_{t}^{mix},\pmb{x}_{t - \gamma}^{mix}\leftarrow \pmb{x}_{0}^{ori},\pmb{x}_{0}^{obj}$ based on Equ 6-7; + $\hat{\pmb{x}}_{t - \gamma}^{mix}\longleftarrow \pmb{x}_{t}^{mix},\epsilon_{\theta}^{(t)}$ based on Equ 8; +Take gradient descent step on + $\nabla_{\theta}\left\| \pmb{x}_{t - \gamma}^{mix} - \hat{\pmb{x}}_{t - \gamma}^{mix}\right\| ^2$ based on Equ 9 +until converged; + +can be obtained using Equation 6 and 7. Finally, we define the new optimization objective as follows + +$$ +\min _ {\theta} \mathbb {E} _ {\gamma \sim \operatorname {U n i f o r m} (1, \gamma_ {m}), t} \left\| \boldsymbol {x} _ {t - \gamma} ^ {m i x} - p _ {\theta} \left(\hat {\boldsymbol {x}} _ {t - \gamma} ^ {m i x} \mid \boldsymbol {x} _ {t} ^ {m i x}\right) \right\| _ {2} ^ {2}. \tag {9} +$$ + +# 4.2. Self-Rectifying Attention + +We have established a self-rectifying diffusion pathway to guide the model toward object removal. However, information leakage from the mask's shape can still lead to artifacts during the early stages of denoising, resulting in latent state shifts. The self-attention mechanism tends to give the masked region stronger attention to itself rather than to the background. This phenomenon can continuously amplify artifacts along the reverse sampling path, ultimately leading to a deviation from the object removal direction. An intuitive solution is to alter the current attention layers for path calibration to mitigate the above risk of generating unexpected objects due to deviations from specific states. + +Based on our observations, we believe that the generation of the foreground region should rely more on backgrounds rather than focusing on itself. Additionally, the background region remains visible and thus should not be affected by the content of the foregrounds. By altering the self-attention activation, it is possible to ignore the negative effects of artifacts while emphasizing the background, thereby further enhancing the coherence and elimination of the generated content. Thus, we propose a simple yet effective Self-Rectifying Attention mechanism to replace the standard self-attention mechanism. Specifically, we first downsample and flatten the image mask $\mathbf{M}$ to obtain the corresponding mask vector $m\in \{0,1\}^{wh}$ . We then design an extended mask $m^{\prime}\in \{-inf,1\}^{wh\times wh}$ as follows + +$$ +m _ {i, j} ^ {\prime} = \left\{ \begin{array}{l l} 1, & m _ {i} = 0 \text {o r} m _ {j} = 0 \\ - i n f, & \text {e l s e} \end{array} \right. \tag {10} +$$ + +The extended mask is subsequently applied directly to the corresponding attention activations, effectively suppressing + +objects within the masked regions while enhancing background features. This is represented mathematically as + +$$ +\operatorname {S R A} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {\top}}{\sqrt {d}} \cdot m ^ {\prime}\right) \mathbf {V}. \tag {11} +$$ + +This mechanism enhances the model's capacity to perceive background information without any additional computational cost, while also effectively diminishing the interference of artifacts on the final generated results. Consequently, this improvement ensures that the sampling process can calibrate itself to the target direction of object removal. + +# 5. Experiments + +# 5.1. Experimental Setup + +Benchmark Datasets. We conducted a thorough evaluation of our proposed pipeline utilizing the publicly available OpenImages V5 segmentation dataset [18]. Each instance within this dataset comprises the original image, a corresponding segmentation mask, segmentation bounding boxes, and associated class labels, thereby enabling a rigorous comparative analysis against several baseline models. For our experiments, we randomly selected a subset of 10,000 samples from the OpenImages V5 test set. + +Comparison Baselines. We have chosen several cutting-edge image inpainting methods as our baselines for comparison. These include two mask-guided approaches: SD2-Inpaint [27] and LaMa [33], alongside two text-guided techniques: Inst-Inpaint [45] and PowerPaint [52]. + +Evaluation Metrics. In accordance with LaMa [33], we employ two primary evaluation metrics: Fréchet Inception Distance (FID) [11] and Learned Perceptual Image Patch Similarity (LPIPS) [47]. These metrics are adept at assessing the overall visual coherence of the inpainted images. To provide a more nuanced evaluation of the quality of content generated within the masked regions, we introduce the Local FID metric [39], which allows for a detailed assessment of local visual fidelity. Furthermore, we augment our evaluation framework by incorporating analyses derived from GPT-4o [24] along with expert human annotations, thereby examining the effectiveness of these erasure-targeted models in eliminating objects and artifacts. + +Training Details. Our proposed pipeline builds upon the foundations established by SD2-Inpaint and is implemented using PyTorch along with the Diffusers library. During the training phase, we employed the Adam [16] optimizer with a learning rate set to $3 \times 10^{-6}$ on the OpenImages V5 training set. To simplify parameters, the increasing sequence $\lambda_{:T}$ follows the same schedule as the sequence $1 - \bar{\alpha}_{:T}$ . For loss computation, we ensured that the time intervals between two timestamps did not exceed $\gamma_{m} = 100$ . All experiments were conducted on NVIDIA A100 GPUs. + +![](images/66a19e63e37ad046e2679e958ac26eac83779cd461df6111808b1824b14ebf27.jpg) +Figure 3. Qualitative results of OpenImages V5 dataset compared among SD2-Inpaint [27], SD2-Inpaint with prompt guidance [27], PowerPaint [52], Inst-Inpaint [45], LaMa [33], and our approach. + +
MethodFID↓LPIPS↓Local FID↓
SD2-Inpaint3.8050.3018.852
SD2-Inpaint*†4.0190.3087.194
PowerPaint6.0270.28910.021
Inst-Inpaint11.4230.41043.472
LaMa7.5330.2196.091
EraDiff (ours)6.5400.1923.799
+ +Table 1. Quantitative assessment of various erase inpainting models on the OpenImages V5 dataset. Optimal results are highlighted in bold, with runner-up performance underlined. + +# 5.2. Qualitative and Quantitative Comparisons + +The quantitative evaluation results of our proposed method, in comparison to several established models, are summarized in Table 1. Additionally, we present the experimental findings when the prompt is used as supplementary guid + +ance for SD2-Inpaint (abbreviated as SD2-Inpaint*, consistent throughout the paper). As indicated in the table, our model is in the mid-range of all baseline models in terms of the FID score, yet it significantly surpasses the others in the Local FID metric. This observation highlighting its remarkable ability to generate visually coherent results specifically within the designated erased regions, all the while maintaining a commendable level of visual coherence throughout the entire image. Additionally, our method records the highest performance on the LPIPS metric, suggesting that the images produced exhibit enhanced visual fidelity following the removal process. Figure 3 illustrates relevant visualization results, clearly demonstrating the model's robustness in addressing challenges such as the presence of objects within the image that closely resemble the target erasure region, as well as extensive areas that necessitate removal. + +It is essential to highlight that both the FID and LPIPS metrics primarily evaluate the visual coherence and aesthetic quality of the final generated images. However, these + +
MethodSuperiorComparableInferior
SD2-Inpaint1.03%18.07%80.90%
SD2-Inpaint*2.20%24.79%73.01%
LaMa13.17%35.29%51.54%
+ +Table 2. Quantitative results of OpenImages V5 dataset among SD2-Inpaint, SD2-Inpaint*, LaMa, and EraDiff. This table delineates a comparative analysis of the elimination performance results obtained by these methodologies relative to ours, highlighting whether their outcomes are superior, comparable, or inferior to those achieved by our approach. + +![](images/d19dbdd762df7979d9831daf6c2ff8d527184d3c0c4c7ade567ddd53a6494853.jpg) +Figure 4. Results from the user study. EraDiff demonstrates enhanced performance, as indicated by its higher mean scores in both elimination and coherence evaluations. + +metrics do not provide a direct assessment of the effectiveness regarding object or artifact elimination in the denoising process. This limitation is clearly illustrated in Figure 3, which demonstrates that despite the SD2-Inpaint model achieving the highest FID score among all evaluated models, it frequently fails to adequately eliminate unwanted objects or artifacts from the designated masked areas. + +To evaluate the effectiveness of the baseline models and our proposed method in eliminating objects and artifacts, we systematically selected the top-performing models: SD2-Inpaint, SD2-Inpaint*, and LaMa, as indicated in Table 1. A comparative analysis was performed through pairwise comparisons against our method, utilizing GPT-4o to generate objective evaluations focused on identifying superior results among the competing models. The outcomes of this analysis are described in Table 2. Furthermore, we undertook a user study that involved 20 experts that was tasked with appraising the effectiveness of the erasure and the visual aesthetics of the generated results, using a scoring system ranging from 1 to 10; where a score of 1 reflects poor performance and a score of 10 indicates exceptional quality. The outcomes of this assessment are illustrated in Fig- + +![](images/10383b492f5c1611b226cc85d3fbfde99ccefc8216bde19f98f8813dd27ca33c.jpg) +Figure 5. Visualization of EraDiff's performance across a diverse array of in-the-wild scenarios: animated imagery, e-commerce content, oil paintings, and glasses-free 3D visuals. + +ure 4. The results from both experimental methodologies distinctly demonstrate that our proposed approach significantly outperforms the competitors in effectively eliminating target objects. This conclusion is further corroborated by the finding that, despite the lower FID score recorded for SD2-Inpaint, its efficacy in object removal does not surpass that of our innovative model. + +Furthermore, we find that EraDiff shows superior performance across a diverse array of in-the-wild scenarios, as illustrated in Figure 5. A more detailed analysis of this will be provided in the appendix. + +# 5.3. Ablation Study + +To evaluate the impact of calibrating the sampling pathway on the performance of erase inpainting, we conducted four extensive experiments: each focusing on the removal of CRO, SRA, and the simultaneous exclusion of both CRO and SRA, all based on the EraDiff model. Furthermore, we assessed the model's performance without the incorporation of the mix-up strategy during training, wherein the value of $\lambda_{t}$ was held constant. For our evaluation, we employed the GPT-4o to quantify the success rate of object elimination. Results are detailed in Table 3. The results indicate that the removal of either the CRO or SRA components, as well as their simultaneous exclusion, leads to a substantial deterioration in both the visual coherence of the erased images and the efficacy of object elimination. Particularly noteworthy is the fact that the omission of CRO results in a more pronounced decline in overall performance. Additionally, the absence of the mix-up strategy produced considerable fluctuations in training loss and impeded model convergence. The instability observed can be attributed to + +
MethodLocal FID↓GPT score↑
EraDiff3.79983.43%
-w/o CRO5.71372.96%
-w/o SRA4.95078.54%
-w/o CRO ∪ SRA8.85227.80%
-w/o mix-upNaNNaN
+ +Table 3. Results of the ablation study highlight the individual and combined effects of CRO and SRA methodologies. + +![](images/0f81a14e564eefb95cabdcf7545742082a795d608bdf07351a210edccd6753c0.jpg) +Figure 6. Visual examples for the ablation study comparing baseline, baseline with CRO, baseline with SRA, and baseline with both CRO and SRA, displayed left to right. + +the limited noise present at earlier time steps, which complicates the model's ability to accurately predict the entirety of the masked region as noise. Figure 6 presents the visual outcomes of these ablation experiments, clearly illustrating the effectiveness of the calibration sampling pathway in enhancing the erasure task. In this regard, SRA serves a corrective function, effectively addressing specific anomalous denoising artifacts. + +# 5.4. How the CRO and SRA work? + +To further elucidate the roles of the CRO and SRA methodologies in the denoising process of erase inpainting, we established two types of extended experimental setups. + +To investigate the mechanism underlying CRO, we established an extremely challenging condition by setting the denoising strength to 0.6. This configuration allows for significant leakage of original image information, including the target objects intended for removal. This leakage can lead to considerable disruptions, resulting in artifacts within the erased regions and potentially even the restoration of the erasure-targeted objects. As illustrated in Figure 7, we observed the denoising processes of the EraDiff alongside the baseline model (i.e., SD2-Inpaint) under this condition. Notably, as the denoising process progressed, the EraDiff model effectively concealed the object artifacts in the erased region, whereas the baseline model highlighted it. These experimental findings provide further validation of our earlier discussion in Section 4, asserting that EraDiff accomplishes the erasure task by traversing a sample pathway that closely + +![](images/7c45982b5cf794fe1b2c6d47b1dea7664de05df4ee2ee135528461cdc8707e15.jpg) +Figure 7. Comparison of EraDiff's and baseline model's denoising process with strength set to 0.6. + +![](images/e984b43fc0ef25c5bec9de0c1fb1b7a723ffeb592280176c85bcf9257b47c65e.jpg) +Figure 8. Visualization of heatmaps representing attention block outputs in the presence and absence of the SRA mechanism. + +resembles $x_{t}^{mix}$ + +To reveal the mechanism of the SRA approach, we examined a misleading scenario where multiple similar objects exist within an image, necessitating the removal of one specific object. The heatmaps generated before and after the attention block for both the EraDiff and the baseline model are presented in Figure 8. The results indicate that the baseline model tends to focus on similar objects, while our proposed method shifts its attention to the background. This strategic focus allows the model to extract critical information from the background rather than from the foreground, thereby facilitating the elimination of the target object while maintaining coherence between the erased region and the surrounding background. + +# 6. Conclusion + +In this paper, we present the EraDiff, which enhances object elimination while maintaining visual coherence in erase inpainting. By introducing a CRO paradigm, EraDiff establishes innovative diffusion pathways that facilitate a gradual removal of objects, allowing the model to better understand the erasure intent. Moreover, the SRA mechanism effectively reduces artifacts during the sampling process. With inclusive experiments, we demonstrate that these advancements significantly improve performance in challenging scenarios, positioning EraDiff as a valuable contribution to the field of erase inpainting. + +# References + +[1] Fazil Altinel, Mete Ozay, and Takayuki Okatani. Deep structured energy-based image inpainting. In ICPR, pages 423-428. IEEE Computer Society, 2018. 3 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 3 +[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: text-driven layered image and video editing. In ECCV, pages 707-723. Springer, 2022. 2 +[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B. Goldman. Patchmatch: a randomized correspondence algorithm for structural image editing. ACM TOG, 28(3):24, 2009. 2 +[5] Chenjie Cao, Qiaole Dong, and Yanwei Fu. ZITS++: image inpainting by improving the incremental transformer on structural priors. IEEE TPAMI, 45(10):12667-12684, 2023. 2, 3 +[6] Benlei Cui, Xue-Mei Dong, Qiaoqiao Zhan, Jiangtao Peng, and Weiwei Sun. LitedealthWisenet: A lightweight network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2022. 2 +[7] Ding Ding, Sundaresh Ram, and Jeffrey J. Rodríguez. Image inpainting using nonlocal texture matching and nonlinear filtering. IEEE TIP, 28(4):1705-1719, 2019. 3 +[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In CVPR, 2022. 2, 3 +[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, Dong Chen, and Baining Guo. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, pages 12709-12720. IEEE, 2024. 3 +[10] James Hays and Alexei A. Efros. Scene completion using millions of photographs. ACM TOG, 26(3):4, 2007. 2 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5 +[12] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, and Ying Shan. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR. IEEE, 2024. 3 +[13] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM TOG, 36(4):107:1-107:14, 2017. 3 +[14] Jitesh Jain, Yuqian Zhou, Ning Yu, and Humphrey Shi. Keys to better image inpainting: Structure and texture go hand in hand. In WACV, pages 208-217, 2023. 2 +[15] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T. Freeman, David Salesin, Brian Curless, and Ce Liu. SLIDE: single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12498-12507. IEEE, 2021. 2 +[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 + +[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3 +[18] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Tom Duerig, and Vittorio Ferrari. The open images dataset V4: unified image classification, object detection, and visual relationship detection at scale. CoRR, abs/1811.00982, 2018. 2, 5 +[19] Anat Levin, Assaf Zomet, and Yair Weiss. Learning how to inpaint from global image statistics. In ICCV, pages 305-312, 2003. 2 +[20] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, pages 7757-7765. Computer Vision Foundation / IEEE, 2020. 3 +[21] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. ZONE: zero-shot instruction-guided local editing. In CVPR, pages 6254-6263. IEEE, 2024. 3 +[22] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. MAT: mask-aware transformer for large hole image inpainting. In CVPR, pages 10748-10758, 2022. 2, 3 +[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11451-11461. IEEE, 2022. 3 +[24] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-07-10. 5 +[25] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544. IEEE Computer Society, 2016. 3 +[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 3 +[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 2, 3, 5, 6 +[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 3 +[29] Andranik Sargsyan, Shant Navasardyan, Xingqian Xu, and Humphrey Shi. MI-GAN: A simple baseline for image inpainting on mobile devices. In ICCV, pages 7301-7311, 2023. 2, 3 +[30] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879. IEEE, 2024. 3 +[31] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3 +[32] Wenhao Sun, Benlei Cui, Xue-Mei Dong, and Jingqun Tang. Attentive eraser: Unleashing diffusion model's object re + +moyal potential via self-attention redirection guidance. In AAAI, 2025. 2 +[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 3172–3182, 2022. 2, 3, 5, 6 +[34] Jingqun Tang, Su Qiao, Benlei Cui, Yuhang Ma, Sheng Zhang, and Dimitrios Kanoulas. You can even annotate text with voice: Transcription-only-supervised text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4154-4163, New York, NY, USA, 2022. Association for Computing Machinery. 2 +[35] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369. IEEE, 2023. 3 +[36] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Learning 3d geometry and feature consistent gaussian splatt-ting for object removal. In ECCV, pages 1-17. Springer, 2024. 2 +[37] Daniel Winter, Matan Cohen, Shlomi Fruchtner, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. In ECCV, pages 112-129. Springer, 2024. 2 +[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 3 +[39] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 5 +[40] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin C. K. Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. CoRR, abs/2312.03771, 2023. 3 +[41] Xingqian Xu, Shant Navasardyan, Vahram Tadevosyan, Andranik Sargsyan, Yadong Mu, and Humphrey Shi. Image completion with heterogeneously filtered spectral hints. In WACV, pages 4580-4590, 2023. 2 +[42] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 2 +[43] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 3 +[44] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, pages 7505-7514, 2020. 2, 3 + +[45] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. CoRR, abs/2304.03246, 2023. 3, 5, 6 +[46] Donggeun Yoon and Donghyeon Cho. CORE-MPI: consistency object removal with embedding multiplane image. In CVPR, pages 20081-20090. IEEE, 2024. 2 +[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5 +[48] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I-Chao Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In ICLR, 2021. 2 +[49] Hao Zhou, Chuanping Hu, Chongyang Zhang, and Shengyang Shen. Visual relationship recognition via language and position guided attention. In ICASSP, 2019. 2 +[50] Hao Zhou, Chongyang Zhang, Yan Luo, Yanjun Chen, and Chuanping Hu. Embracing uncertainty: Decoupling and debias for robust temporal grounding. In CVPR, pages 8445-8454, 2021. 2 +[51] Tong Zhou, Changxing Ding, Shaowen Lin, Xinchao Wang, and Dacheng Tao. Learning oracle attention for high-fidelity face completion. In CVPR, pages 7677-7686, 2020. 3 +[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 5, 6 \ No newline at end of file diff --git a/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/images.zip b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f4add2b24f7804b4f15be694ff155866d6ee0800 --- /dev/null +++ b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aabcf68317ced40f7912332a747356a29be8658cad302c286d2c4dce7600e10f +size 877330 diff --git a/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/layout.json b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..94ecb18d9c395a70713300783544bece4fe6f17a --- /dev/null +++ b/2025/Erase Diffusion_ Empowering Object Removal Through Calibrating Diffusion Pathways/layout.json @@ -0,0 +1,7576 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 102, + 517, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 102, + 517, + 139 + ], + "spans": [ + { + "bbox": [ + 95, + 102, + 517, + 139 + ], + "type": "text", + "content": "Erase Diffusion: Empowering Object Removal Through Calibrating Diffusion Pathways" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 160, + 452, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 160, + 452, + 190 + ], + "spans": [ + { + "bbox": [ + 159, + 160, + 452, + 190 + ], + "type": "text", + "content": "Yi Liu*, Hao Zhou*, Benlei Cui†, Wenxiang Shang, Ran Lin Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 191, + 507, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 191, + 507, + 204 + ], + "spans": [ + { + "bbox": [ + 100, + 191, + 507, + 204 + ], + "type": "text", + "content": "{ly402089,baishu.zh,cuibenlei.cbl,shangwenxiang.swx,linran.lr09}@taobao.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 232, + 176, + 312 + ], + "blocks": [ + { + "bbox": [ + 58, + 232, + 176, + 312 + ], + "lines": [ + { + "bbox": [ + 58, + 232, + 176, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 232, + 176, + 312 + ], + "type": "image", + "image_path": "9012fe4d2c76d6425a890207ec4dda4fca8af07c8402537114a83df786fa68e2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 313, + 176, + 393 + ], + "blocks": [ + { + "bbox": [ + 58, + 313, + 176, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 313, + 176, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 313, + 176, + 393 + ], + "type": "image", + "image_path": "9cf332200c98842a53ab2c8c9aac846ae491cf3249e33a069ad25cfabd9344e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 180, + 232, + 303, + 393 + ], + "blocks": [ + { + "bbox": [ + 180, + 232, + 303, + 393 + ], + "lines": [ + { + "bbox": [ + 180, + 232, + 303, + 393 + ], + "spans": [ + { + "bbox": [ + 180, + 232, + 303, + 393 + ], + "type": "image", + "image_path": "a34409901c6394bf90a22338c1058e9efee906a2bb5f94268d252e96aa3936a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 429, + 393 + ], + "blocks": [ + { + "bbox": [ + 307, + 232, + 429, + 393 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 429, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 429, + 393 + ], + "type": "image", + "image_path": "c4df7140a17c3ba39d51bae3a8f2580ba3948b56ec3f7ae8416f02349d668b78.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 433, + 232, + 552, + 312 + ], + "blocks": [ + { + "bbox": [ + 433, + 232, + 552, + 312 + ], + "lines": [ + { + "bbox": [ + 433, + 232, + 552, + 312 + ], + "spans": [ + { + "bbox": [ + 433, + 232, + 552, + 312 + ], + "type": "image", + "image_path": "7c4558ad4aa3d86c08071d1759c17b99380e815b154b2c34675817f1dcabe776.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 433, + 313, + 552, + 393 + ], + "blocks": [ + { + "bbox": [ + 433, + 313, + 552, + 393 + ], + "lines": [ + { + "bbox": [ + 433, + 313, + 552, + 393 + ], + "spans": [ + { + "bbox": [ + 433, + 313, + 552, + 393 + ], + "type": "image", + "image_path": "7b085af2e69fc4034c70413fb5df5a155ebc8e10ad8d83d0a1482e9c6eb45742.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 397, + 178, + 517 + ], + "blocks": [ + { + "bbox": [ + 58, + 397, + 178, + 517 + ], + "lines": [ + { + "bbox": [ + 58, + 397, + 178, + 517 + ], + "spans": [ + { + "bbox": [ + 58, + 397, + 178, + 517 + ], + "type": "image", + "image_path": "bbad07ff923cf89d006fe593ae8ae0ec30a826e844348a4dc68009846188a954.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 532, + 555, + 566 + ], + "lines": [ + { + "bbox": [ + 55, + 532, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 555, + 566 + ], + "type": "text", + "content": "Figure 1. Diverse erase inpainting results produced by our proposed EraDiff, where images before and after removal are presented in pairs, and the areas to be erased in the original images have been marked. The EraDiff can eliminate targets in various complex real-world scenes while ensuring visual coherence in the generated images." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 181, + 397, + 302, + 517 + ], + "blocks": [ + { + "bbox": [ + 181, + 397, + 302, + 517 + ], + "lines": [ + { + "bbox": [ + 181, + 397, + 302, + 517 + ], + "spans": [ + { + "bbox": [ + 181, + 397, + 302, + 517 + ], + "type": "image", + "image_path": "bcd965ade37fa1876b3bb3addedb71b91e0b63033a5ec47d5b80eb70a72c2817.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 397, + 427, + 517 + ], + "blocks": [ + { + "bbox": [ + 306, + 397, + 427, + 517 + ], + "lines": [ + { + "bbox": [ + 306, + 397, + 427, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 427, + 517 + ], + "type": "image", + "image_path": "5f64599edb3f1420c16da5fbf8bd07cc4dcc8e4bf50ab8b783d83a8e393c14f8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 431, + 397, + 552, + 517 + ], + "blocks": [ + { + "bbox": [ + 431, + 397, + 552, + 517 + ], + "lines": [ + { + "bbox": [ + 431, + 397, + 552, + 517 + ], + "spans": [ + { + "bbox": [ + 431, + 397, + 552, + 517 + ], + "type": "image", + "image_path": "50a1fe9e7a07290dac8a2e84117ed3890217db55ba42f54f9e1ab5c9401a75eb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 151, + 575, + 200, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 575, + 200, + 588 + ], + "spans": [ + { + "bbox": [ + 151, + 575, + 200, + 588 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 601, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 601, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 601, + 295, + 685 + ], + "type": "text", + "content": "Erase inpainting, or object removal, aims to precisely remove target objects within masked regions while preserving the overall consistency of the surrounding content. Despite diffusion-based methods have made significant strides in the field of image inpainting, challenges remain regarding the emergence of unexpected objects or artifacts. We assert that the inexact diffusion pathways established by existing" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 577, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 555, + 708 + ], + "type": "text", + "content": "standard optimization paradigms constrain the efficacy of object removal. To tackle these challenges, we propose a novel Erase Diffusion, termed EraDiff, aimed at unleashing the potential power of standard diffusion in the context of object removal. In contrast to standard diffusion, the EraDiff adapts both the optimization paradigm and the network to improve the coherence and elimination of the erasure results. We first introduce a Chain-Rectifying Optimization (CRO) paradigm, a sophisticated diffusion process specifically designed to align with the objectives of erasure. This paradigm establishes innovative diffusion transition" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 196, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 196, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 196, + 703 + ], + "type": "text", + "content": "* These authors contributed equally to this work." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 56, + 703, + 127, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 127, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 127, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2418" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "pathways that simulate the gradual elimination of objects during optimization, allowing the model to accurately capture the intent of object removal. Furthermore, to mitigate deviations caused by artifacts during the sampling pathways, we develop a simple yet effective Self-Rectifying Attention (SRA) mechanism. The SRA calibrates the sampling pathways by altering self-attention activation, allowing the model to effectively bypass artifacts while further enhancing the coherence of the generated content. With this design, our proposed EraDiff achieves state-of-the-art performance on the OpenImages V5 dataset and demonstrates significant superiority in real-world scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 136, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 136, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 136, + 251 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 426 + ], + "type": "text", + "content": "Image inpainting is a crucial technique in computer vision, aimed at reconstructing missing or damaged regions while maintaining visual coherence and contextual integrity [15, 32]. As one specialized form of image inpainting, object removal, also called erase inpainting, has broader applications in various fields such as social media [3, 50], advertising design [36], and image processing [37, 46]. However, in comparison to general image inpainting, an ideal erase inpainting model must address two critical challenges. Coherence: The model should seamlessly inpaint the masked area, ensuring consistency in lighting, content, and other relevant aspects. Elimination: The model should accurately remove objects within the masked area while preventing the generation of extraneous elements or artifacts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 569 + ], + "type": "text", + "content": "Prior erase inpainting methods primarily relied on nonparametric patch sampling techniques [4, 10, 19] or Generative Adversarial Networks (GANs) [5, 8, 14, 22, 29, 33, 41, 44, 48]. These approaches often resort to filling large masked regions with repetitive patches, resulting in a lack of coherence in the generated images. In recent years, methods based on Latent Diffusion Models (LDMs) [27] have demonstrated superiority in generating more natural images. However, these methods struggle to eliminate target objects. For example, when a user seeks to erase a piece of pizza in Figure 1, the LDMs may mistakenly generate another piece of pizza instead of the expected clean plate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "We attribute the above issues to the fact that standard diffusion pathways in most existing methods are not suitable for erase tasks. During optimization, most diffusion-based erase inpainting models take the original image with added noise and randomly generated masks as input, aiming to recover the original image in the presence of noise. However, this standard training paradigm only establishes a denoising process that transitions from random noise to clear images, without addressing the specific goal of object removal. Consequently, the model may follow a denoising pathway from noise to an image with objects. Based on this observation, we argue that an ideal erase inpainting model should" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "establish a diffusion pathway directly from objects to backgrounds to ensure the definitive removal of unwanted elements. Furthermore, in the early stages of denoising, the reconstructed regions are significantly influenced by the shape of the masks and the level of noise. This impact may lead to deviations in the early latent states, resulting in the emergence of artifacts. The standard self-attention mechanism tends to incorrectly regard the features of these artifacts as important information, which gradually amplifies the deviations during the subsequent denoising process. Ultimately, it leads to unexpected objects in the generated images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 553, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 562 + ], + "type": "text", + "content": "In this paper, we present a novel erase diffusion model, termed EraDiff, specifically designed for object removal, which effectively unleashes the potential erasure power of standard diffusion. First, we introduce a Chain-Rectifying Optimization (CRO) paradigm that supports the establishment of new diffusion pathways from noise to backgrounds. Specifically, we develop a dynamic image synthesis strategy that enables the generation of a variety of dynamic images at different time steps without the need for extra data. These synthesized dynamic images effectively simulate the gradual elimination of objects during the denoising process, yielding intermediate latent states corresponding to different time steps. Next, we introduce a new dedicated optimization objective for erase inpainting. This objective guides the establishment of transitions among multiple intermediate latent states. By applying the CRO paradigm, the model learns to accurately identify the intention to erase and improve content coherence. Furthermore, to address potential deviations caused by artifacts during the early denoising process, we design a Self-Rectifying Attention (SRA) mechanism that explicitly guides the model in executing object removal more effectively. By altering self-attention activation, the SRA enhances background features while rectifying its incorrect reliance on artifacts. This replacement of the standard attention mechanism results in final generated images that appear more realistic and devoid of unexpected objects. To the best of our knowledge, our proposed method achieves state-of-the-art performance in both the public OpenImages V5 [18] dataset and large-scale real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 573, + 400, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 573, + 400, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 573, + 400, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "Image Inpainting. Image inpainting, which involves the reconstruction of large-scale missing regions, has garnered substantial attention within the field. Traditional image inpainting methods employ heuristic patch-based propagation algorithms to borrow texture and structure from neighboring regions of the corrupted areas [4, 10, 19]. However, these methods often struggle to effectively restore areas with complex structures or semantics [42]. To tackle this issue, numerous studies have harnessed the power of deep neural networks [6, 34, 49] and proposed various en" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2419" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "hancements to GANs [25] aimed at improving the coherence of inpainted regions. Subsequent research has extensively examined different facets, including semantic context and texture [1, 7, 13], edges and contours [5, 8], as well as the design of hand-engineered architectures [20, 51]. Notably, Suvorov et al. [33] introduced FFCs into residual modules, providing a remedy for enhancing the perceptual quality of GAN-generated outputs. More recently, promising efforts have been made in applying LDMs to inpainting tasks, showcasing their potential to achieve impressive results [9, 23, 26, 31, 35, 38]. However, general image inpainting models may generate unexpected objects in the masked areas, making these LDM-based methods unsuitable for effective object removal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 296, + 515 + ], + "type": "text", + "content": "Erase Inpainting. Erase inpainting, one specialized form of image inpainting, focuses on removing unwanted content from images. The typical setup for erase inpainting is to provide an image along with an object mask as input, ultimately generating a clean background that excludes the specified object. Many studies have concentrated on improvements in model architecture [22, 40, 43, 44] and loss functions [29, 33]. While these methods effectively improve the coherence of masked regions, they still struggle with object removal. Recent works [2, 12, 21, 30, 45] have attempted to leverage textual prompts to locate erase targets within images. However, relying on textual cues to accurately identify areas for erasure has proven to be unstable. Models fail to follow textual instructions, which can lead to unfavorable results. As a result, these approaches face challenges in large-scale practical applications. Through in-depth investigation, we find that the struggle to effectively erase objects stems from the inexact construction of the erase diffusion chain. Our proposed method attempts to directly establish diffusion pathways from the noise to the background. This approach enables effective object removal without relying on additional textual information and produces one clean and natural background." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 524, + 141, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 141, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 141, + 537 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": "Latent Diffusion Models. In this paper, we utilize latent diffusion models (LDMs) as our erase inpainting model. LDMs are probabilistic models designed to learn a data distribution " + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "inline_equation", + "content": "p(\\pmb{x})" + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": " by gradually denoising a normally distributed variable, which corresponds to learning the reverse sampling process of a fixed Markov Chain of length " + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": ". Specifically, the forward diffusion process progressively adds noise to the initial latent state " + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 646, + 295, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 646, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 94, + 646, + 295, + 660 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\boldsymbol {I}\\right), \\tag {1}", + "image_path": "e9af3c3f0f7de431216345b416b6eaa369201a44340d7c180d72729828216098.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "inline_equation", + "content": "1 - \\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "text", + "content": " is the variance schedule at step " + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "text", + "content": " and represents the level of noise, with " + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s" + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 690, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 295, + 714 + ], + "type": "text", + "content": "According to the DDIM algorithm, the reverse sampling process can be defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 341, + 86, + 553, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 86, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 341, + 86, + 553, + 140 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} _ {p r e v} = \\sqrt {\\bar {\\alpha} _ {p r e v}} \\left(\\frac {\\boldsymbol {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) \\tag {2} \\\\ + \\sqrt {1 - \\bar {\\alpha} _ {p r e v} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t}) + \\sigma_ {t} \\epsilon_ {t}, \\\\ \\end{array}", + "image_path": "f2e84439b1caa631db94210dc1c4470869949081f1ed9f6bc2820f186278f1df.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "inline_equation", + "content": "\\epsilon_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "text", + "content": " is standard Gaussian noise. Here, " + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "text", + "content": " are hyper-parameters, and the reverse sampling process can be considered deterministic as " + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "inline_equation", + "content": "\\sigma_t \\to 0" + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta^{(t)}(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "text", + "content": " is trained to predict the noise added to " + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 313, + 144, + 555, + 206 + ], + "type": "text", + "content": " by minimizing the following optimization objective" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 210, + 553, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 210, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 367, + 210, + 553, + 233 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (\\mathbf {0}, I)} \\left\\| \\epsilon - \\epsilon_ {\\theta} ^ {(t)} \\left(\\boldsymbol {x} _ {t}\\right) \\right\\| _ {2} ^ {2}. \\tag {3}", + "image_path": "7a2a1dbed4f239150487af7b0867c15e4b18debdefdc8134ea013c645caa4700.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "text", + "content": "Structure of Erasing Ipainting. Most advanced erasing inpainting methods are based on diffusion models. In this paper, we focus on the SD2-Inpaint model[27], which mainly consists of a Variational AutoEncoder (VAE) [17] and a U-Net network [28]. The VAE facilitates the transformation of input images into latent space. For Erasing Ipainting, the inputs typically consist of a noisy image " + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t^{ori} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "text", + "content": ", a random binary mask " + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "text", + "content": ", and the corresponding masked image " + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}' = \\boldsymbol{x}_0^{ori} \\odot \\mathbf{M}" + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "text", + "content": ". The U-Net network predicts the associated noise at timestamp " + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 237, + 555, + 392 + ], + "type": "text", + "content": " based on the Equation 3. This methodology enables the model to establish a diffusion chain between the original image and Gaussian noise, allowing for seamlessly inpainting masked regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "content": "Additionally, the self-attention mechanism within U-Net networks plays a significant role in the erasing process. This mechanism aggregates information from the entire image, thereby controlling the generation of features in the masked regions during the reverse sampling process. Given one latent feature map " + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "content": " are the height, width, and channel dimensions of " + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "content": " respectively, the self-attention process can be represented as follows" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 493, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 493, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 365, + 493, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} = \\ell_ {Q} (\\boldsymbol {z}), \\ell_ {K} (\\boldsymbol {z}), \\ell_ {V} (\\boldsymbol {z}), \\tag {4}", + "image_path": "87688c30a09352bf2828c68cb4a1abed760a6cb3ce319d4c4ede5b011a791805.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 512, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 512, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 351, + 512, + 553, + 539 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V}, \\tag {5}", + "image_path": "af45e19849959fdd0016086a3fbad9569384c981d836a09fa816d764d87e3a3d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 541, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 541, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\ell_{Q},\\ell_{K},\\ell_{V}" + }, + { + "bbox": [ + 313, + 541, + 554, + 564 + ], + "type": "text", + "content": " are learnable linear layers, and " + }, + { + "bbox": [ + 313, + 541, + 554, + 564 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 541, + 554, + 564 + ], + "type": "text", + "content": " denotes the scaling factor." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 574, + 370, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 574, + 370, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 574, + 370, + 586 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Overview. The overall architecture of our proposed Erase Diffusion, termed EraDiff, is shown in Figure 2, consisting of a Chain-Rectifying Optimization paradigm (see Section 4.1) and a Self-Rectifying Attention mechanism (see Section 4.2). This CRO paradigm introduces the design of dynamic latent states along with a dedicated optimization objective, thereby establishing novel diffusion chains that better align with the erasure objectives. Meanwhile, the SRA mechanism effectively guides the sampling process to prevent state deviations caused by artifacts." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2420" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 281 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 281 + ], + "type": "image", + "image_path": "daf41fa0b78b4a11615ca8cf467fbe269d34071c6fae5815d754f1db6f9ed420.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 290, + 296, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 456 + ], + "type": "text", + "content": "Figure 2. The overview of our proposed Erase Diffusion, termed EraDiff. Left: Dynamic image synthesis. Each image is initially transformed using techniques like matting, scaling, and copy-pasting. A mix-up strategy then synthesizes a series of dynamic images " + }, + { + "bbox": [ + 55, + 290, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\{\\tilde{x}_t^{mix}\\}" + }, + { + "bbox": [ + 55, + 290, + 296, + 456 + ], + "type": "text", + "content": " that simulate the gradual fading of the object. Top: Chain-Rectifying Optimization (CRO). The standard sampling pathway is prone to generating artifacts (black dashed lines). In contrast, we establish a new sampling path for erasing (red dashed lines) that better aligns the reverse sampling trajectory with a clear background. Bottom: Self-Rectifying Attention (SRA). The standard self-attention mechanism may inadvertently amplify artifacts, diverging from the expected diffusion pathway. By modifying the attention activation, we guide the model to bypass artifact regions, enhancing its focus on the background and ensuring a more accurate erase sampling path." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 479, + 224, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 224, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 224, + 493 + ], + "type": "text", + "content": "4.1. Chain-Rectifying Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "In the optimization process, existing erase inpainting methods typically employ randomly generated or object-based masks, predicting the noise in the masked regions using Equation 3. These masked regions may cover objects, and the model's optimization objective is to reconstruct these objects rather than eliminate them. As discussed above, such standard diffusion chains may inadvertently encourage the model to generate unexpected artifacts. An effective solution is to dedicate diffusion paths for the erase inpainting task between objects and backgrounds. By simulating the gradual fading of objects, this approach can suppress their emergence during the sampling process. Even if the sampling process does not start with standard Gaussian noise (a common enhancement trick), it can still help eliminate leaked objects within the masked region. Thus, the main challenge of this approach is to develop new latent states within the erase diffusion chains and to optimize the model to manipulate transitions among these latent states." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "Dynamic Latent States. To design dedicated diffusion chains for erasure between objects and backgrounds, it is essential to obtain corresponding image pairs. However, existing public datasets often lack these paired samples, highlighting the need for an innovative data synthesis strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{ori}" + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "text", + "content": " denote the original image, and a trained matting model is employed to segment the primary object within it. We then apply various image transformation techniques, including rotation and scaling, to modify the segmented objects. These transformed objects are randomly pasted onto the background region, resulting in a new synthesized image, " + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{obj}" + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "text", + "content": ". This method allows us to construct a training dataset of object-background image pairs at a relatively low cost. Furthermore, to effectively simulate the gradual fading of the objects, we input dynamic images " + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_t^{mix}" + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "text", + "content": " for each time step " + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 133, + 555, + 265 + ], + "type": "text", + "content": " during the training process," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 367, + 273, + 555, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 273, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 367, + 273, + 555, + 287 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {x}} _ {t} ^ {m i x} = (1 - \\lambda_ {t}) \\boldsymbol {x} _ {0} ^ {o r i} + \\lambda_ {t} \\boldsymbol {x} _ {0} ^ {o b j}, \\tag {6}", + "image_path": "7ed72c3b6e74558ee9192dbb721c5080a7e1687486b7fa2ef0366a4c15c547c2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "text", + "content": "where the decreasing sequence " + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{:T} \\in [0,1]^T" + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "text", + "content": " controls the mix-up level of object-background image pairs. In the case of " + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_0^{mix}" + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "text", + "content": " corresponds to the original image " + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{ori}" + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "text", + "content": ". These dynamic mix-up images assist the model in understanding the smooth transition from objects to backgrounds. Ultimately, based on Equation 1, we can derive a set of new latent states " + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t^{mix}" + }, + { + "bbox": [ + 313, + 297, + 554, + 381 + ], + "type": "text", + "content": " for the self-rectifying diffusion chains," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 390, + 555, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 390, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 367, + 390, + 555, + 404 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} ^ {m i x} = \\sqrt {\\bar {\\alpha} _ {t}} \\tilde {\\boldsymbol {x}} _ {t} ^ {m i x} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {7}", + "image_path": "d7dfd4ee2ff4789c13e2aaa4e44b922b54a4941bb889a19ce7b0cfd9cd992f74.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 413, + 397, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 397, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 397, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 413, + 397, + 425 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "text", + "content": "Optimization Objective. The traditional diffusion process starts from a fixed " + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^{ori}" + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "text", + "content": ", and each intermediate latent state " + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t^{ori}" + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "text", + "content": " can be obtained directly using Equation 1. However, in the new self-rectifying diffusion chains, each latent state " + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t^{mix}" + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "text", + "content": " is derived from the dynamic image " + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{x}}_t^{mix}" + }, + { + "bbox": [ + 313, + 426, + 554, + 544 + ], + "type": "text", + "content": ", making it infeasible to apply the standard optimization objective in Equation 3. To mitigate this challenge, we propose a novel optimization objective that enhances the alignment between the model's predicted distribution and the true distribution, as illustrated in Algorithm 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "text", + "content": "Specifically, the new loss function aims to minimize the distance between the model-predicted latent states " + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_t^{mix}" + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "text", + "content": " and the true states " + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t^{mix}" + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "text", + "content": ". It allows the model to gradually alter its parameters to adapt to the new distribution shifts. Given one latent state " + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t^{mix}" + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "text", + "content": ", we can derive the model-predicted latent state " + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{t - \\gamma}^{mix}" + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "text", + "content": " at the previous " + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 545, + 554, + 619 + ], + "type": "text", + "content": " time step," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 628, + 553, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 688 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\theta} (\\hat {\\boldsymbol {x}} _ {t - \\gamma} ^ {m i x} | \\boldsymbol {x} _ {t} ^ {m i x}) = \\sqrt {\\bar {\\alpha} _ {t - \\gamma}} \\left(\\frac {\\boldsymbol {x} _ {t} ^ {m i x} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\boldsymbol {x} _ {t} ^ {m i x})}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) \\\\ + \\sqrt {1 - \\bar {\\alpha} _ {t - \\gamma}} \\epsilon_ {\\theta} ^ {(t)} \\left(\\boldsymbol {x} _ {t} ^ {m i x}\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "9fefd64076304e20958b8b9ad3e60ee76bf75bb632b3a099df7356a79cd671bd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\gamma \\in (0,\\gamma_{m})" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": ". Given the original image " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{ori}" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": " and the synthesized image " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{obj}" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": ", the corresponding true state " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - \\gamma}^{mix}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2421" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 69, + 92, + 294, + 235 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 247, + 87 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 247, + 87 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 247, + 87 + ], + "type": "text", + "content": "Algorithm 1: Chain-Rectifying Optimization" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "lines": [ + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "text", + "content": "repeat \n" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{ori}\\sim q(x)" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "t\\sim" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "text", + "content": " Uniform " + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "(\\{1,\\dots ,T\\})" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "\\gamma \\sim" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "text", + "content": " Uniform " + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "(\\{1,\\ldots ,\\gamma_{m}\\})" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{obj}\\gets \\pmb{x}_0^{ori}" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "text", + "content": " image transformation; \n" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t}^{mix},\\pmb{x}_{t - \\gamma}^{mix}\\leftarrow \\pmb{x}_{0}^{ori},\\pmb{x}_{0}^{obj}" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "text", + "content": " based on Equ 6-7; \n" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{t - \\gamma}^{mix}\\longleftarrow \\pmb{x}_{t}^{mix},\\epsilon_{\\theta}^{(t)}" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "text", + "content": " based on Equ 8; \nTake gradient descent step on \n" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\left\\| \\pmb{x}_{t - \\gamma}^{mix} - \\hat{\\pmb{x}}_{t - \\gamma}^{mix}\\right\\| ^2" + }, + { + "bbox": [ + 69, + 92, + 294, + 235 + ], + "type": "text", + "content": " based on Equ 9 \nuntil converged;" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 259, + 295, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 295, + 284 + ], + "type": "text", + "content": "can be obtained using Equation 6 and 7. Finally, we define the new optimization objective as follows" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 290, + 296, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 290, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 63, + 290, + 296, + 310 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\gamma \\sim \\operatorname {U n i f o r m} (1, \\gamma_ {m}), t} \\left\\| \\boldsymbol {x} _ {t - \\gamma} ^ {m i x} - p _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {t - \\gamma} ^ {m i x} \\mid \\boldsymbol {x} _ {t} ^ {m i x}\\right) \\right\\| _ {2} ^ {2}. \\tag {9}", + "image_path": "e356f5041447bd7d2cdbcbc3380fa9dbbda7b0d08b866509db5ef868a9cdc58e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 195, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 195, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 195, + 329 + ], + "type": "text", + "content": "4.2. Self-Rectifying Attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": "We have established a self-rectifying diffusion pathway to guide the model toward object removal. However, information leakage from the mask's shape can still lead to artifacts during the early stages of denoising, resulting in latent state shifts. The self-attention mechanism tends to give the masked region stronger attention to itself rather than to the background. This phenomenon can continuously amplify artifacts along the reverse sampling path, ultimately leading to a deviation from the object removal direction. An intuitive solution is to alter the current attention layers for path calibration to mitigate the above risk of generating unexpected objects due to deviations from specific states." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "text", + "content": "Based on our observations, we believe that the generation of the foreground region should rely more on backgrounds rather than focusing on itself. Additionally, the background region remains visible and thus should not be affected by the content of the foregrounds. By altering the self-attention activation, it is possible to ignore the negative effects of artifacts while emphasizing the background, thereby further enhancing the coherence and elimination of the generated content. Thus, we propose a simple yet effective Self-Rectifying Attention mechanism to replace the standard self-attention mechanism. Specifically, we first downsample and flatten the image mask " + }, + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "text", + "content": " to obtain the corresponding mask vector " + }, + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "inline_equation", + "content": "m\\in \\{0,1\\}^{wh}" + }, + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "text", + "content": ". We then design an extended mask " + }, + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "inline_equation", + "content": "m^{\\prime}\\in \\{-inf,1\\}^{wh\\times wh}" + }, + { + "bbox": [ + 55, + 478, + 296, + 646 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 651, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 651, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 97, + 651, + 295, + 684 + ], + "type": "interline_equation", + "content": "m _ {i, j} ^ {\\prime} = \\left\\{ \\begin{array}{l l} 1, & m _ {i} = 0 \\text {o r} m _ {j} = 0 \\\\ - i n f, & \\text {e l s e} \\end{array} \\right. \\tag {10}", + "image_path": "5ef0c527eb83459d88157a0d863507144eb6c0b814a3b1780590dec317df0868.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The extended mask is subsequently applied directly to the corresponding attention activations, effectively suppressing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "objects within the masked regions while enhancing background features. This is represented mathematically as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 329, + 106, + 555, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 106, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 329, + 106, + 555, + 135 + ], + "type": "interline_equation", + "content": "\\operatorname {S R A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}} \\cdot m ^ {\\prime}\\right) \\mathbf {V}. \\tag {11}", + "image_path": "aad490f5560c64e16b1105ac3be6a5b6fdfb3299d888d045700127510b344374.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 144, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 216 + ], + "type": "text", + "content": "This mechanism enhances the model's capacity to perceive background information without any additional computational cost, while also effectively diminishing the interference of artifacts on the final generated results. Consequently, this improvement ensures that the sampling process can calibrate itself to the target direction of object removal." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 229, + 395, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 229, + 395, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 229, + 395, + 243 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 250, + 430, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 250, + 430, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 250, + 430, + 262 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 268, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 376 + ], + "type": "text", + "content": "Benchmark Datasets. We conducted a thorough evaluation of our proposed pipeline utilizing the publicly available OpenImages V5 segmentation dataset [18]. Each instance within this dataset comprises the original image, a corresponding segmentation mask, segmentation bounding boxes, and associated class labels, thereby enabling a rigorous comparative analysis against several baseline models. For our experiments, we randomly selected a subset of 10,000 samples from the OpenImages V5 test set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "content": "Comparison Baselines. We have chosen several cutting-edge image inpainting methods as our baselines for comparison. These include two mask-guided approaches: SD2-Inpaint [27] and LaMa [33], alongside two text-guided techniques: Inst-Inpaint [45] and PowerPaint [52]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "type": "text", + "content": "Evaluation Metrics. In accordance with LaMa [33], we employ two primary evaluation metrics: Fréchet Inception Distance (FID) [11] and Learned Perceptual Image Patch Similarity (LPIPS) [47]. These metrics are adept at assessing the overall visual coherence of the inpainted images. To provide a more nuanced evaluation of the quality of content generated within the masked regions, we introduce the Local FID metric [39], which allows for a detailed assessment of local visual fidelity. Furthermore, we augment our evaluation framework by incorporating analyses derived from GPT-4o [24] along with expert human annotations, thereby examining the effectiveness of these erasure-targeted models in eliminating objects and artifacts." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Training Details. Our proposed pipeline builds upon the foundations established by SD2-Inpaint and is implemented using PyTorch along with the Diffusers library. During the training phase, we employed the Adam [16] optimizer with a learning rate set to " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " on the OpenImages V5 training set. To simplify parameters, the increasing sequence " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{:T}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " follows the same schedule as the sequence " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "1 - \\bar{\\alpha}_{:T}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". For loss computation, we ensured that the time intervals between two timestamps did not exceed " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{m} = 100" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". All experiments were conducted on NVIDIA A100 GPUs." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2422" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 553, + 439 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 439 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 439 + ], + "type": "image", + "image_path": "66a19e63e37ad046e2679e958ac26eac83779cd461df6111808b1824b14ebf27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 448, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 448, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 555, + 472 + ], + "type": "text", + "content": "Figure 3. Qualitative results of OpenImages V5 dataset compared among SD2-Inpaint [27], SD2-Inpaint with prompt guidance [27], PowerPaint [52], Inst-Inpaint [45], LaMa [33], and our approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 483, + 294, + 581 + ], + "blocks": [ + { + "bbox": [ + 56, + 483, + 294, + 581 + ], + "lines": [ + { + "bbox": [ + 56, + 483, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 294, + 581 + ], + "type": "table", + "html": "
MethodFID↓LPIPS↓Local FID↓
SD2-Inpaint3.8050.3018.852
SD2-Inpaint*†4.0190.3087.194
PowerPaint6.0270.28910.021
Inst-Inpaint11.4230.41043.472
LaMa7.5330.2196.091
EraDiff (ours)6.5400.1923.799
", + "image_path": "89892a31795a499aff256d3f7c1fe903ef07903a2f9a898906acdaffdf7a1f0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 589, + 295, + 623 + ], + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 623 + ], + "type": "text", + "content": "Table 1. Quantitative assessment of various erase inpainting models on the OpenImages V5 dataset. Optimal results are highlighted in bold, with runner-up performance underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 644, + 278, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 278, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 278, + 658 + ], + "type": "text", + "content": "5.2. Qualitative and Quantitative Comparisons" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "The quantitative evaluation results of our proposed method, in comparison to several established models, are summarized in Table 1. Additionally, we present the experimental findings when the prompt is used as supplementary guid" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 485, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 676 + ], + "type": "text", + "content": "ance for SD2-Inpaint (abbreviated as SD2-Inpaint*, consistent throughout the paper). As indicated in the table, our model is in the mid-range of all baseline models in terms of the FID score, yet it significantly surpasses the others in the Local FID metric. This observation highlighting its remarkable ability to generate visually coherent results specifically within the designated erased regions, all the while maintaining a commendable level of visual coherence throughout the entire image. Additionally, our method records the highest performance on the LPIPS metric, suggesting that the images produced exhibit enhanced visual fidelity following the removal process. Figure 3 illustrates relevant visualization results, clearly demonstrating the model's robustness in addressing challenges such as the presence of objects within the image that closely resemble the target erasure region, as well as extensive areas that necessitate removal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "It is essential to highlight that both the FID and LPIPS metrics primarily evaluate the visual coherence and aesthetic quality of the final generated images. However, these" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2423" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 68, + 294, + 129 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 294, + 129 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 294, + 129 + ], + "type": "table", + "html": "
MethodSuperiorComparableInferior
SD2-Inpaint1.03%18.07%80.90%
SD2-Inpaint*2.20%24.79%73.01%
LaMa13.17%35.29%51.54%
", + "image_path": "575226ac7d2200ae76735e5e81f2bed82c464e5c8576796747ce2bfd9a278b16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 137, + 295, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 204 + ], + "type": "text", + "content": "Table 2. Quantitative results of OpenImages V5 dataset among SD2-Inpaint, SD2-Inpaint*, LaMa, and EraDiff. This table delineates a comparative analysis of the elimination performance results obtained by these methodologies relative to ours, highlighting whether their outcomes are superior, comparable, or inferior to those achieved by our approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 217, + 291, + 392 + ], + "blocks": [ + { + "bbox": [ + 56, + 217, + 291, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 217, + 291, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 291, + 392 + ], + "type": "image", + "image_path": "d19dbdd762df7979d9831daf6c2ff8d527184d3c0c4c7ade567ddd53a6494853.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 295, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 434 + ], + "type": "text", + "content": "Figure 4. Results from the user study. EraDiff demonstrates enhanced performance, as indicated by its higher mean scores in both elimination and coherence evaluations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": "metrics do not provide a direct assessment of the effectiveness regarding object or artifact elimination in the denoising process. This limitation is clearly illustrated in Figure 3, which demonstrates that despite the SD2-Inpaint model achieving the highest FID score among all evaluated models, it frequently fails to adequately eliminate unwanted objects or artifacts from the designated masked areas." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": "To evaluate the effectiveness of the baseline models and our proposed method in eliminating objects and artifacts, we systematically selected the top-performing models: SD2-Inpaint, SD2-Inpaint*, and LaMa, as indicated in Table 1. A comparative analysis was performed through pairwise comparisons against our method, utilizing GPT-4o to generate objective evaluations focused on identifying superior results among the competing models. The outcomes of this analysis are described in Table 2. Furthermore, we undertook a user study that involved 20 experts that was tasked with appraising the effectiveness of the erasure and the visual aesthetics of the generated results, using a scoring system ranging from 1 to 10; where a score of 1 reflects poor performance and a score of 10 indicates exceptional quality. The outcomes of this assessment are illustrated in Fig-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 269 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 269 + ], + "type": "image", + "image_path": "10383b492f5c1611b226cc85d3fbfde99ccefc8216bde19f98f8813dd27ca33c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "text", + "content": "Figure 5. Visualization of EraDiff's performance across a diverse array of in-the-wild scenarios: animated imagery, e-commerce content, oil paintings, and glasses-free 3D visuals." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 324, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 407 + ], + "type": "text", + "content": "ure 4. The results from both experimental methodologies distinctly demonstrate that our proposed approach significantly outperforms the competitors in effectively eliminating target objects. This conclusion is further corroborated by the finding that, despite the lower FID score recorded for SD2-Inpaint, its efficacy in object removal does not surpass that of our innovative model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "text", + "content": "Furthermore, we find that EraDiff shows superior performance across a diverse array of in-the-wild scenarios, as illustrated in Figure 5. A more detailed analysis of this will be provided in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 408, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 408, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 408, + 480 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "To evaluate the impact of calibrating the sampling pathway on the performance of erase inpainting, we conducted four extensive experiments: each focusing on the removal of CRO, SRA, and the simultaneous exclusion of both CRO and SRA, all based on the EraDiff model. Furthermore, we assessed the model's performance without the incorporation of the mix-up strategy during training, wherein the value of " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " was held constant. For our evaluation, we employed the GPT-4o to quantify the success rate of object elimination. Results are detailed in Table 3. The results indicate that the removal of either the CRO or SRA components, as well as their simultaneous exclusion, leads to a substantial deterioration in both the visual coherence of the erased images and the efficacy of object elimination. Particularly noteworthy is the fact that the omission of CRO results in a more pronounced decline in overall performance. Additionally, the absence of the mix-up strategy produced considerable fluctuations in training loss and impeded model convergence. The instability observed can be attributed to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2424" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 286, + 156 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 286, + 156 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 286, + 156 + ], + "type": "table", + "html": "
MethodLocal FID↓GPT score↑
EraDiff3.79983.43%
-w/o CRO5.71372.96%
-w/o SRA4.95078.54%
-w/o CRO ∪ SRA8.85227.80%
-w/o mix-upNaNNaN
", + "image_path": "7edab3f66294605d48a6f8b08850865f18ab8c5410d8fa29b403b4d371581944.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 296, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 187 + ], + "type": "text", + "content": "Table 3. Results of the ablation study highlight the individual and combined effects of CRO and SRA methodologies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 196, + 296, + 305 + ], + "blocks": [ + { + "bbox": [ + 58, + 196, + 296, + 305 + ], + "lines": [ + { + "bbox": [ + 58, + 196, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 58, + 196, + 296, + 305 + ], + "type": "image", + "image_path": "0f81a14e564eefb95cabdcf7545742082a795d608bdf07351a210edccd6753c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 312, + 296, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 346 + ], + "type": "text", + "content": "Figure 6. Visual examples for the ablation study comparing baseline, baseline with CRO, baseline with SRA, and baseline with both CRO and SRA, displayed left to right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 365, + 295, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 461 + ], + "type": "text", + "content": "the limited noise present at earlier time steps, which complicates the model's ability to accurately predict the entirety of the masked region as noise. Figure 6 presents the visual outcomes of these ablation experiments, clearly illustrating the effectiveness of the calibration sampling pathway in enhancing the erasure task. In this regard, SRA serves a corrective function, effectively addressing specific anomalous denoising artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 222, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 222, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 222, + 480 + ], + "type": "text", + "content": "5.4. How the CRO and SRA work?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": "To further elucidate the roles of the CRO and SRA methodologies in the denoising process of erase inpainting, we established two types of extended experimental setups." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "content": "To investigate the mechanism underlying CRO, we established an extremely challenging condition by setting the denoising strength to 0.6. This configuration allows for significant leakage of original image information, including the target objects intended for removal. This leakage can lead to considerable disruptions, resulting in artifacts within the erased regions and potentially even the restoration of the erasure-targeted objects. As illustrated in Figure 7, we observed the denoising processes of the EraDiff alongside the baseline model (i.e., SD2-Inpaint) under this condition. Notably, as the denoising process progressed, the EraDiff model effectively concealed the object artifacts in the erased region, whereas the baseline model highlighted it. These experimental findings provide further validation of our earlier discussion in Section 4, asserting that EraDiff accomplishes the erasure task by traversing a sample pathway that closely" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 554, + 161 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 554, + 161 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 554, + 161 + ], + "type": "image", + "image_path": "7c45982b5cf794fe1b2c6d47b1dea7664de05df4ee2ee135528461cdc8707e15.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 168, + 555, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 192 + ], + "type": "text", + "content": "Figure 7. Comparison of EraDiff's and baseline model's denoising process with strength set to 0.6." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 201, + 553, + 330 + ], + "blocks": [ + { + "bbox": [ + 316, + 201, + 553, + 330 + ], + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 330 + ], + "type": "image", + "image_path": "e984b43fc0ef25c5bec9de0c1fb1b7a723ffeb592280176c85bcf9257b47c65e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "type": "text", + "content": "Figure 8. Visualization of heatmaps representing attention block outputs in the presence and absence of the SRA mechanism." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 382, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 382, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 382, + 392 + ], + "type": "text", + "content": "resembles " + }, + { + "bbox": [ + 313, + 380, + 382, + 392 + ], + "type": "inline_equation", + "content": "x_{t}^{mix}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 393, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 548 + ], + "type": "text", + "content": "To reveal the mechanism of the SRA approach, we examined a misleading scenario where multiple similar objects exist within an image, necessitating the removal of one specific object. The heatmaps generated before and after the attention block for both the EraDiff and the baseline model are presented in Figure 8. The results indicate that the baseline model tends to focus on similar objects, while our proposed method shifts its attention to the background. This strategic focus allows the model to extract critical information from the background rather than from the foreground, thereby facilitating the elimination of the target object while maintaining coherence between the erased region and the surrounding background." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we present the EraDiff, which enhances object elimination while maintaining visual coherence in erase inpainting. By introducing a CRO paradigm, EraDiff establishes innovative diffusion pathways that facilitate a gradual removal of objects, allowing the model to better understand the erasure intent. Moreover, the SRA mechanism effectively reduces artifacts during the sampling process. With inclusive experiments, we demonstrate that these advancements significantly improve performance in challenging scenarios, positioning EraDiff as a valuable contribution to the field of erase inpainting." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2425" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Fazil Altinel, Mete Ozay, and Takayuki Okatani. Deep structured energy-based image inpainting. In ICPR, pages 423-428. IEEE Computer Society, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 296, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 296, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 296, + 146 + ], + "type": "text", + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 295, + 179 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: text-driven layered image and video editing. In ECCV, pages 707-723. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "text", + "content": "[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B. Goldman. Patchmatch: a randomized correspondence algorithm for structural image editing. ACM TOG, 28(3):24, 2009. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "type": "text", + "content": "[5] Chenjie Cao, Qiaole Dong, and Yanwei Fu. ZITS++: image inpainting by improving the incremental transformer on structural priors. IEEE TPAMI, 45(10):12667-12684, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 269, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 295, + 313 + ], + "type": "text", + "content": "[6] Benlei Cui, Xue-Mei Dong, Qiaoqiao Zhan, Jiangtao Peng, and Weiwei Sun. LitedealthWisenet: A lightweight network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 315, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 346 + ], + "type": "text", + "content": "[7] Ding Ding, Sundaresh Ram, and Jeffrey J. Rodríguez. Image inpainting using nonlocal texture matching and nonlinear filtering. IEEE TIP, 28(4):1705-1719, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 379 + ], + "type": "text", + "content": "[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 380, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 435 + ], + "type": "text", + "content": "[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, Dong Chen, and Baining Guo. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, pages 12709-12720. IEEE, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 436, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 436, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 57, + 436, + 295, + 458 + ], + "type": "text", + "content": "[10] James Hays and Alexei A. Efros. Scene completion using millions of photographs. ACM TOG, 26(3):4, 2007. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 459, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 295, + 502 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "text", + "content": "[12] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, and Ying Shan. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR. IEEE, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "type": "text", + "content": "[13] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM TOG, 36(4):107:1-107:14, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "text", + "content": "[14] Jitesh Jain, Yuqian Zhou, Ning Yu, and Humphrey Shi. Keys to better image inpainting: Structure and texture go hand in hand. In WACV, pages 208-217, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 625, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 689 + ], + "type": "text", + "content": "[15] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T. Freeman, David Salesin, Brian Curless, and Ce Liu. SLIDE: single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12498-12507. IEEE, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "type": "text", + "content": "[18] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Tom Duerig, and Vittorio Ferrari. The open images dataset V4: unified image classification, object detection, and visual relationship detection at scale. CoRR, abs/1811.00982, 2018. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[19] Anat Levin, Assaf Zomet, and Yair Weiss. Learning how to inpaint from global image statistics. In ICCV, pages 305-312, 2003. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[20] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, pages 7757-7765. Computer Vision Foundation / IEEE, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[21] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. ZONE: zero-shot instruction-guided local editing. In CVPR, pages 6254-6263. IEEE, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "type": "text", + "content": "[22] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. MAT: mask-aware transformer for large hole image inpainting. In CVPR, pages 10748-10758, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11451-11461. IEEE, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "text", + "content": "[24] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-07-10. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[25] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544. IEEE Computer Society, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "text", + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[29] Andranik Sargsyan, Shant Navasardyan, Xingqian Xu, and Humphrey Shi. MI-GAN: A simple baseline for image inpainting on mobile devices. In ICCV, pages 7301-7311, 2023. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[30] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879. IEEE, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[31] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[32] Wenhao Sun, Benlei Cui, Xue-Mei Dong, and Jingqun Tang. Attentive eraser: Unleashing diffusion model's object re" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2426" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "moyal potential via self-attention redirection guidance. In AAAI, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 163 + ], + "type": "text", + "content": "[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 3172–3182, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 295, + 232 + ], + "type": "text", + "content": "[34] Jingqun Tang, Su Qiao, Benlei Cui, Yuhang Ma, Sheng Zhang, and Dimitrios Kanoulas. You can even annotate text with voice: Transcription-only-supervised text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4154-4163, New York, NY, USA, 2022. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 233, + 294, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 309 + ], + "type": "text", + "content": "[35] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369. IEEE, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 312, + 294, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 294, + 355 + ], + "type": "text", + "content": "[36] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Learning 3d geometry and feature consistent gaussian splatt-ting for object removal. In ECCV, pages 1-17. Springer, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 357, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 357, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 294, + 402 + ], + "type": "text", + "content": "[37] Daniel Winter, Matan Cohen, Shlomi Fruchtner, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. In ECCV, pages 112-129. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "text", + "content": "[38] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 450, + 294, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 450, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 56, + 450, + 294, + 493 + ], + "type": "text", + "content": "[39] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437. IEEE, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 496, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 496, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 496, + 294, + 540 + ], + "type": "text", + "content": "[40] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin C. K. Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. CoRR, abs/2312.03771, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 586 + ], + "type": "text", + "content": "[41] Xingqian Xu, Shant Navasardyan, Vahram Tadevosyan, Andranik Sargsyan, Yadong Mu, and Humphrey Shi. Image completion with heterogeneously filtered spectral hints. In WACV, pages 4580-4590, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 588, + 294, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 294, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 294, + 631 + ], + "type": "text", + "content": "[42] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "text", + "content": "[43] Siyuan Yang, Lu Zhang, Liqian Ma, Yu Liu, Jingjing Fu, and You He. Magicremover: Tuning-free text-guided image inpainting with diffusion models. CoRR, abs/2310.02848, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[44] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, pages 7505-7514, 2020. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 397 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "text", + "content": "[45] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. CoRR, abs/2304.03246, 2023. 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[46] Donggeun Yoon and Donghyeon Cho. CORE-MPI: consistency object removal with embedding multiplane image. In CVPR, pages 20081-20090. IEEE, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "text", + "content": "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 197, + 553, + 239 + ], + "type": "text", + "content": "[48] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I-Chao Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In ICLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "type": "text", + "content": "[49] Hao Zhou, Chuanping Hu, Chongyang Zhang, and Shengyang Shen. Visual relationship recognition via language and position guided attention. In ICASSP, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "text", + "content": "[50] Hao Zhou, Chongyang Zhang, Yan Luo, Yanjun Chen, and Chuanping Hu. Embracing uncertainty: Decoupling and debias for robust temporal grounding. In CVPR, pages 8445-8454, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 320, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 320, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 320, + 553, + 354 + ], + "type": "text", + "content": "[51] Tong Zhou, Changxing Ding, Shaowen Lin, Xinchao Wang, and Dacheng Tao. Learning oracle attention for high-fidelity face completion. In CVPR, pages 7677-7686, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "type": "text", + "content": "[52] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 5, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2427" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_content_list.json b/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c1c7da6847b69bb996ab869e365f287b65bfc376 --- /dev/null +++ b/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_content_list.json @@ -0,0 +1,1645 @@ +[ + { + "type": "text", + "text": "Erasing Undesirable Influence in Diffusion Models", + "text_level": 1, + "bbox": [ + 241, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jing Wu $^{1}$ , Trung Le $^{1}$ , Munawar Hayat $^{2}$ , Mehrtash Harandi $^{1}$ $^{1}$ Monash University, Melbourne, VIC, Australia, $^{2}$ Qualcomm, San Deigo, CA, US \n{jing.wu1, trunglm, mehrtash.harandi}@monash.edu, hayat@qti.Qualcomm.com", + "bbox": [ + 174, + 180, + 823, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17ae78a31740d4b9de3e311438bca2c72819f358973e92ee65de28f126d1caf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 286, + 232, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/112ca5af0c57a515ab46f0c02b96e87d6ccb29b94fdd4b546684b020963e5578.jpg", + "image_caption": [ + "Figure 1. Top to Bottom: generated samples by SD v1.4 and model scrubbed by our method, when erasing the concept of 'nudity'. Our method can avoid NSFW (not safe for work) content while preserving model utility." + ], + "image_footnote": [], + "bbox": [ + 96, + 385, + 232, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46651389b1cebd84f1d8b5c41e8306608112e2649f0608ef1ffb4cd97acadb8e.jpg", + "image_caption": [ + "I2P prompts" + ], + "image_footnote": [], + "bbox": [ + 233, + 287, + 367, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/775af8103bcd653a1c4507557af163384bb954287a16316b8aab955358cea7d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 385, + 367, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d079d5f12be466c8d22260e31dbe7cd289d4bac356b0bd7cf634f60f48e4119.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 286, + 500, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94b147510b9c83c4d921141cc1c58f5f87b9de82e0ee20960bc34fb8fe6eab1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 385, + 500, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22f0da26f62de723d6de6c75a1720fd07c6dec1d63395f6eeb0ed676cfb20db1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 286, + 643, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12136886ef4f30ac41cb0bff03c0d05e5bf4aa5a115ca05d76dbc8f7329ed1b1.jpg", + "image_caption": [ + "COCO prompts" + ], + "image_footnote": [], + "bbox": [ + 645, + 286, + 772, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/444ae9d1d59ca02f32f037433f72ad07f919a10e9e0734206a955068ae06af6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 286, + 901, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c847e1f63febefb6ff5cfd5330894d12ceb8294736b4c33a09340279ed2c3f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 385, + 901, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/718c3957d05a8fe75a0f23ddf41b912c9155eee4d46ea66b0699d27c88ebebab.jpg", + "image_caption": [ + "Added by authors for publication" + ], + "image_footnote": [], + "bbox": [ + 622, + 491, + 681, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 564, + 326, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models are highly effective at generating high-quality images but pose risks, such as the unintentional generation of NSFW (not safe for work) content. Although various techniques have been proposed to mitigate unwanted influences in diffusion models while preserving overall performance, achieving a balance between these goals remains challenging. In this work, we introduce EraseDiff, an algorithm designed to preserve the utility of the diffusion model on retained data while removing the unwanted information associated with the data to be forgotten. Our approach formulates this task as a constrained optimization problem using the value function, resulting in a natural first-order algorithm for solving the optimization problem. By altering the generative process to deviate away from the groundtruth denoising trajectory, we update parameters for preservation while controlling constraint reduction to ensure effective erasure, striking an optimal trade-off. Extensive experiments and thorough comparisons with state-of-the-art algorithms demonstrate that EraseDiff effectively preserves the model's utility, efficacy, and efficiency.", + "bbox": [ + 86, + 598, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "WARNING: This paper contains sexually explicit imagery that may be offensive in nature.", + "bbox": [ + 511, + 565, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 631, + 645, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion Models [30, 49, 60] are now the method of choice in deep generative models, owing to their high-quality output, stability, and ease of training procedure. This has facilitated their successful integration into commercial applications such as midjourney. Unfortunately, the ease of use associated with diffusion models brings forth significant privacy risks. Studies have shown that these models can memorize and regenerate individual images from their training datasets [10, 58, 59]. Beyond privacy, diffusion models are susceptible to misuse and can generate NSFW digital content [48, 52, 54]. In this context, individuals whose images are used for training might request the removal of their private data. In particular, data protection regulations like the European Union General Data Protection Regulation (GDPR) [64] and the California Consumer Privacy Act (CCPA) [24] grant users the right to be", + "bbox": [ + 509, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28263", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forgotten, obligating companies to expunge data pertaining to a user upon receiving a request for deletion. These legal provisions grant data owners the right to remove their data from trained models and eliminate its influence on said models [4, 11, 23, 25, 44, 56, 61, 62, 69].", + "bbox": [ + 88, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A straightforward solution is to retrain the model from scratch after excluding the data that needs to be forgotten. However, the removal of pertinent data followed by retraining diffusion models from scratch demands substantial resources and is often deemed impractical. A version of the stable diffusion model trained on subsets of the LAION-5B dataset [55] costs approximately 150,000 GPU hours with 256 A100 GPUs1. Existing research on erasing unwanted influence has primarily focused on classification tasks [4, 7, 11, 22, 23, 25, 33, 44, 56, 68]. Despite substantial progress, prior methods developed in classification are observed to be ineffective for generation tasks [16]. Consequently, there is a pressing need for the development of methods capable of scrubbing data from diffusion models without necessitating complete retraining.", + "bbox": [ + 89, + 167, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a handful of studies [6, 16, 20, 21, 27, 28, 36, 41, 71] target unlearning in diffusion models, with a primary focus on the text-to-image models [20, 21, 71]. Broadly, these methods aim to achieve two main objectives: erasing data influence and preserving overall model performance. However, as demonstrated by Bui et al. [5], balancing this trade-off remains challenging.", + "bbox": [ + 89, + 393, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose EraseDiff, an algorithm tailored to balance the overall performance of diffusion models with the erasure of undesirable information. Drawing inspiration from optimization-based meta-learning algorithms [18, 42] that enable fast adaptation to new learning tasks, we formulate this challenge as a bi-level optimization problem, where the \"inner\" optimization focus on erasing undesirable influence and the \"outer\" objective seeks to preserve model performance. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively. However, this nested optimization can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with non-convex functions, making it difficult to achieve a stable solution [39]. Therefore, we further reformulate the problem as a constrained optimization problem using the value function [40, 46, 70], which facilitates a natural first-order solution [39], allows us to optimize preservation and erasure in a unified manner. This approach achieves a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. We benchmark EraseDiff on various scenarios, encompassing unlearning of classes on CIFAR-10 [35] with Denoising Diffusion Probabilistic Models (DDPM) [30], classes on", + "bbox": [ + 91, + 501, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imagenette [31] and concepts on the I2P dataset [54] with stable diffusion. Our empirical findings show that EraseDiff is $11 \\times$ faster than Heng and Soh's method [28] and $2 \\times$ faster than Fan's method [16] when forgetting on DDPM while achieving better unlearning results across several metrics. The results demonstrate that EraseDiff is capable of effectively erasing data influence in diffusion models, ranging from specific classes to the concept of nudity.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 511, + 227, + 640, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we outline the components of the models we evaluate, including DDPM and latent diffusion models [49]. Throughout the paper, we denote scalars, and vectors/matrices by lowercase and bold symbols, respectively (e.g., $a$ , $a$ , $A$ ).", + "bbox": [ + 511, + 252, + 903, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DDPM. (1) Diffusion: DDPM gradually diffuses the data distribution $\\mathbb{R}^d\\ni \\mathbf{x}_0\\sim q(\\mathbf{x})$ into the standard Gaussian distribution $\\mathbb{R}^d\\ni \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)$ with $T$ time steps, i.e., $q(\\mathbf{x}_t|\\mathbf{x}_{t - 1}) = \\mathcal{N}(\\mathbf{x}_t;\\sqrt{\\alpha_t}\\mathbf{x}_{t - 1},(1 - \\alpha_t)\\mathbf{I}_d)$ , where $\\alpha_{t} = 1 - \\beta_{t}$ and $\\{\\beta_t\\}_{t = 1}^T$ are the pre-defined variance schedule. The diffusion takes the form $\\mathbf{x}_t$ as $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon$ , where $\\bar{\\alpha}_t = \\prod_{i = 1}^{t}\\alpha_{i}$ . (2) Training: A model $\\epsilon_{\\theta}(\\cdot)$ with parameters $\\pmb {\\theta}\\in \\mathbb{R}^n$ is trained to learn the reverse process $p_{\\pmb{\\theta}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)\\approx q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)$ . Given $\\mathbf{x}_0\\sim q(\\mathbf{x})$ and time step $t\\in [1,T]$ , the simplified training objective is to minimize the distance between $\\pmb{\\epsilon}$ and the predicted $\\epsilon_{t}$ given $\\mathbf{x}_0$ at time $t$ , i.e., $\\| \\epsilon -\\epsilon_{\\pmb{\\theta}}(\\mathbf{x}_t,t)\\|$ . (3) Sampling: after training the model, we could obtain the learnable backward distribution $p_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_{t}) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\pmb{\\mu}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t),\\pmb{\\Sigma}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t))$ , where $\\pmb{\\mu}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t) = \\frac{1}{\\sqrt{\\alpha_{t}}} (\\mathbf{x}_{t} - \\frac{\\beta_{t}}{\\sqrt{1 - \\alpha_{t}}}\\epsilon_{\\pmb{\\theta}}(\\mathbf{x}_{t},t))$ and $\\pmb{\\Sigma}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t) = \\frac{(1 - \\bar{\\alpha}_{t - 1})\\beta_{t}}{1 - \\bar{\\alpha}_{t}}$ . Then, given $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)$ , $\\mathbf{x}_0$ could be obtained via sampling from $p_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_{t})$ from $t = T$ to $t = 1$ step by step.", + "bbox": [ + 511, + 349, + 903, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latent diffusion models. Latent diffusion models apply the diffusion models in the latent space $\\mathbf{z}$ of a pre-trained variational autoencoder. The noise would be added to $\\mathbf{z} = \\varepsilon(\\mathbf{x})$ , instead of the data $\\mathbf{x}$ , and the denoised output would be transformed to image space with the decoder. Besides, text embeddings generated by models like CLIP are used as conditioning inputs.", + "bbox": [ + 511, + 647, + 905, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Diffusion Unlearning", + "text_level": 1, + "bbox": [ + 511, + 768, + 714, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $\\mathcal{D} = \\{\\mathbf{x}_i, c_i\\}_i^N$ be a dataset of images $\\mathbf{x}_i$ associated with label $c_i$ representing the class. $\\mathcal{C} = \\{1, \\dots, C\\}$ denotes the label space where $C$ is the total number of classes and $c_i \\in \\mathcal{C}$ . We split the training data $\\mathcal{D}$ into the forgetting data $\\mathcal{D}_f \\subset \\mathcal{D}$ and its complement, remaining data $\\mathcal{D}_r = \\mathcal{D} \\setminus \\mathcal{D}_f$ . The forgetting data has label space $\\mathcal{C}_f \\subseteq \\mathcal{C}$ , and the remaining label space is denoted as $\\mathcal{C}_r = \\mathcal{C} \\setminus \\mathcal{C}_f$ .", + "bbox": [ + 511, + 794, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://stablediffusion.gitbook.io/overview/stable-diffusion-overview/technology/training-procedures", + "bbox": [ + 89, + 875, + 429, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28264", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Training objective", + "text_level": 1, + "bbox": [ + 89, + 90, + 267, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to scrub the information about $\\mathcal{D}_f$ carried by the diffusion models while maintaining the model utility over the remaining data $\\mathcal{D}_r$ . To achieve this, we adopt different training objectives for $\\mathcal{D}_r$ and $\\mathcal{D}_f$ as follows.", + "bbox": [ + 89, + 112, + 483, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the remaining data $\\mathcal{D}_r$ , we fine-tune the diffusion models with the original objective:", + "bbox": [ + 89, + 172, + 482, + 203 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\in \\mathcal {N} \\left(\\mathbf {0}, \\mathbf {I} _ {d}\\right), \\left(\\mathbf {x} _ {0}, c\\right) \\sim \\mathcal {D} _ {r} \\times \\mathcal {C} _ {r}} [ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {t} | c\\right) \\| ^ {\\frac {2}{2}} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 209, + 482, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\epsilon}$ . For the forgetting data $\\mathcal{D}_f$ , we aim to let the models fail to generate meaningful images corresponding to $\\mathcal{C}_f$ and thus propose:", + "bbox": [ + 89, + 250, + 483, + 296 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}) = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\in \\mathcal {N} \\left(\\mathbf {0}, \\mathbf {I} _ {d}\\right), \\left(\\mathbf {x} _ {0}, c\\right) \\sim \\mathcal {D} _ {f} \\times \\mathcal {C} _ {f}} \\left[ \\| \\boldsymbol {\\epsilon} _ {f} - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {t} | c\\right) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 301, + 486, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{f} = \\epsilon_{\\theta}(\\mathbf{x}_{t}|c_{m})$ and $c_{m}\\neq c$ so that the denoised image $\\mathbf{x}_0$ is not related to the forgetting class/concept $c$ [16, 28]. With this, we hinder the approximator $\\epsilon_{\\theta}$ to guide the denoising process to obtain meaningful examples for the forgetting data example $\\mathbf{x}_0\\sim \\mathcal{D}_f$", + "bbox": [ + 89, + 343, + 483, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To erase the undesirable influence of $\\mathcal{D}_f$ and preserve the overall performance, it is common to form", + "bbox": [ + 89, + 419, + 482, + 448 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) + \\lambda \\mathcal {L} _ {f} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 455, + 482, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\lambda > 0$ as the optimization objective (see for example [16]). However, training could be hindered due to the conflicting gradients between the erasing and preservation objectives, preventing a balanced trade-off between erasure and preservation [38]. To address this, rather than scalarizing the two objectives, we consider a framework based on optimization-based meta-learning algorithm [47] that allows iteratively updates to optimize each objective:", + "bbox": [ + 89, + 479, + 483, + 601 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r})\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 608, + 336, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "s.t. $\\pmb {\\theta}\\in \\underset {\\phi}{\\arg \\min}\\mathcal{L}_f(\\pmb {\\phi};\\mathcal{D}_f)$ (4)", + "bbox": [ + 184, + 628, + 482, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the outer objective minimizes the remaining loss $\\mathcal{L}_r$ (i.e., preserving model utility), the inner optimization minimizes the forgetting loss $\\mathcal{L}_f$ (i.e., erasing) with initialization $\\mathbb{R}^n\\ni \\phi_{\\mathrm{init}} = \\pmb{\\theta}$ . Given $\\pmb{\\theta}$ , the inner optimization on $\\phi$ aims to minimize the forgetting data influence, with the goal of achieving effective erasure while preserving model utility. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively.", + "bbox": [ + 89, + 659, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the above framework allows for iterative updates to address the conflicting objectives of erasure and preservation, it still relies on nested optimization, which can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with nonconvex functions, making it difficult to achieve a stable solution [39]. To further streamline the optimization process,", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we adopt a value function approach [40, 46, 70] that reformulates the problem as a single-constrained optimization:", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r})\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 143, + 723, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "s.t. $\\mathcal{L}_f(\\pmb {\\theta};\\mathcal{D}_f) - \\min_{\\phi}\\mathcal{L}_f(\\phi ;\\mathcal{D}_f)\\leq 0,$ (5)", + "bbox": [ + 571, + 162, + 903, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi$ is initialized at $\\theta$ , leverages the value function to encapsulate the influence of data erasure directly as a constraint on $\\mathcal{L}_f$ . This avoids the need for a nested loop by capturing the forgetting objective as a constraint and provides a natural first-order solution, as the constrained formulation allows us to optimize preservation and erasure in a unified manner.", + "bbox": [ + 511, + 185, + 906, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Solution", + "text_level": 1, + "bbox": [ + 513, + 299, + 614, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To solve Eq. (5), let us first denote $g(\\pmb{\\theta}) \\coloneqq \\mathcal{L}_f(\\pmb{\\theta};\\mathcal{D}_f) - \\min_{\\phi}\\mathcal{L}_f(\\phi ;\\mathcal{D}_f)$ . Our goal is to erase undesirable influence while preserving the overall model performance, hence the update vector $\\delta_t$ for updating the model should aid in minimizing $\\mathcal{L}_r(\\pmb{\\theta};\\mathcal{D}_r)$ and $g(\\pmb{\\theta})$ simultaneously. In other words, suppose that the current solution for Eq. (5) is $\\pmb{\\theta}_t$ , we aim to update $\\pmb{\\theta}_{t + 1} = \\pmb{\\theta}_t - \\eta \\pmb{\\delta}_t$ where $\\eta$ is sufficiently small, so that $\\mathcal{L}_r(\\pmb{\\theta}_{t + 1};\\mathcal{D}_r)$ decreases (i.e., preserve model utility) and $g(\\pmb{\\theta}_{t + 1})$ decreases (i.e., erasure). To this end, we aim to find the update vector $\\delta_t$ by:", + "bbox": [ + 511, + 320, + 906, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\delta} _ {t} \\in \\frac {1}{2} \\operatorname {a r g m i n} _ {\\boldsymbol {\\delta}} \\left\\| \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} \\left(\\boldsymbol {\\theta} _ {t}; \\mathcal {D} _ {r}\\right) - \\boldsymbol {\\delta} \\right\\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 478, + 861, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "s.t. $\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_t)^\\top \\pmb{\\delta} \\geq a_t > 0.$ (6)", + "bbox": [ + 557, + 508, + 903, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This will ensure that the update $\\delta_t$ is close to $\\nabla_\\theta \\mathcal{L}_r(\\pmb{\\theta}_t; \\mathcal{D}_r)$ and decreases $g(\\pmb{\\theta}_t)$ until it reaches stationary. Because $g(\\pmb{\\theta}_{t+1}) - g(\\pmb{\\theta}_t) \\approx -\\eta \\nabla_\\theta g(\\pmb{\\theta}_t)^\\top \\delta \\leq -\\eta a_t < 0$ for some scalar $a_t > 0$ , we can ensure that $g(\\pmb{\\theta}_{t+1}) < g(\\pmb{\\theta}_t)$ for small step size $\\eta > 0$ . This means that the update $\\delta_t$ can ensure to minimize $\\mathcal{L}_f(\\pmb{\\theta}; \\mathcal{D}_f)$ as long as it does not conflict with descent of $\\mathcal{L}_r(\\pmb{\\theta}; \\mathcal{D}_r)$ .", + "bbox": [ + 511, + 532, + 905, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To find the solution to the optimization problem in Eq. (6), the following theorem is developed:", + "bbox": [ + 511, + 641, + 905, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 3.1. The optimal solution of the optimization problem in Eq. (6) is $\\delta^{*} = \\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r}) + \\lambda_{t}\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})$ where $\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}$ .", + "bbox": [ + 511, + 676, + 906, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide the proof in §7 in the Appendix. This provides the solution to the optimization problem by constructing the update vector $\\delta$ to balance two competing objectives. The variable $a_{t}$ adjusts the weight of the forgetting objective, ensuring that the update vector $\\delta$ decreases the remaining loss without violating the erasure goal, and achieves the dual goals of maintaining utility and achieving erasing. In practice, we can choose $a_{t} = \\eta \\| \\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}$ , and we start from $\\phi^0 = \\pmb{\\theta}_t$ and use gradient descend in $K$ steps with the learning rate $\\xi > 0$ to reach $\\phi^K$ , namely $\\phi^{k + 1} = \\phi^k - \\xi \\nabla_\\phi \\mathcal{L}_f(\\phi^k; \\mathcal{D}_f)$ and $k = 0, \\dots, K - 1$ .", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28265", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 EraseDiff: Erasing undesirable influence in diffusion models.", + "bbox": [ + 91, + 90, + 480, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Well-trained model with parameters $\\theta_0$ , forgetting data $\\mathcal{D}_f$ and remaining data $\\mathcal{D}_r$ , outer iteration number $T$ and inner iteration number $K$ , learning rate $\\eta$ .", + "bbox": [ + 91, + 122, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Parameters $\\theta^{*}$ for the scrubbed model.", + "bbox": [ + 91, + 167, + 411, + 181 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for iteration $t$ in $T$ do", + "2: $\\phi^0 = \\pmb {\\theta}_t$", + "3: Get $\\phi^K$ by $K$ steps of gradient descent on $\\mathcal{L}_f(\\phi; \\mathcal{D}_f)$ starting from $\\phi^0$ .", + "4: Set $g(\\pmb{\\theta}_t) = \\mathcal{L}_f(\\pmb{\\theta}_t; \\mathcal{D}_f) - \\mathcal{L}_f(\\phi^K; \\mathcal{D}_f)$ .", + "5: Update the model: $\\pmb{\\theta}_{t + 1} = \\pmb{\\theta}_t - \\eta (\\nabla_{\\pmb{\\theta}_t}\\mathcal{L}_r(\\pmb {\\theta}_t;\\mathcal{D}_r) + \\lambda_t\\nabla_{\\pmb{\\theta}_t}g(\\pmb {\\theta}_t;\\phi^K))$", + "6: where $\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})^{T}\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}$ .", + "7: end for" + ], + "bbox": [ + 101, + 184, + 480, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Analysis", + "text_level": 1, + "bbox": [ + 91, + 353, + 192, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can characterize the solution of our algorithm as follows and the proof can be found in §7 in the Appendix:", + "bbox": [ + 89, + 376, + 482, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.2 (Pareto optimality). The stationary point obtained by our algorithm is Pareto optimal of the problem $\\min_{\\pmb{\\theta}}[\\mathcal{L}_r(\\pmb{\\theta};\\mathcal{D}_r),\\mathcal{L}_f(\\pmb{\\theta};\\mathcal{D}_f)]$", + "bbox": [ + 89, + 417, + 483, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This asserts that the solution obtained by our algorithm is Pareto optimal for the problem of minimizing both objectives, which implies that the solution obtained by the algorithm ensures a balanced trade-off between preserving model utility and erasing undesirable influences.", + "bbox": [ + 88, + 476, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further take DDPM with CIFAR-10 when forgetting the 'airplane' as an example to show that our proposed method helps alleviate the gradient conflict which prevents a balanced trade-off between erasure and preservation. Fig. 2 presents the cosine similarity between the update vector $\\delta$ and the gradient $g_{r} = \\nabla_{\\theta}\\mathcal{L}_{r}(\\theta ;\\mathcal{D}_{r})$ for preservation, and the cosine similarity between the update vector $\\delta$ and the gradient $g_{f} = \\nabla \\mathcal{L}_{f}(\\theta ;\\mathcal{D}_{f})$ for erasing. The cosine similarity represents the alignment between the update vector and the gradients associated with the preservation and erasure. Higher positive values indicate alignment, meaning that the update vector $\\delta$ is directed similarly to the respective gradient, whereas negative values indicate misalignment or conflict. In particular, negative values suggest a high degree of opposition between the update vector and the respective gradient, which can signify competing objectives between preservation and erasure during the optimization process.", + "bbox": [ + 89, + 551, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MOO (Multi-Objective Optimization) denotes the naive integration of erasing and preservation as stated in Eq. (3). For the vanilla MOO, Fig. 2 shows a clear alternating pattern in cosine similarity values between the update vector $\\delta$ and the gradients $g_{r}$ and $g_{f}$ . Specifically, when the cosine", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b5c502d8fe4a19dd98c2fc18080b49c434130f3803283a2c87acbc90e92fab90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 95, + 895, + 286 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10377f078ff0c8c707ad539046dee7ac2a2c2efa1cef9481f0d73191f5482acc.jpg", + "image_caption": [ + "Figure 2. Top to bottom: cosine similarity between the update vector $\\delta$ and the preservation gradient $g_{r}$ , followed by the cosine similarity between $\\delta$ and the erasing gradient $g_{f}$ . Positive values indicate alignment, while negative values suggest conflict. This visualization illustrates how well the update vector aligns with the objectives of preservation and erasure over successive iterations." + ], + "image_footnote": [], + "bbox": [ + 531, + 291, + 895, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "similarity between $\\delta$ and $g_{r}$ is greater than 0, the similarity between $\\delta$ and $g_{f}$ tends to be less than 0, and vice versa. This pattern suggests that MOO experiences gradient conflict, as it cannot effectively balance the two objectives of preservation and erasure, preventing MOO from achieving a harmonious update that supports both goals simultaneously. In contrast, EraseDiff mostly shows positive cosine similarity values between the update vector $\\delta$ and both the preservation gradient $g_{r}$ and the erasing gradient $g_{f}$ . This indicates that EraseDiff aligns the update direction with both objectives, suggesting it manages to avoid significant gradient conflict. By maintaining positive alignment, EraseDiff appears to balance preservation and erasure more effectively, leading to better cooperation between objectives.", + "bbox": [ + 511, + 614, + 906, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Related Work", + "text_level": 1, + "bbox": [ + 511, + 843, + 653, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memorization in generative models. Privacy of generative models has been studied extensively for GANs [17, 43,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28266", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "66] and generative language models [8, 9, 32, 63]. These generative models often risk replicating from their training data. Recently, several studies [10, 58, 59, 65] investigated these data replication behaviors in diffusion models, raising concerns about the privacy and copyright issues. Possible mitigation strategies are deduplicating and randomizing conditional information [58, 59], or training models with differential privacy (DP) [1, 13-15]. However, leveraging DP-SGD [1] may cause training to diverge [10].", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Malicious misuse. Diffusion models usually use training data from varied open sources and when such unfiltered data is employed, there is a risk of it being tainted [12] or manipulated [48], resulting in inappropriate generation [54]. They also risk the imitation of copyrighted content, e.g., mimicking the artistic style [20, 57]. To counter inappropriate generation, data censoring [3, 19, 45, 53] where excluding black-listed images before training, and safety guidance where diffusion models will be updated away from the inappropriate/undesired concept [20, 54] are proposed. [34] uses human feedback to annotate the generations for getting soft tokens that quantify the harmfulness of the images, then fine-tunes the model using self-distillation to achieve harmony between preservation and erasure. Shan et al. [57] propose protecting artistic style by adding barely perceptible perturbations to the artworks before public release. Yet, Rando et al. [48] argue that DMs can still generate content that bypasses the filter. Chen et al. [12] highlight the susceptibility of DMs to poison attacks, where target images are generated with specific triggers.", + "bbox": [ + 91, + 229, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Machine unlearning. Removing data directly involves retraining the model from scratch, which is inefficient and impractical. Thus, to reduce the computational overhead, efficient machines unlearning methods [4, 7, 11, 22, 23, 25, 33, 44, 50, 56, 61, 68] have been proposed. Several studies [5, 16, 20, 21, 27, 28, 67, 71] recently introduce unlearning in diffusion models. Most of them [20, 21, 27, 71] mainly focus on text-to-image models and high-level visual concept erasure. Heng and Soh [28] adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) from continual learning to perform unlearning effectively without access to the training data. Heng and Soh's method can be applied to a wide range of generative models, however, it needs the computation of FIM for different datasets and models, which may lead to significant computational demands. Fan et al. [16] propose a very potent unlearning algorithm called SalUn that shifts attention to important parameters w.r.t. the forgetting data. SalUn can perform effectively across image classification and generation tasks.", + "bbox": [ + 89, + 534, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we introduce a simple yet effective unlearning algorithm for diffusion models by formulating the problem as a constrained optimization problem, to achieve a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. Below, we will show", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that our algorithm is not only faster than Heng and Soh's method [28] and Fan's method [16], but even outperforms these methods in terms of the trade-off between the forgetting and preserving model utility.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 164, + 638, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate EraseDiff in various scenarios, including removing images with specific classes/concepts, to answer the following research questions (RQs): (i) Can typical machine unlearning methods be applied to diffusion models? (ii) Is EraseDiff able to remove the influence of $\\mathcal{D}_f$ in the diffusion models? (iii) Is EraseDiff able to preserve the model utility while removing $\\mathcal{D}_f$ ? (iv) Is EraseDiff efficient in removing the data? (v) How does EraseDiff perform on the public well-trained models?", + "bbox": [ + 511, + 189, + 906, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 334, + 596, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiments are reported on CIFAR-10 [35] with DDPM, Imagenette [31] with Stable Diffusion (SD) for class-wise forgetting, I2P [54] dataset with SD for concept-wise forgetting. For all SD experiments, we use the open-source SD v1.4 [49] checkpoint as the pre-trained model. Implementation details and additional results like visualizations of generated images can be found in §8 and §9.", + "bbox": [ + 511, + 356, + 905, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We primarily benchmark against the following baselines commonly used in machine unlearning: (i) Unscrubbed, (ii) Finetune (FT) [23], (iii) NegGrad (NG) [23], (iv) BlindSpot [61], (v) ESD [20], (vi) FMN [71], (vii) Selective Amnesia (SA) [28] and (viii) the SOTA machine unlearning algorithm SalUn [16].", + "bbox": [ + 511, + 462, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Several metrics are utilized to evaluate the algorithms: (i) Frechet Inception Distance (FID) [29]: the widely-used metric for assessing the quality of generated images. (ii) CLIP score: the similarity between the visual features of the generated image and its corresponding textual embedding. (iii) $P_{\\psi}(\\mathbf{y} = c_f|\\mathbf{x}_f)$ [28]: the classification rate of a pre-trained classifier $P_{\\psi}(\\mathbf{y}|\\mathbf{x})$ , with a ResNet architecture [26] used to classify generated images conditioned on the forgetting classes. A lower classification value indicates superior unlearning performance. (iv) Precision and Recall: A low FID may indicate high precision (realistic images) but low recall (small variations) [37, 51]. Kynkänniemi et al. [37] shows that generative models claim to optimize FID (high fidelity) but always sacrifice variation (low diversity). Hence, we include metric precision (fidelity) and recall (diversity) to express the quality of the generated samples, to provide explicit visibility of the tradeoff between sample quality and variety.", + "bbox": [ + 511, + 551, + 906, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Results on DDPM", + "text_level": 1, + "bbox": [ + 511, + 833, + 689, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following SA, we aim to forget the 'airplane' class on CIFAR-10. Here, we replace $\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)$ with $\\epsilon_{f} = \\epsilon_{\\theta}(\\mathbf{x}_{t}|c_{m})$ like random labelling used in [16] where $c_{m}\\neq c$", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28267", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results on CIFAR10 with DDPM when forgetting the 'airplane' class. $P_{\\psi}(\\mathbf{y} = c_f|\\mathbf{x}_f)$ indicate the probability of the forgotten class (i.e., the effectiveness of erasing). Precision and Recall demonstrate the fidelity and diversity [37, 51], and FID scores are computed between the generated 45K images and the corresponding ground truth images with the same labels from $\\mathcal{D}_r$ (i.e., preserving model utility). SA excels in class-wise forgetting but struggles to perform concept-wise forgetting as shown in Fig. 3 and Tab. 2. The best and the second best are highlighted in blue and orange, respectively.", + "bbox": [ + 89, + 88, + 906, + 159 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d18d88d6b9c7292a953de00d484b9f742099c5f31852be764c642abe6e2b3b0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UnscrubbedFT [23]NG [23]BlindSpot [61]SA [28]SalUn [16]EraseDiffrlEraseDiffnoise
FID ↓9.638.2176.739.128.199.168.667.61
Precision (fidelity) ↑0.400.430.080.410.430.410.430.43
Recall (diversity) ↑0.790.770.610.780.750.760.770.72
Pψ(y = cf|xf)↓0.970.960.610.900.060.070.240.22
", + "bbox": [ + 102, + 169, + 893, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1736ccbff23f646147064b068bc6a26b1f06e1358197e68507ddc6d68ddf9dfb.jpg", + "image_caption": [ + "Figure 3. Quantity of nudity content detected using the NudeNet classifier from I2P data. Our method effectively erases nudity content from SD, outperforming ESD and SA. Note that Fig. 4 and Tab. 2 together presents the trade-off between erasing and preservation." + ], + "image_footnote": [], + "bbox": [ + 99, + 280, + 380, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b80ded473f9a468bd3de678469e126074cd2ade9028534305a8083ef3cc4e020.jpg", + "image_caption": [ + "Table 2. Evaluation of $30\\mathrm{K}$ generated images by SD when erasing 'nudity'. The FID score is measured compared to validation data, while the CLIP similarity score evaluates the alignment between generated images and the corresponding prompts. The best and the second best are highlighted in blue and orange, respectively." + ], + "image_footnote": [], + "bbox": [ + 388, + 280, + 552, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3f99bc74074be9ed65566e4bf084f34cbffaddd925b8a8920c7c8b7a7a49359.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 281, + 723, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0369b42997825dc6c18aed22fc3078892a901db3cceabcf224f59d0493a1c38f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 281, + 897, + 436 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/74e5044fb53d52704a9f0e3f36a039e22404aad84e49b5b2a16c8efa8ac3b36c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ESD [20]SA [28]SalUn [16]EraseDiff
FID ↓15.7625.5825.0617.01
CLIP ↑30.3331.0328.9130.58
", + "bbox": [ + 109, + 580, + 465, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "denoted as $\\text{EraseDiff}_{\\text{rl}}$ . We also try to use $\\epsilon_f = \\mathcal{U}(\\mathbf{0}, \\mathbf{I}_d)$ like SA, denoted as $\\text{EraseDiff}_{\\text{noise}}$ . Note that the choice of replacement for forgotten classes is flexible and is not the primary focus of this work. For further discussion on the choice of substitution strategies, please refer to [6, 41].", + "bbox": [ + 89, + 672, + 482, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results are presented in Tab. 1. Firstly, from Tab. 1, we can conclude that traditional machine unlearning methods designed for image classification or regression tasks fall short in effectively performing forgetting for DDPM. Finetune and BlindSpot suffer from under-forgetting (i.e., the generated image quality is good but the probability of generated images belonging to the forgetting class approaching the value of the unscrubbed model), and NegGrad suffers from over-forgetting (the probability of generated images belonging to the forgetting class is decreased compared to", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that of the unscrubbed model but the generated image quality drops significantly).", + "bbox": [ + 511, + 503, + 903, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, comparing SA and SalUn's unlearning methods, SA achieves an FID score of 8.19 but sacrifices variation (decreased recall). Also, note that SA introduces excessive computational resource requirements and time consumption [28, 72]. Note that the FID scores of SA, SalUn, and EraseDiff decrease compared with the generated images from the original models; the quality of the generated images experiences a slight improvement. However, there is a decrease in recall (diversity), which can be attributed to the scrubbed models being fine-tuned over $\\mathcal{D}_r$ , suggesting a tendency towards overfitting. Regarding forgetting, SalUn achieves a smaller probability of the generated images classified as the forgetting class than ours; yet, the FID score is larger than ours, and images generated by EraseDiffr1 present better diversity and fidelity.", + "bbox": [ + 511, + 534, + 906, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results on Stable Diffusion", + "text_level": 1, + "bbox": [ + 511, + 771, + 756, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply EraseDiff to perform class-wise forgetting from Imagenette and erase the 'nudity' concept with SD v1.4. When forgetting 'nudity', we have no access to the training data; instead, we generate $\\sim 400$ images with the prompts $c_{f} = \\{\\text{'nudity', 'naked', 'erotic', 'sexual'}\\}$ .", + "bbox": [ + 511, + 794, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Forget nudity. 4703 images are generated using I2P prompts, and 1K images are generated using the prompts", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28268", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77f150cff340b7bd1a389bc17e00771d50359d87ce9cbe07b6d9838e1b9760e2.jpg", + "image_caption": [ + "Table 3. P forgetting", + "{nudity, nak, sexual} The quality of nu-dity content is detected using the NudeNet classifier [2]. In Fig. 3, the number in the y-axis denotes the number of exposed body parts generated by the SD v1.4 model. Fig. 3 presents the percentage change in exposed body parts w.r.t. SD v1.4. In §9, we provide the number of exposed body parts counted in all generated images with different thresholds. Here, our algorithm replaces $\\epsilon_{f}$ with $\\epsilon_{\\theta}(\\mathbf{x}_i|c_m)$ where $c_{m}$ is 'a photo of poker'. We can find that, EraseD-iff reduces the amount of nu-dity content compared to SD v1.4, ESD, and SA, particularly on sensitive content like Female/Male interests and female/Male Genitalia. While SalUn excels a 'forgetting', our algorithm demonstrates a significant improvement in the quality of generated images as shown in Tab. 2. Tab. 2 presented results evaluating the utility of scrubbed models. The FID and CLIP scores are" + ], + "image_footnote": [], + "bbox": [ + 151, + 45, + 986, + 694 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0d6c7c35f43ceedc354c7e64bdcc9066f2fba320467025a689a34fd5d17e7b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 915, + 983, + 1000 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/803feed9f157b6661ac68f6b1a065a08c3150f18b57b048210ab86218fef10ec.jpg", + "table_caption": [ + "Table 4. Computational overhead. Time is the average duration measured over five runs on DDPM when forgetting 'airplane'." + ], + "table_footnote": [], + "table_body": "
Memory (MiB)Time (min.)Complexity
SA3352.3140.00O(n2)
SalUn4336.228.17O(n)
EraseDiff3360.312.70O(n)
", + "bbox": [ + 104, + 128, + 470, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "approach, clearly indicate the advantages of our method. We also present results when improving the forgetting ability of SalUn in §9. However, note that this enhancement comes with a drop in the FID score of the generated images. Our method, while slightly better than SalUn on average across 10 classes, demonstrates a more balanced trade-off between erasing and preservation, indicating that it achieves a favorable balance in preserving fidelity while enhancing erasing performance.", + "bbox": [ + 88, + 231, + 483, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Computational efficiency", + "text_level": 1, + "bbox": [ + 89, + 378, + 318, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we measure the computational complexity of unlearning algorithms. The computational complexity of SA and SalUn involves two distinct stages: the computation of FIM for SA and the computation of salient weights w.r.t. $\\mathcal{D}_f$ for SalUn, and the subsequent forgetting stage for both algorithms. We consider the maximum memory usage across both stages, the metric 'Time' is exclusively associated with the duration of the forgetting stage for unlearning algorithms. Tab. 4 show that EraseDiff outperforms SA and SalUn in terms of efficiency, achieving a speed increase of $\\sim 11\\times$ than SA and $\\sim 2\\times$ than SalUn. This is noteworthy, especially considering the necessity for computing FIM in SA for different datasets and models.", + "bbox": [ + 88, + 401, + 483, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 608, + 241, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further investigate the influence of the number of iterations $K$ that approximate $\\min \\mathcal{L}_f(\\phi ;\\mathcal{D}_f)$ , and the step size $\\eta$ that controls the weight of forgetting and preserving model utility. Here, we replace $\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)$ with $\\epsilon_{f}\\in \\mathcal{U}(\\mathbf{0},\\mathbf{I}_{d})$ . Note that for different hyperparameters in Fig. 5 (a), the average entropy of the classifier's output distribution given $\\mathbf{x}_f$ , which is $H(P_{\\psi}(\\mathbf{y}|\\mathbf{x}_f)) = -\\mathbb{E}[\\sum_iP_{\\psi}(\\mathbf{y} = c_i|\\mathbf{x})\\log_eP_{\\psi}(\\mathbf{y} = c_i|\\mathbf{x})]$ , remains close to 2.02. This indicates that the scrubbed models become uncertain about the images conditioned on the forgetting class, effectively erasing the information about $\\mathcal{D}_f$ . Below, we will further demonstrate the influence on the model utility. In practice, we have $\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\theta}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\theta}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\} = \\max \\{0,\\eta -\\frac{\\nabla_{\\theta}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\theta}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}$ , we can see that $\\eta$ determines the extent to which the update direction for forgetting can deviate from that for preserving model utility. A larger $\\eta$ would allow for more deviation in the updating, thus pri", + "bbox": [ + 91, + 630, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbb94cf6ab7a3076aeebc533cd9b12764298cc2921b6e7cf3f2b7874e58ba66b.jpg", + "image_caption": [ + "Figure 5. (a) Ablation results. (b) Potential incomplete erasures." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 908, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "oritizing forgetting over preserving model utility. In Fig. 5 (a), the FID score tends to increase (i.e., image quality drop) as the step size $\\eta$ increases, indicating that larger $\\eta$ leads to greater deviations from the direction that preserves the model utility. Furthermore, the number of iterations $K$ determines how closely the approximation $\\phi^K$ will approach $\\arg \\min_{\\phi} \\mathcal{L}_f(\\phi; \\mathcal{D}_f)$ . Hence, a larger number of iterations $K$ leads to more thorough erasure, which is also supported by the results shown in Fig. 5 (a), as increasing $K$ correlates with an increase in the FID score.", + "bbox": [ + 511, + 315, + 906, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion, limitations and broader impacts", + "text_level": 1, + "bbox": [ + 511, + 479, + 906, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we explored erasing undesirable influence in diffusion models and proposed an efficient method EraseDiff to achieve a balanced trade-off between erasing and preservation. Comprehensive experiments on diffusion models demonstrate the proposed algorithm's effectiveness in data removal, its efficacy in preserving the model utility, and its efficiency in erasure.", + "bbox": [ + 511, + 505, + 905, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, our scrubbed model may have incomplete erasures, potentially arises due to the choice of the replacements. In Fig. 5 (b), with random labeling, generated images conditioned on the forgetting class 'tench' by our scrubbed model may preserve some characteristics similar to that close to 'tench' visually; instead, replacing the erased class 'tench' with other class such as 'dog' could improve in erasure. Besides, the scrubbed models could be biased or be misused for censorship or exploitation. This includes using technology to selectively remove or amplify NSFW content in various ways. They might also falsely assure users that NSFW content is fully removed when it isn't. To address this, advanced privacy-preserving training techniques are in demand to enhance the security and fairness of the models. Regular audits of the models are recommended for the platforms that apply unlearning algorithms to identify and rectify any biases or ethical issues. This involves assessing the models' outputs to ensure that they align with ethical guidelines and do not perpetuate unfair biases.", + "bbox": [ + 511, + 611, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28270", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mehrtash Harandi is supported by the Australian Research Council (ARC) Discovery Program DP250100262. The authors gratefully acknowledge the anonymous reviewers for their insightful feedback and valuable suggestions, which have significantly improved the quality of this work.", + "bbox": [ + 89, + 113, + 483, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308-318, 2016. 5", + "[2] P Bedapudi. Nudenet: Neural nets for nudity classification, detection and selective censoring, 2019. 7, 2", + "[3] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1536-1546. IEEE, 2021. 5", + "[4] Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159, 2021. 2, 5", + "[5] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 2, 5", + "[6] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 2, 6", + "[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480, 2015. 2, 5", + "[8] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633-2650, 2021. 5", + "[9] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022. 5", + "[10] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 5", + "[11] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 2, 5" + ], + "bbox": [ + 93, + 236, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 5", + "[13] Tim Dockhorn, Tianshi Cao, Arash Vahdat, and Karsten Kreis. Differentially private diffusion models. arXiv preprint arXiv:2210.09929, 2022. 5", + "[14] Cynthia Dwork. Differential privacy: A survey of results. In International conference on theory and applications of models of computation, pages 1-19. Springer, 2008.", + "[15] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, New York, NY, USA, March 4-7, 2006. Proceedings 3, pages 265-284. Springer, 2006. 5", + "[16] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Dennis Wei, Eric Wong, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3, 5, 6, 7", + "[17] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701–6710, 2021. 4", + "[18] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2", + "[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 5", + "[20] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In 2023 IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5, 6, 7", + "[21] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 5", + "[22] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2, 5", + "[23] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309, 2020. 2, 5, 6", + "[24] Eric Goldman. An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper, 2020. 1", + "[25] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28271", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. In Proceedings of the 37th International Conference on Machine Learning, pages 3832-3842. PMLR, 2020. 2, 5", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[27] Alvin Heng and Harold Soh. Continual learning for forgetting in deep generative models. 2023. 2, 5", + "[28] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 5, 6", + "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[31] Jeremy Howard and Sylvain Gugger. Fastai: A layered api for deep learning. Information, 11(2):108, 2020. 2, 5", + "[32] Matthew Jagielski, Om Thakkar, Florian Tramer, Daphne Ippolito, Katherine Lee, Nicholas Carlini, Eric Wallace, Shuang Song, Abhradeep Thakurta, Nicolas Papernot, et al. Measuring forgetting of memorized training examples. arXiv preprint arXiv:2207.00099, 2022. 5", + "[33] Masayuki Karasuyama and Ichiro Takeuchi. Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks, 21(7):1048-1059, 2010. 2, 5", + "[34] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Safeguard text-to-image diffusion models with human feedback inversion. In European Conference on Computer Vision, pages 128-145. Springer, 2024. 5", + "[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5", + "[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 2", + "[37] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5, 6", + "[38] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems (NeurIPS), 34:18878-18890, 2021. 3", + "[39] Bo Liu, Mao Ye, Stephen Wright, Peter Stone, and Qiang Liu. Bome! bilevel optimization made easy: A simple first-order approach. Advances in Neural Information Processing Systems, 35:17248-17262, 2022. 2, 3", + "[40] Risheng Liu, Xuan Liu, Xiaoming Yuan, Shangzhi Zeng, and Jin Zhang. A value-function-based interior-point method" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for non-convex bi-level optimization. In International conference on machine learning, pages 6882-6892. PMLR, 2021. 2, 3", + "[41] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 2, 6", + "[42] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In International conference on machine learning, pages 2113-2122. PMLR, 2015. 2", + "[43] Casey Meehan, Kamalika Chaudhuri, and Sanjoy Das-gupta. A non-parametric test to detect data-copying in generative models. In International Conference on Artificial Intelligence and Statistics, 2020. 4", + "[44] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N. Ravi. Deep unlearning via randomized conditionally independent hessenians. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10412-10421, 2022. 2, 5", + "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5", + "[46] Jiri V Outrata. On the numerical solution of a class of stackelberg problems. Zeitschrift für Operations Research, 34: 255-277, 1990. 2, 3", + "[47] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3", + "[48] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1, 5", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5", + "[50] Enrique Romero, Ignacio Barrio, and Lluis Belanche. Incremental and decremental learning for linear support vector machines. In International Conference on Artificial Neural Networks, pages 209-218. Springer, 2007. 5", + "[51] Mehdi SM Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. Advances in neural information processing systems, 31, 2018. 5, 6", + "[52] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023.1", + "[53] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content?" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28272", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350-1361, 2022. 5", + "[54] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 5", + "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[56] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS), 34: 18075-18086, 2021. 2, 5", + "[57] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 5", + "[58] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 1, 5", + "[59] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 5", + "[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1", + "[61] Ayush Kumar Tarun, Vikram Singh Chandawat, Murari Mandal, and Mohan Kankanhalli. Deep regression unlearning. In International Conference on Machine Learning, pages 33921-33939. PMLR, 2023. 2, 5, 6", + "[62] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[63] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 35: 38274-38290, 2022. 5", + "[64] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1", + "[65] Nikhil Vyas, Sham Kakade, and Boaz Barak. Provable copyright protection for generative models. arXiv preprint arXiv:2302.10870, 2023. 5", + "[66] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. This person (probably) exists. identity member-" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ship attacks against gan generated faces. arXiv preprint arXiv:2107.06018, 2021.5", + "[67] Jing Wu and Mehrtash Harandi. Scissorhands: Scrub data influence via connection sensitivity in networks. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 5", + "[68] Yinjun Wu, Edgar Dobriban, and Susan Davidson. DeltaGrad: Rapid retraining of machine learning models. In Proceedings of the 37th International Conference on Machine Learning, pages 10355-10366. PMLR, 2020. 2, 5", + "[69] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 87-103. Springer, 2022. 2", + "[70] Jane J Ye and DL Zhu. Optimality conditions for bilevel programming problems. Optimization, 33(1):9-27, 1995. 2, 3", + "[71] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2, 5, 7", + "[72] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024.6" + ], + "bbox": [ + 516, + 92, + 903, + 484 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28273", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_model.json b/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_model.json new file mode 100644 index 0000000000000000000000000000000000000000..344bbd2870cb02e721b5078349ab3fd5ac6f68da --- /dev/null +++ b/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.243, + 0.131, + 0.757, + 0.154 + ], + "angle": 0, + "content": "Erasing Undesirable Influence in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.181, + 0.825, + 0.236 + ], + "angle": 0, + "content": "Jing Wu\\(^{1}\\), Trung Le\\(^{1}\\), Munawar Hayat\\(^{2}\\), Mehrtash Harandi\\(^{1}\\) \n\\(^{1}\\)Monash University, Melbourne, VIC, Australia, \\(^{2}\\)Qualcomm, San Deigo, CA, US \n{jing.wu1, trunglm, mehrtash.harandi}@monash.edu, hayat@qti.Qualcomm.com" + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.27, + 0.35, + 0.287 + ], + "angle": 0, + "content": "I2P prompts" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.287, + 0.233, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.386, + 0.233, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.288, + 0.369, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.386, + 0.368, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.287, + 0.5, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.386, + 0.5, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.27, + 0.776, + 0.287 + ], + "angle": 0, + "content": "COCO prompts" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.287, + 0.645, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.287, + 0.773, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.287, + 0.903, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.386, + 0.903, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.492, + 0.683, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.493, + 0.902, + 0.51 + ], + "angle": 0, + "content": "Added by authors for publication" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.523, + 0.908, + 0.553 + ], + "angle": 0, + "content": "Figure 1. Top to Bottom: generated samples by SD v1.4 and model scrubbed by our method, when erasing the concept of 'nudity'. Our method can avoid NSFW (not safe for work) content while preserving model utility." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.565, + 0.327, + 0.58 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.599, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Diffusion models are highly effective at generating high-quality images but pose risks, such as the unintentional generation of NSFW (not safe for work) content. Although various techniques have been proposed to mitigate unwanted influences in diffusion models while preserving overall performance, achieving a balance between these goals remains challenging. In this work, we introduce EraseDiff, an algorithm designed to preserve the utility of the diffusion model on retained data while removing the unwanted information associated with the data to be forgotten. Our approach formulates this task as a constrained optimization problem using the value function, resulting in a natural first-order algorithm for solving the optimization problem. By altering the generative process to deviate away from the groundtruth denoising trajectory, we update parameters for preservation while controlling constraint reduction to ensure effective erasure, striking an optimal trade-off. Extensive experiments and thorough comparisons with state-of-the-art algorithms demonstrate that EraseDiff effectively preserves the model's utility, efficacy, and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.566, + 0.907, + 0.597 + ], + "angle": 0, + "content": "WARNING: This paper contains sexually explicit imagery that may be offensive in nature." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.632, + 0.646, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion Models [30, 49, 60] are now the method of choice in deep generative models, owing to their high-quality output, stability, and ease of training procedure. This has facilitated their successful integration into commercial applications such as midjourney. Unfortunately, the ease of use associated with diffusion models brings forth significant privacy risks. Studies have shown that these models can memorize and regenerate individual images from their training datasets [10, 58, 59]. Beyond privacy, diffusion models are susceptible to misuse and can generate NSFW digital content [48, 52, 54]. In this context, individuals whose images are used for training might request the removal of their private data. In particular, data protection regulations like the European Union General Data Protection Regulation (GDPR) [64] and the California Consumer Privacy Act (CCPA) [24] grant users the right to be" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28263" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "forgotten, obligating companies to expunge data pertaining to a user upon receiving a request for deletion. These legal provisions grant data owners the right to remove their data from trained models and eliminate its influence on said models [4, 11, 23, 25, 44, 56, 61, 62, 69]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.394 + ], + "angle": 0, + "content": "A straightforward solution is to retrain the model from scratch after excluding the data that needs to be forgotten. However, the removal of pertinent data followed by retraining diffusion models from scratch demands substantial resources and is often deemed impractical. A version of the stable diffusion model trained on subsets of the LAION-5B dataset [55] costs approximately 150,000 GPU hours with 256 A100 GPUs1. Existing research on erasing unwanted influence has primarily focused on classification tasks [4, 7, 11, 22, 23, 25, 33, 44, 56, 68]. Despite substantial progress, prior methods developed in classification are observed to be ineffective for generation tasks [16]. Consequently, there is a pressing need for the development of methods capable of scrubbing data from diffusion models without necessitating complete retraining." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Recently, a handful of studies [6, 16, 20, 21, 27, 28, 36, 41, 71] target unlearning in diffusion models, with a primary focus on the text-to-image models [20, 21, 71]. Broadly, these methods aim to achieve two main objectives: erasing data influence and preserving overall model performance. However, as demonstrated by Bui et al. [5], balancing this trade-off remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.502, + 0.483, + 0.864 + ], + "angle": 0, + "content": "In this work, we propose EraseDiff, an algorithm tailored to balance the overall performance of diffusion models with the erasure of undesirable information. Drawing inspiration from optimization-based meta-learning algorithms [18, 42] that enable fast adaptation to new learning tasks, we formulate this challenge as a bi-level optimization problem, where the \"inner\" optimization focus on erasing undesirable influence and the \"outer\" objective seeks to preserve model performance. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively. However, this nested optimization can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with non-convex functions, making it difficult to achieve a stable solution [39]. Therefore, we further reformulate the problem as a constrained optimization problem using the value function [40, 46, 70], which facilitates a natural first-order solution [39], allows us to optimize preservation and erasure in a unified manner. This approach achieves a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. We benchmark EraseDiff on various scenarios, encompassing unlearning of classes on CIFAR-10 [35] with Denoising Diffusion Probabilistic Models (DDPM) [30], classes on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Imagenette [31] and concepts on the I2P dataset [54] with stable diffusion. Our empirical findings show that EraseDiff is \\(11 \\times\\) faster than Heng and Soh's method [28] and \\(2 \\times\\) faster than Fan's method [16] when forgetting on DDPM while achieving better unlearning results across several metrics. The results demonstrate that EraseDiff is capable of effectively erasing data influence in diffusion models, ranging from specific classes to the concept of nudity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.228, + 0.642, + 0.244 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.905, + 0.331 + ], + "angle": 0, + "content": "In this section, we outline the components of the models we evaluate, including DDPM and latent diffusion models [49]. Throughout the paper, we denote scalars, and vectors/matrices by lowercase and bold symbols, respectively (e.g., \\(a\\), \\(a\\), \\(A\\))." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.35, + 0.905, + 0.629 + ], + "angle": 0, + "content": "DDPM. (1) Diffusion: DDPM gradually diffuses the data distribution \\(\\mathbb{R}^d\\ni \\mathbf{x}_0\\sim q(\\mathbf{x})\\) into the standard Gaussian distribution \\(\\mathbb{R}^d\\ni \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)\\) with \\(T\\) time steps, i.e., \\(q(\\mathbf{x}_t|\\mathbf{x}_{t - 1}) = \\mathcal{N}(\\mathbf{x}_t;\\sqrt{\\alpha_t}\\mathbf{x}_{t - 1},(1 - \\alpha_t)\\mathbf{I}_d)\\), where \\(\\alpha_{t} = 1 - \\beta_{t}\\) and \\(\\{\\beta_t\\}_{t = 1}^T\\) are the pre-defined variance schedule. The diffusion takes the form \\(\\mathbf{x}_t\\) as \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon\\), where \\(\\bar{\\alpha}_t = \\prod_{i = 1}^{t}\\alpha_{i}\\). (2) Training: A model \\(\\epsilon_{\\theta}(\\cdot)\\) with parameters \\(\\pmb {\\theta}\\in \\mathbb{R}^n\\) is trained to learn the reverse process \\(p_{\\pmb{\\theta}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)\\approx q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)\\). Given \\(\\mathbf{x}_0\\sim q(\\mathbf{x})\\) and time step \\(t\\in [1,T]\\), the simplified training objective is to minimize the distance between \\(\\pmb{\\epsilon}\\) and the predicted \\(\\epsilon_{t}\\) given \\(\\mathbf{x}_0\\) at time \\(t\\), i.e., \\(\\| \\epsilon -\\epsilon_{\\pmb{\\theta}}(\\mathbf{x}_t,t)\\|\\). (3) Sampling: after training the model, we could obtain the learnable backward distribution \\(p_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_{t}) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\pmb{\\mu}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t),\\pmb{\\Sigma}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t))\\), where \\(\\pmb{\\mu}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t) = \\frac{1}{\\sqrt{\\alpha_{t}}} (\\mathbf{x}_{t} - \\frac{\\beta_{t}}{\\sqrt{1 - \\alpha_{t}}}\\epsilon_{\\pmb{\\theta}}(\\mathbf{x}_{t},t))\\) and \\(\\pmb{\\Sigma}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t) = \\frac{(1 - \\bar{\\alpha}_{t - 1})\\beta_{t}}{1 - \\bar{\\alpha}_{t}}\\). Then, given \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)\\), \\(\\mathbf{x}_0\\) could be obtained via sampling from \\(p_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_{t})\\) from \\(t = T\\) to \\(t = 1\\) step by step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.906, + 0.756 + ], + "angle": 0, + "content": "Latent diffusion models. Latent diffusion models apply the diffusion models in the latent space \\(\\mathbf{z}\\) of a pre-trained variational autoencoder. The noise would be added to \\(\\mathbf{z} = \\varepsilon(\\mathbf{x})\\), instead of the data \\(\\mathbf{x}\\), and the denoised output would be transformed to image space with the decoder. Besides, text embeddings generated by models like CLIP are used as conditioning inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.715, + 0.787 + ], + "angle": 0, + "content": "3. Diffusion Unlearning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Let \\(\\mathcal{D} = \\{\\mathbf{x}_i, c_i\\}_i^N\\) be a dataset of images \\(\\mathbf{x}_i\\) associated with label \\(c_i\\) representing the class. \\(\\mathcal{C} = \\{1, \\dots, C\\}\\) denotes the label space where \\(C\\) is the total number of classes and \\(c_i \\in \\mathcal{C}\\). We split the training data \\(\\mathcal{D}\\) into the forgetting data \\(\\mathcal{D}_f \\subset \\mathcal{D}\\) and its complement, remaining data \\(\\mathcal{D}_r = \\mathcal{D} \\setminus \\mathcal{D}_f\\). The forgetting data has label space \\(\\mathcal{C}_f \\subseteq \\mathcal{C}\\), and the remaining label space is denoted as \\(\\mathcal{C}_r = \\mathcal{C} \\setminus \\mathcal{C}_f\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.431, + 0.901 + ], + "angle": 0, + "content": "1https://stablediffusion.gitbook.io/overview/stable-diffusion-overview/technology/training-procedures" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28264" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.269, + 0.108 + ], + "angle": 0, + "content": "3.1. Training objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.174 + ], + "angle": 0, + "content": "Our goal is to scrub the information about \\(\\mathcal{D}_f\\) carried by the diffusion models while maintaining the model utility over the remaining data \\(\\mathcal{D}_r\\). To achieve this, we adopt different training objectives for \\(\\mathcal{D}_r\\) and \\(\\mathcal{D}_f\\) as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.174, + 0.483, + 0.204 + ], + "angle": 0, + "content": "For the remaining data \\(\\mathcal{D}_r\\), we fine-tune the diffusion models with the original objective:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.21, + 0.483, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\in \\mathcal {N} \\left(\\mathbf {0}, \\mathbf {I} _ {d}\\right), \\left(\\mathbf {x} _ {0}, c\\right) \\sim \\mathcal {D} _ {r} \\times \\mathcal {C} _ {r}} [ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {t} | c\\right) \\| ^ {\\frac {2}{2}} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.25, + 0.485, + 0.297 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\epsilon}\\). For the forgetting data \\(\\mathcal{D}_f\\), we aim to let the models fail to generate meaningful images corresponding to \\(\\mathcal{C}_f\\) and thus propose:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.303, + 0.487, + 0.336 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}) = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\in \\mathcal {N} \\left(\\mathbf {0}, \\mathbf {I} _ {d}\\right), \\left(\\mathbf {x} _ {0}, c\\right) \\sim \\mathcal {D} _ {f} \\times \\mathcal {C} _ {f}} \\left[ \\| \\boldsymbol {\\epsilon} _ {f} - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {t} | c\\right) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.484, + 0.419 + ], + "angle": 0, + "content": "where \\(\\epsilon_{f} = \\epsilon_{\\theta}(\\mathbf{x}_{t}|c_{m})\\) and \\(c_{m}\\neq c\\) so that the denoised image \\(\\mathbf{x}_0\\) is not related to the forgetting class/concept \\(c\\) [16, 28]. With this, we hinder the approximator \\(\\epsilon_{\\theta}\\) to guide the denoising process to obtain meaningful examples for the forgetting data example \\(\\mathbf{x}_0\\sim \\mathcal{D}_f\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.483, + 0.449 + ], + "angle": 0, + "content": "To erase the undesirable influence of \\(\\mathcal{D}_f\\) and preserve the overall performance, it is common to form" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.457, + 0.483, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) + \\lambda \\mathcal {L} _ {f} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.602 + ], + "angle": 0, + "content": "with \\(\\lambda > 0\\) as the optimization objective (see for example [16]). However, training could be hindered due to the conflicting gradients between the erasing and preservation objectives, preventing a balanced trade-off between erasure and preservation [38]. To address this, rather than scalarizing the two objectives, we consider a framework based on optimization-based meta-learning algorithm [47] that allows iteratively updates to optimize each objective:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.609, + 0.338, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.629, + 0.483, + 0.653 + ], + "angle": 0, + "content": "s.t. \\(\\pmb {\\theta}\\in \\underset {\\phi}{\\arg \\min}\\mathcal{L}_f(\\pmb {\\phi};\\mathcal{D}_f)\\) (4)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.795 + ], + "angle": 0, + "content": "where the outer objective minimizes the remaining loss \\(\\mathcal{L}_r\\) (i.e., preserving model utility), the inner optimization minimizes the forgetting loss \\(\\mathcal{L}_f\\) (i.e., erasing) with initialization \\(\\mathbb{R}^n\\ni \\phi_{\\mathrm{init}} = \\pmb{\\theta}\\). Given \\(\\pmb{\\theta}\\), the inner optimization on \\(\\phi\\) aims to minimize the forgetting data influence, with the goal of achieving effective erasure while preserving model utility. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While the above framework allows for iterative updates to address the conflicting objectives of erasure and preservation, it still relies on nested optimization, which can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with nonconvex functions, making it difficult to achieve a stable solution [39]. To further streamline the optimization process," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "we adopt a value function approach [40, 46, 70] that reformulates the problem as a single-constrained optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.144, + 0.724, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.164, + 0.905, + 0.18 + ], + "angle": 0, + "content": "s.t. \\(\\mathcal{L}_f(\\pmb {\\theta};\\mathcal{D}_f) - \\min_{\\phi}\\mathcal{L}_f(\\phi ;\\mathcal{D}_f)\\leq 0,\\) (5)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.907, + 0.291 + ], + "angle": 0, + "content": "where \\(\\phi\\) is initialized at \\(\\theta\\), leverages the value function to encapsulate the influence of data erasure directly as a constraint on \\(\\mathcal{L}_f\\). This avoids the need for a nested loop by capturing the forgetting objective as a constraint and provides a natural first-order solution, as the constrained formulation allows us to optimize preservation and erasure in a unified manner." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.3, + 0.615, + 0.314 + ], + "angle": 0, + "content": "3.2. Solution" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.321, + 0.907, + 0.473 + ], + "angle": 0, + "content": "To solve Eq. (5), let us first denote \\( g(\\pmb{\\theta}) \\coloneqq \\mathcal{L}_f(\\pmb{\\theta};\\mathcal{D}_f) - \\min_{\\phi}\\mathcal{L}_f(\\phi ;\\mathcal{D}_f) \\). Our goal is to erase undesirable influence while preserving the overall model performance, hence the update vector \\( \\delta_t \\) for updating the model should aid in minimizing \\( \\mathcal{L}_r(\\pmb{\\theta};\\mathcal{D}_r) \\) and \\( g(\\pmb{\\theta}) \\) simultaneously. In other words, suppose that the current solution for Eq. (5) is \\( \\pmb{\\theta}_t \\), we aim to update \\( \\pmb{\\theta}_{t + 1} = \\pmb{\\theta}_t - \\eta \\pmb{\\delta}_t \\) where \\( \\eta \\) is sufficiently small, so that \\( \\mathcal{L}_r(\\pmb{\\theta}_{t + 1};\\mathcal{D}_r) \\) decreases (i.e., preserve model utility) and \\( g(\\pmb{\\theta}_{t + 1}) \\) decreases (i.e., erasure). To this end, we aim to find the update vector \\( \\delta_t \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.479, + 0.862, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\delta} _ {t} \\in \\frac {1}{2} \\operatorname {a r g m i n} _ {\\boldsymbol {\\delta}} \\left\\| \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} \\left(\\boldsymbol {\\theta} _ {t}; \\mathcal {D} _ {r}\\right) - \\boldsymbol {\\delta} \\right\\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.51, + 0.905, + 0.527 + ], + "angle": 0, + "content": "s.t. \\(\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_t)^\\top \\pmb{\\delta} \\geq a_t > 0.\\) (6)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.533, + 0.906, + 0.642 + ], + "angle": 0, + "content": "This will ensure that the update \\(\\delta_t\\) is close to \\(\\nabla_\\theta \\mathcal{L}_r(\\pmb{\\theta}_t; \\mathcal{D}_r)\\) and decreases \\(g(\\pmb{\\theta}_t)\\) until it reaches stationary. Because \\(g(\\pmb{\\theta}_{t+1}) - g(\\pmb{\\theta}_t) \\approx -\\eta \\nabla_\\theta g(\\pmb{\\theta}_t)^\\top \\delta \\leq -\\eta a_t < 0\\) for some scalar \\(a_t > 0\\), we can ensure that \\(g(\\pmb{\\theta}_{t+1}) < g(\\pmb{\\theta}_t)\\) for small step size \\(\\eta > 0\\). This means that the update \\(\\delta_t\\) can ensure to minimize \\(\\mathcal{L}_f(\\pmb{\\theta}; \\mathcal{D}_f)\\) as long as it does not conflict with descent of \\(\\mathcal{L}_r(\\pmb{\\theta}; \\mathcal{D}_r)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.642, + 0.906, + 0.672 + ], + "angle": 0, + "content": "To find the solution to the optimization problem in Eq. (6), the following theorem is developed:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.678, + 0.907, + 0.733 + ], + "angle": 0, + "content": "Theorem 3.1. The optimal solution of the optimization problem in Eq. (6) is \\(\\delta^{*} = \\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r}) + \\lambda_{t}\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})\\) where \\(\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We provide the proof in §7 in the Appendix. This provides the solution to the optimization problem by constructing the update vector \\(\\delta\\) to balance two competing objectives. The variable \\(a_{t}\\) adjusts the weight of the forgetting objective, ensuring that the update vector \\(\\delta\\) decreases the remaining loss without violating the erasure goal, and achieves the dual goals of maintaining utility and achieving erasing. In practice, we can choose \\(a_{t} = \\eta \\| \\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}\\), and we start from \\(\\phi^0 = \\pmb{\\theta}_t\\) and use gradient descend in \\(K\\) steps with the learning rate \\(\\xi > 0\\) to reach \\(\\phi^K\\), namely \\(\\phi^{k + 1} = \\phi^k - \\xi \\nabla_\\phi \\mathcal{L}_f(\\phi^k; \\mathcal{D}_f)\\) and \\(k = 0, \\dots, K - 1\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28265" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.091, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Algorithm 1 EraseDiff: Erasing undesirable influence in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.123, + 0.483, + 0.169 + ], + "angle": 0, + "content": "Input: Well-trained model with parameters \\(\\theta_0\\), forgetting data \\(\\mathcal{D}_f\\) and remaining data \\(\\mathcal{D}_r\\), outer iteration number \\(T\\) and inner iteration number \\(K\\), learning rate \\(\\eta\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.169, + 0.413, + 0.183 + ], + "angle": 0, + "content": "Output: Parameters \\(\\theta^{*}\\) for the scrubbed model." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.185, + 0.267, + 0.197 + ], + "angle": 0, + "content": "1: for iteration \\(t\\) in \\(T\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.199, + 0.2, + 0.213 + ], + "angle": 0, + "content": "2: \\(\\phi^0 = \\pmb {\\theta}_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.214, + 0.482, + 0.244 + ], + "angle": 0, + "content": "3: Get \\(\\phi^K\\) by \\(K\\) steps of gradient descent on \\(\\mathcal{L}_f(\\phi; \\mathcal{D}_f)\\) starting from \\(\\phi^0\\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.245, + 0.406, + 0.259 + ], + "angle": 0, + "content": "4: Set \\(g(\\pmb{\\theta}_t) = \\mathcal{L}_f(\\pmb{\\theta}_t; \\mathcal{D}_f) - \\mathcal{L}_f(\\phi^K; \\mathcal{D}_f)\\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.26, + 0.482, + 0.29 + ], + "angle": 0, + "content": "5: Update the model: \\(\\pmb{\\theta}_{t + 1} = \\pmb{\\theta}_t - \\eta (\\nabla_{\\pmb{\\theta}_t}\\mathcal{L}_r(\\pmb {\\theta}_t;\\mathcal{D}_r) + \\lambda_t\\nabla_{\\pmb{\\theta}_t}g(\\pmb {\\theta}_t;\\phi^K))\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.29, + 0.446, + 0.313 + ], + "angle": 0, + "content": "6: where \\(\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})^{T}\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.313, + 0.174, + 0.323 + ], + "angle": 0, + "content": "7: end for" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.185, + 0.482, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.354, + 0.193, + 0.37 + ], + "angle": 0, + "content": "3.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.377, + 0.483, + 0.408 + ], + "angle": 0, + "content": "We can characterize the solution of our algorithm as follows and the proof can be found in §7 in the Appendix:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.466 + ], + "angle": 0, + "content": "Theorem 3.2 (Pareto optimality). The stationary point obtained by our algorithm is Pareto optimal of the problem \\(\\min_{\\pmb{\\theta}}[\\mathcal{L}_r(\\pmb{\\theta};\\mathcal{D}_r),\\mathcal{L}_f(\\pmb{\\theta};\\mathcal{D}_f)]\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.484, + 0.552 + ], + "angle": 0, + "content": "This asserts that the solution obtained by our algorithm is Pareto optimal for the problem of minimizing both objectives, which implies that the solution obtained by the algorithm ensures a balanced trade-off between preserving model utility and erasing undesirable influences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We further take DDPM with CIFAR-10 when forgetting the 'airplane' as an example to show that our proposed method helps alleviate the gradient conflict which prevents a balanced trade-off between erasure and preservation. Fig. 2 presents the cosine similarity between the update vector \\(\\delta\\) and the gradient \\(g_{r} = \\nabla_{\\theta}\\mathcal{L}_{r}(\\theta ;\\mathcal{D}_{r})\\) for preservation, and the cosine similarity between the update vector \\(\\delta\\) and the gradient \\(g_{f} = \\nabla \\mathcal{L}_{f}(\\theta ;\\mathcal{D}_{f})\\) for erasing. The cosine similarity represents the alignment between the update vector and the gradients associated with the preservation and erasure. Higher positive values indicate alignment, meaning that the update vector \\(\\delta\\) is directed similarly to the respective gradient, whereas negative values indicate misalignment or conflict. In particular, negative values suggest a high degree of opposition between the update vector and the respective gradient, which can signify competing objectives between preservation and erasure during the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "MOO (Multi-Objective Optimization) denotes the naive integration of erasing and preservation as stated in Eq. (3). For the vanilla MOO, Fig. 2 shows a clear alternating pattern in cosine similarity values between the update vector \\(\\delta\\) and the gradients \\(g_{r}\\) and \\(g_{f}\\). Specifically, when the cosine" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.097, + 0.897, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.292, + 0.897, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.502, + 0.906, + 0.586 + ], + "angle": 0, + "content": "Figure 2. Top to bottom: cosine similarity between the update vector \\(\\delta\\) and the preservation gradient \\(g_{r}\\), followed by the cosine similarity between \\(\\delta\\) and the erasing gradient \\(g_{f}\\). Positive values indicate alignment, while negative values suggest conflict. This visualization illustrates how well the update vector aligns with the objectives of preservation and erasure over successive iterations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.827 + ], + "angle": 0, + "content": "similarity between \\(\\delta\\) and \\(g_{r}\\) is greater than 0, the similarity between \\(\\delta\\) and \\(g_{f}\\) tends to be less than 0, and vice versa. This pattern suggests that MOO experiences gradient conflict, as it cannot effectively balance the two objectives of preservation and erasure, preventing MOO from achieving a harmonious update that supports both goals simultaneously. In contrast, EraseDiff mostly shows positive cosine similarity values between the update vector \\(\\delta\\) and both the preservation gradient \\(g_{r}\\) and the erasing gradient \\(g_{f}\\). This indicates that EraseDiff aligns the update direction with both objectives, suggesting it manages to avoid significant gradient conflict. By maintaining positive alignment, EraseDiff appears to balance preservation and erasure more effectively, leading to better cooperation between objectives." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.844, + 0.655, + 0.86 + ], + "angle": 0, + "content": "4. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Memorization in generative models. Privacy of generative models has been studied extensively for GANs [17, 43," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28266" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "66] and generative language models [8, 9, 32, 63]. These generative models often risk replicating from their training data. Recently, several studies [10, 58, 59, 65] investigated these data replication behaviors in diffusion models, raising concerns about the privacy and copyright issues. Possible mitigation strategies are deduplicating and randomizing conditional information [58, 59], or training models with differential privacy (DP) [1, 13-15]. However, leveraging DP-SGD [1] may cause training to diverge [10]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.23, + 0.485, + 0.534 + ], + "angle": 0, + "content": "Malicious misuse. Diffusion models usually use training data from varied open sources and when such unfiltered data is employed, there is a risk of it being tainted [12] or manipulated [48], resulting in inappropriate generation [54]. They also risk the imitation of copyrighted content, e.g., mimicking the artistic style [20, 57]. To counter inappropriate generation, data censoring [3, 19, 45, 53] where excluding black-listed images before training, and safety guidance where diffusion models will be updated away from the inappropriate/undesired concept [20, 54] are proposed. [34] uses human feedback to annotate the generations for getting soft tokens that quantify the harmfulness of the images, then fine-tunes the model using self-distillation to achieve harmony between preservation and erasure. Shan et al. [57] propose protecting artistic style by adding barely perceptible perturbations to the artworks before public release. Yet, Rando et al. [48] argue that DMs can still generate content that bypasses the filter. Chen et al. [12] highlight the susceptibility of DMs to poison attacks, where target images are generated with specific triggers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Machine unlearning. Removing data directly involves retraining the model from scratch, which is inefficient and impractical. Thus, to reduce the computational overhead, efficient machines unlearning methods [4, 7, 11, 22, 23, 25, 33, 44, 50, 56, 61, 68] have been proposed. Several studies [5, 16, 20, 21, 27, 28, 67, 71] recently introduce unlearning in diffusion models. Most of them [20, 21, 27, 71] mainly focus on text-to-image models and high-level visual concept erasure. Heng and Soh [28] adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) from continual learning to perform unlearning effectively without access to the training data. Heng and Soh's method can be applied to a wide range of generative models, however, it needs the computation of FIM for different datasets and models, which may lead to significant computational demands. Fan et al. [16] propose a very potent unlearning algorithm called SalUn that shifts attention to important parameters w.r.t. the forgetting data. SalUn can perform effectively across image classification and generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we introduce a simple yet effective unlearning algorithm for diffusion models by formulating the problem as a constrained optimization problem, to achieve a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. Below, we will show" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "that our algorithm is not only faster than Heng and Soh's method [28] and Fan's method [16], but even outperforms these methods in terms of the trade-off between the forgetting and preserving model utility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.165, + 0.64, + 0.183 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.907, + 0.326 + ], + "angle": 0, + "content": "We evaluate EraseDiff in various scenarios, including removing images with specific classes/concepts, to answer the following research questions (RQs): (i) Can typical machine unlearning methods be applied to diffusion models? (ii) Is EraseDiff able to remove the influence of \\(\\mathcal{D}_f\\) in the diffusion models? (iii) Is EraseDiff able to preserve the model utility while removing \\(\\mathcal{D}_f\\)? (iv) Is EraseDiff efficient in removing the data? (v) How does EraseDiff perform on the public well-trained models?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.335, + 0.597, + 0.351 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Experiments are reported on CIFAR-10 [35] with DDPM, Imagenette [31] with Stable Diffusion (SD) for class-wise forgetting, I2P [54] dataset with SD for concept-wise forgetting. For all SD experiments, we use the open-source SD v1.4 [49] checkpoint as the pre-trained model. Implementation details and additional results like visualizations of generated images can be found in §8 and §9." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.553 + ], + "angle": 0, + "content": "Baselines. We primarily benchmark against the following baselines commonly used in machine unlearning: (i) Unscrubbed, (ii) Finetune (FT) [23], (iii) NegGrad (NG) [23], (iv) BlindSpot [61], (v) ESD [20], (vi) FMN [71], (vii) Selective Amnesia (SA) [28] and (viii) the SOTA machine unlearning algorithm SalUn [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.826 + ], + "angle": 0, + "content": "Metrics. Several metrics are utilized to evaluate the algorithms: (i) Frechet Inception Distance (FID) [29]: the widely-used metric for assessing the quality of generated images. (ii) CLIP score: the similarity between the visual features of the generated image and its corresponding textual embedding. (iii) \\( P_{\\psi}(\\mathbf{y} = c_f|\\mathbf{x}_f) \\) [28]: the classification rate of a pre-trained classifier \\( P_{\\psi}(\\mathbf{y}|\\mathbf{x}) \\), with a ResNet architecture [26] used to classify generated images conditioned on the forgetting classes. A lower classification value indicates superior unlearning performance. (iv) Precision and Recall: A low FID may indicate high precision (realistic images) but low recall (small variations) [37, 51]. Kynkänniemi et al. [37] shows that generative models claim to optimize FID (high fidelity) but always sacrifice variation (low diversity). Hence, we include metric precision (fidelity) and recall (diversity) to express the quality of the generated samples, to provide explicit visibility of the tradeoff between sample quality and variety." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.691, + 0.849 + ], + "angle": 0, + "content": "5.2. Results on DDPM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Following SA, we aim to forget the 'airplane' class on CIFAR-10. Here, we replace \\(\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)\\) with \\(\\epsilon_{f} = \\epsilon_{\\theta}(\\mathbf{x}_{t}|c_{m})\\) like random labelling used in [16] where \\(c_{m}\\neq c\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28267" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.16 + ], + "angle": 0, + "content": "Table 1. Results on CIFAR10 with DDPM when forgetting the 'airplane' class. \\( P_{\\psi}(\\mathbf{y} = c_f|\\mathbf{x}_f) \\) indicate the probability of the forgotten class (i.e., the effectiveness of erasing). Precision and Recall demonstrate the fidelity and diversity [37, 51], and FID scores are computed between the generated 45K images and the corresponding ground truth images with the same labels from \\( \\mathcal{D}_r \\) (i.e., preserving model utility). SA excels in class-wise forgetting but struggles to perform concept-wise forgetting as shown in Fig. 3 and Tab. 2. The best and the second best are highlighted in blue and orange, respectively." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.17, + 0.895, + 0.272 + ], + "angle": 0, + "content": "
UnscrubbedFT [23]NG [23]BlindSpot [61]SA [28]SalUn [16]EraseDiffrlEraseDiffnoise
FID ↓9.638.2176.739.128.199.168.667.61
Precision (fidelity) ↑0.400.430.080.410.430.410.430.43
Recall (diversity) ↑0.790.770.610.780.750.760.770.72
Pψ(y = cf|xf)↓0.970.960.610.900.060.070.240.22
" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.281, + 0.382, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.281, + 0.553, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.282, + 0.725, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.282, + 0.898, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.448, + 0.908, + 0.478 + ], + "angle": 0, + "content": "Figure 3. Quantity of nudity content detected using the NudeNet classifier from I2P data. Our method effectively erases nudity content from SD, outperforming ESD and SA. Note that Fig. 4 and Tab. 2 together presents the trade-off between erasing and preservation." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Table 2. Evaluation of \\(30\\mathrm{K}\\) generated images by SD when erasing 'nudity'. The FID score is measured compared to validation data, while the CLIP similarity score evaluates the alignment between generated images and the corresponding prompts. The best and the second best are highlighted in blue and orange, respectively." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.581, + 0.466, + 0.648 + ], + "angle": 0, + "content": "
ESD [20]SA [28]SalUn [16]EraseDiff
FID ↓15.7625.5825.0617.01
CLIP ↑30.3331.0328.9130.58
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.749 + ], + "angle": 0, + "content": "denoted as \\( \\text{EraseDiff}_{\\text{rl}} \\). We also try to use \\( \\epsilon_f = \\mathcal{U}(\\mathbf{0}, \\mathbf{I}_d) \\) like SA, denoted as \\( \\text{EraseDiff}_{\\text{noise}} \\). Note that the choice of replacement for forgotten classes is flexible and is not the primary focus of this work. For further discussion on the choice of substitution strategies, please refer to [6, 41]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Results are presented in Tab. 1. Firstly, from Tab. 1, we can conclude that traditional machine unlearning methods designed for image classification or regression tasks fall short in effectively performing forgetting for DDPM. Finetune and BlindSpot suffer from under-forgetting (i.e., the generated image quality is good but the probability of generated images belonging to the forgetting class approaching the value of the unscrubbed model), and NegGrad suffers from over-forgetting (the probability of generated images belonging to the forgetting class is decreased compared to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.504, + 0.905, + 0.534 + ], + "angle": 0, + "content": "that of the unscrubbed model but the generated image quality drops significantly)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Then, comparing SA and SalUn's unlearning methods, SA achieves an FID score of 8.19 but sacrifices variation (decreased recall). Also, note that SA introduces excessive computational resource requirements and time consumption [28, 72]. Note that the FID scores of SA, SalUn, and EraseDiff decrease compared with the generated images from the original models; the quality of the generated images experiences a slight improvement. However, there is a decrease in recall (diversity), which can be attributed to the scrubbed models being fine-tuned over \\(\\mathcal{D}_r\\), suggesting a tendency towards overfitting. Regarding forgetting, SalUn achieves a smaller probability of the generated images classified as the forgetting class than ours; yet, the FID score is larger than ours, and images generated by EraseDiffr1 present better diversity and fidelity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.757, + 0.787 + ], + "angle": 0, + "content": "5.3. Results on Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.87 + ], + "angle": 0, + "content": "We apply EraseDiff to perform class-wise forgetting from Imagenette and erase the 'nudity' concept with SD v1.4. When forgetting 'nudity', we have no access to the training data; instead, we generate \\(\\sim 400\\) images with the prompts \\(c_{f} = \\{\\text{'nudity', 'naked', 'erotic', 'sexual'}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Forget nudity. 4703 images are generated using I2P prompts, and 1K images are generated using the prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28268" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.089, + 0.155, + 0.118 + ], + "angle": 0, + "content": "Table 3. P forgetting" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.046, + 0.987, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.9 + ], + "angle": 0, + "content": "{nudity, nak, sexual} The quality of nu-dity content is detected using the NudeNet classifier [2]. In Fig. 3, the number in the y-axis denotes the number of exposed body parts generated by the SD v1.4 model. Fig. 3 presents the percentage change in exposed body parts w.r.t. SD v1.4. In §9, we provide the number of exposed body parts counted in all generated images with different thresholds. Here, our algorithm replaces \\(\\epsilon_{f}\\) with \\(\\epsilon_{\\theta}(\\mathbf{x}_i|c_m)\\) where \\(c_{m}\\) is 'a photo of poker'. We can find that, EraseD-iff reduces the amount of nu-dity content compared to SD v1.4, ESD, and SA, particularly on sensitive content like Female/Male interests and female/Male Genitalia. While SalUn excels a 'forgetting', our algorithm demonstrates a significant improvement in the quality of generated images as shown in Tab. 2. Tab. 2 presented results evaluating the utility of scrubbed models. The FID and CLIP scores are" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.916, + 0.984, + 1.0 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.12 + ], + "angle": 0, + "content": "Table 4. Computational overhead. Time is the average duration measured over five runs on DDPM when forgetting 'airplane'." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.129, + 0.472, + 0.207 + ], + "angle": 0, + "content": "
Memory (MiB)Time (min.)Complexity
SA3352.3140.00O(n2)
SalUn4336.228.17O(n)
EraseDiff3360.312.70O(n)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.232, + 0.485, + 0.369 + ], + "angle": 0, + "content": "approach, clearly indicate the advantages of our method. We also present results when improving the forgetting ability of SalUn in §9. However, note that this enhancement comes with a drop in the FID score of the generated images. Our method, while slightly better than SalUn on average across 10 classes, demonstrates a more balanced trade-off between erasing and preservation, indicating that it achieves a favorable balance in preserving fidelity while enhancing erasing performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.379, + 0.32, + 0.396 + ], + "angle": 0, + "content": "5.4. Computational efficiency" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.485, + 0.598 + ], + "angle": 0, + "content": "Finally, we measure the computational complexity of unlearning algorithms. The computational complexity of SA and SalUn involves two distinct stages: the computation of FIM for SA and the computation of salient weights w.r.t. \\(\\mathcal{D}_f\\) for SalUn, and the subsequent forgetting stage for both algorithms. We consider the maximum memory usage across both stages, the metric 'Time' is exclusively associated with the duration of the forgetting stage for unlearning algorithms. Tab. 4 show that EraseDiff outperforms SA and SalUn in terms of efficiency, achieving a speed increase of \\(\\sim 11\\times\\) than SA and \\(\\sim 2\\times\\) than SalUn. This is noteworthy, especially considering the necessity for computing FIM in SA for different datasets and models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.609, + 0.242, + 0.625 + ], + "angle": 0, + "content": "5.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.631, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We further investigate the influence of the number of iterations \\(K\\) that approximate \\(\\min \\mathcal{L}_f(\\phi ;\\mathcal{D}_f)\\), and the step size \\(\\eta\\) that controls the weight of forgetting and preserving model utility. Here, we replace \\(\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)\\) with \\(\\epsilon_{f}\\in \\mathcal{U}(\\mathbf{0},\\mathbf{I}_{d})\\). Note that for different hyperparameters in Fig. 5 (a), the average entropy of the classifier's output distribution given \\(\\mathbf{x}_f\\), which is \\(H(P_{\\psi}(\\mathbf{y}|\\mathbf{x}_f)) = -\\mathbb{E}[\\sum_iP_{\\psi}(\\mathbf{y} = c_i|\\mathbf{x})\\log_eP_{\\psi}(\\mathbf{y} = c_i|\\mathbf{x})]\\), remains close to 2.02. This indicates that the scrubbed models become uncertain about the images conditioned on the forgetting class, effectively erasing the information about \\(\\mathcal{D}_f\\). Below, we will further demonstrate the influence on the model utility. In practice, we have \\(\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\theta}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\theta}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\} = \\max \\{0,\\eta -\\frac{\\nabla_{\\theta}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\theta}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}\\), we can see that \\(\\eta\\) determines the extent to which the update direction for forgetting can deviate from that for preserving model utility. A larger \\(\\eta\\) would allow for more deviation in the updating, thus pri" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.91, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.275, + 0.9, + 0.29 + ], + "angle": 0, + "content": "Figure 5. (a) Ablation results. (b) Potential incomplete erasures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.468 + ], + "angle": 0, + "content": "oritizing forgetting over preserving model utility. In Fig. 5 (a), the FID score tends to increase (i.e., image quality drop) as the step size \\(\\eta\\) increases, indicating that larger \\(\\eta\\) leads to greater deviations from the direction that preserves the model utility. Furthermore, the number of iterations \\(K\\) determines how closely the approximation \\(\\phi^K\\) will approach \\(\\arg \\min_{\\phi} \\mathcal{L}_f(\\phi; \\mathcal{D}_f)\\). Hence, a larger number of iterations \\(K\\) leads to more thorough erasure, which is also supported by the results shown in Fig. 5 (a), as increasing \\(K\\) correlates with an increase in the FID score." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.907, + 0.498 + ], + "angle": 0, + "content": "6. Conclusion, limitations and broader impacts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.611 + ], + "angle": 0, + "content": "In this work, we explored erasing undesirable influence in diffusion models and proposed an efficient method EraseDiff to achieve a balanced trade-off between erasing and preservation. Comprehensive experiments on diffusion models demonstrate the proposed algorithm's effectiveness in data removal, its efficacy in preserving the model utility, and its efficiency in erasure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.899 + ], + "angle": 0, + "content": "However, our scrubbed model may have incomplete erasures, potentially arises due to the choice of the replacements. In Fig. 5 (b), with random labeling, generated images conditioned on the forgetting class 'tench' by our scrubbed model may preserve some characteristics similar to that close to 'tench' visually; instead, replacing the erased class 'tench' with other class such as 'dog' could improve in erasure. Besides, the scrubbed models could be biased or be misused for censorship or exploitation. This includes using technology to selectively remove or amplify NSFW content in various ways. They might also falsely assure users that NSFW content is fully removed when it isn't. To address this, advanced privacy-preserving training techniques are in demand to enhance the security and fairness of the models. Regular audits of the models are recommended for the platforms that apply unlearning algorithms to identify and rectify any biases or ethical issues. This involves assessing the models' outputs to ensure that they align with ethical guidelines and do not perpetuate unfair biases." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28270" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.186 + ], + "angle": 0, + "content": "Mehrtash Harandi is supported by the Australian Research Council (ARC) Discovery Program DP250100262. The authors gratefully acknowledge the anonymous reviewers for their insightful feedback and valuable suggestions, which have significantly improved the quality of this work." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308-318, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.309, + 0.483, + 0.336 + ], + "angle": 0, + "content": "[2] P Bedapudi. Nudenet: Neural nets for nudity classification, detection and selective censoring, 2019. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.338, + 0.483, + 0.393 + ], + "angle": 0, + "content": "[3] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1536-1546. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.395, + 0.482, + 0.463 + ], + "angle": 0, + "content": "[4] Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.465, + 0.482, + 0.52 + ], + "angle": 0, + "content": "[5] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[6] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[8] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633-2650, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[9] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[10] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[12] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Tim Dockhorn, Tianshi Cao, Arash Vahdat, and Karsten Kreis. Differentially private diffusion models. arXiv preprint arXiv:2210.09929, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[14] Cynthia Dwork. Differential privacy: A survey of results. In International conference on theory and applications of models of computation, pages 1-19. Springer, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[15] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, New York, NY, USA, March 4-7, 2006. Proceedings 3, pages 265-284. Springer, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[16] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Dennis Wei, Eric Wong, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[17] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701–6710, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.496 + ], + "angle": 0, + "content": "[18] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[20] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In 2023 IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[21] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[22] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[23] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[24] Eric Goldman. An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28271" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "models. In Proceedings of the 37th International Conference on Machine Learning, pages 3832-3842. PMLR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[27] Alvin Heng and Harold Soh. Continual learning for forgetting in deep generative models. 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[28] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.334, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.378, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[31] Jeremy Howard and Sylvain Gugger. Fastai: A layered api for deep learning. Information, 11(2):108, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[32] Matthew Jagielski, Om Thakkar, Florian Tramer, Daphne Ippolito, Katherine Lee, Nicholas Carlini, Eric Wallace, Shuang Song, Abhradeep Thakurta, Nicolas Papernot, et al. Measuring forgetting of memorized training examples. arXiv preprint arXiv:2207.00099, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.477, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[33] Masayuki Karasuyama and Ichiro Takeuchi. Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks, 21(7):1048-1059, 2010. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[34] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Safeguard text-to-image diffusion models with human feedback inversion. In European Conference on Computer Vision, pages 128-145. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.604, + 0.482, + 0.632 + ], + "angle": 0, + "content": "[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[37] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[38] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems (NeurIPS), 34:18878-18890, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[39] Bo Liu, Mao Ye, Stephen Wright, Peter Stone, and Qiang Liu. Bome! bilevel optimization made easy: A simple first-order approach. Advances in Neural Information Processing Systems, 35:17248-17262, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Risheng Liu, Xuan Liu, Xiaoming Yuan, Shangzhi Zeng, and Jin Zhang. A value-function-based interior-point method" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "for non-convex bi-level optimization. In International conference on machine learning, pages 6882-6892. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[41] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[42] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In International conference on machine learning, pages 2113-2122. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[43] Casey Meehan, Kamalika Chaudhuri, and Sanjoy Das-gupta. A non-parametric test to detect data-copying in generative models. In International Conference on Artificial Intelligence and Statistics, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[44] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N. Ravi. Deep unlearning via randomized conditionally independent hessenians. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10412-10421, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[46] Jiri V Outrata. On the numerical solution of a class of stackelberg problems. Zeitschrift für Operations Research, 34: 255-277, 1990. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[47] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.904, + 0.617 + ], + "angle": 0, + "content": "[48] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[50] Enrique Romero, Ignacio Barrio, and Lluis Belanche. Incremental and decremental learning for linear support vector machines. In International Conference on Artificial Neural Networks, pages 209-218. Springer, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[51] Mehdi SM Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. Advances in neural information processing systems, 31, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[52] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content?" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28272" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350-1361, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[54] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[56] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS), 34: 18075-18086, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[57] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[58] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[59] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[61] Ayush Kumar Tarun, Vikram Singh Chandawat, Murari Mandal, and Mohan Kankanhalli. Deep regression unlearning. In International Conference on Machine Learning, pages 33921-33939. PMLR, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[62] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[63] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 35: 38274-38290, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[64] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[65] Nikhil Vyas, Sham Kakade, and Boaz Barak. Provable copyright protection for generative models. arXiv preprint arXiv:2302.10870, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[66] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. This person (probably) exists. identity member-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ship attacks against gan generated faces. arXiv preprint arXiv:2107.06018, 2021.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[67] Jing Wu and Mehrtash Harandi. Scissorhands: Scrub data influence via connection sensitivity in networks. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[68] Yinjun Wu, Edgar Dobriban, and Susan Davidson. DeltaGrad: Rapid retraining of machine learning models. In Proceedings of the 37th International Conference on Machine Learning, pages 10355-10366. PMLR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[69] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 87-103. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[70] Jane J Ye and DL Zhu. Optimality conditions for bilevel programming problems. Optimization, 33(1):9-27, 1995. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[71] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[72] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024.6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28273" + } + ] +] \ No newline at end of file diff --git a/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_origin.pdf b/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9ebd5027809dfc6f4606a631a6a4c358881925b1 --- /dev/null +++ b/2025/Erasing Undesirable Influence in Diffusion Models/cfcabb09-71fe-461a-a521-c3d740140c91_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e956f75ae99129b83784db6a6a63539d6ba5adf520aeccf3f39ac6cb54d94f6 +size 6331280 diff --git a/2025/Erasing Undesirable Influence in Diffusion Models/full.md b/2025/Erasing Undesirable Influence in Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..24dc59b2ab4461dfe10f8a257000bc3b3f42843c --- /dev/null +++ b/2025/Erasing Undesirable Influence in Diffusion Models/full.md @@ -0,0 +1,338 @@ +# Erasing Undesirable Influence in Diffusion Models + +Jing Wu $^{1}$ , Trung Le $^{1}$ , Munawar Hayat $^{2}$ , Mehrtash Harandi $^{1}$ $^{1}$ Monash University, Melbourne, VIC, Australia, $^{2}$ Qualcomm, San Deigo, CA, US +{jing.wu1, trunglm, mehrtash.harandi}@monash.edu, hayat@qti.Qualcomm.com + +![](images/17ae78a31740d4b9de3e311438bca2c72819f358973e92ee65de28f126d1caf4.jpg) + +![](images/112ca5af0c57a515ab46f0c02b96e87d6ccb29b94fdd4b546684b020963e5578.jpg) +Figure 1. Top to Bottom: generated samples by SD v1.4 and model scrubbed by our method, when erasing the concept of 'nudity'. Our method can avoid NSFW (not safe for work) content while preserving model utility. + +![](images/46651389b1cebd84f1d8b5c41e8306608112e2649f0608ef1ffb4cd97acadb8e.jpg) +I2P prompts + +![](images/775af8103bcd653a1c4507557af163384bb954287a16316b8aab955358cea7d4.jpg) + +![](images/5d079d5f12be466c8d22260e31dbe7cd289d4bac356b0bd7cf634f60f48e4119.jpg) + +![](images/94b147510b9c83c4d921141cc1c58f5f87b9de82e0ee20960bc34fb8fe6eab1e.jpg) + +![](images/22f0da26f62de723d6de6c75a1720fd07c6dec1d63395f6eeb0ed676cfb20db1.jpg) + +![](images/12136886ef4f30ac41cb0bff03c0d05e5bf4aa5a115ca05d76dbc8f7329ed1b1.jpg) +COCO prompts + +![](images/444ae9d1d59ca02f32f037433f72ad07f919a10e9e0734206a955068ae06af6b.jpg) + +![](images/3c847e1f63febefb6ff5cfd5330894d12ceb8294736b4c33a09340279ed2c3f1.jpg) + +![](images/718c3957d05a8fe75a0f23ddf41b912c9155eee4d46ea66b0699d27c88ebebab.jpg) +Added by authors for publication + +# Abstract + +Diffusion models are highly effective at generating high-quality images but pose risks, such as the unintentional generation of NSFW (not safe for work) content. Although various techniques have been proposed to mitigate unwanted influences in diffusion models while preserving overall performance, achieving a balance between these goals remains challenging. In this work, we introduce EraseDiff, an algorithm designed to preserve the utility of the diffusion model on retained data while removing the unwanted information associated with the data to be forgotten. Our approach formulates this task as a constrained optimization problem using the value function, resulting in a natural first-order algorithm for solving the optimization problem. By altering the generative process to deviate away from the groundtruth denoising trajectory, we update parameters for preservation while controlling constraint reduction to ensure effective erasure, striking an optimal trade-off. Extensive experiments and thorough comparisons with state-of-the-art algorithms demonstrate that EraseDiff effectively preserves the model's utility, efficacy, and efficiency. + +WARNING: This paper contains sexually explicit imagery that may be offensive in nature. + +# 1. Introduction + +Diffusion Models [30, 49, 60] are now the method of choice in deep generative models, owing to their high-quality output, stability, and ease of training procedure. This has facilitated their successful integration into commercial applications such as midjourney. Unfortunately, the ease of use associated with diffusion models brings forth significant privacy risks. Studies have shown that these models can memorize and regenerate individual images from their training datasets [10, 58, 59]. Beyond privacy, diffusion models are susceptible to misuse and can generate NSFW digital content [48, 52, 54]. In this context, individuals whose images are used for training might request the removal of their private data. In particular, data protection regulations like the European Union General Data Protection Regulation (GDPR) [64] and the California Consumer Privacy Act (CCPA) [24] grant users the right to be + +forgotten, obligating companies to expunge data pertaining to a user upon receiving a request for deletion. These legal provisions grant data owners the right to remove their data from trained models and eliminate its influence on said models [4, 11, 23, 25, 44, 56, 61, 62, 69]. + +A straightforward solution is to retrain the model from scratch after excluding the data that needs to be forgotten. However, the removal of pertinent data followed by retraining diffusion models from scratch demands substantial resources and is often deemed impractical. A version of the stable diffusion model trained on subsets of the LAION-5B dataset [55] costs approximately 150,000 GPU hours with 256 A100 GPUs1. Existing research on erasing unwanted influence has primarily focused on classification tasks [4, 7, 11, 22, 23, 25, 33, 44, 56, 68]. Despite substantial progress, prior methods developed in classification are observed to be ineffective for generation tasks [16]. Consequently, there is a pressing need for the development of methods capable of scrubbing data from diffusion models without necessitating complete retraining. + +Recently, a handful of studies [6, 16, 20, 21, 27, 28, 36, 41, 71] target unlearning in diffusion models, with a primary focus on the text-to-image models [20, 21, 71]. Broadly, these methods aim to achieve two main objectives: erasing data influence and preserving overall model performance. However, as demonstrated by Bui et al. [5], balancing this trade-off remains challenging. + +In this work, we propose EraseDiff, an algorithm tailored to balance the overall performance of diffusion models with the erasure of undesirable information. Drawing inspiration from optimization-based meta-learning algorithms [18, 42] that enable fast adaptation to new learning tasks, we formulate this challenge as a bi-level optimization problem, where the "inner" optimization focus on erasing undesirable influence and the "outer" objective seeks to preserve model performance. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively. However, this nested optimization can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with non-convex functions, making it difficult to achieve a stable solution [39]. Therefore, we further reformulate the problem as a constrained optimization problem using the value function [40, 46, 70], which facilitates a natural first-order solution [39], allows us to optimize preservation and erasure in a unified manner. This approach achieves a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. We benchmark EraseDiff on various scenarios, encompassing unlearning of classes on CIFAR-10 [35] with Denoising Diffusion Probabilistic Models (DDPM) [30], classes on + +Imagenette [31] and concepts on the I2P dataset [54] with stable diffusion. Our empirical findings show that EraseDiff is $11 \times$ faster than Heng and Soh's method [28] and $2 \times$ faster than Fan's method [16] when forgetting on DDPM while achieving better unlearning results across several metrics. The results demonstrate that EraseDiff is capable of effectively erasing data influence in diffusion models, ranging from specific classes to the concept of nudity. + +# 2. Background + +In this section, we outline the components of the models we evaluate, including DDPM and latent diffusion models [49]. Throughout the paper, we denote scalars, and vectors/matrices by lowercase and bold symbols, respectively (e.g., $a$ , $a$ , $A$ ). + +DDPM. (1) Diffusion: DDPM gradually diffuses the data distribution $\mathbb{R}^d\ni \mathbf{x}_0\sim q(\mathbf{x})$ into the standard Gaussian distribution $\mathbb{R}^d\ni \epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I}_d)$ with $T$ time steps, i.e., $q(\mathbf{x}_t|\mathbf{x}_{t - 1}) = \mathcal{N}(\mathbf{x}_t;\sqrt{\alpha_t}\mathbf{x}_{t - 1},(1 - \alpha_t)\mathbf{I}_d)$ , where $\alpha_{t} = 1 - \beta_{t}$ and $\{\beta_t\}_{t = 1}^T$ are the pre-defined variance schedule. The diffusion takes the form $\mathbf{x}_t$ as $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t}\epsilon$ , where $\bar{\alpha}_t = \prod_{i = 1}^{t}\alpha_{i}$ . (2) Training: A model $\epsilon_{\theta}(\cdot)$ with parameters $\pmb {\theta}\in \mathbb{R}^n$ is trained to learn the reverse process $p_{\pmb{\theta}}(\mathbf{x}_{t - 1}|\mathbf{x}_t)\approx q(\mathbf{x}_{t - 1}|\mathbf{x}_t)$ . Given $\mathbf{x}_0\sim q(\mathbf{x})$ and time step $t\in [1,T]$ , the simplified training objective is to minimize the distance between $\pmb{\epsilon}$ and the predicted $\epsilon_{t}$ given $\mathbf{x}_0$ at time $t$ , i.e., $\| \epsilon -\epsilon_{\pmb{\theta}}(\mathbf{x}_t,t)\|$ . (3) Sampling: after training the model, we could obtain the learnable backward distribution $p_{\pmb{\theta}^{*}}(\mathbf{x}_{t - 1}|\mathbf{x}_{t}) = \mathcal{N}(\mathbf{x}_{t - 1};\pmb{\mu}_{\pmb{\theta}^{*}}(\mathbf{x}_{t},t),\pmb{\Sigma}_{\pmb{\theta}^{*}}(\mathbf{x}_{t},t))$ , where $\pmb{\mu}_{\pmb{\theta}^{*}}(\mathbf{x}_{t},t) = \frac{1}{\sqrt{\alpha_{t}}} (\mathbf{x}_{t} - \frac{\beta_{t}}{\sqrt{1 - \alpha_{t}}}\epsilon_{\pmb{\theta}}(\mathbf{x}_{t},t))$ and $\pmb{\Sigma}_{\pmb{\theta}^{*}}(\mathbf{x}_{t},t) = \frac{(1 - \bar{\alpha}_{t - 1})\beta_{t}}{1 - \bar{\alpha}_{t}}$ . Then, given $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I}_d)$ , $\mathbf{x}_0$ could be obtained via sampling from $p_{\pmb{\theta}^{*}}(\mathbf{x}_{t - 1}|\mathbf{x}_{t})$ from $t = T$ to $t = 1$ step by step. + +Latent diffusion models. Latent diffusion models apply the diffusion models in the latent space $\mathbf{z}$ of a pre-trained variational autoencoder. The noise would be added to $\mathbf{z} = \varepsilon(\mathbf{x})$ , instead of the data $\mathbf{x}$ , and the denoised output would be transformed to image space with the decoder. Besides, text embeddings generated by models like CLIP are used as conditioning inputs. + +# 3. Diffusion Unlearning + +Let $\mathcal{D} = \{\mathbf{x}_i, c_i\}_i^N$ be a dataset of images $\mathbf{x}_i$ associated with label $c_i$ representing the class. $\mathcal{C} = \{1, \dots, C\}$ denotes the label space where $C$ is the total number of classes and $c_i \in \mathcal{C}$ . We split the training data $\mathcal{D}$ into the forgetting data $\mathcal{D}_f \subset \mathcal{D}$ and its complement, remaining data $\mathcal{D}_r = \mathcal{D} \setminus \mathcal{D}_f$ . The forgetting data has label space $\mathcal{C}_f \subseteq \mathcal{C}$ , and the remaining label space is denoted as $\mathcal{C}_r = \mathcal{C} \setminus \mathcal{C}_f$ . + +# 3.1. Training objective + +Our goal is to scrub the information about $\mathcal{D}_f$ carried by the diffusion models while maintaining the model utility over the remaining data $\mathcal{D}_r$ . To achieve this, we adopt different training objectives for $\mathcal{D}_r$ and $\mathcal{D}_f$ as follows. + +For the remaining data $\mathcal{D}_r$ , we fine-tune the diffusion models with the original objective: + +$$ +\mathcal {L} _ {r} (\boldsymbol {\theta}; \mathcal {D} _ {r}) = \mathbb {E} _ {t, \boldsymbol {\epsilon} \in \mathcal {N} \left(\mathbf {0}, \mathbf {I} _ {d}\right), \left(\mathbf {x} _ {0}, c\right) \sim \mathcal {D} _ {r} \times \mathcal {C} _ {r}} [ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} \left(\mathbf {x} _ {t} | c\right) \| ^ {\frac {2}{2}} ], \tag {1} +$$ + +where $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t}\pmb {\epsilon}$ . For the forgetting data $\mathcal{D}_f$ , we aim to let the models fail to generate meaningful images corresponding to $\mathcal{C}_f$ and thus propose: + +$$ +\mathcal {L} _ {f} (\boldsymbol {\theta}; \mathcal {D} _ {f}) = \mathbb {E} _ {t, \boldsymbol {\epsilon} \in \mathcal {N} \left(\mathbf {0}, \mathbf {I} _ {d}\right), \left(\mathbf {x} _ {0}, c\right) \sim \mathcal {D} _ {f} \times \mathcal {C} _ {f}} \left[ \| \boldsymbol {\epsilon} _ {f} - \epsilon_ {\boldsymbol {\theta}} \left(\mathbf {x} _ {t} | c\right) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\epsilon_{f} = \epsilon_{\theta}(\mathbf{x}_{t}|c_{m})$ and $c_{m}\neq c$ so that the denoised image $\mathbf{x}_0$ is not related to the forgetting class/concept $c$ [16, 28]. With this, we hinder the approximator $\epsilon_{\theta}$ to guide the denoising process to obtain meaningful examples for the forgetting data example $\mathbf{x}_0\sim \mathcal{D}_f$ + +To erase the undesirable influence of $\mathcal{D}_f$ and preserve the overall performance, it is common to form + +$$ +\mathcal {L} _ {r} (\boldsymbol {\theta}; \mathcal {D} _ {r}) + \lambda \mathcal {L} _ {f} (\boldsymbol {\theta}; \mathcal {D} _ {f}), \tag {3} +$$ + +with $\lambda > 0$ as the optimization objective (see for example [16]). However, training could be hindered due to the conflicting gradients between the erasing and preservation objectives, preventing a balanced trade-off between erasure and preservation [38]. To address this, rather than scalarizing the two objectives, we consider a framework based on optimization-based meta-learning algorithm [47] that allows iteratively updates to optimize each objective: + +$$ +\min _ {\boldsymbol {\theta}} \mathcal {L} _ {r} (\boldsymbol {\theta}; \mathcal {D} _ {r}) +$$ + +s.t. $\pmb {\theta}\in \underset {\phi}{\arg \min}\mathcal{L}_f(\pmb {\phi};\mathcal{D}_f)$ (4) + +where the outer objective minimizes the remaining loss $\mathcal{L}_r$ (i.e., preserving model utility), the inner optimization minimizes the forgetting loss $\mathcal{L}_f$ (i.e., erasing) with initialization $\mathbb{R}^n\ni \phi_{\mathrm{init}} = \pmb{\theta}$ . Given $\pmb{\theta}$ , the inner optimization on $\phi$ aims to minimize the forgetting data influence, with the goal of achieving effective erasure while preserving model utility. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively. + +While the above framework allows for iterative updates to address the conflicting objectives of erasure and preservation, it still relies on nested optimization, which can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with nonconvex functions, making it difficult to achieve a stable solution [39]. To further streamline the optimization process, + +we adopt a value function approach [40, 46, 70] that reformulates the problem as a single-constrained optimization: + +$$ +\min _ {\boldsymbol {\theta}} \mathcal {L} _ {r} (\boldsymbol {\theta}; \mathcal {D} _ {r}) +$$ + +s.t. $\mathcal{L}_f(\pmb {\theta};\mathcal{D}_f) - \min_{\phi}\mathcal{L}_f(\phi ;\mathcal{D}_f)\leq 0,$ (5) + +where $\phi$ is initialized at $\theta$ , leverages the value function to encapsulate the influence of data erasure directly as a constraint on $\mathcal{L}_f$ . This avoids the need for a nested loop by capturing the forgetting objective as a constraint and provides a natural first-order solution, as the constrained formulation allows us to optimize preservation and erasure in a unified manner. + +# 3.2. Solution + +To solve Eq. (5), let us first denote $g(\pmb{\theta}) \coloneqq \mathcal{L}_f(\pmb{\theta};\mathcal{D}_f) - \min_{\phi}\mathcal{L}_f(\phi ;\mathcal{D}_f)$ . Our goal is to erase undesirable influence while preserving the overall model performance, hence the update vector $\delta_t$ for updating the model should aid in minimizing $\mathcal{L}_r(\pmb{\theta};\mathcal{D}_r)$ and $g(\pmb{\theta})$ simultaneously. In other words, suppose that the current solution for Eq. (5) is $\pmb{\theta}_t$ , we aim to update $\pmb{\theta}_{t + 1} = \pmb{\theta}_t - \eta \pmb{\delta}_t$ where $\eta$ is sufficiently small, so that $\mathcal{L}_r(\pmb{\theta}_{t + 1};\mathcal{D}_r)$ decreases (i.e., preserve model utility) and $g(\pmb{\theta}_{t + 1})$ decreases (i.e., erasure). To this end, we aim to find the update vector $\delta_t$ by: + +$$ +\boldsymbol {\delta} _ {t} \in \frac {1}{2} \operatorname {a r g m i n} _ {\boldsymbol {\delta}} \left\| \nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {r} \left(\boldsymbol {\theta} _ {t}; \mathcal {D} _ {r}\right) - \boldsymbol {\delta} \right\| _ {2} ^ {2}, +$$ + +s.t. $\nabla_{\pmb{\theta}}g(\pmb{\theta}_t)^\top \pmb{\delta} \geq a_t > 0.$ (6) + +This will ensure that the update $\delta_t$ is close to $\nabla_\theta \mathcal{L}_r(\pmb{\theta}_t; \mathcal{D}_r)$ and decreases $g(\pmb{\theta}_t)$ until it reaches stationary. Because $g(\pmb{\theta}_{t+1}) - g(\pmb{\theta}_t) \approx -\eta \nabla_\theta g(\pmb{\theta}_t)^\top \delta \leq -\eta a_t < 0$ for some scalar $a_t > 0$ , we can ensure that $g(\pmb{\theta}_{t+1}) < g(\pmb{\theta}_t)$ for small step size $\eta > 0$ . This means that the update $\delta_t$ can ensure to minimize $\mathcal{L}_f(\pmb{\theta}; \mathcal{D}_f)$ as long as it does not conflict with descent of $\mathcal{L}_r(\pmb{\theta}; \mathcal{D}_r)$ . + +To find the solution to the optimization problem in Eq. (6), the following theorem is developed: + +Theorem 3.1. The optimal solution of the optimization problem in Eq. (6) is $\delta^{*} = \nabla_{\pmb{\theta}}\mathcal{L}_{r}(\pmb{\theta}_{t};\mathcal{D}_{r}) + \lambda_{t}\nabla_{\pmb{\theta}}g(\pmb{\theta}_{t})$ where $\lambda_{t} = \max \{0,\frac{a_{t} - \nabla_{\pmb{\theta}}g(\pmb{\theta}_{t})^{\top}\nabla_{\pmb{\theta}}\mathcal{L}_{r}(\pmb{\theta}_{t};\mathcal{D}_{r})}{\|\nabla_{\pmb{\theta}}g(\pmb{\theta}_{t})\|_{2}^{2}}\}$ . + +We provide the proof in §7 in the Appendix. This provides the solution to the optimization problem by constructing the update vector $\delta$ to balance two competing objectives. The variable $a_{t}$ adjusts the weight of the forgetting objective, ensuring that the update vector $\delta$ decreases the remaining loss without violating the erasure goal, and achieves the dual goals of maintaining utility and achieving erasing. In practice, we can choose $a_{t} = \eta \| \nabla_{\theta}g(\pmb{\theta}_{t})\|_{2}^{2}$ , and we start from $\phi^0 = \pmb{\theta}_t$ and use gradient descend in $K$ steps with the learning rate $\xi > 0$ to reach $\phi^K$ , namely $\phi^{k + 1} = \phi^k - \xi \nabla_\phi \mathcal{L}_f(\phi^k; \mathcal{D}_f)$ and $k = 0, \dots, K - 1$ . + +Algorithm 1 EraseDiff: Erasing undesirable influence in diffusion models. + +Input: Well-trained model with parameters $\theta_0$ , forgetting data $\mathcal{D}_f$ and remaining data $\mathcal{D}_r$ , outer iteration number $T$ and inner iteration number $K$ , learning rate $\eta$ . + +Output: Parameters $\theta^{*}$ for the scrubbed model. + +1: for iteration $t$ in $T$ do +2: $\phi^0 = \pmb {\theta}_t$ +3: Get $\phi^K$ by $K$ steps of gradient descent on $\mathcal{L}_f(\phi; \mathcal{D}_f)$ starting from $\phi^0$ . +4: Set $g(\pmb{\theta}_t) = \mathcal{L}_f(\pmb{\theta}_t; \mathcal{D}_f) - \mathcal{L}_f(\phi^K; \mathcal{D}_f)$ . +5: Update the model: $\pmb{\theta}_{t + 1} = \pmb{\theta}_t - \eta (\nabla_{\pmb{\theta}_t}\mathcal{L}_r(\pmb {\theta}_t;\mathcal{D}_r) + \lambda_t\nabla_{\pmb{\theta}_t}g(\pmb {\theta}_t;\phi^K))$ +6: where $\lambda_{t} = \max \{0,\frac{a_{t} - \nabla_{\pmb{\theta}}g(\pmb{\theta}_{t})^{T}\nabla_{\pmb{\theta}}\mathcal{L}_{r}(\pmb{\theta}_{t};\mathcal{D}_{r})}{\|\nabla_{\pmb{\theta}}g(\pmb{\theta}_{t})\|_{2}^{2}}\}$ . +7: end for + +# 3.3. Analysis + +We can characterize the solution of our algorithm as follows and the proof can be found in §7 in the Appendix: + +Theorem 3.2 (Pareto optimality). The stationary point obtained by our algorithm is Pareto optimal of the problem $\min_{\pmb{\theta}}[\mathcal{L}_r(\pmb{\theta};\mathcal{D}_r),\mathcal{L}_f(\pmb{\theta};\mathcal{D}_f)]$ + +This asserts that the solution obtained by our algorithm is Pareto optimal for the problem of minimizing both objectives, which implies that the solution obtained by the algorithm ensures a balanced trade-off between preserving model utility and erasing undesirable influences. + +We further take DDPM with CIFAR-10 when forgetting the 'airplane' as an example to show that our proposed method helps alleviate the gradient conflict which prevents a balanced trade-off between erasure and preservation. Fig. 2 presents the cosine similarity between the update vector $\delta$ and the gradient $g_{r} = \nabla_{\theta}\mathcal{L}_{r}(\theta ;\mathcal{D}_{r})$ for preservation, and the cosine similarity between the update vector $\delta$ and the gradient $g_{f} = \nabla \mathcal{L}_{f}(\theta ;\mathcal{D}_{f})$ for erasing. The cosine similarity represents the alignment between the update vector and the gradients associated with the preservation and erasure. Higher positive values indicate alignment, meaning that the update vector $\delta$ is directed similarly to the respective gradient, whereas negative values indicate misalignment or conflict. In particular, negative values suggest a high degree of opposition between the update vector and the respective gradient, which can signify competing objectives between preservation and erasure during the optimization process. + +MOO (Multi-Objective Optimization) denotes the naive integration of erasing and preservation as stated in Eq. (3). For the vanilla MOO, Fig. 2 shows a clear alternating pattern in cosine similarity values between the update vector $\delta$ and the gradients $g_{r}$ and $g_{f}$ . Specifically, when the cosine + +![](images/b5c502d8fe4a19dd98c2fc18080b49c434130f3803283a2c87acbc90e92fab90.jpg) + +![](images/10377f078ff0c8c707ad539046dee7ac2a2c2efa1cef9481f0d73191f5482acc.jpg) +Figure 2. Top to bottom: cosine similarity between the update vector $\delta$ and the preservation gradient $g_{r}$ , followed by the cosine similarity between $\delta$ and the erasing gradient $g_{f}$ . Positive values indicate alignment, while negative values suggest conflict. This visualization illustrates how well the update vector aligns with the objectives of preservation and erasure over successive iterations. + +similarity between $\delta$ and $g_{r}$ is greater than 0, the similarity between $\delta$ and $g_{f}$ tends to be less than 0, and vice versa. This pattern suggests that MOO experiences gradient conflict, as it cannot effectively balance the two objectives of preservation and erasure, preventing MOO from achieving a harmonious update that supports both goals simultaneously. In contrast, EraseDiff mostly shows positive cosine similarity values between the update vector $\delta$ and both the preservation gradient $g_{r}$ and the erasing gradient $g_{f}$ . This indicates that EraseDiff aligns the update direction with both objectives, suggesting it manages to avoid significant gradient conflict. By maintaining positive alignment, EraseDiff appears to balance preservation and erasure more effectively, leading to better cooperation between objectives. + +# 4. Related Work + +Memorization in generative models. Privacy of generative models has been studied extensively for GANs [17, 43, + +66] and generative language models [8, 9, 32, 63]. These generative models often risk replicating from their training data. Recently, several studies [10, 58, 59, 65] investigated these data replication behaviors in diffusion models, raising concerns about the privacy and copyright issues. Possible mitigation strategies are deduplicating and randomizing conditional information [58, 59], or training models with differential privacy (DP) [1, 13-15]. However, leveraging DP-SGD [1] may cause training to diverge [10]. + +Malicious misuse. Diffusion models usually use training data from varied open sources and when such unfiltered data is employed, there is a risk of it being tainted [12] or manipulated [48], resulting in inappropriate generation [54]. They also risk the imitation of copyrighted content, e.g., mimicking the artistic style [20, 57]. To counter inappropriate generation, data censoring [3, 19, 45, 53] where excluding black-listed images before training, and safety guidance where diffusion models will be updated away from the inappropriate/undesired concept [20, 54] are proposed. [34] uses human feedback to annotate the generations for getting soft tokens that quantify the harmfulness of the images, then fine-tunes the model using self-distillation to achieve harmony between preservation and erasure. Shan et al. [57] propose protecting artistic style by adding barely perceptible perturbations to the artworks before public release. Yet, Rando et al. [48] argue that DMs can still generate content that bypasses the filter. Chen et al. [12] highlight the susceptibility of DMs to poison attacks, where target images are generated with specific triggers. + +Machine unlearning. Removing data directly involves retraining the model from scratch, which is inefficient and impractical. Thus, to reduce the computational overhead, efficient machines unlearning methods [4, 7, 11, 22, 23, 25, 33, 44, 50, 56, 61, 68] have been proposed. Several studies [5, 16, 20, 21, 27, 28, 67, 71] recently introduce unlearning in diffusion models. Most of them [20, 21, 27, 71] mainly focus on text-to-image models and high-level visual concept erasure. Heng and Soh [28] adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) from continual learning to perform unlearning effectively without access to the training data. Heng and Soh's method can be applied to a wide range of generative models, however, it needs the computation of FIM for different datasets and models, which may lead to significant computational demands. Fan et al. [16] propose a very potent unlearning algorithm called SalUn that shifts attention to important parameters w.r.t. the forgetting data. SalUn can perform effectively across image classification and generation tasks. + +In this work, we introduce a simple yet effective unlearning algorithm for diffusion models by formulating the problem as a constrained optimization problem, to achieve a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. Below, we will show + +that our algorithm is not only faster than Heng and Soh's method [28] and Fan's method [16], but even outperforms these methods in terms of the trade-off between the forgetting and preserving model utility. + +# 5. Experiment + +We evaluate EraseDiff in various scenarios, including removing images with specific classes/concepts, to answer the following research questions (RQs): (i) Can typical machine unlearning methods be applied to diffusion models? (ii) Is EraseDiff able to remove the influence of $\mathcal{D}_f$ in the diffusion models? (iii) Is EraseDiff able to preserve the model utility while removing $\mathcal{D}_f$ ? (iv) Is EraseDiff efficient in removing the data? (v) How does EraseDiff perform on the public well-trained models? + +# 5.1. Setup + +Experiments are reported on CIFAR-10 [35] with DDPM, Imagenette [31] with Stable Diffusion (SD) for class-wise forgetting, I2P [54] dataset with SD for concept-wise forgetting. For all SD experiments, we use the open-source SD v1.4 [49] checkpoint as the pre-trained model. Implementation details and additional results like visualizations of generated images can be found in §8 and §9. + +Baselines. We primarily benchmark against the following baselines commonly used in machine unlearning: (i) Unscrubbed, (ii) Finetune (FT) [23], (iii) NegGrad (NG) [23], (iv) BlindSpot [61], (v) ESD [20], (vi) FMN [71], (vii) Selective Amnesia (SA) [28] and (viii) the SOTA machine unlearning algorithm SalUn [16]. + +Metrics. Several metrics are utilized to evaluate the algorithms: (i) Frechet Inception Distance (FID) [29]: the widely-used metric for assessing the quality of generated images. (ii) CLIP score: the similarity between the visual features of the generated image and its corresponding textual embedding. (iii) $P_{\psi}(\mathbf{y} = c_f|\mathbf{x}_f)$ [28]: the classification rate of a pre-trained classifier $P_{\psi}(\mathbf{y}|\mathbf{x})$ , with a ResNet architecture [26] used to classify generated images conditioned on the forgetting classes. A lower classification value indicates superior unlearning performance. (iv) Precision and Recall: A low FID may indicate high precision (realistic images) but low recall (small variations) [37, 51]. Kynkänniemi et al. [37] shows that generative models claim to optimize FID (high fidelity) but always sacrifice variation (low diversity). Hence, we include metric precision (fidelity) and recall (diversity) to express the quality of the generated samples, to provide explicit visibility of the tradeoff between sample quality and variety. + +# 5.2. Results on DDPM + +Following SA, we aim to forget the 'airplane' class on CIFAR-10. Here, we replace $\epsilon \in \mathcal{N}(\mathbf{0},\mathbf{I}_d)$ with $\epsilon_{f} = \epsilon_{\theta}(\mathbf{x}_{t}|c_{m})$ like random labelling used in [16] where $c_{m}\neq c$ + +Table 1. Results on CIFAR10 with DDPM when forgetting the 'airplane' class. $P_{\psi}(\mathbf{y} = c_f|\mathbf{x}_f)$ indicate the probability of the forgotten class (i.e., the effectiveness of erasing). Precision and Recall demonstrate the fidelity and diversity [37, 51], and FID scores are computed between the generated 45K images and the corresponding ground truth images with the same labels from $\mathcal{D}_r$ (i.e., preserving model utility). SA excels in class-wise forgetting but struggles to perform concept-wise forgetting as shown in Fig. 3 and Tab. 2. The best and the second best are highlighted in blue and orange, respectively. + +
UnscrubbedFT [23]NG [23]BlindSpot [61]SA [28]SalUn [16]EraseDiffrlEraseDiffnoise
FID ↓9.638.2176.739.128.199.168.667.61
Precision (fidelity) ↑0.400.430.080.410.430.410.430.43
Recall (diversity) ↑0.790.770.610.780.750.760.770.72
Pψ(y = cf|xf)↓0.970.960.610.900.060.070.240.22
+ +![](images/1736ccbff23f646147064b068bc6a26b1f06e1358197e68507ddc6d68ddf9dfb.jpg) +Figure 3. Quantity of nudity content detected using the NudeNet classifier from I2P data. Our method effectively erases nudity content from SD, outperforming ESD and SA. Note that Fig. 4 and Tab. 2 together presents the trade-off between erasing and preservation. + +![](images/b80ded473f9a468bd3de678469e126074cd2ade9028534305a8083ef3cc4e020.jpg) +Table 2. Evaluation of $30\mathrm{K}$ generated images by SD when erasing 'nudity'. The FID score is measured compared to validation data, while the CLIP similarity score evaluates the alignment between generated images and the corresponding prompts. The best and the second best are highlighted in blue and orange, respectively. + +![](images/f3f99bc74074be9ed65566e4bf084f34cbffaddd925b8a8920c7c8b7a7a49359.jpg) + +![](images/0369b42997825dc6c18aed22fc3078892a901db3cceabcf224f59d0493a1c38f.jpg) + +
ESD [20]SA [28]SalUn [16]EraseDiff
FID ↓15.7625.5825.0617.01
CLIP ↑30.3331.0328.9130.58
+ +denoted as $\text{EraseDiff}_{\text{rl}}$ . We also try to use $\epsilon_f = \mathcal{U}(\mathbf{0}, \mathbf{I}_d)$ like SA, denoted as $\text{EraseDiff}_{\text{noise}}$ . Note that the choice of replacement for forgotten classes is flexible and is not the primary focus of this work. For further discussion on the choice of substitution strategies, please refer to [6, 41]. + +Results are presented in Tab. 1. Firstly, from Tab. 1, we can conclude that traditional machine unlearning methods designed for image classification or regression tasks fall short in effectively performing forgetting for DDPM. Finetune and BlindSpot suffer from under-forgetting (i.e., the generated image quality is good but the probability of generated images belonging to the forgetting class approaching the value of the unscrubbed model), and NegGrad suffers from over-forgetting (the probability of generated images belonging to the forgetting class is decreased compared to + +that of the unscrubbed model but the generated image quality drops significantly). + +Then, comparing SA and SalUn's unlearning methods, SA achieves an FID score of 8.19 but sacrifices variation (decreased recall). Also, note that SA introduces excessive computational resource requirements and time consumption [28, 72]. Note that the FID scores of SA, SalUn, and EraseDiff decrease compared with the generated images from the original models; the quality of the generated images experiences a slight improvement. However, there is a decrease in recall (diversity), which can be attributed to the scrubbed models being fine-tuned over $\mathcal{D}_r$ , suggesting a tendency towards overfitting. Regarding forgetting, SalUn achieves a smaller probability of the generated images classified as the forgetting class than ours; yet, the FID score is larger than ours, and images generated by EraseDiffr1 present better diversity and fidelity. + +# 5.3. Results on Stable Diffusion + +We apply EraseDiff to perform class-wise forgetting from Imagenette and erase the 'nudity' concept with SD v1.4. When forgetting 'nudity', we have no access to the training data; instead, we generate $\sim 400$ images with the prompts $c_{f} = \{\text{'nudity', 'naked', 'erotic', 'sexual'}\}$ . + +Forget nudity. 4703 images are generated using I2P prompts, and 1K images are generated using the prompts + +![](images/77f150cff340b7bd1a389bc17e00771d50359d87ce9cbe07b6d9838e1b9760e2.jpg) +Table 3. P forgetting +{nudity, nak, sexual} The quality of nu-dity content is detected using the NudeNet classifier [2]. In Fig. 3, the number in the y-axis denotes the number of exposed body parts generated by the SD v1.4 model. Fig. 3 presents the percentage change in exposed body parts w.r.t. SD v1.4. In §9, we provide the number of exposed body parts counted in all generated images with different thresholds. Here, our algorithm replaces $\epsilon_{f}$ with $\epsilon_{\theta}(\mathbf{x}_i|c_m)$ where $c_{m}$ is 'a photo of poker'. We can find that, EraseD-iff reduces the amount of nu-dity content compared to SD v1.4, ESD, and SA, particularly on sensitive content like Female/Male interests and female/Male Genitalia. While SalUn excels a 'forgetting', our algorithm demonstrates a significant improvement in the quality of generated images as shown in Tab. 2. Tab. 2 presented results evaluating the utility of scrubbed models. The FID and CLIP scores are + +![](images/e0d6c7c35f43ceedc354c7e64bdcc9066f2fba320467025a689a34fd5d17e7b5.jpg) + +Table 4. Computational overhead. Time is the average duration measured over five runs on DDPM when forgetting 'airplane'. + +
Memory (MiB)Time (min.)Complexity
SA3352.3140.00O(n2)
SalUn4336.228.17O(n)
EraseDiff3360.312.70O(n)
+ +approach, clearly indicate the advantages of our method. We also present results when improving the forgetting ability of SalUn in §9. However, note that this enhancement comes with a drop in the FID score of the generated images. Our method, while slightly better than SalUn on average across 10 classes, demonstrates a more balanced trade-off between erasing and preservation, indicating that it achieves a favorable balance in preserving fidelity while enhancing erasing performance. + +# 5.4. Computational efficiency + +Finally, we measure the computational complexity of unlearning algorithms. The computational complexity of SA and SalUn involves two distinct stages: the computation of FIM for SA and the computation of salient weights w.r.t. $\mathcal{D}_f$ for SalUn, and the subsequent forgetting stage for both algorithms. We consider the maximum memory usage across both stages, the metric 'Time' is exclusively associated with the duration of the forgetting stage for unlearning algorithms. Tab. 4 show that EraseDiff outperforms SA and SalUn in terms of efficiency, achieving a speed increase of $\sim 11\times$ than SA and $\sim 2\times$ than SalUn. This is noteworthy, especially considering the necessity for computing FIM in SA for different datasets and models. + +# 5.5. Ablation study + +We further investigate the influence of the number of iterations $K$ that approximate $\min \mathcal{L}_f(\phi ;\mathcal{D}_f)$ , and the step size $\eta$ that controls the weight of forgetting and preserving model utility. Here, we replace $\epsilon \in \mathcal{N}(\mathbf{0},\mathbf{I}_d)$ with $\epsilon_{f}\in \mathcal{U}(\mathbf{0},\mathbf{I}_{d})$ . Note that for different hyperparameters in Fig. 5 (a), the average entropy of the classifier's output distribution given $\mathbf{x}_f$ , which is $H(P_{\psi}(\mathbf{y}|\mathbf{x}_f)) = -\mathbb{E}[\sum_iP_{\psi}(\mathbf{y} = c_i|\mathbf{x})\log_eP_{\psi}(\mathbf{y} = c_i|\mathbf{x})]$ , remains close to 2.02. This indicates that the scrubbed models become uncertain about the images conditioned on the forgetting class, effectively erasing the information about $\mathcal{D}_f$ . Below, we will further demonstrate the influence on the model utility. In practice, we have $\lambda_{t} = \max \{0,\frac{a_{t} - \nabla_{\theta}g(\pmb{\theta}_{t})^{\top}\nabla_{\theta}\mathcal{L}_{r}(\pmb{\theta}_{t};\mathcal{D}_{r})}{\|\nabla_{\theta}g(\pmb{\theta}_{t})\|_{2}^{2}}\} = \max \{0,\eta -\frac{\nabla_{\theta}g(\pmb{\theta}_{t})^{\top}\nabla_{\theta}\mathcal{L}_{r}(\pmb{\theta}_{t};\mathcal{D}_{r})}{\|\nabla_{\theta}g(\pmb{\theta}_{t})\|_{2}^{2}}\}$ , we can see that $\eta$ determines the extent to which the update direction for forgetting can deviate from that for preserving model utility. A larger $\eta$ would allow for more deviation in the updating, thus pri + +![](images/bbb94cf6ab7a3076aeebc533cd9b12764298cc2921b6e7cf3f2b7874e58ba66b.jpg) +Figure 5. (a) Ablation results. (b) Potential incomplete erasures. + +oritizing forgetting over preserving model utility. In Fig. 5 (a), the FID score tends to increase (i.e., image quality drop) as the step size $\eta$ increases, indicating that larger $\eta$ leads to greater deviations from the direction that preserves the model utility. Furthermore, the number of iterations $K$ determines how closely the approximation $\phi^K$ will approach $\arg \min_{\phi} \mathcal{L}_f(\phi; \mathcal{D}_f)$ . Hence, a larger number of iterations $K$ leads to more thorough erasure, which is also supported by the results shown in Fig. 5 (a), as increasing $K$ correlates with an increase in the FID score. + +# 6. Conclusion, limitations and broader impacts + +In this work, we explored erasing undesirable influence in diffusion models and proposed an efficient method EraseDiff to achieve a balanced trade-off between erasing and preservation. Comprehensive experiments on diffusion models demonstrate the proposed algorithm's effectiveness in data removal, its efficacy in preserving the model utility, and its efficiency in erasure. + +However, our scrubbed model may have incomplete erasures, potentially arises due to the choice of the replacements. In Fig. 5 (b), with random labeling, generated images conditioned on the forgetting class 'tench' by our scrubbed model may preserve some characteristics similar to that close to 'tench' visually; instead, replacing the erased class 'tench' with other class such as 'dog' could improve in erasure. Besides, the scrubbed models could be biased or be misused for censorship or exploitation. This includes using technology to selectively remove or amplify NSFW content in various ways. They might also falsely assure users that NSFW content is fully removed when it isn't. To address this, advanced privacy-preserving training techniques are in demand to enhance the security and fairness of the models. Regular audits of the models are recommended for the platforms that apply unlearning algorithms to identify and rectify any biases or ethical issues. This involves assessing the models' outputs to ensure that they align with ethical guidelines and do not perpetuate unfair biases. + +# Acknowledgements + +Mehrtash Harandi is supported by the Australian Research Council (ARC) Discovery Program DP250100262. The authors gratefully acknowledge the anonymous reviewers for their insightful feedback and valuable suggestions, which have significantly improved the quality of this work. + +# References + +[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308-318, 2016. 5 +[2] P Bedapudi. Nudenet: Neural nets for nudity classification, detection and selective censoring, 2019. 7, 2 +[3] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1536-1546. IEEE, 2021. 5 +[4] Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159, 2021. 2, 5 +[5] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 2, 5 +[6] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 2, 6 +[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480, 2015. 2, 5 +[8] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633-2650, 2021. 5 +[9] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022. 5 +[10] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 5 +[11] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 2, 5 + +[12] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 5 +[13] Tim Dockhorn, Tianshi Cao, Arash Vahdat, and Karsten Kreis. Differentially private diffusion models. arXiv preprint arXiv:2210.09929, 2022. 5 +[14] Cynthia Dwork. Differential privacy: A survey of results. In International conference on theory and applications of models of computation, pages 1-19. Springer, 2008. +[15] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, New York, NY, USA, March 4-7, 2006. Proceedings 3, pages 265-284. Springer, 2006. 5 +[16] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Dennis Wei, Eric Wong, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3, 5, 6, 7 +[17] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701–6710, 2021. 4 +[18] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2 +[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 5 +[20] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In 2023 IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5, 6, 7 +[21] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 5 +[22] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2, 5 +[23] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309, 2020. 2, 5, 6 +[24] Eric Goldman. An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper, 2020. 1 +[25] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning + +models. In Proceedings of the 37th International Conference on Machine Learning, pages 3832-3842. PMLR, 2020. 2, 5 +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[27] Alvin Heng and Harold Soh. Continual learning for forgetting in deep generative models. 2023. 2, 5 +[28] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 5, 6 +[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[31] Jeremy Howard and Sylvain Gugger. Fastai: A layered api for deep learning. Information, 11(2):108, 2020. 2, 5 +[32] Matthew Jagielski, Om Thakkar, Florian Tramer, Daphne Ippolito, Katherine Lee, Nicholas Carlini, Eric Wallace, Shuang Song, Abhradeep Thakurta, Nicolas Papernot, et al. Measuring forgetting of memorized training examples. arXiv preprint arXiv:2207.00099, 2022. 5 +[33] Masayuki Karasuyama and Ichiro Takeuchi. Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks, 21(7):1048-1059, 2010. 2, 5 +[34] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Safeguard text-to-image diffusion models with human feedback inversion. In European Conference on Computer Vision, pages 128-145. Springer, 2024. 5 +[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5 +[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 2 +[37] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5, 6 +[38] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems (NeurIPS), 34:18878-18890, 2021. 3 +[39] Bo Liu, Mao Ye, Stephen Wright, Peter Stone, and Qiang Liu. Bome! bilevel optimization made easy: A simple first-order approach. Advances in Neural Information Processing Systems, 35:17248-17262, 2022. 2, 3 +[40] Risheng Liu, Xuan Liu, Xiaoming Yuan, Shangzhi Zeng, and Jin Zhang. A value-function-based interior-point method + +for non-convex bi-level optimization. In International conference on machine learning, pages 6882-6892. PMLR, 2021. 2, 3 +[41] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 2, 6 +[42] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In International conference on machine learning, pages 2113-2122. PMLR, 2015. 2 +[43] Casey Meehan, Kamalika Chaudhuri, and Sanjoy Das-gupta. A non-parametric test to detect data-copying in generative models. In International Conference on Artificial Intelligence and Statistics, 2020. 4 +[44] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N. Ravi. Deep unlearning via randomized conditionally independent hessenians. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10412-10421, 2022. 2, 5 +[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5 +[46] Jiri V Outrata. On the numerical solution of a class of stackelberg problems. Zeitschrift für Operations Research, 34: 255-277, 1990. 2, 3 +[47] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3 +[48] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1, 5 +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5 +[50] Enrique Romero, Ignacio Barrio, and Lluis Belanche. Incremental and decremental learning for linear support vector machines. In International Conference on Artificial Neural Networks, pages 209-218. Springer, 2007. 5 +[51] Mehdi SM Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. Advances in neural information processing systems, 31, 2018. 5, 6 +[52] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023.1 +[53] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? + +In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350-1361, 2022. 5 +[54] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 5 +[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[56] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS), 34: 18075-18086, 2021. 2, 5 +[57] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 5 +[58] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 1, 5 +[59] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 5 +[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1 +[61] Ayush Kumar Tarun, Vikram Singh Chandawat, Murari Mandal, and Mohan Kankanhalli. Deep regression unlearning. In International Conference on Machine Learning, pages 33921-33939. PMLR, 2023. 2, 5, 6 +[62] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[63] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 35: 38274-38290, 2022. 5 +[64] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1 +[65] Nikhil Vyas, Sham Kakade, and Boaz Barak. Provable copyright protection for generative models. arXiv preprint arXiv:2302.10870, 2023. 5 +[66] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. This person (probably) exists. identity member- + +ship attacks against gan generated faces. arXiv preprint arXiv:2107.06018, 2021.5 +[67] Jing Wu and Mehrtash Harandi. Scissorhands: Scrub data influence via connection sensitivity in networks. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 5 +[68] Yinjun Wu, Edgar Dobriban, and Susan Davidson. DeltaGrad: Rapid retraining of machine learning models. In Proceedings of the 37th International Conference on Machine Learning, pages 10355-10366. PMLR, 2020. 2, 5 +[69] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 87-103. Springer, 2022. 2 +[70] Jane J Ye and DL Zhu. Optimality conditions for bilevel programming problems. Optimization, 33(1):9-27, 1995. 2, 3 +[71] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2, 5, 7 +[72] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024.6 \ No newline at end of file diff --git a/2025/Erasing Undesirable Influence in Diffusion Models/images.zip b/2025/Erasing Undesirable Influence in Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6dacd23fb449c2bce0da528115f5140b05f1c026 --- /dev/null +++ b/2025/Erasing Undesirable Influence in Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e631c829b902c6ac3f81eed58760027ba9ff252561b769d484d3be66f7d14b1 +size 785564 diff --git a/2025/Erasing Undesirable Influence in Diffusion Models/layout.json b/2025/Erasing Undesirable Influence in Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5a069195c839f5bf2b5d721d242a1b85faf290a4 --- /dev/null +++ b/2025/Erasing Undesirable Influence in Diffusion Models/layout.json @@ -0,0 +1,11436 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "text", + "content": "Erasing Undesirable Influence in Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "text", + "content": "Jing Wu" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "text", + "content": ", Trung Le" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "text", + "content": ", Munawar Hayat" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "text", + "content": ", Mehrtash Harandi" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "text", + "content": "Monash University, Melbourne, VIC, Australia, " + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 143, + 504, + 186 + ], + "type": "text", + "content": "Qualcomm, San Deigo, CA, US \n{jing.wu1, trunglm, mehrtash.harandi}@monash.edu, hayat@qti.Qualcomm.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 227, + 142, + 305 + ], + "blocks": [ + { + "bbox": [ + 59, + 227, + 142, + 305 + ], + "lines": [ + { + "bbox": [ + 59, + 227, + 142, + 305 + ], + "spans": [ + { + "bbox": [ + 59, + 227, + 142, + 305 + ], + "type": "image", + "image_path": "17ae78a31740d4b9de3e311438bca2c72819f358973e92ee65de28f126d1caf4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 305, + 142, + 387 + ], + "blocks": [ + { + "bbox": [ + 59, + 305, + 142, + 387 + ], + "lines": [ + { + "bbox": [ + 59, + 305, + 142, + 387 + ], + "spans": [ + { + "bbox": [ + 59, + 305, + 142, + 387 + ], + "type": "image", + "image_path": "112ca5af0c57a515ab46f0c02b96e87d6ccb29b94fdd4b546684b020963e5578.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 414, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 54, + 414, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. Top to Bottom: generated samples by SD v1.4 and model scrubbed by our method, when erasing the concept of 'nudity'. Our method can avoid NSFW (not safe for work) content while preserving model utility." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 228, + 225, + 305 + ], + "blocks": [ + { + "bbox": [ + 153, + 213, + 214, + 227 + ], + "lines": [ + { + "bbox": [ + 153, + 213, + 214, + 227 + ], + "spans": [ + { + "bbox": [ + 153, + 213, + 214, + 227 + ], + "type": "text", + "content": "I2P prompts" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 228, + 225, + 305 + ], + "lines": [ + { + "bbox": [ + 143, + 228, + 225, + 305 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 225, + 305 + ], + "type": "image", + "image_path": "46651389b1cebd84f1d8b5c41e8306608112e2649f0608ef1ffb4cd97acadb8e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 143, + 305, + 225, + 388 + ], + "blocks": [ + { + "bbox": [ + 143, + 305, + 225, + 388 + ], + "lines": [ + { + "bbox": [ + 143, + 305, + 225, + 388 + ], + "spans": [ + { + "bbox": [ + 143, + 305, + 225, + 388 + ], + "type": "image", + "image_path": "775af8103bcd653a1c4507557af163384bb954287a16316b8aab955358cea7d4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 227, + 227, + 306, + 304 + ], + "blocks": [ + { + "bbox": [ + 227, + 227, + 306, + 304 + ], + "lines": [ + { + "bbox": [ + 227, + 227, + 306, + 304 + ], + "spans": [ + { + "bbox": [ + 227, + 227, + 306, + 304 + ], + "type": "image", + "image_path": "5d079d5f12be466c8d22260e31dbe7cd289d4bac356b0bd7cf634f60f48e4119.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 227, + 305, + 306, + 388 + ], + "blocks": [ + { + "bbox": [ + 227, + 305, + 306, + 388 + ], + "lines": [ + { + "bbox": [ + 227, + 305, + 306, + 388 + ], + "spans": [ + { + "bbox": [ + 227, + 305, + 306, + 388 + ], + "type": "image", + "image_path": "94b147510b9c83c4d921141cc1c58f5f87b9de82e0ee20960bc34fb8fe6eab1e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 227, + 394, + 388 + ], + "blocks": [ + { + "bbox": [ + 317, + 227, + 394, + 388 + ], + "lines": [ + { + "bbox": [ + 317, + 227, + 394, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 227, + 394, + 388 + ], + "type": "image", + "image_path": "22f0da26f62de723d6de6c75a1720fd07c6dec1d63395f6eeb0ed676cfb20db1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 227, + 473, + 387 + ], + "blocks": [ + { + "bbox": [ + 398, + 213, + 474, + 227 + ], + "lines": [ + { + "bbox": [ + 398, + 213, + 474, + 227 + ], + "spans": [ + { + "bbox": [ + 398, + 213, + 474, + 227 + ], + "type": "text", + "content": "COCO prompts" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 227, + 473, + 387 + ], + "lines": [ + { + "bbox": [ + 395, + 227, + 473, + 387 + ], + "spans": [ + { + "bbox": [ + 395, + 227, + 473, + 387 + ], + "type": "image", + "image_path": "12136886ef4f30ac41cb0bff03c0d05e5bf4aa5a115ca05d76dbc8f7329ed1b1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 474, + 227, + 552, + 305 + ], + "blocks": [ + { + "bbox": [ + 474, + 227, + 552, + 305 + ], + "lines": [ + { + "bbox": [ + 474, + 227, + 552, + 305 + ], + "spans": [ + { + "bbox": [ + 474, + 227, + 552, + 305 + ], + "type": "image", + "image_path": "444ae9d1d59ca02f32f037433f72ad07f919a10e9e0734206a955068ae06af6b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 474, + 305, + 552, + 388 + ], + "blocks": [ + { + "bbox": [ + 474, + 305, + 552, + 388 + ], + "lines": [ + { + "bbox": [ + 474, + 305, + 552, + 388 + ], + "spans": [ + { + "bbox": [ + 474, + 305, + 552, + 388 + ], + "type": "image", + "image_path": "3c847e1f63febefb6ff5cfd5330894d12ceb8294736b4c33a09340279ed2c3f1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 381, + 389, + 417, + 404 + ], + "blocks": [ + { + "bbox": [ + 381, + 389, + 417, + 404 + ], + "lines": [ + { + "bbox": [ + 381, + 389, + 417, + 404 + ], + "spans": [ + { + "bbox": [ + 381, + 389, + 417, + 404 + ], + "type": "image", + "image_path": "718c3957d05a8fe75a0f23ddf41b912c9155eee4d46ea66b0699d27c88ebebab.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 390, + 552, + 403 + ], + "lines": [ + { + "bbox": [ + 419, + 390, + 552, + 403 + ], + "spans": [ + { + "bbox": [ + 419, + 390, + 552, + 403 + ], + "type": "text", + "content": "Added by authors for publication" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 152, + 447, + 200, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 447, + 200, + 459 + ], + "spans": [ + { + "bbox": [ + 152, + 447, + 200, + 459 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 53, + 474, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 474, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 474, + 297, + 715 + ], + "type": "text", + "content": "Diffusion models are highly effective at generating high-quality images but pose risks, such as the unintentional generation of NSFW (not safe for work) content. Although various techniques have been proposed to mitigate unwanted influences in diffusion models while preserving overall performance, achieving a balance between these goals remains challenging. In this work, we introduce EraseDiff, an algorithm designed to preserve the utility of the diffusion model on retained data while removing the unwanted information associated with the data to be forgotten. Our approach formulates this task as a constrained optimization problem using the value function, resulting in a natural first-order algorithm for solving the optimization problem. By altering the generative process to deviate away from the groundtruth denoising trajectory, we update parameters for preservation while controlling constraint reduction to ensure effective erasure, striking an optimal trade-off. Extensive experiments and thorough comparisons with state-of-the-art algorithms demonstrate that EraseDiff effectively preserves the model's utility, efficacy, and efficiency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 448, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 472 + ], + "type": "text", + "content": "WARNING: This paper contains sexually explicit imagery that may be offensive in nature." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 500, + 395, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 395, + 511 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 395, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": "Diffusion Models [30, 49, 60] are now the method of choice in deep generative models, owing to their high-quality output, stability, and ease of training procedure. This has facilitated their successful integration into commercial applications such as midjourney. Unfortunately, the ease of use associated with diffusion models brings forth significant privacy risks. Studies have shown that these models can memorize and regenerate individual images from their training datasets [10, 58, 59]. Beyond privacy, diffusion models are susceptible to misuse and can generate NSFW digital content [48, 52, 54]. In this context, individuals whose images are used for training might request the removal of their private data. In particular, data protection regulations like the European Union General Data Protection Regulation (GDPR) [64] and the California Consumer Privacy Act (CCPA) [24] grant users the right to be" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28263" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": "forgotten, obligating companies to expunge data pertaining to a user upon receiving a request for deletion. These legal provisions grant data owners the right to remove their data from trained models and eliminate its influence on said models [4, 11, 23, 25, 44, 56, 61, 62, 69]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "type": "text", + "content": "A straightforward solution is to retrain the model from scratch after excluding the data that needs to be forgotten. However, the removal of pertinent data followed by retraining diffusion models from scratch demands substantial resources and is often deemed impractical. A version of the stable diffusion model trained on subsets of the LAION-5B dataset [55] costs approximately 150,000 GPU hours with 256 A100 GPUs1. Existing research on erasing unwanted influence has primarily focused on classification tasks [4, 7, 11, 22, 23, 25, 33, 44, 56, 68]. Despite substantial progress, prior methods developed in classification are observed to be ineffective for generation tasks [16]. Consequently, there is a pressing need for the development of methods capable of scrubbing data from diffusion models without necessitating complete retraining." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 396 + ], + "type": "text", + "content": "Recently, a handful of studies [6, 16, 20, 21, 27, 28, 36, 41, 71] target unlearning in diffusion models, with a primary focus on the text-to-image models [20, 21, 71]. Broadly, these methods aim to achieve two main objectives: erasing data influence and preserving overall model performance. However, as demonstrated by Bui et al. [5], balancing this trade-off remains challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 397, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 295, + 684 + ], + "type": "text", + "content": "In this work, we propose EraseDiff, an algorithm tailored to balance the overall performance of diffusion models with the erasure of undesirable information. Drawing inspiration from optimization-based meta-learning algorithms [18, 42] that enable fast adaptation to new learning tasks, we formulate this challenge as a bi-level optimization problem, where the \"inner\" optimization focus on erasing undesirable influence and the \"outer\" objective seeks to preserve model performance. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively. However, this nested optimization can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with non-convex functions, making it difficult to achieve a stable solution [39]. Therefore, we further reformulate the problem as a constrained optimization problem using the value function [40, 46, 70], which facilitates a natural first-order solution [39], allows us to optimize preservation and erasure in a unified manner. This approach achieves a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. We benchmark EraseDiff on various scenarios, encompassing unlearning of classes on CIFAR-10 [35] with Denoising Diffusion Probabilistic Models (DDPM) [30], classes on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "Imagenette [31] and concepts on the I2P dataset [54] with stable diffusion. Our empirical findings show that EraseDiff is " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "11 \\times" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": " faster than Heng and Soh's method [28] and " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": " faster than Fan's method [16] when forgetting on DDPM while achieving better unlearning results across several metrics. The results demonstrate that EraseDiff is capable of effectively erasing data influence in diffusion models, ranging from specific classes to the concept of nudity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 392, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 392, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 392, + 193 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "text", + "content": "In this section, we outline the components of the models we evaluate, including DDPM and latent diffusion models [49]. Throughout the paper, we denote scalars, and vectors/matrices by lowercase and bold symbols, respectively (e.g., " + }, + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 200, + 553, + 262 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": "DDPM. (1) Diffusion: DDPM gradually diffuses the data distribution " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d\\ni \\mathbf{x}_0\\sim q(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " into the standard Gaussian distribution " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d\\ni \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " time steps, i.e., " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t|\\mathbf{x}_{t - 1}) = \\mathcal{N}(\\mathbf{x}_t;\\sqrt{\\alpha_t}\\mathbf{x}_{t - 1},(1 - \\alpha_t)\\mathbf{I}_d)" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " are the pre-defined variance schedule. The diffusion takes the form " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i = 1}^{t}\\alpha_{i}" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ". (2) Training: A model " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\pmb {\\theta}\\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " is trained to learn the reverse process " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "p_{\\pmb{\\theta}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)\\approx q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\sim q(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " and time step " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "t\\in [1,T]" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ", the simplified training objective is to minimize the distance between " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " and the predicted " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\| \\epsilon -\\epsilon_{\\pmb{\\theta}}(\\mathbf{x}_t,t)\\|" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ". (3) Sampling: after training the model, we could obtain the learnable backward distribution " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "p_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_{t}) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\pmb{\\mu}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t),\\pmb{\\Sigma}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t))" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t) = \\frac{1}{\\sqrt{\\alpha_{t}}} (\\mathbf{x}_{t} - \\frac{\\beta_{t}}{\\sqrt{1 - \\alpha_{t}}}\\epsilon_{\\pmb{\\theta}}(\\mathbf{x}_{t},t))" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t},t) = \\frac{(1 - \\bar{\\alpha}_{t - 1})\\beta_{t}}{1 - \\bar{\\alpha}_{t}}" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ". Then, given " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " could be obtained via sampling from " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "p_{\\pmb{\\theta}^{*}}(\\mathbf{x}_{t - 1}|\\mathbf{x}_{t})" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 277, + 553, + 498 + ], + "type": "text", + "content": " step by step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "text", + "content": "Latent diffusion models. Latent diffusion models apply the diffusion models in the latent space " + }, + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "text", + "content": " of a pre-trained variational autoencoder. The noise would be added to " + }, + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\varepsilon(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "text", + "content": ", instead of the data " + }, + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 513, + 554, + 598 + ], + "type": "text", + "content": ", and the denoised output would be transformed to image space with the decoder. Besides, text embeddings generated by models like CLIP are used as conditioning inputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 609, + 437, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 437, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 437, + 623 + ], + "type": "text", + "content": "3. Diffusion Unlearning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{x}_i, c_i\\}_i^N" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": " be a dataset of images " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": " associated with label " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": " representing the class. " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{1, \\dots, C\\}" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": " denotes the label space where " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": " is the total number of classes and " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "c_i \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": ". We split the training data " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": " into the forgetting data " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f \\subset \\mathcal{D}" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": " and its complement, remaining data " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r = \\mathcal{D} \\setminus \\mathcal{D}_f" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": ". The forgetting data has label space " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_f \\subseteq \\mathcal{C}" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": ", and the remaining label space is denoted as " + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r = \\mathcal{C} \\setminus \\mathcal{C}_f" + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 263, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 263, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 263, + 713 + ], + "type": "text", + "content": "1https://stablediffusion.gitbook.io/overview/stable-diffusion-overview/technology/training-procedures" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28264" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 164, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 164, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 164, + 85 + ], + "type": "text", + "content": "3.1. Training objective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": "Our goal is to scrub the information about " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": " carried by the diffusion models while maintaining the model utility over the remaining data " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": ". To achieve this, we adopt different training objectives for " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": " as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "text", + "content": "For the remaining data " + }, + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "text", + "content": ", we fine-tune the diffusion models with the original objective:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 166, + 295, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 166, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 58, + 166, + 295, + 191 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\in \\mathcal {N} \\left(\\mathbf {0}, \\mathbf {I} _ {d}\\right), \\left(\\mathbf {x} _ {0}, c\\right) \\sim \\mathcal {D} _ {r} \\times \\mathcal {C} _ {r}} [ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {t} | c\\right) \\| ^ {\\frac {2}{2}} ], \\tag {1}", + "image_path": "731c2ef941deb0b48655349b7314e25041a36ff924c0ef64c1022828af59527f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\epsilon}" + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "content": ". For the forgetting data " + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "content": ", we aim to let the models fail to generate meaningful images corresponding to " + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_f" + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "content": " and thus propose:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 298, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 298, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 298, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}) = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\in \\mathcal {N} \\left(\\mathbf {0}, \\mathbf {I} _ {d}\\right), \\left(\\mathbf {x} _ {0}, c\\right) \\sim \\mathcal {D} _ {f} \\times \\mathcal {C} _ {f}} \\left[ \\| \\boldsymbol {\\epsilon} _ {f} - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {t} | c\\right) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "fe762d5229d8f452e8db1776d5d1a8ffdec756af630d8461c1f1b437db4afd6a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\epsilon_{f} = \\epsilon_{\\theta}(\\mathbf{x}_{t}|c_{m})" + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "inline_equation", + "content": "c_{m}\\neq c" + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "text", + "content": " so that the denoised image " + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "text", + "content": " is not related to the forgetting class/concept " + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "text", + "content": " [16, 28]. With this, we hinder the approximator " + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "text", + "content": " to guide the denoising process to obtain meaningful examples for the forgetting data example " + }, + { + "bbox": [ + 55, + 272, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\sim \\mathcal{D}_f" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 332, + 295, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 295, + 355 + ], + "type": "text", + "content": "To erase the undesirable influence of " + }, + { + "bbox": [ + 55, + 332, + 295, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 55, + 332, + 295, + 355 + ], + "type": "text", + "content": " and preserve the overall performance, it is common to form" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 361, + 295, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 361, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 121, + 361, + 295, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) + \\lambda \\mathcal {L} _ {f} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}), \\tag {3}", + "image_path": "8b88bbdb55d724f62b1e0f8fe8a9f5b1ed693dfa484cbcdfadd8d085df2cfdcc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 380, + 296, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 476 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 380, + 296, + 476 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 55, + 380, + 296, + 476 + ], + "type": "text", + "content": " as the optimization objective (see for example [16]). However, training could be hindered due to the conflicting gradients between the erasing and preservation objectives, preventing a balanced trade-off between erasure and preservation [38]. To address this, rather than scalarizing the two objectives, we consider a framework based on optimization-based meta-learning algorithm [47] that allows iteratively updates to optimize each objective:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 482, + 206, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 482, + 206, + 495 + ], + "spans": [ + { + "bbox": [ + 136, + 482, + 206, + 495 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r})", + "image_path": "1c19599b9bc4be422163d20a824c19c2e98aef52a7958258e38d38ff0a7cfaaf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 498, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 498, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 113, + 498, + 295, + 517 + ], + "type": "text", + "content": "s.t. " + }, + { + "bbox": [ + 113, + 498, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\pmb {\\theta}\\in \\underset {\\phi}{\\arg \\min}\\mathcal{L}_f(\\pmb {\\phi};\\mathcal{D}_f)" + }, + { + "bbox": [ + 113, + 498, + 295, + 517 + ], + "type": "text", + "content": " (4)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "text", + "content": "where the outer objective minimizes the remaining loss " + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "text", + "content": " (i.e., preserving model utility), the inner optimization minimizes the forgetting loss " + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "text", + "content": " (i.e., erasing) with initialization " + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n\\ni \\phi_{\\mathrm{init}} = \\pmb{\\theta}" + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "text", + "content": ", the inner optimization on " + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 522, + 295, + 629 + ], + "type": "text", + "content": " aims to minimize the forgetting data influence, with the goal of achieving effective erasure while preserving model utility. The outer objective and inner optimization are interdependent, iterating between preservation and erasure to balance the trade-off effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "While the above framework allows for iterative updates to address the conflicting objectives of erasure and preservation, it still relies on nested optimization, which can be challenging to optimize efficiently. The inner optimization may converge to a saddle point or struggle with nonconvex functions, making it difficult to achieve a stable solution [39]. To further streamline the optimization process," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "we adopt a value function approach [40, 46, 70] that reformulates the problem as a single-constrained optimization:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 114, + 443, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 114, + 443, + 126 + ], + "spans": [ + { + "bbox": [ + 373, + 114, + 443, + 126 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r})", + "image_path": "2c2556e462c64634fa3aacce9c4618ae40b6b6a4256e52d6c1bb1571434be0b1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 350, + 129, + 553, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 129, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 350, + 129, + 553, + 142 + ], + "type": "text", + "content": "s.t. " + }, + { + "bbox": [ + 350, + 129, + 553, + 142 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f(\\pmb {\\theta};\\mathcal{D}_f) - \\min_{\\phi}\\mathcal{L}_f(\\phi ;\\mathcal{D}_f)\\leq 0," + }, + { + "bbox": [ + 350, + 129, + 553, + 142 + ], + "type": "text", + "content": " (5)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "text", + "content": " is initialized at " + }, + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "text", + "content": ", leverages the value function to encapsulate the influence of data erasure directly as a constraint on " + }, + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 313, + 147, + 555, + 230 + ], + "type": "text", + "content": ". This avoids the need for a nested loop by capturing the forgetting objective as a constraint and provides a natural first-order solution, as the constrained formulation allows us to optimize preservation and erasure in a unified manner." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 237, + 376, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 237, + 376, + 248 + ], + "spans": [ + { + "bbox": [ + 314, + 237, + 376, + 248 + ], + "type": "text", + "content": "3.2. Solution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": "To solve Eq. (5), let us first denote " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}) \\coloneqq \\mathcal{L}_f(\\pmb{\\theta};\\mathcal{D}_f) - \\min_{\\phi}\\mathcal{L}_f(\\phi ;\\mathcal{D}_f)" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": ". Our goal is to erase undesirable influence while preserving the overall model performance, hence the update vector " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " for updating the model should aid in minimizing " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r(\\pmb{\\theta};\\mathcal{D}_r)" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta})" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " simultaneously. In other words, suppose that the current solution for Eq. (5) is " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_t" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": ", we aim to update " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{t + 1} = \\pmb{\\theta}_t - \\eta \\pmb{\\delta}_t" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " is sufficiently small, so that " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r(\\pmb{\\theta}_{t + 1};\\mathcal{D}_r)" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " decreases (i.e., preserve model utility) and " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}_{t + 1})" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " decreases (i.e., erasure). To this end, we aim to find the update vector " + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 313, + 254, + 555, + 374 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 363, + 379, + 527, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 379, + 527, + 401 + ], + "spans": [ + { + "bbox": [ + 363, + 379, + 527, + 401 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\delta} _ {t} \\in \\frac {1}{2} \\operatorname {a r g m i n} _ {\\boldsymbol {\\delta}} \\left\\| \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {r} \\left(\\boldsymbol {\\theta} _ {t}; \\mathcal {D} _ {r}\\right) - \\boldsymbol {\\delta} \\right\\| _ {2} ^ {2},", + "image_path": "63db8eae1141219878567c5d33df2c897a701ee47d619d66c6d648311d6e11d6.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 341, + 403, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 403, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 341, + 403, + 553, + 417 + ], + "type": "text", + "content": "s.t. " + }, + { + "bbox": [ + 341, + 403, + 553, + 417 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_t)^\\top \\pmb{\\delta} \\geq a_t > 0." + }, + { + "bbox": [ + 341, + 403, + 553, + 417 + ], + "type": "text", + "content": " (6)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": "This will ensure that the update " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": " is close to " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\nabla_\\theta \\mathcal{L}_r(\\pmb{\\theta}_t; \\mathcal{D}_r)" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": " and decreases " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}_t)" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": " until it reaches stationary. Because " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}_{t+1}) - g(\\pmb{\\theta}_t) \\approx -\\eta \\nabla_\\theta g(\\pmb{\\theta}_t)^\\top \\delta \\leq -\\eta a_t < 0" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": " for some scalar " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "a_t > 0" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": ", we can ensure that " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}_{t+1}) < g(\\pmb{\\theta}_t)" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": " for small step size " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\eta > 0" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": ". This means that the update " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": " can ensure to minimize " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f(\\pmb{\\theta}; \\mathcal{D}_f)" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": " as long as it does not conflict with descent of " + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r(\\pmb{\\theta}; \\mathcal{D}_r)" + }, + { + "bbox": [ + 313, + 422, + 554, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 508, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 532 + ], + "type": "text", + "content": "To find the solution to the optimization problem in Eq. (6), the following theorem is developed:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 536, + 555, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 580 + ], + "type": "text", + "content": "Theorem 3.1. The optimal solution of the optimization problem in Eq. (6) is " + }, + { + "bbox": [ + 313, + 536, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\delta^{*} = \\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r}) + \\lambda_{t}\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})" + }, + { + "bbox": [ + 313, + 536, + 555, + 580 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 536, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}" + }, + { + "bbox": [ + 313, + 536, + 555, + 580 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We provide the proof in §7 in the Appendix. This provides the solution to the optimization problem by constructing the update vector " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " to balance two competing objectives. The variable " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " adjusts the weight of the forgetting objective, ensuring that the update vector " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " decreases the remaining loss without violating the erasure goal, and achieves the dual goals of maintaining utility and achieving erasing. In practice, we can choose " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "a_{t} = \\eta \\| \\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": ", and we start from " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi^0 = \\pmb{\\theta}_t" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " and use gradient descend in " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " steps with the learning rate " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\xi > 0" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " to reach " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi^K" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi^{k + 1} = \\phi^k - \\xi \\nabla_\\phi \\mathcal{L}_f(\\phi^k; \\mathcal{D}_f)" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "k = 0, \\dots, K - 1" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28265" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 95 + ], + "type": "text", + "content": "Algorithm 1 EraseDiff: Erasing undesirable influence in diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "content": "Input: Well-trained model with parameters " + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "content": ", forgetting data " + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "content": " and remaining data " + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "content": ", outer iteration number " + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "content": " and inner iteration number " + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 56, + 97, + 295, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 133, + 252, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 252, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 252, + 144 + ], + "type": "text", + "content": "Output: Parameters " + }, + { + "bbox": [ + 56, + 133, + 252, + 144 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 56, + 133, + 252, + 144 + ], + "type": "text", + "content": " for the scrubbed model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 146, + 294, + 255 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 62, + 146, + 163, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 146, + 163, + 156 + ], + "spans": [ + { + "bbox": [ + 62, + 146, + 163, + 156 + ], + "type": "text", + "content": "1: for iteration " + }, + { + "bbox": [ + 62, + 146, + 163, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 62, + 146, + 163, + 156 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 62, + 146, + 163, + 156 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 62, + 146, + 163, + 156 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 157, + 122, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 157, + 122, + 168 + ], + "spans": [ + { + "bbox": [ + 62, + 157, + 122, + 168 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 62, + 157, + 122, + 168 + ], + "type": "inline_equation", + "content": "\\phi^0 = \\pmb {\\theta}_t" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "text", + "content": "3: Get " + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\phi^K" + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "text", + "content": " steps of gradient descent on " + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f(\\phi; \\mathcal{D}_f)" + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "text", + "content": " starting from " + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\phi^0" + }, + { + "bbox": [ + 62, + 169, + 294, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 194, + 248, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 248, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 248, + 205 + ], + "type": "text", + "content": "4: Set " + }, + { + "bbox": [ + 62, + 194, + 248, + 205 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}_t) = \\mathcal{L}_f(\\pmb{\\theta}_t; \\mathcal{D}_f) - \\mathcal{L}_f(\\phi^K; \\mathcal{D}_f)" + }, + { + "bbox": [ + 62, + 194, + 248, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 205, + 294, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 229 + ], + "type": "text", + "content": "5: Update the model: " + }, + { + "bbox": [ + 62, + 205, + 294, + 229 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{t + 1} = \\pmb{\\theta}_t - \\eta (\\nabla_{\\pmb{\\theta}_t}\\mathcal{L}_r(\\pmb {\\theta}_t;\\mathcal{D}_r) + \\lambda_t\\nabla_{\\pmb{\\theta}_t}g(\\pmb {\\theta}_t;\\phi^K))" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 229, + 272, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 229, + 272, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 229, + 272, + 247 + ], + "type": "text", + "content": "6: where " + }, + { + "bbox": [ + 62, + 229, + 272, + 247 + ], + "type": "inline_equation", + "content": "\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})^{T}\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\pmb{\\theta}}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}" + }, + { + "bbox": [ + 62, + 229, + 272, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 247, + 106, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 106, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 106, + 255 + ], + "type": "text", + "content": "7: end for" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 280, + 118, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 118, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 118, + 293 + ], + "type": "text", + "content": "3.3. Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 298, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 295, + 323 + ], + "type": "text", + "content": "We can characterize the solution of our algorithm as follows and the proof can be found in §7 in the Appendix:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 331, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 369 + ], + "type": "text", + "content": "Theorem 3.2 (Pareto optimality). The stationary point obtained by our algorithm is Pareto optimal of the problem " + }, + { + "bbox": [ + 55, + 331, + 296, + 369 + ], + "type": "inline_equation", + "content": "\\min_{\\pmb{\\theta}}[\\mathcal{L}_r(\\pmb{\\theta};\\mathcal{D}_r),\\mathcal{L}_f(\\pmb{\\theta};\\mathcal{D}_f)]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 377, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 296, + 437 + ], + "type": "text", + "content": "This asserts that the solution obtained by our algorithm is Pareto optimal for the problem of minimizing both objectives, which implies that the solution obtained by the algorithm ensures a balanced trade-off between preserving model utility and erasing undesirable influences." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": "We further take DDPM with CIFAR-10 when forgetting the 'airplane' as an example to show that our proposed method helps alleviate the gradient conflict which prevents a balanced trade-off between erasure and preservation. Fig. 2 presents the cosine similarity between the update vector " + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": " and the gradient " + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "g_{r} = \\nabla_{\\theta}\\mathcal{L}_{r}(\\theta ;\\mathcal{D}_{r})" + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": " for preservation, and the cosine similarity between the update vector " + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": " and the gradient " + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "g_{f} = \\nabla \\mathcal{L}_{f}(\\theta ;\\mathcal{D}_{f})" + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": " for erasing. The cosine similarity represents the alignment between the update vector and the gradients associated with the preservation and erasure. Higher positive values indicate alignment, meaning that the update vector " + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": " is directed similarly to the respective gradient, whereas negative values indicate misalignment or conflict. In particular, negative values suggest a high degree of opposition between the update vector and the respective gradient, which can signify competing objectives between preservation and erasure during the optimization process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "MOO (Multi-Objective Optimization) denotes the naive integration of erasing and preservation as stated in Eq. (3). For the vanilla MOO, Fig. 2 shows a clear alternating pattern in cosine similarity values between the update vector " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and the gradients " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "g_{r}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "g_{f}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ". Specifically, when the cosine" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 322, + 76, + 548, + 227 + ], + "blocks": [ + { + "bbox": [ + 322, + 76, + 548, + 227 + ], + "lines": [ + { + "bbox": [ + 322, + 76, + 548, + 227 + ], + "spans": [ + { + "bbox": [ + 322, + 76, + 548, + 227 + ], + "type": "image", + "image_path": "b5c502d8fe4a19dd98c2fc18080b49c434130f3803283a2c87acbc90e92fab90.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 325, + 231, + 548, + 382 + ], + "blocks": [ + { + "bbox": [ + 325, + 231, + 548, + 382 + ], + "lines": [ + { + "bbox": [ + 325, + 231, + 548, + 382 + ], + "spans": [ + { + "bbox": [ + 325, + 231, + 548, + 382 + ], + "type": "image", + "image_path": "10377f078ff0c8c707ad539046dee7ac2a2c2efa1cef9481f0d73191f5482acc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "text", + "content": "Figure 2. Top to bottom: cosine similarity between the update vector " + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "text", + "content": " and the preservation gradient " + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "inline_equation", + "content": "g_{r}" + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "text", + "content": ", followed by the cosine similarity between " + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "text", + "content": " and the erasing gradient " + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "inline_equation", + "content": "g_{f}" + }, + { + "bbox": [ + 313, + 397, + 554, + 464 + ], + "type": "text", + "content": ". Positive values indicate alignment, while negative values suggest conflict. This visualization illustrates how well the update vector aligns with the objectives of preservation and erasure over successive iterations." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": "similarity between " + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "inline_equation", + "content": "g_{r}" + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": " is greater than 0, the similarity between " + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "inline_equation", + "content": "g_{f}" + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": " tends to be less than 0, and vice versa. This pattern suggests that MOO experiences gradient conflict, as it cannot effectively balance the two objectives of preservation and erasure, preventing MOO from achieving a harmonious update that supports both goals simultaneously. In contrast, EraseDiff mostly shows positive cosine similarity values between the update vector " + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": " and both the preservation gradient " + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "inline_equation", + "content": "g_{r}" + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": " and the erasing gradient " + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "inline_equation", + "content": "g_{f}" + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": ". This indicates that EraseDiff aligns the update direction with both objectives, suggesting it manages to avoid significant gradient conflict. By maintaining positive alignment, EraseDiff appears to balance preservation and erasure more effectively, leading to better cooperation between objectives." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 668, + 400, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 668, + 400, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 668, + 400, + 681 + ], + "type": "text", + "content": "4. Related Work" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Memorization in generative models. Privacy of generative models has been studied extensively for GANs [17, 43," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28266" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "66] and generative language models [8, 9, 32, 63]. These generative models often risk replicating from their training data. Recently, several studies [10, 58, 59, 65] investigated these data replication behaviors in diffusion models, raising concerns about the privacy and copyright issues. Possible mitigation strategies are deduplicating and randomizing conditional information [58, 59], or training models with differential privacy (DP) [1, 13-15]. However, leveraging DP-SGD [1] may cause training to diverge [10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 296, + 422 + ], + "type": "text", + "content": "Malicious misuse. Diffusion models usually use training data from varied open sources and when such unfiltered data is employed, there is a risk of it being tainted [12] or manipulated [48], resulting in inappropriate generation [54]. They also risk the imitation of copyrighted content, e.g., mimicking the artistic style [20, 57]. To counter inappropriate generation, data censoring [3, 19, 45, 53] where excluding black-listed images before training, and safety guidance where diffusion models will be updated away from the inappropriate/undesired concept [20, 54] are proposed. [34] uses human feedback to annotate the generations for getting soft tokens that quantify the harmfulness of the images, then fine-tunes the model using self-distillation to achieve harmony between preservation and erasure. Shan et al. [57] propose protecting artistic style by adding barely perceptible perturbations to the artworks before public release. Yet, Rando et al. [48] argue that DMs can still generate content that bypasses the filter. Chen et al. [12] highlight the susceptibility of DMs to poison attacks, where target images are generated with specific triggers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": "Machine unlearning. Removing data directly involves retraining the model from scratch, which is inefficient and impractical. Thus, to reduce the computational overhead, efficient machines unlearning methods [4, 7, 11, 22, 23, 25, 33, 44, 50, 56, 61, 68] have been proposed. Several studies [5, 16, 20, 21, 27, 28, 67, 71] recently introduce unlearning in diffusion models. Most of them [20, 21, 27, 71] mainly focus on text-to-image models and high-level visual concept erasure. Heng and Soh [28] adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) from continual learning to perform unlearning effectively without access to the training data. Heng and Soh's method can be applied to a wide range of generative models, however, it needs the computation of FIM for different datasets and models, which may lead to significant computational demands. Fan et al. [16] propose a very potent unlearning algorithm called SalUn that shifts attention to important parameters w.r.t. the forgetting data. SalUn can perform effectively across image classification and generation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In this work, we introduce a simple yet effective unlearning algorithm for diffusion models by formulating the problem as a constrained optimization problem, to achieve a fine-tuned balance between preservation and targeted erasure, yielding an optimal trade-off. Below, we will show" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "that our algorithm is not only faster than Heng and Soh's method [28] and Fan's method [16], but even outperforms these methods in terms of the trade-off between the forgetting and preserving model utility." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 130, + 391, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 391, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 391, + 144 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "text", + "content": "We evaluate EraseDiff in various scenarios, including removing images with specific classes/concepts, to answer the following research questions (RQs): (i) Can typical machine unlearning methods be applied to diffusion models? (ii) Is EraseDiff able to remove the influence of " + }, + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "text", + "content": " in the diffusion models? (iii) Is EraseDiff able to preserve the model utility while removing " + }, + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "text", + "content": "? (iv) Is EraseDiff efficient in removing the data? (v) How does EraseDiff perform on the public well-trained models?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 265, + 365, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 365, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 365, + 277 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "type": "text", + "content": "Experiments are reported on CIFAR-10 [35] with DDPM, Imagenette [31] with Stable Diffusion (SD) for class-wise forgetting, I2P [54] dataset with SD for concept-wise forgetting. For all SD experiments, we use the open-source SD v1.4 [49] checkpoint as the pre-trained model. Implementation details and additional results like visualizations of generated images can be found in §8 and §9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 366, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 437 + ], + "type": "text", + "content": "Baselines. We primarily benchmark against the following baselines commonly used in machine unlearning: (i) Unscrubbed, (ii) Finetune (FT) [23], (iii) NegGrad (NG) [23], (iv) BlindSpot [61], (v) ESD [20], (vi) FMN [71], (vii) Selective Amnesia (SA) [28] and (viii) the SOTA machine unlearning algorithm SalUn [16]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "text", + "content": "Metrics. Several metrics are utilized to evaluate the algorithms: (i) Frechet Inception Distance (FID) [29]: the widely-used metric for assessing the quality of generated images. (ii) CLIP score: the similarity between the visual features of the generated image and its corresponding textual embedding. (iii) " + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "inline_equation", + "content": "P_{\\psi}(\\mathbf{y} = c_f|\\mathbf{x}_f)" + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "text", + "content": " [28]: the classification rate of a pre-trained classifier " + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "inline_equation", + "content": "P_{\\psi}(\\mathbf{y}|\\mathbf{x})" + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "text", + "content": ", with a ResNet architecture [26] used to classify generated images conditioned on the forgetting classes. A lower classification value indicates superior unlearning performance. (iv) Precision and Recall: A low FID may indicate high precision (realistic images) but low recall (small variations) [37, 51]. Kynkänniemi et al. [37] shows that generative models claim to optimize FID (high fidelity) but always sacrifice variation (low diversity). Hence, we include metric precision (fidelity) and recall (diversity) to express the quality of the generated samples, to provide explicit visibility of the tradeoff between sample quality and variety." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 660, + 422, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 422, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 422, + 672 + ], + "type": "text", + "content": "5.2. Results on DDPM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Following SA, we aim to forget the 'airplane' class on CIFAR-10. Here, we replace " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{f} = \\epsilon_{\\theta}(\\mathbf{x}_{t}|c_{m})" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " like random labelling used in [16] where " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "c_{m}\\neq c" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28267" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Results on CIFAR10 with DDPM when forgetting the 'airplane' class. " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "P_{\\psi}(\\mathbf{y} = c_f|\\mathbf{x}_f)" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " indicate the probability of the forgotten class (i.e., the effectiveness of erasing). Precision and Recall demonstrate the fidelity and diversity [37, 51], and FID scores are computed between the generated 45K images and the corresponding ground truth images with the same labels from " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " (i.e., preserving model utility). SA excels in class-wise forgetting but struggles to perform concept-wise forgetting as shown in Fig. 3 and Tab. 2. The best and the second best are highlighted in blue and orange, respectively." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 134, + 547, + 215 + ], + "blocks": [ + { + "bbox": [ + 63, + 134, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 63, + 134, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 63, + 134, + 547, + 215 + ], + "type": "table", + "html": "
UnscrubbedFT [23]NG [23]BlindSpot [61]SA [28]SalUn [16]EraseDiffrlEraseDiffnoise
FID ↓9.638.2176.739.128.199.168.667.61
Precision (fidelity) ↑0.400.430.080.410.430.410.430.43
Recall (diversity) ↑0.790.770.610.780.750.760.770.72
Pψ(y = cf|xf)↓0.970.960.610.900.060.070.240.22
", + "image_path": "d18d88d6b9c7292a953de00d484b9f742099c5f31852be764c642abe6e2b3b0e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 222, + 233, + 345 + ], + "blocks": [ + { + "bbox": [ + 61, + 222, + 233, + 345 + ], + "lines": [ + { + "bbox": [ + 61, + 222, + 233, + 345 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 233, + 345 + ], + "type": "image", + "image_path": "1736ccbff23f646147064b068bc6a26b1f06e1358197e68507ddc6d68ddf9dfb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 354, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 555, + 378 + ], + "type": "text", + "content": "Figure 3. Quantity of nudity content detected using the NudeNet classifier from I2P data. Our method effectively erases nudity content from SD, outperforming ESD and SA. Note that Fig. 4 and Tab. 2 together presents the trade-off between erasing and preservation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 222, + 338, + 346 + ], + "blocks": [ + { + "bbox": [ + 238, + 222, + 338, + 346 + ], + "lines": [ + { + "bbox": [ + 238, + 222, + 338, + 346 + ], + "spans": [ + { + "bbox": [ + 238, + 222, + 338, + 346 + ], + "type": "image", + "image_path": "b80ded473f9a468bd3de678469e126074cd2ade9028534305a8083ef3cc4e020.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 396, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 452 + ], + "type": "text", + "content": "Table 2. Evaluation of " + }, + { + "bbox": [ + 55, + 396, + 295, + 452 + ], + "type": "inline_equation", + "content": "30\\mathrm{K}" + }, + { + "bbox": [ + 55, + 396, + 295, + 452 + ], + "type": "text", + "content": " generated images by SD when erasing 'nudity'. The FID score is measured compared to validation data, while the CLIP similarity score evaluates the alignment between generated images and the corresponding prompts. The best and the second best are highlighted in blue and orange, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 341, + 223, + 443, + 346 + ], + "blocks": [ + { + "bbox": [ + 341, + 223, + 443, + 346 + ], + "lines": [ + { + "bbox": [ + 341, + 223, + 443, + 346 + ], + "spans": [ + { + "bbox": [ + 341, + 223, + 443, + 346 + ], + "type": "image", + "image_path": "f3f99bc74074be9ed65566e4bf084f34cbffaddd925b8a8920c7c8b7a7a49359.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 446, + 223, + 549, + 346 + ], + "blocks": [ + { + "bbox": [ + 446, + 223, + 549, + 346 + ], + "lines": [ + { + "bbox": [ + 446, + 223, + 549, + 346 + ], + "spans": [ + { + "bbox": [ + 446, + 223, + 549, + 346 + ], + "type": "image", + "image_path": "0369b42997825dc6c18aed22fc3078892a901db3cceabcf224f59d0493a1c38f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 67, + 460, + 285, + 513 + ], + "blocks": [ + { + "bbox": [ + 67, + 460, + 285, + 513 + ], + "lines": [ + { + "bbox": [ + 67, + 460, + 285, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 285, + 513 + ], + "type": "table", + "html": "
ESD [20]SA [28]SalUn [16]EraseDiff
FID ↓15.7625.5825.0617.01
CLIP ↑30.3331.0328.9130.58
", + "image_path": "74e5044fb53d52704a9f0e3f36a039e22404aad84e49b5b2a16c8efa8ac3b36c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": "denoted as " + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\text{EraseDiff}_{\\text{rl}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": ". We also try to use " + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\epsilon_f = \\mathcal{U}(\\mathbf{0}, \\mathbf{I}_d)" + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": " like SA, denoted as " + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\text{EraseDiff}_{\\text{noise}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": ". Note that the choice of replacement for forgotten classes is flexible and is not the primary focus of this work. For further discussion on the choice of substitution strategies, please refer to [6, 41]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Results are presented in Tab. 1. Firstly, from Tab. 1, we can conclude that traditional machine unlearning methods designed for image classification or regression tasks fall short in effectively performing forgetting for DDPM. Finetune and BlindSpot suffer from under-forgetting (i.e., the generated image quality is good but the probability of generated images belonging to the forgetting class approaching the value of the unscrubbed model), and NegGrad suffers from over-forgetting (the probability of generated images belonging to the forgetting class is decreased compared to" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "type": "text", + "content": "that of the unscrubbed model but the generated image quality drops significantly)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 602 + ], + "type": "text", + "content": "Then, comparing SA and SalUn's unlearning methods, SA achieves an FID score of 8.19 but sacrifices variation (decreased recall). Also, note that SA introduces excessive computational resource requirements and time consumption [28, 72]. Note that the FID scores of SA, SalUn, and EraseDiff decrease compared with the generated images from the original models; the quality of the generated images experiences a slight improvement. However, there is a decrease in recall (diversity), which can be attributed to the scrubbed models being fine-tuned over " + }, + { + "bbox": [ + 313, + 423, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 313, + 423, + 555, + 602 + ], + "type": "text", + "content": ", suggesting a tendency towards overfitting. Regarding forgetting, SalUn achieves a smaller probability of the generated images classified as the forgetting class than ours; yet, the FID score is larger than ours, and images generated by EraseDiffr1 present better diversity and fidelity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 463, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 463, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 463, + 623 + ], + "type": "text", + "content": "5.3. Results on Stable Diffusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "content": "We apply EraseDiff to perform class-wise forgetting from Imagenette and erase the 'nudity' concept with SD v1.4. When forgetting 'nudity', we have no access to the training data; instead, we generate " + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\sim 400" + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "content": " images with the prompts " + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "inline_equation", + "content": "c_{f} = \\{\\text{'nudity', 'naked', 'erotic', 'sexual'}\\}" + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Forget nudity. 4703 images are generated using I2P prompts, and 1K images are generated using the prompts" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28268" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 36, + 604, + 550 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 94, + 93 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 94, + 93 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 94, + 93 + ], + "type": "text", + "content": "Table 3. P forgetting" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 36, + 604, + 550 + ], + "lines": [ + { + "bbox": [ + 93, + 36, + 604, + 550 + ], + "spans": [ + { + "bbox": [ + 93, + 36, + 604, + 550 + ], + "type": "image", + "image_path": "77f150cff340b7bd1a389bc17e00771d50359d87ce9cbe07b6d9838e1b9760e2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "type": "text", + "content": "{nudity, nak, sexual} The quality of nu-dity content is detected using the NudeNet classifier [2]. In Fig. 3, the number in the y-axis denotes the number of exposed body parts generated by the SD v1.4 model. Fig. 3 presents the percentage change in exposed body parts w.r.t. SD v1.4. In §9, we provide the number of exposed body parts counted in all generated images with different thresholds. Here, our algorithm replaces " + }, + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon_{f}" + }, + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_i|c_m)" + }, + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 55, + 521, + 295, + 712 + ], + "type": "text", + "content": " is 'a photo of poker'. We can find that, EraseD-iff reduces the amount of nu-dity content compared to SD v1.4, ESD, and SA, particularly on sensitive content like Female/Male interests and female/Male Genitalia. While SalUn excels a 'forgetting', our algorithm demonstrates a significant improvement in the quality of generated images as shown in Tab. 2. Tab. 2 presented results evaluating the utility of scrubbed models. The FID and CLIP scores are" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 725, + 602, + 792 + ], + "blocks": [ + { + "bbox": [ + 55, + 725, + 602, + 792 + ], + "lines": [ + { + "bbox": [ + 55, + 725, + 602, + 792 + ], + "spans": [ + { + "bbox": [ + 55, + 725, + 602, + 792 + ], + "type": "image", + "image_path": "e0d6c7c35f43ceedc354c7e64bdcc9066f2fba320467025a689a34fd5d17e7b5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 102, + 288, + 163 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "type": "text", + "content": "Table 4. Computational overhead. Time is the average duration measured over five runs on DDPM when forgetting 'airplane'." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 102, + 288, + 163 + ], + "lines": [ + { + "bbox": [ + 64, + 102, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 288, + 163 + ], + "type": "table", + "html": "
Memory (MiB)Time (min.)Complexity
SA3352.3140.00O(n2)
SalUn4336.228.17O(n)
EraseDiff3360.312.70O(n)
", + "image_path": "803feed9f157b6661ac68f6b1a065a08c3150f18b57b048210ab86218fef10ec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 183, + 296, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 183, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 54, + 183, + 296, + 292 + ], + "type": "text", + "content": "approach, clearly indicate the advantages of our method. We also present results when improving the forgetting ability of SalUn in §9. However, note that this enhancement comes with a drop in the FID score of the generated images. Our method, while slightly better than SalUn on average across 10 classes, demonstrates a more balanced trade-off between erasing and preservation, indicating that it achieves a favorable balance in preserving fidelity while enhancing erasing performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 195, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 195, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 195, + 313 + ], + "type": "text", + "content": "5.4. Computational efficiency" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "text", + "content": "Finally, we measure the computational complexity of unlearning algorithms. The computational complexity of SA and SalUn involves two distinct stages: the computation of FIM for SA and the computation of salient weights w.r.t. " + }, + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "text", + "content": " for SalUn, and the subsequent forgetting stage for both algorithms. We consider the maximum memory usage across both stages, the metric 'Time' is exclusively associated with the duration of the forgetting stage for unlearning algorithms. Tab. 4 show that EraseDiff outperforms SA and SalUn in terms of efficiency, achieving a speed increase of " + }, + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\sim 11\\times" + }, + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "text", + "content": " than SA and " + }, + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\sim 2\\times" + }, + { + "bbox": [ + 54, + 318, + 296, + 473 + ], + "type": "text", + "content": " than SalUn. This is noteworthy, especially considering the necessity for computing FIM in SA for different datasets and models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 482, + 148, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 148, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 148, + 495 + ], + "type": "text", + "content": "5.5. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": "We further investigate the influence of the number of iterations " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": " that approximate " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\min \\mathcal{L}_f(\\phi ;\\mathcal{D}_f)" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": ", and the step size " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": " that controls the weight of forgetting and preserving model utility. Here, we replace " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{f}\\in \\mathcal{U}(\\mathbf{0},\\mathbf{I}_{d})" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": ". Note that for different hyperparameters in Fig. 5 (a), the average entropy of the classifier's output distribution given " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_f" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "H(P_{\\psi}(\\mathbf{y}|\\mathbf{x}_f)) = -\\mathbb{E}[\\sum_iP_{\\psi}(\\mathbf{y} = c_i|\\mathbf{x})\\log_eP_{\\psi}(\\mathbf{y} = c_i|\\mathbf{x})]" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": ", remains close to 2.02. This indicates that the scrubbed models become uncertain about the images conditioned on the forgetting class, effectively erasing the information about " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": ". Below, we will further demonstrate the influence on the model utility. In practice, we have " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{t} = \\max \\{0,\\frac{a_{t} - \\nabla_{\\theta}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\theta}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\} = \\max \\{0,\\eta -\\frac{\\nabla_{\\theta}g(\\pmb{\\theta}_{t})^{\\top}\\nabla_{\\theta}\\mathcal{L}_{r}(\\pmb{\\theta}_{t};\\mathcal{D}_{r})}{\\|\\nabla_{\\theta}g(\\pmb{\\theta}_{t})\\|_{2}^{2}}\\}" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": ", we can see that " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": " determines the extent to which the update direction for forgetting can deviate from that for preserving model utility. A larger " + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 56, + 499, + 297, + 714 + ], + "type": "text", + "content": " would allow for more deviation in the updating, thus pri" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 556, + 206 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 556, + 206 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 556, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 556, + 206 + ], + "type": "image", + "image_path": "bbb94cf6ab7a3076aeebc533cd9b12764298cc2921b6e7cf3f2b7874e58ba66b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 217, + 550, + 229 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 550, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 550, + 229 + ], + "type": "text", + "content": "Figure 5. (a) Ablation results. (b) Potential incomplete erasures." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": "oritizing forgetting over preserving model utility. In Fig. 5 (a), the FID score tends to increase (i.e., image quality drop) as the step size " + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": " increases, indicating that larger " + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": " leads to greater deviations from the direction that preserves the model utility. Furthermore, the number of iterations " + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": " determines how closely the approximation " + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\phi^K" + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": " will approach " + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\phi} \\mathcal{L}_f(\\phi; \\mathcal{D}_f)" + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": ". Hence, a larger number of iterations " + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": " leads to more thorough erasure, which is also supported by the results shown in Fig. 5 (a), as increasing " + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": " correlates with an increase in the FID score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 555, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 394 + ], + "type": "text", + "content": "6. Conclusion, limitations and broader impacts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "text", + "content": "In this work, we explored erasing undesirable influence in diffusion models and proposed an efficient method EraseDiff to achieve a balanced trade-off between erasing and preservation. Comprehensive experiments on diffusion models demonstrate the proposed algorithm's effectiveness in data removal, its efficacy in preserving the model utility, and its efficiency in erasure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 712 + ], + "type": "text", + "content": "However, our scrubbed model may have incomplete erasures, potentially arises due to the choice of the replacements. In Fig. 5 (b), with random labeling, generated images conditioned on the forgetting class 'tench' by our scrubbed model may preserve some characteristics similar to that close to 'tench' visually; instead, replacing the erased class 'tench' with other class such as 'dog' could improve in erasure. Besides, the scrubbed models could be biased or be misused for censorship or exploitation. This includes using technology to selectively remove or amplify NSFW content in various ways. They might also falsely assure users that NSFW content is fully removed when it isn't. To address this, advanced privacy-preserving training techniques are in demand to enhance the security and fairness of the models. Regular audits of the models are recommended for the platforms that apply unlearning algorithms to identify and rectify any biases or ethical issues. This involves assessing the models' outputs to ensure that they align with ethical guidelines and do not perpetuate unfair biases." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28270" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "content": "Mehrtash Harandi is supported by the Australian Research Council (ARC) Discovery Program DP250100262. The authors gratefully acknowledge the anonymous reviewers for their insightful feedback and valuable suggestions, which have significantly improved the quality of this work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 187, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 296, + 243 + ], + "type": "text", + "content": "[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308-318, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 244, + 295, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 244, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 295, + 266 + ], + "type": "text", + "content": "[2] P Bedapudi. Nudenet: Neural nets for nudity classification, detection and selective censoring, 2019. 7, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 267, + 295, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 267, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 62, + 267, + 295, + 311 + ], + "type": "text", + "content": "[3] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1536-1546. IEEE, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 312, + 294, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 294, + 366 + ], + "type": "text", + "content": "[4] Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159, 2021. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 368, + 294, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 368, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 368, + 294, + 411 + ], + "type": "text", + "content": "[5] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 413, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 294, + 456 + ], + "type": "text", + "content": "[6] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "type": "text", + "content": "[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480, 2015. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 555 + ], + "type": "text", + "content": "[8] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633-2650, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "type": "text", + "content": "[9] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 294, + 658 + ], + "type": "text", + "content": "[10] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 714 + ], + "type": "text", + "content": "[11] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[12] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[13] Tim Dockhorn, Tianshi Cao, Arash Vahdat, and Karsten Kreis. Differentially private diffusion models. arXiv preprint arXiv:2210.09929, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 184 + ], + "type": "text", + "content": "[14] Cynthia Dwork. Differential privacy: A survey of results. In International conference on theory and applications of models of computation, pages 1-19. Springer, 2008." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 248 + ], + "type": "text", + "content": "[15] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, New York, NY, USA, March 4-7, 2006. Proceedings 3, pages 265-284. Springer, 2006. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[16] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Dennis Wei, Eric Wong, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "text", + "content": "[17] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701–6710, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[18] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 394, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 470 + ], + "type": "text", + "content": "[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[20] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In 2023 IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "text", + "content": "[21] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "text", + "content": "[22] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "text", + "content": "[23] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309, 2020. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "text", + "content": "[24] Eric Goldman. An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[25] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28271" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "models. In Proceedings of the 37th International Conference on Machine Learning, pages 3832-3842. PMLR, 2020. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "type": "text", + "content": "[27] Alvin Heng and Harold Soh. Continual learning for forgetting in deep generative models. 2023. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "text", + "content": "[28] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "type": "text", + "content": "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 264, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 295, + 298 + ], + "type": "text", + "content": "[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 299, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 299, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 57, + 299, + 294, + 320 + ], + "type": "text", + "content": "[31] Jeremy Howard and Sylvain Gugger. Fastai: A layered api for deep learning. Information, 11(2):108, 2020. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 322, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 295, + 376 + ], + "type": "text", + "content": "[32] Matthew Jagielski, Om Thakkar, Florian Tramer, Daphne Ippolito, Katherine Lee, Nicholas Carlini, Eric Wallace, Shuang Song, Abhradeep Thakurta, Nicolas Papernot, et al. Measuring forgetting of memorized training examples. arXiv preprint arXiv:2207.00099, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 377, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 377, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 57, + 377, + 294, + 421 + ], + "type": "text", + "content": "[33] Masayuki Karasuyama and Ichiro Takeuchi. Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks, 21(7):1048-1059, 2010. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "text", + "content": "[34] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Safeguard text-to-image diffusion models with human feedback inversion. In European Conference on Computer Vision, pages 128-145. Springer, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 478, + 294, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 478, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 57, + 478, + 294, + 500 + ], + "type": "text", + "content": "[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "text", + "content": "[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 557, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 601 + ], + "type": "text", + "content": "[37] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 646 + ], + "type": "text", + "content": "[38] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems (NeurIPS), 34:18878-18890, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[39] Bo Liu, Mao Ye, Stephen Wright, Peter Stone, and Qiang Liu. Bome! bilevel optimization made easy: A simple first-order approach. Advances in Neural Information Processing Systems, 35:17248-17262, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[40] Risheng Liu, Xuan Liu, Xiaoming Yuan, Shangzhi Zeng, and Jin Zhang. A value-function-based interior-point method" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "for non-convex bi-level optimization. In International conference on machine learning, pages 6882-6892. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[41] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "text", + "content": "[42] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In International conference on machine learning, pages 2113-2122. PMLR, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[43] Casey Meehan, Kamalika Chaudhuri, and Sanjoy Das-gupta. A non-parametric test to detect data-copying in generative models. In International Conference on Artificial Intelligence and Statistics, 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[44] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N. Ravi. Deep unlearning via randomized conditionally independent hessenians. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10412-10421, 2022. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "text", + "content": "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "text", + "content": "[46] Jiri V Outrata. On the numerical solution of a class of stackelberg problems. Zeitschrift für Operations Research, 34: 255-277, 1990. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "text", + "content": "[47] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 455, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 488 + ], + "type": "text", + "content": "[48] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[50] Enrique Romero, Ignacio Barrio, and Lluis Belanche. Incremental and decremental learning for linear support vector machines. In International Conference on Artificial Neural Networks, pages 209-218. Springer, 2007. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[51] Mehdi SM Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. Advances in neural information processing systems, 31, 2018. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[52] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023.1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[53] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content?" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28272" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350-1361, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[54] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 229 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 284 + ], + "type": "text", + "content": "[56] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS), 34: 18075-18086, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "type": "text", + "content": "[57] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 387 + ], + "type": "text", + "content": "[58] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "text", + "content": "[59] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 464 + ], + "type": "text", + "content": "[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 466, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 296, + 510 + ], + "type": "text", + "content": "[61] Ayush Kumar Tarun, Vikram Singh Chandawat, Murari Mandal, and Mohan Kankanhalli. Deep regression unlearning. In International Conference on Machine Learning, pages 33921-33939. PMLR, 2023. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 296, + 555 + ], + "type": "text", + "content": "[62] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 609 + ], + "type": "text", + "content": "[63] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 35: 38274-38290, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 296, + 655 + ], + "type": "text", + "content": "[64] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "type": "text", + "content": "[65] Nikhil Vyas, Sham Kakade, and Boaz Barak. Provable copyright protection for generative models. arXiv preprint arXiv:2302.10870, 2023. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[66] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. This person (probably) exists. identity member-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ship attacks against gan generated faces. arXiv preprint arXiv:2107.06018, 2021.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[67] Jing Wu and Mehrtash Harandi. Scissorhands: Scrub data influence via connection sensitivity in networks. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[68] Yinjun Wu, Edgar Dobriban, and Susan Davidson. DeltaGrad: Rapid retraining of machine learning models. In Proceedings of the 37th International Conference on Machine Learning, pages 10355-10366. PMLR, 2020. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[69] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 87-103. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 283 + ], + "type": "text", + "content": "[70] Jane J Ye and DL Zhu. Optimality conditions for bilevel programming problems. Optimization, 33(1):9-27, 1995. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[71] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[72] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024.6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28273" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_content_list.json b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d7a7c92a24884148cf0c7e2387fdb52598eec3f5 --- /dev/null +++ b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "Escaping Plato's Cave: Towards the Alignment of 3D and Text Latent Spaces", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Souhail Hadgi", + "bbox": [ + 258, + 179, + 383, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luca Moschella $^{2,*}$", + "bbox": [ + 424, + 180, + 571, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrea Santilli2", + "bbox": [ + 606, + 181, + 735, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diego Gomez", + "bbox": [ + 259, + 199, + 380, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qixing Huang", + "bbox": [ + 423, + 199, + 544, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emanuele Rodolà2", + "bbox": [ + 584, + 199, + 735, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simone Melzi3", + "bbox": [ + 343, + 217, + 462, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maks Ovsjanikov1", + "bbox": [ + 504, + 217, + 653, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "École polytechnique", + "bbox": [ + 225, + 234, + 403, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sapienza University of Rome", + "bbox": [ + 522, + 234, + 769, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Milano-Bicocca", + "bbox": [ + 199, + 253, + 455, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4The University of Texas at Austin", + "bbox": [ + 496, + 253, + 797, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works have shown that, when trained at scale, uni-modal 2D vision and text encoders converge to learned features that share remarkable structural properties, despite arising from different representations. However, the role of 3D encoders with respect to other modalities remains unexplored. Furthermore, existing 3D foundation models that leverage large datasets are typically trained with explicit alignment objectives with respect to frozen encoders from other representations. In this work, we investigate the possibility of a posteriori alignment of representations obtained from uni-modal 3D encoders compared to text-based feature spaces. We show that naive post-training feature alignment of uni-modal text and 3D encoders results in limited performance. We then focus on extracting subspaces of the corresponding feature spaces and discover that by projecting learned representations onto well-chosen lower-dimensional subspaces the quality of alignment becomes significantly higher, leading to improved accuracy on matching and retrieval tasks. Our analysis further sheds light on the nature of these shared subspaces, which roughly separate between semantic and geometric data representations. Overall, ours is the first work that helps to establish a baseline for posttraining alignment of 3D uni-modal and text feature spaces, and helps to highlight both the shared and unique properties of 3D data compared to other representations.", + "bbox": [ + 88, + 335, + 485, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 743, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in multi-modal learning, particularly in vision-language models such as CLIP [47], have sparked interest in extending these successes to the 3D domain. Most current approaches primarily focus on training 3D encoders through triplet-based learning with pre-trained 2D vision and language encoders [32, 64, 69], leveraging new large-scale datasets such as Objaverse [9, 10], and showing promising", + "bbox": [ + 89, + 768, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "results in tasks such as zero-shot shape recognition.", + "bbox": [ + 511, + 305, + 854, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While CLIP [47] was trained with explicit alignment objectives between text and image representations, recent work has observed that even when trained independently, vision and text encoders tend to exhibit significant similarities [21, 37]. In particular, the learned latent spaces of pure (uni-modal) text and vision encoders have similar proximity structure [21] and can be aligned relatively easily after training with a small number of known anchor correspondences [2, 37, 43]. Furthermore, the degree of similarity in learned features across modalities strongly correlates with the quality of performance in various downstream tasks. One prominent interpretation of these results is that given sufficient scale of training data and model complexity, different representations are converging to a shared underlying structure of the physical world. This has given rise to the Platonic Representation Hypothesis [21], where different representations are viewed as projections of reality on particular modalities.", + "bbox": [ + 511, + 321, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At the same time, as the physical world is inherently (at least) 3-dimensional, a natural question is how the structure of uni-modal vision or text encoders relates to the features learned directly from 3D data. This question raises several challenges: first, large-scale 3D datasets have only recently become available [9, 10]; second, most existing 3D \"foundation models\" are trained with explicit alignment objectives with respect to frozen 2D and text encoders, which limits the utility of post-training comparison. Finally, there is a lack of universally agreed-upon architectures and training objectives for 3D data, and many commonly-used architectures tend to have limited generalization power [62].", + "bbox": [ + 511, + 580, + 908, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we initiate the first study on the relation between 3D and language representations. We formulate the task of post-training alignment between a range of 3D and text encoders and study the accuracy and utility of several alignment strategies.", + "bbox": [ + 511, + 763, + 906, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our first insight is that when trained on pure 3D data with self-supervised objectives, 3D encoders tend to lead to representations that align only very weakly to text-based representations. We believe that this observation already", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Currently at Apple.", + "bbox": [ + 109, + 886, + 220, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19825", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46ffe3e9717f4e2077b73030cfc9ed91b0269e1ca365e2d139c2b0f60b639825.jpg", + "image_caption": [ + "Figure 1. A visual overview of the proposed approach. is illustrated. From left to right: We begin with two distinct input collections—one consisting of 3D shapes and the other of textual prompts. In the blue box, independent, frozen uni-modal encoders map each modality into separate, high-dimensional latent spaces, shown in the red box. A dimensionality reduction procedure is applied to project these learned spaces into low-dimensional subspaces, represented in the green box. Finally, an alignment method registers the two low-dimensional subspaces, enabling cross-modal applications such as shape retrieval, with examples depicted in the yellow box." + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Highlights the difficulty of uni-modal training and sheds light on the scarcity of \"pure\" 3D foundation models.", + "bbox": [ + 88, + 459, + 482, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we reveal that while 3D and text latent spaces are not naturally aligned, effective cross-modal translation can be achieved through subspace projection and alignment. Our key insight is that by identifying and operating in correlated subspaces, we can improve the latent space alignment of 3D and text encoders without the need for expensive joint training. To achieve this, we introduce a simple but effective approach that combines Canonical Correlation Analysis (CCA) and existing alignment approaches such as affine translation [36] and local CKA [37]. This approach and our subsequent analysis extend previous works aimed at comparing learned feature spaces, but shows that careful subspace alignment can reveal subtle but important similarities, which are otherwise obscured in global comparisons. We provide a visual overview of the proposed approach in Fig. 1.", + "bbox": [ + 91, + 493, + 485, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 109, + 739, + 464, + 753 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We extend the analysis of vision-text uni-modal latent space alignment to 3D uni-modal encoders and text, highlighting the limited similarity between these latent spaces and the low efficacy of current alignment approaches.", + "2. We propose an efficient approach for cross-modal alignment between text and 3D features that combines CCA for subspace selection and existing alignment methods. This approach improves alignment performance with minimal computational overhead, as demonstrated through" + ], + "bbox": [ + 89, + 758, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "experiments in matching and retrieval tasks. Our method establishes a baseline for 3D-text cross-modal understanding, providing an alternative to explicit multimodal pretraining for cross-modal tasks.", + "bbox": [ + 532, + 460, + 908, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We observe a complementary structure between the spanned subspaces and the original feature spaces, enabling a distinction between geometric and semantic representations.", + "bbox": [ + 513, + 525, + 908, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 597, + 653, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal representation learning. Multimodal representation learning has surged in recent years, driven by the success of image-language models [25, 30, 47, 52] that enable seamless cross-modal applications. These models serve as the backbone for tasks spanning from text-based image retrieval to generating high-quality visuals from natural language prompts. By aligning visual and linguistic features in a shared latent space, these models have paved the way for advanced interactions between modalities, setting the stage for applications where visual and textual information are jointly processed and understood.", + "bbox": [ + 511, + 628, + 908, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these advancements, vision-language models have recently been adapted for 3D point cloud representation, where 3D-image-text triplets [32, 63, 64, 69] enable contrastive pre-training. These models leverage powerful techniques such as momentum contrast [19] to align representations across modalities, allowing point cloud encoders to be pre-trained in a multimodal context. Additionally, tech-", + "bbox": [ + 511, + 795, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19826", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "niques that apply vision-language models directly to 2D projections of point clouds [68, 70] have expanded cross-modal applications to 3D, including zero-shot shape classification, where models trained on 2D-image-text data demonstrate strong performance on 3D-related tasks.", + "bbox": [ + 88, + 90, + 483, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the domain of multimodal synthesis, recent efforts in 2D and 3D model generation have leveraged diffusion models to tackle complex generation tasks. These models use robust priors, often trained on vast datasets, which allow them to create high-fidelity outputs in scenarios such as novel view synthesis and realistic 3D reconstructions from single RGB images [33, 65]. By combining synthetic data with diffusion-based priors, these frameworks achieve impressive zero-shot generalization and geometry-consistent 3D synthesis, opening new avenues for applications where generating realistic 3D content from limited information is critical.", + "bbox": [ + 89, + 170, + 483, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representational similarity. The study of representation similarity across neural networks has seen a significant rise in interest, spurred largely by seminal works originating in neuroscience and computational cognitive science. These fields have long been invested in understanding the nature and alignment of cognitive representations, providing a foundational basis that has influenced the current trajectory in machine learning [42, 56]. Based on these insights, various works [1, 31, 40, 41, 43, 44, 59] provide evidence of an intrinsic connection between independently trained networks; the similarity is especially notable among large-scale models, with works such as [39, 40, 55] exploring the phenomenon. In the computer vision and pattern recognition areas, the line of work on similarity-based representations, pioneered by Duin and Pekalska [45], has also been seminal. This line of research, which examines data through the lens of similarity rather than feature attributes, has provided robust frameworks for classification and clustering [5], enabling models to generalize across patterns and variations in complex datasets. Although mostly empirical in nature, these observations find theoretical support in the study of harmonics in neural networks weights [38], Independent Component Analysis [22, 23, 26, 50] and Independent Mechanism Analysis [13, 14, 54]. These works suggest that, when capturing the same underlying data generative factors, deep learning models may converge towards similar structures despite their complexity and non-linearity.", + "bbox": [ + 91, + 362, + 483, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent space alignment. More recently, a range of approaches has been developed to align latent spaces within the same modality [3, 7, 12], as well as across different modalities [36, 58, 67], particularly between visual and textual domains. Techniques such as Procrustes analysis [60] and several similarity metrics [27, 53, 57, 61], including centered kernel alignment (CKA) [8, 28, 37], have proven instrumental in aligning representations. These methods offer", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "various strategies for quantifying similarity between feature spaces, allowing us to examine cross-modal interactions at a deeper level. HHowever, these methods often focus on aligning entire latent spaces, potentially missing meaningful similarities confined to specific subspaces.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach builds upon CCA [20, 49], a pivotal tool in pattern recognition and multi-view learning applications [16, 51]. This technique identifies maximally correlated subspaces, enabling more refined alignment across modalities by isolating core, mutually relevant components. Leveraging CCA, alignment methods can extend into complex domains, such as connecting 3D and textual latent spaces, as we demonstrate in the following.", + "bbox": [ + 511, + 167, + 908, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 300, + 604, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare the similarity between latent (feature) spaces of various 3D and text encoders, introducing a new approach that builds on existing alignment methods to improve their effectiveness. Below, we outline the tools used to measure and align these latent spaces.", + "bbox": [ + 511, + 325, + 906, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 409, + 653, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Centered Kernel Alignment. CKA is a similarity measure frequently used in recent studies [8, 28] to compare representations in neural network feature spaces. Given feature matrices $\\mathbf{X} \\in \\mathbb{R}^{n \\times p}$ and $\\mathbf{Y} \\in \\mathbb{R}^{n \\times q}$ , we apply kernel functions $k$ and $l$ to obtain kernel representations $\\mathbf{K} = k(\\mathbf{X}, \\mathbf{X}) \\in \\mathbb{R}^{n \\times n}$ and $\\mathbf{L} = l(\\mathbf{Y}, \\mathbf{Y}) \\in \\mathbb{R}^{n \\times n}$ . CKA is defined as:", + "bbox": [ + 511, + 436, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C K A} (\\mathbf {K}, \\mathbf {L}) = \\frac {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {L})}{\\sqrt {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {K}) \\operatorname {H S I C} (\\mathbf {L} , \\mathbf {L})}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 547, + 906, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where HSIC is the Hilbert-Schmidt Independence Criterion [15] and can be written as", + "bbox": [ + 511, + 590, + 905, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {H S I C} (\\mathbf {K}, \\mathbf {L}) = \\frac {1}{(n - 1) ^ {2}} \\operatorname {t r} (\\mathbf {K H L H}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 628, + 906, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{H} = \\mathbf{I} - \\frac{1}{n}\\mathbf{1}\\mathbf{1}^{\\top}$", + "bbox": [ + 511, + 669, + 671, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Canonical Correlation Analysis. CCA [20, 49] is a statistical method that finds linear projections of two datasets, maximizing their correlation in a shared latent space.", + "bbox": [ + 511, + 691, + 908, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, given two sets of zero-centered variables $\\mathbf{X} \\in \\mathbb{R}^{n \\times d_1}$ and $\\mathbf{Y} \\in \\mathbb{R}^{n \\times d_2}$ , CCA seeks two projection matrices $\\mathbf{W}_{\\mathbf{X}} \\in \\mathbb{R}^{d_1 \\times k}$ and $\\mathbf{W}_{\\mathbf{Y}} \\in \\mathbb{R}^{d_2 \\times k}$ that map $\\mathbf{X}$ and $\\mathbf{Y}$ into a common $k$ -dimensional space, maximizing the correlation between the projections $\\mathbf{X}\\mathbf{W}_{\\mathbf{X}}$ and $\\mathbf{Y}\\mathbf{W}_{\\mathbf{Y}}$ . The optimization can be expressed as:", + "bbox": [ + 511, + 737, + 908, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbf {W} _ {\\mathbf {X}}, \\mathbf {W} _ {\\mathbf {Y}}} \\operatorname {c o r r} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {X}}, \\mathbf {Y} \\mathbf {W} _ {\\mathbf {Y}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 837, + 906, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\operatorname{corr}(\\cdot, \\cdot)$ represents the correlation between the projected variables.", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Alignment approaches", + "text_level": 1, + "bbox": [ + 89, + 90, + 302, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent developments in latent space alignment have introduced various methods. In this work, we examine both the affine transformation approach in [36] and the CKA-based matching approach from [37], observing their limited effectiveness in 3D-text latent space alignment, and propose a method to address these limitations.", + "bbox": [ + 89, + 112, + 485, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Space Translation through Affine Transformation [36]. It is possible to estimate an affine transformation $T$ that maps one latent space $\\mathcal{X}$ onto another latent space $\\mathcal{Y}$ such that $T(\\mathbf{x}) = \\mathbf{R}\\mathbf{x} + \\mathbf{b}, \\forall \\mathbf{x} \\in \\mathcal{X}$ . To compute $T$ , we assume access to an anchor subset consisting of ground-truth paired samples from both latent spaces. These anchors enable us to determine $T$ by optimizing through gradient descent or, if $\\mathbf{b} = \\mathbf{0}$ by minimizing the least squares error. It assumes that $\\mathcal{X}$ and $\\mathcal{Y}$ share the same dimensionality and are normalized to zero mean and unit variance. These constraints can be enforced by zero-padding the smaller latent space.", + "bbox": [ + 89, + 209, + 485, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local CKA-based retrieval and matching [37]. The CKA metric computed between two sets of features is sensitive to ordering and maximized when ground-truth pairs are aligned, this insight can be used to match unseen data by including them in a well-aligned anchors set. Formally, starting from aligned set of features $\\mathbf{X}_A$ and $\\mathbf{Y}_A$ , we compute for a query pair $(\\mathbf{x}_q,\\mathbf{y}_q)$ its local CKA defined as:", + "bbox": [ + 89, + 381, + 483, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o c a l C K A} \\left(\\mathbf {x} _ {q}, \\mathbf {y} _ {q}\\right) = \\operatorname {C K A} \\left(\\mathbf {K} _ {\\left[ \\mathbf {X} _ {A}, \\mathbf {x} _ {q} \\right]}, \\mathbf {K} _ {\\left[ \\mathbf {Y} _ {A}, \\mathbf {y} _ {q} \\right]}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 500, + 483, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\mathbf{X},\\mathbf{x}]$ denotes the column-wise concatenation of matrix $\\mathbf{X}$ and vector $\\mathbf{x}$ . Local CKA calculates similarity in a pairwise manner, accounting for anchor alignment while being sensitive to ordering among query pairs. We introduce all possible query pairs, where correctly matched pairs exhibit the highest local CKA.", + "bbox": [ + 89, + 527, + 485, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Proposed method.", + "text_level": 1, + "bbox": [ + 89, + 630, + 266, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to reliably align the latent spaces of pre-trained 3D and text encoders. Given a dataset of $n$ caption-point cloud pairs, each embedded in the corresponding feature space, and represented as matrices $\\mathbf{X} \\in \\mathbb{R}^{n \\times p}$ and $\\mathbf{Y} \\in \\mathbb{R}^{n \\times q}$ respectively, we first select a subset of anchor pairs that will guide our translation process, denoted as $\\mathbf{X}_A \\in \\mathbb{R}^{n_A \\times p}$ and $\\mathbf{Y}_A \\in \\mathbb{R}^{n_A \\times q}$ . Building upon the mathematical foundations introduced in Section 3.2, we develop a pipeline that combines dimensionality reduction with the previously introduced alignment methods.", + "bbox": [ + 89, + 652, + 483, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Common Subspace Projection. In our work, we show that 3D and text latent spaces can effectively be aligned in low-dimensional connected subspaces. We begin by applying CCA to the anchor pairs to identify a shared $k$ -dimensional subspace (with $k < p, q$ ) that connects point cloud and text latent spaces. This yields projection matrices $\\mathbf{W}_{\\mathbf{X}^A} \\in \\mathbb{R}^{p \\times k}$", + "bbox": [ + 89, + 809, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\mathbf{W}_{\\mathbf{Y}^A}\\in \\mathbb{R}^{q\\times k}$ . All samples are then projected into this reduced space:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {r} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {X} _ {A}}, \\quad \\mathbf {Y} ^ {r} = \\mathbf {Y} \\mathbf {W} _ {\\mathbf {Y} _ {A}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 133, + 906, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, we refer to the reduced anchors as $\\mathbf{X}_A^r\\in \\mathbb{R}^{n_A\\times k}$ and $\\mathbf{Y}_A^r\\in \\mathbb{R}^{n_A\\times k}$ .", + "bbox": [ + 511, + 164, + 906, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our experiments show that projecting 3D and text latent spaces into a lower-dimensional, shared subspace improves alignment by isolating features that are highly correlated across modalities", + "bbox": [ + 511, + 194, + 906, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alignment of projected latent spaces. Given the projected latent spaces, we can apply either of the previously described alignment methods in the reduced space. For the affine transformation approach, we learn a mapping between $\\mathbf{X}^r$ and $\\mathbf{Y}^r$ using the projected anchor pairs $\\mathbf{X}_A^r$ , $\\mathbf{Y}_A^r$ to optimize the transformation parameters $\\mathbf{R}$ and $\\mathbf{b}$ :", + "bbox": [ + 511, + 260, + 908, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT \\left(\\mathbf {X} ^ {r}\\right) = \\mathbf {R X} ^ {r} + \\mathbf {b}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 364, + 906, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alternatively, we can employ the local CKA-based matching approach in the projected space. For a query pair $(\\mathbf{x}_q^r,\\mathbf{y}_q^r)$ we compute its local CKA using the projected anchor sets:", + "bbox": [ + 511, + 393, + 908, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o c a l C K A} \\left(\\mathbf {x} _ {q} ^ {r}, \\mathbf {y} _ {q} ^ {r}\\right) = \\operatorname {C K A} \\left(\\mathbf {K} _ {\\left[ \\mathbf {X} _ {A} ^ {r}, \\mathbf {x} _ {q} ^ {r} \\right]}, \\mathbf {K} _ {\\left[ \\mathbf {Y} _ {A} ^ {r}, \\mathbf {y} _ {q} ^ {r} \\right]}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 467, + 906, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 511, + 702, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Pre-training Dataset", + "text_level": 1, + "bbox": [ + 511, + 536, + 707, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior works in point cloud pre-training, particularly within uni-modal frameworks, have relied heavily on ShapeNet [4], a dataset of 51,300 annotated 3D synthetic shapes spanning 55 categories. ShapeNet has been instrumental in advancing foundational methods, yet it remains limited by the relatively narrow scope of categories. With the release of Objaverse [9], which includes over 800,000 shapes across diverse real-world categories, a new standard for large-scale representation learning in the 3D domain has emerged. Objaverse's extensive shape diversity makes it ideal for both uni-modal and multi-modal learning. Despite these advantages, there is a lack of works that explore uni-modal pre-training specifically on Objaverse.", + "bbox": [ + 511, + 558, + 908, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.Encoders", + "text_level": 1, + "bbox": [ + 511, + 766, + 624, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We explore both multi-modal and uni-modal 3D and text encoders across varying levels of model complexity.", + "bbox": [ + 511, + 787, + 906, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-modal 3D Encoders. For multi-modal pre-training, we use OpenShape, ULIP-2 and Uni3D [32] pre-trained models, trained on point cloud, image, and text triplets of Objverse with a contrastive pre-text task to align 3D encoders with frozen CLIP encoders. We adopt the Point-BERT-based", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19828", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/822f3f9f748c447b426df4d0f1561685bb536bdf672782a5669c80af17c66cd1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method3D EncoderCLIPRoBERTaBERT
Matching accuracyTop-5 retrievalMatching accuracyTop-5 retrievalMatching accuracyTop-5 retrieval
Multi-modal 3D encoder
Affine + Subspace ProjectionOpenShape67.685.655.275.845.070.6
Affine + Subspace ProjectionULIP-265.685.247.270.635.259.8
Affine + Subspace ProjectionUni3D61.481.645.863.834.447.2
Uni-modal 3D encoder
AffinePointBert15.823.67.815.66.413.4
AffineSparseConv11.034.46.020.04.216.2
AffinePointnet18.421.88.010.29.612.2
Affine + Subspace Projection (Ours)PointBert30.842.223.228.415.618.2
Affine + Subspace Projection (Ours)SparseConv21.445.619.216.815.815.0
Affine + Subspace Projection (Ours)Pointnet25.236.622.420.816.616.0
Local CKAPointBert5.815.21.81.41.04.39
Local CKASparseConv3.413.61.791.60.63.8
Local CKAPointnet6.618.02.42.01.05.0
Local CKA + Subspace Projection (Ours)PointBert29.460.1917.042.415.037.2
Local CKA + Subspace Projection (Ours)SparseConv19.056.915.834.010.030.8
Local CKA + Subspace Projection (Ours)Pointnet26.853.018.640.214.338
", + "bbox": [ + 91, + 87, + 906, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Matching and retrieval performance across 3D and text encoders using different alignment approaches. We use 30,000 anchors for subspace projection and affine transformation approaches, and 1,000 anchors for local CKA. A query set of 500 is uniformly sampled, with results averaged over 3 different seeds. The subspace dimension is fixed at 50. Our approach (Ours) consistently demonstrates improved matching and retrieval performance, with multi-modal 3D encoders setting the upper bound for performance. Additional top- $k$ retrieval metrics are provided in the supplementary.", + "bbox": [ + 89, + 301, + 906, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "variant of each model [66]. We primarily focus on OpenShape for simplicity but generalize results to ULIP-2 and Uni3D.", + "bbox": [ + 89, + 393, + 482, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Uni-modal 3D Encoders. For the uni-modal setup, we pre-train PointBERT on Objaverse using its original pretext tasks, which include masked point reconstruction and a uni-modal contrastive loss to encourage robust shape representation. We also explore two additional architectures: a Sparse convolution (MinkowskiNet [6]) model and a simpler architecture in PointNet [46], each pre-trained using a shape-level contrastive learning method [18]. This approach contrasts different partial views of input shapes. Across all 3D encoders in this setup, we fix the latent dimension at 512 to maintain consistency in representation space comparisons.", + "bbox": [ + 88, + 446, + 482, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text encoders. We use the text encoder from OpenCLIP ViTbigG-14 [24], chosen to match text encoder used in OpenShape. Additionally, we examine alignment with purely unimodal text encoders by including BERT [11] and RoBERTa [34], we also evaluate the alignment with T5 [48] in the supplementary.", + "bbox": [ + 89, + 619, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Across all pre-training setups, parameters are kept consistent to facilitate direct comparisons. We detail the technical details in the supplementary.", + "bbox": [ + 89, + 712, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Downstream tasks", + "text_level": 1, + "bbox": [ + 89, + 771, + 267, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our alignment approach on Objaverse-LVIS [17], a human-verified test subset of Objaverse that contains 1,156 object categories. This subset is specifically reserved for evaluation and is unseen during pre-training. The captions are generated with Cap3D [35], which provides enhanced descriptive text for each 3D shape. Our evaluation framework consists of two main tasks: matching and retrieval. The", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "matching task involves finding the correct permutation of captions for perfect matching given a shuffled set of query images and their corresponding captions; we utilize the linear sum assignment approach to perform this task [29]. For the retrieval task, the model must identify the correct 3D object from the query set based on a text caption. These tasks are particularly effective in measuring the cross-modal capabilities of encoders, and have been evaluated in prior Vision-Text studies [37]. While our main results emphasize the matching and top-5 retrieval tasks, we provide evaluation of top-1 and top-10 retrieval metrics in the supplementary.", + "bbox": [ + 511, + 393, + 906, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 579, + 599, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Are 3D and Text Latent Spaces similar?", + "text_level": 1, + "bbox": [ + 511, + 604, + 857, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the inherent similarity between 3D and text feature spaces without alignment, we compute linear CKA scores for both unimodal and multimodal encoders. The results are shown in Fig. 2.", + "bbox": [ + 511, + 628, + 906, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison of 3D-Text and Vision-Text alignment. Prior work [21] reports CKA alignment values ranging from approximately $30\\%$ to $48\\%$ between different uni-modal vision and text encoders (see Figure 13 in [21]). In stark contrast, we find that the default alignment between 3D and text latent spaces is significantly weaker, with a maximum score of 0.12 observed for the uni-modal PointBERT and CLIP pair. This substantial gap underscores a key insight: unlike vision and text encoders, 3D encoders did not converge to structures similar to text.", + "bbox": [ + 511, + 696, + 908, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment favors multi-modality. We observe that 3D multi-modal encoders demonstrate the highest CKA score with text encoders, particularly with CLIP's text encoder.", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19829", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e3b8a1a21c9eb56a2b848c4343eed59d22494912e782f00a154a42b39b3ae965.jpg", + "image_caption": [ + "Figure 2. Linear CKA scores between text and 3D encoders without alignment. Higher scores reflect stronger alignment between encoder pairs, with the strongest alignment observed between multi-modal 3D encoders (OpenShape) and CLIP text encoder due to their shared training on aligned representations. Uni-modal 3D encoders show significantly lower alignment with text encoders, although slightly higher with the CLIP text encoder." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 480, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a430ff55c377ff41811a4a29ef7e986c858fac148530f041b4a6c7ea1b05d6f7.jpg", + "image_caption": [ + "Figure 3. Linear CKA scores between text and 3D encoders after affine translation. Affine translation results in a consistent increase in similarity for both 3D-to-text and text-to-3D directions." + ], + "image_footnote": [], + "bbox": [ + 96, + 450, + 480, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This behavior is expected, given that 3D multi-modal encoders are explicitly trained to align with CLIP latent spaces, which share the same textual modality. Even among unimodal 3D encoders, alignment with the CLIP text encoder is notably higher (e.g. 0.12 for PointBERT-CLIP) compared to alignment with uni-modal text encoders like RoBERTa (e.g., 0.04 for PointBERT-RoBERTa). This suggests that the visual understanding embedded in CLIP's text encoder extends beyond image and text domains to include 3D representations,", + "bbox": [ + 89, + 763, + 486, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "despite the lack of explicit alignment during pre-training.", + "bbox": [ + 511, + 90, + 890, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Latent Space Alignment results", + "text_level": 1, + "bbox": [ + 511, + 118, + 790, + 133 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of the alignment approaches outlined in Sec. 3 using matching and retrieval tasks, and analyze their effectiveness in aligning 3D and text encoders. Unless otherwise specified, we fix the subspace dimension to $d = 50$ and the number of anchors to 30,000. For downstream tasks, we uniformly sample a query set of size 500 and average results over 3 different seeds. For the affine transformation approach, we present results for the text-to-3D direction, noting that similar performance is achieved in the 3D-to-text direction as seen in Fig. 3.", + "bbox": [ + 511, + 141, + 908, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing approaches enable limited alignment We first assess whether previous successful approaches—affine transformation and local CKA—can achieve meaningful 3D-text alignment. As shown in Tab. 1, both methods yield modest alignment improvements over the unaligned baseline, where uni-modal 3D encoders start near zero in matching and retrieval tasks. These results suggest a small alignment shift. However, even with this improvement, alignment remains significantly lower than the alignment achieved in vision-text benchmarks [37]. This finding hints at the limits of uni-modal 3D encoders in achieving similar Vision-Text alignment performance, and might necessitate a different approach to align their latent space with text encoders.", + "bbox": [ + 511, + 297, + 910, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importance of subspace projection (Ours). While aligning the latent spaces of 3D and text encoders achieved some success with existing methods, results remained well below vision-text alignment benchmarks. Motivated by this limitation, we propose aligning lower-dimensional subspaces, based on the hypothesis that 3D and text representations might intersect within a shared latent subspace. Using a validation set, we report in Fig. 4 the impact of subspace dimension, obtained through our CCA approach (see Sec. 3.3) combined with the affine transformation on the top-5 retrieval accuracy of uni-modal PointBert. While affine alignment alone yields better performance at higher dimensions, our subspace projection approach significantly outperforms it when the subspace dimension is reduced. This finding indicates that alignment quality is optimized within carefully chosen lower-dimensional spaces, reinforcing our hypothesis that 3D-text alignment is more successful within targeted subspaces. The results in Tab. 1 further validate this approach, showing a substantial increase in matching and retrieval performance across all 3D and text encoder pairs, outperforming both alignment approaches when they operate on the original latent space.", + "bbox": [ + 511, + 501, + 908, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-modal encoders as an upper bound of performance. We include 3D multi-modal encoders for two main reasons: (1) As an upper bound for alignment performance, (2) To study how alignment degrades across different text", + "bbox": [ + 511, + 839, + 911, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19830", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "encoders, including those that were not used during pretraining. In this context, the inclusion of 'Affine + Subspace Projection' multi-modal results in Tab. 1 shows how using our approach significantly narrows the gap between multimodal and uni-modal performance. However, to illustrate point 1. above, we also measured with cosine similarity the alignment of OpenShape (a multi-modal 3D encoder) and the same CLIP text encoder used during pre-training, achieving 0.94 for top-5 retrieval accuracy, an expected result since 3D and text encoders are explicitly aligned during pre-training.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D encoder complexity's low impact on alignment. Point-BERT outperforms the more complex SparseConv among uni-modal 3D encoders, suggesting that increasing model complexity alone does not guarantee improved alignment in 3D-text tasks. Surprisingly, even PointNet—a relatively simple model—achieves similar alignment scores, showing that factors other than model complexity may play a pivotal role in 3D-text alignment. This observation contrasts with vision-text alignment results [21, 37], where complex models typically leverage large datasets more effectively. Our findings thus indicate a distinctive aspect of 3D-text alignment: model simplicity does not impede, and may even aid, the interpretability and compatibility of learned features for cross-modal alignment.", + "bbox": [ + 89, + 260, + 483, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different alignment techniques have different strengths. Our experiments reveal that different alignment techniques offer complementary strengths across tasks. For instance, affine transformation proves particularly effective for matching tasks across all 3D encoders, while local CKA shows superior performance in top-5 retrieval accuracy. This suggests that while some methods excel in precision tasks, others might better capture broader semantic nuances, making them more suitable for retrieval. Together, these observations imply that a hybrid approach could leverage the unique strengths of each method, opening up promising directions for future cross-modal applications.", + "bbox": [ + 89, + 489, + 483, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling of our approach. In Fig. 5, we explore the scalability of our approach by analyzing how performance responds to increasing the number of anchors. Notably, our subspace projection method scales effectively with anchor count, reaching a plateau before requiring the full dataset (over 800,000 shapes). This scalability highlights the approach's efficiency in learning robust 3D-text mappings with a limited subset of anchor pairs. However, we observe a leveling off in performance gains beyond a certain anchor count, likely due to the constraints imposed by the low-dimensional subspace. This suggests that while our approach is computationally efficient and data-efficient, its reliance on a fixed lower-dimensional subspace could limit its adaptability to larger, more diverse datasets.", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c832cb4e78c788f2cf6dcbec1a08c1cba97b07f87b1f7df234c4033bc7208dc.jpg", + "image_caption": [ + "Figure 4. Impact of subspace dimensionality on retrieval performance. Comparison of three approaches: our proposed CCA + affine translation method (blue), affine translation without subspace projection (red), and baseline feature space alignment without transformation (orange). Results are shown for the uni-modal PointBERT and CLIP text encoder, with generalizations to other encoders provided in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 707, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9065bb67ee67e43d30b785a71fa1390c78ab6621ac0bbc35628e80157872791.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 903, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50266cd140248909d4f28fcc4073c59f6a56340e34d95dbd174a240c12897d62.jpg", + "image_caption": [ + "Figure 5. Effect of anchor set size on retrieval performance. Validation set results with different subspace dimensions show that retrieval performance improves as the anchor subset size increases but eventually reaches a plateau. Results are shown for the unimodal PointBERT and CLIP text encoder." + ], + "image_footnote": [], + "bbox": [ + 519, + 323, + 898, + 478 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41977d9bdfc5fe4789fbc00ff5b3f9d5c3e1fa433ae05099d05dacd5a80dfaae.jpg", + "image_caption": [ + "Figure 6. Pearson correlation between shape query chamfer distances and pairwise distances in the projected text latent subspace. We observe a higher Pearson correlation in the projected text latent subspace with optimal subspace dimension. Results are shown for the uni-modal PointBERT and CLIP text encoders." + ], + "image_footnote": [], + "bbox": [ + 521, + 590, + 898, + 776 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19831", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eab150a7a7dacf55cc3ce2cc95ad954123cf061de63ca241fb312bb39382336d.jpg", + "image_caption": [ + "Figure 7. Shape retrieval comparison between original and reduced latent spaces. For a given query shape, we retrieve the closest match based on cosine similarity. Results demonstrate a higher semantic understanding in the reduced 3D latent subspace compared to the original latent space. Shown for uni-modal Point-BERT and CLIP." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 478, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Geometries vs. semantics.", + "text_level": 1, + "bbox": [ + 89, + 674, + 323, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our quantitative evaluation shows that low-dimensional subspace projection significantly improves latent space alignment in the 3D-Text setting. To better understand the characteristics of these subspaces spanned relative to the original spaces, we analyze the increase in semantic and geometric knowledge within the projected spaces.", + "bbox": [ + 89, + 696, + 485, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Increased geometric awareness of the text latent subspace. To quantify the increase in geometric awareness within the projected text subspace, we compute the Pearson correlation between the Chamfer distances of a query set of 500 shapes and the pairwise distances between feature vectors within the text subspace and compare these results to those obtained from the original text latent space. Results", + "bbox": [ + 89, + 795, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 6 show that the optimal subspace dimension yields a high correlation with Chamfer distances, signaling an increase in the sensitivity of the text encoder to geometric properties when projected into this specific subspace.", + "bbox": [ + 511, + 90, + 908, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Increased semantic understanding of the 3D latent subspace. To assess the improved semantic capacity of the 3D subspace, we conducted a qualitative analysis of shape retrieval performance in both the reduced and the original latent spaces (Fig. 7). Given a set of query shapes, we observe that the reduced subspace retrieves shapes with stronger semantic similarity (e.g., retrieving animals for a query sheep shape), while the original latent space primarily favors geometric resemblance (e.g. retrieving stacked square shapes when querying a '3D model of a CPU'). This shift suggests that our subspace projection method enables the 3D encoder to capture semantic relationships that are absent in the full latent space. This effectively enables cross-modal applications and explains its higher matching and retrieval performance.", + "bbox": [ + 511, + 159, + 908, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion, Limitations and Future Work", + "text_level": 1, + "bbox": [ + 511, + 388, + 888, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present the first study investigating latent space alignment between 3D and text pre-trained encoders. Building on the hypothesis that these modalities share semantic connections within lower-dimensional subspaces, we propose an effective approach combining CCA projection with affine transformation estimation to translate between modalities' latent spaces. Our empirical results show that optimal cross-modal performance is achieved through low-dimensional subspace projection, and our method successfully improves alignment across diverse 3D and text uni-modal encoders. While CLIP-based multi-modal encoders establish performance upper bounds, we enable significant cross-modal capabilities in uni-modal encoders previously limited to single-modality tasks. We also demonstrate that semantic understanding can be extracted from geometry-aware latent spaces of uni-modal 3D encoders.", + "bbox": [ + 511, + 415, + 908, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our work focused on Objaverse, which is the first large-scale 3D dataset, it would be interesting to consider how scaling on Objaverse-XL [10] would affect the alignment quality between 3D and text encoders. Moreover, in this work we do not distinguish object-level vs. scene-level annotations, and decomposing objects or scenes into their composing blocks could shed light onto the compositionality of the learned representations. Finally, the subspace alignment method that we introduce in this work can be broadly applicable to other representations as well. In the future, we plan to use it to investigate the limitations of alignment observed in other representations. In particular, even when trained at significantly higher data scales, images and text representations do not align perfectly [21], and it would be interesting to reveal the unique and complementary nature of different modalities via subspace analysis.", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19832", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, as well as gifts from Ansys and Adobe Research. This work was also supported by the Galileo 2022 fellowship from the Università Italo Francese/ Université Franco Italiène (UIF/UFI) within the project G22_4 titled “Multimodal Artificial Intelligence for 3D shape analysis, modeling and applications”.", + "bbox": [ + 89, + 90, + 485, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 257 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Bonheme and Marek Grzes. How do variational autoencoders learn? insights from representational similarity. arXiv preprint arXiv:2205.08399, 2022. 3", + "[2] Irene Cannistraci, Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, and Emanuele Rodola. Bootstrapping parallel anchors for relative representations. In The First Tiny Papers Track at ICLR 2023, Tiny Papers at ICLR 2023, 2023. 1", + "[3] Irene Cannistraci, Luca Moschella, Marco Fumero, Valentino Maiorca, and Emanuele Rodolà. From bricks to bridges: Product of invariances to enhance latent space communication. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[4] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4", + "[5] Yihua Chen, Eric K. Garcia, Maya R. Gupta, Ali Rahimi, and Luca Cazzanti. Similarity-based classification: Concepts and algorithms. Journal of Machine Learning Research, 10(27): 747-776, 2009. 3", + "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 5", + "[7] Donato Crisostomi, Irene Cannistraci, Luca Moschella, Pietro Barbiero, Marco Ciccone, Pietro Lio, and Emanuele Rodola. From charts to atlas: Merging latent spaces into one. In NeurIPS 2023 Workshop on Symmetry and Geometry in Neural Representations, 2023. 3", + "[8] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 3", + "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. 2023 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2022. 1, 4", + "[10] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli" + ], + "bbox": [ + 93, + 265, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 1, 8", + "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[12] Marco Fumero, Marco Pegoraro, Valentino Maiorca, Francesco Locatello, and Emanuele Rodolà. Latent functional maps: a spectral framework for representation alignment. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3", + "[13] Shubhangi Ghosh, Luigi Gresele, Julius von Kugelgen, Michel Besserve, and Bernhard Scholkopf. Independent mechanism analysis and the manifold hypothesis, 2023. 3", + "[14] Luigi Gresele, Julius Von Kugelgen, Vincent Stimper, Bernhard Scholkopf, and Michel Besserve. Independent mechanism analysis, a new concept? In Advances in Neural Information Processing Systems, 2021. 3", + "[15] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3", + "[16] Chenfeng Guo and Dongrui Wu. Canonical correlation analysis (cca) based multi-view learning: An overview. arXiv preprint arXiv:1907.01693, 2019. 3", + "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5", + "[18] Souhail Hadgi, Lei Li, and Maks Ovsjanikov. To supervise or not to supervise: Understanding and addressing the key challenges of 3d transfer learning. arXiv preprint arXiv:2403.17869, 2024.5", + "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arxiv e-prints, art. arXiv preprint arXiv:1911.05722, 2019. 2", + "[20] Harold Hotelling. Relations between two sets of variates. In Breakthroughs in statistics: methodology and distribution, pages 162-190. Springer, 1992. 3", + "[21] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987, 2024. 1, 5, 7, 8", + "[22] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA. In Advances in Neural Information Processing Systems. Curran Associates, Inc. 3", + "[23] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, pages 859–868. PMLR. 3", + "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19833", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2", + "[26] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR. 3", + "[27] Max Klabunde, Tobias Schumacher, Markus Strohmaier, and Florian Lemmerich. Similarity of neural network models: A survey of functional and representational measures, 2024. 3", + "[28] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 3", + "[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5", + "[30] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2", + "[31] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John Hopcroft. Convergent learning: Do different neural networks learn the same representations? arXiv preprint arXiv:1511.07543, 2015. 3", + "[32] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in neural information processing systems, 36, 2024. 1, 2, 4", + "[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. 2023. 3", + "[34] Yinhan Liu. *Roberta: A robustly optimized bert pretraining approach.* arXiv preprint arXiv:1907.11692, 364, 2019. 5", + "[35] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[36] Valentino Maiorca, Luca Moschella, Antonio Norelli, Marco Fumero, Francesco Locatello, and Emanuele Rodola. Latent space translation via semantic alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4", + "[37] Mayug Maniparambili, Raiymbek Akshulakov, Yasser Abdelaziz Dahou Djilali, Mohamed El Amine Seddik, Sanath Narayan, Karttikeya Mangalam, and Noel E O'Connor. Do vision and language encoders represent the world similarly? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14334-14343, 2024. 1, 2, 3, 4, 5, 6, 7", + "[38] Giovanni Luca Marchetti and Christopher Hillar. Harmonics of learning: Universal fourier features emerge" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in invariant networks. https://synthical.com/article/03032df5-f9c5-43a6-8253-d4af1d67e0d4, 2023.3", + "[39] Raghav Mehta, Vítor Albiero, Li Chen, Ivan Evtimov, Tamar Glaser, Zhiheng Li, and Tal Hassner. You only need a good embeddings extractor to fix spurious correlations, 2022. 3", + "[40] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3", + "[41] Ari S. Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, page 5732-5741, Red Hook, NY, USA, 2018. Curran Associates Inc. 3", + "[42] Luca Moschella. Latent communication in artificial neural networks, 2024. 3", + "[43] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In The Eleventh International Conference on Learning Representations, 2023. 1, 3", + "[44] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. Advances in Neural Information Processing Systems, 36:15303-15319, 2023. 3", + "[45] E. Pekalska and Robert Duin. Automatic pattern recognition by similarity representations. *Electronics Letters*, 37:159 - 160, 2001. 3", + "[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 5", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5", + "[49] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3", + "[50] Geoffrey Roeder, Luke Metz, and Durk Kingma. On Linear Identifiability of Learned Representations. In Proceedings of the 38th International Conference on Machine Learning, pages 9030-9039. PMLR. 3", + "[51] Jan Rupnik and John Shawe-Taylor. Multi-view canonical correlation analysis. In Conference on data mining and data warehouses (SiKDD 2010), pages 1-4, 2010. 3" + ], + "bbox": [ + 516, + 92, + 908, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19834", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[53] Mahdiyar Shahbazi, Ali Shirali, Hamid Aghajan, and Hamed Nili. Using distance on the riemannian manifold to compare representations in brain and in models. NeuroImage, 239: 118271, 2021. 3", + "[54] Joanna Sliwa, Shubhangi Ghosh, Vincent Stimper, Luigi Gresele, and Bernhard Scholkopf. Probing the robustness of independent mechanism analysis for representation learning. In UAI 2022 Workshop on Causal Representation Learning, 2022. 3", + "[55] Gowthami Somepalli, Liam Fowl, Arpit Bansal, Ping Yeh-Chiang, Yehuda Dar, Richard Baraniuk, Micah Goldblum, and Tom Goldstein. Can neural nets learn the same model twice? investigating reproducibility and double descent from the decision boundary perspective. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13689-13698, 2022. 3", + "[56] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. Love, Erin Grant, Iris Groen, Jascha Achterberg, Joshua B. Tenenbaum, Katherine M. Collins, Katherine L. Hermann, Kerem Oktar, Klaus Greff, Martin N. Hebart, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas P. O'Connell, Thomas Unterthiner, Andrew K. Lampinen, Klaus-Robert Müller, Mariya Toneva, and Thomas L. Griffiths. Getting aligned on representational alignment, 2023. 3", + "[57] Shuai Tang, Wesley J. Maddox, Charlie Dickens, Tom Diethe, and Andreas Damianou. Similarity of neural networks with gradients, 2020. 3", + "[58] Thomas Theodoridis, Theocharis Chatzis, Vassilios Solachidis, Kosmas Dimitropoulos, and Petros Daras. Cross-modal variational alignment of latent spaces. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4127-4136, 2020. 3", + "[59] Ivan Vulic, Sebastian Ruder, and Anders Søgaard. Are all good word vector spaces isomorphic? arXiv preprint arXiv:2004.04070, 2020.3", + "[60] Chang Wang and Sridhar Mahadevan. Manifold alignment using procrastes analysis. In Proceedings of the 25th international conference on Machine learning, pages 1120-1127, 2008. 3", + "[61] Alex H Williams, Erin Kunz, Simon Kornblith, and Scott Linderman. Generalized shape metrics on neural representations. In Advances in Neural Information Processing Systems, 2021. 3", + "[62] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 2", + "[64] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27091–27101, 2024. 1, 2", + "[65] Yuxuan Xue, Xianghui Xie, Riccardo Marin, and Gerard. Pons-Moll. Human 3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models. 2024. 3", + "[66] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, and Jie Zhou. Pre-training 3d point cloud transformers with masked point modeling. 2022 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19291-19300, 2021. 5", + "[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18102-18112, 2022. 3", + "[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3", + "[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. arXiv preprint arXiv:2310.06773, 2023. 1, 2", + "[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 3(4), 2022. 3" + ], + "bbox": [ + 516, + 90, + 906, + 626 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19835", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_model.json b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c4607717186dbe34704ea7cb59172174133c057 --- /dev/null +++ b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "Escaping Plato's Cave: Towards the Alignment of 3D and Text Latent Spaces" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.18, + 0.385, + 0.199 + ], + "angle": 0, + "content": "Souhail Hadgi" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.181, + 0.573, + 0.199 + ], + "angle": 0, + "content": "Luca Moschella\\(^{2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.182, + 0.736, + 0.198 + ], + "angle": 0, + "content": "Andrea Santilli2" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.2, + 0.381, + 0.217 + ], + "angle": 0, + "content": "Diego Gomez" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.2, + 0.545, + 0.217 + ], + "angle": 0, + "content": "Qixing Huang" + }, + { + "type": "text", + "bbox": [ + 0.585, + 0.2, + 0.736, + 0.216 + ], + "angle": 0, + "content": "Emanuele Rodolà2" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.218, + 0.464, + 0.232 + ], + "angle": 0, + "content": "Simone Melzi3" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.218, + 0.654, + 0.235 + ], + "angle": 0, + "content": "Maks Ovsjanikov1" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.236, + 0.404, + 0.252 + ], + "angle": 0, + "content": "École polytechnique" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.236, + 0.771, + 0.252 + ], + "angle": 0, + "content": "Sapienza University of Rome" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.254, + 0.457, + 0.269 + ], + "angle": 0, + "content": "3University of Milano-Bicocca" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.254, + 0.798, + 0.269 + ], + "angle": 0, + "content": "4The University of Texas at Austin" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.716 + ], + "angle": 0, + "content": "Recent works have shown that, when trained at scale, uni-modal 2D vision and text encoders converge to learned features that share remarkable structural properties, despite arising from different representations. However, the role of 3D encoders with respect to other modalities remains unexplored. Furthermore, existing 3D foundation models that leverage large datasets are typically trained with explicit alignment objectives with respect to frozen encoders from other representations. In this work, we investigate the possibility of a posteriori alignment of representations obtained from uni-modal 3D encoders compared to text-based feature spaces. We show that naive post-training feature alignment of uni-modal text and 3D encoders results in limited performance. We then focus on extracting subspaces of the corresponding feature spaces and discover that by projecting learned representations onto well-chosen lower-dimensional subspaces the quality of alignment becomes significantly higher, leading to improved accuracy on matching and retrieval tasks. Our analysis further sheds light on the nature of these shared subspaces, which roughly separate between semantic and geometric data representations. Overall, ours is the first work that helps to establish a baseline for posttraining alignment of 3D uni-modal and text feature spaces, and helps to highlight both the shared and unique properties of 3D data compared to other representations." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.744, + 0.222, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Recent advances in multi-modal learning, particularly in vision-language models such as CLIP [47], have sparked interest in extending these successes to the 3D domain. Most current approaches primarily focus on training 3D encoders through triplet-based learning with pre-trained 2D vision and language encoders [32, 64, 69], leveraging new large-scale datasets such as Objaverse [9, 10], and showing promising" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.306, + 0.855, + 0.321 + ], + "angle": 0, + "content": "results in tasks such as zero-shot shape recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.908, + 0.58 + ], + "angle": 0, + "content": "While CLIP [47] was trained with explicit alignment objectives between text and image representations, recent work has observed that even when trained independently, vision and text encoders tend to exhibit significant similarities [21, 37]. In particular, the learned latent spaces of pure (uni-modal) text and vision encoders have similar proximity structure [21] and can be aligned relatively easily after training with a small number of known anchor correspondences [2, 37, 43]. Furthermore, the degree of similarity in learned features across modalities strongly correlates with the quality of performance in various downstream tasks. One prominent interpretation of these results is that given sufficient scale of training data and model complexity, different representations are converging to a shared underlying structure of the physical world. This has given rise to the Platonic Representation Hypothesis [21], where different representations are viewed as projections of reality on particular modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.763 + ], + "angle": 0, + "content": "At the same time, as the physical world is inherently (at least) 3-dimensional, a natural question is how the structure of uni-modal vision or text encoders relates to the features learned directly from 3D data. This question raises several challenges: first, large-scale 3D datasets have only recently become available [9, 10]; second, most existing 3D \"foundation models\" are trained with explicit alignment objectives with respect to frozen 2D and text encoders, which limits the utility of post-training comparison. Finally, there is a lack of universally agreed-upon architectures and training objectives for 3D data, and many commonly-used architectures tend to have limited generalization power [62]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this work, we initiate the first study on the relation between 3D and language representations. We formulate the task of post-training alignment between a range of 3D and text encoders and study the accuracy and utility of several alignment strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Our first insight is that when trained on pure 3D data with self-supervised objectives, 3D encoders tend to lead to representations that align only very weakly to text-based representations. We believe that this observation already" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.221, + 0.901 + ], + "angle": 0, + "content": "*Currently at Apple." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19825" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.905, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.364, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 1. A visual overview of the proposed approach. is illustrated. From left to right: We begin with two distinct input collections—one consisting of 3D shapes and the other of textual prompts. In the blue box, independent, frozen uni-modal encoders map each modality into separate, high-dimensional latent spaces, shown in the red box. A dimensionality reduction procedure is applied to project these learned spaces into low-dimensional subspaces, represented in the green box. Finally, an alignment method registers the two low-dimensional subspaces, enabling cross-modal applications such as shape retrieval, with examples depicted in the yellow box." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Highlights the difficulty of uni-modal training and sheds light on the scarcity of \"pure\" 3D foundation models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.494, + 0.486, + 0.736 + ], + "angle": 0, + "content": "Additionally, we reveal that while 3D and text latent spaces are not naturally aligned, effective cross-modal translation can be achieved through subspace projection and alignment. Our key insight is that by identifying and operating in correlated subspaces, we can improve the latent space alignment of 3D and text encoders without the need for expensive joint training. To achieve this, we introduce a simple but effective approach that combines Canonical Correlation Analysis (CCA) and existing alignment approaches such as affine translation [36] and local CKA [37]. This approach and our subsequent analysis extend previous works aimed at comparing learned feature spaces, but shows that careful subspace alignment can reveal subtle but important similarities, which are otherwise obscured in global comparisons. We provide a visual overview of the proposed approach in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.74, + 0.465, + 0.755 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.759, + 0.483, + 0.821 + ], + "angle": 0, + "content": "1. We extend the analysis of vision-text uni-modal latent space alignment to 3D uni-modal encoders and text, highlighting the limited similarity between these latent spaces and the low efficacy of current alignment approaches." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.486, + 0.902 + ], + "angle": 0, + "content": "2. We propose an efficient approach for cross-modal alignment between text and 3D features that combines CCA for subspace selection and existing alignment methods. This approach improves alignment performance with minimal computational overhead, as demonstrated through" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.759, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.461, + 0.909, + 0.521 + ], + "angle": 0, + "content": "experiments in matching and retrieval tasks. Our method establishes a baseline for 3D-text cross-modal understanding, providing an alternative to explicit multimodal pretraining for cross-modal tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.526, + 0.91, + 0.587 + ], + "angle": 0, + "content": "3. We observe a complementary structure between the spanned subspaces and the original feature spaces, enabling a distinction between geometric and semantic representations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.598, + 0.655, + 0.614 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Multi-modal representation learning. Multimodal representation learning has surged in recent years, driven by the success of image-language models [25, 30, 47, 52] that enable seamless cross-modal applications. These models serve as the backbone for tasks spanning from text-based image retrieval to generating high-quality visuals from natural language prompts. By aligning visual and linguistic features in a shared latent space, these models have paved the way for advanced interactions between modalities, setting the stage for applications where visual and textual information are jointly processed and understood." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Building on these advancements, vision-language models have recently been adapted for 3D point cloud representation, where 3D-image-text triplets [32, 63, 64, 69] enable contrastive pre-training. These models leverage powerful techniques such as momentum contrast [19] to align representations across modalities, allowing point cloud encoders to be pre-trained in a multimodal context. Additionally, tech-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19826" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "niques that apply vision-language models directly to 2D projections of point clouds [68, 70] have expanded cross-modal applications to 3D, including zero-shot shape classification, where models trained on 2D-image-text data demonstrate strong performance on 3D-related tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.484, + 0.352 + ], + "angle": 0, + "content": "In the domain of multimodal synthesis, recent efforts in 2D and 3D model generation have leveraged diffusion models to tackle complex generation tasks. These models use robust priors, often trained on vast datasets, which allow them to create high-fidelity outputs in scenarios such as novel view synthesis and realistic 3D reconstructions from single RGB images [33, 65]. By combining synthetic data with diffusion-based priors, these frameworks achieve impressive zero-shot generalization and geometry-consistent 3D synthesis, opening new avenues for applications where generating realistic 3D content from limited information is critical." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.363, + 0.485, + 0.771 + ], + "angle": 0, + "content": "Representational similarity. The study of representation similarity across neural networks has seen a significant rise in interest, spurred largely by seminal works originating in neuroscience and computational cognitive science. These fields have long been invested in understanding the nature and alignment of cognitive representations, providing a foundational basis that has influenced the current trajectory in machine learning [42, 56]. Based on these insights, various works [1, 31, 40, 41, 43, 44, 59] provide evidence of an intrinsic connection between independently trained networks; the similarity is especially notable among large-scale models, with works such as [39, 40, 55] exploring the phenomenon. In the computer vision and pattern recognition areas, the line of work on similarity-based representations, pioneered by Duin and Pekalska [45], has also been seminal. This line of research, which examines data through the lens of similarity rather than feature attributes, has provided robust frameworks for classification and clustering [5], enabling models to generalize across patterns and variations in complex datasets. Although mostly empirical in nature, these observations find theoretical support in the study of harmonics in neural networks weights [38], Independent Component Analysis [22, 23, 26, 50] and Independent Mechanism Analysis [13, 14, 54]. These works suggest that, when capturing the same underlying data generative factors, deep learning models may converge towards similar structures despite their complexity and non-linearity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Latent space alignment. More recently, a range of approaches has been developed to align latent spaces within the same modality [3, 7, 12], as well as across different modalities [36, 58, 67], particularly between visual and textual domains. Techniques such as Procrustes analysis [60] and several similarity metrics [27, 53, 57, 61], including centered kernel alignment (CKA) [8, 28, 37], have proven instrumental in aligning representations. These methods offer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "various strategies for quantifying similarity between feature spaces, allowing us to examine cross-modal interactions at a deeper level. HHowever, these methods often focus on aligning entire latent spaces, potentially missing meaningful similarities confined to specific subspaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.909, + 0.289 + ], + "angle": 0, + "content": "Our approach builds upon CCA [20, 49], a pivotal tool in pattern recognition and multi-view learning applications [16, 51]. This technique identifies maximally correlated subspaces, enabling more refined alignment across modalities by isolating core, mutually relevant components. Leveraging CCA, alignment methods can extend into complex domains, such as connecting 3D and textual latent spaces, as we demonstrate in the following." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.301, + 0.605, + 0.317 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.402 + ], + "angle": 0, + "content": "We compare the similarity between latent (feature) spaces of various 3D and text encoders, introducing a new approach that builds on existing alignment methods to improve their effectiveness. Below, we outline the tools used to measure and align these latent spaces." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.654, + 0.424 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Centered Kernel Alignment. CKA is a similarity measure frequently used in recent studies [8, 28] to compare representations in neural network feature spaces. Given feature matrices \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times p}\\) and \\(\\mathbf{Y} \\in \\mathbb{R}^{n \\times q}\\), we apply kernel functions \\(k\\) and \\(l\\) to obtain kernel representations \\(\\mathbf{K} = k(\\mathbf{X}, \\mathbf{X}) \\in \\mathbb{R}^{n \\times n}\\) and \\(\\mathbf{L} = l(\\mathbf{Y}, \\mathbf{Y}) \\in \\mathbb{R}^{n \\times n}\\). CKA is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.549, + 0.907, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C K A} (\\mathbf {K}, \\mathbf {L}) = \\frac {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {L})}{\\sqrt {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {K}) \\operatorname {H S I C} (\\mathbf {L} , \\mathbf {L})}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.592, + 0.906, + 0.622 + ], + "angle": 0, + "content": "where HSIC is the Hilbert-Schmidt Independence Criterion [15] and can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.63, + 0.907, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\operatorname {H S I C} (\\mathbf {K}, \\mathbf {L}) = \\frac {1}{(n - 1) ^ {2}} \\operatorname {t r} (\\mathbf {K H L H}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.67, + 0.672, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathbf{H} = \\mathbf{I} - \\frac{1}{n}\\mathbf{1}\\mathbf{1}^{\\top}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.909, + 0.737 + ], + "angle": 0, + "content": "Canonical Correlation Analysis. CCA [20, 49] is a statistical method that finds linear projections of two datasets, maximizing their correlation in a shared latent space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.909, + 0.829 + ], + "angle": 0, + "content": "Formally, given two sets of zero-centered variables \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times d_1}\\) and \\(\\mathbf{Y} \\in \\mathbb{R}^{n \\times d_2}\\), CCA seeks two projection matrices \\(\\mathbf{W}_{\\mathbf{X}} \\in \\mathbb{R}^{d_1 \\times k}\\) and \\(\\mathbf{W}_{\\mathbf{Y}} \\in \\mathbb{R}^{d_2 \\times k}\\) that map \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\) into a common \\(k\\)-dimensional space, maximizing the correlation between the projections \\(\\mathbf{X}\\mathbf{W}_{\\mathbf{X}}\\) and \\(\\mathbf{Y}\\mathbf{W}_{\\mathbf{Y}}\\). The optimization can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.838, + 0.907, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbf {W} _ {\\mathbf {X}}, \\mathbf {W} _ {\\mathbf {Y}}} \\operatorname {c o r r} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {X}}, \\mathbf {Y} \\mathbf {W} _ {\\mathbf {Y}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\(\\operatorname{corr}(\\cdot, \\cdot)\\) represents the correlation between the projected variables." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19827" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.303, + 0.108 + ], + "angle": 0, + "content": "3.2. Alignment approaches" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.204 + ], + "angle": 0, + "content": "Recent developments in latent space alignment have introduced various methods. In this work, we examine both the affine transformation approach in [36] and the CKA-based matching approach from [37], observing their limited effectiveness in 3D-text latent space alignment, and propose a method to address these limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.21, + 0.486, + 0.378 + ], + "angle": 0, + "content": "Latent Space Translation through Affine Transformation [36]. It is possible to estimate an affine transformation \\( T \\) that maps one latent space \\( \\mathcal{X} \\) onto another latent space \\( \\mathcal{Y} \\) such that \\( T(\\mathbf{x}) = \\mathbf{R}\\mathbf{x} + \\mathbf{b}, \\forall \\mathbf{x} \\in \\mathcal{X} \\). To compute \\( T \\), we assume access to an anchor subset consisting of ground-truth paired samples from both latent spaces. These anchors enable us to determine \\( T \\) by optimizing through gradient descent or, if \\( \\mathbf{b} = \\mathbf{0} \\) by minimizing the least squares error. It assumes that \\( \\mathcal{X} \\) and \\( \\mathcal{Y} \\) share the same dimensionality and are normalized to zero mean and unit variance. These constraints can be enforced by zero-padding the smaller latent space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.485, + 0.489 + ], + "angle": 0, + "content": "Local CKA-based retrieval and matching [37]. The CKA metric computed between two sets of features is sensitive to ordering and maximized when ground-truth pairs are aligned, this insight can be used to match unseen data by including them in a well-aligned anchors set. Formally, starting from aligned set of features \\(\\mathbf{X}_A\\) and \\(\\mathbf{Y}_A\\), we compute for a query pair \\((\\mathbf{x}_q,\\mathbf{y}_q)\\) its local CKA defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.501, + 0.484, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o c a l C K A} \\left(\\mathbf {x} _ {q}, \\mathbf {y} _ {q}\\right) = \\operatorname {C K A} \\left(\\mathbf {K} _ {\\left[ \\mathbf {X} _ {A}, \\mathbf {x} _ {q} \\right]}, \\mathbf {K} _ {\\left[ \\mathbf {Y} _ {A}, \\mathbf {y} _ {q} \\right]}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.486, + 0.622 + ], + "angle": 0, + "content": "where \\([\\mathbf{X},\\mathbf{x}]\\) denotes the column-wise concatenation of matrix \\(\\mathbf{X}\\) and vector \\(\\mathbf{x}\\). Local CKA calculates similarity in a pairwise manner, accounting for anchor alignment while being sensitive to ordering among query pairs. We introduce all possible query pairs, where correctly matched pairs exhibit the highest local CKA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.267, + 0.647 + ], + "angle": 0, + "content": "3.3. Proposed method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.485, + 0.805 + ], + "angle": 0, + "content": "Our goal is to reliably align the latent spaces of pre-trained 3D and text encoders. Given a dataset of \\( n \\) caption-point cloud pairs, each embedded in the corresponding feature space, and represented as matrices \\( \\mathbf{X} \\in \\mathbb{R}^{n \\times p} \\) and \\( \\mathbf{Y} \\in \\mathbb{R}^{n \\times q} \\) respectively, we first select a subset of anchor pairs that will guide our translation process, denoted as \\( \\mathbf{X}_A \\in \\mathbb{R}^{n_A \\times p} \\) and \\( \\mathbf{Y}_A \\in \\mathbb{R}^{n_A \\times q} \\). Building upon the mathematical foundations introduced in Section 3.2, we develop a pipeline that combines dimensionality reduction with the previously introduced alignment methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Common Subspace Projection. In our work, we show that 3D and text latent spaces can effectively be aligned in low-dimensional connected subspaces. We begin by applying CCA to the anchor pairs to identify a shared \\(k\\)-dimensional subspace (with \\(k < p, q\\)) that connects point cloud and text latent spaces. This yields projection matrices \\(\\mathbf{W}_{\\mathbf{X}^A} \\in \\mathbb{R}^{p \\times k}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.907, + 0.122 + ], + "angle": 0, + "content": "and \\(\\mathbf{W}_{\\mathbf{Y}^A}\\in \\mathbb{R}^{q\\times k}\\). All samples are then projected into this reduced space:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.135, + 0.907, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {r} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {X} _ {A}}, \\quad \\mathbf {Y} ^ {r} = \\mathbf {Y} \\mathbf {W} _ {\\mathbf {Y} _ {A}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.165, + 0.907, + 0.195 + ], + "angle": 0, + "content": "In particular, we refer to the reduced anchors as \\(\\mathbf{X}_A^r\\in \\mathbb{R}^{n_A\\times k}\\) and \\(\\mathbf{Y}_A^r\\in \\mathbb{R}^{n_A\\times k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.195, + 0.907, + 0.255 + ], + "angle": 0, + "content": "Our experiments show that projecting 3D and text latent spaces into a lower-dimensional, shared subspace improves alignment by isolating features that are highly correlated across modalities" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.261, + 0.909, + 0.352 + ], + "angle": 0, + "content": "Alignment of projected latent spaces. Given the projected latent spaces, we can apply either of the previously described alignment methods in the reduced space. For the affine transformation approach, we learn a mapping between \\(\\mathbf{X}^r\\) and \\(\\mathbf{Y}^r\\) using the projected anchor pairs \\(\\mathbf{X}_A^r\\), \\(\\mathbf{Y}_A^r\\) to optimize the transformation parameters \\(\\mathbf{R}\\) and \\(\\mathbf{b}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.365, + 0.907, + 0.382 + ], + "angle": 0, + "content": "\\[\nT \\left(\\mathbf {X} ^ {r}\\right) = \\mathbf {R X} ^ {r} + \\mathbf {b}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.909, + 0.44 + ], + "angle": 0, + "content": "Alternatively, we can employ the local CKA-based matching approach in the projected space. For a query pair \\((\\mathbf{x}_q^r,\\mathbf{y}_q^r)\\) we compute its local CKA using the projected anchor sets:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.468, + 0.907, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o c a l C K A} \\left(\\mathbf {x} _ {q} ^ {r}, \\mathbf {y} _ {q} ^ {r}\\right) = \\operatorname {C K A} \\left(\\mathbf {K} _ {\\left[ \\mathbf {X} _ {A} ^ {r}, \\mathbf {x} _ {q} ^ {r} \\right]}, \\mathbf {K} _ {\\left[ \\mathbf {Y} _ {A} ^ {r}, \\mathbf {y} _ {q} ^ {r} \\right]}\\right), \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.704, + 0.53 + ], + "angle": 0, + "content": "4. Experimental setup" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.537, + 0.709, + 0.554 + ], + "angle": 0, + "content": "4.1. Pre-training Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.909, + 0.756 + ], + "angle": 0, + "content": "Prior works in point cloud pre-training, particularly within uni-modal frameworks, have relied heavily on ShapeNet [4], a dataset of 51,300 annotated 3D synthetic shapes spanning 55 categories. ShapeNet has been instrumental in advancing foundational methods, yet it remains limited by the relatively narrow scope of categories. With the release of Objaverse [9], which includes over 800,000 shapes across diverse real-world categories, a new standard for large-scale representation learning in the 3D domain has emerged. Objaverse's extensive shape diversity makes it ideal for both uni-modal and multi-modal learning. Despite these advantages, there is a lack of works that explore uni-modal pre-training specifically on Objaverse." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.767, + 0.625, + 0.782 + ], + "angle": 0, + "content": "4.2.Encoders" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.907, + 0.821 + ], + "angle": 0, + "content": "We explore both multi-modal and uni-modal 3D and text encoders across varying levels of model complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Multi-modal 3D Encoders. For multi-modal pre-training, we use OpenShape, ULIP-2 and Uni3D [32] pre-trained models, trained on point cloud, image, and text triplets of Objverse with a contrastive pre-text task to align 3D encoders with frozen CLIP encoders. We adopt the Point-BERT-based" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19828" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.088, + 0.907, + 0.293 + ], + "angle": 0, + "content": "
Method3D EncoderCLIPRoBERTaBERT
Matching accuracyTop-5 retrievalMatching accuracyTop-5 retrievalMatching accuracyTop-5 retrieval
Multi-modal 3D encoder
Affine + Subspace ProjectionOpenShape67.685.655.275.845.070.6
Affine + Subspace ProjectionULIP-265.685.247.270.635.259.8
Affine + Subspace ProjectionUni3D61.481.645.863.834.447.2
Uni-modal 3D encoder
AffinePointBert15.823.67.815.66.413.4
AffineSparseConv11.034.46.020.04.216.2
AffinePointnet18.421.88.010.29.612.2
Affine + Subspace Projection (Ours)PointBert30.842.223.228.415.618.2
Affine + Subspace Projection (Ours)SparseConv21.445.619.216.815.815.0
Affine + Subspace Projection (Ours)Pointnet25.236.622.420.816.616.0
Local CKAPointBert5.815.21.81.41.04.39
Local CKASparseConv3.413.61.791.60.63.8
Local CKAPointnet6.618.02.42.01.05.0
Local CKA + Subspace Projection (Ours)PointBert29.460.1917.042.415.037.2
Local CKA + Subspace Projection (Ours)SparseConv19.056.915.834.010.030.8
Local CKA + Subspace Projection (Ours)Pointnet26.853.018.640.214.338
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.303, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Table 1. Matching and retrieval performance across 3D and text encoders using different alignment approaches. We use 30,000 anchors for subspace projection and affine transformation approaches, and 1,000 anchors for local CKA. A query set of 500 is uniformly sampled, with results averaged over 3 different seeds. The subspace dimension is fixed at 50. Our approach (Ours) consistently demonstrates improved matching and retrieval performance, with multi-modal 3D encoders setting the upper bound for performance. Additional top-\\(k\\) retrieval metrics are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.44 + ], + "angle": 0, + "content": "variant of each model [66]. We primarily focus on OpenShape for simplicity but generalize results to ULIP-2 and Uni3D." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Uni-modal 3D Encoders. For the uni-modal setup, we pre-train PointBERT on Objaverse using its original pretext tasks, which include masked point reconstruction and a uni-modal contrastive loss to encourage robust shape representation. We also explore two additional architectures: a Sparse convolution (MinkowskiNet [6]) model and a simpler architecture in PointNet [46], each pre-trained using a shape-level contrastive learning method [18]. This approach contrasts different partial views of input shapes. Across all 3D encoders in this setup, we fix the latent dimension at 512 to maintain consistency in representation space comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.711 + ], + "angle": 0, + "content": "Text encoders. We use the text encoder from OpenCLIP ViTbigG-14 [24], chosen to match text encoder used in OpenShape. Additionally, we examine alignment with purely unimodal text encoders by including BERT [11] and RoBERTa [34], we also evaluate the alignment with T5 [48] in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Across all pre-training setups, parameters are kept consistent to facilitate direct comparisons. We detail the technical details in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.268, + 0.786 + ], + "angle": 0, + "content": "4.3. Downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate our alignment approach on Objaverse-LVIS [17], a human-verified test subset of Objaverse that contains 1,156 object categories. This subset is specifically reserved for evaluation and is unseen during pre-training. The captions are generated with Cap3D [35], which provides enhanced descriptive text for each 3D shape. Our evaluation framework consists of two main tasks: matching and retrieval. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.562 + ], + "angle": 0, + "content": "matching task involves finding the correct permutation of captions for perfect matching given a shuffled set of query images and their corresponding captions; we utilize the linear sum assignment approach to perform this task [29]. For the retrieval task, the model must identify the correct 3D object from the query set based on a text caption. These tasks are particularly effective in measuring the cross-modal capabilities of encoders, and have been evaluated in prior Vision-Text studies [37]. While our main results emphasize the matching and top-5 retrieval tasks, we provide evaluation of top-1 and top-10 retrieval metrics in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.58, + 0.6, + 0.594 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.859, + 0.622 + ], + "angle": 0, + "content": "5.1. Are 3D and Text Latent Spaces similar?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.691 + ], + "angle": 0, + "content": "To evaluate the inherent similarity between 3D and text feature spaces without alignment, we compute linear CKA scores for both unimodal and multimodal encoders. The results are shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.848 + ], + "angle": 0, + "content": "Comparison of 3D-Text and Vision-Text alignment. Prior work [21] reports CKA alignment values ranging from approximately \\(30\\%\\) to \\(48\\%\\) between different uni-modal vision and text encoders (see Figure 13 in [21]). In stark contrast, we find that the default alignment between 3D and text latent spaces is significantly weaker, with a maximum score of 0.12 observed for the uni-modal PointBERT and CLIP pair. This substantial gap underscores a key insight: unlike vision and text encoders, 3D encoders did not converge to structures similar to text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Alignment favors multi-modality. We observe that 3D multi-modal encoders demonstrate the highest CKA score with text encoders, particularly with CLIP's text encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.482, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.488, + 0.434 + ], + "angle": 0, + "content": "Figure 2. Linear CKA scores between text and 3D encoders without alignment. Higher scores reflect stronger alignment between encoder pairs, with the strongest alignment observed between multi-modal 3D encoders (OpenShape) and CLIP text encoder due to their shared training on aligned representations. Uni-modal 3D encoders show significantly lower alignment with text encoders, although slightly higher with the CLIP text encoder." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.452, + 0.482, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.693, + 0.486, + 0.737 + ], + "angle": 0, + "content": "Figure 3. Linear CKA scores between text and 3D encoders after affine translation. Affine translation results in a consistent increase in similarity for both 3D-to-text and text-to-3D directions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.487, + 0.904 + ], + "angle": 0, + "content": "This behavior is expected, given that 3D multi-modal encoders are explicitly trained to align with CLIP latent spaces, which share the same textual modality. Even among unimodal 3D encoders, alignment with the CLIP text encoder is notably higher (e.g. 0.12 for PointBERT-CLIP) compared to alignment with uni-modal text encoders like RoBERTa (e.g., 0.04 for PointBERT-RoBERTa). This suggests that the visual understanding embedded in CLIP's text encoder extends beyond image and text domains to include 3D representations," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.108 + ], + "angle": 0, + "content": "despite the lack of explicit alignment during pre-training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.119, + 0.792, + 0.135 + ], + "angle": 0, + "content": "5.2. Latent Space Alignment results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.142, + 0.909, + 0.295 + ], + "angle": 0, + "content": "We evaluate the performance of the alignment approaches outlined in Sec. 3 using matching and retrieval tasks, and analyze their effectiveness in aligning 3D and text encoders. Unless otherwise specified, we fix the subspace dimension to \\( d = 50 \\) and the number of anchors to 30,000. For downstream tasks, we uniformly sample a query set of size 500 and average results over 3 different seeds. For the affine transformation approach, we present results for the text-to-3D direction, noting that similar performance is achieved in the 3D-to-text direction as seen in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.911, + 0.496 + ], + "angle": 0, + "content": "Existing approaches enable limited alignment We first assess whether previous successful approaches—affine transformation and local CKA—can achieve meaningful 3D-text alignment. As shown in Tab. 1, both methods yield modest alignment improvements over the unaligned baseline, where uni-modal 3D encoders start near zero in matching and retrieval tasks. These results suggest a small alignment shift. However, even with this improvement, alignment remains significantly lower than the alignment achieved in vision-text benchmarks [37]. This finding hints at the limits of uni-modal 3D encoders in achieving similar Vision-Text alignment performance, and might necessitate a different approach to align their latent space with text encoders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.835 + ], + "angle": 0, + "content": "Importance of subspace projection (Ours). While aligning the latent spaces of 3D and text encoders achieved some success with existing methods, results remained well below vision-text alignment benchmarks. Motivated by this limitation, we propose aligning lower-dimensional subspaces, based on the hypothesis that 3D and text representations might intersect within a shared latent subspace. Using a validation set, we report in Fig. 4 the impact of subspace dimension, obtained through our CCA approach (see Sec. 3.3) combined with the affine transformation on the top-5 retrieval accuracy of uni-modal PointBert. While affine alignment alone yields better performance at higher dimensions, our subspace projection approach significantly outperforms it when the subspace dimension is reduced. This finding indicates that alignment quality is optimized within carefully chosen lower-dimensional spaces, reinforcing our hypothesis that 3D-text alignment is more successful within targeted subspaces. The results in Tab. 1 further validate this approach, showing a substantial increase in matching and retrieval performance across all 3D and text encoder pairs, outperforming both alignment approaches when they operate on the original latent space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.912, + 0.903 + ], + "angle": 0, + "content": "Multi-modal encoders as an upper bound of performance. We include 3D multi-modal encoders for two main reasons: (1) As an upper bound for alignment performance, (2) To study how alignment degrades across different text" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19830" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "encoders, including those that were not used during pretraining. In this context, the inclusion of 'Affine + Subspace Projection' multi-modal results in Tab. 1 shows how using our approach significantly narrows the gap between multimodal and uni-modal performance. However, to illustrate point 1. above, we also measured with cosine similarity the alignment of OpenShape (a multi-modal 3D encoder) and the same CLIP text encoder used during pre-training, achieving 0.94 for top-5 retrieval accuracy, an expected result since 3D and text encoders are explicitly aligned during pre-training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.473 + ], + "angle": 0, + "content": "3D encoder complexity's low impact on alignment. Point-BERT outperforms the more complex SparseConv among uni-modal 3D encoders, suggesting that increasing model complexity alone does not guarantee improved alignment in 3D-text tasks. Surprisingly, even PointNet—a relatively simple model—achieves similar alignment scores, showing that factors other than model complexity may play a pivotal role in 3D-text alignment. This observation contrasts with vision-text alignment results [21, 37], where complex models typically leverage large datasets more effectively. Our findings thus indicate a distinctive aspect of 3D-text alignment: model simplicity does not impede, and may even aid, the interpretability and compatibility of learned features for cross-modal alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Different alignment techniques have different strengths. Our experiments reveal that different alignment techniques offer complementary strengths across tasks. For instance, affine transformation proves particularly effective for matching tasks across all 3D encoders, while local CKA shows superior performance in top-5 retrieval accuracy. This suggests that while some methods excel in precision tasks, others might better capture broader semantic nuances, making them more suitable for retrieval. Together, these observations imply that a hybrid approach could leverage the unique strengths of each method, opening up promising directions for future cross-modal applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Scaling of our approach. In Fig. 5, we explore the scalability of our approach by analyzing how performance responds to increasing the number of anchors. Notably, our subspace projection method scales effectively with anchor count, reaching a plateau before requiring the full dataset (over 800,000 shapes). This scalability highlights the approach's efficiency in learning robust 3D-text mappings with a limited subset of anchor pairs. However, we observe a leveling off in performance gains beyond a certain anchor count, likely due to the constraints imposed by the low-dimensional subspace. This suggests that while our approach is computationally efficient and data-efficient, its reliance on a fixed lower-dimensional subspace could limit its adaptability to larger, more diverse datasets." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.709, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.092, + 0.904, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.2, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Impact of subspace dimensionality on retrieval performance. Comparison of three approaches: our proposed CCA + affine translation method (blue), affine translation without subspace projection (red), and baseline feature space alignment without transformation (orange). Results are shown for the uni-modal PointBERT and CLIP text encoder, with generalizations to other encoders provided in the supplementary." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.324, + 0.9, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.495, + 0.909, + 0.565 + ], + "angle": 0, + "content": "Figure 5. Effect of anchor set size on retrieval performance. Validation set results with different subspace dimensions show that retrieval performance improves as the anchor subset size increases but eventually reaches a plateau. Results are shown for the unimodal PointBERT and CLIP text encoder." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.592, + 0.899, + 0.777 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.794, + 0.908, + 0.864 + ], + "angle": 0, + "content": "Figure 6. Pearson correlation between shape query chamfer distances and pairwise distances in the projected text latent subspace. We observe a higher Pearson correlation in the projected text latent subspace with optimal subspace dimension. Results are shown for the uni-modal PointBERT and CLIP text encoders." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19831" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.48, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.563, + 0.486, + 0.649 + ], + "angle": 0, + "content": "Figure 7. Shape retrieval comparison between original and reduced latent spaces. For a given query shape, we retrieve the closest match based on cosine similarity. Results demonstrate a higher semantic understanding in the reduced 3D latent subspace compared to the original latent space. Shown for uni-modal Point-BERT and CLIP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.325, + 0.691 + ], + "angle": 0, + "content": "5.3. Geometries vs. semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.486, + 0.79 + ], + "angle": 0, + "content": "Our quantitative evaluation shows that low-dimensional subspace projection significantly improves latent space alignment in the 3D-Text setting. To better understand the characteristics of these subspaces spanned relative to the original spaces, we analyze the increase in semantic and geometric knowledge within the projected spaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Increased geometric awareness of the text latent subspace. To quantify the increase in geometric awareness within the projected text subspace, we compute the Pearson correlation between the Chamfer distances of a query set of 500 shapes and the pairwise distances between feature vectors within the text subspace and compare these results to those obtained from the original text latent space. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.153 + ], + "angle": 0, + "content": "in Fig. 6 show that the optimal subspace dimension yields a high correlation with Chamfer distances, signaling an increase in the sensitivity of the text encoder to geometric properties when projected into this specific subspace." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.16, + 0.909, + 0.372 + ], + "angle": 0, + "content": "Increased semantic understanding of the 3D latent subspace. To assess the improved semantic capacity of the 3D subspace, we conducted a qualitative analysis of shape retrieval performance in both the reduced and the original latent spaces (Fig. 7). Given a set of query shapes, we observe that the reduced subspace retrieves shapes with stronger semantic similarity (e.g., retrieving animals for a query sheep shape), while the original latent space primarily favors geometric resemblance (e.g. retrieving stacked square shapes when querying a '3D model of a CPU'). This shift suggests that our subspace projection method enables the 3D encoder to capture semantic relationships that are absent in the full latent space. This effectively enables cross-modal applications and explains its higher matching and retrieval performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.89, + 0.407 + ], + "angle": 0, + "content": "6. Conclusion, Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.909, + 0.657 + ], + "angle": 0, + "content": "In this work, we present the first study investigating latent space alignment between 3D and text pre-trained encoders. Building on the hypothesis that these modalities share semantic connections within lower-dimensional subspaces, we propose an effective approach combining CCA projection with affine transformation estimation to translate between modalities' latent spaces. Our empirical results show that optimal cross-modal performance is achieved through low-dimensional subspace projection, and our method successfully improves alignment across diverse 3D and text uni-modal encoders. While CLIP-based multi-modal encoders establish performance upper bounds, we enable significant cross-modal capabilities in uni-modal encoders previously limited to single-modality tasks. We also demonstrate that semantic understanding can be extracted from geometry-aware latent spaces of uni-modal 3D encoders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Although our work focused on Objaverse, which is the first large-scale 3D dataset, it would be interesting to consider how scaling on Objaverse-XL [10] would affect the alignment quality between 3D and text encoders. Moreover, in this work we do not distinguish object-level vs. scene-level annotations, and decomposing objects or scenes into their composing blocks could shed light onto the compositionality of the learned representations. Finally, the subspace alignment method that we introduce in this work can be broadly applicable to other representations as well. In the future, we plan to use it to investigate the limitations of alignment observed in other representations. In particular, even when trained at significantly higher data scales, images and text representations do not align perfectly [21], and it would be interesting to reveal the unique and complementary nature of different modalities via subspace analysis." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19832" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, as well as gifts from Ansys and Adobe Research. This work was also supported by the Galileo 2022 fellowship from the Università Italo Francese/ Université Franco Italiène (UIF/UFI) within the project G22_4 titled “Multimodal Artificial Intelligence for 3D shape analysis, modeling and applications”." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.241, + 0.188, + 0.258 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.266, + 0.484, + 0.308 + ], + "angle": 0, + "content": "[1] Lisa Bonheme and Marek Grzes. How do variational autoencoders learn? insights from representational similarity. arXiv preprint arXiv:2205.08399, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.484, + 0.377 + ], + "angle": 0, + "content": "[2] Irene Cannistraci, Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, and Emanuele Rodola. Bootstrapping parallel anchors for relative representations. In The First Tiny Papers Track at ICLR 2023, Tiny Papers at ICLR 2023, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.38, + 0.484, + 0.448 + ], + "angle": 0, + "content": "[3] Irene Cannistraci, Luca Moschella, Marco Fumero, Valentino Maiorca, and Emanuele Rodolà. From bricks to bridges: Product of invariances to enhance latent space communication. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[4] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.521, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[5] Yihua Chen, Eric K. Garcia, Maya R. Gupta, Ali Rahimi, and Luca Cazzanti. Similarity-based classification: Concepts and algorithms. Journal of Machine Learning Research, 10(27): 747-776, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[7] Donato Crisostomi, Irene Cannistraci, Luca Moschella, Pietro Barbiero, Marco Ciccone, Pietro Lio, and Emanuele Rodola. From charts to atlas: Merging latent spaces into one. In NeurIPS 2023 Workshop on Symmetry and Geometry in Neural Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. 2023 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.266, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. arXiv preprint arXiv:2307.05663, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.189 + ], + "angle": 0, + "content": "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[12] Marco Fumero, Marco Pegoraro, Valentino Maiorca, Francesco Locatello, and Emanuele Rodolà. Latent functional maps: a spectral framework for representation alignment. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[13] Shubhangi Ghosh, Luigi Gresele, Julius von Kugelgen, Michel Besserve, and Bernhard Scholkopf. Independent mechanism analysis and the manifold hypothesis, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[14] Luigi Gresele, Julius Von Kugelgen, Vincent Stimper, Bernhard Scholkopf, and Michel Besserve. Independent mechanism analysis, a new concept? In Advances in Neural Information Processing Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[15] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.907, + 0.455 + ], + "angle": 0, + "content": "[16] Chenfeng Guo and Dongrui Wu. Canonical correlation analysis (cca) based multi-view learning: An overview. arXiv preprint arXiv:1907.01693, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.906, + 0.511 + ], + "angle": 0, + "content": "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[18] Souhail Hadgi, Lei Li, and Maks Ovsjanikov. To supervise or not to supervise: Understanding and addressing the key challenges of 3d transfer learning. arXiv preprint arXiv:2403.17869, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.907, + 0.621 + ], + "angle": 0, + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arxiv e-prints, art. arXiv preprint arXiv:1911.05722, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.907, + 0.665 + ], + "angle": 0, + "content": "[20] Harold Hotelling. Relations between two sets of variates. In Breakthroughs in statistics: methodology and distribution, pages 162-190. Springer, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[21] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987, 2024. 1, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[22] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA. In Advances in Neural Information Processing Systems. Curran Associates, Inc. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[23] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, pages 859–868. PMLR. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19833" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.174 + ], + "angle": 0, + "content": "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[26] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[27] Max Klabunde, Tobias Schumacher, Markus Strohmaier, and Florian Lemmerich. Similarity of neural network models: A survey of functional and representational measures, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[28] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[30] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[31] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John Hopcroft. Convergent learning: Do different neural networks learn the same representations? arXiv preprint arXiv:1511.07543, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[32] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in neural information processing systems, 36, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[34] Yinhan Liu. *Roberta: A robustly optimized bert pretraining approach.* arXiv preprint arXiv:1907.11692, 364, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[35] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[36] Valentino Maiorca, Luca Moschella, Antonio Norelli, Marco Fumero, Francesco Locatello, and Emanuele Rodola. Latent space translation via semantic alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[37] Mayug Maniparambili, Raiymbek Akshulakov, Yasser Abdelaziz Dahou Djilali, Mohamed El Amine Seddik, Sanath Narayan, Karttikeya Mangalam, and Noel E O'Connor. Do vision and language encoders represent the world similarly? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14334-14343, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Giovanni Luca Marchetti and Christopher Hillar. Harmonics of learning: Universal fourier features emerge" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.909, + 0.134 + ], + "angle": 0, + "content": "in invariant networks. https://synthical.com/article/03032df5-f9c5-43a6-8253-d4af1d67e0d4, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[39] Raghav Mehta, Vítor Albiero, Li Chen, Ivan Evtimov, Tamar Glaser, Zhiheng Li, and Tal Hassner. You only need a good embeddings extractor to fix spurious correlations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[40] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[41] Ari S. Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, page 5732-5741, Red Hook, NY, USA, 2018. Curran Associates Inc. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[42] Luca Moschella. Latent communication in artificial neural networks, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.908, + 0.398 + ], + "angle": 0, + "content": "[43] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In The Eleventh International Conference on Learning Representations, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.908, + 0.468 + ], + "angle": 0, + "content": "[44] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. Advances in Neural Information Processing Systems, 36:15303-15319, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.908, + 0.51 + ], + "angle": 0, + "content": "[45] E. Pekalska and Robert Duin. Automatic pattern recognition by similarity representations. *Electronics Letters*, 37:159 - 160, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.908, + 0.579 + ], + "angle": 0, + "content": "[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.908, + 0.663 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.908, + 0.801 + ], + "angle": 0, + "content": "[49] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[50] Geoffrey Roeder, Luke Metz, and Durk Kingma. On Linear Identifiability of Learned Representations. In Proceedings of the 38th International Conference on Machine Learning, pages 9030-9039. PMLR. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.861, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[51] Jan Rupnik and John Shawe-Taylor. Multi-view canonical correlation analysis. In Conference on data mining and data warehouses (SiKDD 2010), pages 1-4, 2010. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19834" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[53] Mahdiyar Shahbazi, Ali Shirali, Hamid Aghajan, and Hamed Nili. Using distance on the riemannian manifold to compare representations in brain and in models. NeuroImage, 239: 118271, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[54] Joanna Sliwa, Shubhangi Ghosh, Vincent Stimper, Luigi Gresele, and Bernhard Scholkopf. Probing the robustness of independent mechanism analysis for representation learning. In UAI 2022 Workshop on Causal Representation Learning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[55] Gowthami Somepalli, Liam Fowl, Arpit Bansal, Ping Yeh-Chiang, Yehuda Dar, Richard Baraniuk, Micah Goldblum, and Tom Goldstein. Can neural nets learn the same model twice? investigating reproducibility and double descent from the decision boundary perspective. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13689-13698, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[56] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. Love, Erin Grant, Iris Groen, Jascha Achterberg, Joshua B. Tenenbaum, Katherine M. Collins, Katherine L. Hermann, Kerem Oktar, Klaus Greff, Martin N. Hebart, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas P. O'Connell, Thomas Unterthiner, Andrew K. Lampinen, Klaus-Robert Müller, Mariya Toneva, and Thomas L. Griffiths. Getting aligned on representational alignment, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[57] Shuai Tang, Wesley J. Maddox, Charlie Dickens, Tom Diethe, and Andreas Damianou. Similarity of neural networks with gradients, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[58] Thomas Theodoridis, Theocharis Chatzis, Vassilios Solachidis, Kosmas Dimitropoulos, and Petros Daras. Cross-modal variational alignment of latent spaces. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4127-4136, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[59] Ivan Vulic, Sebastian Ruder, and Anders Søgaard. Are all good word vector spaces isomorphic? arXiv preprint arXiv:2004.04070, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[60] Chang Wang and Sridhar Mahadevan. Manifold alignment using procrastes analysis. In Proceedings of the 25th international conference on Machine learning, pages 1120-1127, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[61] Alex H Williams, Erin Kunz, Simon Kornblith, and Scott Linderman. Generalized shape metrics on neural representations. In Advances in Neural Information Processing Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[62] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.26 + ], + "angle": 0, + "content": "[64] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27091–27101, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[65] Yuxuan Xue, Xianghui Xie, Riccardo Marin, and Gerard. Pons-Moll. Human 3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.907, + 0.372 + ], + "angle": 0, + "content": "[66] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, and Jie Zhou. Pre-training 3d point cloud transformers with masked point modeling. 2022 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19291-19300, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.907, + 0.444 + ], + "angle": 0, + "content": "[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18102-18112, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.907, + 0.514 + ], + "angle": 0, + "content": "[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.907, + 0.57 + ], + "angle": 0, + "content": "[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. arXiv preprint arXiv:2310.06773, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.906, + 0.627 + ], + "angle": 0, + "content": "[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 3(4), 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19835" + } + ] +] \ No newline at end of file diff --git a/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_origin.pdf b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb1b70fa2e930125dc73fbb6e16cd4fb2e5ef263 --- /dev/null +++ b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/85c228c2-88b3-4b42-91cd-0b9c6fff9635_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75e59fe6efc35bec59e6510b795bb436e939be20b7c97884c2fce341bc8503a4 +size 1957058 diff --git a/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/full.md b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18373096ace3c626ec5f2772530f97ad084152e1 --- /dev/null +++ b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/full.md @@ -0,0 +1,329 @@ +# Escaping Plato's Cave: Towards the Alignment of 3D and Text Latent Spaces + +Souhail Hadgi + +Luca Moschella $^{2,*}$ + +Andrea Santilli2 + +Diego Gomez + +Qixing Huang + +Emanuele Rodolà2 + +Simone Melzi3 + +Maks Ovsjanikov1 + +École polytechnique + +Sapienza University of Rome + +3University of Milano-Bicocca + +4The University of Texas at Austin + +# Abstract + +Recent works have shown that, when trained at scale, uni-modal 2D vision and text encoders converge to learned features that share remarkable structural properties, despite arising from different representations. However, the role of 3D encoders with respect to other modalities remains unexplored. Furthermore, existing 3D foundation models that leverage large datasets are typically trained with explicit alignment objectives with respect to frozen encoders from other representations. In this work, we investigate the possibility of a posteriori alignment of representations obtained from uni-modal 3D encoders compared to text-based feature spaces. We show that naive post-training feature alignment of uni-modal text and 3D encoders results in limited performance. We then focus on extracting subspaces of the corresponding feature spaces and discover that by projecting learned representations onto well-chosen lower-dimensional subspaces the quality of alignment becomes significantly higher, leading to improved accuracy on matching and retrieval tasks. Our analysis further sheds light on the nature of these shared subspaces, which roughly separate between semantic and geometric data representations. Overall, ours is the first work that helps to establish a baseline for posttraining alignment of 3D uni-modal and text feature spaces, and helps to highlight both the shared and unique properties of 3D data compared to other representations. + +# 1. Introduction + +Recent advances in multi-modal learning, particularly in vision-language models such as CLIP [47], have sparked interest in extending these successes to the 3D domain. Most current approaches primarily focus on training 3D encoders through triplet-based learning with pre-trained 2D vision and language encoders [32, 64, 69], leveraging new large-scale datasets such as Objaverse [9, 10], and showing promising + +results in tasks such as zero-shot shape recognition. + +While CLIP [47] was trained with explicit alignment objectives between text and image representations, recent work has observed that even when trained independently, vision and text encoders tend to exhibit significant similarities [21, 37]. In particular, the learned latent spaces of pure (uni-modal) text and vision encoders have similar proximity structure [21] and can be aligned relatively easily after training with a small number of known anchor correspondences [2, 37, 43]. Furthermore, the degree of similarity in learned features across modalities strongly correlates with the quality of performance in various downstream tasks. One prominent interpretation of these results is that given sufficient scale of training data and model complexity, different representations are converging to a shared underlying structure of the physical world. This has given rise to the Platonic Representation Hypothesis [21], where different representations are viewed as projections of reality on particular modalities. + +At the same time, as the physical world is inherently (at least) 3-dimensional, a natural question is how the structure of uni-modal vision or text encoders relates to the features learned directly from 3D data. This question raises several challenges: first, large-scale 3D datasets have only recently become available [9, 10]; second, most existing 3D "foundation models" are trained with explicit alignment objectives with respect to frozen 2D and text encoders, which limits the utility of post-training comparison. Finally, there is a lack of universally agreed-upon architectures and training objectives for 3D data, and many commonly-used architectures tend to have limited generalization power [62]. + +In this work, we initiate the first study on the relation between 3D and language representations. We formulate the task of post-training alignment between a range of 3D and text encoders and study the accuracy and utility of several alignment strategies. + +Our first insight is that when trained on pure 3D data with self-supervised objectives, 3D encoders tend to lead to representations that align only very weakly to text-based representations. We believe that this observation already + +![](images/46ffe3e9717f4e2077b73030cfc9ed91b0269e1ca365e2d139c2b0f60b639825.jpg) +Figure 1. A visual overview of the proposed approach. is illustrated. From left to right: We begin with two distinct input collections—one consisting of 3D shapes and the other of textual prompts. In the blue box, independent, frozen uni-modal encoders map each modality into separate, high-dimensional latent spaces, shown in the red box. A dimensionality reduction procedure is applied to project these learned spaces into low-dimensional subspaces, represented in the green box. Finally, an alignment method registers the two low-dimensional subspaces, enabling cross-modal applications such as shape retrieval, with examples depicted in the yellow box. + +Highlights the difficulty of uni-modal training and sheds light on the scarcity of "pure" 3D foundation models. + +Additionally, we reveal that while 3D and text latent spaces are not naturally aligned, effective cross-modal translation can be achieved through subspace projection and alignment. Our key insight is that by identifying and operating in correlated subspaces, we can improve the latent space alignment of 3D and text encoders without the need for expensive joint training. To achieve this, we introduce a simple but effective approach that combines Canonical Correlation Analysis (CCA) and existing alignment approaches such as affine translation [36] and local CKA [37]. This approach and our subsequent analysis extend previous works aimed at comparing learned feature spaces, but shows that careful subspace alignment can reveal subtle but important similarities, which are otherwise obscured in global comparisons. We provide a visual overview of the proposed approach in Fig. 1. + +To summarize, our main contributions are as follows: + +1. We extend the analysis of vision-text uni-modal latent space alignment to 3D uni-modal encoders and text, highlighting the limited similarity between these latent spaces and the low efficacy of current alignment approaches. +2. We propose an efficient approach for cross-modal alignment between text and 3D features that combines CCA for subspace selection and existing alignment methods. This approach improves alignment performance with minimal computational overhead, as demonstrated through + +experiments in matching and retrieval tasks. Our method establishes a baseline for 3D-text cross-modal understanding, providing an alternative to explicit multimodal pretraining for cross-modal tasks. + +3. We observe a complementary structure between the spanned subspaces and the original feature spaces, enabling a distinction between geometric and semantic representations. + +# 2. Related Work + +Multi-modal representation learning. Multimodal representation learning has surged in recent years, driven by the success of image-language models [25, 30, 47, 52] that enable seamless cross-modal applications. These models serve as the backbone for tasks spanning from text-based image retrieval to generating high-quality visuals from natural language prompts. By aligning visual and linguistic features in a shared latent space, these models have paved the way for advanced interactions between modalities, setting the stage for applications where visual and textual information are jointly processed and understood. + +Building on these advancements, vision-language models have recently been adapted for 3D point cloud representation, where 3D-image-text triplets [32, 63, 64, 69] enable contrastive pre-training. These models leverage powerful techniques such as momentum contrast [19] to align representations across modalities, allowing point cloud encoders to be pre-trained in a multimodal context. Additionally, tech- + +niques that apply vision-language models directly to 2D projections of point clouds [68, 70] have expanded cross-modal applications to 3D, including zero-shot shape classification, where models trained on 2D-image-text data demonstrate strong performance on 3D-related tasks. + +In the domain of multimodal synthesis, recent efforts in 2D and 3D model generation have leveraged diffusion models to tackle complex generation tasks. These models use robust priors, often trained on vast datasets, which allow them to create high-fidelity outputs in scenarios such as novel view synthesis and realistic 3D reconstructions from single RGB images [33, 65]. By combining synthetic data with diffusion-based priors, these frameworks achieve impressive zero-shot generalization and geometry-consistent 3D synthesis, opening new avenues for applications where generating realistic 3D content from limited information is critical. + +Representational similarity. The study of representation similarity across neural networks has seen a significant rise in interest, spurred largely by seminal works originating in neuroscience and computational cognitive science. These fields have long been invested in understanding the nature and alignment of cognitive representations, providing a foundational basis that has influenced the current trajectory in machine learning [42, 56]. Based on these insights, various works [1, 31, 40, 41, 43, 44, 59] provide evidence of an intrinsic connection between independently trained networks; the similarity is especially notable among large-scale models, with works such as [39, 40, 55] exploring the phenomenon. In the computer vision and pattern recognition areas, the line of work on similarity-based representations, pioneered by Duin and Pekalska [45], has also been seminal. This line of research, which examines data through the lens of similarity rather than feature attributes, has provided robust frameworks for classification and clustering [5], enabling models to generalize across patterns and variations in complex datasets. Although mostly empirical in nature, these observations find theoretical support in the study of harmonics in neural networks weights [38], Independent Component Analysis [22, 23, 26, 50] and Independent Mechanism Analysis [13, 14, 54]. These works suggest that, when capturing the same underlying data generative factors, deep learning models may converge towards similar structures despite their complexity and non-linearity. + +Latent space alignment. More recently, a range of approaches has been developed to align latent spaces within the same modality [3, 7, 12], as well as across different modalities [36, 58, 67], particularly between visual and textual domains. Techniques such as Procrustes analysis [60] and several similarity metrics [27, 53, 57, 61], including centered kernel alignment (CKA) [8, 28, 37], have proven instrumental in aligning representations. These methods offer + +various strategies for quantifying similarity between feature spaces, allowing us to examine cross-modal interactions at a deeper level. HHowever, these methods often focus on aligning entire latent spaces, potentially missing meaningful similarities confined to specific subspaces. + +Our approach builds upon CCA [20, 49], a pivotal tool in pattern recognition and multi-view learning applications [16, 51]. This technique identifies maximally correlated subspaces, enabling more refined alignment across modalities by isolating core, mutually relevant components. Leveraging CCA, alignment methods can extend into complex domains, such as connecting 3D and textual latent spaces, as we demonstrate in the following. + +# 3. Method + +We compare the similarity between latent (feature) spaces of various 3D and text encoders, introducing a new approach that builds on existing alignment methods to improve their effectiveness. Below, we outline the tools used to measure and align these latent spaces. + +# 3.1. Preliminaries + +Centered Kernel Alignment. CKA is a similarity measure frequently used in recent studies [8, 28] to compare representations in neural network feature spaces. Given feature matrices $\mathbf{X} \in \mathbb{R}^{n \times p}$ and $\mathbf{Y} \in \mathbb{R}^{n \times q}$ , we apply kernel functions $k$ and $l$ to obtain kernel representations $\mathbf{K} = k(\mathbf{X}, \mathbf{X}) \in \mathbb{R}^{n \times n}$ and $\mathbf{L} = l(\mathbf{Y}, \mathbf{Y}) \in \mathbb{R}^{n \times n}$ . CKA is defined as: + +$$ +\operatorname {C K A} (\mathbf {K}, \mathbf {L}) = \frac {\operatorname {H S I C} (\mathbf {K} , \mathbf {L})}{\sqrt {\operatorname {H S I C} (\mathbf {K} , \mathbf {K}) \operatorname {H S I C} (\mathbf {L} , \mathbf {L})}} \tag {1} +$$ + +where HSIC is the Hilbert-Schmidt Independence Criterion [15] and can be written as + +$$ +\operatorname {H S I C} (\mathbf {K}, \mathbf {L}) = \frac {1}{(n - 1) ^ {2}} \operatorname {t r} (\mathbf {K H L H}), \tag {2} +$$ + +where $\mathbf{H} = \mathbf{I} - \frac{1}{n}\mathbf{1}\mathbf{1}^{\top}$ + +Canonical Correlation Analysis. CCA [20, 49] is a statistical method that finds linear projections of two datasets, maximizing their correlation in a shared latent space. + +Formally, given two sets of zero-centered variables $\mathbf{X} \in \mathbb{R}^{n \times d_1}$ and $\mathbf{Y} \in \mathbb{R}^{n \times d_2}$ , CCA seeks two projection matrices $\mathbf{W}_{\mathbf{X}} \in \mathbb{R}^{d_1 \times k}$ and $\mathbf{W}_{\mathbf{Y}} \in \mathbb{R}^{d_2 \times k}$ that map $\mathbf{X}$ and $\mathbf{Y}$ into a common $k$ -dimensional space, maximizing the correlation between the projections $\mathbf{X}\mathbf{W}_{\mathbf{X}}$ and $\mathbf{Y}\mathbf{W}_{\mathbf{Y}}$ . The optimization can be expressed as: + +$$ +\max _ {\mathbf {W} _ {\mathbf {X}}, \mathbf {W} _ {\mathbf {Y}}} \operatorname {c o r r} \left(\mathbf {X} \mathbf {W} _ {\mathbf {X}}, \mathbf {Y} \mathbf {W} _ {\mathbf {Y}}\right), \tag {3} +$$ + +where $\operatorname{corr}(\cdot, \cdot)$ represents the correlation between the projected variables. + +# 3.2. Alignment approaches + +Recent developments in latent space alignment have introduced various methods. In this work, we examine both the affine transformation approach in [36] and the CKA-based matching approach from [37], observing their limited effectiveness in 3D-text latent space alignment, and propose a method to address these limitations. + +Latent Space Translation through Affine Transformation [36]. It is possible to estimate an affine transformation $T$ that maps one latent space $\mathcal{X}$ onto another latent space $\mathcal{Y}$ such that $T(\mathbf{x}) = \mathbf{R}\mathbf{x} + \mathbf{b}, \forall \mathbf{x} \in \mathcal{X}$ . To compute $T$ , we assume access to an anchor subset consisting of ground-truth paired samples from both latent spaces. These anchors enable us to determine $T$ by optimizing through gradient descent or, if $\mathbf{b} = \mathbf{0}$ by minimizing the least squares error. It assumes that $\mathcal{X}$ and $\mathcal{Y}$ share the same dimensionality and are normalized to zero mean and unit variance. These constraints can be enforced by zero-padding the smaller latent space. + +Local CKA-based retrieval and matching [37]. The CKA metric computed between two sets of features is sensitive to ordering and maximized when ground-truth pairs are aligned, this insight can be used to match unseen data by including them in a well-aligned anchors set. Formally, starting from aligned set of features $\mathbf{X}_A$ and $\mathbf{Y}_A$ , we compute for a query pair $(\mathbf{x}_q,\mathbf{y}_q)$ its local CKA defined as: + +$$ +\operatorname {l o c a l C K A} \left(\mathbf {x} _ {q}, \mathbf {y} _ {q}\right) = \operatorname {C K A} \left(\mathbf {K} _ {\left[ \mathbf {X} _ {A}, \mathbf {x} _ {q} \right]}, \mathbf {K} _ {\left[ \mathbf {Y} _ {A}, \mathbf {y} _ {q} \right]}\right), \tag {4} +$$ + +where $[\mathbf{X},\mathbf{x}]$ denotes the column-wise concatenation of matrix $\mathbf{X}$ and vector $\mathbf{x}$ . Local CKA calculates similarity in a pairwise manner, accounting for anchor alignment while being sensitive to ordering among query pairs. We introduce all possible query pairs, where correctly matched pairs exhibit the highest local CKA. + +# 3.3. Proposed method. + +Our goal is to reliably align the latent spaces of pre-trained 3D and text encoders. Given a dataset of $n$ caption-point cloud pairs, each embedded in the corresponding feature space, and represented as matrices $\mathbf{X} \in \mathbb{R}^{n \times p}$ and $\mathbf{Y} \in \mathbb{R}^{n \times q}$ respectively, we first select a subset of anchor pairs that will guide our translation process, denoted as $\mathbf{X}_A \in \mathbb{R}^{n_A \times p}$ and $\mathbf{Y}_A \in \mathbb{R}^{n_A \times q}$ . Building upon the mathematical foundations introduced in Section 3.2, we develop a pipeline that combines dimensionality reduction with the previously introduced alignment methods. + +Common Subspace Projection. In our work, we show that 3D and text latent spaces can effectively be aligned in low-dimensional connected subspaces. We begin by applying CCA to the anchor pairs to identify a shared $k$ -dimensional subspace (with $k < p, q$ ) that connects point cloud and text latent spaces. This yields projection matrices $\mathbf{W}_{\mathbf{X}^A} \in \mathbb{R}^{p \times k}$ + +and $\mathbf{W}_{\mathbf{Y}^A}\in \mathbb{R}^{q\times k}$ . All samples are then projected into this reduced space: + +$$ +\mathbf {X} ^ {r} = \mathbf {X} \mathbf {W} _ {\mathbf {X} _ {A}}, \quad \mathbf {Y} ^ {r} = \mathbf {Y} \mathbf {W} _ {\mathbf {Y} _ {A}}. \tag {5} +$$ + +In particular, we refer to the reduced anchors as $\mathbf{X}_A^r\in \mathbb{R}^{n_A\times k}$ and $\mathbf{Y}_A^r\in \mathbb{R}^{n_A\times k}$ . + +Our experiments show that projecting 3D and text latent spaces into a lower-dimensional, shared subspace improves alignment by isolating features that are highly correlated across modalities + +Alignment of projected latent spaces. Given the projected latent spaces, we can apply either of the previously described alignment methods in the reduced space. For the affine transformation approach, we learn a mapping between $\mathbf{X}^r$ and $\mathbf{Y}^r$ using the projected anchor pairs $\mathbf{X}_A^r$ , $\mathbf{Y}_A^r$ to optimize the transformation parameters $\mathbf{R}$ and $\mathbf{b}$ : + +$$ +T \left(\mathbf {X} ^ {r}\right) = \mathbf {R X} ^ {r} + \mathbf {b}, \tag {6} +$$ + +Alternatively, we can employ the local CKA-based matching approach in the projected space. For a query pair $(\mathbf{x}_q^r,\mathbf{y}_q^r)$ we compute its local CKA using the projected anchor sets: + +$$ +\operatorname {l o c a l C K A} \left(\mathbf {x} _ {q} ^ {r}, \mathbf {y} _ {q} ^ {r}\right) = \operatorname {C K A} \left(\mathbf {K} _ {\left[ \mathbf {X} _ {A} ^ {r}, \mathbf {x} _ {q} ^ {r} \right]}, \mathbf {K} _ {\left[ \mathbf {Y} _ {A} ^ {r}, \mathbf {y} _ {q} ^ {r} \right]}\right), \tag {7} +$$ + +# 4. Experimental setup + +# 4.1. Pre-training Dataset + +Prior works in point cloud pre-training, particularly within uni-modal frameworks, have relied heavily on ShapeNet [4], a dataset of 51,300 annotated 3D synthetic shapes spanning 55 categories. ShapeNet has been instrumental in advancing foundational methods, yet it remains limited by the relatively narrow scope of categories. With the release of Objaverse [9], which includes over 800,000 shapes across diverse real-world categories, a new standard for large-scale representation learning in the 3D domain has emerged. Objaverse's extensive shape diversity makes it ideal for both uni-modal and multi-modal learning. Despite these advantages, there is a lack of works that explore uni-modal pre-training specifically on Objaverse. + +# 4.2.Encoders + +We explore both multi-modal and uni-modal 3D and text encoders across varying levels of model complexity. + +Multi-modal 3D Encoders. For multi-modal pre-training, we use OpenShape, ULIP-2 and Uni3D [32] pre-trained models, trained on point cloud, image, and text triplets of Objverse with a contrastive pre-text task to align 3D encoders with frozen CLIP encoders. We adopt the Point-BERT-based + +
Method3D EncoderCLIPRoBERTaBERT
Matching accuracyTop-5 retrievalMatching accuracyTop-5 retrievalMatching accuracyTop-5 retrieval
Multi-modal 3D encoder
Affine + Subspace ProjectionOpenShape67.685.655.275.845.070.6
Affine + Subspace ProjectionULIP-265.685.247.270.635.259.8
Affine + Subspace ProjectionUni3D61.481.645.863.834.447.2
Uni-modal 3D encoder
AffinePointBert15.823.67.815.66.413.4
AffineSparseConv11.034.46.020.04.216.2
AffinePointnet18.421.88.010.29.612.2
Affine + Subspace Projection (Ours)PointBert30.842.223.228.415.618.2
Affine + Subspace Projection (Ours)SparseConv21.445.619.216.815.815.0
Affine + Subspace Projection (Ours)Pointnet25.236.622.420.816.616.0
Local CKAPointBert5.815.21.81.41.04.39
Local CKASparseConv3.413.61.791.60.63.8
Local CKAPointnet6.618.02.42.01.05.0
Local CKA + Subspace Projection (Ours)PointBert29.460.1917.042.415.037.2
Local CKA + Subspace Projection (Ours)SparseConv19.056.915.834.010.030.8
Local CKA + Subspace Projection (Ours)Pointnet26.853.018.640.214.338
+ +Table 1. Matching and retrieval performance across 3D and text encoders using different alignment approaches. We use 30,000 anchors for subspace projection and affine transformation approaches, and 1,000 anchors for local CKA. A query set of 500 is uniformly sampled, with results averaged over 3 different seeds. The subspace dimension is fixed at 50. Our approach (Ours) consistently demonstrates improved matching and retrieval performance, with multi-modal 3D encoders setting the upper bound for performance. Additional top- $k$ retrieval metrics are provided in the supplementary. + +variant of each model [66]. We primarily focus on OpenShape for simplicity but generalize results to ULIP-2 and Uni3D. + +Uni-modal 3D Encoders. For the uni-modal setup, we pre-train PointBERT on Objaverse using its original pretext tasks, which include masked point reconstruction and a uni-modal contrastive loss to encourage robust shape representation. We also explore two additional architectures: a Sparse convolution (MinkowskiNet [6]) model and a simpler architecture in PointNet [46], each pre-trained using a shape-level contrastive learning method [18]. This approach contrasts different partial views of input shapes. Across all 3D encoders in this setup, we fix the latent dimension at 512 to maintain consistency in representation space comparisons. + +Text encoders. We use the text encoder from OpenCLIP ViTbigG-14 [24], chosen to match text encoder used in OpenShape. Additionally, we examine alignment with purely unimodal text encoders by including BERT [11] and RoBERTa [34], we also evaluate the alignment with T5 [48] in the supplementary. + +Across all pre-training setups, parameters are kept consistent to facilitate direct comparisons. We detail the technical details in the supplementary. + +# 4.3. Downstream tasks + +We evaluate our alignment approach on Objaverse-LVIS [17], a human-verified test subset of Objaverse that contains 1,156 object categories. This subset is specifically reserved for evaluation and is unseen during pre-training. The captions are generated with Cap3D [35], which provides enhanced descriptive text for each 3D shape. Our evaluation framework consists of two main tasks: matching and retrieval. The + +matching task involves finding the correct permutation of captions for perfect matching given a shuffled set of query images and their corresponding captions; we utilize the linear sum assignment approach to perform this task [29]. For the retrieval task, the model must identify the correct 3D object from the query set based on a text caption. These tasks are particularly effective in measuring the cross-modal capabilities of encoders, and have been evaluated in prior Vision-Text studies [37]. While our main results emphasize the matching and top-5 retrieval tasks, we provide evaluation of top-1 and top-10 retrieval metrics in the supplementary. + +# 5. Results + +# 5.1. Are 3D and Text Latent Spaces similar? + +To evaluate the inherent similarity between 3D and text feature spaces without alignment, we compute linear CKA scores for both unimodal and multimodal encoders. The results are shown in Fig. 2. + +Comparison of 3D-Text and Vision-Text alignment. Prior work [21] reports CKA alignment values ranging from approximately $30\%$ to $48\%$ between different uni-modal vision and text encoders (see Figure 13 in [21]). In stark contrast, we find that the default alignment between 3D and text latent spaces is significantly weaker, with a maximum score of 0.12 observed for the uni-modal PointBERT and CLIP pair. This substantial gap underscores a key insight: unlike vision and text encoders, 3D encoders did not converge to structures similar to text. + +Alignment favors multi-modality. We observe that 3D multi-modal encoders demonstrate the highest CKA score with text encoders, particularly with CLIP's text encoder. + +![](images/e3b8a1a21c9eb56a2b848c4343eed59d22494912e782f00a154a42b39b3ae965.jpg) +Figure 2. Linear CKA scores between text and 3D encoders without alignment. Higher scores reflect stronger alignment between encoder pairs, with the strongest alignment observed between multi-modal 3D encoders (OpenShape) and CLIP text encoder due to their shared training on aligned representations. Uni-modal 3D encoders show significantly lower alignment with text encoders, although slightly higher with the CLIP text encoder. + +![](images/a430ff55c377ff41811a4a29ef7e986c858fac148530f041b4a6c7ea1b05d6f7.jpg) +Figure 3. Linear CKA scores between text and 3D encoders after affine translation. Affine translation results in a consistent increase in similarity for both 3D-to-text and text-to-3D directions. + +This behavior is expected, given that 3D multi-modal encoders are explicitly trained to align with CLIP latent spaces, which share the same textual modality. Even among unimodal 3D encoders, alignment with the CLIP text encoder is notably higher (e.g. 0.12 for PointBERT-CLIP) compared to alignment with uni-modal text encoders like RoBERTa (e.g., 0.04 for PointBERT-RoBERTa). This suggests that the visual understanding embedded in CLIP's text encoder extends beyond image and text domains to include 3D representations, + +despite the lack of explicit alignment during pre-training. + +# 5.2. Latent Space Alignment results + +We evaluate the performance of the alignment approaches outlined in Sec. 3 using matching and retrieval tasks, and analyze their effectiveness in aligning 3D and text encoders. Unless otherwise specified, we fix the subspace dimension to $d = 50$ and the number of anchors to 30,000. For downstream tasks, we uniformly sample a query set of size 500 and average results over 3 different seeds. For the affine transformation approach, we present results for the text-to-3D direction, noting that similar performance is achieved in the 3D-to-text direction as seen in Fig. 3. + +Existing approaches enable limited alignment We first assess whether previous successful approaches—affine transformation and local CKA—can achieve meaningful 3D-text alignment. As shown in Tab. 1, both methods yield modest alignment improvements over the unaligned baseline, where uni-modal 3D encoders start near zero in matching and retrieval tasks. These results suggest a small alignment shift. However, even with this improvement, alignment remains significantly lower than the alignment achieved in vision-text benchmarks [37]. This finding hints at the limits of uni-modal 3D encoders in achieving similar Vision-Text alignment performance, and might necessitate a different approach to align their latent space with text encoders. + +Importance of subspace projection (Ours). While aligning the latent spaces of 3D and text encoders achieved some success with existing methods, results remained well below vision-text alignment benchmarks. Motivated by this limitation, we propose aligning lower-dimensional subspaces, based on the hypothesis that 3D and text representations might intersect within a shared latent subspace. Using a validation set, we report in Fig. 4 the impact of subspace dimension, obtained through our CCA approach (see Sec. 3.3) combined with the affine transformation on the top-5 retrieval accuracy of uni-modal PointBert. While affine alignment alone yields better performance at higher dimensions, our subspace projection approach significantly outperforms it when the subspace dimension is reduced. This finding indicates that alignment quality is optimized within carefully chosen lower-dimensional spaces, reinforcing our hypothesis that 3D-text alignment is more successful within targeted subspaces. The results in Tab. 1 further validate this approach, showing a substantial increase in matching and retrieval performance across all 3D and text encoder pairs, outperforming both alignment approaches when they operate on the original latent space. + +Multi-modal encoders as an upper bound of performance. We include 3D multi-modal encoders for two main reasons: (1) As an upper bound for alignment performance, (2) To study how alignment degrades across different text + +encoders, including those that were not used during pretraining. In this context, the inclusion of 'Affine + Subspace Projection' multi-modal results in Tab. 1 shows how using our approach significantly narrows the gap between multimodal and uni-modal performance. However, to illustrate point 1. above, we also measured with cosine similarity the alignment of OpenShape (a multi-modal 3D encoder) and the same CLIP text encoder used during pre-training, achieving 0.94 for top-5 retrieval accuracy, an expected result since 3D and text encoders are explicitly aligned during pre-training. + +3D encoder complexity's low impact on alignment. Point-BERT outperforms the more complex SparseConv among uni-modal 3D encoders, suggesting that increasing model complexity alone does not guarantee improved alignment in 3D-text tasks. Surprisingly, even PointNet—a relatively simple model—achieves similar alignment scores, showing that factors other than model complexity may play a pivotal role in 3D-text alignment. This observation contrasts with vision-text alignment results [21, 37], where complex models typically leverage large datasets more effectively. Our findings thus indicate a distinctive aspect of 3D-text alignment: model simplicity does not impede, and may even aid, the interpretability and compatibility of learned features for cross-modal alignment. + +Different alignment techniques have different strengths. Our experiments reveal that different alignment techniques offer complementary strengths across tasks. For instance, affine transformation proves particularly effective for matching tasks across all 3D encoders, while local CKA shows superior performance in top-5 retrieval accuracy. This suggests that while some methods excel in precision tasks, others might better capture broader semantic nuances, making them more suitable for retrieval. Together, these observations imply that a hybrid approach could leverage the unique strengths of each method, opening up promising directions for future cross-modal applications. + +Scaling of our approach. In Fig. 5, we explore the scalability of our approach by analyzing how performance responds to increasing the number of anchors. Notably, our subspace projection method scales effectively with anchor count, reaching a plateau before requiring the full dataset (over 800,000 shapes). This scalability highlights the approach's efficiency in learning robust 3D-text mappings with a limited subset of anchor pairs. However, we observe a leveling off in performance gains beyond a certain anchor count, likely due to the constraints imposed by the low-dimensional subspace. This suggests that while our approach is computationally efficient and data-efficient, its reliance on a fixed lower-dimensional subspace could limit its adaptability to larger, more diverse datasets. + +![](images/8c832cb4e78c788f2cf6dcbec1a08c1cba97b07f87b1f7df234c4033bc7208dc.jpg) +Figure 4. Impact of subspace dimensionality on retrieval performance. Comparison of three approaches: our proposed CCA + affine translation method (blue), affine translation without subspace projection (red), and baseline feature space alignment without transformation (orange). Results are shown for the uni-modal PointBERT and CLIP text encoder, with generalizations to other encoders provided in the supplementary. + +![](images/a9065bb67ee67e43d30b785a71fa1390c78ab6621ac0bbc35628e80157872791.jpg) + +![](images/50266cd140248909d4f28fcc4073c59f6a56340e34d95dbd174a240c12897d62.jpg) +Figure 5. Effect of anchor set size on retrieval performance. Validation set results with different subspace dimensions show that retrieval performance improves as the anchor subset size increases but eventually reaches a plateau. Results are shown for the unimodal PointBERT and CLIP text encoder. + +![](images/41977d9bdfc5fe4789fbc00ff5b3f9d5c3e1fa433ae05099d05dacd5a80dfaae.jpg) +Figure 6. Pearson correlation between shape query chamfer distances and pairwise distances in the projected text latent subspace. We observe a higher Pearson correlation in the projected text latent subspace with optimal subspace dimension. Results are shown for the uni-modal PointBERT and CLIP text encoders. + +![](images/eab150a7a7dacf55cc3ce2cc95ad954123cf061de63ca241fb312bb39382336d.jpg) +Figure 7. Shape retrieval comparison between original and reduced latent spaces. For a given query shape, we retrieve the closest match based on cosine similarity. Results demonstrate a higher semantic understanding in the reduced 3D latent subspace compared to the original latent space. Shown for uni-modal Point-BERT and CLIP. + +# 5.3. Geometries vs. semantics. + +Our quantitative evaluation shows that low-dimensional subspace projection significantly improves latent space alignment in the 3D-Text setting. To better understand the characteristics of these subspaces spanned relative to the original spaces, we analyze the increase in semantic and geometric knowledge within the projected spaces. + +Increased geometric awareness of the text latent subspace. To quantify the increase in geometric awareness within the projected text subspace, we compute the Pearson correlation between the Chamfer distances of a query set of 500 shapes and the pairwise distances between feature vectors within the text subspace and compare these results to those obtained from the original text latent space. Results + +in Fig. 6 show that the optimal subspace dimension yields a high correlation with Chamfer distances, signaling an increase in the sensitivity of the text encoder to geometric properties when projected into this specific subspace. + +Increased semantic understanding of the 3D latent subspace. To assess the improved semantic capacity of the 3D subspace, we conducted a qualitative analysis of shape retrieval performance in both the reduced and the original latent spaces (Fig. 7). Given a set of query shapes, we observe that the reduced subspace retrieves shapes with stronger semantic similarity (e.g., retrieving animals for a query sheep shape), while the original latent space primarily favors geometric resemblance (e.g. retrieving stacked square shapes when querying a '3D model of a CPU'). This shift suggests that our subspace projection method enables the 3D encoder to capture semantic relationships that are absent in the full latent space. This effectively enables cross-modal applications and explains its higher matching and retrieval performance. + +# 6. Conclusion, Limitations and Future Work + +In this work, we present the first study investigating latent space alignment between 3D and text pre-trained encoders. Building on the hypothesis that these modalities share semantic connections within lower-dimensional subspaces, we propose an effective approach combining CCA projection with affine transformation estimation to translate between modalities' latent spaces. Our empirical results show that optimal cross-modal performance is achieved through low-dimensional subspace projection, and our method successfully improves alignment across diverse 3D and text uni-modal encoders. While CLIP-based multi-modal encoders establish performance upper bounds, we enable significant cross-modal capabilities in uni-modal encoders previously limited to single-modality tasks. We also demonstrate that semantic understanding can be extracted from geometry-aware latent spaces of uni-modal 3D encoders. + +Although our work focused on Objaverse, which is the first large-scale 3D dataset, it would be interesting to consider how scaling on Objaverse-XL [10] would affect the alignment quality between 3D and text encoders. Moreover, in this work we do not distinguish object-level vs. scene-level annotations, and decomposing objects or scenes into their composing blocks could shed light onto the compositionality of the learned representations. Finally, the subspace alignment method that we introduce in this work can be broadly applicable to other representations as well. In the future, we plan to use it to investigate the limitations of alignment observed in other representations. In particular, even when trained at significantly higher data scales, images and text representations do not align perfectly [21], and it would be interesting to reveal the unique and complementary nature of different modalities via subspace analysis. + +Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, as well as gifts from Ansys and Adobe Research. This work was also supported by the Galileo 2022 fellowship from the Università Italo Francese/ Université Franco Italiène (UIF/UFI) within the project G22_4 titled “Multimodal Artificial Intelligence for 3D shape analysis, modeling and applications”. + +# References + +[1] Lisa Bonheme and Marek Grzes. How do variational autoencoders learn? insights from representational similarity. arXiv preprint arXiv:2205.08399, 2022. 3 +[2] Irene Cannistraci, Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, and Emanuele Rodola. Bootstrapping parallel anchors for relative representations. In The First Tiny Papers Track at ICLR 2023, Tiny Papers at ICLR 2023, 2023. 1 +[3] Irene Cannistraci, Luca Moschella, Marco Fumero, Valentino Maiorca, and Emanuele Rodolà. From bricks to bridges: Product of invariances to enhance latent space communication. In The Twelfth International Conference on Learning Representations, 2024. 3 +[4] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4 +[5] Yihua Chen, Eric K. Garcia, Maya R. Gupta, Ali Rahimi, and Luca Cazzanti. Similarity-based classification: Concepts and algorithms. Journal of Machine Learning Research, 10(27): 747-776, 2009. 3 +[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 5 +[7] Donato Crisostomi, Irene Cannistraci, Luca Moschella, Pietro Barbiero, Marco Ciccone, Pietro Lio, and Emanuele Rodola. From charts to atlas: Merging latent spaces into one. In NeurIPS 2023 Workshop on Symmetry and Geometry in Neural Representations, 2023. 3 +[8] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 3 +[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. 2023 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2022. 1, 4 +[10] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli + +VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 1, 8 +[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[12] Marco Fumero, Marco Pegoraro, Valentino Maiorca, Francesco Locatello, and Emanuele Rodolà. Latent functional maps: a spectral framework for representation alignment. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3 +[13] Shubhangi Ghosh, Luigi Gresele, Julius von Kugelgen, Michel Besserve, and Bernhard Scholkopf. Independent mechanism analysis and the manifold hypothesis, 2023. 3 +[14] Luigi Gresele, Julius Von Kugelgen, Vincent Stimper, Bernhard Scholkopf, and Michel Besserve. Independent mechanism analysis, a new concept? In Advances in Neural Information Processing Systems, 2021. 3 +[15] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3 +[16] Chenfeng Guo and Dongrui Wu. Canonical correlation analysis (cca) based multi-view learning: An overview. arXiv preprint arXiv:1907.01693, 2019. 3 +[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5 +[18] Souhail Hadgi, Lei Li, and Maks Ovsjanikov. To supervise or not to supervise: Understanding and addressing the key challenges of 3d transfer learning. arXiv preprint arXiv:2403.17869, 2024.5 +[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arxiv e-prints, art. arXiv preprint arXiv:1911.05722, 2019. 2 +[20] Harold Hotelling. Relations between two sets of variates. In Breakthroughs in statistics: methodology and distribution, pages 162-190. Springer, 1992. 3 +[21] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987, 2024. 1, 5, 7, 8 +[22] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA. In Advances in Neural Information Processing Systems. Curran Associates, Inc. 3 +[23] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, pages 859–868. PMLR. 3 +[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5 + +[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[26] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR. 3 +[27] Max Klabunde, Tobias Schumacher, Markus Strohmaier, and Florian Lemmerich. Similarity of neural network models: A survey of functional and representational measures, 2024. 3 +[28] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 3 +[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5 +[30] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2 +[31] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John Hopcroft. Convergent learning: Do different neural networks learn the same representations? arXiv preprint arXiv:1511.07543, 2015. 3 +[32] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in neural information processing systems, 36, 2024. 1, 2, 4 +[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. 2023. 3 +[34] Yinhan Liu. *Roberta: A robustly optimized bert pretraining approach.* arXiv preprint arXiv:1907.11692, 364, 2019. 5 +[35] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 5 +[36] Valentino Maiorca, Luca Moschella, Antonio Norelli, Marco Fumero, Francesco Locatello, and Emanuele Rodola. Latent space translation via semantic alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4 +[37] Mayug Maniparambili, Raiymbek Akshulakov, Yasser Abdelaziz Dahou Djilali, Mohamed El Amine Seddik, Sanath Narayan, Karttikeya Mangalam, and Noel E O'Connor. Do vision and language encoders represent the world similarly? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14334-14343, 2024. 1, 2, 3, 4, 5, 6, 7 +[38] Giovanni Luca Marchetti and Christopher Hillar. Harmonics of learning: Universal fourier features emerge + +in invariant networks. https://synthical.com/article/03032df5-f9c5-43a6-8253-d4af1d67e0d4, 2023.3 +[39] Raghav Mehta, Vítor Albiero, Li Chen, Ivan Evtimov, Tamar Glaser, Zhiheng Li, and Tal Hassner. You only need a good embeddings extractor to fix spurious correlations, 2022. 3 +[40] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3 +[41] Ari S. Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, page 5732-5741, Red Hook, NY, USA, 2018. Curran Associates Inc. 3 +[42] Luca Moschella. Latent communication in artificial neural networks, 2024. 3 +[43] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In The Eleventh International Conference on Learning Representations, 2023. 1, 3 +[44] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. Advances in Neural Information Processing Systems, 36:15303-15319, 2023. 3 +[45] E. Pekalska and Robert Duin. Automatic pattern recognition by similarity representations. *Electronics Letters*, 37:159 - 160, 2001. 3 +[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 5 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5 +[49] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3 +[50] Geoffrey Roeder, Luke Metz, and Durk Kingma. On Linear Identifiability of Learned Representations. In Proceedings of the 38th International Conference on Machine Learning, pages 9030-9039. PMLR. 3 +[51] Jan Rupnik and John Shawe-Taylor. Multi-view canonical correlation analysis. In Conference on data mining and data warehouses (SiKDD 2010), pages 1-4, 2010. 3 + +[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[53] Mahdiyar Shahbazi, Ali Shirali, Hamid Aghajan, and Hamed Nili. Using distance on the riemannian manifold to compare representations in brain and in models. NeuroImage, 239: 118271, 2021. 3 +[54] Joanna Sliwa, Shubhangi Ghosh, Vincent Stimper, Luigi Gresele, and Bernhard Scholkopf. Probing the robustness of independent mechanism analysis for representation learning. In UAI 2022 Workshop on Causal Representation Learning, 2022. 3 +[55] Gowthami Somepalli, Liam Fowl, Arpit Bansal, Ping Yeh-Chiang, Yehuda Dar, Richard Baraniuk, Micah Goldblum, and Tom Goldstein. Can neural nets learn the same model twice? investigating reproducibility and double descent from the decision boundary perspective. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13689-13698, 2022. 3 +[56] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. Love, Erin Grant, Iris Groen, Jascha Achterberg, Joshua B. Tenenbaum, Katherine M. Collins, Katherine L. Hermann, Kerem Oktar, Klaus Greff, Martin N. Hebart, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas P. O'Connell, Thomas Unterthiner, Andrew K. Lampinen, Klaus-Robert Müller, Mariya Toneva, and Thomas L. Griffiths. Getting aligned on representational alignment, 2023. 3 +[57] Shuai Tang, Wesley J. Maddox, Charlie Dickens, Tom Diethe, and Andreas Damianou. Similarity of neural networks with gradients, 2020. 3 +[58] Thomas Theodoridis, Theocharis Chatzis, Vassilios Solachidis, Kosmas Dimitropoulos, and Petros Daras. Cross-modal variational alignment of latent spaces. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4127-4136, 2020. 3 +[59] Ivan Vulic, Sebastian Ruder, and Anders Søgaard. Are all good word vector spaces isomorphic? arXiv preprint arXiv:2004.04070, 2020.3 +[60] Chang Wang and Sridhar Mahadevan. Manifold alignment using procrastes analysis. In Proceedings of the 25th international conference on Machine learning, pages 1120-1127, 2008. 3 +[61] Alex H Williams, Erin Kunz, Simon Kornblith, and Scott Linderman. Generalized shape metrics on neural representations. In Advances in Neural Information Processing Systems, 2021. 3 +[62] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1 + +[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 2 +[64] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27091–27101, 2024. 1, 2 +[65] Yuxuan Xue, Xianghui Xie, Riccardo Marin, and Gerard. Pons-Moll. Human 3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models. 2024. 3 +[66] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, and Jie Zhou. Pre-training 3d point cloud transformers with masked point modeling. 2022 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19291-19300, 2021. 5 +[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18102-18112, 2022. 3 +[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3 +[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. arXiv preprint arXiv:2310.06773, 2023. 1, 2 +[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 3(4), 2022. 3 \ No newline at end of file diff --git a/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/images.zip b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..55cbbf993f5e426759a6d0e2d1a14b52a5938969 --- /dev/null +++ b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:128db1d4293d12805d26623b9fc172ee2dbd12c0c2dfeae26407cf96df03edf5 +size 365399 diff --git a/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/layout.json b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..72581526d37f96fb135a1ea697fb85ff150a1715 --- /dev/null +++ b/2025/Escaping Plato's Cave_ Towards the Alignment of 3D and Text Latent Spaces/layout.json @@ -0,0 +1,8328 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "Escaping Plato's Cave: Towards the Alignment of 3D and Text Latent Spaces" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 142, + 235, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 142, + 235, + 157 + ], + "spans": [ + { + "bbox": [ + 158, + 142, + 235, + 157 + ], + "type": "text", + "content": "Souhail Hadgi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 143, + 350, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 143, + 350, + 157 + ], + "spans": [ + { + "bbox": [ + 260, + 143, + 350, + 157 + ], + "type": "text", + "content": "Luca Moschella" + }, + { + "bbox": [ + 260, + 143, + 350, + 157 + ], + "type": "inline_equation", + "content": "^{2,*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 371, + 144, + 450, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 144, + 450, + 156 + ], + "spans": [ + { + "bbox": [ + 371, + 144, + 450, + 156 + ], + "type": "text", + "content": "Andrea Santilli2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 158, + 233, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 158, + 233, + 171 + ], + "spans": [ + { + "bbox": [ + 159, + 158, + 233, + 171 + ], + "type": "text", + "content": "Diego Gomez" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 259, + 158, + 333, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 158, + 333, + 171 + ], + "spans": [ + { + "bbox": [ + 259, + 158, + 333, + 171 + ], + "type": "text", + "content": "Qixing Huang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 158, + 450, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 158, + 450, + 171 + ], + "spans": [ + { + "bbox": [ + 358, + 158, + 450, + 171 + ], + "type": "text", + "content": "Emanuele Rodolà2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 210, + 172, + 283, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 172, + 283, + 183 + ], + "spans": [ + { + "bbox": [ + 210, + 172, + 283, + 183 + ], + "type": "text", + "content": "Simone Melzi3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 172, + 400, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 172, + 400, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 172, + 400, + 186 + ], + "type": "text", + "content": "Maks Ovsjanikov1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 186, + 247, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 247, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 247, + 199 + ], + "type": "text", + "content": "École polytechnique" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 186, + 471, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 471, + 199 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 471, + 199 + ], + "type": "text", + "content": "Sapienza University of Rome" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 122, + 201, + 279, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 201, + 279, + 213 + ], + "spans": [ + { + "bbox": [ + 122, + 201, + 279, + 213 + ], + "type": "text", + "content": "3University of Milano-Bicocca" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 201, + 488, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 488, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 488, + 213 + ], + "type": "text", + "content": "4The University of Texas at Austin" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 266, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 567 + ], + "type": "text", + "content": "Recent works have shown that, when trained at scale, uni-modal 2D vision and text encoders converge to learned features that share remarkable structural properties, despite arising from different representations. However, the role of 3D encoders with respect to other modalities remains unexplored. Furthermore, existing 3D foundation models that leverage large datasets are typically trained with explicit alignment objectives with respect to frozen encoders from other representations. In this work, we investigate the possibility of a posteriori alignment of representations obtained from uni-modal 3D encoders compared to text-based feature spaces. We show that naive post-training feature alignment of uni-modal text and 3D encoders results in limited performance. We then focus on extracting subspaces of the corresponding feature spaces and discover that by projecting learned representations onto well-chosen lower-dimensional subspaces the quality of alignment becomes significantly higher, leading to improved accuracy on matching and retrieval tasks. Our analysis further sheds light on the nature of these shared subspaces, which roughly separate between semantic and geometric data representations. Overall, ours is the first work that helps to establish a baseline for posttraining alignment of 3D uni-modal and text feature spaces, and helps to highlight both the shared and unique properties of 3D data compared to other representations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 609, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 694 + ], + "type": "text", + "content": "Recent advances in multi-modal learning, particularly in vision-language models such as CLIP [47], have sparked interest in extending these successes to the 3D domain. Most current approaches primarily focus on training 3D encoders through triplet-based learning with pre-trained 2D vision and language encoders [32, 64, 69], leveraging new large-scale datasets such as Objaverse [9, 10], and showing promising" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 242, + 523, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 523, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 523, + 254 + ], + "type": "text", + "content": "results in tasks such as zero-shot shape recognition." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 255, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 459 + ], + "type": "text", + "content": "While CLIP [47] was trained with explicit alignment objectives between text and image representations, recent work has observed that even when trained independently, vision and text encoders tend to exhibit significant similarities [21, 37]. In particular, the learned latent spaces of pure (uni-modal) text and vision encoders have similar proximity structure [21] and can be aligned relatively easily after training with a small number of known anchor correspondences [2, 37, 43]. Furthermore, the degree of similarity in learned features across modalities strongly correlates with the quality of performance in various downstream tasks. One prominent interpretation of these results is that given sufficient scale of training data and model complexity, different representations are converging to a shared underlying structure of the physical world. This has given rise to the Platonic Representation Hypothesis [21], where different representations are viewed as projections of reality on particular modalities." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": "At the same time, as the physical world is inherently (at least) 3-dimensional, a natural question is how the structure of uni-modal vision or text encoders relates to the features learned directly from 3D data. This question raises several challenges: first, large-scale 3D datasets have only recently become available [9, 10]; second, most existing 3D \"foundation models\" are trained with explicit alignment objectives with respect to frozen 2D and text encoders, which limits the utility of post-training comparison. Finally, there is a lack of universally agreed-upon architectures and training objectives for 3D data, and many commonly-used architectures tend to have limited generalization power [62]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 665 + ], + "type": "text", + "content": "In this work, we initiate the first study on the relation between 3D and language representations. We formulate the task of post-training alignment between a range of 3D and text encoders and study the accuracy and utility of several alignment strategies." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Our first insight is that when trained on pure 3D data with self-supervised objectives, 3D encoders tend to lead to representations that align only very weakly to text-based representations. We believe that this observation already" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 135, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 135, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 135, + 713 + ], + "type": "text", + "content": "*Currently at Apple." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19825" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 553, + 277 + ], + "type": "image", + "image_path": "46ffe3e9717f4e2077b73030cfc9ed91b0269e1ca365e2d139c2b0f60b639825.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": "Figure 1. A visual overview of the proposed approach. is illustrated. From left to right: We begin with two distinct input collections—one consisting of 3D shapes and the other of textual prompts. In the blue box, independent, frozen uni-modal encoders map each modality into separate, high-dimensional latent spaces, shown in the red box. A dimensionality reduction procedure is applied to project these learned spaces into low-dimensional subspaces, represented in the green box. Finally, an alignment method registers the two low-dimensional subspaces, enabling cross-modal applications such as shape retrieval, with examples depicted in the yellow box." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 364, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 295, + 389 + ], + "type": "text", + "content": "Highlights the difficulty of uni-modal training and sheds light on the scarcity of \"pure\" 3D foundation models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 391, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 391, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 391, + 297, + 582 + ], + "type": "text", + "content": "Additionally, we reveal that while 3D and text latent spaces are not naturally aligned, effective cross-modal translation can be achieved through subspace projection and alignment. Our key insight is that by identifying and operating in correlated subspaces, we can improve the latent space alignment of 3D and text encoders without the need for expensive joint training. To achieve this, we introduce a simple but effective approach that combines Canonical Correlation Analysis (CCA) and existing alignment approaches such as affine translation [36] and local CKA [37]. This approach and our subsequent analysis extend previous works aimed at comparing learned feature spaces, but shows that careful subspace alignment can reveal subtle but important similarities, which are otherwise obscured in global comparisons. We provide a visual overview of the proposed approach in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 586, + 284, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 586, + 284, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 586, + 284, + 597 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 601, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 601, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 650 + ], + "type": "text", + "content": "1. We extend the analysis of vision-text uni-modal latent space alignment to 3D uni-modal encoders and text, highlighting the limited similarity between these latent spaces and the low efficacy of current alignment approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 297, + 714 + ], + "type": "text", + "content": "2. We propose an efficient approach for cross-modal alignment between text and 3D features that combines CCA for subspace selection and existing alignment methods. This approach improves alignment performance with minimal computational overhead, as demonstrated through" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 365, + 556, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 365, + 556, + 412 + ], + "spans": [ + { + "bbox": [ + 326, + 365, + 556, + 412 + ], + "type": "text", + "content": "experiments in matching and retrieval tasks. Our method establishes a baseline for 3D-text cross-modal understanding, providing an alternative to explicit multimodal pretraining for cross-modal tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 416, + 556, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 416, + 556, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 416, + 556, + 464 + ], + "type": "text", + "content": "3. We observe a complementary structure between the spanned subspaces and the original feature spaces, enabling a distinction between geometric and semantic representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 473, + 400, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 400, + 486 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 400, + 486 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "text", + "content": "Multi-modal representation learning. Multimodal representation learning has surged in recent years, driven by the success of image-language models [25, 30, 47, 52] that enable seamless cross-modal applications. These models serve as the backbone for tasks spanning from text-based image retrieval to generating high-quality visuals from natural language prompts. By aligning visual and linguistic features in a shared latent space, these models have paved the way for advanced interactions between modalities, setting the stage for applications where visual and textual information are jointly processed and understood." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "type": "text", + "content": "Building on these advancements, vision-language models have recently been adapted for 3D point cloud representation, where 3D-image-text triplets [32, 63, 64, 69] enable contrastive pre-training. These models leverage powerful techniques such as momentum contrast [19] to align representations across modalities, allowing point cloud encoders to be pre-trained in a multimodal context. Additionally, tech-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19826" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 296, + 133 + ], + "type": "text", + "content": "niques that apply vision-language models directly to 2D projections of point clouds [68, 70] have expanded cross-modal applications to 3D, including zero-shot shape classification, where models trained on 2D-image-text data demonstrate strong performance on 3D-related tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 135, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 296, + 278 + ], + "type": "text", + "content": "In the domain of multimodal synthesis, recent efforts in 2D and 3D model generation have leveraged diffusion models to tackle complex generation tasks. These models use robust priors, often trained on vast datasets, which allow them to create high-fidelity outputs in scenarios such as novel view synthesis and realistic 3D reconstructions from single RGB images [33, 65]. By combining synthetic data with diffusion-based priors, these frameworks achieve impressive zero-shot generalization and geometry-consistent 3D synthesis, opening new avenues for applications where generating realistic 3D content from limited information is critical." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 287, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 610 + ], + "type": "text", + "content": "Representational similarity. The study of representation similarity across neural networks has seen a significant rise in interest, spurred largely by seminal works originating in neuroscience and computational cognitive science. These fields have long been invested in understanding the nature and alignment of cognitive representations, providing a foundational basis that has influenced the current trajectory in machine learning [42, 56]. Based on these insights, various works [1, 31, 40, 41, 43, 44, 59] provide evidence of an intrinsic connection between independently trained networks; the similarity is especially notable among large-scale models, with works such as [39, 40, 55] exploring the phenomenon. In the computer vision and pattern recognition areas, the line of work on similarity-based representations, pioneered by Duin and Pekalska [45], has also been seminal. This line of research, which examines data through the lens of similarity rather than feature attributes, has provided robust frameworks for classification and clustering [5], enabling models to generalize across patterns and variations in complex datasets. Although mostly empirical in nature, these observations find theoretical support in the study of harmonics in neural networks weights [38], Independent Component Analysis [22, 23, 26, 50] and Independent Mechanism Analysis [13, 14, 54]. These works suggest that, when capturing the same underlying data generative factors, deep learning models may converge towards similar structures despite their complexity and non-linearity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "Latent space alignment. More recently, a range of approaches has been developed to align latent spaces within the same modality [3, 7, 12], as well as across different modalities [36, 58, 67], particularly between visual and textual domains. Techniques such as Procrustes analysis [60] and several similarity metrics [27, 53, 57, 61], including centered kernel alignment (CKA) [8, 28, 37], have proven instrumental in aligning representations. These methods offer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "various strategies for quantifying similarity between feature spaces, allowing us to examine cross-modal interactions at a deeper level. HHowever, these methods often focus on aligning entire latent spaces, potentially missing meaningful similarities confined to specific subspaces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 133, + 556, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 556, + 228 + ], + "type": "text", + "content": "Our approach builds upon CCA [20, 49], a pivotal tool in pattern recognition and multi-view learning applications [16, 51]. This technique identifies maximally correlated subspaces, enabling more refined alignment across modalities by isolating core, mutually relevant components. Leveraging CCA, alignment methods can extend into complex domains, such as connecting 3D and textual latent spaces, as we demonstrate in the following." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 238, + 370, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 370, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 370, + 251 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 555, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 318 + ], + "type": "text", + "content": "We compare the similarity between latent (feature) spaces of various 3D and text encoders, introducing a new approach that builds on existing alignment methods to improve their effectiveness. Below, we outline the tools used to measure and align these latent spaces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 324, + 400, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 400, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 400, + 335 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "content": "Centered Kernel Alignment. CKA is a similarity measure frequently used in recent studies [8, 28] to compare representations in neural network feature spaces. Given feature matrices " + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times p}" + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{n \\times q}" + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "content": ", we apply kernel functions " + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "content": " to obtain kernel representations " + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = k(\\mathbf{X}, \\mathbf{X}) \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = l(\\mathbf{Y}, \\mathbf{Y}) \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 313, + 346, + 555, + 430 + ], + "type": "text", + "content": ". CKA is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 434, + 555, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 434, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 343, + 434, + 555, + 464 + ], + "type": "interline_equation", + "content": "\\operatorname {C K A} (\\mathbf {K}, \\mathbf {L}) = \\frac {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {L})}{\\sqrt {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {K}) \\operatorname {H S I C} (\\mathbf {L} , \\mathbf {L})}} \\tag {1}", + "image_path": "2ffa772ccb3bae549f6cb08154a63bdbcf643c6d7cbb6ad756e99f6b82d102f1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "type": "text", + "content": "where HSIC is the Hilbert-Schmidt Independence Criterion [15] and can be written as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 355, + 498, + 555, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 498, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 355, + 498, + 555, + 525 + ], + "type": "interline_equation", + "content": "\\operatorname {H S I C} (\\mathbf {K}, \\mathbf {L}) = \\frac {1}{(n - 1) ^ {2}} \\operatorname {t r} (\\mathbf {K H L H}), \\tag {2}", + "image_path": "5583fb6decbf587487e306ce842bfd01335e2cd5e9ebe004cff65dbe1a4d0c74.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 530, + 411, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 411, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 411, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 530, + 411, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\mathbf{I} - \\frac{1}{n}\\mathbf{1}\\mathbf{1}^{\\top}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 548, + 556, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 556, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 556, + 583 + ], + "type": "text", + "content": "Canonical Correlation Analysis. CCA [20, 49] is a statistical method that finds linear projections of two datasets, maximizing their correlation in a shared latent space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": "Formally, given two sets of zero-centered variables " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times d_1}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{n \\times d_2}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": ", CCA seeks two projection matrices " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{X}} \\in \\mathbb{R}^{d_1 \\times k}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Y}} \\in \\mathbb{R}^{d_2 \\times k}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": " that map " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": " into a common " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": "-dimensional space, maximizing the correlation between the projections " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{W}_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}\\mathbf{W}_{\\mathbf{Y}}" + }, + { + "bbox": [ + 313, + 584, + 556, + 656 + ], + "type": "text", + "content": ". The optimization can be expressed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 663, + 555, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 663, + 555, + 683 + ], + "spans": [ + { + "bbox": [ + 370, + 663, + 555, + 683 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbf {W} _ {\\mathbf {X}}, \\mathbf {W} _ {\\mathbf {Y}}} \\operatorname {c o r r} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {X}}, \\mathbf {Y} \\mathbf {W} _ {\\mathbf {Y}}\\right), \\tag {3}", + "image_path": "9a3014109f6b3589cc84221124887ace344279d0e09e45c9eaebb09ba29fb4b4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{corr}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " represents the correlation between the projected variables." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19827" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 185, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 185, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 185, + 85 + ], + "type": "text", + "content": "3.2. Alignment approaches" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 161 + ], + "type": "text", + "content": "Recent developments in latent space alignment have introduced various methods. In this work, we examine both the affine transformation approach in [36] and the CKA-based matching approach from [37], observing their limited effectiveness in 3D-text latent space alignment, and propose a method to address these limitations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": "Latent Space Translation through Affine Transformation [36]. It is possible to estimate an affine transformation " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": " that maps one latent space " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": " onto another latent space " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "T(\\mathbf{x}) = \\mathbf{R}\\mathbf{x} + \\mathbf{b}, \\forall \\mathbf{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": ". To compute " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": ", we assume access to an anchor subset consisting of ground-truth paired samples from both latent spaces. These anchors enable us to determine " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": " by optimizing through gradient descent or, if " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = \\mathbf{0}" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": " by minimizing the least squares error. It assumes that " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 166, + 297, + 299 + ], + "type": "text", + "content": " share the same dimensionality and are normalized to zero mean and unit variance. These constraints can be enforced by zero-padding the smaller latent space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "text", + "content": "Local CKA-based retrieval and matching [37]. The CKA metric computed between two sets of features is sensitive to ordering and maximized when ground-truth pairs are aligned, this insight can be used to match unseen data by including them in a well-aligned anchors set. Formally, starting from aligned set of features " + }, + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A" + }, + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_A" + }, + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "text", + "content": ", we compute for a query pair " + }, + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_q,\\mathbf{y}_q)" + }, + { + "bbox": [ + 55, + 302, + 296, + 387 + ], + "type": "text", + "content": " its local CKA defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 396, + 296, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 396, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 67, + 396, + 296, + 412 + ], + "type": "interline_equation", + "content": "\\operatorname {l o c a l C K A} \\left(\\mathbf {x} _ {q}, \\mathbf {y} _ {q}\\right) = \\operatorname {C K A} \\left(\\mathbf {K} _ {\\left[ \\mathbf {X} _ {A}, \\mathbf {x} _ {q} \\right]}, \\mathbf {K} _ {\\left[ \\mathbf {Y} _ {A}, \\mathbf {y} _ {q} \\right]}\\right), \\tag {4}", + "image_path": "a669b6fbc1d583d4eba04659d551bfe5f2746413e0959a96a86a57e5eb42bd03.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "inline_equation", + "content": "[\\mathbf{X},\\mathbf{x}]" + }, + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "text", + "content": " denotes the column-wise concatenation of matrix " + }, + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "text", + "content": " and vector " + }, + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 418, + 297, + 492 + ], + "type": "text", + "content": ". Local CKA calculates similarity in a pairwise manner, accounting for anchor alignment while being sensitive to ordering among query pairs. We introduce all possible query pairs, where correctly matched pairs exhibit the highest local CKA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 499, + 163, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 163, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 163, + 512 + ], + "type": "text", + "content": "3.3. Proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "text", + "content": "Our goal is to reliably align the latent spaces of pre-trained 3D and text encoders. Given a dataset of " + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "text", + "content": " caption-point cloud pairs, each embedded in the corresponding feature space, and represented as matrices " + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times p}" + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{n \\times q}" + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "text", + "content": " respectively, we first select a subset of anchor pairs that will guide our translation process, denoted as " + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A \\in \\mathbb{R}^{n_A \\times p}" + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_A \\in \\mathbb{R}^{n_A \\times q}" + }, + { + "bbox": [ + 55, + 517, + 296, + 637 + ], + "type": "text", + "content": ". Building upon the mathematical foundations introduced in Section 3.2, we develop a pipeline that combines dimensionality reduction with the previously introduced alignment methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "Common Subspace Projection. In our work, we show that 3D and text latent spaces can effectively be aligned in low-dimensional connected subspaces. We begin by applying CCA to the anchor pairs to identify a shared " + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "-dimensional subspace (with " + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "inline_equation", + "content": "k < p, q" + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": ") that connects point cloud and text latent spaces. This yields projection matrices " + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{X}^A} \\in \\mathbb{R}^{p \\times k}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Y}^A}\\in \\mathbb{R}^{q\\times k}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ". All samples are then projected into this reduced space:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 362, + 106, + 555, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 106, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 362, + 106, + 555, + 121 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {r} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {X} _ {A}}, \\quad \\mathbf {Y} ^ {r} = \\mathbf {Y} \\mathbf {W} _ {\\mathbf {Y} _ {A}}. \\tag {5}", + "image_path": "237e4c5c85091e09fb936cc46643827fa748431313ce21adde789b35dcf71b5d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 130, + 555, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 555, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 555, + 154 + ], + "type": "text", + "content": "In particular, we refer to the reduced anchors as " + }, + { + "bbox": [ + 313, + 130, + 555, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A^r\\in \\mathbb{R}^{n_A\\times k}" + }, + { + "bbox": [ + 313, + 130, + 555, + 154 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 130, + 555, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_A^r\\in \\mathbb{R}^{n_A\\times k}" + }, + { + "bbox": [ + 313, + 130, + 555, + 154 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 154, + 555, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 201 + ], + "type": "text", + "content": "Our experiments show that projecting 3D and text latent spaces into a lower-dimensional, shared subspace improves alignment by isolating features that are highly correlated across modalities" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "content": "Alignment of projected latent spaces. Given the projected latent spaces, we can apply either of the previously described alignment methods in the reduced space. For the affine transformation approach, we learn a mapping between " + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^r" + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^r" + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "content": " using the projected anchor pairs " + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A^r" + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_A^r" + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "content": " to optimize the transformation parameters " + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 206, + 556, + 278 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 390, + 289, + 555, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 289, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 390, + 289, + 555, + 302 + ], + "type": "interline_equation", + "content": "T \\left(\\mathbf {X} ^ {r}\\right) = \\mathbf {R X} ^ {r} + \\mathbf {b}, \\tag {6}", + "image_path": "b67f53b398017a69a2975f83f01bcd54c0b746ba794b8386a5d43af2fb9b264f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 312, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 348 + ], + "type": "text", + "content": "Alternatively, we can employ the local CKA-based matching approach in the projected space. For a query pair " + }, + { + "bbox": [ + 313, + 312, + 556, + 348 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_q^r,\\mathbf{y}_q^r)" + }, + { + "bbox": [ + 313, + 312, + 556, + 348 + ], + "type": "text", + "content": " we compute its local CKA using the projected anchor sets:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 327, + 370, + 555, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 370, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 327, + 370, + 555, + 386 + ], + "type": "interline_equation", + "content": "\\operatorname {l o c a l C K A} \\left(\\mathbf {x} _ {q} ^ {r}, \\mathbf {y} _ {q} ^ {r}\\right) = \\operatorname {C K A} \\left(\\mathbf {K} _ {\\left[ \\mathbf {X} _ {A} ^ {r}, \\mathbf {x} _ {q} ^ {r} \\right]}, \\mathbf {K} _ {\\left[ \\mathbf {Y} _ {A} ^ {r}, \\mathbf {y} _ {q} ^ {r} \\right]}\\right), \\tag {7}", + "image_path": "f52f57a81b2e734300a3378ad86387e22d46c3dbad1ed882eab4333ad01d60de.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 405, + 430, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 430, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 430, + 419 + ], + "type": "text", + "content": "4. Experimental setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 425, + 433, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 433, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 433, + 438 + ], + "type": "text", + "content": "4.1. Pre-training Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 442, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 556, + 598 + ], + "type": "text", + "content": "Prior works in point cloud pre-training, particularly within uni-modal frameworks, have relied heavily on ShapeNet [4], a dataset of 51,300 annotated 3D synthetic shapes spanning 55 categories. ShapeNet has been instrumental in advancing foundational methods, yet it remains limited by the relatively narrow scope of categories. With the release of Objaverse [9], which includes over 800,000 shapes across diverse real-world categories, a new standard for large-scale representation learning in the 3D domain has emerged. Objaverse's extensive shape diversity makes it ideal for both uni-modal and multi-modal learning. Despite these advantages, there is a lack of works that explore uni-modal pre-training specifically on Objaverse." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 607, + 382, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 382, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 382, + 619 + ], + "type": "text", + "content": "4.2.Encoders" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 624, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 650 + ], + "type": "text", + "content": "We explore both multi-modal and uni-modal 3D and text encoders across varying levels of model complexity." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Multi-modal 3D Encoders. For multi-modal pre-training, we use OpenShape, ULIP-2 and Uni3D [32] pre-trained models, trained on point cloud, image, and text triplets of Objverse with a contrastive pre-text task to align 3D encoders with frozen CLIP encoders. We adopt the Point-BERT-based" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19828" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 232 + ], + "type": "table", + "html": "
Method3D EncoderCLIPRoBERTaBERT
Matching accuracyTop-5 retrievalMatching accuracyTop-5 retrievalMatching accuracyTop-5 retrieval
Multi-modal 3D encoder
Affine + Subspace ProjectionOpenShape67.685.655.275.845.070.6
Affine + Subspace ProjectionULIP-265.685.247.270.635.259.8
Affine + Subspace ProjectionUni3D61.481.645.863.834.447.2
Uni-modal 3D encoder
AffinePointBert15.823.67.815.66.413.4
AffineSparseConv11.034.46.020.04.216.2
AffinePointnet18.421.88.010.29.612.2
Affine + Subspace Projection (Ours)PointBert30.842.223.228.415.618.2
Affine + Subspace Projection (Ours)SparseConv21.445.619.216.815.815.0
Affine + Subspace Projection (Ours)Pointnet25.236.622.420.816.616.0
Local CKAPointBert5.815.21.81.41.04.39
Local CKASparseConv3.413.61.791.60.63.8
Local CKAPointnet6.618.02.42.01.05.0
Local CKA + Subspace Projection (Ours)PointBert29.460.1917.042.415.037.2
Local CKA + Subspace Projection (Ours)SparseConv19.056.915.834.010.030.8
Local CKA + Subspace Projection (Ours)Pointnet26.853.018.640.214.338
", + "image_path": "822f3f9f748c447b426df4d0f1561685bb536bdf672782a5669c80af17c66cd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 297 + ], + "type": "text", + "content": "Table 1. Matching and retrieval performance across 3D and text encoders using different alignment approaches. We use 30,000 anchors for subspace projection and affine transformation approaches, and 1,000 anchors for local CKA. A query set of 500 is uniformly sampled, with results averaged over 3 different seeds. The subspace dimension is fixed at 50. Our approach (Ours) consistently demonstrates improved matching and retrieval performance, with multi-modal 3D encoders setting the upper bound for performance. Additional top-" + }, + { + "bbox": [ + 55, + 239, + 555, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 239, + 555, + 297 + ], + "type": "text", + "content": " retrieval metrics are provided in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 312, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 348 + ], + "type": "text", + "content": "variant of each model [66]. We primarily focus on OpenShape for simplicity but generalize results to ULIP-2 and Uni3D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": "Uni-modal 3D Encoders. For the uni-modal setup, we pre-train PointBERT on Objaverse using its original pretext tasks, which include masked point reconstruction and a uni-modal contrastive loss to encourage robust shape representation. We also explore two additional architectures: a Sparse convolution (MinkowskiNet [6]) model and a simpler architecture in PointNet [46], each pre-trained using a shape-level contrastive learning method [18]. This approach contrasts different partial views of input shapes. Across all 3D encoders in this setup, we fix the latent dimension at 512 to maintain consistency in representation space comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 563 + ], + "type": "text", + "content": "Text encoders. We use the text encoder from OpenCLIP ViTbigG-14 [24], chosen to match text encoder used in OpenShape. Additionally, we examine alignment with purely unimodal text encoders by including BERT [11] and RoBERTa [34], we also evaluate the alignment with T5 [48] in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 601 + ], + "type": "text", + "content": "Across all pre-training setups, parameters are kept consistent to facilitate direct comparisons. We detail the technical details in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 164, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 164, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 164, + 622 + ], + "type": "text", + "content": "4.3. Downstream tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "We evaluate our alignment approach on Objaverse-LVIS [17], a human-verified test subset of Objaverse that contains 1,156 object categories. This subset is specifically reserved for evaluation and is unseen during pre-training. The captions are generated with Cap3D [35], which provides enhanced descriptive text for each 3D shape. Our evaluation framework consists of two main tasks: matching and retrieval. The" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 445 + ], + "type": "text", + "content": "matching task involves finding the correct permutation of captions for perfect matching given a shuffled set of query images and their corresponding captions; we utilize the linear sum assignment approach to perform this task [29]. For the retrieval task, the model must identify the correct 3D object from the query set based on a text caption. These tasks are particularly effective in measuring the cross-modal capabilities of encoders, and have been evaluated in prior Vision-Text studies [37]. While our main results emphasize the matching and top-5 retrieval tasks, we provide evaluation of top-1 and top-10 retrieval metrics in the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 459, + 367, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 459, + 367, + 470 + ], + "spans": [ + { + "bbox": [ + 314, + 459, + 367, + 470 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 525, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 525, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 525, + 492 + ], + "type": "text", + "content": "5.1. Are 3D and Text Latent Spaces similar?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 547 + ], + "type": "text", + "content": "To evaluate the inherent similarity between 3D and text feature spaces without alignment, we compute linear CKA scores for both unimodal and multimodal encoders. The results are shown in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 556, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 671 + ], + "type": "text", + "content": "Comparison of 3D-Text and Vision-Text alignment. Prior work [21] reports CKA alignment values ranging from approximately " + }, + { + "bbox": [ + 313, + 552, + 556, + 671 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 671 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 552, + 556, + 671 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 671 + ], + "type": "text", + "content": " between different uni-modal vision and text encoders (see Figure 13 in [21]). In stark contrast, we find that the default alignment between 3D and text latent spaces is significantly weaker, with a maximum score of 0.12 observed for the uni-modal PointBERT and CLIP pair. This substantial gap underscores a key insight: unlike vision and text encoders, 3D encoders did not converge to structures similar to text." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Alignment favors multi-modality. We observe that 3D multi-modal encoders demonstrate the highest CKA score with text encoders, particularly with CLIP's text encoder." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19829" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 294, + 253 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 294, + 253 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 294, + 253 + ], + "type": "image", + "image_path": "e3b8a1a21c9eb56a2b848c4343eed59d22494912e782f00a154a42b39b3ae965.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 263, + 298, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 298, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 298, + 343 + ], + "type": "text", + "content": "Figure 2. Linear CKA scores between text and 3D encoders without alignment. Higher scores reflect stronger alignment between encoder pairs, with the strongest alignment observed between multi-modal 3D encoders (OpenShape) and CLIP text encoder due to their shared training on aligned representations. Uni-modal 3D encoders show significantly lower alignment with text encoders, although slightly higher with the CLIP text encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 357, + 294, + 538 + ], + "blocks": [ + { + "bbox": [ + 59, + 357, + 294, + 538 + ], + "lines": [ + { + "bbox": [ + 59, + 357, + 294, + 538 + ], + "spans": [ + { + "bbox": [ + 59, + 357, + 294, + 538 + ], + "type": "image", + "image_path": "a430ff55c377ff41811a4a29ef7e986c858fac148530f041b4a6c7ea1b05d6f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 548, + 297, + 583 + ], + "lines": [ + { + "bbox": [ + 55, + 548, + 297, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 297, + 583 + ], + "type": "text", + "content": "Figure 3. Linear CKA scores between text and 3D encoders after affine translation. Affine translation results in a consistent increase in similarity for both 3D-to-text and text-to-3D directions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 298, + 715 + ], + "type": "text", + "content": "This behavior is expected, given that 3D multi-modal encoders are explicitly trained to align with CLIP latent spaces, which share the same textual modality. Even among unimodal 3D encoders, alignment with the CLIP text encoder is notably higher (e.g. 0.12 for PointBERT-CLIP) compared to alignment with uni-modal text encoders like RoBERTa (e.g., 0.04 for PointBERT-RoBERTa). This suggests that the visual understanding embedded in CLIP's text encoder extends beyond image and text domains to include 3D representations," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 545, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 85 + ], + "type": "text", + "content": "despite the lack of explicit alignment during pre-training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 94, + 484, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 94, + 484, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 94, + 484, + 106 + ], + "type": "text", + "content": "5.2. Latent Space Alignment results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 112, + 556, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 112, + 556, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 112, + 556, + 233 + ], + "type": "text", + "content": "We evaluate the performance of the alignment approaches outlined in Sec. 3 using matching and retrieval tasks, and analyze their effectiveness in aligning 3D and text encoders. Unless otherwise specified, we fix the subspace dimension to " + }, + { + "bbox": [ + 313, + 112, + 556, + 233 + ], + "type": "inline_equation", + "content": "d = 50" + }, + { + "bbox": [ + 313, + 112, + 556, + 233 + ], + "type": "text", + "content": " and the number of anchors to 30,000. For downstream tasks, we uniformly sample a query set of size 500 and average results over 3 different seeds. For the affine transformation approach, we present results for the text-to-3D direction, noting that similar performance is achieved in the 3D-to-text direction as seen in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 236, + 557, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 557, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 557, + 392 + ], + "type": "text", + "content": "Existing approaches enable limited alignment We first assess whether previous successful approaches—affine transformation and local CKA—can achieve meaningful 3D-text alignment. As shown in Tab. 1, both methods yield modest alignment improvements over the unaligned baseline, where uni-modal 3D encoders start near zero in matching and retrieval tasks. These results suggest a small alignment shift. However, even with this improvement, alignment remains significantly lower than the alignment achieved in vision-text benchmarks [37]. This finding hints at the limits of uni-modal 3D encoders in achieving similar Vision-Text alignment performance, and might necessitate a different approach to align their latent space with text encoders." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 397, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 661 + ], + "type": "text", + "content": "Importance of subspace projection (Ours). While aligning the latent spaces of 3D and text encoders achieved some success with existing methods, results remained well below vision-text alignment benchmarks. Motivated by this limitation, we propose aligning lower-dimensional subspaces, based on the hypothesis that 3D and text representations might intersect within a shared latent subspace. Using a validation set, we report in Fig. 4 the impact of subspace dimension, obtained through our CCA approach (see Sec. 3.3) combined with the affine transformation on the top-5 retrieval accuracy of uni-modal PointBert. While affine alignment alone yields better performance at higher dimensions, our subspace projection approach significantly outperforms it when the subspace dimension is reduced. This finding indicates that alignment quality is optimized within carefully chosen lower-dimensional spaces, reinforcing our hypothesis that 3D-text alignment is more successful within targeted subspaces. The results in Tab. 1 further validate this approach, showing a substantial increase in matching and retrieval performance across all 3D and text encoder pairs, outperforming both alignment approaches when they operate on the original latent space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 558, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 558, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 558, + 715 + ], + "type": "text", + "content": "Multi-modal encoders as an upper bound of performance. We include 3D multi-modal encoders for two main reasons: (1) As an upper bound for alignment performance, (2) To study how alignment degrades across different text" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19830" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "encoders, including those that were not used during pretraining. In this context, the inclusion of 'Affine + Subspace Projection' multi-modal results in Tab. 1 shows how using our approach significantly narrows the gap between multimodal and uni-modal performance. However, to illustrate point 1. above, we also measured with cosine similarity the alignment of OpenShape (a multi-modal 3D encoder) and the same CLIP text encoder used during pre-training, achieving 0.94 for top-5 retrieval accuracy, an expected result since 3D and text encoders are explicitly aligned during pre-training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 374 + ], + "type": "text", + "content": "3D encoder complexity's low impact on alignment. Point-BERT outperforms the more complex SparseConv among uni-modal 3D encoders, suggesting that increasing model complexity alone does not guarantee improved alignment in 3D-text tasks. Surprisingly, even PointNet—a relatively simple model—achieves similar alignment scores, showing that factors other than model complexity may play a pivotal role in 3D-text alignment. This observation contrasts with vision-text alignment results [21, 37], where complex models typically leverage large datasets more effectively. Our findings thus indicate a distinctive aspect of 3D-text alignment: model simplicity does not impede, and may even aid, the interpretability and compatibility of learned features for cross-modal alignment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 388, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 533 + ], + "type": "text", + "content": "Different alignment techniques have different strengths. Our experiments reveal that different alignment techniques offer complementary strengths across tasks. For instance, affine transformation proves particularly effective for matching tasks across all 3D encoders, while local CKA shows superior performance in top-5 retrieval accuracy. This suggests that while some methods excel in precision tasks, others might better capture broader semantic nuances, making them more suitable for retrieval. Together, these observations imply that a hybrid approach could leverage the unique strengths of each method, opening up promising directions for future cross-modal applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Scaling of our approach. In Fig. 5, we explore the scalability of our approach by analyzing how performance responds to increasing the number of anchors. Notably, our subspace projection method scales effectively with anchor count, reaching a plateau before requiring the full dataset (over 800,000 shapes). This scalability highlights the approach's efficiency in learning robust 3D-text mappings with a limited subset of anchor pairs. However, we observe a leveling off in performance gains beyond a certain anchor count, likely due to the constraints imposed by the low-dimensional subspace. This suggests that while our approach is computationally efficient and data-efficient, its reliance on a fixed lower-dimensional subspace could limit its adaptability to larger, more diverse datasets." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 433, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 433, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 433, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 433, + 148 + ], + "type": "image", + "image_path": "8c832cb4e78c788f2cf6dcbec1a08c1cba97b07f87b1f7df234c4033bc7208dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 158, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 236 + ], + "type": "text", + "content": "Figure 4. Impact of subspace dimensionality on retrieval performance. Comparison of three approaches: our proposed CCA + affine translation method (blue), affine translation without subspace projection (red), and baseline feature space alignment without transformation (orange). Results are shown for the uni-modal PointBERT and CLIP text encoder, with generalizations to other encoders provided in the supplementary." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 553, + 148 + ], + "type": "image", + "image_path": "a9065bb67ee67e43d30b785a71fa1390c78ab6621ac0bbc35628e80157872791.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 256, + 550, + 379 + ], + "blocks": [ + { + "bbox": [ + 318, + 256, + 550, + 379 + ], + "lines": [ + { + "bbox": [ + 318, + 256, + 550, + 379 + ], + "spans": [ + { + "bbox": [ + 318, + 256, + 550, + 379 + ], + "type": "image", + "image_path": "50266cd140248909d4f28fcc4073c59f6a56340e34d95dbd174a240c12897d62.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 556, + 447 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 447 + ], + "type": "text", + "content": "Figure 5. Effect of anchor set size on retrieval performance. Validation set results with different subspace dimensions show that retrieval performance improves as the anchor subset size increases but eventually reaches a plateau. Results are shown for the unimodal PointBERT and CLIP text encoder." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 468, + 550, + 615 + ], + "blocks": [ + { + "bbox": [ + 319, + 468, + 550, + 615 + ], + "lines": [ + { + "bbox": [ + 319, + 468, + 550, + 615 + ], + "spans": [ + { + "bbox": [ + 319, + 468, + 550, + 615 + ], + "type": "image", + "image_path": "41977d9bdfc5fe4789fbc00ff5b3f9d5c3e1fa433ae05099d05dacd5a80dfaae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 628, + 555, + 684 + ], + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 684 + ], + "type": "text", + "content": "Figure 6. Pearson correlation between shape query chamfer distances and pairwise distances in the projected text latent subspace. We observe a higher Pearson correlation in the projected text latent subspace with optimal subspace dimension. Results are shown for the uni-modal PointBERT and CLIP text encoders." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19831" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 293, + 427 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 293, + 427 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 293, + 427 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 293, + 427 + ], + "type": "image", + "image_path": "eab150a7a7dacf55cc3ce2cc95ad954123cf061de63ca241fb312bb39382336d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 445, + 297, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 297, + 514 + ], + "type": "text", + "content": "Figure 7. Shape retrieval comparison between original and reduced latent spaces. For a given query shape, we retrieve the closest match based on cosine similarity. Results demonstrate a higher semantic understanding in the reduced 3D latent subspace compared to the original latent space. Shown for uni-modal Point-BERT and CLIP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 534, + 198, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 198, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 198, + 547 + ], + "type": "text", + "content": "5.3. Geometries vs. semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 552, + 297, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 297, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 297, + 625 + ], + "type": "text", + "content": "Our quantitative evaluation shows that low-dimensional subspace projection significantly improves latent space alignment in the 3D-Text setting. To better understand the characteristics of these subspaces spanned relative to the original spaces, we analyze the increase in semantic and geometric knowledge within the projected spaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": "Increased geometric awareness of the text latent subspace. To quantify the increase in geometric awareness within the projected text subspace, we compute the Pearson correlation between the Chamfer distances of a query set of 500 shapes and the pairwise distances between feature vectors within the text subspace and compare these results to those obtained from the original text latent space. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "type": "text", + "content": "in Fig. 6 show that the optimal subspace dimension yields a high correlation with Chamfer distances, signaling an increase in the sensitivity of the text encoder to geometric properties when projected into this specific subspace." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 126, + 556, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 556, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 556, + 294 + ], + "type": "text", + "content": "Increased semantic understanding of the 3D latent subspace. To assess the improved semantic capacity of the 3D subspace, we conducted a qualitative analysis of shape retrieval performance in both the reduced and the original latent spaces (Fig. 7). Given a set of query shapes, we observe that the reduced subspace retrieves shapes with stronger semantic similarity (e.g., retrieving animals for a query sheep shape), while the original latent space primarily favors geometric resemblance (e.g. retrieving stacked square shapes when querying a '3D model of a CPU'). This shift suggests that our subspace projection method enables the 3D encoder to capture semantic relationships that are absent in the full latent space. This effectively enables cross-modal applications and explains its higher matching and retrieval performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 308, + 544, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 544, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 544, + 322 + ], + "type": "text", + "content": "6. Conclusion, Limitations and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 329, + 556, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 520 + ], + "type": "text", + "content": "In this work, we present the first study investigating latent space alignment between 3D and text pre-trained encoders. Building on the hypothesis that these modalities share semantic connections within lower-dimensional subspaces, we propose an effective approach combining CCA projection with affine transformation estimation to translate between modalities' latent spaces. Our empirical results show that optimal cross-modal performance is achieved through low-dimensional subspace projection, and our method successfully improves alignment across diverse 3D and text uni-modal encoders. While CLIP-based multi-modal encoders establish performance upper bounds, we enable significant cross-modal capabilities in uni-modal encoders previously limited to single-modality tasks. We also demonstrate that semantic understanding can be extracted from geometry-aware latent spaces of uni-modal 3D encoders." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "Although our work focused on Objaverse, which is the first large-scale 3D dataset, it would be interesting to consider how scaling on Objaverse-XL [10] would affect the alignment quality between 3D and text encoders. Moreover, in this work we do not distinguish object-level vs. scene-level annotations, and decomposing objects or scenes into their composing blocks could shed light onto the compositionality of the learned representations. Finally, the subspace alignment method that we introduce in this work can be broadly applicable to other representations as well. In the future, we plan to use it to investigate the limitations of alignment observed in other representations. In particular, even when trained at significantly higher data scales, images and text representations do not align perfectly [21], and it would be interesting to reveal the unique and complementary nature of different modalities via subspace analysis." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19832" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, as well as gifts from Ansys and Adobe Research. This work was also supported by the Galileo 2022 fellowship from the Università Italo Francese/ Université Franco Italiène (UIF/UFI) within the project G22_4 titled “Multimodal Artificial Intelligence for 3D shape analysis, modeling and applications”." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 190, + 115, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 204 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 210, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "text", + "content": "[1] Lisa Bonheme and Marek Grzes. How do variational autoencoders learn? insights from representational similarity. arXiv preprint arXiv:2205.08399, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 245, + 296, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 296, + 298 + ], + "type": "text", + "content": "[2] Irene Cannistraci, Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, and Emanuele Rodola. Bootstrapping parallel anchors for relative representations. In The First Tiny Papers Track at ICLR 2023, Tiny Papers at ICLR 2023, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 300, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 296, + 354 + ], + "type": "text", + "content": "[3] Irene Cannistraci, Luca Moschella, Marco Fumero, Valentino Maiorca, and Emanuele Rodolà. From bricks to bridges: Product of invariances to enhance latent space communication. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 357, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 409 + ], + "type": "text", + "content": "[4] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 412, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 412, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 412, + 296, + 456 + ], + "type": "text", + "content": "[5] Yihua Chen, Eric K. Garcia, Maya R. Gupta, Ali Rahimi, and Luca Cazzanti. Similarity-based classification: Concepts and algorithms. Journal of Machine Learning Research, 10(27): 747-776, 2009. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 457, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 511 + ], + "type": "text", + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "text", + "content": "[7] Donato Crisostomi, Irene Cannistraci, Luca Moschella, Pietro Barbiero, Marco Ciccone, Pietro Lio, and Emanuele Rodola. From charts to atlas: Merging latent spaces into one. In NeurIPS 2023 Workshop on Symmetry and Geometry in Neural Representations, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "type": "text", + "content": "[8] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 678 + ], + "type": "text", + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. 2023 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2022. 1, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[10] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663, 2023. 1, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 554, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 149 + ], + "type": "text", + "content": "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 151, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 205 + ], + "type": "text", + "content": "[12] Marco Fumero, Marco Pegoraro, Valentino Maiorca, Francesco Locatello, and Emanuele Rodolà. Latent functional maps: a spectral framework for representation alignment. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 239 + ], + "type": "text", + "content": "[13] Shubhangi Ghosh, Luigi Gresele, Julius von Kugelgen, Michel Besserve, and Bernhard Scholkopf. Independent mechanism analysis and the manifold hypothesis, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 555, + 282 + ], + "type": "text", + "content": "[14] Luigi Gresele, Julius Von Kugelgen, Vincent Stimper, Bernhard Scholkopf, and Michel Besserve. Independent mechanism analysis, a new concept? In Advances in Neural Information Processing Systems, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "text", + "content": "[15] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 328, + 555, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 555, + 360 + ], + "type": "text", + "content": "[16] Chenfeng Guo and Dongrui Wu. Canonical correlation analysis (cca) based multi-view learning: An overview. arXiv preprint arXiv:1907.01693, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 554, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 404 + ], + "type": "text", + "content": "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 405, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 555, + 448 + ], + "type": "text", + "content": "[18] Souhail Hadgi, Lei Li, and Maks Ovsjanikov. To supervise or not to supervise: Understanding and addressing the key challenges of 3d transfer learning. arXiv preprint arXiv:2403.17869, 2024.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 491 + ], + "type": "text", + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arxiv e-prints, art. arXiv preprint arXiv:1911.05722, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 555, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 555, + 526 + ], + "type": "text", + "content": "[20] Harold Hotelling. Relations between two sets of variates. In Breakthroughs in statistics: methodology and distribution, pages 162-190. Springer, 1992. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 558 + ], + "type": "text", + "content": "[21] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987, 2024. 1, 5, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "text", + "content": "[22] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA. In Advances in Neural Information Processing Systems. Curran Associates, Inc. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 658 + ], + "type": "text", + "content": "[23] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, pages 859–868. PMLR. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 659, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 555, + 713 + ], + "type": "text", + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19833" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 137 + ], + "type": "text", + "content": "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "text", + "content": "[26] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "type": "text", + "content": "[27] Max Klabunde, Tobias Schumacher, Markus Strohmaier, and Florian Lemmerich. Similarity of neural network models: A survey of functional and representational measures, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "text", + "content": "[28] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 307 + ], + "type": "text", + "content": "[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 374 + ], + "type": "text", + "content": "[30] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 420 + ], + "type": "text", + "content": "[31] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John Hopcroft. Convergent learning: Do different neural networks learn the same representations? arXiv preprint arXiv:1511.07543, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "type": "text", + "content": "[32] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in neural information processing systems, 36, 2024. 1, 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "type": "text", + "content": "[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 533 + ], + "type": "text", + "content": "[34] Yinhan Liu. *Roberta: A robustly optimized bert pretraining approach.* arXiv preprint arXiv:1907.11692, 364, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 567 + ], + "type": "text", + "content": "[35] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 569, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 295, + 612 + ], + "type": "text", + "content": "[36] Valentino Maiorca, Luca Moschella, Antonio Norelli, Marco Fumero, Francesco Locatello, and Emanuele Rodola. Latent space translation via semantic alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 689 + ], + "type": "text", + "content": "[37] Mayug Maniparambili, Raiymbek Akshulakov, Yasser Abdelaziz Dahou Djilali, Mohamed El Amine Seddik, Sanath Narayan, Karttikeya Mangalam, and Noel E O'Connor. Do vision and language encoders represent the world similarly? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14334-14343, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[38] Giovanni Luca Marchetti and Christopher Hillar. Harmonics of learning: Universal fourier features emerge" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 556, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 556, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 556, + 106 + ], + "type": "text", + "content": "in invariant networks. https://synthical.com/article/03032df5-f9c5-43a6-8253-d4af1d67e0d4, 2023.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "text", + "content": "[39] Raghav Mehta, Vítor Albiero, Li Chen, Ivan Evtimov, Tamar Glaser, Zhiheng Li, and Tal Hassner. You only need a good embeddings extractor to fix spurious correlations, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 172 + ], + "type": "text", + "content": "[40] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 237 + ], + "type": "text", + "content": "[41] Ari S. Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, page 5732-5741, Red Hook, NY, USA, 2018. Curran Associates Inc. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 259 + ], + "type": "text", + "content": "[42] Luca Moschella. Latent communication in artificial neural networks, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 555, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 315 + ], + "type": "text", + "content": "[43] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In The Eleventh International Conference on Learning Representations, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 317, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 555, + 370 + ], + "type": "text", + "content": "[44] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. Advances in Neural Information Processing Systems, 36:15303-15319, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "text", + "content": "[45] E. Pekalska and Robert Duin. Automatic pattern recognition by similarity representations. *Electronics Letters*, 37:159 - 160, 2001. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 405, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 555, + 458 + ], + "type": "text", + "content": "[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 525 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 581 + ], + "type": "text", + "content": "[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 634 + ], + "type": "text", + "content": "[49] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 680 + ], + "type": "text", + "content": "[50] Geoffrey Roeder, Luke Metz, and Durk Kingma. On Linear Identifiability of Learned Representations. In Proceedings of the 38th International Conference on Machine Learning, pages 9030-9039. PMLR. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[51] Jan Rupnik and John Shawe-Taylor. Multi-view canonical correlation analysis. In Conference on data mining and data warehouses (SiKDD 2010), pages 1-4, 2010. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19834" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "type": "text", + "content": "[53] Mahdiyar Shahbazi, Ali Shirali, Hamid Aghajan, and Hamed Nili. Using distance on the riemannian manifold to compare representations in brain and in models. NeuroImage, 239: 118271, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 237 + ], + "type": "text", + "content": "[54] Joanna Sliwa, Shubhangi Ghosh, Vincent Stimper, Luigi Gresele, and Bernhard Scholkopf. Probing the robustness of independent mechanism analysis for representation learning. In UAI 2022 Workshop on Causal Representation Learning, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 316 + ], + "type": "text", + "content": "[55] Gowthami Somepalli, Liam Fowl, Arpit Bansal, Ping Yeh-Chiang, Yehuda Dar, Richard Baraniuk, Micah Goldblum, and Tom Goldstein. Can neural nets learn the same model twice? investigating reproducibility and double descent from the decision boundary perspective. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13689-13698, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 316, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 426 + ], + "type": "text", + "content": "[56] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. Love, Erin Grant, Iris Groen, Jascha Achterberg, Joshua B. Tenenbaum, Katherine M. Collins, Katherine L. Hermann, Kerem Oktar, Klaus Greff, Martin N. Hebart, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas P. O'Connell, Thomas Unterthiner, Andrew K. Lampinen, Klaus-Robert Müller, Mariya Toneva, and Thomas L. Griffiths. Getting aligned on representational alignment, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "text", + "content": "[57] Shuai Tang, Wesley J. Maddox, Charlie Dickens, Tom Diethe, and Andreas Damianou. Similarity of neural networks with gradients, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 460, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 524 + ], + "type": "text", + "content": "[58] Thomas Theodoridis, Theocharis Chatzis, Vassilios Solachidis, Kosmas Dimitropoulos, and Petros Daras. Cross-modal variational alignment of latent spaces. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4127-4136, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 525, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 558 + ], + "type": "text", + "content": "[59] Ivan Vulic, Sebastian Ruder, and Anders Søgaard. Are all good word vector spaces isomorphic? arXiv preprint arXiv:2004.04070, 2020.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 559, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 601 + ], + "type": "text", + "content": "[60] Chang Wang and Sridhar Mahadevan. Manifold alignment using procrastes analysis. In Proceedings of the 25th international conference on Machine learning, pages 1120-1127, 2008. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "type": "text", + "content": "[61] Alex H Williams, Erin Kunz, Simon Kornblith, and Scott Linderman. Generalized shape metrics on neural representations. In Advances in Neural Information Processing Systems, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 713 + ], + "type": "text", + "content": "[62] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 496 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 205 + ], + "type": "text", + "content": "[64] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27091–27101, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 207, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 239 + ], + "type": "text", + "content": "[65] Yuxuan Xue, Xianghui Xie, Riccardo Marin, and Gerard. Pons-Moll. Human 3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models. 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 240, + 555, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 555, + 294 + ], + "type": "text", + "content": "[66] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, and Jie Zhou. Pre-training 3d point cloud transformers with masked point modeling. 2022 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19291-19300, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 297, + 555, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 351 + ], + "type": "text", + "content": "[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18102-18112, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 353, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 555, + 407 + ], + "type": "text", + "content": "[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 408, + 555, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 555, + 451 + ], + "type": "text", + "content": "[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. arXiv preprint arXiv:2310.06773, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 453, + 554, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 554, + 496 + ], + "type": "text", + "content": "[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 3(4), 2022. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19835" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_content_list.json b/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1a5b5c3d4fa5a0b14975b880139ca5e12f71225b --- /dev/null +++ b/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_content_list.json @@ -0,0 +1,1956 @@ +[ + { + "type": "text", + "text": "Estimating Body and Hand Motion in an Ego-sensed World", + "text_level": 1, + "bbox": [ + 212, + 130, + 785, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brent Yi $^{1}$ Vickie Ye $^{1}$ Maya Zheng $^{1}$ Yunqi Li $^{2}$ Lea Müller $^{1}$ Georgios Pavlakos $^{3}$ Yi Ma $^{1}$ Jitendra Malik $^{1}$ Angjoo Kanazawa $^{1}$ UC Berkeley ${}^{2}$ ShanghaiTech ${}^{3}$ UT Austin", + "bbox": [ + 230, + 170, + 767, + 243 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/736b3a79b52c98a5c596117a87cd1888334df4458f4e6ca710a1c6af83829da1.jpg", + "image_caption": [ + "Figure 1. EgoAllo. We present a system that estimates human body pose, height, and hand parameters from egocentric SLAM poses and images. Outputs capture the wearer's actions in the allocentric reference frame of the scene, which we visualize here with 3D reconstructions." + ], + "image_footnote": [], + "bbox": [ + 93, + 265, + 617, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ecba775f6884bd43051b09d83eacbc8f6c2c586bcf2eb6223483ba1e7f9e770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 266, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 250, + 561, + 325, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present EgoAllo, a system for human motion estimation from a head-mounted device. Using only egocentric SLAM poses and images, EgoAllo guides sampling from a conditional diffusion model to estimate 3D body pose, height, and hand parameters that capture a device wearer's actions in the allocentric coordinate frame of the scene. To achieve this, our key insight is in representation: we propose spatial and temporal invariance criteria for improving model performance, from which we derive a head motion conditioning parameterization that improves estimation by up to $18\\%$ . We also show how the bodies estimated by our system can improve hand estimation: the resulting kinematic and temporal constraints can reduce world-frame errors in single-frame estimates by $40\\%$ .", + "bbox": [ + 88, + 597, + 485, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 828, + 215, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Head-mounted devices are becoming increasingly mainstream. In addition to offering new challenges for 3D scene understanding [18, 57, 68, 111], egocentric sensors from these devices are", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unique in that their outputs are coupled to a human wearer's motion in the world. Using these sensors to understand the wearer in addition to the scene around them is essential for applications in augmented reality, robotics, and assistive technologies.", + "bbox": [ + 511, + 564, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We therefore introduce EgoAllo, a system that uses egocentric inputs to estimate the wearer and their motion in the world, or allocentric, coordinate frame. We take as input sensed metric SLAM head poses and egocentric video from devices like Project Aria [61]. We then output estimates of human body pose, height, and hand motion parameters.", + "bbox": [ + 511, + 626, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This is a difficult task: while body parts like hands occasionally appear in egocentric frames, most body parameters are never directly observed. Estimated body pose, body height, and hand parameters must be consistent with both the sensed egomotion and visual observations. This setting differs from most prior works in egocentric human motion estimation [7, 48], which focus on body pose and do not address the challenges of height and hand motion.", + "bbox": [ + 511, + 718, + 908, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our proposed system uses a head pose-conditioned diffusion model as a motion prior, as well as a Levenberg-Marquardt guidance optimizer for sampling hand-body sequence that align with image observations. Our results are enabled by a key", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7072", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "insight: that the representation used for head pose conditioning is critical for accurate full-body motion estimation. We study choices for this representation by (1) identifying desirable spatial and temporal invariance properties that are not fulfilled by existing systems and (2) using these properties to derive improved parameterizations for our motion prior.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We systematically evaluate our system on four datasets. For body estimation, we find that improving the conditioning parameterization leads to an accuracy improvement between $4.9\\%$ and $17.9\\%$ . Furthermore, we observe that the resulting system can improve hand estimation, reducing world-frame errors by over $40\\%$ compared to single-frame estimates. Code, model, and more results can be found on our project webpage.", + "bbox": [ + 89, + 181, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 301, + 225, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D human recovery from external visual inputs. A large body of work has addressed estimating the parameters of human body models like SCAPE [2] or SMPL and its variants [52, 63, 79] from third-person visual inputs, where human subjects are observed from the view of outside cameras. The majority of these works focus on extracting 3D representations from single images, for example by lifting 2D keypoint observations to 3D [58], via end-to-end regression [20, 31, 33, 40, 60, 62, 77], via optimization [19, 44, 63], or by exploiting synergies between regression and optimization [42]. When multiple frames are available in the form of a video, temporal context and tracking can also be incorporated [16, 34, 39, 64, 65, 71, 108]. The inputs (images) and outputs (human meshes) of many of these systems are similar to the egocentric setting addressed by EgoAllo, but egocentric devices present unique challenges because the body being estimated is typically behind the outwards-facing cameras used as input.", + "bbox": [ + 89, + 328, + 482, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Priors for human motion. The primary challenge of ego-sensed human motion estimation is limited observability; a prior is required to resolve ambiguities. For human motion, these priors are typically framed as unconditional distributions over plausible human motion. Distributions can be represented either by modeling the physical constraints of our world [6, 50, 67, 73] or by learning generative models of human motion directly from data. For learning unconditional priors, classical data-driven approaches include fitting mixtures-of-Gaussians to 3D keypoint trajectories [25], while modern approaches include training variational autoencoders [38, 75] to model either autoregressive transitions [15, 51, 74] or full spatiotemporal sequences [22]. After training, these priors can be applied to estimation problems in iterative optimization frameworks [41, 74, 102]. EgoAllo is built on the same intuition as these methods, but follows previous work in ego-sensed motion estimation and uses a task-specific conditional prior.", + "bbox": [ + 89, + 590, + 482, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising diffusion for human motion. The core of EgoAllo is a denoising diffusion model [23, 69, 84] from which we can sample 3D human body motion. While diffusion mod", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els are primarily known for their success in text-conditioned image generation [78, 80], they have also enabled advances in human motion synthesis conditioned on modalities like text [35, 36, 112], music [1, 93], poses [35, 48], and object geometry [43, 47, 49]. EgoAllo adopts a similar conditional diffusion approach, while specifically studying the design of conditioning parameters used for ego-sensed human motion estimation. The iterative nature of denoising diffusion also enables guidance [11, 13, 30, 35, 87, 114], where denoising steps are steered to satisfy a desired objective. We use guidance to incorporate observations like visual hand pose observations during test-time.", + "bbox": [ + 511, + 90, + 906, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human motion from egocentric observations. EgoAllo builds on intuition from several prior works in egocentric sensing for human motion estimation. Many rely on fisheye cameras that place the wearer's body into the field of view [27, 76, 91, 91, 92, 95, 96, 98]. Other approaches rely on body-mounted cameras [83], simulation-based physical plausibility [54, 106, 107], body- and hand-mounted inertial sensors [37, 46, 104, 105], handheld controllers [7, 28, 29], and interaction cues from other humans [59]. Concurrent works have also used the Nymeria [55] dataset for egocentric motion with language description outputs [24], as well for online settings with scene geometry and CLIP [70] feature inputs [21]. Most relevantly, EgoEgo [48] demonstrates how human body poses can be estimated offline without body observability assumptions. The authors accomplish this by carefully integrating several components: a monocular SLAM system [89], a pose-conditioned gravity vector regression network, an optical flow feature-conditioned head orientation and scale regression network, and a head pose-conditioned body diffusion model. EgoAllo differs in both inputs—we study conditioning parameters computed from the metric SLAM poses provided by devices like Project Aria [85]—and outputs—we consider body height variation and hand poses.", + "bbox": [ + 511, + 268, + 908, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditioning for ego-sensed poses. Prior works vary in how head pose information is parameterized and used as neural network input. AvATARPoser [28] and BoDiffusion [7] parameterize head pose as four components: world-frame orientation, orientation deltas, world-frame position, and world-frame position deltas. These works focus on settings with VR controller input, and parameterize controller pose inputs the same way. EgoEgo [48]'s diffusion model uses only absolute head positions and orientations, but similar to HuMoR [74], in implementation defines a per-sequence canonical coordinate frame to ensure that all input trajectories passed to the model are aligned with the same initial $xy$ position and forward direction. In our work, we refer to this as sequence canonicalization. Finally, EgoPoser [29] proposes a similar scheme that aligns initial positions for both head pose and controller pose inputs. We propose an alternative to these parameterizations that is motivated by the robustness and generalization benefits of invariance, as observed in prior work for designing both", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/db89cba97a9938764e3ab6e3b4031644470fdfb31a685e983ad268b1d6b1905f.jpg", + "image_caption": [ + "Figure 2. Overview of components in EgoAllo. We restrict the diffusion model to local body parameters (Section 3.1.1). An invariant parameterization $g(\\cdot)$ (Section 3.1.2) of SLAM poses is used to condition a diffusion model. These can be placed into the global coordinate frame via global alignment (Section 3.2.1) to input poses. When available, egocentric video is used for hand detection via HaMeR [66], which can be incorporated into samples via guidance (Section 3.2.2)." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 897, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "representations [10, 53, 81, 97, 103, 110, 115] and neural network architectures [8, 9, 12, 14, 32, 45, 90, 100, 101, 109]. Specifically, we introduce in Section 3.1.2 a parameterization that is invariant to both spatial and temporal shifts.", + "bbox": [ + 88, + 375, + 482, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 448, + 178, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We study the problem of using sensors from an egocentric device to estimate the actions of the wearer in an allocentric coordinate frame. We assume a flat floor and two inputs: poses from the device's SLAM system and egocentric video.", + "bbox": [ + 88, + 474, + 482, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our system uses head pose information to condition a diffusion-based prior over body pose and height, and incorporates visual hand observations during sampling. This allows it to benefit from both 3D human motion capture datasets [56], which are used for the motion prior, and from large-scale image datasets [66], which are used for hand estimates.", + "bbox": [ + 88, + 534, + 482, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Ego-conditioned motion diffusion", + "text_level": 1, + "bbox": [ + 89, + 633, + 382, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation: we use $\\mathbf{T}_{\\mathrm{A,B}} = (\\mathbf{R}_{\\mathrm{A,B}},\\mathbf{p}_{\\mathrm{A,B}})$ to denote an SE(3) transform to frame A from frame B, composed of rotation $(\\mathbf{R}_{\\mathrm{A,B}})$ and position $(\\mathbf{p}_{\\mathrm{A,B}})$ terms. Temporal steps $t$ are superscripted and diffusion noise steps $n$ are subscripted. $\\vec{x}_0^t$ thus refers to the $t$ -th timestep of a clean $(n = 0)$ human motion sequence.", + "bbox": [ + 88, + 656, + 482, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an observation window of $T$ timesteps, EgoAllo's motion prior is a diffusion model that aims to capture the distribution of human motions $\\vec{x}_0 = \\{\\vec{x}_0^1,\\dots,\\vec{x}_0^T\\}$ conditioned on head pose encodings $\\vec{c} = \\{\\vec{c}^1,\\dots,\\vec{c}^T\\}$ . For each timestep $t$ , we represent human motion in the form of SMPL-H [52, 79] model parameters $\\{\\mathbf{T}_{\\mathrm{world,root}}^t,\\Theta^t,\\beta \\}$ : root transforms $\\mathbf{T}_{\\mathrm{world,root}}^t\\in \\mathrm{SE}(3)$ , where the person's root frame is located at their pelvis, local joint rotation matrices $\\Theta^t\\in \\mathbb{R}^{51\\times 3\\times 3}$ , and time-invariant shape $\\beta \\in \\mathbb{R}^{16}$ .", + "bbox": [ + 88, + 732, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dependencies between local joint rotations, body size variation, and global motion make this learning task a", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "challenging one. Our key insight is that this difficulty can be reduced by designing parameterizations with desirable invariance properties. Spatial and temporal invariances allow the model to focus on the essential structure of motion, without being affected by irrelevant shifts in position or time.", + "bbox": [ + 511, + 375, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Diffusion output representation", + "text_level": 1, + "bbox": [ + 511, + 465, + 779, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As output, we sample body and hand joint rotations, body shapes, and binary contact predictions $\\vec{x}_0^t = \\{\\Theta^t,\\beta^t,\\psi_{j = 1\\dots 21}^t\\}$ , where body shape $\\beta^t$ is supervised to be equal for all timesteps and $\\psi_j^t$ is a per-joint contact indicator. Notably, these parameters are all local—we discuss how outputs can be placed into the allocentric coordinate frame in Section 3.2.1.", + "bbox": [ + 511, + 488, + 906, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We choose this output set for three main reasons. (1) Body shape encodes the wearer's height, which is critical for grounding in the metric-scale geometry of the scene. This is rarely considered by prior work: with the exception of [29], existing methods [7, 28, 48] otherwise produce outputs using a fixed \"mean\" human shape. (2) Contact predictions enable losses for common problems like foot skating, which are discussed in Section 3.2.2. (3) Finally, local bodies are invariant to the global coordinate frame. As we discuss next, the conditioning parameterization for the model can therefore also be invariant to arbitrary transformations along the floor plane.", + "bbox": [ + 511, + 582, + 905, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Invariant conditioning", + "text_level": 1, + "bbox": [ + 511, + 763, + 718, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of our conditioning representation is to map raw SLAM poses (head motion) to a parameterization that is amenable to learning for the diffusion model.", + "bbox": [ + 511, + 786, + 905, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Raw inputs. To capture the head motion at each time step, we assume as input poses of a central pupil frame (CPF), which the SLAM systems of devices like Project Aria can provide with millimeter-level accuracy [85]. For time 1...T, we", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7074", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "reparameterize these poses for conditioning using a function $g$ :", + "bbox": [ + 89, + 90, + 482, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} = \\left(\\mathbf {R} _ {\\text {w o r l d , c p f}} ^ {t}, \\mathbf {p} _ {\\text {w o r l d , c p f}} ^ {t}\\right) \\in \\mathrm {S E} (3), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 116, + 482, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\vec {c} ^ {1}, \\dots , \\vec {c} ^ {T} \\} = g (\\{\\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {1}, \\dots , \\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {T} \\}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 137, + 483, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The CPF frame differs from prior works that condition on a coordinate frame attached to the SMPL human model's \"head joint\" [7, 28, 29, 48]. The offset between this head joint and the device pose depends on the head shape captured by $\\beta^t$ , and is thus difficult to precompute in our setting.", + "bbox": [ + 89, + 166, + 482, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encode absolute height, we assume that the world frame's $+z$ -axis faces upwards, and that the ground is located at $z = 0$ . Ground parameters are directly available in the training data [56]; at test time, we can also extract these parameters from sparse SLAM points via RANSAC (Appendix A.3.5).", + "bbox": [ + 89, + 241, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Invariance goals. As discussed in Section 2, prior work varies in how the function $g$ is implemented. To understand how choices impact learning, we propose two invariance properties for head motion representations. Each reduces representational redundancy, which eases the learning problem. We provide more detailed explanations and visualizations of these properties in Appendix A.1.", + "bbox": [ + 89, + 323, + 483, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Invariance 1 (Spatial) Global transformations along the floor plane should not affect a person's local motion. Given $\\mathbf{T}_{xy} \\in SE(3)$ restricted to the $XY$ plane, $g$ should fulfill $g(\\{\\mathbf{T}_{xy}\\mathbf{T}_{world, cpf}^t\\}_t) = g(\\{\\mathbf{T}_{world, cpf}^t\\}_t) \\forall \\mathbf{T}_{xy}$ .", + "bbox": [ + 88, + 438, + 483, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Invariance 2 (Temporal) If the same human motion occurs at different points in time, the extracted head motion representation should remain consistent. This can be expressed as temporal shift equivariance. Let $\\vec{c}^t$ be as defined in Equation 2. For any shift $\\delta$ such that $\\{\\vec{c}_{shift}^{1},\\dots,\\vec{c}_{shift}^{T}\\} = g(\\{\\mathbf{T}_{world,cpf}^{1 + \\delta},\\dots,\\mathbf{T}_{world,cpf}^{T + \\delta}\\})$ , $g$ should satisfy $\\vec{c}_{shift}^{t} = \\vec{c}^{t + \\delta}$ for overlapping timesteps.", + "bbox": [ + 89, + 510, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "No parameterization used by existing work satisfies both of these properties. The sequence canonicalization approach of EgoEgo [48] achieves spatial invariance (Invariance 1), but inserts a sequence-wide dependency on the first timestep of each window that results in a violation of Invariance 2. The absolute poses and pose deltas used by [7, 28] satisfy Invariance 2, but not Invariance 1. Finally, the relative positions considered by [29] are neither spatially nor temporally invariant.", + "bbox": [ + 89, + 614, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Invariant conditioning. We propose a formulation for $g$ that achieves both invariance properties by locally canonicalizing head motion with respect to the floor at each timestep. We build on the relative motion of the CPF frame at each time $t$ , which respects both Invariance 1 and 2:", + "bbox": [ + 89, + 741, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {T} _ {\\mathrm {c p f}} ^ {t - 1, t} = \\left(\\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {t - 1}\\right) ^ {- 1} \\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 825, + 483, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Importantly, the translation component of this transformation is in the local frame. This is distinct from world-frame position deltas [7, 28, 29], which still violate Invariance 1.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/abcbbc04c56a29d9b6148f942092360c06258d94a86b387c001f31a861ac83b5.jpg", + "image_caption": [ + "Figure 3. Locally canonicalized coordinate frames. We compute our invariant conditioning parameterization (Equation 4) using transformations computed from three coordinate frames. Following [85], the CPF has the $z$ -axis forward. Following HuMoR [74], the world and canonical $z$ -axes point up. Canonical frames are computed by projecting the CPF frame origin to the ground plane, then aligning the canonical $y$ -axis to the CPF forward direction. The axes are color-coded as $x$ (red), $y$ (green), and $z$ (blue)." + ], + "image_footnote": [], + "bbox": [ + 563, + 75, + 857, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relative transforms alone do not encode information relative to the scene or floor: full trajectories can even be flipped upside down without impacting $\\Delta \\mathbf{T}_{\\mathrm{cpf}}^{t - 1,t}$ . We therefore propose to ground relative motion to the floor plane with a transformation between the CPF frame and a per-timestep canonical frame, which is computed by projecting the CPF frame to the floor. This encodes head height and orientation. Our full representation then becomes:", + "bbox": [ + 511, + 431, + 906, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\vec {c} ^ {t} = \\underbrace {\\left\\{\\Delta \\mathbf {T} _ {\\mathrm {c p f}} ^ {t - 1 , t} , \\quad \\left(\\mathbf {T} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t}\\right) ^ {- 1} \\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} \\right\\}} _ {\\text {I n v a r i a n t i m p l e m e n t a t i o n o f} g (\\cdot)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 561, + 906, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We visualize an example of a canonical frame in Figure 3 and our full representation in Appendix A.1. Canonical frames are positioned by projecting the CPF origin to the floor plane; given standard bases $\\mathbf{e}_{\\{x,y,z\\}}$ , we compute:", + "bbox": [ + 511, + 617, + 905, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t} = \\left[ \\begin{array}{l l l} \\mathbf {e} _ {x} & \\mathbf {e} _ {y} & \\vec {0} \\end{array} \\right] ^ {\\top} \\mathbf {p} _ {\\text {w o r l d , c p f}} ^ {t}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 689, + 906, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For orientation, we align the canonical frame's local $z$ -axis parallel to the world $z$ -axis and its local $y$ -axis toward the \"forward\" direction $\\vec{v}^t$ of the CPF frame. With $\\mathbf{R}_{z}(\\cdot):\\mathbb{R}\\to \\mathrm{SO}(3)$ constructing a $z$ -axis rotation and $\\mathbf{e}_{\\{x,y,z\\}}$ again as standard bases, we compute this as:", + "bbox": [ + 511, + 720, + 906, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\vec {v} ^ {t} = \\mathbf {R} _ {\\text {w o r l d , c p f}} ^ {t} \\mathbf {e} _ {z}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 806, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t} = \\mathbf {R} _ {z} \\left(- \\arctan 2 \\left(\\mathbf {e} _ {x} ^ {\\top} \\vec {v} ^ {t}, \\mathbf {e} _ {y} ^ {\\top} \\vec {v} ^ {t}\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 827, + 906, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This canonical frame definition is an important departure from prior work. While EgoEgo [48] and HuMoR [74] use similar canonical frames, they only compute one per sequence. Instead,", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we compute Equations 5 and 7 at every timestep. This enables floor plane grounding without sacrificing Invariance 2.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Connection to prior work. EgoAllo's invariant conditioning shares similarities with the spatial and temporal normalization proposed by EgoPoser [29]. (1) EgoPoser's temporal normalization subtracts positions by the first timestep's position. This is a position-only version EgoEgo [48]'s sequence canonicalization. It does not consider rotation, and therefore only partially fulfills Invariance 1. Like EgoEgo, it also does not fulfill Invariance 2. In Appendix A.2.3, we observe that this approach only has a small impact when adapted to EgoAllo's problem setting. (2) EgoPoser's spatial normalization subtracts the head's XY position from the hand pose inputs. EgoAllo does not use hand pose as network input, so this is not applicable in our setting.", + "bbox": [ + 89, + 121, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Estimation via sampling", + "text_level": 1, + "bbox": [ + 89, + 311, + 313, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use our local body representation and invariant conditioning strategies to train a motion prior in the form of a denoising diffusion model [23]. Given diffusion step $n = N \\ldots 1$ , we follow [72] and approximate the denoising process as:", + "bbox": [ + 89, + 334, + 483, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\vec {x} _ {n - 1} | \\vec {x} _ {n}, \\vec {c}) = \\mathcal {N} \\big (\\mu_ {\\theta} (\\vec {x} _ {n}, n, \\vec {c}), \\sigma_ {n} ^ {2} \\mathbf {I} \\big), \\qquad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 406, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where a transformer [94] $\\mu_{\\theta}$ is trained to predict the posterior mean from noised sample $\\vec{x}_n$ and conditioning $\\vec{c}$ . With noise-dependent weight term $w_{n}$ , the loss can be written as:", + "bbox": [ + 89, + 435, + 483, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\vec {x} _ {0}} \\mathbb {E} _ {n \\sim \\mathcal {U}} \\left[ w _ {n} \\| \\mu_ {\\theta} (\\vec {x} _ {n}, n, \\vec {c}) - \\vec {x} _ {0} \\| ^ {2} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 491, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training, we estimate human motions by following DDIM [86] for sampling. The final EgoAllo sampling procedure includes several additional components: a global alignment phase, guidance losses for physical constraints and visual hand observations, and a path fusion [3] approach for longer sequence lengths. We describe these below.", + "bbox": [ + 89, + 525, + 483, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Global alignment", + "text_level": 1, + "bbox": [ + 89, + 625, + 256, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To place sampled bodies into the allocentric coordinate system, we compute the absolute pose of the SMPL-H root as:", + "bbox": [ + 89, + 642, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {\\text {w o r l d , r o o t}} ^ {t} = \\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} \\mathbf {T} _ {\\text {c p f , r o o t}} ^ {(\\Theta^ {t}, \\beta^ {t})}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 684, + 483, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_{\\mathrm{cpf,root}}^{(\\Theta^t,\\beta^t)}$ computes the transform between the root of the human and their CPF frame for a given set of local pose and shape parameters. Similar processes are applied in [7, 28, 29]. In contrast to directly outputting absolute body transformations from the diffusion model [48], this guarantees exact alignment between estimates and input SLAM sequences.", + "bbox": [ + 89, + 718, + 483, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Guidance losses", + "text_level": 1, + "bbox": [ + 89, + 821, + 246, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our diffusion model learns a distribution of human motion conditioned on the central pupil frame motion. At test time, we incorporate constraints from physical priors and visual hand observations via guidance [11, 30, 114]. Similar to [35, 47],", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we accomplish this by applying costs to the joint rotations $\\Theta = \\{\\Theta^1,\\dots,\\Theta^T\\}$ predicted by $\\mu_{\\theta}(\\vec{x}_n,n,\\vec{c})$ . We treat the body shape $\\beta^{t}$ and contacts $\\psi_{j = 1\\dots 21}^{t}$ as fixed and optimize over body and finger pose to minimize hand observation, skating, and prior costs with a Levenberg-Marquardt optimizer:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\text {g u i d a n c e}} ^ {(\\Theta)} = \\mathcal {E} _ {\\text {h a n d s}} ^ {(\\Theta)} + \\mathcal {E} _ {\\text {s k a t e}} ^ {(\\Theta)} + \\mathcal {E} _ {\\text {p r i o r}} ^ {(\\Theta)}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 172, + 906, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by running HaMeR on the egocentric image corresponding to each timestep $t$ . When detected, this produces 3D hand estimates in the form of MANO [79] joint parameters and camera-centric 3D hand keypoints $\\hat{\\mathbf{p}}_{\\mathrm{camera},j}^{t}$ for hand joint set $j \\in \\mathcal{H}$ . Optionally, wrist and palm poses can also be estimated using Project Aria's Machine Perception Services [85]. With each subscripted $\\lambda$ indicating a scalar weighting term, we have:", + "bbox": [ + 511, + 199, + 905, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\text {h a n d s}} ^ {(\\Theta)} = \\lambda_ {\\text {h a n d s 3 D}} \\mathcal {E} _ {\\text {h a n d s 3 D}} ^ {(\\Theta)} + \\lambda_ {\\text {r e p r o j}} \\mathcal {E} _ {\\text {r e p r o j}} ^ {(\\Theta)}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 311, + 906, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The 3D objective $\\mathcal{E}_{\\mathrm{hands3D}}^{(\\Theta)}$ minimizes the distance between the detected hand parameters and the corresponding SMPL-H hand parameters, in terms of wrist pose and local joint rotations. With $\\Pi_K$ as projection with camera intrinsics $K$ , $\\mathbf{p}_{\\mathrm{world},j}^{(\\Theta^t)} \\in \\mathbb{R}^3$ as the world position for joint $j$ at time $t$ , and $\\mathbf{T}_{\\mathrm{camera,cpf}}$ from the device calibration, the reprojection cost is:", + "bbox": [ + 511, + 342, + 906, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\text {r e p r o j}} ^ {(\\Theta)} = \\sum_ {t, j \\in \\mathcal {H}} | | \\Pi_ {K} \\left(\\mathbf {p} _ {\\text {c a m e r a}, j} ^ {(\\Theta^ {t})}\\right) - \\Pi_ {K} \\left(\\hat {\\mathbf {p}} _ {\\text {c a m e r a}, j} ^ {t}\\right) | | _ {2} ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 444, + 906, + 478 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {\\text {c a m e r a}, j} ^ {(\\Theta^ {t})} = \\mathbf {T} _ {\\text {c a m e r a , c p f}} \\left(\\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t}\\right) ^ {- 1} \\mathbf {p} _ {\\text {w o r l d}, j} ^ {(\\Theta^ {t})}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 481, + 906, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reduce foot skating, we use contact predictions to apply a skating cost [74, 102] for each time $t$ and joint $j$ :", + "bbox": [ + 511, + 508, + 905, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\text {s k a t e}} ^ {(\\Theta)} = \\sum_ {t, j} \\lambda_ {\\text {s k a t e}} \\| \\frac {1}{2} \\left(\\psi_ {j} ^ {t} + \\psi_ {j} ^ {t - 1}\\right) \\left(\\mathbf {p} _ {\\text {w o r l d}, j} ^ {t} - \\mathbf {p} _ {\\text {w o r l d}, j} ^ {t - 1}\\right) \\| _ {2} ^ {2}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 546, + 905, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we minimize a prior cost $\\mathcal{E}_{\\mathrm{prior}}^{(\\Theta)}$ . This cost penalizes deviations between joint rotations $\\Theta^t$ and rotations $\\hat{\\Theta}^t$ from the denoiser $\\mu_{\\theta}(\\vec{x}_n,n,\\vec{c})$ , while also encouraging smoothness. We detail the full list of terms used for this in Appendix A.3.4.", + "bbox": [ + 511, + 590, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Sequence length extrapolation", + "text_level": 1, + "bbox": [ + 511, + 662, + 769, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For longer sequences at test time, we draw on existing methods in compositional generation for both image [3, 113] and human motion [4, 82] diffusion models. We train our motion prior using subsequences of up to length 128; when input observations exceed this length at test time, we split into windows with a 32-timestep overlap between neighbors. We then run our model $\\mu_{\\theta}(\\vec{x}_n,\\vec{c},n)$ on windows in parallel. Diffusion paths for overlapping regions are fused following MultiDiffusion [3] after each denoising step.", + "bbox": [ + 511, + 681, + 906, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 830, + 638, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a series of experiments to evaluate EgoAllo's conditioning parameterization, body estimation accuracy, and hand estimation performance.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7076", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/751f316ffcaf71ddb2efe96f8514124c1006a53a3368c46a471b2e8acce43079.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConditioningSeqlenInvariance 1/2MPJPE ↓% DiffPA-MPJPE ↓% DiffGND ↑
EgoAllo (Eq. 4)32✓ /✓129.8±1.1—109.8±1.1—0.98±0.00
Absolute+Local Relative32P /✓133.0±1.12.4%113.6±1.23.4%0.95±0.00
Absolute+Global Deltas [7,28]32X /✓136.2±1.14.9%118.3±1.27.7%0.93±0.01
Sequence Canonicalization [48]32✓ /X153.1±1.517.9%128.7±1.517.1%0.76±0.01
Absolute32X /✓159.9±1.223.2%141.0±1.328.4%0.89±0.01
EgoAllo (Eq. 4)128✓ /✓119.7±1.3—101.1±1.3—1.00±0.00
Absolute+Local Relative128P /✓124.5±1.34.0%104.9±1.43.8%1.00±0.00
Absolute+Global Deltas [7,28]128X /✓127.4±1.36.4%109.8±1.48.6%0.99±0.00
Sequence Canonicalization [48]128✓ /X134.0±1.811.9%112.1±1.610.9%0.88±0.02
Absolute128X /✓148.3±1.523.9%131.2±1.629.8%0.96±0.01
", + "bbox": [ + 91, + 89, + 903, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Motion prior conditioning comparison. We train and evaluate otherwise identical models using four conditioning parameterizations on AMASS [56] test set sequences, using sequences of length 32 and 128. Parameterizations vary in their spatial (1) and temporal (2) invariance properties, which we loosely classify as following completely ( $\\checkmark$ ), partially ( $\\mathsf{P}$ ), or not at all ( $\\times$ ). The conditioning parameterization used by EgoAllo reduces errors by almost $18\\%$ compared to the sequence canonicalization approach used by the most relevant related work [48].", + "bbox": [ + 89, + 280, + 906, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. To train EgoAllo models used in our experiments, we need sequences containing human body and hand pose parameters, body shapes, and device SLAM poses $\\mathbf{T}_{\\mathrm{world,cpf}}^t$ . Similar to prior work [7, 28, 48], we train EgoAllo using AMASS [56] with synthesized device poses. We annotate train split sequences by anchoring a central pupil frame between vertices corresponding to the left and right pupils in the blend skinned mesh, and at train time sample sequences between length 32 and 128.", + "bbox": [ + 88, + 356, + 482, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We evaluate with four datasets. We use AMASS [56], RICH [26], and Aria Digital Twins (ADT) [61] for body estimation evaluation, and EgoExo4D [17] for hand estimation evaluation. AMASS and RICH do not include egocentric data; we annotate these with synthetic device poses using the same procedure we use for training. ADT and EgoExo4D both include egocentric images and SLAM poses captured using Project Aria glasses [85], which we use directly.", + "bbox": [ + 88, + 486, + 482, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To quantify performance, we report four metrics: (1) MPJPE is a world-frame mean per-joint position error (millimeters). (2) PA-MPJPE is the Procrustes-aligned mean per-joint position error in millimeters, where joint positions are aligned on a per-timestep basis before error are computed. (3) GND is a grounding metric, designed in response to a phenomena where ego-sensed humans \"float\" above the ground. Given a human body trajectory, this metric contains a simple binary indicator of whether the feet of the human ever touch the ground plane. (4) $\\mathbf{T}_{\\mathrm{head}}$ is the average SMPL head joint position error in millimeters.", + "bbox": [ + 88, + 616, + 482, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Body estimation", + "text_level": 1, + "bbox": [ + 89, + 800, + 251, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our first set of experiments, we evaluate body estimation from only device SLAM poses, without considering images or hands. This setting allows us to isolate the advantages of our body motion prior, while directly comparing against methods that do not consider hands.", + "bbox": [ + 89, + 825, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Invariant conditioning evaluation", + "text_level": 1, + "bbox": [ + 511, + 356, + 792, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by evaluating the importance of the spatial and temporal invariance criteria discussed in Section 3.1.2. We do this by comparing five implementations of the conditioning $g$ : (1) EgoAllo is the final invariant representation that we propose in Equation 4. (2) Absolute+Local Relative appends absolute poses with the relative pose deltas written in Equation 3. (3) Absolute+Global Deltas appends absolute poses with relative orientation and the world-frame position deltas used by [7, 28]. (4) Sequence Canonicalization uses the alignment approach implemented by [48], which violates temporal invariance. (5) Absolute naively conditions on absolute poses, which violate spatial invariance.", + "bbox": [ + 511, + 375, + 906, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train conditional diffusion models with otherwise identical architecture using each parameterization, and then evaluate on the AMASS [56] test set. Metrics and percent differences compared to EgoAllo are reported in Table 1.", + "bbox": [ + 511, + 556, + 906, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, we find that the choice of conditioning parameterization makes a dramatic impact on estimation accuracy. We observe accuracy improve consistently as invariance properties are incorporated into the representation. Compared to EgoAllo, Absolute conditioning increases MPJPE by over $23\\%$ for both shorter (length 32) and longer (length 128) sequences. Compared to EgoAllo, SeqCanonical conditioning increases MPJPE by nearly $18\\%$ for length 32 sequences and $12\\%$ for length 128 sequences.", + "bbox": [ + 511, + 618, + 908, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Comparisons against baselines", + "text_level": 1, + "bbox": [ + 511, + 761, + 772, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further study EgoAllo's body estimation quality, we compare against three baselines. (1) NoShape. First, NoShape refers to a variation of EgoAllo that turns off shape estimation, and thus cannot estimate the wearer's height. (2) EgoEgo. We also compare against the human motion diffusion model from EgoEgo [48]. This is similar to EgoAllo, but considers only the SMPL \"mean\" body shape and uses sequence canonicalized coordinates for conditioning and as model output. (3) VAE+Opt.", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7077", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/96e17e5803918f0e186ddc83057ea0b016244293400a2a15b76c4ba9e2959677.jpg", + "table_caption": [ + "AMASS [56]" + ], + "table_footnote": [], + "table_body": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑Thead ↓
EgoAllo32129.8±1.1109.8±1.10.98±0.006.4±0.1
NoShape32138.1±1.1118.8±1.10.94±0.0144.7±0.4
EgoEgo32184.0±1.5158.6±1.60.81±0.0145.2±1.0
VAE+Opt32199.5±1.3191.4±1.40.49±0.0178.0±1.5
EgoAllo128119.7±1.3101.1±1.31.0±0.006.2±0.1
NoShape128128.1±1.3110.3±1.40.98±0.0144.6±0.7
EgoEgo128167.4±2.1145.8±2.00.92±0.0154.9±1.9
VAE+Opt128205.3±2.6192.3±2.80.75±0.0267.8±3.1
", + "bbox": [ + 94, + 108, + 478, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b123743caef541c13437857d45292ecc0cee9404f2846963833d5c1a3d77222.jpg", + "table_caption": [ + "RICH [26]" + ], + "table_footnote": [], + "table_body": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\\( T_{head} \\) ↓
EgoAllo32193.7±3.4174.8±3.60.95±0.018.8±0.2
NoShape32200.9±3.3183.3±3.60.73±0.0244.9±0.9
EgoEgo32215.4±3.9192.9±4.00.73±0.0256.2±2.9
VAE+Opt32352.0±6.7354.8±6.50.59±0.02319.3±11.6
EgoAllo128176.2±5.6160.1±5.90.96±0.028.9±0.3
NoShape128185.7±5.5169.9±5.80.82±0.0345.8±1.6
EgoEgo128207.8±6.9187.8±6.80.88±0.0366.5±5.4
VAE+Opt128319.8±10.1323.8±10.50.75±0.04274.4±17.6
", + "bbox": [ + 96, + 252, + 478, + 380 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf7c5deabaef79dfa0733c695ca158ee3f2f8f8b381c5cf2a16b12b9d55e90f5.jpg", + "table_caption": [ + "Aria Digital Twins [61]" + ], + "table_footnote": [], + "table_body": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\\( T_{head} \\)↓
EgoAllo32173.5±1.1146.1±1.10.88±0.01-
NoShape32178.5±1.1153.0±1.10.89±0.01-
EgoEgo32212.5±1.4181.3±1.60.64±0.01-
VAE+Opt32284.9±1.6283.9±1.90.63±0.01-
EgoAllo128155.1±1.6129.3±1.60.94±0.01-
NoShape128163.7±1.6140.0±1.60.96±0.01-
EgoEgo128182.6±2.3153.9±2.60.73±0.02-
VAE+Opt128290.8±3.8282.5±4.40.7±0.02-
", + "bbox": [ + 94, + 397, + 478, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Body estimation performance, compared against a baseline without shape prediction, EgoEgo [48], and VAE+Opt [74, 102]. We exclude the $\\mathbf{T}_{\\mathrm{head}}$ metric for ADT because the Biomech57 head joints used by ADT are not directly comparable to the SMPL-H head joints used by our model.", + "bbox": [ + 89, + 545, + 483, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we compare against an approach based on the SLAHMR [102] framework for human motion estimation from exocentric video. A key advantage of SLAHMR is that it uses an unconditional motion prior [74] in an optimization framework. It can therefore be adapted to new settings without re-training—we keep the same body pose and shape variables as the original pipeline, but replace the exocentric keypoint [71] cost with an egocentric CPF pose alignment cost.", + "bbox": [ + 89, + 633, + 482, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to differences in problem formulation, many existing methods for egocentric human motion estimation are difficult to directly compare. This is particularly true when they have different inputs, such as fisheye cameras [91, 95, 96], wrist-mounted sensors [46], or handheld controller poses [7, 28, 29]. Additionally, prior works like EgoEgo [48] do not incorporate vision inputs for hand estimation. For fairness, we restrict all methods in this section to only CPF or head pose as input.", + "bbox": [ + 89, + 756, + 483, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EgoAllo improves body motion estimates. We report metrics", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1dd517f31a214f45b04cf06f7e824d6a01737aa765229080bec506d6b7e7b73.jpg", + "image_caption": [ + "(a) Ground-truth" + ], + "image_footnote": [], + "bbox": [ + 535, + 101, + 696, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2da2e150065507518628a2247b7d7d78001b2fc6b0c5a050d9b3c92a6a194ae0.jpg", + "image_caption": [ + "(b) EgoAllo" + ], + "image_footnote": [], + "bbox": [ + 725, + 93, + 890, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30685c4828fe440c3df429c5eb042a642232ee2915171b997b860880d7c3bbab.jpg", + "image_caption": [ + "(c) EgoEgo [48]" + ], + "image_footnote": [], + "bbox": [ + 527, + 189, + 702, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fc72bfc266b0a35cc57f96ad7ae2d72897133c0c504bdd32fdd0f350ef3dbbf.jpg", + "image_caption": [ + "(d) VAE+Opt", + "Figure 4. Egocentric human motion estimation for a running sequence. We show the ground-truth, an output from EgoAllo, and outputs from two baselines. The glasses CAD model is placed at the conditioning transformation $\\mathbf{T}_{\\mathrm{world,cpf}}$ ." + ], + "image_footnote": [], + "bbox": [ + 728, + 191, + 890, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Table 2 and visualize example outputs in Figure 4. We find that EgoAllo enables significant estimation improvements across all datasets, including accuracy improvements of $20\\sim 30\\%$ over EgoEgo for both shorter and longer evaluation sequences. We found shape estimation critical for producing metric-scale, grounded estimates of human body motion, with the head aligned to input SLAM poses and the feet planted on the observed ground plane. This is evident in qualitative results, improved grounding metrics, and in the $6\\sim 7\\%$ MPJPE gap between EgoAllo and the NoShape ablation.", + "bbox": [ + 511, + 380, + 906, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VAE optimization converges poorly. Optimization-based estimation approaches have been effective for settings with keypoint costs [74, 102], but we found convergence difficult in our less constrained setting. In Table 2, we observe poor generalization: VAE+Opt performs competitively on the AMASS test set, but performance deteriorates dramatically when evaluating on RICH or ADT. VAE+Opt outputs in Figure 4 also look overly smoothed, without the same expressiveness as the conditional predictions of EgoAllo or EgoEgo [48]. This highlights the advantage of using a conditional diffusion model problem for this estimation problem.", + "bbox": [ + 511, + 542, + 908, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shape estimation evaluation. To better understand the shape estimation characteristics of EgoAllo, we compare against against the \"mean\" shape used by EgoEgo and the NoShape ablation. On the AMASS test set, we find: EgoAllo slightly improves overall shape (19mm → 18mm mean vertex-to-vertex error) and produces much better height (52mm → 32mm mean height error), but is not able to generalize in terms of body weight (5kg → 8kg mean weight error). The body shape is inferred from the wearer's head pose, which intuitively provides strong height constraints but is less correlated with weight. Accurate height is key for proper scene placement, as reflected by both the MPJPE and GND metrics.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7078", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Hand estimation", + "text_level": 1, + "bbox": [ + 89, + 90, + 254, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate hands estimated by EgoAllo, we run HaMeR on the segment of the EgoExo4D [17] validation set that is labeled with 3D hand pose keypoints. We quantitatively compare four hand estimation methods in Table 3. In (1) HaMeR [66], we use HaMeR out-of-the-box on undistorted egocentric RGB images. We do not assume bounding boxes as input; instead, we follow the HaMeR demo code and compute crops using ViTPose [99]. (2) EgoAllo-NoReproj uses all loss terms except for the reprojection loss (Equation 14). Hand guidance is done directly using the 3D wrist poses predicted by HaMeR. (3) EgoAllo-Mono is the same as EgoAllo-NoReproj, but guides hands using the reprojection loss. This accounts for the scale ambiguities that are inherent to the single-frame HaMeR estimates. Finally, we show an extension of our method. (4) EgoAllo-Wrist3D extends EgoAllo-Mono with 3D wrist pose losses from Project Aria's Machine Perception Services [85]—unlike HaMeR, which assumes monocular input, this is a separate sensing setup that also leverages a pair of SLAM cameras that are unique to Project Aria. For fairness across settings, we compute metrics only on timesteps where HaMeR estimates are available.", + "bbox": [ + 89, + 113, + 483, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results are provided in Table 3. While HaMeR's local poses (PA-MPJPE) are slightly better, EgoAllo's hand-body estimation significantly improves how well hands are estimated in the world coordinate system. Compared to HaMeR, EgoAllo-Mono drops MPJPE from $237.90\\mathrm{mm} \\rightarrow 131.45\\mathrm{mm}$ . Reprojection-based guidance is important: despite using the same inputs, EgoAllo-NoReproj outputs are worse than EgoAllo-Mono in both MPJPE and PA-MPJPE. Finally, incorporating more accurate wrist pose estimates from the extra cameras on Project Aria drops errors $131.45\\mathrm{mm} \\rightarrow 60.08\\mathrm{mm}$ . This highlights both room for improvement for EgoAllo and how EgoAllo's guidance-based formulation makes it easy to extend when more sensors are available.", + "bbox": [ + 89, + 417, + 483, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitatively, we observed that high hand estimation errors in naive monocular estimation with HaMeR are explained by a combination of detection failures and monocular ambiguities. Even when detections succeed, the scale and distance of monocular HaMeR estimates are often incorrect or flicker in between frames. EgoAllo improves on this by applying kinematic constraints and a diffusion-based prior for plausible body motion—we provide examples of HaMeR estimates rendered jointly with EgoAllo outputs in Figure 5.", + "bbox": [ + 89, + 599, + 483, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Dicussion", + "text_level": 1, + "bbox": [ + 89, + 753, + 191, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. While the core contributions of EgoAllo are general, the current implementation of our system has a few limitations that we hope to explore in future work. First, diffusion model guidance is a test-time optimization process that depends on hyperparameters and incurs a runtime cost. In the future, it may be possible to bootstrap using outputs from our model to train a feedforward model that avoids this step. Success for hand guidance also still depends on reasonable", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5d0e6e5c24724b83f082f64773dd82388892094b92a20c923e29805a629d954.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 612, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6dc08720293e3599f6e0283137f1207700143f7389ba4d6db44da6d7260963e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 88, + 717, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e05d4df87330c2592576531aae940510d7869503b762d80c542d64b9bfc57259.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 88, + 810, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/067d87f430e88fcaad7a74dbafb94eb9548a93818e24efba0428cfbdbf8e9c4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 88, + 903, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff32739cb8fa7d88e1a63cd5adb7c329d98bfbbad622b0567df87878edef0758.jpg", + "image_caption": [ + "Figure 5. Body estimation improves hand estimation. We show raw outputs from HaMeR [66] in blue and hand-body estimations from EgoAllo in purple. Top: improved scene interaction during touchscreen operation with EgoAllo-Mono. We know a priori that the fingers are contacting the screen in this sequence. Bottom: qualitative examples from EgoExo [17] evaluation, showing the differences between monocular hands and EgoAllo-Wrist3D estimates." + ], + "image_footnote": [], + "bbox": [ + 516, + 175, + 645, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb2139961d402a1d8399156aa9d44ba260dd3b487a9f1c4d3a3e7338d5ad9e9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 175, + 769, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/556860c8857b3de77f3dee4eca39873535ca320e08a5bbd1ad93cdcd1e448f26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 175, + 903, + 329 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8b200b4fd84a8d3a8947d2847f6aad7ffafd0bb00b74847bf2f2be728d3f1db5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPE ↓PA-MPJPE ↓
HaMeR237.90±1.8913.04±1.89
EgoAllo-NoReproj143.20±0.4214.75±0.42
EgoAllo-Mono131.45±0.3914.71±0.39
EgoAllo-Wrist3D60.08±0.2614.38±0.26
", + "bbox": [ + 542, + 453, + 879, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Hand estimation errors in millimeters. EgoAllo's hand-body estimation can constrain and resolve ambiguities in noisy outputs from HaMeR, which we observe can reduce MPJPE for hands by over $40\\%$ . EgoAllo-Wrist3D incorporates 3D wrist poses provided from the grayscale cameras on Project Aria.", + "bbox": [ + 511, + 580, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "monocular hand estimates. Errors like left/right flipping or spurious detections can therefore cause estimation failure. Finally, we assume flat floors. This is in part because our training data [56] includes floor planes but no detailed scene geometry. As a result, our method will fail in settings like hills or staircases. In the future, we hope to extend our insights to data with more detailed scene information, which concurrent work has highlighted the usefulness of in informing human body estimation [21].", + "bbox": [ + 511, + 666, + 908, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We presented EgoAllo, a system for estimating human motion using a head-mounted device. EgoAllo jointly estimates human body pose, height, and hand parameters from only egocentric SLAM poses and images. Results highlight the importance of spatial and temporal invariance in conditioning for this problem, while demonstrating how estimated bodies can be used to improve hand estimation.", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7079", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This project was funded in part by NSF:CNS-2235013 and IARPA DOI/IBC No. 140D0423C0035. JM was supported by ONR MURI N00014-21-1-2801. YM acknowledges support from the joint Simons Foundation-NSF DMS grant #2031899, the ONR grant N00014-22-1-2102, the NSF grant #2402951, and partial support from TBSI, InnoHK, and the University of Hong Kong. BY is supported by the National Science Foundation Graduate Research Fellowship Program under Grant DGE 2146752. The authors would also like to thank Hongsuk Choi, Michael Taylor, Tyler Bonnen, Songwei Ge, Chung Min Kim, and Justin Kerr for insightful technical discussion and suggestions, as well as Jiaman Li for helpful answers to questions about EgoEgo.", + "bbox": [ + 89, + 90, + 485, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 303, + 183, + 319 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2", + "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. ACM New York, NY, USA, 2005. 2", + "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 5, 1", + "[4] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 5", + "[5] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 3", + "[6] Marcus A Brubaker, David J Fleet, and Aaron Hertzmann. Physics-based person tracking using the anthropomorphic walker. International journal of computer vision, 87(1-2): 140-155, 2010. 2", + "[7] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4221-4231, 2023, 1, 2, 3, 4, 5, 6, 7", + "[8] R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE conference on computer vision and pattern recognition (CVPR), pages 77-85. IEEE, 2017. 3", + "[9] Haiwei Chen, Shichen Liu, Weikai Chen, Hao Li, and Randall Hill. Equivariant point network for 3d point cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14514-14523, 2021. 3", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning" + ], + "bbox": [ + 99, + 329, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3", + "[11] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4800-4810, 2023. 2, 5", + "[12] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[14] Haiwen Feng, Peter Kulits, Shichen Liu, Michael J Black, and Victoria Fernandez Abrevaya. Generalizing neural human fitting to unseen poses with articulated se (3) equivariance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7977-7988, 2023. 3", + "[15] Saeed Ghorbani, Calden Wloka, Ali Etemad, Marcus A Brubaker, and Nikolaus F Troje. Probabilistic character motion synthesis using a hierarchical deep latent variable model. In Computer Graphics Forum, pages 225-239. Wiley Online Library, 2020. 2", + "[16] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In Int. Conf. Comput. Vis., 2023. 2", + "[17] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 6, 8, 3", + "[18] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. arXiv preprint arXiv:2403.18118, 2024. 1", + "[19] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In Int. Conf. Comput. Vis., pages 1381-1388, IEEE, 2009. 2", + "[20] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10884-10894, 2019. 2", + "[21] Vladimir Guzov, Yifeng Jiang, Fangzhou Hong, Gerard Pons-Moll, Richard Newcombe, C Karen Liu, Yuting Ye, and Lingni Ma. Hmd: Environment-aware motion generation from single egocentric head-mounted device. arXiv preprint arXiv:2409.13426, 2024. 2, 8, 1", + "[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5", + "[24] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7080", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Nicholas Howe, Michael Leventon, and William Freeman. Bayesian reconstruction of 3d human motion from single-camera video. Advances in neural information processing systems, 12, 1999. 2", + "[26] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13274-13285, 2022. 6, 7, 3", + "[27] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2", + "[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 2, 3, 4, 5, 6, 7", + "[29] Jiaxi Jiang, Paul Streli, Manuel Meier, Andreas Fender, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2, 3, 4, 5, 7, 1", + "[30] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6142-6152, 2024. 2, 5", + "[31] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards in-the-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2", + "[32] Angjoo Kanazawa, Abhishek Sharma, and David Jacobs. Locally scale-invariant convolutional neural networks. arXiv preprint arXiv:1412.5104, 2014. 3", + "[33] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7122-7131, 2018. 2", + "[34] Angjoo Kanazawa, Jason Y Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5614-5623, 2019. 2", + "[35] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 2, 5", + "[36] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2", + "[37] Meejin Kim and Sukwon Lee. Fusion poser: 3d human pose estimation using sparse imus and head trackers in real time. Sensors, 22(13):4846, 2022. 2", + "[38] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5253-5263, 2020. 2", + "[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021. 2", + "[41] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2", + "[42] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 2", + "[43] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023. 2", + "[44] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6050-6059, 2017. 2", + "[45] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 3", + "[46] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. arXiv preprint arXiv:2401.00847, 2024. 2, 7", + "[47] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 2, 5", + "[48] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17142-17151, 2023. 1, 2, 3, 4, 5, 6, 7", + "[49] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2", + "[50] Zongmian Li, Jiri Sedlar, Justin Carpentier, Ivan Laptev, Nicolas Mansard, and Josef Sivic. Estimating 3d motion and forces of person-object interactions from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8649, 2019. 2", + "[51] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 2", + "[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi-person linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851-866, 2023. 2, 3", + "[53] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3", + "[54] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7081", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Lingni Ma, Yuting Ye, Fangzhou Hong, Vladimir Guzov, Yifeng Jiang, Rowan Postyeni, Luis Pesqueira, Alexander Gamino, Vijay Baiyya, Hyo Jin Kim, et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. arXiv preprint arXiv:2406.09905, 2024. 2", + "[56] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5442-5451, 2019. 3, 4, 6, 7, 8", + "[57] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Ecoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1", + "[58] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Int. Conf. Comput. Vis., pages 2640-2649, 2017. 2", + "[59] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2", + "[60] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In 2018 international conference on 3D vision (3DV), pages 484-494. IEEE, 2018. 2", + "[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 1, 6, 7, 3", + "[62] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 459-468, 2018. 2", + "[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10975-10985, 2019. 2", + "[64] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 2", + "[65] Georgios Pavlakos, Ethan Weber, Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In Eur. Conf. Comput. Vis., pages 732-749. Springer, 2022. 2", + "[66] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In *arxiv*, 2023. 3, 8", + "[67] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. 2", + "[68] Chiara Plizzari, Shubham Goel, Toby Perrett, Jacob Chalk, Angjoo Kanazawa, and Dima Damen. Spatial cognition from" + ], + "bbox": [ + 99, + 92, + 485, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "egocentric video: Out of sight, not out of mind. arXiv preprint arXiv:2404.05072, 2024. 1", + "[69] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2", + "[70] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[71] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2740-2749, 2022. 2, 7", + "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 5", + "[73] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 2", + "[74] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 4, 5, 7, 1", + "[75] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 2", + "[76] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2", + "[77] Gregory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net: Localization-classification-regression for human pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3433-3441, 2017. 2", + "[78] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[79] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2, 3, 5", + "[80] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7082", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Silvio Savarese and Li Fei-Fei. 3d generic object categorization, localization and pose estimation. In 2007 IEEE 11th international conference on computer vision, pages 1-8. IEEE, 2007. 3", + "[82] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 5", + "[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K Hodgins. Motion capture from body-mounted cameras. In ACM SIGGRAPH 2011 papers, pages 1-10. ACM New York, NY, USA, 2011. 2", + "[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[85] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2, 3, 4, 5, 6, 8", + "[86] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 3", + "[87] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[88] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2", + "[89] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2", + "[90] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3", + "[91] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7728-7738, 2019. 2, 7", + "[92] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2", + "[93] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2", + "[94] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 2", + "[95] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11500-11509, 2021. 2, 7" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[96] Jian Wang, Zhe Cao, Diogo Luvizon, Lingjie Liu, Kripasindhu Sarkar, Danhang Tang, Thabo Beeler, and Christian Theobalt. Egocentric whole-body motion capture with fisheyevit and diffusion-based motion refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 777-787, 2024. 2, 7", + "[97] Laurenz Wiskott and Terrence J Sejnowski. Slow feature analysis: Unsupervised learning of invariances. Neural computation, 14(4):715-770, 2002. 3", + "[98] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 2", + "[99] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246, 2022.8", + "[100] Jingyun Yang, Congyue Deng, Jimmy Wu, Rika Antonova, Leonidas Guibas, and Jeannette Bohg. Equivact: Sim(3)-equivariant visuomotor policies beyond rigid object manipulation, 2023. 3", + "[101] Jingyun Yang, Zi-ang Cao, Congyue Deng, Rika Antonova, Shuran Song, and Jeannette Bohg. Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning. arXiv preprint arXiv:2407.01479, 2024. 3", + "[102] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21222-21232, 2023. 2, 5, 7", + "[103] Brent Yi, Weijia Zeng, Sam Buchanan, and Yi Ma. Canonical factors for hybrid neural fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3414-3426, 2023. 3", + "[104] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2", + "[105] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. arXiv preprint arXiv:2305.01599, 2023. 2", + "[106] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2", + "[107] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2", + "[108] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[109] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems, 30, 2017. 3" + ], + "bbox": [ + 517, + 92, + 906, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7083", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[110] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. arXiv preprint arXiv:2305.03462, 2023. 3", + "[111] Daiwei Zhang, Gengyan Li, Jiajie Li, Mickaël Bressieux, Otmar Hilliges, Marc Pollefeys, Luc Van Gool, and Xi Wang. Egogaussian: Dynamic scene understanding from egocentric video with 3d gaussian splatting. arXiv preprint arXiv:2406.19811, 2024. 1", + "[112] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2", + "[113] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 5", + "[114] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 2, 5", + "[115] Zhengdong Zhang, Arvind Ganesh, Xiao Liang, and Yi Ma. Tilt: Transform invariant low-rank textures. International journal of computer vision, 99:1-24, 2012. 3" + ], + "bbox": [ + 89, + 92, + 483, + 430 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "7084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_model.json b/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94b92cb33e2e0985a9b702065bbcb8f4990cd2f3 --- /dev/null +++ b/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_model.json @@ -0,0 +1,3174 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.213, + 0.131, + 0.787, + 0.154 + ], + "angle": 0, + "content": "Estimating Body and Hand Motion in an Ego-sensed World" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.171, + 0.768, + 0.244 + ], + "angle": 0, + "content": "Brent Yi\\(^{1}\\) Vickie Ye\\(^{1}\\) Maya Zheng\\(^{1}\\) Yunqi Li\\(^{2}\\) Lea Müller\\(^{1}\\) Georgios Pavlakos\\(^{3}\\) Yi Ma\\(^{1}\\) Jitendra Malik\\(^{1}\\) Angjoo Kanazawa\\(^{1}\\) UC Berkeley \\({}^{2}\\)ShanghaiTech \\({}^{3}\\)UT Austin" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.266, + 0.618, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.267, + 0.907, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.51, + 0.908, + 0.539 + ], + "angle": 0, + "content": "Figure 1. EgoAllo. We present a system that estimates human body pose, height, and hand parameters from egocentric SLAM poses and images. Outputs capture the wearer's actions in the allocentric reference frame of the scene, which we visualize here with 3D reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.563, + 0.326, + 0.579 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.598, + 0.486, + 0.795 + ], + "angle": 0, + "content": "We present EgoAllo, a system for human motion estimation from a head-mounted device. Using only egocentric SLAM poses and images, EgoAllo guides sampling from a conditional diffusion model to estimate 3D body pose, height, and hand parameters that capture a device wearer's actions in the allocentric coordinate frame of the scene. To achieve this, our key insight is in representation: we propose spatial and temporal invariance criteria for improving model performance, from which we derive a head motion conditioning parameterization that improves estimation by up to \\(18\\%\\). We also show how the bodies estimated by our system can improve hand estimation: the resulting kinematic and temporal constraints can reduce world-frame errors in single-frame estimates by \\(40\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.216, + 0.844 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Head-mounted devices are becoming increasingly mainstream. In addition to offering new challenges for 3D scene understanding [18, 57, 68, 111], egocentric sensors from these devices are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.626 + ], + "angle": 0, + "content": "unique in that their outputs are coupled to a human wearer's motion in the world. Using these sensors to understand the wearer in addition to the scene around them is essential for applications in augmented reality, robotics, and assistive technologies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.717 + ], + "angle": 0, + "content": "We therefore introduce EgoAllo, a system that uses egocentric inputs to estimate the wearer and their motion in the world, or allocentric, coordinate frame. We take as input sensed metric SLAM head poses and egocentric video from devices like Project Aria [61]. We then output estimates of human body pose, height, and hand motion parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.909, + 0.839 + ], + "angle": 0, + "content": "This is a difficult task: while body parts like hands occasionally appear in egocentric frames, most body parameters are never directly observed. Estimated body pose, body height, and hand parameters must be consistent with both the sensed egomotion and visual observations. This setting differs from most prior works in egocentric human motion estimation [7, 48], which focus on body pose and do not address the challenges of height and hand motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our proposed system uses a head pose-conditioned diffusion model as a motion prior, as well as a Levenberg-Marquardt guidance optimizer for sampling hand-body sequence that align with image observations. Our results are enabled by a key" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "insight: that the representation used for head pose conditioning is critical for accurate full-body motion estimation. We study choices for this representation by (1) identifying desirable spatial and temporal invariance properties that are not fulfilled by existing systems and (2) using these properties to derive improved parameterizations for our motion prior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.288 + ], + "angle": 0, + "content": "We systematically evaluate our system on four datasets. For body estimation, we find that improving the conditioning parameterization leads to an accuracy improvement between \\(4.9\\%\\) and \\(17.9\\%\\). Furthermore, we observe that the resulting system can improve hand estimation, reducing world-frame errors by over \\(40\\%\\) compared to single-frame estimates. Code, model, and more results can be found on our project webpage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.303, + 0.226, + 0.318 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.483, + 0.584 + ], + "angle": 0, + "content": "3D human recovery from external visual inputs. A large body of work has addressed estimating the parameters of human body models like SCAPE [2] or SMPL and its variants [52, 63, 79] from third-person visual inputs, where human subjects are observed from the view of outside cameras. The majority of these works focus on extracting 3D representations from single images, for example by lifting 2D keypoint observations to 3D [58], via end-to-end regression [20, 31, 33, 40, 60, 62, 77], via optimization [19, 44, 63], or by exploiting synergies between regression and optimization [42]. When multiple frames are available in the form of a video, temporal context and tracking can also be incorporated [16, 34, 39, 64, 65, 71, 108]. The inputs (images) and outputs (human meshes) of many of these systems are similar to the egocentric setting addressed by EgoAllo, but egocentric devices present unique challenges because the body being estimated is typically behind the outwards-facing cameras used as input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.849 + ], + "angle": 0, + "content": "Priors for human motion. The primary challenge of ego-sensed human motion estimation is limited observability; a prior is required to resolve ambiguities. For human motion, these priors are typically framed as unconditional distributions over plausible human motion. Distributions can be represented either by modeling the physical constraints of our world [6, 50, 67, 73] or by learning generative models of human motion directly from data. For learning unconditional priors, classical data-driven approaches include fitting mixtures-of-Gaussians to 3D keypoint trajectories [25], while modern approaches include training variational autoencoders [38, 75] to model either autoregressive transitions [15, 51, 74] or full spatiotemporal sequences [22]. After training, these priors can be applied to estimation problems in iterative optimization frameworks [41, 74, 102]. EgoAllo is built on the same intuition as these methods, but follows previous work in ego-sensed motion estimation and uses a task-specific conditional prior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Denoising diffusion for human motion. The core of EgoAllo is a denoising diffusion model [23, 69, 84] from which we can sample 3D human body motion. While diffusion mod" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.258 + ], + "angle": 0, + "content": "els are primarily known for their success in text-conditioned image generation [78, 80], they have also enabled advances in human motion synthesis conditioned on modalities like text [35, 36, 112], music [1, 93], poses [35, 48], and object geometry [43, 47, 49]. EgoAllo adopts a similar conditional diffusion approach, while specifically studying the design of conditioning parameters used for ego-sensed human motion estimation. The iterative nature of denoising diffusion also enables guidance [11, 13, 30, 35, 87, 114], where denoising steps are steered to satisfy a desired objective. We use guidance to incorporate observations like visual hand pose observations during test-time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.909, + 0.618 + ], + "angle": 0, + "content": "Human motion from egocentric observations. EgoAllo builds on intuition from several prior works in egocentric sensing for human motion estimation. Many rely on fisheye cameras that place the wearer's body into the field of view [27, 76, 91, 91, 92, 95, 96, 98]. Other approaches rely on body-mounted cameras [83], simulation-based physical plausibility [54, 106, 107], body- and hand-mounted inertial sensors [37, 46, 104, 105], handheld controllers [7, 28, 29], and interaction cues from other humans [59]. Concurrent works have also used the Nymeria [55] dataset for egocentric motion with language description outputs [24], as well for online settings with scene geometry and CLIP [70] feature inputs [21]. Most relevantly, EgoEgo [48] demonstrates how human body poses can be estimated offline without body observability assumptions. The authors accomplish this by carefully integrating several components: a monocular SLAM system [89], a pose-conditioned gravity vector regression network, an optical flow feature-conditioned head orientation and scale regression network, and a head pose-conditioned body diffusion model. EgoAllo differs in both inputs—we study conditioning parameters computed from the metric SLAM poses provided by devices like Project Aria [85]—and outputs—we consider body height variation and hand poses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Conditioning for ego-sensed poses. Prior works vary in how head pose information is parameterized and used as neural network input. AvATARPoser [28] and BoDiffusion [7] parameterize head pose as four components: world-frame orientation, orientation deltas, world-frame position, and world-frame position deltas. These works focus on settings with VR controller input, and parameterize controller pose inputs the same way. EgoEgo [48]'s diffusion model uses only absolute head positions and orientations, but similar to HuMoR [74], in implementation defines a per-sequence canonical coordinate frame to ensure that all input trajectories passed to the model are aligned with the same initial \\( xy \\) position and forward direction. In our work, we refer to this as sequence canonicalization. Finally, EgoPoser [29] proposes a similar scheme that aligns initial positions for both head pose and controller pose inputs. We propose an alternative to these parameterizations that is motivated by the robustness and generalization benefits of invariance, as observed in prior work for designing both" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.898, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.293, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Overview of components in EgoAllo. We restrict the diffusion model to local body parameters (Section 3.1.1). An invariant parameterization \\( g(\\cdot) \\) (Section 3.1.2) of SLAM poses is used to condition a diffusion model. These can be placed into the global coordinate frame via global alignment (Section 3.2.1) to input poses. When available, egocentric video is used for hand detection via HaMeR [66], which can be incorporated into samples via guidance (Section 3.2.2)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.376, + 0.483, + 0.437 + ], + "angle": 0, + "content": "representations [10, 53, 81, 97, 103, 110, 115] and neural network architectures [8, 9, 12, 14, 32, 45, 90, 100, 101, 109]. Specifically, we introduce in Section 3.1.2 a parameterization that is invariant to both spatial and temporal shifts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.449, + 0.179, + 0.465 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.534 + ], + "angle": 0, + "content": "We study the problem of using sensors from an egocentric device to estimate the actions of the wearer in an allocentric coordinate frame. We assume a flat floor and two inputs: poses from the device's SLAM system and egocentric video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Our system uses head pose information to condition a diffusion-based prior over body pose and height, and incorporates visual hand observations during sampling. This allows it to benefit from both 3D human motion capture datasets [56], which are used for the motion prior, and from large-scale image datasets [66], which are used for hand estimates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.383, + 0.65 + ], + "angle": 0, + "content": "3.1. Ego-conditioned motion diffusion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.657, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Notation: we use \\(\\mathbf{T}_{\\mathrm{A,B}} = (\\mathbf{R}_{\\mathrm{A,B}},\\mathbf{p}_{\\mathrm{A,B}})\\) to denote an SE(3) transform to frame A from frame B, composed of rotation \\((\\mathbf{R}_{\\mathrm{A,B}})\\) and position \\((\\mathbf{p}_{\\mathrm{A,B}})\\) terms. Temporal steps \\(t\\) are superscripted and diffusion noise steps \\(n\\) are subscripted. \\(\\vec{x}_0^t\\) thus refers to the \\(t\\)-th timestep of a clean \\((n = 0)\\) human motion sequence." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.733, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Given an observation window of \\(T\\) timesteps, EgoAllo's motion prior is a diffusion model that aims to capture the distribution of human motions \\(\\vec{x}_0 = \\{\\vec{x}_0^1,\\dots,\\vec{x}_0^T\\}\\) conditioned on head pose encodings \\(\\vec{c} = \\{\\vec{c}^1,\\dots,\\vec{c}^T\\}\\). For each timestep \\(t\\), we represent human motion in the form of SMPL-H [52, 79] model parameters \\(\\{\\mathbf{T}_{\\mathrm{world,root}}^t,\\Theta^t,\\beta \\}\\): root transforms \\(\\mathbf{T}_{\\mathrm{world,root}}^t\\in \\mathrm{SE}(3)\\), where the person's root frame is located at their pelvis, local joint rotation matrices \\(\\Theta^t\\in \\mathbb{R}^{51\\times 3\\times 3}\\), and time-invariant shape \\(\\beta \\in \\mathbb{R}^{16}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Dependencies between local joint rotations, body size variation, and global motion make this learning task a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.452 + ], + "angle": 0, + "content": "challenging one. Our key insight is that this difficulty can be reduced by designing parameterizations with desirable invariance properties. Spatial and temporal invariances allow the model to focus on the essential structure of motion, without being affected by irrelevant shifts in position or time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.78, + 0.483 + ], + "angle": 0, + "content": "3.1.1. Diffusion output representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.579 + ], + "angle": 0, + "content": "As output, we sample body and hand joint rotations, body shapes, and binary contact predictions \\(\\vec{x}_0^t = \\{\\Theta^t,\\beta^t,\\psi_{j = 1\\dots 21}^t\\}\\), where body shape \\(\\beta^t\\) is supervised to be equal for all timesteps and \\(\\psi_j^t\\) is a per-joint contact indicator. Notably, these parameters are all local—we discuss how outputs can be placed into the allocentric coordinate frame in Section 3.2.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.749 + ], + "angle": 0, + "content": "We choose this output set for three main reasons. (1) Body shape encodes the wearer's height, which is critical for grounding in the metric-scale geometry of the scene. This is rarely considered by prior work: with the exception of [29], existing methods [7, 28, 48] otherwise produce outputs using a fixed \"mean\" human shape. (2) Contact predictions enable losses for common problems like foot skating, which are discussed in Section 3.2.2. (3) Finally, local bodies are invariant to the global coordinate frame. As we discuss next, the conditioning parameterization for the model can therefore also be invariant to arbitrary transformations along the floor plane." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.719, + 0.78 + ], + "angle": 0, + "content": "3.1.2. Invariant conditioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.906, + 0.832 + ], + "angle": 0, + "content": "The goal of our conditioning representation is to map raw SLAM poses (head motion) to a parameterization that is amenable to learning for the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Raw inputs. To capture the head motion at each time step, we assume as input poses of a central pupil frame (CPF), which the SLAM systems of devices like Project Aria can provide with millimeter-level accuracy [85]. For time 1...T, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.108 + ], + "angle": 0, + "content": "reparameterize these poses for conditioning using a function \\( g \\):" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.117, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} = \\left(\\mathbf {R} _ {\\text {w o r l d , c p f}} ^ {t}, \\mathbf {p} _ {\\text {w o r l d , c p f}} ^ {t}\\right) \\in \\mathrm {S E} (3), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.138, + 0.484, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\{\\vec {c} ^ {1}, \\dots , \\vec {c} ^ {T} \\} = g (\\{\\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {1}, \\dots , \\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {T} \\}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.483, + 0.241 + ], + "angle": 0, + "content": "The CPF frame differs from prior works that condition on a coordinate frame attached to the SMPL human model's \"head joint\" [7, 28, 29, 48]. The offset between this head joint and the device pose depends on the head shape captured by \\(\\beta^t\\), and is thus difficult to precompute in our setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.484, + 0.318 + ], + "angle": 0, + "content": "To encode absolute height, we assume that the world frame's \\(+z\\)-axis faces upwards, and that the ground is located at \\(z = 0\\). Ground parameters are directly available in the training data [56]; at test time, we can also extract these parameters from sparse SLAM points via RANSAC (Appendix A.3.5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.43 + ], + "angle": 0, + "content": "Invariance goals. As discussed in Section 2, prior work varies in how the function \\( g \\) is implemented. To understand how choices impact learning, we propose two invariance properties for head motion representations. Each reduces representational redundancy, which eases the learning problem. We provide more detailed explanations and visualizations of these properties in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.484, + 0.503 + ], + "angle": 0, + "content": "Invariance 1 (Spatial) Global transformations along the floor plane should not affect a person's local motion. Given \\(\\mathbf{T}_{xy} \\in SE(3)\\) restricted to the \\(XY\\) plane, \\(g\\) should fulfill \\(g(\\{\\mathbf{T}_{xy}\\mathbf{T}_{world, cpf}^t\\}_t) = g(\\{\\mathbf{T}_{world, cpf}^t\\}_t) \\forall \\mathbf{T}_{xy}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.485, + 0.606 + ], + "angle": 0, + "content": "Invariance 2 (Temporal) If the same human motion occurs at different points in time, the extracted head motion representation should remain consistent. This can be expressed as temporal shift equivariance. Let \\(\\vec{c}^t\\) be as defined in Equation 2. For any shift \\(\\delta\\) such that \\(\\{\\vec{c}_{shift}^{1},\\dots,\\vec{c}_{shift}^{T}\\} = g(\\{\\mathbf{T}_{world,cpf}^{1 + \\delta},\\dots,\\mathbf{T}_{world,cpf}^{T + \\delta}\\})\\), \\(g\\) should satisfy \\(\\vec{c}_{shift}^{t} = \\vec{c}^{t + \\delta}\\) for overlapping timesteps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.736 + ], + "angle": 0, + "content": "No parameterization used by existing work satisfies both of these properties. The sequence canonicalization approach of EgoEgo [48] achieves spatial invariance (Invariance 1), but inserts a sequence-wide dependency on the first timestep of each window that results in a violation of Invariance 2. The absolute poses and pose deltas used by [7, 28] satisfy Invariance 2, but not Invariance 1. Finally, the relative positions considered by [29] are neither spatially nor temporally invariant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.817 + ], + "angle": 0, + "content": "Invariant conditioning. We propose a formulation for \\( g \\) that achieves both invariance properties by locally canonicalizing head motion with respect to the floor at each timestep. We build on the relative motion of the CPF frame at each time \\( t \\), which respects both Invariance 1 and 2:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.827, + 0.484, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {T} _ {\\mathrm {c p f}} ^ {t - 1, t} = \\left(\\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {t - 1}\\right) ^ {- 1} \\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Importantly, the translation component of this transformation is in the local frame. This is distinct from world-frame position deltas [7, 28, 29], which still violate Invariance 1." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.076, + 0.858, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Figure 3. Locally canonicalized coordinate frames. We compute our invariant conditioning parameterization (Equation 4) using transformations computed from three coordinate frames. Following [85], the CPF has the \\(z\\)-axis forward. Following HuMoR [74], the world and canonical \\(z\\)-axes point up. Canonical frames are computed by projecting the CPF frame origin to the ground plane, then aligning the canonical \\(y\\)-axis to the CPF forward direction. The axes are color-coded as \\(x\\) (red), \\(y\\) (green), and \\(z\\) (blue)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Relative transforms alone do not encode information relative to the scene or floor: full trajectories can even be flipped upside down without impacting \\(\\Delta \\mathbf{T}_{\\mathrm{cpf}}^{t - 1,t}\\). We therefore propose to ground relative motion to the floor plane with a transformation between the CPF frame and a per-timestep canonical frame, which is computed by projecting the CPF frame to the floor. This encodes head height and orientation. Our full representation then becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.562, + 0.907, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\vec {c} ^ {t} = \\underbrace {\\left\\{\\Delta \\mathbf {T} _ {\\mathrm {c p f}} ^ {t - 1 , t} , \\quad \\left(\\mathbf {T} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t}\\right) ^ {- 1} \\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} \\right\\}} _ {\\text {I n v a r i a n t i m p l e m e n t a t i o n o f} g (\\cdot)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.906, + 0.681 + ], + "angle": 0, + "content": "We visualize an example of a canonical frame in Figure 3 and our full representation in Appendix A.1. Canonical frames are positioned by projecting the CPF origin to the floor plane; given standard bases \\(\\mathbf{e}_{\\{x,y,z\\}}\\), we compute:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.69, + 0.907, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t} = \\left[ \\begin{array}{l l l} \\mathbf {e} _ {x} & \\mathbf {e} _ {y} & \\vec {0} \\end{array} \\right] ^ {\\top} \\mathbf {p} _ {\\text {w o r l d , c p f}} ^ {t}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.722, + 0.907, + 0.797 + ], + "angle": 0, + "content": "For orientation, we align the canonical frame's local \\(z\\)-axis parallel to the world \\(z\\)-axis and its local \\(y\\)-axis toward the \"forward\" direction \\(\\vec{v}^t\\) of the CPF frame. With \\(\\mathbf{R}_{z}(\\cdot):\\mathbb{R}\\to \\mathrm{SO}(3)\\) constructing a \\(z\\)-axis rotation and \\(\\mathbf{e}_{\\{x,y,z\\}}\\) again as standard bases, we compute this as:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.807, + 0.907, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\vec {v} ^ {t} = \\mathbf {R} _ {\\text {w o r l d , c p f}} ^ {t} \\mathbf {e} _ {z}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.828, + 0.907, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t} = \\mathbf {R} _ {z} \\left(- \\arctan 2 \\left(\\mathbf {e} _ {x} ^ {\\top} \\vec {v} ^ {t}, \\mathbf {e} _ {y} ^ {\\top} \\vec {v} ^ {t}\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This canonical frame definition is an important departure from prior work. While EgoEgo [48] and HuMoR [74] use similar canonical frames, they only compute one per sequence. Instead," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "we compute Equations 5 and 7 at every timestep. This enables floor plane grounding without sacrificing Invariance 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.304 + ], + "angle": 0, + "content": "Connection to prior work. EgoAllo's invariant conditioning shares similarities with the spatial and temporal normalization proposed by EgoPoser [29]. (1) EgoPoser's temporal normalization subtracts positions by the first timestep's position. This is a position-only version EgoEgo [48]'s sequence canonicalization. It does not consider rotation, and therefore only partially fulfills Invariance 1. Like EgoEgo, it also does not fulfill Invariance 2. In Appendix A.2.3, we observe that this approach only has a small impact when adapted to EgoAllo's problem setting. (2) EgoPoser's spatial normalization subtracts the head's XY position from the hand pose inputs. EgoAllo does not use hand pose as network input, so this is not applicable in our setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.313, + 0.315, + 0.33 + ], + "angle": 0, + "content": "3.2. Estimation via sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.396 + ], + "angle": 0, + "content": "We use our local body representation and invariant conditioning strategies to train a motion prior in the form of a denoising diffusion model [23]. Given diffusion step \\( n = N \\ldots 1 \\), we follow [72] and approximate the denoising process as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.407, + 0.484, + 0.427 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\vec {x} _ {n - 1} | \\vec {x} _ {n}, \\vec {c}) = \\mathcal {N} \\big (\\mu_ {\\theta} (\\vec {x} _ {n}, n, \\vec {c}), \\sigma_ {n} ^ {2} \\mathbf {I} \\big), \\qquad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.482 + ], + "angle": 0, + "content": "where a transformer [94] \\(\\mu_{\\theta}\\) is trained to predict the posterior mean from noised sample \\(\\vec{x}_n\\) and conditioning \\(\\vec{c}\\). With noise-dependent weight term \\(w_{n}\\), the loss can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.492, + 0.484, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\vec {x} _ {0}} \\mathbb {E} _ {n \\sim \\mathcal {U}} \\left[ w _ {n} \\| \\mu_ {\\theta} (\\vec {x} _ {n}, n, \\vec {c}) - \\vec {x} _ {0} \\| ^ {2} \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.617 + ], + "angle": 0, + "content": "After training, we estimate human motions by following DDIM [86] for sampling. The final EgoAllo sampling procedure includes several additional components: a global alignment phase, guidance losses for physical constraints and visual hand observations, and a path fusion [3] approach for longer sequence lengths. We describe these below." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.258, + 0.64 + ], + "angle": 0, + "content": "3.2.1. Global alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.674 + ], + "angle": 0, + "content": "To place sampled bodies into the allocentric coordinate system, we compute the absolute pose of the SMPL-H root as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.685, + 0.484, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {\\text {w o r l d , r o o t}} ^ {t} = \\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} \\mathbf {T} _ {\\text {c p f , r o o t}} ^ {(\\Theta^ {t}, \\beta^ {t})}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.814 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_{\\mathrm{cpf,root}}^{(\\Theta^t,\\beta^t)}\\) computes the transform between the root of the human and their CPF frame for a given set of local pose and shape parameters. Similar processes are applied in [7, 28, 29]. In contrast to directly outputting absolute body transformations from the diffusion model [48], this guarantees exact alignment between estimates and input SLAM sequences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.248, + 0.835 + ], + "angle": 0, + "content": "3.2.2. Guidance losses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our diffusion model learns a distribution of human motion conditioned on the central pupil frame motion. At test time, we incorporate constraints from physical priors and visual hand observations via guidance [11, 30, 114]. Similar to [35, 47]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "we accomplish this by applying costs to the joint rotations \\(\\Theta = \\{\\Theta^1,\\dots,\\Theta^T\\}\\) predicted by \\(\\mu_{\\theta}(\\vec{x}_n,n,\\vec{c})\\). We treat the body shape \\(\\beta^{t}\\) and contacts \\(\\psi_{j = 1\\dots 21}^{t}\\) as fixed and optimize over body and finger pose to minimize hand observation, skating, and prior costs with a Levenberg-Marquardt optimizer:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.173, + 0.907, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\text {g u i d a n c e}} ^ {(\\Theta)} = \\mathcal {E} _ {\\text {h a n d s}} ^ {(\\Theta)} + \\mathcal {E} _ {\\text {s k a t e}} ^ {(\\Theta)} + \\mathcal {E} _ {\\text {p r i o r}} ^ {(\\Theta)}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.906, + 0.306 + ], + "angle": 0, + "content": "We begin by running HaMeR on the egocentric image corresponding to each timestep \\( t \\). When detected, this produces 3D hand estimates in the form of MANO [79] joint parameters and camera-centric 3D hand keypoints \\( \\hat{\\mathbf{p}}_{\\mathrm{camera},j}^{t} \\) for hand joint set \\( j \\in \\mathcal{H} \\). Optionally, wrist and palm poses can also be estimated using Project Aria's Machine Perception Services [85]. With each subscripted \\( \\lambda \\) indicating a scalar weighting term, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.312, + 0.907, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\text {h a n d s}} ^ {(\\Theta)} = \\lambda_ {\\text {h a n d s 3 D}} \\mathcal {E} _ {\\text {h a n d s 3 D}} ^ {(\\Theta)} + \\lambda_ {\\text {r e p r o j}} \\mathcal {E} _ {\\text {r e p r o j}} ^ {(\\Theta)}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.438 + ], + "angle": 0, + "content": "The 3D objective \\(\\mathcal{E}_{\\mathrm{hands3D}}^{(\\Theta)}\\) minimizes the distance between the detected hand parameters and the corresponding SMPL-H hand parameters, in terms of wrist pose and local joint rotations. With \\(\\Pi_K\\) as projection with camera intrinsics \\(K\\), \\(\\mathbf{p}_{\\mathrm{world},j}^{(\\Theta^t)} \\in \\mathbb{R}^3\\) as the world position for joint \\(j\\) at time \\(t\\), and \\(\\mathbf{T}_{\\mathrm{camera,cpf}}\\) from the device calibration, the reprojection cost is:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.445, + 0.907, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\text {r e p r o j}} ^ {(\\Theta)} = \\sum_ {t, j \\in \\mathcal {H}} | | \\Pi_ {K} \\left(\\mathbf {p} _ {\\text {c a m e r a}, j} ^ {(\\Theta^ {t})}\\right) - \\Pi_ {K} \\left(\\hat {\\mathbf {p}} _ {\\text {c a m e r a}, j} ^ {t}\\right) | | _ {2} ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.482, + 0.907, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {\\text {c a m e r a}, j} ^ {(\\Theta^ {t})} = \\mathbf {T} _ {\\text {c a m e r a , c p f}} \\left(\\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t}\\right) ^ {- 1} \\mathbf {p} _ {\\text {w o r l d}, j} ^ {(\\Theta^ {t})}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.541 + ], + "angle": 0, + "content": "To reduce foot skating, we use contact predictions to apply a skating cost [74, 102] for each time \\( t \\) and joint \\( j \\):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.547, + 0.906, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\text {s k a t e}} ^ {(\\Theta)} = \\sum_ {t, j} \\lambda_ {\\text {s k a t e}} \\| \\frac {1}{2} \\left(\\psi_ {j} ^ {t} + \\psi_ {j} ^ {t - 1}\\right) \\left(\\mathbf {p} _ {\\text {w o r l d}, j} ^ {t} - \\mathbf {p} _ {\\text {w o r l d}, j} ^ {t - 1}\\right) \\| _ {2} ^ {2}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Finally, we minimize a prior cost \\(\\mathcal{E}_{\\mathrm{prior}}^{(\\Theta)}\\). This cost penalizes deviations between joint rotations \\(\\Theta^t\\) and rotations \\(\\hat{\\Theta}^t\\) from the denoiser \\(\\mu_{\\theta}(\\vec{x}_n,n,\\vec{c})\\), while also encouraging smoothness. We detail the full list of terms used for this in Appendix A.3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.771, + 0.679 + ], + "angle": 0, + "content": "3.2.3. Sequence length extrapolation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.819 + ], + "angle": 0, + "content": "For longer sequences at test time, we draw on existing methods in compositional generation for both image [3, 113] and human motion [4, 82] diffusion models. We train our motion prior using subsequences of up to length 128; when input observations exceed this length at test time, we split into windows with a 32-timestep overlap between neighbors. We then run our model \\(\\mu_{\\theta}(\\vec{x}_n,\\vec{c},n)\\) on windows in parallel. Diffusion paths for overlapping regions are fused following MultiDiffusion [3] after each denoising step." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.64, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct a series of experiments to evaluate EgoAllo's conditioning parameterization, body estimation accuracy, and hand estimation performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7076" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.09, + 0.905, + 0.272 + ], + "angle": 0, + "content": "
ConditioningSeqlenInvariance 1/2MPJPE ↓% DiffPA-MPJPE ↓% DiffGND ↑
EgoAllo (Eq. 4)32✓ /✓129.8±1.1—109.8±1.1—0.98±0.00
Absolute+Local Relative32P /✓133.0±1.12.4%113.6±1.23.4%0.95±0.00
Absolute+Global Deltas [7,28]32X /✓136.2±1.14.9%118.3±1.27.7%0.93±0.01
Sequence Canonicalization [48]32✓ /X153.1±1.517.9%128.7±1.517.1%0.76±0.01
Absolute32X /✓159.9±1.223.2%141.0±1.328.4%0.89±0.01
EgoAllo (Eq. 4)128✓ /✓119.7±1.3—101.1±1.3—1.00±0.00
Absolute+Local Relative128P /✓124.5±1.34.0%104.9±1.43.8%1.00±0.00
Absolute+Global Deltas [7,28]128X /✓127.4±1.36.4%109.8±1.48.6%0.99±0.00
Sequence Canonicalization [48]128✓ /X134.0±1.811.9%112.1±1.610.9%0.88±0.02
Absolute128X /✓148.3±1.523.9%131.2±1.629.8%0.96±0.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.281, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Table 1. Motion prior conditioning comparison. We train and evaluate otherwise identical models using four conditioning parameterizations on AMASS [56] test set sequences, using sequences of length 32 and 128. Parameterizations vary in their spatial (1) and temporal (2) invariance properties, which we loosely classify as following completely (\\(\\checkmark\\)), partially (\\(\\mathsf{P}\\)), or not at all (\\(\\times\\)). The conditioning parameterization used by EgoAllo reduces errors by almost \\(18\\%\\) compared to the sequence canonicalization approach used by the most relevant related work [48]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.357, + 0.483, + 0.478 + ], + "angle": 0, + "content": "Training. To train EgoAllo models used in our experiments, we need sequences containing human body and hand pose parameters, body shapes, and device SLAM poses \\(\\mathbf{T}_{\\mathrm{world,cpf}}^t\\). Similar to prior work [7, 28, 48], we train EgoAllo using AMASS [56] with synthesized device poses. We annotate train split sequences by anchoring a central pupil frame between vertices corresponding to the left and right pupils in the blend skinned mesh, and at train time sample sequences between length 32 and 128." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.483, + 0.609 + ], + "angle": 0, + "content": "Evaluation. We evaluate with four datasets. We use AMASS [56], RICH [26], and Aria Digital Twins (ADT) [61] for body estimation evaluation, and EgoExo4D [17] for hand estimation evaluation. AMASS and RICH do not include egocentric data; we annotate these with synthetic device poses using the same procedure we use for training. ADT and EgoExo4D both include egocentric images and SLAM poses captured using Project Aria glasses [85], which we use directly." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.617, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Metrics. To quantify performance, we report four metrics: (1) MPJPE is a world-frame mean per-joint position error (millimeters). (2) PA-MPJPE is the Procrustes-aligned mean per-joint position error in millimeters, where joint positions are aligned on a per-timestep basis before error are computed. (3) GND is a grounding metric, designed in response to a phenomena where ego-sensed humans \"float\" above the ground. Given a human body trajectory, this metric contains a simple binary indicator of whether the feet of the human ever touch the ground plane. (4) \\(\\mathbf{T}_{\\mathrm{head}}\\) is the average SMPL head joint position error in millimeters." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.801, + 0.253, + 0.817 + ], + "angle": 0, + "content": "4.1. Body estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.9 + ], + "angle": 0, + "content": "In our first set of experiments, we evaluate body estimation from only device SLAM poses, without considering images or hands. This setting allows us to isolate the advantages of our body motion prior, while directly comparing against methods that do not consider hands." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.357, + 0.794, + 0.373 + ], + "angle": 0, + "content": "4.1.1. Invariant conditioning evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.908, + 0.557 + ], + "angle": 0, + "content": "We begin by evaluating the importance of the spatial and temporal invariance criteria discussed in Section 3.1.2. We do this by comparing five implementations of the conditioning \\( g \\): (1) EgoAllo is the final invariant representation that we propose in Equation 4. (2) Absolute+Local Relative appends absolute poses with the relative pose deltas written in Equation 3. (3) Absolute+Global Deltas appends absolute poses with relative orientation and the world-frame position deltas used by [7, 28]. (4) Sequence Canonicalization uses the alignment approach implemented by [48], which violates temporal invariance. (5) Absolute naively conditions on absolute poses, which violate spatial invariance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.618 + ], + "angle": 0, + "content": "We train conditional diffusion models with otherwise identical architecture using each parameterization, and then evaluate on the AMASS [56] test set. Metrics and percent differences compared to EgoAllo are reported in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.909, + 0.755 + ], + "angle": 0, + "content": "Overall, we find that the choice of conditioning parameterization makes a dramatic impact on estimation accuracy. We observe accuracy improve consistently as invariance properties are incorporated into the representation. Compared to EgoAllo, Absolute conditioning increases MPJPE by over \\(23\\%\\) for both shorter (length 32) and longer (length 128) sequences. Compared to EgoAllo, SeqCanonical conditioning increases MPJPE by nearly \\(18\\%\\) for length 32 sequences and \\(12\\%\\) for length 128 sequences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.774, + 0.777 + ], + "angle": 0, + "content": "4.1.2. Comparisons against baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To further study EgoAllo's body estimation quality, we compare against three baselines. (1) NoShape. First, NoShape refers to a variation of EgoAllo that turns off shape estimation, and thus cannot estimate the wearer's height. (2) EgoEgo. We also compare against the human motion diffusion model from EgoEgo [48]. This is similar to EgoAllo, but considers only the SMPL \"mean\" body shape and uses sequence canonicalized coordinates for conditioning and as model output. (3) VAE+Opt." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7077" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.095, + 0.173, + 0.107 + ], + "angle": 0, + "content": "AMASS [56]" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.109, + 0.48, + 0.236 + ], + "angle": 0, + "content": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑Thead ↓
EgoAllo32129.8±1.1109.8±1.10.98±0.006.4±0.1
NoShape32138.1±1.1118.8±1.10.94±0.0144.7±0.4
EgoEgo32184.0±1.5158.6±1.60.81±0.0145.2±1.0
VAE+Opt32199.5±1.3191.4±1.40.49±0.0178.0±1.5
EgoAllo128119.7±1.3101.1±1.31.0±0.006.2±0.1
NoShape128128.1±1.3110.3±1.40.98±0.0144.6±0.7
EgoEgo128167.4±2.1145.8±2.00.92±0.0154.9±1.9
VAE+Opt128205.3±2.6192.3±2.80.75±0.0267.8±3.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.239, + 0.162, + 0.251 + ], + "angle": 0, + "content": "RICH [26]" + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.253, + 0.48, + 0.381 + ], + "angle": 0, + "content": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\\( T_{head} \\) ↓
EgoAllo32193.7±3.4174.8±3.60.95±0.018.8±0.2
NoShape32200.9±3.3183.3±3.60.73±0.0244.9±0.9
EgoEgo32215.4±3.9192.9±4.00.73±0.0256.2±2.9
VAE+Opt32352.0±6.7354.8±6.50.59±0.02319.3±11.6
EgoAllo128176.2±5.6160.1±5.90.96±0.028.9±0.3
NoShape128185.7±5.5169.9±5.80.82±0.0345.8±1.6
EgoEgo128207.8±6.9187.8±6.80.88±0.0366.5±5.4
VAE+Opt128319.8±10.1323.8±10.50.75±0.04274.4±17.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.384, + 0.228, + 0.396 + ], + "angle": 0, + "content": "Aria Digital Twins [61]" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.398, + 0.48, + 0.526 + ], + "angle": 0, + "content": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\\( T_{head} \\)↓
EgoAllo32173.5±1.1146.1±1.10.88±0.01-
NoShape32178.5±1.1153.0±1.10.89±0.01-
EgoEgo32212.5±1.4181.3±1.60.64±0.01-
VAE+Opt32284.9±1.6283.9±1.90.63±0.01-
EgoAllo128155.1±1.6129.3±1.60.94±0.01-
NoShape128163.7±1.6140.0±1.60.96±0.01-
EgoEgo128182.6±2.3153.9±2.60.73±0.02-
VAE+Opt128290.8±3.8282.5±4.40.7±0.02-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.616 + ], + "angle": 0, + "content": "Table 2. Body estimation performance, compared against a baseline without shape prediction, EgoEgo [48], and VAE+Opt [74, 102]. We exclude the \\(\\mathbf{T}_{\\mathrm{head}}\\) metric for ADT because the Biomech57 head joints used by ADT are not directly comparable to the SMPL-H head joints used by our model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.755 + ], + "angle": 0, + "content": "Finally, we compare against an approach based on the SLAHMR [102] framework for human motion estimation from exocentric video. A key advantage of SLAHMR is that it uses an unconditional motion prior [74] in an optimization framework. It can therefore be adapted to new settings without re-training—we keep the same body pose and shape variables as the original pipeline, but replace the exocentric keypoint [71] cost with an egocentric CPF pose alignment cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Due to differences in problem formulation, many existing methods for egocentric human motion estimation are difficult to directly compare. This is particularly true when they have different inputs, such as fisheye cameras [91, 95, 96], wrist-mounted sensors [46], or handheld controller poses [7, 28, 29]. Additionally, prior works like EgoEgo [48] do not incorporate vision inputs for hand estimation. For fairness, we restrict all methods in this section to only CPF or head pose as input." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "EgoAllo improves body motion estimates. We report metrics" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.102, + 0.697, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.174, + 0.655, + 0.186 + ], + "angle": 0, + "content": "(a) Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.094, + 0.892, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.174, + 0.84, + 0.187 + ], + "angle": 0, + "content": "(b) EgoAllo" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.19, + 0.703, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.272, + 0.655, + 0.285 + ], + "angle": 0, + "content": "(c) EgoEgo [48]" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.192, + 0.891, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.272, + 0.844, + 0.284 + ], + "angle": 0, + "content": "(d) VAE+Opt" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.354 + ], + "angle": 0, + "content": "Figure 4. Egocentric human motion estimation for a running sequence. We show the ground-truth, an output from EgoAllo, and outputs from two baselines. The glasses CAD model is placed at the conditioning transformation \\(\\mathbf{T}_{\\mathrm{world,cpf}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.533 + ], + "angle": 0, + "content": "in Table 2 and visualize example outputs in Figure 4. We find that EgoAllo enables significant estimation improvements across all datasets, including accuracy improvements of \\(20\\sim 30\\%\\) over EgoEgo for both shorter and longer evaluation sequences. We found shape estimation critical for producing metric-scale, grounded estimates of human body motion, with the head aligned to input SLAM poses and the feet planted on the observed ground plane. This is evident in qualitative results, improved grounding metrics, and in the \\(6\\sim 7\\%\\) MPJPE gap between EgoAllo and the NoShape ablation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.909, + 0.71 + ], + "angle": 0, + "content": "VAE optimization converges poorly. Optimization-based estimation approaches have been effective for settings with keypoint costs [74, 102], but we found convergence difficult in our less constrained setting. In Table 2, we observe poor generalization: VAE+Opt performs competitively on the AMASS test set, but performance deteriorates dramatically when evaluating on RICH or ADT. VAE+Opt outputs in Figure 4 also look overly smoothed, without the same expressiveness as the conditional predictions of EgoAllo or EgoEgo [48]. This highlights the advantage of using a conditional diffusion model problem for this estimation problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Shape estimation evaluation. To better understand the shape estimation characteristics of EgoAllo, we compare against against the \"mean\" shape used by EgoEgo and the NoShape ablation. On the AMASS test set, we find: EgoAllo slightly improves overall shape (19mm → 18mm mean vertex-to-vertex error) and produces much better height (52mm → 32mm mean height error), but is not able to generalize in terms of body weight (5kg → 8kg mean weight error). The body shape is inferred from the wearer's head pose, which intuitively provides strong height constraints but is less correlated with weight. Accurate height is key for proper scene placement, as reflected by both the MPJPE and GND metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7078" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.255, + 0.106 + ], + "angle": 0, + "content": "4.2. Hand estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.417 + ], + "angle": 0, + "content": "To evaluate hands estimated by EgoAllo, we run HaMeR on the segment of the EgoExo4D [17] validation set that is labeled with 3D hand pose keypoints. We quantitatively compare four hand estimation methods in Table 3. In (1) HaMeR [66], we use HaMeR out-of-the-box on undistorted egocentric RGB images. We do not assume bounding boxes as input; instead, we follow the HaMeR demo code and compute crops using ViTPose [99]. (2) EgoAllo-NoReproj uses all loss terms except for the reprojection loss (Equation 14). Hand guidance is done directly using the 3D wrist poses predicted by HaMeR. (3) EgoAllo-Mono is the same as EgoAllo-NoReproj, but guides hands using the reprojection loss. This accounts for the scale ambiguities that are inherent to the single-frame HaMeR estimates. Finally, we show an extension of our method. (4) EgoAllo-Wrist3D extends EgoAllo-Mono with 3D wrist pose losses from Project Aria's Machine Perception Services [85]—unlike HaMeR, which assumes monocular input, this is a separate sensing setup that also leverages a pair of SLAM cameras that are unique to Project Aria. For fairness across settings, we compute metrics only on timesteps where HaMeR estimates are available." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Results are provided in Table 3. While HaMeR's local poses (PA-MPJPE) are slightly better, EgoAllo's hand-body estimation significantly improves how well hands are estimated in the world coordinate system. Compared to HaMeR, EgoAllo-Mono drops MPJPE from \\(237.90\\mathrm{mm} \\rightarrow 131.45\\mathrm{mm}\\). Reprojection-based guidance is important: despite using the same inputs, EgoAllo-NoReproj outputs are worse than EgoAllo-Mono in both MPJPE and PA-MPJPE. Finally, incorporating more accurate wrist pose estimates from the extra cameras on Project Aria drops errors \\(131.45\\mathrm{mm} \\rightarrow 60.08\\mathrm{mm}\\). This highlights both room for improvement for EgoAllo and how EgoAllo's guidance-based formulation makes it easy to extend when more sensors are available." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Qualitatively, we observed that high hand estimation errors in naive monocular estimation with HaMeR are explained by a combination of detection failures and monocular ambiguities. Even when detections succeed, the scale and distance of monocular HaMeR estimates are often incorrect or flicker in between frames. EgoAllo improves on this by applying kinematic constraints and a diffusion-based prior for plausible body motion—we provide examples of HaMeR estimates rendered jointly with EgoAllo outputs in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.192, + 0.769 + ], + "angle": 0, + "content": "5. Dicussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Limitations and future work. While the core contributions of EgoAllo are general, the current implementation of our system has a few limitations that we hope to explore in future work. First, diffusion model guidance is a test-time optimization process that depends on hyperparameters and incurs a runtime cost. In the future, it may be possible to bootstrap using outputs from our model to train a feedforward model that avoids this step. Success for hand guidance also still depends on reasonable" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.614, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.089, + 0.718, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.089, + 0.812, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.089, + 0.905, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.176, + 0.646, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.176, + 0.771, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.176, + 0.905, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 5. Body estimation improves hand estimation. We show raw outputs from HaMeR [66] in blue and hand-body estimations from EgoAllo in purple. Top: improved scene interaction during touchscreen operation with EgoAllo-Mono. We know a priori that the fingers are contacting the screen in this sequence. Bottom: qualitative examples from EgoExo [17] evaluation, showing the differences between monocular hands and EgoAllo-Wrist3D estimates." + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.454, + 0.88, + 0.561 + ], + "angle": 0, + "content": "
MethodMPJPE ↓PA-MPJPE ↓
HaMeR237.90±1.8913.04±1.89
EgoAllo-NoReproj143.20±0.4214.75±0.42
EgoAllo-Mono131.45±0.3914.71±0.39
EgoAllo-Wrist3D60.08±0.2614.38±0.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Table 3. Hand estimation errors in millimeters. EgoAllo's hand-body estimation can constrain and resolve ambiguities in noisy outputs from HaMeR, which we observe can reduce MPJPE for hands by over \\(40\\%\\). EgoAllo-Wrist3D incorporates 3D wrist poses provided from the grayscale cameras on Project Aria." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.789 + ], + "angle": 0, + "content": "monocular hand estimates. Errors like left/right flipping or spurious detections can therefore cause estimation failure. Finally, we assume flat floors. This is in part because our training data [56] includes floor planes but no detailed scene geometry. As a result, our method will fail in settings like hills or staircases. In the future, we hope to extend our insights to data with more detailed scene information, which concurrent work has highlighted the usefulness of in informing human body estimation [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Conclusion. We presented EgoAllo, a system for estimating human motion using a head-mounted device. EgoAllo jointly estimates human body pose, height, and hand parameters from only egocentric SLAM poses and images. Results highlight the importance of spatial and temporal invariance in conditioning for this problem, while demonstrating how estimated bodies can be used to improve hand estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7079" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.289 + ], + "angle": 0, + "content": "Acknowledgments. This project was funded in part by NSF:CNS-2235013 and IARPA DOI/IBC No. 140D0423C0035. JM was supported by ONR MURI N00014-21-1-2801. YM acknowledges support from the joint Simons Foundation-NSF DMS grant #2031899, the ONR grant N00014-22-1-2102, the NSF grant #2402951, and partial support from TBSI, InnoHK, and the University of Hong Kong. BY is supported by the National Science Foundation Graduate Research Fellowship Program under Grant DGE 2146752. The authors would also like to thank Hongsuk Choi, Michael Taylor, Tyler Bonnen, Songwei Ge, Chung Min Kim, and Justin Kerr for insightful technical discussion and suggestions, as well as Jiaman Li for helpful answers to questions about EgoEgo." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.304, + 0.184, + 0.32 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.387, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. ACM New York, NY, USA, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.444, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.488, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[4] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.545, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[5] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.616, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[6] Marcus A Brubaker, David J Fleet, and Aaron Hertzmann. Physics-based person tracking using the anthropomorphic walker. International journal of computer vision, 87(1-2): 140-155, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.673, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[7] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4221-4231, 2023, 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.758, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[8] R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE conference on computer vision and pattern recognition (CVPR), pages 77-85. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.816, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[9] Haiwei Chen, Shichen Liu, Weikai Chen, Hao Li, and Randall Hill. Equivariant point network for 3d point cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14514-14523, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.33, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.121, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[11] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4800-4810, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.233, + 0.904, + 0.274 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[14] Haiwen Feng, Peter Kulits, Shichen Liu, Michael J Black, and Victoria Fernandez Abrevaya. Generalizing neural human fitting to unseen poses with articulated se (3) equivariance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7977-7988, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.344, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[15] Saeed Ghorbani, Calden Wloka, Ali Etemad, Marcus A Brubaker, and Nikolaus F Troje. Probabilistic character motion synthesis using a hierarchical deep latent variable model. In Computer Graphics Forum, pages 225-239. Wiley Online Library, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[16] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In Int. Conf. Comput. Vis., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[17] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[18] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. arXiv preprint arXiv:2403.18118, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.595, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[19] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In Int. Conf. Comput. Vis., pages 1381-1388, IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.65, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[20] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10884-10894, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.692, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[21] Vladimir Guzov, Yifeng Jiang, Fangzhou Hong, Gerard Pons-Moll, Richard Newcombe, C Karen Liu, Yuting Ye, and Lingni Ma. Hmd: Environment-aware motion generation from single egocentric head-mounted device. arXiv preprint arXiv:2409.13426, 2024. 2, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.762, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7080" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[25] Nicholas Howe, Michael Leventon, and William Freeman. Bayesian reconstruction of 3d human motion from single-camera video. Advances in neural information processing systems, 12, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[26] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13274-13285, 2022. 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.221, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[27] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.348, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[29] Jiaxi Jiang, Paul Streli, Manuel Meier, Andreas Fender, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2, 3, 4, 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.405, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[30] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6142-6152, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[31] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards in-the-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[32] Angjoo Kanazawa, Abhishek Sharma, and David Jacobs. Locally scale-invariant convolutional neural networks. arXiv preprint arXiv:1412.5104, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[33] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7122-7131, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[34] Angjoo Kanazawa, Jason Y Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5614-5623, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.742 + ], + "angle": 0, + "content": "[35] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[36] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[37] Meejin Kim and Sukwon Lee. Fusion poser: 3d human pose estimation using sparse imus and head trackers in real time. Sensors, 22(13):4846, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.845, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[38] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5253-5263, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[41] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[42] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[43] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[44] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6050-6059, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.419, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[45] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[46] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. arXiv preprint arXiv:2401.00847, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.505, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[47] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[48] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17142-17151, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.591, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[49] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.633, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[50] Zongmian Li, Jiri Sedlar, Justin Carpentier, Ivan Laptev, Nicolas Mansard, and Josef Sivic. Estimating 3d motion and forces of person-object interactions from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8649, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.704, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[51] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.747, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi-person linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851-866, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.803, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[53] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7081" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[55] Lingni Ma, Yuting Ye, Fangzhou Hong, Vladimir Guzov, Yifeng Jiang, Rowan Postyeni, Luis Pesqueira, Alexander Gamino, Vijay Baiyya, Hyo Jin Kim, et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. arXiv preprint arXiv:2406.09905, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[56] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5442-5451, 2019. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.219, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[57] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Ecoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[58] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Int. Conf. Comput. Vis., pages 2640-2649, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[59] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[60] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In 2018 international conference on 3D vision (3DV), pages 484-494. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.483, + 0.551 + ], + "angle": 0, + "content": "[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 1, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[62] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 459-468, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10975-10985, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[64] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[65] Georgios Pavlakos, Ethan Weber, Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In Eur. Conf. Comput. Vis., pages 732-749. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[66] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In *arxiv*, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[67] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[68] Chiara Plizzari, Shubham Goel, Toby Perrett, Jacob Chalk, Angjoo Kanazawa, and Dima Damen. Spatial cognition from" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "egocentric video: Out of sight, not out of mind. arXiv preprint arXiv:2404.05072, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.123, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[69] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.195, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[70] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.267, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[71] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2740-2749, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.325, + 0.905, + 0.367 + ], + "angle": 0, + "content": "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.369, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[73] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[74] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 4, 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.513, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[75] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.572, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[76] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.643, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[77] Gregory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net: Localization-classification-regression for human pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3433-3441, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.702, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[78] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.773, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[79] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[80] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7082" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.135 + ], + "angle": 0, + "content": "[81] Silvio Savarese and Li Fei-Fei. 3d generic object categorization, localization and pose estimation. In 2007 IEEE 11th international conference on computer vision, pages 1-8. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.485, + 0.176 + ], + "angle": 0, + "content": "[82] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.179, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K Hodgins. Motion capture from body-mounted cameras. In ACM SIGGRAPH 2011 papers, pages 1-10. ACM New York, NY, USA, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.236, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[85] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[86] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[87] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[88] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[89] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[90] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[91] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7728-7738, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[92] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[93] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[94] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[95] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11500-11509, 2021. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[96] Jian Wang, Zhe Cao, Diogo Luvizon, Lingjie Liu, Kripasindhu Sarkar, Danhang Tang, Thabo Beeler, and Christian Theobalt. Egocentric whole-body motion capture with fisheyevit and diffusion-based motion refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 777-787, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.179, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[97] Laurenz Wiskott and Terrence J Sejnowski. Slow feature analysis: Unsupervised learning of invariances. Neural computation, 14(4):715-770, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.223, + 0.907, + 0.292 + ], + "angle": 0, + "content": "[98] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.295, + 0.907, + 0.336 + ], + "angle": 0, + "content": "[99] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.339, + 0.907, + 0.393 + ], + "angle": 0, + "content": "[100] Jingyun Yang, Congyue Deng, Jimmy Wu, Rika Antonova, Leonidas Guibas, and Jeannette Bohg. Equivact: Sim(3)-equivariant visuomotor policies beyond rigid object manipulation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.397, + 0.907, + 0.451 + ], + "angle": 0, + "content": "[101] Jingyun Yang, Zi-ang Cao, Congyue Deng, Rika Antonova, Shuran Song, and Jeannette Bohg. Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning. arXiv preprint arXiv:2407.01479, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.454, + 0.906, + 0.509 + ], + "angle": 0, + "content": "[102] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21222-21232, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[103] Brent Yi, Weijia Zeng, Sam Buchanan, and Yi Ma. Canonical factors for hybrid neural fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3414-3426, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.57, + 0.906, + 0.611 + ], + "angle": 0, + "content": "[104] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.614, + 0.907, + 0.681 + ], + "angle": 0, + "content": "[105] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. arXiv preprint arXiv:2305.01599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.686, + 0.906, + 0.728 + ], + "angle": 0, + "content": "[106] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.907, + 0.784 + ], + "angle": 0, + "content": "[107] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[108] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[109] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7083" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[110] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. arXiv preprint arXiv:2305.03462, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[111] Daiwei Zhang, Gengyan Li, Jiajie Li, Mickaël Bressieux, Otmar Hilliges, Marc Pollefeys, Luc Van Gool, and Xi Wang. Egogaussian: Dynamic scene understanding from egocentric video with 3d gaussian splatting. arXiv preprint arXiv:2406.19811, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[112] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.249, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[113] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.319, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[114] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.39, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[115] Zhengdong Zhang, Arvind Ganesh, Xiao Liang, and Yi Ma. Tilt: Transform invariant low-rank textures. International journal of computer vision, 99:1-24, 2012. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7084" + } + ] +] \ No newline at end of file diff --git a/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_origin.pdf b/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e68c83700ac6e3010f3814245588e896e84cff5a --- /dev/null +++ b/2025/Estimating Body and Hand Motion in an Ego-sensed World/71e1babb-034c-40be-8709-39f13bf4244a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93855a04b7831a8b4c8c429bd3ab0a6c6a364275440f30e25c639cfa98b646e8 +size 6622327 diff --git a/2025/Estimating Body and Hand Motion in an Ego-sensed World/full.md b/2025/Estimating Body and Hand Motion in an Ego-sensed World/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9ed0bbd869e1064d6f51135c2b013f61a8b816e9 --- /dev/null +++ b/2025/Estimating Body and Hand Motion in an Ego-sensed World/full.md @@ -0,0 +1,435 @@ +# Estimating Body and Hand Motion in an Ego-sensed World + +Brent Yi $^{1}$ Vickie Ye $^{1}$ Maya Zheng $^{1}$ Yunqi Li $^{2}$ Lea Müller $^{1}$ Georgios Pavlakos $^{3}$ Yi Ma $^{1}$ Jitendra Malik $^{1}$ Angjoo Kanazawa $^{1}$ UC Berkeley ${}^{2}$ ShanghaiTech ${}^{3}$ UT Austin + +![](images/736b3a79b52c98a5c596117a87cd1888334df4458f4e6ca710a1c6af83829da1.jpg) +Figure 1. EgoAllo. We present a system that estimates human body pose, height, and hand parameters from egocentric SLAM poses and images. Outputs capture the wearer's actions in the allocentric reference frame of the scene, which we visualize here with 3D reconstructions. + +![](images/8ecba775f6884bd43051b09d83eacbc8f6c2c586bcf2eb6223483ba1e7f9e770.jpg) + +# Abstract + +We present EgoAllo, a system for human motion estimation from a head-mounted device. Using only egocentric SLAM poses and images, EgoAllo guides sampling from a conditional diffusion model to estimate 3D body pose, height, and hand parameters that capture a device wearer's actions in the allocentric coordinate frame of the scene. To achieve this, our key insight is in representation: we propose spatial and temporal invariance criteria for improving model performance, from which we derive a head motion conditioning parameterization that improves estimation by up to $18\%$ . We also show how the bodies estimated by our system can improve hand estimation: the resulting kinematic and temporal constraints can reduce world-frame errors in single-frame estimates by $40\%$ . + +# 1. Introduction + +Head-mounted devices are becoming increasingly mainstream. In addition to offering new challenges for 3D scene understanding [18, 57, 68, 111], egocentric sensors from these devices are + +unique in that their outputs are coupled to a human wearer's motion in the world. Using these sensors to understand the wearer in addition to the scene around them is essential for applications in augmented reality, robotics, and assistive technologies. + +We therefore introduce EgoAllo, a system that uses egocentric inputs to estimate the wearer and their motion in the world, or allocentric, coordinate frame. We take as input sensed metric SLAM head poses and egocentric video from devices like Project Aria [61]. We then output estimates of human body pose, height, and hand motion parameters. + +This is a difficult task: while body parts like hands occasionally appear in egocentric frames, most body parameters are never directly observed. Estimated body pose, body height, and hand parameters must be consistent with both the sensed egomotion and visual observations. This setting differs from most prior works in egocentric human motion estimation [7, 48], which focus on body pose and do not address the challenges of height and hand motion. + +Our proposed system uses a head pose-conditioned diffusion model as a motion prior, as well as a Levenberg-Marquardt guidance optimizer for sampling hand-body sequence that align with image observations. Our results are enabled by a key + +insight: that the representation used for head pose conditioning is critical for accurate full-body motion estimation. We study choices for this representation by (1) identifying desirable spatial and temporal invariance properties that are not fulfilled by existing systems and (2) using these properties to derive improved parameterizations for our motion prior. + +We systematically evaluate our system on four datasets. For body estimation, we find that improving the conditioning parameterization leads to an accuracy improvement between $4.9\%$ and $17.9\%$ . Furthermore, we observe that the resulting system can improve hand estimation, reducing world-frame errors by over $40\%$ compared to single-frame estimates. Code, model, and more results can be found on our project webpage. + +# 2. Related Work + +3D human recovery from external visual inputs. A large body of work has addressed estimating the parameters of human body models like SCAPE [2] or SMPL and its variants [52, 63, 79] from third-person visual inputs, where human subjects are observed from the view of outside cameras. The majority of these works focus on extracting 3D representations from single images, for example by lifting 2D keypoint observations to 3D [58], via end-to-end regression [20, 31, 33, 40, 60, 62, 77], via optimization [19, 44, 63], or by exploiting synergies between regression and optimization [42]. When multiple frames are available in the form of a video, temporal context and tracking can also be incorporated [16, 34, 39, 64, 65, 71, 108]. The inputs (images) and outputs (human meshes) of many of these systems are similar to the egocentric setting addressed by EgoAllo, but egocentric devices present unique challenges because the body being estimated is typically behind the outwards-facing cameras used as input. + +Priors for human motion. The primary challenge of ego-sensed human motion estimation is limited observability; a prior is required to resolve ambiguities. For human motion, these priors are typically framed as unconditional distributions over plausible human motion. Distributions can be represented either by modeling the physical constraints of our world [6, 50, 67, 73] or by learning generative models of human motion directly from data. For learning unconditional priors, classical data-driven approaches include fitting mixtures-of-Gaussians to 3D keypoint trajectories [25], while modern approaches include training variational autoencoders [38, 75] to model either autoregressive transitions [15, 51, 74] or full spatiotemporal sequences [22]. After training, these priors can be applied to estimation problems in iterative optimization frameworks [41, 74, 102]. EgoAllo is built on the same intuition as these methods, but follows previous work in ego-sensed motion estimation and uses a task-specific conditional prior. + +Denoising diffusion for human motion. The core of EgoAllo is a denoising diffusion model [23, 69, 84] from which we can sample 3D human body motion. While diffusion mod + +els are primarily known for their success in text-conditioned image generation [78, 80], they have also enabled advances in human motion synthesis conditioned on modalities like text [35, 36, 112], music [1, 93], poses [35, 48], and object geometry [43, 47, 49]. EgoAllo adopts a similar conditional diffusion approach, while specifically studying the design of conditioning parameters used for ego-sensed human motion estimation. The iterative nature of denoising diffusion also enables guidance [11, 13, 30, 35, 87, 114], where denoising steps are steered to satisfy a desired objective. We use guidance to incorporate observations like visual hand pose observations during test-time. + +Human motion from egocentric observations. EgoAllo builds on intuition from several prior works in egocentric sensing for human motion estimation. Many rely on fisheye cameras that place the wearer's body into the field of view [27, 76, 91, 91, 92, 95, 96, 98]. Other approaches rely on body-mounted cameras [83], simulation-based physical plausibility [54, 106, 107], body- and hand-mounted inertial sensors [37, 46, 104, 105], handheld controllers [7, 28, 29], and interaction cues from other humans [59]. Concurrent works have also used the Nymeria [55] dataset for egocentric motion with language description outputs [24], as well for online settings with scene geometry and CLIP [70] feature inputs [21]. Most relevantly, EgoEgo [48] demonstrates how human body poses can be estimated offline without body observability assumptions. The authors accomplish this by carefully integrating several components: a monocular SLAM system [89], a pose-conditioned gravity vector regression network, an optical flow feature-conditioned head orientation and scale regression network, and a head pose-conditioned body diffusion model. EgoAllo differs in both inputs—we study conditioning parameters computed from the metric SLAM poses provided by devices like Project Aria [85]—and outputs—we consider body height variation and hand poses. + +Conditioning for ego-sensed poses. Prior works vary in how head pose information is parameterized and used as neural network input. AvATARPoser [28] and BoDiffusion [7] parameterize head pose as four components: world-frame orientation, orientation deltas, world-frame position, and world-frame position deltas. These works focus on settings with VR controller input, and parameterize controller pose inputs the same way. EgoEgo [48]'s diffusion model uses only absolute head positions and orientations, but similar to HuMoR [74], in implementation defines a per-sequence canonical coordinate frame to ensure that all input trajectories passed to the model are aligned with the same initial $xy$ position and forward direction. In our work, we refer to this as sequence canonicalization. Finally, EgoPoser [29] proposes a similar scheme that aligns initial positions for both head pose and controller pose inputs. We propose an alternative to these parameterizations that is motivated by the robustness and generalization benefits of invariance, as observed in prior work for designing both + +![](images/db89cba97a9938764e3ab6e3b4031644470fdfb31a685e983ad268b1d6b1905f.jpg) +Figure 2. Overview of components in EgoAllo. We restrict the diffusion model to local body parameters (Section 3.1.1). An invariant parameterization $g(\cdot)$ (Section 3.1.2) of SLAM poses is used to condition a diffusion model. These can be placed into the global coordinate frame via global alignment (Section 3.2.1) to input poses. When available, egocentric video is used for hand detection via HaMeR [66], which can be incorporated into samples via guidance (Section 3.2.2). + +representations [10, 53, 81, 97, 103, 110, 115] and neural network architectures [8, 9, 12, 14, 32, 45, 90, 100, 101, 109]. Specifically, we introduce in Section 3.1.2 a parameterization that is invariant to both spatial and temporal shifts. + +# 3. Method + +We study the problem of using sensors from an egocentric device to estimate the actions of the wearer in an allocentric coordinate frame. We assume a flat floor and two inputs: poses from the device's SLAM system and egocentric video. + +Our system uses head pose information to condition a diffusion-based prior over body pose and height, and incorporates visual hand observations during sampling. This allows it to benefit from both 3D human motion capture datasets [56], which are used for the motion prior, and from large-scale image datasets [66], which are used for hand estimates. + +# 3.1. Ego-conditioned motion diffusion + +Notation: we use $\mathbf{T}_{\mathrm{A,B}} = (\mathbf{R}_{\mathrm{A,B}},\mathbf{p}_{\mathrm{A,B}})$ to denote an SE(3) transform to frame A from frame B, composed of rotation $(\mathbf{R}_{\mathrm{A,B}})$ and position $(\mathbf{p}_{\mathrm{A,B}})$ terms. Temporal steps $t$ are superscripted and diffusion noise steps $n$ are subscripted. $\vec{x}_0^t$ thus refers to the $t$ -th timestep of a clean $(n = 0)$ human motion sequence. + +Given an observation window of $T$ timesteps, EgoAllo's motion prior is a diffusion model that aims to capture the distribution of human motions $\vec{x}_0 = \{\vec{x}_0^1,\dots,\vec{x}_0^T\}$ conditioned on head pose encodings $\vec{c} = \{\vec{c}^1,\dots,\vec{c}^T\}$ . For each timestep $t$ , we represent human motion in the form of SMPL-H [52, 79] model parameters $\{\mathbf{T}_{\mathrm{world,root}}^t,\Theta^t,\beta \}$ : root transforms $\mathbf{T}_{\mathrm{world,root}}^t\in \mathrm{SE}(3)$ , where the person's root frame is located at their pelvis, local joint rotation matrices $\Theta^t\in \mathbb{R}^{51\times 3\times 3}$ , and time-invariant shape $\beta \in \mathbb{R}^{16}$ . + +Dependencies between local joint rotations, body size variation, and global motion make this learning task a + +challenging one. Our key insight is that this difficulty can be reduced by designing parameterizations with desirable invariance properties. Spatial and temporal invariances allow the model to focus on the essential structure of motion, without being affected by irrelevant shifts in position or time. + +# 3.1.1. Diffusion output representation + +As output, we sample body and hand joint rotations, body shapes, and binary contact predictions $\vec{x}_0^t = \{\Theta^t,\beta^t,\psi_{j = 1\dots 21}^t\}$ , where body shape $\beta^t$ is supervised to be equal for all timesteps and $\psi_j^t$ is a per-joint contact indicator. Notably, these parameters are all local—we discuss how outputs can be placed into the allocentric coordinate frame in Section 3.2.1. + +We choose this output set for three main reasons. (1) Body shape encodes the wearer's height, which is critical for grounding in the metric-scale geometry of the scene. This is rarely considered by prior work: with the exception of [29], existing methods [7, 28, 48] otherwise produce outputs using a fixed "mean" human shape. (2) Contact predictions enable losses for common problems like foot skating, which are discussed in Section 3.2.2. (3) Finally, local bodies are invariant to the global coordinate frame. As we discuss next, the conditioning parameterization for the model can therefore also be invariant to arbitrary transformations along the floor plane. + +# 3.1.2. Invariant conditioning + +The goal of our conditioning representation is to map raw SLAM poses (head motion) to a parameterization that is amenable to learning for the diffusion model. + +Raw inputs. To capture the head motion at each time step, we assume as input poses of a central pupil frame (CPF), which the SLAM systems of devices like Project Aria can provide with millimeter-level accuracy [85]. For time 1...T, we + +reparameterize these poses for conditioning using a function $g$ : + +$$ +\mathbf {T} _ {\text {w o r l d , c p f}} ^ {t} = \left(\mathbf {R} _ {\text {w o r l d , c p f}} ^ {t}, \mathbf {p} _ {\text {w o r l d , c p f}} ^ {t}\right) \in \mathrm {S E} (3), \tag {1} +$$ + +$$ +\{\vec {c} ^ {1}, \dots , \vec {c} ^ {T} \} = g (\{\mathbf {T} _ {\mathrm {w o r l d}, \mathrm {c p f}} ^ {1}, \dots , \mathbf {T} _ {\mathrm {w o r l d}, \mathrm {c p f}} ^ {T} \}). \tag {2} +$$ + +The CPF frame differs from prior works that condition on a coordinate frame attached to the SMPL human model's "head joint" [7, 28, 29, 48]. The offset between this head joint and the device pose depends on the head shape captured by $\beta^t$ , and is thus difficult to precompute in our setting. + +To encode absolute height, we assume that the world frame's $+z$ -axis faces upwards, and that the ground is located at $z = 0$ . Ground parameters are directly available in the training data [56]; at test time, we can also extract these parameters from sparse SLAM points via RANSAC (Appendix A.3.5). + +Invariance goals. As discussed in Section 2, prior work varies in how the function $g$ is implemented. To understand how choices impact learning, we propose two invariance properties for head motion representations. Each reduces representational redundancy, which eases the learning problem. We provide more detailed explanations and visualizations of these properties in Appendix A.1. + +Invariance 1 (Spatial) Global transformations along the floor plane should not affect a person's local motion. Given $\mathbf{T}_{xy} \in SE(3)$ restricted to the $XY$ plane, $g$ should fulfill $g(\{\mathbf{T}_{xy}\mathbf{T}_{world, cpf}^t\}_t) = g(\{\mathbf{T}_{world, cpf}^t\}_t) \forall \mathbf{T}_{xy}$ . + +Invariance 2 (Temporal) If the same human motion occurs at different points in time, the extracted head motion representation should remain consistent. This can be expressed as temporal shift equivariance. Let $\vec{c}^t$ be as defined in Equation 2. For any shift $\delta$ such that $\{\vec{c}_{shift}^{1},\dots,\vec{c}_{shift}^{T}\} = g(\{\mathbf{T}_{world,cpf}^{1 + \delta},\dots,\mathbf{T}_{world,cpf}^{T + \delta}\})$ , $g$ should satisfy $\vec{c}_{shift}^{t} = \vec{c}^{t + \delta}$ for overlapping timesteps. + +No parameterization used by existing work satisfies both of these properties. The sequence canonicalization approach of EgoEgo [48] achieves spatial invariance (Invariance 1), but inserts a sequence-wide dependency on the first timestep of each window that results in a violation of Invariance 2. The absolute poses and pose deltas used by [7, 28] satisfy Invariance 2, but not Invariance 1. Finally, the relative positions considered by [29] are neither spatially nor temporally invariant. + +Invariant conditioning. We propose a formulation for $g$ that achieves both invariance properties by locally canonicalizing head motion with respect to the floor at each timestep. We build on the relative motion of the CPF frame at each time $t$ , which respects both Invariance 1 and 2: + +$$ +\Delta \mathbf {T} _ {\mathrm {c p f}} ^ {t - 1, t} = \left(\mathbf {T} _ {\mathrm {w o r l d}, \mathrm {c p f}} ^ {t - 1}\right) ^ {- 1} \mathbf {T} _ {\mathrm {w o r l d}, \mathrm {c p f}} ^ {t}. \tag {3} +$$ + +Importantly, the translation component of this transformation is in the local frame. This is distinct from world-frame position deltas [7, 28, 29], which still violate Invariance 1. + +![](images/abcbbc04c56a29d9b6148f942092360c06258d94a86b387c001f31a861ac83b5.jpg) +Figure 3. Locally canonicalized coordinate frames. We compute our invariant conditioning parameterization (Equation 4) using transformations computed from three coordinate frames. Following [85], the CPF has the $z$ -axis forward. Following HuMoR [74], the world and canonical $z$ -axes point up. Canonical frames are computed by projecting the CPF frame origin to the ground plane, then aligning the canonical $y$ -axis to the CPF forward direction. The axes are color-coded as $x$ (red), $y$ (green), and $z$ (blue). + +Relative transforms alone do not encode information relative to the scene or floor: full trajectories can even be flipped upside down without impacting $\Delta \mathbf{T}_{\mathrm{cpf}}^{t - 1,t}$ . We therefore propose to ground relative motion to the floor plane with a transformation between the CPF frame and a per-timestep canonical frame, which is computed by projecting the CPF frame to the floor. This encodes head height and orientation. Our full representation then becomes: + +$$ +\vec {c} ^ {t} = \underbrace {\left\{\Delta \mathbf {T} _ {\mathrm {c p f}} ^ {t - 1 , t} , \quad \left(\mathbf {T} _ {\text {w o r l d , c a n o n i c a l}} ^ {t}\right) ^ {- 1} \mathbf {T} _ {\text {w o r l d , c p f}} ^ {t} \right\}} _ {\text {I n v a r i a n t i m p l e m e n t a t i o n o f} g (\cdot)} \tag {4} +$$ + +We visualize an example of a canonical frame in Figure 3 and our full representation in Appendix A.1. Canonical frames are positioned by projecting the CPF origin to the floor plane; given standard bases $\mathbf{e}_{\{x,y,z\}}$ , we compute: + +$$ +\mathbf {p} _ {\text {w o r l d , c a n o n i c a l}} ^ {t} = \left[ \begin{array}{l l l} \mathbf {e} _ {x} & \mathbf {e} _ {y} & \vec {0} \end{array} \right] ^ {\top} \mathbf {p} _ {\text {w o r l d , c p f}} ^ {t}. \tag {5} +$$ + +For orientation, we align the canonical frame's local $z$ -axis parallel to the world $z$ -axis and its local $y$ -axis toward the "forward" direction $\vec{v}^t$ of the CPF frame. With $\mathbf{R}_{z}(\cdot):\mathbb{R}\to \mathrm{SO}(3)$ constructing a $z$ -axis rotation and $\mathbf{e}_{\{x,y,z\}}$ again as standard bases, we compute this as: + +$$ +\vec {v} ^ {t} = \mathbf {R} _ {\text {w o r l d , c p f}} ^ {t} \mathbf {e} _ {z}, \tag {6} +$$ + +$$ +\mathbf {R} _ {\text {w o r l d , c a n o n i c a l}} ^ {t} = \mathbf {R} _ {z} \left(- \arctan 2 \left(\mathbf {e} _ {x} ^ {\top} \vec {v} ^ {t}, \mathbf {e} _ {y} ^ {\top} \vec {v} ^ {t}\right)\right). \tag {7} +$$ + +This canonical frame definition is an important departure from prior work. While EgoEgo [48] and HuMoR [74] use similar canonical frames, they only compute one per sequence. Instead, + +we compute Equations 5 and 7 at every timestep. This enables floor plane grounding without sacrificing Invariance 2. + +Connection to prior work. EgoAllo's invariant conditioning shares similarities with the spatial and temporal normalization proposed by EgoPoser [29]. (1) EgoPoser's temporal normalization subtracts positions by the first timestep's position. This is a position-only version EgoEgo [48]'s sequence canonicalization. It does not consider rotation, and therefore only partially fulfills Invariance 1. Like EgoEgo, it also does not fulfill Invariance 2. In Appendix A.2.3, we observe that this approach only has a small impact when adapted to EgoAllo's problem setting. (2) EgoPoser's spatial normalization subtracts the head's XY position from the hand pose inputs. EgoAllo does not use hand pose as network input, so this is not applicable in our setting. + +# 3.2. Estimation via sampling + +We use our local body representation and invariant conditioning strategies to train a motion prior in the form of a denoising diffusion model [23]. Given diffusion step $n = N \ldots 1$ , we follow [72] and approximate the denoising process as: + +$$ +p _ {\theta} (\vec {x} _ {n - 1} | \vec {x} _ {n}, \vec {c}) = \mathcal {N} \big (\mu_ {\theta} (\vec {x} _ {n}, n, \vec {c}), \sigma_ {n} ^ {2} \mathbf {I} \big), \qquad (8) +$$ + +where a transformer [94] $\mu_{\theta}$ is trained to predict the posterior mean from noised sample $\vec{x}_n$ and conditioning $\vec{c}$ . With noise-dependent weight term $w_{n}$ , the loss can be written as: + +$$ +\min _ {\theta} \mathbb {E} _ {\vec {x} _ {0}} \mathbb {E} _ {n \sim \mathcal {U}} \left[ w _ {n} \| \mu_ {\theta} (\vec {x} _ {n}, n, \vec {c}) - \vec {x} _ {0} \| ^ {2} \right]. \tag {9} +$$ + +After training, we estimate human motions by following DDIM [86] for sampling. The final EgoAllo sampling procedure includes several additional components: a global alignment phase, guidance losses for physical constraints and visual hand observations, and a path fusion [3] approach for longer sequence lengths. We describe these below. + +# 3.2.1. Global alignment + +To place sampled bodies into the allocentric coordinate system, we compute the absolute pose of the SMPL-H root as: + +$$ +\mathbf {T} _ {\text {w o r l d , r o o t}} ^ {t} = \mathbf {T} _ {\text {w o r l d , c p f}} ^ {t} \mathbf {T} _ {\text {c p f , r o o t}} ^ {(\Theta^ {t}, \beta^ {t})}, \tag {10} +$$ + +where $\mathbf{T}_{\mathrm{cpf,root}}^{(\Theta^t,\beta^t)}$ computes the transform between the root of the human and their CPF frame for a given set of local pose and shape parameters. Similar processes are applied in [7, 28, 29]. In contrast to directly outputting absolute body transformations from the diffusion model [48], this guarantees exact alignment between estimates and input SLAM sequences. + +# 3.2.2. Guidance losses + +Our diffusion model learns a distribution of human motion conditioned on the central pupil frame motion. At test time, we incorporate constraints from physical priors and visual hand observations via guidance [11, 30, 114]. Similar to [35, 47], + +we accomplish this by applying costs to the joint rotations $\Theta = \{\Theta^1,\dots,\Theta^T\}$ predicted by $\mu_{\theta}(\vec{x}_n,n,\vec{c})$ . We treat the body shape $\beta^{t}$ and contacts $\psi_{j = 1\dots 21}^{t}$ as fixed and optimize over body and finger pose to minimize hand observation, skating, and prior costs with a Levenberg-Marquardt optimizer: + +$$ +\mathcal {E} _ {\text {g u i d a n c e}} ^ {(\Theta)} = \mathcal {E} _ {\text {h a n d s}} ^ {(\Theta)} + \mathcal {E} _ {\text {s k a t e}} ^ {(\Theta)} + \mathcal {E} _ {\text {p r i o r}} ^ {(\Theta)}. \tag {11} +$$ + +We begin by running HaMeR on the egocentric image corresponding to each timestep $t$ . When detected, this produces 3D hand estimates in the form of MANO [79] joint parameters and camera-centric 3D hand keypoints $\hat{\mathbf{p}}_{\mathrm{camera},j}^{t}$ for hand joint set $j \in \mathcal{H}$ . Optionally, wrist and palm poses can also be estimated using Project Aria's Machine Perception Services [85]. With each subscripted $\lambda$ indicating a scalar weighting term, we have: + +$$ +\mathcal {E} _ {\text {h a n d s}} ^ {(\Theta)} = \lambda_ {\text {h a n d s 3 D}} \mathcal {E} _ {\text {h a n d s 3 D}} ^ {(\Theta)} + \lambda_ {\text {r e p r o j}} \mathcal {E} _ {\text {r e p r o j}} ^ {(\Theta)}. \tag {12} +$$ + +The 3D objective $\mathcal{E}_{\mathrm{hands3D}}^{(\Theta)}$ minimizes the distance between the detected hand parameters and the corresponding SMPL-H hand parameters, in terms of wrist pose and local joint rotations. With $\Pi_K$ as projection with camera intrinsics $K$ , $\mathbf{p}_{\mathrm{world},j}^{(\Theta^t)} \in \mathbb{R}^3$ as the world position for joint $j$ at time $t$ , and $\mathbf{T}_{\mathrm{camera,cpf}}$ from the device calibration, the reprojection cost is: + +$$ +\mathcal {E} _ {\text {r e p r o j}} ^ {(\Theta)} = \sum_ {t, j \in \mathcal {H}} | | \Pi_ {K} \left(\mathbf {p} _ {\text {c a m e r a}, j} ^ {(\Theta^ {t})}\right) - \Pi_ {K} \left(\hat {\mathbf {p}} _ {\text {c a m e r a}, j} ^ {t}\right) | | _ {2} ^ {2}, \tag {13} +$$ + +$$ +\mathbf {p} _ {\text {c a m e r a}, j} ^ {(\Theta^ {t})} = \mathbf {T} _ {\text {c a m e r a , c p f}} \left(\mathbf {T} _ {\text {w o r l d , c p f}} ^ {t}\right) ^ {- 1} \mathbf {p} _ {\text {w o r l d}, j} ^ {(\Theta^ {t})}. \tag {14} +$$ + +To reduce foot skating, we use contact predictions to apply a skating cost [74, 102] for each time $t$ and joint $j$ : + +$$ +\mathcal {E} _ {\text {s k a t e}} ^ {(\Theta)} = \sum_ {t, j} \lambda_ {\text {s k a t e}} \| \frac {1}{2} \left(\psi_ {j} ^ {t} + \psi_ {j} ^ {t - 1}\right) \left(\mathbf {p} _ {\text {w o r l d}, j} ^ {t} - \mathbf {p} _ {\text {w o r l d}, j} ^ {t - 1}\right) \| _ {2} ^ {2}. \tag {15} +$$ + +Finally, we minimize a prior cost $\mathcal{E}_{\mathrm{prior}}^{(\Theta)}$ . This cost penalizes deviations between joint rotations $\Theta^t$ and rotations $\hat{\Theta}^t$ from the denoiser $\mu_{\theta}(\vec{x}_n,n,\vec{c})$ , while also encouraging smoothness. We detail the full list of terms used for this in Appendix A.3.4. + +# 3.2.3. Sequence length extrapolation + +For longer sequences at test time, we draw on existing methods in compositional generation for both image [3, 113] and human motion [4, 82] diffusion models. We train our motion prior using subsequences of up to length 128; when input observations exceed this length at test time, we split into windows with a 32-timestep overlap between neighbors. We then run our model $\mu_{\theta}(\vec{x}_n,\vec{c},n)$ on windows in parallel. Diffusion paths for overlapping regions are fused following MultiDiffusion [3] after each denoising step. + +# 4. Experiments + +We conduct a series of experiments to evaluate EgoAllo's conditioning parameterization, body estimation accuracy, and hand estimation performance. + +
ConditioningSeqlenInvariance 1/2MPJPE ↓% DiffPA-MPJPE ↓% DiffGND ↑
EgoAllo (Eq. 4)32✓ /✓129.8±1.1—109.8±1.1—0.98±0.00
Absolute+Local Relative32P /✓133.0±1.12.4%113.6±1.23.4%0.95±0.00
Absolute+Global Deltas [7,28]32X /✓136.2±1.14.9%118.3±1.27.7%0.93±0.01
Sequence Canonicalization [48]32✓ /X153.1±1.517.9%128.7±1.517.1%0.76±0.01
Absolute32X /✓159.9±1.223.2%141.0±1.328.4%0.89±0.01
EgoAllo (Eq. 4)128✓ /✓119.7±1.3—101.1±1.3—1.00±0.00
Absolute+Local Relative128P /✓124.5±1.34.0%104.9±1.43.8%1.00±0.00
Absolute+Global Deltas [7,28]128X /✓127.4±1.36.4%109.8±1.48.6%0.99±0.00
Sequence Canonicalization [48]128✓ /X134.0±1.811.9%112.1±1.610.9%0.88±0.02
Absolute128X /✓148.3±1.523.9%131.2±1.629.8%0.96±0.01
+ +Table 1. Motion prior conditioning comparison. We train and evaluate otherwise identical models using four conditioning parameterizations on AMASS [56] test set sequences, using sequences of length 32 and 128. Parameterizations vary in their spatial (1) and temporal (2) invariance properties, which we loosely classify as following completely ( $\checkmark$ ), partially ( $\mathsf{P}$ ), or not at all ( $\times$ ). The conditioning parameterization used by EgoAllo reduces errors by almost $18\%$ compared to the sequence canonicalization approach used by the most relevant related work [48]. + +Training. To train EgoAllo models used in our experiments, we need sequences containing human body and hand pose parameters, body shapes, and device SLAM poses $\mathbf{T}_{\mathrm{world,cpf}}^t$ . Similar to prior work [7, 28, 48], we train EgoAllo using AMASS [56] with synthesized device poses. We annotate train split sequences by anchoring a central pupil frame between vertices corresponding to the left and right pupils in the blend skinned mesh, and at train time sample sequences between length 32 and 128. + +Evaluation. We evaluate with four datasets. We use AMASS [56], RICH [26], and Aria Digital Twins (ADT) [61] for body estimation evaluation, and EgoExo4D [17] for hand estimation evaluation. AMASS and RICH do not include egocentric data; we annotate these with synthetic device poses using the same procedure we use for training. ADT and EgoExo4D both include egocentric images and SLAM poses captured using Project Aria glasses [85], which we use directly. + +Metrics. To quantify performance, we report four metrics: (1) MPJPE is a world-frame mean per-joint position error (millimeters). (2) PA-MPJPE is the Procrustes-aligned mean per-joint position error in millimeters, where joint positions are aligned on a per-timestep basis before error are computed. (3) GND is a grounding metric, designed in response to a phenomena where ego-sensed humans "float" above the ground. Given a human body trajectory, this metric contains a simple binary indicator of whether the feet of the human ever touch the ground plane. (4) $\mathbf{T}_{\mathrm{head}}$ is the average SMPL head joint position error in millimeters. + +# 4.1. Body estimation + +In our first set of experiments, we evaluate body estimation from only device SLAM poses, without considering images or hands. This setting allows us to isolate the advantages of our body motion prior, while directly comparing against methods that do not consider hands. + +# 4.1.1. Invariant conditioning evaluation + +We begin by evaluating the importance of the spatial and temporal invariance criteria discussed in Section 3.1.2. We do this by comparing five implementations of the conditioning $g$ : (1) EgoAllo is the final invariant representation that we propose in Equation 4. (2) Absolute+Local Relative appends absolute poses with the relative pose deltas written in Equation 3. (3) Absolute+Global Deltas appends absolute poses with relative orientation and the world-frame position deltas used by [7, 28]. (4) Sequence Canonicalization uses the alignment approach implemented by [48], which violates temporal invariance. (5) Absolute naively conditions on absolute poses, which violate spatial invariance. + +We train conditional diffusion models with otherwise identical architecture using each parameterization, and then evaluate on the AMASS [56] test set. Metrics and percent differences compared to EgoAllo are reported in Table 1. + +Overall, we find that the choice of conditioning parameterization makes a dramatic impact on estimation accuracy. We observe accuracy improve consistently as invariance properties are incorporated into the representation. Compared to EgoAllo, Absolute conditioning increases MPJPE by over $23\%$ for both shorter (length 32) and longer (length 128) sequences. Compared to EgoAllo, SeqCanonical conditioning increases MPJPE by nearly $18\%$ for length 32 sequences and $12\%$ for length 128 sequences. + +# 4.1.2. Comparisons against baselines + +To further study EgoAllo's body estimation quality, we compare against three baselines. (1) NoShape. First, NoShape refers to a variation of EgoAllo that turns off shape estimation, and thus cannot estimate the wearer's height. (2) EgoEgo. We also compare against the human motion diffusion model from EgoEgo [48]. This is similar to EgoAllo, but considers only the SMPL "mean" body shape and uses sequence canonicalized coordinates for conditioning and as model output. (3) VAE+Opt. + +AMASS [56] + +
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑Thead ↓
EgoAllo32129.8±1.1109.8±1.10.98±0.006.4±0.1
NoShape32138.1±1.1118.8±1.10.94±0.0144.7±0.4
EgoEgo32184.0±1.5158.6±1.60.81±0.0145.2±1.0
VAE+Opt32199.5±1.3191.4±1.40.49±0.0178.0±1.5
EgoAllo128119.7±1.3101.1±1.31.0±0.006.2±0.1
NoShape128128.1±1.3110.3±1.40.98±0.0144.6±0.7
EgoEgo128167.4±2.1145.8±2.00.92±0.0154.9±1.9
VAE+Opt128205.3±2.6192.3±2.80.75±0.0267.8±3.1
+ +RICH [26] + +
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\( T_{head} \) ↓
EgoAllo32193.7±3.4174.8±3.60.95±0.018.8±0.2
NoShape32200.9±3.3183.3±3.60.73±0.0244.9±0.9
EgoEgo32215.4±3.9192.9±4.00.73±0.0256.2±2.9
VAE+Opt32352.0±6.7354.8±6.50.59±0.02319.3±11.6
EgoAllo128176.2±5.6160.1±5.90.96±0.028.9±0.3
NoShape128185.7±5.5169.9±5.80.82±0.0345.8±1.6
EgoEgo128207.8±6.9187.8±6.80.88±0.0366.5±5.4
VAE+Opt128319.8±10.1323.8±10.50.75±0.04274.4±17.6
+ +Aria Digital Twins [61] + +
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\( T_{head} \)↓
EgoAllo32173.5±1.1146.1±1.10.88±0.01-
NoShape32178.5±1.1153.0±1.10.89±0.01-
EgoEgo32212.5±1.4181.3±1.60.64±0.01-
VAE+Opt32284.9±1.6283.9±1.90.63±0.01-
EgoAllo128155.1±1.6129.3±1.60.94±0.01-
NoShape128163.7±1.6140.0±1.60.96±0.01-
EgoEgo128182.6±2.3153.9±2.60.73±0.02-
VAE+Opt128290.8±3.8282.5±4.40.7±0.02-
+ +Table 2. Body estimation performance, compared against a baseline without shape prediction, EgoEgo [48], and VAE+Opt [74, 102]. We exclude the $\mathbf{T}_{\mathrm{head}}$ metric for ADT because the Biomech57 head joints used by ADT are not directly comparable to the SMPL-H head joints used by our model. + +Finally, we compare against an approach based on the SLAHMR [102] framework for human motion estimation from exocentric video. A key advantage of SLAHMR is that it uses an unconditional motion prior [74] in an optimization framework. It can therefore be adapted to new settings without re-training—we keep the same body pose and shape variables as the original pipeline, but replace the exocentric keypoint [71] cost with an egocentric CPF pose alignment cost. + +Due to differences in problem formulation, many existing methods for egocentric human motion estimation are difficult to directly compare. This is particularly true when they have different inputs, such as fisheye cameras [91, 95, 96], wrist-mounted sensors [46], or handheld controller poses [7, 28, 29]. Additionally, prior works like EgoEgo [48] do not incorporate vision inputs for hand estimation. For fairness, we restrict all methods in this section to only CPF or head pose as input. + +EgoAllo improves body motion estimates. We report metrics + +![](images/a1dd517f31a214f45b04cf06f7e824d6a01737aa765229080bec506d6b7e7b73.jpg) +(a) Ground-truth + +![](images/2da2e150065507518628a2247b7d7d78001b2fc6b0c5a050d9b3c92a6a194ae0.jpg) +(b) EgoAllo + +![](images/30685c4828fe440c3df429c5eb042a642232ee2915171b997b860880d7c3bbab.jpg) +(c) EgoEgo [48] + +![](images/7fc72bfc266b0a35cc57f96ad7ae2d72897133c0c504bdd32fdd0f350ef3dbbf.jpg) +(d) VAE+Opt +Figure 4. Egocentric human motion estimation for a running sequence. We show the ground-truth, an output from EgoAllo, and outputs from two baselines. The glasses CAD model is placed at the conditioning transformation $\mathbf{T}_{\mathrm{world,cpf}}$ . + +in Table 2 and visualize example outputs in Figure 4. We find that EgoAllo enables significant estimation improvements across all datasets, including accuracy improvements of $20\sim 30\%$ over EgoEgo for both shorter and longer evaluation sequences. We found shape estimation critical for producing metric-scale, grounded estimates of human body motion, with the head aligned to input SLAM poses and the feet planted on the observed ground plane. This is evident in qualitative results, improved grounding metrics, and in the $6\sim 7\%$ MPJPE gap between EgoAllo and the NoShape ablation. + +VAE optimization converges poorly. Optimization-based estimation approaches have been effective for settings with keypoint costs [74, 102], but we found convergence difficult in our less constrained setting. In Table 2, we observe poor generalization: VAE+Opt performs competitively on the AMASS test set, but performance deteriorates dramatically when evaluating on RICH or ADT. VAE+Opt outputs in Figure 4 also look overly smoothed, without the same expressiveness as the conditional predictions of EgoAllo or EgoEgo [48]. This highlights the advantage of using a conditional diffusion model problem for this estimation problem. + +Shape estimation evaluation. To better understand the shape estimation characteristics of EgoAllo, we compare against against the "mean" shape used by EgoEgo and the NoShape ablation. On the AMASS test set, we find: EgoAllo slightly improves overall shape (19mm → 18mm mean vertex-to-vertex error) and produces much better height (52mm → 32mm mean height error), but is not able to generalize in terms of body weight (5kg → 8kg mean weight error). The body shape is inferred from the wearer's head pose, which intuitively provides strong height constraints but is less correlated with weight. Accurate height is key for proper scene placement, as reflected by both the MPJPE and GND metrics. + +# 4.2. Hand estimation + +To evaluate hands estimated by EgoAllo, we run HaMeR on the segment of the EgoExo4D [17] validation set that is labeled with 3D hand pose keypoints. We quantitatively compare four hand estimation methods in Table 3. In (1) HaMeR [66], we use HaMeR out-of-the-box on undistorted egocentric RGB images. We do not assume bounding boxes as input; instead, we follow the HaMeR demo code and compute crops using ViTPose [99]. (2) EgoAllo-NoReproj uses all loss terms except for the reprojection loss (Equation 14). Hand guidance is done directly using the 3D wrist poses predicted by HaMeR. (3) EgoAllo-Mono is the same as EgoAllo-NoReproj, but guides hands using the reprojection loss. This accounts for the scale ambiguities that are inherent to the single-frame HaMeR estimates. Finally, we show an extension of our method. (4) EgoAllo-Wrist3D extends EgoAllo-Mono with 3D wrist pose losses from Project Aria's Machine Perception Services [85]—unlike HaMeR, which assumes monocular input, this is a separate sensing setup that also leverages a pair of SLAM cameras that are unique to Project Aria. For fairness across settings, we compute metrics only on timesteps where HaMeR estimates are available. + +Results are provided in Table 3. While HaMeR's local poses (PA-MPJPE) are slightly better, EgoAllo's hand-body estimation significantly improves how well hands are estimated in the world coordinate system. Compared to HaMeR, EgoAllo-Mono drops MPJPE from $237.90\mathrm{mm} \rightarrow 131.45\mathrm{mm}$ . Reprojection-based guidance is important: despite using the same inputs, EgoAllo-NoReproj outputs are worse than EgoAllo-Mono in both MPJPE and PA-MPJPE. Finally, incorporating more accurate wrist pose estimates from the extra cameras on Project Aria drops errors $131.45\mathrm{mm} \rightarrow 60.08\mathrm{mm}$ . This highlights both room for improvement for EgoAllo and how EgoAllo's guidance-based formulation makes it easy to extend when more sensors are available. + +Qualitatively, we observed that high hand estimation errors in naive monocular estimation with HaMeR are explained by a combination of detection failures and monocular ambiguities. Even when detections succeed, the scale and distance of monocular HaMeR estimates are often incorrect or flicker in between frames. EgoAllo improves on this by applying kinematic constraints and a diffusion-based prior for plausible body motion—we provide examples of HaMeR estimates rendered jointly with EgoAllo outputs in Figure 5. + +# 5. Dicussion + +Limitations and future work. While the core contributions of EgoAllo are general, the current implementation of our system has a few limitations that we hope to explore in future work. First, diffusion model guidance is a test-time optimization process that depends on hyperparameters and incurs a runtime cost. In the future, it may be possible to bootstrap using outputs from our model to train a feedforward model that avoids this step. Success for hand guidance also still depends on reasonable + +![](images/b5d0e6e5c24724b83f082f64773dd82388892094b92a20c923e29805a629d954.jpg) + +![](images/6dc08720293e3599f6e0283137f1207700143f7389ba4d6db44da6d7260963e9.jpg) + +![](images/e05d4df87330c2592576531aae940510d7869503b762d80c542d64b9bfc57259.jpg) + +![](images/067d87f430e88fcaad7a74dbafb94eb9548a93818e24efba0428cfbdbf8e9c4e.jpg) + +![](images/ff32739cb8fa7d88e1a63cd5adb7c329d98bfbbad622b0567df87878edef0758.jpg) +Figure 5. Body estimation improves hand estimation. We show raw outputs from HaMeR [66] in blue and hand-body estimations from EgoAllo in purple. Top: improved scene interaction during touchscreen operation with EgoAllo-Mono. We know a priori that the fingers are contacting the screen in this sequence. Bottom: qualitative examples from EgoExo [17] evaluation, showing the differences between monocular hands and EgoAllo-Wrist3D estimates. + +![](images/fb2139961d402a1d8399156aa9d44ba260dd3b487a9f1c4d3a3e7338d5ad9e9a.jpg) + +![](images/556860c8857b3de77f3dee4eca39873535ca320e08a5bbd1ad93cdcd1e448f26.jpg) + +
MethodMPJPE ↓PA-MPJPE ↓
HaMeR237.90±1.8913.04±1.89
EgoAllo-NoReproj143.20±0.4214.75±0.42
EgoAllo-Mono131.45±0.3914.71±0.39
EgoAllo-Wrist3D60.08±0.2614.38±0.26
+ +Table 3. Hand estimation errors in millimeters. EgoAllo's hand-body estimation can constrain and resolve ambiguities in noisy outputs from HaMeR, which we observe can reduce MPJPE for hands by over $40\%$ . EgoAllo-Wrist3D incorporates 3D wrist poses provided from the grayscale cameras on Project Aria. + +monocular hand estimates. Errors like left/right flipping or spurious detections can therefore cause estimation failure. Finally, we assume flat floors. This is in part because our training data [56] includes floor planes but no detailed scene geometry. As a result, our method will fail in settings like hills or staircases. In the future, we hope to extend our insights to data with more detailed scene information, which concurrent work has highlighted the usefulness of in informing human body estimation [21]. + +Conclusion. We presented EgoAllo, a system for estimating human motion using a head-mounted device. EgoAllo jointly estimates human body pose, height, and hand parameters from only egocentric SLAM poses and images. Results highlight the importance of spatial and temporal invariance in conditioning for this problem, while demonstrating how estimated bodies can be used to improve hand estimation. + +Acknowledgments. This project was funded in part by NSF:CNS-2235013 and IARPA DOI/IBC No. 140D0423C0035. JM was supported by ONR MURI N00014-21-1-2801. YM acknowledges support from the joint Simons Foundation-NSF DMS grant #2031899, the ONR grant N00014-22-1-2102, the NSF grant #2402951, and partial support from TBSI, InnoHK, and the University of Hong Kong. BY is supported by the National Science Foundation Graduate Research Fellowship Program under Grant DGE 2146752. The authors would also like to thank Hongsuk Choi, Michael Taylor, Tyler Bonnen, Songwei Ge, Chung Min Kim, and Justin Kerr for insightful technical discussion and suggestions, as well as Jiaman Li for helpful answers to questions about EgoEgo. + +# References + +[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2 +[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. ACM New York, NY, USA, 2005. 2 +[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 5, 1 +[4] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 5 +[5] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 3 +[6] Marcus A Brubaker, David J Fleet, and Aaron Hertzmann. Physics-based person tracking using the anthropomorphic walker. International journal of computer vision, 87(1-2): 140-155, 2010. 2 +[7] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4221-4231, 2023, 1, 2, 3, 4, 5, 6, 7 +[8] R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE conference on computer vision and pattern recognition (CVPR), pages 77-85. IEEE, 2017. 3 +[9] Haiwei Chen, Shichen Liu, Weikai Chen, Hao Li, and Randall Hill. Equivariant point network for 3d point cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14514-14523, 2021. 3 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning + +of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3 +[11] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4800-4810, 2023. 2, 5 +[12] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[14] Haiwen Feng, Peter Kulits, Shichen Liu, Michael J Black, and Victoria Fernandez Abrevaya. Generalizing neural human fitting to unseen poses with articulated se (3) equivariance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7977-7988, 2023. 3 +[15] Saeed Ghorbani, Calden Wloka, Ali Etemad, Marcus A Brubaker, and Nikolaus F Troje. Probabilistic character motion synthesis using a hierarchical deep latent variable model. In Computer Graphics Forum, pages 225-239. Wiley Online Library, 2020. 2 +[16] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In Int. Conf. Comput. Vis., 2023. 2 +[17] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 6, 8, 3 +[18] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. arXiv preprint arXiv:2403.18118, 2024. 1 +[19] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In Int. Conf. Comput. Vis., pages 1381-1388, IEEE, 2009. 2 +[20] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10884-10894, 2019. 2 +[21] Vladimir Guzov, Yifeng Jiang, Fangzhou Hong, Gerard Pons-Moll, Richard Newcombe, C Karen Liu, Yuting Ye, and Lingni Ma. Hmd: Environment-aware motion generation from single egocentric head-mounted device. arXiv preprint arXiv:2409.13426, 2024. 2, 8, 1 +[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5 +[24] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2 + +[25] Nicholas Howe, Michael Leventon, and William Freeman. Bayesian reconstruction of 3d human motion from single-camera video. Advances in neural information processing systems, 12, 1999. 2 +[26] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13274-13285, 2022. 6, 7, 3 +[27] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2 +[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 2, 3, 4, 5, 6, 7 +[29] Jiaxi Jiang, Paul Streli, Manuel Meier, Andreas Fender, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2, 3, 4, 5, 7, 1 +[30] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6142-6152, 2024. 2, 5 +[31] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards in-the-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2 +[32] Angjoo Kanazawa, Abhishek Sharma, and David Jacobs. Locally scale-invariant convolutional neural networks. arXiv preprint arXiv:1412.5104, 2014. 3 +[33] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7122-7131, 2018. 2 +[34] Angjoo Kanazawa, Jason Y Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5614-5623, 2019. 2 +[35] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 2, 5 +[36] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2 +[37] Meejin Kim and Sukwon Lee. Fusion poser: 3d human pose estimation using sparse imus and head trackers in real time. Sensors, 22(13):4846, 2022. 2 +[38] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape + +estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5253-5263, 2020. 2 +[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021. 2 +[41] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2 +[42] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 2 +[43] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023. 2 +[44] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6050-6059, 2017. 2 +[45] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 3 +[46] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. arXiv preprint arXiv:2401.00847, 2024. 2, 7 +[47] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 2, 5 +[48] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17142-17151, 2023. 1, 2, 3, 4, 5, 6, 7 +[49] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2 +[50] Zongmian Li, Jiri Sedlar, Justin Carpentier, Ivan Laptev, Nicolas Mansard, and Josef Sivic. Estimating 3d motion and forces of person-object interactions from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8649, 2019. 2 +[51] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 2 +[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi-person linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851-866, 2023. 2, 3 +[53] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3 +[54] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2 + +[55] Lingni Ma, Yuting Ye, Fangzhou Hong, Vladimir Guzov, Yifeng Jiang, Rowan Postyeni, Luis Pesqueira, Alexander Gamino, Vijay Baiyya, Hyo Jin Kim, et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. arXiv preprint arXiv:2406.09905, 2024. 2 +[56] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5442-5451, 2019. 3, 4, 6, 7, 8 +[57] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Ecoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1 +[58] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Int. Conf. Comput. Vis., pages 2640-2649, 2017. 2 +[59] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2 +[60] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In 2018 international conference on 3D vision (3DV), pages 484-494. IEEE, 2018. 2 +[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 1, 6, 7, 3 +[62] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 459-468, 2018. 2 +[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10975-10985, 2019. 2 +[64] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 2 +[65] Georgios Pavlakos, Ethan Weber, Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In Eur. Conf. Comput. Vis., pages 732-749. Springer, 2022. 2 +[66] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In *arxiv*, 2023. 3, 8 +[67] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. 2 +[68] Chiara Plizzari, Shubham Goel, Toby Perrett, Jacob Chalk, Angjoo Kanazawa, and Dima Damen. Spatial cognition from + +egocentric video: Out of sight, not out of mind. arXiv preprint arXiv:2404.05072, 2024. 1 +[69] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2 +[70] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[71] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2740-2749, 2022. 2, 7 +[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 5 +[73] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 2 +[74] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 4, 5, 7, 1 +[75] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 2 +[76] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2 +[77] Gregory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net: Localization-classification-regression for human pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3433-3441, 2017. 2 +[78] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[79] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2, 3, 5 +[80] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 + +[81] Silvio Savarese and Li Fei-Fei. 3d generic object categorization, localization and pose estimation. In 2007 IEEE 11th international conference on computer vision, pages 1-8. IEEE, 2007. 3 +[82] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 5 +[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K Hodgins. Motion capture from body-mounted cameras. In ACM SIGGRAPH 2011 papers, pages 1-10. ACM New York, NY, USA, 2011. 2 +[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[85] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2, 3, 4, 5, 6, 8 +[86] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 3 +[87] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[88] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2 +[89] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2 +[90] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3 +[91] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7728-7738, 2019. 2, 7 +[92] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2 +[93] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2 +[94] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 2 +[95] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11500-11509, 2021. 2, 7 + +[96] Jian Wang, Zhe Cao, Diogo Luvizon, Lingjie Liu, Kripasindhu Sarkar, Danhang Tang, Thabo Beeler, and Christian Theobalt. Egocentric whole-body motion capture with fisheyevit and diffusion-based motion refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 777-787, 2024. 2, 7 +[97] Laurenz Wiskott and Terrence J Sejnowski. Slow feature analysis: Unsupervised learning of invariances. Neural computation, 14(4):715-770, 2002. 3 +[98] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 2 +[99] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246, 2022.8 +[100] Jingyun Yang, Congyue Deng, Jimmy Wu, Rika Antonova, Leonidas Guibas, and Jeannette Bohg. Equivact: Sim(3)-equivariant visuomotor policies beyond rigid object manipulation, 2023. 3 +[101] Jingyun Yang, Zi-ang Cao, Congyue Deng, Rika Antonova, Shuran Song, and Jeannette Bohg. Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning. arXiv preprint arXiv:2407.01479, 2024. 3 +[102] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21222-21232, 2023. 2, 5, 7 +[103] Brent Yi, Weijia Zeng, Sam Buchanan, and Yi Ma. Canonical factors for hybrid neural fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3414-3426, 2023. 3 +[104] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2 +[105] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. arXiv preprint arXiv:2305.01599, 2023. 2 +[106] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2 +[107] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2 +[108] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[109] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems, 30, 2017. 3 + +[110] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. arXiv preprint arXiv:2305.03462, 2023. 3 +[111] Daiwei Zhang, Gengyan Li, Jiajie Li, Mickaël Bressieux, Otmar Hilliges, Marc Pollefeys, Luc Van Gool, and Xi Wang. Egogaussian: Dynamic scene understanding from egocentric video with 3d gaussian splatting. arXiv preprint arXiv:2406.19811, 2024. 1 +[112] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2 +[113] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 5 +[114] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 2, 5 +[115] Zhengdong Zhang, Arvind Ganesh, Xiao Liang, and Yi Ma. Tilt: Transform invariant low-rank textures. International journal of computer vision, 99:1-24, 2012. 3 \ No newline at end of file diff --git a/2025/Estimating Body and Hand Motion in an Ego-sensed World/images.zip b/2025/Estimating Body and Hand Motion in an Ego-sensed World/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..de23371fc940924eb5670732224b600eb71788b3 --- /dev/null +++ b/2025/Estimating Body and Hand Motion in an Ego-sensed World/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b02a21f2f5cca4d236352f2fc1794d75b172baca72133c78c4f1bf15cda588b1 +size 607723 diff --git a/2025/Estimating Body and Hand Motion in an Ego-sensed World/layout.json b/2025/Estimating Body and Hand Motion in an Ego-sensed World/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f98d6f0f5e1f62ec0f9d14d52a17b4d2817eb655 --- /dev/null +++ b/2025/Estimating Body and Hand Motion in an Ego-sensed World/layout.json @@ -0,0 +1,11936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "type": "text", + "content": "Estimating Body and Hand Motion in an Ego-sensed World" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": "Brent Yi" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Vickie Ye" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Maya Zheng" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Yunqi Li" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Lea Müller" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Georgios Pavlakos" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Yi Ma" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Jitendra Malik" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " Angjoo Kanazawa" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": " UC Berkeley " + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": "ShanghaiTech " + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 141, + 135, + 470, + 193 + ], + "type": "text", + "content": "UT Austin" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 210, + 378, + 396 + ], + "blocks": [ + { + "bbox": [ + 57, + 210, + 378, + 396 + ], + "lines": [ + { + "bbox": [ + 57, + 210, + 378, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 210, + 378, + 396 + ], + "type": "image", + "image_path": "736b3a79b52c98a5c596117a87cd1888334df4458f4e6ca710a1c6af83829da1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 403, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. EgoAllo. We present a system that estimates human body pose, height, and hand parameters from egocentric SLAM poses and images. Outputs capture the wearer's actions in the allocentric reference frame of the scene, which we visualize here with 3D reconstructions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 211, + 555, + 396 + ], + "blocks": [ + { + "bbox": [ + 391, + 211, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 391, + 211, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 391, + 211, + 555, + 396 + ], + "type": "image", + "image_path": "8ecba775f6884bd43051b09d83eacbc8f6c2c586bcf2eb6223483ba1e7f9e770.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 445, + 199, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 445, + 199, + 458 + ], + "spans": [ + { + "bbox": [ + 153, + 445, + 199, + 458 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 473, + 297, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 473, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 473, + 297, + 629 + ], + "type": "text", + "content": "We present EgoAllo, a system for human motion estimation from a head-mounted device. Using only egocentric SLAM poses and images, EgoAllo guides sampling from a conditional diffusion model to estimate 3D body pose, height, and hand parameters that capture a device wearer's actions in the allocentric coordinate frame of the scene. To achieve this, our key insight is in representation: we propose spatial and temporal invariance criteria for improving model performance, from which we derive a head motion conditioning parameterization that improves estimation by up to " + }, + { + "bbox": [ + 54, + 473, + 297, + 629 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 54, + 473, + 297, + 629 + ], + "type": "text", + "content": ". We also show how the bodies estimated by our system can improve hand estimation: the resulting kinematic and temporal constraints can reduce world-frame errors in single-frame estimates by " + }, + { + "bbox": [ + 54, + 473, + 297, + 629 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 54, + 473, + 297, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 656, + 132, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 132, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 132, + 668 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Head-mounted devices are becoming increasingly mainstream. In addition to offering new challenges for 3D scene understanding [18, 57, 68, 111], egocentric sensors from these devices are" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 495 + ], + "type": "text", + "content": "unique in that their outputs are coupled to a human wearer's motion in the world. Using these sensors to understand the wearer in addition to the scene around them is essential for applications in augmented reality, robotics, and assistive technologies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 496, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 567 + ], + "type": "text", + "content": "We therefore introduce EgoAllo, a system that uses egocentric inputs to estimate the wearer and their motion in the world, or allocentric, coordinate frame. We take as input sensed metric SLAM head poses and egocentric video from devices like Project Aria [61]. We then output estimates of human body pose, height, and hand motion parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 569, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 556, + 664 + ], + "type": "text", + "content": "This is a difficult task: while body parts like hands occasionally appear in egocentric frames, most body parameters are never directly observed. Estimated body pose, body height, and hand parameters must be consistent with both the sensed egomotion and visual observations. This setting differs from most prior works in egocentric human motion estimation [7, 48], which focus on body pose and do not address the challenges of height and hand motion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Our proposed system uses a head pose-conditioned diffusion model as a motion prior, as well as a Levenberg-Marquardt guidance optimizer for sampling hand-body sequence that align with image observations. Our results are enabled by a key" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7072" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "insight: that the representation used for head pose conditioning is critical for accurate full-body motion estimation. We study choices for this representation by (1) identifying desirable spatial and temporal invariance properties that are not fulfilled by existing systems and (2) using these properties to derive improved parameterizations for our motion prior." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "text", + "content": "We systematically evaluate our system on four datasets. For body estimation, we find that improving the conditioning parameterization leads to an accuracy improvement between " + }, + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "inline_equation", + "content": "17.9\\%" + }, + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "text", + "content": ". Furthermore, we observe that the resulting system can improve hand estimation, reducing world-frame errors by over " + }, + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 144, + 294, + 228 + ], + "type": "text", + "content": " compared to single-frame estimates. Code, model, and more results can be found on our project webpage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 138, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 138, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 138, + 251 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 260, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 295, + 462 + ], + "type": "text", + "content": "3D human recovery from external visual inputs. A large body of work has addressed estimating the parameters of human body models like SCAPE [2] or SMPL and its variants [52, 63, 79] from third-person visual inputs, where human subjects are observed from the view of outside cameras. The majority of these works focus on extracting 3D representations from single images, for example by lifting 2D keypoint observations to 3D [58], via end-to-end regression [20, 31, 33, 40, 60, 62, 77], via optimization [19, 44, 63], or by exploiting synergies between regression and optimization [42]. When multiple frames are available in the form of a video, temporal context and tracking can also be incorporated [16, 34, 39, 64, 65, 71, 108]. The inputs (images) and outputs (human meshes) of many of these systems are similar to the egocentric setting addressed by EgoAllo, but egocentric devices present unique challenges because the body being estimated is typically behind the outwards-facing cameras used as input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 672 + ], + "type": "text", + "content": "Priors for human motion. The primary challenge of ego-sensed human motion estimation is limited observability; a prior is required to resolve ambiguities. For human motion, these priors are typically framed as unconditional distributions over plausible human motion. Distributions can be represented either by modeling the physical constraints of our world [6, 50, 67, 73] or by learning generative models of human motion directly from data. For learning unconditional priors, classical data-driven approaches include fitting mixtures-of-Gaussians to 3D keypoint trajectories [25], while modern approaches include training variational autoencoders [38, 75] to model either autoregressive transitions [15, 51, 74] or full spatiotemporal sequences [22]. After training, these priors can be applied to estimation problems in iterative optimization frameworks [41, 74, 102]. EgoAllo is built on the same intuition as these methods, but follows previous work in ego-sensed motion estimation and uses a task-specific conditional prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Denoising diffusion for human motion. The core of EgoAllo is a denoising diffusion model [23, 69, 84] from which we can sample 3D human body motion. While diffusion mod" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "els are primarily known for their success in text-conditioned image generation [78, 80], they have also enabled advances in human motion synthesis conditioned on modalities like text [35, 36, 112], music [1, 93], poses [35, 48], and object geometry [43, 47, 49]. EgoAllo adopts a similar conditional diffusion approach, while specifically studying the design of conditioning parameters used for ego-sensed human motion estimation. The iterative nature of denoising diffusion also enables guidance [11, 13, 30, 35, 87, 114], where denoising steps are steered to satisfy a desired objective. We use guidance to incorporate observations like visual hand pose observations during test-time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 213, + 556, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 556, + 489 + ], + "type": "text", + "content": "Human motion from egocentric observations. EgoAllo builds on intuition from several prior works in egocentric sensing for human motion estimation. Many rely on fisheye cameras that place the wearer's body into the field of view [27, 76, 91, 91, 92, 95, 96, 98]. Other approaches rely on body-mounted cameras [83], simulation-based physical plausibility [54, 106, 107], body- and hand-mounted inertial sensors [37, 46, 104, 105], handheld controllers [7, 28, 29], and interaction cues from other humans [59]. Concurrent works have also used the Nymeria [55] dataset for egocentric motion with language description outputs [24], as well for online settings with scene geometry and CLIP [70] feature inputs [21]. Most relevantly, EgoEgo [48] demonstrates how human body poses can be estimated offline without body observability assumptions. The authors accomplish this by carefully integrating several components: a monocular SLAM system [89], a pose-conditioned gravity vector regression network, an optical flow feature-conditioned head orientation and scale regression network, and a head pose-conditioned body diffusion model. EgoAllo differs in both inputs—we study conditioning parameters computed from the metric SLAM poses provided by devices like Project Aria [85]—and outputs—we consider body height variation and hand poses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "Conditioning for ego-sensed poses. Prior works vary in how head pose information is parameterized and used as neural network input. AvATARPoser [28] and BoDiffusion [7] parameterize head pose as four components: world-frame orientation, orientation deltas, world-frame position, and world-frame position deltas. These works focus on settings with VR controller input, and parameterize controller pose inputs the same way. EgoEgo [48]'s diffusion model uses only absolute head positions and orientations, but similar to HuMoR [74], in implementation defines a per-sequence canonical coordinate frame to ensure that all input trajectories passed to the model are aligned with the same initial " + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": " position and forward direction. In our work, we refer to this as sequence canonicalization. Finally, EgoPoser [29] proposes a similar scheme that aligns initial positions for both head pose and controller pose inputs. We propose an alternative to these parameterizations that is motivated by the robustness and generalization benefits of invariance, as observed in prior work for designing both" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7073" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 549, + 219 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 549, + 219 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 549, + 219 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 549, + 219 + ], + "type": "image", + "image_path": "db89cba97a9938764e3ab6e3b4031644470fdfb31a685e983ad268b1d6b1905f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Overview of components in EgoAllo. We restrict the diffusion model to local body parameters (Section 3.1.1). An invariant parameterization " + }, + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "type": "text", + "content": " (Section 3.1.2) of SLAM poses is used to condition a diffusion model. These can be placed into the global coordinate frame via global alignment (Section 3.2.1) to input poses. When available, egocentric video is used for hand detection via HaMeR [66], which can be incorporated into samples via guidance (Section 3.2.2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 297, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 295, + 346 + ], + "type": "text", + "content": "representations [10, 53, 81, 97, 103, 110, 115] and neural network architectures [8, 9, 12, 14, 32, 45, 90, 100, 101, 109]. Specifically, we introduce in Section 3.1.2 a parameterization that is invariant to both spatial and temporal shifts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 109, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 109, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 109, + 368 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 376, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 422 + ], + "type": "text", + "content": "We study the problem of using sensors from an egocentric device to estimate the actions of the wearer in an allocentric coordinate frame. We assume a flat floor and two inputs: poses from the device's SLAM system and egocentric video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 423, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 495 + ], + "type": "text", + "content": "Our system uses head pose information to condition a diffusion-based prior over body pose and height, and incorporates visual hand observations during sampling. This allows it to benefit from both 3D human motion capture datasets [56], which are used for the motion prior, and from large-scale image datasets [66], which are used for hand estimates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 234, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 234, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 234, + 514 + ], + "type": "text", + "content": "3.1. Ego-conditioned motion diffusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": "Notation: we use " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{A,B}} = (\\mathbf{R}_{\\mathrm{A,B}},\\mathbf{p}_{\\mathrm{A,B}})" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": " to denote an SE(3) transform to frame A from frame B, composed of rotation " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_{\\mathrm{A,B}})" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": " and position " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_{\\mathrm{A,B}})" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": " terms. Temporal steps " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": " are superscripted and diffusion noise steps " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": " are subscripted. " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\vec{x}_0^t" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": " thus refers to the " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": "-th timestep of a clean " + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "(n = 0)" + }, + { + "bbox": [ + 54, + 520, + 295, + 579 + ], + "type": "text", + "content": " human motion sequence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": "Given an observation window of " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": " timesteps, EgoAllo's motion prior is a diffusion model that aims to capture the distribution of human motions " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\vec{x}_0 = \\{\\vec{x}_0^1,\\dots,\\vec{x}_0^T\\}" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": " conditioned on head pose encodings " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\vec{c} = \\{\\vec{c}^1,\\dots,\\vec{c}^T\\}" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": ". For each timestep " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": ", we represent human motion in the form of SMPL-H [52, 79] model parameters " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}_{\\mathrm{world,root}}^t,\\Theta^t,\\beta \\}" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": ": root transforms " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{world,root}}^t\\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": ", where the person's root frame is located at their pelvis, local joint rotation matrices " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\Theta^t\\in \\mathbb{R}^{51\\times 3\\times 3}" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": ", and time-invariant shape " + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 54, + 580, + 295, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Dependencies between local joint rotations, body size variation, and global motion make this learning task a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 297, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 357 + ], + "type": "text", + "content": "challenging one. Our key insight is that this difficulty can be reduced by designing parameterizations with desirable invariance properties. Spatial and temporal invariances allow the model to focus on the essential structure of motion, without being affected by irrelevant shifts in position or time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 477, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 477, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 477, + 382 + ], + "type": "text", + "content": "3.1.1. Diffusion output representation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "text", + "content": "As output, we sample body and hand joint rotations, body shapes, and binary contact predictions " + }, + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "inline_equation", + "content": "\\vec{x}_0^t = \\{\\Theta^t,\\beta^t,\\psi_{j = 1\\dots 21}^t\\}" + }, + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "text", + "content": ", where body shape " + }, + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "inline_equation", + "content": "\\beta^t" + }, + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "text", + "content": " is supervised to be equal for all timesteps and " + }, + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "inline_equation", + "content": "\\psi_j^t" + }, + { + "bbox": [ + 313, + 387, + 555, + 458 + ], + "type": "text", + "content": " is a per-joint contact indicator. Notably, these parameters are all local—we discuss how outputs can be placed into the allocentric coordinate frame in Section 3.2.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 593 + ], + "type": "text", + "content": "We choose this output set for three main reasons. (1) Body shape encodes the wearer's height, which is critical for grounding in the metric-scale geometry of the scene. This is rarely considered by prior work: with the exception of [29], existing methods [7, 28, 48] otherwise produce outputs using a fixed \"mean\" human shape. (2) Contact predictions enable losses for common problems like foot skating, which are discussed in Section 3.2.2. (3) Finally, local bodies are invariant to the global coordinate frame. As we discuss next, the conditioning parameterization for the model can therefore also be invariant to arbitrary transformations along the floor plane." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 440, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 440, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 440, + 617 + ], + "type": "text", + "content": "3.1.2. Invariant conditioning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 623, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 554, + 658 + ], + "type": "text", + "content": "The goal of our conditioning representation is to map raw SLAM poses (head motion) to a parameterization that is amenable to learning for the diffusion model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Raw inputs. To capture the head motion at each time step, we assume as input poses of a central pupil frame (CPF), which the SLAM systems of devices like Project Aria can provide with millimeter-level accuracy [85]. For time 1...T, we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7074" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": "reparameterize these poses for conditioning using a function " + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 92, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 92, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 102, + 92, + 295, + 107 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} = \\left(\\mathbf {R} _ {\\text {w o r l d , c p f}} ^ {t}, \\mathbf {p} _ {\\text {w o r l d , c p f}} ^ {t}\\right) \\in \\mathrm {S E} (3), \\tag {1}", + "image_path": "ed6a0324baa931f31f46b2386eb072304ae75b0320898f87ac5a736e9c9b810a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 109, + 296, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 109, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 91, + 109, + 296, + 125 + ], + "type": "interline_equation", + "content": "\\{\\vec {c} ^ {1}, \\dots , \\vec {c} ^ {T} \\} = g (\\{\\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {1}, \\dots , \\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {T} \\}). \\tag {2}", + "image_path": "6acee50592289449366938610dfb35e7356744cbb6c24e51eac0ded00f9ddb1b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 132, + 295, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 295, + 190 + ], + "type": "text", + "content": "The CPF frame differs from prior works that condition on a coordinate frame attached to the SMPL human model's \"head joint\" [7, 28, 29, 48]. The offset between this head joint and the device pose depends on the head shape captured by " + }, + { + "bbox": [ + 55, + 132, + 295, + 190 + ], + "type": "inline_equation", + "content": "\\beta^t" + }, + { + "bbox": [ + 55, + 132, + 295, + 190 + ], + "type": "text", + "content": ", and is thus difficult to precompute in our setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": "To encode absolute height, we assume that the world frame's " + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "inline_equation", + "content": "+z" + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": "-axis faces upwards, and that the ground is located at " + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "inline_equation", + "content": "z = 0" + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": ". Ground parameters are directly available in the training data [56]; at test time, we can also extract these parameters from sparse SLAM points via RANSAC (Appendix A.3.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "type": "text", + "content": "Invariance goals. As discussed in Section 2, prior work varies in how the function " + }, + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "type": "text", + "content": " is implemented. To understand how choices impact learning, we propose two invariance properties for head motion representations. Each reduces representational redundancy, which eases the learning problem. We provide more detailed explanations and visualizations of these properties in Appendix A.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "text", + "content": "Invariance 1 (Spatial) Global transformations along the floor plane should not affect a person's local motion. Given " + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{xy} \\in SE(3)" + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "text", + "content": " restricted to the " + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "inline_equation", + "content": "XY" + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "text", + "content": " plane, " + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "text", + "content": " should fulfill " + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "inline_equation", + "content": "g(\\{\\mathbf{T}_{xy}\\mathbf{T}_{world, cpf}^t\\}_t) = g(\\{\\mathbf{T}_{world, cpf}^t\\}_t) \\forall \\mathbf{T}_{xy}" + }, + { + "bbox": [ + 54, + 347, + 296, + 398 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "text", + "content": "Invariance 2 (Temporal) If the same human motion occurs at different points in time, the extracted head motion representation should remain consistent. This can be expressed as temporal shift equivariance. Let " + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\vec{c}^t" + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "text", + "content": " be as defined in Equation 2. For any shift " + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\{\\vec{c}_{shift}^{1},\\dots,\\vec{c}_{shift}^{T}\\} = g(\\{\\mathbf{T}_{world,cpf}^{1 + \\delta},\\dots,\\mathbf{T}_{world,cpf}^{T + \\delta}\\})" + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "text", + "content": " should satisfy " + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\vec{c}_{shift}^{t} = \\vec{c}^{t + \\delta}" + }, + { + "bbox": [ + 55, + 404, + 296, + 479 + ], + "type": "text", + "content": " for overlapping timesteps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "content": "No parameterization used by existing work satisfies both of these properties. The sequence canonicalization approach of EgoEgo [48] achieves spatial invariance (Invariance 1), but inserts a sequence-wide dependency on the first timestep of each window that results in a violation of Invariance 2. The absolute poses and pose deltas used by [7, 28] satisfy Invariance 2, but not Invariance 1. Finally, the relative positions considered by [29] are neither spatially nor temporally invariant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 587, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 647 + ], + "type": "text", + "content": "Invariant conditioning. We propose a formulation for " + }, + { + "bbox": [ + 55, + 587, + 296, + 647 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 587, + 296, + 647 + ], + "type": "text", + "content": " that achieves both invariance properties by locally canonicalizing head motion with respect to the floor at each timestep. We build on the relative motion of the CPF frame at each time " + }, + { + "bbox": [ + 55, + 587, + 296, + 647 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 587, + 296, + 647 + ], + "type": "text", + "content": ", which respects both Invariance 1 and 2:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 654, + 296, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 654, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 108, + 654, + 296, + 671 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {T} _ {\\mathrm {c p f}} ^ {t - 1, t} = \\left(\\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {t - 1}\\right) ^ {- 1} \\mathbf {T} _ {\\mathrm {w o r l d}, \\mathrm {c p f}} ^ {t}. \\tag {3}", + "image_path": "e555820d8c748de69402462833d7f56408af6f86424c5e0aa03400dc6b3d62de.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Importantly, the translation component of this transformation is in the local frame. This is distinct from world-frame position deltas [7, 28, 29], which still violate Invariance 1." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 345, + 60, + 525, + 235 + ], + "blocks": [ + { + "bbox": [ + 345, + 60, + 525, + 235 + ], + "lines": [ + { + "bbox": [ + 345, + 60, + 525, + 235 + ], + "spans": [ + { + "bbox": [ + 345, + 60, + 525, + 235 + ], + "type": "image", + "image_path": "abcbbc04c56a29d9b6148f942092360c06258d94a86b387c001f31a861ac83b5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "text", + "content": "Figure 3. Locally canonicalized coordinate frames. We compute our invariant conditioning parameterization (Equation 4) using transformations computed from three coordinate frames. Following [85], the CPF has the " + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "text", + "content": "-axis forward. Following HuMoR [74], the world and canonical " + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "text", + "content": "-axes point up. Canonical frames are computed by projecting the CPF frame origin to the ground plane, then aligning the canonical " + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "text", + "content": "-axis to the CPF forward direction. The axes are color-coded as " + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "text", + "content": " (red), " + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "text", + "content": " (green), and " + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 243, + 555, + 332 + ], + "type": "text", + "content": " (blue)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "content": "Relative transforms alone do not encode information relative to the scene or floor: full trajectories can even be flipped upside down without impacting " + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}_{\\mathrm{cpf}}^{t - 1,t}" + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "content": ". We therefore propose to ground relative motion to the floor plane with a transformation between the CPF frame and a per-timestep canonical frame, which is computed by projecting the CPF frame to the floor. This encodes head height and orientation. Our full representation then becomes:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 340, + 445, + 555, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 445, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 340, + 445, + 555, + 482 + ], + "type": "interline_equation", + "content": "\\vec {c} ^ {t} = \\underbrace {\\left\\{\\Delta \\mathbf {T} _ {\\mathrm {c p f}} ^ {t - 1 , t} , \\quad \\left(\\mathbf {T} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t}\\right) ^ {- 1} \\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} \\right\\}} _ {\\text {I n v a r i a n t i m p l e m e n t a t i o n o f} g (\\cdot)} \\tag {4}", + "image_path": "9e21e42d9a3ae8505e413736b623c70f714055ffbb2878b6adc8ecfce3e2a9de.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 489, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 554, + 539 + ], + "type": "text", + "content": "We visualize an example of a canonical frame in Figure 3 and our full representation in Appendix A.1. Canonical frames are positioned by projecting the CPF origin to the floor plane; given standard bases " + }, + { + "bbox": [ + 313, + 489, + 554, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\{x,y,z\\}}" + }, + { + "bbox": [ + 313, + 489, + 554, + 539 + ], + "type": "text", + "content": ", we compute:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 355, + 546, + 555, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 546, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 355, + 546, + 555, + 564 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t} = \\left[ \\begin{array}{l l l} \\mathbf {e} _ {x} & \\mathbf {e} _ {y} & \\vec {0} \\end{array} \\right] ^ {\\top} \\mathbf {p} _ {\\text {w o r l d , c p f}} ^ {t}. \\tag {5}", + "image_path": "9f91e9c2ecf5b662cb939ec8e239b56bd8bc75324cc6e9fdc7bdb81870184b38.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": "For orientation, we align the canonical frame's local " + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": "-axis parallel to the world " + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": "-axis and its local " + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": "-axis toward the \"forward\" direction " + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\vec{v}^t" + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": " of the CPF frame. With " + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{z}(\\cdot):\\mathbb{R}\\to \\mathrm{SO}(3)" + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": " constructing a " + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": "-axis rotation and " + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\{x,y,z\\}}" + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": " again as standard bases, we compute this as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 387, + 639, + 555, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 639, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 387, + 639, + 555, + 653 + ], + "type": "interline_equation", + "content": "\\vec {v} ^ {t} = \\mathbf {R} _ {\\text {w o r l d , c p f}} ^ {t} \\mathbf {e} _ {z}, \\tag {6}", + "image_path": "d0d36daa1182df2f862e0ba6d94380c923c86c86118fc65c98232347ac10e7a8.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 655, + 555, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 655, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 345, + 655, + 555, + 671 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {\\text {w o r l d , c a n o n i c a l}} ^ {t} = \\mathbf {R} _ {z} \\left(- \\arctan 2 \\left(\\mathbf {e} _ {x} ^ {\\top} \\vec {v} ^ {t}, \\mathbf {e} _ {y} ^ {\\top} \\vec {v} ^ {t}\\right)\\right). \\tag {7}", + "image_path": "7065abbb732472dfe5533858382155a25ad78521306021f0ff062cf987c8eeff.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "This canonical frame definition is an important departure from prior work. While EgoEgo [48] and HuMoR [74] use similar canonical frames, they only compute one per sequence. Instead," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7075" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "we compute Equations 5 and 7 at every timestep. This enables floor plane grounding without sacrificing Invariance 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": "Connection to prior work. EgoAllo's invariant conditioning shares similarities with the spatial and temporal normalization proposed by EgoPoser [29]. (1) EgoPoser's temporal normalization subtracts positions by the first timestep's position. This is a position-only version EgoEgo [48]'s sequence canonicalization. It does not consider rotation, and therefore only partially fulfills Invariance 1. Like EgoEgo, it also does not fulfill Invariance 2. In Appendix A.2.3, we observe that this approach only has a small impact when adapted to EgoAllo's problem setting. (2) EgoPoser's spatial normalization subtracts the head's XY position from the hand pose inputs. EgoAllo does not use hand pose as network input, so this is not applicable in our setting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 192, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 192, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 192, + 261 + ], + "type": "text", + "content": "3.2. Estimation via sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "type": "text", + "content": "We use our local body representation and invariant conditioning strategies to train a motion prior in the form of a denoising diffusion model [23]. Given diffusion step " + }, + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "type": "inline_equation", + "content": "n = N \\ldots 1" + }, + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "type": "text", + "content": ", we follow [72] and approximate the denoising process as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 322, + 296, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 322, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 100, + 322, + 296, + 338 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\vec {x} _ {n - 1} | \\vec {x} _ {n}, \\vec {c}) = \\mathcal {N} \\big (\\mu_ {\\theta} (\\vec {x} _ {n}, n, \\vec {c}), \\sigma_ {n} ^ {2} \\mathbf {I} \\big), \\qquad (8)", + "image_path": "596e3b80f813a56c60bb201a7d4f269d678962887cf2b9e27f79f35a6c1ee5a7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": "where a transformer [94] " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": " is trained to predict the posterior mean from noised sample " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\vec{x}_n" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": " and conditioning " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\vec{c}" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": ". With noise-dependent weight term " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": ", the loss can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 95, + 389, + 296, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 389, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 95, + 389, + 296, + 407 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\vec {x} _ {0}} \\mathbb {E} _ {n \\sim \\mathcal {U}} \\left[ w _ {n} \\| \\mu_ {\\theta} (\\vec {x} _ {n}, n, \\vec {c}) - \\vec {x} _ {0} \\| ^ {2} \\right]. \\tag {9}", + "image_path": "3e24ae5411badb67c31675cf561f4db3d69aaf8e29ec02023aa7258bca9e3ab3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 416, + 296, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 488 + ], + "type": "text", + "content": "After training, we estimate human motions by following DDIM [86] for sampling. The final EgoAllo sampling procedure includes several additional components: a global alignment phase, guidance losses for physical constraints and visual hand observations, and a path fusion [3] approach for longer sequence lengths. We describe these below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 157, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 157, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 157, + 506 + ], + "type": "text", + "content": "3.2.1. Global alignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 509, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 533 + ], + "type": "text", + "content": "To place sampled bodies into the allocentric coordinate system, we compute the absolute pose of the SMPL-H root as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 542, + 296, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 542, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 116, + 542, + 296, + 561 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {\\text {w o r l d , r o o t}} ^ {t} = \\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t} \\mathbf {T} _ {\\text {c p f , r o o t}} ^ {(\\Theta^ {t}, \\beta^ {t})}, \\tag {10}", + "image_path": "20e75a3f26c5333c910c66988964358ae8d7b5ae3f0a2ddade8556a528a0b515.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 569, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{cpf,root}}^{(\\Theta^t,\\beta^t)}" + }, + { + "bbox": [ + 55, + 569, + 296, + 644 + ], + "type": "text", + "content": " computes the transform between the root of the human and their CPF frame for a given set of local pose and shape parameters. Similar processes are applied in [7, 28, 29]. In contrast to directly outputting absolute body transformations from the diffusion model [48], this guarantees exact alignment between estimates and input SLAM sequences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 651, + 151, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 151, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 151, + 661 + ], + "type": "text", + "content": "3.2.2. Guidance losses" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Our diffusion model learns a distribution of human motion conditioned on the central pupil frame motion. At test time, we incorporate constraints from physical priors and visual hand observations via guidance [11, 30, 114]. Similar to [35, 47]," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "we accomplish this by applying costs to the joint rotations " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\Theta^1,\\dots,\\Theta^T\\}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " predicted by " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\vec{x}_n,n,\\vec{c})" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". We treat the body shape " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\beta^{t}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and contacts " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\psi_{j = 1\\dots 21}^{t}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " as fixed and optimize over body and finger pose to minimize hand observation, skating, and prior costs with a Levenberg-Marquardt optimizer:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 374, + 137, + 555, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 137, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 374, + 137, + 555, + 155 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\text {g u i d a n c e}} ^ {(\\Theta)} = \\mathcal {E} _ {\\text {h a n d s}} ^ {(\\Theta)} + \\mathcal {E} _ {\\text {s k a t e}} ^ {(\\Theta)} + \\mathcal {E} _ {\\text {p r i o r}} ^ {(\\Theta)}. \\tag {11}", + "image_path": "8a0cc0191e4b9e6f7aee57c7484a09c6e66c23a8a9bb43a747bb65f558d24b2a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "text", + "content": "We begin by running HaMeR on the egocentric image corresponding to each timestep " + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "text", + "content": ". When detected, this produces 3D hand estimates in the form of MANO [79] joint parameters and camera-centric 3D hand keypoints " + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{\\mathrm{camera},j}^{t}" + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "text", + "content": " for hand joint set " + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "inline_equation", + "content": "j \\in \\mathcal{H}" + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "text", + "content": ". Optionally, wrist and palm poses can also be estimated using Project Aria's Machine Perception Services [85]. With each subscripted " + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 158, + 554, + 242 + ], + "type": "text", + "content": " indicating a scalar weighting term, we have:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 247, + 555, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 247, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 360, + 247, + 555, + 265 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\text {h a n d s}} ^ {(\\Theta)} = \\lambda_ {\\text {h a n d s 3 D}} \\mathcal {E} _ {\\text {h a n d s 3 D}} ^ {(\\Theta)} + \\lambda_ {\\text {r e p r o j}} \\mathcal {E} _ {\\text {r e p r o j}} ^ {(\\Theta)}. \\tag {12}", + "image_path": "63ae244bc255636859bead5d60e8e0948722b95df56c120485bc5f67ff049fcb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": "The 3D objective " + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{hands3D}}^{(\\Theta)}" + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": " minimizes the distance between the detected hand parameters and the corresponding SMPL-H hand parameters, in terms of wrist pose and local joint rotations. With " + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\Pi_K" + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": " as projection with camera intrinsics " + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{world},j}^{(\\Theta^t)} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": " as the world position for joint " + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{camera,cpf}}" + }, + { + "bbox": [ + 313, + 271, + 555, + 346 + ], + "type": "text", + "content": " from the device calibration, the reprojection cost is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 337, + 352, + 555, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 352, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 337, + 352, + 555, + 379 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\text {r e p r o j}} ^ {(\\Theta)} = \\sum_ {t, j \\in \\mathcal {H}} | | \\Pi_ {K} \\left(\\mathbf {p} _ {\\text {c a m e r a}, j} ^ {(\\Theta^ {t})}\\right) - \\Pi_ {K} \\left(\\hat {\\mathbf {p}} _ {\\text {c a m e r a}, j} ^ {t}\\right) | | _ {2} ^ {2}, \\tag {13}", + "image_path": "d6f280ad2dd2061896025deb31e1e681553cec7f9caea6d1d21605600b55236a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 328, + 381, + 555, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 381, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 328, + 381, + 555, + 399 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {\\text {c a m e r a}, j} ^ {(\\Theta^ {t})} = \\mathbf {T} _ {\\text {c a m e r a , c p f}} \\left(\\mathbf {T} _ {\\text {w o r l d , c p f}} ^ {t}\\right) ^ {- 1} \\mathbf {p} _ {\\text {w o r l d}, j} ^ {(\\Theta^ {t})}. \\tag {14}", + "image_path": "0375326933d377bc74aadc1ba2890dd9bb3b9e2f9faf51c596933ea66dd57416.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 403, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 428 + ], + "type": "text", + "content": "To reduce foot skating, we use contact predictions to apply a skating cost [74, 102] for each time " + }, + { + "bbox": [ + 313, + 403, + 554, + 428 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 403, + 554, + 428 + ], + "type": "text", + "content": " and joint " + }, + { + "bbox": [ + 313, + 403, + 554, + 428 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 403, + 554, + 428 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 433, + 554, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 433, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 318, + 433, + 554, + 461 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\text {s k a t e}} ^ {(\\Theta)} = \\sum_ {t, j} \\lambda_ {\\text {s k a t e}} \\| \\frac {1}{2} \\left(\\psi_ {j} ^ {t} + \\psi_ {j} ^ {t - 1}\\right) \\left(\\mathbf {p} _ {\\text {w o r l d}, j} ^ {t} - \\mathbf {p} _ {\\text {w o r l d}, j} ^ {t - 1}\\right) \\| _ {2} ^ {2}. \\tag {15}", + "image_path": "48b857bc4ff6b5e92d81ede7306f3319971796441e54b193344efc8c8b96ce4a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "text", + "content": "Finally, we minimize a prior cost " + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{prior}}^{(\\Theta)}" + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "text", + "content": ". This cost penalizes deviations between joint rotations " + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "text", + "content": " and rotations " + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta}^t" + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "text", + "content": " from the denoiser " + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\vec{x}_n,n,\\vec{c})" + }, + { + "bbox": [ + 313, + 468, + 554, + 521 + ], + "type": "text", + "content": ", while also encouraging smoothness. We detail the full list of terms used for this in Appendix A.3.4." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 525, + 471, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 471, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 471, + 537 + ], + "type": "text", + "content": "3.2.3. Sequence length extrapolation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "text", + "content": "For longer sequences at test time, we draw on existing methods in compositional generation for both image [3, 113] and human motion [4, 82] diffusion models. We train our motion prior using subsequences of up to length 128; when input observations exceed this length at test time, we split into windows with a 32-timestep overlap between neighbors. We then run our model " + }, + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\vec{x}_n,\\vec{c},n)" + }, + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "text", + "content": " on windows in parallel. Diffusion paths for overlapping regions are fused following MultiDiffusion [3] after each denoising step." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 658, + 391, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 391, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 391, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We conduct a series of experiments to evaluate EgoAllo's conditioning parameterization, body estimation accuracy, and hand estimation performance." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7076" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 553, + 215 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 553, + 215 + ], + "type": "table", + "html": "
ConditioningSeqlenInvariance 1/2MPJPE ↓% DiffPA-MPJPE ↓% DiffGND ↑
EgoAllo (Eq. 4)32✓ /✓129.8±1.1—109.8±1.1—0.98±0.00
Absolute+Local Relative32P /✓133.0±1.12.4%113.6±1.23.4%0.95±0.00
Absolute+Global Deltas [7,28]32X /✓136.2±1.14.9%118.3±1.27.7%0.93±0.01
Sequence Canonicalization [48]32✓ /X153.1±1.517.9%128.7±1.517.1%0.76±0.01
Absolute32X /✓159.9±1.223.2%141.0±1.328.4%0.89±0.01
EgoAllo (Eq. 4)128✓ /✓119.7±1.3—101.1±1.3—1.00±0.00
Absolute+Local Relative128P /✓124.5±1.34.0%104.9±1.43.8%1.00±0.00
Absolute+Global Deltas [7,28]128X /✓127.4±1.36.4%109.8±1.48.6%0.99±0.00
Sequence Canonicalization [48]128✓ /X134.0±1.811.9%112.1±1.610.9%0.88±0.02
Absolute128X /✓148.3±1.523.9%131.2±1.629.8%0.96±0.01
", + "image_path": "751f316ffcaf71ddb2efe96f8514124c1006a53a3368c46a471b2e8acce43079.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "text", + "content": "Table 1. Motion prior conditioning comparison. We train and evaluate otherwise identical models using four conditioning parameterizations on AMASS [56] test set sequences, using sequences of length 32 and 128. Parameterizations vary in their spatial (1) and temporal (2) invariance properties, which we loosely classify as following completely (" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "text", + "content": "), partially (" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathsf{P}" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "text", + "content": "), or not at all (" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "text", + "content": "). The conditioning parameterization used by EgoAllo reduces errors by almost " + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "text", + "content": " compared to the sequence canonicalization approach used by the most relevant related work [48]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 282, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 282, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 295, + 378 + ], + "type": "text", + "content": "Training. To train EgoAllo models used in our experiments, we need sequences containing human body and hand pose parameters, body shapes, and device SLAM poses " + }, + { + "bbox": [ + 54, + 282, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{world,cpf}}^t" + }, + { + "bbox": [ + 54, + 282, + 295, + 378 + ], + "type": "text", + "content": ". Similar to prior work [7, 28, 48], we train EgoAllo using AMASS [56] with synthesized device poses. We annotate train split sequences by anchoring a central pupil frame between vertices corresponding to the left and right pupils in the blend skinned mesh, and at train time sample sequences between length 32 and 128." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 385, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 295, + 482 + ], + "type": "text", + "content": "Evaluation. We evaluate with four datasets. We use AMASS [56], RICH [26], and Aria Digital Twins (ADT) [61] for body estimation evaluation, and EgoExo4D [17] for hand estimation evaluation. AMASS and RICH do not include egocentric data; we annotate these with synthetic device poses using the same procedure we use for training. ADT and EgoExo4D both include egocentric images and SLAM poses captured using Project Aria glasses [85], which we use directly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 488, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 488, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 54, + 488, + 295, + 620 + ], + "type": "text", + "content": "Metrics. To quantify performance, we report four metrics: (1) MPJPE is a world-frame mean per-joint position error (millimeters). (2) PA-MPJPE is the Procrustes-aligned mean per-joint position error in millimeters, where joint positions are aligned on a per-timestep basis before error are computed. (3) GND is a grounding metric, designed in response to a phenomena where ego-sensed humans \"float\" above the ground. Given a human body trajectory, this metric contains a simple binary indicator of whether the feet of the human ever touch the ground plane. (4) " + }, + { + "bbox": [ + 54, + 488, + 295, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{head}}" + }, + { + "bbox": [ + 54, + 488, + 295, + 620 + ], + "type": "text", + "content": " is the average SMPL head joint position error in millimeters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 634, + 154, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 154, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 154, + 647 + ], + "type": "text", + "content": "4.1. Body estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "type": "text", + "content": "In our first set of experiments, we evaluate body estimation from only device SLAM poses, without considering images or hands. This setting allows us to isolate the advantages of our body motion prior, while directly comparing against methods that do not consider hands." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 282, + 485, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 485, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 485, + 295 + ], + "type": "text", + "content": "4.1.1. Invariant conditioning evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": "We begin by evaluating the importance of the spatial and temporal invariance criteria discussed in Section 3.1.2. We do this by comparing five implementations of the conditioning " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ": (1) EgoAllo is the final invariant representation that we propose in Equation 4. (2) Absolute+Local Relative appends absolute poses with the relative pose deltas written in Equation 3. (3) Absolute+Global Deltas appends absolute poses with relative orientation and the world-frame position deltas used by [7, 28]. (4) Sequence Canonicalization uses the alignment approach implemented by [48], which violates temporal invariance. (5) Absolute naively conditions on absolute poses, which violate spatial invariance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "content": "We train conditional diffusion models with otherwise identical architecture using each parameterization, and then evaluate on the AMASS [56] test set. Metrics and percent differences compared to EgoAllo are reported in Table 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "text", + "content": "Overall, we find that the choice of conditioning parameterization makes a dramatic impact on estimation accuracy. We observe accuracy improve consistently as invariance properties are incorporated into the representation. Compared to EgoAllo, Absolute conditioning increases MPJPE by over " + }, + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "text", + "content": " for both shorter (length 32) and longer (length 128) sequences. Compared to EgoAllo, SeqCanonical conditioning increases MPJPE by nearly " + }, + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "text", + "content": " for length 32 sequences and " + }, + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 313, + 490, + 556, + 597 + ], + "type": "text", + "content": " for length 128 sequences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 603, + 473, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 473, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 473, + 615 + ], + "type": "text", + "content": "4.1.2. Comparisons against baselines" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "To further study EgoAllo's body estimation quality, we compare against three baselines. (1) NoShape. First, NoShape refers to a variation of EgoAllo that turns off shape estimation, and thus cannot estimate the wearer's height. (2) EgoEgo. We also compare against the human motion diffusion model from EgoEgo [48]. This is similar to EgoAllo, but considers only the SMPL \"mean\" body shape and uses sequence canonicalized coordinates for conditioning and as model output. (3) VAE+Opt." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7077" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 86, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 61, + 75, + 105, + 84 + ], + "lines": [ + { + "bbox": [ + 61, + 75, + 105, + 84 + ], + "spans": [ + { + "bbox": [ + 61, + 75, + 105, + 84 + ], + "type": "text", + "content": "AMASS [56]" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 86, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 86, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 86, + 293, + 186 + ], + "type": "table", + "html": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑Thead ↓
EgoAllo32129.8±1.1109.8±1.10.98±0.006.4±0.1
NoShape32138.1±1.1118.8±1.10.94±0.0144.7±0.4
EgoEgo32184.0±1.5158.6±1.60.81±0.0145.2±1.0
VAE+Opt32199.5±1.3191.4±1.40.49±0.0178.0±1.5
EgoAllo128119.7±1.3101.1±1.31.0±0.006.2±0.1
NoShape128128.1±1.3110.3±1.40.98±0.0144.6±0.7
EgoEgo128167.4±2.1145.8±2.00.92±0.0154.9±1.9
VAE+Opt128205.3±2.6192.3±2.80.75±0.0267.8±3.1
", + "image_path": "96e17e5803918f0e186ddc83057ea0b016244293400a2a15b76c4ba9e2959677.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 200, + 293, + 301 + ], + "blocks": [ + { + "bbox": [ + 61, + 189, + 99, + 198 + ], + "lines": [ + { + "bbox": [ + 61, + 189, + 99, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 189, + 99, + 198 + ], + "type": "text", + "content": "RICH [26]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 200, + 293, + 301 + ], + "lines": [ + { + "bbox": [ + 59, + 200, + 293, + 301 + ], + "spans": [ + { + "bbox": [ + 59, + 200, + 293, + 301 + ], + "type": "table", + "html": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\\( T_{head} \\) ↓
EgoAllo32193.7±3.4174.8±3.60.95±0.018.8±0.2
NoShape32200.9±3.3183.3±3.60.73±0.0244.9±0.9
EgoEgo32215.4±3.9192.9±4.00.73±0.0256.2±2.9
VAE+Opt32352.0±6.7354.8±6.50.59±0.02319.3±11.6
EgoAllo128176.2±5.6160.1±5.90.96±0.028.9±0.3
NoShape128185.7±5.5169.9±5.80.82±0.0345.8±1.6
EgoEgo128207.8±6.9187.8±6.80.88±0.0366.5±5.4
VAE+Opt128319.8±10.1323.8±10.50.75±0.04274.4±17.6
", + "image_path": "0b123743caef541c13437857d45292ecc0cee9404f2846963833d5c1a3d77222.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 315, + 293, + 416 + ], + "blocks": [ + { + "bbox": [ + 61, + 304, + 139, + 313 + ], + "lines": [ + { + "bbox": [ + 61, + 304, + 139, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 139, + 313 + ], + "type": "text", + "content": "Aria Digital Twins [61]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 315, + 293, + 416 + ], + "lines": [ + { + "bbox": [ + 58, + 315, + 293, + 416 + ], + "spans": [ + { + "bbox": [ + 58, + 315, + 293, + 416 + ], + "type": "table", + "html": "
MethodSeqMPJPE ↓PA-MPJPE ↓GND ↑\\( T_{head} \\)↓
EgoAllo32173.5±1.1146.1±1.10.88±0.01-
NoShape32178.5±1.1153.0±1.10.89±0.01-
EgoEgo32212.5±1.4181.3±1.60.64±0.01-
VAE+Opt32284.9±1.6283.9±1.90.63±0.01-
EgoAllo128155.1±1.6129.3±1.60.94±0.01-
NoShape128163.7±1.6140.0±1.60.96±0.01-
EgoEgo128182.6±2.3153.9±2.60.73±0.02-
VAE+Opt128290.8±3.8282.5±4.40.7±0.02-
", + "image_path": "cf7c5deabaef79dfa0733c695ca158ee3f2f8f8b381c5cf2a16b12b9d55e90f5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "type": "text", + "content": "Table 2. Body estimation performance, compared against a baseline without shape prediction, EgoEgo [48], and VAE+Opt [74, 102]. We exclude the " + }, + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{head}}" + }, + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "type": "text", + "content": " metric for ADT because the Biomech57 head joints used by ADT are not directly comparable to the SMPL-H head joints used by our model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": "Finally, we compare against an approach based on the SLAHMR [102] framework for human motion estimation from exocentric video. A key advantage of SLAHMR is that it uses an unconditional motion prior [74] in an optimization framework. It can therefore be adapted to new settings without re-training—we keep the same body pose and shape variables as the original pipeline, but replace the exocentric keypoint [71] cost with an egocentric CPF pose alignment cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": "Due to differences in problem formulation, many existing methods for egocentric human motion estimation are difficult to directly compare. This is particularly true when they have different inputs, such as fisheye cameras [91, 95, 96], wrist-mounted sensors [46], or handheld controller poses [7, 28, 29]. Additionally, prior works like EgoEgo [48] do not incorporate vision inputs for hand estimation. For fairness, we restrict all methods in this section to only CPF or head pose as input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "EgoAllo improves body motion estimates. We report metrics" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 80, + 426, + 131 + ], + "blocks": [ + { + "bbox": [ + 328, + 80, + 426, + 131 + ], + "lines": [ + { + "bbox": [ + 328, + 80, + 426, + 131 + ], + "spans": [ + { + "bbox": [ + 328, + 80, + 426, + 131 + ], + "type": "image", + "image_path": "a1dd517f31a214f45b04cf06f7e824d6a01737aa765229080bec506d6b7e7b73.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 137, + 400, + 147 + ], + "lines": [ + { + "bbox": [ + 348, + 137, + 400, + 147 + ], + "spans": [ + { + "bbox": [ + 348, + 137, + 400, + 147 + ], + "type": "text", + "content": "(a) Ground-truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 444, + 74, + 545, + 126 + ], + "blocks": [ + { + "bbox": [ + 444, + 74, + 545, + 126 + ], + "lines": [ + { + "bbox": [ + 444, + 74, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 444, + 74, + 545, + 126 + ], + "type": "image", + "image_path": "2da2e150065507518628a2247b7d7d78001b2fc6b0c5a050d9b3c92a6a194ae0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 137, + 514, + 148 + ], + "lines": [ + { + "bbox": [ + 475, + 137, + 514, + 148 + ], + "spans": [ + { + "bbox": [ + 475, + 137, + 514, + 148 + ], + "type": "text", + "content": "(b) EgoAllo" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 150, + 430, + 201 + ], + "blocks": [ + { + "bbox": [ + 323, + 150, + 430, + 201 + ], + "lines": [ + { + "bbox": [ + 323, + 150, + 430, + 201 + ], + "spans": [ + { + "bbox": [ + 323, + 150, + 430, + 201 + ], + "type": "image", + "image_path": "30685c4828fe440c3df429c5eb042a642232ee2915171b997b860880d7c3bbab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 215, + 400, + 225 + ], + "lines": [ + { + "bbox": [ + 349, + 215, + 400, + 225 + ], + "spans": [ + { + "bbox": [ + 349, + 215, + 400, + 225 + ], + "type": "text", + "content": "(c) EgoEgo [48]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 446, + 152, + 545, + 201 + ], + "blocks": [ + { + "bbox": [ + 446, + 152, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 446, + 152, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 446, + 152, + 545, + 201 + ], + "type": "image", + "image_path": "7fc72bfc266b0a35cc57f96ad7ae2d72897133c0c504bdd32fdd0f350ef3dbbf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 215, + 516, + 224 + ], + "lines": [ + { + "bbox": [ + 473, + 215, + 516, + 224 + ], + "spans": [ + { + "bbox": [ + 473, + 215, + 516, + 224 + ], + "type": "text", + "content": "(d) VAE+Opt" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 235, + 555, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 280 + ], + "type": "text", + "content": "Figure 4. Egocentric human motion estimation for a running sequence. We show the ground-truth, an output from EgoAllo, and outputs from two baselines. The glasses CAD model is placed at the conditioning transformation " + }, + { + "bbox": [ + 313, + 235, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{world,cpf}}" + }, + { + "bbox": [ + 313, + 235, + 555, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 301, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 422 + ], + "type": "text", + "content": "in Table 2 and visualize example outputs in Figure 4. We find that EgoAllo enables significant estimation improvements across all datasets, including accuracy improvements of " + }, + { + "bbox": [ + 313, + 301, + 555, + 422 + ], + "type": "inline_equation", + "content": "20\\sim 30\\%" + }, + { + "bbox": [ + 313, + 301, + 555, + 422 + ], + "type": "text", + "content": " over EgoEgo for both shorter and longer evaluation sequences. We found shape estimation critical for producing metric-scale, grounded estimates of human body motion, with the head aligned to input SLAM poses and the feet planted on the observed ground plane. This is evident in qualitative results, improved grounding metrics, and in the " + }, + { + "bbox": [ + 313, + 301, + 555, + 422 + ], + "type": "inline_equation", + "content": "6\\sim 7\\%" + }, + { + "bbox": [ + 313, + 301, + 555, + 422 + ], + "type": "text", + "content": " MPJPE gap between EgoAllo and the NoShape ablation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "type": "text", + "content": "VAE optimization converges poorly. Optimization-based estimation approaches have been effective for settings with keypoint costs [74, 102], but we found convergence difficult in our less constrained setting. In Table 2, we observe poor generalization: VAE+Opt performs competitively on the AMASS test set, but performance deteriorates dramatically when evaluating on RICH or ADT. VAE+Opt outputs in Figure 4 also look overly smoothed, without the same expressiveness as the conditional predictions of EgoAllo or EgoEgo [48]. This highlights the advantage of using a conditional diffusion model problem for this estimation problem." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Shape estimation evaluation. To better understand the shape estimation characteristics of EgoAllo, we compare against against the \"mean\" shape used by EgoEgo and the NoShape ablation. On the AMASS test set, we find: EgoAllo slightly improves overall shape (19mm → 18mm mean vertex-to-vertex error) and produces much better height (52mm → 32mm mean height error), but is not able to generalize in terms of body weight (5kg → 8kg mean weight error). The body shape is inferred from the wearer's head pose, which intuitively provides strong height constraints but is less correlated with weight. Accurate height is key for proper scene placement, as reflected by both the MPJPE and GND metrics." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7078" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 156, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 156, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 156, + 83 + ], + "type": "text", + "content": "4.2. Hand estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 330 + ], + "type": "text", + "content": "To evaluate hands estimated by EgoAllo, we run HaMeR on the segment of the EgoExo4D [17] validation set that is labeled with 3D hand pose keypoints. We quantitatively compare four hand estimation methods in Table 3. In (1) HaMeR [66], we use HaMeR out-of-the-box on undistorted egocentric RGB images. We do not assume bounding boxes as input; instead, we follow the HaMeR demo code and compute crops using ViTPose [99]. (2) EgoAllo-NoReproj uses all loss terms except for the reprojection loss (Equation 14). Hand guidance is done directly using the 3D wrist poses predicted by HaMeR. (3) EgoAllo-Mono is the same as EgoAllo-NoReproj, but guides hands using the reprojection loss. This accounts for the scale ambiguities that are inherent to the single-frame HaMeR estimates. Finally, we show an extension of our method. (4) EgoAllo-Wrist3D extends EgoAllo-Mono with 3D wrist pose losses from Project Aria's Machine Perception Services [85]—unlike HaMeR, which assumes monocular input, this is a separate sensing setup that also leverages a pair of SLAM cameras that are unique to Project Aria. For fairness across settings, we compute metrics only on timesteps where HaMeR estimates are available." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "text", + "content": "Results are provided in Table 3. While HaMeR's local poses (PA-MPJPE) are slightly better, EgoAllo's hand-body estimation significantly improves how well hands are estimated in the world coordinate system. Compared to HaMeR, EgoAllo-Mono drops MPJPE from " + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "inline_equation", + "content": "237.90\\mathrm{mm} \\rightarrow 131.45\\mathrm{mm}" + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "text", + "content": ". Reprojection-based guidance is important: despite using the same inputs, EgoAllo-NoReproj outputs are worse than EgoAllo-Mono in both MPJPE and PA-MPJPE. Finally, incorporating more accurate wrist pose estimates from the extra cameras on Project Aria drops errors " + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "inline_equation", + "content": "131.45\\mathrm{mm} \\rightarrow 60.08\\mathrm{mm}" + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "text", + "content": ". This highlights both room for improvement for EgoAllo and how EgoAllo's guidance-based formulation makes it easy to extend when more sensors are available." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 475, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 584 + ], + "type": "text", + "content": "Qualitatively, we observed that high hand estimation errors in naive monocular estimation with HaMeR are explained by a combination of detection failures and monocular ambiguities. Even when detections succeed, the scale and distance of monocular HaMeR estimates are often incorrect or flicker in between frames. EgoAllo improves on this by applying kinematic constraints and a diffusion-based prior for plausible body motion—we provide examples of HaMeR estimates rendered jointly with EgoAllo outputs in Figure 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 597, + 117, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 117, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 117, + 609 + ], + "type": "text", + "content": "5. Dicussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Limitations and future work. While the core contributions of EgoAllo are general, the current implementation of our system has a few limitations that we hope to explore in future work. First, diffusion model guidance is a test-time optimization process that depends on hyperparameters and incurs a runtime cost. In the future, it may be possible to bootstrap using outputs from our model to train a feedforward model that avoids this step. Success for hand guidance also still depends on reasonable" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 375, + 137 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 375, + 137 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 375, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 375, + 137 + ], + "type": "image", + "image_path": "b5d0e6e5c24724b83f082f64773dd82388892094b92a20c923e29805a629d954.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 439, + 137 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 439, + 137 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 439, + 137 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 439, + 137 + ], + "type": "image", + "image_path": "6dc08720293e3599f6e0283137f1207700143f7389ba4d6db44da6d7260963e9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 70, + 496, + 137 + ], + "blocks": [ + { + "bbox": [ + 440, + 70, + 496, + 137 + ], + "lines": [ + { + "bbox": [ + 440, + 70, + 496, + 137 + ], + "spans": [ + { + "bbox": [ + 440, + 70, + 496, + 137 + ], + "type": "image", + "image_path": "e05d4df87330c2592576531aae940510d7869503b762d80c542d64b9bfc57259.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 498, + 70, + 553, + 137 + ], + "blocks": [ + { + "bbox": [ + 498, + 70, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 498, + 70, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 498, + 70, + 553, + 137 + ], + "type": "image", + "image_path": "067d87f430e88fcaad7a74dbafb94eb9548a93818e24efba0428cfbdbf8e9c4e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 139, + 395, + 261 + ], + "blocks": [ + { + "bbox": [ + 316, + 139, + 395, + 261 + ], + "lines": [ + { + "bbox": [ + 316, + 139, + 395, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 395, + 261 + ], + "type": "image", + "image_path": "ff32739cb8fa7d88e1a63cd5adb7c329d98bfbbad622b0567df87878edef0758.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 269, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 346 + ], + "type": "text", + "content": "Figure 5. Body estimation improves hand estimation. We show raw outputs from HaMeR [66] in blue and hand-body estimations from EgoAllo in purple. Top: improved scene interaction during touchscreen operation with EgoAllo-Mono. We know a priori that the fingers are contacting the screen in this sequence. Bottom: qualitative examples from EgoExo [17] evaluation, showing the differences between monocular hands and EgoAllo-Wrist3D estimates." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 395, + 139, + 471, + 261 + ], + "blocks": [ + { + "bbox": [ + 395, + 139, + 471, + 261 + ], + "lines": [ + { + "bbox": [ + 395, + 139, + 471, + 261 + ], + "spans": [ + { + "bbox": [ + 395, + 139, + 471, + 261 + ], + "type": "image", + "image_path": "fb2139961d402a1d8399156aa9d44ba260dd3b487a9f1c4d3a3e7338d5ad9e9a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 473, + 139, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 473, + 139, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 473, + 139, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 473, + 139, + 553, + 261 + ], + "type": "image", + "image_path": "556860c8857b3de77f3dee4eca39873535ca320e08a5bbd1ad93cdcd1e448f26.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 332, + 359, + 538, + 444 + ], + "blocks": [ + { + "bbox": [ + 332, + 359, + 538, + 444 + ], + "lines": [ + { + "bbox": [ + 332, + 359, + 538, + 444 + ], + "spans": [ + { + "bbox": [ + 332, + 359, + 538, + 444 + ], + "type": "table", + "html": "
MethodMPJPE ↓PA-MPJPE ↓
HaMeR237.90±1.8913.04±1.89
EgoAllo-NoReproj143.20±0.4214.75±0.42
EgoAllo-Mono131.45±0.3914.71±0.39
EgoAllo-Wrist3D60.08±0.2614.38±0.26
", + "image_path": "8b200b4fd84a8d3a8947d2847f6aad7ffafd0bb00b74847bf2f2be728d3f1db5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 515 + ], + "type": "text", + "content": "Table 3. Hand estimation errors in millimeters. EgoAllo's hand-body estimation can constrain and resolve ambiguities in noisy outputs from HaMeR, which we observe can reduce MPJPE for hands by over " + }, + { + "bbox": [ + 313, + 460, + 555, + 515 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 460, + 555, + 515 + ], + "type": "text", + "content": ". EgoAllo-Wrist3D incorporates 3D wrist poses provided from the grayscale cameras on Project Aria." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 528, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 624 + ], + "type": "text", + "content": "monocular hand estimates. Errors like left/right flipping or spurious detections can therefore cause estimation failure. Finally, we assume flat floors. This is in part because our training data [56] includes floor planes but no detailed scene geometry. As a result, our method will fail in settings like hills or staircases. In the future, we hope to extend our insights to data with more detailed scene information, which concurrent work has highlighted the usefulness of in informing human body estimation [21]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Conclusion. We presented EgoAllo, a system for estimating human motion using a head-mounted device. EgoAllo jointly estimates human body pose, height, and hand parameters from only egocentric SLAM poses and images. Results highlight the importance of spatial and temporal invariance in conditioning for this problem, while demonstrating how estimated bodies can be used to improve hand estimation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7079" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "content": "Acknowledgments. This project was funded in part by NSF:CNS-2235013 and IARPA DOI/IBC No. 140D0423C0035. JM was supported by ONR MURI N00014-21-1-2801. YM acknowledges support from the joint Simons Foundation-NSF DMS grant #2031899, the ONR grant N00014-22-1-2102, the NSF grant #2402951, and partial support from TBSI, InnoHK, and the University of Hong Kong. BY is supported by the National Science Foundation Graduate Research Fellowship Program under Grant DGE 2146752. The authors would also like to thank Hongsuk Choi, Michael Taylor, Tyler Bonnen, Songwei Ge, Chung Min Kim, and Justin Kerr for insightful technical discussion and suggestions, as well as Jiaman Li for helpful answers to questions about EgoEgo." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 112, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 112, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 112, + 253 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 261, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 65, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 65, + 261, + 294, + 304 + ], + "type": "text", + "content": "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 306, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 306, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 65, + 306, + 295, + 350 + ], + "type": "text", + "content": "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. ACM New York, NY, USA, 2005. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 351, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 351, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 65, + 351, + 295, + 384 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 5, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 386, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 386, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 65, + 386, + 295, + 430 + ], + "type": "text", + "content": "[4] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 431, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 431, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 65, + 431, + 295, + 485 + ], + "type": "text", + "content": "[5] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 487, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 487, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 65, + 487, + 295, + 531 + ], + "type": "text", + "content": "[6] Marcus A Brubaker, David J Fleet, and Aaron Hertzmann. Physics-based person tracking using the anthropomorphic walker. International journal of computer vision, 87(1-2): 140-155, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 533, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 533, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 65, + 533, + 295, + 597 + ], + "type": "text", + "content": "[7] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4221-4231, 2023, 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 600, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 600, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 65, + 600, + 295, + 644 + ], + "type": "text", + "content": "[8] R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE conference on computer vision and pattern recognition (CVPR), pages 77-85. IEEE, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 65, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 65, + 646, + 295, + 689 + ], + "type": "text", + "content": "[9] Haiwei Chen, Shichen Liu, Weikai Chen, Hao Li, and Randall Hill. Equivariant point network for 3d point cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14514-14523, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 95, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 95, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 320, + 95, + 555, + 149 + ], + "type": "text", + "content": "[11] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4800-4810, 2023. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 151, + 553, + 183 + ], + "type": "text", + "content": "[12] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 184, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 553, + 217 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 553, + 271 + ], + "type": "text", + "content": "[14] Haiwen Feng, Peter Kulits, Shichen Liu, Michael J Black, and Victoria Fernandez Abrevaya. Generalizing neural human fitting to unseen poses with articulated se (3) equivariance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7977-7988, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 272, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 272, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 553, + 327 + ], + "type": "text", + "content": "[15] Saeed Ghorbani, Calden Wloka, Ali Etemad, Marcus A Brubaker, and Nikolaus F Troje. Probabilistic character motion synthesis using a hierarchical deep latent variable model. In Computer Graphics Forum, pages 225-239. Wiley Online Library, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 553, + 370 + ], + "type": "text", + "content": "[16] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In Int. Conf. Comput. Vis., 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 320, + 372, + 553, + 426 + ], + "type": "text", + "content": "[17] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 6, 8, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 553, + 469 + ], + "type": "text", + "content": "[18] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. arXiv preprint arXiv:2403.18118, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 471, + 553, + 514 + ], + "type": "text", + "content": "[19] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In Int. Conf. Comput. Vis., pages 1381-1388, IEEE, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 514, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 514, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 320, + 514, + 553, + 548 + ], + "type": "text", + "content": "[20] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10884-10894, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 320, + 548, + 553, + 602 + ], + "type": "text", + "content": "[21] Vladimir Guzov, Yifeng Jiang, Fangzhou Hong, Gerard Pons-Moll, Richard Newcombe, C Karen Liu, Yuting Ye, and Lingni Ma. Hmd: Environment-aware motion generation from single egocentric head-mounted device. arXiv preprint arXiv:2409.13426, 2024. 2, 8, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 603, + 553, + 635 + ], + "type": "text", + "content": "[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 669 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7080" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "text", + "content": "[25] Nicholas Howe, Michael Leventon, and William Freeman. Bayesian reconstruction of 3d human motion from single-camera video. Advances in neural information processing systems, 12, 1999. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 173 + ], + "type": "text", + "content": "[26] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13274-13285, 2022. 6, 7, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 295, + 217 + ], + "type": "text", + "content": "[27] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 295, + 274 + ], + "type": "text", + "content": "[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 275, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 275, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 275, + 295, + 317 + ], + "type": "text", + "content": "[29] Jiaxi Jiang, Paul Streli, Manuel Meier, Andreas Fender, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2, 3, 4, 5, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 320, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 320, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 320, + 295, + 374 + ], + "type": "text", + "content": "[30] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6142-6152, 2024. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "type": "text", + "content": "[31] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards in-the-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 421, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 295, + 453 + ], + "type": "text", + "content": "[32] Angjoo Kanazawa, Abhishek Sharma, and David Jacobs. Locally scale-invariant convolutional neural networks. arXiv preprint arXiv:1412.5104, 2014. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 295, + 497 + ], + "type": "text", + "content": "[33] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7122-7131, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 295, + 533 + ], + "type": "text", + "content": "[34] Angjoo Kanazawa, Jason Y Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5614-5623, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 534, + 295, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 587 + ], + "type": "text", + "content": "[35] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 632 + ], + "type": "text", + "content": "[36] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 635, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 666 + ], + "type": "text", + "content": "[37] Meejin Kim and Sukwon Lee. Fusion poser: 3d human pose estimation using sparse imus and head trackers in real time. Sensors, 22(13):4846, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 669, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 669, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 669, + 295, + 690 + ], + "type": "text", + "content": "[38] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "text", + "content": "[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5253-5263, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 140 + ], + "type": "text", + "content": "[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 184 + ], + "type": "text", + "content": "[41] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 553, + 229 + ], + "type": "text", + "content": "[42] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 231, + 553, + 274 + ], + "type": "text", + "content": "[43] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 276, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 553, + 329 + ], + "type": "text", + "content": "[44] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6050-6059, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 331, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 331, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 553, + 364 + ], + "type": "text", + "content": "[45] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 553, + 398 + ], + "type": "text", + "content": "[46] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. arXiv preprint arXiv:2401.00847, 2024. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 399, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 553, + 432 + ], + "type": "text", + "content": "[47] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 553, + 466 + ], + "type": "text", + "content": "[48] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17142-17151, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 468, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 468, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 320, + 468, + 553, + 499 + ], + "type": "text", + "content": "[49] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 501, + 553, + 555 + ], + "type": "text", + "content": "[50] Zongmian Li, Jiri Sedlar, Justin Carpentier, Ivan Laptev, Nicolas Mansard, and Josef Sivic. Estimating 3d motion and forces of person-object interactions from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8649, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 557, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 553, + 590 + ], + "type": "text", + "content": "[51] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 591, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 591, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 591, + 553, + 634 + ], + "type": "text", + "content": "[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi-person linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851-866, 2023. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 635, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 666 + ], + "type": "text", + "content": "[53] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "text", + "content": "[54] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7081" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "type": "text", + "content": "[55] Lingni Ma, Yuting Ye, Fangzhou Hong, Vladimir Guzov, Yifeng Jiang, Rowan Postyeni, Luis Pesqueira, Alexander Gamino, Vijay Baiyya, Hyo Jin Kim, et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. arXiv preprint arXiv:2406.09905, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 295, + 172 + ], + "type": "text", + "content": "[56] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5442-5451, 2019. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 173, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 173, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 173, + 296, + 227 + ], + "type": "text", + "content": "[57] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Ecoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 228, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 261 + ], + "type": "text", + "content": "[58] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Int. Conf. Comput. Vis., pages 2640-2649, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "text", + "content": "[59] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 316, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 369 + ], + "type": "text", + "content": "[60] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In 2018 international conference on 3D vision (3DV), pages 484-494. IEEE, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 371, + 295, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 295, + 436 + ], + "type": "text", + "content": "[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 1, 6, 7, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 437, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 480 + ], + "type": "text", + "content": "[62] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 459-468, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "text", + "content": "[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10975-10985, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 537, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 569 + ], + "type": "text", + "content": "[64] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 570, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 295, + 613 + ], + "type": "text", + "content": "[65] Georgios Pavlakos, Ethan Weber, Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In Eur. Conf. Comput. Vis., pages 732-749. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 614, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 295, + 646 + ], + "type": "text", + "content": "[66] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In *arxiv*, 2023. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "type": "text", + "content": "[67] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "text", + "content": "[68] Chiara Plizzari, Shubham Goel, Toby Perrett, Jacob Chalk, Angjoo Kanazawa, and Dima Damen. Spatial cognition from" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "egocentric video: Out of sight, not out of mind. arXiv preprint arXiv:2404.05072, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 97, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 97, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 320, + 97, + 553, + 152 + ], + "type": "text", + "content": "[69] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 154, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 154, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 320, + 154, + 553, + 209 + ], + "type": "text", + "content": "[70] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 211, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 211, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 320, + 211, + 553, + 255 + ], + "type": "text", + "content": "[71] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2740-2749, 2022. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 257, + 553, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 257, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 320, + 257, + 553, + 290 + ], + "type": "text", + "content": "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 292, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 292, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 320, + 292, + 553, + 346 + ], + "type": "text", + "content": "[73] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 553, + 403 + ], + "type": "text", + "content": "[74] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 4, 5, 7, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "type": "text", + "content": "[75] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 320, + 453, + 553, + 507 + ], + "type": "text", + "content": "[76] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 320, + 509, + 553, + 552 + ], + "type": "text", + "content": "[77] Gregory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net: Localization-classification-regression for human pose. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3433-3441, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 555, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 555, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 320, + 555, + 553, + 609 + ], + "type": "text", + "content": "[78] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 612, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 612, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 320, + 612, + 553, + 644 + ], + "type": "text", + "content": "[79] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "type": "text", + "content": "[80] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7082" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 106 + ], + "type": "text", + "content": "[81] Silvio Savarese and Li Fei-Fei. 3d generic object categorization, localization and pose estimation. In 2007 IEEE 11th international conference on computer vision, pages 1-8. IEEE, 2007. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 296, + 139 + ], + "type": "text", + "content": "[82] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 141, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 141, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 295, + 184 + ], + "type": "text", + "content": "[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K Hodgins. Motion capture from body-mounted cameras. In ACM SIGGRAPH 2011 papers, pages 1-10. ACM New York, NY, USA, 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 230 + ], + "type": "text", + "content": "[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 285 + ], + "type": "text", + "content": "[85] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 287, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 295, + 319 + ], + "type": "text", + "content": "[86] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 321, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 295, + 364 + ], + "type": "text", + "content": "[87] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 366, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 295, + 399 + ], + "type": "text", + "content": "[88] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 400, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 295, + 432 + ], + "type": "text", + "content": "[89] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "type": "text", + "content": "[90] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 479, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 295, + 522 + ], + "type": "text", + "content": "[91] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7728-7738, 2019. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 524, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 295, + 578 + ], + "type": "text", + "content": "[92] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 295, + 623 + ], + "type": "text", + "content": "[93] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 295, + 668 + ], + "type": "text", + "content": "[94] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "text", + "content": "[95] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11500-11509, 2021. 2, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 139 + ], + "type": "text", + "content": "[96] Jian Wang, Zhe Cao, Diogo Luvizon, Lingjie Liu, Kripasindhu Sarkar, Danhang Tang, Thabo Beeler, and Christian Theobalt. Egocentric whole-body motion capture with fisheyevit and diffusion-based motion refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 777-787, 2024. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 141, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 554, + 174 + ], + "type": "text", + "content": "[97] Laurenz Wiskott and Terrence J Sejnowski. Slow feature analysis: Unsupervised learning of invariances. Neural computation, 14(4):715-770, 2002. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 176, + 555, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 176, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 320, + 176, + 555, + 231 + ], + "type": "text", + "content": "[98] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 233, + 555, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 233, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 320, + 233, + 555, + 266 + ], + "type": "text", + "content": "[99] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246, 2022.8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 268, + 555, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 268, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 555, + 311 + ], + "type": "text", + "content": "[100] Jingyun Yang, Congyue Deng, Jimmy Wu, Rika Antonova, Leonidas Guibas, and Jeannette Bohg. Equivact: Sim(3)-equivariant visuomotor policies beyond rigid object manipulation, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 314, + 555, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 314, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 314, + 555, + 357 + ], + "type": "text", + "content": "[101] Jingyun Yang, Zi-ang Cao, Congyue Deng, Rika Antonova, Shuran Song, and Jeannette Bohg. Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning. arXiv preprint arXiv:2407.01479, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 359, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 359, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 359, + 554, + 403 + ], + "type": "text", + "content": "[102] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21222-21232, 2023. 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 405, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 554, + 449 + ], + "type": "text", + "content": "[103] Brent Yi, Weijia Zeng, Sam Buchanan, and Yi Ma. Canonical factors for hybrid neural fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3414-3426, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 451, + 554, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 451, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 451, + 554, + 483 + ], + "type": "text", + "content": "[104] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 486, + 555, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 486, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 317, + 486, + 555, + 539 + ], + "type": "text", + "content": "[105] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. arXiv preprint arXiv:2305.01599, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 543, + 554, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 543, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 317, + 543, + 554, + 576 + ], + "type": "text", + "content": "[106] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 578, + 555, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 555, + 620 + ], + "type": "text", + "content": "[107] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 667 + ], + "type": "text", + "content": "[108] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "text", + "content": "[109] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7083" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[110] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. arXiv preprint arXiv:2305.03462, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "type": "text", + "content": "[111] Daiwei Zhang, Gengyan Li, Jiajie Li, Mickaël Bressieux, Otmar Hilliges, Marc Pollefeys, Luc Van Gool, and Xi Wang. Egogaussian: Dynamic scene understanding from egocentric video with 3d gaussian splatting. arXiv preprint arXiv:2406.19811, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 296, + 195 + ], + "type": "text", + "content": "[112] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 197, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 197, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 197, + 295, + 251 + ], + "type": "text", + "content": "[113] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 252, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 252, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 295, + 306 + ], + "type": "text", + "content": "[114] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7989-8000, 2023. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 308, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 295, + 341 + ], + "type": "text", + "content": "[115] Zhengdong Zhang, Arvind Ganesh, Xiao Liang, and Yi Ma. Tilt: Transform invariant low-rank textures. International journal of computer vision, 99:1-24, 2012. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7084" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_content_list.json b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f4f193673435417d3e85ee6fb39d3323df1ad7b4 --- /dev/null +++ b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "Ev-3DOD: Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras", + "text_level": 1, + "bbox": [ + 151, + 128, + 846, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hoonhee Cho*, Jae-Young Kang*, Youngho Kim, and Kuk-Jin Yoon KAIST", + "bbox": [ + 228, + 203, + 769, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gnsgnsgml,mickeykang,kmax2001,kjyoon}@kaist.ac.kr", + "bbox": [ + 274, + 241, + 718, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e63f2f6992edaa7daed8436d54be8bd4c5979af92b981593186c691844ac6ae3.jpg", + "image_caption": [ + "Figure 1. Comparison between the proposed Ev-3DOD and conventional 3D detection methods. Fixed frame rate sensors (e.g., LiDAR, frame camera) inevitably have periods of \"blind time\" $(t_{0\\rightarrow 1})$ when they lack data between active timestamps (0 and 1). Consequently, as shown in (a), conventional methods cannot detect objects during blind time, while other objects like vehicles can undergo significant motion within these gaps. In contrast, as illustrated in (b), our method, Ev-3DOD, leverages the high temporal resolution of the event camera, enabling accurate object detection even during the blind times of other sensors." + ], + "image_footnote": [], + "bbox": [ + 98, + 279, + 367, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2be3700a8221c22bd980928c3d6da76f63ea82f142b6881b969ce148df3640f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 287, + 607, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/115f667de511432d6fcb088013b37a0e501177ccbb0f550566a3d8d130f7074d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 287, + 879, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 590, + 328, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting 3D objects in point clouds plays a crucial role in autonomous driving systems. Recently, advanced multimodal methods incorporating camera information have achieved notable performance. For a safe and effective autonomous driving system, algorithms that excel not only in accuracy but also in speed and low latency are essential. However, existing algorithms fail to meet these requirements due to the latency and bandwidth limitations of fixed frame rate sensors, e.g., LiDAR and camera. To address this limitation, we introduce asynchronous event cameras into 3D object detection for the first time. We leverage their high temporal resolution and low bandwidth to enable high-speed 3D object detection. Our method enables detection even during inter-frame intervals when synchronized data is unavailable, by retrieving previous 3D information through the event camera. Furthermore, we introduce the first event-based 3D object detection datasets, EvWaymo and DSEC-3DOD, both of which include ground", + "bbox": [ + 88, + 628, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "truth 3D bounding boxes at 100 FPS, establishing the first benchmarks for event-based 3D detectors. The code and dataset are available at https://github.com/mickeykang16/Ev3DOD.", + "bbox": [ + 511, + 593, + 908, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 684, + 645, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection is a crucial task in autonomous driving. Multi-modal sensors like LiDAR and cameras are commonly used to understand scenes and handle complex road scenarios. This sensor setup consists of synchronized sensors with a fixed frame rate and has limited speed due to bandwidth limitations. As shown in Fig. 1 (a), the blind time, which is the period when sensor information is unavailable (e.g., 100 milliseconds), is critical as cars can experience large motions as 3 meters during this time. Therefore, temporal resolution is a crucial factor that can directly influence accident risk.", + "bbox": [ + 511, + 710, + 908, + 876 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The first two authors contributed equally.", + "bbox": [ + 531, + 886, + 758, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27197", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, event cameras [18] have introduced a new paradigm in perceptual research. Event cameras provide brightness changes as a stream of data with sub-milisecond latency, making them robust to motion blur [41, 106] and enabling low bandwidth [22, 94]. Event cameras overcome the trade-off between bandwidth and blind time but have the drawback of sacrificing intensity information. To address this, recent work [12, 13, 92, 120] has combined event cameras with frame cameras to complement the sparsity of event cameras while overcoming the fixed frame rate sensor's limits.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim for accurate 3D object detection in blind time through a multi-modal approach that combines LiDAR and frame-based sensors with an event camera, leveraging its high temporal resolution. As shown in Fig. 1 (b), event data is available during blind times, while the LiDAR and image data from the most recent active timestamp are accessible. Therefore, there is a misalignment in timestamps between the fixed frame sensor data and the current blind time. Then, this raises the question: How can we estimate 3D motion using only event data, especially when data from other sensors is unavailable at the current time?", + "bbox": [ + 89, + 258, + 480, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we propose a virtual 3D event fusion (V3D-EF) that transfers past 3D information to align with the present time using the event camera. As extensively covered in previous research [14, 24, 40, 89], event data inherently contains motion information on an image plane. Based on this observation, we hypothesized that providing a 3D cue associated with the initial timestamp to the event information would enable accurate estimation of the 3D motion vector. To this end, we project event features into a 3D voxel space to create virtual 3D event features. In this process, events are projected only onto regions where point cloud data exists, aligning with the point cloud data. Then, we aggregate the event features with the voxel features to estimate an implicit motion field that captures the motion inherent within the grid. The generated implicit motion field encapsulates both the 3D spatial information from voxels and the motion information from events, enabling the calculation of a 3D motion vector from the most recent active timestamp to the present blind time.", + "bbox": [ + 89, + 441, + 482, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another crucial factor in 3D object detection is the score for each bounding box. The model considers both box proposal scores at the active timestamp and the motion score of each predicted box in blind time. In blind time, the model aligns 3D information of voxel features and motion information of events to learn the confidence of box motion.", + "bbox": [ + 89, + 732, + 480, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A major hurdle in 3D detection research using event cameras is the lack of suitable datasets. To address this, we present two datasets: Ev-Waymo, which includes synthetic event data, and DSEC-3DOD, which contains real event data. Unlike existing public 3D object detection datasets,", + "bbox": [ + 89, + 825, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ours include annotations not only for active timestamps with 3D LiDAR sensor data but also for blind times, providing annotation in 100 FPS frame rate.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 150, + 663, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-modality 3D Object Detection. Given that cameras offer significant cost advantages over LiDAR sensors, many researchers have focused on developing methods that leverage camera-based systems for 3D object detection using image-only inputs [7, 10, 26, 30, 33, 38, 46, 54, 60-62, 64, 71, 91, 97, 99, 115]. Image-based 3D object detection faces challenges due to the lack of direct depth information. To overcome this challenge, several studies [7, 70, 76, 83, 98, 115] have employed depth estimation techniques to generate pseudo 3D point clouds or elevate 2D features into 3D space for object detection, alongside proposals for transformer-based architectures [35, 99, 121] that utilize 3D object queries and 3D-2D correspondences. Nonetheless, accurately estimating depth from images remains difficult, resulting in image-based methods performing worse than LiDAR-based approaches.", + "bbox": [ + 511, + 175, + 903, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR-based 3D object detection can reliably estimate 3D bounding boxes using point clouds captured from LiDAR sensors. Current LiDAR-based detection methods can be divided into three categories based on different point cloud encoding formats: point-based methods [17, 32, 79, 80, 82, 86, 88, 108, 112, 124], voxel-based methods [4, 15, 16, 25, 29, 29, 31, 43, 49, 55, 65, 66, 85, 100, 103, 111, 117, 125], and point-voxel fusion networks [6, 28, 68, 87, 107]. While LiDAR detection is advantageous in various conditions, it often struggles in areas with sparse LiDAR data. Therefore, integrating the geometric benefits of point clouds with the semantic richness of images remains crucial for improving performance.", + "bbox": [ + 511, + 417, + 903, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modality 3D Object Detection. Multi-modal sensor fusion [5, 34, 36, 47, 48, 56, 57, 59, 90, 101, 102, 113], where different sensors complement each other, typically utilizes a combination of cameras and LiDAR. Multimodal 3D object detection has gained attention for improving accuracy over unimodal methods. Fusion approaches are categorized into early [3, 11, 42, 81, 95, 104], middle [51-53, 58, 63, 77, 78, 109, 122], and late fusions [1, 47, 69], with middle fusion now preferred due to its robustness to calibration errors and enhanced feature interaction. Crossmodal fusion has achieved significant performance improvements by enhancing the sparse features of point clouds with semantic information from images. However, there are still unexplored areas concerning safety in autonomous driving. LiDAR and cameras have limited time resolution (i.e., $10 - 20\\mathrm{Hz}$ ) due to the relatively high bandwidths. To enhance safety, algorithms that operate at high speeds are essential. To address this, we introduce an asynchronous event camera with high temporal resolution for 3D object", + "bbox": [ + 511, + 614, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27198", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2cba8d8d590ff446bb95c44596e9b699280aa311b7afbaffef249b015571c0b9.jpg", + "image_caption": [ + "Figure 2. The overall pipeline for event-based 3D object detection (Ev-3DOD). (a): At active timestamp 0, LiDAR and image data are available. Therefore, we utilize an RGB-LiDAR Region Proposal Network (RPN) to extract voxel features and 3D bounding boxes at the active timestamp. (b): To predict the bounding box during blind time, $0 \\leq t < 1$ , we estimate the 3D motion and confidence score using event features. For computational efficiency, we design the process to compute (a) only once before the next active timestamp while performing iterative computations solely for (b)." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 883, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "detection for the first time. The proposed method performs multi-modal fusion between synchronized sensors and asynchronous events to increase the operational frequency of 3D detection.", + "bbox": [ + 88, + 431, + 480, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based Object Detection. Event cameras break through the limitations of bandwidth requirements and perceptual latency with the trade-off of sparse data and loss of intensity information. To make the most of these characteristics, prior work has designed models to adapt to event data using graph neural networks [2, 19, 20, 50, 84], spiking neural networks [110, 119], and sparse neural networks [67, 74]. Recently, effective dense neural networks [27, 37, 45, 72, 75, 96, 127] and transformer architectures [22, 44, 73, 74] have been developed to leverage event cameras, aiming to achieve both low latency and high performance. These advances enable real-time operation with impressive accuracy, demonstrating the practical applicability of event cameras in real-world scenarios. To the best of our knowledge, we are the first to extend event-based object detection from 2D to 3D space. By employing a multimodal fusion approach of the event camera and 3D sensors (i.e., LiDAR), we aim to overcome the latency and persistent interval time barriers of 3D object detections.", + "bbox": [ + 91, + 492, + 483, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 792, + 181, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition.", + "text_level": 1, + "bbox": [ + 89, + 816, + 277, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR and camera data are only available at active timestamps 0 and 1, where 1 represents invalid future information in an online process. Our goal is to predict 3D bounding boxes during $0 \\leq t < 1$ , including the blind time", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "when synchronized sensors are unavailable. Given a pair of synchronized 3D points $P_0 = \\{(x_l, y_l, z_l)\\}_{l=1}^N$ and image $I_0 \\in \\mathbb{R}^{3 \\times H \\times W}$ at active timestamp 0, we leverage event stream $\\mathcal{E}_{0 \\to t} = \\{(i, j, p, \\tau) | 0 \\leq \\tau < t\\}$ to detect 3D bounding box set $B_t$ at an arbitrary time $0 \\leq t < 1$ .", + "bbox": [ + 511, + 431, + 906, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Framework Overview", + "text_level": 1, + "bbox": [ + 511, + 516, + 718, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in (a) of Fig. 2, at active timestamp 0, both LiDAR and image data are available. RGB-LiDAR Region Proposal Network (RPN) generates bounding box proposals from rich 3D information in active time. We begin by voxelizing the point cloud and generating voxel features, denoted as $\\mathbf{V}_0\\in \\mathbb{R}^{D_X\\times D_Y\\times D_Z\\times C}$ , where $D_{X}$ , $D_{Y}$ , and $D_Z$ represent the grid size, and $C$ denotes the channel dimension, using a 3D backbone [125]. Then, using existing region proposals [114] and refinement networks [47] with image features, we generate bounding box proposals, $B_0 = \\{B_0^1,B_0^2,\\dots ,B_0^n\\}$ and box proposal confidence $p_0 = \\{p_0^1,p_0^2,\\dots ,p_0^n\\} \\in [0,1]$ for the active timestamp 0. RGB-LiDAR-based multi-modal networks have been widely studied, leading to many advanced models capable of accurate estimation at active timestamps. We tuned the model with fewer parameters and a reduced channel dimension to improve efficiency.", + "bbox": [ + 509, + 537, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our emphasis is on achieving precise estimation throughout the blind time, spanning the interval from active timestamp 0 to the subsequent timestamp 1. To infer bounding boxes at an arbitrary inter-frame time $t$ , we first convert the event stream, $\\mathcal{E}_{0\\rightarrow t}$ into voxel grids [126] as $\\mathbb{E}_t$ . Then, we use a feature encoder to produce event features, $\\mathbf{E}_t\\in \\mathbb{R}^{H / 4\\times W / 4\\times C}$ , where $H$ and $W$ is the spa", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27199", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tial dimensions of the events. Our objective is to estimate the 3D motion of objects during blind time. To achieve this, we utilize the bounding box set $B_{0}$ and voxel features $\\mathbf{V}_0$ from the active timestamp 0. During blind times, we share the voxel features, $\\mathbf{V}_0$ , generated at the active timestamp, ensuring that voxel features are computed only once to maintain a cost-effective structure. Through the Virtual 3D Event Fusion (V3D-EF), we generate an implicit motion field for each bounding box by integrating event and voxel features, encapsulating motion information. Using this field, we compute the 3D motion and confidence score, enabling the estimation of 3D boxes at the blind time $t$ .", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Virtual 3D Event Fusion for Motion Prediction", + "text_level": 1, + "bbox": [ + 89, + 282, + 482, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The event feature continuously captures each pixel's motion information over time on the 2D image plane. However, estimating 3D motion using only events is a challenge. We hypothesize that providing each event feature with adequate 3D information would allow for effective 3D motion estimation. To achieve this, we aim to incorporate a 3D cue into the event features. Since both event features and voxel features are sparse, it is crucial to align each voxels with the corresponding events accurately. Directly using the voxel features would cause misalignment with the actual 3D geometry. Thus, we adopt the voxel point centroid method employed in previous works [32].", + "bbox": [ + 89, + 306, + 483, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first consider the non-empty voxel features as $\\mathbf{V}_0 = \\{V_0^k = \\{h_0^k,f_0^k\\} \\mid k = 1,\\dots ,N_\\nu \\}$ , where $h_0^k$ is the 3D voxel index, $f_0^k$ is the associated voxel feature vector, and $N_{\\mathcal{V}}$ is the number of non-empty voxels. Then, points within each voxel are grouped into a set $\\mathcal{N}(V_0^k)$ by determining their voxel index $h_0^k$ based on their spatial coordinates $X_{i} = (x_{i},y_{i},z_{i})$ of points. The centroid of each voxel feature is subsequently computed as:", + "bbox": [ + 89, + 488, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {0} ^ {k} = \\frac {1}{\\left| \\mathcal {N} \\left(V _ {0} ^ {k}\\right) \\right|} \\sum_ {X _ {i} \\in \\mathcal {N} \\left(V _ {0} ^ {k}\\right)} X _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 619, + 482, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The centroid $c_0^k$ is calculated for non-empty voxels, ensuring it corresponds only to indices with existing points, with the goal of aligning these centroids to the event data. As shown in the top of Fig. 3, we propose a method to project events into a virtual voxel space based on the coordinates of the centroids. The projection onto the image plane from 3D space can be obtained as: $\\mathbf{p}_0^k = K \\cdot E \\cdot c_0^k$ , where $K$ and $E$ denote the intrinsic and extrinsic parameters, $\\cdot$ operation is matrix multiplication. Therefore, we calculate the event coordinates corresponding to each centroid, $c_0^k$ , and gather the event features, $\\mathbf{E}_t$ , at these coordinates, $\\mathbf{p}_0^k$ , to create virtual 3D event features, $\\mathbf{V}_t^E = \\{\\mathbf{E}_t(\\mathbf{p}_0^k)\\}_{k=1}^{N_\\nu} \\in \\mathbb{R}^{N_\\nu \\times C}$ .", + "bbox": [ + 89, + 672, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we aim to estimate the 3D motion of bounding boxes from the active timestamp $0$ up to the current time $t$ , we consider fusion only for the features corresponding", + "bbox": [ + 89, + 856, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4fea34eff8b6df80bcb906bf14b904c30978e89f1b4e2aa1c020bc200ee82d21.jpg", + "image_caption": [ + "Figure 3. Virtual 3D Event Fusion (V3D-EF). We generate an implicit motion field for each bounding box by applying the ROI pooling separately to the virtual 3D event features and the voxel features. Virtual 3D event features represent the 3D projection of 2D events. None-empty voxels of the voxel feature, $\\mathbf{V}_0$ , are colored." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 898, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the box proposal. Given the bounding box proposals, $B_0$ , at active timestamp 0, we divide box proposals into $S \\times S \\times S$ regular sub-voxels. Then, we independently perform the ROI pooling [15] on the virtual 3D event features $\\mathbf{V}_t^E$ , and the voxel features $\\mathbf{V}_0$ , corresponding to each subvoxel. This generates grid features for the bounding boxes, denoted as $\\tilde{\\mathbf{V}}_t^E \\in \\mathbb{R}^{n \\times S^3 \\times C}$ and $\\tilde{\\mathbf{V}}_0 \\in \\mathbb{R}^{n \\times S^3 \\times C}$ , where $n$ is the number of box proposals. The generated features are then concatenated and passed through an MLP to produce the implicit motion field, $\\mathcal{M}_t = \\{\\mathcal{M}_t^0, \\mathcal{M}_t^1, \\dots, \\mathcal{M}_t^n\\}$ . The implicit motion field is created by aligning 3D events with their corresponding voxel features, encapsulating motion information for the bounding box. Using an MLP, we estimate the explicit 3D motion. The explicit box motion $\\mathbf{M}_{0 \\rightarrow t}^i = (dx^i, dy^i, dz^i, d\\alpha^i)$ , defined in box local coordinate of $B_0^i$ , contains spatial shift and yaw rotation. The final box location is calculated as $B_t^i = g(B_0^i, \\mathbf{M}_{0 \\rightarrow t}^i)$ , where $g$ is a box shifting operation.", + "bbox": [ + 511, + 422, + 906, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Motion Confidence Estimator", + "text_level": 1, + "bbox": [ + 511, + 710, + 777, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in previous studies [114, 125], self-awareness of the confidence of predicted bounding boxes is a key factor. The simplest choice of final confidence score in our setting would be the score of active timestamp 0. However, the final score should take into account the object's motion prediction during the blind time. If the object experiences challenging motion during the blind time, a reliable bounding box at the active timestamp 0 could become less reliable. Therefore, the final score should consider two different sources. The first aspect is the score estimated at the active timestamp, based on LiDAR and images. The second", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27200", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e8306ee1b72f89ba50f37e89d5338ba97ff78a409e54ff2fa9f2db5ad88e9618.jpg", + "image_caption": [ + "Figure 4. Ev-Waymo Dataset samples. Synthetic event data, label overlaid image, and LiDAR from left to right." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 477, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aspect is the confidence score based on motion during the blind time using events.", + "bbox": [ + 89, + 213, + 482, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The implicit motion field from Sec. 3.3 contains 3D motion information through alignment between the point cloud and events. We additionally employ an MLP specifically for calculating motion confidence. Box-wise motion confidence is estimated and normalized, resulting in $p_{0\\rightarrow t} = \\{p_{0\\rightarrow t}^{1},p_{0\\rightarrow t}^{2},\\dots ,p_{0\\rightarrow t}^{n}\\} \\in [0,1]$ . The final confidence score for box proposal $i$ is calculated by considering both the score at the active timestamp and the motion confidence:", + "bbox": [ + 89, + 243, + 483, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} ^ {i} = p _ {0} ^ {i} \\cdot p _ {0 \\rightarrow t} ^ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 378, + 480, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Objectives", + "text_level": 1, + "bbox": [ + 89, + 402, + 279, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For RGB-LiDAR RPN in the active timestamp, we follow RPN loss from [47, 105, 114]. For blind time prediction, box regression loss from [114] is applied between the predicted 3D box $B_{t}^{i}$ and the ground truth $\\hat{B}_{t}^{i}$ . The target assigned to the motion confidence branch is an IoU-related value, as:", + "bbox": [ + 89, + 425, + 483, + 515 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {0 \\to t} ^ {i} (\\mathrm {I o U} _ {i}) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f I o U _ {i} < \\theta_ {L}}, \\\\ \\frac {\\mathrm {I o U} _ {i} - \\theta_ {L}}{\\theta_ {H} - \\theta_ {L}} & \\text {i f \\theta_ {L} \\leq I o U _ {i} < \\theta_ {H}}, \\\\ 1 & \\text {i f I o U _ {i} \\geq \\theta_ {H}}, \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 523, + 444, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathrm{IoU}_i$ is the IoU between the $i$ -th prediction and the corresponding ground truth box, while $\\theta_H$ and $\\theta_L$ are foreground and background IoU thresholds. Binary Cross Entropy Loss is exploited here for confidence prediction. The losses of our detect head are computed as:", + "bbox": [ + 89, + 589, + 483, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {R P N} + \\lambda_ {1} \\cdot \\mathcal {L} _ {r e g} + \\lambda_ {2} \\cdot \\mathcal {L} _ {s c o r e} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 680, + 482, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Event-based 3D Object Detection Datasets", + "text_level": 1, + "bbox": [ + 89, + 708, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This study is the first to explore the use of event cameras for 3D detection. To validate our model's effectiveness, we introduce two event 3D detection datasets: Ev-Waymo, which consists of synthetic events, and DSEC-3DOD with real events. We first obtained accurate 3D bounding box annotations for the active time, when both LiDAR and image data are available, and then processed the blind time.", + "bbox": [ + 89, + 733, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Annotation for Active Time at 10 FPS", + "text_level": 1, + "bbox": [ + 89, + 848, + 416, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ev-Waymo Dataset is a synthetic event dataset based on the Waymo Open Dataset (WOD) [93]. WOD provides", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad2e992570a22964759e6efed473190c28a1a597f9c7e478e059765260d59afc.jpg", + "image_caption": [ + "Figure 5. DSEC-3DOD Dataset samples. Event data, label overlaid image, and accumulated LiDAR from left to right." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 893, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sequential pairs of synchronized LiDAR, image data, and 3D box annotations at 10 FPS. As shown in Fig. 4, event streams are generated using a widely-adopted event simulator, Vid2E [21]. The Ev-Waymo evaluation follows the same metrics as WOD, calculating both average precision (AP) and average precision weighted by heading (APH). Adhering to the DSEC-3DOD setting, Ev-Waymo employs the front camera and a cropped LiDAR aligned to the corresponding field of view.", + "bbox": [ + 511, + 296, + 905, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DSEC-3DOD. The DSEC dataset [23] is a stereo event dataset that includes LiDAR, image, and real event data captured in challenging environments. We sampled a portion of the \"zurich_city\" sequence, providing LiDAR and image data at 10 FPS. We hired annotation experts to label 3D bounding boxes at 10 FPS, ensuring alignment with the sensor data. As shown in Fig. 5, DSEC-3DOD contains real events, incorporating noise and capturing challenging environments, unlike simulated events in Ev-Waymo. During the annotation process, we accumulated LiDAR points to precisely generate ground truth 3D bounding boxes.", + "bbox": [ + 511, + 431, + 905, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation on DSEC-3DOD, we used the same evaluation metrics as Ev-Waymo but include only vehicle and pedestrian classes due to a lack of cyclist data, with all objects categorized under a single difficulty level.", + "bbox": [ + 511, + 598, + 905, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Annotation for Blind Time at 100 FPS", + "text_level": 1, + "bbox": [ + 511, + 667, + 841, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For training and evaluation on 3D detection during blind time, we created 3D box annotations for periods when Li-DAR and images were absent. We linearly interpolated the original 10 FPS annotations, dividing blind times into 10 evenly spaced intervals between consecutive frames, resulting in temporally dense annotations at 100 FPS. To ensure accurate bounding box ground truths in instances where Li-DAR and image data are unavailable, we employed event-based video frame interpolation [39] and point cloud interpolation [123] methods to generate sensor data for blind time. Data annotation experts refined interpolated bounding boxes based on the generated sensor data. The Ev-Waymo dataset consists of 80 sequences comprising a total of 157K labeled scenes, providing 100 FPS detection ground truth.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27201", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a963eee106f7d129f65437d7a699a469dfbf09aafc729111f8ab7eba122d2816.jpg", + "table_caption": [ + "Table 1. Performance comparison on the Ev-Waymo dataset for 3D vehicle (IoU = 0.7), pedestrian (IoU = 0.5) and cyclist (IoU = 0.5) detection following previous protocols. Evaluated on 100 FPS ground truth annotation for $t = \\{0,0.1,\\dots,0.9\\}$ . Offline represents results obtained by interpolating data using sensor data from active timestamps 0 and 1. Online represents results obtained using data accessible at present. Specifically, online evaluation of previous works uses only the data from active timestamp 0, while we also utilize event data from active timestamp 0 up to the present time." + ], + "table_footnote": [], + "table_body": "
Interpolation MethodsMethods3D Detection ModalityALL (mAP/mAPH)VEH (AP/APH)PED (AP/APH)CYC (AP/APH)
L2L1L2L1L2L1L2
OfflineN-PCI [123]VoxelNeXt [9]L53.61/50.6762.96/62.5159.10/58.6768.43/60.7762.85/55.6944.55/43.1538.88/37.65
HEDNet [118]L50.52/47.5159.57/59.0955.90/55.4563.28/55.3258.03/50.6243.14/41.8037.63/36.46
N-PCI [123] + EMA [116]Focals Conv [8]L + I42.53/40.4155.49/55.0254.45/53.9948.07/43.0646.54/41.6927.70/26.6026.59/25.54
LoGoNet [47]L + I53.29/50.4963.48/63.0560.15/59.7367.85/60.2862.43/55.3742.40/41.3737.29/36.38
N-PCI [123] + CBM [39]Focals Conv [8]L + I42.51/40.3955.48/55.0154.44/53.9848.07/43.0646.54/41.6827.67/26.5726.56/25.50
LoGoNet [47]L + I53.57/50.7563.60/63.1660.27/59.8568.14/60.5362.77/55.6742.85/41.8137.66/36.74
OnlineN/AVoxelNeXt [9]L33.32/31.7044.40/44.1041.78/41.4940.52/36.2336.93/32.9624.40/23.7321.24/20.66
HEDNet [118]L31.57/29.9042.03/41.7139.32/39.0238.86/34.4335.67/31.5322.64/21.9919.72/19.14
Focals Conv [8]L + I26.27/25.0137.31/37.0136.60/36.3129.20/26.1628.41/25.4414.30/13.7813.79/13.29
LoGoNet [47]L + I33.27/31.7544.14/43.8741.73/41.4739.98/35.8436.48/32.6724.71/24.1521.59/21.10
Ev-3DOD (Ours)L + I + E48.06/45.6060.30/59.9559.19/58.8557.40/50.7855.30/48.9331.08/30.3829.69/29.03
", + "bbox": [ + 99, + 161, + 898, + 339 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bd631d55e2438c9214c8979624d0d0cb5c08a7926a5e6d4e5f5059cbc2149882.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons of our method with other offline and online evaluations on the Ev-Waymo dataset. $t = 0$ represents the active time, while $t = 0.3, 0.6, 0.9$ denotes the blind times. For easier understanding, images at active timestamps 0 and 1 are overlaid." + ], + "image_footnote": [], + "bbox": [ + 96, + 349, + 897, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The DSEC 3D dataset consists of 178 sequences and includes 54K labeled scenes. For evaluation in blind time, we utilize all ground truth boxes at 100 FPS. More details of the datasets are described in the supplementary material.", + "bbox": [ + 89, + 681, + 482, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 758, + 223, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 785, + 266, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We target causal 3D detections to detect objects during blind time. For comparison, we present two types of evaluation methods for conventional models. First is the online evaluation, where the network is inferred at any given blind time using only accessible data. Conventional methods can only process the most recent sensor data at active timestamp", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "0. In contrast, our approach additionally incorporates event data from active timestamp 0 up to the current blind time. Second, offline evaluation addresses the challenge of applying conventional fixed frame rate approaches directly to blind time by utilizing future data. In other words, data for the blind time is interpolated using active timestamps 0 and 1, where timestamp 1 represents future data. Consequently, these offline evaluations serve as a performance oracle for online methods. More implementation details are provided in the supplementary material.", + "bbox": [ + 511, + 681, + 906, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 845, + 717, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ recent state-of-the-art methods, including LiDAR-based [9, 118] and LiDAR-camera-based meth", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27202", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b768e3601e10ea993b74d5807fb2bfe7af21b507c185309a39aa54a451f17651.jpg", + "table_caption": [ + "Table 2. Results on the DSEC-3DOD dataset for 3D vehicle (IoU = 0.7), and pedestrian (IoU = 0.5) detection." + ], + "table_footnote": [], + "table_body": "
Interpolation MethodsMethods3D Detection ModalityALLVEHPED
mAPmAPHAPAPHAPAPH
OfflineN-PCI [123]VoxelNeXt [9]L36.2932.4345.3544.9327.2319.93
HEDNet [118]L37.7330.7341.8241.3533.6420.1
N-PCI [123] + EMA [116]Focals Conv [8]L + I30.2925.8535.5235.0425.0516.65
LoGoNet [47]L + I37.3031.3944.6444.0629.9618.72
N-PCI [123] + CBM [39]Focals Conv [8]L + I30.2825.8635.5635.0824.9916.64
LoGoNet [47]L + I37.3531.4544.7644.1729.9418.73
OnlineN/AVoxelNeXt [9]L13.9412.9021.4221.236.464.57
HEDNet [118]L13.7511.6818.4818.299.025.07
Focals Conv [8]L + I10.819.6115.7415.555.883.67
LoGoNet [47]L + I13.2911.9119.5019.327.074.49
Ev-3DOD (Ours)L + I + E31.1726.5441.6541.2020.6811.88
", + "bbox": [ + 99, + 104, + 898, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d785061edc9d593e3dd21e232b1527bbee327bb8292696d8953f11e990f14ac4.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons of our method with other offline and online evaluations on the DSEC-3DOD dataset. $t = 0.3, 0.6, 0.9$ denotes the blind times." + ], + "image_footnote": [], + "bbox": [ + 94, + 280, + 477, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods [8, 47]. For offline image interpolation, we use frame-based, EMA [116], and event-based method, CBM [39]. NeuralPCI [123] is adopted for point cloud interpolation.", + "bbox": [ + 89, + 590, + 482, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ev-Waymo. Table 1 shows the quantitative result of Ev-Waymo. Comparing methods in the online setting, previous methods [8, 9, 47, 118], which rely on fixed frame rate sensors (i.e., LiDAR and Image), cannot detect object motion during blind times. This limitation leads to reduced performance when evaluated with 100 FPS ground truth. In contrast, our approach leverages high-temporal resolution of event data to estimate object motion during blind time, significantly outperforming existing online methods. Notably, Ev-3DOD is comparable to offline oracles and even surpasses certain approaches, such as the combination of NeuralPCI [123] + EMA [116] and Focals Conv [8].", + "bbox": [ + 89, + 640, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 shows comparisons between the results of online, offline, and our method during blind time. Existing online methods fail to align prediction with the ground truth during blind time. In contrast, Ev-3DOD, despite being an online approach, aligns well with the ground truth and", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca34d303fc19602b3545c1e455ac03566705a207869623df6ba391ed6260a3f9.jpg", + "table_caption": [ + "Table 3. Ablation study of the proposed components. V3D-EF and MCE denote Virtual 3D Event Fusion (Sec. 3.3) and the motion confidence estimator (Sec. 3.4), respectively." + ], + "table_footnote": [], + "table_body": "
V3D-EFMCEL2
mAPmAPH
33.3031.68
✓46.5544.22
✓✓48.0645.60
", + "bbox": [ + 571, + 327, + 848, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstrates performance comparable to offline methods. The right graph in Fig. 6 further demonstrates our method's superiority, showing that despite performance degradation with increasing blind time, the gap remains comparable to offline methods, while the online approach exhibits a much larger pixel difference.", + "bbox": [ + 511, + 412, + 906, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DSEC-3DOD. As shown in Table 2, we evaluate our model on the DSEC-3DOD real event dataset, which includes fast motion and challenging lighting conditions that significantly degrade performance during blind time. Our model effectively estimates the 3D motion of objects from events, achieving a substantial improvement over other online methods. Notably, it demonstrates performance comparable to offline oracle methods that utilize future information. The analysis of computational complexity is provided in the supplementary material.", + "bbox": [ + 511, + 503, + 905, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7 shows blind time detection inference for fast-moving objects in the DSEC dataset. Under rapid motion, the quality of the interpolated pseudo-sensor data may deteriorate, occasionally resulting in detection failures, even when future information is utilized. In contrast, the proposed Ev-3DOD accurately detects objects in blind time, predicting motion from events.", + "bbox": [ + 511, + 655, + 905, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 511, + 771, + 769, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present an ablation study on key components and various analyses on the Ev-Waymo dataset. In V3D-EF analyses, the motion confidence estimator (MCE) module is excluded for clear comparisons.", + "bbox": [ + 511, + 794, + 905, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Each Component. In Table 3, we conduct the ablation study on the key components of Ev-3DOD. We evaluated a baseline and progressively added modules to", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/520be6e819998ea0cab577bec34c45b92aff99da68caaf7a658aecfe91469082.jpg", + "table_caption": [ + "Table 4. Effectiveness of the non-empty mask in the V3D-EF module, as defined in Eq. (1)." + ], + "table_footnote": [], + "table_body": "
MethodL2
mAPmAPH
w/o Non-empty Mask42.5740.50
w/ Non-empty Mask46.5544.22
", + "bbox": [ + 153, + 119, + 426, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d212269a88b8173a1d682cf24b06a19b8d893e9cce13606db12614898e70dff1.jpg", + "table_caption": [ + "Table 5. Comparison of results across different modalities and fusion variants of the V3D-EF module." + ], + "table_footnote": [], + "table_body": "
MethodL2
mAPmAPH
Base Model33.3031.68
+ LiDAR37.0435.25
+ Event42.2340.17
+ LiDAR + Event (fusion before pooling)43.9541.87
+ LiDAR + Event (fusion after pooling)46.5544.22
", + "bbox": [ + 106, + 223, + 470, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ae3125fd5d0942b66963a5f76085387bd963d136aaf524c2db6f185e9646496.jpg", + "image_caption": [ + "Figure 8. Detection performance as time elapses since the active timestamp. We compare ours with offline and online approaches." + ], + "image_footnote": [], + "bbox": [ + 98, + 342, + 473, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "observe improvements. The motion confidence estimator is dependent on V3D-EF, so it was not evaluated separately. V3D-EF, which explicitly estimates motion, has a direct impact on performance, resulting in significant improvements of $+13.25\\%$ in mAP and $+12.54\\%$ in mAPH. Additionally, the motion confidence estimator refines prediction scores at each blind time, demonstrating its effectiveness with further gains of $+1.51\\%$ and $+1.38\\%$ , respectively.", + "bbox": [ + 88, + 518, + 482, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Non-empty Mask in V3D-EF. In the proposed V3D-EF, we employ a non-empty mask to achieve accurate and efficient alignment between sparse events and voxel features. To demonstrate the effectiveness of this non-empty mask, we conduct a comparative experiment in which events are projected onto all voxels within the V3D-EF module. The results in Table 4 confirm that the non-empty mask enables effective alignment between the two modalities.", + "bbox": [ + 88, + 642, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Variations of the V3D-EF Module. Our core component, the V3D-EF module, provides multiple design options. To analyze these configurations, we developed and evaluated several variants, with the results summarized in Table 5. We experimented with generating the implicit motion field using each modality, i.e., LiDAR and event sensors, and found a performance drop compared to the multi-modal approach. This decline is likely due to the limitations of each modality: LiDAR alone cannot capture motion information during", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dda0b179bad76df24d0a1e645aa931daa0aeb96a2c0bec24b0238cdd7253157b.jpg", + "table_caption": [ + "Table 6. Evaluation results of the model trained with 100 FPS annotations using sliced events at 200 and 300 FPS." + ], + "table_footnote": [], + "table_body": "
Test SetL2
mAPmAPH
100 FPS48.0645.60
200 FPS47.8745.44
300 FPS47.8145.39
", + "bbox": [ + 588, + 119, + 834, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "blind times, and event features cannot accurately estimate refined 3D motion. Additionally, multi-modal fusion before pooling decreased performance. We believe this is because fusing before pooling leads to integration between voxels and events outside of box proposal regions, resulting in the use of redundant and possibly misaligned information.", + "bbox": [ + 511, + 205, + 906, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection Performance per Elapsed Blind Time. Figure 8 shows the performance over time elapsed since the active timestamp. Due to object motion, the performance of online detection deteriorates rapidly as the time distance from the active timestamp increases. In contrast, offline methods sustain robust performance throughout the blind time by leveraging future data, with only a slight performance drop at the center of the blind time as it moves further from the active timestamps located at both ends. The proposed model, relying solely on event data, achieves a temporal degradation pattern comparable to the offline method.", + "bbox": [ + 511, + 296, + 908, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to Higher Frame-rate. To demonstrate the generalization ability to different frame rates, we generate 200 and 300 FPS annotations for arbitrary timestamps by interpolating a pre-constructed 100 FPS dataset. Without additional training, we apply the model trained on 100 FPS. As shown in Table 6, our model effectively infers at arbitrary timestamps with minimal degradation.", + "bbox": [ + 511, + 464, + 908, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 585, + 640, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes Ev-3DOD, a novel method to utilize an event camera in 3D detection for detecting objects in blind time. To transfer information from active times to the current blind time, we estimate 3D motion based on event data. To effectively fuse sparse LiDAR data with events, we propose a Virtual 3D Event Fusion (V3D-EF) and introduce a motion confidence estimator to define confidence for 3D detection during blind times. For this study, we introduce the first event-based 3D detection dataset, emphasizing its 100 FPS annotations. We hope this work showcases the potential of neuromorphic cameras, inspiring future research.", + "bbox": [ + 511, + 612, + 906, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. Our work is the first to address 3D detection during the blind time, establishing it as a pioneering study in this area. However, one limitation is the lack of recurrent processing of past event information to enhance temporal interaction. As a result, discontinuities may occur when fixed frame rate sensors update at active timestamps. We consider improving the interaction between consecutive event data a crucial direction for future work.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27204", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank 42dot for funding this research. We also appreciate the support and valuable discussions from the members of the 42dot research team. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1090-1099, 2022. 2", + "[2] Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Bourtsoulatze, and Yiannis Andreopoulos. Graph-based object classification for neuromorphic vision sensing. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 491-501, 2019. 3", + "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2", + "[4] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2", + "[5] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 172-181, 2022. 2", + "[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2", + "[7] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2", + "[8] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5428-5437, 2022. 6, 7", + "[9] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 6, 7", + "[10] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 887-897, 2022. 2" + ], + "bbox": [ + 99, + 243, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Deformable feature aggregation for dynamic multi-modal 3d object detection. In European conference on computer vision, pages 628-644. Springer, 2022. 2", + "[12] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2", + "[13] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2", + "[14] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In European Conference on Computer Vision, pages 294–314. Springer, 2024. 2", + "[15] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2, 4", + "[16] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8458-8468, 2022. 2", + "[17] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1761-1769, 2024. 2", + "[18] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44 (1):154-180, 2020. 2", + "[19] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. ArXiv, abs/2211.12324, 2022. 3", + "[20] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629:1034 - 1040, 2024. 3", + "[21] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 5", + "[22] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13884-13893, 2022. 2, 3", + "[23] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 5", + "[24] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[25] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 2", + "[26] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3153-3163, 2021. 2", + "[27] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22867-22876, 2023. 3", + "[28] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11873-11882, 2020. 2", + "[29] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8417-8427, 2022. 2", + "[30] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2", + "[31] Cheng-Ju Ho, Chen-Hsuan Tai, Yen-Yu Lin, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Diffusion-ss3d: Diffusion model for semi-supervised 3d object detection. Advances in Neural Information Processing Systems, 36:49100-49112, 2023. 2", + "[32] Jordan SK Hu, Tianshu Kuai, and Steven L Waslander. Point density-aware voxels for lidar 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8469-8478, 2022. 2, 4", + "[33] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2", + "[34] Junjie Huang, Yun Ye, Zhujin Liang, Yi Shan, and Dalong Du. Detecting as labeling: Rethinking lidar-camera fusion in 3d object detection. In European Conference on Computer Vision, pages 439-455. Springer, 2025. 2", + "[35] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2", + "[36] Yu Huang, Chenzhuang Du, Zihui Xue, Xuanyao Chen, Hang Zhao, and Longbo Huang. What makes multi-modal" + ], + "bbox": [ + 99, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning better than single (provably). Advances in Neural Information Processing Systems, 34:10944-10956, 2021. 2", + "[37] Massimiliano Iacono, Stefan Weber, Arren J. Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9, 2018. 3", + "[38] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformer. In Proceedings of the AAAI conference on Artificial Intelligence, pages 1042-1050, 2023. 2", + "[39] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 5, 6, 7", + "[40] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2", + "[41] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2", + "[42] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2", + "[43] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2", + "[44] Dianze Li, Jianing Li, and Yonghong Tian. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:14020-14037, 2023. 3", + "[45] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 3", + "[46] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2", + "[47] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 2, 3, 5, 6, 7", + "[48] Xiaotian Li, Baojie Fan, Jiandong Tian, and Huijie Fan. Gafusion: Adaptive fusing lidar and camera with multiple guidance for 3d object detection. In Proceedings of" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27206", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21209-21218, 2024. 2", + "[49] Xiang Li, Junbo Yin, Wei Li, Chengzhong Xu, Ruigang Yang, and Jianbing Shen. Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3208-3215, 2024. 2", + "[50] Yijin Li, Han Zhou, Bangbang Yang, Yexin Zhang, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Graph-based asynchronous event processing for rapid object recognition. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 914-923, 2021. 3", + "[51] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems, 35:18442-18455, 2022. 2", + "[52] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022.", + "[53] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17182-17191, 2022. 2", + "[54] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2", + "[55] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7546-7555, 2021. 2", + "[56] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2", + "[57] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7345-7353, 2019. 2", + "[58] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 2", + "[59] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2" + ], + "bbox": [ + 99, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Xianpeng Liu, Ce Zheng, Ming Qian, Nan Xue, Chen Chen, Zhebin Zhang, Chen Li, and Tianfu Wu. Multi-view attentive contextualization for multi-view 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16688-16698, 2024. 2", + "[61] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022.", + "[62] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 996-997, 2020. 2", + "[63] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2", + "[64] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 2", + "[65] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2723-2732, 2021. 2", + "[66] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Ji-ashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3164–3173, 2021. 2", + "[67] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. ArXiv, abs/2003.09148, 2020. 3", + "[68] Zhenwei Miao, Jikai Chen, Hongyu Pan, Ruiwen Zhang, Kaixuan Liu, Peihan Hao, Jun Zhu, Yang Wang, and Xin Zhan. Pvgnet: A bottom-up one-stage 3d object detector with integrated multi-level features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3279-3288, 2021. 2", + "[69] Su Pang, Daniel Morris, and Hayden Radha. Clocs: Camera-lidar object candidates fusion for 3d object detection. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10386-10393. IEEE, 2020. 2", + "[70] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 2", + "[71] Wanli Peng, Hao Pan, He Liu, and Yi Sun. Ida-3d: Instance-depth-aware 3d object detection from stereo vision" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13015-13024, 2020. 2", + "[72] Yan Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In AAAI Conference on Artificial Intelligence, 2023. 3", + "[73] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6015-6025, 2023. 3", + "[74] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16794-16804, 2024. 3", + "[75] Etienne Perot, Pierre de Tournemire, Davide Oscar Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. *ArXiv*, abs/2009.13436, 2020. 3", + "[76] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2", + "[77] AJ Piergiovanni, Vincent Casser, Michael S Ryoo, and Anelia Angelova. 4d-net for learned multi-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15435-15445, 2021. 2", + "[78] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2", + "[79] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[80] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2", + "[81] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2", + "[82] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2", + "[83] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2" + ], + "bbox": [ + 99, + 90, + 483, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[85] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2743-2752, 2021. 2", + "[86] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2", + "[87] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2", + "[88] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1711-1719, 2020. 2", + "[89] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2", + "[90] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 2", + "[91] Jiaming Sun, Linghao Chen, Yiming Xie, Siyu Zhang, Qin-hong Jiang, Xiaowei Zhou, and Hujun Bao. Disp r-cnn: Stereo 3d object detection via shape prior guided instance disparity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10548-10557, 2020. 2", + "[92] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2", + "[93] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2443-2451, 2019. 5", + "[94] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "27208", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2", + "[96] Dongsheng Wang, Xu Jia, Yang Zhang, Xinyu Zhang, Yaoyuan Wang, Ziyang Zhang, D. Wang, and Huchuan Lu. Dual memory aggregation network for event-based object detection with learnable representation. In AAAI Conference on Artificial Intelligence, 2023. 3", + "[97] Wenjie Wang, Yehao Lu, Guangcong Zheng, Shuigen Zhan, Xiaqing Ye, Zichang Tan, Jingdong Wang, Gaoang Wang, and Xi Li. Bevspread: Spread voxel pooling for bird's-eye-view representation in vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14718-14727, 2024. 2", + "[98] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2", + "[99] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2", + "[100] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 2", + "[101] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multimodal sparse representations for multi-sensor 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 17545-17556, 2023. 2", + "[102] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 2", + "[103] Jianyun Xu, Zhenwei Miao, Da Zhang, Hongyu Pan, Kaixuan Liu, Peihan Hao, Jun Zhu, Zhengyang Sun, Hongmin Li, and Xin Zhan. Int: Towards infinite-frames 3d detection with an efficient framework. In European Conference on Computer Vision, pages 193–209. Springer, 2022. 2", + "[104] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Zhou Bin, and Liangjun Zhang. Fusionpainting: Multimodal fusion with adaptive attention for 3d object detection. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3047-3054. IEEE, 2021. 2", + "[105] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 5", + "[106] Yixin Yang, Jinxiu Liang, Bohan Yu, Yan Chen, Jimmy S Ren, and Boxin Shi. Latency correction for event-guided" + ], + "bbox": [ + 93, + 92, + 483, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deblurring and frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24977-24986, 2024. 2", + "[107] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 2", + "[108] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2", + "[109] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. Advances in Neural Information Processing Systems, 35:1992-2005, 2022. 2", + "[110] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhao-Xu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10201-10210, 2021. 3", + "[111] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022. 2", + "[112] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In European conference on computer vision, pages 17-33. Springer, 2022. 2", + "[113] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 2", + "[114] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 4, 5", + "[115] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kili-ian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. arXiv preprint arXiv:1906.06310, 2019. 2", + "[116] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 6, 7", + "[117] Gang Zhang, Junnan Chen, Guohuan Gao, Jianmin Li, Si Liu, and Xiaolin Hu. Safdnet: A simple and effective network for fully sparse 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14477-14486, 2024. 2", + "[118] Gang Zhang, Chen Junnan, Guohuan Gao, Jianmin Li, and Xiaolin Hu. Hednet: A hierarchical encoder-decoder net" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "27209", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "work for 3d object detection in point clouds. Advances in Neural Information Processing Systems, 36, 2024. 6, 7", + "[119] Jiqing Zhang, B. Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8791-8800, 2022. 3", + "[120] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9781-9790, 2023. 2", + "[121] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2", + "[122] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 2", + "[123] Zehan Zheng, Danni Wu, Ruisi Lu, Fan Lu, Guang Chen, and Changjun Jiang. Neuralpci: Spatio-temporal neural field for 3d point cloud multi-frame non-linear interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 909–918, 2023. 5, 6, 7", + "[124] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1839-1849, 2020. 2", + "[125] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2, 3, 4", + "[126] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3", + "[127] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12800-12810, 2023. 3" + ], + "bbox": [ + 91, + 92, + 482, + 767 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "27210", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_model.json b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b73b85b02071facbb588f4250e30b7fcb45345b --- /dev/null +++ b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_model.json @@ -0,0 +1,3033 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.153, + 0.13, + 0.848, + 0.175 + ], + "angle": 0, + "content": "Ev-3DOD: Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.204, + 0.771, + 0.238 + ], + "angle": 0, + "content": "Hoonhee Cho*, Jae-Young Kang*, Youngho Kim, and Kuk-Jin Yoon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.242, + 0.719, + 0.258 + ], + "angle": 0, + "content": "{gnsgnsgml,mickeykang,kmax2001,kjyoon}@kaist.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.28, + 0.368, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.288, + 0.608, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.288, + 0.88, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.509, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Figure 1. Comparison between the proposed Ev-3DOD and conventional 3D detection methods. Fixed frame rate sensors (e.g., LiDAR, frame camera) inevitably have periods of \"blind time\" \\((t_{0\\rightarrow 1})\\) when they lack data between active timestamps (0 and 1). Consequently, as shown in (a), conventional methods cannot detect objects during blind time, while other objects like vehicles can undergo significant motion within these gaps. In contrast, as illustrated in (b), our method, Ev-3DOD, leverages the high temporal resolution of the event camera, enabling accurate object detection even during the blind times of other sensors." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.592, + 0.33, + 0.608 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Detecting 3D objects in point clouds plays a crucial role in autonomous driving systems. Recently, advanced multimodal methods incorporating camera information have achieved notable performance. For a safe and effective autonomous driving system, algorithms that excel not only in accuracy but also in speed and low latency are essential. However, existing algorithms fail to meet these requirements due to the latency and bandwidth limitations of fixed frame rate sensors, e.g., LiDAR and camera. To address this limitation, we introduce asynchronous event cameras into 3D object detection for the first time. We leverage their high temporal resolution and low bandwidth to enable high-speed 3D object detection. Our method enables detection even during inter-frame intervals when synchronized data is unavailable, by retrieving previous 3D information through the event camera. Furthermore, we introduce the first event-based 3D object detection datasets, EvWaymo and DSEC-3DOD, both of which include ground" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.909, + 0.655 + ], + "angle": 0, + "content": "truth 3D bounding boxes at 100 FPS, establishing the first benchmarks for event-based 3D detectors. The code and dataset are available at https://github.com/mickeykang16/Ev3DOD." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.685, + 0.646, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.909, + 0.877 + ], + "angle": 0, + "content": "3D object detection is a crucial task in autonomous driving. Multi-modal sensors like LiDAR and cameras are commonly used to understand scenes and handle complex road scenarios. This sensor setup consists of synchronized sensors with a fixed frame rate and has limited speed due to bandwidth limitations. As shown in Fig. 1 (a), the blind time, which is the period when sensor information is unavailable (e.g., 100 milliseconds), is critical as cars can experience large motions as 3 meters during this time. Therefore, temporal resolution is a crucial factor that can directly influence accident risk." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.759, + 0.902 + ], + "angle": 0, + "content": "*The first two authors contributed equally." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "Recently, event cameras [18] have introduced a new paradigm in perceptual research. Event cameras provide brightness changes as a stream of data with sub-milisecond latency, making them robust to motion blur [41, 106] and enabling low bandwidth [22, 94]. Event cameras overcome the trade-off between bandwidth and blind time but have the drawback of sacrificing intensity information. To address this, recent work [12, 13, 92, 120] has combined event cameras with frame cameras to complement the sparsity of event cameras while overcoming the fixed frame rate sensor's limits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.482, + 0.44 + ], + "angle": 0, + "content": "In this paper, we aim for accurate 3D object detection in blind time through a multi-modal approach that combines LiDAR and frame-based sensors with an event camera, leveraging its high temporal resolution. As shown in Fig. 1 (b), event data is available during blind times, while the LiDAR and image data from the most recent active timestamp are accessible. Therefore, there is a misalignment in timestamps between the fixed frame sensor data and the current blind time. Then, this raises the question: How can we estimate 3D motion using only event data, especially when data from other sensors is unavailable at the current time?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.731 + ], + "angle": 0, + "content": "To address this, we propose a virtual 3D event fusion (V3D-EF) that transfers past 3D information to align with the present time using the event camera. As extensively covered in previous research [14, 24, 40, 89], event data inherently contains motion information on an image plane. Based on this observation, we hypothesized that providing a 3D cue associated with the initial timestamp to the event information would enable accurate estimation of the 3D motion vector. To this end, we project event features into a 3D voxel space to create virtual 3D event features. In this process, events are projected only onto regions where point cloud data exists, aligning with the point cloud data. Then, we aggregate the event features with the voxel features to estimate an implicit motion field that captures the motion inherent within the grid. The generated implicit motion field encapsulates both the 3D spatial information from voxels and the motion information from events, enabling the calculation of a 3D motion vector from the most recent active timestamp to the present blind time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.482, + 0.822 + ], + "angle": 0, + "content": "Another crucial factor in 3D object detection is the score for each bounding box. The model considers both box proposal scores at the active timestamp and the motion score of each predicted box in blind time. In blind time, the model aligns 3D information of voxel features and motion information of events to learn the confidence of box motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": "A major hurdle in 3D detection research using event cameras is the lack of suitable datasets. To address this, we present two datasets: Ev-Waymo, which includes synthetic event data, and DSEC-3DOD, which contains real event data. Unlike existing public 3D object detection datasets," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "ours include annotations not only for active timestamps with 3D LiDAR sensor data but also for blind times, providing annotation in 100 FPS frame rate." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.151, + 0.664, + 0.166 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.176, + 0.905, + 0.417 + ], + "angle": 0, + "content": "Single-modality 3D Object Detection. Given that cameras offer significant cost advantages over LiDAR sensors, many researchers have focused on developing methods that leverage camera-based systems for 3D object detection using image-only inputs [7, 10, 26, 30, 33, 38, 46, 54, 60-62, 64, 71, 91, 97, 99, 115]. Image-based 3D object detection faces challenges due to the lack of direct depth information. To overcome this challenge, several studies [7, 70, 76, 83, 98, 115] have employed depth estimation techniques to generate pseudo 3D point clouds or elevate 2D features into 3D space for object detection, alongside proposals for transformer-based architectures [35, 99, 121] that utilize 3D object queries and 3D-2D correspondences. Nonetheless, accurately estimating depth from images remains difficult, resulting in image-based methods performing worse than LiDAR-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.905, + 0.614 + ], + "angle": 0, + "content": "LiDAR-based 3D object detection can reliably estimate 3D bounding boxes using point clouds captured from LiDAR sensors. Current LiDAR-based detection methods can be divided into three categories based on different point cloud encoding formats: point-based methods [17, 32, 79, 80, 82, 86, 88, 108, 112, 124], voxel-based methods [4, 15, 16, 25, 29, 29, 31, 43, 49, 55, 65, 66, 85, 100, 103, 111, 117, 125], and point-voxel fusion networks [6, 28, 68, 87, 107]. While LiDAR detection is advantageous in various conditions, it often struggles in areas with sparse LiDAR data. Therefore, integrating the geometric benefits of point clouds with the semantic richness of images remains crucial for improving performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Multi-modality 3D Object Detection. Multi-modal sensor fusion [5, 34, 36, 47, 48, 56, 57, 59, 90, 101, 102, 113], where different sensors complement each other, typically utilizes a combination of cameras and LiDAR. Multimodal 3D object detection has gained attention for improving accuracy over unimodal methods. Fusion approaches are categorized into early [3, 11, 42, 81, 95, 104], middle [51-53, 58, 63, 77, 78, 109, 122], and late fusions [1, 47, 69], with middle fusion now preferred due to its robustness to calibration errors and enhanced feature interaction. Crossmodal fusion has achieved significant performance improvements by enhancing the sparse features of point clouds with semantic information from images. However, there are still unexplored areas concerning safety in autonomous driving. LiDAR and cameras have limited time resolution (i.e., \\(10 - 20\\mathrm{Hz}\\)) due to the relatively high bandwidths. To enhance safety, algorithms that operate at high speeds are essential. To address this, we introduce an asynchronous event camera with high temporal resolution for 3D object" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27198" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.087, + 0.885, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.336, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline for event-based 3D object detection (Ev-3DOD). (a): At active timestamp 0, LiDAR and image data are available. Therefore, we utilize an RGB-LiDAR Region Proposal Network (RPN) to extract voxel features and 3D bounding boxes at the active timestamp. (b): To predict the bounding box during blind time, \\(0 \\leq t < 1\\), we estimate the 3D motion and confidence score using event features. For computational efficiency, we design the process to compute (a) only once before the next active timestamp while performing iterative computations solely for (b)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.433, + 0.482, + 0.492 + ], + "angle": 0, + "content": "detection for the first time. The proposed method performs multi-modal fusion between synchronized sensors and asynchronous events to increase the operational frequency of 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.493, + 0.484, + 0.781 + ], + "angle": 0, + "content": "Event-based Object Detection. Event cameras break through the limitations of bandwidth requirements and perceptual latency with the trade-off of sparse data and loss of intensity information. To make the most of these characteristics, prior work has designed models to adapt to event data using graph neural networks [2, 19, 20, 50, 84], spiking neural networks [110, 119], and sparse neural networks [67, 74]. Recently, effective dense neural networks [27, 37, 45, 72, 75, 96, 127] and transformer architectures [22, 44, 73, 74] have been developed to leverage event cameras, aiming to achieve both low latency and high performance. These advances enable real-time operation with impressive accuracy, demonstrating the practical applicability of event cameras in real-world scenarios. To the best of our knowledge, we are the first to extend event-based object detection from 2D to 3D space. By employing a multimodal fusion approach of the event camera and 3D sensors (i.e., LiDAR), we aim to overcome the latency and persistent interval time barriers of 3D object detections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.182, + 0.809 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.279, + 0.833 + ], + "angle": 0, + "content": "3.1. Problem Definition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "LiDAR and camera data are only available at active timestamps 0 and 1, where 1 represents invalid future information in an online process. Our goal is to predict 3D bounding boxes during \\(0 \\leq t < 1\\), including the blind time" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.907, + 0.51 + ], + "angle": 0, + "content": "when synchronized sensors are unavailable. Given a pair of synchronized 3D points \\( P_0 = \\{(x_l, y_l, z_l)\\}_{l=1}^N \\) and image \\( I_0 \\in \\mathbb{R}^{3 \\times H \\times W} \\) at active timestamp 0, we leverage event stream \\( \\mathcal{E}_{0 \\to t} = \\{(i, j, p, \\tau) | 0 \\leq \\tau < t\\} \\) to detect 3D bounding box set \\( B_t \\) at an arbitrary time \\( 0 \\leq t < 1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.72, + 0.532 + ], + "angle": 0, + "content": "3.2. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.539, + 0.907, + 0.795 + ], + "angle": 0, + "content": "As illustrated in (a) of Fig. 2, at active timestamp 0, both LiDAR and image data are available. RGB-LiDAR Region Proposal Network (RPN) generates bounding box proposals from rich 3D information in active time. We begin by voxelizing the point cloud and generating voxel features, denoted as \\(\\mathbf{V}_0\\in \\mathbb{R}^{D_X\\times D_Y\\times D_Z\\times C}\\), where \\(D_{X}\\), \\(D_{Y}\\), and \\(D_Z\\) represent the grid size, and \\(C\\) denotes the channel dimension, using a 3D backbone [125]. Then, using existing region proposals [114] and refinement networks [47] with image features, we generate bounding box proposals, \\(B_0 = \\{B_0^1,B_0^2,\\dots ,B_0^n\\}\\) and box proposal confidence \\(p_0 = \\{p_0^1,p_0^2,\\dots ,p_0^n\\} \\in [0,1]\\) for the active timestamp 0. RGB-LiDAR-based multi-modal networks have been widely studied, leading to many advanced models capable of accurate estimation at active timestamps. We tuned the model with fewer parameters and a reduced channel dimension to improve efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our emphasis is on achieving precise estimation throughout the blind time, spanning the interval from active timestamp 0 to the subsequent timestamp 1. To infer bounding boxes at an arbitrary inter-frame time \\( t \\), we first convert the event stream, \\( \\mathcal{E}_{0\\rightarrow t} \\) into voxel grids [126] as \\( \\mathbb{E}_t \\). Then, we use a feature encoder to produce event features, \\( \\mathbf{E}_t\\in \\mathbb{R}^{H / 4\\times W / 4\\times C} \\), where \\( H \\) and \\( W \\) is the spa" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27199" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "tial dimensions of the events. Our objective is to estimate the 3D motion of objects during blind time. To achieve this, we utilize the bounding box set \\( B_{0} \\) and voxel features \\( \\mathbf{V}_0 \\) from the active timestamp 0. During blind times, we share the voxel features, \\( \\mathbf{V}_0 \\), generated at the active timestamp, ensuring that voxel features are computed only once to maintain a cost-effective structure. Through the Virtual 3D Event Fusion (V3D-EF), we generate an implicit motion field for each bounding box by integrating event and voxel features, encapsulating motion information. Using this field, we compute the 3D motion and confidence score, enabling the estimation of 3D boxes at the blind time \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.284, + 0.483, + 0.3 + ], + "angle": 0, + "content": "3.3. Virtual 3D Event Fusion for Motion Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.488 + ], + "angle": 0, + "content": "The event feature continuously captures each pixel's motion information over time on the 2D image plane. However, estimating 3D motion using only events is a challenge. We hypothesize that providing each event feature with adequate 3D information would allow for effective 3D motion estimation. To achieve this, we aim to incorporate a 3D cue into the event features. Since both event features and voxel features are sparse, it is crucial to align each voxels with the corresponding events accurately. Directly using the voxel features would cause misalignment with the actual 3D geometry. Thus, we adopt the voxel point centroid method employed in previous works [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.61 + ], + "angle": 0, + "content": "Specifically, we first consider the non-empty voxel features as \\(\\mathbf{V}_0 = \\{V_0^k = \\{h_0^k,f_0^k\\} \\mid k = 1,\\dots ,N_\\nu \\}\\), where \\(h_0^k\\) is the 3D voxel index, \\(f_0^k\\) is the associated voxel feature vector, and \\(N_{\\mathcal{V}}\\) is the number of non-empty voxels. Then, points within each voxel are grouped into a set \\(\\mathcal{N}(V_0^k)\\) by determining their voxel index \\(h_0^k\\) based on their spatial coordinates \\(X_{i} = (x_{i},y_{i},z_{i})\\) of points. The centroid of each voxel feature is subsequently computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.621, + 0.483, + 0.661 + ], + "angle": 0, + "content": "\\[\nc _ {0} ^ {k} = \\frac {1}{\\left| \\mathcal {N} \\left(V _ {0} ^ {k}\\right) \\right|} \\sum_ {X _ {i} \\in \\mathcal {N} \\left(V _ {0} ^ {k}\\right)} X _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.856 + ], + "angle": 0, + "content": "The centroid \\( c_0^k \\) is calculated for non-empty voxels, ensuring it corresponds only to indices with existing points, with the goal of aligning these centroids to the event data. As shown in the top of Fig. 3, we propose a method to project events into a virtual voxel space based on the coordinates of the centroids. The projection onto the image plane from 3D space can be obtained as: \\( \\mathbf{p}_0^k = K \\cdot E \\cdot c_0^k \\), where \\( K \\) and \\( E \\) denote the intrinsic and extrinsic parameters, \\( \\cdot \\) operation is matrix multiplication. Therefore, we calculate the event coordinates corresponding to each centroid, \\( c_0^k \\), and gather the event features, \\( \\mathbf{E}_t \\), at these coordinates, \\( \\mathbf{p}_0^k \\), to create virtual 3D event features, \\( \\mathbf{V}_t^E = \\{\\mathbf{E}_t(\\mathbf{p}_0^k)\\}_{k=1}^{N_\\nu} \\in \\mathbb{R}^{N_\\nu \\times C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Since we aim to estimate the 3D motion of bounding boxes from the active timestamp \\(0\\) up to the current time \\(t\\), we consider fusion only for the features corresponding" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.9, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.317, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Figure 3. Virtual 3D Event Fusion (V3D-EF). We generate an implicit motion field for each bounding box by applying the ROI pooling separately to the virtual 3D event features and the voxel features. Virtual 3D event features represent the 3D projection of 2D events. None-empty voxels of the voxel feature, \\(\\mathbf{V}_0\\), are colored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.907, + 0.697 + ], + "angle": 0, + "content": "to the box proposal. Given the bounding box proposals, \\(B_0\\), at active timestamp 0, we divide box proposals into \\(S \\times S \\times S\\) regular sub-voxels. Then, we independently perform the ROI pooling [15] on the virtual 3D event features \\(\\mathbf{V}_t^E\\), and the voxel features \\(\\mathbf{V}_0\\), corresponding to each subvoxel. This generates grid features for the bounding boxes, denoted as \\(\\tilde{\\mathbf{V}}_t^E \\in \\mathbb{R}^{n \\times S^3 \\times C}\\) and \\(\\tilde{\\mathbf{V}}_0 \\in \\mathbb{R}^{n \\times S^3 \\times C}\\), where \\(n\\) is the number of box proposals. The generated features are then concatenated and passed through an MLP to produce the implicit motion field, \\(\\mathcal{M}_t = \\{\\mathcal{M}_t^0, \\mathcal{M}_t^1, \\dots, \\mathcal{M}_t^n\\}\\). The implicit motion field is created by aligning 3D events with their corresponding voxel features, encapsulating motion information for the bounding box. Using an MLP, we estimate the explicit 3D motion. The explicit box motion \\(\\mathbf{M}_{0 \\rightarrow t}^i = (dx^i, dy^i, dz^i, d\\alpha^i)\\), defined in box local coordinate of \\(B_0^i\\), contains spatial shift and yaw rotation. The final box location is calculated as \\(B_t^i = g(B_0^i, \\mathbf{M}_{0 \\rightarrow t}^i)\\), where \\(g\\) is a box shifting operation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.779, + 0.726 + ], + "angle": 0, + "content": "3.4. Motion Confidence Estimator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As discussed in previous studies [114, 125], self-awareness of the confidence of predicted bounding boxes is a key factor. The simplest choice of final confidence score in our setting would be the score of active timestamp 0. However, the final score should take into account the object's motion prediction during the blind time. If the object experiences challenging motion during the blind time, a reliable bounding box at the active timestamp 0 could become less reliable. Therefore, the final score should consider two different sources. The first aspect is the score estimated at the active timestamp, based on LiDAR and images. The second" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27200" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.478, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.199 + ], + "angle": 0, + "content": "Figure 4. Ev-Waymo Dataset samples. Synthetic event data, label overlaid image, and LiDAR from left to right." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.483, + 0.244 + ], + "angle": 0, + "content": "aspect is the confidence score based on motion during the blind time using events." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.378 + ], + "angle": 0, + "content": "The implicit motion field from Sec. 3.3 contains 3D motion information through alignment between the point cloud and events. We additionally employ an MLP specifically for calculating motion confidence. Box-wise motion confidence is estimated and normalized, resulting in \\( p_{0\\rightarrow t} = \\{p_{0\\rightarrow t}^{1},p_{0\\rightarrow t}^{2},\\dots ,p_{0\\rightarrow t}^{n}\\} \\in [0,1] \\). The final confidence score for box proposal \\( i \\) is calculated by considering both the score at the active timestamp and the motion confidence:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.379, + 0.482, + 0.397 + ], + "angle": 0, + "content": "\\[\np _ {t} ^ {i} = p _ {0} ^ {i} \\cdot p _ {0 \\rightarrow t} ^ {i} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.403, + 0.281, + 0.42 + ], + "angle": 0, + "content": "3.5. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.516 + ], + "angle": 0, + "content": "For RGB-LiDAR RPN in the active timestamp, we follow RPN loss from [47, 105, 114]. For blind time prediction, box regression loss from [114] is applied between the predicted 3D box \\( B_{t}^{i} \\) and the ground truth \\( \\hat{B}_{t}^{i} \\). The target assigned to the motion confidence branch is an IoU-related value, as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.525, + 0.445, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {0 \\to t} ^ {i} (\\mathrm {I o U} _ {i}) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f I o U _ {i} < \\theta_ {L}}, \\\\ \\frac {\\mathrm {I o U} _ {i} - \\theta_ {L}}{\\theta_ {H} - \\theta_ {L}} & \\text {i f \\theta_ {L} \\leq I o U _ {i} < \\theta_ {H}}, \\\\ 1 & \\text {i f I o U _ {i} \\geq \\theta_ {H}}, \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.667 + ], + "angle": 0, + "content": "\\(\\mathrm{IoU}_i\\) is the IoU between the \\(i\\)-th prediction and the corresponding ground truth box, while \\(\\theta_H\\) and \\(\\theta_L\\) are foreground and background IoU thresholds. Binary Cross Entropy Loss is exploited here for confidence prediction. The losses of our detect head are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.681, + 0.483, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {R P N} + \\lambda_ {1} \\cdot \\mathcal {L} _ {r e g} + \\lambda_ {2} \\cdot \\mathcal {L} _ {s c o r e} \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.709, + 0.47, + 0.727 + ], + "angle": 0, + "content": "4. Event-based 3D Object Detection Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.484, + 0.84 + ], + "angle": 0, + "content": "This study is the first to explore the use of event cameras for 3D detection. To validate our model's effectiveness, we introduce two event 3D detection datasets: Ev-Waymo, which consists of synthetic events, and DSEC-3DOD with real events. We first obtained accurate 3D bounding box annotations for the active time, when both LiDAR and image data are available, and then processed the blind time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.418, + 0.865 + ], + "angle": 0, + "content": "4.1. Annotation for Active Time at 10 FPS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ev-Waymo Dataset is a synthetic event dataset based on the Waymo Open Dataset (WOD) [93]. WOD provides" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.091, + 0.895, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.247, + 0.905, + 0.275 + ], + "angle": 0, + "content": "Figure 5. DSEC-3DOD Dataset samples. Event data, label overlaid image, and accumulated LiDAR from left to right." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.906, + 0.432 + ], + "angle": 0, + "content": "sequential pairs of synchronized LiDAR, image data, and 3D box annotations at 10 FPS. As shown in Fig. 4, event streams are generated using a widely-adopted event simulator, Vid2E [21]. The Ev-Waymo evaluation follows the same metrics as WOD, calculating both average precision (AP) and average precision weighted by heading (APH). Adhering to the DSEC-3DOD setting, Ev-Waymo employs the front camera and a cropped LiDAR aligned to the corresponding field of view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.906, + 0.599 + ], + "angle": 0, + "content": "DSEC-3DOD. The DSEC dataset [23] is a stereo event dataset that includes LiDAR, image, and real event data captured in challenging environments. We sampled a portion of the \"zurich_city\" sequence, providing LiDAR and image data at 10 FPS. We hired annotation experts to label 3D bounding boxes at 10 FPS, ensuring alignment with the sensor data. As shown in Fig. 5, DSEC-3DOD contains real events, incorporating noise and capturing challenging environments, unlike simulated events in Ev-Waymo. During the annotation process, we accumulated LiDAR points to precisely generate ground truth 3D bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.66 + ], + "angle": 0, + "content": "For evaluation on DSEC-3DOD, we used the same evaluation metrics as Ev-Waymo but include only vehicle and pedestrian classes due to a lack of cyclist data, with all objects categorized under a single difficulty level." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.842, + 0.683 + ], + "angle": 0, + "content": "4.2. Annotation for Blind Time at 100 FPS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For training and evaluation on 3D detection during blind time, we created 3D box annotations for periods when Li-DAR and images were absent. We linearly interpolated the original 10 FPS annotations, dividing blind times into 10 evenly spaced intervals between consecutive frames, resulting in temporally dense annotations at 100 FPS. To ensure accurate bounding box ground truths in instances where Li-DAR and image data are unavailable, we employed event-based video frame interpolation [39] and point cloud interpolation [123] methods to generate sensor data for blind time. Data annotation experts refined interpolated bounding boxes based on the generated sensor data. The Ev-Waymo dataset consists of 80 sequences comprising a total of 157K labeled scenes, providing 100 FPS detection ground truth." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27201" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 1. Performance comparison on the Ev-Waymo dataset for 3D vehicle (IoU = 0.7), pedestrian (IoU = 0.5) and cyclist (IoU = 0.5) detection following previous protocols. Evaluated on 100 FPS ground truth annotation for \\( t = \\{0,0.1,\\dots,0.9\\} \\). Offline represents results obtained by interpolating data using sensor data from active timestamps 0 and 1. Online represents results obtained using data accessible at present. Specifically, online evaluation of previous works uses only the data from active timestamp 0, while we also utilize event data from active timestamp 0 up to the present time." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.162, + 0.9, + 0.34 + ], + "angle": 0, + "content": "
Interpolation MethodsMethods3D Detection ModalityALL (mAP/mAPH)VEH (AP/APH)PED (AP/APH)CYC (AP/APH)
L2L1L2L1L2L1L2
OfflineN-PCI [123]VoxelNeXt [9]L53.61/50.6762.96/62.5159.10/58.6768.43/60.7762.85/55.6944.55/43.1538.88/37.65
HEDNet [118]L50.52/47.5159.57/59.0955.90/55.4563.28/55.3258.03/50.6243.14/41.8037.63/36.46
N-PCI [123] + EMA [116]Focals Conv [8]L + I42.53/40.4155.49/55.0254.45/53.9948.07/43.0646.54/41.6927.70/26.6026.59/25.54
LoGoNet [47]L + I53.29/50.4963.48/63.0560.15/59.7367.85/60.2862.43/55.3742.40/41.3737.29/36.38
N-PCI [123] + CBM [39]Focals Conv [8]L + I42.51/40.3955.48/55.0154.44/53.9848.07/43.0646.54/41.6827.67/26.5726.56/25.50
LoGoNet [47]L + I53.57/50.7563.60/63.1660.27/59.8568.14/60.5362.77/55.6742.85/41.8137.66/36.74
OnlineN/AVoxelNeXt [9]L33.32/31.7044.40/44.1041.78/41.4940.52/36.2336.93/32.9624.40/23.7321.24/20.66
HEDNet [118]L31.57/29.9042.03/41.7139.32/39.0238.86/34.4335.67/31.5322.64/21.9919.72/19.14
Focals Conv [8]L + I26.27/25.0137.31/37.0136.60/36.3129.20/26.1628.41/25.4414.30/13.7813.79/13.29
LoGoNet [47]L + I33.27/31.7544.14/43.8741.73/41.4739.98/35.8436.48/32.6724.71/24.1521.59/21.10
Ev-3DOD (Ours)L + I + E48.06/45.6060.30/59.9559.19/58.8557.40/50.7855.30/48.9331.08/30.3829.69/29.03
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.351, + 0.898, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.626, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of our method with other offline and online evaluations on the Ev-Waymo dataset. \\( t = 0 \\) represents the active time, while \\( t = 0.3, 0.6, 0.9 \\) denotes the blind times. For easier understanding, images at active timestamps 0 and 1 are overlaid." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.743 + ], + "angle": 0, + "content": "The DSEC 3D dataset consists of 178 sequences and includes 54K labeled scenes. For evaluation in blind time, we utilize all ground truth boxes at 100 FPS. More details of the datasets are described in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.76, + 0.224, + 0.778 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.786, + 0.267, + 0.803 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We target causal 3D detections to detect objects during blind time. For comparison, we present two types of evaluation methods for conventional models. First is the online evaluation, where the network is inferred at any given blind time using only accessible data. Conventional methods can only process the most recent sensor data at active timestamp" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.907, + 0.834 + ], + "angle": 0, + "content": "0. In contrast, our approach additionally incorporates event data from active timestamp 0 up to the current blind time. Second, offline evaluation addresses the challenge of applying conventional fixed frame rate approaches directly to blind time by utilizing future data. In other words, data for the blind time is interpolated using active timestamps 0 and 1, where timestamp 1 represents future data. Consequently, these offline evaluations serve as a performance oracle for online methods. More implementation details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.718, + 0.864 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We employ recent state-of-the-art methods, including LiDAR-based [9, 118] and LiDAR-camera-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27202" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.173, + 0.089, + 0.823, + 0.103 + ], + "angle": 0, + "content": "Table 2. Results on the DSEC-3DOD dataset for 3D vehicle (IoU = 0.7), and pedestrian (IoU = 0.5) detection." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.105, + 0.9, + 0.262 + ], + "angle": 0, + "content": "
Interpolation MethodsMethods3D Detection ModalityALLVEHPED
mAPmAPHAPAPHAPAPH
OfflineN-PCI [123]VoxelNeXt [9]L36.2932.4345.3544.9327.2319.93
HEDNet [118]L37.7330.7341.8241.3533.6420.1
N-PCI [123] + EMA [116]Focals Conv [8]L + I30.2925.8535.5235.0425.0516.65
LoGoNet [47]L + I37.3031.3944.6444.0629.9618.72
N-PCI [123] + CBM [39]Focals Conv [8]L + I30.2825.8635.5635.0824.9916.64
LoGoNet [47]L + I37.3531.4544.7644.1729.9418.73
OnlineN/AVoxelNeXt [9]L13.9412.9021.4221.236.464.57
HEDNet [118]L13.7511.6818.4818.299.025.07
Focals Conv [8]L + I10.819.6115.7415.555.883.67
LoGoNet [47]L + I13.2911.9119.5019.327.074.49
Ev-3DOD (Ours)L + I + E31.1726.5441.6541.2020.6811.88
" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.281, + 0.478, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons of our method with other offline and online evaluations on the DSEC-3DOD dataset. \\( t = 0.3, 0.6, 0.9 \\) denotes the blind times." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.639 + ], + "angle": 0, + "content": "ods [8, 47]. For offline image interpolation, we use frame-based, EMA [116], and event-based method, CBM [39]. NeuralPCI [123] is adopted for point cloud interpolation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Ev-Waymo. Table 1 shows the quantitative result of Ev-Waymo. Comparing methods in the online setting, previous methods [8, 9, 47, 118], which rely on fixed frame rate sensors (i.e., LiDAR and Image), cannot detect object motion during blind times. This limitation leads to reduced performance when evaluated with 100 FPS ground truth. In contrast, our approach leverages high-temporal resolution of event data to estimate object motion during blind time, significantly outperforming existing online methods. Notably, Ev-3DOD is comparable to offline oracles and even surpasses certain approaches, such as the combination of NeuralPCI [123] + EMA [116] and Focals Conv [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Figure 6 shows comparisons between the results of online, offline, and our method during blind time. Existing online methods fail to align prediction with the ground truth during blind time. In contrast, Ev-3DOD, despite being an online approach, aligns well with the ground truth and" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.281, + 0.907, + 0.324 + ], + "angle": 0, + "content": "Table 3. Ablation study of the proposed components. V3D-EF and MCE denote Virtual 3D Event Fusion (Sec. 3.3) and the motion confidence estimator (Sec. 3.4), respectively." + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.328, + 0.849, + 0.401 + ], + "angle": 0, + "content": "
V3D-EFMCEL2
mAPmAPH
33.3031.68
✓46.5544.22
✓✓48.0645.60
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.504 + ], + "angle": 0, + "content": "demonstrates performance comparable to offline methods. The right graph in Fig. 6 further demonstrates our method's superiority, showing that despite performance degradation with increasing blind time, the gap remains comparable to offline methods, while the online approach exhibits a much larger pixel difference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.656 + ], + "angle": 0, + "content": "DSEC-3DOD. As shown in Table 2, we evaluate our model on the DSEC-3DOD real event dataset, which includes fast motion and challenging lighting conditions that significantly degrade performance during blind time. Our model effectively estimates the 3D motion of objects from events, achieving a substantial improvement over other online methods. Notably, it demonstrates performance comparable to offline oracle methods that utilize future information. The analysis of computational complexity is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Figure 7 shows blind time detection inference for fast-moving objects in the DSEC dataset. Under rapid motion, the quality of the interpolated pseudo-sensor data may deteriorate, occasionally resulting in detection failures, even when future information is utilized. In contrast, the proposed Ev-3DOD accurately detects objects in blind time, predicting motion from events." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.771, + 0.789 + ], + "angle": 0, + "content": "5.3. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.855 + ], + "angle": 0, + "content": "We present an ablation study on key components and various analyses on the Ev-Waymo dataset. In V3D-EF analyses, the motion confidence estimator (MCE) module is excluded for clear comparisons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Each Component. In Table 3, we conduct the ablation study on the key components of Ev-3DOD. We evaluated a baseline and progressively added modules to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27203" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 4. Effectiveness of the non-empty mask in the V3D-EF module, as defined in Eq. (1)." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.121, + 0.427, + 0.182 + ], + "angle": 0, + "content": "
MethodL2
mAPmAPH
w/o Non-empty Mask42.5740.50
w/ Non-empty Mask46.5544.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.195, + 0.483, + 0.223 + ], + "angle": 0, + "content": "Table 5. Comparison of results across different modalities and fusion variants of the V3D-EF module." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.224, + 0.472, + 0.326 + ], + "angle": 0, + "content": "
MethodL2
mAPmAPH
Base Model33.3031.68
+ LiDAR37.0435.25
+ Event42.2340.17
+ LiDAR + Event (fusion before pooling)43.9541.87
+ LiDAR + Event (fusion after pooling)46.5544.22
" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.343, + 0.475, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Figure 8. Detection performance as time elapses since the active timestamp. We compare ours with offline and online approaches." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.483, + 0.641 + ], + "angle": 0, + "content": "observe improvements. The motion confidence estimator is dependent on V3D-EF, so it was not evaluated separately. V3D-EF, which explicitly estimates motion, has a direct impact on performance, resulting in significant improvements of \\(+13.25\\%\\) in mAP and \\(+12.54\\%\\) in mAPH. Additionally, the motion confidence estimator refines prediction scores at each blind time, demonstrating its effectiveness with further gains of \\(+1.51\\%\\) and \\(+1.38\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Non-empty Mask in V3D-EF. In the proposed V3D-EF, we employ a non-empty mask to achieve accurate and efficient alignment between sparse events and voxel features. To demonstrate the effectiveness of this non-empty mask, we conduct a comparative experiment in which events are projected onto all voxels within the V3D-EF module. The results in Table 4 confirm that the non-empty mask enables effective alignment between the two modalities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Variations of the V3D-EF Module. Our core component, the V3D-EF module, provides multiple design options. To analyze these configurations, we developed and evaluated several variants, with the results summarized in Table 5. We experimented with generating the implicit motion field using each modality, i.e., LiDAR and event sensors, and found a performance drop compared to the multi-modal approach. This decline is likely due to the limitations of each modality: LiDAR alone cannot capture motion information during" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 6. Evaluation results of the model trained with 100 FPS annotations using sliced events at 200 and 300 FPS." + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.12, + 0.835, + 0.193 + ], + "angle": 0, + "content": "
Test SetL2
mAPmAPH
100 FPS48.0645.60
200 FPS47.8745.44
300 FPS47.8145.39
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.907, + 0.296 + ], + "angle": 0, + "content": "blind times, and event features cannot accurately estimate refined 3D motion. Additionally, multi-modal fusion before pooling decreased performance. We believe this is because fusing before pooling leads to integration between voxels and events outside of box proposal regions, resulting in the use of redundant and possibly misaligned information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.909, + 0.465 + ], + "angle": 0, + "content": "Detection Performance per Elapsed Blind Time. Figure 8 shows the performance over time elapsed since the active timestamp. Due to object motion, the performance of online detection deteriorates rapidly as the time distance from the active timestamp increases. In contrast, offline methods sustain robust performance throughout the blind time by leveraging future data, with only a slight performance drop at the center of the blind time as it moves further from the active timestamps located at both ends. The proposed model, relying solely on event data, achieves a temporal degradation pattern comparable to the offline method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.909, + 0.571 + ], + "angle": 0, + "content": "Generalization to Higher Frame-rate. To demonstrate the generalization ability to different frame rates, we generate 200 and 300 FPS annotations for arbitrary timestamps by interpolating a pre-constructed 100 FPS dataset. Without additional training, we apply the model trained on 100 FPS. As shown in Table 6, our model effectively infers at arbitrary timestamps with minimal degradation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.642, + 0.604 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.78 + ], + "angle": 0, + "content": "This paper proposes Ev-3DOD, a novel method to utilize an event camera in 3D detection for detecting objects in blind time. To transfer information from active times to the current blind time, we estimate 3D motion based on event data. To effectively fuse sparse LiDAR data with events, we propose a Virtual 3D Event Fusion (V3D-EF) and introduce a motion confidence estimator to define confidence for 3D detection during blind times. For this study, we introduce the first event-based 3D detection dataset, emphasizing its 100 FPS annotations. We hope this work showcases the potential of neuromorphic cameras, inspiring future research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitation and Future Work. Our work is the first to address 3D detection during the blind time, establishing it as a pioneering study in this area. However, one limitation is the lack of recurrent processing of past event information to enhance temporal interaction. As a result, discontinuities may occur when fixed frame rate sensors update at active timestamps. We consider improving the interaction between consecutive event data a crucial direction for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27204" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.206 + ], + "angle": 0, + "content": "We thank 42dot for funding this research. We also appreciate the support and valuable discussions from the members of the 42dot research team. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.244, + 0.483, + 0.326 + ], + "angle": 0, + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1090-1099, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.328, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[2] Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Bourtsoulatze, and Yiannis Andreopoulos. Graph-based object classification for neuromorphic vision sensing. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 491-501, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.398, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.468, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[4] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.538, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[5] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 172-181, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.608, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.65, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[7] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.706, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[8] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5428-5437, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[9] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 887-897, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.244, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[11] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Deformable feature aggregation for dynamic multi-modal 3d object detection. In European conference on computer vision, pages 628-644. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[12] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[13] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[14] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In European Conference on Computer Vision, pages 294–314. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[15] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.386, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[16] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8458-8468, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.469, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[17] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1761-1769, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.538, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[18] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44 (1):154-180, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.622, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[19] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. ArXiv, abs/2211.12324, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.664, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[20] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629:1034 - 1040, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.705, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[21] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.762, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13884-13893, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[23] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27205" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[26] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3153-3163, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[27] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22867-22876, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[28] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11873-11882, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[29] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8417-8427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[30] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[31] Cheng-Ju Ho, Chen-Hsuan Tai, Yen-Yu Lin, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Diffusion-ss3d: Diffusion model for semi-supervised 3d object detection. Advances in Neural Information Processing Systems, 36:49100-49112, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[32] Jordan SK Hu, Tianshu Kuai, and Steven L Waslander. Point density-aware voxels for lidar 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8469-8478, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[33] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[34] Junjie Huang, Yun Ye, Zhujin Liang, Yi Shan, and Dalong Du. Detecting as labeling: Rethinking lidar-camera fusion in 3d object detection. In European Conference on Computer Vision, pages 439-455. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[35] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Yu Huang, Chenzhuang Du, Zihui Xue, Xuanyao Chen, Hang Zhao, and Longbo Huang. What makes multi-modal" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "learning better than single (provably). Advances in Neural Information Processing Systems, 34:10944-10956, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.123, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[37] Massimiliano Iacono, Stefan Weber, Arren J. Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.194, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[38] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformer. In Proceedings of the AAAI conference on Artificial Intelligence, pages 1042-1050, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.265, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[39] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.335, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[40] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.392, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[41] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.462, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[42] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.908, + 0.602 + ], + "angle": 0, + "content": "[43] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[44] Dianze Li, Jianing Li, and Yonghong Tian. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:14020-14037, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.661, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[45] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[46] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.776, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[47] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[48] Xiaotian Li, Baojie Fan, Jiandong Tian, and Huijie Fan. Gafusion: Adaptive fusing lidar and camera with multiple guidance for 3d object detection. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27206" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21209-21218, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[49] Xiang Li, Junbo Yin, Wei Li, Chengzhong Xu, Ruigang Yang, and Jianbing Shen. Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3208-3215, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[50] Yijin Li, Han Zhou, Bangbang Yang, Yexin Zhang, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Graph-based asynchronous event processing for rapid object recognition. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 914-923, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[51] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems, 35:18442-18455, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.335, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[52] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[53] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17182-17191, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[54] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[55] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7546-7555, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[56] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[57] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7345-7353, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[58] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[59] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[60] Xianpeng Liu, Ce Zheng, Ming Qian, Nan Xue, Chen Chen, Zhebin Zhang, Chen Li, and Tianfu Wu. Multi-view attentive contextualization for multi-view 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16688-16698, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.178, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[61] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[62] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 996-997, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.306, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[63] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.39, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[64] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.462, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[65] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2723-2732, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[66] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Ji-ashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3164–3173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[67] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. ArXiv, abs/2003.09148, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[68] Zhenwei Miao, Jikai Chen, Hongyu Pan, Ruiwen Zhang, Kaixuan Liu, Peihan Hao, Jun Zhu, Yang Wang, and Xin Zhan. Pvgnet: A bottom-up one-stage 3d object detector with integrated multi-level features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3279-3288, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.732, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[69] Su Pang, Daniel Morris, and Hayden Radha. Clocs: Camera-lidar object candidates fusion for 3d object detection. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10386-10393. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[70] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Wanli Peng, Hao Pan, He Liu, and Yi Sun. Ida-3d: Instance-depth-aware 3d object detection from stereo vision" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27207" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13015-13024, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[72] Yan Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In AAAI Conference on Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[73] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6015-6025, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[74] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16794-16804, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[75] Etienne Perot, Pierre de Tournemire, Davide Oscar Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. *ArXiv*, abs/2009.13436, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[76] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.449, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[77] AJ Piergiovanni, Vincent Casser, Michael S Ryoo, and Anelia Angelova. 4d-net for learned multi-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15435-15445, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[78] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[79] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[80] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[81] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[82] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[83] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[84] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[85] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2743-2752, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.219, + 0.905, + 0.285 + ], + "angle": 0, + "content": "[86] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.288, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[87] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[88] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1711-1719, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.415, + 0.904, + 0.467 + ], + "angle": 0, + "content": "[89] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[90] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.527, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[91] Jiaming Sun, Linghao Chen, Yiming Xie, Siyu Zhang, Qin-hong Jiang, Xiaowei Zhou, and Hujun Bao. Disp r-cnn: Stereo 3d object detection via shape prior guided instance disparity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10548-10557, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.609, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[92] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.68, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[93] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2443-2451, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[94] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27208" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[95] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[96] Dongsheng Wang, Xu Jia, Yang Zhang, Xinyu Zhang, Yaoyuan Wang, Ziyang Zhang, D. Wang, and Huchuan Lu. Dual memory aggregation network for event-based object detection with learnable representation. In AAAI Conference on Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[97] Wenjie Wang, Yehao Lu, Guangcong Zheng, Shuigen Zhan, Xiaqing Ye, Zichang Tan, Jingdong Wang, Gaoang Wang, and Xi Li. Bevspread: Spread voxel pooling for bird's-eye-view representation in vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14718-14727, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[98] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[99] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.483, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[100] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.552, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[101] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multimodal sparse representations for multi-sensor 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 17545-17556, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.637, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[102] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.693, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[103] Jianyun Xu, Zhenwei Miao, Da Zhang, Hongyu Pan, Kaixuan Liu, Peihan Hao, Jun Zhu, Zhengyang Sun, Hongmin Li, and Xin Zhan. Int: Towards infinite-frames 3d detection with an efficient framework. In European Conference on Computer Vision, pages 193–209. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[104] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Zhou Bin, and Liangjun Zhang. Fusionpainting: Multimodal fusion with adaptive attention for 3d object detection. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3047-3054. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[105] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[106] Yixin Yang, Jinxiu Liang, Bohan Yu, Yan Chen, Jimmy S Ren, and Boxin Shi. Latency correction for event-guided" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "deblurring and frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24977-24986, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[107] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[108] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[109] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. Advances in Neural Information Processing Systems, 35:1992-2005, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.303, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[110] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhao-Xu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10201-10210, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[111] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.443, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[112] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In European conference on computer vision, pages 17-33. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[113] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[114] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[115] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kili-ian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. arXiv preprint arXiv:1906.06310, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[116] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[117] Gang Zhang, Junnan Chen, Guohuan Gao, Jianmin Li, Si Liu, and Xiaolin Hu. Safdnet: A simple and effective network for fully sparse 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14477-14486, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[118] Gang Zhang, Chen Junnan, Guohuan Gao, Jianmin Li, and Xiaolin Hu. Hednet: A hierarchical encoder-decoder net" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27209" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "work for 3d object detection in point clouds. Advances in Neural Information Processing Systems, 36, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[119] Jiqing Zhang, B. Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8791-8800, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[120] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9781-9790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[121] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[122] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[123] Zehan Zheng, Danni Wu, Ruisi Lu, Fan Lu, Guang Chen, and Changjun Jiang. Neuralpci: Spatio-temporal neural field for 3d point cloud multi-frame non-linear interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 909–918, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[124] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1839-1849, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[125] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[126] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.482, + 0.768 + ], + "angle": 0, + "content": "[127] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12800-12810, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27210" + } + ] +] \ No newline at end of file diff --git a/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_origin.pdf b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4cdfc14a83bdf3c51f04d1636fbcf862d7ea900d --- /dev/null +++ b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/e0858502-8bdd-48cc-ae6e-4de96b374d49_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9255f28b28f55746e9a97f0968a43329274ceac116472561b985deef3f01c0a6 +size 5957995 diff --git a/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/full.md b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c219cb9fcf18a743e8c19dc30a60f100df3ab597 --- /dev/null +++ b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/full.md @@ -0,0 +1,367 @@ +# Ev-3DOD: Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras + +Hoonhee Cho*, Jae-Young Kang*, Youngho Kim, and Kuk-Jin Yoon KAIST + +{gnsgnsgml,mickeykang,kmax2001,kjyoon}@kaist.ac.kr + +![](images/e63f2f6992edaa7daed8436d54be8bd4c5979af92b981593186c691844ac6ae3.jpg) +Figure 1. Comparison between the proposed Ev-3DOD and conventional 3D detection methods. Fixed frame rate sensors (e.g., LiDAR, frame camera) inevitably have periods of "blind time" $(t_{0\rightarrow 1})$ when they lack data between active timestamps (0 and 1). Consequently, as shown in (a), conventional methods cannot detect objects during blind time, while other objects like vehicles can undergo significant motion within these gaps. In contrast, as illustrated in (b), our method, Ev-3DOD, leverages the high temporal resolution of the event camera, enabling accurate object detection even during the blind times of other sensors. + +![](images/f2be3700a8221c22bd980928c3d6da76f63ea82f142b6881b969ce148df3640f.jpg) + +![](images/115f667de511432d6fcb088013b37a0e501177ccbb0f550566a3d8d130f7074d.jpg) + +# Abstract + +Detecting 3D objects in point clouds plays a crucial role in autonomous driving systems. Recently, advanced multimodal methods incorporating camera information have achieved notable performance. For a safe and effective autonomous driving system, algorithms that excel not only in accuracy but also in speed and low latency are essential. However, existing algorithms fail to meet these requirements due to the latency and bandwidth limitations of fixed frame rate sensors, e.g., LiDAR and camera. To address this limitation, we introduce asynchronous event cameras into 3D object detection for the first time. We leverage their high temporal resolution and low bandwidth to enable high-speed 3D object detection. Our method enables detection even during inter-frame intervals when synchronized data is unavailable, by retrieving previous 3D information through the event camera. Furthermore, we introduce the first event-based 3D object detection datasets, EvWaymo and DSEC-3DOD, both of which include ground + +truth 3D bounding boxes at 100 FPS, establishing the first benchmarks for event-based 3D detectors. The code and dataset are available at https://github.com/mickeykang16/Ev3DOD. + +# 1. Introduction + +3D object detection is a crucial task in autonomous driving. Multi-modal sensors like LiDAR and cameras are commonly used to understand scenes and handle complex road scenarios. This sensor setup consists of synchronized sensors with a fixed frame rate and has limited speed due to bandwidth limitations. As shown in Fig. 1 (a), the blind time, which is the period when sensor information is unavailable (e.g., 100 milliseconds), is critical as cars can experience large motions as 3 meters during this time. Therefore, temporal resolution is a crucial factor that can directly influence accident risk. + +Recently, event cameras [18] have introduced a new paradigm in perceptual research. Event cameras provide brightness changes as a stream of data with sub-milisecond latency, making them robust to motion blur [41, 106] and enabling low bandwidth [22, 94]. Event cameras overcome the trade-off between bandwidth and blind time but have the drawback of sacrificing intensity information. To address this, recent work [12, 13, 92, 120] has combined event cameras with frame cameras to complement the sparsity of event cameras while overcoming the fixed frame rate sensor's limits. + +In this paper, we aim for accurate 3D object detection in blind time through a multi-modal approach that combines LiDAR and frame-based sensors with an event camera, leveraging its high temporal resolution. As shown in Fig. 1 (b), event data is available during blind times, while the LiDAR and image data from the most recent active timestamp are accessible. Therefore, there is a misalignment in timestamps between the fixed frame sensor data and the current blind time. Then, this raises the question: How can we estimate 3D motion using only event data, especially when data from other sensors is unavailable at the current time? + +To address this, we propose a virtual 3D event fusion (V3D-EF) that transfers past 3D information to align with the present time using the event camera. As extensively covered in previous research [14, 24, 40, 89], event data inherently contains motion information on an image plane. Based on this observation, we hypothesized that providing a 3D cue associated with the initial timestamp to the event information would enable accurate estimation of the 3D motion vector. To this end, we project event features into a 3D voxel space to create virtual 3D event features. In this process, events are projected only onto regions where point cloud data exists, aligning with the point cloud data. Then, we aggregate the event features with the voxel features to estimate an implicit motion field that captures the motion inherent within the grid. The generated implicit motion field encapsulates both the 3D spatial information from voxels and the motion information from events, enabling the calculation of a 3D motion vector from the most recent active timestamp to the present blind time. + +Another crucial factor in 3D object detection is the score for each bounding box. The model considers both box proposal scores at the active timestamp and the motion score of each predicted box in blind time. In blind time, the model aligns 3D information of voxel features and motion information of events to learn the confidence of box motion. + +A major hurdle in 3D detection research using event cameras is the lack of suitable datasets. To address this, we present two datasets: Ev-Waymo, which includes synthetic event data, and DSEC-3DOD, which contains real event data. Unlike existing public 3D object detection datasets, + +ours include annotations not only for active timestamps with 3D LiDAR sensor data but also for blind times, providing annotation in 100 FPS frame rate. + +# 2. Related Works + +Single-modality 3D Object Detection. Given that cameras offer significant cost advantages over LiDAR sensors, many researchers have focused on developing methods that leverage camera-based systems for 3D object detection using image-only inputs [7, 10, 26, 30, 33, 38, 46, 54, 60-62, 64, 71, 91, 97, 99, 115]. Image-based 3D object detection faces challenges due to the lack of direct depth information. To overcome this challenge, several studies [7, 70, 76, 83, 98, 115] have employed depth estimation techniques to generate pseudo 3D point clouds or elevate 2D features into 3D space for object detection, alongside proposals for transformer-based architectures [35, 99, 121] that utilize 3D object queries and 3D-2D correspondences. Nonetheless, accurately estimating depth from images remains difficult, resulting in image-based methods performing worse than LiDAR-based approaches. + +LiDAR-based 3D object detection can reliably estimate 3D bounding boxes using point clouds captured from LiDAR sensors. Current LiDAR-based detection methods can be divided into three categories based on different point cloud encoding formats: point-based methods [17, 32, 79, 80, 82, 86, 88, 108, 112, 124], voxel-based methods [4, 15, 16, 25, 29, 29, 31, 43, 49, 55, 65, 66, 85, 100, 103, 111, 117, 125], and point-voxel fusion networks [6, 28, 68, 87, 107]. While LiDAR detection is advantageous in various conditions, it often struggles in areas with sparse LiDAR data. Therefore, integrating the geometric benefits of point clouds with the semantic richness of images remains crucial for improving performance. + +Multi-modality 3D Object Detection. Multi-modal sensor fusion [5, 34, 36, 47, 48, 56, 57, 59, 90, 101, 102, 113], where different sensors complement each other, typically utilizes a combination of cameras and LiDAR. Multimodal 3D object detection has gained attention for improving accuracy over unimodal methods. Fusion approaches are categorized into early [3, 11, 42, 81, 95, 104], middle [51-53, 58, 63, 77, 78, 109, 122], and late fusions [1, 47, 69], with middle fusion now preferred due to its robustness to calibration errors and enhanced feature interaction. Crossmodal fusion has achieved significant performance improvements by enhancing the sparse features of point clouds with semantic information from images. However, there are still unexplored areas concerning safety in autonomous driving. LiDAR and cameras have limited time resolution (i.e., $10 - 20\mathrm{Hz}$ ) due to the relatively high bandwidths. To enhance safety, algorithms that operate at high speeds are essential. To address this, we introduce an asynchronous event camera with high temporal resolution for 3D object + +![](images/2cba8d8d590ff446bb95c44596e9b699280aa311b7afbaffef249b015571c0b9.jpg) +Figure 2. The overall pipeline for event-based 3D object detection (Ev-3DOD). (a): At active timestamp 0, LiDAR and image data are available. Therefore, we utilize an RGB-LiDAR Region Proposal Network (RPN) to extract voxel features and 3D bounding boxes at the active timestamp. (b): To predict the bounding box during blind time, $0 \leq t < 1$ , we estimate the 3D motion and confidence score using event features. For computational efficiency, we design the process to compute (a) only once before the next active timestamp while performing iterative computations solely for (b). + +detection for the first time. The proposed method performs multi-modal fusion between synchronized sensors and asynchronous events to increase the operational frequency of 3D detection. + +Event-based Object Detection. Event cameras break through the limitations of bandwidth requirements and perceptual latency with the trade-off of sparse data and loss of intensity information. To make the most of these characteristics, prior work has designed models to adapt to event data using graph neural networks [2, 19, 20, 50, 84], spiking neural networks [110, 119], and sparse neural networks [67, 74]. Recently, effective dense neural networks [27, 37, 45, 72, 75, 96, 127] and transformer architectures [22, 44, 73, 74] have been developed to leverage event cameras, aiming to achieve both low latency and high performance. These advances enable real-time operation with impressive accuracy, demonstrating the practical applicability of event cameras in real-world scenarios. To the best of our knowledge, we are the first to extend event-based object detection from 2D to 3D space. By employing a multimodal fusion approach of the event camera and 3D sensors (i.e., LiDAR), we aim to overcome the latency and persistent interval time barriers of 3D object detections. + +# 3. Method + +# 3.1. Problem Definition. + +LiDAR and camera data are only available at active timestamps 0 and 1, where 1 represents invalid future information in an online process. Our goal is to predict 3D bounding boxes during $0 \leq t < 1$ , including the blind time + +when synchronized sensors are unavailable. Given a pair of synchronized 3D points $P_0 = \{(x_l, y_l, z_l)\}_{l=1}^N$ and image $I_0 \in \mathbb{R}^{3 \times H \times W}$ at active timestamp 0, we leverage event stream $\mathcal{E}_{0 \to t} = \{(i, j, p, \tau) | 0 \leq \tau < t\}$ to detect 3D bounding box set $B_t$ at an arbitrary time $0 \leq t < 1$ . + +# 3.2. Framework Overview + +As illustrated in (a) of Fig. 2, at active timestamp 0, both LiDAR and image data are available. RGB-LiDAR Region Proposal Network (RPN) generates bounding box proposals from rich 3D information in active time. We begin by voxelizing the point cloud and generating voxel features, denoted as $\mathbf{V}_0\in \mathbb{R}^{D_X\times D_Y\times D_Z\times C}$ , where $D_{X}$ , $D_{Y}$ , and $D_Z$ represent the grid size, and $C$ denotes the channel dimension, using a 3D backbone [125]. Then, using existing region proposals [114] and refinement networks [47] with image features, we generate bounding box proposals, $B_0 = \{B_0^1,B_0^2,\dots ,B_0^n\}$ and box proposal confidence $p_0 = \{p_0^1,p_0^2,\dots ,p_0^n\} \in [0,1]$ for the active timestamp 0. RGB-LiDAR-based multi-modal networks have been widely studied, leading to many advanced models capable of accurate estimation at active timestamps. We tuned the model with fewer parameters and a reduced channel dimension to improve efficiency. + +Our emphasis is on achieving precise estimation throughout the blind time, spanning the interval from active timestamp 0 to the subsequent timestamp 1. To infer bounding boxes at an arbitrary inter-frame time $t$ , we first convert the event stream, $\mathcal{E}_{0\rightarrow t}$ into voxel grids [126] as $\mathbb{E}_t$ . Then, we use a feature encoder to produce event features, $\mathbf{E}_t\in \mathbb{R}^{H / 4\times W / 4\times C}$ , where $H$ and $W$ is the spa + +tial dimensions of the events. Our objective is to estimate the 3D motion of objects during blind time. To achieve this, we utilize the bounding box set $B_{0}$ and voxel features $\mathbf{V}_0$ from the active timestamp 0. During blind times, we share the voxel features, $\mathbf{V}_0$ , generated at the active timestamp, ensuring that voxel features are computed only once to maintain a cost-effective structure. Through the Virtual 3D Event Fusion (V3D-EF), we generate an implicit motion field for each bounding box by integrating event and voxel features, encapsulating motion information. Using this field, we compute the 3D motion and confidence score, enabling the estimation of 3D boxes at the blind time $t$ . + +# 3.3. Virtual 3D Event Fusion for Motion Prediction + +The event feature continuously captures each pixel's motion information over time on the 2D image plane. However, estimating 3D motion using only events is a challenge. We hypothesize that providing each event feature with adequate 3D information would allow for effective 3D motion estimation. To achieve this, we aim to incorporate a 3D cue into the event features. Since both event features and voxel features are sparse, it is crucial to align each voxels with the corresponding events accurately. Directly using the voxel features would cause misalignment with the actual 3D geometry. Thus, we adopt the voxel point centroid method employed in previous works [32]. + +Specifically, we first consider the non-empty voxel features as $\mathbf{V}_0 = \{V_0^k = \{h_0^k,f_0^k\} \mid k = 1,\dots ,N_\nu \}$ , where $h_0^k$ is the 3D voxel index, $f_0^k$ is the associated voxel feature vector, and $N_{\mathcal{V}}$ is the number of non-empty voxels. Then, points within each voxel are grouped into a set $\mathcal{N}(V_0^k)$ by determining their voxel index $h_0^k$ based on their spatial coordinates $X_{i} = (x_{i},y_{i},z_{i})$ of points. The centroid of each voxel feature is subsequently computed as: + +$$ +c _ {0} ^ {k} = \frac {1}{\left| \mathcal {N} \left(V _ {0} ^ {k}\right) \right|} \sum_ {X _ {i} \in \mathcal {N} \left(V _ {0} ^ {k}\right)} X _ {i}. \tag {1} +$$ + +The centroid $c_0^k$ is calculated for non-empty voxels, ensuring it corresponds only to indices with existing points, with the goal of aligning these centroids to the event data. As shown in the top of Fig. 3, we propose a method to project events into a virtual voxel space based on the coordinates of the centroids. The projection onto the image plane from 3D space can be obtained as: $\mathbf{p}_0^k = K \cdot E \cdot c_0^k$ , where $K$ and $E$ denote the intrinsic and extrinsic parameters, $\cdot$ operation is matrix multiplication. Therefore, we calculate the event coordinates corresponding to each centroid, $c_0^k$ , and gather the event features, $\mathbf{E}_t$ , at these coordinates, $\mathbf{p}_0^k$ , to create virtual 3D event features, $\mathbf{V}_t^E = \{\mathbf{E}_t(\mathbf{p}_0^k)\}_{k=1}^{N_\nu} \in \mathbb{R}^{N_\nu \times C}$ . + +Since we aim to estimate the 3D motion of bounding boxes from the active timestamp $0$ up to the current time $t$ , we consider fusion only for the features corresponding + +![](images/4fea34eff8b6df80bcb906bf14b904c30978e89f1b4e2aa1c020bc200ee82d21.jpg) +Figure 3. Virtual 3D Event Fusion (V3D-EF). We generate an implicit motion field for each bounding box by applying the ROI pooling separately to the virtual 3D event features and the voxel features. Virtual 3D event features represent the 3D projection of 2D events. None-empty voxels of the voxel feature, $\mathbf{V}_0$ , are colored. + +to the box proposal. Given the bounding box proposals, $B_0$ , at active timestamp 0, we divide box proposals into $S \times S \times S$ regular sub-voxels. Then, we independently perform the ROI pooling [15] on the virtual 3D event features $\mathbf{V}_t^E$ , and the voxel features $\mathbf{V}_0$ , corresponding to each subvoxel. This generates grid features for the bounding boxes, denoted as $\tilde{\mathbf{V}}_t^E \in \mathbb{R}^{n \times S^3 \times C}$ and $\tilde{\mathbf{V}}_0 \in \mathbb{R}^{n \times S^3 \times C}$ , where $n$ is the number of box proposals. The generated features are then concatenated and passed through an MLP to produce the implicit motion field, $\mathcal{M}_t = \{\mathcal{M}_t^0, \mathcal{M}_t^1, \dots, \mathcal{M}_t^n\}$ . The implicit motion field is created by aligning 3D events with their corresponding voxel features, encapsulating motion information for the bounding box. Using an MLP, we estimate the explicit 3D motion. The explicit box motion $\mathbf{M}_{0 \rightarrow t}^i = (dx^i, dy^i, dz^i, d\alpha^i)$ , defined in box local coordinate of $B_0^i$ , contains spatial shift and yaw rotation. The final box location is calculated as $B_t^i = g(B_0^i, \mathbf{M}_{0 \rightarrow t}^i)$ , where $g$ is a box shifting operation. + +# 3.4. Motion Confidence Estimator + +As discussed in previous studies [114, 125], self-awareness of the confidence of predicted bounding boxes is a key factor. The simplest choice of final confidence score in our setting would be the score of active timestamp 0. However, the final score should take into account the object's motion prediction during the blind time. If the object experiences challenging motion during the blind time, a reliable bounding box at the active timestamp 0 could become less reliable. Therefore, the final score should consider two different sources. The first aspect is the score estimated at the active timestamp, based on LiDAR and images. The second + +![](images/e8306ee1b72f89ba50f37e89d5338ba97ff78a409e54ff2fa9f2db5ad88e9618.jpg) +Figure 4. Ev-Waymo Dataset samples. Synthetic event data, label overlaid image, and LiDAR from left to right. + +aspect is the confidence score based on motion during the blind time using events. + +The implicit motion field from Sec. 3.3 contains 3D motion information through alignment between the point cloud and events. We additionally employ an MLP specifically for calculating motion confidence. Box-wise motion confidence is estimated and normalized, resulting in $p_{0\rightarrow t} = \{p_{0\rightarrow t}^{1},p_{0\rightarrow t}^{2},\dots ,p_{0\rightarrow t}^{n}\} \in [0,1]$ . The final confidence score for box proposal $i$ is calculated by considering both the score at the active timestamp and the motion confidence: + +$$ +p _ {t} ^ {i} = p _ {0} ^ {i} \cdot p _ {0 \rightarrow t} ^ {i} \tag {2} +$$ + +# 3.5. Training Objectives + +For RGB-LiDAR RPN in the active timestamp, we follow RPN loss from [47, 105, 114]. For blind time prediction, box regression loss from [114] is applied between the predicted 3D box $B_{t}^{i}$ and the ground truth $\hat{B}_{t}^{i}$ . The target assigned to the motion confidence branch is an IoU-related value, as: + +$$ +\hat {p} _ {0 \to t} ^ {i} (\mathrm {I o U} _ {i}) = \left\{ \begin{array}{l l} 0 & \text {i f I o U _ {i} < \theta_ {L}}, \\ \frac {\mathrm {I o U} _ {i} - \theta_ {L}}{\theta_ {H} - \theta_ {L}} & \text {i f \theta_ {L} \leq I o U _ {i} < \theta_ {H}}, \\ 1 & \text {i f I o U _ {i} \geq \theta_ {H}}, \end{array} \right. +$$ + +$\mathrm{IoU}_i$ is the IoU between the $i$ -th prediction and the corresponding ground truth box, while $\theta_H$ and $\theta_L$ are foreground and background IoU thresholds. Binary Cross Entropy Loss is exploited here for confidence prediction. The losses of our detect head are computed as: + +$$ +\mathcal {L} = \mathcal {L} _ {R P N} + \lambda_ {1} \cdot \mathcal {L} _ {r e g} + \lambda_ {2} \cdot \mathcal {L} _ {s c o r e} \tag {3} +$$ + +# 4. Event-based 3D Object Detection Datasets + +This study is the first to explore the use of event cameras for 3D detection. To validate our model's effectiveness, we introduce two event 3D detection datasets: Ev-Waymo, which consists of synthetic events, and DSEC-3DOD with real events. We first obtained accurate 3D bounding box annotations for the active time, when both LiDAR and image data are available, and then processed the blind time. + +# 4.1. Annotation for Active Time at 10 FPS + +Ev-Waymo Dataset is a synthetic event dataset based on the Waymo Open Dataset (WOD) [93]. WOD provides + +![](images/ad2e992570a22964759e6efed473190c28a1a597f9c7e478e059765260d59afc.jpg) +Figure 5. DSEC-3DOD Dataset samples. Event data, label overlaid image, and accumulated LiDAR from left to right. + +sequential pairs of synchronized LiDAR, image data, and 3D box annotations at 10 FPS. As shown in Fig. 4, event streams are generated using a widely-adopted event simulator, Vid2E [21]. The Ev-Waymo evaluation follows the same metrics as WOD, calculating both average precision (AP) and average precision weighted by heading (APH). Adhering to the DSEC-3DOD setting, Ev-Waymo employs the front camera and a cropped LiDAR aligned to the corresponding field of view. + +DSEC-3DOD. The DSEC dataset [23] is a stereo event dataset that includes LiDAR, image, and real event data captured in challenging environments. We sampled a portion of the "zurich_city" sequence, providing LiDAR and image data at 10 FPS. We hired annotation experts to label 3D bounding boxes at 10 FPS, ensuring alignment with the sensor data. As shown in Fig. 5, DSEC-3DOD contains real events, incorporating noise and capturing challenging environments, unlike simulated events in Ev-Waymo. During the annotation process, we accumulated LiDAR points to precisely generate ground truth 3D bounding boxes. + +For evaluation on DSEC-3DOD, we used the same evaluation metrics as Ev-Waymo but include only vehicle and pedestrian classes due to a lack of cyclist data, with all objects categorized under a single difficulty level. + +# 4.2. Annotation for Blind Time at 100 FPS + +For training and evaluation on 3D detection during blind time, we created 3D box annotations for periods when Li-DAR and images were absent. We linearly interpolated the original 10 FPS annotations, dividing blind times into 10 evenly spaced intervals between consecutive frames, resulting in temporally dense annotations at 100 FPS. To ensure accurate bounding box ground truths in instances where Li-DAR and image data are unavailable, we employed event-based video frame interpolation [39] and point cloud interpolation [123] methods to generate sensor data for blind time. Data annotation experts refined interpolated bounding boxes based on the generated sensor data. The Ev-Waymo dataset consists of 80 sequences comprising a total of 157K labeled scenes, providing 100 FPS detection ground truth. + +Table 1. Performance comparison on the Ev-Waymo dataset for 3D vehicle (IoU = 0.7), pedestrian (IoU = 0.5) and cyclist (IoU = 0.5) detection following previous protocols. Evaluated on 100 FPS ground truth annotation for $t = \{0,0.1,\dots,0.9\}$ . Offline represents results obtained by interpolating data using sensor data from active timestamps 0 and 1. Online represents results obtained using data accessible at present. Specifically, online evaluation of previous works uses only the data from active timestamp 0, while we also utilize event data from active timestamp 0 up to the present time. + +
Interpolation MethodsMethods3D Detection ModalityALL (mAP/mAPH)VEH (AP/APH)PED (AP/APH)CYC (AP/APH)
L2L1L2L1L2L1L2
OfflineN-PCI [123]VoxelNeXt [9]L53.61/50.6762.96/62.5159.10/58.6768.43/60.7762.85/55.6944.55/43.1538.88/37.65
HEDNet [118]L50.52/47.5159.57/59.0955.90/55.4563.28/55.3258.03/50.6243.14/41.8037.63/36.46
N-PCI [123] + EMA [116]Focals Conv [8]L + I42.53/40.4155.49/55.0254.45/53.9948.07/43.0646.54/41.6927.70/26.6026.59/25.54
LoGoNet [47]L + I53.29/50.4963.48/63.0560.15/59.7367.85/60.2862.43/55.3742.40/41.3737.29/36.38
N-PCI [123] + CBM [39]Focals Conv [8]L + I42.51/40.3955.48/55.0154.44/53.9848.07/43.0646.54/41.6827.67/26.5726.56/25.50
LoGoNet [47]L + I53.57/50.7563.60/63.1660.27/59.8568.14/60.5362.77/55.6742.85/41.8137.66/36.74
OnlineN/AVoxelNeXt [9]L33.32/31.7044.40/44.1041.78/41.4940.52/36.2336.93/32.9624.40/23.7321.24/20.66
HEDNet [118]L31.57/29.9042.03/41.7139.32/39.0238.86/34.4335.67/31.5322.64/21.9919.72/19.14
Focals Conv [8]L + I26.27/25.0137.31/37.0136.60/36.3129.20/26.1628.41/25.4414.30/13.7813.79/13.29
LoGoNet [47]L + I33.27/31.7544.14/43.8741.73/41.4739.98/35.8436.48/32.6724.71/24.1521.59/21.10
Ev-3DOD (Ours)L + I + E48.06/45.6060.30/59.9559.19/58.8557.40/50.7855.30/48.9331.08/30.3829.69/29.03
+ +![](images/bd631d55e2438c9214c8979624d0d0cb5c08a7926a5e6d4e5f5059cbc2149882.jpg) +Figure 6. Qualitative comparisons of our method with other offline and online evaluations on the Ev-Waymo dataset. $t = 0$ represents the active time, while $t = 0.3, 0.6, 0.9$ denotes the blind times. For easier understanding, images at active timestamps 0 and 1 are overlaid. + +The DSEC 3D dataset consists of 178 sequences and includes 54K labeled scenes. For evaluation in blind time, we utilize all ground truth boxes at 100 FPS. More details of the datasets are described in the supplementary material. + +# 5. Experiments + +# 5.1. Experiment Setup + +We target causal 3D detections to detect objects during blind time. For comparison, we present two types of evaluation methods for conventional models. First is the online evaluation, where the network is inferred at any given blind time using only accessible data. Conventional methods can only process the most recent sensor data at active timestamp + +0. In contrast, our approach additionally incorporates event data from active timestamp 0 up to the current blind time. Second, offline evaluation addresses the challenge of applying conventional fixed frame rate approaches directly to blind time by utilizing future data. In other words, data for the blind time is interpolated using active timestamps 0 and 1, where timestamp 1 represents future data. Consequently, these offline evaluations serve as a performance oracle for online methods. More implementation details are provided in the supplementary material. + +# 5.2. Experimental Results + +We employ recent state-of-the-art methods, including LiDAR-based [9, 118] and LiDAR-camera-based meth + +Table 2. Results on the DSEC-3DOD dataset for 3D vehicle (IoU = 0.7), and pedestrian (IoU = 0.5) detection. + +
Interpolation MethodsMethods3D Detection ModalityALLVEHPED
mAPmAPHAPAPHAPAPH
OfflineN-PCI [123]VoxelNeXt [9]L36.2932.4345.3544.9327.2319.93
HEDNet [118]L37.7330.7341.8241.3533.6420.1
N-PCI [123] + EMA [116]Focals Conv [8]L + I30.2925.8535.5235.0425.0516.65
LoGoNet [47]L + I37.3031.3944.6444.0629.9618.72
N-PCI [123] + CBM [39]Focals Conv [8]L + I30.2825.8635.5635.0824.9916.64
LoGoNet [47]L + I37.3531.4544.7644.1729.9418.73
OnlineN/AVoxelNeXt [9]L13.9412.9021.4221.236.464.57
HEDNet [118]L13.7511.6818.4818.299.025.07
Focals Conv [8]L + I10.819.6115.7415.555.883.67
LoGoNet [47]L + I13.2911.9119.5019.327.074.49
Ev-3DOD (Ours)L + I + E31.1726.5441.6541.2020.6811.88
+ +![](images/d785061edc9d593e3dd21e232b1527bbee327bb8292696d8953f11e990f14ac4.jpg) +Figure 7. Qualitative comparisons of our method with other offline and online evaluations on the DSEC-3DOD dataset. $t = 0.3, 0.6, 0.9$ denotes the blind times. + +ods [8, 47]. For offline image interpolation, we use frame-based, EMA [116], and event-based method, CBM [39]. NeuralPCI [123] is adopted for point cloud interpolation. + +Ev-Waymo. Table 1 shows the quantitative result of Ev-Waymo. Comparing methods in the online setting, previous methods [8, 9, 47, 118], which rely on fixed frame rate sensors (i.e., LiDAR and Image), cannot detect object motion during blind times. This limitation leads to reduced performance when evaluated with 100 FPS ground truth. In contrast, our approach leverages high-temporal resolution of event data to estimate object motion during blind time, significantly outperforming existing online methods. Notably, Ev-3DOD is comparable to offline oracles and even surpasses certain approaches, such as the combination of NeuralPCI [123] + EMA [116] and Focals Conv [8]. + +Figure 6 shows comparisons between the results of online, offline, and our method during blind time. Existing online methods fail to align prediction with the ground truth during blind time. In contrast, Ev-3DOD, despite being an online approach, aligns well with the ground truth and + +Table 3. Ablation study of the proposed components. V3D-EF and MCE denote Virtual 3D Event Fusion (Sec. 3.3) and the motion confidence estimator (Sec. 3.4), respectively. + +
V3D-EFMCEL2
mAPmAPH
33.3031.68
✓46.5544.22
✓✓48.0645.60
+ +demonstrates performance comparable to offline methods. The right graph in Fig. 6 further demonstrates our method's superiority, showing that despite performance degradation with increasing blind time, the gap remains comparable to offline methods, while the online approach exhibits a much larger pixel difference. + +DSEC-3DOD. As shown in Table 2, we evaluate our model on the DSEC-3DOD real event dataset, which includes fast motion and challenging lighting conditions that significantly degrade performance during blind time. Our model effectively estimates the 3D motion of objects from events, achieving a substantial improvement over other online methods. Notably, it demonstrates performance comparable to offline oracle methods that utilize future information. The analysis of computational complexity is provided in the supplementary material. + +Figure 7 shows blind time detection inference for fast-moving objects in the DSEC dataset. Under rapid motion, the quality of the interpolated pseudo-sensor data may deteriorate, occasionally resulting in detection failures, even when future information is utilized. In contrast, the proposed Ev-3DOD accurately detects objects in blind time, predicting motion from events. + +# 5.3. Ablation Study and Analysis + +We present an ablation study on key components and various analyses on the Ev-Waymo dataset. In V3D-EF analyses, the motion confidence estimator (MCE) module is excluded for clear comparisons. + +Effectiveness of Each Component. In Table 3, we conduct the ablation study on the key components of Ev-3DOD. We evaluated a baseline and progressively added modules to + +Table 4. Effectiveness of the non-empty mask in the V3D-EF module, as defined in Eq. (1). + +
MethodL2
mAPmAPH
w/o Non-empty Mask42.5740.50
w/ Non-empty Mask46.5544.22
+ +Table 5. Comparison of results across different modalities and fusion variants of the V3D-EF module. + +
MethodL2
mAPmAPH
Base Model33.3031.68
+ LiDAR37.0435.25
+ Event42.2340.17
+ LiDAR + Event (fusion before pooling)43.9541.87
+ LiDAR + Event (fusion after pooling)46.5544.22
+ +![](images/8ae3125fd5d0942b66963a5f76085387bd963d136aaf524c2db6f185e9646496.jpg) +Figure 8. Detection performance as time elapses since the active timestamp. We compare ours with offline and online approaches. + +observe improvements. The motion confidence estimator is dependent on V3D-EF, so it was not evaluated separately. V3D-EF, which explicitly estimates motion, has a direct impact on performance, resulting in significant improvements of $+13.25\%$ in mAP and $+12.54\%$ in mAPH. Additionally, the motion confidence estimator refines prediction scores at each blind time, demonstrating its effectiveness with further gains of $+1.51\%$ and $+1.38\%$ , respectively. + +Non-empty Mask in V3D-EF. In the proposed V3D-EF, we employ a non-empty mask to achieve accurate and efficient alignment between sparse events and voxel features. To demonstrate the effectiveness of this non-empty mask, we conduct a comparative experiment in which events are projected onto all voxels within the V3D-EF module. The results in Table 4 confirm that the non-empty mask enables effective alignment between the two modalities. + +Variations of the V3D-EF Module. Our core component, the V3D-EF module, provides multiple design options. To analyze these configurations, we developed and evaluated several variants, with the results summarized in Table 5. We experimented with generating the implicit motion field using each modality, i.e., LiDAR and event sensors, and found a performance drop compared to the multi-modal approach. This decline is likely due to the limitations of each modality: LiDAR alone cannot capture motion information during + +Table 6. Evaluation results of the model trained with 100 FPS annotations using sliced events at 200 and 300 FPS. + +
Test SetL2
mAPmAPH
100 FPS48.0645.60
200 FPS47.8745.44
300 FPS47.8145.39
+ +blind times, and event features cannot accurately estimate refined 3D motion. Additionally, multi-modal fusion before pooling decreased performance. We believe this is because fusing before pooling leads to integration between voxels and events outside of box proposal regions, resulting in the use of redundant and possibly misaligned information. + +Detection Performance per Elapsed Blind Time. Figure 8 shows the performance over time elapsed since the active timestamp. Due to object motion, the performance of online detection deteriorates rapidly as the time distance from the active timestamp increases. In contrast, offline methods sustain robust performance throughout the blind time by leveraging future data, with only a slight performance drop at the center of the blind time as it moves further from the active timestamps located at both ends. The proposed model, relying solely on event data, achieves a temporal degradation pattern comparable to the offline method. + +Generalization to Higher Frame-rate. To demonstrate the generalization ability to different frame rates, we generate 200 and 300 FPS annotations for arbitrary timestamps by interpolating a pre-constructed 100 FPS dataset. Without additional training, we apply the model trained on 100 FPS. As shown in Table 6, our model effectively infers at arbitrary timestamps with minimal degradation. + +# 6. Conclusions + +This paper proposes Ev-3DOD, a novel method to utilize an event camera in 3D detection for detecting objects in blind time. To transfer information from active times to the current blind time, we estimate 3D motion based on event data. To effectively fuse sparse LiDAR data with events, we propose a Virtual 3D Event Fusion (V3D-EF) and introduce a motion confidence estimator to define confidence for 3D detection during blind times. For this study, we introduce the first event-based 3D detection dataset, emphasizing its 100 FPS annotations. We hope this work showcases the potential of neuromorphic cameras, inspiring future research. + +Limitation and Future Work. Our work is the first to address 3D detection during the blind time, establishing it as a pioneering study in this area. However, one limitation is the lack of recurrent processing of past event information to enhance temporal interaction. As a result, discontinuities may occur when fixed frame rate sensors update at active timestamps. We consider improving the interaction between consecutive event data a crucial direction for future work. + +# 7. Acknowledgements + +We thank 42dot for funding this research. We also appreciate the support and valuable discussions from the members of the 42dot research team. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636). + +# References + +[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1090-1099, 2022. 2 +[2] Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Bourtsoulatze, and Yiannis Andreopoulos. Graph-based object classification for neuromorphic vision sensing. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 491-501, 2019. 3 +[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2 +[4] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2 +[5] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 172-181, 2022. 2 +[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2 +[7] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2 +[8] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5428-5437, 2022. 6, 7 +[9] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 6, 7 +[10] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 887-897, 2022. 2 + +[11] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Deformable feature aggregation for dynamic multi-modal 3d object detection. In European conference on computer vision, pages 628-644. Springer, 2022. 2 +[12] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2 +[13] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2 +[14] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In European Conference on Computer Vision, pages 294–314. Springer, 2024. 2 +[15] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2, 4 +[16] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8458-8468, 2022. 2 +[17] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1761-1769, 2024. 2 +[18] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44 (1):154-180, 2020. 2 +[19] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. ArXiv, abs/2211.12324, 2022. 3 +[20] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629:1034 - 1040, 2024. 3 +[21] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 5 +[22] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13884-13893, 2022. 2, 3 +[23] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 5 +[24] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event + +cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[25] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 2 +[26] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3153-3163, 2021. 2 +[27] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22867-22876, 2023. 3 +[28] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11873-11882, 2020. 2 +[29] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8417-8427, 2022. 2 +[30] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2 +[31] Cheng-Ju Ho, Chen-Hsuan Tai, Yen-Yu Lin, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Diffusion-ss3d: Diffusion model for semi-supervised 3d object detection. Advances in Neural Information Processing Systems, 36:49100-49112, 2023. 2 +[32] Jordan SK Hu, Tianshu Kuai, and Steven L Waslander. Point density-aware voxels for lidar 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8469-8478, 2022. 2, 4 +[33] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2 +[34] Junjie Huang, Yun Ye, Zhujin Liang, Yi Shan, and Dalong Du. Detecting as labeling: Rethinking lidar-camera fusion in 3d object detection. In European Conference on Computer Vision, pages 439-455. Springer, 2025. 2 +[35] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2 +[36] Yu Huang, Chenzhuang Du, Zihui Xue, Xuanyao Chen, Hang Zhao, and Longbo Huang. What makes multi-modal + +learning better than single (provably). Advances in Neural Information Processing Systems, 34:10944-10956, 2021. 2 +[37] Massimiliano Iacono, Stefan Weber, Arren J. Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9, 2018. 3 +[38] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformer. In Proceedings of the AAAI conference on Artificial Intelligence, pages 1042-1050, 2023. 2 +[39] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 5, 6, 7 +[40] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2 +[41] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2 +[42] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2 +[43] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2 +[44] Dianze Li, Jianing Li, and Yonghong Tian. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:14020-14037, 2023. 3 +[45] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 3 +[46] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2 +[47] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 2, 3, 5, 6, 7 +[48] Xiaotian Li, Baojie Fan, Jiandong Tian, and Huijie Fan. Gafusion: Adaptive fusing lidar and camera with multiple guidance for 3d object detection. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21209-21218, 2024. 2 +[49] Xiang Li, Junbo Yin, Wei Li, Chengzhong Xu, Ruigang Yang, and Jianbing Shen. Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3208-3215, 2024. 2 +[50] Yijin Li, Han Zhou, Bangbang Yang, Yexin Zhang, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Graph-based asynchronous event processing for rapid object recognition. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 914-923, 2021. 3 +[51] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems, 35:18442-18455, 2022. 2 +[52] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022. +[53] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17182-17191, 2022. 2 +[54] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2 +[55] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7546-7555, 2021. 2 +[56] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2 +[57] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7345-7353, 2019. 2 +[58] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 2 +[59] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2 + +[60] Xianpeng Liu, Ce Zheng, Ming Qian, Nan Xue, Chen Chen, Zhebin Zhang, Chen Li, and Tianfu Wu. Multi-view attentive contextualization for multi-view 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16688-16698, 2024. 2 +[61] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. +[62] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 996-997, 2020. 2 +[63] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2 +[64] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 2 +[65] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2723-2732, 2021. 2 +[66] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Ji-ashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3164–3173, 2021. 2 +[67] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. ArXiv, abs/2003.09148, 2020. 3 +[68] Zhenwei Miao, Jikai Chen, Hongyu Pan, Ruiwen Zhang, Kaixuan Liu, Peihan Hao, Jun Zhu, Yang Wang, and Xin Zhan. Pvgnet: A bottom-up one-stage 3d object detector with integrated multi-level features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3279-3288, 2021. 2 +[69] Su Pang, Daniel Morris, and Hayden Radha. Clocs: Camera-lidar object candidates fusion for 3d object detection. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10386-10393. IEEE, 2020. 2 +[70] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 2 +[71] Wanli Peng, Hao Pan, He Liu, and Yi Sun. Ida-3d: Instance-depth-aware 3d object detection from stereo vision + +for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13015-13024, 2020. 2 +[72] Yan Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In AAAI Conference on Artificial Intelligence, 2023. 3 +[73] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6015-6025, 2023. 3 +[74] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16794-16804, 2024. 3 +[75] Etienne Perot, Pierre de Tournemire, Davide Oscar Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. *ArXiv*, abs/2009.13436, 2020. 3 +[76] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2 +[77] AJ Piergiovanni, Vincent Casser, Michael S Ryoo, and Anelia Angelova. 4d-net for learned multi-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15435-15445, 2021. 2 +[78] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2 +[79] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[80] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2 +[81] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2 +[82] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2 +[83] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2 + +[84] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3 +[85] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2743-2752, 2021. 2 +[86] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2 +[87] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2 +[88] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1711-1719, 2020. 2 +[89] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2 +[90] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 2 +[91] Jiaming Sun, Linghao Chen, Yiming Xie, Siyu Zhang, Qin-hong Jiang, Xiaowei Zhou, and Hujun Bao. Disp r-cnn: Stereo 3d object detection via shape prior guided instance disparity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10548-10557, 2020. 2 +[92] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2 +[93] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2443-2451, 2019. 5 +[94] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2 + +[95] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2 +[96] Dongsheng Wang, Xu Jia, Yang Zhang, Xinyu Zhang, Yaoyuan Wang, Ziyang Zhang, D. Wang, and Huchuan Lu. Dual memory aggregation network for event-based object detection with learnable representation. In AAAI Conference on Artificial Intelligence, 2023. 3 +[97] Wenjie Wang, Yehao Lu, Guangcong Zheng, Shuigen Zhan, Xiaqing Ye, Zichang Tan, Jingdong Wang, Gaoang Wang, and Xi Li. Bevspread: Spread voxel pooling for bird's-eye-view representation in vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14718-14727, 2024. 2 +[98] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2 +[99] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2 +[100] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 2 +[101] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multimodal sparse representations for multi-sensor 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 17545-17556, 2023. 2 +[102] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 2 +[103] Jianyun Xu, Zhenwei Miao, Da Zhang, Hongyu Pan, Kaixuan Liu, Peihan Hao, Jun Zhu, Zhengyang Sun, Hongmin Li, and Xin Zhan. Int: Towards infinite-frames 3d detection with an efficient framework. In European Conference on Computer Vision, pages 193–209. Springer, 2022. 2 +[104] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Zhou Bin, and Liangjun Zhang. Fusionpainting: Multimodal fusion with adaptive attention for 3d object detection. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3047-3054. IEEE, 2021. 2 +[105] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 5 +[106] Yixin Yang, Jinxiu Liang, Bohan Yu, Yan Chen, Jimmy S Ren, and Boxin Shi. Latency correction for event-guided + +deblurring and frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24977-24986, 2024. 2 +[107] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 2 +[108] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2 +[109] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. Advances in Neural Information Processing Systems, 35:1992-2005, 2022. 2 +[110] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhao-Xu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10201-10210, 2021. 3 +[111] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022. 2 +[112] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In European conference on computer vision, pages 17-33. Springer, 2022. 2 +[113] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 2 +[114] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 4, 5 +[115] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kili-ian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. arXiv preprint arXiv:1906.06310, 2019. 2 +[116] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 6, 7 +[117] Gang Zhang, Junnan Chen, Guohuan Gao, Jianmin Li, Si Liu, and Xiaolin Hu. Safdnet: A simple and effective network for fully sparse 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14477-14486, 2024. 2 +[118] Gang Zhang, Chen Junnan, Guohuan Gao, Jianmin Li, and Xiaolin Hu. Hednet: A hierarchical encoder-decoder net + +work for 3d object detection in point clouds. Advances in Neural Information Processing Systems, 36, 2024. 6, 7 +[119] Jiqing Zhang, B. Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8791-8800, 2022. 3 +[120] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9781-9790, 2023. 2 +[121] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2 +[122] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 2 +[123] Zehan Zheng, Danni Wu, Ruisi Lu, Fan Lu, Guang Chen, and Changjun Jiang. Neuralpci: Spatio-temporal neural field for 3d point cloud multi-frame non-linear interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 909–918, 2023. 5, 6, 7 +[124] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1839-1849, 2020. 2 +[125] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2, 3, 4 +[126] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3 +[127] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12800-12810, 2023. 3 \ No newline at end of file diff --git a/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/images.zip b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5286775c3e215535c094b0dec134f87512ac7da9 --- /dev/null +++ b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8928a21afa516c25906884ddc0959769ec5085c0ed60e18c8cfbf89eb0df5b86 +size 834679 diff --git a/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/layout.json b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac7d69e46bd03802a8e74a345c2537e5d966512a --- /dev/null +++ b/2025/Ev-3DOD_ Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras/layout.json @@ -0,0 +1,10612 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 102, + 518, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 102, + 518, + 138 + ], + "spans": [ + { + "bbox": [ + 93, + 102, + 518, + 138 + ], + "type": "text", + "content": "Ev-3DOD: Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 161, + 471, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 161, + 471, + 188 + ], + "spans": [ + { + "bbox": [ + 140, + 161, + 471, + 188 + ], + "type": "text", + "content": "Hoonhee Cho*, Jae-Young Kang*, Youngho Kim, and Kuk-Jin Yoon KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 191, + 440, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 191, + 440, + 204 + ], + "spans": [ + { + "bbox": [ + 168, + 191, + 440, + 204 + ], + "type": "text", + "content": "{gnsgnsgml,mickeykang,kmax2001,kjyoon}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 221, + 225, + 401 + ], + "blocks": [ + { + "bbox": [ + 60, + 221, + 225, + 401 + ], + "lines": [ + { + "bbox": [ + 60, + 221, + 225, + 401 + ], + "spans": [ + { + "bbox": [ + 60, + 221, + 225, + 401 + ], + "type": "image", + "image_path": "e63f2f6992edaa7daed8436d54be8bd4c5979af92b981593186c691844ac6ae3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 403, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 54, + 403, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. Comparison between the proposed Ev-3DOD and conventional 3D detection methods. Fixed frame rate sensors (e.g., LiDAR, frame camera) inevitably have periods of \"blind time\" " + }, + { + "bbox": [ + 54, + 403, + 555, + 460 + ], + "type": "inline_equation", + "content": "(t_{0\\rightarrow 1})" + }, + { + "bbox": [ + 54, + 403, + 555, + 460 + ], + "type": "text", + "content": " when they lack data between active timestamps (0 and 1). Consequently, as shown in (a), conventional methods cannot detect objects during blind time, while other objects like vehicles can undergo significant motion within these gaps. In contrast, as illustrated in (b), our method, Ev-3DOD, leverages the high temporal resolution of the event camera, enabling accurate object detection even during the blind times of other sensors." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 244, + 228, + 372, + 395 + ], + "blocks": [ + { + "bbox": [ + 244, + 228, + 372, + 395 + ], + "lines": [ + { + "bbox": [ + 244, + 228, + 372, + 395 + ], + "spans": [ + { + "bbox": [ + 244, + 228, + 372, + 395 + ], + "type": "image", + "image_path": "f2be3700a8221c22bd980928c3d6da76f63ea82f142b6881b969ce148df3640f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 396, + 228, + 538, + 395 + ], + "blocks": [ + { + "bbox": [ + 396, + 228, + 538, + 395 + ], + "lines": [ + { + "bbox": [ + 396, + 228, + 538, + 395 + ], + "spans": [ + { + "bbox": [ + 396, + 228, + 538, + 395 + ], + "type": "image", + "image_path": "115f667de511432d6fcb088013b37a0e501177ccbb0f550566a3d8d130f7074d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 468, + 201, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 468, + 201, + 481 + ], + "spans": [ + { + "bbox": [ + 151, + 468, + 201, + 481 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "type": "text", + "content": "Detecting 3D objects in point clouds plays a crucial role in autonomous driving systems. Recently, advanced multimodal methods incorporating camera information have achieved notable performance. For a safe and effective autonomous driving system, algorithms that excel not only in accuracy but also in speed and low latency are essential. However, existing algorithms fail to meet these requirements due to the latency and bandwidth limitations of fixed frame rate sensors, e.g., LiDAR and camera. To address this limitation, we introduce asynchronous event cameras into 3D object detection for the first time. We leverage their high temporal resolution and low bandwidth to enable high-speed 3D object detection. Our method enables detection even during inter-frame intervals when synchronized data is unavailable, by retrieving previous 3D information through the event camera. Furthermore, we introduce the first event-based 3D object detection datasets, EvWaymo and DSEC-3DOD, both of which include ground" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 470, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 556, + 518 + ], + "type": "text", + "content": "truth 3D bounding boxes at 100 FPS, establishing the first benchmarks for event-based 3D detectors. The code and dataset are available at https://github.com/mickeykang16/Ev3DOD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 542, + 395, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 542, + 395, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 542, + 395, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 556, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 556, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 556, + 694 + ], + "type": "text", + "content": "3D object detection is a crucial task in autonomous driving. Multi-modal sensors like LiDAR and cameras are commonly used to understand scenes and handle complex road scenarios. This sensor setup consists of synchronized sensors with a fixed frame rate and has limited speed due to bandwidth limitations. As shown in Fig. 1 (a), the blind time, which is the period when sensor information is unavailable (e.g., 100 milliseconds), is critical as cars can experience large motions as 3 meters during this time. Therefore, temporal resolution is a crucial factor that can directly influence accident risk." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 325, + 702, + 464, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 464, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 464, + 714 + ], + "type": "text", + "content": "*The first two authors contributed equally." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27197" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "Recently, event cameras [18] have introduced a new paradigm in perceptual research. Event cameras provide brightness changes as a stream of data with sub-milisecond latency, making them robust to motion blur [41, 106] and enabling low bandwidth [22, 94]. Event cameras overcome the trade-off between bandwidth and blind time but have the drawback of sacrificing intensity information. To address this, recent work [12, 13, 92, 120] has combined event cameras with frame cameras to complement the sparsity of event cameras while overcoming the fixed frame rate sensor's limits." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 294, + 348 + ], + "type": "text", + "content": "In this paper, we aim for accurate 3D object detection in blind time through a multi-modal approach that combines LiDAR and frame-based sensors with an event camera, leveraging its high temporal resolution. As shown in Fig. 1 (b), event data is available during blind times, while the LiDAR and image data from the most recent active timestamp are accessible. Therefore, there is a misalignment in timestamps between the fixed frame sensor data and the current blind time. Then, this raises the question: How can we estimate 3D motion using only event data, especially when data from other sensors is unavailable at the current time?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 350, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 578 + ], + "type": "text", + "content": "To address this, we propose a virtual 3D event fusion (V3D-EF) that transfers past 3D information to align with the present time using the event camera. As extensively covered in previous research [14, 24, 40, 89], event data inherently contains motion information on an image plane. Based on this observation, we hypothesized that providing a 3D cue associated with the initial timestamp to the event information would enable accurate estimation of the 3D motion vector. To this end, we project event features into a 3D voxel space to create virtual 3D event features. In this process, events are projected only onto regions where point cloud data exists, aligning with the point cloud data. Then, we aggregate the event features with the voxel features to estimate an implicit motion field that captures the motion inherent within the grid. The generated implicit motion field encapsulates both the 3D spatial information from voxels and the motion information from events, enabling the calculation of a 3D motion vector from the most recent active timestamp to the present blind time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 580, + 294, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 294, + 651 + ], + "type": "text", + "content": "Another crucial factor in 3D object detection is the score for each bounding box. The model considers both box proposal scores at the active timestamp and the motion score of each predicted box in blind time. In blind time, the model aligns 3D information of voxel features and motion information of events to learn the confidence of box motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": "A major hurdle in 3D detection research using event cameras is the lack of suitable datasets. To address this, we present two datasets: Ev-Waymo, which includes synthetic event data, and DSEC-3DOD, which contains real event data. Unlike existing public 3D object detection datasets," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "ours include annotations not only for active timestamps with 3D LiDAR sensor data but also for blind times, providing annotation in 100 FPS frame rate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 119, + 406, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 119, + 406, + 131 + ], + "spans": [ + { + "bbox": [ + 314, + 119, + 406, + 131 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 139, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 553, + 330 + ], + "type": "text", + "content": "Single-modality 3D Object Detection. Given that cameras offer significant cost advantages over LiDAR sensors, many researchers have focused on developing methods that leverage camera-based systems for 3D object detection using image-only inputs [7, 10, 26, 30, 33, 38, 46, 54, 60-62, 64, 71, 91, 97, 99, 115]. Image-based 3D object detection faces challenges due to the lack of direct depth information. To overcome this challenge, several studies [7, 70, 76, 83, 98, 115] have employed depth estimation techniques to generate pseudo 3D point clouds or elevate 2D features into 3D space for object detection, alongside proposals for transformer-based architectures [35, 99, 121] that utilize 3D object queries and 3D-2D correspondences. Nonetheless, accurately estimating depth from images remains difficult, resulting in image-based methods performing worse than LiDAR-based approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 486 + ], + "type": "text", + "content": "LiDAR-based 3D object detection can reliably estimate 3D bounding boxes using point clouds captured from LiDAR sensors. Current LiDAR-based detection methods can be divided into three categories based on different point cloud encoding formats: point-based methods [17, 32, 79, 80, 82, 86, 88, 108, 112, 124], voxel-based methods [4, 15, 16, 25, 29, 29, 31, 43, 49, 55, 65, 66, 85, 100, 103, 111, 117, 125], and point-voxel fusion networks [6, 28, 68, 87, 107]. While LiDAR detection is advantageous in various conditions, it often struggles in areas with sparse LiDAR data. Therefore, integrating the geometric benefits of point clouds with the semantic richness of images remains crucial for improving performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 487, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 713 + ], + "type": "text", + "content": "Multi-modality 3D Object Detection. Multi-modal sensor fusion [5, 34, 36, 47, 48, 56, 57, 59, 90, 101, 102, 113], where different sensors complement each other, typically utilizes a combination of cameras and LiDAR. Multimodal 3D object detection has gained attention for improving accuracy over unimodal methods. Fusion approaches are categorized into early [3, 11, 42, 81, 95, 104], middle [51-53, 58, 63, 77, 78, 109, 122], and late fusions [1, 47, 69], with middle fusion now preferred due to its robustness to calibration errors and enhanced feature interaction. Crossmodal fusion has achieved significant performance improvements by enhancing the sparse features of point clouds with semantic information from images. However, there are still unexplored areas concerning safety in autonomous driving. LiDAR and cameras have limited time resolution (i.e., " + }, + { + "bbox": [ + 313, + 487, + 553, + 713 + ], + "type": "inline_equation", + "content": "10 - 20\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 487, + 553, + 713 + ], + "type": "text", + "content": ") due to the relatively high bandwidths. To enhance safety, algorithms that operate at high speeds are essential. To address this, we introduce an asynchronous event camera with high temporal resolution for 3D object" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27198" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 541, + 262 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 541, + 262 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 541, + 262 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 541, + 262 + ], + "type": "image", + "image_path": "2cba8d8d590ff446bb95c44596e9b699280aa311b7afbaffef249b015571c0b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 322 + ], + "type": "text", + "content": "Figure 2. The overall pipeline for event-based 3D object detection (Ev-3DOD). (a): At active timestamp 0, LiDAR and image data are available. Therefore, we utilize an RGB-LiDAR Region Proposal Network (RPN) to extract voxel features and 3D bounding boxes at the active timestamp. (b): To predict the bounding box during blind time, " + }, + { + "bbox": [ + 54, + 266, + 555, + 322 + ], + "type": "inline_equation", + "content": "0 \\leq t < 1" + }, + { + "bbox": [ + 54, + 266, + 555, + 322 + ], + "type": "text", + "content": ", we estimate the 3D motion and confidence score using event features. For computational efficiency, we design the process to compute (a) only once before the next active timestamp while performing iterative computations solely for (b)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 342, + 294, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 294, + 389 + ], + "type": "text", + "content": "detection for the first time. The proposed method performs multi-modal fusion between synchronized sensors and asynchronous events to increase the operational frequency of 3D detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 390, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 390, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 56, + 390, + 296, + 618 + ], + "type": "text", + "content": "Event-based Object Detection. Event cameras break through the limitations of bandwidth requirements and perceptual latency with the trade-off of sparse data and loss of intensity information. To make the most of these characteristics, prior work has designed models to adapt to event data using graph neural networks [2, 19, 20, 50, 84], spiking neural networks [110, 119], and sparse neural networks [67, 74]. Recently, effective dense neural networks [27, 37, 45, 72, 75, 96, 127] and transformer architectures [22, 44, 73, 74] have been developed to leverage event cameras, aiming to achieve both low latency and high performance. These advances enable real-time operation with impressive accuracy, demonstrating the practical applicability of event cameras in real-world scenarios. To the best of our knowledge, we are the first to extend event-based object detection from 2D to 3D space. By employing a multimodal fusion approach of the event camera and 3D sensors (i.e., LiDAR), we aim to overcome the latency and persistent interval time barriers of 3D object detections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 170, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 170, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 170, + 659 + ], + "type": "text", + "content": "3.1. Problem Definition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "LiDAR and camera data are only available at active timestamps 0 and 1, where 1 represents invalid future information in an online process. Our goal is to predict 3D bounding boxes during " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "0 \\leq t < 1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", including the blind time" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": "when synchronized sensors are unavailable. Given a pair of synchronized 3D points " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "P_0 = \\{(x_l, y_l, z_l)\\}_{l=1}^N" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "I_0 \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " at active timestamp 0, we leverage event stream " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{0 \\to t} = \\{(i, j, p, \\tau) | 0 \\leq \\tau < t\\}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " to detect 3D bounding box set " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "B_t" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " at an arbitrary time " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "0 \\leq t < 1" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 409, + 440, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 440, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 440, + 421 + ], + "type": "text", + "content": "3.2. Framework Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": "As illustrated in (a) of Fig. 2, at active timestamp 0, both LiDAR and image data are available. RGB-LiDAR Region Proposal Network (RPN) generates bounding box proposals from rich 3D information in active time. We begin by voxelizing the point cloud and generating voxel features, denoted as " + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_0\\in \\mathbb{R}^{D_X\\times D_Y\\times D_Z\\times C}" + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "D_{X}" + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "D_{Y}" + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "D_Z" + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": " represent the grid size, and " + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": " denotes the channel dimension, using a 3D backbone [125]. Then, using existing region proposals [114] and refinement networks [47] with image features, we generate bounding box proposals, " + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "B_0 = \\{B_0^1,B_0^2,\\dots ,B_0^n\\}" + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": " and box proposal confidence " + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "p_0 = \\{p_0^1,p_0^2,\\dots ,p_0^n\\} \\in [0,1]" + }, + { + "bbox": [ + 312, + 426, + 555, + 629 + ], + "type": "text", + "content": " for the active timestamp 0. RGB-LiDAR-based multi-modal networks have been widely studied, leading to many advanced models capable of accurate estimation at active timestamps. We tuned the model with fewer parameters and a reduced channel dimension to improve efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Our emphasis is on achieving precise estimation throughout the blind time, spanning the interval from active timestamp 0 to the subsequent timestamp 1. To infer bounding boxes at an arbitrary inter-frame time " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", we first convert the event stream, " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{0\\rightarrow t}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " into voxel grids [126] as " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_t" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ". Then, we use a feature encoder to produce event features, " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t\\in \\mathbb{R}^{H / 4\\times W / 4\\times C}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " is the spa" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27199" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "tial dimensions of the events. Our objective is to estimate the 3D motion of objects during blind time. To achieve this, we utilize the bounding box set " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "B_{0}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " and voxel features " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " from the active timestamp 0. During blind times, we share the voxel features, " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ", generated at the active timestamp, ensuring that voxel features are computed only once to maintain a cost-effective structure. Through the Virtual 3D Event Fusion (V3D-EF), we generate an implicit motion field for each bounding box by integrating event and voxel features, encapsulating motion information. Using this field, we compute the 3D motion and confidence score, enabling the estimation of 3D boxes at the blind time " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 224, + 295, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 237 + ], + "type": "text", + "content": "3.3. Virtual 3D Event Fusion for Motion Prediction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 386 + ], + "type": "text", + "content": "The event feature continuously captures each pixel's motion information over time on the 2D image plane. However, estimating 3D motion using only events is a challenge. We hypothesize that providing each event feature with adequate 3D information would allow for effective 3D motion estimation. To achieve this, we aim to incorporate a 3D cue into the event features. Since both event features and voxel features are sparse, it is crucial to align each voxels with the corresponding events accurately. Directly using the voxel features would cause misalignment with the actual 3D geometry. Thus, we adopt the voxel point centroid method employed in previous works [32]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": "Specifically, we first consider the non-empty voxel features as " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_0 = \\{V_0^k = \\{h_0^k,f_0^k\\} \\mid k = 1,\\dots ,N_\\nu \\}" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "h_0^k" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " is the 3D voxel index, " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "f_0^k" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " is the associated voxel feature vector, and " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " is the number of non-empty voxels. Then, points within each voxel are grouped into a set " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(V_0^k)" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " by determining their voxel index " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "h_0^k" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " based on their spatial coordinates " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "X_{i} = (x_{i},y_{i},z_{i})" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " of points. The centroid of each voxel feature is subsequently computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 491, + 295, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 491, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 114, + 491, + 295, + 523 + ], + "type": "interline_equation", + "content": "c _ {0} ^ {k} = \\frac {1}{\\left| \\mathcal {N} \\left(V _ {0} ^ {k}\\right) \\right|} \\sum_ {X _ {i} \\in \\mathcal {N} \\left(V _ {0} ^ {k}\\right)} X _ {i}. \\tag {1}", + "image_path": "fb6ac71865f142eab0728c04211c499db87361cc04f4578f4912da212802fd3b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": "The centroid " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "c_0^k" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": " is calculated for non-empty voxels, ensuring it corresponds only to indices with existing points, with the goal of aligning these centroids to the event data. As shown in the top of Fig. 3, we propose a method to project events into a virtual voxel space based on the coordinates of the centroids. The projection onto the image plane from 3D space can be obtained as: " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_0^k = K \\cdot E \\cdot c_0^k" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": " denote the intrinsic and extrinsic parameters, " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": " operation is matrix multiplication. Therefore, we calculate the event coordinates corresponding to each centroid, " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "c_0^k" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": ", and gather the event features, " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": ", at these coordinates, " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_0^k" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": ", to create virtual 3D event features, " + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t^E = \\{\\mathbf{E}_t(\\mathbf{p}_0^k)\\}_{k=1}^{N_\\nu} \\in \\mathbb{R}^{N_\\nu \\times C}" + }, + { + "bbox": [ + 55, + 533, + 296, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "content": "Since we aim to estimate the 3D motion of bounding boxes from the active timestamp " + }, + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "content": " up to the current time " + }, + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "content": ", we consider fusion only for the features corresponding" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 550, + 243 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 550, + 243 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 550, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 550, + 243 + ], + "type": "image", + "image_path": "4fea34eff8b6df80bcb906bf14b904c30978e89f1b4e2aa1c020bc200ee82d21.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 316 + ], + "type": "text", + "content": "Figure 3. Virtual 3D Event Fusion (V3D-EF). We generate an implicit motion field for each bounding box by applying the ROI pooling separately to the virtual 3D event features and the voxel features. Virtual 3D event features represent the 3D projection of 2D events. None-empty voxels of the voxel feature, " + }, + { + "bbox": [ + 313, + 251, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_0" + }, + { + "bbox": [ + 313, + 251, + 555, + 316 + ], + "type": "text", + "content": ", are colored." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": "to the box proposal. Given the bounding box proposals, " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "B_0" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ", at active timestamp 0, we divide box proposals into " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "S \\times S \\times S" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": " regular sub-voxels. Then, we independently perform the ROI pooling [15] on the virtual 3D event features " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t^E" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ", and the voxel features " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_0" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ", corresponding to each subvoxel. This generates grid features for the bounding boxes, denoted as " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_t^E \\in \\mathbb{R}^{n \\times S^3 \\times C}" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_0 \\in \\mathbb{R}^{n \\times S^3 \\times C}" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": " is the number of box proposals. The generated features are then concatenated and passed through an MLP to produce the implicit motion field, " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t = \\{\\mathcal{M}_t^0, \\mathcal{M}_t^1, \\dots, \\mathcal{M}_t^n\\}" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ". The implicit motion field is created by aligning 3D events with their corresponding voxel features, encapsulating motion information for the bounding box. Using an MLP, we estimate the explicit 3D motion. The explicit box motion " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{0 \\rightarrow t}^i = (dx^i, dy^i, dz^i, d\\alpha^i)" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ", defined in box local coordinate of " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "B_0^i" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ", contains spatial shift and yaw rotation. The final box location is calculated as " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "B_t^i = g(B_0^i, \\mathbf{M}_{0 \\rightarrow t}^i)" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 335, + 555, + 552 + ], + "type": "text", + "content": " is a box shifting operation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 476, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 476, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 476, + 574 + ], + "type": "text", + "content": "3.4. Motion Confidence Estimator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "As discussed in previous studies [114, 125], self-awareness of the confidence of predicted bounding boxes is a key factor. The simplest choice of final confidence score in our setting would be the score of active timestamp 0. However, the final score should take into account the object's motion prediction during the blind time. If the object experiences challenging motion during the blind time, a reliable bounding box at the active timestamp 0 could become less reliable. Therefore, the final score should consider two different sources. The first aspect is the score estimated at the active timestamp, based on LiDAR and images. The second" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27200" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 292, + 133 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 292, + 133 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 292, + 133 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 292, + 133 + ], + "type": "image", + "image_path": "e8306ee1b72f89ba50f37e89d5338ba97ff78a409e54ff2fa9f2db5ad88e9618.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "type": "text", + "content": "Figure 4. Ev-Waymo Dataset samples. Synthetic event data, label overlaid image, and LiDAR from left to right." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 295, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 295, + 193 + ], + "type": "text", + "content": "aspect is the confidence score based on motion during the blind time using events." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 193, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 299 + ], + "type": "text", + "content": "The implicit motion field from Sec. 3.3 contains 3D motion information through alignment between the point cloud and events. We additionally employ an MLP specifically for calculating motion confidence. Box-wise motion confidence is estimated and normalized, resulting in " + }, + { + "bbox": [ + 55, + 193, + 296, + 299 + ], + "type": "inline_equation", + "content": "p_{0\\rightarrow t} = \\{p_{0\\rightarrow t}^{1},p_{0\\rightarrow t}^{2},\\dots ,p_{0\\rightarrow t}^{n}\\} \\in [0,1]" + }, + { + "bbox": [ + 55, + 193, + 296, + 299 + ], + "type": "text", + "content": ". The final confidence score for box proposal " + }, + { + "bbox": [ + 55, + 193, + 296, + 299 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 193, + 296, + 299 + ], + "type": "text", + "content": " is calculated by considering both the score at the active timestamp and the motion confidence:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 300, + 294, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 300, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 143, + 300, + 294, + 314 + ], + "type": "interline_equation", + "content": "p _ {t} ^ {i} = p _ {0} ^ {i} \\cdot p _ {0 \\rightarrow t} ^ {i} \\tag {2}", + "image_path": "d648dd64b59849208871af5a1758eb0625ed02258ce04ce4cf2063550d48ff7e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 171, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 171, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 171, + 332 + ], + "type": "text", + "content": "3.5. Training Objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 337, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 408 + ], + "type": "text", + "content": "For RGB-LiDAR RPN in the active timestamp, we follow RPN loss from [47, 105, 114]. For blind time prediction, box regression loss from [114] is applied between the predicted 3D box " + }, + { + "bbox": [ + 55, + 337, + 296, + 408 + ], + "type": "inline_equation", + "content": "B_{t}^{i}" + }, + { + "bbox": [ + 55, + 337, + 296, + 408 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 55, + 337, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\hat{B}_{t}^{i}" + }, + { + "bbox": [ + 55, + 337, + 296, + 408 + ], + "type": "text", + "content": ". The target assigned to the motion confidence branch is an IoU-related value, as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 415, + 272, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 415, + 272, + 460 + ], + "spans": [ + { + "bbox": [ + 77, + 415, + 272, + 460 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {0 \\to t} ^ {i} (\\mathrm {I o U} _ {i}) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f I o U _ {i} < \\theta_ {L}}, \\\\ \\frac {\\mathrm {I o U} _ {i} - \\theta_ {L}}{\\theta_ {H} - \\theta_ {L}} & \\text {i f \\theta_ {L} \\leq I o U _ {i} < \\theta_ {H}}, \\\\ 1 & \\text {i f I o U _ {i} \\geq \\theta_ {H}}, \\end{array} \\right.", + "image_path": "dc0d2a5311492c9c15fc73055786a56757844f7c2effe247332adff7f7da8179.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_i" + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": " is the IoU between the " + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": "-th prediction and the corresponding ground truth box, while " + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\theta_H" + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\theta_L" + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": " are foreground and background IoU thresholds. Binary Cross Entropy Loss is exploited here for confidence prediction. The losses of our detect head are computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 539, + 295, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 539, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 97, + 539, + 295, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {R P N} + \\lambda_ {1} \\cdot \\mathcal {L} _ {r e g} + \\lambda_ {2} \\cdot \\mathcal {L} _ {s c o r e} \\tag {3}", + "image_path": "a552f14d18a3ac649a381a256c8d6415d889dc1cab3fbdb38508fffd4f4ae034.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 561, + 287, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 287, + 575 + ], + "type": "text", + "content": "4. Event-based 3D Object Detection Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 581, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 665 + ], + "type": "text", + "content": "This study is the first to explore the use of event cameras for 3D detection. To validate our model's effectiveness, we introduce two event 3D detection datasets: Ev-Waymo, which consists of synthetic events, and DSEC-3DOD with real events. We first obtained accurate 3D bounding box annotations for the active time, when both LiDAR and image data are available, and then processed the blind time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 255, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 255, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 255, + 685 + ], + "type": "text", + "content": "4.1. Annotation for Active Time at 10 FPS" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Ev-Waymo Dataset is a synthetic event dataset based on the Waymo Open Dataset (WOD) [93]. WOD provides" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 547, + 193 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 547, + 193 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 547, + 193 + ], + "type": "image", + "image_path": "ad2e992570a22964759e6efed473190c28a1a597f9c7e478e059765260d59afc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 195, + 553, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 195, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 553, + 217 + ], + "type": "text", + "content": "Figure 5. DSEC-3DOD Dataset samples. Event data, label overlaid image, and accumulated LiDAR from left to right." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 235, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 554, + 342 + ], + "type": "text", + "content": "sequential pairs of synchronized LiDAR, image data, and 3D box annotations at 10 FPS. As shown in Fig. 4, event streams are generated using a widely-adopted event simulator, Vid2E [21]. The Ev-Waymo evaluation follows the same metrics as WOD, calculating both average precision (AP) and average precision weighted by heading (APH). Adhering to the DSEC-3DOD setting, Ev-Waymo employs the front camera and a cropped LiDAR aligned to the corresponding field of view." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 342, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 474 + ], + "type": "text", + "content": "DSEC-3DOD. The DSEC dataset [23] is a stereo event dataset that includes LiDAR, image, and real event data captured in challenging environments. We sampled a portion of the \"zurich_city\" sequence, providing LiDAR and image data at 10 FPS. We hired annotation experts to label 3D bounding boxes at 10 FPS, ensuring alignment with the sensor data. As shown in Fig. 5, DSEC-3DOD contains real events, incorporating noise and capturing challenging environments, unlike simulated events in Ev-Waymo. During the annotation process, we accumulated LiDAR points to precisely generate ground truth 3D bounding boxes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "text", + "content": "For evaluation on DSEC-3DOD, we used the same evaluation metrics as Ev-Waymo but include only vehicle and pedestrian classes due to a lack of cyclist data, with all objects categorized under a single difficulty level." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 529, + 515, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 515, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 515, + 540 + ], + "type": "text", + "content": "4.2. Annotation for Blind Time at 100 FPS" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "For training and evaluation on 3D detection during blind time, we created 3D box annotations for periods when Li-DAR and images were absent. We linearly interpolated the original 10 FPS annotations, dividing blind times into 10 evenly spaced intervals between consecutive frames, resulting in temporally dense annotations at 100 FPS. To ensure accurate bounding box ground truths in instances where Li-DAR and image data are unavailable, we employed event-based video frame interpolation [39] and point cloud interpolation [123] methods to generate sensor data for blind time. Data annotation experts refined interpolated bounding boxes based on the generated sensor data. The Ev-Waymo dataset consists of 80 sequences comprising a total of 157K labeled scenes, providing 100 FPS detection ground truth." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27201" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 128, + 550, + 269 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Performance comparison on the Ev-Waymo dataset for 3D vehicle (IoU = 0.7), pedestrian (IoU = 0.5) and cyclist (IoU = 0.5) detection following previous protocols. Evaluated on 100 FPS ground truth annotation for " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "t = \\{0,0.1,\\dots,0.9\\}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": ". Offline represents results obtained by interpolating data using sensor data from active timestamps 0 and 1. Online represents results obtained using data accessible at present. Specifically, online evaluation of previous works uses only the data from active timestamp 0, while we also utilize event data from active timestamp 0 up to the present time." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 128, + 550, + 269 + ], + "lines": [ + { + "bbox": [ + 61, + 128, + 550, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 550, + 269 + ], + "type": "table", + "html": "
Interpolation MethodsMethods3D Detection ModalityALL (mAP/mAPH)VEH (AP/APH)PED (AP/APH)CYC (AP/APH)
L2L1L2L1L2L1L2
OfflineN-PCI [123]VoxelNeXt [9]L53.61/50.6762.96/62.5159.10/58.6768.43/60.7762.85/55.6944.55/43.1538.88/37.65
HEDNet [118]L50.52/47.5159.57/59.0955.90/55.4563.28/55.3258.03/50.6243.14/41.8037.63/36.46
N-PCI [123] + EMA [116]Focals Conv [8]L + I42.53/40.4155.49/55.0254.45/53.9948.07/43.0646.54/41.6927.70/26.6026.59/25.54
LoGoNet [47]L + I53.29/50.4963.48/63.0560.15/59.7367.85/60.2862.43/55.3742.40/41.3737.29/36.38
N-PCI [123] + CBM [39]Focals Conv [8]L + I42.51/40.3955.48/55.0154.44/53.9848.07/43.0646.54/41.6827.67/26.5726.56/25.50
LoGoNet [47]L + I53.57/50.7563.60/63.1660.27/59.8568.14/60.5362.77/55.6742.85/41.8137.66/36.74
OnlineN/AVoxelNeXt [9]L33.32/31.7044.40/44.1041.78/41.4940.52/36.2336.93/32.9624.40/23.7321.24/20.66
HEDNet [118]L31.57/29.9042.03/41.7139.32/39.0238.86/34.4335.67/31.5322.64/21.9919.72/19.14
Focals Conv [8]L + I26.27/25.0137.31/37.0136.60/36.3129.20/26.1628.41/25.4414.30/13.7813.79/13.29
LoGoNet [47]L + I33.27/31.7544.14/43.8741.73/41.4739.98/35.8436.48/32.6724.71/24.1521.59/21.10
Ev-3DOD (Ours)L + I + E48.06/45.6060.30/59.9559.19/58.8557.40/50.7855.30/48.9331.08/30.3829.69/29.03
", + "image_path": "a963eee106f7d129f65437d7a699a469dfbf09aafc729111f8ab7eba122d2816.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 277, + 549, + 492 + ], + "blocks": [ + { + "bbox": [ + 59, + 277, + 549, + 492 + ], + "lines": [ + { + "bbox": [ + 59, + 277, + 549, + 492 + ], + "spans": [ + { + "bbox": [ + 59, + 277, + 549, + 492 + ], + "type": "image", + "image_path": "bd631d55e2438c9214c8979624d0d0cb5c08a7926a5e6d4e5f5059cbc2149882.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 495, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 555, + 518 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of our method with other offline and online evaluations on the Ev-Waymo dataset. " + }, + { + "bbox": [ + 55, + 495, + 555, + 518 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 495, + 555, + 518 + ], + "type": "text", + "content": " represents the active time, while " + }, + { + "bbox": [ + 55, + 495, + 555, + 518 + ], + "type": "inline_equation", + "content": "t = 0.3, 0.6, 0.9" + }, + { + "bbox": [ + 55, + 495, + 555, + 518 + ], + "type": "text", + "content": " denotes the blind times. For easier understanding, images at active timestamps 0 and 1 are overlaid." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "text", + "content": "The DSEC 3D dataset consists of 178 sequences and includes 54K labeled scenes. For evaluation in blind time, we utilize all ground truth boxes at 100 FPS. More details of the datasets are described in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 601, + 137, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 137, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 137, + 616 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 622, + 163, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 163, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 163, + 635 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "We target causal 3D detections to detect objects during blind time. For comparison, we present two types of evaluation methods for conventional models. First is the online evaluation, where the network is inferred at any given blind time using only accessible data. Conventional methods can only process the most recent sensor data at active timestamp" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 540, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 660 + ], + "type": "text", + "content": "0. In contrast, our approach additionally incorporates event data from active timestamp 0 up to the current blind time. Second, offline evaluation addresses the challenge of applying conventional fixed frame rate approaches directly to blind time by utilizing future data. In other words, data for the blind time is interpolated using active timestamps 0 and 1, where timestamp 1 represents future data. Consequently, these offline evaluations serve as a performance oracle for online methods. More implementation details are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 670, + 439, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 439, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 439, + 684 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We employ recent state-of-the-art methods, including LiDAR-based [9, 118] and LiDAR-camera-based meth" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27202" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 83, + 550, + 207 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 503, + 81 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 503, + 81 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 503, + 81 + ], + "type": "text", + "content": "Table 2. Results on the DSEC-3DOD dataset for 3D vehicle (IoU = 0.7), and pedestrian (IoU = 0.5) detection." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 83, + 550, + 207 + ], + "lines": [ + { + "bbox": [ + 61, + 83, + 550, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 83, + 550, + 207 + ], + "type": "table", + "html": "
Interpolation MethodsMethods3D Detection ModalityALLVEHPED
mAPmAPHAPAPHAPAPH
OfflineN-PCI [123]VoxelNeXt [9]L36.2932.4345.3544.9327.2319.93
HEDNet [118]L37.7330.7341.8241.3533.6420.1
N-PCI [123] + EMA [116]Focals Conv [8]L + I30.2925.8535.5235.0425.0516.65
LoGoNet [47]L + I37.3031.3944.6444.0629.9618.72
N-PCI [123] + CBM [39]Focals Conv [8]L + I30.2825.8635.5635.0824.9916.64
LoGoNet [47]L + I37.3531.4544.7644.1729.9418.73
OnlineN/AVoxelNeXt [9]L13.9412.9021.4221.236.464.57
HEDNet [118]L13.7511.6818.4818.299.025.07
Focals Conv [8]L + I10.819.6115.7415.555.883.67
LoGoNet [47]L + I13.2911.9119.5019.327.074.49
Ev-3DOD (Ours)L + I + E31.1726.5441.6541.2020.6811.88
", + "image_path": "b768e3601e10ea993b74d5807fb2bfe7af21b507c185309a39aa54a451f17651.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 222, + 292, + 416 + ], + "blocks": [ + { + "bbox": [ + 58, + 222, + 292, + 416 + ], + "lines": [ + { + "bbox": [ + 58, + 222, + 292, + 416 + ], + "spans": [ + { + "bbox": [ + 58, + 222, + 292, + 416 + ], + "type": "image", + "image_path": "d785061edc9d593e3dd21e232b1527bbee327bb8292696d8953f11e990f14ac4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 417, + 295, + 450 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 450 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons of our method with other offline and online evaluations on the DSEC-3DOD dataset. " + }, + { + "bbox": [ + 55, + 417, + 295, + 450 + ], + "type": "inline_equation", + "content": "t = 0.3, 0.6, 0.9" + }, + { + "bbox": [ + 55, + 417, + 295, + 450 + ], + "type": "text", + "content": " denotes the blind times." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 506 + ], + "type": "text", + "content": "ods [8, 47]. For offline image interpolation, we use frame-based, EMA [116], and event-based method, CBM [39]. NeuralPCI [123] is adopted for point cloud interpolation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 651 + ], + "type": "text", + "content": "Ev-Waymo. Table 1 shows the quantitative result of Ev-Waymo. Comparing methods in the online setting, previous methods [8, 9, 47, 118], which rely on fixed frame rate sensors (i.e., LiDAR and Image), cannot detect object motion during blind times. This limitation leads to reduced performance when evaluated with 100 FPS ground truth. In contrast, our approach leverages high-temporal resolution of event data to estimate object motion during blind time, significantly outperforming existing online methods. Notably, Ev-3DOD is comparable to offline oracles and even surpasses certain approaches, such as the combination of NeuralPCI [123] + EMA [116] and Focals Conv [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Figure 6 shows comparisons between the results of online, offline, and our method during blind time. Existing online methods fail to align prediction with the ground truth during blind time. In contrast, Ev-3DOD, despite being an online approach, aligns well with the ground truth and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 350, + 259, + 519, + 317 + ], + "blocks": [ + { + "bbox": [ + 313, + 222, + 555, + 256 + ], + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 256 + ], + "type": "text", + "content": "Table 3. Ablation study of the proposed components. V3D-EF and MCE denote Virtual 3D Event Fusion (Sec. 3.3) and the motion confidence estimator (Sec. 3.4), respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 259, + 519, + 317 + ], + "lines": [ + { + "bbox": [ + 350, + 259, + 519, + 317 + ], + "spans": [ + { + "bbox": [ + 350, + 259, + 519, + 317 + ], + "type": "table", + "html": "
V3D-EFMCEL2
mAPmAPH
33.3031.68
✓46.5544.22
✓✓48.0645.60
", + "image_path": "ca34d303fc19602b3545c1e455ac03566705a207869623df6ba391ed6260a3f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": "demonstrates performance comparable to offline methods. The right graph in Fig. 6 further demonstrates our method's superiority, showing that despite performance degradation with increasing blind time, the gap remains comparable to offline methods, while the online approach exhibits a much larger pixel difference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 519 + ], + "type": "text", + "content": "DSEC-3DOD. As shown in Table 2, we evaluate our model on the DSEC-3DOD real event dataset, which includes fast motion and challenging lighting conditions that significantly degrade performance during blind time. Our model effectively estimates the 3D motion of objects from events, achieving a substantial improvement over other online methods. Notably, it demonstrates performance comparable to offline oracle methods that utilize future information. The analysis of computational complexity is provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 519, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 604 + ], + "type": "text", + "content": "Figure 7 shows blind time detection inference for fast-moving objects in the DSEC dataset. Under rapid motion, the quality of the interpolated pseudo-sensor data may deteriorate, occasionally resulting in detection failures, even when future information is utilized. In contrast, the proposed Ev-3DOD accurately detects objects in blind time, predicting motion from events." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 611, + 471, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 471, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 471, + 624 + ], + "type": "text", + "content": "5.3. Ablation Study and Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": "We present an ablation study on key components and various analyses on the Ev-Waymo dataset. In V3D-EF analyses, the motion confidence estimator (MCE) module is excluded for clear comparisons." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Effectiveness of Each Component. In Table 3, we conduct the ablation study on the key components of Ev-3DOD. We evaluated a baseline and progressively added modules to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27203" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 95, + 261, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 4. Effectiveness of the non-empty mask in the V3D-EF module, as defined in Eq. (1)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 95, + 261, + 144 + ], + "lines": [ + { + "bbox": [ + 94, + 95, + 261, + 144 + ], + "spans": [ + { + "bbox": [ + 94, + 95, + 261, + 144 + ], + "type": "table", + "html": "
MethodL2
mAPmAPH
w/o Non-empty Mask42.5740.50
w/ Non-empty Mask46.5544.22
", + "image_path": "520be6e819998ea0cab577bec34c45b92aff99da68caaf7a658aecfe91469082.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 177, + 288, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 154, + 295, + 176 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 295, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 295, + 176 + ], + "type": "text", + "content": "Table 5. Comparison of results across different modalities and fusion variants of the V3D-EF module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 177, + 288, + 258 + ], + "lines": [ + { + "bbox": [ + 65, + 177, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 65, + 177, + 288, + 258 + ], + "type": "table", + "html": "
MethodL2
mAPmAPH
Base Model33.3031.68
+ LiDAR37.0435.25
+ Event42.2340.17
+ LiDAR + Event (fusion before pooling)43.9541.87
+ LiDAR + Event (fusion after pooling)46.5544.22
", + "image_path": "d212269a88b8173a1d682cf24b06a19b8d893e9cce13606db12614898e70dff1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 271, + 290, + 370 + ], + "blocks": [ + { + "bbox": [ + 60, + 271, + 290, + 370 + ], + "lines": [ + { + "bbox": [ + 60, + 271, + 290, + 370 + ], + "spans": [ + { + "bbox": [ + 60, + 271, + 290, + 370 + ], + "type": "image", + "image_path": "8ae3125fd5d0942b66963a5f76085387bd963d136aaf524c2db6f185e9646496.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 295, + 397 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 397 + ], + "type": "text", + "content": "Figure 8. Detection performance as time elapses since the active timestamp. We compare ours with offline and online approaches." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "text", + "content": "observe improvements. The motion confidence estimator is dependent on V3D-EF, so it was not evaluated separately. V3D-EF, which explicitly estimates motion, has a direct impact on performance, resulting in significant improvements of " + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "inline_equation", + "content": "+13.25\\%" + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "text", + "content": " in mAP and " + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "inline_equation", + "content": "+12.54\\%" + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "text", + "content": " in mAPH. Additionally, the motion confidence estimator refines prediction scores at each blind time, demonstrating its effectiveness with further gains of " + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "inline_equation", + "content": "+1.51\\%" + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "inline_equation", + "content": "+1.38\\%" + }, + { + "bbox": [ + 54, + 411, + 295, + 507 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 509, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 604 + ], + "type": "text", + "content": "Non-empty Mask in V3D-EF. In the proposed V3D-EF, we employ a non-empty mask to achieve accurate and efficient alignment between sparse events and voxel features. To demonstrate the effectiveness of this non-empty mask, we conduct a comparative experiment in which events are projected onto all voxels within the V3D-EF module. The results in Table 4 confirm that the non-empty mask enables effective alignment between the two modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "Variations of the V3D-EF Module. Our core component, the V3D-EF module, provides multiple design options. To analyze these configurations, we developed and evaluated several variants, with the results summarized in Table 5. We experimented with generating the implicit motion field using each modality, i.e., LiDAR and event sensors, and found a performance drop compared to the multi-modal approach. This decline is likely due to the limitations of each modality: LiDAR alone cannot capture motion information during" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 360, + 95, + 511, + 152 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 6. Evaluation results of the model trained with 100 FPS annotations using sliced events at 200 and 300 FPS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 360, + 95, + 511, + 152 + ], + "lines": [ + { + "bbox": [ + 360, + 95, + 511, + 152 + ], + "spans": [ + { + "bbox": [ + 360, + 95, + 511, + 152 + ], + "type": "table", + "html": "
Test SetL2
mAPmAPH
100 FPS48.0645.60
200 FPS47.8745.44
300 FPS47.8145.39
", + "image_path": "dda0b179bad76df24d0a1e645aa931daa0aeb96a2c0bec24b0238cdd7253157b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 163, + 555, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 234 + ], + "type": "text", + "content": "blind times, and event features cannot accurately estimate refined 3D motion. Additionally, multi-modal fusion before pooling decreased performance. We believe this is because fusing before pooling leads to integration between voxels and events outside of box proposal regions, resulting in the use of redundant and possibly misaligned information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 235, + 556, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 556, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 556, + 368 + ], + "type": "text", + "content": "Detection Performance per Elapsed Blind Time. Figure 8 shows the performance over time elapsed since the active timestamp. Due to object motion, the performance of online detection deteriorates rapidly as the time distance from the active timestamp increases. In contrast, offline methods sustain robust performance throughout the blind time by leveraging future data, with only a slight performance drop at the center of the blind time as it moves further from the active timestamps located at both ends. The proposed model, relying solely on event data, achieves a temporal degradation pattern comparable to the offline method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 368, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 556, + 452 + ], + "type": "text", + "content": "Generalization to Higher Frame-rate. To demonstrate the generalization ability to different frame rates, we generate 200 and 300 FPS annotations for arbitrary timestamps by interpolating a pre-constructed 100 FPS dataset. Without additional training, we apply the model trained on 100 FPS. As shown in Table 6, our model effectively infers at arbitrary timestamps with minimal degradation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 464, + 392, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 392, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 392, + 478 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": "This paper proposes Ev-3DOD, a novel method to utilize an event camera in 3D detection for detecting objects in blind time. To transfer information from active times to the current blind time, we estimate 3D motion based on event data. To effectively fuse sparse LiDAR data with events, we propose a Virtual 3D Event Fusion (V3D-EF) and introduce a motion confidence estimator to define confidence for 3D detection during blind times. For this study, we introduce the first event-based 3D detection dataset, emphasizing its 100 FPS annotations. We hope this work showcases the potential of neuromorphic cameras, inspiring future research." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Limitation and Future Work. Our work is the first to address 3D detection during the blind time, establishing it as a pioneering study in this area. However, one limitation is the lack of recurrent processing of past event information to enhance temporal interaction. As a result, discontinuities may occur when fixed frame rate sensors update at active timestamps. We consider improving the interaction between consecutive event data a crucial direction for future work." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27204" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "We thank 42dot for funding this research. We also appreciate the support and valuable discussions from the members of the 42dot research team. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 193, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 193, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 193, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 66, + 193, + 295, + 258 + ], + "type": "text", + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1090-1099, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 259, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 259, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 66, + 259, + 295, + 314 + ], + "type": "text", + "content": "[2] Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Bourtsoulatze, and Yiannis Andreopoulos. Graph-based object classification for neuromorphic vision sensing. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 491-501, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 315, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 315, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 66, + 315, + 294, + 369 + ], + "type": "text", + "content": "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 370, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 370, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 66, + 370, + 294, + 425 + ], + "type": "text", + "content": "[4] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 66, + 426, + 294, + 480 + ], + "type": "text", + "content": "[5] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 172-181, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 66, + 481, + 294, + 514 + ], + "type": "text", + "content": "[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 514, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 514, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 66, + 514, + 294, + 558 + ], + "type": "text", + "content": "[7] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 66, + 559, + 294, + 612 + ], + "type": "text", + "content": "[8] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5428-5437, 2022. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 66, + 614, + 294, + 669 + ], + "type": "text", + "content": "[9] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 670, + 294, + 713 + ], + "type": "text", + "content": "[10] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 887-897, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "type": "text", + "content": "[11] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Deformable feature aggregation for dynamic multi-modal 3d object detection. In European conference on computer vision, pages 628-644. Springer, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "text", + "content": "[12] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 553, + 205 + ], + "type": "text", + "content": "[13] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 320, + 206, + 553, + 249 + ], + "type": "text", + "content": "[14] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In European Conference on Computer Vision, pages 294–314. Springer, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "type": "text", + "content": "[15] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 305, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 305, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 320, + 305, + 553, + 370 + ], + "type": "text", + "content": "[16] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8458-8468, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 371, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 371, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 320, + 371, + 553, + 425 + ], + "type": "text", + "content": "[17] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1761-1769, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 426, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 553, + 491 + ], + "type": "text", + "content": "[18] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44 (1):154-180, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 492, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 492, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 320, + 492, + 553, + 525 + ], + "type": "text", + "content": "[19] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. ArXiv, abs/2211.12324, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 525, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 525, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 320, + 525, + 553, + 557 + ], + "type": "text", + "content": "[20] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629:1034 - 1040, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 558, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 558, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 320, + 558, + 553, + 602 + ], + "type": "text", + "content": "[21] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 603, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 603, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 320, + 603, + 553, + 647 + ], + "type": "text", + "content": "[22] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13884-13893, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "text", + "content": "[23] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27205" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "text", + "content": "[25] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 295, + 218 + ], + "type": "text", + "content": "[26] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3153-3163, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 295, + 274 + ], + "type": "text", + "content": "[27] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22867-22876, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 276, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 296, + 330 + ], + "type": "text", + "content": "[28] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11873-11882, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 332, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 295, + 385 + ], + "type": "text", + "content": "[29] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8417-8427, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 431 + ], + "type": "text", + "content": "[30] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 486 + ], + "type": "text", + "content": "[31] Cheng-Ju Ho, Chen-Hsuan Tai, Yen-Yu Lin, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Diffusion-ss3d: Diffusion model for semi-supervised 3d object detection. Advances in Neural Information Processing Systems, 36:49100-49112, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 489, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 295, + 542 + ], + "type": "text", + "content": "[32] Jordan SK Hu, Tianshu Kuai, and Steven L Waslander. Point density-aware voxels for lidar 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8469-8478, 2022. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 545, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 588 + ], + "type": "text", + "content": "[33] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 634 + ], + "type": "text", + "content": "[34] Junjie Huang, Yun Ye, Zhujin Liang, Yi Shan, and Dalong Du. Detecting as labeling: Rethinking lidar-camera fusion in 3d object detection. In European Conference on Computer Vision, pages 439-455. Springer, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 689 + ], + "type": "text", + "content": "[35] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "text", + "content": "[36] Yu Huang, Chenzhuang Du, Zihui Xue, Xuanyao Chen, Hang Zhao, and Longbo Huang. What makes multi-modal" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 554, + 95 + ], + "type": "text", + "content": "learning better than single (provably). Advances in Neural Information Processing Systems, 34:10944-10956, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 97, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 97, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 97, + 554, + 151 + ], + "type": "text", + "content": "[37] Massimiliano Iacono, Stefan Weber, Arren J. Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 153, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 153, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 320, + 153, + 554, + 207 + ], + "type": "text", + "content": "[38] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformer. In Proceedings of the AAAI conference on Artificial Intelligence, pages 1042-1050, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 554, + 262 + ], + "type": "text", + "content": "[39] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 265, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 265, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 265, + 554, + 308 + ], + "type": "text", + "content": "[40] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 310, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 310, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 321, + 310, + 554, + 363 + ], + "type": "text", + "content": "[41] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 365, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 554, + 420 + ], + "type": "text", + "content": "[42] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 422, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 555, + 476 + ], + "type": "text", + "content": "[43] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 478, + 554, + 521 + ], + "type": "text", + "content": "[44] Dianze Li, Jianing Li, and Yonghong Tian. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:14020-14037, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 321, + 523, + 554, + 567 + ], + "type": "text", + "content": "[45] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 554, + 612 + ], + "type": "text", + "content": "[46] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 614, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 614, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 614, + 554, + 678 + ], + "type": "text", + "content": "[47] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 555, + 713 + ], + "type": "text", + "content": "[48] Xiaotian Li, Baojie Fan, Jiandong Tian, and Huijie Fan. Gafusion: Adaptive fusing lidar and camera with multiple guidance for 3d object detection. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27206" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21209-21218, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "text", + "content": "[49] Xiang Li, Junbo Yin, Wei Li, Chengzhong Xu, Ruigang Yang, and Jianbing Shen. Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3208-3215, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 295, + 218 + ], + "type": "text", + "content": "[50] Yijin Li, Han Zhou, Bangbang Yang, Yexin Zhang, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Graph-based asynchronous event processing for rapid object recognition. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 914-923, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 295, + 263 + ], + "type": "text", + "content": "[51] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems, 35:18442-18455, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 265, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 265, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 265, + 295, + 319 + ], + "type": "text", + "content": "[52] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 321, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 295, + 387 + ], + "type": "text", + "content": "[53] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17182-17191, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 443 + ], + "type": "text", + "content": "[54] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 295, + 488 + ], + "type": "text", + "content": "[55] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7546-7555, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 295, + 533 + ], + "type": "text", + "content": "[56] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 535, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 589 + ], + "type": "text", + "content": "[57] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7345-7353, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 591, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 645 + ], + "type": "text", + "content": "[58] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 712 + ], + "type": "text", + "content": "[59] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 137 + ], + "type": "text", + "content": "[60] Xianpeng Liu, Ce Zheng, Ming Qian, Nan Xue, Chen Chen, Zhebin Zhang, Chen Li, and Tianfu Wu. Multi-view attentive contextualization for multi-view 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16688-16698, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 140, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 553, + 184 + ], + "type": "text", + "content": "[61] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "type": "text", + "content": "[62] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 996-997, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 242, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 242, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 321, + 242, + 553, + 306 + ], + "type": "text", + "content": "[63] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 308, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 308, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 321, + 308, + 553, + 364 + ], + "type": "text", + "content": "[64] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 553, + 420 + ], + "type": "text", + "content": "[65] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2723-2732, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "type": "text", + "content": "[66] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Ji-ashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3164–3173, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "text", + "content": "[67] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. ArXiv, abs/2003.09148, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 512, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 578 + ], + "type": "text", + "content": "[68] Zhenwei Miao, Jikai Chen, Hongyu Pan, Ruiwen Zhang, Kaixuan Liu, Peihan Hao, Jun Zhu, Yang Wang, and Xin Zhan. Pvgnet: A bottom-up one-stage 3d object detector with integrated multi-level features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3279-3288, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 579, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 579, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 321, + 579, + 553, + 632 + ], + "type": "text", + "content": "[69] Su Pang, Daniel Morris, and Hayden Radha. Clocs: Camera-lidar object candidates fusion for 3d object detection. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10386-10393. IEEE, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "type": "text", + "content": "[70] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "type": "text", + "content": "[71] Wanli Peng, Hao Pan, He Liu, and Yi Sun. Ida-3d: Instance-depth-aware 3d object detection from stereo vision" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27207" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "type": "text", + "content": "for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13015-13024, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 151 + ], + "type": "text", + "content": "[72] Yan Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In AAAI Conference on Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "type": "text", + "content": "[73] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6015-6025, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 251 + ], + "type": "text", + "content": "[74] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16794-16804, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 294, + 297 + ], + "type": "text", + "content": "[75] Etienne Perot, Pierre de Tournemire, Davide Oscar Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. *ArXiv*, abs/2009.13436, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 294, + 354 + ], + "type": "text", + "content": "[76] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 355, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 355, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 355, + 294, + 399 + ], + "type": "text", + "content": "[77] AJ Piergiovanni, Vincent Casser, Michael S Ryoo, and Anelia Angelova. 4d-net for learned multi-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15435-15445, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 400, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 294, + 453 + ], + "type": "text", + "content": "[78] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "text", + "content": "[79] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "text", + "content": "[80] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "type": "text", + "content": "[81] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 657 + ], + "type": "text", + "content": "[82] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "type": "text", + "content": "[83] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[84] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 171 + ], + "type": "text", + "content": "[85] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2743-2752, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 173, + 553, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 173, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 553, + 225 + ], + "type": "text", + "content": "[86] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 553, + 282 + ], + "type": "text", + "content": "[87] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 320, + 284, + 553, + 327 + ], + "type": "text", + "content": "[88] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1711-1719, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 328, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 328, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 328, + 553, + 369 + ], + "type": "text", + "content": "[89] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 320, + 372, + 553, + 415 + ], + "type": "text", + "content": "[90] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 417, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 417, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 320, + 417, + 553, + 480 + ], + "type": "text", + "content": "[91] Jiaming Sun, Linghao Chen, Yiming Xie, Siyu Zhang, Qin-hong Jiang, Xiaowei Zhou, and Hujun Bao. Disp r-cnn: Stereo 3d object detection via shape prior guided instance disparity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10548-10557, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 553, + 536 + ], + "type": "text", + "content": "[92] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 538, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 538, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 320, + 538, + 553, + 647 + ], + "type": "text", + "content": "[93] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2443-2451, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "type": "text", + "content": "[94] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27208" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 126 + ], + "type": "text", + "content": "[95] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 294, + 183 + ], + "type": "text", + "content": "[96] Dongsheng Wang, Xu Jia, Yang Zhang, Xinyu Zhang, Yaoyuan Wang, Ziyang Zhang, D. Wang, and Huchuan Lu. Dual memory aggregation network for event-based object detection with learnable representation. In AAAI Conference on Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 184, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 258 + ], + "type": "text", + "content": "[97] Wenjie Wang, Yehao Lu, Guangcong Zheng, Shuigen Zhan, Xiaqing Ye, Zichang Tan, Jingdong Wang, Gaoang Wang, and Xi Li. Bevspread: Spread voxel pooling for bird's-eye-view representation in vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14718-14727, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 261, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 326 + ], + "type": "text", + "content": "[98] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "type": "text", + "content": "[99] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 382, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 382, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 57, + 382, + 294, + 435 + ], + "type": "text", + "content": "[100] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 437, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 294, + 502 + ], + "type": "text", + "content": "[101] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multimodal sparse representations for multi-sensor 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 17545-17556, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 504, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 504, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 57, + 504, + 294, + 547 + ], + "type": "text", + "content": "[102] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 602 + ], + "type": "text", + "content": "[103] Jianyun Xu, Zhenwei Miao, Da Zhang, Hongyu Pan, Kaixuan Liu, Peihan Hao, Jun Zhu, Zhengyang Sun, Hongmin Li, and Xin Zhan. Int: Towards infinite-frames 3d detection with an efficient framework. In European Conference on Computer Vision, pages 193–209. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 604, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 294, + 658 + ], + "type": "text", + "content": "[104] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Zhou Bin, and Liangjun Zhang. Fusionpainting: Multimodal fusion with adaptive attention for 3d object detection. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3047-3054. IEEE, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[105] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[106] Yixin Yang, Jinxiu Liang, Bohan Yu, Yan Chen, Jimmy S Ren, and Boxin Shi. Latency correction for event-guided" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "text", + "content": "deblurring and frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24977-24986, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[107] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "text", + "content": "[108] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 238 + ], + "type": "text", + "content": "[109] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. Advances in Neural Information Processing Systems, 35:1992-2005, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "type": "text", + "content": "[110] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhao-Xu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10201-10210, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[111] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 350, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 404 + ], + "type": "text", + "content": "[112] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In European conference on computer vision, pages 17-33. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 405, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 469 + ], + "type": "text", + "content": "[113] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[114] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "text", + "content": "[115] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kili-ian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. arXiv preprint arXiv:1906.06310, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 571, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 635 + ], + "type": "text", + "content": "[116] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[117] Gang Zhang, Junnan Chen, Guohuan Gao, Jianmin Li, Si Liu, and Xiaolin Hu. Safdnet: A simple and effective network for fully sparse 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14477-14486, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[118] Gang Zhang, Chen Junnan, Guohuan Gao, Jianmin Li, and Xiaolin Hu. Hednet: A hierarchical encoder-decoder net" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27209" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 608 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "work for 3d object detection in point clouds. Advances in Neural Information Processing Systems, 36, 2024. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[119] Jiqing Zhang, B. Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8791-8800, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[120] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9781-9790, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[121] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[122] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "text", + "content": "[123] Zehan Zheng, Danni Wu, Ruisi Lu, Fan Lu, Guang Chen, and Changjun Jiang. Neuralpci: Spatio-temporal neural field for 3d point cloud multi-frame non-linear interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 909–918, 2023. 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "type": "text", + "content": "[124] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1839-1849, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "text", + "content": "[125] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "text", + "content": "[126] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 554, + 294, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 294, + 608 + ], + "type": "text", + "content": "[127] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12800-12810, 2023. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27210" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_content_list.json b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2b5a4242ecb60d896afcc2783209c1626330e42 --- /dev/null +++ b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuoyan Wei $^{1,2}$ Feng Li $^{3,*}$ Shengeng Tang $^{3}$ Yao Zhao $^{1,2}$ Huihui Bai $^{4,1,2,*}$", + "bbox": [ + 192, + 202, + 808, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Science, Beijing Jiaotong University", + "bbox": [ + 256, + 220, + 741, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing Key Laboratory of Advanced Information Science and Network Technology", + "bbox": [ + 163, + 238, + 833, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer Science and Engineering, Hefei University of Technology", + "bbox": [ + 184, + 256, + 812, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Tangshan Research Institute of Beijing Jiaotong University", + "bbox": [ + 263, + 273, + 735, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shuoyan.wei, yzhao, hhbai}@bjtu.edu.cn, {fengli, tangsg}@hfut.edu.cn", + "bbox": [ + 199, + 294, + 797, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continuous space-time video super-resolution (C-STVSR) endeavors to upscale videos simultaneously at arbitrary spatial and temporal scales, which has recently garnered increasing interest. However, prevailing methods struggle to yield satisfactory videos at out-of-distribution spatial and temporal scales. On the other hand, event streams characterized by high temporal resolution and high dynamic range, exhibit compelling promise in vision tasks. This paper presents EvEnhancer, an innovative approach that marries the unique advantages of event streams to elevate effectiveness, efficiency, and generalizability for C-STVSR. Our approach hinges on two pivotal components: 1) Event-adapted synthesis capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories, enabling the adaptive interpolation and fusion of informative spatiotemporal features; 2) Local implicit video transformer integrates local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations utilized to generate plausible videos at arbitrary resolutions and frame rates. Experiments show that EvEnhancer achieves superiority on synthetic and real-world datasets and preferable generalizability on out-of-distribution scales against state-of-the-art methods. Code is available at https://github.com/W-Shuoyan/EvEnhancer.", + "bbox": [ + 86, + 376, + 483, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) is a fundamental low-level vision task, aiming to recover high-resolution (HR) videos from low-resolution (LR) inputs. Nevertheless, video resources are mostly archived with reduced spatial resolution", + "bbox": [ + 89, + 811, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/efa3564cdc2fd103e4f5882e81c7e5a60e155fa4a1eaf20d70c4e9fa238fc66d.jpg", + "image_caption": [ + "Adobe240", + "VideoINR (11.31M)" + ], + "image_footnote": [], + "bbox": [ + 516, + 343, + 712, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c8f16765429f5d249a256317d9d7e04a55a297f4267a652581fca79134589b3.jpg", + "image_caption": [ + "age", + "-Center: reference & center frames\n-Average: all frames" + ], + "image_footnote": [], + "bbox": [ + 599, + 473, + 620, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f629148fd724f95f0bbeae5a269678787b1e6fa57fb4a725ed2ac179215fa274.jpg", + "image_caption": [ + "MoTIF (12.55M)", + "EvEnhancer-light (5.81M)" + ], + "image_footnote": [], + "bbox": [ + 655, + 465, + 669, + 472 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/319189ecb6fc2ee177ec7a9473192ffdbf61c5d264c280d84f05721b7ce36aa6.jpg", + "image_caption": [ + "t: temporal scale factor" + ], + "image_footnote": [], + "bbox": [ + 717, + 343, + 903, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5dadf5fae3067a50d3bd228910ee8088b46192682332860bad8b94b411458e9.jpg", + "image_caption": [ + "(a) Different datasets (In-dist.)" + ], + "image_footnote": [], + "bbox": [ + 516, + 518, + 643, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/778cc400e320508dd656f461854df8a26c9a593a14839b02c021dc0bef0ba73b.jpg", + "image_caption": [ + "Overlapping LR Frames", + "LR Events", + "(c) Visualization comparison", + "Figure 1. Performance comparison of different C-STVSR methods. (a) PSNR (dB) comparison for in-distribution (In-dist.) spatiotemporal upsampling scale (temporal scale $t = 8$ , spatial scale $s = 4$ ) on the different datasets including GoPro [51], Adobe240 [64] and BS-ERGB [69]. (b) PSNR (dB) comparison for different spatiotemporal upsampling scales on GoPro. (c) Visualization comparison for In-dist. scale on GoPro." + ], + "image_footnote": [], + "bbox": [ + 516, + 585, + 643, + 643 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44be04deb02d1728963b276f557fe97c0d4c6570c3a982121b756c9820c9768a.jpg", + "image_caption": [ + "EvEnhancer (6.55M)", + "(b) Different upsampling scales", + "VideoINR[7]" + ], + "image_footnote": [], + "bbox": [ + 650, + 518, + 774, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d391d6829b7b4e017a931cc7043e40e9c81fcde632ed0c4f28072a229f16f93.jpg", + "image_caption": [ + "32.90 dB/0.9337", + "EvEnhancer (Ours)" + ], + "image_footnote": [], + "bbox": [ + 650, + 587, + 774, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/633db7ff7f2780199452f34c4cce262f02a4ef1b33b448a9de60de467e14bc61.jpg", + "image_caption": [ + "$s$ : spatial scale factor", + "MoTIF[6]" + ], + "image_footnote": [], + "bbox": [ + 777, + 518, + 903, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfb82ef11da97bdbc69fa22b73a658d58134ad675821fdc0fce78ee556420959.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 777, + 587, + 903, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and constrained frame rates (i.e., temporal resolution) practically, necessitating the conversion to HR and high-frame-rate (HFR) ones for various applications [40, 73, 76, 80, 85, 89] and user experiences [18, 58, 77].", + "bbox": [ + 511, + 791, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this problem, a naive approach is to cascade separate VSR [2, 3, 14, 39, 41, 75] and video frame interpolation (VFI) [9, 37, 53, 59, 82] or in turn, which over-", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17755", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "looks the correlations between two sub-tasks, leading to suboptimal exploitation of spatiotemporal information in videos. Space-time video super-resolution (STVSR) [15, 15, 19, 20, 33, 70, 71, 79] integrates VSR and VFI into a unified framework that super-resolves the videos with intermediate frame interpolation via one-stage process. Both the above approaches are limited to fixed discrete space and time magnifications, cannot provide arbitrary controls. Some researchers study continuous VFI [1, 24, 29, 88] or continuous VSR [22, 43, 49, 61] methods, where the former is capable of variable frames interpolation while the latter enables arbitrary spatial upsampling including non-integer scales. However, these methods operate in isolation, failing to simultaneously ensure spatial and temporal flexibility.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, research on continuous STVSR (C-STVSR) [6, 7, 50] algorithms has emerged, which typically learn video implicit neural representation (INR) to decode 3D spatiotemporal coordinates into RGB values, thus generating HR while HFR videos arbitrarily. These methods mainly decouple the learning for video INR into spatial INR and temporal INR to predict continuous spatial and temporal feature domains separately. This fails to fully capitalize on the spatiotemporal dependencies in the continuous domain, thus hindering the reconstruction performance, particularly at scales falling outside the training distribution. Besides, existing methods [6, 7] approximate intermediate motion fields between consecutive frames captured from conventional frame-based cameras, which is insufficient to infer accurate inter-frame temporal features when motions are vast or non-linear. Recently, event cameras famous as bio-inspired asynchronous sensors with high temporal resolution and high dynamic range, providing a promising solution to fulfill the inter-frame motion fields in video reconstruction [13, 16, 30, 45, 49, 68, 87]. Nonetheless, the potential of events in C-STVSR remains largely unexplored.", + "bbox": [ + 91, + 305, + 482, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of the above discussion, we propose an innovative event-driven approach EvEnhancer, which adeptly integrates the spatiotemporal synergies between frames and events to learn continuous video representations for effective and generalized C-STVSR. First, we present an event-adapted synthesis module (EASM) that executes event-modulated alignment and bidirectional recurrent compensation to explore the event information for long-term holistic motion trajectory modeling. It incorporates motion cues between input frames with event-based modulation forward and backward to acquire latent inter-frame features. We then recurrently propagate the event stream across time and fuse it with aligned frame features in both directions to maximize the gathering of temporal information.", + "bbox": [ + 89, + 626, + 482, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the decoupling strategies [6, 7, 50] for video INR, we propose a local implicit video transformer (LIVT) that integrates local implicit video neural function with spatiotemporal attention. LIVT calculates cross-scale", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D attention based on the queried space-time coordinates in the continuous domain and the features from EASM over a local sampling grid. In this way, we can directly and efficiently obtain an accurate unified video INR that predicts the RGB values of queried coordinates, enabling to super-resolve videos in arbitrary frames and resolutions. As shown in Figure 1, we conduct experiments with synthetic GoPro [51] and Adobe240 [64] datasets and the real-world BS-ERGB [68] dataset. Experiments demonstrate that EvEnhancer achieves superior effectiveness and efficiency in super-resolving videos in arbitrary spatial and temporal scales within training distribution (In-Dist.) against existing state-of-the-art methods. We also validate its preferable generalizability at out-of-distribution (OOD) scales. The main contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EvEnhancer that subtly marries the unique advantages of event streams with video frames for C-STVSR. Extensive experiments validate its superior performance over recent state-of-the-art methods.", + "- We propose the event-adapted synthesis module (EASM) that enables long-term holistic motion trajectory modeling to acquire informative spatiotemporal features.", + "- We propose the local implicit video transformer (LIVT) that integrates local implicit video neural function with cross-scale spatiotemporal attention to learn accurate unify continuous video INR efficiently." + ], + "bbox": [ + 514, + 316, + 903, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 497, + 653, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Super-Resolution (VSR). Conventional VSR aims to reconstruct HR videos from LR counterparts, which heavily relies on motion estimation and compensation between frames. Earlier methods address this problem by learning frame-level optical flows [39, 72]. Later methods apply deformable convolution [2, 67, 75] or 3D convolution [23, 27] to conduct implicit feature-level motion compensation, exhibiting improved effectiveness and efficiency. Some methods draw inspiration from attention mechanism [28, 74], which incorporate spatial-temporal attention [42, 84, 86, 90] or video transformer [44, 46, 57, 91] to model inter-frame correlations. However, these methods can only upsample at fixed scales.", + "bbox": [ + 511, + 523, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Frame Interpolation (VFI). VFI involves interpolating potential frames between consecutive video frames to enhance the temporal resolution. Flow-based methods [32, 56, 82] leverage optical flow networks [10, 26, 65] to estimate the bidirectional flows between frames. Kernel-based methods [8, 53, 54] utilize local convolution to map interframe motion, which synthesizes target pixels by convolving over input frames. In addition, some recent works focus on splatting-based methods [21, 52, 55] or transformer-based methods [47, 48, 63, 88], showing significant advances. These methods almost learn pixel-level motions for interpolation, which is effective in linear motion but strug", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17756", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gles with complex non-linear motion or occlusions.", + "bbox": [ + 89, + 90, + 431, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Space-Time Video Super-Resolution (STVSR). STVSR can be seen as a combination of VSR and VFI, which reconstruct HR videos spatially and temporally. In literature, most existing methods including traditional [60, 62] and deep learning methods [12, 25, 79] are prone to achieve the STVSR with fixed spatial and temporal scales, denoted by F-STVSR. To mitigate the temporal scale constraints, TMNet [81] proposes controlled frame interpolation with time temporal modulation. SAFA [25] introduces a scale adaptive feature aggregation network that iteratively estimates motion with trainable block-wise scale selection. Though these methods achieve temporal flexibility to some extent, they still fail in arbitrary spatial scales.", + "bbox": [ + 88, + 109, + 480, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based VSR & VFI. Event cameras are characterized by high temporal resolution, high dynamic range, and low latency, promoting many VFI [34, 66, 68, 69, 78] and VSR [13, 30, 49] works. E-VSR [30] introduces the pioneering event-based VSR method that utilizes event-based asynchronous interpolation to synthesize asynchronous neighboring frames. EGVSR [49] unleashes the potential of events in arbitrary spatial VSR, which learns implicit neural representations from both RGB frames and events. EvTexture [31] regards the voxelized events as high-frequency signals to improve the textures in RGB frames, yielding promising VSR reconstruction. As for event-based VFI, TimeLens [68] and TimeLens++ [69] innovatively integrate deformation and synthesis processes. CBMNet [34] utilizes cross-modality information to directly estimate inter-frame motion fields. REFID [66], as a flexible method, exhibits robustness even with blurred reference frames. Notably, these methods achieve upsampling only in the spatial or temporal domain independently, and few studies address a unified approach for both, i.e., event-based STVSR.", + "bbox": [ + 91, + 308, + 483, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Neural Representation (INR) in STVSR. As a method of continuous representation of signals, INR is a common means of continuous super-resolution, including image [4, 5, 17, 38, 83] and video [6, 7, 49, 50] tasks. Inspired by LIIF [5], VideoINR [7] pioneers continuous STVSR (C-STVSR) by learning separate spatial and temporal INRs. Building on this, MoTIF [6] introduces forward spatiotemporal INR to enhance motion field estimation, but it retains the paradigm of decoupling video INR across multiple domains, which lacks robustness across varying scales. Similarly, HR-INR [50], a recent event-based C-STVSR method, also adheres to the same decoupling paradigm as VideoINR [7]. Besides, HR-INR conducts holistic event-frame extraction for long-range motion dependencies and requires repeatedly extracting input frames to realize frame temporal super-resolution, introducing redundancy. Instead, we break the inherent paradigm. We devise the EASM that executes event-modulated alignment and bidirectional recurrent compensation for long-", + "bbox": [ + 91, + 613, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "term motion trajectory modeling. Based on this, our proposed LIVT achieves a unified spatiotemporal INR through the high temporal resolution grids generated from events, thus achieving better generalization with lower weights.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 164, + 687, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given LR and low-frame-rete (LFR) videos and corresponding event streams as inputs, we present an event-driven C-STVSR approach to reconstruct HR and HFR videos at arbitrary spatial and temporal scales. Before elaborating on the components of our methodology, we first provide the details for the event representation in this work.", + "bbox": [ + 511, + 188, + 903, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Event Representation", + "text_level": 1, + "bbox": [ + 511, + 287, + 718, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event cameras asynchronously capture the logarithmic intensity change $\\Delta \\log (I_{x,y,\\tau})$ of each pixel at coordinate $(x,y)$ within the RGB image $I_{x,y,\\tau}$ , where $\\tau$ denote the timestamp of an event stream $\\mathcal{E} = (x,y,\\tau,p)$ . $p$ represents the polarity of $\\mathcal{E}$ triggered at $(x,y)$ , with 1 and -1 representing the change directions respectively. Compared to conventional RGB images, a raw event stream constituted by large amounts of sparse points, which is capable of continuously capturing intensity change with high temporal resolution. However, processing such an unstructured sparse format in deep networks is non-trivial. Following existing event-based method [66], we represent the event stream into a 3D spatiotemporal voxel grid $\\mathcal{V} \\in \\mathbb{R}^{H \\times W \\times (M + 1)}$ linearly, where $H \\times W$ denotes the spatial size, and then discretize the temporal dimension into $M$ event segments $[E_1,E_2,\\dots,E_M]$ . Each segment records the intensity information within a time window of $\\mathcal{E}$ , which is often set to be a small constant (see Figure 2).", + "bbox": [ + 511, + 310, + 906, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 511, + 590, + 624, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{I}^{LR} = I_0^{LR}, I_1^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}$ denote two LR RGB frames and $\\mathcal{E}_{0 \\to 1}$ denote their corresponding event stream. We propose EvEnhancer that addresses the task of super-resolving $\\mathbf{I}^{LR}$ to $(t + 1)$ HR video frames with any spatial scale $s \\geq 1$ and temporal scale $t \\geq 1$ . The overall framework is illustrated in Figure 2, which consists of an event-adapted synthesis module (EASM) and a local implicit video transformer (LIVT).", + "bbox": [ + 511, + 609, + 903, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generally, we first extract initial features on frames and event segments, resulting in RGB frame features $F_0^{LR}$ , $F_1^{LR}$ , and event feature $\\mathbf{F}_{0\\rightarrow 1}^E = \\{F_{0\\rightarrow 1}^{E,m}\\}_{m = 1}^M$ for $M$ segments. Considering that the latent inter-frame features between RGB frames can be captured in both forward and backward directions, we also reverse the time and polarity of $\\nu$ to derive a backward event feature $\\mathbf{F}_{1\\rightarrow 0}^E$ . Then, EASM models long-term holistic motion trajectories with the guidance of events to generate a spatiotemporal feature sequence $\\widehat{\\mathbf{F}}$ with high temporal resolution. Finally, LIVT integrates cross-scale spatiotemporal attention with", + "bbox": [ + 511, + 733, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/259889dbcadf88b12d6f91d908a1c1664478b4c9c83c986fa5d3b4138bd52733.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our EvEnhancer, an event-driven C-STVSR method which consists of an event-adapted synthesis module (EASM), and a local implicit video transformer (LIVT), where EASM contains two steps: (a) event-modulated alignment (EMA), and (b) bidirectional recurrent compensation (BRC). \"CS\": channel squeeze, \"FFN\": feed-forward neural network, \"Res\": residual blocks." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "local implicit function to learn continuous video INR $\\mathbf{F}_{st}$ that decodes arbitrary space-time coordinates into RGB values, thus yielding HR video frames $\\mathbf{I}_{st}^{HR}$ with HFR.", + "bbox": [ + 89, + 501, + 482, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Event-Adapted Synthesis Module", + "text_level": 1, + "bbox": [ + 89, + 559, + 383, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As depicted in Figure 2, EASM contains two steps: 1) Event-modulated alignment incorporates motion cues between input frames with event modulation to acquire latent inter-frame features. Based on the event features and reverse ones, we deploy such alignment in parallel to model the motions forward $(^{+} + )$ and backward $(^{-})$ . 2) Bidirectional recurrent compensation propagates the event stream across time and fuses it with acquired features in both directions to maximize the gathering of temporal information.", + "bbox": [ + 89, + 580, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event-Modulated Alignment (EMA). Given input frame features $F_0^{LR}$ and $F_1^{LR}$ , our goal here is to infer the features of intermediate frames according to $M$ event segment features. Motivated by the effectiveness of the PCD module in video restoration [75, 79], we construct a customized pyramid and event-modulated structure that progressively learns the motion offsets and feature alignment. In the following, we take the forward alignment as an example for clarity.", + "bbox": [ + 89, + 719, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for the forward alignment, as shown in Figure 2(a), we implement a 3-level pyramid by downsampling the frame and event features $(F_0^{LR}, F_1^{LR}, F_{0\\rightarrow 1}^{E,m})$ to 3 different spatial scales $\\times 1$ , $\\times \\frac{1}{2}$ , $\\times \\frac{1}{4}$ via stride convolutions.", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each level $l$ , we first calculate an initial motion vector by a convolutional block (\"Conv+LeakyReLU (LReLU)\") on concatenated frame features, which is then fused with the learned offset $O_{m}^{(l-1)}$ from the $(l-1)$ -th level to obtain a motion vector $MV_{m}^{(l)}$ near the $m$ -th timestamp. Then, we design an event modulation block (EMB) to generate modulated $\\widetilde{MV}_{m}^{(l)}$ . The modulation is controlled by the event feature $F_{0\\rightarrow 1}^{E,m}$ . Compared to previous PCD-based methods [75, 79, 81], our event-based approach can furnish more deterministic motion cues to facilitate frame synthesis, formulated as", + "bbox": [ + 511, + 501, + 906, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {M V} _ {m} ^ {(l)} = \\mathcal {M} ^ {(l)} \\left(M V _ {m} ^ {(l)}, F _ {0 \\rightarrow 1} ^ {E, m (l)}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 686, + 906, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{0\\rightarrow 1}^{E,m(l)}$ denote the downsampled event feature at level $l$ . Then, we can learn the offset $O_{m}^{(l)}$ from $MV_{m}^{(l)}$ and $\\widetilde{MV}_{m}^{(l)}$ . Following [75], we use the deformable convolutional network (DCN) [92] for feature alignment and interpolation. The deformable alignments at different levels are cascaded and aggregated coarse-to-fine for accuracy improvement via the pyramid workflow.", + "bbox": [ + 511, + 722, + 906, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, as there are $M$ segment features $\\{F_{0\\rightarrow 1}^{E,m}\\}_{m = 1}^{M}$ by the event-modulated alignment, we can produce $M$ intermediate frame features $\\mathbf{F}^{+} = [F_{1}^{LR + },F_{2}^{LR + },\\dots ,F_{M}^{LR + }]$ . Similarly, for the backward alignment, we get $\\mathbf{F}^{-} =$", + "bbox": [ + 511, + 838, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17758", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$[F_1^{LR-}, F_2^{LR-}, \\dots, F_M^{LR-}]$ . The two are integrated by channel concatenation with a convolutional block, forming the final features $\\mathbf{F} = \\{F_{\\tau}^{LR}\\}$ with the input features $F_0^{LR}$ , $F_1^{LR}$ , where the timestamp $\\tau = 0, \\frac{1}{M+1}, \\dots, \\frac{M}{M+1}, 1$ is aligned with the input frames and $M$ event segments.", + "bbox": [ + 89, + 88, + 480, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bidirectional Recurrent Compensation (BRC). Given that the information encapsulated within each event segment is confined to a very narrow time window, after the modulated alignment, we propose bidirectional recurrent compensation that iteratively fuses the event and frame features to model the global temporal information inherent in two modalities, as shown in Figure 2(b). In each recurrent iteration, the compensation block receives the event feature $F_{\\tau}^{E}$ from the current timestamp $\\tau$ , the forward hidden state from $h_{\\tau - \\Delta \\tau}^{f}$ the previous timestamp $\\tau - \\Delta \\tau$ , and the backward hidden state $h_{\\tau + \\Delta \\tau}^{b}$ from the next timestamp $\\tau + \\Delta \\tau$ , and the aligned frame feature $F_{\\tau}^{LR} \\in \\mathbf{F}$ . Here, $\\tau \\pm \\Delta \\tau$ denotes the former timestamp and the latter one. Hence, the recurrent fusion process at timestamp $\\tau$ is formulated as", + "bbox": [ + 89, + 167, + 482, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\tau} ^ {b}, h _ {\\tau} ^ {b} = \\mathcal {F} ^ {b} \\left(\\mathcal {C} ^ {b} \\left(F _ {\\tau} ^ {L R}, F _ {\\tau} ^ {E}\\right), h _ {\\tau + \\Delta \\tau} ^ {b}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 391, + 480, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\tau} ^ {f}, h _ {\\tau} ^ {f} = \\mathcal {F} ^ {f} (\\mathcal {C} ^ {f} (F _ {\\tau} ^ {L R}, F _ {\\tau} ^ {E}), h _ {\\tau - \\Delta \\tau} ^ {f}, r _ {\\tau} ^ {b}),\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 412, + 424, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_{\\tau}^{b}$ and $h_{\\tau}^{b}$ denotes the output and hidden states respectively in the backward direction, and $r_{\\tau}^{f}$ and $h_{\\tau}^{f}$ corresponds to the forward. $\\mathcal{C}^f$ and $\\mathcal{C}^b$ calculate the channel attention on event and frame features during the forward and backward recurrent iterations respectively. By recurrent propagating the event information across time, we can comprehensively exploit the high temporal resolution properties from events to refine the frame features and maximize the gathering of temporal information. Finally, a bidirectional feature fusion is run by the element-wise addition between the input feature fed into BRC and its output bidirectional features to retain valid information.", + "bbox": [ + 89, + 444, + 483, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Local Implicit Video Transformer", + "text_level": 1, + "bbox": [ + 89, + 636, + 387, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the EASM, we obtain the discrete temporal feature sequence $\\widehat{\\mathbf{F}}$ . The other key in this method is to learn continuous video INR that specifies the feature at the queried space-time coordinate $(\\mathcal{T},x,y)$ are transferred to any target spatial scale $s$ and temporal scale $t$ in Figure 2. Previous C-STVSR methods [6, 7] learn spatial and temporal INR separately to realize video INR, which fails to fully capitalize on the spatiotemporal dependencies in the continuous domain. In this work, we propose LIVT which calculates cross-scale 3D attention based on the queried space-time coordinates in the continuous domain over a local sampling grid to learn an accurate unified video INR directly.", + "bbox": [ + 89, + 659, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assuming the temporal length and spatial size of the local grid in LR space are $T^G$ and $H^{G} \\times W^{G}$ , respectively. For the full temporal range [0, 1], to generate a frame at the timestamp $\\mathcal{T} \\in [0,1]$ , we only need to select the $T^G$ nearest", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f7f5341d0d6739882ddc280bf9e2a1da01e7f26423948059adadb7647b95ccf.jpg", + "image_caption": [ + "Figure 3. The structure of our local implicit video transformer (LIVT), which integrates the 3D local spatiotemporal attention with implicit neural function to learn continuous video INR to reconstruct HR and HFR video frames." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 906, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "elements to $\\mathcal{T}$ along the time axis. Let the candidate selection of the timestamps $\\tau_{i} = 0,\\frac{1}{M + 1},\\dots,\\frac{M}{M + 1},1$ , which in actual align with the $M$ event segments and input frames $\\mathbf{I}^{LR}$ . We denote the set of candidate timestamps $\\{\\tau_i\\}$ as $\\mathbf{S}_{\\tau}\\subseteq [0,\\frac{1}{M + 1},\\dots,\\frac{M}{M + 1},1]$ , and the size of the $|\\mathbf{S}_{\\tau}| = T^G$ which is selected following the rule as", + "bbox": [ + 511, + 364, + 906, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {\\tau} = \\underset {\\mathbf {S} _ {\\tau}} {\\arg \\min } \\sum_ {\\tau_ {i} \\in \\mathbf {S} _ {\\tau}} | \\tau_ {i} - \\mathcal {T} |. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 465, + 903, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, as shown in Figure 3, with the LR feature sequence $\\widehat{\\mathbf{F}}$ from EASM, we encode them into three embedding spaces through 3D convolution: key $\\{K\\}$ , query $\\{Q\\}$ , and value $\\{V\\}$ . Notably, we do not calculate the spatiotemporal attention across the overall time range. Based on Eq. 3, we conduct temporal selection on $K$ , $Q$ , and $V$ , which allows the learning for 3D local attention in a subtemporal space near the target timestamp $\\mathcal{T}$ . On top of this, we employ 3D trilinear sampling on $Q$ to get the query $q_{\\mathcal{T}} \\in \\mathbb{R}^{s^2 HW \\times 1 \\times C}$ for a certain frame, where $C$ is the channel number. However, since both $k_{\\mathcal{T}}$ and $v_{\\mathcal{T}}$ have size $s^2 HW \\times T^G H^G W^G \\times C$ , it will suffer from expensive computational costs if we directly calculate the attention. To solve this problem, we propose to unfold the sequence along the temporal dimension and calculate cross-scale attention separately on $T^G$ embeddings. Thus, we sample the key $k_{\\tau_i}$ and value $v_{\\tau_i}$ from $K$ and $V$ respectively using 2D nearest local sampling at each time slice $\\tau_i$ . Moreover, we encode and reshape the spatiotemporal relative coordinates $(\\delta \\tau, \\delta x, \\delta y) \\in (-1, 1)$ from each query point to all pixel points within its local grid, obtaining the spatiotemporal bias $b_{\\tau_i} \\in \\mathbb{R}^{s^2 HW \\times H^G W^G \\times 3}$ .", + "bbox": [ + 511, + 505, + 906, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compute the inner product between $q_{\\mathcal{T}}$ and $k_{\\tau_i}$ and add it with the $b_{\\tau_i}$ to obtain the local attention map at the timestamp $\\mathcal{T}$ after a Softmax operation. After the matrix multiplication with $v_{\\tau_i}$ , we can derive the output embedding", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17759", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d308602f6e8dee4661fa1e011eaf1f255c7ad8249bce5d77b339df5b65b13904.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons for in-distribution (In-dist.) spatiotemporal upsampling scales ( $t = 4$ , $f = 8$ ). † indicates the model trained on the same Adobe240 dataset [64] as ours, and ‡ indicates the model fine-tuned on the Adobe240 dataset. **Bold** and **underline** indicate the best and the second-best performance, respectively. Metrics: PSNR (dB) / SSIM." + ], + "table_footnote": [], + "table_body": "
VFI MethodVSR MethodEventsGoPro-CenterGoPro-AverageAdobe240-CenterAdobe240-AverageParams (M)
TimeLens [69]EGVSR [49]✓28.41 / 0.830727.42 / 0.807726.64 / 0.767525.13 / 0.729872.20+2.45
TimeLens [69]EvTexture [31]✓30.50 / 0.878428.51 / 0.847828.80 / 0.833726.12 / 0.778972.20+8.90
CBMNet-L [34]EGVSR [49]✓26.14 / 0.767423.55 / 0.701825.15 / 0.731122.89 / 0.673522.23+2.45
CBMNet-L [34]EvTexture [31]✓27.58 / 0.796623.92 / 0.716326.74 / 0.777723.38 / 0.698622.23+8.90
REFID [66]†EGVSR [49]✓28.74 / 0.836427.97 / 0.817127.28 / 0.784426.55 / 0.763715.91+2.45
REFID [66]†EvTexture [31]✓30.86 / 0.878429.14 / 0.848929.42 / 0.842427.69 / 0.805915.91+8.90
Zooming Slow-Mo [25]†30.69 / 0.8847-30.26 / 0.8821-11.10
TMNet [81]‡30.14 / 0.869228.83 / 0.851429.41 / 0.852428.30 / 0.835412.26
SAFA [25]†31.28 / 0.889430.22 / 0.876130.97 / 0.887830.13 / 0.87824.94
VideoINR-fixed [7]†30.73 / 0.8850-30.21 / 0.8805-11.31
VideoINR [7]†30.26 / 0.879229.41 / 0.866929.92 / 0.874629.27 / 0.865111.31
MoTIF [6]†31.04 / 0.887730.04 / 0.877330.63 / 0.883929.82 / 0.875012.55
HR-INR [50]†✓31.97 / 0.929832.13 / 0.937131.26 / 0.924631.11 / 0.92168.27
EvEnhancer-light✓33.11 / 0.924232.73 / 0.920331.90 / 0.903331.47 / 0.898831.47 / 0.89885.81
EvEnhancer✓33.52 / 0.929533.30 / 0.927932.43 / 0.912932.18 / 0.911632.18 / 0.91166.55
", + "bbox": [ + 109, + 141, + 885, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\widetilde{z}_{\\tau_i}$ . This procedure is described as", + "bbox": [ + 89, + 368, + 323, + 385 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {z} _ {\\tau_ {i}} = \\operatorname {S o f t m a x} \\left(\\frac {q _ {\\mathcal {T}} k _ {\\tau_ {i}} ^ {\\mathrm {T}}}{\\sqrt {C}} + b _ {\\tau_ {i}}\\right) \\otimes v _ {\\tau_ {i}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 391, + 482, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We concatenate the output embeddings $\\widetilde{z}_{\\tau_i}$ along the channel dimension to construct the complete continuous video INR $\\widetilde{Z}_{\\mathcal{T}}$ to yield the features corresponding to continuous space-time coordinates. Following [4, 5, 7], we decode the features into RGB values using a 5-layer MLP with the previous query $q_{\\mathcal{T}}$ and cell decoding to reconstruct the target RGB frame $\\hat{I}_{\\mathcal{T}}^{HR}$ .", + "bbox": [ + 89, + 433, + 483, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 551, + 223, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 575, + 281, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. In our experiments, we train all the models on the Adobe240 dataset [64] which contains 133 video sequences, where 100 for training, 16 for validation, and 17 for testing. We evaluate the performance of our models on both synthetic Adobe240 [64] and GoPro [51] datasets and real-world BS-ERGB [69] and ALPIX-VSR [49] datasets. For the synthetic events, we use the event simulation method vid2e [11] to generate simulated event data. Given a spatial scale $s$ and a temporal scale $t$ , we select consecutive $(t + 1)$ frames as a clip and select the first and the last frames for $s: 1$ bicubic downsampling as the LR and LFR inputs fed into the network to generate HR and HFR frame sequences, and the original $(t + 1)$ frames are groundtruth (GT). Following [31], the event voxels are bicubic-downsampled as the LR event inputs, thus ensuring the spatial size is aligned with the LR frames.", + "bbox": [ + 89, + 598, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Referring VideoINR [7], the training for our model comprises two stages. In the first stage, we set $s = 4$ and $t = 8$ , training for 450K iterations on Adobe240 [64]. In the second stage, we uni-", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formly sample $s$ in [1, 4] and fine-tune for another 150K iterations on the same dataset. We use the Adam optimizer [35] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ , and apply cosine annealing to decay the learning rate from $1 \\times 10^{-4}$ to $1 \\times 10^{-7}$ over 150K iterations. For data augmentation, we randomly crop $32 \\times 32$ spatial regions from downsampled images and event segments with random rotations and horizontal flips. The model is optimized with Charbonnier loss [36] using $\\epsilon^{2} = 1 \\times 10^{-6}$ . In this work, we implement two types of variants, the full model EvEnhancer and a lightweight version EvEnhancer-light. To measure our models, we calculate the PSNR (dB) and SSIM on the Y channel.", + "bbox": [ + 511, + 368, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 560, + 880, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We comprehensively consider frame-based and event-based one-stage or two-stage state-of-the-art methods for comparison: 1) two-stage event-based STVSR methods consisting of event-based VFI (TimeLens [68], REFID [66], CBMNet-L [34]) and event-based VSR (EGVSR [49], EvTexture [31]); 2) one-stage frame-based F-STVSR methods (Zooming Slow-Mo[79], TMNet [81], SAFA [25]); 3) one-stage frame-based C-STVSR methods (VideoINR [7], MoTIF [6]); 4) one-stage event-based C-STVSR methods (HR-INR [50]). For fair comparisons, we use the models of most VFI methods and all STVSR methods trained or fine-tuned on the same Adobe240 dataset [64] as ours. Notably, due to the limitation that the HR-INR source code is not yet available, the results of HR-INR are derived from its original paper.", + "bbox": [ + 511, + 582, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. Table 1 reports the results for the F-STVSR task, where the spatial and temporal scales are the same as the training distribution (In-Dist.). \"Center\" represents the average performance of the reconstructed left, right, and center frames. \"Average\" represents the average performance of all reconstructed frames. Notably,", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/13ac7cce6e9a5740e21fccf22440aef52dc05398fac7304b39b4fdc572e43afd.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons for out-of-distribution (OOD) spatiotemporal upsampling scales on the GoPro dataset [51]." + ], + "table_footnote": [], + "table_body": "
Temporal ScaleSpatial ScaleTMNet [81]‡SAFA [25]†VideoINR [7]†MoTIF [6]†HR-INR [50]†EvEnhancer -lightEvEnhancer
t = 6s = 430.49 / 0.886131.68 / 0.906830.78 / 0.895431.56 / 0.9064-33.10 / 0.926633.41 / 0.9300
t = 6s = 6--25.56 / 0.767129.36 / 0.8505-29.84 / 0.860330.12 / 0.8675
t = 6s = 12--24.02 / 0.690025.81 / 0.7330-25.41 / 0.726125.50 / 0.7323
t = 12s = 426.38 / 0.793127.95 / 0.824927.32 / 0.814127.77 / 0.823028.87 / 0.885431.17 / 0.896932.07 / 0.9116
t = 12s = 6--24.68 / 0.735826.78 / 0.790827.14 / 0.817328.94 / 0.842529.54 / 0.8579
t = 12s = 12--23.70 / 0.683024.72 / 0.7108-25.25 / 0.724825.45 / 0.7345
t = 16s = 424.72 / 0.752626.38 / 0.783225.81 / 0.773925.98 / 0.775827.29 / 0.855630.15 / 0.876230.93 / 0.8918
t = 16s = 6--23.86 / 0.712325.34 / 0.752726.09 / 0.795428.34 / 0.828328.86 / 0.8434
t = 16s = 12--22.88 / 0.665923.88 / 0.6923-25.10 / 0.721725.25 / 0.7310
t = 6s = 1--32.34 / 0.954534.77 / 0.969638.53 / 0.973537.68 / 0.966738.80 / 0.9714
", + "bbox": [ + 94, + 114, + 901, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bfc3249321239f23d882d22ec458660b909be80fe18a02543c09a795c052f9e.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons for different temporal upsampling scales $t$ (fixed $s = 4$ ) on the BS-ERGB dataset [69]." + ], + "table_footnote": [], + "table_body": "
MethodEventst=4t=6t=8
TimeLens[69]+EGVSR[49]✓22.98/0.676222.01/0.659221.33/0.6467
TimeLens[69]+EvTexture[31]✓23.58/0.708822.48/0.688721.74/0.6746
CBMNet-L[34]+EGVSR[49]✓22.74/0.685621.93/0.670521.29/0.6657
CBMNet-L[34]+EvTexture[31]✓23.40/0.719922.47/0.702021.78/0.6962
REFID[66]†+EGVSR[49]✓--22.76/0.6478
REFID[66]†+EvTexture[31]✓--22.77/0.6527
SAFA[25]†24.32/0.731423.23/0.712022.41/0.6954
VideoINR[7]†24.06/0.729023.00/0.709822.23/0.6945
MoTIF[6]†24.21/0.729623.15/0.709922.37/0.6939
EvEnhancer-light✓25.32/0.731124.54/0.718424.16/0.7100
EvEnhancer✓25.44/0.733824.70/0.721524.29/0.7138
", + "bbox": [ + 91, + 323, + 480, + 467 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d22523dc1eff32991136bdf40917f9eef16db26a60098910dfcec396fb70d6a.jpg", + "table_caption": [ + "Table 4. $2\\times$ VSR performance on the ALPIX-VSR dataset [49]." + ], + "table_footnote": [], + "table_body": "
MethodVideoINR [7]MoTIF [6]EGVSR [49]EvEnhancer
PSNR / SSIM32.53 / 0.938338.61 / 0.963639.31 / 0.963540.84 / 0.9786
", + "bbox": [ + 93, + 506, + 478, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VideoINR-fixed is trained specifically for single-frame interpolation. We can see that both our EvEnhancer and EvEnhancer-light significantly outperform existing methods in all settings with much fewer model parameters, which achieves more than 1 dB PSNR improvements. Then, we also evaluate our models on the C-STVSR task, where spatial and temporal scales are out of the training distribution. The results are illustrated in Table 2. It can be observed that both our models exceed almost all existing methods at various spatial and temporal scales, especially in the challenging large temporal scale, achieving superior effectiveness and generalization ability. In particular, our EvEnhancer-light and EvEnhancer use lighter-weight models and perform better than HR-INR [50] at most temporal and spatial scales, especially at OOD scales.", + "bbox": [ + 88, + 551, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further validate the effectiveness of our method on the real-world BS-ERGB dataset under the setting of fixed spatial scale $s = 4$ but varying temporal scales $t$ . The comparisons in Table 3 demonstrate that our method yields the best quantitative performance under all conditions. Compared to cascaded event-based VFI and VSR, our EvEnhancer simultaneously learns temporal interpolation and spatial VSR, thus achieving superior performance. Due to", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ed0ea1f2701b710c782415af7a06e615ae5c7cac0b98117cacf0939d28345fca.jpg", + "table_caption": [ + "Table 5. TFLOPs comparisons of C-STVSR models at different upsampling temporal scales $t$ (fixed $s = 4$ ). * indicates the repeated iterations during inference on the NVIDIA GeForce RTX 3090 GPU due to their large computations." + ], + "table_footnote": [], + "table_body": "
Temporal ScaleVideoINR [7] (11.31M)MoTIF [6] (12.55M)EvEnhancer -light (5.81M)EvEnhancer (6.55M)
t = 22.0112.0431.8763.398
t = 42.3954.086*2.4724.647
t = 62.7796.129*3.0675.885
t = 83.1636.129*3.6637.129
t = 123.93210.22*4.8549.617
t = 164.70012.26*6.04412.10
", + "bbox": [ + 516, + 349, + 901, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the high temporal resolution of events, equipped by EASM, our EvEnhancer enables holistic motion trajectory modeling, thus outperforming frame-based methods. Benefiting from the cross-scale attention in LIVT, it also performs preferably at OOD scales. To validate the effectiveness of our EvEnhancer for real-world LR events, we conduct experiments on the real-world ALPIX-VSR dataset [49] (image resolution: $3264 \\times 2448$ , event resolution: $1632 \\times 1224$ ) for $2 \\times$ VSR. As shown in Table 4, EvEnhancer significantly surpasses all methods without additional fine-tuning.", + "bbox": [ + 509, + 470, + 906, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Complexity. To validate the model efficiency of the proposed methods, besides the model parameter comparison in Table 1, we compare the TFLOPs with existing C-STVSR methods in Table 5, which are calculated on the upsampling of a clip with a resolution of $180 \\times 320$ by a spital upsampling scale $s = 4$ at different temporal upsampling scales $t$ . VideoINR shows the fewest TFLOPs but more parameters and the worst performance (see Table 1). At the low temporal scale ( $\\leq 4$ ), MoTIF is computationally efficient than ours. However, as the temporal scale increases, its computations increase and exceed both our models.", + "bbox": [ + 509, + 625, + 908, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. With the qualitative results in Figure 4, compared to VideoINR [7] and MoTIF [6], our models reconstruct the faithful HR frames with clearer textures. By computing the difference maps between generated frames and corresponding ground truth frames, we can see that our methods are more excellent at tackling non-linear motions and recovering more details.", + "bbox": [ + 509, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f65700759f696021fdab42b743e8e4cd5bda3f73b27412d45eebaad6c7a14308.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison for In-dist. scale $(t = 8, s = 4)$ on the GoPro dataset [51]. We compare the normalized absolute difference maps (yellow boxes) for the same regions in each frame as in the GT frames." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 893, + 371 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8ab95dd8c19e9d28631ea008ed9592a6ccd097239be30998decceaa0e3554a11.jpg", + "table_caption": [ + "Table 6. Ablations on the designs of event-adapted synthesis module (EASM). \"fwd.\" and \"bwd.\" denote the forward and backward directions, respectively. Metrics: PSNR (dB) / SSIM." + ], + "table_footnote": [], + "table_body": "
EMABRCBFIn-Dist.OOD
fwd. & bwd.32.50 / 0.916628.86 / 0.8407
fwd. & bwd.✓33.13 / 0.925229.42 / 0.8539
fwd.✓✓33.20 / 0.926229.47 / 0.8556
fwd. & bwd.✓✓33.30 / 0.927929.54 / 0.8579
", + "bbox": [ + 129, + 470, + 444, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 558, + 256, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies by keeping the same training strategy on the Adobe240 dataset [64], and testing on the GoPro dataset [51] at both In-Dist. $(t = 8, s = 4)$ and OOD $(t = 12, s = 6)$ spatiotemporal scales to validate the effectiveness of our methodology design.", + "bbox": [ + 89, + 580, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Event-Adapted Synthesis Module. Table 6 presents the ablations of EASM, which involves the event-modulated alignment (EMA), bidirectional recurrent compensation (BRC), and bidirectional feature fusion (BF). The model with only EMA achieves the worst performance. When we introduce BRC in the model, one can see the performance improves. Besides, with the integration BRC, BF, and EMA, the model achieves the best at both In-Dist. and OOD scales. We can also observe that the model with only a single forward alignment can suffer from performance degradation. The results demonstrate the effects of EASM.", + "bbox": [ + 89, + 657, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Local Implicit Video Transformer. Table 7 investigates the effect of LIVT. We first implement the model with a 2D decoupling strategy as videoINR [7] but introduce the local attention, concatenate cell decoding, and query, one can see that the model achieves better performance than the model", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/59fc303cf78f1fb5ddb32ea3f71574532c9cb043c8b2e188283d80480485ad64.jpg", + "table_caption": [ + "Table 7. Ablations on the designs of local implicit video transformer (LIVT). \"LA\": local attention, \"CE\": concatenate cell decoding, \"PQ\": concatenate the previous query. Metrics: PSNR (dB) / SSIM." + ], + "table_footnote": [], + "table_body": "
INR ModeLACEPQIn-dist.OOD
2D decoupling✓✓✓33.26 / 0.927229.26 / 0.8525
3D unification27.32 / 0.797525.03 / 0.7032
3D unification✓33.27 / 0.927629.47 / 0.8565
3D unification✓✓33.30 / 0.928229.49 / 0.8570
3D unification✓✓✓33.30 / 0.927929.54 / 0.8579
", + "bbox": [ + 537, + 484, + 879, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "without any components. When we replace the 2D decoupling with our unified 3D video INR, the model produces results with higher PSNR and SSIM, especially at OOD scales. In addition, the concatenation of the cell [5] and previous query [7] brings a larger gain at OOD scales than In-Dist., which validates their effects.", + "bbox": [ + 511, + 585, + 906, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce EvEnhancer, a novel approach that amalgamates the unique advantages of event streams to enhance the effectiveness, efficiency, and generalizability of C-STVSR. Our EvEnhancer is underpinned by two critical components: 1) the EASM capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories; and 2) the LIVT, which integrates a local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations. Experiments indicate that EvEnhancer significantly surpasses current state-of-the-art methods.", + "bbox": [ + 511, + 719, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by the National Natural Science Foundation of China (62302141, 62331003, 62120106009), the Joint Funds of the National Natural Science Foundation of China (U23A20314), the Fundamental Research Funds for the Central Universities (JZ2024HGTB0255), Beijing Natural Science Foundation (L223022), and the Natural Science Foundation of Hebei Province (F2024105029).", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 247, + 187, + 263 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2", + "[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2", + "[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1", + "[4] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 3, 6", + "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 6, 8", + "[6] Yi-Hsin Chen, Si-Cun Chen, Yen-Yu Lin, and Wen-Hsiao Peng. Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video superresolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23131-23141, 2023. 1, 2, 3, 5, 6, 7, 8", + "[7] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2047-2057, 2022. 1, 2, 3, 5, 6, 7, 8", + "[8] Xianhang Cheng and Zhenzhong Chen. Video frame interpolation via deformable separable convolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10607-10614, 2020. 2", + "[9] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at" + ], + "bbox": [ + 99, + 272, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 1", + "[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 2", + "[11] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In CVPR, pages 3586-3595, 2020. 6", + "[12] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 3", + "[13] Jin Han, Yixin Yang, Chu Zhou, Chao Xu, and Boxin Shi. Evintsr-net: Event guided multiple latent frames reconstruction and super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4882-4891, 2021. 2, 3", + "[14] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3897-3906, 2019. 1", + "[15] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Space-time-aware multi-resolution video enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2859-2868, 2020. 2", + "[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 2", + "[17] Zongyao He and Zhi Jin. Latent modulated function for computational optimal continuous image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26026-26035, 2024. 3", + "[18] Qiqi Hou, Abhijay Ghildyal, and Feng Liu. A perceptual quality metric for video frame interpolation. In European Conference on Computer Vision, pages 234-253. Springer, 2022. 1", + "[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 2", + "[20] Mengshun Hu, Kui Jiang, Zheng Wang, Xiang Bai, and Ruimin Hu. Cycmunet+: Cycle-projected mutual learning" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for spatial-temporal video super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[21] Ping Hu, Simon Niklaus, Lu Zhang, Stan Sclaroff, and Kate Saenko. Video frame interpolation with many-to-many splating and spatial selective refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[22] Cong Huang, Jiahao Li, Lei Chu, Dong Liu, and Yan Lu. Arbitrary-scale video super-resolution guided by dynamic context. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2302, 2024. 2", + "[23] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 40(4):1015-1028, 2017. 2", + "[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2", + "[25] Zhewei Huang, Ailin Huang, Xiaotao Hu, Chen Hu, Jun Xu, and Shuchang Zhou. Scale-adaptive feature aggregation for efficient space-time video super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4228-4239, 2024. 3, 6, 7", + "[26] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 2", + "[27] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 2", + "[28] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. arXiv preprint arXiv:2008.05765, 2020. 2", + "[29] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2", + "[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2, 3", + "[31] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. EvTexture: Event-driven Texture Enhancement for Video Super-Resolution. In Proceedings of the 41st International Conference on Machine Learning, pages 22817–22839. PMLR, 2024. 3, 6, 7", + "[32] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 2" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Soo Ye Kim, Jihyong Oh, and Munchurl Kim. Fisr: Deep joint frame interpolation and super-resolution with a multiscale temporal loss. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11278-11286, 2020. 2", + "[34] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 3, 6, 7", + "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[36] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 6", + "[37] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5316-5325, 2020. 1", + "[38] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1929-1938, 2022. 3", + "[39] Feng Li, Huihui Bai, and Yao Zhao. Learning a deep dual attention network for video super-resolution. IEEE transactions on image processing, 29:4474-4488, 2020. 1, 2", + "[40] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 1", + "[41] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10522-10531, 2019. 1", + "[42] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 2", + "[43] Zekun Li, Hongying Liu, Fanhua Shang, Yuanyuan Liu, Liang Wan, and Wei Feng. Savsr: Arbitrary-scale video super-resolution via a learned scale-adaptive network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3288-3296, 2024. 2", + "[44] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2", + "[45] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17764", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning event-driven video deblurring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 695-710. Springer, 2020. 2", + "[46] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2", + "[47] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19125-19134, 2024. 2", + "[48] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 2", + "[49] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1557-1567, 2023. 2, 3, 6, 7", + "[50] Yunfan Lu, Zipeng Wang, Yusheng Wang, and Hui Xiong. Hr-inr: Continuous space-time video super-resolution via event camera. arXiv preprint arXiv:2405.13389, 2024. 2, 3, 6, 7", + "[51] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 6, 7, 8", + "[52] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 2", + "[53] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE international conference on computer vision, pages 261-270, 2017. 1, 2", + "[54] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive convolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 670-679, 2017. 2", + "[55] Simon Niklaus, Ping Hu, and Jiawen Chen. Splitting-based synthesis for video frame interpolation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 713-723, 2023. 2", + "[56] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14539-14548, 2021. 2", + "[57] Zhongwei Qiu, Huan Yang, Jianlong Fu, Daochang Liu, Chang Xu, and Dongmei Fu. Learning degradation-robust spatiotemporal frequency-transformer for video superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Nasrin Rahimi and A Murat Tekalp. Spatio-temporal perception-distortion trade-off in learned video sr. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1400-1404. IEEE, 2023. 1", + "[59] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF international conference on computer vision, pages 892-900, 2019. 1", + "[60] Oded Shahar, Alon Faktor, and Michal Irani. Space-time super-resolution from a single video. IEEE, 2011. 3", + "[61] Wei Shang, Dongwei Ren, Wanying Zhang, Yuming Fang, Wangmeng Zuo, and Kede Ma. Arbitrary-scale video superresolution with structural and textural priors. In European Conference on Computer Vision, pages 73-90. Springer, 2025. 2", + "[62] Eli Shechtman, Yaron Caspi, and Michal Irani. Space-time super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(4):531-545, 2005. 3", + "[63] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 2", + "[64] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 1, 2, 6, 8", + "[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2", + "[66] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 3, 6, 7", + "[67] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2", + "[68] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 2, 3, 6", + "[69] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1, 3, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17765", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Hai Wang, Wenming Yang, Qingmin Liao, and Jie Zhou. Bi- rstu: Bidirectional recurrent upsampling network for spacetime video super-resolution. IEEE Transactions on Multimedia, 2022. 2", + "[71] Hai Wang, Xiaoyu Xiang, Yapeng Tian, Wenming Yang, and Qingmin Liao. Stdan: deformable attention network for space-time video super-resolution. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through hr optical flow estimation. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 514–529. Springer, 2019. 2", + "[73] Li Wang, Dong Li, Yousong Zhu, Lu Tian, and Yi Shan. Dual super-resolution learning for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3774-3783, 2020. 1", + "[74] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2", + "[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1, 2, 4", + "[76] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 1", + "[77] Guangyang Wu, Xin Tao, Changlin Li, Wenyi Wang, Xiao-hong Liu, and Qingqing Zheng. Perception-oriented video frame interpolation via asymmetric blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2753-2762, 2024. 1", + "[78] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In European Conference on Computer Vision, pages 267-283. Springer, 2022. 3", + "[79] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3370-3379, 2020. 2, 3, 4, 6", + "[80] Chenming Xu, Meiqin Liu, Chao Yao, Weisi Lin, and Yao Zhao. Ibvc: Interpolation-driven b-frame video compression. Pattern Recognition, 153:110465, 2024. 1", + "[81] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6388-6397, 2021. 3, 4, 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2", + "[83] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 3", + "[84] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 2", + "[85] Jun-Sang Yoo, Hongjae Lee, and Seung-Won Jung. Video object segmentation-aware video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12322-12333, 2023. 1", + "[86] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17834-17843, 2022. 2", + "[87] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S Ren. Training weakly supervised video frame interpolation with events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14589-14598, 2021. 2", + "[88] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 2", + "[89] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8799-8808, 2019. 1", + "[90] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 2", + "[91] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 2", + "[92] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 4" + ], + "bbox": [ + 516, + 90, + 903, + 782 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17766", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_model.json b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5939e62d2de1d809f127fcf9f917970c6e180427 --- /dev/null +++ b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_model.json @@ -0,0 +1,2798 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.203, + 0.809, + 0.222 + ], + "angle": 0, + "content": "Shuoyan Wei\\(^{1,2}\\) Feng Li\\(^{3,*}\\) Shengeng Tang\\(^{3}\\) Yao Zhao\\(^{1,2}\\) Huihui Bai\\(^{4,1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.222, + 0.743, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Information Science, Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.239, + 0.834, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Beijing Key Laboratory of Advanced Information Science and Network Technology" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.257, + 0.813, + 0.276 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Computer Science and Engineering, Hefei University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.275, + 0.736, + 0.294 + ], + "angle": 0, + "content": "\\(^{4}\\)Tangshan Research Institute of Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.295, + 0.799, + 0.31 + ], + "angle": 0, + "content": "{shuoyan.wei, yzhao, hhbai}@bjtu.edu.cn, {fengli, tangsg}@hfut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.377, + 0.485, + 0.754 + ], + "angle": 0, + "content": "Continuous space-time video super-resolution (C-STVSR) endeavors to upscale videos simultaneously at arbitrary spatial and temporal scales, which has recently garnered increasing interest. However, prevailing methods struggle to yield satisfactory videos at out-of-distribution spatial and temporal scales. On the other hand, event streams characterized by high temporal resolution and high dynamic range, exhibit compelling promise in vision tasks. This paper presents EvEnhancer, an innovative approach that marries the unique advantages of event streams to elevate effectiveness, efficiency, and generalizability for C-STVSR. Our approach hinges on two pivotal components: 1) Event-adapted synthesis capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories, enabling the adaptive interpolation and fusion of informative spatiotemporal features; 2) Local implicit video transformer integrates local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations utilized to generate plausible videos at arbitrary resolutions and frame rates. Experiments show that EvEnhancer achieves superiority on synthetic and real-world datasets and preferable generalizability on out-of-distribution scales against state-of-the-art methods. Code is available at https://github.com/W-Shuoyan/EvEnhancer." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Video super-resolution (VSR) is a fundamental low-level vision task, aiming to recover high-resolution (HR) videos from low-resolution (LR) inputs. Nevertheless, video resources are mostly archived with reduced spatial resolution" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.344, + 0.714, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.467, + 0.635, + 0.474 + ], + "angle": 0, + "content": "Adobe240" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.474, + 0.621, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.474, + 0.635, + 0.482 + ], + "angle": 0, + "content": "age" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.486, + 0.647, + 0.503 + ], + "angle": 0, + "content": "-Center: reference & center frames\n-Average: all frames" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.466, + 0.671, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.467, + 0.746, + 0.474 + ], + "angle": 0, + "content": "VideoINR (11.31M)" + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.474, + 0.735, + 0.484 + ], + "angle": 0, + "content": "MoTIF (12.55M)" + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.486, + 0.773, + 0.494 + ], + "angle": 0, + "content": "EvEnhancer-light (5.81M)" + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.494, + 0.753, + 0.503 + ], + "angle": 0, + "content": "EvEnhancer (6.55M)" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.344, + 0.905, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.486, + 0.902, + 0.494 + ], + "angle": 0, + "content": "t: temporal scale factor" + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.494, + 0.902, + 0.503 + ], + "angle": 0, + "content": "\\(s\\) : spatial scale factor" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.507, + 0.697, + 0.518 + ], + "angle": 0, + "content": "(a) Different datasets (In-dist.)" + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.507, + 0.885, + 0.518 + ], + "angle": 0, + "content": "(b) Different upsampling scales" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.52, + 0.644, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.579, + 0.636, + 0.588 + ], + "angle": 0, + "content": "Overlapping LR Frames" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.587, + 0.644, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.648, + 0.605, + 0.657 + ], + "angle": 0, + "content": "LR Events" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.52, + 0.776, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.579, + 0.743, + 0.588 + ], + "angle": 0, + "content": "VideoINR[7]" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.588, + 0.776, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.636, + 0.752, + 0.644 + ], + "angle": 0, + "content": "32.90 dB/0.9337" + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.648, + 0.762, + 0.658 + ], + "angle": 0, + "content": "EvEnhancer (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.66, + 0.786, + 0.671 + ], + "angle": 0, + "content": "(c) Visualization comparison" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.52, + 0.905, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.579, + 0.865, + 0.588 + ], + "angle": 0, + "content": "MoTIF[6]" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.588, + 0.904, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.833, + 0.648, + 0.85, + 0.656 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.677, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Figure 1. Performance comparison of different C-STVSR methods. (a) PSNR (dB) comparison for in-distribution (In-dist.) spatiotemporal upsampling scale (temporal scale \\( t = 8 \\), spatial scale \\( s = 4 \\)) on the different datasets including GoPro [51], Adobe240 [64] and BS-ERGB [69]. (b) PSNR (dB) comparison for different spatiotemporal upsampling scales on GoPro. (c) Visualization comparison for In-dist. scale on GoPro." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.907, + 0.854 + ], + "angle": 0, + "content": "and constrained frame rates (i.e., temporal resolution) practically, necessitating the conversion to HR and high-frame-rate (HFR) ones for various applications [40, 73, 76, 80, 85, 89] and user experiences [18, 58, 77]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To address this problem, a naive approach is to cascade separate VSR [2, 3, 14, 39, 41, 75] and video frame interpolation (VFI) [9, 37, 53, 59, 82] or in turn, which over-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "looks the correlations between two sub-tasks, leading to suboptimal exploitation of spatiotemporal information in videos. Space-time video super-resolution (STVSR) [15, 15, 19, 20, 33, 70, 71, 79] integrates VSR and VFI into a unified framework that super-resolves the videos with intermediate frame interpolation via one-stage process. Both the above approaches are limited to fixed discrete space and time magnifications, cannot provide arbitrary controls. Some researchers study continuous VFI [1, 24, 29, 88] or continuous VSR [22, 43, 49, 61] methods, where the former is capable of variable frames interpolation while the latter enables arbitrary spatial upsampling including non-integer scales. However, these methods operate in isolation, failing to simultaneously ensure spatial and temporal flexibility." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.306, + 0.483, + 0.624 + ], + "angle": 0, + "content": "To achieve this, research on continuous STVSR (C-STVSR) [6, 7, 50] algorithms has emerged, which typically learn video implicit neural representation (INR) to decode 3D spatiotemporal coordinates into RGB values, thus generating HR while HFR videos arbitrarily. These methods mainly decouple the learning for video INR into spatial INR and temporal INR to predict continuous spatial and temporal feature domains separately. This fails to fully capitalize on the spatiotemporal dependencies in the continuous domain, thus hindering the reconstruction performance, particularly at scales falling outside the training distribution. Besides, existing methods [6, 7] approximate intermediate motion fields between consecutive frames captured from conventional frame-based cameras, which is insufficient to infer accurate inter-frame temporal features when motions are vast or non-linear. Recently, event cameras famous as bio-inspired asynchronous sensors with high temporal resolution and high dynamic range, providing a promising solution to fulfill the inter-frame motion fields in video reconstruction [13, 16, 30, 45, 49, 68, 87]. Nonetheless, the potential of events in C-STVSR remains largely unexplored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.838 + ], + "angle": 0, + "content": "In light of the above discussion, we propose an innovative event-driven approach EvEnhancer, which adeptly integrates the spatiotemporal synergies between frames and events to learn continuous video representations for effective and generalized C-STVSR. First, we present an event-adapted synthesis module (EASM) that executes event-modulated alignment and bidirectional recurrent compensation to explore the event information for long-term holistic motion trajectory modeling. It incorporates motion cues between input frames with event-based modulation forward and backward to acquire latent inter-frame features. We then recurrently propagate the event stream across time and fuse it with aligned frame features in both directions to maximize the gathering of temporal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In contrast to the decoupling strategies [6, 7, 50] for video INR, we propose a local implicit video transformer (LIVT) that integrates local implicit video neural function with spatiotemporal attention. LIVT calculates cross-scale" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.317 + ], + "angle": 0, + "content": "3D attention based on the queried space-time coordinates in the continuous domain and the features from EASM over a local sampling grid. In this way, we can directly and efficiently obtain an accurate unified video INR that predicts the RGB values of queried coordinates, enabling to super-resolve videos in arbitrary frames and resolutions. As shown in Figure 1, we conduct experiments with synthetic GoPro [51] and Adobe240 [64] datasets and the real-world BS-ERGB [68] dataset. Experiments demonstrate that EvEnhancer achieves superior effectiveness and efficiency in super-resolving videos in arbitrary spatial and temporal scales within training distribution (In-Dist.) against existing state-of-the-art methods. We also validate its preferable generalizability at out-of-distribution (OOD) scales. The main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.318, + 0.905, + 0.378 + ], + "angle": 0, + "content": "- We propose EvEnhancer that subtly marries the unique advantages of event streams with video frames for C-STVSR. Extensive experiments validate its superior performance over recent state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.379, + 0.905, + 0.424 + ], + "angle": 0, + "content": "- We propose the event-adapted synthesis module (EASM) that enables long-term holistic motion trajectory modeling to acquire informative spatiotemporal features." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.424, + 0.905, + 0.484 + ], + "angle": 0, + "content": "- We propose the local implicit video transformer (LIVT) that integrates local implicit video neural function with cross-scale spatiotemporal attention to learn accurate unify continuous video INR efficiently." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.318, + 0.905, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.655, + 0.513 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.719 + ], + "angle": 0, + "content": "Video Super-Resolution (VSR). Conventional VSR aims to reconstruct HR videos from LR counterparts, which heavily relies on motion estimation and compensation between frames. Earlier methods address this problem by learning frame-level optical flows [39, 72]. Later methods apply deformable convolution [2, 67, 75] or 3D convolution [23, 27] to conduct implicit feature-level motion compensation, exhibiting improved effectiveness and efficiency. Some methods draw inspiration from attention mechanism [28, 74], which incorporate spatial-temporal attention [42, 84, 86, 90] or video transformer [44, 46, 57, 91] to model inter-frame correlations. However, these methods can only upsample at fixed scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Video Frame Interpolation (VFI). VFI involves interpolating potential frames between consecutive video frames to enhance the temporal resolution. Flow-based methods [32, 56, 82] leverage optical flow networks [10, 26, 65] to estimate the bidirectional flows between frames. Kernel-based methods [8, 53, 54] utilize local convolution to map interframe motion, which synthesizes target pixels by convolving over input frames. In addition, some recent works focus on splatting-based methods [21, 52, 55] or transformer-based methods [47, 48, 63, 88], showing significant advances. These methods almost learn pixel-level motions for interpolation, which is effective in linear motion but strug" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17756" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.432, + 0.106 + ], + "angle": 0, + "content": "gles with complex non-linear motion or occlusions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.11, + 0.482, + 0.306 + ], + "angle": 0, + "content": "Space-Time Video Super-Resolution (STVSR). STVSR can be seen as a combination of VSR and VFI, which reconstruct HR videos spatially and temporally. In literature, most existing methods including traditional [60, 62] and deep learning methods [12, 25, 79] are prone to achieve the STVSR with fixed spatial and temporal scales, denoted by F-STVSR. To mitigate the temporal scale constraints, TMNet [81] proposes controlled frame interpolation with time temporal modulation. SAFA [25] introduces a scale adaptive feature aggregation network that iteratively estimates motion with trainable block-wise scale selection. Though these methods achieve temporal flexibility to some extent, they still fail in arbitrary spatial scales." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.309, + 0.485, + 0.612 + ], + "angle": 0, + "content": "Event-based VSR & VFI. Event cameras are characterized by high temporal resolution, high dynamic range, and low latency, promoting many VFI [34, 66, 68, 69, 78] and VSR [13, 30, 49] works. E-VSR [30] introduces the pioneering event-based VSR method that utilizes event-based asynchronous interpolation to synthesize asynchronous neighboring frames. EGVSR [49] unleashes the potential of events in arbitrary spatial VSR, which learns implicit neural representations from both RGB frames and events. EvTexture [31] regards the voxelized events as high-frequency signals to improve the textures in RGB frames, yielding promising VSR reconstruction. As for event-based VFI, TimeLens [68] and TimeLens++ [69] innovatively integrate deformation and synthesis processes. CBMNet [34] utilizes cross-modality information to directly estimate inter-frame motion fields. REFID [66], as a flexible method, exhibits robustness even with blurred reference frames. Notably, these methods achieve upsampling only in the spatial or temporal domain independently, and few studies address a unified approach for both, i.e., event-based STVSR." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implicit Neural Representation (INR) in STVSR. As a method of continuous representation of signals, INR is a common means of continuous super-resolution, including image [4, 5, 17, 38, 83] and video [6, 7, 49, 50] tasks. Inspired by LIIF [5], VideoINR [7] pioneers continuous STVSR (C-STVSR) by learning separate spatial and temporal INRs. Building on this, MoTIF [6] introduces forward spatiotemporal INR to enhance motion field estimation, but it retains the paradigm of decoupling video INR across multiple domains, which lacks robustness across varying scales. Similarly, HR-INR [50], a recent event-based C-STVSR method, also adheres to the same decoupling paradigm as VideoINR [7]. Besides, HR-INR conducts holistic event-frame extraction for long-range motion dependencies and requires repeatedly extracting input frames to realize frame temporal super-resolution, introducing redundancy. Instead, we break the inherent paradigm. We devise the EASM that executes event-modulated alignment and bidirectional recurrent compensation for long-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "term motion trajectory modeling. Based on this, our proposed LIVT achieves a unified spatiotemporal INR through the high temporal resolution grids generated from events, thus achieving better generalization with lower weights." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.165, + 0.688, + 0.182 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.905, + 0.28 + ], + "angle": 0, + "content": "Given LR and low-frame-rete (LFR) videos and corresponding event streams as inputs, we present an event-driven C-STVSR approach to reconstruct HR and HFR videos at arbitrary spatial and temporal scales. Before elaborating on the components of our methodology, we first provide the details for the event representation in this work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.719, + 0.305 + ], + "angle": 0, + "content": "3.1. Event Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Event cameras asynchronously capture the logarithmic intensity change \\(\\Delta \\log (I_{x,y,\\tau})\\) of each pixel at coordinate \\((x,y)\\) within the RGB image \\(I_{x,y,\\tau}\\), where \\(\\tau\\) denote the timestamp of an event stream \\(\\mathcal{E} = (x,y,\\tau,p)\\). \\(p\\) represents the polarity of \\(\\mathcal{E}\\) triggered at \\((x,y)\\), with 1 and -1 representing the change directions respectively. Compared to conventional RGB images, a raw event stream constituted by large amounts of sparse points, which is capable of continuously capturing intensity change with high temporal resolution. However, processing such an unstructured sparse format in deep networks is non-trivial. Following existing event-based method [66], we represent the event stream into a 3D spatiotemporal voxel grid \\(\\mathcal{V} \\in \\mathbb{R}^{H \\times W \\times (M + 1)}\\) linearly, where \\(H \\times W\\) denotes the spatial size, and then discretize the temporal dimension into \\(M\\) event segments \\([E_1,E_2,\\dots,E_M]\\). Each segment records the intensity information within a time window of \\(\\mathcal{E}\\), which is often set to be a small constant (see Figure 2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.625, + 0.605 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Let \\(\\mathbf{I}^{LR} = I_0^{LR}, I_1^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}\\) denote two LR RGB frames and \\(\\mathcal{E}_{0 \\to 1}\\) denote their corresponding event stream. We propose EvEnhancer that addresses the task of super-resolving \\(\\mathbf{I}^{LR}\\) to \\((t + 1)\\) HR video frames with any spatial scale \\(s \\geq 1\\) and temporal scale \\(t \\geq 1\\). The overall framework is illustrated in Figure 2, which consists of an event-adapted synthesis module (EASM) and a local implicit video transformer (LIVT)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Generally, we first extract initial features on frames and event segments, resulting in RGB frame features \\(F_0^{LR}\\), \\(F_1^{LR}\\), and event feature \\(\\mathbf{F}_{0\\rightarrow 1}^E = \\{F_{0\\rightarrow 1}^{E,m}\\}_{m = 1}^M\\) for \\(M\\) segments. Considering that the latent inter-frame features between RGB frames can be captured in both forward and backward directions, we also reverse the time and polarity of \\(\\nu\\) to derive a backward event feature \\(\\mathbf{F}_{1\\rightarrow 0}^E\\). Then, EASM models long-term holistic motion trajectories with the guidance of events to generate a spatiotemporal feature sequence \\(\\widehat{\\mathbf{F}}\\) with high temporal resolution. Finally, LIVT integrates cross-scale spatiotemporal attention with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17757" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.451, + 0.908, + 0.494 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our EvEnhancer, an event-driven C-STVSR method which consists of an event-adapted synthesis module (EASM), and a local implicit video transformer (LIVT), where EASM contains two steps: (a) event-modulated alignment (EMA), and (b) bidirectional recurrent compensation (BRC). \"CS\": channel squeeze, \"FFN\": feed-forward neural network, \"Res\": residual blocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.548 + ], + "angle": 0, + "content": "local implicit function to learn continuous video INR \\(\\mathbf{F}_{st}\\) that decodes arbitrary space-time coordinates into RGB values, thus yielding HR video frames \\(\\mathbf{I}_{st}^{HR}\\) with HFR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.384, + 0.575 + ], + "angle": 0, + "content": "3.3. Event-Adapted Synthesis Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.717 + ], + "angle": 0, + "content": "As depicted in Figure 2, EASM contains two steps: 1) Event-modulated alignment incorporates motion cues between input frames with event modulation to acquire latent inter-frame features. Based on the event features and reverse ones, we deploy such alignment in parallel to model the motions forward \\((^{+} + )\\) and backward \\((^{-})\\). 2) Bidirectional recurrent compensation propagates the event stream across time and fuses it with acquired features in both directions to maximize the gathering of temporal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Event-Modulated Alignment (EMA). Given input frame features \\( F_0^{LR} \\) and \\( F_1^{LR} \\), our goal here is to infer the features of intermediate frames according to \\( M \\) event segment features. Motivated by the effectiveness of the PCD module in video restoration [75, 79], we construct a customized pyramid and event-modulated structure that progressively learns the motion offsets and feature alignment. In the following, we take the forward alignment as an example for clarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Specifically, for the forward alignment, as shown in Figure 2(a), we implement a 3-level pyramid by downsampling the frame and event features \\((F_0^{LR}, F_1^{LR}, F_{0\\rightarrow 1}^{E,m})\\) to 3 different spatial scales \\(\\times 1\\), \\(\\times \\frac{1}{2}\\), \\(\\times \\frac{1}{4}\\) via stride convolutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.676 + ], + "angle": 0, + "content": "At each level \\(l\\), we first calculate an initial motion vector by a convolutional block (\"Conv+LeakyReLU (LReLU)\") on concatenated frame features, which is then fused with the learned offset \\(O_{m}^{(l-1)}\\) from the \\((l-1)\\)-th level to obtain a motion vector \\(MV_{m}^{(l)}\\) near the \\(m\\)-th timestamp. Then, we design an event modulation block (EMB) to generate modulated \\(\\widetilde{MV}_{m}^{(l)}\\). The modulation is controlled by the event feature \\(F_{0\\rightarrow 1}^{E,m}\\). Compared to previous PCD-based methods [75, 79, 81], our event-based approach can furnish more deterministic motion cues to facilitate frame synthesis, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.687, + 0.907, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\widetilde {M V} _ {m} ^ {(l)} = \\mathcal {M} ^ {(l)} \\left(M V _ {m} ^ {(l)}, F _ {0 \\rightarrow 1} ^ {E, m (l)}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.907, + 0.839 + ], + "angle": 0, + "content": "where \\( F_{0\\rightarrow 1}^{E,m(l)} \\) denote the downsampled event feature at level \\( l \\). Then, we can learn the offset \\( O_{m}^{(l)} \\) from \\( MV_{m}^{(l)} \\) and \\( \\widetilde{MV}_{m}^{(l)} \\). Following [75], we use the deformable convolutional network (DCN) [92] for feature alignment and interpolation. The deformable alignments at different levels are cascaded and aggregated coarse-to-fine for accuracy improvement via the pyramid workflow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Notably, as there are \\(M\\) segment features \\(\\{F_{0\\rightarrow 1}^{E,m}\\}_{m = 1}^{M}\\) by the event-modulated alignment, we can produce \\(M\\) intermediate frame features \\(\\mathbf{F}^{+} = [F_{1}^{LR + },F_{2}^{LR + },\\dots ,F_{M}^{LR + }]\\). Similarly, for the backward alignment, we get \\(\\mathbf{F}^{-} =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17758" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.089, + 0.482, + 0.167 + ], + "angle": 0, + "content": "\\([F_1^{LR-}, F_2^{LR-}, \\dots, F_M^{LR-}]\\). The two are integrated by channel concatenation with a convolutional block, forming the final features \\(\\mathbf{F} = \\{F_{\\tau}^{LR}\\}\\) with the input features \\(F_0^{LR}\\), \\(F_1^{LR}\\), where the timestamp \\(\\tau = 0, \\frac{1}{M+1}, \\dots, \\frac{M}{M+1}, 1\\) is aligned with the input frames and \\(M\\) event segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Bidirectional Recurrent Compensation (BRC). Given that the information encapsulated within each event segment is confined to a very narrow time window, after the modulated alignment, we propose bidirectional recurrent compensation that iteratively fuses the event and frame features to model the global temporal information inherent in two modalities, as shown in Figure 2(b). In each recurrent iteration, the compensation block receives the event feature \\( F_{\\tau}^{E} \\) from the current timestamp \\( \\tau \\), the forward hidden state from \\( h_{\\tau - \\Delta \\tau}^{f} \\) the previous timestamp \\( \\tau - \\Delta \\tau \\), and the backward hidden state \\( h_{\\tau + \\Delta \\tau}^{b} \\) from the next timestamp \\( \\tau + \\Delta \\tau \\), and the aligned frame feature \\( F_{\\tau}^{LR} \\in \\mathbf{F} \\). Here, \\( \\tau \\pm \\Delta \\tau \\) denotes the former timestamp and the latter one. Hence, the recurrent fusion process at timestamp \\( \\tau \\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.392, + 0.482, + 0.418 + ], + "angle": 0, + "content": "\\[\nr _ {\\tau} ^ {b}, h _ {\\tau} ^ {b} = \\mathcal {F} ^ {b} \\left(\\mathcal {C} ^ {b} \\left(F _ {\\tau} ^ {L R}, F _ {\\tau} ^ {E}\\right), h _ {\\tau + \\Delta \\tau} ^ {b}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.414, + 0.425, + 0.434 + ], + "angle": 0, + "content": "\\[\nr _ {\\tau} ^ {f}, h _ {\\tau} ^ {f} = \\mathcal {F} ^ {f} (\\mathcal {C} ^ {f} (F _ {\\tau} ^ {L R}, F _ {\\tau} ^ {E}), h _ {\\tau - \\Delta \\tau} ^ {f}, r _ {\\tau} ^ {b}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.627 + ], + "angle": 0, + "content": "where \\( r_{\\tau}^{b} \\) and \\( h_{\\tau}^{b} \\) denotes the output and hidden states respectively in the backward direction, and \\( r_{\\tau}^{f} \\) and \\( h_{\\tau}^{f} \\) corresponds to the forward. \\( \\mathcal{C}^f \\) and \\( \\mathcal{C}^b \\) calculate the channel attention on event and frame features during the forward and backward recurrent iterations respectively. By recurrent propagating the event information across time, we can comprehensively exploit the high temporal resolution properties from events to refine the frame features and maximize the gathering of temporal information. Finally, a bidirectional feature fusion is run by the element-wise addition between the input feature fed into BRC and its output bidirectional features to retain valid information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.388, + 0.654 + ], + "angle": 0, + "content": "3.4. Local Implicit Video Transformer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.84 + ], + "angle": 0, + "content": "After the EASM, we obtain the discrete temporal feature sequence \\(\\widehat{\\mathbf{F}}\\). The other key in this method is to learn continuous video INR that specifies the feature at the queried space-time coordinate \\((\\mathcal{T},x,y)\\) are transferred to any target spatial scale \\(s\\) and temporal scale \\(t\\) in Figure 2. Previous C-STVSR methods [6, 7] learn spatial and temporal INR separately to realize video INR, which fails to fully capitalize on the spatiotemporal dependencies in the continuous domain. In this work, we propose LIVT which calculates cross-scale 3D attention based on the queried space-time coordinates in the continuous domain over a local sampling grid to learn an accurate unified video INR directly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Assuming the temporal length and spatial size of the local grid in LR space are \\( T^G \\) and \\( H^{G} \\times W^{G} \\), respectively. For the full temporal range [0, 1], to generate a frame at the timestamp \\( \\mathcal{T} \\in [0,1] \\), we only need to select the \\( T^G \\) nearest" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.907, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.301, + 0.907, + 0.357 + ], + "angle": 0, + "content": "Figure 3. The structure of our local implicit video transformer (LIVT), which integrates the 3D local spatiotemporal attention with implicit neural function to learn continuous video INR to reconstruct HR and HFR video frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.907, + 0.457 + ], + "angle": 0, + "content": "elements to \\(\\mathcal{T}\\) along the time axis. Let the candidate selection of the timestamps \\(\\tau_{i} = 0,\\frac{1}{M + 1},\\dots,\\frac{M}{M + 1},1\\) , which in actual align with the \\(M\\) event segments and input frames \\(\\mathbf{I}^{LR}\\) . We denote the set of candidate timestamps \\(\\{\\tau_i\\}\\) as \\(\\mathbf{S}_{\\tau}\\subseteq [0,\\frac{1}{M + 1},\\dots,\\frac{M}{M + 1},1]\\) , and the size of the \\(|\\mathbf{S}_{\\tau}| = T^G\\) which is selected following the rule as" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.467, + 0.905, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {\\tau} = \\underset {\\mathbf {S} _ {\\tau}} {\\arg \\min } \\sum_ {\\tau_ {i} \\in \\mathbf {S} _ {\\tau}} | \\tau_ {i} - \\mathcal {T} |. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Then, as shown in Figure 3, with the LR feature sequence \\(\\widehat{\\mathbf{F}}\\) from EASM, we encode them into three embedding spaces through 3D convolution: key \\(\\{K\\}\\), query \\(\\{Q\\}\\), and value \\(\\{V\\}\\). Notably, we do not calculate the spatiotemporal attention across the overall time range. Based on Eq. 3, we conduct temporal selection on \\(K\\), \\(Q\\), and \\(V\\), which allows the learning for 3D local attention in a subtemporal space near the target timestamp \\(\\mathcal{T}\\). On top of this, we employ 3D trilinear sampling on \\(Q\\) to get the query \\(q_{\\mathcal{T}} \\in \\mathbb{R}^{s^2 HW \\times 1 \\times C}\\) for a certain frame, where \\(C\\) is the channel number. However, since both \\(k_{\\mathcal{T}}\\) and \\(v_{\\mathcal{T}}\\) have size \\(s^2 HW \\times T^G H^G W^G \\times C\\), it will suffer from expensive computational costs if we directly calculate the attention. To solve this problem, we propose to unfold the sequence along the temporal dimension and calculate cross-scale attention separately on \\(T^G\\) embeddings. Thus, we sample the key \\(k_{\\tau_i}\\) and value \\(v_{\\tau_i}\\) from \\(K\\) and \\(V\\) respectively using 2D nearest local sampling at each time slice \\(\\tau_i\\). Moreover, we encode and reshape the spatiotemporal relative coordinates \\((\\delta \\tau, \\delta x, \\delta y) \\in (-1, 1)\\) from each query point to all pixel points within its local grid, obtaining the spatiotemporal bias \\(b_{\\tau_i} \\in \\mathbb{R}^{s^2 HW \\times H^G W^G \\times 3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We compute the inner product between \\( q_{\\mathcal{T}} \\) and \\( k_{\\tau_i} \\) and add it with the \\( b_{\\tau_i} \\) to obtain the local attention map at the timestamp \\( \\mathcal{T} \\) after a Softmax operation. After the matrix multiplication with \\( v_{\\tau_i} \\), we can derive the output embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17759" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons for in-distribution (In-dist.) spatiotemporal upsampling scales (\\( t = 4 \\), \\( f = 8 \\)). † indicates the model trained on the same Adobe240 dataset [64] as ours, and ‡ indicates the model fine-tuned on the Adobe240 dataset. **Bold** and **underline** indicate the best and the second-best performance, respectively. Metrics: PSNR (dB) / SSIM." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.142, + 0.887, + 0.354 + ], + "angle": 0, + "content": "
VFI MethodVSR MethodEventsGoPro-CenterGoPro-AverageAdobe240-CenterAdobe240-AverageParams (M)
TimeLens [69]EGVSR [49]✓28.41 / 0.830727.42 / 0.807726.64 / 0.767525.13 / 0.729872.20+2.45
TimeLens [69]EvTexture [31]✓30.50 / 0.878428.51 / 0.847828.80 / 0.833726.12 / 0.778972.20+8.90
CBMNet-L [34]EGVSR [49]✓26.14 / 0.767423.55 / 0.701825.15 / 0.731122.89 / 0.673522.23+2.45
CBMNet-L [34]EvTexture [31]✓27.58 / 0.796623.92 / 0.716326.74 / 0.777723.38 / 0.698622.23+8.90
REFID [66]†EGVSR [49]✓28.74 / 0.836427.97 / 0.817127.28 / 0.784426.55 / 0.763715.91+2.45
REFID [66]†EvTexture [31]✓30.86 / 0.878429.14 / 0.848929.42 / 0.842427.69 / 0.805915.91+8.90
Zooming Slow-Mo [25]†30.69 / 0.8847-30.26 / 0.8821-11.10
TMNet [81]‡30.14 / 0.869228.83 / 0.851429.41 / 0.852428.30 / 0.835412.26
SAFA [25]†31.28 / 0.889430.22 / 0.876130.97 / 0.887830.13 / 0.87824.94
VideoINR-fixed [7]†30.73 / 0.8850-30.21 / 0.8805-11.31
VideoINR [7]†30.26 / 0.879229.41 / 0.866929.92 / 0.874629.27 / 0.865111.31
MoTIF [6]†31.04 / 0.887730.04 / 0.877330.63 / 0.883929.82 / 0.875012.55
HR-INR [50]†✓31.97 / 0.929832.13 / 0.937131.26 / 0.924631.11 / 0.92168.27
EvEnhancer-light✓33.11 / 0.924232.73 / 0.920331.90 / 0.903331.47 / 0.898831.47 / 0.89885.81
EvEnhancer✓33.52 / 0.929533.30 / 0.927932.43 / 0.912932.18 / 0.911632.18 / 0.91166.55
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.369, + 0.325, + 0.386 + ], + "angle": 0, + "content": "\\(\\widetilde{z}_{\\tau_i}\\). This procedure is described as" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.392, + 0.483, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\widetilde {z} _ {\\tau_ {i}} = \\operatorname {S o f t m a x} \\left(\\frac {q _ {\\mathcal {T}} k _ {\\tau_ {i}} ^ {\\mathrm {T}}}{\\sqrt {C}} + b _ {\\tau_ {i}}\\right) \\otimes v _ {\\tau_ {i}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.542 + ], + "angle": 0, + "content": "We concatenate the output embeddings \\(\\widetilde{z}_{\\tau_i}\\) along the channel dimension to construct the complete continuous video INR \\(\\widetilde{Z}_{\\mathcal{T}}\\) to yield the features corresponding to continuous space-time coordinates. Following [4, 5, 7], we decode the features into RGB values using a 5-layer MLP with the previous query \\(q_{\\mathcal{T}}\\) and cell decoding to reconstruct the target RGB frame \\(\\hat{I}_{\\mathcal{T}}^{HR}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.224, + 0.57 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.282, + 0.594 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Dataset. In our experiments, we train all the models on the Adobe240 dataset [64] which contains 133 video sequences, where 100 for training, 16 for validation, and 17 for testing. We evaluate the performance of our models on both synthetic Adobe240 [64] and GoPro [51] datasets and real-world BS-ERGB [69] and ALPIX-VSR [49] datasets. For the synthetic events, we use the event simulation method vid2e [11] to generate simulated event data. Given a spatial scale \\( s \\) and a temporal scale \\( t \\), we select consecutive \\( (t + 1) \\) frames as a clip and select the first and the last frames for \\( s: 1 \\) bicubic downsampling as the LR and LFR inputs fed into the network to generate HR and HFR frame sequences, and the original \\( (t + 1) \\) frames are groundtruth (GT). Following [31], the event voxels are bicubic-downsampled as the LR event inputs, thus ensuring the spatial size is aligned with the LR frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Referring VideoINR [7], the training for our model comprises two stages. In the first stage, we set \\( s = 4 \\) and \\( t = 8 \\), training for 450K iterations on Adobe240 [64]. In the second stage, we uni-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.907, + 0.552 + ], + "angle": 0, + "content": "formly sample \\( s \\) in [1, 4] and fine-tune for another 150K iterations on the same dataset. We use the Adam optimizer [35] with \\( \\beta_{1} = 0.9 \\) and \\( \\beta_{2} = 0.999 \\), and apply cosine annealing to decay the learning rate from \\( 1 \\times 10^{-4} \\) to \\( 1 \\times 10^{-7} \\) over 150K iterations. For data augmentation, we randomly crop \\( 32 \\times 32 \\) spatial regions from downsampled images and event segments with random rotations and horizontal flips. The model is optimized with Charbonnier loss [36] using \\( \\epsilon^{2} = 1 \\times 10^{-6} \\). In this work, we implement two types of variants, the full model EvEnhancer and a lightweight version EvEnhancer-light. To measure our models, we calculate the PSNR (dB) and SSIM on the Y channel." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.882, + 0.577 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.81 + ], + "angle": 0, + "content": "We comprehensively consider frame-based and event-based one-stage or two-stage state-of-the-art methods for comparison: 1) two-stage event-based STVSR methods consisting of event-based VFI (TimeLens [68], REFID [66], CBMNet-L [34]) and event-based VSR (EGVSR [49], EvTexture [31]); 2) one-stage frame-based F-STVSR methods (Zooming Slow-Mo[79], TMNet [81], SAFA [25]); 3) one-stage frame-based C-STVSR methods (VideoINR [7], MoTIF [6]); 4) one-stage event-based C-STVSR methods (HR-INR [50]). For fair comparisons, we use the models of most VFI methods and all STVSR methods trained or fine-tuned on the same Adobe240 dataset [64] as ours. Notably, due to the limitation that the HR-INR source code is not yet available, the results of HR-INR are derived from its original paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. Table 1 reports the results for the F-STVSR task, where the spatial and temporal scales are the same as the training distribution (In-Dist.). \"Center\" represents the average performance of the reconstructed left, right, and center frames. \"Average\" represents the average performance of all reconstructed frames. Notably," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17760" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.089, + 0.87, + 0.104 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons for out-of-distribution (OOD) spatiotemporal upsampling scales on the GoPro dataset [51]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.902, + 0.273 + ], + "angle": 0, + "content": "
Temporal ScaleSpatial ScaleTMNet [81]‡SAFA [25]†VideoINR [7]†MoTIF [6]†HR-INR [50]†EvEnhancer -lightEvEnhancer
t = 6s = 430.49 / 0.886131.68 / 0.906830.78 / 0.895431.56 / 0.9064-33.10 / 0.926633.41 / 0.9300
t = 6s = 6--25.56 / 0.767129.36 / 0.8505-29.84 / 0.860330.12 / 0.8675
t = 6s = 12--24.02 / 0.690025.81 / 0.7330-25.41 / 0.726125.50 / 0.7323
t = 12s = 426.38 / 0.793127.95 / 0.824927.32 / 0.814127.77 / 0.823028.87 / 0.885431.17 / 0.896932.07 / 0.9116
t = 12s = 6--24.68 / 0.735826.78 / 0.790827.14 / 0.817328.94 / 0.842529.54 / 0.8579
t = 12s = 12--23.70 / 0.683024.72 / 0.7108-25.25 / 0.724825.45 / 0.7345
t = 16s = 424.72 / 0.752626.38 / 0.783225.81 / 0.773925.98 / 0.775827.29 / 0.855630.15 / 0.876230.93 / 0.8918
t = 16s = 6--23.86 / 0.712325.34 / 0.752726.09 / 0.795428.34 / 0.828328.86 / 0.8434
t = 16s = 12--22.88 / 0.665923.88 / 0.6923-25.10 / 0.721725.25 / 0.7310
t = 6s = 1--32.34 / 0.954534.77 / 0.969638.53 / 0.973537.68 / 0.966738.80 / 0.9714
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.285, + 0.482, + 0.313 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons for different temporal upsampling scales \\( t \\) (fixed \\( s = 4 \\) ) on the BS-ERGB dataset [69]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.324, + 0.482, + 0.468 + ], + "angle": 0, + "content": "
MethodEventst=4t=6t=8
TimeLens[69]+EGVSR[49]✓22.98/0.676222.01/0.659221.33/0.6467
TimeLens[69]+EvTexture[31]✓23.58/0.708822.48/0.688721.74/0.6746
CBMNet-L[34]+EGVSR[49]✓22.74/0.685621.93/0.670521.29/0.6657
CBMNet-L[34]+EvTexture[31]✓23.40/0.719922.47/0.702021.78/0.6962
REFID[66]†+EGVSR[49]✓--22.76/0.6478
REFID[66]†+EvTexture[31]✓--22.77/0.6527
SAFA[25]†24.32/0.731423.23/0.712022.41/0.6954
VideoINR[7]†24.06/0.729023.00/0.709822.23/0.6945
MoTIF[6]†24.21/0.729623.15/0.709922.37/0.6939
EvEnhancer-light✓25.32/0.731124.54/0.718424.16/0.7100
EvEnhancer✓25.44/0.733824.70/0.721524.29/0.7138
" + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.482, + 0.478, + 0.497 + ], + "angle": 0, + "content": "Table 4. \\(2\\times\\) VSR performance on the ALPIX-VSR dataset [49]." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.507, + 0.48, + 0.535 + ], + "angle": 0, + "content": "
MethodVideoINR [7]MoTIF [6]EGVSR [49]EvEnhancer
PSNR / SSIM32.53 / 0.938338.61 / 0.963639.31 / 0.963540.84 / 0.9786
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.78 + ], + "angle": 0, + "content": "VideoINR-fixed is trained specifically for single-frame interpolation. We can see that both our EvEnhancer and EvEnhancer-light significantly outperform existing methods in all settings with much fewer model parameters, which achieves more than 1 dB PSNR improvements. Then, we also evaluate our models on the C-STVSR task, where spatial and temporal scales are out of the training distribution. The results are illustrated in Table 2. It can be observed that both our models exceed almost all existing methods at various spatial and temporal scales, especially in the challenging large temporal scale, achieving superior effectiveness and generalization ability. In particular, our EvEnhancer-light and EvEnhancer use lighter-weight models and perform better than HR-INR [50] at most temporal and spatial scales, especially at OOD scales." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We further validate the effectiveness of our method on the real-world BS-ERGB dataset under the setting of fixed spatial scale \\( s = 4 \\) but varying temporal scales \\( t \\). The comparisons in Table 3 demonstrate that our method yields the best quantitative performance under all conditions. Compared to cascaded event-based VFI and VSR, our EvEnhancer simultaneously learns temporal interpolation and spatial VSR, thus achieving superior performance. Due to" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.284, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Table 5. TFLOPs comparisons of C-STVSR models at different upsampling temporal scales \\( t \\) (fixed \\( s = 4 \\)). * indicates the repeated iterations during inference on the NVIDIA GeForce RTX 3090 GPU due to their large computations." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.351, + 0.902, + 0.448 + ], + "angle": 0, + "content": "
Temporal ScaleVideoINR [7] (11.31M)MoTIF [6] (12.55M)EvEnhancer -light (5.81M)EvEnhancer (6.55M)
t = 22.0112.0431.8763.398
t = 42.3954.086*2.4724.647
t = 62.7796.129*3.0675.885
t = 83.1636.129*3.6637.129
t = 123.93210.22*4.8549.617
t = 164.70012.26*6.04412.10
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.471, + 0.907, + 0.623 + ], + "angle": 0, + "content": "the high temporal resolution of events, equipped by EASM, our EvEnhancer enables holistic motion trajectory modeling, thus outperforming frame-based methods. Benefiting from the cross-scale attention in LIVT, it also performs preferably at OOD scales. To validate the effectiveness of our EvEnhancer for real-world LR events, we conduct experiments on the real-world ALPIX-VSR dataset [49] (image resolution: \\(3264 \\times 2448\\), event resolution: \\(1632 \\times 1224\\)) for \\(2 \\times\\) VSR. As shown in Table 4, EvEnhancer significantly surpasses all methods without additional fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.626, + 0.909, + 0.792 + ], + "angle": 0, + "content": "Model Complexity. To validate the model efficiency of the proposed methods, besides the model parameter comparison in Table 1, we compare the TFLOPs with existing C-STVSR methods in Table 5, which are calculated on the upsampling of a clip with a resolution of \\(180 \\times 320\\) by a spital upsampling scale \\(s = 4\\) at different temporal upsampling scales \\(t\\). VideoINR shows the fewest TFLOPs but more parameters and the worst performance (see Table 1). At the low temporal scale (\\(\\leq 4\\)), MoTIF is computationally efficient than ours. However, as the temporal scale increases, its computations increase and exceed both our models." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparison. With the qualitative results in Figure 4, compared to VideoINR [7] and MoTIF [6], our models reconstruct the faithful HR frames with clearer textures. By computing the difference maps between generated frames and corresponding ground truth frames, we can see that our methods are more excellent at tackling non-linear motions and recovering more details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17761" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.895, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.376, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison for In-dist. scale \\((t = 8, s = 4)\\) on the GoPro dataset [51]. We compare the normalized absolute difference maps (yellow boxes) for the same regions in each frame as in the GT frames." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.419, + 0.487, + 0.462 + ], + "angle": 0, + "content": "Table 6. Ablations on the designs of event-adapted synthesis module (EASM). \"fwd.\" and \"bwd.\" denote the forward and backward directions, respectively. Metrics: PSNR (dB) / SSIM." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.471, + 0.445, + 0.54 + ], + "angle": 0, + "content": "
EMABRCBFIn-Dist.OOD
fwd. & bwd.32.50 / 0.916628.86 / 0.8407
fwd. & bwd.✓33.13 / 0.925229.42 / 0.8539
fwd.✓✓33.20 / 0.926229.47 / 0.8556
fwd. & bwd.✓✓33.30 / 0.927929.54 / 0.8579
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.258, + 0.573 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.658 + ], + "angle": 0, + "content": "We conduct ablation studies by keeping the same training strategy on the Adobe240 dataset [64], and testing on the GoPro dataset [51] at both In-Dist. \\((t = 8, s = 4)\\) and OOD \\((t = 12, s = 6)\\) spatiotemporal scales to validate the effectiveness of our methodology design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Event-Adapted Synthesis Module. Table 6 presents the ablations of EASM, which involves the event-modulated alignment (EMA), bidirectional recurrent compensation (BRC), and bidirectional feature fusion (BF). The model with only EMA achieves the worst performance. When we introduce BRC in the model, one can see the performance improves. Besides, with the integration BRC, BF, and EMA, the model achieves the best at both In-Dist. and OOD scales. We can also observe that the model with only a single forward alignment can suffer from performance degradation. The results demonstrate the effects of EASM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Local Implicit Video Transformer. Table 7 investigates the effect of LIVT. We first implement the model with a 2D decoupling strategy as videoINR [7] but introduce the local attention, concatenate cell decoding, and query, one can see that the model achieves better performance than the model" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.419, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Table 7. Ablations on the designs of local implicit video transformer (LIVT). \"LA\": local attention, \"CE\": concatenate cell decoding, \"PQ\": concatenate the previous query. Metrics: PSNR (dB) / SSIM." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.485, + 0.88, + 0.566 + ], + "angle": 0, + "content": "
INR ModeLACEPQIn-dist.OOD
2D decoupling✓✓✓33.26 / 0.927229.26 / 0.8525
3D unification27.32 / 0.797525.03 / 0.7032
3D unification✓33.27 / 0.927629.47 / 0.8565
3D unification✓✓33.30 / 0.928229.49 / 0.8570
3D unification✓✓✓33.30 / 0.927929.54 / 0.8579
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.907, + 0.678 + ], + "angle": 0, + "content": "without any components. When we replace the 2D decoupling with our unified 3D video INR, the model produces results with higher PSNR and SSIM, especially at OOD scales. In addition, the concatenation of the cell [5] and previous query [7] brings a larger gain at OOD scales than In-Dist., which validates their effects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we introduce EvEnhancer, a novel approach that amalgamates the unique advantages of event streams to enhance the effectiveness, efficiency, and generalizability of C-STVSR. Our EvEnhancer is underpinned by two critical components: 1) the EASM capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories; and 2) the LIVT, which integrates a local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations. Experiments indicate that EvEnhancer significantly surpasses current state-of-the-art methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17762" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.237 + ], + "angle": 0, + "content": "This work was partly supported by the National Natural Science Foundation of China (62302141, 62331003, 62120106009), the Joint Funds of the National Natural Science Foundation of China (U23A20314), the Fundamental Research Funds for the Central Universities (JZ2024HGTB0255), Beijing Natural Science Foundation (L223022), and the Natural Science Foundation of Hebei Province (F2024105029)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.248, + 0.188, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.273, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[4] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[6] Yi-Hsin Chen, Si-Cun Chen, Yen-Yu Lin, and Wen-Hsiao Peng. Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video superresolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23131-23141, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[7] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2047-2057, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[8] Xianhang Cheng and Zhenzhong Chen. Video frame interpolation via deformable separable convolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10607-10614, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.273, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.246 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[11] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In CVPR, pages 3586-3595, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.907, + 0.375 + ], + "angle": 0, + "content": "[12] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[13] Jin Han, Yixin Yang, Chu Zhou, Chao Xu, and Boxin Shi. Evintsr-net: Event guided multiple latent frames reconstruction and super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4882-4891, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[14] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3897-3906, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.586 + ], + "angle": 0, + "content": "[15] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Space-time-aware multi-resolution video enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2859-2868, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.907, + 0.672 + ], + "angle": 0, + "content": "[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[17] Zongyao He and Zhi Jin. Latent modulated function for computational optimal continuous image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26026-26035, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[18] Qiqi Hou, Abhijay Ghildyal, and Feng Liu. A perceptual quality metric for video frame interpolation. In European Conference on Computer Vision, pages 234-253. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[20] Mengshun Hu, Kui Jiang, Zheng Wang, Xiang Bai, and Ruimin Hu. Cycmunet+: Cycle-projected mutual learning" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17763" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "for spatial-temporal video super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[21] Ping Hu, Simon Niklaus, Lu Zhang, Stan Sclaroff, and Kate Saenko. Video frame interpolation with many-to-many splating and spatial selective refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[22] Cong Huang, Jiahao Li, Lei Chu, Dong Liu, and Yan Lu. Arbitrary-scale video super-resolution guided by dynamic context. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2302, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.236, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[23] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 40(4):1015-1028, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[25] Zhewei Huang, Ailin Huang, Xiaotao Hu, Chen Hu, Jun Xu, and Shuchang Zhou. Scale-adaptive feature aggregation for efficient space-time video super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4228-4239, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[26] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.483, + 0.561 + ], + "angle": 0, + "content": "[27] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[28] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. arXiv preprint arXiv:2008.05765, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[29] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[31] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. EvTexture: Event-driven Texture Enhancement for Video Super-Resolution. In Proceedings of the 41st International Conference on Machine Learning, pages 22817–22839. PMLR, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[32] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[33] Soo Ye Kim, Jihyong Oh, and Munchurl Kim. Fisr: Deep joint frame interpolation and super-resolution with a multiscale temporal loss. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11278-11286, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[34] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[36] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[37] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5316-5325, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[38] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1929-1938, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[39] Feng Li, Huihui Bai, and Yao Zhao. Learning a deep dual attention network for video super-resolution. IEEE transactions on image processing, 29:4474-4488, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[40] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[41] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10522-10531, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[42] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[43] Zekun Li, Hongying Liu, Fanhua Shang, Yuanyuan Liu, Liang Wan, and Wei Feng. Savsr: Arbitrary-scale video super-resolution via a learned scale-adaptive network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3288-3296, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[44] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[45] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17764" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.148 + ], + "angle": 0, + "content": "Learning event-driven video deblurring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 695-710. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[46] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[47] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19125-19134, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[48] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[49] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1557-1567, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[50] Yunfan Lu, Zipeng Wang, Yusheng Wang, and Hui Xiong. Hr-inr: Continuous space-time video super-resolution via event camera. arXiv preprint arXiv:2405.13389, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[51] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[52] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[53] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE international conference on computer vision, pages 261-270, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[54] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive convolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 670-679, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[55] Simon Niklaus, Ping Hu, and Jiawen Chen. Splitting-based synthesis for video frame interpolation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 713-723, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[56] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14539-14548, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[57] Zhongwei Qiu, Huan Yang, Jianlong Fu, Daochang Liu, Chang Xu, and Dongmei Fu. Learning degradation-robust spatiotemporal frequency-transformer for video superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[58] Nasrin Rahimi and A Murat Tekalp. Spatio-temporal perception-distortion trade-off in learned video sr. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1400-1404. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[59] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF international conference on computer vision, pages 892-900, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[60] Oded Shahar, Alon Faktor, and Michal Irani. Space-time super-resolution from a single video. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[61] Wei Shang, Dongwei Ren, Wanying Zhang, Yuming Fang, Wangmeng Zuo, and Kede Ma. Arbitrary-scale video superresolution with structural and textural priors. In European Conference on Computer Vision, pages 73-90. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[62] Eli Shechtman, Yaron Caspi, and Michal Irani. Space-time super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(4):531-545, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[63] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[64] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[66] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[67] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[68] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[69] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17765" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[70] Hai Wang, Wenming Yang, Qingmin Liao, and Jie Zhou. Bi- rstu: Bidirectional recurrent upsampling network for spacetime video super-resolution. IEEE Transactions on Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[71] Hai Wang, Xiaoyu Xiang, Yapeng Tian, Wenming Yang, and Qingmin Liao. Stdan: deformable attention network for space-time video super-resolution. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.21, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through hr optical flow estimation. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 514–529. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[73] Li Wang, Dong Li, Yousong Zhu, Lu Tian, and Yi Shan. Dual super-resolution learning for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3774-3783, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[74] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.412, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[76] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.556, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[77] Guangyang Wu, Xin Tao, Changlin Li, Wenyi Wang, Xiao-hong Liu, and Qingqing Zheng. Perception-oriented video frame interpolation via asymmetric blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2753-2762, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[78] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In European Conference on Computer Vision, pages 267-283. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.783 + ], + "angle": 0, + "content": "[79] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3370-3379, 2020. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.786, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[80] Chenming Xu, Meiqin Liu, Chao Yao, Weisi Lin, and Yao Zhao. Ibvc: Interpolation-driven b-frame video compression. Pattern Recognition, 153:110465, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[81] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6388-6397, 2021. 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[82] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[83] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[84] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[85] Jun-Sang Yoo, Hongjae Lee, and Seung-Won Jung. Video object segmentation-aware video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12322-12333, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[86] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17834-17843, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[87] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S Ren. Training weakly supervised video frame interpolation with events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14589-14598, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[88] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[89] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8799-8808, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[90] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[91] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[92] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17766" + } + ] +] \ No newline at end of file diff --git a/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_origin.pdf b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2bf7f33c41a66e45e814dce23c539953e8fca1ab --- /dev/null +++ b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/84be2ed9-42d5-47be-92a0-ca505e0f2685_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:325acbc5fdda01bc5c4f2e342e5358bbc33b7663e8041e3aaa8ea71de287505d +size 8447168 diff --git a/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/full.md b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f90f43bc8746441ad6c191c17cb3056e3b6e6827 --- /dev/null +++ b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/full.md @@ -0,0 +1,354 @@ +# EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events + +Shuoyan Wei $^{1,2}$ Feng Li $^{3,*}$ Shengeng Tang $^{3}$ Yao Zhao $^{1,2}$ Huihui Bai $^{4,1,2,*}$ + +$^{1}$ Institute of Information Science, Beijing Jiaotong University + +$^{2}$ Beijing Key Laboratory of Advanced Information Science and Network Technology + +$^{3}$ School of Computer Science and Engineering, Hefei University of Technology + +$^{4}$ Tangshan Research Institute of Beijing Jiaotong University + +{shuoyan.wei, yzhao, hhbai}@bjtu.edu.cn, {fengli, tangsg}@hfut.edu.cn + +# Abstract + +Continuous space-time video super-resolution (C-STVSR) endeavors to upscale videos simultaneously at arbitrary spatial and temporal scales, which has recently garnered increasing interest. However, prevailing methods struggle to yield satisfactory videos at out-of-distribution spatial and temporal scales. On the other hand, event streams characterized by high temporal resolution and high dynamic range, exhibit compelling promise in vision tasks. This paper presents EvEnhancer, an innovative approach that marries the unique advantages of event streams to elevate effectiveness, efficiency, and generalizability for C-STVSR. Our approach hinges on two pivotal components: 1) Event-adapted synthesis capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories, enabling the adaptive interpolation and fusion of informative spatiotemporal features; 2) Local implicit video transformer integrates local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations utilized to generate plausible videos at arbitrary resolutions and frame rates. Experiments show that EvEnhancer achieves superiority on synthetic and real-world datasets and preferable generalizability on out-of-distribution scales against state-of-the-art methods. Code is available at https://github.com/W-Shuoyan/EvEnhancer. + +# 1. Introduction + +Video super-resolution (VSR) is a fundamental low-level vision task, aiming to recover high-resolution (HR) videos from low-resolution (LR) inputs. Nevertheless, video resources are mostly archived with reduced spatial resolution + +![](images/efa3564cdc2fd103e4f5882e81c7e5a60e155fa4a1eaf20d70c4e9fa238fc66d.jpg) +Adobe240 +VideoINR (11.31M) + +![](images/2c8f16765429f5d249a256317d9d7e04a55a297f4267a652581fca79134589b3.jpg) +age +-Center: reference & center frames +-Average: all frames + +![](images/f629148fd724f95f0bbeae5a269678787b1e6fa57fb4a725ed2ac179215fa274.jpg) +MoTIF (12.55M) +EvEnhancer-light (5.81M) + +![](images/319189ecb6fc2ee177ec7a9473192ffdbf61c5d264c280d84f05721b7ce36aa6.jpg) +t: temporal scale factor + +![](images/d5dadf5fae3067a50d3bd228910ee8088b46192682332860bad8b94b411458e9.jpg) +(a) Different datasets (In-dist.) + +![](images/778cc400e320508dd656f461854df8a26c9a593a14839b02c021dc0bef0ba73b.jpg) +Overlapping LR Frames +LR Events +(c) Visualization comparison +Figure 1. Performance comparison of different C-STVSR methods. (a) PSNR (dB) comparison for in-distribution (In-dist.) spatiotemporal upsampling scale (temporal scale $t = 8$ , spatial scale $s = 4$ ) on the different datasets including GoPro [51], Adobe240 [64] and BS-ERGB [69]. (b) PSNR (dB) comparison for different spatiotemporal upsampling scales on GoPro. (c) Visualization comparison for In-dist. scale on GoPro. + +![](images/44be04deb02d1728963b276f557fe97c0d4c6570c3a982121b756c9820c9768a.jpg) +EvEnhancer (6.55M) +(b) Different upsampling scales +VideoINR[7] + +![](images/7d391d6829b7b4e017a931cc7043e40e9c81fcde632ed0c4f28072a229f16f93.jpg) +32.90 dB/0.9337 +EvEnhancer (Ours) + +![](images/633db7ff7f2780199452f34c4cce262f02a4ef1b33b448a9de60de467e14bc61.jpg) +$s$ : spatial scale factor +MoTIF[6] + +![](images/bfb82ef11da97bdbc69fa22b73a658d58134ad675821fdc0fce78ee556420959.jpg) +GT + +and constrained frame rates (i.e., temporal resolution) practically, necessitating the conversion to HR and high-frame-rate (HFR) ones for various applications [40, 73, 76, 80, 85, 89] and user experiences [18, 58, 77]. + +To address this problem, a naive approach is to cascade separate VSR [2, 3, 14, 39, 41, 75] and video frame interpolation (VFI) [9, 37, 53, 59, 82] or in turn, which over- + +looks the correlations between two sub-tasks, leading to suboptimal exploitation of spatiotemporal information in videos. Space-time video super-resolution (STVSR) [15, 15, 19, 20, 33, 70, 71, 79] integrates VSR and VFI into a unified framework that super-resolves the videos with intermediate frame interpolation via one-stage process. Both the above approaches are limited to fixed discrete space and time magnifications, cannot provide arbitrary controls. Some researchers study continuous VFI [1, 24, 29, 88] or continuous VSR [22, 43, 49, 61] methods, where the former is capable of variable frames interpolation while the latter enables arbitrary spatial upsampling including non-integer scales. However, these methods operate in isolation, failing to simultaneously ensure spatial and temporal flexibility. + +To achieve this, research on continuous STVSR (C-STVSR) [6, 7, 50] algorithms has emerged, which typically learn video implicit neural representation (INR) to decode 3D spatiotemporal coordinates into RGB values, thus generating HR while HFR videos arbitrarily. These methods mainly decouple the learning for video INR into spatial INR and temporal INR to predict continuous spatial and temporal feature domains separately. This fails to fully capitalize on the spatiotemporal dependencies in the continuous domain, thus hindering the reconstruction performance, particularly at scales falling outside the training distribution. Besides, existing methods [6, 7] approximate intermediate motion fields between consecutive frames captured from conventional frame-based cameras, which is insufficient to infer accurate inter-frame temporal features when motions are vast or non-linear. Recently, event cameras famous as bio-inspired asynchronous sensors with high temporal resolution and high dynamic range, providing a promising solution to fulfill the inter-frame motion fields in video reconstruction [13, 16, 30, 45, 49, 68, 87]. Nonetheless, the potential of events in C-STVSR remains largely unexplored. + +In light of the above discussion, we propose an innovative event-driven approach EvEnhancer, which adeptly integrates the spatiotemporal synergies between frames and events to learn continuous video representations for effective and generalized C-STVSR. First, we present an event-adapted synthesis module (EASM) that executes event-modulated alignment and bidirectional recurrent compensation to explore the event information for long-term holistic motion trajectory modeling. It incorporates motion cues between input frames with event-based modulation forward and backward to acquire latent inter-frame features. We then recurrently propagate the event stream across time and fuse it with aligned frame features in both directions to maximize the gathering of temporal information. + +In contrast to the decoupling strategies [6, 7, 50] for video INR, we propose a local implicit video transformer (LIVT) that integrates local implicit video neural function with spatiotemporal attention. LIVT calculates cross-scale + +3D attention based on the queried space-time coordinates in the continuous domain and the features from EASM over a local sampling grid. In this way, we can directly and efficiently obtain an accurate unified video INR that predicts the RGB values of queried coordinates, enabling to super-resolve videos in arbitrary frames and resolutions. As shown in Figure 1, we conduct experiments with synthetic GoPro [51] and Adobe240 [64] datasets and the real-world BS-ERGB [68] dataset. Experiments demonstrate that EvEnhancer achieves superior effectiveness and efficiency in super-resolving videos in arbitrary spatial and temporal scales within training distribution (In-Dist.) against existing state-of-the-art methods. We also validate its preferable generalizability at out-of-distribution (OOD) scales. The main contributions are as follows: + +- We propose EvEnhancer that subtly marries the unique advantages of event streams with video frames for C-STVSR. Extensive experiments validate its superior performance over recent state-of-the-art methods. +- We propose the event-adapted synthesis module (EASM) that enables long-term holistic motion trajectory modeling to acquire informative spatiotemporal features. +- We propose the local implicit video transformer (LIVT) that integrates local implicit video neural function with cross-scale spatiotemporal attention to learn accurate unify continuous video INR efficiently. + +# 2. Related Work + +Video Super-Resolution (VSR). Conventional VSR aims to reconstruct HR videos from LR counterparts, which heavily relies on motion estimation and compensation between frames. Earlier methods address this problem by learning frame-level optical flows [39, 72]. Later methods apply deformable convolution [2, 67, 75] or 3D convolution [23, 27] to conduct implicit feature-level motion compensation, exhibiting improved effectiveness and efficiency. Some methods draw inspiration from attention mechanism [28, 74], which incorporate spatial-temporal attention [42, 84, 86, 90] or video transformer [44, 46, 57, 91] to model inter-frame correlations. However, these methods can only upsample at fixed scales. + +Video Frame Interpolation (VFI). VFI involves interpolating potential frames between consecutive video frames to enhance the temporal resolution. Flow-based methods [32, 56, 82] leverage optical flow networks [10, 26, 65] to estimate the bidirectional flows between frames. Kernel-based methods [8, 53, 54] utilize local convolution to map interframe motion, which synthesizes target pixels by convolving over input frames. In addition, some recent works focus on splatting-based methods [21, 52, 55] or transformer-based methods [47, 48, 63, 88], showing significant advances. These methods almost learn pixel-level motions for interpolation, which is effective in linear motion but strug + +gles with complex non-linear motion or occlusions. + +Space-Time Video Super-Resolution (STVSR). STVSR can be seen as a combination of VSR and VFI, which reconstruct HR videos spatially and temporally. In literature, most existing methods including traditional [60, 62] and deep learning methods [12, 25, 79] are prone to achieve the STVSR with fixed spatial and temporal scales, denoted by F-STVSR. To mitigate the temporal scale constraints, TMNet [81] proposes controlled frame interpolation with time temporal modulation. SAFA [25] introduces a scale adaptive feature aggregation network that iteratively estimates motion with trainable block-wise scale selection. Though these methods achieve temporal flexibility to some extent, they still fail in arbitrary spatial scales. + +Event-based VSR & VFI. Event cameras are characterized by high temporal resolution, high dynamic range, and low latency, promoting many VFI [34, 66, 68, 69, 78] and VSR [13, 30, 49] works. E-VSR [30] introduces the pioneering event-based VSR method that utilizes event-based asynchronous interpolation to synthesize asynchronous neighboring frames. EGVSR [49] unleashes the potential of events in arbitrary spatial VSR, which learns implicit neural representations from both RGB frames and events. EvTexture [31] regards the voxelized events as high-frequency signals to improve the textures in RGB frames, yielding promising VSR reconstruction. As for event-based VFI, TimeLens [68] and TimeLens++ [69] innovatively integrate deformation and synthesis processes. CBMNet [34] utilizes cross-modality information to directly estimate inter-frame motion fields. REFID [66], as a flexible method, exhibits robustness even with blurred reference frames. Notably, these methods achieve upsampling only in the spatial or temporal domain independently, and few studies address a unified approach for both, i.e., event-based STVSR. + +Implicit Neural Representation (INR) in STVSR. As a method of continuous representation of signals, INR is a common means of continuous super-resolution, including image [4, 5, 17, 38, 83] and video [6, 7, 49, 50] tasks. Inspired by LIIF [5], VideoINR [7] pioneers continuous STVSR (C-STVSR) by learning separate spatial and temporal INRs. Building on this, MoTIF [6] introduces forward spatiotemporal INR to enhance motion field estimation, but it retains the paradigm of decoupling video INR across multiple domains, which lacks robustness across varying scales. Similarly, HR-INR [50], a recent event-based C-STVSR method, also adheres to the same decoupling paradigm as VideoINR [7]. Besides, HR-INR conducts holistic event-frame extraction for long-range motion dependencies and requires repeatedly extracting input frames to realize frame temporal super-resolution, introducing redundancy. Instead, we break the inherent paradigm. We devise the EASM that executes event-modulated alignment and bidirectional recurrent compensation for long- + +term motion trajectory modeling. Based on this, our proposed LIVT achieves a unified spatiotemporal INR through the high temporal resolution grids generated from events, thus achieving better generalization with lower weights. + +# 3. Proposed Method + +Given LR and low-frame-rete (LFR) videos and corresponding event streams as inputs, we present an event-driven C-STVSR approach to reconstruct HR and HFR videos at arbitrary spatial and temporal scales. Before elaborating on the components of our methodology, we first provide the details for the event representation in this work. + +# 3.1. Event Representation + +Event cameras asynchronously capture the logarithmic intensity change $\Delta \log (I_{x,y,\tau})$ of each pixel at coordinate $(x,y)$ within the RGB image $I_{x,y,\tau}$ , where $\tau$ denote the timestamp of an event stream $\mathcal{E} = (x,y,\tau,p)$ . $p$ represents the polarity of $\mathcal{E}$ triggered at $(x,y)$ , with 1 and -1 representing the change directions respectively. Compared to conventional RGB images, a raw event stream constituted by large amounts of sparse points, which is capable of continuously capturing intensity change with high temporal resolution. However, processing such an unstructured sparse format in deep networks is non-trivial. Following existing event-based method [66], we represent the event stream into a 3D spatiotemporal voxel grid $\mathcal{V} \in \mathbb{R}^{H \times W \times (M + 1)}$ linearly, where $H \times W$ denotes the spatial size, and then discretize the temporal dimension into $M$ event segments $[E_1,E_2,\dots,E_M]$ . Each segment records the intensity information within a time window of $\mathcal{E}$ , which is often set to be a small constant (see Figure 2). + +# 3.2. Overview + +Let $\mathbf{I}^{LR} = I_0^{LR}, I_1^{LR} \in \mathbb{R}^{H \times W \times 3}$ denote two LR RGB frames and $\mathcal{E}_{0 \to 1}$ denote their corresponding event stream. We propose EvEnhancer that addresses the task of super-resolving $\mathbf{I}^{LR}$ to $(t + 1)$ HR video frames with any spatial scale $s \geq 1$ and temporal scale $t \geq 1$ . The overall framework is illustrated in Figure 2, which consists of an event-adapted synthesis module (EASM) and a local implicit video transformer (LIVT). + +Generally, we first extract initial features on frames and event segments, resulting in RGB frame features $F_0^{LR}$ , $F_1^{LR}$ , and event feature $\mathbf{F}_{0\rightarrow 1}^E = \{F_{0\rightarrow 1}^{E,m}\}_{m = 1}^M$ for $M$ segments. Considering that the latent inter-frame features between RGB frames can be captured in both forward and backward directions, we also reverse the time and polarity of $\nu$ to derive a backward event feature $\mathbf{F}_{1\rightarrow 0}^E$ . Then, EASM models long-term holistic motion trajectories with the guidance of events to generate a spatiotemporal feature sequence $\widehat{\mathbf{F}}$ with high temporal resolution. Finally, LIVT integrates cross-scale spatiotemporal attention with + +![](images/259889dbcadf88b12d6f91d908a1c1664478b4c9c83c986fa5d3b4138bd52733.jpg) +Figure 2. The overall architecture of our EvEnhancer, an event-driven C-STVSR method which consists of an event-adapted synthesis module (EASM), and a local implicit video transformer (LIVT), where EASM contains two steps: (a) event-modulated alignment (EMA), and (b) bidirectional recurrent compensation (BRC). "CS": channel squeeze, "FFN": feed-forward neural network, "Res": residual blocks. + +local implicit function to learn continuous video INR $\mathbf{F}_{st}$ that decodes arbitrary space-time coordinates into RGB values, thus yielding HR video frames $\mathbf{I}_{st}^{HR}$ with HFR. + +# 3.3. Event-Adapted Synthesis Module + +As depicted in Figure 2, EASM contains two steps: 1) Event-modulated alignment incorporates motion cues between input frames with event modulation to acquire latent inter-frame features. Based on the event features and reverse ones, we deploy such alignment in parallel to model the motions forward $(^{+} + )$ and backward $(^{-})$ . 2) Bidirectional recurrent compensation propagates the event stream across time and fuses it with acquired features in both directions to maximize the gathering of temporal information. + +Event-Modulated Alignment (EMA). Given input frame features $F_0^{LR}$ and $F_1^{LR}$ , our goal here is to infer the features of intermediate frames according to $M$ event segment features. Motivated by the effectiveness of the PCD module in video restoration [75, 79], we construct a customized pyramid and event-modulated structure that progressively learns the motion offsets and feature alignment. In the following, we take the forward alignment as an example for clarity. + +Specifically, for the forward alignment, as shown in Figure 2(a), we implement a 3-level pyramid by downsampling the frame and event features $(F_0^{LR}, F_1^{LR}, F_{0\rightarrow 1}^{E,m})$ to 3 different spatial scales $\times 1$ , $\times \frac{1}{2}$ , $\times \frac{1}{4}$ via stride convolutions. + +At each level $l$ , we first calculate an initial motion vector by a convolutional block ("Conv+LeakyReLU (LReLU)") on concatenated frame features, which is then fused with the learned offset $O_{m}^{(l-1)}$ from the $(l-1)$ -th level to obtain a motion vector $MV_{m}^{(l)}$ near the $m$ -th timestamp. Then, we design an event modulation block (EMB) to generate modulated $\widetilde{MV}_{m}^{(l)}$ . The modulation is controlled by the event feature $F_{0\rightarrow 1}^{E,m}$ . Compared to previous PCD-based methods [75, 79, 81], our event-based approach can furnish more deterministic motion cues to facilitate frame synthesis, formulated as + +$$ +\widetilde {M V} _ {m} ^ {(l)} = \mathcal {M} ^ {(l)} \left(M V _ {m} ^ {(l)}, F _ {0 \rightarrow 1} ^ {E, m (l)}\right) \tag {1} +$$ + +where $F_{0\rightarrow 1}^{E,m(l)}$ denote the downsampled event feature at level $l$ . Then, we can learn the offset $O_{m}^{(l)}$ from $MV_{m}^{(l)}$ and $\widetilde{MV}_{m}^{(l)}$ . Following [75], we use the deformable convolutional network (DCN) [92] for feature alignment and interpolation. The deformable alignments at different levels are cascaded and aggregated coarse-to-fine for accuracy improvement via the pyramid workflow. + +Notably, as there are $M$ segment features $\{F_{0\rightarrow 1}^{E,m}\}_{m = 1}^{M}$ by the event-modulated alignment, we can produce $M$ intermediate frame features $\mathbf{F}^{+} = [F_{1}^{LR + },F_{2}^{LR + },\dots ,F_{M}^{LR + }]$ . Similarly, for the backward alignment, we get $\mathbf{F}^{-} =$ + +$[F_1^{LR-}, F_2^{LR-}, \dots, F_M^{LR-}]$ . The two are integrated by channel concatenation with a convolutional block, forming the final features $\mathbf{F} = \{F_{\tau}^{LR}\}$ with the input features $F_0^{LR}$ , $F_1^{LR}$ , where the timestamp $\tau = 0, \frac{1}{M+1}, \dots, \frac{M}{M+1}, 1$ is aligned with the input frames and $M$ event segments. + +Bidirectional Recurrent Compensation (BRC). Given that the information encapsulated within each event segment is confined to a very narrow time window, after the modulated alignment, we propose bidirectional recurrent compensation that iteratively fuses the event and frame features to model the global temporal information inherent in two modalities, as shown in Figure 2(b). In each recurrent iteration, the compensation block receives the event feature $F_{\tau}^{E}$ from the current timestamp $\tau$ , the forward hidden state from $h_{\tau - \Delta \tau}^{f}$ the previous timestamp $\tau - \Delta \tau$ , and the backward hidden state $h_{\tau + \Delta \tau}^{b}$ from the next timestamp $\tau + \Delta \tau$ , and the aligned frame feature $F_{\tau}^{LR} \in \mathbf{F}$ . Here, $\tau \pm \Delta \tau$ denotes the former timestamp and the latter one. Hence, the recurrent fusion process at timestamp $\tau$ is formulated as + +$$ +r _ {\tau} ^ {b}, h _ {\tau} ^ {b} = \mathcal {F} ^ {b} \left(\mathcal {C} ^ {b} \left(F _ {\tau} ^ {L R}, F _ {\tau} ^ {E}\right), h _ {\tau + \Delta \tau} ^ {b}\right), \tag {2} +$$ + +$$ +r _ {\tau} ^ {f}, h _ {\tau} ^ {f} = \mathcal {F} ^ {f} (\mathcal {C} ^ {f} (F _ {\tau} ^ {L R}, F _ {\tau} ^ {E}), h _ {\tau - \Delta \tau} ^ {f}, r _ {\tau} ^ {b}), +$$ + +where $r_{\tau}^{b}$ and $h_{\tau}^{b}$ denotes the output and hidden states respectively in the backward direction, and $r_{\tau}^{f}$ and $h_{\tau}^{f}$ corresponds to the forward. $\mathcal{C}^f$ and $\mathcal{C}^b$ calculate the channel attention on event and frame features during the forward and backward recurrent iterations respectively. By recurrent propagating the event information across time, we can comprehensively exploit the high temporal resolution properties from events to refine the frame features and maximize the gathering of temporal information. Finally, a bidirectional feature fusion is run by the element-wise addition between the input feature fed into BRC and its output bidirectional features to retain valid information. + +# 3.4. Local Implicit Video Transformer + +After the EASM, we obtain the discrete temporal feature sequence $\widehat{\mathbf{F}}$ . The other key in this method is to learn continuous video INR that specifies the feature at the queried space-time coordinate $(\mathcal{T},x,y)$ are transferred to any target spatial scale $s$ and temporal scale $t$ in Figure 2. Previous C-STVSR methods [6, 7] learn spatial and temporal INR separately to realize video INR, which fails to fully capitalize on the spatiotemporal dependencies in the continuous domain. In this work, we propose LIVT which calculates cross-scale 3D attention based on the queried space-time coordinates in the continuous domain over a local sampling grid to learn an accurate unified video INR directly. + +Assuming the temporal length and spatial size of the local grid in LR space are $T^G$ and $H^{G} \times W^{G}$ , respectively. For the full temporal range [0, 1], to generate a frame at the timestamp $\mathcal{T} \in [0,1]$ , we only need to select the $T^G$ nearest + +![](images/9f7f5341d0d6739882ddc280bf9e2a1da01e7f26423948059adadb7647b95ccf.jpg) +Figure 3. The structure of our local implicit video transformer (LIVT), which integrates the 3D local spatiotemporal attention with implicit neural function to learn continuous video INR to reconstruct HR and HFR video frames. + +elements to $\mathcal{T}$ along the time axis. Let the candidate selection of the timestamps $\tau_{i} = 0,\frac{1}{M + 1},\dots,\frac{M}{M + 1},1$ , which in actual align with the $M$ event segments and input frames $\mathbf{I}^{LR}$ . We denote the set of candidate timestamps $\{\tau_i\}$ as $\mathbf{S}_{\tau}\subseteq [0,\frac{1}{M + 1},\dots,\frac{M}{M + 1},1]$ , and the size of the $|\mathbf{S}_{\tau}| = T^G$ which is selected following the rule as + +$$ +\mathbf {S} _ {\tau} = \underset {\mathbf {S} _ {\tau}} {\arg \min } \sum_ {\tau_ {i} \in \mathbf {S} _ {\tau}} | \tau_ {i} - \mathcal {T} |. \tag {3} +$$ + +Then, as shown in Figure 3, with the LR feature sequence $\widehat{\mathbf{F}}$ from EASM, we encode them into three embedding spaces through 3D convolution: key $\{K\}$ , query $\{Q\}$ , and value $\{V\}$ . Notably, we do not calculate the spatiotemporal attention across the overall time range. Based on Eq. 3, we conduct temporal selection on $K$ , $Q$ , and $V$ , which allows the learning for 3D local attention in a subtemporal space near the target timestamp $\mathcal{T}$ . On top of this, we employ 3D trilinear sampling on $Q$ to get the query $q_{\mathcal{T}} \in \mathbb{R}^{s^2 HW \times 1 \times C}$ for a certain frame, where $C$ is the channel number. However, since both $k_{\mathcal{T}}$ and $v_{\mathcal{T}}$ have size $s^2 HW \times T^G H^G W^G \times C$ , it will suffer from expensive computational costs if we directly calculate the attention. To solve this problem, we propose to unfold the sequence along the temporal dimension and calculate cross-scale attention separately on $T^G$ embeddings. Thus, we sample the key $k_{\tau_i}$ and value $v_{\tau_i}$ from $K$ and $V$ respectively using 2D nearest local sampling at each time slice $\tau_i$ . Moreover, we encode and reshape the spatiotemporal relative coordinates $(\delta \tau, \delta x, \delta y) \in (-1, 1)$ from each query point to all pixel points within its local grid, obtaining the spatiotemporal bias $b_{\tau_i} \in \mathbb{R}^{s^2 HW \times H^G W^G \times 3}$ . + +We compute the inner product between $q_{\mathcal{T}}$ and $k_{\tau_i}$ and add it with the $b_{\tau_i}$ to obtain the local attention map at the timestamp $\mathcal{T}$ after a Softmax operation. After the matrix multiplication with $v_{\tau_i}$ , we can derive the output embedding + +Table 1. Quantitative comparisons for in-distribution (In-dist.) spatiotemporal upsampling scales ( $t = 4$ , $f = 8$ ). † indicates the model trained on the same Adobe240 dataset [64] as ours, and ‡ indicates the model fine-tuned on the Adobe240 dataset. **Bold** and **underline** indicate the best and the second-best performance, respectively. Metrics: PSNR (dB) / SSIM. + +
VFI MethodVSR MethodEventsGoPro-CenterGoPro-AverageAdobe240-CenterAdobe240-AverageParams (M)
TimeLens [69]EGVSR [49]✓28.41 / 0.830727.42 / 0.807726.64 / 0.767525.13 / 0.729872.20+2.45
TimeLens [69]EvTexture [31]✓30.50 / 0.878428.51 / 0.847828.80 / 0.833726.12 / 0.778972.20+8.90
CBMNet-L [34]EGVSR [49]✓26.14 / 0.767423.55 / 0.701825.15 / 0.731122.89 / 0.673522.23+2.45
CBMNet-L [34]EvTexture [31]✓27.58 / 0.796623.92 / 0.716326.74 / 0.777723.38 / 0.698622.23+8.90
REFID [66]†EGVSR [49]✓28.74 / 0.836427.97 / 0.817127.28 / 0.784426.55 / 0.763715.91+2.45
REFID [66]†EvTexture [31]✓30.86 / 0.878429.14 / 0.848929.42 / 0.842427.69 / 0.805915.91+8.90
Zooming Slow-Mo [25]†30.69 / 0.8847-30.26 / 0.8821-11.10
TMNet [81]‡30.14 / 0.869228.83 / 0.851429.41 / 0.852428.30 / 0.835412.26
SAFA [25]†31.28 / 0.889430.22 / 0.876130.97 / 0.887830.13 / 0.87824.94
VideoINR-fixed [7]†30.73 / 0.8850-30.21 / 0.8805-11.31
VideoINR [7]†30.26 / 0.879229.41 / 0.866929.92 / 0.874629.27 / 0.865111.31
MoTIF [6]†31.04 / 0.887730.04 / 0.877330.63 / 0.883929.82 / 0.875012.55
HR-INR [50]†✓31.97 / 0.929832.13 / 0.937131.26 / 0.924631.11 / 0.92168.27
EvEnhancer-light✓33.11 / 0.924232.73 / 0.920331.90 / 0.903331.47 / 0.898831.47 / 0.89885.81
EvEnhancer✓33.52 / 0.929533.30 / 0.927932.43 / 0.912932.18 / 0.911632.18 / 0.91166.55
+ +$\widetilde{z}_{\tau_i}$ . This procedure is described as + +$$ +\widetilde {z} _ {\tau_ {i}} = \operatorname {S o f t m a x} \left(\frac {q _ {\mathcal {T}} k _ {\tau_ {i}} ^ {\mathrm {T}}}{\sqrt {C}} + b _ {\tau_ {i}}\right) \otimes v _ {\tau_ {i}}. \tag {4} +$$ + +We concatenate the output embeddings $\widetilde{z}_{\tau_i}$ along the channel dimension to construct the complete continuous video INR $\widetilde{Z}_{\mathcal{T}}$ to yield the features corresponding to continuous space-time coordinates. Following [4, 5, 7], we decode the features into RGB values using a 5-layer MLP with the previous query $q_{\mathcal{T}}$ and cell decoding to reconstruct the target RGB frame $\hat{I}_{\mathcal{T}}^{HR}$ . + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. In our experiments, we train all the models on the Adobe240 dataset [64] which contains 133 video sequences, where 100 for training, 16 for validation, and 17 for testing. We evaluate the performance of our models on both synthetic Adobe240 [64] and GoPro [51] datasets and real-world BS-ERGB [69] and ALPIX-VSR [49] datasets. For the synthetic events, we use the event simulation method vid2e [11] to generate simulated event data. Given a spatial scale $s$ and a temporal scale $t$ , we select consecutive $(t + 1)$ frames as a clip and select the first and the last frames for $s: 1$ bicubic downsampling as the LR and LFR inputs fed into the network to generate HR and HFR frame sequences, and the original $(t + 1)$ frames are groundtruth (GT). Following [31], the event voxels are bicubic-downsampled as the LR event inputs, thus ensuring the spatial size is aligned with the LR frames. + +Implementation Details. Referring VideoINR [7], the training for our model comprises two stages. In the first stage, we set $s = 4$ and $t = 8$ , training for 450K iterations on Adobe240 [64]. In the second stage, we uni- + +formly sample $s$ in [1, 4] and fine-tune for another 150K iterations on the same dataset. We use the Adam optimizer [35] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ , and apply cosine annealing to decay the learning rate from $1 \times 10^{-4}$ to $1 \times 10^{-7}$ over 150K iterations. For data augmentation, we randomly crop $32 \times 32$ spatial regions from downsampled images and event segments with random rotations and horizontal flips. The model is optimized with Charbonnier loss [36] using $\epsilon^{2} = 1 \times 10^{-6}$ . In this work, we implement two types of variants, the full model EvEnhancer and a lightweight version EvEnhancer-light. To measure our models, we calculate the PSNR (dB) and SSIM on the Y channel. + +# 4.2. Comparison with State-of-the-Art Methods + +We comprehensively consider frame-based and event-based one-stage or two-stage state-of-the-art methods for comparison: 1) two-stage event-based STVSR methods consisting of event-based VFI (TimeLens [68], REFID [66], CBMNet-L [34]) and event-based VSR (EGVSR [49], EvTexture [31]); 2) one-stage frame-based F-STVSR methods (Zooming Slow-Mo[79], TMNet [81], SAFA [25]); 3) one-stage frame-based C-STVSR methods (VideoINR [7], MoTIF [6]); 4) one-stage event-based C-STVSR methods (HR-INR [50]). For fair comparisons, we use the models of most VFI methods and all STVSR methods trained or fine-tuned on the same Adobe240 dataset [64] as ours. Notably, due to the limitation that the HR-INR source code is not yet available, the results of HR-INR are derived from its original paper. + +Quantitative Comparison. Table 1 reports the results for the F-STVSR task, where the spatial and temporal scales are the same as the training distribution (In-Dist.). "Center" represents the average performance of the reconstructed left, right, and center frames. "Average" represents the average performance of all reconstructed frames. Notably, + +Table 2. Quantitative comparisons for out-of-distribution (OOD) spatiotemporal upsampling scales on the GoPro dataset [51]. + +
Temporal ScaleSpatial ScaleTMNet [81]‡SAFA [25]†VideoINR [7]†MoTIF [6]†HR-INR [50]†EvEnhancer -lightEvEnhancer
t = 6s = 430.49 / 0.886131.68 / 0.906830.78 / 0.895431.56 / 0.9064-33.10 / 0.926633.41 / 0.9300
t = 6s = 6--25.56 / 0.767129.36 / 0.8505-29.84 / 0.860330.12 / 0.8675
t = 6s = 12--24.02 / 0.690025.81 / 0.7330-25.41 / 0.726125.50 / 0.7323
t = 12s = 426.38 / 0.793127.95 / 0.824927.32 / 0.814127.77 / 0.823028.87 / 0.885431.17 / 0.896932.07 / 0.9116
t = 12s = 6--24.68 / 0.735826.78 / 0.790827.14 / 0.817328.94 / 0.842529.54 / 0.8579
t = 12s = 12--23.70 / 0.683024.72 / 0.7108-25.25 / 0.724825.45 / 0.7345
t = 16s = 424.72 / 0.752626.38 / 0.783225.81 / 0.773925.98 / 0.775827.29 / 0.855630.15 / 0.876230.93 / 0.8918
t = 16s = 6--23.86 / 0.712325.34 / 0.752726.09 / 0.795428.34 / 0.828328.86 / 0.8434
t = 16s = 12--22.88 / 0.665923.88 / 0.6923-25.10 / 0.721725.25 / 0.7310
t = 6s = 1--32.34 / 0.954534.77 / 0.969638.53 / 0.973537.68 / 0.966738.80 / 0.9714
+ +Table 3. Quantitative comparisons for different temporal upsampling scales $t$ (fixed $s = 4$ ) on the BS-ERGB dataset [69]. + +
MethodEventst=4t=6t=8
TimeLens[69]+EGVSR[49]✓22.98/0.676222.01/0.659221.33/0.6467
TimeLens[69]+EvTexture[31]✓23.58/0.708822.48/0.688721.74/0.6746
CBMNet-L[34]+EGVSR[49]✓22.74/0.685621.93/0.670521.29/0.6657
CBMNet-L[34]+EvTexture[31]✓23.40/0.719922.47/0.702021.78/0.6962
REFID[66]†+EGVSR[49]✓--22.76/0.6478
REFID[66]†+EvTexture[31]✓--22.77/0.6527
SAFA[25]†24.32/0.731423.23/0.712022.41/0.6954
VideoINR[7]†24.06/0.729023.00/0.709822.23/0.6945
MoTIF[6]†24.21/0.729623.15/0.709922.37/0.6939
EvEnhancer-light✓25.32/0.731124.54/0.718424.16/0.7100
EvEnhancer✓25.44/0.733824.70/0.721524.29/0.7138
+ +Table 4. $2\times$ VSR performance on the ALPIX-VSR dataset [49]. + +
MethodVideoINR [7]MoTIF [6]EGVSR [49]EvEnhancer
PSNR / SSIM32.53 / 0.938338.61 / 0.963639.31 / 0.963540.84 / 0.9786
+ +VideoINR-fixed is trained specifically for single-frame interpolation. We can see that both our EvEnhancer and EvEnhancer-light significantly outperform existing methods in all settings with much fewer model parameters, which achieves more than 1 dB PSNR improvements. Then, we also evaluate our models on the C-STVSR task, where spatial and temporal scales are out of the training distribution. The results are illustrated in Table 2. It can be observed that both our models exceed almost all existing methods at various spatial and temporal scales, especially in the challenging large temporal scale, achieving superior effectiveness and generalization ability. In particular, our EvEnhancer-light and EvEnhancer use lighter-weight models and perform better than HR-INR [50] at most temporal and spatial scales, especially at OOD scales. + +We further validate the effectiveness of our method on the real-world BS-ERGB dataset under the setting of fixed spatial scale $s = 4$ but varying temporal scales $t$ . The comparisons in Table 3 demonstrate that our method yields the best quantitative performance under all conditions. Compared to cascaded event-based VFI and VSR, our EvEnhancer simultaneously learns temporal interpolation and spatial VSR, thus achieving superior performance. Due to + +Table 5. TFLOPs comparisons of C-STVSR models at different upsampling temporal scales $t$ (fixed $s = 4$ ). * indicates the repeated iterations during inference on the NVIDIA GeForce RTX 3090 GPU due to their large computations. + +
Temporal ScaleVideoINR [7] (11.31M)MoTIF [6] (12.55M)EvEnhancer -light (5.81M)EvEnhancer (6.55M)
t = 22.0112.0431.8763.398
t = 42.3954.086*2.4724.647
t = 62.7796.129*3.0675.885
t = 83.1636.129*3.6637.129
t = 123.93210.22*4.8549.617
t = 164.70012.26*6.04412.10
+ +the high temporal resolution of events, equipped by EASM, our EvEnhancer enables holistic motion trajectory modeling, thus outperforming frame-based methods. Benefiting from the cross-scale attention in LIVT, it also performs preferably at OOD scales. To validate the effectiveness of our EvEnhancer for real-world LR events, we conduct experiments on the real-world ALPIX-VSR dataset [49] (image resolution: $3264 \times 2448$ , event resolution: $1632 \times 1224$ ) for $2 \times$ VSR. As shown in Table 4, EvEnhancer significantly surpasses all methods without additional fine-tuning. + +Model Complexity. To validate the model efficiency of the proposed methods, besides the model parameter comparison in Table 1, we compare the TFLOPs with existing C-STVSR methods in Table 5, which are calculated on the upsampling of a clip with a resolution of $180 \times 320$ by a spital upsampling scale $s = 4$ at different temporal upsampling scales $t$ . VideoINR shows the fewest TFLOPs but more parameters and the worst performance (see Table 1). At the low temporal scale ( $\leq 4$ ), MoTIF is computationally efficient than ours. However, as the temporal scale increases, its computations increase and exceed both our models. + +Qualitative Comparison. With the qualitative results in Figure 4, compared to VideoINR [7] and MoTIF [6], our models reconstruct the faithful HR frames with clearer textures. By computing the difference maps between generated frames and corresponding ground truth frames, we can see that our methods are more excellent at tackling non-linear motions and recovering more details. + +![](images/f65700759f696021fdab42b743e8e4cd5bda3f73b27412d45eebaad6c7a14308.jpg) +Figure 4. Qualitative comparison for In-dist. scale $(t = 8, s = 4)$ on the GoPro dataset [51]. We compare the normalized absolute difference maps (yellow boxes) for the same regions in each frame as in the GT frames. + +Table 6. Ablations on the designs of event-adapted synthesis module (EASM). "fwd." and "bwd." denote the forward and backward directions, respectively. Metrics: PSNR (dB) / SSIM. + +
EMABRCBFIn-Dist.OOD
fwd. & bwd.32.50 / 0.916628.86 / 0.8407
fwd. & bwd.✓33.13 / 0.925229.42 / 0.8539
fwd.✓✓33.20 / 0.926229.47 / 0.8556
fwd. & bwd.✓✓33.30 / 0.927929.54 / 0.8579
+ +# 4.3. Ablation Studies + +We conduct ablation studies by keeping the same training strategy on the Adobe240 dataset [64], and testing on the GoPro dataset [51] at both In-Dist. $(t = 8, s = 4)$ and OOD $(t = 12, s = 6)$ spatiotemporal scales to validate the effectiveness of our methodology design. + +Event-Adapted Synthesis Module. Table 6 presents the ablations of EASM, which involves the event-modulated alignment (EMA), bidirectional recurrent compensation (BRC), and bidirectional feature fusion (BF). The model with only EMA achieves the worst performance. When we introduce BRC in the model, one can see the performance improves. Besides, with the integration BRC, BF, and EMA, the model achieves the best at both In-Dist. and OOD scales. We can also observe that the model with only a single forward alignment can suffer from performance degradation. The results demonstrate the effects of EASM. + +Local Implicit Video Transformer. Table 7 investigates the effect of LIVT. We first implement the model with a 2D decoupling strategy as videoINR [7] but introduce the local attention, concatenate cell decoding, and query, one can see that the model achieves better performance than the model + +Table 7. Ablations on the designs of local implicit video transformer (LIVT). "LA": local attention, "CE": concatenate cell decoding, "PQ": concatenate the previous query. Metrics: PSNR (dB) / SSIM. + +
INR ModeLACEPQIn-dist.OOD
2D decoupling✓✓✓33.26 / 0.927229.26 / 0.8525
3D unification27.32 / 0.797525.03 / 0.7032
3D unification✓33.27 / 0.927629.47 / 0.8565
3D unification✓✓33.30 / 0.928229.49 / 0.8570
3D unification✓✓✓33.30 / 0.927929.54 / 0.8579
+ +without any components. When we replace the 2D decoupling with our unified 3D video INR, the model produces results with higher PSNR and SSIM, especially at OOD scales. In addition, the concatenation of the cell [5] and previous query [7] brings a larger gain at OOD scales than In-Dist., which validates their effects. + +# 5. Conclusion + +In this paper, we introduce EvEnhancer, a novel approach that amalgamates the unique advantages of event streams to enhance the effectiveness, efficiency, and generalizability of C-STVSR. Our EvEnhancer is underpinned by two critical components: 1) the EASM capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories; and 2) the LIVT, which integrates a local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations. Experiments indicate that EvEnhancer significantly surpasses current state-of-the-art methods. + +# Acknowledgment + +This work was partly supported by the National Natural Science Foundation of China (62302141, 62331003, 62120106009), the Joint Funds of the National Natural Science Foundation of China (U23A20314), the Fundamental Research Funds for the Central Universities (JZ2024HGTB0255), Beijing Natural Science Foundation (L223022), and the Natural Science Foundation of Hebei Province (F2024105029). + +# References + +[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2 +[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2 +[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1 +[4] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 3, 6 +[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 6, 8 +[6] Yi-Hsin Chen, Si-Cun Chen, Yen-Yu Lin, and Wen-Hsiao Peng. Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video superresolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23131-23141, 2023. 1, 2, 3, 5, 6, 7, 8 +[7] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2047-2057, 2022. 1, 2, 3, 5, 6, 7, 8 +[8] Xianhang Cheng and Zhenzhong Chen. Video frame interpolation via deformable separable convolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10607-10614, 2020. 2 +[9] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at + +once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 1 +[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 2 +[11] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In CVPR, pages 3586-3595, 2020. 6 +[12] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 3 +[13] Jin Han, Yixin Yang, Chu Zhou, Chao Xu, and Boxin Shi. Evintsr-net: Event guided multiple latent frames reconstruction and super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4882-4891, 2021. 2, 3 +[14] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3897-3906, 2019. 1 +[15] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Space-time-aware multi-resolution video enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2859-2868, 2020. 2 +[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 2 +[17] Zongyao He and Zhi Jin. Latent modulated function for computational optimal continuous image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26026-26035, 2024. 3 +[18] Qiqi Hou, Abhijay Ghildyal, and Feng Liu. A perceptual quality metric for video frame interpolation. In European Conference on Computer Vision, pages 234-253. Springer, 2022. 1 +[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 2 +[20] Mengshun Hu, Kui Jiang, Zheng Wang, Xiang Bai, and Ruimin Hu. Cycmunet+: Cycle-projected mutual learning + +for spatial-temporal video super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[21] Ping Hu, Simon Niklaus, Lu Zhang, Stan Sclaroff, and Kate Saenko. Video frame interpolation with many-to-many splating and spatial selective refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[22] Cong Huang, Jiahao Li, Lei Chu, Dong Liu, and Yan Lu. Arbitrary-scale video super-resolution guided by dynamic context. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2302, 2024. 2 +[23] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 40(4):1015-1028, 2017. 2 +[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2 +[25] Zhewei Huang, Ailin Huang, Xiaotao Hu, Chen Hu, Jun Xu, and Shuchang Zhou. Scale-adaptive feature aggregation for efficient space-time video super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4228-4239, 2024. 3, 6, 7 +[26] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 2 +[27] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 2 +[28] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. arXiv preprint arXiv:2008.05765, 2020. 2 +[29] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2 +[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2, 3 +[31] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. EvTexture: Event-driven Texture Enhancement for Video Super-Resolution. In Proceedings of the 41st International Conference on Machine Learning, pages 22817–22839. PMLR, 2024. 3, 6, 7 +[32] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 2 + +[33] Soo Ye Kim, Jihyong Oh, and Munchurl Kim. Fisr: Deep joint frame interpolation and super-resolution with a multiscale temporal loss. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11278-11286, 2020. 2 +[34] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 3, 6, 7 +[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[36] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 6 +[37] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5316-5325, 2020. 1 +[38] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1929-1938, 2022. 3 +[39] Feng Li, Huihui Bai, and Yao Zhao. Learning a deep dual attention network for video super-resolution. IEEE transactions on image processing, 29:4474-4488, 2020. 1, 2 +[40] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 1 +[41] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10522-10531, 2019. 1 +[42] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 2 +[43] Zekun Li, Hongying Liu, Fanhua Shang, Yuanyuan Liu, Liang Wan, and Wei Feng. Savsr: Arbitrary-scale video super-resolution via a learned scale-adaptive network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3288-3296, 2024. 2 +[44] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2 +[45] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. + +Learning event-driven video deblurring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 695-710. Springer, 2020. 2 +[46] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2 +[47] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19125-19134, 2024. 2 +[48] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 2 +[49] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1557-1567, 2023. 2, 3, 6, 7 +[50] Yunfan Lu, Zipeng Wang, Yusheng Wang, and Hui Xiong. Hr-inr: Continuous space-time video super-resolution via event camera. arXiv preprint arXiv:2405.13389, 2024. 2, 3, 6, 7 +[51] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 6, 7, 8 +[52] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 2 +[53] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE international conference on computer vision, pages 261-270, 2017. 1, 2 +[54] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive convolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 670-679, 2017. 2 +[55] Simon Niklaus, Ping Hu, and Jiawen Chen. Splitting-based synthesis for video frame interpolation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 713-723, 2023. 2 +[56] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14539-14548, 2021. 2 +[57] Zhongwei Qiu, Huan Yang, Jianlong Fu, Daochang Liu, Chang Xu, and Dongmei Fu. Learning degradation-robust spatiotemporal frequency-transformer for video superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 + +[58] Nasrin Rahimi and A Murat Tekalp. Spatio-temporal perception-distortion trade-off in learned video sr. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1400-1404. IEEE, 2023. 1 +[59] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF international conference on computer vision, pages 892-900, 2019. 1 +[60] Oded Shahar, Alon Faktor, and Michal Irani. Space-time super-resolution from a single video. IEEE, 2011. 3 +[61] Wei Shang, Dongwei Ren, Wanying Zhang, Yuming Fang, Wangmeng Zuo, and Kede Ma. Arbitrary-scale video superresolution with structural and textural priors. In European Conference on Computer Vision, pages 73-90. Springer, 2025. 2 +[62] Eli Shechtman, Yaron Caspi, and Michal Irani. Space-time super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(4):531-545, 2005. 3 +[63] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 2 +[64] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 1, 2, 6, 8 +[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2 +[66] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 3, 6, 7 +[67] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2 +[68] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 2, 3, 6 +[69] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1, 3, 6, 7 + +[70] Hai Wang, Wenming Yang, Qingmin Liao, and Jie Zhou. Bi- rstu: Bidirectional recurrent upsampling network for spacetime video super-resolution. IEEE Transactions on Multimedia, 2022. 2 +[71] Hai Wang, Xiaoyu Xiang, Yapeng Tian, Wenming Yang, and Qingmin Liao. Stdan: deformable attention network for space-time video super-resolution. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through hr optical flow estimation. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 514–529. Springer, 2019. 2 +[73] Li Wang, Dong Li, Yousong Zhu, Lu Tian, and Yi Shan. Dual super-resolution learning for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3774-3783, 2020. 1 +[74] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2 +[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1, 2, 4 +[76] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 1 +[77] Guangyang Wu, Xin Tao, Changlin Li, Wenyi Wang, Xiao-hong Liu, and Qingqing Zheng. Perception-oriented video frame interpolation via asymmetric blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2753-2762, 2024. 1 +[78] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In European Conference on Computer Vision, pages 267-283. Springer, 2022. 3 +[79] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3370-3379, 2020. 2, 3, 4, 6 +[80] Chenming Xu, Meiqin Liu, Chao Yao, Weisi Lin, and Yao Zhao. Ibvc: Interpolation-driven b-frame video compression. Pattern Recognition, 153:110465, 2024. 1 +[81] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6388-6397, 2021. 3, 4, 6, 7 + +[82] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2 +[83] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 3 +[84] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 2 +[85] Jun-Sang Yoo, Hongjae Lee, and Seung-Won Jung. Video object segmentation-aware video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12322-12333, 2023. 1 +[86] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17834-17843, 2022. 2 +[87] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S Ren. Training weakly supervised video frame interpolation with events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14589-14598, 2021. 2 +[88] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 2 +[89] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8799-8808, 2019. 1 +[90] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 2 +[91] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 2 +[92] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 4 \ No newline at end of file diff --git a/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/images.zip b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..606c37f7b9df8a31f7d170b698a4a9bf6a21bfe6 --- /dev/null +++ b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0411184d4cd00f55e58f12cb1dac63d060e71d2ca1c72127412329d5cf1e0337 +size 886828 diff --git a/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/layout.json b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..870070438b771df11934a70f8d5f2a9d139a059f --- /dev/null +++ b/2025/EvEnhancer_ Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events/layout.json @@ -0,0 +1,11785 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "text", + "content": "EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "spans": [ + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "text", + "content": "Shuoyan Wei" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "text", + "content": " Feng Li" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "text", + "content": " Shengeng Tang" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "text", + "content": " Yao Zhao" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "text", + "content": " Huihui Bai" + }, + { + "bbox": [ + 118, + 160, + 495, + 175 + ], + "type": "inline_equation", + "content": "^{4,1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 175, + 454, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 175, + 454, + 190 + ], + "spans": [ + { + "bbox": [ + 157, + 175, + 454, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 175, + 454, + 190 + ], + "type": "text", + "content": "Institute of Information Science, Beijing Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 189, + 510, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 189, + 510, + 204 + ], + "spans": [ + { + "bbox": [ + 100, + 189, + 510, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 189, + 510, + 204 + ], + "type": "text", + "content": "Beijing Key Laboratory of Advanced Information Science and Network Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 203, + 497, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 203, + 497, + 218 + ], + "spans": [ + { + "bbox": [ + 113, + 203, + 497, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 203, + 497, + 218 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Hefei University of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 161, + 217, + 450, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 217, + 450, + 232 + ], + "spans": [ + { + "bbox": [ + 161, + 217, + 450, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 161, + 217, + 450, + 232 + ], + "type": "text", + "content": "Tangshan Research Institute of Beijing Jiaotong University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 233, + 488, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 233, + 488, + 245 + ], + "spans": [ + { + "bbox": [ + 122, + 233, + 488, + 245 + ], + "type": "text", + "content": "{shuoyan.wei, yzhao, hhbai}@bjtu.edu.cn, {fengli, tangsg}@hfut.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 298, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 296, + 597 + ], + "type": "text", + "content": "Continuous space-time video super-resolution (C-STVSR) endeavors to upscale videos simultaneously at arbitrary spatial and temporal scales, which has recently garnered increasing interest. However, prevailing methods struggle to yield satisfactory videos at out-of-distribution spatial and temporal scales. On the other hand, event streams characterized by high temporal resolution and high dynamic range, exhibit compelling promise in vision tasks. This paper presents EvEnhancer, an innovative approach that marries the unique advantages of event streams to elevate effectiveness, efficiency, and generalizability for C-STVSR. Our approach hinges on two pivotal components: 1) Event-adapted synthesis capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories, enabling the adaptive interpolation and fusion of informative spatiotemporal features; 2) Local implicit video transformer integrates local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations utilized to generate plausible videos at arbitrary resolutions and frame rates. Experiments show that EvEnhancer achieves superiority on synthetic and real-world datasets and preferable generalizability on out-of-distribution scales against state-of-the-art methods. Code is available at https://github.com/W-Shuoyan/EvEnhancer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "content": "Video super-resolution (VSR) is a fundamental low-level vision task, aiming to recover high-resolution (HR) videos from low-resolution (LR) inputs. Nevertheless, video resources are mostly archived with reduced spatial resolution" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 272, + 436, + 370 + ], + "blocks": [ + { + "bbox": [ + 316, + 272, + 436, + 370 + ], + "lines": [ + { + "bbox": [ + 316, + 272, + 436, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 436, + 370 + ], + "type": "image", + "image_path": "efa3564cdc2fd103e4f5882e81c7e5a60e155fa4a1eaf20d70c4e9fa238fc66d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 369, + 388, + 375 + ], + "lines": [ + { + "bbox": [ + 365, + 369, + 388, + 375 + ], + "spans": [ + { + "bbox": [ + 365, + 369, + 388, + 375 + ], + "type": "text", + "content": "Adobe240" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 369, + 456, + 375 + ], + "lines": [ + { + "bbox": [ + 410, + 369, + 456, + 375 + ], + "spans": [ + { + "bbox": [ + 410, + 369, + 456, + 375 + ], + "type": "text", + "content": "VideoINR (11.31M)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 367, + 375, + 380, + 381 + ], + "blocks": [ + { + "bbox": [ + 367, + 375, + 380, + 381 + ], + "lines": [ + { + "bbox": [ + 367, + 375, + 380, + 381 + ], + "spans": [ + { + "bbox": [ + 367, + 375, + 380, + 381 + ], + "type": "image", + "image_path": "2c8f16765429f5d249a256317d9d7e04a55a297f4267a652581fca79134589b3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 375, + 388, + 381 + ], + "lines": [ + { + "bbox": [ + 380, + 375, + 388, + 381 + ], + "spans": [ + { + "bbox": [ + 380, + 375, + 388, + 381 + ], + "type": "text", + "content": "age" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 384, + 395, + 398 + ], + "lines": [ + { + "bbox": [ + 317, + 384, + 395, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 395, + 398 + ], + "type": "text", + "content": "-Center: reference & center frames\n-Average: all frames" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 401, + 369, + 410, + 374 + ], + "blocks": [ + { + "bbox": [ + 401, + 369, + 410, + 374 + ], + "lines": [ + { + "bbox": [ + 401, + 369, + 410, + 374 + ], + "spans": [ + { + "bbox": [ + 401, + 369, + 410, + 374 + ], + "type": "image", + "image_path": "f629148fd724f95f0bbeae5a269678787b1e6fa57fb4a725ed2ac179215fa274.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 375, + 449, + 383 + ], + "lines": [ + { + "bbox": [ + 411, + 375, + 449, + 383 + ], + "spans": [ + { + "bbox": [ + 411, + 375, + 449, + 383 + ], + "type": "text", + "content": "MoTIF (12.55M)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 384, + 473, + 391 + ], + "lines": [ + { + "bbox": [ + 411, + 384, + 473, + 391 + ], + "spans": [ + { + "bbox": [ + 411, + 384, + 473, + 391 + ], + "type": "text", + "content": "EvEnhancer-light (5.81M)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 439, + 272, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 439, + 272, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 439, + 272, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 439, + 272, + 553, + 369 + ], + "type": "image", + "image_path": "319189ecb6fc2ee177ec7a9473192ffdbf61c5d264c280d84f05721b7ce36aa6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 384, + 552, + 391 + ], + "lines": [ + { + "bbox": [ + 501, + 384, + 552, + 391 + ], + "spans": [ + { + "bbox": [ + 501, + 384, + 552, + 391 + ], + "type": "text", + "content": "t: temporal scale factor" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 316, + 411, + 394, + 456 + ], + "blocks": [ + { + "bbox": [ + 328, + 401, + 426, + 410 + ], + "lines": [ + { + "bbox": [ + 328, + 401, + 426, + 410 + ], + "spans": [ + { + "bbox": [ + 328, + 401, + 426, + 410 + ], + "type": "text", + "content": "(a) Different datasets (In-dist.)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 411, + 394, + 456 + ], + "lines": [ + { + "bbox": [ + 316, + 411, + 394, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 394, + 456 + ], + "type": "image", + "image_path": "d5dadf5fae3067a50d3bd228910ee8088b46192682332860bad8b94b411458e9.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 316, + 464, + 394, + 510 + ], + "blocks": [ + { + "bbox": [ + 320, + 458, + 389, + 465 + ], + "lines": [ + { + "bbox": [ + 320, + 458, + 389, + 465 + ], + "spans": [ + { + "bbox": [ + 320, + 458, + 389, + 465 + ], + "type": "text", + "content": "Overlapping LR Frames" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 464, + 394, + 510 + ], + "lines": [ + { + "bbox": [ + 316, + 464, + 394, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 394, + 510 + ], + "type": "image", + "image_path": "778cc400e320508dd656f461854df8a26c9a593a14839b02c021dc0bef0ba73b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 513, + 370, + 520 + ], + "lines": [ + { + "bbox": [ + 339, + 513, + 370, + 520 + ], + "spans": [ + { + "bbox": [ + 339, + 513, + 370, + 520 + ], + "type": "text", + "content": "LR Events" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 522, + 481, + 531 + ], + "lines": [ + { + "bbox": [ + 388, + 522, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 388, + 522, + 481, + 531 + ], + "type": "text", + "content": "(c) Visualization comparison" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 536, + 555, + 613 + ], + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 613 + ], + "type": "text", + "content": "Figure 1. Performance comparison of different C-STVSR methods. (a) PSNR (dB) comparison for in-distribution (In-dist.) spatiotemporal upsampling scale (temporal scale " + }, + { + "bbox": [ + 313, + 536, + 555, + 613 + ], + "type": "inline_equation", + "content": "t = 8" + }, + { + "bbox": [ + 313, + 536, + 555, + 613 + ], + "type": "text", + "content": ", spatial scale " + }, + { + "bbox": [ + 313, + 536, + 555, + 613 + ], + "type": "inline_equation", + "content": "s = 4" + }, + { + "bbox": [ + 313, + 536, + 555, + 613 + ], + "type": "text", + "content": ") on the different datasets including GoPro [51], Adobe240 [64] and BS-ERGB [69]. (b) PSNR (dB) comparison for different spatiotemporal upsampling scales on GoPro. (c) Visualization comparison for In-dist. scale on GoPro." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 398, + 411, + 474, + 458 + ], + "blocks": [ + { + "bbox": [ + 411, + 391, + 460, + 398 + ], + "lines": [ + { + "bbox": [ + 411, + 391, + 460, + 398 + ], + "spans": [ + { + "bbox": [ + 411, + 391, + 460, + 398 + ], + "type": "text", + "content": "EvEnhancer (6.55M)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 401, + 541, + 410 + ], + "lines": [ + { + "bbox": [ + 440, + 401, + 541, + 410 + ], + "spans": [ + { + "bbox": [ + 440, + 401, + 541, + 410 + ], + "type": "text", + "content": "(b) Different upsampling scales" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 398, + 411, + 474, + 458 + ], + "lines": [ + { + "bbox": [ + 398, + 411, + 474, + 458 + ], + "spans": [ + { + "bbox": [ + 398, + 411, + 474, + 458 + ], + "type": "image", + "image_path": "44be04deb02d1728963b276f557fe97c0d4c6570c3a982121b756c9820c9768a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 458, + 454, + 465 + ], + "lines": [ + { + "bbox": [ + 417, + 458, + 454, + 465 + ], + "spans": [ + { + "bbox": [ + 417, + 458, + 454, + 465 + ], + "type": "text", + "content": "VideoINR[7]" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 398, + 465, + 474, + 503 + ], + "blocks": [ + { + "bbox": [ + 398, + 465, + 474, + 503 + ], + "lines": [ + { + "bbox": [ + 398, + 465, + 474, + 503 + ], + "spans": [ + { + "bbox": [ + 398, + 465, + 474, + 503 + ], + "type": "image", + "image_path": "7d391d6829b7b4e017a931cc7043e40e9c81fcde632ed0c4f28072a229f16f93.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 503, + 460, + 510 + ], + "lines": [ + { + "bbox": [ + 413, + 503, + 460, + 510 + ], + "spans": [ + { + "bbox": [ + 413, + 503, + 460, + 510 + ], + "type": "text", + "content": "32.90 dB/0.9337" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 513, + 466, + 521 + ], + "lines": [ + { + "bbox": [ + 406, + 513, + 466, + 521 + ], + "spans": [ + { + "bbox": [ + 406, + 513, + 466, + 521 + ], + "type": "text", + "content": "EvEnhancer (Ours)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 476, + 411, + 553, + 458 + ], + "blocks": [ + { + "bbox": [ + 501, + 391, + 552, + 398 + ], + "lines": [ + { + "bbox": [ + 501, + 391, + 552, + 398 + ], + "spans": [ + { + "bbox": [ + 501, + 391, + 552, + 398 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 501, + 391, + 552, + 398 + ], + "type": "text", + "content": " : spatial scale factor" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 411, + 553, + 458 + ], + "lines": [ + { + "bbox": [ + 476, + 411, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 476, + 411, + 553, + 458 + ], + "type": "image", + "image_path": "633db7ff7f2780199452f34c4cce262f02a4ef1b33b448a9de60de467e14bc61.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 458, + 529, + 465 + ], + "lines": [ + { + "bbox": [ + 500, + 458, + 529, + 465 + ], + "spans": [ + { + "bbox": [ + 500, + 458, + 529, + 465 + ], + "type": "text", + "content": "MoTIF[6]" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 476, + 465, + 553, + 510 + ], + "blocks": [ + { + "bbox": [ + 476, + 465, + 553, + 510 + ], + "lines": [ + { + "bbox": [ + 476, + 465, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 476, + 465, + 553, + 510 + ], + "type": "image", + "image_path": "bfb82ef11da97bdbc69fa22b73a658d58134ad675821fdc0fce78ee556420959.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 513, + 520, + 519 + ], + "lines": [ + { + "bbox": [ + 509, + 513, + 520, + 519 + ], + "spans": [ + { + "bbox": [ + 509, + 513, + 520, + 519 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 627, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 555, + 676 + ], + "type": "text", + "content": "and constrained frame rates (i.e., temporal resolution) practically, necessitating the conversion to HR and high-frame-rate (HFR) ones for various applications [40, 73, 76, 80, 85, 89] and user experiences [18, 58, 77]." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "To address this problem, a naive approach is to cascade separate VSR [2, 3, 14, 39, 41, 75] and video frame interpolation (VFI) [9, 37, 53, 59, 82] or in turn, which over-" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17755" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "looks the correlations between two sub-tasks, leading to suboptimal exploitation of spatiotemporal information in videos. Space-time video super-resolution (STVSR) [15, 15, 19, 20, 33, 70, 71, 79] integrates VSR and VFI into a unified framework that super-resolves the videos with intermediate frame interpolation via one-stage process. Both the above approaches are limited to fixed discrete space and time magnifications, cannot provide arbitrary controls. Some researchers study continuous VFI [1, 24, 29, 88] or continuous VSR [22, 43, 49, 61] methods, where the former is capable of variable frames interpolation while the latter enables arbitrary spatial upsampling including non-integer scales. However, these methods operate in isolation, failing to simultaneously ensure spatial and temporal flexibility." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 242, + 295, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 295, + 494 + ], + "type": "text", + "content": "To achieve this, research on continuous STVSR (C-STVSR) [6, 7, 50] algorithms has emerged, which typically learn video implicit neural representation (INR) to decode 3D spatiotemporal coordinates into RGB values, thus generating HR while HFR videos arbitrarily. These methods mainly decouple the learning for video INR into spatial INR and temporal INR to predict continuous spatial and temporal feature domains separately. This fails to fully capitalize on the spatiotemporal dependencies in the continuous domain, thus hindering the reconstruction performance, particularly at scales falling outside the training distribution. Besides, existing methods [6, 7] approximate intermediate motion fields between consecutive frames captured from conventional frame-based cameras, which is insufficient to infer accurate inter-frame temporal features when motions are vast or non-linear. Recently, event cameras famous as bio-inspired asynchronous sensors with high temporal resolution and high dynamic range, providing a promising solution to fulfill the inter-frame motion fields in video reconstruction [13, 16, 30, 45, 49, 68, 87]. Nonetheless, the potential of events in C-STVSR remains largely unexplored." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 496, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 663 + ], + "type": "text", + "content": "In light of the above discussion, we propose an innovative event-driven approach EvEnhancer, which adeptly integrates the spatiotemporal synergies between frames and events to learn continuous video representations for effective and generalized C-STVSR. First, we present an event-adapted synthesis module (EASM) that executes event-modulated alignment and bidirectional recurrent compensation to explore the event information for long-term holistic motion trajectory modeling. It incorporates motion cues between input frames with event-based modulation forward and backward to acquire latent inter-frame features. We then recurrently propagate the event stream across time and fuse it with aligned frame features in both directions to maximize the gathering of temporal information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "In contrast to the decoupling strategies [6, 7, 50] for video INR, we propose a local implicit video transformer (LIVT) that integrates local implicit video neural function with spatiotemporal attention. LIVT calculates cross-scale" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "3D attention based on the queried space-time coordinates in the continuous domain and the features from EASM over a local sampling grid. In this way, we can directly and efficiently obtain an accurate unified video INR that predicts the RGB values of queried coordinates, enabling to super-resolve videos in arbitrary frames and resolutions. As shown in Figure 1, we conduct experiments with synthetic GoPro [51] and Adobe240 [64] datasets and the real-world BS-ERGB [68] dataset. Experiments demonstrate that EvEnhancer achieves superior effectiveness and efficiency in super-resolving videos in arbitrary spatial and temporal scales within training distribution (In-Dist.) against existing state-of-the-art methods. We also validate its preferable generalizability at out-of-distribution (OOD) scales. The main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 251, + 553, + 383 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 315, + 251, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 251, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 251, + 553, + 299 + ], + "type": "text", + "content": "- We propose EvEnhancer that subtly marries the unique advantages of event streams with video frames for C-STVSR. Extensive experiments validate its superior performance over recent state-of-the-art methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 300, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 300, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 553, + 335 + ], + "type": "text", + "content": "- We propose the event-adapted synthesis module (EASM) that enables long-term holistic motion trajectory modeling to acquire informative spatiotemporal features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 335, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 335, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 553, + 383 + ], + "type": "text", + "content": "- We propose the local implicit video transformer (LIVT) that integrates local implicit video neural function with cross-scale spatiotemporal attention to learn accurate unify continuous video INR efficiently." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 394, + 400, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 400, + 406 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 400, + 406 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 569 + ], + "type": "text", + "content": "Video Super-Resolution (VSR). Conventional VSR aims to reconstruct HR videos from LR counterparts, which heavily relies on motion estimation and compensation between frames. Earlier methods address this problem by learning frame-level optical flows [39, 72]. Later methods apply deformable convolution [2, 67, 75] or 3D convolution [23, 27] to conduct implicit feature-level motion compensation, exhibiting improved effectiveness and efficiency. Some methods draw inspiration from attention mechanism [28, 74], which incorporate spatial-temporal attention [42, 84, 86, 90] or video transformer [44, 46, 57, 91] to model inter-frame correlations. However, these methods can only upsample at fixed scales." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Video Frame Interpolation (VFI). VFI involves interpolating potential frames between consecutive video frames to enhance the temporal resolution. Flow-based methods [32, 56, 82] leverage optical flow networks [10, 26, 65] to estimate the bidirectional flows between frames. Kernel-based methods [8, 53, 54] utilize local convolution to map interframe motion, which synthesizes target pixels by convolving over input frames. In addition, some recent works focus on splatting-based methods [21, 52, 55] or transformer-based methods [47, 48, 63, 88], showing significant advances. These methods almost learn pixel-level motions for interpolation, which is effective in linear motion but strug" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17756" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 264, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 264, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 264, + 83 + ], + "type": "text", + "content": "gles with complex non-linear motion or occlusions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 87, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 87, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 87, + 294, + 242 + ], + "type": "text", + "content": "Space-Time Video Super-Resolution (STVSR). STVSR can be seen as a combination of VSR and VFI, which reconstruct HR videos spatially and temporally. In literature, most existing methods including traditional [60, 62] and deep learning methods [12, 25, 79] are prone to achieve the STVSR with fixed spatial and temporal scales, denoted by F-STVSR. To mitigate the temporal scale constraints, TMNet [81] proposes controlled frame interpolation with time temporal modulation. SAFA [25] introduces a scale adaptive feature aggregation network that iteratively estimates motion with trainable block-wise scale selection. Though these methods achieve temporal flexibility to some extent, they still fail in arbitrary spatial scales." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 244, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 296, + 484 + ], + "type": "text", + "content": "Event-based VSR & VFI. Event cameras are characterized by high temporal resolution, high dynamic range, and low latency, promoting many VFI [34, 66, 68, 69, 78] and VSR [13, 30, 49] works. E-VSR [30] introduces the pioneering event-based VSR method that utilizes event-based asynchronous interpolation to synthesize asynchronous neighboring frames. EGVSR [49] unleashes the potential of events in arbitrary spatial VSR, which learns implicit neural representations from both RGB frames and events. EvTexture [31] regards the voxelized events as high-frequency signals to improve the textures in RGB frames, yielding promising VSR reconstruction. As for event-based VFI, TimeLens [68] and TimeLens++ [69] innovatively integrate deformation and synthesis processes. CBMNet [34] utilizes cross-modality information to directly estimate inter-frame motion fields. REFID [66], as a flexible method, exhibits robustness even with blurred reference frames. Notably, these methods achieve upsampling only in the spatial or temporal domain independently, and few studies address a unified approach for both, i.e., event-based STVSR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 296, + 714 + ], + "type": "text", + "content": "Implicit Neural Representation (INR) in STVSR. As a method of continuous representation of signals, INR is a common means of continuous super-resolution, including image [4, 5, 17, 38, 83] and video [6, 7, 49, 50] tasks. Inspired by LIIF [5], VideoINR [7] pioneers continuous STVSR (C-STVSR) by learning separate spatial and temporal INRs. Building on this, MoTIF [6] introduces forward spatiotemporal INR to enhance motion field estimation, but it retains the paradigm of decoupling video INR across multiple domains, which lacks robustness across varying scales. Similarly, HR-INR [50], a recent event-based C-STVSR method, also adheres to the same decoupling paradigm as VideoINR [7]. Besides, HR-INR conducts holistic event-frame extraction for long-range motion dependencies and requires repeatedly extracting input frames to realize frame temporal super-resolution, introducing redundancy. Instead, we break the inherent paradigm. We devise the EASM that executes event-modulated alignment and bidirectional recurrent compensation for long-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "term motion trajectory modeling. Based on this, our proposed LIVT achieves a unified spatiotemporal INR through the high temporal resolution grids generated from events, thus achieving better generalization with lower weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 130, + 421, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 421, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 421, + 144 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 149, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 553, + 221 + ], + "type": "text", + "content": "Given LR and low-frame-rete (LFR) videos and corresponding event streams as inputs, we present an event-driven C-STVSR approach to reconstruct HR and HFR videos at arbitrary spatial and temporal scales. Before elaborating on the components of our methodology, we first provide the details for the event representation in this work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 440, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 440, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 440, + 241 + ], + "type": "text", + "content": "3.1. Event Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": "Event cameras asynchronously capture the logarithmic intensity change " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\Delta \\log (I_{x,y,\\tau})" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " of each pixel at coordinate " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " within the RGB image " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "I_{x,y,\\tau}" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " denote the timestamp of an event stream " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = (x,y,\\tau,p)" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " represents the polarity of " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " triggered at " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": ", with 1 and -1 representing the change directions respectively. Compared to conventional RGB images, a raw event stream constituted by large amounts of sparse points, which is capable of continuously capturing intensity change with high temporal resolution. However, processing such an unstructured sparse format in deep networks is non-trivial. Following existing event-based method [66], we represent the event stream into a 3D spatiotemporal voxel grid " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{H \\times W \\times (M + 1)}" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " linearly, where " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " denotes the spatial size, and then discretize the temporal dimension into " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": " event segments " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "[E_1,E_2,\\dots,E_M]" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": ". Each segment records the intensity information within a time window of " + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 246, + 555, + 461 + ], + "type": "text", + "content": ", which is often set to be a small constant (see Figure 2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 382, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 382, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 382, + 479 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{LR} = I_0^{LR}, I_1^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "content": " denote two LR RGB frames and " + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{0 \\to 1}" + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "content": " denote their corresponding event stream. We propose EvEnhancer that addresses the task of super-resolving " + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{LR}" + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "content": " HR video frames with any spatial scale " + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "inline_equation", + "content": "s \\geq 1" + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "content": " and temporal scale " + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "inline_equation", + "content": "t \\geq 1" + }, + { + "bbox": [ + 313, + 483, + 553, + 581 + ], + "type": "text", + "content": ". The overall framework is illustrated in Figure 2, which consists of an event-adapted synthesis module (EASM) and a local implicit video transformer (LIVT)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": "Generally, we first extract initial features on frames and event segments, resulting in RGB frame features " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_0^{LR}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_1^{LR}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ", and event feature " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{0\\rightarrow 1}^E = \\{F_{0\\rightarrow 1}^{E,m}\\}_{m = 1}^M" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " segments. Considering that the latent inter-frame features between RGB frames can be captured in both forward and backward directions, we also reverse the time and polarity of " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " to derive a backward event feature " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{1\\rightarrow 0}^E" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ". Then, EASM models long-term holistic motion trajectories with the guidance of events to generate a spatiotemporal feature sequence " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{F}}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " with high temporal resolution. Finally, LIVT integrates cross-scale spatiotemporal attention with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17757" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 349 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 349 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 349 + ], + "type": "image", + "image_path": "259889dbcadf88b12d6f91d908a1c1664478b4c9c83c986fa5d3b4138bd52733.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 357, + 555, + 391 + ], + "lines": [ + { + "bbox": [ + 54, + 357, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 555, + 391 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our EvEnhancer, an event-driven C-STVSR method which consists of an event-adapted synthesis module (EASM), and a local implicit video transformer (LIVT), where EASM contains two steps: (a) event-modulated alignment (EMA), and (b) bidirectional recurrent compensation (BRC). \"CS\": channel squeeze, \"FFN\": feed-forward neural network, \"Res\": residual blocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "text", + "content": "local implicit function to learn continuous video INR " + }, + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{st}" + }, + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "text", + "content": " that decodes arbitrary space-time coordinates into RGB values, thus yielding HR video frames " + }, + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{st}^{HR}" + }, + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "text", + "content": " with HFR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 235, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 235, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 235, + 455 + ], + "type": "text", + "content": "3.3. Event-Adapted Synthesis Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": "As depicted in Figure 2, EASM contains two steps: 1) Event-modulated alignment incorporates motion cues between input frames with event modulation to acquire latent inter-frame features. Based on the event features and reverse ones, we deploy such alignment in parallel to model the motions forward " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "(^{+} + )" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": " and backward " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "(^{-})" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": ". 2) Bidirectional recurrent compensation propagates the event stream across time and fuses it with acquired features in both directions to maximize the gathering of temporal information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": "Event-Modulated Alignment (EMA). Given input frame features " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "F_0^{LR}" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "F_1^{LR}" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": ", our goal here is to infer the features of intermediate frames according to " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": " event segment features. Motivated by the effectiveness of the PCD module in video restoration [75, 79], we construct a customized pyramid and event-modulated structure that progressively learns the motion offsets and feature alignment. In the following, we take the forward alignment as an example for clarity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "Specifically, for the forward alignment, as shown in Figure 2(a), we implement a 3-level pyramid by downsampling the frame and event features " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "(F_0^{LR}, F_1^{LR}, F_{0\\rightarrow 1}^{E,m})" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": " to 3 different spatial scales " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\times 1" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\times \\frac{1}{2}" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\times \\frac{1}{4}" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": " via stride convolutions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": "At each level " + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": ", we first calculate an initial motion vector by a convolutional block (\"Conv+LeakyReLU (LReLU)\") on concatenated frame features, which is then fused with the learned offset " + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "inline_equation", + "content": "O_{m}^{(l-1)}" + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "inline_equation", + "content": "(l-1)" + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": "-th level to obtain a motion vector " + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "inline_equation", + "content": "MV_{m}^{(l)}" + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": " near the " + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": "-th timestamp. Then, we design an event modulation block (EMB) to generate modulated " + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\widetilde{MV}_{m}^{(l)}" + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": ". The modulation is controlled by the event feature " + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow 1}^{E,m}" + }, + { + "bbox": [ + 313, + 397, + 555, + 535 + ], + "type": "text", + "content": ". Compared to previous PCD-based methods [75, 79, 81], our event-based approach can furnish more deterministic motion cues to facilitate frame synthesis, formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 365, + 544, + 555, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 544, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 365, + 544, + 555, + 563 + ], + "type": "interline_equation", + "content": "\\widetilde {M V} _ {m} ^ {(l)} = \\mathcal {M} ^ {(l)} \\left(M V _ {m} ^ {(l)}, F _ {0 \\rightarrow 1} ^ {E, m (l)}\\right) \\tag {1}", + "image_path": "978ec5f91b229f9e603c5ac10f6137c2198879b85a613cfce42ff89de57f3105.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow 1}^{E,m(l)}" + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "text", + "content": " denote the downsampled event feature at level " + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "text", + "content": ". Then, we can learn the offset " + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "inline_equation", + "content": "O_{m}^{(l)}" + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "inline_equation", + "content": "MV_{m}^{(l)}" + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\widetilde{MV}_{m}^{(l)}" + }, + { + "bbox": [ + 313, + 572, + 555, + 664 + ], + "type": "text", + "content": ". Following [75], we use the deformable convolutional network (DCN) [92] for feature alignment and interpolation. The deformable alignments at different levels are cascaded and aggregated coarse-to-fine for accuracy improvement via the pyramid workflow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "text", + "content": "Notably, as there are " + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "text", + "content": " segment features " + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{F_{0\\rightarrow 1}^{E,m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "text", + "content": " by the event-modulated alignment, we can produce " + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "text", + "content": " intermediate frame features " + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{+} = [F_{1}^{LR + },F_{2}^{LR + },\\dots ,F_{M}^{LR + }]" + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "text", + "content": ". Similarly, for the backward alignment, we get " + }, + { + "bbox": [ + 313, + 664, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{-} =" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17758" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "inline_equation", + "content": "[F_1^{LR-}, F_2^{LR-}, \\dots, F_M^{LR-}]" + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "text", + "content": ". The two are integrated by channel concatenation with a convolutional block, forming the final features " + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{F_{\\tau}^{LR}\\}" + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "text", + "content": " with the input features " + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "inline_equation", + "content": "F_0^{LR}" + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "inline_equation", + "content": "F_1^{LR}" + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "text", + "content": ", where the timestamp " + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "inline_equation", + "content": "\\tau = 0, \\frac{1}{M+1}, \\dots, \\frac{M}{M+1}, 1" + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "text", + "content": " is aligned with the input frames and " + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 70, + 294, + 132 + ], + "type": "text", + "content": " event segments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": "Bidirectional Recurrent Compensation (BRC). Given that the information encapsulated within each event segment is confined to a very narrow time window, after the modulated alignment, we propose bidirectional recurrent compensation that iteratively fuses the event and frame features to model the global temporal information inherent in two modalities, as shown in Figure 2(b). In each recurrent iteration, the compensation block receives the event feature " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "F_{\\tau}^{E}" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": " from the current timestamp " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": ", the forward hidden state from " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "h_{\\tau - \\Delta \\tau}^{f}" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": " the previous timestamp " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\tau - \\Delta \\tau" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": ", and the backward hidden state " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "h_{\\tau + \\Delta \\tau}^{b}" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": " from the next timestamp " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\tau + \\Delta \\tau" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": ", and the aligned frame feature " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "F_{\\tau}^{LR} \\in \\mathbf{F}" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\tau \\pm \\Delta \\tau" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": " denotes the former timestamp and the latter one. Hence, the recurrent fusion process at timestamp " + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 133, + 295, + 301 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 310, + 294, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 310, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 89, + 310, + 294, + 331 + ], + "type": "interline_equation", + "content": "r _ {\\tau} ^ {b}, h _ {\\tau} ^ {b} = \\mathcal {F} ^ {b} \\left(\\mathcal {C} ^ {b} \\left(F _ {\\tau} ^ {L R}, F _ {\\tau} ^ {E}\\right), h _ {\\tau + \\Delta \\tau} ^ {b}\\right), \\tag {2}", + "image_path": "d8796fc4ffd5360b880bdddf70a6c50388276d6d0a05ef37c5537959c4cf17aa.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 327, + 260, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 327, + 260, + 343 + ], + "spans": [ + { + "bbox": [ + 91, + 327, + 260, + 343 + ], + "type": "interline_equation", + "content": "r _ {\\tau} ^ {f}, h _ {\\tau} ^ {f} = \\mathcal {F} ^ {f} (\\mathcal {C} ^ {f} (F _ {\\tau} ^ {L R}, F _ {\\tau} ^ {E}), h _ {\\tau - \\Delta \\tau} ^ {f}, r _ {\\tau} ^ {b}),", + "image_path": "8911778a2db61d93cad4f3fd520f2a8f0b8324930feff819d572930ebf8cbd1f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "inline_equation", + "content": "r_{\\tau}^{b}" + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "inline_equation", + "content": "h_{\\tau}^{b}" + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "content": " denotes the output and hidden states respectively in the backward direction, and " + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "inline_equation", + "content": "r_{\\tau}^{f}" + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "inline_equation", + "content": "h_{\\tau}^{f}" + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "content": " corresponds to the forward. " + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^f" + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^b" + }, + { + "bbox": [ + 55, + 352, + 296, + 496 + ], + "type": "text", + "content": " calculate the channel attention on event and frame features during the forward and backward recurrent iterations respectively. By recurrent propagating the event information across time, we can comprehensively exploit the high temporal resolution properties from events to refine the frame features and maximize the gathering of temporal information. Finally, a bidirectional feature fusion is run by the element-wise addition between the input feature fed into BRC and its output bidirectional features to retain valid information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 237, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 237, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 237, + 517 + ], + "type": "text", + "content": "3.4. Local Implicit Video Transformer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": "After the EASM, we obtain the discrete temporal feature sequence " + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": ". The other key in this method is to learn continuous video INR that specifies the feature at the queried space-time coordinate " + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "inline_equation", + "content": "(\\mathcal{T},x,y)" + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": " are transferred to any target spatial scale " + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": " and temporal scale " + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": " in Figure 2. Previous C-STVSR methods [6, 7] learn spatial and temporal INR separately to realize video INR, which fails to fully capitalize on the spatiotemporal dependencies in the continuous domain. In this work, we propose LIVT which calculates cross-scale 3D attention based on the queried space-time coordinates in the continuous domain over a local sampling grid to learn an accurate unified video INR directly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Assuming the temporal length and spatial size of the local grid in LR space are " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "T^G" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "H^{G} \\times W^{G}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", respectively. For the full temporal range [0, 1], to generate a frame at the timestamp " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\in [0,1]" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", we only need to select the " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "T^G" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " nearest" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 555, + 233 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 555, + 233 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 555, + 233 + ], + "type": "image", + "image_path": "9f7f5341d0d6739882ddc280bf9e2a1da01e7f26423948059adadb7647b95ccf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 282 + ], + "type": "text", + "content": "Figure 3. The structure of our local implicit video transformer (LIVT), which integrates the 3D local spatiotemporal attention with implicit neural function to learn continuous video INR to reconstruct HR and HFR video frames." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": "elements to " + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": " along the time axis. Let the candidate selection of the timestamps " + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\tau_{i} = 0,\\frac{1}{M + 1},\\dots,\\frac{M}{M + 1},1" + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": " , which in actual align with the " + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": " event segments and input frames " + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{LR}" + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": " . We denote the set of candidate timestamps " + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\{\\tau_i\\}" + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\tau}\\subseteq [0,\\frac{1}{M + 1},\\dots,\\frac{M}{M + 1},1]" + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": " , and the size of the " + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "inline_equation", + "content": "|\\mathbf{S}_{\\tau}| = T^G" + }, + { + "bbox": [ + 313, + 289, + 555, + 361 + ], + "type": "text", + "content": " which is selected following the rule as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 369, + 553, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 369, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 373, + 369, + 553, + 396 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {\\tau} = \\underset {\\mathbf {S} _ {\\tau}} {\\arg \\min } \\sum_ {\\tau_ {i} \\in \\mathbf {S} _ {\\tau}} | \\tau_ {i} - \\mathcal {T} |. \\tag {3}", + "image_path": "b0e0c1c941b5edc0510d6b20d3880040ce5cca71f2a7b683cd68d5c2b341ad7e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": "Then, as shown in Figure 3, with the LR feature sequence " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{F}}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " from EASM, we encode them into three embedding spaces through 3D convolution: key " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{K\\}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ", query " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{Q\\}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{V\\}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ". Notably, we do not calculate the spatiotemporal attention across the overall time range. Based on Eq. 3, we conduct temporal selection on " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ", which allows the learning for 3D local attention in a subtemporal space near the target timestamp " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ". On top of this, we employ 3D trilinear sampling on " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " to get the query " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "q_{\\mathcal{T}} \\in \\mathbb{R}^{s^2 HW \\times 1 \\times C}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " for a certain frame, where " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " is the channel number. However, since both " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "k_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "v_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " have size " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "s^2 HW \\times T^G H^G W^G \\times C" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ", it will suffer from expensive computational costs if we directly calculate the attention. To solve this problem, we propose to unfold the sequence along the temporal dimension and calculate cross-scale attention separately on " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "T^G" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " embeddings. Thus, we sample the key " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "k_{\\tau_i}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "v_{\\tau_i}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " respectively using 2D nearest local sampling at each time slice " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\tau_i" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": ". Moreover, we encode and reshape the spatiotemporal relative coordinates " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "(\\delta \\tau, \\delta x, \\delta y) \\in (-1, 1)" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": " from each query point to all pixel points within its local grid, obtaining the spatiotemporal bias " + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "inline_equation", + "content": "b_{\\tau_i} \\in \\mathbb{R}^{s^2 HW \\times H^G W^G \\times 3}" + }, + { + "bbox": [ + 313, + 400, + 555, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "We compute the inner product between " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "q_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "k_{\\tau_i}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " and add it with the " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "b_{\\tau_i}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " to obtain the local attention map at the timestamp " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " after a Softmax operation. After the matrix multiplication with " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "v_{\\tau_i}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ", we can derive the output embedding" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17759" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 112, + 542, + 280 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons for in-distribution (In-dist.) spatiotemporal upsampling scales (" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "t = 4" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "f = 8" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "). † indicates the model trained on the same Adobe240 dataset [64] as ours, and ‡ indicates the model fine-tuned on the Adobe240 dataset. **Bold** and **underline** indicate the best and the second-best performance, respectively. Metrics: PSNR (dB) / SSIM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 112, + 542, + 280 + ], + "lines": [ + { + "bbox": [ + 67, + 112, + 542, + 280 + ], + "spans": [ + { + "bbox": [ + 67, + 112, + 542, + 280 + ], + "type": "table", + "html": "
VFI MethodVSR MethodEventsGoPro-CenterGoPro-AverageAdobe240-CenterAdobe240-AverageParams (M)
TimeLens [69]EGVSR [49]✓28.41 / 0.830727.42 / 0.807726.64 / 0.767525.13 / 0.729872.20+2.45
TimeLens [69]EvTexture [31]✓30.50 / 0.878428.51 / 0.847828.80 / 0.833726.12 / 0.778972.20+8.90
CBMNet-L [34]EGVSR [49]✓26.14 / 0.767423.55 / 0.701825.15 / 0.731122.89 / 0.673522.23+2.45
CBMNet-L [34]EvTexture [31]✓27.58 / 0.796623.92 / 0.716326.74 / 0.777723.38 / 0.698622.23+8.90
REFID [66]†EGVSR [49]✓28.74 / 0.836427.97 / 0.817127.28 / 0.784426.55 / 0.763715.91+2.45
REFID [66]†EvTexture [31]✓30.86 / 0.878429.14 / 0.848929.42 / 0.842427.69 / 0.805915.91+8.90
Zooming Slow-Mo [25]†30.69 / 0.8847-30.26 / 0.8821-11.10
TMNet [81]‡30.14 / 0.869228.83 / 0.851429.41 / 0.852428.30 / 0.835412.26
SAFA [25]†31.28 / 0.889430.22 / 0.876130.97 / 0.887830.13 / 0.87824.94
VideoINR-fixed [7]†30.73 / 0.8850-30.21 / 0.8805-11.31
VideoINR [7]†30.26 / 0.879229.41 / 0.866929.92 / 0.874629.27 / 0.865111.31
MoTIF [6]†31.04 / 0.887730.04 / 0.877330.63 / 0.883929.82 / 0.875012.55
HR-INR [50]†✓31.97 / 0.929832.13 / 0.937131.26 / 0.924631.11 / 0.92168.27
EvEnhancer-light✓33.11 / 0.924232.73 / 0.920331.90 / 0.903331.47 / 0.898831.47 / 0.89885.81
EvEnhancer✓33.52 / 0.929533.30 / 0.927932.43 / 0.912932.18 / 0.911632.18 / 0.91166.55
", + "image_path": "d308602f6e8dee4661fa1e011eaf1f255c7ad8249bce5d77b339df5b65b13904.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 292, + 198, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 198, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 198, + 305 + ], + "type": "inline_equation", + "content": "\\widetilde{z}_{\\tau_i}" + }, + { + "bbox": [ + 55, + 292, + 198, + 305 + ], + "type": "text", + "content": ". This procedure is described as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 310, + 295, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 310, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 97, + 310, + 295, + 338 + ], + "type": "interline_equation", + "content": "\\widetilde {z} _ {\\tau_ {i}} = \\operatorname {S o f t m a x} \\left(\\frac {q _ {\\mathcal {T}} k _ {\\tau_ {i}} ^ {\\mathrm {T}}}{\\sqrt {C}} + b _ {\\tau_ {i}}\\right) \\otimes v _ {\\tau_ {i}}. \\tag {4}", + "image_path": "24336bb19f1073da734c575ab5326f7c6a270b11d5a419e2e66f9e3dd671372c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "text", + "content": "We concatenate the output embeddings " + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\widetilde{z}_{\\tau_i}" + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "text", + "content": " along the channel dimension to construct the complete continuous video INR " + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\widetilde{Z}_{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "text", + "content": " to yield the features corresponding to continuous space-time coordinates. Following [4, 5, 7], we decode the features into RGB values using a 5-layer MLP with the previous query " + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "inline_equation", + "content": "q_{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "text", + "content": " and cell decoding to reconstruct the target RGB frame " + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathcal{T}}^{HR}" + }, + { + "bbox": [ + 55, + 343, + 296, + 429 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 137, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 137, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 137, + 451 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 172, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 172, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 172, + 470 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "content": "Dataset. In our experiments, we train all the models on the Adobe240 dataset [64] which contains 133 video sequences, where 100 for training, 16 for validation, and 17 for testing. We evaluate the performance of our models on both synthetic Adobe240 [64] and GoPro [51] datasets and real-world BS-ERGB [69] and ALPIX-VSR [49] datasets. For the synthetic events, we use the event simulation method vid2e [11] to generate simulated event data. Given a spatial scale " + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "content": " and a temporal scale " + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "content": ", we select consecutive " + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "content": " frames as a clip and select the first and the last frames for " + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "inline_equation", + "content": "s: 1" + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "content": " bicubic downsampling as the LR and LFR inputs fed into the network to generate HR and HFR frame sequences, and the original " + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "content": " frames are groundtruth (GT). Following [31], the event voxels are bicubic-downsampled as the LR event inputs, thus ensuring the spatial size is aligned with the LR frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. Referring VideoINR [7], the training for our model comprises two stages. In the first stage, we set " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "s = 4" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t = 8" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", training for 450K iterations on Adobe240 [64]. In the second stage, we uni-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": "formly sample " + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": " in [1, 4] and fine-tune for another 150K iterations on the same dataset. We use the Adam optimizer [35] with " + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": ", and apply cosine annealing to decay the learning rate from " + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-7}" + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": " over 150K iterations. For data augmentation, we randomly crop " + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": " spatial regions from downsampled images and event segments with random rotations and horizontal flips. The model is optimized with Charbonnier loss [36] using " + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon^{2} = 1 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 292, + 555, + 437 + ], + "type": "text", + "content": ". In this work, we implement two types of variants, the full model EvEnhancer and a lightweight version EvEnhancer-light. To measure our models, we calculate the PSNR (dB) and SSIM on the Y channel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 444, + 539, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 539, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 539, + 456 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": "We comprehensively consider frame-based and event-based one-stage or two-stage state-of-the-art methods for comparison: 1) two-stage event-based STVSR methods consisting of event-based VFI (TimeLens [68], REFID [66], CBMNet-L [34]) and event-based VSR (EGVSR [49], EvTexture [31]); 2) one-stage frame-based F-STVSR methods (Zooming Slow-Mo[79], TMNet [81], SAFA [25]); 3) one-stage frame-based C-STVSR methods (VideoINR [7], MoTIF [6]); 4) one-stage event-based C-STVSR methods (HR-INR [50]). For fair comparisons, we use the models of most VFI methods and all STVSR methods trained or fine-tuned on the same Adobe240 dataset [64] as ours. Notably, due to the limitation that the HR-INR source code is not yet available, the results of HR-INR are derived from its original paper." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. Table 1 reports the results for the F-STVSR task, where the spatial and temporal scales are the same as the training distribution (In-Dist.). \"Center\" represents the average performance of the reconstructed left, right, and center frames. \"Average\" represents the average performance of all reconstructed frames. Notably," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17760" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 552, + 216 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 532, + 82 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 532, + 82 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 532, + 82 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons for out-of-distribution (OOD) spatiotemporal upsampling scales on the GoPro dataset [51]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 552, + 216 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 552, + 216 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 552, + 216 + ], + "type": "table", + "html": "
Temporal ScaleSpatial ScaleTMNet [81]‡SAFA [25]†VideoINR [7]†MoTIF [6]†HR-INR [50]†EvEnhancer -lightEvEnhancer
t = 6s = 430.49 / 0.886131.68 / 0.906830.78 / 0.895431.56 / 0.9064-33.10 / 0.926633.41 / 0.9300
t = 6s = 6--25.56 / 0.767129.36 / 0.8505-29.84 / 0.860330.12 / 0.8675
t = 6s = 12--24.02 / 0.690025.81 / 0.7330-25.41 / 0.726125.50 / 0.7323
t = 12s = 426.38 / 0.793127.95 / 0.824927.32 / 0.814127.77 / 0.823028.87 / 0.885431.17 / 0.896932.07 / 0.9116
t = 12s = 6--24.68 / 0.735826.78 / 0.790827.14 / 0.817328.94 / 0.842529.54 / 0.8579
t = 12s = 12--23.70 / 0.683024.72 / 0.7108-25.25 / 0.724825.45 / 0.7345
t = 16s = 424.72 / 0.752626.38 / 0.783225.81 / 0.773925.98 / 0.775827.29 / 0.855630.15 / 0.876230.93 / 0.8918
t = 16s = 6--23.86 / 0.712325.34 / 0.752726.09 / 0.795428.34 / 0.828328.86 / 0.8434
t = 16s = 12--22.88 / 0.665923.88 / 0.6923-25.10 / 0.721725.25 / 0.7310
t = 6s = 1--32.34 / 0.954534.77 / 0.969638.53 / 0.973537.68 / 0.966738.80 / 0.9714
", + "image_path": "13ac7cce6e9a5740e21fccf22440aef52dc05398fac7304b39b4fdc572e43afd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 256, + 294, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 225, + 294, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 225, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 294, + 247 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons for different temporal upsampling scales " + }, + { + "bbox": [ + 55, + 225, + 294, + 247 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 225, + 294, + 247 + ], + "type": "text", + "content": " (fixed " + }, + { + "bbox": [ + 55, + 225, + 294, + 247 + ], + "type": "inline_equation", + "content": "s = 4" + }, + { + "bbox": [ + 55, + 225, + 294, + 247 + ], + "type": "text", + "content": " ) on the BS-ERGB dataset [69]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 256, + 294, + 370 + ], + "lines": [ + { + "bbox": [ + 56, + 256, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 294, + 370 + ], + "type": "table", + "html": "
MethodEventst=4t=6t=8
TimeLens[69]+EGVSR[49]✓22.98/0.676222.01/0.659221.33/0.6467
TimeLens[69]+EvTexture[31]✓23.58/0.708822.48/0.688721.74/0.6746
CBMNet-L[34]+EGVSR[49]✓22.74/0.685621.93/0.670521.29/0.6657
CBMNet-L[34]+EvTexture[31]✓23.40/0.719922.47/0.702021.78/0.6962
REFID[66]†+EGVSR[49]✓--22.76/0.6478
REFID[66]†+EvTexture[31]✓--22.77/0.6527
SAFA[25]†24.32/0.731423.23/0.712022.41/0.6954
VideoINR[7]†24.06/0.729023.00/0.709822.23/0.6945
MoTIF[6]†24.21/0.729623.15/0.709922.37/0.6939
EvEnhancer-light✓25.32/0.731124.54/0.718424.16/0.7100
EvEnhancer✓25.44/0.733824.70/0.721524.29/0.7138
", + "image_path": "9bfc3249321239f23d882d22ec458660b909be80fe18a02543c09a795c052f9e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 401, + 293, + 423 + ], + "blocks": [ + { + "bbox": [ + 58, + 381, + 292, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 381, + 292, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 381, + 292, + 393 + ], + "type": "text", + "content": "Table 4. " + }, + { + "bbox": [ + 58, + 381, + 292, + 393 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 58, + 381, + 292, + 393 + ], + "type": "text", + "content": " VSR performance on the ALPIX-VSR dataset [49]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 401, + 293, + 423 + ], + "lines": [ + { + "bbox": [ + 57, + 401, + 293, + 423 + ], + "spans": [ + { + "bbox": [ + 57, + 401, + 293, + 423 + ], + "type": "table", + "html": "
MethodVideoINR [7]MoTIF [6]EGVSR [49]EvEnhancer
PSNR / SSIM32.53 / 0.938338.61 / 0.963639.31 / 0.963540.84 / 0.9786
", + "image_path": "7d22523dc1eff32991136bdf40917f9eef16db26a60098910dfcec396fb70d6a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 437, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 617 + ], + "type": "text", + "content": "VideoINR-fixed is trained specifically for single-frame interpolation. We can see that both our EvEnhancer and EvEnhancer-light significantly outperform existing methods in all settings with much fewer model parameters, which achieves more than 1 dB PSNR improvements. Then, we also evaluate our models on the C-STVSR task, where spatial and temporal scales are out of the training distribution. The results are illustrated in Table 2. It can be observed that both our models exceed almost all existing methods at various spatial and temporal scales, especially in the challenging large temporal scale, achieving superior effectiveness and generalization ability. In particular, our EvEnhancer-light and EvEnhancer use lighter-weight models and perform better than HR-INR [50] at most temporal and spatial scales, especially at OOD scales." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "We further validate the effectiveness of our method on the real-world BS-ERGB dataset under the setting of fixed spatial scale " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "s = 4" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " but varying temporal scales " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": ". The comparisons in Table 3 demonstrate that our method yields the best quantitative performance under all conditions. Compared to cascaded event-based VFI and VSR, our EvEnhancer simultaneously learns temporal interpolation and spatial VSR, thus achieving superior performance. Due to" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 277, + 552, + 354 + ], + "blocks": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "text", + "content": "Table 5. TFLOPs comparisons of C-STVSR models at different upsampling temporal scales " + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "text", + "content": " (fixed " + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "inline_equation", + "content": "s = 4" + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "text", + "content": "). * indicates the repeated iterations during inference on the NVIDIA GeForce RTX 3090 GPU due to their large computations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 277, + 552, + 354 + ], + "lines": [ + { + "bbox": [ + 316, + 277, + 552, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 552, + 354 + ], + "type": "table", + "html": "
Temporal ScaleVideoINR [7] (11.31M)MoTIF [6] (12.55M)EvEnhancer -light (5.81M)EvEnhancer (6.55M)
t = 22.0112.0431.8763.398
t = 42.3954.086*2.4724.647
t = 62.7796.129*3.0675.885
t = 83.1636.129*3.6637.129
t = 123.93210.22*4.8549.617
t = 164.70012.26*6.04412.10
", + "image_path": "ed0ea1f2701b710c782415af7a06e615ae5c7cac0b98117cacf0939d28345fca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "text", + "content": "the high temporal resolution of events, equipped by EASM, our EvEnhancer enables holistic motion trajectory modeling, thus outperforming frame-based methods. Benefiting from the cross-scale attention in LIVT, it also performs preferably at OOD scales. To validate the effectiveness of our EvEnhancer for real-world LR events, we conduct experiments on the real-world ALPIX-VSR dataset [49] (image resolution: " + }, + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "inline_equation", + "content": "3264 \\times 2448" + }, + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "text", + "content": ", event resolution: " + }, + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "inline_equation", + "content": "1632 \\times 1224" + }, + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "text", + "content": ") for " + }, + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 312, + 373, + 555, + 493 + ], + "type": "text", + "content": " VSR. As shown in Table 4, EvEnhancer significantly surpasses all methods without additional fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "spans": [ + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "text", + "content": "Model Complexity. To validate the model efficiency of the proposed methods, besides the model parameter comparison in Table 1, we compare the TFLOPs with existing C-STVSR methods in Table 5, which are calculated on the upsampling of a clip with a resolution of " + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "inline_equation", + "content": "180 \\times 320" + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "text", + "content": " by a spital upsampling scale " + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "inline_equation", + "content": "s = 4" + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "text", + "content": " at different temporal upsampling scales " + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "text", + "content": ". VideoINR shows the fewest TFLOPs but more parameters and the worst performance (see Table 1). At the low temporal scale (" + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "inline_equation", + "content": "\\leq 4" + }, + { + "bbox": [ + 312, + 495, + 556, + 627 + ], + "type": "text", + "content": "), MoTIF is computationally efficient than ours. However, as the temporal scale increases, its computations increase and exceed both our models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 629, + 556, + 715 + ], + "type": "text", + "content": "Qualitative Comparison. With the qualitative results in Figure 4, compared to VideoINR [7] and MoTIF [6], our models reconstruct the faithful HR frames with clearer textures. By computing the difference maps between generated frames and corresponding ground truth frames, we can see that our methods are more excellent at tackling non-linear motions and recovering more details." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17761" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 547, + 294 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 547, + 294 + ], + "type": "image", + "image_path": "f65700759f696021fdab42b743e8e4cd5bda3f73b27412d45eebaad6c7a14308.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 297, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 555, + 322 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison for In-dist. scale " + }, + { + "bbox": [ + 55, + 297, + 555, + 322 + ], + "type": "inline_equation", + "content": "(t = 8, s = 4)" + }, + { + "bbox": [ + 55, + 297, + 555, + 322 + ], + "type": "text", + "content": " on the GoPro dataset [51]. We compare the normalized absolute difference maps (yellow boxes) for the same regions in each frame as in the GT frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 79, + 373, + 272, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 331, + 298, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 298, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 298, + 365 + ], + "type": "text", + "content": "Table 6. Ablations on the designs of event-adapted synthesis module (EASM). \"fwd.\" and \"bwd.\" denote the forward and backward directions, respectively. Metrics: PSNR (dB) / SSIM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 373, + 272, + 427 + ], + "lines": [ + { + "bbox": [ + 79, + 373, + 272, + 427 + ], + "spans": [ + { + "bbox": [ + 79, + 373, + 272, + 427 + ], + "type": "table", + "html": "
EMABRCBFIn-Dist.OOD
fwd. & bwd.32.50 / 0.916628.86 / 0.8407
fwd. & bwd.✓33.13 / 0.925229.42 / 0.8539
fwd.✓✓33.20 / 0.926229.47 / 0.8556
fwd. & bwd.✓✓33.30 / 0.927929.54 / 0.8579
", + "image_path": "8ab95dd8c19e9d28631ea008ed9592a6ccd097239be30998decceaa0e3554a11.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 442, + 157, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 157, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 157, + 453 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 521 + ], + "type": "text", + "content": "We conduct ablation studies by keeping the same training strategy on the Adobe240 dataset [64], and testing on the GoPro dataset [51] at both In-Dist. " + }, + { + "bbox": [ + 55, + 460, + 295, + 521 + ], + "type": "inline_equation", + "content": "(t = 8, s = 4)" + }, + { + "bbox": [ + 55, + 460, + 295, + 521 + ], + "type": "text", + "content": " and OOD " + }, + { + "bbox": [ + 55, + 460, + 295, + 521 + ], + "type": "inline_equation", + "content": "(t = 12, s = 6)" + }, + { + "bbox": [ + 55, + 460, + 295, + 521 + ], + "type": "text", + "content": " spatiotemporal scales to validate the effectiveness of our methodology design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 652 + ], + "type": "text", + "content": "Event-Adapted Synthesis Module. Table 6 presents the ablations of EASM, which involves the event-modulated alignment (EMA), bidirectional recurrent compensation (BRC), and bidirectional feature fusion (BF). The model with only EMA achieves the worst performance. When we introduce BRC in the model, one can see the performance improves. Besides, with the integration BRC, BF, and EMA, the model achieves the best at both In-Dist. and OOD scales. We can also observe that the model with only a single forward alignment can suffer from performance degradation. The results demonstrate the effects of EASM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Local Implicit Video Transformer. Table 7 investigates the effect of LIVT. We first implement the model with a 2D decoupling strategy as videoINR [7] but introduce the local attention, concatenate cell decoding, and query, one can see that the model achieves better performance than the model" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 329, + 384, + 538, + 448 + ], + "blocks": [ + { + "bbox": [ + 313, + 331, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 374 + ], + "type": "text", + "content": "Table 7. Ablations on the designs of local implicit video transformer (LIVT). \"LA\": local attention, \"CE\": concatenate cell decoding, \"PQ\": concatenate the previous query. Metrics: PSNR (dB) / SSIM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 384, + 538, + 448 + ], + "lines": [ + { + "bbox": [ + 329, + 384, + 538, + 448 + ], + "spans": [ + { + "bbox": [ + 329, + 384, + 538, + 448 + ], + "type": "table", + "html": "
INR ModeLACEPQIn-dist.OOD
2D decoupling✓✓✓33.26 / 0.927229.26 / 0.8525
3D unification27.32 / 0.797525.03 / 0.7032
3D unification✓33.27 / 0.927629.47 / 0.8565
3D unification✓✓33.30 / 0.928229.49 / 0.8570
3D unification✓✓✓33.30 / 0.927929.54 / 0.8579
", + "image_path": "59fc303cf78f1fb5ddb32ea3f71574532c9cb043c8b2e188283d80480485ad64.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 536 + ], + "type": "text", + "content": "without any components. When we replace the 2D decoupling with our unified 3D video INR, the model produces results with higher PSNR and SSIM, especially at OOD scales. In addition, the concatenation of the cell [5] and previous query [7] brings a larger gain at OOD scales than In-Dist., which validates their effects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we introduce EvEnhancer, a novel approach that amalgamates the unique advantages of event streams to enhance the effectiveness, efficiency, and generalizability of C-STVSR. Our EvEnhancer is underpinned by two critical components: 1) the EASM capitalizes on the spatiotemporal correlations between frames and events to discern and learn long-term motion trajectories; and 2) the LIVT, which integrates a local implicit video neural function with cross-scale spatiotemporal attention to learn continuous video representations. Experiments indicate that EvEnhancer significantly surpasses current state-of-the-art methods." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17762" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "This work was partly supported by the National Natural Science Foundation of China (62302141, 62331003, 62120106009), the Joint Funds of the National Natural Science Foundation of China (U23A20314), the Fundamental Research Funds for the Central Universities (JZ2024HGTB0255), Beijing Natural Science Foundation (L223022), and the Natural Science Foundation of Hebei Province (F2024105029)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 216, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "type": "text", + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 326 + ], + "type": "text", + "content": "[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 327, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 296, + 381 + ], + "type": "text", + "content": "[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 383, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 296, + 446 + ], + "type": "text", + "content": "[4] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 502 + ], + "type": "text", + "content": "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 504, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 296, + 568 + ], + "type": "text", + "content": "[6] Yi-Hsin Chen, Si-Cun Chen, Yen-Yu Lin, and Wen-Hsiao Peng. Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video superresolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23131-23141, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 570, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 645 + ], + "type": "text", + "content": "[7] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2047-2057, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "text", + "content": "[8] Xianhang Cheng and Zhenzhong Chen. Video frame interpolation via deformable separable convolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10607-10614, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "text", + "content": "[9] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 128 + ], + "type": "text", + "content": "once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 197, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 239 + ], + "type": "text", + "content": "[11] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In CVPR, pages 3586-3595, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 243, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 297 + ], + "type": "text", + "content": "[12] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "text", + "content": "[13] Jin Han, Yixin Yang, Chu Zhou, Chao Xu, and Boxin Shi. Evintsr-net: Event guided multiple latent frames reconstruction and super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4882-4891, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "text", + "content": "[14] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3897-3906, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 411, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 464 + ], + "type": "text", + "content": "[15] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Space-time-aware multi-resolution video enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2859-2868, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 467, + 555, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 532 + ], + "type": "text", + "content": "[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "type": "text", + "content": "[17] Zongyao He and Zhi Jin. Latent modulated function for computational optimal continuous image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26026-26035, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 579, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 554, + 622 + ], + "type": "text", + "content": "[18] Qiqi Hou, Abhijay Ghildyal, and Feng Liu. A perceptual quality metric for video frame interpolation. In European Conference on Computer Vision, pages 234-253. Springer, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "text", + "content": "[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[20] Mengshun Hu, Kui Jiang, Zheng Wang, Xiang Bai, and Ruimin Hu. Cycmunet+: Cycle-projected mutual learning" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17763" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "for spatial-temporal video super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[21] Ping Hu, Simon Niklaus, Lu Zhang, Stan Sclaroff, and Kate Saenko. Video frame interpolation with many-to-many splating and spatial selective refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[22] Cong Huang, Jiahao Li, Lei Chu, Dong Liu, and Yan Lu. Arbitrary-scale video super-resolution guided by dynamic context. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2302, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 295, + 230 + ], + "type": "text", + "content": "[23] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 40(4):1015-1028, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "type": "text", + "content": "[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 331 + ], + "type": "text", + "content": "[25] Zhewei Huang, Ailin Huang, Xiaotao Hu, Chen Hu, Jun Xu, and Shuchang Zhou. Scale-adaptive feature aggregation for efficient space-time video super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4228-4239, 2024. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "type": "text", + "content": "[26] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 295, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 295, + 444 + ], + "type": "text", + "content": "[27] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 478 + ], + "type": "text", + "content": "[28] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. arXiv preprint arXiv:2008.05765, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 544 + ], + "type": "text", + "content": "[29] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 601 + ], + "type": "text", + "content": "[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "type": "text", + "content": "[31] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. EvTexture: Event-driven Texture Enhancement for Video Super-Resolution. In Proceedings of the 41st International Conference on Machine Learning, pages 22817–22839. PMLR, 2024. 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[32] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[33] Soo Ye Kim, Jihyong Oh, and Munchurl Kim. Fisr: Deep joint frame interpolation and super-resolution with a multiscale temporal loss. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11278-11286, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "text", + "content": "[34] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[36] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[37] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5316-5325, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 321, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 555, + 364 + ], + "type": "text", + "content": "[38] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1929-1938, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "text", + "content": "[39] Feng Li, Huihui Bai, and Yao Zhao. Learning a deep dual attention network for video super-resolution. IEEE transactions on image processing, 29:4474-4488, 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "text", + "content": "[40] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[41] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10522-10531, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 513, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 577 + ], + "type": "text", + "content": "[42] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "type": "text", + "content": "[43] Zekun Li, Hongying Liu, Fanhua Shang, Yuanyuan Liu, Liang Wan, and Wei Feng. Savsr: Arbitrary-scale video super-resolution via a learned scale-adaptive network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3288-3296, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[44] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[45] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17764" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "text", + "content": "Learning event-driven video deblurring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 695-710. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[46] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "text", + "content": "[47] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19125-19134, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "text", + "content": "[48] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 330 + ], + "type": "text", + "content": "[49] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1557-1567, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 374 + ], + "type": "text", + "content": "[50] Yunfan Lu, Zipeng Wang, Yusheng Wang, and Hui Xiong. Hr-inr: Continuous space-time video super-resolution via event camera. arXiv preprint arXiv:2405.13389, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "text", + "content": "[51] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 476 + ], + "type": "text", + "content": "[52] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "text", + "content": "[53] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE international conference on computer vision, pages 261-270, 2017. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 567 + ], + "type": "text", + "content": "[54] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive convolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 670-679, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "text", + "content": "[55] Simon Niklaus, Ping Hu, and Jiawen Chen. Splitting-based synthesis for video frame interpolation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 713-723, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 657 + ], + "type": "text", + "content": "[56] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14539-14548, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[57] Zhongwei Qiu, Huan Yang, Jianlong Fu, Daochang Liu, Chang Xu, and Dongmei Fu. Learning degradation-robust spatiotemporal frequency-transformer for video superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[58] Nasrin Rahimi and A Murat Tekalp. Spatio-temporal perception-distortion trade-off in learned video sr. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1400-1404. IEEE, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "text", + "content": "[59] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF international conference on computer vision, pages 892-900, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "text", + "content": "[60] Oded Shahar, Alon Faktor, and Michal Irani. Space-time super-resolution from a single video. IEEE, 2011. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "text", + "content": "[61] Wei Shang, Dongwei Ren, Wanying Zhang, Yuming Fang, Wangmeng Zuo, and Kede Ma. Arbitrary-scale video superresolution with structural and textural priors. In European Conference on Computer Vision, pages 73-90. Springer, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "text", + "content": "[62] Eli Shechtman, Yaron Caspi, and Michal Irani. Space-time super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(4):531-545, 2005. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "text", + "content": "[63] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[64] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 1, 2, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 405, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 458 + ], + "type": "text", + "content": "[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "text", + "content": "[66] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "text", + "content": "[67] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 645 + ], + "type": "text", + "content": "[68] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[69] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17765" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[70] Hai Wang, Wenming Yang, Qingmin Liao, and Jie Zhou. Bi- rstu: Bidirectional recurrent upsampling network for spacetime video super-resolution. IEEE Transactions on Multimedia, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "text", + "content": "[71] Hai Wang, Xiaoyu Xiang, Yapeng Tian, Wenming Yang, and Qingmin Liao. Stdan: deformable attention network for space-time video super-resolution. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 166, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 294, + 232 + ], + "type": "text", + "content": "[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through hr optical flow estimation. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 514–529. Springer, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 234, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 278 + ], + "type": "text", + "content": "[73] Li Wang, Dong Li, Yousong Zhu, Lu Tian, and Yi Shan. Dual super-resolution learning for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3774-3783, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "type": "text", + "content": "[74] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 326, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 294, + 381 + ], + "type": "text", + "content": "[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[76] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 440, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 440, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 440, + 294, + 495 + ], + "type": "text", + "content": "[77] Guangyang Wu, Xin Tao, Changlin Li, Wenyi Wang, Xiao-hong Liu, and Qingqing Zheng. Perception-oriented video frame interpolation via asymmetric blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2753-2762, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 553 + ], + "type": "text", + "content": "[78] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In European Conference on Computer Vision, pages 267-283. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 620 + ], + "type": "text", + "content": "[79] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3370-3379, 2020. 2, 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 622, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 294, + 656 + ], + "type": "text", + "content": "[80] Chenming Xu, Meiqin Liu, Chao Yao, Weisi Lin, and Yao Zhao. Ibvc: Interpolation-driven b-frame video compression. Pattern Recognition, 153:110465, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[81] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6388-6397, 2021. 3, 4, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 620 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[82] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[83] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[84] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[85] Jun-Sang Yoo, Hongjae Lee, and Seung-Won Jung. Video object segmentation-aware video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12322-12333, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[86] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17834-17843, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[87] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S Ren. Training weakly supervised video frame interpolation with events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14589-14598, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "type": "text", + "content": "[88] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "type": "text", + "content": "[89] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8799-8808, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 465, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 553, + 518 + ], + "type": "text", + "content": "[90] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "text", + "content": "[91] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "text", + "content": "[92] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 4" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17766" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_content_list.json b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d932ff4e5cbc7223a5156ac2745fdbf10ef03973 --- /dev/null +++ b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_content_list.json @@ -0,0 +1,1688 @@ +[ + { + "type": "text", + "text": "EvOcc: Accurate Semantic Occupancy for Automated Driving Using Evidence Theory", + "text_level": 1, + "bbox": [ + 261, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonas Käblé $^{1,2}$ Sascha Wirges $^{1}$ $^{1}$ Bosch Center for Artificial Intelligence", + "bbox": [ + 176, + 202, + 495, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "firstname_lastname@de.bosch.com", + "bbox": [ + 218, + 242, + 495, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maxim Tatarchenko1 Eddy Ilg2 \n2University of Technology Nuremberg", + "bbox": [ + 506, + 202, + 815, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eddy.ilg@utn.de", + "bbox": [ + 509, + 242, + 645, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present EvOcc, a novel framework for evidential semantic occupancy mapping. It consists of two parts: (1) an evidential approach for calculating the ground-truth 3D semantic occupancy maps from noisy LiDAR measurements, and (2) a method for training image-based occupancy estimation models through a new loss formulation. In contrast to state-of-the-art semantic occupancy maps, our approach explicitly models the uncertainty introduced by unobserved spaces or contradicting measurements, and we show that using it results in significantly stronger models. Evaluated as ray-based mIoU, our evidential semantic occupancy mapping approach improves over the baselines by at least $15.8\\%$ for the ground truth and $5.5\\%$ for the trained model. Overall, we make a significant contribution towards more detailed and uncertainty-aware 3D environment understanding and safe operation in autonomous driving.", + "bbox": [ + 88, + 324, + 485, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 599, + 220, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A critical component for safe automated driving is accurate and reliable environment perception, which includes understanding the scene geometry as well as interpreting its semantic content. While various 3D representations have been explored in the past, semantic voxel grids have become a recent de facto standard for ML-based environment perception [15, 31, 32, 34, 35].", + "bbox": [ + 89, + 627, + 485, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For a long time, automated driving relied on LiDAR sensors due to their low and distance-independent measurement errors. However, recent methods attempt to reconstruct the 3D environment from camera images alone [10-12, 15, 17, 19-21, 24], since LiDAR is sparse, expensive, and redundant. This makes predicting accurate environment maps from images an important task, and further brings up the problem of obtaining ground-truth training data. The latter is usually produced by recording LiDAR measurements, and employing heuristics to convert them into semantic occupancy maps. However, such reconstructed occupancy maps", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4817f692080b6b73f5f72534e4bffde3718da7596df33fd253fe3662e6a4c53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 290, + 905, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1011a8cf653cb5a2f062ae2732285b1309358a8a08a7f8cca39aff296f369c11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 405, + 710, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73d08c7bbdf454641943c857213fa40d6579c46fbfda248dfd6aca0442925c42.jpg", + "image_caption": [ + "Figure 1. Semantic occupancy and birds-eye-view uncertainty. Our model uses multi-view camera images as input and predicts semantic occupancy. We visualize the ground-truth semantic occupancy obtained from LiDAR scans on left and the predicted semantic occupancy on the right. Below we show the GT uncertainty obtained from our semantic basic belief assignment (BBA) using the LiDAR scans as well as the predicted uncertainties. High uncertainties (e.g., due to occlusions) are colored in red, whereas areas of low uncertainty are colored in blue." + ], + "image_footnote": [], + "bbox": [ + 517, + 516, + 710, + 628 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72416b7e24146e874ca4ad1dfa61dd2c27f258ecdf4d931f24f157568093cb3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 405, + 903, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ef55350608bfcc444883b360bb94bdfd9cf0ff522d5a945459841de98bdd764.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 516, + 903, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are often inaccurate. A recent work by Käblé et al. [15] has shown that the accuracy can in fact be improved by a margin of $30 - 50\\%$ when simple heuristics are replaced with the framework of evidential mapping. Different from a classical probabilistic interpretation, this approach explicitly models conflicting hypotheses and unknown states, which is crucial for representing noisy real-world measurements.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27467", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, the authors only model the voxel occupancy state from LiDAR scans, without any notion of semantic classes.", + "bbox": [ + 89, + 90, + 480, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present $EvOcc$ , an evidence theory (ET) [3, 28] based framework for (1) accurately reconstructing semantic occupancy maps from LiDAR scans (see Fig. 2) and (2) using these as ground-truth to train multi-camera vision models to predict evidential occupancy maps (see Fig. 1). To this end, we extend the evidential mapping framework from Käblé et al. [15] to include semantic class labels. We derive a novel basic belief assignment (BBA) formulation that can handle noisy and partially annotated LiDAR measurements. We also employ a new loss for training semantic occupancy prediction models with our ground-truth data.", + "bbox": [ + 88, + 137, + 482, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic occupancy maps produced by our approach are far superior to those from state-of-the-art semantic occupancy methods [31, 32, 34, 35], which we demonstrate both qualitatively and quantitatively. Using our reconstructed evidential occupancy maps as training data, we are able to train evidential occupancy models that significantly outperform those trained on previously available ground truth. Note that this large improvement is achieved by only changing the model supervision, while keeping the model architecture the same. In summary, we make the following contributions:", + "bbox": [ + 89, + 319, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a complete 3D evidential mapping framework for semantic occupancy consisting of two parts: occupancy GT generation and image-based occupancy estimation.", + "- We apply a novel BBA formulation based on noisy, partially annotated LiDAR measurements, and employ a novel loss that directly uses the obtained evidential semantic occupancy maps as supervision.", + "- We demonstrate both qualitatively and quantitatively, that the semantic occupancy maps produced by our trained model and mapping approach are significantly more accurate than those from state-of-the-art methods in terms of agreement with LiDAR measurements (see Sec. 5.3)." + ], + "bbox": [ + 89, + 472, + 483, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our code is available on GitHub1.", + "bbox": [ + 89, + 652, + 315, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 684, + 232, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Occupancy Datasets. Training perception models for automated driving requires large-scale multimodal datasets, such as KITTI [8], nuScenes [2], Waymo [30] and Agro-verse2 [36]. These datasets record the surrounding environment of the vehicle with multiple onboard sensors, such as LiDAR and cameras. Additionally, they provide annotations for the raw sensor data for tasks like 3D object detection [17, 19, 23, 24] and semantic LiDAR segmentation [12, 37]. The task of camera-only semantic occupancy prediction [18, 31, 32, 34, 35] is to predict the occupancy state and class of each voxel in the environment of the vehi", + "bbox": [ + 89, + 710, + 482, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cle, given the onboard camera images.", + "bbox": [ + 511, + 90, + 767, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A common process for obtaining ground-truth (GT) 3D occupancy grids is aggregating LiDAR measurements over multiple time steps under the consideration of moving objects [31, 32, 34, 35]. The resulting aggregated point clouds are then partitioned into dense voxel grids, where each voxel containing a point is labeled according to the point's annotation. Finally, ray casting is used to determine the free space labels between the LiDAR origin and the first occupied voxel. The quality of the occupancy grids can be improved through additional processing steps. For instance, Occ3D [31] and SurroundOcc [35] propose to make the point cloud more dense using mesh and surface reconstruction [13, 33]. Furthermore, Occ3D enforces consistency between the camera images and the occupancy grid through image-guided voxel refinement. OpenOccupancy [34] improves the label quality with additional human supervision and their Augment and Purifying process. Käblé et al. [15] leverage evidence theory (ET) and achieve the current state of the art by proposing to model the occupancy state of each voxel based on the number of transmissions and reflections through it. Our work builds up on [15] and proposes a framework using ET that can model beliefs for semantic classes.", + "bbox": [ + 511, + 106, + 906, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Semantic Occupancy Prediction. The task of 3D semantic occupancy prediction is to estimate the state of each voxel in the environment of the vehicle using only camera images. A voxel can either be free or occupied by a specific class. The CVPR 2023 Occupancy Prediction Challenge [31, 32] is a popular benchmark for this task. The winner of this challenge, FB-OCC [20, 21], extends BEVDet [11] by a forward-backward view transformation module. TPVFormer [12] addresses the high memory and compute requirements for a dense 3D grid and projects the 3D space onto 3D planes. Additional depth supervision for the process of splatting 2D image features into the 3D space has been introduced by BEVDepth [17]. BEVDet4D [10] aggregates additional information over time for the task of 3D object detection. In this work, we use BEVDet4D-Occ-R50 as a baseline model, which is an adaptation of BEVDet4D specifically tailored for the task of semantic occupancy prediction and commonly used as reference model. We show that training this model on the new ground truth obtained with our approach gives superior results.", + "bbox": [ + 511, + 458, + 908, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Occupancy Grid Mapping. Static occupancy grid mapping using a Bayesian interpretation of probabilities is a well-explored area in robotics [7, 25]. The state of a voxel cell is modeled as a Bernoulli-distributed random variable and can either be occupied or free. Since the state of the cell is defined by a single probability, the model does not consider the uncertainty in the measurements to estimate the cell state, nor does it handle conflicts between different mea", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 107, + 886, + 415, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27468", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30475cc4e711c9c4644d176754cd70ca8ed7683fcd00ff827bd5987cd1d5ff22.jpg", + "image_caption": [ + "Figure 2. Overview. A scheme of our evidential framework for semantic occupancy mapping." + ], + "image_footnote": [], + "bbox": [ + 102, + 85, + 903, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "suresments. To circumvent these limitations, Yang and Aitken [39] propose to model the state of a cell using ET [3, 28], which can explicitly represent the uncertainty. Richter et al. [27] and Li et al. [16] model the 2D birds-eye-view (BEV) occupancy maps using ET. More recent works extend the mapping task to 3D and consider dynamic environments as well as semantics. There are some specialized models for automated driving scenarios. Richter et al. [26] split dynamic traffic scenes into two evidential BEV grid maps containing the ground plane and everything on top of the ground. Kim et al. [14] present a 3D semantic occupancy mapping approach, which fuses unreliable semantic predictions using ET. Furthermore, improvements in the mapping process for dynamic scenes can be achieved by considering the motion of objects using scene flow from LiDAR measurements [1]. A framework for 3D occupancy grid mapping based on ET within a dynamic environment is presented in [15]. However, the presented method is limited to occupancy maps without semantic information.", + "bbox": [ + 89, + 295, + 485, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 598, + 230, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Evidence Theory", + "text_level": 1, + "bbox": [ + 89, + 623, + 256, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we briefly revisit the fundamentals of evidence theory (ET), or Dempster-Shafer theory [3, 28], used to model the semantic occupancy state of individual voxels (see Sec. 4.1). ET provides a flexible framework for reasoning about the uncertainty, be it total uncertainty or partial knowledge about a system [29]. A single hypothesis is described by a subset of the frame of discernment $\\Omega$ , and the BBA", + "bbox": [ + 89, + 646, + 485, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {m} \\colon 2 ^ {\\Omega} \\rightarrow [ 0, 1 ] \\quad \\text {w i t h} \\quad \\sum_ {\\omega \\subseteq \\Omega} \\mathrm {m} (\\omega) = 1 \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 763, + 483, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "assigns a normalized belief mass $\\mathrm{m}(\\omega)$ to each hypothesis $\\omega \\subseteq \\Omega$ .", + "bbox": [ + 89, + 809, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, the belief mass represents the proportion of relevant and available evidence supporting the hypothesis $\\omega$ but not any particular subset of it. The classical probability of a hypothesis $\\omega$ lies between two values known as belief", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and plausibility. The belief", + "bbox": [ + 511, + 295, + 697, + 310 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {b e l} (\\omega) = \\sum_ {X \\subseteq \\omega} \\mathrm {m} (X) \\leq \\operatorname {p r o b} (\\omega) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 321, + 906, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of a hypothesis $\\omega$ is the sum of all belief masses supporting the hypothesis $\\omega$ . It serves as a lower bound of the probability that a certain hypothesis is true. The plausibility", + "bbox": [ + 511, + 364, + 906, + 411 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {p l} (\\omega) = 1 - \\sum_ {X \\cap \\omega = \\emptyset} \\mathrm {m} (X) \\geq \\operatorname {p r o b} (\\omega) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 422, + 906, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is an upper bound for this probability. It can be defined as the complement of all belief masses not supporting the hypothesis $\\omega$ . The uncertainty", + "bbox": [ + 511, + 465, + 905, + 511 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {u} (\\omega) = \\operatorname {p l} (\\omega) - \\operatorname {b e l} (\\omega) \\in [ 0, 1 ] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 522, + 906, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of a hypothesis $\\omega$ can then be defined as the discrepancy between plausibility and belief [6].", + "bbox": [ + 511, + 551, + 905, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spherical Sensor Mapping", + "text_level": 1, + "bbox": [ + 511, + 590, + 754, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Käblé et al. [15] describe an approach to determine the state of a voxel based on the transmissions, and reflections through it. These are measured by a LiDAR sensor and provide evidence for a voxel being free or occupied by a certain class. Different from their work, we address the multi-class case here.", + "bbox": [ + 511, + 613, + 905, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reflections Grid Map. Instead of mapping all reflections into a single spherical grid map, we create a separate spherical reflections grid map $r_{\\mathrm{ci}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)$ for each class $\\mathbf{c}_i$ . On top of that, we have one additional reflections grid map $r_{\\mathrm{unlabeled}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)$ for all unlabeled reflections2. This results in a total of $C + 1$ spherical reflection grid maps, where $C$ is the number of classes.", + "bbox": [ + 511, + 722, + 906, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transmissions Grid Map. Since transmissions represent free space and are independent of the reflections' semantic", + "bbox": [ + 511, + 847, + 906, + 877 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2In nuScenes [2] only one of every 10 LiDAR scans is labeled.", + "bbox": [ + 529, + 886, + 862, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27469", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "classes, we only produce a single spherical transmission map $t^{\\mathrm{ph}}(\\rho ,\\phi ,\\theta)$ . For a given voxel cell located at $(\\rho ,\\theta ,\\phi)$ , the number of transmissions:", + "bbox": [ + 89, + 90, + 482, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt ^ {\\mathrm {s p h}} (\\rho , \\phi , \\theta) = \\sum_ {\\rho^ {\\prime} > \\rho} \\underbrace {\\left(r _ {\\text {u n l a b e l e d}} ^ {\\mathrm {s p h}} \\left(\\rho^ {\\prime} , \\phi , \\theta\\right) + \\sum_ {i = 1} ^ {C} r _ {\\mathrm {c} _ {i}} ^ {\\mathrm {s p h}} \\left(\\rho^ {\\prime} , \\phi , \\theta\\right)\\right)} _ {\\text {S u m o v e r l a b e l e d a n d u n l a b e l e d r e f l e c t i o n s}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 138, + 488, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is given by the sum over all reflections with the radius $\\rho'$ greater than $\\rho$ of the voxel under consideration. This can be efficiently implemented as a cumulative sum from the maximal to the minimal radial distance.", + "bbox": [ + 89, + 220, + 483, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grid Map Normalization. We normalize the number of reflections and transmissions per voxel by their respective volume. We obtain the normalized reflections $\\hat{r}_{\\mathrm{ci}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)$ and $\\hat{r}_{\\mathrm{unlabled}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)$ as well as the normalized transmissions $\\hat{t}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)$ for each time step separately. Details can be found in the Supplemental in Sec. A.", + "bbox": [ + 89, + 297, + 483, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Frame Aggregation. We aggregate the reflections and transmissions from multiple time steps into a single Cartesian voxel grid. Using known per-object transformations between the reference frame at time step $n$ and frames at all other time steps $k$ , we can accumulate the final per-voxel values as:", + "bbox": [ + 89, + 407, + 483, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {r} _ {\\mathrm {c} _ {i}, n} (x, y, z) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} g \\left(\\hat {r} _ {\\mathrm {c} _ {i}, k} ^ {\\text {s p h}} (\\rho , \\phi , \\theta)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 501, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g$ applies the motion compensation and transforms spherical coordinates into Cartesian ones. The same applies to the unlabeled reflections $\\hat{r}_{\\mathrm{unlabled}}(x,y,z)$ and transmissions $\\hat{t}(x,y,z)$ .", + "bbox": [ + 89, + 547, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 619, + 181, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we present our EvOcc framework split into two stages. The first stage (Sec. 4.1) reconstructs accurate semantic occupancy maps from sparse LiDAR scans with partially annotated semantic labels. The second stage (Sec. 4.2) provides models trained to predict semantic occupancy with significantly improved accuracy compared to the state of the art.", + "bbox": [ + 89, + 645, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Evidential Semantic Occupancy Mapping", + "text_level": 1, + "bbox": [ + 89, + 742, + 446, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to the commonly used Bayesian approach [10-12, 19-22, 31] for modeling semantic occupancy grids, ET has an advantage of directly representing uncertainties.", + "bbox": [ + 89, + 763, + 483, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, we need to define the frame of discernment, that models possible states, which we here define per voxel as being free or occupied by a certain class:", + "bbox": [ + 89, + 810, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega \\stackrel {\\text {d e f}} {=} \\left\\{\\underset {C \\text {l a b e l e d c l a s s e s}} {\\mathrm {f r e e}}, \\underbrace {\\mathrm {c} _ {1} , \\mathrm {c} _ {2} , \\dots , \\mathrm {c} _ {C}} \\right\\} = \\left\\{\\mathrm {f}, \\mathrm {c} _ {1}, \\mathrm {c} _ {2}, \\dots , \\mathrm {c} _ {C} \\right\\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 861, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/861b4c48065a8764f9dd42ea0e7113fdc4b5c42e49eae94005787733fba669b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Voxel stateHypothesis ωDirect Measurement z(ω)
FreeF = {f}z(F) = αtˆt
Occupied byithclassCi = {ci}z(Ci) = αrˆrci
OccupiedO = ∪iCiz(O) = αrˆunlabeled
UncertainΩ = O∪Fz(Ω) = 0
", + "bbox": [ + 514, + 88, + 903, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Overview of modeled hypotheses and direct measurements. The table shows our selected hypotheses from the power set $2^{\\Omega}$ . We define the hypothesis $\\mathcal{F}$ to model free voxels. Each labeled class has its own hypothesis $\\mathcal{C}_i$ . The hypothesis $\\mathcal{O}$ models voxels, which are known to be occupied with unknown semantics. The hypothesis $\\Omega$ describes total uncertainty about the voxel state. We directly associate every LiDAR measurement processed by $z(\\cdot)$ to the corresponding hypotheses. There are no measurements for $\\Omega$ . $\\alpha_t$ and $\\alpha_r$ act as sensor dependent-hyperparameters.", + "bbox": [ + 511, + 172, + 906, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The set of hypotheses is then the power set of these states. Obviously, simply modeling the belief mass for each possible hypothesis $\\omega \\in 2^{\\Omega}$ quickly becomes infeasible since the number of hypotheses $N = |2^{\\Omega}| = 2^{|\\Omega|}$ grows exponentially. For instance, the nuScenes dataset [2] contains 17 different classes resulting in a cardinality of $|\\Omega_{\\mathrm{nuScenes}}| = 18$ and $N_{\\mathrm{nuScenes}} = 2^{18} = 262$ , 144 possible hypotheses. Therefore, we propose an approximation through the most important subset that contains the \"free\" hypothesis $\\mathcal{F}$ , the \"occupied by class $i$ \" hypotheses $\\mathcal{C}_i$ , the \"generally occupied\" $\\mathcal{O} = \\cup_{i=1}^{C} \\{c_i\\}$ hypothesis, and the $\\Omega$ hypothesis, which corresponds to general uncertainty.", + "bbox": [ + 511, + 325, + 906, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The partly annotated LiDAR measurements are represented as several Cartesian voxel grids (see Sec. 3.2): transmissions $\\hat{t}$ , reflections from a certain class $\\hat{r}_{c_i}$ and unlabeled reflections $\\hat{r}_{\\mathrm{unlabled}}$ . As the next step, we need to define how to map these measurements to the hypotheses established above. To that end, we introduce a simple measurement function $z\\colon 2^{\\Omega}\\to [0,\\inf)$ , which defines sensor measurements, directly providing evidence for a hypothesis but not any particular subset of it. The measurement function $z$ returns a value greater than or equal to zero corresponding to the number of transmissions or reflections supporting the hypothesis in question, scaled with factors $\\alpha_{t}\\in \\mathbb{R}_{\\geq 0}$ and $\\alpha_{r}\\in \\mathbb{R}_{\\geq 0}$ . These factors act as sensor-dependent hyperparameters, similar to $p_{\\mathrm{FN}}$ and $p_{\\mathrm{FP}}$ in [15]. We provide an overview of the hypotheses and the associated direct measurements in Table 1.", + "bbox": [ + 511, + 507, + 906, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to model the belief of a hypothesis $\\omega \\subset \\Omega$ by considering its supporting and conflicting measurements defined as follows:", + "bbox": [ + 511, + 750, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Supporting measurements $z(X)$ provide evidence for hypothesis $\\omega$ being true if $X \\subseteq \\omega$ . Example: a measurement of a certain class $z(\\mathcal{C}_i)$ supports the hypothesis of being (generally) occupied, since the $\\mathcal{C}_i$ is a subset of the occupied hypothesis $\\mathcal{O}$ .", + "- Conflicting measurements $z(X)$ disprove the hypothesis $\\omega$ if $X \\cap \\omega = \\emptyset$ , i.e., when $X$ and $\\omega$ are mutually exclusive" + ], + "bbox": [ + 513, + 795, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27470", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and cannot be true at the same time. Example: a free measurement $z(\\mathcal{F})$ conflicts with the occupied hypothesis, since $\\mathcal{F}$ is disjoint to the occupied hypothesis $\\mathcal{O}$ .", + "bbox": [ + 102, + 90, + 483, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By summing over all supporting and conflicting measurements individually, we obtain the belief of a hypothesis $\\omega$ as:", + "bbox": [ + 89, + 137, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {b e l} (\\omega) = \\left(1 - \\exp \\left(- \\underbrace {\\sum_ {X \\subseteq \\omega} z (X)} _ {\\text {S u p p o r t i n g m e a s u r e m e n t s}}\\right)\\right) \\exp \\left(- \\underbrace {\\sum_ {X \\cap \\omega = \\emptyset} z (X)} _ {\\text {C o n f l i c t i n g m e a s u r e m e n t s}}\\right) (8) \\\\ = \\left(1 - \\prod_ {X \\subseteq \\omega} e ^ {- z (X)}\\right) \\prod_ {X \\cap \\omega = \\emptyset} e ^ {- z (X)} \\quad \\text {f o r} \\quad \\omega \\subset \\Omega . (9) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 188, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Measurements that are neither supporting nor conflicting do not provide any concrete evidence about the specific hypothesis in question and therefore do not contribute to the belief. Additionally, the belief for $\\omega = \\Omega$ is equal to one by definition. Intuitively, this formulation ensures that the belief (acting as a lower bound that a hypothesis is true) can only be greater than zero if there is any measurement supporting the hypothesis. Furthermore, the belief of a hypothesis can only reach the maximum of one if there are no conflicting measurements.", + "bbox": [ + 89, + 378, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To calculate the belief masses $\\mathbf{m}(\\omega)$ from the beliefs $\\mathrm{bel}(\\omega)$ , we make use of Eq. (2) by evaluating it for each hypothesis $\\omega_{i}$ . This can be formulated as a linear system:", + "bbox": [ + 89, + 530, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} = \\left( \\begin{array}{c} \\operatorname {b e l} \\left(\\omega_ {1}\\right) \\\\ \\operatorname {b e l} \\left(\\omega_ {2}\\right) \\\\ \\vdots \\\\ \\operatorname {b e l} \\left(\\omega_ {n}\\right) \\end{array} \\right) = \\mathbf {S} \\left( \\begin{array}{c} \\mathrm {m} \\left(\\omega_ {1}\\right) \\\\ \\mathrm {m} \\left(\\omega_ {2}\\right) \\\\ \\vdots \\\\ \\mathrm {m} \\left(\\omega_ {n}\\right) \\end{array} \\right) = \\mathbf {S m}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 599, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} \\in \\mathbb {R} ^ {n \\times n}: \\mathbf {S} _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\omega_ {j} \\subseteq \\omega_ {i} \\\\ 0 & \\text {o t h e r w i s e} , \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 674, + 403, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{b}\\in \\mathbb{R}^n$ : belief vector,", + "bbox": [ + 179, + 717, + 336, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{m} \\in \\mathbb{R}^n$ : belief mass vector.", + "bbox": [ + 174, + 736, + 372, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The belief mass vector $\\mathbf{m}$ contains the belief masses $\\mathrm{m}(\\omega_j)$ for all hypotheses. The matrix $\\mathbf{S}$ has ones as elements if the belief mass $\\mathrm{m}(\\omega_j)$ contributes to the belief $\\mathrm{bel}(\\omega_i)$ (i.e., $\\omega_j \\subseteq \\omega_i$ ), and zeros otherwise. For a full definition of the entries, please see the supplemental in Sec. B. $\\mathbf{S}$ is always invertible (see supplemental in Sec. C), and we can calculate the belief mass vector as $\\mathbf{m} = \\mathbf{S}^{-1}\\mathbf{b}$ . In fact, our derived BBA is a generalization of the BBA proposed in [15] where there are only two conflicting measurements, namely", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "transmissions and reflections (without any class information), and the scale values are set to $\\alpha_{r} = -\\log (p_{\\mathrm{FP}})$ and $\\alpha_{t} = -\\log (p_{\\mathrm{FN}})$ .", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evidential Semantic Occupancy Prediction", + "text_level": 1, + "bbox": [ + 511, + 143, + 877, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Towards the goal of predicting semantic occupancy grids from multi-view camera images, we present a novel loss formulation which allows to train models directly with the basic belief assignments obtained from our evidential grid mapping approach.", + "bbox": [ + 511, + 166, + 905, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we use an off-the-shelf occupancy prediction model [10] and only adapt the model outputs. Without changing the network architecture, we interpret its softmax output as a belief mass vector, and supervise it with the ground truth belief masses calculated according to Sec. 4.1. Training such a model requires a custom BBA-appropriate loss function. Several works attempted to define a cross-entropy analogue for BBAs [4-6, 38]. We adopt a differentiable formulation by Dezert and Dambreville [6], and directly use it as a loss between the predicted belief masses $\\mathbf{m}^{\\mathrm{pred}}$ and the ground-truth belief masses $\\mathbf{m}^{\\mathrm{gt}}$ .", + "bbox": [ + 511, + 242, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a single voxel $i$ , the BBA cross-entropy is defined as", + "bbox": [ + 531, + 409, + 905, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right) \\tag {11} \\\\ = \\sum_ {X \\subseteq \\Omega} \\left[ - \\mathrm {m} _ {i} ^ {\\mathrm {g t}} (X) \\left(1 - \\mathrm {u} _ {i} ^ {\\mathrm {g t}} (X)\\right) \\log \\left(\\mathrm {m} _ {i} ^ {\\mathrm {p r e d}} (X)\\right) \\right. \\\\ \\left. \\left. + \\mathrm {u} _ {i} ^ {\\mathrm {g t}} (X) \\left(1 - \\mathrm {m} _ {i} ^ {\\mathrm {p r e d}} (X)\\right) \\right] \\right.. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 443, + 905, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the belief mass vector $\\mathbf{m}_i^{\\mathrm{gt}}$ and $\\mathbf{m}_i^{\\mathrm{pred}}$ the uncertainty vector $\\mathbf{u}_i^{\\mathrm{gt}}$ is needed, which is defined as the difference between the plausibility and belief (see Eq. (4)). This can be easily calculated from $\\mathbf{m}_i^{\\mathrm{gt}}$ using Eq. (2) and Eq. (3).", + "bbox": [ + 511, + 545, + 906, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to its classical counterpart, the definition of the BBA cross-entropy is not symmetric, but it always satisfies the inequality:", + "bbox": [ + 511, + 606, + 905, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right) \\geq \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {g t}}\\right) \\geq 0. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 659, + 905, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, the left equality is achieved, if and only if the two BBAs are equal, i.e., $\\mathbf{m}_i^{\\mathrm{gt}} = \\mathbf{m}_i^{\\mathrm{pred}}$ . Additionally, if the BBAs $\\mathbf{m}_i^{\\mathrm{gt}}$ and $\\mathbf{m}_i^{\\mathrm{pred}}$ are Bayesian (i.e., $\\operatorname{prob}(\\omega) = \\operatorname{bel}(\\omega) = \\operatorname{pl}(\\omega)$ ) and therefore $\\mathfrak{u}(\\omega) = 0$ , $\\forall \\omega \\subseteq \\Omega$ , the $\\mathrm{CE}_{\\mathrm{BBA}}(\\mathbf{m}_i^{\\mathrm{gt}}, \\mathbf{m}_i^{\\mathrm{pred}})$ becomes the classical cross-entropy.", + "bbox": [ + 511, + 686, + 905, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have observed that the distribution of belief masses in the ground-truth is highly imbalanced (see supplemental in Fig. 1). To compensate for that, we introduce the per-voxel weight $w_{i}$ and balance the contribution of individual voxels based on their frequency. The final loss is a weighted sum of the BBAs cross-entropy values of all voxels:", + "bbox": [ + 511, + 766, + 905, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i} w _ {i} \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 864, + 905, + 896 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c79d15bdfc860da2564fd8b6372ed8c8e212e6f41c81b4f2863b9633905610f.jpg", + "image_caption": [ + "Figure 3. Semantic occupancy grid maps. We visualize the ground-truth semantic occupancy grid maps from Occ3D [31] and our method on samples from the nuScenes [2] dataset. Semantic labels are assigned to each voxel based on the BBA (see Sec. 4.2). Additionally, we visualize predictions of BEVDet4D-Occ [10] trained with the semantic occupancy data of Occ3D [31] (baseline) and trained with our framework. Both models use the camera images as input. In the first and second example, the baseline model predicts arbitrary flying voxels, whereas our predictions are much cleaner. In the third example, our model predicts the shape of the trees much better. The last example shows how our GT occupancy is better capable of separating the pedestrians." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 290, + 276 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c19e57d0699cd586f0480eaec120b60311f5a80e20427f0f0b7592df0472351f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 88, + 496, + 276 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3da5e910afb6463b5209ba47215cad12cb64ec249279f08d382e71f39a164abf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 700, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b12cba98482901fbbe2a6a12baf9decaa8ab1a1e74ed4ee8f1a61bc2955f9744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 88, + 906, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88c3cb7bc1e590a04e56bdb57b44b927302d34ddc3b2e62bdfe61bd67dc2b7a5.jpg", + "image_caption": [ + "Figure 4. Birds-eye-view uncertainty. For each of the three scenes, we visualize the input camera images on the top. For each scene, we show the predicted semantics of our model on the left. On the right we depict the predicted uncertainties $\\mathfrak{m}(\\Omega)$ averaged along the height dimension. The model predicts high uncertainties (colored in red) mainly in occluded areas, e.g., behind walls. Clearly visible areas obtain much lower uncertainties colored in blue." + ], + "image_footnote": [], + "bbox": [ + 93, + 383, + 361, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8eb1db8c4dcc784ff0e062f7e012709e14c71315077cd2257312ea10eef206db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 383, + 632, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5af6c9053af8ddbd36951d7c194a1ff56db68d549a569a15aee5e54472ef6efb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 383, + 903, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The weight $w_{i}$ is calculated from the inverse square root of the mean BBA over all voxels $\\mathrm{m}_{\\mathrm{mean}}^{\\mathrm{gt}}$ :", + "bbox": [ + 89, + 633, + 485, + 665 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\sum_ {X \\subseteq \\Omega} \\frac {\\mathrm {m} _ {i} ^ {\\mathrm {g t}} (X)}{\\sqrt {\\mathrm {m} _ {\\text {m e a n}} ^ {\\mathrm {g t}} (X)}} \\quad \\text {w i t h} \\quad \\mathbf {m} _ {\\text {m e a n}} ^ {\\mathrm {g t}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbf {m} _ {i} ^ {\\mathrm {g t}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 672, + 483, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This per-voxel weighting increases the contribution of rare voxels and decreases the contribution of very common voxels, e.g., those with total uncertainty $(\\mathrm{m}_i^{\\mathrm{gt}}(\\Omega) = 1)$ .", + "bbox": [ + 89, + 746, + 483, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decision-making based on BBAs. In order to finally predict each voxel's state, we need to make an informed decision based on the BBA of this voxel. To this end, we calculate the relative entropy $\\mathrm{D_{BBA}(\\mathbf{m}_i||\\mathbf{m}) = CE_{BBA}(\\mathbf{m}_i,\\mathbf{m}) - CE_{BBA}(\\mathbf{m}_i,\\mathbf{m}_i)}$ between our predicted BBA $\\mathbf{m}$ and BBAs $\\mathbf{m}_i$ , where $\\mathbf{m}_i$ is focused on a specific hypothesis such that", + "bbox": [ + 89, + 809, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{m}_i(\\omega) = 1$ if $\\omega = \\omega_{i}$ and 0 otherwise. The BBA $\\mathbf{m}_i$ and the respective hypothesis $\\omega_{i}$ , which minimizes the relative entropy, is selected as the decision [6]:", + "bbox": [ + 511, + 633, + 906, + 679 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\omega^ {*} = \\omega_ {i ^ {*}} \\quad \\text {w i t h} \\quad i ^ {*} = \\underset {i} {\\arg \\min } \\mathrm {D} _ {\\mathrm {B B A}} (\\mathbf {m} _ {i} \\| \\mathbf {m}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 686, + 906, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our particular case, there are two decisions to be made. First, we need to decide if a voxel is free or occupied by anything (class-agnostic). Second, we need to determine the semantic classes of all occupied voxels. This leads to the following decision rules:", + "bbox": [ + 511, + 719, + 906, + 794 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Occupancy state: We define $\\mathrm{m_{occ}}(\\omega) = 1$ if $\\omega = \\mathcal{O}$ and 0 otherwise. Similarly, we define $\\mathrm{m}_{\\mathrm{free}}(\\omega) = 1$ if $\\omega = \\mathcal{F}$ and 0 otherwise. A voxel is considered occupied if $\\mathrm{D_{BBA}}(\\mathbf{m_{occ}}||\\mathbf{m}) < \\mathrm{D_{BBA}}(\\mathbf{m_{free}}||\\mathbf{m})$ and free otherwise.", + "- Semantic class: For each class $i$ , we define $\\mathrm{m}_i(\\omega) = 1$ if $\\omega = \\mathbf{c}_i$ , and 0 otherwise. An occupied voxel is assigned the semantic class $c_i$ if $\\mathrm{D_{BBA}}(\\mathbf{m}_i||\\mathbf{m}) < \\mathrm{D_{BBA}}(\\mathbf{m}_j||\\mathbf{m})$ ." + ], + "bbox": [ + 513, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27472", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for all $j \\neq i$ , i.e., the relative entropy between the BBA $\\mathbf{m}_i$ and the BBA $\\mathbf{m}$ is minimal.", + "bbox": [ + 102, + 90, + 482, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that this decision step is only applied before the final evaluation, which requires a discrete class for each voxel. During model training we directly use the belief masses to supervise our model.", + "bbox": [ + 89, + 121, + 483, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 196, + 225, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our method on the nuScenes dataset [2], which is most common in the field of occupancy prediction [15, 31, 32, 34, 35]. Additional results for the Waymo dataset [30] are provided in the supplementary material in Sec. F.", + "bbox": [ + 89, + 220, + 485, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 89, + 292, + 282, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The evaluation of occupancy grids is a challenging task, since a perfect ground-truth occupancy grid is not available. However, we can evaluate the quality of different occupancy grid maps by comparing them to the LiDAR measurements, which provide the best common reference and source of ground-truth. We follow the evaluation protocol of [15] and report additional error metrics for the semantic labels.", + "bbox": [ + 89, + 314, + 483, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth Evaluation. To evaluate the geometry of the occupancy grid maps, we compare the measured depth of the LiDAR $d_{i}^{\\mathrm{lidar}}$ scans with the rendered depth from the occupancy grid $d_{i}^{\\mathrm{pred}}$ . To do so, we cast a ray from the LiDAR sensor origin in the direction of the $i^{\\mathrm{th}}$ LiDAR measurement. We terminate the ray at the first intersection with the surface of a voxel predicted to be occupied. We then extract the length of this rendered ray until the intersection point and call it the rendered depth $d_{i}^{\\mathrm{pred}}$ . Finally, we compare the rendered depth $d_{i}^{\\mathrm{pred}}$ to the measured depth $d_{i}^{\\mathrm{lidar}}$ of the LiDAR using established depth metrics. In Tab. 2 we report the mean absolute error (MAE) and root mean squared error (RMSE), as well as the percentage of points with a relative error $\\delta$ below a certain threshold.", + "bbox": [ + 89, + 439, + 485, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Evaluation. Similar to the depth evaluation, we cast a ray from the LiDAR sensor origin in the direction of each LiDAR measurement. Instead of comparing the depths, we compare the semantic labels of the first occupied voxels along each ray $l_i^{\\mathrm{pred}}$ with the semantic labels of the labeled LiDAR scans $l_i^{\\mathrm{lidar}}$ given in the nuScenes dataset [2]. We report the intersection-over-union (IoU) per class, as well as the mean IoU (mIoU) averaged across all classes in Tab. 3. Note that this metric does not take the depth errors into account: if the LiDAR ray happens to have the same label as the first voxel they intersect, this prediction will be counted as correct even though the depth might be wrong.", + "bbox": [ + 89, + 674, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantify both, the depth and the class accuracy at the same time, we calculate the fraction of points which have a relative depth error below a certain threshold $\\delta_{\\mathrm{cls}}$ and have", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the correct semantic label as given by the labeled LiDAR scans. This metric is stricter than the relative depth accuracy, as it requires the correct semantic label in addition to the correct depth.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Model Training", + "text_level": 1, + "bbox": [ + 511, + 167, + 669, + 184 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the impact of the more accurate GT data provided by our method, we compare BEVDet4D-Occ [10, 11] trained on our semantic occupancy data and our proposed supervision to the same model trained on the commonly used Occ3D [31] dataset. We use ResNet-50 [9] as the backbone for BEVDet4D-Occ [10, 11]. The model predicts semantic occupancy grid maps, using the surround-view camera images as model inputs. Importantly, we do not change the model architecture, but only the loss and the number of outputs in the last layer to match the number of hypotheses. Additionally, we use our generated semantic occupancy data instead of Occ3D's data for supervision.", + "bbox": [ + 511, + 191, + 906, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BEVDet4D-Occ uses additional depth supervision during training, which is not required for our occupancy GT. We start training from a model checkpoint trained on Occ3D [31]. We continue training for 24 epochs using the same hyperparameters, but only changing the model supervision to our evidential formulation. Finally, we reduce the learning rate by a factor of 10 and continue training for 24 more epochs. The baseline model follows the same training schedule, but is supervised with Occ3D's data.", + "bbox": [ + 511, + 376, + 908, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evaluation of Semantic Occupancy Grids", + "text_level": 1, + "bbox": [ + 511, + 529, + 867, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Occupancy Ground Truth. We compare our semantic occupancy grids with those of SurroundOcc [35], Scene as Occupancy [32], Occ3D [31], and Käble et al. [15] on the nuScenes dataset [2]. The former methods provide semantic occupancy grids, while the last method provides the state of the art only for binary occupancy grids. Therefore, we can only report the depth evaluation results for [15] and the additional semantic evaluation as described in Sec. 5.1 for all other methods. We show qualitative results in Fig. 3, visualizing the generated occupancy GT from Occ3D [31] and our method. Additionally, we show the corresponding model predictions as well as the front camera images. The depth evaluation results are summarized in Tab. 2. Our method consistently outperforms all baselines across different metrics and voxel grid resolutions. Through our extended hyperparameter search space for $\\alpha_{r}$ and $\\alpha_{t}$ and novel decision-making scheme, we even outperform [15] in terms of geometric accuracy. Tab. 3 provides quantitative results for the ray-based IoU for the semantic classes. Also in this case our method consistently outperforms all the baselines, which validates that our approach for semantic occupancy using evidence theory is superior to simpler heuristic-based methods.", + "bbox": [ + 511, + 553, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27473", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/981b3b2781e3d19469b345591978ec1a080247aa414d512f916df56253ade108.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Voxel sizeRMSEδ<1.25δ<1.252δ<1.253δclsδcls2
in min min min %in %in %in %in %in %in %in %in %in %in %
Occupancy GTSurroundOcc [35]0.52.805.3456.281.188.649.869.072.6
Scn. as Occ. [32]0.52.925.1746.773.486.641.462.370.4
Occ3D [31]0.41.894.3381.589.493.074.278.780.1
Käblle et al. [15]0.41.253.4287.994.696.8---
EvOcc (ours)0.41.212.9890.195.897.683.486.687.2
OpenOccupancy [34]0.22.084.3570.683.689.565.174.277.3
Käblle et al. [15]0.20.912.9992.696.297.7---
EvOcc (ours)0.20.842.6293.896.998.387.188.588.9
Pred.Occ3D [31]0.42.334.3158.889.694.350.474.376.4
EvOcc (ours)0.41.664.1888.694.396.776.679.580.4
", + "bbox": [ + 192, + 88, + 803, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/79b4e3ac60a260efb17420f3be94f6a34be1a8af6203448a8494487eb646258a.jpg", + "table_caption": [ + "Table 2. Depth evaluation. We compare different methods to generate occupancy GT data as well as trained models on the nuScenes [2] dataset, following the evaluation procedure described Sec. 5.1. The voxel size and depth errors are in meters, while the remaining numbers are in percent. Our method provides occupancy maps that match the LiDAR scans much better compared to other methods using a comparable voxel size. We see the same for BEVDet4D-Occ [10] trained with our evidential framework as compared to training on Occ3D [31]." + ], + "table_footnote": [], + "table_body": "
Voxel sizeBarrierBicycleBusCarConstruction vehicleMotorcyclePedestrianTraffic coneTrailerTruckDriveable surfaceOther flatSidewalkTerrainManmadeVegetationMean
Occupancy GTSurroundOcc [35]0.549.612.967.048.134.728.119.39.142.558.179.931.844.445.348.647.441.7
Scn. as Occ. [32]0.550.013.970.752.143.830.320.68.852.763.877.528.738.639.449.748.043.0
Occ3D [31]0.465.519.074.258.048.532.724.614.653.766.484.850.658.956.555.954.551.2
EvOcc (ours)0.472.934.785.177.171.857.545.224.880.083.289.059.666.668.977.777.267.0
OpenOcc. [34]0.266.226.677.967.856.447.136.916.957.574.481.844.351.050.856.453.454.1
EvOcc (ours)0.282.145.289.783.378.269.356.837.184.987.989.070.976.176.582.781.374.4
Pred:Occ3D [31]0.453.615.368.755.233.126.323.510.339.552.882.031.343.143.256.255.543.1
EvOcc (ours)0.461.35.076.669.238.832.125.613.346.963.286.944.053.756.858.646.248.6
", + "bbox": [ + 91, + 385, + 906, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Semantic evaluation. We compare the semantic occupancy grid maps of different semantic occupancy GT reconstructions, as well as those from trained models, to the LiDAR scans of the nuScenes [2] dataset following the evaluation procedure from Sec. 5.1. The voxel size is in meters, while the remaining numbers are in percent. Our method consistently yields more accurate GT semantic occupancy grid maps and leads to better trained models. All numbers are in percent.", + "bbox": [ + 89, + 573, + 906, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic Occupancy Prediction. We compare the model predictions of the models trained on our semantic occupancy data with the models trained on the Occ3D [31] data. We evaluate the same metrics for the model predictions as for the occupancy GT (see Sec. 5.1). We report the quantitative results in Tab. 2 and Tab. 3. BEVDet4D-Occ trained on our data yields more accurate predictions in terms of geometry and semantics. Qualitative results of the model semantic predictions and corresponding GT occupancy grids are shown in Fig. 3. Additionally, we visualize the predicted semantics together with the aggregated uncertainty $\\mathrm{m}(\\Omega)$ along the height dimension in Fig. 4. Furthermore, we compare our ground-truth uncertainty and the predicted uncertainty in Fig. 1.", + "bbox": [ + 89, + 656, + 485, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 654, + 633, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented a novel approach to obtain high-quality 3D semantic occupancy maps using ET. Our approach introduces a new BBA formulation for modeling LiDAR measurements with partial semantic annotations. We compared our generated ground truth to that of other state-of-the-art semantic occupancy datasets and showed that our data matches the LiDAR measurements better by a large margin in terms of geometry and semantics. We also demonstrated that using our ground truth for training the semantic occupancy prediction model BEVDet4D-Occ significantly improves the model's performance. In conclusion, our work sets a new state of the art for semantic occupancy ground truth generation and model training, and therewith makes an important contribution towards autonomous driving.", + "bbox": [ + 511, + 679, + 908, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27474", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stefan Andreas Baur, David Josef Emmerichs, Frank Moosmann, Peter Pinggera, Björn Ommer, and Andreas Geiger. SLIM: Self-supervised LiDAR scene flow and motion segmentation. In ICCV, 2021. 3", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3, 4, 6, 7, 8, 1", + "[3] Arthur P Dempster. A generalization of bayesian inference. Journal of the Royal Statistical Society: Series B (Methodological), 1968. 2, 3", + "[4] Yong Deng. Uncertainty measure in evidence theory. Science China Information Sciences, 63, 2020. 5", + "[5] Jean Dezert. An effective measure of uncertainty of basic belief assignments. 2022 25th International Conference on Information Fusion (FUSION), pages 1-8, 2022.", + "[6] Jean Dezert and Frédéric Dambreville. Cross-entropy and relative entropy of basic belief assignments. 2023 26th International Conference on Information Fusion (FUSION), pages 1–8, 2023. 3, 5, 6", + "[7] Alberto Elfes. Using occupancy grids for mobile robot perception and navigation. Computer, 22(6):46-57, 1989. 2", + "[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In CVPR, 2012. 2", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7", + "[10] Junjie Huang and Guan Huang. BEVDet4D: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2, 4, 5, 6, 7, 8", + "[11] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2, 7", + "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, 2023. 1, 2, 4", + "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, page 0, 2006. 2", + "[14] J. Kim et al. Evidential semantic mapping in off-road environments with uncertainty-aware bayesian kernel inference. IROS, 2024. 3", + "[15] Jonas Käblé, Sascha Wirges, Maxim Tatarchenko, and Eddy Ilg. Accurate training data for occupancy map prediction in automated driving using evidence theory. In CVPR, 2024. 1, 2, 3, 4, 5, 7, 8", + "[16] J. Li et al. Segm: A novel semantic evidential grid map by fusing multiple sensors. In PRCV, 2020. 3", + "[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of reliable depth for multi-view 3D object detection. In AAAI, pages 1477-1485, 2023. 1, 2" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: Monocular 3D semantic scene completion benchmark in street views. arXiv preprint arXiv:2306.09001, 2023. 2", + "[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, Cham, 2022. 1, 2, 4", + "[20] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M. Alvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation, 2023. 2", + "[21] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M. Alvarez. FB-BEV: BEV representation from forward-backward view transformations. In ICCV, pages 6919-6928, 2023. 1, 2", + "[22] Haisong Liu, Yang Chen, Haiguang Wang, Zetong Yang, Tianyu Li, Jia Zeng, Li Chen, Hongyang Li, and Limin Wang. Fully sparse 3d occupancy prediction. In ECCV, 2023. 4", + "[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position embedding transformation for multi-view 3D object detection. In ECCV, pages 531-548. Springer, 2022. 2", + "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2", + "[25] H. P. Moravec. Sensor fusion in certainty grids for mobile robots. Sensor Devices and Systems for Robotics, 9(2):253-276, 1989. 2", + "[26] Sven Richter, Frank Bieder, Sascha Wirges, and Christoph Stiller. A dual evidential top-view representation to model the semantic environment of automated vehicles. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2023. 3", + "[27] S. Richter et al. Semantic evidential grid mapping using monocular and stereo cameras. Sensors, 2021. 3", + "[28] Glenn Shafer. A mathematical theory of evidence. 1976. 2, 3", + "[29] Philippe Smets. Data fusion in the transferable belief model. Proceedings of the Third International Conference on Information Fusion, 1:PS21-PS33 vol.1, 2000. 3", + "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. CVPR, 2019. 2, 7", + "[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3D: A large-scale 3d occupancy prediction benchmark for autonomous driving. In NeurIPS, 2023. 1, 2, 4, 6, 7, 8", + "[32] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27475", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Hongyang Li. Scene as occupancy. In ICCV, 2023. 1, 2, 7, 8", + "[33] Ignacio Vizzo, Tiziano Guadagnino, Jens Behley, and Cyril Stachniss. VDBFusion: Flexible and efficient TSDF integration of range sensor data. Sensors, 22(3):1296, 2022. 2", + "[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. OpenOccupancy: A large scale benchmark for surrounding semantic occupancy perception. In ICCV, 2023. 1, 2, 7, 8", + "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. SurroundOcc: Multi-camera 3d occupancy prediction for autonomous driving. In CVPR, 2023. 1, 2, 7, 8", + "[36] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. Curran, 2021. 2", + "[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 2", + "[38] Hangyu Yan and Yong Deng. An improved belief entropy in evidence theory. IEEE Access, 8:57505-57516, 2020. 5", + "[39] T. Yang and V. Aitken. Evidential mapping for mobile robots with range sensors. IEEE Transactions on Instrumentation and Measurement, 55:1422-1429, 2006. 3" + ], + "bbox": [ + 91, + 90, + 483, + 501 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27476", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_model.json b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bca605089bbbcbece1651a83e258e629255ce2fc --- /dev/null +++ b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.263, + 0.131, + 0.738, + 0.177 + ], + "angle": 0, + "content": "EvOcc: Accurate Semantic Occupancy for Automated Driving Using Evidence Theory" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.203, + 0.496, + 0.24 + ], + "angle": 0, + "content": "Jonas Käblé\\(^{1,2}\\) Sascha Wirges\\(^{1}\\) \n\\(^{1}\\)Bosch Center for Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.243, + 0.496, + 0.256 + ], + "angle": 0, + "content": "firstname_lastname@de.bosch.com" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.203, + 0.816, + 0.24 + ], + "angle": 0, + "content": "Maxim Tatarchenko1 Eddy Ilg2 \n2University of Technology Nuremberg" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.243, + 0.646, + 0.256 + ], + "angle": 0, + "content": "eddy.ilg@utn.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.568 + ], + "angle": 0, + "content": "We present EvOcc, a novel framework for evidential semantic occupancy mapping. It consists of two parts: (1) an evidential approach for calculating the ground-truth 3D semantic occupancy maps from noisy LiDAR measurements, and (2) a method for training image-based occupancy estimation models through a new loss formulation. In contrast to state-of-the-art semantic occupancy maps, our approach explicitly models the uncertainty introduced by unobserved spaces or contradicting measurements, and we show that using it results in significantly stronger models. Evaluated as ray-based mIoU, our evidential semantic occupancy mapping approach improves over the baselines by at least \\(15.8\\%\\) for the ground truth and \\(5.5\\%\\) for the trained model. Overall, we make a significant contribution towards more detailed and uncertainty-aware 3D environment understanding and safe operation in autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.222, + 0.617 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.486, + 0.734 + ], + "angle": 0, + "content": "A critical component for safe automated driving is accurate and reliable environment perception, which includes understanding the scene geometry as well as interpreting its semantic content. While various 3D representations have been explored in the past, semantic voxel grids have become a recent de facto standard for ML-based environment perception [15, 31, 32, 34, 35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "For a long time, automated driving relied on LiDAR sensors due to their low and distance-independent measurement errors. However, recent methods attempt to reconstruct the 3D environment from camera images alone [10-12, 15, 17, 19-21, 24], since LiDAR is sparse, expensive, and redundant. This makes predicting accurate environment maps from images an important task, and further brings up the problem of obtaining ground-truth training data. The latter is usually produced by recording LiDAR measurements, and employing heuristics to convert them into semantic occupancy maps. However, such reconstructed occupancy maps" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.906, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.406, + 0.712, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.517, + 0.712, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.406, + 0.905, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.517, + 0.905, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.641, + 0.909, + 0.767 + ], + "angle": 0, + "content": "Figure 1. Semantic occupancy and birds-eye-view uncertainty. Our model uses multi-view camera images as input and predicts semantic occupancy. We visualize the ground-truth semantic occupancy obtained from LiDAR scans on left and the predicted semantic occupancy on the right. Below we show the GT uncertainty obtained from our semantic basic belief assignment (BBA) using the LiDAR scans as well as the predicted uncertainties. High uncertainties (e.g., due to occlusions) are colored in red, whereas areas of low uncertainty are colored in blue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "are often inaccurate. A recent work by Käblé et al. [15] has shown that the accuracy can in fact be improved by a margin of \\(30 - 50\\%\\) when simple heuristics are replaced with the framework of evidential mapping. Different from a classical probabilistic interpretation, this approach explicitly models conflicting hypotheses and unknown states, which is crucial for representing noisy real-world measurements." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27467" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "Nevertheless, the authors only model the voxel occupancy state from LiDAR scans, without any notion of semantic classes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.138, + 0.483, + 0.317 + ], + "angle": 0, + "content": "In this work, we present \\(EvOcc\\), an evidence theory (ET) [3, 28] based framework for (1) accurately reconstructing semantic occupancy maps from LiDAR scans (see Fig. 2) and (2) using these as ground-truth to train multi-camera vision models to predict evidential occupancy maps (see Fig. 1). To this end, we extend the evidential mapping framework from Käblé et al. [15] to include semantic class labels. We derive a novel basic belief assignment (BBA) formulation that can handle noisy and partially annotated LiDAR measurements. We also employ a new loss for training semantic occupancy prediction models with our ground-truth data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.472 + ], + "angle": 0, + "content": "Semantic occupancy maps produced by our approach are far superior to those from state-of-the-art semantic occupancy methods [31, 32, 34, 35], which we demonstrate both qualitatively and quantitatively. Using our reconstructed evidential occupancy maps as training data, we are able to train evidential occupancy models that significantly outperform those trained on previously available ground truth. Note that this large improvement is achieved by only changing the model supervision, while keeping the model architecture the same. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.473, + 0.483, + 0.517 + ], + "angle": 0, + "content": "- We present a complete 3D evidential mapping framework for semantic occupancy consisting of two parts: occupancy GT generation and image-based occupancy estimation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.518, + 0.484, + 0.577 + ], + "angle": 0, + "content": "- We apply a novel BBA formulation based on noisy, partially annotated LiDAR measurements, and employ a novel loss that directly uses the obtained evidential semantic occupancy maps as supervision." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.578, + 0.483, + 0.651 + ], + "angle": 0, + "content": "- We demonstrate both qualitatively and quantitatively, that the semantic occupancy maps produced by our trained model and mapping approach are significantly more accurate than those from state-of-the-art methods in terms of agreement with LiDAR measurements (see Sec. 5.3)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.473, + 0.484, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.654, + 0.316, + 0.668 + ], + "angle": 0, + "content": "Our code is available on GitHub1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.233, + 0.7 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.876 + ], + "angle": 0, + "content": "3D Occupancy Datasets. Training perception models for automated driving requires large-scale multimodal datasets, such as KITTI [8], nuScenes [2], Waymo [30] and Agro-verse2 [36]. These datasets record the surrounding environment of the vehicle with multiple onboard sensors, such as LiDAR and cameras. Additionally, they provide annotations for the raw sensor data for tasks like 3D object detection [17, 19, 23, 24] and semantic LiDAR segmentation [12, 37]. The task of camera-only semantic occupancy prediction [18, 31, 32, 34, 35] is to predict the occupancy state and class of each voxel in the environment of the vehi" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.768, + 0.107 + ], + "angle": 0, + "content": "cle, given the onboard camera images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.908, + 0.439 + ], + "angle": 0, + "content": "A common process for obtaining ground-truth (GT) 3D occupancy grids is aggregating LiDAR measurements over multiple time steps under the consideration of moving objects [31, 32, 34, 35]. The resulting aggregated point clouds are then partitioned into dense voxel grids, where each voxel containing a point is labeled according to the point's annotation. Finally, ray casting is used to determine the free space labels between the LiDAR origin and the first occupied voxel. The quality of the occupancy grids can be improved through additional processing steps. For instance, Occ3D [31] and SurroundOcc [35] propose to make the point cloud more dense using mesh and surface reconstruction [13, 33]. Furthermore, Occ3D enforces consistency between the camera images and the occupancy grid through image-guided voxel refinement. OpenOccupancy [34] improves the label quality with additional human supervision and their Augment and Purifying process. Käblé et al. [15] leverage evidence theory (ET) and achieve the current state of the art by proposing to model the occupancy state of each voxel based on the number of transmissions and reflections through it. Our work builds up on [15] and proposes a framework using ET that can model beliefs for semantic classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.761 + ], + "angle": 0, + "content": "3D Semantic Occupancy Prediction. The task of 3D semantic occupancy prediction is to estimate the state of each voxel in the environment of the vehicle using only camera images. A voxel can either be free or occupied by a specific class. The CVPR 2023 Occupancy Prediction Challenge [31, 32] is a popular benchmark for this task. The winner of this challenge, FB-OCC [20, 21], extends BEVDet [11] by a forward-backward view transformation module. TPVFormer [12] addresses the high memory and compute requirements for a dense 3D grid and projects the 3D space onto 3D planes. Additional depth supervision for the process of splatting 2D image features into the 3D space has been introduced by BEVDepth [17]. BEVDet4D [10] aggregates additional information over time for the task of 3D object detection. In this work, we use BEVDet4D-Occ-R50 as a baseline model, which is an adaptation of BEVDet4D specifically tailored for the task of semantic occupancy prediction and commonly used as reference model. We show that training this model on the new ground truth obtained with our approach gives superior results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Occupancy Grid Mapping. Static occupancy grid mapping using a Bayesian interpretation of probabilities is a well-explored area in robotics [7, 25]. The state of a voxel cell is modeled as a Bernoulli-distributed random variable and can either be occupied or free. Since the state of the cell is defined by a single probability, the model does not consider the uncertainty in the measurements to estimate the cell state, nor does it handle conflicts between different mea" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.416, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27468" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.087, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.255, + 0.778, + 0.272 + ], + "angle": 0, + "content": "Figure 2. Overview. A scheme of our evidential framework for semantic occupancy mapping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.486, + 0.584 + ], + "angle": 0, + "content": "suresments. To circumvent these limitations, Yang and Aitken [39] propose to model the state of a cell using ET [3, 28], which can explicitly represent the uncertainty. Richter et al. [27] and Li et al. [16] model the 2D birds-eye-view (BEV) occupancy maps using ET. More recent works extend the mapping task to 3D and consider dynamic environments as well as semantics. There are some specialized models for automated driving scenarios. Richter et al. [26] split dynamic traffic scenes into two evidential BEV grid maps containing the ground plane and everything on top of the ground. Kim et al. [14] present a 3D semantic occupancy mapping approach, which fuses unreliable semantic predictions using ET. Furthermore, improvements in the mapping process for dynamic scenes can be achieved by considering the motion of objects using scene flow from LiDAR measurements [1]. A framework for 3D occupancy grid mapping based on ET within a dynamic environment is presented in [15]. However, the presented method is limited to occupancy maps without semantic information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.599, + 0.231, + 0.615 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.624, + 0.258, + 0.641 + ], + "angle": 0, + "content": "3.1. Evidence Theory" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.486, + 0.753 + ], + "angle": 0, + "content": "Here, we briefly revisit the fundamentals of evidence theory (ET), or Dempster-Shafer theory [3, 28], used to model the semantic occupancy state of individual voxels (see Sec. 4.1). ET provides a flexible framework for reasoning about the uncertainty, be it total uncertainty or partial knowledge about a system [29]. A single hypothesis is described by a subset of the frame of discernment \\(\\Omega\\), and the BBA" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.764, + 0.484, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathrm {m} \\colon 2 ^ {\\Omega} \\rightarrow [ 0, 1 ] \\quad \\text {w i t h} \\quad \\sum_ {\\omega \\subseteq \\Omega} \\mathrm {m} (\\omega) = 1 \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.839 + ], + "angle": 0, + "content": "assigns a normalized belief mass \\(\\mathrm{m}(\\omega)\\) to each hypothesis \\(\\omega \\subseteq \\Omega\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Intuitively, the belief mass represents the proportion of relevant and available evidence supporting the hypothesis \\(\\omega\\) but not any particular subset of it. The classical probability of a hypothesis \\(\\omega\\) lies between two values known as belief" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.698, + 0.311 + ], + "angle": 0, + "content": "and plausibility. The belief" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.322, + 0.907, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\operatorname {b e l} (\\omega) = \\sum_ {X \\subseteq \\omega} \\mathrm {m} (X) \\leq \\operatorname {p r o b} (\\omega) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.412 + ], + "angle": 0, + "content": "of a hypothesis \\(\\omega\\) is the sum of all belief masses supporting the hypothesis \\(\\omega\\). It serves as a lower bound of the probability that a certain hypothesis is true. The plausibility" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.423, + 0.907, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\operatorname {p l} (\\omega) = 1 - \\sum_ {X \\cap \\omega = \\emptyset} \\mathrm {m} (X) \\geq \\operatorname {p r o b} (\\omega) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.906, + 0.512 + ], + "angle": 0, + "content": "is an upper bound for this probability. It can be defined as the complement of all belief masses not supporting the hypothesis \\(\\omega\\). The uncertainty" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.523, + 0.907, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mathrm {u} (\\omega) = \\operatorname {p l} (\\omega) - \\operatorname {b e l} (\\omega) \\in [ 0, 1 ] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.583 + ], + "angle": 0, + "content": "of a hypothesis \\(\\omega\\) can then be defined as the discrepancy between plausibility and belief [6]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.755, + 0.608 + ], + "angle": 0, + "content": "3.2. Spherical Sensor Mapping" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Käblé et al. [15] describe an approach to determine the state of a voxel based on the transmissions, and reflections through it. These are measured by a LiDAR sensor and provide evidence for a voxel being free or occupied by a certain class. Different from their work, we address the multi-class case here." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.907, + 0.829 + ], + "angle": 0, + "content": "Reflections Grid Map. Instead of mapping all reflections into a single spherical grid map, we create a separate spherical reflections grid map \\(r_{\\mathrm{ci}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)\\) for each class \\(\\mathbf{c}_i\\). On top of that, we have one additional reflections grid map \\(r_{\\mathrm{unlabeled}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)\\) for all unlabeled reflections2. This results in a total of \\(C + 1\\) spherical reflection grid maps, where \\(C\\) is the number of classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.848, + 0.907, + 0.878 + ], + "angle": 0, + "content": "Transmissions Grid Map. Since transmissions represent free space and are independent of the reflections' semantic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.864, + 0.901 + ], + "angle": 0, + "content": "2In nuScenes [2] only one of every 10 LiDAR scans is labeled." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "classes, we only produce a single spherical transmission map \\(t^{\\mathrm{ph}}(\\rho ,\\phi ,\\theta)\\). For a given voxel cell located at \\((\\rho ,\\theta ,\\phi)\\), the number of transmissions:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.14, + 0.49, + 0.214 + ], + "angle": 0, + "content": "\\[\nt ^ {\\mathrm {s p h}} (\\rho , \\phi , \\theta) = \\sum_ {\\rho^ {\\prime} > \\rho} \\underbrace {\\left(r _ {\\text {u n l a b e l e d}} ^ {\\mathrm {s p h}} \\left(\\rho^ {\\prime} , \\phi , \\theta\\right) + \\sum_ {i = 1} ^ {C} r _ {\\mathrm {c} _ {i}} ^ {\\mathrm {s p h}} \\left(\\rho^ {\\prime} , \\phi , \\theta\\right)\\right)} _ {\\text {S u m o v e r l a b e l e d a n d u n l a b e l e d r e f l e c t i o n s}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.485, + 0.281 + ], + "angle": 0, + "content": "is given by the sum over all reflections with the radius \\(\\rho'\\) greater than \\(\\rho\\) of the voxel under consideration. This can be efficiently implemented as a cumulative sum from the maximal to the minimal radial distance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.391 + ], + "angle": 0, + "content": "Grid Map Normalization. We normalize the number of reflections and transmissions per voxel by their respective volume. We obtain the normalized reflections \\(\\hat{r}_{\\mathrm{ci}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)\\) and \\(\\hat{r}_{\\mathrm{unlabled}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)\\) as well as the normalized transmissions \\(\\hat{t}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)\\) for each time step separately. Details can be found in the Supplemental in Sec. A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Multi-Frame Aggregation. We aggregate the reflections and transmissions from multiple time steps into a single Cartesian voxel grid. Using known per-object transformations between the reference frame at time step \\( n \\) and frames at all other time steps \\( k \\), we can accumulate the final per-voxel values as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.502, + 0.484, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\hat {r} _ {\\mathrm {c} _ {i}, n} (x, y, z) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} g \\left(\\hat {r} _ {\\mathrm {c} _ {i}, k} ^ {\\text {s p h}} (\\rho , \\phi , \\theta)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.61 + ], + "angle": 0, + "content": "where \\( g \\) applies the motion compensation and transforms spherical coordinates into Cartesian ones. The same applies to the unlabeled reflections \\( \\hat{r}_{\\mathrm{unlabled}}(x,y,z) \\) and transmissions \\( \\hat{t}(x,y,z) \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.182, + 0.635 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.736 + ], + "angle": 0, + "content": "Here we present our EvOcc framework split into two stages. The first stage (Sec. 4.1) reconstructs accurate semantic occupancy maps from sparse LiDAR scans with partially annotated semantic labels. The second stage (Sec. 4.2) provides models trained to predict semantic occupancy with significantly improved accuracy compared to the state of the art." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.447, + 0.76 + ], + "angle": 0, + "content": "4.1. Evidential Semantic Occupancy Mapping" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.811 + ], + "angle": 0, + "content": "In contrast to the commonly used Bayesian approach [10-12, 19-22, 31] for modeling semantic occupancy grids, ET has an advantage of directly representing uncertainties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Initially, we need to define the frame of discernment, that models possible states, which we here define per voxel as being free or occupied by a certain class:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.862, + 0.484, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\Omega \\stackrel {\\text {d e f}} {=} \\left\\{\\underset {C \\text {l a b e l e d c l a s s e s}} {\\mathrm {f r e e}}, \\underbrace {\\mathrm {c} _ {1} , \\mathrm {c} _ {2} , \\dots , \\mathrm {c} _ {C}} \\right\\} = \\left\\{\\mathrm {f}, \\mathrm {c} _ {1}, \\mathrm {c} _ {2}, \\dots , \\mathrm {c} _ {C} \\right\\}. \\tag {7}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.905, + 0.163 + ], + "angle": 0, + "content": "
Voxel stateHypothesis ωDirect Measurement z(ω)
FreeF = {f}z(F) = αtˆt
Occupied byithclassCi = {ci}z(Ci) = αrˆrci
OccupiedO = ∪iCiz(O) = αrˆunlabeled
UncertainΩ = O∪Fz(Ω) = 0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.173, + 0.908, + 0.299 + ], + "angle": 0, + "content": "Table 1. Overview of modeled hypotheses and direct measurements. The table shows our selected hypotheses from the power set \\(2^{\\Omega}\\). We define the hypothesis \\(\\mathcal{F}\\) to model free voxels. Each labeled class has its own hypothesis \\(\\mathcal{C}_i\\). The hypothesis \\(\\mathcal{O}\\) models voxels, which are known to be occupied with unknown semantics. The hypothesis \\(\\Omega\\) describes total uncertainty about the voxel state. We directly associate every LiDAR measurement processed by \\(z(\\cdot)\\) to the corresponding hypotheses. There are no measurements for \\(\\Omega\\). \\(\\alpha_t\\) and \\(\\alpha_r\\) act as sensor dependent-hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.908, + 0.508 + ], + "angle": 0, + "content": "The set of hypotheses is then the power set of these states. Obviously, simply modeling the belief mass for each possible hypothesis \\(\\omega \\in 2^{\\Omega}\\) quickly becomes infeasible since the number of hypotheses \\(N = |2^{\\Omega}| = 2^{|\\Omega|}\\) grows exponentially. For instance, the nuScenes dataset [2] contains 17 different classes resulting in a cardinality of \\(|\\Omega_{\\mathrm{nuScenes}}| = 18\\) and \\(N_{\\mathrm{nuScenes}} = 2^{18} = 262\\), 144 possible hypotheses. Therefore, we propose an approximation through the most important subset that contains the \"free\" hypothesis \\(\\mathcal{F}\\), the \"occupied by class \\(i\\)\" hypotheses \\(\\mathcal{C}_i\\), the \"generally occupied\" \\(\\mathcal{O} = \\cup_{i=1}^{C} \\{c_i\\}\\) hypothesis, and the \\(\\Omega\\) hypothesis, which corresponds to general uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.908, + 0.748 + ], + "angle": 0, + "content": "The partly annotated LiDAR measurements are represented as several Cartesian voxel grids (see Sec. 3.2): transmissions \\(\\hat{t}\\), reflections from a certain class \\(\\hat{r}_{c_i}\\) and unlabeled reflections \\(\\hat{r}_{\\mathrm{unlabled}}\\). As the next step, we need to define how to map these measurements to the hypotheses established above. To that end, we introduce a simple measurement function \\(z\\colon 2^{\\Omega}\\to [0,\\inf)\\), which defines sensor measurements, directly providing evidence for a hypothesis but not any particular subset of it. The measurement function \\(z\\) returns a value greater than or equal to zero corresponding to the number of transmissions or reflections supporting the hypothesis in question, scaled with factors \\(\\alpha_{t}\\in \\mathbb{R}_{\\geq 0}\\) and \\(\\alpha_{r}\\in \\mathbb{R}_{\\geq 0}\\). These factors act as sensor-dependent hyperparameters, similar to \\(p_{\\mathrm{FN}}\\) and \\(p_{\\mathrm{FP}}\\) in [15]. We provide an overview of the hypotheses and the associated direct measurements in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.794 + ], + "angle": 0, + "content": "We propose to model the belief of a hypothesis \\(\\omega \\subset \\Omega\\) by considering its supporting and conflicting measurements defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.796, + 0.908, + 0.87 + ], + "angle": 0, + "content": "- Supporting measurements \\( z(X) \\) provide evidence for hypothesis \\( \\omega \\) being true if \\( X \\subseteq \\omega \\). Example: a measurement of a certain class \\( z(\\mathcal{C}_i) \\) supports the hypothesis of being (generally) occupied, since the \\( \\mathcal{C}_i \\) is a subset of the occupied hypothesis \\( \\mathcal{O} \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Conflicting measurements \\( z(X) \\) disprove the hypothesis \\( \\omega \\) if \\( X \\cap \\omega = \\emptyset \\), i.e., when \\( X \\) and \\( \\omega \\) are mutually exclusive" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.796, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.484, + 0.137 + ], + "angle": 0, + "content": "and cannot be true at the same time. Example: a free measurement \\( z(\\mathcal{F}) \\) conflicts with the occupied hypothesis, since \\( \\mathcal{F} \\) is disjoint to the occupied hypothesis \\( \\mathcal{O} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.182 + ], + "angle": 0, + "content": "By summing over all supporting and conflicting measurements individually, we obtain the belief of a hypothesis \\(\\omega\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.189, + 0.484, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {b e l} (\\omega) = \\left(1 - \\exp \\left(- \\underbrace {\\sum_ {X \\subseteq \\omega} z (X)} _ {\\text {S u p p o r t i n g m e a s u r e m e n t s}}\\right)\\right) \\exp \\left(- \\underbrace {\\sum_ {X \\cap \\omega = \\emptyset} z (X)} _ {\\text {C o n f l i c t i n g m e a s u r e m e n t s}}\\right) (8) \\\\ = \\left(1 - \\prod_ {X \\subseteq \\omega} e ^ {- z (X)}\\right) \\prod_ {X \\cap \\omega = \\emptyset} e ^ {- z (X)} \\quad \\text {f o r} \\quad \\omega \\subset \\Omega . (9) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.529 + ], + "angle": 0, + "content": "Measurements that are neither supporting nor conflicting do not provide any concrete evidence about the specific hypothesis in question and therefore do not contribute to the belief. Additionally, the belief for \\(\\omega = \\Omega\\) is equal to one by definition. Intuitively, this formulation ensures that the belief (acting as a lower bound that a hypothesis is true) can only be greater than zero if there is any measurement supporting the hypothesis. Furthermore, the belief of a hypothesis can only reach the maximum of one if there are no conflicting measurements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.577 + ], + "angle": 0, + "content": "To calculate the belief masses \\(\\mathbf{m}(\\omega)\\) from the beliefs \\(\\mathrm{bel}(\\omega)\\), we make use of Eq. (2) by evaluating it for each hypothesis \\(\\omega_{i}\\). This can be formulated as a linear system:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.6, + 0.484, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} = \\left( \\begin{array}{c} \\operatorname {b e l} \\left(\\omega_ {1}\\right) \\\\ \\operatorname {b e l} \\left(\\omega_ {2}\\right) \\\\ \\vdots \\\\ \\operatorname {b e l} \\left(\\omega_ {n}\\right) \\end{array} \\right) = \\mathbf {S} \\left( \\begin{array}{c} \\mathrm {m} \\left(\\omega_ {1}\\right) \\\\ \\mathrm {m} \\left(\\omega_ {2}\\right) \\\\ \\vdots \\\\ \\mathrm {m} \\left(\\omega_ {n}\\right) \\end{array} \\right) = \\mathbf {S m}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.675, + 0.405, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} \\in \\mathbb {R} ^ {n \\times n}: \\mathbf {S} _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\omega_ {j} \\subseteq \\omega_ {i} \\\\ 0 & \\text {o t h e r w i s e} , \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.718, + 0.338, + 0.732 + ], + "angle": 0, + "content": "\\(\\mathbf{b}\\in \\mathbb{R}^n\\) : belief vector," + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.737, + 0.374, + 0.751 + ], + "angle": 0, + "content": "\\(\\mathbf{m} \\in \\mathbb{R}^n\\): belief mass vector." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The belief mass vector \\(\\mathbf{m}\\) contains the belief masses \\(\\mathrm{m}(\\omega_j)\\) for all hypotheses. The matrix \\(\\mathbf{S}\\) has ones as elements if the belief mass \\(\\mathrm{m}(\\omega_j)\\) contributes to the belief \\(\\mathrm{bel}(\\omega_i)\\) (i.e., \\(\\omega_j \\subseteq \\omega_i\\)), and zeros otherwise. For a full definition of the entries, please see the supplemental in Sec. B. \\(\\mathbf{S}\\) is always invertible (see supplemental in Sec. C), and we can calculate the belief mass vector as \\(\\mathbf{m} = \\mathbf{S}^{-1}\\mathbf{b}\\). In fact, our derived BBA is a generalization of the BBA proposed in [15] where there are only two conflicting measurements, namely" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "transmissions and reflections (without any class information), and the scale values are set to \\(\\alpha_{r} = -\\log (p_{\\mathrm{FP}})\\) and \\(\\alpha_{t} = -\\log (p_{\\mathrm{FN}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.145, + 0.879, + 0.162 + ], + "angle": 0, + "content": "4.2. Evidential Semantic Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.906, + 0.243 + ], + "angle": 0, + "content": "Towards the goal of predicting semantic occupancy grids from multi-view camera images, we present a novel loss formulation which allows to train models directly with the basic belief assignments obtained from our evidential grid mapping approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Therefore, we use an off-the-shelf occupancy prediction model [10] and only adapt the model outputs. Without changing the network architecture, we interpret its softmax output as a belief mass vector, and supervise it with the ground truth belief masses calculated according to Sec. 4.1. Training such a model requires a custom BBA-appropriate loss function. Several works attempted to define a cross-entropy analogue for BBAs [4-6, 38]. We adopt a differentiable formulation by Dezert and Dambreville [6], and directly use it as a loss between the predicted belief masses \\(\\mathbf{m}^{\\mathrm{pred}}\\) and the ground-truth belief masses \\(\\mathbf{m}^{\\mathrm{gt}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.41, + 0.906, + 0.424 + ], + "angle": 0, + "content": "For a single voxel \\(i\\), the BBA cross-entropy is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.444, + 0.906, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right) \\tag {11} \\\\ = \\sum_ {X \\subseteq \\Omega} \\left[ - \\mathrm {m} _ {i} ^ {\\mathrm {g t}} (X) \\left(1 - \\mathrm {u} _ {i} ^ {\\mathrm {g t}} (X)\\right) \\log \\left(\\mathrm {m} _ {i} ^ {\\mathrm {p r e d}} (X)\\right) \\right. \\\\ \\left. \\left. + \\mathrm {u} _ {i} ^ {\\mathrm {g t}} (X) \\left(1 - \\mathrm {m} _ {i} ^ {\\mathrm {p r e d}} (X)\\right) \\right] \\right.. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.607 + ], + "angle": 0, + "content": "In addition to the belief mass vector \\(\\mathbf{m}_i^{\\mathrm{gt}}\\) and \\(\\mathbf{m}_i^{\\mathrm{pred}}\\) the uncertainty vector \\(\\mathbf{u}_i^{\\mathrm{gt}}\\) is needed, which is defined as the difference between the plausibility and belief (see Eq. (4)). This can be easily calculated from \\(\\mathbf{m}_i^{\\mathrm{gt}}\\) using Eq. (2) and Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.653 + ], + "angle": 0, + "content": "Similar to its classical counterpart, the definition of the BBA cross-entropy is not symmetric, but it always satisfies the inequality:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.66, + 0.906, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right) \\geq \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {g t}}\\right) \\geq 0. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.767 + ], + "angle": 0, + "content": "Importantly, the left equality is achieved, if and only if the two BBAs are equal, i.e., \\(\\mathbf{m}_i^{\\mathrm{gt}} = \\mathbf{m}_i^{\\mathrm{pred}}\\). Additionally, if the BBAs \\(\\mathbf{m}_i^{\\mathrm{gt}}\\) and \\(\\mathbf{m}_i^{\\mathrm{pred}}\\) are Bayesian (i.e., \\(\\operatorname{prob}(\\omega) = \\operatorname{bel}(\\omega) = \\operatorname{pl}(\\omega)\\)) and therefore \\(\\mathfrak{u}(\\omega) = 0\\), \\(\\forall \\omega \\subseteq \\Omega\\), the \\(\\mathrm{CE}_{\\mathrm{BBA}}(\\mathbf{m}_i^{\\mathrm{gt}}, \\mathbf{m}_i^{\\mathrm{pred}})\\) becomes the classical cross-entropy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.906, + 0.857 + ], + "angle": 0, + "content": "We have observed that the distribution of belief masses in the ground-truth is highly imbalanced (see supplemental in Fig. 1). To compensate for that, we introduce the per-voxel weight \\( w_{i} \\) and balance the contribution of individual voxels based on their frequency. The final loss is a weighted sum of the BBAs cross-entropy values of all voxels:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.866, + 0.906, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i} w _ {i} \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right). \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.292, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.089, + 0.498, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.702, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.089, + 0.908, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.289, + 0.908, + 0.373 + ], + "angle": 0, + "content": "Figure 3. Semantic occupancy grid maps. We visualize the ground-truth semantic occupancy grid maps from Occ3D [31] and our method on samples from the nuScenes [2] dataset. Semantic labels are assigned to each voxel based on the BBA (see Sec. 4.2). Additionally, we visualize predictions of BEVDet4D-Occ [10] trained with the semantic occupancy data of Occ3D [31] (baseline) and trained with our framework. Both models use the camera images as input. In the first and second example, the baseline model predicts arbitrary flying voxels, whereas our predictions are much cleaner. In the third example, our model predicts the shape of the trees much better. The last example shows how our GT occupancy is better capable of separating the pedestrians." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.384, + 0.362, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.384, + 0.633, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.384, + 0.905, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.552, + 0.908, + 0.607 + ], + "angle": 0, + "content": "Figure 4. Birds-eye-view uncertainty. For each of the three scenes, we visualize the input camera images on the top. For each scene, we show the predicted semantics of our model on the left. On the right we depict the predicted uncertainties \\(\\mathfrak{m}(\\Omega)\\) averaged along the height dimension. The model predicts high uncertainties (colored in red) mainly in occluded areas, e.g., behind walls. Clearly visible areas obtain much lower uncertainties colored in blue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.486, + 0.666 + ], + "angle": 0, + "content": "The weight \\( w_{i} \\) is calculated from the inverse square root of the mean BBA over all voxels \\( \\mathrm{m}_{\\mathrm{mean}}^{\\mathrm{gt}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.674, + 0.484, + 0.737 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\sum_ {X \\subseteq \\Omega} \\frac {\\mathrm {m} _ {i} ^ {\\mathrm {g t}} (X)}{\\sqrt {\\mathrm {m} _ {\\text {m e a n}} ^ {\\mathrm {g t}} (X)}} \\quad \\text {w i t h} \\quad \\mathbf {m} _ {\\text {m e a n}} ^ {\\mathrm {g t}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbf {m} _ {i} ^ {\\mathrm {g t}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.485, + 0.793 + ], + "angle": 0, + "content": "This per-voxel weighting increases the contribution of rare voxels and decreases the contribution of very common voxels, e.g., those with total uncertainty \\((\\mathrm{m}_i^{\\mathrm{gt}}(\\Omega) = 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Decision-making based on BBAs. In order to finally predict each voxel's state, we need to make an informed decision based on the BBA of this voxel. To this end, we calculate the relative entropy \\(\\mathrm{D_{BBA}(\\mathbf{m}_i||\\mathbf{m}) = CE_{BBA}(\\mathbf{m}_i,\\mathbf{m}) - CE_{BBA}(\\mathbf{m}_i,\\mathbf{m}_i)}\\) between our predicted BBA \\(\\mathbf{m}\\) and BBAs \\(\\mathbf{m}_i\\), where \\(\\mathbf{m}_i\\) is focused on a specific hypothesis such that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.68 + ], + "angle": 0, + "content": "\\(\\mathbf{m}_i(\\omega) = 1\\) if \\(\\omega = \\omega_{i}\\) and 0 otherwise. The BBA \\(\\mathbf{m}_i\\) and the respective hypothesis \\(\\omega_{i}\\), which minimizes the relative entropy, is selected as the decision [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.688, + 0.907, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\omega^ {*} = \\omega_ {i ^ {*}} \\quad \\text {w i t h} \\quad i ^ {*} = \\underset {i} {\\arg \\min } \\mathrm {D} _ {\\mathrm {B B A}} (\\mathbf {m} _ {i} \\| \\mathbf {m}). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.908, + 0.795 + ], + "angle": 0, + "content": "In our particular case, there are two decisions to be made. First, we need to decide if a voxel is free or occupied by anything (class-agnostic). Second, we need to determine the semantic classes of all occupied voxels. This leads to the following decision rules:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.796, + 0.907, + 0.856 + ], + "angle": 0, + "content": "- Occupancy state: We define \\(\\mathrm{m_{occ}}(\\omega) = 1\\) if \\(\\omega = \\mathcal{O}\\) and 0 otherwise. Similarly, we define \\(\\mathrm{m}_{\\mathrm{free}}(\\omega) = 1\\) if \\(\\omega = \\mathcal{F}\\) and 0 otherwise. A voxel is considered occupied if \\(\\mathrm{D_{BBA}}(\\mathbf{m_{occ}}||\\mathbf{m}) < \\mathrm{D_{BBA}}(\\mathbf{m_{free}}||\\mathbf{m})\\) and free otherwise." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "- Semantic class: For each class \\( i \\), we define \\( \\mathrm{m}_i(\\omega) = 1 \\) if \\( \\omega = \\mathbf{c}_i \\), and 0 otherwise. An occupied voxel is assigned the semantic class \\( c_i \\) if \\( \\mathrm{D_{BBA}}(\\mathbf{m}_i||\\mathbf{m}) < \\mathrm{D_{BBA}}(\\mathbf{m}_j||\\mathbf{m}) \\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "for all \\( j \\neq i \\), i.e., the relative entropy between the BBA \\( \\mathbf{m}_i \\) and the BBA \\( \\mathbf{m} \\) is minimal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.485, + 0.184 + ], + "angle": 0, + "content": "Note that this decision step is only applied before the final evaluation, which requires a discrete class for each voxel. During model training we directly use the belief masses to supervise our model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.197, + 0.226, + 0.214 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.486, + 0.285 + ], + "angle": 0, + "content": "We evaluate our method on the nuScenes dataset [2], which is most common in the field of occupancy prediction [15, 31, 32, 34, 35]. Additional results for the Waymo dataset [30] are provided in the supplementary material in Sec. F." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.293, + 0.283, + 0.308 + ], + "angle": 0, + "content": "5.1. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.485, + 0.421 + ], + "angle": 0, + "content": "The evaluation of occupancy grids is a challenging task, since a perfect ground-truth occupancy grid is not available. However, we can evaluate the quality of different occupancy grid maps by comparing them to the LiDAR measurements, which provide the best common reference and source of ground-truth. We follow the evaluation protocol of [15] and report additional error metrics for the semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.486, + 0.655 + ], + "angle": 0, + "content": "Depth Evaluation. To evaluate the geometry of the occupancy grid maps, we compare the measured depth of the LiDAR \\(d_{i}^{\\mathrm{lidar}}\\) scans with the rendered depth from the occupancy grid \\(d_{i}^{\\mathrm{pred}}\\). To do so, we cast a ray from the LiDAR sensor origin in the direction of the \\(i^{\\mathrm{th}}\\) LiDAR measurement. We terminate the ray at the first intersection with the surface of a voxel predicted to be occupied. We then extract the length of this rendered ray until the intersection point and call it the rendered depth \\(d_{i}^{\\mathrm{pred}}\\). Finally, we compare the rendered depth \\(d_{i}^{\\mathrm{pred}}\\) to the measured depth \\(d_{i}^{\\mathrm{lidar}}\\) of the LiDAR using established depth metrics. In Tab. 2 we report the mean absolute error (MAE) and root mean squared error (RMSE), as well as the percentage of points with a relative error \\(\\delta\\) below a certain threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Semantic Evaluation. Similar to the depth evaluation, we cast a ray from the LiDAR sensor origin in the direction of each LiDAR measurement. Instead of comparing the depths, we compare the semantic labels of the first occupied voxels along each ray \\( l_i^{\\mathrm{pred}} \\) with the semantic labels of the labeled LiDAR scans \\( l_i^{\\mathrm{lidar}} \\) given in the nuScenes dataset [2]. We report the intersection-over-union (IoU) per class, as well as the mean IoU (mIoU) averaged across all classes in Tab. 3. Note that this metric does not take the depth errors into account: if the LiDAR ray happens to have the same label as the first voxel they intersect, this prediction will be counted as correct even though the depth might be wrong." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "To quantify both, the depth and the class accuracy at the same time, we calculate the fraction of points which have a relative depth error below a certain threshold \\(\\delta_{\\mathrm{cls}}\\) and have" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.152 + ], + "angle": 0, + "content": "the correct semantic label as given by the labeled LiDAR scans. This metric is stricter than the relative depth accuracy, as it requires the correct semantic label in addition to the correct depth." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.169, + 0.671, + 0.185 + ], + "angle": 0, + "content": "5.2. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.193, + 0.907, + 0.375 + ], + "angle": 0, + "content": "To demonstrate the impact of the more accurate GT data provided by our method, we compare BEVDet4D-Occ [10, 11] trained on our semantic occupancy data and our proposed supervision to the same model trained on the commonly used Occ3D [31] dataset. We use ResNet-50 [9] as the backbone for BEVDet4D-Occ [10, 11]. The model predicts semantic occupancy grid maps, using the surround-view camera images as model inputs. Importantly, we do not change the model architecture, but only the loss and the number of outputs in the last layer to match the number of hypotheses. Additionally, we use our generated semantic occupancy data instead of Occ3D's data for supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.909, + 0.513 + ], + "angle": 0, + "content": "BEVDet4D-Occ uses additional depth supervision during training, which is not required for our occupancy GT. We start training from a model checkpoint trained on Occ3D [31]. We continue training for 24 epochs using the same hyperparameters, but only changing the model supervision to our evidential formulation. Finally, we reduce the learning rate by a factor of 10 and continue training for 24 more epochs. The baseline model follows the same training schedule, but is supervised with Occ3D's data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.868, + 0.546 + ], + "angle": 0, + "content": "5.3. Evaluation of Semantic Occupancy Grids" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Semantic Occupancy Ground Truth. We compare our semantic occupancy grids with those of SurroundOcc [35], Scene as Occupancy [32], Occ3D [31], and Käble et al. [15] on the nuScenes dataset [2]. The former methods provide semantic occupancy grids, while the last method provides the state of the art only for binary occupancy grids. Therefore, we can only report the depth evaluation results for [15] and the additional semantic evaluation as described in Sec. 5.1 for all other methods. We show qualitative results in Fig. 3, visualizing the generated occupancy GT from Occ3D [31] and our method. Additionally, we show the corresponding model predictions as well as the front camera images. The depth evaluation results are summarized in Tab. 2. Our method consistently outperforms all baselines across different metrics and voxel grid resolutions. Through our extended hyperparameter search space for \\(\\alpha_{r}\\) and \\(\\alpha_{t}\\) and novel decision-making scheme, we even outperform [15] in terms of geometric accuracy. Tab. 3 provides quantitative results for the ray-based IoU for the semantic classes. Also in this case our method consistently outperforms all the baselines, which validates that our approach for semantic occupancy using evidence theory is superior to simpler heuristic-based methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27473" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.194, + 0.089, + 0.805, + 0.306 + ], + "angle": 0, + "content": "
Voxel sizeRMSEδ<1.25δ<1.252δ<1.253δclsδcls2
in min min min %in %in %in %in %in %in %in %in %in %in %
Occupancy GTSurroundOcc [35]0.52.805.3456.281.188.649.869.072.6
Scn. as Occ. [32]0.52.925.1746.773.486.641.462.370.4
Occ3D [31]0.41.894.3381.589.493.074.278.780.1
Käblle et al. [15]0.41.253.4287.994.696.8---
EvOcc (ours)0.41.212.9890.195.897.683.486.687.2
OpenOccupancy [34]0.22.084.3570.683.689.565.174.277.3
Käblle et al. [15]0.20.912.9992.696.297.7---
EvOcc (ours)0.20.842.6293.896.998.387.188.588.9
Pred.Occ3D [31]0.42.334.3158.889.694.350.474.376.4
EvOcc (ours)0.41.664.1888.694.396.776.679.580.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.317, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Table 2. Depth evaluation. We compare different methods to generate occupancy GT data as well as trained models on the nuScenes [2] dataset, following the evaluation procedure described Sec. 5.1. The voxel size and depth errors are in meters, while the remaining numbers are in percent. Our method provides occupancy maps that match the LiDAR scans much better compared to other methods using a comparable voxel size. We see the same for BEVDet4D-Occ [10] trained with our evidential framework as compared to training on Occ3D [31]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.386, + 0.907, + 0.564 + ], + "angle": 0, + "content": "
Voxel sizeBarrierBicycleBusCarConstruction vehicleMotorcyclePedestrianTraffic coneTrailerTruckDriveable surfaceOther flatSidewalkTerrainManmadeVegetationMean
Occupancy GTSurroundOcc [35]0.549.612.967.048.134.728.119.39.142.558.179.931.844.445.348.647.441.7
Scn. as Occ. [32]0.550.013.970.752.143.830.320.68.852.763.877.528.738.639.449.748.043.0
Occ3D [31]0.465.519.074.258.048.532.724.614.653.766.484.850.658.956.555.954.551.2
EvOcc (ours)0.472.934.785.177.171.857.545.224.880.083.289.059.666.668.977.777.267.0
OpenOcc. [34]0.266.226.677.967.856.447.136.916.957.574.481.844.351.050.856.453.454.1
EvOcc (ours)0.282.145.289.783.378.269.356.837.184.987.989.070.976.176.582.781.374.4
Pred:Occ3D [31]0.453.615.368.755.233.126.323.510.339.552.882.031.343.143.256.255.543.1
EvOcc (ours)0.461.35.076.669.238.832.125.613.346.963.286.944.053.756.858.646.248.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.574, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Table 3. Semantic evaluation. We compare the semantic occupancy grid maps of different semantic occupancy GT reconstructions, as well as those from trained models, to the LiDAR scans of the nuScenes [2] dataset following the evaluation procedure from Sec. 5.1. The voxel size is in meters, while the remaining numbers are in percent. Our method consistently yields more accurate GT semantic occupancy grid maps and leads to better trained models. All numbers are in percent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.486, + 0.869 + ], + "angle": 0, + "content": "Semantic Occupancy Prediction. We compare the model predictions of the models trained on our semantic occupancy data with the models trained on the Occ3D [31] data. We evaluate the same metrics for the model predictions as for the occupancy GT (see Sec. 5.1). We report the quantitative results in Tab. 2 and Tab. 3. BEVDet4D-Occ trained on our data yields more accurate predictions in terms of geometry and semantics. Qualitative results of the model semantic predictions and corresponding GT occupancy grids are shown in Fig. 3. Additionally, we visualize the predicted semantics together with the aggregated uncertainty \\(\\mathrm{m}(\\Omega)\\) along the height dimension in Fig. 4. Furthermore, we compare our ground-truth uncertainty and the predicted uncertainty in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.655, + 0.634, + 0.671 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.909, + 0.893 + ], + "angle": 0, + "content": "In this work, we presented a novel approach to obtain high-quality 3D semantic occupancy maps using ET. Our approach introduces a new BBA formulation for modeling LiDAR measurements with partial semantic annotations. We compared our generated ground truth to that of other state-of-the-art semantic occupancy datasets and showed that our data matches the LiDAR measurements better by a large margin in terms of geometry and semantics. We also demonstrated that using our ground truth for training the semantic occupancy prediction model BEVDet4D-Occ significantly improves the model's performance. In conclusion, our work sets a new state of the art for semantic occupancy ground truth generation and model training, and therewith makes an important contribution towards autonomous driving." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27474" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.17 + ], + "angle": 0, + "content": "[1] Stefan Andreas Baur, David Josef Emmerichs, Frank Moosmann, Peter Pinggera, Björn Ommer, and Andreas Geiger. SLIM: Self-supervised LiDAR scene flow and motion segmentation. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.485, + 0.239 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3, 4, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.241, + 0.485, + 0.282 + ], + "angle": 0, + "content": "[3] Arthur P Dempster. A generalization of bayesian inference. Journal of the Royal Statistical Society: Series B (Methodological), 1968. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.283, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[4] Yong Deng. Uncertainty measure in evidence theory. Science China Information Sciences, 63, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.353 + ], + "angle": 0, + "content": "[5] Jean Dezert. An effective measure of uncertainty of basic belief assignments. 2022 25th International Conference on Information Fusion (FUSION), pages 1-8, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[6] Jean Dezert and Frédéric Dambreville. Cross-entropy and relative entropy of basic belief assignments. 2023 26th International Conference on Information Fusion (FUSION), pages 1–8, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[7] Alberto Elfes. Using occupancy grids for mobile robot perception and navigation. Computer, 22(6):46-57, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In CVPR, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.484, + 0.519 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.522, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[10] Junjie Huang and Guan Huang. BEVDet4D: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.565, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[11] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, page 0, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[14] J. Kim et al. Evidential semantic mapping in off-road environments with uncertainty-aware bayesian kernel inference. IROS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[15] Jonas Käblé, Sascha Wirges, Maxim Tatarchenko, and Eddy Ilg. Accurate training data for occupancy map prediction in automated driving using evidence theory. In CVPR, 2024. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[16] J. Li et al. Segm: A novel semantic evidential grid map by fusing multiple sensors. In PRCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of reliable depth for multi-view 3D object detection. In AAAI, pages 1477-1485, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.174 + ], + "angle": 0, + "content": "[18] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: Monocular 3D semantic scene completion benchmark in street views. arXiv preprint arXiv:2306.09001, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.176, + 0.908, + 0.243 + ], + "angle": 0, + "content": "[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, Cham, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.907, + 0.299 + ], + "angle": 0, + "content": "[20] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M. Alvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.907, + 0.356 + ], + "angle": 0, + "content": "[21] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M. Alvarez. FB-BEV: BEV representation from forward-backward view transformations. In ICCV, pages 6919-6928, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.907, + 0.398 + ], + "angle": 0, + "content": "[22] Haisong Liu, Yang Chen, Haiguang Wang, Zetong Yang, Tianyu Li, Jia Zeng, Li Chen, Hongyang Li, and Limin Wang. Fully sparse 3d occupancy prediction. In ECCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position embedding transformation for multi-view 3D object detection. In ECCV, pages 531-548. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.522 + ], + "angle": 0, + "content": "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.524, + 0.907, + 0.564 + ], + "angle": 0, + "content": "[25] H. P. Moravec. Sensor fusion in certainty grids for mobile robots. Sensor Devices and Systems for Robotics, 9(2):253-276, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.621 + ], + "angle": 0, + "content": "[26] Sven Richter, Frank Bieder, Sascha Wirges, and Christoph Stiller. A dual evidential top-view representation to model the semantic environment of automated vehicles. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.649 + ], + "angle": 0, + "content": "[27] S. Richter et al. Semantic evidential grid mapping using monocular and stereo cameras. Sensors, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.906, + 0.664 + ], + "angle": 0, + "content": "[28] Glenn Shafer. A mathematical theory of evidence. 1976. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[29] Philippe Smets. Data fusion in the transferable belief model. Proceedings of the Third International Conference on Information Fusion, 1:PS21-PS33 vol.1, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.908, + 0.816 + ], + "angle": 0, + "content": "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. CVPR, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3D: A large-scale 3d occupancy prediction benchmark for autonomous driving. In NeurIPS, 2023. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[32] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "and Hongyang Li. Scene as occupancy. In ICCV, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[33] Ignacio Vizzo, Tiziano Guadagnino, Jens Behley, and Cyril Stachniss. VDBFusion: Flexible and efficient TSDF integration of range sensor data. Sensors, 22(3):1296, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. OpenOccupancy: A large scale benchmark for surrounding semantic occupancy perception. In ICCV, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. SurroundOcc: Multi-camera 3d occupancy prediction for autonomous driving. In CVPR, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[36] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. Curran, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[38] Hangyu Yan and Yong Deng. An improved belief entropy in evidence theory. IEEE Access, 8:57505-57516, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[39] T. Yang and V. Aitken. Evidential mapping for mobile robots with range sensors. IEEE Transactions on Instrumentation and Measurement, 55:1422-1429, 2006. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27476" + } + ] +] \ No newline at end of file diff --git a/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_origin.pdf b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..67a2ea305ffbbe810f0a05adba8307668072e866 --- /dev/null +++ b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/7c502132-611b-4140-a7e5-11680d62e1b1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55dfa0ce9f98b8d79642e138cd8b4904869f2886e1e0753857a9fd88b160e40b +size 6125866 diff --git a/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/full.md b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aebe1f57517447ec80f67ac3878174efea9eaed2 --- /dev/null +++ b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/full.md @@ -0,0 +1,331 @@ +# EvOcc: Accurate Semantic Occupancy for Automated Driving Using Evidence Theory + +Jonas Käblé $^{1,2}$ Sascha Wirges $^{1}$ $^{1}$ Bosch Center for Artificial Intelligence + +firstname_lastname@de.bosch.com + +Maxim Tatarchenko1 Eddy Ilg2 +2University of Technology Nuremberg + +eddy.ilg@utn.de + +# Abstract + +We present EvOcc, a novel framework for evidential semantic occupancy mapping. It consists of two parts: (1) an evidential approach for calculating the ground-truth 3D semantic occupancy maps from noisy LiDAR measurements, and (2) a method for training image-based occupancy estimation models through a new loss formulation. In contrast to state-of-the-art semantic occupancy maps, our approach explicitly models the uncertainty introduced by unobserved spaces or contradicting measurements, and we show that using it results in significantly stronger models. Evaluated as ray-based mIoU, our evidential semantic occupancy mapping approach improves over the baselines by at least $15.8\%$ for the ground truth and $5.5\%$ for the trained model. Overall, we make a significant contribution towards more detailed and uncertainty-aware 3D environment understanding and safe operation in autonomous driving. + +# 1. Introduction + +A critical component for safe automated driving is accurate and reliable environment perception, which includes understanding the scene geometry as well as interpreting its semantic content. While various 3D representations have been explored in the past, semantic voxel grids have become a recent de facto standard for ML-based environment perception [15, 31, 32, 34, 35]. + +For a long time, automated driving relied on LiDAR sensors due to their low and distance-independent measurement errors. However, recent methods attempt to reconstruct the 3D environment from camera images alone [10-12, 15, 17, 19-21, 24], since LiDAR is sparse, expensive, and redundant. This makes predicting accurate environment maps from images an important task, and further brings up the problem of obtaining ground-truth training data. The latter is usually produced by recording LiDAR measurements, and employing heuristics to convert them into semantic occupancy maps. However, such reconstructed occupancy maps + +![](images/b4817f692080b6b73f5f72534e4bffde3718da7596df33fd253fe3662e6a4c53.jpg) + +![](images/1011a8cf653cb5a2f062ae2732285b1309358a8a08a7f8cca39aff296f369c11.jpg) + +![](images/73d08c7bbdf454641943c857213fa40d6579c46fbfda248dfd6aca0442925c42.jpg) +Figure 1. Semantic occupancy and birds-eye-view uncertainty. Our model uses multi-view camera images as input and predicts semantic occupancy. We visualize the ground-truth semantic occupancy obtained from LiDAR scans on left and the predicted semantic occupancy on the right. Below we show the GT uncertainty obtained from our semantic basic belief assignment (BBA) using the LiDAR scans as well as the predicted uncertainties. High uncertainties (e.g., due to occlusions) are colored in red, whereas areas of low uncertainty are colored in blue. + +![](images/72416b7e24146e874ca4ad1dfa61dd2c27f258ecdf4d931f24f157568093cb3c.jpg) + +![](images/5ef55350608bfcc444883b360bb94bdfd9cf0ff522d5a945459841de98bdd764.jpg) + +are often inaccurate. A recent work by Käblé et al. [15] has shown that the accuracy can in fact be improved by a margin of $30 - 50\%$ when simple heuristics are replaced with the framework of evidential mapping. Different from a classical probabilistic interpretation, this approach explicitly models conflicting hypotheses and unknown states, which is crucial for representing noisy real-world measurements. + +Nevertheless, the authors only model the voxel occupancy state from LiDAR scans, without any notion of semantic classes. + +In this work, we present $EvOcc$ , an evidence theory (ET) [3, 28] based framework for (1) accurately reconstructing semantic occupancy maps from LiDAR scans (see Fig. 2) and (2) using these as ground-truth to train multi-camera vision models to predict evidential occupancy maps (see Fig. 1). To this end, we extend the evidential mapping framework from Käblé et al. [15] to include semantic class labels. We derive a novel basic belief assignment (BBA) formulation that can handle noisy and partially annotated LiDAR measurements. We also employ a new loss for training semantic occupancy prediction models with our ground-truth data. + +Semantic occupancy maps produced by our approach are far superior to those from state-of-the-art semantic occupancy methods [31, 32, 34, 35], which we demonstrate both qualitatively and quantitatively. Using our reconstructed evidential occupancy maps as training data, we are able to train evidential occupancy models that significantly outperform those trained on previously available ground truth. Note that this large improvement is achieved by only changing the model supervision, while keeping the model architecture the same. In summary, we make the following contributions: + +- We present a complete 3D evidential mapping framework for semantic occupancy consisting of two parts: occupancy GT generation and image-based occupancy estimation. +- We apply a novel BBA formulation based on noisy, partially annotated LiDAR measurements, and employ a novel loss that directly uses the obtained evidential semantic occupancy maps as supervision. +- We demonstrate both qualitatively and quantitatively, that the semantic occupancy maps produced by our trained model and mapping approach are significantly more accurate than those from state-of-the-art methods in terms of agreement with LiDAR measurements (see Sec. 5.3). + +Our code is available on GitHub1. + +# 2. Related Work + +3D Occupancy Datasets. Training perception models for automated driving requires large-scale multimodal datasets, such as KITTI [8], nuScenes [2], Waymo [30] and Agro-verse2 [36]. These datasets record the surrounding environment of the vehicle with multiple onboard sensors, such as LiDAR and cameras. Additionally, they provide annotations for the raw sensor data for tasks like 3D object detection [17, 19, 23, 24] and semantic LiDAR segmentation [12, 37]. The task of camera-only semantic occupancy prediction [18, 31, 32, 34, 35] is to predict the occupancy state and class of each voxel in the environment of the vehi + +cle, given the onboard camera images. + +A common process for obtaining ground-truth (GT) 3D occupancy grids is aggregating LiDAR measurements over multiple time steps under the consideration of moving objects [31, 32, 34, 35]. The resulting aggregated point clouds are then partitioned into dense voxel grids, where each voxel containing a point is labeled according to the point's annotation. Finally, ray casting is used to determine the free space labels between the LiDAR origin and the first occupied voxel. The quality of the occupancy grids can be improved through additional processing steps. For instance, Occ3D [31] and SurroundOcc [35] propose to make the point cloud more dense using mesh and surface reconstruction [13, 33]. Furthermore, Occ3D enforces consistency between the camera images and the occupancy grid through image-guided voxel refinement. OpenOccupancy [34] improves the label quality with additional human supervision and their Augment and Purifying process. Käblé et al. [15] leverage evidence theory (ET) and achieve the current state of the art by proposing to model the occupancy state of each voxel based on the number of transmissions and reflections through it. Our work builds up on [15] and proposes a framework using ET that can model beliefs for semantic classes. + +3D Semantic Occupancy Prediction. The task of 3D semantic occupancy prediction is to estimate the state of each voxel in the environment of the vehicle using only camera images. A voxel can either be free or occupied by a specific class. The CVPR 2023 Occupancy Prediction Challenge [31, 32] is a popular benchmark for this task. The winner of this challenge, FB-OCC [20, 21], extends BEVDet [11] by a forward-backward view transformation module. TPVFormer [12] addresses the high memory and compute requirements for a dense 3D grid and projects the 3D space onto 3D planes. Additional depth supervision for the process of splatting 2D image features into the 3D space has been introduced by BEVDepth [17]. BEVDet4D [10] aggregates additional information over time for the task of 3D object detection. In this work, we use BEVDet4D-Occ-R50 as a baseline model, which is an adaptation of BEVDet4D specifically tailored for the task of semantic occupancy prediction and commonly used as reference model. We show that training this model on the new ground truth obtained with our approach gives superior results. + +Occupancy Grid Mapping. Static occupancy grid mapping using a Bayesian interpretation of probabilities is a well-explored area in robotics [7, 25]. The state of a voxel cell is modeled as a Bernoulli-distributed random variable and can either be occupied or free. Since the state of the cell is defined by a single probability, the model does not consider the uncertainty in the measurements to estimate the cell state, nor does it handle conflicts between different mea + +![](images/30475cc4e711c9c4644d176754cd70ca8ed7683fcd00ff827bd5987cd1d5ff22.jpg) +Figure 2. Overview. A scheme of our evidential framework for semantic occupancy mapping. + +suresments. To circumvent these limitations, Yang and Aitken [39] propose to model the state of a cell using ET [3, 28], which can explicitly represent the uncertainty. Richter et al. [27] and Li et al. [16] model the 2D birds-eye-view (BEV) occupancy maps using ET. More recent works extend the mapping task to 3D and consider dynamic environments as well as semantics. There are some specialized models for automated driving scenarios. Richter et al. [26] split dynamic traffic scenes into two evidential BEV grid maps containing the ground plane and everything on top of the ground. Kim et al. [14] present a 3D semantic occupancy mapping approach, which fuses unreliable semantic predictions using ET. Furthermore, improvements in the mapping process for dynamic scenes can be achieved by considering the motion of objects using scene flow from LiDAR measurements [1]. A framework for 3D occupancy grid mapping based on ET within a dynamic environment is presented in [15]. However, the presented method is limited to occupancy maps without semantic information. + +# 3. Preliminaries + +# 3.1. Evidence Theory + +Here, we briefly revisit the fundamentals of evidence theory (ET), or Dempster-Shafer theory [3, 28], used to model the semantic occupancy state of individual voxels (see Sec. 4.1). ET provides a flexible framework for reasoning about the uncertainty, be it total uncertainty or partial knowledge about a system [29]. A single hypothesis is described by a subset of the frame of discernment $\Omega$ , and the BBA + +$$ +\mathrm {m} \colon 2 ^ {\Omega} \rightarrow [ 0, 1 ] \quad \text {w i t h} \quad \sum_ {\omega \subseteq \Omega} \mathrm {m} (\omega) = 1 \tag {1} +$$ + +assigns a normalized belief mass $\mathrm{m}(\omega)$ to each hypothesis $\omega \subseteq \Omega$ . + +Intuitively, the belief mass represents the proportion of relevant and available evidence supporting the hypothesis $\omega$ but not any particular subset of it. The classical probability of a hypothesis $\omega$ lies between two values known as belief + +and plausibility. The belief + +$$ +\operatorname {b e l} (\omega) = \sum_ {X \subseteq \omega} \mathrm {m} (X) \leq \operatorname {p r o b} (\omega) \tag {2} +$$ + +of a hypothesis $\omega$ is the sum of all belief masses supporting the hypothesis $\omega$ . It serves as a lower bound of the probability that a certain hypothesis is true. The plausibility + +$$ +\operatorname {p l} (\omega) = 1 - \sum_ {X \cap \omega = \emptyset} \mathrm {m} (X) \geq \operatorname {p r o b} (\omega) \tag {3} +$$ + +is an upper bound for this probability. It can be defined as the complement of all belief masses not supporting the hypothesis $\omega$ . The uncertainty + +$$ +\mathrm {u} (\omega) = \operatorname {p l} (\omega) - \operatorname {b e l} (\omega) \in [ 0, 1 ] \tag {4} +$$ + +of a hypothesis $\omega$ can then be defined as the discrepancy between plausibility and belief [6]. + +# 3.2. Spherical Sensor Mapping + +Käblé et al. [15] describe an approach to determine the state of a voxel based on the transmissions, and reflections through it. These are measured by a LiDAR sensor and provide evidence for a voxel being free or occupied by a certain class. Different from their work, we address the multi-class case here. + +Reflections Grid Map. Instead of mapping all reflections into a single spherical grid map, we create a separate spherical reflections grid map $r_{\mathrm{ci}}^{\mathrm{sph}}(\rho, \phi, \theta)$ for each class $\mathbf{c}_i$ . On top of that, we have one additional reflections grid map $r_{\mathrm{unlabeled}}^{\mathrm{sph}}(\rho, \phi, \theta)$ for all unlabeled reflections2. This results in a total of $C + 1$ spherical reflection grid maps, where $C$ is the number of classes. + +Transmissions Grid Map. Since transmissions represent free space and are independent of the reflections' semantic + +classes, we only produce a single spherical transmission map $t^{\mathrm{ph}}(\rho ,\phi ,\theta)$ . For a given voxel cell located at $(\rho ,\theta ,\phi)$ , the number of transmissions: + +$$ +t ^ {\mathrm {s p h}} (\rho , \phi , \theta) = \sum_ {\rho^ {\prime} > \rho} \underbrace {\left(r _ {\text {u n l a b e l e d}} ^ {\mathrm {s p h}} \left(\rho^ {\prime} , \phi , \theta\right) + \sum_ {i = 1} ^ {C} r _ {\mathrm {c} _ {i}} ^ {\mathrm {s p h}} \left(\rho^ {\prime} , \phi , \theta\right)\right)} _ {\text {S u m o v e r l a b e l e d a n d u n l a b e l e d r e f l e c t i o n s}} \tag {5} +$$ + +is given by the sum over all reflections with the radius $\rho'$ greater than $\rho$ of the voxel under consideration. This can be efficiently implemented as a cumulative sum from the maximal to the minimal radial distance. + +Grid Map Normalization. We normalize the number of reflections and transmissions per voxel by their respective volume. We obtain the normalized reflections $\hat{r}_{\mathrm{ci}}^{\mathrm{sph}}(\rho, \phi, \theta)$ and $\hat{r}_{\mathrm{unlabled}}^{\mathrm{sph}}(\rho, \phi, \theta)$ as well as the normalized transmissions $\hat{t}^{\mathrm{sph}}(\rho, \phi, \theta)$ for each time step separately. Details can be found in the Supplemental in Sec. A. + +Multi-Frame Aggregation. We aggregate the reflections and transmissions from multiple time steps into a single Cartesian voxel grid. Using known per-object transformations between the reference frame at time step $n$ and frames at all other time steps $k$ , we can accumulate the final per-voxel values as: + +$$ +\hat {r} _ {\mathrm {c} _ {i}, n} (x, y, z) = \frac {1}{K} \sum_ {k = 1} ^ {K} g \left(\hat {r} _ {\mathrm {c} _ {i}, k} ^ {\text {s p h}} (\rho , \phi , \theta)\right), \tag {6} +$$ + +where $g$ applies the motion compensation and transforms spherical coordinates into Cartesian ones. The same applies to the unlabeled reflections $\hat{r}_{\mathrm{unlabled}}(x,y,z)$ and transmissions $\hat{t}(x,y,z)$ . + +# 4. Method + +Here we present our EvOcc framework split into two stages. The first stage (Sec. 4.1) reconstructs accurate semantic occupancy maps from sparse LiDAR scans with partially annotated semantic labels. The second stage (Sec. 4.2) provides models trained to predict semantic occupancy with significantly improved accuracy compared to the state of the art. + +# 4.1. Evidential Semantic Occupancy Mapping + +In contrast to the commonly used Bayesian approach [10-12, 19-22, 31] for modeling semantic occupancy grids, ET has an advantage of directly representing uncertainties. + +Initially, we need to define the frame of discernment, that models possible states, which we here define per voxel as being free or occupied by a certain class: + +$$ +\Omega \stackrel {\text {d e f}} {=} \left\{\underset {C \text {l a b e l e d c l a s s e s}} {\mathrm {f r e e}}, \underbrace {\mathrm {c} _ {1} , \mathrm {c} _ {2} , \dots , \mathrm {c} _ {C}} \right\} = \left\{\mathrm {f}, \mathrm {c} _ {1}, \mathrm {c} _ {2}, \dots , \mathrm {c} _ {C} \right\}. \tag {7} +$$ + +
Voxel stateHypothesis ωDirect Measurement z(ω)
FreeF = {f}z(F) = αtˆt
Occupied byithclassCi = {ci}z(Ci) = αrˆrci
OccupiedO = ∪iCiz(O) = αrˆunlabeled
UncertainΩ = O∪Fz(Ω) = 0
+ +Table 1. Overview of modeled hypotheses and direct measurements. The table shows our selected hypotheses from the power set $2^{\Omega}$ . We define the hypothesis $\mathcal{F}$ to model free voxels. Each labeled class has its own hypothesis $\mathcal{C}_i$ . The hypothesis $\mathcal{O}$ models voxels, which are known to be occupied with unknown semantics. The hypothesis $\Omega$ describes total uncertainty about the voxel state. We directly associate every LiDAR measurement processed by $z(\cdot)$ to the corresponding hypotheses. There are no measurements for $\Omega$ . $\alpha_t$ and $\alpha_r$ act as sensor dependent-hyperparameters. + +The set of hypotheses is then the power set of these states. Obviously, simply modeling the belief mass for each possible hypothesis $\omega \in 2^{\Omega}$ quickly becomes infeasible since the number of hypotheses $N = |2^{\Omega}| = 2^{|\Omega|}$ grows exponentially. For instance, the nuScenes dataset [2] contains 17 different classes resulting in a cardinality of $|\Omega_{\mathrm{nuScenes}}| = 18$ and $N_{\mathrm{nuScenes}} = 2^{18} = 262$ , 144 possible hypotheses. Therefore, we propose an approximation through the most important subset that contains the "free" hypothesis $\mathcal{F}$ , the "occupied by class $i$ " hypotheses $\mathcal{C}_i$ , the "generally occupied" $\mathcal{O} = \cup_{i=1}^{C} \{c_i\}$ hypothesis, and the $\Omega$ hypothesis, which corresponds to general uncertainty. + +The partly annotated LiDAR measurements are represented as several Cartesian voxel grids (see Sec. 3.2): transmissions $\hat{t}$ , reflections from a certain class $\hat{r}_{c_i}$ and unlabeled reflections $\hat{r}_{\mathrm{unlabled}}$ . As the next step, we need to define how to map these measurements to the hypotheses established above. To that end, we introduce a simple measurement function $z\colon 2^{\Omega}\to [0,\inf)$ , which defines sensor measurements, directly providing evidence for a hypothesis but not any particular subset of it. The measurement function $z$ returns a value greater than or equal to zero corresponding to the number of transmissions or reflections supporting the hypothesis in question, scaled with factors $\alpha_{t}\in \mathbb{R}_{\geq 0}$ and $\alpha_{r}\in \mathbb{R}_{\geq 0}$ . These factors act as sensor-dependent hyperparameters, similar to $p_{\mathrm{FN}}$ and $p_{\mathrm{FP}}$ in [15]. We provide an overview of the hypotheses and the associated direct measurements in Table 1. + +We propose to model the belief of a hypothesis $\omega \subset \Omega$ by considering its supporting and conflicting measurements defined as follows: + +- Supporting measurements $z(X)$ provide evidence for hypothesis $\omega$ being true if $X \subseteq \omega$ . Example: a measurement of a certain class $z(\mathcal{C}_i)$ supports the hypothesis of being (generally) occupied, since the $\mathcal{C}_i$ is a subset of the occupied hypothesis $\mathcal{O}$ . +- Conflicting measurements $z(X)$ disprove the hypothesis $\omega$ if $X \cap \omega = \emptyset$ , i.e., when $X$ and $\omega$ are mutually exclusive + +and cannot be true at the same time. Example: a free measurement $z(\mathcal{F})$ conflicts with the occupied hypothesis, since $\mathcal{F}$ is disjoint to the occupied hypothesis $\mathcal{O}$ . + +By summing over all supporting and conflicting measurements individually, we obtain the belief of a hypothesis $\omega$ as: + +$$ +\begin{array}{l} \operatorname {b e l} (\omega) = \left(1 - \exp \left(- \underbrace {\sum_ {X \subseteq \omega} z (X)} _ {\text {S u p p o r t i n g m e a s u r e m e n t s}}\right)\right) \exp \left(- \underbrace {\sum_ {X \cap \omega = \emptyset} z (X)} _ {\text {C o n f l i c t i n g m e a s u r e m e n t s}}\right) (8) \\ = \left(1 - \prod_ {X \subseteq \omega} e ^ {- z (X)}\right) \prod_ {X \cap \omega = \emptyset} e ^ {- z (X)} \quad \text {f o r} \quad \omega \subset \Omega . (9) \\ \end{array} +$$ + +Measurements that are neither supporting nor conflicting do not provide any concrete evidence about the specific hypothesis in question and therefore do not contribute to the belief. Additionally, the belief for $\omega = \Omega$ is equal to one by definition. Intuitively, this formulation ensures that the belief (acting as a lower bound that a hypothesis is true) can only be greater than zero if there is any measurement supporting the hypothesis. Furthermore, the belief of a hypothesis can only reach the maximum of one if there are no conflicting measurements. + +To calculate the belief masses $\mathbf{m}(\omega)$ from the beliefs $\mathrm{bel}(\omega)$ , we make use of Eq. (2) by evaluating it for each hypothesis $\omega_{i}$ . This can be formulated as a linear system: + +$$ +\mathbf {b} = \left( \begin{array}{c} \operatorname {b e l} \left(\omega_ {1}\right) \\ \operatorname {b e l} \left(\omega_ {2}\right) \\ \vdots \\ \operatorname {b e l} \left(\omega_ {n}\right) \end{array} \right) = \mathbf {S} \left( \begin{array}{c} \mathrm {m} \left(\omega_ {1}\right) \\ \mathrm {m} \left(\omega_ {2}\right) \\ \vdots \\ \mathrm {m} \left(\omega_ {n}\right) \end{array} \right) = \mathbf {S m}, \tag {10} +$$ + +$$ +\mathbf {S} \in \mathbb {R} ^ {n \times n}: \mathbf {S} _ {i j} = \left\{ \begin{array}{l l} 1 & \text {i f} \omega_ {j} \subseteq \omega_ {i} \\ 0 & \text {o t h e r w i s e} , \end{array} \right. +$$ + +$\mathbf{b}\in \mathbb{R}^n$ : belief vector, + +$\mathbf{m} \in \mathbb{R}^n$ : belief mass vector. + +The belief mass vector $\mathbf{m}$ contains the belief masses $\mathrm{m}(\omega_j)$ for all hypotheses. The matrix $\mathbf{S}$ has ones as elements if the belief mass $\mathrm{m}(\omega_j)$ contributes to the belief $\mathrm{bel}(\omega_i)$ (i.e., $\omega_j \subseteq \omega_i$ ), and zeros otherwise. For a full definition of the entries, please see the supplemental in Sec. B. $\mathbf{S}$ is always invertible (see supplemental in Sec. C), and we can calculate the belief mass vector as $\mathbf{m} = \mathbf{S}^{-1}\mathbf{b}$ . In fact, our derived BBA is a generalization of the BBA proposed in [15] where there are only two conflicting measurements, namely + +transmissions and reflections (without any class information), and the scale values are set to $\alpha_{r} = -\log (p_{\mathrm{FP}})$ and $\alpha_{t} = -\log (p_{\mathrm{FN}})$ . + +# 4.2. Evidential Semantic Occupancy Prediction + +Towards the goal of predicting semantic occupancy grids from multi-view camera images, we present a novel loss formulation which allows to train models directly with the basic belief assignments obtained from our evidential grid mapping approach. + +Therefore, we use an off-the-shelf occupancy prediction model [10] and only adapt the model outputs. Without changing the network architecture, we interpret its softmax output as a belief mass vector, and supervise it with the ground truth belief masses calculated according to Sec. 4.1. Training such a model requires a custom BBA-appropriate loss function. Several works attempted to define a cross-entropy analogue for BBAs [4-6, 38]. We adopt a differentiable formulation by Dezert and Dambreville [6], and directly use it as a loss between the predicted belief masses $\mathbf{m}^{\mathrm{pred}}$ and the ground-truth belief masses $\mathbf{m}^{\mathrm{gt}}$ . + +For a single voxel $i$ , the BBA cross-entropy is defined as + +$$ +\begin{array}{l} \mathrm {C E} _ {\mathrm {B B A}} \left(\mathbf {m} _ {i} ^ {\mathrm {g t}}, \mathbf {m} _ {i} ^ {\mathrm {p r e d}}\right) \tag {11} \\ = \sum_ {X \subseteq \Omega} \left[ - \mathrm {m} _ {i} ^ {\mathrm {g t}} (X) \left(1 - \mathrm {u} _ {i} ^ {\mathrm {g t}} (X)\right) \log \left(\mathrm {m} _ {i} ^ {\mathrm {p r e d}} (X)\right) \right. \\ \left. \left. + \mathrm {u} _ {i} ^ {\mathrm {g t}} (X) \left(1 - \mathrm {m} _ {i} ^ {\mathrm {p r e d}} (X)\right) \right] \right.. \\ \end{array} +$$ + +In addition to the belief mass vector $\mathbf{m}_i^{\mathrm{gt}}$ and $\mathbf{m}_i^{\mathrm{pred}}$ the uncertainty vector $\mathbf{u}_i^{\mathrm{gt}}$ is needed, which is defined as the difference between the plausibility and belief (see Eq. (4)). This can be easily calculated from $\mathbf{m}_i^{\mathrm{gt}}$ using Eq. (2) and Eq. (3). + +Similar to its classical counterpart, the definition of the BBA cross-entropy is not symmetric, but it always satisfies the inequality: + +$$ +\mathrm {C E} _ {\mathrm {B B A}} \left(\mathbf {m} _ {i} ^ {\mathrm {g t}}, \mathbf {m} _ {i} ^ {\mathrm {p r e d}}\right) \geq \mathrm {C E} _ {\mathrm {B B A}} \left(\mathbf {m} _ {i} ^ {\mathrm {g t}}, \mathbf {m} _ {i} ^ {\mathrm {g t}}\right) \geq 0. \tag {12} +$$ + +Importantly, the left equality is achieved, if and only if the two BBAs are equal, i.e., $\mathbf{m}_i^{\mathrm{gt}} = \mathbf{m}_i^{\mathrm{pred}}$ . Additionally, if the BBAs $\mathbf{m}_i^{\mathrm{gt}}$ and $\mathbf{m}_i^{\mathrm{pred}}$ are Bayesian (i.e., $\operatorname{prob}(\omega) = \operatorname{bel}(\omega) = \operatorname{pl}(\omega)$ ) and therefore $\mathfrak{u}(\omega) = 0$ , $\forall \omega \subseteq \Omega$ , the $\mathrm{CE}_{\mathrm{BBA}}(\mathbf{m}_i^{\mathrm{gt}}, \mathbf{m}_i^{\mathrm{pred}})$ becomes the classical cross-entropy. + +We have observed that the distribution of belief masses in the ground-truth is highly imbalanced (see supplemental in Fig. 1). To compensate for that, we introduce the per-voxel weight $w_{i}$ and balance the contribution of individual voxels based on their frequency. The final loss is a weighted sum of the BBAs cross-entropy values of all voxels: + +$$ +\mathcal {L} = \sum_ {i} w _ {i} \mathrm {C E} _ {\mathrm {B B A}} \left(\mathbf {m} _ {i} ^ {\mathrm {g t}}, \mathbf {m} _ {i} ^ {\mathrm {p r e d}}\right). \tag {13} +$$ + +![](images/6c79d15bdfc860da2564fd8b6372ed8c8e212e6f41c81b4f2863b9633905610f.jpg) +Figure 3. Semantic occupancy grid maps. We visualize the ground-truth semantic occupancy grid maps from Occ3D [31] and our method on samples from the nuScenes [2] dataset. Semantic labels are assigned to each voxel based on the BBA (see Sec. 4.2). Additionally, we visualize predictions of BEVDet4D-Occ [10] trained with the semantic occupancy data of Occ3D [31] (baseline) and trained with our framework. Both models use the camera images as input. In the first and second example, the baseline model predicts arbitrary flying voxels, whereas our predictions are much cleaner. In the third example, our model predicts the shape of the trees much better. The last example shows how our GT occupancy is better capable of separating the pedestrians. + +![](images/c19e57d0699cd586f0480eaec120b60311f5a80e20427f0f0b7592df0472351f.jpg) + +![](images/3da5e910afb6463b5209ba47215cad12cb64ec249279f08d382e71f39a164abf.jpg) + +![](images/b12cba98482901fbbe2a6a12baf9decaa8ab1a1e74ed4ee8f1a61bc2955f9744.jpg) + +![](images/88c3cb7bc1e590a04e56bdb57b44b927302d34ddc3b2e62bdfe61bd67dc2b7a5.jpg) +Figure 4. Birds-eye-view uncertainty. For each of the three scenes, we visualize the input camera images on the top. For each scene, we show the predicted semantics of our model on the left. On the right we depict the predicted uncertainties $\mathfrak{m}(\Omega)$ averaged along the height dimension. The model predicts high uncertainties (colored in red) mainly in occluded areas, e.g., behind walls. Clearly visible areas obtain much lower uncertainties colored in blue. + +![](images/8eb1db8c4dcc784ff0e062f7e012709e14c71315077cd2257312ea10eef206db.jpg) + +![](images/5af6c9053af8ddbd36951d7c194a1ff56db68d549a569a15aee5e54472ef6efb.jpg) + +The weight $w_{i}$ is calculated from the inverse square root of the mean BBA over all voxels $\mathrm{m}_{\mathrm{mean}}^{\mathrm{gt}}$ : + +$$ +w _ {i} = \sum_ {X \subseteq \Omega} \frac {\mathrm {m} _ {i} ^ {\mathrm {g t}} (X)}{\sqrt {\mathrm {m} _ {\text {m e a n}} ^ {\mathrm {g t}} (X)}} \quad \text {w i t h} \quad \mathbf {m} _ {\text {m e a n}} ^ {\mathrm {g t}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathbf {m} _ {i} ^ {\mathrm {g t}}. \tag {14} +$$ + +This per-voxel weighting increases the contribution of rare voxels and decreases the contribution of very common voxels, e.g., those with total uncertainty $(\mathrm{m}_i^{\mathrm{gt}}(\Omega) = 1)$ . + +Decision-making based on BBAs. In order to finally predict each voxel's state, we need to make an informed decision based on the BBA of this voxel. To this end, we calculate the relative entropy $\mathrm{D_{BBA}(\mathbf{m}_i||\mathbf{m}) = CE_{BBA}(\mathbf{m}_i,\mathbf{m}) - CE_{BBA}(\mathbf{m}_i,\mathbf{m}_i)}$ between our predicted BBA $\mathbf{m}$ and BBAs $\mathbf{m}_i$ , where $\mathbf{m}_i$ is focused on a specific hypothesis such that + +$\mathbf{m}_i(\omega) = 1$ if $\omega = \omega_{i}$ and 0 otherwise. The BBA $\mathbf{m}_i$ and the respective hypothesis $\omega_{i}$ , which minimizes the relative entropy, is selected as the decision [6]: + +$$ +\omega^ {*} = \omega_ {i ^ {*}} \quad \text {w i t h} \quad i ^ {*} = \underset {i} {\arg \min } \mathrm {D} _ {\mathrm {B B A}} (\mathbf {m} _ {i} \| \mathbf {m}). \tag {15} +$$ + +In our particular case, there are two decisions to be made. First, we need to decide if a voxel is free or occupied by anything (class-agnostic). Second, we need to determine the semantic classes of all occupied voxels. This leads to the following decision rules: + +- Occupancy state: We define $\mathrm{m_{occ}}(\omega) = 1$ if $\omega = \mathcal{O}$ and 0 otherwise. Similarly, we define $\mathrm{m}_{\mathrm{free}}(\omega) = 1$ if $\omega = \mathcal{F}$ and 0 otherwise. A voxel is considered occupied if $\mathrm{D_{BBA}}(\mathbf{m_{occ}}||\mathbf{m}) < \mathrm{D_{BBA}}(\mathbf{m_{free}}||\mathbf{m})$ and free otherwise. +- Semantic class: For each class $i$ , we define $\mathrm{m}_i(\omega) = 1$ if $\omega = \mathbf{c}_i$ , and 0 otherwise. An occupied voxel is assigned the semantic class $c_i$ if $\mathrm{D_{BBA}}(\mathbf{m}_i||\mathbf{m}) < \mathrm{D_{BBA}}(\mathbf{m}_j||\mathbf{m})$ . + +for all $j \neq i$ , i.e., the relative entropy between the BBA $\mathbf{m}_i$ and the BBA $\mathbf{m}$ is minimal. + +Note that this decision step is only applied before the final evaluation, which requires a discrete class for each voxel. During model training we directly use the belief masses to supervise our model. + +# 5. Experiments + +We evaluate our method on the nuScenes dataset [2], which is most common in the field of occupancy prediction [15, 31, 32, 34, 35]. Additional results for the Waymo dataset [30] are provided in the supplementary material in Sec. F. + +# 5.1. Evaluation Protocol + +The evaluation of occupancy grids is a challenging task, since a perfect ground-truth occupancy grid is not available. However, we can evaluate the quality of different occupancy grid maps by comparing them to the LiDAR measurements, which provide the best common reference and source of ground-truth. We follow the evaluation protocol of [15] and report additional error metrics for the semantic labels. + +Depth Evaluation. To evaluate the geometry of the occupancy grid maps, we compare the measured depth of the LiDAR $d_{i}^{\mathrm{lidar}}$ scans with the rendered depth from the occupancy grid $d_{i}^{\mathrm{pred}}$ . To do so, we cast a ray from the LiDAR sensor origin in the direction of the $i^{\mathrm{th}}$ LiDAR measurement. We terminate the ray at the first intersection with the surface of a voxel predicted to be occupied. We then extract the length of this rendered ray until the intersection point and call it the rendered depth $d_{i}^{\mathrm{pred}}$ . Finally, we compare the rendered depth $d_{i}^{\mathrm{pred}}$ to the measured depth $d_{i}^{\mathrm{lidar}}$ of the LiDAR using established depth metrics. In Tab. 2 we report the mean absolute error (MAE) and root mean squared error (RMSE), as well as the percentage of points with a relative error $\delta$ below a certain threshold. + +Semantic Evaluation. Similar to the depth evaluation, we cast a ray from the LiDAR sensor origin in the direction of each LiDAR measurement. Instead of comparing the depths, we compare the semantic labels of the first occupied voxels along each ray $l_i^{\mathrm{pred}}$ with the semantic labels of the labeled LiDAR scans $l_i^{\mathrm{lidar}}$ given in the nuScenes dataset [2]. We report the intersection-over-union (IoU) per class, as well as the mean IoU (mIoU) averaged across all classes in Tab. 3. Note that this metric does not take the depth errors into account: if the LiDAR ray happens to have the same label as the first voxel they intersect, this prediction will be counted as correct even though the depth might be wrong. + +To quantify both, the depth and the class accuracy at the same time, we calculate the fraction of points which have a relative depth error below a certain threshold $\delta_{\mathrm{cls}}$ and have + +the correct semantic label as given by the labeled LiDAR scans. This metric is stricter than the relative depth accuracy, as it requires the correct semantic label in addition to the correct depth. + +# 5.2. Model Training + +To demonstrate the impact of the more accurate GT data provided by our method, we compare BEVDet4D-Occ [10, 11] trained on our semantic occupancy data and our proposed supervision to the same model trained on the commonly used Occ3D [31] dataset. We use ResNet-50 [9] as the backbone for BEVDet4D-Occ [10, 11]. The model predicts semantic occupancy grid maps, using the surround-view camera images as model inputs. Importantly, we do not change the model architecture, but only the loss and the number of outputs in the last layer to match the number of hypotheses. Additionally, we use our generated semantic occupancy data instead of Occ3D's data for supervision. + +BEVDet4D-Occ uses additional depth supervision during training, which is not required for our occupancy GT. We start training from a model checkpoint trained on Occ3D [31]. We continue training for 24 epochs using the same hyperparameters, but only changing the model supervision to our evidential formulation. Finally, we reduce the learning rate by a factor of 10 and continue training for 24 more epochs. The baseline model follows the same training schedule, but is supervised with Occ3D's data. + +# 5.3. Evaluation of Semantic Occupancy Grids + +Semantic Occupancy Ground Truth. We compare our semantic occupancy grids with those of SurroundOcc [35], Scene as Occupancy [32], Occ3D [31], and Käble et al. [15] on the nuScenes dataset [2]. The former methods provide semantic occupancy grids, while the last method provides the state of the art only for binary occupancy grids. Therefore, we can only report the depth evaluation results for [15] and the additional semantic evaluation as described in Sec. 5.1 for all other methods. We show qualitative results in Fig. 3, visualizing the generated occupancy GT from Occ3D [31] and our method. Additionally, we show the corresponding model predictions as well as the front camera images. The depth evaluation results are summarized in Tab. 2. Our method consistently outperforms all baselines across different metrics and voxel grid resolutions. Through our extended hyperparameter search space for $\alpha_{r}$ and $\alpha_{t}$ and novel decision-making scheme, we even outperform [15] in terms of geometric accuracy. Tab. 3 provides quantitative results for the ray-based IoU for the semantic classes. Also in this case our method consistently outperforms all the baselines, which validates that our approach for semantic occupancy using evidence theory is superior to simpler heuristic-based methods. + +
Voxel sizeRMSEδ<1.25δ<1.252δ<1.253δclsδcls2
in min min min %in %in %in %in %in %in %in %in %in %in %
Occupancy GTSurroundOcc [35]0.52.805.3456.281.188.649.869.072.6
Scn. as Occ. [32]0.52.925.1746.773.486.641.462.370.4
Occ3D [31]0.41.894.3381.589.493.074.278.780.1
Käblle et al. [15]0.41.253.4287.994.696.8---
EvOcc (ours)0.41.212.9890.195.897.683.486.687.2
OpenOccupancy [34]0.22.084.3570.683.689.565.174.277.3
Käblle et al. [15]0.20.912.9992.696.297.7---
EvOcc (ours)0.20.842.6293.896.998.387.188.588.9
Pred.Occ3D [31]0.42.334.3158.889.694.350.474.376.4
EvOcc (ours)0.41.664.1888.694.396.776.679.580.4
+ +Table 2. Depth evaluation. We compare different methods to generate occupancy GT data as well as trained models on the nuScenes [2] dataset, following the evaluation procedure described Sec. 5.1. The voxel size and depth errors are in meters, while the remaining numbers are in percent. Our method provides occupancy maps that match the LiDAR scans much better compared to other methods using a comparable voxel size. We see the same for BEVDet4D-Occ [10] trained with our evidential framework as compared to training on Occ3D [31]. + +
Voxel sizeBarrierBicycleBusCarConstruction vehicleMotorcyclePedestrianTraffic coneTrailerTruckDriveable surfaceOther flatSidewalkTerrainManmadeVegetationMean
Occupancy GTSurroundOcc [35]0.549.612.967.048.134.728.119.39.142.558.179.931.844.445.348.647.441.7
Scn. as Occ. [32]0.550.013.970.752.143.830.320.68.852.763.877.528.738.639.449.748.043.0
Occ3D [31]0.465.519.074.258.048.532.724.614.653.766.484.850.658.956.555.954.551.2
EvOcc (ours)0.472.934.785.177.171.857.545.224.880.083.289.059.666.668.977.777.267.0
OpenOcc. [34]0.266.226.677.967.856.447.136.916.957.574.481.844.351.050.856.453.454.1
EvOcc (ours)0.282.145.289.783.378.269.356.837.184.987.989.070.976.176.582.781.374.4
Pred:Occ3D [31]0.453.615.368.755.233.126.323.510.339.552.882.031.343.143.256.255.543.1
EvOcc (ours)0.461.35.076.669.238.832.125.613.346.963.286.944.053.756.858.646.248.6
+ +Table 3. Semantic evaluation. We compare the semantic occupancy grid maps of different semantic occupancy GT reconstructions, as well as those from trained models, to the LiDAR scans of the nuScenes [2] dataset following the evaluation procedure from Sec. 5.1. The voxel size is in meters, while the remaining numbers are in percent. Our method consistently yields more accurate GT semantic occupancy grid maps and leads to better trained models. All numbers are in percent. + +Semantic Occupancy Prediction. We compare the model predictions of the models trained on our semantic occupancy data with the models trained on the Occ3D [31] data. We evaluate the same metrics for the model predictions as for the occupancy GT (see Sec. 5.1). We report the quantitative results in Tab. 2 and Tab. 3. BEVDet4D-Occ trained on our data yields more accurate predictions in terms of geometry and semantics. Qualitative results of the model semantic predictions and corresponding GT occupancy grids are shown in Fig. 3. Additionally, we visualize the predicted semantics together with the aggregated uncertainty $\mathrm{m}(\Omega)$ along the height dimension in Fig. 4. Furthermore, we compare our ground-truth uncertainty and the predicted uncertainty in Fig. 1. + +# 6. Conclusion + +In this work, we presented a novel approach to obtain high-quality 3D semantic occupancy maps using ET. Our approach introduces a new BBA formulation for modeling LiDAR measurements with partial semantic annotations. We compared our generated ground truth to that of other state-of-the-art semantic occupancy datasets and showed that our data matches the LiDAR measurements better by a large margin in terms of geometry and semantics. We also demonstrated that using our ground truth for training the semantic occupancy prediction model BEVDet4D-Occ significantly improves the model's performance. In conclusion, our work sets a new state of the art for semantic occupancy ground truth generation and model training, and therewith makes an important contribution towards autonomous driving. + +# References + +[1] Stefan Andreas Baur, David Josef Emmerichs, Frank Moosmann, Peter Pinggera, Björn Ommer, and Andreas Geiger. SLIM: Self-supervised LiDAR scene flow and motion segmentation. In ICCV, 2021. 3 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3, 4, 6, 7, 8, 1 +[3] Arthur P Dempster. A generalization of bayesian inference. Journal of the Royal Statistical Society: Series B (Methodological), 1968. 2, 3 +[4] Yong Deng. Uncertainty measure in evidence theory. Science China Information Sciences, 63, 2020. 5 +[5] Jean Dezert. An effective measure of uncertainty of basic belief assignments. 2022 25th International Conference on Information Fusion (FUSION), pages 1-8, 2022. +[6] Jean Dezert and Frédéric Dambreville. Cross-entropy and relative entropy of basic belief assignments. 2023 26th International Conference on Information Fusion (FUSION), pages 1–8, 2023. 3, 5, 6 +[7] Alberto Elfes. Using occupancy grids for mobile robot perception and navigation. Computer, 22(6):46-57, 1989. 2 +[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In CVPR, 2012. 2 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7 +[10] Junjie Huang and Guan Huang. BEVDet4D: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2, 4, 5, 6, 7, 8 +[11] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2, 7 +[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, 2023. 1, 2, 4 +[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, page 0, 2006. 2 +[14] J. Kim et al. Evidential semantic mapping in off-road environments with uncertainty-aware bayesian kernel inference. IROS, 2024. 3 +[15] Jonas Käblé, Sascha Wirges, Maxim Tatarchenko, and Eddy Ilg. Accurate training data for occupancy map prediction in automated driving using evidence theory. In CVPR, 2024. 1, 2, 3, 4, 5, 7, 8 +[16] J. Li et al. Segm: A novel semantic evidential grid map by fusing multiple sensors. In PRCV, 2020. 3 +[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of reliable depth for multi-view 3D object detection. In AAAI, pages 1477-1485, 2023. 1, 2 + +[18] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: Monocular 3D semantic scene completion benchmark in street views. arXiv preprint arXiv:2306.09001, 2023. 2 +[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, Cham, 2022. 1, 2, 4 +[20] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M. Alvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation, 2023. 2 +[21] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M. Alvarez. FB-BEV: BEV representation from forward-backward view transformations. In ICCV, pages 6919-6928, 2023. 1, 2 +[22] Haisong Liu, Yang Chen, Haiguang Wang, Zetong Yang, Tianyu Li, Jia Zeng, Li Chen, Hongyang Li, and Limin Wang. Fully sparse 3d occupancy prediction. In ECCV, 2023. 4 +[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position embedding transformation for multi-view 3D object detection. In ECCV, pages 531-548. Springer, 2022. 2 +[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2 +[25] H. P. Moravec. Sensor fusion in certainty grids for mobile robots. Sensor Devices and Systems for Robotics, 9(2):253-276, 1989. 2 +[26] Sven Richter, Frank Bieder, Sascha Wirges, and Christoph Stiller. A dual evidential top-view representation to model the semantic environment of automated vehicles. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2023. 3 +[27] S. Richter et al. Semantic evidential grid mapping using monocular and stereo cameras. Sensors, 2021. 3 +[28] Glenn Shafer. A mathematical theory of evidence. 1976. 2, 3 +[29] Philippe Smets. Data fusion in the transferable belief model. Proceedings of the Third International Conference on Information Fusion, 1:PS21-PS33 vol.1, 2000. 3 +[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. CVPR, 2019. 2, 7 +[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3D: A large-scale 3d occupancy prediction benchmark for autonomous driving. In NeurIPS, 2023. 1, 2, 4, 6, 7, 8 +[32] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, + +and Hongyang Li. Scene as occupancy. In ICCV, 2023. 1, 2, 7, 8 +[33] Ignacio Vizzo, Tiziano Guadagnino, Jens Behley, and Cyril Stachniss. VDBFusion: Flexible and efficient TSDF integration of range sensor data. Sensors, 22(3):1296, 2022. 2 +[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. OpenOccupancy: A large scale benchmark for surrounding semantic occupancy perception. In ICCV, 2023. 1, 2, 7, 8 +[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. SurroundOcc: Multi-camera 3d occupancy prediction for autonomous driving. In CVPR, 2023. 1, 2, 7, 8 +[36] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. Curran, 2021. 2 +[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 2 +[38] Hangyu Yan and Yong Deng. An improved belief entropy in evidence theory. IEEE Access, 8:57505-57516, 2020. 5 +[39] T. Yang and V. Aitken. Evidential mapping for mobile robots with range sensors. IEEE Transactions on Instrumentation and Measurement, 55:1422-1429, 2006. 3 \ No newline at end of file diff --git a/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/images.zip b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c9d7ef2188519c4d0492c1c91143b8a71c771bef --- /dev/null +++ b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a44b59c9bfa2f3bd521b7b59cdcaaf9af3bae4bf80eb2528ec16727c2809345 +size 672000 diff --git a/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/layout.json b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4269ad2bad85b2da55c796464a73ac6a4a7fcd8c --- /dev/null +++ b/2025/EvOcc_ Accurate Semantic Occupancy for Automated Driving Using Evidence Theory/layout.json @@ -0,0 +1,9441 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "text", + "content": "EvOcc: Accurate Semantic Occupancy for Automated Driving Using Evidence Theory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "type": "text", + "content": "Jonas Käblé" + }, + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "type": "text", + "content": " Sascha Wirges" + }, + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 160, + 303, + 190 + ], + "type": "text", + "content": "Bosch Center for Artificial Intelligence" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 192, + 303, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 192, + 303, + 202 + ], + "spans": [ + { + "bbox": [ + 134, + 192, + 303, + 202 + ], + "type": "text", + "content": "firstname_lastname@de.bosch.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 310, + 160, + 499, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 160, + 499, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 160, + 499, + 190 + ], + "type": "text", + "content": "Maxim Tatarchenko1 Eddy Ilg2 \n2University of Technology Nuremberg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 192, + 395, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 192, + 395, + 202 + ], + "spans": [ + { + "bbox": [ + 312, + 192, + 395, + 202 + ], + "type": "text", + "content": "eddy.ilg@utn.de" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 257, + 297, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 449 + ], + "type": "text", + "content": "We present EvOcc, a novel framework for evidential semantic occupancy mapping. It consists of two parts: (1) an evidential approach for calculating the ground-truth 3D semantic occupancy maps from noisy LiDAR measurements, and (2) a method for training image-based occupancy estimation models through a new loss formulation. In contrast to state-of-the-art semantic occupancy maps, our approach explicitly models the uncertainty introduced by unobserved spaces or contradicting measurements, and we show that using it results in significantly stronger models. Evaluated as ray-based mIoU, our evidential semantic occupancy mapping approach improves over the baselines by at least " + }, + { + "bbox": [ + 54, + 257, + 297, + 449 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 54, + 257, + 297, + 449 + ], + "type": "text", + "content": " for the ground truth and " + }, + { + "bbox": [ + 54, + 257, + 297, + 449 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 54, + 257, + 297, + 449 + ], + "type": "text", + "content": " for the trained model. Overall, we make a significant contribution towards more detailed and uncertainty-aware 3D environment understanding and safe operation in autonomous driving." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 475, + 135, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 135, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 135, + 488 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 497, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 297, + 581 + ], + "type": "text", + "content": "A critical component for safe automated driving is accurate and reliable environment perception, which includes understanding the scene geometry as well as interpreting its semantic content. While various 3D representations have been explored in the past, semantic voxel grids have become a recent de facto standard for ML-based environment perception [15, 31, 32, 34, 35]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "For a long time, automated driving relied on LiDAR sensors due to their low and distance-independent measurement errors. However, recent methods attempt to reconstruct the 3D environment from camera images alone [10-12, 15, 17, 19-21, 24], since LiDAR is sparse, expensive, and redundant. This makes predicting accurate environment maps from images an important task, and further brings up the problem of obtaining ground-truth training data. The latter is usually produced by recording LiDAR measurements, and employing heuristics to convert them into semantic occupancy maps. However, such reconstructed occupancy maps" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 554, + 321 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 554, + 321 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 554, + 321 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 554, + 321 + ], + "type": "image", + "image_path": "b4817f692080b6b73f5f72534e4bffde3718da7596df33fd253fe3662e6a4c53.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 321, + 435, + 409 + ], + "blocks": [ + { + "bbox": [ + 317, + 321, + 435, + 409 + ], + "lines": [ + { + "bbox": [ + 317, + 321, + 435, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 435, + 409 + ], + "type": "image", + "image_path": "1011a8cf653cb5a2f062ae2732285b1309358a8a08a7f8cca39aff296f369c11.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 409, + 435, + 498 + ], + "blocks": [ + { + "bbox": [ + 317, + 409, + 435, + 498 + ], + "lines": [ + { + "bbox": [ + 317, + 409, + 435, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 409, + 435, + 498 + ], + "type": "image", + "image_path": "73d08c7bbdf454641943c857213fa40d6579c46fbfda248dfd6aca0442925c42.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 507, + 556, + 607 + ], + "lines": [ + { + "bbox": [ + 313, + 507, + 556, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 556, + 607 + ], + "type": "text", + "content": "Figure 1. Semantic occupancy and birds-eye-view uncertainty. Our model uses multi-view camera images as input and predicts semantic occupancy. We visualize the ground-truth semantic occupancy obtained from LiDAR scans on left and the predicted semantic occupancy on the right. Below we show the GT uncertainty obtained from our semantic basic belief assignment (BBA) using the LiDAR scans as well as the predicted uncertainties. High uncertainties (e.g., due to occlusions) are colored in red, whereas areas of low uncertainty are colored in blue." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 435, + 321, + 553, + 409 + ], + "blocks": [ + { + "bbox": [ + 435, + 321, + 553, + 409 + ], + "lines": [ + { + "bbox": [ + 435, + 321, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 435, + 321, + 553, + 409 + ], + "type": "image", + "image_path": "72416b7e24146e874ca4ad1dfa61dd2c27f258ecdf4d931f24f157568093cb3c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 409, + 553, + 498 + ], + "blocks": [ + { + "bbox": [ + 435, + 409, + 553, + 498 + ], + "lines": [ + { + "bbox": [ + 435, + 409, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 435, + 409, + 553, + 498 + ], + "type": "image", + "image_path": "5ef55350608bfcc444883b360bb94bdfd9cf0ff522d5a945459841de98bdd764.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "are often inaccurate. A recent work by Käblé et al. [15] has shown that the accuracy can in fact be improved by a margin of " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "30 - 50\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " when simple heuristics are replaced with the framework of evidential mapping. Different from a classical probabilistic interpretation, this approach explicitly models conflicting hypotheses and unknown states, which is crucial for representing noisy real-world measurements." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27467" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": "Nevertheless, the authors only model the voxel occupancy state from LiDAR scans, without any notion of semantic classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 109, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 109, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 109, + 295, + 251 + ], + "type": "text", + "content": "In this work, we present " + }, + { + "bbox": [ + 54, + 109, + 295, + 251 + ], + "type": "inline_equation", + "content": "EvOcc" + }, + { + "bbox": [ + 54, + 109, + 295, + 251 + ], + "type": "text", + "content": ", an evidence theory (ET) [3, 28] based framework for (1) accurately reconstructing semantic occupancy maps from LiDAR scans (see Fig. 2) and (2) using these as ground-truth to train multi-camera vision models to predict evidential occupancy maps (see Fig. 1). To this end, we extend the evidential mapping framework from Käblé et al. [15] to include semantic class labels. We derive a novel basic belief assignment (BBA) formulation that can handle noisy and partially annotated LiDAR measurements. We also employ a new loss for training semantic occupancy prediction models with our ground-truth data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 253, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 373 + ], + "type": "text", + "content": "Semantic occupancy maps produced by our approach are far superior to those from state-of-the-art semantic occupancy methods [31, 32, 34, 35], which we demonstrate both qualitatively and quantitatively. Using our reconstructed evidential occupancy maps as training data, we are able to train evidential occupancy models that significantly outperform those trained on previously available ground truth. Note that this large improvement is achieved by only changing the model supervision, while keeping the model architecture the same. In summary, we make the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 515 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 374, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 409 + ], + "type": "text", + "content": "- We present a complete 3D evidential mapping framework for semantic occupancy consisting of two parts: occupancy GT generation and image-based occupancy estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 410, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 296, + 456 + ], + "type": "text", + "content": "- We apply a novel BBA formulation based on noisy, partially annotated LiDAR measurements, and employ a novel loss that directly uses the obtained evidential semantic occupancy maps as supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 515 + ], + "type": "text", + "content": "- We demonstrate both qualitatively and quantitatively, that the semantic occupancy maps produced by our trained model and mapping approach are significantly more accurate than those from state-of-the-art methods in terms of agreement with LiDAR measurements (see Sec. 5.3)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 517, + 193, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 193, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 193, + 529 + ], + "type": "text", + "content": "Our code is available on GitHub1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 542, + 142, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 142, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 142, + 554 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 563, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 693 + ], + "type": "text", + "content": "3D Occupancy Datasets. Training perception models for automated driving requires large-scale multimodal datasets, such as KITTI [8], nuScenes [2], Waymo [30] and Agro-verse2 [36]. These datasets record the surrounding environment of the vehicle with multiple onboard sensors, such as LiDAR and cameras. Additionally, they provide annotations for the raw sensor data for tasks like 3D object detection [17, 19, 23, 24] and semantic LiDAR segmentation [12, 37]. The task of camera-only semantic occupancy prediction [18, 31, 32, 34, 35] is to predict the occupancy state and class of each voxel in the environment of the vehi" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 470, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 470, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 470, + 84 + ], + "type": "text", + "content": "cle, given the onboard camera images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 84, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 555, + 347 + ], + "type": "text", + "content": "A common process for obtaining ground-truth (GT) 3D occupancy grids is aggregating LiDAR measurements over multiple time steps under the consideration of moving objects [31, 32, 34, 35]. The resulting aggregated point clouds are then partitioned into dense voxel grids, where each voxel containing a point is labeled according to the point's annotation. Finally, ray casting is used to determine the free space labels between the LiDAR origin and the first occupied voxel. The quality of the occupancy grids can be improved through additional processing steps. For instance, Occ3D [31] and SurroundOcc [35] propose to make the point cloud more dense using mesh and surface reconstruction [13, 33]. Furthermore, Occ3D enforces consistency between the camera images and the occupancy grid through image-guided voxel refinement. OpenOccupancy [34] improves the label quality with additional human supervision and their Augment and Purifying process. Käblé et al. [15] leverage evidence theory (ET) and achieve the current state of the art by proposing to model the occupancy state of each voxel based on the number of transmissions and reflections through it. Our work builds up on [15] and proposes a framework using ET that can model beliefs for semantic classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 602 + ], + "type": "text", + "content": "3D Semantic Occupancy Prediction. The task of 3D semantic occupancy prediction is to estimate the state of each voxel in the environment of the vehicle using only camera images. A voxel can either be free or occupied by a specific class. The CVPR 2023 Occupancy Prediction Challenge [31, 32] is a popular benchmark for this task. The winner of this challenge, FB-OCC [20, 21], extends BEVDet [11] by a forward-backward view transformation module. TPVFormer [12] addresses the high memory and compute requirements for a dense 3D grid and projects the 3D space onto 3D planes. Additional depth supervision for the process of splatting 2D image features into the 3D space has been introduced by BEVDepth [17]. BEVDet4D [10] aggregates additional information over time for the task of 3D object detection. In this work, we use BEVDet4D-Occ-R50 as a baseline model, which is an adaptation of BEVDet4D specifically tailored for the task of semantic occupancy prediction and commonly used as reference model. We show that training this model on the new ground truth obtained with our approach gives superior results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Occupancy Grid Mapping. Static occupancy grid mapping using a Bayesian interpretation of probabilities is a well-explored area in robotics [7, 25]. The state of a voxel cell is modeled as a Bernoulli-distributed random variable and can either be occupied or free. Since the state of the cell is defined by a single probability, the model does not consider the uncertainty in the measurements to estimate the cell state, nor does it handle conflicts between different mea" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 254, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 254, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 254, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27468" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 68, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 63, + 68, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 63, + 68, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 63, + 68, + 553, + 196 + ], + "type": "image", + "image_path": "30475cc4e711c9c4644d176754cd70ca8ed7683fcd00ff827bd5987cd1d5ff22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 201, + 476, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 201, + 476, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 476, + 215 + ], + "type": "text", + "content": "Figure 2. Overview. A scheme of our evidential framework for semantic occupancy mapping." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 297, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 297, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 297, + 462 + ], + "type": "text", + "content": "suresments. To circumvent these limitations, Yang and Aitken [39] propose to model the state of a cell using ET [3, 28], which can explicitly represent the uncertainty. Richter et al. [27] and Li et al. [16] model the 2D birds-eye-view (BEV) occupancy maps using ET. More recent works extend the mapping task to 3D and consider dynamic environments as well as semantics. There are some specialized models for automated driving scenarios. Richter et al. [26] split dynamic traffic scenes into two evidential BEV grid maps containing the ground plane and everything on top of the ground. Kim et al. [14] present a 3D semantic occupancy mapping approach, which fuses unreliable semantic predictions using ET. Furthermore, improvements in the mapping process for dynamic scenes can be achieved by considering the motion of objects using scene flow from LiDAR measurements [1]. A framework for 3D occupancy grid mapping based on ET within a dynamic environment is presented in [15]. However, the presented method is limited to occupancy maps without semantic information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 141, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 141, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 141, + 487 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 494, + 157, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 157, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 157, + 507 + ], + "type": "text", + "content": "3.1. Evidence Theory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 512, + 297, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 297, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 297, + 596 + ], + "type": "text", + "content": "Here, we briefly revisit the fundamentals of evidence theory (ET), or Dempster-Shafer theory [3, 28], used to model the semantic occupancy state of individual voxels (see Sec. 4.1). ET provides a flexible framework for reasoning about the uncertainty, be it total uncertainty or partial knowledge about a system [29]. A single hypothesis is described by a subset of the frame of discernment " + }, + { + "bbox": [ + 55, + 512, + 297, + 596 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 512, + 297, + 596 + ], + "type": "text", + "content": ", and the BBA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 605, + 296, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 605, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 93, + 605, + 296, + 631 + ], + "type": "interline_equation", + "content": "\\mathrm {m} \\colon 2 ^ {\\Omega} \\rightarrow [ 0, 1 ] \\quad \\text {w i t h} \\quad \\sum_ {\\omega \\subseteq \\Omega} \\mathrm {m} (\\omega) = 1 \\tag {1}", + "image_path": "ef7a6a117bfbcff9572127f6521dc265f3c1c42c8ab616c4abc7444e0b7f3edb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 664 + ], + "type": "text", + "content": "assigns a normalized belief mass " + }, + { + "bbox": [ + 55, + 641, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{m}(\\omega)" + }, + { + "bbox": [ + 55, + 641, + 295, + 664 + ], + "type": "text", + "content": " to each hypothesis " + }, + { + "bbox": [ + 55, + 641, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\omega \\subseteq \\Omega" + }, + { + "bbox": [ + 55, + 641, + 295, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "Intuitively, the belief mass represents the proportion of relevant and available evidence supporting the hypothesis " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": " but not any particular subset of it. The classical probability of a hypothesis " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": " lies between two values known as belief" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 234, + 427, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 427, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 427, + 246 + ], + "type": "text", + "content": "and plausibility. The belief" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 255, + 555, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 255, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 365, + 255, + 555, + 281 + ], + "type": "interline_equation", + "content": "\\operatorname {b e l} (\\omega) = \\sum_ {X \\subseteq \\omega} \\mathrm {m} (X) \\leq \\operatorname {p r o b} (\\omega) \\tag {2}", + "image_path": "ae4848f5752581f0ec9917a9d7c48238c31072dea052c3f43ca1624ddb3a315a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "content": "of a hypothesis " + }, + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "content": " is the sum of all belief masses supporting the hypothesis " + }, + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "content": ". It serves as a lower bound of the probability that a certain hypothesis is true. The plausibility" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 335, + 555, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 335, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 354, + 335, + 555, + 360 + ], + "type": "interline_equation", + "content": "\\operatorname {p l} (\\omega) = 1 - \\sum_ {X \\cap \\omega = \\emptyset} \\mathrm {m} (X) \\geq \\operatorname {p r o b} (\\omega) \\tag {3}", + "image_path": "a9b5f077dce36c671a1e6ab9baaf9a639b5837d729fb5e7cd837e7b50b7358d8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 369, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 405 + ], + "type": "text", + "content": "is an upper bound for this probability. It can be defined as the complement of all belief masses not supporting the hypothesis " + }, + { + "bbox": [ + 313, + 369, + 554, + 405 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 369, + 554, + 405 + ], + "type": "text", + "content": ". The uncertainty" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 368, + 414, + 555, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 414, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 368, + 414, + 555, + 428 + ], + "type": "interline_equation", + "content": "\\mathrm {u} (\\omega) = \\operatorname {p l} (\\omega) - \\operatorname {b e l} (\\omega) \\in [ 0, 1 ] \\tag {4}", + "image_path": "f9a17805deb8ebe8e63ff24873d76790bd660acc57123eb5224b9afb15f348b9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "type": "text", + "content": "of a hypothesis " + }, + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "type": "text", + "content": " can then be defined as the discrepancy between plausibility and belief [6]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 468, + 462, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 462, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 462, + 481 + ], + "type": "text", + "content": "3.2. Spherical Sensor Mapping" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": "Käblé et al. [15] describe an approach to determine the state of a voxel based on the transmissions, and reflections through it. These are measured by a LiDAR sensor and provide evidence for a voxel being free or occupied by a certain class. Different from their work, we address the multi-class case here." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "text", + "content": "Reflections Grid Map. Instead of mapping all reflections into a single spherical grid map, we create a separate spherical reflections grid map " + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{ci}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)" + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "text", + "content": ". On top of that, we have one additional reflections grid map " + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{unlabeled}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)" + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "text", + "content": " for all unlabeled reflections2. This results in a total of " + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "text", + "content": " spherical reflection grid maps, where " + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 572, + 555, + 656 + ], + "type": "text", + "content": " is the number of classes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 671, + 555, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 555, + 695 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 555, + 695 + ], + "type": "text", + "content": "Transmissions Grid Map. Since transmissions represent free space and are independent of the reflections' semantic" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 528, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 528, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 528, + 713 + ], + "type": "text", + "content": "2In nuScenes [2] only one of every 10 LiDAR scans is labeled." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27469" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": "classes, we only produce a single spherical transmission map " + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "inline_equation", + "content": "t^{\\mathrm{ph}}(\\rho ,\\phi ,\\theta)" + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": ". For a given voxel cell located at " + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "inline_equation", + "content": "(\\rho ,\\theta ,\\phi)" + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": ", the number of transmissions:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 299, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 299, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 299, + 169 + ], + "type": "interline_equation", + "content": "t ^ {\\mathrm {s p h}} (\\rho , \\phi , \\theta) = \\sum_ {\\rho^ {\\prime} > \\rho} \\underbrace {\\left(r _ {\\text {u n l a b e l e d}} ^ {\\mathrm {s p h}} \\left(\\rho^ {\\prime} , \\phi , \\theta\\right) + \\sum_ {i = 1} ^ {C} r _ {\\mathrm {c} _ {i}} ^ {\\mathrm {s p h}} \\left(\\rho^ {\\prime} , \\phi , \\theta\\right)\\right)} _ {\\text {S u m o v e r l a b e l e d a n d u n l a b e l e d r e f l e c t i o n s}} \\tag {5}", + "image_path": "c03c6887f834804b03b283f4547c6139da06a4d49a8755f131110fc392e019e4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 222 + ], + "type": "text", + "content": "is given by the sum over all reflections with the radius " + }, + { + "bbox": [ + 55, + 175, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\rho'" + }, + { + "bbox": [ + 55, + 175, + 296, + 222 + ], + "type": "text", + "content": " greater than " + }, + { + "bbox": [ + 55, + 175, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 175, + 296, + 222 + ], + "type": "text", + "content": " of the voxel under consideration. This can be efficiently implemented as a cumulative sum from the maximal to the minimal radial distance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "text", + "content": "Grid Map Normalization. We normalize the number of reflections and transmissions per voxel by their respective volume. We obtain the normalized reflections " + }, + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\hat{r}_{\\mathrm{ci}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)" + }, + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\hat{r}_{\\mathrm{unlabled}}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)" + }, + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "text", + "content": " as well as the normalized transmissions " + }, + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\hat{t}^{\\mathrm{sph}}(\\rho, \\phi, \\theta)" + }, + { + "bbox": [ + 55, + 236, + 296, + 309 + ], + "type": "text", + "content": " for each time step separately. Details can be found in the Supplemental in Sec. A." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 393 + ], + "type": "text", + "content": "Multi-Frame Aggregation. We aggregate the reflections and transmissions from multiple time steps into a single Cartesian voxel grid. Using known per-object transformations between the reference frame at time step " + }, + { + "bbox": [ + 55, + 323, + 296, + 393 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 323, + 296, + 393 + ], + "type": "text", + "content": " and frames at all other time steps " + }, + { + "bbox": [ + 55, + 323, + 296, + 393 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 323, + 296, + 393 + ], + "type": "text", + "content": ", we can accumulate the final per-voxel values as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 397, + 296, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 397, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 92, + 397, + 296, + 430 + ], + "type": "interline_equation", + "content": "\\hat {r} _ {\\mathrm {c} _ {i}, n} (x, y, z) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} g \\left(\\hat {r} _ {\\mathrm {c} _ {i}, k} ^ {\\text {s p h}} (\\rho , \\phi , \\theta)\\right), \\tag {6}", + "image_path": "2e556109367b7b4d027a06d91cf5e07b167219c54c35df5888288ef9aa1161c3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "text", + "content": " applies the motion compensation and transforms spherical coordinates into Cartesian ones. The same applies to the unlabeled reflections " + }, + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\hat{r}_{\\mathrm{unlabled}}(x,y,z)" + }, + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "text", + "content": " and transmissions " + }, + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\hat{t}(x,y,z)" + }, + { + "bbox": [ + 55, + 434, + 296, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 111, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 111, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 111, + 502 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": "Here we present our EvOcc framework split into two stages. The first stage (Sec. 4.1) reconstructs accurate semantic occupancy maps from sparse LiDAR scans with partially annotated semantic labels. The second stage (Sec. 4.2) provides models trained to predict semantic occupancy with significantly improved accuracy compared to the state of the art." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 588, + 273, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 273, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 273, + 601 + ], + "type": "text", + "content": "4.1. Evidential Semantic Occupancy Mapping" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 642 + ], + "type": "text", + "content": "In contrast to the commonly used Bayesian approach [10-12, 19-22, 31] for modeling semantic occupancy grids, ET has an advantage of directly representing uncertainties." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "text", + "content": "Initially, we need to define the frame of discernment, that models possible states, which we here define per voxel as being free or occupied by a certain class:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 682, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 682, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 682, + 296, + 712 + ], + "type": "interline_equation", + "content": "\\Omega \\stackrel {\\text {d e f}} {=} \\left\\{\\underset {C \\text {l a b e l e d c l a s s e s}} {\\mathrm {f r e e}}, \\underbrace {\\mathrm {c} _ {1} , \\mathrm {c} _ {2} , \\dots , \\mathrm {c} _ {C}} \\right\\} = \\left\\{\\mathrm {f}, \\mathrm {c} _ {1}, \\mathrm {c} _ {2}, \\dots , \\mathrm {c} _ {C} \\right\\}. \\tag {7}", + "image_path": "11ecb6d69348ac66c180d93ce0b0089720eb126c26377ebea50e9051d3b861f2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 129 + ], + "type": "table", + "html": "
Voxel stateHypothesis ωDirect Measurement z(ω)
FreeF = {f}z(F) = αtˆt
Occupied byithclassCi = {ci}z(Ci) = αrˆrci
OccupiedO = ∪iCiz(O) = αrˆunlabeled
UncertainΩ = O∪Fz(Ω) = 0
", + "image_path": "861b4c48065a8764f9dd42ea0e7113fdc4b5c42e49eae94005787733fba669b4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": "Table 1. Overview of modeled hypotheses and direct measurements. The table shows our selected hypotheses from the power set " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "2^{\\Omega}" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": ". We define the hypothesis " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": " to model free voxels. Each labeled class has its own hypothesis " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": ". The hypothesis " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": " models voxels, which are known to be occupied with unknown semantics. The hypothesis " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": " describes total uncertainty about the voxel state. We directly associate every LiDAR measurement processed by " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "z(\\cdot)" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": " to the corresponding hypotheses. There are no measurements for " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\alpha_r" + }, + { + "bbox": [ + 313, + 137, + 555, + 236 + ], + "type": "text", + "content": " act as sensor dependent-hyperparameters." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": "The set of hypotheses is then the power set of these states. Obviously, simply modeling the belief mass for each possible hypothesis " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\omega \\in 2^{\\Omega}" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": " quickly becomes infeasible since the number of hypotheses " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "N = |2^{\\Omega}| = 2^{|\\Omega|}" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": " grows exponentially. For instance, the nuScenes dataset [2] contains 17 different classes resulting in a cardinality of " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "|\\Omega_{\\mathrm{nuScenes}}| = 18" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{nuScenes}} = 2^{18} = 262" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": ", 144 possible hypotheses. Therefore, we propose an approximation through the most important subset that contains the \"free\" hypothesis " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": ", the \"occupied by class " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": "\" hypotheses " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": ", the \"generally occupied\" " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\cup_{i=1}^{C} \\{c_i\\}" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": " hypothesis, and the " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": " hypothesis, which corresponds to general uncertainty." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": "The partly annotated LiDAR measurements are represented as several Cartesian voxel grids (see Sec. 3.2): transmissions " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": ", reflections from a certain class " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\hat{r}_{c_i}" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": " and unlabeled reflections " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\hat{r}_{\\mathrm{unlabled}}" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": ". As the next step, we need to define how to map these measurements to the hypotheses established above. To that end, we introduce a simple measurement function " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "z\\colon 2^{\\Omega}\\to [0,\\inf)" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": ", which defines sensor measurements, directly providing evidence for a hypothesis but not any particular subset of it. The measurement function " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": " returns a value greater than or equal to zero corresponding to the number of transmissions or reflections supporting the hypothesis in question, scaled with factors " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\alpha_{t}\\in \\mathbb{R}_{\\geq 0}" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\alpha_{r}\\in \\mathbb{R}_{\\geq 0}" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": ". These factors act as sensor-dependent hyperparameters, similar to " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{FN}}" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{FP}}" + }, + { + "bbox": [ + 313, + 402, + 555, + 592 + ], + "type": "text", + "content": " in [15]. We provide an overview of the hypotheses and the associated direct measurements in Table 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "type": "text", + "content": "We propose to model the belief of a hypothesis " + }, + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\omega \\subset \\Omega" + }, + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "type": "text", + "content": " by considering its supporting and conflicting measurements defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 630, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "content": "- Supporting measurements " + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "inline_equation", + "content": "z(X)" + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "content": " provide evidence for hypothesis " + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "content": " being true if " + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "inline_equation", + "content": "X \\subseteq \\omega" + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "content": ". Example: a measurement of a certain class " + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "inline_equation", + "content": "z(\\mathcal{C}_i)" + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "content": " supports the hypothesis of being (generally) occupied, since the " + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "content": " is a subset of the occupied hypothesis " + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 314, + 630, + 555, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": "- Conflicting measurements " + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "z(X)" + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": " disprove the hypothesis " + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "X \\cap \\omega = \\emptyset" + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": ", i.e., when " + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": " are mutually exclusive" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27470" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "text", + "content": "and cannot be true at the same time. Example: a free measurement " + }, + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "inline_equation", + "content": "z(\\mathcal{F})" + }, + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "text", + "content": " conflicts with the occupied hypothesis, since " + }, + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "text", + "content": " is disjoint to the occupied hypothesis " + }, + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 63, + 72, + 296, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 144 + ], + "type": "text", + "content": "By summing over all supporting and conflicting measurements individually, we obtain the belief of a hypothesis " + }, + { + "bbox": [ + 55, + 109, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 109, + 296, + 144 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 149, + 296, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 291 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {b e l} (\\omega) = \\left(1 - \\exp \\left(- \\underbrace {\\sum_ {X \\subseteq \\omega} z (X)} _ {\\text {S u p p o r t i n g m e a s u r e m e n t s}}\\right)\\right) \\exp \\left(- \\underbrace {\\sum_ {X \\cap \\omega = \\emptyset} z (X)} _ {\\text {C o n f l i c t i n g m e a s u r e m e n t s}}\\right) (8) \\\\ = \\left(1 - \\prod_ {X \\subseteq \\omega} e ^ {- z (X)}\\right) \\prod_ {X \\cap \\omega = \\emptyset} e ^ {- z (X)} \\quad \\text {f o r} \\quad \\omega \\subset \\Omega . (9) \\\\ \\end{array}", + "image_path": "98b271fb5f0c33296d9d1247fdc73bb3c9f5be91904bc2d36bf369b935e009d6.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "content": "Measurements that are neither supporting nor conflicting do not provide any concrete evidence about the specific hypothesis in question and therefore do not contribute to the belief. Additionally, the belief for " + }, + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\omega = \\Omega" + }, + { + "bbox": [ + 55, + 300, + 296, + 418 + ], + "type": "text", + "content": " is equal to one by definition. Intuitively, this formulation ensures that the belief (acting as a lower bound that a hypothesis is true) can only be greater than zero if there is any measurement supporting the hypothesis. Furthermore, the belief of a hypothesis can only reach the maximum of one if there are no conflicting measurements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "text", + "content": "To calculate the belief masses " + }, + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{m}(\\omega)" + }, + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "text", + "content": " from the beliefs " + }, + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{bel}(\\omega)" + }, + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "text", + "content": ", we make use of Eq. (2) by evaluating it for each hypothesis " + }, + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 55, + 420, + 296, + 456 + ], + "type": "text", + "content": ". This can be formulated as a linear system:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 475, + 296, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 475, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 88, + 475, + 296, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {b} = \\left( \\begin{array}{c} \\operatorname {b e l} \\left(\\omega_ {1}\\right) \\\\ \\operatorname {b e l} \\left(\\omega_ {2}\\right) \\\\ \\vdots \\\\ \\operatorname {b e l} \\left(\\omega_ {n}\\right) \\end{array} \\right) = \\mathbf {S} \\left( \\begin{array}{c} \\mathrm {m} \\left(\\omega_ {1}\\right) \\\\ \\mathrm {m} \\left(\\omega_ {2}\\right) \\\\ \\vdots \\\\ \\mathrm {m} \\left(\\omega_ {n}\\right) \\end{array} \\right) = \\mathbf {S m}, \\tag {10}", + "image_path": "8ec3cc60b4b80c38e7e8ec4954851e05cbf6301ed030020db718bcac15f916d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 534, + 247, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 534, + 247, + 565 + ], + "spans": [ + { + "bbox": [ + 99, + 534, + 247, + 565 + ], + "type": "interline_equation", + "content": "\\mathbf {S} \\in \\mathbb {R} ^ {n \\times n}: \\mathbf {S} _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\omega_ {j} \\subseteq \\omega_ {i} \\\\ 0 & \\text {o t h e r w i s e} , \\end{array} \\right.", + "image_path": "e50328f71c14a575e905ff4bf8c1b1f35e21423307d7b8eee160a4dd98c2b7cc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 568, + 206, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 568, + 206, + 579 + ], + "spans": [ + { + "bbox": [ + 110, + 568, + 206, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{b}\\in \\mathbb{R}^n" + }, + { + "bbox": [ + 110, + 568, + 206, + 579 + ], + "type": "text", + "content": " : belief vector," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 583, + 228, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 583, + 228, + 594 + ], + "spans": [ + { + "bbox": [ + 107, + 583, + 228, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 107, + 583, + 228, + 594 + ], + "type": "text", + "content": ": belief mass vector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "The belief mass vector " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " contains the belief masses " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{m}(\\omega_j)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " for all hypotheses. The matrix " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " has ones as elements if the belief mass " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{m}(\\omega_j)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " contributes to the belief " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{bel}(\\omega_i)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\omega_j \\subseteq \\omega_i" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "), and zeros otherwise. For a full definition of the entries, please see the supplemental in Sec. B. " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " is always invertible (see supplemental in Sec. C), and we can calculate the belief mass vector as " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m} = \\mathbf{S}^{-1}\\mathbf{b}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". In fact, our derived BBA is a generalization of the BBA proposed in [15] where there are only two conflicting measurements, namely" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "transmissions and reflections (without any class information), and the scale values are set to " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\alpha_{r} = -\\log (p_{\\mathrm{FP}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = -\\log (p_{\\mathrm{FN}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 114, + 537, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 537, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 537, + 128 + ], + "type": "text", + "content": "4.2. Evidential Semantic Occupancy Prediction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 132, + 554, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 554, + 192 + ], + "type": "text", + "content": "Towards the goal of predicting semantic occupancy grids from multi-view camera images, we present a novel loss formulation which allows to train models directly with the basic belief assignments obtained from our evidential grid mapping approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "text", + "content": "Therefore, we use an off-the-shelf occupancy prediction model [10] and only adapt the model outputs. Without changing the network architecture, we interpret its softmax output as a belief mass vector, and supervise it with the ground truth belief masses calculated according to Sec. 4.1. Training such a model requires a custom BBA-appropriate loss function. Several works attempted to define a cross-entropy analogue for BBAs [4-6, 38]. We adopt a differentiable formulation by Dezert and Dambreville [6], and directly use it as a loss between the predicted belief masses " + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "text", + "content": " and the ground-truth belief masses " + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 324, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 324, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 325, + 324, + 554, + 335 + ], + "type": "text", + "content": "For a single voxel " + }, + { + "bbox": [ + 325, + 324, + 554, + 335 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 325, + 324, + 554, + 335 + ], + "type": "text", + "content": ", the BBA cross-entropy is defined as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 351, + 554, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 351, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 332, + 351, + 554, + 423 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right) \\tag {11} \\\\ = \\sum_ {X \\subseteq \\Omega} \\left[ - \\mathrm {m} _ {i} ^ {\\mathrm {g t}} (X) \\left(1 - \\mathrm {u} _ {i} ^ {\\mathrm {g t}} (X)\\right) \\log \\left(\\mathrm {m} _ {i} ^ {\\mathrm {p r e d}} (X)\\right) \\right. \\\\ \\left. \\left. + \\mathrm {u} _ {i} ^ {\\mathrm {g t}} (X) \\left(1 - \\mathrm {m} _ {i} ^ {\\mathrm {p r e d}} (X)\\right) \\right] \\right.. \\\\ \\end{array}", + "image_path": "ed0b3534307cc062941788cb669cd061068afc01e08def19788472b6939c27c3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "text", + "content": "In addition to the belief mass vector " + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i^{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "text", + "content": " the uncertainty vector " + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "text", + "content": " is needed, which is defined as the difference between the plausibility and belief (see Eq. (4)). This can be easily calculated from " + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 432, + 555, + 480 + ], + "type": "text", + "content": " using Eq. (2) and Eq. (3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 480, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 517 + ], + "type": "text", + "content": "Similar to its classical counterpart, the definition of the BBA cross-entropy is not symmetric, but it always satisfies the inequality:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 332, + 522, + 554, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 522, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 332, + 522, + 554, + 538 + ], + "type": "interline_equation", + "content": "\\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right) \\geq \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {g t}}\\right) \\geq 0. \\tag {12}", + "image_path": "0ee4885e96d59471a94c49f0618af2f28be05351e9d57b14dcba7f8e39d843f0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": "Importantly, the left equality is achieved, if and only if the two BBAs are equal, i.e., " + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i^{\\mathrm{gt}} = \\mathbf{m}_i^{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": ". Additionally, if the BBAs " + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i^{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": " are Bayesian (i.e., " + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\operatorname{prob}(\\omega) = \\operatorname{bel}(\\omega) = \\operatorname{pl}(\\omega)" + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": ") and therefore " + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\mathfrak{u}(\\omega) = 0" + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\forall \\omega \\subseteq \\Omega" + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}_{\\mathrm{BBA}}(\\mathbf{m}_i^{\\mathrm{gt}}, \\mathbf{m}_i^{\\mathrm{pred}})" + }, + { + "bbox": [ + 313, + 544, + 554, + 607 + ], + "type": "text", + "content": " becomes the classical cross-entropy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 607, + 554, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 554, + 678 + ], + "type": "text", + "content": "We have observed that the distribution of belief masses in the ground-truth is highly imbalanced (see supplemental in Fig. 1). To compensate for that, we introduce the per-voxel weight " + }, + { + "bbox": [ + 313, + 607, + 554, + 678 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 607, + 554, + 678 + ], + "type": "text", + "content": " and balance the contribution of individual voxels based on their frequency. The final loss is a weighted sum of the BBAs cross-entropy values of all voxels:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 367, + 685, + 554, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 685, + 554, + 710 + ], + "spans": [ + { + "bbox": [ + 367, + 685, + 554, + 710 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i} w _ {i} \\mathrm {C E} _ {\\mathrm {B B A}} \\left(\\mathbf {m} _ {i} ^ {\\mathrm {g t}}, \\mathbf {m} _ {i} ^ {\\mathrm {p r e d}}\\right). \\tag {13}", + "image_path": "d990ea3164adc3c96f44d9ac8e90380bfdf40fc0cc74c6f43a12c112f46875d6.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27471" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 178, + 219 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 178, + 219 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 178, + 219 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 178, + 219 + ], + "type": "image", + "image_path": "6c79d15bdfc860da2564fd8b6372ed8c8e212e6f41c81b4f2863b9633905610f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 295 + ], + "type": "text", + "content": "Figure 3. Semantic occupancy grid maps. We visualize the ground-truth semantic occupancy grid maps from Occ3D [31] and our method on samples from the nuScenes [2] dataset. Semantic labels are assigned to each voxel based on the BBA (see Sec. 4.2). Additionally, we visualize predictions of BEVDet4D-Occ [10] trained with the semantic occupancy data of Occ3D [31] (baseline) and trained with our framework. Both models use the camera images as input. In the first and second example, the baseline model predicts arbitrary flying voxels, whereas our predictions are much cleaner. In the third example, our model predicts the shape of the trees much better. The last example shows how our GT occupancy is better capable of separating the pedestrians." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 70, + 304, + 219 + ], + "blocks": [ + { + "bbox": [ + 182, + 70, + 304, + 219 + ], + "lines": [ + { + "bbox": [ + 182, + 70, + 304, + 219 + ], + "spans": [ + { + "bbox": [ + 182, + 70, + 304, + 219 + ], + "type": "image", + "image_path": "c19e57d0699cd586f0480eaec120b60311f5a80e20427f0f0b7592df0472351f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 429, + 218 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 429, + 218 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 429, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 429, + 218 + ], + "type": "image", + "image_path": "3da5e910afb6463b5209ba47215cad12cb64ec249279f08d382e71f39a164abf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 555, + 218 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 555, + 218 + ], + "type": "image", + "image_path": "b12cba98482901fbbe2a6a12baf9decaa8ab1a1e74ed4ee8f1a61bc2955f9744.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 304, + 221, + 430 + ], + "blocks": [ + { + "bbox": [ + 57, + 304, + 221, + 430 + ], + "lines": [ + { + "bbox": [ + 57, + 304, + 221, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 221, + 430 + ], + "type": "image", + "image_path": "88c3cb7bc1e590a04e56bdb57b44b927302d34ddc3b2e62bdfe61bd67dc2b7a5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 437, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 555, + 480 + ], + "type": "text", + "content": "Figure 4. Birds-eye-view uncertainty. For each of the three scenes, we visualize the input camera images on the top. For each scene, we show the predicted semantics of our model on the left. On the right we depict the predicted uncertainties " + }, + { + "bbox": [ + 55, + 437, + 555, + 480 + ], + "type": "inline_equation", + "content": "\\mathfrak{m}(\\Omega)" + }, + { + "bbox": [ + 55, + 437, + 555, + 480 + ], + "type": "text", + "content": " averaged along the height dimension. The model predicts high uncertainties (colored in red) mainly in occluded areas, e.g., behind walls. Clearly visible areas obtain much lower uncertainties colored in blue." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 223, + 304, + 387, + 429 + ], + "blocks": [ + { + "bbox": [ + 223, + 304, + 387, + 429 + ], + "lines": [ + { + "bbox": [ + 223, + 304, + 387, + 429 + ], + "spans": [ + { + "bbox": [ + 223, + 304, + 387, + 429 + ], + "type": "image", + "image_path": "8eb1db8c4dcc784ff0e062f7e012709e14c71315077cd2257312ea10eef206db.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 388, + 304, + 553, + 429 + ], + "blocks": [ + { + "bbox": [ + 388, + 304, + 553, + 429 + ], + "lines": [ + { + "bbox": [ + 388, + 304, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 388, + 304, + 553, + 429 + ], + "type": "image", + "image_path": "5af6c9053af8ddbd36951d7c194a1ff56db68d549a569a15aee5e54472ef6efb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 297, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 297, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 297, + 527 + ], + "type": "text", + "content": "The weight " + }, + { + "bbox": [ + 55, + 502, + 297, + 527 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 502, + 297, + 527 + ], + "type": "text", + "content": " is calculated from the inverse square root of the mean BBA over all voxels " + }, + { + "bbox": [ + 55, + 502, + 297, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{m}_{\\mathrm{mean}}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 55, + 502, + 297, + 527 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 533, + 296, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 533, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 59, + 533, + 296, + 583 + ], + "type": "interline_equation", + "content": "w _ {i} = \\sum_ {X \\subseteq \\Omega} \\frac {\\mathrm {m} _ {i} ^ {\\mathrm {g t}} (X)}{\\sqrt {\\mathrm {m} _ {\\text {m e a n}} ^ {\\mathrm {g t}} (X)}} \\quad \\text {w i t h} \\quad \\mathbf {m} _ {\\text {m e a n}} ^ {\\mathrm {g t}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbf {m} _ {i} ^ {\\mathrm {g t}}. \\tag {14}", + "image_path": "4977da78cef186e25874fd144ba5dfc8c3efde61a26f3ec883cf4d697068e79f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 591, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 628 + ], + "type": "text", + "content": "This per-voxel weighting increases the contribution of rare voxels and decreases the contribution of very common voxels, e.g., those with total uncertainty " + }, + { + "bbox": [ + 55, + 591, + 296, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{m}_i^{\\mathrm{gt}}(\\Omega) = 1)" + }, + { + "bbox": [ + 55, + 591, + 296, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": "Decision-making based on BBAs. In order to finally predict each voxel's state, we need to make an informed decision based on the BBA of this voxel. To this end, we calculate the relative entropy " + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{D_{BBA}(\\mathbf{m}_i||\\mathbf{m}) = CE_{BBA}(\\mathbf{m}_i,\\mathbf{m}) - CE_{BBA}(\\mathbf{m}_i,\\mathbf{m}_i)}" + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": " between our predicted BBA " + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": " and BBAs " + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": " is focused on a specific hypothesis such that" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i(\\omega) = 1" + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\omega = \\omega_{i}" + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "text", + "content": " and 0 otherwise. The BBA " + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "text", + "content": " and the respective hypothesis " + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 502, + 555, + 538 + ], + "type": "text", + "content": ", which minimizes the relative entropy, is selected as the decision [6]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 544, + 555, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 544, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 328, + 544, + 555, + 564 + ], + "type": "interline_equation", + "content": "\\omega^ {*} = \\omega_ {i ^ {*}} \\quad \\text {w i t h} \\quad i ^ {*} = \\underset {i} {\\arg \\min } \\mathrm {D} _ {\\mathrm {B B A}} (\\mathbf {m} _ {i} \\| \\mathbf {m}). \\tag {15}", + "image_path": "3c53c8a503514f3984230d4f91f4527aaffaa5fa2b6990fadfbb6c6e22f47b3b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 629 + ], + "type": "text", + "content": "In our particular case, there are two decisions to be made. First, we need to decide if a voxel is free or occupied by anything (class-agnostic). Second, we need to determine the semantic classes of all occupied voxels. This leads to the following decision rules:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 630, + 556, + 715 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "text", + "content": "- Occupancy state: We define " + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{m_{occ}}(\\omega) = 1" + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\omega = \\mathcal{O}" + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "text", + "content": " and 0 otherwise. Similarly, we define " + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{m}_{\\mathrm{free}}(\\omega) = 1" + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\omega = \\mathcal{F}" + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "text", + "content": " and 0 otherwise. A voxel is considered occupied if " + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{D_{BBA}}(\\mathbf{m_{occ}}||\\mathbf{m}) < \\mathrm{D_{BBA}}(\\mathbf{m_{free}}||\\mathbf{m})" + }, + { + "bbox": [ + 314, + 630, + 555, + 677 + ], + "type": "text", + "content": " and free otherwise." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "text", + "content": "- Semantic class: For each class " + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{m}_i(\\omega) = 1" + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\omega = \\mathbf{c}_i" + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "text", + "content": ", and 0 otherwise. An occupied voxel is assigned the semantic class " + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{D_{BBA}}(\\mathbf{m}_i||\\mathbf{m}) < \\mathrm{D_{BBA}}(\\mathbf{m}_j||\\mathbf{m})" + }, + { + "bbox": [ + 314, + 677, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27472" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": "for all " + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": ", i.e., the relative entropy between the BBA " + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": " and the BBA " + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": " is minimal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "type": "text", + "content": "Note that this decision step is only applied before the final evaluation, which requires a discrete class for each voxel. During model training we directly use the belief masses to supervise our model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 138, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 138, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 138, + 169 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 297, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 297, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 297, + 225 + ], + "type": "text", + "content": "We evaluate our method on the nuScenes dataset [2], which is most common in the field of occupancy prediction [15, 31, 32, 34, 35]. Additional results for the Waymo dataset [30] are provided in the supplementary material in Sec. F." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 173, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 173, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 173, + 243 + ], + "type": "text", + "content": "5.1. Evaluation Protocol" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 249, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 333 + ], + "type": "text", + "content": "The evaluation of occupancy grids is a challenging task, since a perfect ground-truth occupancy grid is not available. However, we can evaluate the quality of different occupancy grid maps by comparing them to the LiDAR measurements, which provide the best common reference and source of ground-truth. We follow the evaluation protocol of [15] and report additional error metrics for the semantic labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": "Depth Evaluation. To evaluate the geometry of the occupancy grid maps, we compare the measured depth of the LiDAR " + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "inline_equation", + "content": "d_{i}^{\\mathrm{lidar}}" + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": " scans with the rendered depth from the occupancy grid " + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "inline_equation", + "content": "d_{i}^{\\mathrm{pred}}" + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": ". To do so, we cast a ray from the LiDAR sensor origin in the direction of the " + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": " LiDAR measurement. We terminate the ray at the first intersection with the surface of a voxel predicted to be occupied. We then extract the length of this rendered ray until the intersection point and call it the rendered depth " + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "inline_equation", + "content": "d_{i}^{\\mathrm{pred}}" + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": ". Finally, we compare the rendered depth " + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "inline_equation", + "content": "d_{i}^{\\mathrm{pred}}" + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": " to the measured depth " + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "inline_equation", + "content": "d_{i}^{\\mathrm{lidar}}" + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": " of the LiDAR using established depth metrics. In Tab. 2 we report the mean absolute error (MAE) and root mean squared error (RMSE), as well as the percentage of points with a relative error " + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 348, + 297, + 518 + ], + "type": "text", + "content": " below a certain threshold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "text", + "content": "Semantic Evaluation. Similar to the depth evaluation, we cast a ray from the LiDAR sensor origin in the direction of each LiDAR measurement. Instead of comparing the depths, we compare the semantic labels of the first occupied voxels along each ray " + }, + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "inline_equation", + "content": "l_i^{\\mathrm{pred}}" + }, + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "text", + "content": " with the semantic labels of the labeled LiDAR scans " + }, + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "inline_equation", + "content": "l_i^{\\mathrm{lidar}}" + }, + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "text", + "content": " given in the nuScenes dataset [2]. We report the intersection-over-union (IoU) per class, as well as the mean IoU (mIoU) averaged across all classes in Tab. 3. Note that this metric does not take the depth errors into account: if the LiDAR ray happens to have the same label as the first voxel they intersect, this prediction will be counted as correct even though the depth might be wrong." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To quantify both, the depth and the class accuracy at the same time, we calculate the fraction of points which have a relative depth error below a certain threshold " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and have" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "the correct semantic label as given by the labeled LiDAR scans. This metric is stricter than the relative depth accuracy, as it requires the correct semantic label in addition to the correct depth." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 410, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 410, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 410, + 146 + ], + "type": "text", + "content": "5.2. Model Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 152, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 297 + ], + "type": "text", + "content": "To demonstrate the impact of the more accurate GT data provided by our method, we compare BEVDet4D-Occ [10, 11] trained on our semantic occupancy data and our proposed supervision to the same model trained on the commonly used Occ3D [31] dataset. We use ResNet-50 [9] as the backbone for BEVDet4D-Occ [10, 11]. The model predicts semantic occupancy grid maps, using the surround-view camera images as model inputs. Importantly, we do not change the model architecture, but only the loss and the number of outputs in the last layer to match the number of hypotheses. Additionally, we use our generated semantic occupancy data instead of Occ3D's data for supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 298, + 556, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 556, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 556, + 406 + ], + "type": "text", + "content": "BEVDet4D-Occ uses additional depth supervision during training, which is not required for our occupancy GT. We start training from a model checkpoint trained on Occ3D [31]. We continue training for 24 epochs using the same hyperparameters, but only changing the model supervision to our evidential formulation. Finally, we reduce the learning rate by a factor of 10 and continue training for 24 more epochs. The baseline model follows the same training schedule, but is supervised with Occ3D's data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 419, + 531, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 531, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 531, + 432 + ], + "type": "text", + "content": "5.3. Evaluation of Semantic Occupancy Grids" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "text", + "content": "Semantic Occupancy Ground Truth. We compare our semantic occupancy grids with those of SurroundOcc [35], Scene as Occupancy [32], Occ3D [31], and Käble et al. [15] on the nuScenes dataset [2]. The former methods provide semantic occupancy grids, while the last method provides the state of the art only for binary occupancy grids. Therefore, we can only report the depth evaluation results for [15] and the additional semantic evaluation as described in Sec. 5.1 for all other methods. We show qualitative results in Fig. 3, visualizing the generated occupancy GT from Occ3D [31] and our method. Additionally, we show the corresponding model predictions as well as the front camera images. The depth evaluation results are summarized in Tab. 2. Our method consistently outperforms all baselines across different metrics and voxel grid resolutions. Through our extended hyperparameter search space for " + }, + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{r}" + }, + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "text", + "content": " and novel decision-making scheme, we even outperform [15] in terms of geometric accuracy. Tab. 3 provides quantitative results for the ray-based IoU for the semantic classes. Also in this case our method consistently outperforms all the baselines, which validates that our approach for semantic occupancy using evidence theory is superior to simpler heuristic-based methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27473" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 70, + 492, + 242 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 492, + 242 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 492, + 242 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 492, + 242 + ], + "type": "table", + "html": "
Voxel sizeRMSEδ<1.25δ<1.252δ<1.253δclsδcls2
in min min min %in %in %in %in %in %in %in %in %in %in %
Occupancy GTSurroundOcc [35]0.52.805.3456.281.188.649.869.072.6
Scn. as Occ. [32]0.52.925.1746.773.486.641.462.370.4
Occ3D [31]0.41.894.3381.589.493.074.278.780.1
Käblle et al. [15]0.41.253.4287.994.696.8---
EvOcc (ours)0.41.212.9890.195.897.683.486.687.2
OpenOccupancy [34]0.22.084.3570.683.689.565.174.277.3
Käblle et al. [15]0.20.912.9992.696.297.7---
EvOcc (ours)0.20.842.6293.896.998.387.188.588.9
Pred.Occ3D [31]0.42.334.3158.889.694.350.474.376.4
EvOcc (ours)0.41.664.1888.694.396.776.679.580.4
", + "image_path": "981b3b2781e3d19469b345591978ec1a080247aa414d512f916df56253ade108.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 305, + 555, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": "Table 2. Depth evaluation. We compare different methods to generate occupancy GT data as well as trained models on the nuScenes [2] dataset, following the evaluation procedure described Sec. 5.1. The voxel size and depth errors are in meters, while the remaining numbers are in percent. Our method provides occupancy maps that match the LiDAR scans much better compared to other methods using a comparable voxel size. We see the same for BEVDet4D-Occ [10] trained with our evidential framework as compared to training on Occ3D [31]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 305, + 555, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 305, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 555, + 446 + ], + "type": "table", + "html": "
Voxel sizeBarrierBicycleBusCarConstruction vehicleMotorcyclePedestrianTraffic coneTrailerTruckDriveable surfaceOther flatSidewalkTerrainManmadeVegetationMean
Occupancy GTSurroundOcc [35]0.549.612.967.048.134.728.119.39.142.558.179.931.844.445.348.647.441.7
Scn. as Occ. [32]0.550.013.970.752.143.830.320.68.852.763.877.528.738.639.449.748.043.0
Occ3D [31]0.465.519.074.258.048.532.724.614.653.766.484.850.658.956.555.954.551.2
EvOcc (ours)0.472.934.785.177.171.857.545.224.880.083.289.059.666.668.977.777.267.0
OpenOcc. [34]0.266.226.677.967.856.447.136.916.957.574.481.844.351.050.856.453.454.1
EvOcc (ours)0.282.145.289.783.378.269.356.837.184.987.989.070.976.176.582.781.374.4
Pred:Occ3D [31]0.453.615.368.755.233.126.323.510.339.552.882.031.343.143.256.255.543.1
EvOcc (ours)0.461.35.076.669.238.832.125.613.346.963.286.944.053.756.858.646.248.6
", + "image_path": "79b4e3ac60a260efb17420f3be94f6a34be1a8af6203448a8494487eb646258a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 454, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 454, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 555, + 499 + ], + "type": "text", + "content": "Table 3. Semantic evaluation. We compare the semantic occupancy grid maps of different semantic occupancy GT reconstructions, as well as those from trained models, to the LiDAR scans of the nuScenes [2] dataset following the evaluation procedure from Sec. 5.1. The voxel size is in meters, while the remaining numbers are in percent. Our method consistently yields more accurate GT semantic occupancy grid maps and leads to better trained models. All numbers are in percent." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 520, + 297, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 297, + 688 + ], + "type": "text", + "content": "Semantic Occupancy Prediction. We compare the model predictions of the models trained on our semantic occupancy data with the models trained on the Occ3D [31] data. We evaluate the same metrics for the model predictions as for the occupancy GT (see Sec. 5.1). We report the quantitative results in Tab. 2 and Tab. 3. BEVDet4D-Occ trained on our data yields more accurate predictions in terms of geometry and semantics. Qualitative results of the model semantic predictions and corresponding GT occupancy grids are shown in Fig. 3. Additionally, we visualize the predicted semantics together with the aggregated uncertainty " + }, + { + "bbox": [ + 55, + 520, + 297, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{m}(\\Omega)" + }, + { + "bbox": [ + 55, + 520, + 297, + 688 + ], + "type": "text", + "content": " along the height dimension in Fig. 4. Furthermore, we compare our ground-truth uncertainty and the predicted uncertainty in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 518, + 388, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 518, + 388, + 531 + ], + "spans": [ + { + "bbox": [ + 314, + 518, + 388, + 531 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 538, + 556, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 556, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 556, + 707 + ], + "type": "text", + "content": "In this work, we presented a novel approach to obtain high-quality 3D semantic occupancy maps using ET. Our approach introduces a new BBA formulation for modeling LiDAR measurements with partial semantic annotations. We compared our generated ground truth to that of other state-of-the-art semantic occupancy datasets and showed that our data matches the LiDAR measurements better by a large margin in terms of geometry and semantics. We also demonstrated that using our ground truth for training the semantic occupancy prediction model BEVDet4D-Occ significantly improves the model's performance. In conclusion, our work sets a new state of the art for semantic occupancy ground truth generation and model training, and therewith makes an important contribution towards autonomous driving." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27474" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "text", + "content": "[1] Stefan Andreas Baur, David Josef Emmerichs, Frank Moosmann, Peter Pinggera, Björn Ommer, and Andreas Geiger. SLIM: Self-supervised LiDAR scene flow and motion segmentation. In ICCV, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 189 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3, 4, 6, 7, 8, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 190, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 190, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 190, + 296, + 223 + ], + "type": "text", + "content": "[3] Arthur P Dempster. A generalization of bayesian inference. Journal of the Royal Statistical Society: Series B (Methodological), 1968. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 246 + ], + "type": "text", + "content": "[4] Yong Deng. Uncertainty measure in evidence theory. Science China Information Sciences, 63, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "text", + "content": "[5] Jean Dezert. An effective measure of uncertainty of basic belief assignments. 2022 25th International Conference on Information Fusion (FUSION), pages 1-8, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[6] Jean Dezert and Frédéric Dambreville. Cross-entropy and relative entropy of basic belief assignments. 2023 26th International Conference on Information Fusion (FUSION), pages 1–8, 2023. 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 324, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 296, + 346 + ], + "type": "text", + "content": "[7] Alberto Elfes. Using occupancy grids for mobile robot perception and navigation. Computer, 22(6):46-57, 1989. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 346, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 379 + ], + "type": "text", + "content": "[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In CVPR, 2012. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 380, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 411 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 413, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 413, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 57, + 413, + 295, + 445 + ], + "type": "text", + "content": "[10] Junjie Huang and Guan Huang. BEVDet4D: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 447, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 447, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 57, + 447, + 295, + 489 + ], + "type": "text", + "content": "[11] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 491, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 295, + 524 + ], + "type": "text", + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, 2023. 1, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "text", + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, page 0, 2006. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 569, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 601 + ], + "type": "text", + "content": "[14] J. Kim et al. Evidential semantic mapping in off-road environments with uncertainty-aware bayesian kernel inference. IROS, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "text", + "content": "[15] Jonas Käblé, Sascha Wirges, Maxim Tatarchenko, and Eddy Ilg. Accurate training data for occupancy map prediction in automated driving using evidence theory. In CVPR, 2024. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 669 + ], + "type": "text", + "content": "[16] J. Li et al. Segm: A novel semantic evidential grid map by fusing multiple sensors. In PRCV, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of reliable depth for multi-view 3D object detection. In AAAI, pages 1477-1485, 2023. 1, 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "type": "text", + "content": "[18] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: Monocular 3D semantic scene completion benchmark in street views. arXiv preprint arXiv:2306.09001, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 139, + 555, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 139, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 555, + 192 + ], + "type": "text", + "content": "[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, Cham, 2022. 1, 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 194, + 555, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 555, + 236 + ], + "type": "text", + "content": "[20] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M. Alvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 238, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 555, + 281 + ], + "type": "text", + "content": "[21] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M. Alvarez. FB-BEV: BEV representation from forward-backward view transformations. In ICCV, pages 6919-6928, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 283, + 555, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 555, + 315 + ], + "type": "text", + "content": "[22] Haisong Liu, Yang Chen, Haiguang Wang, Zetong Yang, Tianyu Li, Jia Zeng, Li Chen, Hongyang Li, and Limin Wang. Fully sparse 3d occupancy prediction. In ECCV, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 316, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 348 + ], + "type": "text", + "content": "[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position embedding transformation for multi-view 3D object detection. In ECCV, pages 531-548. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 350, + 555, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 413 + ], + "type": "text", + "content": "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 415, + 555, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 446 + ], + "type": "text", + "content": "[25] H. P. Moravec. Sensor fusion in certainty grids for mobile robots. Sensor Devices and Systems for Robotics, 9(2):253-276, 1989. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 449, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 491 + ], + "type": "text", + "content": "[26] Sven Richter, Frank Bieder, Sascha Wirges, and Christoph Stiller. A dual evidential top-view representation to model the semantic environment of automated vehicles. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 493, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 514 + ], + "type": "text", + "content": "[27] S. Richter et al. Semantic evidential grid mapping using monocular and stereo cameras. Sensors, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 514, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 554, + 525 + ], + "type": "text", + "content": "[28] Glenn Shafer. A mathematical theory of evidence. 1976. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 526, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 555, + 557 + ], + "type": "text", + "content": "[29] Philippe Smets. Data fusion in the transferable belief model. Proceedings of the Third International Conference on Information Fusion, 1:PS21-PS33 vol.1, 2000. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 559, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 646 + ], + "type": "text", + "content": "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. CVPR, 2019. 2, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "text", + "content": "[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3D: A large-scale 3d occupancy prediction benchmark for autonomous driving. In NeurIPS, 2023. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[32] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin," + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27475" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 397 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "and Hongyang Li. Scene as occupancy. In ICCV, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[33] Ignacio Vizzo, Tiziano Guadagnino, Jens Behley, and Cyril Stachniss. VDBFusion: Flexible and efficient TSDF integration of range sensor data. Sensors, 22(3):1296, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 184 + ], + "type": "text", + "content": "[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. OpenOccupancy: A large scale benchmark for surrounding semantic occupancy perception. In ICCV, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 219 + ], + "type": "text", + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. SurroundOcc: Multi-camera 3d occupancy prediction for autonomous driving. In CVPR, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 297 + ], + "type": "text", + "content": "[36] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. Curran, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "type": "text", + "content": "[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "type": "text", + "content": "[38] Hangyu Yan and Yong Deng. An improved belief entropy in evidence theory. IEEE Access, 8:57505-57516, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "text", + "content": "[39] T. Yang and V. Aitken. Evidential mapping for mobile robots with range sensors. IEEE Transactions on Instrumentation and Measurement, 55:1422-1429, 2006. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27476" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_content_list.json b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..34768c1bb02769bc2d97c5c794116fa10b85b143 --- /dev/null +++ b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_content_list.json @@ -0,0 +1,1468 @@ +[ + { + "type": "text", + "text": "Eval3D: Interpretable and Fine-grained Evaluation for 3D Generation", + "text_level": 1, + "bbox": [ + 143, + 130, + 852, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shivam Duggal\\*1 Yushi Hu\\*2 Oscar Michel3 Aniruddha Kembhavi4 William T. Freeman1 Noah A. Smith2,5 Ranjay Krishna2,5 Antonio Torralba1 Ali Farhadi2,5 Wei-Chiu Ma6", + "bbox": [ + 114, + 179, + 879, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Massachusetts Institute of Technology 2University of Washington 3New York University 4Wayve 5Allen Institute for AI 6Cornell University", + "bbox": [ + 130, + 232, + 864, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a31224bffda5f833cbf114807fef4e21b412f738e41b98a18f9a4fe00a357813.jpg", + "image_caption": [ + "Figure 1. Challenges of 3D generation: (1) Structural inconsistency: lack of globally-coherent 3D shape; (2) Text-3D misalignment: failure to meet the requirements of the input text-prompt; (3) Semantic inconsistency: content change and incoherent semantics; (4) Geometric inconsistency: misaligned geometry and texture." + ], + "image_footnote": [], + "bbox": [ + 109, + 280, + 880, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 520, + 326, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the unprecedented progress in the field of 3D generation, current systems still often fail to produce high-quality 3D assets that are visually appealing and geometrically and semantically consistent across multiple viewpoints. To effectively assess the quality of the generated 3D data, there is a need for a reliable 3D evaluation tool. Unfortunately, existing 3D evaluation metrics often overlook the geometric quality of generated assets or merely rely on black-box multimodal large language models for coarse assessment. In this paper, we introduce Eval3D, a fine-grained, interpretable evaluation tool that can faithfully evaluate the quality of generated 3D assets based on various distinct yet complementary criteria. Our key observation is that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. We thus leverage a diverse set of models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. Compared to prior work, Eval3D provides pixel-wise measurement, enables accurate 3D spatial feedback, and aligns more closely with human judgments. We comprehensively evaluate existing 3D generation models using Eval3D and highlight the limitations and challenges", + "bbox": [ + 88, + 553, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of current models. Project page: http://eval3d.github.io.", + "bbox": [ + 511, + 522, + 880, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 547, + 643, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality 3D assets play a pivotal role across various domains including gaming, film, augmented/virtual reality, and robotics [24, 54, 65]. They empower content creation applications to deliver visually compelling and immersive experiences and enable simulators to realistically reproduce the real world. Unfortunately, 3D asset creation is complex and tedious, requiring a lot of careful manual effort. Automatic (text/image-to-)3D generation, despite fast progress over the last few years, still falls short in key aspects including geometry, semantic consistency, etc.", + "bbox": [ + 509, + 565, + 908, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What makes a good (text/image-to-)3D generation? 3D assets are not created to merely be rendered onto a 2D image; rather, they are meant to live within a rich 3D world – the assets may be viewed from various distinct viewpoints, and their 3D geometry will affect how they interact with the surroundings (e.g., collision, reflection, etc). Therefore, a high-quality 3D asset should not only be visually appealing and faithful to its text/image input across multiple individual viewpoints, but its geometry and appearance should also be coherent and plausible when observed as a whole. In the absence of these characteristics, the utility of the generated", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13326", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7a1e000e80b2f2cf3e833389166e77c350723a3498bf748527ebe28c5c1835b.jpg", + "image_caption": [ + "Figure 2. Eval3D offers interpretable, fine-grained, and human-aligned metrics to assess the quality of 3D generations from various aspects. We utilize a diverse array of foundation models and tools to measure the consistency among different representations of generated 3D assets." + ], + "image_footnote": [], + "bbox": [ + 153, + 58, + 844, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "assets degrades drastically. Consider the generated 3D assets in Fig. 1. While the corgi (Fig. 1 (left)) looks reasonable from the front and the back, its overall 3D shape is quite poor and implausible – notice the Janus issue (i.e., presenting multiple faces). Similarly, while the fried chickens and waffles (Fig. 1(right)) appear faithful from one viewpoint, the 3D asset suffers from semantic and geometric inconsistency across views – as the object rotates, the maple syrup transform into a jar, and the fried chicken pieces merge with the waffles. These 3D-specific structural and consistency challenges continue to pose significant difficulties to existing (text/image-to-)3D generation models.", + "bbox": [ + 88, + 386, + 485, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What makes a good 3D evaluation metric? To advance (text/image-to-)3D asset generation and to effectively quantify the aforementioned issues, there is a need for a reliable 3D evaluation tool. Unfortunately, while significant progress has been made in evaluating 2D text-to-image models [10, 21, 59, 62, 67], the development of their 3D counterparts is lagging behind. Existing 3D evaluation metrics typically rely on black-box multimodal large language models (LLMs), such as GPT-4V, for assessment [61], or often overlook the geometric quality of generated assets [19]. A comprehensive 3D metric should be able to: (a) rate the consistency between the input text/image and the 3D generation and detect misalignments such as a missing object or part, wrong attribute, etc., (b) measure the structural and semantic consistency of the generated asset across the input and the output as well as across multiple views of the 3D asset, and (c) localize flaws and inconsistencies in the produced asset. Such a comprehensive metric would not only permit fine-grained comparisons between different 3D generations but also provide feedback on where and what to improve.", + "bbox": [ + 89, + 598, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method: We introduce Eval3D, an interpretable, and fine-grained evaluation tool that measures the quality of generated 3D assets from multiple dimensions. We build Eval3D based on the observation that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. For instance, to measure the geometric consistency of the generated chicken and waffles (i.e., the red region in Fig. 1(right)), one can derive the surface normal analytically from the 3D object, and then compare it with the normal estimated by a foundation model (e.g., Depth Anything [64]) from rendered RGB images. Based on 2D visual priors, waffles usually possess a grid-like 3D geometry. However, the grid-like patterns do not exist in the normal map derived from 3D. By computing the differences, we can automatically identify those regions with inconsistent surface normals and localize where the generative model leverages texture maps to \"cheat.\" We adopt a diverse set of open-source foundation models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. To aggregate the measurements across multiple viewpoints, we further present several simple yet effective 2D and 3D-based fusion mechanisms. With these fine-grained, pixel-wise measurements, we can not only easily locate the problematic regions in 2D images (Fig. 3), but also lift the inconsistency map back to 3D and enable accurate spatial artifact localization (Fig. 5). To our knowledge, we are the first to provide such spatial feedback.", + "bbox": [ + 511, + 386, + 908, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To promote the use of our new evaluation tool, we curate a diverse set of prompts and images from multiple sources [18, 19, 51]. We evaluate a variety of (text/image-to-)3D generation models on our data and densely annotate human preferences based on different criteria to verify how well our metrics align with human assessment. We show that Eval3D", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13327", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5061bc2869063ff17218d4cd0cac07416692365b8fe45dc9ddcd7df336f02e8b.jpg", + "image_caption": [ + "Figure 3. Geometry inconsistency evaluates texture-geometry misalignment by comparing 3D rendered normal and image-based normal. Bright-yellow indicates large discrepancy." + ], + "image_footnote": [], + "bbox": [ + 150, + 73, + 843, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is not only able to automatically detect various generation artifacts (e.g., the Janus issue), but also accurately localize them (Fig. 3). Surprisingly, we also find that while existing state-of-the-art generations are visually appealing, they often still suffer from geometric or semantic inconsistencies. Finally, we demonstrate that although the accuracy of Eval3D is dependent on the accuracy of the foundation models, it consistently aligns more closely with, or is comparable to, human judgments across all dimensions than prior work. As foundation models will improve over time, we hypothesize that Eval3D will become increasingly reliable.", + "bbox": [ + 88, + 266, + 483, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 462, + 232, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative models for 3D asset generation: Early work on 3D asset generation trained probabilistic generative models directly on 3D data, most commonly using assets from ShapeNet [6]. While numerous generative modeling frameworks (e.g., GANs [8, 9, 16, 17, 26, 60], VAEs [4, 8, 12, 17, 40], flow models [5, 63]) and 3D representations (e.g., point clouds [5, 37, 63], meshes [16, 16], voxels [4, 9, 60], and implicit functions [8, 26, 32, 40]) have been explored, the quality and diversity of 3D generations are constrained by the quantity and characteristics of the available data. With this in mind, recent dominant approaches turn to leverage supervision from foundation models trained on massive collections of 2D images [44-46]. During generation, they propagate the gradient from multi-view 2D space back to 3D and iteratively update the underlying 3D representations [8, 23, 25, 34, 36, 43]. Since foundation models encode rich priors of our world, this significantly improves the quality of the generated objects. To further improve multi-view consistency [3, 30, 48, 58, 69], object geometry [7, 28, 53], or even the optimization process [33, 57, 68], researchers have proposed a variety of enhancements, such as multi-view diffusion [49]. While a general trend of progress is evident, the lack of evaluation tools makes it unclear which methods are superior and in which dimensions. Our framework addresses this confusion by providing a general metric to measure progress, which can be decomposed to identify performance across critical factors relevant to 3D generation.", + "bbox": [ + 91, + 493, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Evaluating 3D generation:** Perhaps closest to our work are GPT-4V [61] and T3-Bench [19]. GPT-4V [61] is an ELOB-based method that uses the eponymous multimodal LLM to compare the quality of two generated objects, while T3-Bench [19] employs a vision-language model (VLM) to evaluate the semantic adherence of generated objects to the prompt. While we share similar insights, there are several key differences: first, unlike GPT-4V, which provides a pairwise preference, Eval3D offers an absolute metric, making it easier to quantify and compare performance across different models and tasks. Additionally, both GPT-4V and T3 rely mainly on language outputs for evaluation. The reliance on the inherently imprecise nature of language makes localizing the source of error difficult. This issue is particularly problematic for detecting multi-view inconsistencies, as each view may independently seem satisfactory to the VLM. In contrast, our method incorporates semantic judgments from VLMs [11, 21, 50] and uses pretrained computer vision models [39, 64] to provide precise and localized estimates of geometric and structural deviations at multiple scales of granularity. These error measurements can be back-projected to 3D, making Eval3D more informative.", + "bbox": [ + 511, + 266, + 908, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Eval3D", + "text_level": 1, + "bbox": [ + 511, + 608, + 601, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our aim is to develop an evaluation tool that can faithfully assess the quality of (generated) 3D assets based on various distinct yet complementary criteria. Unlike 2D images, which are merely projections of our visual world, these generated assets live in 3D. Their geometry influences how they interact with the environment, and they can be viewed from multiple angles. Therefore, it is crucial to ensure that the appearance of these 3D generations is coherent and visually appealing from various viewpoints, that their geometry is high-quality and consistent with their RGB counterparts, and that the generated content remains faithful to the input conditions (e.g., text, image).", + "bbox": [ + 511, + 627, + 908, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Towards this goal, we devise five distinct measurements to evaluate the quality of generated 3D assets from different aspects. Our key observation is that many of the desired 3D properties can be predicted from one another if the (generated) 3D asset is of high-quality; if the cues are unpredictable, then the 3D generation often has artifacts or is not", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13328", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "coherent. Building upon this intuition, we consolidate a diverse set of foundation models [1, 38, 39, 62, 64] and tools to compute and predict a wide range of visual cues and measure their consistency. The measurements across multiple viewpoints are then aggregated to reflect the quality of the 3D generation with respect to different aspects. For each critic, we scale the score to lie within $0 - 100\\%$ .", + "bbox": [ + 89, + 90, + 483, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notation: Let $t$ be the input text prompt and $\\mathcal{I}^{\\mathrm{GT}}$ be the optional image condition. Let $\\mathcal{O}_t$ be the text-conditioned 3D asset, where the underlying representation could be either a mesh $\\mathcal{O}_t^{\\mathrm{mesh}}$ or an implicit representation $\\mathcal{O}_t^{\\mathrm{imp}}$ . Let $\\{\\mathcal{I}_i\\}_{i=1}^N$ denote the rendered RGB images of the generated asset $\\mathcal{O}_t$ at viewpoint $v_i$ . Let $f$ denote the foundation models or tools that we adopt to extract visual cues.", + "bbox": [ + 89, + 196, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Geometric Consistency", + "text_level": 1, + "bbox": [ + 89, + 313, + 303, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate how good the geometry of an asset is, one naive way is to consume the 3D asset and output a score. Unfortunately, it is unclear what constitutes a good geometry and there is no such data or model that allows us to directly predict the geometric quality.", + "bbox": [ + 89, + 334, + 483, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we take a proxy and measure the consistency between the surface normal analytically derived from the 3D asset and the normal predicted by a foundation model from 2D images. The intuition is that a good 3D geometry should be reflective of its visual appearance. If the derived geometry does not align with the normal predicted from RGB (e.g., the chicken waffle in Fig. 1, right), then one of the components is incorrect. Specifically, we first compute the analytical surface normal of the 3D asset by taking the normal of each face in the mesh, or by taking the negative gradient of the density [2, 35] or signed distance field [55, 66], depending on the 3D representation. Then for each viewpoint, we exploit Depth Anything [64] to compute the monocular depth and convert it into normal. Empirically we find it more robust than direct surface normal estimation [13]. Finally, we project the analytical normal to each view, measure the pixel-wise angular difference, and compute the amount of inliers:", + "bbox": [ + 89, + 410, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric consistency $=$", + "bbox": [ + 112, + 695, + 282, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N _ {p}} \\sum_ {p} \\mathbb {1} \\left[ \\operatorname {a r c c o s} \\left(\\mathbf {n} _ {p} ^ {\\text {a n a l}} \\cdot \\mathbf {n} _ {p} ^ {\\text {p r e d}}\\right) < \\delta^ {\\text {n o r m}} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 714, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we iterate over all pixels $p$ . $\\mathbf{n}^{\\mathrm{anal}}$ indicates that the normal of pixel $p$ is analytically derived, while $\\mathbf{n}^{\\mathrm{pred}}$ suggests it is predicted. $N_{p}$ denotes the number of valid rendered pixels. In practice, we set the threshold $\\delta^{\\mathrm{norm}}$ to $23^{\\circ}$ based on a hold-out validation set.", + "bbox": [ + 89, + 762, + 483, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Semantic Consistency", + "text_level": 1, + "bbox": [ + 89, + 848, + 294, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another crucial factor for a 3D generation to be appealing and useful is that the underlying content and semantics", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "remain consistent and coherent across different viewpoints. Unlike visual appearance, which may be affected by extrinsic factors such as lighting and change across views, the semantics of an asset should be invariant to different perspectives. To quantify the degree to which the semantics change across views, we capitalize on large vision foundation models that encode rich semantics.", + "bbox": [ + 511, + 90, + 906, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first extract DINO features for every rendered image $\\mathcal{F}_i^{\\mathrm{DINO}} = f^{\\mathrm{DINO}}(\\mathcal{I}_i)$ . Then, for each 3D point $x$ on the surface of the asset $\\mathcal{O}$ , we project it onto 2D feature maps $\\{\\mathcal{F}_i^{\\mathrm{DINO}}\\}$ to retrieve its corresponding features $\\{\\mathcal{F}_i^{\\mathrm{DINO}}(\\pi_{v_i}(x))\\}$ . Here, $\\pi$ is the projection operator. Since a point may be visible from multiple images, we compute the variance of its DINO features. Specifically, we calculate the variance along each feature dimension separately across viewpoints and then take the mean of these variances across all feature dimensions. Finally, we compute the ratio of vertices whose variance is below a certain threshold. The intuition is that if the content and semantics of a 3D asset remain relatively stable across different viewpoints, its DINO features should be consistent. In practice, for computational efficiency, we only consider the vertices of the 3D mesh $x^{\\mathrm{vert}}$ as 3D points:", + "bbox": [ + 511, + 196, + 908, + 438 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S e m a n t i c c o n s i s t e n c y} = \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 448, + 906, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N _ {\\text {v e r t}}} \\sum_ {x ^ {\\text {v e r t}}} \\mathbb {1} \\left[ \\operatorname {m e a n} \\left(\\operatorname {V a r} \\left(\\left\\{\\mathcal {F} _ {i} ^ {\\mathrm {D I N O}} \\left(\\pi_ {v _ {i}} \\left(x ^ {\\text {v e r t}}\\right) \\right.\\right)\\right) < \\delta^ {\\mathrm {D I N O}} \\right] \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 465, + 901, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We set $\\delta^{\\mathrm{DINO}}$ to the 70th percentile of average DINO variance based on a hold-out validation set.", + "bbox": [ + 511, + 511, + 905, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Structural Consistency", + "text_level": 1, + "bbox": [ + 511, + 546, + 725, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A generated asset that looks plausible from a single view does not necessarily imply that it remains coherent when observed as a whole. The third criterion we are interested in measuring is whether the overall structure of the generated asset is coherent and plausible.", + "bbox": [ + 511, + 568, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We frame this assessment as a novel view synthesis (NVS) problem. Our intuition is that if the generated asset is overall plausible, one should be able to synthesize, to some degree, what other views might look like. For instance, given an image of the back of the head, one should be capable of predicting the face, although there might be variations in the size of the eyes or nose compared to the actual generation. On the other hand, if the generation suffers from artifacts or noise, it would be difficult to effectively conduct cross-view prediction, and the discrepancies could be significant.", + "bbox": [ + 511, + 643, + 906, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this motivation in mind, given a rendered RGB image $\\mathcal{I}_i$ as input, we first utilize Stable-Zero123 [1], a diffusion-based NVS model, to predict what an image from another unobserved viewpoint might look like, denoted as $\\mathcal{I}_{i\\rightarrow j} = f^{0123}(\\mathcal{I}_i;v_i,v_j)$ . We then measure the difference between the predicted image $\\mathcal{I}_{i\\rightarrow j}$ and the image rendered from the generated asset $\\mathcal{I}_j$ . Since we care more about the", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13329", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59250fd286ab4c24094ae601e98dc57a67321c176ad3c183ad777e2767297bda.jpg", + "image_caption": [ + "Figure 4. Structural consistency evaluates the geometric coherence of the generated 3D assets by comparing rendered views with the predictions from a novel view synthesis model (Zero-123) across various rotations. We utilize DreamSim to assess image similarity." + ], + "image_footnote": [], + "bbox": [ + 150, + 78, + 848, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "perceptual similarity rather than pixel-wise differences, we employ DreamSim [14], a human-aligned perceptual metric, for this similarity assessment (see Fig. 4 for overview). Finally, we repeat the same operation for multiple views $j$ and then take the average. In practice, to increase robustness, we consider different images $\\mathcal{I}_i$ as input and select the best one:", + "bbox": [ + 89, + 311, + 485, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Structural consistency $=$", + "bbox": [ + 114, + 409, + 282, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {i} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\left(1 - f ^ {\\text {D r e a m S i m}} \\left(\\mathcal {I} _ {i \\rightarrow j} ^ {\\text {p r e d}}, \\mathcal {I} _ {j}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 428, + 483, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We take as input both the canonical view (rendered at $0^{\\circ}$ ), and the rendered image from $90^{\\circ}$ . We predict novel viewpoints at an interval of $90^{\\circ}$ .", + "bbox": [ + 89, + 474, + 485, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Text-3D Alignment", + "text_level": 1, + "bbox": [ + 89, + 527, + 274, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another often overlooked aspect of 3D evaluation is the fidelity of the generated 3D asset to the input conditions. For instance, are all the mentioned objects being generated? Do the relationships among the objects align with what the text describes? The alignment issue is critical as it reflects how accurately the 3D generation adheres to the instructions.", + "bbox": [ + 89, + 549, + 483, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively measure the text-3D alignment, we draw inspiration from TIFA [21], an evaluation metric for text-to-image generation, and adapt it for assessing text-to-3D generation. The key idea is to measure the consistency among large language models (LLMs) and multimodal LLMs across multiple views, and use this consistency as a proxy for evaluation. Specifically, we first employ a LLM to generate $M$ multiple-choice question-answer tuples $\\{Q_j,C_j,A_j\\}_{j = 1}^M$ based on the input text prompt $t$ , where $Q_{j}$ is a question, $C_j$ is a set of answer choices, and $A_{j}\\in C_{j}$ is the gold answer. Then, for each question $Q_{j}$ , we employ a multimodal LLM to produce an answer based on each input image $A_{j}^{v_{i},\\mathrm{VQA}} = f^{\\mathrm{MLLM}}(\\mathcal{I}_{i},Q_{j})$ . If the answer from the multimodal LLM $A_{j}^{v_{i},\\mathrm{VQA}}$ aligns with the answer derived from the LLM $A_{j}$ , we conclude that the specific property in question is aligned at such viewpoint $v_{i}$ . For example, given a text prompt $t = \"\\text{A statue of a dog}\"$ ,", + "bbox": [ + 89, + 641, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the LLM could generate question-answer tuples such as {\"Is there a statue?\", \"Yes\", \"No\", \"Yes\"} and view-dependent alignment is assessed based on whether both the text and the image confirm the presence of a statue (see Fig. 2). To aggregate individual scores, we repeat the same procedure for all questions $\\{Q_j\\}_{j=1}^M$ and across all viewpoints $\\{v_i\\}_{i=1}^N$ :", + "bbox": [ + 511, + 311, + 906, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-3D alignment =", + "bbox": [ + 524, + 433, + 671, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{M} \\sum_ {j = 1} ^ {M} \\operatorname {A n y} _ {v _ {i}} \\left(\\operatorname {A l l} _ {v \\in \\operatorname {A d j} \\left(v _ {i}\\right)} \\left(\\mathbb {1} \\left[ A _ {j} = = A _ {j} ^ {v, \\mathrm {V Q A}} \\right]\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 453, + 906, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A question is considered globally aligned if and only if, for some viewpoint $v_{i}$ , the answers of all its neighboring views are aligned. We find that this method is effective for dealing with occlusions and the noise of multimodal LLM. In practice, we adopt LLaVA-NeXT-7B [27] as our VQA model, and we follow DSG [10] for generating and evaluating our question-answer pairs.", + "bbox": [ + 511, + 508, + 908, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Aesthetic", + "text_level": 1, + "bbox": [ + 513, + 625, + 622, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, given that a generated 3D asset can be viewed from a wide range of angles, it is crucial to ensure that it appears visually appealing across all possible views. In this paper, we exploit two different approaches, with varying computational cost, to compute the aesthetic score $f^{\\text{aesthetic}}(\\{\\mathcal{I}_i\\})$ . First, following [61], we use GPT-4o [38] to conduct pairwise quality comparisons of 3D assets and compute the ELO score of each model. We improve the prompts and normalize the ELO score to 0-100% (by comparing with 3D generation models in Sec. 4.2). While effective, the closed-source nature and the high computational cost resulting from combinatorial comparison may create barriers to large-scale evaluation. As an alternative, we employ ImageReward [62], an open-source aesthetic estimator, to measure the aesthetic quality of each rendered image $f^{\\text{ImgReward}}(\\mathcal{I}_i)$ and then compute the average. We compare both implementations in Sec. 4 and discuss their respective strengths.", + "bbox": [ + 511, + 648, + 908, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13330", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/75d8ec71c8150b0f58c278378c317b16ba80c31467aad19ad40a05596deef841.jpg", + "image_caption": [ + "Generated 3D asset", + "Figure 5. 3D inconsistency maps: The proposed 3D metrics, semantic and geometric consistencies, allow fine-grained localization of the artifacts (eg: Janus issue: multiple nose / face, inconsistent hand geometry, arbitrary surface patterns on the back) by fusing / computing the metrics in 3D space." + ], + "image_footnote": [], + "bbox": [ + 140, + 95, + 450, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Localizing 3D Artifacts", + "text_level": 1, + "bbox": [ + 89, + 296, + 305, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike previous works [19, 61], our proposed geometric and semantic consistency measures operate in the 3D space, allowing us to not only classify the generation as an artifact but also to localize the artifacts on the extracted 3D mesh. For the semantic consistency, artifacts can be localized by highlighting outlier vertices—specifically, mesh vertices with DINO feature variance exceeding a certain threshold. Fig. 5 illustrates an example where vertices with high semantic inconsistency (Eq. 2) significantly contribute to the generated 3D inconsistencies like the Janus issue and extraneous geometries. For the geometric consistency, we back-project the pixel-wise cosine distance map between the predicted and rendered geometric normals onto the 3D mesh, highlighting vertices where the mean or maximum cosine distance across different viewpoints exceeds a threshold. Finally, as shown in Fig. 5, fusing the proposed semantic or geometric consistency metrics to the 3D space could enable estimating 3D uncertainty at novel viewpoints. Such consistency cues can be leveraged as a self-supervised refinement metric to enhance the generated geometry.", + "bbox": [ + 89, + 320, + 483, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 633, + 223, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we begin by discussing the curated Eval3D evaluation benchmark. (Sec. 4.1) We then outline the experimental details, covering the targeted 3D generative models, experimental setup, and human annotation process. (Sec. 4.2) Finally, we present our experimental results. (Sec. 4.3) We first assess the agreements between Eval3D and human annotations, showing that Eval3D aligns well with humans on all evaluation dimensions. Then we benchmark text/image-to-3D algorithms with Eval3D and discuss our findings.", + "bbox": [ + 89, + 654, + 485, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Eval3D Benchmark", + "text_level": 1, + "bbox": [ + 89, + 803, + 277, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To thoroughly evaluate 3D generation algorithms, we curate the Eval3D benchmark. Unlike prior work, which provides only text prompts, Eval3D additionally includes a scene graph, dense human annotations, and optionally a carefully selected SDXL image suitable for image-to-3D generation.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/751561f0965350bf3589f736afd0c9b71595eed1013552cf742c3ffb98186dd3.jpg", + "image_caption": [ + "Text prompt = \"A panda rowing a boat in a pond.\"", + "MV-Dream", + "Figure 6. Evaluation examples from Eval3D. While the generation from MV-Dream appears visually appealing at first glance, both the boat and the paddle are missing (reflected in text-3D alignment). There are also geometric artifacts on the left side of the panda (reflected in geometric consistency). Please refer to the Appendix for detailed breakdown of each metric." + ], + "image_footnote": [], + "bbox": [ + 534, + 88, + 702, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Structural consistency: 82.22", + "bbox": [ + 535, + 229, + 694, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-3D alignment: 50", + "bbox": [ + 553, + 243, + 676, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic consistency: 76.8", + "bbox": [ + 540, + 256, + 689, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometric consistency: 82.85", + "bbox": [ + 535, + 267, + 694, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aesthetic score: 84.33", + "bbox": [ + 555, + 279, + 674, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d11054cb8046dd34b6296a167f1c454568ef9f6e1dad289c431a0f1bac2d9357.jpg", + "image_caption": [ + "TextMesh" + ], + "image_footnote": [], + "bbox": [ + 712, + 88, + 880, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Structural consistency: 50.12", + "bbox": [ + 718, + 231, + 875, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-3D alignment: $50\\%$", + "bbox": [ + 733, + 243, + 862, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic consistency: $24.40\\%$", + "bbox": [ + 715, + 255, + 879, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometric consistency: 81.91%", + "bbox": [ + 715, + 267, + 879, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aesthetic score: 26.09", + "bbox": [ + 736, + 279, + 856, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt source: Eval3D contains 160 text prompts for text-to-3D evaluation. The prompts come from diverse sources, including threestudio [18] and the GPT-generated prompts from prior works [19, 51, 61]. We curate the prompt source such that 80 prompts contain a single object, and 80 prompts contain multiple objects. For image-to-3D evaluation, we use the 20 image-text pairs from Dreamcraft3D [51].", + "bbox": [ + 511, + 388, + 906, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene graphs: To facilitate further investigation, we also provide the scene graph for each prompt. Following prior work [10], the scene graph contains 5 kinds of elements: entity (e.g., \"statue\"), attribute (e.g., colors, materials), relation (e.g., \"on top of\"), action (e.g., \"sit\"), and others (e.g., art style). Please refer to Appendix for more details.", + "bbox": [ + 511, + 494, + 906, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Annotations: We provide dense expert annotations on all 3D evaluation criteria discussed in Sec. 3 for 160 prompts across six 3D generation models. We note that our annotation is an order of magnitude larger than prior work. For example, [61] only contains pairwise comparison for 3 prompts for each model pair. [19] annotates quality and alignment for 90 prompts, ignoring geometric, semantic, and structural consistency. See Sec. 4.2 and Appendix for more details.", + "bbox": [ + 511, + 585, + 908, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Details", + "text_level": 1, + "bbox": [ + 511, + 714, + 712, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Target 3D generation models: We evaluate the proposed metrics on six state-of-the-art text-to-3D algorithms (i.e., DreamFusion [41], Magic3D [28], ProlificDreamer [56], TextMesh [52], and MVDream [49]) and two image-to-3D algorithms (i.e., Magic123 [42] and DreamCraft3D [51]). These algorithms represent a diverse set of text/image-to-3D generative models, featuring different 3D representations (NeRF, NeRF + Tetrahedral Mesh, Gaussian Splitting), various optimization strategies (single-stage vs. multi-stage), and different generative (diffusion) priors (single view vs. multi-view). Consequently, the generated assets exhibit a", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13331", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0b33daaf32d735cae59ad75dbc1397c395c79a2bb01c4dfd8b152cf7f1448a00.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ApproachGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
CLIP-Score [20]43.6-64.062.074.1
ImageReward [62]44.3-63.775.479.5
T3-Bench [19]---76.556.7
GPT4V [61]46.9-68.985.672.8
Eval3D (Ours)83.068.069.287.488.7
", + "bbox": [ + 125, + 78, + 867, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Pairwise rating agreements between human and evaluation metrics. We measure the probability that the automatic evaluation makes the same decision as human annotators when comparing two 3D assets. Eval3D achieves the best alignment with humans.", + "bbox": [ + 89, + 179, + 906, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "wide range of artifacts, enabling a thorough study of fine-grained 3D evaluation. We leverage the implementations from ThreeStudio [18]. Please refer to Appendix for details. Setup: We render each generated 3D asset from 120 different viewpoints, with cameras positioned at an elevation of $15^{\\circ}$ and an azimuth varying from $0^{\\circ}$ to $360^{\\circ}$ degrees. We render RGB images, surface normals, and opacity maps at a resolution of $512 \\times 512$ . To compute semantic consistency, we extract the underlying 3D mesh using either marching cubes [31] or DMTet [47]. We upsample DINO features to $256 \\times 256$ usingFeatUp [15] and then fuse them with the extracted 3D mesh. Due to cost considerations, we use 12 viewpoints for the text-3D alignment metric.", + "bbox": [ + 88, + 223, + 485, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human annotations: Previous human annotation data either lacked coverage for all 3D evaluation dimensions (e.g., only focuses on aesthetics and alignment, ignoring geometry and structure) or included a very limited number of annotations. To address these issues, we performed dense expert annotations on all 3D assets generated by six text-to-3D models based on different criteria discussed in Sec. 3. We carefully designed the annotation process for each metric to ensure high inter-annotator agreement. Take geometric consistency for example. For each text prompt, we showed annotators videos of RGB and rendered surface normals, placed side-by-side for all six algorithms. Then we asked annotators to rank the generations based on how well the texture aligns with the surface normals. Ties were allowed. We adopted a similar annotation procedure for aesthetics, but showing only the RGB videos. As for semantic and structural consistency, since it is difficult to rank (e.g., whether a two-faced dog is better or worse than a three-legged astronaut), we instead framed it as a yes/no question and asked annotators whether they observed any inconsistencies. Finally, for text-3D alignment, annotators looked at the RGB videos and answered a series of multi-choice questions generated automatically from the text prompt, as discussed in [10]. We direct the readers to the Appendix for more comprehensive details on our annotation process.", + "bbox": [ + 91, + 421, + 485, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 804, + 294, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human alignment: An optimal 3D evaluation metric should align closely with human judgment. We assess this alignment by comparing human annotations with ours and baseline metrics. As shown in Tab. 1, Eval3D outperforms all baseline evaluation metrics across all dimensions. Specif", + "bbox": [ + 89, + 824, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ically, our text-3D metric aligns approximately $16\\%$ better with human annotations than GPT-4V [61]. Our geometric consistency metric, which is analogous to GPT-4V's texture-geometry alignment [61], achieves $36\\%$ better alignment with human judgments compared to all other baselines. Importantly, our geometric and semantic consistency metrics are also interpretable and fine-grained, allowing us to localize potential artifacts (see Figs. 5 and 3). We note that we are the first to capture the multi-view semantic consistency.", + "bbox": [ + 511, + 223, + 908, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmarking text-to-3D models: Tab. 2 shows the scores (scaled to $0 - 100\\%$ ) of each text-to-3D model under various criteria. MVDream [49] generally performs the best, ranking 1st or 2nd on all metrics, except for geometry consistency. Upon inspecting the MVDream generations, we find that Eval3D successfully captures its weaknesses: despite their high overall quality, MVDream assets suffer from holes and frequently hallucinate details. DreamFusion [41], TextMesh [52], and Magic3D [28] perform poorly across most metrics. They produce low-resolution geometries, which result in low aesthetic scores. Additionally, they do not accurately adhere to the text prompts, achieving only $40\\% - 65\\%$ in text-3D alignment. Nevertheless, their geometric consistency is comparable to, or better than, that of MVDream. We hypothesize that this is due to their tendency to generate simpler and more convex geometries. ProlificDreamer achieves high scores in aesthetics and text-3D alignment, yet it ranks last or second-to-last in all three geometric metrics. In contrast, previous studies [19, 61] rate ProlificDreamer highly. Upon inspecting these outputs, we observe that Eval3D more effectively highlights the weaknesses of ProlificDreamer. Fig 5 showcases a generation by ProlificDreamer that, despite its aesthetic sharpness and alignment with the text prompt, suffers from geometric, structural, and semantic inconsistencies. Gaussian Splatting ranks the lowest on geometric consistency. This validates [22]'s finding that Gaussian-Splatting-based methods generate extremely noisy geometry. Interestingly, despite the noisy geometries, the generations rank highly in terms of semantic consistency measured by DINOv2 [39] and multi-view structural consistency measured using Zero123 [29]. We show a few qualitative examples as well as their evaluation results in Fig. 6. Please refer to the Appendix for those of the other text-to-3D algorithms.", + "bbox": [ + 511, + 364, + 908, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmarking image-to-3D models: Unlike previous stud", + "bbox": [ + 513, + 885, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13332", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d13b298abeb85c767c8a316cebf4c8e3dbc7dc7547a6cd2e5237406621caafe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AlgorithmGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
TextMesh [52]83.174.169.618.943.2
DreamFusion [41]88.367.177.624.250.8
Magic3D [28]94.161.579.351.264.2
ProlificDreamer [56]77.159.575.160.676.9
Gaussian Splitting [68]67.984.881.862.366.7
MVDream [49]85.874.381.182.567.7
", + "bbox": [ + 102, + 75, + 890, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/24c33e6f8609cd7a1c8c0932b97b9ba712d06f37426e48caebc76acda0895106.jpg", + "table_caption": [ + "Table 2. Eval3D results of Text-to-3D models. We benchmark the text-to-3D models discussed in Sec. 4.2 on all the evaluation criteria in Eval3D. For each criterion, the score ranges from 0 to $100\\%$ . Higher values indicate better performance." + ], + "table_footnote": [], + "table_body": "
AlgorithmInputGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
Magic3D [28]97.167.280.325.990.2
MVDream [49]89.776.183.476.880.5
Magic123 [42]✓97.670.782.039.288.4
DreamCraft3D [51]✓97.082.881.369.191.9
", + "bbox": [ + 98, + 229, + 895, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Eval3D results of image-to-3D models. We evaluate Magic123 [42] and DreamCraft3D [51], both of which take images and texts as inputs. We also show the results of two text-to-3D models (Magic3D [28]) and MVDream [49]) using the same text prompts as inputs.", + "bbox": [ + 89, + 316, + 906, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ies [19, 61], we also evaluate state-of-the-art image+text-to-3D generative algorithms (i.e., Magic123[42] and DreamCraft3D[51]). To validate to which degree an input image could help, we compare these algorithms with two text-only 3D generation algorithms, Magic3D [28] and MVDream [49]. As shown in Tab. 3, text+image-to-3D generations better align with the text prompts and perform comparable on geometric and structural consistency metrics. In particular, although Magic123 and Magic3D share similar generation pipelines, Magic123, with the aid of image guidance, achieves improved semantic and structural consistency. We conjecture that the improvement stems from the use of diffusion models to generate novel views from the input image, which serve as priors and help avoid structural inconsistencies like the Janus problem.", + "bbox": [ + 88, + 353, + 485, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 89, + 590, + 316, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Open-source vs. closed-source evaluation: Compared to previous work [61], which exclusively capitalized on closed-source multimodal LLMs for assessment, we propose a fully open-source solution for 3D evaluation. Additionally, we design Eval3D to be flexible, allowing foundation models and tools to be easily swapped with alternatives possessing similar capabilities. For instance, one could replace the open-source LLaVA model [27] with the closed-source GPT-4o for the text-3D alignment metric, or alternatively, select either GPT-4o or ImageReward [62] for aesthetic measurement. We experimented with these models and found that for text-3D alignment, our human alignment is similar using two models (88.7% using LLaVA, 91.5% using GPT-4o), all substantially outperforming prior metrics. While GPT-4o aligns more closely with human judgments than ImageReward does, the overall ranking of 3D algorithms assessed by the two is similar, with an edit distance of 1. These results suggest that open-source models are effective for evaluation.", + "bbox": [ + 89, + 611, + 485, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: The foundation models used in the paper are", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "not perfect and may introduce errors. For example, StableZero123 [29] may synthesize incorrect novel-view scenes; DepthAnything [64] may produce coarse predictions; and multimodal LLMs like LLaVA [27] may generate wrong answers. However, as shown in our experiments, Eval3D still aligns better with humans than all available automatic 3D evaluations, while being significantly more interpretable. Furthermore, our proposed framework is designed to support seamless integration of new and improved foundation models. As foundation models continue to improve, we anticipate even stronger alignment with human evaluations.", + "bbox": [ + 511, + 353, + 908, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader impact: Eval3D can be used to identify and filter out inaccurate and unrealistic 3D assets, and potentially improve existing 3D generation algorithms. It could have a significant impact on the entertainment and robotics industries where high-quality 3D assets are crucial. However, it could also propagate the biases and limitations of foundation models into evaluation metrics. Additionally, it may affect employment within creative fields, as there could be a reduced need for human workers.", + "bbox": [ + 511, + 523, + 910, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion: We present Eval3D, an interpretable and fine-grained evaluation framework for text/image-to-3D generation algorithms. Eval3D leverages the consistency among various vision and language foundation models to accurately evaluate generated 3D assets along multiple dimensions. It aligns with human judgment better than available alternatives and enables pixel-wise and spatial localization of artifacts in the generated 3D assets. We believe Eval3D can serve as a reliable framework for evaluating, filtering, and potentially improving 3D generation algorithms.", + "bbox": [ + 511, + 667, + 908, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: The paper was partially supported by a gift from Ai2, an Amazon Science sponsored research grant, and DARPA TIAMAT program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA.", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13333", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stability AI. Stable-zero123: Quality 3d object generation from single images, 2024. 4", + "[2] Mark Boss, Raphael Braun, V. Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. Nerd: Neural reflectance decomposition from image collections. ICCV, 2020. 4", + "[3] Emmanuelle Bourigault and Pauline Bourigault. Mvdiff: Scalable and flexible multi-view diffusion for 3d object reconstruction from single-view. arXiv preprint arXiv:2405.03894, 2024. 3", + "[4] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236, 2016. 3", + "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 3", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3", + "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22246–22256, 2023. 3", + "[8] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3", + "[9] Zhiqin Chen, Vladimir G Kim, Matthew Fisher, Noam Aigerman, Hao Zhang, and Siddhartha Chaudhuri. Decor-gan: 3d shape detailization by conditional refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15740-15749, 2021. 3", + "[10] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian Scene Graph: Improving Reliability in Fine-Grained Evaluation for Text-to-Image Generation. In ICLR, 2024. 2, 5, 6, 7", + "[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation, 2024. 3", + "[12] Nam Anh Dinh, Haochen Wang, Greg Shakhnarovich, and Rana Hanocka. Loopdraw: a loop-based autoregressive model for shape synthesis and editing. ArXiv, abs/2212.04981, 2022.3", + "[13] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 4" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Stephanie Fu*, Netanel Tamir*, Shobhita Sundaram*, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv, 2023. 5", + "[15] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feup: A model-agnostic framework for features at any resolution. In The Twelfth International Conference on Learning Representations, 2024. 7", + "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3", + "[17] William Gao, April Wang, Gal Metzer, Raymond A Yeh, and Rana Hanocka. Tetgan: A convolutional neural network for tetrahedral mesh generation. arXiv preprint arXiv:2210.05735, 2022. 3", + "[18] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. three-studio: A unified framework for 3d content generation. 2, 6, 7", + "[19] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv, 2023. 2, 3, 6, 7, 8", + "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 7", + "[21] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 2, 3, 5", + "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 7", + "[23] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 867-876, 2022. 3", + "[24] Brian Karis. Journey to nanite. https://www.highperformancegraphics.org/slides22/Journey_to_Nanite.pdf, 2022. Presented at High Performance Graphics 2022. 1", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 3", + "[26] Marian Kleineberg, Matthias Fey, and Frank Weichert. Adversarial generation of continuous implicit shape representations. arXiv preprint arXiv:2002.00349, 2020. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13334", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, and Li Chunyuan Liu, Ziwei. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 5, 8", + "[28] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3, 6, 7, 8", + "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ICCV, 2023. 7, 8", + "[30] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3", + "[31] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 7", + "[32] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3", + "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3", + "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3", + "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf. Communications of the ACM, 2020. 4", + "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022. 3", + "[38] OpenAI. Hello gpt-4o. 2024. 4, 5", + "[39] Maxime Oquab, Timoth'ee Darcet, Théo Moutakanni, Huy Q. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russ Howes, PoYao (Bernie) Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Huijiao Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 4, 7" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 3", + "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. 2023. 6, 7, 8", + "[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 6, 8", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[45] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021.", + "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3", + "[47] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 7", + "[48] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3", + "[49] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and X. Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv, 2023. 3, 6, 7, 8", + "[50] Jiao Sun, Deqing Fu, Yushi Hu, Su Wang, Royi Rassin, Da-Cheng Juan, Dana Alon, Charles Herrmann, Sjoerd van Steenkiste, Ranjay Krishna, et al. Dreamsync: Aligning text-to-image generation with image understanding feedback. In Synthetic Data for Computer Vision Workshop @ CVPR 2024, 2023. 3", + "[51] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13335", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d generation with bootstrapped diffusion prior. arXiv, 2023. 2, 6, 8", + "[52] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 6, 7, 8", + "[53] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3", + "[54] Luke Villapaz. Gta 5 costs 265 million to develop and market, making it the most expensive video game ever produced: Report. International Business Times, 2013. 1", + "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv, 2021. 4", + "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2024. 6, 8", + "[57] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[58] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3", + "[59] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with geo: On metrics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 2", + "[60] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3", + "[61] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. arXiv, 2024. 2, 3, 5, 6, 7, 8", + "[62] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv, 2023. 2, 4, 5, 7, 8", + "[63] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3", + "[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4, 8" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Yue Yang, Fan-Yun Sun, Luca Weihs, Eli VanderBilt, Alvaro Herrasti, Winson Han, Jiajun Wu, Nick Haber, Ranjay Krishna, Lingjie Liu, et al. Holodeck: Language guided generation of 3d embodied ai environments. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024), pages 20-25. IEEE/CVF, 2024. 1", + "[66] Lior Yariv, Matan Atzmon, and Yaron Lipman. Universal differentiable renderer for implicit neural representations. arXiv, 2020. 4", + "[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[68] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3, 8", + "[69] Hongxiang Zhao, Xili Dai, Jianan Wang, Shengbang Tong, Jingyuan Zhang, Weida Wang, Lei Zhang, and Yi Ma. Ctrl123: Consistent novel view synthesis via closed-loop transcription. arXiv preprint arXiv:2403.10953, 2024. 3" + ], + "bbox": [ + 516, + 90, + 906, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13336", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_model.json b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c04ac3faf321d6bc66d2c4eb21a591e5e5f03603 --- /dev/null +++ b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.854, + 0.152 + ], + "angle": 0, + "content": "Eval3D: Interpretable and Fine-grained Evaluation for 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.18, + 0.88, + 0.217 + ], + "angle": 0, + "content": "Shivam Duggal\\*1 Yushi Hu\\*2 Oscar Michel3 Aniruddha Kembhavi4 William T. Freeman1 Noah A. Smith2,5 Ranjay Krishna2,5 Antonio Torralba1 Ali Farhadi2,5 Wei-Chiu Ma6" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.233, + 0.865, + 0.271 + ], + "angle": 0, + "content": "1Massachusetts Institute of Technology 2University of Washington 3New York University 4Wayve 5Allen Institute for AI 6Cornell University" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.281, + 0.882, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.468, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Figure 1. Challenges of 3D generation: (1) Structural inconsistency: lack of globally-coherent 3D shape; (2) Text-3D misalignment: failure to meet the requirements of the input text-prompt; (3) Semantic inconsistency: content change and incoherent semantics; (4) Geometric inconsistency: misaligned geometry and texture." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.521, + 0.327, + 0.536 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Despite the unprecedented progress in the field of 3D generation, current systems still often fail to produce high-quality 3D assets that are visually appealing and geometrically and semantically consistent across multiple viewpoints. To effectively assess the quality of the generated 3D data, there is a need for a reliable 3D evaluation tool. Unfortunately, existing 3D evaluation metrics often overlook the geometric quality of generated assets or merely rely on black-box multimodal large language models for coarse assessment. In this paper, we introduce Eval3D, a fine-grained, interpretable evaluation tool that can faithfully evaluate the quality of generated 3D assets based on various distinct yet complementary criteria. Our key observation is that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. We thus leverage a diverse set of models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. Compared to prior work, Eval3D provides pixel-wise measurement, enables accurate 3D spatial feedback, and aligns more closely with human judgments. We comprehensively evaluate existing 3D generation models using Eval3D and highlight the limitations and challenges" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.882, + 0.538 + ], + "angle": 0, + "content": "of current models. Project page: http://eval3d.github.io." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.548, + 0.645, + 0.562 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.566, + 0.909, + 0.718 + ], + "angle": 0, + "content": "High-quality 3D assets play a pivotal role across various domains including gaming, film, augmented/virtual reality, and robotics [24, 54, 65]. They empower content creation applications to deliver visually compelling and immersive experiences and enable simulators to realistically reproduce the real world. Unfortunately, 3D asset creation is complex and tedious, requiring a lot of careful manual effort. Automatic (text/image-to-)3D generation, despite fast progress over the last few years, still falls short in key aspects including geometry, semantic consistency, etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "What makes a good (text/image-to-)3D generation? 3D assets are not created to merely be rendered onto a 2D image; rather, they are meant to live within a rich 3D world – the assets may be viewed from various distinct viewpoints, and their 3D geometry will affect how they interact with the surroundings (e.g., collision, reflection, etc). Therefore, a high-quality 3D asset should not only be visually appealing and faithful to its text/image input across multiple individual viewpoints, but its geometry and appearance should also be coherent and plausible when observed as a whole. In the absence of these characteristics, the utility of the generated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13326" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.059, + 0.845, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.35, + 0.908, + 0.38 + ], + "angle": 0, + "content": "Figure 2. Eval3D offers interpretable, fine-grained, and human-aligned metrics to assess the quality of 3D generations from various aspects. We utilize a diverse array of foundation models and tools to measure the consistency among different representations of generated 3D assets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.486, + 0.569 + ], + "angle": 0, + "content": "assets degrades drastically. Consider the generated 3D assets in Fig. 1. While the corgi (Fig. 1 (left)) looks reasonable from the front and the back, its overall 3D shape is quite poor and implausible – notice the Janus issue (i.e., presenting multiple faces). Similarly, while the fried chickens and waffles (Fig. 1(right)) appear faithful from one viewpoint, the 3D asset suffers from semantic and geometric inconsistency across views – as the object rotates, the maple syrup transform into a jar, and the fried chicken pieces merge with the waffles. These 3D-specific structural and consistency challenges continue to pose significant difficulties to existing (text/image-to-)3D generation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.486, + 0.901 + ], + "angle": 0, + "content": "What makes a good 3D evaluation metric? To advance (text/image-to-)3D asset generation and to effectively quantify the aforementioned issues, there is a need for a reliable 3D evaluation tool. Unfortunately, while significant progress has been made in evaluating 2D text-to-image models [10, 21, 59, 62, 67], the development of their 3D counterparts is lagging behind. Existing 3D evaluation metrics typically rely on black-box multimodal large language models (LLMs), such as GPT-4V, for assessment [61], or often overlook the geometric quality of generated assets [19]. A comprehensive 3D metric should be able to: (a) rate the consistency between the input text/image and the 3D generation and detect misalignments such as a missing object or part, wrong attribute, etc., (b) measure the structural and semantic consistency of the generated asset across the input and the output as well as across multiple views of the 3D asset, and (c) localize flaws and inconsistencies in the produced asset. Such a comprehensive metric would not only permit fine-grained comparisons between different 3D generations but also provide feedback on where and what to improve." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Our method: We introduce Eval3D, an interpretable, and fine-grained evaluation tool that measures the quality of generated 3D assets from multiple dimensions. We build Eval3D based on the observation that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. For instance, to measure the geometric consistency of the generated chicken and waffles (i.e., the red region in Fig. 1(right)), one can derive the surface normal analytically from the 3D object, and then compare it with the normal estimated by a foundation model (e.g., Depth Anything [64]) from rendered RGB images. Based on 2D visual priors, waffles usually possess a grid-like 3D geometry. However, the grid-like patterns do not exist in the normal map derived from 3D. By computing the differences, we can automatically identify those regions with inconsistent surface normals and localize where the generative model leverages texture maps to \"cheat.\" We adopt a diverse set of open-source foundation models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. To aggregate the measurements across multiple viewpoints, we further present several simple yet effective 2D and 3D-based fusion mechanisms. With these fine-grained, pixel-wise measurements, we can not only easily locate the problematic regions in 2D images (Fig. 3), but also lift the inconsistency map back to 3D and enable accurate spatial artifact localization (Fig. 5). To our knowledge, we are the first to provide such spatial feedback." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To promote the use of our new evaluation tool, we curate a diverse set of prompts and images from multiple sources [18, 19, 51]. We evaluate a variety of (text/image-to-)3D generation models on our data and densely annotate human preferences based on different criteria to verify how well our metrics align with human assessment. We show that Eval3D" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13327" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.074, + 0.844, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.23, + 0.908, + 0.259 + ], + "angle": 0, + "content": "Figure 3. Geometry inconsistency evaluates texture-geometry misalignment by comparing 3D rendered normal and image-based normal. Bright-yellow indicates large discrepancy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.267, + 0.485, + 0.435 + ], + "angle": 0, + "content": "is not only able to automatically detect various generation artifacts (e.g., the Janus issue), but also accurately localize them (Fig. 3). Surprisingly, we also find that while existing state-of-the-art generations are visually appealing, they often still suffer from geometric or semantic inconsistencies. Finally, we demonstrate that although the accuracy of Eval3D is dependent on the accuracy of the foundation models, it consistently aligns more closely with, or is comparable to, human judgments across all dimensions than prior work. As foundation models will improve over time, we hypothesize that Eval3D will become increasingly reliable." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.233, + 0.479 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.494, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Generative models for 3D asset generation: Early work on 3D asset generation trained probabilistic generative models directly on 3D data, most commonly using assets from ShapeNet [6]. While numerous generative modeling frameworks (e.g., GANs [8, 9, 16, 17, 26, 60], VAEs [4, 8, 12, 17, 40], flow models [5, 63]) and 3D representations (e.g., point clouds [5, 37, 63], meshes [16, 16], voxels [4, 9, 60], and implicit functions [8, 26, 32, 40]) have been explored, the quality and diversity of 3D generations are constrained by the quantity and characteristics of the available data. With this in mind, recent dominant approaches turn to leverage supervision from foundation models trained on massive collections of 2D images [44-46]. During generation, they propagate the gradient from multi-view 2D space back to 3D and iteratively update the underlying 3D representations [8, 23, 25, 34, 36, 43]. Since foundation models encode rich priors of our world, this significantly improves the quality of the generated objects. To further improve multi-view consistency [3, 30, 48, 58, 69], object geometry [7, 28, 53], or even the optimization process [33, 57, 68], researchers have proposed a variety of enhancements, such as multi-view diffusion [49]. While a general trend of progress is evident, the lack of evaluation tools makes it unclear which methods are superior and in which dimensions. Our framework addresses this confusion by providing a general metric to measure progress, which can be decomposed to identify performance across critical factors relevant to 3D generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.909, + 0.6 + ], + "angle": 0, + "content": "**Evaluating 3D generation:** Perhaps closest to our work are GPT-4V [61] and T3-Bench [19]. GPT-4V [61] is an ELOB-based method that uses the eponymous multimodal LLM to compare the quality of two generated objects, while T3-Bench [19] employs a vision-language model (VLM) to evaluate the semantic adherence of generated objects to the prompt. While we share similar insights, there are several key differences: first, unlike GPT-4V, which provides a pairwise preference, Eval3D offers an absolute metric, making it easier to quantify and compare performance across different models and tasks. Additionally, both GPT-4V and T3 rely mainly on language outputs for evaluation. The reliance on the inherently imprecise nature of language makes localizing the source of error difficult. This issue is particularly problematic for detecting multi-view inconsistencies, as each view may independently seem satisfactory to the VLM. In contrast, our method incorporates semantic judgments from VLMs [11, 21, 50] and uses pretrained computer vision models [39, 64] to provide precise and localized estimates of geometric and structural deviations at multiple scales of granularity. These error measurements can be back-projected to 3D, making Eval3D more informative." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.609, + 0.602, + 0.624 + ], + "angle": 0, + "content": "3. Eval3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Our aim is to develop an evaluation tool that can faithfully assess the quality of (generated) 3D assets based on various distinct yet complementary criteria. Unlike 2D images, which are merely projections of our visual world, these generated assets live in 3D. Their geometry influences how they interact with the environment, and they can be viewed from multiple angles. Therefore, it is crucial to ensure that the appearance of these 3D generations is coherent and visually appealing from various viewpoints, that their geometry is high-quality and consistent with their RGB counterparts, and that the generated content remains faithful to the input conditions (e.g., text, image)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Towards this goal, we devise five distinct measurements to evaluate the quality of generated 3D assets from different aspects. Our key observation is that many of the desired 3D properties can be predicted from one another if the (generated) 3D asset is of high-quality; if the cues are unpredictable, then the 3D generation often has artifacts or is not" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13328" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.196 + ], + "angle": 0, + "content": "coherent. Building upon this intuition, we consolidate a diverse set of foundation models [1, 38, 39, 62, 64] and tools to compute and predict a wide range of visual cues and measure their consistency. The measurements across multiple viewpoints are then aggregated to reflect the quality of the 3D generation with respect to different aspects. For each critic, we scale the score to lie within \\(0 - 100\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.484, + 0.304 + ], + "angle": 0, + "content": "Notation: Let \\( t \\) be the input text prompt and \\( \\mathcal{I}^{\\mathrm{GT}} \\) be the optional image condition. Let \\( \\mathcal{O}_t \\) be the text-conditioned 3D asset, where the underlying representation could be either a mesh \\( \\mathcal{O}_t^{\\mathrm{mesh}} \\) or an implicit representation \\( \\mathcal{O}_t^{\\mathrm{imp}} \\). Let \\( \\{\\mathcal{I}_i\\}_{i=1}^N \\) denote the rendered RGB images of the generated asset \\( \\mathcal{O}_t \\) at viewpoint \\( v_i \\). Let \\( f \\) denote the foundation models or tools that we adopt to extract visual cues." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.314, + 0.305, + 0.33 + ], + "angle": 0, + "content": "3.1. Geometric Consistency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.412 + ], + "angle": 0, + "content": "To evaluate how good the geometry of an asset is, one naive way is to consume the 3D asset and output a score. Unfortunately, it is unclear what constitutes a good geometry and there is no such data or model that allows us to directly predict the geometric quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.681 + ], + "angle": 0, + "content": "In this work, we take a proxy and measure the consistency between the surface normal analytically derived from the 3D asset and the normal predicted by a foundation model from 2D images. The intuition is that a good 3D geometry should be reflective of its visual appearance. If the derived geometry does not align with the normal predicted from RGB (e.g., the chicken waffle in Fig. 1, right), then one of the components is incorrect. Specifically, we first compute the analytical surface normal of the 3D asset by taking the normal of each face in the mesh, or by taking the negative gradient of the density [2, 35] or signed distance field [55, 66], depending on the 3D representation. Then for each viewpoint, we exploit Depth Anything [64] to compute the monocular depth and convert it into normal. Empirically we find it more robust than direct surface normal estimation [13]. Finally, we project the analytical normal to each view, measure the pixel-wise angular difference, and compute the amount of inliers:" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.696, + 0.284, + 0.71 + ], + "angle": 0, + "content": "Geometric consistency \\( = \\)" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.715, + 0.484, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N _ {p}} \\sum_ {p} \\mathbb {1} \\left[ \\operatorname {a r c c o s} \\left(\\mathbf {n} _ {p} ^ {\\text {a n a l}} \\cdot \\mathbf {n} _ {p} ^ {\\text {p r e d}}\\right) < \\delta^ {\\text {n o r m}} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Here, we iterate over all pixels \\( p \\). \\( \\mathbf{n}^{\\mathrm{anal}} \\) indicates that the normal of pixel \\( p \\) is analytically derived, while \\( \\mathbf{n}^{\\mathrm{pred}} \\) suggests it is predicted. \\( N_{p} \\) denotes the number of valid rendered pixels. In practice, we set the threshold \\( \\delta^{\\mathrm{norm}} \\) to \\( 23^{\\circ} \\) based on a hold-out validation set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.295, + 0.866 + ], + "angle": 0, + "content": "3.2. Semantic Consistency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Another crucial factor for a 3D generation to be appealing and useful is that the underlying content and semantics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.196 + ], + "angle": 0, + "content": "remain consistent and coherent across different viewpoints. Unlike visual appearance, which may be affected by extrinsic factors such as lighting and change across views, the semantics of an asset should be invariant to different perspectives. To quantify the degree to which the semantics change across views, we capitalize on large vision foundation models that encode rich semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.909, + 0.439 + ], + "angle": 0, + "content": "We first extract DINO features for every rendered image \\(\\mathcal{F}_i^{\\mathrm{DINO}} = f^{\\mathrm{DINO}}(\\mathcal{I}_i)\\). Then, for each 3D point \\(x\\) on the surface of the asset \\(\\mathcal{O}\\), we project it onto 2D feature maps \\(\\{\\mathcal{F}_i^{\\mathrm{DINO}}\\}\\) to retrieve its corresponding features \\(\\{\\mathcal{F}_i^{\\mathrm{DINO}}(\\pi_{v_i}(x))\\}\\). Here, \\(\\pi\\) is the projection operator. Since a point may be visible from multiple images, we compute the variance of its DINO features. Specifically, we calculate the variance along each feature dimension separately across viewpoints and then take the mean of these variances across all feature dimensions. Finally, we compute the ratio of vertices whose variance is below a certain threshold. The intuition is that if the content and semantics of a 3D asset remain relatively stable across different viewpoints, its DINO features should be consistent. In practice, for computational efficiency, we only consider the vertices of the 3D mesh \\(x^{\\mathrm{vert}}\\) as 3D points:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.449, + 0.907, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\text {S e m a n t i c c o n s i s t e n c y} = \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.467, + 0.903, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N _ {\\text {v e r t}}} \\sum_ {x ^ {\\text {v e r t}}} \\mathbb {1} \\left[ \\operatorname {m e a n} \\left(\\operatorname {V a r} \\left(\\left\\{\\mathcal {F} _ {i} ^ {\\mathrm {D I N O}} \\left(\\pi_ {v _ {i}} \\left(x ^ {\\text {v e r t}}\\right) \\right.\\right)\\right) < \\delta^ {\\mathrm {D I N O}} \\right] \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.906, + 0.542 + ], + "angle": 0, + "content": "We set \\(\\delta^{\\mathrm{DINO}}\\) to the 70th percentile of average DINO variance based on a hold-out validation set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.727, + 0.563 + ], + "angle": 0, + "content": "3.3. Structural Consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.644 + ], + "angle": 0, + "content": "A generated asset that looks plausible from a single view does not necessarily imply that it remains coherent when observed as a whole. The third criterion we are interested in measuring is whether the overall structure of the generated asset is coherent and plausible." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.795 + ], + "angle": 0, + "content": "We frame this assessment as a novel view synthesis (NVS) problem. Our intuition is that if the generated asset is overall plausible, one should be able to synthesize, to some degree, what other views might look like. For instance, given an image of the back of the head, one should be capable of predicting the face, although there might be variations in the size of the eyes or nose compared to the actual generation. On the other hand, if the generation suffers from artifacts or noise, it would be difficult to effectively conduct cross-view prediction, and the discrepancies could be significant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "With this motivation in mind, given a rendered RGB image \\(\\mathcal{I}_i\\) as input, we first utilize Stable-Zero123 [1], a diffusion-based NVS model, to predict what an image from another unobserved viewpoint might look like, denoted as \\(\\mathcal{I}_{i\\rightarrow j} = f^{0123}(\\mathcal{I}_i;v_i,v_j)\\). We then measure the difference between the predicted image \\(\\mathcal{I}_{i\\rightarrow j}\\) and the image rendered from the generated asset \\(\\mathcal{I}_j\\). Since we care more about the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13329" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.079, + 0.849, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.275, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Structural consistency evaluates the geometric coherence of the generated 3D assets by comparing rendered views with the predictions from a novel view synthesis model (Zero-123) across various rotations. We utilize DreamSim to assess image similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.486, + 0.404 + ], + "angle": 0, + "content": "perceptual similarity rather than pixel-wise differences, we employ DreamSim [14], a human-aligned perceptual metric, for this similarity assessment (see Fig. 4 for overview). Finally, we repeat the same operation for multiple views \\( j \\) and then take the average. In practice, to increase robustness, we consider different images \\( \\mathcal{I}_i \\) as input and select the best one:" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.41, + 0.283, + 0.424 + ], + "angle": 0, + "content": "Structural consistency \\( = \\)" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.429, + 0.484, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\max _ {i} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\left(1 - f ^ {\\text {D r e a m S i m}} \\left(\\mathcal {I} _ {i \\rightarrow j} ^ {\\text {p r e d}}, \\mathcal {I} _ {j}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.486, + 0.522 + ], + "angle": 0, + "content": "We take as input both the canonical view (rendered at \\(0^{\\circ}\\)), and the rendered image from \\(90^{\\circ}\\). We predict novel viewpoints at an interval of \\(90^{\\circ}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.275, + 0.545 + ], + "angle": 0, + "content": "3.4. Text-3D Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Another often overlooked aspect of 3D evaluation is the fidelity of the generated 3D asset to the input conditions. For instance, are all the mentioned objects being generated? Do the relationships among the objects align with what the text describes? The alignment issue is critical as it reflects how accurately the 3D generation adheres to the instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To effectively measure the text-3D alignment, we draw inspiration from TIFA [21], an evaluation metric for text-to-image generation, and adapt it for assessing text-to-3D generation. The key idea is to measure the consistency among large language models (LLMs) and multimodal LLMs across multiple views, and use this consistency as a proxy for evaluation. Specifically, we first employ a LLM to generate \\(M\\) multiple-choice question-answer tuples \\(\\{Q_j,C_j,A_j\\}_{j = 1}^M\\) based on the input text prompt \\(t\\), where \\(Q_{j}\\) is a question, \\(C_j\\) is a set of answer choices, and \\(A_{j}\\in C_{j}\\) is the gold answer. Then, for each question \\(Q_{j}\\), we employ a multimodal LLM to produce an answer based on each input image \\(A_{j}^{v_{i},\\mathrm{VQA}} = f^{\\mathrm{MLLM}}(\\mathcal{I}_{i},Q_{j})\\). If the answer from the multimodal LLM \\(A_{j}^{v_{i},\\mathrm{VQA}}\\) aligns with the answer derived from the LLM \\(A_{j}\\), we conclude that the specific property in question is aligned at such viewpoint \\(v_{i}\\). For example, given a text prompt \\(t = \"\\text{A statue of a dog}\"\\)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.907, + 0.421 + ], + "angle": 0, + "content": "the LLM could generate question-answer tuples such as {\"Is there a statue?\", \"Yes\", \"No\", \"Yes\"} and view-dependent alignment is assessed based on whether both the text and the image confirm the presence of a statue (see Fig. 2). To aggregate individual scores, we repeat the same procedure for all questions \\(\\{Q_j\\}_{j=1}^M\\) and across all viewpoints \\(\\{v_i\\}_{i=1}^N\\):" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.434, + 0.672, + 0.45 + ], + "angle": 0, + "content": "Text-3D alignment =" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.454, + 0.907, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{M} \\sum_ {j = 1} ^ {M} \\operatorname {A n y} _ {v _ {i}} \\left(\\operatorname {A l l} _ {v \\in \\operatorname {A d j} \\left(v _ {i}\\right)} \\left(\\mathbb {1} \\left[ A _ {j} = = A _ {j} ^ {v, \\mathrm {V Q A}} \\right]\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.909, + 0.617 + ], + "angle": 0, + "content": "A question is considered globally aligned if and only if, for some viewpoint \\( v_{i} \\), the answers of all its neighboring views are aligned. We find that this method is effective for dealing with occlusions and the noise of multimodal LLM. In practice, we adopt LLaVA-NeXT-7B [27] as our VQA model, and we follow DSG [10] for generating and evaluating our question-answer pairs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.625, + 0.623, + 0.64 + ], + "angle": 0, + "content": "3.5. Aesthetic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.909, + 0.906 + ], + "angle": 0, + "content": "Finally, given that a generated 3D asset can be viewed from a wide range of angles, it is crucial to ensure that it appears visually appealing across all possible views. In this paper, we exploit two different approaches, with varying computational cost, to compute the aesthetic score \\( f^{\\text{aesthetic}}(\\{\\mathcal{I}_i\\}) \\). First, following [61], we use GPT-4o [38] to conduct pairwise quality comparisons of 3D assets and compute the ELO score of each model. We improve the prompts and normalize the ELO score to 0-100% (by comparing with 3D generation models in Sec. 4.2). While effective, the closed-source nature and the high computational cost resulting from combinatorial comparison may create barriers to large-scale evaluation. As an alternative, we employ ImageReward [62], an open-source aesthetic estimator, to measure the aesthetic quality of each rendered image \\( f^{\\text{ImgReward}}(\\mathcal{I}_i) \\) and then compute the average. We compare both implementations in Sec. 4 and discuss their respective strengths." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13330" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.082, + 0.307, + 0.094 + ], + "angle": 0, + "content": "Generated 3D asset" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.097, + 0.452, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.213, + 0.486, + 0.283 + ], + "angle": 0, + "content": "Figure 5. 3D inconsistency maps: The proposed 3D metrics, semantic and geometric consistencies, allow fine-grained localization of the artifacts (eg: Janus issue: multiple nose / face, inconsistent hand geometry, arbitrary surface patterns on the back) by fusing / computing the metrics in 3D space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.297, + 0.307, + 0.314 + ], + "angle": 0, + "content": "3.6. Localizing 3D Artifacts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.485, + 0.624 + ], + "angle": 0, + "content": "Unlike previous works [19, 61], our proposed geometric and semantic consistency measures operate in the 3D space, allowing us to not only classify the generation as an artifact but also to localize the artifacts on the extracted 3D mesh. For the semantic consistency, artifacts can be localized by highlighting outlier vertices—specifically, mesh vertices with DINO feature variance exceeding a certain threshold. Fig. 5 illustrates an example where vertices with high semantic inconsistency (Eq. 2) significantly contribute to the generated 3D inconsistencies like the Janus issue and extraneous geometries. For the geometric consistency, we back-project the pixel-wise cosine distance map between the predicted and rendered geometric normals onto the 3D mesh, highlighting vertices where the mean or maximum cosine distance across different viewpoints exceeds a threshold. Finally, as shown in Fig. 5, fusing the proposed semantic or geometric consistency metrics to the 3D space could enable estimating 3D uncertainty at novel viewpoints. Such consistency cues can be leveraged as a self-supervised refinement metric to enhance the generated geometry." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.224, + 0.652 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.486, + 0.791 + ], + "angle": 0, + "content": "In this section, we begin by discussing the curated Eval3D evaluation benchmark. (Sec. 4.1) We then outline the experimental details, covering the targeted 3D generative models, experimental setup, and human annotation process. (Sec. 4.2) Finally, we present our experimental results. (Sec. 4.3) We first assess the agreements between Eval3D and human annotations, showing that Eval3D aligns well with humans on all evaluation dimensions. Then we benchmark text/image-to-3D algorithms with Eval3D and discuss our findings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.279, + 0.818 + ], + "angle": 0, + "content": "4.1. Eval3D Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "To thoroughly evaluate 3D generation algorithms, we curate the Eval3D benchmark. Unlike prior work, which provides only text prompts, Eval3D additionally includes a scene graph, dense human annotations, and optionally a carefully selected SDXL image suitable for image-to-3D generation." + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.072, + 0.872, + 0.087 + ], + "angle": 0, + "content": "Text prompt = \"A panda rowing a boat in a pond.\"" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.089, + 0.703, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.22, + 0.653, + 0.23 + ], + "angle": 0, + "content": "MV-Dream" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.231, + 0.696, + 0.243 + ], + "angle": 0, + "content": "Structural consistency: 82.22" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.244, + 0.678, + 0.256 + ], + "angle": 0, + "content": "Text-3D alignment: 50" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.257, + 0.69, + 0.267 + ], + "angle": 0, + "content": "Semantic consistency: 76.8" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.268, + 0.695, + 0.279 + ], + "angle": 0, + "content": "Geometric consistency: 82.85" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.28, + 0.676, + 0.289 + ], + "angle": 0, + "content": "Aesthetic score: 84.33" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.089, + 0.882, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.22, + 0.828, + 0.23 + ], + "angle": 0, + "content": "TextMesh" + }, + { + "type": "text", + "bbox": [ + 0.719, + 0.232, + 0.877, + 0.243 + ], + "angle": 0, + "content": "Structural consistency: 50.12" + }, + { + "type": "text", + "bbox": [ + 0.734, + 0.244, + 0.863, + 0.255 + ], + "angle": 0, + "content": "Text-3D alignment: \\(50\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.256, + 0.88, + 0.267 + ], + "angle": 0, + "content": "Semantic consistency: \\(24.40\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.268, + 0.88, + 0.279 + ], + "angle": 0, + "content": "Geometric consistency: 81.91%" + }, + { + "type": "text", + "bbox": [ + 0.738, + 0.28, + 0.857, + 0.289 + ], + "angle": 0, + "content": "Aesthetic score: 26.09" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.294, + 0.907, + 0.377 + ], + "angle": 0, + "content": "Figure 6. Evaluation examples from Eval3D. While the generation from MV-Dream appears visually appealing at first glance, both the boat and the paddle are missing (reflected in text-3D alignment). There are also geometric artifacts on the left side of the panda (reflected in geometric consistency). Please refer to the Appendix for detailed breakdown of each metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Prompt source: Eval3D contains 160 text prompts for text-to-3D evaluation. The prompts come from diverse sources, including threestudio [18] and the GPT-generated prompts from prior works [19, 51, 61]. We curate the prompt source such that 80 prompts contain a single object, and 80 prompts contain multiple objects. For image-to-3D evaluation, we use the 20 image-text pairs from Dreamcraft3D [51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.586 + ], + "angle": 0, + "content": "Scene graphs: To facilitate further investigation, we also provide the scene graph for each prompt. Following prior work [10], the scene graph contains 5 kinds of elements: entity (e.g., \"statue\"), attribute (e.g., colors, materials), relation (e.g., \"on top of\"), action (e.g., \"sit\"), and others (e.g., art style). Please refer to Appendix for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.909, + 0.707 + ], + "angle": 0, + "content": "Annotations: We provide dense expert annotations on all 3D evaluation criteria discussed in Sec. 3 for 160 prompts across six 3D generation models. We note that our annotation is an order of magnitude larger than prior work. For example, [61] only contains pairwise comparison for 3 prompts for each model pair. [19] annotates quality and alignment for 90 prompts, ignoring geometric, semantic, and structural consistency. See Sec. 4.2 and Appendix for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.715, + 0.714, + 0.731 + ], + "angle": 0, + "content": "4.2. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Target 3D generation models: We evaluate the proposed metrics on six state-of-the-art text-to-3D algorithms (i.e., DreamFusion [41], Magic3D [28], ProlificDreamer [56], TextMesh [52], and MVDream [49]) and two image-to-3D algorithms (i.e., Magic123 [42] and DreamCraft3D [51]). These algorithms represent a diverse set of text/image-to-3D generative models, featuring different 3D representations (NeRF, NeRF + Tetrahedral Mesh, Gaussian Splitting), various optimization strategies (single-stage vs. multi-stage), and different generative (diffusion) priors (single view vs. multi-view). Consequently, the generated assets exhibit a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13331" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.079, + 0.868, + 0.177 + ], + "angle": 0, + "content": "
ApproachGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
CLIP-Score [20]43.6-64.062.074.1
ImageReward [62]44.3-63.775.479.5
T3-Bench [19]---76.556.7
GPT4V [61]46.9-68.985.672.8
Eval3D (Ours)83.068.069.287.488.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.18, + 0.907, + 0.21 + ], + "angle": 0, + "content": "Table 1. Pairwise rating agreements between human and evaluation metrics. We measure the probability that the automatic evaluation makes the same decision as human annotators when comparing two 3D assets. Eval3D achieves the best alignment with humans." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.224, + 0.486, + 0.42 + ], + "angle": 0, + "content": "wide range of artifacts, enabling a thorough study of fine-grained 3D evaluation. We leverage the implementations from ThreeStudio [18]. Please refer to Appendix for details. Setup: We render each generated 3D asset from 120 different viewpoints, with cameras positioned at an elevation of \\(15^{\\circ}\\) and an azimuth varying from \\(0^{\\circ}\\) to \\(360^{\\circ}\\) degrees. We render RGB images, surface normals, and opacity maps at a resolution of \\(512 \\times 512\\). To compute semantic consistency, we extract the underlying 3D mesh using either marching cubes [31] or DMTet [47]. We upsample DINO features to \\(256 \\times 256\\) usingFeatUp [15] and then fuse them with the extracted 3D mesh. Due to cost considerations, we use 12 viewpoints for the text-3D alignment metric." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.422, + 0.486, + 0.799 + ], + "angle": 0, + "content": "Human annotations: Previous human annotation data either lacked coverage for all 3D evaluation dimensions (e.g., only focuses on aesthetics and alignment, ignoring geometry and structure) or included a very limited number of annotations. To address these issues, we performed dense expert annotations on all 3D assets generated by six text-to-3D models based on different criteria discussed in Sec. 3. We carefully designed the annotation process for each metric to ensure high inter-annotator agreement. Take geometric consistency for example. For each text prompt, we showed annotators videos of RGB and rendered surface normals, placed side-by-side for all six algorithms. Then we asked annotators to rank the generations based on how well the texture aligns with the surface normals. Ties were allowed. We adopted a similar annotation procedure for aesthetics, but showing only the RGB videos. As for semantic and structural consistency, since it is difficult to rank (e.g., whether a two-faced dog is better or worse than a three-legged astronaut), we instead framed it as a yes/no question and asked annotators whether they observed any inconsistencies. Finally, for text-3D alignment, annotators looked at the RGB videos and answered a series of multi-choice questions generated automatically from the text prompt, as discussed in [10]. We direct the readers to the Appendix for more comprehensive details on our annotation process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.805, + 0.295, + 0.822 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Human alignment: An optimal 3D evaluation metric should align closely with human judgment. We assess this alignment by comparing human annotations with ours and baseline metrics. As shown in Tab. 1, Eval3D outperforms all baseline evaluation metrics across all dimensions. Specif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.909, + 0.362 + ], + "angle": 0, + "content": "ically, our text-3D metric aligns approximately \\(16\\%\\) better with human annotations than GPT-4V [61]. Our geometric consistency metric, which is analogous to GPT-4V's texture-geometry alignment [61], achieves \\(36\\%\\) better alignment with human judgments compared to all other baselines. Importantly, our geometric and semantic consistency metrics are also interpretable and fine-grained, allowing us to localize potential artifacts (see Figs. 5 and 3). We note that we are the first to capture the multi-view semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.909, + 0.882 + ], + "angle": 0, + "content": "Benchmarking text-to-3D models: Tab. 2 shows the scores (scaled to \\(0 - 100\\%\\)) of each text-to-3D model under various criteria. MVDream [49] generally performs the best, ranking 1st or 2nd on all metrics, except for geometry consistency. Upon inspecting the MVDream generations, we find that Eval3D successfully captures its weaknesses: despite their high overall quality, MVDream assets suffer from holes and frequently hallucinate details. DreamFusion [41], TextMesh [52], and Magic3D [28] perform poorly across most metrics. They produce low-resolution geometries, which result in low aesthetic scores. Additionally, they do not accurately adhere to the text prompts, achieving only \\(40\\% - 65\\%\\) in text-3D alignment. Nevertheless, their geometric consistency is comparable to, or better than, that of MVDream. We hypothesize that this is due to their tendency to generate simpler and more convex geometries. ProlificDreamer achieves high scores in aesthetics and text-3D alignment, yet it ranks last or second-to-last in all three geometric metrics. In contrast, previous studies [19, 61] rate ProlificDreamer highly. Upon inspecting these outputs, we observe that Eval3D more effectively highlights the weaknesses of ProlificDreamer. Fig 5 showcases a generation by ProlificDreamer that, despite its aesthetic sharpness and alignment with the text prompt, suffers from geometric, structural, and semantic inconsistencies. Gaussian Splatting ranks the lowest on geometric consistency. This validates [22]'s finding that Gaussian-Splatting-based methods generate extremely noisy geometry. Interestingly, despite the noisy geometries, the generations rank highly in terms of semantic consistency measured by DINOv2 [39] and multi-view structural consistency measured using Zero123 [29]. We show a few qualitative examples as well as their evaluation results in Fig. 6. Please refer to the Appendix for those of the other text-to-3D algorithms." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.91, + 0.902 + ], + "angle": 0, + "content": "Benchmarking image-to-3D models: Unlike previous stud" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13332" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.076, + 0.891, + 0.188 + ], + "angle": 0, + "content": "
AlgorithmGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
TextMesh [52]83.174.169.618.943.2
DreamFusion [41]88.367.177.624.250.8
Magic3D [28]94.161.579.351.264.2
ProlificDreamer [56]77.159.575.160.676.9
Gaussian Splitting [68]67.984.881.862.366.7
MVDream [49]85.874.381.182.567.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.194, + 0.907, + 0.222 + ], + "angle": 0, + "content": "Table 2. Eval3D results of Text-to-3D models. We benchmark the text-to-3D models discussed in Sec. 4.2 on all the evaluation criteria in Eval3D. For each criterion, the score ranges from 0 to \\(100\\%\\). Higher values indicate better performance." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.23, + 0.897, + 0.314 + ], + "angle": 0, + "content": "
AlgorithmInputGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
Magic3D [28]97.167.280.325.990.2
MVDream [49]89.776.183.476.880.5
Magic123 [42]✓97.670.782.039.288.4
DreamCraft3D [51]✓97.082.881.369.191.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.317, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Table 3. Eval3D results of image-to-3D models. We evaluate Magic123 [42] and DreamCraft3D [51], both of which take images and texts as inputs. We also show the results of two text-to-3D models (Magic3D [28]) and MVDream [49]) using the same text prompts as inputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.486, + 0.581 + ], + "angle": 0, + "content": "ies [19, 61], we also evaluate state-of-the-art image+text-to-3D generative algorithms (i.e., Magic123[42] and DreamCraft3D[51]). To validate to which degree an input image could help, we compare these algorithms with two text-only 3D generation algorithms, Magic3D [28] and MVDream [49]. As shown in Tab. 3, text+image-to-3D generations better align with the text prompts and perform comparable on geometric and structural consistency metrics. In particular, although Magic123 and Magic3D share similar generation pipelines, Magic123, with the aid of image guidance, achieves improved semantic and structural consistency. We conjecture that the improvement stems from the use of diffusion models to generate novel views from the input image, which serve as priors and help avoid structural inconsistencies like the Janus problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.317, + 0.608 + ], + "angle": 0, + "content": "5. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Open-source vs. closed-source evaluation: Compared to previous work [61], which exclusively capitalized on closed-source multimodal LLMs for assessment, we propose a fully open-source solution for 3D evaluation. Additionally, we design Eval3D to be flexible, allowing foundation models and tools to be easily swapped with alternatives possessing similar capabilities. For instance, one could replace the open-source LLaVA model [27] with the closed-source GPT-4o for the text-3D alignment metric, or alternatively, select either GPT-4o or ImageReward [62] for aesthetic measurement. We experimented with these models and found that for text-3D alignment, our human alignment is similar using two models (88.7% using LLaVA, 91.5% using GPT-4o), all substantially outperforming prior metrics. While GPT-4o aligns more closely with human judgments than ImageReward does, the overall ranking of 3D algorithms assessed by the two is similar, with an edit distance of 1. These results suggest that open-source models are effective for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Limitations: The foundation models used in the paper are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.909, + 0.52 + ], + "angle": 0, + "content": "not perfect and may introduce errors. For example, StableZero123 [29] may synthesize incorrect novel-view scenes; DepthAnything [64] may produce coarse predictions; and multimodal LLMs like LLaVA [27] may generate wrong answers. However, as shown in our experiments, Eval3D still aligns better with humans than all available automatic 3D evaluations, while being significantly more interpretable. Furthermore, our proposed framework is designed to support seamless integration of new and improved foundation models. As foundation models continue to improve, we anticipate even stronger alignment with human evaluations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.911, + 0.662 + ], + "angle": 0, + "content": "Broader impact: Eval3D can be used to identify and filter out inaccurate and unrealistic 3D assets, and potentially improve existing 3D generation algorithms. It could have a significant impact on the entertainment and robotics industries where high-quality 3D assets are crucial. However, it could also propagate the biases and limitations of foundation models into evaluation metrics. Additionally, it may affect employment within creative fields, as there could be a reduced need for human workers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.909, + 0.82 + ], + "angle": 0, + "content": "Conclusion: We present Eval3D, an interpretable and fine-grained evaluation framework for text/image-to-3D generation algorithms. Eval3D leverages the consistency among various vision and language foundation models to accurately evaluate generated 3D assets along multiple dimensions. It aligns with human judgment better than available alternatives and enables pixel-wise and spatial localization of artifacts in the generated 3D assets. We believe Eval3D can serve as a reliable framework for evaluating, filtering, and potentially improving 3D generation algorithms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment: The paper was partially supported by a gift from Ai2, an Amazon Science sponsored research grant, and DARPA TIAMAT program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13333" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Stability AI. Stable-zero123: Quality 3d object generation from single images, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.185 + ], + "angle": 0, + "content": "[2] Mark Boss, Raphael Braun, V. Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. Nerd: Neural reflectance decomposition from image collections. ICCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.187, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[3] Emmanuelle Bourigault and Pauline Bourigault. Mvdiff: Scalable and flexible multi-view diffusion for 3d object reconstruction from single-view. arXiv preprint arXiv:2405.03894, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[4] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.384, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22246–22256, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.59 + ], + "angle": 0, + "content": "[8] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Vladimir G Kim, Matthew Fisher, Noam Aigerman, Hao Zhang, and Siddhartha Chaudhuri. Decor-gan: 3d shape detailization by conditional refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15740-15749, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[10] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian Scene Graph: Improving Reliability in Fine-Grained Evaluation for Text-to-Image Generation. In ICLR, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[12] Nam Anh Dinh, Haochen Wang, Greg Shakhnarovich, and Rana Hanocka. Loopdraw: a loop-based autoregressive model for shape synthesis and editing. ArXiv, abs/2212.04981, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[14] Stephanie Fu*, Netanel Tamir*, Shobhita Sundaram*, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[15] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feup: A model-agnostic framework for features at any resolution. In The Twelfth International Conference on Learning Representations, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.223, + 0.906, + 0.291 + ], + "angle": 0, + "content": "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.295, + 0.907, + 0.349 + ], + "angle": 0, + "content": "[17] William Gao, April Wang, Gal Metzer, Raymond A Yeh, and Rana Hanocka. Tetgan: A convolutional neural network for tetrahedral mesh generation. arXiv preprint arXiv:2210.05735, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.353, + 0.907, + 0.421 + ], + "angle": 0, + "content": "[18] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. three-studio: A unified framework for 3d content generation. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.425, + 0.907, + 0.48 + ], + "angle": 0, + "content": "[19] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.528, + 0.907, + 0.609 + ], + "angle": 0, + "content": "[21] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.613, + 0.907, + 0.669 + ], + "angle": 0, + "content": "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.672, + 0.907, + 0.739 + ], + "angle": 0, + "content": "[23] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 867-876, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.743, + 0.907, + 0.798 + ], + "angle": 0, + "content": "[24] Brian Karis. Journey to nanite. https://www.highperformancegraphics.org/slides22/Journey_to_Nanite.pdf, 2022. Presented at High Performance Graphics 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Marian Kleineberg, Matthias Fey, and Frank Weichert. Adversarial generation of continuous implicit shape representations. arXiv preprint arXiv:2002.00349, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13334" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[27] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, and Li Chunyuan Liu, Ziwei. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.204 + ], + "angle": 0, + "content": "[28] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ICCV, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[30] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[31] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.387, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[32] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.484, + 0.593 + ], + "angle": 0, + "content": "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.637 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf. Communications of the ACM, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.638, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.329, + 0.776 + ], + "angle": 0, + "content": "[38] OpenAI. Hello gpt-4o. 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[39] Maxime Oquab, Timoth'ee Darcet, Théo Moutakanni, Huy Q. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russ Howes, PoYao (Bernie) Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Huijiao Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[40] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.908, + 0.192 + ], + "angle": 0, + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.908, + 0.289 + ], + "angle": 0, + "content": "[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.908, + 0.375 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.908, + 0.446 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.908, + 0.516 + ], + "angle": 0, + "content": "[45] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.908, + 0.614 + ], + "angle": 0, + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.908, + 0.673 + ], + "angle": 0, + "content": "[47] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.908, + 0.742 + ], + "angle": 0, + "content": "[48] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.908, + 0.786 + ], + "angle": 0, + "content": "[49] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and X. Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.908, + 0.87 + ], + "angle": 0, + "content": "[50] Jiao Sun, Deqing Fu, Yushi Hu, Su Wang, Royi Rassin, Da-Cheng Juan, Dana Alon, Charles Herrmann, Sjoerd van Steenkiste, Ranjay Krishna, et al. Dreamsync: Aligning text-to-image generation with image understanding feedback. In Synthetic Data for Computer Vision Workshop @ CVPR 2024, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[51] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13335" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "3d generation with bootstrapped diffusion prior. arXiv, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[52] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[53] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[54] Luke Villapaz. Gta 5 costs 265 million to develop and market, making it the most expensive video game ever produced: Report. International Business Times, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.484, + 0.333 + ], + "angle": 0, + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[57] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[58] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[59] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with geo: On metrics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[60] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[61] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. arXiv, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[62] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv, 2023. 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[63] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[65] Yue Yang, Fan-Yun Sun, Luca Weihs, Eli VanderBilt, Alvaro Herrasti, Winson Han, Jiajun Wu, Nick Haber, Ranjay Krishna, Lingjie Liu, et al. Holodeck: Language guided generation of 3d embodied ai environments. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024), pages 20-25. IEEE/CVF, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.908, + 0.217 + ], + "angle": 0, + "content": "[66] Lior Yariv, Matan Atzmon, and Yaron Lipman. Universal differentiable renderer for implicit neural representations. arXiv, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.908, + 0.289 + ], + "angle": 0, + "content": "[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.908, + 0.358 + ], + "angle": 0, + "content": "[68] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.908, + 0.416 + ], + "angle": 0, + "content": "[69] Hongxiang Zhao, Xili Dai, Jianan Wang, Shengbang Tong, Jingyuan Zhang, Weida Wang, Lei Zhang, and Yi Ma. Ctrl123: Consistent novel view synthesis via closed-loop transcription. arXiv preprint arXiv:2403.10953, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13336" + } + ] +] \ No newline at end of file diff --git a/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_origin.pdf b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d5fbd30557396722c9f2c9044ed713a7c8c8d910 --- /dev/null +++ b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/ec87f8a0-6f47-4f99-9af3-0e617994c3ad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8a49e6834c5d376e7d703576924ea2e11cca373efd492b81bc2326785489183 +size 4405735 diff --git a/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/full.md b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b66b55e96388611d640ebcd0061b6873d84ca7a5 --- /dev/null +++ b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/full.md @@ -0,0 +1,299 @@ +# Eval3D: Interpretable and Fine-grained Evaluation for 3D Generation + +Shivam Duggal\*1 Yushi Hu\*2 Oscar Michel3 Aniruddha Kembhavi4 William T. Freeman1 Noah A. Smith2,5 Ranjay Krishna2,5 Antonio Torralba1 Ali Farhadi2,5 Wei-Chiu Ma6 + +1Massachusetts Institute of Technology 2University of Washington 3New York University 4Wayve 5Allen Institute for AI 6Cornell University + +![](images/a31224bffda5f833cbf114807fef4e21b412f738e41b98a18f9a4fe00a357813.jpg) +Figure 1. Challenges of 3D generation: (1) Structural inconsistency: lack of globally-coherent 3D shape; (2) Text-3D misalignment: failure to meet the requirements of the input text-prompt; (3) Semantic inconsistency: content change and incoherent semantics; (4) Geometric inconsistency: misaligned geometry and texture. + +# Abstract + +Despite the unprecedented progress in the field of 3D generation, current systems still often fail to produce high-quality 3D assets that are visually appealing and geometrically and semantically consistent across multiple viewpoints. To effectively assess the quality of the generated 3D data, there is a need for a reliable 3D evaluation tool. Unfortunately, existing 3D evaluation metrics often overlook the geometric quality of generated assets or merely rely on black-box multimodal large language models for coarse assessment. In this paper, we introduce Eval3D, a fine-grained, interpretable evaluation tool that can faithfully evaluate the quality of generated 3D assets based on various distinct yet complementary criteria. Our key observation is that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. We thus leverage a diverse set of models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. Compared to prior work, Eval3D provides pixel-wise measurement, enables accurate 3D spatial feedback, and aligns more closely with human judgments. We comprehensively evaluate existing 3D generation models using Eval3D and highlight the limitations and challenges + +of current models. Project page: http://eval3d.github.io. + +# 1. Introduction + +High-quality 3D assets play a pivotal role across various domains including gaming, film, augmented/virtual reality, and robotics [24, 54, 65]. They empower content creation applications to deliver visually compelling and immersive experiences and enable simulators to realistically reproduce the real world. Unfortunately, 3D asset creation is complex and tedious, requiring a lot of careful manual effort. Automatic (text/image-to-)3D generation, despite fast progress over the last few years, still falls short in key aspects including geometry, semantic consistency, etc. + +What makes a good (text/image-to-)3D generation? 3D assets are not created to merely be rendered onto a 2D image; rather, they are meant to live within a rich 3D world – the assets may be viewed from various distinct viewpoints, and their 3D geometry will affect how they interact with the surroundings (e.g., collision, reflection, etc). Therefore, a high-quality 3D asset should not only be visually appealing and faithful to its text/image input across multiple individual viewpoints, but its geometry and appearance should also be coherent and plausible when observed as a whole. In the absence of these characteristics, the utility of the generated + +![](images/d7a1e000e80b2f2cf3e833389166e77c350723a3498bf748527ebe28c5c1835b.jpg) +Figure 2. Eval3D offers interpretable, fine-grained, and human-aligned metrics to assess the quality of 3D generations from various aspects. We utilize a diverse array of foundation models and tools to measure the consistency among different representations of generated 3D assets. + +assets degrades drastically. Consider the generated 3D assets in Fig. 1. While the corgi (Fig. 1 (left)) looks reasonable from the front and the back, its overall 3D shape is quite poor and implausible – notice the Janus issue (i.e., presenting multiple faces). Similarly, while the fried chickens and waffles (Fig. 1(right)) appear faithful from one viewpoint, the 3D asset suffers from semantic and geometric inconsistency across views – as the object rotates, the maple syrup transform into a jar, and the fried chicken pieces merge with the waffles. These 3D-specific structural and consistency challenges continue to pose significant difficulties to existing (text/image-to-)3D generation models. + +What makes a good 3D evaluation metric? To advance (text/image-to-)3D asset generation and to effectively quantify the aforementioned issues, there is a need for a reliable 3D evaluation tool. Unfortunately, while significant progress has been made in evaluating 2D text-to-image models [10, 21, 59, 62, 67], the development of their 3D counterparts is lagging behind. Existing 3D evaluation metrics typically rely on black-box multimodal large language models (LLMs), such as GPT-4V, for assessment [61], or often overlook the geometric quality of generated assets [19]. A comprehensive 3D metric should be able to: (a) rate the consistency between the input text/image and the 3D generation and detect misalignments such as a missing object or part, wrong attribute, etc., (b) measure the structural and semantic consistency of the generated asset across the input and the output as well as across multiple views of the 3D asset, and (c) localize flaws and inconsistencies in the produced asset. Such a comprehensive metric would not only permit fine-grained comparisons between different 3D generations but also provide feedback on where and what to improve. + +Our method: We introduce Eval3D, an interpretable, and fine-grained evaluation tool that measures the quality of generated 3D assets from multiple dimensions. We build Eval3D based on the observation that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. For instance, to measure the geometric consistency of the generated chicken and waffles (i.e., the red region in Fig. 1(right)), one can derive the surface normal analytically from the 3D object, and then compare it with the normal estimated by a foundation model (e.g., Depth Anything [64]) from rendered RGB images. Based on 2D visual priors, waffles usually possess a grid-like 3D geometry. However, the grid-like patterns do not exist in the normal map derived from 3D. By computing the differences, we can automatically identify those regions with inconsistent surface normals and localize where the generative model leverages texture maps to "cheat." We adopt a diverse set of open-source foundation models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. To aggregate the measurements across multiple viewpoints, we further present several simple yet effective 2D and 3D-based fusion mechanisms. With these fine-grained, pixel-wise measurements, we can not only easily locate the problematic regions in 2D images (Fig. 3), but also lift the inconsistency map back to 3D and enable accurate spatial artifact localization (Fig. 5). To our knowledge, we are the first to provide such spatial feedback. + +To promote the use of our new evaluation tool, we curate a diverse set of prompts and images from multiple sources [18, 19, 51]. We evaluate a variety of (text/image-to-)3D generation models on our data and densely annotate human preferences based on different criteria to verify how well our metrics align with human assessment. We show that Eval3D + +![](images/5061bc2869063ff17218d4cd0cac07416692365b8fe45dc9ddcd7df336f02e8b.jpg) +Figure 3. Geometry inconsistency evaluates texture-geometry misalignment by comparing 3D rendered normal and image-based normal. Bright-yellow indicates large discrepancy. + +is not only able to automatically detect various generation artifacts (e.g., the Janus issue), but also accurately localize them (Fig. 3). Surprisingly, we also find that while existing state-of-the-art generations are visually appealing, they often still suffer from geometric or semantic inconsistencies. Finally, we demonstrate that although the accuracy of Eval3D is dependent on the accuracy of the foundation models, it consistently aligns more closely with, or is comparable to, human judgments across all dimensions than prior work. As foundation models will improve over time, we hypothesize that Eval3D will become increasingly reliable. + +# 2. Related Work + +Generative models for 3D asset generation: Early work on 3D asset generation trained probabilistic generative models directly on 3D data, most commonly using assets from ShapeNet [6]. While numerous generative modeling frameworks (e.g., GANs [8, 9, 16, 17, 26, 60], VAEs [4, 8, 12, 17, 40], flow models [5, 63]) and 3D representations (e.g., point clouds [5, 37, 63], meshes [16, 16], voxels [4, 9, 60], and implicit functions [8, 26, 32, 40]) have been explored, the quality and diversity of 3D generations are constrained by the quantity and characteristics of the available data. With this in mind, recent dominant approaches turn to leverage supervision from foundation models trained on massive collections of 2D images [44-46]. During generation, they propagate the gradient from multi-view 2D space back to 3D and iteratively update the underlying 3D representations [8, 23, 25, 34, 36, 43]. Since foundation models encode rich priors of our world, this significantly improves the quality of the generated objects. To further improve multi-view consistency [3, 30, 48, 58, 69], object geometry [7, 28, 53], or even the optimization process [33, 57, 68], researchers have proposed a variety of enhancements, such as multi-view diffusion [49]. While a general trend of progress is evident, the lack of evaluation tools makes it unclear which methods are superior and in which dimensions. Our framework addresses this confusion by providing a general metric to measure progress, which can be decomposed to identify performance across critical factors relevant to 3D generation. + +**Evaluating 3D generation:** Perhaps closest to our work are GPT-4V [61] and T3-Bench [19]. GPT-4V [61] is an ELOB-based method that uses the eponymous multimodal LLM to compare the quality of two generated objects, while T3-Bench [19] employs a vision-language model (VLM) to evaluate the semantic adherence of generated objects to the prompt. While we share similar insights, there are several key differences: first, unlike GPT-4V, which provides a pairwise preference, Eval3D offers an absolute metric, making it easier to quantify and compare performance across different models and tasks. Additionally, both GPT-4V and T3 rely mainly on language outputs for evaluation. The reliance on the inherently imprecise nature of language makes localizing the source of error difficult. This issue is particularly problematic for detecting multi-view inconsistencies, as each view may independently seem satisfactory to the VLM. In contrast, our method incorporates semantic judgments from VLMs [11, 21, 50] and uses pretrained computer vision models [39, 64] to provide precise and localized estimates of geometric and structural deviations at multiple scales of granularity. These error measurements can be back-projected to 3D, making Eval3D more informative. + +# 3. Eval3D + +Our aim is to develop an evaluation tool that can faithfully assess the quality of (generated) 3D assets based on various distinct yet complementary criteria. Unlike 2D images, which are merely projections of our visual world, these generated assets live in 3D. Their geometry influences how they interact with the environment, and they can be viewed from multiple angles. Therefore, it is crucial to ensure that the appearance of these 3D generations is coherent and visually appealing from various viewpoints, that their geometry is high-quality and consistent with their RGB counterparts, and that the generated content remains faithful to the input conditions (e.g., text, image). + +Towards this goal, we devise five distinct measurements to evaluate the quality of generated 3D assets from different aspects. Our key observation is that many of the desired 3D properties can be predicted from one another if the (generated) 3D asset is of high-quality; if the cues are unpredictable, then the 3D generation often has artifacts or is not + +coherent. Building upon this intuition, we consolidate a diverse set of foundation models [1, 38, 39, 62, 64] and tools to compute and predict a wide range of visual cues and measure their consistency. The measurements across multiple viewpoints are then aggregated to reflect the quality of the 3D generation with respect to different aspects. For each critic, we scale the score to lie within $0 - 100\%$ . + +Notation: Let $t$ be the input text prompt and $\mathcal{I}^{\mathrm{GT}}$ be the optional image condition. Let $\mathcal{O}_t$ be the text-conditioned 3D asset, where the underlying representation could be either a mesh $\mathcal{O}_t^{\mathrm{mesh}}$ or an implicit representation $\mathcal{O}_t^{\mathrm{imp}}$ . Let $\{\mathcal{I}_i\}_{i=1}^N$ denote the rendered RGB images of the generated asset $\mathcal{O}_t$ at viewpoint $v_i$ . Let $f$ denote the foundation models or tools that we adopt to extract visual cues. + +# 3.1. Geometric Consistency + +To evaluate how good the geometry of an asset is, one naive way is to consume the 3D asset and output a score. Unfortunately, it is unclear what constitutes a good geometry and there is no such data or model that allows us to directly predict the geometric quality. + +In this work, we take a proxy and measure the consistency between the surface normal analytically derived from the 3D asset and the normal predicted by a foundation model from 2D images. The intuition is that a good 3D geometry should be reflective of its visual appearance. If the derived geometry does not align with the normal predicted from RGB (e.g., the chicken waffle in Fig. 1, right), then one of the components is incorrect. Specifically, we first compute the analytical surface normal of the 3D asset by taking the normal of each face in the mesh, or by taking the negative gradient of the density [2, 35] or signed distance field [55, 66], depending on the 3D representation. Then for each viewpoint, we exploit Depth Anything [64] to compute the monocular depth and convert it into normal. Empirically we find it more robust than direct surface normal estimation [13]. Finally, we project the analytical normal to each view, measure the pixel-wise angular difference, and compute the amount of inliers: + +Geometric consistency $=$ + +$$ +\frac {1}{N _ {p}} \sum_ {p} \mathbb {1} \left[ \operatorname {a r c c o s} \left(\mathbf {n} _ {p} ^ {\text {a n a l}} \cdot \mathbf {n} _ {p} ^ {\text {p r e d}}\right) < \delta^ {\text {n o r m}} \right] \tag {1} +$$ + +Here, we iterate over all pixels $p$ . $\mathbf{n}^{\mathrm{anal}}$ indicates that the normal of pixel $p$ is analytically derived, while $\mathbf{n}^{\mathrm{pred}}$ suggests it is predicted. $N_{p}$ denotes the number of valid rendered pixels. In practice, we set the threshold $\delta^{\mathrm{norm}}$ to $23^{\circ}$ based on a hold-out validation set. + +# 3.2. Semantic Consistency + +Another crucial factor for a 3D generation to be appealing and useful is that the underlying content and semantics + +remain consistent and coherent across different viewpoints. Unlike visual appearance, which may be affected by extrinsic factors such as lighting and change across views, the semantics of an asset should be invariant to different perspectives. To quantify the degree to which the semantics change across views, we capitalize on large vision foundation models that encode rich semantics. + +We first extract DINO features for every rendered image $\mathcal{F}_i^{\mathrm{DINO}} = f^{\mathrm{DINO}}(\mathcal{I}_i)$ . Then, for each 3D point $x$ on the surface of the asset $\mathcal{O}$ , we project it onto 2D feature maps $\{\mathcal{F}_i^{\mathrm{DINO}}\}$ to retrieve its corresponding features $\{\mathcal{F}_i^{\mathrm{DINO}}(\pi_{v_i}(x))\}$ . Here, $\pi$ is the projection operator. Since a point may be visible from multiple images, we compute the variance of its DINO features. Specifically, we calculate the variance along each feature dimension separately across viewpoints and then take the mean of these variances across all feature dimensions. Finally, we compute the ratio of vertices whose variance is below a certain threshold. The intuition is that if the content and semantics of a 3D asset remain relatively stable across different viewpoints, its DINO features should be consistent. In practice, for computational efficiency, we only consider the vertices of the 3D mesh $x^{\mathrm{vert}}$ as 3D points: + +$$ +\text {S e m a n t i c c o n s i s t e n c y} = \tag {2} +$$ + +$$ +\frac {1}{N _ {\text {v e r t}}} \sum_ {x ^ {\text {v e r t}}} \mathbb {1} \left[ \operatorname {m e a n} \left(\operatorname {V a r} \left(\left\{\mathcal {F} _ {i} ^ {\mathrm {D I N O}} \left(\pi_ {v _ {i}} \left(x ^ {\text {v e r t}}\right) \right.\right)\right) < \delta^ {\mathrm {D I N O}} \right] \right. +$$ + +We set $\delta^{\mathrm{DINO}}$ to the 70th percentile of average DINO variance based on a hold-out validation set. + +# 3.3. Structural Consistency + +A generated asset that looks plausible from a single view does not necessarily imply that it remains coherent when observed as a whole. The third criterion we are interested in measuring is whether the overall structure of the generated asset is coherent and plausible. + +We frame this assessment as a novel view synthesis (NVS) problem. Our intuition is that if the generated asset is overall plausible, one should be able to synthesize, to some degree, what other views might look like. For instance, given an image of the back of the head, one should be capable of predicting the face, although there might be variations in the size of the eyes or nose compared to the actual generation. On the other hand, if the generation suffers from artifacts or noise, it would be difficult to effectively conduct cross-view prediction, and the discrepancies could be significant. + +With this motivation in mind, given a rendered RGB image $\mathcal{I}_i$ as input, we first utilize Stable-Zero123 [1], a diffusion-based NVS model, to predict what an image from another unobserved viewpoint might look like, denoted as $\mathcal{I}_{i\rightarrow j} = f^{0123}(\mathcal{I}_i;v_i,v_j)$ . We then measure the difference between the predicted image $\mathcal{I}_{i\rightarrow j}$ and the image rendered from the generated asset $\mathcal{I}_j$ . Since we care more about the + +![](images/59250fd286ab4c24094ae601e98dc57a67321c176ad3c183ad777e2767297bda.jpg) +Figure 4. Structural consistency evaluates the geometric coherence of the generated 3D assets by comparing rendered views with the predictions from a novel view synthesis model (Zero-123) across various rotations. We utilize DreamSim to assess image similarity. + +perceptual similarity rather than pixel-wise differences, we employ DreamSim [14], a human-aligned perceptual metric, for this similarity assessment (see Fig. 4 for overview). Finally, we repeat the same operation for multiple views $j$ and then take the average. In practice, to increase robustness, we consider different images $\mathcal{I}_i$ as input and select the best one: + +Structural consistency $=$ + +$$ +\max _ {i} \frac {1}{N} \sum_ {j = 1} ^ {N} \left(1 - f ^ {\text {D r e a m S i m}} \left(\mathcal {I} _ {i \rightarrow j} ^ {\text {p r e d}}, \mathcal {I} _ {j}\right)\right) \tag {3} +$$ + +We take as input both the canonical view (rendered at $0^{\circ}$ ), and the rendered image from $90^{\circ}$ . We predict novel viewpoints at an interval of $90^{\circ}$ . + +# 3.4. Text-3D Alignment + +Another often overlooked aspect of 3D evaluation is the fidelity of the generated 3D asset to the input conditions. For instance, are all the mentioned objects being generated? Do the relationships among the objects align with what the text describes? The alignment issue is critical as it reflects how accurately the 3D generation adheres to the instructions. + +To effectively measure the text-3D alignment, we draw inspiration from TIFA [21], an evaluation metric for text-to-image generation, and adapt it for assessing text-to-3D generation. The key idea is to measure the consistency among large language models (LLMs) and multimodal LLMs across multiple views, and use this consistency as a proxy for evaluation. Specifically, we first employ a LLM to generate $M$ multiple-choice question-answer tuples $\{Q_j,C_j,A_j\}_{j = 1}^M$ based on the input text prompt $t$ , where $Q_{j}$ is a question, $C_j$ is a set of answer choices, and $A_{j}\in C_{j}$ is the gold answer. Then, for each question $Q_{j}$ , we employ a multimodal LLM to produce an answer based on each input image $A_{j}^{v_{i},\mathrm{VQA}} = f^{\mathrm{MLLM}}(\mathcal{I}_{i},Q_{j})$ . If the answer from the multimodal LLM $A_{j}^{v_{i},\mathrm{VQA}}$ aligns with the answer derived from the LLM $A_{j}$ , we conclude that the specific property in question is aligned at such viewpoint $v_{i}$ . For example, given a text prompt $t = "\text{A statue of a dog}"$ , + +the LLM could generate question-answer tuples such as {"Is there a statue?", "Yes", "No", "Yes"} and view-dependent alignment is assessed based on whether both the text and the image confirm the presence of a statue (see Fig. 2). To aggregate individual scores, we repeat the same procedure for all questions $\{Q_j\}_{j=1}^M$ and across all viewpoints $\{v_i\}_{i=1}^N$ : + +Text-3D alignment = + +$$ +\frac {1}{M} \sum_ {j = 1} ^ {M} \operatorname {A n y} _ {v _ {i}} \left(\operatorname {A l l} _ {v \in \operatorname {A d j} \left(v _ {i}\right)} \left(\mathbb {1} \left[ A _ {j} = = A _ {j} ^ {v, \mathrm {V Q A}} \right]\right)\right) \tag {4} +$$ + +A question is considered globally aligned if and only if, for some viewpoint $v_{i}$ , the answers of all its neighboring views are aligned. We find that this method is effective for dealing with occlusions and the noise of multimodal LLM. In practice, we adopt LLaVA-NeXT-7B [27] as our VQA model, and we follow DSG [10] for generating and evaluating our question-answer pairs. + +# 3.5. Aesthetic + +Finally, given that a generated 3D asset can be viewed from a wide range of angles, it is crucial to ensure that it appears visually appealing across all possible views. In this paper, we exploit two different approaches, with varying computational cost, to compute the aesthetic score $f^{\text{aesthetic}}(\{\mathcal{I}_i\})$ . First, following [61], we use GPT-4o [38] to conduct pairwise quality comparisons of 3D assets and compute the ELO score of each model. We improve the prompts and normalize the ELO score to 0-100% (by comparing with 3D generation models in Sec. 4.2). While effective, the closed-source nature and the high computational cost resulting from combinatorial comparison may create barriers to large-scale evaluation. As an alternative, we employ ImageReward [62], an open-source aesthetic estimator, to measure the aesthetic quality of each rendered image $f^{\text{ImgReward}}(\mathcal{I}_i)$ and then compute the average. We compare both implementations in Sec. 4 and discuss their respective strengths. + +![](images/75d8ec71c8150b0f58c278378c317b16ba80c31467aad19ad40a05596deef841.jpg) +Generated 3D asset +Figure 5. 3D inconsistency maps: The proposed 3D metrics, semantic and geometric consistencies, allow fine-grained localization of the artifacts (eg: Janus issue: multiple nose / face, inconsistent hand geometry, arbitrary surface patterns on the back) by fusing / computing the metrics in 3D space. + +# 3.6. Localizing 3D Artifacts + +Unlike previous works [19, 61], our proposed geometric and semantic consistency measures operate in the 3D space, allowing us to not only classify the generation as an artifact but also to localize the artifacts on the extracted 3D mesh. For the semantic consistency, artifacts can be localized by highlighting outlier vertices—specifically, mesh vertices with DINO feature variance exceeding a certain threshold. Fig. 5 illustrates an example where vertices with high semantic inconsistency (Eq. 2) significantly contribute to the generated 3D inconsistencies like the Janus issue and extraneous geometries. For the geometric consistency, we back-project the pixel-wise cosine distance map between the predicted and rendered geometric normals onto the 3D mesh, highlighting vertices where the mean or maximum cosine distance across different viewpoints exceeds a threshold. Finally, as shown in Fig. 5, fusing the proposed semantic or geometric consistency metrics to the 3D space could enable estimating 3D uncertainty at novel viewpoints. Such consistency cues can be leveraged as a self-supervised refinement metric to enhance the generated geometry. + +# 4. Experiments + +In this section, we begin by discussing the curated Eval3D evaluation benchmark. (Sec. 4.1) We then outline the experimental details, covering the targeted 3D generative models, experimental setup, and human annotation process. (Sec. 4.2) Finally, we present our experimental results. (Sec. 4.3) We first assess the agreements between Eval3D and human annotations, showing that Eval3D aligns well with humans on all evaluation dimensions. Then we benchmark text/image-to-3D algorithms with Eval3D and discuss our findings. + +# 4.1. Eval3D Benchmark + +To thoroughly evaluate 3D generation algorithms, we curate the Eval3D benchmark. Unlike prior work, which provides only text prompts, Eval3D additionally includes a scene graph, dense human annotations, and optionally a carefully selected SDXL image suitable for image-to-3D generation. + +![](images/751561f0965350bf3589f736afd0c9b71595eed1013552cf742c3ffb98186dd3.jpg) +Text prompt = "A panda rowing a boat in a pond." +MV-Dream +Figure 6. Evaluation examples from Eval3D. While the generation from MV-Dream appears visually appealing at first glance, both the boat and the paddle are missing (reflected in text-3D alignment). There are also geometric artifacts on the left side of the panda (reflected in geometric consistency). Please refer to the Appendix for detailed breakdown of each metric. + +Structural consistency: 82.22 + +Text-3D alignment: 50 + +Semantic consistency: 76.8 + +Geometric consistency: 82.85 + +Aesthetic score: 84.33 + +![](images/d11054cb8046dd34b6296a167f1c454568ef9f6e1dad289c431a0f1bac2d9357.jpg) +TextMesh + +Structural consistency: 50.12 + +Text-3D alignment: $50\%$ + +Semantic consistency: $24.40\%$ + +Geometric consistency: 81.91% + +Aesthetic score: 26.09 + +Prompt source: Eval3D contains 160 text prompts for text-to-3D evaluation. The prompts come from diverse sources, including threestudio [18] and the GPT-generated prompts from prior works [19, 51, 61]. We curate the prompt source such that 80 prompts contain a single object, and 80 prompts contain multiple objects. For image-to-3D evaluation, we use the 20 image-text pairs from Dreamcraft3D [51]. + +Scene graphs: To facilitate further investigation, we also provide the scene graph for each prompt. Following prior work [10], the scene graph contains 5 kinds of elements: entity (e.g., "statue"), attribute (e.g., colors, materials), relation (e.g., "on top of"), action (e.g., "sit"), and others (e.g., art style). Please refer to Appendix for more details. + +Annotations: We provide dense expert annotations on all 3D evaluation criteria discussed in Sec. 3 for 160 prompts across six 3D generation models. We note that our annotation is an order of magnitude larger than prior work. For example, [61] only contains pairwise comparison for 3 prompts for each model pair. [19] annotates quality and alignment for 90 prompts, ignoring geometric, semantic, and structural consistency. See Sec. 4.2 and Appendix for more details. + +# 4.2. Experimental Details + +Target 3D generation models: We evaluate the proposed metrics on six state-of-the-art text-to-3D algorithms (i.e., DreamFusion [41], Magic3D [28], ProlificDreamer [56], TextMesh [52], and MVDream [49]) and two image-to-3D algorithms (i.e., Magic123 [42] and DreamCraft3D [51]). These algorithms represent a diverse set of text/image-to-3D generative models, featuring different 3D representations (NeRF, NeRF + Tetrahedral Mesh, Gaussian Splitting), various optimization strategies (single-stage vs. multi-stage), and different generative (diffusion) priors (single view vs. multi-view). Consequently, the generated assets exhibit a + +
ApproachGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
CLIP-Score [20]43.6-64.062.074.1
ImageReward [62]44.3-63.775.479.5
T3-Bench [19]---76.556.7
GPT4V [61]46.9-68.985.672.8
Eval3D (Ours)83.068.069.287.488.7
+ +Table 1. Pairwise rating agreements between human and evaluation metrics. We measure the probability that the automatic evaluation makes the same decision as human annotators when comparing two 3D assets. Eval3D achieves the best alignment with humans. + +wide range of artifacts, enabling a thorough study of fine-grained 3D evaluation. We leverage the implementations from ThreeStudio [18]. Please refer to Appendix for details. Setup: We render each generated 3D asset from 120 different viewpoints, with cameras positioned at an elevation of $15^{\circ}$ and an azimuth varying from $0^{\circ}$ to $360^{\circ}$ degrees. We render RGB images, surface normals, and opacity maps at a resolution of $512 \times 512$ . To compute semantic consistency, we extract the underlying 3D mesh using either marching cubes [31] or DMTet [47]. We upsample DINO features to $256 \times 256$ usingFeatUp [15] and then fuse them with the extracted 3D mesh. Due to cost considerations, we use 12 viewpoints for the text-3D alignment metric. + +Human annotations: Previous human annotation data either lacked coverage for all 3D evaluation dimensions (e.g., only focuses on aesthetics and alignment, ignoring geometry and structure) or included a very limited number of annotations. To address these issues, we performed dense expert annotations on all 3D assets generated by six text-to-3D models based on different criteria discussed in Sec. 3. We carefully designed the annotation process for each metric to ensure high inter-annotator agreement. Take geometric consistency for example. For each text prompt, we showed annotators videos of RGB and rendered surface normals, placed side-by-side for all six algorithms. Then we asked annotators to rank the generations based on how well the texture aligns with the surface normals. Ties were allowed. We adopted a similar annotation procedure for aesthetics, but showing only the RGB videos. As for semantic and structural consistency, since it is difficult to rank (e.g., whether a two-faced dog is better or worse than a three-legged astronaut), we instead framed it as a yes/no question and asked annotators whether they observed any inconsistencies. Finally, for text-3D alignment, annotators looked at the RGB videos and answered a series of multi-choice questions generated automatically from the text prompt, as discussed in [10]. We direct the readers to the Appendix for more comprehensive details on our annotation process. + +# 4.3. Experimental Results + +Human alignment: An optimal 3D evaluation metric should align closely with human judgment. We assess this alignment by comparing human annotations with ours and baseline metrics. As shown in Tab. 1, Eval3D outperforms all baseline evaluation metrics across all dimensions. Specif + +ically, our text-3D metric aligns approximately $16\%$ better with human annotations than GPT-4V [61]. Our geometric consistency metric, which is analogous to GPT-4V's texture-geometry alignment [61], achieves $36\%$ better alignment with human judgments compared to all other baselines. Importantly, our geometric and semantic consistency metrics are also interpretable and fine-grained, allowing us to localize potential artifacts (see Figs. 5 and 3). We note that we are the first to capture the multi-view semantic consistency. + +Benchmarking text-to-3D models: Tab. 2 shows the scores (scaled to $0 - 100\%$ ) of each text-to-3D model under various criteria. MVDream [49] generally performs the best, ranking 1st or 2nd on all metrics, except for geometry consistency. Upon inspecting the MVDream generations, we find that Eval3D successfully captures its weaknesses: despite their high overall quality, MVDream assets suffer from holes and frequently hallucinate details. DreamFusion [41], TextMesh [52], and Magic3D [28] perform poorly across most metrics. They produce low-resolution geometries, which result in low aesthetic scores. Additionally, they do not accurately adhere to the text prompts, achieving only $40\% - 65\%$ in text-3D alignment. Nevertheless, their geometric consistency is comparable to, or better than, that of MVDream. We hypothesize that this is due to their tendency to generate simpler and more convex geometries. ProlificDreamer achieves high scores in aesthetics and text-3D alignment, yet it ranks last or second-to-last in all three geometric metrics. In contrast, previous studies [19, 61] rate ProlificDreamer highly. Upon inspecting these outputs, we observe that Eval3D more effectively highlights the weaknesses of ProlificDreamer. Fig 5 showcases a generation by ProlificDreamer that, despite its aesthetic sharpness and alignment with the text prompt, suffers from geometric, structural, and semantic inconsistencies. Gaussian Splatting ranks the lowest on geometric consistency. This validates [22]'s finding that Gaussian-Splatting-based methods generate extremely noisy geometry. Interestingly, despite the noisy geometries, the generations rank highly in terms of semantic consistency measured by DINOv2 [39] and multi-view structural consistency measured using Zero123 [29]. We show a few qualitative examples as well as their evaluation results in Fig. 6. Please refer to the Appendix for those of the other text-to-3D algorithms. + +Benchmarking image-to-3D models: Unlike previous stud + +
AlgorithmGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
TextMesh [52]83.174.169.618.943.2
DreamFusion [41]88.367.177.624.250.8
Magic3D [28]94.161.579.351.264.2
ProlificDreamer [56]77.159.575.160.676.9
Gaussian Splitting [68]67.984.881.862.366.7
MVDream [49]85.874.381.182.567.7
+ +Table 2. Eval3D results of Text-to-3D models. We benchmark the text-to-3D models discussed in Sec. 4.2 on all the evaluation criteria in Eval3D. For each criterion, the score ranges from 0 to $100\%$ . Higher values indicate better performance. + +
AlgorithmInputGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
Magic3D [28]97.167.280.325.990.2
MVDream [49]89.776.183.476.880.5
Magic123 [42]✓97.670.782.039.288.4
DreamCraft3D [51]✓97.082.881.369.191.9
+ +Table 3. Eval3D results of image-to-3D models. We evaluate Magic123 [42] and DreamCraft3D [51], both of which take images and texts as inputs. We also show the results of two text-to-3D models (Magic3D [28]) and MVDream [49]) using the same text prompts as inputs. + +ies [19, 61], we also evaluate state-of-the-art image+text-to-3D generative algorithms (i.e., Magic123[42] and DreamCraft3D[51]). To validate to which degree an input image could help, we compare these algorithms with two text-only 3D generation algorithms, Magic3D [28] and MVDream [49]. As shown in Tab. 3, text+image-to-3D generations better align with the text prompts and perform comparable on geometric and structural consistency metrics. In particular, although Magic123 and Magic3D share similar generation pipelines, Magic123, with the aid of image guidance, achieves improved semantic and structural consistency. We conjecture that the improvement stems from the use of diffusion models to generate novel views from the input image, which serve as priors and help avoid structural inconsistencies like the Janus problem. + +# 5. Analysis and Discussion + +Open-source vs. closed-source evaluation: Compared to previous work [61], which exclusively capitalized on closed-source multimodal LLMs for assessment, we propose a fully open-source solution for 3D evaluation. Additionally, we design Eval3D to be flexible, allowing foundation models and tools to be easily swapped with alternatives possessing similar capabilities. For instance, one could replace the open-source LLaVA model [27] with the closed-source GPT-4o for the text-3D alignment metric, or alternatively, select either GPT-4o or ImageReward [62] for aesthetic measurement. We experimented with these models and found that for text-3D alignment, our human alignment is similar using two models (88.7% using LLaVA, 91.5% using GPT-4o), all substantially outperforming prior metrics. While GPT-4o aligns more closely with human judgments than ImageReward does, the overall ranking of 3D algorithms assessed by the two is similar, with an edit distance of 1. These results suggest that open-source models are effective for evaluation. + +Limitations: The foundation models used in the paper are + +not perfect and may introduce errors. For example, StableZero123 [29] may synthesize incorrect novel-view scenes; DepthAnything [64] may produce coarse predictions; and multimodal LLMs like LLaVA [27] may generate wrong answers. However, as shown in our experiments, Eval3D still aligns better with humans than all available automatic 3D evaluations, while being significantly more interpretable. Furthermore, our proposed framework is designed to support seamless integration of new and improved foundation models. As foundation models continue to improve, we anticipate even stronger alignment with human evaluations. + +Broader impact: Eval3D can be used to identify and filter out inaccurate and unrealistic 3D assets, and potentially improve existing 3D generation algorithms. It could have a significant impact on the entertainment and robotics industries where high-quality 3D assets are crucial. However, it could also propagate the biases and limitations of foundation models into evaluation metrics. Additionally, it may affect employment within creative fields, as there could be a reduced need for human workers. + +Conclusion: We present Eval3D, an interpretable and fine-grained evaluation framework for text/image-to-3D generation algorithms. Eval3D leverages the consistency among various vision and language foundation models to accurately evaluate generated 3D assets along multiple dimensions. It aligns with human judgment better than available alternatives and enables pixel-wise and spatial localization of artifacts in the generated 3D assets. We believe Eval3D can serve as a reliable framework for evaluating, filtering, and potentially improving 3D generation algorithms. + +Acknowledgment: The paper was partially supported by a gift from Ai2, an Amazon Science sponsored research grant, and DARPA TIAMAT program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA. + +# References + +[1] Stability AI. Stable-zero123: Quality 3d object generation from single images, 2024. 4 +[2] Mark Boss, Raphael Braun, V. Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. Nerd: Neural reflectance decomposition from image collections. ICCV, 2020. 4 +[3] Emmanuelle Bourigault and Pauline Bourigault. Mvdiff: Scalable and flexible multi-view diffusion for 3d object reconstruction from single-view. arXiv preprint arXiv:2405.03894, 2024. 3 +[4] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236, 2016. 3 +[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 3 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3 +[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22246–22256, 2023. 3 +[8] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3 +[9] Zhiqin Chen, Vladimir G Kim, Matthew Fisher, Noam Aigerman, Hao Zhang, and Siddhartha Chaudhuri. Decor-gan: 3d shape detailization by conditional refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15740-15749, 2021. 3 +[10] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian Scene Graph: Improving Reliability in Fine-Grained Evaluation for Text-to-Image Generation. In ICLR, 2024. 2, 5, 6, 7 +[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation, 2024. 3 +[12] Nam Anh Dinh, Haochen Wang, Greg Shakhnarovich, and Rana Hanocka. Loopdraw: a loop-based autoregressive model for shape synthesis and editing. ArXiv, abs/2212.04981, 2022.3 +[13] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 4 + +[14] Stephanie Fu*, Netanel Tamir*, Shobhita Sundaram*, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv, 2023. 5 +[15] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feup: A model-agnostic framework for features at any resolution. In The Twelfth International Conference on Learning Representations, 2024. 7 +[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3 +[17] William Gao, April Wang, Gal Metzer, Raymond A Yeh, and Rana Hanocka. Tetgan: A convolutional neural network for tetrahedral mesh generation. arXiv preprint arXiv:2210.05735, 2022. 3 +[18] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. three-studio: A unified framework for 3d content generation. 2, 6, 7 +[19] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv, 2023. 2, 3, 6, 7, 8 +[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 7 +[21] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 2, 3, 5 +[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 7 +[23] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 867-876, 2022. 3 +[24] Brian Karis. Journey to nanite. https://www.highperformancegraphics.org/slides22/Journey_to_Nanite.pdf, 2022. Presented at High Performance Graphics 2022. 1 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 3 +[26] Marian Kleineberg, Matthias Fey, and Frank Weichert. Adversarial generation of continuous implicit shape representations. arXiv preprint arXiv:2002.00349, 2020. 3 + +[27] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, and Li Chunyuan Liu, Ziwei. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 5, 8 +[28] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3, 6, 7, 8 +[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ICCV, 2023. 7, 8 +[30] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3 +[31] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 7 +[32] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3 +[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3 +[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3 +[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf. Communications of the ACM, 2020. 4 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022. 3 +[38] OpenAI. Hello gpt-4o. 2024. 4, 5 +[39] Maxime Oquab, Timoth'ee Darcet, Théo Moutakanni, Huy Q. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russ Howes, PoYao (Bernie) Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Huijiao Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 4, 7 + +[40] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 3 +[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. 2023. 6, 7, 8 +[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 6, 8 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[45] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. +[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3 +[47] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 7 +[48] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3 +[49] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and X. Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv, 2023. 3, 6, 7, 8 +[50] Jiao Sun, Deqing Fu, Yushi Hu, Su Wang, Royi Rassin, Da-Cheng Juan, Dana Alon, Charles Herrmann, Sjoerd van Steenkiste, Ranjay Krishna, et al. Dreamsync: Aligning text-to-image generation with image understanding feedback. In Synthetic Data for Computer Vision Workshop @ CVPR 2024, 2023. 3 +[51] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical + +3d generation with bootstrapped diffusion prior. arXiv, 2023. 2, 6, 8 +[52] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 6, 7, 8 +[53] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3 +[54] Luke Villapaz. Gta 5 costs 265 million to develop and market, making it the most expensive video game ever produced: Report. International Business Times, 2013. 1 +[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv, 2021. 4 +[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2024. 6, 8 +[57] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[58] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3 +[59] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with geo: On metrics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 2 +[60] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3 +[61] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. arXiv, 2024. 2, 3, 5, 6, 7, 8 +[62] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv, 2023. 2, 4, 5, 7, 8 +[63] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3 +[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4, 8 + +[65] Yue Yang, Fan-Yun Sun, Luca Weihs, Eli VanderBilt, Alvaro Herrasti, Winson Han, Jiajun Wu, Nick Haber, Ranjay Krishna, Lingjie Liu, et al. Holodeck: Language guided generation of 3d embodied ai environments. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024), pages 20-25. IEEE/CVF, 2024. 1 +[66] Lior Yariv, Matan Atzmon, and Yaron Lipman. Universal differentiable renderer for implicit neural representations. arXiv, 2020. 4 +[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[68] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3, 8 +[69] Hongxiang Zhao, Xili Dai, Jianan Wang, Shengbang Tong, Jingyuan Zhang, Weida Wang, Lei Zhang, and Yi Ma. Ctrl123: Consistent novel view synthesis via closed-loop transcription. arXiv preprint arXiv:2403.10953, 2024. 3 \ No newline at end of file diff --git a/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/images.zip b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0bddc669da64534432a4ddca50009c670a80a6ce --- /dev/null +++ b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55b9fb8c7b60b935f32f620ae79810b5029c69fd134d00845db3f5627088d36d +size 521953 diff --git a/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/layout.json b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c3fb46f57bf7a5244d7224bee8030f982c341067 --- /dev/null +++ b/2025/Eval3D_ Interpretable and Fine-grained Evaluation for 3D Generation/layout.json @@ -0,0 +1,8034 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 522, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 522, + 120 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 522, + 120 + ], + "type": "text", + "content": "Eval3D: Interpretable and Fine-grained Evaluation for 3D Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 538, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 538, + 171 + ], + "type": "text", + "content": "Shivam Duggal\\*1 Yushi Hu\\*2 Oscar Michel3 Aniruddha Kembhavi4 William T. Freeman1 Noah A. Smith2,5 Ranjay Krishna2,5 Antonio Torralba1 Ali Farhadi2,5 Wei-Chiu Ma6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 184, + 529, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 184, + 529, + 214 + ], + "spans": [ + { + "bbox": [ + 80, + 184, + 529, + 214 + ], + "type": "text", + "content": "1Massachusetts Institute of Technology 2University of Washington 3New York University 4Wayve 5Allen Institute for AI 6Cornell University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 222, + 539, + 369 + ], + "blocks": [ + { + "bbox": [ + 67, + 222, + 539, + 369 + ], + "lines": [ + { + "bbox": [ + 67, + 222, + 539, + 369 + ], + "spans": [ + { + "bbox": [ + 67, + 222, + 539, + 369 + ], + "type": "image", + "image_path": "a31224bffda5f833cbf114807fef4e21b412f738e41b98a18f9a4fe00a357813.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 370, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 555, + 404 + ], + "type": "text", + "content": "Figure 1. Challenges of 3D generation: (1) Structural inconsistency: lack of globally-coherent 3D shape; (2) Text-3D misalignment: failure to meet the requirements of the input text-prompt; (3) Semantic inconsistency: content change and incoherent semantics; (4) Geometric inconsistency: misaligned geometry and texture." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "spans": [ + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 438, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 297, + 714 + ], + "type": "text", + "content": "Despite the unprecedented progress in the field of 3D generation, current systems still often fail to produce high-quality 3D assets that are visually appealing and geometrically and semantically consistent across multiple viewpoints. To effectively assess the quality of the generated 3D data, there is a need for a reliable 3D evaluation tool. Unfortunately, existing 3D evaluation metrics often overlook the geometric quality of generated assets or merely rely on black-box multimodal large language models for coarse assessment. In this paper, we introduce Eval3D, a fine-grained, interpretable evaluation tool that can faithfully evaluate the quality of generated 3D assets based on various distinct yet complementary criteria. Our key observation is that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. We thus leverage a diverse set of models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. Compared to prior work, Eval3D provides pixel-wise measurement, enables accurate 3D spatial feedback, and aligns more closely with human judgments. We comprehensively evaluate existing 3D generation models using Eval3D and highlight the limitations and challenges" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 414, + 539, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 539, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 539, + 426 + ], + "type": "text", + "content": "of current models. Project page: http://eval3d.github.io." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 434, + 394, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 394, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 394, + 445 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 448, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 448, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 448, + 556, + 568 + ], + "type": "text", + "content": "High-quality 3D assets play a pivotal role across various domains including gaming, film, augmented/virtual reality, and robotics [24, 54, 65]. They empower content creation applications to deliver visually compelling and immersive experiences and enable simulators to realistically reproduce the real world. Unfortunately, 3D asset creation is complex and tedious, requiring a lot of careful manual effort. Automatic (text/image-to-)3D generation, despite fast progress over the last few years, still falls short in key aspects including geometry, semantic consistency, etc." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "What makes a good (text/image-to-)3D generation? 3D assets are not created to merely be rendered onto a 2D image; rather, they are meant to live within a rich 3D world – the assets may be viewed from various distinct viewpoints, and their 3D geometry will affect how they interact with the surroundings (e.g., collision, reflection, etc). Therefore, a high-quality 3D asset should not only be visually appealing and faithful to its text/image input across multiple individual viewpoints, but its geometry and appearance should also be coherent and plausible when observed as a whole. In the absence of these characteristics, the utility of the generated" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13326" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 46, + 517, + 270 + ], + "blocks": [ + { + "bbox": [ + 94, + 46, + 517, + 270 + ], + "lines": [ + { + "bbox": [ + 94, + 46, + 517, + 270 + ], + "spans": [ + { + "bbox": [ + 94, + 46, + 517, + 270 + ], + "type": "image", + "image_path": "d7a1e000e80b2f2cf3e833389166e77c350723a3498bf748527ebe28c5c1835b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 300 + ], + "type": "text", + "content": "Figure 2. Eval3D offers interpretable, fine-grained, and human-aligned metrics to assess the quality of 3D generations from various aspects. We utilize a diverse array of foundation models and tools to measure the consistency among different representations of generated 3D assets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 306, + 297, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 297, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 297, + 450 + ], + "type": "text", + "content": "assets degrades drastically. Consider the generated 3D assets in Fig. 1. While the corgi (Fig. 1 (left)) looks reasonable from the front and the back, its overall 3D shape is quite poor and implausible – notice the Janus issue (i.e., presenting multiple faces). Similarly, while the fried chickens and waffles (Fig. 1(right)) appear faithful from one viewpoint, the 3D asset suffers from semantic and geometric inconsistency across views – as the object rotates, the maple syrup transform into a jar, and the fried chicken pieces merge with the waffles. These 3D-specific structural and consistency challenges continue to pose significant difficulties to existing (text/image-to-)3D generation models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 297, + 713 + ], + "type": "text", + "content": "What makes a good 3D evaluation metric? To advance (text/image-to-)3D asset generation and to effectively quantify the aforementioned issues, there is a need for a reliable 3D evaluation tool. Unfortunately, while significant progress has been made in evaluating 2D text-to-image models [10, 21, 59, 62, 67], the development of their 3D counterparts is lagging behind. Existing 3D evaluation metrics typically rely on black-box multimodal large language models (LLMs), such as GPT-4V, for assessment [61], or often overlook the geometric quality of generated assets [19]. A comprehensive 3D metric should be able to: (a) rate the consistency between the input text/image and the 3D generation and detect misalignments such as a missing object or part, wrong attribute, etc., (b) measure the structural and semantic consistency of the generated asset across the input and the output as well as across multiple views of the 3D asset, and (c) localize flaws and inconsistencies in the produced asset. Such a comprehensive metric would not only permit fine-grained comparisons between different 3D generations but also provide feedback on where and what to improve." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 306, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 641 + ], + "type": "text", + "content": "Our method: We introduce Eval3D, an interpretable, and fine-grained evaluation tool that measures the quality of generated 3D assets from multiple dimensions. We build Eval3D based on the observation that many desired properties of 3D generation, such as semantic and geometric consistency, can be effectively captured by measuring the consistency among various foundation models and tools. For instance, to measure the geometric consistency of the generated chicken and waffles (i.e., the red region in Fig. 1(right)), one can derive the surface normal analytically from the 3D object, and then compare it with the normal estimated by a foundation model (e.g., Depth Anything [64]) from rendered RGB images. Based on 2D visual priors, waffles usually possess a grid-like 3D geometry. However, the grid-like patterns do not exist in the normal map derived from 3D. By computing the differences, we can automatically identify those regions with inconsistent surface normals and localize where the generative model leverages texture maps to \"cheat.\" We adopt a diverse set of open-source foundation models and tools as probes to evaluate the inconsistency of generated 3D assets across different aspects. To aggregate the measurements across multiple viewpoints, we further present several simple yet effective 2D and 3D-based fusion mechanisms. With these fine-grained, pixel-wise measurements, we can not only easily locate the problematic regions in 2D images (Fig. 3), but also lift the inconsistency map back to 3D and enable accurate spatial artifact localization (Fig. 5). To our knowledge, we are the first to provide such spatial feedback." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "To promote the use of our new evaluation tool, we curate a diverse set of prompts and images from multiple sources [18, 19, 51]. We evaluate a variety of (text/image-to-)3D generation models on our data and densely annotate human preferences based on different criteria to verify how well our metrics align with human assessment. We show that Eval3D" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13327" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 58, + 516, + 178 + ], + "blocks": [ + { + "bbox": [ + 92, + 58, + 516, + 178 + ], + "lines": [ + { + "bbox": [ + 92, + 58, + 516, + 178 + ], + "spans": [ + { + "bbox": [ + 92, + 58, + 516, + 178 + ], + "type": "image", + "image_path": "5061bc2869063ff17218d4cd0cac07416692365b8fe45dc9ddcd7df336f02e8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 555, + 205 + ], + "type": "text", + "content": "Figure 3. Geometry inconsistency evaluates texture-geometry misalignment by comparing 3D rendered normal and image-based normal. Bright-yellow indicates large discrepancy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 211, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 211, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 296, + 344 + ], + "type": "text", + "content": "is not only able to automatically detect various generation artifacts (e.g., the Janus issue), but also accurately localize them (Fig. 3). Surprisingly, we also find that while existing state-of-the-art generations are visually appealing, they often still suffer from geometric or semantic inconsistencies. Finally, we demonstrate that although the accuracy of Eval3D is dependent on the accuracy of the foundation models, it consistently aligns more closely with, or is comparable to, human judgments across all dimensions than prior work. As foundation models will improve over time, we hypothesize that Eval3D will become increasingly reliable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 142, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 142, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 142, + 379 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 391, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 391, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 391, + 296, + 713 + ], + "type": "text", + "content": "Generative models for 3D asset generation: Early work on 3D asset generation trained probabilistic generative models directly on 3D data, most commonly using assets from ShapeNet [6]. While numerous generative modeling frameworks (e.g., GANs [8, 9, 16, 17, 26, 60], VAEs [4, 8, 12, 17, 40], flow models [5, 63]) and 3D representations (e.g., point clouds [5, 37, 63], meshes [16, 16], voxels [4, 9, 60], and implicit functions [8, 26, 32, 40]) have been explored, the quality and diversity of 3D generations are constrained by the quantity and characteristics of the available data. With this in mind, recent dominant approaches turn to leverage supervision from foundation models trained on massive collections of 2D images [44-46]. During generation, they propagate the gradient from multi-view 2D space back to 3D and iteratively update the underlying 3D representations [8, 23, 25, 34, 36, 43]. Since foundation models encode rich priors of our world, this significantly improves the quality of the generated objects. To further improve multi-view consistency [3, 30, 48, 58, 69], object geometry [7, 28, 53], or even the optimization process [33, 57, 68], researchers have proposed a variety of enhancements, such as multi-view diffusion [49]. While a general trend of progress is evident, the lack of evaluation tools makes it unclear which methods are superior and in which dimensions. Our framework addresses this confusion by providing a general metric to measure progress, which can be decomposed to identify performance across critical factors relevant to 3D generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 211, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 556, + 475 + ], + "type": "text", + "content": "**Evaluating 3D generation:** Perhaps closest to our work are GPT-4V [61] and T3-Bench [19]. GPT-4V [61] is an ELOB-based method that uses the eponymous multimodal LLM to compare the quality of two generated objects, while T3-Bench [19] employs a vision-language model (VLM) to evaluate the semantic adherence of generated objects to the prompt. While we share similar insights, there are several key differences: first, unlike GPT-4V, which provides a pairwise preference, Eval3D offers an absolute metric, making it easier to quantify and compare performance across different models and tasks. Additionally, both GPT-4V and T3 rely mainly on language outputs for evaluation. The reliance on the inherently imprecise nature of language makes localizing the source of error difficult. This issue is particularly problematic for detecting multi-view inconsistencies, as each view may independently seem satisfactory to the VLM. In contrast, our method incorporates semantic judgments from VLMs [11, 21, 50] and uses pretrained computer vision models [39, 64] to provide precise and localized estimates of geometric and structural deviations at multiple scales of granularity. These error measurements can be back-projected to 3D, making Eval3D more informative." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 482, + 368, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 368, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 368, + 494 + ], + "type": "text", + "content": "3. Eval3D" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 497, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 640 + ], + "type": "text", + "content": "Our aim is to develop an evaluation tool that can faithfully assess the quality of (generated) 3D assets based on various distinct yet complementary criteria. Unlike 2D images, which are merely projections of our visual world, these generated assets live in 3D. Their geometry influences how they interact with the environment, and they can be viewed from multiple angles. Therefore, it is crucial to ensure that the appearance of these 3D generations is coherent and visually appealing from various viewpoints, that their geometry is high-quality and consistent with their RGB counterparts, and that the generated content remains faithful to the input conditions (e.g., text, image)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "Towards this goal, we devise five distinct measurements to evaluate the quality of generated 3D assets from different aspects. Our key observation is that many of the desired 3D properties can be predicted from one another if the (generated) 3D asset is of high-quality; if the cues are unpredictable, then the 3D generation often has artifacts or is not" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13328" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "content": "coherent. Building upon this intuition, we consolidate a diverse set of foundation models [1, 38, 39, 62, 64] and tools to compute and predict a wide range of visual cues and measure their consistency. The measurements across multiple viewpoints are then aggregated to reflect the quality of the 3D generation with respect to different aspects. For each critic, we scale the score to lie within " + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "inline_equation", + "content": "0 - 100\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": "Notation: Let " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": " be the input text prompt and " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": " be the optional image condition. Let " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": " be the text-conditioned 3D asset, where the underlying representation could be either a mesh " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t^{\\mathrm{mesh}}" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": " or an implicit representation " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t^{\\mathrm{imp}}" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": " denote the rendered RGB images of the generated asset " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": " at viewpoint " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 156, + 296, + 240 + ], + "type": "text", + "content": " denote the foundation models or tools that we adopt to extract visual cues." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 248, + 186, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 186, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 186, + 261 + ], + "type": "text", + "content": "3.1. Geometric Consistency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 326 + ], + "type": "text", + "content": "To evaluate how good the geometry of an asset is, one naive way is to consume the 3D asset and output a score. Unfortunately, it is unclear what constitutes a good geometry and there is no such data or model that allows us to directly predict the geometric quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 325, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 539 + ], + "type": "text", + "content": "In this work, we take a proxy and measure the consistency between the surface normal analytically derived from the 3D asset and the normal predicted by a foundation model from 2D images. The intuition is that a good 3D geometry should be reflective of its visual appearance. If the derived geometry does not align with the normal predicted from RGB (e.g., the chicken waffle in Fig. 1, right), then one of the components is incorrect. Specifically, we first compute the analytical surface normal of the 3D asset by taking the normal of each face in the mesh, or by taking the negative gradient of the density [2, 35] or signed distance field [55, 66], depending on the 3D representation. Then for each viewpoint, we exploit Depth Anything [64] to compute the monocular depth and convert it into normal. Empirically we find it more robust than direct surface normal estimation [13]. Finally, we project the analytical normal to each view, measure the pixel-wise angular difference, and compute the amount of inliers:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 551, + 173, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 551, + 173, + 562 + ], + "spans": [ + { + "bbox": [ + 69, + 551, + 173, + 562 + ], + "type": "text", + "content": "Geometric consistency " + }, + { + "bbox": [ + 69, + 551, + 173, + 562 + ], + "type": "inline_equation", + "content": "=" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 566, + 296, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 566, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 111, + 566, + 296, + 594 + ], + "type": "interline_equation", + "content": "\\frac {1}{N _ {p}} \\sum_ {p} \\mathbb {1} \\left[ \\operatorname {a r c c o s} \\left(\\mathbf {n} _ {p} ^ {\\text {a n a l}} \\cdot \\mathbf {n} _ {p} ^ {\\text {p r e d}}\\right) < \\delta^ {\\text {n o r m}} \\right] \\tag {1}", + "image_path": "9a63fd7a2b1416d2b3d797a6d863e0919650ff18e256106c5ddbf46b69d4b5ea.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": "Here, we iterate over all pixels " + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^{\\mathrm{anal}}" + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": " indicates that the normal of pixel " + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": " is analytically derived, while " + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^{\\mathrm{pred}}" + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": " suggests it is predicted. " + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": " denotes the number of valid rendered pixels. In practice, we set the threshold " + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\delta^{\\mathrm{norm}}" + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "inline_equation", + "content": "23^{\\circ}" + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": " based on a hold-out validation set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 180, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 180, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 180, + 685 + ], + "type": "text", + "content": "3.2. Semantic Consistency" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Another crucial factor for a 3D generation to be appealing and useful is that the underlying content and semantics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 155 + ], + "type": "text", + "content": "remain consistent and coherent across different viewpoints. Unlike visual appearance, which may be affected by extrinsic factors such as lighting and change across views, the semantics of an asset should be invariant to different perspectives. To quantify the degree to which the semantics change across views, we capitalize on large vision foundation models that encode rich semantics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": "We first extract DINO features for every rendered image " + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^{\\mathrm{DINO}} = f^{\\mathrm{DINO}}(\\mathcal{I}_i)" + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": ". Then, for each 3D point " + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": " on the surface of the asset " + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": ", we project it onto 2D feature maps " + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_i^{\\mathrm{DINO}}\\}" + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": " to retrieve its corresponding features " + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_i^{\\mathrm{DINO}}(\\pi_{v_i}(x))\\}" + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": " is the projection operator. Since a point may be visible from multiple images, we compute the variance of its DINO features. Specifically, we calculate the variance along each feature dimension separately across viewpoints and then take the mean of these variances across all feature dimensions. Finally, we compute the ratio of vertices whose variance is below a certain threshold. The intuition is that if the content and semantics of a 3D asset remain relatively stable across different viewpoints, its DINO features should be consistent. In practice, for computational efficiency, we only consider the vertices of the 3D mesh " + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{vert}}" + }, + { + "bbox": [ + 313, + 156, + 556, + 347 + ], + "type": "text", + "content": " as 3D points:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 355, + 555, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 355, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 315, + 355, + 555, + 366 + ], + "type": "interline_equation", + "content": "\\text {S e m a n t i c c o n s i s t e n c y} = \\tag {2}", + "image_path": "445d6da4374751bff7d02ddc9778e99b6f2f334f5a836873ae62b085291d2965.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 334, + 369, + 552, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 369, + 552, + 397 + ], + "spans": [ + { + "bbox": [ + 334, + 369, + 552, + 397 + ], + "type": "interline_equation", + "content": "\\frac {1}{N _ {\\text {v e r t}}} \\sum_ {x ^ {\\text {v e r t}}} \\mathbb {1} \\left[ \\operatorname {m e a n} \\left(\\operatorname {V a r} \\left(\\left\\{\\mathcal {F} _ {i} ^ {\\mathrm {D I N O}} \\left(\\pi_ {v _ {i}} \\left(x ^ {\\text {v e r t}}\\right) \\right.\\right)\\right) < \\delta^ {\\mathrm {D I N O}} \\right] \\right.", + "image_path": "d734c563c9c76752165fb8fca0b867dcb2c42f3417a48e4adca0b381766324e0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 405, + 554, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 429 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 313, + 405, + 554, + 429 + ], + "type": "inline_equation", + "content": "\\delta^{\\mathrm{DINO}}" + }, + { + "bbox": [ + 313, + 405, + 554, + 429 + ], + "type": "text", + "content": " to the 70th percentile of average DINO variance based on a hold-out validation set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 433, + 444, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 444, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 444, + 445 + ], + "type": "text", + "content": "3.3. Structural Consistency" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 450, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 510 + ], + "type": "text", + "content": "A generated asset that looks plausible from a single view does not necessarily imply that it remains coherent when observed as a whole. The third criterion we are interested in measuring is whether the overall structure of the generated asset is coherent and plausible." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 510, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 629 + ], + "type": "text", + "content": "We frame this assessment as a novel view synthesis (NVS) problem. Our intuition is that if the generated asset is overall plausible, one should be able to synthesize, to some degree, what other views might look like. For instance, given an image of the back of the head, one should be capable of predicting the face, although there might be variations in the size of the eyes or nose compared to the actual generation. On the other hand, if the generation suffers from artifacts or noise, it would be difficult to effectively conduct cross-view prediction, and the discrepancies could be significant." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "With this motivation in mind, given a rendered RGB image " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " as input, we first utilize Stable-Zero123 [1], a diffusion-based NVS model, to predict what an image from another unobserved viewpoint might look like, denoted as " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{i\\rightarrow j} = f^{0123}(\\mathcal{I}_i;v_i,v_j)" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". We then measure the difference between the predicted image " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{i\\rightarrow j}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " and the image rendered from the generated asset " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_j" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". Since we care more about the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13329" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 62, + 519, + 214 + ], + "blocks": [ + { + "bbox": [ + 92, + 62, + 519, + 214 + ], + "lines": [ + { + "bbox": [ + 92, + 62, + 519, + 214 + ], + "spans": [ + { + "bbox": [ + 92, + 62, + 519, + 214 + ], + "type": "image", + "image_path": "59250fd286ab4c24094ae601e98dc57a67321c176ad3c183ad777e2767297bda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 217, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 217, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 555, + 240 + ], + "type": "text", + "content": "Figure 4. Structural consistency evaluates the geometric coherence of the generated 3D assets by comparing rendered views with the predictions from a novel view synthesis model (Zero-123) across various rotations. We utilize DreamSim to assess image similarity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 297, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 297, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 297, + 319 + ], + "type": "text", + "content": "perceptual similarity rather than pixel-wise differences, we employ DreamSim [14], a human-aligned perceptual metric, for this similarity assessment (see Fig. 4 for overview). Finally, we repeat the same operation for multiple views " + }, + { + "bbox": [ + 55, + 247, + 297, + 319 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 247, + 297, + 319 + ], + "type": "text", + "content": " and then take the average. In practice, to increase robustness, we consider different images " + }, + { + "bbox": [ + 55, + 247, + 297, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 55, + 247, + 297, + 319 + ], + "type": "text", + "content": " as input and select the best one:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 324, + 173, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 324, + 173, + 335 + ], + "spans": [ + { + "bbox": [ + 70, + 324, + 173, + 335 + ], + "type": "text", + "content": "Structural consistency " + }, + { + "bbox": [ + 70, + 324, + 173, + 335 + ], + "type": "inline_equation", + "content": "=" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 339, + 296, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 339, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 110, + 339, + 296, + 372 + ], + "type": "interline_equation", + "content": "\\max _ {i} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\left(1 - f ^ {\\text {D r e a m S i m}} \\left(\\mathcal {I} _ {i \\rightarrow j} ^ {\\text {p r e d}}, \\mathcal {I} _ {j}\\right)\\right) \\tag {3}", + "image_path": "39b96dc849c0a0bb2fe94607ccdae7c34bcec651c94ebabeb9cc8765191b0cd1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": "We take as input both the canonical view (rendered at " + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": "), and the rendered image from " + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": ". We predict novel viewpoints at an interval of " + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 168, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 168, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 168, + 431 + ], + "type": "text", + "content": "3.4. Text-3D Alignment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 435, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 507 + ], + "type": "text", + "content": "Another often overlooked aspect of 3D evaluation is the fidelity of the generated 3D asset to the input conditions. For instance, are all the mentioned objects being generated? Do the relationships among the objects align with what the text describes? The alignment issue is critical as it reflects how accurately the 3D generation adheres to the instructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": "To effectively measure the text-3D alignment, we draw inspiration from TIFA [21], an evaluation metric for text-to-image generation, and adapt it for assessing text-to-3D generation. The key idea is to measure the consistency among large language models (LLMs) and multimodal LLMs across multiple views, and use this consistency as a proxy for evaluation. Specifically, we first employ a LLM to generate " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": " multiple-choice question-answer tuples " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{Q_j,C_j,A_j\\}_{j = 1}^M" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": " based on the input text prompt " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "Q_{j}" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": " is a question, " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": " is a set of answer choices, and " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "A_{j}\\in C_{j}" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": " is the gold answer. Then, for each question " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "Q_{j}" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": ", we employ a multimodal LLM to produce an answer based on each input image " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "A_{j}^{v_{i},\\mathrm{VQA}} = f^{\\mathrm{MLLM}}(\\mathcal{I}_{i},Q_{j})" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": ". If the answer from the multimodal LLM " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "A_{j}^{v_{i},\\mathrm{VQA}}" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": " aligns with the answer derived from the LLM " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": ", we conclude that the specific property in question is aligned at such viewpoint " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": ". For example, given a text prompt " + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "inline_equation", + "content": "t = \"\\text{A statue of a dog}\"" + }, + { + "bbox": [ + 55, + 508, + 297, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 247, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 333 + ], + "type": "text", + "content": "the LLM could generate question-answer tuples such as {\"Is there a statue?\", \"Yes\", \"No\", \"Yes\"} and view-dependent alignment is assessed based on whether both the text and the image confirm the presence of a statue (see Fig. 2). To aggregate individual scores, we repeat the same procedure for all questions " + }, + { + "bbox": [ + 313, + 247, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\{Q_j\\}_{j=1}^M" + }, + { + "bbox": [ + 313, + 247, + 555, + 333 + ], + "type": "text", + "content": " and across all viewpoints " + }, + { + "bbox": [ + 313, + 247, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\{v_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 247, + 555, + 333 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 343, + 411, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 343, + 411, + 356 + ], + "spans": [ + { + "bbox": [ + 321, + 343, + 411, + 356 + ], + "type": "text", + "content": "Text-3D alignment =" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 359, + 555, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 359, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 354, + 359, + 555, + 393 + ], + "type": "interline_equation", + "content": "\\frac {1}{M} \\sum_ {j = 1} ^ {M} \\operatorname {A n y} _ {v _ {i}} \\left(\\operatorname {A l l} _ {v \\in \\operatorname {A d j} \\left(v _ {i}\\right)} \\left(\\mathbb {1} \\left[ A _ {j} = = A _ {j} ^ {v, \\mathrm {V Q A}} \\right]\\right)\\right) \\tag {4}", + "image_path": "c21cefba23bfd917a72d99d174c86a9a911cd86b02b8c996fb296776db83774a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 403, + 556, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 488 + ], + "type": "text", + "content": "A question is considered globally aligned if and only if, for some viewpoint " + }, + { + "bbox": [ + 313, + 403, + 556, + 488 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 403, + 556, + 488 + ], + "type": "text", + "content": ", the answers of all its neighboring views are aligned. We find that this method is effective for dealing with occlusions and the noise of multimodal LLM. In practice, we adopt LLaVA-NeXT-7B [27] as our VQA model, and we follow DSG [10] for generating and evaluating our question-answer pairs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 495, + 381, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 381, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 381, + 506 + ], + "type": "text", + "content": "3.5. Aesthetic" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 556, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 717 + ], + "type": "text", + "content": "Finally, given that a generated 3D asset can be viewed from a wide range of angles, it is crucial to ensure that it appears visually appealing across all possible views. In this paper, we exploit two different approaches, with varying computational cost, to compute the aesthetic score " + }, + { + "bbox": [ + 313, + 514, + 556, + 717 + ], + "type": "inline_equation", + "content": "f^{\\text{aesthetic}}(\\{\\mathcal{I}_i\\})" + }, + { + "bbox": [ + 313, + 514, + 556, + 717 + ], + "type": "text", + "content": ". First, following [61], we use GPT-4o [38] to conduct pairwise quality comparisons of 3D assets and compute the ELO score of each model. We improve the prompts and normalize the ELO score to 0-100% (by comparing with 3D generation models in Sec. 4.2). While effective, the closed-source nature and the high computational cost resulting from combinatorial comparison may create barriers to large-scale evaluation. As an alternative, we employ ImageReward [62], an open-source aesthetic estimator, to measure the aesthetic quality of each rendered image " + }, + { + "bbox": [ + 313, + 514, + 556, + 717 + ], + "type": "inline_equation", + "content": "f^{\\text{ImgReward}}(\\mathcal{I}_i)" + }, + { + "bbox": [ + 313, + 514, + 556, + 717 + ], + "type": "text", + "content": " and then compute the average. We compare both implementations in Sec. 4 and discuss their respective strengths." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13330" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 76, + 276, + 163 + ], + "blocks": [ + { + "bbox": [ + 114, + 64, + 187, + 74 + ], + "lines": [ + { + "bbox": [ + 114, + 64, + 187, + 74 + ], + "spans": [ + { + "bbox": [ + 114, + 64, + 187, + 74 + ], + "type": "text", + "content": "Generated 3D asset" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 76, + 276, + 163 + ], + "lines": [ + { + "bbox": [ + 86, + 76, + 276, + 163 + ], + "spans": [ + { + "bbox": [ + 86, + 76, + 276, + 163 + ], + "type": "image", + "image_path": "75d8ec71c8150b0f58c278378c317b16ba80c31467aad19ad40a05596deef841.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 168, + 297, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 297, + 224 + ], + "type": "text", + "content": "Figure 5. 3D inconsistency maps: The proposed 3D metrics, semantic and geometric consistencies, allow fine-grained localization of the artifacts (eg: Janus issue: multiple nose / face, inconsistent hand geometry, arbitrary surface patterns on the back) by fusing / computing the metrics in 3D space." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 235, + 187, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 187, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 187, + 248 + ], + "type": "text", + "content": "3.6. Localizing 3D Artifacts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 254, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 494 + ], + "type": "text", + "content": "Unlike previous works [19, 61], our proposed geometric and semantic consistency measures operate in the 3D space, allowing us to not only classify the generation as an artifact but also to localize the artifacts on the extracted 3D mesh. For the semantic consistency, artifacts can be localized by highlighting outlier vertices—specifically, mesh vertices with DINO feature variance exceeding a certain threshold. Fig. 5 illustrates an example where vertices with high semantic inconsistency (Eq. 2) significantly contribute to the generated 3D inconsistencies like the Janus issue and extraneous geometries. For the geometric consistency, we back-project the pixel-wise cosine distance map between the predicted and rendered geometric normals onto the 3D mesh, highlighting vertices where the mean or maximum cosine distance across different viewpoints exceeds a threshold. Finally, as shown in Fig. 5, fusing the proposed semantic or geometric consistency metrics to the 3D space could enable estimating 3D uncertainty at novel viewpoints. Such consistency cues can be leveraged as a self-supervised refinement metric to enhance the generated geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 137, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 137, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 137, + 516 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 297, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 297, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 297, + 626 + ], + "type": "text", + "content": "In this section, we begin by discussing the curated Eval3D evaluation benchmark. (Sec. 4.1) We then outline the experimental details, covering the targeted 3D generative models, experimental setup, and human annotation process. (Sec. 4.2) Finally, we present our experimental results. (Sec. 4.3) We first assess the agreements between Eval3D and human annotations, showing that Eval3D aligns well with humans on all evaluation dimensions. Then we benchmark text/image-to-3D algorithms with Eval3D and discuss our findings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 636, + 170, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 170, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 170, + 647 + ], + "type": "text", + "content": "4.1. Eval3D Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "To thoroughly evaluate 3D generation algorithms, we curate the Eval3D benchmark. Unlike prior work, which provides only text prompts, Eval3D additionally includes a scene graph, dense human annotations, and optionally a carefully selected SDXL image suitable for image-to-3D generation." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 327, + 70, + 430, + 172 + ], + "blocks": [ + { + "bbox": [ + 336, + 57, + 533, + 68 + ], + "lines": [ + { + "bbox": [ + 336, + 57, + 533, + 68 + ], + "spans": [ + { + "bbox": [ + 336, + 57, + 533, + 68 + ], + "type": "text", + "content": "Text prompt = \"A panda rowing a boat in a pond.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 327, + 70, + 430, + 172 + ], + "lines": [ + { + "bbox": [ + 327, + 70, + 430, + 172 + ], + "spans": [ + { + "bbox": [ + 327, + 70, + 430, + 172 + ], + "type": "image", + "image_path": "751561f0965350bf3589f736afd0c9b71595eed1013552cf742c3ffb98186dd3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 174, + 399, + 182 + ], + "lines": [ + { + "bbox": [ + 357, + 174, + 399, + 182 + ], + "spans": [ + { + "bbox": [ + 357, + 174, + 399, + 182 + ], + "type": "text", + "content": "MV-Dream" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 232, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 298 + ], + "type": "text", + "content": "Figure 6. Evaluation examples from Eval3D. While the generation from MV-Dream appears visually appealing at first glance, both the boat and the paddle are missing (reflected in text-3D alignment). There are also geometric artifacts on the left side of the panda (reflected in geometric consistency). Please refer to the Appendix for detailed breakdown of each metric." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 182, + 425, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 182, + 425, + 192 + ], + "spans": [ + { + "bbox": [ + 328, + 182, + 425, + 192 + ], + "type": "text", + "content": "Structural consistency: 82.22" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 193, + 414, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 193, + 414, + 202 + ], + "spans": [ + { + "bbox": [ + 339, + 193, + 414, + 202 + ], + "type": "text", + "content": "Text-3D alignment: 50" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 203, + 422, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 203, + 422, + 211 + ], + "spans": [ + { + "bbox": [ + 331, + 203, + 422, + 211 + ], + "type": "text", + "content": "Semantic consistency: 76.8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 328, + 212, + 425, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 212, + 425, + 220 + ], + "spans": [ + { + "bbox": [ + 328, + 212, + 425, + 220 + ], + "type": "text", + "content": "Geometric consistency: 82.85" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 340, + 221, + 413, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 221, + 413, + 228 + ], + "spans": [ + { + "bbox": [ + 340, + 221, + 413, + 228 + ], + "type": "text", + "content": "Aesthetic score: 84.33" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 70, + 539, + 172 + ], + "blocks": [ + { + "bbox": [ + 436, + 70, + 539, + 172 + ], + "lines": [ + { + "bbox": [ + 436, + 70, + 539, + 172 + ], + "spans": [ + { + "bbox": [ + 436, + 70, + 539, + 172 + ], + "type": "image", + "image_path": "d11054cb8046dd34b6296a167f1c454568ef9f6e1dad289c431a0f1bac2d9357.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 174, + 506, + 182 + ], + "lines": [ + { + "bbox": [ + 470, + 174, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 470, + 174, + 506, + 182 + ], + "type": "text", + "content": "TextMesh" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 440, + 183, + 536, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 183, + 536, + 192 + ], + "spans": [ + { + "bbox": [ + 440, + 183, + 536, + 192 + ], + "type": "text", + "content": "Structural consistency: 50.12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 449, + 193, + 528, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 449, + 193, + 528, + 201 + ], + "spans": [ + { + "bbox": [ + 449, + 193, + 528, + 201 + ], + "type": "text", + "content": "Text-3D alignment: " + }, + { + "bbox": [ + 449, + 193, + 528, + 201 + ], + "type": "inline_equation", + "content": "50\\%" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 438, + 202, + 538, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 202, + 538, + 211 + ], + "spans": [ + { + "bbox": [ + 438, + 202, + 538, + 211 + ], + "type": "text", + "content": "Semantic consistency: " + }, + { + "bbox": [ + 438, + 202, + 538, + 211 + ], + "type": "inline_equation", + "content": "24.40\\%" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 438, + 212, + 538, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 212, + 538, + 220 + ], + "spans": [ + { + "bbox": [ + 438, + 212, + 538, + 220 + ], + "type": "text", + "content": "Geometric consistency: 81.91%" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 451, + 221, + 524, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 221, + 524, + 228 + ], + "spans": [ + { + "bbox": [ + 451, + 221, + 524, + 228 + ], + "type": "text", + "content": "Aesthetic score: 26.09" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 308, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 392 + ], + "type": "text", + "content": "Prompt source: Eval3D contains 160 text prompts for text-to-3D evaluation. The prompts come from diverse sources, including threestudio [18] and the GPT-generated prompts from prior works [19, 51, 61]. We curate the prompt source such that 80 prompts contain a single object, and 80 prompts contain multiple objects. For image-to-3D evaluation, we use the 20 image-text pairs from Dreamcraft3D [51]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 392, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 464 + ], + "type": "text", + "content": "Scene graphs: To facilitate further investigation, we also provide the scene graph for each prompt. Following prior work [10], the scene graph contains 5 kinds of elements: entity (e.g., \"statue\"), attribute (e.g., colors, materials), relation (e.g., \"on top of\"), action (e.g., \"sit\"), and others (e.g., art style). Please refer to Appendix for more details." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 464, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 559 + ], + "type": "text", + "content": "Annotations: We provide dense expert annotations on all 3D evaluation criteria discussed in Sec. 3 for 160 prompts across six 3D generation models. We note that our annotation is an order of magnitude larger than prior work. For example, [61] only contains pairwise comparison for 3 prompts for each model pair. [19] annotates quality and alignment for 90 prompts, ignoring geometric, semantic, and structural consistency. See Sec. 4.2 and Appendix for more details." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 566, + 436, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 436, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 436, + 578 + ], + "type": "text", + "content": "4.2. Experimental Details" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Target 3D generation models: We evaluate the proposed metrics on six state-of-the-art text-to-3D algorithms (i.e., DreamFusion [41], Magic3D [28], ProlificDreamer [56], TextMesh [52], and MVDream [49]) and two image-to-3D algorithms (i.e., Magic123 [42] and DreamCraft3D [51]). These algorithms represent a diverse set of text/image-to-3D generative models, featuring different 3D representations (NeRF, NeRF + Tetrahedral Mesh, Gaussian Splitting), various optimization strategies (single-stage vs. multi-stage), and different generative (diffusion) priors (single view vs. multi-view). Consequently, the generated assets exhibit a" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13331" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 62, + 531, + 140 + ], + "blocks": [ + { + "bbox": [ + 77, + 62, + 531, + 140 + ], + "lines": [ + { + "bbox": [ + 77, + 62, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 62, + 531, + 140 + ], + "type": "table", + "html": "
ApproachGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
CLIP-Score [20]43.6-64.062.074.1
ImageReward [62]44.3-63.775.479.5
T3-Bench [19]---76.556.7
GPT4V [61]46.9-68.985.672.8
Eval3D (Ours)83.068.069.287.488.7
", + "image_path": "0b33daaf32d735cae59ad75dbc1397c395c79a2bb01c4dfd8b152cf7f1448a00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 142, + 555, + 166 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 555, + 166 + ], + "type": "text", + "content": "Table 1. Pairwise rating agreements between human and evaluation metrics. We measure the probability that the automatic evaluation makes the same decision as human annotators when comparing two 3D assets. Eval3D achieves the best alignment with humans." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "spans": [ + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "text", + "content": "wide range of artifacts, enabling a thorough study of fine-grained 3D evaluation. We leverage the implementations from ThreeStudio [18]. Please refer to Appendix for details. Setup: We render each generated 3D asset from 120 different viewpoints, with cameras positioned at an elevation of " + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "text", + "content": " and an azimuth varying from " + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "text", + "content": " degrees. We render RGB images, surface normals, and opacity maps at a resolution of " + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "text", + "content": ". To compute semantic consistency, we extract the underlying 3D mesh using either marching cubes [31] or DMTet [47]. We upsample DINO features to " + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 177, + 297, + 332 + ], + "type": "text", + "content": " usingFeatUp [15] and then fuse them with the extracted 3D mesh. Due to cost considerations, we use 12 viewpoints for the text-3D alignment metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 334, + 297, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 334, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 334, + 297, + 632 + ], + "type": "text", + "content": "Human annotations: Previous human annotation data either lacked coverage for all 3D evaluation dimensions (e.g., only focuses on aesthetics and alignment, ignoring geometry and structure) or included a very limited number of annotations. To address these issues, we performed dense expert annotations on all 3D assets generated by six text-to-3D models based on different criteria discussed in Sec. 3. We carefully designed the annotation process for each metric to ensure high inter-annotator agreement. Take geometric consistency for example. For each text prompt, we showed annotators videos of RGB and rendered surface normals, placed side-by-side for all six algorithms. Then we asked annotators to rank the generations based on how well the texture aligns with the surface normals. Ties were allowed. We adopted a similar annotation procedure for aesthetics, but showing only the RGB videos. As for semantic and structural consistency, since it is difficult to rank (e.g., whether a two-faced dog is better or worse than a three-legged astronaut), we instead framed it as a yes/no question and asked annotators whether they observed any inconsistencies. Finally, for text-3D alignment, annotators looked at the RGB videos and answered a series of multi-choice questions generated automatically from the text prompt, as discussed in [10]. We direct the readers to the Appendix for more comprehensive details on our annotation process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 637, + 180, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 180, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 180, + 651 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": "Human alignment: An optimal 3D evaluation metric should align closely with human judgment. We assess this alignment by comparing human annotations with ours and baseline metrics. As shown in Tab. 1, Eval3D outperforms all baseline evaluation metrics across all dimensions. Specif" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 177, + 556, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 556, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 556, + 286 + ], + "type": "text", + "content": "ically, our text-3D metric aligns approximately " + }, + { + "bbox": [ + 313, + 177, + 556, + 286 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 313, + 177, + 556, + 286 + ], + "type": "text", + "content": " better with human annotations than GPT-4V [61]. Our geometric consistency metric, which is analogous to GPT-4V's texture-geometry alignment [61], achieves " + }, + { + "bbox": [ + 313, + 177, + 556, + 286 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 313, + 177, + 556, + 286 + ], + "type": "text", + "content": " better alignment with human judgments compared to all other baselines. Importantly, our geometric and semantic consistency metrics are also interpretable and fine-grained, allowing us to localize potential artifacts (see Figs. 5 and 3). We note that we are the first to capture the multi-view semantic consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 289, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 556, + 698 + ], + "type": "text", + "content": "Benchmarking text-to-3D models: Tab. 2 shows the scores (scaled to " + }, + { + "bbox": [ + 313, + 289, + 556, + 698 + ], + "type": "inline_equation", + "content": "0 - 100\\%" + }, + { + "bbox": [ + 313, + 289, + 556, + 698 + ], + "type": "text", + "content": ") of each text-to-3D model under various criteria. MVDream [49] generally performs the best, ranking 1st or 2nd on all metrics, except for geometry consistency. Upon inspecting the MVDream generations, we find that Eval3D successfully captures its weaknesses: despite their high overall quality, MVDream assets suffer from holes and frequently hallucinate details. DreamFusion [41], TextMesh [52], and Magic3D [28] perform poorly across most metrics. They produce low-resolution geometries, which result in low aesthetic scores. Additionally, they do not accurately adhere to the text prompts, achieving only " + }, + { + "bbox": [ + 313, + 289, + 556, + 698 + ], + "type": "inline_equation", + "content": "40\\% - 65\\%" + }, + { + "bbox": [ + 313, + 289, + 556, + 698 + ], + "type": "text", + "content": " in text-3D alignment. Nevertheless, their geometric consistency is comparable to, or better than, that of MVDream. We hypothesize that this is due to their tendency to generate simpler and more convex geometries. ProlificDreamer achieves high scores in aesthetics and text-3D alignment, yet it ranks last or second-to-last in all three geometric metrics. In contrast, previous studies [19, 61] rate ProlificDreamer highly. Upon inspecting these outputs, we observe that Eval3D more effectively highlights the weaknesses of ProlificDreamer. Fig 5 showcases a generation by ProlificDreamer that, despite its aesthetic sharpness and alignment with the text prompt, suffers from geometric, structural, and semantic inconsistencies. Gaussian Splatting ranks the lowest on geometric consistency. This validates [22]'s finding that Gaussian-Splatting-based methods generate extremely noisy geometry. Interestingly, despite the noisy geometries, the generations rank highly in terms of semantic consistency measured by DINOv2 [39] and multi-view structural consistency measured using Zero123 [29]. We show a few qualitative examples as well as their evaluation results in Fig. 6. Please refer to the Appendix for those of the other text-to-3D algorithms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 701, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 556, + 714 + ], + "type": "text", + "content": "Benchmarking image-to-3D models: Unlike previous stud" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13332" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 60, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 63, + 60, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 63, + 60, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 63, + 60, + 545, + 148 + ], + "type": "table", + "html": "
AlgorithmGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
TextMesh [52]83.174.169.618.943.2
DreamFusion [41]88.367.177.624.250.8
Magic3D [28]94.161.579.351.264.2
ProlificDreamer [56]77.159.575.160.676.9
Gaussian Splitting [68]67.984.881.862.366.7
MVDream [49]85.874.381.182.567.7
", + "image_path": "5d13b298abeb85c767c8a316cebf4c8e3dbc7dc7547a6cd2e5237406621caafe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 182, + 548, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 153, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 555, + 175 + ], + "type": "text", + "content": "Table 2. Eval3D results of Text-to-3D models. We benchmark the text-to-3D models discussed in Sec. 4.2 on all the evaluation criteria in Eval3D. For each criterion, the score ranges from 0 to " + }, + { + "bbox": [ + 55, + 153, + 555, + 175 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 153, + 555, + 175 + ], + "type": "text", + "content": ". Higher values indicate better performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 182, + 548, + 248 + ], + "lines": [ + { + "bbox": [ + 60, + 182, + 548, + 248 + ], + "spans": [ + { + "bbox": [ + 60, + 182, + 548, + 248 + ], + "type": "table", + "html": "
AlgorithmInputGeometric Consis. ↑Semantic Consis. ↑Structural Consis. ↑Aesthetics ↑Text-3D Align. ↑
Magic3D [28]97.167.280.325.990.2
MVDream [49]89.776.183.476.880.5
Magic123 [42]✓97.670.782.039.288.4
DreamCraft3D [51]✓97.082.881.369.191.9
", + "image_path": "24c33e6f8609cd7a1c8c0932b97b9ba712d06f37426e48caebc76acda0895106.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "type": "text", + "content": "Table 3. Eval3D results of image-to-3D models. We evaluate Magic123 [42] and DreamCraft3D [51], both of which take images and texts as inputs. We also show the results of two text-to-3D models (Magic3D [28]) and MVDream [49]) using the same text prompts as inputs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 280, + 297, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 297, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 297, + 460 + ], + "type": "text", + "content": "ies [19, 61], we also evaluate state-of-the-art image+text-to-3D generative algorithms (i.e., Magic123[42] and DreamCraft3D[51]). To validate to which degree an input image could help, we compare these algorithms with two text-only 3D generation algorithms, Magic3D [28] and MVDream [49]. As shown in Tab. 3, text+image-to-3D generations better align with the text prompts and perform comparable on geometric and structural consistency metrics. In particular, although Magic123 and Magic3D share similar generation pipelines, Magic123, with the aid of image guidance, achieves improved semantic and structural consistency. We conjecture that the improvement stems from the use of diffusion models to generate novel views from the input image, which serve as priors and help avoid structural inconsistencies like the Janus problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 194, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 194, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 194, + 481 + ], + "type": "text", + "content": "5. Analysis and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 297, + 700 + ], + "type": "text", + "content": "Open-source vs. closed-source evaluation: Compared to previous work [61], which exclusively capitalized on closed-source multimodal LLMs for assessment, we propose a fully open-source solution for 3D evaluation. Additionally, we design Eval3D to be flexible, allowing foundation models and tools to be easily swapped with alternatives possessing similar capabilities. For instance, one could replace the open-source LLaVA model [27] with the closed-source GPT-4o for the text-3D alignment metric, or alternatively, select either GPT-4o or ImageReward [62] for aesthetic measurement. We experimented with these models and found that for text-3D alignment, our human alignment is similar using two models (88.7% using LLaVA, 91.5% using GPT-4o), all substantially outperforming prior metrics. While GPT-4o aligns more closely with human judgments than ImageReward does, the overall ranking of 3D algorithms assessed by the two is similar, with an edit distance of 1. These results suggest that open-source models are effective for evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Limitations: The foundation models used in the paper are" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "text", + "content": "not perfect and may introduce errors. For example, StableZero123 [29] may synthesize incorrect novel-view scenes; DepthAnything [64] may produce coarse predictions; and multimodal LLMs like LLaVA [27] may generate wrong answers. However, as shown in our experiments, Eval3D still aligns better with humans than all available automatic 3D evaluations, while being significantly more interpretable. Furthermore, our proposed framework is designed to support seamless integration of new and improved foundation models. As foundation models continue to improve, we anticipate even stronger alignment with human evaluations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 557, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 557, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 557, + 524 + ], + "type": "text", + "content": "Broader impact: Eval3D can be used to identify and filter out inaccurate and unrealistic 3D assets, and potentially improve existing 3D generation algorithms. It could have a significant impact on the entertainment and robotics industries where high-quality 3D assets are crucial. However, it could also propagate the biases and limitations of foundation models into evaluation metrics. Additionally, it may affect employment within creative fields, as there could be a reduced need for human workers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 529, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 649 + ], + "type": "text", + "content": "Conclusion: We present Eval3D, an interpretable and fine-grained evaluation framework for text/image-to-3D generation algorithms. Eval3D leverages the consistency among various vision and language foundation models to accurately evaluate generated 3D assets along multiple dimensions. It aligns with human judgment better than available alternatives and enables pixel-wise and spatial localization of artifacts in the generated 3D assets. We believe Eval3D can serve as a reliable framework for evaluating, filtering, and potentially improving 3D generation algorithms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgment: The paper was partially supported by a gift from Ai2, an Amazon Science sponsored research grant, and DARPA TIAMAT program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13333" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Stability AI. Stable-zero123: Quality 3d object generation from single images, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 146 + ], + "type": "text", + "content": "[2] Mark Boss, Raphael Braun, V. Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. Nerd: Neural reflectance decomposition from image collections. ICCV, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 296, + 190 + ], + "type": "text", + "content": "[3] Emmanuelle Bourigault and Pauline Bourigault. Mvdiff: Scalable and flexible multi-view diffusion for 3d object reconstruction from single-view. arXiv preprint arXiv:2405.03894, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 192, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 296, + 235 + ], + "type": "text", + "content": "[4] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 302 + ], + "type": "text", + "content": "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 304, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 296, + 356 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "type": "text", + "content": "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22246–22256, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 296, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 467 + ], + "type": "text", + "content": "[8] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 296, + 524 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Vladimir G Kim, Matthew Fisher, Noam Aigerman, Hao Zhang, and Siddhartha Chaudhuri. Decor-gan: 3d shape detailization by conditional refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15740-15749, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "text", + "content": "[10] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian Scene Graph: Improving Reliability in Fine-Grained Evaluation for Text-to-Image Generation. In ICLR, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "text", + "content": "[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 677 + ], + "type": "text", + "content": "[12] Nam Anh Dinh, Haochen Wang, Greg Shakhnarovich, and Rana Hanocka. Loopdraw: a loop-based autoregressive model for shape synthesis and editing. ArXiv, abs/2212.04981, 2022.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[13] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[14] Stephanie Fu*, Netanel Tamir*, Shobhita Sundaram*, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "type": "text", + "content": "[15] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feup: A model-agnostic framework for features at any resolution. In The Twelfth International Conference on Learning Representations, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 176, + 554, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 176, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 176, + 554, + 230 + ], + "type": "text", + "content": "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 233, + 555, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 233, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 233, + 555, + 276 + ], + "type": "text", + "content": "[17] William Gao, April Wang, Gal Metzer, Raymond A Yeh, and Rana Hanocka. Tetgan: A convolutional neural network for tetrahedral mesh generation. arXiv preprint arXiv:2210.05735, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 279, + 555, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 279, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 317, + 279, + 555, + 333 + ], + "type": "text", + "content": "[18] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. three-studio: A unified framework for 3d content generation. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 336, + 555, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 336, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 555, + 380 + ], + "type": "text", + "content": "[19] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 382, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 554, + 415 + ], + "type": "text", + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 418, + 555, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 555, + 482 + ], + "type": "text", + "content": "[21] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 485, + 555, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 485, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 485, + 555, + 529 + ], + "type": "text", + "content": "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 532, + 555, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 555, + 585 + ], + "type": "text", + "content": "[23] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 867-876, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 588, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 588, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 588, + 555, + 632 + ], + "type": "text", + "content": "[24] Brian Karis. Journey to nanite. https://www.highperformancegraphics.org/slides22/Journey_to_Nanite.pdf, 2022. Presented at High Performance Graphics 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 677 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[26] Marian Kleineberg, Matthias Fey, and Frank Weichert. Adversarial generation of continuous implicit shape representations. arXiv preprint arXiv:2002.00349, 2020. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13334" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[27] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, and Li Chunyuan Liu, Ziwei. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "text", + "content": "[28] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "type": "text", + "content": "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ICCV, 2023. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 196, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 249 + ], + "type": "text", + "content": "[30] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 305 + ], + "type": "text", + "content": "[31] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 306, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 296, + 360 + ], + "type": "text", + "content": "[32] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "text", + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 296, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 296, + 469 + ], + "type": "text", + "content": "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 296, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 504 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf. Communications of the ACM, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 505, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 296, + 557 + ], + "type": "text", + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "type": "text", + "content": "[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 201, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 201, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 201, + 614 + ], + "type": "text", + "content": "[38] OpenAI. Hello gpt-4o. 2024. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 615, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 296, + 713 + ], + "type": "text", + "content": "[39] Maxime Oquab, Timoth'ee Darcet, Théo Moutakanni, Huy Q. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russ Howes, PoYao (Bernie) Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Huijiao Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 4, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[40] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 152 + ], + "type": "text", + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. 2023. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 153, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 228 + ], + "type": "text", + "content": "[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 297 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 353 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "type": "text", + "content": "[45] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 555, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 555, + 486 + ], + "type": "text", + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 488, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 555, + 533 + ], + "type": "text", + "content": "[47] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 534, + 555, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 587 + ], + "type": "text", + "content": "[48] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 622 + ], + "type": "text", + "content": "[49] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and X. Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "type": "text", + "content": "[50] Jiao Sun, Deqing Fu, Yushi Hu, Su Wang, Royi Rassin, Da-Cheng Juan, Dana Alon, Charles Herrmann, Sjoerd van Steenkiste, Ranjay Krishna, et al. Dreamsync: Aligning text-to-image generation with image understanding feedback. In Synthetic Data for Computer Vision Workshop @ CVPR 2024, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[51] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13335" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 94 + ], + "type": "text", + "content": "3d generation with bootstrapped diffusion prior. arXiv, 2023. 2, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[52] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[53] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 219 + ], + "type": "text", + "content": "[54] Luke Villapaz. Gta 5 costs 265 million to develop and market, making it the most expensive video game ever produced: Report. International Business Times, 2013. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 296, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 263 + ], + "type": "text", + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 296, + 308 + ], + "type": "text", + "content": "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2024. 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 364 + ], + "type": "text", + "content": "[57] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 296, + 409 + ], + "type": "text", + "content": "[58] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 475 + ], + "type": "text", + "content": "[59] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with geo: On metrics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 521 + ], + "type": "text", + "content": "[60] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "type": "text", + "content": "[61] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. arXiv, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "text", + "content": "[62] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv, 2023. 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "text", + "content": "[63] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 329 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[65] Yue Yang, Fan-Yun Sun, Luca Weihs, Eli VanderBilt, Alvaro Herrasti, Winson Han, Jiajun Wu, Nick Haber, Ranjay Krishna, Lingjie Liu, et al. Holodeck: Language guided generation of 3d embodied ai environments. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024), pages 20-25. IEEE/CVF, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 171 + ], + "type": "text", + "content": "[66] Lior Yariv, Matan Atzmon, and Yaron Lipman. Universal differentiable renderer for implicit neural representations. arXiv, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "text", + "content": "[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 555, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 283 + ], + "type": "text", + "content": "[68] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 555, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 555, + 329 + ], + "type": "text", + "content": "[69] Hongxiang Zhao, Xili Dai, Jianan Wang, Shengbang Tong, Jingyuan Zhang, Weida Wang, Lei Zhang, and Yi Ma. Ctrl123: Consistent novel view synthesis via closed-loop transcription. arXiv preprint arXiv:2403.10953, 2024. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13336" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_content_list.json b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..34b8661c805475c624573e2d2b9bf8a399fdbca4 --- /dev/null +++ b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_content_list.json @@ -0,0 +1,1476 @@ +[ + { + "type": "text", + "text": "Evaluating Model Perception of Color Illusions in Photorealistic Scenes", + "text_level": 1, + "bbox": [ + 140, + 130, + 856, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingjun Mao Zineng Tang Alane Suhr University of California, Berkeley {lingjun, terran, suhr}@berkeley.edu", + "bbox": [ + 290, + 180, + 705, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study the perception of color illusions by vision-language models. Color illusion, where a person's visual system perceives color differently from actual color, is well-studied in human vision. However, it remains underexplored whether vision-language models (VLMs), trained on large-scale human data, exhibit similar perceptual biases when confronted with such color illusions. We propose an automated framework for generating color illusion images, resulting in RCID (Realistic Color Illusion Dataset), a dataset of 19,000 realistic illusion images. Our experiments show that all studied VLMs exhibit perceptual biases similar human vision. Finally, we train a model to distinguish both human perception and actual pixel differences.", + "bbox": [ + 89, + 301, + 485, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 526, + 220, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human visual perception does not always align perfectly with the real world [4, 12, 17]; we are easily influenced by contextual factors such as lighting conditions or surrounding objects and patterns. Figure 1 shows several examples of color illusions, where human judgment of object color and brightness does not align with actual pixel values. Currently, many large vision-language models (VLMs) [9, 22, 23, 39] exhibit behavioral similarities to human visual perception in various aspects, including object recognition [41, 44], scene understanding [24, 36], and spatial reasoning [5]. But because the human visual system behaves idiosyncratically when encountering these illusions, we aim to study whether these models perceive such illusions in the same way humans do. Prior work [34, 43] has explored this question by sourcing illusion images from the internet. However, this approach has a drawback: the number of illusion images on the web is limited, and most images (e.g., $60\\%$ for the IllusionVQA dataset [34]) are well-known examples of these illusions; thus, VLMs have likely memorized humanlike behavioral responses to them. Additionally, the limited scale also restricts the depth and variety of analyses that can be conducted.", + "bbox": [ + 89, + 551, + 482, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, we propose an automatic", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "framework for generating realistic color illusion images, covering many scenarios with illusions that we might genuinely encounter in our real life. We cover three different types of color illusions: contrast (Figure 1a), stripe (Figure 1b), and filter (Figure 1c) illusions. For contrast and stripe illusions, we first generate simple arrangements of shapes exhibiting specific illusions, which are then processed through ControlNet [42] to generate more realistic images. For filter illusions, we directly apply filters to the original MS COCO 2017 images. In the end, we produce a dataset consisting of total 19,000 images, half of which include an illusion, and the other half being a control set without illusions. Each image is paired with a synthetically-generated natural language question, asking either about color differences between specific regions or directly inquiring about the color of a particular object in the image.", + "bbox": [ + 511, + 268, + 906, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We test recent VLMs using our generated illusion images. We find that, after appropriate fine-tuning, they perform well on non-illusion images, but their accuracy drops significantly on illusion images. Some of their responses exhibit visual biases similar to those of humans, while some are entirely incorrect, failing to align with either pixel values or human perception. We also observe that external prompts, such as Chain of Thought (CoT) [40] and few-shot examples, can influence VLM performance to some extent. Furthermore, fine-tuning on illusion images enhances the models' ability to understand these color illusions. Based on this, we propose a foundational baseline method that enables models to capture both human perceptual biases and true pixel values. Finally, we investigate the underlying causes of VLM susceptibility to color illusions, including biases from both the visual system and prior knowledge.", + "bbox": [ + 511, + 517, + 908, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In conclusion, our contributions are as follows: (1) We propose an automated framework for generating realistic illusion images and create a large, realistic dataset of color illusion images, named RCID (Realistic Color Illusion Dataset), to enhance the fairness and accuracy of model testing. (2) We investigate underlying mechanisms of color illusions in VLMs, highlighting the combined influence of the vision system and prior knowledge. We also explore how external prompts and instruction tuning", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7805", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/93d6999a295cd279ef6593dd55a3802f5f6ba1d2704828692f34cbcdc220645c.jpg", + "image_caption": [ + "(a) Contrast Illusion" + ], + "image_footnote": [ + "Q1: Based on pixel values, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A1: III" + ], + "bbox": [ + 96, + 90, + 352, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/83489c9454b9cd8d35735ada0a6d705d43e9b67c7e93f4a90ae96e77062ad520.jpg", + "image_caption": [ + "(b) Stripe Illusion" + ], + "image_footnote": [ + "Q2: According to human perception, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A2: II", + "Q1: Evaluated by pixel values, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A1: III", + "Q2: According to human perception, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A2: II" + ], + "bbox": [ + 370, + 90, + 625, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11fdeb0e7c043d2cca6d53d5d537103304357b4f89412e1fb5859ac1236715bd.jpg", + "image_caption": [ + "(c) Filter Illusion", + "Figure 1. Examples of color illusions. (a) Contrast illusion: The two green squares on the left and right are identical in color, but most people perceive the square on the right as darker due to the brighter background. (b) Stripe illusion: The left and right walls seem to be different colors, but this is an illusion caused by the dark stripes interfering with our perception. (c) Filter illusion: The player's uniform appears red, but in reality, there are no red pixels present." + ], + "image_footnote": [ + "Q1: According to pixel values, what color is the athlete's uniform?", + "A1: ['Gray', 'Yellowish-green', 'Greenish-yellow', 'Green']", + "Q2: Through human eyes, what color is the athlete's uniform?", + "A2: ['Red']" + ], + "bbox": [ + 645, + 90, + 900, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "impact the models' performance on these illusions. (3) We propose a simple training method that enables models to understand human perception while also recognizing the actual pixel values. Our code and data are released under an open-source license at the following URL: https://github.com/mao1207/RCID.", + "bbox": [ + 89, + 420, + 483, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 526, + 232, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual illusion is a phenomenon where human perception differs from the actual environment [4, 12, 17], often triggered by specific visual stimuli [7]. These illusions offer valuable insights into human perception [16, 31] and have been widely used in artworks. [13, 25]. Many studies [28, 29] have attempted to explore the causes of visual illusions, with the bottom-up and top-down theories being the most widely accepted explanations. Proponents [33] of the bottom-up theory argue that low-level visual signals, such as contrast, brightness, edges, and motion, are correlated with human perception of illusions. However, supporters of the top-down theory contend that prior knowledge is also a major factor in these illusions. Evidence shows that people in industrialized societies are more susceptible to the Müller-Lyer illusion due to being misled by implied arrow meanings, and individuals from language communities with fewer color terms are less affected by color illusions [1, 30].", + "bbox": [ + 88, + 551, + 482, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by illusion phenomena in human visual perception, recent research has begun to explore whether models are also susceptible to such illusions. Early research [14, 15] suggests that traditional vision models, such as CNNs, are influenced by visual illusions in a way similar to humans. More recent studies [34, 43] have shown", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that VLMs also exhibit similar perceptual biases. However, images on which these models have been tested are either sourced from the internet or manual compositions of simple shapes (e.g., two squares of different colors), making it easy for the VLM to infer results based on their knowledge rather than actual observation. Additionally, the number of these images is quite limited. To address these limitations, we develop a large-scale dataset that embeds visual illusions into realistic scenes, enabling robust evaluation in complex, real-world contexts. Our study fills a gap by investigating the mechanisms behind VLMs' susceptibility to illusions.", + "bbox": [ + 511, + 420, + 906, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Illusion Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 599, + 784, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The construction of our dataset involves three steps (Figure 2): (1) Image Generation. For contrast and stripe illusions, we use procedural code to generate simple illusion images, which are then processed by ControlNet to create realistic illusion images. For filter illusions, we directly apply contrasting color filters to the original images. Each type of illusion also includes a corresponding control group without any illusions for comparison. (2) Question Generation. We use GPT-4o to generate image-specific questions that are designed to evaluate the model's understanding of the illusion. (3) Human Feedback. We collect human participants' feedback on these images and adjust the original classification of \"illusion\" and \"non-illusion\" based on whether participants are deceived.", + "bbox": [ + 511, + 626, + 906, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Contrast and Stripe Illusion Image Generation", + "text_level": 1, + "bbox": [ + 511, + 847, + 906, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ControlNet Training Stage: We begin by training a ControlNet to map from simple illusion images to photoreal-", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7806", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0d37c77073981aa85fc1557a1c734a0403527d23a35aac9b5febcd6cd5f0c6cd.jpg", + "image_caption": [ + "Figure 2. Process for generating our dataset." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 625, + 373 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/11b47c1997316f14a1763f4f25a9c1a97c24261663e3a41841ce5955f18cfa6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 90, + 861, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "istic images with illusions. We train using modified image-caption pairs $(I, T)$ from the MS COCO 2017 dataset. First, we quantize each original image $I$ into a $10 \\times 10$ grid $G$ , where each cell $G_{x,y}$ represents the average color of its corresponding region:", + "bbox": [ + 89, + 414, + 483, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG _ {x, y} = \\frac {1}{\\left| R _ {x , y} \\right|} \\sum_ {p \\in R _ {x, y}} I (p)\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 498, + 370, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $R_{x,y}$ denotes the set of pixels in the $(x,y)$ -th grid cell, and $I(p)$ is the color value of pixel $p$ in the original image.", + "bbox": [ + 89, + 536, + 482, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, Gaussian noise is added to the quantized image $x_0 = G$ , producing a noisy image $x_{t}$ at time step $t$ . The model, conditioned on the text prompt $T$ and color distribution $G$ , is trained to predict the added noise $\\epsilon_{\\theta}$ by minimizing the expected squared error:", + "bbox": [ + 89, + 566, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {x _ {0}, t, C, \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, t, T, G\\right) \\| _ {2} ^ {2} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 654, + 426, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through this loss, the model learns to progressively refine $x_{t}$ into a realistic image during inference, guided by color distribution and the text prompt.", + "bbox": [ + 89, + 674, + 482, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Procedural Generation Stage: We use a procedural function $f$ to generate simple illusion images, denoted as $S$ , which consist of basic shapes and colors, on which we will condition the trained ControlNet model.", + "bbox": [ + 89, + 719, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For Contrast Illusions, we randomly select two base colors from RGB space: $C_b$ for the background and $C_f$ for the foreground (squares). For a color $C = (r, g, b)$ , we use the function $p$ to adjust its brightness by $p(C, \\mu) = (r \\cdot \\mu, g \\cdot \\mu, b \\cdot \\mu)$ , where $\\mu$ scales each RGB component. The background is divided into a darker region $p(C_b, \\mu_{b1})$ and a brighter region $p(C_b, \\mu_{b2})$ , where $\\mu_{b1} < 1$ and $\\mu_{b2} > 1$ . Additionally, two foreground squares, colored $p(C_f, \\mu_{f1})$", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $p(C_f, \\mu_{f2})$ , are symmetrically placed at random positions $(x_1, y_1)$ and $(x_2, y_2)$ on the background.", + "bbox": [ + 511, + 414, + 903, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If human perception differs from the actual pixel values, the image is classified as an illusion; otherwise, it belongs to the control group (no illusion). We observe that a darker background enhances the perceived brightness of the foreground color, with this effect being strongest when $C_b$ and $C_f$ are similar and weakening as the contrast increases. Section 4.2.3 validates this conclusion. Based on this principle, we generate theoretical illusion and non-illusion images, which are later adjusted through human validation. To increase image diversity, we apply random noise $\\eta$ to introduce irregularities at the square edges or make the boundaries less smooth. The generated image $S_i$ for the contrast illusion can be represented as:", + "bbox": [ + 511, + 445, + 906, + 641 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} S = f \\left(p \\left(C _ {b}, \\mu_ {b 1}\\right), p \\left(C _ {b}, \\mu_ {b 2}\\right), p \\left(C _ {f}, \\mu_ {f 1}\\right), p \\left(C _ {f}, \\mu_ {f 2}\\right), \\right. \\\\ \\left(x _ {1}, y _ {1}\\right), \\left(x _ {2}, y _ {2}\\right), \\eta) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 660, + 877, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, for Stripe Illusions, we randomly generate the background color $C_b$ and the stripe color $C_s$ . The colored stripes alternate with black stripes, and their colors are adjusted to $p(C_s, \\mu_{s1})$ and $p(C_s, \\mu_{s2})$ . Based on an initial estimate of human perception of $\\mu_{s1}$ and $\\mu_{s2}$ , the images are classified as either illusion or non-illusion. The generated image $S$ for the stripe illusion is represented as:", + "bbox": [ + 511, + 702, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS = f \\left(C _ {b}, p \\left(C _ {s}, \\mu_ {s 1}\\right), p \\left(C _ {s}, \\mu_ {s 2}\\right), \\theta , N, \\eta\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 820, + 841, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here the stripe direction $\\theta$ (horizontal, vertical, or diagonal) and the number of stripes $N$ are randomly selected. Random noise $\\eta$ is applied to create irregularities, such as curved stripe arrangements or slight misalignments.", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7807", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2b1589ddc6a1a224083e9820436a3e05b1b5aab15edecaaa22abc63b166d0f75.jpg", + "image_caption": [ + "Figure 3. Data statistics of RCID (Realistic Color Illusion Dataset)." + ], + "image_footnote": [], + "bbox": [ + 135, + 92, + 395, + 327 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9777cdd4936f9d48191a56a9bfdb9601251f91584ffe8331bd541006a2e95c69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 93, + 656, + 325 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ad33710260e74fbb7719319735fc58980bd3a1623c78febbb896824569eb0bb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 95, + 854, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ControlNet Generating Stage: We pair each simple illusion image $S$ with a randomly selected caption $T$ from the MS COCO 2017 dataset, which together serve as input to our model. Starting from the quantized image $x_0 = S$ , the model recursively denoises over multiple time steps following a diffusion-based approach, gradually refining the image to produce a realistic output $I'$ that retains the intended illusion effects. The final image $I'$ is obtained by summing the contributions from all time steps $T$ :", + "bbox": [ + 88, + 367, + 483, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {\\prime} = \\sum_ {t = 0} ^ {T} \\left(x _ {t} + \\epsilon_ {\\theta} \\left(x _ {t}, t, C, S\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 513, + 382, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Filter Illusion Image Generation", + "text_level": 1, + "bbox": [ + 89, + 559, + 375, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate filter illusion images, we first select images from MS COCO 2017 that predominantly contain a specific color $C_{\\mathrm{d}}$ (e.g. red, blue, yellow). We then apply a contrasting color filter $C_{\\mathrm{f}}$ which overlays the original image in the HSV (Hue, Saturation, Value) color space to suppress $C_{\\mathrm{d}}$ by shifting the hue $H$ of filter and ensuring no pixels fall within $C_{\\mathrm{d}}$ 's HSV range. Despite this, the filtered regions may still visually evoke the original color.", + "bbox": [ + 89, + 580, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Question Generation", + "text_level": 1, + "bbox": [ + 89, + 712, + 287, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each image $I'$ , we use GPT-4o to generate an image-grounded question $\\mathcal{Q}$ that asks human respondents to compare colors between two regions or identify the color of a specific object. For example, in Figure 1a, the generated questions refer to the walls on the left and right of the baseball player. We then prefix the question with two distinct prompts: one focused on \"human perception\" (e.g., \"Based on human perception\", \"Through human eye\") and the other on \"pixel values\", each associated with a different answer depending on whether the image contains an illusion. For contrast and stripe illusions, answer options are appended", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the question. Examples are shown in Figure 1.", + "bbox": [ + 511, + 367, + 836, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Human Validation and Selection", + "text_level": 1, + "bbox": [ + 511, + 395, + 797, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure that our generated images truly contain illusions, we conduct a survey on Prolific with a total of 241 participants. Each participant is shown 400 images, including both illusion and non-illusion examples. We use respondents' answers as the human perception ground truth. For every image in our dataset, we assign 5 participants. If more than 3 participants are deceived by the illusion, the image is classified as an illusion image. Conversely, if all 5 participants provide answers consistent with the pixel values, the image is classified as non-illusion and added to the control group. On average, we pay $21 USD per hour.", + "bbox": [ + 511, + 417, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Dataset Details", + "text_level": 1, + "bbox": [ + 511, + 598, + 665, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Realistic Color Illusion Dataset (RCID, Figure 3) contains 19,000 images, comprising 8,000 contrast, 9,000 stripe, and 2,000 filter images. We divide our dataset into 9,500 images for training, 4,750 for development, and the remaining 4,750 for the test set. Each illusion type has distinct structures and subcategories.", + "bbox": [ + 511, + 621, + 905, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Evaluating Vision-Language Models", + "text_level": 1, + "bbox": [ + 511, + 729, + 841, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 755, + 702, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this experiment, we evaluate different models' performance on color illusions and explore the impact of external prompts and instruction-tuning on them. We also provide a foundational baseline model based on LLaVA-1.5 (7B) for understanding color illusions.", + "bbox": [ + 511, + 777, + 905, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Language Models. We test recent models with our generated illusion dataset to examine whether VLMs exhibit perceptual responses similar to humans. To ensure", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7808", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8945c71c48938d8c53e3b6a9d27721ab28facce79191c992e7a121126b1f2dbc.jpg", + "image_caption": [ + "Figure 4. This figure shows the proportion of different model responses across three types of illusions (Contrast, Filter, and Stripe) on our development set. For non-illusion images, we report the proportions of \"Accurate\" and \"Wrong\" responses. For illusion images, we categorize responses as \"No Illusion\" (consistent with pixel values), \"Human like,\" and \"N/A.\" Each image is evaluated using two types of prompts: one based on pixel values and the other based on human perception." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 861, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that their performance on illusion images reflects actual susceptibility to illusions rather than challenges in identifying specified regions and comparing color brightness [11, 37], we focus on open-source models, and fine-tune them on our generated non-illusion simple images and a small set of realistic images to enhance their ability in answering questions about relative colors of different regions in an image. During training, we randomly apply prompts such as 'based on human perception,' 'based on pixel values,' or provide no specific guidance prompt at all. Our models tested include LLaVA [22], InstructBLIP [9], CogVLM [39], InternVL2 [6], Qwen-VL2 [2], OFA [38] and Unified-IO [26]. During inference for contrast and stripe illusions, questions are presented in a multiple-choice format, while filter illusion questions require open-ended responses about the color of specific objects in the image, as illustrated in Figure 1. We also include experiments with closed-source models like GPT-4o and Gemini in the appendix. While their overall color comparison ability is weak, there is still a significant drop in performance on illusion images.", + "bbox": [ + 91, + 417, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For images containing illusions, we use the same metrics as in GVIL [43]. We track the Human-like rate, which measures alignment between human perception and VLM responses by calculating the percentage of cases where the model's responses match human answers. When the prompt explicitly includes \"Based on pixel values,\" this can also be referred to as Deception Rate. The No-Illusion rate indicates the proportion of examples where the model's responses align with actual pixel values. If the model's answer aligns with neither the human response nor the pixel values, we categorize those instances as Not Applicable", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$(N / A)$ . For images without illusions, where human perception matches pixel values, we classify model responses as \"accurate\" or \"inaccurate\".", + "bbox": [ + 511, + 417, + 906, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 470, + 709, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate a range of open-source vision-language models using generated illusion and non-illusion images in our development set, with questions explicitly asking for color judgments 'Based on pixel values' or 'Based on human perception'. The results show that, after fine-tuning on non-illusion images, these models achieve high accuracy $(>75\\%)$ on non-illusion images, while their accuracy on illusion images is significantly lower. We find that explicitly querying for color judgments 'Based on pixel values' or 'Based on human perception' does not lead to significant changes in model performance. In addition, models are likely to produce responses that are completely inaccurate, matching neither pixel values nor human perception. This suggests that while the models are misled by color illusions to some extent, they still struggle to fully model human perception. Detailed results can be found in Figure 4.", + "bbox": [ + 509, + 492, + 906, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inter-Annotator Agreement. We refine our final set of illusion images through human validation, retaining only those that deceive at least 3 out of 5 human participants. Before this selection, the average Fleiss' kappa score among the five respondents on the illusion images is 0.648, which increases to 0.806 after filtering. $21.6\\%$ of the selected images deceive all five participants. We use this stricter subset of illusion images as a challenging test set for our models. The results show a slight increase $(< 5\\%)$ in the proportion of Human-Like responses by the VLMs across all three types of illusions, but overall, the difference remains", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7809", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8622480d8734251f951485471d565e8d726bb7a4498f067dba62334aa2997831.jpg", + "image_caption": [ + "Figure 5. Deception rates of humans and VLMs across different structural patterns." + ], + "image_footnote": [], + "bbox": [ + 137, + 89, + 459, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db77b3c389d0b6826f4cde811605a978ee0c6d0bc2655b118c8b745f8a372b19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 857, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "minimal, indicating that our filtering criterion (more than 3 humans are deceived) is robust.", + "bbox": [ + 89, + 227, + 482, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generated Image Realism. As some generated illusion images appear realistic while others still exhibit an \"AI-looking\" appearance, we assess LLaVA-7B and two human participants on three 500-image sets with varying realism levels. Our results show no significant correlation between realism and accuracy, with $r = 0.084$ for human participants and $r = 0.091$ for the model.", + "bbox": [ + 88, + 257, + 482, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Size. To investigate the effect of model size, we evaluate the performance of two groups of models, OFA and Unified-IO. These groups consist of models of varying sizes but share the same overall architecture. We conduct the experiment on contrast illusions, and explicitly emphasize \"Based on pixel values\" in the questions. Our result shows, with the model size increases, the proportion of human-like responses also increases, while the proportion of responses consistent with actual pixel values decreases. We present OFA's performance in Figure 6 and provide Unified-IO's performance details in the appendix.", + "bbox": [ + 88, + 363, + 483, + 530 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/216538a85f9186ed5fe6dfc5a8cc2d33237d1a0d21d82515e2e5e2f2846ba7ca.jpg", + "image_caption": [ + "Figure 6. Proportions of 'No Illusion,' 'Human Like,' and 'N/A' responses for OFA models of different sizes on contrast illusions." + ], + "image_footnote": [], + "bbox": [ + 138, + 542, + 431, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Illusion Design. We explore a range of visual factors that may influence the strength of color illusions and compare whether these factors affect human perception and VLMs in the same way. We focus on three potential influencing factors: the orientation of the illusion, the contrast between foreground and background colors (for contrast illusions only), and the number of stripes (for filter illusions only). Overall, our findings indicate that these factors significantly impact the strength of the illusion. For example, altering certain factors, such as increasing the color contrast between the foreground and background, can turn a non-illusion image into an illusion image. And these effects are", + "bbox": [ + 88, + 719, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "consistent across both humans and VLMs. In these experiments, we use LLaVA-1.5 (7B) as our tested model.", + "bbox": [ + 511, + 227, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Illusion orientation. We randomly generate a set of images with various structural arrangements to analyze how the location of the illusion affects humans and VLMs. For contrast illusions, the images feature left-right and up-down arrangements, while for stripe illusions, they include horizontal, vertical, and diagonal stripes (Figure 3). We evaluate the error rates of both humans and VLMs across these structures. As illustrated in Figure 5, the results show that for contrast illusions, both humans and VLMs are more frequently deceived by left-right arrangements, while for stripe illusions, they are more often misled by horizontal and diagonal stripes. We find that humans can quickly and accurately perceive color differences between left and right, but are slightly less sensitive to top-bottom differences.", + "bbox": [ + 511, + 258, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Color distance between the foreground and background. For contrast illusions, we calculate the correlation between the color distance in RGB space between the foreground and background, and the illusion's intensity. Figure 7a shows that as the foreground and background colors become more similar, the contrast illusion becomes most pronounced for both humans and VLMs. Conversely, when the foreground and background colors are distinctly different, humans and VLMs are better at distinguishing them.", + "bbox": [ + 511, + 473, + 908, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image spatial frequency. For stripe illusions, we examine the effect of image spatial frequency on the strength of the illusion. After converting the image to the frequency domain using the Fourier transform, we compute spatial frequency as the ratio of high-frequency energy to the total image energy [20]. Our findings suggest that the stripe illusion becomes stronger for both humans and models as image frequency increases, which typically corresponds to a higher number of stripes, as shown in Figure 7b.", + "bbox": [ + 511, + 611, + 908, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompting Methods. We explore whether external prompts could alter VLM responses. In contrast to our main experiments shown in Figure 4, here we aim to identify the model's bias towards predicting answers which are human-like or based on pixel values without specifying this in the question posed to the model. We consider several prompting methods: a simple prompt that only contains questions about color comparisons; chain of thought (CoT), where the model is prompted to first consider if any factors in the image might affect color perception; emphasizing the illusion", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7810", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aeeb23cc6577ca98251b59e994562fc64513269902c938a70aaf62b3c139331b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Illusion CategoryContrast IllusionFilter IllusionStripe Illusion
PromptsPixel-BasedHuman-BasedPixel-BasedHuman-BasedPixel-BasedHuman-Based
LLaVA-1.5 (7B)35.4 (41.2)32.4 (37.4)20.0 (67.8)19.8 (64.6)30.3 (35.2)28.3 (40.1)
Mixed Training (Ours)83.9 (6.9)8.1 (79.0)68.0 (24.3)12.1 (82.3)84.5 (9.8)15.4 (76.7)
", + "bbox": [ + 91, + 88, + 906, + 170 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Model's improved understanding of illusions with mixed training. Two prompts are used: \"based on pixel value\" and \"based on human perception.\" The values represent the proportion of no-illusion responses, with parentheses indicating Human-like rates.", + "bbox": [ + 89, + 179, + 906, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50a24b23e665a3f59d4888f96941c609bc5b81745d38cc24e1de86d31aa4853a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 215, + 413, + 369 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9e627f4c7f41615f5c80bb43f4e4dfdfd71a636cdf7147c7b83b0cecd41e091.jpg", + "image_caption": [ + "(a) Influence of background–foreground color difference on intensity of contrast illusion." + ], + "image_footnote": [], + "bbox": [ + 151, + 398, + 419, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08432bf5060508bbda552ffa5bc21c7b36bd3179fbf9481d5c08d7fe12a1e89d.jpg", + "image_caption": [ + "(b) Influence of image spatial frequency on intensity of stripe illusions." + ], + "image_footnote": [], + "bbox": [ + 160, + 516, + 236, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45078f9bf8f90dae33064fd5dcb8abbf06862f265d760cdf5504028573e30514.jpg", + "image_caption": [ + "Figure 7. Error rates of humans and LLaVA across different conditions in color illusions." + ], + "image_footnote": [], + "bbox": [ + 251, + 516, + 326, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9b587f6424873e68977342d702bc6f0b3a91229aa9c5b68bd0b4e8b6c9b3719.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 516, + 418, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by directly informing the model that the image contains an illusion without specifying its type; and providing few-shot examples, using no-illusion (NI) or human-like (HL) answers as 3-shot examples. Because we use the LLaVA-7B model here, which doesn't support multi-image input, we combine the few-shot examples and the query image into a single composite image before feeding it to the model [22].", + "bbox": [ + 88, + 628, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 8 show that external prompts have minimal effect on filter illusions. But for contrast and stripe illusions, the effect is more pronounced, particularly when no-illusion answers are used as few-shot examples, increasing the proportion of no-illusion responses by approximately $25\\%$ . However, using human-like responses as few-shot examples does not effectively enhance the model's reproduction of human illusion perception, the proportion of human-like responses even decreases slightly for contrast illusions. Additionally, both direct emphasis and CoT prompts partially increase the proportion of no-illusion responses, al", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec8c8951fe5d5517b09e1692cce15ed61c2ce4eb56123f0cb784cff0920dbac5.jpg", + "image_caption": [ + "Figure 8. Effect of different prompts and fine-tuning on VLM response preferences across illusion Types. \"NI\" denotes datasets using no-illusion examples, while \"HL\" represents datasets with human-like examples." + ], + "image_footnote": [], + "bbox": [ + 545, + 220, + 870, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "though the effect of direct emphasis is relatively weak.", + "bbox": [ + 511, + 439, + 872, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-Tuning. First, we fine-tune two separate LLaVA-7B models on illusion images, with one model trained with responses according to pixel values (NI), and the other trained on human-like responses (HL). The bottom two rows of Figure 8 show a significant effect on the test set, particularly when fine-tuning on no-illusion responses. For both contrast and stripe illusions, fine-tuning on no-illusion responses increases the proportion of no-illusion answers to over $90\\%$ . In contrast, fine-tuning on human-like responses has a weaker effect, suggesting that even with human-like responses, the model struggles to fully grasp human perception in the context of illusion images.", + "bbox": [ + 511, + 455, + 906, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we experiment with a mixed training approach that combines both types of responses. We train on question-answer pairs for both types of responses, where questions asking \"Based on pixel values\" are paired with pixel-value responses, those asking \"Based on human perception\" are paired with a human-like response. Table 1 shows the influence of this mixed training on the test set, compared to the base model. These results confirm that training model can effectively distinguish between pixel-based and human perception cues, yielding responses aligned with user's query.", + "bbox": [ + 511, + 638, + 905, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Why Color Illusions Affect VLMs", + "text_level": 1, + "bbox": [ + 511, + 806, + 821, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Perceptual Biases from the Visual System", + "text_level": 1, + "bbox": [ + 511, + 830, + 864, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate whether VLMs' perceptual biases originate from their visual components, we test a range of purely visual models, including ResNet [18], ViT [10], VGG [35],", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7811", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06e5b4a86c0a5c9843ac7089a1fb783e7a949c705d731fa75ad0aa03f0b8fed5.jpg", + "image_caption": [ + "(a) Performance of purely vision models on non-illusion and illusion images." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 277, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f4753d4e6f317f040583acd4ada89f5305dfa03bbd7bd8a8bd76a11068c7579.jpg", + "image_caption": [ + "(b) Accuracy of ViT on non-illusion and illusion images over each training epochs." + ], + "image_footnote": [], + "bbox": [ + 295, + 93, + 478, + 184 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MobileNet [19] and Large Vision Model (LVM) [3]. We evaluate these models on both simple, rectangle-based contrast illusion images and generated realistic illusion images. For simple illusion images, we generate backgrounds with varying brightness on both sides and place a rectangle on each side. The vision model is trained as a three-way classifier on 6,000 non-illusion images to predict which rectangle appears darker, or if their colors are identical, stopping when the loss shows no significant decrease. It is then tested on 1,000 non-illusion and 1,000 illusion images. The results are shown in Figure 9a. For realistic illusion images, we train the model on the training set of our generated dataset and evaluate it on the test set, shown in the appendix. While overall accuracy (the proportion of classifications aligning with actual pixel values) on realistic illusion images drops compared to simple rectangle-based images, in both cases, accuracy on illusion images remains significantly lower than on non-illusion images. We also observe that during training, accuracy on non-illusion images quickly improves, while accuracy on illusion images fluctuates significantly. We use ViT as an example in Figure 9b. These results suggest that even purely vision-based models can be influenced by color illusions.", + "bbox": [ + 89, + 258, + 483, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Perceptual Biases from Prior Knowledge", + "text_level": 1, + "bbox": [ + 89, + 622, + 439, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Language. We test the multilingual vision-language model PALO [32] on contrast illusions, posing questions in English, Chinese, and Russian to analyze error rates on images containing target regions across several colors. We hypothesize that colors with more frequent textual modifiers (e.g., dark blue, sky blue) lead to responses aligned with pixel values rather than human perception. To validate this, we first analyze color descriptor diversity in the training data. For example, in English, $50.7\\%$ of red instances include modifiers like dark red or bright red, while $49.3\\%$ use red alone. Figure 10 shows the correlation between color descriptor diversity and pixel-value accuracy for English, Chinese, and Russian. We find that richer descriptions improve model sensitivity to that color, reducing susceptibility to color illusions in that color range.", + "bbox": [ + 89, + 645, + 482, + 872 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b7269aa4b36779b053b034e63ea25ae8a9124a0ba10e2a1700acc077a3257a0.jpg", + "image_caption": [ + "Figure 9. Perceptual biases in purely vision models.", + "Figure 10. Comparison of color term accuracy and frequency across languages for the vision-language model. Each shape and color represents a specific color term (e.g., blue, red, gray)." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Commonsense Knowledge. Humans are often deceived by visual illusions due to our prior knowledge. For instance, research suggests that in filter illusions, humans perceive colors that aren't actually present due to a phenomenon called chromatic adaptation [27]. This allows us to use surrounding context to infer the presence of a filter, mentally \"removing\" it to perceive the original color. One example is that we always recognize a piece of white paper as white, regardless of the lighting conditions.", + "bbox": [ + 511, + 232, + 906, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To examine whether commonsense knowledge similarly influences VLMs' susceptibility to illusions, we categorized filter illusion images into three groups: single objects, multiple-object compositions, and complex real-world scenes, then analyzed the deception rates for both humans and VLMs across these groups. On non-illusion images, VLMs perform well across all three categories, suggesting that scene complexity does not affect their basic color recognition ability. However, for illusion images, our results in Figure 11 show that both humans and VLMs are less likely to be deceived by single-object images, while complex scenes lead to the highest deception rates, underscoring the role of prior knowledge in visual illusions.", + "bbox": [ + 511, + 369, + 908, + 566 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3251dc85eda73f9872f1c406931bba288ce32be8ba757a6cc8b12546a6151c9.jpg", + "image_caption": [ + "Figure 11. Deception rates of VLMs and human participants across different types of filter illusions." + ], + "image_footnote": [], + "bbox": [ + 550, + 578, + 867, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments show that current VLMs, influenced by their visual systems and prior knowledge, can exhibit visual illusions similar to humans. This raises a key question: as models become more human-like, might they also inherit human perceptual biases? In fields like biomedicine [8, 21], such biases could result in flawed visual judgments, while in human-centric tasks like image generation, understanding human perception can benefit. Therefore, VLM behavior should be tailored to application needs, with careful consideration of human-like perceptual biases.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1In Russian, blue has 100% descriptor diversity due to distinct terms for light blue (goluboj) and deep blue (sinij).", + "bbox": [ + 89, + 875, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7812", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adam Alter. Drunk tank pink: And other unexpected forces that shape how we think, feel, and behave. 2014. 2", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023.5", + "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. CVPR, 2024. 8", + "[4] Claus-Christian Carbon. Understanding human perception by human-made illusions. Frontiers in human neuroscience, 2014. 1, 2", + "[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: Endowing vision-language models with spatial reasoning capabilities. CVPR, 2024. 1", + "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. CVPR, 2024. 5", + "[7] Fan L Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C Aoki, Jin Hirano, and Yukiyasu Kamitani. Reconstructing visual illusory experiences from human brain activity. Science Advances, 2023. 2", + "[8] Hejie Cui, Lingjun Mao, Xin Liang, Jieyu Zhang, Hui Ren, Quanzheng Li, Xiang Li, and Carl Yang. Biomedical visual instruction tuning with clinician preference alignment. NeurIPS, 2024. 8", + "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 1, 5", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7", + "[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. BLINK: Multimodal large language models can see but not perceive. ECCV, 2024. 5", + "[12] Maurizio Gentilucci, Sergio Chieffi, Elena Daprati, M Cristina Saetti, and Ivan Toni. Visual illusion and action. Neuropsychologia, 1996. 1, 2", + "[13] Ernst Hans Gombrich. Art and illusion: A study in the psychology of pictorial representation-millennium edition. 2023. 2", + "[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. CVPR, 2019. 2" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmio, and Jesus Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 2020. 2", + "[16] Richard Langton Gregory. Perceptual illusions and brain models. Proceedings of the Royal Society of London. Series B. Biological Sciences, 1968. 2", + "[17] Richard L Gregory. Visual illusions. Scientific American, 1968. 1, 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016. 7", + "[19] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8", + "[20] Yan Ke, Xiaou Tang, and Feng Jing. The design of high-level features for photo quality assessment. CVPR, 2006. 6", + "[21] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. NeurIPS, 2023. 8", + "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeuRIPS, 2023. 1, 5, 7", + "[23] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. CVPR, 2024. 1", + "[24] Sichao Liu, Jianjing Zhang, Robert X. Gao, Xi Vincent Wang, and Lihui Wang. Vision-language model-driven scene understanding and robotic object manipulation. IEEE CASE, 2024. 1", + "[25] Margaret S Livingstone. Art, illusion and the visual system. Scientific American, 1988. 2", + "[26] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-IO 2: Scaling autoregressive multimodal models with vision, language, audio, and action. CVPR, 2023. 5", + "[27] D. L. MacAdam. Chromatic adaptation*. J. Opt. Soc. Am., 1956. 8", + "[28] O Maddaluno, A Facchin, D Zavagno, N Bolognini, R Daini, et al. Top-down and bottom-up neuromodulation over two different visual illusions. Atti del 40th European Conference on Visual Perception, 2018. 2", + "[29] Glenn E Meyer and Susan Petry. Top-down and bottom-up: The illusory contour as a microcosm of issues in perception. 1987. 2", + "[30] William L Phillips. Cross-cultural differences in visual perception of color, illusions, depth, and pictures. *Cross-Cultural Psychology: Contemporary Themes and Perspectives*, 2019. 2", + "[31] Jiang Qiu, Hong Li, Qinglin Zhang, Qiang Liu, and Fenghua Zhang. The muller-lyer illusion seen by the brain: An event-related brain potentials study. Biological psychology, 2008. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Hanoona Rasheed, Muhammad Maaz, Abdelrahman Shaker, Salman Khan, Hisham Cholakal, Rao M. Anwer, Tim Baldwin, Michael Felsberg, and Fahad S. Khan. Palo: A large multilingual multimodal language model. WACV, 2025. 8", + "[33] Alireza Saeedi, Kun Wang, Ghazaleh Nikpourian, Andreas Bartels, Nikos K Logothetis, Nelson K Totah, and Masataka Watanabe. Brightness illusions drive a neuronal response in the primary visual cortex under top-down modulation. Nature Communications, 2024. 2", + "[34] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 1, 2", + "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2015. 7", + "[36] Zineng Tang, Lingjun Mao, and Alane Suhr. Grounding language in multi-perspective referential communication. EMNLP, 2024. 1", + "[37] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? exploring the visual shortcomings of multimodal llms. CVPR, 2024. 5", + "[38] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. PMLR, 2022. 5", + "[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogVLM: Visual expert for pretrained language models. NeuRIPS, 2024. 1, 5", + "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. NeuRIPS, 2022. 1", + "[41] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. arXiv preprint arXiv:2304.00685, 2024. 1", + "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. CVPR, 2023. 1", + "[43] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding Visual Illusions in Language: Do vision-language models perceive illusions like humans? EMNLP, 2023. 1, 2, 5", + "[44] Zeliang Zhang, Xin Liang, Mingqian Feng, Susan Liang, and Chenliang Xu. Will the inclusion of generated data amplify bias across generations in future image classification models? arXiv preprint arXiv:2410.10160, 2024.1" + ], + "bbox": [ + 91, + 92, + 480, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7814", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_model.json b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..55e76c9ccdd237f9fd6670b58dd2c3216fe2ecfb --- /dev/null +++ b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.857, + 0.152 + ], + "angle": 0, + "content": "Evaluating Model Perception of Color Illusions in Photorealistic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.181, + 0.706, + 0.235 + ], + "angle": 0, + "content": "Lingjun Mao Zineng Tang Alane Suhr University of California, Berkeley {lingjun, terran, suhr}@berkeley.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.499 + ], + "angle": 0, + "content": "We study the perception of color illusions by vision-language models. Color illusion, where a person's visual system perceives color differently from actual color, is well-studied in human vision. However, it remains underexplored whether vision-language models (VLMs), trained on large-scale human data, exhibit similar perceptual biases when confronted with such color illusions. We propose an automated framework for generating color illusion images, resulting in RCID (Realistic Color Illusion Dataset), a dataset of 19,000 realistic illusion images. Our experiments show that all studied VLMs exhibit perceptual biases similar human vision. Finally, we train a model to distinguish both human perception and actual pixel differences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.222, + 0.543 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Human visual perception does not always align perfectly with the real world [4, 12, 17]; we are easily influenced by contextual factors such as lighting conditions or surrounding objects and patterns. Figure 1 shows several examples of color illusions, where human judgment of object color and brightness does not align with actual pixel values. Currently, many large vision-language models (VLMs) [9, 22, 23, 39] exhibit behavioral similarities to human visual perception in various aspects, including object recognition [41, 44], scene understanding [24, 36], and spatial reasoning [5]. But because the human visual system behaves idiosyncratically when encountering these illusions, we aim to study whether these models perceive such illusions in the same way humans do. Prior work [34, 43] has explored this question by sourcing illusion images from the internet. However, this approach has a drawback: the number of illusion images on the web is limited, and most images (e.g., \\(60\\%\\) for the IllusionVQA dataset [34]) are well-known examples of these illusions; thus, VLMs have likely memorized humanlike behavioral responses to them. Additionally, the limited scale also restricts the depth and variety of analyses that can be conducted." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To address these limitations, we propose an automatic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.513 + ], + "angle": 0, + "content": "framework for generating realistic color illusion images, covering many scenarios with illusions that we might genuinely encounter in our real life. We cover three different types of color illusions: contrast (Figure 1a), stripe (Figure 1b), and filter (Figure 1c) illusions. For contrast and stripe illusions, we first generate simple arrangements of shapes exhibiting specific illusions, which are then processed through ControlNet [42] to generate more realistic images. For filter illusions, we directly apply filters to the original MS COCO 2017 images. In the end, we produce a dataset consisting of total 19,000 images, half of which include an illusion, and the other half being a control set without illusions. Each image is paired with a synthetically-generated natural language question, asking either about color differences between specific regions or directly inquiring about the color of a particular object in the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.909, + 0.761 + ], + "angle": 0, + "content": "We test recent VLMs using our generated illusion images. We find that, after appropriate fine-tuning, they perform well on non-illusion images, but their accuracy drops significantly on illusion images. Some of their responses exhibit visual biases similar to those of humans, while some are entirely incorrect, failing to align with either pixel values or human perception. We also observe that external prompts, such as Chain of Thought (CoT) [40] and few-shot examples, can influence VLM performance to some extent. Furthermore, fine-tuning on illusion images enhances the models' ability to understand these color illusions. Based on this, we propose a foundational baseline method that enables models to capture both human perceptual biases and true pixel values. Finally, we investigate the underlying causes of VLM susceptibility to color illusions, including biases from both the visual system and prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In conclusion, our contributions are as follows: (1) We propose an automated framework for generating realistic illusion images and create a large, realistic dataset of color illusion images, named RCID (Realistic Color Illusion Dataset), to enhance the fairness and accuracy of model testing. (2) We investigate underlying mechanisms of color illusions in VLMs, highlighting the combined influence of the vision system and prior knowledge. We also explore how external prompts and instruction tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7805" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.353, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.22, + 0.352, + 0.263 + ], + "angle": 0, + "content": "Q1: Based on pixel values, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A1: III" + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.271, + 0.349, + 0.314 + ], + "angle": 0, + "content": "Q2: According to human perception, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A2: II" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.22, + 0.352, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.321, + 0.279, + 0.333 + ], + "angle": 0, + "content": "(a) Contrast Illusion" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.092, + 0.627, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.375, + 0.221, + 0.621, + 0.263 + ], + "angle": 0, + "content": "Q1: Evaluated by pixel values, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A1: III" + }, + { + "type": "image_footnote", + "bbox": [ + 0.374, + 0.271, + 0.622, + 0.313 + ], + "angle": 0, + "content": "Q2: According to human perception, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A2: II" + }, + { + "type": "list", + "bbox": [ + 0.374, + 0.221, + 0.622, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.321, + 0.547, + 0.334 + ], + "angle": 0, + "content": "(b) Stripe Illusion" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.092, + 0.901, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.649, + 0.22, + 0.899, + 0.24 + ], + "angle": 0, + "content": "Q1: According to pixel values, what color is the athlete's uniform?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.651, + 0.241, + 0.877, + 0.261 + ], + "angle": 0, + "content": "A1: ['Gray', 'Yellowish-green', 'Greenish-yellow', 'Green']" + }, + { + "type": "image_footnote", + "bbox": [ + 0.651, + 0.27, + 0.879, + 0.29 + ], + "angle": 0, + "content": "Q2: Through human eyes, what color is the athlete's uniform?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.652, + 0.291, + 0.705, + 0.301 + ], + "angle": 0, + "content": "A2: ['Red']" + }, + { + "type": "list", + "bbox": [ + 0.649, + 0.22, + 0.899, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.321, + 0.82, + 0.333 + ], + "angle": 0, + "content": "(c) Filter Illusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.342, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Figure 1. Examples of color illusions. (a) Contrast illusion: The two green squares on the left and right are identical in color, but most people perceive the square on the right as darker due to the brighter background. (b) Stripe illusion: The left and right walls seem to be different colors, but this is an illusion caused by the dark stripes interfering with our perception. (c) Filter illusion: The player's uniform appears red, but in reality, there are no red pixels present." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.512 + ], + "angle": 0, + "content": "impact the models' performance on these illusions. (3) We propose a simple training method that enables models to understand human perception while also recognizing the actual pixel values. Our code and data are released under an open-source license at the following URL: https://github.com/mao1207/RCID." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.233, + 0.543 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Visual illusion is a phenomenon where human perception differs from the actual environment [4, 12, 17], often triggered by specific visual stimuli [7]. These illusions offer valuable insights into human perception [16, 31] and have been widely used in artworks. [13, 25]. Many studies [28, 29] have attempted to explore the causes of visual illusions, with the bottom-up and top-down theories being the most widely accepted explanations. Proponents [33] of the bottom-up theory argue that low-level visual signals, such as contrast, brightness, edges, and motion, are correlated with human perception of illusions. However, supporters of the top-down theory contend that prior knowledge is also a major factor in these illusions. Evidence shows that people in industrialized societies are more susceptible to the Müller-Lyer illusion due to being misled by implied arrow meanings, and individuals from language communities with fewer color terms are less affected by color illusions [1, 30]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Motivated by illusion phenomena in human visual perception, recent research has begun to explore whether models are also susceptible to such illusions. Early research [14, 15] suggests that traditional vision models, such as CNNs, are influenced by visual illusions in a way similar to humans. More recent studies [34, 43] have shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.588 + ], + "angle": 0, + "content": "that VLMs also exhibit similar perceptual biases. However, images on which these models have been tested are either sourced from the internet or manual compositions of simple shapes (e.g., two squares of different colors), making it easy for the VLM to infer results based on their knowledge rather than actual observation. Additionally, the number of these images is quite limited. To address these limitations, we develop a large-scale dataset that embeds visual illusions into realistic scenes, enabling robust evaluation in complex, real-world contexts. Our study fills a gap by investigating the mechanisms behind VLMs' susceptibility to illusions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.785, + 0.617 + ], + "angle": 0, + "content": "3. Illusion Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.84 + ], + "angle": 0, + "content": "The construction of our dataset involves three steps (Figure 2): (1) Image Generation. For contrast and stripe illusions, we use procedural code to generate simple illusion images, which are then processed by ControlNet to create realistic illusion images. For filter illusions, we directly apply contrasting color filters to the original images. Each type of illusion also includes a corresponding control group without any illusions for comparison. (2) Question Generation. We use GPT-4o to generate image-specific questions that are designed to evaluate the model's understanding of the illusion. (3) Human Feedback. We collect human participants' feedback on these images and adjust the original classification of \"illusion\" and \"non-illusion\" based on whether participants are deceived." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.907, + 0.865 + ], + "angle": 0, + "content": "3.1. Contrast and Stripe Illusion Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "ControlNet Training Stage: We begin by training a ControlNet to map from simple illusion images to photoreal-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7806" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.09, + 0.627, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.091, + 0.862, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.388, + 0.63, + 0.403 + ], + "angle": 0, + "content": "Figure 2. Process for generating our dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.492 + ], + "angle": 0, + "content": "istic images with illusions. We train using modified image-caption pairs \\((I, T)\\) from the MS COCO 2017 dataset. First, we quantize each original image \\(I\\) into a \\(10 \\times 10\\) grid \\(G\\), where each cell \\(G_{x,y}\\) represents the average color of its corresponding region:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.499, + 0.372, + 0.534 + ], + "angle": 0, + "content": "\\[\nG _ {x, y} = \\frac {1}{\\left| R _ {x , y} \\right|} \\sum_ {p \\in R _ {x, y}} I (p)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Here, \\( R_{x,y} \\) denotes the set of pixels in the \\( (x,y) \\)-th grid cell, and \\( I(p) \\) is the color value of pixel \\( p \\) in the original image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.645 + ], + "angle": 0, + "content": "During training, Gaussian noise is added to the quantized image \\( x_0 = G \\), producing a noisy image \\( x_{t} \\) at time step \\( t \\). The model, conditioned on the text prompt \\( T \\) and color distribution \\( G \\), is trained to predict the added noise \\( \\epsilon_{\\theta} \\) by minimizing the expected squared error:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.655, + 0.427, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {x _ {0}, t, C, \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, t, T, G\\right) \\| _ {2} ^ {2} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Through this loss, the model learns to progressively refine \\( x_{t} \\) into a realistic image during inference, guided by color distribution and the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Procedural Generation Stage: We use a procedural function \\( f \\) to generate simple illusion images, denoted as \\( S \\), which consist of basic shapes and colors, on which we will condition the trained ControlNet model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.904 + ], + "angle": 0, + "content": "For Contrast Illusions, we randomly select two base colors from RGB space: \\( C_b \\) for the background and \\( C_f \\) for the foreground (squares). For a color \\( C = (r, g, b) \\), we use the function \\( p \\) to adjust its brightness by \\( p(C, \\mu) = (r \\cdot \\mu, g \\cdot \\mu, b \\cdot \\mu) \\), where \\( \\mu \\) scales each RGB component. The background is divided into a darker region \\( p(C_b, \\mu_{b1}) \\) and a brighter region \\( p(C_b, \\mu_{b2}) \\), where \\( \\mu_{b1} < 1 \\) and \\( \\mu_{b2} > 1 \\). Additionally, two foreground squares, colored \\( p(C_f, \\mu_{f1}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.905, + 0.445 + ], + "angle": 0, + "content": "and \\(p(C_f, \\mu_{f2})\\), are symmetrically placed at random positions \\((x_1, y_1)\\) and \\((x_2, y_2)\\) on the background." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.642 + ], + "angle": 0, + "content": "If human perception differs from the actual pixel values, the image is classified as an illusion; otherwise, it belongs to the control group (no illusion). We observe that a darker background enhances the perceived brightness of the foreground color, with this effect being strongest when \\( C_b \\) and \\( C_f \\) are similar and weakening as the contrast increases. Section 4.2.3 validates this conclusion. Based on this principle, we generate theoretical illusion and non-illusion images, which are later adjusted through human validation. To increase image diversity, we apply random noise \\( \\eta \\) to introduce irregularities at the square edges or make the boundaries less smooth. The generated image \\( S_i \\) for the contrast illusion can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.661, + 0.879, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} S = f \\left(p \\left(C _ {b}, \\mu_ {b 1}\\right), p \\left(C _ {b}, \\mu_ {b 2}\\right), p \\left(C _ {f}, \\mu_ {f 1}\\right), p \\left(C _ {f}, \\mu_ {f 2}\\right), \\right. \\\\ \\left(x _ {1}, y _ {1}\\right), \\left(x _ {2}, y _ {2}\\right), \\eta) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Similarly, for Stripe Illusions, we randomly generate the background color \\( C_b \\) and the stripe color \\( C_s \\). The colored stripes alternate with black stripes, and their colors are adjusted to \\( p(C_s, \\mu_{s1}) \\) and \\( p(C_s, \\mu_{s2}) \\). Based on an initial estimate of human perception of \\( \\mu_{s1} \\) and \\( \\mu_{s2} \\), the images are classified as either illusion or non-illusion. The generated image \\( S \\) for the stripe illusion is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.821, + 0.843, + 0.837 + ], + "angle": 0, + "content": "\\[\nS = f \\left(C _ {b}, p \\left(C _ {s}, \\mu_ {s 1}\\right), p \\left(C _ {s}, \\mu_ {s 2}\\right), \\theta , N, \\eta\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Here the stripe direction \\(\\theta\\) (horizontal, vertical, or diagonal) and the number of stripes \\(N\\) are randomly selected. Random noise \\(\\eta\\) is applied to create irregularities, such as curved stripe arrangements or slight misalignments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.093, + 0.396, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.094, + 0.657, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.096, + 0.855, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.341, + 0.698, + 0.355 + ], + "angle": 0, + "content": "Figure 3. Data statistics of RCID (Realistic Color Illusion Dataset)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.368, + 0.484, + 0.504 + ], + "angle": 0, + "content": "ControlNet Generating Stage: We pair each simple illusion image \\( S \\) with a randomly selected caption \\( T \\) from the MS COCO 2017 dataset, which together serve as input to our model. Starting from the quantized image \\( x_0 = S \\), the model recursively denoises over multiple time steps following a diffusion-based approach, gradually refining the image to produce a realistic output \\( I' \\) that retains the intended illusion effects. The final image \\( I' \\) is obtained by summing the contributions from all time steps \\( T \\):" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.514, + 0.383, + 0.551 + ], + "angle": 0, + "content": "\\[\nI ^ {\\prime} = \\sum_ {t = 0} ^ {T} \\left(x _ {t} + \\epsilon_ {\\theta} \\left(x _ {t}, t, C, S\\right)\\right)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.377, + 0.575 + ], + "angle": 0, + "content": "3.2. Filter Illusion Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.704 + ], + "angle": 0, + "content": "To generate filter illusion images, we first select images from MS COCO 2017 that predominantly contain a specific color \\( C_{\\mathrm{d}} \\) (e.g. red, blue, yellow). We then apply a contrasting color filter \\( C_{\\mathrm{f}} \\) which overlays the original image in the HSV (Hue, Saturation, Value) color space to suppress \\( C_{\\mathrm{d}} \\) by shifting the hue \\( H \\) of filter and ensuring no pixels fall within \\( C_{\\mathrm{d}} \\)'s HSV range. Despite this, the filtered regions may still visually evoke the original color." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.288, + 0.729 + ], + "angle": 0, + "content": "3.3. Question Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For each image \\(I'\\), we use GPT-4o to generate an image-grounded question \\(\\mathcal{Q}\\) that asks human respondents to compare colors between two regions or identify the color of a specific object. For example, in Figure 1a, the generated questions refer to the walls on the left and right of the baseball player. We then prefix the question with two distinct prompts: one focused on \"human perception\" (e.g., \"Based on human perception\", \"Through human eye\") and the other on \"pixel values\", each associated with a different answer depending on whether the image contains an illusion. For contrast and stripe illusions, answer options are appended" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.838, + 0.384 + ], + "angle": 0, + "content": "to the question. Examples are shown in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.396, + 0.799, + 0.411 + ], + "angle": 0, + "content": "3.4. Human Validation and Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.587 + ], + "angle": 0, + "content": "To ensure that our generated images truly contain illusions, we conduct a survey on Prolific with a total of 241 participants. Each participant is shown 400 images, including both illusion and non-illusion examples. We use respondents' answers as the human perception ground truth. For every image in our dataset, we assign 5 participants. If more than 3 participants are deceived by the illusion, the image is classified as an illusion image. Conversely, if all 5 participants provide answers consistent with the pixel values, the image is classified as non-illusion and added to the control group. On average, we pay $21 USD per hour." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.666, + 0.614 + ], + "angle": 0, + "content": "3.5. Dataset Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.906, + 0.713 + ], + "angle": 0, + "content": "Realistic Color Illusion Dataset (RCID, Figure 3) contains 19,000 images, comprising 8,000 contrast, 9,000 stripe, and 2,000 filter images. We divide our dataset into 9,500 images for training, 4,750 for development, and the remaining 4,750 for the test set. Each illusion type has distinct structures and subcategories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.73, + 0.843, + 0.747 + ], + "angle": 0, + "content": "4. Evaluating Vision-Language Models" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.704, + 0.772 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.906, + 0.854 + ], + "angle": 0, + "content": "In this experiment, we evaluate different models' performance on color illusions and explore the impact of external prompts and instruction-tuning on them. We also provide a foundational baseline model based on LLaVA-1.5 (7B) for understanding color illusions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Vision-Language Models. We test recent models with our generated illusion dataset to examine whether VLMs exhibit perceptual responses similar to humans. To ensure" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.862, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.346, + 0.908, + 0.404 + ], + "angle": 0, + "content": "Figure 4. This figure shows the proportion of different model responses across three types of illusions (Contrast, Filter, and Stripe) on our development set. For non-illusion images, we report the proportions of \"Accurate\" and \"Wrong\" responses. For illusion images, we categorize responses as \"No Illusion\" (consistent with pixel values), \"Human like,\" and \"N/A.\" Each image is evaluated using two types of prompts: one based on pixel values and the other based on human perception." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.419, + 0.485, + 0.722 + ], + "angle": 0, + "content": "that their performance on illusion images reflects actual susceptibility to illusions rather than challenges in identifying specified regions and comparing color brightness [11, 37], we focus on open-source models, and fine-tune them on our generated non-illusion simple images and a small set of realistic images to enhance their ability in answering questions about relative colors of different regions in an image. During training, we randomly apply prompts such as 'based on human perception,' 'based on pixel values,' or provide no specific guidance prompt at all. Our models tested include LLaVA [22], InstructBLIP [9], CogVLM [39], InternVL2 [6], Qwen-VL2 [2], OFA [38] and Unified-IO [26]. During inference for contrast and stripe illusions, questions are presented in a multiple-choice format, while filter illusion questions require open-ended responses about the color of specific objects in the image, as illustrated in Figure 1. We also include experiments with closed-source models like GPT-4o and Gemini in the appendix. While their overall color comparison ability is weak, there is still a significant drop in performance on illusion images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Metrics. For images containing illusions, we use the same metrics as in GVIL [43]. We track the Human-like rate, which measures alignment between human perception and VLM responses by calculating the percentage of cases where the model's responses match human answers. When the prompt explicitly includes \"Based on pixel values,\" this can also be referred to as Deception Rate. The No-Illusion rate indicates the proportion of examples where the model's responses align with actual pixel values. If the model's answer aligns with neither the human response nor the pixel values, we categorize those instances as Not Applicable" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.464 + ], + "angle": 0, + "content": "\\((N / A)\\). For images without illusions, where human perception matches pixel values, we classify model responses as \"accurate\" or \"inaccurate\"." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.71, + 0.487 + ], + "angle": 0, + "content": "4.2. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.493, + 0.907, + 0.735 + ], + "angle": 0, + "content": "We evaluate a range of open-source vision-language models using generated illusion and non-illusion images in our development set, with questions explicitly asking for color judgments 'Based on pixel values' or 'Based on human perception'. The results show that, after fine-tuning on non-illusion images, these models achieve high accuracy \\((>75\\%)\\) on non-illusion images, while their accuracy on illusion images is significantly lower. We find that explicitly querying for color judgments 'Based on pixel values' or 'Based on human perception' does not lead to significant changes in model performance. In addition, models are likely to produce responses that are completely inaccurate, matching neither pixel values nor human perception. This suggests that while the models are misled by color illusions to some extent, they still struggle to fully model human perception. Detailed results can be found in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Inter-Annotator Agreement. We refine our final set of illusion images through human validation, retaining only those that deceive at least 3 out of 5 human participants. Before this selection, the average Fleiss' kappa score among the five respondents on the illusion images is 0.648, which increases to 0.806 after filtering. \\(21.6\\%\\) of the selected images deceive all five participants. We use this stricter subset of illusion images as a challenging test set for our models. The results show a slight increase \\((< 5\\%)\\) in the proportion of Human-Like responses by the VLMs across all three types of illusions, but overall, the difference remains" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.09, + 0.46, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.858, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.2, + 0.745, + 0.215 + ], + "angle": 0, + "content": "Figure 5. Deception rates of humans and VLMs across different structural patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.483, + 0.257 + ], + "angle": 0, + "content": "minimal, indicating that our filtering criterion (more than 3 humans are deceived) is robust." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.258, + 0.483, + 0.363 + ], + "angle": 0, + "content": "Generated Image Realism. As some generated illusion images appear realistic while others still exhibit an \"AI-looking\" appearance, we assess LLaVA-7B and two human participants on three 500-image sets with varying realism levels. Our results show no significant correlation between realism and accuracy, with \\( r = 0.084 \\) for human participants and \\( r = 0.091 \\) for the model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.484, + 0.531 + ], + "angle": 0, + "content": "Model Size. To investigate the effect of model size, we evaluate the performance of two groups of models, OFA and Unified-IO. These groups consist of models of varying sizes but share the same overall architecture. We conduct the experiment on contrast illusions, and explicitly emphasize \"Based on pixel values\" in the questions. Our result shows, with the model size increases, the proportion of human-like responses also increases, while the proportion of responses consistent with actual pixel values decreases. We present OFA's performance in Figure 6 and provide Unified-IO's performance details in the appendix." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.544, + 0.433, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.71 + ], + "angle": 0, + "content": "Figure 6. Proportions of 'No Illusion,' 'Human Like,' and 'N/A' responses for OFA models of different sizes on contrast illusions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Illusion Design. We explore a range of visual factors that may influence the strength of color illusions and compare whether these factors affect human perception and VLMs in the same way. We focus on three potential influencing factors: the orientation of the illusion, the contrast between foreground and background colors (for contrast illusions only), and the number of stripes (for filter illusions only). Overall, our findings indicate that these factors significantly impact the strength of the illusion. For example, altering certain factors, such as increasing the color contrast between the foreground and background, can turn a non-illusion image into an illusion image. And these effects are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.258 + ], + "angle": 0, + "content": "consistent across both humans and VLMs. In these experiments, we use LLaVA-1.5 (7B) as our tested model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Illusion orientation. We randomly generate a set of images with various structural arrangements to analyze how the location of the illusion affects humans and VLMs. For contrast illusions, the images feature left-right and up-down arrangements, while for stripe illusions, they include horizontal, vertical, and diagonal stripes (Figure 3). We evaluate the error rates of both humans and VLMs across these structures. As illustrated in Figure 5, the results show that for contrast illusions, both humans and VLMs are more frequently deceived by left-right arrangements, while for stripe illusions, they are more often misled by horizontal and diagonal stripes. We find that humans can quickly and accurately perceive color differences between left and right, but are slightly less sensitive to top-bottom differences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.909, + 0.61 + ], + "angle": 0, + "content": "Color distance between the foreground and background. For contrast illusions, we calculate the correlation between the color distance in RGB space between the foreground and background, and the illusion's intensity. Figure 7a shows that as the foreground and background colors become more similar, the contrast illusion becomes most pronounced for both humans and VLMs. Conversely, when the foreground and background colors are distinctly different, humans and VLMs are better at distinguishing them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Image spatial frequency. For stripe illusions, we examine the effect of image spatial frequency on the strength of the illusion. After converting the image to the frequency domain using the Fourier transform, we compute spatial frequency as the ratio of high-frequency energy to the total image energy [20]. Our findings suggest that the stripe illusion becomes stronger for both humans and models as image frequency increases, which typically corresponds to a higher number of stripes, as shown in Figure 7b." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Prompting Methods. We explore whether external prompts could alter VLM responses. In contrast to our main experiments shown in Figure 4, here we aim to identify the model's bias towards predicting answers which are human-like or based on pixel values without specifying this in the question posed to the model. We consider several prompting methods: a simple prompt that only contains questions about color comparisons; chain of thought (CoT), where the model is prompted to first consider if any factors in the image might affect color perception; emphasizing the illusion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7810" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.171 + ], + "angle": 0, + "content": "
Illusion CategoryContrast IllusionFilter IllusionStripe Illusion
PromptsPixel-BasedHuman-BasedPixel-BasedHuman-BasedPixel-BasedHuman-Based
LLaVA-1.5 (7B)35.4 (41.2)32.4 (37.4)20.0 (67.8)19.8 (64.6)30.3 (35.2)28.3 (40.1)
Mixed Training (Ours)83.9 (6.9)8.1 (79.0)68.0 (24.3)12.1 (82.3)84.5 (9.8)15.4 (76.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.18, + 0.907, + 0.208 + ], + "angle": 0, + "content": "Table 1. Model's improved understanding of illusions with mixed training. Two prompts are used: \"based on pixel value\" and \"based on human perception.\" The values represent the proportion of no-illusion responses, with parentheses indicating Human-like rates." + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.217, + 0.415, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.375, + 0.424, + 0.398 + ], + "angle": 0, + "content": "(a) Influence of background–foreground color difference on intensity of contrast illusion." + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.4, + 0.42, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.517, + 0.237, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.517, + 0.328, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.517, + 0.419, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.557, + 0.424, + 0.581 + ], + "angle": 0, + "content": "(b) Influence of image spatial frequency on intensity of stripe illusions." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.617 + ], + "angle": 0, + "content": "Figure 7. Error rates of humans and LLaVA across different conditions in color illusions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.734 + ], + "angle": 0, + "content": "by directly informing the model that the image contains an illusion without specifying its type; and providing few-shot examples, using no-illusion (NI) or human-like (HL) answers as 3-shot examples. Because we use the LLaVA-7B model here, which doesn't support multi-image input, we combine the few-shot examples and the query image into a single composite image before feeding it to the model [22]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Figure 8 show that external prompts have minimal effect on filter illusions. But for contrast and stripe illusions, the effect is more pronounced, particularly when no-illusion answers are used as few-shot examples, increasing the proportion of no-illusion responses by approximately \\(25\\%\\). However, using human-like responses as few-shot examples does not effectively enhance the model's reproduction of human illusion perception, the proportion of human-like responses even decreases slightly for contrast illusions. Additionally, both direct emphasis and CoT prompts partially increase the proportion of no-illusion responses, al" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.221, + 0.871, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.371, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Figure 8. Effect of different prompts and fine-tuning on VLM response preferences across illusion Types. \"NI\" denotes datasets using no-illusion examples, while \"HL\" represents datasets with human-like examples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.874, + 0.455 + ], + "angle": 0, + "content": "though the effect of direct emphasis is relatively weak." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Fine-Tuning. First, we fine-tune two separate LLaVA-7B models on illusion images, with one model trained with responses according to pixel values (NI), and the other trained on human-like responses (HL). The bottom two rows of Figure 8 show a significant effect on the test set, particularly when fine-tuning on no-illusion responses. For both contrast and stripe illusions, fine-tuning on no-illusion responses increases the proportion of no-illusion answers to over \\(90\\%\\). In contrast, fine-tuning on human-like responses has a weaker effect, suggesting that even with human-like responses, the model struggles to fully grasp human perception in the context of illusion images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Next, we experiment with a mixed training approach that combines both types of responses. We train on question-answer pairs for both types of responses, where questions asking \"Based on pixel values\" are paired with pixel-value responses, those asking \"Based on human perception\" are paired with a human-like response. Table 1 shows the influence of this mixed training on the test set, compared to the base model. These results confirm that training model can effectively distinguish between pixel-based and human perception cues, yielding responses aligned with user's query." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.822, + 0.823 + ], + "angle": 0, + "content": "5. Why Color Illusions Affect VLMs" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.866, + 0.848 + ], + "angle": 0, + "content": "5.1. Perceptual Biases from the Visual System" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To investigate whether VLMs' perceptual biases originate from their visual components, we test a range of purely visual models, including ResNet [18], ViT [10], VGG [35]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.278, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.188, + 0.282, + 0.224 + ], + "angle": 0, + "content": "(a) Performance of purely vision models on non-illusion and illusion images." + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.094, + 0.48, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.188, + 0.483, + 0.224 + ], + "angle": 0, + "content": "(b) Accuracy of ViT on non-illusion and illusion images over each training epochs." + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.233, + 0.442, + 0.247 + ], + "angle": 0, + "content": "Figure 9. Perceptual biases in purely vision models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.485, + 0.609 + ], + "angle": 0, + "content": "MobileNet [19] and Large Vision Model (LVM) [3]. We evaluate these models on both simple, rectangle-based contrast illusion images and generated realistic illusion images. For simple illusion images, we generate backgrounds with varying brightness on both sides and place a rectangle on each side. The vision model is trained as a three-way classifier on 6,000 non-illusion images to predict which rectangle appears darker, or if their colors are identical, stopping when the loss shows no significant decrease. It is then tested on 1,000 non-illusion and 1,000 illusion images. The results are shown in Figure 9a. For realistic illusion images, we train the model on the training set of our generated dataset and evaluate it on the test set, shown in the appendix. While overall accuracy (the proportion of classifications aligning with actual pixel values) on realistic illusion images drops compared to simple rectangle-based images, in both cases, accuracy on illusion images remains significantly lower than on non-illusion images. We also observe that during training, accuracy on non-illusion images quickly improves, while accuracy on illusion images fluctuates significantly. We use ViT as an example in Figure 9b. These results suggest that even purely vision-based models can be influenced by color illusions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.623, + 0.44, + 0.639 + ], + "angle": 0, + "content": "5.2. Perceptual Biases from Prior Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.873 + ], + "angle": 0, + "content": "Influence of Language. We test the multilingual vision-language model PALO [32] on contrast illusions, posing questions in English, Chinese, and Russian to analyze error rates on images containing target regions across several colors. We hypothesize that colors with more frequent textual modifiers (e.g., dark blue, sky blue) lead to responses aligned with pixel values rather than human perception. To validate this, we first analyze color descriptor diversity in the training data. For example, in English, \\(50.7\\%\\) of red instances include modifiers like dark red or bright red, while \\(49.3\\%\\) use red alone. Figure 10 shows the correlation between color descriptor diversity and pixel-value accuracy for English, Chinese, and Russian. We find that richer descriptions improve model sensitivity to that color, reducing susceptibility to color illusions in that color range." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.903 + ], + "angle": 0, + "content": "1In Russian, blue has 100% descriptor diversity due to distinct terms for light blue (goluboj) and deep blue (sinij)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.181, + 0.907, + 0.223 + ], + "angle": 0, + "content": "Figure 10. Comparison of color term accuracy and frequency across languages for the vision-language model. Each shape and color represents a specific color term (e.g., blue, red, gray)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.907, + 0.369 + ], + "angle": 0, + "content": "Influence of Commonsense Knowledge. Humans are often deceived by visual illusions due to our prior knowledge. For instance, research suggests that in filter illusions, humans perceive colors that aren't actually present due to a phenomenon called chromatic adaptation [27]. This allows us to use surrounding context to infer the presence of a filter, mentally \"removing\" it to perceive the original color. One example is that we always recognize a piece of white paper as white, regardless of the lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.909, + 0.568 + ], + "angle": 0, + "content": "To examine whether commonsense knowledge similarly influences VLMs' susceptibility to illusions, we categorized filter illusion images into three groups: single objects, multiple-object compositions, and complex real-world scenes, then analyzed the deception rates for both humans and VLMs across these groups. On non-illusion images, VLMs perform well across all three categories, suggesting that scene complexity does not affect their basic color recognition ability. However, for illusion images, our results in Figure 11 show that both humans and VLMs are less likely to be deceived by single-object images, while complex scenes lead to the highest deception rates, underscoring the role of prior knowledge in visual illusions." + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.579, + 0.868, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.682, + 0.909, + 0.71 + ], + "angle": 0, + "content": "Figure 11. Deception rates of VLMs and human participants across different types of filter illusions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our experiments show that current VLMs, influenced by their visual systems and prior knowledge, can exhibit visual illusions similar to humans. This raises a key question: as models become more human-like, might they also inherit human perceptual biases? In fields like biomedicine [8, 21], such biases could result in flawed visual judgments, while in human-centric tasks like image generation, understanding human perception can benefit. Therefore, VLM behavior should be tailored to application needs, with careful consideration of human-like perceptual biases." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7812" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.142 + ], + "angle": 0, + "content": "[1] Adam Alter. Drunk tank pink: And other unexpected forces that shape how we think, feel, and behave. 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[4] Claus-Christian Carbon. Understanding human perception by human-made illusions. Frontiers in human neuroscience, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: Endowing vision-language models with spatial reasoning capabilities. CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.506 + ], + "angle": 0, + "content": "[7] Fan L Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C Aoki, Jin Hirano, and Yukiyasu Kamitani. Reconstructing visual illusory experiences from human brain activity. Science Advances, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[8] Hejie Cui, Lingjun Mao, Xin Liang, Jieyu Zhang, Hui Ren, Quanzheng Li, Xiang Li, and Carl Yang. Biomedical visual instruction tuning with clinician preference alignment. NeurIPS, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. BLINK: Multimodal large language models can see but not perceive. ECCV, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[12] Maurizio Gentilucci, Sergio Chieffi, Elena Daprati, M Cristina Saetti, and Ivan Toni. Visual illusion and action. Neuropsychologia, 1996. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Ernst Hans Gombrich. Art and illusion: A study in the psychology of pictorial representation-millennium edition. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmio, and Jesus Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[16] Richard Langton Gregory. Perceptual illusions and brain models. Proceedings of the Royal Society of London. Series B. Biological Sciences, 1968. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[17] Richard L Gregory. Visual illusions. Scientific American, 1968. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[19] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.362 + ], + "angle": 0, + "content": "[20] Yan Ke, Xiaou Tang, and Feng Jing. The design of high-level features for photo quality assessment. CVPR, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.364, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[21] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. NeurIPS, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.904, + 0.461 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeuRIPS, 2023. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.904, + 0.502 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[24] Sichao Liu, Jianjing Zhang, Robert X. Gao, Xi Vincent Wang, and Lihui Wang. Vision-language model-driven scene understanding and robotic object manipulation. IEEE CASE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.904, + 0.589 + ], + "angle": 0, + "content": "[25] Margaret S Livingstone. Art, illusion and the visual system. Scientific American, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[26] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-IO 2: Scaling autoregressive multimodal models with vision, language, audio, and action. CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.904, + 0.687 + ], + "angle": 0, + "content": "[27] D. L. MacAdam. Chromatic adaptation*. J. Opt. Soc. Am., 1956. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[28] O Maddaluno, A Facchin, D Zavagno, N Bolognini, R Daini, et al. Top-down and bottom-up neuromodulation over two different visual illusions. Atti del 40th European Conference on Visual Perception, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[29] Glenn E Meyer and Susan Petry. Top-down and bottom-up: The illusory contour as a microcosm of issues in perception. 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[30] William L Phillips. Cross-cultural differences in visual perception of color, illusions, depth, and pictures. *Cross-Cultural Psychology: Contemporary Themes and Perspectives*, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[31] Jiang Qiu, Hong Li, Qinglin Zhang, Qiang Liu, and Fenghua Zhang. The muller-lyer illusion seen by the brain: An event-related brain potentials study. Biological psychology, 2008. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7813" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[32] Hanoona Rasheed, Muhammad Maaz, Abdelrahman Shaker, Salman Khan, Hisham Cholakal, Rao M. Anwer, Tim Baldwin, Michael Felsberg, and Fahad S. Khan. Palo: A large multilingual multimodal language model. WACV, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[33] Alireza Saeedi, Kun Wang, Ghazaleh Nikpourian, Andreas Bartels, Nikos K Logothetis, Nelson K Totah, and Masataka Watanabe. Brightness illusions drive a neuronal response in the primary visual cortex under top-down modulation. Nature Communications, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[34] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[36] Zineng Tang, Lingjun Mao, and Alane Suhr. Grounding language in multi-perspective referential communication. EMNLP, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[37] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? exploring the visual shortcomings of multimodal llms. CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[38] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogVLM: Visual expert for pretrained language models. NeuRIPS, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. NeuRIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[41] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. arXiv preprint arXiv:2304.00685, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[43] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding Visual Illusions in Language: Do vision-language models perceive illusions like humans? EMNLP, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[44] Zeliang Zhang, Xin Liang, Mingqian Feng, Susan Liang, and Chenliang Xu. Will the inclusion of generated data amplify bias across generations in future image classification models? arXiv preprint arXiv:2410.10160, 2024.1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7814" + } + ] +] \ No newline at end of file diff --git a/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_origin.pdf b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..04ac8735bf9bbd204abc8231ad76e782f6078666 --- /dev/null +++ b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/9bb16cc9-5b88-4a24-a8e1-499494d36b2e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:270c0a0381a7e84cd3431edb565d5a00296867ed65b115a5047798f33927d425 +size 9603141 diff --git a/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/full.md b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4faf423093979fd8d07104b9dd6ae1cca6c8fddd --- /dev/null +++ b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/full.md @@ -0,0 +1,288 @@ +# Evaluating Model Perception of Color Illusions in Photorealistic Scenes + +Lingjun Mao Zineng Tang Alane Suhr University of California, Berkeley {lingjun, terran, suhr}@berkeley.edu + +# Abstract + +We study the perception of color illusions by vision-language models. Color illusion, where a person's visual system perceives color differently from actual color, is well-studied in human vision. However, it remains underexplored whether vision-language models (VLMs), trained on large-scale human data, exhibit similar perceptual biases when confronted with such color illusions. We propose an automated framework for generating color illusion images, resulting in RCID (Realistic Color Illusion Dataset), a dataset of 19,000 realistic illusion images. Our experiments show that all studied VLMs exhibit perceptual biases similar human vision. Finally, we train a model to distinguish both human perception and actual pixel differences. + +# 1. Introduction + +Human visual perception does not always align perfectly with the real world [4, 12, 17]; we are easily influenced by contextual factors such as lighting conditions or surrounding objects and patterns. Figure 1 shows several examples of color illusions, where human judgment of object color and brightness does not align with actual pixel values. Currently, many large vision-language models (VLMs) [9, 22, 23, 39] exhibit behavioral similarities to human visual perception in various aspects, including object recognition [41, 44], scene understanding [24, 36], and spatial reasoning [5]. But because the human visual system behaves idiosyncratically when encountering these illusions, we aim to study whether these models perceive such illusions in the same way humans do. Prior work [34, 43] has explored this question by sourcing illusion images from the internet. However, this approach has a drawback: the number of illusion images on the web is limited, and most images (e.g., $60\%$ for the IllusionVQA dataset [34]) are well-known examples of these illusions; thus, VLMs have likely memorized humanlike behavioral responses to them. Additionally, the limited scale also restricts the depth and variety of analyses that can be conducted. + +To address these limitations, we propose an automatic + +framework for generating realistic color illusion images, covering many scenarios with illusions that we might genuinely encounter in our real life. We cover three different types of color illusions: contrast (Figure 1a), stripe (Figure 1b), and filter (Figure 1c) illusions. For contrast and stripe illusions, we first generate simple arrangements of shapes exhibiting specific illusions, which are then processed through ControlNet [42] to generate more realistic images. For filter illusions, we directly apply filters to the original MS COCO 2017 images. In the end, we produce a dataset consisting of total 19,000 images, half of which include an illusion, and the other half being a control set without illusions. Each image is paired with a synthetically-generated natural language question, asking either about color differences between specific regions or directly inquiring about the color of a particular object in the image. + +We test recent VLMs using our generated illusion images. We find that, after appropriate fine-tuning, they perform well on non-illusion images, but their accuracy drops significantly on illusion images. Some of their responses exhibit visual biases similar to those of humans, while some are entirely incorrect, failing to align with either pixel values or human perception. We also observe that external prompts, such as Chain of Thought (CoT) [40] and few-shot examples, can influence VLM performance to some extent. Furthermore, fine-tuning on illusion images enhances the models' ability to understand these color illusions. Based on this, we propose a foundational baseline method that enables models to capture both human perceptual biases and true pixel values. Finally, we investigate the underlying causes of VLM susceptibility to color illusions, including biases from both the visual system and prior knowledge. + +In conclusion, our contributions are as follows: (1) We propose an automated framework for generating realistic illusion images and create a large, realistic dataset of color illusion images, named RCID (Realistic Color Illusion Dataset), to enhance the fairness and accuracy of model testing. (2) We investigate underlying mechanisms of color illusions in VLMs, highlighting the combined influence of the vision system and prior knowledge. We also explore how external prompts and instruction tuning + +(a) Contrast Illusion +![](images/93d6999a295cd279ef6593dd55a3802f5f6ba1d2704828692f34cbcdc220645c.jpg) +Q1: Based on pixel values, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A1: III + +(b) Stripe Illusion +![](images/83489c9454b9cd8d35735ada0a6d705d43e9b67c7e93f4a90ae96e77062ad520.jpg) +Q2: According to human perception, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A2: II +Q1: Evaluated by pixel values, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A1: III +Q2: According to human perception, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A2: II + +(c) Filter Illusion +Figure 1. Examples of color illusions. (a) Contrast illusion: The two green squares on the left and right are identical in color, but most people perceive the square on the right as darker due to the brighter background. (b) Stripe illusion: The left and right walls seem to be different colors, but this is an illusion caused by the dark stripes interfering with our perception. (c) Filter illusion: The player's uniform appears red, but in reality, there are no red pixels present. +![](images/11fdeb0e7c043d2cca6d53d5d537103304357b4f89412e1fb5859ac1236715bd.jpg) +Q1: According to pixel values, what color is the athlete's uniform? +A1: ['Gray', 'Yellowish-green', 'Greenish-yellow', 'Green'] +Q2: Through human eyes, what color is the athlete's uniform? +A2: ['Red'] + +impact the models' performance on these illusions. (3) We propose a simple training method that enables models to understand human perception while also recognizing the actual pixel values. Our code and data are released under an open-source license at the following URL: https://github.com/mao1207/RCID. + +# 2. Related Work + +Visual illusion is a phenomenon where human perception differs from the actual environment [4, 12, 17], often triggered by specific visual stimuli [7]. These illusions offer valuable insights into human perception [16, 31] and have been widely used in artworks. [13, 25]. Many studies [28, 29] have attempted to explore the causes of visual illusions, with the bottom-up and top-down theories being the most widely accepted explanations. Proponents [33] of the bottom-up theory argue that low-level visual signals, such as contrast, brightness, edges, and motion, are correlated with human perception of illusions. However, supporters of the top-down theory contend that prior knowledge is also a major factor in these illusions. Evidence shows that people in industrialized societies are more susceptible to the Müller-Lyer illusion due to being misled by implied arrow meanings, and individuals from language communities with fewer color terms are less affected by color illusions [1, 30]. + +Motivated by illusion phenomena in human visual perception, recent research has begun to explore whether models are also susceptible to such illusions. Early research [14, 15] suggests that traditional vision models, such as CNNs, are influenced by visual illusions in a way similar to humans. More recent studies [34, 43] have shown + +that VLMs also exhibit similar perceptual biases. However, images on which these models have been tested are either sourced from the internet or manual compositions of simple shapes (e.g., two squares of different colors), making it easy for the VLM to infer results based on their knowledge rather than actual observation. Additionally, the number of these images is quite limited. To address these limitations, we develop a large-scale dataset that embeds visual illusions into realistic scenes, enabling robust evaluation in complex, real-world contexts. Our study fills a gap by investigating the mechanisms behind VLMs' susceptibility to illusions. + +# 3. Illusion Dataset Construction + +The construction of our dataset involves three steps (Figure 2): (1) Image Generation. For contrast and stripe illusions, we use procedural code to generate simple illusion images, which are then processed by ControlNet to create realistic illusion images. For filter illusions, we directly apply contrasting color filters to the original images. Each type of illusion also includes a corresponding control group without any illusions for comparison. (2) Question Generation. We use GPT-4o to generate image-specific questions that are designed to evaluate the model's understanding of the illusion. (3) Human Feedback. We collect human participants' feedback on these images and adjust the original classification of "illusion" and "non-illusion" based on whether participants are deceived. + +# 3.1. Contrast and Stripe Illusion Image Generation + +ControlNet Training Stage: We begin by training a ControlNet to map from simple illusion images to photoreal- + +![](images/0d37c77073981aa85fc1557a1c734a0403527d23a35aac9b5febcd6cd5f0c6cd.jpg) +Figure 2. Process for generating our dataset. + +![](images/11b47c1997316f14a1763f4f25a9c1a97c24261663e3a41841ce5955f18cfa6a.jpg) + +istic images with illusions. We train using modified image-caption pairs $(I, T)$ from the MS COCO 2017 dataset. First, we quantize each original image $I$ into a $10 \times 10$ grid $G$ , where each cell $G_{x,y}$ represents the average color of its corresponding region: + +$$ +G _ {x, y} = \frac {1}{\left| R _ {x , y} \right|} \sum_ {p \in R _ {x, y}} I (p) +$$ + +Here, $R_{x,y}$ denotes the set of pixels in the $(x,y)$ -th grid cell, and $I(p)$ is the color value of pixel $p$ in the original image. + +During training, Gaussian noise is added to the quantized image $x_0 = G$ , producing a noisy image $x_{t}$ at time step $t$ . The model, conditioned on the text prompt $T$ and color distribution $G$ , is trained to predict the added noise $\epsilon_{\theta}$ by minimizing the expected squared error: + +$$ +\mathcal {L} = \mathbb {E} _ {x _ {0}, t, C, \epsilon \sim \mathcal {N} (0, I)} \left[ \| \epsilon - \epsilon_ {\theta} \left(x _ {t}, t, T, G\right) \| _ {2} ^ {2} \right] +$$ + +Through this loss, the model learns to progressively refine $x_{t}$ into a realistic image during inference, guided by color distribution and the text prompt. + +Procedural Generation Stage: We use a procedural function $f$ to generate simple illusion images, denoted as $S$ , which consist of basic shapes and colors, on which we will condition the trained ControlNet model. + +For Contrast Illusions, we randomly select two base colors from RGB space: $C_b$ for the background and $C_f$ for the foreground (squares). For a color $C = (r, g, b)$ , we use the function $p$ to adjust its brightness by $p(C, \mu) = (r \cdot \mu, g \cdot \mu, b \cdot \mu)$ , where $\mu$ scales each RGB component. The background is divided into a darker region $p(C_b, \mu_{b1})$ and a brighter region $p(C_b, \mu_{b2})$ , where $\mu_{b1} < 1$ and $\mu_{b2} > 1$ . Additionally, two foreground squares, colored $p(C_f, \mu_{f1})$ + +and $p(C_f, \mu_{f2})$ , are symmetrically placed at random positions $(x_1, y_1)$ and $(x_2, y_2)$ on the background. + +If human perception differs from the actual pixel values, the image is classified as an illusion; otherwise, it belongs to the control group (no illusion). We observe that a darker background enhances the perceived brightness of the foreground color, with this effect being strongest when $C_b$ and $C_f$ are similar and weakening as the contrast increases. Section 4.2.3 validates this conclusion. Based on this principle, we generate theoretical illusion and non-illusion images, which are later adjusted through human validation. To increase image diversity, we apply random noise $\eta$ to introduce irregularities at the square edges or make the boundaries less smooth. The generated image $S_i$ for the contrast illusion can be represented as: + +$$ +\begin{array}{l} S = f \left(p \left(C _ {b}, \mu_ {b 1}\right), p \left(C _ {b}, \mu_ {b 2}\right), p \left(C _ {f}, \mu_ {f 1}\right), p \left(C _ {f}, \mu_ {f 2}\right), \right. \\ \left(x _ {1}, y _ {1}\right), \left(x _ {2}, y _ {2}\right), \eta) \\ \end{array} +$$ + +Similarly, for Stripe Illusions, we randomly generate the background color $C_b$ and the stripe color $C_s$ . The colored stripes alternate with black stripes, and their colors are adjusted to $p(C_s, \mu_{s1})$ and $p(C_s, \mu_{s2})$ . Based on an initial estimate of human perception of $\mu_{s1}$ and $\mu_{s2}$ , the images are classified as either illusion or non-illusion. The generated image $S$ for the stripe illusion is represented as: + +$$ +S = f \left(C _ {b}, p \left(C _ {s}, \mu_ {s 1}\right), p \left(C _ {s}, \mu_ {s 2}\right), \theta , N, \eta\right) +$$ + +Here the stripe direction $\theta$ (horizontal, vertical, or diagonal) and the number of stripes $N$ are randomly selected. Random noise $\eta$ is applied to create irregularities, such as curved stripe arrangements or slight misalignments. + +![](images/2b1589ddc6a1a224083e9820436a3e05b1b5aab15edecaaa22abc63b166d0f75.jpg) +Figure 3. Data statistics of RCID (Realistic Color Illusion Dataset). + +![](images/9777cdd4936f9d48191a56a9bfdb9601251f91584ffe8331bd541006a2e95c69.jpg) + +![](images/ad33710260e74fbb7719319735fc58980bd3a1623c78febbb896824569eb0bb8.jpg) + +ControlNet Generating Stage: We pair each simple illusion image $S$ with a randomly selected caption $T$ from the MS COCO 2017 dataset, which together serve as input to our model. Starting from the quantized image $x_0 = S$ , the model recursively denoises over multiple time steps following a diffusion-based approach, gradually refining the image to produce a realistic output $I'$ that retains the intended illusion effects. The final image $I'$ is obtained by summing the contributions from all time steps $T$ : + +$$ +I ^ {\prime} = \sum_ {t = 0} ^ {T} \left(x _ {t} + \epsilon_ {\theta} \left(x _ {t}, t, C, S\right)\right) +$$ + +# 3.2. Filter Illusion Image Generation + +To generate filter illusion images, we first select images from MS COCO 2017 that predominantly contain a specific color $C_{\mathrm{d}}$ (e.g. red, blue, yellow). We then apply a contrasting color filter $C_{\mathrm{f}}$ which overlays the original image in the HSV (Hue, Saturation, Value) color space to suppress $C_{\mathrm{d}}$ by shifting the hue $H$ of filter and ensuring no pixels fall within $C_{\mathrm{d}}$ 's HSV range. Despite this, the filtered regions may still visually evoke the original color. + +# 3.3. Question Generation + +For each image $I'$ , we use GPT-4o to generate an image-grounded question $\mathcal{Q}$ that asks human respondents to compare colors between two regions or identify the color of a specific object. For example, in Figure 1a, the generated questions refer to the walls on the left and right of the baseball player. We then prefix the question with two distinct prompts: one focused on "human perception" (e.g., "Based on human perception", "Through human eye") and the other on "pixel values", each associated with a different answer depending on whether the image contains an illusion. For contrast and stripe illusions, answer options are appended + +to the question. Examples are shown in Figure 1. + +# 3.4. Human Validation and Selection + +To ensure that our generated images truly contain illusions, we conduct a survey on Prolific with a total of 241 participants. Each participant is shown 400 images, including both illusion and non-illusion examples. We use respondents' answers as the human perception ground truth. For every image in our dataset, we assign 5 participants. If more than 3 participants are deceived by the illusion, the image is classified as an illusion image. Conversely, if all 5 participants provide answers consistent with the pixel values, the image is classified as non-illusion and added to the control group. On average, we pay $21 USD per hour. + +# 3.5. Dataset Details + +Realistic Color Illusion Dataset (RCID, Figure 3) contains 19,000 images, comprising 8,000 contrast, 9,000 stripe, and 2,000 filter images. We divide our dataset into 9,500 images for training, 4,750 for development, and the remaining 4,750 for the test set. Each illusion type has distinct structures and subcategories. + +# 4. Evaluating Vision-Language Models + +# 4.1. Experimental Setup + +In this experiment, we evaluate different models' performance on color illusions and explore the impact of external prompts and instruction-tuning on them. We also provide a foundational baseline model based on LLaVA-1.5 (7B) for understanding color illusions. + +Vision-Language Models. We test recent models with our generated illusion dataset to examine whether VLMs exhibit perceptual responses similar to humans. To ensure + +![](images/8945c71c48938d8c53e3b6a9d27721ab28facce79191c992e7a121126b1f2dbc.jpg) +Figure 4. This figure shows the proportion of different model responses across three types of illusions (Contrast, Filter, and Stripe) on our development set. For non-illusion images, we report the proportions of "Accurate" and "Wrong" responses. For illusion images, we categorize responses as "No Illusion" (consistent with pixel values), "Human like," and "N/A." Each image is evaluated using two types of prompts: one based on pixel values and the other based on human perception. + +that their performance on illusion images reflects actual susceptibility to illusions rather than challenges in identifying specified regions and comparing color brightness [11, 37], we focus on open-source models, and fine-tune them on our generated non-illusion simple images and a small set of realistic images to enhance their ability in answering questions about relative colors of different regions in an image. During training, we randomly apply prompts such as 'based on human perception,' 'based on pixel values,' or provide no specific guidance prompt at all. Our models tested include LLaVA [22], InstructBLIP [9], CogVLM [39], InternVL2 [6], Qwen-VL2 [2], OFA [38] and Unified-IO [26]. During inference for contrast and stripe illusions, questions are presented in a multiple-choice format, while filter illusion questions require open-ended responses about the color of specific objects in the image, as illustrated in Figure 1. We also include experiments with closed-source models like GPT-4o and Gemini in the appendix. While their overall color comparison ability is weak, there is still a significant drop in performance on illusion images. + +Metrics. For images containing illusions, we use the same metrics as in GVIL [43]. We track the Human-like rate, which measures alignment between human perception and VLM responses by calculating the percentage of cases where the model's responses match human answers. When the prompt explicitly includes "Based on pixel values," this can also be referred to as Deception Rate. The No-Illusion rate indicates the proportion of examples where the model's responses align with actual pixel values. If the model's answer aligns with neither the human response nor the pixel values, we categorize those instances as Not Applicable + +$(N / A)$ . For images without illusions, where human perception matches pixel values, we classify model responses as "accurate" or "inaccurate". + +# 4.2. Results and Analysis + +We evaluate a range of open-source vision-language models using generated illusion and non-illusion images in our development set, with questions explicitly asking for color judgments 'Based on pixel values' or 'Based on human perception'. The results show that, after fine-tuning on non-illusion images, these models achieve high accuracy $(>75\%)$ on non-illusion images, while their accuracy on illusion images is significantly lower. We find that explicitly querying for color judgments 'Based on pixel values' or 'Based on human perception' does not lead to significant changes in model performance. In addition, models are likely to produce responses that are completely inaccurate, matching neither pixel values nor human perception. This suggests that while the models are misled by color illusions to some extent, they still struggle to fully model human perception. Detailed results can be found in Figure 4. + +Inter-Annotator Agreement. We refine our final set of illusion images through human validation, retaining only those that deceive at least 3 out of 5 human participants. Before this selection, the average Fleiss' kappa score among the five respondents on the illusion images is 0.648, which increases to 0.806 after filtering. $21.6\%$ of the selected images deceive all five participants. We use this stricter subset of illusion images as a challenging test set for our models. The results show a slight increase $(< 5\%)$ in the proportion of Human-Like responses by the VLMs across all three types of illusions, but overall, the difference remains + +![](images/8622480d8734251f951485471d565e8d726bb7a4498f067dba62334aa2997831.jpg) +Figure 5. Deception rates of humans and VLMs across different structural patterns. + +![](images/db77b3c389d0b6826f4cde811605a978ee0c6d0bc2655b118c8b745f8a372b19.jpg) + +minimal, indicating that our filtering criterion (more than 3 humans are deceived) is robust. + +Generated Image Realism. As some generated illusion images appear realistic while others still exhibit an "AI-looking" appearance, we assess LLaVA-7B and two human participants on three 500-image sets with varying realism levels. Our results show no significant correlation between realism and accuracy, with $r = 0.084$ for human participants and $r = 0.091$ for the model. + +Model Size. To investigate the effect of model size, we evaluate the performance of two groups of models, OFA and Unified-IO. These groups consist of models of varying sizes but share the same overall architecture. We conduct the experiment on contrast illusions, and explicitly emphasize "Based on pixel values" in the questions. Our result shows, with the model size increases, the proportion of human-like responses also increases, while the proportion of responses consistent with actual pixel values decreases. We present OFA's performance in Figure 6 and provide Unified-IO's performance details in the appendix. + +![](images/216538a85f9186ed5fe6dfc5a8cc2d33237d1a0d21d82515e2e5e2f2846ba7ca.jpg) +Figure 6. Proportions of 'No Illusion,' 'Human Like,' and 'N/A' responses for OFA models of different sizes on contrast illusions. + +Illusion Design. We explore a range of visual factors that may influence the strength of color illusions and compare whether these factors affect human perception and VLMs in the same way. We focus on three potential influencing factors: the orientation of the illusion, the contrast between foreground and background colors (for contrast illusions only), and the number of stripes (for filter illusions only). Overall, our findings indicate that these factors significantly impact the strength of the illusion. For example, altering certain factors, such as increasing the color contrast between the foreground and background, can turn a non-illusion image into an illusion image. And these effects are + +consistent across both humans and VLMs. In these experiments, we use LLaVA-1.5 (7B) as our tested model. + +Illusion orientation. We randomly generate a set of images with various structural arrangements to analyze how the location of the illusion affects humans and VLMs. For contrast illusions, the images feature left-right and up-down arrangements, while for stripe illusions, they include horizontal, vertical, and diagonal stripes (Figure 3). We evaluate the error rates of both humans and VLMs across these structures. As illustrated in Figure 5, the results show that for contrast illusions, both humans and VLMs are more frequently deceived by left-right arrangements, while for stripe illusions, they are more often misled by horizontal and diagonal stripes. We find that humans can quickly and accurately perceive color differences between left and right, but are slightly less sensitive to top-bottom differences. + +Color distance between the foreground and background. For contrast illusions, we calculate the correlation between the color distance in RGB space between the foreground and background, and the illusion's intensity. Figure 7a shows that as the foreground and background colors become more similar, the contrast illusion becomes most pronounced for both humans and VLMs. Conversely, when the foreground and background colors are distinctly different, humans and VLMs are better at distinguishing them. + +Image spatial frequency. For stripe illusions, we examine the effect of image spatial frequency on the strength of the illusion. After converting the image to the frequency domain using the Fourier transform, we compute spatial frequency as the ratio of high-frequency energy to the total image energy [20]. Our findings suggest that the stripe illusion becomes stronger for both humans and models as image frequency increases, which typically corresponds to a higher number of stripes, as shown in Figure 7b. + +Prompting Methods. We explore whether external prompts could alter VLM responses. In contrast to our main experiments shown in Figure 4, here we aim to identify the model's bias towards predicting answers which are human-like or based on pixel values without specifying this in the question posed to the model. We consider several prompting methods: a simple prompt that only contains questions about color comparisons; chain of thought (CoT), where the model is prompted to first consider if any factors in the image might affect color perception; emphasizing the illusion + +
Illusion CategoryContrast IllusionFilter IllusionStripe Illusion
PromptsPixel-BasedHuman-BasedPixel-BasedHuman-BasedPixel-BasedHuman-Based
LLaVA-1.5 (7B)35.4 (41.2)32.4 (37.4)20.0 (67.8)19.8 (64.6)30.3 (35.2)28.3 (40.1)
Mixed Training (Ours)83.9 (6.9)8.1 (79.0)68.0 (24.3)12.1 (82.3)84.5 (9.8)15.4 (76.7)
+ +Table 1. Model's improved understanding of illusions with mixed training. Two prompts are used: "based on pixel value" and "based on human perception." The values represent the proportion of no-illusion responses, with parentheses indicating Human-like rates. + +![](images/50a24b23e665a3f59d4888f96941c609bc5b81745d38cc24e1de86d31aa4853a.jpg) + +![](images/c9e627f4c7f41615f5c80bb43f4e4dfdfd71a636cdf7147c7b83b0cecd41e091.jpg) +(a) Influence of background–foreground color difference on intensity of contrast illusion. + +![](images/08432bf5060508bbda552ffa5bc21c7b36bd3179fbf9481d5c08d7fe12a1e89d.jpg) +(b) Influence of image spatial frequency on intensity of stripe illusions. + +![](images/45078f9bf8f90dae33064fd5dcb8abbf06862f265d760cdf5504028573e30514.jpg) +Figure 7. Error rates of humans and LLaVA across different conditions in color illusions. + +![](images/f9b587f6424873e68977342d702bc6f0b3a91229aa9c5b68bd0b4e8b6c9b3719.jpg) + +by directly informing the model that the image contains an illusion without specifying its type; and providing few-shot examples, using no-illusion (NI) or human-like (HL) answers as 3-shot examples. Because we use the LLaVA-7B model here, which doesn't support multi-image input, we combine the few-shot examples and the query image into a single composite image before feeding it to the model [22]. + +Figure 8 show that external prompts have minimal effect on filter illusions. But for contrast and stripe illusions, the effect is more pronounced, particularly when no-illusion answers are used as few-shot examples, increasing the proportion of no-illusion responses by approximately $25\%$ . However, using human-like responses as few-shot examples does not effectively enhance the model's reproduction of human illusion perception, the proportion of human-like responses even decreases slightly for contrast illusions. Additionally, both direct emphasis and CoT prompts partially increase the proportion of no-illusion responses, al + +![](images/ec8c8951fe5d5517b09e1692cce15ed61c2ce4eb56123f0cb784cff0920dbac5.jpg) +Figure 8. Effect of different prompts and fine-tuning on VLM response preferences across illusion Types. "NI" denotes datasets using no-illusion examples, while "HL" represents datasets with human-like examples. + +though the effect of direct emphasis is relatively weak. + +Fine-Tuning. First, we fine-tune two separate LLaVA-7B models on illusion images, with one model trained with responses according to pixel values (NI), and the other trained on human-like responses (HL). The bottom two rows of Figure 8 show a significant effect on the test set, particularly when fine-tuning on no-illusion responses. For both contrast and stripe illusions, fine-tuning on no-illusion responses increases the proportion of no-illusion answers to over $90\%$ . In contrast, fine-tuning on human-like responses has a weaker effect, suggesting that even with human-like responses, the model struggles to fully grasp human perception in the context of illusion images. + +Next, we experiment with a mixed training approach that combines both types of responses. We train on question-answer pairs for both types of responses, where questions asking "Based on pixel values" are paired with pixel-value responses, those asking "Based on human perception" are paired with a human-like response. Table 1 shows the influence of this mixed training on the test set, compared to the base model. These results confirm that training model can effectively distinguish between pixel-based and human perception cues, yielding responses aligned with user's query. + +# 5. Why Color Illusions Affect VLMs + +# 5.1. Perceptual Biases from the Visual System + +To investigate whether VLMs' perceptual biases originate from their visual components, we test a range of purely visual models, including ResNet [18], ViT [10], VGG [35], + +![](images/06e5b4a86c0a5c9843ac7089a1fb783e7a949c705d731fa75ad0aa03f0b8fed5.jpg) +(a) Performance of purely vision models on non-illusion and illusion images. + +![](images/2f4753d4e6f317f040583acd4ada89f5305dfa03bbd7bd8a8bd76a11068c7579.jpg) +(b) Accuracy of ViT on non-illusion and illusion images over each training epochs. + +MobileNet [19] and Large Vision Model (LVM) [3]. We evaluate these models on both simple, rectangle-based contrast illusion images and generated realistic illusion images. For simple illusion images, we generate backgrounds with varying brightness on both sides and place a rectangle on each side. The vision model is trained as a three-way classifier on 6,000 non-illusion images to predict which rectangle appears darker, or if their colors are identical, stopping when the loss shows no significant decrease. It is then tested on 1,000 non-illusion and 1,000 illusion images. The results are shown in Figure 9a. For realistic illusion images, we train the model on the training set of our generated dataset and evaluate it on the test set, shown in the appendix. While overall accuracy (the proportion of classifications aligning with actual pixel values) on realistic illusion images drops compared to simple rectangle-based images, in both cases, accuracy on illusion images remains significantly lower than on non-illusion images. We also observe that during training, accuracy on non-illusion images quickly improves, while accuracy on illusion images fluctuates significantly. We use ViT as an example in Figure 9b. These results suggest that even purely vision-based models can be influenced by color illusions. + +# 5.2. Perceptual Biases from Prior Knowledge + +Influence of Language. We test the multilingual vision-language model PALO [32] on contrast illusions, posing questions in English, Chinese, and Russian to analyze error rates on images containing target regions across several colors. We hypothesize that colors with more frequent textual modifiers (e.g., dark blue, sky blue) lead to responses aligned with pixel values rather than human perception. To validate this, we first analyze color descriptor diversity in the training data. For example, in English, $50.7\%$ of red instances include modifiers like dark red or bright red, while $49.3\%$ use red alone. Figure 10 shows the correlation between color descriptor diversity and pixel-value accuracy for English, Chinese, and Russian. We find that richer descriptions improve model sensitivity to that color, reducing susceptibility to color illusions in that color range. + +![](images/8b7269aa4b36779b053b034e63ea25ae8a9124a0ba10e2a1700acc077a3257a0.jpg) +Figure 9. Perceptual biases in purely vision models. +Figure 10. Comparison of color term accuracy and frequency across languages for the vision-language model. Each shape and color represents a specific color term (e.g., blue, red, gray). + +Influence of Commonsense Knowledge. Humans are often deceived by visual illusions due to our prior knowledge. For instance, research suggests that in filter illusions, humans perceive colors that aren't actually present due to a phenomenon called chromatic adaptation [27]. This allows us to use surrounding context to infer the presence of a filter, mentally "removing" it to perceive the original color. One example is that we always recognize a piece of white paper as white, regardless of the lighting conditions. + +To examine whether commonsense knowledge similarly influences VLMs' susceptibility to illusions, we categorized filter illusion images into three groups: single objects, multiple-object compositions, and complex real-world scenes, then analyzed the deception rates for both humans and VLMs across these groups. On non-illusion images, VLMs perform well across all three categories, suggesting that scene complexity does not affect their basic color recognition ability. However, for illusion images, our results in Figure 11 show that both humans and VLMs are less likely to be deceived by single-object images, while complex scenes lead to the highest deception rates, underscoring the role of prior knowledge in visual illusions. + +![](images/e3251dc85eda73f9872f1c406931bba288ce32be8ba757a6cc8b12546a6151c9.jpg) +Figure 11. Deception rates of VLMs and human participants across different types of filter illusions. + +# 6. Conclusion + +Our experiments show that current VLMs, influenced by their visual systems and prior knowledge, can exhibit visual illusions similar to humans. This raises a key question: as models become more human-like, might they also inherit human perceptual biases? In fields like biomedicine [8, 21], such biases could result in flawed visual judgments, while in human-centric tasks like image generation, understanding human perception can benefit. Therefore, VLM behavior should be tailored to application needs, with careful consideration of human-like perceptual biases. + +# References + +[1] Adam Alter. Drunk tank pink: And other unexpected forces that shape how we think, feel, and behave. 2014. 2 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023.5 +[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. CVPR, 2024. 8 +[4] Claus-Christian Carbon. Understanding human perception by human-made illusions. Frontiers in human neuroscience, 2014. 1, 2 +[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: Endowing vision-language models with spatial reasoning capabilities. CVPR, 2024. 1 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. CVPR, 2024. 5 +[7] Fan L Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C Aoki, Jin Hirano, and Yukiyasu Kamitani. Reconstructing visual illusory experiences from human brain activity. Science Advances, 2023. 2 +[8] Hejie Cui, Lingjun Mao, Xin Liang, Jieyu Zhang, Hui Ren, Quanzheng Li, Xiang Li, and Carl Yang. Biomedical visual instruction tuning with clinician preference alignment. NeurIPS, 2024. 8 +[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 1, 5 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7 +[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. BLINK: Multimodal large language models can see but not perceive. ECCV, 2024. 5 +[12] Maurizio Gentilucci, Sergio Chieffi, Elena Daprati, M Cristina Saetti, and Ivan Toni. Visual illusion and action. Neuropsychologia, 1996. 1, 2 +[13] Ernst Hans Gombrich. Art and illusion: A study in the psychology of pictorial representation-millennium edition. 2023. 2 +[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. CVPR, 2019. 2 + +[15] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmio, and Jesus Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 2020. 2 +[16] Richard Langton Gregory. Perceptual illusions and brain models. Proceedings of the Royal Society of London. Series B. Biological Sciences, 1968. 2 +[17] Richard L Gregory. Visual illusions. Scientific American, 1968. 1, 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016. 7 +[19] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8 +[20] Yan Ke, Xiaou Tang, and Feng Jing. The design of high-level features for photo quality assessment. CVPR, 2006. 6 +[21] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. NeurIPS, 2023. 8 +[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeuRIPS, 2023. 1, 5, 7 +[23] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. CVPR, 2024. 1 +[24] Sichao Liu, Jianjing Zhang, Robert X. Gao, Xi Vincent Wang, and Lihui Wang. Vision-language model-driven scene understanding and robotic object manipulation. IEEE CASE, 2024. 1 +[25] Margaret S Livingstone. Art, illusion and the visual system. Scientific American, 1988. 2 +[26] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-IO 2: Scaling autoregressive multimodal models with vision, language, audio, and action. CVPR, 2023. 5 +[27] D. L. MacAdam. Chromatic adaptation*. J. Opt. Soc. Am., 1956. 8 +[28] O Maddaluno, A Facchin, D Zavagno, N Bolognini, R Daini, et al. Top-down and bottom-up neuromodulation over two different visual illusions. Atti del 40th European Conference on Visual Perception, 2018. 2 +[29] Glenn E Meyer and Susan Petry. Top-down and bottom-up: The illusory contour as a microcosm of issues in perception. 1987. 2 +[30] William L Phillips. Cross-cultural differences in visual perception of color, illusions, depth, and pictures. *Cross-Cultural Psychology: Contemporary Themes and Perspectives*, 2019. 2 +[31] Jiang Qiu, Hong Li, Qinglin Zhang, Qiang Liu, and Fenghua Zhang. The muller-lyer illusion seen by the brain: An event-related brain potentials study. Biological psychology, 2008. 2 + +[32] Hanoona Rasheed, Muhammad Maaz, Abdelrahman Shaker, Salman Khan, Hisham Cholakal, Rao M. Anwer, Tim Baldwin, Michael Felsberg, and Fahad S. Khan. Palo: A large multilingual multimodal language model. WACV, 2025. 8 +[33] Alireza Saeedi, Kun Wang, Ghazaleh Nikpourian, Andreas Bartels, Nikos K Logothetis, Nelson K Totah, and Masataka Watanabe. Brightness illusions drive a neuronal response in the primary visual cortex under top-down modulation. Nature Communications, 2024. 2 +[34] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 1, 2 +[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2015. 7 +[36] Zineng Tang, Lingjun Mao, and Alane Suhr. Grounding language in multi-perspective referential communication. EMNLP, 2024. 1 +[37] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? exploring the visual shortcomings of multimodal llms. CVPR, 2024. 5 +[38] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. PMLR, 2022. 5 +[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogVLM: Visual expert for pretrained language models. NeuRIPS, 2024. 1, 5 +[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. NeuRIPS, 2022. 1 +[41] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. arXiv preprint arXiv:2304.00685, 2024. 1 +[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. CVPR, 2023. 1 +[43] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding Visual Illusions in Language: Do vision-language models perceive illusions like humans? EMNLP, 2023. 1, 2, 5 +[44] Zeliang Zhang, Xin Liang, Mingqian Feng, Susan Liang, and Chenliang Xu. Will the inclusion of generated data amplify bias across generations in future image classification models? arXiv preprint arXiv:2410.10160, 2024.1 \ No newline at end of file diff --git a/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/images.zip b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2c4971350f26282c48f9c19c874bda33978ff128 --- /dev/null +++ b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4268bd102f30a7ceb756ceafc19b42d125039a6480ac2d78406d2e544c57eb0e +size 573081 diff --git a/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/layout.json b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f4b8b11304b84bd3172868d40c3e7daf174acad --- /dev/null +++ b/2025/Evaluating Model Perception of Color Illusions in Photorealistic Scenes/layout.json @@ -0,0 +1,7958 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 524, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 524, + 120 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 524, + 120 + ], + "type": "text", + "content": "Evaluating Model Perception of Color Illusions in Photorealistic Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 143, + 432, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 143, + 432, + 186 + ], + "spans": [ + { + "bbox": [ + 178, + 143, + 432, + 186 + ], + "type": "text", + "content": "Lingjun Mao Zineng Tang Alane Suhr University of California, Berkeley {lingjun, terran, suhr}@berkeley.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 297, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 395 + ], + "type": "text", + "content": "We study the perception of color illusions by vision-language models. Color illusion, where a person's visual system perceives color differently from actual color, is well-studied in human vision. However, it remains underexplored whether vision-language models (VLMs), trained on large-scale human data, exhibit similar perceptual biases when confronted with such color illusions. We propose an automated framework for generating color illusion images, resulting in RCID (Realistic Color Illusion Dataset), a dataset of 19,000 realistic illusion images. Our experiments show that all studied VLMs exhibit perceptual biases similar human vision. Finally, we train a model to distinguish both human perception and actual pixel differences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 417, + 135, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 135, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 135, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 700 + ], + "type": "text", + "content": "Human visual perception does not always align perfectly with the real world [4, 12, 17]; we are easily influenced by contextual factors such as lighting conditions or surrounding objects and patterns. Figure 1 shows several examples of color illusions, where human judgment of object color and brightness does not align with actual pixel values. Currently, many large vision-language models (VLMs) [9, 22, 23, 39] exhibit behavioral similarities to human visual perception in various aspects, including object recognition [41, 44], scene understanding [24, 36], and spatial reasoning [5]. But because the human visual system behaves idiosyncratically when encountering these illusions, we aim to study whether these models perceive such illusions in the same way humans do. Prior work [34, 43] has explored this question by sourcing illusion images from the internet. However, this approach has a drawback: the number of illusion images on the web is limited, and most images (e.g., " + }, + { + "bbox": [ + 55, + 437, + 295, + 700 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 437, + 295, + 700 + ], + "type": "text", + "content": " for the IllusionVQA dataset [34]) are well-known examples of these illusions; thus, VLMs have likely memorized humanlike behavioral responses to them. Additionally, the limited scale also restricts the depth and variety of analyses that can be conducted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "To address these limitations, we propose an automatic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "content": "framework for generating realistic color illusion images, covering many scenarios with illusions that we might genuinely encounter in our real life. We cover three different types of color illusions: contrast (Figure 1a), stripe (Figure 1b), and filter (Figure 1c) illusions. For contrast and stripe illusions, we first generate simple arrangements of shapes exhibiting specific illusions, which are then processed through ControlNet [42] to generate more realistic images. For filter illusions, we directly apply filters to the original MS COCO 2017 images. In the end, we produce a dataset consisting of total 19,000 images, half of which include an illusion, and the other half being a control set without illusions. Each image is paired with a synthetically-generated natural language question, asking either about color differences between specific regions or directly inquiring about the color of a particular object in the image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 556, + 602 + ], + "type": "text", + "content": "We test recent VLMs using our generated illusion images. We find that, after appropriate fine-tuning, they perform well on non-illusion images, but their accuracy drops significantly on illusion images. Some of their responses exhibit visual biases similar to those of humans, while some are entirely incorrect, failing to align with either pixel values or human perception. We also observe that external prompts, such as Chain of Thought (CoT) [40] and few-shot examples, can influence VLM performance to some extent. Furthermore, fine-tuning on illusion images enhances the models' ability to understand these color illusions. Based on this, we propose a foundational baseline method that enables models to capture both human perceptual biases and true pixel values. Finally, we investigate the underlying causes of VLM susceptibility to color illusions, including biases from both the visual system and prior knowledge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "In conclusion, our contributions are as follows: (1) We propose an automated framework for generating realistic illusion images and create a large, realistic dataset of color illusion images, named RCID (Realistic Color Illusion Dataset), to enhance the fairness and accuracy of model testing. (2) We investigate underlying mechanisms of color illusions in VLMs, highlighting the combined influence of the vision system and prior knowledge. We also explore how external prompts and instruction tuning" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7805" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 216, + 171 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 216, + 171 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 216, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 216, + 171 + ], + "type": "image", + "image_path": "93d6999a295cd279ef6593dd55a3802f5f6ba1d2704828692f34cbcdc220645c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 174, + 215, + 208 + ], + "lines": [ + { + "bbox": [ + 59, + 174, + 215, + 208 + ], + "spans": [ + { + "bbox": [ + 59, + 174, + 215, + 208 + ], + "type": "text", + "content": "Q1: Based on pixel values, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A1: III" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 104, + 254, + 170, + 263 + ], + "lines": [ + { + "bbox": [ + 104, + 254, + 170, + 263 + ], + "spans": [ + { + "bbox": [ + 104, + 254, + 170, + 263 + ], + "type": "text", + "content": "(a) Contrast Illusion" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 383, + 171 + ], + "blocks": [ + { + "bbox": [ + 59, + 214, + 213, + 248 + ], + "lines": [ + { + "bbox": [ + 59, + 214, + 213, + 248 + ], + "spans": [ + { + "bbox": [ + 59, + 214, + 213, + 248 + ], + "type": "text", + "content": "Q2: According to human perception, how do the colors of the walls on the left and right of the baseball player compare? Here are some options you can choose: I. A is darker II. B is darker III. They are exactly same A2: II" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 227, + 72, + 383, + 171 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 383, + 171 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 383, + 171 + ], + "type": "image", + "image_path": "83489c9454b9cd8d35735ada0a6d705d43e9b67c7e93f4a90ae96e77062ad520.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 175, + 380, + 208 + ], + "lines": [ + { + "bbox": [ + 229, + 175, + 380, + 208 + ], + "spans": [ + { + "bbox": [ + 229, + 175, + 380, + 208 + ], + "type": "text", + "content": "Q1: Evaluated by pixel values, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A1: III" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 228, + 214, + 380, + 247 + ], + "lines": [ + { + "bbox": [ + 228, + 214, + 380, + 247 + ], + "spans": [ + { + "bbox": [ + 228, + 214, + 380, + 247 + ], + "type": "text", + "content": "Q2: According to human perception, how do the colors of the pink walls on the left and right compare? Here are some options you can choose: I. Left is darker II. Right is darker III. They are exactly same A2: II" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 276, + 254, + 334, + 264 + ], + "lines": [ + { + "bbox": [ + 276, + 254, + 334, + 264 + ], + "spans": [ + { + "bbox": [ + 276, + 254, + 334, + 264 + ], + "type": "text", + "content": "(b) Stripe Illusion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 551, + 171 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 551, + 171 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 551, + 171 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 551, + 171 + ], + "type": "image", + "image_path": "11fdeb0e7c043d2cca6d53d5d537103304357b4f89412e1fb5859ac1236715bd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 174, + 550, + 190 + ], + "lines": [ + { + "bbox": [ + 397, + 174, + 550, + 190 + ], + "spans": [ + { + "bbox": [ + 397, + 174, + 550, + 190 + ], + "type": "text", + "content": "Q1: According to pixel values, what color is the athlete's uniform?" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 398, + 190, + 536, + 206 + ], + "lines": [ + { + "bbox": [ + 398, + 190, + 536, + 206 + ], + "spans": [ + { + "bbox": [ + 398, + 190, + 536, + 206 + ], + "type": "text", + "content": "A1: ['Gray', 'Yellowish-green', 'Greenish-yellow', 'Green']" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 398, + 213, + 537, + 229 + ], + "lines": [ + { + "bbox": [ + 398, + 213, + 537, + 229 + ], + "spans": [ + { + "bbox": [ + 398, + 213, + 537, + 229 + ], + "type": "text", + "content": "Q2: Through human eyes, what color is the athlete's uniform?" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 399, + 230, + 431, + 238 + ], + "lines": [ + { + "bbox": [ + 399, + 230, + 431, + 238 + ], + "spans": [ + { + "bbox": [ + 399, + 230, + 431, + 238 + ], + "type": "text", + "content": "A2: ['Red']" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 445, + 254, + 501, + 263 + ], + "lines": [ + { + "bbox": [ + 445, + 254, + 501, + 263 + ], + "spans": [ + { + "bbox": [ + 445, + 254, + 501, + 263 + ], + "type": "text", + "content": "(c) Filter Illusion" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 270, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 315 + ], + "type": "text", + "content": "Figure 1. Examples of color illusions. (a) Contrast illusion: The two green squares on the left and right are identical in color, but most people perceive the square on the right as darker due to the brighter background. (b) Stripe illusion: The left and right walls seem to be different colors, but this is an illusion caused by the dark stripes interfering with our perception. (c) Filter illusion: The player's uniform appears red, but in reality, there are no red pixels present." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 333, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 405 + ], + "type": "text", + "content": "impact the models' performance on these illusions. (3) We propose a simple training method that enables models to understand human perception while also recognizing the actual pixel values. Our code and data are released under an open-source license at the following URL: https://github.com/mao1207/RCID." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 417, + 142, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 142, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 142, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 437, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 641 + ], + "type": "text", + "content": "Visual illusion is a phenomenon where human perception differs from the actual environment [4, 12, 17], often triggered by specific visual stimuli [7]. These illusions offer valuable insights into human perception [16, 31] and have been widely used in artworks. [13, 25]. Many studies [28, 29] have attempted to explore the causes of visual illusions, with the bottom-up and top-down theories being the most widely accepted explanations. Proponents [33] of the bottom-up theory argue that low-level visual signals, such as contrast, brightness, edges, and motion, are correlated with human perception of illusions. However, supporters of the top-down theory contend that prior knowledge is also a major factor in these illusions. Evidence shows that people in industrialized societies are more susceptible to the Müller-Lyer illusion due to being misled by implied arrow meanings, and individuals from language communities with fewer color terms are less affected by color illusions [1, 30]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Motivated by illusion phenomena in human visual perception, recent research has begun to explore whether models are also susceptible to such illusions. Early research [14, 15] suggests that traditional vision models, such as CNNs, are influenced by visual illusions in a way similar to humans. More recent studies [34, 43] have shown" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "content": "that VLMs also exhibit similar perceptual biases. However, images on which these models have been tested are either sourced from the internet or manual compositions of simple shapes (e.g., two squares of different colors), making it easy for the VLM to infer results based on their knowledge rather than actual observation. Additionally, the number of these images is quite limited. To address these limitations, we develop a large-scale dataset that embeds visual illusions into realistic scenes, enabling robust evaluation in complex, real-world contexts. Our study fills a gap by investigating the mechanisms behind VLMs' susceptibility to illusions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 475, + 480, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 480, + 488 + ], + "type": "text", + "content": "3. Illusion Dataset Construction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 496, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 665 + ], + "type": "text", + "content": "The construction of our dataset involves three steps (Figure 2): (1) Image Generation. For contrast and stripe illusions, we use procedural code to generate simple illusion images, which are then processed by ControlNet to create realistic illusion images. For filter illusions, we directly apply contrasting color filters to the original images. Each type of illusion also includes a corresponding control group without any illusions for comparison. (2) Question Generation. We use GPT-4o to generate image-specific questions that are designed to evaluate the model's understanding of the illusion. (3) Human Feedback. We collect human participants' feedback on these images and adjust the original classification of \"illusion\" and \"non-illusion\" based on whether participants are deceived." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 671, + 555, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 555, + 685 + ], + "type": "text", + "content": "3.1. Contrast and Stripe Illusion Image Generation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "ControlNet Training Stage: We begin by training a ControlNet to map from simple illusion images to photoreal-" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7806" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 383, + 296 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 383, + 296 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 383, + 296 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 383, + 296 + ], + "type": "image", + "image_path": "0d37c77073981aa85fc1557a1c734a0403527d23a35aac9b5febcd6cd5f0c6cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 307, + 385, + 319 + ], + "lines": [ + { + "bbox": [ + 223, + 307, + 385, + 319 + ], + "spans": [ + { + "bbox": [ + 223, + 307, + 385, + 319 + ], + "type": "text", + "content": "Figure 2. Process for generating our dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 385, + 72, + 527, + 296 + ], + "blocks": [ + { + "bbox": [ + 385, + 72, + 527, + 296 + ], + "lines": [ + { + "bbox": [ + 385, + 72, + 527, + 296 + ], + "spans": [ + { + "bbox": [ + 385, + 72, + 527, + 296 + ], + "type": "image", + "image_path": "11b47c1997316f14a1763f4f25a9c1a97c24261663e3a41841ce5955f18cfa6a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "text", + "content": "istic images with illusions. We train using modified image-caption pairs " + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "inline_equation", + "content": "(I, T)" + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "text", + "content": " from the MS COCO 2017 dataset. First, we quantize each original image " + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "inline_equation", + "content": "10 \\times 10" + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "text", + "content": " grid " + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "text", + "content": ", where each cell " + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "inline_equation", + "content": "G_{x,y}" + }, + { + "bbox": [ + 55, + 328, + 296, + 389 + ], + "type": "text", + "content": " represents the average color of its corresponding region:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 395, + 227, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 395, + 227, + 422 + ], + "spans": [ + { + "bbox": [ + 123, + 395, + 227, + 422 + ], + "type": "interline_equation", + "content": "G _ {x, y} = \\frac {1}{\\left| R _ {x , y} \\right|} \\sum_ {p \\in R _ {x, y}} I (p)", + "image_path": "a79e44eec1488350516237f7f5deb309e9ef65297a06be1c6491d3ab6da93cb7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "inline_equation", + "content": "R_{x,y}" + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "text", + "content": " denotes the set of pixels in the " + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "text", + "content": "-th grid cell, and " + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "inline_equation", + "content": "I(p)" + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "text", + "content": " is the color value of pixel " + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 425, + 295, + 450 + ], + "type": "text", + "content": " in the original image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": "During training, Gaussian noise is added to the quantized image " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "x_0 = G" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": ", producing a noisy image " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": ". The model, conditioned on the text prompt " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": " and color distribution " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": ", is trained to predict the added noise " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": " by minimizing the expected squared error:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 518, + 261, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 518, + 261, + 533 + ], + "spans": [ + { + "bbox": [ + 88, + 518, + 261, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {x _ {0}, t, C, \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, t, T, G\\right) \\| _ {2} ^ {2} \\right]", + "image_path": "cab3370f0da45243b4127a359486150fef077c749caba7ef4f1f58b6ce838d19.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "content": "Through this loss, the model learns to progressively refine " + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "content": " into a realistic image during inference, guided by color distribution and the text prompt." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "Procedural Generation Stage: We use a procedural function " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " to generate simple illusion images, denoted as " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": ", which consist of basic shapes and colors, on which we will condition the trained ControlNet model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "For Contrast Illusions, we randomly select two base colors from RGB space: " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " for the background and " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " for the foreground (squares). For a color " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "C = (r, g, b)" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": ", we use the function " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " to adjust its brightness by " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(C, \\mu) = (r \\cdot \\mu, g \\cdot \\mu, b \\cdot \\mu)" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " scales each RGB component. The background is divided into a darker region " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(C_b, \\mu_{b1})" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " and a brighter region " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(C_b, \\mu_{b2})" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{b1} < 1" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{b2} > 1" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": ". Additionally, two foreground squares, colored " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(C_f, \\mu_{f1})" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "inline_equation", + "content": "p(C_f, \\mu_{f2})" + }, + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "text", + "content": ", are symmetrically placed at random positions " + }, + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "inline_equation", + "content": "(x_1, y_1)" + }, + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "inline_equation", + "content": "(x_2, y_2)" + }, + { + "bbox": [ + 313, + 328, + 553, + 352 + ], + "type": "text", + "content": " on the background." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "text", + "content": "If human perception differs from the actual pixel values, the image is classified as an illusion; otherwise, it belongs to the control group (no illusion). We observe that a darker background enhances the perceived brightness of the foreground color, with this effect being strongest when " + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "text", + "content": " are similar and weakening as the contrast increases. Section 4.2.3 validates this conclusion. Based on this principle, we generate theoretical illusion and non-illusion images, which are later adjusted through human validation. To increase image diversity, we apply random noise " + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "text", + "content": " to introduce irregularities at the square edges or make the boundaries less smooth. The generated image " + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 313, + 353, + 555, + 508 + ], + "type": "text", + "content": " for the contrast illusion can be represented as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 523, + 537, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 523, + 537, + 550 + ], + "spans": [ + { + "bbox": [ + 330, + 523, + 537, + 550 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} S = f \\left(p \\left(C _ {b}, \\mu_ {b 1}\\right), p \\left(C _ {b}, \\mu_ {b 2}\\right), p \\left(C _ {f}, \\mu_ {f 1}\\right), p \\left(C _ {f}, \\mu_ {f 2}\\right), \\right. \\\\ \\left(x _ {1}, y _ {1}\\right), \\left(x _ {2}, y _ {2}\\right), \\eta) \\\\ \\end{array}", + "image_path": "8e1eb78e1a683d9a61263024dc4dab9dc8dc0f9fbcb3f5ce44b1918491893fe5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": "Similarly, for Stripe Illusions, we randomly generate the background color " + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": " and the stripe color " + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "inline_equation", + "content": "C_s" + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": ". The colored stripes alternate with black stripes, and their colors are adjusted to " + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "inline_equation", + "content": "p(C_s, \\mu_{s1})" + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "inline_equation", + "content": "p(C_s, \\mu_{s2})" + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": ". Based on an initial estimate of human perception of " + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mu_{s1}" + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mu_{s2}" + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": ", the images are classified as either illusion or non-illusion. The generated image " + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": " for the stripe illusion is represented as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 650, + 515, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 650, + 515, + 662 + ], + "spans": [ + { + "bbox": [ + 352, + 650, + 515, + 662 + ], + "type": "interline_equation", + "content": "S = f \\left(C _ {b}, p \\left(C _ {s}, \\mu_ {s 1}\\right), p \\left(C _ {s}, \\mu_ {s 2}\\right), \\theta , N, \\eta\\right)", + "image_path": "7d988f3028baf59a7f2961d77c64b728971af632351d01743a95e946332916ff.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Here the stripe direction " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " (horizontal, vertical, or diagonal) and the number of stripes " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " are randomly selected. Random noise " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " is applied to create irregularities, such as curved stripe arrangements or slight misalignments." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7807" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 73, + 242, + 259 + ], + "blocks": [ + { + "bbox": [ + 83, + 73, + 242, + 259 + ], + "lines": [ + { + "bbox": [ + 83, + 73, + 242, + 259 + ], + "spans": [ + { + "bbox": [ + 83, + 73, + 242, + 259 + ], + "type": "image", + "image_path": "2b1589ddc6a1a224083e9820436a3e05b1b5aab15edecaaa22abc63b166d0f75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 270, + 427, + 281 + ], + "lines": [ + { + "bbox": [ + 181, + 270, + 427, + 281 + ], + "spans": [ + { + "bbox": [ + 181, + 270, + 427, + 281 + ], + "type": "text", + "content": "Figure 3. Data statistics of RCID (Realistic Color Illusion Dataset)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 248, + 74, + 402, + 258 + ], + "blocks": [ + { + "bbox": [ + 248, + 74, + 402, + 258 + ], + "lines": [ + { + "bbox": [ + 248, + 74, + 402, + 258 + ], + "spans": [ + { + "bbox": [ + 248, + 74, + 402, + 258 + ], + "type": "image", + "image_path": "9777cdd4936f9d48191a56a9bfdb9601251f91584ffe8331bd541006a2e95c69.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 408, + 76, + 523, + 258 + ], + "blocks": [ + { + "bbox": [ + 408, + 76, + 523, + 258 + ], + "lines": [ + { + "bbox": [ + 408, + 76, + 523, + 258 + ], + "spans": [ + { + "bbox": [ + 408, + 76, + 523, + 258 + ], + "type": "image", + "image_path": "ad33710260e74fbb7719319735fc58980bd3a1623c78febbb896824569eb0bb8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "content": "ControlNet Generating Stage: We pair each simple illusion image " + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "content": " with a randomly selected caption " + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "content": " from the MS COCO 2017 dataset, which together serve as input to our model. Starting from the quantized image " + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "inline_equation", + "content": "x_0 = S" + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "content": ", the model recursively denoises over multiple time steps following a diffusion-based approach, gradually refining the image to produce a realistic output " + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "content": " that retains the intended illusion effects. The final image " + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "content": " is obtained by summing the contributions from all time steps " + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 291, + 296, + 399 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 407, + 234, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 407, + 234, + 436 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 234, + 436 + ], + "type": "interline_equation", + "content": "I ^ {\\prime} = \\sum_ {t = 0} ^ {T} \\left(x _ {t} + \\epsilon_ {\\theta} \\left(x _ {t}, t, C, S\\right)\\right)", + "image_path": "0403cb44aa52c646143b06ebb304615234a656cea1776f922534e7c865a4c73c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 443, + 230, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 230, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 230, + 455 + ], + "type": "text", + "content": "3.2. Filter Illusion Image Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "text", + "content": "To generate filter illusion images, we first select images from MS COCO 2017 that predominantly contain a specific color " + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "text", + "content": " (e.g. red, blue, yellow). We then apply a contrasting color filter " + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "text", + "content": " which overlays the original image in the HSV (Hue, Saturation, Value) color space to suppress " + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "text", + "content": " by shifting the hue " + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "text", + "content": " of filter and ensuring no pixels fall within " + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 460, + 296, + 557 + ], + "type": "text", + "content": "'s HSV range. Despite this, the filtered regions may still visually evoke the original color." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 564, + 176, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 176, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 176, + 577 + ], + "type": "text", + "content": "3.3. Question Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "For each image " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", we use GPT-4o to generate an image-grounded question " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " that asks human respondents to compare colors between two regions or identify the color of a specific object. For example, in Figure 1a, the generated questions refer to the walls on the left and right of the baseball player. We then prefix the question with two distinct prompts: one focused on \"human perception\" (e.g., \"Based on human perception\", \"Through human eye\") and the other on \"pixel values\", each associated with a different answer depending on whether the image contains an illusion. For contrast and stripe illusions, answer options are appended" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 291, + 512, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 512, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 512, + 304 + ], + "type": "text", + "content": "to the question. Examples are shown in Figure 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 313, + 488, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 488, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 488, + 325 + ], + "type": "text", + "content": "3.4. Human Validation and Selection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 464 + ], + "type": "text", + "content": "To ensure that our generated images truly contain illusions, we conduct a survey on Prolific with a total of 241 participants. Each participant is shown 400 images, including both illusion and non-illusion examples. We use respondents' answers as the human perception ground truth. For every image in our dataset, we assign 5 participants. If more than 3 participants are deceived by the illusion, the image is classified as an illusion image. Conversely, if all 5 participants provide answers consistent with the pixel values, the image is classified as non-illusion and added to the control group. On average, we pay $21 USD per hour." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 407, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 407, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 407, + 486 + ], + "type": "text", + "content": "3.5. Dataset Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "text", + "content": "Realistic Color Illusion Dataset (RCID, Figure 3) contains 19,000 images, comprising 8,000 contrast, 9,000 stripe, and 2,000 filter images. We divide our dataset into 9,500 images for training, 4,750 for development, and the remaining 4,750 for the test set. Each illusion type has distinct structures and subcategories." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 578, + 515, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 515, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 515, + 591 + ], + "type": "text", + "content": "4. Evaluating Vision-Language Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 598, + 430, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 430, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 430, + 611 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 616, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 676 + ], + "type": "text", + "content": "In this experiment, we evaluate different models' performance on color illusions and explore the impact of external prompts and instruction-tuning on them. We also provide a foundational baseline model based on LLaVA-1.5 (7B) for understanding color illusions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Vision-Language Models. We test recent models with our generated illusion dataset to examine whether VLMs exhibit perceptual responses similar to humans. To ensure" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7808" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 527, + 264 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 527, + 264 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 527, + 264 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 527, + 264 + ], + "type": "image", + "image_path": "8945c71c48938d8c53e3b6a9d27721ab28facce79191c992e7a121126b1f2dbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 274, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 319 + ], + "type": "text", + "content": "Figure 4. This figure shows the proportion of different model responses across three types of illusions (Contrast, Filter, and Stripe) on our development set. For non-illusion images, we report the proportions of \"Accurate\" and \"Wrong\" responses. For illusion images, we categorize responses as \"No Illusion\" (consistent with pixel values), \"Human like,\" and \"N/A.\" Each image is evaluated using two types of prompts: one based on pixel values and the other based on human perception." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 331, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 571 + ], + "type": "text", + "content": "that their performance on illusion images reflects actual susceptibility to illusions rather than challenges in identifying specified regions and comparing color brightness [11, 37], we focus on open-source models, and fine-tune them on our generated non-illusion simple images and a small set of realistic images to enhance their ability in answering questions about relative colors of different regions in an image. During training, we randomly apply prompts such as 'based on human perception,' 'based on pixel values,' or provide no specific guidance prompt at all. Our models tested include LLaVA [22], InstructBLIP [9], CogVLM [39], InternVL2 [6], Qwen-VL2 [2], OFA [38] and Unified-IO [26]. During inference for contrast and stripe illusions, questions are presented in a multiple-choice format, while filter illusion questions require open-ended responses about the color of specific objects in the image, as illustrated in Figure 1. We also include experiments with closed-source models like GPT-4o and Gemini in the appendix. While their overall color comparison ability is weak, there is still a significant drop in performance on illusion images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Metrics. For images containing illusions, we use the same metrics as in GVIL [43]. We track the Human-like rate, which measures alignment between human perception and VLM responses by calculating the percentage of cases where the model's responses match human answers. When the prompt explicitly includes \"Based on pixel values,\" this can also be referred to as Deception Rate. The No-Illusion rate indicates the proportion of examples where the model's responses align with actual pixel values. If the model's answer aligns with neither the human response nor the pixel values, we categorize those instances as Not Applicable" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 331, + 555, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 367 + ], + "type": "inline_equation", + "content": "(N / A)" + }, + { + "bbox": [ + 313, + 331, + 555, + 367 + ], + "type": "text", + "content": ". For images without illusions, where human perception matches pixel values, we classify model responses as \"accurate\" or \"inaccurate\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 373, + 434, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 434, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 434, + 385 + ], + "type": "text", + "content": "4.2. Results and Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "type": "text", + "content": "We evaluate a range of open-source vision-language models using generated illusion and non-illusion images in our development set, with questions explicitly asking for color judgments 'Based on pixel values' or 'Based on human perception'. The results show that, after fine-tuning on non-illusion images, these models achieve high accuracy " + }, + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "(>75\\%)" + }, + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "type": "text", + "content": " on non-illusion images, while their accuracy on illusion images is significantly lower. We find that explicitly querying for color judgments 'Based on pixel values' or 'Based on human perception' does not lead to significant changes in model performance. In addition, models are likely to produce responses that are completely inaccurate, matching neither pixel values nor human perception. This suggests that while the models are misled by color illusions to some extent, they still struggle to fully model human perception. Detailed results can be found in Figure 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Inter-Annotator Agreement. We refine our final set of illusion images through human validation, retaining only those that deceive at least 3 out of 5 human participants. Before this selection, the average Fleiss' kappa score among the five respondents on the illusion images is 0.648, which increases to 0.806 after filtering. " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "21.6\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " of the selected images deceive all five participants. We use this stricter subset of illusion images as a challenging test set for our models. The results show a slight increase " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "(< 5\\%)" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " in the proportion of Human-Like responses by the VLMs across all three types of illusions, but overall, the difference remains" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7809" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 71, + 281, + 149 + ], + "blocks": [ + { + "bbox": [ + 84, + 71, + 281, + 149 + ], + "lines": [ + { + "bbox": [ + 84, + 71, + 281, + 149 + ], + "spans": [ + { + "bbox": [ + 84, + 71, + 281, + 149 + ], + "type": "image", + "image_path": "8622480d8734251f951485471d565e8d726bb7a4498f067dba62334aa2997831.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 158, + 455, + 170 + ], + "lines": [ + { + "bbox": [ + 153, + 158, + 455, + 170 + ], + "spans": [ + { + "bbox": [ + 153, + 158, + 455, + 170 + ], + "type": "text", + "content": "Figure 5. Deception rates of humans and VLMs across different structural patterns." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 525, + 149 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 525, + 149 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 525, + 149 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 525, + 149 + ], + "type": "image", + "image_path": "db77b3c389d0b6826f4cde811605a978ee0c6d0bc2655b118c8b745f8a372b19.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 295, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 203 + ], + "type": "text", + "content": "minimal, indicating that our filtering criterion (more than 3 humans are deceived) is robust." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 204, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 204, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 295, + 287 + ], + "type": "text", + "content": "Generated Image Realism. As some generated illusion images appear realistic while others still exhibit an \"AI-looking\" appearance, we assess LLaVA-7B and two human participants on three 500-image sets with varying realism levels. Our results show no significant correlation between realism and accuracy, with " + }, + { + "bbox": [ + 54, + 204, + 295, + 287 + ], + "type": "inline_equation", + "content": "r = 0.084" + }, + { + "bbox": [ + 54, + 204, + 295, + 287 + ], + "type": "text", + "content": " for human participants and " + }, + { + "bbox": [ + 54, + 204, + 295, + 287 + ], + "type": "inline_equation", + "content": "r = 0.091" + }, + { + "bbox": [ + 54, + 204, + 295, + 287 + ], + "type": "text", + "content": " for the model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 288, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 296, + 420 + ], + "type": "text", + "content": "Model Size. To investigate the effect of model size, we evaluate the performance of two groups of models, OFA and Unified-IO. These groups consist of models of varying sizes but share the same overall architecture. We conduct the experiment on contrast illusions, and explicitly emphasize \"Based on pixel values\" in the questions. Our result shows, with the model size increases, the proportion of human-like responses also increases, while the proportion of responses consistent with actual pixel values decreases. We present OFA's performance in Figure 6 and provide Unified-IO's performance details in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 85, + 430, + 264, + 536 + ], + "blocks": [ + { + "bbox": [ + 85, + 430, + 264, + 536 + ], + "lines": [ + { + "bbox": [ + 85, + 430, + 264, + 536 + ], + "spans": [ + { + "bbox": [ + 85, + 430, + 264, + 536 + ], + "type": "image", + "image_path": "216538a85f9186ed5fe6dfc5a8cc2d33237d1a0d21d82515e2e5e2f2846ba7ca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 539, + 295, + 562 + ], + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 562 + ], + "type": "text", + "content": "Figure 6. Proportions of 'No Illusion,' 'Human Like,' and 'N/A' responses for OFA models of different sizes on contrast illusions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": "Illusion Design. We explore a range of visual factors that may influence the strength of color illusions and compare whether these factors affect human perception and VLMs in the same way. We focus on three potential influencing factors: the orientation of the illusion, the contrast between foreground and background colors (for contrast illusions only), and the number of stripes (for filter illusions only). Overall, our findings indicate that these factors significantly impact the strength of the illusion. For example, altering certain factors, such as increasing the color contrast between the foreground and background, can turn a non-illusion image into an illusion image. And these effects are" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "content": "consistent across both humans and VLMs. In these experiments, we use LLaVA-1.5 (7B) as our tested model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 373 + ], + "type": "text", + "content": "Illusion orientation. We randomly generate a set of images with various structural arrangements to analyze how the location of the illusion affects humans and VLMs. For contrast illusions, the images feature left-right and up-down arrangements, while for stripe illusions, they include horizontal, vertical, and diagonal stripes (Figure 3). We evaluate the error rates of both humans and VLMs across these structures. As illustrated in Figure 5, the results show that for contrast illusions, both humans and VLMs are more frequently deceived by left-right arrangements, while for stripe illusions, they are more often misled by horizontal and diagonal stripes. We find that humans can quickly and accurately perceive color differences between left and right, but are slightly less sensitive to top-bottom differences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 556, + 483 + ], + "type": "text", + "content": "Color distance between the foreground and background. For contrast illusions, we calculate the correlation between the color distance in RGB space between the foreground and background, and the illusion's intensity. Figure 7a shows that as the foreground and background colors become more similar, the contrast illusion becomes most pronounced for both humans and VLMs. Conversely, when the foreground and background colors are distinctly different, humans and VLMs are better at distinguishing them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "text", + "content": "Image spatial frequency. For stripe illusions, we examine the effect of image spatial frequency on the strength of the illusion. After converting the image to the frequency domain using the Fourier transform, we compute spatial frequency as the ratio of high-frequency energy to the total image energy [20]. Our findings suggest that the stripe illusion becomes stronger for both humans and models as image frequency increases, which typically corresponds to a higher number of stripes, as shown in Figure 7b." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Prompting Methods. We explore whether external prompts could alter VLM responses. In contrast to our main experiments shown in Figure 4, here we aim to identify the model's bias towards predicting answers which are human-like or based on pixel values without specifying this in the question posed to the model. We consider several prompting methods: a simple prompt that only contains questions about color comparisons; chain of thought (CoT), where the model is prompted to first consider if any factors in the image might affect color perception; emphasizing the illusion" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7810" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 135 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 135 + ], + "type": "table", + "html": "
Illusion CategoryContrast IllusionFilter IllusionStripe Illusion
PromptsPixel-BasedHuman-BasedPixel-BasedHuman-BasedPixel-BasedHuman-Based
LLaVA-1.5 (7B)35.4 (41.2)32.4 (37.4)20.0 (67.8)19.8 (64.6)30.3 (35.2)28.3 (40.1)
Mixed Training (Ours)83.9 (6.9)8.1 (79.0)68.0 (24.3)12.1 (82.3)84.5 (9.8)15.4 (76.7)
", + "image_path": "aeeb23cc6577ca98251b59e994562fc64513269902c938a70aaf62b3c139331b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 142, + 555, + 164 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 555, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 555, + 164 + ], + "type": "text", + "content": "Table 1. Model's improved understanding of illusions with mixed training. Two prompts are used: \"based on pixel value\" and \"based on human perception.\" The values represent the proportion of no-illusion responses, with parentheses indicating Human-like rates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 93, + 171, + 253, + 293 + ], + "blocks": [ + { + "bbox": [ + 93, + 171, + 253, + 293 + ], + "lines": [ + { + "bbox": [ + 93, + 171, + 253, + 293 + ], + "spans": [ + { + "bbox": [ + 93, + 171, + 253, + 293 + ], + "type": "image", + "image_path": "50a24b23e665a3f59d4888f96941c609bc5b81745d38cc24e1de86d31aa4853a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 93, + 316, + 257, + 408 + ], + "blocks": [ + { + "bbox": [ + 91, + 297, + 259, + 315 + ], + "lines": [ + { + "bbox": [ + 91, + 297, + 259, + 315 + ], + "spans": [ + { + "bbox": [ + 91, + 297, + 259, + 315 + ], + "type": "text", + "content": "(a) Influence of background–foreground color difference on intensity of contrast illusion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 316, + 257, + 408 + ], + "lines": [ + { + "bbox": [ + 93, + 316, + 257, + 408 + ], + "spans": [ + { + "bbox": [ + 93, + 316, + 257, + 408 + ], + "type": "image", + "image_path": "c9e627f4c7f41615f5c80bb43f4e4dfdfd71a636cdf7147c7b83b0cecd41e091.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 98, + 409, + 145, + 437 + ], + "blocks": [ + { + "bbox": [ + 98, + 409, + 145, + 437 + ], + "lines": [ + { + "bbox": [ + 98, + 409, + 145, + 437 + ], + "spans": [ + { + "bbox": [ + 98, + 409, + 145, + 437 + ], + "type": "image", + "image_path": "08432bf5060508bbda552ffa5bc21c7b36bd3179fbf9481d5c08d7fe12a1e89d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 441, + 259, + 460 + ], + "lines": [ + { + "bbox": [ + 91, + 441, + 259, + 460 + ], + "spans": [ + { + "bbox": [ + 91, + 441, + 259, + 460 + ], + "type": "text", + "content": "(b) Influence of image spatial frequency on intensity of stripe illusions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 154, + 409, + 200, + 437 + ], + "blocks": [ + { + "bbox": [ + 154, + 409, + 200, + 437 + ], + "lines": [ + { + "bbox": [ + 154, + 409, + 200, + 437 + ], + "spans": [ + { + "bbox": [ + 154, + 409, + 200, + 437 + ], + "type": "image", + "image_path": "45078f9bf8f90dae33064fd5dcb8abbf06862f265d760cdf5504028573e30514.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 467, + 295, + 488 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 488 + ], + "type": "text", + "content": "Figure 7. Error rates of humans and LLaVA across different conditions in color illusions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 209, + 409, + 256, + 437 + ], + "blocks": [ + { + "bbox": [ + 209, + 409, + 256, + 437 + ], + "lines": [ + { + "bbox": [ + 209, + 409, + 256, + 437 + ], + "spans": [ + { + "bbox": [ + 209, + 409, + 256, + 437 + ], + "type": "image", + "image_path": "f9b587f6424873e68977342d702bc6f0b3a91229aa9c5b68bd0b4e8b6c9b3719.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 498, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 581 + ], + "type": "text", + "content": "by directly informing the model that the image contains an illusion without specifying its type; and providing few-shot examples, using no-illusion (NI) or human-like (HL) answers as 3-shot examples. Because we use the LLaVA-7B model here, which doesn't support multi-image input, we combine the few-shot examples and the query image into a single composite image before feeding it to the model [22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Figure 8 show that external prompts have minimal effect on filter illusions. But for contrast and stripe illusions, the effect is more pronounced, particularly when no-illusion answers are used as few-shot examples, increasing the proportion of no-illusion responses by approximately " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": ". However, using human-like responses as few-shot examples does not effectively enhance the model's reproduction of human illusion perception, the proportion of human-like responses even decreases slightly for contrast illusions. Additionally, both direct emphasis and CoT prompts partially increase the proportion of no-illusion responses, al" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 175, + 533, + 286 + ], + "blocks": [ + { + "bbox": [ + 334, + 175, + 533, + 286 + ], + "lines": [ + { + "bbox": [ + 334, + 175, + 533, + 286 + ], + "spans": [ + { + "bbox": [ + 334, + 175, + 533, + 286 + ], + "type": "image", + "image_path": "ec8c8951fe5d5517b09e1692cce15ed61c2ce4eb56123f0cb784cff0920dbac5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 293, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 338 + ], + "type": "text", + "content": "Figure 8. Effect of different prompts and fine-tuning on VLM response preferences across illusion Types. \"NI\" denotes datasets using no-illusion examples, while \"HL\" represents datasets with human-like examples." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 348, + 534, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 534, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 534, + 360 + ], + "type": "text", + "content": "though the effect of direct emphasis is relatively weak." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "type": "text", + "content": "Fine-Tuning. First, we fine-tune two separate LLaVA-7B models on illusion images, with one model trained with responses according to pixel values (NI), and the other trained on human-like responses (HL). The bottom two rows of Figure 8 show a significant effect on the test set, particularly when fine-tuning on no-illusion responses. For both contrast and stripe illusions, fine-tuning on no-illusion responses increases the proportion of no-illusion answers to over " + }, + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "type": "text", + "content": ". In contrast, fine-tuning on human-like responses has a weaker effect, suggesting that even with human-like responses, the model struggles to fully grasp human perception in the context of illusion images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 506, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 626 + ], + "type": "text", + "content": "Next, we experiment with a mixed training approach that combines both types of responses. We train on question-answer pairs for both types of responses, where questions asking \"Based on pixel values\" are paired with pixel-value responses, those asking \"Based on human perception\" are paired with a human-like response. Table 1 shows the influence of this mixed training on the test set, compared to the base model. These results confirm that training model can effectively distinguish between pixel-based and human perception cues, yielding responses aligned with user's query." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 639, + 503, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 503, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 503, + 651 + ], + "type": "text", + "content": "5. Why Color Illusions Affect VLMs" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 658, + 529, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 529, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 529, + 671 + ], + "type": "text", + "content": "5.1. Perceptual Biases from the Visual System" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "To investigate whether VLMs' perceptual biases originate from their visual components, we test a range of purely visual models, including ResNet [18], ViT [10], VGG [35]," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7811" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 170, + 146 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 146 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 146 + ], + "type": "image", + "image_path": "06e5b4a86c0a5c9843ac7089a1fb783e7a949c705d731fa75ad0aa03f0b8fed5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 148, + 172, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 148, + 172, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 172, + 177 + ], + "type": "text", + "content": "(a) Performance of purely vision models on non-illusion and illusion images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 74, + 293, + 146 + ], + "blocks": [ + { + "bbox": [ + 181, + 74, + 293, + 146 + ], + "lines": [ + { + "bbox": [ + 181, + 74, + 293, + 146 + ], + "spans": [ + { + "bbox": [ + 181, + 74, + 293, + 146 + ], + "type": "image", + "image_path": "2f4753d4e6f317f040583acd4ada89f5305dfa03bbd7bd8a8bd76a11068c7579.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 148, + 295, + 177 + ], + "lines": [ + { + "bbox": [ + 179, + 148, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 179, + 148, + 295, + 177 + ], + "type": "text", + "content": "(b) Accuracy of ViT on non-illusion and illusion images over each training epochs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 482 + ], + "type": "text", + "content": "MobileNet [19] and Large Vision Model (LVM) [3]. We evaluate these models on both simple, rectangle-based contrast illusion images and generated realistic illusion images. For simple illusion images, we generate backgrounds with varying brightness on both sides and place a rectangle on each side. The vision model is trained as a three-way classifier on 6,000 non-illusion images to predict which rectangle appears darker, or if their colors are identical, stopping when the loss shows no significant decrease. It is then tested on 1,000 non-illusion and 1,000 illusion images. The results are shown in Figure 9a. For realistic illusion images, we train the model on the training set of our generated dataset and evaluate it on the test set, shown in the appendix. While overall accuracy (the proportion of classifications aligning with actual pixel values) on realistic illusion images drops compared to simple rectangle-based images, in both cases, accuracy on illusion images remains significantly lower than on non-illusion images. We also observe that during training, accuracy on non-illusion images quickly improves, while accuracy on illusion images fluctuates significantly. We use ViT as an example in Figure 9b. These results suggest that even purely vision-based models can be influenced by color illusions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 269, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 269, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 269, + 506 + ], + "type": "text", + "content": "5.2. Perceptual Biases from Prior Knowledge" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": "Influence of Language. We test the multilingual vision-language model PALO [32] on contrast illusions, posing questions in English, Chinese, and Russian to analyze error rates on images containing target regions across several colors. We hypothesize that colors with more frequent textual modifiers (e.g., dark blue, sky blue) lead to responses aligned with pixel values rather than human perception. To validate this, we first analyze color descriptor diversity in the training data. For example, in English, " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": " of red instances include modifiers like dark red or bright red, while " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "49.3\\%" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": " use red alone. Figure 10 shows the correlation between color descriptor diversity and pixel-value accuracy for English, Chinese, and Russian. We find that richer descriptions improve model sensitivity to that color, reducing susceptibility to color illusions in that color range." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 80, + 184, + 270, + 195 + ], + "lines": [ + { + "bbox": [ + 80, + 184, + 270, + 195 + ], + "spans": [ + { + "bbox": [ + 80, + 184, + 270, + 195 + ], + "type": "text", + "content": "Figure 9. Perceptual biases in purely vision models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 72, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 133 + ], + "type": "image", + "image_path": "8b7269aa4b36779b053b034e63ea25ae8a9124a0ba10e2a1700acc077a3257a0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 143, + 555, + 176 + ], + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 176 + ], + "type": "text", + "content": "Figure 10. Comparison of color term accuracy and frequency across languages for the vision-language model. Each shape and color represents a specific color term (e.g., blue, red, gray)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 184, + 555, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 292 + ], + "type": "text", + "content": "Influence of Commonsense Knowledge. Humans are often deceived by visual illusions due to our prior knowledge. For instance, research suggests that in filter illusions, humans perceive colors that aren't actually present due to a phenomenon called chromatic adaptation [27]. This allows us to use surrounding context to infer the presence of a filter, mentally \"removing\" it to perceive the original color. One example is that we always recognize a piece of white paper as white, regardless of the lighting conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": "To examine whether commonsense knowledge similarly influences VLMs' susceptibility to illusions, we categorized filter illusion images into three groups: single objects, multiple-object compositions, and complex real-world scenes, then analyzed the deception rates for both humans and VLMs across these groups. On non-illusion images, VLMs perform well across all three categories, suggesting that scene complexity does not affect their basic color recognition ability. However, for illusion images, our results in Figure 11 show that both humans and VLMs are less likely to be deceived by single-object images, while complex scenes lead to the highest deception rates, underscoring the role of prior knowledge in visual illusions." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 337, + 458, + 531, + 531 + ], + "blocks": [ + { + "bbox": [ + 337, + 458, + 531, + 531 + ], + "lines": [ + { + "bbox": [ + 337, + 458, + 531, + 531 + ], + "spans": [ + { + "bbox": [ + 337, + 458, + 531, + 531 + ], + "type": "image", + "image_path": "e3251dc85eda73f9872f1c406931bba288ce32be8ba757a6cc8b12546a6151c9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 540, + 556, + 562 + ], + "lines": [ + { + "bbox": [ + 313, + 540, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 556, + 562 + ], + "type": "text", + "content": "Figure 11. Deception rates of VLMs and human participants across different types of filter illusions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Our experiments show that current VLMs, influenced by their visual systems and prior knowledge, can exhibit visual illusions similar to humans. This raises a key question: as models become more human-like, might they also inherit human perceptual biases? In fields like biomedicine [8, 21], such biases could result in flawed visual judgments, while in human-centric tasks like image generation, understanding human perception can benefit. Therefore, VLM behavior should be tailored to application needs, with careful consideration of human-like perceptual biases." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "text", + "content": "1In Russian, blue has 100% descriptor diversity due to distinct terms for light blue (goluboj) and deep blue (sinij)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7812" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "text", + "content": "[1] Adam Alter. Drunk tank pink: And other unexpected forces that shape how we think, feel, and behave. 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 168 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "type": "text", + "content": "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. CVPR, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "text", + "content": "[4] Claus-Christian Carbon. Understanding human perception by human-made illusions. Frontiers in human neuroscience, 2014. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "type": "text", + "content": "[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: Endowing vision-language models with spatial reasoning capabilities. CVPR, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "type": "text", + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. CVPR, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 294, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 400 + ], + "type": "text", + "content": "[7] Fan L Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C Aoki, Jin Hirano, and Yukiyasu Kamitani. Reconstructing visual illusory experiences from human brain activity. Science Advances, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "text", + "content": "[8] Hejie Cui, Lingjun Mao, Xin Liang, Jieyu Zhang, Hui Ren, Quanzheng Li, Xiang Li, and Carl Yang. Biomedical visual instruction tuning with clinician preference alignment. NeurIPS, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 448, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 294, + 501 + ], + "type": "text", + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 294, + 567 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 294, + 612 + ], + "type": "text", + "content": "[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. BLINK: Multimodal large language models can see but not perceive. ECCV, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "text", + "content": "[12] Maurizio Gentilucci, Sergio Chieffi, Elena Daprati, M Cristina Saetti, and Ivan Toni. Visual illusion and action. Neuropsychologia, 1996. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[13] Ernst Hans Gombrich. Art and illusion: A study in the psychology of pictorial representation-millennium edition. 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. CVPR, 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmio, and Jesus Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[16] Richard Langton Gregory. Perceptual illusions and brain models. Proceedings of the Royal Society of London. Series B. Biological Sciences, 1968. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 173 + ], + "type": "text", + "content": "[17] Richard L Gregory. Visual illusions. Scientific American, 1968. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 206 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[19] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "text", + "content": "[20] Yan Ke, Xiaou Tang, and Feng Jing. The design of high-level features for photo quality assessment. CVPR, 2006. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "type": "text", + "content": "[21] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. NeurIPS, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeuRIPS, 2023. 1, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 366, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 397 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. CVPR, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "text", + "content": "[24] Sichao Liu, Jianjing Zhang, Robert X. Gao, Xi Vincent Wang, and Lihui Wang. Vision-language model-driven scene understanding and robotic object manipulation. IEEE CASE, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "text", + "content": "[25] Margaret S Livingstone. Art, illusion and the visual system. Scientific American, 1988. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "text", + "content": "[26] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-IO 2: Scaling autoregressive multimodal models with vision, language, audio, and action. CVPR, 2023. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "type": "text", + "content": "[27] D. L. MacAdam. Chromatic adaptation*. J. Opt. Soc. Am., 1956. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "text", + "content": "[28] O Maddaluno, A Facchin, D Zavagno, N Bolognini, R Daini, et al. Top-down and bottom-up neuromodulation over two different visual illusions. Atti del 40th European Conference on Visual Perception, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 591, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 622 + ], + "type": "text", + "content": "[29] Glenn E Meyer and Susan Petry. Top-down and bottom-up: The illusory contour as a microcosm of issues in perception. 1987. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[30] William L Phillips. Cross-cultural differences in visual perception of color, illusions, depth, and pictures. *Cross-Cultural Psychology: Contemporary Themes and Perspectives*, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Jiang Qiu, Hong Li, Qinglin Zhang, Qiang Liu, and Fenghua Zhang. The muller-lyer illusion seen by the brain: An event-related brain potentials study. Biological psychology, 2008. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7813" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 643 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[32] Hanoona Rasheed, Muhammad Maaz, Abdelrahman Shaker, Salman Khan, Hisham Cholakal, Rao M. Anwer, Tim Baldwin, Michael Felsberg, and Fahad S. Khan. Palo: A large multilingual multimodal language model. WACV, 2025. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[33] Alireza Saeedi, Kun Wang, Ghazaleh Nikpourian, Andreas Bartels, Nikos K Logothetis, Nelson K Totah, and Masataka Watanabe. Brightness illusions drive a neuronal response in the primary visual cortex under top-down modulation. Nature Communications, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[34] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2015. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "text", + "content": "[36] Zineng Tang, Lingjun Mao, and Alane Suhr. Grounding language in multi-perspective referential communication. EMNLP, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "text", + "content": "[37] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? exploring the visual shortcomings of multimodal llms. CVPR, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[38] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. PMLR, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "text", + "content": "[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogVLM: Visual expert for pretrained language models. NeuRIPS, 2024. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 486 + ], + "type": "text", + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. NeuRIPS, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 520 + ], + "type": "text", + "content": "[41] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. arXiv preprint arXiv:2304.00685, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 522, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 553 + ], + "type": "text", + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. CVPR, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "text", + "content": "[43] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding Visual Illusions in Language: Do vision-language models perceive illusions like humans? EMNLP, 2023. 1, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "text", + "content": "[44] Zeliang Zhang, Xin Liang, Mingqian Feng, Susan Liang, and Chenliang Xu. Will the inclusion of generated data amplify bias across generations in future image classification models? arXiv preprint arXiv:2410.10160, 2024.1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7814" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_content_list.json b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4cb0f983ebe0abb84917e994d6984cc56d718de6 --- /dev/null +++ b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_content_list.json @@ -0,0 +1,1653 @@ +[ + { + "type": "text", + "text": "Evaluating Vision-Language Models as Evaluators in Path Planning", + "text_level": 1, + "bbox": [ + 156, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohamed Aghzal1, Xiang Yue2, Erion Plaku3*, Ziyu Yao1 \n1George Mason University, 2Carnegie Mellon University, 3National Science Foundation {maghzal, ziyuyao}@gmu.edu, xyue2@andrew.cmu.edu, eplaku@nsf.gov", + "bbox": [ + 151, + 179, + 843, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their promise to perform complex reasoning, large language models (LLMs) have been shown to have limited effectiveness in end-to-end planning. This has inspired an intriguing question: if these models cannot plan well, can they still contribute to the planning framework as a helpful plan evaluator? In this work, we generalize this question to consider LLMs augmented with visual understanding, i.e., Vision-Language Models (VLMs). We introduce PATHEVAL, a novel benchmark evaluating VLMs as plan evaluators in complex path-planning scenarios. Succeeding in the benchmark requires a VLM to be able to abstract traits of optimal paths from the scenario description, demonstrate precise low-level perception on each path, and integrate this information to decide the better path. Our analysis of state-of-the-art VLMs reveals that these models face significant challenges on the benchmark. We observe that the VLMs can precisely abstract given scenarios to identify the desired traits and exhibit mixed performance in integrating the provided information. Yet, their vision component presents a critical bottleneck, with models struggling to perceive low-level details about a path. Our experimental results show that this issue cannot be trivially addressed via end-to-end fine-tuning; rather, task-specific discriminative adaptation of these vision encoders is needed for these VLMs to become effective path evaluators.[12]", + "bbox": [ + 89, + 301, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 710, + 220, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The impressive capabilities of Large Language Models (LLMs) [6, 8, 71] and Vision-Language Models (VLMs) [82] have led to an increasing interest in applying them to automated motion planning and navigation tasks [50, 51]. However, the inherent limitations of these models in long-", + "bbox": [ + 89, + 737, + 483, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "horizon planning have rendered them ineffective as end-to-end motion planners [2, 3, 11, 75]. This has made researchers wonder: if these models cannot be good motion planners themselves, can they still support a motion planning framework? Intuitively, using these models still holds the promise of significantly enhancing the motion planning framework, as they have learned extensive factual and commonsense knowledge that could benefit planning during their pre-training. As a result, there has been an emerging paradigm exploring how these models can be leveraged in combination with traditional methods [27].", + "bbox": [ + 511, + 268, + 906, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One particularly interesting approach within this line of work involves using these models as plan evaluators. Motivated by the intuition that \"evaluation is easier than generation\" [31], several efforts have explored leveraging these models as \"critics\" to assess the quality of generated plans [5, 21, 73, 83]. However, most of these efforts have focused on scenarios that require only limited, high-level visual perception, without necessitating fine-grained or precise perceptual abilities. On the other hand, while there have been prior works similarly investigating VLMs' low-level perception [26, 64], studies specifically about the use of these models in planning remain limited. Tasks such as motion planning often require fine-grained visual understanding in highly specific contexts, while also drawing on broad commonsense knowledge acquired during pre-training. Hence, there is a pressing need to investigate the potential of VLMs to understand both low-level visual details and leverage these visual signals for high-level reasoning.", + "bbox": [ + 511, + 441, + 908, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we explore whether we can utilize VLMs as evaluators in highly intricate continuous path planning problems. We introduce PATHEVAL (Figure 1), a controllable benchmark designed around path planning in complex environments under diverse commonsense decision-making scenarios (e.g., A firefighting robot has to cover as much of the area as possible to extinguish fires and scout for survivors, thus needs to prioritize paths with more coverage and higher clearance). Traditionally, encoding constraints in a planning algorithm requires significant human effort in hand-crafting scenario-specific criteria for path evaluation, such that the constraints can be injected into the planner", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The work by E. Plaku is supported by (while serving at) the National Science Foundation. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. \n1Source code: https://github.com/MohamedAghzal/PathEval \n2Dataset: https://huggingface.co/datasets/maghzal/PathEval", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6886", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Example of an instance of PATHEVAL", + "text_level": 1, + "bbox": [ + 116, + 92, + 372, + 107 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0cd1b2523106083092d842b736f974c7c1231d5c5a89f94d796fae60e61758db.jpg", + "image_caption": [ + "2D" + ], + "image_footnote": [], + "bbox": [ + 116, + 119, + 390, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0838274835244e4bdc4acc68be840e06f4ef52695e7f8d98d7b664cd06a6bd3e.jpg", + "image_caption": [ + "3D" + ], + "image_footnote": [], + "bbox": [ + 428, + 133, + 645, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9583cea04a42a866a2085509284da1af944969998b3d5988bc342156a8d6e63.jpg", + "image_caption": [ + "Figure 1. Example of PATHEVAL. The benchmark consists of 14,550 instances of path pairs rendered in both 2D and 3D and mapped to 15 decision-making scenarios. Success on this task is tied to three distinct levels: 1) Attribute abstraction: recognizing what aspects make a path ideal, 2) Low-level perception: extracting the required attributes for each path from the images, and 2) Information integration: synthesizing the collected information to make a decision. We test a set of VLMs on the task and find that they struggle particularly with low-level perception. Incorrect answers by different models are shown (explanations indicating misperception are highlighted)." + ], + "image_footnote": [], + "bbox": [ + 656, + 133, + 872, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Path Descriptors:", + "text_level": 1, + "bbox": [ + 116, + 256, + 218, + 268 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Path 1: Min. clearance: 1.07, Max. clearance: 8.91, Avg. clearance: 2.62, Path length: 243.46, Smoothness: 843.07, Number of sharp turns: 4, Maximum angle: 124.35.", + "Path 2: Min. clearance: 1.57, Max. clearance: 8.91, Avg. clearance: 3.18, Path length: 243.75, Smoothness: 809.65, Number of sharp turns: 1, Maximum angle: 93.12." + ], + "bbox": [ + 116, + 268, + 879, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scenario #1: The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic.", + "bbox": [ + 116, + 327, + 879, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Required Descriptors: Smoothness $(\\downarrow)$ , Number of Sharp turns $(\\downarrow)$ , Maximum angle $(\\downarrow)$", + "bbox": [ + 116, + 351, + 591, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ground Truth: Path 2", + "bbox": [ + 116, + 364, + 246, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scenario #2: An autonomous firefighting robot is designed to navigate and operate within burning buildings to extinguish fires and rescue trapped individuals. It should explore as much of the area as possible, while maintaining a safe distance from the walls to avoid damage.", + "bbox": [ + 116, + 375, + 879, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Required Descriptors: Avg. Clearance $(\\uparrow)$ , Min. Clearance $(\\uparrow)$ , Path length $(\\uparrow)$", + "bbox": [ + 116, + 398, + 544, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ground Truth: Path 1", + "bbox": [ + 116, + 412, + 241, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scenario #3: As the vehicle is traversing a warzone, it must stay concealed from enemy operatives, making use of covers like walls and avoiding open spaces as much as possible. It should also reach its target as quickly as possible.", + "bbox": [ + 116, + 422, + 879, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Required Descriptors: Avg. Clearance (↓), Max. Clearance (↓), Path length (↓)", + "bbox": [ + 116, + 446, + 545, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ground Truth: Path 2", + "bbox": [ + 116, + 459, + 246, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt", + "text_level": 1, + "bbox": [ + 116, + 483, + 163, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic. Which path better achieves the task? Path 1 is on the left side and Path 2 is on the right side. Your answer should follow the format below:", + "bbox": [ + 116, + 494, + 879, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer: Path 1 or Path 2.", + "bbox": [ + 116, + 531, + 256, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explanation: Why you chose the path (1 or 2).", + "bbox": [ + 116, + 542, + 364, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GPT-40", + "text_level": 1, + "bbox": [ + 116, + 566, + 165, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer: Path 1.", + "bbox": [ + 116, + 579, + 209, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explanation: Path 1 has fewer sharp turns and maintains longer straight segments compared to Path 2. This is more suitable for a large truck navigating through heavy traffic, as it minimizes the need for difficult maneuvers and allows for smoother driving. Path 2, on the other hand, has more frequent and sharper turns, which would be more challenging for a large truck to navigate.", + "bbox": [ + 114, + 590, + 879, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLaVA-NeXT-7b", + "text_level": 1, + "bbox": [ + 116, + 638, + 217, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer: Path 1.", + "bbox": [ + 116, + 650, + 210, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explanation: Path 1 is better suited for a large truck because it has fewer sharp turns (90 degrees or larger) compared to Path 2. The truck would have to make only one 90-degree turn, which is easier to execute than the two 90-degree turns in Path 2. Additionally, Path 1 has more straight sections, which would be more efficient for a large truck to navigate through heavy traffic. The fewer turns and the greater number of straight sections make Path 1 the more suitable route for a large truck.", + "bbox": [ + 114, + 662, + 879, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "during the search process [36, 52, 54]. However, this not only is a tedious and costly process but also cannot scale up to handle the countless intricate scenarios in meeting real-life planning needs. An effective VLM as a \"reward model\"", + "bbox": [ + 89, + 825, + 485, + 886 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relieving humans from customizing evaluation criteria for specific scenarios could enable scalable, general-purpose planners that can adapt to diverse commonsense scenarios described in natural language.", + "bbox": [ + 511, + 825, + 906, + 886 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6887", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In PATHEVAL, a VLM is tasked with comparing two paths within a given decision-making context and selecting the one that better satisfies the constraints outlined by the scenario. Success on this task requires effective performance across three distinct levels: 1) attribute abstraction: recognizing the attributes that define a favorable path in a particular decision-making context; 2) low-level perception: demonstrating precise low-level perception to determine which path performs better based on the given criteria; and 3) information integration: integrating and synthesizing the perceived information to produce an answer.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using this benchmark, we analyze the performance of 9 state-of-the-art (SOTA) VLMs, including both the closed-source GPT-4o and GPT-4o-mini [47] and 7 different open-source VLMs of different sizes (i.e., LLaVA-NeXT 7b and 13b [39], Qwen2-VL-7b [69], LLaVA-OneVision-7b [32], LLaMA-3.2-11b [17], and Intern-VL2 8b and 40b [12]). We find that these models struggle with the path evaluation task (e.g., Qwen2-VL-7b achieving only $50.2\\%$ accuracy). However, when providing these VLMs with verbalized path specifications, their performance significantly improves (e.g., $74.2\\%$ accuracy for Qwen2-VL-7b), which reveals a potential vision bottleneck of these VLMs. Our further analysis confirms these models' weakness in low-level perception, especially when they are tasked to perceive the clearance of a path with respect to surrounding obstacles, and this weakness could be more prominent when the environment and the path representation become more complex. We discover the source of this weakness from the vision encoders used by these VLMs, yet simply fine-tuning the VLMs end-to-end with the vision encoders does not address the issue. Rather, our experiments suggest performing task-specific discriminative adaptation of these vision encoders.", + "bbox": [ + 91, + 258, + 486, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 608, + 232, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Vision-Language Models", + "text_level": 1, + "bbox": [ + 89, + 633, + 316, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The outstanding success of decoder-only LLMs [6, 48, 65] has driven the development of Vision-Language Models (VLMs), which extend LLMs with a vision component in an attempt to generalize their performance into a multi-modal setting [12, 17, 32, 37, 38, 62, 69]. VLMs are designed for tasks that require unifying visual and linguistic representations (e.g. visual question answering [4]). Typically, this is achieved by connecting a vision encoder to a language decoder and projecting the representations into a shared space, allowing visual elements and linguistic components to be linked in a semantically meaningful manner. One of the earliest successful models at achieving this was the CLIP encoder [55], which was trained using contrastive learning to learn representations that map images and their corresponding textual descriptions. Several varieties of CLIP have then been introduced [25, 56, 76, 80]. These mod", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "els, while showing tremendous promises have shown several limitations when it comes to visual tasks [57]; as such several works [64, 72] have sought ways to improve such representations by combining them with vision foundation models such as DINO [8, 49]. In this work, we contribute to the research of VLMs with a new benchmark, namely PATHEVAL, focusing on evaluating VLMs as evaluators for path planning. This benchmark tests VLMs seamlessly on their commonsense understanding (i.e., being able to abstract critical concepts expressed in the described planning scenarios), low-level perception (i.e., precisely perceiving details about paths in complex environments), and the ability to reason about the collected textual and visual information for decision-making.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Automated Planning with VLMs", + "text_level": 1, + "bbox": [ + 511, + 313, + 799, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language planning promises more flexible planning frameworks and enhanced human-AI interaction. Therefore, designing systems that can effectively understand natural language instruction and leverage perceptual inputs to conduct planning tasks has been a topic of interest in recent years [20, 61, 70]. The rise of VLMs has led many to investigate the use of these models as vision-language planning agents [15, 16, 50, 51]. However, existing literature highlights the limitations of LLMs in spatial reasoning [2, 26, 74] as well as long-horizon planning [3, 67, 68]. This pushed researchers to explore alternative ways to incorporate VLMs and LLMs into planning frameworks more reliably [27, 58]. One potential direction is their use as plan evaluators, either through the generation of reward functions [23, 33, 73], or by using them directly as off-the-shelf critics [5, 21, 83]. The success of such frameworks assumes perfect perception and that the models can accurately perceive visual information and reason about it in order to produce an answer; nevertheless, it has been shown that the representations used by these models fail in highly intricate visual settings [64]. Several works have explored the use of VLMs as well as LLMs for path and motion planning [2, 3, 7, 9, 11, 14, 44, 75], however, to the knowledge of the authors there is no work that explores the use of VLMs as path critics in this context. Accordingly, we aim to evaluate the ability of VLMs to serve as evaluators in cases requiring navigation in complex environments while adhering to decision-making constraints specified in natural language.", + "bbox": [ + 511, + 337, + 906, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Vision-Language Model Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 771, + 834, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The introduction of multimodal models has prompted the development of several benchmarks that are capable of assessing the performance of these models on visual reasoning tasks such as visual question-answering datasets [4, 18, 19, 24, 42, 43, 84]. However, the rise of foundation models has produced the need for a more holistic evaluation of the perceptual and reasoning capabilities of large", + "bbox": [ + 511, + 795, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6888", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLMs, leading to benchmarks such as MM-Vet [77], MM-Bench [40], MMMU [45, 78, 79, 81] and OmniBench [35]. Several benchmarks have also specifically been designed to assess the perception capabilities of these models and explore the limitations associated with visual hallucinations and optical illusions [10, 22, 34, 60, 64]. Our proposed benchmark provides a flexible yet challenging framework for interleavedly assessing the low-level perception and reasoning capabilities of VLMs.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. The PATHEVAL Benchmark", + "text_level": 1, + "bbox": [ + 89, + 241, + 349, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the need to evaluate VLMs as path evaluators in real-world planning scenarios, we introduce PATHEVAL, a controllable and extensible benchmark focused on path planning [28-30, 41] in complex environments under a diverse set of decision-making constraints. We list the decision-making scenarios as well as the descriptors they attempt to optimize in Appendix A. In total, PATHEVAL includes 14,550 tasks over more than 1,150 distinct environments and 15 distinct scenarios. Below, we introduce a formal description of the task and the dataset construction.", + "bbox": [ + 89, + 267, + 482, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Task Formulation", + "text_level": 1, + "bbox": [ + 89, + 428, + 264, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given two paths, $P_{1}$ and $P_{2}$ , and a scenario $S$ , the objective is to determine which path better satisfies the scenario's optimization criteria. Each scenario $S$ is a high-level description that aims to optimize over a set of path descriptors (or metrics) $\\mathcal{M} = \\{m_1, m_2, \\dots, m_k\\}$ , where each descriptor $m_j : P \\to \\mathbb{R}$ evaluates a specific property of a path (e.g., length, smoothness, or proximity to obstacles). A VLM $\\mathcal{V}$ is presented with two images presenting $P_{1}$ and $P_{2}$ in the same environment, respectively. The model must then decide which path better satisfies the scenario's criteria. To explore the sensitivity of VLMs to the way how a path is presented, PATHEVAL includes both the 2D and 3D images of the path illustration. The model is also prompted to generate an explanation to justify its choice.", + "bbox": [ + 89, + 452, + 482, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Environment and Path Generation", + "text_level": 1, + "bbox": [ + 89, + 674, + 393, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Environment Generation: An environment, as shown in Figure 1, is defined by a set of walls $\\mathcal{O} = \\{O_1,O_2,\\dots ,O_n\\}$ , where each wall $O_{i}$ represents an obstacle in the 2D space. Each wall is a closed geometric shape described by its vertices, and the set $\\mathcal{O}$ forms the obstacles that the path must avoid. In this work, we leverage the environments of Plaku et. al. [53], which consists of four types of obstacle arrangements: (1) rings, where the environments are structured as mazes with circular walls, (2) waves, which consist of wavy horizontal obstacles, (3) mazes, which consist of both vertical and horizontal walls forming a complex maze structure, as well as (4) random, which consist of randomly placed obstacles.", + "bbox": [ + 89, + 698, + 482, + 893 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Path Synthesis via the Randomly-exploring Rapid Tree (RRT) algorithm: To generate path candidates in PATHVAL, we leverage the RRT path planning algorithm [30]. Starting from the initial location in the environment, the algorithm works by building a tree which expands itself by randomly selecting the next location in the environment while avoiding obstacles, until it reaches the goal. In this work, we use the Open Motion Planning Library (OMPL) [59] and implement the RRT-Connect algorithm [29]. We note that while we use RRT in the current benchmark, our codebase is adaptable and can incorporate most path-planning algorithms provided by OMPL. We encourage future research building on our benchmark to experiment with other algorithms as well.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Path Descriptors", + "text_level": 1, + "bbox": [ + 511, + 319, + 679, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect the following descriptors $\\mathcal{M}$ for each of a generated path: Minimum Clearance measures the smallest distance between any point on the path and the nearest obstacle; Maximum Clearance measures the largest distance between any point on the path and the nearest obstacle; Average Clearance computes the average distance between all points on the path and the nearest obstacle; Path Length is calculated by summing up the Euclidean distances between consecutive points on the path; Smoothness is defined as the sum of the angles between consecutive segments of the path, measuring how smoothly the path changes direction; Number of Sharp Turns counts the number of turns in the path where the angle between consecutive segments exceeds 90 degrees; and Maximum Angle denotes the largest angle between any two consecutive segments of the path. The three Clearance metrics and Path Length share the same measuring unit, i.e., one grid size; Smoothness and Maximum Angle are measured by degree; and the Number of Sharp Turns is an integer count. We include the formula of each descriptor in Appendix B.", + "bbox": [ + 511, + 344, + 906, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Natural Language Descriptions of Scenarios", + "text_level": 1, + "bbox": [ + 511, + 662, + 885, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To create a sufficiently challenging path-planning evaluation benchmark, we design a total of 15 decision-making scenarios that aim to optimize different combinations of the path descriptors. For instance, Scenario #2 (as shown in Figure 1) requires searching through an area affected with fire in search for survivors, and the agent thus must cover as much ground as possible. In contrast, Scenario #3 indicates that the path is to be executed within a warzone; as a result, the vehicle has to remain hidden and take the shortest route. As such, given the same set of paths, the one minimizing the path length is favored by Scenario #3 while Scenario #2 needs to maximize this value. A complete summary of the 15 scenarios, along with the descriptors each scenario aims to optimize, is presented in Tables 6-7 in Appendix A.", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6889", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Task Pairs Selection", + "text_level": 1, + "bbox": [ + 91, + 90, + 281, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each environment we synthesized, we ran the RRT planner 30 times to generate different paths. Upon eliminating paths that did not reach the goal, we selected path pairs that exhibited the greatest dissimilarity in terms of path descriptors. Specifically, we first represented each path using a (7-dimension) vector of its path descriptor values. Given that each path descriptor ranged in a dramatically different scale, we normalized the vector by performing the min-max scaling, i.e., scaling each value $x$ in the vector to $x' = \\frac{x - x_{\\min}}{x_{\\max} - x_{\\min}}$ , where $x$ represents each value in the vector, and $x_{\\min}$ and $x_{\\max}$ are the minimum and maximum values of each descriptor across paths sampled from the same environment, respectively. We then measured the distance between two paths by calculating the Euclidean distance between their normalized path descriptor vectors, and selected 5 path pairs with the largest distances. Each path is included in only one pair to avoid redundancy. We repeated the same selection procedure for a total of 1,150 environments.", + "bbox": [ + 89, + 114, + 483, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upon performing this generation, we matched each pair with all fifteen scenarios; however, we only kept pairs where there was a significant difference in at least one of the descriptors required for the particular scenario. In other words, we ensure that the absolute difference is greater than a predefined threshold (0.8 for clearance descriptors, 50 for path length, 90 for smoothness, 1 for the number of sharp turns, and 30 for maximum angle) for at least one of the required descriptors. This makes it more likely that the difference is more noticeable to the naked eye, and thus the paths can be compared by visual inspection.", + "bbox": [ + 89, + 387, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We constructed our final evaluation benchmark by randomly selecting 70 task pairs from each scenario, resulting in 1,050 pairs in total. The remaining task pairs (a total of 13,500) are used as the training set to facilitate the fine-tuning experiments in Section 5.", + "bbox": [ + 89, + 555, + 483, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Can SOTA VLMs Evaluate Planned Paths?", + "text_level": 1, + "bbox": [ + 89, + 647, + 478, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order for VLMs to perform successfully in our benchmark, they need to perform successfully at three different levels, i.e., recognizing the critical descriptors required by each scenario (Attribute Abstraction), exhibiting sharp low-level perception to precisely assess each path's properties (Low-level Perception), and integrating the prior information to make a rational decision on the better path (Information Integration). Among them, the first two levels reflect parallel properties that serve as a foundation for the third level. In this section, we evaluated a set of 9 VLMs on PATHEVAL in a zero-shot manner and analyzed their capabilities at these three levels. These VLMs include (1) two closed-source VLMs, i.e., GPT-4o and GPT-4o-mini [47], and (2) seven open-source VLMs with various sizes, including LLaVA-NeXT-7b and LLaVA-NeXT-13b [38, 39],", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaVA-OneVision-7b [32], Qwen2-VL-7b [69], LLaMA3.2-11b [17], and Intern-VL2-8b and Intern-VL2-40b [12]. We include all prompt scripts used in this Section in Appendix C.1.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Overview of VLMs on PATHEVAL", + "text_level": 1, + "bbox": [ + 511, + 161, + 802, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The performance of the 9 VLMs on PATHEVAL is showcased in Table 1. We notice that all of the models, except GPT-4o, fail to perform significantly better than a simple random baseline, indicating significant limitations. For GPT-4o, we also notice a $4\\%$ higher accuracy on 2D images compared to prompting with 3D inputs. This observation indicates that the model is prone to visual illusions introduced by the 3D images, when it has to rely on solely the image for decision-making (although we observe an opposite effect of 2D vs. 3D when verbalized descriptor values are provided). In the remaining section, we will further break down these models' capabilities to gain a deeper understanding of their failure on this task.", + "bbox": [ + 511, + 184, + 906, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Providing verbalized path information yields better task accuracy, implying reasonable VLM performance in attribute abstraction. In Table 1, we further show the performance of each model when we explicitly list the value for each descriptor as part of the language prompt (i.e., \"PATHEVAL w Desc.). We notice a $11.1\\% -27.1\\%$ improvement across most models, indicating that when given low-level details, the models can better filter out the information and make better comparisons. This points out that the bottleneck for these VLMs's better performance lies in their inability to accurately perceive low-level information about the paths (we discuss this in more detail in Section 4.2), whereas these models generally have a reasonable capability in abstracting the critical attributes for decision making in various scenarios. In Appendix D.1, we include an experiment where we explicitly query each VLM to identify the critical path metrics for each scenario; the result corroborates our hypothesis. In particular, we find that for most models, the success rate for identifying a required descriptor is over $92\\%$ . Finally, a surprising observation happens to LLaVa-NeXT-7b. We notice that this model suffers particularly severely from hallucination; even when the textual descriptor values are provided and when the model can correctly identify critical path metrics based on our analysis in Appendix D.1, it fails to pick the better paths. We show an example of its explanation when textual descriptors are provided in Figure 5 of Appendix E.", + "bbox": [ + 511, + 383, + 908, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VLMs exhibit mixed performance in integrating visual and textual path information. We take a further look into the model performance when both the image and the textual descriptor values are provided, and contrast it with their performance when only the textual path descriptions are provided (i.e., \"Desc Only\"). Interest-", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6890", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/984d8d013dbf83d3fdae80fbd9b0f329d1f5b0c131822a5cbd10647454017ec0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model2D3DDesc Only
PATHEVALPATHEVAL w Desc.PATHEVALPATHEVAL w Desc.
GPT-4o-mini0.5200.7500.5080.7450.680
GPT-4o0.6650.8600.6240.8950.894
LLaVa-NeXT-7b0.5010.5240.4990.5170.514
Qwen2-VL-7b0.5020.7310.5110.7420.737
LLaVA-OneVision-7b0.5050.7180.5090.7390.721
Intern-VL2-8b0.4890.6540.5050.6910.648
LLaMa-3.2-11b0.4800.6950.4600.6800.686
LLaVa-NeXT-13b0.5090.6200.4940.6010.630
Intern-VL2-40b0.5060.6880.4960.7170.679
Random Baseline0.5000.5000.5000.5000.500
", + "bbox": [ + 176, + 88, + 823, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ingly, we observe mixed information integration performance from these models. For GPT-4o, Qwen2-VL-7b, LLaVA-OneVision-7b, LLaMa-3.2-11b, and LLaVa-NeXT-13b, their performance based on only descriptor values has no obvious difference to their best performance when (2D or 3D) images are also provided. This observation implies that these models do not benefit from the additional image information when the textual path descriptions are provided. Instead, sometimes the images (e.g., 2D images for GPT-4o and 3D images for LLaVa-NeXT-13b) may confuse their understanding of the textual descriptors, resulting in a worse accuracy compared to Desc Only. For GPT-4o-mini, InternVL2-8b, and Intern-VL2-40b, however, providing both the visual and textual path information offers $\\sim 4\\% - 7\\%$ performance gain over Desc Only, indicating better information integration from these VLMs.", + "bbox": [ + 88, + 337, + 480, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Appendix D.2, we present a breakdown of GPT-4o's performance by scenarios. We show that performance varies greatly from one scenario to the other. Interestingly, we observe that GPT-4o may overuse their commonsense knowledge. For instance, in the case of Scenario #2 shown in Figure 1, where the agent is required to maximize the path length for better coverage, GPT-4o still favors the shorter path. This scenario represents a counterfactual situation as models are often trained or instructed to seek the shortest paths. Evaluating VLMs in such counterfactual scenarios allows us to effectively probe their task understanding and reasoning, which we consider to be an important direction for future research.", + "bbox": [ + 88, + 579, + 482, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The reasoning of VLMs can be unreliable. Limited by the vision bottleneck, we noticed these VLMs fabricating seemingly plausible explanations to justify their path evaluation, despite the fact that they could not actually perceive the necessary path details to perform the task. This fabrication echos findings from other recent work, where LLMs were shown to produce reasoning traces that do not", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3bd26b1765a79830567d5a329b19bc75c7ffb4fb22cef58b3cbc097505e5337.jpg", + "table_caption": [ + "Table 1. Accuracy of VLMs on PATHEVAL based on 2D and 3D environment images. To investigate the potential vision bottleneck of VLMs, we additionally present each model's accuracy when explicit descriptor values are provided in the language prompts (\"PATHEVAL w Desc\"). The last column (\"Desc Only\") shows the model performance when only the textual descriptor values are provided." + ], + "table_footnote": [], + "table_body": "
ModelDefaultFlippedRandom IDs
GPT-4o-mini34/101622/102694/956
GPT-4o278/772258/792291/759
LLaVa-NeXT-7b1028/221041/9580/470
Qwen2-VL-7b21/1029130/920127/923
LLaVA-OneVision-7b438/612440/610262/788
Intern-VL2-8b433/610430/620418/632
LLaMA-3.2-11b*601/251620/264731/319
LLaVa-NeXT-13b795/255939/111305/745
Intern-VL2-40b394/656410/639510/540
Ground truth labels530/520
", + "bbox": [ + 529, + 334, + 890, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance on the 2D case (#of times first path is chosen / #of times second path is chosen) when we flip the path order or replace their default names with random IDs. *There are several cases where LLaMA-3.2-11b does not follow the required format and/or does not give an answer, we omit those cases from this table.", + "bbox": [ + 511, + 503, + 906, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "accurately reflect the process of how the models reach an answer [1, 66]. To gain further insights, we performed an analysis comparing model performance on PATHEVAL with cases that consist of: 1) flipping the order of paths in the pair, and 2) assigning random IDs to the paths (e.g. instead of referring to them as \"Path 1\" and \"Path 2\", we use a random sequence such as \"Path Xu2q\" and \"Path fP48\"). The results presented in Table 2 showcase that VLMs demonstrate bias for a particular label, when they actually do not have the capability to resolve the task. For example, when no matter the default or the flipping labels are used, LLaVA-NeXT-7b consistently selects Path $198\\%$ of the time and fabricates incorrect observations of the two paths in its explanations (Figure 1) to support this choice. As we discussed earlier, LLaVA-NeXT-7b is particularly prone to hallucination in explanations, leading to its random-guess performance with or without textual descriptors. Introducing random IDs as path names mitigates this bias for LLaVa-NeXT-7b (although the model still obtains a close-to-random accuracy on PATHEVAL) but does not seem to help other models dramatically.", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6891", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a41bd53061424851923f31e618bcfd17edd1e2fb0a9db002a2d5c8a9fa76cbb7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DescriptorTest Setε1ε2ε3
Min. Clearance0.46/0.460.50/0.460.74/0.700.86/0.74
Max. Clearance0.44/0.460.41/0.490.46/0.550.50/0.60
Avg. Clearance0.53/0.550.50/0.520.70/0.570.73/0.60
Path Length0.58/0.700.86/0.910.92/0.860.94/0.94
Smoothness0.74/0.720.86/0.820.90/0.900.90/0.89
# of Sharp Turns0.76/0.730.80/0.840.77/0.700.60/0.55
Max. Angle0.71/0.700.82/0.840.86/0.880.94/0.96
", + "bbox": [ + 106, + 88, + 470, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. GPT-4o's fine-grained perception accuracy (2D/3D) on the test set of PATHEVAL and three additionally synthesized datasets with increasing metric differences.", + "bbox": [ + 89, + 205, + 482, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. What Can Vision-Language Models See?", + "text_level": 1, + "bbox": [ + 89, + 256, + 441, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The previous subsection clearly highlights the vision component as the bottleneck for path evaluation on PATHEVAL. In this subsection, we conduct an analysis of the visual perception capabilities of VLMs. We focus our analysis on GPT-4o since it is the only model that performs substantially better than the random baseline in the case where no descriptors are required.", + "bbox": [ + 89, + 281, + 483, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT-4o can perceive paths that are significantly different. In order to get a better understanding of the model's perception capability, we break down its performance in terms of perception by individual metric. Specifically, we prompt GPT-4o to select which path in the pair provides a lower value on each individual metric and report its accuracy. We perform this analysis on both the task pairs in the test set of PATHEVAL and an additionally synthesized dataset consisting of task pairs with three levels (denoted as $\\epsilon_{1}$ , $\\epsilon_{2}$ , and $\\epsilon_{3}$ ) of increasing differences in their descriptor values. We describe the data generation process for this dataset in more detail in Appendix C.2.", + "bbox": [ + 89, + 387, + 483, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results for both experiments are shown in Table 3. Upon evaluating the performance on the test set of PATHEVAL by individual metrics, we notice that GPT-4o particularly struggles with Clearance metrics. These metrics typically require a lower level of perception and is naturally more challenging to discriminate than other metrics. On the other hand, Smoothness, Number of Sharpest Turns, and Max Angle appear to be easier for the model to capture. We also note that GPT-4o perceives the path length much more easily in a 3D environment presentation. Furthermore, GPT-4o's performance increases as we increase the descriptor difference between paths. This, however, is not the case for the Number of Sharp Turns. We conjecture that when we increase the number of sharp turns, we also enforce the challenge of \"counting\" the number of satisfying turns, which VLMs have been shown with limitations [46].", + "bbox": [ + 89, + 570, + 483, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does segment complexity affect performance? We look into whether the complexity of the path segment is the key reason for GPT-4o's limited perception of clearance. To this end, we test GPT-4o on segments of varying complexity (i.e., points, straight lines, and curves), in a simplified", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64370fd06c0e1b6c0ffc05b4ad3b3dd2d3fac2b321af24f07d50de566c670467.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 87, + 616, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/797dcb26cc65e4acc5a8548ba8507480512eeccb1d2e575eb31105b9c6389929.jpg", + "image_caption": [ + "(a) Point" + ], + "image_footnote": [], + "bbox": [ + 517, + 183, + 612, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/223e1262a5584fe660daa68ba08978ce7562dd5eeade35746fed784978de8428.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 88, + 759, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6b5e322480982896d78b549393ea442747ad58630c462ebe76d636c4ff6422c.jpg", + "image_caption": [ + "(b) Straight line" + ], + "image_footnote": [], + "bbox": [ + 663, + 184, + 756, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3293b2cd816cf29ad2a3b7463952ed1bc871b3cc637a029db4988eafd99a361.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 88, + 903, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf60e29ab12647331ed9e4a6f5a0d1019199b600f144e02495db1d762791a9cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 184, + 901, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bbab9b6506d01862c50749d8e46a9e7912d85bfc0a817618f028adc644cdf826.jpg", + "image_caption": [ + "(c) Curve", + "(d) GPT-40 performance on the distance to obstacle under different segment complexities.", + "Figure 2. Example segment complexity test cases in simplified environments and performance across the various settings." + ], + "image_footnote": [], + "bbox": [ + 540, + 243, + 879, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "environment with only one rectangular obstacle at the center (Figure 2), and evaluate its accuracy in identifying segments that are closer to the obstacle. For individual points, the clearance is defined by the perpendicular distance from the point to the obstacle; for straight lines and curves, we consider a path closer if one of its endpoints is closer to the obstacle. For each segment type, we synthesize 100 pairs by first randomly generating 200 segments and then pairing each with the segment with the greatest distance difference from the obstacle relative to it (i.e., maximizing the absolute difference between the distances of the two segments from the obstacle). This increases the likelihood that the distance is significant enough to be perceivable. The average difference in the clearances of the pairs of segments are 14.76, 14.42, and 14.28 for points, lines, and curves respectively.", + "bbox": [ + 511, + 378, + 905, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Figure 2d show that GPT-4o can perform better in very easy scenarios; however, it struggles more as the segment complexity increases. For instance, the model was able to identify the closer points in $77\\%$ of the cases; however, when considering straight and curved lines, its performance drops to $64\\%$ and $59\\%$ , respectively. A surprising observation is that, in the case of curved lines, GPT-4o's performance is dramatically better in 3D images. As shown in Appendix C.2, the average Clearance difference of path pairs in PATHEVAL is merely 0.12 - 1.31. As the paths in PATHEVAL are much more complicated than the curves in this experiment, it is expected that GPT-4o exhibits difficulty in judging paths' clearances. The complexity of the environments (compared to a single square obstacle) could add challenges.", + "bbox": [ + 511, + 604, + 905, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Fine-tuning a Path Evaluator", + "text_level": 1, + "bbox": [ + 511, + 844, + 782, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One intuitive question is whether simply fine-tuning the VLMs can relieve their vision bottleneck. To answer this", + "bbox": [ + 511, + 869, + 905, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6892", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4be98c2f0b85047d8c927d2f81bdb50e3c3e89c606694d9a908cca4e5d351a26.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFrozenFine-tuned
AccuracyAvg. Cosine SimilarityAccuracyAvg. Cosine Similarity
clip-vit-base-patch320.5100.9140.7830.514
clip-vit-large-patch14-3360.4980.9070.7490.548
OpenCLIP-vit-B-320.5400.8830.7430.475
siglip-base-patch16-2240.5290.8950.7310.612
dino-vit-160.4950.9110.7630.754
dino-v2-base0.5100.7610.7210.681
", + "bbox": [ + 207, + 88, + 790, + 203 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/faec76c6d170612426fffd7fc2d278b8264e9dc1c7e7b3ec652ad756a743685d.jpg", + "table_caption": [ + "Table 4. Probing accuracy and average cosine similarity between distinct path images (2D) when the vision encoder is frozen or fine-tuned." + ], + "table_footnote": [], + "table_body": "
SettingFrozenTunable
Image Only0.520.51
Image w Descriptors0.96-
Image Only (Random IDs)0.480.52
", + "bbox": [ + 143, + 234, + 431, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Fine-tuned LLaVA-NeXT-7b performance (2D) when we keep the vision encoder frozen or tunable.", + "bbox": [ + 89, + 309, + 482, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "question, we experiment with LLaVA-NeXT-7b and fine-tune it on the training set (13,500 pairs) of PATHEVAL. We focus on the 2D case for the set of experiments described in this section. We consider three separate settings for training: (1) training with only images as input, (2) training with images and textual descriptor values as input, and (3) the same setting as (1) with using random IDs as target labels. Details in experimental setup are included in Appendix C.3.", + "bbox": [ + 89, + 347, + 482, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Overall Performance", + "text_level": 1, + "bbox": [ + 89, + 479, + 290, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning does not help with vision-language mapping. Table 5 shows that the model fails to learn meaningful patterns in the data, even after training 50 epochs. However, when the textual descriptors are provided as input, the model can easily learn the function achieving $96\\%$ accuracy (a $45\\%$ improvement from the zero-shot setting). This shows that the model is unable to extract the same descriptor values from the image input. Unfreezing the encoder for fine-tuning also does not provide any significant improvement. We include a further discussion in Appendix D.3. The results point to a limitation in the vision model's ability to encode the images, which we will investigate next.", + "bbox": [ + 89, + 503, + 482, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Understanding the Visual Representations", + "text_level": 1, + "bbox": [ + 89, + 695, + 450, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to further understand the limitations of the vision component, we conduct an analysis to better understand how well different vision encoder models can differentiate between different paths in the dataset. To this end, we first apply a linear probe to see how easily distinguishable different images are. Specifically, given a pair of images, we first use the vision encoder to extract high-dimensional feature representations for both images. These features are then concatenated and passed through a simple binary classification layer (i.e., the probe). The probe is trained to predict a label of 1 if the images are the same and 0 otherwise. We experiment with various SOTA vision encoders, namely CLIP", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[55] base and large varieties, LAION-OpenCLIP [25, 56], SigLip [80], DINO [8], and DINO-v2 [49] and analyze how well their learned representations capture visual similarities and differences. We use a set of 1,000 randomly sampled path pairs with a balanced label distribution to train the probe, and look at whether the model can learn to distinguish between these paths. We also perform an experiment where we fine-tune the vision encoder along with the probe. In this setting, our goal is to gauge if carefully fine-tuning a vision encoder can potentially improve the model's performance in low-level perception. Finally, in both settings, we present the average cosine similarity between distinct paths.", + "bbox": [ + 511, + 237, + 903, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vision encoders cannot distinguish between paths. From Table 4, it can be seen that vision encoder models are unable to provide representations that are significantly different for the probe to tell if they are the same. This is further supported by the high values for the average cosine similarity across all models.", + "bbox": [ + 511, + 421, + 903, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning the encoders on a discrimination task can help disentangle the visual representations. By unfreezing the encoder weights and training them to identify whether two paths are identical, we enhance their adaptability to the task. The results in Table 4 demonstrate that this approach effectively disentangles the learned representations, resulting in significantly improved performance and increased separability, as evidenced by the notably lower cosine similarity between non-identical paths. The results thus imply the need for carefully fine-tuning task-specific vision encoders for path evaluation on PATHEVAL.", + "bbox": [ + 511, + 515, + 906, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 705, + 633, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work explored the use of VLMs as evaluators in path-planning scenarios. We evaluated a number of VLMs on our proposed benchmark, PATHEVAL, and found that these models struggle with low-level perception. Specifically, we find that visual encoders used by SOTA models are unable to discern the differences between different paths in intricate scenarios. We hope that PATHEVAL will inspire researchers to further explore ways to improve the visual capabilities of VLMs and contribute to finding better ways to incorporate foundation models for developing more flexible, robust, and scalable planning paradigms.", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6893", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project was supported by resources provided by the Office of Research Computing at George Mason University (URL: https://orc.gmu.edu) and funded in part by grants from the National Science Foundation (Award Number 2018631).", + "bbox": [ + 89, + 116, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 208, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chirag Agarwal, Sree Harsha Tanneru, and Himabindu Lakkaraju. Faithfulness vs. Plausibility: On the (Un)Reliability of Explanations from Large Language Models, 2024. 6", + "[2] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can Large Language Models be Good Path Planners? A Benchmark and Investigation on Spatial-temporal Reasoning. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1, 3", + "[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Look Further Ahead: Testing the Limits of GPT-4 in Path Planning. IEEE CASE, 2024. 1, 3", + "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In International Conference on Computer Vision (ICCV), 2015. 3", + "[5] Kate Baumli, Satinder Baveja, Feryal Behbahani, Harris Chan, Gheorghe Comanici, Sebastian Flennerhag, Maxime Gazeau, Kristian Holsheimer, Dan Horgan, Michael Laskin, Clare Lyle, Hussain Masoom, Kay McKinney, Volodymyr Mnih, Alexander Neitz, Dmitry Nikulin, Fabio Pardo, Jack Parker-Holder, John Quan, Tim Roktaschel, Himanshu Sahni, Tom Schaul, Yannick Schroecker, Stephen Spencer, Richie Steigerwald, Luyu Wang, and Lei Zhang. Vision-Language Models as a Source of Rewards, 2024. 1, 3", + "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1, 3", + "[7] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21819–21830, 2024. 3", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 3, 8" + ], + "bbox": [ + 99, + 233, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee Wong. MapGPT: Map-guided prompting with adaptive path planning for vision-and-language navigation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9796-9810, Bangkok, Thailand, 2024. Association for Computational Linguistics. 3", + "[10] Xuweiyi Chen, Ziqiao Ma, Xuejun Zhang, Sihan Xu, Shengyi Qian, Jianing Yang, David Fouhey, and Joyce Chai. Multi-object hallucination in vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4", + "[11] Yongchao Chen, Jacob Arkin, Charles Dawson, Yang Zhang, Nicholas Roy, and Chuchu Fan. Autotamp: Autoregressive task and motion planning with llms as translators and checkers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6695-6702, 2024. 1, 3", + "[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. arXiv preprint arXiv:2312.14238, 2023. 3, 5", + "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 4", + "[14] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 3", + "[15] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 3", + "[16] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, brian richter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B. Tenenbaum, Leslie Pack Kaelbling, Andy Zeng, and Jonathan Tompson. Video language planning. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[17] Abhimanyu Dubey et. al. The Llama 3 Herd of Models, 2024. 3, 5", + "[18] Difei Gao, Ruiping Wang, Shiguang Shan, and Xilin Chen. Cric: A vqa dataset for compositional reasoning on vision and commonsense. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5561-5578, 2022. 3", + "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6894", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[20] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 2022. 3", + "[21] Lin Guan, Yifan Zhou, Denis Liu, Yantian Zha, Heni Ben Amor, and Subbarao Kambhampati. Task success is not enough: Investigating the use of video-language models as behavior critics for catching undesirable agent behaviors. In First Conference on Language Modeling, 2024. 1, 3", + "[22] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models, 2023. 4", + "[23] Xu Han, Qiannan Yang, Xianda Chen, Xiaowen Chu, and Meixin Zhu. Generating and evolving reward functions for highway driving with large language models, 2024. 3", + "[24] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3, 8", + "[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's \"up\" with vision-language models? Investigating their struggle with spatial reasoning. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 3", + "[27] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Audit Verma, Kaya Stechly, Siddhant Bhambri, Lucas Paul Saldyt, and Anil B Murthy. Position: LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks. In _Forty-first International Conference on Machine Learning_, 2024. 1, 3", + "[28] L.E. Kavraki, P. Svestka, J.-C. Latombe, and M.H. Overmars. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and Automation, 12(4):566-580, 1996. 4", + "[29] J.J. Kuffner and S.M. LaValle. RRT-connect: An efficient approach to single-query path planning. In Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065), pages 995-1001 vol.2, 2000. 4", + "[30] Steven M. LaValle. Rapidly-exploring random trees: a new tool for path planning. The annual research report, 1998. 4", + "[31] Jan Leike. Why i'm optimistic about our alignment approach: Evaluation is easier than generation." + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "https://aligned.substack.com/i/88447351/evaluation-is-easier-than-generation,2022.Accessed:2024-11.1", + "[32] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5", + "[33] Hao Li, Xue Yang, Zhaokai Wang, Xizhou Zhu, Jie Zhou, Yu Qiao, Xiaogang Wang, Hongsheng Li, Lewei Lu, and Jifeng Dai. Auto mc-reward: Automated dense reward design with large language models for minecraft. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 4", + "[35] Yizhi Li, Ge Zhang, Yinghao Ma, Ruibin Yuan, Kang Zhu, Hangyu Guo, Yiming Liang, Jiaheng Liu, Zekun Wang, Jian Yang, Siwei Wu, Xingwei Qu, Jinjie Shi, Xinyue Zhang, Zhenzhu Yang, Xiangzhou Wang, Zhaoxiang Zhang, Zachary Liu, Emmanouil Benetos, Wenhao Huang, and Chenghua Lin. OmniBench: Towards The Future of Universal Omni-Language Models, 2024. 4", + "[36] Hui Sheng Lim, Shuangshuang Fan, Christopher K.H. Chin, Shuhong Chai, Neil Bose, and Eonjoo Kim. Constrained path planning of autonomous underwater vehicle using selectively-hybridized particle swarm optimization algorithms. IFAC-PapersOnLine, 52(21):315-322, 2019. 12th IFAC Conference on Control Applications in Marine Systems, Robotics, and Vehicles CAMS 2019. 2", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3", + "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2024. 3, 5", + "[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 3, 5", + "[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 4", + "[41] Tomás Lozano-Pérez and Michael A. Wesley. An algorithm for planning collision-free paths among polyhedral obstacles. Communications of the ACM, 22(10):560-570, 1979. 4", + "[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning Representations, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6895", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[44] Silin Meng, Yiwei Wang, Cheng-Fu Yang, Nanyun Peng, and Kai-Wei Chang. LLM-a*: Large language model enhanced incremental heuristic search on path planning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1087-1102, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3", + "[45] Shota Onohara, Atsuyuki Miyai, Yuki Imajuku, Kazuki Egashira, Jeonghun Baek, Xiang Yue, Graham Neubig, and Kiyoharu Aizawa. Jmmm: A japanese massive multidiscipline multimodal understanding benchmark for culture-aware evaluation, 2024. 4", + "[46] OpenAI. GPT-4V(ision) System Card. 2023. 7", + "[47] OpenAI. GPT-4o System Card, 2024. 3, 5", + "[48] OpenAI. GPT-4 Technical Report, 2024. 3", + "[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 8", + "[50] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 950-974, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 3", + "[51] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14760-14769, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 1, 3", + "[52] Clment Petres, Yan Pailhas, Pedro Patron, Yvan Petillot, Jonathan Evans, and David Lane. Path planning for autonomous underwater vehicles. IEEE Transactions on Robotics, 23(2):331-341, 2007. 2", + "[53] Evis Plaku, Erion Plaku, and Patricio Simari. Clearance-driven motion planning for mobile robots with differential constraints. Robotica, 36(7):971-993, 2018. 4", + "[54] Patrick A. Plonski, Pratap Tokekar, and Volkan Isler. Energy-Efficient Path Planning for Solar-Powered Mobile Robots, pages 717-731. Springer International Publishing, Heidelberg, 2013. 2", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pam Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Proceedings of" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the 38th International Conference on Machine Learning, 2021. 3, 8", + "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3, 8", + "[57] Jie-Jing Shao, Jiang-Xin Shi, Xiao-Wen Yang, Lan-Zhe Guo, and Yu-Feng Li. Investigating the Limitation of CLIP Models: The Worst-Performing Categories, 2023. 3", + "[58] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B. Tenenbaum, Leslie Pack Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models, 2023. 3", + "[59] Loan A. Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics & Automation Magazine, 19(4):72-82, 2012. 4", + "[60] Yinan Sun, Zicheng Zhang, Haoning Wu, Xiaohong Liu, Weisi Lin, Guangtao Zhai, and Xiongkuo Min. Explore the hallucination on low-level perception for mllms, 2024. 4", + "[61] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3", + "[62] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 3", + "[63] Shengbang Tong, Erik Jones, and Jacob Steinhardt. Mass-producing failures of multimodal systems with language models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 4", + "[64] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs, 2024. 1, 3, 4", + "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6896", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 3", + "[66] Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language models don't always say what they think: unfaithful explanations in chain-of-thought prompting. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 6", + "[67] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large Language Models Still Can't Plan (A Benchmark for LLMs on Planning and Reasoning about Change). In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 3", + "[68] Karthik Valmeekam, Kaya Stechly, and Subbarao Kambhampati. LLMs Still Can't Plan; Can LRMs? A Preliminary Evaluation of OpenAI's o1 on PlanBench, 2024. 3", + "[69] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5", + "[70] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6622-6631, 2019. 3", + "[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, 2022. 1", + "[72] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. ECCV, 2024. 3", + "[73] Tianbao Xie, Siheng Zhao, Chen Henry Wu, Yitao Liu, Qian Luo, Victor Zhong, Yanchao Yang, and Tao Yu. Text2Reward: Reward Shaping with Language Models for Reinforcement Learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 3", + "[74] Yutaro Yamada, Yihan Bao, Andrew Kyle Lampinen, Jungo Kasai, and Ilker Yildirim. Evaluating spatial understanding of large language models. Transactions on Machine Learning Research, 2024. 3", + "[75] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1, 3", + "[76] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024. 4", + "[78] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI. In Proceedings of CVPR, 2024. 4", + "[79] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark. arXiv preprint arXiv:2409.02813, 2024. 4", + "[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 8", + "[81] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, and Jie Fu. CMMU: A Chinese Massive Multi-discipline Multimodal Understanding Benchmark, 2024. 4", + "[82] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models, 2024. 1", + "[83] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy Improvement using Language Feedback Models, 2024. 1, 3", + "[84] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7W: Grounded Question Answering in Images. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3" + ], + "bbox": [ + 516, + 90, + 903, + 652 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_model.json b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..da2fe617ad689ce85c1aa27a5c20012694840473 --- /dev/null +++ b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_model.json @@ -0,0 +1,2589 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.842, + 0.154 + ], + "angle": 0, + "content": "Evaluating Vision-Language Models as Evaluators in Path Planning" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.18, + 0.844, + 0.234 + ], + "angle": 0, + "content": "Mohamed Aghzal1, Xiang Yue2, Erion Plaku3*, Ziyu Yao1 \n1George Mason University, 2Carnegie Mellon University, 3National Science Foundation {maghzal, ziyuyao}@gmu.edu, xyue2@andrew.cmu.edu, eplaku@nsf.gov" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.486, + 0.681 + ], + "angle": 0, + "content": "Despite their promise to perform complex reasoning, large language models (LLMs) have been shown to have limited effectiveness in end-to-end planning. This has inspired an intriguing question: if these models cannot plan well, can they still contribute to the planning framework as a helpful plan evaluator? In this work, we generalize this question to consider LLMs augmented with visual understanding, i.e., Vision-Language Models (VLMs). We introduce PATHEVAL, a novel benchmark evaluating VLMs as plan evaluators in complex path-planning scenarios. Succeeding in the benchmark requires a VLM to be able to abstract traits of optimal paths from the scenario description, demonstrate precise low-level perception on each path, and integrate this information to decide the better path. Our analysis of state-of-the-art VLMs reveals that these models face significant challenges on the benchmark. We observe that the VLMs can precisely abstract given scenarios to identify the desired traits and exhibit mixed performance in integrating the provided information. Yet, their vision component presents a critical bottleneck, with models struggling to perceive low-level details about a path. Our experimental results show that this issue cannot be trivially addressed via end-to-end fine-tuning; rather, task-specific discriminative adaptation of these vision encoders is needed for these VLMs to become effective path evaluators.[12]" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.712, + 0.222, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.484, + 0.815 + ], + "angle": 0, + "content": "The impressive capabilities of Large Language Models (LLMs) [6, 8, 71] and Vision-Language Models (VLMs) [82] have led to an increasing interest in applying them to automated motion planning and navigation tasks [50, 51]. However, the inherent limitations of these models in long-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.438 + ], + "angle": 0, + "content": "horizon planning have rendered them ineffective as end-to-end motion planners [2, 3, 11, 75]. This has made researchers wonder: if these models cannot be good motion planners themselves, can they still support a motion planning framework? Intuitively, using these models still holds the promise of significantly enhancing the motion planning framework, as they have learned extensive factual and commonsense knowledge that could benefit planning during their pre-training. As a result, there has been an emerging paradigm exploring how these models can be leveraged in combination with traditional methods [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.715 + ], + "angle": 0, + "content": "One particularly interesting approach within this line of work involves using these models as plan evaluators. Motivated by the intuition that \"evaluation is easier than generation\" [31], several efforts have explored leveraging these models as \"critics\" to assess the quality of generated plans [5, 21, 73, 83]. However, most of these efforts have focused on scenarios that require only limited, high-level visual perception, without necessitating fine-grained or precise perceptual abilities. On the other hand, while there have been prior works similarly investigating VLMs' low-level perception [26, 64], studies specifically about the use of these models in planning remain limited. Tasks such as motion planning often require fine-grained visual understanding in highly specific contexts, while also drawing on broad commonsense knowledge acquired during pre-training. Hence, there is a pressing need to investigate the potential of VLMs to understand both low-level visual details and leverage these visual signals for high-level reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In this work, we explore whether we can utilize VLMs as evaluators in highly intricate continuous path planning problems. We introduce PATHEVAL (Figure 1), a controllable benchmark designed around path planning in complex environments under diverse commonsense decision-making scenarios (e.g., A firefighting robot has to cover as much of the area as possible to extinguish fires and scout for survivors, thus needs to prioritize paths with more coverage and higher clearance). Traditionally, encoding constraints in a planning algorithm requires significant human effort in hand-crafting scenario-specific criteria for path evaluation, such that the constraints can be injected into the planner" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.902 + ], + "angle": 0, + "content": "*The work by E. Plaku is supported by (while serving at) the National Science Foundation. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. \n1Source code: https://github.com/MohamedAghzal/PathEval \n2Dataset: https://huggingface.co/datasets/maghzal/PathEval" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6886" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.117, + 0.093, + 0.373, + 0.108 + ], + "angle": 0, + "content": "Example of an instance of PATHEVAL" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.121, + 0.391, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.238, + 0.264, + 0.25 + ], + "angle": 0, + "content": "2D" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.135, + 0.646, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.135, + 0.874, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.236, + 0.662, + 0.248 + ], + "angle": 0, + "content": "3D" + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.257, + 0.22, + 0.269 + ], + "angle": 0, + "content": "Path Descriptors:" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.269, + 0.88, + 0.292 + ], + "angle": 0, + "content": "Path 1: Min. clearance: 1.07, Max. clearance: 8.91, Avg. clearance: 2.62, Path length: 243.46, Smoothness: 843.07, Number of sharp turns: 4, Maximum angle: 124.35." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.292, + 0.88, + 0.317 + ], + "angle": 0, + "content": "Path 2: Min. clearance: 1.57, Max. clearance: 8.91, Avg. clearance: 3.18, Path length: 243.75, Smoothness: 809.65, Number of sharp turns: 1, Maximum angle: 93.12." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.269, + 0.88, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.328, + 0.88, + 0.352 + ], + "angle": 0, + "content": "Scenario #1: The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.352, + 0.592, + 0.365 + ], + "angle": 0, + "content": "Required Descriptors: Smoothness \\((\\downarrow)\\), Number of Sharp turns \\((\\downarrow)\\), Maximum angle \\((\\downarrow)\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.365, + 0.247, + 0.375 + ], + "angle": 0, + "content": "Ground Truth: Path 2" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.376, + 0.88, + 0.4 + ], + "angle": 0, + "content": "Scenario #2: An autonomous firefighting robot is designed to navigate and operate within burning buildings to extinguish fires and rescue trapped individuals. It should explore as much of the area as possible, while maintaining a safe distance from the walls to avoid damage." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.4, + 0.545, + 0.413 + ], + "angle": 0, + "content": "Required Descriptors: Avg. Clearance \\((\\uparrow)\\), Min. Clearance \\((\\uparrow)\\), Path length \\((\\uparrow)\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.413, + 0.242, + 0.424 + ], + "angle": 0, + "content": "Ground Truth: Path 1" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.424, + 0.88, + 0.448 + ], + "angle": 0, + "content": "Scenario #3: As the vehicle is traversing a warzone, it must stay concealed from enemy operatives, making use of covers like walls and avoiding open spaces as much as possible. It should also reach its target as quickly as possible." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.448, + 0.546, + 0.46 + ], + "angle": 0, + "content": "Required Descriptors: Avg. Clearance (↓), Max. Clearance (↓), Path length (↓)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.46, + 0.247, + 0.471 + ], + "angle": 0, + "content": "Ground Truth: Path 2" + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.484, + 0.165, + 0.495 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.496, + 0.88, + 0.531 + ], + "angle": 0, + "content": "The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic. Which path better achieves the task? Path 1 is on the left side and Path 2 is on the right side. Your answer should follow the format below:" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.532, + 0.258, + 0.543 + ], + "angle": 0, + "content": "Answer: Path 1 or Path 2." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.544, + 0.365, + 0.556 + ], + "angle": 0, + "content": "Explanation: Why you chose the path (1 or 2)." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.567, + 0.166, + 0.578 + ], + "angle": 0, + "content": "GPT-40" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.58, + 0.21, + 0.59 + ], + "angle": 0, + "content": "Answer: Path 1." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.591, + 0.88, + 0.629 + ], + "angle": 0, + "content": "Explanation: Path 1 has fewer sharp turns and maintains longer straight segments compared to Path 2. This is more suitable for a large truck navigating through heavy traffic, as it minimizes the need for difficult maneuvers and allows for smoother driving. Path 2, on the other hand, has more frequent and sharper turns, which would be more challenging for a large truck to navigate." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.639, + 0.218, + 0.65 + ], + "angle": 0, + "content": "LLaVA-NeXT-7b" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.651, + 0.212, + 0.662 + ], + "angle": 0, + "content": "Answer: Path 1." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.663, + 0.88, + 0.711 + ], + "angle": 0, + "content": "Explanation: Path 1 is better suited for a large truck because it has fewer sharp turns (90 degrees or larger) compared to Path 2. The truck would have to make only one 90-degree turn, which is easier to execute than the two 90-degree turns in Path 2. Additionally, Path 1 has more straight sections, which would be more efficient for a large truck to navigate through heavy traffic. The fewer turns and the greater number of straight sections make Path 1 the more suitable route for a large truck." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.742, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Figure 1. Example of PATHEVAL. The benchmark consists of 14,550 instances of path pairs rendered in both 2D and 3D and mapped to 15 decision-making scenarios. Success on this task is tied to three distinct levels: 1) Attribute abstraction: recognizing what aspects make a path ideal, 2) Low-level perception: extracting the required attributes for each path from the images, and 2) Information integration: synthesizing the collected information to make a decision. We test a set of VLMs on the task and find that they struggle particularly with low-level perception. Incorrect answers by different models are shown (explanations indicating misperception are highlighted)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.486, + 0.887 + ], + "angle": 0, + "content": "during the search process [36, 52, 54]. However, this not only is a tedious and costly process but also cannot scale up to handle the countless intricate scenarios in meeting real-life planning needs. An effective VLM as a \"reward model\"" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.887 + ], + "angle": 0, + "content": "relieving humans from customizing evaluation criteria for specific scenarios could enable scalable, general-purpose planners that can adapt to diverse commonsense scenarios described in natural language." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6887" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "In PATHEVAL, a VLM is tasked with comparing two paths within a given decision-making context and selecting the one that better satisfies the constraints outlined by the scenario. Success on this task requires effective performance across three distinct levels: 1) attribute abstraction: recognizing the attributes that define a favorable path in a particular decision-making context; 2) low-level perception: demonstrating precise low-level perception to determine which path performs better based on the given criteria; and 3) information integration: integrating and synthesizing the perceived information to produce an answer." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.26, + 0.487, + 0.593 + ], + "angle": 0, + "content": "Using this benchmark, we analyze the performance of 9 state-of-the-art (SOTA) VLMs, including both the closed-source GPT-4o and GPT-4o-mini [47] and 7 different open-source VLMs of different sizes (i.e., LLaVA-NeXT 7b and 13b [39], Qwen2-VL-7b [69], LLaVA-OneVision-7b [32], LLaMA-3.2-11b [17], and Intern-VL2 8b and 40b [12]). We find that these models struggle with the path evaluation task (e.g., Qwen2-VL-7b achieving only \\(50.2\\%\\) accuracy). However, when providing these VLMs with verbalized path specifications, their performance significantly improves (e.g., \\(74.2\\%\\) accuracy for Qwen2-VL-7b), which reveals a potential vision bottleneck of these VLMs. Our further analysis confirms these models' weakness in low-level perception, especially when they are tasked to perceive the clearance of a path with respect to surrounding obstacles, and this weakness could be more prominent when the environment and the path representation become more complex. We discover the source of this weakness from the vision encoders used by these VLMs, yet simply fine-tuning the VLMs end-to-end with the vision encoders does not address the issue. Rather, our experiments suggest performing task-specific discriminative adaptation of these vision encoders." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.609, + 0.233, + 0.625 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.318, + 0.65 + ], + "angle": 0, + "content": "2.1. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The outstanding success of decoder-only LLMs [6, 48, 65] has driven the development of Vision-Language Models (VLMs), which extend LLMs with a vision component in an attempt to generalize their performance into a multi-modal setting [12, 17, 32, 37, 38, 62, 69]. VLMs are designed for tasks that require unifying visual and linguistic representations (e.g. visual question answering [4]). Typically, this is achieved by connecting a vision encoder to a language decoder and projecting the representations into a shared space, allowing visual elements and linguistic components to be linked in a semantically meaningful manner. One of the earliest successful models at achieving this was the CLIP encoder [55], which was trained using contrastive learning to learn representations that map images and their corresponding textual descriptions. Several varieties of CLIP have then been introduced [25, 56, 76, 80]. These mod" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "els, while showing tremendous promises have shown several limitations when it comes to visual tasks [57]; as such several works [64, 72] have sought ways to improve such representations by combining them with vision foundation models such as DINO [8, 49]. In this work, we contribute to the research of VLMs with a new benchmark, namely PATHEVAL, focusing on evaluating VLMs as evaluators for path planning. This benchmark tests VLMs seamlessly on their commonsense understanding (i.e., being able to abstract critical concepts expressed in the described planning scenarios), low-level perception (i.e., precisely perceiving details about paths in complex environments), and the ability to reason about the collected textual and visual information for decision-making." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.8, + 0.33 + ], + "angle": 0, + "content": "2.2. Automated Planning with VLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.908, + 0.761 + ], + "angle": 0, + "content": "Vision-language planning promises more flexible planning frameworks and enhanced human-AI interaction. Therefore, designing systems that can effectively understand natural language instruction and leverage perceptual inputs to conduct planning tasks has been a topic of interest in recent years [20, 61, 70]. The rise of VLMs has led many to investigate the use of these models as vision-language planning agents [15, 16, 50, 51]. However, existing literature highlights the limitations of LLMs in spatial reasoning [2, 26, 74] as well as long-horizon planning [3, 67, 68]. This pushed researchers to explore alternative ways to incorporate VLMs and LLMs into planning frameworks more reliably [27, 58]. One potential direction is their use as plan evaluators, either through the generation of reward functions [23, 33, 73], or by using them directly as off-the-shelf critics [5, 21, 83]. The success of such frameworks assumes perfect perception and that the models can accurately perceive visual information and reason about it in order to produce an answer; nevertheless, it has been shown that the representations used by these models fail in highly intricate visual settings [64]. Several works have explored the use of VLMs as well as LLMs for path and motion planning [2, 3, 7, 9, 11, 14, 44, 75], however, to the knowledge of the authors there is no work that explores the use of VLMs as path critics in this context. Accordingly, we aim to evaluate the ability of VLMs to serve as evaluators in cases requiring navigation in complex environments while adhering to decision-making constraints specified in natural language." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.835, + 0.787 + ], + "angle": 0, + "content": "2.3. Vision-Language Model Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.903 + ], + "angle": 0, + "content": "The introduction of multimodal models has prompted the development of several benchmarks that are capable of assessing the performance of these models on visual reasoning tasks such as visual question-answering datasets [4, 18, 19, 24, 42, 43, 84]. However, the rise of foundation models has produced the need for a more holistic evaluation of the perceptual and reasoning capabilities of large" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6888" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "VLMs, leading to benchmarks such as MM-Vet [77], MM-Bench [40], MMMU [45, 78, 79, 81] and OmniBench [35]. Several benchmarks have also specifically been designed to assess the perception capabilities of these models and explore the limitations associated with visual hallucinations and optical illusions [10, 22, 34, 60, 64]. Our proposed benchmark provides a flexible yet challenging framework for interleavedly assessing the low-level perception and reasoning capabilities of VLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.242, + 0.351, + 0.259 + ], + "angle": 0, + "content": "3. The PATHEVAL Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.419 + ], + "angle": 0, + "content": "Motivated by the need to evaluate VLMs as path evaluators in real-world planning scenarios, we introduce PATHEVAL, a controllable and extensible benchmark focused on path planning [28-30, 41] in complex environments under a diverse set of decision-making constraints. We list the decision-making scenarios as well as the descriptors they attempt to optimize in Appendix A. In total, PATHEVAL includes 14,550 tasks over more than 1,150 distinct environments and 15 distinct scenarios. Below, we introduce a formal description of the task and the dataset construction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.429, + 0.266, + 0.445 + ], + "angle": 0, + "content": "3.1. Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.664 + ], + "angle": 0, + "content": "Given two paths, \\( P_{1} \\) and \\( P_{2} \\), and a scenario \\( S \\), the objective is to determine which path better satisfies the scenario's optimization criteria. Each scenario \\( S \\) is a high-level description that aims to optimize over a set of path descriptors (or metrics) \\( \\mathcal{M} = \\{m_1, m_2, \\dots, m_k\\} \\), where each descriptor \\( m_j : P \\to \\mathbb{R} \\) evaluates a specific property of a path (e.g., length, smoothness, or proximity to obstacles). A VLM \\( \\mathcal{V} \\) is presented with two images presenting \\( P_{1} \\) and \\( P_{2} \\) in the same environment, respectively. The model must then decide which path better satisfies the scenario's criteria. To explore the sensitivity of VLMs to the way how a path is presented, PATHEVAL includes both the 2D and 3D images of the path illustration. The model is also prompted to generate an explanation to justify its choice." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.394, + 0.691 + ], + "angle": 0, + "content": "3.2. Environment and Path Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.895 + ], + "angle": 0, + "content": "Environment Generation: An environment, as shown in Figure 1, is defined by a set of walls \\(\\mathcal{O} = \\{O_1,O_2,\\dots ,O_n\\}\\), where each wall \\(O_{i}\\) represents an obstacle in the 2D space. Each wall is a closed geometric shape described by its vertices, and the set \\(\\mathcal{O}\\) forms the obstacles that the path must avoid. In this work, we leverage the environments of Plaku et. al. [53], which consists of four types of obstacle arrangements: (1) rings, where the environments are structured as mazes with circular walls, (2) waves, which consist of wavy horizontal obstacles, (3) mazes, which consist of both vertical and horizontal walls forming a complex maze structure, as well as (4) random, which consist of randomly placed obstacles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Path Synthesis via the Randomly-exploring Rapid Tree (RRT) algorithm: To generate path candidates in PATHVAL, we leverage the RRT path planning algorithm [30]. Starting from the initial location in the environment, the algorithm works by building a tree which expands itself by randomly selecting the next location in the environment while avoiding obstacles, until it reaches the goal. In this work, we use the Open Motion Planning Library (OMPL) [59] and implement the RRT-Connect algorithm [29]. We note that while we use RRT in the current benchmark, our codebase is adaptable and can incorporate most path-planning algorithms provided by OMPL. We encourage future research building on our benchmark to experiment with other algorithms as well." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.681, + 0.336 + ], + "angle": 0, + "content": "3.3. Path Descriptors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.648 + ], + "angle": 0, + "content": "We collect the following descriptors \\(\\mathcal{M}\\) for each of a generated path: Minimum Clearance measures the smallest distance between any point on the path and the nearest obstacle; Maximum Clearance measures the largest distance between any point on the path and the nearest obstacle; Average Clearance computes the average distance between all points on the path and the nearest obstacle; Path Length is calculated by summing up the Euclidean distances between consecutive points on the path; Smoothness is defined as the sum of the angles between consecutive segments of the path, measuring how smoothly the path changes direction; Number of Sharp Turns counts the number of turns in the path where the angle between consecutive segments exceeds 90 degrees; and Maximum Angle denotes the largest angle between any two consecutive segments of the path. The three Clearance metrics and Path Length share the same measuring unit, i.e., one grid size; Smoothness and Maximum Angle are measured by degree; and the Number of Sharp Turns is an integer count. We include the formula of each descriptor in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.887, + 0.681 + ], + "angle": 0, + "content": "3.4. Natural Language Descriptions of Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To create a sufficiently challenging path-planning evaluation benchmark, we design a total of 15 decision-making scenarios that aim to optimize different combinations of the path descriptors. For instance, Scenario #2 (as shown in Figure 1) requires searching through an area affected with fire in search for survivors, and the agent thus must cover as much ground as possible. In contrast, Scenario #3 indicates that the path is to be executed within a warzone; as a result, the vehicle has to remain hidden and take the shortest route. As such, given the same set of paths, the one minimizing the path length is favored by Scenario #3 while Scenario #2 needs to maximize this value. A complete summary of the 15 scenarios, along with the descriptors each scenario aims to optimize, is presented in Tables 6-7 in Appendix A." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6889" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.282, + 0.106 + ], + "angle": 0, + "content": "3.5. Task Pairs Selection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.386 + ], + "angle": 0, + "content": "For each environment we synthesized, we ran the RRT planner 30 times to generate different paths. Upon eliminating paths that did not reach the goal, we selected path pairs that exhibited the greatest dissimilarity in terms of path descriptors. Specifically, we first represented each path using a (7-dimension) vector of its path descriptor values. Given that each path descriptor ranged in a dramatically different scale, we normalized the vector by performing the min-max scaling, i.e., scaling each value \\( x \\) in the vector to \\( x' = \\frac{x - x_{\\min}}{x_{\\max} - x_{\\min}} \\), where \\( x \\) represents each value in the vector, and \\( x_{\\min} \\) and \\( x_{\\max} \\) are the minimum and maximum values of each descriptor across paths sampled from the same environment, respectively. We then measured the distance between two paths by calculating the Euclidean distance between their normalized path descriptor vectors, and selected 5 path pairs with the largest distances. Each path is included in only one pair to avoid redundancy. We repeated the same selection procedure for a total of 1,150 environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.554 + ], + "angle": 0, + "content": "Upon performing this generation, we matched each pair with all fifteen scenarios; however, we only kept pairs where there was a significant difference in at least one of the descriptors required for the particular scenario. In other words, we ensure that the absolute difference is greater than a predefined threshold (0.8 for clearance descriptors, 50 for path length, 90 for smoothness, 1 for the number of sharp turns, and 30 for maximum angle) for at least one of the required descriptors. This makes it more likely that the difference is more noticeable to the naked eye, and thus the paths can be compared by visual inspection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.632 + ], + "angle": 0, + "content": "We constructed our final evaluation benchmark by randomly selecting 70 task pairs from each scenario, resulting in 1,050 pairs in total. The remaining task pairs (a total of 13,500) are used as the training set to facilitate the fine-tuning experiments in Section 5." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.648, + 0.479, + 0.665 + ], + "angle": 0, + "content": "4. Can SOTA VLMs Evaluate Planned Paths?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In order for VLMs to perform successfully in our benchmark, they need to perform successfully at three different levels, i.e., recognizing the critical descriptors required by each scenario (Attribute Abstraction), exhibiting sharp low-level perception to precisely assess each path's properties (Low-level Perception), and integrating the prior information to make a rational decision on the better path (Information Integration). Among them, the first two levels reflect parallel properties that serve as a foundation for the third level. In this section, we evaluated a set of 9 VLMs on PATHEVAL in a zero-shot manner and analyzed their capabilities at these three levels. These VLMs include (1) two closed-source VLMs, i.e., GPT-4o and GPT-4o-mini [47], and (2) seven open-source VLMs with various sizes, including LLaVA-NeXT-7b and LLaVA-NeXT-13b [38, 39]," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "LLaVA-OneVision-7b [32], Qwen2-VL-7b [69], LLaMA3.2-11b [17], and Intern-VL2-8b and Intern-VL2-40b [12]. We include all prompt scripts used in this Section in Appendix C.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.162, + 0.803, + 0.177 + ], + "angle": 0, + "content": "4.1. Overview of VLMs on PATHEVAL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.38 + ], + "angle": 0, + "content": "The performance of the 9 VLMs on PATHEVAL is showcased in Table 1. We notice that all of the models, except GPT-4o, fail to perform significantly better than a simple random baseline, indicating significant limitations. For GPT-4o, we also notice a \\(4\\%\\) higher accuracy on 2D images compared to prompting with 3D inputs. This observation indicates that the model is prone to visual illusions introduced by the 3D images, when it has to rely on solely the image for decision-making (although we observe an opposite effect of 2D vs. 3D when verbalized descriptor values are provided). In the remaining section, we will further break down these models' capabilities to gain a deeper understanding of their failure on this task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Providing verbalized path information yields better task accuracy, implying reasonable VLM performance in attribute abstraction. In Table 1, we further show the performance of each model when we explicitly list the value for each descriptor as part of the language prompt (i.e., \"PATHEVAL w Desc.). We notice a \\(11.1\\% -27.1\\%\\) improvement across most models, indicating that when given low-level details, the models can better filter out the information and make better comparisons. This points out that the bottleneck for these VLMs's better performance lies in their inability to accurately perceive low-level information about the paths (we discuss this in more detail in Section 4.2), whereas these models generally have a reasonable capability in abstracting the critical attributes for decision making in various scenarios. In Appendix D.1, we include an experiment where we explicitly query each VLM to identify the critical path metrics for each scenario; the result corroborates our hypothesis. In particular, we find that for most models, the success rate for identifying a required descriptor is over \\(92\\%\\). Finally, a surprising observation happens to LLaVa-NeXT-7b. We notice that this model suffers particularly severely from hallucination; even when the textual descriptor values are provided and when the model can correctly identify critical path metrics based on our analysis in Appendix D.1, it fails to pick the better paths. We show an example of its explanation when textual descriptors are provided in Figure 5 of Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "VLMs exhibit mixed performance in integrating visual and textual path information. We take a further look into the model performance when both the image and the textual descriptor values are provided, and contrast it with their performance when only the textual path descriptions are provided (i.e., \"Desc Only\"). Interest-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6890" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.178, + 0.089, + 0.825, + 0.269 + ], + "angle": 0, + "content": "
Model2D3DDesc Only
PATHEVALPATHEVAL w Desc.PATHEVALPATHEVAL w Desc.
GPT-4o-mini0.5200.7500.5080.7450.680
GPT-4o0.6650.8600.6240.8950.894
LLaVa-NeXT-7b0.5010.5240.4990.5170.514
Qwen2-VL-7b0.5020.7310.5110.7420.737
LLaVA-OneVision-7b0.5050.7180.5090.7390.721
Intern-VL2-8b0.4890.6540.5050.6910.648
LLaMa-3.2-11b0.4800.6950.4600.6800.686
LLaVa-NeXT-13b0.5090.6200.4940.6010.630
Intern-VL2-40b0.5060.6880.4960.7170.679
Random Baseline0.5000.5000.5000.5000.500
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.278, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Table 1. Accuracy of VLMs on PATHEVAL based on 2D and 3D environment images. To investigate the potential vision bottleneck of VLMs, we additionally present each model's accuracy when explicit descriptor values are provided in the language prompts (\"PATHEVAL w Desc\"). The last column (\"Desc Only\") shows the model performance when only the textual descriptor values are provided." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.482, + 0.579 + ], + "angle": 0, + "content": "ingly, we observe mixed information integration performance from these models. For GPT-4o, Qwen2-VL-7b, LLaVA-OneVision-7b, LLaMa-3.2-11b, and LLaVa-NeXT-13b, their performance based on only descriptor values has no obvious difference to their best performance when (2D or 3D) images are also provided. This observation implies that these models do not benefit from the additional image information when the textual path descriptions are provided. Instead, sometimes the images (e.g., 2D images for GPT-4o and 3D images for LLaVa-NeXT-13b) may confuse their understanding of the textual descriptors, resulting in a worse accuracy compared to Desc Only. For GPT-4o-mini, InternVL2-8b, and Intern-VL2-40b, however, providing both the visual and textual path information offers \\(\\sim 4\\% - 7\\%\\) performance gain over Desc Only, indicating better information integration from these VLMs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.58, + 0.483, + 0.777 + ], + "angle": 0, + "content": "In Appendix D.2, we present a breakdown of GPT-4o's performance by scenarios. We show that performance varies greatly from one scenario to the other. Interestingly, we observe that GPT-4o may overuse their commonsense knowledge. For instance, in the case of Scenario #2 shown in Figure 1, where the agent is required to maximize the path length for better coverage, GPT-4o still favors the shorter path. This scenario represents a counterfactual situation as models are often trained or instructed to seek the shortest paths. Evaluating VLMs in such counterfactual scenarios allows us to effectively probe their task understanding and reasoning, which we consider to be an important direction for future research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The reasoning of VLMs can be unreliable. Limited by the vision bottleneck, we noticed these VLMs fabricating seemingly plausible explanations to justify their path evaluation, despite the fact that they could not actually perceive the necessary path details to perform the task. This fabrication echos findings from other recent work, where LLMs were shown to produce reasoning traces that do not" + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.335, + 0.891, + 0.495 + ], + "angle": 0, + "content": "
ModelDefaultFlippedRandom IDs
GPT-4o-mini34/101622/102694/956
GPT-4o278/772258/792291/759
LLaVa-NeXT-7b1028/221041/9580/470
Qwen2-VL-7b21/1029130/920127/923
LLaVA-OneVision-7b438/612440/610262/788
Intern-VL2-8b433/610430/620418/632
LLaMA-3.2-11b*601/251620/264731/319
LLaVa-NeXT-13b795/255939/111305/745
Intern-VL2-40b394/656410/639510/540
Ground truth labels530/520
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.505, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Table 2. Performance on the 2D case (#of times first path is chosen / #of times second path is chosen) when we flip the path order or replace their default names with random IDs. *There are several cases where LLaMA-3.2-11b does not follow the required format and/or does not give an answer, we omit those cases from this table." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "accurately reflect the process of how the models reach an answer [1, 66]. To gain further insights, we performed an analysis comparing model performance on PATHEVAL with cases that consist of: 1) flipping the order of paths in the pair, and 2) assigning random IDs to the paths (e.g. instead of referring to them as \"Path 1\" and \"Path 2\", we use a random sequence such as \"Path Xu2q\" and \"Path fP48\"). The results presented in Table 2 showcase that VLMs demonstrate bias for a particular label, when they actually do not have the capability to resolve the task. For example, when no matter the default or the flipping labels are used, LLaVA-NeXT-7b consistently selects Path \\(198\\%\\) of the time and fabricates incorrect observations of the two paths in its explanations (Figure 1) to support this choice. As we discussed earlier, LLaVA-NeXT-7b is particularly prone to hallucination in explanations, leading to its random-guess performance with or without textual descriptors. Introducing random IDs as path names mitigates this bias for LLaVa-NeXT-7b (although the model still obtains a close-to-random accuracy on PATHEVAL) but does not seem to help other models dramatically." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "6891" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.472, + 0.196 + ], + "angle": 0, + "content": "
DescriptorTest Setε1ε2ε3
Min. Clearance0.46/0.460.50/0.460.74/0.700.86/0.74
Max. Clearance0.44/0.460.41/0.490.46/0.550.50/0.60
Avg. Clearance0.53/0.550.50/0.520.70/0.570.73/0.60
Path Length0.58/0.700.86/0.910.92/0.860.94/0.94
Smoothness0.74/0.720.86/0.820.90/0.900.90/0.89
# of Sharp Turns0.76/0.730.80/0.840.77/0.700.60/0.55
Max. Angle0.71/0.700.82/0.840.86/0.880.94/0.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.206, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Table 3. GPT-4o's fine-grained perception accuracy (2D/3D) on the test set of PATHEVAL and three additionally synthesized datasets with increasing metric differences." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.257, + 0.442, + 0.274 + ], + "angle": 0, + "content": "4.2. What Can Vision-Language Models See?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.386 + ], + "angle": 0, + "content": "The previous subsection clearly highlights the vision component as the bottleneck for path evaluation on PATHEVAL. In this subsection, we conduct an analysis of the visual perception capabilities of VLMs. We focus our analysis on GPT-4o since it is the only model that performs substantially better than the random baseline in the case where no descriptors are required." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.569 + ], + "angle": 0, + "content": "GPT-4o can perceive paths that are significantly different. In order to get a better understanding of the model's perception capability, we break down its performance in terms of perception by individual metric. Specifically, we prompt GPT-4o to select which path in the pair provides a lower value on each individual metric and report its accuracy. We perform this analysis on both the task pairs in the test set of PATHEVAL and an additionally synthesized dataset consisting of task pairs with three levels (denoted as \\(\\epsilon_{1}\\), \\(\\epsilon_{2}\\), and \\(\\epsilon_{3}\\)) of increasing differences in their descriptor values. We describe the data generation process for this dataset in more detail in Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.813 + ], + "angle": 0, + "content": "The results for both experiments are shown in Table 3. Upon evaluating the performance on the test set of PATHEVAL by individual metrics, we notice that GPT-4o particularly struggles with Clearance metrics. These metrics typically require a lower level of perception and is naturally more challenging to discriminate than other metrics. On the other hand, Smoothness, Number of Sharpest Turns, and Max Angle appear to be easier for the model to capture. We also note that GPT-4o perceives the path length much more easily in a 3D environment presentation. Furthermore, GPT-4o's performance increases as we increase the descriptor difference between paths. This, however, is not the case for the Number of Sharp Turns. We conjecture that when we increase the number of sharp turns, we also enforce the challenge of \"counting\" the number of satisfying turns, which VLMs have been shown with limitations [46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Does segment complexity affect performance? We look into whether the complexity of the path segment is the key reason for GPT-4o's limited perception of clearance. To this end, we test GPT-4o on segments of varying complexity (i.e., points, straight lines, and curves), in a simplified" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.617, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.184, + 0.613, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.226, + 0.59, + 0.238 + ], + "angle": 0, + "content": "(a) Point" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.089, + 0.761, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.185, + 0.757, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.226, + 0.753, + 0.239 + ], + "angle": 0, + "content": "(b) Straight line" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.089, + 0.905, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.185, + 0.902, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.226, + 0.882, + 0.238 + ], + "angle": 0, + "content": "(c) Curve" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.244, + 0.88, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.315, + 0.905, + 0.34 + ], + "angle": 0, + "content": "(d) GPT-40 performance on the distance to obstacle under different segment complexities." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.351, + 0.906, + 0.379 + ], + "angle": 0, + "content": "Figure 2. Example segment complexity test cases in simplified environments and performance across the various settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.605 + ], + "angle": 0, + "content": "environment with only one rectangular obstacle at the center (Figure 2), and evaluate its accuracy in identifying segments that are closer to the obstacle. For individual points, the clearance is defined by the perpendicular distance from the point to the obstacle; for straight lines and curves, we consider a path closer if one of its endpoints is closer to the obstacle. For each segment type, we synthesize 100 pairs by first randomly generating 200 segments and then pairing each with the segment with the greatest distance difference from the obstacle relative to it (i.e., maximizing the absolute difference between the distances of the two segments from the obstacle). This increases the likelihood that the distance is significant enough to be perceivable. The average difference in the clearances of the pairs of segments are 14.76, 14.42, and 14.28 for points, lines, and curves respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.833 + ], + "angle": 0, + "content": "The results in Figure 2d show that GPT-4o can perform better in very easy scenarios; however, it struggles more as the segment complexity increases. For instance, the model was able to identify the closer points in \\(77\\%\\) of the cases; however, when considering straight and curved lines, its performance drops to \\(64\\%\\) and \\(59\\%\\), respectively. A surprising observation is that, in the case of curved lines, GPT-4o's performance is dramatically better in 3D images. As shown in Appendix C.2, the average Clearance difference of path pairs in PATHEVAL is merely 0.12 - 1.31. As the paths in PATHEVAL are much more complicated than the curves in this experiment, it is expected that GPT-4o exhibits difficulty in judging paths' clearances. The complexity of the environments (compared to a single square obstacle) could add challenges." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.783, + 0.862 + ], + "angle": 0, + "content": "5. Fine-tuning a Path Evaluator" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "One intuitive question is whether simply fine-tuning the VLMs can relieve their vision bottleneck. To answer this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6892" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.209, + 0.089, + 0.791, + 0.204 + ], + "angle": 0, + "content": "
ModelFrozenFine-tuned
AccuracyAvg. Cosine SimilarityAccuracyAvg. Cosine Similarity
clip-vit-base-patch320.5100.9140.7830.514
clip-vit-large-patch14-3360.4980.9070.7490.548
OpenCLIP-vit-B-320.5400.8830.7430.475
siglip-base-patch16-2240.5290.8950.7310.612
dino-vit-160.4950.9110.7630.754
dino-v2-base0.5100.7610.7210.681
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.215, + 0.905, + 0.229 + ], + "angle": 0, + "content": "Table 4. Probing accuracy and average cosine similarity between distinct path images (2D) when the vision encoder is frozen or fine-tuned." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.235, + 0.432, + 0.3 + ], + "angle": 0, + "content": "
SettingFrozenTunable
Image Only0.520.51
Image w Descriptors0.96-
Image Only (Random IDs)0.480.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.31, + 0.483, + 0.339 + ], + "angle": 0, + "content": "Table 5. Fine-tuned LLaVA-NeXT-7b performance (2D) when we keep the vision encoder frozen or tunable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.469 + ], + "angle": 0, + "content": "question, we experiment with LLaVA-NeXT-7b and fine-tune it on the training set (13,500 pairs) of PATHEVAL. We focus on the 2D case for the set of experiments described in this section. We consider three separate settings for training: (1) training with only images as input, (2) training with images and textual descriptor values as input, and (3) the same setting as (1) with using random IDs as target labels. Details in experimental setup are included in Appendix C.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.48, + 0.291, + 0.495 + ], + "angle": 0, + "content": "5.1. Overall Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Fine-tuning does not help with vision-language mapping. Table 5 shows that the model fails to learn meaningful patterns in the data, even after training 50 epochs. However, when the textual descriptors are provided as input, the model can easily learn the function achieving \\(96\\%\\) accuracy (a \\(45\\%\\) improvement from the zero-shot setting). This shows that the model is unable to extract the same descriptor values from the image input. Unfreezing the encoder for fine-tuning also does not provide any significant improvement. We include a further discussion in Appendix D.3. The results point to a limitation in the vision model's ability to encode the images, which we will investigate next." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.696, + 0.452, + 0.712 + ], + "angle": 0, + "content": "5.2. Understanding the Visual Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In order to further understand the limitations of the vision component, we conduct an analysis to better understand how well different vision encoder models can differentiate between different paths in the dataset. To this end, we first apply a linear probe to see how easily distinguishable different images are. Specifically, given a pair of images, we first use the vision encoder to extract high-dimensional feature representations for both images. These features are then concatenated and passed through a simple binary classification layer (i.e., the probe). The probe is trained to predict a label of 1 if the images are the same and 0 otherwise. We experiment with various SOTA vision encoders, namely CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[55] base and large varieties, LAION-OpenCLIP [25, 56], SigLip [80], DINO [8], and DINO-v2 [49] and analyze how well their learned representations capture visual similarities and differences. We use a set of 1,000 randomly sampled path pairs with a balanced label distribution to train the probe, and look at whether the model can learn to distinguish between these paths. We also perform an experiment where we fine-tune the vision encoder along with the probe. In this setting, our goal is to gauge if carefully fine-tuning a vision encoder can potentially improve the model's performance in low-level perception. Finally, in both settings, we present the average cosine similarity between distinct paths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.905, + 0.512 + ], + "angle": 0, + "content": "Vision encoders cannot distinguish between paths. From Table 4, it can be seen that vision encoder models are unable to provide representations that are significantly different for the probe to tell if they are the same. This is further supported by the high values for the average cosine similarity across all models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.682 + ], + "angle": 0, + "content": "Fine-tuning the encoders on a discrimination task can help disentangle the visual representations. By unfreezing the encoder weights and training them to identify whether two paths are identical, we enhance their adaptability to the task. The results in Table 4 demonstrate that this approach effectively disentangles the learned representations, resulting in significantly improved performance and increased separability, as evidenced by the notably lower cosine similarity between non-identical paths. The results thus imply the need for carefully fine-tuning task-specific vision encoders for path evaluation on PATHEVAL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.707, + 0.634, + 0.722 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This work explored the use of VLMs as evaluators in path-planning scenarios. We evaluated a number of VLMs on our proposed benchmark, PATHEVAL, and found that these models struggle with low-level perception. Specifically, we find that visual encoders used by SOTA models are unable to discern the differences between different paths in intricate scenarios. We hope that PATHEVAL will inspire researchers to further explore ways to improve the visual capabilities of VLMs and contribute to finding better ways to incorporate foundation models for developing more flexible, robust, and scalable planning paradigms." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6893" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This project was supported by resources provided by the Office of Research Computing at George Mason University (URL: https://orc.gmu.edu) and funded in part by grants from the National Science Foundation (Award Number 2018631)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.209, + 0.188, + 0.225 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[1] Chirag Agarwal, Sree Harsha Tanneru, and Himabindu Lakkaraju. Faithfulness vs. Plausibility: On the (Un)Reliability of Explanations from Large Language Models, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.292, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[2] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can Large Language Models be Good Path Planners? A Benchmark and Investigation on Spatial-temporal Reasoning. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.363, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Look Further Ahead: Testing the Limits of GPT-4 in Path Planning. IEEE CASE, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.407, + 0.484, + 0.463 + ], + "angle": 0, + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In International Conference on Computer Vision (ICCV), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.465, + 0.484, + 0.59 + ], + "angle": 0, + "content": "[5] Kate Baumli, Satinder Baveja, Feryal Behbahani, Harris Chan, Gheorghe Comanici, Sebastian Flennerhag, Maxime Gazeau, Kristian Holsheimer, Dan Horgan, Michael Laskin, Clare Lyle, Hussain Masoom, Kay McKinney, Volodymyr Mnih, Alexander Neitz, Dmitry Nikulin, Fabio Pardo, Jack Parker-Holder, John Quan, Tim Roktaschel, Himanshu Sahni, Tom Schaul, Yannick Schroecker, Stephen Spencer, Richie Steigerwald, Luyu Wang, and Lei Zhang. Vision-Language Models as a Source of Rewards, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.592, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.733, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[7] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21819–21830, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.234, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.203 + ], + "angle": 0, + "content": "[9] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee Wong. MapGPT: Map-guided prompting with adaptive path planning for vision-and-language navigation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9796-9810, Bangkok, Thailand, 2024. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[10] Xuweiyi Chen, Ziqiao Ma, Xuejun Zhang, Sihan Xu, Shengyi Qian, Jianing Yang, David Fouhey, and Joyce Chai. Multi-object hallucination in vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[11] Yongchao Chen, Jacob Arkin, Charles Dawson, Yang Zhang, Nicholas Roy, and Chuchu Fan. Autotamp: Autoregressive task and motion planning with llms as translators and checkers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6695-6702, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.906, + 0.43 + ], + "angle": 0, + "content": "[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. arXiv preprint arXiv:2312.14238, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[14] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[15] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.906, + 0.783 + ], + "angle": 0, + "content": "[16] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, brian richter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B. Tenenbaum, Leslie Pack Kaelbling, Andy Zeng, and Jonathan Tompson. Video language planning. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[17] Abhimanyu Dubey et. al. The Llama 3 Herd of Models, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[18] Difei Gao, Ruiping Wang, Shiguang Shan, and Xilin Chen. Cric: A vqa dataset for compositional reasoning on vision and commonsense. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5561-5578, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6894" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[20] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[21] Lin Guan, Yifan Zhou, Denis Liu, Yantian Zha, Heni Ben Amor, and Subbarao Kambhampati. Task success is not enough: Investigating the use of video-language models as behavior critics for catching undesirable agent behaviors. In First Conference on Language Modeling, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[22] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[23] Xu Han, Qiannan Yang, Xianda Chen, Xiaowen Chu, and Meixin Zhu. Generating and evolving reward functions for highway driving with large language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[24] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.566, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's \"up\" with vision-language models? Investigating their struggle with spatial reasoning. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[27] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Audit Verma, Kaya Stechly, Siddhant Bhambri, Lucas Paul Saldyt, and Anil B Murthy. Position: LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks. In _Forty-first International Conference on Machine Learning_, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[28] L.E. Kavraki, P. Svestka, J.-C. Latombe, and M.H. Overmars. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and Automation, 12(4):566-580, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[29] J.J. Kuffner and S.M. LaValle. RRT-connect: An efficient approach to single-query path planning. In Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065), pages 995-1001 vol.2, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[30] Steven M. LaValle. Rapidly-exploring random trees: a new tool for path planning. The annual research report, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[31] Jan Leike. Why i'm optimistic about our alignment approach: Evaluation is easier than generation." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "https://aligned.substack.com/i/88447351/evaluation-is-easier-than-generation,2022.Accessed:2024-11.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[32] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[33] Hao Li, Xue Yang, Zhaokai Wang, Xizhou Zhu, Jie Zhou, Yu Qiao, Xiaogang Wang, Hongsheng Li, Lewei Lu, and Jifeng Dai. Auto mc-reward: Automated dense reward design with large language models for minecraft. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[35] Yizhi Li, Ge Zhang, Yinghao Ma, Ruibin Yuan, Kang Zhu, Hangyu Guo, Yiming Liang, Jiaheng Liu, Zekun Wang, Jian Yang, Siwei Wu, Xingwei Qu, Jinjie Shi, Xinyue Zhang, Zhenzhu Yang, Xiangzhou Wang, Zhaoxiang Zhang, Zachary Liu, Emmanouil Benetos, Wenhao Huang, and Chenghua Lin. OmniBench: Towards The Future of Universal Omni-Language Models, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[36] Hui Sheng Lim, Shuangshuang Fan, Christopher K.H. Chin, Shuhong Chai, Neil Bose, and Eonjoo Kim. Constrained path planning of autonomous underwater vehicle using selectively-hybridized particle swarm optimization algorithms. IFAC-PapersOnLine, 52(21):315-322, 2019. 12th IFAC Conference on Control Applications in Marine Systems, Robotics, and Vehicles CAMS 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[41] Tomás Lozano-Pérez and Michael A. Wesley. An algorithm for planning collision-free paths among polyhedral obstacles. Communications of the ACM, 22(10):560-570, 1979. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6895" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[43] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[44] Silin Meng, Yiwei Wang, Cheng-Fu Yang, Nanyun Peng, and Kai-Wei Chang. LLM-a*: Large language model enhanced incremental heuristic search on path planning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1087-1102, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[45] Shota Onohara, Atsuyuki Miyai, Yuki Imajuku, Kazuki Egashira, Jeonghun Baek, Xiang Yue, Graham Neubig, and Kiyoharu Aizawa. Jmmm: A japanese massive multidiscipline multimodal understanding benchmark for culture-aware evaluation, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.319, + 0.406, + 0.332 + ], + "angle": 0, + "content": "[46] OpenAI. GPT-4V(ision) System Card. 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.334, + 0.378, + 0.348 + ], + "angle": 0, + "content": "[47] OpenAI. GPT-4o System Card, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.35, + 0.38, + 0.363 + ], + "angle": 0, + "content": "[48] OpenAI. GPT-4 Technical Report, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.365, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.483, + 0.575 + ], + "angle": 0, + "content": "[50] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 950-974, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[51] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14760-14769, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[52] Clment Petres, Yan Pailhas, Pedro Patron, Yvan Petillot, Jonathan Evans, and David Lane. Path planning for autonomous underwater vehicles. IEEE Transactions on Robotics, 23(2):331-341, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[53] Evis Plaku, Erion Plaku, and Patricio Simari. Clearance-driven motion planning for mobile robots with differential constraints. Robotica, 36(7):971-993, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[54] Patrick A. Plonski, Pratap Tokekar, and Volkan Isler. Energy-Efficient Path Planning for Solar-Powered Mobile Robots, pages 717-731. Springer International Publishing, Heidelberg, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pam Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "the 38th International Conference on Machine Learning, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[57] Jie-Jing Shao, Jiang-Xin Shi, Xiao-Wen Yang, Lan-Zhe Guo, and Yu-Feng Li. Investigating the Limitation of CLIP Models: The Worst-Performing Categories, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[58] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B. Tenenbaum, Leslie Pack Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[59] Loan A. Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics & Automation Magazine, 19(4):72-82, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[60] Yinan Sun, Zicheng Zhang, Haoning Wu, Xiaohong Liu, Weisi Lin, Guangtao Zhai, and Xiongkuo Min. Explore the hallucination on low-level perception for mllms, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.907, + 0.537 + ], + "angle": 0, + "content": "[61] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[62] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[63] Shengbang Tong, Erik Jones, and Jacob Steinhardt. Mass-producing failures of multimodal systems with language models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[64] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.68, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6896" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.161 + ], + "angle": 0, + "content": "Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[66] Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language models don't always say what they think: unfaithful explanations in chain-of-thought prompting. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[67] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large Language Models Still Can't Plan (A Benchmark for LLMs on Planning and Reasoning about Change). In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[68] Karthik Valmeekam, Kaya Stechly, and Subbarao Kambhampati. LLMs Still Can't Plan; Can LRMs? A Preliminary Evaluation of OpenAI's o1 on PlanBench, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[69] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[70] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6622-6631, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[72] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[73] Tianbao Xie, Siheng Zhao, Chen Henry Wu, Yitao Liu, Qian Luo, Victor Zhong, Yanchao Yang, and Tao Yu. Text2Reward: Reward Shaping with Language Models for Reinforcement Learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[74] Yutaro Yamada, Yihan Bao, Andrew Kyle Lampinen, Jungo Kasai, and Ilker Yildirim. Evaluating spatial understanding of large language models. Transactions on Machine Learning Research, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[75] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[76] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[78] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI. In Proceedings of CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[79] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark. arXiv preprint arXiv:2409.02813, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[81] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, and Jie Fu. CMMU: A Chinese Massive Multi-discipline Multimodal Understanding Benchmark, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.905, + 0.556 + ], + "angle": 0, + "content": "[82] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.557, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[83] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy Improvement using Language Feedback Models, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.654 + ], + "angle": 0, + "content": "[84] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7W: Grounded Question Answering in Images. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6897" + } + ] +] \ No newline at end of file diff --git a/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_origin.pdf b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..04b93f23fa0a1da53f2e3f51138bba7371d180f8 --- /dev/null +++ b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/ea2b9d98-7774-421e-b29a-893befcc86ee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c0e3ecdda8035a440d6e4cc27c3993bbe636af3e28b02826ea0180c5853b6eb +size 1312072 diff --git a/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/full.md b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ecd1ecfd261a8a6c3c646465d4675798e7d6fe11 --- /dev/null +++ b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/full.md @@ -0,0 +1,329 @@ +# Evaluating Vision-Language Models as Evaluators in Path Planning + +Mohamed Aghzal1, Xiang Yue2, Erion Plaku3*, Ziyu Yao1 +1George Mason University, 2Carnegie Mellon University, 3National Science Foundation {maghzal, ziyuyao}@gmu.edu, xyue2@andrew.cmu.edu, eplaku@nsf.gov + +# Abstract + +Despite their promise to perform complex reasoning, large language models (LLMs) have been shown to have limited effectiveness in end-to-end planning. This has inspired an intriguing question: if these models cannot plan well, can they still contribute to the planning framework as a helpful plan evaluator? In this work, we generalize this question to consider LLMs augmented with visual understanding, i.e., Vision-Language Models (VLMs). We introduce PATHEVAL, a novel benchmark evaluating VLMs as plan evaluators in complex path-planning scenarios. Succeeding in the benchmark requires a VLM to be able to abstract traits of optimal paths from the scenario description, demonstrate precise low-level perception on each path, and integrate this information to decide the better path. Our analysis of state-of-the-art VLMs reveals that these models face significant challenges on the benchmark. We observe that the VLMs can precisely abstract given scenarios to identify the desired traits and exhibit mixed performance in integrating the provided information. Yet, their vision component presents a critical bottleneck, with models struggling to perceive low-level details about a path. Our experimental results show that this issue cannot be trivially addressed via end-to-end fine-tuning; rather, task-specific discriminative adaptation of these vision encoders is needed for these VLMs to become effective path evaluators.[12] + +# 1. Introduction + +The impressive capabilities of Large Language Models (LLMs) [6, 8, 71] and Vision-Language Models (VLMs) [82] have led to an increasing interest in applying them to automated motion planning and navigation tasks [50, 51]. However, the inherent limitations of these models in long- + +horizon planning have rendered them ineffective as end-to-end motion planners [2, 3, 11, 75]. This has made researchers wonder: if these models cannot be good motion planners themselves, can they still support a motion planning framework? Intuitively, using these models still holds the promise of significantly enhancing the motion planning framework, as they have learned extensive factual and commonsense knowledge that could benefit planning during their pre-training. As a result, there has been an emerging paradigm exploring how these models can be leveraged in combination with traditional methods [27]. + +One particularly interesting approach within this line of work involves using these models as plan evaluators. Motivated by the intuition that "evaluation is easier than generation" [31], several efforts have explored leveraging these models as "critics" to assess the quality of generated plans [5, 21, 73, 83]. However, most of these efforts have focused on scenarios that require only limited, high-level visual perception, without necessitating fine-grained or precise perceptual abilities. On the other hand, while there have been prior works similarly investigating VLMs' low-level perception [26, 64], studies specifically about the use of these models in planning remain limited. Tasks such as motion planning often require fine-grained visual understanding in highly specific contexts, while also drawing on broad commonsense knowledge acquired during pre-training. Hence, there is a pressing need to investigate the potential of VLMs to understand both low-level visual details and leverage these visual signals for high-level reasoning. + +In this work, we explore whether we can utilize VLMs as evaluators in highly intricate continuous path planning problems. We introduce PATHEVAL (Figure 1), a controllable benchmark designed around path planning in complex environments under diverse commonsense decision-making scenarios (e.g., A firefighting robot has to cover as much of the area as possible to extinguish fires and scout for survivors, thus needs to prioritize paths with more coverage and higher clearance). Traditionally, encoding constraints in a planning algorithm requires significant human effort in hand-crafting scenario-specific criteria for path evaluation, such that the constraints can be injected into the planner + +# Example of an instance of PATHEVAL + +![](images/0cd1b2523106083092d842b736f974c7c1231d5c5a89f94d796fae60e61758db.jpg) +2D + +![](images/0838274835244e4bdc4acc68be840e06f4ef52695e7f8d98d7b664cd06a6bd3e.jpg) +3D + +![](images/e9583cea04a42a866a2085509284da1af944969998b3d5988bc342156a8d6e63.jpg) +Figure 1. Example of PATHEVAL. The benchmark consists of 14,550 instances of path pairs rendered in both 2D and 3D and mapped to 15 decision-making scenarios. Success on this task is tied to three distinct levels: 1) Attribute abstraction: recognizing what aspects make a path ideal, 2) Low-level perception: extracting the required attributes for each path from the images, and 2) Information integration: synthesizing the collected information to make a decision. We test a set of VLMs on the task and find that they struggle particularly with low-level perception. Incorrect answers by different models are shown (explanations indicating misperception are highlighted). + +# Path Descriptors: + +Path 1: Min. clearance: 1.07, Max. clearance: 8.91, Avg. clearance: 2.62, Path length: 243.46, Smoothness: 843.07, Number of sharp turns: 4, Maximum angle: 124.35. +Path 2: Min. clearance: 1.57, Max. clearance: 8.91, Avg. clearance: 3.18, Path length: 243.75, Smoothness: 809.65, Number of sharp turns: 1, Maximum angle: 93.12. + +Scenario #1: The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic. + +Required Descriptors: Smoothness $(\downarrow)$ , Number of Sharp turns $(\downarrow)$ , Maximum angle $(\downarrow)$ + +Ground Truth: Path 2 + +Scenario #2: An autonomous firefighting robot is designed to navigate and operate within burning buildings to extinguish fires and rescue trapped individuals. It should explore as much of the area as possible, while maintaining a safe distance from the walls to avoid damage. + +Required Descriptors: Avg. Clearance $(\uparrow)$ , Min. Clearance $(\uparrow)$ , Path length $(\uparrow)$ + +Ground Truth: Path 1 + +Scenario #3: As the vehicle is traversing a warzone, it must stay concealed from enemy operatives, making use of covers like walls and avoiding open spaces as much as possible. It should also reach its target as quickly as possible. + +Required Descriptors: Avg. Clearance (↓), Max. Clearance (↓), Path length (↓) + +Ground Truth: Path 2 + +# Prompt + +The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic. Which path better achieves the task? Path 1 is on the left side and Path 2 is on the right side. Your answer should follow the format below: + +Answer: Path 1 or Path 2. + +Explanation: Why you chose the path (1 or 2). + +# GPT-40 + +Answer: Path 1. + +Explanation: Path 1 has fewer sharp turns and maintains longer straight segments compared to Path 2. This is more suitable for a large truck navigating through heavy traffic, as it minimizes the need for difficult maneuvers and allows for smoother driving. Path 2, on the other hand, has more frequent and sharper turns, which would be more challenging for a large truck to navigate. + +# LLaVA-NeXT-7b + +Answer: Path 1. + +Explanation: Path 1 is better suited for a large truck because it has fewer sharp turns (90 degrees or larger) compared to Path 2. The truck would have to make only one 90-degree turn, which is easier to execute than the two 90-degree turns in Path 2. Additionally, Path 1 has more straight sections, which would be more efficient for a large truck to navigate through heavy traffic. The fewer turns and the greater number of straight sections make Path 1 the more suitable route for a large truck. + +during the search process [36, 52, 54]. However, this not only is a tedious and costly process but also cannot scale up to handle the countless intricate scenarios in meeting real-life planning needs. An effective VLM as a "reward model" + +relieving humans from customizing evaluation criteria for specific scenarios could enable scalable, general-purpose planners that can adapt to diverse commonsense scenarios described in natural language. + +In PATHEVAL, a VLM is tasked with comparing two paths within a given decision-making context and selecting the one that better satisfies the constraints outlined by the scenario. Success on this task requires effective performance across three distinct levels: 1) attribute abstraction: recognizing the attributes that define a favorable path in a particular decision-making context; 2) low-level perception: demonstrating precise low-level perception to determine which path performs better based on the given criteria; and 3) information integration: integrating and synthesizing the perceived information to produce an answer. + +Using this benchmark, we analyze the performance of 9 state-of-the-art (SOTA) VLMs, including both the closed-source GPT-4o and GPT-4o-mini [47] and 7 different open-source VLMs of different sizes (i.e., LLaVA-NeXT 7b and 13b [39], Qwen2-VL-7b [69], LLaVA-OneVision-7b [32], LLaMA-3.2-11b [17], and Intern-VL2 8b and 40b [12]). We find that these models struggle with the path evaluation task (e.g., Qwen2-VL-7b achieving only $50.2\%$ accuracy). However, when providing these VLMs with verbalized path specifications, their performance significantly improves (e.g., $74.2\%$ accuracy for Qwen2-VL-7b), which reveals a potential vision bottleneck of these VLMs. Our further analysis confirms these models' weakness in low-level perception, especially when they are tasked to perceive the clearance of a path with respect to surrounding obstacles, and this weakness could be more prominent when the environment and the path representation become more complex. We discover the source of this weakness from the vision encoders used by these VLMs, yet simply fine-tuning the VLMs end-to-end with the vision encoders does not address the issue. Rather, our experiments suggest performing task-specific discriminative adaptation of these vision encoders. + +# 2. Related Work + +# 2.1. Vision-Language Models + +The outstanding success of decoder-only LLMs [6, 48, 65] has driven the development of Vision-Language Models (VLMs), which extend LLMs with a vision component in an attempt to generalize their performance into a multi-modal setting [12, 17, 32, 37, 38, 62, 69]. VLMs are designed for tasks that require unifying visual and linguistic representations (e.g. visual question answering [4]). Typically, this is achieved by connecting a vision encoder to a language decoder and projecting the representations into a shared space, allowing visual elements and linguistic components to be linked in a semantically meaningful manner. One of the earliest successful models at achieving this was the CLIP encoder [55], which was trained using contrastive learning to learn representations that map images and their corresponding textual descriptions. Several varieties of CLIP have then been introduced [25, 56, 76, 80]. These mod + +els, while showing tremendous promises have shown several limitations when it comes to visual tasks [57]; as such several works [64, 72] have sought ways to improve such representations by combining them with vision foundation models such as DINO [8, 49]. In this work, we contribute to the research of VLMs with a new benchmark, namely PATHEVAL, focusing on evaluating VLMs as evaluators for path planning. This benchmark tests VLMs seamlessly on their commonsense understanding (i.e., being able to abstract critical concepts expressed in the described planning scenarios), low-level perception (i.e., precisely perceiving details about paths in complex environments), and the ability to reason about the collected textual and visual information for decision-making. + +# 2.2. Automated Planning with VLMs + +Vision-language planning promises more flexible planning frameworks and enhanced human-AI interaction. Therefore, designing systems that can effectively understand natural language instruction and leverage perceptual inputs to conduct planning tasks has been a topic of interest in recent years [20, 61, 70]. The rise of VLMs has led many to investigate the use of these models as vision-language planning agents [15, 16, 50, 51]. However, existing literature highlights the limitations of LLMs in spatial reasoning [2, 26, 74] as well as long-horizon planning [3, 67, 68]. This pushed researchers to explore alternative ways to incorporate VLMs and LLMs into planning frameworks more reliably [27, 58]. One potential direction is their use as plan evaluators, either through the generation of reward functions [23, 33, 73], or by using them directly as off-the-shelf critics [5, 21, 83]. The success of such frameworks assumes perfect perception and that the models can accurately perceive visual information and reason about it in order to produce an answer; nevertheless, it has been shown that the representations used by these models fail in highly intricate visual settings [64]. Several works have explored the use of VLMs as well as LLMs for path and motion planning [2, 3, 7, 9, 11, 14, 44, 75], however, to the knowledge of the authors there is no work that explores the use of VLMs as path critics in this context. Accordingly, we aim to evaluate the ability of VLMs to serve as evaluators in cases requiring navigation in complex environments while adhering to decision-making constraints specified in natural language. + +# 2.3. Vision-Language Model Benchmarks + +The introduction of multimodal models has prompted the development of several benchmarks that are capable of assessing the performance of these models on visual reasoning tasks such as visual question-answering datasets [4, 18, 19, 24, 42, 43, 84]. However, the rise of foundation models has produced the need for a more holistic evaluation of the perceptual and reasoning capabilities of large + +VLMs, leading to benchmarks such as MM-Vet [77], MM-Bench [40], MMMU [45, 78, 79, 81] and OmniBench [35]. Several benchmarks have also specifically been designed to assess the perception capabilities of these models and explore the limitations associated with visual hallucinations and optical illusions [10, 22, 34, 60, 64]. Our proposed benchmark provides a flexible yet challenging framework for interleavedly assessing the low-level perception and reasoning capabilities of VLMs. + +# 3. The PATHEVAL Benchmark + +Motivated by the need to evaluate VLMs as path evaluators in real-world planning scenarios, we introduce PATHEVAL, a controllable and extensible benchmark focused on path planning [28-30, 41] in complex environments under a diverse set of decision-making constraints. We list the decision-making scenarios as well as the descriptors they attempt to optimize in Appendix A. In total, PATHEVAL includes 14,550 tasks over more than 1,150 distinct environments and 15 distinct scenarios. Below, we introduce a formal description of the task and the dataset construction. + +# 3.1. Task Formulation + +Given two paths, $P_{1}$ and $P_{2}$ , and a scenario $S$ , the objective is to determine which path better satisfies the scenario's optimization criteria. Each scenario $S$ is a high-level description that aims to optimize over a set of path descriptors (or metrics) $\mathcal{M} = \{m_1, m_2, \dots, m_k\}$ , where each descriptor $m_j : P \to \mathbb{R}$ evaluates a specific property of a path (e.g., length, smoothness, or proximity to obstacles). A VLM $\mathcal{V}$ is presented with two images presenting $P_{1}$ and $P_{2}$ in the same environment, respectively. The model must then decide which path better satisfies the scenario's criteria. To explore the sensitivity of VLMs to the way how a path is presented, PATHEVAL includes both the 2D and 3D images of the path illustration. The model is also prompted to generate an explanation to justify its choice. + +# 3.2. Environment and Path Generation + +Environment Generation: An environment, as shown in Figure 1, is defined by a set of walls $\mathcal{O} = \{O_1,O_2,\dots ,O_n\}$ , where each wall $O_{i}$ represents an obstacle in the 2D space. Each wall is a closed geometric shape described by its vertices, and the set $\mathcal{O}$ forms the obstacles that the path must avoid. In this work, we leverage the environments of Plaku et. al. [53], which consists of four types of obstacle arrangements: (1) rings, where the environments are structured as mazes with circular walls, (2) waves, which consist of wavy horizontal obstacles, (3) mazes, which consist of both vertical and horizontal walls forming a complex maze structure, as well as (4) random, which consist of randomly placed obstacles. + +Path Synthesis via the Randomly-exploring Rapid Tree (RRT) algorithm: To generate path candidates in PATHVAL, we leverage the RRT path planning algorithm [30]. Starting from the initial location in the environment, the algorithm works by building a tree which expands itself by randomly selecting the next location in the environment while avoiding obstacles, until it reaches the goal. In this work, we use the Open Motion Planning Library (OMPL) [59] and implement the RRT-Connect algorithm [29]. We note that while we use RRT in the current benchmark, our codebase is adaptable and can incorporate most path-planning algorithms provided by OMPL. We encourage future research building on our benchmark to experiment with other algorithms as well. + +# 3.3. Path Descriptors + +We collect the following descriptors $\mathcal{M}$ for each of a generated path: Minimum Clearance measures the smallest distance between any point on the path and the nearest obstacle; Maximum Clearance measures the largest distance between any point on the path and the nearest obstacle; Average Clearance computes the average distance between all points on the path and the nearest obstacle; Path Length is calculated by summing up the Euclidean distances between consecutive points on the path; Smoothness is defined as the sum of the angles between consecutive segments of the path, measuring how smoothly the path changes direction; Number of Sharp Turns counts the number of turns in the path where the angle between consecutive segments exceeds 90 degrees; and Maximum Angle denotes the largest angle between any two consecutive segments of the path. The three Clearance metrics and Path Length share the same measuring unit, i.e., one grid size; Smoothness and Maximum Angle are measured by degree; and the Number of Sharp Turns is an integer count. We include the formula of each descriptor in Appendix B. + +# 3.4. Natural Language Descriptions of Scenarios + +To create a sufficiently challenging path-planning evaluation benchmark, we design a total of 15 decision-making scenarios that aim to optimize different combinations of the path descriptors. For instance, Scenario #2 (as shown in Figure 1) requires searching through an area affected with fire in search for survivors, and the agent thus must cover as much ground as possible. In contrast, Scenario #3 indicates that the path is to be executed within a warzone; as a result, the vehicle has to remain hidden and take the shortest route. As such, given the same set of paths, the one minimizing the path length is favored by Scenario #3 while Scenario #2 needs to maximize this value. A complete summary of the 15 scenarios, along with the descriptors each scenario aims to optimize, is presented in Tables 6-7 in Appendix A. + +# 3.5. Task Pairs Selection + +For each environment we synthesized, we ran the RRT planner 30 times to generate different paths. Upon eliminating paths that did not reach the goal, we selected path pairs that exhibited the greatest dissimilarity in terms of path descriptors. Specifically, we first represented each path using a (7-dimension) vector of its path descriptor values. Given that each path descriptor ranged in a dramatically different scale, we normalized the vector by performing the min-max scaling, i.e., scaling each value $x$ in the vector to $x' = \frac{x - x_{\min}}{x_{\max} - x_{\min}}$ , where $x$ represents each value in the vector, and $x_{\min}$ and $x_{\max}$ are the minimum and maximum values of each descriptor across paths sampled from the same environment, respectively. We then measured the distance between two paths by calculating the Euclidean distance between their normalized path descriptor vectors, and selected 5 path pairs with the largest distances. Each path is included in only one pair to avoid redundancy. We repeated the same selection procedure for a total of 1,150 environments. + +Upon performing this generation, we matched each pair with all fifteen scenarios; however, we only kept pairs where there was a significant difference in at least one of the descriptors required for the particular scenario. In other words, we ensure that the absolute difference is greater than a predefined threshold (0.8 for clearance descriptors, 50 for path length, 90 for smoothness, 1 for the number of sharp turns, and 30 for maximum angle) for at least one of the required descriptors. This makes it more likely that the difference is more noticeable to the naked eye, and thus the paths can be compared by visual inspection. + +We constructed our final evaluation benchmark by randomly selecting 70 task pairs from each scenario, resulting in 1,050 pairs in total. The remaining task pairs (a total of 13,500) are used as the training set to facilitate the fine-tuning experiments in Section 5. + +# 4. Can SOTA VLMs Evaluate Planned Paths? + +In order for VLMs to perform successfully in our benchmark, they need to perform successfully at three different levels, i.e., recognizing the critical descriptors required by each scenario (Attribute Abstraction), exhibiting sharp low-level perception to precisely assess each path's properties (Low-level Perception), and integrating the prior information to make a rational decision on the better path (Information Integration). Among them, the first two levels reflect parallel properties that serve as a foundation for the third level. In this section, we evaluated a set of 9 VLMs on PATHEVAL in a zero-shot manner and analyzed their capabilities at these three levels. These VLMs include (1) two closed-source VLMs, i.e., GPT-4o and GPT-4o-mini [47], and (2) seven open-source VLMs with various sizes, including LLaVA-NeXT-7b and LLaVA-NeXT-13b [38, 39], + +LLaVA-OneVision-7b [32], Qwen2-VL-7b [69], LLaMA3.2-11b [17], and Intern-VL2-8b and Intern-VL2-40b [12]. We include all prompt scripts used in this Section in Appendix C.1. + +# 4.1. Overview of VLMs on PATHEVAL + +The performance of the 9 VLMs on PATHEVAL is showcased in Table 1. We notice that all of the models, except GPT-4o, fail to perform significantly better than a simple random baseline, indicating significant limitations. For GPT-4o, we also notice a $4\%$ higher accuracy on 2D images compared to prompting with 3D inputs. This observation indicates that the model is prone to visual illusions introduced by the 3D images, when it has to rely on solely the image for decision-making (although we observe an opposite effect of 2D vs. 3D when verbalized descriptor values are provided). In the remaining section, we will further break down these models' capabilities to gain a deeper understanding of their failure on this task. + +Providing verbalized path information yields better task accuracy, implying reasonable VLM performance in attribute abstraction. In Table 1, we further show the performance of each model when we explicitly list the value for each descriptor as part of the language prompt (i.e., "PATHEVAL w Desc.). We notice a $11.1\% -27.1\%$ improvement across most models, indicating that when given low-level details, the models can better filter out the information and make better comparisons. This points out that the bottleneck for these VLMs's better performance lies in their inability to accurately perceive low-level information about the paths (we discuss this in more detail in Section 4.2), whereas these models generally have a reasonable capability in abstracting the critical attributes for decision making in various scenarios. In Appendix D.1, we include an experiment where we explicitly query each VLM to identify the critical path metrics for each scenario; the result corroborates our hypothesis. In particular, we find that for most models, the success rate for identifying a required descriptor is over $92\%$ . Finally, a surprising observation happens to LLaVa-NeXT-7b. We notice that this model suffers particularly severely from hallucination; even when the textual descriptor values are provided and when the model can correctly identify critical path metrics based on our analysis in Appendix D.1, it fails to pick the better paths. We show an example of its explanation when textual descriptors are provided in Figure 5 of Appendix E. + +VLMs exhibit mixed performance in integrating visual and textual path information. We take a further look into the model performance when both the image and the textual descriptor values are provided, and contrast it with their performance when only the textual path descriptions are provided (i.e., "Desc Only"). Interest- + +
Model2D3DDesc Only
PATHEVALPATHEVAL w Desc.PATHEVALPATHEVAL w Desc.
GPT-4o-mini0.5200.7500.5080.7450.680
GPT-4o0.6650.8600.6240.8950.894
LLaVa-NeXT-7b0.5010.5240.4990.5170.514
Qwen2-VL-7b0.5020.7310.5110.7420.737
LLaVA-OneVision-7b0.5050.7180.5090.7390.721
Intern-VL2-8b0.4890.6540.5050.6910.648
LLaMa-3.2-11b0.4800.6950.4600.6800.686
LLaVa-NeXT-13b0.5090.6200.4940.6010.630
Intern-VL2-40b0.5060.6880.4960.7170.679
Random Baseline0.5000.5000.5000.5000.500
+ +ingly, we observe mixed information integration performance from these models. For GPT-4o, Qwen2-VL-7b, LLaVA-OneVision-7b, LLaMa-3.2-11b, and LLaVa-NeXT-13b, their performance based on only descriptor values has no obvious difference to their best performance when (2D or 3D) images are also provided. This observation implies that these models do not benefit from the additional image information when the textual path descriptions are provided. Instead, sometimes the images (e.g., 2D images for GPT-4o and 3D images for LLaVa-NeXT-13b) may confuse their understanding of the textual descriptors, resulting in a worse accuracy compared to Desc Only. For GPT-4o-mini, InternVL2-8b, and Intern-VL2-40b, however, providing both the visual and textual path information offers $\sim 4\% - 7\%$ performance gain over Desc Only, indicating better information integration from these VLMs. + +In Appendix D.2, we present a breakdown of GPT-4o's performance by scenarios. We show that performance varies greatly from one scenario to the other. Interestingly, we observe that GPT-4o may overuse their commonsense knowledge. For instance, in the case of Scenario #2 shown in Figure 1, where the agent is required to maximize the path length for better coverage, GPT-4o still favors the shorter path. This scenario represents a counterfactual situation as models are often trained or instructed to seek the shortest paths. Evaluating VLMs in such counterfactual scenarios allows us to effectively probe their task understanding and reasoning, which we consider to be an important direction for future research. + +The reasoning of VLMs can be unreliable. Limited by the vision bottleneck, we noticed these VLMs fabricating seemingly plausible explanations to justify their path evaluation, despite the fact that they could not actually perceive the necessary path details to perform the task. This fabrication echos findings from other recent work, where LLMs were shown to produce reasoning traces that do not + +Table 1. Accuracy of VLMs on PATHEVAL based on 2D and 3D environment images. To investigate the potential vision bottleneck of VLMs, we additionally present each model's accuracy when explicit descriptor values are provided in the language prompts ("PATHEVAL w Desc"). The last column ("Desc Only") shows the model performance when only the textual descriptor values are provided. + +
ModelDefaultFlippedRandom IDs
GPT-4o-mini34/101622/102694/956
GPT-4o278/772258/792291/759
LLaVa-NeXT-7b1028/221041/9580/470
Qwen2-VL-7b21/1029130/920127/923
LLaVA-OneVision-7b438/612440/610262/788
Intern-VL2-8b433/610430/620418/632
LLaMA-3.2-11b*601/251620/264731/319
LLaVa-NeXT-13b795/255939/111305/745
Intern-VL2-40b394/656410/639510/540
Ground truth labels530/520
+ +Table 2. Performance on the 2D case (#of times first path is chosen / #of times second path is chosen) when we flip the path order or replace their default names with random IDs. *There are several cases where LLaMA-3.2-11b does not follow the required format and/or does not give an answer, we omit those cases from this table. + +accurately reflect the process of how the models reach an answer [1, 66]. To gain further insights, we performed an analysis comparing model performance on PATHEVAL with cases that consist of: 1) flipping the order of paths in the pair, and 2) assigning random IDs to the paths (e.g. instead of referring to them as "Path 1" and "Path 2", we use a random sequence such as "Path Xu2q" and "Path fP48"). The results presented in Table 2 showcase that VLMs demonstrate bias for a particular label, when they actually do not have the capability to resolve the task. For example, when no matter the default or the flipping labels are used, LLaVA-NeXT-7b consistently selects Path $198\%$ of the time and fabricates incorrect observations of the two paths in its explanations (Figure 1) to support this choice. As we discussed earlier, LLaVA-NeXT-7b is particularly prone to hallucination in explanations, leading to its random-guess performance with or without textual descriptors. Introducing random IDs as path names mitigates this bias for LLaVa-NeXT-7b (although the model still obtains a close-to-random accuracy on PATHEVAL) but does not seem to help other models dramatically. + +
DescriptorTest Setε1ε2ε3
Min. Clearance0.46/0.460.50/0.460.74/0.700.86/0.74
Max. Clearance0.44/0.460.41/0.490.46/0.550.50/0.60
Avg. Clearance0.53/0.550.50/0.520.70/0.570.73/0.60
Path Length0.58/0.700.86/0.910.92/0.860.94/0.94
Smoothness0.74/0.720.86/0.820.90/0.900.90/0.89
# of Sharp Turns0.76/0.730.80/0.840.77/0.700.60/0.55
Max. Angle0.71/0.700.82/0.840.86/0.880.94/0.96
+ +Table 3. GPT-4o's fine-grained perception accuracy (2D/3D) on the test set of PATHEVAL and three additionally synthesized datasets with increasing metric differences. + +# 4.2. What Can Vision-Language Models See? + +The previous subsection clearly highlights the vision component as the bottleneck for path evaluation on PATHEVAL. In this subsection, we conduct an analysis of the visual perception capabilities of VLMs. We focus our analysis on GPT-4o since it is the only model that performs substantially better than the random baseline in the case where no descriptors are required. + +GPT-4o can perceive paths that are significantly different. In order to get a better understanding of the model's perception capability, we break down its performance in terms of perception by individual metric. Specifically, we prompt GPT-4o to select which path in the pair provides a lower value on each individual metric and report its accuracy. We perform this analysis on both the task pairs in the test set of PATHEVAL and an additionally synthesized dataset consisting of task pairs with three levels (denoted as $\epsilon_{1}$ , $\epsilon_{2}$ , and $\epsilon_{3}$ ) of increasing differences in their descriptor values. We describe the data generation process for this dataset in more detail in Appendix C.2. + +The results for both experiments are shown in Table 3. Upon evaluating the performance on the test set of PATHEVAL by individual metrics, we notice that GPT-4o particularly struggles with Clearance metrics. These metrics typically require a lower level of perception and is naturally more challenging to discriminate than other metrics. On the other hand, Smoothness, Number of Sharpest Turns, and Max Angle appear to be easier for the model to capture. We also note that GPT-4o perceives the path length much more easily in a 3D environment presentation. Furthermore, GPT-4o's performance increases as we increase the descriptor difference between paths. This, however, is not the case for the Number of Sharp Turns. We conjecture that when we increase the number of sharp turns, we also enforce the challenge of "counting" the number of satisfying turns, which VLMs have been shown with limitations [46]. + +Does segment complexity affect performance? We look into whether the complexity of the path segment is the key reason for GPT-4o's limited perception of clearance. To this end, we test GPT-4o on segments of varying complexity (i.e., points, straight lines, and curves), in a simplified + +![](images/64370fd06c0e1b6c0ffc05b4ad3b3dd2d3fac2b321af24f07d50de566c670467.jpg) + +![](images/797dcb26cc65e4acc5a8548ba8507480512eeccb1d2e575eb31105b9c6389929.jpg) +(a) Point + +![](images/223e1262a5584fe660daa68ba08978ce7562dd5eeade35746fed784978de8428.jpg) + +![](images/c6b5e322480982896d78b549393ea442747ad58630c462ebe76d636c4ff6422c.jpg) +(b) Straight line + +![](images/e3293b2cd816cf29ad2a3b7463952ed1bc871b3cc637a029db4988eafd99a361.jpg) + +![](images/cf60e29ab12647331ed9e4a6f5a0d1019199b600f144e02495db1d762791a9cf.jpg) + +![](images/bbab9b6506d01862c50749d8e46a9e7912d85bfc0a817618f028adc644cdf826.jpg) +(c) Curve +(d) GPT-40 performance on the distance to obstacle under different segment complexities. +Figure 2. Example segment complexity test cases in simplified environments and performance across the various settings. + +environment with only one rectangular obstacle at the center (Figure 2), and evaluate its accuracy in identifying segments that are closer to the obstacle. For individual points, the clearance is defined by the perpendicular distance from the point to the obstacle; for straight lines and curves, we consider a path closer if one of its endpoints is closer to the obstacle. For each segment type, we synthesize 100 pairs by first randomly generating 200 segments and then pairing each with the segment with the greatest distance difference from the obstacle relative to it (i.e., maximizing the absolute difference between the distances of the two segments from the obstacle). This increases the likelihood that the distance is significant enough to be perceivable. The average difference in the clearances of the pairs of segments are 14.76, 14.42, and 14.28 for points, lines, and curves respectively. + +The results in Figure 2d show that GPT-4o can perform better in very easy scenarios; however, it struggles more as the segment complexity increases. For instance, the model was able to identify the closer points in $77\%$ of the cases; however, when considering straight and curved lines, its performance drops to $64\%$ and $59\%$ , respectively. A surprising observation is that, in the case of curved lines, GPT-4o's performance is dramatically better in 3D images. As shown in Appendix C.2, the average Clearance difference of path pairs in PATHEVAL is merely 0.12 - 1.31. As the paths in PATHEVAL are much more complicated than the curves in this experiment, it is expected that GPT-4o exhibits difficulty in judging paths' clearances. The complexity of the environments (compared to a single square obstacle) could add challenges. + +# 5. Fine-tuning a Path Evaluator + +One intuitive question is whether simply fine-tuning the VLMs can relieve their vision bottleneck. To answer this + +
ModelFrozenFine-tuned
AccuracyAvg. Cosine SimilarityAccuracyAvg. Cosine Similarity
clip-vit-base-patch320.5100.9140.7830.514
clip-vit-large-patch14-3360.4980.9070.7490.548
OpenCLIP-vit-B-320.5400.8830.7430.475
siglip-base-patch16-2240.5290.8950.7310.612
dino-vit-160.4950.9110.7630.754
dino-v2-base0.5100.7610.7210.681
+ +Table 4. Probing accuracy and average cosine similarity between distinct path images (2D) when the vision encoder is frozen or fine-tuned. + +
SettingFrozenTunable
Image Only0.520.51
Image w Descriptors0.96-
Image Only (Random IDs)0.480.52
+ +Table 5. Fine-tuned LLaVA-NeXT-7b performance (2D) when we keep the vision encoder frozen or tunable. + +question, we experiment with LLaVA-NeXT-7b and fine-tune it on the training set (13,500 pairs) of PATHEVAL. We focus on the 2D case for the set of experiments described in this section. We consider three separate settings for training: (1) training with only images as input, (2) training with images and textual descriptor values as input, and (3) the same setting as (1) with using random IDs as target labels. Details in experimental setup are included in Appendix C.3. + +# 5.1. Overall Performance + +Fine-tuning does not help with vision-language mapping. Table 5 shows that the model fails to learn meaningful patterns in the data, even after training 50 epochs. However, when the textual descriptors are provided as input, the model can easily learn the function achieving $96\%$ accuracy (a $45\%$ improvement from the zero-shot setting). This shows that the model is unable to extract the same descriptor values from the image input. Unfreezing the encoder for fine-tuning also does not provide any significant improvement. We include a further discussion in Appendix D.3. The results point to a limitation in the vision model's ability to encode the images, which we will investigate next. + +# 5.2. Understanding the Visual Representations + +In order to further understand the limitations of the vision component, we conduct an analysis to better understand how well different vision encoder models can differentiate between different paths in the dataset. To this end, we first apply a linear probe to see how easily distinguishable different images are. Specifically, given a pair of images, we first use the vision encoder to extract high-dimensional feature representations for both images. These features are then concatenated and passed through a simple binary classification layer (i.e., the probe). The probe is trained to predict a label of 1 if the images are the same and 0 otherwise. We experiment with various SOTA vision encoders, namely CLIP + +[55] base and large varieties, LAION-OpenCLIP [25, 56], SigLip [80], DINO [8], and DINO-v2 [49] and analyze how well their learned representations capture visual similarities and differences. We use a set of 1,000 randomly sampled path pairs with a balanced label distribution to train the probe, and look at whether the model can learn to distinguish between these paths. We also perform an experiment where we fine-tune the vision encoder along with the probe. In this setting, our goal is to gauge if carefully fine-tuning a vision encoder can potentially improve the model's performance in low-level perception. Finally, in both settings, we present the average cosine similarity between distinct paths. + +Vision encoders cannot distinguish between paths. From Table 4, it can be seen that vision encoder models are unable to provide representations that are significantly different for the probe to tell if they are the same. This is further supported by the high values for the average cosine similarity across all models. + +Fine-tuning the encoders on a discrimination task can help disentangle the visual representations. By unfreezing the encoder weights and training them to identify whether two paths are identical, we enhance their adaptability to the task. The results in Table 4 demonstrate that this approach effectively disentangles the learned representations, resulting in significantly improved performance and increased separability, as evidenced by the notably lower cosine similarity between non-identical paths. The results thus imply the need for carefully fine-tuning task-specific vision encoders for path evaluation on PATHEVAL. + +# 6. Conclusion + +This work explored the use of VLMs as evaluators in path-planning scenarios. We evaluated a number of VLMs on our proposed benchmark, PATHEVAL, and found that these models struggle with low-level perception. Specifically, we find that visual encoders used by SOTA models are unable to discern the differences between different paths in intricate scenarios. We hope that PATHEVAL will inspire researchers to further explore ways to improve the visual capabilities of VLMs and contribute to finding better ways to incorporate foundation models for developing more flexible, robust, and scalable planning paradigms. + +# Acknowledgments + +This project was supported by resources provided by the Office of Research Computing at George Mason University (URL: https://orc.gmu.edu) and funded in part by grants from the National Science Foundation (Award Number 2018631). + +# References + +[1] Chirag Agarwal, Sree Harsha Tanneru, and Himabindu Lakkaraju. Faithfulness vs. Plausibility: On the (Un)Reliability of Explanations from Large Language Models, 2024. 6 +[2] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can Large Language Models be Good Path Planners? A Benchmark and Investigation on Spatial-temporal Reasoning. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1, 3 +[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Look Further Ahead: Testing the Limits of GPT-4 in Path Planning. IEEE CASE, 2024. 1, 3 +[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In International Conference on Computer Vision (ICCV), 2015. 3 +[5] Kate Baumli, Satinder Baveja, Feryal Behbahani, Harris Chan, Gheorghe Comanici, Sebastian Flennerhag, Maxime Gazeau, Kristian Holsheimer, Dan Horgan, Michael Laskin, Clare Lyle, Hussain Masoom, Kay McKinney, Volodymyr Mnih, Alexander Neitz, Dmitry Nikulin, Fabio Pardo, Jack Parker-Holder, John Quan, Tim Roktaschel, Himanshu Sahni, Tom Schaul, Yannick Schroecker, Stephen Spencer, Richie Steigerwald, Luyu Wang, and Lei Zhang. Vision-Language Models as a Source of Rewards, 2024. 1, 3 +[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1, 3 +[7] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21819–21830, 2024. 3 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 3, 8 + +[9] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee Wong. MapGPT: Map-guided prompting with adaptive path planning for vision-and-language navigation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9796-9810, Bangkok, Thailand, 2024. Association for Computational Linguistics. 3 +[10] Xuweiyi Chen, Ziqiao Ma, Xuejun Zhang, Sihan Xu, Shengyi Qian, Jianing Yang, David Fouhey, and Joyce Chai. Multi-object hallucination in vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4 +[11] Yongchao Chen, Jacob Arkin, Charles Dawson, Yang Zhang, Nicholas Roy, and Chuchu Fan. Autotamp: Autoregressive task and motion planning with llms as translators and checkers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6695-6702, 2024. 1, 3 +[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. arXiv preprint arXiv:2312.14238, 2023. 3, 5 +[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 4 +[14] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 3 +[15] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 3 +[16] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, brian richter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B. Tenenbaum, Leslie Pack Kaelbling, Andy Zeng, and Jonathan Tompson. Video language planning. In The Twelfth International Conference on Learning Representations, 2024. 3 +[17] Abhimanyu Dubey et. al. The Llama 3 Herd of Models, 2024. 3, 5 +[18] Difei Gao, Ruiping Wang, Shiguang Shan, and Xilin Chen. Cric: A vqa dataset for compositional reasoning on vision and commonsense. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5561-5578, 2022. 3 +[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating + +the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[20] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 2022. 3 +[21] Lin Guan, Yifan Zhou, Denis Liu, Yantian Zha, Heni Ben Amor, and Subbarao Kambhampati. Task success is not enough: Investigating the use of video-language models as behavior critics for catching undesirable agent behaviors. In First Conference on Language Modeling, 2024. 1, 3 +[22] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models, 2023. 4 +[23] Xu Han, Qiannan Yang, Xianda Chen, Xiaowen Chu, and Meixin Zhu. Generating and evolving reward functions for highway driving with large language models, 2024. 3 +[24] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3, 8 +[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's "up" with vision-language models? Investigating their struggle with spatial reasoning. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 3 +[27] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Audit Verma, Kaya Stechly, Siddhant Bhambri, Lucas Paul Saldyt, and Anil B Murthy. Position: LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks. In _Forty-first International Conference on Machine Learning_, 2024. 1, 3 +[28] L.E. Kavraki, P. Svestka, J.-C. Latombe, and M.H. Overmars. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and Automation, 12(4):566-580, 1996. 4 +[29] J.J. Kuffner and S.M. LaValle. RRT-connect: An efficient approach to single-query path planning. In Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065), pages 995-1001 vol.2, 2000. 4 +[30] Steven M. LaValle. Rapidly-exploring random trees: a new tool for path planning. The annual research report, 1998. 4 +[31] Jan Leike. Why i'm optimistic about our alignment approach: Evaluation is easier than generation. + +https://aligned.substack.com/i/88447351/evaluation-is-easier-than-generation,2022.Accessed:2024-11.1 +[32] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5 +[33] Hao Li, Xue Yang, Zhaokai Wang, Xizhou Zhu, Jie Zhou, Yu Qiao, Xiaogang Wang, Hongsheng Li, Lewei Lu, and Jifeng Dai. Auto mc-reward: Automated dense reward design with large language models for minecraft. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 4 +[35] Yizhi Li, Ge Zhang, Yinghao Ma, Ruibin Yuan, Kang Zhu, Hangyu Guo, Yiming Liang, Jiaheng Liu, Zekun Wang, Jian Yang, Siwei Wu, Xingwei Qu, Jinjie Shi, Xinyue Zhang, Zhenzhu Yang, Xiangzhou Wang, Zhaoxiang Zhang, Zachary Liu, Emmanouil Benetos, Wenhao Huang, and Chenghua Lin. OmniBench: Towards The Future of Universal Omni-Language Models, 2024. 4 +[36] Hui Sheng Lim, Shuangshuang Fan, Christopher K.H. Chin, Shuhong Chai, Neil Bose, and Eonjoo Kim. Constrained path planning of autonomous underwater vehicle using selectively-hybridized particle swarm optimization algorithms. IFAC-PapersOnLine, 52(21):315-322, 2019. 12th IFAC Conference on Control Applications in Marine Systems, Robotics, and Vehicles CAMS 2019. 2 +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3 +[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2024. 3, 5 +[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 3, 5 +[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 4 +[41] Tomás Lozano-Pérez and Michael A. Wesley. An algorithm for planning collision-free paths among polyhedral obstacles. Communications of the ACM, 22(10):560-570, 1979. 4 +[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning Representations, 2024. 3 + +[43] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[44] Silin Meng, Yiwei Wang, Cheng-Fu Yang, Nanyun Peng, and Kai-Wei Chang. LLM-a*: Large language model enhanced incremental heuristic search on path planning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1087-1102, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3 +[45] Shota Onohara, Atsuyuki Miyai, Yuki Imajuku, Kazuki Egashira, Jeonghun Baek, Xiang Yue, Graham Neubig, and Kiyoharu Aizawa. Jmmm: A japanese massive multidiscipline multimodal understanding benchmark for culture-aware evaluation, 2024. 4 +[46] OpenAI. GPT-4V(ision) System Card. 2023. 7 +[47] OpenAI. GPT-4o System Card, 2024. 3, 5 +[48] OpenAI. GPT-4 Technical Report, 2024. 3 +[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 8 +[50] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 950-974, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 3 +[51] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14760-14769, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 1, 3 +[52] Clment Petres, Yan Pailhas, Pedro Patron, Yvan Petillot, Jonathan Evans, and David Lane. Path planning for autonomous underwater vehicles. IEEE Transactions on Robotics, 23(2):331-341, 2007. 2 +[53] Evis Plaku, Erion Plaku, and Patricio Simari. Clearance-driven motion planning for mobile robots with differential constraints. Robotica, 36(7):971-993, 2018. 4 +[54] Patrick A. Plonski, Pratap Tokekar, and Volkan Isler. Energy-Efficient Path Planning for Solar-Powered Mobile Robots, pages 717-731. Springer International Publishing, Heidelberg, 2013. 2 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pam Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Proceedings of + +the 38th International Conference on Machine Learning, 2021. 3, 8 +[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3, 8 +[57] Jie-Jing Shao, Jiang-Xin Shi, Xiao-Wen Yang, Lan-Zhe Guo, and Yu-Feng Li. Investigating the Limitation of CLIP Models: The Worst-Performing Categories, 2023. 3 +[58] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B. Tenenbaum, Leslie Pack Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models, 2023. 3 +[59] Loan A. Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics & Automation Magazine, 19(4):72-82, 2012. 4 +[60] Yinan Sun, Zicheng Zhang, Haoning Wu, Xiaohong Liu, Weisi Lin, Guangtao Zhai, and Xiongkuo Min. Explore the hallucination on low-level perception for mllms, 2024. 4 +[61] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3 +[62] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 3 +[63] Shengbang Tong, Erik Jones, and Jacob Steinhardt. Mass-producing failures of multimodal systems with language models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 4 +[64] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs, 2024. 1, 3, 4 +[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams + +Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 3 +[66] Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language models don't always say what they think: unfaithful explanations in chain-of-thought prompting. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 6 +[67] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large Language Models Still Can't Plan (A Benchmark for LLMs on Planning and Reasoning about Change). In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 3 +[68] Karthik Valmeekam, Kaya Stechly, and Subbarao Kambhampati. LLMs Still Can't Plan; Can LRMs? A Preliminary Evaluation of OpenAI's o1 on PlanBench, 2024. 3 +[69] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5 +[70] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6622-6631, 2019. 3 +[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, 2022. 1 +[72] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. ECCV, 2024. 3 +[73] Tianbao Xie, Siheng Zhao, Chen Henry Wu, Yitao Liu, Qian Luo, Victor Zhong, Yanchao Yang, and Tao Yu. Text2Reward: Reward Shaping with Language Models for Reinforcement Learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 3 +[74] Yutaro Yamada, Yihan Bao, Andrew Kyle Lampinen, Jungo Kasai, and Ilker Yildirim. Evaluating spatial understanding of large language models. Transactions on Machine Learning Research, 2024. 3 +[75] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1, 3 +[76] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3 + +[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024. 4 +[78] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI. In Proceedings of CVPR, 2024. 4 +[79] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark. arXiv preprint arXiv:2409.02813, 2024. 4 +[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 8 +[81] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, and Jie Fu. CMMU: A Chinese Massive Multi-discipline Multimodal Understanding Benchmark, 2024. 4 +[82] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models, 2024. 1 +[83] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy Improvement using Language Feedback Models, 2024. 1, 3 +[84] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7W: Grounded Question Answering in Images. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3 \ No newline at end of file diff --git a/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/images.zip b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cffdf308ad0017032870a47b58fced482363d5f1 --- /dev/null +++ b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aeda7b0ab4b359066c151ad6673d3a578a83414a9a716af80ac33af6ff925d14 +size 236689 diff --git a/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/layout.json b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f273174a0e4577c450968a64b91f447938b999bf --- /dev/null +++ b/2025/Evaluating Vision-Language Models as Evaluators in Path Planning/layout.json @@ -0,0 +1,8621 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 515, + 121 + ], + "type": "text", + "content": "Evaluating Vision-Language Models as Evaluators in Path Planning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 142, + 516, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 142, + 516, + 185 + ], + "spans": [ + { + "bbox": [ + 93, + 142, + 516, + 185 + ], + "type": "text", + "content": "Mohamed Aghzal1, Xiang Yue2, Erion Plaku3*, Ziyu Yao1 \n1George Mason University, 2Carnegie Mellon University, 3National Science Foundation {maghzal, ziyuyao}@gmu.edu, xyue2@andrew.cmu.edu, eplaku@nsf.gov" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 539 + ], + "type": "text", + "content": "Despite their promise to perform complex reasoning, large language models (LLMs) have been shown to have limited effectiveness in end-to-end planning. This has inspired an intriguing question: if these models cannot plan well, can they still contribute to the planning framework as a helpful plan evaluator? In this work, we generalize this question to consider LLMs augmented with visual understanding, i.e., Vision-Language Models (VLMs). We introduce PATHEVAL, a novel benchmark evaluating VLMs as plan evaluators in complex path-planning scenarios. Succeeding in the benchmark requires a VLM to be able to abstract traits of optimal paths from the scenario description, demonstrate precise low-level perception on each path, and integrate this information to decide the better path. Our analysis of state-of-the-art VLMs reveals that these models face significant challenges on the benchmark. We observe that the VLMs can precisely abstract given scenarios to identify the desired traits and exhibit mixed performance in integrating the provided information. Yet, their vision component presents a critical bottleneck, with models struggling to perceive low-level details about a path. Our experimental results show that this issue cannot be trivially addressed via end-to-end fine-tuning; rather, task-specific discriminative adaptation of these vision encoders is needed for these VLMs to become effective path evaluators.[12]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 584, + 296, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 296, + 645 + ], + "type": "text", + "content": "The impressive capabilities of Large Language Models (LLMs) [6, 8, 71] and Vision-Language Models (VLMs) [82] have led to an increasing interest in applying them to automated motion planning and navigation tasks [50, 51]. However, the inherent limitations of these models in long-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 213, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 346 + ], + "type": "text", + "content": "horizon planning have rendered them ineffective as end-to-end motion planners [2, 3, 11, 75]. This has made researchers wonder: if these models cannot be good motion planners themselves, can they still support a motion planning framework? Intuitively, using these models still holds the promise of significantly enhancing the motion planning framework, as they have learned extensive factual and commonsense knowledge that could benefit planning during their pre-training. As a result, there has been an emerging paradigm exploring how these models can be leveraged in combination with traditional methods [27]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 556, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 566 + ], + "type": "text", + "content": "One particularly interesting approach within this line of work involves using these models as plan evaluators. Motivated by the intuition that \"evaluation is easier than generation\" [31], several efforts have explored leveraging these models as \"critics\" to assess the quality of generated plans [5, 21, 73, 83]. However, most of these efforts have focused on scenarios that require only limited, high-level visual perception, without necessitating fine-grained or precise perceptual abilities. On the other hand, while there have been prior works similarly investigating VLMs' low-level perception [26, 64], studies specifically about the use of these models in planning remain limited. Tasks such as motion planning often require fine-grained visual understanding in highly specific contexts, while also drawing on broad commonsense knowledge acquired during pre-training. Hence, there is a pressing need to investigate the potential of VLMs to understand both low-level visual details and leverage these visual signals for high-level reasoning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "In this work, we explore whether we can utilize VLMs as evaluators in highly intricate continuous path planning problems. We introduce PATHEVAL (Figure 1), a controllable benchmark designed around path planning in complex environments under diverse commonsense decision-making scenarios (e.g., A firefighting robot has to cover as much of the area as possible to extinguish fires and scout for survivors, thus needs to prioritize paths with more coverage and higher clearance). Traditionally, encoding constraints in a planning algorithm requires significant human effort in hand-crafting scenario-specific criteria for path evaluation, such that the constraints can be injected into the planner" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "*The work by E. Plaku is supported by (while serving at) the National Science Foundation. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. \n1Source code: https://github.com/MohamedAghzal/PathEval \n2Dataset: https://huggingface.co/datasets/maghzal/PathEval" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6886" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 71, + 73, + 228, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 73, + 228, + 85 + ], + "spans": [ + { + "bbox": [ + 71, + 73, + 228, + 85 + ], + "type": "text", + "content": "Example of an instance of PATHEVAL" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 95, + 239, + 180 + ], + "blocks": [ + { + "bbox": [ + 71, + 95, + 239, + 180 + ], + "lines": [ + { + "bbox": [ + 71, + 95, + 239, + 180 + ], + "spans": [ + { + "bbox": [ + 71, + 95, + 239, + 180 + ], + "type": "image", + "image_path": "0cd1b2523106083092d842b736f974c7c1231d5c5a89f94d796fae60e61758db.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 188, + 161, + 198 + ], + "lines": [ + { + "bbox": [ + 148, + 188, + 161, + 198 + ], + "spans": [ + { + "bbox": [ + 148, + 188, + 161, + 198 + ], + "type": "text", + "content": "2D" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 262, + 106, + 395, + 171 + ], + "blocks": [ + { + "bbox": [ + 262, + 106, + 395, + 171 + ], + "lines": [ + { + "bbox": [ + 262, + 106, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 262, + 106, + 395, + 171 + ], + "type": "image", + "image_path": "0838274835244e4bdc4acc68be840e06f4ef52695e7f8d98d7b664cd06a6bd3e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 186, + 405, + 196 + ], + "lines": [ + { + "bbox": [ + 392, + 186, + 405, + 196 + ], + "spans": [ + { + "bbox": [ + 392, + 186, + 405, + 196 + ], + "type": "text", + "content": "3D" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 402, + 106, + 534, + 171 + ], + "blocks": [ + { + "bbox": [ + 402, + 106, + 534, + 171 + ], + "lines": [ + { + "bbox": [ + 402, + 106, + 534, + 171 + ], + "spans": [ + { + "bbox": [ + 402, + 106, + 534, + 171 + ], + "type": "image", + "image_path": "e9583cea04a42a866a2085509284da1af944969998b3d5988bc342156a8d6e63.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 587, + 555, + 643 + ], + "lines": [ + { + "bbox": [ + 55, + 587, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 555, + 643 + ], + "type": "text", + "content": "Figure 1. Example of PATHEVAL. The benchmark consists of 14,550 instances of path pairs rendered in both 2D and 3D and mapped to 15 decision-making scenarios. Success on this task is tied to three distinct levels: 1) Attribute abstraction: recognizing what aspects make a path ideal, 2) Low-level perception: extracting the required attributes for each path from the images, and 2) Information integration: synthesizing the collected information to make a decision. We test a set of VLMs on the task and find that they struggle particularly with low-level perception. Incorrect answers by different models are shown (explanations indicating misperception are highlighted)." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 203, + 134, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 203, + 134, + 213 + ], + "spans": [ + { + "bbox": [ + 71, + 203, + 134, + 213 + ], + "type": "text", + "content": "Path Descriptors:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 213, + 538, + 251 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 71, + 213, + 538, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 213, + 538, + 231 + ], + "spans": [ + { + "bbox": [ + 71, + 213, + 538, + 231 + ], + "type": "text", + "content": "Path 1: Min. clearance: 1.07, Max. clearance: 8.91, Avg. clearance: 2.62, Path length: 243.46, Smoothness: 843.07, Number of sharp turns: 4, Maximum angle: 124.35." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 231, + 538, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 231, + 538, + 251 + ], + "spans": [ + { + "bbox": [ + 71, + 231, + 538, + 251 + ], + "type": "text", + "content": "Path 2: Min. clearance: 1.57, Max. clearance: 8.91, Avg. clearance: 3.18, Path length: 243.75, Smoothness: 809.65, Number of sharp turns: 1, Maximum angle: 93.12." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 71, + 259, + 538, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 259, + 538, + 278 + ], + "spans": [ + { + "bbox": [ + 71, + 259, + 538, + 278 + ], + "type": "text", + "content": "Scenario #1: The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "spans": [ + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "type": "text", + "content": "Required Descriptors: Smoothness " + }, + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "type": "text", + "content": ", Number of Sharp turns " + }, + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "type": "text", + "content": ", Maximum angle " + }, + { + "bbox": [ + 71, + 278, + 362, + 289 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 289, + 151, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 289, + 151, + 297 + ], + "spans": [ + { + "bbox": [ + 71, + 289, + 151, + 297 + ], + "type": "text", + "content": "Ground Truth: Path 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 297, + 538, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 297, + 538, + 316 + ], + "spans": [ + { + "bbox": [ + 71, + 297, + 538, + 316 + ], + "type": "text", + "content": "Scenario #2: An autonomous firefighting robot is designed to navigate and operate within burning buildings to extinguish fires and rescue trapped individuals. It should explore as much of the area as possible, while maintaining a safe distance from the walls to avoid damage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "spans": [ + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "type": "text", + "content": "Required Descriptors: Avg. Clearance " + }, + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "type": "text", + "content": ", Min. Clearance " + }, + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "type": "text", + "content": ", Path length " + }, + { + "bbox": [ + 71, + 316, + 333, + 327 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 71, + 327, + 148, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 327, + 148, + 335 + ], + "spans": [ + { + "bbox": [ + 71, + 327, + 148, + 335 + ], + "type": "text", + "content": "Ground Truth: Path 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 71, + 335, + 538, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 335, + 538, + 354 + ], + "spans": [ + { + "bbox": [ + 71, + 335, + 538, + 354 + ], + "type": "text", + "content": "Scenario #3: As the vehicle is traversing a warzone, it must stay concealed from enemy operatives, making use of covers like walls and avoiding open spaces as much as possible. It should also reach its target as quickly as possible." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 71, + 354, + 334, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 354, + 334, + 364 + ], + "spans": [ + { + "bbox": [ + 71, + 354, + 334, + 364 + ], + "type": "text", + "content": "Required Descriptors: Avg. Clearance (↓), Max. Clearance (↓), Path length (↓)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 71, + 364, + 151, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 364, + 151, + 373 + ], + "spans": [ + { + "bbox": [ + 71, + 364, + 151, + 373 + ], + "type": "text", + "content": "Ground Truth: Path 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 71, + 383, + 100, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 383, + 100, + 392 + ], + "spans": [ + { + "bbox": [ + 71, + 383, + 100, + 392 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 71, + 392, + 538, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 392, + 538, + 420 + ], + "spans": [ + { + "bbox": [ + 71, + 392, + 538, + 420 + ], + "type": "text", + "content": "The agent navigating this maze is a large truck, so sharp turns (90 degrees or larger) are harder to make. It should also stay on a straight line (unless it is making a turn) as it is driving through heavy traffic. Which path better achieves the task? Path 1 is on the left side and Path 2 is on the right side. Your answer should follow the format below:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 71, + 421, + 157, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 421, + 157, + 430 + ], + "spans": [ + { + "bbox": [ + 71, + 421, + 157, + 430 + ], + "type": "text", + "content": "Answer: Path 1 or Path 2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 71, + 430, + 223, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 430, + 223, + 440 + ], + "spans": [ + { + "bbox": [ + 71, + 430, + 223, + 440 + ], + "type": "text", + "content": "Explanation: Why you chose the path (1 or 2)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 71, + 449, + 101, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 449, + 101, + 457 + ], + "spans": [ + { + "bbox": [ + 71, + 449, + 101, + 457 + ], + "type": "text", + "content": "GPT-40" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 71, + 459, + 128, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 459, + 128, + 467 + ], + "spans": [ + { + "bbox": [ + 71, + 459, + 128, + 467 + ], + "type": "text", + "content": "Answer: Path 1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 70, + 468, + 538, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 468, + 538, + 498 + ], + "spans": [ + { + "bbox": [ + 70, + 468, + 538, + 498 + ], + "type": "text", + "content": "Explanation: Path 1 has fewer sharp turns and maintains longer straight segments compared to Path 2. This is more suitable for a large truck navigating through heavy traffic, as it minimizes the need for difficult maneuvers and allows for smoother driving. Path 2, on the other hand, has more frequent and sharper turns, which would be more challenging for a large truck to navigate." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 71, + 506, + 133, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 506, + 133, + 514 + ], + "spans": [ + { + "bbox": [ + 71, + 506, + 133, + 514 + ], + "type": "text", + "content": "LLaVA-NeXT-7b" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 71, + 515, + 129, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 515, + 129, + 524 + ], + "spans": [ + { + "bbox": [ + 71, + 515, + 129, + 524 + ], + "type": "text", + "content": "Answer: Path 1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 70, + 525, + 538, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 525, + 538, + 563 + ], + "spans": [ + { + "bbox": [ + 70, + 525, + 538, + 563 + ], + "type": "text", + "content": "Explanation: Path 1 is better suited for a large truck because it has fewer sharp turns (90 degrees or larger) compared to Path 2. The truck would have to make only one 90-degree turn, which is easier to execute than the two 90-degree turns in Path 2. Additionally, Path 1 has more straight sections, which would be more efficient for a large truck to navigate through heavy traffic. The fewer turns and the greater number of straight sections make Path 1 the more suitable route for a large truck." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 654, + 297, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 297, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 297, + 702 + ], + "type": "text", + "content": "during the search process [36, 52, 54]. However, this not only is a tedious and costly process but also cannot scale up to handle the countless intricate scenarios in meeting real-life planning needs. An effective VLM as a \"reward model\"" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 654, + 555, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 702 + ], + "type": "text", + "content": "relieving humans from customizing evaluation criteria for specific scenarios could enable scalable, general-purpose planners that can adapt to diverse commonsense scenarios described in natural language." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6887" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "In PATHEVAL, a VLM is tasked with comparing two paths within a given decision-making context and selecting the one that better satisfies the constraints outlined by the scenario. Success on this task requires effective performance across three distinct levels: 1) attribute abstraction: recognizing the attributes that define a favorable path in a particular decision-making context; 2) low-level perception: demonstrating precise low-level perception to determine which path performs better based on the given criteria; and 3) information integration: integrating and synthesizing the perceived information to produce an answer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "text", + "content": "Using this benchmark, we analyze the performance of 9 state-of-the-art (SOTA) VLMs, including both the closed-source GPT-4o and GPT-4o-mini [47] and 7 different open-source VLMs of different sizes (i.e., LLaVA-NeXT 7b and 13b [39], Qwen2-VL-7b [69], LLaVA-OneVision-7b [32], LLaMA-3.2-11b [17], and Intern-VL2 8b and 40b [12]). We find that these models struggle with the path evaluation task (e.g., Qwen2-VL-7b achieving only " + }, + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "inline_equation", + "content": "50.2\\%" + }, + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "text", + "content": " accuracy). However, when providing these VLMs with verbalized path specifications, their performance significantly improves (e.g., " + }, + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "inline_equation", + "content": "74.2\\%" + }, + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "text", + "content": " accuracy for Qwen2-VL-7b), which reveals a potential vision bottleneck of these VLMs. Our further analysis confirms these models' weakness in low-level perception, especially when they are tasked to perceive the clearance of a path with respect to surrounding obstacles, and this weakness could be more prominent when the environment and the path representation become more complex. We discover the source of this weakness from the vision encoders used by these VLMs, yet simply fine-tuning the VLMs end-to-end with the vision encoders does not address the issue. Rather, our experiments suggest performing task-specific discriminative adaptation of these vision encoders." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 482, + 142, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 142, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 142, + 495 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 502, + 194, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 194, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 194, + 514 + ], + "type": "text", + "content": "2.1. Vision-Language Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "The outstanding success of decoder-only LLMs [6, 48, 65] has driven the development of Vision-Language Models (VLMs), which extend LLMs with a vision component in an attempt to generalize their performance into a multi-modal setting [12, 17, 32, 37, 38, 62, 69]. VLMs are designed for tasks that require unifying visual and linguistic representations (e.g. visual question answering [4]). Typically, this is achieved by connecting a vision encoder to a language decoder and projecting the representations into a shared space, allowing visual elements and linguistic components to be linked in a semantically meaningful manner. One of the earliest successful models at achieving this was the CLIP encoder [55], which was trained using contrastive learning to learn representations that map images and their corresponding textual descriptions. Several varieties of CLIP have then been introduced [25, 56, 76, 80]. These mod" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "els, while showing tremendous promises have shown several limitations when it comes to visual tasks [57]; as such several works [64, 72] have sought ways to improve such representations by combining them with vision foundation models such as DINO [8, 49]. In this work, we contribute to the research of VLMs with a new benchmark, namely PATHEVAL, focusing on evaluating VLMs as evaluators for path planning. This benchmark tests VLMs seamlessly on their commonsense understanding (i.e., being able to abstract critical concepts expressed in the described planning scenarios), low-level perception (i.e., precisely perceiving details about paths in complex environments), and the ability to reason about the collected textual and visual information for decision-making." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 248, + 489, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 489, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 489, + 261 + ], + "type": "text", + "content": "2.2. Automated Planning with VLMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 267, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 602 + ], + "type": "text", + "content": "Vision-language planning promises more flexible planning frameworks and enhanced human-AI interaction. Therefore, designing systems that can effectively understand natural language instruction and leverage perceptual inputs to conduct planning tasks has been a topic of interest in recent years [20, 61, 70]. The rise of VLMs has led many to investigate the use of these models as vision-language planning agents [15, 16, 50, 51]. However, existing literature highlights the limitations of LLMs in spatial reasoning [2, 26, 74] as well as long-horizon planning [3, 67, 68]. This pushed researchers to explore alternative ways to incorporate VLMs and LLMs into planning frameworks more reliably [27, 58]. One potential direction is their use as plan evaluators, either through the generation of reward functions [23, 33, 73], or by using them directly as off-the-shelf critics [5, 21, 83]. The success of such frameworks assumes perfect perception and that the models can accurately perceive visual information and reason about it in order to produce an answer; nevertheless, it has been shown that the representations used by these models fail in highly intricate visual settings [64]. Several works have explored the use of VLMs as well as LLMs for path and motion planning [2, 3, 7, 9, 11, 14, 44, 75], however, to the knowledge of the authors there is no work that explores the use of VLMs as path critics in this context. Accordingly, we aim to evaluate the ability of VLMs to serve as evaluators in cases requiring navigation in complex environments while adhering to decision-making constraints specified in natural language." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 611, + 511, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 511, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 511, + 623 + ], + "type": "text", + "content": "2.3. Vision-Language Model Benchmarks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "type": "text", + "content": "The introduction of multimodal models has prompted the development of several benchmarks that are capable of assessing the performance of these models on visual reasoning tasks such as visual question-answering datasets [4, 18, 19, 24, 42, 43, 84]. However, the rise of foundation models has produced the need for a more holistic evaluation of the perceptual and reasoning capabilities of large" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6888" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "VLMs, leading to benchmarks such as MM-Vet [77], MM-Bench [40], MMMU [45, 78, 79, 81] and OmniBench [35]. Several benchmarks have also specifically been designed to assess the perception capabilities of these models and explore the limitations associated with visual hallucinations and optical illusions [10, 22, 34, 60, 64]. Our proposed benchmark provides a flexible yet challenging framework for interleavedly assessing the low-level perception and reasoning capabilities of VLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 191, + 214, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 214, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 214, + 205 + ], + "type": "text", + "content": "3. The PATHEVAL Benchmark" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 331 + ], + "type": "text", + "content": "Motivated by the need to evaluate VLMs as path evaluators in real-world planning scenarios, we introduce PATHEVAL, a controllable and extensible benchmark focused on path planning [28-30, 41] in complex environments under a diverse set of decision-making constraints. We list the decision-making scenarios as well as the descriptors they attempt to optimize in Appendix A. In total, PATHEVAL includes 14,550 tasks over more than 1,150 distinct environments and 15 distinct scenarios. Below, we introduce a formal description of the task and the dataset construction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 162, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 162, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 162, + 352 + ], + "type": "text", + "content": "3.1. Task Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": "Given two paths, " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": ", and a scenario " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": ", the objective is to determine which path better satisfies the scenario's optimization criteria. Each scenario " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": " is a high-level description that aims to optimize over a set of path descriptors (or metrics) " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{m_1, m_2, \\dots, m_k\\}" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": ", where each descriptor " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "m_j : P \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": " evaluates a specific property of a path (e.g., length, smoothness, or proximity to obstacles). A VLM " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": " is presented with two images presenting " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 55, + 358, + 295, + 525 + ], + "type": "text", + "content": " in the same environment, respectively. The model must then decide which path better satisfies the scenario's criteria. To explore the sensitivity of VLMs to the way how a path is presented, PATHEVAL includes both the 2D and 3D images of the path illustration. The model is also prompted to generate an explanation to justify its choice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 241, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 241, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 241, + 547 + ], + "type": "text", + "content": "3.2. Environment and Path Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "text", + "content": "Environment Generation: An environment, as shown in Figure 1, is defined by a set of walls " + }, + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{O_1,O_2,\\dots ,O_n\\}" + }, + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "text", + "content": ", where each wall " + }, + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "text", + "content": " represents an obstacle in the 2D space. Each wall is a closed geometric shape described by its vertices, and the set " + }, + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 553, + 295, + 708 + ], + "type": "text", + "content": " forms the obstacles that the path must avoid. In this work, we leverage the environments of Plaku et. al. [53], which consists of four types of obstacle arrangements: (1) rings, where the environments are structured as mazes with circular walls, (2) waves, which consist of wavy horizontal obstacles, (3) mazes, which consist of both vertical and horizontal walls forming a complex maze structure, as well as (4) random, which consist of randomly placed obstacles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "Path Synthesis via the Randomly-exploring Rapid Tree (RRT) algorithm: To generate path candidates in PATHVAL, we leverage the RRT path planning algorithm [30]. Starting from the initial location in the environment, the algorithm works by building a tree which expands itself by randomly selecting the next location in the environment while avoiding obstacles, until it reaches the goal. In this work, we use the Open Motion Planning Library (OMPL) [59] and implement the RRT-Connect algorithm [29]. We note that while we use RRT in the current benchmark, our codebase is adaptable and can incorporate most path-planning algorithms provided by OMPL. We encourage future research building on our benchmark to experiment with other algorithms as well." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 253, + 416, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 416, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 416, + 266 + ], + "type": "text", + "content": "3.3. Path Descriptors" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 273, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 513 + ], + "type": "text", + "content": "We collect the following descriptors " + }, + { + "bbox": [ + 313, + 273, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 273, + 555, + 513 + ], + "type": "text", + "content": " for each of a generated path: Minimum Clearance measures the smallest distance between any point on the path and the nearest obstacle; Maximum Clearance measures the largest distance between any point on the path and the nearest obstacle; Average Clearance computes the average distance between all points on the path and the nearest obstacle; Path Length is calculated by summing up the Euclidean distances between consecutive points on the path; Smoothness is defined as the sum of the angles between consecutive segments of the path, measuring how smoothly the path changes direction; Number of Sharp Turns counts the number of turns in the path where the angle between consecutive segments exceeds 90 degrees; and Maximum Angle denotes the largest angle between any two consecutive segments of the path. The three Clearance metrics and Path Length share the same measuring unit, i.e., one grid size; Smoothness and Maximum Angle are measured by degree; and the Number of Sharp Turns is an integer count. We include the formula of each descriptor in Appendix B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 525, + 542, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 542, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 542, + 539 + ], + "type": "text", + "content": "3.4. Natural Language Descriptions of Scenarios" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "To create a sufficiently challenging path-planning evaluation benchmark, we design a total of 15 decision-making scenarios that aim to optimize different combinations of the path descriptors. For instance, Scenario #2 (as shown in Figure 1) requires searching through an area affected with fire in search for survivors, and the agent thus must cover as much ground as possible. In contrast, Scenario #3 indicates that the path is to be executed within a warzone; as a result, the vehicle has to remain hidden and take the shortest route. As such, given the same set of paths, the one minimizing the path length is favored by Scenario #3 while Scenario #2 needs to maximize this value. A complete summary of the 15 scenarios, along with the descriptors each scenario aims to optimize, is presented in Tables 6-7 in Appendix A." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6889" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 172, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 172, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 172, + 83 + ], + "type": "text", + "content": "3.5. Task Pairs Selection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "text", + "content": "For each environment we synthesized, we ran the RRT planner 30 times to generate different paths. Upon eliminating paths that did not reach the goal, we selected path pairs that exhibited the greatest dissimilarity in terms of path descriptors. Specifically, we first represented each path using a (7-dimension) vector of its path descriptor values. Given that each path descriptor ranged in a dramatically different scale, we normalized the vector by performing the min-max scaling, i.e., scaling each value " + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "text", + "content": " in the vector to " + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "inline_equation", + "content": "x' = \\frac{x - x_{\\min}}{x_{\\max} - x_{\\min}}" + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "text", + "content": " represents each value in the vector, and " + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "inline_equation", + "content": "x_{\\min}" + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "inline_equation", + "content": "x_{\\max}" + }, + { + "bbox": [ + 55, + 91, + 296, + 305 + ], + "type": "text", + "content": " are the minimum and maximum values of each descriptor across paths sampled from the same environment, respectively. We then measured the distance between two paths by calculating the Euclidean distance between their normalized path descriptor vectors, and selected 5 path pairs with the largest distances. Each path is included in only one pair to avoid redundancy. We repeated the same selection procedure for a total of 1,150 environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 307, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 438 + ], + "type": "text", + "content": "Upon performing this generation, we matched each pair with all fifteen scenarios; however, we only kept pairs where there was a significant difference in at least one of the descriptors required for the particular scenario. In other words, we ensure that the absolute difference is greater than a predefined threshold (0.8 for clearance descriptors, 50 for path length, 90 for smoothness, 1 for the number of sharp turns, and 30 for maximum angle) for at least one of the required descriptors. This makes it more likely that the difference is more noticeable to the naked eye, and thus the paths can be compared by visual inspection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 440, + 296, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 500 + ], + "type": "text", + "content": "We constructed our final evaluation benchmark by randomly selecting 70 task pairs from each scenario, resulting in 1,050 pairs in total. The remaining task pairs (a total of 13,500) are used as the training set to facilitate the fine-tuning experiments in Section 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 513, + 293, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 293, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 293, + 526 + ], + "type": "text", + "content": "4. Can SOTA VLMs Evaluate Planned Paths?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "In order for VLMs to perform successfully in our benchmark, they need to perform successfully at three different levels, i.e., recognizing the critical descriptors required by each scenario (Attribute Abstraction), exhibiting sharp low-level perception to precisely assess each path's properties (Low-level Perception), and integrating the prior information to make a rational decision on the better path (Information Integration). Among them, the first two levels reflect parallel properties that serve as a foundation for the third level. In this section, we evaluated a set of 9 VLMs on PATHEVAL in a zero-shot manner and analyzed their capabilities at these three levels. These VLMs include (1) two closed-source VLMs, i.e., GPT-4o and GPT-4o-mini [47], and (2) seven open-source VLMs with various sizes, including LLaVA-NeXT-7b and LLaVA-NeXT-13b [38, 39]," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "LLaVA-OneVision-7b [32], Qwen2-VL-7b [69], LLaMA3.2-11b [17], and Intern-VL2-8b and Intern-VL2-40b [12]. We include all prompt scripts used in this Section in Appendix C.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 128, + 491, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 491, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 491, + 140 + ], + "type": "text", + "content": "4.1. Overview of VLMs on PATHEVAL" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 146, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 300 + ], + "type": "text", + "content": "The performance of the 9 VLMs on PATHEVAL is showcased in Table 1. We notice that all of the models, except GPT-4o, fail to perform significantly better than a simple random baseline, indicating significant limitations. For GPT-4o, we also notice a " + }, + { + "bbox": [ + 313, + 146, + 555, + 300 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 146, + 555, + 300 + ], + "type": "text", + "content": " higher accuracy on 2D images compared to prompting with 3D inputs. This observation indicates that the model is prone to visual illusions introduced by the 3D images, when it has to rely on solely the image for decision-making (although we observe an opposite effect of 2D vs. 3D when verbalized descriptor values are provided). In the remaining section, we will further break down these models' capabilities to gain a deeper understanding of their failure on this task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 556, + 628 + ], + "type": "text", + "content": "Providing verbalized path information yields better task accuracy, implying reasonable VLM performance in attribute abstraction. In Table 1, we further show the performance of each model when we explicitly list the value for each descriptor as part of the language prompt (i.e., \"PATHEVAL w Desc.). We notice a " + }, + { + "bbox": [ + 313, + 304, + 556, + 628 + ], + "type": "inline_equation", + "content": "11.1\\% -27.1\\%" + }, + { + "bbox": [ + 313, + 304, + 556, + 628 + ], + "type": "text", + "content": " improvement across most models, indicating that when given low-level details, the models can better filter out the information and make better comparisons. This points out that the bottleneck for these VLMs's better performance lies in their inability to accurately perceive low-level information about the paths (we discuss this in more detail in Section 4.2), whereas these models generally have a reasonable capability in abstracting the critical attributes for decision making in various scenarios. In Appendix D.1, we include an experiment where we explicitly query each VLM to identify the critical path metrics for each scenario; the result corroborates our hypothesis. In particular, we find that for most models, the success rate for identifying a required descriptor is over " + }, + { + "bbox": [ + 313, + 304, + 556, + 628 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 313, + 304, + 556, + 628 + ], + "type": "text", + "content": ". Finally, a surprising observation happens to LLaVa-NeXT-7b. We notice that this model suffers particularly severely from hallucination; even when the textual descriptor values are provided and when the model can correctly identify critical path metrics based on our analysis in Appendix D.1, it fails to pick the better paths. We show an example of its explanation when textual descriptors are provided in Figure 5 of Appendix E." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "VLMs exhibit mixed performance in integrating visual and textual path information. We take a further look into the model performance when both the image and the textual descriptor values are provided, and contrast it with their performance when only the textual path descriptions are provided (i.e., \"Desc Only\"). Interest-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6890" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 504, + 213 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 504, + 213 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 504, + 213 + ], + "type": "table", + "html": "
Model2D3DDesc Only
PATHEVALPATHEVAL w Desc.PATHEVALPATHEVAL w Desc.
GPT-4o-mini0.5200.7500.5080.7450.680
GPT-4o0.6650.8600.6240.8950.894
LLaVa-NeXT-7b0.5010.5240.4990.5170.514
Qwen2-VL-7b0.5020.7310.5110.7420.737
LLaVA-OneVision-7b0.5050.7180.5090.7390.721
Intern-VL2-8b0.4890.6540.5050.6910.648
LLaMa-3.2-11b0.4800.6950.4600.6800.686
LLaVa-NeXT-13b0.5090.6200.4940.6010.630
Intern-VL2-40b0.5060.6880.4960.7170.679
Random Baseline0.5000.5000.5000.5000.500
", + "image_path": "984d8d013dbf83d3fdae80fbd9b0f329d1f5b0c131822a5cbd10647454017ec0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 267, + 294, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 294, + 458 + ], + "type": "text", + "content": "ingly, we observe mixed information integration performance from these models. For GPT-4o, Qwen2-VL-7b, LLaVA-OneVision-7b, LLaMa-3.2-11b, and LLaVa-NeXT-13b, their performance based on only descriptor values has no obvious difference to their best performance when (2D or 3D) images are also provided. This observation implies that these models do not benefit from the additional image information when the textual path descriptions are provided. Instead, sometimes the images (e.g., 2D images for GPT-4o and 3D images for LLaVa-NeXT-13b) may confuse their understanding of the textual descriptors, resulting in a worse accuracy compared to Desc Only. For GPT-4o-mini, InternVL2-8b, and Intern-VL2-40b, however, providing both the visual and textual path information offers " + }, + { + "bbox": [ + 54, + 267, + 294, + 458 + ], + "type": "inline_equation", + "content": "\\sim 4\\% - 7\\%" + }, + { + "bbox": [ + 54, + 267, + 294, + 458 + ], + "type": "text", + "content": " performance gain over Desc Only, indicating better information integration from these VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 459, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 459, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 54, + 459, + 295, + 615 + ], + "type": "text", + "content": "In Appendix D.2, we present a breakdown of GPT-4o's performance by scenarios. We show that performance varies greatly from one scenario to the other. Interestingly, we observe that GPT-4o may overuse their commonsense knowledge. For instance, in the case of Scenario #2 shown in Figure 1, where the agent is required to maximize the path length for better coverage, GPT-4o still favors the shorter path. This scenario represents a counterfactual situation as models are often trained or instructed to seek the shortest paths. Evaluating VLMs in such counterfactual scenarios allows us to effectively probe their task understanding and reasoning, which we consider to be an important direction for future research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "The reasoning of VLMs can be unreliable. Limited by the vision bottleneck, we noticed these VLMs fabricating seemingly plausible explanations to justify their path evaluation, despite the fact that they could not actually perceive the necessary path details to perform the task. This fabrication echos findings from other recent work, where LLMs were shown to produce reasoning traces that do not" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 324, + 265, + 545, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "type": "text", + "content": "Table 1. Accuracy of VLMs on PATHEVAL based on 2D and 3D environment images. To investigate the potential vision bottleneck of VLMs, we additionally present each model's accuracy when explicit descriptor values are provided in the language prompts (\"PATHEVAL w Desc\"). The last column (\"Desc Only\") shows the model performance when only the textual descriptor values are provided." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 265, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 324, + 265, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 324, + 265, + 545, + 392 + ], + "type": "table", + "html": "
ModelDefaultFlippedRandom IDs
GPT-4o-mini34/101622/102694/956
GPT-4o278/772258/792291/759
LLaVa-NeXT-7b1028/221041/9580/470
Qwen2-VL-7b21/1029130/920127/923
LLaVA-OneVision-7b438/612440/610262/788
Intern-VL2-8b433/610430/620418/632
LLaMA-3.2-11b*601/251620/264731/319
LLaVa-NeXT-13b795/255939/111305/745
Intern-VL2-40b394/656410/639510/540
Ground truth labels530/520
", + "image_path": "c3bd26b1765a79830567d5a329b19bc75c7ffb4fb22cef58b3cbc097505e5337.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 399, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 449 + ], + "type": "text", + "content": "Table 2. Performance on the 2D case (#of times first path is chosen / #of times second path is chosen) when we flip the path order or replace their default names with random IDs. *There are several cases where LLaMA-3.2-11b does not follow the required format and/or does not give an answer, we omit those cases from this table." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "accurately reflect the process of how the models reach an answer [1, 66]. To gain further insights, we performed an analysis comparing model performance on PATHEVAL with cases that consist of: 1) flipping the order of paths in the pair, and 2) assigning random IDs to the paths (e.g. instead of referring to them as \"Path 1\" and \"Path 2\", we use a random sequence such as \"Path Xu2q\" and \"Path fP48\"). The results presented in Table 2 showcase that VLMs demonstrate bias for a particular label, when they actually do not have the capability to resolve the task. For example, when no matter the default or the flipping labels are used, LLaVA-NeXT-7b consistently selects Path " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "198\\%" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " of the time and fabricates incorrect observations of the two paths in its explanations (Figure 1) to support this choice. As we discussed earlier, LLaVA-NeXT-7b is particularly prone to hallucination in explanations, leading to its random-guess performance with or without textual descriptors. Introducing random IDs as path names mitigates this bias for LLaVa-NeXT-7b (although the model still obtains a close-to-random accuracy on PATHEVAL) but does not seem to help other models dramatically." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6891" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 288, + 155 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 288, + 155 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 288, + 155 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 288, + 155 + ], + "type": "table", + "html": "
DescriptorTest Setε1ε2ε3
Min. Clearance0.46/0.460.50/0.460.74/0.700.86/0.74
Max. Clearance0.44/0.460.41/0.490.46/0.550.50/0.60
Avg. Clearance0.53/0.550.50/0.520.70/0.570.73/0.60
Path Length0.58/0.700.86/0.910.92/0.860.94/0.94
Smoothness0.74/0.720.86/0.820.90/0.900.90/0.89
# of Sharp Turns0.76/0.730.80/0.840.77/0.700.60/0.55
Max. Angle0.71/0.700.82/0.840.86/0.880.94/0.96
", + "image_path": "a41bd53061424851923f31e618bcfd17edd1e2fb0a9db002a2d5c8a9fa76cbb7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "type": "text", + "content": "Table 3. GPT-4o's fine-grained perception accuracy (2D/3D) on the test set of PATHEVAL and three additionally synthesized datasets with increasing metric differences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 203, + 270, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 270, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 270, + 217 + ], + "type": "text", + "content": "4.2. What Can Vision-Language Models See?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 305 + ], + "type": "text", + "content": "The previous subsection clearly highlights the vision component as the bottleneck for path evaluation on PATHEVAL. In this subsection, we conduct an analysis of the visual perception capabilities of VLMs. We focus our analysis on GPT-4o since it is the only model that performs substantially better than the random baseline in the case where no descriptors are required." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "text", + "content": "GPT-4o can perceive paths that are significantly different. In order to get a better understanding of the model's perception capability, we break down its performance in terms of perception by individual metric. Specifically, we prompt GPT-4o to select which path in the pair provides a lower value on each individual metric and report its accuracy. We perform this analysis on both the task pairs in the test set of PATHEVAL and an additionally synthesized dataset consisting of task pairs with three levels (denoted as " + }, + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon_{1}" + }, + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon_{2}" + }, + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon_{3}" + }, + { + "bbox": [ + 55, + 307, + 296, + 450 + ], + "type": "text", + "content": ") of increasing differences in their descriptor values. We describe the data generation process for this dataset in more detail in Appendix C.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 452, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 643 + ], + "type": "text", + "content": "The results for both experiments are shown in Table 3. Upon evaluating the performance on the test set of PATHEVAL by individual metrics, we notice that GPT-4o particularly struggles with Clearance metrics. These metrics typically require a lower level of perception and is naturally more challenging to discriminate than other metrics. On the other hand, Smoothness, Number of Sharpest Turns, and Max Angle appear to be easier for the model to capture. We also note that GPT-4o perceives the path length much more easily in a 3D environment presentation. Furthermore, GPT-4o's performance increases as we increase the descriptor difference between paths. This, however, is not the case for the Number of Sharp Turns. We conjecture that when we increase the number of sharp turns, we also enforce the challenge of \"counting\" the number of satisfying turns, which VLMs have been shown with limitations [46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Does segment complexity affect performance? We look into whether the complexity of the path segment is the key reason for GPT-4o's limited perception of clearance. To this end, we test GPT-4o on segments of varying complexity (i.e., points, straight lines, and curves), in a simplified" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 377, + 131 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 377, + 131 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 377, + 131 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 377, + 131 + ], + "type": "image", + "image_path": "64370fd06c0e1b6c0ffc05b4ad3b3dd2d3fac2b321af24f07d50de566c670467.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 145, + 375, + 175 + ], + "blocks": [ + { + "bbox": [ + 317, + 145, + 375, + 175 + ], + "lines": [ + { + "bbox": [ + 317, + 145, + 375, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 145, + 375, + 175 + ], + "type": "image", + "image_path": "797dcb26cc65e4acc5a8548ba8507480512eeccb1d2e575eb31105b9c6389929.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 178, + 361, + 188 + ], + "lines": [ + { + "bbox": [ + 331, + 178, + 361, + 188 + ], + "spans": [ + { + "bbox": [ + 331, + 178, + 361, + 188 + ], + "type": "text", + "content": "(a) Point" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 404, + 70, + 465, + 131 + ], + "blocks": [ + { + "bbox": [ + 404, + 70, + 465, + 131 + ], + "lines": [ + { + "bbox": [ + 404, + 70, + 465, + 131 + ], + "spans": [ + { + "bbox": [ + 404, + 70, + 465, + 131 + ], + "type": "image", + "image_path": "223e1262a5584fe660daa68ba08978ce7562dd5eeade35746fed784978de8428.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 406, + 146, + 463, + 175 + ], + "blocks": [ + { + "bbox": [ + 406, + 146, + 463, + 175 + ], + "lines": [ + { + "bbox": [ + 406, + 146, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 406, + 146, + 463, + 175 + ], + "type": "image", + "image_path": "c6b5e322480982896d78b549393ea442747ad58630c462ebe76d636c4ff6422c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 178, + 460, + 189 + ], + "lines": [ + { + "bbox": [ + 408, + 178, + 460, + 189 + ], + "spans": [ + { + "bbox": [ + 408, + 178, + 460, + 189 + ], + "type": "text", + "content": "(b) Straight line" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 493, + 70, + 553, + 131 + ], + "blocks": [ + { + "bbox": [ + 493, + 70, + 553, + 131 + ], + "lines": [ + { + "bbox": [ + 493, + 70, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 493, + 70, + 553, + 131 + ], + "type": "image", + "image_path": "e3293b2cd816cf29ad2a3b7463952ed1bc871b3cc637a029db4988eafd99a361.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 495, + 146, + 552, + 175 + ], + "blocks": [ + { + "bbox": [ + 495, + 146, + 552, + 175 + ], + "lines": [ + { + "bbox": [ + 495, + 146, + 552, + 175 + ], + "spans": [ + { + "bbox": [ + 495, + 146, + 552, + 175 + ], + "type": "image", + "image_path": "cf60e29ab12647331ed9e4a6f5a0d1019199b600f144e02495db1d762791a9cf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 331, + 193, + 538, + 243 + ], + "blocks": [ + { + "bbox": [ + 507, + 178, + 539, + 188 + ], + "lines": [ + { + "bbox": [ + 507, + 178, + 539, + 188 + ], + "spans": [ + { + "bbox": [ + 507, + 178, + 539, + 188 + ], + "type": "text", + "content": "(c) Curve" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 193, + 538, + 243 + ], + "lines": [ + { + "bbox": [ + 331, + 193, + 538, + 243 + ], + "spans": [ + { + "bbox": [ + 331, + 193, + 538, + 243 + ], + "type": "image", + "image_path": "bbab9b6506d01862c50749d8e46a9e7912d85bfc0a817618f028adc644cdf826.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 249, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 314, + 249, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 553, + 269 + ], + "type": "text", + "content": "(d) GPT-40 performance on the distance to obstacle under different segment complexities." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "type": "text", + "content": "Figure 2. Example segment complexity test cases in simplified environments and performance across the various settings." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 300, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 479 + ], + "type": "text", + "content": "environment with only one rectangular obstacle at the center (Figure 2), and evaluate its accuracy in identifying segments that are closer to the obstacle. For individual points, the clearance is defined by the perpendicular distance from the point to the obstacle; for straight lines and curves, we consider a path closer if one of its endpoints is closer to the obstacle. For each segment type, we synthesize 100 pairs by first randomly generating 200 segments and then pairing each with the segment with the greatest distance difference from the obstacle relative to it (i.e., maximizing the absolute difference between the distances of the two segments from the obstacle). This increases the likelihood that the distance is significant enough to be perceivable. The average difference in the clearances of the pairs of segments are 14.76, 14.42, and 14.28 for points, lines, and curves respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "text", + "content": "The results in Figure 2d show that GPT-4o can perform better in very easy scenarios; however, it struggles more as the segment complexity increases. For instance, the model was able to identify the closer points in " + }, + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "text", + "content": " of the cases; however, when considering straight and curved lines, its performance drops to " + }, + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 659 + ], + "type": "text", + "content": ", respectively. A surprising observation is that, in the case of curved lines, GPT-4o's performance is dramatically better in 3D images. As shown in Appendix C.2, the average Clearance difference of path pairs in PATHEVAL is merely 0.12 - 1.31. As the paths in PATHEVAL are much more complicated than the curves in this experiment, it is expected that GPT-4o exhibits difficulty in judging paths' clearances. The complexity of the environments (compared to a single square obstacle) could add challenges." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 669, + 479, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 479, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 479, + 682 + ], + "type": "text", + "content": "5. Fine-tuning a Path Evaluator" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "One intuitive question is whether simply fine-tuning the VLMs can relieve their vision bottleneck. To answer this" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6892" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 127, + 70, + 484, + 161 + ], + "blocks": [ + { + "bbox": [ + 127, + 70, + 484, + 161 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 484, + 161 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 484, + 161 + ], + "type": "table", + "html": "
ModelFrozenFine-tuned
AccuracyAvg. Cosine SimilarityAccuracyAvg. Cosine Similarity
clip-vit-base-patch320.5100.9140.7830.514
clip-vit-large-patch14-3360.4980.9070.7490.548
OpenCLIP-vit-B-320.5400.8830.7430.475
siglip-base-patch16-2240.5290.8950.7310.612
dino-vit-160.4950.9110.7630.754
dino-v2-base0.5100.7610.7210.681
", + "image_path": "4be98c2f0b85047d8c927d2f81bdb50e3c3e89c606694d9a908cca4e5d351a26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 88, + 186, + 264, + 237 + ], + "blocks": [ + { + "bbox": [ + 55, + 170, + 553, + 181 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 553, + 181 + ], + "type": "text", + "content": "Table 4. Probing accuracy and average cosine similarity between distinct path images (2D) when the vision encoder is frozen or fine-tuned." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 186, + 264, + 237 + ], + "lines": [ + { + "bbox": [ + 88, + 186, + 264, + 237 + ], + "spans": [ + { + "bbox": [ + 88, + 186, + 264, + 237 + ], + "type": "table", + "html": "
SettingFrozenTunable
Image Only0.520.51
Image w Descriptors0.96-
Image Only (Random IDs)0.480.52
", + "image_path": "faec76c6d170612426fffd7fc2d278b8264e9dc1c7e7b3ec652ad756a743685d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 295, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 295, + 268 + ], + "type": "text", + "content": "Table 5. Fine-tuned LLaVA-NeXT-7b performance (2D) when we keep the vision encoder frozen or tunable." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 275, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 371 + ], + "type": "text", + "content": "question, we experiment with LLaVA-NeXT-7b and fine-tune it on the training set (13,500 pairs) of PATHEVAL. We focus on the 2D case for the set of experiments described in this section. We consider three separate settings for training: (1) training with only images as input, (2) training with images and textual descriptor values as input, and (3) the same setting as (1) with using random IDs as target labels. Details in experimental setup are included in Appendix C.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 178, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 178, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 178, + 392 + ], + "type": "text", + "content": "5.1. Overall Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 542 + ], + "type": "text", + "content": "Fine-tuning does not help with vision-language mapping. Table 5 shows that the model fails to learn meaningful patterns in the data, even after training 50 epochs. However, when the textual descriptors are provided as input, the model can easily learn the function achieving " + }, + { + "bbox": [ + 55, + 399, + 295, + 542 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 55, + 399, + 295, + 542 + ], + "type": "text", + "content": " accuracy (a " + }, + { + "bbox": [ + 55, + 399, + 295, + 542 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 55, + 399, + 295, + 542 + ], + "type": "text", + "content": " improvement from the zero-shot setting). This shows that the model is unable to extract the same descriptor values from the image input. Unfreezing the encoder for fine-tuning also does not provide any significant improvement. We include a further discussion in Appendix D.3. The results point to a limitation in the vision model's ability to encode the images, which we will investigate next." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 276, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 276, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 276, + 563 + ], + "type": "text", + "content": "5.2. Understanding the Visual Representations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "In order to further understand the limitations of the vision component, we conduct an analysis to better understand how well different vision encoder models can differentiate between different paths in the dataset. To this end, we first apply a linear probe to see how easily distinguishable different images are. Specifically, given a pair of images, we first use the vision encoder to extract high-dimensional feature representations for both images. These features are then concatenated and passed through a simple binary classification layer (i.e., the probe). The probe is trained to predict a label of 1 if the images are the same and 0 otherwise. We experiment with various SOTA vision encoders, namely CLIP" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 188, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 332 + ], + "type": "text", + "content": "[55] base and large varieties, LAION-OpenCLIP [25, 56], SigLip [80], DINO [8], and DINO-v2 [49] and analyze how well their learned representations capture visual similarities and differences. We use a set of 1,000 randomly sampled path pairs with a balanced label distribution to train the probe, and look at whether the model can learn to distinguish between these paths. We also perform an experiment where we fine-tune the vision encoder along with the probe. In this setting, our goal is to gauge if carefully fine-tuning a vision encoder can potentially improve the model's performance in low-level perception. Finally, in both settings, we present the average cosine similarity between distinct paths." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 334, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 405 + ], + "type": "text", + "content": "Vision encoders cannot distinguish between paths. From Table 4, it can be seen that vision encoder models are unable to provide representations that are significantly different for the probe to tell if they are the same. This is further supported by the high values for the average cosine similarity across all models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 408, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 540 + ], + "type": "text", + "content": "Fine-tuning the encoders on a discrimination task can help disentangle the visual representations. By unfreezing the encoder weights and training them to identify whether two paths are identical, we enhance their adaptability to the task. The results in Table 4 demonstrate that this approach effectively disentangles the learned representations, resulting in significantly improved performance and increased separability, as evidenced by the notably lower cosine similarity between non-identical paths. The results thus imply the need for carefully fine-tuning task-specific vision encoders for path evaluation on PATHEVAL." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "This work explored the use of VLMs as evaluators in path-planning scenarios. We evaluated a number of VLMs on our proposed benchmark, PATHEVAL, and found that these models struggle with low-level perception. Specifically, we find that visual encoders used by SOTA models are unable to discern the differences between different paths in intricate scenarios. We hope that PATHEVAL will inspire researchers to further explore ways to improve the visual capabilities of VLMs and contribute to finding better ways to incorporate foundation models for developing more flexible, robust, and scalable planning paradigms." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6893" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 152 + ], + "type": "text", + "content": "This project was supported by resources provided by the Office of Research Computing at George Mason University (URL: https://orc.gmu.edu) and funded in part by grants from the National Science Foundation (Award Number 2018631)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 185, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 185, + 296, + 228 + ], + "type": "text", + "content": "[1] Chirag Agarwal, Sree Harsha Tanneru, and Himabindu Lakkaraju. Faithfulness vs. Plausibility: On the (Un)Reliability of Explanations from Large Language Models, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 231, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 231, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 231, + 296, + 285 + ], + "type": "text", + "content": "[2] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can Large Language Models be Good Path Planners? A Benchmark and Investigation on Spatial-temporal Reasoning. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 287, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 287, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 287, + 295, + 319 + ], + "type": "text", + "content": "[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Look Further Ahead: Testing the Limits of GPT-4 in Path Planning. IEEE CASE, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 322, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 322, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 296, + 366 + ], + "type": "text", + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In International Conference on Computer Vision (ICCV), 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 368, + 296, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 368, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 61, + 368, + 296, + 467 + ], + "type": "text", + "content": "[5] Kate Baumli, Satinder Baveja, Feryal Behbahani, Harris Chan, Gheorghe Comanici, Sebastian Flennerhag, Maxime Gazeau, Kristian Holsheimer, Dan Horgan, Michael Laskin, Clare Lyle, Hussain Masoom, Kay McKinney, Volodymyr Mnih, Alexander Neitz, Dmitry Nikulin, Fabio Pardo, Jack Parker-Holder, John Quan, Tim Roktaschel, Himanshu Sahni, Tom Schaul, Yannick Schroecker, Stephen Spencer, Richie Steigerwald, Luyu Wang, and Lei Zhang. Vision-Language Models as a Source of Rewards, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 468, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 468, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 61, + 468, + 296, + 578 + ], + "type": "text", + "content": "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 580, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 580, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 61, + 580, + 296, + 657 + ], + "type": "text", + "content": "[7] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21819–21830, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 296, + 713 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 3, 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 160 + ], + "type": "text", + "content": "[9] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee Wong. MapGPT: Map-guided prompting with adaptive path planning for vision-and-language navigation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9796-9810, Bangkok, Thailand, 2024. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "text", + "content": "[10] Xuweiyi Chen, Ziqiao Ma, Xuejun Zhang, Sihan Xu, Shengyi Qian, Jianing Yang, David Fouhey, and Joyce Chai. Multi-object hallucination in vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 219, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 274 + ], + "type": "text", + "content": "[11] Yongchao Chen, Jacob Arkin, Charles Dawson, Yang Zhang, Nicholas Roy, and Chuchu Fan. Autotamp: Autoregressive task and motion planning with llms as translators and checkers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6695-6702, 2024. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 275, + 554, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 554, + 340 + ], + "type": "text", + "content": "[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. arXiv preprint arXiv:2312.14238, 2023. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "type": "text", + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "text", + "content": "[14] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 455, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 554, + 553 + ], + "type": "text", + "content": "[15] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 555, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 554, + 620 + ], + "type": "text", + "content": "[16] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, brian richter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B. Tenenbaum, Leslie Pack Kaelbling, Andy Zeng, and Jonathan Tompson. Video language planning. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 622, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 643 + ], + "type": "text", + "content": "[17] Abhimanyu Dubey et. al. The Llama 3 Herd of Models, 2024. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[18] Difei Gao, Ruiping Wang, Shiguang Shan, and Xilin Chen. Cric: A vqa dataset for compositional reasoning on vision and commonsense. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5561-5578, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6894" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 172 + ], + "type": "text", + "content": "[20] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "type": "text", + "content": "[21] Lin Guan, Yifan Zhou, Denis Liu, Yantian Zha, Heni Ben Amor, and Subbarao Kambhampati. Task success is not enough: Investigating the use of video-language models as behavior critics for catching undesirable agent behaviors. In First Conference on Language Modeling, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 293 + ], + "type": "text", + "content": "[22] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 326 + ], + "type": "text", + "content": "[23] Xu Han, Qiannan Yang, Xianda Chen, Xiaowen Chu, and Meixin Zhu. Generating and evolving reward functions for highway driving with large language models, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "text", + "content": "[24] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 295, + 446 + ], + "type": "text", + "content": "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 448, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 448, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 448, + 295, + 502 + ], + "type": "text", + "content": "[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's \"up\" with vision-language models? Investigating their struggle with spatial reasoning. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 568 + ], + "type": "text", + "content": "[27] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Audit Verma, Kaya Stechly, Siddhant Bhambri, Lucas Paul Saldyt, and Anil B Murthy. Position: LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks. In _Forty-first International Conference on Machine Learning_, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 613 + ], + "type": "text", + "content": "[28] L.E. Kavraki, P. Svestka, J.-C. Latombe, and M.H. Overmars. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and Automation, 12(4):566-580, 1996. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "type": "text", + "content": "[29] J.J. Kuffner and S.M. LaValle. RRT-connect: An efficient approach to single-query path planning. In Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065), pages 995-1001 vol.2, 2000. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 691 + ], + "type": "text", + "content": "[30] Steven M. LaValle. Rapidly-exploring random trees: a new tool for path planning. The annual research report, 1998. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[31] Jan Leike. Why i'm optimistic about our alignment approach: Evaluation is easier than generation." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "https://aligned.substack.com/i/88447351/evaluation-is-easier-than-generation,2022.Accessed:2024-11.1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "text", + "content": "[32] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[33] Hao Li, Xue Yang, Zhaokai Wang, Xizhou Zhu, Jie Zhou, Yu Qiao, Xiaogang Wang, Hongsheng Li, Lewei Lu, and Jifeng Dai. Auto mc-reward: Automated dense reward design with large language models for minecraft. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 342 + ], + "type": "text", + "content": "[35] Yizhi Li, Ge Zhang, Yinghao Ma, Ruibin Yuan, Kang Zhu, Hangyu Guo, Yiming Liang, Jiaheng Liu, Zekun Wang, Jian Yang, Siwei Wu, Xingwei Qu, Jinjie Shi, Xinyue Zhang, Zhenzhu Yang, Xiangzhou Wang, Zhaoxiang Zhang, Zachary Liu, Emmanouil Benetos, Wenhao Huang, and Chenghua Lin. OmniBench: Towards The Future of Universal Omni-Language Models, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 419 + ], + "type": "text", + "content": "[36] Hui Sheng Lim, Shuangshuang Fan, Christopher K.H. Chin, Shuhong Chai, Neil Bose, and Eonjoo Kim. Constrained path planning of autonomous underwater vehicle using selectively-hybridized particle swarm optimization algorithms. IFAC-PapersOnLine, 52(21):315-322, 2019. 12th IFAC Conference on Control Applications in Marine Systems, Robotics, and Vehicles CAMS 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2024. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 499, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 533 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 534, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 600 + ], + "type": "text", + "content": "[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[41] Tomás Lozano-Pérez and Michael A. Wesley. An algorithm for planning collision-free paths among polyhedral obstacles. Communications of the ACM, 22(10):560-570, 1979. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6895" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[43] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 195 + ], + "type": "text", + "content": "[44] Silin Meng, Yiwei Wang, Cheng-Fu Yang, Nanyun Peng, and Kai-Wei Chang. LLM-a*: Large language model enhanced incremental heuristic search on path planning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1087-1102, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 250 + ], + "type": "text", + "content": "[45] Shota Onohara, Atsuyuki Miyai, Yuki Imajuku, Kazuki Egashira, Jeonghun Baek, Xiang Yue, Graham Neubig, and Kiyoharu Aizawa. Jmmm: A japanese massive multidiscipline multimodal understanding benchmark for culture-aware evaluation, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 252, + 248, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 252, + 248, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 248, + 262 + ], + "type": "text", + "content": "[46] OpenAI. GPT-4V(ision) System Card. 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 264, + 231, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 231, + 275 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 231, + 275 + ], + "type": "text", + "content": "[47] OpenAI. GPT-4o System Card, 2024. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 277, + 232, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 277, + 232, + 287 + ], + "spans": [ + { + "bbox": [ + 57, + 277, + 232, + 287 + ], + "type": "text", + "content": "[48] OpenAI. GPT-4 Technical Report, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 289, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 388 + ], + "type": "text", + "content": "[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 295, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 295, + 455 + ], + "type": "text", + "content": "[50] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 950-974, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 531 + ], + "type": "text", + "content": "[51] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14760-14769, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "text", + "content": "[52] Clment Petres, Yan Pailhas, Pedro Patron, Yvan Petillot, Jonathan Evans, and David Lane. Path planning for autonomous underwater vehicles. IEEE Transactions on Robotics, 23(2):331-341, 2007. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 612 + ], + "type": "text", + "content": "[53] Evis Plaku, Erion Plaku, and Patricio Simari. Clearance-driven motion planning for mobile robots with differential constraints. Robotica, 36(7):971-993, 2018. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 657 + ], + "type": "text", + "content": "[54] Patrick A. Plonski, Pratap Tokekar, and Volkan Isler. Energy-Efficient Path Planning for Solar-Powered Mobile Robots, pages 717-731. Springer International Publishing, Heidelberg, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pam Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "the 38th International Conference on Machine Learning, 2021. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 193 + ], + "type": "text", + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[57] Jie-Jing Shao, Jiang-Xin Shi, Xiao-Wen Yang, Lan-Zhe Guo, and Yu-Feng Li. Investigating the Limitation of CLIP Models: The Worst-Performing Categories, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[58] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B. Tenenbaum, Leslie Pack Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[59] Loan A. Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics & Automation Magazine, 19(4):72-82, 2012. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "type": "text", + "content": "[60] Yinan Sun, Zicheng Zhang, Haoning Wu, Xiaohong Liu, Weisi Lin, Guangtao Zhai, and Xiongkuo Min. Explore the hallucination on low-level perception for mllms, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 339, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 555, + 425 + ], + "type": "text", + "content": "[61] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "type": "text", + "content": "[62] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "text", + "content": "[63] Shengbang Tong, Erik Jones, and Jacob Steinhardt. Mass-producing failures of multimodal systems with language models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "type": "text", + "content": "[64] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs, 2024. 1, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 538, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 538, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 538, + 553, + 713 + ], + "type": "text", + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6896" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 127 + ], + "type": "text", + "content": "Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 193 + ], + "type": "text", + "content": "[66] Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language models don't always say what they think: unfaithful explanations in chain-of-thought prompting. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "text", + "content": "[67] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large Language Models Still Can't Plan (A Benchmark for LLMs on Planning and Reasoning about Change). In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "text", + "content": "[68] Karthik Valmeekam, Kaya Stechly, and Subbarao Kambhampati. LLMs Still Can't Plan; Can LRMs? A Preliminary Evaluation of OpenAI's o1 on PlanBench, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 358 + ], + "type": "text", + "content": "[69] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 425 + ], + "type": "text", + "content": "[70] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6622-6631, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "text", + "content": "[72] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. ECCV, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "text", + "content": "[73] Tianbao Xie, Siheng Zhao, Chen Henry Wu, Yitao Liu, Qian Luo, Victor Zhong, Yanchao Yang, and Tao Yu. Text2Reward: Reward Shaping with Language Models for Reinforcement Learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[74] Yutaro Yamada, Yihan Bao, Andrew Kyle Lampinen, Jungo Kasai, and Ilker Yildirim. Evaluating spatial understanding of large language models. Transactions on Machine Learning Research, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 666 + ], + "type": "text", + "content": "[75] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[76] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 517 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 216 + ], + "type": "text", + "content": "[78] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI. In Proceedings of CVPR, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "text", + "content": "[79] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark. arXiv preprint arXiv:2409.02813, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 284, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 328 + ], + "type": "text", + "content": "[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 329, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 553, + 406 + ], + "type": "text", + "content": "[81] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, and Jie Fu. CMMU: A Chinese Massive Multi-discipline Multimodal Understanding Benchmark, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 407, + 553, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 440 + ], + "type": "text", + "content": "[82] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 441, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 473 + ], + "type": "text", + "content": "[83] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy Improvement using Language Feedback Models, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 475, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 517 + ], + "type": "text", + "content": "[84] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7W: Grounded Question Answering in Images. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6897" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_content_list.json b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a99f37f46fa02b55c096a5a83816e67d7f61062 --- /dev/null +++ b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_content_list.json @@ -0,0 +1,1806 @@ +[ + { + "type": "text", + "text": "Event Ellipsometer: Event-based Mueller-Matrix Video Imaging", + "text_level": 1, + "bbox": [ + 173, + 130, + 825, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ryota Maeda $^{1,2}$ Yunseong Moon $^{1}$ Seung-Hwan Baek $^{1}$ $^{1}$ POSTECH $^{2}$ University of Hyogo", + "bbox": [ + 253, + 179, + 743, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 257, + 326, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Light-matter interactions modify both the intensity and polarization state of light. Changes in polarization, represented by a Mueller matrix, encode detailed scene information. Existing optical ellipsometers capture Mueller-matrix images; however, they are often limited to capturing static scenes due to long acquisition times. Here, we introduce Event Ellipsometer, a method for acquiring a Mueller-matrix video for dynamic scenes. Our imaging system employs fast-rotating quarter-wave plates (QWPs) in front of a light source and an event camera that asynchronously captures intensity changes induced by the rotating QWPs. We develop an ellipsometric-event image formation model, a calibration method, and an ellipsometric-event reconstruction method. We experimentally demonstrate that Event Ellipsometer enables Mueller-matrix video imaging at 30 fps, extending ellipsometry to dynamic scenes.", + "bbox": [ + 89, + 289, + 483, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 556, + 220, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Polarization describes the oscillation of the electric field in light waves, encoding valuable information about the scenes with which light interacts. The polarization state of light can be represented as a Stokes vector $\\mathbf{s} \\in \\mathbb{R}^{4 \\times 1}$ [17]. Polarimetric image analysis focuses on capturing complete or partial forms of the per-pixel Stokes vector and has been extensively studied for shape-from-polarization [31, 36], diffuse-specular separation [19], reflection removal [41], seeing through scattering [42], and transparent object segmentation [32].", + "bbox": [ + 89, + 583, + 482, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ellipsometry advances polarimetric imaging by analyzing the polarization state of light captured under varying polarization states of illumination. This allows for acquiring polarimetric reflectance, represented as a Mueller matrix $\\mathbf{M} \\in \\mathbb{R}^{4 \\times 4}$ [9], which comprehensively characterizes how light-matter interactions alter the polarization state of incident light. Ellipsometry has been widely used in material science [30, 38], biology [1, 23], and has recently gained attention in computer vision and graphics for 3D shape and reflection analysis [6, 7, 43], material acquisition [8], light transport decomposition [5, 35], and photoelasticity analy-", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75b8dfbe7a8dcc63c0d043df6f7ac2b727225229cd7f24508b188991ab2a1416.jpg", + "image_caption": [ + "(a) Overview diagram of Event Ellipsometer" + ], + "image_footnote": [], + "bbox": [ + 516, + 257, + 906, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3a693dcb625d5226236bd6a6200934911058c7dbfb4bf65c27a425afa108ee1.jpg", + "image_caption": [ + "(b) Capture configuration" + ], + "image_footnote": [], + "bbox": [ + 516, + 398, + 735, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9a0c624fe9ac22f84299a4f97bfea94be6a5de132f22f80ec54d2a184626339.jpg", + "image_caption": [ + "HDR imaging Tape detection (c) Applications", + "Figure 1. Overview of Event Ellipsometer. (a) Our imaging system captures the Mueller matrix at 30 fps from event streams induced by the continuously rotating QWPs. (b) With our experimental prototype, (c) we demonstrate ellipsometric analysis for dynamic scenes and various applications." + ], + "image_footnote": [], + "bbox": [ + 736, + 398, + 821, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86d21b84eee3a9365a10fb85e8e8957c9709facf300fb66ce9dd09ecac0a7c5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 398, + 903, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tape detection cations", + "bbox": [ + 828, + 534, + 897, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sis [18].", + "bbox": [ + 511, + 652, + 570, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the rich information provided by ellipsometry, its applications have been mainly limited to static scenes because conventional methods require capturing multiple images while mechanically rotating polarizing optics. Typically, more than 20 rotation angles [8] are necessary, leading to long acquisition times and making it impractical for capturing dynamic scenes. Although single-shot Mueller matrix imaging techniques using conventional sensors exist [12, 33, 49], they assume planar target scenes at fixed distances, significantly sacrifice sensor resolution, or use custom-fabricated nano- or micro-optical elements. Additionally, conventional intensity sensors are limited in capturing high-dynamic range (HDR) scenes, often requiring the capture of additional images with multiple exposures.", + "bbox": [ + 509, + 670, + 906, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present Event Ellipsometer, a Mueller-", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21804", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "matrix imaging method capable of capturing dynamic and HDR scenes. Departing from using conventional intensity sensors, we utilize an event camera, which asynchronously records intensity changes. We equip fast-rotating QWPs and linear polarizers (LPs) in front of both the event camera and an LED light source. The rotating QWPs modulate the polarization state of light emitted and received by the imaging system, resulting in intensity changes captured as events.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We develop an ellipsometric-event image formation model that relates the time differences of adjacent events to the Mueller matrix, normalized by its first element. Using this model, we propose a two-stage Mueller-matrix reconstruction method consisting of per-pixel estimation and spatiotemporal propagation. We incorporate physical validity constraints to handle outliers from sensor noise and scene motion. We also devise calibration methods.", + "bbox": [ + 89, + 227, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimentally, we demonstrate Mueller-matrix imaging at 30 fps, achieving a mean-squared error of 0.045 for materials with known Mueller matrices. Unlike previous single-shot methods [49], Event Ellipsometer can capture non-planar objects and does not compromise spatial resolution. Furthermore, our method is capable of capturing HDR scenes without the need for additional measurement with different exposures.", + "bbox": [ + 89, + 347, + 480, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 469, + 410, + 483 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Event Ellipsometer, enabling Mueller-matrix imaging for dynamic scenes at 30 fps.", + "- We develop an imaging system using an event camera and a light source, each equipped with synchronized fast-rotating QWPs.", + "- We formulate an ellipsometric-event image formation model relating event streams to the Mueller matrix, along with a calibration method and a robust ellipsometric-event reconstruction algorithm.", + "- We experimentally show the high accuracy of Event Ellipsometer on samples with known ground truth, and demonstrate ellipsometric analysis of dynamic scenes with applications on photoelasticity, human capture, HDR imaging, and tape detection." + ], + "bbox": [ + 91, + 484, + 480, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 709, + 232, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imaging with Polarized Illumination Polarimetric imaging with polarized illumination exploits polarization-dependent light transport to extract material and geometric scene properties. Previous methods often capture scenes with single or a few polarized illuminations and fit parametric models to the captured images [14, 27, 28, 46]. However, this approach is limited by the representation power of the parametric models, thus cannot reveal the true polarimetric reflectance of real-world scenes. Ellipsometry extends polarimetric imaging by directly capturing the Mueller matrix, providing a comprehensive characteriza", + "bbox": [ + 89, + 734, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion of material polarization properties [4, 17, 20]. It has applications in material science [30, 38], biology [1, 23], and has recently been applied to computer vision and graphics for tasks such as 3D shape analysis [6, 7, 43], material acquisition [8], and light transport decomposition [5, 35]. However, conventional ellipsometry techniques are unsuitable for dynamic scenes because they require capturing multiple images while mechanically rotating polarizing optics and capturing each with an intensity sensor, leading to long acquisition. Single-shot Mueller-matrix imaging methods have been proposed [12, 33, 49], however these approaches often compromise spatial resolution and are limited to planar scenes at fixed distances, restricting their applicability to general dynamic scenes.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event Cameras for Photometric Analysis Event cameras offer a high temporal resolution of microseconds and a HDR compared to standard frame-based intensity cameras [21]. These advantages have been leveraged for photometric analysis by combining continuously modulated optical light sources or filters. For instance, intensity-modulated illumination with an event camera enables radiance estimation [15, 24] and bispectral photometry [45]. Yu et al. [48] demonstrated event-based photometric stereo for normal estimation with a mechanically-rotating point light source. Hawks et al. [25] and Muglikar et al. [37] explored passive linear-polarization imaging by rotating a linear polarizer in front of an event camera. Existing event-based vision methods cannot capture full polarization reflectance properties as a Mueller-matrix image, limiting their utility.", + "bbox": [ + 511, + 324, + 903, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imaging with Rotating Optical Elements Rotating optical components have been used for HDR and multispectral imaging [40] and high-speed video reconstruction [13] using a conventional intensity camera. Recently, rotating optics with event cameras have been demonstrated to leverage asynchronous operation and high temporal resolution of the event cameras [25, 26, 37]. These methods rotate optical elements only in front of an event camera without illumination modulation.", + "bbox": [ + 511, + 571, + 903, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Imaging System", + "text_level": 1, + "bbox": [ + 511, + 723, + 671, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental Prototype We build an imaging system using an LED light source (Thorlabs MCWHLP3), fast-rotating QWPs (Edmund Optics WP140HE), fixed LPs (Thorlabs WP25M-VIS), and an event camera (Prophesee EVK4). Figure 2(a) and (c) depict our setup, where a pair of a QWP and an LP is placed in front of the event camera and the light source, respectively. We rotate the QWPs so that the event camera detects the event streams caused by the rotating QWPs, from which the normalized Mueller matrix is reconstructed. We can configure the illumination", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21805", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c03fb221471db728db7ed2545b1af70a9754bf0e18549e23b37f33e4449ae44.jpg", + "image_caption": [ + "(a) Schematic diagram" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 207 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3a90a26a4009ce56153453a59a00e14c9a9bbb365c2d68d6d56726799edf44ab.jpg", + "image_caption": [ + "(b) Timeline of QWPs rotation and event measurement", + "Figure 2. Imaging system of Event Ellipsometer. (a) Schematic diagram illustrating the optical arrangement and hardware operation. (b) Timeline showing the rotation of two QWPs and the event measurement. (c) Our hardware prototype. The system can move the light source and camera position for use in (d) Reflection mode or (e) Transmission mode." + ], + "image_footnote": [], + "bbox": [ + 91, + 220, + 480, + 393 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87c6155bb1edeb05151d21244a090246f9840b39a24a34c24b4b811e7a017b59.jpg", + "image_caption": [ + "(c) Hardware prototype" + ], + "image_footnote": [], + "bbox": [ + 483, + 88, + 802, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a15b567aef125c9a17b3b485302740ec647279efdc17e75189b98d01d6ff3f79.jpg", + "image_caption": [ + "QWP film" + ], + "image_footnote": [], + "bbox": [ + 805, + 88, + 903, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e1b3059d4828ee7022366407b074de23b5c015aa5ade8be29b7aef105caa29e0.jpg", + "image_caption": [ + "(d) Reflection mode" + ], + "image_footnote": [], + "bbox": [ + 483, + 271, + 694, + 393 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3195701c7db6a0391f4fa754e1788354108206742ef8f5038b534b92144a4dc.jpg", + "image_caption": [ + "(e) Transmission mode" + ], + "image_footnote": [], + "bbox": [ + 694, + 271, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the camera modules mounted on different breadboards in both the reflection model and the transmission mode as shown in Figure 2(d) and (e). Our system can be seen as a combination of the optical dual rotating retarder [3] and the event-based vision. For a complete list of components, refer to the Supplemental Material.", + "bbox": [ + 88, + 470, + 482, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rotating QWPs and Angle Encoder We rotate the QWP on the camera side five times faster than the one on the light source side [44]. We denote the QWP angles of the light source and the camera as $\\theta_{1,t} = \\omega t + i_1, \\theta_{2,t} = 5\\omega t + i_2$ where $\\omega = 30\\pi$ rad/sec is the angular velocity of the motor driving the light-source QWP. $i_1$ and $i_2$ are the initial QWP angles. One challenge here is to rotate the QWPs at such high speeds while recording their angles $\\theta_{1,t}$ and $\\theta_{2,t}$ over time $t$ . To this end, we use two independently controlled brushed direct current (DC) motors rotating the QWP films, resulting in an affordable configuration. Also, we develop custom angle encoders using a 3D-printed occluder, a photointerrupter, and an Arduino microcontroller. The occluder rotates at the same speed as the QWP and blocks the light path within the photointerrupter at every $\\pi$ rotation. The microcontroller detects such change at a microsecond resolution and emits a hardware trigger to the event camera. Figure 2(c) shows the angle encoder.", + "bbox": [ + 89, + 579, + 483, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame of Mueller Matrix Given the rotation speed of $\\omega$ , every $\\pi$ rotation of the light-source QWP takes $\\pi/\\omega \\approx$", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "33 ms. We set this as the effective temporal duration of each frame $f$ of the Mueller-matrix video which we aim to reconstruct. That is, as shown in Figure 2(b), each frame $f \\in \\{1, \\dots, F\\}$ of the Mueller matrix is estimated based on the events measured during the temporal duration $\\pi / \\omega$ . $F$ is the number of frames in a reconstructed Mueller-matrix video.", + "bbox": [ + 511, + 470, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Image Formation", + "text_level": 1, + "bbox": [ + 511, + 590, + 684, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we relate the event-camera measurements to the scene Mueller matrix which we aim to reconstruct.", + "bbox": [ + 511, + 616, + 905, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Polarimetric Modulation The LED light source emits unpolarized light of Stokes vector $\\mathbf{s} = [1,0,0,0]^{\\intercal}$ . At a time $t$ , the light passes through the LP and the QWP rotated by angle $\\theta_{1,t}$ , resulting in the Stokes vector $\\mathbf{Q}(\\theta_{1,t})\\mathbf{L}(0)\\mathbf{s}$ , where $\\mathbf{Q}(\\theta_{1,t})$ and $\\mathbf{L}(0)$ denote the Mueller matrices of a QWP rotated by $\\theta_{1,t}$ and an LP at $0^{\\circ}$ , respectively. The light then interacts with the target scene, undergoing polarization change represented by the scene Mueller matrix $\\mathbf{M}$ . After interaction, the light passes through another QWP at angle $\\theta_{2,t}$ and an LP in front of the event camera, resulting in the time-varying intensity $I_{t}$ incident on the sensor:", + "bbox": [ + 511, + 664, + 906, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} = [ \\mathbf {L} (0) \\mathbf {Q} (\\theta_ {2, t}) \\mathbf {M} \\mathbf {Q} (\\theta_ {1, t}) \\mathbf {L} (0) \\mathbf {s} ] _ {0}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 842, + 903, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\left[\\cdot\\right]_{0}$ denotes intensity, the first element of the Stokes vector.", + "bbox": [ + 511, + 869, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21806", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We rearrange Equation (1) and derive a matrix-vector form:", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} = \\mathbf {A} _ {t} \\hat {\\mathbf {M}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 122, + 480, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{M}} = [\\mathbf{M}_{00},\\mathbf{M}_{01},\\dots ,\\mathbf{M}_{33}]^{\\intercal}\\in \\mathbb{R}^{16\\times 1}$ is the vectorized form of $\\mathbf{M}$ and $\\mathbf{A}_t\\in \\mathbb{R}^{1\\times 16}$ is the system matrix defined as", + "bbox": [ + 89, + 147, + 482, + 193 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {A} _ {t} = [ 1, \\alpha_ {1} ^ {2}, \\alpha_ {1} \\alpha_ {2}, \\alpha_ {2}, \\alpha_ {3} ^ {2}, \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2}, \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2}, \\alpha_ {2} \\alpha_ {3} ^ {2}, \\\\ \\alpha_ {3} \\alpha_ {4}, \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4}, \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, - \\alpha_ {4} \\\\ \\left. - \\alpha_ {1} ^ {2} \\alpha_ {4}, - \\alpha_ {1} \\alpha_ {2} \\alpha_ {4}, - \\alpha_ {2} \\alpha_ {4} \\right], \\text {w h e r e} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 205, + 447, + 265 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {1} = \\cos (2 i _ {1} + 2 \\omega t), \\alpha_ {2} = \\sin (2 i _ {1} + 2 \\omega t),\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 268, + 424, + 284 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {3} = \\cos (2 i _ {2} + 1 0 \\omega t), \\alpha_ {4} = \\sin (2 i _ {2} + 1 0 \\omega t). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 287, + 480, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the full derivation, refer to the Supplemental Document.", + "bbox": [ + 89, + 316, + 480, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differential Intensity The event camera triggers events based on the temporal change of the logarithm of photocurrent [21]. Analytically differentiating the logarithm of Equation (2) with respect to time $t$ , we obtain", + "bbox": [ + 89, + 354, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\log I _ {t}}{\\partial t} = \\frac {\\frac {\\partial I _ {t}}{\\partial t}}{I _ {t}} = \\frac {\\frac {\\partial \\mathbf {A} _ {t}}{\\partial t} \\hat {\\mathbf {M}}}{\\mathbf {A} _ {t} \\hat {\\mathbf {M}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 426, + 482, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\frac{\\partial\\mathbf{A}_t}{\\partial t}$ is given as", + "bbox": [ + 89, + 476, + 240, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial \\mathbf {A} _ {t}}{\\partial t} = [ 0, - 4 \\alpha_ {1} \\alpha_ {2}, 2 \\alpha_ {1} ^ {2} - 2 \\alpha_ {2} ^ {2}, 2 \\alpha_ {1}, - 2 0 \\alpha_ {3} \\alpha_ {4}, \\\\ - 2 0 \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4} - 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2}, 2 \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2} - 2 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4} - 2 \\alpha_ {2} ^ {2} \\alpha_ {3} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 506, + 491, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n2 \\alpha_ {1} \\alpha_ {3} ^ {2} - 2 0 \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, 1 0 \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {4} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 558, + 328, + 577 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} 1 0 \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {1} ^ {2} \\alpha_ {4} ^ {2} - 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, \\\\ 2 \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4} + 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {4} ^ {2} - 2 \\alpha_ {2} ^ {2} \\alpha_ {3} \\alpha_ {4}, \\\\ 2 \\alpha_ {1} \\alpha_ {3} \\alpha_ {4} + 1 0 \\alpha_ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {2} \\alpha_ {4} ^ {2}, - 1 0 \\alpha_ {3}, - 1 0 \\alpha_ {1} ^ {2} \\alpha_ {3} + 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {4}, \\\\ - 2 \\alpha_ {1} ^ {2} \\alpha_ {4} - 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} + 2 \\alpha_ {2} ^ {2} \\alpha_ {4}, - 2 \\alpha_ {1} \\alpha_ {4} - 1 0 \\alpha_ {2} \\alpha_ {3} ]. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 578, + 509, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Events from Polarimetric Modulation For a pixel, we denote the time difference between consecutive events as $\\Delta t_{k}$ , where $k\\in \\{1,\\dots ,K\\}$ is the event index in each frame $f$ . $K$ is the number of detected events at that pixel in the frame $f$ . The time difference is known to be related to the change of photocurrent according to Taylor's expansion [21]:", + "bbox": [ + 89, + 694, + 482, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\log I _ {t _ {k}}}{\\partial t} = \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 799, + 480, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where event polarity $p_k \\in \\{+1, -1\\}$ is the sign of intensity change, and $C$ is a constant threshold of the event camera.", + "bbox": [ + 89, + 839, + 480, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By combining Equations (4) and (6), we relate the measured per-pixel events to the Mueller matrix, resulting in the", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50bc1e749a9554ccd8774ced66bcaf79744b8f6c330feabd040c8c067a8d1c88.jpg", + "image_caption": [ + "Figure 3. Overview of our Mueller-matrix reconstruction pipeline. This method consists of two steps: (1) per-pixel reconstruction and (2) propagation and refinement." + ], + "image_footnote": [], + "bbox": [ + 513, + 89, + 903, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image formation model:", + "bbox": [ + 511, + 231, + 674, + 244 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {p _ {k} C}{\\Delta t _ {k}} = \\frac {\\frac {\\partial \\mathbf {A} _ {t _ {k}}}{\\partial t _ {k}} \\hat {\\mathbf {M}}}{\\mathbf {A} _ {t _ {k}} \\hat {\\mathbf {M}}} \\\\ \\Rightarrow \\left(\\frac {\\partial \\mathbf {A} _ {t _ {k}}}{\\partial t _ {k}} - \\frac {p _ {k} C}{\\Delta t _ {k}} \\mathbf {A} _ {t _ {k}}\\right) \\hat {\\mathbf {M}} = 0 \\\\ \\Rightarrow \\mathbf {B} _ {t _ {k}} \\dot {\\mathbf {M}} = 0, \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 255, + 903, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_{t_k} \\in \\mathbb{R}^{1 \\times 16}$ is the system matrix.", + "bbox": [ + 511, + 362, + 787, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Reconstruction", + "text_level": 1, + "bbox": [ + 513, + 391, + 665, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The image formation model in Equation (7) allows us to reconstruct the Mueller matrix $\\hat{\\mathbf{M}}$ . Our reconstruction method, depicted in Figure 3, consists of two main steps: (1) per-pixel reconstruction and (2) spatio-temporal propagation.", + "bbox": [ + 511, + 416, + 903, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Per-pixel Reconstruction", + "text_level": 1, + "bbox": [ + 513, + 500, + 740, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this step, our goal is to estimate a Mueller matrix in a computationally efficient manner and being robust to noisy events.", + "bbox": [ + 511, + 523, + 903, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we stack the vectors $\\mathbf{B}_{t_k}$ for all $k$ , forming a matrix $\\mathbf{B} \\in \\mathbb{R}^{K \\times 16}$ . We then solve for the vectorized Mueller matrix $\\hat{\\mathbf{M}}$ by minimizing the weighted least-squares problem:", + "bbox": [ + 511, + 569, + 903, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\hat {\\mathbf {M}}} {\\text {m i n i m i z e}} \\| \\mathbf {W D B M} \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 627, + 903, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W} = \\mathrm{diag}(\\mathbf{w})$ is a diagonal matrix of weights $\\mathbf{w} = [w_1, \\ldots, w_K] \\in \\mathbb{R}^{1 \\times K}$ , initialized as ones: $w_k = 1$ . The matrix $\\mathbf{D} = \\mathrm{diag}([\\Delta t_1, \\ldots, \\Delta t_K])$ is a weight matrix that adjusts the influence of each event according to sampling density. This weighting ensures that more densely sampled events at small $\\Delta t$ do not dominate the model fitting. We solve Equation (8) using singular value decomposition (SVD) to obtain $\\hat{\\mathbf{M}}$ while avoiding the trivial zero solution.", + "bbox": [ + 511, + 657, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, we filter the reconstructed Mueller matrix $\\hat{\\mathbf{M}}$ to ensure physical validity. Sensor noise and rapid scene dynamics can make solving Equation (8) challenging, often resulting in physically invalid Mueller matrices. These invalid matrices can trap the optimization in local minima. To address this, we apply Cloude's Mueller matrix filtering [16] to project the estimated matrix onto the space of", + "bbox": [ + 511, + 794, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21807", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d257cb698b40f9e6bad5b05ddc08e647a7d1ccc9d6dea2e23c78b5de3f71fa6.jpg", + "image_caption": [ + "Figure 4. Synthetic data evaluation result. (a) The rendered images include two materials: blue silicone and brass. (b) The plot shows the error (mean absolute error) of the reconstructed Mueller matrix over the number of iterations. (c) Ground truth Mueller matrix. (d)&(e) Top: the reconstructed Mueller-matrix images for the SVD initialization and the full-stage, respectively. Pixels with insufficient event counts are visualized in white. Bottom: a plot of the differentiation of log intensity and fitted line with our method." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 908, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "physically valid Mueller matrices:", + "bbox": [ + 89, + 407, + 318, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} \\leftarrow \\operatorname {C l o u d e F i l t e r} (\\hat {\\mathbf {M}}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 434, + 480, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where CloudeFilter $(\\cdot)$ denotes the filtering function.", + "bbox": [ + 89, + 462, + 434, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Third, we recalculate the weight vector $\\mathbf{w}$ using the physically valid Mueller matrix $\\hat{\\mathbf{M}}$ :", + "bbox": [ + 89, + 478, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {k} = 1 / \\max \\left(\\left| \\mathbf {B} _ {t _ {k}} \\hat {\\mathbf {M}} \\right|, \\epsilon\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 520, + 480, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a small positive constant to avoid division by zero sets as $10^{-6}$ in our experiments. This weight reduces the influence of outlier events and enables robust estimation.", + "bbox": [ + 89, + 547, + 482, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We iterate these three steps—solving the weighted least-squares problem, applying physical validity filtering, and recomputing the weights—until convergence, resulting in the Mueller matrix estimate $\\hat{\\mathbf{M}}$ for each pixel $(x,y)$ and frame $f$ . We set the maximum number of iterations to 5 in our experiments", + "bbox": [ + 89, + 594, + 483, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Spatio-temporal Propagation and Refinement", + "text_level": 1, + "bbox": [ + 89, + 694, + 475, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we start by denoting the estimated Mueller matrix at a pixel $(x,y)$ and frame $f$ as $\\hat{\\mathbf{M}}_{xyf}$ and the corresponding system matrix as $\\mathbf{B}_{xyf}$ , and weight matrix as $\\mathbf{D}_{xyz}$ . This second stage refines the Mueller matrix $\\hat{\\mathbf{M}}_{xyf}$ estimated from the first stage using spatio-temporal neighboring pixels. This process is inspired by PatchMatchStereo from stereoscopic depth estimation [10, 11, 22].", + "bbox": [ + 89, + 717, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we perform propagation. We consider neighboring pixels in both spatial and temporal axes, denoted by $(x', y', f') \\in \\mathcal{N}(x, y, f)$ . The set $\\mathcal{N}$ defines a spatio-temporal neighborhood extended from the red-black spatial-only neighborhood [22]. For each pixel at a frame", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$(x,y,f)$ , we update its Mueller matrix if a neighboring pixel $(x',y',f')$ offers a lower error:", + "bbox": [ + 511, + 407, + 906, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {x y f} \\leftarrow \\hat {\\mathbf {M}} _ {x ^ {\\prime} y ^ {\\prime} f ^ {\\prime}},\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 446, + 653, + 464 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {i f} \\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x y f} \\| _ {1} > \\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x ^ {\\prime} y ^ {\\prime} f ^ {\\prime}} \\| _ {1}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 467, + 903, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we refine the Mueller matrix by applying random perturbations and accepting the perturbation if it reduces the error:", + "bbox": [ + 511, + 493, + 905, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {x y f} \\leftarrow \\hat {\\mathbf {M}} _ {x y f} ^ {\\text {p e r t u r b}}, \\text {w h e r e}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 544, + 696, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {x y f} ^ {\\text {p e r t u r b}} = \\operatorname {C l o u d e F i l t e r} \\left(\\hat {\\mathbf {M}} _ {x y f} \\odot (\\mathbf {1} + \\sigma \\mathbf {N})\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 566, + 903, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x y f} \\right\\| _ {1} > \\left\\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\mathbf {M} _ {x y f} ^ {\\text {p e r t u r b}} \\right\\| _ {1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 590, + 903, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{M}}_{xyf}^{\\mathrm{perturb}}$ is the perturbed Mueller matrix, $\\odot$ denotes element-wise multiplication, $\\mathbf{1}$ is an all-ones matrix, $\\mathbf{N} \\in \\mathbb{R}^{4 \\times 4}$ contains random values from a standard normal distribution, and $\\sigma$ is a scalar set to 0.01. This multiplicative perturbation preserves the relative scale of the Mueller matrix elements, allowing larger elements to undergo proportionally larger adjustments while smaller elements change minimally. We apply Cloude's filtering to the perturbed Mueller matrix to ensure physical validity.", + "bbox": [ + 511, + 619, + 905, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We repeat the propagation and refinement steps iteratively for a fixed number of iterations set to 10 in our experiments to obtain the final Mueller matrix $\\hat{\\mathbf{M}}_{xyf}$ .", + "bbox": [ + 511, + 757, + 905, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Calibration", + "text_level": 1, + "bbox": [ + 511, + 814, + 635, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform one-time calibration of the system parameters: contrast threshold $(C)$ and QWP offset angles $(i_1$ and $i_2)$ . Calibration details are provided in the Supplementary Material.", + "bbox": [ + 511, + 839, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21808", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b9874e3c86ddcd40057fcf07f0f0e4a7726ce6886b40d2910e99a921e9b5067.jpg", + "image_caption": [ + "Figure 5. Assessment of reconstructed Mueller matrix on real data. (a) Evaluation with known optical elements. We show the corresponding mean squared errors (MSEs). (b) Measurement on an in-the-wild metal plate induces strong diagonal components." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 547, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95c7017207b1055cbc2971ecf99283d25592a95faf33ddf9487348def26644c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 88, + 905, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contrast Threshold We calibrate the per-pixel contrast threshold $C$ of an event sensor by analyzing pixel responses under controlled illumination stimuli. We increase and decrease the LED intensity linearly over time, and the event camera directly captures the LED light, generating events. We then fit the contrast threshold $C$ to the events per each pixel using Equation (6).", + "bbox": [ + 88, + 329, + 480, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "QWP Offset Angle We obtain the offset angles of the fast axis of the QWP in our system: $i_1$ and $i_2$ by capturing another QWP sample with known fast axis. Given the known Mueller matrix of the reference QWP, we find the best offset angles that minimize the reconstruction error of Equation (6).", + "bbox": [ + 88, + 452, + 482, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7. Results", + "text_level": 1, + "bbox": [ + 89, + 556, + 178, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation We implemented our reconstruction algorithm in $\\mathrm{C + + }$ using OpenMP CPU parallelization. It takes 62 seconds for per-pixel reconstruction (Section 5.1), and 68 seconds for the propagation and refinement processes (Section 5.2), measured for processing $500\\times 500$ -resolution event image sequence of 30 frames, with an average event rate of $153\\mathrm{MEv / s}$ , on the AMD Ryzen 7 7800X3D 8-core processor.", + "bbox": [ + 88, + 580, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Validation on Synthetic Data To evaluate our method, we render synthetic data in three steps. First, we mimic our capture system by replacing the event camera with an intensity camera in Mitsuba3 [29], and render an image sequence of a synthetic scene for each QWP angle with interval $\\omega t = 0.01^{\\circ}$ . We use the real-world polarimetric BRDFs to construct the synthetic scene [8]. Then, the rendered images are converted to event streams by simulating event-camera processing based on DVS-Voltmeter [34]. Last, we add Gaussian noise with a standard deviation of 0.5 and replace $5\\%$ of true events with values sampled from a Gaussian distribution with a standard deviation of 5.0. Figure 4", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shows the synthetic data and reconstructed Mueller matrix using our method. While initial SVD-based estimation suffers from noise with a reconstruction error of 0.11, our full reconstruction improves accuracy with a reconstruction error of 0.04.", + "bbox": [ + 511, + 329, + 906, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Validation on Real Data Figure 5(a) shows that our reconstruction closely matches the corresponding pseudo ground truth of real-world samples: air, linear polarizer $(0^{\\circ}$ and $45^{\\circ})$ , quarter-wave-plate $(0^{\\circ}$ and $45^{\\circ})$ . Figure 5(b) shows the result on an in-the-wild sample: a metal plate. The strong diagonal components in this result indicate the preservation of polarization, which is characteristic of reflections from metal surfaces. Notably, these Mueller matrix images are captured in just $33\\mathrm{ms}$ , significantly faster than the conventional ellipsometer, which requires several minutes due to its frame-based imaging principle.", + "bbox": [ + 511, + 429, + 908, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Photoelasticity of Transparent Gelatine Photoelasticity is an optical property whereby dielectric materials exhibit birefringence under deformation. Imaging photoelasticity has applications in mechanical stress and material analysis [39, 47]. Analyzing Mueller-matrix images reveals such stress distributions even in transparent objects [18]. Our system enables observing photoelasticity at video rates. Figure 6 shows Mueller-matrix images of a gelatine disk, revealing stress-dependent fringe patterns.", + "bbox": [ + 511, + 619, + 908, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transparent Tape Detection Detecting transparent objects is challenging for conventional cameras. Sticky tape, a common transparent material used for sealing cardboard, exhibits birefringent properties due to the molecular structure of its stretched plastic. Figure 7 shows the measurement result of sticky tape on a cardboard box. The reconstructed Mueller matrix shows the birefringent property, and we can clearly recognize the tape region. This", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21809", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2a796269c4579178bf5d1109ea761d6a9520c24de15f1bbf640c151a7b9397d.jpg", + "image_caption": [ + "(a) Measurement setup" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 339, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dfd408613e57492f96aaceffb9b904fddee952a68e00ea26fa7fcb2cc5a87b3.jpg", + "image_caption": [ + "Before applying force" + ], + "image_footnote": [], + "bbox": [ + 344, + 87, + 622, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69de5e4f17a7b054dd2a6372a85846bc76ba1f987393359ecf5adcf39a9de424.jpg", + "image_caption": [ + "After applying force", + "(b) Reconstructed Mueller matrix -1 +1" + ], + "image_footnote": [], + "bbox": [ + 625, + 87, + 905, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/402acf9ad9adc834b4606f7f0399564f92aa2e629080718b1ae390ac3ea33f80.jpg", + "image_caption": [ + "Figure 6. Photoelasticity analysis. (a) Experimental setup for measuring a gelatine disk in transmission mode. We gradually apply force for the gelatine disk to observe changes in photoelastic properties. (b) Reconstructed Mueller matrix images with different forces, revealing complex stress-dependent polarimetric patterns. Notably, as force increases, the Mueller matrix image shows denser fringe patterns.", + "(a) Target object" + ], + "image_footnote": [], + "bbox": [ + 91, + 327, + 300, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/962460144edd04345c48b98a143598e506e502ece3e65a9d4dd3ba5ff234e69f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 329, + 467, + 445 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9decf049b0c3c899c1ed388802f1744dbaee633dd22b1c4f1e26c9e8b5f37499.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 460, + 236, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd815b372e8f21768bac318cacb8c14d957b5da6ecccfcc7f712aef0e46840dd.jpg", + "image_caption": [ + "(b) Comparison" + ], + "image_footnote": [], + "bbox": [ + 93, + 555, + 236, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8b1f8a35d10105438711d3063a39782856bd8e90e61981fefe0f2c36c3dfe8c.jpg", + "image_caption": [ + "(c) Reconstructed Mueller matrix" + ], + "image_footnote": [], + "bbox": [ + 241, + 460, + 482, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2f7ae8724d6bbe20d66dd441421974b8003f13ce8b0fe6337fa8710283cd790.jpg", + "image_caption": [ + "Figure 7. Transparent tape detection. (a) The target object is transparent sticky tape on a box. (b)&(c) The Reconstructed Mueller matrix reveals the tape region, which is difficult to see with a conventional RGB camera." + ], + "image_footnote": [], + "bbox": [ + 299, + 659, + 478, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstration shows the potential for inspecting dynamically moving sealed boxes in automated industrial processes.", + "bbox": [ + 89, + 755, + 482, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic Human Capture Ellipsometers, as a comprehensive polarization imaging technique for human capture, can reveal hidden properties of the human face and hair that are difficult to analyze in conventional imaging and non-ellipsometric polarization imaging [2, 19]. In Figure 8, we", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstrate the Mueller-matrix reconstruction of dynamic facial expressions and hair movements, revealing intricate polarization properties. The reconstructed Mueller matrix for the face exhibits strong diagonal components associated with specular reflection, while the non-diagonal regions display weakly polarized reflection with a dependency on surface normal along the face edges. The hair result shows polarization property in the specular highlight regions.", + "bbox": [ + 511, + 330, + 906, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HDR Mueller-matrix Imaging Our event-based ellipsometer enables Mueller-matrix imaging of HDR scenes. Figure 9 shows reconstruction results on a scene with specular and dark diffuse reflections. While conventional ellipsometers require multiple exposures to prevent overexposure, our method captures the Mueller matrix in HDR scenes at 30 fps without the need for additional measurement.", + "bbox": [ + 511, + 484, + 906, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8. Discussion", + "text_level": 1, + "bbox": [ + 511, + 627, + 625, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Normalized Mueller-Matrix Imaging Normalized Mueller matrix has been extensively used as it provides rich scene information [5, 7, 18]. Since event cameras only detect intensity changes, not intensity itself, our method also reconstructs a normalized version of Mueller matrix per pixel, $\\mathbf{M} / \\mathbf{M}(0,0)$ . We leave reconstructing full Mueller-matrix as a future work that might be accomplished using an event-intensity hybrid imaging system.", + "bbox": [ + 511, + 655, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "End-to-end Real-time Pipeline Although our capture speed is fast, our current Mueller-matrix reconstruction relies solely on CPU acceleration, resulting in a non-realtime pipeline from capture to reconstruction. Implementing GPU acceleration could enable on-the-fly capture, reconstruction, and visualization [22, 48].", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21810", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1604f053161511fba6b9020f246bb5f1f1bdda08910b9714e0121bdf77abd7ae.jpg", + "image_caption": [ + "Figure 8. Mueller matrix acquisition for capturing dynamic human (a) face and (b) hair, demonstrating the capture of diffuse and specular polarimetric responses." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 908, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63a874a97f0afd956c1560d17b1e60692542d22c446ecbcd3e7075f33b7d28e1.jpg", + "image_caption": [ + "(a) Target object", + "Figure 9. Mueller matrix measurement for a HDR scene. (a) The target scene contains regions with both strong specular reflection and black diffuse reflection. (b) Our proposed method achieves an accurate Mueller-matrix image for the HDR scene in just $33~\\mathrm{ms}$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 402, + 285, + 627 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c13ff77dab3feb86078999efb4ef9d13eac5ba120f60f18b4e6589db59b5ec5.jpg", + "image_caption": [ + "(b) Reconstructed Mueller matrix" + ], + "image_footnote": [], + "bbox": [ + 285, + 402, + 483, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Event Bandwidth and Sensitivity The ideal setting of an event camera is to have a low contrast threshold and a short refractory period, to capture subtle intensity changes. However, this produces many events, exceeding the transmission bandwidth, resulting in lost or delayed events [21]. In contrast, when the contrast threshold is high, weakly polarized phenomena cannot be detected as events. Given this tradeoff, we defined a region of interest (ROI) in the event sensor and empirically configured the camera settings to find the best parameters.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DC Motor Synchronization For faster rotation speed and precise synchronization between two motors, it is a viable alternative to use brushless DC motors, also known as an electronically commutated (EC) motors. However, this increases system building costs.", + "bbox": [ + 511, + 405, + 906, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gradient Descent Gradient descent in an automatic differentiation framework offers an alternative approach for Mueller-matrix reconstruction. However, the number of events differs across pixels, which complicates the construction of a dense tensor for efficient computation on modern GPUs.", + "bbox": [ + 511, + 520, + 906, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 635, + 633, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced Event Ellipsometer, a Mueller-matrix video imaging method that combines an event camera and a light source with fast-rotating QWPs. Our image formation model, calibration, and reconstruction method enable robust Mueller-matrix imaging at 30 fps. We validate our method on synthetic and real data, and demonstrate Mueller-matrix imaging on photoelasticity, dynamic human hair and face capture, HDR imaging, and transparent tape detection.", + "bbox": [ + 511, + 664, + 906, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532, 2022R1A6A1A03052954, RS-2024-00437866), and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52.", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21811", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hans Arwin. Application of ellipsometry techniques to biological materials. Thin Solid Films, 519(9):2589-2592, 2011. 1, 2", + "[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 7", + "[3] RMA Azzam. Photopolarimetric measurement of the mueller matrix by fourier analysis of a single detected signal. Optics Letters, 2(6):148-150, 1978. 3", + "[4] Rasheed MA Azzam. Stokes-vector and mueller-matrix polarimetry. JOSA A, 33(7):1396-1408, 2016. 2", + "[5] Seung-Hwan Baek and Felix Heide. Polarimetric spatiotemporal light transport probing. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 7", + "[6] Seung-Hwan Baek and Felix Heide. All-photon polarimetric time-of-flight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17876-17885, 2022. 1, 2", + "[7] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM Trans. Graph., 37(6):268, 2018. 1, 2, 7", + "[8] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph., 39(4):139, 2020. 1, 2, 6", + "[9] Seung-Hwan Baek, Ramesh Raskar, Jinwei Ye, Akshit Dave, Achuta Kadambi, and Huajin Chen. Polarization-based visual computing. In ACM SIGGRAPH 2023 Courses, pages 1-1, 2023. 1", + "[10] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 5", + "[11] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In *Bmvc*, pages 1-11, 2011. 5", + "[12] Qizhi Cao, Min Jiang, Chenling Jia, Siyue Jiang, Jing Zhang, Baoli Yao, Mingwu Jin, Edward Dehoog, Lian Duan, Huahua Wang, et al. Snapshot imaging mueller matrix polarimeter using modified savart polariscopes. Applied Optics, 62(8):2124-2129, 2023. 1, 2", + "[13] Dorian Chan, Mark Sheinin, and Matthew O'Toole. Spincam: High-speed imaging via a rotating point-spread function. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10789-10799, 2023. 2", + "[14] Tongbo Chen, Hendrik PA Lensch, Christian Fuchs, and Hans-Peter Seidel. Polarization and phase-shifting for 3d scanning of translucent objects. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2", + "[15] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proceedings of the IEEE/CVF Conference on Computer" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 14760-14770, 2021. 2", + "[16] Shane R Cloude. Conditions for the physical realizability of matrix operators in polarimetry. In *Polarization Considerations for Optical Systems II*, pages 177–187. SPIE, 1990. 4", + "[17] Edward Collett. Field guide to polarization. Spie Bellingham, 2005. 1, 2", + "[18] Akshat Dave, Tianyi Zhang, Aaron Young, Ramesh Raskar, Wolfgang Heidrich, and Ashok Veeraraghavan. Nest: Neural stress tensor tomography by leveraging 3d photoelasticity. arXiv preprint arXiv:2406.10212, 2024. 1, 6, 7", + "[19] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 7", + "[20] Hiroyuki Fujiwara. Spectroscopic ellipsometry: principles and applications. John Wiley & Sons, 2007. 2", + "[21] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 4, 8", + "[22] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE international conference on computer vision, pages 873-881, 2015. 5, 7", + "[23] Nirmalya Ghosh and I Alex Vitkin. Tissue polarimetry: concepts, challenges, applications, and outlook. Journal of biomedical optics, 16(11):110801-110801, 2011. 1, 2", + "[24] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20616-20625, 2023. 2", + "[25] Michael Hawks and Michael Dexter. Event-based imaging polarimeter. Optical Engineering, 61(5):053101-053101, 2022. 2", + "[26] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 2", + "[27] Inseung Hwang, Daniel S Jeon, Adolfo Munoz, Diego Gutierrez, Xin Tong, and Min H Kim. Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 2", + "[28] Tomoki Ichikawa, Shohei Nobuhara, and Ko Nishino. Spiders: Structured polarization for invisible depth and reflectance sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25077-25085, 2024. 2", + "[29] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsubishi 3 renderer, 2022. https://mitsuba-renderer.org. 6" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21812", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] GE Jellison, Donovan N Leonard, Lawrence M Anovitz, Chad M Parish, Eliot D Specht, and TM Rosseel. Crystallographic orientation of uniaxial calcite and dolomite determined using reflection generalized ellipsometry. Journal of Applied Physics, 124(22), 2018. 1, 2", + "[31] Achuta Kadambi, Vage Taamazyan, Boxin Shi, and Ramesh Raskar. Polarized 3d: High-quality depth sensing with polarization cues. In Proceedings of the IEEE international conference on computer vision, pages 3370-3378, 2015. 1", + "[32] Agastya Kalra, Vage Taamazyan, Supreeth Krishna Rao, Kartik Venkataraman, Ramesh Raskar, and Achuta Kadambi. Deep polarization cues for transparent object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8602-8611, 2020. 1", + "[33] Michael W Kudenov, Michael J Escuti, Nathan Hagen, Eustace L Dereniak, and Kazuhiko Oka. Snapshot imaging mueller matrix polarimeter using polarization gratings. Optics letters, 37(8):1367-1369, 2012. 1, 2", + "[34] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In European Conference on Computer Vision, pages 578-593. Springer, 2022. 6", + "[35] Ryota Maeda and Shinsaku Hiura. Polarimetric light transport analysis for specular inter-reflection. IEEE Transactions on Computational Imaging, 10:876-887, 2024. 1, 2", + "[36] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 982-987. IEEE, 2003. 1", + "[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2", + "[38] David A Ramsey and Kenneth C Ludema. The influences of roughness on film thickness measurements by mueller matrix ellipsometry. Review of scientific instruments, 65(9):2874-2881, 1994. 1, 2", + "[39] Michele Scafidi, Giuseppe Pitarresi, Andrea Toscano, Giovanni Petrucci, Sabina Alessi, and Augusto Ajovalasit. Review of photoelastic image analysis applied to structural birefringent materials: glass and polymers. Optical Engineering, 54(8):081206-081206, 2015. 6", + "[40] Yoav Y Schechner and Shree K Nayar. Uncontrolled modulation imaging. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-II. IEEE, 2004. 2", + "[41] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, pages 814-819. IEEE, 1999. 1", + "[42] Yoav Y Schechner, Srinivasa G Narasimhan, and Shree K Nayar. Instant dehazing of images using polarization. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I. IEEE, 2001. 1" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Dominik Scheuble, Chenyang Lei, Seung-Hwan Baek, Mario Bijelic, and Felix Heide. Polarization wavefront lidar: Learning large scene reconstruction from polarized wavefronts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21241-21250, 2024. 1, 2", + "[44] Matthew H Smith. Optimization of a dual-rotating-retarder mueller matrix polarimeter. Applied optics, 41(13):2488-2493, 2002. 3", + "[45] Tsuyoshi Takatani, Yuzuha Ito, Ayaka Ebisu, Yinqiang Zheng, and Takahito Aoto. Event-based bispectral photometry using temporally modulated illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15647, 2021. 2", + "[46] Tali Treibitz and Yoav Y Schechner. Active polarization scattering. IEEE transactions on pattern analysis and machine intelligence, 31(3):385-399, 2008. 2", + "[47] Li Wang, Yang Ju, Heping Xie, Guowei Ma, Lingtao Mao, and Kexin He. The mechanical and photoelastic properties of 3d printable stress-visualized materials. Scientific reports, 7(1):10918, 2017. 6", + "[48] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7", + "[49] Aun Zaidi, Noah A Rubin, Maryna L Meretska, Lisa W Li, Ahmed H Dorrah, Joon-Suh Park, and Federico Capasso. Metasurface-enabled single-shot and complete mueller matrix imaging. Nature Photonics, pages 1-9, 2024. 1, 2" + ], + "bbox": [ + 514, + 92, + 905, + 501 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_model.json b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6d48d2fa1c6cf914865612b05065281f73d65ddf --- /dev/null +++ b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_model.json @@ -0,0 +1,2519 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.826, + 0.154 + ], + "angle": 0, + "content": "Event Ellipsometer: Event-based Mueller-Matrix Video Imaging" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.18, + 0.744, + 0.225 + ], + "angle": 0, + "content": "Ryota Maeda\\(^{1,2}\\) Yunseong Moon\\(^{1}\\) Seung-Hwan Baek\\(^{1}\\) \n\\(^{1}\\)POSTECH \\(^{2}\\)University of Hyogo" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.258, + 0.327, + 0.274 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.532 + ], + "angle": 0, + "content": "Light-matter interactions modify both the intensity and polarization state of light. Changes in polarization, represented by a Mueller matrix, encode detailed scene information. Existing optical ellipsometers capture Mueller-matrix images; however, they are often limited to capturing static scenes due to long acquisition times. Here, we introduce Event Ellipsometer, a method for acquiring a Mueller-matrix video for dynamic scenes. Our imaging system employs fast-rotating quarter-wave plates (QWPs) in front of a light source and an event camera that asynchronously captures intensity changes induced by the rotating QWPs. We develop an ellipsometric-event image formation model, a calibration method, and an ellipsometric-event reconstruction method. We experimentally demonstrate that Event Ellipsometer enables Mueller-matrix video imaging at 30 fps, extending ellipsometry to dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.222, + 0.573 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Polarization describes the oscillation of the electric field in light waves, encoding valuable information about the scenes with which light interacts. The polarization state of light can be represented as a Stokes vector \\(\\mathbf{s} \\in \\mathbb{R}^{4 \\times 1}\\) [17]. Polarimetric image analysis focuses on capturing complete or partial forms of the per-pixel Stokes vector and has been extensively studied for shape-from-polarization [31, 36], diffuse-specular separation [19], reflection removal [41], seeing through scattering [42], and transparent object segmentation [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ellipsometry advances polarimetric imaging by analyzing the polarization state of light captured under varying polarization states of illumination. This allows for acquiring polarimetric reflectance, represented as a Mueller matrix \\(\\mathbf{M} \\in \\mathbb{R}^{4 \\times 4}\\) [9], which comprehensively characterizes how light-matter interactions alter the polarization state of incident light. Ellipsometry has been widely used in material science [30, 38], biology [1, 23], and has recently gained attention in computer vision and graphics for 3D shape and reflection analysis [6, 7, 43], material acquisition [8], light transport decomposition [5, 35], and photoelasticity analy-" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.258, + 0.907, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.383, + 0.843, + 0.396 + ], + "angle": 0, + "content": "(a) Overview diagram of Event Ellipsometer" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.399, + 0.736, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.536, + 0.702, + 0.547 + ], + "angle": 0, + "content": "(b) Capture configuration" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.399, + 0.822, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.535, + 0.898, + 0.556 + ], + "angle": 0, + "content": "HDR imaging Tape detection (c) Applications" + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.399, + 0.905, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.535, + 0.898, + 0.555 + ], + "angle": 0, + "content": "Tape detection cations" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Figure 1. Overview of Event Ellipsometer. (a) Our imaging system captures the Mueller matrix at 30 fps from event streams induced by the continuously rotating QWPs. (b) With our experimental prototype, (c) we demonstrate ellipsometric analysis for dynamic scenes and various applications." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.653, + 0.571, + 0.668 + ], + "angle": 0, + "content": "sis [18]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.671, + 0.907, + 0.883 + ], + "angle": 0, + "content": "Despite the rich information provided by ellipsometry, its applications have been mainly limited to static scenes because conventional methods require capturing multiple images while mechanically rotating polarizing optics. Typically, more than 20 rotation angles [8] are necessary, leading to long acquisition times and making it impractical for capturing dynamic scenes. Although single-shot Mueller matrix imaging techniques using conventional sensors exist [12, 33, 49], they assume planar target scenes at fixed distances, significantly sacrifice sensor resolution, or use custom-fabricated nano- or micro-optical elements. Additionally, conventional intensity sensors are limited in capturing high-dynamic range (HDR) scenes, often requiring the capture of additional images with multiple exposures." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present Event Ellipsometer, a Mueller-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21804" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "matrix imaging method capable of capturing dynamic and HDR scenes. Departing from using conventional intensity sensors, we utilize an event camera, which asynchronously records intensity changes. We equip fast-rotating QWPs and linear polarizers (LPs) in front of both the event camera and an LED light source. The rotating QWPs modulate the polarization state of light emitted and received by the imaging system, resulting in intensity changes captured as events." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.348 + ], + "angle": 0, + "content": "We develop an ellipsometric-event image formation model that relates the time differences of adjacent events to the Mueller matrix, normalized by its first element. Using this model, we propose a two-stage Mueller-matrix reconstruction method consisting of per-pixel estimation and spatiotemporal propagation. We incorporate physical validity constraints to handle outliers from sensor noise and scene motion. We also devise calibration methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.482, + 0.469 + ], + "angle": 0, + "content": "Experimentally, we demonstrate Mueller-matrix imaging at 30 fps, achieving a mean-squared error of 0.045 for materials with known Mueller matrices. Unlike previous single-shot methods [49], Event Ellipsometer can capture non-planar objects and does not compromise spatial resolution. Furthermore, our method is capable of capturing HDR scenes without the need for additional measurement with different exposures." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.47, + 0.411, + 0.484 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.485, + 0.481, + 0.514 + ], + "angle": 0, + "content": "- We propose Event Ellipsometer, enabling Mueller-matrix imaging for dynamic scenes at 30 fps." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.515, + 0.481, + 0.559 + ], + "angle": 0, + "content": "- We develop an imaging system using an event camera and a light source, each equipped with synchronized fast-rotating QWPs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.56, + 0.481, + 0.62 + ], + "angle": 0, + "content": "- We formulate an ellipsometric-event image formation model relating event streams to the Mueller matrix, along with a calibration method and a robust ellipsometric-event reconstruction algorithm." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.621, + 0.481, + 0.696 + ], + "angle": 0, + "content": "- We experimentally show the high accuracy of Event Ellipsometer on samples with known ground truth, and demonstrate ellipsometric analysis of dynamic scenes with applications on photoelasticity, human capture, HDR imaging, and tape detection." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.485, + 0.481, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.233, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Imaging with Polarized Illumination Polarimetric imaging with polarized illumination exploits polarization-dependent light transport to extract material and geometric scene properties. Previous methods often capture scenes with single or a few polarized illuminations and fit parametric models to the captured images [14, 27, 28, 46]. However, this approach is limited by the representation power of the parametric models, thus cannot reveal the true polarimetric reflectance of real-world scenes. Ellipsometry extends polarimetric imaging by directly capturing the Mueller matrix, providing a comprehensive characteriza" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "tion of material polarization properties [4, 17, 20]. It has applications in material science [30, 38], biology [1, 23], and has recently been applied to computer vision and graphics for tasks such as 3D shape analysis [6, 7, 43], material acquisition [8], and light transport decomposition [5, 35]. However, conventional ellipsometry techniques are unsuitable for dynamic scenes because they require capturing multiple images while mechanically rotating polarizing optics and capturing each with an intensity sensor, leading to long acquisition. Single-shot Mueller-matrix imaging methods have been proposed [12, 33, 49], however these approaches often compromise spatial resolution and are limited to planar scenes at fixed distances, restricting their applicability to general dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.905, + 0.552 + ], + "angle": 0, + "content": "Event Cameras for Photometric Analysis Event cameras offer a high temporal resolution of microseconds and a HDR compared to standard frame-based intensity cameras [21]. These advantages have been leveraged for photometric analysis by combining continuously modulated optical light sources or filters. For instance, intensity-modulated illumination with an event camera enables radiance estimation [15, 24] and bispectral photometry [45]. Yu et al. [48] demonstrated event-based photometric stereo for normal estimation with a mechanically-rotating point light source. Hawks et al. [25] and Muglikar et al. [37] explored passive linear-polarization imaging by rotating a linear polarizer in front of an event camera. Existing event-based vision methods cannot capture full polarization reflectance properties as a Mueller-matrix image, limiting their utility." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.905, + 0.708 + ], + "angle": 0, + "content": "Imaging with Rotating Optical Elements Rotating optical components have been used for HDR and multispectral imaging [40] and high-speed video reconstruction [13] using a conventional intensity camera. Recently, rotating optics with event cameras have been demonstrated to leverage asynchronous operation and high temporal resolution of the event cameras [25, 26, 37]. These methods rotate optical elements only in front of an event camera without illumination modulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.673, + 0.741 + ], + "angle": 0, + "content": "3. Imaging System" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Experimental Prototype We build an imaging system using an LED light source (Thorlabs MCWHLP3), fast-rotating QWPs (Edmund Optics WP140HE), fixed LPs (Thorlabs WP25M-VIS), and an event camera (Prophesee EVK4). Figure 2(a) and (c) depict our setup, where a pair of a QWP and an LP is placed in front of the event camera and the light source, respectively. We rotate the QWPs so that the event camera detects the event streams caused by the rotating QWPs, from which the normalized Mueller matrix is reconstructed. We can configure the illumination" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21805" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.208, + 0.356, + 0.22 + ], + "angle": 0, + "content": "(a) Schematic diagram" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.221, + 0.482, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.394, + 0.451, + 0.407 + ], + "angle": 0, + "content": "(b) Timeline of QWPs rotation and event measurement" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.803, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.256, + 0.776, + 0.268 + ], + "angle": 0, + "content": "(c) Hardware prototype" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.089, + 0.905, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.248, + 0.885, + 0.26 + ], + "angle": 0, + "content": "QWP film" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.272, + 0.695, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.395, + 0.65, + 0.407 + ], + "angle": 0, + "content": "(d) Reflection mode" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.272, + 0.905, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.395, + 0.868, + 0.407 + ], + "angle": 0, + "content": "(e) Transmission mode" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.421, + 0.907, + 0.464 + ], + "angle": 0, + "content": "Figure 2. Imaging system of Event Ellipsometer. (a) Schematic diagram illustrating the optical arrangement and hardware operation. (b) Timeline showing the rotation of two QWPs and the event measurement. (c) Our hardware prototype. The system can move the light source and camera position for use in (d) Reflection mode or (e) Transmission mode." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.483, + 0.563 + ], + "angle": 0, + "content": "and the camera modules mounted on different breadboards in both the reflection model and the transmission mode as shown in Figure 2(d) and (e). Our system can be seen as a combination of the optical dual rotating retarder [3] and the event-based vision. For a complete list of components, refer to the Supplemental Material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.853 + ], + "angle": 0, + "content": "Rotating QWPs and Angle Encoder We rotate the QWP on the camera side five times faster than the one on the light source side [44]. We denote the QWP angles of the light source and the camera as \\(\\theta_{1,t} = \\omega t + i_1, \\theta_{2,t} = 5\\omega t + i_2\\) where \\(\\omega = 30\\pi\\) rad/sec is the angular velocity of the motor driving the light-source QWP. \\(i_1\\) and \\(i_2\\) are the initial QWP angles. One challenge here is to rotate the QWPs at such high speeds while recording their angles \\(\\theta_{1,t}\\) and \\(\\theta_{2,t}\\) over time \\(t\\). To this end, we use two independently controlled brushed direct current (DC) motors rotating the QWP films, resulting in an affordable configuration. Also, we develop custom angle encoders using a 3D-printed occluder, a photointerrupter, and an Arduino microcontroller. The occluder rotates at the same speed as the QWP and blocks the light path within the photointerrupter at every \\(\\pi\\) rotation. The microcontroller detects such change at a microsecond resolution and emits a hardware trigger to the event camera. Figure 2(c) shows the angle encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Frame of Mueller Matrix Given the rotation speed of \\(\\omega\\), every \\(\\pi\\) rotation of the light-source QWP takes \\(\\pi/\\omega \\approx\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.577 + ], + "angle": 0, + "content": "33 ms. We set this as the effective temporal duration of each frame \\( f \\) of the Mueller-matrix video which we aim to reconstruct. That is, as shown in Figure 2(b), each frame \\( f \\in \\{1, \\dots, F\\} \\) of the Mueller matrix is estimated based on the events measured during the temporal duration \\( \\pi / \\omega \\). \\( F \\) is the number of frames in a reconstructed Mueller-matrix video." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.686, + 0.608 + ], + "angle": 0, + "content": "4. Image Formation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.906, + 0.646 + ], + "angle": 0, + "content": "Here, we relate the event-camera measurements to the scene Mueller matrix which we aim to reconstruct." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.832 + ], + "angle": 0, + "content": "Polarimetric Modulation The LED light source emits unpolarized light of Stokes vector \\(\\mathbf{s} = [1,0,0,0]^{\\intercal}\\). At a time \\(t\\), the light passes through the LP and the QWP rotated by angle \\(\\theta_{1,t}\\), resulting in the Stokes vector \\(\\mathbf{Q}(\\theta_{1,t})\\mathbf{L}(0)\\mathbf{s}\\), where \\(\\mathbf{Q}(\\theta_{1,t})\\) and \\(\\mathbf{L}(0)\\) denote the Mueller matrices of a QWP rotated by \\(\\theta_{1,t}\\) and an LP at \\(0^{\\circ}\\), respectively. The light then interacts with the target scene, undergoing polarization change represented by the scene Mueller matrix \\(\\mathbf{M}\\). After interaction, the light passes through another QWP at angle \\(\\theta_{2,t}\\) and an LP in front of the event camera, resulting in the time-varying intensity \\(I_{t}\\) incident on the sensor:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.843, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\nI _ {t} = [ \\mathbf {L} (0) \\mathbf {Q} (\\theta_ {2, t}) \\mathbf {M} \\mathbf {Q} (\\theta_ {1, t}) \\mathbf {L} (0) \\mathbf {s} ] _ {0}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "where \\(\\left[\\cdot\\right]_{0}\\) denotes intensity, the first element of the Stokes vector." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "We rearrange Equation (1) and derive a matrix-vector form:" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.123, + 0.482, + 0.14 + ], + "angle": 0, + "content": "\\[\nI _ {t} = \\mathbf {A} _ {t} \\hat {\\mathbf {M}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.148, + 0.483, + 0.194 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{M}} = [\\mathbf{M}_{00},\\mathbf{M}_{01},\\dots ,\\mathbf{M}_{33}]^{\\intercal}\\in \\mathbb{R}^{16\\times 1}\\) is the vectorized form of \\(\\mathbf{M}\\) and \\(\\mathbf{A}_t\\in \\mathbb{R}^{1\\times 16}\\) is the system matrix defined as" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.206, + 0.449, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {A} _ {t} = [ 1, \\alpha_ {1} ^ {2}, \\alpha_ {1} \\alpha_ {2}, \\alpha_ {2}, \\alpha_ {3} ^ {2}, \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2}, \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2}, \\alpha_ {2} \\alpha_ {3} ^ {2}, \\\\ \\alpha_ {3} \\alpha_ {4}, \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4}, \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, - \\alpha_ {4} \\\\ \\left. - \\alpha_ {1} ^ {2} \\alpha_ {4}, - \\alpha_ {1} \\alpha_ {2} \\alpha_ {4}, - \\alpha_ {2} \\alpha_ {4} \\right], \\text {w h e r e} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.269, + 0.425, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {1} = \\cos (2 i _ {1} + 2 \\omega t), \\alpha_ {2} = \\sin (2 i _ {1} + 2 \\omega t),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.288, + 0.482, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {3} = \\cos (2 i _ {2} + 1 0 \\omega t), \\alpha_ {4} = \\sin (2 i _ {2} + 1 0 \\omega t). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.482, + 0.332 + ], + "angle": 0, + "content": "For the full derivation, refer to the Supplemental Document." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.485, + 0.415 + ], + "angle": 0, + "content": "Differential Intensity The event camera triggers events based on the temporal change of the logarithm of photocurrent [21]. Analytically differentiating the logarithm of Equation (2) with respect to time \\( t \\), we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.428, + 0.483, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\log I _ {t}}{\\partial t} = \\frac {\\frac {\\partial I _ {t}}{\\partial t}}{I _ {t}} = \\frac {\\frac {\\partial \\mathbf {A} _ {t}}{\\partial t} \\hat {\\mathbf {M}}}{\\mathbf {A} _ {t} \\hat {\\mathbf {M}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.477, + 0.241, + 0.496 + ], + "angle": 0, + "content": "where \\(\\frac{\\partial\\mathbf{A}_t}{\\partial t}\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.507, + 0.492, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial \\mathbf {A} _ {t}}{\\partial t} = [ 0, - 4 \\alpha_ {1} \\alpha_ {2}, 2 \\alpha_ {1} ^ {2} - 2 \\alpha_ {2} ^ {2}, 2 \\alpha_ {1}, - 2 0 \\alpha_ {3} \\alpha_ {4}, \\\\ - 2 0 \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4} - 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2}, 2 \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2} - 2 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4} - 2 \\alpha_ {2} ^ {2} \\alpha_ {3} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.559, + 0.329, + 0.578 + ], + "angle": 0, + "content": "\\[\n2 \\alpha_ {1} \\alpha_ {3} ^ {2} - 2 0 \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, 1 0 \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {4} ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.579, + 0.511, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} 1 0 \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {1} ^ {2} \\alpha_ {4} ^ {2} - 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, \\\\ 2 \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4} + 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {4} ^ {2} - 2 \\alpha_ {2} ^ {2} \\alpha_ {3} \\alpha_ {4}, \\\\ 2 \\alpha_ {1} \\alpha_ {3} \\alpha_ {4} + 1 0 \\alpha_ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {2} \\alpha_ {4} ^ {2}, - 1 0 \\alpha_ {3}, - 1 0 \\alpha_ {1} ^ {2} \\alpha_ {3} + 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {4}, \\\\ - 2 \\alpha_ {1} ^ {2} \\alpha_ {4} - 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} + 2 \\alpha_ {2} ^ {2} \\alpha_ {4}, - 2 \\alpha_ {1} \\alpha_ {4} - 1 0 \\alpha_ {2} \\alpha_ {3} ]. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.8 + ], + "angle": 0, + "content": "Events from Polarimetric Modulation For a pixel, we denote the time difference between consecutive events as \\(\\Delta t_{k}\\), where \\(k\\in \\{1,\\dots ,K\\}\\) is the event index in each frame \\(f\\). \\(K\\) is the number of detected events at that pixel in the frame \\(f\\). The time difference is known to be related to the change of photocurrent according to Taylor's expansion [21]:" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.8, + 0.482, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\log I _ {t _ {k}}}{\\partial t} = \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.482, + 0.87 + ], + "angle": 0, + "content": "where event polarity \\( p_k \\in \\{+1, -1\\} \\) is the sign of intensity change, and \\( C \\) is a constant threshold of the event camera." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "By combining Equations (4) and (6), we relate the measured per-pixel events to the Mueller matrix, resulting in the" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.09, + 0.905, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.175, + 0.905, + 0.217 + ], + "angle": 0, + "content": "Figure 3. Overview of our Mueller-matrix reconstruction pipeline. This method consists of two steps: (1) per-pixel reconstruction and (2) propagation and refinement." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.676, + 0.246 + ], + "angle": 0, + "content": "image formation model:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.256, + 0.905, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {p _ {k} C}{\\Delta t _ {k}} = \\frac {\\frac {\\partial \\mathbf {A} _ {t _ {k}}}{\\partial t _ {k}} \\hat {\\mathbf {M}}}{\\mathbf {A} _ {t _ {k}} \\hat {\\mathbf {M}}} \\\\ \\Rightarrow \\left(\\frac {\\partial \\mathbf {A} _ {t _ {k}}}{\\partial t _ {k}} - \\frac {p _ {k} C}{\\Delta t _ {k}} \\mathbf {A} _ {t _ {k}}\\right) \\hat {\\mathbf {M}} = 0 \\\\ \\Rightarrow \\mathbf {B} _ {t _ {k}} \\dot {\\mathbf {M}} = 0, \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.363, + 0.789, + 0.38 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_{t_k} \\in \\mathbb{R}^{1 \\times 16}\\) is the system matrix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.392, + 0.666, + 0.407 + ], + "angle": 0, + "content": "5. Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.905, + 0.493 + ], + "angle": 0, + "content": "The image formation model in Equation (7) allows us to reconstruct the Mueller matrix \\(\\hat{\\mathbf{M}}\\). Our reconstruction method, depicted in Figure 3, consists of two main steps: (1) per-pixel reconstruction and (2) spatio-temporal propagation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.501, + 0.741, + 0.517 + ], + "angle": 0, + "content": "5.1. Per-pixel Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.525, + 0.905, + 0.568 + ], + "angle": 0, + "content": "In this step, our goal is to estimate a Mueller matrix in a computationally efficient manner and being robust to noisy events." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.905, + 0.616 + ], + "angle": 0, + "content": "First, we stack the vectors \\(\\mathbf{B}_{t_k}\\) for all \\(k\\), forming a matrix \\(\\mathbf{B} \\in \\mathbb{R}^{K \\times 16}\\). We then solve for the vectorized Mueller matrix \\(\\hat{\\mathbf{M}}\\) by minimizing the weighted least-squares problem:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.628, + 0.905, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\underset {\\hat {\\mathbf {M}}} {\\text {m i n i m i z e}} \\| \\mathbf {W D B M} \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.794 + ], + "angle": 0, + "content": "where \\(\\mathbf{W} = \\mathrm{diag}(\\mathbf{w})\\) is a diagonal matrix of weights \\(\\mathbf{w} = [w_1, \\ldots, w_K] \\in \\mathbb{R}^{1 \\times K}\\), initialized as ones: \\(w_k = 1\\). The matrix \\(\\mathbf{D} = \\mathrm{diag}([\\Delta t_1, \\ldots, \\Delta t_K])\\) is a weight matrix that adjusts the influence of each event according to sampling density. This weighting ensures that more densely sampled events at small \\(\\Delta t\\) do not dominate the model fitting. We solve Equation (8) using singular value decomposition (SVD) to obtain \\(\\hat{\\mathbf{M}}\\) while avoiding the trivial zero solution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Second, we filter the reconstructed Mueller matrix \\(\\hat{\\mathbf{M}}\\) to ensure physical validity. Sensor noise and rapid scene dynamics can make solving Equation (8) challenging, often resulting in physically invalid Mueller matrices. These invalid matrices can trap the optimization in local minima. To address this, we apply Cloude's Mueller matrix filtering [16] to project the estimated matrix onto the space of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.91, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.335, + 0.908, + 0.392 + ], + "angle": 0, + "content": "Figure 4. Synthetic data evaluation result. (a) The rendered images include two materials: blue silicone and brass. (b) The plot shows the error (mean absolute error) of the reconstructed Mueller matrix over the number of iterations. (c) Ground truth Mueller matrix. (d)&(e) Top: the reconstructed Mueller-matrix images for the SVD initialization and the full-stage, respectively. Pixels with insufficient event counts are visualized in white. Bottom: a plot of the differentiation of log intensity and fitted line with our method." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.408, + 0.319, + 0.423 + ], + "angle": 0, + "content": "physically valid Mueller matrices:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.435, + 0.482, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} \\leftarrow \\operatorname {C l o u d e F i l t e r} (\\hat {\\mathbf {M}}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.435, + 0.478 + ], + "angle": 0, + "content": "where CloudeFilter \\((\\cdot)\\) denotes the filtering function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.509 + ], + "angle": 0, + "content": "Third, we recalculate the weight vector \\(\\mathbf{w}\\) using the physically valid Mueller matrix \\(\\hat{\\mathbf{M}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.521, + 0.482, + 0.539 + ], + "angle": 0, + "content": "\\[\nw _ {k} = 1 / \\max \\left(\\left| \\mathbf {B} _ {t _ {k}} \\hat {\\mathbf {M}} \\right|, \\epsilon\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.594 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a small positive constant to avoid division by zero sets as \\(10^{-6}\\) in our experiments. This weight reduces the influence of outlier events and enables robust estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.686 + ], + "angle": 0, + "content": "We iterate these three steps—solving the weighted least-squares problem, applying physical validity filtering, and recomputing the weights—until convergence, resulting in the Mueller matrix estimate \\(\\hat{\\mathbf{M}}\\) for each pixel \\((x,y)\\) and frame \\(f\\). We set the maximum number of iterations to 5 in our experiments" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.695, + 0.477, + 0.711 + ], + "angle": 0, + "content": "5.2. Spatio-temporal Propagation and Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Here, we start by denoting the estimated Mueller matrix at a pixel \\((x,y)\\) and frame \\(f\\) as \\(\\hat{\\mathbf{M}}_{xyf}\\) and the corresponding system matrix as \\(\\mathbf{B}_{xyf}\\), and weight matrix as \\(\\mathbf{D}_{xyz}\\). This second stage refines the Mueller matrix \\(\\hat{\\mathbf{M}}_{xyf}\\) estimated from the first stage using spatio-temporal neighboring pixels. This process is inspired by PatchMatchStereo from stereoscopic depth estimation [10, 11, 22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "First, we perform propagation. We consider neighboring pixels in both spatial and temporal axes, denoted by \\((x', y', f') \\in \\mathcal{N}(x, y, f)\\). The set \\(\\mathcal{N}\\) defines a spatio-temporal neighborhood extended from the red-black spatial-only neighborhood [22]. For each pixel at a frame" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.408, + 0.907, + 0.439 + ], + "angle": 0, + "content": "\\((x,y,f)\\), we update its Mueller matrix if a neighboring pixel \\((x',y',f')\\) offers a lower error:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.447, + 0.655, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {x y f} \\leftarrow \\hat {\\mathbf {M}} _ {x ^ {\\prime} y ^ {\\prime} f ^ {\\prime}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.468, + 0.905, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\text {i f} \\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x y f} \\| _ {1} > \\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x ^ {\\prime} y ^ {\\prime} f ^ {\\prime}} \\| _ {1}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.538 + ], + "angle": 0, + "content": "Next, we refine the Mueller matrix by applying random perturbations and accepting the perturbation if it reduces the error:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.545, + 0.697, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {x y f} \\leftarrow \\hat {\\mathbf {M}} _ {x y f} ^ {\\text {p e r t u r b}}, \\text {w h e r e}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.568, + 0.905, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {x y f} ^ {\\text {p e r t u r b}} = \\operatorname {C l o u d e F i l t e r} \\left(\\hat {\\mathbf {M}} _ {x y f} \\odot (\\mathbf {1} + \\sigma \\mathbf {N})\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.591, + 0.905, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x y f} \\right\\| _ {1} > \\left\\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\mathbf {M} _ {x y f} ^ {\\text {p e r t u r b}} \\right\\| _ {1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.906, + 0.758 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{M}}_{xyf}^{\\mathrm{perturb}}\\) is the perturbed Mueller matrix, \\(\\odot\\) denotes element-wise multiplication, \\(\\mathbf{1}\\) is an all-ones matrix, \\(\\mathbf{N} \\in \\mathbb{R}^{4 \\times 4}\\) contains random values from a standard normal distribution, and \\(\\sigma\\) is a scalar set to 0.01. This multiplicative perturbation preserves the relative scale of the Mueller matrix elements, allowing larger elements to undergo proportionally larger adjustments while smaller elements change minimally. We apply Cloude's filtering to the perturbed Mueller matrix to ensure physical validity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.906, + 0.805 + ], + "angle": 0, + "content": "We repeat the propagation and refinement steps iteratively for a fixed number of iterations set to 10 in our experiments to obtain the final Mueller matrix \\(\\hat{\\mathbf{M}}_{xyf}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.637, + 0.831 + ], + "angle": 0, + "content": "6. Calibration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.9 + ], + "angle": 0, + "content": "We perform one-time calibration of the system parameters: contrast threshold \\((C)\\) and QWP offset angles \\((i_1\\) and \\(i_2)\\). Calibration details are provided in the Supplementary Material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.549, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.089, + 0.906, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.276, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Figure 5. Assessment of reconstructed Mueller matrix on real data. (a) Evaluation with known optical elements. We show the corresponding mean squared errors (MSEs). (b) Measurement on an in-the-wild metal plate induces strong diagonal components." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.33, + 0.482, + 0.437 + ], + "angle": 0, + "content": "Contrast Threshold We calibrate the per-pixel contrast threshold \\(C\\) of an event sensor by analyzing pixel responses under controlled illumination stimuli. We increase and decrease the LED intensity linearly over time, and the event camera directly captures the LED light, generating events. We then fit the contrast threshold \\(C\\) to the events per each pixel using Equation (6)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.453, + 0.483, + 0.545 + ], + "angle": 0, + "content": "QWP Offset Angle We obtain the offset angles of the fast axis of the QWP in our system: \\( i_1 \\) and \\( i_2 \\) by capturing another QWP sample with known fast axis. Given the known Mueller matrix of the reference QWP, we find the best offset angles that minimize the reconstruction error of Equation (6)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.179, + 0.573 + ], + "angle": 0, + "content": "7. Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Implementation We implemented our reconstruction algorithm in \\(\\mathrm{C + + }\\) using OpenMP CPU parallelization. It takes 62 seconds for per-pixel reconstruction (Section 5.1), and 68 seconds for the propagation and refinement processes (Section 5.2), measured for processing \\(500\\times 500\\) -resolution event image sequence of 30 frames, with an average event rate of \\(153\\mathrm{MEv / s}\\), on the AMD Ryzen 7 7800X3D 8-core processor." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Validation on Synthetic Data To evaluate our method, we render synthetic data in three steps. First, we mimic our capture system by replacing the event camera with an intensity camera in Mitsuba3 [29], and render an image sequence of a synthetic scene for each QWP angle with interval \\(\\omega t = 0.01^{\\circ}\\). We use the real-world polarimetric BRDFs to construct the synthetic scene [8]. Then, the rendered images are converted to event streams by simulating event-camera processing based on DVS-Voltmeter [34]. Last, we add Gaussian noise with a standard deviation of 0.5 and replace \\(5\\%\\) of true events with values sampled from a Gaussian distribution with a standard deviation of 5.0. Figure 4" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.406 + ], + "angle": 0, + "content": "shows the synthetic data and reconstructed Mueller matrix using our method. While initial SVD-based estimation suffers from noise with a reconstruction error of 0.11, our full reconstruction improves accuracy with a reconstruction error of 0.04." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.909, + 0.598 + ], + "angle": 0, + "content": "Validation on Real Data Figure 5(a) shows that our reconstruction closely matches the corresponding pseudo ground truth of real-world samples: air, linear polarizer \\((0^{\\circ}\\) and \\(45^{\\circ})\\), quarter-wave-plate \\((0^{\\circ}\\) and \\(45^{\\circ})\\). Figure 5(b) shows the result on an in-the-wild sample: a metal plate. The strong diagonal components in this result indicate the preservation of polarization, which is characteristic of reflections from metal surfaces. Notably, these Mueller matrix images are captured in just \\(33\\mathrm{ms}\\), significantly faster than the conventional ellipsometer, which requires several minutes due to its frame-based imaging principle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.909, + 0.758 + ], + "angle": 0, + "content": "Photoelasticity of Transparent Gelatine Photoelasticity is an optical property whereby dielectric materials exhibit birefringence under deformation. Imaging photoelasticity has applications in mechanical stress and material analysis [39, 47]. Analyzing Mueller-matrix images reveals such stress distributions even in transparent objects [18]. Our system enables observing photoelasticity at video rates. Figure 6 shows Mueller-matrix images of a gelatine disk, revealing stress-dependent fringe patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Transparent Tape Detection Detecting transparent objects is challenging for conventional cameras. Sticky tape, a common transparent material used for sealing cardboard, exhibits birefringent properties due to the molecular structure of its stretched plastic. Figure 7 shows the measurement result of sticky tape on a cardboard box. The reconstructed Mueller matrix shows the birefringent property, and we can clearly recognize the tape region. This" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.341, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.236, + 0.285, + 0.249 + ], + "angle": 0, + "content": "(a) Measurement setup" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.088, + 0.624, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.234, + 0.55, + 0.245 + ], + "angle": 0, + "content": "Before applying force" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.088, + 0.906, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.234, + 0.817, + 0.246 + ], + "angle": 0, + "content": "After applying force" + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.248, + 0.904, + 0.261 + ], + "angle": 0, + "content": "(b) Reconstructed Mueller matrix -1 +1" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 6. Photoelasticity analysis. (a) Experimental setup for measuring a gelatine disk in transmission mode. We gradually apply force for the gelatine disk to observe changes in photoelastic properties. (b) Reconstructed Mueller matrix images with different forces, revealing complex stress-dependent polarimetric patterns. Notably, as force increases, the Mueller matrix image shows denser fringe patterns." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.328, + 0.301, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.33, + 0.468, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.448, + 0.337, + 0.461 + ], + "angle": 0, + "content": "(a) Target object" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.462, + 0.238, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.556, + 0.238, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.647, + 0.212, + 0.66 + ], + "angle": 0, + "content": "(b) Comparison" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.462, + 0.483, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.647, + 0.462, + 0.658 + ], + "angle": 0, + "content": "(c) Reconstructed Mueller matrix" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.66, + 0.48, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.739 + ], + "angle": 0, + "content": "Figure 7. Transparent tape detection. (a) The target object is transparent sticky tape on a box. (b)&(c) The Reconstructed Mueller matrix reveals the tape region, which is difficult to see with a conventional RGB camera." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.801 + ], + "angle": 0, + "content": "demonstration shows the potential for inspecting dynamically moving sealed boxes in automated industrial processes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Dynamic Human Capture Ellipsometers, as a comprehensive polarization imaging technique for human capture, can reveal hidden properties of the human face and hair that are difficult to analyze in conventional imaging and non-ellipsometric polarization imaging [2, 19]. In Figure 8, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.907, + 0.453 + ], + "angle": 0, + "content": "demonstrate the Mueller-matrix reconstruction of dynamic facial expressions and hair movements, revealing intricate polarization properties. The reconstructed Mueller matrix for the face exhibits strong diagonal components associated with specular reflection, while the non-diagonal regions display weakly polarized reflection with a dependency on surface normal along the face edges. The hair result shows polarization property in the specular highlight regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.605 + ], + "angle": 0, + "content": "HDR Mueller-matrix Imaging Our event-based ellipsometer enables Mueller-matrix imaging of HDR scenes. Figure 9 shows reconstruction results on a scene with specular and dark diffuse reflections. While conventional ellipsometers require multiple exposures to prevent overexposure, our method captures the Mueller matrix in HDR scenes at 30 fps without the need for additional measurement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.627, + 0.643 + ], + "angle": 0, + "content": "8. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Normalized Mueller-Matrix Imaging Normalized Mueller matrix has been extensively used as it provides rich scene information [5, 7, 18]. Since event cameras only detect intensity changes, not intensity itself, our method also reconstructs a normalized version of Mueller matrix per pixel, \\(\\mathbf{M} / \\mathbf{M}(0,0)\\). We leave reconstructing full Mueller-matrix as a future work that might be accomplished using an event-intensity hybrid imaging system." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "End-to-end Real-time Pipeline Although our capture speed is fast, our current Mueller-matrix reconstruction relies solely on CPU acceleration, resulting in a non-realtime pipeline from capture to reconstruction. Implementing GPU acceleration could enable on-the-fly capture, reconstruction, and visualization [22, 48]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.91, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.362, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 8. Mueller matrix acquisition for capturing dynamic human (a) face and (b) hair, demonstrating the capture of diffuse and specular polarimetric responses." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.404, + 0.287, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.63, + 0.24, + 0.643 + ], + "angle": 0, + "content": "(a) Target object" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.404, + 0.484, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.63, + 0.482, + 0.642 + ], + "angle": 0, + "content": "(b) Reconstructed Mueller matrix" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.71 + ], + "angle": 0, + "content": "Figure 9. Mueller matrix measurement for a HDR scene. (a) The target scene contains regions with both strong specular reflection and black diffuse reflection. (b) Our proposed method achieves an accurate Mueller-matrix image for the HDR scene in just \\(33~\\mathrm{ms}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Event Bandwidth and Sensitivity The ideal setting of an event camera is to have a low contrast threshold and a short refractory period, to capture subtle intensity changes. However, this produces many events, exceeding the transmission bandwidth, resulting in lost or delayed events [21]. In contrast, when the contrast threshold is high, weakly polarized phenomena cannot be detected as events. Given this tradeoff, we defined a region of interest (ROI) in the event sensor and empirically configured the camera settings to find the best parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.907, + 0.482 + ], + "angle": 0, + "content": "DC Motor Synchronization For faster rotation speed and precise synchronization between two motors, it is a viable alternative to use brushless DC motors, also known as an electronically commutated (EC) motors. However, this increases system building costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Gradient Descent Gradient descent in an automatic differentiation framework offers an alternative approach for Mueller-matrix reconstruction. However, the number of events differs across pixels, which complicates the construction of a dense tensor for efficient computation on modern GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.634, + 0.652 + ], + "angle": 0, + "content": "9. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.786 + ], + "angle": 0, + "content": "We have introduced Event Ellipsometer, a Mueller-matrix video imaging method that combines an event camera and a light source with fast-rotating QWPs. Our image formation model, calibration, and reconstruction method enable robust Mueller-matrix imaging at 30 fps. We validate our method on synthetic and real data, and demonstrate Mueller-matrix imaging on photoelasticity, dynamic human hair and face capture, HDR imaging, and transparent tape detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532, 2022R1A6A1A03052954, RS-2024-00437866), and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21811" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Hans Arwin. Application of ellipsometry techniques to biological materials. Thin Solid Films, 519(9):2589-2592, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[3] RMA Azzam. Photopolarimetric measurement of the mueller matrix by fourier analysis of a single detected signal. Optics Letters, 2(6):148-150, 1978. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[4] Rasheed MA Azzam. Stokes-vector and mueller-matrix polarimetry. JOSA A, 33(7):1396-1408, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[5] Seung-Hwan Baek and Felix Heide. Polarimetric spatiotemporal light transport probing. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[6] Seung-Hwan Baek and Felix Heide. All-photon polarimetric time-of-flight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17876-17885, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.444 + ], + "angle": 0, + "content": "[7] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM Trans. Graph., 37(6):268, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[8] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph., 39(4):139, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.483, + 0.558 + ], + "angle": 0, + "content": "[9] Seung-Hwan Baek, Ramesh Raskar, Jinwei Ye, Akshit Dave, Achuta Kadambi, and Huajin Chen. Polarization-based visual computing. In ACM SIGGRAPH 2023 Courses, pages 1-1, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.561, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[10] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[11] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In *Bmvc*, pages 1-11, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[12] Qizhi Cao, Min Jiang, Chenling Jia, Siyue Jiang, Jing Zhang, Baoli Yao, Mingwu Jin, Edward Dehoog, Lian Duan, Huahua Wang, et al. Snapshot imaging mueller matrix polarimeter using modified savart polariscopes. Applied Optics, 62(8):2124-2129, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[13] Dorian Chan, Mark Sheinin, and Matthew O'Toole. Spincam: High-speed imaging via a rotating point-spread function. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10789-10799, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[14] Tongbo Chen, Hendrik PA Lensch, Christian Fuchs, and Hans-Peter Seidel. Polarization and phase-shifting for 3d scanning of translucent objects. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proceedings of the IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 14760-14770, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[16] Shane R Cloude. Conditions for the physical realizability of matrix operators in polarimetry. In *Polarization Considerations for Optical Systems II*, pages 177–187. SPIE, 1990. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[17] Edward Collett. Field guide to polarization. Spie Bellingham, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[18] Akshat Dave, Tianyi Zhang, Aaron Young, Ramesh Raskar, Wolfgang Heidrich, and Ashok Veeraraghavan. Nest: Neural stress tensor tomography by leveraging 3d photoelasticity. arXiv preprint arXiv:2406.10212, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[19] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.331, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[20] Hiroyuki Fujiwara. Spectroscopic ellipsometry: principles and applications. John Wiley & Sons, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.359, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[21] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[22] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE international conference on computer vision, pages 873-881, 2015. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[23] Nirmalya Ghosh and I Alex Vitkin. Tissue polarimetry: concepts, challenges, applications, and outlook. Journal of biomedical optics, 16(11):110801-110801, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[24] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20616-20625, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[25] Michael Hawks and Michael Dexter. Event-based imaging polarimeter. Optical Engineering, 61(5):053101-053101, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[26] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[27] Inseung Hwang, Daniel S Jeon, Adolfo Munoz, Diego Gutierrez, Xin Tong, and Min H Kim. Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[28] Tomoki Ichikawa, Shohei Nobuhara, and Ko Nishino. Spiders: Structured polarization for invisible depth and reflectance sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25077-25085, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsubishi 3 renderer, 2022. https://mitsuba-renderer.org. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21812" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[30] GE Jellison, Donovan N Leonard, Lawrence M Anovitz, Chad M Parish, Eliot D Specht, and TM Rosseel. Crystallographic orientation of uniaxial calcite and dolomite determined using reflection generalized ellipsometry. Journal of Applied Physics, 124(22), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[31] Achuta Kadambi, Vage Taamazyan, Boxin Shi, and Ramesh Raskar. Polarized 3d: High-quality depth sensing with polarization cues. In Proceedings of the IEEE international conference on computer vision, pages 3370-3378, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[32] Agastya Kalra, Vage Taamazyan, Supreeth Krishna Rao, Kartik Venkataraman, Ramesh Raskar, and Achuta Kadambi. Deep polarization cues for transparent object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8602-8611, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[33] Michael W Kudenov, Michael J Escuti, Nathan Hagen, Eustace L Dereniak, and Kazuhiko Oka. Snapshot imaging mueller matrix polarimeter using polarization gratings. Optics letters, 37(8):1367-1369, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[34] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In European Conference on Computer Vision, pages 578-593. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[35] Ryota Maeda and Shinsaku Hiura. Polarimetric light transport analysis for specular inter-reflection. IEEE Transactions on Computational Imaging, 10:876-887, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[36] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 982-987. IEEE, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.483, + 0.636 + ], + "angle": 0, + "content": "[38] David A Ramsey and Kenneth C Ludema. The influences of roughness on film thickness measurements by mueller matrix ellipsometry. Review of scientific instruments, 65(9):2874-2881, 1994. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.705 + ], + "angle": 0, + "content": "[39] Michele Scafidi, Giuseppe Pitarresi, Andrea Toscano, Giovanni Petrucci, Sabina Alessi, and Augusto Ajovalasit. Review of photoelastic image analysis applied to structural birefringent materials: glass and polymers. Optical Engineering, 54(8):081206-081206, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[40] Yoav Y Schechner and Shree K Nayar. Uncontrolled modulation imaging. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-II. IEEE, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[41] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, pages 814-819. IEEE, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[42] Yoav Y Schechner, Srinivasa G Narasimhan, and Shree K Nayar. Instant dehazing of images using polarization. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I. IEEE, 2001. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.174 + ], + "angle": 0, + "content": "[43] Dominik Scheuble, Chenyang Lei, Seung-Hwan Baek, Mario Bijelic, and Felix Heide. Polarization wavefront lidar: Learning large scene reconstruction from polarized wavefronts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21241-21250, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.176, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[44] Matthew H Smith. Optimization of a dual-rotating-retarder mueller matrix polarimeter. Applied optics, 41(13):2488-2493, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.218, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[45] Tsuyoshi Takatani, Yuzuha Ito, Ayaka Ebisu, Yinqiang Zheng, and Takahito Aoto. Event-based bispectral photometry using temporally modulated illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15647, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.29, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[46] Tali Treibitz and Yoav Y Schechner. Active polarization scattering. IEEE transactions on pattern analysis and machine intelligence, 31(3):385-399, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.332, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[47] Li Wang, Yang Ju, Heping Xie, Guowei Ma, Lingtao Mao, and Kexin He. The mechanical and photoelastic properties of 3d printable stress-visualized materials. Scientific reports, 7(1):10918, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.389, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[48] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.446, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[49] Aun Zaidi, Noah A Rubin, Maryna L Meretska, Lisa W Li, Ahmed H Dorrah, Joon-Suh Park, and Federico Capasso. Metasurface-enabled single-shot and complete mueller matrix imaging. Nature Photonics, pages 1-9, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21813" + } + ] +] \ No newline at end of file diff --git a/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_origin.pdf b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..669b65ca6c1eb5139faf77bcb6e14b4adecec17a --- /dev/null +++ b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/00131384-06b2-4638-a55f-16332464dd84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21540f1d6f15bcd37c3f72a63b8dbae1ee51cfe88b6ea2cdc31e8ef0665de1f5 +size 7060372 diff --git a/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/full.md b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d84945e36072e5506685e684ca9584b983b1625 --- /dev/null +++ b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/full.md @@ -0,0 +1,381 @@ +# Event Ellipsometer: Event-based Mueller-Matrix Video Imaging + +Ryota Maeda $^{1,2}$ Yunseong Moon $^{1}$ Seung-Hwan Baek $^{1}$ $^{1}$ POSTECH $^{2}$ University of Hyogo + +# Abstract + +Light-matter interactions modify both the intensity and polarization state of light. Changes in polarization, represented by a Mueller matrix, encode detailed scene information. Existing optical ellipsometers capture Mueller-matrix images; however, they are often limited to capturing static scenes due to long acquisition times. Here, we introduce Event Ellipsometer, a method for acquiring a Mueller-matrix video for dynamic scenes. Our imaging system employs fast-rotating quarter-wave plates (QWPs) in front of a light source and an event camera that asynchronously captures intensity changes induced by the rotating QWPs. We develop an ellipsometric-event image formation model, a calibration method, and an ellipsometric-event reconstruction method. We experimentally demonstrate that Event Ellipsometer enables Mueller-matrix video imaging at 30 fps, extending ellipsometry to dynamic scenes. + +# 1. Introduction + +Polarization describes the oscillation of the electric field in light waves, encoding valuable information about the scenes with which light interacts. The polarization state of light can be represented as a Stokes vector $\mathbf{s} \in \mathbb{R}^{4 \times 1}$ [17]. Polarimetric image analysis focuses on capturing complete or partial forms of the per-pixel Stokes vector and has been extensively studied for shape-from-polarization [31, 36], diffuse-specular separation [19], reflection removal [41], seeing through scattering [42], and transparent object segmentation [32]. + +Ellipsometry advances polarimetric imaging by analyzing the polarization state of light captured under varying polarization states of illumination. This allows for acquiring polarimetric reflectance, represented as a Mueller matrix $\mathbf{M} \in \mathbb{R}^{4 \times 4}$ [9], which comprehensively characterizes how light-matter interactions alter the polarization state of incident light. Ellipsometry has been widely used in material science [30, 38], biology [1, 23], and has recently gained attention in computer vision and graphics for 3D shape and reflection analysis [6, 7, 43], material acquisition [8], light transport decomposition [5, 35], and photoelasticity analy- + +![](images/75b8dfbe7a8dcc63c0d043df6f7ac2b727225229cd7f24508b188991ab2a1416.jpg) +(a) Overview diagram of Event Ellipsometer + +![](images/b3a693dcb625d5226236bd6a6200934911058c7dbfb4bf65c27a425afa108ee1.jpg) +(b) Capture configuration + +![](images/e9a0c624fe9ac22f84299a4f97bfea94be6a5de132f22f80ec54d2a184626339.jpg) +HDR imaging Tape detection (c) Applications +Figure 1. Overview of Event Ellipsometer. (a) Our imaging system captures the Mueller matrix at 30 fps from event streams induced by the continuously rotating QWPs. (b) With our experimental prototype, (c) we demonstrate ellipsometric analysis for dynamic scenes and various applications. + +![](images/86d21b84eee3a9365a10fb85e8e8957c9709facf300fb66ce9dd09ecac0a7c5b.jpg) + +Tape detection cations + +sis [18]. + +Despite the rich information provided by ellipsometry, its applications have been mainly limited to static scenes because conventional methods require capturing multiple images while mechanically rotating polarizing optics. Typically, more than 20 rotation angles [8] are necessary, leading to long acquisition times and making it impractical for capturing dynamic scenes. Although single-shot Mueller matrix imaging techniques using conventional sensors exist [12, 33, 49], they assume planar target scenes at fixed distances, significantly sacrifice sensor resolution, or use custom-fabricated nano- or micro-optical elements. Additionally, conventional intensity sensors are limited in capturing high-dynamic range (HDR) scenes, often requiring the capture of additional images with multiple exposures. + +In this paper, we present Event Ellipsometer, a Mueller- + +matrix imaging method capable of capturing dynamic and HDR scenes. Departing from using conventional intensity sensors, we utilize an event camera, which asynchronously records intensity changes. We equip fast-rotating QWPs and linear polarizers (LPs) in front of both the event camera and an LED light source. The rotating QWPs modulate the polarization state of light emitted and received by the imaging system, resulting in intensity changes captured as events. + +We develop an ellipsometric-event image formation model that relates the time differences of adjacent events to the Mueller matrix, normalized by its first element. Using this model, we propose a two-stage Mueller-matrix reconstruction method consisting of per-pixel estimation and spatiotemporal propagation. We incorporate physical validity constraints to handle outliers from sensor noise and scene motion. We also devise calibration methods. + +Experimentally, we demonstrate Mueller-matrix imaging at 30 fps, achieving a mean-squared error of 0.045 for materials with known Mueller matrices. Unlike previous single-shot methods [49], Event Ellipsometer can capture non-planar objects and does not compromise spatial resolution. Furthermore, our method is capable of capturing HDR scenes without the need for additional measurement with different exposures. + +In summary, our contributions are as follows: + +- We propose Event Ellipsometer, enabling Mueller-matrix imaging for dynamic scenes at 30 fps. +- We develop an imaging system using an event camera and a light source, each equipped with synchronized fast-rotating QWPs. +- We formulate an ellipsometric-event image formation model relating event streams to the Mueller matrix, along with a calibration method and a robust ellipsometric-event reconstruction algorithm. +- We experimentally show the high accuracy of Event Ellipsometer on samples with known ground truth, and demonstrate ellipsometric analysis of dynamic scenes with applications on photoelasticity, human capture, HDR imaging, and tape detection. + +# 2. Related Work + +Imaging with Polarized Illumination Polarimetric imaging with polarized illumination exploits polarization-dependent light transport to extract material and geometric scene properties. Previous methods often capture scenes with single or a few polarized illuminations and fit parametric models to the captured images [14, 27, 28, 46]. However, this approach is limited by the representation power of the parametric models, thus cannot reveal the true polarimetric reflectance of real-world scenes. Ellipsometry extends polarimetric imaging by directly capturing the Mueller matrix, providing a comprehensive characteriza + +tion of material polarization properties [4, 17, 20]. It has applications in material science [30, 38], biology [1, 23], and has recently been applied to computer vision and graphics for tasks such as 3D shape analysis [6, 7, 43], material acquisition [8], and light transport decomposition [5, 35]. However, conventional ellipsometry techniques are unsuitable for dynamic scenes because they require capturing multiple images while mechanically rotating polarizing optics and capturing each with an intensity sensor, leading to long acquisition. Single-shot Mueller-matrix imaging methods have been proposed [12, 33, 49], however these approaches often compromise spatial resolution and are limited to planar scenes at fixed distances, restricting their applicability to general dynamic scenes. + +Event Cameras for Photometric Analysis Event cameras offer a high temporal resolution of microseconds and a HDR compared to standard frame-based intensity cameras [21]. These advantages have been leveraged for photometric analysis by combining continuously modulated optical light sources or filters. For instance, intensity-modulated illumination with an event camera enables radiance estimation [15, 24] and bispectral photometry [45]. Yu et al. [48] demonstrated event-based photometric stereo for normal estimation with a mechanically-rotating point light source. Hawks et al. [25] and Muglikar et al. [37] explored passive linear-polarization imaging by rotating a linear polarizer in front of an event camera. Existing event-based vision methods cannot capture full polarization reflectance properties as a Mueller-matrix image, limiting their utility. + +Imaging with Rotating Optical Elements Rotating optical components have been used for HDR and multispectral imaging [40] and high-speed video reconstruction [13] using a conventional intensity camera. Recently, rotating optics with event cameras have been demonstrated to leverage asynchronous operation and high temporal resolution of the event cameras [25, 26, 37]. These methods rotate optical elements only in front of an event camera without illumination modulation. + +# 3. Imaging System + +Experimental Prototype We build an imaging system using an LED light source (Thorlabs MCWHLP3), fast-rotating QWPs (Edmund Optics WP140HE), fixed LPs (Thorlabs WP25M-VIS), and an event camera (Prophesee EVK4). Figure 2(a) and (c) depict our setup, where a pair of a QWP and an LP is placed in front of the event camera and the light source, respectively. We rotate the QWPs so that the event camera detects the event streams caused by the rotating QWPs, from which the normalized Mueller matrix is reconstructed. We can configure the illumination + +![](images/9c03fb221471db728db7ed2545b1af70a9754bf0e18549e23b37f33e4449ae44.jpg) +(a) Schematic diagram + +![](images/3a90a26a4009ce56153453a59a00e14c9a9bbb365c2d68d6d56726799edf44ab.jpg) +(b) Timeline of QWPs rotation and event measurement +Figure 2. Imaging system of Event Ellipsometer. (a) Schematic diagram illustrating the optical arrangement and hardware operation. (b) Timeline showing the rotation of two QWPs and the event measurement. (c) Our hardware prototype. The system can move the light source and camera position for use in (d) Reflection mode or (e) Transmission mode. + +![](images/87c6155bb1edeb05151d21244a090246f9840b39a24a34c24b4b811e7a017b59.jpg) +(c) Hardware prototype + +![](images/a15b567aef125c9a17b3b485302740ec647279efdc17e75189b98d01d6ff3f79.jpg) +QWP film + +![](images/e1b3059d4828ee7022366407b074de23b5c015aa5ade8be29b7aef105caa29e0.jpg) +(d) Reflection mode + +![](images/b3195701c7db6a0391f4fa754e1788354108206742ef8f5038b534b92144a4dc.jpg) +(e) Transmission mode + +and the camera modules mounted on different breadboards in both the reflection model and the transmission mode as shown in Figure 2(d) and (e). Our system can be seen as a combination of the optical dual rotating retarder [3] and the event-based vision. For a complete list of components, refer to the Supplemental Material. + +Rotating QWPs and Angle Encoder We rotate the QWP on the camera side five times faster than the one on the light source side [44]. We denote the QWP angles of the light source and the camera as $\theta_{1,t} = \omega t + i_1, \theta_{2,t} = 5\omega t + i_2$ where $\omega = 30\pi$ rad/sec is the angular velocity of the motor driving the light-source QWP. $i_1$ and $i_2$ are the initial QWP angles. One challenge here is to rotate the QWPs at such high speeds while recording their angles $\theta_{1,t}$ and $\theta_{2,t}$ over time $t$ . To this end, we use two independently controlled brushed direct current (DC) motors rotating the QWP films, resulting in an affordable configuration. Also, we develop custom angle encoders using a 3D-printed occluder, a photointerrupter, and an Arduino microcontroller. The occluder rotates at the same speed as the QWP and blocks the light path within the photointerrupter at every $\pi$ rotation. The microcontroller detects such change at a microsecond resolution and emits a hardware trigger to the event camera. Figure 2(c) shows the angle encoder. + +Frame of Mueller Matrix Given the rotation speed of $\omega$ , every $\pi$ rotation of the light-source QWP takes $\pi/\omega \approx$ + +33 ms. We set this as the effective temporal duration of each frame $f$ of the Mueller-matrix video which we aim to reconstruct. That is, as shown in Figure 2(b), each frame $f \in \{1, \dots, F\}$ of the Mueller matrix is estimated based on the events measured during the temporal duration $\pi / \omega$ . $F$ is the number of frames in a reconstructed Mueller-matrix video. + +# 4. Image Formation + +Here, we relate the event-camera measurements to the scene Mueller matrix which we aim to reconstruct. + +Polarimetric Modulation The LED light source emits unpolarized light of Stokes vector $\mathbf{s} = [1,0,0,0]^{\intercal}$ . At a time $t$ , the light passes through the LP and the QWP rotated by angle $\theta_{1,t}$ , resulting in the Stokes vector $\mathbf{Q}(\theta_{1,t})\mathbf{L}(0)\mathbf{s}$ , where $\mathbf{Q}(\theta_{1,t})$ and $\mathbf{L}(0)$ denote the Mueller matrices of a QWP rotated by $\theta_{1,t}$ and an LP at $0^{\circ}$ , respectively. The light then interacts with the target scene, undergoing polarization change represented by the scene Mueller matrix $\mathbf{M}$ . After interaction, the light passes through another QWP at angle $\theta_{2,t}$ and an LP in front of the event camera, resulting in the time-varying intensity $I_{t}$ incident on the sensor: + +$$ +I _ {t} = [ \mathbf {L} (0) \mathbf {Q} (\theta_ {2, t}) \mathbf {M} \mathbf {Q} (\theta_ {1, t}) \mathbf {L} (0) \mathbf {s} ] _ {0}, \tag {1} +$$ + +where $\left[\cdot\right]_{0}$ denotes intensity, the first element of the Stokes vector. + +We rearrange Equation (1) and derive a matrix-vector form: + +$$ +I _ {t} = \mathbf {A} _ {t} \hat {\mathbf {M}}, \tag {2} +$$ + +where $\hat{\mathbf{M}} = [\mathbf{M}_{00},\mathbf{M}_{01},\dots ,\mathbf{M}_{33}]^{\intercal}\in \mathbb{R}^{16\times 1}$ is the vectorized form of $\mathbf{M}$ and $\mathbf{A}_t\in \mathbb{R}^{1\times 16}$ is the system matrix defined as + +$$ +\begin{array}{l} \mathbf {A} _ {t} = [ 1, \alpha_ {1} ^ {2}, \alpha_ {1} \alpha_ {2}, \alpha_ {2}, \alpha_ {3} ^ {2}, \alpha_ {1} ^ {2} \alpha_ {3} ^ {2}, \alpha_ {1} \alpha_ {2} \alpha_ {3} ^ {2}, \alpha_ {2} \alpha_ {3} ^ {2}, \\ \alpha_ {3} \alpha_ {4}, \alpha_ {1} ^ {2} \alpha_ {3} \alpha_ {4}, \alpha_ {1} \alpha_ {2} \alpha_ {3} \alpha_ {4}, \alpha_ {2} \alpha_ {3} \alpha_ {4}, - \alpha_ {4} \\ \left. - \alpha_ {1} ^ {2} \alpha_ {4}, - \alpha_ {1} \alpha_ {2} \alpha_ {4}, - \alpha_ {2} \alpha_ {4} \right], \text {w h e r e} \\ \end{array} +$$ + +$$ +\alpha_ {1} = \cos (2 i _ {1} + 2 \omega t), \alpha_ {2} = \sin (2 i _ {1} + 2 \omega t), +$$ + +$$ +\alpha_ {3} = \cos (2 i _ {2} + 1 0 \omega t), \alpha_ {4} = \sin (2 i _ {2} + 1 0 \omega t). \tag {3} +$$ + +For the full derivation, refer to the Supplemental Document. + +Differential Intensity The event camera triggers events based on the temporal change of the logarithm of photocurrent [21]. Analytically differentiating the logarithm of Equation (2) with respect to time $t$ , we obtain + +$$ +\frac {\partial \log I _ {t}}{\partial t} = \frac {\frac {\partial I _ {t}}{\partial t}}{I _ {t}} = \frac {\frac {\partial \mathbf {A} _ {t}}{\partial t} \hat {\mathbf {M}}}{\mathbf {A} _ {t} \hat {\mathbf {M}}}, \tag {4} +$$ + +where $\frac{\partial\mathbf{A}_t}{\partial t}$ is given as + +$$ +\begin{array}{l} \frac {\partial \mathbf {A} _ {t}}{\partial t} = [ 0, - 4 \alpha_ {1} \alpha_ {2}, 2 \alpha_ {1} ^ {2} - 2 \alpha_ {2} ^ {2}, 2 \alpha_ {1}, - 2 0 \alpha_ {3} \alpha_ {4}, \\ - 2 0 \alpha_ {1} ^ {2} \alpha_ {3} \alpha_ {4} - 4 \alpha_ {1} \alpha_ {2} \alpha_ {3} ^ {2}, 2 \alpha_ {1} ^ {2} \alpha_ {3} ^ {2} - 2 0 \alpha_ {1} \alpha_ {2} \alpha_ {3} \alpha_ {4} - 2 \alpha_ {2} ^ {2} \alpha_ {3} ^ {2}, \\ \end{array} +$$ + +$$ +2 \alpha_ {1} \alpha_ {3} ^ {2} - 2 0 \alpha_ {2} \alpha_ {3} \alpha_ {4}, 1 0 \alpha_ {3} ^ {2} - 1 0 \alpha_ {4} ^ {2}, +$$ + +$$ +\begin{array}{l} 1 0 \alpha_ {1} ^ {2} \alpha_ {3} ^ {2} - 1 0 \alpha_ {1} ^ {2} \alpha_ {4} ^ {2} - 4 \alpha_ {1} \alpha_ {2} \alpha_ {3} \alpha_ {4}, \\ 2 \alpha_ {1} ^ {2} \alpha_ {3} \alpha_ {4} + 1 0 \alpha_ {1} \alpha_ {2} \alpha_ {3} ^ {2} - 1 0 \alpha_ {1} \alpha_ {2} \alpha_ {4} ^ {2} - 2 \alpha_ {2} ^ {2} \alpha_ {3} \alpha_ {4}, \\ 2 \alpha_ {1} \alpha_ {3} \alpha_ {4} + 1 0 \alpha_ {2} \alpha_ {3} ^ {2} - 1 0 \alpha_ {2} \alpha_ {4} ^ {2}, - 1 0 \alpha_ {3}, - 1 0 \alpha_ {1} ^ {2} \alpha_ {3} + 4 \alpha_ {1} \alpha_ {2} \alpha_ {4}, \\ - 2 \alpha_ {1} ^ {2} \alpha_ {4} - 1 0 \alpha_ {1} \alpha_ {2} \alpha_ {3} + 2 \alpha_ {2} ^ {2} \alpha_ {4}, - 2 \alpha_ {1} \alpha_ {4} - 1 0 \alpha_ {2} \alpha_ {3} ]. \tag {5} \\ \end{array} +$$ + +Events from Polarimetric Modulation For a pixel, we denote the time difference between consecutive events as $\Delta t_{k}$ , where $k\in \{1,\dots ,K\}$ is the event index in each frame $f$ . $K$ is the number of detected events at that pixel in the frame $f$ . The time difference is known to be related to the change of photocurrent according to Taylor's expansion [21]: + +$$ +\frac {\partial \log I _ {t _ {k}}}{\partial t} = \frac {p _ {k} C}{\Delta t _ {k}}, \tag {6} +$$ + +where event polarity $p_k \in \{+1, -1\}$ is the sign of intensity change, and $C$ is a constant threshold of the event camera. + +By combining Equations (4) and (6), we relate the measured per-pixel events to the Mueller matrix, resulting in the + +![](images/50bc1e749a9554ccd8774ced66bcaf79744b8f6c330feabd040c8c067a8d1c88.jpg) +Figure 3. Overview of our Mueller-matrix reconstruction pipeline. This method consists of two steps: (1) per-pixel reconstruction and (2) propagation and refinement. + +image formation model: + +$$ +\begin{array}{l} \frac {p _ {k} C}{\Delta t _ {k}} = \frac {\frac {\partial \mathbf {A} _ {t _ {k}}}{\partial t _ {k}} \hat {\mathbf {M}}}{\mathbf {A} _ {t _ {k}} \hat {\mathbf {M}}} \\ \Rightarrow \left(\frac {\partial \mathbf {A} _ {t _ {k}}}{\partial t _ {k}} - \frac {p _ {k} C}{\Delta t _ {k}} \mathbf {A} _ {t _ {k}}\right) \hat {\mathbf {M}} = 0 \\ \Rightarrow \mathbf {B} _ {t _ {k}} \dot {\mathbf {M}} = 0, \tag {7} \\ \end{array} +$$ + +where $\mathbf{B}_{t_k} \in \mathbb{R}^{1 \times 16}$ is the system matrix. + +# 5. Reconstruction + +The image formation model in Equation (7) allows us to reconstruct the Mueller matrix $\hat{\mathbf{M}}$ . Our reconstruction method, depicted in Figure 3, consists of two main steps: (1) per-pixel reconstruction and (2) spatio-temporal propagation. + +# 5.1. Per-pixel Reconstruction + +In this step, our goal is to estimate a Mueller matrix in a computationally efficient manner and being robust to noisy events. + +First, we stack the vectors $\mathbf{B}_{t_k}$ for all $k$ , forming a matrix $\mathbf{B} \in \mathbb{R}^{K \times 16}$ . We then solve for the vectorized Mueller matrix $\hat{\mathbf{M}}$ by minimizing the weighted least-squares problem: + +$$ +\underset {\hat {\mathbf {M}}} {\text {m i n i m i z e}} \| \mathbf {W D B M} \| _ {2} ^ {2}, \tag {8} +$$ + +where $\mathbf{W} = \mathrm{diag}(\mathbf{w})$ is a diagonal matrix of weights $\mathbf{w} = [w_1, \ldots, w_K] \in \mathbb{R}^{1 \times K}$ , initialized as ones: $w_k = 1$ . The matrix $\mathbf{D} = \mathrm{diag}([\Delta t_1, \ldots, \Delta t_K])$ is a weight matrix that adjusts the influence of each event according to sampling density. This weighting ensures that more densely sampled events at small $\Delta t$ do not dominate the model fitting. We solve Equation (8) using singular value decomposition (SVD) to obtain $\hat{\mathbf{M}}$ while avoiding the trivial zero solution. + +Second, we filter the reconstructed Mueller matrix $\hat{\mathbf{M}}$ to ensure physical validity. Sensor noise and rapid scene dynamics can make solving Equation (8) challenging, often resulting in physically invalid Mueller matrices. These invalid matrices can trap the optimization in local minima. To address this, we apply Cloude's Mueller matrix filtering [16] to project the estimated matrix onto the space of + +![](images/1d257cb698b40f9e6bad5b05ddc08e647a7d1ccc9d6dea2e23c78b5de3f71fa6.jpg) +Figure 4. Synthetic data evaluation result. (a) The rendered images include two materials: blue silicone and brass. (b) The plot shows the error (mean absolute error) of the reconstructed Mueller matrix over the number of iterations. (c) Ground truth Mueller matrix. (d)&(e) Top: the reconstructed Mueller-matrix images for the SVD initialization and the full-stage, respectively. Pixels with insufficient event counts are visualized in white. Bottom: a plot of the differentiation of log intensity and fitted line with our method. + +physically valid Mueller matrices: + +$$ +\hat {\mathbf {M}} \leftarrow \operatorname {C l o u d e F i l t e r} (\hat {\mathbf {M}}), \tag {9} +$$ + +where CloudeFilter $(\cdot)$ denotes the filtering function. + +Third, we recalculate the weight vector $\mathbf{w}$ using the physically valid Mueller matrix $\hat{\mathbf{M}}$ : + +$$ +w _ {k} = 1 / \max \left(\left| \mathbf {B} _ {t _ {k}} \hat {\mathbf {M}} \right|, \epsilon\right), \tag {10} +$$ + +where $\epsilon$ is a small positive constant to avoid division by zero sets as $10^{-6}$ in our experiments. This weight reduces the influence of outlier events and enables robust estimation. + +We iterate these three steps—solving the weighted least-squares problem, applying physical validity filtering, and recomputing the weights—until convergence, resulting in the Mueller matrix estimate $\hat{\mathbf{M}}$ for each pixel $(x,y)$ and frame $f$ . We set the maximum number of iterations to 5 in our experiments + +# 5.2. Spatio-temporal Propagation and Refinement + +Here, we start by denoting the estimated Mueller matrix at a pixel $(x,y)$ and frame $f$ as $\hat{\mathbf{M}}_{xyf}$ and the corresponding system matrix as $\mathbf{B}_{xyf}$ , and weight matrix as $\mathbf{D}_{xyz}$ . This second stage refines the Mueller matrix $\hat{\mathbf{M}}_{xyf}$ estimated from the first stage using spatio-temporal neighboring pixels. This process is inspired by PatchMatchStereo from stereoscopic depth estimation [10, 11, 22]. + +First, we perform propagation. We consider neighboring pixels in both spatial and temporal axes, denoted by $(x', y', f') \in \mathcal{N}(x, y, f)$ . The set $\mathcal{N}$ defines a spatio-temporal neighborhood extended from the red-black spatial-only neighborhood [22]. For each pixel at a frame + +$(x,y,f)$ , we update its Mueller matrix if a neighboring pixel $(x',y',f')$ offers a lower error: + +$$ +\hat {\mathbf {M}} _ {x y f} \leftarrow \hat {\mathbf {M}} _ {x ^ {\prime} y ^ {\prime} f ^ {\prime}}, +$$ + +$$ +\text {i f} \| \mathbf {D} _ {x y z} \mathbf {B} _ {x y f} \hat {\mathbf {M}} _ {x y f} \| _ {1} > \| \mathbf {D} _ {x y z} \mathbf {B} _ {x y f} \hat {\mathbf {M}} _ {x ^ {\prime} y ^ {\prime} f ^ {\prime}} \| _ {1}. \tag {11} +$$ + +Next, we refine the Mueller matrix by applying random perturbations and accepting the perturbation if it reduces the error: + +$$ +\hat {\mathbf {M}} _ {x y f} \leftarrow \hat {\mathbf {M}} _ {x y f} ^ {\text {p e r t u r b}}, \text {w h e r e} +$$ + +$$ +\hat {\mathbf {M}} _ {x y f} ^ {\text {p e r t u r b}} = \operatorname {C l o u d e F i l t e r} \left(\hat {\mathbf {M}} _ {x y f} \odot (\mathbf {1} + \sigma \mathbf {N})\right), \tag {12} +$$ + +$$ +\left\| \mathbf {D} _ {x y z} \mathbf {B} _ {x y f} \hat {\mathbf {M}} _ {x y f} \right\| _ {1} > \left\| \mathbf {D} _ {x y z} \mathbf {B} _ {x y f} \mathbf {M} _ {x y f} ^ {\text {p e r t u r b}} \right\| _ {1}, \tag {13} +$$ + +where $\hat{\mathbf{M}}_{xyf}^{\mathrm{perturb}}$ is the perturbed Mueller matrix, $\odot$ denotes element-wise multiplication, $\mathbf{1}$ is an all-ones matrix, $\mathbf{N} \in \mathbb{R}^{4 \times 4}$ contains random values from a standard normal distribution, and $\sigma$ is a scalar set to 0.01. This multiplicative perturbation preserves the relative scale of the Mueller matrix elements, allowing larger elements to undergo proportionally larger adjustments while smaller elements change minimally. We apply Cloude's filtering to the perturbed Mueller matrix to ensure physical validity. + +We repeat the propagation and refinement steps iteratively for a fixed number of iterations set to 10 in our experiments to obtain the final Mueller matrix $\hat{\mathbf{M}}_{xyf}$ . + +# 6. Calibration + +We perform one-time calibration of the system parameters: contrast threshold $(C)$ and QWP offset angles $(i_1$ and $i_2)$ . Calibration details are provided in the Supplementary Material. + +![](images/9b9874e3c86ddcd40057fcf07f0f0e4a7726ce6886b40d2910e99a921e9b5067.jpg) +Figure 5. Assessment of reconstructed Mueller matrix on real data. (a) Evaluation with known optical elements. We show the corresponding mean squared errors (MSEs). (b) Measurement on an in-the-wild metal plate induces strong diagonal components. + +![](images/95c7017207b1055cbc2971ecf99283d25592a95faf33ddf9487348def26644c6.jpg) + +Contrast Threshold We calibrate the per-pixel contrast threshold $C$ of an event sensor by analyzing pixel responses under controlled illumination stimuli. We increase and decrease the LED intensity linearly over time, and the event camera directly captures the LED light, generating events. We then fit the contrast threshold $C$ to the events per each pixel using Equation (6). + +QWP Offset Angle We obtain the offset angles of the fast axis of the QWP in our system: $i_1$ and $i_2$ by capturing another QWP sample with known fast axis. Given the known Mueller matrix of the reference QWP, we find the best offset angles that minimize the reconstruction error of Equation (6). + +# 7. Results + +Implementation We implemented our reconstruction algorithm in $\mathrm{C + + }$ using OpenMP CPU parallelization. It takes 62 seconds for per-pixel reconstruction (Section 5.1), and 68 seconds for the propagation and refinement processes (Section 5.2), measured for processing $500\times 500$ -resolution event image sequence of 30 frames, with an average event rate of $153\mathrm{MEv / s}$ , on the AMD Ryzen 7 7800X3D 8-core processor. + +Validation on Synthetic Data To evaluate our method, we render synthetic data in three steps. First, we mimic our capture system by replacing the event camera with an intensity camera in Mitsuba3 [29], and render an image sequence of a synthetic scene for each QWP angle with interval $\omega t = 0.01^{\circ}$ . We use the real-world polarimetric BRDFs to construct the synthetic scene [8]. Then, the rendered images are converted to event streams by simulating event-camera processing based on DVS-Voltmeter [34]. Last, we add Gaussian noise with a standard deviation of 0.5 and replace $5\%$ of true events with values sampled from a Gaussian distribution with a standard deviation of 5.0. Figure 4 + +shows the synthetic data and reconstructed Mueller matrix using our method. While initial SVD-based estimation suffers from noise with a reconstruction error of 0.11, our full reconstruction improves accuracy with a reconstruction error of 0.04. + +Validation on Real Data Figure 5(a) shows that our reconstruction closely matches the corresponding pseudo ground truth of real-world samples: air, linear polarizer $(0^{\circ}$ and $45^{\circ})$ , quarter-wave-plate $(0^{\circ}$ and $45^{\circ})$ . Figure 5(b) shows the result on an in-the-wild sample: a metal plate. The strong diagonal components in this result indicate the preservation of polarization, which is characteristic of reflections from metal surfaces. Notably, these Mueller matrix images are captured in just $33\mathrm{ms}$ , significantly faster than the conventional ellipsometer, which requires several minutes due to its frame-based imaging principle. + +Photoelasticity of Transparent Gelatine Photoelasticity is an optical property whereby dielectric materials exhibit birefringence under deformation. Imaging photoelasticity has applications in mechanical stress and material analysis [39, 47]. Analyzing Mueller-matrix images reveals such stress distributions even in transparent objects [18]. Our system enables observing photoelasticity at video rates. Figure 6 shows Mueller-matrix images of a gelatine disk, revealing stress-dependent fringe patterns. + +Transparent Tape Detection Detecting transparent objects is challenging for conventional cameras. Sticky tape, a common transparent material used for sealing cardboard, exhibits birefringent properties due to the molecular structure of its stretched plastic. Figure 7 shows the measurement result of sticky tape on a cardboard box. The reconstructed Mueller matrix shows the birefringent property, and we can clearly recognize the tape region. This + +![](images/a2a796269c4579178bf5d1109ea761d6a9520c24de15f1bbf640c151a7b9397d.jpg) +(a) Measurement setup + +![](images/6dfd408613e57492f96aaceffb9b904fddee952a68e00ea26fa7fcb2cc5a87b3.jpg) +Before applying force + +![](images/69de5e4f17a7b054dd2a6372a85846bc76ba1f987393359ecf5adcf39a9de424.jpg) +After applying force +(b) Reconstructed Mueller matrix -1 +1 + +![](images/402acf9ad9adc834b4606f7f0399564f92aa2e629080718b1ae390ac3ea33f80.jpg) +Figure 6. Photoelasticity analysis. (a) Experimental setup for measuring a gelatine disk in transmission mode. We gradually apply force for the gelatine disk to observe changes in photoelastic properties. (b) Reconstructed Mueller matrix images with different forces, revealing complex stress-dependent polarimetric patterns. Notably, as force increases, the Mueller matrix image shows denser fringe patterns. +(a) Target object + +![](images/962460144edd04345c48b98a143598e506e502ece3e65a9d4dd3ba5ff234e69f.jpg) + +![](images/9decf049b0c3c899c1ed388802f1744dbaee633dd22b1c4f1e26c9e8b5f37499.jpg) + +![](images/bd815b372e8f21768bac318cacb8c14d957b5da6ecccfcc7f712aef0e46840dd.jpg) +(b) Comparison + +![](images/c8b1f8a35d10105438711d3063a39782856bd8e90e61981fefe0f2c36c3dfe8c.jpg) +(c) Reconstructed Mueller matrix + +![](images/f2f7ae8724d6bbe20d66dd441421974b8003f13ce8b0fe6337fa8710283cd790.jpg) +Figure 7. Transparent tape detection. (a) The target object is transparent sticky tape on a box. (b)&(c) The Reconstructed Mueller matrix reveals the tape region, which is difficult to see with a conventional RGB camera. + +demonstration shows the potential for inspecting dynamically moving sealed boxes in automated industrial processes. + +Dynamic Human Capture Ellipsometers, as a comprehensive polarization imaging technique for human capture, can reveal hidden properties of the human face and hair that are difficult to analyze in conventional imaging and non-ellipsometric polarization imaging [2, 19]. In Figure 8, we + +demonstrate the Mueller-matrix reconstruction of dynamic facial expressions and hair movements, revealing intricate polarization properties. The reconstructed Mueller matrix for the face exhibits strong diagonal components associated with specular reflection, while the non-diagonal regions display weakly polarized reflection with a dependency on surface normal along the face edges. The hair result shows polarization property in the specular highlight regions. + +HDR Mueller-matrix Imaging Our event-based ellipsometer enables Mueller-matrix imaging of HDR scenes. Figure 9 shows reconstruction results on a scene with specular and dark diffuse reflections. While conventional ellipsometers require multiple exposures to prevent overexposure, our method captures the Mueller matrix in HDR scenes at 30 fps without the need for additional measurement. + +# 8. Discussion + +Normalized Mueller-Matrix Imaging Normalized Mueller matrix has been extensively used as it provides rich scene information [5, 7, 18]. Since event cameras only detect intensity changes, not intensity itself, our method also reconstructs a normalized version of Mueller matrix per pixel, $\mathbf{M} / \mathbf{M}(0,0)$ . We leave reconstructing full Mueller-matrix as a future work that might be accomplished using an event-intensity hybrid imaging system. + +End-to-end Real-time Pipeline Although our capture speed is fast, our current Mueller-matrix reconstruction relies solely on CPU acceleration, resulting in a non-realtime pipeline from capture to reconstruction. Implementing GPU acceleration could enable on-the-fly capture, reconstruction, and visualization [22, 48]. + +![](images/1604f053161511fba6b9020f246bb5f1f1bdda08910b9714e0121bdf77abd7ae.jpg) +Figure 8. Mueller matrix acquisition for capturing dynamic human (a) face and (b) hair, demonstrating the capture of diffuse and specular polarimetric responses. + +![](images/63a874a97f0afd956c1560d17b1e60692542d22c446ecbcd3e7075f33b7d28e1.jpg) +(a) Target object +Figure 9. Mueller matrix measurement for a HDR scene. (a) The target scene contains regions with both strong specular reflection and black diffuse reflection. (b) Our proposed method achieves an accurate Mueller-matrix image for the HDR scene in just $33~\mathrm{ms}$ . + +![](images/9c13ff77dab3feb86078999efb4ef9d13eac5ba120f60f18b4e6589db59b5ec5.jpg) +(b) Reconstructed Mueller matrix + +Event Bandwidth and Sensitivity The ideal setting of an event camera is to have a low contrast threshold and a short refractory period, to capture subtle intensity changes. However, this produces many events, exceeding the transmission bandwidth, resulting in lost or delayed events [21]. In contrast, when the contrast threshold is high, weakly polarized phenomena cannot be detected as events. Given this tradeoff, we defined a region of interest (ROI) in the event sensor and empirically configured the camera settings to find the best parameters. + +DC Motor Synchronization For faster rotation speed and precise synchronization between two motors, it is a viable alternative to use brushless DC motors, also known as an electronically commutated (EC) motors. However, this increases system building costs. + +Gradient Descent Gradient descent in an automatic differentiation framework offers an alternative approach for Mueller-matrix reconstruction. However, the number of events differs across pixels, which complicates the construction of a dense tensor for efficient computation on modern GPUs. + +# 9. Conclusion + +We have introduced Event Ellipsometer, a Mueller-matrix video imaging method that combines an event camera and a light source with fast-rotating QWPs. Our image formation model, calibration, and reconstruction method enable robust Mueller-matrix imaging at 30 fps. We validate our method on synthetic and real data, and demonstrate Mueller-matrix imaging on photoelasticity, dynamic human hair and face capture, HDR imaging, and transparent tape detection. + +Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532, 2022R1A6A1A03052954, RS-2024-00437866), and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52. + +# References + +[1] Hans Arwin. Application of ellipsometry techniques to biological materials. Thin Solid Films, 519(9):2589-2592, 2011. 1, 2 +[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 7 +[3] RMA Azzam. Photopolarimetric measurement of the mueller matrix by fourier analysis of a single detected signal. Optics Letters, 2(6):148-150, 1978. 3 +[4] Rasheed MA Azzam. Stokes-vector and mueller-matrix polarimetry. JOSA A, 33(7):1396-1408, 2016. 2 +[5] Seung-Hwan Baek and Felix Heide. Polarimetric spatiotemporal light transport probing. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 7 +[6] Seung-Hwan Baek and Felix Heide. All-photon polarimetric time-of-flight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17876-17885, 2022. 1, 2 +[7] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM Trans. Graph., 37(6):268, 2018. 1, 2, 7 +[8] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph., 39(4):139, 2020. 1, 2, 6 +[9] Seung-Hwan Baek, Ramesh Raskar, Jinwei Ye, Akshit Dave, Achuta Kadambi, and Huajin Chen. Polarization-based visual computing. In ACM SIGGRAPH 2023 Courses, pages 1-1, 2023. 1 +[10] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 5 +[11] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In *Bmvc*, pages 1-11, 2011. 5 +[12] Qizhi Cao, Min Jiang, Chenling Jia, Siyue Jiang, Jing Zhang, Baoli Yao, Mingwu Jin, Edward Dehoog, Lian Duan, Huahua Wang, et al. Snapshot imaging mueller matrix polarimeter using modified savart polariscopes. Applied Optics, 62(8):2124-2129, 2023. 1, 2 +[13] Dorian Chan, Mark Sheinin, and Matthew O'Toole. Spincam: High-speed imaging via a rotating point-spread function. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10789-10799, 2023. 2 +[14] Tongbo Chen, Hendrik PA Lensch, Christian Fuchs, and Hans-Peter Seidel. Polarization and phase-shifting for 3d scanning of translucent objects. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2 +[15] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition, pages 14760-14770, 2021. 2 +[16] Shane R Cloude. Conditions for the physical realizability of matrix operators in polarimetry. In *Polarization Considerations for Optical Systems II*, pages 177–187. SPIE, 1990. 4 +[17] Edward Collett. Field guide to polarization. Spie Bellingham, 2005. 1, 2 +[18] Akshat Dave, Tianyi Zhang, Aaron Young, Ramesh Raskar, Wolfgang Heidrich, and Ashok Veeraraghavan. Nest: Neural stress tensor tomography by leveraging 3d photoelasticity. arXiv preprint arXiv:2406.10212, 2024. 1, 6, 7 +[19] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 7 +[20] Hiroyuki Fujiwara. Spectroscopic ellipsometry: principles and applications. John Wiley & Sons, 2007. 2 +[21] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 4, 8 +[22] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE international conference on computer vision, pages 873-881, 2015. 5, 7 +[23] Nirmalya Ghosh and I Alex Vitkin. Tissue polarimetry: concepts, challenges, applications, and outlook. Journal of biomedical optics, 16(11):110801-110801, 2011. 1, 2 +[24] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20616-20625, 2023. 2 +[25] Michael Hawks and Michael Dexter. Event-based imaging polarimeter. Optical Engineering, 61(5):053101-053101, 2022. 2 +[26] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 2 +[27] Inseung Hwang, Daniel S Jeon, Adolfo Munoz, Diego Gutierrez, Xin Tong, and Min H Kim. Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 2 +[28] Tomoki Ichikawa, Shohei Nobuhara, and Ko Nishino. Spiders: Structured polarization for invisible depth and reflectance sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25077-25085, 2024. 2 +[29] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsubishi 3 renderer, 2022. https://mitsuba-renderer.org. 6 + +[30] GE Jellison, Donovan N Leonard, Lawrence M Anovitz, Chad M Parish, Eliot D Specht, and TM Rosseel. Crystallographic orientation of uniaxial calcite and dolomite determined using reflection generalized ellipsometry. Journal of Applied Physics, 124(22), 2018. 1, 2 +[31] Achuta Kadambi, Vage Taamazyan, Boxin Shi, and Ramesh Raskar. Polarized 3d: High-quality depth sensing with polarization cues. In Proceedings of the IEEE international conference on computer vision, pages 3370-3378, 2015. 1 +[32] Agastya Kalra, Vage Taamazyan, Supreeth Krishna Rao, Kartik Venkataraman, Ramesh Raskar, and Achuta Kadambi. Deep polarization cues for transparent object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8602-8611, 2020. 1 +[33] Michael W Kudenov, Michael J Escuti, Nathan Hagen, Eustace L Dereniak, and Kazuhiko Oka. Snapshot imaging mueller matrix polarimeter using polarization gratings. Optics letters, 37(8):1367-1369, 2012. 1, 2 +[34] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In European Conference on Computer Vision, pages 578-593. Springer, 2022. 6 +[35] Ryota Maeda and Shinsaku Hiura. Polarimetric light transport analysis for specular inter-reflection. IEEE Transactions on Computational Imaging, 10:876-887, 2024. 1, 2 +[36] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 982-987. IEEE, 2003. 1 +[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2 +[38] David A Ramsey and Kenneth C Ludema. The influences of roughness on film thickness measurements by mueller matrix ellipsometry. Review of scientific instruments, 65(9):2874-2881, 1994. 1, 2 +[39] Michele Scafidi, Giuseppe Pitarresi, Andrea Toscano, Giovanni Petrucci, Sabina Alessi, and Augusto Ajovalasit. Review of photoelastic image analysis applied to structural birefringent materials: glass and polymers. Optical Engineering, 54(8):081206-081206, 2015. 6 +[40] Yoav Y Schechner and Shree K Nayar. Uncontrolled modulation imaging. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-II. IEEE, 2004. 2 +[41] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, pages 814-819. IEEE, 1999. 1 +[42] Yoav Y Schechner, Srinivasa G Narasimhan, and Shree K Nayar. Instant dehazing of images using polarization. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I. IEEE, 2001. 1 + +[43] Dominik Scheuble, Chenyang Lei, Seung-Hwan Baek, Mario Bijelic, and Felix Heide. Polarization wavefront lidar: Learning large scene reconstruction from polarized wavefronts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21241-21250, 2024. 1, 2 +[44] Matthew H Smith. Optimization of a dual-rotating-retarder mueller matrix polarimeter. Applied optics, 41(13):2488-2493, 2002. 3 +[45] Tsuyoshi Takatani, Yuzuha Ito, Ayaka Ebisu, Yinqiang Zheng, and Takahito Aoto. Event-based bispectral photometry using temporally modulated illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15647, 2021. 2 +[46] Tali Treibitz and Yoav Y Schechner. Active polarization scattering. IEEE transactions on pattern analysis and machine intelligence, 31(3):385-399, 2008. 2 +[47] Li Wang, Yang Ju, Heping Xie, Guowei Ma, Lingtao Mao, and Kexin He. The mechanical and photoelastic properties of 3d printable stress-visualized materials. Scientific reports, 7(1):10918, 2017. 6 +[48] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7 +[49] Aun Zaidi, Noah A Rubin, Maryna L Meretska, Lisa W Li, Ahmed H Dorrah, Joon-Suh Park, and Federico Capasso. Metasurface-enabled single-shot and complete mueller matrix imaging. Nature Photonics, pages 1-9, 2024. 1, 2 \ No newline at end of file diff --git a/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/images.zip b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ba467ceabd2561aa43e7900150dab5eeec98f1d --- /dev/null +++ b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a28feffd787e52eb26600f928a3d11d92f5b5aa01a1576672a6ec31271425a1 +size 805250 diff --git a/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/layout.json b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fde89cf78c5e133a6276899b24e12a04e61033a7 --- /dev/null +++ b/2025/Event Ellipsometer_ Event-based Mueller-Matrix Video Imaging/layout.json @@ -0,0 +1,9853 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 505, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 121 + ], + "type": "text", + "content": "Event Ellipsometer: Event-based Mueller-Matrix Video Imaging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "spans": [ + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "text", + "content": "Ryota Maeda" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "text", + "content": " Yunseong Moon" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "text", + "content": " Seung-Hwan Baek" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "text", + "content": "POSTECH " + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 155, + 142, + 455, + 178 + ], + "type": "text", + "content": "University of Hyogo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 204, + 200, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 204, + 200, + 217 + ], + "spans": [ + { + "bbox": [ + 151, + 204, + 200, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 229, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 421 + ], + "type": "text", + "content": "Light-matter interactions modify both the intensity and polarization state of light. Changes in polarization, represented by a Mueller matrix, encode detailed scene information. Existing optical ellipsometers capture Mueller-matrix images; however, they are often limited to capturing static scenes due to long acquisition times. Here, we introduce Event Ellipsometer, a method for acquiring a Mueller-matrix video for dynamic scenes. Our imaging system employs fast-rotating quarter-wave plates (QWPs) in front of a light source and an event camera that asynchronously captures intensity changes induced by the rotating QWPs. We develop an ellipsometric-event image formation model, a calibration method, and an ellipsometric-event reconstruction method. We experimentally demonstrate that Event Ellipsometer enables Mueller-matrix video imaging at 30 fps, extending ellipsometry to dynamic scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 135, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 135, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 135, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": "Polarization describes the oscillation of the electric field in light waves, encoding valuable information about the scenes with which light interacts. The polarization state of light can be represented as a Stokes vector " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{4 \\times 1}" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": " [17]. Polarimetric image analysis focuses on capturing complete or partial forms of the per-pixel Stokes vector and has been extensively studied for shape-from-polarization [31, 36], diffuse-specular separation [19], reflection removal [41], seeing through scattering [42], and transparent object segmentation [32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Ellipsometry advances polarimetric imaging by analyzing the polarization state of light captured under varying polarization states of illumination. This allows for acquiring polarimetric reflectance, represented as a Mueller matrix " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " [9], which comprehensively characterizes how light-matter interactions alter the polarization state of incident light. Ellipsometry has been widely used in material science [30, 38], biology [1, 23], and has recently gained attention in computer vision and graphics for 3D shape and reflection analysis [6, 7, 43], material acquisition [8], light transport decomposition [5, 35], and photoelasticity analy-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 204, + 555, + 301 + ], + "blocks": [ + { + "bbox": [ + 316, + 204, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 316, + 204, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 316, + 204, + 555, + 301 + ], + "type": "image", + "image_path": "75b8dfbe7a8dcc63c0d043df6f7ac2b727225229cd7f24508b188991ab2a1416.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 303, + 515, + 313 + ], + "lines": [ + { + "bbox": [ + 353, + 303, + 515, + 313 + ], + "spans": [ + { + "bbox": [ + 353, + 303, + 515, + 313 + ], + "type": "text", + "content": "(a) Overview diagram of Event Ellipsometer" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 316, + 450, + 423 + ], + "blocks": [ + { + "bbox": [ + 316, + 316, + 450, + 423 + ], + "lines": [ + { + "bbox": [ + 316, + 316, + 450, + 423 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 450, + 423 + ], + "type": "image", + "image_path": "b3a693dcb625d5226236bd6a6200934911058c7dbfb4bf65c27a425afa108ee1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 424, + 429, + 433 + ], + "lines": [ + { + "bbox": [ + 336, + 424, + 429, + 433 + ], + "spans": [ + { + "bbox": [ + 336, + 424, + 429, + 433 + ], + "type": "text", + "content": "(b) Capture configuration" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 451, + 316, + 503, + 423 + ], + "blocks": [ + { + "bbox": [ + 451, + 316, + 503, + 423 + ], + "lines": [ + { + "bbox": [ + 451, + 316, + 503, + 423 + ], + "spans": [ + { + "bbox": [ + 451, + 316, + 503, + 423 + ], + "type": "image", + "image_path": "e9a0c624fe9ac22f84299a4f97bfea94be6a5de132f22f80ec54d2a184626339.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 423, + 549, + 440 + ], + "lines": [ + { + "bbox": [ + 455, + 423, + 549, + 440 + ], + "spans": [ + { + "bbox": [ + 455, + 423, + 549, + 440 + ], + "type": "text", + "content": "HDR imaging Tape detection (c) Applications" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 449, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 506 + ], + "type": "text", + "content": "Figure 1. Overview of Event Ellipsometer. (a) Our imaging system captures the Mueller matrix at 30 fps from event streams induced by the continuously rotating QWPs. (b) With our experimental prototype, (c) we demonstrate ellipsometric analysis for dynamic scenes and various applications." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 503, + 316, + 553, + 423 + ], + "blocks": [ + { + "bbox": [ + 503, + 316, + 553, + 423 + ], + "lines": [ + { + "bbox": [ + 503, + 316, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 503, + 316, + 553, + 423 + ], + "type": "image", + "image_path": "86d21b84eee3a9365a10fb85e8e8957c9709facf300fb66ce9dd09ecac0a7c5b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 507, + 423, + 549, + 439 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 507, + 423, + 549, + 439 + ], + "spans": [ + { + "bbox": [ + 507, + 423, + 549, + 439 + ], + "type": "text", + "content": "Tape detection cations" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "bbox": [ + 313, + 517, + 349, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 349, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 349, + 529 + ], + "type": "text", + "content": "sis [18]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 531, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 531, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 312, + 531, + 555, + 699 + ], + "type": "text", + "content": "Despite the rich information provided by ellipsometry, its applications have been mainly limited to static scenes because conventional methods require capturing multiple images while mechanically rotating polarizing optics. Typically, more than 20 rotation angles [8] are necessary, leading to long acquisition times and making it impractical for capturing dynamic scenes. Although single-shot Mueller matrix imaging techniques using conventional sensors exist [12, 33, 49], they assume planar target scenes at fixed distances, significantly sacrifice sensor resolution, or use custom-fabricated nano- or micro-optical elements. Additionally, conventional intensity sensors are limited in capturing high-dynamic range (HDR) scenes, often requiring the capture of additional images with multiple exposures." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "In this paper, we present Event Ellipsometer, a Mueller-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21804" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "matrix imaging method capable of capturing dynamic and HDR scenes. Departing from using conventional intensity sensors, we utilize an event camera, which asynchronously records intensity changes. We equip fast-rotating QWPs and linear polarizers (LPs) in front of both the event camera and an LED light source. The rotating QWPs modulate the polarization state of light emitted and received by the imaging system, resulting in intensity changes captured as events." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 275 + ], + "type": "text", + "content": "We develop an ellipsometric-event image formation model that relates the time differences of adjacent events to the Mueller matrix, normalized by its first element. Using this model, we propose a two-stage Mueller-matrix reconstruction method consisting of per-pixel estimation and spatiotemporal propagation. We incorporate physical validity constraints to handle outliers from sensor noise and scene motion. We also devise calibration methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 294, + 371 + ], + "type": "text", + "content": "Experimentally, we demonstrate Mueller-matrix imaging at 30 fps, achieving a mean-squared error of 0.045 for materials with known Mueller matrices. Unlike previous single-shot methods [49], Event Ellipsometer can capture non-planar objects and does not compromise spatial resolution. Furthermore, our method is capable of capturing HDR scenes without the need for additional measurement with different exposures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 372, + 251, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 372, + 251, + 383 + ], + "spans": [ + { + "bbox": [ + 67, + 372, + 251, + 383 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 384, + 294, + 551 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 384, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 407 + ], + "type": "text", + "content": "- We propose Event Ellipsometer, enabling Mueller-matrix imaging for dynamic scenes at 30 fps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 407, + 294, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 442 + ], + "type": "text", + "content": "- We develop an imaging system using an event camera and a light source, each equipped with synchronized fast-rotating QWPs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 443, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 491 + ], + "type": "text", + "content": "- We formulate an ellipsometric-event image formation model relating event streams to the Mueller matrix, along with a calibration method and a robust ellipsometric-event reconstruction algorithm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 491, + 294, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 551 + ], + "type": "text", + "content": "- We experimentally show the high accuracy of Event Ellipsometer on samples with known ground truth, and demonstrate ellipsometric analysis of dynamic scenes with applications on photoelasticity, human capture, HDR imaging, and tape detection." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 562, + 142, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 142, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 142, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 294, + 713 + ], + "type": "text", + "content": "Imaging with Polarized Illumination Polarimetric imaging with polarized illumination exploits polarization-dependent light transport to extract material and geometric scene properties. Previous methods often capture scenes with single or a few polarized illuminations and fit parametric models to the captured images [14, 27, 28, 46]. However, this approach is limited by the representation power of the parametric models, thus cannot reveal the true polarimetric reflectance of real-world scenes. Ellipsometry extends polarimetric imaging by directly capturing the Mueller matrix, providing a comprehensive characteriza" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "tion of material polarization properties [4, 17, 20]. It has applications in material science [30, 38], biology [1, 23], and has recently been applied to computer vision and graphics for tasks such as 3D shape analysis [6, 7, 43], material acquisition [8], and light transport decomposition [5, 35]. However, conventional ellipsometry techniques are unsuitable for dynamic scenes because they require capturing multiple images while mechanically rotating polarizing optics and capturing each with an intensity sensor, leading to long acquisition. Single-shot Mueller-matrix imaging methods have been proposed [12, 33, 49], however these approaches often compromise spatial resolution and are limited to planar scenes at fixed distances, restricting their applicability to general dynamic scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 257, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 437 + ], + "type": "text", + "content": "Event Cameras for Photometric Analysis Event cameras offer a high temporal resolution of microseconds and a HDR compared to standard frame-based intensity cameras [21]. These advantages have been leveraged for photometric analysis by combining continuously modulated optical light sources or filters. For instance, intensity-modulated illumination with an event camera enables radiance estimation [15, 24] and bispectral photometry [45]. Yu et al. [48] demonstrated event-based photometric stereo for normal estimation with a mechanically-rotating point light source. Hawks et al. [25] and Muglikar et al. [37] explored passive linear-polarization imaging by rotating a linear polarizer in front of an event camera. Existing event-based vision methods cannot capture full polarization reflectance properties as a Mueller-matrix image, limiting their utility." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 553, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 553, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 553, + 560 + ], + "type": "text", + "content": "Imaging with Rotating Optical Elements Rotating optical components have been used for HDR and multispectral imaging [40] and high-speed video reconstruction [13] using a conventional intensity camera. Recently, rotating optics with event cameras have been demonstrated to leverage asynchronous operation and high temporal resolution of the event cameras [25, 26, 37]. These methods rotate optical elements only in front of an event camera without illumination modulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 573, + 411, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 411, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 411, + 586 + ], + "type": "text", + "content": "3. Imaging System" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "Experimental Prototype We build an imaging system using an LED light source (Thorlabs MCWHLP3), fast-rotating QWPs (Edmund Optics WP140HE), fixed LPs (Thorlabs WP25M-VIS), and an event camera (Prophesee EVK4). Figure 2(a) and (c) depict our setup, where a pair of a QWP and an LP is placed in front of the event camera and the light source, respectively. We rotate the QWPs so that the event camera detects the event streams caused by the rotating QWPs, from which the normalized Mueller matrix is reconstructed. We can configure the illumination" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21805" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 164 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 164 + ], + "type": "image", + "image_path": "9c03fb221471db728db7ed2545b1af70a9754bf0e18549e23b37f33e4449ae44.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 164, + 217, + 174 + ], + "lines": [ + { + "bbox": [ + 135, + 164, + 217, + 174 + ], + "spans": [ + { + "bbox": [ + 135, + 164, + 217, + 174 + ], + "type": "text", + "content": "(a) Schematic diagram" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 175, + 294, + 312 + ], + "blocks": [ + { + "bbox": [ + 56, + 175, + 294, + 312 + ], + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 312 + ], + "type": "image", + "image_path": "3a90a26a4009ce56153453a59a00e14c9a9bbb365c2d68d6d56726799edf44ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 312, + 276, + 322 + ], + "lines": [ + { + "bbox": [ + 78, + 312, + 276, + 322 + ], + "spans": [ + { + "bbox": [ + 78, + 312, + 276, + 322 + ], + "type": "text", + "content": "(b) Timeline of QWPs rotation and event measurement" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 333, + 555, + 367 + ], + "lines": [ + { + "bbox": [ + 54, + 333, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 555, + 367 + ], + "type": "text", + "content": "Figure 2. Imaging system of Event Ellipsometer. (a) Schematic diagram illustrating the optical arrangement and hardware operation. (b) Timeline showing the rotation of two QWPs and the event measurement. (c) Our hardware prototype. The system can move the light source and camera position for use in (d) Reflection mode or (e) Transmission mode." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 491, + 201 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 491, + 201 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 491, + 201 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 491, + 201 + ], + "type": "image", + "image_path": "87c6155bb1edeb05151d21244a090246f9840b39a24a34c24b4b811e7a017b59.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 202, + 474, + 212 + ], + "lines": [ + { + "bbox": [ + 389, + 202, + 474, + 212 + ], + "spans": [ + { + "bbox": [ + 389, + 202, + 474, + 212 + ], + "type": "text", + "content": "(c) Hardware prototype" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 493, + 70, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 493, + 70, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 493, + 70, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 493, + 70, + 553, + 198 + ], + "type": "image", + "image_path": "a15b567aef125c9a17b3b485302740ec647279efdc17e75189b98d01d6ff3f79.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 196, + 541, + 205 + ], + "lines": [ + { + "bbox": [ + 507, + 196, + 541, + 205 + ], + "spans": [ + { + "bbox": [ + 507, + 196, + 541, + 205 + ], + "type": "text", + "content": "QWP film" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 215, + 425, + 312 + ], + "blocks": [ + { + "bbox": [ + 296, + 215, + 425, + 312 + ], + "lines": [ + { + "bbox": [ + 296, + 215, + 425, + 312 + ], + "spans": [ + { + "bbox": [ + 296, + 215, + 425, + 312 + ], + "type": "image", + "image_path": "e1b3059d4828ee7022366407b074de23b5c015aa5ade8be29b7aef105caa29e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 312, + 397, + 322 + ], + "lines": [ + { + "bbox": [ + 324, + 312, + 397, + 322 + ], + "spans": [ + { + "bbox": [ + 324, + 312, + 397, + 322 + ], + "type": "text", + "content": "(d) Reflection mode" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 215, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 425, + 215, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 425, + 215, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 425, + 215, + 553, + 312 + ], + "type": "image", + "image_path": "b3195701c7db6a0391f4fa754e1788354108206742ef8f5038b534b92144a4dc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 312, + 531, + 322 + ], + "lines": [ + { + "bbox": [ + 448, + 312, + 531, + 322 + ], + "spans": [ + { + "bbox": [ + 448, + 312, + 531, + 322 + ], + "type": "text", + "content": "(e) Transmission mode" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 373, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 445 + ], + "type": "text", + "content": "and the camera modules mounted on different breadboards in both the reflection model and the transmission mode as shown in Figure 2(d) and (e). Our system can be seen as a combination of the optical dual rotating retarder [3] and the event-based vision. For a complete list of components, refer to the Supplemental Material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": "Rotating QWPs and Angle Encoder We rotate the QWP on the camera side five times faster than the one on the light source side [44]. We denote the QWP angles of the light source and the camera as " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\theta_{1,t} = \\omega t + i_1, \\theta_{2,t} = 5\\omega t + i_2" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\omega = 30\\pi" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": " rad/sec is the angular velocity of the motor driving the light-source QWP. " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "i_1" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "i_2" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": " are the initial QWP angles. One challenge here is to rotate the QWPs at such high speeds while recording their angles " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\theta_{1,t}" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\theta_{2,t}" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": " over time " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": ". To this end, we use two independently controlled brushed direct current (DC) motors rotating the QWP films, resulting in an affordable configuration. Also, we develop custom angle encoders using a 3D-printed occluder, a photointerrupter, and an Arduino microcontroller. The occluder rotates at the same speed as the QWP and blocks the light path within the photointerrupter at every " + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 459, + 296, + 675 + ], + "type": "text", + "content": " rotation. The microcontroller detects such change at a microsecond resolution and emits a hardware trigger to the event camera. Figure 2(c) shows the angle encoder." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Frame of Mueller Matrix Given the rotation speed of " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ", every " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " rotation of the light-source QWP takes " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pi/\\omega \\approx" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "text", + "content": "33 ms. We set this as the effective temporal duration of each frame " + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "text", + "content": " of the Mueller-matrix video which we aim to reconstruct. That is, as shown in Figure 2(b), each frame " + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "inline_equation", + "content": "f \\in \\{1, \\dots, F\\}" + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "text", + "content": " of the Mueller matrix is estimated based on the events measured during the temporal duration " + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\pi / \\omega" + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 373, + 555, + 456 + ], + "type": "text", + "content": " is the number of frames in a reconstructed Mueller-matrix video." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 419, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 419, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 419, + 481 + ], + "type": "text", + "content": "4. Image Formation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 488, + 554, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 511 + ], + "type": "text", + "content": "Here, we relate the event-camera measurements to the scene Mueller matrix which we aim to reconstruct." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": "Polarimetric Modulation The LED light source emits unpolarized light of Stokes vector " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = [1,0,0,0]^{\\intercal}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": ". At a time " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": ", the light passes through the LP and the QWP rotated by angle " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\theta_{1,t}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": ", resulting in the Stokes vector " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}(\\theta_{1,t})\\mathbf{L}(0)\\mathbf{s}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}(\\theta_{1,t})" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(0)" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": " denote the Mueller matrices of a QWP rotated by " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\theta_{1,t}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": " and an LP at " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": ", respectively. The light then interacts with the target scene, undergoing polarization change represented by the scene Mueller matrix " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": ". After interaction, the light passes through another QWP at angle " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\theta_{2,t}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": " and an LP in front of the event camera, resulting in the time-varying intensity " + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 526, + 555, + 658 + ], + "type": "text", + "content": " incident on the sensor:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 357, + 667, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 667, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 357, + 667, + 553, + 681 + ], + "type": "interline_equation", + "content": "I _ {t} = [ \\mathbf {L} (0) \\mathbf {Q} (\\theta_ {2, t}) \\mathbf {M} \\mathbf {Q} (\\theta_ {1, t}) \\mathbf {L} (0) \\mathbf {s} ] _ {0}, \\tag {1}", + "image_path": "679c220d3efef265a96dd5e4eb7592bfa57b87196f3705a2e908fcaf73a1d620.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "inline_equation", + "content": "\\left[\\cdot\\right]_{0}" + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": " denotes intensity, the first element of the Stokes vector." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21806" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "We rearrange Equation (1) and derive a matrix-vector form:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 97, + 294, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 97, + 294, + 110 + ], + "spans": [ + { + "bbox": [ + 150, + 97, + 294, + 110 + ], + "type": "interline_equation", + "content": "I _ {t} = \\mathbf {A} _ {t} \\hat {\\mathbf {M}}, \\tag {2}", + "image_path": "f29fd8ebab6367292098929561976c58991673ffcc0afbfac91f5a5b35c3320f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}} = [\\mathbf{M}_{00},\\mathbf{M}_{01},\\dots ,\\mathbf{M}_{33}]^{\\intercal}\\in \\mathbb{R}^{16\\times 1}" + }, + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "text", + "content": " is the vectorized form of " + }, + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t\\in \\mathbb{R}^{1\\times 16}" + }, + { + "bbox": [ + 55, + 117, + 295, + 153 + ], + "type": "text", + "content": " is the system matrix defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 163, + 274, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 163, + 274, + 210 + ], + "spans": [ + { + "bbox": [ + 77, + 163, + 274, + 210 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {A} _ {t} = [ 1, \\alpha_ {1} ^ {2}, \\alpha_ {1} \\alpha_ {2}, \\alpha_ {2}, \\alpha_ {3} ^ {2}, \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2}, \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2}, \\alpha_ {2} \\alpha_ {3} ^ {2}, \\\\ \\alpha_ {3} \\alpha_ {4}, \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4}, \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, - \\alpha_ {4} \\\\ \\left. - \\alpha_ {1} ^ {2} \\alpha_ {4}, - \\alpha_ {1} \\alpha_ {2} \\alpha_ {4}, - \\alpha_ {2} \\alpha_ {4} \\right], \\text {w h e r e} \\\\ \\end{array}", + "image_path": "acd780d960e13bdae97ef61e58087d8548e0e3e4503770fc32bf8ad265d46b11.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 213, + 260, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 213, + 260, + 225 + ], + "spans": [ + { + "bbox": [ + 78, + 213, + 260, + 225 + ], + "type": "interline_equation", + "content": "\\alpha_ {1} = \\cos (2 i _ {1} + 2 \\omega t), \\alpha_ {2} = \\sin (2 i _ {1} + 2 \\omega t),", + "image_path": "da083428a45d95a2f13983e8d3b27b5105da4180f93e7dfc3f99c154964f008c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 228, + 294, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 228, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 78, + 228, + 294, + 240 + ], + "type": "interline_equation", + "content": "\\alpha_ {3} = \\cos (2 i _ {2} + 1 0 \\omega t), \\alpha_ {4} = \\sin (2 i _ {2} + 1 0 \\omega t). \\tag {3}", + "image_path": "c30331189e3f7d4ed34130ae9e6a8d8221ee6a689d87d18a179e79396de166ce.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 251, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 262 + ], + "type": "text", + "content": "For the full derivation, refer to the Supplemental Document." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "text", + "content": "Differential Intensity The event camera triggers events based on the temporal change of the logarithm of photocurrent [21]. Analytically differentiating the logarithm of Equation (2) with respect to time " + }, + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "text", + "content": ", we obtain" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 338, + 295, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 338, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 119, + 338, + 295, + 368 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\log I _ {t}}{\\partial t} = \\frac {\\frac {\\partial I _ {t}}{\\partial t}}{I _ {t}} = \\frac {\\frac {\\partial \\mathbf {A} _ {t}}{\\partial t} \\hat {\\mathbf {M}}}{\\mathbf {A} _ {t} \\hat {\\mathbf {M}}}, \\tag {4}", + "image_path": "e1aa8c79d21802f4c6fc4b4e6e8a9d2f55c43e5701e9979de129d603d2357356.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 377, + 147, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 147, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 147, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 377, + 147, + 392 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathbf{A}_t}{\\partial t}" + }, + { + "bbox": [ + 55, + 377, + 147, + 392 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 401, + 301, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 401, + 301, + 441 + ], + "spans": [ + { + "bbox": [ + 58, + 401, + 301, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial \\mathbf {A} _ {t}}{\\partial t} = [ 0, - 4 \\alpha_ {1} \\alpha_ {2}, 2 \\alpha_ {1} ^ {2} - 2 \\alpha_ {2} ^ {2}, 2 \\alpha_ {1}, - 2 0 \\alpha_ {3} \\alpha_ {4}, \\\\ - 2 0 \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4} - 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2}, 2 \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2} - 2 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4} - 2 \\alpha_ {2} ^ {2} \\alpha_ {3} ^ {2}, \\\\ \\end{array}", + "image_path": "6afb9ae302d63039f37e799141ab8b296bed9825961786905f36125f8c90657f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 442, + 201, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 442, + 201, + 457 + ], + "spans": [ + { + "bbox": [ + 58, + 442, + 201, + 457 + ], + "type": "interline_equation", + "content": "2 \\alpha_ {1} \\alpha_ {3} ^ {2} - 2 0 \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, 1 0 \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {4} ^ {2},", + "image_path": "5fd3e9c328c1e98ad91a8c0a5ffbef35cb03caa06d03af1ced6713da6ceb4174.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 458, + 312, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 458, + 312, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 458, + 312, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} 1 0 \\alpha_ {1} ^ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {1} ^ {2} \\alpha_ {4} ^ {2} - 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} \\alpha_ {4}, \\\\ 2 \\alpha_ {1} ^ {2} \\alpha_ {3} \\alpha_ {4} + 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {4} ^ {2} - 2 \\alpha_ {2} ^ {2} \\alpha_ {3} \\alpha_ {4}, \\\\ 2 \\alpha_ {1} \\alpha_ {3} \\alpha_ {4} + 1 0 \\alpha_ {2} \\alpha_ {3} ^ {2} - 1 0 \\alpha_ {2} \\alpha_ {4} ^ {2}, - 1 0 \\alpha_ {3}, - 1 0 \\alpha_ {1} ^ {2} \\alpha_ {3} + 4 \\alpha_ {1} \\alpha_ {2} \\alpha_ {4}, \\\\ - 2 \\alpha_ {1} ^ {2} \\alpha_ {4} - 1 0 \\alpha_ {1} \\alpha_ {2} \\alpha_ {3} + 2 \\alpha_ {2} ^ {2} \\alpha_ {4}, - 2 \\alpha_ {1} \\alpha_ {4} - 1 0 \\alpha_ {2} \\alpha_ {3} ]. \\tag {5} \\\\ \\end{array}", + "image_path": "f9c53bf58c4e4b0c56b82022330bce603b8729e972cb1b09e4bc39143e5ccb11.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "text", + "content": "Events from Polarimetric Modulation For a pixel, we denote the time difference between consecutive events as " + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "inline_equation", + "content": "\\Delta t_{k}" + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "inline_equation", + "content": "k\\in \\{1,\\dots ,K\\}" + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "text", + "content": " is the event index in each frame " + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "text", + "content": " is the number of detected events at that pixel in the frame " + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 550, + 295, + 633 + ], + "type": "text", + "content": ". The time difference is known to be related to the change of photocurrent according to Taylor's expansion [21]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 633, + 294, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 294, + 659 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 294, + 659 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\log I _ {t _ {k}}}{\\partial t} = \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {6}", + "image_path": "4c55b3c7cfab37b1d0768ff50927a8e66ea96493e4c8696ba0a4ef31127227dd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 689 + ], + "type": "text", + "content": "where event polarity " + }, + { + "bbox": [ + 55, + 665, + 294, + 689 + ], + "type": "inline_equation", + "content": "p_k \\in \\{+1, -1\\}" + }, + { + "bbox": [ + 55, + 665, + 294, + 689 + ], + "type": "text", + "content": " is the sign of intensity change, and " + }, + { + "bbox": [ + 55, + 665, + 294, + 689 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 665, + 294, + 689 + ], + "type": "text", + "content": " is a constant threshold of the event camera." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "By combining Equations (4) and (6), we relate the measured per-pixel events to the Mueller matrix, resulting in the" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 314, + 71, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 553, + 129 + ], + "type": "image", + "image_path": "50bc1e749a9554ccd8774ced66bcaf79744b8f6c330feabd040c8c067a8d1c88.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 138, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 313, + 138, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 553, + 171 + ], + "type": "text", + "content": "Figure 3. Overview of our Mueller-matrix reconstruction pipeline. This method consists of two steps: (1) per-pixel reconstruction and (2) propagation and refinement." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 183, + 413, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 413, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 413, + 194 + ], + "type": "text", + "content": "image formation model:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 202, + 553, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 202, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 367, + 202, + 553, + 279 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {p _ {k} C}{\\Delta t _ {k}} = \\frac {\\frac {\\partial \\mathbf {A} _ {t _ {k}}}{\\partial t _ {k}} \\hat {\\mathbf {M}}}{\\mathbf {A} _ {t _ {k}} \\hat {\\mathbf {M}}} \\\\ \\Rightarrow \\left(\\frac {\\partial \\mathbf {A} _ {t _ {k}}}{\\partial t _ {k}} - \\frac {p _ {k} C}{\\Delta t _ {k}} \\mathbf {A} _ {t _ {k}}\\right) \\hat {\\mathbf {M}} = 0 \\\\ \\Rightarrow \\mathbf {B} _ {t _ {k}} \\dot {\\mathbf {M}} = 0, \\tag {7} \\\\ \\end{array}", + "image_path": "97c20452970646fc119f90f2ee63368a728b284358086714200947b8ef48f994.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 287, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 482, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 287, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t_k} \\in \\mathbb{R}^{1 \\times 16}" + }, + { + "bbox": [ + 313, + 287, + 482, + 300 + ], + "type": "text", + "content": " is the system matrix." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 310, + 407, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 310, + 407, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 310, + 407, + 322 + ], + "type": "text", + "content": "5. Reconstruction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 330, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 390 + ], + "type": "text", + "content": "The image formation model in Equation (7) allows us to reconstruct the Mueller matrix " + }, + { + "bbox": [ + 313, + 330, + 553, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 330, + 553, + 390 + ], + "type": "text", + "content": ". Our reconstruction method, depicted in Figure 3, consists of two main steps: (1) per-pixel reconstruction and (2) spatio-temporal propagation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 396, + 453, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 453, + 409 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 453, + 409 + ], + "type": "text", + "content": "5.1. Per-pixel Reconstruction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 415, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 449 + ], + "type": "text", + "content": "In this step, our goal is to estimate a Mueller matrix in a computationally efficient manner and being robust to noisy events." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "content": "First, we stack the vectors " + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t_k}" + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "content": ", forming a matrix " + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{K \\times 16}" + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "content": ". We then solve for the vectorized Mueller matrix " + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "content": " by minimizing the weighted least-squares problem:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 385, + 497, + 553, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 497, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 385, + 497, + 553, + 518 + ], + "type": "interline_equation", + "content": "\\underset {\\hat {\\mathbf {M}}} {\\text {m i n i m i z e}} \\| \\mathbf {W D B M} \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "51f528f9ac09827533cd9bac9adf9b6c1a4654b15c7018bc752c34fca999ad50.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = \\mathrm{diag}(\\mathbf{w})" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " is a diagonal matrix of weights " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = [w_1, \\ldots, w_K] \\in \\mathbb{R}^{1 \\times K}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": ", initialized as ones: " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "w_k = 1" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": ". The matrix " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = \\mathrm{diag}([\\Delta t_1, \\ldots, \\Delta t_K])" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " is a weight matrix that adjusts the influence of each event according to sampling density. This weighting ensures that more densely sampled events at small " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " do not dominate the model fitting. We solve Equation (8) using singular value decomposition (SVD) to obtain " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " while avoiding the trivial zero solution." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": "Second, we filter the reconstructed Mueller matrix " + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": " to ensure physical validity. Sensor noise and rapid scene dynamics can make solving Equation (8) challenging, often resulting in physically invalid Mueller matrices. These invalid matrices can trap the optimization in local minima. To address this, we apply Cloude's Mueller matrix filtering [16] to project the estimated matrix onto the space of" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21807" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 556, + 252 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 556, + 252 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 556, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 556, + 252 + ], + "type": "image", + "image_path": "1d257cb698b40f9e6bad5b05ddc08e647a7d1ccc9d6dea2e23c78b5de3f71fa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 265, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 54, + 265, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 555, + 310 + ], + "type": "text", + "content": "Figure 4. Synthetic data evaluation result. (a) The rendered images include two materials: blue silicone and brass. (b) The plot shows the error (mean absolute error) of the reconstructed Mueller matrix over the number of iterations. (c) Ground truth Mueller matrix. (d)&(e) Top: the reconstructed Mueller-matrix images for the SVD initialization and the full-stage, respectively. Pixels with insufficient event counts are visualized in white. Bottom: a plot of the differentiation of log intensity and fitted line with our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 195, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 195, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 195, + 335 + ], + "type": "text", + "content": "physically valid Mueller matrices:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 344, + 294, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 344, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 124, + 344, + 294, + 357 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} \\leftarrow \\operatorname {C l o u d e F i l t e r} (\\hat {\\mathbf {M}}), \\tag {9}", + "image_path": "b97009d631a400b41101d0a35755912096444837bec51d976324490f3aeb1e8a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 266, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 266, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 266, + 378 + ], + "type": "text", + "content": "where CloudeFilter " + }, + { + "bbox": [ + 55, + 366, + 266, + 378 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 55, + 366, + 266, + 378 + ], + "type": "text", + "content": " denotes the filtering function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "text", + "content": "Third, we recalculate the weight vector " + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "text", + "content": " using the physically valid Mueller matrix " + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 412, + 294, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 412, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 121, + 412, + 294, + 426 + ], + "type": "interline_equation", + "content": "w _ {k} = 1 / \\max \\left(\\left| \\mathbf {B} _ {t _ {k}} \\hat {\\mathbf {M}} \\right|, \\epsilon\\right), \\tag {10}", + "image_path": "06c9344f31c0d7d627f2647a2718e0e71990880a2af6f58bebb802d51aab255a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "text", + "content": " is a small positive constant to avoid division by zero sets as " + }, + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "text", + "content": " in our experiments. This weight reduces the influence of outlier events and enables robust estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "text", + "content": "We iterate these three steps—solving the weighted least-squares problem, applying physical validity filtering, and recomputing the weights—until convergence, resulting in the Mueller matrix estimate " + }, + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "text", + "content": " for each pixel " + }, + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "text", + "content": " and frame " + }, + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 471, + 296, + 543 + ], + "type": "text", + "content": ". We set the maximum number of iterations to 5 in our experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 550, + 291, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 291, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 291, + 563 + ], + "type": "text", + "content": "5.2. Spatio-temporal Propagation and Refinement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "content": "Here, we start by denoting the estimated Mueller matrix at a pixel " + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "content": " and frame " + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{xyf}" + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "content": " and the corresponding system matrix as " + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{xyf}" + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "content": ", and weight matrix as " + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{xyz}" + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "content": ". This second stage refines the Mueller matrix " + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{xyf}" + }, + { + "bbox": [ + 55, + 568, + 295, + 653 + ], + "type": "text", + "content": " estimated from the first stage using spatio-temporal neighboring pixels. This process is inspired by PatchMatchStereo from stereoscopic depth estimation [10, 11, 22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "First, we perform propagation. We consider neighboring pixels in both spatial and temporal axes, denoted by " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x', y', f') \\in \\mathcal{N}(x, y, f)" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ". The set " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " defines a spatio-temporal neighborhood extended from the red-black spatial-only neighborhood [22]. For each pixel at a frame" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 323, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 347 + ], + "type": "inline_equation", + "content": "(x,y,f)" + }, + { + "bbox": [ + 313, + 323, + 555, + 347 + ], + "type": "text", + "content": ", we update its Mueller matrix if a neighboring pixel " + }, + { + "bbox": [ + 313, + 323, + 555, + 347 + ], + "type": "inline_equation", + "content": "(x',y',f')" + }, + { + "bbox": [ + 313, + 323, + 555, + 347 + ], + "type": "text", + "content": " offers a lower error:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 354, + 400, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 354, + 400, + 368 + ], + "spans": [ + { + "bbox": [ + 320, + 354, + 400, + 368 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {x y f} \\leftarrow \\hat {\\mathbf {M}} _ {x ^ {\\prime} y ^ {\\prime} f ^ {\\prime}},", + "image_path": "faade2eb245b17f5c0d815a2c0cada45be6807463daca42fc55e3fd7991295cf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 370, + 553, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 370, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 320, + 370, + 553, + 385 + ], + "type": "interline_equation", + "content": "\\text {i f} \\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x y f} \\| _ {1} > \\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x ^ {\\prime} y ^ {\\prime} f ^ {\\prime}} \\| _ {1}. \\tag {11}", + "image_path": "3cfa3ba9c09947e168d10b00bfdb846f9bdcbf1bdafafe537a0af20895cc8ec4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 391, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 426 + ], + "type": "text", + "content": "Next, we refine the Mueller matrix by applying random perturbations and accepting the perturbation if it reduces the error:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 431, + 426, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 431, + 426, + 449 + ], + "spans": [ + { + "bbox": [ + 320, + 431, + 426, + 449 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {x y f} \\leftarrow \\hat {\\mathbf {M}} _ {x y f} ^ {\\text {p e r t u r b}}, \\text {w h e r e}", + "image_path": "7e3b2f8d3b967a97e6197a1bf40d6d387fdf9e26e03b5d90b2cff6c058947d4e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 449, + 553, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 449, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 321, + 449, + 553, + 467 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {x y f} ^ {\\text {p e r t u r b}} = \\operatorname {C l o u d e F i l t e r} \\left(\\hat {\\mathbf {M}} _ {x y f} \\odot (\\mathbf {1} + \\sigma \\mathbf {N})\\right), \\tag {12}", + "image_path": "7ba7fc07e556cec34e2cf33c6d54beb911a753b5ead6e160b9cf52e3d2ce4ca8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 468, + 553, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 468, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 321, + 468, + 553, + 485 + ], + "type": "interline_equation", + "content": "\\left\\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\hat {\\mathbf {M}} _ {x y f} \\right\\| _ {1} > \\left\\| \\mathbf {D} _ {x y z} \\mathbf {B} _ {x y f} \\mathbf {M} _ {x y f} ^ {\\text {p e r t u r b}} \\right\\| _ {1}, \\tag {13}", + "image_path": "75c489db4eab749bef4f84983313bda554f2f2917c26f3f53b8b11cf5567cb60.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{xyf}^{\\mathrm{perturb}}" + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "text", + "content": " is the perturbed Mueller matrix, " + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "text", + "content": " denotes element-wise multiplication, " + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "text", + "content": " is an all-ones matrix, " + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{N} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "text", + "content": " contains random values from a standard normal distribution, and " + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 491, + 554, + 600 + ], + "type": "text", + "content": " is a scalar set to 0.01. This multiplicative perturbation preserves the relative scale of the Mueller matrix elements, allowing larger elements to undergo proportionally larger adjustments while smaller elements change minimally. We apply Cloude's filtering to the perturbed Mueller matrix to ensure physical validity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 600, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 554, + 637 + ], + "type": "text", + "content": "We repeat the propagation and refinement steps iteratively for a fixed number of iterations set to 10 in our experiments to obtain the final Mueller matrix " + }, + { + "bbox": [ + 313, + 600, + 554, + 637 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{xyf}" + }, + { + "bbox": [ + 313, + 600, + 554, + 637 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 645, + 389, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 389, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 389, + 658 + ], + "type": "text", + "content": "6. Calibration" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": "We perform one-time calibration of the system parameters: contrast threshold " + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": " and QWP offset angles " + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "inline_equation", + "content": "(i_1" + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "inline_equation", + "content": "i_2)" + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": ". Calibration details are provided in the Supplementary Material." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21808" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 335, + 210 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 335, + 210 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 335, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 335, + 210 + ], + "type": "image", + "image_path": "9b9874e3c86ddcd40057fcf07f0f0e4a7726ce6886b40d2910e99a921e9b5067.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 218, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 218, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 218, + 555, + 243 + ], + "type": "text", + "content": "Figure 5. Assessment of reconstructed Mueller matrix on real data. (a) Evaluation with known optical elements. We show the corresponding mean squared errors (MSEs). (b) Measurement on an in-the-wild metal plate induces strong diagonal components." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 337, + 70, + 554, + 209 + ], + "blocks": [ + { + "bbox": [ + 337, + 70, + 554, + 209 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 554, + 209 + ], + "type": "image", + "image_path": "95c7017207b1055cbc2971ecf99283d25592a95faf33ddf9487348def26644c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 261, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 261, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 294, + 346 + ], + "type": "text", + "content": "Contrast Threshold We calibrate the per-pixel contrast threshold " + }, + { + "bbox": [ + 54, + 261, + 294, + 346 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 261, + 294, + 346 + ], + "type": "text", + "content": " of an event sensor by analyzing pixel responses under controlled illumination stimuli. We increase and decrease the LED intensity linearly over time, and the event camera directly captures the LED light, generating events. We then fit the contrast threshold " + }, + { + "bbox": [ + 54, + 261, + 294, + 346 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 261, + 294, + 346 + ], + "type": "text", + "content": " to the events per each pixel using Equation (6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 358, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 358, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 295, + 431 + ], + "type": "text", + "content": "QWP Offset Angle We obtain the offset angles of the fast axis of the QWP in our system: " + }, + { + "bbox": [ + 54, + 358, + 295, + 431 + ], + "type": "inline_equation", + "content": "i_1" + }, + { + "bbox": [ + 54, + 358, + 295, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 358, + 295, + 431 + ], + "type": "inline_equation", + "content": "i_2" + }, + { + "bbox": [ + 54, + 358, + 295, + 431 + ], + "type": "text", + "content": " by capturing another QWP sample with known fast axis. Given the known Mueller matrix of the reference QWP, we find the best offset angles that minimize the reconstruction error of Equation (6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 109, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 109, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 109, + 453 + ], + "type": "text", + "content": "7. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "text", + "content": "Implementation We implemented our reconstruction algorithm in " + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{C + + }" + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "text", + "content": " using OpenMP CPU parallelization. It takes 62 seconds for per-pixel reconstruction (Section 5.1), and 68 seconds for the propagation and refinement processes (Section 5.2), measured for processing " + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "inline_equation", + "content": "500\\times 500" + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "text", + "content": " -resolution event image sequence of 30 frames, with an average event rate of " + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "inline_equation", + "content": "153\\mathrm{MEv / s}" + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "text", + "content": ", on the AMD Ryzen 7 7800X3D 8-core processor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "Validation on Synthetic Data To evaluate our method, we render synthetic data in three steps. First, we mimic our capture system by replacing the event camera with an intensity camera in Mitsuba3 [29], and render an image sequence of a synthetic scene for each QWP angle with interval " + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\omega t = 0.01^{\\circ}" + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": ". We use the real-world polarimetric BRDFs to construct the synthetic scene [8]. Then, the rendered images are converted to event streams by simulating event-camera processing based on DVS-Voltmeter [34]. Last, we add Gaussian noise with a standard deviation of 0.5 and replace " + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": " of true events with values sampled from a Gaussian distribution with a standard deviation of 5.0. Figure 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "content": "shows the synthetic data and reconstructed Mueller matrix using our method. While initial SVD-based estimation suffers from noise with a reconstruction error of 0.11, our full reconstruction improves accuracy with a reconstruction error of 0.04." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "text", + "content": "Validation on Real Data Figure 5(a) shows that our reconstruction closely matches the corresponding pseudo ground truth of real-world samples: air, linear polarizer " + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "inline_equation", + "content": "(0^{\\circ}" + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "inline_equation", + "content": "45^{\\circ})" + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "text", + "content": ", quarter-wave-plate " + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "inline_equation", + "content": "(0^{\\circ}" + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "inline_equation", + "content": "45^{\\circ})" + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "text", + "content": ". Figure 5(b) shows the result on an in-the-wild sample: a metal plate. The strong diagonal components in this result indicate the preservation of polarization, which is characteristic of reflections from metal surfaces. Notably, these Mueller matrix images are captured in just " + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "inline_equation", + "content": "33\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 340, + 556, + 473 + ], + "type": "text", + "content": ", significantly faster than the conventional ellipsometer, which requires several minutes due to its frame-based imaging principle." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 556, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 600 + ], + "type": "text", + "content": "Photoelasticity of Transparent Gelatine Photoelasticity is an optical property whereby dielectric materials exhibit birefringence under deformation. Imaging photoelasticity has applications in mechanical stress and material analysis [39, 47]. Analyzing Mueller-matrix images reveals such stress distributions even in transparent objects [18]. Our system enables observing photoelasticity at video rates. Figure 6 shows Mueller-matrix images of a gelatine disk, revealing stress-dependent fringe patterns." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Transparent Tape Detection Detecting transparent objects is challenging for conventional cameras. Sticky tape, a common transparent material used for sealing cardboard, exhibits birefringent properties due to the molecular structure of its stretched plastic. Figure 7 shows the measurement result of sticky tape on a cardboard box. The reconstructed Mueller matrix shows the birefringent property, and we can clearly recognize the tape region. This" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21809" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 208, + 185 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 208, + 185 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 208, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 208, + 185 + ], + "type": "image", + "image_path": "a2a796269c4579178bf5d1109ea761d6a9520c24de15f1bbf640c151a7b9397d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 186, + 174, + 197 + ], + "lines": [ + { + "bbox": [ + 91, + 186, + 174, + 197 + ], + "spans": [ + { + "bbox": [ + 91, + 186, + 174, + 197 + ], + "type": "text", + "content": "(a) Measurement setup" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 69, + 381, + 184 + ], + "blocks": [ + { + "bbox": [ + 211, + 69, + 381, + 184 + ], + "lines": [ + { + "bbox": [ + 211, + 69, + 381, + 184 + ], + "spans": [ + { + "bbox": [ + 211, + 69, + 381, + 184 + ], + "type": "image", + "image_path": "6dfd408613e57492f96aaceffb9b904fddee952a68e00ea26fa7fcb2cc5a87b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 185, + 336, + 194 + ], + "lines": [ + { + "bbox": [ + 265, + 185, + 336, + 194 + ], + "spans": [ + { + "bbox": [ + 265, + 185, + 336, + 194 + ], + "type": "text", + "content": "Before applying force" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 69, + 554, + 184 + ], + "blocks": [ + { + "bbox": [ + 383, + 69, + 554, + 184 + ], + "lines": [ + { + "bbox": [ + 383, + 69, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 383, + 69, + 554, + 184 + ], + "type": "image", + "image_path": "69de5e4f17a7b054dd2a6372a85846bc76ba1f987393359ecf5adcf39a9de424.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 185, + 500, + 194 + ], + "lines": [ + { + "bbox": [ + 434, + 185, + 500, + 194 + ], + "spans": [ + { + "bbox": [ + 434, + 185, + 500, + 194 + ], + "type": "text", + "content": "After applying force" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 196, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 321, + 196, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 321, + 196, + 553, + 206 + ], + "type": "text", + "content": "(b) Reconstructed Mueller matrix -1 +1" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 259, + 184, + 354 + ], + "blocks": [ + { + "bbox": [ + 55, + 216, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 555, + 251 + ], + "type": "text", + "content": "Figure 6. Photoelasticity analysis. (a) Experimental setup for measuring a gelatine disk in transmission mode. We gradually apply force for the gelatine disk to observe changes in photoelastic properties. (b) Reconstructed Mueller matrix images with different forces, revealing complex stress-dependent polarimetric patterns. Notably, as force increases, the Mueller matrix image shows denser fringe patterns." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 259, + 184, + 354 + ], + "lines": [ + { + "bbox": [ + 56, + 259, + 184, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 184, + 354 + ], + "type": "image", + "image_path": "402acf9ad9adc834b4606f7f0399564f92aa2e629080718b1ae390ac3ea33f80.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 354, + 206, + 365 + ], + "lines": [ + { + "bbox": [ + 146, + 354, + 206, + 365 + ], + "spans": [ + { + "bbox": [ + 146, + 354, + 206, + 365 + ], + "type": "text", + "content": "(a) Target object" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 193, + 261, + 286, + 353 + ], + "blocks": [ + { + "bbox": [ + 193, + 261, + 286, + 353 + ], + "lines": [ + { + "bbox": [ + 193, + 261, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 193, + 261, + 286, + 353 + ], + "type": "image", + "image_path": "962460144edd04345c48b98a143598e506e502ece3e65a9d4dd3ba5ff234e69f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 56, + 365, + 145, + 437 + ], + "blocks": [ + { + "bbox": [ + 56, + 365, + 145, + 437 + ], + "lines": [ + { + "bbox": [ + 56, + 365, + 145, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 145, + 437 + ], + "type": "image", + "image_path": "9decf049b0c3c899c1ed388802f1744dbaee633dd22b1c4f1e26c9e8b5f37499.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 57, + 440, + 145, + 511 + ], + "blocks": [ + { + "bbox": [ + 57, + 440, + 145, + 511 + ], + "lines": [ + { + "bbox": [ + 57, + 440, + 145, + 511 + ], + "spans": [ + { + "bbox": [ + 57, + 440, + 145, + 511 + ], + "type": "image", + "image_path": "bd815b372e8f21768bac318cacb8c14d957b5da6ecccfcc7f712aef0e46840dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 512, + 129, + 522 + ], + "lines": [ + { + "bbox": [ + 71, + 512, + 129, + 522 + ], + "spans": [ + { + "bbox": [ + 71, + 512, + 129, + 522 + ], + "type": "text", + "content": "(b) Comparison" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 148, + 365, + 295, + 511 + ], + "blocks": [ + { + "bbox": [ + 148, + 365, + 295, + 511 + ], + "lines": [ + { + "bbox": [ + 148, + 365, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 148, + 365, + 295, + 511 + ], + "type": "image", + "image_path": "c8b1f8a35d10105438711d3063a39782856bd8e90e61981fefe0f2c36c3dfe8c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 512, + 282, + 521 + ], + "lines": [ + { + "bbox": [ + 161, + 512, + 282, + 521 + ], + "spans": [ + { + "bbox": [ + 161, + 512, + 282, + 521 + ], + "type": "text", + "content": "(c) Reconstructed Mueller matrix" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 183, + 522, + 293, + 532 + ], + "blocks": [ + { + "bbox": [ + 183, + 522, + 293, + 532 + ], + "lines": [ + { + "bbox": [ + 183, + 522, + 293, + 532 + ], + "spans": [ + { + "bbox": [ + 183, + 522, + 293, + 532 + ], + "type": "image", + "image_path": "f2f7ae8724d6bbe20d66dd441421974b8003f13ce8b0fe6337fa8710283cd790.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 540, + 295, + 585 + ], + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 585 + ], + "type": "text", + "content": "Figure 7. Transparent tape detection. (a) The target object is transparent sticky tape on a box. (b)&(c) The Reconstructed Mueller matrix reveals the tape region, which is difficult to see with a conventional RGB camera." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "type": "text", + "content": "demonstration shows the potential for inspecting dynamically moving sealed boxes in automated industrial processes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Dynamic Human Capture Ellipsometers, as a comprehensive polarization imaging technique for human capture, can reveal hidden properties of the human face and hair that are difficult to analyze in conventional imaging and non-ellipsometric polarization imaging [2, 19]. In Figure 8, we" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 262, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 358 + ], + "type": "text", + "content": "demonstrate the Mueller-matrix reconstruction of dynamic facial expressions and hair movements, revealing intricate polarization properties. The reconstructed Mueller matrix for the face exhibits strong diagonal components associated with specular reflection, while the non-diagonal regions display weakly polarized reflection with a dependency on surface normal along the face edges. The hair result shows polarization property in the specular highlight regions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 384, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 479 + ], + "type": "text", + "content": "HDR Mueller-matrix Imaging Our event-based ellipsometer enables Mueller-matrix imaging of HDR scenes. Figure 9 shows reconstruction results on a scene with specular and dark diffuse reflections. While conventional ellipsometers require multiple exposures to prevent overexposure, our method captures the Mueller matrix in HDR scenes at 30 fps without the need for additional measurement." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 497, + 383, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 383, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 383, + 509 + ], + "type": "text", + "content": "8. Discussion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "text", + "content": "Normalized Mueller-Matrix Imaging Normalized Mueller matrix has been extensively used as it provides rich scene information [5, 7, 18]. Since event cameras only detect intensity changes, not intensity itself, our method also reconstructs a normalized version of Mueller matrix per pixel, " + }, + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{M} / \\mathbf{M}(0,0)" + }, + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "text", + "content": ". We leave reconstructing full Mueller-matrix as a future work that might be accomplished using an event-intensity hybrid imaging system." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "End-to-end Real-time Pipeline Although our capture speed is fast, our current Mueller-matrix reconstruction relies solely on CPU acceleration, resulting in a non-realtime pipeline from capture to reconstruction. Implementing GPU acceleration could enable on-the-fly capture, reconstruction, and visualization [22, 48]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21810" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 556, + 279 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 556, + 279 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 556, + 279 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 556, + 279 + ], + "type": "image", + "image_path": "1604f053161511fba6b9020f246bb5f1f1bdda08910b9714e0121bdf77abd7ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 286, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 555, + 309 + ], + "type": "text", + "content": "Figure 8. Mueller matrix acquisition for capturing dynamic human (a) face and (b) hair, demonstrating the capture of diffuse and specular polarimetric responses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 319, + 175, + 497 + ], + "blocks": [ + { + "bbox": [ + 56, + 319, + 175, + 497 + ], + "lines": [ + { + "bbox": [ + 56, + 319, + 175, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 175, + 497 + ], + "type": "image", + "image_path": "63a874a97f0afd956c1560d17b1e60692542d22c446ecbcd3e7075f33b7d28e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 498, + 146, + 509 + ], + "lines": [ + { + "bbox": [ + 86, + 498, + 146, + 509 + ], + "spans": [ + { + "bbox": [ + 86, + 498, + 146, + 509 + ], + "type": "text", + "content": "(a) Target object" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 518, + 295, + 562 + ], + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 562 + ], + "type": "text", + "content": "Figure 9. Mueller matrix measurement for a HDR scene. (a) The target scene contains regions with both strong specular reflection and black diffuse reflection. (b) Our proposed method achieves an accurate Mueller-matrix image for the HDR scene in just " + }, + { + "bbox": [ + 55, + 518, + 295, + 562 + ], + "type": "inline_equation", + "content": "33~\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 518, + 295, + 562 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 319, + 296, + 497 + ], + "blocks": [ + { + "bbox": [ + 175, + 319, + 296, + 497 + ], + "lines": [ + { + "bbox": [ + 175, + 319, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 175, + 319, + 296, + 497 + ], + "type": "image", + "image_path": "9c13ff77dab3feb86078999efb4ef9d13eac5ba120f60f18b4e6589db59b5ec5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 498, + 294, + 508 + ], + "lines": [ + { + "bbox": [ + 173, + 498, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 173, + 498, + 294, + 508 + ], + "type": "text", + "content": "(b) Reconstructed Mueller matrix" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Event Bandwidth and Sensitivity The ideal setting of an event camera is to have a low contrast threshold and a short refractory period, to capture subtle intensity changes. However, this produces many events, exceeding the transmission bandwidth, resulting in lost or delayed events [21]. In contrast, when the contrast threshold is high, weakly polarized phenomena cannot be detected as events. Given this tradeoff, we defined a region of interest (ROI) in the event sensor and empirically configured the camera settings to find the best parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 321, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 381 + ], + "type": "text", + "content": "DC Motor Synchronization For faster rotation speed and precise synchronization between two motors, it is a viable alternative to use brushless DC motors, also known as an electronically commutated (EC) motors. However, this increases system building costs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 412, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 483 + ], + "type": "text", + "content": "Gradient Descent Gradient descent in an automatic differentiation framework offers an alternative approach for Mueller-matrix reconstruction. However, the number of events differs across pixels, which complicates the construction of a dense tensor for efficient computation on modern GPUs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 503, + 388, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 388, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 388, + 516 + ], + "type": "text", + "content": "9. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 526, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 622 + ], + "type": "text", + "content": "We have introduced Event Ellipsometer, a Mueller-matrix video imaging method that combines an event camera and a light source with fast-rotating QWPs. Our image formation model, calibration, and reconstruction method enable robust Mueller-matrix imaging at 30 fps. We validate our method on synthetic and real data, and demonstrate Mueller-matrix imaging on photoelasticity, dynamic human hair and face capture, HDR imaging, and transparent tape detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532, 2022R1A6A1A03052954, RS-2024-00437866), and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21811" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Hans Arwin. Application of ellipsometry techniques to biological materials. Thin Solid Films, 519(9):2589-2592, 2011. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "text", + "content": "[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "type": "text", + "content": "[3] RMA Azzam. Photopolarimetric measurement of the mueller matrix by fourier analysis of a single detected signal. Optics Letters, 2(6):148-150, 1978. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 237 + ], + "type": "text", + "content": "[4] Rasheed MA Azzam. Stokes-vector and mueller-matrix polarimetry. JOSA A, 33(7):1396-1408, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 272 + ], + "type": "text", + "content": "[5] Seung-Hwan Baek and Felix Heide. Polarimetric spatiotemporal light transport probing. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "text", + "content": "[6] Seung-Hwan Baek and Felix Heide. All-photon polarimetric time-of-flight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17876-17885, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 295, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 351 + ], + "type": "text", + "content": "[7] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM Trans. Graph., 37(6):268, 2018. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "type": "text", + "content": "[8] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph., 39(4):139, 2020. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 398, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 295, + 441 + ], + "type": "text", + "content": "[9] Seung-Hwan Baek, Ramesh Raskar, Jinwei Ye, Akshit Dave, Achuta Kadambi, and Huajin Chen. Polarization-based visual computing. In ACM SIGGRAPH 2023 Courses, pages 1-1, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "type": "text", + "content": "[10] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "type": "text", + "content": "[11] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In *Bmvc*, pages 1-11, 2011. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 523, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 295, + 577 + ], + "type": "text", + "content": "[12] Qizhi Cao, Min Jiang, Chenling Jia, Siyue Jiang, Jing Zhang, Baoli Yao, Mingwu Jin, Edward Dehoog, Lian Duan, Huahua Wang, et al. Snapshot imaging mueller matrix polarimeter using modified savart polariscopes. Applied Optics, 62(8):2124-2129, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 623 + ], + "type": "text", + "content": "[13] Dorian Chan, Mark Sheinin, and Matthew O'Toole. Spincam: High-speed imaging via a rotating point-spread function. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10789-10799, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "text", + "content": "[14] Tongbo Chen, Hendrik PA Lensch, Christian Fuchs, and Hans-Peter Seidel. Polarization and phase-shifting for 3d scanning of translucent objects. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[15] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proceedings of the IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 14760-14770, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "type": "text", + "content": "[16] Shane R Cloude. Conditions for the physical realizability of matrix operators in polarimetry. In *Polarization Considerations for Optical Systems II*, pages 177–187. SPIE, 1990. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "type": "text", + "content": "[17] Edward Collett. Field guide to polarization. Spie Bellingham, 2005. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "text", + "content": "[18] Akshat Dave, Tianyi Zhang, Aaron Young, Ramesh Raskar, Wolfgang Heidrich, and Ashok Veeraraghavan. Nest: Neural stress tensor tomography by leveraging 3d photoelasticity. arXiv preprint arXiv:2406.10212, 2024. 1, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "type": "text", + "content": "[19] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 262, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 553, + 282 + ], + "type": "text", + "content": "[20] Hiroyuki Fujiwara. Spectroscopic ellipsometry: principles and applications. John Wiley & Sons, 2007. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 284, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 284, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 553, + 347 + ], + "type": "text", + "content": "[21] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 4, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "type": "text", + "content": "[22] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE international conference on computer vision, pages 873-881, 2015. 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 394, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 555, + 426 + ], + "type": "text", + "content": "[23] Nirmalya Ghosh and I Alex Vitkin. Tissue polarimetry: concepts, challenges, applications, and outlook. Journal of biomedical optics, 16(11):110801-110801, 2011. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "type": "text", + "content": "[24] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20616-20625, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "text", + "content": "[25] Michael Hawks and Michael Dexter. Event-based imaging polarimeter. Optical Engineering, 61(5):053101-053101, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "type": "text", + "content": "[26] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "text", + "content": "[27] Inseung Hwang, Daniel S Jeon, Adolfo Munoz, Diego Gutierrez, Xin Tong, and Min H Kim. Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 615, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 668 + ], + "type": "text", + "content": "[28] Tomoki Ichikawa, Shohei Nobuhara, and Ko Nishino. Spiders: Structured polarization for invisible depth and reflectance sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25077-25085, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[29] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsubishi 3 renderer, 2022. https://mitsuba-renderer.org. 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21812" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[30] GE Jellison, Donovan N Leonard, Lawrence M Anovitz, Chad M Parish, Eliot D Specht, and TM Rosseel. Crystallographic orientation of uniaxial calcite and dolomite determined using reflection generalized ellipsometry. Journal of Applied Physics, 124(22), 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 172 + ], + "type": "text", + "content": "[31] Achuta Kadambi, Vage Taamazyan, Boxin Shi, and Ramesh Raskar. Polarized 3d: High-quality depth sensing with polarization cues. In Proceedings of the IEEE international conference on computer vision, pages 3370-3378, 2015. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "type": "text", + "content": "[32] Agastya Kalra, Vage Taamazyan, Supreeth Krishna Rao, Kartik Venkataraman, Ramesh Raskar, and Achuta Kadambi. Deep polarization cues for transparent object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8602-8611, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 295, + 282 + ], + "type": "text", + "content": "[33] Michael W Kudenov, Michael J Escuti, Nathan Hagen, Eustace L Dereniak, and Kazuhiko Oka. Snapshot imaging mueller matrix polarimeter using polarization gratings. Optics letters, 37(8):1367-1369, 2012. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 295, + 327 + ], + "type": "text", + "content": "[34] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In European Conference on Computer Vision, pages 578-593. Springer, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 360 + ], + "type": "text", + "content": "[35] Ryota Maeda and Shinsaku Hiura. Polarimetric light transport analysis for specular inter-reflection. IEEE Transactions on Computational Imaging, 10:876-887, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "text", + "content": "[36] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 982-987. IEEE, 2003. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "type": "text", + "content": "[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 295, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 295, + 503 + ], + "type": "text", + "content": "[38] David A Ramsey and Kenneth C Ludema. The influences of roughness on film thickness measurements by mueller matrix ellipsometry. Review of scientific instruments, 65(9):2874-2881, 1994. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 295, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 558 + ], + "type": "text", + "content": "[39] Michele Scafidi, Giuseppe Pitarresi, Andrea Toscano, Giovanni Petrucci, Sabina Alessi, and Augusto Ajovalasit. Review of photoelastic image analysis applied to structural birefringent materials: glass and polymers. Optical Engineering, 54(8):081206-081206, 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "text", + "content": "[40] Yoav Y Schechner and Shree K Nayar. Uncontrolled modulation imaging. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-II. IEEE, 2004. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "text", + "content": "[41] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, pages 814-819. IEEE, 1999. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[42] Yoav Y Schechner, Srinivasa G Narasimhan, and Shree K Nayar. Instant dehazing of images using polarization. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I. IEEE, 2001. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 397 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 73, + 554, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 554, + 137 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 554, + 137 + ], + "type": "text", + "content": "[43] Dominik Scheuble, Chenyang Lei, Seung-Hwan Baek, Mario Bijelic, and Felix Heide. Polarization wavefront lidar: Learning large scene reconstruction from polarized wavefronts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21241-21250, 2024. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 139, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 139, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 139, + 554, + 172 + ], + "type": "text", + "content": "[44] Matthew H Smith. Optimization of a dual-rotating-retarder mueller matrix polarimeter. Applied optics, 41(13):2488-2493, 2002. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 172, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 172, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 315, + 172, + 554, + 229 + ], + "type": "text", + "content": "[45] Tsuyoshi Takatani, Yuzuha Ito, Ayaka Ebisu, Yinqiang Zheng, and Takahito Aoto. Event-based bispectral photometry using temporally modulated illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15647, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 229, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 229, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 229, + 554, + 262 + ], + "type": "text", + "content": "[46] Tali Treibitz and Yoav Y Schechner. Active polarization scattering. IEEE transactions on pattern analysis and machine intelligence, 31(3):385-399, 2008. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 262, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 262, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 315, + 262, + 554, + 306 + ], + "type": "text", + "content": "[47] Li Wang, Yang Ju, Heping Xie, Guowei Ma, Lingtao Mao, and Kexin He. The mechanical and photoelastic properties of 3d printable stress-visualized materials. Scientific reports, 7(1):10918, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 308, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 554, + 352 + ], + "type": "text", + "content": "[48] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 353, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 353, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 554, + 397 + ], + "type": "text", + "content": "[49] Aun Zaidi, Noah A Rubin, Maryna L Meretska, Lisa W Li, Ahmed H Dorrah, Joon-Suh Park, and Federico Capasso. Metasurface-enabled single-shot and complete mueller matrix imaging. Nature Photonics, pages 1-9, 2024. 1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21813" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_content_list.json b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8de1f1f29d638efa8acfb6f709bca7798dc718ca --- /dev/null +++ b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_content_list.json @@ -0,0 +1,2028 @@ +[ + { + "type": "text", + "text": "Event fields: Capturing light fields at high speed, resolution, and dynamic range", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyuan Qu Dartmouth College", + "bbox": [ + 114, + 181, + 271, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zihao Zou \nUNC Chapel Hill", + "bbox": [ + 307, + 181, + 449, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vivek Boominathan Rice University", + "bbox": [ + 485, + 181, + 647, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Praneeth Chakravarthula \nUNC Chapel Hill", + "bbox": [ + 683, + 181, + 880, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adithya Pediredla Dartmouth College", + "bbox": [ + 419, + 223, + 573, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/098c5afda1924c14b4ad7c758b3330a88746163e43b5ef86b36c43b82da882a7.jpg", + "image_caption": [ + "Figure 1. We introduce Event Fields, a novel framework that leverages event cameras to capture light fields at high speed, resolution, and dynamic range using innovative optical designs. Event fields integrate the unique advantages of event sensing—such as low latency, data efficiency, and power-saving capabilities—with the detailed angular information provided by light field sensing. We demonstrate the versatility of the event fields through three application scenarios: slow motion refocus (SloMoRF), instant true depth estimation, and high dynamic range (HDR) light field imaging." + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 390, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/542431f7ab310f6ad4c9d2f7d41f5a09f622f56f30b307a044ff5b6d81ef728c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 273, + 903, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 536, + 326, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras, which feature pixels that independently respond to changes in brightness, are becoming increasingly popular in high-speed applications due to their lower latency, reduced bandwidth requirements, and enhanced dynamic range compared to traditional frame-based cameras. Numerous imaging and vision techniques have leveraged event cameras for high-speed scene understanding by capturing high-framerate, high-dynamic range videos, primarily utilizing the temporal advantages inherent to event cameras. Additionally, imaging and vision techniques have utilized the light field—a complementary dimension to temporal information—for enhanced scene understanding.", + "bbox": [ + 88, + 568, + 482, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose \"Event Fields\", a new approach that utilizes innovative optical designs for event cameras to capture light fields at high speed. We develop the underlying mathematical framework for Event Fields and introduce two foundational frameworks to capture them practically: spatial multiplexing to capture temporal derivatives and temporal multiplexing to capture angular derivatives. To realize these, we design two complementary optical setups—one using a kaleidoscope for spatial multiplexing and another using a galvanometer for temporal multiplexing. We", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "evaluate the performance of both designs using a custom-built simulator and real hardware prototypes, showcasing their distinct benefits. Our event fields unlock the full advantages of typical light fields—like post-capture refocusing and depth estimation—now supercharged for high-speed and high-dynamic range scenes. This novel light-sensing paradigm opens doors to new applications in photography, robotics, and AR/VR, and presents fresh challenges in rendering and machine learning.", + "bbox": [ + 511, + 537, + 906, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 691, + 643, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Capturing high-speed visual phenomena in real-time is essential for advancing science and engineering. Traditional high-speed frame cameras like Phantom, Photron, and IDT can achieve impressive frame rates (up to 100,000 fps) but are bulky, have finite capture duration, and are expensive (exceeding $100,000), restricting accessibility. In contrast, event cameras offer a more affordable and efficient solution with high temporal resolution (down to $1\\mu \\mathrm{s}$ ) and continuous data acquisition, suitable for real-time applications. Unlike traditional cameras, which require substantial bandwidth (e.g., 100,000 fps, HD capture requires 100 GBps), event cameras capture only changes in the scene using asyn", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26910", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chronous pixel readouts, reducing bandwidth demands and enabling continuous streaming. This approach has recently been harnessed for high-speed applications, including optical flow, object tracking, high-dynamic range imaging, structured light scanning, and photometric stereo [6, 10].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we extend event camera capabilities by introducing angular capture, allowing for post-capture manipulation of high-speed video—a previously unexplored functionality. Capturing angular radiance information is known as light field imaging [24, 25, 53], enabling features such as post-capture refocusing and depth estimation. We propose \"Event Fields\" that record radiance derivatives across angular and temporal dimensions, extending light field imaging principles to high-speed, high-dynamic-range scenarios (see Fig. 1). To capture these Event Fields, we introduce two complementary hardware systems: a kaleidoscope-based design and a galvanometer-based design, each with unique advantages for this purpose.", + "bbox": [ + 89, + 170, + 482, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The kaleidoscope design (see Fig. 3), inspired by Manakov et al. [30], captures angle-resolved time derivatives by placing a rectangular kaleidoscope in the optical path. This approach allows for capturing angular information affordably and simply at $3 \\times 3$ views. The galvanometer design (see Fig. 4) uses a galvanometer positioned in the optical path to capture angular derivatives directly. By controlling the mirrors at $250\\mathrm{Hz}$ , the incoming rays are shifted along a curve, enabling the capture of both angular and temporal derivatives. As angular derivatives are sparse [21, 56], this design does not significantly increase bandwidth requirements. Table 1 compares the relative advantages of these two hardware approaches.", + "bbox": [ + 89, + 371, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To systematically evaluate the relative merits of the hardware designs, we build a Blender plugin to render event-fields with physics based light simulation. We also build hardware imaging prototypes with kaleidoscope and galvanometer, and algorithms to demonstrate new imaging capabilities in challenging environments. These include post-capture refocusing on dynamic scenes (e.g., a rotating fan at 480 RPM), focus tracking of nef darts traveling at $20\\mathrm{m / s}$ , high-speed depth reconstruction, and high-resolution, high dynamic range light fields.", + "bbox": [ + 89, + 571, + 482, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. To state explicitly, the manuscript makes the following contributions:", + "bbox": [ + 89, + 727, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Introduces event fields; spatial and temporal multiplexing techniques to capture them.", + "- Builds a physics-based simulator and two hardware prototypes to systematically evaluate the merits and limitations of the multiplexing techniques.", + "- New applications and related algorithms, including slow motion refocus (SloMoRF), true-depth estimation, and high dynamic range light field imaging.", + "- Open source simulator, data, and processing codes [43]." + ], + "bbox": [ + 89, + 762, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/1fe09fcc3e7f30b98165aea82225a7390b67c2c7a9d09c8c41755ba1757cf13c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KaleidoscopeGalvanometer
DerivativesTemp.Ang. + Temp.
Recon. PerformanceScene dep.Scene independ.
ColorYesNo
Static LF CaptureNoYes
Dynamic LF captureYesYes
Refocus FidelityChoppySmooth
Bandwidth RequiredLowHigh
", + "bbox": [ + 519, + 89, + 901, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Relative advantages of kaleidoscope vs. galvanometer design for capturing event fields.", + "bbox": [ + 511, + 229, + 903, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 265, + 653, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Light field sensing: One of the first approaches to light field capture [54] involves using multiple sensors to gather angular information from various perspectives. However, this setup is bulky, costly, and requires high bandwidth communication. An alternative is time-sequential capture, which mounts a sensor on a mechanical gantry to capture the light field of a static scene [12, 25]. Rather than moving the sensor itself, a mirror can be used to redirect the view. Ihrke et al. [19] proposed using a movable planar mirror to shift the view temporally. However, this method requires individual calibration for each image, adding complexity. Fujii et al. [8] uses a galvanometer and pairs it with a high-speed camera. Inspired by their approach, we developed a similar design, precisely controlling the galvanometer's movement. A key distinction is that our galvanometer operates continuously, while their system selects discrete positions from a lookup table, making it slower.", + "bbox": [ + 511, + 291, + 906, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to time-sequential capture, angular information can also be measured with spatial multiplexing. One way to achieve this is through a microlens array [39], which has been widely adopted in commercial light field cameras like Raytrix [44] and Lytro [28]. While effective, microlens arrays are challenging to manufacture and install. Taguchi et al. [48] suggested using a mirror ball to expand the camera's field of view, though this setup results in the camera being visible in the mirror ball's reflection. Alternatively, Manakov et al. [30] introduced a kaleidoscope-like mirror array that can be attached to a standard camera without modification. Inspired by this design, we adapted it to convert a conventional event camera into a device capable of capturing the event field.", + "bbox": [ + 511, + 551, + 908, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event camera designs: From the first neuromorphic sensor by Mahowald et al. [29] to the recent Sony IMX636, event sensing saw exponential growth. We briefly review designs most relevant to our manuscript and refer the readers to recent surveys [6, 10]. Event cameras are regularly used in robotics and autonomous navigation for optical flow computation [2, 11, 52], feature tracking [49], and motion segmentation [34, 57], but event cameras cannot detect edges in the direction of the camera motion. Inspired by microsac", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26911", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cades in human eye movement, He et al. [17] placed a rotating wedge mask before the event camera aperture to capture edges in all directions. Similar to them, our galvanometer design detects edges oriented in all directions, independent of camera and object motion, as shown in Fig. 6. Our galvanometer design also finds inspiration from the designs of Mugalikar et al. [37], who used a rotating polarizer in the aperture plane to capture high-fidelity polarization information for shape-from-polarization and Yu et al. [55], who used a rotating light source and event camera for fast photometric stereo.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Habuchi et al. [16] captured light fields for static scenes by placing an LCOS device in the aperture plane of event cameras and using CNN to recover the light field. While the technique is quite novel, it is limited to static scenes due to the LCOS response time (22 ms). Guo et al. [14] placed a microlens array (MLA) on the event sensor to capture the light field for dynamic scenes and showed microscopic imaging applications with their prototype. We built our prototypes for macroscopic applications, and our kaleidoscope design is functionally same as MLA design. However, the galvanometer design can capture angular information even for static scenes without trading off the spatial resolution like Guo et al. [14].", + "bbox": [ + 89, + 257, + 482, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event processing algorithms: Event cameras capture scene information at higher temporal resolution (in $\\mu$ s) and dynamic range (140 dB) compared to standard cameras (60 dB) [10]. These features enable solutions to challenges where standard frame-based sensors struggle, such as high-speed motion estimation [9], high frame rate reconstruction [45], and high dynamic range imaging [45]. Besides, event cameras and standard frame-based cameras are complementary; event cameras provide high-speed, sparse data, while standard cameras capture color and fine details. By fusing these two types of sensors, algorithms like Timelens [50] and Timelens++ [51] effectively interpolate RGB frames using event information. We adapt these event processing algorithms for processing event field data.", + "bbox": [ + 89, + 454, + 482, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other event-based fusions: Apart from RGB cameras, other computational imaging systems can also be made fast by fusing with event sensors [58]. Event-based structured light sensing techniques [18, 23, 31, 32, 35, 47] reconstruct 3D scenes at high speed by replacing standard intensity sensors with event cameras. Event cameras are also used for foveated depth sensing [36] to scan moving objects at a higher spatial resolution. Muglikar et al. [38] combined single-photon avalanche diodes (SPADs) with event cameras to achieve high-speed, low-light, low-bandwidth imaging capabilities.", + "bbox": [ + 89, + 665, + 482, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Technical Background", + "text_level": 1, + "bbox": [ + 89, + 844, + 302, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will briefly review the image formation models for light field and event cameras that are useful in developing the", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mathematical models for event fields.", + "bbox": [ + 513, + 90, + 761, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Light Field", + "text_level": 1, + "bbox": [ + 513, + 114, + 635, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The plenoptic model of light $L(\\vec{x}, \\vec{\\omega}, \\lambda, t)$ is a comprehensive representation that captures all the dimensions of light emitted by a scene [1, 33]. Here, $\\vec{x} = (x, y, z)$ represents the spatial coordinates of the observer through which the light passes; $\\vec{\\omega} = (\\omega_x, \\omega_y)$ are the angular coordinates; $\\lambda$ is the wavelength, and $t$ is the time. At a given time, the measurement of a monochromatic light is 5D function. Levoy and Hanrahan [25] and Gortler et al. [12] showed that the 5D representation has redundancies and have reduced it to 4D by assuming that the light field is measured in free space. Under this assumption, as the radiance of a ray remains a constant in a straight line, the static monochromatic light field is a 4D function $L(\\vec{u}, \\vec{s})$ , where $\\vec{u} = (u, v)$ is a point on the virtual sensor and $\\vec{s} = (s, t)$ is a point on the virtual aperture.", + "bbox": [ + 511, + 137, + 903, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To synthesize an image from a given light field, we integrate the light field over the aperture as:", + "bbox": [ + 511, + 364, + 903, + 395 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, y) = \\int_ {\\vec {s}} L (\\vec {u}, \\vec {s}) \\mathrm {d} \\vec {s}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 405, + 903, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Light field allows us to change the focus plane after imaging by computationally changing the distance between the sensor and aperture planes [3, 25]. If $d_0$ is the original distance between the planes and $d$ is the distance required to focus at a given depth, the light field that focuses at $d$ will be $L_{d}(\\vec{u},\\vec{s}) = L(\\vec{u} +(d_{0} / d)\\vec{s},\\vec{s})$ . The refocused image can be synthesized by integrating angles, as shown in Eq. (1).", + "bbox": [ + 511, + 446, + 903, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.Event Camera", + "text_level": 1, + "bbox": [ + 513, + 561, + 661, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An event camera asynchronously records the brightness $(B\\dot{=} \\log (\\int_{\\vec{\\omega}}Ld\\vec{\\omega}))$ change event for each pixel independently [4, 26, 29, 42]. Specifically, in an ideal and noise-free scenario, a pixel $\\mathbf{x} = (x_k,y_k)^T$ will be triggered at time $t_k$ as soon as the brightness change reaches a predefined threshold $C$ since the last triggered event. The polarity $p_k\\in \\{+1, - 1\\}$ , which is the sign of the brightness change, will be recorded for each trigger. Mathematically,", + "bbox": [ + 511, + 584, + 903, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta B \\left(\\mathbf {x} _ {k}, t _ {k}\\right) \\doteq B \\left(\\mathbf {x} _ {k}, t _ {k}\\right) - B \\left(\\mathbf {x} _ {k}, t _ {k} - \\Delta t _ {k}\\right) = p _ {k} C\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 718, + 885, + 733 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial B \\left(\\mathbf {x} _ {k} , t _ {k}\\right)}{\\partial t} \\approx \\frac {p _ {k} C}{\\Delta t _ {k}} \\Rightarrow \\frac {\\partial L \\left(\\mathbf {x} _ {k} , t _ {k}\\right)}{\\partial t} \\approx L \\frac {p _ {k} C}{\\Delta t _ {k}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 737, + 903, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, event cameras capture a high-speed temporal derivative slice $(\\mathbf{x}, t_k)$ of the plenoptic function.", + "bbox": [ + 511, + 777, + 903, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the related work, we referred to several techniques that excel at extracting high-speed spatial information from sparse event streams. These algorithms can do so as the scene information typically changes slowly [10], making the temporal gradient $\\frac{\\partial L(\\mathbf{x}_k,t_k)}{\\partial t}$ inherently sparse. This redundancy also applies to angular dimension as light field", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26912", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is mostly smooth [21, 56] and hence, $\\frac{\\partial L(\\vec{x},\\vec{\\omega},\\vec{s},\\lambda,t)}{\\partial\\vec{\\omega}}$ and $\\frac{\\partial L(\\vec{x},\\vec{\\omega},\\vec{s},\\lambda,t)}{\\partial t}$ are sparse. We exploit this sparsity to capture event fields with a single event camera in the next section.", + "bbox": [ + 89, + 88, + 482, + 140 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Event Field", + "text_level": 1, + "bbox": [ + 89, + 152, + 210, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned briefly in Sec. 3.2, an event camera's photodetector measures brightness by taking the logarithm of radiance integrated over all angles. This integration results in a loss of angular information. Our objective, through event fields is to retain this angular information $\\vec{\\omega}$ . To achieve this goal, we need to multiplex angular information into either the spatial or temporal dimensions that the event camera captures. We advocate for and design both spatial and temporal multiplexing techniques, as they offer complementary advantages.", + "bbox": [ + 89, + 176, + 483, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Spatial Multiplexing", + "text_level": 1, + "bbox": [ + 89, + 335, + 284, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We encode discrete angular information $(\\omega_{i})$ into the spatial dimension using a mapping function defined as:", + "bbox": [ + 89, + 359, + 482, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\omega_ {i}, \\mathbf {x} _ {\\mathbf {s}} \\left(\\omega_ {i}\\right)\\right) \\mapsto \\mathbf {x}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 398, + 480, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_{\\mathbf{s}}(\\omega_{i})$ represents the spatial coordinate for each measured view $(\\omega_{i})$ . Let the number of discrete angular views be $\\mathbf{n}$ , the spatial resolution of the sensor be $\\mathbf{r}$ , and $\\mathbf{r}$ is an integral multiple of $\\mathbf{n}$ . The specific mapping function $\\mathbf{x}_{\\mathbf{s}}$ varies based on the hardware design. The two commonly used mapping functions are:", + "bbox": [ + 89, + 422, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Multi-sensor array [54] or kaleidoscope [30]: In this design, the mapping function is $\\mathbf{x}_{\\mathbf{s}} = \\mathbf{x} \\mod (\\mathbf{r} / \\mathbf{n})$ .", + "- Microlens array [39]: In this design, the mapping function is $\\mathbf{x}_{\\mathbf{s}} = \\mathbf{x} / / \\mathbf{n}$ ." + ], + "bbox": [ + 89, + 513, + 482, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By applying this mapping function to Equation (2), the measurements are:", + "bbox": [ + 89, + 575, + 482, + 603 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial B}{\\partial t} (\\underbrace {\\mathbf {x} _ {\\mathbf {s}} , \\omega} _ {\\mathbf {x} _ {k}}, t _ {k}) \\approx \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 608, + 480, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the event camera will now measure the temporal brightness changes across both spatial and angular dimensions. Note that the brightness derivatives are still related to time, and we lose spatial resolution (by a factor of $n$ ) to improve angular resolution.", + "bbox": [ + 89, + 659, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Temporal Multiplexing", + "text_level": 1, + "bbox": [ + 89, + 742, + 305, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encode angular information in the temporal dimension, we steer the incoming rays along a closed periodic curve $\\mathcal{C}(t)$ with period $T$ . This steering could be accomplished by either moving the camera or placing steerable galvanometer in the optical path of the camera. The galvanometer can be steered fast along Lissajous curves [8, 41]; hence, we use it in all our experiments.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The temporal multiplexing does not sacrifice spatial resolution but sacrifices temporal resolution and always detects", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86b90fa7ac7414ed93aaa2daaa20d00b87b480c3f9fbfac983165ae3c05527c4.jpg", + "image_caption": [ + "(a) Plugin UI", + "Figure 2. Blender plugin for physics-based event field rendering." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 717, + 258 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb8e23a5fd75e89c232c185f70276e8fa49648c9c9eedce900347b340876faa2.jpg", + "image_caption": [ + "(b) Event simulation" + ], + "image_footnote": [], + "bbox": [ + 720, + 88, + 905, + 169 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88d04abcbab4db534f7df759d63506c2a98a970444bb041d75c639e9fc631d20.jpg", + "image_caption": [ + "(c) Event field" + ], + "image_footnote": [], + "bbox": [ + 720, + 180, + 903, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more events than spatial multiplexing. At any given time instant, the angular view will be at $\\mathcal{C}(t \\bmod T)$ and the temporally varying light field is $L(\\mathbf{x}, \\mathcal{C}(t \\bmod T), t)$ . For static scenes, the event camera measurement is", + "bbox": [ + 511, + 299, + 903, + 358 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial B (\\mathbf {x} , \\mathcal {C} (t))}{\\partial t} = \\left. \\frac {\\partial B (\\mathbf {x} , \\omega)}{\\partial \\omega} \\frac {\\partial \\mathcal {C} (t)}{\\partial t} \\right| _ {t = t _ {k}} \\approx \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 364, + 903, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, temporal multiplexing technique measures the angular derivatives of the light field $\\frac{\\partial L(\\mathbf{x},\\omega)}{\\partial\\omega}$ along the scanning curve.", + "bbox": [ + 511, + 407, + 903, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While we restricted the scanning curves to be Lissajous curves, specifically circles in most of the results, in principle, the scanning curve can be any periodic curve, including space-filling curves [15, 46]. Space-filling curves result in dense angular light-field capture at the cost of loosing temporal information. However, typically light field is highly redundant and smooth and a circular scan captures most of the light-field diversity as we show in the results.", + "bbox": [ + 511, + 455, + 905, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Simulator and Hardware Design", + "text_level": 1, + "bbox": [ + 511, + 588, + 808, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To validate our approach, we built a physics-based simulator and two hardware prototypes, kaleidoscope for spatial multiplexing and galvanometer for temporal multiplexing. The simulator allows us to precisely control the scene and design parameters, enabling an accurate and fair comparison between the two multiplexing techniques. Hardware prototypes show the real-world feasibility of our designs as well as new capabilities.", + "bbox": [ + 511, + 613, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Simulator", + "text_level": 1, + "bbox": [ + 511, + 742, + 625, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have used a straightforward approach to simulating events, which involves rendering the full plenoptic function, calculating brightness by propagating the plenoptic data through optics, and tracking the brightness difference from the previous event. A new event is triggered for each pixel when this difference exceeds a manually set threshold.", + "bbox": [ + 511, + 763, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We chose Blender [7] as our backend as it is opensource, physically based, supports GPU acceleration (Cycles), Python APIs, and brandishes a user-friendly GUI. We", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26913", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6dd46700a738085abdfa4ab4767a2f944cc7455ace35316672506cea806efe91.jpg", + "image_caption": [ + "(a) Ray sketch" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50b2831ae951dd4128427ed6f7cf7fe73ba816fd94926515f2b60beaaed1ec89.jpg", + "image_caption": [ + "(b) Hardware setup" + ], + "image_footnote": [], + "bbox": [ + 91, + 157, + 480, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/823beea596d251a87a5722b774b475a27a2a852abb00dec7f33fe54511801098.jpg", + "image_caption": [ + "Figure 3. Kaleidoscope design. (a) Illustration shows rays from various angles being directed towards the event camera. (b) The main lens focuses the scene onto the kaleidoscope, splitting the view into multiple angles. The beam splitter directs these multiple views to the RGB camera for standard light-field imaging and the event camera for event-field imaging.", + "(a) Ray sketch", + "Figure 4. Galvanometer design. (a) Ray sketch illustrating the angular separation achieved by controlling mirror angles in the galvanometer setup. (b) Hardware setup showing the galvanometer mirror system integrated with the event camera. This design allows for capturing angular derivative events by dynamically controlling the mirror position." + ], + "image_footnote": [], + "bbox": [ + 91, + 359, + 267, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/859904120910af13932cb06d70c1d9731da781e99f5723cf2e65fe8375a7de07.jpg", + "image_caption": [ + "(b) Hardware setup" + ], + "image_footnote": [], + "bbox": [ + 271, + 359, + 482, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "developed a Blender plugin to render event fields for either spatial or temporal multiplexing design, along with a simple GUI to facilitate community use. The GUI and sample simulation results are presented in Fig. 2.", + "bbox": [ + 89, + 577, + 482, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Kaleidoscope for Spatial Multiplexing", + "text_level": 1, + "bbox": [ + 89, + 648, + 416, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A kaleidoscope is a common optical instrument composed of multiple first-surface mirrors. Manakov et al. [30] introduced this optical device as an add-on for sensor fusion. As shown in Fig. 3, this kaleidoscope is a rectangular design made of four mirrors. The lens at the front forms an image at the entrance of the kaleidoscope. After multiple reflections within the kaleidoscope, the rear camera captures the image, resulting in several flipped images, each with a different virtual camera location. Therefore, the light field information is spatially multiplexed onto the camera sensor.", + "bbox": [ + 89, + 671, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Galvanometer for Temporal Multiplexing", + "text_level": 1, + "bbox": [ + 89, + 832, + 446, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the kaleidoscope design offers a simple solution without moving parts, it has two key limitations. First, although it provides angularly separated events, the events", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "still represent the derivative of light intensity over time. Second, the spatial resolution of current event cameras is not very high (1 megapixel), and this resolution would further decrease with spatial multiplexing. Temporal multiplexing offers an alternate approach.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For temporal multiplexing, we place a galvanometer, as shown in Fig. 4, in the optical path of the event camera. A galvanometer consists of two mirrors that can be precisely adjusted at high speed (Eg. GVS012 from Thorlabs can scan a thousand Lissajous curves per second). Moving mirrors allow us to rotate the camera virtually at a high speed and capture the event field. In this design, the captured information represents the derivative of light intensity with respect to angle, rather than time, as we discussed in Sec. 4.2.", + "bbox": [ + 511, + 167, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 513, + 316, + 599, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed techniques on various scenes. We first summarize the observations and then provide details.", + "bbox": [ + 511, + 340, + 903, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Simulations demonstrate that each design is optimal in different contexts. The kaleidoscope (spatial multiplexing) excels in capturing dynamic scenes with high frame rates but suffers from lower spatial resolution and is ineffective for static scenes, as no events are generated without motion. The galvanometer (temporal multiplexing) captures both static and dynamic scenes at full spatial resolution by recording angular derivatives but introduces motion blur if the object motion is significant.", + "bbox": [ + 511, + 371, + 905, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-world applications demonstrate each design's post-capture refocusing capabilities. Specifically, with the galvanometer setup, we show high-resolution refocusing of static and dynamic parts of the scene. With the kaleidoscope setup, we perform sensor fusion with an RGB camera, providing vibrant, high-speed refocusing for dynamic scenes. These results underscore the complementary nature of the two designs: the kaleidoscope for high-speed dynamic capture and the galvanometer for detailed, high spatial-angular light field recovery.", + "bbox": [ + 511, + 507, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following sections, we compare our proposed hardware designs (see Section 6.1), followed by applications for each setup. Specifically, we explore HDR light fields (Section 6.2) and slow-motion refocusing (Section 6.3) using the kaleidoscope configuration, as well as high-speed post-capture refocusing (Section 6.4) and instant true depth estimation (Section 6.5) using the galvanometer-based design.", + "bbox": [ + 511, + 657, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Kaleidoscope vs. Galvanometer Designs", + "text_level": 1, + "bbox": [ + 511, + 787, + 856, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practical applications, differences in optical properties—such as focal lengths—make it challenging to compare two cameras under identical conditions. The simulator, however, allows us to control these variables, enabling a fair comparison of both designs. We, therefore, used Blender to create a scene containing swirling water and moving objects", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26914", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ed24bd2bc9e9ba0b8472bc30d936e258785ab47d19c7e16cd02810038c5327a.jpg", + "image_caption": [ + "(a) Ground truth", + "Figure 5. (Simulation) comparison of spatial (kaleidoscope) and temporal (galvanometer) multiplexing techniques. (a) A simulated scene with swirling water and objects, with a stationary box. (b) Refocused image using the kaleidoscope design, showing reduced spatial resolution. However, the image quality is independent of scene motion. The static wooden box is not reconstructed. (c) Refocused images using the galvanometer design at different object movement speeds $(1\\times,2\\times,4\\times,$ and $8\\times)$ while maintaining a constant scanning speed. $1\\times$ corresponds to about 2.5 pixels per scan for the highlighted object. As the scene motion increases, the highlighted object, unlike in the kaleidoscopic design, suffers from motion blur. The static wooden boxes are perfectly reconstructed, unlike the kaleidoscopic design." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 228, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1eba0882cab0721c7c39d884d1fc89e229b5464d4f02d647785fc6ac5e5a422b.jpg", + "image_caption": [ + "(b) Kaleidoscope" + ], + "image_footnote": [], + "bbox": [ + 228, + 88, + 362, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44c57d8c6e6a88c733751cc1355de01903d79921b708ce674e157e346f92d1e2.jpg", + "image_caption": [ + "$\\therefore m : x = 1$ 或 ${3x} + {4y} + 1 = 0$" + ], + "image_footnote": [], + "bbox": [ + 364, + 89, + 498, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e60ca239c1ab5b1c113110a9c66e3d958e604534e9f4672f232fa131a47a101.jpg", + "image_caption": [ + "(c) Galvanometer" + ], + "image_footnote": [], + "bbox": [ + 498, + 89, + 633, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3770a4352c4c5efabc26a0bdec7c88bfb47ffc79df026630d39a2dcb311acd61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 89, + 767, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bfd7ab5e273ff5f747de5775503a5198ef781a020dee2b30d7179449d109fdf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 89, + 903, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8a513dfb20b468368cda8b74a09f220ee439f8c57ae423b09618ff9a00a5ab44.jpg", + "image_caption": [ + "(a) Scene" + ], + "image_footnote": [], + "bbox": [ + 93, + 282, + 218, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ac3b43a8af3f4f557cdea92ee0553db7d2a3c5ebba450872cd4ea3e95941350.jpg", + "image_caption": [ + "(b) Kaleidoscope" + ], + "image_footnote": [], + "bbox": [ + 223, + 282, + 354, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7956fe40f31be6cb50f7413fdd6ba6571a81ccd3720bcc99fdd2a6087e30abb9.jpg", + "image_caption": [ + "(c) Galvanometer" + ], + "image_footnote": [], + "bbox": [ + 354, + 282, + 480, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7781054d13be25ba622b52a449f446168ccfc0becd61ab29dfe57aeaa8d33390.jpg", + "image_caption": [ + "Figure 6. (Experimental) comparison of multiplexing techniques. (a) Reference scene with a grid pattern. (b) When the grid moves horizontally, the kaleidoscope design cannot capture horizontal lines as intensity does not change along this axis. (c) Using a galvanometer that scans in a circular motion ensures complete capture, avoiding information loss in the angular domain.", + "(a) Short exposure", + "Figure 7. High dynamic range (HDR) light field. A high dynamic range scene is captured with a standard RGB camera using (a) short exposure and (b) long exposure, and both fail to capture the full dynamic range. In contrast, (c) the HDR light field reconstructed from the event field effectively captures the complete dynamic range of the scene." + ], + "image_footnote": [], + "bbox": [ + 93, + 450, + 220, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66a298453a73c0e08f52cf32653edb2a5901148849ba2270d1ebafabbb3ae7f0.jpg", + "image_caption": [ + "(b) Long exposure" + ], + "image_footnote": [], + "bbox": [ + 223, + 452, + 352, + 512 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acdb913614ed441d2d25a34dce2e30afea51b3994540c610afa71b56004d7f82.jpg", + "image_caption": [ + "(c) HDR light field" + ], + "image_footnote": [], + "bbox": [ + 352, + 452, + 480, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "within a stationary box (see Fig. 5) to assess the strengths and limitations of each design.", + "bbox": [ + 89, + 621, + 482, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first significant difference is that the galvanometer design can fully reconstruct scene information by capturing angular derivatives, whereas the kaleidoscope design can only capture moving objects, as no events are generated in static regions. The stationary wooden box is clearly visible in the galvanometer design but is absent in the kaleidoscope design. We also observe this in a real experiment (Fig. 6), where a rectangular mesh is translated horizontally, and the kaleidoscope fails to reconstruct horizontal lines.", + "bbox": [ + 89, + 655, + 482, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A second key difference is that the galvanometer design multiplexes the event field into the time domain, making it susceptible to motion blur from rapidly moving objects. This is visible in the zoomed insets in Fig. 5. The kaleidoscope design, by contrast, avoids motion blur but instead suffers from lower spatial resolution. However, both designs can capturing and refocusing dynamic objects.", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. HDR Light Fields", + "text_level": 1, + "bbox": [ + 513, + 284, + 687, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Event fields are not only high-speed but also high-dynamic range light fields. We demonstrate this in Fig. 7, where the scene containing bright LED light in the foreground and diffuse texture in the background is captured using both an RGB camera and the event camera through a kaleidoscope. The event fields have a large dynamic range and capture the spatial structure of both LEDs and the background texture, whereas the RGB camera either captures the light sources (short exposure) or the background (long exposure).", + "bbox": [ + 511, + 306, + 906, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. SloMoRF: Slow Motion Refocusing", + "text_level": 1, + "bbox": [ + 511, + 455, + 820, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We augment the kaleidoscope design with a co-located RGB camera to capture and reconstruct slow-motion RGB frames with dynamic refocusing capability, called SloMoRF. The hardware setup, shown in Fig. 3, features a beam splitter for identical kaleidoscope views for RGB and event camera without interfering with the main lens.", + "bbox": [ + 511, + 479, + 905, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize the TimeLens algorithm [50] for sensor fusion1. Using the captured event data, we interpolate frames to generate a $5 \\times$ higher frame-rate RGB video (from 120 to 720 fps). The interpolated results are presented in Fig. 8.", + "bbox": [ + 511, + 570, + 905, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This approach allows us to generate a high-speed color light field. We treat this light field as unstructured as we found it empirically easier than calibrating the views. To align the images, we estimate the necessary shift at a specified patch using a basic pattern-matching algorithm. Note that this shift computation is only needed once for the entire video sequence as long as the target focal plane remains fixed. Figure 8 shows refocusing results across three focal planes. Additionally, the captured event field allows for focus tracking on a high-speed object moving along the depth. We demonstrate this capability in Fig. 9, where we visually compare sequence of images that focus on a fast-moving toy dart vs. focusing solely on the background. These results demonstrate the high frame rate, color, and refocusing capabilities of our kaleidoscope design when combined with the state-of-the-art sensor fusion algorithms.", + "bbox": [ + 511, + 632, + 906, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ code for the updated TimeLens++ [51] is not publicly available", + "bbox": [ + 531, + 886, + 870, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26915", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb33bd42cc41b131abcf7fe0a42c4e0250630c68d0f8d618162e5e321341e4c0.jpg", + "image_caption": [ + "(a) Scene", + "(b) Captured frame (with aligned events)" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 251, + 260 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90e65e349a4af66a24d58c54e194ed748da79af488be8ad830f35e192abe0c6e.jpg", + "image_caption": [ + "Frame 19" + ], + "image_footnote": [], + "bbox": [ + 251, + 88, + 380, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/904f98013420f21178a1aa8c47a5f09623e32975064fa4a7e8f58d75430122b3.jpg", + "image_caption": [ + "1st interpolated", + "(c) Refocus with interpolated frames" + ], + "image_footnote": [], + "bbox": [ + 383, + 88, + 511, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6d1cf29d2fdd0ac59fd9c741866429c4a3e6bb6f1ec36389ee18ec622697905.jpg", + "image_caption": [ + "3rd interpolated" + ], + "image_footnote": [], + "bbox": [ + 513, + 88, + 640, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07067c492c2cf50501d9a50c1f71446f626788a57ebd02562df0cb53cc6b8319.jpg", + "image_caption": [ + "5th interpolated" + ], + "image_footnote": [], + "bbox": [ + 642, + 88, + 772, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/842a969a96ff0334f4d83ea8bf90032e8b336bae661158ca2ba4926bbe4caee7.jpg", + "image_caption": [ + "Frame 20" + ], + "image_footnote": [], + "bbox": [ + 774, + 88, + 903, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59ecbe89c21831a2b98979d831edf6e25c28d706977e95d26c536f85a7e1f1f0.jpg", + "image_caption": [ + "Figure 8. SlowMoRF: Slow motion refocusing with kaleidoscope and sensor fusion. (a) Original setup with Lego pieces floating in a water tank as a wooden brick is thrown in, with CVPR in the background, (b) Captured raw frames with aligned events using kaleidoscope design, and (c) Interpolated refocused frames at three focal planes (back, mid, and front). Between the frames, the fall of the wooden block and changes in water splash and caustics can be observed in the front refocus. Mid and back refocus show sharp Lego toys and CVPR. This setup achieves high-speed (720 fps) refocusing with enhanced clarity across focal depths. The supplementary contains the full video.", + "(a) Refocus on the bullet" + ], + "image_footnote": [], + "bbox": [ + 91, + 378, + 222, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6ff010469d5b771780b691cd597a4560909f4bc096566cbd57aa2c7ab54ad2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 380, + 352, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b2dc2885cb5471ff7398ad3cc5275d9610971667e2b4890909c8b9860d0178f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 380, + 480, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc487c107a240e654d62835e933e6046e524f389be0f8dcca9679ae859d03fc2.jpg", + "image_caption": [ + "(b) Refocus on the background", + "Figure 9. High-speed refocusing on a toy dart. The original video was captured at 120 fps, interpolating seven frames between each pair of original frames, resulting in a 960 fps high-speed light field video. The images shown here are all the middle (the fourth) interpolated frames, demonstrating refocusing on two planes: (a) the moving dart and (b) the background. Current high-speed videography cannot adjust focus during capture, so dynamic objects must remain at a fixed focal plane to stay sharp. Our technique eliminates this limitation and can focus on dynamic object." + ], + "image_footnote": [], + "bbox": [ + 93, + 449, + 222, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8966ffc4b44ca2314a2467ffc09c39a5472a47d2cfe03aef8e38f5330c6e0c36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 449, + 352, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/390a2d3a2cca213bde6f1e2acd506402a20598d235bd8b3388ba2a40a8879bcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 449, + 480, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. High-Speed Post-capture Refocusing", + "text_level": 1, + "bbox": [ + 89, + 669, + 408, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we test the post-capture refocusing ability of the galvanometer setup on dynamic scenes. Compared to unstructured event field captured by kaleidoscope design, our galvanometer design produces a fully structured event field after calibration, as we precisely control the galvanometer with a signal generator to determine the angle relative to time. The supplementary materials provide calibration details.", + "bbox": [ + 89, + 696, + 482, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using E2VID, we generate a high-speed video at 10,000 fps (0.1 ms per frame) from the captured event field data. The galvanometer is operating at $250\\mathrm{Hz}$ . Therefore, we reconstruct the light field at 250 fps and 40 views, the fastest recorded so far at a megapixel resolution.", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6047bd58e4da46426b4ce53b7f965386231a186521235a2536d87b799113d92e.jpg", + "image_caption": [ + "(a) Grayscale reconstructed images" + ], + "image_footnote": [], + "bbox": [ + 514, + 380, + 709, + 484 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/034afb9cd748ca08e532e0b704a59c1231d0732977c4e1b978fb8ae076219389.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 380, + 903, + 484 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac7386f86882bdfb9e07bc6a3268771f7e6ff69fbc8002f4ec9e4c932468f41e.jpg", + "image_caption": [ + "(b) Scene captured by ordinary camera (60Hz)", + "Figure 10. Dynamic scene refocusing. Using our galvanometer design at $250\\mathrm{Hz}$ , we capture a scene with a rotating fan marked with \"CVPR\", spinning at $300\\mathrm{rpm}$ . (a) Using prior knowledge of the scanned mirror's location, we refocus on the background or the rotating fan in the foreground. (b) An ordinary camera captures the scene at $60\\mathrm{Hz}$ , focusing on the background and foreground from different viewpoints. The camera lacks refocus ability and causes motion blur on the fan due to the slower shutter speed." + ], + "image_footnote": [], + "bbox": [ + 514, + 496, + 709, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2e7124817fc22ac5b71f79dab14ab5fdbbce15d8013bd4a416633e2b060b70a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 496, + 903, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using this light field and equations in Sec. 3.1, we can do post-capture refocusing. Note that even though E2VID generates noisy frames at high speed (due to low light), the noise gets averaged out when we synthesize a refocused image. In Fig. 10, we show refocusing results on the background texture (left figure) or foreground fan (right figure) and can observe that they are noise-free. We further stress-tested our hardware setup in Fig. 11 and noticed that the event camera begins to lose events due to the readout bandwidth limit when the galvanometer operates beyond $250\\mathrm{Hz}$ .", + "bbox": [ + 511, + 746, + 906, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26916", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41072534c48c6633677789e53b10f0ab81a6898c82d177d1db7c895730a9fda7.jpg", + "image_caption": [ + "Event frame Front refocus Back refocus", + "Figure 11. Effect of galvo scan frequency for refocusing accuracy. The scene consists of two static objects positioned in the foreground and background, along with a fan (highlighted in yellow boxes) in the foreground rotating at $480\\mathrm{RPM}$ . The static objects (highlighted in red and blue boxes) remain sharp up to a scanning frequency of $200\\mathrm{Hz}$ but become increasingly blurry as the scan speed increases. This counter intuitive phenomenon occurs as event camera reaches its maximum bandwidth capacity, resulting in random event loss, as seen in the Event frame row. However, the text on the rotating fan becomes clearer at higher scan speeds due to reduced motion blur." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 903, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e00f44c110d3db29a87ccb121aa6c6404f3aaaf5a8a00b360c0985c9f108905e.jpg", + "image_caption": [ + "Figure 12. Instant true depth estimation at $100\\mathrm{Hz}$ . Depth mapping of a running person, shown in three frames with color-coded depth scale (in inches), highlighting the system's capability for rapid, accurate depth measurement in dynamic scenes." + ], + "image_footnote": [], + "bbox": [ + 91, + 385, + 478, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Instant True Depth Estimation", + "text_level": 1, + "bbox": [ + 89, + 585, + 362, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to refocusing, our galvanometer design enables real-time true depth estimation. Unlike SLAM-based methods [5], our approach provides instant true depth estimation based on the event field. As outlined in Sec. 6.4, our galvanometer design generates a structured light field, requiring only a single calibration step to determine the depth-to-disparity ratio. We employ the \"depth from focus\" algorithm [13] to calculate the depth of each pixel. This involves first constructing a focal stack by refocusing the light field at various depths. For each pixel, we determine the focal stack image where the pixel appears sharpest, and the depth associated with that image is the depth of the pixel.", + "bbox": [ + 89, + 618, + 482, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By adopting this approach, we obtain the instant true depth at the same frequency as the galvanometer's scanning speed. As shown in Fig. 12, a $100\\mathrm{Hz}$ galvo scan successfully captures a person running away from the camera, offering sparse yet real-time depth information of the letters on their shirt at a rate of 100 fps.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 387, + 777, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this manuscript, we introduce \"event field\", a new approach for capturing high-speed, high resolution, HDR light fields using two optical configurations—kaleidoscope and galvanometer—combined with an event camera. Our kaleidoscope design enables capturing temporal derivatives for multiple views, while the galvanometer design captures angular derivatives along curves. We validate both designs through simulations and real-world data, demonstrating their unique advantages.", + "bbox": [ + 511, + 419, + 906, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments showcase several potential applications. Most notably, our setup enables high-speed refocusing on fast-moving objects, real-time true depth estimation, and high-dynamic-range light field capture. Supplementary PDF and video provides further visualizations of these results across additional scenes.", + "bbox": [ + 511, + 561, + 906, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our approach shows several new possibilities, it also creates new research problems in the graphics, imaging, and vision fields. Our current brute-force simulation of the event field could benefit from optimal importance sampling techniques that improve the rendering speed by orders of magnitude similar to previous importance techniques for computational cameras [20, 22, 27, 40]. Additionally, as shown in Fig. 11, static parts of the scene do not require high bandwidth. By implementing adaptive scanning (e.g., slower scans for static parts, faster scans for dynamic areas) with a MEMS array, we can enhance bandwidth efficiency. We currently use pre-trained E2VID and Timelens models, which are optimized for time-sequential event streams. Generating sufficient event field data, possibly through an optimal simulator and training models specifically for our event fields, would yield a significant performance boost.", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26917", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This project is funded by NSF awards 2403122, 2326904, 2107454, and 2107313 and ARPA-H award D24AC00296.", + "We thank Laurence Rubin at Boston Craft Works, Joseph Poissant at Dartmouth Engineering Machine Shop, and Hannah Machesky at Two Way Mirrors for providing mirrors and polishing expertise for our kaleidoscope design.", + "We also extend our gratitude to the RISC lab members, Atul Agarwal, Dhawal Sirikonda, Kedari Chowtoori, and Sarah Friday, for their generous assistance in capturing the data and their 3D printing and laser cutting skills." + ], + "bbox": [ + 89, + 114, + 482, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 291, + 187, + 306 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Vision and Modeling Group, Media Laboratory, Massachusetts Institute of ..., 1991. 3", + "[2] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2", + "[3] Tom E Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE transactions on pattern analysis and machine intelligence, 34(5):972-986, 2011. 3", + "[4] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 3", + "[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 8", + "[6] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-iliidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. arXiv preprint arXiv:2408.13627, 2024. 2", + "[7] Blender Online Community. Blender - a 3d modelling and rendering package, 2024. 4", + "[8] Toshiaki Fujii and Masayuki Tanimoto. A real-time ray-space acquisition system. In Stereoscopic Displays and Virtual Reality Systems XI, pages 179-187. SPIE, 2004. 2, 4", + "[9] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. IEEE transactions on pattern analysis and machine intelligence, 40(10):2402-2412, 2017. 3", + "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 3" + ], + "bbox": [ + 93, + 316, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 2", + "[12] Steven J. Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F. Cohen. The lumigraph. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 43-54, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3", + "[13] Paul Grossmann. Depth from focus. Pattern recognition letters, 5(1):63-69, 1987. 8", + "[14] Ruipeng Guo, Qianwan Yang, Andrew S Chang, Guorong Hu, Joseph Greene, Christopher V Gabel, Sixian You, and Lei Tian. Eventlfm: Event camera integrated fourier light field microscopy for ultrafast 3d imaging. Light: Science & Applications, 13(1):144, 2024. 3", + "[15] Mohit Gupta, Andreas Velten, Shree K Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM Transactions on Graphics (TOG), 37(2):1-18, 2018. 4", + "[16] Shuji Habuchi, Keita Takahashi, Chihiro Tsutake, Toshiaki Fujii, and Hajime Nagahara. Time-efficient light-field acquisition using coded aperture and events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24923-24933, 2024. 3", + "[17] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 3", + "[18] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. High-speed structured light based 3d scanning using an event camera. Optics Express, 29(22):35864-35876, 2021. 3", + "[19] Ivo Ihrke, Timo Stich, Heiko Gottschlich, Marcus Magnor, and Hans-Peter Seidel. Fast incident light field acquisition and rendering. Journal of WSCG, 2008. 2", + "[20] Adrian Jarabo, Julio Marco, Adolfo Munoz, Raul Buisan, Wojciech Jarosz, and Diego Gutierrez. A framework for transient rendering. ACM Transactions on Graphics (ToG), 33 (6):1-10, 2014. 8", + "[21] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 2, 4", + "[22] Juhyeon Kim, Wojciech Jarosz, Ioannis Gkioulekas, and Adithya Pediredla. Doppler time-of-flight rendering. ACM Trans. Graph., 42(6), 2023. 8", + "[23] T Leroux, S-H Ieng, and Ryad Benosman. Event-based structured light for depth reconstruction using frequency tagged light patterns. arXiv preprint arXiv:1811.10771, 2018.3", + "[24] M. Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2", + "[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 31-42, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3" + ], + "bbox": [ + 514, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26918", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 db $15\\mu s$ latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 3", + "[27] Yang Liu, Shaojie Jiao, and Wojciech Jarosz. Temporally sliced photon primitives for time-of-flight rendering. In Computer Graphics Forum, pages 29-40. Wiley Online Library, 2022. 8", + "[28] Lytro. Lytro. https://www.lytro.com/, 2017. [Online; accessed 6-Oct-2024]. 2", + "[29] Misha Mahowald and Misha Mahowald. The silicon retina. An Analog VLSI System for Stereoscopic Vision, pages 4-65, 1994. 2, 3", + "[30] Alkhazur Manakov, John Restrepo, Oliver Klehm, Ramon Hegedus, Elmar Eisemann, Hans-Peter Seidel, and Ivo Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM Transactions on Graphics, 32(4):47-1, 2013. 2, 4, 5", + "[31] Ashish Rao Mangalore, Chandra Sekhar Seelamantula, and Chetan Singh Thakur. Neuromorphic fringe projection profilometry. IEEE Signal Processing Letters, 27:1510-1514, 2020. 3", + "[32] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 3", + "[33] Leonard McMillan and Gary Bishop. Plenoptic modeling: An image-based rendering system. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 433-440. Association for Computing Machinery, 2023. 3", + "[34] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14414-14423, 2020. 2", + "[35] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 2021 International Conference on 3D Vision (3DV), pages 1165-1174. IEEE, 2021. 3", + "[36] Manasi Muglikar, Diederik Paul Moeys, and Davide Scaramuzza. Event guided depth sensing. In 2021 International Conference on 3D Vision (3DV), pages 385-393. IEEE, 2021. 3", + "[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 3", + "[38] Manasi Muglikar, Siddharth Somasundaram, Akshat Dave, Edoardo Charbon, Ramesh Raskar, and Davide Scaramuzza. Event cameras meet spads for high-speed, low-bandwidth imaging. arXiv preprint arXiv:2404.11511, 2024. 3", + "[39] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 4", + "[40] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 8", + "[41] Luboš Poláček, Jana Jurmanová, and Zdeněk Navrátil. Laser galvo mirrors: perfect instrument for the demonstration of lissajous figures. Physics Education, 54(5):055002, 2019. 4", + "[42] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 3", + "[43] Ziyuan Qu. Event field. https://dartmouth-risc-lab.github.io/event-field-website/, 2024.2", + "[44] A Raytrix. 3d light field camera technology, 2017. 2", + "[45] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 3", + "[46] Hans Sagan. *Space-filling curves*. Springer Science & Business Media, 2012. 4", + "[47] Dhawal Sirikonda, Praneeth Chakravarthula, Ioannis Gkioulekas, and Adithya Pediredla. Structured light with a million light planes per second. arXiv preprint arXiv:2411.18597, 2024. 3", + "[48] Yuichi Taguchi, Amit Agrawal, Srikumar Ramalingam, and Ashok Veeraraghavan. Axial light field for curved mirrors: Reflect your perspective, widen your view. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 499-506. IEEE, 2010. 2", + "[49] David Tedaldi, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Feature detection and tracking with the dynamic and active-pixel vision sensor (davis). In 2016 Second International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7. IEEE, 2016. 2", + "[50] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 3, 6", + "[51] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 3, 6", + "[52] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2", + "[53] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 2", + "[54] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2, 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26919", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9602-9611, 2024. 3", + "[56] Mingyuan Zhao, Hao Sheng, Da Yang, Sizhe Wang, Ruixuan Cong, Zhenglong Cui, Rongshan Chen, Tun Wang, Shuai Wang, Yang Huang, et al. A survey for light field superresolution. High-Confidence Computing, page 100206, 2024. 2, 4", + "[57] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE transactions on neural networks and learning systems, 34(8):4868-4880, 2021. 2", + "[58] Zihao Zou, Ziyuan Qu, Xi Peng, Vivek Boominathan, Adithya Pediredla, and Praneeth Chakravarthula. High-speed dynamic 3d imaging with sensor fusion splatting. arXiv preprint arXiv:2502.04630, 2025. 3" + ], + "bbox": [ + 91, + 90, + 485, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26920", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_model.json b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..359ffe1bb0a2cadcc00982c944abfa18ab37a559 --- /dev/null +++ b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_model.json @@ -0,0 +1,2994 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.153 + ], + "angle": 0, + "content": "Event fields: Capturing light fields at high speed, resolution, and dynamic range" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.182, + 0.272, + 0.218 + ], + "angle": 0, + "content": "Ziyuan Qu Dartmouth College" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.182, + 0.45, + 0.217 + ], + "angle": 0, + "content": "Zihao Zou \nUNC Chapel Hill" + }, + { + "type": "text", + "bbox": [ + 0.486, + 0.182, + 0.648, + 0.218 + ], + "angle": 0, + "content": "Vivek Boominathan Rice University" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.182, + 0.882, + 0.218 + ], + "angle": 0, + "content": "Praneeth Chakravarthula \nUNC Chapel Hill" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.224, + 0.574, + 0.26 + ], + "angle": 0, + "content": "Adithya Pediredla Dartmouth College" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.275, + 0.391, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.275, + 0.905, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.454, + 0.908, + 0.525 + ], + "angle": 0, + "content": "Figure 1. We introduce Event Fields, a novel framework that leverages event cameras to capture light fields at high speed, resolution, and dynamic range using innovative optical designs. Event fields integrate the unique advantages of event sensing—such as low latency, data efficiency, and power-saving capabilities—with the detailed angular information provided by light field sensing. We demonstrate the versatility of the event fields through three application scenarios: slow motion refocus (SloMoRF), instant true depth estimation, and high dynamic range (HDR) light field imaging." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.537, + 0.327, + 0.552 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.569, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Event cameras, which feature pixels that independently respond to changes in brightness, are becoming increasingly popular in high-speed applications due to their lower latency, reduced bandwidth requirements, and enhanced dynamic range compared to traditional frame-based cameras. Numerous imaging and vision techniques have leveraged event cameras for high-speed scene understanding by capturing high-framerate, high-dynamic range videos, primarily utilizing the temporal advantages inherent to event cameras. Additionally, imaging and vision techniques have utilized the light field—a complementary dimension to temporal information—for enhanced scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose \"Event Fields\", a new approach that utilizes innovative optical designs for event cameras to capture light fields at high speed. We develop the underlying mathematical framework for Event Fields and introduce two foundational frameworks to capture them practically: spatial multiplexing to capture temporal derivatives and temporal multiplexing to capture angular derivatives. To realize these, we design two complementary optical setups—one using a kaleidoscope for spatial multiplexing and another using a galvanometer for temporal multiplexing. We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.675 + ], + "angle": 0, + "content": "evaluate the performance of both designs using a custom-built simulator and real hardware prototypes, showcasing their distinct benefits. Our event fields unlock the full advantages of typical light fields—like post-capture refocusing and depth estimation—now supercharged for high-speed and high-dynamic range scenes. This novel light-sensing paradigm opens doors to new applications in photography, robotics, and AR/VR, and presents fresh challenges in rendering and machine learning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.693, + 0.645, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Capturing high-speed visual phenomena in real-time is essential for advancing science and engineering. Traditional high-speed frame cameras like Phantom, Photron, and IDT can achieve impressive frame rates (up to 100,000 fps) but are bulky, have finite capture duration, and are expensive (exceeding $100,000), restricting accessibility. In contrast, event cameras offer a more affordable and efficient solution with high temporal resolution (down to \\(1\\mu \\mathrm{s}\\)) and continuous data acquisition, suitable for real-time applications. Unlike traditional cameras, which require substantial bandwidth (e.g., 100,000 fps, HD capture requires 100 GBps), event cameras capture only changes in the scene using asyn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26910" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "chronous pixel readouts, reducing bandwidth demands and enabling continuous streaming. This approach has recently been harnessed for high-speed applications, including optical flow, object tracking, high-dynamic range imaging, structured light scanning, and photometric stereo [6, 10]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.483, + 0.368 + ], + "angle": 0, + "content": "In this work, we extend event camera capabilities by introducing angular capture, allowing for post-capture manipulation of high-speed video—a previously unexplored functionality. Capturing angular radiance information is known as light field imaging [24, 25, 53], enabling features such as post-capture refocusing and depth estimation. We propose \"Event Fields\" that record radiance derivatives across angular and temporal dimensions, extending light field imaging principles to high-speed, high-dynamic-range scenarios (see Fig. 1). To capture these Event Fields, we introduce two complementary hardware systems: a kaleidoscope-based design and a galvanometer-based design, each with unique advantages for this purpose." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.569 + ], + "angle": 0, + "content": "The kaleidoscope design (see Fig. 3), inspired by Manakov et al. [30], captures angle-resolved time derivatives by placing a rectangular kaleidoscope in the optical path. This approach allows for capturing angular information affordably and simply at \\(3 \\times 3\\) views. The galvanometer design (see Fig. 4) uses a galvanometer positioned in the optical path to capture angular derivatives directly. By controlling the mirrors at \\(250\\mathrm{Hz}\\), the incoming rays are shifted along a curve, enabling the capture of both angular and temporal derivatives. As angular derivatives are sparse [21, 56], this design does not significantly increase bandwidth requirements. Table 1 compares the relative advantages of these two hardware approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.724 + ], + "angle": 0, + "content": "To systematically evaluate the relative merits of the hardware designs, we build a Blender plugin to render event-fields with physics based light simulation. We also build hardware imaging prototypes with kaleidoscope and galvanometer, and algorithms to demonstrate new imaging capabilities in challenging environments. These include post-capture refocusing on dynamic scenes (e.g., a rotating fan at 480 RPM), focus tracking of nef darts traveling at \\(20\\mathrm{m / s}\\), high-speed depth reconstruction, and high-resolution, high dynamic range light fields." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Contributions. To state explicitly, the manuscript makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.763, + 0.483, + 0.793 + ], + "angle": 0, + "content": "- Introduces event fields; spatial and temporal multiplexing techniques to capture them." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.794, + 0.483, + 0.838 + ], + "angle": 0, + "content": "- Builds a physics-based simulator and two hardware prototypes to systematically evaluate the merits and limitations of the multiplexing techniques." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.839, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- New applications and related algorithms, including slow motion refocus (SloMoRF), true-depth estimation, and high dynamic range light field imaging." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.884, + 0.473, + 0.899 + ], + "angle": 0, + "content": "- Open source simulator, data, and processing codes [43]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.763, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.09, + 0.903, + 0.226 + ], + "angle": 0, + "content": "
KaleidoscopeGalvanometer
DerivativesTemp.Ang. + Temp.
Recon. PerformanceScene dep.Scene independ.
ColorYesNo
Static LF CaptureNoYes
Dynamic LF captureYesYes
Refocus FidelityChoppySmooth
Bandwidth RequiredLowHigh
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.231, + 0.905, + 0.259 + ], + "angle": 0, + "content": "Table 1. Relative advantages of kaleidoscope vs. galvanometer design for capturing event fields." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.266, + 0.655, + 0.281 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.55 + ], + "angle": 0, + "content": "Light field sensing: One of the first approaches to light field capture [54] involves using multiple sensors to gather angular information from various perspectives. However, this setup is bulky, costly, and requires high bandwidth communication. An alternative is time-sequential capture, which mounts a sensor on a mechanical gantry to capture the light field of a static scene [12, 25]. Rather than moving the sensor itself, a mirror can be used to redirect the view. Ihrke et al. [19] proposed using a movable planar mirror to shift the view temporally. However, this method requires individual calibration for each image, adding complexity. Fujii et al. [8] uses a galvanometer and pairs it with a high-speed camera. Inspired by their approach, we developed a similar design, precisely controlling the galvanometer's movement. A key distinction is that our galvanometer operates continuously, while their system selects discrete positions from a lookup table, making it slower." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.909, + 0.763 + ], + "angle": 0, + "content": "In addition to time-sequential capture, angular information can also be measured with spatial multiplexing. One way to achieve this is through a microlens array [39], which has been widely adopted in commercial light field cameras like Raytrix [44] and Lytro [28]. While effective, microlens arrays are challenging to manufacture and install. Taguchi et al. [48] suggested using a mirror ball to expand the camera's field of view, though this setup results in the camera being visible in the mirror ball's reflection. Alternatively, Manakov et al. [30] introduced a kaleidoscope-like mirror array that can be attached to a standard camera without modification. Inspired by this design, we adapted it to convert a conventional event camera into a device capable of capturing the event field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Event camera designs: From the first neuromorphic sensor by Mahowald et al. [29] to the recent Sony IMX636, event sensing saw exponential growth. We briefly review designs most relevant to our manuscript and refer the readers to recent surveys [6, 10]. Event cameras are regularly used in robotics and autonomous navigation for optical flow computation [2, 11, 52], feature tracking [49], and motion segmentation [34, 57], but event cameras cannot detect edges in the direction of the camera motion. Inspired by microsac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26911" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "cades in human eye movement, He et al. [17] placed a rotating wedge mask before the event camera aperture to capture edges in all directions. Similar to them, our galvanometer design detects edges oriented in all directions, independent of camera and object motion, as shown in Fig. 6. Our galvanometer design also finds inspiration from the designs of Mugalikar et al. [37], who used a rotating polarizer in the aperture plane to capture high-fidelity polarization information for shape-from-polarization and Yu et al. [55], who used a rotating light source and event camera for fast photometric stereo." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.454 + ], + "angle": 0, + "content": "Habuchi et al. [16] captured light fields for static scenes by placing an LCOS device in the aperture plane of event cameras and using CNN to recover the light field. While the technique is quite novel, it is limited to static scenes due to the LCOS response time (22 ms). Guo et al. [14] placed a microlens array (MLA) on the event sensor to capture the light field for dynamic scenes and showed microscopic imaging applications with their prototype. We built our prototypes for macroscopic applications, and our kaleidoscope design is functionally same as MLA design. However, the galvanometer design can capture angular information even for static scenes without trading off the spatial resolution like Guo et al. [14]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.666 + ], + "angle": 0, + "content": "Event processing algorithms: Event cameras capture scene information at higher temporal resolution (in \\(\\mu\\)s) and dynamic range (140 dB) compared to standard cameras (60 dB) [10]. These features enable solutions to challenges where standard frame-based sensors struggle, such as high-speed motion estimation [9], high frame rate reconstruction [45], and high dynamic range imaging [45]. Besides, event cameras and standard frame-based cameras are complementary; event cameras provide high-speed, sparse data, while standard cameras capture color and fine details. By fusing these two types of sensors, algorithms like Timelens [50] and Timelens++ [51] effectively interpolate RGB frames using event information. We adapt these event processing algorithms for processing event field data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.832 + ], + "angle": 0, + "content": "Other event-based fusions: Apart from RGB cameras, other computational imaging systems can also be made fast by fusing with event sensors [58]. Event-based structured light sensing techniques [18, 23, 31, 32, 35, 47] reconstruct 3D scenes at high speed by replacing standard intensity sensors with event cameras. Event cameras are also used for foveated depth sensing [36] to scan moving objects at a higher spatial resolution. Muglikar et al. [38] combined single-photon avalanche diodes (SPADs) with event cameras to achieve high-speed, low-light, low-bandwidth imaging capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.303, + 0.862 + ], + "angle": 0, + "content": "3. Technical Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We will briefly review the image formation models for light field and event cameras that are useful in developing the" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.763, + 0.105 + ], + "angle": 0, + "content": "mathematical models for event fields." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.636, + 0.132 + ], + "angle": 0, + "content": "3.1. Light Field" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.905, + 0.365 + ], + "angle": 0, + "content": "The plenoptic model of light \\( L(\\vec{x}, \\vec{\\omega}, \\lambda, t) \\) is a comprehensive representation that captures all the dimensions of light emitted by a scene [1, 33]. Here, \\( \\vec{x} = (x, y, z) \\) represents the spatial coordinates of the observer through which the light passes; \\( \\vec{\\omega} = (\\omega_x, \\omega_y) \\) are the angular coordinates; \\( \\lambda \\) is the wavelength, and \\( t \\) is the time. At a given time, the measurement of a monochromatic light is 5D function. Levoy and Hanrahan [25] and Gortler et al. [12] showed that the 5D representation has redundancies and have reduced it to 4D by assuming that the light field is measured in free space. Under this assumption, as the radiance of a ray remains a constant in a straight line, the static monochromatic light field is a 4D function \\( L(\\vec{u}, \\vec{s}) \\), where \\( \\vec{u} = (u, v) \\) is a point on the virtual sensor and \\( \\vec{s} = (s, t) \\) is a point on the virtual aperture." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.905, + 0.396 + ], + "angle": 0, + "content": "To synthesize an image from a given light field, we integrate the light field over the aperture as:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.406, + 0.905, + 0.438 + ], + "angle": 0, + "content": "\\[\nI (x, y) = \\int_ {\\vec {s}} L (\\vec {u}, \\vec {s}) \\mathrm {d} \\vec {s}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.905, + 0.553 + ], + "angle": 0, + "content": "Light field allows us to change the focus plane after imaging by computationally changing the distance between the sensor and aperture planes [3, 25]. If \\(d_0\\) is the original distance between the planes and \\(d\\) is the distance required to focus at a given depth, the light field that focuses at \\(d\\) will be \\(L_{d}(\\vec{u},\\vec{s}) = L(\\vec{u} +(d_{0} / d)\\vec{s},\\vec{s})\\). The refocused image can be synthesized by integrating angles, as shown in Eq. (1)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.563, + 0.662, + 0.577 + ], + "angle": 0, + "content": "3.2.Event Camera" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.905, + 0.706 + ], + "angle": 0, + "content": "An event camera asynchronously records the brightness \\((B\\dot{=} \\log (\\int_{\\vec{\\omega}}Ld\\vec{\\omega}))\\) change event for each pixel independently [4, 26, 29, 42]. Specifically, in an ideal and noise-free scenario, a pixel \\(\\mathbf{x} = (x_k,y_k)^T\\) will be triggered at time \\(t_k\\) as soon as the brightness change reaches a predefined threshold \\(C\\) since the last triggered event. The polarity \\(p_k\\in \\{+1, - 1\\}\\), which is the sign of the brightness change, will be recorded for each trigger. Mathematically," + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.719, + 0.887, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\Delta B \\left(\\mathbf {x} _ {k}, t _ {k}\\right) \\doteq B \\left(\\mathbf {x} _ {k}, t _ {k}\\right) - B \\left(\\mathbf {x} _ {k}, t _ {k} - \\Delta t _ {k}\\right) = p _ {k} C\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.738, + 0.905, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial B \\left(\\mathbf {x} _ {k} , t _ {k}\\right)}{\\partial t} \\approx \\frac {p _ {k} C}{\\Delta t _ {k}} \\Rightarrow \\frac {\\partial L \\left(\\mathbf {x} _ {k} , t _ {k}\\right)}{\\partial t} \\approx L \\frac {p _ {k} C}{\\Delta t _ {k}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.779, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Therefore, event cameras capture a high-speed temporal derivative slice \\((\\mathbf{x}, t_k)\\) of the plenoptic function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the related work, we referred to several techniques that excel at extracting high-speed spatial information from sparse event streams. These algorithms can do so as the scene information typically changes slowly [10], making the temporal gradient \\(\\frac{\\partial L(\\mathbf{x}_k,t_k)}{\\partial t}\\) inherently sparse. This redundancy also applies to angular dimension as light field" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26912" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.141 + ], + "angle": 0, + "content": "is mostly smooth [21, 56] and hence, \\(\\frac{\\partial L(\\vec{x},\\vec{\\omega},\\vec{s},\\lambda,t)}{\\partial\\vec{\\omega}}\\) and \\(\\frac{\\partial L(\\vec{x},\\vec{\\omega},\\vec{s},\\lambda,t)}{\\partial t}\\) are sparse. We exploit this sparsity to capture event fields with a single event camera in the next section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.153, + 0.212, + 0.167 + ], + "angle": 0, + "content": "4. Event Field" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.178, + 0.484, + 0.33 + ], + "angle": 0, + "content": "As mentioned briefly in Sec. 3.2, an event camera's photodetector measures brightness by taking the logarithm of radiance integrated over all angles. This integration results in a loss of angular information. Our objective, through event fields is to retain this angular information \\(\\vec{\\omega}\\). To achieve this goal, we need to multiplex angular information into either the spatial or temporal dimensions that the event camera captures. We advocate for and design both spatial and temporal multiplexing techniques, as they offer complementary advantages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.337, + 0.285, + 0.354 + ], + "angle": 0, + "content": "4.1. Spatial Multiplexing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.483, + 0.39 + ], + "angle": 0, + "content": "We encode discrete angular information \\((\\omega_{i})\\) into the spatial dimension using a mapping function defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.399, + 0.482, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\left(\\omega_ {i}, \\mathbf {x} _ {\\mathbf {s}} \\left(\\omega_ {i}\\right)\\right) \\mapsto \\mathbf {x}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.514 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_{\\mathbf{s}}(\\omega_{i})\\) represents the spatial coordinate for each measured view \\((\\omega_{i})\\). Let the number of discrete angular views be \\(\\mathbf{n}\\), the spatial resolution of the sensor be \\(\\mathbf{r}\\), and \\(\\mathbf{r}\\) is an integral multiple of \\(\\mathbf{n}\\). The specific mapping function \\(\\mathbf{x}_{\\mathbf{s}}\\) varies based on the hardware design. The two commonly used mapping functions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.515, + 0.483, + 0.545 + ], + "angle": 0, + "content": "- Multi-sensor array [54] or kaleidoscope [30]: In this design, the mapping function is \\( \\mathbf{x}_{\\mathbf{s}} = \\mathbf{x} \\mod (\\mathbf{r} / \\mathbf{n}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.546, + 0.483, + 0.575 + ], + "angle": 0, + "content": "- Microlens array [39]: In this design, the mapping function is \\( \\mathbf{x}_{\\mathbf{s}} = \\mathbf{x} / / \\mathbf{n} \\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.515, + 0.483, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.604 + ], + "angle": 0, + "content": "By applying this mapping function to Equation (2), the measurements are:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.609, + 0.482, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial B}{\\partial t} (\\underbrace {\\mathbf {x} _ {\\mathbf {s}} , \\omega} _ {\\mathbf {x} _ {k}}, t _ {k}) \\approx \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.735 + ], + "angle": 0, + "content": "and the event camera will now measure the temporal brightness changes across both spatial and angular dimensions. Note that the brightness derivatives are still related to time, and we lose spatial resolution (by a factor of \\( n \\)) to improve angular resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.306, + 0.76 + ], + "angle": 0, + "content": "4.2. Temporal Multiplexing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "To encode angular information in the temporal dimension, we steer the incoming rays along a closed periodic curve \\(\\mathcal{C}(t)\\) with period \\(T\\). This steering could be accomplished by either moving the camera or placing steerable galvanometer in the optical path of the camera. The galvanometer can be steered fast along Lissajous curves [8, 41]; hence, we use it in all our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The temporal multiplexing does not sacrifice spatial resolution but sacrifices temporal resolution and always detects" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.718, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.26, + 0.658, + 0.271 + ], + "angle": 0, + "content": "(a) Plugin UI" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.089, + 0.906, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.169, + 0.874, + 0.181 + ], + "angle": 0, + "content": "(b) Event simulation" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.181, + 0.905, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.259, + 0.856, + 0.27 + ], + "angle": 0, + "content": "(c) Event field" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.277, + 0.903, + 0.292 + ], + "angle": 0, + "content": "Figure 2. Blender plugin for physics-based event field rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.905, + 0.359 + ], + "angle": 0, + "content": "more events than spatial multiplexing. At any given time instant, the angular view will be at \\(\\mathcal{C}(t \\bmod T)\\) and the temporally varying light field is \\(L(\\mathbf{x}, \\mathcal{C}(t \\bmod T), t)\\). For static scenes, the event camera measurement is" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.366, + 0.905, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial B (\\mathbf {x} , \\mathcal {C} (t))}{\\partial t} = \\left. \\frac {\\partial B (\\mathbf {x} , \\omega)}{\\partial \\omega} \\frac {\\partial \\mathcal {C} (t)}{\\partial t} \\right| _ {t = t _ {k}} \\approx \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.905, + 0.455 + ], + "angle": 0, + "content": "Therefore, temporal multiplexing technique measures the angular derivatives of the light field \\(\\frac{\\partial L(\\mathbf{x},\\omega)}{\\partial\\omega}\\) along the scanning curve." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.576 + ], + "angle": 0, + "content": "While we restricted the scanning curves to be Lissajous curves, specifically circles in most of the results, in principle, the scanning curve can be any periodic curve, including space-filling curves [15, 46]. Space-filling curves result in dense angular light-field capture at the cost of loosing temporal information. However, typically light field is highly redundant and smooth and a circular scan captures most of the light-field diversity as we show in the results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.81, + 0.607 + ], + "angle": 0, + "content": "5. Simulator and Hardware Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.735 + ], + "angle": 0, + "content": "To validate our approach, we built a physics-based simulator and two hardware prototypes, kaleidoscope for spatial multiplexing and galvanometer for temporal multiplexing. The simulator allows us to precisely control the scene and design parameters, enabling an accurate and fair comparison between the two multiplexing techniques. Hardware prototypes show the real-world feasibility of our designs as well as new capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.627, + 0.757 + ], + "angle": 0, + "content": "5.1. Simulator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.855 + ], + "angle": 0, + "content": "We have used a straightforward approach to simulating events, which involves rendering the full plenoptic function, calculating brightness by propagating the plenoptic data through optics, and tracking the brightness difference from the previous event. A new event is triggered for each pixel when this difference exceeds a manually set threshold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We chose Blender [7] as our backend as it is opensource, physically based, supports GPU acceleration (Cycles), Python APIs, and brandishes a user-friendly GUI. We" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26913" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.146, + 0.331, + 0.159 + ], + "angle": 0, + "content": "(a) Ray sketch" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.159, + 0.482, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.251, + 0.346, + 0.264 + ], + "angle": 0, + "content": "(b) Hardware setup" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.352 + ], + "angle": 0, + "content": "Figure 3. Kaleidoscope design. (a) Illustration shows rays from various angles being directed towards the event camera. (b) The main lens focuses the scene onto the kaleidoscope, splitting the view into multiple angles. The beam splitter directs these multiple views to the RGB camera for standard light-field imaging and the event camera for event-field imaging." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.361, + 0.269, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.467, + 0.221, + 0.48 + ], + "angle": 0, + "content": "(a) Ray sketch" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.361, + 0.483, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.468, + 0.435, + 0.48 + ], + "angle": 0, + "content": "(b) Hardware setup" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Figure 4. Galvanometer design. (a) Ray sketch illustrating the angular separation achieved by controlling mirror angles in the galvanometer setup. (b) Hardware setup showing the galvanometer mirror system integrated with the event camera. This design allows for capturing angular derivative events by dynamically controlling the mirror position." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.639 + ], + "angle": 0, + "content": "developed a Blender plugin to render event fields for either spatial or temporal multiplexing design, along with a simple GUI to facilitate community use. The GUI and sample simulation results are presented in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.418, + 0.666 + ], + "angle": 0, + "content": "5.2. Kaleidoscope for Spatial Multiplexing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.822 + ], + "angle": 0, + "content": "A kaleidoscope is a common optical instrument composed of multiple first-surface mirrors. Manakov et al. [30] introduced this optical device as an add-on for sensor fusion. As shown in Fig. 3, this kaleidoscope is a rectangular design made of four mirrors. The lens at the front forms an image at the entrance of the kaleidoscope. After multiple reflections within the kaleidoscope, the rear camera captures the image, resulting in several flipped images, each with a different virtual camera location. Therefore, the light field information is spatially multiplexed onto the camera sensor." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.447, + 0.85 + ], + "angle": 0, + "content": "5.3. Galvanometer for Temporal Multiplexing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While the kaleidoscope design offers a simple solution without moving parts, it has two key limitations. First, although it provides angularly separated events, the events" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "still represent the derivative of light intensity over time. Second, the spatial resolution of current event cameras is not very high (1 megapixel), and this resolution would further decrease with spatial multiplexing. Temporal multiplexing offers an alternate approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.304 + ], + "angle": 0, + "content": "For temporal multiplexing, we place a galvanometer, as shown in Fig. 4, in the optical path of the event camera. A galvanometer consists of two mirrors that can be precisely adjusted at high speed (Eg. GVS012 from Thorlabs can scan a thousand Lissajous curves per second). Moving mirrors allow us to rotate the camera virtually at a high speed and capture the event field. In this design, the captured information represents the derivative of light intensity with respect to angle, rather than time, as we discussed in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.317, + 0.6, + 0.332 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.905, + 0.371 + ], + "angle": 0, + "content": "We evaluate the proposed techniques on various scenes. We first summarize the observations and then provide details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.906, + 0.508 + ], + "angle": 0, + "content": "Simulations demonstrate that each design is optimal in different contexts. The kaleidoscope (spatial multiplexing) excels in capturing dynamic scenes with high frame rates but suffers from lower spatial resolution and is ineffective for static scenes, as no events are generated without motion. The galvanometer (temporal multiplexing) captures both static and dynamic scenes at full spatial resolution by recording angular derivatives but introduces motion blur if the object motion is significant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Real-world applications demonstrate each design's post-capture refocusing capabilities. Specifically, with the galvanometer setup, we show high-resolution refocusing of static and dynamic parts of the scene. With the kaleidoscope setup, we perform sensor fusion with an RGB camera, providing vibrant, high-speed refocusing for dynamic scenes. These results underscore the complementary nature of the two designs: the kaleidoscope for high-speed dynamic capture and the galvanometer for detailed, high spatial-angular light field recovery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.78 + ], + "angle": 0, + "content": "In the following sections, we compare our proposed hardware designs (see Section 6.1), followed by applications for each setup. Specifically, we explore HDR light fields (Section 6.2) and slow-motion refocusing (Section 6.3) using the kaleidoscope configuration, as well as high-speed post-capture refocusing (Section 6.4) and instant true depth estimation (Section 6.5) using the galvanometer-based design." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.857, + 0.805 + ], + "angle": 0, + "content": "6.1. Kaleidoscope vs. Galvanometer Designs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In practical applications, differences in optical properties—such as focal lengths—make it challenging to compare two cameras under identical conditions. The simulator, however, allows us to control these variables, enabling a fair comparison of both designs. We, therefore, used Blender to create a scene containing swirling water and moving objects" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26914" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.229, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.164, + 0.21, + 0.175 + ], + "angle": 0, + "content": "(a) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.089, + 0.363, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.164, + 0.348, + 0.176 + ], + "angle": 0, + "content": "(b) Kaleidoscope" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.09, + 0.499, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.151, + 0.442, + 0.162 + ], + "angle": 0, + "content": "\\( \\therefore m : x = 1 \\) 或 \\( {3x} + {4y} + 1 = 0 \\)" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.09, + 0.634, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.09, + 0.769, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.09, + 0.905, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.164, + 0.689, + 0.175 + ], + "angle": 0, + "content": "(c) Galvanometer" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.181, + 0.907, + 0.266 + ], + "angle": 0, + "content": "Figure 5. (Simulation) comparison of spatial (kaleidoscope) and temporal (galvanometer) multiplexing techniques. (a) A simulated scene with swirling water and objects, with a stationary box. (b) Refocused image using the kaleidoscope design, showing reduced spatial resolution. However, the image quality is independent of scene motion. The static wooden box is not reconstructed. (c) Refocused images using the galvanometer design at different object movement speeds \\((1\\times,2\\times,4\\times,\\) and \\(8\\times)\\) while maintaining a constant scanning speed. \\(1\\times\\) corresponds to about 2.5 pixels per scan for the highlighted object. As the scene motion increases, the highlighted object, unlike in the kaleidoscopic design, suffers from motion blur. The static wooden boxes are perfectly reconstructed, unlike the kaleidoscopic design." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.284, + 0.22, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.337, + 0.188, + 0.347 + ], + "angle": 0, + "content": "(a) Scene" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.283, + 0.355, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.337, + 0.341, + 0.348 + ], + "angle": 0, + "content": "(b) Kaleidoscope" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.283, + 0.482, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.337, + 0.471, + 0.348 + ], + "angle": 0, + "content": "(c) Galvanometer" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.437 + ], + "angle": 0, + "content": "Figure 6. (Experimental) comparison of multiplexing techniques. (a) Reference scene with a grid pattern. (b) When the grid moves horizontally, the kaleidoscope design cannot capture horizontal lines as intensity does not change along this axis. (c) Using a galvanometer that scans in a circular motion ensures complete capture, avoiding information loss in the angular domain." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.452, + 0.222, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.514, + 0.213, + 0.526 + ], + "angle": 0, + "content": "(a) Short exposure" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.453, + 0.353, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.514, + 0.343, + 0.526 + ], + "angle": 0, + "content": "(b) Long exposure" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.453, + 0.482, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.514, + 0.474, + 0.526 + ], + "angle": 0, + "content": "(c) HDR light field" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Figure 7. High dynamic range (HDR) light field. A high dynamic range scene is captured with a standard RGB camera using (a) short exposure and (b) long exposure, and both fail to capture the full dynamic range. In contrast, (c) the HDR light field reconstructed from the event field effectively captures the complete dynamic range of the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.653 + ], + "angle": 0, + "content": "within a stationary box (see Fig. 5) to assess the strengths and limitations of each design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.791 + ], + "angle": 0, + "content": "The first significant difference is that the galvanometer design can fully reconstruct scene information by capturing angular derivatives, whereas the kaleidoscope design can only capture moving objects, as no events are generated in static regions. The stationary wooden box is clearly visible in the galvanometer design but is absent in the kaleidoscope design. We also observe this in a real experiment (Fig. 6), where a rectangular mesh is translated horizontally, and the kaleidoscope fails to reconstruct horizontal lines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "A second key difference is that the galvanometer design multiplexes the event field into the time domain, making it susceptible to motion blur from rapidly moving objects. This is visible in the zoomed insets in Fig. 5. The kaleidoscope design, by contrast, avoids motion blur but instead suffers from lower spatial resolution. However, both designs can capturing and refocusing dynamic objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.285, + 0.688, + 0.301 + ], + "angle": 0, + "content": "6.2. HDR Light Fields" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Event fields are not only high-speed but also high-dynamic range light fields. We demonstrate this in Fig. 7, where the scene containing bright LED light in the foreground and diffuse texture in the background is captured using both an RGB camera and the event camera through a kaleidoscope. The event fields have a large dynamic range and capture the spatial structure of both LEDs and the background texture, whereas the RGB camera either captures the light sources (short exposure) or the background (long exposure)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.456, + 0.821, + 0.472 + ], + "angle": 0, + "content": "6.3. SloMoRF: Slow Motion Refocusing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.906, + 0.57 + ], + "angle": 0, + "content": "We augment the kaleidoscope design with a co-located RGB camera to capture and reconstruct slow-motion RGB frames with dynamic refocusing capability, called SloMoRF. The hardware setup, shown in Fig. 3, features a beam splitter for identical kaleidoscope views for RGB and event camera without interfering with the main lens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.632 + ], + "angle": 0, + "content": "We utilize the TimeLens algorithm [50] for sensor fusion1. Using the captured event data, we interpolate frames to generate a \\( 5 \\times \\) higher frame-rate RGB video (from 120 to 720 fps). The interpolated results are presented in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.875 + ], + "angle": 0, + "content": "This approach allows us to generate a high-speed color light field. We treat this light field as unstructured as we found it empirically easier than calibrating the views. To align the images, we estimate the necessary shift at a specified patch using a basic pattern-matching algorithm. Note that this shift computation is only needed once for the entire video sequence as long as the target focal plane remains fixed. Figure 8 shows refocusing results across three focal planes. Additionally, the captured event field allows for focus tracking on a high-speed object moving along the depth. We demonstrate this capability in Fig. 9, where we visually compare sequence of images that focus on a fast-moving toy dart vs. focusing solely on the background. These results demonstrate the high frame rate, color, and refocusing capabilities of our kaleidoscope design when combined with the state-of-the-art sensor fusion algorithms." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.871, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)code for the updated TimeLens++ [51] is not publicly available" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26915" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.252, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.185, + 0.193, + 0.196 + ], + "angle": 0, + "content": "(a) Scene" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.262, + 0.228, + 0.288 + ], + "angle": 0, + "content": "(b) Captured frame (with aligned events)" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.089, + 0.382, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.261, + 0.345, + 0.272 + ], + "angle": 0, + "content": "Frame 19" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.089, + 0.513, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.261, + 0.493, + 0.273 + ], + "angle": 0, + "content": "1st interpolated" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.089, + 0.642, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.261, + 0.625, + 0.273 + ], + "angle": 0, + "content": "3rd interpolated" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.089, + 0.774, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.261, + 0.755, + 0.273 + ], + "angle": 0, + "content": "5th interpolated" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.089, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.261, + 0.868, + 0.272 + ], + "angle": 0, + "content": "Frame 20" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.275, + 0.686, + 0.288 + ], + "angle": 0, + "content": "(c) Refocus with interpolated frames" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Figure 8. SlowMoRF: Slow motion refocusing with kaleidoscope and sensor fusion. (a) Original setup with Lego pieces floating in a water tank as a wooden brick is thrown in, with CVPR in the background, (b) Captured raw frames with aligned events using kaleidoscope design, and (c) Interpolated refocused frames at three focal planes (back, mid, and front). Between the frames, the fall of the wooden block and changes in water splash and caustics can be observed in the front refocus. Mid and back refocus show sharp Lego toys and CVPR. This setup achieves high-speed (720 fps) refocusing with enhanced clarity across focal depths. The supplementary contains the full video." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.38, + 0.223, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.381, + 0.353, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.381, + 0.482, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.439, + 0.361, + 0.449 + ], + "angle": 0, + "content": "(a) Refocus on the bullet" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.45, + 0.223, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.45, + 0.353, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.45, + 0.482, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.508, + 0.38, + 0.52 + ], + "angle": 0, + "content": "(b) Refocus on the background" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Figure 9. High-speed refocusing on a toy dart. The original video was captured at 120 fps, interpolating seven frames between each pair of original frames, resulting in a 960 fps high-speed light field video. The images shown here are all the middle (the fourth) interpolated frames, demonstrating refocusing on two planes: (a) the moving dart and (b) the background. Current high-speed videography cannot adjust focus during capture, so dynamic objects must remain at a fixed focal plane to stay sharp. Our technique eliminates this limitation and can focus on dynamic object." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.409, + 0.687 + ], + "angle": 0, + "content": "6.4. High-Speed Post-capture Refocusing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.818 + ], + "angle": 0, + "content": "In this section, we test the post-capture refocusing ability of the galvanometer setup on dynamic scenes. Compared to unstructured event field captured by kaleidoscope design, our galvanometer design produces a fully structured event field after calibration, as we precisely control the galvanometer with a signal generator to determine the angle relative to time. The supplementary materials provide calibration details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Using E2VID, we generate a high-speed video at 10,000 fps (0.1 ms per frame) from the captured event field data. The galvanometer is operating at \\(250\\mathrm{Hz}\\). Therefore, we reconstruct the light field at 250 fps and 40 views, the fastest recorded so far at a megapixel resolution." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.381, + 0.71, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.381, + 0.905, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.485, + 0.815, + 0.497 + ], + "angle": 0, + "content": "(a) Grayscale reconstructed images" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.497, + 0.71, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.497, + 0.905, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.61, + 0.848, + 0.623 + ], + "angle": 0, + "content": "(b) Scene captured by ordinary camera (60Hz)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.74 + ], + "angle": 0, + "content": "Figure 10. Dynamic scene refocusing. Using our galvanometer design at \\(250\\mathrm{Hz}\\), we capture a scene with a rotating fan marked with \"CVPR\", spinning at \\(300\\mathrm{rpm}\\). (a) Using prior knowledge of the scanned mirror's location, we refocus on the background or the rotating fan in the foreground. (b) An ordinary camera captures the scene at \\(60\\mathrm{Hz}\\), focusing on the background and foreground from different viewpoints. The camera lacks refocus ability and causes motion blur on the fan due to the slower shutter speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.907, + 0.899 + ], + "angle": 0, + "content": "Using this light field and equations in Sec. 3.1, we can do post-capture refocusing. Note that even though E2VID generates noisy frames at high speed (due to low light), the noise gets averaged out when we synthesize a refocused image. In Fig. 10, we show refocusing results on the background texture (left figure) or foreground fan (right figure) and can observe that they are noise-free. We further stress-tested our hardware setup in Fig. 11 and noticed that the event camera begins to lose events due to the readout bandwidth limit when the galvanometer operates beyond \\(250\\mathrm{Hz}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26916" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.089, + 0.113, + 0.277 + ], + "angle": 270, + "content": "Event frame Front refocus Back refocus" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.905, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.299, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 11. Effect of galvo scan frequency for refocusing accuracy. The scene consists of two static objects positioned in the foreground and background, along with a fan (highlighted in yellow boxes) in the foreground rotating at \\(480\\mathrm{RPM}\\). The static objects (highlighted in red and blue boxes) remain sharp up to a scanning frequency of \\(200\\mathrm{Hz}\\) but become increasingly blurry as the scan speed increases. This counter intuitive phenomenon occurs as event camera reaches its maximum bandwidth capacity, resulting in random event loss, as seen in the Event frame row. However, the text on the rotating fan becomes clearer at higher scan speeds due to reduced motion blur." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.386, + 0.48, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.568 + ], + "angle": 0, + "content": "Figure 12. Instant true depth estimation at \\(100\\mathrm{Hz}\\). Depth mapping of a running person, shown in three frames with color-coded depth scale (in inches), highlighting the system's capability for rapid, accurate depth measurement in dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.364, + 0.604 + ], + "angle": 0, + "content": "6.5. Instant True Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.483, + 0.801 + ], + "angle": 0, + "content": "In addition to refocusing, our galvanometer design enables real-time true depth estimation. Unlike SLAM-based methods [5], our approach provides instant true depth estimation based on the event field. As outlined in Sec. 6.4, our galvanometer design generates a structured light field, requiring only a single calibration step to determine the depth-to-disparity ratio. We employ the \"depth from focus\" algorithm [13] to calculate the depth of each pixel. This involves first constructing a focal stack by refocusing the light field at various depths. For each pixel, we determine the focal stack image where the pixel appears sharpest, and the depth associated with that image is the depth of the pixel." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "By adopting this approach, we obtain the instant true depth at the same frequency as the galvanometer's scanning speed. As shown in Fig. 12, a \\(100\\mathrm{Hz}\\) galvo scan successfully captures a person running away from the camera, offering sparse yet real-time depth information of the letters on their shirt at a rate of 100 fps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.388, + 0.779, + 0.403 + ], + "angle": 0, + "content": "7. Discussion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.556 + ], + "angle": 0, + "content": "In this manuscript, we introduce \"event field\", a new approach for capturing high-speed, high resolution, HDR light fields using two optical configurations—kaleidoscope and galvanometer—combined with an event camera. Our kaleidoscope design enables capturing temporal derivatives for multiple views, while the galvanometer design captures angular derivatives along curves. We validate both designs through simulations and real-world data, demonstrating their unique advantages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.652 + ], + "angle": 0, + "content": "Our experiments showcase several potential applications. Most notably, our setup enables high-speed refocusing on fast-moving objects, real-time true depth estimation, and high-dynamic-range light field capture. Supplementary PDF and video provides further visualizations of these results across additional scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "While our approach shows several new possibilities, it also creates new research problems in the graphics, imaging, and vision fields. Our current brute-force simulation of the event field could benefit from optimal importance sampling techniques that improve the rendering speed by orders of magnitude similar to previous importance techniques for computational cameras [20, 22, 27, 40]. Additionally, as shown in Fig. 11, static parts of the scene do not require high bandwidth. By implementing adaptive scanning (e.g., slower scans for static parts, faster scans for dynamic areas) with a MEMS array, we can enhance bandwidth efficiency. We currently use pre-trained E2VID and Timelens models, which are optimized for time-sequential event streams. Generating sufficient event field data, possibly through an optimal simulator and training models specifically for our event fields, would yield a significant performance boost." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26917" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.116, + 0.482, + 0.146 + ], + "angle": 0, + "content": "This project is funded by NSF awards 2403122, 2326904, 2107454, and 2107313 and ARPA-H award D24AC00296." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.147, + 0.483, + 0.207 + ], + "angle": 0, + "content": "We thank Laurence Rubin at Boston Craft Works, Joseph Poissant at Dartmouth Engineering Machine Shop, and Hannah Machesky at Two Way Mirrors for providing mirrors and polishing expertise for our kaleidoscope design." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.208, + 0.482, + 0.264 + ], + "angle": 0, + "content": "We also extend our gratitude to the RISC lab members, Atul Agarwal, Dhawal Sirikonda, Kedari Chowtoori, and Sarah Friday, for their generous assistance in capturing the data and their 3D printing and laser cutting skills." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.292, + 0.188, + 0.308 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[1] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Vision and Modeling Group, Media Laboratory, Massachusetts Institute of ..., 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.375, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[2] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[3] Tom E Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE transactions on pattern analysis and machine intelligence, 34(5):972-986, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[4] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[6] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-iliidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. arXiv preprint arXiv:2408.13627, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[7] Blender Online Community. Blender - a 3d modelling and rendering package, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[8] Toshiaki Fujii and Masayuki Tanimoto. A real-time ray-space acquisition system. In Stereoscopic Displays and Virtual Reality Systems XI, pages 179-187. SPIE, 2004. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[9] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. IEEE transactions on pattern analysis and machine intelligence, 40(10):2402-2412, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.317, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[12] Steven J. Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F. Cohen. The lumigraph. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 43-54, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[13] Paul Grossmann. Depth from focus. Pattern recognition letters, 5(1):63-69, 1987. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[14] Ruipeng Guo, Qianwan Yang, Andrew S Chang, Guorong Hu, Joseph Greene, Christopher V Gabel, Sixian You, and Lei Tian. Eventlfm: Event camera integrated fourier light field microscopy for ultrafast 3d imaging. Light: Science & Applications, 13(1):144, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[15] Mohit Gupta, Andreas Velten, Shree K Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM Transactions on Graphics (TOG), 37(2):1-18, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[16] Shuji Habuchi, Keita Takahashi, Chihiro Tsutake, Toshiaki Fujii, and Hajime Nagahara. Time-efficient light-field acquisition using coded aperture and events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24923-24933, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[17] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[18] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. High-speed structured light based 3d scanning using an event camera. Optics Express, 29(22):35864-35876, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Ivo Ihrke, Timo Stich, Heiko Gottschlich, Marcus Magnor, and Hans-Peter Seidel. Fast incident light field acquisition and rendering. Journal of WSCG, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[20] Adrian Jarabo, Julio Marco, Adolfo Munoz, Raul Buisan, Wojciech Jarosz, and Diego Gutierrez. A framework for transient rendering. ACM Transactions on Graphics (ToG), 33 (6):1-10, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[21] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.745 + ], + "angle": 0, + "content": "[22] Juhyeon Kim, Wojciech Jarosz, Ioannis Gkioulekas, and Adithya Pediredla. Doppler time-of-flight rendering. ACM Trans. Graph., 42(6), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[23] T Leroux, S-H Ieng, and Ryad Benosman. Event-based structured light for depth reconstruction using frequency tagged light patterns. arXiv preprint arXiv:1811.10771, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.803, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[24] M. Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 31-42, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26918" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 db \\(15\\mu s\\) latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[27] Yang Liu, Shaojie Jiao, and Wojciech Jarosz. Temporally sliced photon primitives for time-of-flight rendering. In Computer Graphics Forum, pages 29-40. Wiley Online Library, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.205, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[28] Lytro. Lytro. https://www.lytro.com/, 2017. [Online; accessed 6-Oct-2024]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.233, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[29] Misha Mahowald and Misha Mahowald. The silicon retina. An Analog VLSI System for Stereoscopic Vision, pages 4-65, 1994. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[30] Alkhazur Manakov, John Restrepo, Oliver Klehm, Ramon Hegedus, Elmar Eisemann, Hans-Peter Seidel, and Ivo Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM Transactions on Graphics, 32(4):47-1, 2013. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[31] Ashish Rao Mangalore, Chandra Sekhar Seelamantula, and Chetan Singh Thakur. Neuromorphic fringe projection profilometry. IEEE Signal Processing Letters, 27:1510-1514, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[32] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[33] Leonard McMillan and Gary Bishop. Plenoptic modeling: An image-based rendering system. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 433-440. Association for Computing Machinery, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[34] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14414-14423, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[35] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 2021 International Conference on 3D Vision (3DV), pages 1165-1174. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[36] Manasi Muglikar, Diederik Paul Moeys, and Davide Scaramuzza. Event guided depth sensing. In 2021 International Conference on 3D Vision (3DV), pages 385-393. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[38] Manasi Muglikar, Siddharth Somasundaram, Akshat Dave, Edoardo Charbon, Ramesh Raskar, and Davide Scaramuzza. Event cameras meet spads for high-speed, low-bandwidth imaging. arXiv preprint arXiv:2404.11511, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[39] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[41] Luboš Poláček, Jana Jurmanová, and Zdeněk Navrátil. Laser galvo mirrors: perfect instrument for the demonstration of lissajous figures. Physics Education, 54(5):055002, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[42] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[43] Ziyuan Qu. Event field. https://dartmouth-risc-lab.github.io/event-field-website/, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.861, + 0.288 + ], + "angle": 0, + "content": "[44] A Raytrix. 3d light field camera technology, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[45] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[46] Hans Sagan. *Space-filling curves*. Springer Science & Business Media, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[47] Dhawal Sirikonda, Praneeth Chakravarthula, Ioannis Gkioulekas, and Adithya Pediredla. Structured light with a million light planes per second. arXiv preprint arXiv:2411.18597, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[48] Yuichi Taguchi, Amit Agrawal, Srikumar Ramalingam, and Ashok Veeraraghavan. Axial light field for curved mirrors: Reflect your perspective, widen your view. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 499-506. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[49] David Tedaldi, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Feature detection and tracking with the dynamic and active-pixel vision sensor (davis). In 2016 Second International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[50] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[51] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[52] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[53] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26919" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[55] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9602-9611, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[56] Mingyuan Zhao, Hao Sheng, Da Yang, Sizhe Wang, Ruixuan Cong, Zhenglong Cui, Rongshan Chen, Tun Wang, Shuai Wang, Yang Huang, et al. A survey for light field superresolution. High-Confidence Computing, page 100206, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[57] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE transactions on neural networks and learning systems, 34(8):4868-4880, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[58] Zihao Zou, Ziyuan Qu, Xi Peng, Vivek Boominathan, Adithya Pediredla, and Praneeth Chakravarthula. High-speed dynamic 3d imaging with sensor fusion splatting. arXiv preprint arXiv:2502.04630, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26920" + } + ] +] \ No newline at end of file diff --git a/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_origin.pdf b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4ef59752e8888b5272ff96c5113127512a6e6667 --- /dev/null +++ b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/b83374cf-6c00-4bd9-b132-43495f3f9c2e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8332b40f83cc5bd7f1b87f3e03aa94c38531d20ce623ab8e6b09e3801b2d9053 +size 6913431 diff --git a/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/full.md b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/full.md new file mode 100644 index 0000000000000000000000000000000000000000..505cce94121d41fcd5bb9ecac599da7fae0c2146 --- /dev/null +++ b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/full.md @@ -0,0 +1,407 @@ +# Event fields: Capturing light fields at high speed, resolution, and dynamic range + +Ziyuan Qu Dartmouth College + +Zihao Zou +UNC Chapel Hill + +Vivek Boominathan Rice University + +Praneeth Chakravarthula +UNC Chapel Hill + +Adithya Pediredla Dartmouth College + +![](images/098c5afda1924c14b4ad7c758b3330a88746163e43b5ef86b36c43b82da882a7.jpg) +Figure 1. We introduce Event Fields, a novel framework that leverages event cameras to capture light fields at high speed, resolution, and dynamic range using innovative optical designs. Event fields integrate the unique advantages of event sensing—such as low latency, data efficiency, and power-saving capabilities—with the detailed angular information provided by light field sensing. We demonstrate the versatility of the event fields through three application scenarios: slow motion refocus (SloMoRF), instant true depth estimation, and high dynamic range (HDR) light field imaging. + +![](images/542431f7ab310f6ad4c9d2f7d41f5a09f622f56f30b307a044ff5b6d81ef728c.jpg) + +# Abstract + +Event cameras, which feature pixels that independently respond to changes in brightness, are becoming increasingly popular in high-speed applications due to their lower latency, reduced bandwidth requirements, and enhanced dynamic range compared to traditional frame-based cameras. Numerous imaging and vision techniques have leveraged event cameras for high-speed scene understanding by capturing high-framerate, high-dynamic range videos, primarily utilizing the temporal advantages inherent to event cameras. Additionally, imaging and vision techniques have utilized the light field—a complementary dimension to temporal information—for enhanced scene understanding. + +In this work, we propose "Event Fields", a new approach that utilizes innovative optical designs for event cameras to capture light fields at high speed. We develop the underlying mathematical framework for Event Fields and introduce two foundational frameworks to capture them practically: spatial multiplexing to capture temporal derivatives and temporal multiplexing to capture angular derivatives. To realize these, we design two complementary optical setups—one using a kaleidoscope for spatial multiplexing and another using a galvanometer for temporal multiplexing. We + +evaluate the performance of both designs using a custom-built simulator and real hardware prototypes, showcasing their distinct benefits. Our event fields unlock the full advantages of typical light fields—like post-capture refocusing and depth estimation—now supercharged for high-speed and high-dynamic range scenes. This novel light-sensing paradigm opens doors to new applications in photography, robotics, and AR/VR, and presents fresh challenges in rendering and machine learning. + +# 1. Introduction + +Capturing high-speed visual phenomena in real-time is essential for advancing science and engineering. Traditional high-speed frame cameras like Phantom, Photron, and IDT can achieve impressive frame rates (up to 100,000 fps) but are bulky, have finite capture duration, and are expensive (exceeding $100,000), restricting accessibility. In contrast, event cameras offer a more affordable and efficient solution with high temporal resolution (down to $1\mu \mathrm{s}$ ) and continuous data acquisition, suitable for real-time applications. Unlike traditional cameras, which require substantial bandwidth (e.g., 100,000 fps, HD capture requires 100 GBps), event cameras capture only changes in the scene using asyn + +chronous pixel readouts, reducing bandwidth demands and enabling continuous streaming. This approach has recently been harnessed for high-speed applications, including optical flow, object tracking, high-dynamic range imaging, structured light scanning, and photometric stereo [6, 10]. + +In this work, we extend event camera capabilities by introducing angular capture, allowing for post-capture manipulation of high-speed video—a previously unexplored functionality. Capturing angular radiance information is known as light field imaging [24, 25, 53], enabling features such as post-capture refocusing and depth estimation. We propose "Event Fields" that record radiance derivatives across angular and temporal dimensions, extending light field imaging principles to high-speed, high-dynamic-range scenarios (see Fig. 1). To capture these Event Fields, we introduce two complementary hardware systems: a kaleidoscope-based design and a galvanometer-based design, each with unique advantages for this purpose. + +The kaleidoscope design (see Fig. 3), inspired by Manakov et al. [30], captures angle-resolved time derivatives by placing a rectangular kaleidoscope in the optical path. This approach allows for capturing angular information affordably and simply at $3 \times 3$ views. The galvanometer design (see Fig. 4) uses a galvanometer positioned in the optical path to capture angular derivatives directly. By controlling the mirrors at $250\mathrm{Hz}$ , the incoming rays are shifted along a curve, enabling the capture of both angular and temporal derivatives. As angular derivatives are sparse [21, 56], this design does not significantly increase bandwidth requirements. Table 1 compares the relative advantages of these two hardware approaches. + +To systematically evaluate the relative merits of the hardware designs, we build a Blender plugin to render event-fields with physics based light simulation. We also build hardware imaging prototypes with kaleidoscope and galvanometer, and algorithms to demonstrate new imaging capabilities in challenging environments. These include post-capture refocusing on dynamic scenes (e.g., a rotating fan at 480 RPM), focus tracking of nef darts traveling at $20\mathrm{m / s}$ , high-speed depth reconstruction, and high-resolution, high dynamic range light fields. + +Contributions. To state explicitly, the manuscript makes the following contributions: + +- Introduces event fields; spatial and temporal multiplexing techniques to capture them. +- Builds a physics-based simulator and two hardware prototypes to systematically evaluate the merits and limitations of the multiplexing techniques. +- New applications and related algorithms, including slow motion refocus (SloMoRF), true-depth estimation, and high dynamic range light field imaging. +- Open source simulator, data, and processing codes [43]. + +
KaleidoscopeGalvanometer
DerivativesTemp.Ang. + Temp.
Recon. PerformanceScene dep.Scene independ.
ColorYesNo
Static LF CaptureNoYes
Dynamic LF captureYesYes
Refocus FidelityChoppySmooth
Bandwidth RequiredLowHigh
+ +Table 1. Relative advantages of kaleidoscope vs. galvanometer design for capturing event fields. + +# 2. Related Work + +Light field sensing: One of the first approaches to light field capture [54] involves using multiple sensors to gather angular information from various perspectives. However, this setup is bulky, costly, and requires high bandwidth communication. An alternative is time-sequential capture, which mounts a sensor on a mechanical gantry to capture the light field of a static scene [12, 25]. Rather than moving the sensor itself, a mirror can be used to redirect the view. Ihrke et al. [19] proposed using a movable planar mirror to shift the view temporally. However, this method requires individual calibration for each image, adding complexity. Fujii et al. [8] uses a galvanometer and pairs it with a high-speed camera. Inspired by their approach, we developed a similar design, precisely controlling the galvanometer's movement. A key distinction is that our galvanometer operates continuously, while their system selects discrete positions from a lookup table, making it slower. + +In addition to time-sequential capture, angular information can also be measured with spatial multiplexing. One way to achieve this is through a microlens array [39], which has been widely adopted in commercial light field cameras like Raytrix [44] and Lytro [28]. While effective, microlens arrays are challenging to manufacture and install. Taguchi et al. [48] suggested using a mirror ball to expand the camera's field of view, though this setup results in the camera being visible in the mirror ball's reflection. Alternatively, Manakov et al. [30] introduced a kaleidoscope-like mirror array that can be attached to a standard camera without modification. Inspired by this design, we adapted it to convert a conventional event camera into a device capable of capturing the event field. + +Event camera designs: From the first neuromorphic sensor by Mahowald et al. [29] to the recent Sony IMX636, event sensing saw exponential growth. We briefly review designs most relevant to our manuscript and refer the readers to recent surveys [6, 10]. Event cameras are regularly used in robotics and autonomous navigation for optical flow computation [2, 11, 52], feature tracking [49], and motion segmentation [34, 57], but event cameras cannot detect edges in the direction of the camera motion. Inspired by microsac + +cades in human eye movement, He et al. [17] placed a rotating wedge mask before the event camera aperture to capture edges in all directions. Similar to them, our galvanometer design detects edges oriented in all directions, independent of camera and object motion, as shown in Fig. 6. Our galvanometer design also finds inspiration from the designs of Mugalikar et al. [37], who used a rotating polarizer in the aperture plane to capture high-fidelity polarization information for shape-from-polarization and Yu et al. [55], who used a rotating light source and event camera for fast photometric stereo. + +Habuchi et al. [16] captured light fields for static scenes by placing an LCOS device in the aperture plane of event cameras and using CNN to recover the light field. While the technique is quite novel, it is limited to static scenes due to the LCOS response time (22 ms). Guo et al. [14] placed a microlens array (MLA) on the event sensor to capture the light field for dynamic scenes and showed microscopic imaging applications with their prototype. We built our prototypes for macroscopic applications, and our kaleidoscope design is functionally same as MLA design. However, the galvanometer design can capture angular information even for static scenes without trading off the spatial resolution like Guo et al. [14]. + +Event processing algorithms: Event cameras capture scene information at higher temporal resolution (in $\mu$ s) and dynamic range (140 dB) compared to standard cameras (60 dB) [10]. These features enable solutions to challenges where standard frame-based sensors struggle, such as high-speed motion estimation [9], high frame rate reconstruction [45], and high dynamic range imaging [45]. Besides, event cameras and standard frame-based cameras are complementary; event cameras provide high-speed, sparse data, while standard cameras capture color and fine details. By fusing these two types of sensors, algorithms like Timelens [50] and Timelens++ [51] effectively interpolate RGB frames using event information. We adapt these event processing algorithms for processing event field data. + +Other event-based fusions: Apart from RGB cameras, other computational imaging systems can also be made fast by fusing with event sensors [58]. Event-based structured light sensing techniques [18, 23, 31, 32, 35, 47] reconstruct 3D scenes at high speed by replacing standard intensity sensors with event cameras. Event cameras are also used for foveated depth sensing [36] to scan moving objects at a higher spatial resolution. Muglikar et al. [38] combined single-photon avalanche diodes (SPADs) with event cameras to achieve high-speed, low-light, low-bandwidth imaging capabilities. + +# 3. Technical Background + +We will briefly review the image formation models for light field and event cameras that are useful in developing the + +mathematical models for event fields. + +# 3.1. Light Field + +The plenoptic model of light $L(\vec{x}, \vec{\omega}, \lambda, t)$ is a comprehensive representation that captures all the dimensions of light emitted by a scene [1, 33]. Here, $\vec{x} = (x, y, z)$ represents the spatial coordinates of the observer through which the light passes; $\vec{\omega} = (\omega_x, \omega_y)$ are the angular coordinates; $\lambda$ is the wavelength, and $t$ is the time. At a given time, the measurement of a monochromatic light is 5D function. Levoy and Hanrahan [25] and Gortler et al. [12] showed that the 5D representation has redundancies and have reduced it to 4D by assuming that the light field is measured in free space. Under this assumption, as the radiance of a ray remains a constant in a straight line, the static monochromatic light field is a 4D function $L(\vec{u}, \vec{s})$ , where $\vec{u} = (u, v)$ is a point on the virtual sensor and $\vec{s} = (s, t)$ is a point on the virtual aperture. + +To synthesize an image from a given light field, we integrate the light field over the aperture as: + +$$ +I (x, y) = \int_ {\vec {s}} L (\vec {u}, \vec {s}) \mathrm {d} \vec {s}. \tag {1} +$$ + +Light field allows us to change the focus plane after imaging by computationally changing the distance between the sensor and aperture planes [3, 25]. If $d_0$ is the original distance between the planes and $d$ is the distance required to focus at a given depth, the light field that focuses at $d$ will be $L_{d}(\vec{u},\vec{s}) = L(\vec{u} +(d_{0} / d)\vec{s},\vec{s})$ . The refocused image can be synthesized by integrating angles, as shown in Eq. (1). + +# 3.2.Event Camera + +An event camera asynchronously records the brightness $(B\dot{=} \log (\int_{\vec{\omega}}Ld\vec{\omega}))$ change event for each pixel independently [4, 26, 29, 42]. Specifically, in an ideal and noise-free scenario, a pixel $\mathbf{x} = (x_k,y_k)^T$ will be triggered at time $t_k$ as soon as the brightness change reaches a predefined threshold $C$ since the last triggered event. The polarity $p_k\in \{+1, - 1\}$ , which is the sign of the brightness change, will be recorded for each trigger. Mathematically, + +$$ +\Delta B \left(\mathbf {x} _ {k}, t _ {k}\right) \doteq B \left(\mathbf {x} _ {k}, t _ {k}\right) - B \left(\mathbf {x} _ {k}, t _ {k} - \Delta t _ {k}\right) = p _ {k} C +$$ + +$$ +\frac {\partial B \left(\mathbf {x} _ {k} , t _ {k}\right)}{\partial t} \approx \frac {p _ {k} C}{\Delta t _ {k}} \Rightarrow \frac {\partial L \left(\mathbf {x} _ {k} , t _ {k}\right)}{\partial t} \approx L \frac {p _ {k} C}{\Delta t _ {k}}. \tag {2} +$$ + +Therefore, event cameras capture a high-speed temporal derivative slice $(\mathbf{x}, t_k)$ of the plenoptic function. + +In the related work, we referred to several techniques that excel at extracting high-speed spatial information from sparse event streams. These algorithms can do so as the scene information typically changes slowly [10], making the temporal gradient $\frac{\partial L(\mathbf{x}_k,t_k)}{\partial t}$ inherently sparse. This redundancy also applies to angular dimension as light field + +is mostly smooth [21, 56] and hence, $\frac{\partial L(\vec{x},\vec{\omega},\vec{s},\lambda,t)}{\partial\vec{\omega}}$ and $\frac{\partial L(\vec{x},\vec{\omega},\vec{s},\lambda,t)}{\partial t}$ are sparse. We exploit this sparsity to capture event fields with a single event camera in the next section. + +# 4. Event Field + +As mentioned briefly in Sec. 3.2, an event camera's photodetector measures brightness by taking the logarithm of radiance integrated over all angles. This integration results in a loss of angular information. Our objective, through event fields is to retain this angular information $\vec{\omega}$ . To achieve this goal, we need to multiplex angular information into either the spatial or temporal dimensions that the event camera captures. We advocate for and design both spatial and temporal multiplexing techniques, as they offer complementary advantages. + +# 4.1. Spatial Multiplexing + +We encode discrete angular information $(\omega_{i})$ into the spatial dimension using a mapping function defined as: + +$$ +\left(\omega_ {i}, \mathbf {x} _ {\mathbf {s}} \left(\omega_ {i}\right)\right) \mapsto \mathbf {x}, \tag {3} +$$ + +where $\mathbf{x}_{\mathbf{s}}(\omega_{i})$ represents the spatial coordinate for each measured view $(\omega_{i})$ . Let the number of discrete angular views be $\mathbf{n}$ , the spatial resolution of the sensor be $\mathbf{r}$ , and $\mathbf{r}$ is an integral multiple of $\mathbf{n}$ . The specific mapping function $\mathbf{x}_{\mathbf{s}}$ varies based on the hardware design. The two commonly used mapping functions are: + +- Multi-sensor array [54] or kaleidoscope [30]: In this design, the mapping function is $\mathbf{x}_{\mathbf{s}} = \mathbf{x} \mod (\mathbf{r} / \mathbf{n})$ . +- Microlens array [39]: In this design, the mapping function is $\mathbf{x}_{\mathbf{s}} = \mathbf{x} / / \mathbf{n}$ . + +By applying this mapping function to Equation (2), the measurements are: + +$$ +\frac {\partial B}{\partial t} (\underbrace {\mathbf {x} _ {\mathbf {s}} , \omega} _ {\mathbf {x} _ {k}}, t _ {k}) \approx \frac {p _ {k} C}{\Delta t _ {k}}, \tag {4} +$$ + +and the event camera will now measure the temporal brightness changes across both spatial and angular dimensions. Note that the brightness derivatives are still related to time, and we lose spatial resolution (by a factor of $n$ ) to improve angular resolution. + +# 4.2. Temporal Multiplexing + +To encode angular information in the temporal dimension, we steer the incoming rays along a closed periodic curve $\mathcal{C}(t)$ with period $T$ . This steering could be accomplished by either moving the camera or placing steerable galvanometer in the optical path of the camera. The galvanometer can be steered fast along Lissajous curves [8, 41]; hence, we use it in all our experiments. + +The temporal multiplexing does not sacrifice spatial resolution but sacrifices temporal resolution and always detects + +![](images/86b90fa7ac7414ed93aaa2daaa20d00b87b480c3f9fbfac983165ae3c05527c4.jpg) +(a) Plugin UI +Figure 2. Blender plugin for physics-based event field rendering. + +![](images/bb8e23a5fd75e89c232c185f70276e8fa49648c9c9eedce900347b340876faa2.jpg) +(b) Event simulation + +![](images/88d04abcbab4db534f7df759d63506c2a98a970444bb041d75c639e9fc631d20.jpg) +(c) Event field + +more events than spatial multiplexing. At any given time instant, the angular view will be at $\mathcal{C}(t \bmod T)$ and the temporally varying light field is $L(\mathbf{x}, \mathcal{C}(t \bmod T), t)$ . For static scenes, the event camera measurement is + +$$ +\frac {\partial B (\mathbf {x} , \mathcal {C} (t))}{\partial t} = \left. \frac {\partial B (\mathbf {x} , \omega)}{\partial \omega} \frac {\partial \mathcal {C} (t)}{\partial t} \right| _ {t = t _ {k}} \approx \frac {p _ {k} C}{\Delta t _ {k}}, \tag {5} +$$ + +Therefore, temporal multiplexing technique measures the angular derivatives of the light field $\frac{\partial L(\mathbf{x},\omega)}{\partial\omega}$ along the scanning curve. + +While we restricted the scanning curves to be Lissajous curves, specifically circles in most of the results, in principle, the scanning curve can be any periodic curve, including space-filling curves [15, 46]. Space-filling curves result in dense angular light-field capture at the cost of loosing temporal information. However, typically light field is highly redundant and smooth and a circular scan captures most of the light-field diversity as we show in the results. + +# 5. Simulator and Hardware Design + +To validate our approach, we built a physics-based simulator and two hardware prototypes, kaleidoscope for spatial multiplexing and galvanometer for temporal multiplexing. The simulator allows us to precisely control the scene and design parameters, enabling an accurate and fair comparison between the two multiplexing techniques. Hardware prototypes show the real-world feasibility of our designs as well as new capabilities. + +# 5.1. Simulator + +We have used a straightforward approach to simulating events, which involves rendering the full plenoptic function, calculating brightness by propagating the plenoptic data through optics, and tracking the brightness difference from the previous event. A new event is triggered for each pixel when this difference exceeds a manually set threshold. + +We chose Blender [7] as our backend as it is opensource, physically based, supports GPU acceleration (Cycles), Python APIs, and brandishes a user-friendly GUI. We + +![](images/6dd46700a738085abdfa4ab4767a2f944cc7455ace35316672506cea806efe91.jpg) +(a) Ray sketch + +![](images/50b2831ae951dd4128427ed6f7cf7fe73ba816fd94926515f2b60beaaed1ec89.jpg) +(b) Hardware setup + +![](images/823beea596d251a87a5722b774b475a27a2a852abb00dec7f33fe54511801098.jpg) +Figure 3. Kaleidoscope design. (a) Illustration shows rays from various angles being directed towards the event camera. (b) The main lens focuses the scene onto the kaleidoscope, splitting the view into multiple angles. The beam splitter directs these multiple views to the RGB camera for standard light-field imaging and the event camera for event-field imaging. +(a) Ray sketch +Figure 4. Galvanometer design. (a) Ray sketch illustrating the angular separation achieved by controlling mirror angles in the galvanometer setup. (b) Hardware setup showing the galvanometer mirror system integrated with the event camera. This design allows for capturing angular derivative events by dynamically controlling the mirror position. + +![](images/859904120910af13932cb06d70c1d9731da781e99f5723cf2e65fe8375a7de07.jpg) +(b) Hardware setup + +developed a Blender plugin to render event fields for either spatial or temporal multiplexing design, along with a simple GUI to facilitate community use. The GUI and sample simulation results are presented in Fig. 2. + +# 5.2. Kaleidoscope for Spatial Multiplexing + +A kaleidoscope is a common optical instrument composed of multiple first-surface mirrors. Manakov et al. [30] introduced this optical device as an add-on for sensor fusion. As shown in Fig. 3, this kaleidoscope is a rectangular design made of four mirrors. The lens at the front forms an image at the entrance of the kaleidoscope. After multiple reflections within the kaleidoscope, the rear camera captures the image, resulting in several flipped images, each with a different virtual camera location. Therefore, the light field information is spatially multiplexed onto the camera sensor. + +# 5.3. Galvanometer for Temporal Multiplexing + +While the kaleidoscope design offers a simple solution without moving parts, it has two key limitations. First, although it provides angularly separated events, the events + +still represent the derivative of light intensity over time. Second, the spatial resolution of current event cameras is not very high (1 megapixel), and this resolution would further decrease with spatial multiplexing. Temporal multiplexing offers an alternate approach. + +For temporal multiplexing, we place a galvanometer, as shown in Fig. 4, in the optical path of the event camera. A galvanometer consists of two mirrors that can be precisely adjusted at high speed (Eg. GVS012 from Thorlabs can scan a thousand Lissajous curves per second). Moving mirrors allow us to rotate the camera virtually at a high speed and capture the event field. In this design, the captured information represents the derivative of light intensity with respect to angle, rather than time, as we discussed in Sec. 4.2. + +# 6. Results + +We evaluate the proposed techniques on various scenes. We first summarize the observations and then provide details. + +Simulations demonstrate that each design is optimal in different contexts. The kaleidoscope (spatial multiplexing) excels in capturing dynamic scenes with high frame rates but suffers from lower spatial resolution and is ineffective for static scenes, as no events are generated without motion. The galvanometer (temporal multiplexing) captures both static and dynamic scenes at full spatial resolution by recording angular derivatives but introduces motion blur if the object motion is significant. + +Real-world applications demonstrate each design's post-capture refocusing capabilities. Specifically, with the galvanometer setup, we show high-resolution refocusing of static and dynamic parts of the scene. With the kaleidoscope setup, we perform sensor fusion with an RGB camera, providing vibrant, high-speed refocusing for dynamic scenes. These results underscore the complementary nature of the two designs: the kaleidoscope for high-speed dynamic capture and the galvanometer for detailed, high spatial-angular light field recovery. + +In the following sections, we compare our proposed hardware designs (see Section 6.1), followed by applications for each setup. Specifically, we explore HDR light fields (Section 6.2) and slow-motion refocusing (Section 6.3) using the kaleidoscope configuration, as well as high-speed post-capture refocusing (Section 6.4) and instant true depth estimation (Section 6.5) using the galvanometer-based design. + +# 6.1. Kaleidoscope vs. Galvanometer Designs + +In practical applications, differences in optical properties—such as focal lengths—make it challenging to compare two cameras under identical conditions. The simulator, however, allows us to control these variables, enabling a fair comparison of both designs. We, therefore, used Blender to create a scene containing swirling water and moving objects + +![](images/1ed24bd2bc9e9ba0b8472bc30d936e258785ab47d19c7e16cd02810038c5327a.jpg) +(a) Ground truth +Figure 5. (Simulation) comparison of spatial (kaleidoscope) and temporal (galvanometer) multiplexing techniques. (a) A simulated scene with swirling water and objects, with a stationary box. (b) Refocused image using the kaleidoscope design, showing reduced spatial resolution. However, the image quality is independent of scene motion. The static wooden box is not reconstructed. (c) Refocused images using the galvanometer design at different object movement speeds $(1\times,2\times,4\times,$ and $8\times)$ while maintaining a constant scanning speed. $1\times$ corresponds to about 2.5 pixels per scan for the highlighted object. As the scene motion increases, the highlighted object, unlike in the kaleidoscopic design, suffers from motion blur. The static wooden boxes are perfectly reconstructed, unlike the kaleidoscopic design. + +![](images/1eba0882cab0721c7c39d884d1fc89e229b5464d4f02d647785fc6ac5e5a422b.jpg) +(b) Kaleidoscope + +![](images/44c57d8c6e6a88c733751cc1355de01903d79921b708ce674e157e346f92d1e2.jpg) +$\therefore m : x = 1$ 或 ${3x} + {4y} + 1 = 0$ + +![](images/6e60ca239c1ab5b1c113110a9c66e3d958e604534e9f4672f232fa131a47a101.jpg) +(c) Galvanometer + +![](images/3770a4352c4c5efabc26a0bdec7c88bfb47ffc79df026630d39a2dcb311acd61.jpg) + +![](images/bfd7ab5e273ff5f747de5775503a5198ef781a020dee2b30d7179449d109fdf1.jpg) + +![](images/8a513dfb20b468368cda8b74a09f220ee439f8c57ae423b09618ff9a00a5ab44.jpg) +(a) Scene + +![](images/1ac3b43a8af3f4f557cdea92ee0553db7d2a3c5ebba450872cd4ea3e95941350.jpg) +(b) Kaleidoscope + +![](images/7956fe40f31be6cb50f7413fdd6ba6571a81ccd3720bcc99fdd2a6087e30abb9.jpg) +(c) Galvanometer + +![](images/7781054d13be25ba622b52a449f446168ccfc0becd61ab29dfe57aeaa8d33390.jpg) +Figure 6. (Experimental) comparison of multiplexing techniques. (a) Reference scene with a grid pattern. (b) When the grid moves horizontally, the kaleidoscope design cannot capture horizontal lines as intensity does not change along this axis. (c) Using a galvanometer that scans in a circular motion ensures complete capture, avoiding information loss in the angular domain. +(a) Short exposure +Figure 7. High dynamic range (HDR) light field. A high dynamic range scene is captured with a standard RGB camera using (a) short exposure and (b) long exposure, and both fail to capture the full dynamic range. In contrast, (c) the HDR light field reconstructed from the event field effectively captures the complete dynamic range of the scene. + +![](images/66a298453a73c0e08f52cf32653edb2a5901148849ba2270d1ebafabbb3ae7f0.jpg) +(b) Long exposure + +![](images/acdb913614ed441d2d25a34dce2e30afea51b3994540c610afa71b56004d7f82.jpg) +(c) HDR light field + +within a stationary box (see Fig. 5) to assess the strengths and limitations of each design. + +The first significant difference is that the galvanometer design can fully reconstruct scene information by capturing angular derivatives, whereas the kaleidoscope design can only capture moving objects, as no events are generated in static regions. The stationary wooden box is clearly visible in the galvanometer design but is absent in the kaleidoscope design. We also observe this in a real experiment (Fig. 6), where a rectangular mesh is translated horizontally, and the kaleidoscope fails to reconstruct horizontal lines. + +A second key difference is that the galvanometer design multiplexes the event field into the time domain, making it susceptible to motion blur from rapidly moving objects. This is visible in the zoomed insets in Fig. 5. The kaleidoscope design, by contrast, avoids motion blur but instead suffers from lower spatial resolution. However, both designs can capturing and refocusing dynamic objects. + +# 6.2. HDR Light Fields + +Event fields are not only high-speed but also high-dynamic range light fields. We demonstrate this in Fig. 7, where the scene containing bright LED light in the foreground and diffuse texture in the background is captured using both an RGB camera and the event camera through a kaleidoscope. The event fields have a large dynamic range and capture the spatial structure of both LEDs and the background texture, whereas the RGB camera either captures the light sources (short exposure) or the background (long exposure). + +# 6.3. SloMoRF: Slow Motion Refocusing + +We augment the kaleidoscope design with a co-located RGB camera to capture and reconstruct slow-motion RGB frames with dynamic refocusing capability, called SloMoRF. The hardware setup, shown in Fig. 3, features a beam splitter for identical kaleidoscope views for RGB and event camera without interfering with the main lens. + +We utilize the TimeLens algorithm [50] for sensor fusion1. Using the captured event data, we interpolate frames to generate a $5 \times$ higher frame-rate RGB video (from 120 to 720 fps). The interpolated results are presented in Fig. 8. + +This approach allows us to generate a high-speed color light field. We treat this light field as unstructured as we found it empirically easier than calibrating the views. To align the images, we estimate the necessary shift at a specified patch using a basic pattern-matching algorithm. Note that this shift computation is only needed once for the entire video sequence as long as the target focal plane remains fixed. Figure 8 shows refocusing results across three focal planes. Additionally, the captured event field allows for focus tracking on a high-speed object moving along the depth. We demonstrate this capability in Fig. 9, where we visually compare sequence of images that focus on a fast-moving toy dart vs. focusing solely on the background. These results demonstrate the high frame rate, color, and refocusing capabilities of our kaleidoscope design when combined with the state-of-the-art sensor fusion algorithms. + +![](images/fb33bd42cc41b131abcf7fe0a42c4e0250630c68d0f8d618162e5e321341e4c0.jpg) +(a) Scene +(b) Captured frame (with aligned events) + +![](images/90e65e349a4af66a24d58c54e194ed748da79af488be8ad830f35e192abe0c6e.jpg) +Frame 19 + +![](images/904f98013420f21178a1aa8c47a5f09623e32975064fa4a7e8f58d75430122b3.jpg) +1st interpolated +(c) Refocus with interpolated frames + +![](images/f6d1cf29d2fdd0ac59fd9c741866429c4a3e6bb6f1ec36389ee18ec622697905.jpg) +3rd interpolated + +![](images/07067c492c2cf50501d9a50c1f71446f626788a57ebd02562df0cb53cc6b8319.jpg) +5th interpolated + +![](images/842a969a96ff0334f4d83ea8bf90032e8b336bae661158ca2ba4926bbe4caee7.jpg) +Frame 20 + +![](images/59ecbe89c21831a2b98979d831edf6e25c28d706977e95d26c536f85a7e1f1f0.jpg) +Figure 8. SlowMoRF: Slow motion refocusing with kaleidoscope and sensor fusion. (a) Original setup with Lego pieces floating in a water tank as a wooden brick is thrown in, with CVPR in the background, (b) Captured raw frames with aligned events using kaleidoscope design, and (c) Interpolated refocused frames at three focal planes (back, mid, and front). Between the frames, the fall of the wooden block and changes in water splash and caustics can be observed in the front refocus. Mid and back refocus show sharp Lego toys and CVPR. This setup achieves high-speed (720 fps) refocusing with enhanced clarity across focal depths. The supplementary contains the full video. +(a) Refocus on the bullet + +![](images/c6ff010469d5b771780b691cd597a4560909f4bc096566cbd57aa2c7ab54ad2e.jpg) + +![](images/1b2dc2885cb5471ff7398ad3cc5275d9610971667e2b4890909c8b9860d0178f.jpg) + +![](images/bc487c107a240e654d62835e933e6046e524f389be0f8dcca9679ae859d03fc2.jpg) +(b) Refocus on the background +Figure 9. High-speed refocusing on a toy dart. The original video was captured at 120 fps, interpolating seven frames between each pair of original frames, resulting in a 960 fps high-speed light field video. The images shown here are all the middle (the fourth) interpolated frames, demonstrating refocusing on two planes: (a) the moving dart and (b) the background. Current high-speed videography cannot adjust focus during capture, so dynamic objects must remain at a fixed focal plane to stay sharp. Our technique eliminates this limitation and can focus on dynamic object. + +![](images/8966ffc4b44ca2314a2467ffc09c39a5472a47d2cfe03aef8e38f5330c6e0c36.jpg) + +![](images/390a2d3a2cca213bde6f1e2acd506402a20598d235bd8b3388ba2a40a8879bcb.jpg) + +# 6.4. High-Speed Post-capture Refocusing + +In this section, we test the post-capture refocusing ability of the galvanometer setup on dynamic scenes. Compared to unstructured event field captured by kaleidoscope design, our galvanometer design produces a fully structured event field after calibration, as we precisely control the galvanometer with a signal generator to determine the angle relative to time. The supplementary materials provide calibration details. + +Using E2VID, we generate a high-speed video at 10,000 fps (0.1 ms per frame) from the captured event field data. The galvanometer is operating at $250\mathrm{Hz}$ . Therefore, we reconstruct the light field at 250 fps and 40 views, the fastest recorded so far at a megapixel resolution. + +![](images/6047bd58e4da46426b4ce53b7f965386231a186521235a2536d87b799113d92e.jpg) +(a) Grayscale reconstructed images + +![](images/034afb9cd748ca08e532e0b704a59c1231d0732977c4e1b978fb8ae076219389.jpg) + +![](images/ac7386f86882bdfb9e07bc6a3268771f7e6ff69fbc8002f4ec9e4c932468f41e.jpg) +(b) Scene captured by ordinary camera (60Hz) +Figure 10. Dynamic scene refocusing. Using our galvanometer design at $250\mathrm{Hz}$ , we capture a scene with a rotating fan marked with "CVPR", spinning at $300\mathrm{rpm}$ . (a) Using prior knowledge of the scanned mirror's location, we refocus on the background or the rotating fan in the foreground. (b) An ordinary camera captures the scene at $60\mathrm{Hz}$ , focusing on the background and foreground from different viewpoints. The camera lacks refocus ability and causes motion blur on the fan due to the slower shutter speed. + +![](images/e2e7124817fc22ac5b71f79dab14ab5fdbbce15d8013bd4a416633e2b060b70a.jpg) + +Using this light field and equations in Sec. 3.1, we can do post-capture refocusing. Note that even though E2VID generates noisy frames at high speed (due to low light), the noise gets averaged out when we synthesize a refocused image. In Fig. 10, we show refocusing results on the background texture (left figure) or foreground fan (right figure) and can observe that they are noise-free. We further stress-tested our hardware setup in Fig. 11 and noticed that the event camera begins to lose events due to the readout bandwidth limit when the galvanometer operates beyond $250\mathrm{Hz}$ . + +![](images/41072534c48c6633677789e53b10f0ab81a6898c82d177d1db7c895730a9fda7.jpg) +Event frame Front refocus Back refocus +Figure 11. Effect of galvo scan frequency for refocusing accuracy. The scene consists of two static objects positioned in the foreground and background, along with a fan (highlighted in yellow boxes) in the foreground rotating at $480\mathrm{RPM}$ . The static objects (highlighted in red and blue boxes) remain sharp up to a scanning frequency of $200\mathrm{Hz}$ but become increasingly blurry as the scan speed increases. This counter intuitive phenomenon occurs as event camera reaches its maximum bandwidth capacity, resulting in random event loss, as seen in the Event frame row. However, the text on the rotating fan becomes clearer at higher scan speeds due to reduced motion blur. + +![](images/e00f44c110d3db29a87ccb121aa6c6404f3aaaf5a8a00b360c0985c9f108905e.jpg) +Figure 12. Instant true depth estimation at $100\mathrm{Hz}$ . Depth mapping of a running person, shown in three frames with color-coded depth scale (in inches), highlighting the system's capability for rapid, accurate depth measurement in dynamic scenes. + +# 6.5. Instant True Depth Estimation + +In addition to refocusing, our galvanometer design enables real-time true depth estimation. Unlike SLAM-based methods [5], our approach provides instant true depth estimation based on the event field. As outlined in Sec. 6.4, our galvanometer design generates a structured light field, requiring only a single calibration step to determine the depth-to-disparity ratio. We employ the "depth from focus" algorithm [13] to calculate the depth of each pixel. This involves first constructing a focal stack by refocusing the light field at various depths. For each pixel, we determine the focal stack image where the pixel appears sharpest, and the depth associated with that image is the depth of the pixel. + +By adopting this approach, we obtain the instant true depth at the same frequency as the galvanometer's scanning speed. As shown in Fig. 12, a $100\mathrm{Hz}$ galvo scan successfully captures a person running away from the camera, offering sparse yet real-time depth information of the letters on their shirt at a rate of 100 fps. + +# 7. Discussion and Future Work + +In this manuscript, we introduce "event field", a new approach for capturing high-speed, high resolution, HDR light fields using two optical configurations—kaleidoscope and galvanometer—combined with an event camera. Our kaleidoscope design enables capturing temporal derivatives for multiple views, while the galvanometer design captures angular derivatives along curves. We validate both designs through simulations and real-world data, demonstrating their unique advantages. + +Our experiments showcase several potential applications. Most notably, our setup enables high-speed refocusing on fast-moving objects, real-time true depth estimation, and high-dynamic-range light field capture. Supplementary PDF and video provides further visualizations of these results across additional scenes. + +While our approach shows several new possibilities, it also creates new research problems in the graphics, imaging, and vision fields. Our current brute-force simulation of the event field could benefit from optimal importance sampling techniques that improve the rendering speed by orders of magnitude similar to previous importance techniques for computational cameras [20, 22, 27, 40]. Additionally, as shown in Fig. 11, static parts of the scene do not require high bandwidth. By implementing adaptive scanning (e.g., slower scans for static parts, faster scans for dynamic areas) with a MEMS array, we can enhance bandwidth efficiency. We currently use pre-trained E2VID and Timelens models, which are optimized for time-sequential event streams. Generating sufficient event field data, possibly through an optimal simulator and training models specifically for our event fields, would yield a significant performance boost. + +# Acknowledgment + +This project is funded by NSF awards 2403122, 2326904, 2107454, and 2107313 and ARPA-H award D24AC00296. +We thank Laurence Rubin at Boston Craft Works, Joseph Poissant at Dartmouth Engineering Machine Shop, and Hannah Machesky at Two Way Mirrors for providing mirrors and polishing expertise for our kaleidoscope design. +We also extend our gratitude to the RISC lab members, Atul Agarwal, Dhawal Sirikonda, Kedari Chowtoori, and Sarah Friday, for their generous assistance in capturing the data and their 3D printing and laser cutting skills. + +# References + +[1] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Vision and Modeling Group, Media Laboratory, Massachusetts Institute of ..., 1991. 3 +[2] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2 +[3] Tom E Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE transactions on pattern analysis and machine intelligence, 34(5):972-986, 2011. 3 +[4] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 3 +[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 8 +[6] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-iliidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. arXiv preprint arXiv:2408.13627, 2024. 2 +[7] Blender Online Community. Blender - a 3d modelling and rendering package, 2024. 4 +[8] Toshiaki Fujii and Masayuki Tanimoto. A real-time ray-space acquisition system. In Stereoscopic Displays and Virtual Reality Systems XI, pages 179-187. SPIE, 2004. 2, 4 +[9] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. IEEE transactions on pattern analysis and machine intelligence, 40(10):2402-2412, 2017. 3 +[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 3 + +[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 2 +[12] Steven J. Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F. Cohen. The lumigraph. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 43-54, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3 +[13] Paul Grossmann. Depth from focus. Pattern recognition letters, 5(1):63-69, 1987. 8 +[14] Ruipeng Guo, Qianwan Yang, Andrew S Chang, Guorong Hu, Joseph Greene, Christopher V Gabel, Sixian You, and Lei Tian. Eventlfm: Event camera integrated fourier light field microscopy for ultrafast 3d imaging. Light: Science & Applications, 13(1):144, 2024. 3 +[15] Mohit Gupta, Andreas Velten, Shree K Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM Transactions on Graphics (TOG), 37(2):1-18, 2018. 4 +[16] Shuji Habuchi, Keita Takahashi, Chihiro Tsutake, Toshiaki Fujii, and Hajime Nagahara. Time-efficient light-field acquisition using coded aperture and events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24923-24933, 2024. 3 +[17] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 3 +[18] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. High-speed structured light based 3d scanning using an event camera. Optics Express, 29(22):35864-35876, 2021. 3 +[19] Ivo Ihrke, Timo Stich, Heiko Gottschlich, Marcus Magnor, and Hans-Peter Seidel. Fast incident light field acquisition and rendering. Journal of WSCG, 2008. 2 +[20] Adrian Jarabo, Julio Marco, Adolfo Munoz, Raul Buisan, Wojciech Jarosz, and Diego Gutierrez. A framework for transient rendering. ACM Transactions on Graphics (ToG), 33 (6):1-10, 2014. 8 +[21] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 2, 4 +[22] Juhyeon Kim, Wojciech Jarosz, Ioannis Gkioulekas, and Adithya Pediredla. Doppler time-of-flight rendering. ACM Trans. Graph., 42(6), 2023. 8 +[23] T Leroux, S-H Ieng, and Ryad Benosman. Event-based structured light for depth reconstruction using frequency tagged light patterns. arXiv preprint arXiv:1811.10771, 2018.3 +[24] M. Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2 +[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 31-42, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3 + +[26] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 db $15\mu s$ latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 3 +[27] Yang Liu, Shaojie Jiao, and Wojciech Jarosz. Temporally sliced photon primitives for time-of-flight rendering. In Computer Graphics Forum, pages 29-40. Wiley Online Library, 2022. 8 +[28] Lytro. Lytro. https://www.lytro.com/, 2017. [Online; accessed 6-Oct-2024]. 2 +[29] Misha Mahowald and Misha Mahowald. The silicon retina. An Analog VLSI System for Stereoscopic Vision, pages 4-65, 1994. 2, 3 +[30] Alkhazur Manakov, John Restrepo, Oliver Klehm, Ramon Hegedus, Elmar Eisemann, Hans-Peter Seidel, and Ivo Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM Transactions on Graphics, 32(4):47-1, 2013. 2, 4, 5 +[31] Ashish Rao Mangalore, Chandra Sekhar Seelamantula, and Chetan Singh Thakur. Neuromorphic fringe projection profilometry. IEEE Signal Processing Letters, 27:1510-1514, 2020. 3 +[32] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 3 +[33] Leonard McMillan and Gary Bishop. Plenoptic modeling: An image-based rendering system. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 433-440. Association for Computing Machinery, 2023. 3 +[34] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14414-14423, 2020. 2 +[35] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 2021 International Conference on 3D Vision (3DV), pages 1165-1174. IEEE, 2021. 3 +[36] Manasi Muglikar, Diederik Paul Moeys, and Davide Scaramuzza. Event guided depth sensing. In 2021 International Conference on 3D Vision (3DV), pages 385-393. IEEE, 2021. 3 +[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 3 +[38] Manasi Muglikar, Siddharth Somasundaram, Akshat Dave, Edoardo Charbon, Ramesh Raskar, and Davide Scaramuzza. Event cameras meet spads for high-speed, low-bandwidth imaging. arXiv preprint arXiv:2404.11511, 2024. 3 +[39] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 4 +[40] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa + +Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 8 +[41] Luboš Poláček, Jana Jurmanová, and Zdeněk Navrátil. Laser galvo mirrors: perfect instrument for the demonstration of lissajous figures. Physics Education, 54(5):055002, 2019. 4 +[42] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 3 +[43] Ziyuan Qu. Event field. https://dartmouth-risc-lab.github.io/event-field-website/, 2024.2 +[44] A Raytrix. 3d light field camera technology, 2017. 2 +[45] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 3 +[46] Hans Sagan. *Space-filling curves*. Springer Science & Business Media, 2012. 4 +[47] Dhawal Sirikonda, Praneeth Chakravarthula, Ioannis Gkioulekas, and Adithya Pediredla. Structured light with a million light planes per second. arXiv preprint arXiv:2411.18597, 2024. 3 +[48] Yuichi Taguchi, Amit Agrawal, Srikumar Ramalingam, and Ashok Veeraraghavan. Axial light field for curved mirrors: Reflect your perspective, widen your view. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 499-506. IEEE, 2010. 2 +[49] David Tedaldi, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Feature detection and tracking with the dynamic and active-pixel vision sensor (davis). In 2016 Second International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7. IEEE, 2016. 2 +[50] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 3, 6 +[51] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 3, 6 +[52] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2 +[53] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 2 +[54] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2, 4 + +[55] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9602-9611, 2024. 3 +[56] Mingyuan Zhao, Hao Sheng, Da Yang, Sizhe Wang, Ruixuan Cong, Zhenglong Cui, Rongshan Chen, Tun Wang, Shuai Wang, Yang Huang, et al. A survey for light field superresolution. High-Confidence Computing, page 100206, 2024. 2, 4 +[57] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE transactions on neural networks and learning systems, 34(8):4868-4880, 2021. 2 +[58] Zihao Zou, Ziyuan Qu, Xi Peng, Vivek Boominathan, Adithya Pediredla, and Praneeth Chakravarthula. High-speed dynamic 3d imaging with sensor fusion splatting. arXiv preprint arXiv:2502.04630, 2025. 3 \ No newline at end of file diff --git a/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/images.zip b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..37f4a5e0d18549643cd265d5dbe5679a91dde88e --- /dev/null +++ b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8d1dd1e73515ed5caa6939d4d717b5e00ba61f2f69d0c8261178fe2ce40fd76 +size 669805 diff --git a/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/layout.json b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42b00f97beee2e5655af47ab591f8a9f2363153d --- /dev/null +++ b/2025/Event Fields_ Capturing Light Fields at High Speed, Resolution, and Dynamic Range/layout.json @@ -0,0 +1,10602 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "text", + "content": "Event fields: Capturing light fields at high speed, resolution, and dynamic range" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 144, + 166, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 144, + 166, + 172 + ], + "spans": [ + { + "bbox": [ + 70, + 144, + 166, + 172 + ], + "type": "text", + "content": "Ziyuan Qu Dartmouth College" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 144, + 275, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 144, + 275, + 171 + ], + "spans": [ + { + "bbox": [ + 188, + 144, + 275, + 171 + ], + "type": "text", + "content": "Zihao Zou \nUNC Chapel Hill" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 144, + 396, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 144, + 396, + 172 + ], + "spans": [ + { + "bbox": [ + 297, + 144, + 396, + 172 + ], + "type": "text", + "content": "Vivek Boominathan Rice University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 418, + 144, + 539, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 144, + 539, + 172 + ], + "spans": [ + { + "bbox": [ + 418, + 144, + 539, + 172 + ], + "type": "text", + "content": "Praneeth Chakravarthula \nUNC Chapel Hill" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 257, + 177, + 351, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 177, + 351, + 205 + ], + "spans": [ + { + "bbox": [ + 257, + 177, + 351, + 205 + ], + "type": "text", + "content": "Adithya Pediredla Dartmouth College" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 239, + 354 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 239, + 354 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 239, + 354 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 239, + 354 + ], + "type": "image", + "image_path": "098c5afda1924c14b4ad7c758b3330a88746163e43b5ef86b36c43b82da882a7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 359, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 359, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 555, + 415 + ], + "type": "text", + "content": "Figure 1. We introduce Event Fields, a novel framework that leverages event cameras to capture light fields at high speed, resolution, and dynamic range using innovative optical designs. Event fields integrate the unique advantages of event sensing—such as low latency, data efficiency, and power-saving capabilities—with the detailed angular information provided by light field sensing. We demonstrate the versatility of the event fields through three application scenarios: slow motion refocus (SloMoRF), instant true depth estimation, and high dynamic range (HDR) light field imaging." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 242, + 217, + 553, + 354 + ], + "blocks": [ + { + "bbox": [ + 242, + 217, + 553, + 354 + ], + "lines": [ + { + "bbox": [ + 242, + 217, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 242, + 217, + 553, + 354 + ], + "type": "image", + "image_path": "542431f7ab310f6ad4c9d2f7d41f5a09f622f56f30b307a044ff5b6d81ef728c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 425, + 200, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 425, + 200, + 437 + ], + "spans": [ + { + "bbox": [ + 152, + 425, + 200, + 437 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 450, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 450, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 295, + 593 + ], + "type": "text", + "content": "Event cameras, which feature pixels that independently respond to changes in brightness, are becoming increasingly popular in high-speed applications due to their lower latency, reduced bandwidth requirements, and enhanced dynamic range compared to traditional frame-based cameras. Numerous imaging and vision techniques have leveraged event cameras for high-speed scene understanding by capturing high-framerate, high-dynamic range videos, primarily utilizing the temporal advantages inherent to event cameras. Additionally, imaging and vision techniques have utilized the light field—a complementary dimension to temporal information—for enhanced scene understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "In this work, we propose \"Event Fields\", a new approach that utilizes innovative optical designs for event cameras to capture light fields at high speed. We develop the underlying mathematical framework for Event Fields and introduce two foundational frameworks to capture them practically: spatial multiplexing to capture temporal derivatives and temporal multiplexing to capture angular derivatives. To realize these, we design two complementary optical setups—one using a kaleidoscope for spatial multiplexing and another using a galvanometer for temporal multiplexing. We" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 534 + ], + "type": "text", + "content": "evaluate the performance of both designs using a custom-built simulator and real hardware prototypes, showcasing their distinct benefits. Our event fields unlock the full advantages of typical light fields—like post-capture refocusing and depth estimation—now supercharged for high-speed and high-dynamic range scenes. This novel light-sensing paradigm opens doors to new applications in photography, robotics, and AR/VR, and presents fresh challenges in rendering and machine learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 548, + 394, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 394, + 561 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 394, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "Capturing high-speed visual phenomena in real-time is essential for advancing science and engineering. Traditional high-speed frame cameras like Phantom, Photron, and IDT can achieve impressive frame rates (up to 100,000 fps) but are bulky, have finite capture duration, and are expensive (exceeding $100,000), restricting accessibility. In contrast, event cameras offer a more affordable and efficient solution with high temporal resolution (down to " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "1\\mu \\mathrm{s}" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": ") and continuous data acquisition, suitable for real-time applications. Unlike traditional cameras, which require substantial bandwidth (e.g., 100,000 fps, HD capture requires 100 GBps), event cameras capture only changes in the scene using asyn" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26910" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "chronous pixel readouts, reducing bandwidth demands and enabling continuous streaming. This approach has recently been harnessed for high-speed applications, including optical flow, object tracking, high-dynamic range imaging, structured light scanning, and photometric stereo [6, 10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 135, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 295, + 291 + ], + "type": "text", + "content": "In this work, we extend event camera capabilities by introducing angular capture, allowing for post-capture manipulation of high-speed video—a previously unexplored functionality. Capturing angular radiance information is known as light field imaging [24, 25, 53], enabling features such as post-capture refocusing and depth estimation. We propose \"Event Fields\" that record radiance derivatives across angular and temporal dimensions, extending light field imaging principles to high-speed, high-dynamic-range scenarios (see Fig. 1). To capture these Event Fields, we introduce two complementary hardware systems: a kaleidoscope-based design and a galvanometer-based design, each with unique advantages for this purpose." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 294, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 450 + ], + "type": "text", + "content": "The kaleidoscope design (see Fig. 3), inspired by Manakov et al. [30], captures angle-resolved time derivatives by placing a rectangular kaleidoscope in the optical path. This approach allows for capturing angular information affordably and simply at " + }, + { + "bbox": [ + 55, + 294, + 295, + 450 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 294, + 295, + 450 + ], + "type": "text", + "content": " views. The galvanometer design (see Fig. 4) uses a galvanometer positioned in the optical path to capture angular derivatives directly. By controlling the mirrors at " + }, + { + "bbox": [ + 55, + 294, + 295, + 450 + ], + "type": "inline_equation", + "content": "250\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 294, + 295, + 450 + ], + "type": "text", + "content": ", the incoming rays are shifted along a curve, enabling the capture of both angular and temporal derivatives. As angular derivatives are sparse [21, 56], this design does not significantly increase bandwidth requirements. Table 1 compares the relative advantages of these two hardware approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 573 + ], + "type": "text", + "content": "To systematically evaluate the relative merits of the hardware designs, we build a Blender plugin to render event-fields with physics based light simulation. We also build hardware imaging prototypes with kaleidoscope and galvanometer, and algorithms to demonstrate new imaging capabilities in challenging environments. These include post-capture refocusing on dynamic scenes (e.g., a rotating fan at 480 RPM), focus tracking of nef darts traveling at " + }, + { + "bbox": [ + 55, + 453, + 295, + 573 + ], + "type": "inline_equation", + "content": "20\\mathrm{m / s}" + }, + { + "bbox": [ + 55, + 453, + 295, + 573 + ], + "type": "text", + "content": ", high-speed depth reconstruction, and high-resolution, high dynamic range light fields." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "type": "text", + "content": "Contributions. To state explicitly, the manuscript makes the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 604, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 604, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 628 + ], + "type": "text", + "content": "- Introduces event fields; spatial and temporal multiplexing techniques to capture them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "content": "- Builds a physics-based simulator and two hardware prototypes to systematically evaluate the merits and limitations of the multiplexing techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 664, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 295, + 700 + ], + "type": "text", + "content": "- New applications and related algorithms, including slow motion refocus (SloMoRF), true-depth estimation, and high dynamic range light field imaging." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 700, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 289, + 712 + ], + "type": "text", + "content": "- Open source simulator, data, and processing codes [43]." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 552, + 178 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 552, + 178 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 552, + 178 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 552, + 178 + ], + "type": "table", + "html": "
KaleidoscopeGalvanometer
DerivativesTemp.Ang. + Temp.
Recon. PerformanceScene dep.Scene independ.
ColorYesNo
Static LF CaptureNoYes
Dynamic LF captureYesYes
Refocus FidelityChoppySmooth
Bandwidth RequiredLowHigh
", + "image_path": "1fe09fcc3e7f30b98165aea82225a7390b67c2c7a9d09c8c41755ba1757cf13c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 182, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 205 + ], + "type": "text", + "content": "Table 1. Relative advantages of kaleidoscope vs. galvanometer design for capturing event fields." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 210, + 400, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 210, + 400, + 222 + ], + "spans": [ + { + "bbox": [ + 314, + 210, + 400, + 222 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 231, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 435 + ], + "type": "text", + "content": "Light field sensing: One of the first approaches to light field capture [54] involves using multiple sensors to gather angular information from various perspectives. However, this setup is bulky, costly, and requires high bandwidth communication. An alternative is time-sequential capture, which mounts a sensor on a mechanical gantry to capture the light field of a static scene [12, 25]. Rather than moving the sensor itself, a mirror can be used to redirect the view. Ihrke et al. [19] proposed using a movable planar mirror to shift the view temporally. However, this method requires individual calibration for each image, adding complexity. Fujii et al. [8] uses a galvanometer and pairs it with a high-speed camera. Inspired by their approach, we developed a similar design, precisely controlling the galvanometer's movement. A key distinction is that our galvanometer operates continuously, while their system selects discrete positions from a lookup table, making it slower." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 604 + ], + "type": "text", + "content": "In addition to time-sequential capture, angular information can also be measured with spatial multiplexing. One way to achieve this is through a microlens array [39], which has been widely adopted in commercial light field cameras like Raytrix [44] and Lytro [28]. While effective, microlens arrays are challenging to manufacture and install. Taguchi et al. [48] suggested using a mirror ball to expand the camera's field of view, though this setup results in the camera being visible in the mirror ball's reflection. Alternatively, Manakov et al. [30] introduced a kaleidoscope-like mirror array that can be attached to a standard camera without modification. Inspired by this design, we adapted it to convert a conventional event camera into a device capable of capturing the event field." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Event camera designs: From the first neuromorphic sensor by Mahowald et al. [29] to the recent Sony IMX636, event sensing saw exponential growth. We briefly review designs most relevant to our manuscript and refer the readers to recent surveys [6, 10]. Event cameras are regularly used in robotics and autonomous navigation for optical flow computation [2, 11, 52], feature tracking [49], and motion segmentation [34, 57], but event cameras cannot detect edges in the direction of the camera motion. Inspired by microsac" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26911" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "cades in human eye movement, He et al. [17] placed a rotating wedge mask before the event camera aperture to capture edges in all directions. Similar to them, our galvanometer design detects edges oriented in all directions, independent of camera and object motion, as shown in Fig. 6. Our galvanometer design also finds inspiration from the designs of Mugalikar et al. [37], who used a rotating polarizer in the aperture plane to capture high-fidelity polarization information for shape-from-polarization and Yu et al. [55], who used a rotating light source and event camera for fast photometric stereo." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 295, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 359 + ], + "type": "text", + "content": "Habuchi et al. [16] captured light fields for static scenes by placing an LCOS device in the aperture plane of event cameras and using CNN to recover the light field. While the technique is quite novel, it is limited to static scenes due to the LCOS response time (22 ms). Guo et al. [14] placed a microlens array (MLA) on the event sensor to capture the light field for dynamic scenes and showed microscopic imaging applications with their prototype. We built our prototypes for macroscopic applications, and our kaleidoscope design is functionally same as MLA design. However, the galvanometer design can capture angular information even for static scenes without trading off the spatial resolution like Guo et al. [14]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 527 + ], + "type": "text", + "content": "Event processing algorithms: Event cameras capture scene information at higher temporal resolution (in " + }, + { + "bbox": [ + 55, + 360, + 295, + 527 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 360, + 295, + 527 + ], + "type": "text", + "content": "s) and dynamic range (140 dB) compared to standard cameras (60 dB) [10]. These features enable solutions to challenges where standard frame-based sensors struggle, such as high-speed motion estimation [9], high frame rate reconstruction [45], and high dynamic range imaging [45]. Besides, event cameras and standard frame-based cameras are complementary; event cameras provide high-speed, sparse data, while standard cameras capture color and fine details. By fusing these two types of sensors, algorithms like Timelens [50] and Timelens++ [51] effectively interpolate RGB frames using event information. We adapt these event processing algorithms for processing event field data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 527, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 658 + ], + "type": "text", + "content": "Other event-based fusions: Apart from RGB cameras, other computational imaging systems can also be made fast by fusing with event sensors [58]. Event-based structured light sensing techniques [18, 23, 31, 32, 35, 47] reconstruct 3D scenes at high speed by replacing standard intensity sensors with event cameras. Event cameras are also used for foveated depth sensing [36] to scan moving objects at a higher spatial resolution. Muglikar et al. [38] combined single-photon avalanche diodes (SPADs) with event cameras to achieve high-speed, low-light, low-bandwidth imaging capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 669, + 185, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 185, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 185, + 682 + ], + "type": "text", + "content": "3. Technical Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We will briefly review the image formation models for light field and event cameras that are useful in developing the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 466, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 466, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 466, + 83 + ], + "type": "text", + "content": "mathematical models for event fields." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 91, + 389, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 389, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 389, + 104 + ], + "type": "text", + "content": "3.1. Light Field" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": "The plenoptic model of light " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "L(\\vec{x}, \\vec{\\omega}, \\lambda, t)" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": " is a comprehensive representation that captures all the dimensions of light emitted by a scene [1, 33]. Here, " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\vec{x} = (x, y, z)" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": " represents the spatial coordinates of the observer through which the light passes; " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\vec{\\omega} = (\\omega_x, \\omega_y)" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": " are the angular coordinates; " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": " is the wavelength, and " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": " is the time. At a given time, the measurement of a monochromatic light is 5D function. Levoy and Hanrahan [25] and Gortler et al. [12] showed that the 5D representation has redundancies and have reduced it to 4D by assuming that the light field is measured in free space. Under this assumption, as the radiance of a ray remains a constant in a straight line, the static monochromatic light field is a 4D function " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "L(\\vec{u}, \\vec{s})" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\vec{u} = (u, v)" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": " is a point on the virtual sensor and " + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\vec{s} = (s, t)" + }, + { + "bbox": [ + 313, + 109, + 553, + 289 + ], + "type": "text", + "content": " is a point on the virtual aperture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "text", + "content": "To synthesize an image from a given light field, we integrate the light field over the aperture as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 384, + 321, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 321, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 384, + 321, + 553, + 346 + ], + "type": "interline_equation", + "content": "I (x, y) = \\int_ {\\vec {s}} L (\\vec {u}, \\vec {s}) \\mathrm {d} \\vec {s}. \\tag {1}", + "image_path": "343b8c25ce31789618949f7b47a239b146dce1a0cafc9f2fb9125e8043b1ecf0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "content": "Light field allows us to change the focus plane after imaging by computationally changing the distance between the sensor and aperture planes [3, 25]. If " + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "inline_equation", + "content": "d_0" + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "content": " is the original distance between the planes and " + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "content": " is the distance required to focus at a given depth, the light field that focuses at " + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "content": " will be " + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "inline_equation", + "content": "L_{d}(\\vec{u},\\vec{s}) = L(\\vec{u} +(d_{0} / d)\\vec{s},\\vec{s})" + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "content": ". The refocused image can be synthesized by integrating angles, as shown in Eq. (1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 445, + 405, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 445, + 405, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 445, + 405, + 456 + ], + "type": "text", + "content": "3.2.Event Camera" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "text", + "content": "An event camera asynchronously records the brightness " + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "inline_equation", + "content": "(B\\dot{=} \\log (\\int_{\\vec{\\omega}}Ld\\vec{\\omega}))" + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "text", + "content": " change event for each pixel independently [4, 26, 29, 42]. Specifically, in an ideal and noise-free scenario, a pixel " + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x_k,y_k)^T" + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "text", + "content": " will be triggered at time " + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "text", + "content": " as soon as the brightness change reaches a predefined threshold " + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "text", + "content": " since the last triggered event. The polarity " + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "inline_equation", + "content": "p_k\\in \\{+1, - 1\\}" + }, + { + "bbox": [ + 313, + 463, + 553, + 559 + ], + "type": "text", + "content": ", which is the sign of the brightness change, will be recorded for each trigger. Mathematically," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 569, + 542, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 569, + 542, + 581 + ], + "spans": [ + { + "bbox": [ + 328, + 569, + 542, + 581 + ], + "type": "interline_equation", + "content": "\\Delta B \\left(\\mathbf {x} _ {k}, t _ {k}\\right) \\doteq B \\left(\\mathbf {x} _ {k}, t _ {k}\\right) - B \\left(\\mathbf {x} _ {k}, t _ {k} - \\Delta t _ {k}\\right) = p _ {k} C", + "image_path": "0f5d0c1af02eee90b1167789ab72d434089b2f222c177446bd1244c59eea9336.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 328, + 584, + 553, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 584, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 328, + 584, + 553, + 609 + ], + "type": "interline_equation", + "content": "\\frac {\\partial B \\left(\\mathbf {x} _ {k} , t _ {k}\\right)}{\\partial t} \\approx \\frac {p _ {k} C}{\\Delta t _ {k}} \\Rightarrow \\frac {\\partial L \\left(\\mathbf {x} _ {k} , t _ {k}\\right)}{\\partial t} \\approx L \\frac {p _ {k} C}{\\Delta t _ {k}}. \\tag {2}", + "image_path": "c2b971661d85cd44abaf8d62e0786948c0d9618d5949d29261a42b1677c31971.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "type": "text", + "content": "Therefore, event cameras capture a high-speed temporal derivative slice " + }, + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, t_k)" + }, + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "type": "text", + "content": " of the plenoptic function." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "In the related work, we referred to several techniques that excel at extracting high-speed spatial information from sparse event streams. These algorithms can do so as the scene information typically changes slowly [10], making the temporal gradient " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L(\\mathbf{x}_k,t_k)}{\\partial t}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " inherently sparse. This redundancy also applies to angular dimension as light field" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26912" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 111 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 111 + ], + "type": "text", + "content": "is mostly smooth [21, 56] and hence, " + }, + { + "bbox": [ + 55, + 70, + 295, + 111 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L(\\vec{x},\\vec{\\omega},\\vec{s},\\lambda,t)}{\\partial\\vec{\\omega}}" + }, + { + "bbox": [ + 55, + 70, + 295, + 111 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 295, + 111 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L(\\vec{x},\\vec{\\omega},\\vec{s},\\lambda,t)}{\\partial t}" + }, + { + "bbox": [ + 55, + 70, + 295, + 111 + ], + "type": "text", + "content": " are sparse. We exploit this sparsity to capture event fields with a single event camera in the next section." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 129, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 129, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 129, + 132 + ], + "type": "text", + "content": "4. Event Field" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 261 + ], + "type": "text", + "content": "As mentioned briefly in Sec. 3.2, an event camera's photodetector measures brightness by taking the logarithm of radiance integrated over all angles. This integration results in a loss of angular information. Our objective, through event fields is to retain this angular information " + }, + { + "bbox": [ + 55, + 140, + 296, + 261 + ], + "type": "inline_equation", + "content": "\\vec{\\omega}" + }, + { + "bbox": [ + 55, + 140, + 296, + 261 + ], + "type": "text", + "content": ". To achieve this goal, we need to multiplex angular information into either the spatial or temporal dimensions that the event camera captures. We advocate for and design both spatial and temporal multiplexing techniques, as they offer complementary advantages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 174, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 174, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 174, + 280 + ], + "type": "text", + "content": "4.1. Spatial Multiplexing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "type": "text", + "content": "We encode discrete angular information " + }, + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "type": "inline_equation", + "content": "(\\omega_{i})" + }, + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "type": "text", + "content": " into the spatial dimension using a mapping function defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 316, + 294, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 316, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 137, + 316, + 294, + 328 + ], + "type": "interline_equation", + "content": "\\left(\\omega_ {i}, \\mathbf {x} _ {\\mathbf {s}} \\left(\\omega_ {i}\\right)\\right) \\mapsto \\mathbf {x}, \\tag {3}", + "image_path": "d87fe495302ecb7a94599361375aedde592919c009a30c44bb305b3bd2c2434f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{s}}(\\omega_{i})" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": " represents the spatial coordinate for each measured view " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "(\\omega_{i})" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": ". Let the number of discrete angular views be " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": ", the spatial resolution of the sensor be " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": " is an integral multiple of " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": ". The specific mapping function " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{s}}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": " varies based on the hardware design. The two commonly used mapping functions are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 407, + 295, + 455 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 407, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 431 + ], + "type": "text", + "content": "- Multi-sensor array [54] or kaleidoscope [30]: In this design, the mapping function is " + }, + { + "bbox": [ + 55, + 407, + 295, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{s}} = \\mathbf{x} \\mod (\\mathbf{r} / \\mathbf{n})" + }, + { + "bbox": [ + 55, + 407, + 295, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 295, + 455 + ], + "type": "text", + "content": "- Microlens array [39]: In this design, the mapping function is " + }, + { + "bbox": [ + 56, + 432, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{s}} = \\mathbf{x} / / \\mathbf{n}" + }, + { + "bbox": [ + 56, + 432, + 295, + 455 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 456, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 478 + ], + "type": "text", + "content": "By applying this mapping function to Equation (2), the measurements are:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 482, + 294, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 482, + 294, + 516 + ], + "spans": [ + { + "bbox": [ + 127, + 482, + 294, + 516 + ], + "type": "interline_equation", + "content": "\\frac {\\partial B}{\\partial t} (\\underbrace {\\mathbf {x} _ {\\mathbf {s}} , \\omega} _ {\\mathbf {x} _ {k}}, t _ {k}) \\approx \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {4}", + "image_path": "237aca0cd6c34d6dcb141435dfbceadfda270b225439f156daa8b6929534a6ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": "and the event camera will now measure the temporal brightness changes across both spatial and angular dimensions. Note that the brightness derivatives are still related to time, and we lose spatial resolution (by a factor of " + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": ") to improve angular resolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "text", + "content": "4.2. Temporal Multiplexing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "To encode angular information in the temporal dimension, we steer the incoming rays along a closed periodic curve " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(t)" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": " with period " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": ". This steering could be accomplished by either moving the camera or placing steerable galvanometer in the optical path of the camera. The galvanometer can be steered fast along Lissajous curves [8, 41]; hence, we use it in all our experiments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The temporal multiplexing does not sacrifice spatial resolution but sacrifices temporal resolution and always detects" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 439, + 205 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 439, + 205 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 439, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 439, + 205 + ], + "type": "image", + "image_path": "86b90fa7ac7414ed93aaa2daaa20d00b87b480c3f9fbfac983165ae3c05527c4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 205, + 402, + 214 + ], + "lines": [ + { + "bbox": [ + 353, + 205, + 402, + 214 + ], + "spans": [ + { + "bbox": [ + 353, + 205, + 402, + 214 + ], + "type": "text", + "content": "(a) Plugin UI" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 219, + 552, + 231 + ], + "lines": [ + { + "bbox": [ + 313, + 219, + 552, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 552, + 231 + ], + "type": "text", + "content": "Figure 2. Blender plugin for physics-based event field rendering." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 441, + 70, + 554, + 134 + ], + "blocks": [ + { + "bbox": [ + 441, + 70, + 554, + 134 + ], + "lines": [ + { + "bbox": [ + 441, + 70, + 554, + 134 + ], + "spans": [ + { + "bbox": [ + 441, + 70, + 554, + 134 + ], + "type": "image", + "image_path": "bb8e23a5fd75e89c232c185f70276e8fa49648c9c9eedce900347b340876faa2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 133, + 534, + 143 + ], + "lines": [ + { + "bbox": [ + 460, + 133, + 534, + 143 + ], + "spans": [ + { + "bbox": [ + 460, + 133, + 534, + 143 + ], + "type": "text", + "content": "(b) Event simulation" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 441, + 143, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 441, + 143, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 441, + 143, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 441, + 143, + 553, + 205 + ], + "type": "image", + "image_path": "88d04abcbab4db534f7df759d63506c2a98a970444bb041d75c639e9fc631d20.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 205, + 523, + 213 + ], + "lines": [ + { + "bbox": [ + 470, + 205, + 523, + 213 + ], + "spans": [ + { + "bbox": [ + 470, + 205, + 523, + 213 + ], + "type": "text", + "content": "(c) Event field" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 237, + 553, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 553, + 284 + ], + "type": "text", + "content": "more events than spatial multiplexing. At any given time instant, the angular view will be at " + }, + { + "bbox": [ + 313, + 237, + 553, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(t \\bmod T)" + }, + { + "bbox": [ + 313, + 237, + 553, + 284 + ], + "type": "text", + "content": " and the temporally varying light field is " + }, + { + "bbox": [ + 313, + 237, + 553, + 284 + ], + "type": "inline_equation", + "content": "L(\\mathbf{x}, \\mathcal{C}(t \\bmod T), t)" + }, + { + "bbox": [ + 313, + 237, + 553, + 284 + ], + "type": "text", + "content": ". For static scenes, the event camera measurement is" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 332, + 289, + 553, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 289, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 332, + 289, + 553, + 318 + ], + "type": "interline_equation", + "content": "\\frac {\\partial B (\\mathbf {x} , \\mathcal {C} (t))}{\\partial t} = \\left. \\frac {\\partial B (\\mathbf {x} , \\omega)}{\\partial \\omega} \\frac {\\partial \\mathcal {C} (t)}{\\partial t} \\right| _ {t = t _ {k}} \\approx \\frac {p _ {k} C}{\\Delta t _ {k}}, \\tag {5}", + "image_path": "0bd92dc4adb9bc91f7de37a4b19318098f423f5daab83c19939448190b078107.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "type": "text", + "content": "Therefore, temporal multiplexing technique measures the angular derivatives of the light field " + }, + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L(\\mathbf{x},\\omega)}{\\partial\\omega}" + }, + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "type": "text", + "content": " along the scanning curve." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": "While we restricted the scanning curves to be Lissajous curves, specifically circles in most of the results, in principle, the scanning curve can be any periodic curve, including space-filling curves [15, 46]. Space-filling curves result in dense angular light-field capture at the cost of loosing temporal information. However, typically light field is highly redundant and smooth and a circular scan captures most of the light-field diversity as we show in the results." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 466, + 495, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 495, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 495, + 480 + ], + "type": "text", + "content": "5. Simulator and Hardware Design" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "type": "text", + "content": "To validate our approach, we built a physics-based simulator and two hardware prototypes, kaleidoscope for spatial multiplexing and galvanometer for temporal multiplexing. The simulator allows us to precisely control the scene and design parameters, enabling an accurate and fair comparison between the two multiplexing techniques. Hardware prototypes show the real-world feasibility of our designs as well as new capabilities." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 588, + 383, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 383, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 383, + 599 + ], + "type": "text", + "content": "5.1. Simulator" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": "We have used a straightforward approach to simulating events, which involves rendering the full plenoptic function, calculating brightness by propagating the plenoptic data through optics, and tracking the brightness difference from the previous event. A new event is triggered for each pixel when this difference exceeds a manually set threshold." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We chose Blender [7] as our backend as it is opensource, physically based, supports GPU acceleration (Cycles), Python APIs, and brandishes a user-friendly GUI. We" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26913" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 118 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 118 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 118 + ], + "type": "image", + "image_path": "6dd46700a738085abdfa4ab4767a2f944cc7455ace35316672506cea806efe91.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 115, + 202, + 125 + ], + "lines": [ + { + "bbox": [ + 149, + 115, + 202, + 125 + ], + "spans": [ + { + "bbox": [ + 149, + 115, + 202, + 125 + ], + "type": "text", + "content": "(a) Ray sketch" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 125, + 294, + 198 + ], + "blocks": [ + { + "bbox": [ + 56, + 125, + 294, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 294, + 198 + ], + "type": "image", + "image_path": "50b2831ae951dd4128427ed6f7cf7fe73ba816fd94926515f2b60beaaed1ec89.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 198, + 211, + 209 + ], + "lines": [ + { + "bbox": [ + 140, + 198, + 211, + 209 + ], + "spans": [ + { + "bbox": [ + 140, + 198, + 211, + 209 + ], + "type": "text", + "content": "(b) Hardware setup" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 285, + 164, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 295, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 278 + ], + "type": "text", + "content": "Figure 3. Kaleidoscope design. (a) Illustration shows rays from various angles being directed towards the event camera. (b) The main lens focuses the scene onto the kaleidoscope, splitting the view into multiple angles. The beam splitter directs these multiple views to the RGB camera for standard light-field imaging and the event camera for event-field imaging." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 285, + 164, + 369 + ], + "lines": [ + { + "bbox": [ + 56, + 285, + 164, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 164, + 369 + ], + "type": "image", + "image_path": "823beea596d251a87a5722b774b475a27a2a852abb00dec7f33fe54511801098.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 369, + 135, + 380 + ], + "lines": [ + { + "bbox": [ + 82, + 369, + 135, + 380 + ], + "spans": [ + { + "bbox": [ + 82, + 369, + 135, + 380 + ], + "type": "text", + "content": "(a) Ray sketch" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 385, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 452 + ], + "type": "text", + "content": "Figure 4. Galvanometer design. (a) Ray sketch illustrating the angular separation achieved by controlling mirror angles in the galvanometer setup. (b) Hardware setup showing the galvanometer mirror system integrated with the event camera. This design allows for capturing angular derivative events by dynamically controlling the mirror position." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 166, + 285, + 295, + 369 + ], + "blocks": [ + { + "bbox": [ + 166, + 285, + 295, + 369 + ], + "lines": [ + { + "bbox": [ + 166, + 285, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 166, + 285, + 295, + 369 + ], + "type": "image", + "image_path": "859904120910af13932cb06d70c1d9731da781e99f5723cf2e65fe8375a7de07.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 370, + 266, + 380 + ], + "lines": [ + { + "bbox": [ + 195, + 370, + 266, + 380 + ], + "spans": [ + { + "bbox": [ + 195, + 370, + 266, + 380 + ], + "type": "text", + "content": "(b) Hardware setup" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 457, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 506 + ], + "type": "text", + "content": "developed a Blender plugin to render event fields for either spatial or temporal multiplexing design, along with a simple GUI to facilitate community use. The GUI and sample simulation results are presented in Fig. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 514, + 255, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 255, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 255, + 527 + ], + "type": "text", + "content": "5.2. Kaleidoscope for Spatial Multiplexing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "type": "text", + "content": "A kaleidoscope is a common optical instrument composed of multiple first-surface mirrors. Manakov et al. [30] introduced this optical device as an add-on for sensor fusion. As shown in Fig. 3, this kaleidoscope is a rectangular design made of four mirrors. The lens at the front forms an image at the entrance of the kaleidoscope. After multiple reflections within the kaleidoscope, the rear camera captures the image, resulting in several flipped images, each with a different virtual camera location. Therefore, the light field information is spatially multiplexed onto the camera sensor." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 659, + 273, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 273, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 273, + 673 + ], + "type": "text", + "content": "5.3. Galvanometer for Temporal Multiplexing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "While the kaleidoscope design offers a simple solution without moving parts, it has two key limitations. First, although it provides angularly separated events, the events" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "still represent the derivative of light intensity over time. Second, the spatial resolution of current event cameras is not very high (1 megapixel), and this resolution would further decrease with spatial multiplexing. Temporal multiplexing offers an alternate approach." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "type": "text", + "content": "For temporal multiplexing, we place a galvanometer, as shown in Fig. 4, in the optical path of the event camera. A galvanometer consists of two mirrors that can be precisely adjusted at high speed (Eg. GVS012 from Thorlabs can scan a thousand Lissajous curves per second). Moving mirrors allow us to rotate the camera virtually at a high speed and capture the event field. In this design, the captured information represents the derivative of light intensity with respect to angle, rather than time, as we discussed in Sec. 4.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 251, + 367, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 367, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 367, + 262 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "text", + "content": "We evaluate the proposed techniques on various scenes. We first summarize the observations and then provide details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 294, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 554, + 402 + ], + "type": "text", + "content": "Simulations demonstrate that each design is optimal in different contexts. The kaleidoscope (spatial multiplexing) excels in capturing dynamic scenes with high frame rates but suffers from lower spatial resolution and is ineffective for static scenes, as no events are generated without motion. The galvanometer (temporal multiplexing) captures both static and dynamic scenes at full spatial resolution by recording angular derivatives but introduces motion blur if the object motion is significant." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": "Real-world applications demonstrate each design's post-capture refocusing capabilities. Specifically, with the galvanometer setup, we show high-resolution refocusing of static and dynamic parts of the scene. With the kaleidoscope setup, we perform sensor fusion with an RGB camera, providing vibrant, high-speed refocusing for dynamic scenes. These results underscore the complementary nature of the two designs: the kaleidoscope for high-speed dynamic capture and the galvanometer for detailed, high spatial-angular light field recovery." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "type": "text", + "content": "In the following sections, we compare our proposed hardware designs (see Section 6.1), followed by applications for each setup. Specifically, we explore HDR light fields (Section 6.2) and slow-motion refocusing (Section 6.3) using the kaleidoscope configuration, as well as high-speed post-capture refocusing (Section 6.4) and instant true depth estimation (Section 6.5) using the galvanometer-based design." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 624, + 524, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 524, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 524, + 637 + ], + "type": "text", + "content": "6.1. Kaleidoscope vs. Galvanometer Designs" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "In practical applications, differences in optical properties—such as focal lengths—make it challenging to compare two cameras under identical conditions. The simulator, however, allows us to control these variables, enabling a fair comparison of both designs. We, therefore, used Blender to create a scene containing swirling water and moving objects" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26914" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 140, + 129 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 140, + 129 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 140, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 140, + 129 + ], + "type": "image", + "image_path": "1ed24bd2bc9e9ba0b8472bc30d936e258785ab47d19c7e16cd02810038c5327a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 129, + 128, + 138 + ], + "lines": [ + { + "bbox": [ + 69, + 129, + 128, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 129, + 128, + 138 + ], + "type": "text", + "content": "(a) Ground truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "type": "text", + "content": "Figure 5. (Simulation) comparison of spatial (kaleidoscope) and temporal (galvanometer) multiplexing techniques. (a) A simulated scene with swirling water and objects, with a stationary box. (b) Refocused image using the kaleidoscope design, showing reduced spatial resolution. However, the image quality is independent of scene motion. The static wooden box is not reconstructed. (c) Refocused images using the galvanometer design at different object movement speeds " + }, + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "type": "inline_equation", + "content": "(1\\times,2\\times,4\\times," + }, + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "type": "inline_equation", + "content": "8\\times)" + }, + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "type": "text", + "content": " while maintaining a constant scanning speed. " + }, + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 55, + 143, + 555, + 210 + ], + "type": "text", + "content": " corresponds to about 2.5 pixels per scan for the highlighted object. As the scene motion increases, the highlighted object, unlike in the kaleidoscopic design, suffers from motion blur. The static wooden boxes are perfectly reconstructed, unlike the kaleidoscopic design." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 140, + 70, + 222, + 129 + ], + "blocks": [ + { + "bbox": [ + 140, + 70, + 222, + 129 + ], + "lines": [ + { + "bbox": [ + 140, + 70, + 222, + 129 + ], + "spans": [ + { + "bbox": [ + 140, + 70, + 222, + 129 + ], + "type": "image", + "image_path": "1eba0882cab0721c7c39d884d1fc89e229b5464d4f02d647785fc6ac5e5a422b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 129, + 212, + 139 + ], + "lines": [ + { + "bbox": [ + 149, + 129, + 212, + 139 + ], + "spans": [ + { + "bbox": [ + 149, + 129, + 212, + 139 + ], + "type": "text", + "content": "(b) Kaleidoscope" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 71, + 305, + 129 + ], + "blocks": [ + { + "bbox": [ + 223, + 71, + 305, + 129 + ], + "lines": [ + { + "bbox": [ + 223, + 71, + 305, + 129 + ], + "spans": [ + { + "bbox": [ + 223, + 71, + 305, + 129 + ], + "type": "image", + "image_path": "44c57d8c6e6a88c733751cc1355de01903d79921b708ce674e157e346f92d1e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 119, + 270, + 128 + ], + "lines": [ + { + "bbox": [ + 256, + 119, + 270, + 128 + ], + "spans": [ + { + "bbox": [ + 256, + 119, + 270, + 128 + ], + "type": "inline_equation", + "content": "\\therefore m : x = 1" + }, + { + "bbox": [ + 256, + 119, + 270, + 128 + ], + "type": "text", + "content": " 或 " + }, + { + "bbox": [ + 256, + 119, + 270, + 128 + ], + "type": "inline_equation", + "content": "{3x} + {4y} + 1 = 0" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 71, + 388, + 129 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 388, + 129 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 388, + 129 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 388, + 129 + ], + "type": "image", + "image_path": "6e60ca239c1ab5b1c113110a9c66e3d958e604534e9f4672f232fa131a47a101.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 129, + 421, + 138 + ], + "lines": [ + { + "bbox": [ + 358, + 129, + 421, + 138 + ], + "spans": [ + { + "bbox": [ + 358, + 129, + 421, + 138 + ], + "type": "text", + "content": "(c) Galvanometer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 71, + 470, + 129 + ], + "blocks": [ + { + "bbox": [ + 389, + 71, + 470, + 129 + ], + "lines": [ + { + "bbox": [ + 389, + 71, + 470, + 129 + ], + "spans": [ + { + "bbox": [ + 389, + 71, + 470, + 129 + ], + "type": "image", + "image_path": "3770a4352c4c5efabc26a0bdec7c88bfb47ffc79df026630d39a2dcb311acd61.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 471, + 71, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 471, + 71, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 471, + 71, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 471, + 71, + 553, + 129 + ], + "type": "image", + "image_path": "bfd7ab5e273ff5f747de5775503a5198ef781a020dee2b30d7179449d109fdf1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 224, + 134, + 266 + ], + "blocks": [ + { + "bbox": [ + 57, + 224, + 134, + 266 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 134, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 134, + 266 + ], + "type": "image", + "image_path": "8a513dfb20b468368cda8b74a09f220ee439f8c57ae423b09618ff9a00a5ab44.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 266, + 115, + 274 + ], + "lines": [ + { + "bbox": [ + 80, + 266, + 115, + 274 + ], + "spans": [ + { + "bbox": [ + 80, + 266, + 115, + 274 + ], + "type": "text", + "content": "(a) Scene" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 137, + 224, + 217, + 266 + ], + "blocks": [ + { + "bbox": [ + 137, + 224, + 217, + 266 + ], + "lines": [ + { + "bbox": [ + 137, + 224, + 217, + 266 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 217, + 266 + ], + "type": "image", + "image_path": "1ac3b43a8af3f4f557cdea92ee0553db7d2a3c5ebba450872cd4ea3e95941350.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 266, + 208, + 275 + ], + "lines": [ + { + "bbox": [ + 146, + 266, + 208, + 275 + ], + "spans": [ + { + "bbox": [ + 146, + 266, + 208, + 275 + ], + "type": "text", + "content": "(b) Kaleidoscope" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 217, + 224, + 294, + 266 + ], + "blocks": [ + { + "bbox": [ + 217, + 224, + 294, + 266 + ], + "lines": [ + { + "bbox": [ + 217, + 224, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 217, + 224, + 294, + 266 + ], + "type": "image", + "image_path": "7956fe40f31be6cb50f7413fdd6ba6571a81ccd3720bcc99fdd2a6087e30abb9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 266, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 223, + 266, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 223, + 266, + 288, + 275 + ], + "type": "text", + "content": "(c) Galvanometer" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 57, + 357, + 135, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 280, + 296, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 346 + ], + "type": "text", + "content": "Figure 6. (Experimental) comparison of multiplexing techniques. (a) Reference scene with a grid pattern. (b) When the grid moves horizontally, the kaleidoscope design cannot capture horizontal lines as intensity does not change along this axis. (c) Using a galvanometer that scans in a circular motion ensures complete capture, avoiding information loss in the angular domain." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 357, + 135, + 407 + ], + "lines": [ + { + "bbox": [ + 57, + 357, + 135, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 357, + 135, + 407 + ], + "type": "image", + "image_path": "7781054d13be25ba622b52a449f446168ccfc0becd61ab29dfe57aeaa8d33390.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 407, + 130, + 416 + ], + "lines": [ + { + "bbox": [ + 63, + 407, + 130, + 416 + ], + "spans": [ + { + "bbox": [ + 63, + 407, + 130, + 416 + ], + "type": "text", + "content": "(a) Short exposure" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 420, + 296, + 486 + ], + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 486 + ], + "type": "text", + "content": "Figure 7. High dynamic range (HDR) light field. A high dynamic range scene is captured with a standard RGB camera using (a) short exposure and (b) long exposure, and both fail to capture the full dynamic range. In contrast, (c) the HDR light field reconstructed from the event field effectively captures the complete dynamic range of the scene." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 137, + 358, + 216, + 406 + ], + "blocks": [ + { + "bbox": [ + 137, + 358, + 216, + 406 + ], + "lines": [ + { + "bbox": [ + 137, + 358, + 216, + 406 + ], + "spans": [ + { + "bbox": [ + 137, + 358, + 216, + 406 + ], + "type": "image", + "image_path": "66a298453a73c0e08f52cf32653edb2a5901148849ba2270d1ebafabbb3ae7f0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 407, + 209, + 416 + ], + "lines": [ + { + "bbox": [ + 142, + 407, + 209, + 416 + ], + "spans": [ + { + "bbox": [ + 142, + 407, + 209, + 416 + ], + "type": "text", + "content": "(b) Long exposure" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 216, + 358, + 294, + 406 + ], + "blocks": [ + { + "bbox": [ + 216, + 358, + 294, + 406 + ], + "lines": [ + { + "bbox": [ + 216, + 358, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 216, + 358, + 294, + 406 + ], + "type": "image", + "image_path": "acdb913614ed441d2d25a34dce2e30afea51b3994540c610afa71b56004d7f82.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 407, + 290, + 416 + ], + "lines": [ + { + "bbox": [ + 220, + 407, + 290, + 416 + ], + "spans": [ + { + "bbox": [ + 220, + 407, + 290, + 416 + ], + "type": "text", + "content": "(c) HDR light field" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 492, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 517 + ], + "type": "text", + "content": "within a stationary box (see Fig. 5) to assess the strengths and limitations of each design." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 519, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 626 + ], + "type": "text", + "content": "The first significant difference is that the galvanometer design can fully reconstruct scene information by capturing angular derivatives, whereas the kaleidoscope design can only capture moving objects, as no events are generated in static regions. The stationary wooden box is clearly visible in the galvanometer design but is absent in the kaleidoscope design. We also observe this in a real experiment (Fig. 6), where a rectangular mesh is translated horizontally, and the kaleidoscope fails to reconstruct horizontal lines." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "A second key difference is that the galvanometer design multiplexes the event field into the time domain, making it susceptible to motion blur from rapidly moving objects. This is visible in the zoomed insets in Fig. 5. The kaleidoscope design, by contrast, avoids motion blur but instead suffers from lower spatial resolution. However, both designs can capturing and refocusing dynamic objects." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 225, + 421, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 225, + 421, + 238 + ], + "spans": [ + { + "bbox": [ + 314, + 225, + 421, + 238 + ], + "type": "text", + "content": "6.2. HDR Light Fields" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 243, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 352 + ], + "type": "text", + "content": "Event fields are not only high-speed but also high-dynamic range light fields. We demonstrate this in Fig. 7, where the scene containing bright LED light in the foreground and diffuse texture in the background is captured using both an RGB camera and the event camera through a kaleidoscope. The event fields have a large dynamic range and capture the spatial structure of both LEDs and the background texture, whereas the RGB camera either captures the light sources (short exposure) or the background (long exposure)." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 361, + 502, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 502, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 502, + 373 + ], + "type": "text", + "content": "6.3. SloMoRF: Slow Motion Refocusing" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 380, + 554, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 451 + ], + "type": "text", + "content": "We augment the kaleidoscope design with a co-located RGB camera to capture and reconstruct slow-motion RGB frames with dynamic refocusing capability, called SloMoRF. The hardware setup, shown in Fig. 3, features a beam splitter for identical kaleidoscope views for RGB and event camera without interfering with the main lens." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "type": "text", + "content": "We utilize the TimeLens algorithm [50] for sensor fusion1. Using the captured event data, we interpolate frames to generate a " + }, + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "type": "text", + "content": " higher frame-rate RGB video (from 120 to 720 fps). The interpolated results are presented in Fig. 8." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 501, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 693 + ], + "type": "text", + "content": "This approach allows us to generate a high-speed color light field. We treat this light field as unstructured as we found it empirically easier than calibrating the views. To align the images, we estimate the necessary shift at a specified patch using a basic pattern-matching algorithm. Note that this shift computation is only needed once for the entire video sequence as long as the target focal plane remains fixed. Figure 8 shows refocusing results across three focal planes. Additionally, the captured event field allows for focus tracking on a high-speed object moving along the depth. We demonstrate this capability in Fig. 9, where we visually compare sequence of images that focus on a fast-moving toy dart vs. focusing solely on the background. These results demonstrate the high frame rate, color, and refocusing capabilities of our kaleidoscope design when combined with the state-of-the-art sensor fusion algorithms." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 533, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 533, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 533, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 325, + 702, + 533, + 713 + ], + "type": "text", + "content": "code for the updated TimeLens++ [51] is not publicly available" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26915" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 154, + 206 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 154, + 206 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 154, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 154, + 206 + ], + "type": "image", + "image_path": "fb33bd42cc41b131abcf7fe0a42c4e0250630c68d0f8d618162e5e321341e4c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 146, + 118, + 155 + ], + "lines": [ + { + "bbox": [ + 82, + 146, + 118, + 155 + ], + "spans": [ + { + "bbox": [ + 82, + 146, + 118, + 155 + ], + "type": "text", + "content": "(a) Scene" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 207, + 139, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 207, + 139, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 139, + 228 + ], + "type": "text", + "content": "(b) Captured frame (with aligned events)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 233, + 205 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 233, + 205 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 233, + 205 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 233, + 205 + ], + "type": "image", + "image_path": "90e65e349a4af66a24d58c54e194ed748da79af488be8ad830f35e192abe0c6e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 206, + 211, + 215 + ], + "lines": [ + { + "bbox": [ + 176, + 206, + 211, + 215 + ], + "spans": [ + { + "bbox": [ + 176, + 206, + 211, + 215 + ], + "type": "text", + "content": "Frame 19" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 70, + 313, + 205 + ], + "blocks": [ + { + "bbox": [ + 235, + 70, + 313, + 205 + ], + "lines": [ + { + "bbox": [ + 235, + 70, + 313, + 205 + ], + "spans": [ + { + "bbox": [ + 235, + 70, + 313, + 205 + ], + "type": "image", + "image_path": "904f98013420f21178a1aa8c47a5f09623e32975064fa4a7e8f58d75430122b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 206, + 301, + 216 + ], + "lines": [ + { + "bbox": [ + 246, + 206, + 301, + 216 + ], + "spans": [ + { + "bbox": [ + 246, + 206, + 301, + 216 + ], + "type": "text", + "content": "1st interpolated" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 217, + 419, + 228 + ], + "lines": [ + { + "bbox": [ + 286, + 217, + 419, + 228 + ], + "spans": [ + { + "bbox": [ + 286, + 217, + 419, + 228 + ], + "type": "text", + "content": "(c) Refocus with interpolated frames" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 70, + 392, + 205 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 392, + 205 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 392, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 392, + 205 + ], + "type": "image", + "image_path": "f6d1cf29d2fdd0ac59fd9c741866429c4a3e6bb6f1ec36389ee18ec622697905.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 206, + 382, + 216 + ], + "lines": [ + { + "bbox": [ + 323, + 206, + 382, + 216 + ], + "spans": [ + { + "bbox": [ + 323, + 206, + 382, + 216 + ], + "type": "text", + "content": "3rd interpolated" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 393, + 70, + 473, + 205 + ], + "blocks": [ + { + "bbox": [ + 393, + 70, + 473, + 205 + ], + "lines": [ + { + "bbox": [ + 393, + 70, + 473, + 205 + ], + "spans": [ + { + "bbox": [ + 393, + 70, + 473, + 205 + ], + "type": "image", + "image_path": "07067c492c2cf50501d9a50c1f71446f626788a57ebd02562df0cb53cc6b8319.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 206, + 462, + 216 + ], + "lines": [ + { + "bbox": [ + 404, + 206, + 462, + 216 + ], + "spans": [ + { + "bbox": [ + 404, + 206, + 462, + 216 + ], + "type": "text", + "content": "5th interpolated" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 474, + 70, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 474, + 70, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 474, + 70, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 474, + 70, + 553, + 205 + ], + "type": "image", + "image_path": "842a969a96ff0334f4d83ea8bf90032e8b336bae661158ca2ba4926bbe4caee7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 206, + 531, + 215 + ], + "lines": [ + { + "bbox": [ + 496, + 206, + 531, + 215 + ], + "spans": [ + { + "bbox": [ + 496, + 206, + 531, + 215 + ], + "type": "text", + "content": "Frame 20" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 56, + 300, + 136, + 347 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 288 + ], + "type": "text", + "content": "Figure 8. SlowMoRF: Slow motion refocusing with kaleidoscope and sensor fusion. (a) Original setup with Lego pieces floating in a water tank as a wooden brick is thrown in, with CVPR in the background, (b) Captured raw frames with aligned events using kaleidoscope design, and (c) Interpolated refocused frames at three focal planes (back, mid, and front). Between the frames, the fall of the wooden block and changes in water splash and caustics can be observed in the front refocus. Mid and back refocus show sharp Lego toys and CVPR. This setup achieves high-speed (720 fps) refocusing with enhanced clarity across focal depths. The supplementary contains the full video." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 300, + 136, + 347 + ], + "lines": [ + { + "bbox": [ + 56, + 300, + 136, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 136, + 347 + ], + "type": "image", + "image_path": "59ecbe89c21831a2b98979d831edf6e25c28d706977e95d26c536f85a7e1f1f0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 347, + 220, + 355 + ], + "lines": [ + { + "bbox": [ + 130, + 347, + 220, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 220, + 355 + ], + "type": "text", + "content": "(a) Refocus on the bullet" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 136, + 301, + 216, + 347 + ], + "blocks": [ + { + "bbox": [ + 136, + 301, + 216, + 347 + ], + "lines": [ + { + "bbox": [ + 136, + 301, + 216, + 347 + ], + "spans": [ + { + "bbox": [ + 136, + 301, + 216, + 347 + ], + "type": "image", + "image_path": "c6ff010469d5b771780b691cd597a4560909f4bc096566cbd57aa2c7ab54ad2e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 216, + 301, + 294, + 347 + ], + "blocks": [ + { + "bbox": [ + 216, + 301, + 294, + 347 + ], + "lines": [ + { + "bbox": [ + 216, + 301, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 216, + 301, + 294, + 347 + ], + "type": "image", + "image_path": "1b2dc2885cb5471ff7398ad3cc5275d9610971667e2b4890909c8b9860d0178f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 57, + 356, + 136, + 402 + ], + "blocks": [ + { + "bbox": [ + 57, + 356, + 136, + 402 + ], + "lines": [ + { + "bbox": [ + 57, + 356, + 136, + 402 + ], + "spans": [ + { + "bbox": [ + 57, + 356, + 136, + 402 + ], + "type": "image", + "image_path": "bc487c107a240e654d62835e933e6046e524f389be0f8dcca9679ae859d03fc2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 402, + 232, + 411 + ], + "lines": [ + { + "bbox": [ + 119, + 402, + 232, + 411 + ], + "spans": [ + { + "bbox": [ + 119, + 402, + 232, + 411 + ], + "type": "text", + "content": "(b) Refocus on the background" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 415, + 296, + 515 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 515 + ], + "type": "text", + "content": "Figure 9. High-speed refocusing on a toy dart. The original video was captured at 120 fps, interpolating seven frames between each pair of original frames, resulting in a 960 fps high-speed light field video. The images shown here are all the middle (the fourth) interpolated frames, demonstrating refocusing on two planes: (a) the moving dart and (b) the background. Current high-speed videography cannot adjust focus during capture, so dynamic objects must remain at a fixed focal plane to stay sharp. Our technique eliminates this limitation and can focus on dynamic object." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 136, + 356, + 216, + 401 + ], + "blocks": [ + { + "bbox": [ + 136, + 356, + 216, + 401 + ], + "lines": [ + { + "bbox": [ + 136, + 356, + 216, + 401 + ], + "spans": [ + { + "bbox": [ + 136, + 356, + 216, + 401 + ], + "type": "image", + "image_path": "8966ffc4b44ca2314a2467ffc09c39a5472a47d2cfe03aef8e38f5330c6e0c36.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 216, + 356, + 294, + 401 + ], + "blocks": [ + { + "bbox": [ + 216, + 356, + 294, + 401 + ], + "lines": [ + { + "bbox": [ + 216, + 356, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 216, + 356, + 294, + 401 + ], + "type": "image", + "image_path": "390a2d3a2cca213bde6f1e2acd506402a20598d235bd8b3388ba2a40a8879bcb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 530, + 250, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 250, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 250, + 544 + ], + "type": "text", + "content": "6.4. High-Speed Post-capture Refocusing" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 552, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 647 + ], + "type": "text", + "content": "In this section, we test the post-capture refocusing ability of the galvanometer setup on dynamic scenes. Compared to unstructured event field captured by kaleidoscope design, our galvanometer design produces a fully structured event field after calibration, as we precisely control the galvanometer with a signal generator to determine the angle relative to time. The supplementary materials provide calibration details." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Using E2VID, we generate a high-speed video at 10,000 fps (0.1 ms per frame) from the captured event field data. The galvanometer is operating at " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "250\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": ". Therefore, we reconstruct the light field at 250 fps and 40 views, the fastest recorded so far at a megapixel resolution." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 315, + 301, + 434, + 384 + ], + "blocks": [ + { + "bbox": [ + 315, + 301, + 434, + 384 + ], + "lines": [ + { + "bbox": [ + 315, + 301, + 434, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 301, + 434, + 384 + ], + "type": "image", + "image_path": "6047bd58e4da46426b4ce53b7f965386231a186521235a2536d87b799113d92e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 384, + 498, + 393 + ], + "lines": [ + { + "bbox": [ + 371, + 384, + 498, + 393 + ], + "spans": [ + { + "bbox": [ + 371, + 384, + 498, + 393 + ], + "type": "text", + "content": "(a) Grayscale reconstructed images" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 435, + 301, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 435, + 301, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 435, + 301, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 435, + 301, + 553, + 384 + ], + "type": "image", + "image_path": "034afb9cd748ca08e532e0b704a59c1231d0732977c4e1b978fb8ae076219389.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 315, + 393, + 434, + 483 + ], + "blocks": [ + { + "bbox": [ + 315, + 393, + 434, + 483 + ], + "lines": [ + { + "bbox": [ + 315, + 393, + 434, + 483 + ], + "spans": [ + { + "bbox": [ + 315, + 393, + 434, + 483 + ], + "type": "image", + "image_path": "ac7386f86882bdfb9e07bc6a3268771f7e6ff69fbc8002f4ec9e4c932468f41e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 483, + 518, + 493 + ], + "lines": [ + { + "bbox": [ + 350, + 483, + 518, + 493 + ], + "spans": [ + { + "bbox": [ + 350, + 483, + 518, + 493 + ], + "type": "text", + "content": "(b) Scene captured by ordinary camera (60Hz)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "type": "text", + "content": "Figure 10. Dynamic scene refocusing. Using our galvanometer design at " + }, + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "type": "inline_equation", + "content": "250\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "type": "text", + "content": ", we capture a scene with a rotating fan marked with \"CVPR\", spinning at " + }, + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "type": "inline_equation", + "content": "300\\mathrm{rpm}" + }, + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "type": "text", + "content": ". (a) Using prior knowledge of the scanned mirror's location, we refocus on the background or the rotating fan in the foreground. (b) An ordinary camera captures the scene at " + }, + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "type": "inline_equation", + "content": "60\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 497, + 555, + 586 + ], + "type": "text", + "content": ", focusing on the background and foreground from different viewpoints. The camera lacks refocus ability and causes motion blur on the fan due to the slower shutter speed." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 434, + 393, + 553, + 483 + ], + "blocks": [ + { + "bbox": [ + 434, + 393, + 553, + 483 + ], + "lines": [ + { + "bbox": [ + 434, + 393, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 434, + 393, + 553, + 483 + ], + "type": "image", + "image_path": "e2e7124817fc22ac5b71f79dab14ab5fdbbce15d8013bd4a416633e2b060b70a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 591, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 555, + 712 + ], + "type": "text", + "content": "Using this light field and equations in Sec. 3.1, we can do post-capture refocusing. Note that even though E2VID generates noisy frames at high speed (due to low light), the noise gets averaged out when we synthesize a refocused image. In Fig. 10, we show refocusing results on the background texture (left figure) or foreground fan (right figure) and can observe that they are noise-free. We further stress-tested our hardware setup in Fig. 11 and noticed that the event camera begins to lose events due to the readout bandwidth limit when the galvanometer operates beyond " + }, + { + "bbox": [ + 313, + 591, + 555, + 712 + ], + "type": "inline_equation", + "content": "250\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 591, + 555, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26916" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 69, + 219 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 69, + 219 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 69, + 219 + ], + "type": "text", + "content": "Event frame Front refocus Back refocus" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 70, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 553, + 232 + ], + "type": "image", + "image_path": "41072534c48c6633677789e53b10f0ab81a6898c82d177d1db7c895730a9fda7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 293 + ], + "type": "text", + "content": "Figure 11. Effect of galvo scan frequency for refocusing accuracy. The scene consists of two static objects positioned in the foreground and background, along with a fan (highlighted in yellow boxes) in the foreground rotating at " + }, + { + "bbox": [ + 55, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "480\\mathrm{RPM}" + }, + { + "bbox": [ + 55, + 236, + 555, + 293 + ], + "type": "text", + "content": ". The static objects (highlighted in red and blue boxes) remain sharp up to a scanning frequency of " + }, + { + "bbox": [ + 55, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "200\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 236, + 555, + 293 + ], + "type": "text", + "content": " but become increasingly blurry as the scan speed increases. This counter intuitive phenomenon occurs as event camera reaches its maximum bandwidth capacity, resulting in random event loss, as seen in the Event frame row. However, the text on the rotating fan becomes clearer at higher scan speeds due to reduced motion blur." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 305, + 293, + 401 + ], + "blocks": [ + { + "bbox": [ + 56, + 305, + 293, + 401 + ], + "lines": [ + { + "bbox": [ + 56, + 305, + 293, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 293, + 401 + ], + "type": "image", + "image_path": "e00f44c110d3db29a87ccb121aa6c6404f3aaaf5a8a00b360c0985c9f108905e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 405, + 295, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 449 + ], + "type": "text", + "content": "Figure 12. Instant true depth estimation at " + }, + { + "bbox": [ + 55, + 405, + 295, + 449 + ], + "type": "inline_equation", + "content": "100\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 405, + 295, + 449 + ], + "type": "text", + "content": ". Depth mapping of a running person, shown in three frames with color-coded depth scale (in inches), highlighting the system's capability for rapid, accurate depth measurement in dynamic scenes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 464, + 222, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 222, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 222, + 478 + ], + "type": "text", + "content": "6.5. Instant True Depth Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 490, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 634 + ], + "type": "text", + "content": "In addition to refocusing, our galvanometer design enables real-time true depth estimation. Unlike SLAM-based methods [5], our approach provides instant true depth estimation based on the event field. As outlined in Sec. 6.4, our galvanometer design generates a structured light field, requiring only a single calibration step to determine the depth-to-disparity ratio. We employ the \"depth from focus\" algorithm [13] to calculate the depth of each pixel. This involves first constructing a focal stack by refocusing the light field at various depths. For each pixel, we determine the focal stack image where the pixel appears sharpest, and the depth associated with that image is the depth of the pixel." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "By adopting this approach, we obtain the instant true depth at the same frequency as the galvanometer's scanning speed. As shown in Fig. 12, a " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "100\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " galvo scan successfully captures a person running away from the camera, offering sparse yet real-time depth information of the letters on their shirt at a rate of 100 fps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 307, + 476, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 476, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 476, + 319 + ], + "type": "text", + "content": "7. Discussion and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 332, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 440 + ], + "type": "text", + "content": "In this manuscript, we introduce \"event field\", a new approach for capturing high-speed, high resolution, HDR light fields using two optical configurations—kaleidoscope and galvanometer—combined with an event camera. Our kaleidoscope design enables capturing temporal derivatives for multiple views, while the galvanometer design captures angular derivatives along curves. We validate both designs through simulations and real-world data, demonstrating their unique advantages." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 516 + ], + "type": "text", + "content": "Our experiments showcase several potential applications. Most notably, our setup enables high-speed refocusing on fast-moving objects, real-time true depth estimation, and high-dynamic-range light field capture. Supplementary PDF and video provides further visualizations of these results across additional scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "While our approach shows several new possibilities, it also creates new research problems in the graphics, imaging, and vision fields. Our current brute-force simulation of the event field could benefit from optimal importance sampling techniques that improve the rendering speed by orders of magnitude similar to previous importance techniques for computational cameras [20, 22, 27, 40]. Additionally, as shown in Fig. 11, static parts of the scene do not require high bandwidth. By implementing adaptive scanning (e.g., slower scans for static parts, faster scans for dynamic areas) with a MEMS array, we can enhance bandwidth efficiency. We currently use pre-trained E2VID and Timelens models, which are optimized for time-sequential event streams. Generating sufficient event field data, possibly through an optimal simulator and training models specifically for our event fields, would yield a significant performance boost." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26917" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 209 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 91, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 115 + ], + "type": "text", + "content": "This project is funded by NSF awards 2403122, 2326904, 2107454, and 2107313 and ARPA-H award D24AC00296." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 116, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 116, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 116, + 295, + 163 + ], + "type": "text", + "content": "We thank Laurence Rubin at Boston Craft Works, Joseph Poissant at Dartmouth Engineering Machine Shop, and Hannah Machesky at Two Way Mirrors for providing mirrors and polishing expertise for our kaleidoscope design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 209 + ], + "type": "text", + "content": "We also extend our gratitude to the RISC lab members, Atul Agarwal, Dhawal Sirikonda, Kedari Chowtoori, and Sarah Friday, for their generous assistance in capturing the data and their 3D printing and laser cutting skills." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 56, + 231, + 115, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 115, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 115, + 243 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 251, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 251, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 293 + ], + "type": "text", + "content": "[1] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Vision and Modeling Group, Media Laboratory, Massachusetts Institute of ..., 1991. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 297, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 296, + 339 + ], + "type": "text", + "content": "[2] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "text", + "content": "[3] Tom E Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE transactions on pattern analysis and machine intelligence, 34(5):972-986, 2011. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "text", + "content": "[4] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "text", + "content": "[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "text", + "content": "[6] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-iliidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. arXiv preprint arXiv:2408.13627, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 534, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 555 + ], + "type": "text", + "content": "[7] Blender Online Community. Blender - a 3d modelling and rendering package, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 557, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 295, + 590 + ], + "type": "text", + "content": "[8] Toshiaki Fujii and Masayuki Tanimoto. A real-time ray-space acquisition system. In Stereoscopic Displays and Virtual Reality Systems XI, pages 179-187. SPIE, 2004. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 591, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 645 + ], + "type": "text", + "content": "[9] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. IEEE transactions on pattern analysis and machine intelligence, 40(10):2402-2412, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "text", + "content": "[12] Steven J. Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F. Cohen. The lumigraph. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 43-54, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 195 + ], + "type": "text", + "content": "[13] Paul Grossmann. Depth from focus. Pattern recognition letters, 5(1):63-69, 1987. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 251 + ], + "type": "text", + "content": "[14] Ruipeng Guo, Qianwan Yang, Andrew S Chang, Guorong Hu, Joseph Greene, Christopher V Gabel, Sixian You, and Lei Tian. Eventlfm: Event camera integrated fourier light field microscopy for ultrafast 3d imaging. Light: Science & Applications, 13(1):144, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 253, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 296 + ], + "type": "text", + "content": "[15] Mohit Gupta, Andreas Velten, Shree K Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM Transactions on Graphics (TOG), 37(2):1-18, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 352 + ], + "type": "text", + "content": "[16] Shuji Habuchi, Keita Takahashi, Chihiro Tsutake, Toshiaki Fujii, and Hajime Nagahara. Time-efficient light-field acquisition using coded aperture and events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24923-24933, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 397 + ], + "type": "text", + "content": "[17] Botao He, Ze Wang, Yuan Zhou, Jingxi Chen, Chahat Deep Singh, Haojia Li, Yuman Gao, Shaojie Shen, Kaiwei Wang, Yanjun Cao, et al. Microsaccade-inspired event camera for robotics. Science Robotics, 9(90):eadj8124, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "text", + "content": "[18] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. High-speed structured light based 3d scanning using an event camera. Optics Express, 29(22):35864-35876, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[19] Ivo Ihrke, Timo Stich, Heiko Gottschlich, Marcus Magnor, and Hans-Peter Seidel. Fast incident light field acquisition and rendering. Journal of WSCG, 2008. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 467, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 554, + 510 + ], + "type": "text", + "content": "[20] Adrian Jarabo, Julio Marco, Adolfo Munoz, Raul Buisan, Wojciech Jarosz, and Diego Gutierrez. A framework for transient rendering. ACM Transactions on Graphics (ToG), 33 (6):1-10, 2014. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "text", + "content": "[21] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 557, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 590 + ], + "type": "text", + "content": "[22] Juhyeon Kim, Wojciech Jarosz, Ioannis Gkioulekas, and Adithya Pediredla. Doppler time-of-flight rendering. ACM Trans. Graph., 42(6), 2023. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 591, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 554, + 633 + ], + "type": "text", + "content": "[23] T Leroux, S-H Ieng, and Ryad Benosman. Event-based structured light for depth reconstruction using frequency tagged light patterns. arXiv preprint arXiv:1811.10771, 2018.3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 635, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 635, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 315, + 635, + 553, + 657 + ], + "type": "text", + "content": "[24] M. Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 31-42, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26918" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[27] Yang Liu, Shaojie Jiao, and Wojciech Jarosz. Temporally sliced photon primitives for time-of-flight rendering. In Computer Graphics Forum, pages 29-40. Wiley Online Library, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 162, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 162, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 162, + 294, + 183 + ], + "type": "text", + "content": "[28] Lytro. Lytro. https://www.lytro.com/, 2017. [Online; accessed 6-Oct-2024]. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 184, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 184, + 294, + 216 + ], + "type": "text", + "content": "[29] Misha Mahowald and Misha Mahowald. The silicon retina. An Analog VLSI System for Stereoscopic Vision, pages 4-65, 1994. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[30] Alkhazur Manakov, John Restrepo, Oliver Klehm, Ramon Hegedus, Elmar Eisemann, Hans-Peter Seidel, and Ivo Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM Transactions on Graphics, 32(4):47-1, 2013. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 314 + ], + "type": "text", + "content": "[31] Ashish Rao Mangalore, Chandra Sekhar Seelamantula, and Chetan Singh Thakur. Neuromorphic fringe projection profilometry. IEEE Signal Processing Letters, 27:1510-1514, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[32] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "type": "text", + "content": "[33] Leonard McMillan and Gary Bishop. Plenoptic modeling: An image-based rendering system. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 433-440. Association for Computing Machinery, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "text", + "content": "[34] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14414-14423, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 459, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 502 + ], + "type": "text", + "content": "[35] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 2021 International Conference on 3D Vision (3DV), pages 1165-1174. IEEE, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "text", + "content": "[36] Manasi Muglikar, Diederik Paul Moeys, and Davide Scaramuzza. Event guided depth sensing. In 2021 International Conference on 3D Vision (3DV), pages 385-393. IEEE, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[37] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "type": "text", + "content": "[38] Manasi Muglikar, Siddharth Somasundaram, Akshat Dave, Edoardo Charbon, Ramesh Raskar, and Davide Scaramuzza. Event cameras meet spads for high-speed, low-bandwidth imaging. arXiv preprint arXiv:2404.11511, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[39] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "text", + "content": "[40] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "type": "text", + "content": "[41] Luboš Poláček, Jana Jurmanová, and Zdeněk Navrátil. Laser galvo mirrors: perfect instrument for the demonstration of lissajous figures. Physics Education, 54(5):055002, 2019. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "text", + "content": "[42] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 194, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 217 + ], + "type": "text", + "content": "[43] Ziyuan Qu. Event field. https://dartmouth-risc-lab.github.io/event-field-website/, 2024.2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 217, + 526, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 526, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 526, + 228 + ], + "type": "text", + "content": "[44] A Raytrix. 3d light field camera technology, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[45] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "text", + "content": "[46] Hans Sagan. *Space-filling curves*. Springer Science & Business Media, 2012. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "text", + "content": "[47] Dhawal Sirikonda, Praneeth Chakravarthula, Ioannis Gkioulekas, and Adithya Pediredla. Structured light with a million light planes per second. arXiv preprint arXiv:2411.18597, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "text", + "content": "[48] Yuichi Taguchi, Amit Agrawal, Srikumar Ramalingam, and Ashok Veeraraghavan. Axial light field for curved mirrors: Reflect your perspective, widen your view. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 499-506. IEEE, 2010. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 393, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 459 + ], + "type": "text", + "content": "[49] David Tedaldi, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Feature detection and tracking with the dynamic and active-pixel vision sensor (davis). In 2016 Second International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7. IEEE, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 459, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 525 + ], + "type": "text", + "content": "[50] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 525, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 592 + ], + "type": "text", + "content": "[51] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[52] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 678 + ], + "type": "text", + "content": "[53] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[54] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2, 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26919" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 274 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[55] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9602-9611, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "text", + "content": "[56] Mingyuan Zhao, Hao Sheng, Da Yang, Sizhe Wang, Ruixuan Cong, Zhenglong Cui, Rongshan Chen, Tun Wang, Shuai Wang, Yang Huang, et al. A survey for light field superresolution. High-Confidence Computing, page 100206, 2024. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "text", + "content": "[57] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE transactions on neural networks and learning systems, 34(8):4868-4880, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 274 + ], + "type": "text", + "content": "[58] Zihao Zou, Ziyuan Qu, Xi Peng, Vivek Boominathan, Adithya Pediredla, and Praneeth Chakravarthula. High-speed dynamic 3d imaging with sensor fusion splatting. arXiv preprint arXiv:2502.04630, 2025. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26920" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_content_list.json b/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..33b54afa159d8d0b76bc6d79096cdb41b6742af4 --- /dev/null +++ b/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_content_list.json @@ -0,0 +1,1526 @@ +[ + { + "type": "text", + "text": "Event-Equalized Dense Video Captioning", + "text_level": 1, + "bbox": [ + 289, + 130, + 709, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kangyi Wu* Pengna Li* Jingwen Fu Yizhe Li Yang Wu Yuhan Liu Jinjun Wang Sanping Zhou†", + "bbox": [ + 194, + 179, + 803, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University", + "bbox": [ + 205, + 217, + 831, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wukangyi747600, sauerfisch, ful371252069, lyz0515, wuyang_CC, liuyuhan200095}@stu.xjtu.edu.cn jinjun@mail.xjtu.edu.cn spzhou@xjtu.edu.cn", + "bbox": [ + 102, + 272, + 933, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 339, + 326, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense video captioning aims to localize and caption all events in arbitrary untrimmed videos. Although previous methods have achieved appealing results, they still face the issue of temporal bias, i.e., models tend to focus more on events with certain temporal characteristics. Specifically, 1) the temporal distribution of events in training datasets is uneven. Models trained on these datasets will pay less attention to out-of-distribution events. 2) long-duration events have more frame features than short ones and will attract more attention. To address this, we argue that events, with varying temporal characteristics, should be treated equally when it comes to dense video captioning. Intuitively, different events tend to have distinct visual differences due to varied camera views, backgrounds, or subjects. Inspired by that, we intend to utilize visual features to have an approximate perception of possible events and pay equal attention to them. In this paper, we introduce a simple but effective framework, called Event-Equalized Dense Video Captioning $(E^2 DVC)$ to overcome the temporal bias and treat all possible events equally. Experimental results on ActivityNet Captions and YouCook2 dataset validate the effectiveness of the proposed methods and show State-of-the-art (SOTA) performance on dense video captioning.", + "bbox": [ + 86, + 371, + 483, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 748, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, an increasing number of works have begun to focus on video understanding[21, 22, 28, 38, 44, 50, 57] where video captioning is a challenging branch. Conventional video captioning [5, 12, 33, 34, 42] aims to generate a natural sentence to describe the main event of a short video clip. However, they face considerable challenges when ap", + "bbox": [ + 89, + 773, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d226cc13c95c008e8dde0a80b1c84703f77acb24348d907dd3e8b31e680ab12a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 337, + 733, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e807cdabd3fa7777c0601b09753afbfaf8da729408131892783681e7848a4c4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 339, + 885, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04872f2e16adead48cda3838b466cfe7eeafd9e70f46198e5079566bfc629151.jpg", + "image_caption": [ + "(a) Data Event Distribution", + "(b) Event Duration Ratio" + ], + "image_footnote": [], + "bbox": [ + 517, + 463, + 735, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f3f9668bd550a93885f1d6feef72e72f797560689dd1a46ebc0b446340ade0c.jpg", + "image_caption": [ + "Figure 1. Temporal distribution of ground truth and predicted events on ActivityNet Captions training set. The prediction is from PDVC [48]. (a) The distribution of all events in time dimensions. Horizontal and vertical axes represent the normalized center time position and duration respectively. It's obvious that events are distributed unevenly in these two temporal dimensions. (b) The proportion of the number of events in different duration intervals. It's clear that the predicted short-duration events are much fewer than the ground truth." + ], + "image_footnote": [], + "bbox": [ + 736, + 463, + 885, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plied to realistic videos which are usually long, untrimmed, and contain a variety of events. To address the challenge, dense video captioning (DVC) [19] is proposed to automatically localize and caption all events in an untrimmed video.", + "bbox": [ + 511, + 776, + 903, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing DVC methods [6, 17, 18, 30, 60] follow a common pipeline to first extract image features from down-sampled video frames with pre-trained image encoder [15, 35, 39, 46] and then delicately design a framework to ex", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally.", + "bbox": [ + 119, + 875, + 313, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 122, + 888, + 250, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8417", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ploit event representations from them. Although state-of-the-art (SOTA) performance has been achieved, existing methods still face the issue of temporal bias, which means that models will pay more attention to events with certain temporal characteristics while overlooking some other ones. The temporal bias can be divided into two categories based on its causes: 1) Data distribution bias. As shown in Figure 1(a), the predicted events are scattered in the time axes. Events are completely ignored or mis-localized for most timestamps. This is because the models tend to learn from the dataset which part of the video to focus on and the ground truth events are unevenly distributed in this dimension. 2) Event duration bias. This bias results from the duration variation between events. Since short-duration events have fewer frames, their features are easy to be overwhelmed by others. Figure 1(b) shows the proportion of events in different duration intervals. It's obvious that the predicted short-duration events are much fewer than the ground truth. This means that most short-duration events are totally ignored or localized with large deviations.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the issues above, we argue that every event, regardless of its temporal characteristics, should share equal importance when it comes to dense video captioning. To achieve this goal, we need to reduce the influence of temporal information and ensure that attention is evenly distributed across all events. However, the challenge lies in how to perceive possible events in a video without relying on temporal information. Intuitively, different events have large visual differences considering their varied camera views, backgrounds, or subjects [19]. Inspired by this, we intend to utilize the differences in pure visual features to gain an initial perception of segments where events may exist. Equal Attention will be allocated to all these segments.", + "bbox": [ + 89, + 396, + 482, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel Event-Equalized Dense Video Captioning framework $(\\mathrm{E}^2\\mathrm{DVC})$ to overcome the temporal bias and treat each potential event equally. First, a non-trainable Event Perception module (EPM) is devised to utilize agglomerative hierarchical clustering algorithm [7] on visual features to generate video frame clusters. These clusters are segments with potential events. A refinement mechanism is designed to split frames that are discontinuous in time and discard isolated frames. These refined clusters are regarded as pseudo-events. Since the clustering process is non-trainable, the result will not be affected by the uneven data distribution. Additionally, since agglomerative hierarchical clustering is a non-uniform method, frames from short-duration events can also be equally clustered into the same category as the long ones.", + "bbox": [ + 89, + 595, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the pseudo-events, equal attention is required among them. To achieve this, a Pseudo-event Initialization module (PEI) is introduced. Concretely, we calculate the temporal locations of these pseudo-events (center timestamps) and use a temporal encoder to encode them", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as pseudo-event queries. These queries serve as the initial guess of the event's locations to help the model allocate its attention. A novel Event-Enhanced Encoder (EEE) is also proposed to explore frame-frame and frame-event relationships during the encoding process.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our $\\mathrm{E}^2\\mathrm{DVC}$ can determine and localize events with higher precision which further improves the dense video captioning performance. The main contributions of our work can be summarized as:", + "bbox": [ + 511, + 167, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the issue of temporal bias and propose an Event-Equalized Dense Video Captioning framework to treat all events equally.", + "- We propose a non-trainable Event Perception module, a Pseudo-Event Initialization module, and an Event-Enhanced Encoder to ensure all events are noticed. Our components could perceive the input video with pure visual features and allocate equal attention to events with large temporal variations.", + "- Comprehensive experiments have been conducted on the ActivityNet Captions [19] and the YouCook2 [58] datasets to verify the effectiveness of our model in the dense video captioning task." + ], + "bbox": [ + 511, + 227, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 436, + 661, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense Video Captioning. Dense video captioning consists of two subtasks: event localization [9, 11, 25, 26] and event captioning [24, 31, 42]. Based on different pipelines, the methods can be divided into two categories: two-stage methods and end-to-end methods. For two-stage methods, they solve dense video captioning by a \"localize-then-describe\" scheme. These methods first predict a set of event proposals with corresponding time boundaries and then generate a detailed sentence description for each proposal by further extracting semantic knowledge and visual context. Krishna et al. [19] propose the first two-stage dense video captioning model. Wang et al. [43] and Yang et al. [54] focus on mining useful contextual information to improve DVC performance. Iashin et al. [15, 16] construct architectures to utilize any number of modalities for event description. However, these methods lack inter-task interactions and heavily rely on hand-crafted components.", + "bbox": [ + 511, + 462, + 906, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the limitations above, recent methods follow the end-to-end pipeline [3, 6, 18, 23, 30, 36, 53, 60]. Li et al. [23] propose a joint and global optimization of detection and captioning. Islam et al. [17] and Deng et al. [6] solve DVC in hierarchical ways where the former follows a bottom-up structure while the latter proposes a top-down framework. Mun et al. [30] and Zhou et al. [60] leverage visual and linguistic context from prior events for coherent storytelling while the latter proposes an additional memory module and streaming decoding algorithm to support streaming video input. Kim et al. [18] utilize external memory to incorporate prior video-to-text knowledge to gener-", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8418", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a0a7923d0bc4a3e36cfd2e2d29283ea3bce9cebf9491e90e55cfcb7222c66924.jpg", + "image_caption": [ + "Figure 2. The overview of our $\\mathbf{E}^2\\mathbf{DVC}$ . Our model follows an encoder-decoder structure. Given an input video, a pre-trained visual encoder is adopted to extract frame-level features. (a) Event Perception Module (EPM). Based on visual features similarity, these frames are clustered to multiple uneven segments and a refinement mechanism is utilized to discard outliers. The remaining segments are assigned with pseudo-events. (b) Event-Enhanced Encoder (EEE). By incorporate the pseudo-event labels with visual frame features, it could enhance the encoder's ability to explore frame-event and event-event relationships, preventing short-duration events from being overwhelmed. (c) Pseudo-Event Initialization module (PEI). Given the pseudo events, we calculate and encode their temporal locations and use the embedding to initialize the queries, helping the model allocate its attention equally among them. Finally, we obtain the predictions of the event-number, time localization and event captions through heads." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 908, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ate natural descriptions. Yang et al. [53] reformulate sentence boundaries of transcribed speech as pseudo-events to achieve large-scale video data and propose a multi-modal pertaining model trained on this dataset. However, it remains a significant challenge to detect all events and localize them with high precision in a video. We propose a simple yet effective approach to enable the model to treat all events equally thus improving the DVC performance.", + "bbox": [ + 88, + 497, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query Formulation in Transformers. Since Detection Transformer (DETR)[2] has achieved promising performance in object detection without the need for any handcrafted components, many researchers have studied and further modified the model. Among these, research on decoder query formulation has received significant attention. Conditional DETR [29] learns a conditional spatial query by encoding the reference points as the query position embedding to help focus on distinct regions. Anchor DETR [49] and DAB-DETR [27] utilize anchor boxes to initialize queries to improve interpretability and force each query to only pay attention to a specific region in the object detection task. DN-DETR [20] additionally adds ground-truth bounding boxes with noises into decoder queries to reduce bipartite graph matching difficulty and achieve faster convergence. Group DETR [4] utilizes multiple groups of queries to conduct one-to-many assignments in order to speed up training and improve performance. Decoupled DETR [55] discovers that", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using shared queries for classification and localization will lead to suboptimal results. Instead, they introduce a task-aware query generation module to decouple these two tasks. DualDETR [62] presents a joint query initialization strategy to align queries from the instance level and boundary level to obtain better performance in the Temporal Action Detection task. In this paper, we introduce the query formulation strategies into the Dense Video Captioning task. Our $\\mathrm{E}^2\\mathrm{DVC}$ employs a novel Pseudo-Event Initialization module to initialize the decoder queries with pseudo-events to ensure that every possible events are treated equally.", + "bbox": [ + 511, + 497, + 906, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 678, + 604, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to treat each event equally and localize and caption all events without omission. To realize this, we introduce an Event Perception module (Section 3.1) with a delicately devised refinement mechanism to perceive the input video with only visual features available and select those segments where events may exist. Furthermore, A novel Event-Enhanced Encoder (Section 3.2) is also introduced to incorporate the pseudo-event labels into the self-attention framework in order to explore frame-frame and frame-event relationships. Meanwhile, we design an Pseudo Event Initialization module (Section 3.3) to ensure that the model pays equal attention to those pseudo-event locations. The overall framework of $\\mathrm{E}^2\\mathrm{DVC}$ is illustrated in Figure 2,", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8419", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6f7daf472a0e4d00fdf068eb72859d42e6572b90ec973e10ecb6d103643070a.jpg", + "image_caption": [ + "Figure 3. Illustration of the refinement mechanism. Different shapes denote different events. In the same events, frames with different color are temporal consistent. Our refinement mechanism could split the temporal inconsistent events into multiple new consistent events and discard the extremely short events which is short than the defined temporal event threshold." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 486, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which follows a standard parallel encoder-decoder structure with subtask heads. The objective is to generate $\\mathbf{N}$ pairs of events and captions $\\{(t_s^n, t_e^n, Cap^n)\\}_{n=1}^N$ , where $\\mathbf{N}$ denotes the number of events detected in the video. $t_s^n$ and $t_e^n$ denote the start and end timestamps of the $n$ -th event. $Cap^n$ denotes the corresponding caption. The details of task heads and training loss will be introduced in Section 3.4 and 3.5.", + "bbox": [ + 89, + 292, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Event Perception Module", + "text_level": 1, + "bbox": [ + 89, + 406, + 323, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To perceive the whole video with only visual features and pick out those pseudo-events, we devise an event perception module with agglomerative hierarchical clustering. The input video is first sampled in a certain ratio to obtain the image frames $V = \\{v_{i}\\}_{i=1}^{F}$ where $F$ denotes the number of frames. Then, a pre-trained image encoder $E$ [15, 35, 39, 46] is utilized to extract visual features $X = \\{x_{i}\\}_{i=1}^{F}$ from each video frame. Based on the visual feature similarities, we perform a hierarchical clustering algorithm to obtain pseudo-events.", + "bbox": [ + 89, + 429, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we choose the Agglomerative clustering algorithm because it doesn't assume any particular cluster shape and sizes. It could progressively merge the individual frames to generate uneven clusters with different sizes and assign them with pseudo-event labels. Specifically, each visual feature $x_{i}$ is initially considered an individual cluster with one element. We calculate the pairwise distance matrix among all the clusters. The two clusters that are the most close in feature space are combined into a bigger cluster. Then we recalculate the distance matrix. This procedure is iterated until the final number of clusters is equal to $N_{c}$ which is a hyper-parameter which is set to control the granularity of the clustering process. Finally, we obtain a number of clusters $\\mathbb{C}$ which can be formulated as", + "bbox": [ + 89, + 580, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {C} = \\left\\{\\mathbb {C} _ {i} \\right\\} _ {i = 1} ^ {N _ {c}} = \\left\\{\\left\\{x _ {i} ^ {j} \\right\\} _ {j = 1} ^ {N _ {c} ^ {i}} \\right\\} _ {i = 1} ^ {N _ {c}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 795, + 483, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $\\mathbb{C}_i$ denotes the $i$ -th cluster and $N_c^i$ denotes the number of this cluster.", + "bbox": [ + 89, + 824, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, since the timestamps of the frame features are not considered during the clustering process, some clustered frames are temporally inconsistent with others in the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "same cluster. Considering that frames belonging to the same event should be temporally consistent, as illustrated in Figure 3, we utilize a refinement mechanism to split the temporally discontinuous events into multiple new consistent events based on the timestamps of the video frames. Moreover, since an event typically occupies a certain duration, we devise a temporal event threshold $\\tau$ . If the durations of the events are shorter than $\\tau$ , these events would be considered outliers and discarded. The remains are regarded as final pseudo-events. Our framework will pay special attention to these pseudo-event locations with improved localization and captioning performance.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Event-Enhanced Encoder", + "text_level": 1, + "bbox": [ + 511, + 282, + 748, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the pseudo events, we directly assign the video frames with their pseudo labels. In this section, we introduce how to incorporate the assigned labels with visual features into transformer encoder to get better representations. We hope that in the process of encoding frame features, the focus should not only be on the frame-frame relationships but also on the frame-event one. In this way, frame features from short-duration events will not be overwhelmed during the self-attention process.", + "bbox": [ + 511, + 306, + 905, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To realize this, we devise an Event-Enhanced Encoder (EEE) to add pseudo-event labels to the frame feature encoding process. Specifically, for each video frame $v_{i}$ , we have its pseudo-event label $l_{i}$ . The pseudo-event label $l_{i}$ is transformed into a label embedding $le_{i}$ with a label encoder, which consists of a learnable dictionary and a multilayer perceptron. During the encoding process, we concatenate the frame-level features $X = \\{x_{i}\\}_{i=1}^{F}$ with their corresponding label embedding and positional embedding (PE). They are fed into the deformable transformer encoder and a multi-scale convolution layers to extract multi-scale features with frame-frame and frame-event relationships.", + "bbox": [ + 511, + 441, + 906, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pseudo-Event Initialization", + "text_level": 1, + "bbox": [ + 511, + 633, + 759, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the pseudo-events from Event Perception module, we want the model to pay equal attention to their temporal locations. Previously, the decoder aimed to query the event-level features from all frame features directly with $N_{q}$ randomly initialized and learnable event queries. In this way, each event query is associated with multiple random temporal locations and learns which part to focus on from the uneven training dataset, leading to those events with out-of-distribution temporal characteristics easy to be overlooked.", + "bbox": [ + 511, + 657, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we devise a novel Pseudo-Event Initialization (PEI) module to initialize those event queries with pseudo-events. Specifically, given a pseudo-event $E_{i} = \\{x_{i,j}\\}_{j = 1}^{N_{i}}$ in which $L_{i}$ is the number of frames for $N_{i}$ , we calculate its center temporal location $t_i$ (reference points) with:", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} = \\left(t _ {i, 1} + t _ {i, L _ {i}}\\right) / 2, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8420", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $t_{i,j}$ is the timestamp for frame feature $x_{i,j}$ . In this way, for each pseudo-event $E_{i}$ , we have its reference points $t_{i}$ . We calculate the inverse sigmoid $(Inv_{Sig})$ of $t_{i}$ and use temporal encoder $f_{t}$ [40] to get the temporal embedding. Instead of randomly initializing the event queries, we initialize them with the positional embeddings of all pseudo-events. The process can be formulated as:", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i} = f _ {t} \\left(\\operatorname {I n v} _ {\\operatorname {S i g}} \\left(t _ {i}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 204, + 480, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in which $q_{i}$ is the $i$ -th event query. What's more, considering that the pseudo-events may not cover all possible events in the ground truth, we also concatenate $N_{r}$ randomly initialized queries with these pseudo-event queries together. In this way, the event queries explicitly prioritize the locations of these pseudo-events. This significantly enhances the performance of event localization and further improves the event captioning performance as well.", + "bbox": [ + 89, + 226, + 483, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Task Heads", + "text_level": 1, + "bbox": [ + 89, + 354, + 217, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work follows an encoder-decoder framework with a parallel decoding structure [18, 48, 52]. Given the event features, three parallel prediction heads are utilized to do dense video captioning.", + "bbox": [ + 89, + 378, + 482, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Localization Head. The localization head is responsible for box prediction and binary classification for each event query $q_{i}$ . It outputs a set of tuples $\\{t_s^i, t_e^i, c^i\\}_{i=1}^{N_q}$ , where $t_s^i$ and $t_e^i$ denote the start and end timestamps and $c^i$ represents the localization confidence.", + "bbox": [ + 89, + 439, + 483, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Captioning Head. We utilize LSTM [14] as the backbone of our captioning head. At each timestamp $t$ , we feed context features $z_{i,t}$ , event query features $q_i$ and previous words $\\{w_{i,j}\\}_{j = 1}^{t - 1}$ into the LSTM model to get the current word $w_{i,t}$ . As the sentence is generated, the captioning head produces the complete sentence $S_{i} = w_{i,1},\\dots,w_{l,T}$ , where $T$ denotes the length of the sentence.", + "bbox": [ + 89, + 515, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Event Counter. It aims to detect the number of events in the given video. It first compresses the most salient information of event queries $Q$ to a global feature vector with a max-pooling layer and predicts a fixed-size vector $f_{count}$ with an FC layer with softmax activation. In $f_{count}$ , each value refers to the possibility of a specific number. During the inference stage, the event number $N_{event}$ is obtained by $N_{event} = \\arg\\max(f_{count})$ . The final outputs are obtained by selecting $N_{event}$ events with the highest confidence score $c$ which is calculated by:", + "bbox": [ + 89, + 621, + 483, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} = c _ {i} ^ {l o c} + \\frac {\\mu}{M _ {i} ^ {\\gamma}} \\sum_ {t = 1} ^ {M _ {i}} \\log \\left(c _ {i, t} ^ {c a p}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 777, + 482, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in which $M_{i}$ is the number of words for the $i$ -th event caption. $C_{i}^{loc}$ is the localization confidence for the $i$ -th event. $c_{i,t}^{cap}$ is the confidence score for the $t$ -th word for the $i$ -th event. $\\mu$ and $\\gamma$ are two hyper-parameters to control the influence of the captioning confidence.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function", + "text_level": 1, + "bbox": [ + 513, + 90, + 658, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Suppose $N_{q}$ events are predicted with their localizations and captions, Hungarian algorithm [2] is utilized to find the best bipartite matching results between the predicted events and the ground truth, during which the generalized IOU [37] between the predicted temporal segments and ground-truth segments is calculated as $L_{giou}$ and the focal loss between the predicted classification score and the ground-truth label is calculated as $L_{cls}$ . The matched event pairs are selected to calculate the set prediction loss, which is the weighted sum of gIOU loss, classification loss, countering loss, and caption loss. The loss function is formulated as:", + "bbox": [ + 511, + 114, + 906, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\alpha_ {g i o u} L _ {g i o u} + \\alpha_ {c l s} L _ {c l s} + \\alpha_ {e c} L _ {e c} + \\alpha_ {c a p} L _ {c a p}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 294, + 906, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in which $L_{cap}$ is the cross-entropy loss between the predicted word probability and the ground truth normalized by the caption length $M$ . $L_{ec}$ is the cross-entropy loss between the predicted count distribution and the ground truth. More details can be found in Supplementary.", + "bbox": [ + 511, + 321, + 905, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 412, + 645, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 438, + 702, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on the two commonly used dense video captioning benchmark datasets: ActivityNet Captions [19] and YouCook2 [58]. ActivityNet Captions includes over 20000 long videos covering a wide range of human daily activities and actions. Each video spans about 120s and is annotated with 3.7 temporally localized sentences on average. YouCook2 consists of 2000 long untrimmed videos from 89 cooking recipes. Generally, each video lasts approximately 320s and has 7.7 annotated segments with the corresponding sentences. For both datasets, we adhere to the standard split for training, validation, and testing. It's worth noting that we only use those videos that are still online on YouTube, leading to roughly $7\\%$ less videos compared to the original datasets.", + "bbox": [ + 511, + 460, + 906, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For both datasets, we sample video frames at a rate of 1 FPS (one frame per second) and subsample or pad the sequence of frames to reach a total of F frames, where F is set to 100 for ActivityNet Captions and 200 for YouCook2. We use CLIP ViT-L/14 [8, 35] to extract clip visual features for video frames [18]. To further investigate the generalization ability of our model to different frame-level features, we also evaluate our model based on C3D [39] features provided by PDVC [48] and TSN [46] features provided by [59]. We utilize a two-layer deformable transformer that incorporates multi-scale deformable attention across four levels. We adopt the Agglomerative Hierarchical cluster algorithm for generating pseudo events. For ActivityNet Captions, we set the number of event queries to 10, while for YouCook2, we set it", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8421", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/88c2ac469e3a568b7d1a119c301a6202b9e0180ad5f04db8f5bb610b1a392ba7.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art methods for dense video captioning in ActivityNet Captions. B4/M/C is short for BLEU4/Meteoror/CIDEr. * indicates the results of our implementation using the official code. The best performance results are highlighted for each feature type. We also present results from those pretrained models." + ], + "table_footnote": [], + "table_body": "
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
UEDVC [56]C3D✓---5.50
Vid2Seq [53]CLIP✓-8.5030.105.80
OmniVID [45]-✓1.737.5426.005.60
ECHR [47]C3Dx1.827.2014.73.20
E2ESG [61]C3Dx-3.5025.00-
PDVC [48]C3Dx1.657.5025.875.26
E²DVCC3Dx1.797.5426.835.32
MT [59]TSNx0.303.2014.70-
PDVC [48]TSNx1.787.9628.965.44
E²DVCTSNx2.038.0229.915.77
CM² [18]CLIPx2.388.5533.016.18
PDVC* [48]CLIPx2.218.0629.975.92
E²DVCCLIPx2.438.5733.636.13
", + "bbox": [ + 94, + 183, + 480, + 450 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe82df02dd7a3bfae34e1bd597484b85c331258a2436be7ac718600771800aef.jpg", + "table_caption": [ + "Table 2. Comparison with the state-of-the-art methods for dense video captioning on Youcook2. Since C3D features for YouCook2 are not publicly available, we fail to provide the results." + ], + "table_footnote": [], + "table_body": "
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
Vid2Seq [53]CLIP✓-9.3047.107.90
MT [59]TSN×1.155.009.30-
PDVC [48]TSN×0.804.7522.714.42
E²DVCTSN×1.054.7623.154.31
CM² [18]CLIP×1.636.0831.665.34
PDVC* [48]CLIP×1.405.5629.694.92
E²DVCCLIP×1.686.1134.265.39
", + "bbox": [ + 94, + 522, + 480, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to 100. For the event counter, we select a maximum count of 10 for the ActivityNet Captions dataset and 20 for the YouCook2 dataset. We set $N_{c}$ to 5 and $\\tau$ to 4. Other model hyperparameters follow the recipe of PDVC [48].", + "bbox": [ + 89, + 700, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate our model in two aspects: 1) To assess dense captioning performance, we adopt the official evaluation tool from ActivityNet Challenge [13], which use the metrics BLEU4 [32], METEOR [1], and CIDEr [41] to compute the average precision of matched pairs between generated captions and the ground truth. These scores are averaged over IOU at $\\{0.3, 0.5, 0.7, 0.9\\}$ . BLEU4 focuses on n-gram precision, METEOR incorporates semantic matching, and CIDEr emphasizes consen", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/82dec359702b46080d0de375be3d8a92021cadc740febea70cf733b1f8621fc8.jpg", + "table_caption": [ + "Table 3. Comparison to state-of-the-art methods for event Localization. We report the results on ActivityNet Captions and Youcook2 datasets. PT denotes Pretrain. Pre./Rec. is short for Precision/Recall." + ], + "table_footnote": [], + "table_body": "
MethodFeatures PTActivityNetYouCook2
Pre.↑Rec.↑F1↑Pre.↑Rec.↑F1↑
Vid2Seq [53]CLIP✓53.9052.7053.2927.8027.9027.84
E2ESG [61]C3D✘---20.6020.7020.65
MFT [51]TSN✘51.4124.3133.01---
CM² [18]CLIP✘56.8153.7155.2133.3824.7628.43
PDVC* [48]CLIP✘56.3853.2754.7832.3722.8926.81
E²DVCC3D✘57.7154.7056.17---
E²DVCTSN✘57.7755.1456.4234.1325.0128.87
E²DVCCLIP✘57.7054.6756.1434.7524.3628.64
", + "bbox": [ + 517, + 156, + 903, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sus among multiple references. Moreover, we employed SODA_c [10] for measuring the storytelling quality of generated captions. 2) For evaluating localization performance, we reported the average precision, average recall, and their harmonic mean, F1 score across IOU thresholds of $\\{0.3, 0.5, 0.7, 0.9\\}$ .", + "bbox": [ + 511, + 367, + 906, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 511, + 468, + 877, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dense Video Captioning Performance. In Table 1 and Table 2, we compare our $\\mathrm{E}^2\\mathrm{DVC}$ to the state-of-the-art methods for the task of dense video captioning on the ActivityNet Captions and YouCook2 datasets. Compared to baseline PDVC [48], our method achieves substantial gains for three different types of characteristics. From Table 1, our method achieves the best scores over BLEU4, METEOR, CIDEr even compared to the pretrained methods that use extra videos for pretraining. For Table 2, Vid2Seq [53] performs better on YouCook2 dataset than our method. This is because the YouCook2 dataset has fewer videos for training but a more diverse range of action types and semantic descriptions. Vid2seq uses an extra 1 million videos for pretraining, thus having a great advantage over our method. We achieve the best scores among all non-pretrained methods. This validates that by treating each event equally, our method will not overlook those out-of-distribution events and caption them with improved quality.", + "bbox": [ + 511, + 492, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Event Localization Performance. In Table 3, we compare our method with the state-of-the-art methods for the event localization task. For the ActivityNet Captions dataset, our method obtains the highest scores on the Precision, Recall, and F1 metrics among all. On the YouCook2 dataset, we achieve the highest performance over Precision and F1 scores even compared with the pretaining method. This is because our initialized queries will focus on each temporal location with high event possibilities. In this way, few", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8422", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cf625584407ea466224cb531ff46f64ec779bbb0482b275557e6fa599d4a8ff9.jpg", + "table_caption": [ + "Table 4. Ablation on the designed components. We report the results on YouCook2. PEI denotes the Pesudo-Event initialization module. EEE denotes the Event-Enhanced encoder module. Without PEI, the queries will be randomly initialized and are updated during the training process. Without EEE, we directly fuse visual features and positional embedding for each video frame together." + ], + "table_footnote": [], + "table_body": "
PEIEEEBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
××1.405.5629.694.9226.81
×✓1.495.7531.155.1328.04
✓×1.565.8830.925.2728.13
✓✓1.686.1134.265.3928.64
", + "bbox": [ + 94, + 183, + 480, + 275 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da870b60a2153e316086358e92bada0467ee251a481757ff44602fd39b08ae91.jpg", + "table_caption": [ + "Table 5. Ablation on varying number of clusters. We report the results on the Youcook2 dataset. The best and second performance results are highlighted." + ], + "table_footnote": [], + "table_body": "
NcBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.405.5629.694.9226.81
11.455.7431.455.1526.72
21.435.9231.325.1427.66
31.575.9532.575.2727.27
41.515.7830.895.0427.71
51.686.1134.265.3928.64
61.575.9033.275.2628.11
71.295.7830.105.1628.02
", + "bbox": [ + 94, + 340, + 478, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/990419524d1c00fc27c97111cc3df2876621edb6311180fa9d1d624292724328.jpg", + "table_caption": [ + "Table 6. Ablation on varying temporal event threshold $\\tau$ ." + ], + "table_footnote": [], + "table_body": "
τBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.325.5629.595.0027.21
11.355.5731.695.1727.51
21.425.6830.255.0827.70
31.595.7931.765.1227.48
41.686.1134.265.3928.64
51.575.8931.505.2228.41
61.515.8529.515.1828.62
71.465.8331.135.2428.24
", + "bbox": [ + 94, + 507, + 478, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "events will be overlooked. Also, our event-enhanced encoder has the pseudo-event labels as an additional input. This will help the transformer encoder to perceive the visual differences between video frames, thus preventing short-duration videos from being overwhelmed and helping the model obtain more accurate boundaries.", + "bbox": [ + 89, + 656, + 482, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 756, + 254, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Different Components in $\\mathbf{E}^2\\mathbf{DVC}$ . As illustrated in Table 4, we present an ablation study to explore the effectiveness of the Pseudo-Event Initialization module (PEI) and the Event-Enhanced Encoder (EEE). PEI initializes decoder queries with pseudo-event locations to ensure equal attention to all potential events, while EEE prevents those short-duration events from being overwhelmed during encoding. The results show that PEI significantly", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1f0324556ec6b89d8a8a9b61fd46e61eed3653f2b3b46d822efb1e37c2ead14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 87, + 709, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a466ae7a5d9490cd96274fe6d334ec0d478b0ca9a5337b232f260497d9b99ede.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 88, + 866, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07af83a93ffcf908a6d5286ba5bce1a923df98dd7f674a917381b084b379a745.jpg", + "image_caption": [ + "Figure 4. The distribution of the predicted and ground-truth events on ActivityNet Captions. Horizontal and vertical axes represent the normalized center time position and duration respectively. We compare our methods with PDVC [48] and $\\mathrm{CM}^2$ [18]." + ], + "image_footnote": [], + "bbox": [ + 553, + 189, + 709, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05e369ffb0f1c6faf135e839372b34498a444832212272db88d0c89783941de2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 189, + 864, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/148091f76c0a704e7dc01431d57251ef21fda19865b7fd7acf8b820ffa4ff66e.jpg", + "image_caption": [ + "(a) PDVC" + ], + "image_footnote": [], + "bbox": [ + 535, + 372, + 648, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f42a45ff867935008c6cc914cfbf79e4a0d3582fb6c037bd68dde8228a73df5d.jpg", + "image_caption": [ + "(b) $\\mathsf{E}^2\\mathsf{DVC}$" + ], + "image_footnote": [], + "bbox": [ + 658, + 375, + 766, + 457 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9ace2200c2147fd3fe07a68bdfc97960536900a84d7e4f656d6d2dd7eb7a497.jpg", + "image_caption": [ + "(c) GT", + "Figure 5. Different events duration proportion in ActivityNet Captions. Compared with PDVC, our method $\\mathrm{E}^2\\mathrm{DVC}$ could identify more short events and the distribution is more close to GT." + ], + "image_footnote": [], + "bbox": [ + 774, + 375, + 885, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "improves both dense video captioning and event localization performance, enabling more precise event detection and better captions by focusing queries on specific pseudo-event locations. Although the improvement from EEE is smaller, it still outperforms the baseline by enhancing event-level knowledge extraction. Combining PEI and EEE yields the best performance, which demonstrates the effectiveness of our proposed components.", + "bbox": [ + 511, + 547, + 905, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameter Analysis on the Number of Clusters $N_{c}$ . We try different numbers of clusters $N_{c}$ on the YouCook2 dataset. Table 5 shows the performance. The first row is the result from the baseline. When the cluster number is set to 1, all video frames are clustered together as a single pseudo-event, where the model only adds one additional query to focus on the middle of the video. In this case, the performance shows a slight improvement compared to the baseline. As the cluster number increases, video frames will be distinguished more finely based on visual differences. The best performance appears when $N_{c}$ is set to 5. When $N_{c}$ equals to the number of video frames, each frame will be allocated a single query to focus on. The model's attention will be distracted and the real events may be overlooked.", + "bbox": [ + 511, + 671, + 906, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameter Analysis on the Event Threshold $\\tau$ . The pa", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8423", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c95f0f9b42ee94e2ebd541cbd4d958c55ef749eb6a79de49b0ebfbcf42c63cf9.jpg", + "image_caption": [ + "Figure 6. Example visualizations of dense event captioning prediction on ActivityNet Captions. The color of the image border represents the category of the pseudo-event. From top to down, we show the results from the ground truth, the baseline (PDVC) and our method." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "riameter decides the lower-bound of pseudo-events duration. In the Table 6, we presents the results by changing the event threshold $\\tau$ in the YouCook2 dataset. When $\\tau$ is set too small, the performance drops significantly. It's because even isolated frames are treated as pseudo-events and are allocated certain decoder queries to focus on, which distracts the attention of the model. The best performance appears when $\\tau$ is set to 4. As $\\tau$ increases, the overall performance decreases. However, the performance of event localization remains at a high level. This is because long-duration events account for the majority and discarding some short-duration events will not hurt performance on long-duration ones.", + "bbox": [ + 88, + 361, + 480, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 556, + 272, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Event Temporal Distribution. Figure 4 shows the generated event temporal distribution from PDVC [48], $\\mathrm{CM}^2$ [18] and our method $\\mathrm{E}^2\\mathrm{DVC}$ . As can be seen from the figure, our method generates more short-duration events compared with the other two. The distribution predicted by our methods is more similar to the ground truth. We noticed that for long-duration events, we still cannot generate the precise duration. This is because the model is trained by captioning loss and localization loss together. For long-duration events, due to the large number of frames, some localization deviations have little effect on the final captioning performance. Even in some cases, predicting a shorter duration can help in describing the event more delicately, while predicting a longer duration allows for a more comprehensive caption of the event. Therefore, in order to achieve better captioning performance, some localization accuracy may be sacrificed.", + "bbox": [ + 91, + 580, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Events Duration Statistics. In Figure 5, we present the event duration statistics in the ActivityNet Captions dataset. Our proposed $\\mathrm{E}^2\\mathrm{DVC}$ outperforms the baseline PDVC by identifying a greater number of short", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "duration events. Besides, its event duration distribution is closer to the ground truth. This indicates that $\\mathrm{E}^2\\mathrm{DVC}$ is more effective at capturing events of varying lengths, particularly short-duration ones. By better aligning with the GT distribution, $\\mathrm{E}^2\\mathrm{DVC}$ demonstrates improved event localization performance, highlighting its ability to handle a wider range of event durations, leading to more accurate and comprehensive video captioning.", + "bbox": [ + 511, + 361, + 906, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Predicted Example. Figure 6 visualizes an example of dense video captioning predictions of PDVC and our method. Compared with PDVC, our method can localize short-duration events more accurately. This also helps improve the video captioning performance. The reason is that since short-duration events have fewer frames, they are more sensitive to localization deviations. If too many irrelevant frames are localized in the same short events, the description will become inaccurate.", + "bbox": [ + 511, + 484, + 908, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 645, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we propose the Event-Equalized Dense Video Captioning framework to address temporal bias and treat all events equally in video captioning. By leveraging a non-trainable Event Perception module with agglomerative hierarchical clustering, we identify pseudo-events and ensure temporal continuity through a refinement mechanism. The Event-Prior Initialization and Event-Prior Encoder modules allocate equal attention to all events, enhancing the model's precision, particularly for short-duration events. Experimental results on ActivityNet Captions and YouCook2 datasets validate the effectiveness of our approach, showing improved dense video captioning performance. $\\mathrm{E}^2\\mathrm{DVC}$ offers a robust solution for handling diverse event durations and temporal distributions, making it a promising advancement in dense video captioning.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8424", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement.", + "text_level": 1, + "bbox": [ + 89, + 90, + 274, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, and Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80.", + "bbox": [ + 89, + 113, + 485, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 209, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5", + "[3] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2", + "[4] Qiang Chen, Xiaokang Chen, Jian Wang, Shan Zhang, Kun Yao, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6633-6642, 2023. 3", + "[5] Shizhe Chen, Jia Chen, Qin Jin, and Alexander Hauptmann. Video captioning with guidance of multimodal latent topics. In Proceedings of the 25th ACM international conference on Multimedia, pages 1838-1846, 2017. 1", + "[6] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2021. 1, 2", + "[7] Chris Ding and Xiaofeng He. Cluster merging and splitting in hierarchical clustering algorithms. In 2002 IEEE International Conference on Data Mining, 2002. Proceedings., pages 139-146. IEEE, 2002. 2", + "[8] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[9] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep action proposals for action understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 768-784. Springer, 2016. 2", + "[10] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6" + ], + "bbox": [ + 93, + 234, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Jiyang Gao, Kan Chen, and Ram Nevatia. Ctap: Complementary temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 68-83, 2018. 2", + "[12] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1", + "[13] Bernard Ghanem, Juan Carlos Niebles, Cees Snoek, Fabian Caba Heilbron, Humam Alwassel, Victor Escoria, Ranjay Krishna, Shyamal Buch, and Cuong Duc Dao. The activitynet large-scale activity recognition challenge 2018 summary. arXiv preprint arXiv:1808.03766, 2018. 6", + "[14] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 5", + "[15] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. arXiv preprint arXiv:2005.08271, 2020. 1, 2, 4", + "[16] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 2", + "[17] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18198-18208, 2024. 1, 2", + "[18] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 1, 2, 5, 6, 7, 8", + "[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5", + "[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3", + "[21] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1", + "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2025. 1", + "[23] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7492-7500, 2018. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17958, 2022. 2", + "[25] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 2", + "[26] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3889-3898, 2019. 2", + "[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3", + "[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 1", + "[29] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3", + "[30] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6588-6597, 2019. 1, 2", + "[31] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6504-6512, 2017. 2", + "[32] K Papinesi. Bleu: A method for automatic evaluation of machine translation. In Proc. 40th Actual Meeting of the Association for Computational Linguistics (ACL), 2002, pages 311-318, 2002. 6", + "[33] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8347–8356, 2019. 1", + "[34] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 30(8):2617-2633, 2019. 1", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 5", + "[36] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8908-8917, 2019. 2", + "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5", + "[38] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1", + "[39] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 4, 5", + "[40] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5", + "[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6", + "[42] Bairui Wang, Lin Ma, Wei Zhang, and Wei Liu. Reconstruction network for video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7622-7631, 2018. 1, 2", + "[43] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7190-7198, 2018. 2", + "[44] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 1", + "[45] Junke Wang, Dongdong Chen, Chong Luo, Bo He, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Omnivid: A generative framework for universal video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18209-18220, 2024. 6", + "[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 1, 4, 5", + "[47] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1890-1900, 2020. 6", + "[48] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 1, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8426", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based object detection. arXiv preprint arXiv:2109.07107, 3(6), 2021. 3", + "[50] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 1", + "[51] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 468-483, 2018. 6", + "[52] Yifang Xu, Yunzhuo Sun, Benxiang Zhai, Youyao Jia, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 5", + "[53] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2, 3, 6", + "[54] Dali Yang and Chun Yuan. Hierarchical context encoding for events captioning in videos. In 2018 25th IEEE International Conference on Image Processing (ICIP), pages 1288-1292. IEEE, 2018. 2", + "[55] Manyuan Zhang, Guanglu Song, Yu Liu, and Hongsheng Li. Decoupled detr: Spatially disentangling localization and classification for improved end-to-end object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6601-6610, 2023. 3", + "[56] Qi Zhang, Yuqing Song, and Qin Jin. Unifying event detection and captioning as sequence generation via pre-training. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 6", + "[57] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1", + "[58] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2, 5", + "[59] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8739-8748, 2018. 5, 6", + "[60] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 1, 2", + "[61] Wanrong Zhu, Bo Pang, Ashish V Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning" + ], + "bbox": [ + 89, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "as sequence generation. arXiv preprint arXiv:2204.08121, 2022.6", + "[62] Yuhan Zhu, Guozhen Zhang, Jing Tan, Gangshan Wu, and Limin Wang. Dual detrs for multi-label temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18559-18569, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8427", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_model.json b/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8408efbb761d32f47f96e267f25acfdc15bdfb40 --- /dev/null +++ b/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.29, + 0.131, + 0.71, + 0.154 + ], + "angle": 0, + "content": "Event-Equalized Dense Video Captioning" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.18, + 0.804, + 0.217 + ], + "angle": 0, + "content": "Kangyi Wu* Pengna Li* Jingwen Fu Yizhe Li Yang Wu Yuhan Liu Jinjun Wang Sanping Zhou†" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.218, + 0.832, + 0.27 + ], + "angle": 0, + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.273, + 0.934, + 0.305 + ], + "angle": 0, + "content": "{wukangyi747600, sauerfisch, ful371252069, lyz0515, wuyang_CC, liuyuhan200095}@stu.xjtu.edu.cn jinjun@mail.xjtu.edu.cn spzhou@xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.34, + 0.327, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.372, + 0.485, + 0.72 + ], + "angle": 0, + "content": "Dense video captioning aims to localize and caption all events in arbitrary untrimmed videos. Although previous methods have achieved appealing results, they still face the issue of temporal bias, i.e., models tend to focus more on events with certain temporal characteristics. Specifically, 1) the temporal distribution of events in training datasets is uneven. Models trained on these datasets will pay less attention to out-of-distribution events. 2) long-duration events have more frame features than short ones and will attract more attention. To address this, we argue that events, with varying temporal characteristics, should be treated equally when it comes to dense video captioning. Intuitively, different events tend to have distinct visual differences due to varied camera views, backgrounds, or subjects. Inspired by that, we intend to utilize visual features to have an approximate perception of possible events and pay equal attention to them. In this paper, we introduce a simple but effective framework, called Event-Equalized Dense Video Captioning \\((E^2 DVC)\\) to overcome the temporal bias and treat all possible events equally. Experimental results on ActivityNet Captions and YouCook2 dataset validate the effectiveness of the proposed methods and show State-of-the-art (SOTA) performance on dense video captioning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.749, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.483, + 0.866 + ], + "angle": 0, + "content": "In recent years, an increasing number of works have begun to focus on video understanding[21, 22, 28, 38, 44, 50, 57] where video captioning is a challenging branch. Conventional video captioning [5, 12, 33, 34, 42] aims to generate a natural sentence to describe the main event of a short video clip. However, they face considerable challenges when ap" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.338, + 0.735, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.34, + 0.886, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.464, + 0.736, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.589, + 0.698, + 0.6 + ], + "angle": 0, + "content": "(a) Data Event Distribution" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.464, + 0.886, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.589, + 0.871, + 0.6 + ], + "angle": 0, + "content": "(b) Event Duration Ratio" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.752 + ], + "angle": 0, + "content": "Figure 1. Temporal distribution of ground truth and predicted events on ActivityNet Captions training set. The prediction is from PDVC [48]. (a) The distribution of all events in time dimensions. Horizontal and vertical axes represent the normalized center time position and duration respectively. It's obvious that events are distributed unevenly in these two temporal dimensions. (b) The proportion of the number of events in different duration intervals. It's clear that the predicted short-duration events are much fewer than the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.905, + 0.838 + ], + "angle": 0, + "content": "plied to realistic videos which are usually long, untrimmed, and contain a variety of events. To address the challenge, dense video captioning (DVC) [19] is proposed to automatically localize and caption all events in an untrimmed video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Existing DVC methods [6, 17, 18, 30, 60] follow a common pipeline to first extract image features from down-sampled video frames with pre-trained image encoder [15, 35, 39, 46] and then delicately design a framework to ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.12, + 0.876, + 0.314, + 0.888 + ], + "angle": 0, + "content": "* These authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.123, + 0.889, + 0.251, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.12, + 0.876, + 0.314, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8417" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.394 + ], + "angle": 0, + "content": "ploit event representations from them. Although state-of-the-art (SOTA) performance has been achieved, existing methods still face the issue of temporal bias, which means that models will pay more attention to events with certain temporal characteristics while overlooking some other ones. The temporal bias can be divided into two categories based on its causes: 1) Data distribution bias. As shown in Figure 1(a), the predicted events are scattered in the time axes. Events are completely ignored or mis-localized for most timestamps. This is because the models tend to learn from the dataset which part of the video to focus on and the ground truth events are unevenly distributed in this dimension. 2) Event duration bias. This bias results from the duration variation between events. Since short-duration events have fewer frames, their features are easy to be overwhelmed by others. Figure 1(b) shows the proportion of events in different duration intervals. It's obvious that the predicted short-duration events are much fewer than the ground truth. This means that most short-duration events are totally ignored or localized with large deviations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.593 + ], + "angle": 0, + "content": "To address the issues above, we argue that every event, regardless of its temporal characteristics, should share equal importance when it comes to dense video captioning. To achieve this goal, we need to reduce the influence of temporal information and ensure that attention is evenly distributed across all events. However, the challenge lies in how to perceive possible events in a video without relying on temporal information. Intuitively, different events have large visual differences considering their varied camera views, backgrounds, or subjects [19]. Inspired by this, we intend to utilize the differences in pure visual features to gain an initial perception of segments where events may exist. Equal Attention will be allocated to all these segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.822 + ], + "angle": 0, + "content": "In this paper, we propose a novel Event-Equalized Dense Video Captioning framework \\((\\mathrm{E}^2\\mathrm{DVC})\\) to overcome the temporal bias and treat each potential event equally. First, a non-trainable Event Perception module (EPM) is devised to utilize agglomerative hierarchical clustering algorithm [7] on visual features to generate video frame clusters. These clusters are segments with potential events. A refinement mechanism is designed to split frames that are discontinuous in time and discard isolated frames. These refined clusters are regarded as pseudo-events. Since the clustering process is non-trainable, the result will not be affected by the uneven data distribution. Additionally, since agglomerative hierarchical clustering is a non-uniform method, frames from short-duration events can also be equally clustered into the same category as the long ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With the pseudo-events, equal attention is required among them. To achieve this, a Pseudo-event Initialization module (PEI) is introduced. Concretely, we calculate the temporal locations of these pseudo-events (center timestamps) and use a temporal encoder to encode them" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "as pseudo-event queries. These queries serve as the initial guess of the event's locations to help the model allocate its attention. A novel Event-Enhanced Encoder (EEE) is also proposed to explore frame-frame and frame-event relationships during the encoding process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Our \\(\\mathrm{E}^2\\mathrm{DVC}\\) can determine and localize events with higher precision which further improves the dense video captioning performance. The main contributions of our work can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- We identify the issue of temporal bias and propose an Event-Equalized Dense Video Captioning framework to treat all events equally." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- We propose a non-trainable Event Perception module, a Pseudo-Event Initialization module, and an Event-Enhanced Encoder to ensure all events are noticed. Our components could perceive the input video with pure visual features and allocate equal attention to events with large temporal variations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.905, + 0.424 + ], + "angle": 0, + "content": "- Comprehensive experiments have been conducted on the ActivityNet Captions [19] and the YouCook2 [58] datasets to verify the effectiveness of our model in the dense video captioning task." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.663, + 0.453 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.719 + ], + "angle": 0, + "content": "Dense Video Captioning. Dense video captioning consists of two subtasks: event localization [9, 11, 25, 26] and event captioning [24, 31, 42]. Based on different pipelines, the methods can be divided into two categories: two-stage methods and end-to-end methods. For two-stage methods, they solve dense video captioning by a \"localize-then-describe\" scheme. These methods first predict a set of event proposals with corresponding time boundaries and then generate a detailed sentence description for each proposal by further extracting semantic knowledge and visual context. Krishna et al. [19] propose the first two-stage dense video captioning model. Wang et al. [43] and Yang et al. [54] focus on mining useful contextual information to improve DVC performance. Iashin et al. [15, 16] construct architectures to utilize any number of modalities for event description. However, these methods lack inter-task interactions and heavily rely on hand-crafted components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address the limitations above, recent methods follow the end-to-end pipeline [3, 6, 18, 23, 30, 36, 53, 60]. Li et al. [23] propose a joint and global optimization of detection and captioning. Islam et al. [17] and Deng et al. [6] solve DVC in hierarchical ways where the former follows a bottom-up structure while the latter proposes a top-down framework. Mun et al. [30] and Zhou et al. [60] leverage visual and linguistic context from prior events for coherent storytelling while the latter proposes an additional memory module and streaming decoding algorithm to support streaming video input. Kim et al. [18] utilize external memory to incorporate prior video-to-text knowledge to gener-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8418" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.909, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.36, + 0.908, + 0.472 + ], + "angle": 0, + "content": "Figure 2. The overview of our \\(\\mathbf{E}^2\\mathbf{DVC}\\). Our model follows an encoder-decoder structure. Given an input video, a pre-trained visual encoder is adopted to extract frame-level features. (a) Event Perception Module (EPM). Based on visual features similarity, these frames are clustered to multiple uneven segments and a refinement mechanism is utilized to discard outliers. The remaining segments are assigned with pseudo-events. (b) Event-Enhanced Encoder (EEE). By incorporate the pseudo-event labels with visual frame features, it could enhance the encoder's ability to explore frame-event and event-event relationships, preventing short-duration events from being overwhelmed. (c) Pseudo-Event Initialization module (PEI). Given the pseudo events, we calculate and encode their temporal locations and use the embedding to initialize the queries, helping the model allocate its attention equally among them. Finally, we obtain the predictions of the event-number, time localization and event captions through heads." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.498, + 0.484, + 0.62 + ], + "angle": 0, + "content": "ate natural descriptions. Yang et al. [53] reformulate sentence boundaries of transcribed speech as pseudo-events to achieve large-scale video data and propose a multi-modal pertaining model trained on this dataset. However, it remains a significant challenge to detect all events and localize them with high precision in a video. We propose a simple yet effective approach to enable the model to treat all events equally thus improving the DVC performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Query Formulation in Transformers. Since Detection Transformer (DETR)[2] has achieved promising performance in object detection without the need for any handcrafted components, many researchers have studied and further modified the model. Among these, research on decoder query formulation has received significant attention. Conditional DETR [29] learns a conditional spatial query by encoding the reference points as the query position embedding to help focus on distinct regions. Anchor DETR [49] and DAB-DETR [27] utilize anchor boxes to initialize queries to improve interpretability and force each query to only pay attention to a specific region in the object detection task. DN-DETR [20] additionally adds ground-truth bounding boxes with noises into decoder queries to reduce bipartite graph matching difficulty and achieve faster convergence. Group DETR [4] utilizes multiple groups of queries to conduct one-to-many assignments in order to speed up training and improve performance. Decoupled DETR [55] discovers that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.665 + ], + "angle": 0, + "content": "using shared queries for classification and localization will lead to suboptimal results. Instead, they introduce a task-aware query generation module to decouple these two tasks. DualDETR [62] presents a joint query initialization strategy to align queries from the instance level and boundary level to obtain better performance in the Temporal Action Detection task. In this paper, we introduce the query formulation strategies into the Dense Video Captioning task. Our \\(\\mathrm{E}^2\\mathrm{DVC}\\) employs a novel Pseudo-Event Initialization module to initialize the decoder queries with pseudo-events to ensure that every possible events are treated equally." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.605, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our goal is to treat each event equally and localize and caption all events without omission. To realize this, we introduce an Event Perception module (Section 3.1) with a delicately devised refinement mechanism to perceive the input video with only visual features available and select those segments where events may exist. Furthermore, A novel Event-Enhanced Encoder (Section 3.2) is also introduced to incorporate the pseudo-event labels into the self-attention framework in order to explore frame-frame and frame-event relationships. Meanwhile, we design an Pseudo Event Initialization module (Section 3.3) to ensure that the model pays equal attention to those pseudo-event locations. The overall framework of \\(\\mathrm{E}^2\\mathrm{DVC}\\) is illustrated in Figure 2," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8419" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.488, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.187, + 0.486, + 0.272 + ], + "angle": 0, + "content": "Figure 3. Illustration of the refinement mechanism. Different shapes denote different events. In the same events, frames with different color are temporal consistent. Our refinement mechanism could split the temporal inconsistent events into multiple new consistent events and discard the extremely short events which is short than the defined temporal event threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.485, + 0.4 + ], + "angle": 0, + "content": "which follows a standard parallel encoder-decoder structure with subtask heads. The objective is to generate \\( \\mathbf{N} \\) pairs of events and captions \\( \\{(t_s^n, t_e^n, Cap^n)\\}_{n=1}^N \\), where \\( \\mathbf{N} \\) denotes the number of events detected in the video. \\( t_s^n \\) and \\( t_e^n \\) denote the start and end timestamps of the \\( n \\)-th event. \\( Cap^n \\) denotes the corresponding caption. The details of task heads and training loss will be introduced in Section 3.4 and 3.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.407, + 0.325, + 0.423 + ], + "angle": 0, + "content": "3.1. Event Perception Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.581 + ], + "angle": 0, + "content": "To perceive the whole video with only visual features and pick out those pseudo-events, we devise an event perception module with agglomerative hierarchical clustering. The input video is first sampled in a certain ratio to obtain the image frames \\( V = \\{v_{i}\\}_{i=1}^{F} \\) where \\( F \\) denotes the number of frames. Then, a pre-trained image encoder \\( E \\) [15, 35, 39, 46] is utilized to extract visual features \\( X = \\{x_{i}\\}_{i=1}^{F} \\) from each video frame. Based on the visual feature similarities, we perform a hierarchical clustering algorithm to obtain pseudo-events." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.793 + ], + "angle": 0, + "content": "In this paper, we choose the Agglomerative clustering algorithm because it doesn't assume any particular cluster shape and sizes. It could progressively merge the individual frames to generate uneven clusters with different sizes and assign them with pseudo-event labels. Specifically, each visual feature \\( x_{i} \\) is initially considered an individual cluster with one element. We calculate the pairwise distance matrix among all the clusters. The two clusters that are the most close in feature space are combined into a bigger cluster. Then we recalculate the distance matrix. This procedure is iterated until the final number of clusters is equal to \\( N_{c} \\) which is a hyper-parameter which is set to control the granularity of the clustering process. Finally, we obtain a number of clusters \\( \\mathbb{C} \\) which can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.796, + 0.484, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mathbb {C} = \\left\\{\\mathbb {C} _ {i} \\right\\} _ {i = 1} ^ {N _ {c}} = \\left\\{\\left\\{x _ {i} ^ {j} \\right\\} _ {j = 1} ^ {N _ {c} ^ {i}} \\right\\} _ {i = 1} ^ {N _ {c}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.855 + ], + "angle": 0, + "content": "in which \\(\\mathbb{C}_i\\) denotes the \\(i\\)-th cluster and \\(N_c^i\\) denotes the number of this cluster." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Additionally, since the timestamps of the frame features are not considered during the clustering process, some clustered frames are temporally inconsistent with others in the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "same cluster. Considering that frames belonging to the same event should be temporally consistent, as illustrated in Figure 3, we utilize a refinement mechanism to split the temporally discontinuous events into multiple new consistent events based on the timestamps of the video frames. Moreover, since an event typically occupies a certain duration, we devise a temporal event threshold \\(\\tau\\). If the durations of the events are shorter than \\(\\tau\\), these events would be considered outliers and discarded. The remains are regarded as final pseudo-events. Our framework will pay special attention to these pseudo-event locations with improved localization and captioning performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.283, + 0.749, + 0.298 + ], + "angle": 0, + "content": "3.2. Event-Enhanced Encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.906, + 0.442 + ], + "angle": 0, + "content": "After obtaining the pseudo events, we directly assign the video frames with their pseudo labels. In this section, we introduce how to incorporate the assigned labels with visual features into transformer encoder to get better representations. We hope that in the process of encoding frame features, the focus should not only be on the frame-frame relationships but also on the frame-event one. In this way, frame features from short-duration events will not be overwhelmed during the self-attention process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.625 + ], + "angle": 0, + "content": "To realize this, we devise an Event-Enhanced Encoder (EEE) to add pseudo-event labels to the frame feature encoding process. Specifically, for each video frame \\( v_{i} \\), we have its pseudo-event label \\( l_{i} \\). The pseudo-event label \\( l_{i} \\) is transformed into a label embedding \\( le_{i} \\) with a label encoder, which consists of a learnable dictionary and a multilayer perceptron. During the encoding process, we concatenate the frame-level features \\( X = \\{x_{i}\\}_{i=1}^{F} \\) with their corresponding label embedding and positional embedding (PE). They are fed into the deformable transformer encoder and a multi-scale convolution layers to extract multi-scale features with frame-frame and frame-event relationships." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.76, + 0.649 + ], + "angle": 0, + "content": "3.3. Pseudo-Event Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Given the pseudo-events from Event Perception module, we want the model to pay equal attention to their temporal locations. Previously, the decoder aimed to query the event-level features from all frame features directly with \\( N_{q} \\) randomly initialized and learnable event queries. In this way, each event query is associated with multiple random temporal locations and learns which part to focus on from the uneven training dataset, leading to those events with out-of-distribution temporal characteristics easy to be overlooked." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.795, + 0.906, + 0.884 + ], + "angle": 0, + "content": "To address this issue, we devise a novel Pseudo-Event Initialization (PEI) module to initialize those event queries with pseudo-events. Specifically, given a pseudo-event \\( E_{i} = \\{x_{i,j}\\}_{j = 1}^{N_{i}} \\) in which \\( L_{i} \\) is the number of frames for \\( N_{i} \\), we calculate its center temporal location \\( t_i \\) (reference points) with:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\nt _ {i} = \\left(t _ {i, 1} + t _ {i, L _ {i}}\\right) / 2, \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "in which \\( t_{i,j} \\) is the timestamp for frame feature \\( x_{i,j} \\). In this way, for each pseudo-event \\( E_{i} \\), we have its reference points \\( t_{i} \\). We calculate the inverse sigmoid \\( (Inv_{Sig}) \\) of \\( t_{i} \\) and use temporal encoder \\( f_{t} \\) [40] to get the temporal embedding. Instead of randomly initializing the event queries, we initialize them with the positional embeddings of all pseudo-events. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.205, + 0.482, + 0.222 + ], + "angle": 0, + "content": "\\[\nq _ {i} = f _ {t} \\left(\\operatorname {I n v} _ {\\operatorname {S i g}} \\left(t _ {i}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.484, + 0.349 + ], + "angle": 0, + "content": "in which \\( q_{i} \\) is the \\( i \\)-th event query. What's more, considering that the pseudo-events may not cover all possible events in the ground truth, we also concatenate \\( N_{r} \\) randomly initialized queries with these pseudo-event queries together. In this way, the event queries explicitly prioritize the locations of these pseudo-events. This significantly enhances the performance of event localization and further improves the event captioning performance as well." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.355, + 0.218, + 0.37 + ], + "angle": 0, + "content": "3.4. Task Heads" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.439 + ], + "angle": 0, + "content": "Our work follows an encoder-decoder framework with a parallel decoding structure [18, 48, 52]. Given the event features, three parallel prediction heads are utilized to do dense video captioning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.516 + ], + "angle": 0, + "content": "Localization Head. The localization head is responsible for box prediction and binary classification for each event query \\( q_{i} \\). It outputs a set of tuples \\( \\{t_s^i, t_e^i, c^i\\}_{i=1}^{N_q} \\), where \\( t_s^i \\) and \\( t_e^i \\) denote the start and end timestamps and \\( c^i \\) represents the localization confidence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Captioning Head. We utilize LSTM [14] as the backbone of our captioning head. At each timestamp \\( t \\), we feed context features \\( z_{i,t} \\), event query features \\( q_i \\) and previous words \\( \\{w_{i,j}\\}_{j = 1}^{t - 1} \\) into the LSTM model to get the current word \\( w_{i,t} \\). As the sentence is generated, the captioning head produces the complete sentence \\( S_{i} = w_{i,1},\\dots,w_{l,T} \\), where \\( T \\) denotes the length of the sentence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Event Counter. It aims to detect the number of events in the given video. It first compresses the most salient information of event queries \\( Q \\) to a global feature vector with a max-pooling layer and predicts a fixed-size vector \\( f_{count} \\) with an FC layer with softmax activation. In \\( f_{count} \\), each value refers to the possibility of a specific number. During the inference stage, the event number \\( N_{event} \\) is obtained by \\( N_{event} = \\arg\\max(f_{count}) \\). The final outputs are obtained by selecting \\( N_{event} \\) events with the highest confidence score \\( c \\) which is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.779, + 0.483, + 0.82 + ], + "angle": 0, + "content": "\\[\nc _ {i} = c _ {i} ^ {l o c} + \\frac {\\mu}{M _ {i} ^ {\\gamma}} \\sum_ {t = 1} ^ {M _ {i}} \\log \\left(c _ {i, t} ^ {c a p}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "in which \\( M_{i} \\) is the number of words for the \\( i \\)-th event caption. \\( C_{i}^{loc} \\) is the localization confidence for the \\( i \\)-th event. \\( c_{i,t}^{cap} \\) is the confidence score for the \\( t \\)-th word for the \\( i \\)-th event. \\( \\mu \\) and \\( \\gamma \\) are two hyper-parameters to control the influence of the captioning confidence." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.66, + 0.106 + ], + "angle": 0, + "content": "3.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Suppose \\( N_{q} \\) events are predicted with their localizations and captions, Hungarian algorithm [2] is utilized to find the best bipartite matching results between the predicted events and the ground truth, during which the generalized IOU [37] between the predicted temporal segments and ground-truth segments is calculated as \\( L_{giou} \\) and the focal loss between the predicted classification score and the ground-truth label is calculated as \\( L_{cls} \\). The matched event pairs are selected to calculate the set prediction loss, which is the weighted sum of gIOU loss, classification loss, countering loss, and caption loss. The loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.295, + 0.907, + 0.312 + ], + "angle": 0, + "content": "\\[\nL = \\alpha_ {g i o u} L _ {g i o u} + \\alpha_ {c l s} L _ {c l s} + \\alpha_ {e c} L _ {e c} + \\alpha_ {c a p} L _ {c a p}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.906, + 0.4 + ], + "angle": 0, + "content": "in which \\( L_{cap} \\) is the cross-entropy loss between the predicted word probability and the ground truth normalized by the caption length \\( M \\). \\( L_{ec} \\) is the cross-entropy loss between the predicted count distribution and the ground truth. More details can be found in Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.414, + 0.646, + 0.431 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.704, + 0.456 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on the two commonly used dense video captioning benchmark datasets: ActivityNet Captions [19] and YouCook2 [58]. ActivityNet Captions includes over 20000 long videos covering a wide range of human daily activities and actions. Each video spans about 120s and is annotated with 3.7 temporally localized sentences on average. YouCook2 consists of 2000 long untrimmed videos from 89 cooking recipes. Generally, each video lasts approximately 320s and has 7.7 annotated segments with the corresponding sentences. For both datasets, we adhere to the standard split for training, validation, and testing. It's worth noting that we only use those videos that are still online on YouTube, leading to roughly \\(7\\%\\) less videos compared to the original datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For both datasets, we sample video frames at a rate of 1 FPS (one frame per second) and subsample or pad the sequence of frames to reach a total of F frames, where F is set to 100 for ActivityNet Captions and 200 for YouCook2. We use CLIP ViT-L/14 [8, 35] to extract clip visual features for video frames [18]. To further investigate the generalization ability of our model to different frame-level features, we also evaluate our model based on C3D [39] features provided by PDVC [48] and TSN [46] features provided by [59]. We utilize a two-layer deformable transformer that incorporates multi-scale deformable attention across four levels. We adopt the Agglomerative Hierarchical cluster algorithm for generating pseudo events. For ActivityNet Captions, we set the number of event queries to 10, while for YouCook2, we set it" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8421" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.175 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art methods for dense video captioning in ActivityNet Captions. B4/M/C is short for BLEU4/Meteoror/CIDEr. * indicates the results of our implementation using the official code. The best performance results are highlighted for each feature type. We also present results from those pretrained models." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.184, + 0.482, + 0.451 + ], + "angle": 0, + "content": "
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
UEDVC [56]C3D✓---5.50
Vid2Seq [53]CLIP✓-8.5030.105.80
OmniVID [45]-✓1.737.5426.005.60
ECHR [47]C3Dx1.827.2014.73.20
E2ESG [61]C3Dx-3.5025.00-
PDVC [48]C3Dx1.657.5025.875.26
E²DVCC3Dx1.797.5426.835.32
MT [59]TSNx0.303.2014.70-
PDVC [48]TSNx1.787.9628.965.44
E²DVCTSNx2.038.0229.915.77
CM² [18]CLIPx2.388.5533.016.18
PDVC* [48]CLIPx2.218.0629.975.92
E²DVCCLIPx2.438.5733.636.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.469, + 0.485, + 0.514 + ], + "angle": 0, + "content": "Table 2. Comparison with the state-of-the-art methods for dense video captioning on Youcook2. Since C3D features for YouCook2 are not publicly available, we fail to provide the results." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.523, + 0.481, + 0.685 + ], + "angle": 0, + "content": "
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
Vid2Seq [53]CLIP✓-9.3047.107.90
MT [59]TSN×1.155.009.30-
PDVC [48]TSN×0.804.7522.714.42
E²DVCTSN×1.054.7623.154.31
CM² [18]CLIP×1.636.0831.665.34
PDVC* [48]CLIP×1.405.5629.694.92
E²DVCCLIP×1.686.1134.265.39
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.763 + ], + "angle": 0, + "content": "to 100. For the event counter, we select a maximum count of 10 for the ActivityNet Captions dataset and 20 for the YouCook2 dataset. We set \\( N_{c} \\) to 5 and \\( \\tau \\) to 4. Other model hyperparameters follow the recipe of PDVC [48]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate our model in two aspects: 1) To assess dense captioning performance, we adopt the official evaluation tool from ActivityNet Challenge [13], which use the metrics BLEU4 [32], METEOR [1], and CIDEr [41] to compute the average precision of matched pairs between generated captions and the ground truth. These scores are averaged over IOU at \\(\\{0.3, 0.5, 0.7, 0.9\\}\\). BLEU4 focuses on n-gram precision, METEOR incorporates semantic matching, and CIDEr emphasizes consen" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.145 + ], + "angle": 0, + "content": "Table 3. Comparison to state-of-the-art methods for event Localization. We report the results on ActivityNet Captions and Youcook2 datasets. PT denotes Pretrain. Pre./Rec. is short for Precision/Recall." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.157, + 0.905, + 0.355 + ], + "angle": 0, + "content": "
MethodFeatures PTActivityNetYouCook2
Pre.↑Rec.↑F1↑Pre.↑Rec.↑F1↑
Vid2Seq [53]CLIP✓53.9052.7053.2927.8027.9027.84
E2ESG [61]C3D✘---20.6020.7020.65
MFT [51]TSN✘51.4124.3133.01---
CM² [18]CLIP✘56.8153.7155.2133.3824.7628.43
PDVC* [48]CLIP✘56.3853.2754.7832.3722.8926.81
E²DVCC3D✘57.7154.7056.17---
E²DVCTSN✘57.7755.1456.4234.1325.0128.87
E²DVCCLIP✘57.7054.6756.1434.7524.3628.64
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.461 + ], + "angle": 0, + "content": "sus among multiple references. Moreover, we employed SODA_c [10] for measuring the storytelling quality of generated captions. 2) For evaluating localization performance, we reported the average precision, average recall, and their harmonic mean, F1 score across IOU thresholds of \\(\\{0.3, 0.5, 0.7, 0.9\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.879, + 0.485 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Dense Video Captioning Performance. In Table 1 and Table 2, we compare our \\(\\mathrm{E}^2\\mathrm{DVC}\\) to the state-of-the-art methods for the task of dense video captioning on the ActivityNet Captions and YouCook2 datasets. Compared to baseline PDVC [48], our method achieves substantial gains for three different types of characteristics. From Table 1, our method achieves the best scores over BLEU4, METEOR, CIDEr even compared to the pretrained methods that use extra videos for pretraining. For Table 2, Vid2Seq [53] performs better on YouCook2 dataset than our method. This is because the YouCook2 dataset has fewer videos for training but a more diverse range of action types and semantic descriptions. Vid2seq uses an extra 1 million videos for pretraining, thus having a great advantage over our method. We achieve the best scores among all non-pretrained methods. This validates that by treating each event equally, our method will not overlook those out-of-distribution events and caption them with improved quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Event Localization Performance. In Table 3, we compare our method with the state-of-the-art methods for the event localization task. For the ActivityNet Captions dataset, our method obtains the highest scores on the Precision, Recall, and F1 metrics among all. On the YouCook2 dataset, we achieve the highest performance over Precision and F1 scores even compared with the pretaining method. This is because our initialized queries will focus on each temporal location with high event possibilities. In this way, few" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8422" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.173 + ], + "angle": 0, + "content": "Table 4. Ablation on the designed components. We report the results on YouCook2. PEI denotes the Pesudo-Event initialization module. EEE denotes the Event-Enhanced encoder module. Without PEI, the queries will be randomly initialized and are updated during the training process. Without EEE, we directly fuse visual features and positional embedding for each video frame together." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.184, + 0.481, + 0.276 + ], + "angle": 0, + "content": "
PEIEEEBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
××1.405.5629.694.9226.81
×✓1.495.7531.155.1328.04
✓×1.565.8830.925.2728.13
✓✓1.686.1134.265.3928.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.332 + ], + "angle": 0, + "content": "Table 5. Ablation on varying number of clusters. We report the results on the Youcook2 dataset. The best and second performance results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.342, + 0.48, + 0.476 + ], + "angle": 0, + "content": "
NcBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.405.5629.694.9226.81
11.455.7431.455.1526.72
21.435.9231.325.1427.66
31.575.9532.575.2727.27
41.515.7830.895.0427.71
51.686.1134.265.3928.64
61.575.9033.275.2628.11
71.295.7830.105.1628.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.483, + 0.468, + 0.497 + ], + "angle": 0, + "content": "Table 6. Ablation on varying temporal event threshold \\( \\tau \\) ." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.508, + 0.48, + 0.645 + ], + "angle": 0, + "content": "
τBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.325.5629.595.0027.21
11.355.5731.695.1727.51
21.425.6830.255.0827.70
31.595.7931.765.1227.48
41.686.1134.265.3928.64
51.575.8931.505.2228.41
61.515.8529.515.1828.62
71.465.8331.135.2428.24
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.747 + ], + "angle": 0, + "content": "events will be overlooked. Also, our event-enhanced encoder has the pseudo-event labels as an additional input. This will help the transformer encoder to perceive the visual differences between video frames, thus preventing short-duration videos from being overwhelmed and helping the model obtain more accurate boundaries." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.255, + 0.772 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Analysis of Different Components in \\(\\mathbf{E}^2\\mathbf{DVC}\\). As illustrated in Table 4, we present an ablation study to explore the effectiveness of the Pseudo-Event Initialization module (PEI) and the Event-Enhanced Encoder (EEE). PEI initializes decoder queries with pseudo-event locations to ensure equal attention to all potential events, while EEE prevents those short-duration events from being overwhelmed during encoding. The results show that PEI significantly" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.088, + 0.71, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.867, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.19, + 0.71, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.19, + 0.866, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.299, + 0.906, + 0.356 + ], + "angle": 0, + "content": "Figure 4. The distribution of the predicted and ground-truth events on ActivityNet Captions. Horizontal and vertical axes represent the normalized center time position and duration respectively. We compare our methods with PDVC [48] and \\(\\mathrm{CM}^2\\) [18]." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.373, + 0.65, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.462, + 0.615, + 0.472 + ], + "angle": 0, + "content": "(a) PDVC" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.375, + 0.767, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.461, + 0.74, + 0.472 + ], + "angle": 0, + "content": "(b) \\(\\mathsf{E}^2\\mathsf{DVC}\\)" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.375, + 0.887, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.461, + 0.847, + 0.472 + ], + "angle": 0, + "content": "(c) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.483, + 0.906, + 0.525 + ], + "angle": 0, + "content": "Figure 5. Different events duration proportion in ActivityNet Captions. Compared with PDVC, our method \\(\\mathrm{E}^2\\mathrm{DVC}\\) could identify more short events and the distribution is more close to GT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.67 + ], + "angle": 0, + "content": "improves both dense video captioning and event localization performance, enabling more precise event detection and better captions by focusing queries on specific pseudo-event locations. Although the improvement from EEE is smaller, it still outperforms the baseline by enhancing event-level knowledge extraction. Combining PEI and EEE yields the best performance, which demonstrates the effectiveness of our proposed components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.883 + ], + "angle": 0, + "content": "Parameter Analysis on the Number of Clusters \\( N_{c} \\). We try different numbers of clusters \\( N_{c} \\) on the YouCook2 dataset. Table 5 shows the performance. The first row is the result from the baseline. When the cluster number is set to 1, all video frames are clustered together as a single pseudo-event, where the model only adds one additional query to focus on the middle of the video. In this case, the performance shows a slight improvement compared to the baseline. As the cluster number increases, video frames will be distinguished more finely based on visual differences. The best performance appears when \\( N_{c} \\) is set to 5. When \\( N_{c} \\) equals to the number of video frames, each frame will be allocated a single query to focus on. The model's attention will be distracted and the real events may be overlooked." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Parameter Analysis on the Event Threshold \\(\\tau\\). The pa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8423" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Example visualizations of dense event captioning prediction on ActivityNet Captions. The color of the image border represents the category of the pseudo-event. From top to down, we show the results from the ground truth, the baseline (PDVC) and our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.482, + 0.543 + ], + "angle": 0, + "content": "riameter decides the lower-bound of pseudo-events duration. In the Table 6, we presents the results by changing the event threshold \\(\\tau\\) in the YouCook2 dataset. When \\(\\tau\\) is set too small, the performance drops significantly. It's because even isolated frames are treated as pseudo-events and are allocated certain decoder queries to focus on, which distracts the attention of the model. The best performance appears when \\(\\tau\\) is set to 4. As \\(\\tau\\) increases, the overall performance decreases. However, the performance of event localization remains at a high level. This is because long-duration events account for the majority and discarding some short-duration events will not hurt performance on long-duration ones." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.274, + 0.573 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.582, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Visualization of Event Temporal Distribution. Figure 4 shows the generated event temporal distribution from PDVC [48], \\(\\mathrm{CM}^2\\) [18] and our method \\(\\mathrm{E}^2\\mathrm{DVC}\\). As can be seen from the figure, our method generates more short-duration events compared with the other two. The distribution predicted by our methods is more similar to the ground truth. We noticed that for long-duration events, we still cannot generate the precise duration. This is because the model is trained by captioning loss and localization loss together. For long-duration events, due to the large number of frames, some localization deviations have little effect on the final captioning performance. Even in some cases, predicting a shorter duration can help in describing the event more delicately, while predicting a longer duration allows for a more comprehensive caption of the event. Therefore, in order to achieve better captioning performance, some localization accuracy may be sacrificed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Visualization of Events Duration Statistics. In Figure 5, we present the event duration statistics in the ActivityNet Captions dataset. Our proposed \\(\\mathrm{E}^2\\mathrm{DVC}\\) outperforms the baseline PDVC by identifying a greater number of short" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.907, + 0.483 + ], + "angle": 0, + "content": "duration events. Besides, its event duration distribution is closer to the ground truth. This indicates that \\(\\mathrm{E}^2\\mathrm{DVC}\\) is more effective at capturing events of varying lengths, particularly short-duration ones. By better aligning with the GT distribution, \\(\\mathrm{E}^2\\mathrm{DVC}\\) demonstrates improved event localization performance, highlighting its ability to handle a wider range of event durations, leading to more accurate and comprehensive video captioning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.622 + ], + "angle": 0, + "content": "Visualization of Predicted Example. Figure 6 visualizes an example of dense video captioning predictions of PDVC and our method. Compared with PDVC, our method can localize short-duration events more accurately. This also helps improve the video captioning performance. The reason is that since short-duration events have fewer frames, they are more sensitive to localization deviations. If too many irrelevant frames are localized in the same short events, the description will become inaccurate." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.634, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In conclusion, we propose the Event-Equalized Dense Video Captioning framework to address temporal bias and treat all events equally in video captioning. By leveraging a non-trainable Event Perception module with agglomerative hierarchical clustering, we identify pseudo-events and ensure temporal continuity through a refinement mechanism. The Event-Prior Initialization and Event-Prior Encoder modules allocate equal attention to all events, enhancing the model's precision, particularly for short-duration events. Experimental results on ActivityNet Captions and YouCook2 datasets validate the effectiveness of our approach, showing improved dense video captioning performance. \\(\\mathrm{E}^2\\mathrm{DVC}\\) offers a robust solution for handling diverse event durations and temporal distributions, making it a promising advancement in dense video captioning." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8424" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.276, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.184 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, and Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.21, + 0.188, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.235, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.306, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[3] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.513 + ], + "angle": 0, + "content": "[4] Qiang Chen, Xiaokang Chen, Jian Wang, Shan Zhang, Kun Yao, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6633-6642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.514, + 0.482, + 0.568 + ], + "angle": 0, + "content": "[5] Shizhe Chen, Jia Chen, Qin Jin, and Alexander Hauptmann. Video captioning with guidance of multimodal latent topics. In Proceedings of the 25th ACM international conference on Multimedia, pages 1838-1846, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.569, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[6] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[7] Chris Ding and Xiaofeng He. Cluster merging and splitting in hierarchical clustering algorithms. In 2002 IEEE International Conference on Data Mining, 2002. Proceedings., pages 139-146. IEEE, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.694, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.736, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[9] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep action proposals for action understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 768-784. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[10] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.235, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[11] Jiyang Gao, Kan Chen, and Ram Nevatia. Ctap: Complementary temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 68-83, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[12] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[13] Bernard Ghanem, Juan Carlos Niebles, Cees Snoek, Fabian Caba Heilbron, Humam Alwassel, Victor Escoria, Ranjay Krishna, Shyamal Buch, and Cuong Duc Dao. The activitynet large-scale activity recognition challenge 2018 summary. arXiv preprint arXiv:1808.03766, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.3 + ], + "angle": 0, + "content": "[14] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.904, + 0.343 + ], + "angle": 0, + "content": "[15] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. arXiv preprint arXiv:2005.08271, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[16] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[17] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18198-18208, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[18] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[21] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[23] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7492-7500, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8425" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17958, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.176, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[25] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[26] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3889-3898, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[29] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[30] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6588-6597, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[31] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6504-6512, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[32] K Papinesi. Bleu: A method for automatic evaluation of machine translation. In Proc. 40th Actual Meeting of the Association for Computational Linguistics (ACL), 2002, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[33] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8347–8356, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[34] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 30(8):2617-2633, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8908-8917, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[38] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[39] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[40] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.907, + 0.497 + ], + "angle": 0, + "content": "[42] Bairui Wang, Lin Ma, Wei Zhang, and Wei Liu. Reconstruction network for video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7622-7631, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[43] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7190-7198, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[44] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[45] Junke Wang, Dongdong Chen, Chong Luo, Bo He, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Omnivid: A generative framework for universal video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18209-18220, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.832 + ], + "angle": 0, + "content": "[47] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1890-1900, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 1, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8426" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[49] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based object detection. arXiv preprint arXiv:2109.07107, 3(6), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[50] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[51] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 468-483, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[52] Yifang Xu, Yunzhuo Sun, Benxiang Zhai, Youyao Jia, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.334, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[53] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.42, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[54] Dali Yang and Chun Yuan. Hierarchical context encoding for events captioning in videos. In 2018 25th IEEE International Conference on Image Processing (ICIP), pages 1288-1292. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.476, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[55] Manyuan Zhang, Guanglu Song, Yu Liu, and Hongsheng Li. Decoupled detr: Spatially disentangling localization and classification for improved end-to-end object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6601-6610, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.548, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[56] Qi Zhang, Yuqing Song, and Qin Jin. Unifying event detection and captioning as sequence generation via pre-training. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.604, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[57] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[58] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[59] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8739-8748, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[60] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[61] Wanrong Zhu, Bo Pang, Ashish V Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "as sequence generation. arXiv preprint arXiv:2204.08121, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[62] Yuhan Zhu, Guozhen Zhang, Jing Tan, Gangshan Wu, and Limin Wang. Dual detrs for multi-label temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18559-18569, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8427" + } + ] +] \ No newline at end of file diff --git a/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_origin.pdf b/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd9bcf300dd4b11e12f7af084281aa4d77ded9ef --- /dev/null +++ b/2025/Event-Equalized Dense Video Captioning/fd5415a0-125d-45cf-9033-a51e081e5f8e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5548702cc75c924fe2e26126ca56a44583dbcfa16f5acff08bcb0d7372f70120 +size 2573104 diff --git a/2025/Event-Equalized Dense Video Captioning/full.md b/2025/Event-Equalized Dense Video Captioning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..646db6a873930d021d1ae75e827d7bceb06f4b1c --- /dev/null +++ b/2025/Event-Equalized Dense Video Captioning/full.md @@ -0,0 +1,306 @@ +# Event-Equalized Dense Video Captioning + +Kangyi Wu* Pengna Li* Jingwen Fu Yizhe Li Yang Wu Yuhan Liu Jinjun Wang Sanping Zhou† + +National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University + +{wukangyi747600, sauerfisch, ful371252069, lyz0515, wuyang_CC, liuyuhan200095}@stu.xjtu.edu.cn jinjun@mail.xjtu.edu.cn spzhou@xjtu.edu.cn + +# Abstract + +Dense video captioning aims to localize and caption all events in arbitrary untrimmed videos. Although previous methods have achieved appealing results, they still face the issue of temporal bias, i.e., models tend to focus more on events with certain temporal characteristics. Specifically, 1) the temporal distribution of events in training datasets is uneven. Models trained on these datasets will pay less attention to out-of-distribution events. 2) long-duration events have more frame features than short ones and will attract more attention. To address this, we argue that events, with varying temporal characteristics, should be treated equally when it comes to dense video captioning. Intuitively, different events tend to have distinct visual differences due to varied camera views, backgrounds, or subjects. Inspired by that, we intend to utilize visual features to have an approximate perception of possible events and pay equal attention to them. In this paper, we introduce a simple but effective framework, called Event-Equalized Dense Video Captioning $(E^2 DVC)$ to overcome the temporal bias and treat all possible events equally. Experimental results on ActivityNet Captions and YouCook2 dataset validate the effectiveness of the proposed methods and show State-of-the-art (SOTA) performance on dense video captioning. + +# 1. Introduction + +In recent years, an increasing number of works have begun to focus on video understanding[21, 22, 28, 38, 44, 50, 57] where video captioning is a challenging branch. Conventional video captioning [5, 12, 33, 34, 42] aims to generate a natural sentence to describe the main event of a short video clip. However, they face considerable challenges when ap + +![](images/d226cc13c95c008e8dde0a80b1c84703f77acb24348d907dd3e8b31e680ab12a.jpg) + +![](images/e807cdabd3fa7777c0601b09753afbfaf8da729408131892783681e7848a4c4e.jpg) + +![](images/04872f2e16adead48cda3838b466cfe7eeafd9e70f46198e5079566bfc629151.jpg) +(a) Data Event Distribution +(b) Event Duration Ratio + +![](images/7f3f9668bd550a93885f1d6feef72e72f797560689dd1a46ebc0b446340ade0c.jpg) +Figure 1. Temporal distribution of ground truth and predicted events on ActivityNet Captions training set. The prediction is from PDVC [48]. (a) The distribution of all events in time dimensions. Horizontal and vertical axes represent the normalized center time position and duration respectively. It's obvious that events are distributed unevenly in these two temporal dimensions. (b) The proportion of the number of events in different duration intervals. It's clear that the predicted short-duration events are much fewer than the ground truth. + +plied to realistic videos which are usually long, untrimmed, and contain a variety of events. To address the challenge, dense video captioning (DVC) [19] is proposed to automatically localize and caption all events in an untrimmed video. + +Existing DVC methods [6, 17, 18, 30, 60] follow a common pipeline to first extract image features from down-sampled video frames with pre-trained image encoder [15, 35, 39, 46] and then delicately design a framework to ex + +ploit event representations from them. Although state-of-the-art (SOTA) performance has been achieved, existing methods still face the issue of temporal bias, which means that models will pay more attention to events with certain temporal characteristics while overlooking some other ones. The temporal bias can be divided into two categories based on its causes: 1) Data distribution bias. As shown in Figure 1(a), the predicted events are scattered in the time axes. Events are completely ignored or mis-localized for most timestamps. This is because the models tend to learn from the dataset which part of the video to focus on and the ground truth events are unevenly distributed in this dimension. 2) Event duration bias. This bias results from the duration variation between events. Since short-duration events have fewer frames, their features are easy to be overwhelmed by others. Figure 1(b) shows the proportion of events in different duration intervals. It's obvious that the predicted short-duration events are much fewer than the ground truth. This means that most short-duration events are totally ignored or localized with large deviations. + +To address the issues above, we argue that every event, regardless of its temporal characteristics, should share equal importance when it comes to dense video captioning. To achieve this goal, we need to reduce the influence of temporal information and ensure that attention is evenly distributed across all events. However, the challenge lies in how to perceive possible events in a video without relying on temporal information. Intuitively, different events have large visual differences considering their varied camera views, backgrounds, or subjects [19]. Inspired by this, we intend to utilize the differences in pure visual features to gain an initial perception of segments where events may exist. Equal Attention will be allocated to all these segments. + +In this paper, we propose a novel Event-Equalized Dense Video Captioning framework $(\mathrm{E}^2\mathrm{DVC})$ to overcome the temporal bias and treat each potential event equally. First, a non-trainable Event Perception module (EPM) is devised to utilize agglomerative hierarchical clustering algorithm [7] on visual features to generate video frame clusters. These clusters are segments with potential events. A refinement mechanism is designed to split frames that are discontinuous in time and discard isolated frames. These refined clusters are regarded as pseudo-events. Since the clustering process is non-trainable, the result will not be affected by the uneven data distribution. Additionally, since agglomerative hierarchical clustering is a non-uniform method, frames from short-duration events can also be equally clustered into the same category as the long ones. + +With the pseudo-events, equal attention is required among them. To achieve this, a Pseudo-event Initialization module (PEI) is introduced. Concretely, we calculate the temporal locations of these pseudo-events (center timestamps) and use a temporal encoder to encode them + +as pseudo-event queries. These queries serve as the initial guess of the event's locations to help the model allocate its attention. A novel Event-Enhanced Encoder (EEE) is also proposed to explore frame-frame and frame-event relationships during the encoding process. + +Our $\mathrm{E}^2\mathrm{DVC}$ can determine and localize events with higher precision which further improves the dense video captioning performance. The main contributions of our work can be summarized as: + +- We identify the issue of temporal bias and propose an Event-Equalized Dense Video Captioning framework to treat all events equally. +- We propose a non-trainable Event Perception module, a Pseudo-Event Initialization module, and an Event-Enhanced Encoder to ensure all events are noticed. Our components could perceive the input video with pure visual features and allocate equal attention to events with large temporal variations. +- Comprehensive experiments have been conducted on the ActivityNet Captions [19] and the YouCook2 [58] datasets to verify the effectiveness of our model in the dense video captioning task. + +# 2. Related Works + +Dense Video Captioning. Dense video captioning consists of two subtasks: event localization [9, 11, 25, 26] and event captioning [24, 31, 42]. Based on different pipelines, the methods can be divided into two categories: two-stage methods and end-to-end methods. For two-stage methods, they solve dense video captioning by a "localize-then-describe" scheme. These methods first predict a set of event proposals with corresponding time boundaries and then generate a detailed sentence description for each proposal by further extracting semantic knowledge and visual context. Krishna et al. [19] propose the first two-stage dense video captioning model. Wang et al. [43] and Yang et al. [54] focus on mining useful contextual information to improve DVC performance. Iashin et al. [15, 16] construct architectures to utilize any number of modalities for event description. However, these methods lack inter-task interactions and heavily rely on hand-crafted components. + +To address the limitations above, recent methods follow the end-to-end pipeline [3, 6, 18, 23, 30, 36, 53, 60]. Li et al. [23] propose a joint and global optimization of detection and captioning. Islam et al. [17] and Deng et al. [6] solve DVC in hierarchical ways where the former follows a bottom-up structure while the latter proposes a top-down framework. Mun et al. [30] and Zhou et al. [60] leverage visual and linguistic context from prior events for coherent storytelling while the latter proposes an additional memory module and streaming decoding algorithm to support streaming video input. Kim et al. [18] utilize external memory to incorporate prior video-to-text knowledge to gener- + +![](images/a0a7923d0bc4a3e36cfd2e2d29283ea3bce9cebf9491e90e55cfcb7222c66924.jpg) +Figure 2. The overview of our $\mathbf{E}^2\mathbf{DVC}$ . Our model follows an encoder-decoder structure. Given an input video, a pre-trained visual encoder is adopted to extract frame-level features. (a) Event Perception Module (EPM). Based on visual features similarity, these frames are clustered to multiple uneven segments and a refinement mechanism is utilized to discard outliers. The remaining segments are assigned with pseudo-events. (b) Event-Enhanced Encoder (EEE). By incorporate the pseudo-event labels with visual frame features, it could enhance the encoder's ability to explore frame-event and event-event relationships, preventing short-duration events from being overwhelmed. (c) Pseudo-Event Initialization module (PEI). Given the pseudo events, we calculate and encode their temporal locations and use the embedding to initialize the queries, helping the model allocate its attention equally among them. Finally, we obtain the predictions of the event-number, time localization and event captions through heads. + +ate natural descriptions. Yang et al. [53] reformulate sentence boundaries of transcribed speech as pseudo-events to achieve large-scale video data and propose a multi-modal pertaining model trained on this dataset. However, it remains a significant challenge to detect all events and localize them with high precision in a video. We propose a simple yet effective approach to enable the model to treat all events equally thus improving the DVC performance. + +Query Formulation in Transformers. Since Detection Transformer (DETR)[2] has achieved promising performance in object detection without the need for any handcrafted components, many researchers have studied and further modified the model. Among these, research on decoder query formulation has received significant attention. Conditional DETR [29] learns a conditional spatial query by encoding the reference points as the query position embedding to help focus on distinct regions. Anchor DETR [49] and DAB-DETR [27] utilize anchor boxes to initialize queries to improve interpretability and force each query to only pay attention to a specific region in the object detection task. DN-DETR [20] additionally adds ground-truth bounding boxes with noises into decoder queries to reduce bipartite graph matching difficulty and achieve faster convergence. Group DETR [4] utilizes multiple groups of queries to conduct one-to-many assignments in order to speed up training and improve performance. Decoupled DETR [55] discovers that + +using shared queries for classification and localization will lead to suboptimal results. Instead, they introduce a task-aware query generation module to decouple these two tasks. DualDETR [62] presents a joint query initialization strategy to align queries from the instance level and boundary level to obtain better performance in the Temporal Action Detection task. In this paper, we introduce the query formulation strategies into the Dense Video Captioning task. Our $\mathrm{E}^2\mathrm{DVC}$ employs a novel Pseudo-Event Initialization module to initialize the decoder queries with pseudo-events to ensure that every possible events are treated equally. + +# 3. Method + +Our goal is to treat each event equally and localize and caption all events without omission. To realize this, we introduce an Event Perception module (Section 3.1) with a delicately devised refinement mechanism to perceive the input video with only visual features available and select those segments where events may exist. Furthermore, A novel Event-Enhanced Encoder (Section 3.2) is also introduced to incorporate the pseudo-event labels into the self-attention framework in order to explore frame-frame and frame-event relationships. Meanwhile, we design an Pseudo Event Initialization module (Section 3.3) to ensure that the model pays equal attention to those pseudo-event locations. The overall framework of $\mathrm{E}^2\mathrm{DVC}$ is illustrated in Figure 2, + +![](images/a6f7daf472a0e4d00fdf068eb72859d42e6572b90ec973e10ecb6d103643070a.jpg) +Figure 3. Illustration of the refinement mechanism. Different shapes denote different events. In the same events, frames with different color are temporal consistent. Our refinement mechanism could split the temporal inconsistent events into multiple new consistent events and discard the extremely short events which is short than the defined temporal event threshold. + +which follows a standard parallel encoder-decoder structure with subtask heads. The objective is to generate $\mathbf{N}$ pairs of events and captions $\{(t_s^n, t_e^n, Cap^n)\}_{n=1}^N$ , where $\mathbf{N}$ denotes the number of events detected in the video. $t_s^n$ and $t_e^n$ denote the start and end timestamps of the $n$ -th event. $Cap^n$ denotes the corresponding caption. The details of task heads and training loss will be introduced in Section 3.4 and 3.5. + +# 3.1. Event Perception Module + +To perceive the whole video with only visual features and pick out those pseudo-events, we devise an event perception module with agglomerative hierarchical clustering. The input video is first sampled in a certain ratio to obtain the image frames $V = \{v_{i}\}_{i=1}^{F}$ where $F$ denotes the number of frames. Then, a pre-trained image encoder $E$ [15, 35, 39, 46] is utilized to extract visual features $X = \{x_{i}\}_{i=1}^{F}$ from each video frame. Based on the visual feature similarities, we perform a hierarchical clustering algorithm to obtain pseudo-events. + +In this paper, we choose the Agglomerative clustering algorithm because it doesn't assume any particular cluster shape and sizes. It could progressively merge the individual frames to generate uneven clusters with different sizes and assign them with pseudo-event labels. Specifically, each visual feature $x_{i}$ is initially considered an individual cluster with one element. We calculate the pairwise distance matrix among all the clusters. The two clusters that are the most close in feature space are combined into a bigger cluster. Then we recalculate the distance matrix. This procedure is iterated until the final number of clusters is equal to $N_{c}$ which is a hyper-parameter which is set to control the granularity of the clustering process. Finally, we obtain a number of clusters $\mathbb{C}$ which can be formulated as + +$$ +\mathbb {C} = \left\{\mathbb {C} _ {i} \right\} _ {i = 1} ^ {N _ {c}} = \left\{\left\{x _ {i} ^ {j} \right\} _ {j = 1} ^ {N _ {c} ^ {i}} \right\} _ {i = 1} ^ {N _ {c}}, \tag {1} +$$ + +in which $\mathbb{C}_i$ denotes the $i$ -th cluster and $N_c^i$ denotes the number of this cluster. + +Additionally, since the timestamps of the frame features are not considered during the clustering process, some clustered frames are temporally inconsistent with others in the + +same cluster. Considering that frames belonging to the same event should be temporally consistent, as illustrated in Figure 3, we utilize a refinement mechanism to split the temporally discontinuous events into multiple new consistent events based on the timestamps of the video frames. Moreover, since an event typically occupies a certain duration, we devise a temporal event threshold $\tau$ . If the durations of the events are shorter than $\tau$ , these events would be considered outliers and discarded. The remains are regarded as final pseudo-events. Our framework will pay special attention to these pseudo-event locations with improved localization and captioning performance. + +# 3.2. Event-Enhanced Encoder + +After obtaining the pseudo events, we directly assign the video frames with their pseudo labels. In this section, we introduce how to incorporate the assigned labels with visual features into transformer encoder to get better representations. We hope that in the process of encoding frame features, the focus should not only be on the frame-frame relationships but also on the frame-event one. In this way, frame features from short-duration events will not be overwhelmed during the self-attention process. + +To realize this, we devise an Event-Enhanced Encoder (EEE) to add pseudo-event labels to the frame feature encoding process. Specifically, for each video frame $v_{i}$ , we have its pseudo-event label $l_{i}$ . The pseudo-event label $l_{i}$ is transformed into a label embedding $le_{i}$ with a label encoder, which consists of a learnable dictionary and a multilayer perceptron. During the encoding process, we concatenate the frame-level features $X = \{x_{i}\}_{i=1}^{F}$ with their corresponding label embedding and positional embedding (PE). They are fed into the deformable transformer encoder and a multi-scale convolution layers to extract multi-scale features with frame-frame and frame-event relationships. + +# 3.3. Pseudo-Event Initialization + +Given the pseudo-events from Event Perception module, we want the model to pay equal attention to their temporal locations. Previously, the decoder aimed to query the event-level features from all frame features directly with $N_{q}$ randomly initialized and learnable event queries. In this way, each event query is associated with multiple random temporal locations and learns which part to focus on from the uneven training dataset, leading to those events with out-of-distribution temporal characteristics easy to be overlooked. + +To address this issue, we devise a novel Pseudo-Event Initialization (PEI) module to initialize those event queries with pseudo-events. Specifically, given a pseudo-event $E_{i} = \{x_{i,j}\}_{j = 1}^{N_{i}}$ in which $L_{i}$ is the number of frames for $N_{i}$ , we calculate its center temporal location $t_i$ (reference points) with: + +$$ +t _ {i} = \left(t _ {i, 1} + t _ {i, L _ {i}}\right) / 2, \tag {2} +$$ + +in which $t_{i,j}$ is the timestamp for frame feature $x_{i,j}$ . In this way, for each pseudo-event $E_{i}$ , we have its reference points $t_{i}$ . We calculate the inverse sigmoid $(Inv_{Sig})$ of $t_{i}$ and use temporal encoder $f_{t}$ [40] to get the temporal embedding. Instead of randomly initializing the event queries, we initialize them with the positional embeddings of all pseudo-events. The process can be formulated as: + +$$ +q _ {i} = f _ {t} \left(\operatorname {I n v} _ {\operatorname {S i g}} \left(t _ {i}\right)\right), \tag {3} +$$ + +in which $q_{i}$ is the $i$ -th event query. What's more, considering that the pseudo-events may not cover all possible events in the ground truth, we also concatenate $N_{r}$ randomly initialized queries with these pseudo-event queries together. In this way, the event queries explicitly prioritize the locations of these pseudo-events. This significantly enhances the performance of event localization and further improves the event captioning performance as well. + +# 3.4. Task Heads + +Our work follows an encoder-decoder framework with a parallel decoding structure [18, 48, 52]. Given the event features, three parallel prediction heads are utilized to do dense video captioning. + +Localization Head. The localization head is responsible for box prediction and binary classification for each event query $q_{i}$ . It outputs a set of tuples $\{t_s^i, t_e^i, c^i\}_{i=1}^{N_q}$ , where $t_s^i$ and $t_e^i$ denote the start and end timestamps and $c^i$ represents the localization confidence. + +Captioning Head. We utilize LSTM [14] as the backbone of our captioning head. At each timestamp $t$ , we feed context features $z_{i,t}$ , event query features $q_i$ and previous words $\{w_{i,j}\}_{j = 1}^{t - 1}$ into the LSTM model to get the current word $w_{i,t}$ . As the sentence is generated, the captioning head produces the complete sentence $S_{i} = w_{i,1},\dots,w_{l,T}$ , where $T$ denotes the length of the sentence. + +Event Counter. It aims to detect the number of events in the given video. It first compresses the most salient information of event queries $Q$ to a global feature vector with a max-pooling layer and predicts a fixed-size vector $f_{count}$ with an FC layer with softmax activation. In $f_{count}$ , each value refers to the possibility of a specific number. During the inference stage, the event number $N_{event}$ is obtained by $N_{event} = \arg\max(f_{count})$ . The final outputs are obtained by selecting $N_{event}$ events with the highest confidence score $c$ which is calculated by: + +$$ +c _ {i} = c _ {i} ^ {l o c} + \frac {\mu}{M _ {i} ^ {\gamma}} \sum_ {t = 1} ^ {M _ {i}} \log \left(c _ {i, t} ^ {c a p}\right), \tag {4} +$$ + +in which $M_{i}$ is the number of words for the $i$ -th event caption. $C_{i}^{loc}$ is the localization confidence for the $i$ -th event. $c_{i,t}^{cap}$ is the confidence score for the $t$ -th word for the $i$ -th event. $\mu$ and $\gamma$ are two hyper-parameters to control the influence of the captioning confidence. + +# 3.5. Loss Function + +Suppose $N_{q}$ events are predicted with their localizations and captions, Hungarian algorithm [2] is utilized to find the best bipartite matching results between the predicted events and the ground truth, during which the generalized IOU [37] between the predicted temporal segments and ground-truth segments is calculated as $L_{giou}$ and the focal loss between the predicted classification score and the ground-truth label is calculated as $L_{cls}$ . The matched event pairs are selected to calculate the set prediction loss, which is the weighted sum of gIOU loss, classification loss, countering loss, and caption loss. The loss function is formulated as: + +$$ +L = \alpha_ {g i o u} L _ {g i o u} + \alpha_ {c l s} L _ {c l s} + \alpha_ {e c} L _ {e c} + \alpha_ {c a p} L _ {c a p}, \tag {5} +$$ + +in which $L_{cap}$ is the cross-entropy loss between the predicted word probability and the ground truth normalized by the caption length $M$ . $L_{ec}$ is the cross-entropy loss between the predicted count distribution and the ground truth. More details can be found in Supplementary. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate our method on the two commonly used dense video captioning benchmark datasets: ActivityNet Captions [19] and YouCook2 [58]. ActivityNet Captions includes over 20000 long videos covering a wide range of human daily activities and actions. Each video spans about 120s and is annotated with 3.7 temporally localized sentences on average. YouCook2 consists of 2000 long untrimmed videos from 89 cooking recipes. Generally, each video lasts approximately 320s and has 7.7 annotated segments with the corresponding sentences. For both datasets, we adhere to the standard split for training, validation, and testing. It's worth noting that we only use those videos that are still online on YouTube, leading to roughly $7\%$ less videos compared to the original datasets. + +Implementation Details. For both datasets, we sample video frames at a rate of 1 FPS (one frame per second) and subsample or pad the sequence of frames to reach a total of F frames, where F is set to 100 for ActivityNet Captions and 200 for YouCook2. We use CLIP ViT-L/14 [8, 35] to extract clip visual features for video frames [18]. To further investigate the generalization ability of our model to different frame-level features, we also evaluate our model based on C3D [39] features provided by PDVC [48] and TSN [46] features provided by [59]. We utilize a two-layer deformable transformer that incorporates multi-scale deformable attention across four levels. We adopt the Agglomerative Hierarchical cluster algorithm for generating pseudo events. For ActivityNet Captions, we set the number of event queries to 10, while for YouCook2, we set it + +Table 1. Comparison with the state-of-the-art methods for dense video captioning in ActivityNet Captions. B4/M/C is short for BLEU4/Meteoror/CIDEr. * indicates the results of our implementation using the official code. The best performance results are highlighted for each feature type. We also present results from those pretrained models. + +
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
UEDVC [56]C3D✓---5.50
Vid2Seq [53]CLIP✓-8.5030.105.80
OmniVID [45]-✓1.737.5426.005.60
ECHR [47]C3Dx1.827.2014.73.20
E2ESG [61]C3Dx-3.5025.00-
PDVC [48]C3Dx1.657.5025.875.26
E²DVCC3Dx1.797.5426.835.32
MT [59]TSNx0.303.2014.70-
PDVC [48]TSNx1.787.9628.965.44
E²DVCTSNx2.038.0229.915.77
CM² [18]CLIPx2.388.5533.016.18
PDVC* [48]CLIPx2.218.0629.975.92
E²DVCCLIPx2.438.5733.636.13
+ +Table 2. Comparison with the state-of-the-art methods for dense video captioning on Youcook2. Since C3D features for YouCook2 are not publicly available, we fail to provide the results. + +
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
Vid2Seq [53]CLIP✓-9.3047.107.90
MT [59]TSN×1.155.009.30-
PDVC [48]TSN×0.804.7522.714.42
E²DVCTSN×1.054.7623.154.31
CM² [18]CLIP×1.636.0831.665.34
PDVC* [48]CLIP×1.405.5629.694.92
E²DVCCLIP×1.686.1134.265.39
+ +to 100. For the event counter, we select a maximum count of 10 for the ActivityNet Captions dataset and 20 for the YouCook2 dataset. We set $N_{c}$ to 5 and $\tau$ to 4. Other model hyperparameters follow the recipe of PDVC [48]. + +Evaluation Metrics. We evaluate our model in two aspects: 1) To assess dense captioning performance, we adopt the official evaluation tool from ActivityNet Challenge [13], which use the metrics BLEU4 [32], METEOR [1], and CIDEr [41] to compute the average precision of matched pairs between generated captions and the ground truth. These scores are averaged over IOU at $\{0.3, 0.5, 0.7, 0.9\}$ . BLEU4 focuses on n-gram precision, METEOR incorporates semantic matching, and CIDEr emphasizes consen + +Table 3. Comparison to state-of-the-art methods for event Localization. We report the results on ActivityNet Captions and Youcook2 datasets. PT denotes Pretrain. Pre./Rec. is short for Precision/Recall. + +
MethodFeatures PTActivityNetYouCook2
Pre.↑Rec.↑F1↑Pre.↑Rec.↑F1↑
Vid2Seq [53]CLIP✓53.9052.7053.2927.8027.9027.84
E2ESG [61]C3D✘---20.6020.7020.65
MFT [51]TSN✘51.4124.3133.01---
CM² [18]CLIP✘56.8153.7155.2133.3824.7628.43
PDVC* [48]CLIP✘56.3853.2754.7832.3722.8926.81
E²DVCC3D✘57.7154.7056.17---
E²DVCTSN✘57.7755.1456.4234.1325.0128.87
E²DVCCLIP✘57.7054.6756.1434.7524.3628.64
+ +sus among multiple references. Moreover, we employed SODA_c [10] for measuring the storytelling quality of generated captions. 2) For evaluating localization performance, we reported the average precision, average recall, and their harmonic mean, F1 score across IOU thresholds of $\{0.3, 0.5, 0.7, 0.9\}$ . + +# 4.2. Comparison with State-of-the-art Methods + +Dense Video Captioning Performance. In Table 1 and Table 2, we compare our $\mathrm{E}^2\mathrm{DVC}$ to the state-of-the-art methods for the task of dense video captioning on the ActivityNet Captions and YouCook2 datasets. Compared to baseline PDVC [48], our method achieves substantial gains for three different types of characteristics. From Table 1, our method achieves the best scores over BLEU4, METEOR, CIDEr even compared to the pretrained methods that use extra videos for pretraining. For Table 2, Vid2Seq [53] performs better on YouCook2 dataset than our method. This is because the YouCook2 dataset has fewer videos for training but a more diverse range of action types and semantic descriptions. Vid2seq uses an extra 1 million videos for pretraining, thus having a great advantage over our method. We achieve the best scores among all non-pretrained methods. This validates that by treating each event equally, our method will not overlook those out-of-distribution events and caption them with improved quality. + +Event Localization Performance. In Table 3, we compare our method with the state-of-the-art methods for the event localization task. For the ActivityNet Captions dataset, our method obtains the highest scores on the Precision, Recall, and F1 metrics among all. On the YouCook2 dataset, we achieve the highest performance over Precision and F1 scores even compared with the pretaining method. This is because our initialized queries will focus on each temporal location with high event possibilities. In this way, few + +Table 4. Ablation on the designed components. We report the results on YouCook2. PEI denotes the Pesudo-Event initialization module. EEE denotes the Event-Enhanced encoder module. Without PEI, the queries will be randomly initialized and are updated during the training process. Without EEE, we directly fuse visual features and positional embedding for each video frame together. + +
PEIEEEBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
××1.405.5629.694.9226.81
×✓1.495.7531.155.1328.04
✓×1.565.8830.925.2728.13
✓✓1.686.1134.265.3928.64
+ +Table 5. Ablation on varying number of clusters. We report the results on the Youcook2 dataset. The best and second performance results are highlighted. + +
NcBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.405.5629.694.9226.81
11.455.7431.455.1526.72
21.435.9231.325.1427.66
31.575.9532.575.2727.27
41.515.7830.895.0427.71
51.686.1134.265.3928.64
61.575.9033.275.2628.11
71.295.7830.105.1628.02
+ +Table 6. Ablation on varying temporal event threshold $\tau$ . + +
τBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.325.5629.595.0027.21
11.355.5731.695.1727.51
21.425.6830.255.0827.70
31.595.7931.765.1227.48
41.686.1134.265.3928.64
51.575.8931.505.2228.41
61.515.8529.515.1828.62
71.465.8331.135.2428.24
+ +events will be overlooked. Also, our event-enhanced encoder has the pseudo-event labels as an additional input. This will help the transformer encoder to perceive the visual differences between video frames, thus preventing short-duration videos from being overwhelmed and helping the model obtain more accurate boundaries. + +# 4.3. Ablation Studies + +Analysis of Different Components in $\mathbf{E}^2\mathbf{DVC}$ . As illustrated in Table 4, we present an ablation study to explore the effectiveness of the Pseudo-Event Initialization module (PEI) and the Event-Enhanced Encoder (EEE). PEI initializes decoder queries with pseudo-event locations to ensure equal attention to all potential events, while EEE prevents those short-duration events from being overwhelmed during encoding. The results show that PEI significantly + +![](images/a1f0324556ec6b89d8a8a9b61fd46e61eed3653f2b3b46d822efb1e37c2ead14.jpg) + +![](images/a466ae7a5d9490cd96274fe6d334ec0d478b0ca9a5337b232f260497d9b99ede.jpg) + +![](images/07af83a93ffcf908a6d5286ba5bce1a923df98dd7f674a917381b084b379a745.jpg) +Figure 4. The distribution of the predicted and ground-truth events on ActivityNet Captions. Horizontal and vertical axes represent the normalized center time position and duration respectively. We compare our methods with PDVC [48] and $\mathrm{CM}^2$ [18]. + +![](images/05e369ffb0f1c6faf135e839372b34498a444832212272db88d0c89783941de2.jpg) + +![](images/148091f76c0a704e7dc01431d57251ef21fda19865b7fd7acf8b820ffa4ff66e.jpg) +(a) PDVC + +![](images/f42a45ff867935008c6cc914cfbf79e4a0d3582fb6c037bd68dde8228a73df5d.jpg) +(b) $\mathsf{E}^2\mathsf{DVC}$ + +![](images/b9ace2200c2147fd3fe07a68bdfc97960536900a84d7e4f656d6d2dd7eb7a497.jpg) +(c) GT +Figure 5. Different events duration proportion in ActivityNet Captions. Compared with PDVC, our method $\mathrm{E}^2\mathrm{DVC}$ could identify more short events and the distribution is more close to GT. + +improves both dense video captioning and event localization performance, enabling more precise event detection and better captions by focusing queries on specific pseudo-event locations. Although the improvement from EEE is smaller, it still outperforms the baseline by enhancing event-level knowledge extraction. Combining PEI and EEE yields the best performance, which demonstrates the effectiveness of our proposed components. + +Parameter Analysis on the Number of Clusters $N_{c}$ . We try different numbers of clusters $N_{c}$ on the YouCook2 dataset. Table 5 shows the performance. The first row is the result from the baseline. When the cluster number is set to 1, all video frames are clustered together as a single pseudo-event, where the model only adds one additional query to focus on the middle of the video. In this case, the performance shows a slight improvement compared to the baseline. As the cluster number increases, video frames will be distinguished more finely based on visual differences. The best performance appears when $N_{c}$ is set to 5. When $N_{c}$ equals to the number of video frames, each frame will be allocated a single query to focus on. The model's attention will be distracted and the real events may be overlooked. + +Parameter Analysis on the Event Threshold $\tau$ . The pa + +![](images/c95f0f9b42ee94e2ebd541cbd4d958c55ef749eb6a79de49b0ebfbcf42c63cf9.jpg) +Figure 6. Example visualizations of dense event captioning prediction on ActivityNet Captions. The color of the image border represents the category of the pseudo-event. From top to down, we show the results from the ground truth, the baseline (PDVC) and our method. + +riameter decides the lower-bound of pseudo-events duration. In the Table 6, we presents the results by changing the event threshold $\tau$ in the YouCook2 dataset. When $\tau$ is set too small, the performance drops significantly. It's because even isolated frames are treated as pseudo-events and are allocated certain decoder queries to focus on, which distracts the attention of the model. The best performance appears when $\tau$ is set to 4. As $\tau$ increases, the overall performance decreases. However, the performance of event localization remains at a high level. This is because long-duration events account for the majority and discarding some short-duration events will not hurt performance on long-duration ones. + +# 4.4. Qualitative Results + +Visualization of Event Temporal Distribution. Figure 4 shows the generated event temporal distribution from PDVC [48], $\mathrm{CM}^2$ [18] and our method $\mathrm{E}^2\mathrm{DVC}$ . As can be seen from the figure, our method generates more short-duration events compared with the other two. The distribution predicted by our methods is more similar to the ground truth. We noticed that for long-duration events, we still cannot generate the precise duration. This is because the model is trained by captioning loss and localization loss together. For long-duration events, due to the large number of frames, some localization deviations have little effect on the final captioning performance. Even in some cases, predicting a shorter duration can help in describing the event more delicately, while predicting a longer duration allows for a more comprehensive caption of the event. Therefore, in order to achieve better captioning performance, some localization accuracy may be sacrificed. + +Visualization of Events Duration Statistics. In Figure 5, we present the event duration statistics in the ActivityNet Captions dataset. Our proposed $\mathrm{E}^2\mathrm{DVC}$ outperforms the baseline PDVC by identifying a greater number of short + +duration events. Besides, its event duration distribution is closer to the ground truth. This indicates that $\mathrm{E}^2\mathrm{DVC}$ is more effective at capturing events of varying lengths, particularly short-duration ones. By better aligning with the GT distribution, $\mathrm{E}^2\mathrm{DVC}$ demonstrates improved event localization performance, highlighting its ability to handle a wider range of event durations, leading to more accurate and comprehensive video captioning. + +Visualization of Predicted Example. Figure 6 visualizes an example of dense video captioning predictions of PDVC and our method. Compared with PDVC, our method can localize short-duration events more accurately. This also helps improve the video captioning performance. The reason is that since short-duration events have fewer frames, they are more sensitive to localization deviations. If too many irrelevant frames are localized in the same short events, the description will become inaccurate. + +# 5. Conclusion + +In conclusion, we propose the Event-Equalized Dense Video Captioning framework to address temporal bias and treat all events equally in video captioning. By leveraging a non-trainable Event Perception module with agglomerative hierarchical clustering, we identify pseudo-events and ensure temporal continuity through a refinement mechanism. The Event-Prior Initialization and Event-Prior Encoder modules allocate equal attention to all events, enhancing the model's precision, particularly for short-duration events. Experimental results on ActivityNet Captions and YouCook2 datasets validate the effectiveness of our approach, showing improved dense video captioning performance. $\mathrm{E}^2\mathrm{DVC}$ offers a robust solution for handling diverse event durations and temporal distributions, making it a promising advancement in dense video captioning. + +# 6. Acknowledgement. + +This work was supported in part by the National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, and Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80. + +# References + +[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5 +[3] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2 +[4] Qiang Chen, Xiaokang Chen, Jian Wang, Shan Zhang, Kun Yao, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6633-6642, 2023. 3 +[5] Shizhe Chen, Jia Chen, Qin Jin, and Alexander Hauptmann. Video captioning with guidance of multimodal latent topics. In Proceedings of the 25th ACM international conference on Multimedia, pages 1838-1846, 2017. 1 +[6] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2021. 1, 2 +[7] Chris Ding and Xiaofeng He. Cluster merging and splitting in hierarchical clustering algorithms. In 2002 IEEE International Conference on Data Mining, 2002. Proceedings., pages 139-146. IEEE, 2002. 2 +[8] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[9] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep action proposals for action understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 768-784. Springer, 2016. 2 +[10] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6 + +[11] Jiyang Gao, Kan Chen, and Ram Nevatia. Ctap: Complementary temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 68-83, 2018. 2 +[12] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1 +[13] Bernard Ghanem, Juan Carlos Niebles, Cees Snoek, Fabian Caba Heilbron, Humam Alwassel, Victor Escoria, Ranjay Krishna, Shyamal Buch, and Cuong Duc Dao. The activitynet large-scale activity recognition challenge 2018 summary. arXiv preprint arXiv:1808.03766, 2018. 6 +[14] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 5 +[15] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. arXiv preprint arXiv:2005.08271, 2020. 1, 2, 4 +[16] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 2 +[17] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18198-18208, 2024. 1, 2 +[18] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 1, 2, 5, 6, 7, 8 +[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5 +[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3 +[21] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1 +[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2025. 1 +[23] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7492-7500, 2018. 2 + +[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17958, 2022. 2 +[25] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 2 +[26] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3889-3898, 2019. 2 +[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3 +[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 1 +[29] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3 +[30] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6588-6597, 2019. 1, 2 +[31] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6504-6512, 2017. 2 +[32] K Papinesi. Bleu: A method for automatic evaluation of machine translation. In Proc. 40th Actual Meeting of the Association for Computational Linguistics (ACL), 2002, pages 311-318, 2002. 6 +[33] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8347–8356, 2019. 1 +[34] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 30(8):2617-2633, 2019. 1 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 5 +[36] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event + +captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8908-8917, 2019. 2 +[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5 +[38] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1 +[39] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 4, 5 +[40] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5 +[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6 +[42] Bairui Wang, Lin Ma, Wei Zhang, and Wei Liu. Reconstruction network for video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7622-7631, 2018. 1, 2 +[43] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7190-7198, 2018. 2 +[44] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 1 +[45] Junke Wang, Dongdong Chen, Chong Luo, Bo He, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Omnivid: A generative framework for universal video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18209-18220, 2024. 6 +[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 1, 4, 5 +[47] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1890-1900, 2020. 6 +[48] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 1, 5, 6, 7, 8 + +[49] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based object detection. arXiv preprint arXiv:2109.07107, 3(6), 2021. 3 +[50] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 1 +[51] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 468-483, 2018. 6 +[52] Yifang Xu, Yunzhuo Sun, Benxiang Zhai, Youyao Jia, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 5 +[53] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2, 3, 6 +[54] Dali Yang and Chun Yuan. Hierarchical context encoding for events captioning in videos. In 2018 25th IEEE International Conference on Image Processing (ICIP), pages 1288-1292. IEEE, 2018. 2 +[55] Manyuan Zhang, Guanglu Song, Yu Liu, and Hongsheng Li. Decoupled detr: Spatially disentangling localization and classification for improved end-to-end object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6601-6610, 2023. 3 +[56] Qi Zhang, Yuqing Song, and Qin Jin. Unifying event detection and captioning as sequence generation via pre-training. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 6 +[57] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1 +[58] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2, 5 +[59] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8739-8748, 2018. 5, 6 +[60] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 1, 2 +[61] Wanrong Zhu, Bo Pang, Ashish V Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning + +as sequence generation. arXiv preprint arXiv:2204.08121, 2022.6 +[62] Yuhan Zhu, Guozhen Zhang, Jing Tan, Gangshan Wu, and Limin Wang. Dual detrs for multi-label temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18559-18569, 2024. 3 \ No newline at end of file diff --git a/2025/Event-Equalized Dense Video Captioning/images.zip b/2025/Event-Equalized Dense Video Captioning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d782d7331835440848390245673836835415b05e --- /dev/null +++ b/2025/Event-Equalized Dense Video Captioning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ccdc6f09a1437fc55b20a8ae4922f41c418281d69321de8c656a6325bd67e45 +size 569192 diff --git a/2025/Event-Equalized Dense Video Captioning/layout.json b/2025/Event-Equalized Dense Video Captioning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2019c4b9330410ae560c84fdb16f02e1cbd1cc5 --- /dev/null +++ b/2025/Event-Equalized Dense Video Captioning/layout.json @@ -0,0 +1,9133 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 177, + 103, + 434, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 103, + 434, + 121 + ], + "spans": [ + { + "bbox": [ + 177, + 103, + 434, + 121 + ], + "type": "text", + "content": "Event-Equalized Dense Video Captioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 142, + 492, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 142, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 119, + 142, + 492, + 171 + ], + "type": "text", + "content": "Kangyi Wu* Pengna Li* Jingwen Fu Yizhe Li Yang Wu Yuhan Liu Jinjun Wang Sanping Zhou†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 172, + 509, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 172, + 509, + 213 + ], + "spans": [ + { + "bbox": [ + 126, + 172, + 509, + 213 + ], + "type": "text", + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 216, + 571, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 216, + 571, + 241 + ], + "spans": [ + { + "bbox": [ + 63, + 216, + 571, + 241 + ], + "type": "text", + "content": "{wukangyi747600, sauerfisch, ful371252069, lyz0515, wuyang_CC, liuyuhan200095}@stu.xjtu.edu.cn jinjun@mail.xjtu.edu.cn spzhou@xjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "spans": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 294, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 296, + 570 + ], + "type": "text", + "content": "Dense video captioning aims to localize and caption all events in arbitrary untrimmed videos. Although previous methods have achieved appealing results, they still face the issue of temporal bias, i.e., models tend to focus more on events with certain temporal characteristics. Specifically, 1) the temporal distribution of events in training datasets is uneven. Models trained on these datasets will pay less attention to out-of-distribution events. 2) long-duration events have more frame features than short ones and will attract more attention. To address this, we argue that events, with varying temporal characteristics, should be treated equally when it comes to dense video captioning. Intuitively, different events tend to have distinct visual differences due to varied camera views, backgrounds, or subjects. Inspired by that, we intend to utilize visual features to have an approximate perception of possible events and pay equal attention to them. In this paper, we introduce a simple but effective framework, called Event-Equalized Dense Video Captioning " + }, + { + "bbox": [ + 53, + 294, + 296, + 570 + ], + "type": "inline_equation", + "content": "(E^2 DVC)" + }, + { + "bbox": [ + 53, + 294, + 296, + 570 + ], + "type": "text", + "content": " to overcome the temporal bias and treat all possible events equally. Experimental results on ActivityNet Captions and YouCook2 dataset validate the effectiveness of the proposed methods and show State-of-the-art (SOTA) performance on dense video captioning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "content": "In recent years, an increasing number of works have begun to focus on video understanding[21, 22, 28, 38, 44, 50, 57] where video captioning is a challenging branch. Conventional video captioning [5, 12, 33, 34, 42] aims to generate a natural sentence to describe the main event of a short video clip. However, they face considerable challenges when ap" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 267, + 449, + 367 + ], + "blocks": [ + { + "bbox": [ + 315, + 267, + 449, + 367 + ], + "lines": [ + { + "bbox": [ + 315, + 267, + 449, + 367 + ], + "spans": [ + { + "bbox": [ + 315, + 267, + 449, + 367 + ], + "type": "image", + "image_path": "d226cc13c95c008e8dde0a80b1c84703f77acb24348d907dd3e8b31e680ab12a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 451, + 269, + 542, + 357 + ], + "blocks": [ + { + "bbox": [ + 451, + 269, + 542, + 357 + ], + "lines": [ + { + "bbox": [ + 451, + 269, + 542, + 357 + ], + "spans": [ + { + "bbox": [ + 451, + 269, + 542, + 357 + ], + "type": "image", + "image_path": "e807cdabd3fa7777c0601b09753afbfaf8da729408131892783681e7848a4c4e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 367, + 450, + 464 + ], + "blocks": [ + { + "bbox": [ + 317, + 367, + 450, + 464 + ], + "lines": [ + { + "bbox": [ + 317, + 367, + 450, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 367, + 450, + 464 + ], + "type": "image", + "image_path": "04872f2e16adead48cda3838b466cfe7eeafd9e70f46198e5079566bfc629151.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 466, + 427, + 475 + ], + "lines": [ + { + "bbox": [ + 348, + 466, + 427, + 475 + ], + "spans": [ + { + "bbox": [ + 348, + 466, + 427, + 475 + ], + "type": "text", + "content": "(a) Data Event Distribution" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 466, + 533, + 475 + ], + "lines": [ + { + "bbox": [ + 459, + 466, + 533, + 475 + ], + "spans": [ + { + "bbox": [ + 459, + 466, + 533, + 475 + ], + "type": "text", + "content": "(b) Event Duration Ratio" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 451, + 367, + 542, + 455 + ], + "blocks": [ + { + "bbox": [ + 451, + 367, + 542, + 455 + ], + "lines": [ + { + "bbox": [ + 451, + 367, + 542, + 455 + ], + "spans": [ + { + "bbox": [ + 451, + 367, + 542, + 455 + ], + "type": "image", + "image_path": "7f3f9668bd550a93885f1d6feef72e72f797560689dd1a46ebc0b446340ade0c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 496, + 555, + 595 + ], + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 595 + ], + "type": "text", + "content": "Figure 1. Temporal distribution of ground truth and predicted events on ActivityNet Captions training set. The prediction is from PDVC [48]. (a) The distribution of all events in time dimensions. Horizontal and vertical axes represent the normalized center time position and duration respectively. It's obvious that events are distributed unevenly in these two temporal dimensions. (b) The proportion of the number of events in different duration intervals. It's clear that the predicted short-duration events are much fewer than the ground truth." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 615, + 553, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 553, + 663 + ], + "type": "text", + "content": "plied to realistic videos which are usually long, untrimmed, and contain a variety of events. To address the challenge, dense video captioning (DVC) [19] is proposed to automatically localize and caption all events in an untrimmed video." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Existing DVC methods [6, 17, 18, 30, 60] follow a common pipeline to first extract image features from down-sampled video frames with pre-trained image encoder [15, 35, 39, 46] and then delicately design a framework to ex" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 693, + 192, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 693, + 192, + 703 + ], + "spans": [ + { + "bbox": [ + 73, + 693, + 192, + 703 + ], + "type": "text", + "content": "* These authors contributed equally." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 75, + 704, + 153, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 704, + 153, + 712 + ], + "spans": [ + { + "bbox": [ + 75, + 704, + 153, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8417" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "ploit event representations from them. Although state-of-the-art (SOTA) performance has been achieved, existing methods still face the issue of temporal bias, which means that models will pay more attention to events with certain temporal characteristics while overlooking some other ones. The temporal bias can be divided into two categories based on its causes: 1) Data distribution bias. As shown in Figure 1(a), the predicted events are scattered in the time axes. Events are completely ignored or mis-localized for most timestamps. This is because the models tend to learn from the dataset which part of the video to focus on and the ground truth events are unevenly distributed in this dimension. 2) Event duration bias. This bias results from the duration variation between events. Since short-duration events have fewer frames, their features are easy to be overwhelmed by others. Figure 1(b) shows the proportion of events in different duration intervals. It's obvious that the predicted short-duration events are much fewer than the ground truth. This means that most short-duration events are totally ignored or localized with large deviations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 314, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 469 + ], + "type": "text", + "content": "To address the issues above, we argue that every event, regardless of its temporal characteristics, should share equal importance when it comes to dense video captioning. To achieve this goal, we need to reduce the influence of temporal information and ensure that attention is evenly distributed across all events. However, the challenge lies in how to perceive possible events in a video without relying on temporal information. Intuitively, different events have large visual differences considering their varied camera views, backgrounds, or subjects [19]. Inspired by this, we intend to utilize the differences in pure visual features to gain an initial perception of segments where events may exist. Equal Attention will be allocated to all these segments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": "In this paper, we propose a novel Event-Equalized Dense Video Captioning framework " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "(\\mathrm{E}^2\\mathrm{DVC})" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": " to overcome the temporal bias and treat each potential event equally. First, a non-trainable Event Perception module (EPM) is devised to utilize agglomerative hierarchical clustering algorithm [7] on visual features to generate video frame clusters. These clusters are segments with potential events. A refinement mechanism is designed to split frames that are discontinuous in time and discard isolated frames. These refined clusters are regarded as pseudo-events. Since the clustering process is non-trainable, the result will not be affected by the uneven data distribution. Additionally, since agglomerative hierarchical clustering is a non-uniform method, frames from short-duration events can also be equally clustered into the same category as the long ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "With the pseudo-events, equal attention is required among them. To achieve this, a Pseudo-event Initialization module (PEI) is introduced. Concretely, we calculate the temporal locations of these pseudo-events (center timestamps) and use a temporal encoder to encode them" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "as pseudo-event queries. These queries serve as the initial guess of the event's locations to help the model allocate its attention. A novel Event-Enhanced Encoder (EEE) is also proposed to explore frame-frame and frame-event relationships during the encoding process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "type": "text", + "content": "Our " + }, + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "type": "text", + "content": " can determine and localize events with higher precision which further improves the dense video captioning performance. The main contributions of our work can be summarized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 553, + 335 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "type": "text", + "content": "- We identify the issue of temporal bias and propose an Event-Equalized Dense Video Captioning framework to treat all events equally." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": "- We propose a non-trainable Event Perception module, a Pseudo-Event Initialization module, and an Event-Enhanced Encoder to ensure all events are noticed. Our components could perceive the input video with pure visual features and allocate equal attention to events with large temporal variations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 288, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 335 + ], + "type": "text", + "content": "- Comprehensive experiments have been conducted on the ActivityNet Captions [19] and the YouCook2 [58] datasets to verify the effectiveness of our model in the dense video captioning task." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 346, + 405, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 405, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 405, + 358 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 366, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 569 + ], + "type": "text", + "content": "Dense Video Captioning. Dense video captioning consists of two subtasks: event localization [9, 11, 25, 26] and event captioning [24, 31, 42]. Based on different pipelines, the methods can be divided into two categories: two-stage methods and end-to-end methods. For two-stage methods, they solve dense video captioning by a \"localize-then-describe\" scheme. These methods first predict a set of event proposals with corresponding time boundaries and then generate a detailed sentence description for each proposal by further extracting semantic knowledge and visual context. Krishna et al. [19] propose the first two-stage dense video captioning model. Wang et al. [43] and Yang et al. [54] focus on mining useful contextual information to improve DVC performance. Iashin et al. [15, 16] construct architectures to utilize any number of modalities for event description. However, these methods lack inter-task interactions and heavily rely on hand-crafted components." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "To address the limitations above, recent methods follow the end-to-end pipeline [3, 6, 18, 23, 30, 36, 53, 60]. Li et al. [23] propose a joint and global optimization of detection and captioning. Islam et al. [17] and Deng et al. [6] solve DVC in hierarchical ways where the former follows a bottom-up structure while the latter proposes a top-down framework. Mun et al. [30] and Zhou et al. [60] leverage visual and linguistic context from prior events for coherent storytelling while the latter proposes an additional memory module and streaming decoding algorithm to support streaming video input. Kim et al. [18] utilize external memory to incorporate prior video-to-text knowledge to gener-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8418" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 556, + 277 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 556, + 277 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 556, + 277 + ], + "type": "image", + "image_path": "a0a7923d0bc4a3e36cfd2e2d29283ea3bce9cebf9491e90e55cfcb7222c66924.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 373 + ], + "type": "text", + "content": "Figure 2. The overview of our " + }, + { + "bbox": [ + 54, + 285, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^2\\mathbf{DVC}" + }, + { + "bbox": [ + 54, + 285, + 555, + 373 + ], + "type": "text", + "content": ". Our model follows an encoder-decoder structure. Given an input video, a pre-trained visual encoder is adopted to extract frame-level features. (a) Event Perception Module (EPM). Based on visual features similarity, these frames are clustered to multiple uneven segments and a refinement mechanism is utilized to discard outliers. The remaining segments are assigned with pseudo-events. (b) Event-Enhanced Encoder (EEE). By incorporate the pseudo-event labels with visual frame features, it could enhance the encoder's ability to explore frame-event and event-event relationships, preventing short-duration events from being overwhelmed. (c) Pseudo-Event Initialization module (PEI). Given the pseudo events, we calculate and encode their temporal locations and use the embedding to initialize the queries, helping the model allocate its attention equally among them. Finally, we obtain the predictions of the event-number, time localization and event captions through heads." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 394, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 394, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 296, + 491 + ], + "type": "text", + "content": "ate natural descriptions. Yang et al. [53] reformulate sentence boundaries of transcribed speech as pseudo-events to achieve large-scale video data and propose a multi-modal pertaining model trained on this dataset. However, it remains a significant challenge to detect all events and localize them with high precision in a video. We propose a simple yet effective approach to enable the model to treat all events equally thus improving the DVC performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Query Formulation in Transformers. Since Detection Transformer (DETR)[2] has achieved promising performance in object detection without the need for any handcrafted components, many researchers have studied and further modified the model. Among these, research on decoder query formulation has received significant attention. Conditional DETR [29] learns a conditional spatial query by encoding the reference points as the query position embedding to help focus on distinct regions. Anchor DETR [49] and DAB-DETR [27] utilize anchor boxes to initialize queries to improve interpretability and force each query to only pay attention to a specific region in the object detection task. DN-DETR [20] additionally adds ground-truth bounding boxes with noises into decoder queries to reduce bipartite graph matching difficulty and achieve faster convergence. Group DETR [4] utilizes multiple groups of queries to conduct one-to-many assignments in order to speed up training and improve performance. Decoupled DETR [55] discovers that" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 394, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 526 + ], + "type": "text", + "content": "using shared queries for classification and localization will lead to suboptimal results. Instead, they introduce a task-aware query generation module to decouple these two tasks. DualDETR [62] presents a joint query initialization strategy to align queries from the instance level and boundary level to obtain better performance in the Temporal Action Detection task. In this paper, we introduce the query formulation strategies into the Dense Video Captioning task. Our " + }, + { + "bbox": [ + 313, + 394, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 394, + 555, + 526 + ], + "type": "text", + "content": " employs a novel Pseudo-Event Initialization module to initialize the decoder queries with pseudo-events to ensure that every possible events are treated equally." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 537, + 370, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 370, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 370, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Our goal is to treat each event equally and localize and caption all events without omission. To realize this, we introduce an Event Perception module (Section 3.1) with a delicately devised refinement mechanism to perceive the input video with only visual features available and select those segments where events may exist. Furthermore, A novel Event-Enhanced Encoder (Section 3.2) is also introduced to incorporate the pseudo-event labels into the self-attention framework in order to explore frame-frame and frame-event relationships. Meanwhile, we design an Pseudo Event Initialization module (Section 3.3) to ensure that the model pays equal attention to those pseudo-event locations. The overall framework of " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " is illustrated in Figure 2," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8419" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 298, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 298, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 298, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 298, + 140 + ], + "type": "image", + "image_path": "a6f7daf472a0e4d00fdf068eb72859d42e6572b90ec973e10ecb6d103643070a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 148, + 297, + 215 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 297, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 297, + 215 + ], + "type": "text", + "content": "Figure 3. Illustration of the refinement mechanism. Different shapes denote different events. In the same events, frames with different color are temporal consistent. Our refinement mechanism could split the temporal inconsistent events into multiple new consistent events and discard the extremely short events which is short than the defined temporal event threshold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": "which follows a standard parallel encoder-decoder structure with subtask heads. The objective is to generate " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": " pairs of events and captions " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\{(t_s^n, t_e^n, Cap^n)\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": " denotes the number of events detected in the video. " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "t_s^n" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "t_e^n" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": " denote the start and end timestamps of the " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": "-th event. " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "Cap^n" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": " denotes the corresponding caption. The details of task heads and training loss will be introduced in Section 3.4 and 3.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 198, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 198, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 198, + 335 + ], + "type": "text", + "content": "3.1. Event Perception Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": "To perceive the whole video with only visual features and pick out those pseudo-events, we devise an event perception module with agglomerative hierarchical clustering. The input video is first sampled in a certain ratio to obtain the image frames " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "V = \\{v_{i}\\}_{i=1}^{F}" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " denotes the number of frames. Then, a pre-trained image encoder " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " [15, 35, 39, 46] is utilized to extract visual features " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "X = \\{x_{i}\\}_{i=1}^{F}" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " from each video frame. Based on the visual feature similarities, we perform a hierarchical clustering algorithm to obtain pseudo-events." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": "In this paper, we choose the Agglomerative clustering algorithm because it doesn't assume any particular cluster shape and sizes. It could progressively merge the individual frames to generate uneven clusters with different sizes and assign them with pseudo-event labels. Specifically, each visual feature " + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": " is initially considered an individual cluster with one element. We calculate the pairwise distance matrix among all the clusters. The two clusters that are the most close in feature space are combined into a bigger cluster. Then we recalculate the distance matrix. This procedure is iterated until the final number of clusters is equal to " + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": " which is a hyper-parameter which is set to control the granularity of the clustering process. Finally, we obtain a number of clusters " + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": " which can be formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 630, + 296, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 630, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 110, + 630, + 296, + 648 + ], + "type": "interline_equation", + "content": "\\mathbb {C} = \\left\\{\\mathbb {C} _ {i} \\right\\} _ {i = 1} ^ {N _ {c}} = \\left\\{\\left\\{x _ {i} ^ {j} \\right\\} _ {j = 1} ^ {N _ {c} ^ {i}} \\right\\} _ {i = 1} ^ {N _ {c}}, \\tag {1}", + "image_path": "7ba6dd2851630c75733c057c58638a68f9d9dee5888100532533b476d30c39a5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{C}_i" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": "-th cluster and " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "N_c^i" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " denotes the number of this cluster." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Additionally, since the timestamps of the frame features are not considered during the clustering process, some clustered frames are temporally inconsistent with others in the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "same cluster. Considering that frames belonging to the same event should be temporally consistent, as illustrated in Figure 3, we utilize a refinement mechanism to split the temporally discontinuous events into multiple new consistent events based on the timestamps of the video frames. Moreover, since an event typically occupies a certain duration, we devise a temporal event threshold " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ". If the durations of the events are shorter than " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", these events would be considered outliers and discarded. The remains are regarded as final pseudo-events. Our framework will pay special attention to these pseudo-event locations with improved localization and captioning performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 224, + 458, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 458, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 458, + 236 + ], + "type": "text", + "content": "3.2. Event-Enhanced Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 350 + ], + "type": "text", + "content": "After obtaining the pseudo events, we directly assign the video frames with their pseudo labels. In this section, we introduce how to incorporate the assigned labels with visual features into transformer encoder to get better representations. We hope that in the process of encoding frame features, the focus should not only be on the frame-frame relationships but also on the frame-event one. In this way, frame features from short-duration events will not be overwhelmed during the self-attention process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": "To realize this, we devise an Event-Enhanced Encoder (EEE) to add pseudo-event labels to the frame feature encoding process. Specifically, for each video frame " + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": ", we have its pseudo-event label " + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": ". The pseudo-event label " + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": " is transformed into a label embedding " + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "inline_equation", + "content": "le_{i}" + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": " with a label encoder, which consists of a learnable dictionary and a multilayer perceptron. During the encoding process, we concatenate the frame-level features " + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "inline_equation", + "content": "X = \\{x_{i}\\}_{i=1}^{F}" + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": " with their corresponding label embedding and positional embedding (PE). They are fed into the deformable transformer encoder and a multi-scale convolution layers to extract multi-scale features with frame-frame and frame-event relationships." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 465, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 465, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 465, + 514 + ], + "type": "text", + "content": "3.3. Pseudo-Event Initialization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "content": "Given the pseudo-events from Event Perception module, we want the model to pay equal attention to their temporal locations. Previously, the decoder aimed to query the event-level features from all frame features directly with " + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "content": " randomly initialized and learnable event queries. In this way, each event query is associated with multiple random temporal locations and learns which part to focus on from the uneven training dataset, leading to those events with out-of-distribution temporal characteristics easy to be overlooked." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "To address this issue, we devise a novel Pseudo-Event Initialization (PEI) module to initialize those event queries with pseudo-events. Specifically, given a pseudo-event " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "E_{i} = \\{x_{i,j}\\}_{j = 1}^{N_{i}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": " is the number of frames for " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": ", we calculate its center temporal location " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": " (reference points) with:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 390, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 390, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "t _ {i} = \\left(t _ {i, 1} + t _ {i, L _ {i}}\\right) / 2, \\tag {2}", + "image_path": "f996984d9185cdb595668ab4614507b42d3e93fb618df177d3a1ee4968f36252.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8420" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "t_{i,j}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " is the timestamp for frame feature " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "x_{i,j}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". In this way, for each pseudo-event " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", we have its reference points " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". We calculate the inverse sigmoid " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "(Inv_{Sig})" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " and use temporal encoder " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " [40] to get the temporal embedding. Instead of randomly initializing the event queries, we initialize them with the positional embeddings of all pseudo-events. The process can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 162, + 294, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 294, + 175 + ], + "type": "interline_equation", + "content": "q _ {i} = f _ {t} \\left(\\operatorname {I n v} _ {\\operatorname {S i g}} \\left(t _ {i}\\right)\\right), \\tag {3}", + "image_path": "0c3ede0a118174ea7b503ed1f2dbbc27b949f440141ba8caa3db640e3bbd450e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "text", + "content": "-th event query. What's more, considering that the pseudo-events may not cover all possible events in the ground truth, we also concatenate " + }, + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 55, + 179, + 296, + 276 + ], + "type": "text", + "content": " randomly initialized queries with these pseudo-event queries together. In this way, the event queries explicitly prioritize the locations of these pseudo-events. This significantly enhances the performance of event localization and further improves the event captioning performance as well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 133, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 133, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 133, + 293 + ], + "type": "text", + "content": "3.4. Task Heads" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 295, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 347 + ], + "type": "text", + "content": "Our work follows an encoder-decoder framework with a parallel decoding structure [18, 48, 52]. Given the event features, three parallel prediction heads are utilized to do dense video captioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "text", + "content": "Localization Head. The localization head is responsible for box prediction and binary classification for each event query " + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "text", + "content": ". It outputs a set of tuples " + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\{t_s^i, t_e^i, c^i\\}_{i=1}^{N_q}" + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "inline_equation", + "content": "t_s^i" + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "inline_equation", + "content": "t_e^i" + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "text", + "content": " denote the start and end timestamps and " + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "inline_equation", + "content": "c^i" + }, + { + "bbox": [ + 55, + 348, + 296, + 408 + ], + "type": "text", + "content": " represents the localization confidence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": "Captioning Head. We utilize LSTM [14] as the backbone of our captioning head. At each timestamp " + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": ", we feed context features " + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "inline_equation", + "content": "z_{i,t}" + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": ", event query features " + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": " and previous words " + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\{w_{i,j}\\}_{j = 1}^{t - 1}" + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": " into the LSTM model to get the current word " + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "inline_equation", + "content": "w_{i,t}" + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": ". As the sentence is generated, the captioning head produces the complete sentence " + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "inline_equation", + "content": "S_{i} = w_{i,1},\\dots,w_{l,T}" + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 408, + 296, + 491 + ], + "type": "text", + "content": " denotes the length of the sentence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": "Event Counter. It aims to detect the number of events in the given video. It first compresses the most salient information of event queries " + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": " to a global feature vector with a max-pooling layer and predicts a fixed-size vector " + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "inline_equation", + "content": "f_{count}" + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": " with an FC layer with softmax activation. In " + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "inline_equation", + "content": "f_{count}" + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": ", each value refers to the possibility of a specific number. During the inference stage, the event number " + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "inline_equation", + "content": "N_{event}" + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "inline_equation", + "content": "N_{event} = \\arg\\max(f_{count})" + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": ". The final outputs are obtained by selecting " + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "inline_equation", + "content": "N_{event}" + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": " events with the highest confidence score " + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 492, + 296, + 612 + ], + "type": "text", + "content": " which is calculated by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 616, + 295, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 616, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 111, + 616, + 295, + 649 + ], + "type": "interline_equation", + "content": "c _ {i} = c _ {i} ^ {l o c} + \\frac {\\mu}{M _ {i} ^ {\\gamma}} \\sum_ {t = 1} ^ {M _ {i}} \\log \\left(c _ {i, t} ^ {c a p}\\right), \\tag {4}", + "image_path": "cbdab82dc6be7d33c641886e94b269ee8ecb2135d42422d948bc4de092bad57f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is the number of words for the " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "-th event caption. " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "C_{i}^{loc}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is the localization confidence for the " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "-th event. " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{i,t}^{cap}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is the confidence score for the " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "-th word for the " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "-th event. " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " are two hyper-parameters to control the influence of the captioning confidence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 403, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 403, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 403, + 83 + ], + "type": "text", + "content": "3.5. Loss Function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "text", + "content": "Suppose " + }, + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "text", + "content": " events are predicted with their localizations and captions, Hungarian algorithm [2] is utilized to find the best bipartite matching results between the predicted events and the ground truth, during which the generalized IOU [37] between the predicted temporal segments and ground-truth segments is calculated as " + }, + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "inline_equation", + "content": "L_{giou}" + }, + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "text", + "content": " and the focal loss between the predicted classification score and the ground-truth label is calculated as " + }, + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 313, + 91, + 555, + 222 + ], + "type": "text", + "content": ". The matched event pairs are selected to calculate the set prediction loss, which is the weighted sum of gIOU loss, classification loss, countering loss, and caption loss. The loss function is formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 233, + 555, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 233, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 323, + 233, + 555, + 247 + ], + "type": "interline_equation", + "content": "L = \\alpha_ {g i o u} L _ {g i o u} + \\alpha_ {c l s} L _ {c l s} + \\alpha_ {e c} L _ {e c} + \\alpha_ {c a p} L _ {c a p}, \\tag {5}", + "image_path": "4c15ca7c8a89b8349c6728db39380c9ae8e16d508f14775e40502119e889cab0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "inline_equation", + "content": "L_{cap}" + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": " is the cross-entropy loss between the predicted word probability and the ground truth normalized by the caption length " + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "inline_equation", + "content": "L_{ec}" + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": " is the cross-entropy loss between the predicted count distribution and the ground truth. More details can be found in Supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 327, + 395, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 395, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 395, + 341 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 347, + 430, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 430, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 430, + 361 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "type": "text", + "content": "Datasets. We evaluate our method on the two commonly used dense video captioning benchmark datasets: ActivityNet Captions [19] and YouCook2 [58]. ActivityNet Captions includes over 20000 long videos covering a wide range of human daily activities and actions. Each video spans about 120s and is annotated with 3.7 temporally localized sentences on average. YouCook2 consists of 2000 long untrimmed videos from 89 cooking recipes. Generally, each video lasts approximately 320s and has 7.7 annotated segments with the corresponding sentences. For both datasets, we adhere to the standard split for training, validation, and testing. It's worth noting that we only use those videos that are still online on YouTube, leading to roughly " + }, + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "type": "text", + "content": " less videos compared to the original datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Implementation Details. For both datasets, we sample video frames at a rate of 1 FPS (one frame per second) and subsample or pad the sequence of frames to reach a total of F frames, where F is set to 100 for ActivityNet Captions and 200 for YouCook2. We use CLIP ViT-L/14 [8, 35] to extract clip visual features for video frames [18]. To further investigate the generalization ability of our model to different frame-level features, we also evaluate our model based on C3D [39] features provided by PDVC [48] and TSN [46] features provided by [59]. We utilize a two-layer deformable transformer that incorporates multi-scale deformable attention across four levels. We adopt the Agglomerative Hierarchical cluster algorithm for generating pseudo events. For ActivityNet Captions, we set the number of event queries to 10, while for YouCook2, we set it" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8421" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 145, + 294, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art methods for dense video captioning in ActivityNet Captions. B4/M/C is short for BLEU4/Meteoror/CIDEr. * indicates the results of our implementation using the official code. The best performance results are highlighted for each feature type. We also present results from those pretrained models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 145, + 294, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 145, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 145, + 294, + 357 + ], + "type": "table", + "html": "
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
UEDVC [56]C3D✓---5.50
Vid2Seq [53]CLIP✓-8.5030.105.80
OmniVID [45]-✓1.737.5426.005.60
ECHR [47]C3Dx1.827.2014.73.20
E2ESG [61]C3Dx-3.5025.00-
PDVC [48]C3Dx1.657.5025.875.26
E²DVCC3Dx1.797.5426.835.32
MT [59]TSNx0.303.2014.70-
PDVC [48]TSNx1.787.9628.965.44
E²DVCTSNx2.038.0229.915.77
CM² [18]CLIPx2.388.5533.016.18
PDVC* [48]CLIPx2.218.0629.975.92
E²DVCCLIPx2.438.5733.636.13
", + "image_path": "88c2ac469e3a568b7d1a119c301a6202b9e0180ad5f04db8f5bb610b1a392ba7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 414, + 294, + 542 + ], + "blocks": [ + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "type": "text", + "content": "Table 2. Comparison with the state-of-the-art methods for dense video captioning on Youcook2. Since C3D features for YouCook2 are not publicly available, we fail to provide the results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 414, + 294, + 542 + ], + "lines": [ + { + "bbox": [ + 58, + 414, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 58, + 414, + 294, + 542 + ], + "type": "table", + "html": "
MethodFeaturesPretrainB4↑M↑C↑SODA_c↑
Vid2Seq [53]CLIP✓-9.3047.107.90
MT [59]TSN×1.155.009.30-
PDVC [48]TSN×0.804.7522.714.42
E²DVCTSN×1.054.7623.154.31
CM² [18]CLIP×1.636.0831.665.34
PDVC* [48]CLIP×1.405.5629.694.92
E²DVCCLIP×1.686.1134.265.39
", + "image_path": "fe82df02dd7a3bfae34e1bd597484b85c331258a2436be7ac718600771800aef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "text", + "content": "to 100. For the event counter, we select a maximum count of 10 for the ActivityNet Captions dataset and 20 for the YouCook2 dataset. We set " + }, + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "text", + "content": " to 5 and " + }, + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "text", + "content": " to 4. Other model hyperparameters follow the recipe of PDVC [48]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate our model in two aspects: 1) To assess dense captioning performance, we adopt the official evaluation tool from ActivityNet Challenge [13], which use the metrics BLEU4 [32], METEOR [1], and CIDEr [41] to compute the average precision of matched pairs between generated captions and the ground truth. These scores are averaged over IOU at " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{0.3, 0.5, 0.7, 0.9\\}" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": ". BLEU4 focuses on n-gram precision, METEOR incorporates semantic matching, and CIDEr emphasizes consen" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 124, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 114 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 114 + ], + "type": "text", + "content": "Table 3. Comparison to state-of-the-art methods for event Localization. We report the results on ActivityNet Captions and Youcook2 datasets. PT denotes Pretrain. Pre./Rec. is short for Precision/Recall." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 124, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 317, + 124, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 124, + 553, + 281 + ], + "type": "table", + "html": "
MethodFeatures PTActivityNetYouCook2
Pre.↑Rec.↑F1↑Pre.↑Rec.↑F1↑
Vid2Seq [53]CLIP✓53.9052.7053.2927.8027.9027.84
E2ESG [61]C3D✘---20.6020.7020.65
MFT [51]TSN✘51.4124.3133.01---
CM² [18]CLIP✘56.8153.7155.2133.3824.7628.43
PDVC* [48]CLIP✘56.3853.2754.7832.3722.8926.81
E²DVCC3D✘57.7154.7056.17---
E²DVCTSN✘57.7755.1456.4234.1325.0128.87
E²DVCCLIP✘57.7054.6756.1434.7524.3628.64
", + "image_path": "82dec359702b46080d0de375be3d8a92021cadc740febea70cf733b1f8621fc8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 291, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 365 + ], + "type": "text", + "content": "sus among multiple references. Moreover, we employed SODA_c [10] for measuring the storytelling quality of generated captions. 2) For evaluating localization performance, we reported the average precision, average recall, and their harmonic mean, F1 score across IOU thresholds of " + }, + { + "bbox": [ + 313, + 291, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\{0.3, 0.5, 0.7, 0.9\\}" + }, + { + "bbox": [ + 313, + 291, + 555, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 371, + 537, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 537, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 537, + 384 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 390, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 605 + ], + "type": "text", + "content": "Dense Video Captioning Performance. In Table 1 and Table 2, we compare our " + }, + { + "bbox": [ + 313, + 390, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 390, + 555, + 605 + ], + "type": "text", + "content": " to the state-of-the-art methods for the task of dense video captioning on the ActivityNet Captions and YouCook2 datasets. Compared to baseline PDVC [48], our method achieves substantial gains for three different types of characteristics. From Table 1, our method achieves the best scores over BLEU4, METEOR, CIDEr even compared to the pretrained methods that use extra videos for pretraining. For Table 2, Vid2Seq [53] performs better on YouCook2 dataset than our method. This is because the YouCook2 dataset has fewer videos for training but a more diverse range of action types and semantic descriptions. Vid2seq uses an extra 1 million videos for pretraining, thus having a great advantage over our method. We achieve the best scores among all non-pretrained methods. This validates that by treating each event equally, our method will not overlook those out-of-distribution events and caption them with improved quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Event Localization Performance. In Table 3, we compare our method with the state-of-the-art methods for the event localization task. For the ActivityNet Captions dataset, our method obtains the highest scores on the Precision, Recall, and F1 metrics among all. On the YouCook2 dataset, we achieve the highest performance over Precision and F1 scores even compared with the pretaining method. This is because our initialized queries will focus on each temporal location with high event possibilities. In this way, few" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8422" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 145, + 294, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 137 + ], + "type": "text", + "content": "Table 4. Ablation on the designed components. We report the results on YouCook2. PEI denotes the Pesudo-Event initialization module. EEE denotes the Event-Enhanced encoder module. Without PEI, the queries will be randomly initialized and are updated during the training process. Without EEE, we directly fuse visual features and positional embedding for each video frame together." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 145, + 294, + 218 + ], + "lines": [ + { + "bbox": [ + 58, + 145, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 58, + 145, + 294, + 218 + ], + "type": "table", + "html": "
PEIEEEBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
××1.405.5629.694.9226.81
×✓1.495.7531.155.1328.04
✓×1.565.8830.925.2728.13
✓✓1.686.1134.265.3928.64
", + "image_path": "cf625584407ea466224cb531ff46f64ec779bbb0482b275557e6fa599d4a8ff9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 270, + 293, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 228, + 295, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 262 + ], + "type": "text", + "content": "Table 5. Ablation on varying number of clusters. We report the results on the Youcook2 dataset. The best and second performance results are highlighted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 270, + 293, + 376 + ], + "lines": [ + { + "bbox": [ + 58, + 270, + 293, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 270, + 293, + 376 + ], + "type": "table", + "html": "
NcBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.405.5629.694.9226.81
11.455.7431.455.1526.72
21.435.9231.325.1427.66
31.575.9532.575.2727.27
41.515.7830.895.0427.71
51.686.1134.265.3928.64
61.575.9033.275.2628.11
71.295.7830.105.1628.02
", + "image_path": "da870b60a2153e316086358e92bada0467ee251a481757ff44602fd39b08ae91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 402, + 293, + 510 + ], + "blocks": [ + { + "bbox": [ + 64, + 382, + 286, + 393 + ], + "lines": [ + { + "bbox": [ + 64, + 382, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 64, + 382, + 286, + 393 + ], + "type": "text", + "content": "Table 6. Ablation on varying temporal event threshold " + }, + { + "bbox": [ + 64, + 382, + 286, + 393 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 64, + 382, + 286, + 393 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 402, + 293, + 510 + ], + "lines": [ + { + "bbox": [ + 58, + 402, + 293, + 510 + ], + "spans": [ + { + "bbox": [ + 58, + 402, + 293, + 510 + ], + "type": "table", + "html": "
τBLEU4↑METEOR↑CIDEr↑SODA_c↑F1↑
01.325.5629.595.0027.21
11.355.5731.695.1727.51
21.425.6830.255.0827.70
31.595.7931.765.1227.48
41.686.1134.265.3928.64
51.575.8931.505.2228.41
61.515.8529.515.1828.62
71.465.8331.135.2428.24
", + "image_path": "990419524d1c00fc27c97111cc3df2876621edb6311180fa9d1d624292724328.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 591 + ], + "type": "text", + "content": "events will be overlooked. Also, our event-enhanced encoder has the pseudo-event labels as an additional input. This will help the transformer encoder to perceive the visual differences between video frames, thus preventing short-duration videos from being overwhelmed and helping the model obtain more accurate boundaries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 599, + 156, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 156, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 156, + 611 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Analysis of Different Components in " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^2\\mathbf{DVC}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". As illustrated in Table 4, we present an ablation study to explore the effectiveness of the Pseudo-Event Initialization module (PEI) and the Event-Enhanced Encoder (EEE). PEI initializes decoder queries with pseudo-event locations to ensure equal attention to all potential events, while EEE prevents those short-duration events from being overwhelmed during encoding. The results show that PEI significantly" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 339, + 69, + 434, + 148 + ], + "blocks": [ + { + "bbox": [ + 339, + 69, + 434, + 148 + ], + "lines": [ + { + "bbox": [ + 339, + 69, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 339, + 69, + 434, + 148 + ], + "type": "image", + "image_path": "a1f0324556ec6b89d8a8a9b61fd46e61eed3653f2b3b46d822efb1e37c2ead14.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 530, + 148 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 530, + 148 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 530, + 148 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 530, + 148 + ], + "type": "image", + "image_path": "a466ae7a5d9490cd96274fe6d334ec0d478b0ca9a5337b232f260497d9b99ede.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 339, + 150, + 434, + 228 + ], + "blocks": [ + { + "bbox": [ + 339, + 150, + 434, + 228 + ], + "lines": [ + { + "bbox": [ + 339, + 150, + 434, + 228 + ], + "spans": [ + { + "bbox": [ + 339, + 150, + 434, + 228 + ], + "type": "image", + "image_path": "07af83a93ffcf908a6d5286ba5bce1a923df98dd7f674a917381b084b379a745.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 554, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 281 + ], + "type": "text", + "content": "Figure 4. The distribution of the predicted and ground-truth events on ActivityNet Captions. Horizontal and vertical axes represent the normalized center time position and duration respectively. We compare our methods with PDVC [48] and " + }, + { + "bbox": [ + 313, + 236, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{CM}^2" + }, + { + "bbox": [ + 313, + 236, + 554, + 281 + ], + "type": "text", + "content": " [18]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 434, + 150, + 529, + 228 + ], + "blocks": [ + { + "bbox": [ + 434, + 150, + 529, + 228 + ], + "lines": [ + { + "bbox": [ + 434, + 150, + 529, + 228 + ], + "spans": [ + { + "bbox": [ + 434, + 150, + 529, + 228 + ], + "type": "image", + "image_path": "05e369ffb0f1c6faf135e839372b34498a444832212272db88d0c89783941de2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 295, + 397, + 364 + ], + "blocks": [ + { + "bbox": [ + 328, + 295, + 397, + 364 + ], + "lines": [ + { + "bbox": [ + 328, + 295, + 397, + 364 + ], + "spans": [ + { + "bbox": [ + 328, + 295, + 397, + 364 + ], + "type": "image", + "image_path": "148091f76c0a704e7dc01431d57251ef21fda19865b7fd7acf8b820ffa4ff66e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 365, + 376, + 373 + ], + "lines": [ + { + "bbox": [ + 348, + 365, + 376, + 373 + ], + "spans": [ + { + "bbox": [ + 348, + 365, + 376, + 373 + ], + "type": "text", + "content": "(a) PDVC" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 403, + 297, + 469, + 362 + ], + "blocks": [ + { + "bbox": [ + 403, + 297, + 469, + 362 + ], + "lines": [ + { + "bbox": [ + 403, + 297, + 469, + 362 + ], + "spans": [ + { + "bbox": [ + 403, + 297, + 469, + 362 + ], + "type": "image", + "image_path": "f42a45ff867935008c6cc914cfbf79e4a0d3582fb6c037bd68dde8228a73df5d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 365, + 452, + 373 + ], + "lines": [ + { + "bbox": [ + 422, + 365, + 452, + 373 + ], + "spans": [ + { + "bbox": [ + 422, + 365, + 452, + 373 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 422, + 365, + 452, + 373 + ], + "type": "inline_equation", + "content": "\\mathsf{E}^2\\mathsf{DVC}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 474, + 297, + 542, + 363 + ], + "blocks": [ + { + "bbox": [ + 474, + 297, + 542, + 363 + ], + "lines": [ + { + "bbox": [ + 474, + 297, + 542, + 363 + ], + "spans": [ + { + "bbox": [ + 474, + 297, + 542, + 363 + ], + "type": "image", + "image_path": "b9ace2200c2147fd3fe07a68bdfc97960536900a84d7e4f656d6d2dd7eb7a497.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 365, + 518, + 373 + ], + "lines": [ + { + "bbox": [ + 500, + 365, + 518, + 373 + ], + "spans": [ + { + "bbox": [ + 500, + 365, + 518, + 373 + ], + "type": "text", + "content": "(c) GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 382, + 554, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 415 + ], + "type": "text", + "content": "Figure 5. Different events duration proportion in ActivityNet Captions. Compared with PDVC, our method " + }, + { + "bbox": [ + 313, + 382, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 382, + 554, + 415 + ], + "type": "text", + "content": " could identify more short events and the distribution is more close to GT." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 434, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 530 + ], + "type": "text", + "content": "improves both dense video captioning and event localization performance, enabling more precise event detection and better captions by focusing queries on specific pseudo-event locations. Although the improvement from EEE is smaller, it still outperforms the baseline by enhancing event-level knowledge extraction. Combining PEI and EEE yields the best performance, which demonstrates the effectiveness of our proposed components." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "text", + "content": "Parameter Analysis on the Number of Clusters " + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "text", + "content": ". We try different numbers of clusters " + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "text", + "content": " on the YouCook2 dataset. Table 5 shows the performance. The first row is the result from the baseline. When the cluster number is set to 1, all video frames are clustered together as a single pseudo-event, where the model only adds one additional query to focus on the middle of the video. In this case, the performance shows a slight improvement compared to the baseline. As the cluster number increases, video frames will be distinguished more finely based on visual differences. The best performance appears when " + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "text", + "content": " is set to 5. When " + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 532, + 555, + 699 + ], + "type": "text", + "content": " equals to the number of video frames, each frame will be allocated a single query to focus on. The model's attention will be distracted and the real events may be overlooked." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Parameter Analysis on the Event Threshold " + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": ". The pa" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8423" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 230 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 230 + ], + "type": "image", + "image_path": "c95f0f9b42ee94e2ebd541cbd4d958c55ef749eb6a79de49b0ebfbcf42c63cf9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 270 + ], + "type": "text", + "content": "Figure 6. Example visualizations of dense event captioning prediction on ActivityNet Captions. The color of the image border represents the category of the pseudo-event. From top to down, we show the results from the ground truth, the baseline (PDVC) and our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "text", + "content": "riameter decides the lower-bound of pseudo-events duration. In the Table 6, we presents the results by changing the event threshold " + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "text", + "content": " in the YouCook2 dataset. When " + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "text", + "content": " is set too small, the performance drops significantly. It's because even isolated frames are treated as pseudo-events and are allocated certain decoder queries to focus on, which distracts the attention of the model. The best performance appears when " + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "text", + "content": " is set to 4. As " + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 286, + 294, + 430 + ], + "type": "text", + "content": " increases, the overall performance decreases. However, the performance of event localization remains at a high level. This is because long-duration events account for the majority and discarding some short-duration events will not hurt performance on long-duration ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 167, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 167, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 167, + 453 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 460, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 664 + ], + "type": "text", + "content": "Visualization of Event Temporal Distribution. Figure 4 shows the generated event temporal distribution from PDVC [48], " + }, + { + "bbox": [ + 56, + 460, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{CM}^2" + }, + { + "bbox": [ + 56, + 460, + 295, + 664 + ], + "type": "text", + "content": " [18] and our method " + }, + { + "bbox": [ + 56, + 460, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 56, + 460, + 295, + 664 + ], + "type": "text", + "content": ". As can be seen from the figure, our method generates more short-duration events compared with the other two. The distribution predicted by our methods is more similar to the ground truth. We noticed that for long-duration events, we still cannot generate the precise duration. This is because the model is trained by captioning loss and localization loss together. For long-duration events, due to the large number of frames, some localization deviations have little effect on the final captioning performance. Even in some cases, predicting a shorter duration can help in describing the event more delicately, while predicting a longer duration allows for a more comprehensive caption of the event. Therefore, in order to achieve better captioning performance, some localization accuracy may be sacrificed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Visualization of Events Duration Statistics. In Figure 5, we present the event duration statistics in the ActivityNet Captions dataset. Our proposed " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": " outperforms the baseline PDVC by identifying a greater number of short" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 286, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 382 + ], + "type": "text", + "content": "duration events. Besides, its event duration distribution is closer to the ground truth. This indicates that " + }, + { + "bbox": [ + 313, + 286, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 286, + 555, + 382 + ], + "type": "text", + "content": " is more effective at capturing events of varying lengths, particularly short-duration ones. By better aligning with the GT distribution, " + }, + { + "bbox": [ + 313, + 286, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 286, + 555, + 382 + ], + "type": "text", + "content": " demonstrates improved event localization performance, highlighting its ability to handle a wider range of event durations, leading to more accurate and comprehensive video captioning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 384, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 492 + ], + "type": "text", + "content": "Visualization of Predicted Example. Figure 6 visualizes an example of dense video captioning predictions of PDVC and our method. Compared with PDVC, our method can localize short-duration events more accurately. This also helps improve the video captioning performance. The reason is that since short-duration events have fewer frames, they are more sensitive to localization deviations. If too many irrelevant frames are localized in the same short events, the description will become inaccurate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "In conclusion, we propose the Event-Equalized Dense Video Captioning framework to address temporal bias and treat all events equally in video captioning. By leveraging a non-trainable Event Perception module with agglomerative hierarchical clustering, we identify pseudo-events and ensure temporal continuity through a refinement mechanism. The Event-Prior Initialization and Event-Prior Encoder modules allocate equal attention to all events, enhancing the model's precision, particularly for short-duration events. Experimental results on ActivityNet Captions and YouCook2 datasets validate the effectiveness of our approach, showing improved dense video captioning performance. " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{DVC}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " offers a robust solution for handling diverse event durations and temporal distributions, making it a promising advancement in dense video captioning." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8424" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "type": "text", + "content": "6. Acknowledgement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 145 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, and Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 186, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 294, + 240 + ], + "type": "text", + "content": "[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 242, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 242, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 295, + 293 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 294, + 339 + ], + "type": "text", + "content": "[3] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 294, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 406 + ], + "type": "text", + "content": "[4] Qiang Chen, Xiaokang Chen, Jian Wang, Shan Zhang, Kun Yao, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6633-6642, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 407, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 294, + 449 + ], + "type": "text", + "content": "[5] Shizhe Chen, Jia Chen, Qin Jin, and Alexander Hauptmann. Video captioning with guidance of multimodal latent topics. In Proceedings of the 25th ACM international conference on Multimedia, pages 1838-1846, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 450, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 450, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 450, + 294, + 503 + ], + "type": "text", + "content": "[6] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 505, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 294, + 548 + ], + "type": "text", + "content": "[7] Chris Ding and Xiaofeng He. Cluster merging and splitting in hierarchical clustering algorithms. In 2002 IEEE International Conference on Data Mining, 2002. Proceedings., pages 139-146. IEEE, 2002. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 549, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 549, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 549, + 294, + 581 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 582, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 294, + 647 + ], + "type": "text", + "content": "[9] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep action proposals for action understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 768-784. Springer, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "type": "text", + "content": "[10] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[11] Jiyang Gao, Kan Chen, and Ram Nevatia. Ctap: Complementary temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 68-83, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[12] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 216 + ], + "type": "text", + "content": "[13] Bernard Ghanem, Juan Carlos Niebles, Cees Snoek, Fabian Caba Heilbron, Humam Alwassel, Victor Escoria, Ranjay Krishna, Shyamal Buch, and Cuong Duc Dao. The activitynet large-scale activity recognition challenge 2018 summary. arXiv preprint arXiv:1808.03766, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 237 + ], + "type": "text", + "content": "[14] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[15] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. arXiv preprint arXiv:2005.08271, 2020. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "type": "text", + "content": "[16] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 317, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 380 + ], + "type": "text", + "content": "[17] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18198-18208, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 447 + ], + "type": "text", + "content": "[18] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "text", + "content": "[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "type": "text", + "content": "[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 548, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 613 + ], + "type": "text", + "content": "[21] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "type": "text", + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2025. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[23] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7492-7500, 2018. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8425" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "text", + "content": "[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17958, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 139, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 294, + 183 + ], + "type": "text", + "content": "[25] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[26] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3889-3898, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 327 + ], + "type": "text", + "content": "[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "text", + "content": "[29] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[30] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6588-6597, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[31] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6504-6512, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[32] K Papinesi. Bleu: A method for automatic evaluation of machine translation. In Proc. 40th Actual Meeting of the Association for Computational Linguistics (ACL), 2002, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 570 + ], + "type": "text", + "content": "[33] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8347–8356, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "text", + "content": "[34] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 30(8):2617-2633, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8908-8917, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 228 + ], + "type": "text", + "content": "[38] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "text", + "content": "[39] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 304 + ], + "type": "text", + "content": "[40] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "text", + "content": "[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 555, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 393 + ], + "type": "text", + "content": "[42] Bairui Wang, Lin Ma, Wei Zhang, and Wei Liu. Reconstruction network for video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7622-7631, 2018. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[43] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7190-7198, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "text", + "content": "[44] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 505, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 559 + ], + "type": "text", + "content": "[45] Junke Wang, Dongdong Chen, Chong Luo, Bo He, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Omnivid: A generative framework for universal video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18209-18220, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 1, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[47] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1890-1900, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "type": "text", + "content": "[48] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8426" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[49] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based object detection. arXiv preprint arXiv:2109.07107, 3(6), 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "type": "text", + "content": "[50] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 295, + 207 + ], + "type": "text", + "content": "[51] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 468-483, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "type": "text", + "content": "[52] Yifang Xu, Yunzhuo Sun, Benxiang Zhai, Youyao Jia, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 264, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 295, + 330 + ], + "type": "text", + "content": "[53] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 332, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 332, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 57, + 332, + 295, + 374 + ], + "type": "text", + "content": "[54] Dali Yang and Chun Yuan. Hierarchical context encoding for events captioning in videos. In 2018 25th IEEE International Conference on Image Processing (ICIP), pages 1288-1292. IEEE, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 376, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 295, + 432 + ], + "type": "text", + "content": "[55] Manyuan Zhang, Guanglu Song, Yu Liu, and Hongsheng Li. Decoupled detr: Spatially disentangling localization and classification for improved end-to-end object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6601-6610, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 434, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 295, + 476 + ], + "type": "text", + "content": "[56] Qi Zhang, Yuqing Song, and Qin Jin. Unifying event detection and captioning as sequence generation via pre-training. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 478, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 478, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 478, + 295, + 533 + ], + "type": "text", + "content": "[57] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 295, + 578 + ], + "type": "text", + "content": "[58] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "type": "text", + "content": "[59] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8739-8748, 2018. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[60] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[61] Wanrong Zhu, Bo Pang, Ashish V Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 150 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "as sequence generation. arXiv preprint arXiv:2204.08121, 2022.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[62] Yuhan Zhu, Guozhen Zhang, Jing Tan, Gangshan Wu, and Limin Wang. Dual detrs for multi-label temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18559-18569, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8427" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_content_list.json b/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2969900ebb64122543a94e089eeaba07336a17fb --- /dev/null +++ b/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_content_list.json @@ -0,0 +1,2369 @@ +[ + { + "type": "text", + "text": "Event-based Video Super-Resolution via State Space Models", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeyu Xiao Xinchao Wang*", + "bbox": [ + 377, + 181, + 622, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National University of Singapore", + "bbox": [ + 367, + 199, + 630, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zeyuxiao@nus.edu.sg, xinchao@nus.edu.sg", + "bbox": [ + 300, + 219, + 684, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Exploiting temporal correlations is crucial for video super-resolution (VSR). Recent approaches enhance this by incorporating event cameras. In this paper, we introduce MamEVSR, a Mamba-based network for event-based VSR that leverages the selective state space model, Mamba. MamEVSR stands out by offering global receptive field coverage with linear computational complexity, thus addressing the limitations of convolutional neural networks and Transformers. The key components of MamEVSR include: (1) The interleaved Mamba (iMamba) block, which interleaves tokens from adjacent frames and applies multidirectional selective state space modeling, enabling efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity. (2) The cross-modality Mamba (cMamba) block facilitates further interaction and aggregation between event information and the output from the iMamba block. The cMamba block can leverage complementary spatio-temporal information from both modalities and allows MamEVSR to capture finer motion details. Experimental results show that the proposed MamEVSR achieves superior performance on various datasets quantitatively and qualitatively.", + "bbox": [ + 88, + 301, + 485, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) [18, 25-28] is a fundamental task in computer vision aimed at reconstructing high-resolution (HR) videos from low-resolution (LR) inputs. Given its broad applications in areas such as video surveillance [1, 39, 91], high-definition television [2-4, 12-15, 20, 45-47, 57, 79, 83], and satellite imagery [16, 54, 76, 77], VSR has attracted significant interest. The main challenge lies in effectively exploiting temporal information. Advanced VSR techniques often leverage temporal information through sliding window methods [29, 34, 42, 48, 53, 70, 73-75, 88], recurrent architectures [8-10, 36, 40, 61, 86], and Transformer-based designs [49, 50, 59, 63, 65, 66],", + "bbox": [ + 89, + 691, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08953d7a3671b444b9e54eff5a99af9ef529f9288c5fd5bd953a4d38496646fe.jpg", + "image_caption": [ + "Figure 1. Inference time and performance comparison. MamEVSR outperforms advanced methods with high efficiency. Comparisons are performed on the CED dataset. The green circles represent RGB-based VSR methods, while the yellow circles represent event-based VSR methods." + ], + "image_footnote": [], + "bbox": [ + 517, + 267, + 906, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "all demonstrating significant achievements. Recently, researchers have integrated event cameras into VSR to boost performance by harnessing their high temporal resolution and dynamic range capabilities.", + "bbox": [ + 511, + 578, + 906, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras are bio-inspired sensors that asynchronously detect intensity changes at the microsecond level on a per-pixel basis, generating millions of events per second while maintaining robustness in high-dynamic-range lighting. Leveraging these advantages, recent event-based VSR methods [30, 33, 52, 81, 82] have outperformed frame-only approaches by enhancing flow estimation, temporal alignment, and cross-modal fusion using event streams. However, current methods, primarily based on CNNs and attention models, face limitations, including poor adaptability to dynamic inputs, insufficient receptive fields for capturing inter-frame correlations at high resolutions, and high computational costs.", + "bbox": [ + 509, + 641, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, natural language processing (NLP) has recently witnessed the emergence of structured state space models (SSMs) [22]. Theoretically, SSMs combine the benefits of recurrent neural networks (RNNs) and", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12564", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CNNs, leveraging the global receptive field characteristic of RNNs and the computational efficiency of CNNs. One particularly notable SSM is the selective state space model, also known as Mamba [21], which has garnered significant attention within the vision community. Mamba's key feature is its ability to make SSM parameters time-variant (i.e., data-dependent), enabling it to effectively select relevant context within sequences—a crucial factor for enhancing model performance. Inspired by Mamba's capabilities, we explore its potential in addressing the task of Event-based VSR, and we propose the first Mamba-based method for this task, named MamEVSR.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MamEVSR leverages a bidirectional recurrent pipeline to recurrently align and propagate temporally correlated features and integrate cross-modal information from the event stream. Its key components include: (1) The interleaved Mamba (iMamba) block: To facilitate efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, we introduce the iMamba block. Since the original SSMs model processes a single sequence, we merge tokens from two adjacent frames into one sequence for effective inter-frame modeling. By interleaving tokens from both frames and conducting multi-directional SSMs, we enable interactions between adjacent tokens from different propagated frames during sequence modeling. This design ensures that intermediate tokens in the sequence are from their spatio-temporal neighborhood. Stacking multiple iMamba blocks enhances MamEVSR's ability to handle complex inter-frame information exchange and frame alignment. (2) The cross-modality Mamba (cMamba) block: The cMamba block employs a cross-multiplication mechanism to enhance the interaction and aggregation between event information and the output from the iMamba block. This mechanism allows for the effective integration of event data with the temporal-fused features, thereby enriching the model's ability to leverage complementary cross-modal information and improve the overall reconstruction quality. By combining these two blocks, MamEVSR effectively exploits the strengths of both Mamba and event data to achieve superior performance in event-based VSR tasks. Experimental results on benchmark datasets demonstrate the effectiveness of our proposed MamEVSR (see an example in Figure 1).", + "bbox": [ + 91, + 275, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows: (1) To the best of our knowledge, this marks the first successful application of state space models, specifically Mamba, in event-based VSR. (2) We propose the iMamba block and the cMamba block as core components in MamEVSR for event-based VSR. The former enables efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, the latter leverages complementary spatio-temporal information to capture finer motion details. (3) Extensive evaluations on benchmark", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datasets demonstrate MamEVSR's superior performance, establishing a new benchmark for future explorations of Mamba's potential within event-based VSR.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 148, + 653, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video super-resolution. Existing VSR methods are designed to enhance the quality of LR frames by leveraging temporal information through sliding windows and recurrent structures. Sliding-window-based methods align adjacent LR frames with a reference frame to estimate an HR output. Early work focused on explicit optical flow estimation for frame alignment [6, 53, 69, 80, 84]. More recent methods have shifted to implicit alignment using dynamic filters [31], deformable convolutions [70, 75], and attention modules [28, 43]. Recent attention-based methods achieve new state-of-the-art performance [7, 37, 38, 43, 49, 53, 75, 78, 80]. However, these methods face challenges in capturing long-distance temporal features. Recurrent-based methods use RNNs to exploit temporal information across multiple frames. Sajjadi et al. [61] introduce a recurrent framework with optical flow for alignment. BasicVSR [9] and BasicVSR++[10] utilize bi-directional hidden states and advanced grid propagation techniques. PSRT [63] builds on BasicVSR++ with multi-frame self-attention for feature processing. In this paper, we focus on event-based VSR and introduce MamEVSR.", + "bbox": [ + 511, + 175, + 903, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based VSR. Event cameras have the unique capability to measure intensity changes at each pixel independently with microsecond accuracy, making them valuable for VSR. Jing et al. [30] propose a two-stage method that uses events to interpolate low-resolution (LR) video, generating a high-frequency video for subsequent high-resolution (HR) frame reconstruction. Lu et al. [52] present a joint framework learning implicit neural representations from both RGB frames and events, enabling arbitrary-scale VSR. Kai et al. [32] introduce EvTexture, an event-driven texture enhancement network dedicated to restoring textures in VSR through the incorporation of high-frequency event signals. Xiao et al. [82] take inspiration from the parameter-efficient tuning, introducing an event adapter for event-based VSR. In this paper, we introduce MamEVSR, the first Mamba-based method for event-based VSR.", + "bbox": [ + 511, + 492, + 903, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State space models. SSMs [17, 23, 44, 55, 64, 87, 90] are emerging as strong alternatives to Transformers [71] in natural language processing. S4 [23] is introduced for efficient, linear-complexity sequence modeling, followed by S5 [64], which improves parallelization, and GSS [55], which incorporates gated mechanisms. Mamba (S6) [21] stands out with its data-dependent parameters and hardware efficiency, outperforming Transformers in long-sequence tasks. In the computer vision domain, Vim [92] permutes 2D images into sequences for global modeling using bidirectional SSMs. Vmamba [51] extends this to four directions with a hier-", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f559444258a2bb6522a59d852909b5f0fcad5cff72c153cafcd211f9904f650c.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed MamEVSR. MamEVSR follows a typical bi-directional propagation scheme [9], consisting of bi-directional recurrent cells $For(\\cdot)$ and $Back(\\cdot)$ . For clarity, we have omitted the residual connection from the bicubic upsampling of the input, which is added to the final output, in the figure." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "archical design. VideoMamba [41] applies S6 to spatiotemporal video sequences, while MambaIR [24] is the first to use S6 for image restoration, surpassing Transformer-based methods. In this work, we explore the potential of Mamba in event-based VSR.", + "bbox": [ + 89, + 455, + 483, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 553, + 181, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting State Space Models", + "text_level": 1, + "bbox": [ + 89, + 580, + 356, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State space models (SSM) [17, 23, 55, 64] represent a class of sequence-to-sequence modeling systems characterized by constant dynamics over time, a property also known as linear time-invariant. With linear complexity, SSM can effectively capture the inherent dynamics through an implicit mapping to latent states, which can be defined as", + "bbox": [ + 89, + 606, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny (t) = C h (t) + D x (t), \\dot {h} (t) = A h (t) + B x (t). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 714, + 482, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $x(t)\\in \\mathbb{R},h(t)\\in \\mathbb{R}^N$ , and $y(t)\\in \\mathbb{R}$ denotes the input, hidden state, and the output, respectively. $N$ is the state size and $\\dot{h} (t)$ refers to the time derivative of $h(t)$ . Additionally, $A\\in \\mathbb{R}^{N\\times N}$ $B\\in \\mathbb{R}^{N\\times 1}$ $C\\in \\mathbb{R}^{1\\times N}$ ,and $D\\in \\mathbb{R}$ are the system matrices. To process discrete sequences like image and text, SSMs adopt Zero-Order Hold discretization [22] to map the input sequence $\\{x_{1},x_{2},\\dots,x_{K}\\}$ to the output sequence $\\{y_1,y_2,\\dots,y_K\\}$ . Specifically, suppose $\\Delta \\in \\mathbb{R}^{D}$ is the pre-defined timescale parameter to map continuous parameters $A$ $B$ into a discrete space, the dis", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f93695cb651ecb14c2405b43db020db1592d617abdfea9b0c7d10bcf73d6a19.jpg", + "image_caption": [ + "Figure 3. Overview of the forward recurrent cell $\\text{For}(\\cdot)$ , the backward recurrent cell $\\text{Back}(\\cdot)$ , and the reconstructor $R(\\cdot)$ ." + ], + "image_footnote": [], + "bbox": [ + 526, + 92, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cretization process can be formulated as", + "bbox": [ + 511, + 469, + 779, + 483 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {A} = \\exp (\\Delta A), \\bar {B} = (\\Delta A) ^ {- 1} (\\exp (A) - I) \\Delta B, \\bar {C} = C, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 489, + 903, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {k} = \\bar {C} h _ {k} + \\bar {D} x _ {k}, h _ {k} = \\bar {A} h _ {k - 1} + \\bar {B} x _ {k}.\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 512, + 777, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, all the matrices keep the same dimension as the operation iterates. Notably, $\\overline{D}$ , serving as a residual connection, is often discarded in the equation", + "bbox": [ + 511, + 534, + 903, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {k} = \\bar {C} h _ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 673, + 587, + 903, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides, following Mamba [21], the matrix $\\overline{B}$ can be approximated by the first-order Taylor series", + "bbox": [ + 511, + 611, + 903, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {B} = (\\exp (A) - I) A ^ {- 1} B \\approx (\\Delta A) (\\Delta A) ^ {- 1} \\Delta B = \\Delta B. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 648, + 903, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed iMamba and cMamba blocks are built upon Mamba and are specifically designed to address the challenges of event-based VSR tasks.", + "bbox": [ + 511, + 672, + 906, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 511, + 727, + 624, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{I}^{LR} = \\{\\ldots, I_{t-1}^{LR}, I_t^{LR}, I_{t+1}^{LR}, \\ldots\\}$ ( $I_t^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}$ ) be a sequence of LR input frames, $E^{f,b}$ be forward and backward event streams, the goal of MamEVSR is to generate $\\mathcal{I}^{SR} = \\{\\ldots, I_{t-1}^{SR}, I_t^{SR}, I_{t+1}^{SR}, \\ldots\\}$ ( $I_t^{SR} \\in \\mathbb{R}^{sH \\times sW \\times 3}$ ), which should be close to the ground-truth sequence $\\mathcal{I}^{GT} = \\{\\ldots, I_{t-1}^{GT}, I_t^{GT}, I_{t+1}^{GT}, \\ldots\\}$ ( $I_t^{GT} \\in \\mathbb{R}^{sH \\times sW \\times 3}$ ). $T, H,$ and $W$ are the frame number, height, and width, respectively. $s$ is the upscaling factor.", + "bbox": [ + 511, + 747, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Because the event streams are not convenient for observation and processing by convolutional neural networks, we", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12566", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "convert forward and backward event streams as voxel grids $\\mathbf{V}^{f,b}\\in \\mathbb{R}^{(T - 1)\\times H\\times W\\times Bin}$ via the temporal bilinear interpolation scheme", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV (i) = \\sum_ {k} p _ {k} \\max \\left(0, 1 - \\left| (i - 1) - \\frac {t _ {k} - t _ {0}}{t _ {N _ {e}} - t _ {0}} (B i n - 1) \\right|\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 140, + 482, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i \\in \\{1, \\dots, Bin\\}$ represents the $i$ -th time bin. In our experiments, consistent with previous studies [32, 33], we also set $Bin = 5$ . Furthermore, to mitigate the impact of hot pixels, we follow the study [32, 33] and normalize the voxel grid $V$ as", + "bbox": [ + 89, + 179, + 483, + 255 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} (i) = \\min ((V (i), \\eta)) / \\eta , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 258, + 482, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is the 98-th percentile value in the non-zero values of $V$ . In this way, we obtain the normalized voxel grid $\\hat{V} \\in \\mathbb{R}^{H \\times W \\times Bin}$ , which contains rich high-frequency textural information. We represent $\\pmb{V}^{f,b}$ using this voxel grid representation for further processing.", + "bbox": [ + 89, + 282, + 483, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 shows an overview of the proposed MamEVSR. MamEVSR employs bidirectional recurrent cells $For(\\cdot)$ and $Back(\\cdot)$ akin to the scheme proposed in [9]. However, it introduces novel elements, such as extra inputs and specialized modules to harness event streams, which sets it apart from prior approaches. The event voxels and the LR frames are first converted into the feature domain using the feature encoders $(f_{En}^{V})$ and $f_{En}^{I}$ , consisting of $N_{1}$ residual blocks), following which they are fed to the bidirectional recurrent cells. These cells consist of the iMamba block and the cMamba block. This step aims to utilize temporal correspondence from adjacent frames and aggregate cross-modal information from the event stream. As shown in Figure 3, using timestamp $t$ as an example, the process can be expressed as", + "bbox": [ + 89, + 358, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t, f} ^ {c M}, F _ {t, f} ^ {E, c M} = \\operatorname {F o r} \\left(I _ {t - 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {f}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 587, + 482, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t, b} ^ {c M}, F _ {t, b} ^ {E, c M} = B a c k \\left(I _ {t + 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {b}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 609, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, using the forward recurrent cell as an example, at time $t$ , $F_{t-1}^{LR}$ and $F_{t}^{LR}$ are fed to the iMamba block for efficient feature fusion and propagation across adjacent frames, resulting in $F_{t,f}^{LR'}$ . Next, $F_{t,f}^{LR'}$ and $I_{t}^{LR}$ are concatenated along the channel dimension, followed by the residual of $F_{t}^{LR}$ to enhance the representation capacity. The resulting features $F_{t,f}^{iM}$ , along with $F_{t}^{f}$ , are fed to the cMamba block, which facilitates further interaction and aggregation in a cross-modality manner. The outputs of the cMamba block, $F_{t,f}^{cM}$ and $F_{t,f}^{E,cM}$ , are used for further reconstruction. The backward cell operates similarly.", + "bbox": [ + 89, + 631, + 482, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate the super-resolved output $I_{t}^{SR}$ at timestamp $t$ , features $F_{t,f}^{cM}, F_{t,b}^{cM}, F_{t,f}^{E,cM}, F_{t,b}^{E,cM}$ , along with $I_{t}^{LR}$ and $F_{t}^{LR}$ , are fed to the reconstructor, which consists of an iMamba block, $N_{2}$ residual blocks, and a pixel-shuffling operation. The process can be expressed as", + "bbox": [ + 89, + 801, + 483, + 880 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} ^ {S R} = R \\left(F _ {t, f} ^ {c M}, F _ {t, b} ^ {c M}, F _ {t, f} ^ {E, c M}, F _ {t, b} ^ {E, c M}, I _ {t} ^ {L R}, F _ {t} ^ {L R}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Interleaved Mamba Block", + "text_level": 1, + "bbox": [ + 513, + 90, + 750, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Effective inter-frame modeling is essential for VSR [9, 75, 86]. Conventional methods [75] rely on convolutional layers and attention mechanisms to capture temporal dependencies, offering fast inference but struggling with limited receptive fields. More advanced approaches, such as SemanticLens [68], introduce semantic-aware alignment strategies to enhance feature correspondence, but at the cost of increased computational complexity. To achieve a better balance between efficiency and effectiveness, we explore Mamba [21], an SSM, for inter-frame modeling. By leveraging its long-range dependency modeling and selective state updates, Mamba enables robust temporal feature fusion while maintaining computational efficiency, making it well-suited for VSR.", + "bbox": [ + 511, + 114, + 906, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 4(a), we introduce the iMamba block, which follows a Transformer-like structure (i.e., Norm $\\rightarrow$ Attention $\\rightarrow$ Norm $\\rightarrow$ MLP) [89], with two key modifications: (1) We replace the attention mechanism with a custom block that scans feature maps from two input frames. Thanks to the interleaved token reorganization design, this enables global inter-frame modeling with linear complexity. (2) Inspired by MambaIR [24] and MLFSR [19], which address limitations in locality and inter-channel interaction, we substitute the original MLP with a channel-attention block to enhance feature fusion and propagation. Here, Norm indicate the normalization operation and MLP means the multilayer perceptron.", + "bbox": [ + 511, + 328, + 908, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For inter-frame modeling, given two adjacent low-resolution frame features, $F_{t-1}^{LR}$ and $F_{t}^{LR}$ (in the forward direction), both are first fed into the patch embedding module. This is followed by a series of operations, including layer normalization, linear projection, depth-wise convolution, and token reorganization, resulting in the reorganized feature $F_{Re,f}^{LR}$ . The selective SSM-based Mamba block is then applied to independently model each direction. After processing, the sequences are rearranged and merged via token partitioning, another round of layer normalization, and linear operations. Subsequently, a channel attention block with a residual connection is applied to enhance feature propagation and stability. The output of the iMamba block is denoted as $F_{t,f}^{LR'}$ for the forward direction. Similarly, for the backward direction, $F_{t,b}^{LR'}$ is obtained following the same operations.", + "bbox": [ + 511, + 527, + 908, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The interleaved reorganization within the iMamba block offers significant advantages for VSR. By ensuring that aggregated tokens come from both spatial and temporal neighborhoods, this design enhances local modeling, which is essential for capturing fine details between adjacent frames. Minimizing the distance between spatially and temporally adjacent tokens preserves the local context and reduces noise, which is particularly beneficial for maintaining tem-", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2836ea4a022aa052f45f9cee90119e22c9bc40937e05d5fc13891b8ab2dc9c31.jpg", + "image_caption": [ + "Figure 4. The structure of the proposed iMamba block and the cMamba block. The blocks offer specialized mechanisms to handle interframe modeling and cross-modal interactions effectively. See texts for more details." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 478, + 424 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b50cc4799de8371bcf7e460b3612aa8c80b12f87f87903115cf41fac9b0a468b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 88, + 893, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "poral coherence across frames. This improved context modeling directly enhances the quality of fine detail reconstruction, leading to superior video restoration results.", + "bbox": [ + 89, + 465, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Cross-Modality Mamba Block", + "text_level": 1, + "bbox": [ + 89, + 526, + 359, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The fusion of event data and RGB information is crucial for event-based VSR. Event cameras capture high-temporal-resolution motion cues that complement the rich spatial details in RGB frames. By integrating this sparse, high-frequency motion data with RGB content, we can significantly improve motion estimation, edge preservation, and image quality, especially in challenging conditions like low light or fast motion. This fusion enhances temporal consistency and fine detail reconstruction, leading to superior VSR performance. We, therefore, design the cMamba block for cross-modality fusion.", + "bbox": [ + 88, + 549, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking the forward direction as an example, as shown in Figure 4(b), the input features from the RGB frame output of the iMamba block, $F_{t,f}^{iM}$ , and the event feature, $F_{t}^{f}$ , are first processed by linear layers and depth-wise convolutions, generating $F^{RGB}$ and $F^{Event}$ before being fed to the cross selective SSM.", + "bbox": [ + 89, + 717, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the Mamba selection mechanism described in Section 3.1, system matrices $B$ , $C$ , and $\\Delta$ are generated to endow the model with context-aware capabilities, with linear projection layers generating these matrices. According to Eq. 2, matrix $C$ decodes information from the hidden state $h_k$ to produce the output $y_k$ . Inspired by the cross", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "attention mechanism [11], we enable cross-modal interaction between the RGB and event modalities through information exchange between multiple selective scan modules. In particular, the process can be represented as", + "bbox": [ + 511, + 465, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {A} _ {I} = \\exp \\left(\\Delta_ {\\mathrm {I}} A _ {I}\\right), \\quad \\bar {A} _ {E} = \\exp \\left(\\Delta_ {\\mathrm {E}} A _ {E}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 539, + 903, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {B} _ {I} = \\Delta_ {\\mathrm {r g b}} B _ {I}, \\bar {B} _ {E} = \\Delta_ {\\mathrm {x}} B _ {E}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 559, + 903, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {I} ^ {t} = \\bar {A} _ {I} h _ {I} ^ {t - 1} + \\bar {B} _ {I} x _ {I} ^ {t}, h _ {E} ^ {t} = \\bar {A} _ {E} h _ {E} ^ {t - 1} + \\bar {B} _ {E} x _ {E} ^ {t}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 579, + 903, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {I} ^ {t} = C _ {\\times} h _ {I} ^ {t} + D _ {I} x _ {I} ^ {t}, y _ {E} ^ {t} = C _ {\\mathrm {r g b}} h _ {E} ^ {t} + D _ {E} x _ {E} ^ {t}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 599, + 903, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {I} = \\left[ y _ {I} ^ {1}, y _ {I} ^ {2}, \\dots , y _ {I} ^ {l} \\right], y _ {E} = \\left[ y _ {E} ^ {1}, y _ {E} ^ {2}, \\dots , y _ {E} ^ {l} \\right]. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 621, + 903, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $x_{I / E}^{t}$ represents the input at time step $t$ , and $y_{I / E}$ denotes the selective scan output. $C_E$ and $C_I$ are the cross-modal matrices used for recovering the outputs at each time step from the hidden states.", + "bbox": [ + 511, + 651, + 903, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The outputs from the two modality branches—RGB and event—are individually fed into the layer normalization operation, followed by a residual connection. This results in the final cross-modality enhanced feature, which effectively integrates information from both branches, improving feature representation for tasks like VSR. The above process can be denoted as", + "bbox": [ + 511, + 712, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t, f} ^ {\\hat {i} M}, \\hat {F} _ {t} ^ {f} = \\operatorname {C S S M} \\left(\\mathrm {D} \\left(\\operatorname {L i n e a r} \\left(F _ {t, f} ^ {i M}\\right)\\right), \\mathrm {D} \\left(\\operatorname {L i n e a r} \\left(F _ {t} ^ {f}\\right)\\right)\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 825, + 903, + 845 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t, f} ^ {c M} = \\mathrm {L N} \\left(F _ {t, f} ^ {\\hat {i} M}\\right) + F _ {t, f} ^ {i M}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 856, + 903, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t, f} ^ {E, c M} = \\mathrm {L N} \\left(\\hat {F} _ {t} ^ {f}\\right) + F _ {t} ^ {f}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 882, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fb85f2bbb10bec6488894f91ad4a4d561742a600d6aea0d9641f603a426a3703.jpg", + "table_caption": [ + "Table 1. Quantitative comparison in terms of PSNR and SSIM for the $\\times 2$ event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined, respectively. Results are from [32] and [82]." + ], + "table_footnote": [], + "table_body": "
TypeMethodCED
Scene 1Scene 2Scene 3Scene 4Scene 5Scene 6Scene 7Scene 8Scene 9Scene 10Scene 11Average
RGB-based VSRDUF-----------31.09
-----------0.9183
SOF-VSR-----------31.84
-----------0.9226
TDAN35.8332.1231.5735.7335.4237.7528.9132.5435.5530.6735.0933.74
0.95400.93390.94660.95660.95360.94400.90620.90060.95410.93230.95610.9398
RBPN40.0734.1533.8339.5639.4440.3330.3634.9140.0531.5139.0336.66
0.98680.97390.97390.98690.98590.97820.96480.95020.98780.95510.98620.9754
BasicVSR39.3539.8139.7339.6039.4542.7139.1536.9739.3538.4539.4139.57
0.97840.97660.98320.97890.97780.98150.97480.96720.97760.97320.97990.9778
Event-based VSRE-VSR41.0834.7734.4440.4940.3240.8040.8035.1641.0031.7939.9737.32
0.98910.97750.97730.98910.98800.98010.98010.95360.99780.95860.98840.9783
EGVSR38.7838.6838.6739.0638.9341.9638.0336.1438.8437.6838.8638.69
0.97940.97500.98150.97980.97920.98310.97550.96350.97870.97260.98100.9771
EBVSR39.9540.2340.3140.2240.0643.0839.9737.2439.9538.8840.0740.14
0.98110.97800.98490.98160.98050.98300.97540.96890.98020.97570.98040.9801
EvTexture40.3940.5440.7540.6640.4543.2740.5337.5740.3539.2740.5440.52
0.98240.97890.98590.98290.98190.98340.98000.97050.98150.97690.98370.9813
EvTexture+-----------40.57
-----------0.9815
MamEVSR41.1740.6541.1041.4641.3143.2741.7037.7641.1639.3741.5041.14
0.98480.97910.98660.98530.98450.98260.98440.97190.98390.97750.98630.9831
", + "bbox": [ + 94, + 119, + 901, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The features $F_{t,f}^{cM}$ and $F_{t,f}^{E,cM}$ , obtained after cross-modal fusion, are fed into the reconstructor (see Figure 3) for the final image reconstruction. These fused features effectively capture both the high-frequency event data and spatial richness from the RGB frames, enhancing the quality of the reconstructed output by leveraging complementary information from both modalities.", + "bbox": [ + 89, + 407, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 529, + 225, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 553, + 297, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For training, we utilize the REDS [56] and CED [62] datasets. Specifically, the REDS4 subset (clips 000, 011, 015, 020) from the REDS training set is employed as our test set, evaluating performance in the RGB channels. Consistent with prior work [30, 32, 33], we simulate events using the ESIM [60] tool for these clips. These simulated events are subsequently transformed into voxel grids according to Eqs. (5) and (6). Voxels are downsampled through the Bicubic interpolation operation, aligning with the frame downsampling method [32]. Following previous event-based VSR studies [33, 52], we use the CED [62] dataset for training and evaluating on real-world scenes1. The dataset is captured with a DAVIS346 [5] event camera, which outputs temporally synchronized events and frames at a resolution of $346 \\times 260$ . When calculating the metrics (PSNR and SSIM), we exclude boundary 8 pixels and evaluate in the RGB channel.", + "bbox": [ + 88, + 575, + 482, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use 15 frames as input during training, with a mini-batch size of 6 and an input frame resolution of $64 \\times 64$ . We apply data augmentation techniques to the training data, including horizontal flips and random rotations of $90^{\\circ}$ , $180^{\\circ}$ , and $270^{\\circ}$ . MamEVSR is trained for 300K iterations using the Adam optimizer with a Cosine Annealing learning rate scheduler. Network architecture parameters are set to $N_{1} = 2$ and $N_{2} = 45$ . Supervision is provided by the Charbonnier penalty loss [35]:", + "bbox": [ + 511, + 409, + 906, + 546 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sqrt {\\| I ^ {S R} - I ^ {G T} \\| ^ {2} + \\varepsilon^ {2}}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 555, + 906, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\varepsilon$ is set to $1\\times 10^{-3}$ in our experiments. $I^{SR}$ denotes the results generated by MamEVSR, and $I^{GT}$ represents the ground-truth frames. We omit the subscript $t$ for simplicity. The initial learning rate for MamEVSR is $2\\times 10^{-4}$ . Training is conducted on 2 NVIDIA RTX 3090 GPUs.", + "bbox": [ + 511, + 580, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative and Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 511, + 667, + 877, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed MamEVSR with a wide range of potential methods that could be used to address event-based VSR, aiming to explore as many diverse and rich approaches as possible. (1) RGB-based VSR methods: DUF [31], SOF-VSR [72], RBPN [25], EDVR [75], BasicVSR [9], VRT [49], and TTVSR [50]. In particular, we exclude the event stream and solely feed the LR video frames into these VSR networks for reconstruction, resulting in the final reconstructed video output. (2) Event-based VSR methods: we compare our MamEVSR with E-VSR [30], EGVSR [52], EBVSR [33], EvTexture, and EvTexture+ [32] to provide a thorough evaluation.", + "bbox": [ + 511, + 688, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. As illustrated in Table 1, Table 2 and Table 3, MamEVSR consistently outperforms other event", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1The 11 scenes are as follows: People_Dynamic_Wave, Indoors_Foosball_2, Simple_Wires_2, People_Dynamic_Dancing, People_Dynamic_Jumping, Simple_Fruit_Fast, Outdoor_Jumping_Infrared_2, Simple_Carpet_Fast, People_Dynamic_Armroll, Indoors_Kitchen_2, and People_Dynamic_Sitting [82].", + "bbox": [ + 89, + 838, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12569", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e039aaae2fdb63edded9d8f7b495de162e4dfe9b9c4c42ca7fab7d3df672d58c.jpg", + "table_caption": [ + "Table 2. Quantitative comparison in terms of PSNR and SSIM for the $\\times 4$ event-based VSR task on the REDS4 dataset. The best and second-best results are highlighted in bold and underlined, respectively. Results are from [32]." + ], + "table_footnote": [], + "table_body": "
REDS4RGB-based VSREvent-based VSR
DUFEDVRBasicVSRTTVSRVRTEGVSREBVSREvTextureEvTexture+MamEVSR
00027.30/0.793728.01/0.825028.40/0.843428.82/0.856628.85/0.855325.16/0.706628.44/0.844630.72/0.9082-/-30.87/0.9103
01128.38/0.805632.17/0.886432.47/0.897933.47/0.910033.49/0.907226.56/0.772232.55/0.898733.72/0.9145-/-33.87/0.9166
01531.55/0.884634.06/0.920634.18/0.922435.01/0.932535.26/0.933229.83/0.852634.22/0.923535.06/0.9314-/-35.21/0.9335
02027.30/0.816430.09/0.888130.63/0.900031.17/0.909431.16/0.907825.94/0.784630.67/0.900931.65/0.9154-/-31.80/0.9175
Average28.63/0.825131.09/0.880031.42/0.890932.12/0.902132.19/0.900626.87/0.779031.47/0.891932.79/0.917432.93/0.919532.94/0.9195
#Params5.820.66.36.835.62.612.28.910.19.6
", + "bbox": [ + 91, + 118, + 903, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d09f0b4873329567a6631b1831357bbcd55512330fc3c6057fc8dc75609ceba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 234, + 331, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acb116899ce1d040d36f1807e8f5c0cdc161b841509c2ad5fb61c46ae5dc0362.jpg", + "image_caption": [ + "Figure 5. Visual comparisons for $\\times 4$ VSR on REDS4. From left to right in sequence are patches cropped from LR, TDAN, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image." + ], + "image_footnote": [], + "bbox": [ + 89, + 340, + 331, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7991cf06befd2f8de94e9dcf0d36209e4ff71c1b0d0d63f3d467eaedbf72ce9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 234, + 403, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33c278d1f953fa791c490fe92e841118d663045fd4fa54d33c5f2bbf5d338463.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 287, + 403, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3b6f614ee88f573044a9a4d12f25e01d2b1265022f1b7d4df9d325be36918a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 340, + 401, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c974e2a90e3dac74bc19b2661dff6873ebb2e3bcd2185415cde594dc21c056a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 393, + 403, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08e78b722e96c5c7de965787fbfa48fb6f95c7973a988984db95b6f37b63b7f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 234, + 545, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c19d5f18e24148c9920d875daa67e8dd4749ce411b8c9c59d70bbe61d0b7623.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 287, + 545, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/274a0b9ad8a11145a2d3e0cbf5b386a90a86a964ec8fab1262a9db97f1598ae5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 340, + 545, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/481b0805ce2ff67a11ec08687bee33344b13ddb6cb9dbc2492c39e1e4d497cb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 393, + 545, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e839538222076ff7fcfd49928461bce481e321d351445d5e46427d3554540b1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 234, + 617, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7864f660d46bbe2f75f4caeef43f336e4c50150b243e3a98a3f57a2aa988700e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 287, + 616, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a4a928c77448541a3665ae043b34721438641bb9823c3e862dca9f502aa18e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 340, + 616, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1491547083335d285b5f87437d1cced10df9fb56bd261776a0019e12905f9d35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 393, + 616, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d9cdee591e61a97d39062ff1a1f4631a33b31cc4f9eca7599220cf5f4666925.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 234, + 689, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b56ab0024456ca8c96a224ad062a8f9c55db70d862b8ab7ee40c9c13ae05a381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 287, + 689, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bf4793a95ee4230457673546d026f8b75ca635b0df05502351266a215c3a43a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 340, + 689, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7febd35bd4926451a599b8aa85a0ca321743bfe8e1ccf8ea78e92ef796b5fc76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 393, + 689, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9495be13687f1410951638dc0362e526cc4cc404b42e11318998e941159e9b1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 234, + 763, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe36963c897234235fceccc9fc044b7dbeedee7ee578e56e82eaa968365e3d64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 287, + 763, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a404da58f7b9255a2172f5bed087058ac7eb07646e8b1e6d6f0b3b97f54089fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 340, + 763, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dfa976a0da885310210f1f128044dd7028d06ab4ddb2fb9a0d6b3b970343cb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 393, + 763, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f169570344f9fbaef314c6f1efe56c1c7c122f5b2936c1e1cec17b1654d12729.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 234, + 833, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01141179cede12f13f66f286eb3fd08d78bf584290e438bd911084855ad721a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 287, + 833, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/442e0a920102c0a9fc5c09875549abfd85cce755050f77141e2fe92b80fe09b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 340, + 833, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2df144de1dff438f05e90dabd2364b5c85af776a1cfbae4aac2c8cb951de601b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 393, + 833, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa15dfae37e40bdd2bed3073f7f7f29ca317e0c05d1f9167cfec265ff325e715.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 234, + 903, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c346450f86b2fb5d9f1aecdb26141346b116b074fb7a09a0acfcc412cb1dfcb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 287, + 903, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/654a0eeffa6f17e0043ec269032565e925e93c42e3ee79c8d92b07e6678691f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 340, + 903, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16643d135eced12494e57b059d5d25dd09bb5c42aa634b289436dbd900cada6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 393, + 903, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94a772be7c3395a64bcb14289d31f1d6dbb0ef0a216177b74d36627784a2a539.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 484, + 331, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c346c0717b69d01a3537414919f01da01b0fcc215162aceaca3c03e92435bc8b.jpg", + "image_caption": [ + "Figure 6. Visual comparisons for $\\times 4$ VSR on CED. From left to right in sequence are patches cropped from LR, EDVR, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image." + ], + "image_footnote": [], + "bbox": [ + 89, + 590, + 331, + 699 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ccd19546dd981c636109863fb5c163c0cd0f41289c1a10ae56bbe7359304392.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 484, + 403, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05983bf7b295d1de5a4aad09226b83200a5836ba4eec53d7ca13cc511c730ce5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 537, + 403, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b3f49cb3c9347a414f6c44e852d192f43ac13d4ca0a37f2c9dd55a1f74ed4a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 590, + 403, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8530c1e3c29498bcc4609e0cb8072208b1b12387bd83286a7e46cea0db63f21f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 645, + 403, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc02fe49fcae1b52d440280275d41b7ec7abd62052f2eb39b6b48f8f756aa72a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 484, + 545, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b955f09775b530f34c6236508546e55c76642dcacd7b987465e8ff50a7700604.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 537, + 545, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d14e9942c10956e3646a2bf7bc935e552cfa7d5dfeecf68e68f1905c0761254.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 590, + 545, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c62741f328207a46197dbd61504fe22cf5df6895b3be80ffa2720848618a4407.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 645, + 545, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5201d68d23065a7d1ae33c7c562cb1b44cb29fbdab678e1c5c1eb10b3d188408.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 484, + 617, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02892673f4523dd91fa2aa656f6a01afa9ce49ca2cd603ebc6767f013544538e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 537, + 616, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/809d63e656816cf6a98f42fe5ad94f3542dfa339348882f963ac846c57d43aa4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 590, + 616, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86993b96e0cbc6c7f521ab06b1da878e016919d367a7c3abd9463d6f28af2a10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 645, + 616, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4641407073ba7cdafbfe059953aa0419b4e310cfc4eb03eefb878f386862038a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 484, + 689, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6003d5f74d7fb1006fdbfc8bfcd38fe1439ad43c44d54e255f150c22804f852e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 537, + 689, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5d558d88d6462acc8ee4d4f557908d87eded0a6f55f3c1e80b63b9842782fb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 590, + 689, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5686ac1d5cf5d764873550746047bf2dc0b7815c2040beca2b54407e71289e47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 645, + 689, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b4bb252085922e1413b3f1e5f22b5152a1e52e1d13f1baf40998891261de11f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 484, + 763, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb858399e70c51c794427408ae7c43c93bf2318cbd5e7e145dbe717712724cd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 537, + 763, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27c655a80be0105c3595c129c8c937cc7aff8a8180db6405a255adefa66015a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 590, + 763, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/947374c1fc3a132118133a376e87a715de38f63969f683fdbf4de2991ff5753b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 645, + 763, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2909f6e655391dfb8e22b338ad9c070834ee5d08fa3d32014d30b9191f2762b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 484, + 833, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4acf297317579deb35408eeb69164aa825701d0b72147b7b3df5d2cc0156b11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 537, + 833, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a342a2158f94464fa24fe7b44b548a8c72ecaab09ce91cd87c559e6d2378c063.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 590, + 833, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ef2e896ff08d19704d0de4fd00048e088dc9327f843de66482bdab0cff2a9fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 645, + 833, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a48b88e8440db91fc32ba291535049f0c3983ed632067d7306351b69bfb3b178.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 484, + 903, + 537 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0c3f0179bc92a10aacbf6193cdef8e0c38d3327f4652176c2ca17f0ff71c4d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 537, + 903, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d66e8f02f5eadf479549c26c816165970c349df4dc9edb223c4e6fd90e8ee818.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 590, + 903, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0012f14a7ea5261c7334ecefd330e62c7c39e0c84a9a004ce48ec4af90ed3d77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 645, + 903, + 696 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2998b761e630fc361c3812a075700f7f71ec231609948b890c3d541c5d8d21d1.jpg", + "table_caption": [ + "Table 3. Quantitative comparison in terms of PSNR and SSIM for the $\\times 2$ event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined." + ], + "table_footnote": [], + "table_body": "
MethodPSNR (↑)SSIM (↑)MethodPSNR (↑)SSIM (↑)
SOF-VSR31.840.9226EGVSR38.690.9771
TDAN33.740.9398EBVSR40.140.9801
RBPN36.660.9754EvTexture40.520.9813
BasicVSR39.570.9778EvTexture+40.570.9815
E-VSR37.320.9783MamEVSR41.140.9831
", + "bbox": [ + 91, + 779, + 482, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based VSR methods such as E-VSR, EGVSR, EBVSR, EvTexture, and EvTexture+ across various datasets like CED", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and REDS4. For instance, on the CED dataset for the $\\times 2$ event-based VSR task, MamEVSR achieves an average PSNR of $33.74\\mathrm{~dB}$ and an SSIM of 0.9831, surpassing the next best method, EvTexture+, which scores $33.40\\mathrm{~dB}$ and 0.9827 respectively. Compared to RGB-based VSR methods like DUF, TDAN, RBPN, and BasicVSR, MamEVSR demonstrates superior performance. On the CED dataset for the $\\times 2$ event-based VSR task, MamEVSR's average PSNR of $33.74\\mathrm{~dB}$ and SSIM of 0.9831 far exceed those of the top-performing RGB-based method, BasicVSR, which achieves", + "bbox": [ + 511, + 739, + 906, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12570", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9647a26a19f84fc43380940fc38ae8df2ba2f563ebc58f5c47ba05221956f820.jpg", + "table_caption": [ + "Table 4. Ablation study of different components on CED." + ], + "table_footnote": [], + "table_body": "
MethodCED ×2CED ×4
PSNR↑SSIM↑PSNR↑SSIM↑
OursAll residual blocks39.990.979833.530.9082
Full model41.140.983134.030.9189
Core modules(a) w/o iMamba40.170.980333.600.9109
(b) w/o cMamba40.260.980533.630.9110
(c) w/o backward cell40.070.978933.480.9076
(d) w/o forward cell40.080.978933.400.9065
iMamba(e) Residual blocks40.620.981633.680.9111
(f) Deform. conv.40.720.981933.770.9127
(g) Flow warping40.840.982133.800.9129
(h) Concat & attention40.800.982033.770.9126
(i) w/o Int. Reorg.40.990.982633.880.9140
(j) w/o Channal att.41.020.982933.940.9145
cMamba(k) Residual blocks40.690.981833.710.9114
(l) EBVSR-BCS module [33]40.770.981933.800.9139
(m) Cross-modal att. [67]40.980.982633.880.9150
(n) Concat & iMamba40.800.982033.770.9129
(o) w/o Cross SSM40.870.982533.900.9170
Recon.(p) w/o iMamba41.040.982934.000.9179
(q) All iMamba40.890.982633.940.9168
", + "bbox": [ + 94, + 103, + 480, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "31.69 dB and 0.9226 respectively. Across all datasets and magnification factors, MamEVSR consistently delivers superior performance, showcasing its effectiveness in enhancing event-based video super-resolution tasks compared to both event-based and RGB-based VSR methods.", + "bbox": [ + 89, + 367, + 482, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational cost results. We calculate the number of parameters and runtime on the CED dataset. Results are shown in Table 2 and Figure 1. MamEVSR is positioned in the upper-left quadrant, reflecting an optimal trade-off between high performance and low computational cost. It achieves high-quality reconstructions (PSNR) with efficient computation and a modest parameter count, showcasing effective resource utilization.", + "bbox": [ + 89, + 443, + 483, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. We present visual comparison results on REDS4 and CED in Figure 5 and Figure 6, respectively. The LR patch appears blurry and lacks detail as we move through the SR patches; sharpness and clarity noticeably increase. MamEVSR produces results that closely resemble the ground truth, effectively restoring fine details and textures. For example, MamEVSR more accurately reconstructs wheel textures and license plate details, closely matching the ground truth in Figure 5.", + "bbox": [ + 89, + 564, + 483, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 710, + 243, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on CED in terms of PSNR and SSIM. Results are in Table 4.", + "bbox": [ + 89, + 733, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of core components. Removing the iMamba block decreases PSNR and SSIM for both CED $\\times 2$ and CED $\\times 4$ tasks. Omitting the cMamba block has a smaller impact on performance. Disabling either backward or forward cells reduces PSNR and SSIM, especially for CED $\\times 4$ . These results confirm the importance of the iMamba and cMamba blocks, as well as the bidirectional cell structure, in achieving optimal performance.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the iMamba block. Flow warping per", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b3e4bda64bbcb7eb9fd0b9d7563bed7b1fbe44273925cb3fb47040624a15731.jpg", + "image_caption": [ + "Figure 7. A failure case. From left to right in sequence are patches cropped from the ground truth image, EvTexture, and MamEVSR." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "forms better than concatenation and attention. Removing internal reorganization or attention mechanisms within the iMamba block causes slight drops in PSNR and SSIM. The iMamba block plays a critical role in improving MamEVSR's performance.", + "bbox": [ + 511, + 172, + 903, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the cMamba block. Incorporating the EBVSR-CBCS module [33] marginally boosts PSNR and SSIM. Cross-modal attention [67] yields comparable performance gains. Concatenating and using iMamba shows a slight decrease in metrics. Eliminating the cross-state space modeling mechanism within the cMamba block leads to minimal reductions in PSNR and SSIM.", + "bbox": [ + 511, + 247, + 903, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the reconstructor. Removing the iMamba block from the reconstructor slightly improves PSNR and SSIM for both CED $\\times 2$ and CED $\\times 4$ tasks. However, replacing all iMamba blocks with residual blocks leads to a notable drop in performance, particularly for CED $\\times 4$ .", + "bbox": [ + 511, + 353, + 906, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations and Discussions", + "text_level": 1, + "bbox": [ + 511, + 438, + 764, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While MamEVSR demonstrates strong results in event-based VSR, some challenges remain. As shown in Figure 7, MamEVSR struggles to recover particularly fine textures due to significant information loss in RGB data and low-resolution, poorly defined event data. Potential solutions include incorporating VGG loss to generate more realistic textures, as it effectively captures fine details in restoration tasks. Alternatively, using a longer temporal sequence of both event and RGB data could provide richer, continuous information, enhancing texture recovery. Additionally, we plan to design lighter-weight modules for event-based VSR and explore extending MamEVSR to other video restoration tasks. Investigating the effectiveness of other novel architectures [58, 85] for this task would also be insightful.", + "bbox": [ + 511, + 460, + 906, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 681, + 633, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose MamEVSR for event-based VSR. MamEVSR leverages the selective state space model to offer global receptive field coverage with linear complexity. Key components include the iMamba block for efficient bidirectional feature fusion and the cMamba block for integrating event information and capturing finer motion details. Experiments show that MamEVSR outperforms existing methods on various datasets, achieving superior quantitative and qualitative results.", + "bbox": [ + 511, + 703, + 903, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This project is supported by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation.", + "bbox": [ + 511, + 839, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Aakerberg, Kamal Nasrollahi, and Thomas B Moeslund. Real-world super-resolution of face-images from surveillance cameras. IET Image Processing, 16(2):442-452, 2022.", + "[2] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Yichen Wu, Lilun Deng, Yukun Cui, Tao Feng, and Shuang Xu. Taskdriven image fusion with learnable fusion loss. In CVPR, 2025.", + "[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024.", + "[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024.", + "[5] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based davis sensor. In ISCAS, pages 686-689, 2014.", + "[6] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017.", + "[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021.", + "[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In ECCV, 2022.", + "[9] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021.", + "[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022.", + "[11] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021.", + "[12] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, 2023.", + "[13] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Sixiang Chen, Tian Ye, Renjing Pei, Kaiwen Zhou, Fenglong Song, and Lei Zhu. Restoreagent: Autonomous image restoration agent via multimodal large language models. NeurIPS, 2025.", + "[14] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In CVPR, 2024.", + "[15] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV.", + "[16] Michel Deudon, Alfredo Kalaitzis, Israel Goytom, Md Ri" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fat Arefin, Zhichao Lin, Kris Sankaran, Vincent Michalski, Samira E Kahou, Julien Cornebise, and Yoshua Bengio. Highres-net: Recursive fusion for multi-frame super-resolution of satellite imagery. arXiv preprint arXiv:2002.06460, 2020.", + "[17] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In ICLR, 2022.", + "[18] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019.", + "[19] Ruisheng Gao, Zeyu Xiao, and Zhiwei Xiong. Mamba-based light field super-resolution with efficient subspace scanning. In ACCV, 2024.", + "[20] Tomio Goto, Takafumi Fukuoka, Fumiya Nagashima, Satoshi Hirano, and Masaru Sakurai. Super-resolution system for 4k-hdtv. In ICPR, 2014.", + "[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023.", + "[22] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021.", + "[23] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021.", + "[24] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024.", + "[25] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019.", + "[26] Yan Huang, Wei Wang, and Liang Wang. Bidirectional recurrent convolutional networks for multi-frame superresolution. NeurlPS, 28, 2015.", + "[27] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):1015-1028, 2017.", + "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020.", + "[29] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022.", + "[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In CVPR, 2021.", + "[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018.", + "[32] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. Eventure: Event-driven texture enhancement for video super-resolution. arXiv preprint arXiv:2406.13457, 2024.", + "[33] Dachun Kai, Yueyi Zhang, and Xiaoyan Sun. Video super" + ], + "bbox": [ + 516, + 92, + 903, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12572", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "resolution via event-driven temporal alignment. In ICIP, 2023.", + "[34] Soo Ye Kim, Jeongyeon Lim, Taeyoung Na, and Munchurl Kim. 3DSRnet: Video super-resolution using 3d convolutional neural networks. arXiv preprint arXiv:1812.09079, 2018.", + "[35] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017.", + "[36] Junyong Lee, Myeonghee Lee, Sunghyun Cho, and Seungyong Lee. Reference-based video super-resolution using multi-camera video triplets. In CVPR, 2022.", + "[37] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In CVPR, 2023.", + "[38] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023.", + "[39] Bonan Li, Zicheng Zhang, Xingyi Yang, and Xinchao Wang. Focus on neighbors and know the whole: Towards consistent dense multiview text-to-image generator for 3d creation. arXiv preprint arXiv:2408.13149, 2024.", + "[40] Fei Li, Linfeng Zhang, Zikun Liu, Juan Lei, and Zhenbo Li. Multi-frequency representation enhancement with privilege information for video super-resolution. In ICCV, 2023.", + "[41] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024.", + "[42] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019.", + "[43] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-correspondence aggregation network for video super-resolution. In ECCV, 2020.", + "[44] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. NeurIPS, 37, 2025.", + "[45] Zhuoyuan Li, Jiacheng Li, Yao Li, Li Li, Dong Liu, and Feng Wu. In-loop filtering via trained look-up tables. In VCIP, 2024.", + "[46] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024.", + "[47] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024.", + "[48] Wenyi Lian and Wenjing Lian. Sliding window recurrent network for efficient video super-resolution. In ECCV, 2022.", + "[49] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022.", + "[50] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video super-resolution. In CVPR, 2022.", + "[51] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024.", + "[52] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In CVPR, 2023.", + "[53] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xinchao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023.", + "[54] Yimin Luo, Liguo Zhou, Shu Wang, and Zhongyuan Wang. Video satellite imagery super resolution via convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 14(12):2398-2402, 2017.", + "[55] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022.", + "[56] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019.", + "[57] Tianbo Pan, Zidong Cao, and Lin Wang. Srfnet: Monocular depth estimation with fine-grained structure via spatial reliability-oriented fusion of frames and events. In ICRA, 2024.", + "[58] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023.", + "[59] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In ECCV, 2022.", + "[60] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018.", + "[61] Mehdi SM, Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In CVPR, 2018.", + "[62] Cedric Scheerlinck, Henri Rebecq, Timo Stoffregen, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Ced: Color event camera dataset. In CVPRW, 2019.", + "[63] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. NeurlPS, 2022.", + "[64] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In ICLR, 2022.", + "[65]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, 2023.", + "[66] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, 2022.", + "[67] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022.", + "[68] Qi Tang, Yao Zhao, Meiqin Liu, Jian Jin, and Chao Yao. Semantic lens: Instance-centric semantic alignment for video" + ], + "bbox": [ + 516, + 92, + 903, + 887 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12573", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "super-resolution. In AAAI, 2024.", + "[69] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017.", + "[70] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020.", + "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017.", + "[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through HR optical flow estimation. In ACCV, 2018.", + "[73] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, 2023.", + "[74] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, 2024.", + "[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019.", + "[76] Yi Xiao, Xin Su, Qiangqiang Yuan, Denghong Liu, Huanfeng Shen, and Liangpei Zhang. Satellite video superresolution via multiscale deformable convolution alignment and temporal grouping projection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021.", + "[77] Yi Xiao, Qiangqiang Yuan, Kui Jiang, Xianyu Jin, Jiang He, Liangpei Zhang, and Chia-wen Lin. Local-global temporal difference learning for satellite video super-resolution. arXiv preprint arXiv:2304.04421, 2023.", + "[78] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023.", + "[79] Zeyu Xiao, Zhen Cheng, and Zhiwei Xiong. Space-time super-resolution for light field videos. IEEE Transactions on Image Processing, 32:4785-4799, 2023.", + "[80] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, pages 2113-2122, 2021.", + "[81] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACMMM, pages 2409-2418, 2024.", + "[82] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In ECCV, 2024.", + "[83] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, 2023.", + "[84] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025.", + "[85] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. ICLR, 2025.", + "[86] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, pages 4429-4438, 2021.", + "[87] Weihao Yu and Xinchao Wang. Mambaout: Do we really" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "need mamba for vision? CVPR, 2025.", + "[88] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In ECCV, 2022.", + "[89] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022.", + "[90] Guozhen Zhang, Chuxnu Liu, Yutao Cui, Xiaotong Zhao, Kai Ma, and Limin Wang. Vfimamba: Video frame interpolation with state space models. NeurIPS, 37, 2024.", + "[91] Liangpei Zhang, Hongyan Zhang, Huanfeng Shen, and Pingxiang Li. A super-resolution reconstruction algorithm for surveillance images. Signal Processing, 90(3):848-859, 2010.", + "[92] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 349 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12574", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_model.json b/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9447efd55f00cf6e22b2aa0e3f8164eb344a9e2f --- /dev/null +++ b/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_model.json @@ -0,0 +1,3269 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Event-based Video Super-Resolution via State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.182, + 0.623, + 0.2 + ], + "angle": 0, + "content": "Zeyu Xiao Xinchao Wang*" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.2, + 0.631, + 0.218 + ], + "angle": 0, + "content": "National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.22, + 0.686, + 0.234 + ], + "angle": 0, + "content": "zeyuxiao@nus.edu.sg, xinchao@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.303, + 0.486, + 0.635 + ], + "angle": 0, + "content": "Exploiting temporal correlations is crucial for video super-resolution (VSR). Recent approaches enhance this by incorporating event cameras. In this paper, we introduce MamEVSR, a Mamba-based network for event-based VSR that leverages the selective state space model, Mamba. MamEVSR stands out by offering global receptive field coverage with linear computational complexity, thus addressing the limitations of convolutional neural networks and Transformers. The key components of MamEVSR include: (1) The interleaved Mamba (iMamba) block, which interleaves tokens from adjacent frames and applies multidirectional selective state space modeling, enabling efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity. (2) The cross-modality Mamba (cMamba) block facilitates further interaction and aggregation between event information and the output from the iMamba block. The cMamba block can leverage complementary spatio-temporal information from both modalities and allows MamEVSR to capture finer motion details. Experimental results show that the proposed MamEVSR achieves superior performance on various datasets quantitatively and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.222, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Video super-resolution (VSR) [18, 25-28] is a fundamental task in computer vision aimed at reconstructing high-resolution (HR) videos from low-resolution (LR) inputs. Given its broad applications in areas such as video surveillance [1, 39, 91], high-definition television [2-4, 12-15, 20, 45-47, 57, 79, 83], and satellite imagery [16, 54, 76, 77], VSR has attracted significant interest. The main challenge lies in effectively exploiting temporal information. Advanced VSR techniques often leverage temporal information through sliding window methods [29, 34, 42, 48, 53, 70, 73-75, 88], recurrent architectures [8-10, 36, 40, 61, 86], and Transformer-based designs [49, 50, 59, 63, 65, 66]," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.268, + 0.907, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Figure 1. Inference time and performance comparison. MamEVSR outperforms advanced methods with high efficiency. Comparisons are performed on the CED dataset. The green circles represent RGB-based VSR methods, while the yellow circles represent event-based VSR methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.64 + ], + "angle": 0, + "content": "all demonstrating significant achievements. Recently, researchers have integrated event cameras into VSR to boost performance by harnessing their high temporal resolution and dynamic range capabilities." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.642, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Event cameras are bio-inspired sensors that asynchronously detect intensity changes at the microsecond level on a per-pixel basis, generating millions of events per second while maintaining robustness in high-dynamic-range lighting. Leveraging these advantages, recent event-based VSR methods [30, 33, 52, 81, 82] have outperformed frame-only approaches by enhancing flow estimation, temporal alignment, and cross-modal fusion using event streams. However, current methods, primarily based on CNNs and attention models, face limitations, including poor adaptability to dynamic inputs, insufficient receptive fields for capturing inter-frame correlations at high resolutions, and high computational costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "On the other hand, natural language processing (NLP) has recently witnessed the emergence of structured state space models (SSMs) [22]. Theoretically, SSMs combine the benefits of recurrent neural networks (RNNs) and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12564" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "CNNs, leveraging the global receptive field characteristic of RNNs and the computational efficiency of CNNs. One particularly notable SSM is the selective state space model, also known as Mamba [21], which has garnered significant attention within the vision community. Mamba's key feature is its ability to make SSM parameters time-variant (i.e., data-dependent), enabling it to effectively select relevant context within sequences—a crucial factor for enhancing model performance. Inspired by Mamba's capabilities, we explore its potential in addressing the task of Event-based VSR, and we propose the first Mamba-based method for this task, named MamEVSR." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.746 + ], + "angle": 0, + "content": "MamEVSR leverages a bidirectional recurrent pipeline to recurrently align and propagate temporally correlated features and integrate cross-modal information from the event stream. Its key components include: (1) The interleaved Mamba (iMamba) block: To facilitate efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, we introduce the iMamba block. Since the original SSMs model processes a single sequence, we merge tokens from two adjacent frames into one sequence for effective inter-frame modeling. By interleaving tokens from both frames and conducting multi-directional SSMs, we enable interactions between adjacent tokens from different propagated frames during sequence modeling. This design ensures that intermediate tokens in the sequence are from their spatio-temporal neighborhood. Stacking multiple iMamba blocks enhances MamEVSR's ability to handle complex inter-frame information exchange and frame alignment. (2) The cross-modality Mamba (cMamba) block: The cMamba block employs a cross-multiplication mechanism to enhance the interaction and aggregation between event information and the output from the iMamba block. This mechanism allows for the effective integration of event data with the temporal-fused features, thereby enriching the model's ability to leverage complementary cross-modal information and improve the overall reconstruction quality. By combining these two blocks, MamEVSR effectively exploits the strengths of both Mamba and event data to achieve superior performance in event-based VSR tasks. Experimental results on benchmark datasets demonstrate the effectiveness of our proposed MamEVSR (see an example in Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows: (1) To the best of our knowledge, this marks the first successful application of state space models, specifically Mamba, in event-based VSR. (2) We propose the iMamba block and the cMamba block as core components in MamEVSR for event-based VSR. The former enables efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, the latter leverages complementary spatio-temporal information to capture finer motion details. (3) Extensive evaluations on benchmark" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "datasets demonstrate MamEVSR's superior performance, establishing a new benchmark for future explorations of Mamba's potential within event-based VSR." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.15, + 0.655, + 0.166 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.176, + 0.905, + 0.492 + ], + "angle": 0, + "content": "Video super-resolution. Existing VSR methods are designed to enhance the quality of LR frames by leveraging temporal information through sliding windows and recurrent structures. Sliding-window-based methods align adjacent LR frames with a reference frame to estimate an HR output. Early work focused on explicit optical flow estimation for frame alignment [6, 53, 69, 80, 84]. More recent methods have shifted to implicit alignment using dynamic filters [31], deformable convolutions [70, 75], and attention modules [28, 43]. Recent attention-based methods achieve new state-of-the-art performance [7, 37, 38, 43, 49, 53, 75, 78, 80]. However, these methods face challenges in capturing long-distance temporal features. Recurrent-based methods use RNNs to exploit temporal information across multiple frames. Sajjadi et al. [61] introduce a recurrent framework with optical flow for alignment. BasicVSR [9] and BasicVSR++[10] utilize bi-directional hidden states and advanced grid propagation techniques. PSRT [63] builds on BasicVSR++ with multi-frame self-attention for feature processing. In this paper, we focus on event-based VSR and introduce MamEVSR." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.493, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Event-based VSR. Event cameras have the unique capability to measure intensity changes at each pixel independently with microsecond accuracy, making them valuable for VSR. Jing et al. [30] propose a two-stage method that uses events to interpolate low-resolution (LR) video, generating a high-frequency video for subsequent high-resolution (HR) frame reconstruction. Lu et al. [52] present a joint framework learning implicit neural representations from both RGB frames and events, enabling arbitrary-scale VSR. Kai et al. [32] introduce EvTexture, an event-driven texture enhancement network dedicated to restoring textures in VSR through the incorporation of high-frequency event signals. Xiao et al. [82] take inspiration from the parameter-efficient tuning, introducing an event adapter for event-based VSR. In this paper, we introduce MamEVSR, the first Mamba-based method for event-based VSR." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "State space models. SSMs [17, 23, 44, 55, 64, 87, 90] are emerging as strong alternatives to Transformers [71] in natural language processing. S4 [23] is introduced for efficient, linear-complexity sequence modeling, followed by S5 [64], which improves parallelization, and GSS [55], which incorporates gated mechanisms. Mamba (S6) [21] stands out with its data-dependent parameters and hardware efficiency, outperforming Transformers in long-sequence tasks. In the computer vision domain, Vim [92] permutes 2D images into sequences for global modeling using bidirectional SSMs. Vmamba [51] extends this to four directions with a hier-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12565" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed MamEVSR. MamEVSR follows a typical bi-directional propagation scheme [9], consisting of bi-directional recurrent cells \\( For(\\cdot) \\) and \\( Back(\\cdot) \\). For clarity, we have omitted the residual connection from the bicubic upsampling of the input, which is added to the final output, in the figure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.531 + ], + "angle": 0, + "content": "archical design. VideoMamba [41] applies S6 to spatiotemporal video sequences, while MambaIR [24] is the first to use S6 for image restoration, surpassing Transformer-based methods. In this work, we explore the potential of Mamba in event-based VSR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.554, + 0.182, + 0.57 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.581, + 0.357, + 0.598 + ], + "angle": 0, + "content": "3.1. Revisiting State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.699 + ], + "angle": 0, + "content": "State space models (SSM) [17, 23, 55, 64] represent a class of sequence-to-sequence modeling systems characterized by constant dynamics over time, a property also known as linear time-invariant. With linear complexity, SSM can effectively capture the inherent dynamics through an implicit mapping to latent states, which can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.715, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\ny (t) = C h (t) + D x (t), \\dot {h} (t) = A h (t) + B x (t). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Here, \\(x(t)\\in \\mathbb{R},h(t)\\in \\mathbb{R}^N\\) , and \\(y(t)\\in \\mathbb{R}\\) denotes the input, hidden state, and the output, respectively. \\(N\\) is the state size and \\(\\dot{h} (t)\\) refers to the time derivative of \\(h(t)\\) . Additionally, \\(A\\in \\mathbb{R}^{N\\times N}\\) \\(B\\in \\mathbb{R}^{N\\times 1}\\) \\(C\\in \\mathbb{R}^{1\\times N}\\) ,and \\(D\\in \\mathbb{R}\\) are the system matrices. To process discrete sequences like image and text, SSMs adopt Zero-Order Hold discretization [22] to map the input sequence \\(\\{x_{1},x_{2},\\dots,x_{K}\\}\\) to the output sequence \\(\\{y_1,y_2,\\dots,y_K\\}\\) . Specifically, suppose \\(\\Delta \\in \\mathbb{R}^{D}\\) is the pre-defined timescale parameter to map continuous parameters \\(A\\) \\(B\\) into a discrete space, the dis" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.093, + 0.891, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.431, + 0.905, + 0.46 + ], + "angle": 0, + "content": "Figure 3. Overview of the forward recurrent cell \\( \\text{For}(\\cdot) \\), the backward recurrent cell \\( \\text{Back}(\\cdot) \\), and the reconstructor \\( R(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.781, + 0.484 + ], + "angle": 0, + "content": "cretization process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.491, + 0.905, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\bar {A} = \\exp (\\Delta A), \\bar {B} = (\\Delta A) ^ {- 1} (\\exp (A) - I) \\Delta B, \\bar {C} = C, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.513, + 0.779, + 0.527 + ], + "angle": 0, + "content": "\\[\ny _ {k} = \\bar {C} h _ {k} + \\bar {D} x _ {k}, h _ {k} = \\bar {A} h _ {k - 1} + \\bar {B} x _ {k}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.905, + 0.581 + ], + "angle": 0, + "content": "Here, all the matrices keep the same dimension as the operation iterates. Notably, \\(\\overline{D}\\), serving as a residual connection, is often discarded in the equation" + }, + { + "type": "equation", + "bbox": [ + 0.674, + 0.588, + 0.905, + 0.604 + ], + "angle": 0, + "content": "\\[\ny _ {k} = \\bar {C} h _ {k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.905, + 0.643 + ], + "angle": 0, + "content": "Besides, following Mamba [21], the matrix \\(\\overline{B}\\) can be approximated by the first-order Taylor series" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.649, + 0.905, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\bar {B} = (\\exp (A) - I) A ^ {- 1} B \\approx (\\Delta A) (\\Delta A) ^ {- 1} \\Delta B = \\Delta B. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.908, + 0.72 + ], + "angle": 0, + "content": "The proposed iMamba and cMamba blocks are built upon Mamba and are specifically designed to address the challenges of event-based VSR tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.625, + 0.742 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Let \\(\\mathcal{I}^{LR} = \\{\\ldots, I_{t-1}^{LR}, I_t^{LR}, I_{t+1}^{LR}, \\ldots\\}\\) (\\(I_t^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}\\)) be a sequence of LR input frames, \\(E^{f,b}\\) be forward and backward event streams, the goal of MamEVSR is to generate \\(\\mathcal{I}^{SR} = \\{\\ldots, I_{t-1}^{SR}, I_t^{SR}, I_{t+1}^{SR}, \\ldots\\}\\) (\\(I_t^{SR} \\in \\mathbb{R}^{sH \\times sW \\times 3}\\)), which should be close to the ground-truth sequence \\(\\mathcal{I}^{GT} = \\{\\ldots, I_{t-1}^{GT}, I_t^{GT}, I_{t+1}^{GT}, \\ldots\\}\\) (\\(I_t^{GT} \\in \\mathbb{R}^{sH \\times sW \\times 3}\\)). \\(T, H,\\) and \\(W\\) are the frame number, height, and width, respectively. \\(s\\) is the upscaling factor." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Because the event streams are not convenient for observation and processing by convolutional neural networks, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12566" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "convert forward and backward event streams as voxel grids \\(\\mathbf{V}^{f,b}\\in \\mathbb{R}^{(T - 1)\\times H\\times W\\times Bin}\\) via the temporal bilinear interpolation scheme" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.141, + 0.483, + 0.176 + ], + "angle": 0, + "content": "\\[\nV (i) = \\sum_ {k} p _ {k} \\max \\left(0, 1 - \\left| (i - 1) - \\frac {t _ {k} - t _ {0}}{t _ {N _ {e}} - t _ {0}} (B i n - 1) \\right|\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.484, + 0.256 + ], + "angle": 0, + "content": "where \\( i \\in \\{1, \\dots, Bin\\} \\) represents the \\( i \\)-th time bin. In our experiments, consistent with previous studies [32, 33], we also set \\( Bin = 5 \\). Furthermore, to mitigate the impact of hot pixels, we follow the study [32, 33] and normalize the voxel grid \\( V \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.26, + 0.483, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\hat {V} (i) = \\min ((V (i), \\eta)) / \\eta , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.358 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the 98-th percentile value in the non-zero values of \\(V\\). In this way, we obtain the normalized voxel grid \\(\\hat{V} \\in \\mathbb{R}^{H \\times W \\times Bin}\\), which contains rich high-frequency textural information. We represent \\(\\pmb{V}^{f,b}\\) using this voxel grid representation for further processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Figure 2 shows an overview of the proposed MamEVSR. MamEVSR employs bidirectional recurrent cells \\( For(\\cdot) \\) and \\( Back(\\cdot) \\) akin to the scheme proposed in [9]. However, it introduces novel elements, such as extra inputs and specialized modules to harness event streams, which sets it apart from prior approaches. The event voxels and the LR frames are first converted into the feature domain using the feature encoders \\( (f_{En}^{V}) \\) and \\( f_{En}^{I} \\), consisting of \\( N_{1} \\) residual blocks), following which they are fed to the bidirectional recurrent cells. These cells consist of the iMamba block and the cMamba block. This step aims to utilize temporal correspondence from adjacent frames and aggregate cross-modal information from the event stream. As shown in Figure 3, using timestamp \\( t \\) as an example, the process can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.588, + 0.483, + 0.607 + ], + "angle": 0, + "content": "\\[\nF _ {t, f} ^ {c M}, F _ {t, f} ^ {E, c M} = \\operatorname {F o r} \\left(I _ {t - 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {f}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.611, + 0.483, + 0.63 + ], + "angle": 0, + "content": "\\[\nF _ {t, b} ^ {c M}, F _ {t, b} ^ {E, c M} = B a c k \\left(I _ {t + 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {b}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.483, + 0.801 + ], + "angle": 0, + "content": "Specifically, using the forward recurrent cell as an example, at time \\( t \\), \\( F_{t-1}^{LR} \\) and \\( F_{t}^{LR} \\) are fed to the iMamba block for efficient feature fusion and propagation across adjacent frames, resulting in \\( F_{t,f}^{LR'} \\). Next, \\( F_{t,f}^{LR'} \\) and \\( I_{t}^{LR} \\) are concatenated along the channel dimension, followed by the residual of \\( F_{t}^{LR} \\) to enhance the representation capacity. The resulting features \\( F_{t,f}^{iM} \\), along with \\( F_{t}^{f} \\), are fed to the cMamba block, which facilitates further interaction and aggregation in a cross-modality manner. The outputs of the cMamba block, \\( F_{t,f}^{cM} \\) and \\( F_{t,f}^{E,cM} \\), are used for further reconstruction. The backward cell operates similarly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.881 + ], + "angle": 0, + "content": "To generate the super-resolved output \\( I_{t}^{SR} \\) at timestamp \\( t \\), features \\( F_{t,f}^{cM}, F_{t,b}^{cM}, F_{t,f}^{E,cM}, F_{t,b}^{E,cM} \\), along with \\( I_{t}^{LR} \\) and \\( F_{t}^{LR} \\), are fed to the reconstructor, which consists of an iMamba block, \\( N_{2} \\) residual blocks, and a pixel-shuffling operation. The process can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.884, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nI _ {t} ^ {S R} = R \\left(F _ {t, f} ^ {c M}, F _ {t, b} ^ {c M}, F _ {t, f} ^ {E, c M}, F _ {t, b} ^ {E, c M}, I _ {t} ^ {L R}, F _ {t} ^ {L R}\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.75, + 0.106 + ], + "angle": 0, + "content": "3.3. Interleaved Mamba Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.326 + ], + "angle": 0, + "content": "Effective inter-frame modeling is essential for VSR [9, 75, 86]. Conventional methods [75] rely on convolutional layers and attention mechanisms to capture temporal dependencies, offering fast inference but struggling with limited receptive fields. More advanced approaches, such as SemanticLens [68], introduce semantic-aware alignment strategies to enhance feature correspondence, but at the cost of increased computational complexity. To achieve a better balance between efficiency and effectiveness, we explore Mamba [21], an SSM, for inter-frame modeling. By leveraging its long-range dependency modeling and selective state updates, Mamba enables robust temporal feature fusion while maintaining computational efficiency, making it well-suited for VSR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.909, + 0.526 + ], + "angle": 0, + "content": "As shown in Figure 4(a), we introduce the iMamba block, which follows a Transformer-like structure (i.e., Norm \\(\\rightarrow\\) Attention \\(\\rightarrow\\) Norm \\(\\rightarrow\\) MLP) [89], with two key modifications: (1) We replace the attention mechanism with a custom block that scans feature maps from two input frames. Thanks to the interleaved token reorganization design, this enables global inter-frame modeling with linear complexity. (2) Inspired by MambaIR [24] and MLFSR [19], which address limitations in locality and inter-channel interaction, we substitute the original MLP with a channel-attention block to enhance feature fusion and propagation. Here, Norm indicate the normalization operation and MLP means the multilayer perceptron." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.909, + 0.779 + ], + "angle": 0, + "content": "For inter-frame modeling, given two adjacent low-resolution frame features, \\( F_{t-1}^{LR} \\) and \\( F_{t}^{LR} \\) (in the forward direction), both are first fed into the patch embedding module. This is followed by a series of operations, including layer normalization, linear projection, depth-wise convolution, and token reorganization, resulting in the reorganized feature \\( F_{Re,f}^{LR} \\). The selective SSM-based Mamba block is then applied to independently model each direction. After processing, the sequences are rearranged and merged via token partitioning, another round of layer normalization, and linear operations. Subsequently, a channel attention block with a residual connection is applied to enhance feature propagation and stability. The output of the iMamba block is denoted as \\( F_{t,f}^{LR'} \\) for the forward direction. Similarly, for the backward direction, \\( F_{t,b}^{LR'} \\) is obtained following the same operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The interleaved reorganization within the iMamba block offers significant advantages for VSR. By ensuring that aggregated tokens come from both spatial and temporal neighborhoods, this design enhances local modeling, which is essential for capturing fine details between adjacent frames. Minimizing the distance between spatially and temporally adjacent tokens preserves the local context and reduces noise, which is particularly beneficial for maintaining tem-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12567" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.48, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.089, + 0.895, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.427, + 0.905, + 0.457 + ], + "angle": 0, + "content": "Figure 4. The structure of the proposed iMamba block and the cMamba block. The blocks offer specialized mechanisms to handle interframe modeling and cross-modal interactions effectively. See texts for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.514 + ], + "angle": 0, + "content": "poral coherence across frames. This improved context modeling directly enhances the quality of fine detail reconstruction, leading to superior video restoration results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.36, + 0.543 + ], + "angle": 0, + "content": "3.4. Cross-Modality Mamba Block" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.483, + 0.717 + ], + "angle": 0, + "content": "The fusion of event data and RGB information is crucial for event-based VSR. Event cameras capture high-temporal-resolution motion cues that complement the rich spatial details in RGB frames. By integrating this sparse, high-frequency motion data with RGB content, we can significantly improve motion estimation, edge preservation, and image quality, especially in challenging conditions like low light or fast motion. This fusion enhances temporal consistency and fine detail reconstruction, leading to superior VSR performance. We, therefore, design the cMamba block for cross-modality fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Taking the forward direction as an example, as shown in Figure 4(b), the input features from the RGB frame output of the iMamba block, \\( F_{t,f}^{iM} \\), and the event feature, \\( F_{t}^{f} \\), are first processed by linear layers and depth-wise convolutions, generating \\( F^{RGB} \\) and \\( F^{Event} \\) before being fed to the cross selective SSM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Following the Mamba selection mechanism described in Section 3.1, system matrices \\( B \\), \\( C \\), and \\( \\Delta \\) are generated to endow the model with context-aware capabilities, with linear projection layers generating these matrices. According to Eq. 2, matrix \\( C \\) decodes information from the hidden state \\( h_k \\) to produce the output \\( y_k \\). Inspired by the cross" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.905, + 0.529 + ], + "angle": 0, + "content": "attention mechanism [11], we enable cross-modal interaction between the RGB and event modalities through information exchange between multiple selective scan modules. In particular, the process can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.54, + 0.905, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\bar {A} _ {I} = \\exp \\left(\\Delta_ {\\mathrm {I}} A _ {I}\\right), \\quad \\bar {A} _ {E} = \\exp \\left(\\Delta_ {\\mathrm {E}} A _ {E}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.56, + 0.905, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\bar {B} _ {I} = \\Delta_ {\\mathrm {r g b}} B _ {I}, \\bar {B} _ {E} = \\Delta_ {\\mathrm {x}} B _ {E}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.58, + 0.905, + 0.599 + ], + "angle": 0, + "content": "\\[\nh _ {I} ^ {t} = \\bar {A} _ {I} h _ {I} ^ {t - 1} + \\bar {B} _ {I} x _ {I} ^ {t}, h _ {E} ^ {t} = \\bar {A} _ {E} h _ {E} ^ {t - 1} + \\bar {B} _ {E} x _ {E} ^ {t}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.601, + 0.905, + 0.62 + ], + "angle": 0, + "content": "\\[\ny _ {I} ^ {t} = C _ {\\times} h _ {I} ^ {t} + D _ {I} x _ {I} ^ {t}, y _ {E} ^ {t} = C _ {\\mathrm {r g b}} h _ {E} ^ {t} + D _ {E} x _ {E} ^ {t}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.622, + 0.905, + 0.641 + ], + "angle": 0, + "content": "\\[\ny _ {I} = \\left[ y _ {I} ^ {1}, y _ {I} ^ {2}, \\dots , y _ {I} ^ {l} \\right], y _ {E} = \\left[ y _ {E} ^ {1}, y _ {E} ^ {2}, \\dots , y _ {E} ^ {l} \\right]. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.905, + 0.712 + ], + "angle": 0, + "content": "Here, \\( x_{I / E}^{t} \\) represents the input at time step \\( t \\), and \\( y_{I / E} \\) denotes the selective scan output. \\( C_E \\) and \\( C_I \\) are the cross-modal matrices used for recovering the outputs at each time step from the hidden states." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.817 + ], + "angle": 0, + "content": "The outputs from the two modality branches—RGB and event—are individually fed into the layer normalization operation, followed by a residual connection. This results in the final cross-modality enhanced feature, which effectively integrates information from both branches, improving feature representation for tasks like VSR. The above process can be denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.826, + 0.905, + 0.847 + ], + "angle": 0, + "content": "\\[\nF _ {t, f} ^ {\\hat {i} M}, \\hat {F} _ {t} ^ {f} = \\operatorname {C S S M} \\left(\\mathrm {D} \\left(\\operatorname {L i n e a r} \\left(F _ {t, f} ^ {i M}\\right)\\right), \\mathrm {D} \\left(\\operatorname {L i n e a r} \\left(F _ {t} ^ {f}\\right)\\right)\\right) \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.857, + 0.905, + 0.877 + ], + "angle": 0, + "content": "\\[\nF _ {t, f} ^ {c M} = \\mathrm {L N} \\left(F _ {t, f} ^ {\\hat {i} M}\\right) + F _ {t, f} ^ {i M}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.883, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {t, f} ^ {E, c M} = \\mathrm {L N} \\left(\\hat {F} _ {t} ^ {f}\\right) + F _ {t} ^ {f}. \\tag {17}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12568" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison in terms of PSNR and SSIM for the \\( \\times 2 \\) event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined, respectively. Results are from [32] and [82]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.12, + 0.902, + 0.399 + ], + "angle": 0, + "content": "
TypeMethodCED
Scene 1Scene 2Scene 3Scene 4Scene 5Scene 6Scene 7Scene 8Scene 9Scene 10Scene 11Average
RGB-based VSRDUF-----------31.09
-----------0.9183
SOF-VSR-----------31.84
-----------0.9226
TDAN35.8332.1231.5735.7335.4237.7528.9132.5435.5530.6735.0933.74
0.95400.93390.94660.95660.95360.94400.90620.90060.95410.93230.95610.9398
RBPN40.0734.1533.8339.5639.4440.3330.3634.9140.0531.5139.0336.66
0.98680.97390.97390.98690.98590.97820.96480.95020.98780.95510.98620.9754
BasicVSR39.3539.8139.7339.6039.4542.7139.1536.9739.3538.4539.4139.57
0.97840.97660.98320.97890.97780.98150.97480.96720.97760.97320.97990.9778
Event-based VSRE-VSR41.0834.7734.4440.4940.3240.8040.8035.1641.0031.7939.9737.32
0.98910.97750.97730.98910.98800.98010.98010.95360.99780.95860.98840.9783
EGVSR38.7838.6838.6739.0638.9341.9638.0336.1438.8437.6838.8638.69
0.97940.97500.98150.97980.97920.98310.97550.96350.97870.97260.98100.9771
EBVSR39.9540.2340.3140.2240.0643.0839.9737.2439.9538.8840.0740.14
0.98110.97800.98490.98160.98050.98300.97540.96890.98020.97570.98040.9801
EvTexture40.3940.5440.7540.6640.4543.2740.5337.5740.3539.2740.5440.52
0.98240.97890.98590.98290.98190.98340.98000.97050.98150.97690.98370.9813
EvTexture+-----------40.57
-----------0.9815
MamEVSR41.1740.6541.1041.4641.3143.2741.7037.7641.1639.3741.5041.14
0.98480.97910.98660.98530.98450.98260.98440.97190.98390.97750.98630.9831
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.515 + ], + "angle": 0, + "content": "The features \\( F_{t,f}^{cM} \\) and \\( F_{t,f}^{E,cM} \\), obtained after cross-modal fusion, are fed into the reconstructor (see Figure 3) for the final image reconstruction. These fused features effectively capture both the high-frequency event data and spatial richness from the RGB frames, enhancing the quality of the reconstructed output by leveraging complementary information from both modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.53, + 0.226, + 0.547 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.554, + 0.299, + 0.57 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.576, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Datasets. For training, we utilize the REDS [56] and CED [62] datasets. Specifically, the REDS4 subset (clips 000, 011, 015, 020) from the REDS training set is employed as our test set, evaluating performance in the RGB channels. Consistent with prior work [30, 32, 33], we simulate events using the ESIM [60] tool for these clips. These simulated events are subsequently transformed into voxel grids according to Eqs. (5) and (6). Voxels are downsampled through the Bicubic interpolation operation, aligning with the frame downsampling method [32]. Following previous event-based VSR studies [33, 52], we use the CED [62] dataset for training and evaluating on real-world scenes1. The dataset is captured with a DAVIS346 [5] event camera, which outputs temporally synchronized events and frames at a resolution of \\(346 \\times 260\\). When calculating the metrics (PSNR and SSIM), we exclude boundary 8 pixels and evaluate in the RGB channel." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Implementation details. We use 15 frames as input during training, with a mini-batch size of 6 and an input frame resolution of \\(64 \\times 64\\). We apply data augmentation techniques to the training data, including horizontal flips and random rotations of \\(90^{\\circ}\\), \\(180^{\\circ}\\), and \\(270^{\\circ}\\). MamEVSR is trained for 300K iterations using the Adam optimizer with a Cosine Annealing learning rate scheduler. Network architecture parameters are set to \\(N_{1} = 2\\) and \\(N_{2} = 45\\). Supervision is provided by the Charbonnier penalty loss [35]:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.556, + 0.907, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sqrt {\\| I ^ {S R} - I ^ {G T} \\| ^ {2} + \\varepsilon^ {2}}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.658 + ], + "angle": 0, + "content": "where \\(\\varepsilon\\) is set to \\(1\\times 10^{-3}\\) in our experiments. \\(I^{SR}\\) denotes the results generated by MamEVSR, and \\(I^{GT}\\) represents the ground-truth frames. We omit the subscript \\(t\\) for simplicity. The initial learning rate for MamEVSR is \\(2\\times 10^{-4}\\). Training is conducted on 2 NVIDIA RTX 3090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.878, + 0.685 + ], + "angle": 0, + "content": "4.2. Quantitative and Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.87 + ], + "angle": 0, + "content": "We compare the proposed MamEVSR with a wide range of potential methods that could be used to address event-based VSR, aiming to explore as many diverse and rich approaches as possible. (1) RGB-based VSR methods: DUF [31], SOF-VSR [72], RBPN [25], EDVR [75], BasicVSR [9], VRT [49], and TTVSR [50]. In particular, we exclude the event stream and solely feed the LR video frames into these VSR networks for reconstruction, resulting in the final reconstructed video output. (2) Event-based VSR methods: we compare our MamEVSR with E-VSR [30], EGVSR [52], EBVSR [33], EvTexture, and EvTexture+ [32] to provide a thorough evaluation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. As illustrated in Table 1, Table 2 and Table 3, MamEVSR consistently outperforms other event" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.839, + 0.483, + 0.902 + ], + "angle": 0, + "content": "1The 11 scenes are as follows: People_Dynamic_Wave, Indoors_Foosball_2, Simple_Wires_2, People_Dynamic_Dancing, People_Dynamic_Jumping, Simple_Fruit_Fast, Outdoor_Jumping_Infrared_2, Simple_Carpet_Fast, People_Dynamic_Armroll, Indoors_Kitchen_2, and People_Dynamic_Sitting [82]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12569" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison in terms of PSNR and SSIM for the \\( \\times 4 \\) event-based VSR task on the REDS4 dataset. The best and second-best results are highlighted in bold and underlined, respectively. Results are from [32]." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.119, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
REDS4RGB-based VSREvent-based VSR
DUFEDVRBasicVSRTTVSRVRTEGVSREBVSREvTextureEvTexture+MamEVSR
00027.30/0.793728.01/0.825028.40/0.843428.82/0.856628.85/0.855325.16/0.706628.44/0.844630.72/0.9082-/-30.87/0.9103
01128.38/0.805632.17/0.886432.47/0.897933.47/0.910033.49/0.907226.56/0.772232.55/0.898733.72/0.9145-/-33.87/0.9166
01531.55/0.884634.06/0.920634.18/0.922435.01/0.932535.26/0.933229.83/0.852634.22/0.923535.06/0.9314-/-35.21/0.9335
02027.30/0.816430.09/0.888130.63/0.900031.17/0.909431.16/0.907825.94/0.784630.67/0.900931.65/0.9154-/-31.80/0.9175
Average28.63/0.825131.09/0.880031.42/0.890932.12/0.902132.19/0.900626.87/0.779031.47/0.891932.79/0.917432.93/0.919532.94/0.9195
#Params5.820.66.36.835.62.612.28.910.19.6
" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.235, + 0.333, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.342, + 0.333, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.235, + 0.404, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.288, + 0.404, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.342, + 0.403, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.395, + 0.404, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.235, + 0.547, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.288, + 0.547, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.342, + 0.547, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.395, + 0.546, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.235, + 0.618, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.288, + 0.617, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.342, + 0.617, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.395, + 0.617, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.235, + 0.69, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.288, + 0.69, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.342, + 0.69, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.395, + 0.69, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.235, + 0.764, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.288, + 0.764, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.342, + 0.764, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.395, + 0.764, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.235, + 0.834, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.288, + 0.834, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.342, + 0.834, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.395, + 0.834, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.235, + 0.905, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.288, + 0.905, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.342, + 0.905, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.395, + 0.905, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.453, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons for \\(\\times 4\\) VSR on REDS4. From left to right in sequence are patches cropped from LR, TDAN, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.485, + 0.333, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.592, + 0.333, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.485, + 0.404, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.539, + 0.404, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.592, + 0.404, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.646, + 0.404, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.485, + 0.547, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.539, + 0.547, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.592, + 0.547, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.646, + 0.547, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.485, + 0.618, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.539, + 0.617, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.592, + 0.617, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.646, + 0.617, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.485, + 0.69, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.539, + 0.69, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.592, + 0.69, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.646, + 0.69, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.485, + 0.764, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.539, + 0.764, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.592, + 0.764, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.646, + 0.764, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.485, + 0.834, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.539, + 0.834, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.592, + 0.834, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.646, + 0.834, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.485, + 0.905, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.539, + 0.905, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.592, + 0.905, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.646, + 0.905, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.7, + 0.908, + 0.73 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons for \\(\\times 4\\) VSR on CED. From left to right in sequence are patches cropped from LR, EDVR, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.738, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison in terms of PSNR and SSIM for the \\(\\times 2\\) event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.78, + 0.483, + 0.862 + ], + "angle": 0, + "content": "
MethodPSNR (↑)SSIM (↑)MethodPSNR (↑)SSIM (↑)
SOF-VSR31.840.9226EGVSR38.690.9771
TDAN33.740.9398EBVSR40.140.9801
RBPN36.660.9754EvTexture40.520.9813
BasicVSR39.570.9778EvTexture+40.570.9815
E-VSR37.320.9783MamEVSR41.140.9831
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "based VSR methods such as E-VSR, EGVSR, EBVSR, EvTexture, and EvTexture+ across various datasets like CED" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.907, + 0.892 + ], + "angle": 0, + "content": "and REDS4. For instance, on the CED dataset for the \\(\\times 2\\) event-based VSR task, MamEVSR achieves an average PSNR of \\(33.74\\mathrm{~dB}\\) and an SSIM of 0.9831, surpassing the next best method, EvTexture+, which scores \\(33.40\\mathrm{~dB}\\) and 0.9827 respectively. Compared to RGB-based VSR methods like DUF, TDAN, RBPN, and BasicVSR, MamEVSR demonstrates superior performance. On the CED dataset for the \\(\\times 2\\) event-based VSR task, MamEVSR's average PSNR of \\(33.74\\mathrm{~dB}\\) and SSIM of 0.9831 far exceed those of the top-performing RGB-based method, BasicVSR, which achieves" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12570" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.09, + 0.458, + 0.103 + ], + "angle": 0, + "content": "Table 4. Ablation study of different components on CED." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.104, + 0.481, + 0.355 + ], + "angle": 0, + "content": "
MethodCED ×2CED ×4
PSNR↑SSIM↑PSNR↑SSIM↑
OursAll residual blocks39.990.979833.530.9082
Full model41.140.983134.030.9189
Core modules(a) w/o iMamba40.170.980333.600.9109
(b) w/o cMamba40.260.980533.630.9110
(c) w/o backward cell40.070.978933.480.9076
(d) w/o forward cell40.080.978933.400.9065
iMamba(e) Residual blocks40.620.981633.680.9111
(f) Deform. conv.40.720.981933.770.9127
(g) Flow warping40.840.982133.800.9129
(h) Concat & attention40.800.982033.770.9126
(i) w/o Int. Reorg.40.990.982633.880.9140
(j) w/o Channal att.41.020.982933.940.9145
cMamba(k) Residual blocks40.690.981833.710.9114
(l) EBVSR-BCS module [33]40.770.981933.800.9139
(m) Cross-modal att. [67]40.980.982633.880.9150
(n) Concat & iMamba40.800.982033.770.9129
(o) w/o Cross SSM40.870.982533.900.9170
Recon.(p) w/o iMamba41.040.982934.000.9179
(q) All iMamba40.890.982633.940.9168
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.444 + ], + "angle": 0, + "content": "31.69 dB and 0.9226 respectively. Across all datasets and magnification factors, MamEVSR consistently delivers superior performance, showcasing its effectiveness in enhancing event-based video super-resolution tasks compared to both event-based and RGB-based VSR methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.564 + ], + "angle": 0, + "content": "Computational cost results. We calculate the number of parameters and runtime on the CED dataset. Results are shown in Table 2 and Figure 1. MamEVSR is positioned in the upper-left quadrant, reflecting an optimal trade-off between high performance and low computational cost. It achieves high-quality reconstructions (PSNR) with efficient computation and a modest parameter count, showcasing effective resource utilization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Qualitative results. We present visual comparison results on REDS4 and CED in Figure 5 and Figure 6, respectively. The LR patch appears blurry and lacks detail as we move through the SR patches; sharpness and clarity noticeably increase. MamEVSR produces results that closely resemble the ground truth, effectively restoring fine details and textures. For example, MamEVSR more accurately reconstructs wheel textures and license plate details, closely matching the ground truth in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.245, + 0.728 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.763 + ], + "angle": 0, + "content": "We conduct experiments on CED in terms of PSNR and SSIM. Results are in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Effectiveness of core components. Removing the iMamba block decreases PSNR and SSIM for both CED \\(\\times 2\\) and CED \\(\\times 4\\) tasks. Omitting the cMamba block has a smaller impact on performance. Disabling either backward or forward cells reduces PSNR and SSIM, especially for CED \\(\\times 4\\). These results confirm the importance of the iMamba and cMamba blocks, as well as the bidirectional cell structure, in achieving optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of the iMamba block. Flow warping per" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.132, + 0.905, + 0.16 + ], + "angle": 0, + "content": "Figure 7. A failure case. From left to right in sequence are patches cropped from the ground truth image, EvTexture, and MamEVSR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.173, + 0.905, + 0.248 + ], + "angle": 0, + "content": "forms better than concatenation and attention. Removing internal reorganization or attention mechanisms within the iMamba block causes slight drops in PSNR and SSIM. The iMamba block plays a critical role in improving MamEVSR's performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.248, + 0.905, + 0.352 + ], + "angle": 0, + "content": "Effectiveness of the cMamba block. Incorporating the EBVSR-CBCS module [33] marginally boosts PSNR and SSIM. Cross-modal attention [67] yields comparable performance gains. Concatenating and using iMamba shows a slight decrease in metrics. Eliminating the cross-state space modeling mechanism within the cMamba block leads to minimal reductions in PSNR and SSIM." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.354, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Effectiveness of the reconstructor. Removing the iMamba block from the reconstructor slightly improves PSNR and SSIM for both CED \\(\\times 2\\) and CED \\(\\times 4\\) tasks. However, replacing all iMamba blocks with residual blocks leads to a notable drop in performance, particularly for CED \\(\\times 4\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.766, + 0.453 + ], + "angle": 0, + "content": "4.4. Limitations and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.673 + ], + "angle": 0, + "content": "While MamEVSR demonstrates strong results in event-based VSR, some challenges remain. As shown in Figure 7, MamEVSR struggles to recover particularly fine textures due to significant information loss in RGB data and low-resolution, poorly defined event data. Potential solutions include incorporating VGG loss to generate more realistic textures, as it effectively captures fine details in restoration tasks. Alternatively, using a longer temporal sequence of both event and RGB data could provide richer, continuous information, enhancing texture recovery. Additionally, we plan to design lighter-weight modules for event-based VSR and explore extending MamEVSR to other video restoration tasks. Investigating the effectiveness of other novel architectures [58, 85] for this task would also be insightful." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.634, + 0.697 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose MamEVSR for event-based VSR. MamEVSR leverages the selective state space model to offer global receptive field coverage with linear complexity. Key components include the iMamba block for efficient bidirectional feature fusion and the cMamba block for integrating event information and capturing finer motion details. Experiments show that MamEVSR outperforms existing methods on various datasets, achieving superior quantitative and qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.899 + ], + "angle": 0, + "content": "Acknowledgment. This project is supported by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12571" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Andreas Aakerberg, Kamal Nasrollahi, and Thomas B Moeslund. Real-world super-resolution of face-images from surveillance cameras. IET Image Processing, 16(2):442-452, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.17, + 0.484, + 0.223 + ], + "angle": 0, + "content": "[2] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Yichen Wu, Lilun Deng, Yukun Cui, Tao Feng, and Shuang Xu. Taskdriven image fusion with learnable fusion loss. In CVPR, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.224, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[5] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based davis sensor. In ISCAS, pages 686-689, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[6] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[9] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.59, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.643, + 0.483, + 0.683 + ], + "angle": 0, + "content": "[11] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.684, + 0.483, + 0.737 + ], + "angle": 0, + "content": "[12] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.738, + 0.483, + 0.792 + ], + "angle": 0, + "content": "[13] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Sixiang Chen, Tian Ye, Renjing Pei, Kaiwen Zhou, Fenglong Song, and Lei Zhu. Restoreagent: Autonomous image restoration agent via multimodal large language models. NeurIPS, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.792, + 0.483, + 0.846 + ], + "angle": 0, + "content": "[14] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.885 + ], + "angle": 0, + "content": "[15] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.886, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[16] Michel Deudon, Alfredo Kalaitzis, Israel Goytom, Md Ri" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "fat Arefin, Zhichao Lin, Kris Sankaran, Vincent Michalski, Samira E Kahou, Julien Cornebise, and Yoshua Bengio. Highres-net: Recursive fusion for multi-frame super-resolution of satellite imagery. arXiv preprint arXiv:2002.06460, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.215 + ], + "angle": 0, + "content": "[17] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.216, + 0.905, + 0.255 + ], + "angle": 0, + "content": "[18] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.256, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[19] Ruisheng Gao, Zeyu Xiao, and Zhiwei Xiong. Mamba-based light field super-resolution with efficient subspace scanning. In ACCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[20] Tomio Goto, Takafumi Fukuoka, Fumiya Nagashima, Satoshi Hirano, and Masaru Sakurai. Super-resolution system for 4k-hdtv. In ICPR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.337, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[22] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.456 + ], + "angle": 0, + "content": "[23] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[24] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[25] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.552, + 0.905, + 0.592 + ], + "angle": 0, + "content": "[26] Yan Huang, Wei Wang, and Liang Wang. Bidirectional recurrent convolutional networks for multi-frame superresolution. NeurlPS, 28, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.592, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[27] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):1015-1028, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[29] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.741, + 0.905, + 0.781 + ], + "angle": 0, + "content": "[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.781, + 0.905, + 0.834 + ], + "angle": 0, + "content": "[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.835, + 0.905, + 0.875 + ], + "angle": 0, + "content": "[32] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. Eventure: Event-driven texture enhancement for video super-resolution. arXiv preprint arXiv:2406.13457, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.875, + 0.905, + 0.889 + ], + "angle": 0, + "content": "[33] Dachun Kai, Yueyi Zhang, and Xiaoyan Sun. Video super" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12572" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "resolution via event-driven temporal alignment. In ICIP, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.12, + 0.483, + 0.173 + ], + "angle": 0, + "content": "[34] Soo Ye Kim, Jeongyeon Lim, Taeyoung Na, and Munchurl Kim. 3DSRnet: Video super-resolution using 3d convolutional neural networks. arXiv preprint arXiv:1812.09079, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.174, + 0.482, + 0.214 + ], + "angle": 0, + "content": "[35] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.215, + 0.482, + 0.255 + ], + "angle": 0, + "content": "[36] Junyong Lee, Myeonghee Lee, Sunghyun Cho, and Seungyong Lee. Reference-based video super-resolution using multi-camera video triplets. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.255, + 0.482, + 0.295 + ], + "angle": 0, + "content": "[37] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[38] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[39] Bonan Li, Zicheng Zhang, Xingyi Yang, and Xinchao Wang. Focus on neighbors and know the whole: Towards consistent dense multiview text-to-image generator for 3d creation. arXiv preprint arXiv:2408.13149, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[40] Fei Li, Linfeng Zhang, Zikun Liu, Juan Lei, and Zhenbo Li. Multi-frequency representation enhancement with privilege information for video super-resolution. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.43, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[41] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[42] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.524, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[43] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-correspondence aggregation network for video super-resolution. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.565, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[44] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. NeurIPS, 37, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[45] Zhuoyuan Li, Jiacheng Li, Yao Li, Li Li, Dong Liu, and Feng Wu. In-loop filtering via trained look-up tables. In VCIP, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[46] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.713, + 0.482, + 0.767 + ], + "angle": 0, + "content": "[47] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.767, + 0.482, + 0.793 + ], + "angle": 0, + "content": "[48] Wenyi Lian and Wenjing Lian. Sliding window recurrent network for efficient video super-resolution. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.794, + 0.482, + 0.847 + ], + "angle": 0, + "content": "[49] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.848, + 0.482, + 0.886 + ], + "angle": 0, + "content": "[50] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video super-resolution. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.887, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[51] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.134, + 0.905, + 0.187 + ], + "angle": 0, + "content": "[52] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.187, + 0.905, + 0.228 + ], + "angle": 0, + "content": "[53] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xinchao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.228, + 0.905, + 0.281 + ], + "angle": 0, + "content": "[54] Yimin Luo, Liguo Zhou, Shu Wang, and Zhongyuan Wang. Video satellite imagery super resolution via convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 14(12):2398-2402, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.323 + ], + "angle": 0, + "content": "[55] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.323, + 0.905, + 0.377 + ], + "angle": 0, + "content": "[56] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[57] Tianbo Pan, Zidong Cao, and Lin Wang. Srfnet: Monocular depth estimation with fine-grained structure via spatial reliability-oriented fusion of frames and events. In ICRA, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.905, + 0.498 + ], + "angle": 0, + "content": "[58] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[59] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[60] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.566, + 0.905, + 0.605 + ], + "angle": 0, + "content": "[61] Mehdi SM, Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[62] Cedric Scheerlinck, Henri Rebecq, Timo Stoffregen, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Ced: Color event camera dataset. In CVPRW, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[63] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. NeurlPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[64] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.726, + 0.905, + 0.767 + ], + "angle": 0, + "content": "[65]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.767, + 0.905, + 0.806 + ], + "angle": 0, + "content": "[66] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.806, + 0.905, + 0.86 + ], + "angle": 0, + "content": "[67] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.861, + 0.905, + 0.888 + ], + "angle": 0, + "content": "[68] Qi Tang, Yao Zhao, Meiqin Liu, Jian Jin, and Chao Yao. Semantic lens: Instance-centric semantic alignment for video" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12573" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.321, + 0.106 + ], + "angle": 0, + "content": "super-resolution. In AAAI, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.107, + 0.483, + 0.145 + ], + "angle": 0, + "content": "[69] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.147, + 0.483, + 0.186 + ], + "angle": 0, + "content": "[70] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.187, + 0.482, + 0.226 + ], + "angle": 0, + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.227, + 0.482, + 0.268 + ], + "angle": 0, + "content": "[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through HR optical flow estimation. In ACCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[73] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[74] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.416, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[76] Yi Xiao, Xin Su, Qiangqiang Yuan, Denghong Liu, Huanfeng Shen, and Liangpei Zhang. Satellite video superresolution via multiscale deformable convolution alignment and temporal grouping projection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[77] Yi Xiao, Qiangqiang Yuan, Kui Jiang, Xianyu Jin, Jiang He, Liangpei Zhang, and Chia-wen Lin. Local-global temporal difference learning for satellite video super-resolution. arXiv preprint arXiv:2304.04421, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[78] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.578, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[79] Zeyu Xiao, Zhen Cheng, and Zhiwei Xiong. Space-time super-resolution for light field videos. IEEE Transactions on Image Processing, 32:4785-4799, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[80] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, pages 2113-2122, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[81] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACMMM, pages 2409-2418, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.699, + 0.482, + 0.738 + ], + "angle": 0, + "content": "[82] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.739, + 0.482, + 0.779 + ], + "angle": 0, + "content": "[83] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.78, + 0.482, + 0.82 + ], + "angle": 0, + "content": "[84] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.82, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[85] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. ICLR, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.885 + ], + "angle": 0, + "content": "[86] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, pages 4429-4438, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.886, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[87] Weihao Yu and Xinchao Wang. Mambaout: Do we really" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.775, + 0.106 + ], + "angle": 0, + "content": "need mamba for vision? CVPR, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.106, + 0.905, + 0.145 + ], + "angle": 0, + "content": "[88] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.147, + 0.905, + 0.2 + ], + "angle": 0, + "content": "[89] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.201, + 0.905, + 0.24 + ], + "angle": 0, + "content": "[90] Guozhen Zhang, Chuxnu Liu, Yutao Cui, Xiaotong Zhao, Kai Ma, and Limin Wang. Vfimamba: Video frame interpolation with state space models. NeurIPS, 37, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.241, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[91] Liangpei Zhang, Hongyan Zhang, Huanfeng Shen, and Pingxiang Li. A super-resolution reconstruction algorithm for surveillance images. Signal Processing, 90(3):848-859, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[92] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12574" + } + ] +] \ No newline at end of file diff --git a/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_origin.pdf b/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7e22d24856d76fc6bfdd3fdfa7d903d283ccd961 --- /dev/null +++ b/2025/Event-based Video Super-Resolution via State Space Models/c8e5af22-690b-462c-ba33-5d416cd06ebe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21f48e17ab9ad3470c7cbdec578661df95b5c182b39fa7a16c840a949085ebb1 +size 2320103 diff --git a/2025/Event-based Video Super-Resolution via State Space Models/full.md b/2025/Event-based Video Super-Resolution via State Space Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..43c89f6d67d124560f524ea10ada63c66737359d --- /dev/null +++ b/2025/Event-based Video Super-Resolution via State Space Models/full.md @@ -0,0 +1,487 @@ +# Event-based Video Super-Resolution via State Space Models + +Zeyu Xiao Xinchao Wang* + +National University of Singapore + +zeyuxiao@nus.edu.sg, xinchao@nus.edu.sg + +# Abstract + +Exploiting temporal correlations is crucial for video super-resolution (VSR). Recent approaches enhance this by incorporating event cameras. In this paper, we introduce MamEVSR, a Mamba-based network for event-based VSR that leverages the selective state space model, Mamba. MamEVSR stands out by offering global receptive field coverage with linear computational complexity, thus addressing the limitations of convolutional neural networks and Transformers. The key components of MamEVSR include: (1) The interleaved Mamba (iMamba) block, which interleaves tokens from adjacent frames and applies multidirectional selective state space modeling, enabling efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity. (2) The cross-modality Mamba (cMamba) block facilitates further interaction and aggregation between event information and the output from the iMamba block. The cMamba block can leverage complementary spatio-temporal information from both modalities and allows MamEVSR to capture finer motion details. Experimental results show that the proposed MamEVSR achieves superior performance on various datasets quantitatively and qualitatively. + +# 1. Introduction + +Video super-resolution (VSR) [18, 25-28] is a fundamental task in computer vision aimed at reconstructing high-resolution (HR) videos from low-resolution (LR) inputs. Given its broad applications in areas such as video surveillance [1, 39, 91], high-definition television [2-4, 12-15, 20, 45-47, 57, 79, 83], and satellite imagery [16, 54, 76, 77], VSR has attracted significant interest. The main challenge lies in effectively exploiting temporal information. Advanced VSR techniques often leverage temporal information through sliding window methods [29, 34, 42, 48, 53, 70, 73-75, 88], recurrent architectures [8-10, 36, 40, 61, 86], and Transformer-based designs [49, 50, 59, 63, 65, 66], + +![](images/08953d7a3671b444b9e54eff5a99af9ef529f9288c5fd5bd953a4d38496646fe.jpg) +Figure 1. Inference time and performance comparison. MamEVSR outperforms advanced methods with high efficiency. Comparisons are performed on the CED dataset. The green circles represent RGB-based VSR methods, while the yellow circles represent event-based VSR methods. + +all demonstrating significant achievements. Recently, researchers have integrated event cameras into VSR to boost performance by harnessing their high temporal resolution and dynamic range capabilities. + +Event cameras are bio-inspired sensors that asynchronously detect intensity changes at the microsecond level on a per-pixel basis, generating millions of events per second while maintaining robustness in high-dynamic-range lighting. Leveraging these advantages, recent event-based VSR methods [30, 33, 52, 81, 82] have outperformed frame-only approaches by enhancing flow estimation, temporal alignment, and cross-modal fusion using event streams. However, current methods, primarily based on CNNs and attention models, face limitations, including poor adaptability to dynamic inputs, insufficient receptive fields for capturing inter-frame correlations at high resolutions, and high computational costs. + +On the other hand, natural language processing (NLP) has recently witnessed the emergence of structured state space models (SSMs) [22]. Theoretically, SSMs combine the benefits of recurrent neural networks (RNNs) and + +CNNs, leveraging the global receptive field characteristic of RNNs and the computational efficiency of CNNs. One particularly notable SSM is the selective state space model, also known as Mamba [21], which has garnered significant attention within the vision community. Mamba's key feature is its ability to make SSM parameters time-variant (i.e., data-dependent), enabling it to effectively select relevant context within sequences—a crucial factor for enhancing model performance. Inspired by Mamba's capabilities, we explore its potential in addressing the task of Event-based VSR, and we propose the first Mamba-based method for this task, named MamEVSR. + +MamEVSR leverages a bidirectional recurrent pipeline to recurrently align and propagate temporally correlated features and integrate cross-modal information from the event stream. Its key components include: (1) The interleaved Mamba (iMamba) block: To facilitate efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, we introduce the iMamba block. Since the original SSMs model processes a single sequence, we merge tokens from two adjacent frames into one sequence for effective inter-frame modeling. By interleaving tokens from both frames and conducting multi-directional SSMs, we enable interactions between adjacent tokens from different propagated frames during sequence modeling. This design ensures that intermediate tokens in the sequence are from their spatio-temporal neighborhood. Stacking multiple iMamba blocks enhances MamEVSR's ability to handle complex inter-frame information exchange and frame alignment. (2) The cross-modality Mamba (cMamba) block: The cMamba block employs a cross-multiplication mechanism to enhance the interaction and aggregation between event information and the output from the iMamba block. This mechanism allows for the effective integration of event data with the temporal-fused features, thereby enriching the model's ability to leverage complementary cross-modal information and improve the overall reconstruction quality. By combining these two blocks, MamEVSR effectively exploits the strengths of both Mamba and event data to achieve superior performance in event-based VSR tasks. Experimental results on benchmark datasets demonstrate the effectiveness of our proposed MamEVSR (see an example in Figure 1). + +Our contributions can be summarized as follows: (1) To the best of our knowledge, this marks the first successful application of state space models, specifically Mamba, in event-based VSR. (2) We propose the iMamba block and the cMamba block as core components in MamEVSR for event-based VSR. The former enables efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, the latter leverages complementary spatio-temporal information to capture finer motion details. (3) Extensive evaluations on benchmark + +datasets demonstrate MamEVSR's superior performance, establishing a new benchmark for future explorations of Mamba's potential within event-based VSR. + +# 2. Related Work + +Video super-resolution. Existing VSR methods are designed to enhance the quality of LR frames by leveraging temporal information through sliding windows and recurrent structures. Sliding-window-based methods align adjacent LR frames with a reference frame to estimate an HR output. Early work focused on explicit optical flow estimation for frame alignment [6, 53, 69, 80, 84]. More recent methods have shifted to implicit alignment using dynamic filters [31], deformable convolutions [70, 75], and attention modules [28, 43]. Recent attention-based methods achieve new state-of-the-art performance [7, 37, 38, 43, 49, 53, 75, 78, 80]. However, these methods face challenges in capturing long-distance temporal features. Recurrent-based methods use RNNs to exploit temporal information across multiple frames. Sajjadi et al. [61] introduce a recurrent framework with optical flow for alignment. BasicVSR [9] and BasicVSR++[10] utilize bi-directional hidden states and advanced grid propagation techniques. PSRT [63] builds on BasicVSR++ with multi-frame self-attention for feature processing. In this paper, we focus on event-based VSR and introduce MamEVSR. + +Event-based VSR. Event cameras have the unique capability to measure intensity changes at each pixel independently with microsecond accuracy, making them valuable for VSR. Jing et al. [30] propose a two-stage method that uses events to interpolate low-resolution (LR) video, generating a high-frequency video for subsequent high-resolution (HR) frame reconstruction. Lu et al. [52] present a joint framework learning implicit neural representations from both RGB frames and events, enabling arbitrary-scale VSR. Kai et al. [32] introduce EvTexture, an event-driven texture enhancement network dedicated to restoring textures in VSR through the incorporation of high-frequency event signals. Xiao et al. [82] take inspiration from the parameter-efficient tuning, introducing an event adapter for event-based VSR. In this paper, we introduce MamEVSR, the first Mamba-based method for event-based VSR. + +State space models. SSMs [17, 23, 44, 55, 64, 87, 90] are emerging as strong alternatives to Transformers [71] in natural language processing. S4 [23] is introduced for efficient, linear-complexity sequence modeling, followed by S5 [64], which improves parallelization, and GSS [55], which incorporates gated mechanisms. Mamba (S6) [21] stands out with its data-dependent parameters and hardware efficiency, outperforming Transformers in long-sequence tasks. In the computer vision domain, Vim [92] permutes 2D images into sequences for global modeling using bidirectional SSMs. Vmamba [51] extends this to four directions with a hier- + +![](images/f559444258a2bb6522a59d852909b5f0fcad5cff72c153cafcd211f9904f650c.jpg) +Figure 2. Overview of the proposed MamEVSR. MamEVSR follows a typical bi-directional propagation scheme [9], consisting of bi-directional recurrent cells $For(\cdot)$ and $Back(\cdot)$ . For clarity, we have omitted the residual connection from the bicubic upsampling of the input, which is added to the final output, in the figure. + +archical design. VideoMamba [41] applies S6 to spatiotemporal video sequences, while MambaIR [24] is the first to use S6 for image restoration, surpassing Transformer-based methods. In this work, we explore the potential of Mamba in event-based VSR. + +# 3. Method + +# 3.1. Revisiting State Space Models + +State space models (SSM) [17, 23, 55, 64] represent a class of sequence-to-sequence modeling systems characterized by constant dynamics over time, a property also known as linear time-invariant. With linear complexity, SSM can effectively capture the inherent dynamics through an implicit mapping to latent states, which can be defined as + +$$ +y (t) = C h (t) + D x (t), \dot {h} (t) = A h (t) + B x (t). \tag {1} +$$ + +Here, $x(t)\in \mathbb{R},h(t)\in \mathbb{R}^N$ , and $y(t)\in \mathbb{R}$ denotes the input, hidden state, and the output, respectively. $N$ is the state size and $\dot{h} (t)$ refers to the time derivative of $h(t)$ . Additionally, $A\in \mathbb{R}^{N\times N}$ $B\in \mathbb{R}^{N\times 1}$ $C\in \mathbb{R}^{1\times N}$ ,and $D\in \mathbb{R}$ are the system matrices. To process discrete sequences like image and text, SSMs adopt Zero-Order Hold discretization [22] to map the input sequence $\{x_{1},x_{2},\dots,x_{K}\}$ to the output sequence $\{y_1,y_2,\dots,y_K\}$ . Specifically, suppose $\Delta \in \mathbb{R}^{D}$ is the pre-defined timescale parameter to map continuous parameters $A$ $B$ into a discrete space, the dis + +![](images/1f93695cb651ecb14c2405b43db020db1592d617abdfea9b0c7d10bcf73d6a19.jpg) +Figure 3. Overview of the forward recurrent cell $\text{For}(\cdot)$ , the backward recurrent cell $\text{Back}(\cdot)$ , and the reconstructor $R(\cdot)$ . + +cretization process can be formulated as + +$$ +\bar {A} = \exp (\Delta A), \bar {B} = (\Delta A) ^ {- 1} (\exp (A) - I) \Delta B, \bar {C} = C, \tag {2} +$$ + +$$ +y _ {k} = \bar {C} h _ {k} + \bar {D} x _ {k}, h _ {k} = \bar {A} h _ {k - 1} + \bar {B} x _ {k}. +$$ + +Here, all the matrices keep the same dimension as the operation iterates. Notably, $\overline{D}$ , serving as a residual connection, is often discarded in the equation + +$$ +y _ {k} = \bar {C} h _ {k}. \tag {3} +$$ + +Besides, following Mamba [21], the matrix $\overline{B}$ can be approximated by the first-order Taylor series + +$$ +\bar {B} = (\exp (A) - I) A ^ {- 1} B \approx (\Delta A) (\Delta A) ^ {- 1} \Delta B = \Delta B. \tag {4} +$$ + +The proposed iMamba and cMamba blocks are built upon Mamba and are specifically designed to address the challenges of event-based VSR tasks. + +# 3.2. Overview + +Let $\mathcal{I}^{LR} = \{\ldots, I_{t-1}^{LR}, I_t^{LR}, I_{t+1}^{LR}, \ldots\}$ ( $I_t^{LR} \in \mathbb{R}^{H \times W \times 3}$ ) be a sequence of LR input frames, $E^{f,b}$ be forward and backward event streams, the goal of MamEVSR is to generate $\mathcal{I}^{SR} = \{\ldots, I_{t-1}^{SR}, I_t^{SR}, I_{t+1}^{SR}, \ldots\}$ ( $I_t^{SR} \in \mathbb{R}^{sH \times sW \times 3}$ ), which should be close to the ground-truth sequence $\mathcal{I}^{GT} = \{\ldots, I_{t-1}^{GT}, I_t^{GT}, I_{t+1}^{GT}, \ldots\}$ ( $I_t^{GT} \in \mathbb{R}^{sH \times sW \times 3}$ ). $T, H,$ and $W$ are the frame number, height, and width, respectively. $s$ is the upscaling factor. + +Because the event streams are not convenient for observation and processing by convolutional neural networks, we + +convert forward and backward event streams as voxel grids $\mathbf{V}^{f,b}\in \mathbb{R}^{(T - 1)\times H\times W\times Bin}$ via the temporal bilinear interpolation scheme + +$$ +V (i) = \sum_ {k} p _ {k} \max \left(0, 1 - \left| (i - 1) - \frac {t _ {k} - t _ {0}}{t _ {N _ {e}} - t _ {0}} (B i n - 1) \right|\right), \tag {5} +$$ + +where $i \in \{1, \dots, Bin\}$ represents the $i$ -th time bin. In our experiments, consistent with previous studies [32, 33], we also set $Bin = 5$ . Furthermore, to mitigate the impact of hot pixels, we follow the study [32, 33] and normalize the voxel grid $V$ as + +$$ +\hat {V} (i) = \min ((V (i), \eta)) / \eta , \tag {6} +$$ + +where $\eta$ is the 98-th percentile value in the non-zero values of $V$ . In this way, we obtain the normalized voxel grid $\hat{V} \in \mathbb{R}^{H \times W \times Bin}$ , which contains rich high-frequency textural information. We represent $\pmb{V}^{f,b}$ using this voxel grid representation for further processing. + +Figure 2 shows an overview of the proposed MamEVSR. MamEVSR employs bidirectional recurrent cells $For(\cdot)$ and $Back(\cdot)$ akin to the scheme proposed in [9]. However, it introduces novel elements, such as extra inputs and specialized modules to harness event streams, which sets it apart from prior approaches. The event voxels and the LR frames are first converted into the feature domain using the feature encoders $(f_{En}^{V})$ and $f_{En}^{I}$ , consisting of $N_{1}$ residual blocks), following which they are fed to the bidirectional recurrent cells. These cells consist of the iMamba block and the cMamba block. This step aims to utilize temporal correspondence from adjacent frames and aggregate cross-modal information from the event stream. As shown in Figure 3, using timestamp $t$ as an example, the process can be expressed as + +$$ +F _ {t, f} ^ {c M}, F _ {t, f} ^ {E, c M} = \operatorname {F o r} \left(I _ {t - 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {f}\right), \tag {7} +$$ + +$$ +F _ {t, b} ^ {c M}, F _ {t, b} ^ {E, c M} = B a c k \left(I _ {t + 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {b}\right). \tag {8} +$$ + +Specifically, using the forward recurrent cell as an example, at time $t$ , $F_{t-1}^{LR}$ and $F_{t}^{LR}$ are fed to the iMamba block for efficient feature fusion and propagation across adjacent frames, resulting in $F_{t,f}^{LR'}$ . Next, $F_{t,f}^{LR'}$ and $I_{t}^{LR}$ are concatenated along the channel dimension, followed by the residual of $F_{t}^{LR}$ to enhance the representation capacity. The resulting features $F_{t,f}^{iM}$ , along with $F_{t}^{f}$ , are fed to the cMamba block, which facilitates further interaction and aggregation in a cross-modality manner. The outputs of the cMamba block, $F_{t,f}^{cM}$ and $F_{t,f}^{E,cM}$ , are used for further reconstruction. The backward cell operates similarly. + +To generate the super-resolved output $I_{t}^{SR}$ at timestamp $t$ , features $F_{t,f}^{cM}, F_{t,b}^{cM}, F_{t,f}^{E,cM}, F_{t,b}^{E,cM}$ , along with $I_{t}^{LR}$ and $F_{t}^{LR}$ , are fed to the reconstructor, which consists of an iMamba block, $N_{2}$ residual blocks, and a pixel-shuffling operation. The process can be expressed as + +$$ +I _ {t} ^ {S R} = R \left(F _ {t, f} ^ {c M}, F _ {t, b} ^ {c M}, F _ {t, f} ^ {E, c M}, F _ {t, b} ^ {E, c M}, I _ {t} ^ {L R}, F _ {t} ^ {L R}\right). \tag {9} +$$ + +# 3.3. Interleaved Mamba Block + +Effective inter-frame modeling is essential for VSR [9, 75, 86]. Conventional methods [75] rely on convolutional layers and attention mechanisms to capture temporal dependencies, offering fast inference but struggling with limited receptive fields. More advanced approaches, such as SemanticLens [68], introduce semantic-aware alignment strategies to enhance feature correspondence, but at the cost of increased computational complexity. To achieve a better balance between efficiency and effectiveness, we explore Mamba [21], an SSM, for inter-frame modeling. By leveraging its long-range dependency modeling and selective state updates, Mamba enables robust temporal feature fusion while maintaining computational efficiency, making it well-suited for VSR. + +As shown in Figure 4(a), we introduce the iMamba block, which follows a Transformer-like structure (i.e., Norm $\rightarrow$ Attention $\rightarrow$ Norm $\rightarrow$ MLP) [89], with two key modifications: (1) We replace the attention mechanism with a custom block that scans feature maps from two input frames. Thanks to the interleaved token reorganization design, this enables global inter-frame modeling with linear complexity. (2) Inspired by MambaIR [24] and MLFSR [19], which address limitations in locality and inter-channel interaction, we substitute the original MLP with a channel-attention block to enhance feature fusion and propagation. Here, Norm indicate the normalization operation and MLP means the multilayer perceptron. + +For inter-frame modeling, given two adjacent low-resolution frame features, $F_{t-1}^{LR}$ and $F_{t}^{LR}$ (in the forward direction), both are first fed into the patch embedding module. This is followed by a series of operations, including layer normalization, linear projection, depth-wise convolution, and token reorganization, resulting in the reorganized feature $F_{Re,f}^{LR}$ . The selective SSM-based Mamba block is then applied to independently model each direction. After processing, the sequences are rearranged and merged via token partitioning, another round of layer normalization, and linear operations. Subsequently, a channel attention block with a residual connection is applied to enhance feature propagation and stability. The output of the iMamba block is denoted as $F_{t,f}^{LR'}$ for the forward direction. Similarly, for the backward direction, $F_{t,b}^{LR'}$ is obtained following the same operations. + +The interleaved reorganization within the iMamba block offers significant advantages for VSR. By ensuring that aggregated tokens come from both spatial and temporal neighborhoods, this design enhances local modeling, which is essential for capturing fine details between adjacent frames. Minimizing the distance between spatially and temporally adjacent tokens preserves the local context and reduces noise, which is particularly beneficial for maintaining tem- + +![](images/2836ea4a022aa052f45f9cee90119e22c9bc40937e05d5fc13891b8ab2dc9c31.jpg) +Figure 4. The structure of the proposed iMamba block and the cMamba block. The blocks offer specialized mechanisms to handle interframe modeling and cross-modal interactions effectively. See texts for more details. + +![](images/b50cc4799de8371bcf7e460b3612aa8c80b12f87f87903115cf41fac9b0a468b.jpg) + +poral coherence across frames. This improved context modeling directly enhances the quality of fine detail reconstruction, leading to superior video restoration results. + +# 3.4. Cross-Modality Mamba Block + +The fusion of event data and RGB information is crucial for event-based VSR. Event cameras capture high-temporal-resolution motion cues that complement the rich spatial details in RGB frames. By integrating this sparse, high-frequency motion data with RGB content, we can significantly improve motion estimation, edge preservation, and image quality, especially in challenging conditions like low light or fast motion. This fusion enhances temporal consistency and fine detail reconstruction, leading to superior VSR performance. We, therefore, design the cMamba block for cross-modality fusion. + +Taking the forward direction as an example, as shown in Figure 4(b), the input features from the RGB frame output of the iMamba block, $F_{t,f}^{iM}$ , and the event feature, $F_{t}^{f}$ , are first processed by linear layers and depth-wise convolutions, generating $F^{RGB}$ and $F^{Event}$ before being fed to the cross selective SSM. + +Following the Mamba selection mechanism described in Section 3.1, system matrices $B$ , $C$ , and $\Delta$ are generated to endow the model with context-aware capabilities, with linear projection layers generating these matrices. According to Eq. 2, matrix $C$ decodes information from the hidden state $h_k$ to produce the output $y_k$ . Inspired by the cross + +attention mechanism [11], we enable cross-modal interaction between the RGB and event modalities through information exchange between multiple selective scan modules. In particular, the process can be represented as + +$$ +\bar {A} _ {I} = \exp \left(\Delta_ {\mathrm {I}} A _ {I}\right), \quad \bar {A} _ {E} = \exp \left(\Delta_ {\mathrm {E}} A _ {E}\right), \tag {10} +$$ + +$$ +\bar {B} _ {I} = \Delta_ {\mathrm {r g b}} B _ {I}, \bar {B} _ {E} = \Delta_ {\mathrm {x}} B _ {E}, \tag {11} +$$ + +$$ +h _ {I} ^ {t} = \bar {A} _ {I} h _ {I} ^ {t - 1} + \bar {B} _ {I} x _ {I} ^ {t}, h _ {E} ^ {t} = \bar {A} _ {E} h _ {E} ^ {t - 1} + \bar {B} _ {E} x _ {E} ^ {t}, \tag {12} +$$ + +$$ +y _ {I} ^ {t} = C _ {\times} h _ {I} ^ {t} + D _ {I} x _ {I} ^ {t}, y _ {E} ^ {t} = C _ {\mathrm {r g b}} h _ {E} ^ {t} + D _ {E} x _ {E} ^ {t}, \tag {13} +$$ + +$$ +y _ {I} = \left[ y _ {I} ^ {1}, y _ {I} ^ {2}, \dots , y _ {I} ^ {l} \right], y _ {E} = \left[ y _ {E} ^ {1}, y _ {E} ^ {2}, \dots , y _ {E} ^ {l} \right]. \tag {14} +$$ + +Here, $x_{I / E}^{t}$ represents the input at time step $t$ , and $y_{I / E}$ denotes the selective scan output. $C_E$ and $C_I$ are the cross-modal matrices used for recovering the outputs at each time step from the hidden states. + +The outputs from the two modality branches—RGB and event—are individually fed into the layer normalization operation, followed by a residual connection. This results in the final cross-modality enhanced feature, which effectively integrates information from both branches, improving feature representation for tasks like VSR. The above process can be denoted as + +$$ +F _ {t, f} ^ {\hat {i} M}, \hat {F} _ {t} ^ {f} = \operatorname {C S S M} \left(\mathrm {D} \left(\operatorname {L i n e a r} \left(F _ {t, f} ^ {i M}\right)\right), \mathrm {D} \left(\operatorname {L i n e a r} \left(F _ {t} ^ {f}\right)\right)\right) \tag {15} +$$ + +$$ +F _ {t, f} ^ {c M} = \mathrm {L N} \left(F _ {t, f} ^ {\hat {i} M}\right) + F _ {t, f} ^ {i M}, \tag {16} +$$ + +$$ +F _ {t, f} ^ {E, c M} = \mathrm {L N} \left(\hat {F} _ {t} ^ {f}\right) + F _ {t} ^ {f}. \tag {17} +$$ + +Table 1. Quantitative comparison in terms of PSNR and SSIM for the $\times 2$ event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined, respectively. Results are from [32] and [82]. + +
TypeMethodCED
Scene 1Scene 2Scene 3Scene 4Scene 5Scene 6Scene 7Scene 8Scene 9Scene 10Scene 11Average
RGB-based VSRDUF-----------31.09
-----------0.9183
SOF-VSR-----------31.84
-----------0.9226
TDAN35.8332.1231.5735.7335.4237.7528.9132.5435.5530.6735.0933.74
0.95400.93390.94660.95660.95360.94400.90620.90060.95410.93230.95610.9398
RBPN40.0734.1533.8339.5639.4440.3330.3634.9140.0531.5139.0336.66
0.98680.97390.97390.98690.98590.97820.96480.95020.98780.95510.98620.9754
BasicVSR39.3539.8139.7339.6039.4542.7139.1536.9739.3538.4539.4139.57
0.97840.97660.98320.97890.97780.98150.97480.96720.97760.97320.97990.9778
Event-based VSRE-VSR41.0834.7734.4440.4940.3240.8040.8035.1641.0031.7939.9737.32
0.98910.97750.97730.98910.98800.98010.98010.95360.99780.95860.98840.9783
EGVSR38.7838.6838.6739.0638.9341.9638.0336.1438.8437.6838.8638.69
0.97940.97500.98150.97980.97920.98310.97550.96350.97870.97260.98100.9771
EBVSR39.9540.2340.3140.2240.0643.0839.9737.2439.9538.8840.0740.14
0.98110.97800.98490.98160.98050.98300.97540.96890.98020.97570.98040.9801
EvTexture40.3940.5440.7540.6640.4543.2740.5337.5740.3539.2740.5440.52
0.98240.97890.98590.98290.98190.98340.98000.97050.98150.97690.98370.9813
EvTexture+-----------40.57
-----------0.9815
MamEVSR41.1740.6541.1041.4641.3143.2741.7037.7641.1639.3741.5041.14
0.98480.97910.98660.98530.98450.98260.98440.97190.98390.97750.98630.9831
+ +The features $F_{t,f}^{cM}$ and $F_{t,f}^{E,cM}$ , obtained after cross-modal fusion, are fed into the reconstructor (see Figure 3) for the final image reconstruction. These fused features effectively capture both the high-frequency event data and spatial richness from the RGB frames, enhancing the quality of the reconstructed output by leveraging complementary information from both modalities. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. For training, we utilize the REDS [56] and CED [62] datasets. Specifically, the REDS4 subset (clips 000, 011, 015, 020) from the REDS training set is employed as our test set, evaluating performance in the RGB channels. Consistent with prior work [30, 32, 33], we simulate events using the ESIM [60] tool for these clips. These simulated events are subsequently transformed into voxel grids according to Eqs. (5) and (6). Voxels are downsampled through the Bicubic interpolation operation, aligning with the frame downsampling method [32]. Following previous event-based VSR studies [33, 52], we use the CED [62] dataset for training and evaluating on real-world scenes1. The dataset is captured with a DAVIS346 [5] event camera, which outputs temporally synchronized events and frames at a resolution of $346 \times 260$ . When calculating the metrics (PSNR and SSIM), we exclude boundary 8 pixels and evaluate in the RGB channel. + +Implementation details. We use 15 frames as input during training, with a mini-batch size of 6 and an input frame resolution of $64 \times 64$ . We apply data augmentation techniques to the training data, including horizontal flips and random rotations of $90^{\circ}$ , $180^{\circ}$ , and $270^{\circ}$ . MamEVSR is trained for 300K iterations using the Adam optimizer with a Cosine Annealing learning rate scheduler. Network architecture parameters are set to $N_{1} = 2$ and $N_{2} = 45$ . Supervision is provided by the Charbonnier penalty loss [35]: + +$$ +\mathcal {L} = \sqrt {\| I ^ {S R} - I ^ {G T} \| ^ {2} + \varepsilon^ {2}}, \tag {18} +$$ + +where $\varepsilon$ is set to $1\times 10^{-3}$ in our experiments. $I^{SR}$ denotes the results generated by MamEVSR, and $I^{GT}$ represents the ground-truth frames. We omit the subscript $t$ for simplicity. The initial learning rate for MamEVSR is $2\times 10^{-4}$ . Training is conducted on 2 NVIDIA RTX 3090 GPUs. + +# 4.2. Quantitative and Qualitative Comparisons + +We compare the proposed MamEVSR with a wide range of potential methods that could be used to address event-based VSR, aiming to explore as many diverse and rich approaches as possible. (1) RGB-based VSR methods: DUF [31], SOF-VSR [72], RBPN [25], EDVR [75], BasicVSR [9], VRT [49], and TTVSR [50]. In particular, we exclude the event stream and solely feed the LR video frames into these VSR networks for reconstruction, resulting in the final reconstructed video output. (2) Event-based VSR methods: we compare our MamEVSR with E-VSR [30], EGVSR [52], EBVSR [33], EvTexture, and EvTexture+ [32] to provide a thorough evaluation. + +Quantitative results. As illustrated in Table 1, Table 2 and Table 3, MamEVSR consistently outperforms other event + +Table 2. Quantitative comparison in terms of PSNR and SSIM for the $\times 4$ event-based VSR task on the REDS4 dataset. The best and second-best results are highlighted in bold and underlined, respectively. Results are from [32]. + +
REDS4RGB-based VSREvent-based VSR
DUFEDVRBasicVSRTTVSRVRTEGVSREBVSREvTextureEvTexture+MamEVSR
00027.30/0.793728.01/0.825028.40/0.843428.82/0.856628.85/0.855325.16/0.706628.44/0.844630.72/0.9082-/-30.87/0.9103
01128.38/0.805632.17/0.886432.47/0.897933.47/0.910033.49/0.907226.56/0.772232.55/0.898733.72/0.9145-/-33.87/0.9166
01531.55/0.884634.06/0.920634.18/0.922435.01/0.932535.26/0.933229.83/0.852634.22/0.923535.06/0.9314-/-35.21/0.9335
02027.30/0.816430.09/0.888130.63/0.900031.17/0.909431.16/0.907825.94/0.784630.67/0.900931.65/0.9154-/-31.80/0.9175
Average28.63/0.825131.09/0.880031.42/0.890932.12/0.902132.19/0.900626.87/0.779031.47/0.891932.79/0.917432.93/0.919532.94/0.9195
#Params5.820.66.36.835.62.612.28.910.19.6
+ +![](images/9d09f0b4873329567a6631b1831357bbcd55512330fc3c6057fc8dc75609ceba.jpg) + +![](images/acb116899ce1d040d36f1807e8f5c0cdc161b841509c2ad5fb61c46ae5dc0362.jpg) +Figure 5. Visual comparisons for $\times 4$ VSR on REDS4. From left to right in sequence are patches cropped from LR, TDAN, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image. + +![](images/d7991cf06befd2f8de94e9dcf0d36209e4ff71c1b0d0d63f3d467eaedbf72ce9.jpg) + +![](images/33c278d1f953fa791c490fe92e841118d663045fd4fa54d33c5f2bbf5d338463.jpg) + +![](images/e3b6f614ee88f573044a9a4d12f25e01d2b1265022f1b7d4df9d325be36918a1.jpg) + +![](images/5c974e2a90e3dac74bc19b2661dff6873ebb2e3bcd2185415cde594dc21c056a.jpg) + +![](images/08e78b722e96c5c7de965787fbfa48fb6f95c7973a988984db95b6f37b63b7f9.jpg) + +![](images/7c19d5f18e24148c9920d875daa67e8dd4749ce411b8c9c59d70bbe61d0b7623.jpg) + +![](images/274a0b9ad8a11145a2d3e0cbf5b386a90a86a964ec8fab1262a9db97f1598ae5.jpg) + +![](images/481b0805ce2ff67a11ec08687bee33344b13ddb6cb9dbc2492c39e1e4d497cb5.jpg) + +![](images/e839538222076ff7fcfd49928461bce481e321d351445d5e46427d3554540b1a.jpg) + +![](images/7864f660d46bbe2f75f4caeef43f336e4c50150b243e3a98a3f57a2aa988700e.jpg) + +![](images/1a4a928c77448541a3665ae043b34721438641bb9823c3e862dca9f502aa18e6.jpg) + +![](images/1491547083335d285b5f87437d1cced10df9fb56bd261776a0019e12905f9d35.jpg) + +![](images/9d9cdee591e61a97d39062ff1a1f4631a33b31cc4f9eca7599220cf5f4666925.jpg) + +![](images/b56ab0024456ca8c96a224ad062a8f9c55db70d862b8ab7ee40c9c13ae05a381.jpg) + +![](images/3bf4793a95ee4230457673546d026f8b75ca635b0df05502351266a215c3a43a.jpg) + +![](images/7febd35bd4926451a599b8aa85a0ca321743bfe8e1ccf8ea78e92ef796b5fc76.jpg) + +![](images/9495be13687f1410951638dc0362e526cc4cc404b42e11318998e941159e9b1d.jpg) + +![](images/fe36963c897234235fceccc9fc044b7dbeedee7ee578e56e82eaa968365e3d64.jpg) + +![](images/a404da58f7b9255a2172f5bed087058ac7eb07646e8b1e6d6f0b3b97f54089fd.jpg) + +![](images/5dfa976a0da885310210f1f128044dd7028d06ab4ddb2fb9a0d6b3b970343cb8.jpg) + +![](images/f169570344f9fbaef314c6f1efe56c1c7c122f5b2936c1e1cec17b1654d12729.jpg) + +![](images/01141179cede12f13f66f286eb3fd08d78bf584290e438bd911084855ad721a8.jpg) + +![](images/442e0a920102c0a9fc5c09875549abfd85cce755050f77141e2fe92b80fe09b9.jpg) + +![](images/2df144de1dff438f05e90dabd2364b5c85af776a1cfbae4aac2c8cb951de601b.jpg) + +![](images/fa15dfae37e40bdd2bed3073f7f7f29ca317e0c05d1f9167cfec265ff325e715.jpg) + +![](images/c346450f86b2fb5d9f1aecdb26141346b116b074fb7a09a0acfcc412cb1dfcb9.jpg) + +![](images/654a0eeffa6f17e0043ec269032565e925e93c42e3ee79c8d92b07e6678691f7.jpg) + +![](images/16643d135eced12494e57b059d5d25dd09bb5c42aa634b289436dbd900cada6b.jpg) + +![](images/94a772be7c3395a64bcb14289d31f1d6dbb0ef0a216177b74d36627784a2a539.jpg) + +![](images/c346c0717b69d01a3537414919f01da01b0fcc215162aceaca3c03e92435bc8b.jpg) +Figure 6. Visual comparisons for $\times 4$ VSR on CED. From left to right in sequence are patches cropped from LR, EDVR, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image. + +![](images/6ccd19546dd981c636109863fb5c163c0cd0f41289c1a10ae56bbe7359304392.jpg) + +![](images/05983bf7b295d1de5a4aad09226b83200a5836ba4eec53d7ca13cc511c730ce5.jpg) + +![](images/6b3f49cb3c9347a414f6c44e852d192f43ac13d4ca0a37f2c9dd55a1f74ed4a9.jpg) + +![](images/8530c1e3c29498bcc4609e0cb8072208b1b12387bd83286a7e46cea0db63f21f.jpg) + +![](images/dc02fe49fcae1b52d440280275d41b7ec7abd62052f2eb39b6b48f8f756aa72a.jpg) + +![](images/b955f09775b530f34c6236508546e55c76642dcacd7b987465e8ff50a7700604.jpg) + +![](images/5d14e9942c10956e3646a2bf7bc935e552cfa7d5dfeecf68e68f1905c0761254.jpg) + +![](images/c62741f328207a46197dbd61504fe22cf5df6895b3be80ffa2720848618a4407.jpg) + +![](images/5201d68d23065a7d1ae33c7c562cb1b44cb29fbdab678e1c5c1eb10b3d188408.jpg) + +![](images/02892673f4523dd91fa2aa656f6a01afa9ce49ca2cd603ebc6767f013544538e.jpg) + +![](images/809d63e656816cf6a98f42fe5ad94f3542dfa339348882f963ac846c57d43aa4.jpg) + +![](images/86993b96e0cbc6c7f521ab06b1da878e016919d367a7c3abd9463d6f28af2a10.jpg) + +![](images/4641407073ba7cdafbfe059953aa0419b4e310cfc4eb03eefb878f386862038a.jpg) + +![](images/6003d5f74d7fb1006fdbfc8bfcd38fe1439ad43c44d54e255f150c22804f852e.jpg) + +![](images/b5d558d88d6462acc8ee4d4f557908d87eded0a6f55f3c1e80b63b9842782fb4.jpg) + +![](images/5686ac1d5cf5d764873550746047bf2dc0b7815c2040beca2b54407e71289e47.jpg) + +![](images/2b4bb252085922e1413b3f1e5f22b5152a1e52e1d13f1baf40998891261de11f.jpg) + +![](images/cb858399e70c51c794427408ae7c43c93bf2318cbd5e7e145dbe717712724cd8.jpg) + +![](images/27c655a80be0105c3595c129c8c937cc7aff8a8180db6405a255adefa66015a2.jpg) + +![](images/947374c1fc3a132118133a376e87a715de38f63969f683fdbf4de2991ff5753b.jpg) + +![](images/2909f6e655391dfb8e22b338ad9c070834ee5d08fa3d32014d30b9191f2762b4.jpg) + +![](images/c4acf297317579deb35408eeb69164aa825701d0b72147b7b3df5d2cc0156b11.jpg) + +![](images/a342a2158f94464fa24fe7b44b548a8c72ecaab09ce91cd87c559e6d2378c063.jpg) + +![](images/3ef2e896ff08d19704d0de4fd00048e088dc9327f843de66482bdab0cff2a9fc.jpg) + +![](images/a48b88e8440db91fc32ba291535049f0c3983ed632067d7306351b69bfb3b178.jpg) + +![](images/a0c3f0179bc92a10aacbf6193cdef8e0c38d3327f4652176c2ca17f0ff71c4d1.jpg) + +![](images/d66e8f02f5eadf479549c26c816165970c349df4dc9edb223c4e6fd90e8ee818.jpg) + +![](images/0012f14a7ea5261c7334ecefd330e62c7c39e0c84a9a004ce48ec4af90ed3d77.jpg) + +Table 3. Quantitative comparison in terms of PSNR and SSIM for the $\times 2$ event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined. + +
MethodPSNR (↑)SSIM (↑)MethodPSNR (↑)SSIM (↑)
SOF-VSR31.840.9226EGVSR38.690.9771
TDAN33.740.9398EBVSR40.140.9801
RBPN36.660.9754EvTexture40.520.9813
BasicVSR39.570.9778EvTexture+40.570.9815
E-VSR37.320.9783MamEVSR41.140.9831
+ +based VSR methods such as E-VSR, EGVSR, EBVSR, EvTexture, and EvTexture+ across various datasets like CED + +and REDS4. For instance, on the CED dataset for the $\times 2$ event-based VSR task, MamEVSR achieves an average PSNR of $33.74\mathrm{~dB}$ and an SSIM of 0.9831, surpassing the next best method, EvTexture+, which scores $33.40\mathrm{~dB}$ and 0.9827 respectively. Compared to RGB-based VSR methods like DUF, TDAN, RBPN, and BasicVSR, MamEVSR demonstrates superior performance. On the CED dataset for the $\times 2$ event-based VSR task, MamEVSR's average PSNR of $33.74\mathrm{~dB}$ and SSIM of 0.9831 far exceed those of the top-performing RGB-based method, BasicVSR, which achieves + +Table 4. Ablation study of different components on CED. + +
MethodCED ×2CED ×4
PSNR↑SSIM↑PSNR↑SSIM↑
OursAll residual blocks39.990.979833.530.9082
Full model41.140.983134.030.9189
Core modules(a) w/o iMamba40.170.980333.600.9109
(b) w/o cMamba40.260.980533.630.9110
(c) w/o backward cell40.070.978933.480.9076
(d) w/o forward cell40.080.978933.400.9065
iMamba(e) Residual blocks40.620.981633.680.9111
(f) Deform. conv.40.720.981933.770.9127
(g) Flow warping40.840.982133.800.9129
(h) Concat & attention40.800.982033.770.9126
(i) w/o Int. Reorg.40.990.982633.880.9140
(j) w/o Channal att.41.020.982933.940.9145
cMamba(k) Residual blocks40.690.981833.710.9114
(l) EBVSR-BCS module [33]40.770.981933.800.9139
(m) Cross-modal att. [67]40.980.982633.880.9150
(n) Concat & iMamba40.800.982033.770.9129
(o) w/o Cross SSM40.870.982533.900.9170
Recon.(p) w/o iMamba41.040.982934.000.9179
(q) All iMamba40.890.982633.940.9168
+ +31.69 dB and 0.9226 respectively. Across all datasets and magnification factors, MamEVSR consistently delivers superior performance, showcasing its effectiveness in enhancing event-based video super-resolution tasks compared to both event-based and RGB-based VSR methods. + +Computational cost results. We calculate the number of parameters and runtime on the CED dataset. Results are shown in Table 2 and Figure 1. MamEVSR is positioned in the upper-left quadrant, reflecting an optimal trade-off between high performance and low computational cost. It achieves high-quality reconstructions (PSNR) with efficient computation and a modest parameter count, showcasing effective resource utilization. + +Qualitative results. We present visual comparison results on REDS4 and CED in Figure 5 and Figure 6, respectively. The LR patch appears blurry and lacks detail as we move through the SR patches; sharpness and clarity noticeably increase. MamEVSR produces results that closely resemble the ground truth, effectively restoring fine details and textures. For example, MamEVSR more accurately reconstructs wheel textures and license plate details, closely matching the ground truth in Figure 5. + +# 4.3. Ablation Study + +We conduct experiments on CED in terms of PSNR and SSIM. Results are in Table 4. + +Effectiveness of core components. Removing the iMamba block decreases PSNR and SSIM for both CED $\times 2$ and CED $\times 4$ tasks. Omitting the cMamba block has a smaller impact on performance. Disabling either backward or forward cells reduces PSNR and SSIM, especially for CED $\times 4$ . These results confirm the importance of the iMamba and cMamba blocks, as well as the bidirectional cell structure, in achieving optimal performance. + +Effectiveness of the iMamba block. Flow warping per + +![](images/5b3e4bda64bbcb7eb9fd0b9d7563bed7b1fbe44273925cb3fb47040624a15731.jpg) +Figure 7. A failure case. From left to right in sequence are patches cropped from the ground truth image, EvTexture, and MamEVSR. + +forms better than concatenation and attention. Removing internal reorganization or attention mechanisms within the iMamba block causes slight drops in PSNR and SSIM. The iMamba block plays a critical role in improving MamEVSR's performance. + +Effectiveness of the cMamba block. Incorporating the EBVSR-CBCS module [33] marginally boosts PSNR and SSIM. Cross-modal attention [67] yields comparable performance gains. Concatenating and using iMamba shows a slight decrease in metrics. Eliminating the cross-state space modeling mechanism within the cMamba block leads to minimal reductions in PSNR and SSIM. + +Effectiveness of the reconstructor. Removing the iMamba block from the reconstructor slightly improves PSNR and SSIM for both CED $\times 2$ and CED $\times 4$ tasks. However, replacing all iMamba blocks with residual blocks leads to a notable drop in performance, particularly for CED $\times 4$ . + +# 4.4. Limitations and Discussions + +While MamEVSR demonstrates strong results in event-based VSR, some challenges remain. As shown in Figure 7, MamEVSR struggles to recover particularly fine textures due to significant information loss in RGB data and low-resolution, poorly defined event data. Potential solutions include incorporating VGG loss to generate more realistic textures, as it effectively captures fine details in restoration tasks. Alternatively, using a longer temporal sequence of both event and RGB data could provide richer, continuous information, enhancing texture recovery. Additionally, we plan to design lighter-weight modules for event-based VSR and explore extending MamEVSR to other video restoration tasks. Investigating the effectiveness of other novel architectures [58, 85] for this task would also be insightful. + +# 5. Conclusion + +In this paper, we propose MamEVSR for event-based VSR. MamEVSR leverages the selective state space model to offer global receptive field coverage with linear complexity. Key components include the iMamba block for efficient bidirectional feature fusion and the cMamba block for integrating event information and capturing finer motion details. Experiments show that MamEVSR outperforms existing methods on various datasets, achieving superior quantitative and qualitative results. + +Acknowledgment. This project is supported by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation. + +# References + +[1] Andreas Aakerberg, Kamal Nasrollahi, and Thomas B Moeslund. Real-world super-resolution of face-images from surveillance cameras. IET Image Processing, 16(2):442-452, 2022. +[2] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Yichen Wu, Lilun Deng, Yukun Cui, Tao Feng, and Shuang Xu. Taskdriven image fusion with learnable fusion loss. In CVPR, 2025. +[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. +[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024. +[5] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based davis sensor. In ISCAS, pages 686-689, 2014. +[6] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017. +[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. +[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In ECCV, 2022. +[9] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. +[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. +[11] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021. +[12] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, 2023. +[13] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Sixiang Chen, Tian Ye, Renjing Pei, Kaiwen Zhou, Fenglong Song, and Lei Zhu. Restoreagent: Autonomous image restoration agent via multimodal large language models. NeurIPS, 2025. +[14] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In CVPR, 2024. +[15] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV. +[16] Michel Deudon, Alfredo Kalaitzis, Israel Goytom, Md Ri + +fat Arefin, Zhichao Lin, Kris Sankaran, Vincent Michalski, Samira E Kahou, Julien Cornebise, and Yoshua Bengio. Highres-net: Recursive fusion for multi-frame super-resolution of satellite imagery. arXiv preprint arXiv:2002.06460, 2020. +[17] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In ICLR, 2022. +[18] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. +[19] Ruisheng Gao, Zeyu Xiao, and Zhiwei Xiong. Mamba-based light field super-resolution with efficient subspace scanning. In ACCV, 2024. +[20] Tomio Goto, Takafumi Fukuoka, Fumiya Nagashima, Satoshi Hirano, and Masaru Sakurai. Super-resolution system for 4k-hdtv. In ICPR, 2014. +[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. +[22] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. +[23] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021. +[24] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. +[25] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019. +[26] Yan Huang, Wei Wang, and Liang Wang. Bidirectional recurrent convolutional networks for multi-frame superresolution. NeurlPS, 28, 2015. +[27] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):1015-1028, 2017. +[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. +[29] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022. +[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In CVPR, 2021. +[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. +[32] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. Eventure: Event-driven texture enhancement for video super-resolution. arXiv preprint arXiv:2406.13457, 2024. +[33] Dachun Kai, Yueyi Zhang, and Xiaoyan Sun. Video super + +resolution via event-driven temporal alignment. In ICIP, 2023. +[34] Soo Ye Kim, Jeongyeon Lim, Taeyoung Na, and Munchurl Kim. 3DSRnet: Video super-resolution using 3d convolutional neural networks. arXiv preprint arXiv:1812.09079, 2018. +[35] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. +[36] Junyong Lee, Myeonghee Lee, Sunghyun Cho, and Seungyong Lee. Reference-based video super-resolution using multi-camera video triplets. In CVPR, 2022. +[37] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In CVPR, 2023. +[38] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023. +[39] Bonan Li, Zicheng Zhang, Xingyi Yang, and Xinchao Wang. Focus on neighbors and know the whole: Towards consistent dense multiview text-to-image generator for 3d creation. arXiv preprint arXiv:2408.13149, 2024. +[40] Fei Li, Linfeng Zhang, Zikun Liu, Juan Lei, and Zhenbo Li. Multi-frequency representation enhancement with privilege information for video super-resolution. In ICCV, 2023. +[41] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. +[42] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019. +[43] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-correspondence aggregation network for video super-resolution. In ECCV, 2020. +[44] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. NeurIPS, 37, 2025. +[45] Zhuoyuan Li, Jiacheng Li, Yao Li, Li Li, Dong Liu, and Feng Wu. In-loop filtering via trained look-up tables. In VCIP, 2024. +[46] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. +[47] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. +[48] Wenyi Lian and Wenjing Lian. Sliding window recurrent network for efficient video super-resolution. In ECCV, 2022. +[49] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. +[50] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video super-resolution. In CVPR, 2022. +[51] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi + +Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. +[52] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In CVPR, 2023. +[53] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xinchao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. +[54] Yimin Luo, Liguo Zhou, Shu Wang, and Zhongyuan Wang. Video satellite imagery super resolution via convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 14(12):2398-2402, 2017. +[55] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. +[56] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. +[57] Tianbo Pan, Zidong Cao, and Lin Wang. Srfnet: Monocular depth estimation with fine-grained structure via spatial reliability-oriented fusion of frames and events. In ICRA, 2024. +[58] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. +[59] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In ECCV, 2022. +[60] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. +[61] Mehdi SM, Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In CVPR, 2018. +[62] Cedric Scheerlinck, Henri Rebecq, Timo Stoffregen, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Ced: Color event camera dataset. In CVPRW, 2019. +[63] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. NeurlPS, 2022. +[64] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In ICLR, 2022. +[65]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, 2023. +[66] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, 2022. +[67] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022. +[68] Qi Tang, Yao Zhao, Meiqin Liu, Jian Jin, and Chao Yao. Semantic lens: Instance-centric semantic alignment for video + +super-resolution. In AAAI, 2024. +[69] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. +[70] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. +[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. +[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through HR optical flow estimation. In ACCV, 2018. +[73] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, 2023. +[74] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, 2024. +[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. +[76] Yi Xiao, Xin Su, Qiangqiang Yuan, Denghong Liu, Huanfeng Shen, and Liangpei Zhang. Satellite video superresolution via multiscale deformable convolution alignment and temporal grouping projection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021. +[77] Yi Xiao, Qiangqiang Yuan, Kui Jiang, Xianyu Jin, Jiang He, Liangpei Zhang, and Chia-wen Lin. Local-global temporal difference learning for satellite video super-resolution. arXiv preprint arXiv:2304.04421, 2023. +[78] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023. +[79] Zeyu Xiao, Zhen Cheng, and Zhiwei Xiong. Space-time super-resolution for light field videos. IEEE Transactions on Image Processing, 32:4785-4799, 2023. +[80] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, pages 2113-2122, 2021. +[81] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACMMM, pages 2409-2418, 2024. +[82] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In ECCV, 2024. +[83] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, 2023. +[84] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025. +[85] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. ICLR, 2025. +[86] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, pages 4429-4438, 2021. +[87] Weihao Yu and Xinchao Wang. Mambaout: Do we really + +need mamba for vision? CVPR, 2025. +[88] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In ECCV, 2022. +[89] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. +[90] Guozhen Zhang, Chuxnu Liu, Yutao Cui, Xiaotong Zhao, Kai Ma, and Limin Wang. Vfimamba: Video frame interpolation with state space models. NeurIPS, 37, 2024. +[91] Liangpei Zhang, Hongyan Zhang, Huanfeng Shen, and Pingxiang Li. A super-resolution reconstruction algorithm for surveillance images. Signal Processing, 90(3):848-859, 2010. +[92] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. \ No newline at end of file diff --git a/2025/Event-based Video Super-Resolution via State Space Models/images.zip b/2025/Event-based Video Super-Resolution via State Space Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15ccf7c5c3fd569e2f1e21812f0fd1df376f892a --- /dev/null +++ b/2025/Event-based Video Super-Resolution via State Space Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25e05c6b71e23c751b22ad799ef22e2d77606ea9437659ae2be58afd6fe5c0e5 +size 958523 diff --git a/2025/Event-based Video Super-Resolution via State Space Models/layout.json b/2025/Event-based Video Super-Resolution via State Space Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce747a47631a6ef0f0bbb977e70e411d6b438cc1 --- /dev/null +++ b/2025/Event-based Video Super-Resolution via State Space Models/layout.json @@ -0,0 +1,12921 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "text", + "content": "Event-based Video Super-Resolution via State Space Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 144, + 381, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 144, + 381, + 158 + ], + "spans": [ + { + "bbox": [ + 231, + 144, + 381, + 158 + ], + "type": "text", + "content": "Zeyu Xiao Xinchao Wang*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 158, + 386, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 158, + 386, + 172 + ], + "spans": [ + { + "bbox": [ + 225, + 158, + 386, + 172 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 174, + 419, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 174, + 419, + 185 + ], + "spans": [ + { + "bbox": [ + 184, + 174, + 419, + 185 + ], + "type": "text", + "content": "zeyuxiao@nus.edu.sg, xinchao@nus.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 239, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 502 + ], + "type": "text", + "content": "Exploiting temporal correlations is crucial for video super-resolution (VSR). Recent approaches enhance this by incorporating event cameras. In this paper, we introduce MamEVSR, a Mamba-based network for event-based VSR that leverages the selective state space model, Mamba. MamEVSR stands out by offering global receptive field coverage with linear computational complexity, thus addressing the limitations of convolutional neural networks and Transformers. The key components of MamEVSR include: (1) The interleaved Mamba (iMamba) block, which interleaves tokens from adjacent frames and applies multidirectional selective state space modeling, enabling efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity. (2) The cross-modality Mamba (cMamba) block facilitates further interaction and aggregation between event information and the output from the iMamba block. The cMamba block can leverage complementary spatio-temporal information from both modalities and allows MamEVSR to capture finer motion details. Experimental results show that the proposed MamEVSR achieves superior performance on various datasets quantitatively and qualitatively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": "Video super-resolution (VSR) [18, 25-28] is a fundamental task in computer vision aimed at reconstructing high-resolution (HR) videos from low-resolution (LR) inputs. Given its broad applications in areas such as video surveillance [1, 39, 91], high-definition television [2-4, 12-15, 20, 45-47, 57, 79, 83], and satellite imagery [16, 54, 76, 77], VSR has attracted significant interest. The main challenge lies in effectively exploiting temporal information. Advanced VSR techniques often leverage temporal information through sliding window methods [29, 34, 42, 48, 53, 70, 73-75, 88], recurrent architectures [8-10, 36, 40, 61, 86], and Transformer-based designs [49, 50, 59, 63, 65, 66]," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 212, + 555, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 212, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 317, + 212, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 555, + 384 + ], + "type": "image", + "image_path": "08953d7a3671b444b9e54eff5a99af9ef529f9288c5fd5bd953a4d38496646fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 389, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 445 + ], + "type": "text", + "content": "Figure 1. Inference time and performance comparison. MamEVSR outperforms advanced methods with high efficiency. Comparisons are performed on the CED dataset. The green circles represent RGB-based VSR methods, while the yellow circles represent event-based VSR methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 458, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 506 + ], + "type": "text", + "content": "all demonstrating significant achievements. Recently, researchers have integrated event cameras into VSR to boost performance by harnessing their high temporal resolution and dynamic range capabilities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 508, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 508, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 312, + 508, + 556, + 663 + ], + "type": "text", + "content": "Event cameras are bio-inspired sensors that asynchronously detect intensity changes at the microsecond level on a per-pixel basis, generating millions of events per second while maintaining robustness in high-dynamic-range lighting. Leveraging these advantages, recent event-based VSR methods [30, 33, 52, 81, 82] have outperformed frame-only approaches by enhancing flow estimation, temporal alignment, and cross-modal fusion using event streams. However, current methods, primarily based on CNNs and attention models, face limitations, including poor adaptability to dynamic inputs, insufficient receptive fields for capturing inter-frame correlations at high resolutions, and high computational costs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "On the other hand, natural language processing (NLP) has recently witnessed the emergence of structured state space models (SSMs) [22]. Theoretically, SSMs combine the benefits of recurrent neural networks (RNNs) and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12564" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "CNNs, leveraging the global receptive field characteristic of RNNs and the computational efficiency of CNNs. One particularly notable SSM is the selective state space model, also known as Mamba [21], which has garnered significant attention within the vision community. Mamba's key feature is its ability to make SSM parameters time-variant (i.e., data-dependent), enabling it to effectively select relevant context within sequences—a crucial factor for enhancing model performance. Inspired by Mamba's capabilities, we explore its potential in addressing the task of Event-based VSR, and we propose the first Mamba-based method for this task, named MamEVSR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 218, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 590 + ], + "type": "text", + "content": "MamEVSR leverages a bidirectional recurrent pipeline to recurrently align and propagate temporally correlated features and integrate cross-modal information from the event stream. Its key components include: (1) The interleaved Mamba (iMamba) block: To facilitate efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, we introduce the iMamba block. Since the original SSMs model processes a single sequence, we merge tokens from two adjacent frames into one sequence for effective inter-frame modeling. By interleaving tokens from both frames and conducting multi-directional SSMs, we enable interactions between adjacent tokens from different propagated frames during sequence modeling. This design ensures that intermediate tokens in the sequence are from their spatio-temporal neighborhood. Stacking multiple iMamba blocks enhances MamEVSR's ability to handle complex inter-frame information exchange and frame alignment. (2) The cross-modality Mamba (cMamba) block: The cMamba block employs a cross-multiplication mechanism to enhance the interaction and aggregation between event information and the output from the iMamba block. This mechanism allows for the effective integration of event data with the temporal-fused features, thereby enriching the model's ability to leverage complementary cross-modal information and improve the overall reconstruction quality. By combining these two blocks, MamEVSR effectively exploits the strengths of both Mamba and event data to achieve superior performance in event-based VSR tasks. Experimental results on benchmark datasets demonstrate the effectiveness of our proposed MamEVSR (see an example in Figure 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Our contributions can be summarized as follows: (1) To the best of our knowledge, this marks the first successful application of state space models, specifically Mamba, in event-based VSR. (2) We propose the iMamba block and the cMamba block as core components in MamEVSR for event-based VSR. The former enables efficient feature fusion and propagation across bi-directional frames while maintaining linear complexity, the latter leverages complementary spatio-temporal information to capture finer motion details. (3) Extensive evaluations on benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "datasets demonstrate MamEVSR's superior performance, establishing a new benchmark for future explorations of Mamba's potential within event-based VSR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 118, + 400, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 118, + 400, + 131 + ], + "spans": [ + { + "bbox": [ + 314, + 118, + 400, + 131 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 139, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 553, + 389 + ], + "type": "text", + "content": "Video super-resolution. Existing VSR methods are designed to enhance the quality of LR frames by leveraging temporal information through sliding windows and recurrent structures. Sliding-window-based methods align adjacent LR frames with a reference frame to estimate an HR output. Early work focused on explicit optical flow estimation for frame alignment [6, 53, 69, 80, 84]. More recent methods have shifted to implicit alignment using dynamic filters [31], deformable convolutions [70, 75], and attention modules [28, 43]. Recent attention-based methods achieve new state-of-the-art performance [7, 37, 38, 43, 49, 53, 75, 78, 80]. However, these methods face challenges in capturing long-distance temporal features. Recurrent-based methods use RNNs to exploit temporal information across multiple frames. Sajjadi et al. [61] introduce a recurrent framework with optical flow for alignment. BasicVSR [9] and BasicVSR++[10] utilize bi-directional hidden states and advanced grid propagation techniques. PSRT [63] builds on BasicVSR++ with multi-frame self-attention for feature processing. In this paper, we focus on event-based VSR and introduce MamEVSR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 390, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 553, + 581 + ], + "type": "text", + "content": "Event-based VSR. Event cameras have the unique capability to measure intensity changes at each pixel independently with microsecond accuracy, making them valuable for VSR. Jing et al. [30] propose a two-stage method that uses events to interpolate low-resolution (LR) video, generating a high-frequency video for subsequent high-resolution (HR) frame reconstruction. Lu et al. [52] present a joint framework learning implicit neural representations from both RGB frames and events, enabling arbitrary-scale VSR. Kai et al. [32] introduce EvTexture, an event-driven texture enhancement network dedicated to restoring textures in VSR through the incorporation of high-frequency event signals. Xiao et al. [82] take inspiration from the parameter-efficient tuning, introducing an event adapter for event-based VSR. In this paper, we introduce MamEVSR, the first Mamba-based method for event-based VSR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "State space models. SSMs [17, 23, 44, 55, 64, 87, 90] are emerging as strong alternatives to Transformers [71] in natural language processing. S4 [23] is introduced for efficient, linear-complexity sequence modeling, followed by S5 [64], which improves parallelization, and GSS [55], which incorporates gated mechanisms. Mamba (S6) [21] stands out with its data-dependent parameters and hardware efficiency, outperforming Transformers in long-sequence tasks. In the computer vision domain, Vim [92] permutes 2D images into sequences for global modeling using bidirectional SSMs. Vmamba [51] extends this to four directions with a hier-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12565" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 289 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 289 + ], + "type": "image", + "image_path": "f559444258a2bb6522a59d852909b5f0fcad5cff72c153cafcd211f9904f650c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 292, + 296, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 348 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed MamEVSR. MamEVSR follows a typical bi-directional propagation scheme [9], consisting of bi-directional recurrent cells " + }, + { + "bbox": [ + 55, + 292, + 296, + 348 + ], + "type": "inline_equation", + "content": "For(\\cdot)" + }, + { + "bbox": [ + 55, + 292, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 292, + 296, + 348 + ], + "type": "inline_equation", + "content": "Back(\\cdot)" + }, + { + "bbox": [ + 55, + 292, + 296, + 348 + ], + "type": "text", + "content": ". For clarity, we have omitted the residual connection from the bicubic upsampling of the input, which is added to the final output, in the figure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 361, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 420 + ], + "type": "text", + "content": "archical design. VideoMamba [41] applies S6 to spatiotemporal video sequences, while MambaIR [24] is the first to use S6 for image restoration, surpassing Transformer-based methods. In this work, we explore the potential of Mamba in event-based VSR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 111, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 111, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 111, + 451 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 218, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 218, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 218, + 473 + ], + "type": "text", + "content": "3.1. Revisiting State Space Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 553 + ], + "type": "text", + "content": "State space models (SSM) [17, 23, 55, 64] represent a class of sequence-to-sequence modeling systems characterized by constant dynamics over time, a property also known as linear time-invariant. With linear complexity, SSM can effectively capture the inherent dynamics through an implicit mapping to latent states, which can be defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 566, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 566, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 85, + 566, + 295, + 578 + ], + "type": "interline_equation", + "content": "y (t) = C h (t) + D x (t), \\dot {h} (t) = A h (t) + B x (t). \\tag {1}", + "image_path": "60618cb686e7b72bf5bc4ba52f31df2345133d1c3163b9dd7b2d0355c2f98432.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "x(t)\\in \\mathbb{R},h(t)\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "y(t)\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " denotes the input, hidden state, and the output, respectively. " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is the state size and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\dot{h} (t)" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " refers to the time derivative of " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "h(t)" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " . Additionally, " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "B\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "C\\in \\mathbb{R}^{1\\times N}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "D\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " are the system matrices. To process discrete sequences like image and text, SSMs adopt Zero-Order Hold discretization [22] to map the input sequence " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{x_{1},x_{2},\\dots,x_{K}\\}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " to the output sequence " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{y_1,y_2,\\dots,y_K\\}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " . Specifically, suppose " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is the pre-defined timescale parameter to map continuous parameters " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " into a discrete space, the dis" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 73, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 322, + 73, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 322, + 73, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 322, + 73, + 545, + 339 + ], + "type": "image", + "image_path": "1f93695cb651ecb14c2405b43db020db1592d617abdfea9b0c7d10bcf73d6a19.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "lines": [ + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "type": "text", + "content": "Figure 3. Overview of the forward recurrent cell " + }, + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "type": "inline_equation", + "content": "\\text{For}(\\cdot)" + }, + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "type": "text", + "content": ", the backward recurrent cell " + }, + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "type": "inline_equation", + "content": "\\text{Back}(\\cdot)" + }, + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "type": "text", + "content": ", and the reconstructor " + }, + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 313, + 341, + 553, + 364 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 372, + 477, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 477, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 477, + 383 + ], + "type": "text", + "content": "cretization process can be formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 388, + 553, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 553, + 407 + ], + "type": "interline_equation", + "content": "\\bar {A} = \\exp (\\Delta A), \\bar {B} = (\\Delta A) ^ {- 1} (\\exp (A) - I) \\Delta B, \\bar {C} = C, \\tag {2}", + "image_path": "7e0c005d4bdd58366a4c6a4ed090f35b45a686aee34846295f3d0590782628ff.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 406, + 476, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 406, + 476, + 417 + ], + "spans": [ + { + "bbox": [ + 323, + 406, + 476, + 417 + ], + "type": "interline_equation", + "content": "y _ {k} = \\bar {C} h _ {k} + \\bar {D} x _ {k}, h _ {k} = \\bar {A} h _ {k - 1} + \\bar {B} x _ {k}.", + "image_path": "13d3c9a4bdeaa5978528c9cd8ea4bb45db1ef842a56034d875da60cc02883c1c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 553, + 460 + ], + "type": "text", + "content": "Here, all the matrices keep the same dimension as the operation iterates. Notably, " + }, + { + "bbox": [ + 313, + 423, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\overline{D}" + }, + { + "bbox": [ + 313, + 423, + 553, + 460 + ], + "type": "text", + "content": ", serving as a residual connection, is often discarded in the equation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 412, + 465, + 553, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 465, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 412, + 465, + 553, + 478 + ], + "type": "interline_equation", + "content": "y _ {k} = \\bar {C} h _ {k}. \\tag {3}", + "image_path": "54ad73bca3c9269bc1bfe01a8991186d0dd559889b515b78dbcb9eda0072cf3d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 484, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 553, + 509 + ], + "type": "text", + "content": "Besides, following Mamba [21], the matrix " + }, + { + "bbox": [ + 313, + 484, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\overline{B}" + }, + { + "bbox": [ + 313, + 484, + 553, + 509 + ], + "type": "text", + "content": " can be approximated by the first-order Taylor series" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 514, + 553, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 514, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 324, + 514, + 553, + 528 + ], + "type": "interline_equation", + "content": "\\bar {B} = (\\exp (A) - I) A ^ {- 1} B \\approx (\\Delta A) (\\Delta A) ^ {- 1} \\Delta B = \\Delta B. \\tag {4}", + "image_path": "8ba75dbb5135a9c267014de652402c57fd7e9ae413f774de422d31174f62d292.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "content": "The proposed iMamba and cMamba blocks are built upon Mamba and are specifically designed to address the challenges of event-based VSR tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 576, + 382, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 382, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 382, + 587 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{LR} = \\{\\ldots, I_{t-1}^{LR}, I_t^{LR}, I_{t+1}^{LR}, \\ldots\\}" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "I_t^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": ") be a sequence of LR input frames, " + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "E^{f,b}" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": " be forward and backward event streams, the goal of MamEVSR is to generate " + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{SR} = \\{\\ldots, I_{t-1}^{SR}, I_t^{SR}, I_{t+1}^{SR}, \\ldots\\}" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "I_t^{SR} \\in \\mathbb{R}^{sH \\times sW \\times 3}" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": "), which should be close to the ground-truth sequence " + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{GT} = \\{\\ldots, I_{t-1}^{GT}, I_t^{GT}, I_{t+1}^{GT}, \\ldots\\}" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "I_t^{GT} \\in \\mathbb{R}^{sH \\times sW \\times 3}" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "T, H," + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": " are the frame number, height, and width, respectively. " + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 592, + 555, + 689 + ], + "type": "text", + "content": " is the upscaling factor." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Because the event streams are not convenient for observation and processing by convolutional neural networks, we" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12566" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "convert forward and backward event streams as voxel grids " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{f,b}\\in \\mathbb{R}^{(T - 1)\\times H\\times W\\times Bin}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " via the temporal bilinear interpolation scheme" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 111, + 295, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 111, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 111, + 295, + 139 + ], + "type": "interline_equation", + "content": "V (i) = \\sum_ {k} p _ {k} \\max \\left(0, 1 - \\left| (i - 1) - \\frac {t _ {k} - t _ {0}}{t _ {N _ {e}} - t _ {0}} (B i n - 1) \\right|\\right), \\tag {5}", + "image_path": "980843681ee508945cdb73bac222f30e188e12f6dbee016281252bda826d154c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, Bin\\}" + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "text", + "content": "-th time bin. In our experiments, consistent with previous studies [32, 33], we also set " + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "inline_equation", + "content": "Bin = 5" + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "text", + "content": ". Furthermore, to mitigate the impact of hot pixels, we follow the study [32, 33] and normalize the voxel grid " + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 142, + 296, + 202 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 205, + 295, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 205, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 124, + 205, + 295, + 220 + ], + "type": "interline_equation", + "content": "\\hat {V} (i) = \\min ((V (i), \\eta)) / \\eta , \\tag {6}", + "image_path": "160367a1a9c0d1c19b8eca9c06251e396073840b267a58201cfd591d0b09924d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": " is the 98-th percentile value in the non-zero values of " + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": ". In this way, we obtain the normalized voxel grid " + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "inline_equation", + "content": "\\hat{V} \\in \\mathbb{R}^{H \\times W \\times Bin}" + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": ", which contains rich high-frequency textural information. We represent " + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "inline_equation", + "content": "\\pmb{V}^{f,b}" + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": " using this voxel grid representation for further processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "content": "Figure 2 shows an overview of the proposed MamEVSR. MamEVSR employs bidirectional recurrent cells " + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "inline_equation", + "content": "For(\\cdot)" + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "inline_equation", + "content": "Back(\\cdot)" + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "content": " akin to the scheme proposed in [9]. However, it introduces novel elements, such as extra inputs and specialized modules to harness event streams, which sets it apart from prior approaches. The event voxels and the LR frames are first converted into the feature domain using the feature encoders " + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "inline_equation", + "content": "(f_{En}^{V})" + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "inline_equation", + "content": "f_{En}^{I}" + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "content": ", consisting of " + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "content": " residual blocks), following which they are fed to the bidirectional recurrent cells. These cells consist of the iMamba block and the cMamba block. This step aims to utilize temporal correspondence from adjacent frames and aggregate cross-modal information from the event stream. As shown in Figure 3, using timestamp " + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 284, + 296, + 462 + ], + "type": "text", + "content": " as an example, the process can be expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 465, + 295, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 465, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 295, + 480 + ], + "type": "interline_equation", + "content": "F _ {t, f} ^ {c M}, F _ {t, f} ^ {E, c M} = \\operatorname {F o r} \\left(I _ {t - 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {f}\\right), \\tag {7}", + "image_path": "c2f0926af09bf0e104ba1a6bff440a0e35bc5e530e74fe9ba1035afab3e08f2d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 483, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 483, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 102, + 483, + 295, + 498 + ], + "type": "interline_equation", + "content": "F _ {t, b} ^ {c M}, F _ {t, b} ^ {E, c M} = B a c k \\left(I _ {t + 1} ^ {L R}, I _ {t} ^ {L R}, F _ {t} ^ {b}\\right). \\tag {8}", + "image_path": "783306349bd208586de8219acfda9999a9beae36f85687b93d985c9c59ea63ba.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": "Specifically, using the forward recurrent cell as an example, at time " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t-1}^{LR}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t}^{LR}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": " are fed to the iMamba block for efficient feature fusion and propagation across adjacent frames, resulting in " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t,f}^{LR'}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t,f}^{LR'}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "I_{t}^{LR}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": " are concatenated along the channel dimension, followed by the residual of " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t}^{LR}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": " to enhance the representation capacity. The resulting features " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t,f}^{iM}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": ", along with " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t}^{f}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": ", are fed to the cMamba block, which facilitates further interaction and aggregation in a cross-modality manner. The outputs of the cMamba block, " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t,f}^{cM}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "inline_equation", + "content": "F_{t,f}^{E,cM}" + }, + { + "bbox": [ + 55, + 500, + 295, + 634 + ], + "type": "text", + "content": ", are used for further reconstruction. The backward cell operates similarly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "content": "To generate the super-resolved output " + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "inline_equation", + "content": "I_{t}^{SR}" + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "content": " at timestamp " + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "content": ", features " + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "inline_equation", + "content": "F_{t,f}^{cM}, F_{t,b}^{cM}, F_{t,f}^{E,cM}, F_{t,b}^{E,cM}" + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "content": ", along with " + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "inline_equation", + "content": "I_{t}^{LR}" + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "inline_equation", + "content": "F_{t}^{LR}" + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "content": ", are fed to the reconstructor, which consists of an iMamba block, " + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 55, + 635, + 296, + 697 + ], + "type": "text", + "content": " residual blocks, and a pixel-shuffling operation. The process can be expressed as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 79, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "I _ {t} ^ {S R} = R \\left(F _ {t, f} ^ {c M}, F _ {t, b} ^ {c M}, F _ {t, f} ^ {E, c M}, F _ {t, b} ^ {E, c M}, I _ {t} ^ {L R}, F _ {t} ^ {L R}\\right). \\tag {9}", + "image_path": "b5bcb53d132ecb16cbc14a6c34eb54f23b586c0657040e2a89a2a19d3ea5eb3e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 459, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 459, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 459, + 83 + ], + "type": "text", + "content": "3.3. Interleaved Mamba Block" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "type": "text", + "content": "Effective inter-frame modeling is essential for VSR [9, 75, 86]. Conventional methods [75] rely on convolutional layers and attention mechanisms to capture temporal dependencies, offering fast inference but struggling with limited receptive fields. More advanced approaches, such as SemanticLens [68], introduce semantic-aware alignment strategies to enhance feature correspondence, but at the cost of increased computational complexity. To achieve a better balance between efficiency and effectiveness, we explore Mamba [21], an SSM, for inter-frame modeling. By leveraging its long-range dependency modeling and selective state updates, Mamba enables robust temporal feature fusion while maintaining computational efficiency, making it well-suited for VSR." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "text", + "content": "As shown in Figure 4(a), we introduce the iMamba block, which follows a Transformer-like structure (i.e., Norm " + }, + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "text", + "content": " Attention " + }, + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "text", + "content": " Norm " + }, + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 260, + 556, + 416 + ], + "type": "text", + "content": " MLP) [89], with two key modifications: (1) We replace the attention mechanism with a custom block that scans feature maps from two input frames. Thanks to the interleaved token reorganization design, this enables global inter-frame modeling with linear complexity. (2) Inspired by MambaIR [24] and MLFSR [19], which address limitations in locality and inter-channel interaction, we substitute the original MLP with a channel-attention block to enhance feature fusion and propagation. Here, Norm indicate the normalization operation and MLP means the multilayer perceptron." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "text", + "content": "For inter-frame modeling, given two adjacent low-resolution frame features, " + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "inline_equation", + "content": "F_{t-1}^{LR}" + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "inline_equation", + "content": "F_{t}^{LR}" + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "text", + "content": " (in the forward direction), both are first fed into the patch embedding module. This is followed by a series of operations, including layer normalization, linear projection, depth-wise convolution, and token reorganization, resulting in the reorganized feature " + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "inline_equation", + "content": "F_{Re,f}^{LR}" + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "text", + "content": ". The selective SSM-based Mamba block is then applied to independently model each direction. After processing, the sequences are rearranged and merged via token partitioning, another round of layer normalization, and linear operations. Subsequently, a channel attention block with a residual connection is applied to enhance feature propagation and stability. The output of the iMamba block is denoted as " + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "inline_equation", + "content": "F_{t,f}^{LR'}" + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "text", + "content": " for the forward direction. Similarly, for the backward direction, " + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "inline_equation", + "content": "F_{t,b}^{LR'}" + }, + { + "bbox": [ + 313, + 418, + 556, + 616 + ], + "type": "text", + "content": " is obtained following the same operations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "The interleaved reorganization within the iMamba block offers significant advantages for VSR. By ensuring that aggregated tokens come from both spatial and temporal neighborhoods, this design enhances local modeling, which is essential for capturing fine details between adjacent frames. Minimizing the distance between spatially and temporally adjacent tokens preserves the local context and reduces noise, which is particularly beneficial for maintaining tem-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12567" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 293, + 336 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 293, + 336 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 293, + 336 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 293, + 336 + ], + "type": "image", + "image_path": "2836ea4a022aa052f45f9cee90119e22c9bc40937e05d5fc13891b8ab2dc9c31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 553, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 553, + 361 + ], + "type": "text", + "content": "Figure 4. The structure of the proposed iMamba block and the cMamba block. The blocks offer specialized mechanisms to handle interframe modeling and cross-modal interactions effectively. See texts for more details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 70, + 547, + 336 + ], + "blocks": [ + { + "bbox": [ + 294, + 70, + 547, + 336 + ], + "lines": [ + { + "bbox": [ + 294, + 70, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 294, + 70, + 547, + 336 + ], + "type": "image", + "image_path": "b50cc4799de8371bcf7e460b3612aa8c80b12f87f87903115cf41fac9b0a468b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 369, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 407 + ], + "type": "text", + "content": "poral coherence across frames. This improved context modeling directly enhances the quality of fine detail reconstruction, leading to superior video restoration results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 417, + 220, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 220, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 220, + 430 + ], + "type": "text", + "content": "3.4. Cross-Modality Mamba Block" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 435, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 295, + 567 + ], + "type": "text", + "content": "The fusion of event data and RGB information is crucial for event-based VSR. Event cameras capture high-temporal-resolution motion cues that complement the rich spatial details in RGB frames. By integrating this sparse, high-frequency motion data with RGB content, we can significantly improve motion estimation, edge preservation, and image quality, especially in challenging conditions like low light or fast motion. This fusion enhances temporal consistency and fine detail reconstruction, leading to superior VSR performance. We, therefore, design the cMamba block for cross-modality fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "content": "Taking the forward direction as an example, as shown in Figure 4(b), the input features from the RGB frame output of the iMamba block, " + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "inline_equation", + "content": "F_{t,f}^{iM}" + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "content": ", and the event feature, " + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "inline_equation", + "content": "F_{t}^{f}" + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "content": ", are first processed by linear layers and depth-wise convolutions, generating " + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "inline_equation", + "content": "F^{RGB}" + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "inline_equation", + "content": "F^{Event}" + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "content": " before being fed to the cross selective SSM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Following the Mamba selection mechanism described in Section 3.1, system matrices " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " are generated to endow the model with context-aware capabilities, with linear projection layers generating these matrices. According to Eq. 2, matrix " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " decodes information from the hidden state " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_k" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " to produce the output " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_k" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". Inspired by the cross" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 369, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 418 + ], + "type": "text", + "content": "attention mechanism [11], we enable cross-modal interaction between the RGB and event modalities through information exchange between multiple selective scan modules. In particular, the process can be represented as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 427, + 553, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 427, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 351, + 427, + 553, + 441 + ], + "type": "interline_equation", + "content": "\\bar {A} _ {I} = \\exp \\left(\\Delta_ {\\mathrm {I}} A _ {I}\\right), \\quad \\bar {A} _ {E} = \\exp \\left(\\Delta_ {\\mathrm {E}} A _ {E}\\right), \\tag {10}", + "image_path": "cff38ba8add11006746da1d972846bbb5574b7b47b5116f7715e8569066d856b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 371, + 443, + 553, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 443, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 371, + 443, + 553, + 456 + ], + "type": "interline_equation", + "content": "\\bar {B} _ {I} = \\Delta_ {\\mathrm {r g b}} B _ {I}, \\bar {B} _ {E} = \\Delta_ {\\mathrm {x}} B _ {E}, \\tag {11}", + "image_path": "7b119347cee39ea3134dbf60cdc66bfbd2c6da7d13631ae1824041a733630fbf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 459, + 553, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 459, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 324, + 459, + 553, + 474 + ], + "type": "interline_equation", + "content": "h _ {I} ^ {t} = \\bar {A} _ {I} h _ {I} ^ {t - 1} + \\bar {B} _ {I} x _ {I} ^ {t}, h _ {E} ^ {t} = \\bar {A} _ {E} h _ {E} ^ {t - 1} + \\bar {B} _ {E} x _ {E} ^ {t}, \\tag {12}", + "image_path": "a278ba72cb9d594630655d7bb8698e66ace532ca8ac695fa73ddc65d41050729.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 475, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 475, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 332, + 475, + 553, + 491 + ], + "type": "interline_equation", + "content": "y _ {I} ^ {t} = C _ {\\times} h _ {I} ^ {t} + D _ {I} x _ {I} ^ {t}, y _ {E} ^ {t} = C _ {\\mathrm {r g b}} h _ {E} ^ {t} + D _ {E} x _ {E} ^ {t}, \\tag {13}", + "image_path": "e9dce246da56d8abc9d1857c01fca8f685af93044bcab81a008a455b09714b62.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 492, + 553, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 492, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 331, + 492, + 553, + 507 + ], + "type": "interline_equation", + "content": "y _ {I} = \\left[ y _ {I} ^ {1}, y _ {I} ^ {2}, \\dots , y _ {I} ^ {l} \\right], y _ {E} = \\left[ y _ {E} ^ {1}, y _ {E} ^ {2}, \\dots , y _ {E} ^ {l} \\right]. \\tag {14}", + "image_path": "dd4030bb5594a17f8a3d17dd0b7d5639108494b03feaaa1024960ccc14d993c3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "inline_equation", + "content": "x_{I / E}^{t}" + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "text", + "content": " represents the input at time step " + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "inline_equation", + "content": "y_{I / E}" + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "text", + "content": " denotes the selective scan output. " + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "inline_equation", + "content": "C_E" + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "inline_equation", + "content": "C_I" + }, + { + "bbox": [ + 313, + 516, + 553, + 563 + ], + "type": "text", + "content": " are the cross-modal matrices used for recovering the outputs at each time step from the hidden states." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 564, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 647 + ], + "type": "text", + "content": "The outputs from the two modality branches—RGB and event—are individually fed into the layer normalization operation, followed by a residual connection. This results in the final cross-modality enhanced feature, which effectively integrates information from both branches, improving feature representation for tasks like VSR. The above process can be denoted as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 654, + 553, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 654, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 325, + 654, + 553, + 670 + ], + "type": "interline_equation", + "content": "F _ {t, f} ^ {\\hat {i} M}, \\hat {F} _ {t} ^ {f} = \\operatorname {C S S M} \\left(\\mathrm {D} \\left(\\operatorname {L i n e a r} \\left(F _ {t, f} ^ {i M}\\right)\\right), \\mathrm {D} \\left(\\operatorname {L i n e a r} \\left(F _ {t} ^ {f}\\right)\\right)\\right) \\tag {15}", + "image_path": "464917a1a36010f272407879cc091d2dfd23f32d49dcb4c2bfa5af53a23c4f3e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 678, + 553, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 678, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 383, + 678, + 553, + 694 + ], + "type": "interline_equation", + "content": "F _ {t, f} ^ {c M} = \\mathrm {L N} \\left(F _ {t, f} ^ {\\hat {i} M}\\right) + F _ {t, f} ^ {i M}, \\tag {16}", + "image_path": "52ce57bb5729a01bcd481a2206c932e45acbec01b6be5c8c6c5528c0ae18249d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 384, + 699, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 699, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 384, + 699, + 553, + 715 + ], + "type": "interline_equation", + "content": "F _ {t, f} ^ {E, c M} = \\mathrm {L N} \\left(\\hat {F} _ {t} ^ {f}\\right) + F _ {t} ^ {f}. \\tag {17}", + "image_path": "e32f24dca10a434769666b7384096f116e7e053d8eafc705da6a5d83b151cf04.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12568" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 552, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison in terms of PSNR and SSIM for the " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined, respectively. Results are from [32] and [82]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 552, + 316 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 552, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 552, + 316 + ], + "type": "table", + "html": "
TypeMethodCED
Scene 1Scene 2Scene 3Scene 4Scene 5Scene 6Scene 7Scene 8Scene 9Scene 10Scene 11Average
RGB-based VSRDUF-----------31.09
-----------0.9183
SOF-VSR-----------31.84
-----------0.9226
TDAN35.8332.1231.5735.7335.4237.7528.9132.5435.5530.6735.0933.74
0.95400.93390.94660.95660.95360.94400.90620.90060.95410.93230.95610.9398
RBPN40.0734.1533.8339.5639.4440.3330.3634.9140.0531.5139.0336.66
0.98680.97390.97390.98690.98590.97820.96480.95020.98780.95510.98620.9754
BasicVSR39.3539.8139.7339.6039.4542.7139.1536.9739.3538.4539.4139.57
0.97840.97660.98320.97890.97780.98150.97480.96720.97760.97320.97990.9778
Event-based VSRE-VSR41.0834.7734.4440.4940.3240.8040.8035.1641.0031.7939.9737.32
0.98910.97750.97730.98910.98800.98010.98010.95360.99780.95860.98840.9783
EGVSR38.7838.6838.6739.0638.9341.9638.0336.1438.8437.6838.8638.69
0.97940.97500.98150.97980.97920.98310.97550.96350.97870.97260.98100.9771
EBVSR39.9540.2340.3140.2240.0643.0839.9737.2439.9538.8840.0740.14
0.98110.97800.98490.98160.98050.98300.97540.96890.98020.97570.98040.9801
EvTexture40.3940.5440.7540.6640.4543.2740.5337.5740.3539.2740.5440.52
0.98240.97890.98590.98290.98190.98340.98000.97050.98150.97690.98370.9813
EvTexture+-----------40.57
-----------0.9815
MamEVSR41.1740.6541.1041.4641.3143.2741.7037.7641.1639.3741.5041.14
0.98480.97910.98660.98530.98450.98260.98440.97190.98390.97750.98630.9831
", + "image_path": "fb85f2bbb10bec6488894f91ad4a4d561742a600d6aea0d9641f603a426a3703.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": "The features " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "F_{t,f}^{cM}" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "F_{t,f}^{E,cM}" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": ", obtained after cross-modal fusion, are fed into the reconstructor (see Figure 3) for the final image reconstruction. These fused features effectively capture both the high-frequency event data and spatial richness from the RGB frames, enhancing the quality of the reconstructed output by leveraging complementary information from both modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 138, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 138, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 138, + 433 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 182, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 182, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 182, + 451 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 456, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 295, + 659 + ], + "type": "text", + "content": "Datasets. For training, we utilize the REDS [56] and CED [62] datasets. Specifically, the REDS4 subset (clips 000, 011, 015, 020) from the REDS training set is employed as our test set, evaluating performance in the RGB channels. Consistent with prior work [30, 32, 33], we simulate events using the ESIM [60] tool for these clips. These simulated events are subsequently transformed into voxel grids according to Eqs. (5) and (6). Voxels are downsampled through the Bicubic interpolation operation, aligning with the frame downsampling method [32]. Following previous event-based VSR studies [33, 52], we use the CED [62] dataset for training and evaluating on real-world scenes1. The dataset is captured with a DAVIS346 [5] event camera, which outputs temporally synchronized events and frames at a resolution of " + }, + { + "bbox": [ + 54, + 456, + 295, + 659 + ], + "type": "inline_equation", + "content": "346 \\times 260" + }, + { + "bbox": [ + 54, + 456, + 295, + 659 + ], + "type": "text", + "content": ". When calculating the metrics (PSNR and SSIM), we exclude boundary 8 pixels and evaluate in the RGB channel." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "content": "Implementation details. We use 15 frames as input during training, with a mini-batch size of 6 and an input frame resolution of " + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "content": ". We apply data augmentation techniques to the training data, including horizontal flips and random rotations of " + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "content": ". MamEVSR is trained for 300K iterations using the Adam optimizer with a Cosine Annealing learning rate scheduler. Network architecture parameters are set to " + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "inline_equation", + "content": "N_{1} = 2" + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "inline_equation", + "content": "N_{2} = 45" + }, + { + "bbox": [ + 313, + 324, + 555, + 433 + ], + "type": "text", + "content": ". Supervision is provided by the Charbonnier penalty loss [35]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 378, + 440, + 555, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 440, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 378, + 440, + 555, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sqrt {\\| I ^ {S R} - I ^ {G T} \\| ^ {2} + \\varepsilon^ {2}}, \\tag {18}", + "image_path": "58278af9f63734dc9dc7b4b46381836044b2727ef4d12b67acaaae8da66f2bb1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "content": " in our experiments. " + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "inline_equation", + "content": "I^{SR}" + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "content": " denotes the results generated by MamEVSR, and " + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "inline_equation", + "content": "I^{GT}" + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "content": " represents the ground-truth frames. We omit the subscript " + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "content": " for simplicity. The initial learning rate for MamEVSR is " + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "inline_equation", + "content": "2\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 460, + 554, + 521 + ], + "type": "text", + "content": ". Training is conducted on 2 NVIDIA RTX 3090 GPUs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 529, + 537, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 537, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 537, + 542 + ], + "type": "text", + "content": "4.2. Quantitative and Qualitative Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 689 + ], + "type": "text", + "content": "We compare the proposed MamEVSR with a wide range of potential methods that could be used to address event-based VSR, aiming to explore as many diverse and rich approaches as possible. (1) RGB-based VSR methods: DUF [31], SOF-VSR [72], RBPN [25], EDVR [75], BasicVSR [9], VRT [49], and TTVSR [50]. In particular, we exclude the event stream and solely feed the LR video frames into these VSR networks for reconstruction, resulting in the final reconstructed video output. (2) Event-based VSR methods: we compare our MamEVSR with E-VSR [30], EGVSR [52], EBVSR [33], EvTexture, and EvTexture+ [32] to provide a thorough evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Quantitative results. As illustrated in Table 1, Table 2 and Table 3, MamEVSR consistently outperforms other event" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 664, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 295, + 714 + ], + "type": "text", + "content": "1The 11 scenes are as follows: People_Dynamic_Wave, Indoors_Foosball_2, Simple_Wires_2, People_Dynamic_Dancing, People_Dynamic_Jumping, Simple_Fruit_Fast, Outdoor_Jumping_Infrared_2, Simple_Carpet_Fast, People_Dynamic_Armroll, Indoors_Kitchen_2, and People_Dynamic_Sitting [82]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12569" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 94, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 2. Quantitative comparison in terms of PSNR and SSIM for the " + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": " event-based VSR task on the REDS4 dataset. The best and second-best results are highlighted in bold and underlined, respectively. Results are from [32]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 94, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 94, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 553, + 178 + ], + "type": "table", + "html": "
REDS4RGB-based VSREvent-based VSR
DUFEDVRBasicVSRTTVSRVRTEGVSREBVSREvTextureEvTexture+MamEVSR
00027.30/0.793728.01/0.825028.40/0.843428.82/0.856628.85/0.855325.16/0.706628.44/0.844630.72/0.9082-/-30.87/0.9103
01128.38/0.805632.17/0.886432.47/0.897933.47/0.910033.49/0.907226.56/0.772232.55/0.898733.72/0.9145-/-33.87/0.9166
01531.55/0.884634.06/0.920634.18/0.922435.01/0.932535.26/0.933229.83/0.852634.22/0.923535.06/0.9314-/-35.21/0.9335
02027.30/0.816430.09/0.888130.63/0.900031.17/0.909431.16/0.907825.94/0.784630.67/0.900931.65/0.9154-/-31.80/0.9175
Average28.63/0.825131.09/0.880031.42/0.890932.12/0.902132.19/0.900626.87/0.779031.47/0.891932.79/0.917432.93/0.919532.94/0.9195
#Params5.820.66.36.835.62.612.28.910.19.6
", + "image_path": "e039aaae2fdb63edded9d8f7b495de162e4dfe9b9c4c42ca7fab7d3df672d58c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 186, + 203, + 270 + ], + "blocks": [ + { + "bbox": [ + 55, + 186, + 203, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 203, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 203, + 270 + ], + "type": "image", + "image_path": "9d09f0b4873329567a6631b1831357bbcd55512330fc3c6057fc8dc75609ceba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 55, + 270, + 203, + 355 + ], + "blocks": [ + { + "bbox": [ + 55, + 270, + 203, + 355 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 203, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 203, + 355 + ], + "type": "image", + "image_path": "acb116899ce1d040d36f1807e8f5c0cdc161b841509c2ad5fb61c46ae5dc0362.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 358, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 555, + 381 + ], + "type": "text", + "content": "Figure 5. Visual comparisons for " + }, + { + "bbox": [ + 55, + 358, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 358, + 555, + 381 + ], + "type": "text", + "content": " VSR on REDS4. From left to right in sequence are patches cropped from LR, TDAN, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 186, + 247, + 225 + ], + "blocks": [ + { + "bbox": [ + 205, + 186, + 247, + 225 + ], + "lines": [ + { + "bbox": [ + 205, + 186, + 247, + 225 + ], + "spans": [ + { + "bbox": [ + 205, + 186, + 247, + 225 + ], + "type": "image", + "image_path": "d7991cf06befd2f8de94e9dcf0d36209e4ff71c1b0d0d63f3d467eaedbf72ce9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 206, + 228, + 247, + 270 + ], + "blocks": [ + { + "bbox": [ + 206, + 228, + 247, + 270 + ], + "lines": [ + { + "bbox": [ + 206, + 228, + 247, + 270 + ], + "spans": [ + { + "bbox": [ + 206, + 228, + 247, + 270 + ], + "type": "image", + "image_path": "33c278d1f953fa791c490fe92e841118d663045fd4fa54d33c5f2bbf5d338463.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 206, + 270, + 246, + 312 + ], + "blocks": [ + { + "bbox": [ + 206, + 270, + 246, + 312 + ], + "lines": [ + { + "bbox": [ + 206, + 270, + 246, + 312 + ], + "spans": [ + { + "bbox": [ + 206, + 270, + 246, + 312 + ], + "type": "image", + "image_path": "e3b6f614ee88f573044a9a4d12f25e01d2b1265022f1b7d4df9d325be36918a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 205, + 312, + 247, + 354 + ], + "blocks": [ + { + "bbox": [ + 205, + 312, + 247, + 354 + ], + "lines": [ + { + "bbox": [ + 205, + 312, + 247, + 354 + ], + "spans": [ + { + "bbox": [ + 205, + 312, + 247, + 354 + ], + "type": "image", + "image_path": "5c974e2a90e3dac74bc19b2661dff6873ebb2e3bcd2185415cde594dc21c056a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 249, + 186, + 334, + 224 + ], + "blocks": [ + { + "bbox": [ + 249, + 186, + 334, + 224 + ], + "lines": [ + { + "bbox": [ + 249, + 186, + 334, + 224 + ], + "spans": [ + { + "bbox": [ + 249, + 186, + 334, + 224 + ], + "type": "image", + "image_path": "08e78b722e96c5c7de965787fbfa48fb6f95c7973a988984db95b6f37b63b7f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 249, + 228, + 334, + 270 + ], + "blocks": [ + { + "bbox": [ + 249, + 228, + 334, + 270 + ], + "lines": [ + { + "bbox": [ + 249, + 228, + 334, + 270 + ], + "spans": [ + { + "bbox": [ + 249, + 228, + 334, + 270 + ], + "type": "image", + "image_path": "7c19d5f18e24148c9920d875daa67e8dd4749ce411b8c9c59d70bbe61d0b7623.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 249, + 270, + 334, + 312 + ], + "blocks": [ + { + "bbox": [ + 249, + 270, + 334, + 312 + ], + "lines": [ + { + "bbox": [ + 249, + 270, + 334, + 312 + ], + "spans": [ + { + "bbox": [ + 249, + 270, + 334, + 312 + ], + "type": "image", + "image_path": "274a0b9ad8a11145a2d3e0cbf5b386a90a86a964ec8fab1262a9db97f1598ae5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 249, + 312, + 334, + 354 + ], + "blocks": [ + { + "bbox": [ + 249, + 312, + 334, + 354 + ], + "lines": [ + { + "bbox": [ + 249, + 312, + 334, + 354 + ], + "spans": [ + { + "bbox": [ + 249, + 312, + 334, + 354 + ], + "type": "image", + "image_path": "481b0805ce2ff67a11ec08687bee33344b13ddb6cb9dbc2492c39e1e4d497cb5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 335, + 186, + 378, + 224 + ], + "blocks": [ + { + "bbox": [ + 335, + 186, + 378, + 224 + ], + "lines": [ + { + "bbox": [ + 335, + 186, + 378, + 224 + ], + "spans": [ + { + "bbox": [ + 335, + 186, + 378, + 224 + ], + "type": "image", + "image_path": "e839538222076ff7fcfd49928461bce481e321d351445d5e46427d3554540b1a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 335, + 228, + 377, + 270 + ], + "blocks": [ + { + "bbox": [ + 335, + 228, + 377, + 270 + ], + "lines": [ + { + "bbox": [ + 335, + 228, + 377, + 270 + ], + "spans": [ + { + "bbox": [ + 335, + 228, + 377, + 270 + ], + "type": "image", + "image_path": "7864f660d46bbe2f75f4caeef43f336e4c50150b243e3a98a3f57a2aa988700e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 335, + 270, + 377, + 312 + ], + "blocks": [ + { + "bbox": [ + 335, + 270, + 377, + 312 + ], + "lines": [ + { + "bbox": [ + 335, + 270, + 377, + 312 + ], + "spans": [ + { + "bbox": [ + 335, + 270, + 377, + 312 + ], + "type": "image", + "image_path": "1a4a928c77448541a3665ae043b34721438641bb9823c3e862dca9f502aa18e6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 335, + 312, + 377, + 354 + ], + "blocks": [ + { + "bbox": [ + 335, + 312, + 377, + 354 + ], + "lines": [ + { + "bbox": [ + 335, + 312, + 377, + 354 + ], + "spans": [ + { + "bbox": [ + 335, + 312, + 377, + 354 + ], + "type": "image", + "image_path": "1491547083335d285b5f87437d1cced10df9fb56bd261776a0019e12905f9d35.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 378, + 186, + 422, + 224 + ], + "blocks": [ + { + "bbox": [ + 378, + 186, + 422, + 224 + ], + "lines": [ + { + "bbox": [ + 378, + 186, + 422, + 224 + ], + "spans": [ + { + "bbox": [ + 378, + 186, + 422, + 224 + ], + "type": "image", + "image_path": "9d9cdee591e61a97d39062ff1a1f4631a33b31cc4f9eca7599220cf5f4666925.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 378, + 228, + 422, + 270 + ], + "blocks": [ + { + "bbox": [ + 378, + 228, + 422, + 270 + ], + "lines": [ + { + "bbox": [ + 378, + 228, + 422, + 270 + ], + "spans": [ + { + "bbox": [ + 378, + 228, + 422, + 270 + ], + "type": "image", + "image_path": "b56ab0024456ca8c96a224ad062a8f9c55db70d862b8ab7ee40c9c13ae05a381.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 378, + 270, + 422, + 312 + ], + "blocks": [ + { + "bbox": [ + 378, + 270, + 422, + 312 + ], + "lines": [ + { + "bbox": [ + 378, + 270, + 422, + 312 + ], + "spans": [ + { + "bbox": [ + 378, + 270, + 422, + 312 + ], + "type": "image", + "image_path": "3bf4793a95ee4230457673546d026f8b75ca635b0df05502351266a215c3a43a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 378, + 312, + 422, + 354 + ], + "blocks": [ + { + "bbox": [ + 378, + 312, + 422, + 354 + ], + "lines": [ + { + "bbox": [ + 378, + 312, + 422, + 354 + ], + "spans": [ + { + "bbox": [ + 378, + 312, + 422, + 354 + ], + "type": "image", + "image_path": "7febd35bd4926451a599b8aa85a0ca321743bfe8e1ccf8ea78e92ef796b5fc76.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 424, + 186, + 467, + 224 + ], + "blocks": [ + { + "bbox": [ + 424, + 186, + 467, + 224 + ], + "lines": [ + { + "bbox": [ + 424, + 186, + 467, + 224 + ], + "spans": [ + { + "bbox": [ + 424, + 186, + 467, + 224 + ], + "type": "image", + "image_path": "9495be13687f1410951638dc0362e526cc4cc404b42e11318998e941159e9b1d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 424, + 228, + 467, + 270 + ], + "blocks": [ + { + "bbox": [ + 424, + 228, + 467, + 270 + ], + "lines": [ + { + "bbox": [ + 424, + 228, + 467, + 270 + ], + "spans": [ + { + "bbox": [ + 424, + 228, + 467, + 270 + ], + "type": "image", + "image_path": "fe36963c897234235fceccc9fc044b7dbeedee7ee578e56e82eaa968365e3d64.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 424, + 270, + 467, + 312 + ], + "blocks": [ + { + "bbox": [ + 424, + 270, + 467, + 312 + ], + "lines": [ + { + "bbox": [ + 424, + 270, + 467, + 312 + ], + "spans": [ + { + "bbox": [ + 424, + 270, + 467, + 312 + ], + "type": "image", + "image_path": "a404da58f7b9255a2172f5bed087058ac7eb07646e8b1e6d6f0b3b97f54089fd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 424, + 312, + 467, + 354 + ], + "blocks": [ + { + "bbox": [ + 424, + 312, + 467, + 354 + ], + "lines": [ + { + "bbox": [ + 424, + 312, + 467, + 354 + ], + "spans": [ + { + "bbox": [ + 424, + 312, + 467, + 354 + ], + "type": "image", + "image_path": "5dfa976a0da885310210f1f128044dd7028d06ab4ddb2fb9a0d6b3b970343cb8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 468, + 186, + 510, + 224 + ], + "blocks": [ + { + "bbox": [ + 468, + 186, + 510, + 224 + ], + "lines": [ + { + "bbox": [ + 468, + 186, + 510, + 224 + ], + "spans": [ + { + "bbox": [ + 468, + 186, + 510, + 224 + ], + "type": "image", + "image_path": "f169570344f9fbaef314c6f1efe56c1c7c122f5b2936c1e1cec17b1654d12729.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 468, + 228, + 510, + 270 + ], + "blocks": [ + { + "bbox": [ + 468, + 228, + 510, + 270 + ], + "lines": [ + { + "bbox": [ + 468, + 228, + 510, + 270 + ], + "spans": [ + { + "bbox": [ + 468, + 228, + 510, + 270 + ], + "type": "image", + "image_path": "01141179cede12f13f66f286eb3fd08d78bf584290e438bd911084855ad721a8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 468, + 270, + 510, + 312 + ], + "blocks": [ + { + "bbox": [ + 468, + 270, + 510, + 312 + ], + "lines": [ + { + "bbox": [ + 468, + 270, + 510, + 312 + ], + "spans": [ + { + "bbox": [ + 468, + 270, + 510, + 312 + ], + "type": "image", + "image_path": "442e0a920102c0a9fc5c09875549abfd85cce755050f77141e2fe92b80fe09b9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 468, + 312, + 510, + 354 + ], + "blocks": [ + { + "bbox": [ + 468, + 312, + 510, + 354 + ], + "lines": [ + { + "bbox": [ + 468, + 312, + 510, + 354 + ], + "spans": [ + { + "bbox": [ + 468, + 312, + 510, + 354 + ], + "type": "image", + "image_path": "2df144de1dff438f05e90dabd2364b5c85af776a1cfbae4aac2c8cb951de601b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 511, + 186, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 511, + 186, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 511, + 186, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 511, + 186, + 553, + 224 + ], + "type": "image", + "image_path": "fa15dfae37e40bdd2bed3073f7f7f29ca317e0c05d1f9167cfec265ff325e715.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 511, + 228, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 511, + 228, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 511, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 511, + 228, + 553, + 270 + ], + "type": "image", + "image_path": "c346450f86b2fb5d9f1aecdb26141346b116b074fb7a09a0acfcc412cb1dfcb9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 511, + 270, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 511, + 270, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 511, + 270, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 511, + 270, + 553, + 312 + ], + "type": "image", + "image_path": "654a0eeffa6f17e0043ec269032565e925e93c42e3ee79c8d92b07e6678691f7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 511, + 312, + 553, + 354 + ], + "blocks": [ + { + "bbox": [ + 511, + 312, + 553, + 354 + ], + "lines": [ + { + "bbox": [ + 511, + 312, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 511, + 312, + 553, + 354 + ], + "type": "image", + "image_path": "16643d135eced12494e57b059d5d25dd09bb5c42aa634b289436dbd900cada6b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 55, + 384, + 203, + 468 + ], + "blocks": [ + { + "bbox": [ + 55, + 384, + 203, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 203, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 203, + 468 + ], + "type": "image", + "image_path": "94a772be7c3395a64bcb14289d31f1d6dbb0ef0a216177b74d36627784a2a539.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 55, + 468, + 203, + 554 + ], + "blocks": [ + { + "bbox": [ + 55, + 468, + 203, + 554 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 203, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 203, + 554 + ], + "type": "image", + "image_path": "c346c0717b69d01a3537414919f01da01b0fcc215162aceaca3c03e92435bc8b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 554, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 554, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 555, + 578 + ], + "type": "text", + "content": "Figure 6. Visual comparisons for " + }, + { + "bbox": [ + 55, + 554, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 554, + 555, + 578 + ], + "type": "text", + "content": " VSR on CED. From left to right in sequence are patches cropped from LR, EDVR, BasicVSR, EBVSR, EGVSR, EvTexture, MamEVSR, and the ground truth image." + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 205, + 384, + 247, + 426 + ], + "blocks": [ + { + "bbox": [ + 205, + 384, + 247, + 426 + ], + "lines": [ + { + "bbox": [ + 205, + 384, + 247, + 426 + ], + "spans": [ + { + "bbox": [ + 205, + 384, + 247, + 426 + ], + "type": "image", + "image_path": "6ccd19546dd981c636109863fb5c163c0cd0f41289c1a10ae56bbe7359304392.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 206, + 426, + 247, + 468 + ], + "blocks": [ + { + "bbox": [ + 206, + 426, + 247, + 468 + ], + "lines": [ + { + "bbox": [ + 206, + 426, + 247, + 468 + ], + "spans": [ + { + "bbox": [ + 206, + 426, + 247, + 468 + ], + "type": "image", + "image_path": "05983bf7b295d1de5a4aad09226b83200a5836ba4eec53d7ca13cc511c730ce5.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 206, + 468, + 247, + 510 + ], + "blocks": [ + { + "bbox": [ + 206, + 468, + 247, + 510 + ], + "lines": [ + { + "bbox": [ + 206, + 468, + 247, + 510 + ], + "spans": [ + { + "bbox": [ + 206, + 468, + 247, + 510 + ], + "type": "image", + "image_path": "6b3f49cb3c9347a414f6c44e852d192f43ac13d4ca0a37f2c9dd55a1f74ed4a9.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 206, + 511, + 247, + 552 + ], + "blocks": [ + { + "bbox": [ + 206, + 511, + 247, + 552 + ], + "lines": [ + { + "bbox": [ + 206, + 511, + 247, + 552 + ], + "spans": [ + { + "bbox": [ + 206, + 511, + 247, + 552 + ], + "type": "image", + "image_path": "8530c1e3c29498bcc4609e0cb8072208b1b12387bd83286a7e46cea0db63f21f.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 249, + 384, + 334, + 426 + ], + "blocks": [ + { + "bbox": [ + 249, + 384, + 334, + 426 + ], + "lines": [ + { + "bbox": [ + 249, + 384, + 334, + 426 + ], + "spans": [ + { + "bbox": [ + 249, + 384, + 334, + 426 + ], + "type": "image", + "image_path": "dc02fe49fcae1b52d440280275d41b7ec7abd62052f2eb39b6b48f8f756aa72a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 249, + 426, + 334, + 468 + ], + "blocks": [ + { + "bbox": [ + 249, + 426, + 334, + 468 + ], + "lines": [ + { + "bbox": [ + 249, + 426, + 334, + 468 + ], + "spans": [ + { + "bbox": [ + 249, + 426, + 334, + 468 + ], + "type": "image", + "image_path": "b955f09775b530f34c6236508546e55c76642dcacd7b987465e8ff50a7700604.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 249, + 468, + 334, + 510 + ], + "blocks": [ + { + "bbox": [ + 249, + 468, + 334, + 510 + ], + "lines": [ + { + "bbox": [ + 249, + 468, + 334, + 510 + ], + "spans": [ + { + "bbox": [ + 249, + 468, + 334, + 510 + ], + "type": "image", + "image_path": "5d14e9942c10956e3646a2bf7bc935e552cfa7d5dfeecf68e68f1905c0761254.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 249, + 511, + 334, + 552 + ], + "blocks": [ + { + "bbox": [ + 249, + 511, + 334, + 552 + ], + "lines": [ + { + "bbox": [ + 249, + 511, + 334, + 552 + ], + "spans": [ + { + "bbox": [ + 249, + 511, + 334, + 552 + ], + "type": "image", + "image_path": "c62741f328207a46197dbd61504fe22cf5df6895b3be80ffa2720848618a4407.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 335, + 384, + 378, + 426 + ], + "blocks": [ + { + "bbox": [ + 335, + 384, + 378, + 426 + ], + "lines": [ + { + "bbox": [ + 335, + 384, + 378, + 426 + ], + "spans": [ + { + "bbox": [ + 335, + 384, + 378, + 426 + ], + "type": "image", + "image_path": "5201d68d23065a7d1ae33c7c562cb1b44cb29fbdab678e1c5c1eb10b3d188408.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 335, + 426, + 377, + 468 + ], + "blocks": [ + { + "bbox": [ + 335, + 426, + 377, + 468 + ], + "lines": [ + { + "bbox": [ + 335, + 426, + 377, + 468 + ], + "spans": [ + { + "bbox": [ + 335, + 426, + 377, + 468 + ], + "type": "image", + "image_path": "02892673f4523dd91fa2aa656f6a01afa9ce49ca2cd603ebc6767f013544538e.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 335, + 468, + 377, + 510 + ], + "blocks": [ + { + "bbox": [ + 335, + 468, + 377, + 510 + ], + "lines": [ + { + "bbox": [ + 335, + 468, + 377, + 510 + ], + "spans": [ + { + "bbox": [ + 335, + 468, + 377, + 510 + ], + "type": "image", + "image_path": "809d63e656816cf6a98f42fe5ad94f3542dfa339348882f963ac846c57d43aa4.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 335, + 511, + 377, + 552 + ], + "blocks": [ + { + "bbox": [ + 335, + 511, + 377, + 552 + ], + "lines": [ + { + "bbox": [ + 335, + 511, + 377, + 552 + ], + "spans": [ + { + "bbox": [ + 335, + 511, + 377, + 552 + ], + "type": "image", + "image_path": "86993b96e0cbc6c7f521ab06b1da878e016919d367a7c3abd9463d6f28af2a10.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 380, + 384, + 422, + 426 + ], + "blocks": [ + { + "bbox": [ + 380, + 384, + 422, + 426 + ], + "lines": [ + { + "bbox": [ + 380, + 384, + 422, + 426 + ], + "spans": [ + { + "bbox": [ + 380, + 384, + 422, + 426 + ], + "type": "image", + "image_path": "4641407073ba7cdafbfe059953aa0419b4e310cfc4eb03eefb878f386862038a.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 380, + 426, + 422, + 468 + ], + "blocks": [ + { + "bbox": [ + 380, + 426, + 422, + 468 + ], + "lines": [ + { + "bbox": [ + 380, + 426, + 422, + 468 + ], + "spans": [ + { + "bbox": [ + 380, + 426, + 422, + 468 + ], + "type": "image", + "image_path": "6003d5f74d7fb1006fdbfc8bfcd38fe1439ad43c44d54e255f150c22804f852e.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 380, + 468, + 422, + 510 + ], + "blocks": [ + { + "bbox": [ + 380, + 468, + 422, + 510 + ], + "lines": [ + { + "bbox": [ + 380, + 468, + 422, + 510 + ], + "spans": [ + { + "bbox": [ + 380, + 468, + 422, + 510 + ], + "type": "image", + "image_path": "b5d558d88d6462acc8ee4d4f557908d87eded0a6f55f3c1e80b63b9842782fb4.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 380, + 511, + 422, + 552 + ], + "blocks": [ + { + "bbox": [ + 380, + 511, + 422, + 552 + ], + "lines": [ + { + "bbox": [ + 380, + 511, + 422, + 552 + ], + "spans": [ + { + "bbox": [ + 380, + 511, + 422, + 552 + ], + "type": "image", + "image_path": "5686ac1d5cf5d764873550746047bf2dc0b7815c2040beca2b54407e71289e47.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 424, + 384, + 467, + 426 + ], + "blocks": [ + { + "bbox": [ + 424, + 384, + 467, + 426 + ], + "lines": [ + { + "bbox": [ + 424, + 384, + 467, + 426 + ], + "spans": [ + { + "bbox": [ + 424, + 384, + 467, + 426 + ], + "type": "image", + "image_path": "2b4bb252085922e1413b3f1e5f22b5152a1e52e1d13f1baf40998891261de11f.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 424, + 426, + 467, + 468 + ], + "blocks": [ + { + "bbox": [ + 424, + 426, + 467, + 468 + ], + "lines": [ + { + "bbox": [ + 424, + 426, + 467, + 468 + ], + "spans": [ + { + "bbox": [ + 424, + 426, + 467, + 468 + ], + "type": "image", + "image_path": "cb858399e70c51c794427408ae7c43c93bf2318cbd5e7e145dbe717712724cd8.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 424, + 468, + 467, + 510 + ], + "blocks": [ + { + "bbox": [ + 424, + 468, + 467, + 510 + ], + "lines": [ + { + "bbox": [ + 424, + 468, + 467, + 510 + ], + "spans": [ + { + "bbox": [ + 424, + 468, + 467, + 510 + ], + "type": "image", + "image_path": "27c655a80be0105c3595c129c8c937cc7aff8a8180db6405a255adefa66015a2.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 424, + 511, + 467, + 552 + ], + "blocks": [ + { + "bbox": [ + 424, + 511, + 467, + 552 + ], + "lines": [ + { + "bbox": [ + 424, + 511, + 467, + 552 + ], + "spans": [ + { + "bbox": [ + 424, + 511, + 467, + 552 + ], + "type": "image", + "image_path": "947374c1fc3a132118133a376e87a715de38f63969f683fdbf4de2991ff5753b.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 468, + 384, + 510, + 426 + ], + "blocks": [ + { + "bbox": [ + 468, + 384, + 510, + 426 + ], + "lines": [ + { + "bbox": [ + 468, + 384, + 510, + 426 + ], + "spans": [ + { + "bbox": [ + 468, + 384, + 510, + 426 + ], + "type": "image", + "image_path": "2909f6e655391dfb8e22b338ad9c070834ee5d08fa3d32014d30b9191f2762b4.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 468, + 426, + 510, + 468 + ], + "blocks": [ + { + "bbox": [ + 468, + 426, + 510, + 468 + ], + "lines": [ + { + "bbox": [ + 468, + 426, + 510, + 468 + ], + "spans": [ + { + "bbox": [ + 468, + 426, + 510, + 468 + ], + "type": "image", + "image_path": "c4acf297317579deb35408eeb69164aa825701d0b72147b7b3df5d2cc0156b11.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 468, + 468, + 510, + 510 + ], + "blocks": [ + { + "bbox": [ + 468, + 468, + 510, + 510 + ], + "lines": [ + { + "bbox": [ + 468, + 468, + 510, + 510 + ], + "spans": [ + { + "bbox": [ + 468, + 468, + 510, + 510 + ], + "type": "image", + "image_path": "a342a2158f94464fa24fe7b44b548a8c72ecaab09ce91cd87c559e6d2378c063.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 468, + 511, + 510, + 552 + ], + "blocks": [ + { + "bbox": [ + 468, + 511, + 510, + 552 + ], + "lines": [ + { + "bbox": [ + 468, + 511, + 510, + 552 + ], + "spans": [ + { + "bbox": [ + 468, + 511, + 510, + 552 + ], + "type": "image", + "image_path": "3ef2e896ff08d19704d0de4fd00048e088dc9327f843de66482bdab0cff2a9fc.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 511, + 384, + 553, + 426 + ], + "blocks": [ + { + "bbox": [ + 511, + 384, + 553, + 426 + ], + "lines": [ + { + "bbox": [ + 511, + 384, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 511, + 384, + 553, + 426 + ], + "type": "image", + "image_path": "a48b88e8440db91fc32ba291535049f0c3983ed632067d7306351b69bfb3b178.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 511, + 426, + 553, + 468 + ], + "blocks": [ + { + "bbox": [ + 511, + 426, + 553, + 468 + ], + "lines": [ + { + "bbox": [ + 511, + 426, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 511, + 426, + 553, + 468 + ], + "type": "image", + "image_path": "a0c3f0179bc92a10aacbf6193cdef8e0c38d3327f4652176c2ca17f0ff71c4d1.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 511, + 468, + 553, + 510 + ], + "blocks": [ + { + "bbox": [ + 511, + 468, + 553, + 510 + ], + "lines": [ + { + "bbox": [ + 511, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 511, + 468, + 553, + 510 + ], + "type": "image", + "image_path": "d66e8f02f5eadf479549c26c816165970c349df4dc9edb223c4e6fd90e8ee818.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 511, + 511, + 553, + 552 + ], + "blocks": [ + { + "bbox": [ + 511, + 511, + 553, + 552 + ], + "lines": [ + { + "bbox": [ + 511, + 511, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 511, + 511, + 553, + 552 + ], + "type": "image", + "image_path": "0012f14a7ea5261c7334ecefd330e62c7c39e0c84a9a004ce48ec4af90ed3d77.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "table", + "bbox": [ + 56, + 617, + 295, + 682 + ], + "blocks": [ + { + "bbox": [ + 55, + 584, + 295, + 617 + ], + "lines": [ + { + "bbox": [ + 55, + 584, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 295, + 617 + ], + "type": "text", + "content": "Table 3. Quantitative comparison in terms of PSNR and SSIM for the " + }, + { + "bbox": [ + 55, + 584, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 584, + 295, + 617 + ], + "type": "text", + "content": " event-based VSR task on the CED dataset. The best and the second best results are highlighted in bold and underlined." + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 617, + 295, + 682 + ], + "lines": [ + { + "bbox": [ + 56, + 617, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 56, + 617, + 295, + 682 + ], + "type": "table", + "html": "
MethodPSNR (↑)SSIM (↑)MethodPSNR (↑)SSIM (↑)
SOF-VSR31.840.9226EGVSR38.690.9771
TDAN33.740.9398EBVSR40.140.9801
RBPN36.660.9754EvTexture40.520.9813
BasicVSR39.570.9778EvTexture+40.570.9815
E-VSR37.320.9783MamEVSR41.140.9831
", + "image_path": "2998b761e630fc361c3812a075700f7f71ec231609948b890c3d541c5d8d21d1.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "table_body" + } + ], + "index": 65 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "based VSR methods such as E-VSR, EGVSR, EBVSR, EvTexture, and EvTexture+ across various datasets like CED" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "text", + "content": "and REDS4. For instance, on the CED dataset for the " + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "text", + "content": " event-based VSR task, MamEVSR achieves an average PSNR of " + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "inline_equation", + "content": "33.74\\mathrm{~dB}" + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "text", + "content": " and an SSIM of 0.9831, surpassing the next best method, EvTexture+, which scores " + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "inline_equation", + "content": "33.40\\mathrm{~dB}" + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "text", + "content": " and 0.9827 respectively. Compared to RGB-based VSR methods like DUF, TDAN, RBPN, and BasicVSR, MamEVSR demonstrates superior performance. On the CED dataset for the " + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "text", + "content": " event-based VSR task, MamEVSR's average PSNR of " + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "inline_equation", + "content": "33.74\\mathrm{~dB}" + }, + { + "bbox": [ + 313, + 586, + 555, + 706 + ], + "type": "text", + "content": " and SSIM of 0.9831 far exceed those of the top-performing RGB-based method, BasicVSR, which achieves" + } + ] + } + ], + "index": 67 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12570" + } + ] + } + ], + "index": 68 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 82, + 294, + 281 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 280, + 81 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 280, + 81 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 280, + 81 + ], + "type": "text", + "content": "Table 4. Ablation study of different components on CED." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 82, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 82, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 82, + 294, + 281 + ], + "type": "table", + "html": "
MethodCED ×2CED ×4
PSNR↑SSIM↑PSNR↑SSIM↑
OursAll residual blocks39.990.979833.530.9082
Full model41.140.983134.030.9189
Core modules(a) w/o iMamba40.170.980333.600.9109
(b) w/o cMamba40.260.980533.630.9110
(c) w/o backward cell40.070.978933.480.9076
(d) w/o forward cell40.080.978933.400.9065
iMamba(e) Residual blocks40.620.981633.680.9111
(f) Deform. conv.40.720.981933.770.9127
(g) Flow warping40.840.982133.800.9129
(h) Concat & attention40.800.982033.770.9126
(i) w/o Int. Reorg.40.990.982633.880.9140
(j) w/o Channal att.41.020.982933.940.9145
cMamba(k) Residual blocks40.690.981833.710.9114
(l) EBVSR-BCS module [33]40.770.981933.800.9139
(m) Cross-modal att. [67]40.980.982633.880.9150
(n) Concat & iMamba40.800.982033.770.9129
(o) w/o Cross SSM40.870.982533.900.9170
Recon.(p) w/o iMamba41.040.982934.000.9179
(q) All iMamba40.890.982633.940.9168
", + "image_path": "9647a26a19f84fc43380940fc38ae8df2ba2f563ebc58f5c47ba05221956f820.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "content": "31.69 dB and 0.9226 respectively. Across all datasets and magnification factors, MamEVSR consistently delivers superior performance, showcasing its effectiveness in enhancing event-based video super-resolution tasks compared to both event-based and RGB-based VSR methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 446 + ], + "type": "text", + "content": "Computational cost results. We calculate the number of parameters and runtime on the CED dataset. Results are shown in Table 2 and Figure 1. MamEVSR is positioned in the upper-left quadrant, reflecting an optimal trade-off between high performance and low computational cost. It achieves high-quality reconstructions (PSNR) with efficient computation and a modest parameter count, showcasing effective resource utilization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 555 + ], + "type": "text", + "content": "Qualitative results. We present visual comparison results on REDS4 and CED in Figure 5 and Figure 6, respectively. The LR patch appears blurry and lacks detail as we move through the SR patches; sharpness and clarity noticeably increase. MamEVSR produces results that closely resemble the ground truth, effectively restoring fine details and textures. For example, MamEVSR more accurately reconstructs wheel textures and license plate details, closely matching the ground truth in Figure 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 149, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 149, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 149, + 576 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "type": "text", + "content": "We conduct experiments on CED in terms of PSNR and SSIM. Results are in Table 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "Effectiveness of core components. Removing the iMamba block decreases PSNR and SSIM for both CED " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": " and CED " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": " tasks. Omitting the cMamba block has a smaller impact on performance. Disabling either backward or forward cells reduces PSNR and SSIM, especially for CED " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": ". These results confirm the importance of the iMamba and cMamba blocks, as well as the bidirectional cell structure, in achieving optimal performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Effectiveness of the iMamba block. Flow warping per" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 101 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 101 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 101 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 101 + ], + "type": "image", + "image_path": "5b3e4bda64bbcb7eb9fd0b9d7563bed7b1fbe44273925cb3fb47040624a15731.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 104, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 313, + 104, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 553, + 126 + ], + "type": "text", + "content": "Figure 7. A failure case. From left to right in sequence are patches cropped from the ground truth image, EvTexture, and MamEVSR." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 137, + 553, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 196 + ], + "type": "text", + "content": "forms better than concatenation and attention. Removing internal reorganization or attention mechanisms within the iMamba block causes slight drops in PSNR and SSIM. The iMamba block plays a critical role in improving MamEVSR's performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 196, + 553, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 553, + 278 + ], + "type": "text", + "content": "Effectiveness of the cMamba block. Incorporating the EBVSR-CBCS module [33] marginally boosts PSNR and SSIM. Cross-modal attention [67] yields comparable performance gains. Concatenating and using iMamba shows a slight decrease in metrics. Eliminating the cross-state space modeling mechanism within the cMamba block leads to minimal reductions in PSNR and SSIM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "text", + "content": "Effectiveness of the reconstructor. Removing the iMamba block from the reconstructor slightly improves PSNR and SSIM for both CED " + }, + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "text", + "content": " and CED " + }, + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "text", + "content": " tasks. However, replacing all iMamba blocks with residual blocks leads to a notable drop in performance, particularly for CED " + }, + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 280, + 555, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 347, + 468, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 468, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 468, + 358 + ], + "type": "text", + "content": "4.4. Limitations and Discussions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 533 + ], + "type": "text", + "content": "While MamEVSR demonstrates strong results in event-based VSR, some challenges remain. As shown in Figure 7, MamEVSR struggles to recover particularly fine textures due to significant information loss in RGB data and low-resolution, poorly defined event data. Potential solutions include incorporating VGG loss to generate more realistic textures, as it effectively captures fine details in restoration tasks. Alternatively, using a longer temporal sequence of both event and RGB data could provide richer, continuous information, enhancing texture recovery. Additionally, we plan to design lighter-weight modules for event-based VSR and explore extending MamEVSR to other video restoration tasks. Investigating the effectiveness of other novel architectures [58, 85] for this task would also be insightful." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 540, + 388, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 388, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 388, + 552 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 557, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 665 + ], + "type": "text", + "content": "In this paper, we propose MamEVSR for event-based VSR. MamEVSR leverages the selective state space model to offer global receptive field coverage with linear complexity. Key components include the iMamba block for efficient bidirectional feature fusion and the cMamba block for integrating event information and capturing finer motion details. Experiments show that MamEVSR outperforms existing methods on various datasets, achieving superior quantitative and qualitative results." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": "Acknowledgment. This project is supported by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12571" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Andreas Aakerberg, Kamal Nasrollahi, and Thomas B Moeslund. Real-world super-resolution of face-images from surveillance cameras. IET Image Processing, 16(2):442-452, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 134, + 296, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 134, + 296, + 176 + ], + "spans": [ + { + "bbox": [ + 61, + 134, + 296, + 176 + ], + "type": "text", + "content": "[2] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Yichen Wu, Lilun Deng, Yukun Cui, Tao Feng, and Shuang Xu. Taskdriven image fusion with learnable fusion loss. In CVPR, 2025." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 177, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 177, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 177, + 295, + 231 + ], + "type": "text", + "content": "[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "type": "text", + "content": "[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "text", + "content": "[5] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based davis sensor. In ISCAS, pages 686-689, 2014." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 317, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 295, + 360 + ], + "type": "text", + "content": "[6] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 360, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 392 + ], + "type": "text", + "content": "[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 434 + ], + "type": "text", + "content": "[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In ECCV, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 466 + ], + "type": "text", + "content": "[9] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 467, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 467, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 57, + 467, + 295, + 508 + ], + "type": "text", + "content": "[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 509, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 509, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 57, + 509, + 295, + 540 + ], + "type": "text", + "content": "[11] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 541, + 295, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 541, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 57, + 541, + 295, + 583 + ], + "type": "text", + "content": "[12] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 584, + 295, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 584, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 57, + 584, + 295, + 627 + ], + "type": "text", + "content": "[13] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Sixiang Chen, Tian Ye, Renjing Pei, Kaiwen Zhou, Fenglong Song, and Lei Zhu. Restoreagent: Autonomous image restoration agent via multimodal large language models. NeurIPS, 2025." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 627, + 295, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 627, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 57, + 627, + 295, + 670 + ], + "type": "text", + "content": "[14] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In CVPR, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 700 + ], + "type": "text", + "content": "[15] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 701, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 701, + 295, + 712 + ], + "type": "text", + "content": "[16] Michel Deudon, Alfredo Kalaitzis, Israel Goytom, Md Ri" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 704 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "type": "text", + "content": "fat Arefin, Zhichao Lin, Kris Sankaran, Vincent Michalski, Samira E Kahou, Julien Cornebise, and Yoshua Bengio. Highres-net: Recursive fusion for multi-frame super-resolution of satellite imagery. arXiv preprint arXiv:2002.06460, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 128, + 553, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 170 + ], + "type": "text", + "content": "[17] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In ICLR, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 171, + 553, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 171, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 316, + 171, + 553, + 201 + ], + "type": "text", + "content": "[18] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 202, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 202, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 202, + 553, + 233 + ], + "type": "text", + "content": "[19] Ruisheng Gao, Zeyu Xiao, and Zhiwei Xiong. Mamba-based light field super-resolution with efficient subspace scanning. In ACCV, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 234, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 266 + ], + "type": "text", + "content": "[20] Tomio Goto, Takafumi Fukuoka, Fumiya Nagashima, Satoshi Hirano, and Masaru Sakurai. Super-resolution system for 4k-hdtv. In ICPR, 2014." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 266, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 297 + ], + "type": "text", + "content": "[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 298, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 330 + ], + "type": "text", + "content": "[22] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 331, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 361 + ], + "type": "text", + "content": "[23] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[24] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "text", + "content": "[25] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 437, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 468 + ], + "type": "text", + "content": "[26] Yan Huang, Wei Wang, and Liang Wang. Bidirectional recurrent convolutional networks for multi-frame superresolution. NeurlPS, 28, 2015." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "type": "text", + "content": "[27] Yan Huang, Wei Wang, and Liang Wang. Video superresolution via bidirectional recurrent convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):1015-1028, 2017." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "type": "text", + "content": "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "text", + "content": "[29] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 586, + 553, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 586, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 316, + 586, + 553, + 618 + ], + "type": "text", + "content": "[30] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In CVPR, 2021." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 618, + 553, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 618, + 553, + 660 + ], + "spans": [ + { + "bbox": [ + 316, + 618, + 553, + 660 + ], + "type": "text", + "content": "[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 661, + 553, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 661, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 316, + 661, + 553, + 693 + ], + "type": "text", + "content": "[32] Dachun Kai, Jiayao Lu, Yueyi Zhang, and Xiaoyan Sun. Eventure: Event-driven texture enhancement for video super-resolution. arXiv preprint arXiv:2406.13457, 2024." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 693, + 553, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 693, + 553, + 704 + ], + "spans": [ + { + "bbox": [ + 316, + 693, + 553, + 704 + ], + "type": "text", + "content": "[33] Dachun Kai, Yueyi Zhang, and Xiaoyan Sun. Video super" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12572" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "text", + "content": "resolution via event-driven temporal alignment. In ICIP, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 137 + ], + "type": "text", + "content": "[34] Soo Ye Kim, Jeongyeon Lim, Taeyoung Na, and Munchurl Kim. 3DSRnet: Video super-resolution using 3d convolutional neural networks. arXiv preprint arXiv:1812.09079, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 294, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 294, + 169 + ], + "type": "text", + "content": "[35] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 170, + 294, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 294, + 201 + ], + "type": "text", + "content": "[36] Junyong Lee, Myeonghee Lee, Sunghyun Cho, and Seungyong Lee. Reference-based video super-resolution using multi-camera video triplets. In CVPR, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 201, + 294, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 294, + 233 + ], + "type": "text", + "content": "[37] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In CVPR, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "type": "text", + "content": "[38] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "text", + "content": "[39] Bonan Li, Zicheng Zhang, Xingyi Yang, and Xinchao Wang. Focus on neighbors and know the whole: Towards consistent dense multiview text-to-image generator for 3d creation. arXiv preprint arXiv:2408.13149, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "type": "text", + "content": "[40] Fei Li, Linfeng Zhang, Zikun Liu, Juan Lei, and Zhenbo Li. Multi-frequency representation enhancement with privilege information for video super-resolution. In ICCV, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 340, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 340, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 340, + 294, + 382 + ], + "type": "text", + "content": "[41] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 383, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 414 + ], + "type": "text", + "content": "[42] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 415, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 446 + ], + "type": "text", + "content": "[43] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-correspondence aggregation network for video super-resolution. In ECCV, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 447, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 447, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 447, + 294, + 479 + ], + "type": "text", + "content": "[44] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. NeurIPS, 37, 2025." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 479, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 509 + ], + "type": "text", + "content": "[45] Zhuoyuan Li, Jiacheng Li, Yao Li, Li Li, Dong Liu, and Feng Wu. In-loop filtering via trained look-up tables. In VCIP, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "text", + "content": "[46] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 564, + 294, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 564, + 294, + 607 + ], + "spans": [ + { + "bbox": [ + 56, + 564, + 294, + 607 + ], + "type": "text", + "content": "[47] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 607, + 294, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 607, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 607, + 294, + 628 + ], + "type": "text", + "content": "[48] Wenyi Lian and Wenjing Lian. Sliding window recurrent network for efficient video super-resolution. In ECCV, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 628, + 294, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 294, + 670 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 294, + 670 + ], + "type": "text", + "content": "[49] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 671, + 294, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 671, + 294, + 701 + ], + "spans": [ + { + "bbox": [ + 56, + 671, + 294, + 701 + ], + "type": "text", + "content": "[50] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video super-resolution. In CVPR, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "type": "text", + "content": "[51] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 703 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 106, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 148 + ], + "type": "text", + "content": "[52] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In CVPR, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 148, + 553, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 148, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 148, + 553, + 180 + ], + "type": "text", + "content": "[53] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xinchao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 180, + 553, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 553, + 222 + ], + "type": "text", + "content": "[54] Yimin Luo, Liguo Zhou, Shu Wang, and Zhongyuan Wang. Video satellite imagery super resolution via convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 14(12):2398-2402, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "text", + "content": "[55] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 255, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 298 + ], + "type": "text", + "content": "[56] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 298, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 340 + ], + "type": "text", + "content": "[57] Tianbo Pan, Zidong Cao, and Lin Wang. Srfnet: Monocular depth estimation with fine-grained structure via spatial reliability-oriented fusion of frames and events. In ICRA, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 341, + 553, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 553, + 394 + ], + "type": "text", + "content": "[58] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "text", + "content": "[59] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In ECCV, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 426, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 447 + ], + "type": "text", + "content": "[60] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 448, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 448, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 448, + 553, + 479 + ], + "type": "text", + "content": "[61] Mehdi SM, Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In CVPR, 2018." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 479, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 510 + ], + "type": "text", + "content": "[62] Cedric Scheerlinck, Henri Rebecq, Timo Stoffregen, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Ced: Color event camera dataset. In CVPRW, 2019." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "type": "text", + "content": "[63] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. NeurlPS, 2022." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 543, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 574 + ], + "type": "text", + "content": "[64] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In ICLR, 2022." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 574, + 553, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 574, + 553, + 607 + ], + "spans": [ + { + "bbox": [ + 316, + 574, + 553, + 607 + ], + "type": "text", + "content": "[65]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, 2023." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 607, + 553, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 607, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 316, + 607, + 553, + 638 + ], + "type": "text", + "content": "[66] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, 2022." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 638, + 553, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 638, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 316, + 638, + 553, + 681 + ], + "type": "text", + "content": "[67] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 681, + 553, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 703 + ], + "type": "text", + "content": "[68] Qi Tang, Yao Zhao, Meiqin Liu, Jian Jin, and Chao Yao. Semantic lens: Instance-centric semantic alignment for video" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12573" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 77, + 73, + 196, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 196, + 83 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 196, + 83 + ], + "type": "text", + "content": "super-resolution. In AAAI, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 84, + 295, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 84, + 295, + 114 + ], + "spans": [ + { + "bbox": [ + 57, + 84, + 295, + 114 + ], + "type": "text", + "content": "[69] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 116, + 295, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 116, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 56, + 116, + 295, + 147 + ], + "type": "text", + "content": "[70] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 148, + 294, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 294, + 178 + ], + "type": "text", + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 179, + 294, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 179, + 294, + 212 + ], + "spans": [ + { + "bbox": [ + 56, + 179, + 294, + 212 + ], + "type": "text", + "content": "[72] Longguang Wang, Yulan Guo, Zaiping Lin, Xinpu Deng, and Wei An. Learning for video super-resolution through HR optical flow estimation. In ACCV, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 212, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 253 + ], + "type": "text", + "content": "[73] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 255, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 297 + ], + "type": "text", + "content": "[74] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 297, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 328 + ], + "type": "text", + "content": "[75] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 329, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 329, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 329, + 294, + 382 + ], + "type": "text", + "content": "[76] Yi Xiao, Xin Su, Qiangqiang Yuan, Denghong Liu, Huanfeng Shen, and Liangpei Zhang. Satellite video superresolution via multiscale deformable convolution alignment and temporal grouping projection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[77] Yi Xiao, Qiangqiang Yuan, Kui Jiang, Xianyu Jin, Jiang He, Liangpei Zhang, and Chia-wen Lin. Local-global temporal difference learning for satellite video super-resolution. arXiv preprint arXiv:2304.04421, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 426, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 456 + ], + "type": "text", + "content": "[78] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 457, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 294, + 489 + ], + "type": "text", + "content": "[79] Zeyu Xiao, Zhen Cheng, and Zhiwei Xiong. Space-time super-resolution for light field videos. IEEE Transactions on Image Processing, 32:4785-4799, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "text", + "content": "[80] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, pages 2113-2122, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 521, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 553 + ], + "type": "text", + "content": "[81] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACMMM, pages 2409-2418, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 553, + 294, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 294, + 584 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 294, + 584 + ], + "type": "text", + "content": "[82] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In ECCV, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 585, + 294, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 585, + 294, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 585, + 294, + 616 + ], + "type": "text", + "content": "[83] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 617, + 294, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 617, + 294, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 617, + 294, + 649 + ], + "type": "text", + "content": "[84] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 649, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 649, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 649, + 294, + 669 + ], + "type": "text", + "content": "[85] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. ICLR, 2025." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 670, + 294, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 700 + ], + "type": "text", + "content": "[86] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, pages 4429-4438, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 701, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 294, + 712 + ], + "type": "text", + "content": "[87] Weihao Yu and Xinchao Wang. Mambaout: Do we really" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 277 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 474, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 474, + 83 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 474, + 83 + ], + "type": "text", + "content": "need mamba for vision? CVPR, 2025." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 83, + 553, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 83, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 316, + 83, + 553, + 114 + ], + "type": "text", + "content": "[88] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In ECCV, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 116, + 553, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 116, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 553, + 158 + ], + "type": "text", + "content": "[89] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 159, + 553, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 159, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 316, + 159, + 553, + 190 + ], + "type": "text", + "content": "[90] Guozhen Zhang, Chuxnu Liu, Yutao Cui, Xiaotong Zhao, Kai Ma, and Limin Wang. Vfimamba: Video frame interpolation with state space models. NeurIPS, 37, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 190, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 190, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 190, + 553, + 232 + ], + "type": "text", + "content": "[91] Liangpei Zhang, Hongyan Zhang, Huanfeng Shen, and Pingxiang Li. A super-resolution reconstruction algorithm for surveillance images. Signal Processing, 90(3):848-859, 2010." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 232, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 277 + ], + "type": "text", + "content": "[92] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12574" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_content_list.json b/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1dc615596dcf41893b91d28619306c14e254d337 --- /dev/null +++ b/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_content_list.json @@ -0,0 +1,2517 @@ +[ + { + "type": "text", + "text": "EventFly: Event Camera Perception from Ground to the Sky", + "text_level": 1, + "bbox": [ + 191, + 109, + 803, + 131 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingdong Kong $^{1,2}$ , Dongyue Lu $^{1}$ , Xiang Xu $^{3}$ , Lai Xing Ng $^{4}$ , Wei Tsang Ooi $^{1,5}$ , Benoit R. Cottereau $^{5,6}$", + "bbox": [ + 99, + 157, + 903, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore $^{2}$ CNRS@CREATE $^{3}$ Nanjing University of Aeronautics and Astronautics", + "bbox": [ + 174, + 181, + 831, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Institute for Infocomm Research, A*STAR $^{5}$ IPAL, CNRS IRL 2955, Singapore $^{6}$ CerCo, CNRS UMR 5549, Université Toulouse III", + "bbox": [ + 101, + 200, + 901, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: https://event-fly.github.io", + "bbox": [ + 271, + 220, + 727, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5853b6c60761c6587a74967048b16a75ad001ca58c3dcb39334bd3326cf8a81a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 253, + 199, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vehicle", + "text_level": 1, + "bbox": [ + 204, + 257, + 269, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/272e59af0f8980ad3f170e9f0a09e528b01ce5f78d6b110e47c230fc12efdc45.jpg", + "image_caption": [ + "Viewpoint" + ], + "image_footnote": [], + "bbox": [ + 122, + 282, + 155, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0380827b2d44cf256cc5e2a08937d8621f0f10c63d16060b236ccd30a9f9ddac.jpg", + "image_caption": [ + "Speed" + ], + "image_footnote": [], + "bbox": [ + 205, + 284, + 241, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38bf90130853b4f6eafa14489a686655d4c0a7b784d764d6de96368ce090b95a.jpg", + "image_caption": [ + "Stability" + ], + "image_footnote": [], + "bbox": [ + 297, + 284, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fae723af8c1007b1cf8cf9fef9304bf2c34f57bcd00faafe627a5b95d34e627e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 253, + 483, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Drone", + "text_level": 1, + "bbox": [ + 486, + 258, + 540, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b6e06c1788b26e6da125f6e1ed9bd79cc723f8d5a9cffedff2f3ce0862a1792.jpg", + "image_caption": [ + "Viewpoint" + ], + "image_footnote": [], + "bbox": [ + 395, + 282, + 429, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc0b719d722a0d34af55598a5fa6feb11992a3adf105d490b193544aa6f87359.jpg", + "image_caption": [ + "Speed" + ], + "image_footnote": [], + "bbox": [ + 482, + 284, + 517, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12714ec2df0dabace022464e5d44748c26fcc3df4fa96bd3fece33094c9b4b57.jpg", + "image_caption": [ + "Stability" + ], + "image_footnote": [], + "bbox": [ + 571, + 284, + 599, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/93e5fc4ac5e48669deaf86565705b9cbc570be48789efe743c270687af8c09b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 253, + 733, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quadruped", + "text_level": 1, + "bbox": [ + 736, + 258, + 833, + 272 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17546392b375726ada3df0d26c7da180715b3478ab60db8ce698ae92c7e68f27.jpg", + "image_caption": [ + "Viewpoint" + ], + "image_footnote": [], + "bbox": [ + 669, + 282, + 702, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c47e2fc092d85e7050aee94b8af09d6c4e73894daf51e3d493d78f6fa26ee55.jpg", + "image_caption": [ + "Speed" + ], + "image_footnote": [], + "bbox": [ + 754, + 284, + 790, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e63cb70da619ea54e6a9a2f69cea3994b725304e8303b7be77a65583bdc7744.jpg", + "image_caption": [ + "Stability" + ], + "image_footnote": [], + "bbox": [ + 844, + 284, + 872, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/547e8d07899970ee724622fa5773af2ed3589fe6c77876aa3e5232900cfb2e80.jpg", + "image_caption": [ + "High", + "Medium", + "Medium", + "Low" + ], + "image_footnote": [], + "bbox": [ + 99, + 325, + 122, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/722497fff7d34657c0c3588afe73e091ac960e0df45cb05cadd6dd1039a0d9b9.jpg", + "image_caption": [ + "Fast", + "Medium ", + "G", + "Slow" + ], + "image_footnote": [], + "bbox": [ + 181, + 325, + 204, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2dc7517d49fb0dc1d09ee9ec1a3377475b1a80feea9223a51633793297b04a89.jpg", + "image_caption": [ + "Smooth", + "Medium", + "Medium", + "Unstable" + ], + "image_footnote": [], + "bbox": [ + 267, + 325, + 290, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7430a70af36122e6c18667f04af799155d96ade4050cf6c1feffe5b2dedd8492.jpg", + "image_caption": [ + "High", + "Medium", + "1", + "Low" + ], + "image_footnote": [], + "bbox": [ + 370, + 325, + 392, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e662280e6c6a9d42bda8875c4a36491c55f5bdbc4228c488d4d0d2d4bfbb8585.jpg", + "image_caption": [ + "Fast", + "Medium", + "。", + "Slow" + ], + "image_footnote": [], + "bbox": [ + 459, + 325, + 475, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/090f5d8154999e61ba71d5469c3ce4d016f97c5af25f637f0ef5eb3420b3c34f.jpg", + "image_caption": [ + "Smooth", + "Medium", + "hckable", + "Unstable" + ], + "image_footnote": [], + "bbox": [ + 535, + 325, + 550, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/589250169122ebb40635a3b5bcdaf7866afcc060f323b3ee93d07df75f51a282.jpg", + "image_caption": [ + "High", + "Medium", + "Low" + ], + "image_footnote": [], + "bbox": [ + 645, + 325, + 661, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f834fdae78fc0aa751cdfd0b48310b47b7b4f47c74f22491914de0600ae7d16.jpg", + "image_caption": [ + "Fast", + "Med", + "Slow" + ], + "image_footnote": [], + "bbox": [ + 727, + 325, + 743, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c683c3f4d0f92985f10ab440e3c60db10c280370b861fb0a2fffc9508f8f7d1.jpg", + "image_caption": [ + "Smooth", + "Medium", + "Unstable" + ], + "image_footnote": [], + "bbox": [ + 818, + 325, + 834, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e96fe90ac2ca6c14ea19373eae78d99b907c76fb4a1d3757f17630dbfa480a0.jpg", + "image_caption": [ + "Event Data Distribution" + ], + "image_footnote": [], + "bbox": [ + 99, + 407, + 179, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f29ea8a809278b472a63f846333db188c71f1b1edd09e28d06ede1d8569077fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 409, + 263, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48a278e813bd6ad73baaa9338462d23ce56dfb2a224f3406650e42f7910df55a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 409, + 349, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1cadc832c89bfb512c3223c2a34c5df55f2b854686a9481059492584c0d8fc04.jpg", + "image_caption": [ + "Event Semantic Distribution", + "Road" + ], + "image_footnote": [], + "bbox": [ + 99, + 497, + 179, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6c2e20de6a9d57ab420640ed802932506d6bbb8ddf59260b783564626f1ceb7.jpg", + "image_caption": [ + "Car" + ], + "image_footnote": [], + "bbox": [ + 184, + 497, + 264, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62b1ac51548c559d026477cff98ed06cc400e996f40ebada43d6780cf08bd661.jpg", + "image_caption": [ + "Sky" + ], + "image_footnote": [], + "bbox": [ + 271, + 497, + 349, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/311e1fa70f9d60671194957c69ed89db0762192e158333567d8c2ede21cefd09.jpg", + "image_caption": [ + "Building", + "Figure 1. Key discrepancies across the vehicle $(\\mathcal{P}^{\\mathrm{v}})$ , drone $(\\mathcal{P}^{\\mathrm{d}})$ , and quadruped $(\\mathcal{P}^{\\mathrm{q}})$ platforms. We compare the core attributes (viewpoint, speed, stability), event data distributions, and semantic distributions across event data acquired by three platforms, highlighting the challenges of adapting event camera perception to diverse operational contexts. These variations motivate the need for a robust cross-platform adaptation framework to harmonize event-based dense perception across distinct environmental setups and conditions." + ], + "image_footnote": [], + "bbox": [ + 99, + 549, + 179, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6cb210a35f1456ccf925008373ce6704efc63026634adf41197d81f8d1ae31ab.jpg", + "image_caption": [ + "Terrain" + ], + "image_footnote": [], + "bbox": [ + 184, + 549, + 264, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/493289b3f99d0d0a4c9f16a4f22b068563b08cf7ec593c840b21dc9d9fc73f7f.jpg", + "image_caption": [ + "Vegetation" + ], + "image_footnote": [], + "bbox": [ + 271, + 549, + 349, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb89c73a093a17b69cc9e0b11e62f36afba2bc92f76038d905e9dcadf78f6428.jpg", + "image_caption": [ + "Event Data Distribution" + ], + "image_footnote": [], + "bbox": [ + 372, + 410, + 450, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a06390f3d03a721f4c74b49185dffb5fe74c489558deadf7aacab7a52425152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 410, + 535, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbf364fa6b5bfeba3b5aad3495ff5cb4c88761a1bdc4284d957e7eafeddf3915.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 410, + 622, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9135fa7238a4726e06951fd84212a590fa40cf29ce9d01518912f78d92916ab8.jpg", + "image_caption": [ + "Event Semantic Distribution", + "Road" + ], + "image_footnote": [], + "bbox": [ + 372, + 497, + 450, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b44515befee8a13e04a62f4467188f29a82e1439eabf2c8ac95419a830fd449.jpg", + "image_caption": [ + "Car" + ], + "image_footnote": [], + "bbox": [ + 459, + 497, + 537, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edca595c31516659e530990cc0418a03f96d8633d5cec85cc348d916a9ab8caf.jpg", + "image_caption": [ + "Sky" + ], + "image_footnote": [], + "bbox": [ + 545, + 497, + 622, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db50f4a6193090f94f116710b8425e58af58a833fe459ac3533b57770578957c.jpg", + "image_caption": [ + "Building" + ], + "image_footnote": [], + "bbox": [ + 372, + 550, + 450, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0261deb16bbf0a1b8ee8af8f5fd5c8c60415dac3e4d7bf649fbe3cc084185f7.jpg", + "image_caption": [ + "Terrain" + ], + "image_footnote": [], + "bbox": [ + 459, + 550, + 537, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a70b7b37848f10af007f67bee10b431fc7d5030f810c66f51cfa9ac12d25dd19.jpg", + "image_caption": [ + "Vegetation" + ], + "image_footnote": [], + "bbox": [ + 545, + 550, + 622, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a52df0b455925d426b87f547e8ba74c4fa88c17b7c0975acafb160e8cae3422c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 410, + 725, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca5ab5c620c9479febd96acf155b9de61db38187a92d37697d33550cc215b4e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 410, + 810, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5915ff6a6a67ba116d7fbf2da64c3b83e811ab4b374950d5c9a07f810323f22.jpg", + "image_caption": [ + "Event Data Distribution" + ], + "image_footnote": [], + "bbox": [ + 818, + 409, + 897, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab513ef08af6e1b4c0c87a701de10c42a84ca4736b79ec3ebe737cca0bb45380.jpg", + "image_caption": [ + "Event Semantic Distribution", + "Road" + ], + "image_footnote": [], + "bbox": [ + 647, + 497, + 725, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a9dc4af66216b52316b7fc5b46a05e9155428b5bb702b6f6c60cc56dc981d9e.jpg", + "image_caption": [ + "Car" + ], + "image_footnote": [], + "bbox": [ + 733, + 497, + 810, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b96fd6e9c260adaad00800abbea46a130b08e63c24901c7f6de4802f276495d5.jpg", + "image_caption": [ + "Sky" + ], + "image_footnote": [], + "bbox": [ + 818, + 497, + 897, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0072c8948b2e6cb856f182f469503f64651ab3beea9922e71bb19d7e66599b4f.jpg", + "image_caption": [ + "Terrain" + ], + "image_footnote": [], + "bbox": [ + 733, + 550, + 812, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3532366df72d2686d0bdd7ff51635c8800d64c8ffb8f942901d368198dba5eac.jpg", + "image_caption": [ + "Vegetation" + ], + "image_footnote": [], + "bbox": [ + 818, + 550, + 897, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 672, + 326, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-platform adaptation in event-based dense perception is crucial for deploying event cameras across diverse settings, such as vehicles, drones, and quadrupeds, each with unique motion dynamics, viewpoints, and class distributions. In this work, we introduce EventFly, a framework for robust cross-platform adaptation in event camera perception. Our approach comprises three key components: i) Event Activation Prior (EAP), which identifies high-activation regions in the target domain to minimize prediction entropy, fostering confident, domain-adaptive predictions; ii) EventBlend, a data-mixing strategy that integrates source and target event voxel grids based on EAP-driven similarity and density maps, enhancing feature alignment;", + "bbox": [ + 88, + 704, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and iii) EventMatch, a dual-discriminator technique that aligns features from source, target, and blended domains for better domain-invariant learning. To holistically assess cross-platform adaptation abilities, we introduce EXPo, a large-scale benchmark with diverse samples across vehicle, drone, and quadruped platforms. Extensive experiments validate our effectiveness, demonstrating substantial gains over popular adaptation methods. We hope this work can pave the way for more adaptive, high-performing event perception across diverse and complex environments.", + "bbox": [ + 511, + 674, + 906, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 842, + 643, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras, with their asynchronous operation and high temporal resolution, provide great advantages over conven-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 20 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 17, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1472", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tional frame-based sensors by capturing precise pixel intensity changes as they occur [24, 26, 43, 44, 71]. These attributes make event cameras ideally suited for real-time applications in dynamic environments, such as autonomous driving, aerial navigation, and robotic perception [56, 74].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite this potential, event cameras remain primarily deployed on the vehicle platform [7, 9]. Based on a recent survey study [11], 12 out of 15 public event camera perception datasets are collected exclusively from ground vehicles [2, 13, 29, 64, 93, 100]. In contrast, datasets involving alternative robot platforms, such as drones [4, 16, 85] or mobile robots [5, 95, 102], are limited, making robust event-based perception across diverse domains challenging [66, 70]. To enable versatile applications, a cross-platform approach is vital for adapting the perception models to the new, real-world, and challenging operational contexts.", + "bbox": [ + 89, + 169, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-platform adaptation for event-based dense perception presents significant technical challenges. As can be seen from Fig. 1, each robot platform - vehicle, drone, or quadruped - exhibits unique motion patterns, environmental interactions, and perspective-based dynamics that create distinct activation patterns in the event data [22, 53, 65, 73]. For example, ground vehicles capture event activity near surfaces [26, 29, 100], while drones tend to capture high-altitude landscape features with sparse ground details [12, 42, 52, 56, 80]. Conventional domain adaptation methods commonly used for frame-based sensors are not well-suited to handle the distinct spatial-temporal nuances of event camera data, which demand tailored alignment strategies [48, 50, 60, 78]. A specialized framework for cross-platform adaptation in event-based dense perception is, therefore, crucial for reliable, context-sensitive perception across varied robot sensing conditions.", + "bbox": [ + 89, + 335, + 482, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce EventFly, a robust framework for cross-platform adaptation in event camera dense perception. We bridge domain-specific gaps in event-camera perception by leveraging platform-specific activation patterns and aligning feature distributions across domains. Our design consists of three core components, each addressing a specific challenge in cross-platform event data alignment:", + "bbox": [ + 89, + 595, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "■ Event Activation Prior (EAP). Event cameras on different platforms reveal distinctive high-activation regions shaped by platform-specific dynamics. EAP leverages these by identifying frequently activated areas in the target domain and minimizing prediction entropy within these regions. This promotes confident, domain-adaptive predictions that align the model to platform-specific event patterns, enhancing performance in target-relevant contexts.", + "bbox": [ + 89, + 702, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-Platform Consistency Regularization. To bridge the gaps in-between different robot platforms, we propose EventBlend to create hybrid event voxel grids by combining source and target event data in a spatially structured way. Guided by EAP-driven similarity and density maps, this", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method selectively integrates features based on shared activation patterns, focusing on areas where platform-specific details overlap. This fusion retains platform-specific cues while enhancing cross-domain feature consistency, balancing reliable source information with context-sensitive target features for more robust cross-platform adaptations.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-Platform Adversarial Training. To further align domain distributions, we propose EventMatch, a dual-discriminator approach across source, target, and blended representations. One discriminator enforces alignment between source and blended domains, while the other softly adapts blended features toward the target, particularly in high-activation regions. This layered approach supports robust, domain-adaptive learning that generalizes well across platforms, balancing domain-invariant features with target-specific adaptations for improved perception accuracy.", + "bbox": [ + 511, + 181, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To rigorously assess the cross-platform adaptation performance, we establish EXPo, a large-scale Event-based Xcross-Platform dense perception benchmark with around 90k event data from three platforms: vehicle, drone, and quadruped. From extensive benchmark experiments, our approach consistently demonstrates robust adaptation capabilities, achieving on average $23.8\\%$ higher accuracy and $77.1\\%$ better mIoU across platforms compared to source-only training. When evaluated against prior adaptation methods, we outperform by significant margins across almost all semantic classes, highlighting the scalability and effectiveness in diverse, challenging operational contexts.", + "bbox": [ + 511, + 333, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 532, + 515, + 870, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EventFly, a novel framework designed for cross-platform adaptation in event camera perception, facilitating robustness under diverse event data dynamics. To our knowledge, this is the first work proposed to address this critical gap in event-based perception tasks.", + "- We introduce Event Activation Prior (EAP), EventBlend, and EventMatch - a set of tailored techniques that utilize platform-specific activation patterns, spatial data mixing, and dual-domain feature alignment to tackle the unique challenges of event-based cross-platform adaptation.", + "- We establish a large-scale benchmark, EXPo, for cross-platform adaptation in event-based perception, comprising rich collections of samples from vehicle, drone, and quadruped domains, providing a strong foundation for further research in adaptive event camera applications." + ], + "bbox": [ + 513, + 529, + 903, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 768, + 653, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event Camera Perception. Recent works harnessed event cameras for diverse perception tasks in dynamic settings, such as detection [25, 26, 28, 55, 105], segmentation [1, 32, 47, 75], depth estimation [15, 30, 34, 58, 99, 101], and visual odometry [10, 18, 35, 59]. Object detection with event cameras involves locating objects by leveraging the high temporal resolution of event streams, enabling", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1473", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "effective scene parsing even under challenging conditions [49, 63, 98, 103]. Meanwhile, semantic segmentation aims to enhance scene understanding, which is vital for safe navigation in autonomous systems [15, 45, 57, 83]. Additionally, recent research has focused on integrating event data with complementary modalities to improve accuracy and mitigate the sparse nature of event data [3, 14, 27, 33, 84]. Our work extends this line by focusing on robust cross-platform adaptation, ensuring that event-based perception models can generalize across different robotic platforms.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Domain Adaptation. Transferring knowledge from a labeled source domain to an unlabeled target domain addresses the challenge of limited annotated data in diverse real-world cases [69, 96]. Traditional methods in this area leverage strategies like domain adversarial training [79], entropy minimization [61, 68, 81, 82], contrastive learning [37, 87, 92], domain mixing [46, 51, 77, 97], and self-training [8, 38, 94, 104]. However, despite promising performances, these approaches are primarily designed for frame-based data [17, 36, 91], which lacks the unique spatiotemporal properties of the stream-based event camera data. Our approach adapts these domain adaptation principles specifically for event-based vision perception, addressing the distinct challenges posed by event camera data.", + "bbox": [ + 89, + 242, + 482, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Adaptation from Event Camera. Recent works have begun to address domain shifts for event-based perception. Approaches like Ev-Transfer [57], ESS [75], and HPL-ESS [40] focus on transferring knowledge from RGB frames to event data, enabling effective event-based perception through frame-to-event domain adaptation. Other studies leverage event cameras to aid adaptation in low-light or night-time conditions for conventional RGB sensors [21, 39, 86, 89]. There are also works that explored efficient learning utilizing shared representations across modalities [20, 41, 88]. To the best of our knowledge, our work is the first to tackle cross-platform adaptation across multiple event camera platforms, such as vehicles, drones, and quadrapeds, focusing on the unique motion and environmental interactions inherent to each platform.", + "bbox": [ + 89, + 454, + 482, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 693, + 225, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work tackles the problem of cross-platform adaptation for event-based dense perception across three distinct platforms: vehicle, drone, and quadruped, denoted as $\\mathcal{P}^{\\mathrm{v}}$ , $\\mathcal{P}^{\\mathrm{d}}$ , and $\\mathcal{P}^{\\mathrm{q}}$ , respectively. Each platform exhibits unique spatial and temporal characteristics in the captured event data, influenced by platform-specific perspectives, motion patterns, semantics, and environments.", + "bbox": [ + 89, + 718, + 480, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 833, + 230, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event cameras capture continuous, asynchronous streams [6], where each event $\\mathbf{e}_i = (e_i^x,e_i^y,e_i^t,e_i^p)$ consists of pixel coordinates $(e_i^x,e_i^y)$ , timestamp $e_i^t$ , and polarity $e_i^p\\in$", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\{-1, + 1\\}$ , indicating either a decrease or increase in brightness. This format allows event cameras to capture high-speed motion and scene changes in dynamic environments.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event Data Representation. To efficiently process event data, we convert raw events into voxel grid representations, denoted as $\\mathbf{V}$ . This structured format regularizes the event data, making it compatible with learning-based methods and consistent across different domains. To construct each voxel grid $\\mathbf{V}_i\\in \\mathbb{R}^{T\\times H\\times W}$ , where $T$ is the number of temporal bins, and $H$ and $W$ are the spatial dimensions of the event sensor, we divide the event stream into fixed time intervals, with positive and negative events accumulated in separate bins along the temporal dimension. This polarity-sensitive encoding captures spatial and temporal dynamics of events that reflect both motion directionality and density:", + "bbox": [ + 511, + 136, + 906, + 318 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {\\mathbf {e} _ {j} \\in \\varepsilon_ {i}} e _ {j} ^ {p} \\delta \\left(e _ {j} ^ {x} - e ^ {x}\\right) \\delta \\left(e _ {j} ^ {y} - e ^ {y}\\right) \\max \\left\\{1 - \\left| \\hat {e} _ {j} ^ {t} - e ^ {t} \\right|, 0 \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 328, + 906, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta$ is the Kronecker delta function; $\\hat{e}_j^t = (C - 1)\\frac{e_j^t - e_0^t}{\\Delta T}$ normalizes the event timestamps. Here, $\\Delta T$ is the duration of each time window, and $e_0^t$ is the initial timestamp.", + "bbox": [ + 511, + 364, + 903, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Platform Adaptation. We aim to address the challenge of adapting across three distinct domains, represented by $\\mathcal{D} = \\{\\mathcal{P}^{\\mathrm{v}},\\mathcal{P}^{\\mathrm{d}},\\mathcal{P}^{\\mathrm{q}}\\}$ , respectively. Each domain consists of event voxel grids $\\mathbf{V}$ and associated labels $y\\in \\mathbb{R}^{H\\times W}$ , where each pixel label belongs to one of $C$ semantic classes. In our cross-platform adaptation setting, we assume access to fully labeled data from the source platform while the target platform has only unlabeled data. The goal is to leverage the labeled source data and unlabeled target data to improve event-based perception performance on the target domain. $^1$", + "bbox": [ + 511, + 416, + 905, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Event Activation Prior", + "text_level": 1, + "bbox": [ + 513, + 575, + 723, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-platform adaptation in event-based perception hinges on aligning platform-specific activation patterns, which are influenced by domain-unique motions, and environments. We introduce Event Activation Prior (EAP), a regularization technique that encourages confident, low-entropy predictions in target-domain regions with high event activation.", + "bbox": [ + 511, + 597, + 905, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In each platform, event data often exhibits recurrent activation patterns in specific regions, driven by platform-specific motion and scene dynamics (see Fig. 1). For instance, lower regions in a vehicle domain capture details like road surfaces and obstacles, while upper regions in a drone domain capture landscape features and environmental context. Such regions, denoted as $\\mathbf{S} \\subset \\{0,1,\\dots,H-1\\} \\times \\{0,1,\\dots,W-1\\}$ and along with the conditional entropy $H(y_{\\mathbf{S}}|\\mathbf{V}_{\\mathbf{S}},\\mathbf{S})$ , typically have consistent semantic patterns, allowing for more confident predictions. By leveraging these high-activation areas, we aim to minimize the", + "bbox": [ + 511, + 688, + 906, + 854 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "Without loss of generality, our subsequent explanations focus on the adaptation from vehicle domain $\\mathcal{P}^{\\mathrm{v}}$ to drone domain $\\mathcal{P}^{\\mathrm{d}}$ , although the proposed framework generalizes to any pair of domains in $\\mathcal{D}$ .", + "bbox": [ + 511, + 862, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1474", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2a27ad1640cc13ebb94d1818ddcae1cdd246e138d3ba21c7319041f6b6e8f1ed.jpg", + "image_caption": [ + "Figure 2. Overview of the EventFly framework. Guided by the EAP principle (Sec. 3.2), the pair of source and target event data $\\mathbf{V}_i^{\\mathrm{v}}$ and $\\mathbf{V}_i^{\\mathrm{d}}$ are mixed via the EventBlend operation (Sec. 3.3), where the blending mask $\\mathcal{M}_i$ is obtained by measuring the similarities between density maps $\\mathbf{D}_i^{\\mathrm{v}}$ and $\\tilde{\\mathbf{D}}^{\\mathrm{d}}$ . The features $\\mathbf{F}^{\\mathrm{v}}$ , $\\mathbf{F}^{\\mathrm{d}}$ , and $\\tilde{\\mathbf{F}}$ from the source, target, and blended domains are then used for EventMatch (Sec. 3.4). This facilitates learning an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive)." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 883, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "entropy of predictions in such regions, thereby guiding the model to produce more confident outputs aligned with the unique characteristics of the target platform domain.", + "bbox": [ + 88, + 470, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formulating EAP. To incorporate EAP in training, we define it as a regularization that minimizes prediction uncertainty in these high-activation event areas. For each subregion S, EAP aims to minimize the conditional entropy $H(y_{\\mathbf{S}}|\\mathbf{V}_{\\mathbf{S}},\\mathbf{S})$ , focusing on domain-specific high-activation regions. We enforce the expectation constraint over the model parameters $\\theta$ as $\\mathbb{E}_{\\theta}[H(\\mathbf{V}_{\\mathbf{S}},y_{\\mathbf{S}}|\\mathbf{S})]\\leq c$ , where $c$ is a small constant that encourages high-confidence predictions. This is transformed into a prior on parameters $\\theta$ through the maximum entropy principles [31], yielding:", + "bbox": [ + 89, + 517, + 482, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (\\theta) \\propto \\exp (- \\lambda H (\\mathbf {V} _ {\\mathbf {S}}, y _ {\\mathbf {S}} | \\mathbf {S})) \\propto \\exp (- \\lambda H (y _ {\\mathbf {S}} | \\mathbf {V} _ {\\mathbf {S}}, \\mathbf {S})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 676, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda > 0$ is the Lagrange multiplier corresponding to constant $c$ , which balances the effect of EAP on training.", + "bbox": [ + 89, + 707, + 482, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Empirical Estimation of EAP. To apply EAP, we empirically estimate the entropy in high-activation event regions by aggregating over activations in the target platform domain, where activation patterns are highly informative. Using an empirical estimator, we approximate $H(y|\\mathbf{V},\\mathbf{S})$ as:", + "bbox": [ + 89, + 737, + 482, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\text {e m p}} (y | \\mathbf {V}, \\mathbf {S}) = \\mathbb {E} _ {\\mathbf {V}, y, \\mathbf {S}} \\left[ \\hat {P} (y | \\mathbf {V}, \\mathbf {S}) \\log \\hat {P} (y | \\mathbf {V}, \\mathbf {S}) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 820, + 480, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{P}(y|\\mathbf{V},\\mathbf{S})$ is the empirical prediction probability conditioned on the voxel grid $\\mathbf{V}$ and restricted to region $\\mathbf{S}$ . By minimizing $H_{\\mathrm{emp}}(y|\\mathbf{V},\\mathbf{S})$ , we encourage confident", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predictions within these regions, aligning the model's predictions with the target domain activation patterns.", + "bbox": [ + 511, + 470, + 906, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Integrating EAP into Training. To incorporate the EAP in Eq. (2) into cross-platform adaptation, we define the overall objective as a maximum-a-posteriori (MAP) estimation:", + "bbox": [ + 511, + 502, + 906, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC (\\theta) = \\mathcal {L} (\\theta) - \\lambda H _ {\\mathrm {e m p}} (y | \\mathbf {V}, \\mathbf {S}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 556, + 903, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\theta)$ represents the supervised loss on source data. $H_{\\mathrm{emp}}(y|\\mathbf{V},\\mathbf{S})$ minimizes uncertainty in the target domain by leveraging EAP over high-activation regions. By focusing on these regions, the event camera perception model tends to effectively learn to adapt its predictions to align with platform-specific activation cues in the target domain, achieving robust adaptation across varied platforms.", + "bbox": [ + 511, + 583, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.EventBlend", + "text_level": 1, + "bbox": [ + 511, + 696, + 638, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-platform adaptation in event-based perception requires a selective integration of data from both source and target domains. Here, we introduce EventBlend, a data-mixing strategy designed to construct hybrid voxel grids by blending event data from the source and target domains, guided by high-activation regions identified through EAP. As shown in Fig. 2, by constructing these blended voxel grids, we aim to enhance the generalization ability across domain shifts, enabling the model to adapt effectively to the target distribution while leveraging reliable source labels.", + "bbox": [ + 511, + 719, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event Density Maps. To identify regions where the source and target domains exhibit similar or divergent activation", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1475", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "patterns, we employ density maps, which highlight areas of frequent event activity. For each source event data with voxel grid $\\mathbf{V}_i^{\\mathbf{v}}\\in \\mathbb{R}^{T\\times H\\times W}$ , we calculate a corresponding source density map, $\\mathbf{D}_i^{\\mathbf{v}}\\in \\mathbb{R}^{H\\times W}$ , by summing activations along the temporal axis $T$ for each spatial location $(\\mu ,\\nu)$ :", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {i} ^ {\\mathbf {v}} (\\mu , \\nu) = \\sum_ {t = 1} ^ {T} | \\mathbf {V} _ {i} ^ {\\mathbf {v}} (t, \\mu , \\nu) |. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 174, + 482, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the target domain, we pre-compute an aggregated density map, $\\widetilde{\\mathbf{D}}^{\\mathrm{d}}\\in \\mathbb{R}^{H\\times W}$ , by summing activations over all target samples $\\mathbf{V}_j^{\\mathrm{d}}$ , where $j$ indexes each target sample:", + "bbox": [ + 89, + 203, + 483, + 251 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {D}} ^ {\\mathrm {d}} (\\mu , \\nu) = \\frac {1}{N ^ {\\mathrm {d}}} \\sum_ {j} \\sum_ {t = 1} ^ {T} \\left| \\mathbf {V} _ {j} ^ {\\mathrm {d}} (t, \\mu , \\nu) \\right|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 257, + 482, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N^{\\mathrm{d}}$ denotes the total number of samples in the target domain. These density maps highlight regions with significant activity, which aids in identifying areas where the domains may share similar activation patterns, consistent with the high-activation regions established by EAP.", + "bbox": [ + 89, + 291, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarity Map & Binary Mask. To guide selective blending, for each source sample, we construct a similarity map $\\mathbf{SIM}_i\\in \\mathbb{R}^{H\\times W}$ to quantify spatial overlap in activation patterns between two domains. This map is defined as:", + "bbox": [ + 89, + 367, + 483, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S I M} _ {i} (\\mu , \\nu) = 1 - \\left| \\mathbf {D} _ {i} ^ {\\mathbf {v}} (\\mu , \\nu) - \\tilde {\\mathbf {D}} ^ {\\mathrm {d}} (\\mu , \\nu) \\right|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 434, + 482, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, high values in $\\mathbf{SIM}_i$ correspond to regions where source and target domains exhibit similar activation intensities, aligning with the high-activation areas specified by EAP in Eq. (2). Next, we derive a binary mask $\\mathcal{M}_i\\in$ $\\{0,1\\}^{H\\times W}$ by applying a threshold $\\tau$ to the similarity map:", + "bbox": [ + 89, + 455, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {i} (\\mu , \\nu) = \\left\\{ \\begin{array}{l l} \\mathbf {1}, & \\text {i f} \\mathbf {S I M} _ {i} (\\mu , \\nu) \\geq \\tau , \\\\ \\mathbf {0}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 539, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This mask directs the blending process by specifying which regions to retain from the source and which to adapt from the target. A mask value of 1 indicates that activations in $(\\mu, \\nu)$ from the source domain should be retained, while a value of 0 specifies a shift to the target domain activations. It is worth mentioning that for the selections of $\\mathbf{D}_i^{\\mathrm{v}}$ and $\\tilde{\\mathbf{D}}^{\\mathrm{d}}$ in Eq. (5) and Eq. (6), we only consider activated pixel locations. That is to say, the similarity score does not count cases where both source and target are non-activated.", + "bbox": [ + 89, + 585, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Constructing Blended Event Voxel Grids. For a pair of source and target samples $\\mathbf{V}_i^{\\mathrm{v}}$ and $\\mathbf{V}_i^{\\mathrm{d}}$ , we create a new event voxel grid $\\tilde{\\mathbf{V}}_i \\in \\mathbb{R}^{T \\times H \\times W}$ , using $\\mathcal{M}_i$ . For each spatial location $(\\mu, \\nu)$ , temporal sequences are selectively copied from either source and target based on $\\mathcal{M}_i(\\mu, \\nu)$ :", + "bbox": [ + 89, + 722, + 483, + 799 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {V}} _ {i} (:, \\mu , \\nu) = \\left\\{ \\begin{array}{l l} \\mathbf {V} _ {i} ^ {\\mathbf {v}} (:, \\mu , \\nu), & \\text {i f} \\mathcal {M} (\\mu , \\nu) = \\mathbf {1}, \\\\ \\mathbf {V} _ {i} ^ {\\mathrm {d}} (:, \\mu , \\nu), & \\text {i f} \\mathcal {M} (\\mu , \\nu) = \\mathbf {0}. \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 806, + 482, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This selective copying approach allows $\\tilde{\\mathbf{V}}_i$ to incorporate source-domain stability (i.e., human annotations) in high-similarity regions while adapting to target-specific patterns", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in low-similarity areas. Finally, we generate the label map $\\tilde{y}_i\\in \\mathbb{R}^{H\\times W}$ for $\\tilde{\\mathbf{V}}_i$ by combining ground truth labels from the source and pseudo-labels from the target. Here, the pseudo-labels can be obtained offline from a source-pretrained model or generated online using a mean teacher framework [76]. During adaptation, we use $\\tilde{\\mathbf{V}}_i$ and $\\tilde{y}_i$ for supervised learning. The blended data contains activations and labels from both domains. We leverage EAP-aligned high-activation areas, i.e., Eq. (4), where source-target consistency is high, facilitating robust cross-domain adaptation.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.EventMatch", + "text_level": 1, + "bbox": [ + 513, + 252, + 645, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In cross-platform adaptation, domain shifts manifest as discrepancies in spatial activation patterns, motion characteristics, and perspectives unique to each platform. The EAP (cf. Sec. 3.2) identifies high-activation areas for stable adaptation, while EventBlend (cf. Sec. 3.3) generates hybrid voxel grids that capture shared spatial structures across domains. However, without explicit feature alignment, the model might still struggle to fully adapt to platform-specific idiosyncrasies, especially in target-specific contexts.", + "bbox": [ + 511, + 273, + 906, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we propose the EventMatch technique to align the blended domain as an intermediary, enabling robust domain-invariant feature learning by maintaining source reliability while selectively adapting to target-specific characteristics. We introduce two fully convolutional discriminators, $\\sigma_{1}$ and $\\sigma_{2}$ , each responsible for a distinct aspect of feature alignment:", + "bbox": [ + 511, + 411, + 905, + 516 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Source & Blended (via $\\sigma_{1}$ ): We align the blended voxel grids $\\tilde{\\mathbf{V}}$ with the source domain, encouraging them to retain reliable source-domain characteristics. This alignment preserves robustness and ensures that the blended features maintain patterns learned from the ground truth supervised labels, i.e., $y$ , from the source domain.", + "- Blended & Target (via $\\sigma_{2}$ ): Rather than the strict alignment, we regularize the blended features toward a soft alignment with the target, selectively adapting target-relevant characteristics. This alignment is focused on high-activation regions identified by EAP, allowing the blended features to bridge the domain gap in target-specific contexts without sacrificing source consistency." + ], + "bbox": [ + 513, + 517, + 905, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This dual-discriminator setup allows the blended features to act as an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive).", + "bbox": [ + 511, + 713, + 905, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Domain Adversarial Training. Each discriminator learns to distinguish between the feature representations from its respective domains, guiding the model to produce domain-agnostic features. Given feature outputs $\\mathbf{F}^{\\mathrm{v}}$ , $\\mathbf{F}^{\\mathrm{d}}$ , and $\\tilde{\\mathbf{F}}$ for the source, target, and blended domains, respectively:", + "bbox": [ + 511, + 758, + 905, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Source-Blended Alignment: Discriminator $\\sigma_{1}$ classifies $\\mathbf{F}^{\\vee}$ as source and $\\tilde{\\mathbf{F}}$ as blended, maximizing $\\mathcal{L}_{\\sigma_1}$ as:", + "bbox": [ + 513, + 835, + 905, + 866 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n- \\sum_ {\\mu , \\nu} \\left[ \\left(1 - z _ {1}\\right) \\log \\sigma_ {1} (\\tilde {\\mathbf {F}}) + z _ {1} \\log \\sigma_ {1} \\left(\\mathbf {F} ^ {\\mathrm {v}}\\right) \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 876, + 905, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1476", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/00399096352dbfce608a591df080634c69df823a9b637a5d9149f782f9b5bcbd.jpg", + "table_caption": [ + "Table 1. Benchmark results of platform adaptation from vehicle $(\\mathcal{P}^{\\mathrm{v}})$ to drone $(\\mathcal{P}^{\\mathrm{d}})$ . Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are highlighted in colors." + ], + "table_footnote": [], + "table_body": "
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o43.6933.8115.0411.8148.7111.570.928.4213.3325.488.1831.5114.880.042.41
AdaptSegNet [79]49.1435.3821.1612.1529.3723.570.170.4813.4538.2317.8548.7329.4235.550.40
CBST [104]57.9541.1824.3116.0233.0524.430.003.0818.2456.3216.8456.1523.6135.650.00
IntraDA [61]57.3738.8523.5815.9132.3123.170.004.9014.9156.7018.6754.9420.7133.080.00
DACS [77]59.8142.0127.0716.1435.1626.120.184.1118.4955.6421.7456.8134.6944.730.05
MIC [38]63.1145.6028.8717.4641.4025.190.0110.1122.8659.2520.8458.8633.9544.180.90
PLSR [94]64.6145.9329.6917.9942.0930.060.009.7523.3262.4820.6560.1531.6944.272.06
EventFly (Ours)69.1748.2032.6720.0146.6430.551.2710.9125.5067.1724.2161.0141.3044.546.21
Target •79.5752.2542.9023.3074.4839.407.100.3331.6771.9631.6467.8757.5166.1423.79
", + "bbox": [ + 94, + 122, + 906, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z_{1} = 1$ for source and $z_{1} = 0$ for blended. This alignment ensures that the blended features preserve the supervised stability from the source.", + "bbox": [ + 102, + 311, + 480, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Blended-Target Adaptation: Discriminator $\\sigma_{2}$ adapts $\\tilde{\\mathbf{F}}$ toward high-activation regions in the target, where target-relevant properties are critical. It classifies $\\tilde{\\mathbf{F}}$ as blended and $\\mathbf{F}^{\\mathrm{d}}$ as target, maximizing $\\mathcal{L}_{\\sigma_2}$ as:", + "bbox": [ + 89, + 357, + 482, + 419 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n- \\sum_ {\\mu , \\nu} \\left[ \\left(1 - z _ {2}\\right) \\log \\sigma_ {2} \\left(\\mathbf {F} ^ {\\mathrm {d}}\\right) + z _ {2} \\log \\sigma_ {2} (\\tilde {\\mathbf {F}}) \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 426, + 480, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z_{2} = 1$ for target and $z_{2} = 0$ for blended. Here, $\\sigma_{2}$ does not force complete target alignment but rather emphasizes adaptation in target-relevant areas, enabling $\\tilde{\\mathbf{F}}$ to maintain a balanced representation across domains.", + "bbox": [ + 102, + 463, + 482, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimization. To achieve consistent cross-domain alignment, the backbone network is trained adversarially against both discriminators. The goal is to produce $\\tilde{\\mathbf{F}}$ that each discriminator finds challenging to classify distinctly, thereby promoting a domain-agnostic representation:", + "bbox": [ + 89, + 523, + 482, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d v}} = - \\sum_ {\\mu , \\nu} \\left[ \\log \\sigma_ {1} (\\tilde {\\mathbf {F}}) + \\log \\sigma_ {2} (\\tilde {\\mathbf {F}}) \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 607, + 482, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This objective ensures that the backbone network produces features for blended data that capture both source robustness and target adaptability, guided by high-activation areas where source-target consistency is most beneficial. To achieve this goal, we set two weight coefficients, $\\phi_{1}$ and $\\phi_{2}$ , to balance between these two objectives.", + "bbox": [ + 89, + 643, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, we employ a min-max optimization framework to iteratively update the backbone, $\\sigma_{1}$ , and $\\sigma_{2}$ :", + "bbox": [ + 89, + 734, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Step 1: Update $\\sigma_{1}$ and $\\sigma_{2}$ to improve the classification of the source, blended, and target features.", + "- Step 2: Update the backbone network to minimize $\\mathcal{L}_{\\mathrm{adv}}$ , encouraging it to generate intermediary features that support robust cross-domain generalization." + ], + "bbox": [ + 89, + 765, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through this adversarial process, EventMatch allows the model to bridge the domain gap effectively, with blended features serving as a reliable intermediary that balances source stability with target-specific adaptability.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Framework. As shown in Fig. 2, the source and blended domains share the same network, while the network from the target domain is updated via exponential-moving average (EMA), encouraging consistency among domains.", + "bbox": [ + 511, + 311, + 906, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. The EXPo Benchmark", + "text_level": 1, + "bbox": [ + 511, + 381, + 723, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate robust event camera perception across diverse platforms and environments, we establish a large-scale event-based cross-platform adaptation benchmark: EXPo. Our benchmark is based on [12]. We include 89, 228 frames captured from three distinct platforms - vehicle, drone, and quadruped - spanning 21 sequences: 6 from the vehicle, 7 from the drone, and 8 from the quadruped. Following DSEC-Semantic [75], we support 11 semantic classes relevant to real-world event-based perception and cover varied environments including city, urban, suburban, and rural scenes. Due to space limits, more details are in Appendix.", + "bbox": [ + 511, + 402, + 906, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 582, + 643, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 511, + 606, + 611, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our framework is implemented using PyTorch [62]. The backbone network is from E2VID [67], while the event segmentation head is from ESS [75]. The hyperparameters $\\tau$ , $\\phi_{1}$ and $\\phi_{2}$ are set to 0.4, 1e-3, and 2e-3, respectively. The model is optimized using AdamW [54] and OneCycle learning rate scheduler [72] for 100k iterations. The learning rate and batch size are set to 1e-3 and 8. Due to the lack of baselines, we reproduce popular frame-based adaptation methods for adversarial training, contrastive learning, and self-training [38, 61, 77, 79, 94, 104]. To ensure fair comparisons, all methods adopt the same backbone and training iterations. The models are trained using NVIDIA RTX 4090 GPUs.", + "bbox": [ + 511, + 628, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following conventions, we use Accuracy (Acc) and Intersection-over-Union (IoU) as the main metrics for evaluation. We also report mean Acc (mAcc), mean IoU (mIoU), and frequency-weighted IoU (fIoU) for a holistic comparison among different adaptation methods.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1477", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/95c16d9f3819da026bb5ca62eb49f1b865d051c95309bc65b394a41e1a7755da.jpg", + "table_caption": [ + "Table 2. Benchmark results of platform adaptation from vehicle $(\\mathcal{P}^{\\mathrm{v}})$ to quadruped $(\\mathcal{P}^{\\mathrm{q}})$ . Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are in colors." + ], + "table_footnote": [], + "table_body": "
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o66.5939.7325.1516.5263.0139.263.8817.8810.1251.679.2768.0212.350.240.99
AdaptSegNet [79]67.2548.7332.7914.8945.0045.8830.0034.9212.2255.5015.8573.8416.0731.350.00
CBST [104]69.2549.5835.0614.9547.3954.6834.2736.8313.7856.1518.1374.2316.1834.060.00
IntraDA [61]68.2948.9134.2514.8243.7555.3632.6433.3911.6055.3117.0076.0020.3031.400.00
DACS [77]69.5553.8836.5114.6643.7257.2738.4335.4214.0257.1018.4376.1624.7936.210.00
MIC [38]70.7849.2236.9315.6051.7151.7333.5438.109.4454.2720.7474.4029.7941.780.70
PLSR [94]70.9153.6537.5715.2549.0453.2837.5436.6412.9157.6025.2975.9224.9239.850.24
EventFly (Ours)73.4254.1440.0515.7850.0761.3339.1741.9712.8359.1423.5179.8027.2642.652.86
Target •80.0260.5549.8419.5874.8056.2346.0855.2821.7959.9030.3177.2458.3862.475.81
", + "bbox": [ + 94, + 125, + 906, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d621328f44ad87ad0dbce66dfbdf07630da0a31a37f18ae04fc19c799463ee39.jpg", + "table_caption": [ + "Table 3. Benchmark results of platform adaptation among the vehicle $(\\mathcal{P}^{\\mathrm{v}})$ , drone $(\\mathcal{P}^{\\mathrm{d}})$ , and quadruped $(\\mathcal{P}^{\\mathrm{q}})$ platforms, respectively. A total of six cross-platform adaptation settings are considered in our benchmark. Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are highlighted in colors." + ], + "table_footnote": [], + "table_body": "
MethodPv → PdPv → PqPd → PvPd → PQPq → PvPq → PdAverage
AccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoU
Source-Only o43.6915.0466.5925.1557.9120.7966.8323.0657.4921.3052.6216.8557.5220.37
AdaptSegNet [79]49.1421.1667.2532.7968.2929.5567.5733.9966.7430.6557.0720.9662.8828.18
DACS [77]59.8127.0769.5536.5171.7836.1067.7336.1171.2034.7860.7424.5066.8032.51
MIC [38]63.1128.8770.7836.9372.4636.8867.2936.2772.4635.2264.4926.1168.4333.38
PLSR [94]64.6129.6970.9137.5772.4637.1867.8336.2172.9336.3863.5727.3468.7234.06
EventFly (Ours)69.1732.6773.4240.0575.5039.9269.6837.3773.9337.7065.7828.7971.2536.08
Target •79.5742.9080.0249.8486.1255.9380.0249.8486.1255.9379.5742.9081.9049.56
", + "bbox": [ + 94, + 351, + 906, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparative Study", + "text_level": 1, + "bbox": [ + 89, + 497, + 277, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adapt from Vehicle to Drone. Tab. 1 presents the results of transferring vehicle to drone, where there is a huge domain gap in between. The source-only (15.04% mIoU) and supervised target (42.90% mIoU) models validate this gap. Prior adaptation methods show good results in closing the domain gap, while our EventFly surpasses all competitors for almost all semantic classes, demonstrating strong robustness enhancements for event-based cross-platform adaptation.", + "bbox": [ + 88, + 520, + 482, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adapt from Vehicle to Quadruped. We compare the adaptation methods from vehicle to quadruped in Tab. 2. Since the data and semantic distributions of these two platforms are closer (compared to drone), the adaptation methods show larger performance gains. We achieve a promising $40.05\\%$ mIoU under this setting, which corresponds to a $59.2\\%$ relative improvement over the source-only baseline.", + "bbox": [ + 88, + 642, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Assessments. Fig. 3 visually compares the abilities of different cross-platform adaptation methods. As can be seen, the large domain gap between source and target platforms causes huge performance degradation, resulting in messy semantic predictions (source-only). While state-of-the-art adaptation methods [38, 94] restore the performance to some extent, their predictions still suffer from ambiguous object and background contours. Differently, our EventFly generates more accurate scene semantics, credited to the superior robustness from the EAP-driven designs.", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Platform Adaptations. Since our EXPo benchmark consists of three platforms, we benchmark state-of-the-art methods across a total of six adaptation settings, where each platform interchangeably serve as the source/target domains. As shown in Tab. 3, we observe discrepancies in adaptation difficulties. Specifically, the drone domain is the most difficult to adapt to, mainly due to its unique motion patterns and perspective-based dynamic. Our approach exhibits better robustness than prior arts, achieving the best metrics across all six settings. Such a strong generalization ability is crucial for the deployment of event camera perception algorithms in real-world environments.", + "bbox": [ + 511, + 498, + 906, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 694, + 666, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Analysis. Our EventFly framework compromises three key components: EventBlend, EventMatch, and the dual-backbone network. As shown in Tab. 4, each contribution provides distinct performance improvements. Since our designs are well-motivated by EAP, the combinations of any two (row $\\# c$ ) and all (row $\\# d$ ) of them further yield better cross-platform adaptation performance.", + "bbox": [ + 511, + 717, + 905, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain Blending Techniques. To validate that our EAP-driven EventBlend operation indeed encourages domain regularization effect, we compare the heuristic data mixing techniques [19, 90] and recent domain mixing methods [23, 77]. As shown in Fig. 4, conventional methods, which", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1478", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a8585a447e0f2af9a175bca200cf29852f4f59e30ab08c245c597e5c292f420.jpg", + "image_caption": [ + "Figure 3. Qualitative assessments of adaptation from vehicle to drone (the first 3 rows), and from vehicle to quadruped (the last 3 rows). We use grayscaled event images for better visibility. The RGB frames are for reference purposes only. Best viewed in colors." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 419 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ae61baa0bcd517cf09014618dc128da072de917116deef5aacb231c4d52afe8f.jpg", + "table_caption": [ + "Table 4. Ablation results of each component in cross-platform adaptation from vehicle $(\\mathcal{P}^{\\mathrm{v}})$ to drone $(\\mathcal{P}^{\\mathrm{d}})$ and vehicle $(\\mathcal{P}^{\\mathrm{v}})$ to quadruped $(\\mathcal{P}^{\\mathrm{q}})$ , respectively. Dual: The dual-branch network. Match: The EventMatch technique. Blend: The EventBlend operation. All scores are given in percentage (\\%). We compare settings from the (a) source-only, (b) single-component effects, (c) double-component effects, (d) full configuration, and (e) supervised target results, respectively, in this ablation study." + ], + "table_footnote": [], + "table_body": "
#DualMatchBlendPvAccPd mIoUPvAccPq mIoU
(a)XXX43.6915.0466.5925.15
(b)X✓X59.8625.2867.7433.66
XX✓61.8527.9169.5536.82
(c)✓✓X66.5030.2470.9038.10
✓X✓68.4331.9672.6739.11
(d)✓✓✓69.1732.6773.4240.05
(e)Target79.5742.9080.0249.84
", + "bbox": [ + 94, + 577, + 483, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "neglect the event activation discrepancies, show sub-par performance in bridging the domains. Guided by EAP, our domain blending technique facilitates entropy minimization in high-activation regions, which in turn brings robust feature adaptation across heterogeneous platforms.", + "bbox": [ + 89, + 733, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarity Threshold. The strength of blending source and target platforms is vital in our framework. Therefore, we study the effect of the threshold parameter $\\tau$ . Notably, the 0 (i.e., source-only) or high ( $>0.7$ ) value of $\\tau$ brings limited gains, as the domain consistency regularization is weak. We found that the best possible trade-offs appear between 0.3", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f344061038c8e4ec8f53cf36f49f6a993f97e0b87c40e240bdcb61b0af0f7fb.jpg", + "image_caption": [ + "Figure 4. Ablation study (mIoU scores in %) on different domain-mixing techniques under the (left) $\\mathcal{P}^{\\mathrm{v}}\\rightarrow \\mathcal{P}^{\\mathrm{d}}$ , (middle) $\\mathcal{P}^{\\mathrm{v}}\\rightarrow \\mathcal{P}^{\\mathrm{q}}$ , and (right) $\\mathcal{P}^{\\mathrm{d}}\\rightarrow \\mathcal{P}^{\\mathrm{v}}$ cross-platform adaptation settings." + ], + "image_footnote": [], + "bbox": [ + 514, + 464, + 906, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and 0.5, further validating the importance of bridging domains for cross-platform adaptation.", + "bbox": [ + 511, + 646, + 905, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented EventFly, a framework for robust cross-platform adaptation in event-based dense perception, enabling robust deployment across the vehicle, drone, and quadruped platforms. Leveraging the proposed Event Activation Prior (EAP), EventBlend, and EventMatch, we addressed the unique adaptation challenges of event stream data collected from different robot platforms. We also introduced EXPo, a large-scale benchmark designed to evaluate cross-platform event perception capabilities. Our approach demonstrated notable improvements over existing methods. We hope this work lays the foundation for more adaptive event-based perception across real-world environments.", + "bbox": [ + 509, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1479", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is under the programme DesCartes and is supported by the National Research Foundation, Prime Minister's Office, Singapore, under its Campus for Research Excellence and Technological Enterprise (CREATE) programme. This work is also supported by the Apple Scholars in AI/ML Ph.D. Fellowship program.", + "bbox": [ + 89, + 116, + 483, + 207 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors would like to sincerely thank the Program Chairs, Area Chairs, and Reviewers for the time and efforts devoted during the review process.", + "bbox": [ + 89, + 208, + 483, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 272, + 187, + 287 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Inigo Alonso and Ana C. Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1-10, 2019.", + "[2] Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. Ddd17: End-to-end davis driving dataset. In International Conference on Machine Learning Workshops, pages 1-9, 2017.", + "[3] Shristi Das Biswas, Adarsh Kosta, Chamika Liyanagedera, Marco Apolinario, and Kaushik Roy. Halsie: Hybrid approach to learning segmentation by simultaneously exploiting image and event modalities. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5952-5962, 2024.", + "[4] Michael Blösch, Stephan Weiss, Davide Scaramuzza, and Roland Siegwart. Vision based MAV navigation in unknown and unstructured environments. In IEEE International Conference on Robotics and Automation, pages 21-28, 2010.", + "[5] Chiara Boretti, Philippe Bich, Fabio Pareschi, Luciano Prono, Riccardo Rovatti, and Gianluca Setti. Pedro: An event-based dataset for person detection in robotics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4065-4070, 2023.", + "[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db 3 μs latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014.", + "[7] Tim Brodermann, David Bruggemann, Christos Sakaridis, Kevin Ta, Odysseas Liagouris, Jason Corkill, and Luc Van Gool. Muses: The multi-sensor semantic perception dataset for driving under uncertainty. arXiv preprint arXiv:2401.12761, 2024.", + "[8] David Brüggemann, Christos Sakaridis, Prune Truong, and Luc Van Gool. Refign: Align and refine for adaptation of semantic segmentation to adverse conditions. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3174-3184, 2023.", + "[9] Spencer Carmichael, Austin Buchan, Mani Ramanagopal, Radhika Ravi, Ram Vasudevan, and Katherine A. Skinner. Dataset and benchmark: Novel sensors for autonomous vehicle perception. International Journal of Robotics Research, 2024." + ], + "bbox": [ + 107, + 297, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In IEEE International Conference on Robotics and Automation, pages 703-710, 2014.", + "[11] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-ilidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. In European Conference on Computer Vision Workshops, 2024.", + "[12] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multirobot, multi-sensor, multi-environment event dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4016-4023, 2023.", + "[13] Peiyu Chen, Weipeng Guan, Feng Huang, Yihan Zhong, Weisong Wen, Li-Ta Hsu, and Peng Lu. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 9(1):407-416, 2024.", + "[14] Zhiwen Chen, Zhiyu Zhu, Yifan Zhang, Junhui Hou, Guangming Shi, and Jinjian Wu. Segment any event streams via weighted adaptation of pivotal tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890-3900, 2024.", + "[15] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023.", + "[16] Fernando Cladera, Kenneth Chaney, M. Ani Hsieh, Camillo J. Taylor, and Vijay Kumar. Evmapper: High altitude orthomapping with event cameras. arXiv preprint arXiv:2409.18120, 2024.", + "[17] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016.", + "[18] Javier Cuadrado, Ulysse Rançon, Benoit R. Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023.", + "[19] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017.", + "[20] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 3942-3951, 2023.", + "[21] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Hue dataset: High-resolution event and frame sequences for low-light vision. In European Conference on Computer Vision Workshops, pages 1-18, 2024.", + "[22] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a $1280 \\times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\\mu \\mathrm{m}$" + ], + "bbox": [ + 522, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1480", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In IEEE International Solid-State Circuits Conference, pages 112-114, 2020.", + "[23] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In British Machine Vision Conference, 2020.", + "[24] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 (1):154-180, 2022.", + "[25] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022.", + "[26] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014): 1034-1040, 2024.", + "[27] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020.", + "[28] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023.", + "[29] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021.", + "[30] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In IEEE International Conference on 3D Vision, pages 197-206, 2021.", + "[31] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems, pages 529-536, 2004.", + "[32] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22867-22876, 2023.", + "[33] Dalia Hareb and Jean Martinet. Evsegsn: Neuromorphic semantic segmentation for event data. arXiv preprint arXiv:2406.14178, 2024.", + "[34] Javier Hidalgo-Carrio, Daniel Gehrig, and Davide Scaramuzza. Learning monocular dense depth from events. In IEEE International Conference on 3D Vision, pages 534-542, 2020.", + "[35] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781-5790, 2022.", + "[36] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for" + ], + "bbox": [ + 99, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "domain-adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022.", + "[37] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision, pages 372–391, 2022.", + "[38] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023.", + "[39] Yuhwan Jeong, Hoonhee Cho, and Kuk-Jin Yoon. Towards robust event-based networks for nighttime via unpaired day-to-night event translation. In European Conference on Computer Vision, pages 286-306, 2024.", + "[40] Linglin Jing, Yiming Ding, Yunpeng Gao, Zhigang Wang, Xu Yan, Dong Wang, Gerald Schaefer, Hui Fang, Bin Zhao, and Xuelong Li. Hpl-ess: Hybrid pseudo-labeling for unsupervised event-based semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23128–23137, 2024.", + "[41] Uday Kamal and Saibal Mukhopadhyay. Efficient learning of event-based dense representation using hierarchical memories with adaptive update. In European Conference on Computer Vision, pages 74-89, 2024.", + "[42] Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 620(7976):982-987, 2023.", + "[43] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In IEEE/CVF International Conference on Computer Vision, pages 2146-2156, 2021.", + "[44] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2496-24976, 2024.", + "[45] Simon Klenk, David Bonello, Lukas Koestler, Nikita Araslanov, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2378-2388, 2024.", + "[46] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023.", + "[47] Lingdong Kong, Youquan Liu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Openness: Event-based semantic scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15686-15698, 2024.", + "[48] Wouter M. Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):766-785, 2021." + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1481", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):14020-14037, 2023.", + "[50] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5743-5762, 2024.", + "[51] Yahao Liu, Jinhong Deng, Xinchen Gao, Wen Li, and Lixin Duan. Bapa-net: Boundary adaptation and prototype alignment for cross-domain semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 8801-8811, 2021.", + "[52] Antonio Loquercio, Ana I. Maqueda, Carlos R. Del-Blanco, and Davide Scaramuzza. Dronet: Learning to fly by driving. IEEE Robotics and Automation Letters, 3(2):1088-1095, 2018.", + "[53] Antonio Loquercio, Elia Kaufmann, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Learning high-speed flight in the wild. Science Robotics, 6(59):5810, 2021.", + "[54] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019.", + "[55] Dongyue Lu, Lingdong Kong, Gim Hee Lee, Camille Simon Chane, and Wei Tsang Ooi. Flexevent: Event camera object detection at arbitrary frequencies. arXiv preprint arXiv:2412.06708, 2024.", + "[56] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso García, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5419-5427, 2018.", + "[57] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. IEEE Robotics and Automation Letters, 7(2):3515-3522, 2022.", + "[58] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021.", + "[59] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34 (6):1425-1440, 2018.", + "[60] Poojan Oza, Vishwanath A. Sindagi, Vibashan VS, and Vishal M. Patel. Unsupervised domain adaptation of object detectors: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4018-4040, 2024.", + "[61] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020.", + "[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison," + ], + "bbox": [ + 99, + 90, + 485, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8026-8037, 2019.", + "[63] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2024.", + "[64] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems, pages 16639-16652, 2020.", + "[65] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1): 259-275, 2010.", + "[66] Ulysse Rançon, Javier Cuadrado-Anibarro, Benoit R. Cottereau, and Timothée Masquelier. Stereospike: Depth learning with a spiking neural network. IEEE Access, 10: 127428-127439, 2022.", + "[67] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):1964-1980, 2019.", + "[68] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8197-8207, 2021.", + "[69] Manuel Schwonberg, Joshua Niemeijer, Jan-Aike Termöhlen, Nico M. Schmidt, Hanno Gottschalk, and Tim Fingscheidt. Survey on unsupervised domain adaptation for semantic segmentation for visual perception in automated driving. IEEE Access, 11:54296-54336, 2023.", + "[70] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 12:51275-51306, 2024.", + "[71] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.", + "[72] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. arXiv preprint arXiv:1708.07120, 2017.", + "[73] Bongki Son, Yunjae Suh, Sungho Kim, Heejae Jung, Jun-Seok Kim, Changwoo Shin, Keunju Park, Kyoobin Lee, Jinman Park, Jooyeon Woo, et al. 4.1 a $640 \\times 480$ dynamic vision sensor with a $9\\mu \\mathrm{m}$ pixel and 300meps address-event representation. In IEEE International Solid-State Circuits Conference, pages 66-67, 2017." + ], + "bbox": [ + 522, + 92, + 903, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1482", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Lea Steffen, Daniel Reichard, Jakob Weinland, Jacques Kaiser, Arne Roennau, and Rüdiger Dillmann. Neuromorphic stereo vision: A survey of bio-inspired sensors and algorithms. Frontiers in Neuroscience, 13:28, 2019.", + "[75] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In European Conference on Computer Vision, pages 341-357, 2022.", + "[76] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 195–1204, 2017.", + "[77] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021.", + "[78] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshops, pages 350-357, 2021.", + "[79] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7472-7481, 2018.", + "[80] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate slam? combining events, images, and imu for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3 (2):994-1001, 2018.", + "[81] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517-2526, 2019.", + "[82] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019.", + "[83] Lin Wang, Yujeong Chae, and Kuk-Jin Yoon. Dual transfer learning for event-based end-task prediction via plug-gable event to image translation. In IEEE/CVF International Conference on Computer Vision, pages 2135-2145, 2021.", + "[84] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 608-619, 2021.", + "[85] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In IEEE/CVF Conference on" + ], + "bbox": [ + 99, + 92, + 480, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 19248-19257, 2024.", + "[86] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyuan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 21572-21581, 2023.", + "[87] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9004-9021, 2023.", + "[88] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In IEEE/CVF International Conference on Computer Vision, pages 10699-10709, 2023.", + "[89] Zhen Yao and Mooi Choo Chuah. Event-guided low-light video semantic segmentation. arXiv preprint arXiv:2411.00639, 2024.", + "[90] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018.", + "[91] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12):14679-14694, 2023.", + "[92] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12414-12424, 2021.", + "[93] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, and Sebastian Scherer. Subt-mrs dataset: Pushing slam towards all-weather environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024.", + "[94] Xingchen Zhao, Niluthpol Chowdhury Mithun, Abhinav Rajvanshi, Han-Pang Chiu, and Supun Samarasekera. Unsupervised domain adaptation for semantic segmentation with pseudo label self-refinement. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2399-2409, 2024.", + "[95] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024.", + "[96] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023.", + "[97] Qianyu Zhou, Zhengyang Feng, Qiqi Gu, Jiangmiao Pang, Guangliang Cheng, Xuequan Lu, Jianping Shi, and" + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1483", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lizhuang Ma. Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):804-817, 2023.", + "[98] Zhuyun Zhou, Zongwei Wu, Rémi Boutteau, Fan Yang, Cédric Demonceaux, and Dominique Gingac. Rgb-event fusion for moving object detection in autonomous driving. In IEEE International Conference on Robotics and Automation, pages 7808-7815, 2023.", + "[99] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Robotics: Science and Systems, 2018.", + "[100] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018.", + "[101] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019.", + "[102] Shifan Zhu, Zixun Xiong, and Donghyun Kim. Cear: Comprehensive event camera dataset for rapid perception of agile quadruped robots. IEEE Robotics and Automation Letters, 9(10):8999-9006, 2024.", + "[103] Rong Zou, Manasi Muglikar, Nico Messikommer, and Davide Scaramuzza. Seeing behind dynamic occlusions with event cameras. arXiv preprint arXiv:2307.15829, 2023.", + "[104] Yang Zou, Zhiding Yu, B. V. K. Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In European Conference on Computer Vision, pages 289-305, 2018.", + "[105] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, pages 12846-128567, 2023." + ], + "bbox": [ + 93, + 92, + 480, + 614 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1484", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_model.json b/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_model.json new file mode 100644 index 0000000000000000000000000000000000000000..30eba10ee3ea6e67d8fc1f71725ed26af1877cac --- /dev/null +++ b/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_model.json @@ -0,0 +1,4230 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.021 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.018, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.11, + 0.804, + 0.132 + ], + "angle": 0, + "content": "EventFly: Event Camera Perception from Ground to the Sky" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.159, + 0.905, + 0.179 + ], + "angle": 0, + "content": "Lingdong Kong\\(^{1,2}\\), Dongyue Lu\\(^{1}\\), Xiang Xu\\(^{3}\\), Lai Xing Ng\\(^{4}\\), Wei Tsang Ooi\\(^{1,5}\\), Benoit R. Cottereau\\(^{5,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.183, + 0.832, + 0.2 + ], + "angle": 0, + "content": "\\(^{1}\\)National University of Singapore \\(^{2}\\)CNRS@CREATE \\(^{3}\\)Nanjing University of Aeronautics and Astronautics" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.201, + 0.903, + 0.217 + ], + "angle": 0, + "content": "\\(^{4}\\)Institute for Infocomm Research, A*STAR \\(^{5}\\)IPAL, CNRS IRL 2955, Singapore \\(^{6}\\)CerCo, CNRS UMR 5549, Université Toulouse III" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.221, + 0.728, + 0.24 + ], + "angle": 0, + "content": "Project Page: https://event-fly.github.io" + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.255, + 0.2, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.258, + 0.27, + 0.272 + ], + "angle": 0, + "content": "Vehicle" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.284, + 0.156, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.307, + 0.171, + 0.318 + ], + "angle": 0, + "content": "Viewpoint" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.285, + 0.243, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.307, + 0.245, + 0.318 + ], + "angle": 0, + "content": "Speed" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.285, + 0.327, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.307, + 0.34, + 0.318 + ], + "angle": 0, + "content": "Stability" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.255, + 0.484, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.487, + 0.259, + 0.542, + 0.272 + ], + "angle": 0, + "content": "Drone" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.284, + 0.43, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.307, + 0.445, + 0.318 + ], + "angle": 0, + "content": "Viewpoint" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.285, + 0.518, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.307, + 0.52, + 0.318 + ], + "angle": 0, + "content": "Speed" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.285, + 0.6, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.307, + 0.612, + 0.318 + ], + "angle": 0, + "content": "Stability" + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.255, + 0.734, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.738, + 0.259, + 0.834, + 0.273 + ], + "angle": 0, + "content": "Quadruped" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.284, + 0.704, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.307, + 0.719, + 0.318 + ], + "angle": 0, + "content": "Viewpoint" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.285, + 0.791, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.307, + 0.794, + 0.318 + ], + "angle": 0, + "content": "Speed" + }, + { + "type": "image", + "bbox": [ + 0.846, + 0.285, + 0.874, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.307, + 0.887, + 0.318 + ], + "angle": 0, + "content": "Stability" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.326, + 0.123, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.331, + 0.144, + 0.343 + ], + "angle": 0, + "content": "High" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.344, + 0.167, + 0.356 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.356, + 0.167, + 0.368 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.368, + 0.144, + 0.375 + ], + "angle": 0, + "content": "Low" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.327, + 0.205, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.331, + 0.231, + 0.342 + ], + "angle": 0, + "content": "Fast" + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.344, + 0.253, + 0.356 + ], + "angle": 0, + "content": "Medium " + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.356, + 0.253, + 0.368 + ], + "angle": 0, + "content": "G" + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.368, + 0.233, + 0.378 + ], + "angle": 0, + "content": "Slow" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.327, + 0.291, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.331, + 0.337, + 0.341 + ], + "angle": 0, + "content": "Smooth" + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.343, + 0.339, + 0.355 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.356, + 0.343, + 0.367 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.367, + 0.344, + 0.377 + ], + "angle": 0, + "content": "Unstable" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.327, + 0.393, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.331, + 0.42, + 0.344 + ], + "angle": 0, + "content": "High" + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.346, + 0.441, + 0.358 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.358, + 0.418, + 0.369 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.369, + 0.418, + 0.377 + ], + "angle": 0, + "content": "Low" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.327, + 0.476, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.331, + 0.505, + 0.342 + ], + "angle": 0, + "content": "Fast" + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.347, + 0.527, + 0.358 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.358, + 0.507, + 0.369 + ], + "angle": 0, + "content": "。" + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.369, + 0.507, + 0.377 + ], + "angle": 0, + "content": "Slow" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.327, + 0.552, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.331, + 0.612, + 0.342 + ], + "angle": 0, + "content": "Smooth" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.346, + 0.613, + 0.358 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.358, + 0.619, + 0.369 + ], + "angle": 0, + "content": "hckable" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.369, + 0.619, + 0.378 + ], + "angle": 0, + "content": "Unstable" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.327, + 0.662, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.331, + 0.696, + 0.344 + ], + "angle": 0, + "content": "High" + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.345, + 0.718, + 0.358 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.366, + 0.693, + 0.379 + ], + "angle": 0, + "content": "Low" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.327, + 0.744, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.331, + 0.778, + 0.343 + ], + "angle": 0, + "content": "Fast" + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.345, + 0.778, + 0.358 + ], + "angle": 0, + "content": "Med" + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.366, + 0.783, + 0.377 + ], + "angle": 0, + "content": "Slow" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.327, + 0.836, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.837, + 0.331, + 0.889, + 0.343 + ], + "angle": 0, + "content": "Smooth" + }, + { + "type": "image_caption", + "bbox": [ + 0.837, + 0.345, + 0.89, + 0.358 + ], + "angle": 0, + "content": "Medium" + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.366, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Unstable" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.392, + 0.302, + 0.404 + ], + "angle": 0, + "content": "Event Data Distribution" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.409, + 0.18, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.41, + 0.264, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.41, + 0.35, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.481, + 0.317, + 0.493 + ], + "angle": 0, + "content": "Event Semantic Distribution" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.498, + 0.18, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.535, + 0.155, + 0.545 + ], + "angle": 0, + "content": "Road" + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.498, + 0.265, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.213, + 0.535, + 0.237, + 0.545 + ], + "angle": 0, + "content": "Car" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.498, + 0.35, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.535, + 0.324, + 0.546 + ], + "angle": 0, + "content": "Sky" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.55, + 0.18, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.587, + 0.163, + 0.598 + ], + "angle": 0, + "content": "Building" + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.55, + 0.265, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.587, + 0.248, + 0.597 + ], + "angle": 0, + "content": "Terrain" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.55, + 0.35, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.587, + 0.343, + 0.598 + ], + "angle": 0, + "content": "Vegetation" + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.392, + 0.576, + 0.405 + ], + "angle": 0, + "content": "Event Data Distribution" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.411, + 0.452, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.411, + 0.537, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.411, + 0.624, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.481, + 0.591, + 0.493 + ], + "angle": 0, + "content": "Event Semantic Distribution" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.498, + 0.452, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.535, + 0.43, + 0.545 + ], + "angle": 0, + "content": "Road" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.498, + 0.538, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.535, + 0.511, + 0.545 + ], + "angle": 0, + "content": "Car" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.498, + 0.624, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.535, + 0.597, + 0.546 + ], + "angle": 0, + "content": "Sky" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.551, + 0.452, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.588, + 0.437, + 0.598 + ], + "angle": 0, + "content": "Building" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.551, + 0.538, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.588, + 0.521, + 0.597 + ], + "angle": 0, + "content": "Terrain" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.551, + 0.624, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.588, + 0.617, + 0.598 + ], + "angle": 0, + "content": "Vegetation" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.392, + 0.85, + 0.404 + ], + "angle": 0, + "content": "Event Data Distribution" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.411, + 0.727, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.411, + 0.811, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.41, + 0.898, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.481, + 0.865, + 0.493 + ], + "angle": 0, + "content": "Event Semantic Distribution" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.498, + 0.727, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.535, + 0.704, + 0.546 + ], + "angle": 0, + "content": "Road" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.498, + 0.812, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.535, + 0.785, + 0.545 + ], + "angle": 0, + "content": "Car" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.498, + 0.898, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.848, + 0.535, + 0.871, + 0.547 + ], + "angle": 0, + "content": "Sky" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.551, + 0.813, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.587, + 0.796, + 0.598 + ], + "angle": 0, + "content": "Terrain" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.551, + 0.898, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.587, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Vegetation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.604, + 0.907, + 0.66 + ], + "angle": 0, + "content": "Figure 1. Key discrepancies across the vehicle \\((\\mathcal{P}^{\\mathrm{v}})\\), drone \\((\\mathcal{P}^{\\mathrm{d}})\\), and quadruped \\((\\mathcal{P}^{\\mathrm{q}})\\) platforms. We compare the core attributes (viewpoint, speed, stability), event data distributions, and semantic distributions across event data acquired by three platforms, highlighting the challenges of adapting event camera perception to diverse operational contexts. These variations motivate the need for a robust cross-platform adaptation framework to harmonize event-based dense perception across distinct environmental setups and conditions." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.673, + 0.327, + 0.689 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Cross-platform adaptation in event-based dense perception is crucial for deploying event cameras across diverse settings, such as vehicles, drones, and quadrupeds, each with unique motion dynamics, viewpoints, and class distributions. In this work, we introduce EventFly, a framework for robust cross-platform adaptation in event camera perception. Our approach comprises three key components: i) Event Activation Prior (EAP), which identifies high-activation regions in the target domain to minimize prediction entropy, fostering confident, domain-adaptive predictions; ii) EventBlend, a data-mixing strategy that integrates source and target event voxel grids based on EAP-driven similarity and density maps, enhancing feature alignment;" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.826 + ], + "angle": 0, + "content": "and iii) EventMatch, a dual-discriminator technique that aligns features from source, target, and blended domains for better domain-invariant learning. To holistically assess cross-platform adaptation abilities, we introduce EXPo, a large-scale benchmark with diverse samples across vehicle, drone, and quadruped platforms. Extensive experiments validate our effectiveness, demonstrating substantial gains over popular adaptation methods. We hope this work can pave the way for more adaptive, high-performing event perception across diverse and complex environments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.843, + 0.645, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Event cameras, with their asynchronous operation and high temporal resolution, provide great advantages over conven-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "tional frame-based sensors by capturing precise pixel intensity changes as they occur [24, 26, 43, 44, 71]. These attributes make event cameras ideally suited for real-time applications in dynamic environments, such as autonomous driving, aerial navigation, and robotic perception [56, 74]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Despite this potential, event cameras remain primarily deployed on the vehicle platform [7, 9]. Based on a recent survey study [11], 12 out of 15 public event camera perception datasets are collected exclusively from ground vehicles [2, 13, 29, 64, 93, 100]. In contrast, datasets involving alternative robot platforms, such as drones [4, 16, 85] or mobile robots [5, 95, 102], are limited, making robust event-based perception across diverse domains challenging [66, 70]. To enable versatile applications, a cross-platform approach is vital for adapting the perception models to the new, real-world, and challenging operational contexts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.594 + ], + "angle": 0, + "content": "Cross-platform adaptation for event-based dense perception presents significant technical challenges. As can be seen from Fig. 1, each robot platform - vehicle, drone, or quadruped - exhibits unique motion patterns, environmental interactions, and perspective-based dynamics that create distinct activation patterns in the event data [22, 53, 65, 73]. For example, ground vehicles capture event activity near surfaces [26, 29, 100], while drones tend to capture high-altitude landscape features with sparse ground details [12, 42, 52, 56, 80]. Conventional domain adaptation methods commonly used for frame-based sensors are not well-suited to handle the distinct spatial-temporal nuances of event camera data, which demand tailored alignment strategies [48, 50, 60, 78]. A specialized framework for cross-platform adaptation in event-based dense perception is, therefore, crucial for reliable, context-sensitive perception across varied robot sensing conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.702 + ], + "angle": 0, + "content": "In this work, we introduce EventFly, a robust framework for cross-platform adaptation in event camera dense perception. We bridge domain-specific gaps in event-camera perception by leveraging platform-specific activation patterns and aligning feature distributions across domains. Our design consists of three core components, each addressing a specific challenge in cross-platform event data alignment:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.824 + ], + "angle": 0, + "content": "■ Event Activation Prior (EAP). Event cameras on different platforms reveal distinctive high-activation regions shaped by platform-specific dynamics. EAP leverages these by identifying frequently activated areas in the target domain and minimizing prediction entropy within these regions. This promotes confident, domain-adaptive predictions that align the model to platform-specific event patterns, enhancing performance in target-relevant contexts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Cross-Platform Consistency Regularization. To bridge the gaps in-between different robot platforms, we propose EventBlend to create hybrid event voxel grids by combining source and target event data in a spatially structured way. Guided by EAP-driven similarity and density maps, this" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "method selectively integrates features based on shared activation patterns, focusing on areas where platform-specific details overlap. This fusion retains platform-specific cues while enhancing cross-domain feature consistency, balancing reliable source information with context-sensitive target features for more robust cross-platform adaptations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.333 + ], + "angle": 0, + "content": "Cross-Platform Adversarial Training. To further align domain distributions, we propose EventMatch, a dual-discriminator approach across source, target, and blended representations. One discriminator enforces alignment between source and blended domains, while the other softly adapts blended features toward the target, particularly in high-activation regions. This layered approach supports robust, domain-adaptive learning that generalizes well across platforms, balancing domain-invariant features with target-specific adaptations for improved perception accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.515 + ], + "angle": 0, + "content": "To rigorously assess the cross-platform adaptation performance, we establish EXPo, a large-scale Event-based Xcross-Platform dense perception benchmark with around 90k event data from three platforms: vehicle, drone, and quadruped. From extensive benchmark experiments, our approach consistently demonstrates robust adaptation capabilities, achieving on average \\(23.8\\%\\) higher accuracy and \\(77.1\\%\\) better mIoU across platforms compared to source-only training. When evaluated against prior adaptation methods, we outperform by significant margins across almost all semantic classes, highlighting the scalability and effectiveness in diverse, challenging operational contexts." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.516, + 0.871, + 0.529 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.53, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- We propose EventFly, a novel framework designed for cross-platform adaptation in event camera perception, facilitating robustness under diverse event data dynamics. To our knowledge, this is the first work proposed to address this critical gap in event-based perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.606, + 0.905, + 0.68 + ], + "angle": 0, + "content": "- We introduce Event Activation Prior (EAP), EventBlend, and EventMatch - a set of tailored techniques that utilize platform-specific activation patterns, spatial data mixing, and dual-domain feature alignment to tackle the unique challenges of event-based cross-platform adaptation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.681, + 0.905, + 0.756 + ], + "angle": 0, + "content": "- We establish a large-scale benchmark, EXPo, for cross-platform adaptation in event-based perception, comprising rich collections of samples from vehicle, drone, and quadruped domains, providing a strong foundation for further research in adaptive event camera applications." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.53, + 0.905, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.655, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Event Camera Perception. Recent works harnessed event cameras for diverse perception tasks in dynamic settings, such as detection [25, 26, 28, 55, 105], segmentation [1, 32, 47, 75], depth estimation [15, 30, 34, 58, 99, 101], and visual odometry [10, 18, 35, 59]. Object detection with event cameras involves locating objects by leveraging the high temporal resolution of event streams, enabling" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1473" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "effective scene parsing even under challenging conditions [49, 63, 98, 103]. Meanwhile, semantic segmentation aims to enhance scene understanding, which is vital for safe navigation in autonomous systems [15, 45, 57, 83]. Additionally, recent research has focused on integrating event data with complementary modalities to improve accuracy and mitigate the sparse nature of event data [3, 14, 27, 33, 84]. Our work extends this line by focusing on robust cross-platform adaptation, ensuring that event-based perception models can generalize across different robotic platforms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.452 + ], + "angle": 0, + "content": "Cross-Domain Adaptation. Transferring knowledge from a labeled source domain to an unlabeled target domain addresses the challenge of limited annotated data in diverse real-world cases [69, 96]. Traditional methods in this area leverage strategies like domain adversarial training [79], entropy minimization [61, 68, 81, 82], contrastive learning [37, 87, 92], domain mixing [46, 51, 77, 97], and self-training [8, 38, 94, 104]. However, despite promising performances, these approaches are primarily designed for frame-based data [17, 36, 91], which lacks the unique spatiotemporal properties of the stream-based event camera data. Our approach adapts these domain adaptation principles specifically for event-based vision perception, addressing the distinct challenges posed by event camera data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Domain Adaptation from Event Camera. Recent works have begun to address domain shifts for event-based perception. Approaches like Ev-Transfer [57], ESS [75], and HPL-ESS [40] focus on transferring knowledge from RGB frames to event data, enabling effective event-based perception through frame-to-event domain adaptation. Other studies leverage event cameras to aid adaptation in low-light or night-time conditions for conventional RGB sensors [21, 39, 86, 89]. There are also works that explored efficient learning utilizing shared representations across modalities [20, 41, 88]. To the best of our knowledge, our work is the first to tackle cross-platform adaptation across multiple event camera platforms, such as vehicles, drones, and quadrapeds, focusing on the unique motion and environmental interactions inherent to each platform." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.226, + 0.71 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.482, + 0.825 + ], + "angle": 0, + "content": "This work tackles the problem of cross-platform adaptation for event-based dense perception across three distinct platforms: vehicle, drone, and quadruped, denoted as \\(\\mathcal{P}^{\\mathrm{v}}\\), \\(\\mathcal{P}^{\\mathrm{d}}\\), and \\(\\mathcal{P}^{\\mathrm{q}}\\), respectively. Each platform exhibits unique spatial and temporal characteristics in the captured event data, influenced by platform-specific perspectives, motion patterns, semantics, and environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.232, + 0.849 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Event cameras capture continuous, asynchronous streams [6], where each event \\(\\mathbf{e}_i = (e_i^x,e_i^y,e_i^t,e_i^p)\\) consists of pixel coordinates \\((e_i^x,e_i^y)\\), timestamp \\(e_i^t\\), and polarity \\(e_i^p\\in\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "\\(\\{-1, + 1\\}\\), indicating either a decrease or increase in brightness. This format allows event cameras to capture high-speed motion and scene changes in dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.319 + ], + "angle": 0, + "content": "Event Data Representation. To efficiently process event data, we convert raw events into voxel grid representations, denoted as \\(\\mathbf{V}\\). This structured format regularizes the event data, making it compatible with learning-based methods and consistent across different domains. To construct each voxel grid \\(\\mathbf{V}_i\\in \\mathbb{R}^{T\\times H\\times W}\\), where \\(T\\) is the number of temporal bins, and \\(H\\) and \\(W\\) are the spatial dimensions of the event sensor, we divide the event stream into fixed time intervals, with positive and negative events accumulated in separate bins along the temporal dimension. This polarity-sensitive encoding captures spatial and temporal dynamics of events that reflect both motion directionality and density:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.329, + 0.907, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\sum_ {\\mathbf {e} _ {j} \\in \\varepsilon_ {i}} e _ {j} ^ {p} \\delta \\left(e _ {j} ^ {x} - e ^ {x}\\right) \\delta \\left(e _ {j} ^ {y} - e ^ {y}\\right) \\max \\left\\{1 - \\left| \\hat {e} _ {j} ^ {t} - e ^ {t} \\right|, 0 \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.905, + 0.416 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the Kronecker delta function; \\(\\hat{e}_j^t = (C - 1)\\frac{e_j^t - e_0^t}{\\Delta T}\\) normalizes the event timestamps. Here, \\(\\Delta T\\) is the duration of each time window, and \\(e_0^t\\) is the initial timestamp." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.906, + 0.568 + ], + "angle": 0, + "content": "Cross-Platform Adaptation. We aim to address the challenge of adapting across three distinct domains, represented by \\(\\mathcal{D} = \\{\\mathcal{P}^{\\mathrm{v}},\\mathcal{P}^{\\mathrm{d}},\\mathcal{P}^{\\mathrm{q}}\\}\\), respectively. Each domain consists of event voxel grids \\(\\mathbf{V}\\) and associated labels \\(y\\in \\mathbb{R}^{H\\times W}\\), where each pixel label belongs to one of \\(C\\) semantic classes. In our cross-platform adaptation setting, we assume access to fully labeled data from the source platform while the target platform has only unlabeled data. The goal is to leverage the labeled source data and unlabeled target data to improve event-based perception performance on the target domain.\\(^1\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.576, + 0.724, + 0.591 + ], + "angle": 0, + "content": "3.2. Event Activation Prior" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Cross-platform adaptation in event-based perception hinges on aligning platform-specific activation patterns, which are influenced by domain-unique motions, and environments. We introduce Event Activation Prior (EAP), a regularization technique that encourages confident, low-entropy predictions in target-domain regions with high event activation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.907, + 0.855 + ], + "angle": 0, + "content": "In each platform, event data often exhibits recurrent activation patterns in specific regions, driven by platform-specific motion and scene dynamics (see Fig. 1). For instance, lower regions in a vehicle domain capture details like road surfaces and obstacles, while upper regions in a drone domain capture landscape features and environmental context. Such regions, denoted as \\(\\mathbf{S} \\subset \\{0,1,\\dots,H-1\\} \\times \\{0,1,\\dots,W-1\\}\\) and along with the conditional entropy \\(H(y_{\\mathbf{S}}|\\mathbf{V}_{\\mathbf{S}},\\mathbf{S})\\), typically have consistent semantic patterns, allowing for more confident predictions. By leveraging these high-activation areas, we aim to minimize the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Without loss of generality, our subsequent explanations focus on the adaptation from vehicle domain \\(\\mathcal{P}^{\\mathrm{v}}\\) to drone domain \\(\\mathcal{P}^{\\mathrm{d}}\\), although the proposed framework generalizes to any pair of domains in \\(\\mathcal{D}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1474" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.397, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Overview of the EventFly framework. Guided by the EAP principle (Sec. 3.2), the pair of source and target event data \\(\\mathbf{V}_i^{\\mathrm{v}}\\) and \\(\\mathbf{V}_i^{\\mathrm{d}}\\) are mixed via the EventBlend operation (Sec. 3.3), where the blending mask \\(\\mathcal{M}_i\\) is obtained by measuring the similarities between density maps \\(\\mathbf{D}_i^{\\mathrm{v}}\\) and \\(\\tilde{\\mathbf{D}}^{\\mathrm{d}}\\). The features \\(\\mathbf{F}^{\\mathrm{v}}\\), \\(\\mathbf{F}^{\\mathrm{d}}\\), and \\(\\tilde{\\mathbf{F}}\\) from the source, target, and blended domains are then used for EventMatch (Sec. 3.4). This facilitates learning an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.483, + 0.517 + ], + "angle": 0, + "content": "entropy of predictions in such regions, thereby guiding the model to produce more confident outputs aligned with the unique characteristics of the target platform domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.67 + ], + "angle": 0, + "content": "Formulating EAP. To incorporate EAP in training, we define it as a regularization that minimizes prediction uncertainty in these high-activation event areas. For each subregion S, EAP aims to minimize the conditional entropy \\( H(y_{\\mathbf{S}}|\\mathbf{V}_{\\mathbf{S}},\\mathbf{S}) \\), focusing on domain-specific high-activation regions. We enforce the expectation constraint over the model parameters \\( \\theta \\) as \\( \\mathbb{E}_{\\theta}[H(\\mathbf{V}_{\\mathbf{S}},y_{\\mathbf{S}}|\\mathbf{S})]\\leq c \\), where \\( c \\) is a small constant that encourages high-confidence predictions. This is transformed into a prior on parameters \\( \\theta \\) through the maximum entropy principles [31], yielding:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.677, + 0.483, + 0.707 + ], + "angle": 0, + "content": "\\[\nP (\\theta) \\propto \\exp (- \\lambda H (\\mathbf {V} _ {\\mathbf {S}}, y _ {\\mathbf {S}} | \\mathbf {S})) \\propto \\exp (- \\lambda H (y _ {\\mathbf {S}} | \\mathbf {V} _ {\\mathbf {S}}, \\mathbf {S})), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.483, + 0.738 + ], + "angle": 0, + "content": "where \\(\\lambda > 0\\) is the Lagrange multiplier corresponding to constant \\(c\\), which balances the effect of EAP on training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.483, + 0.815 + ], + "angle": 0, + "content": "Empirical Estimation of EAP. To apply EAP, we empirically estimate the entropy in high-activation event regions by aggregating over activations in the target platform domain, where activation patterns are highly informative. Using an empirical estimator, we approximate \\( H(y|\\mathbf{V},\\mathbf{S}) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.821, + 0.482, + 0.857 + ], + "angle": 0, + "content": "\\[\nH _ {\\text {e m p}} (y | \\mathbf {V}, \\mathbf {S}) = \\mathbb {E} _ {\\mathbf {V}, y, \\mathbf {S}} \\left[ \\hat {P} (y | \\mathbf {V}, \\mathbf {S}) \\log \\hat {P} (y | \\mathbf {V}, \\mathbf {S}) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{P}(y|\\mathbf{V},\\mathbf{S})\\) is the empirical prediction probability conditioned on the voxel grid \\(\\mathbf{V}\\) and restricted to region \\(\\mathbf{S}\\). By minimizing \\(H_{\\mathrm{emp}}(y|\\mathbf{V},\\mathbf{S})\\), we encourage confident" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.503 + ], + "angle": 0, + "content": "predictions within these regions, aligning the model's predictions with the target domain activation patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Integrating EAP into Training. To incorporate the EAP in Eq. (2) into cross-platform adaptation, we define the overall objective as a maximum-a-posteriori (MAP) estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.558, + 0.905, + 0.575 + ], + "angle": 0, + "content": "\\[\nC (\\theta) = \\mathcal {L} (\\theta) - \\lambda H _ {\\mathrm {e m p}} (y | \\mathbf {V}, \\mathbf {S}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\theta)\\) represents the supervised loss on source data. \\(H_{\\mathrm{emp}}(y|\\mathbf{V},\\mathbf{S})\\) minimizes uncertainty in the target domain by leveraging EAP over high-activation regions. By focusing on these regions, the event camera perception model tends to effectively learn to adapt its predictions to align with platform-specific activation cues in the target domain, achieving robust adaptation across varied platforms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.64, + 0.713 + ], + "angle": 0, + "content": "3.3.EventBlend" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Cross-platform adaptation in event-based perception requires a selective integration of data from both source and target domains. Here, we introduce EventBlend, a data-mixing strategy designed to construct hybrid voxel grids by blending event data from the source and target domains, guided by high-activation regions identified through EAP. As shown in Fig. 2, by constructing these blended voxel grids, we aim to enhance the generalization ability across domain shifts, enabling the model to adapt effectively to the target distribution while leveraging reliable source labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Event Density Maps. To identify regions where the source and target domains exhibit similar or divergent activation" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1475" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.169 + ], + "angle": 0, + "content": "patterns, we employ density maps, which highlight areas of frequent event activity. For each source event data with voxel grid \\(\\mathbf{V}_i^{\\mathbf{v}}\\in \\mathbb{R}^{T\\times H\\times W}\\), we calculate a corresponding source density map, \\(\\mathbf{D}_i^{\\mathbf{v}}\\in \\mathbb{R}^{H\\times W}\\), by summing activations along the temporal axis \\(T\\) for each spatial location \\((\\mu ,\\nu)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.175, + 0.483, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {i} ^ {\\mathbf {v}} (\\mu , \\nu) = \\sum_ {t = 1} ^ {T} | \\mathbf {V} _ {i} ^ {\\mathbf {v}} (t, \\mu , \\nu) |. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.252 + ], + "angle": 0, + "content": "For the target domain, we pre-compute an aggregated density map, \\(\\widetilde{\\mathbf{D}}^{\\mathrm{d}}\\in \\mathbb{R}^{H\\times W}\\), by summing activations over all target samples \\(\\mathbf{V}_j^{\\mathrm{d}}\\), where \\(j\\) indexes each target sample:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.258, + 0.483, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {D}} ^ {\\mathrm {d}} (\\mu , \\nu) = \\frac {1}{N ^ {\\mathrm {d}}} \\sum_ {j} \\sum_ {t = 1} ^ {T} \\left| \\mathbf {V} _ {j} ^ {\\mathrm {d}} (t, \\mu , \\nu) \\right|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.368 + ], + "angle": 0, + "content": "where \\( N^{\\mathrm{d}} \\) denotes the total number of samples in the target domain. These density maps highlight regions with significant activity, which aids in identifying areas where the domains may share similar activation patterns, consistent with the high-activation regions established by EAP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Similarity Map & Binary Mask. To guide selective blending, for each source sample, we construct a similarity map \\(\\mathbf{SIM}_i\\in \\mathbb{R}^{H\\times W}\\) to quantify spatial overlap in activation patterns between two domains. This map is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.435, + 0.483, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S I M} _ {i} (\\mu , \\nu) = 1 - \\left| \\mathbf {D} _ {i} ^ {\\mathbf {v}} (\\mu , \\nu) - \\tilde {\\mathbf {D}} ^ {\\mathrm {d}} (\\mu , \\nu) \\right|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Here, high values in \\(\\mathbf{SIM}_i\\) correspond to regions where source and target domains exhibit similar activation intensities, aligning with the high-activation areas specified by EAP in Eq. (2). Next, we derive a binary mask \\(\\mathcal{M}_i\\in\\) \\(\\{0,1\\}^{H\\times W}\\) by applying a threshold \\(\\tau\\) to the similarity map:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.54, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {i} (\\mu , \\nu) = \\left\\{ \\begin{array}{l l} \\mathbf {1}, & \\text {i f} \\mathbf {S I M} _ {i} (\\mu , \\nu) \\geq \\tau , \\\\ \\mathbf {0}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.722 + ], + "angle": 0, + "content": "This mask directs the blending process by specifying which regions to retain from the source and which to adapt from the target. A mask value of 1 indicates that activations in \\((\\mu, \\nu)\\) from the source domain should be retained, while a value of 0 specifies a shift to the target domain activations. It is worth mentioning that for the selections of \\(\\mathbf{D}_i^{\\mathrm{v}}\\) and \\(\\tilde{\\mathbf{D}}^{\\mathrm{d}}\\) in Eq. (5) and Eq. (6), we only consider activated pixel locations. That is to say, the similarity score does not count cases where both source and target are non-activated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.484, + 0.8 + ], + "angle": 0, + "content": "Constructing Blended Event Voxel Grids. For a pair of source and target samples \\(\\mathbf{V}_i^{\\mathrm{v}}\\) and \\(\\mathbf{V}_i^{\\mathrm{d}}\\), we create a new event voxel grid \\(\\tilde{\\mathbf{V}}_i \\in \\mathbb{R}^{T \\times H \\times W}\\), using \\(\\mathcal{M}_i\\). For each spatial location \\((\\mu, \\nu)\\), temporal sequences are selectively copied from either source and target based on \\(\\mathcal{M}_i(\\mu, \\nu)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.807, + 0.483, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {V}} _ {i} (:, \\mu , \\nu) = \\left\\{ \\begin{array}{l l} \\mathbf {V} _ {i} ^ {\\mathbf {v}} (:, \\mu , \\nu), & \\text {i f} \\mathcal {M} (\\mu , \\nu) = \\mathbf {1}, \\\\ \\mathbf {V} _ {i} ^ {\\mathrm {d}} (:, \\mu , \\nu), & \\text {i f} \\mathcal {M} (\\mu , \\nu) = \\mathbf {0}. \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.903 + ], + "angle": 0, + "content": "This selective copying approach allows \\(\\tilde{\\mathbf{V}}_i\\) to incorporate source-domain stability (i.e., human annotations) in high-similarity regions while adapting to target-specific patterns" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "in low-similarity areas. Finally, we generate the label map \\(\\tilde{y}_i\\in \\mathbb{R}^{H\\times W}\\) for \\(\\tilde{\\mathbf{V}}_i\\) by combining ground truth labels from the source and pseudo-labels from the target. Here, the pseudo-labels can be obtained offline from a source-pretrained model or generated online using a mean teacher framework [76]. During adaptation, we use \\(\\tilde{\\mathbf{V}}_i\\) and \\(\\tilde{y}_i\\) for supervised learning. The blended data contains activations and labels from both domains. We leverage EAP-aligned high-activation areas, i.e., Eq. (4), where source-target consistency is high, facilitating robust cross-domain adaptation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.253, + 0.646, + 0.267 + ], + "angle": 0, + "content": "3.4.EventMatch" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.412 + ], + "angle": 0, + "content": "In cross-platform adaptation, domain shifts manifest as discrepancies in spatial activation patterns, motion characteristics, and perspectives unique to each platform. The EAP (cf. Sec. 3.2) identifies high-activation areas for stable adaptation, while EventBlend (cf. Sec. 3.3) generates hybrid voxel grids that capture shared spatial structures across domains. However, without explicit feature alignment, the model might still struggle to fully adapt to platform-specific idiosyncrasies, especially in target-specific contexts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.906, + 0.517 + ], + "angle": 0, + "content": "To address this, we propose the EventMatch technique to align the blended domain as an intermediary, enabling robust domain-invariant feature learning by maintaining source reliability while selectively adapting to target-specific characteristics. We introduce two fully convolutional discriminators, \\(\\sigma_{1}\\) and \\(\\sigma_{2}\\), each responsible for a distinct aspect of feature alignment:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.518, + 0.906, + 0.608 + ], + "angle": 0, + "content": "- Source & Blended (via \\(\\sigma_{1}\\)): We align the blended voxel grids \\(\\tilde{\\mathbf{V}}\\) with the source domain, encouraging them to retain reliable source-domain characteristics. This alignment preserves robustness and ensures that the blended features maintain patterns learned from the ground truth supervised labels, i.e., \\(y\\), from the source domain." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.608, + 0.906, + 0.713 + ], + "angle": 0, + "content": "- Blended & Target (via \\(\\sigma_{2}\\)): Rather than the strict alignment, we regularize the blended features toward a soft alignment with the target, selectively adapting target-relevant characteristics. This alignment is focused on high-activation regions identified by EAP, allowing the blended features to bridge the domain gap in target-specific contexts without sacrificing source consistency." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.518, + 0.906, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.714, + 0.906, + 0.76 + ], + "angle": 0, + "content": "This dual-discriminator setup allows the blended features to act as an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.76, + 0.906, + 0.835 + ], + "angle": 0, + "content": "Domain Adversarial Training. Each discriminator learns to distinguish between the feature representations from its respective domains, guiding the model to produce domain-agnostic features. Given feature outputs \\(\\mathbf{F}^{\\mathrm{v}}\\), \\(\\mathbf{F}^{\\mathrm{d}}\\), and \\(\\tilde{\\mathbf{F}}\\) for the source, target, and blended domains, respectively:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.836, + 0.906, + 0.867 + ], + "angle": 0, + "content": "- Source-Blended Alignment: Discriminator \\(\\sigma_{1}\\) classifies \\(\\mathbf{F}^{\\vee}\\) as source and \\(\\tilde{\\mathbf{F}}\\) as blended, maximizing \\(\\mathcal{L}_{\\sigma_1}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.877, + 0.906, + 0.906 + ], + "angle": 0, + "content": "\\[\n- \\sum_ {\\mu , \\nu} \\left[ \\left(1 - z _ {1}\\right) \\log \\sigma_ {1} (\\tilde {\\mathbf {F}}) + z _ {1} \\log \\sigma_ {1} \\left(\\mathbf {F} ^ {\\mathrm {v}}\\right) \\right], \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1476" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.121 + ], + "angle": 0, + "content": "Table 1. Benchmark results of platform adaptation from vehicle \\((\\mathcal{P}^{\\mathrm{v}})\\) to drone \\((\\mathcal{P}^{\\mathrm{d}})\\). Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are highlighted in colors." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.123, + 0.907, + 0.294 + ], + "angle": 0, + "content": "
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o43.6933.8115.0411.8148.7111.570.928.4213.3325.488.1831.5114.880.042.41
AdaptSegNet [79]49.1435.3821.1612.1529.3723.570.170.4813.4538.2317.8548.7329.4235.550.40
CBST [104]57.9541.1824.3116.0233.0524.430.003.0818.2456.3216.8456.1523.6135.650.00
IntraDA [61]57.3738.8523.5815.9132.3123.170.004.9014.9156.7018.6754.9420.7133.080.00
DACS [77]59.8142.0127.0716.1435.1626.120.184.1118.4955.6421.7456.8134.6944.730.05
MIC [38]63.1145.6028.8717.4641.4025.190.0110.1122.8659.2520.8458.8633.9544.180.90
PLSR [94]64.6145.9329.6917.9942.0930.060.009.7523.3262.4820.6560.1531.6944.272.06
EventFly (Ours)69.1748.2032.6720.0146.6430.551.2710.9125.5067.1724.2161.0141.3044.546.21
Target •79.5752.2542.9023.3074.4839.407.100.3331.6771.9631.6467.8757.5166.1423.79
" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.313, + 0.482, + 0.357 + ], + "angle": 0, + "content": "where \\( z_{1} = 1 \\) for source and \\( z_{1} = 0 \\) for blended. This alignment ensures that the blended features preserve the supervised stability from the source." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.358, + 0.483, + 0.42 + ], + "angle": 0, + "content": "- Blended-Target Adaptation: Discriminator \\(\\sigma_{2}\\) adapts \\(\\tilde{\\mathbf{F}}\\) toward high-activation regions in the target, where target-relevant properties are critical. It classifies \\(\\tilde{\\mathbf{F}}\\) as blended and \\(\\mathbf{F}^{\\mathrm{d}}\\) as target, maximizing \\(\\mathcal{L}_{\\sigma_2}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.428, + 0.482, + 0.455 + ], + "angle": 0, + "content": "\\[\n- \\sum_ {\\mu , \\nu} \\left[ \\left(1 - z _ {2}\\right) \\log \\sigma_ {2} \\left(\\mathbf {F} ^ {\\mathrm {d}}\\right) + z _ {2} \\log \\sigma_ {2} (\\tilde {\\mathbf {F}}) \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.464, + 0.483, + 0.524 + ], + "angle": 0, + "content": "where \\( z_{2} = 1 \\) for target and \\( z_{2} = 0 \\) for blended. Here, \\( \\sigma_{2} \\) does not force complete target alignment but rather emphasizes adaptation in target-relevant areas, enabling \\( \\tilde{\\mathbf{F}} \\) to maintain a balanced representation across domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.601 + ], + "angle": 0, + "content": "Optimization. To achieve consistent cross-domain alignment, the backbone network is trained adversarially against both discriminators. The goal is to produce \\(\\tilde{\\mathbf{F}}\\) that each discriminator finds challenging to classify distinctly, thereby promoting a domain-agnostic representation:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.608, + 0.483, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d v}} = - \\sum_ {\\mu , \\nu} \\left[ \\log \\sigma_ {1} (\\tilde {\\mathbf {F}}) + \\log \\sigma_ {2} (\\tilde {\\mathbf {F}}) \\right]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.734 + ], + "angle": 0, + "content": "This objective ensures that the backbone network produces features for blended data that capture both source robustness and target adaptability, guided by high-activation areas where source-target consistency is most beneficial. To achieve this goal, we set two weight coefficients, \\(\\phi_{1}\\) and \\(\\phi_{2}\\), to balance between these two objectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.765 + ], + "angle": 0, + "content": "During training, we employ a min-max optimization framework to iteratively update the backbone, \\(\\sigma_{1}\\), and \\(\\sigma_{2}\\):" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.766, + 0.483, + 0.795 + ], + "angle": 0, + "content": "- Step 1: Update \\(\\sigma_{1}\\) and \\(\\sigma_{2}\\) to improve the classification of the source, blended, and target features." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.84 + ], + "angle": 0, + "content": "- Step 2: Update the backbone network to minimize \\(\\mathcal{L}_{\\mathrm{adv}}\\), encouraging it to generate intermediary features that support robust cross-domain generalization." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.766, + 0.483, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Through this adversarial process, EventMatch allows the model to bridge the domain gap effectively, with blended features serving as a reliable intermediary that balances source stability with target-specific adaptability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Overall Framework. As shown in Fig. 2, the source and blended domains share the same network, while the network from the target domain is updated via exponential-moving average (EMA), encouraging consistency among domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.382, + 0.724, + 0.396 + ], + "angle": 0, + "content": "3.5. The EXPo Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.57 + ], + "angle": 0, + "content": "To facilitate robust event camera perception across diverse platforms and environments, we establish a large-scale event-based cross-platform adaptation benchmark: EXPo. Our benchmark is based on [12]. We include 89, 228 frames captured from three distinct platforms - vehicle, drone, and quadruped - spanning 21 sequences: 6 from the vehicle, 7 from the drone, and 8 from the quadruped. Following DSEC-Semantic [75], we support 11 semantic classes relevant to real-world event-based perception and cover varied environments including city, urban, suburban, and rural scenes. Due to space limits, more details are in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.645, + 0.6 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.612, + 0.623 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Implementation Details. Our framework is implemented using PyTorch [62]. The backbone network is from E2VID [67], while the event segmentation head is from ESS [75]. The hyperparameters \\(\\tau\\), \\(\\phi_{1}\\) and \\(\\phi_{2}\\) are set to 0.4, 1e-3, and 2e-3, respectively. The model is optimized using AdamW [54] and OneCycle learning rate scheduler [72] for 100k iterations. The learning rate and batch size are set to 1e-3 and 8. Due to the lack of baselines, we reproduce popular frame-based adaptation methods for adversarial training, contrastive learning, and self-training [38, 61, 77, 79, 94, 104]. To ensure fair comparisons, all methods adopt the same backbone and training iterations. The models are trained using NVIDIA RTX 4090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. Following conventions, we use Accuracy (Acc) and Intersection-over-Union (IoU) as the main metrics for evaluation. We also report mean Acc (mAcc), mean IoU (mIoU), and frequency-weighted IoU (fIoU) for a holistic comparison among different adaptation methods." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1477" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Table 2. Benchmark results of platform adaptation from vehicle \\((\\mathcal{P}^{\\mathrm{v}})\\) to quadruped \\((\\mathcal{P}^{\\mathrm{q}})\\). Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are in colors." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.125, + 0.907, + 0.294 + ], + "angle": 0, + "content": "
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o66.5939.7325.1516.5263.0139.263.8817.8810.1251.679.2768.0212.350.240.99
AdaptSegNet [79]67.2548.7332.7914.8945.0045.8830.0034.9212.2255.5015.8573.8416.0731.350.00
CBST [104]69.2549.5835.0614.9547.3954.6834.2736.8313.7856.1518.1374.2316.1834.060.00
IntraDA [61]68.2948.9134.2514.8243.7555.3632.6433.3911.6055.3117.0076.0020.3031.400.00
DACS [77]69.5553.8836.5114.6643.7257.2738.4335.4214.0257.1018.4376.1624.7936.210.00
MIC [38]70.7849.2236.9315.6051.7151.7333.5438.109.4454.2720.7474.4029.7941.780.70
PLSR [94]70.9153.6537.5715.2549.0453.2837.5436.6412.9157.6025.2975.9224.9239.850.24
EventFly (Ours)73.4254.1440.0515.7850.0761.3339.1741.9712.8359.1423.5179.8027.2642.652.86
Target •80.0260.5549.8419.5874.8056.2346.0855.2821.7959.9030.3177.2458.3862.475.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.304, + 0.907, + 0.35 + ], + "angle": 0, + "content": "Table 3. Benchmark results of platform adaptation among the vehicle \\((\\mathcal{P}^{\\mathrm{v}})\\), drone \\((\\mathcal{P}^{\\mathrm{d}})\\), and quadruped \\((\\mathcal{P}^{\\mathrm{q}})\\) platforms, respectively. A total of six cross-platform adaptation settings are considered in our benchmark. Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are highlighted in colors." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.352, + 0.907, + 0.483 + ], + "angle": 0, + "content": "
MethodPv → PdPv → PqPd → PvPd → PQPq → PvPq → PdAverage
AccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoU
Source-Only o43.6915.0466.5925.1557.9120.7966.8323.0657.4921.3052.6216.8557.5220.37
AdaptSegNet [79]49.1421.1667.2532.7968.2929.5567.5733.9966.7430.6557.0720.9662.8828.18
DACS [77]59.8127.0769.5536.5171.7836.1067.7336.1171.2034.7860.7424.5066.8032.51
MIC [38]63.1128.8770.7836.9372.4636.8867.2936.2772.4635.2264.4926.1168.4333.38
PLSR [94]64.6129.6970.9137.5772.4637.1867.8336.2172.9336.3863.5727.3468.7234.06
EventFly (Ours)69.1732.6773.4240.0575.5039.9269.6837.3773.9337.7065.7828.7971.2536.08
Target •79.5742.9080.0249.8486.1255.9380.0249.8486.1255.9379.5742.9081.9049.56
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.279, + 0.515 + ], + "angle": 0, + "content": "4.2. Comparative Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.521, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Adapt from Vehicle to Drone. Tab. 1 presents the results of transferring vehicle to drone, where there is a huge domain gap in between. The source-only (15.04% mIoU) and supervised target (42.90% mIoU) models validate this gap. Prior adaptation methods show good results in closing the domain gap, while our EventFly surpasses all competitors for almost all semantic classes, demonstrating strong robustness enhancements for event-based cross-platform adaptation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Adapt from Vehicle to Quadruped. We compare the adaptation methods from vehicle to quadruped in Tab. 2. Since the data and semantic distributions of these two platforms are closer (compared to drone), the adaptation methods show larger performance gains. We achieve a promising \\(40.05\\%\\) mIoU under this setting, which corresponds to a \\(59.2\\%\\) relative improvement over the source-only baseline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Qualitative Assessments. Fig. 3 visually compares the abilities of different cross-platform adaptation methods. As can be seen, the large domain gap between source and target platforms causes huge performance degradation, resulting in messy semantic predictions (source-only). While state-of-the-art adaptation methods [38, 94] restore the performance to some extent, their predictions still suffer from ambiguous object and background contours. Differently, our EventFly generates more accurate scene semantics, credited to the superior robustness from the EAP-driven designs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Cross-Platform Adaptations. Since our EXPo benchmark consists of three platforms, we benchmark state-of-the-art methods across a total of six adaptation settings, where each platform interchangeably serve as the source/target domains. As shown in Tab. 3, we observe discrepancies in adaptation difficulties. Specifically, the drone domain is the most difficult to adapt to, mainly due to its unique motion patterns and perspective-based dynamic. Our approach exhibits better robustness than prior arts, achieving the best metrics across all six settings. Such a strong generalization ability is crucial for the deployment of event camera perception algorithms in real-world environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.667, + 0.71 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Component Analysis. Our EventFly framework compromises three key components: EventBlend, EventMatch, and the dual-backbone network. As shown in Tab. 4, each contribution provides distinct performance improvements. Since our designs are well-motivated by EAP, the combinations of any two (row \\(\\# c\\)) and all (row \\(\\# d\\)) of them further yield better cross-platform adaptation performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Domain Blending Techniques. To validate that our EAP-driven EventBlend operation indeed encourages domain regularization effect, we compare the heuristic data mixing techniques [19, 90] and recent domain mixing methods [23, 77]. As shown in Fig. 4, conventional methods, which" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1478" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Figure 3. Qualitative assessments of adaptation from vehicle to drone (the first 3 rows), and from vehicle to quadruped (the last 3 rows). We use grayscaled event images for better visibility. The RGB frames are for reference purposes only. Best viewed in colors." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.576 + ], + "angle": 0, + "content": "Table 4. Ablation results of each component in cross-platform adaptation from vehicle \\((\\mathcal{P}^{\\mathrm{v}})\\) to drone \\((\\mathcal{P}^{\\mathrm{d}})\\) and vehicle \\((\\mathcal{P}^{\\mathrm{v}})\\) to quadruped \\((\\mathcal{P}^{\\mathrm{q}})\\), respectively. Dual: The dual-branch network. Match: The EventMatch technique. Blend: The EventBlend operation. All scores are given in percentage (\\%). We compare settings from the (a) source-only, (b) single-component effects, (c) double-component effects, (d) full configuration, and (e) supervised target results, respectively, in this ablation study." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.578, + 0.484, + 0.721 + ], + "angle": 0, + "content": "
#DualMatchBlendPvAccPd mIoUPvAccPq mIoU
(a)XXX43.6915.0466.5925.15
(b)X✓X59.8625.2867.7433.66
XX✓61.8527.9169.5536.82
(c)✓✓X66.5030.2470.9038.10
✓X✓68.4331.9672.6739.11
(d)✓✓✓69.1732.6773.4240.05
(e)Target79.5742.9080.0249.84
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.809 + ], + "angle": 0, + "content": "neglect the event activation discrepancies, show sub-par performance in bridging the domains. Guided by EAP, our domain blending technique facilitates entropy minimization in high-activation regions, which in turn brings robust feature adaptation across heterogeneous platforms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Similarity Threshold. The strength of blending source and target platforms is vital in our framework. Therefore, we study the effect of the threshold parameter \\(\\tau\\). Notably, the 0 (i.e., source-only) or high (\\(>0.7\\)) value of \\(\\tau\\) brings limited gains, as the domain consistency regularization is weak. We found that the best possible trade-offs appear between 0.3" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.465, + 0.907, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.589, + 0.907, + 0.632 + ], + "angle": 0, + "content": "Figure 4. Ablation study (mIoU scores in %) on different domain-mixing techniques under the (left) \\(\\mathcal{P}^{\\mathrm{v}}\\rightarrow \\mathcal{P}^{\\mathrm{d}}\\), (middle) \\(\\mathcal{P}^{\\mathrm{v}}\\rightarrow \\mathcal{P}^{\\mathrm{q}}\\), and (right) \\(\\mathcal{P}^{\\mathrm{d}}\\rightarrow \\mathcal{P}^{\\mathrm{v}}\\) cross-platform adaptation settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.906, + 0.679 + ], + "angle": 0, + "content": "and 0.5, further validating the importance of bridging domains for cross-platform adaptation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we presented EventFly, a framework for robust cross-platform adaptation in event-based dense perception, enabling robust deployment across the vehicle, drone, and quadruped platforms. Leveraging the proposed Event Activation Prior (EAP), EventBlend, and EventMatch, we addressed the unique adaptation challenges of event stream data collected from different robot platforms. We also introduced EXPo, a large-scale benchmark designed to evaluate cross-platform event perception capabilities. Our approach demonstrated notable improvements over existing methods. We hope this work lays the foundation for more adaptive event-based perception across real-world environments." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1479" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.208 + ], + "angle": 0, + "content": "This work is under the programme DesCartes and is supported by the National Research Foundation, Prime Minister's Office, Singapore, under its Campus for Research Excellence and Technological Enterprise (CREATE) programme. This work is also supported by the Apple Scholars in AI/ML Ph.D. Fellowship program." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.209, + 0.484, + 0.255 + ], + "angle": 0, + "content": "The authors would like to sincerely thank the Program Chairs, Area Chairs, and Reviewers for the time and efforts devoted during the review process." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.273, + 0.188, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.298, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[1] Inigo Alonso and Ana C. Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1-10, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.356, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[2] Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. Ddd17: End-to-end davis driving dataset. In International Conference on Machine Learning Workshops, pages 1-9, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.414, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[3] Shristi Das Biswas, Adarsh Kosta, Chamika Liyanagedera, Marco Apolinario, and Kaushik Roy. Halsie: Hybrid approach to learning segmentation by simultaneously exploiting image and event modalities. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5952-5962, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.5, + 0.483, + 0.556 + ], + "angle": 0, + "content": "[4] Michael Blösch, Stephan Weiss, Davide Scaramuzza, and Roland Siegwart. Vision based MAV navigation in unknown and unstructured environments. In IEEE International Conference on Robotics and Automation, pages 21-28, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.558, + 0.484, + 0.627 + ], + "angle": 0, + "content": "[5] Chiara Boretti, Philippe Bich, Fabio Pareschi, Luciano Prono, Riccardo Rovatti, and Gianluca Setti. Pedro: An event-based dataset for person detection in robotics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4065-4070, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.63, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db 3 μs latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.688, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[7] Tim Brodermann, David Bruggemann, Christos Sakaridis, Kevin Ta, Odysseas Liagouris, Jason Corkill, and Luc Van Gool. Muses: The multi-sensor semantic perception dataset for driving under uncertainty. arXiv preprint arXiv:2401.12761, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.76, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[8] David Brüggemann, Christos Sakaridis, Prune Truong, and Luc Van Gool. Refign: Align and refine for adaptation of semantic segmentation to adverse conditions. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3174-3184, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Spencer Carmichael, Austin Buchan, Mani Ramanagopal, Radhika Ravi, Ram Vasudevan, and Katherine A. Skinner. Dataset and benchmark: Novel sensors for autonomous vehicle perception. International Journal of Robotics Research, 2024." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.298, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "[10] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In IEEE International Conference on Robotics and Automation, pages 703-710, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.137, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[11] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-ilidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. In European Conference on Computer Vision Workshops, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.193, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[12] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multirobot, multi-sensor, multi-environment event dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4016-4023, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[13] Peiyu Chen, Weipeng Guan, Feng Huang, Yihan Zhong, Weisong Wen, Li-Ta Hsu, and Peng Lu. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 9(1):407-416, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.334, + 0.906, + 0.404 + ], + "angle": 0, + "content": "[14] Zhiwen Chen, Zhiyu Zhu, Yifan Zhang, Junhui Hou, Guangming Shi, and Jinjian Wu. Segment any event streams via weighted adaptation of pivotal tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890-3900, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.406, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[15] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.462, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[16] Fernando Cladera, Kenneth Chaney, M. Ani Hsieh, Camillo J. Taylor, and Vijay Kumar. Evmapper: High altitude orthomapping with event cameras. arXiv preprint arXiv:2409.18120, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.519, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[17] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[18] Javier Cuadrado, Ulysse Rançon, Benoit R. Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.661, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[19] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.704, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[20] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 3942-3951, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.775, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[21] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Hue dataset: High-resolution event and frame sequences for low-light vision. In European Conference on Computer Vision Workshops, pages 1-18, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.832, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[22] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a \\(1280 \\times 720\\) back-illuminated stacked temporal contrast event-based vision sensor with \\(4.86\\mu \\mathrm{m}\\)" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1480" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In IEEE International Solid-State Circuits Conference, pages 112-114, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[23] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In British Machine Vision Conference, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.205, + 0.483, + 0.286 + ], + "angle": 0, + "content": "[24] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 (1):154-180, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[25] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[26] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014): 1034-1040, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[27] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[28] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[29] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[30] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In IEEE International Conference on 3D Vision, pages 197-206, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[31] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems, pages 529-536, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[32] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22867-22876, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[33] Dalia Hareb and Jean Martinet. Evsegsn: Neuromorphic semantic segmentation for event data. arXiv preprint arXiv:2406.14178, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[34] Javier Hidalgo-Carrio, Daniel Gehrig, and Davide Scaramuzza. Learning monocular dense depth from events. In IEEE International Conference on 3D Vision, pages 534-542, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[35] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781-5790, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "domain-adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[37] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision, pages 372–391, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[38] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[39] Yuhwan Jeong, Hoonhee Cho, and Kuk-Jin Yoon. Towards robust event-based networks for nighttime via unpaired day-to-night event translation. In European Conference on Computer Vision, pages 286-306, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.317, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[40] Linglin Jing, Yiming Ding, Yunpeng Gao, Zhigang Wang, Xu Yan, Dong Wang, Gerald Schaefer, Hui Fang, Bin Zhao, and Xuelong Li. Hpl-ess: Hybrid pseudo-labeling for unsupervised event-based semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23128–23137, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.401, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[41] Uday Kamal and Saibal Mukhopadhyay. Efficient learning of event-based dense representation using hierarchical memories with adaptive update. In European Conference on Computer Vision, pages 74-89, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.457, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[42] Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 620(7976):982-987, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.512, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[43] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In IEEE/CVF International Conference on Computer Vision, pages 2146-2156, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[44] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2496-24976, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[45] Simon Klenk, David Bonello, Lukas Koestler, Nikita Araslanov, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2378-2388, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.707, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[46] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.776, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[47] Lingdong Kong, Youquan Liu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Openness: Event-based semantic scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15686-15698, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[48] Wouter M. Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):766-785, 2021." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1481" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[49] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):14020-14037, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[50] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5743-5762, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[51] Yahao Liu, Jinhong Deng, Xinchen Gao, Wen Li, and Lixin Duan. Bapa-net: Boundary adaptation and prototype alignment for cross-domain semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 8801-8811, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.277, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[52] Antonio Loquercio, Ana I. Maqueda, Carlos R. Del-Blanco, and Davide Scaramuzza. Dronet: Learning to fly by driving. IEEE Robotics and Automation Letters, 3(2):1088-1095, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[53] Antonio Loquercio, Elia Kaufmann, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Learning high-speed flight in the wild. Science Robotics, 6(59):5810, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[54] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[55] Dongyue Lu, Lingdong Kong, Gim Hee Lee, Camille Simon Chane, and Wei Tsang Ooi. Flexevent: Event camera object detection at arbitrary frequencies. arXiv preprint arXiv:2412.06708, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[56] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso García, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5419-5427, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[57] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. IEEE Robotics and Automation Letters, 7(2):3515-3522, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[58] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[59] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34 (6):1425-1440, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[60] Poojan Oza, Vishwanath A. Sindagi, Vibashan VS, and Vishal M. Patel. Unsupervised domain adaptation of object detectors: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4018-4040, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[61] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison," + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8026-8037, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.177, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[63] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.246, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[64] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems, pages 16639-16652, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.316, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[65] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1): 259-275, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.385, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[66] Ulysse Rançon, Javier Cuadrado-Anibarro, Benoit R. Cottereau, and Timothée Masquelier. Stereospike: Depth learning with a spiking neural network. IEEE Access, 10: 127428-127439, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.442, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[67] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):1964-1980, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.498, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[68] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8197-8207, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[69] Manuel Schwonberg, Joshua Niemeijer, Jan-Aike Termöhlen, Nico M. Schmidt, Hanno Gottschalk, and Tim Fingscheidt. Survey on unsupervised domain adaptation for semantic segmentation for visual perception in automated driving. IEEE Access, 11:54296-54336, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[70] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 12:51275-51306, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[71] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.777, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[72] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. arXiv preprint arXiv:1708.07120, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[73] Bongki Son, Yunjae Suh, Sungho Kim, Heejae Jung, Jun-Seok Kim, Changwoo Shin, Keunju Park, Kyoobin Lee, Jinman Park, Jooyeon Woo, et al. 4.1 a \\(640 \\times 480\\) dynamic vision sensor with a \\(9\\mu \\mathrm{m}\\) pixel and 300meps address-event representation. In IEEE International Solid-State Circuits Conference, pages 66-67, 2017." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1482" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[74] Lea Steffen, Daniel Reichard, Jakob Weinland, Jacques Kaiser, Arne Roennau, and Rüdiger Dillmann. Neuromorphic stereo vision: A survey of bio-inspired sensors and algorithms. Frontiers in Neuroscience, 13:28, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[75] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In European Conference on Computer Vision, pages 341-357, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.208, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[76] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 195–1204, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[77] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[78] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshops, pages 350-357, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.407, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[79] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7472-7481, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.478, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[80] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate slam? combining events, images, and imu for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3 (2):994-1001, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[81] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517-2526, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[82] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[83] Lin Wang, Yujeong Chae, and Kuk-Jin Yoon. Dual transfer learning for event-based end-task prediction via plug-gable event to image translation. In IEEE/CVF International Conference on Computer Vision, pages 2135-2145, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[84] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 608-619, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[85] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 19248-19257, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[86] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyuan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 21572-21581, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[87] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9004-9021, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.261, + 0.904, + 0.301 + ], + "angle": 0, + "content": "[88] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In IEEE/CVF International Conference on Computer Vision, pages 10699-10709, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.303, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[89] Zhen Yao and Mooi Choo Chuah. Event-guided low-light video semantic segmentation. arXiv preprint arXiv:2411.00639, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.345, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[90] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.401, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[91] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12):14679-14694, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.471, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[92] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12414-12424, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.54, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[93] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, and Sebastian Scherer. Subt-mrs dataset: Pushing slam towards all-weather environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.665, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[94] Xingchen Zhao, Niluthpol Chowdhury Mithun, Abhinav Rajvanshi, Han-Pang Chiu, and Supun Samarasekera. Unsupervised domain adaptation for semantic segmentation with pseudo label self-refinement. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2399-2409, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[95] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[96] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[97] Qianyu Zhou, Zhengyang Feng, Qiqi Gu, Jiangmiao Pang, Guangliang Cheng, Xuequan Lu, Jianping Shi, and" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1483" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Lizhuang Ma. Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):804-817, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[98] Zhuyun Zhou, Zongwei Wu, Rémi Boutteau, Fan Yang, Cédric Demonceaux, and Dominique Gingac. Rgb-event fusion for moving object detection in autonomous driving. In IEEE International Conference on Robotics and Automation, pages 7808-7815, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[99] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Robotics: Science and Systems, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.097, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[100] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[101] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[102] Shifan Zhu, Zixun Xiong, and Donghyun Kim. Cear: Comprehensive event camera dataset for rapid perception of agile quadruped robots. IEEE Robotics and Automation Letters, 9(10):8999-9006, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.446, + 0.481, + 0.487 + ], + "angle": 0, + "content": "[103] Rong Zou, Manasi Muglikar, Nico Messikommer, and Davide Scaramuzza. Seeing behind dynamic occlusions with event cameras. arXiv preprint arXiv:2307.15829, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.489, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[104] Yang Zou, Zhiding Yu, B. V. K. Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In European Conference on Computer Vision, pages 289-305, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[105] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, pages 12846-128567, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.482, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1484" + } + ] +] \ No newline at end of file diff --git a/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_origin.pdf b/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0b2e893677cdc80cec71e1315a57b1eac9e80538 --- /dev/null +++ b/2025/EventFly_ Event Camera Perception from Ground to the Sky/ecab4823-da56-483f-8321-47b8a1d3b698_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62842397b524836694d479b6e04a15bad70e695dc30ca19777fecc6ce377ee44 +size 3206752 diff --git a/2025/EventFly_ Event Camera Perception from Ground to the Sky/full.md b/2025/EventFly_ Event Camera Perception from Ground to the Sky/full.md new file mode 100644 index 0000000000000000000000000000000000000000..92d7f5bb53ca18b584f19f2254c8f08196e90986 --- /dev/null +++ b/2025/EventFly_ Event Camera Perception from Ground to the Sky/full.md @@ -0,0 +1,553 @@ +# EventFly: Event Camera Perception from Ground to the Sky + +Lingdong Kong $^{1,2}$ , Dongyue Lu $^{1}$ , Xiang Xu $^{3}$ , Lai Xing Ng $^{4}$ , Wei Tsang Ooi $^{1,5}$ , Benoit R. Cottereau $^{5,6}$ + +$^{1}$ National University of Singapore $^{2}$ CNRS@CREATE $^{3}$ Nanjing University of Aeronautics and Astronautics + +$^{4}$ Institute for Infocomm Research, A*STAR $^{5}$ IPAL, CNRS IRL 2955, Singapore $^{6}$ CerCo, CNRS UMR 5549, Université Toulouse III + +Project Page: https://event-fly.github.io + +![](images/5853b6c60761c6587a74967048b16a75ad001ca58c3dcb39334bd3326cf8a81a.jpg) + +# Vehicle + +![](images/272e59af0f8980ad3f170e9f0a09e528b01ce5f78d6b110e47c230fc12efdc45.jpg) +Viewpoint + +![](images/0380827b2d44cf256cc5e2a08937d8621f0f10c63d16060b236ccd30a9f9ddac.jpg) +Speed + +![](images/38bf90130853b4f6eafa14489a686655d4c0a7b784d764d6de96368ce090b95a.jpg) +Stability + +![](images/fae723af8c1007b1cf8cf9fef9304bf2c34f57bcd00faafe627a5b95d34e627e.jpg) + +# Drone + +![](images/7b6e06c1788b26e6da125f6e1ed9bd79cc723f8d5a9cffedff2f3ce0862a1792.jpg) +Viewpoint + +![](images/fc0b719d722a0d34af55598a5fa6feb11992a3adf105d490b193544aa6f87359.jpg) +Speed + +![](images/12714ec2df0dabace022464e5d44748c26fcc3df4fa96bd3fece33094c9b4b57.jpg) +Stability + +![](images/93e5fc4ac5e48669deaf86565705b9cbc570be48789efe743c270687af8c09b9.jpg) + +# Quadruped + +![](images/17546392b375726ada3df0d26c7da180715b3478ab60db8ce698ae92c7e68f27.jpg) +Viewpoint + +![](images/9c47e2fc092d85e7050aee94b8af09d6c4e73894daf51e3d493d78f6fa26ee55.jpg) +Speed + +![](images/2e63cb70da619ea54e6a9a2f69cea3994b725304e8303b7be77a65583bdc7744.jpg) +Stability + +![](images/547e8d07899970ee724622fa5773af2ed3589fe6c77876aa3e5232900cfb2e80.jpg) +High +Medium +Medium +Low + +![](images/722497fff7d34657c0c3588afe73e091ac960e0df45cb05cadd6dd1039a0d9b9.jpg) +Fast +Medium +G +Slow + +![](images/2dc7517d49fb0dc1d09ee9ec1a3377475b1a80feea9223a51633793297b04a89.jpg) +Smooth +Medium +Medium +Unstable + +![](images/7430a70af36122e6c18667f04af799155d96ade4050cf6c1feffe5b2dedd8492.jpg) +High +Medium +1 +Low + +![](images/e662280e6c6a9d42bda8875c4a36491c55f5bdbc4228c488d4d0d2d4bfbb8585.jpg) +Fast +Medium +。 +Slow + +![](images/090f5d8154999e61ba71d5469c3ce4d016f97c5af25f637f0ef5eb3420b3c34f.jpg) +Smooth +Medium +hckable +Unstable + +![](images/589250169122ebb40635a3b5bcdaf7866afcc060f323b3ee93d07df75f51a282.jpg) +High +Medium +Low + +![](images/1f834fdae78fc0aa751cdfd0b48310b47b7b4f47c74f22491914de0600ae7d16.jpg) +Fast +Med +Slow + +![](images/4c683c3f4d0f92985f10ab440e3c60db10c280370b861fb0a2fffc9508f8f7d1.jpg) +Smooth +Medium +Unstable + +![](images/0e96fe90ac2ca6c14ea19373eae78d99b907c76fb4a1d3757f17630dbfa480a0.jpg) +Event Data Distribution + +![](images/f29ea8a809278b472a63f846333db188c71f1b1edd09e28d06ede1d8569077fb.jpg) + +![](images/48a278e813bd6ad73baaa9338462d23ce56dfb2a224f3406650e42f7910df55a.jpg) + +![](images/1cadc832c89bfb512c3223c2a34c5df55f2b854686a9481059492584c0d8fc04.jpg) +Event Semantic Distribution +Road + +![](images/f6c2e20de6a9d57ab420640ed802932506d6bbb8ddf59260b783564626f1ceb7.jpg) +Car + +![](images/62b1ac51548c559d026477cff98ed06cc400e996f40ebada43d6780cf08bd661.jpg) +Sky + +![](images/311e1fa70f9d60671194957c69ed89db0762192e158333567d8c2ede21cefd09.jpg) +Building +Figure 1. Key discrepancies across the vehicle $(\mathcal{P}^{\mathrm{v}})$ , drone $(\mathcal{P}^{\mathrm{d}})$ , and quadruped $(\mathcal{P}^{\mathrm{q}})$ platforms. We compare the core attributes (viewpoint, speed, stability), event data distributions, and semantic distributions across event data acquired by three platforms, highlighting the challenges of adapting event camera perception to diverse operational contexts. These variations motivate the need for a robust cross-platform adaptation framework to harmonize event-based dense perception across distinct environmental setups and conditions. + +![](images/6cb210a35f1456ccf925008373ce6704efc63026634adf41197d81f8d1ae31ab.jpg) +Terrain + +![](images/493289b3f99d0d0a4c9f16a4f22b068563b08cf7ec593c840b21dc9d9fc73f7f.jpg) +Vegetation + +![](images/bb89c73a093a17b69cc9e0b11e62f36afba2bc92f76038d905e9dcadf78f6428.jpg) +Event Data Distribution + +![](images/0a06390f3d03a721f4c74b49185dffb5fe74c489558deadf7aacab7a52425152.jpg) + +![](images/bbf364fa6b5bfeba3b5aad3495ff5cb4c88761a1bdc4284d957e7eafeddf3915.jpg) + +![](images/9135fa7238a4726e06951fd84212a590fa40cf29ce9d01518912f78d92916ab8.jpg) +Event Semantic Distribution +Road + +![](images/2b44515befee8a13e04a62f4467188f29a82e1439eabf2c8ac95419a830fd449.jpg) +Car + +![](images/edca595c31516659e530990cc0418a03f96d8633d5cec85cc348d916a9ab8caf.jpg) +Sky + +![](images/db50f4a6193090f94f116710b8425e58af58a833fe459ac3533b57770578957c.jpg) +Building + +![](images/b0261deb16bbf0a1b8ee8af8f5fd5c8c60415dac3e4d7bf649fbe3cc084185f7.jpg) +Terrain + +![](images/a70b7b37848f10af007f67bee10b431fc7d5030f810c66f51cfa9ac12d25dd19.jpg) +Vegetation + +![](images/a52df0b455925d426b87f547e8ba74c4fa88c17b7c0975acafb160e8cae3422c.jpg) + +![](images/ca5ab5c620c9479febd96acf155b9de61db38187a92d37697d33550cc215b4e2.jpg) + +![](images/b5915ff6a6a67ba116d7fbf2da64c3b83e811ab4b374950d5c9a07f810323f22.jpg) +Event Data Distribution + +![](images/ab513ef08af6e1b4c0c87a701de10c42a84ca4736b79ec3ebe737cca0bb45380.jpg) +Event Semantic Distribution +Road + +![](images/0a9dc4af66216b52316b7fc5b46a05e9155428b5bb702b6f6c60cc56dc981d9e.jpg) +Car + +![](images/b96fd6e9c260adaad00800abbea46a130b08e63c24901c7f6de4802f276495d5.jpg) +Sky + +![](images/0072c8948b2e6cb856f182f469503f64651ab3beea9922e71bb19d7e66599b4f.jpg) +Terrain + +![](images/3532366df72d2686d0bdd7ff51635c8800d64c8ffb8f942901d368198dba5eac.jpg) +Vegetation + +# Abstract + +Cross-platform adaptation in event-based dense perception is crucial for deploying event cameras across diverse settings, such as vehicles, drones, and quadrupeds, each with unique motion dynamics, viewpoints, and class distributions. In this work, we introduce EventFly, a framework for robust cross-platform adaptation in event camera perception. Our approach comprises three key components: i) Event Activation Prior (EAP), which identifies high-activation regions in the target domain to minimize prediction entropy, fostering confident, domain-adaptive predictions; ii) EventBlend, a data-mixing strategy that integrates source and target event voxel grids based on EAP-driven similarity and density maps, enhancing feature alignment; + +and iii) EventMatch, a dual-discriminator technique that aligns features from source, target, and blended domains for better domain-invariant learning. To holistically assess cross-platform adaptation abilities, we introduce EXPo, a large-scale benchmark with diverse samples across vehicle, drone, and quadruped platforms. Extensive experiments validate our effectiveness, demonstrating substantial gains over popular adaptation methods. We hope this work can pave the way for more adaptive, high-performing event perception across diverse and complex environments. + +# 1. Introduction + +Event cameras, with their asynchronous operation and high temporal resolution, provide great advantages over conven- + +tional frame-based sensors by capturing precise pixel intensity changes as they occur [24, 26, 43, 44, 71]. These attributes make event cameras ideally suited for real-time applications in dynamic environments, such as autonomous driving, aerial navigation, and robotic perception [56, 74]. + +Despite this potential, event cameras remain primarily deployed on the vehicle platform [7, 9]. Based on a recent survey study [11], 12 out of 15 public event camera perception datasets are collected exclusively from ground vehicles [2, 13, 29, 64, 93, 100]. In contrast, datasets involving alternative robot platforms, such as drones [4, 16, 85] or mobile robots [5, 95, 102], are limited, making robust event-based perception across diverse domains challenging [66, 70]. To enable versatile applications, a cross-platform approach is vital for adapting the perception models to the new, real-world, and challenging operational contexts. + +Cross-platform adaptation for event-based dense perception presents significant technical challenges. As can be seen from Fig. 1, each robot platform - vehicle, drone, or quadruped - exhibits unique motion patterns, environmental interactions, and perspective-based dynamics that create distinct activation patterns in the event data [22, 53, 65, 73]. For example, ground vehicles capture event activity near surfaces [26, 29, 100], while drones tend to capture high-altitude landscape features with sparse ground details [12, 42, 52, 56, 80]. Conventional domain adaptation methods commonly used for frame-based sensors are not well-suited to handle the distinct spatial-temporal nuances of event camera data, which demand tailored alignment strategies [48, 50, 60, 78]. A specialized framework for cross-platform adaptation in event-based dense perception is, therefore, crucial for reliable, context-sensitive perception across varied robot sensing conditions. + +In this work, we introduce EventFly, a robust framework for cross-platform adaptation in event camera dense perception. We bridge domain-specific gaps in event-camera perception by leveraging platform-specific activation patterns and aligning feature distributions across domains. Our design consists of three core components, each addressing a specific challenge in cross-platform event data alignment: + +■ Event Activation Prior (EAP). Event cameras on different platforms reveal distinctive high-activation regions shaped by platform-specific dynamics. EAP leverages these by identifying frequently activated areas in the target domain and minimizing prediction entropy within these regions. This promotes confident, domain-adaptive predictions that align the model to platform-specific event patterns, enhancing performance in target-relevant contexts. + +Cross-Platform Consistency Regularization. To bridge the gaps in-between different robot platforms, we propose EventBlend to create hybrid event voxel grids by combining source and target event data in a spatially structured way. Guided by EAP-driven similarity and density maps, this + +method selectively integrates features based on shared activation patterns, focusing on areas where platform-specific details overlap. This fusion retains platform-specific cues while enhancing cross-domain feature consistency, balancing reliable source information with context-sensitive target features for more robust cross-platform adaptations. + +Cross-Platform Adversarial Training. To further align domain distributions, we propose EventMatch, a dual-discriminator approach across source, target, and blended representations. One discriminator enforces alignment between source and blended domains, while the other softly adapts blended features toward the target, particularly in high-activation regions. This layered approach supports robust, domain-adaptive learning that generalizes well across platforms, balancing domain-invariant features with target-specific adaptations for improved perception accuracy. + +To rigorously assess the cross-platform adaptation performance, we establish EXPo, a large-scale Event-based Xcross-Platform dense perception benchmark with around 90k event data from three platforms: vehicle, drone, and quadruped. From extensive benchmark experiments, our approach consistently demonstrates robust adaptation capabilities, achieving on average $23.8\%$ higher accuracy and $77.1\%$ better mIoU across platforms compared to source-only training. When evaluated against prior adaptation methods, we outperform by significant margins across almost all semantic classes, highlighting the scalability and effectiveness in diverse, challenging operational contexts. + +In summary, our main contributions are as follows: + +- We propose EventFly, a novel framework designed for cross-platform adaptation in event camera perception, facilitating robustness under diverse event data dynamics. To our knowledge, this is the first work proposed to address this critical gap in event-based perception tasks. +- We introduce Event Activation Prior (EAP), EventBlend, and EventMatch - a set of tailored techniques that utilize platform-specific activation patterns, spatial data mixing, and dual-domain feature alignment to tackle the unique challenges of event-based cross-platform adaptation. +- We establish a large-scale benchmark, EXPo, for cross-platform adaptation in event-based perception, comprising rich collections of samples from vehicle, drone, and quadruped domains, providing a strong foundation for further research in adaptive event camera applications. + +# 2. Related Work + +Event Camera Perception. Recent works harnessed event cameras for diverse perception tasks in dynamic settings, such as detection [25, 26, 28, 55, 105], segmentation [1, 32, 47, 75], depth estimation [15, 30, 34, 58, 99, 101], and visual odometry [10, 18, 35, 59]. Object detection with event cameras involves locating objects by leveraging the high temporal resolution of event streams, enabling + +effective scene parsing even under challenging conditions [49, 63, 98, 103]. Meanwhile, semantic segmentation aims to enhance scene understanding, which is vital for safe navigation in autonomous systems [15, 45, 57, 83]. Additionally, recent research has focused on integrating event data with complementary modalities to improve accuracy and mitigate the sparse nature of event data [3, 14, 27, 33, 84]. Our work extends this line by focusing on robust cross-platform adaptation, ensuring that event-based perception models can generalize across different robotic platforms. + +Cross-Domain Adaptation. Transferring knowledge from a labeled source domain to an unlabeled target domain addresses the challenge of limited annotated data in diverse real-world cases [69, 96]. Traditional methods in this area leverage strategies like domain adversarial training [79], entropy minimization [61, 68, 81, 82], contrastive learning [37, 87, 92], domain mixing [46, 51, 77, 97], and self-training [8, 38, 94, 104]. However, despite promising performances, these approaches are primarily designed for frame-based data [17, 36, 91], which lacks the unique spatiotemporal properties of the stream-based event camera data. Our approach adapts these domain adaptation principles specifically for event-based vision perception, addressing the distinct challenges posed by event camera data. + +Domain Adaptation from Event Camera. Recent works have begun to address domain shifts for event-based perception. Approaches like Ev-Transfer [57], ESS [75], and HPL-ESS [40] focus on transferring knowledge from RGB frames to event data, enabling effective event-based perception through frame-to-event domain adaptation. Other studies leverage event cameras to aid adaptation in low-light or night-time conditions for conventional RGB sensors [21, 39, 86, 89]. There are also works that explored efficient learning utilizing shared representations across modalities [20, 41, 88]. To the best of our knowledge, our work is the first to tackle cross-platform adaptation across multiple event camera platforms, such as vehicles, drones, and quadrapeds, focusing on the unique motion and environmental interactions inherent to each platform. + +# 3. Methodology + +This work tackles the problem of cross-platform adaptation for event-based dense perception across three distinct platforms: vehicle, drone, and quadruped, denoted as $\mathcal{P}^{\mathrm{v}}$ , $\mathcal{P}^{\mathrm{d}}$ , and $\mathcal{P}^{\mathrm{q}}$ , respectively. Each platform exhibits unique spatial and temporal characteristics in the captured event data, influenced by platform-specific perspectives, motion patterns, semantics, and environments. + +# 3.1. Preliminaries + +Event cameras capture continuous, asynchronous streams [6], where each event $\mathbf{e}_i = (e_i^x,e_i^y,e_i^t,e_i^p)$ consists of pixel coordinates $(e_i^x,e_i^y)$ , timestamp $e_i^t$ , and polarity $e_i^p\in$ + +$\{-1, + 1\}$ , indicating either a decrease or increase in brightness. This format allows event cameras to capture high-speed motion and scene changes in dynamic environments. + +Event Data Representation. To efficiently process event data, we convert raw events into voxel grid representations, denoted as $\mathbf{V}$ . This structured format regularizes the event data, making it compatible with learning-based methods and consistent across different domains. To construct each voxel grid $\mathbf{V}_i\in \mathbb{R}^{T\times H\times W}$ , where $T$ is the number of temporal bins, and $H$ and $W$ are the spatial dimensions of the event sensor, we divide the event stream into fixed time intervals, with positive and negative events accumulated in separate bins along the temporal dimension. This polarity-sensitive encoding captures spatial and temporal dynamics of events that reflect both motion directionality and density: + +$$ +\sum_ {\mathbf {e} _ {j} \in \varepsilon_ {i}} e _ {j} ^ {p} \delta \left(e _ {j} ^ {x} - e ^ {x}\right) \delta \left(e _ {j} ^ {y} - e ^ {y}\right) \max \left\{1 - \left| \hat {e} _ {j} ^ {t} - e ^ {t} \right|, 0 \right\}, \tag {1} +$$ + +where $\delta$ is the Kronecker delta function; $\hat{e}_j^t = (C - 1)\frac{e_j^t - e_0^t}{\Delta T}$ normalizes the event timestamps. Here, $\Delta T$ is the duration of each time window, and $e_0^t$ is the initial timestamp. + +Cross-Platform Adaptation. We aim to address the challenge of adapting across three distinct domains, represented by $\mathcal{D} = \{\mathcal{P}^{\mathrm{v}},\mathcal{P}^{\mathrm{d}},\mathcal{P}^{\mathrm{q}}\}$ , respectively. Each domain consists of event voxel grids $\mathbf{V}$ and associated labels $y\in \mathbb{R}^{H\times W}$ , where each pixel label belongs to one of $C$ semantic classes. In our cross-platform adaptation setting, we assume access to fully labeled data from the source platform while the target platform has only unlabeled data. The goal is to leverage the labeled source data and unlabeled target data to improve event-based perception performance on the target domain. $^1$ + +# 3.2. Event Activation Prior + +Cross-platform adaptation in event-based perception hinges on aligning platform-specific activation patterns, which are influenced by domain-unique motions, and environments. We introduce Event Activation Prior (EAP), a regularization technique that encourages confident, low-entropy predictions in target-domain regions with high event activation. + +In each platform, event data often exhibits recurrent activation patterns in specific regions, driven by platform-specific motion and scene dynamics (see Fig. 1). For instance, lower regions in a vehicle domain capture details like road surfaces and obstacles, while upper regions in a drone domain capture landscape features and environmental context. Such regions, denoted as $\mathbf{S} \subset \{0,1,\dots,H-1\} \times \{0,1,\dots,W-1\}$ and along with the conditional entropy $H(y_{\mathbf{S}}|\mathbf{V}_{\mathbf{S}},\mathbf{S})$ , typically have consistent semantic patterns, allowing for more confident predictions. By leveraging these high-activation areas, we aim to minimize the + +![](images/2a27ad1640cc13ebb94d1818ddcae1cdd246e138d3ba21c7319041f6b6e8f1ed.jpg) +Figure 2. Overview of the EventFly framework. Guided by the EAP principle (Sec. 3.2), the pair of source and target event data $\mathbf{V}_i^{\mathrm{v}}$ and $\mathbf{V}_i^{\mathrm{d}}$ are mixed via the EventBlend operation (Sec. 3.3), where the blending mask $\mathcal{M}_i$ is obtained by measuring the similarities between density maps $\mathbf{D}_i^{\mathrm{v}}$ and $\tilde{\mathbf{D}}^{\mathrm{d}}$ . The features $\mathbf{F}^{\mathrm{v}}$ , $\mathbf{F}^{\mathrm{d}}$ , and $\tilde{\mathbf{F}}$ from the source, target, and blended domains are then used for EventMatch (Sec. 3.4). This facilitates learning an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive). + +entropy of predictions in such regions, thereby guiding the model to produce more confident outputs aligned with the unique characteristics of the target platform domain. + +Formulating EAP. To incorporate EAP in training, we define it as a regularization that minimizes prediction uncertainty in these high-activation event areas. For each subregion S, EAP aims to minimize the conditional entropy $H(y_{\mathbf{S}}|\mathbf{V}_{\mathbf{S}},\mathbf{S})$ , focusing on domain-specific high-activation regions. We enforce the expectation constraint over the model parameters $\theta$ as $\mathbb{E}_{\theta}[H(\mathbf{V}_{\mathbf{S}},y_{\mathbf{S}}|\mathbf{S})]\leq c$ , where $c$ is a small constant that encourages high-confidence predictions. This is transformed into a prior on parameters $\theta$ through the maximum entropy principles [31], yielding: + +$$ +P (\theta) \propto \exp (- \lambda H (\mathbf {V} _ {\mathbf {S}}, y _ {\mathbf {S}} | \mathbf {S})) \propto \exp (- \lambda H (y _ {\mathbf {S}} | \mathbf {V} _ {\mathbf {S}}, \mathbf {S})), \tag {2} +$$ + +where $\lambda > 0$ is the Lagrange multiplier corresponding to constant $c$ , which balances the effect of EAP on training. + +Empirical Estimation of EAP. To apply EAP, we empirically estimate the entropy in high-activation event regions by aggregating over activations in the target platform domain, where activation patterns are highly informative. Using an empirical estimator, we approximate $H(y|\mathbf{V},\mathbf{S})$ as: + +$$ +H _ {\text {e m p}} (y | \mathbf {V}, \mathbf {S}) = \mathbb {E} _ {\mathbf {V}, y, \mathbf {S}} \left[ \hat {P} (y | \mathbf {V}, \mathbf {S}) \log \hat {P} (y | \mathbf {V}, \mathbf {S}) \right], \tag {3} +$$ + +where $\hat{P}(y|\mathbf{V},\mathbf{S})$ is the empirical prediction probability conditioned on the voxel grid $\mathbf{V}$ and restricted to region $\mathbf{S}$ . By minimizing $H_{\mathrm{emp}}(y|\mathbf{V},\mathbf{S})$ , we encourage confident + +predictions within these regions, aligning the model's predictions with the target domain activation patterns. + +Integrating EAP into Training. To incorporate the EAP in Eq. (2) into cross-platform adaptation, we define the overall objective as a maximum-a-posteriori (MAP) estimation: + +$$ +C (\theta) = \mathcal {L} (\theta) - \lambda H _ {\mathrm {e m p}} (y | \mathbf {V}, \mathbf {S}), \tag {4} +$$ + +where $\mathcal{L}(\theta)$ represents the supervised loss on source data. $H_{\mathrm{emp}}(y|\mathbf{V},\mathbf{S})$ minimizes uncertainty in the target domain by leveraging EAP over high-activation regions. By focusing on these regions, the event camera perception model tends to effectively learn to adapt its predictions to align with platform-specific activation cues in the target domain, achieving robust adaptation across varied platforms. + +# 3.3.EventBlend + +Cross-platform adaptation in event-based perception requires a selective integration of data from both source and target domains. Here, we introduce EventBlend, a data-mixing strategy designed to construct hybrid voxel grids by blending event data from the source and target domains, guided by high-activation regions identified through EAP. As shown in Fig. 2, by constructing these blended voxel grids, we aim to enhance the generalization ability across domain shifts, enabling the model to adapt effectively to the target distribution while leveraging reliable source labels. + +Event Density Maps. To identify regions where the source and target domains exhibit similar or divergent activation + +patterns, we employ density maps, which highlight areas of frequent event activity. For each source event data with voxel grid $\mathbf{V}_i^{\mathbf{v}}\in \mathbb{R}^{T\times H\times W}$ , we calculate a corresponding source density map, $\mathbf{D}_i^{\mathbf{v}}\in \mathbb{R}^{H\times W}$ , by summing activations along the temporal axis $T$ for each spatial location $(\mu ,\nu)$ : + +$$ +\mathbf {D} _ {i} ^ {\mathbf {v}} (\mu , \nu) = \sum_ {t = 1} ^ {T} | \mathbf {V} _ {i} ^ {\mathbf {v}} (t, \mu , \nu) |. \tag {5} +$$ + +For the target domain, we pre-compute an aggregated density map, $\widetilde{\mathbf{D}}^{\mathrm{d}}\in \mathbb{R}^{H\times W}$ , by summing activations over all target samples $\mathbf{V}_j^{\mathrm{d}}$ , where $j$ indexes each target sample: + +$$ +\tilde {\mathbf {D}} ^ {\mathrm {d}} (\mu , \nu) = \frac {1}{N ^ {\mathrm {d}}} \sum_ {j} \sum_ {t = 1} ^ {T} \left| \mathbf {V} _ {j} ^ {\mathrm {d}} (t, \mu , \nu) \right|, \tag {6} +$$ + +where $N^{\mathrm{d}}$ denotes the total number of samples in the target domain. These density maps highlight regions with significant activity, which aids in identifying areas where the domains may share similar activation patterns, consistent with the high-activation regions established by EAP. + +Similarity Map & Binary Mask. To guide selective blending, for each source sample, we construct a similarity map $\mathbf{SIM}_i\in \mathbb{R}^{H\times W}$ to quantify spatial overlap in activation patterns between two domains. This map is defined as: + +$$ +\operatorname {S I M} _ {i} (\mu , \nu) = 1 - \left| \mathbf {D} _ {i} ^ {\mathbf {v}} (\mu , \nu) - \tilde {\mathbf {D}} ^ {\mathrm {d}} (\mu , \nu) \right|. \tag {7} +$$ + +Here, high values in $\mathbf{SIM}_i$ correspond to regions where source and target domains exhibit similar activation intensities, aligning with the high-activation areas specified by EAP in Eq. (2). Next, we derive a binary mask $\mathcal{M}_i\in$ $\{0,1\}^{H\times W}$ by applying a threshold $\tau$ to the similarity map: + +$$ +\mathcal {M} _ {i} (\mu , \nu) = \left\{ \begin{array}{l l} \mathbf {1}, & \text {i f} \mathbf {S I M} _ {i} (\mu , \nu) \geq \tau , \\ \mathbf {0}, & \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +This mask directs the blending process by specifying which regions to retain from the source and which to adapt from the target. A mask value of 1 indicates that activations in $(\mu, \nu)$ from the source domain should be retained, while a value of 0 specifies a shift to the target domain activations. It is worth mentioning that for the selections of $\mathbf{D}_i^{\mathrm{v}}$ and $\tilde{\mathbf{D}}^{\mathrm{d}}$ in Eq. (5) and Eq. (6), we only consider activated pixel locations. That is to say, the similarity score does not count cases where both source and target are non-activated. + +Constructing Blended Event Voxel Grids. For a pair of source and target samples $\mathbf{V}_i^{\mathrm{v}}$ and $\mathbf{V}_i^{\mathrm{d}}$ , we create a new event voxel grid $\tilde{\mathbf{V}}_i \in \mathbb{R}^{T \times H \times W}$ , using $\mathcal{M}_i$ . For each spatial location $(\mu, \nu)$ , temporal sequences are selectively copied from either source and target based on $\mathcal{M}_i(\mu, \nu)$ : + +$$ +\tilde {\mathbf {V}} _ {i} (:, \mu , \nu) = \left\{ \begin{array}{l l} \mathbf {V} _ {i} ^ {\mathbf {v}} (:, \mu , \nu), & \text {i f} \mathcal {M} (\mu , \nu) = \mathbf {1}, \\ \mathbf {V} _ {i} ^ {\mathrm {d}} (:, \mu , \nu), & \text {i f} \mathcal {M} (\mu , \nu) = \mathbf {0}. \end{array} \right. \tag {9} +$$ + +This selective copying approach allows $\tilde{\mathbf{V}}_i$ to incorporate source-domain stability (i.e., human annotations) in high-similarity regions while adapting to target-specific patterns + +in low-similarity areas. Finally, we generate the label map $\tilde{y}_i\in \mathbb{R}^{H\times W}$ for $\tilde{\mathbf{V}}_i$ by combining ground truth labels from the source and pseudo-labels from the target. Here, the pseudo-labels can be obtained offline from a source-pretrained model or generated online using a mean teacher framework [76]. During adaptation, we use $\tilde{\mathbf{V}}_i$ and $\tilde{y}_i$ for supervised learning. The blended data contains activations and labels from both domains. We leverage EAP-aligned high-activation areas, i.e., Eq. (4), where source-target consistency is high, facilitating robust cross-domain adaptation. + +# 3.4.EventMatch + +In cross-platform adaptation, domain shifts manifest as discrepancies in spatial activation patterns, motion characteristics, and perspectives unique to each platform. The EAP (cf. Sec. 3.2) identifies high-activation areas for stable adaptation, while EventBlend (cf. Sec. 3.3) generates hybrid voxel grids that capture shared spatial structures across domains. However, without explicit feature alignment, the model might still struggle to fully adapt to platform-specific idiosyncrasies, especially in target-specific contexts. + +To address this, we propose the EventMatch technique to align the blended domain as an intermediary, enabling robust domain-invariant feature learning by maintaining source reliability while selectively adapting to target-specific characteristics. We introduce two fully convolutional discriminators, $\sigma_{1}$ and $\sigma_{2}$ , each responsible for a distinct aspect of feature alignment: + +- Source & Blended (via $\sigma_{1}$ ): We align the blended voxel grids $\tilde{\mathbf{V}}$ with the source domain, encouraging them to retain reliable source-domain characteristics. This alignment preserves robustness and ensures that the blended features maintain patterns learned from the ground truth supervised labels, i.e., $y$ , from the source domain. +- Blended & Target (via $\sigma_{2}$ ): Rather than the strict alignment, we regularize the blended features toward a soft alignment with the target, selectively adapting target-relevant characteristics. This alignment is focused on high-activation regions identified by EAP, allowing the blended features to bridge the domain gap in target-specific contexts without sacrificing source consistency. + +This dual-discriminator setup allows the blended features to act as an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive). + +Domain Adversarial Training. Each discriminator learns to distinguish between the feature representations from its respective domains, guiding the model to produce domain-agnostic features. Given feature outputs $\mathbf{F}^{\mathrm{v}}$ , $\mathbf{F}^{\mathrm{d}}$ , and $\tilde{\mathbf{F}}$ for the source, target, and blended domains, respectively: + +- Source-Blended Alignment: Discriminator $\sigma_{1}$ classifies $\mathbf{F}^{\vee}$ as source and $\tilde{\mathbf{F}}$ as blended, maximizing $\mathcal{L}_{\sigma_1}$ as: + +$$ +- \sum_ {\mu , \nu} \left[ \left(1 - z _ {1}\right) \log \sigma_ {1} (\tilde {\mathbf {F}}) + z _ {1} \log \sigma_ {1} \left(\mathbf {F} ^ {\mathrm {v}}\right) \right], \tag {10} +$$ + +Table 1. Benchmark results of platform adaptation from vehicle $(\mathcal{P}^{\mathrm{v}})$ to drone $(\mathcal{P}^{\mathrm{d}})$ . Target is trained with ground truth from the target domain. All scores are given in percentage (\%). The second best and best scores under each metric are highlighted in colors. + +
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o43.6933.8115.0411.8148.7111.570.928.4213.3325.488.1831.5114.880.042.41
AdaptSegNet [79]49.1435.3821.1612.1529.3723.570.170.4813.4538.2317.8548.7329.4235.550.40
CBST [104]57.9541.1824.3116.0233.0524.430.003.0818.2456.3216.8456.1523.6135.650.00
IntraDA [61]57.3738.8523.5815.9132.3123.170.004.9014.9156.7018.6754.9420.7133.080.00
DACS [77]59.8142.0127.0716.1435.1626.120.184.1118.4955.6421.7456.8134.6944.730.05
MIC [38]63.1145.6028.8717.4641.4025.190.0110.1122.8659.2520.8458.8633.9544.180.90
PLSR [94]64.6145.9329.6917.9942.0930.060.009.7523.3262.4820.6560.1531.6944.272.06
EventFly (Ours)69.1748.2032.6720.0146.6430.551.2710.9125.5067.1724.2161.0141.3044.546.21
Target •79.5752.2542.9023.3074.4839.407.100.3331.6771.9631.6467.8757.5166.1423.79
+ +where $z_{1} = 1$ for source and $z_{1} = 0$ for blended. This alignment ensures that the blended features preserve the supervised stability from the source. + +- Blended-Target Adaptation: Discriminator $\sigma_{2}$ adapts $\tilde{\mathbf{F}}$ toward high-activation regions in the target, where target-relevant properties are critical. It classifies $\tilde{\mathbf{F}}$ as blended and $\mathbf{F}^{\mathrm{d}}$ as target, maximizing $\mathcal{L}_{\sigma_2}$ as: + +$$ +- \sum_ {\mu , \nu} \left[ \left(1 - z _ {2}\right) \log \sigma_ {2} \left(\mathbf {F} ^ {\mathrm {d}}\right) + z _ {2} \log \sigma_ {2} (\tilde {\mathbf {F}}) \right], \tag {11} +$$ + +where $z_{2} = 1$ for target and $z_{2} = 0$ for blended. Here, $\sigma_{2}$ does not force complete target alignment but rather emphasizes adaptation in target-relevant areas, enabling $\tilde{\mathbf{F}}$ to maintain a balanced representation across domains. + +Optimization. To achieve consistent cross-domain alignment, the backbone network is trained adversarially against both discriminators. The goal is to produce $\tilde{\mathbf{F}}$ that each discriminator finds challenging to classify distinctly, thereby promoting a domain-agnostic representation: + +$$ +\mathcal {L} _ {\mathrm {a d v}} = - \sum_ {\mu , \nu} \left[ \log \sigma_ {1} (\tilde {\mathbf {F}}) + \log \sigma_ {2} (\tilde {\mathbf {F}}) \right]. \tag {12} +$$ + +This objective ensures that the backbone network produces features for blended data that capture both source robustness and target adaptability, guided by high-activation areas where source-target consistency is most beneficial. To achieve this goal, we set two weight coefficients, $\phi_{1}$ and $\phi_{2}$ , to balance between these two objectives. + +During training, we employ a min-max optimization framework to iteratively update the backbone, $\sigma_{1}$ , and $\sigma_{2}$ : + +- Step 1: Update $\sigma_{1}$ and $\sigma_{2}$ to improve the classification of the source, blended, and target features. +- Step 2: Update the backbone network to minimize $\mathcal{L}_{\mathrm{adv}}$ , encouraging it to generate intermediary features that support robust cross-domain generalization. + +Through this adversarial process, EventMatch allows the model to bridge the domain gap effectively, with blended features serving as a reliable intermediary that balances source stability with target-specific adaptability. + +Overall Framework. As shown in Fig. 2, the source and blended domains share the same network, while the network from the target domain is updated via exponential-moving average (EMA), encouraging consistency among domains. + +# 3.5. The EXPo Benchmark + +To facilitate robust event camera perception across diverse platforms and environments, we establish a large-scale event-based cross-platform adaptation benchmark: EXPo. Our benchmark is based on [12]. We include 89, 228 frames captured from three distinct platforms - vehicle, drone, and quadruped - spanning 21 sequences: 6 from the vehicle, 7 from the drone, and 8 from the quadruped. Following DSEC-Semantic [75], we support 11 semantic classes relevant to real-world event-based perception and cover varied environments including city, urban, suburban, and rural scenes. Due to space limits, more details are in Appendix. + +# 4. Experiments + +# 4.1. Settings + +Implementation Details. Our framework is implemented using PyTorch [62]. The backbone network is from E2VID [67], while the event segmentation head is from ESS [75]. The hyperparameters $\tau$ , $\phi_{1}$ and $\phi_{2}$ are set to 0.4, 1e-3, and 2e-3, respectively. The model is optimized using AdamW [54] and OneCycle learning rate scheduler [72] for 100k iterations. The learning rate and batch size are set to 1e-3 and 8. Due to the lack of baselines, we reproduce popular frame-based adaptation methods for adversarial training, contrastive learning, and self-training [38, 61, 77, 79, 94, 104]. To ensure fair comparisons, all methods adopt the same backbone and training iterations. The models are trained using NVIDIA RTX 4090 GPUs. + +Evaluation Metrics. Following conventions, we use Accuracy (Acc) and Intersection-over-Union (IoU) as the main metrics for evaluation. We also report mean Acc (mAcc), mean IoU (mIoU), and frequency-weighted IoU (fIoU) for a holistic comparison among different adaptation methods. + +Table 2. Benchmark results of platform adaptation from vehicle $(\mathcal{P}^{\mathrm{v}})$ to quadruped $(\mathcal{P}^{\mathrm{q}})$ . Target is trained with ground truth from the target domain. All scores are given in percentage (\%). The second best and best scores under each metric are in colors. + +
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o66.5939.7325.1516.5263.0139.263.8817.8810.1251.679.2768.0212.350.240.99
AdaptSegNet [79]67.2548.7332.7914.8945.0045.8830.0034.9212.2255.5015.8573.8416.0731.350.00
CBST [104]69.2549.5835.0614.9547.3954.6834.2736.8313.7856.1518.1374.2316.1834.060.00
IntraDA [61]68.2948.9134.2514.8243.7555.3632.6433.3911.6055.3117.0076.0020.3031.400.00
DACS [77]69.5553.8836.5114.6643.7257.2738.4335.4214.0257.1018.4376.1624.7936.210.00
MIC [38]70.7849.2236.9315.6051.7151.7333.5438.109.4454.2720.7474.4029.7941.780.70
PLSR [94]70.9153.6537.5715.2549.0453.2837.5436.6412.9157.6025.2975.9224.9239.850.24
EventFly (Ours)73.4254.1440.0515.7850.0761.3339.1741.9712.8359.1423.5179.8027.2642.652.86
Target •80.0260.5549.8419.5874.8056.2346.0855.2821.7959.9030.3177.2458.3862.475.81
+ +Table 3. Benchmark results of platform adaptation among the vehicle $(\mathcal{P}^{\mathrm{v}})$ , drone $(\mathcal{P}^{\mathrm{d}})$ , and quadruped $(\mathcal{P}^{\mathrm{q}})$ platforms, respectively. A total of six cross-platform adaptation settings are considered in our benchmark. Target is trained with ground truth from the target domain. All scores are given in percentage (\%). The second best and best scores under each metric are highlighted in colors. + +
MethodPv → PdPv → PqPd → PvPd → PQPq → PvPq → PdAverage
AccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoU
Source-Only o43.6915.0466.5925.1557.9120.7966.8323.0657.4921.3052.6216.8557.5220.37
AdaptSegNet [79]49.1421.1667.2532.7968.2929.5567.5733.9966.7430.6557.0720.9662.8828.18
DACS [77]59.8127.0769.5536.5171.7836.1067.7336.1171.2034.7860.7424.5066.8032.51
MIC [38]63.1128.8770.7836.9372.4636.8867.2936.2772.4635.2264.4926.1168.4333.38
PLSR [94]64.6129.6970.9137.5772.4637.1867.8336.2172.9336.3863.5727.3468.7234.06
EventFly (Ours)69.1732.6773.4240.0575.5039.9269.6837.3773.9337.7065.7828.7971.2536.08
Target •79.5742.9080.0249.8486.1255.9380.0249.8486.1255.9379.5742.9081.9049.56
+ +# 4.2. Comparative Study + +Adapt from Vehicle to Drone. Tab. 1 presents the results of transferring vehicle to drone, where there is a huge domain gap in between. The source-only (15.04% mIoU) and supervised target (42.90% mIoU) models validate this gap. Prior adaptation methods show good results in closing the domain gap, while our EventFly surpasses all competitors for almost all semantic classes, demonstrating strong robustness enhancements for event-based cross-platform adaptation. + +Adapt from Vehicle to Quadruped. We compare the adaptation methods from vehicle to quadruped in Tab. 2. Since the data and semantic distributions of these two platforms are closer (compared to drone), the adaptation methods show larger performance gains. We achieve a promising $40.05\%$ mIoU under this setting, which corresponds to a $59.2\%$ relative improvement over the source-only baseline. + +Qualitative Assessments. Fig. 3 visually compares the abilities of different cross-platform adaptation methods. As can be seen, the large domain gap between source and target platforms causes huge performance degradation, resulting in messy semantic predictions (source-only). While state-of-the-art adaptation methods [38, 94] restore the performance to some extent, their predictions still suffer from ambiguous object and background contours. Differently, our EventFly generates more accurate scene semantics, credited to the superior robustness from the EAP-driven designs. + +Cross-Platform Adaptations. Since our EXPo benchmark consists of three platforms, we benchmark state-of-the-art methods across a total of six adaptation settings, where each platform interchangeably serve as the source/target domains. As shown in Tab. 3, we observe discrepancies in adaptation difficulties. Specifically, the drone domain is the most difficult to adapt to, mainly due to its unique motion patterns and perspective-based dynamic. Our approach exhibits better robustness than prior arts, achieving the best metrics across all six settings. Such a strong generalization ability is crucial for the deployment of event camera perception algorithms in real-world environments. + +# 4.3. Ablation Study + +Component Analysis. Our EventFly framework compromises three key components: EventBlend, EventMatch, and the dual-backbone network. As shown in Tab. 4, each contribution provides distinct performance improvements. Since our designs are well-motivated by EAP, the combinations of any two (row $\# c$ ) and all (row $\# d$ ) of them further yield better cross-platform adaptation performance. + +Domain Blending Techniques. To validate that our EAP-driven EventBlend operation indeed encourages domain regularization effect, we compare the heuristic data mixing techniques [19, 90] and recent domain mixing methods [23, 77]. As shown in Fig. 4, conventional methods, which + +![](images/6a8585a447e0f2af9a175bca200cf29852f4f59e30ab08c245c597e5c292f420.jpg) +Figure 3. Qualitative assessments of adaptation from vehicle to drone (the first 3 rows), and from vehicle to quadruped (the last 3 rows). We use grayscaled event images for better visibility. The RGB frames are for reference purposes only. Best viewed in colors. + +Table 4. Ablation results of each component in cross-platform adaptation from vehicle $(\mathcal{P}^{\mathrm{v}})$ to drone $(\mathcal{P}^{\mathrm{d}})$ and vehicle $(\mathcal{P}^{\mathrm{v}})$ to quadruped $(\mathcal{P}^{\mathrm{q}})$ , respectively. Dual: The dual-branch network. Match: The EventMatch technique. Blend: The EventBlend operation. All scores are given in percentage (\%). We compare settings from the (a) source-only, (b) single-component effects, (c) double-component effects, (d) full configuration, and (e) supervised target results, respectively, in this ablation study. + +
#DualMatchBlendPvAccPd mIoUPvAccPq mIoU
(a)XXX43.6915.0466.5925.15
(b)X✓X59.8625.2867.7433.66
XX✓61.8527.9169.5536.82
(c)✓✓X66.5030.2470.9038.10
✓X✓68.4331.9672.6739.11
(d)✓✓✓69.1732.6773.4240.05
(e)Target79.5742.9080.0249.84
+ +neglect the event activation discrepancies, show sub-par performance in bridging the domains. Guided by EAP, our domain blending technique facilitates entropy minimization in high-activation regions, which in turn brings robust feature adaptation across heterogeneous platforms. + +Similarity Threshold. The strength of blending source and target platforms is vital in our framework. Therefore, we study the effect of the threshold parameter $\tau$ . Notably, the 0 (i.e., source-only) or high ( $>0.7$ ) value of $\tau$ brings limited gains, as the domain consistency regularization is weak. We found that the best possible trade-offs appear between 0.3 + +![](images/5f344061038c8e4ec8f53cf36f49f6a993f97e0b87c40e240bdcb61b0af0f7fb.jpg) +Figure 4. Ablation study (mIoU scores in %) on different domain-mixing techniques under the (left) $\mathcal{P}^{\mathrm{v}}\rightarrow \mathcal{P}^{\mathrm{d}}$ , (middle) $\mathcal{P}^{\mathrm{v}}\rightarrow \mathcal{P}^{\mathrm{q}}$ , and (right) $\mathcal{P}^{\mathrm{d}}\rightarrow \mathcal{P}^{\mathrm{v}}$ cross-platform adaptation settings. + +and 0.5, further validating the importance of bridging domains for cross-platform adaptation. + +# 5. Conclusion + +In this work, we presented EventFly, a framework for robust cross-platform adaptation in event-based dense perception, enabling robust deployment across the vehicle, drone, and quadruped platforms. Leveraging the proposed Event Activation Prior (EAP), EventBlend, and EventMatch, we addressed the unique adaptation challenges of event stream data collected from different robot platforms. We also introduced EXPo, a large-scale benchmark designed to evaluate cross-platform event perception capabilities. Our approach demonstrated notable improvements over existing methods. We hope this work lays the foundation for more adaptive event-based perception across real-world environments. + +# Acknowledgments + +This work is under the programme DesCartes and is supported by the National Research Foundation, Prime Minister's Office, Singapore, under its Campus for Research Excellence and Technological Enterprise (CREATE) programme. This work is also supported by the Apple Scholars in AI/ML Ph.D. Fellowship program. + +The authors would like to sincerely thank the Program Chairs, Area Chairs, and Reviewers for the time and efforts devoted during the review process. + +# References + +[1] Inigo Alonso and Ana C. Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1-10, 2019. +[2] Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. Ddd17: End-to-end davis driving dataset. In International Conference on Machine Learning Workshops, pages 1-9, 2017. +[3] Shristi Das Biswas, Adarsh Kosta, Chamika Liyanagedera, Marco Apolinario, and Kaushik Roy. Halsie: Hybrid approach to learning segmentation by simultaneously exploiting image and event modalities. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5952-5962, 2024. +[4] Michael Blösch, Stephan Weiss, Davide Scaramuzza, and Roland Siegwart. Vision based MAV navigation in unknown and unstructured environments. In IEEE International Conference on Robotics and Automation, pages 21-28, 2010. +[5] Chiara Boretti, Philippe Bich, Fabio Pareschi, Luciano Prono, Riccardo Rovatti, and Gianluca Setti. Pedro: An event-based dataset for person detection in robotics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4065-4070, 2023. +[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db 3 μs latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. +[7] Tim Brodermann, David Bruggemann, Christos Sakaridis, Kevin Ta, Odysseas Liagouris, Jason Corkill, and Luc Van Gool. Muses: The multi-sensor semantic perception dataset for driving under uncertainty. arXiv preprint arXiv:2401.12761, 2024. +[8] David Brüggemann, Christos Sakaridis, Prune Truong, and Luc Van Gool. Refign: Align and refine for adaptation of semantic segmentation to adverse conditions. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3174-3184, 2023. +[9] Spencer Carmichael, Austin Buchan, Mani Ramanagopal, Radhika Ravi, Ram Vasudevan, and Katherine A. Skinner. Dataset and benchmark: Novel sensors for autonomous vehicle perception. International Journal of Robotics Research, 2024. + +[10] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In IEEE International Conference on Robotics and Automation, pages 703-710, 2014. +[11] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-ilidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. In European Conference on Computer Vision Workshops, 2024. +[12] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multirobot, multi-sensor, multi-environment event dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4016-4023, 2023. +[13] Peiyu Chen, Weipeng Guan, Feng Huang, Yihan Zhong, Weisong Wen, Li-Ta Hsu, and Peng Lu. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 9(1):407-416, 2024. +[14] Zhiwen Chen, Zhiyu Zhu, Yifan Zhang, Junhui Hou, Guangming Shi, and Jinjian Wu. Segment any event streams via weighted adaptation of pivotal tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890-3900, 2024. +[15] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023. +[16] Fernando Cladera, Kenneth Chaney, M. Ani Hsieh, Camillo J. Taylor, and Vijay Kumar. Evmapper: High altitude orthomapping with event cameras. arXiv preprint arXiv:2409.18120, 2024. +[17] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. +[18] Javier Cuadrado, Ulysse Rançon, Benoit R. Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. +[19] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. +[20] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 3942-3951, 2023. +[21] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Hue dataset: High-resolution event and frame sequences for low-light vision. In European Conference on Computer Vision Workshops, pages 1-18, 2024. +[22] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a $1280 \times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\mu \mathrm{m}$ + +pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In IEEE International Solid-State Circuits Conference, pages 112-114, 2020. +[23] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In British Machine Vision Conference, 2020. +[24] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 (1):154-180, 2022. +[25] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022. +[26] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014): 1034-1040, 2024. +[27] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. +[28] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023. +[29] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. +[30] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In IEEE International Conference on 3D Vision, pages 197-206, 2021. +[31] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems, pages 529-536, 2004. +[32] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22867-22876, 2023. +[33] Dalia Hareb and Jean Martinet. Evsegsn: Neuromorphic semantic segmentation for event data. arXiv preprint arXiv:2406.14178, 2024. +[34] Javier Hidalgo-Carrio, Daniel Gehrig, and Davide Scaramuzza. Learning monocular dense depth from events. In IEEE International Conference on 3D Vision, pages 534-542, 2020. +[35] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781-5790, 2022. +[36] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for + +domain-adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. +[37] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision, pages 372–391, 2022. +[38] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023. +[39] Yuhwan Jeong, Hoonhee Cho, and Kuk-Jin Yoon. Towards robust event-based networks for nighttime via unpaired day-to-night event translation. In European Conference on Computer Vision, pages 286-306, 2024. +[40] Linglin Jing, Yiming Ding, Yunpeng Gao, Zhigang Wang, Xu Yan, Dong Wang, Gerald Schaefer, Hui Fang, Bin Zhao, and Xuelong Li. Hpl-ess: Hybrid pseudo-labeling for unsupervised event-based semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23128–23137, 2024. +[41] Uday Kamal and Saibal Mukhopadhyay. Efficient learning of event-based dense representation using hierarchical memories with adaptive update. In European Conference on Computer Vision, pages 74-89, 2024. +[42] Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 620(7976):982-987, 2023. +[43] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In IEEE/CVF International Conference on Computer Vision, pages 2146-2156, 2021. +[44] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2496-24976, 2024. +[45] Simon Klenk, David Bonello, Lukas Koestler, Nikita Araslanov, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2378-2388, 2024. +[46] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023. +[47] Lingdong Kong, Youquan Liu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Openness: Event-based semantic scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15686-15698, 2024. +[48] Wouter M. Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):766-785, 2021. + +[49] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):14020-14037, 2023. +[50] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5743-5762, 2024. +[51] Yahao Liu, Jinhong Deng, Xinchen Gao, Wen Li, and Lixin Duan. Bapa-net: Boundary adaptation and prototype alignment for cross-domain semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 8801-8811, 2021. +[52] Antonio Loquercio, Ana I. Maqueda, Carlos R. Del-Blanco, and Davide Scaramuzza. Dronet: Learning to fly by driving. IEEE Robotics and Automation Letters, 3(2):1088-1095, 2018. +[53] Antonio Loquercio, Elia Kaufmann, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Learning high-speed flight in the wild. Science Robotics, 6(59):5810, 2021. +[54] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. +[55] Dongyue Lu, Lingdong Kong, Gim Hee Lee, Camille Simon Chane, and Wei Tsang Ooi. Flexevent: Event camera object detection at arbitrary frequencies. arXiv preprint arXiv:2412.06708, 2024. +[56] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso García, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5419-5427, 2018. +[57] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. IEEE Robotics and Automation Letters, 7(2):3515-3522, 2022. +[58] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021. +[59] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34 (6):1425-1440, 2018. +[60] Poojan Oza, Vishwanath A. Sindagi, Vibashan VS, and Vishal M. Patel. Unsupervised domain adaptation of object detectors: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4018-4040, 2024. +[61] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020. +[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, + +Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8026-8037, 2019. +[63] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2024. +[64] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems, pages 16639-16652, 2020. +[65] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1): 259-275, 2010. +[66] Ulysse Rançon, Javier Cuadrado-Anibarro, Benoit R. Cottereau, and Timothée Masquelier. Stereospike: Depth learning with a spiking neural network. IEEE Access, 10: 127428-127439, 2022. +[67] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):1964-1980, 2019. +[68] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8197-8207, 2021. +[69] Manuel Schwonberg, Joshua Niemeijer, Jan-Aike Termöhlen, Nico M. Schmidt, Hanno Gottschalk, and Tim Fingscheidt. Survey on unsupervised domain adaptation for semantic segmentation for visual perception in automated driving. IEEE Access, 11:54296-54336, 2023. +[70] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 12:51275-51306, 2024. +[71] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. +[72] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. arXiv preprint arXiv:1708.07120, 2017. +[73] Bongki Son, Yunjae Suh, Sungho Kim, Heejae Jung, Jun-Seok Kim, Changwoo Shin, Keunju Park, Kyoobin Lee, Jinman Park, Jooyeon Woo, et al. 4.1 a $640 \times 480$ dynamic vision sensor with a $9\mu \mathrm{m}$ pixel and 300meps address-event representation. In IEEE International Solid-State Circuits Conference, pages 66-67, 2017. + +[74] Lea Steffen, Daniel Reichard, Jakob Weinland, Jacques Kaiser, Arne Roennau, and Rüdiger Dillmann. Neuromorphic stereo vision: A survey of bio-inspired sensors and algorithms. Frontiers in Neuroscience, 13:28, 2019. +[75] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In European Conference on Computer Vision, pages 341-357, 2022. +[76] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 195–1204, 2017. +[77] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021. +[78] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshops, pages 350-357, 2021. +[79] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7472-7481, 2018. +[80] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate slam? combining events, images, and imu for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3 (2):994-1001, 2018. +[81] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517-2526, 2019. +[82] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019. +[83] Lin Wang, Yujeong Chae, and Kuk-Jin Yoon. Dual transfer learning for event-based end-task prediction via plug-gable event to image translation. In IEEE/CVF International Conference on Computer Vision, pages 2135-2145, 2021. +[84] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 608-619, 2021. +[85] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 19248-19257, 2024. +[86] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyuan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 21572-21581, 2023. +[87] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9004-9021, 2023. +[88] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In IEEE/CVF International Conference on Computer Vision, pages 10699-10709, 2023. +[89] Zhen Yao and Mooi Choo Chuah. Event-guided low-light video semantic segmentation. arXiv preprint arXiv:2411.00639, 2024. +[90] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. +[91] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12):14679-14694, 2023. +[92] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12414-12424, 2021. +[93] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, and Sebastian Scherer. Subt-mrs dataset: Pushing slam towards all-weather environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024. +[94] Xingchen Zhao, Niluthpol Chowdhury Mithun, Abhinav Rajvanshi, Han-Pang Chiu, and Supun Samarasekera. Unsupervised domain adaptation for semantic segmentation with pseudo label self-refinement. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2399-2409, 2024. +[95] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024. +[96] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023. +[97] Qianyu Zhou, Zhengyang Feng, Qiqi Gu, Jiangmiao Pang, Guangliang Cheng, Xuequan Lu, Jianping Shi, and + +Lizhuang Ma. Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):804-817, 2023. +[98] Zhuyun Zhou, Zongwei Wu, Rémi Boutteau, Fan Yang, Cédric Demonceaux, and Dominique Gingac. Rgb-event fusion for moving object detection in autonomous driving. In IEEE International Conference on Robotics and Automation, pages 7808-7815, 2023. +[99] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Robotics: Science and Systems, 2018. +[100] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. +[101] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. +[102] Shifan Zhu, Zixun Xiong, and Donghyun Kim. Cear: Comprehensive event camera dataset for rapid perception of agile quadruped robots. IEEE Robotics and Automation Letters, 9(10):8999-9006, 2024. +[103] Rong Zou, Manasi Muglikar, Nico Messikommer, and Davide Scaramuzza. Seeing behind dynamic occlusions with event cameras. arXiv preprint arXiv:2307.15829, 2023. +[104] Yang Zou, Zhiding Yu, B. V. K. Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In European Conference on Computer Vision, pages 289-305, 2018. +[105] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, pages 12846-128567, 2023. \ No newline at end of file diff --git a/2025/EventFly_ Event Camera Perception from Ground to the Sky/images.zip b/2025/EventFly_ Event Camera Perception from Ground to the Sky/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a51b9c4b43cc3f23a768723ef09f35b362adca7 --- /dev/null +++ b/2025/EventFly_ Event Camera Perception from Ground to the Sky/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af46cd4ca30bd8af32f220d786e39417c1fe72bd4b648ae46d1f58949f6c471b +size 758706 diff --git a/2025/EventFly_ Event Camera Perception from Ground to the Sky/layout.json b/2025/EventFly_ Event Camera Perception from Ground to the Sky/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..213416dcd3d12ec4800387ff8bfa563e82be8a57 --- /dev/null +++ b/2025/EventFly_ Event Camera Perception from Ground to the Sky/layout.json @@ -0,0 +1,16191 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 87, + 492, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 87, + 492, + 104 + ], + "spans": [ + { + "bbox": [ + 117, + 87, + 492, + 104 + ], + "type": "text", + "content": "EventFly: Event Camera Perception from Ground to the Sky" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "text", + "content": "Lingdong Kong" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "text", + "content": ", Dongyue Lu" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "text", + "content": ", Xiang Xu" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "text", + "content": ", Lai Xing Ng" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "text", + "content": ", Wei Tsang Ooi" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "text", + "content": ", Benoit R. Cottereau" + }, + { + "bbox": [ + 61, + 125, + 553, + 141 + ], + "type": "inline_equation", + "content": "^{5,6}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "spans": [ + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "type": "text", + "content": "CNRS@CREATE " + }, + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 144, + 509, + 158 + ], + "type": "text", + "content": "Nanjing University of Aeronautics and Astronautics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "type": "text", + "content": "Institute for Infocomm Research, A*STAR " + }, + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "type": "text", + "content": "IPAL, CNRS IRL 2955, Singapore " + }, + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 62, + 159, + 552, + 171 + ], + "type": "text", + "content": "CerCo, CNRS UMR 5549, Université Toulouse III" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 166, + 175, + 445, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 175, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 166, + 175, + 445, + 190 + ], + "type": "text", + "content": "Project Page: https://event-fly.github.io" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 107, + 201, + 122, + 218 + ], + "blocks": [ + { + "bbox": [ + 107, + 201, + 122, + 218 + ], + "lines": [ + { + "bbox": [ + 107, + 201, + 122, + 218 + ], + "spans": [ + { + "bbox": [ + 107, + 201, + 122, + 218 + ], + "type": "image", + "image_path": "5853b6c60761c6587a74967048b16a75ad001ca58c3dcb39334bd3326cf8a81a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 204, + 165, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 204, + 165, + 215 + ], + "spans": [ + { + "bbox": [ + 125, + 204, + 165, + 215 + ], + "type": "text", + "content": "Vehicle" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 75, + 224, + 95, + 243 + ], + "blocks": [ + { + "bbox": [ + 75, + 224, + 95, + 243 + ], + "lines": [ + { + "bbox": [ + 75, + 224, + 95, + 243 + ], + "spans": [ + { + "bbox": [ + 75, + 224, + 95, + 243 + ], + "type": "image", + "image_path": "272e59af0f8980ad3f170e9f0a09e528b01ce5f78d6b110e47c230fc12efdc45.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 243, + 104, + 251 + ], + "lines": [ + { + "bbox": [ + 65, + 243, + 104, + 251 + ], + "spans": [ + { + "bbox": [ + 65, + 243, + 104, + 251 + ], + "type": "text", + "content": "Viewpoint" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 126, + 225, + 148, + 243 + ], + "blocks": [ + { + "bbox": [ + 126, + 225, + 148, + 243 + ], + "lines": [ + { + "bbox": [ + 126, + 225, + 148, + 243 + ], + "spans": [ + { + "bbox": [ + 126, + 225, + 148, + 243 + ], + "type": "image", + "image_path": "0380827b2d44cf256cc5e2a08937d8621f0f10c63d16060b236ccd30a9f9ddac.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 243, + 149, + 251 + ], + "lines": [ + { + "bbox": [ + 125, + 243, + 149, + 251 + ], + "spans": [ + { + "bbox": [ + 125, + 243, + 149, + 251 + ], + "type": "text", + "content": "Speed" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 182, + 225, + 200, + 243 + ], + "blocks": [ + { + "bbox": [ + 182, + 225, + 200, + 243 + ], + "lines": [ + { + "bbox": [ + 182, + 225, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 182, + 225, + 200, + 243 + ], + "type": "image", + "image_path": "38bf90130853b4f6eafa14489a686655d4c0a7b784d764d6de96368ce090b95a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 243, + 208, + 251 + ], + "lines": [ + { + "bbox": [ + 175, + 243, + 208, + 251 + ], + "spans": [ + { + "bbox": [ + 175, + 243, + 208, + 251 + ], + "type": "text", + "content": "Stability" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 275, + 201, + 296, + 219 + ], + "blocks": [ + { + "bbox": [ + 275, + 201, + 296, + 219 + ], + "lines": [ + { + "bbox": [ + 275, + 201, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 275, + 201, + 296, + 219 + ], + "type": "image", + "image_path": "fae723af8c1007b1cf8cf9fef9304bf2c34f57bcd00faafe627a5b95d34e627e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 298, + 205, + 331, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 205, + 331, + 215 + ], + "spans": [ + { + "bbox": [ + 298, + 205, + 331, + 215 + ], + "type": "text", + "content": "Drone" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 242, + 224, + 263, + 243 + ], + "blocks": [ + { + "bbox": [ + 242, + 224, + 263, + 243 + ], + "lines": [ + { + "bbox": [ + 242, + 224, + 263, + 243 + ], + "spans": [ + { + "bbox": [ + 242, + 224, + 263, + 243 + ], + "type": "image", + "image_path": "7b6e06c1788b26e6da125f6e1ed9bd79cc723f8d5a9cffedff2f3ce0862a1792.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 243, + 272, + 251 + ], + "lines": [ + { + "bbox": [ + 233, + 243, + 272, + 251 + ], + "spans": [ + { + "bbox": [ + 233, + 243, + 272, + 251 + ], + "type": "text", + "content": "Viewpoint" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 295, + 225, + 317, + 243 + ], + "blocks": [ + { + "bbox": [ + 295, + 225, + 317, + 243 + ], + "lines": [ + { + "bbox": [ + 295, + 225, + 317, + 243 + ], + "spans": [ + { + "bbox": [ + 295, + 225, + 317, + 243 + ], + "type": "image", + "image_path": "fc0b719d722a0d34af55598a5fa6feb11992a3adf105d490b193544aa6f87359.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 243, + 318, + 251 + ], + "lines": [ + { + "bbox": [ + 293, + 243, + 318, + 251 + ], + "spans": [ + { + "bbox": [ + 293, + 243, + 318, + 251 + ], + "type": "text", + "content": "Speed" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 350, + 225, + 367, + 243 + ], + "blocks": [ + { + "bbox": [ + 350, + 225, + 367, + 243 + ], + "lines": [ + { + "bbox": [ + 350, + 225, + 367, + 243 + ], + "spans": [ + { + "bbox": [ + 350, + 225, + 367, + 243 + ], + "type": "image", + "image_path": "12714ec2df0dabace022464e5d44748c26fcc3df4fa96bd3fece33094c9b4b57.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 243, + 374, + 251 + ], + "lines": [ + { + "bbox": [ + 343, + 243, + 374, + 251 + ], + "spans": [ + { + "bbox": [ + 343, + 243, + 374, + 251 + ], + "type": "text", + "content": "Stability" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 432, + 201, + 449, + 219 + ], + "blocks": [ + { + "bbox": [ + 432, + 201, + 449, + 219 + ], + "lines": [ + { + "bbox": [ + 432, + 201, + 449, + 219 + ], + "spans": [ + { + "bbox": [ + 432, + 201, + 449, + 219 + ], + "type": "image", + "image_path": "93e5fc4ac5e48669deaf86565705b9cbc570be48789efe743c270687af8c09b9.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 451, + 205, + 510, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 205, + 510, + 216 + ], + "spans": [ + { + "bbox": [ + 451, + 205, + 510, + 216 + ], + "type": "text", + "content": "Quadruped" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 410, + 224, + 430, + 243 + ], + "blocks": [ + { + "bbox": [ + 410, + 224, + 430, + 243 + ], + "lines": [ + { + "bbox": [ + 410, + 224, + 430, + 243 + ], + "spans": [ + { + "bbox": [ + 410, + 224, + 430, + 243 + ], + "type": "image", + "image_path": "17546392b375726ada3df0d26c7da180715b3478ab60db8ce698ae92c7e68f27.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 243, + 440, + 251 + ], + "lines": [ + { + "bbox": [ + 402, + 243, + 440, + 251 + ], + "spans": [ + { + "bbox": [ + 402, + 243, + 440, + 251 + ], + "type": "text", + "content": "Viewpoint" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 462, + 225, + 484, + 243 + ], + "blocks": [ + { + "bbox": [ + 462, + 225, + 484, + 243 + ], + "lines": [ + { + "bbox": [ + 462, + 225, + 484, + 243 + ], + "spans": [ + { + "bbox": [ + 462, + 225, + 484, + 243 + ], + "type": "image", + "image_path": "9c47e2fc092d85e7050aee94b8af09d6c4e73894daf51e3d493d78f6fa26ee55.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 243, + 485, + 251 + ], + "lines": [ + { + "bbox": [ + 462, + 243, + 485, + 251 + ], + "spans": [ + { + "bbox": [ + 462, + 243, + 485, + 251 + ], + "type": "text", + "content": "Speed" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 517, + 225, + 534, + 243 + ], + "blocks": [ + { + "bbox": [ + 517, + 225, + 534, + 243 + ], + "lines": [ + { + "bbox": [ + 517, + 225, + 534, + 243 + ], + "spans": [ + { + "bbox": [ + 517, + 225, + 534, + 243 + ], + "type": "image", + "image_path": "2e63cb70da619ea54e6a9a2f69cea3994b725304e8303b7be77a65583bdc7744.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 243, + 542, + 251 + ], + "lines": [ + { + "bbox": [ + 511, + 243, + 542, + 251 + ], + "spans": [ + { + "bbox": [ + 511, + 243, + 542, + 251 + ], + "type": "text", + "content": "Stability" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 61, + 258, + 75, + 301 + ], + "blocks": [ + { + "bbox": [ + 61, + 258, + 75, + 301 + ], + "lines": [ + { + "bbox": [ + 61, + 258, + 75, + 301 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 75, + 301 + ], + "type": "image", + "image_path": "547e8d07899970ee724622fa5773af2ed3589fe6c77876aa3e5232900cfb2e80.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 262, + 88, + 271 + ], + "lines": [ + { + "bbox": [ + 70, + 262, + 88, + 271 + ], + "spans": [ + { + "bbox": [ + 70, + 262, + 88, + 271 + ], + "type": "text", + "content": "High" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 272, + 102, + 281 + ], + "lines": [ + { + "bbox": [ + 71, + 272, + 102, + 281 + ], + "spans": [ + { + "bbox": [ + 71, + 272, + 102, + 281 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 281, + 102, + 291 + ], + "lines": [ + { + "bbox": [ + 71, + 281, + 102, + 291 + ], + "spans": [ + { + "bbox": [ + 71, + 281, + 102, + 291 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 291, + 88, + 297 + ], + "lines": [ + { + "bbox": [ + 71, + 291, + 88, + 297 + ], + "spans": [ + { + "bbox": [ + 71, + 291, + 88, + 297 + ], + "type": "text", + "content": "Low" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 111, + 258, + 125, + 301 + ], + "blocks": [ + { + "bbox": [ + 111, + 258, + 125, + 301 + ], + "lines": [ + { + "bbox": [ + 111, + 258, + 125, + 301 + ], + "spans": [ + { + "bbox": [ + 111, + 258, + 125, + 301 + ], + "type": "image", + "image_path": "722497fff7d34657c0c3588afe73e091ac960e0df45cb05cadd6dd1039a0d9b9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 262, + 141, + 270 + ], + "lines": [ + { + "bbox": [ + 126, + 262, + 141, + 270 + ], + "spans": [ + { + "bbox": [ + 126, + 262, + 141, + 270 + ], + "type": "text", + "content": "Fast" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 272, + 154, + 281 + ], + "lines": [ + { + "bbox": [ + 124, + 272, + 154, + 281 + ], + "spans": [ + { + "bbox": [ + 124, + 272, + 154, + 281 + ], + "type": "text", + "content": "Medium " + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 281, + 154, + 291 + ], + "lines": [ + { + "bbox": [ + 124, + 281, + 154, + 291 + ], + "spans": [ + { + "bbox": [ + 124, + 281, + 154, + 291 + ], + "type": "text", + "content": "G" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 291, + 142, + 299 + ], + "lines": [ + { + "bbox": [ + 124, + 291, + 142, + 299 + ], + "spans": [ + { + "bbox": [ + 124, + 291, + 142, + 299 + ], + "type": "text", + "content": "Slow" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 164, + 258, + 178, + 301 + ], + "blocks": [ + { + "bbox": [ + 164, + 258, + 178, + 301 + ], + "lines": [ + { + "bbox": [ + 164, + 258, + 178, + 301 + ], + "spans": [ + { + "bbox": [ + 164, + 258, + 178, + 301 + ], + "type": "image", + "image_path": "2dc7517d49fb0dc1d09ee9ec1a3377475b1a80feea9223a51633793297b04a89.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 262, + 206, + 270 + ], + "lines": [ + { + "bbox": [ + 178, + 262, + 206, + 270 + ], + "spans": [ + { + "bbox": [ + 178, + 262, + 206, + 270 + ], + "type": "text", + "content": "Smooth" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 271, + 207, + 281 + ], + "lines": [ + { + "bbox": [ + 178, + 271, + 207, + 281 + ], + "spans": [ + { + "bbox": [ + 178, + 271, + 207, + 281 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 281, + 209, + 290 + ], + "lines": [ + { + "bbox": [ + 178, + 281, + 209, + 290 + ], + "spans": [ + { + "bbox": [ + 178, + 281, + 209, + 290 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 290, + 210, + 298 + ], + "lines": [ + { + "bbox": [ + 178, + 290, + 210, + 298 + ], + "spans": [ + { + "bbox": [ + 178, + 290, + 210, + 298 + ], + "type": "text", + "content": "Unstable" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 227, + 258, + 240, + 301 + ], + "blocks": [ + { + "bbox": [ + 227, + 258, + 240, + 301 + ], + "lines": [ + { + "bbox": [ + 227, + 258, + 240, + 301 + ], + "spans": [ + { + "bbox": [ + 227, + 258, + 240, + 301 + ], + "type": "image", + "image_path": "7430a70af36122e6c18667f04af799155d96ade4050cf6c1feffe5b2dedd8492.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 262, + 257, + 272 + ], + "lines": [ + { + "bbox": [ + 238, + 262, + 257, + 272 + ], + "spans": [ + { + "bbox": [ + 238, + 262, + 257, + 272 + ], + "type": "text", + "content": "High" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 274, + 269, + 283 + ], + "lines": [ + { + "bbox": [ + 238, + 274, + 269, + 283 + ], + "spans": [ + { + "bbox": [ + 238, + 274, + 269, + 283 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 283, + 255, + 292 + ], + "lines": [ + { + "bbox": [ + 239, + 283, + 255, + 292 + ], + "spans": [ + { + "bbox": [ + 239, + 283, + 255, + 292 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 292, + 255, + 298 + ], + "lines": [ + { + "bbox": [ + 239, + 292, + 255, + 298 + ], + "spans": [ + { + "bbox": [ + 239, + 292, + 255, + 298 + ], + "type": "text", + "content": "Low" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 281, + 258, + 291, + 301 + ], + "blocks": [ + { + "bbox": [ + 281, + 258, + 291, + 301 + ], + "lines": [ + { + "bbox": [ + 281, + 258, + 291, + 301 + ], + "spans": [ + { + "bbox": [ + 281, + 258, + 291, + 301 + ], + "type": "image", + "image_path": "e662280e6c6a9d42bda8875c4a36491c55f5bdbc4228c488d4d0d2d4bfbb8585.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 262, + 309, + 270 + ], + "lines": [ + { + "bbox": [ + 291, + 262, + 309, + 270 + ], + "spans": [ + { + "bbox": [ + 291, + 262, + 309, + 270 + ], + "type": "text", + "content": "Fast" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 291, + 274, + 322, + 283 + ], + "lines": [ + { + "bbox": [ + 291, + 274, + 322, + 283 + ], + "spans": [ + { + "bbox": [ + 291, + 274, + 322, + 283 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 291, + 283, + 310, + 292 + ], + "lines": [ + { + "bbox": [ + 291, + 283, + 310, + 292 + ], + "spans": [ + { + "bbox": [ + 291, + 283, + 310, + 292 + ], + "type": "text", + "content": "。" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 291, + 292, + 310, + 298 + ], + "lines": [ + { + "bbox": [ + 291, + 292, + 310, + 298 + ], + "spans": [ + { + "bbox": [ + 291, + 292, + 310, + 298 + ], + "type": "text", + "content": "Slow" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 328, + 258, + 337, + 301 + ], + "blocks": [ + { + "bbox": [ + 328, + 258, + 337, + 301 + ], + "lines": [ + { + "bbox": [ + 328, + 258, + 337, + 301 + ], + "spans": [ + { + "bbox": [ + 328, + 258, + 337, + 301 + ], + "type": "image", + "image_path": "090f5d8154999e61ba71d5469c3ce4d016f97c5af25f637f0ef5eb3420b3c34f.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 262, + 374, + 270 + ], + "lines": [ + { + "bbox": [ + 343, + 262, + 374, + 270 + ], + "spans": [ + { + "bbox": [ + 343, + 262, + 374, + 270 + ], + "type": "text", + "content": "Smooth" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 274, + 375, + 283 + ], + "lines": [ + { + "bbox": [ + 343, + 274, + 375, + 283 + ], + "spans": [ + { + "bbox": [ + 343, + 274, + 375, + 283 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 283, + 378, + 292 + ], + "lines": [ + { + "bbox": [ + 343, + 283, + 378, + 292 + ], + "spans": [ + { + "bbox": [ + 343, + 283, + 378, + 292 + ], + "type": "text", + "content": "hckable" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 292, + 378, + 299 + ], + "lines": [ + { + "bbox": [ + 343, + 292, + 378, + 299 + ], + "spans": [ + { + "bbox": [ + 343, + 292, + 378, + 299 + ], + "type": "text", + "content": "Unstable" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 395, + 258, + 405, + 301 + ], + "blocks": [ + { + "bbox": [ + 395, + 258, + 405, + 301 + ], + "lines": [ + { + "bbox": [ + 395, + 258, + 405, + 301 + ], + "spans": [ + { + "bbox": [ + 395, + 258, + 405, + 301 + ], + "type": "image", + "image_path": "589250169122ebb40635a3b5bcdaf7866afcc060f323b3ee93d07df75f51a282.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 262, + 425, + 272 + ], + "lines": [ + { + "bbox": [ + 406, + 262, + 425, + 272 + ], + "spans": [ + { + "bbox": [ + 406, + 262, + 425, + 272 + ], + "type": "text", + "content": "High" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 273, + 439, + 283 + ], + "lines": [ + { + "bbox": [ + 406, + 273, + 439, + 283 + ], + "spans": [ + { + "bbox": [ + 406, + 273, + 439, + 283 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 289, + 424, + 300 + ], + "lines": [ + { + "bbox": [ + 406, + 289, + 424, + 300 + ], + "spans": [ + { + "bbox": [ + 406, + 289, + 424, + 300 + ], + "type": "text", + "content": "Low" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 445, + 258, + 455, + 301 + ], + "blocks": [ + { + "bbox": [ + 445, + 258, + 455, + 301 + ], + "lines": [ + { + "bbox": [ + 445, + 258, + 455, + 301 + ], + "spans": [ + { + "bbox": [ + 445, + 258, + 455, + 301 + ], + "type": "image", + "image_path": "1f834fdae78fc0aa751cdfd0b48310b47b7b4f47c74f22491914de0600ae7d16.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 262, + 476, + 271 + ], + "lines": [ + { + "bbox": [ + 455, + 262, + 476, + 271 + ], + "spans": [ + { + "bbox": [ + 455, + 262, + 476, + 271 + ], + "type": "text", + "content": "Fast" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 455, + 273, + 476, + 283 + ], + "lines": [ + { + "bbox": [ + 455, + 273, + 476, + 283 + ], + "spans": [ + { + "bbox": [ + 455, + 273, + 476, + 283 + ], + "type": "text", + "content": "Med" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 455, + 289, + 479, + 298 + ], + "lines": [ + { + "bbox": [ + 455, + 289, + 479, + 298 + ], + "spans": [ + { + "bbox": [ + 455, + 289, + 479, + 298 + ], + "type": "text", + "content": "Slow" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 501, + 258, + 511, + 301 + ], + "blocks": [ + { + "bbox": [ + 501, + 258, + 511, + 301 + ], + "lines": [ + { + "bbox": [ + 501, + 258, + 511, + 301 + ], + "spans": [ + { + "bbox": [ + 501, + 258, + 511, + 301 + ], + "type": "image", + "image_path": "4c683c3f4d0f92985f10ab440e3c60db10c280370b861fb0a2fffc9508f8f7d1.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 262, + 544, + 271 + ], + "lines": [ + { + "bbox": [ + 512, + 262, + 544, + 271 + ], + "spans": [ + { + "bbox": [ + 512, + 262, + 544, + 271 + ], + "type": "text", + "content": "Smooth" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 512, + 273, + 544, + 283 + ], + "lines": [ + { + "bbox": [ + 512, + 273, + 544, + 283 + ], + "spans": [ + { + "bbox": [ + 512, + 273, + 544, + 283 + ], + "type": "text", + "content": "Medium" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 512, + 289, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 512, + 289, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 512, + 289, + 547, + 298 + ], + "type": "text", + "content": "Unstable" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 61, + 323, + 110, + 372 + ], + "blocks": [ + { + "bbox": [ + 90, + 310, + 184, + 319 + ], + "lines": [ + { + "bbox": [ + 90, + 310, + 184, + 319 + ], + "spans": [ + { + "bbox": [ + 90, + 310, + 184, + 319 + ], + "type": "text", + "content": "Event Data Distribution" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 323, + 110, + 372 + ], + "lines": [ + { + "bbox": [ + 61, + 323, + 110, + 372 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 110, + 372 + ], + "type": "image", + "image_path": "0e96fe90ac2ca6c14ea19373eae78d99b907c76fb4a1d3757f17630dbfa480a0.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 113, + 324, + 161, + 372 + ], + "blocks": [ + { + "bbox": [ + 113, + 324, + 161, + 372 + ], + "lines": [ + { + "bbox": [ + 113, + 324, + 161, + 372 + ], + "spans": [ + { + "bbox": [ + 113, + 324, + 161, + 372 + ], + "type": "image", + "image_path": "f29ea8a809278b472a63f846333db188c71f1b1edd09e28d06ede1d8569077fb.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 166, + 324, + 214, + 372 + ], + "blocks": [ + { + "bbox": [ + 166, + 324, + 214, + 372 + ], + "lines": [ + { + "bbox": [ + 166, + 324, + 214, + 372 + ], + "spans": [ + { + "bbox": [ + 166, + 324, + 214, + 372 + ], + "type": "image", + "image_path": "48a278e813bd6ad73baaa9338462d23ce56dfb2a224f3406650e42f7910df55a.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + } + ], + "index": 78 + }, + { + "type": "image", + "bbox": [ + 61, + 394, + 110, + 422 + ], + "blocks": [ + { + "bbox": [ + 82, + 380, + 194, + 390 + ], + "lines": [ + { + "bbox": [ + 82, + 380, + 194, + 390 + ], + "spans": [ + { + "bbox": [ + 82, + 380, + 194, + 390 + ], + "type": "text", + "content": "Event Semantic Distribution" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 394, + 110, + 422 + ], + "lines": [ + { + "bbox": [ + 61, + 394, + 110, + 422 + ], + "spans": [ + { + "bbox": [ + 61, + 394, + 110, + 422 + ], + "type": "image", + "image_path": "1cadc832c89bfb512c3223c2a34c5df55f2b854686a9481059492584c0d8fc04.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 423, + 94, + 431 + ], + "lines": [ + { + "bbox": [ + 76, + 423, + 94, + 431 + ], + "spans": [ + { + "bbox": [ + 76, + 423, + 94, + 431 + ], + "type": "text", + "content": "Road" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_caption" + } + ], + "index": 80 + }, + { + "type": "image", + "bbox": [ + 113, + 394, + 162, + 422 + ], + "blocks": [ + { + "bbox": [ + 113, + 394, + 162, + 422 + ], + "lines": [ + { + "bbox": [ + 113, + 394, + 162, + 422 + ], + "spans": [ + { + "bbox": [ + 113, + 394, + 162, + 422 + ], + "type": "image", + "image_path": "f6c2e20de6a9d57ab420640ed802932506d6bbb8ddf59260b783564626f1ceb7.jpg" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 423, + 145, + 431 + ], + "lines": [ + { + "bbox": [ + 130, + 423, + 145, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 145, + 431 + ], + "type": "text", + "content": "Car" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_caption" + } + ], + "index": 82 + }, + { + "type": "image", + "bbox": [ + 166, + 394, + 214, + 422 + ], + "blocks": [ + { + "bbox": [ + 166, + 394, + 214, + 422 + ], + "lines": [ + { + "bbox": [ + 166, + 394, + 214, + 422 + ], + "spans": [ + { + "bbox": [ + 166, + 394, + 214, + 422 + ], + "type": "image", + "image_path": "62b1ac51548c559d026477cff98ed06cc400e996f40ebada43d6780cf08bd661.jpg" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 423, + 198, + 432 + ], + "lines": [ + { + "bbox": [ + 184, + 423, + 198, + 432 + ], + "spans": [ + { + "bbox": [ + 184, + 423, + 198, + 432 + ], + "type": "text", + "content": "Sky" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_caption" + } + ], + "index": 84 + }, + { + "type": "image", + "bbox": [ + 61, + 435, + 110, + 464 + ], + "blocks": [ + { + "bbox": [ + 61, + 435, + 110, + 464 + ], + "lines": [ + { + "bbox": [ + 61, + 435, + 110, + 464 + ], + "spans": [ + { + "bbox": [ + 61, + 435, + 110, + 464 + ], + "type": "image", + "image_path": "311e1fa70f9d60671194957c69ed89db0762192e158333567d8c2ede21cefd09.jpg" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 464, + 99, + 473 + ], + "lines": [ + { + "bbox": [ + 70, + 464, + 99, + 473 + ], + "spans": [ + { + "bbox": [ + 70, + 464, + 99, + 473 + ], + "type": "text", + "content": "Building" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "lines": [ + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "type": "text", + "content": "Figure 1. Key discrepancies across the vehicle " + }, + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "type": "text", + "content": ", drone " + }, + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{d}})" + }, + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "type": "text", + "content": ", and quadruped " + }, + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{q}})" + }, + { + "bbox": [ + 55, + 478, + 555, + 522 + ], + "type": "text", + "content": " platforms. We compare the core attributes (viewpoint, speed, stability), event data distributions, and semantic distributions across event data acquired by three platforms, highlighting the challenges of adapting event camera perception to diverse operational contexts. These variations motivate the need for a robust cross-platform adaptation framework to harmonize event-based dense perception across distinct environmental setups and conditions." + } + ] + } + ], + "index": 124, + "angle": 0, + "type": "image_caption" + } + ], + "index": 86 + }, + { + "type": "image", + "bbox": [ + 113, + 435, + 162, + 464 + ], + "blocks": [ + { + "bbox": [ + 113, + 435, + 162, + 464 + ], + "lines": [ + { + "bbox": [ + 113, + 435, + 162, + 464 + ], + "spans": [ + { + "bbox": [ + 113, + 435, + 162, + 464 + ], + "type": "image", + "image_path": "6cb210a35f1456ccf925008373ce6704efc63026634adf41197d81f8d1ae31ab.jpg" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 464, + 151, + 472 + ], + "lines": [ + { + "bbox": [ + 124, + 464, + 151, + 472 + ], + "spans": [ + { + "bbox": [ + 124, + 464, + 151, + 472 + ], + "type": "text", + "content": "Terrain" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_caption" + } + ], + "index": 88 + }, + { + "type": "image", + "bbox": [ + 166, + 435, + 214, + 464 + ], + "blocks": [ + { + "bbox": [ + 166, + 435, + 214, + 464 + ], + "lines": [ + { + "bbox": [ + 166, + 435, + 214, + 464 + ], + "spans": [ + { + "bbox": [ + 166, + 435, + 214, + 464 + ], + "type": "image", + "image_path": "493289b3f99d0d0a4c9f16a4f22b068563b08cf7ec593c840b21dc9d9fc73f7f.jpg" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 464, + 209, + 473 + ], + "lines": [ + { + "bbox": [ + 170, + 464, + 209, + 473 + ], + "spans": [ + { + "bbox": [ + 170, + 464, + 209, + 473 + ], + "type": "text", + "content": "Vegetation" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_caption" + } + ], + "index": 90 + }, + { + "type": "image", + "bbox": [ + 228, + 325, + 276, + 371 + ], + "blocks": [ + { + "bbox": [ + 257, + 310, + 352, + 320 + ], + "lines": [ + { + "bbox": [ + 257, + 310, + 352, + 320 + ], + "spans": [ + { + "bbox": [ + 257, + 310, + 352, + 320 + ], + "type": "text", + "content": "Event Data Distribution" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 325, + 276, + 371 + ], + "lines": [ + { + "bbox": [ + 228, + 325, + 276, + 371 + ], + "spans": [ + { + "bbox": [ + 228, + 325, + 276, + 371 + ], + "type": "image", + "image_path": "bb89c73a093a17b69cc9e0b11e62f36afba2bc92f76038d905e9dcadf78f6428.jpg" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_body" + } + ], + "index": 93 + }, + { + "type": "image", + "bbox": [ + 282, + 325, + 328, + 372 + ], + "blocks": [ + { + "bbox": [ + 282, + 325, + 328, + 372 + ], + "lines": [ + { + "bbox": [ + 282, + 325, + 328, + 372 + ], + "spans": [ + { + "bbox": [ + 282, + 325, + 328, + 372 + ], + "type": "image", + "image_path": "0a06390f3d03a721f4c74b49185dffb5fe74c489558deadf7aacab7a52425152.jpg" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_body" + } + ], + "index": 94 + }, + { + "type": "image", + "bbox": [ + 334, + 325, + 381, + 372 + ], + "blocks": [ + { + "bbox": [ + 334, + 325, + 381, + 372 + ], + "lines": [ + { + "bbox": [ + 334, + 325, + 381, + 372 + ], + "spans": [ + { + "bbox": [ + 334, + 325, + 381, + 372 + ], + "type": "image", + "image_path": "bbf364fa6b5bfeba3b5aad3495ff5cb4c88761a1bdc4284d957e7eafeddf3915.jpg" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_body" + } + ], + "index": 95 + }, + { + "type": "image", + "bbox": [ + 228, + 394, + 276, + 422 + ], + "blocks": [ + { + "bbox": [ + 249, + 380, + 361, + 390 + ], + "lines": [ + { + "bbox": [ + 249, + 380, + 361, + 390 + ], + "spans": [ + { + "bbox": [ + 249, + 380, + 361, + 390 + ], + "type": "text", + "content": "Event Semantic Distribution" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 394, + 276, + 422 + ], + "lines": [ + { + "bbox": [ + 228, + 394, + 276, + 422 + ], + "spans": [ + { + "bbox": [ + 228, + 394, + 276, + 422 + ], + "type": "image", + "image_path": "9135fa7238a4726e06951fd84212a590fa40cf29ce9d01518912f78d92916ab8.jpg" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 423, + 263, + 431 + ], + "lines": [ + { + "bbox": [ + 242, + 423, + 263, + 431 + ], + "spans": [ + { + "bbox": [ + 242, + 423, + 263, + 431 + ], + "type": "text", + "content": "Road" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_caption" + } + ], + "index": 97 + }, + { + "type": "image", + "bbox": [ + 281, + 394, + 329, + 422 + ], + "blocks": [ + { + "bbox": [ + 281, + 394, + 329, + 422 + ], + "lines": [ + { + "bbox": [ + 281, + 394, + 329, + 422 + ], + "spans": [ + { + "bbox": [ + 281, + 394, + 329, + 422 + ], + "type": "image", + "image_path": "2b44515befee8a13e04a62f4467188f29a82e1439eabf2c8ac95419a830fd449.jpg" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 423, + 312, + 431 + ], + "lines": [ + { + "bbox": [ + 298, + 423, + 312, + 431 + ], + "spans": [ + { + "bbox": [ + 298, + 423, + 312, + 431 + ], + "type": "text", + "content": "Car" + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "image_caption" + } + ], + "index": 99 + }, + { + "type": "image", + "bbox": [ + 334, + 394, + 381, + 422 + ], + "blocks": [ + { + "bbox": [ + 334, + 394, + 381, + 422 + ], + "lines": [ + { + "bbox": [ + 334, + 394, + 381, + 422 + ], + "spans": [ + { + "bbox": [ + 334, + 394, + 381, + 422 + ], + "type": "image", + "image_path": "edca595c31516659e530990cc0418a03f96d8633d5cec85cc348d916a9ab8caf.jpg" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 423, + 365, + 432 + ], + "lines": [ + { + "bbox": [ + 351, + 423, + 365, + 432 + ], + "spans": [ + { + "bbox": [ + 351, + 423, + 365, + 432 + ], + "type": "text", + "content": "Sky" + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "image_caption" + } + ], + "index": 101 + }, + { + "type": "image", + "bbox": [ + 228, + 436, + 276, + 464 + ], + "blocks": [ + { + "bbox": [ + 228, + 436, + 276, + 464 + ], + "lines": [ + { + "bbox": [ + 228, + 436, + 276, + 464 + ], + "spans": [ + { + "bbox": [ + 228, + 436, + 276, + 464 + ], + "type": "image", + "image_path": "db50f4a6193090f94f116710b8425e58af58a833fe459ac3533b57770578957c.jpg" + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 465, + 267, + 473 + ], + "lines": [ + { + "bbox": [ + 237, + 465, + 267, + 473 + ], + "spans": [ + { + "bbox": [ + 237, + 465, + 267, + 473 + ], + "type": "text", + "content": "Building" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "image_caption" + } + ], + "index": 103 + }, + { + "type": "image", + "bbox": [ + 281, + 436, + 329, + 464 + ], + "blocks": [ + { + "bbox": [ + 281, + 436, + 329, + 464 + ], + "lines": [ + { + "bbox": [ + 281, + 436, + 329, + 464 + ], + "spans": [ + { + "bbox": [ + 281, + 436, + 329, + 464 + ], + "type": "image", + "image_path": "b0261deb16bbf0a1b8ee8af8f5fd5c8c60415dac3e4d7bf649fbe3cc084185f7.jpg" + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 465, + 318, + 472 + ], + "lines": [ + { + "bbox": [ + 291, + 465, + 318, + 472 + ], + "spans": [ + { + "bbox": [ + 291, + 465, + 318, + 472 + ], + "type": "text", + "content": "Terrain" + } + ] + } + ], + "index": 106, + "angle": 0, + "type": "image_caption" + } + ], + "index": 105 + }, + { + "type": "image", + "bbox": [ + 334, + 436, + 381, + 464 + ], + "blocks": [ + { + "bbox": [ + 334, + 436, + 381, + 464 + ], + "lines": [ + { + "bbox": [ + 334, + 436, + 381, + 464 + ], + "spans": [ + { + "bbox": [ + 334, + 436, + 381, + 464 + ], + "type": "image", + "image_path": "a70b7b37848f10af007f67bee10b431fc7d5030f810c66f51cfa9ac12d25dd19.jpg" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 465, + 377, + 473 + ], + "lines": [ + { + "bbox": [ + 337, + 465, + 377, + 473 + ], + "spans": [ + { + "bbox": [ + 337, + 465, + 377, + 473 + ], + "type": "text", + "content": "Vegetation" + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "image_caption" + } + ], + "index": 107 + }, + { + "type": "image", + "bbox": [ + 396, + 325, + 444, + 372 + ], + "blocks": [ + { + "bbox": [ + 396, + 325, + 444, + 372 + ], + "lines": [ + { + "bbox": [ + 396, + 325, + 444, + 372 + ], + "spans": [ + { + "bbox": [ + 396, + 325, + 444, + 372 + ], + "type": "image", + "image_path": "a52df0b455925d426b87f547e8ba74c4fa88c17b7c0975acafb160e8cae3422c.jpg" + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "image_body" + } + ], + "index": 110 + }, + { + "type": "image", + "bbox": [ + 449, + 325, + 496, + 372 + ], + "blocks": [ + { + "bbox": [ + 449, + 325, + 496, + 372 + ], + "lines": [ + { + "bbox": [ + 449, + 325, + 496, + 372 + ], + "spans": [ + { + "bbox": [ + 449, + 325, + 496, + 372 + ], + "type": "image", + "image_path": "ca5ab5c620c9479febd96acf155b9de61db38187a92d37697d33550cc215b4e2.jpg" + } + ] + } + ], + "index": 111, + "angle": 0, + "type": "image_body" + } + ], + "index": 111 + }, + { + "type": "image", + "bbox": [ + 501, + 324, + 549, + 372 + ], + "blocks": [ + { + "bbox": [ + 425, + 310, + 520, + 319 + ], + "lines": [ + { + "bbox": [ + 425, + 310, + 520, + 319 + ], + "spans": [ + { + "bbox": [ + 425, + 310, + 520, + 319 + ], + "type": "text", + "content": "Event Data Distribution" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 501, + 324, + 549, + 372 + ], + "lines": [ + { + "bbox": [ + 501, + 324, + 549, + 372 + ], + "spans": [ + { + "bbox": [ + 501, + 324, + 549, + 372 + ], + "type": "image", + "image_path": "b5915ff6a6a67ba116d7fbf2da64c3b83e811ab4b374950d5c9a07f810323f22.jpg" + } + ] + } + ], + "index": 112, + "angle": 0, + "type": "image_body" + } + ], + "index": 112 + }, + { + "type": "image", + "bbox": [ + 396, + 394, + 444, + 422 + ], + "blocks": [ + { + "bbox": [ + 417, + 380, + 529, + 390 + ], + "lines": [ + { + "bbox": [ + 417, + 380, + 529, + 390 + ], + "spans": [ + { + "bbox": [ + 417, + 380, + 529, + 390 + ], + "type": "text", + "content": "Event Semantic Distribution" + } + ] + } + ], + "index": 113, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 394, + 444, + 422 + ], + "lines": [ + { + "bbox": [ + 396, + 394, + 444, + 422 + ], + "spans": [ + { + "bbox": [ + 396, + 394, + 444, + 422 + ], + "type": "image", + "image_path": "ab513ef08af6e1b4c0c87a701de10c42a84ca4736b79ec3ebe737cca0bb45380.jpg" + } + ] + } + ], + "index": 114, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 423, + 430, + 432 + ], + "lines": [ + { + "bbox": [ + 410, + 423, + 430, + 432 + ], + "spans": [ + { + "bbox": [ + 410, + 423, + 430, + 432 + ], + "type": "text", + "content": "Road" + } + ] + } + ], + "index": 115, + "angle": 0, + "type": "image_caption" + } + ], + "index": 114 + }, + { + "type": "image", + "bbox": [ + 449, + 394, + 496, + 422 + ], + "blocks": [ + { + "bbox": [ + 449, + 394, + 496, + 422 + ], + "lines": [ + { + "bbox": [ + 449, + 394, + 496, + 422 + ], + "spans": [ + { + "bbox": [ + 449, + 394, + 496, + 422 + ], + "type": "image", + "image_path": "0a9dc4af66216b52316b7fc5b46a05e9155428b5bb702b6f6c60cc56dc981d9e.jpg" + } + ] + } + ], + "index": 116, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 423, + 480, + 431 + ], + "lines": [ + { + "bbox": [ + 466, + 423, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 466, + 423, + 480, + 431 + ], + "type": "text", + "content": "Car" + } + ] + } + ], + "index": 117, + "angle": 0, + "type": "image_caption" + } + ], + "index": 116 + }, + { + "type": "image", + "bbox": [ + 501, + 394, + 549, + 422 + ], + "blocks": [ + { + "bbox": [ + 501, + 394, + 549, + 422 + ], + "lines": [ + { + "bbox": [ + 501, + 394, + 549, + 422 + ], + "spans": [ + { + "bbox": [ + 501, + 394, + 549, + 422 + ], + "type": "image", + "image_path": "b96fd6e9c260adaad00800abbea46a130b08e63c24901c7f6de4802f276495d5.jpg" + } + ] + } + ], + "index": 118, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 423, + 533, + 433 + ], + "lines": [ + { + "bbox": [ + 518, + 423, + 533, + 433 + ], + "spans": [ + { + "bbox": [ + 518, + 423, + 533, + 433 + ], + "type": "text", + "content": "Sky" + } + ] + } + ], + "index": 119, + "angle": 0, + "type": "image_caption" + } + ], + "index": 118 + }, + { + "type": "image", + "bbox": [ + 449, + 436, + 497, + 464 + ], + "blocks": [ + { + "bbox": [ + 449, + 436, + 497, + 464 + ], + "lines": [ + { + "bbox": [ + 449, + 436, + 497, + 464 + ], + "spans": [ + { + "bbox": [ + 449, + 436, + 497, + 464 + ], + "type": "image", + "image_path": "0072c8948b2e6cb856f182f469503f64651ab3beea9922e71bb19d7e66599b4f.jpg" + } + ] + } + ], + "index": 120, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 464, + 487, + 473 + ], + "lines": [ + { + "bbox": [ + 459, + 464, + 487, + 473 + ], + "spans": [ + { + "bbox": [ + 459, + 464, + 487, + 473 + ], + "type": "text", + "content": "Terrain" + } + ] + } + ], + "index": 121, + "angle": 0, + "type": "image_caption" + } + ], + "index": 120 + }, + { + "type": "image", + "bbox": [ + 501, + 436, + 549, + 464 + ], + "blocks": [ + { + "bbox": [ + 501, + 436, + 549, + 464 + ], + "lines": [ + { + "bbox": [ + 501, + 436, + 549, + 464 + ], + "spans": [ + { + "bbox": [ + 501, + 436, + 549, + 464 + ], + "type": "image", + "image_path": "3532366df72d2686d0bdd7ff51635c8800d64c8ffb8f942901d368198dba5eac.jpg" + } + ] + } + ], + "index": 122, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 464, + 545, + 473 + ], + "lines": [ + { + "bbox": [ + 505, + 464, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 505, + 464, + 545, + 473 + ], + "type": "text", + "content": "Vegetation" + } + ] + } + ], + "index": 123, + "angle": 0, + "type": "image_caption" + } + ], + "index": 122 + }, + { + "bbox": [ + 152, + 533, + 200, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 533, + 200, + 545 + ], + "spans": [ + { + "bbox": [ + 152, + 533, + 200, + 545 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 125 + }, + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "text", + "content": "Cross-platform adaptation in event-based dense perception is crucial for deploying event cameras across diverse settings, such as vehicles, drones, and quadrupeds, each with unique motion dynamics, viewpoints, and class distributions. In this work, we introduce EventFly, a framework for robust cross-platform adaptation in event camera perception. Our approach comprises three key components: i) Event Activation Prior (EAP), which identifies high-activation regions in the target domain to minimize prediction entropy, fostering confident, domain-adaptive predictions; ii) EventBlend, a data-mixing strategy that integrates source and target event voxel grids based on EAP-driven similarity and density maps, enhancing feature alignment;" + } + ] + } + ], + "index": 126 + }, + { + "bbox": [ + 313, + 534, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 654 + ], + "type": "text", + "content": "and iii) EventMatch, a dual-discriminator technique that aligns features from source, target, and blended domains for better domain-invariant learning. To holistically assess cross-platform adaptation abilities, we introduce EXPo, a large-scale benchmark with diverse samples across vehicle, drone, and quadruped platforms. Extensive experiments validate our effectiveness, demonstrating substantial gains over popular adaptation methods. We hope this work can pave the way for more adaptive, high-performing event perception across diverse and complex environments." + } + ] + } + ], + "index": 127 + }, + { + "bbox": [ + 314, + 667, + 394, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 667, + 394, + 680 + ], + "spans": [ + { + "bbox": [ + 314, + 667, + 394, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 128 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Event cameras, with their asynchronous operation and high temporal resolution, provide great advantages over conven-" + } + ] + } + ], + "index": 129 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1472" + } + ] + } + ], + "index": 130 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "tional frame-based sensors by capturing precise pixel intensity changes as they occur [24, 26, 43, 44, 71]. These attributes make event cameras ideally suited for real-time applications in dynamic environments, such as autonomous driving, aerial navigation, and robotic perception [56, 74]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 266 + ], + "type": "text", + "content": "Despite this potential, event cameras remain primarily deployed on the vehicle platform [7, 9]. Based on a recent survey study [11], 12 out of 15 public event camera perception datasets are collected exclusively from ground vehicles [2, 13, 29, 64, 93, 100]. In contrast, datasets involving alternative robot platforms, such as drones [4, 16, 85] or mobile robots [5, 95, 102], are limited, making robust event-based perception across diverse domains challenging [66, 70]. To enable versatile applications, a cross-platform approach is vital for adapting the perception models to the new, real-world, and challenging operational contexts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 470 + ], + "type": "text", + "content": "Cross-platform adaptation for event-based dense perception presents significant technical challenges. As can be seen from Fig. 1, each robot platform - vehicle, drone, or quadruped - exhibits unique motion patterns, environmental interactions, and perspective-based dynamics that create distinct activation patterns in the event data [22, 53, 65, 73]. For example, ground vehicles capture event activity near surfaces [26, 29, 100], while drones tend to capture high-altitude landscape features with sparse ground details [12, 42, 52, 56, 80]. Conventional domain adaptation methods commonly used for frame-based sensors are not well-suited to handle the distinct spatial-temporal nuances of event camera data, which demand tailored alignment strategies [48, 50, 60, 78]. A specialized framework for cross-platform adaptation in event-based dense perception is, therefore, crucial for reliable, context-sensitive perception across varied robot sensing conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 555 + ], + "type": "text", + "content": "In this work, we introduce EventFly, a robust framework for cross-platform adaptation in event camera dense perception. We bridge domain-specific gaps in event-camera perception by leveraging platform-specific activation patterns and aligning feature distributions across domains. Our design consists of three core components, each addressing a specific challenge in cross-platform event data alignment:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 556, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 652 + ], + "type": "text", + "content": "■ Event Activation Prior (EAP). Event cameras on different platforms reveal distinctive high-activation regions shaped by platform-specific dynamics. EAP leverages these by identifying frequently activated areas in the target domain and minimizing prediction entropy within these regions. This promotes confident, domain-adaptive predictions that align the model to platform-specific event patterns, enhancing performance in target-relevant contexts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Cross-Platform Consistency Regularization. To bridge the gaps in-between different robot platforms, we propose EventBlend to create hybrid event voxel grids by combining source and target event data in a spatially structured way. Guided by EAP-driven similarity and density maps, this" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "method selectively integrates features based on shared activation patterns, focusing on areas where platform-specific details overlap. This fusion retains platform-specific cues while enhancing cross-domain feature consistency, balancing reliable source information with context-sensitive target features for more robust cross-platform adaptations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 263 + ], + "type": "text", + "content": "Cross-Platform Adversarial Training. To further align domain distributions, we propose EventMatch, a dual-discriminator approach across source, target, and blended representations. One discriminator enforces alignment between source and blended domains, while the other softly adapts blended features toward the target, particularly in high-activation regions. This layered approach supports robust, domain-adaptive learning that generalizes well across platforms, balancing domain-invariant features with target-specific adaptations for improved perception accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 264, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 407 + ], + "type": "text", + "content": "To rigorously assess the cross-platform adaptation performance, we establish EXPo, a large-scale Event-based Xcross-Platform dense perception benchmark with around 90k event data from three platforms: vehicle, drone, and quadruped. From extensive benchmark experiments, our approach consistently demonstrates robust adaptation capabilities, achieving on average " + }, + { + "bbox": [ + 313, + 264, + 553, + 407 + ], + "type": "inline_equation", + "content": "23.8\\%" + }, + { + "bbox": [ + 313, + 264, + 553, + 407 + ], + "type": "text", + "content": " higher accuracy and " + }, + { + "bbox": [ + 313, + 264, + 553, + 407 + ], + "type": "inline_equation", + "content": "77.1\\%" + }, + { + "bbox": [ + 313, + 264, + 553, + 407 + ], + "type": "text", + "content": " better mIoU across platforms compared to source-only training. When evaluated against prior adaptation methods, we outperform by significant margins across almost all semantic classes, highlighting the scalability and effectiveness in diverse, challenging operational contexts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 408, + 533, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 408, + 533, + 418 + ], + "spans": [ + { + "bbox": [ + 326, + 408, + 533, + 418 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 419, + 553, + 598 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 419, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 419, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 419, + 553, + 479 + ], + "type": "text", + "content": "- We propose EventFly, a novel framework designed for cross-platform adaptation in event camera perception, facilitating robustness under diverse event data dynamics. To our knowledge, this is the first work proposed to address this critical gap in event-based perception tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "type": "text", + "content": "- We introduce Event Activation Prior (EAP), EventBlend, and EventMatch - a set of tailored techniques that utilize platform-specific activation patterns, spatial data mixing, and dual-domain feature alignment to tackle the unique challenges of event-based cross-platform adaptation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 539, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 553, + 598 + ], + "type": "text", + "content": "- We establish a large-scale benchmark, EXPo, for cross-platform adaptation in event-based perception, comprising rich collections of samples from vehicle, drone, and quadruped domains, providing a strong foundation for further research in adaptive event camera applications." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 609, + 400, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 400, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 400, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Event Camera Perception. Recent works harnessed event cameras for diverse perception tasks in dynamic settings, such as detection [25, 26, 28, 55, 105], segmentation [1, 32, 47, 75], depth estimation [15, 30, 34, 58, 99, 101], and visual odometry [10, 18, 35, 59]. Object detection with event cameras involves locating objects by leveraging the high temporal resolution of event streams, enabling" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1473" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "effective scene parsing even under challenging conditions [49, 63, 98, 103]. Meanwhile, semantic segmentation aims to enhance scene understanding, which is vital for safe navigation in autonomous systems [15, 45, 57, 83]. Additionally, recent research has focused on integrating event data with complementary modalities to improve accuracy and mitigate the sparse nature of event data [3, 14, 27, 33, 84]. Our work extends this line by focusing on robust cross-platform adaptation, ensuring that event-based perception models can generalize across different robotic platforms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 357 + ], + "type": "text", + "content": "Cross-Domain Adaptation. Transferring knowledge from a labeled source domain to an unlabeled target domain addresses the challenge of limited annotated data in diverse real-world cases [69, 96]. Traditional methods in this area leverage strategies like domain adversarial training [79], entropy minimization [61, 68, 81, 82], contrastive learning [37, 87, 92], domain mixing [46, 51, 77, 97], and self-training [8, 38, 94, 104]. However, despite promising performances, these approaches are primarily designed for frame-based data [17, 36, 91], which lacks the unique spatiotemporal properties of the stream-based event camera data. Our approach adapts these domain adaptation principles specifically for event-based vision perception, addressing the distinct challenges posed by event camera data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 539 + ], + "type": "text", + "content": "Domain Adaptation from Event Camera. Recent works have begun to address domain shifts for event-based perception. Approaches like Ev-Transfer [57], ESS [75], and HPL-ESS [40] focus on transferring knowledge from RGB frames to event data, enabling effective event-based perception through frame-to-event domain adaptation. Other studies leverage event cameras to aid adaptation in low-light or night-time conditions for conventional RGB sensors [21, 39, 86, 89]. There are also works that explored efficient learning utilizing shared representations across modalities [20, 41, 88]. To the best of our knowledge, our work is the first to tackle cross-platform adaptation across multiple event camera platforms, such as vehicles, drones, and quadrapeds, focusing on the unique motion and environmental interactions inherent to each platform." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 549, + 138, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 138, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 138, + 562 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "text", + "content": "This work tackles the problem of cross-platform adaptation for event-based dense perception across three distinct platforms: vehicle, drone, and quadruped, denoted as " + }, + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{q}}" + }, + { + "bbox": [ + 55, + 569, + 294, + 653 + ], + "type": "text", + "content": ", respectively. Each platform exhibits unique spatial and temporal characteristics in the captured event data, influenced by platform-specific perspectives, motion patterns, semantics, and environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 660, + 141, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 141, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 141, + 672 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Event cameras capture continuous, asynchronous streams [6], where each event " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i = (e_i^x,e_i^y,e_i^t,e_i^p)" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": " consists of pixel coordinates " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "(e_i^x,e_i^y)" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", timestamp " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "e_i^t" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", and polarity " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "e_i^p\\in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\{-1, + 1\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": ", indicating either a decrease or increase in brightness. This format allows event cameras to capture high-speed motion and scene changes in dynamic environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "text", + "content": "Event Data Representation. To efficiently process event data, we convert raw events into voxel grid representations, denoted as " + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "text", + "content": ". This structured format regularizes the event data, making it compatible with learning-based methods and consistent across different domains. To construct each voxel grid " + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i\\in \\mathbb{R}^{T\\times H\\times W}" + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "text", + "content": " is the number of temporal bins, and " + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 108, + 555, + 252 + ], + "type": "text", + "content": " are the spatial dimensions of the event sensor, we divide the event stream into fixed time intervals, with positive and negative events accumulated in separate bins along the temporal dimension. This polarity-sensitive encoding captures spatial and temporal dynamics of events that reflect both motion directionality and density:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 260, + 555, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 260, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 318, + 260, + 555, + 285 + ], + "type": "interline_equation", + "content": "\\sum_ {\\mathbf {e} _ {j} \\in \\varepsilon_ {i}} e _ {j} ^ {p} \\delta \\left(e _ {j} ^ {x} - e ^ {x}\\right) \\delta \\left(e _ {j} ^ {y} - e ^ {y}\\right) \\max \\left\\{1 - \\left| \\hat {e} _ {j} ^ {t} - e ^ {t} \\right|, 0 \\right\\}, \\tag {1}", + "image_path": "d3f79e0f35f8e950339dd4b160e9126da36a35fff027002ff4087e897ff664f3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "text", + "content": " is the Kronecker delta function; " + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "inline_equation", + "content": "\\hat{e}_j^t = (C - 1)\\frac{e_j^t - e_0^t}{\\Delta T}" + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "text", + "content": " normalizes the event timestamps. Here, " + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "inline_equation", + "content": "\\Delta T" + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "text", + "content": " is the duration of each time window, and " + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "inline_equation", + "content": "e_0^t" + }, + { + "bbox": [ + 313, + 289, + 553, + 329 + ], + "type": "text", + "content": " is the initial timestamp." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "content": "Cross-Platform Adaptation. We aim to address the challenge of adapting across three distinct domains, represented by " + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathcal{P}^{\\mathrm{v}},\\mathcal{P}^{\\mathrm{d}},\\mathcal{P}^{\\mathrm{q}}\\}" + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "content": ", respectively. Each domain consists of event voxel grids " + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "content": " and associated labels " + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "inline_equation", + "content": "y\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "content": ", where each pixel label belongs to one of " + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "content": " semantic classes. In our cross-platform adaptation setting, we assume access to fully labeled data from the source platform while the target platform has only unlabeled data. The goal is to leverage the labeled source data and unlabeled target data to improve event-based perception performance on the target domain." + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "inline_equation", + "content": "^1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 456, + 443, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 443, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 443, + 468 + ], + "type": "text", + "content": "3.2. Event Activation Prior" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "type": "text", + "content": "Cross-platform adaptation in event-based perception hinges on aligning platform-specific activation patterns, which are influenced by domain-unique motions, and environments. We introduce Event Activation Prior (EAP), a regularization technique that encourages confident, low-entropy predictions in target-domain regions with high event activation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": "In each platform, event data often exhibits recurrent activation patterns in specific regions, driven by platform-specific motion and scene dynamics (see Fig. 1). For instance, lower regions in a vehicle domain capture details like road surfaces and obstacles, while upper regions in a drone domain capture landscape features and environmental context. Such regions, denoted as " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\subset \\{0,1,\\dots,H-1\\} \\times \\{0,1,\\dots,W-1\\}" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": " and along with the conditional entropy " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "H(y_{\\mathbf{S}}|\\mathbf{V}_{\\mathbf{S}},\\mathbf{S})" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": ", typically have consistent semantic patterns, allowing for more confident predictions. By leveraging these high-activation areas, we aim to minimize the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "Without loss of generality, our subsequent explanations focus on the adaptation from vehicle domain " + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{v}}" + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": " to drone domain " + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": ", although the proposed framework generalizes to any pair of domains in " + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1474" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 541, + 312 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 312 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 312 + ], + "type": "image", + "image_path": "2a27ad1640cc13ebb94d1818ddcae1cdd246e138d3ba21c7319041f6b6e8f1ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. Overview of the EventFly framework. Guided by the EAP principle (Sec. 3.2), the pair of source and target event data " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i^{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": " are mixed via the EventBlend operation (Sec. 3.3), where the blending mask " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": " is obtained by measuring the similarities between density maps " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i^{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{D}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": ". The features " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 314, + 555, + 361 + ], + "type": "text", + "content": " from the source, target, and blended domains are then used for EventMatch (Sec. 3.4). This facilitates learning an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 409 + ], + "type": "text", + "content": "entropy of predictions in such regions, thereby guiding the model to produce more confident outputs aligned with the unique characteristics of the target platform domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "text", + "content": "Formulating EAP. To incorporate EAP in training, we define it as a regularization that minimizes prediction uncertainty in these high-activation event areas. For each subregion S, EAP aims to minimize the conditional entropy " + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "inline_equation", + "content": "H(y_{\\mathbf{S}}|\\mathbf{V}_{\\mathbf{S}},\\mathbf{S})" + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "text", + "content": ", focusing on domain-specific high-activation regions. We enforce the expectation constraint over the model parameters " + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\theta}[H(\\mathbf{V}_{\\mathbf{S}},y_{\\mathbf{S}}|\\mathbf{S})]\\leq c" + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "text", + "content": " is a small constant that encourages high-confidence predictions. This is transformed into a prior on parameters " + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 410, + 295, + 530 + ], + "type": "text", + "content": " through the maximum entropy principles [31], yielding:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "type": "interline_equation", + "content": "P (\\theta) \\propto \\exp (- \\lambda H (\\mathbf {V} _ {\\mathbf {S}}, y _ {\\mathbf {S}} | \\mathbf {S})) \\propto \\exp (- \\lambda H (y _ {\\mathbf {S}} | \\mathbf {V} _ {\\mathbf {S}}, \\mathbf {S})), \\tag {2}", + "image_path": "339f80244914e711759b69a0e69d5088ba504389b9c3bb09457a3fad2b94e846.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 560, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 295, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 560, + 295, + 584 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 55, + 560, + 295, + 584 + ], + "type": "text", + "content": " is the Lagrange multiplier corresponding to constant " + }, + { + "bbox": [ + 55, + 560, + 295, + 584 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 560, + 295, + 584 + ], + "type": "text", + "content": ", which balances the effect of EAP on training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 584, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 295, + 645 + ], + "type": "text", + "content": "Empirical Estimation of EAP. To apply EAP, we empirically estimate the entropy in high-activation event regions by aggregating over activations in the target platform domain, where activation patterns are highly informative. Using an empirical estimator, we approximate " + }, + { + "bbox": [ + 55, + 584, + 295, + 645 + ], + "type": "inline_equation", + "content": "H(y|\\mathbf{V},\\mathbf{S})" + }, + { + "bbox": [ + 55, + 584, + 295, + 645 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 650, + 294, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 650, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 64, + 650, + 294, + 678 + ], + "type": "interline_equation", + "content": "H _ {\\text {e m p}} (y | \\mathbf {V}, \\mathbf {S}) = \\mathbb {E} _ {\\mathbf {V}, y, \\mathbf {S}} \\left[ \\hat {P} (y | \\mathbf {V}, \\mathbf {S}) \\log \\hat {P} (y | \\mathbf {V}, \\mathbf {S}) \\right], \\tag {3}", + "image_path": "ee664e4a3aaaf2f354c227145d566ccd309461609a21c66638e8e5876b61d68a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{P}(y|\\mathbf{V},\\mathbf{S})" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the empirical prediction probability conditioned on the voxel grid " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and restricted to region " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". By minimizing " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{emp}}(y|\\mathbf{V},\\mathbf{S})" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we encourage confident" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 373, + 555, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 398 + ], + "type": "text", + "content": "predictions within these regions, aligning the model's predictions with the target domain activation patterns." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 398, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 434 + ], + "type": "text", + "content": "Integrating EAP into Training. To incorporate the EAP in Eq. (2) into cross-platform adaptation, we define the overall objective as a maximum-a-posteriori (MAP) estimation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 441, + 553, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 441, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 365, + 441, + 553, + 455 + ], + "type": "interline_equation", + "content": "C (\\theta) = \\mathcal {L} (\\theta) - \\lambda H _ {\\mathrm {e m p}} (y | \\mathbf {V}, \\mathbf {S}), \\tag {4}", + "image_path": "c62057835d417473796185572c95a0aceb5698830784eea6dba71008ff2be723.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta)" + }, + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "text", + "content": " represents the supervised loss on source data. " + }, + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{emp}}(y|\\mathbf{V},\\mathbf{S})" + }, + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "text", + "content": " minimizes uncertainty in the target domain by leveraging EAP over high-activation regions. By focusing on these regions, the event camera perception model tends to effectively learn to adapt its predictions to align with platform-specific activation cues in the target domain, achieving robust adaptation across varied platforms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 391, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 391, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 391, + 564 + ], + "type": "text", + "content": "3.3.EventBlend" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": "Cross-platform adaptation in event-based perception requires a selective integration of data from both source and target domains. Here, we introduce EventBlend, a data-mixing strategy designed to construct hybrid voxel grids by blending event data from the source and target domains, guided by high-activation regions identified through EAP. As shown in Fig. 2, by constructing these blended voxel grids, we aim to enhance the generalization ability across domain shifts, enabling the model to adapt effectively to the target distribution while leveraging reliable source labels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Event Density Maps. To identify regions where the source and target domains exhibit similar or divergent activation" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1475" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "patterns, we employ density maps, which highlight areas of frequent event activity. For each source event data with voxel grid " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i^{\\mathbf{v}}\\in \\mathbb{R}^{T\\times H\\times W}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", we calculate a corresponding source density map, " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i^{\\mathbf{v}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", by summing activations along the temporal axis " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " for each spatial location " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "(\\mu ,\\nu)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 138, + 295, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 138, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 104, + 138, + 295, + 159 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {i} ^ {\\mathbf {v}} (\\mu , \\nu) = \\sum_ {t = 1} ^ {T} | \\mathbf {V} _ {i} ^ {\\mathbf {v}} (t, \\mu , \\nu) |. \\tag {5}", + "image_path": "03142bd497afd84b7fdfacbd4c548948afcf6e876522ccb5af7c1ebf1599eac7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "text", + "content": "For the target domain, we pre-compute an aggregated density map, " + }, + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{D}}^{\\mathrm{d}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "text", + "content": ", by summing activations over all target samples " + }, + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_j^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 161, + 296, + 199 + ], + "type": "text", + "content": " indexes each target sample:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 204, + 295, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 204, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 83, + 204, + 295, + 228 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {D}} ^ {\\mathrm {d}} (\\mu , \\nu) = \\frac {1}{N ^ {\\mathrm {d}}} \\sum_ {j} \\sum_ {t = 1} ^ {T} \\left| \\mathbf {V} _ {j} ^ {\\mathrm {d}} (t, \\mu , \\nu) \\right|, \\tag {6}", + "image_path": "229e79aa32749e4b3dee55420e78f7d7fef3287a090a8e95fe2fc63687ba4ab8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": " denotes the total number of samples in the target domain. These density maps highlight regions with significant activity, which aids in identifying areas where the domains may share similar activation patterns, consistent with the high-activation regions established by EAP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 291, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 339 + ], + "type": "text", + "content": "Similarity Map & Binary Mask. To guide selective blending, for each source sample, we construct a similarity map " + }, + { + "bbox": [ + 55, + 291, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{SIM}_i\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 291, + 296, + 339 + ], + "type": "text", + "content": " to quantify spatial overlap in activation patterns between two domains. This map is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 344, + 295, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 344, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 84, + 344, + 295, + 359 + ], + "type": "interline_equation", + "content": "\\operatorname {S I M} _ {i} (\\mu , \\nu) = 1 - \\left| \\mathbf {D} _ {i} ^ {\\mathbf {v}} (\\mu , \\nu) - \\tilde {\\mathbf {D}} ^ {\\mathrm {d}} (\\mu , \\nu) \\right|. \\tag {7}", + "image_path": "e06c1519952b4fad4392ac995d08e98e19932a376c7be0c972c27a132902c3f8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "text", + "content": "Here, high values in " + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{SIM}_i" + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "text", + "content": " correspond to regions where source and target domains exhibit similar activation intensities, aligning with the high-activation areas specified by EAP in Eq. (2). Next, we derive a binary mask " + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i\\in" + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "text", + "content": " by applying a threshold " + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "text", + "content": " to the similarity map:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 427, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 427, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 91, + 427, + 295, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {i} (\\mu , \\nu) = \\left\\{ \\begin{array}{l l} \\mathbf {1}, & \\text {i f} \\mathbf {S I M} _ {i} (\\mu , \\nu) \\geq \\tau , \\\\ \\mathbf {0}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}", + "image_path": "ea34066924fa5a3be72e04294d95f6873e0cdd61d8977f9b69c61ddde572d2d9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "content": "This mask directs the blending process by specifying which regions to retain from the source and which to adapt from the target. A mask value of 1 indicates that activations in " + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "inline_equation", + "content": "(\\mu, \\nu)" + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "content": " from the source domain should be retained, while a value of 0 specifies a shift to the target domain activations. It is worth mentioning that for the selections of " + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i^{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{D}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "content": " in Eq. (5) and Eq. (6), we only consider activated pixel locations. That is to say, the similarity score does not count cases where both source and target are non-activated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": "Constructing Blended Event Voxel Grids. For a pair of source and target samples " + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i^{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": ", we create a new event voxel grid " + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_i \\in \\mathbb{R}^{T \\times H \\times W}" + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": ". For each spatial location " + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "inline_equation", + "content": "(\\mu, \\nu)" + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": ", temporal sequences are selectively copied from either source and target based on " + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i(\\mu, \\nu)" + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 639, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 639, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 70, + 639, + 295, + 670 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {V}} _ {i} (:, \\mu , \\nu) = \\left\\{ \\begin{array}{l l} \\mathbf {V} _ {i} ^ {\\mathbf {v}} (:, \\mu , \\nu), & \\text {i f} \\mathcal {M} (\\mu , \\nu) = \\mathbf {1}, \\\\ \\mathbf {V} _ {i} ^ {\\mathrm {d}} (:, \\mu , \\nu), & \\text {i f} \\mathcal {M} (\\mu , \\nu) = \\mathbf {0}. \\end{array} \\right. \\tag {9}", + "image_path": "230247f0080b9f3c6c3bb1531503c585a36b1c06386320c204b9d347af0b947b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "This selective copying approach allows " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " to incorporate source-domain stability (i.e., human annotations) in high-similarity regions while adapting to target-specific patterns" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "in low-similarity areas. Finally, we generate the label map " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " by combining ground truth labels from the source and pseudo-labels from the target. Here, the pseudo-labels can be obtained offline from a source-pretrained model or generated online using a mean teacher framework [76]. During adaptation, we use " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " for supervised learning. The blended data contains activations and labels from both domains. We leverage EAP-aligned high-activation areas, i.e., Eq. (4), where source-target consistency is high, facilitating robust cross-domain adaptation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 200, + 395, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 200, + 395, + 211 + ], + "spans": [ + { + "bbox": [ + 314, + 200, + 395, + 211 + ], + "type": "text", + "content": "3.4.EventMatch" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 217, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 326 + ], + "type": "text", + "content": "In cross-platform adaptation, domain shifts manifest as discrepancies in spatial activation patterns, motion characteristics, and perspectives unique to each platform. The EAP (cf. Sec. 3.2) identifies high-activation areas for stable adaptation, while EventBlend (cf. Sec. 3.3) generates hybrid voxel grids that capture shared spatial structures across domains. However, without explicit feature alignment, the model might still struggle to fully adapt to platform-specific idiosyncrasies, especially in target-specific contexts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "text", + "content": "To address this, we propose the EventMatch technique to align the blended domain as an intermediary, enabling robust domain-invariant feature learning by maintaining source reliability while selectively adapting to target-specific characteristics. We introduce two fully convolutional discriminators, " + }, + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "text", + "content": ", each responsible for a distinct aspect of feature alignment:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 410, + 554, + 564 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "text", + "content": "- Source & Blended (via " + }, + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "text", + "content": "): We align the blended voxel grids " + }, + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}" + }, + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "text", + "content": " with the source domain, encouraging them to retain reliable source-domain characteristics. This alignment preserves robustness and ensures that the blended features maintain patterns learned from the ground truth supervised labels, i.e., " + }, + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 314, + 410, + 554, + 481 + ], + "type": "text", + "content": ", from the source domain." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 481, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 554, + 564 + ], + "type": "text", + "content": "- Blended & Target (via " + }, + { + "bbox": [ + 314, + 481, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 314, + 481, + 554, + 564 + ], + "type": "text", + "content": "): Rather than the strict alignment, we regularize the blended features toward a soft alignment with the target, selectively adapting target-relevant characteristics. This alignment is focused on high-activation regions identified by EAP, allowing the blended features to bridge the domain gap in target-specific contexts without sacrificing source consistency." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 565, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 554, + 601 + ], + "type": "text", + "content": "This dual-discriminator setup allows the blended features to act as an intermediary representation that is both robust (source-aligned) and adaptable (target-sensitive)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "text", + "content": "Domain Adversarial Training. Each discriminator learns to distinguish between the feature representations from its respective domains, guiding the model to produce domain-agnostic features. Given feature outputs " + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{v}}" + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "text", + "content": " for the source, target, and blended domains, respectively:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "spans": [ + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "text", + "content": "- Source-Blended Alignment: Discriminator " + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "text", + "content": " classifies " + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\vee}" + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "text", + "content": " as source and " + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "text", + "content": " as blended, maximizing " + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\sigma_1}" + }, + { + "bbox": [ + 314, + 662, + 554, + 686 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 331, + 694, + 554, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 694, + 554, + 717 + ], + "spans": [ + { + "bbox": [ + 331, + 694, + 554, + 717 + ], + "type": "interline_equation", + "content": "- \\sum_ {\\mu , \\nu} \\left[ \\left(1 - z _ {1}\\right) \\log \\sigma_ {1} (\\tilde {\\mathbf {F}}) + z _ {1} \\log \\sigma_ {1} \\left(\\mathbf {F} ^ {\\mathrm {v}}\\right) \\right], \\tag {10}", + "image_path": "6f3b016d5c29ddf63443b7e6748671ce20efebe90efca127827a3645ab045db4.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1476" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 97, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Benchmark results of platform adaptation from vehicle " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " to drone " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{d}})" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": ". Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are highlighted in colors." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 97, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 97, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 97, + 555, + 232 + ], + "type": "table", + "html": "
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o43.6933.8115.0411.8148.7111.570.928.4213.3325.488.1831.5114.880.042.41
AdaptSegNet [79]49.1435.3821.1612.1529.3723.570.170.4813.4538.2317.8548.7329.4235.550.40
CBST [104]57.9541.1824.3116.0233.0524.430.003.0818.2456.3216.8456.1523.6135.650.00
IntraDA [61]57.3738.8523.5815.9132.3123.170.004.9014.9156.7018.6754.9420.7133.080.00
DACS [77]59.8142.0127.0716.1435.1626.120.184.1118.4955.6421.7456.8134.6944.730.05
MIC [38]63.1145.6028.8717.4641.4025.190.0110.1122.8659.2520.8458.8633.9544.180.90
PLSR [94]64.6145.9329.6917.9942.0930.060.009.7523.3262.4820.6560.1531.6944.272.06
EventFly (Ours)69.1748.2032.6720.0146.6430.551.2710.9125.5067.1724.2161.0141.3044.546.21
Target •79.5752.2542.9023.3074.4839.407.100.3331.6771.9631.6467.8757.5166.1423.79
", + "image_path": "00399096352dbfce608a591df080634c69df823a9b637a5d9149f782f9b5bcbd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 247, + 294, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 247, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 63, + 247, + 294, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 247, + 294, + 282 + ], + "type": "inline_equation", + "content": "z_{1} = 1" + }, + { + "bbox": [ + 63, + 247, + 294, + 282 + ], + "type": "text", + "content": " for source and " + }, + { + "bbox": [ + 63, + 247, + 294, + 282 + ], + "type": "inline_equation", + "content": "z_{1} = 0" + }, + { + "bbox": [ + 63, + 247, + 294, + 282 + ], + "type": "text", + "content": " for blended. This alignment ensures that the blended features preserve the supervised stability from the source." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "text", + "content": "- Blended-Target Adaptation: Discriminator " + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "text", + "content": " adapts " + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "text", + "content": " toward high-activation regions in the target, where target-relevant properties are critical. It classifies " + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "text", + "content": " as blended and " + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "text", + "content": " as target, maximizing " + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\sigma_2}" + }, + { + "bbox": [ + 55, + 283, + 295, + 332 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 338, + 294, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 338, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 72, + 338, + 294, + 360 + ], + "type": "interline_equation", + "content": "- \\sum_ {\\mu , \\nu} \\left[ \\left(1 - z _ {2}\\right) \\log \\sigma_ {2} \\left(\\mathbf {F} ^ {\\mathrm {d}}\\right) + z _ {2} \\log \\sigma_ {2} (\\tilde {\\mathbf {F}}) \\right], \\tag {11}", + "image_path": "f1edf98e6ccc26bd0fe3a3169d757065dcfd8a254b3272cdd60efb9990bd9b69.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "inline_equation", + "content": "z_{2} = 1" + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "text", + "content": " for target and " + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "inline_equation", + "content": "z_{2} = 0" + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "text", + "content": " for blended. Here, " + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "text", + "content": " does not force complete target alignment but rather emphasizes adaptation in target-relevant areas, enabling " + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 63, + 367, + 295, + 415 + ], + "type": "text", + "content": " to maintain a balanced representation across domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": "Optimization. To achieve consistent cross-domain alignment, the backbone network is trained adversarially against both discriminators. The goal is to produce " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " that each discriminator finds challenging to classify distinctly, thereby promoting a domain-agnostic representation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 481, + 295, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 481, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 77, + 481, + 295, + 504 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d v}} = - \\sum_ {\\mu , \\nu} \\left[ \\log \\sigma_ {1} (\\tilde {\\mathbf {F}}) + \\log \\sigma_ {2} (\\tilde {\\mathbf {F}}) \\right]. \\tag {12}", + "image_path": "e37eca7ecc22437e152886f100988676b5407131bdfcd2eefa5b5e30ff379058.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": "This objective ensures that the backbone network produces features for blended data that capture both source robustness and target adaptability, guided by high-activation areas where source-target consistency is most beneficial. To achieve this goal, we set two weight coefficients, " + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": ", to balance between these two objectives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "content": "During training, we employ a min-max optimization framework to iteratively update the backbone, " + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 606, + 295, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 606, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 629 + ], + "type": "text", + "content": "- Step 1: Update " + }, + { + "bbox": [ + 55, + 606, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 55, + 606, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 606, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 55, + 606, + 295, + 629 + ], + "type": "text", + "content": " to improve the classification of the source, blended, and target features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 665 + ], + "type": "text", + "content": "- Step 2: Update the backbone network to minimize " + }, + { + "bbox": [ + 55, + 630, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 630, + 295, + 665 + ], + "type": "text", + "content": ", encouraging it to generate intermediary features that support robust cross-domain generalization." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Through this adversarial process, EventMatch allows the model to bridge the domain gap effectively, with blended features serving as a reliable intermediary that balances source stability with target-specific adaptability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "type": "text", + "content": "Overall Framework. As shown in Fig. 2, the source and blended domains share the same network, while the network from the target domain is updated via exponential-moving average (EMA), encouraging consistency among domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 302, + 443, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 443, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 443, + 313 + ], + "type": "text", + "content": "3.5. The EXPo Benchmark" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 319, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 451 + ], + "type": "text", + "content": "To facilitate robust event camera perception across diverse platforms and environments, we establish a large-scale event-based cross-platform adaptation benchmark: EXPo. Our benchmark is based on [12]. We include 89, 228 frames captured from three distinct platforms - vehicle, drone, and quadruped - spanning 21 sequences: 6 from the vehicle, 7 from the drone, and 8 from the quadruped. Following DSEC-Semantic [75], we support 11 semantic classes relevant to real-world event-based perception and cover varied environments including city, urban, suburban, and rural scenes. Due to space limits, more details are in Appendix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 461, + 394, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 394, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 394, + 475 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 480, + 374, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 374, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 374, + 493 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": "Implementation Details. Our framework is implemented using PyTorch [62]. The backbone network is from E2VID [67], while the event segmentation head is from ESS [75]. The hyperparameters " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": " are set to 0.4, 1e-3, and 2e-3, respectively. The model is optimized using AdamW [54] and OneCycle learning rate scheduler [72] for 100k iterations. The learning rate and batch size are set to 1e-3 and 8. Due to the lack of baselines, we reproduce popular frame-based adaptation methods for adversarial training, contrastive learning, and self-training [38, 61, 77, 79, 94, 104]. To ensure fair comparisons, all methods adopt the same backbone and training iterations. The models are trained using NVIDIA RTX 4090 GPUs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. Following conventions, we use Accuracy (Acc) and Intersection-over-Union (IoU) as the main metrics for evaluation. We also report mean Acc (mAcc), mean IoU (mIoU), and frequency-weighted IoU (fIoU) for a holistic comparison among different adaptation methods." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1477" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 99, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 2. Benchmark results of platform adaptation from vehicle " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": " to quadruped " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{q}})" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": ". Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are in colors." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 99, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 99, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 99, + 555, + 232 + ], + "type": "table", + "html": "
MethodAccmAccmIoUfIoUgroundbuildfencepersonpoleroadwalkvegcarwallsign
Source-Only o66.5939.7325.1516.5263.0139.263.8817.8810.1251.679.2768.0212.350.240.99
AdaptSegNet [79]67.2548.7332.7914.8945.0045.8830.0034.9212.2255.5015.8573.8416.0731.350.00
CBST [104]69.2549.5835.0614.9547.3954.6834.2736.8313.7856.1518.1374.2316.1834.060.00
IntraDA [61]68.2948.9134.2514.8243.7555.3632.6433.3911.6055.3117.0076.0020.3031.400.00
DACS [77]69.5553.8836.5114.6643.7257.2738.4335.4214.0257.1018.4376.1624.7936.210.00
MIC [38]70.7849.2236.9315.6051.7151.7333.5438.109.4454.2720.7474.4029.7941.780.70
PLSR [94]70.9153.6537.5715.2549.0453.2837.5436.6412.9157.6025.2975.9224.9239.850.24
EventFly (Ours)73.4254.1440.0515.7850.0761.3339.1741.9712.8359.1423.5179.8027.2642.652.86
Target •80.0260.5549.8419.5874.8056.2346.0855.2821.7959.9030.3177.2458.3862.475.81
", + "image_path": "95c16d9f3819da026bb5ca62eb49f1b865d051c95309bc65b394a41e1a7755da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 278, + 555, + 382 + ], + "blocks": [ + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "type": "text", + "content": "Table 3. Benchmark results of platform adaptation among the vehicle " + }, + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "type": "text", + "content": ", drone " + }, + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{d}})" + }, + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "type": "text", + "content": ", and quadruped " + }, + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{q}})" + }, + { + "bbox": [ + 55, + 240, + 555, + 277 + ], + "type": "text", + "content": " platforms, respectively. A total of six cross-platform adaptation settings are considered in our benchmark. Target is trained with ground truth from the target domain. All scores are given in percentage (\\%). The second best and best scores under each metric are highlighted in colors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 278, + 555, + 382 + ], + "lines": [ + { + "bbox": [ + 58, + 278, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 58, + 278, + 555, + 382 + ], + "type": "table", + "html": "
MethodPv → PdPv → PqPd → PvPd → PQPq → PvPq → PdAverage
AccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoUAccmIoU
Source-Only o43.6915.0466.5925.1557.9120.7966.8323.0657.4921.3052.6216.8557.5220.37
AdaptSegNet [79]49.1421.1667.2532.7968.2929.5567.5733.9966.7430.6557.0720.9662.8828.18
DACS [77]59.8127.0769.5536.5171.7836.1067.7336.1171.2034.7860.7424.5066.8032.51
MIC [38]63.1128.8770.7836.9372.4636.8867.2936.2772.4635.2264.4926.1168.4333.38
PLSR [94]64.6129.6970.9137.5772.4637.1867.8336.2172.9336.3863.5727.3468.7234.06
EventFly (Ours)69.1732.6773.4240.0575.5039.9269.6837.3773.9337.7065.7828.7971.2536.08
Target •79.5742.9080.0249.8486.1255.9380.0249.8486.1255.9379.5742.9081.9049.56
", + "image_path": "d621328f44ad87ad0dbce66dfbdf07630da0a31a37f18ae04fc19c799463ee39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 394, + 170, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 170, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 170, + 407 + ], + "type": "text", + "content": "4.2. Comparative Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 412, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 412, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 412, + 295, + 509 + ], + "type": "text", + "content": "Adapt from Vehicle to Drone. Tab. 1 presents the results of transferring vehicle to drone, where there is a huge domain gap in between. The source-only (15.04% mIoU) and supervised target (42.90% mIoU) models validate this gap. Prior adaptation methods show good results in closing the domain gap, while our EventFly surpasses all competitors for almost all semantic classes, demonstrating strong robustness enhancements for event-based cross-platform adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 509, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 593 + ], + "type": "text", + "content": "Adapt from Vehicle to Quadruped. We compare the adaptation methods from vehicle to quadruped in Tab. 2. Since the data and semantic distributions of these two platforms are closer (compared to drone), the adaptation methods show larger performance gains. We achieve a promising " + }, + { + "bbox": [ + 54, + 509, + 295, + 593 + ], + "type": "inline_equation", + "content": "40.05\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 593 + ], + "type": "text", + "content": " mIoU under this setting, which corresponds to a " + }, + { + "bbox": [ + 54, + 509, + 295, + 593 + ], + "type": "inline_equation", + "content": "59.2\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 593 + ], + "type": "text", + "content": " relative improvement over the source-only baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Qualitative Assessments. Fig. 3 visually compares the abilities of different cross-platform adaptation methods. As can be seen, the large domain gap between source and target platforms causes huge performance degradation, resulting in messy semantic predictions (source-only). While state-of-the-art adaptation methods [38, 94] restore the performance to some extent, their predictions still suffer from ambiguous object and background contours. Differently, our EventFly generates more accurate scene semantics, credited to the superior robustness from the EAP-driven designs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 395, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 538 + ], + "type": "text", + "content": "Cross-Platform Adaptations. Since our EXPo benchmark consists of three platforms, we benchmark state-of-the-art methods across a total of six adaptation settings, where each platform interchangeably serve as the source/target domains. As shown in Tab. 3, we observe discrepancies in adaptation difficulties. Specifically, the drone domain is the most difficult to adapt to, mainly due to its unique motion patterns and perspective-based dynamic. Our approach exhibits better robustness than prior arts, achieving the best metrics across all six settings. Such a strong generalization ability is crucial for the deployment of event camera perception algorithms in real-world environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 550, + 408, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 408, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 408, + 562 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 568, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 652 + ], + "type": "text", + "content": "Component Analysis. Our EventFly framework compromises three key components: EventBlend, EventMatch, and the dual-backbone network. As shown in Tab. 4, each contribution provides distinct performance improvements. Since our designs are well-motivated by EAP, the combinations of any two (row " + }, + { + "bbox": [ + 313, + 568, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\# c" + }, + { + "bbox": [ + 313, + 568, + 554, + 652 + ], + "type": "text", + "content": ") and all (row " + }, + { + "bbox": [ + 313, + 568, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\# d" + }, + { + "bbox": [ + 313, + 568, + 554, + 652 + ], + "type": "text", + "content": ") of them further yield better cross-platform adaptation performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Domain Blending Techniques. To validate that our EAP-driven EventBlend operation indeed encourages domain regularization effect, we compare the heuristic data mixing techniques [19, 90] and recent domain mixing methods [23, 77]. As shown in Fig. 4, conventional methods, which" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1478" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 332 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 332 + ], + "type": "image", + "image_path": "6a8585a447e0f2af9a175bca200cf29852f4f59e30ab08c245c597e5c292f420.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 334, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 555, + 360 + ], + "type": "text", + "content": "Figure 3. Qualitative assessments of adaptation from vehicle to drone (the first 3 rows), and from vehicle to quadruped (the last 3 rows). We use grayscaled event images for better visibility. The RGB frames are for reference purposes only. Best viewed in colors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 457, + 296, + 571 + ], + "blocks": [ + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "text", + "content": "Table 4. Ablation results of each component in cross-platform adaptation from vehicle " + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "text", + "content": " to drone " + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{d}})" + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "text", + "content": " and vehicle " + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "text", + "content": " to quadruped " + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\mathrm{q}})" + }, + { + "bbox": [ + 55, + 367, + 295, + 456 + ], + "type": "text", + "content": ", respectively. Dual: The dual-branch network. Match: The EventMatch technique. Blend: The EventBlend operation. All scores are given in percentage (\\%). We compare settings from the (a) source-only, (b) single-component effects, (c) double-component effects, (d) full configuration, and (e) supervised target results, respectively, in this ablation study." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 457, + 296, + 571 + ], + "lines": [ + { + "bbox": [ + 58, + 457, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 58, + 457, + 296, + 571 + ], + "type": "table", + "html": "
#DualMatchBlendPvAccPd mIoUPvAccPq mIoU
(a)XXX43.6915.0466.5925.15
(b)X✓X59.8625.2867.7433.66
XX✓61.8527.9169.5536.82
(c)✓✓X66.5030.2470.9038.10
✓X✓68.4331.9672.6739.11
(d)✓✓✓69.1732.6773.4240.05
(e)Target79.5742.9080.0249.84
", + "image_path": "ae61baa0bcd517cf09014618dc128da072de917116deef5aacb231c4d52afe8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "type": "text", + "content": "neglect the event activation discrepancies, show sub-par performance in bridging the domains. Guided by EAP, our domain blending technique facilitates entropy minimization in high-activation regions, which in turn brings robust feature adaptation across heterogeneous platforms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Similarity Threshold. The strength of blending source and target platforms is vital in our framework. Therefore, we study the effect of the threshold parameter " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". Notably, the 0 (i.e., source-only) or high (" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": ">0.7" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ") value of " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " brings limited gains, as the domain consistency regularization is weak. We found that the best possible trade-offs appear between 0.3" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 368, + 555, + 464 + ], + "blocks": [ + { + "bbox": [ + 315, + 368, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 315, + 368, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 368, + 555, + 464 + ], + "type": "image", + "image_path": "5f344061038c8e4ec8f53cf36f49f6a993f97e0b87c40e240bdcb61b0af0f7fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "type": "text", + "content": "Figure 4. Ablation study (mIoU scores in %) on different domain-mixing techniques under the (left) " + }, + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{v}}\\rightarrow \\mathcal{P}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "type": "text", + "content": ", (middle) " + }, + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{v}}\\rightarrow \\mathcal{P}^{\\mathrm{q}}" + }, + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "type": "text", + "content": ", and (right) " + }, + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{d}}\\rightarrow \\mathcal{P}^{\\mathrm{v}}" + }, + { + "bbox": [ + 313, + 466, + 555, + 500 + ], + "type": "text", + "content": " cross-platform adaptation settings." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 512, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 554, + 537 + ], + "type": "text", + "content": "and 0.5, further validating the importance of bridging domains for cross-platform adaptation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this work, we presented EventFly, a framework for robust cross-platform adaptation in event-based dense perception, enabling robust deployment across the vehicle, drone, and quadruped platforms. Leveraging the proposed Event Activation Prior (EAP), EventBlend, and EventMatch, we addressed the unique adaptation challenges of event stream data collected from different robot platforms. We also introduced EXPo, a large-scale benchmark designed to evaluate cross-platform event perception capabilities. Our approach demonstrated notable improvements over existing methods. We hope this work lays the foundation for more adaptive event-based perception across real-world environments." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1479" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 164 + ], + "type": "text", + "content": "This work is under the programme DesCartes and is supported by the National Research Foundation, Prime Minister's Office, Singapore, under its Campus for Research Excellence and Technological Enterprise (CREATE) programme. This work is also supported by the Apple Scholars in AI/ML Ph.D. Fellowship program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 165, + 296, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 165, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 296, + 201 + ], + "type": "text", + "content": "The authors would like to sincerely thank the Program Chairs, Area Chairs, and Reviewers for the time and efforts devoted during the review process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 216, + 115, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 115, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 115, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 236, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 236, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 236, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 66, + 236, + 295, + 280 + ], + "type": "text", + "content": "[1] Inigo Alonso and Ana C. Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1-10, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 281, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 281, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 296, + 326 + ], + "type": "text", + "content": "[2] Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. Ddd17: End-to-end davis driving dataset. In International Conference on Machine Learning Workshops, pages 1-9, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 327, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 327, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 66, + 327, + 296, + 392 + ], + "type": "text", + "content": "[3] Shristi Das Biswas, Adarsh Kosta, Chamika Liyanagedera, Marco Apolinario, and Kaushik Roy. Halsie: Hybrid approach to learning segmentation by simultaneously exploiting image and event modalities. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5952-5962, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 396, + 295, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 396, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 66, + 396, + 295, + 440 + ], + "type": "text", + "content": "[4] Michael Blösch, Stephan Weiss, Davide Scaramuzza, and Roland Siegwart. Vision based MAV navigation in unknown and unstructured environments. In IEEE International Conference on Robotics and Automation, pages 21-28, 2010." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 441, + 296, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 441, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 66, + 441, + 296, + 496 + ], + "type": "text", + "content": "[5] Chiara Boretti, Philippe Bich, Fabio Pareschi, Luciano Prono, Riccardo Rovatti, and Gianluca Setti. Pedro: An event-based dataset for person detection in robotics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4065-4070, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "type": "text", + "content": "[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "type": "text", + "content": " 130 db 3 μs latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 544, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 544, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 66, + 544, + 296, + 599 + ], + "type": "text", + "content": "[7] Tim Brodermann, David Bruggemann, Christos Sakaridis, Kevin Ta, Odysseas Liagouris, Jason Corkill, and Luc Van Gool. Muses: The multi-sensor semantic perception dataset for driving under uncertainty. arXiv preprint arXiv:2401.12761, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 601, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 601, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 66, + 601, + 296, + 656 + ], + "type": "text", + "content": "[8] David Brüggemann, Christos Sakaridis, Prune Truong, and Luc Van Gool. Refign: Align and refine for adaptation of semantic segmentation to adverse conditions. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3174-3184, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 658, + 296, + 712 + ], + "type": "text", + "content": "[9] Spencer Carmichael, Austin Buchan, Mani Ramanagopal, Radhika Ravi, Ram Vasudevan, and Katherine A. Skinner. Dataset and benchmark: Novel sensors for autonomous vehicle perception. International Journal of Robotics Research, 2024." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "text", + "content": "[10] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In IEEE International Conference on Robotics and Automation, pages 703-710, 2014." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 108, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 108, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 108, + 554, + 151 + ], + "type": "text", + "content": "[11] Bharatesh Chakravarthi, Aayush Atul Verma, Kostas Dani-ilidis, Cornelia Fermuller, and Yezhou Yang. Recent event camera innovations: A survey. In European Conference on Computer Vision Workshops, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 152, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 554, + 218 + ], + "type": "text", + "content": "[12] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multirobot, multi-sensor, multi-environment event dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4016-4023, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "type": "text", + "content": "[13] Peiyu Chen, Weipeng Guan, Feng Huang, Yihan Zhong, Weisong Wen, Li-Ta Hsu, and Peng Lu. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 9(1):407-416, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 264, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 264, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 264, + 554, + 319 + ], + "type": "text", + "content": "[14] Zhiwen Chen, Zhiyu Zhu, Yifan Zhang, Junhui Hou, Guangming Shi, and Jinjian Wu. Segment any event streams via weighted adaptation of pivotal tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890-3900, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 321, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 554, + 364 + ], + "type": "text", + "content": "[15] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 554, + 409 + ], + "type": "text", + "content": "[16] Fernando Cladera, Kenneth Chaney, M. Ani Hsieh, Camillo J. Taylor, and Vijay Kumar. Evmapper: High altitude orthomapping with event cameras. arXiv preprint arXiv:2409.18120, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 411, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 554, + 477 + ], + "type": "text", + "content": "[17] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 478, + 554, + 521 + ], + "type": "text", + "content": "[18] Javier Cuadrado, Ulysse Rançon, Benoit R. Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 523, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 523, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 523, + 554, + 555 + ], + "type": "text", + "content": "[19] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 557, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 554, + 612 + ], + "type": "text", + "content": "[20] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 3942-3951, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 613, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 613, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 613, + 554, + 657 + ], + "type": "text", + "content": "[21] Burak Ercan, Onur Eker, Aykut Erdem, and Erkut Erdem. Hue dataset: High-resolution event and frame sequences for low-light vision. In European Conference on Computer Vision Workshops, pages 1-18, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 658, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 554, + 714 + ], + "type": "text", + "content": "[22] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a " + }, + { + "bbox": [ + 320, + 658, + 554, + 714 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 320, + 658, + 554, + 714 + ], + "type": "text", + "content": " back-illuminated stacked temporal contrast event-based vision sensor with " + }, + { + "bbox": [ + 320, + 658, + 554, + 714 + ], + "type": "inline_equation", + "content": "4.86\\mu \\mathrm{m}" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1480" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 296, + 116 + ], + "type": "text", + "content": "pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In IEEE International Solid-State Circuits Conference, pages 112-114, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 161 + ], + "type": "text", + "content": "[23] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In British Machine Vision Conference, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 162, + 295, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 295, + 226 + ], + "type": "text", + "content": "[24] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 (1):154-180, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 228, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 259 + ], + "type": "text", + "content": "[25] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[26] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014): 1034-1040, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 294, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 295, + 348 + ], + "type": "text", + "content": "[27] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 295, + 392 + ], + "type": "text", + "content": "[28] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 394, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 295, + 437 + ], + "type": "text", + "content": "[29] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 437, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 481 + ], + "type": "text", + "content": "[30] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In IEEE International Conference on 3D Vision, pages 197-206, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 482, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 295, + 514 + ], + "type": "text", + "content": "[31] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems, pages 529-536, 2004." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 515, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 295, + 568 + ], + "type": "text", + "content": "[32] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22867-22876, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 570, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 295, + 602 + ], + "type": "text", + "content": "[33] Dalia Hareb and Jean Martinet. Evsegsn: Neuromorphic semantic segmentation for event data. arXiv preprint arXiv:2406.14178, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 603, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 295, + 646 + ], + "type": "text", + "content": "[34] Javier Hidalgo-Carrio, Daniel Gehrig, and Davide Scaramuzza. Learning monocular dense depth from events. In IEEE International Conference on 3D Vision, pages 534-542, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "text", + "content": "[35] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781-5790, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "text", + "content": "[36] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "text", + "content": "domain-adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 149 + ], + "type": "text", + "content": "[37] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision, pages 372–391, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 151, + 553, + 205 + ], + "type": "text", + "content": "[38] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 320, + 206, + 553, + 249 + ], + "type": "text", + "content": "[39] Yuhwan Jeong, Hoonhee Cho, and Kuk-Jin Yoon. Towards robust event-based networks for nighttime via unpaired day-to-night event translation. In European Conference on Computer Vision, pages 286-306, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 251, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 251, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 320, + 251, + 553, + 316 + ], + "type": "text", + "content": "[40] Linglin Jing, Yiming Ding, Yunpeng Gao, Zhigang Wang, Xu Yan, Dong Wang, Gerald Schaefer, Hui Fang, Bin Zhao, and Xuelong Li. Hpl-ess: Hybrid pseudo-labeling for unsupervised event-based semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23128–23137, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 317, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 317, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 317, + 553, + 360 + ], + "type": "text", + "content": "[41] Uday Kamal and Saibal Mukhopadhyay. Efficient learning of event-based dense representation using hierarchical memories with adaptive update. In European Conference on Computer Vision, pages 74-89, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 404 + ], + "type": "text", + "content": "[42] Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 620(7976):982-987, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 405, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 405, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 320, + 405, + 553, + 459 + ], + "type": "text", + "content": "[43] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In IEEE/CVF International Conference on Computer Vision, pages 2146-2156, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 553, + 503 + ], + "type": "text", + "content": "[44] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2496-24976, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 320, + 504, + 553, + 558 + ], + "type": "text", + "content": "[45] Simon Klenk, David Bonello, Lukas Koestler, Nikita Araslanov, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2378-2388, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "type": "text", + "content": "[46] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 614, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 614, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 614, + 553, + 669 + ], + "type": "text", + "content": "[47] Lingdong Kong, Youquan Liu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Openness: Event-based semantic scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15686-15698, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "text", + "content": "[48] Wouter M. Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):766-785, 2021." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1481" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "text", + "content": "[49] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):14020-14037, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 162 + ], + "type": "text", + "content": "[50] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5743-5762, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "type": "text", + "content": "[51] Yahao Liu, Jinhong Deng, Xinchen Gao, Wen Li, and Lixin Duan. Bapa-net: Boundary adaptation and prototype alignment for cross-domain semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 8801-8811, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 219, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 295, + 262 + ], + "type": "text", + "content": "[52] Antonio Loquercio, Ana I. Maqueda, Carlos R. Del-Blanco, and Davide Scaramuzza. Dronet: Learning to fly by driving. IEEE Robotics and Automation Letters, 3(2):1088-1095, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "type": "text", + "content": "[53] Antonio Loquercio, Elia Kaufmann, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Learning high-speed flight in the wild. Science Robotics, 6(59):5810, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 309, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 295, + 342 + ], + "type": "text", + "content": "[54] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 343, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 295, + 387 + ], + "type": "text", + "content": "[55] Dongyue Lu, Lingdong Kong, Gim Hee Lee, Camille Simon Chane, and Wei Tsang Ooi. Flexevent: Event camera object detection at arbitrary frequencies. arXiv preprint arXiv:2412.06708, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 443 + ], + "type": "text", + "content": "[56] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso García, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5419-5427, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 295, + 488 + ], + "type": "text", + "content": "[57] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. IEEE Robotics and Automation Letters, 7(2):3515-3522, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 295, + 533 + ], + "type": "text", + "content": "[58] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "type": "text", + "content": "[59] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34 (6):1425-1440, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 295, + 624 + ], + "type": "text", + "content": "[60] Poojan Oza, Vishwanath A. Sindagi, Vibashan VS, and Vishal M. Patel. Unsupervised domain adaptation of object detectors: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4018-4040, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 295, + 679 + ], + "type": "text", + "content": "[61] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 295, + 714 + ], + "type": "text", + "content": "[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 138 + ], + "type": "text", + "content": "Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8026-8037, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 553, + 193 + ], + "type": "text", + "content": "[63] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 194, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 194, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 320, + 194, + 553, + 248 + ], + "type": "text", + "content": "[64] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems, pages 16639-16652, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "type": "text", + "content": "[65] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1): 259-275, 2010." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 304, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 304, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 320, + 304, + 553, + 348 + ], + "type": "text", + "content": "[66] Ulysse Rançon, Javier Cuadrado-Anibarro, Benoit R. Cottereau, and Timothée Masquelier. Stereospike: Depth learning with a spiking neural network. IEEE Access, 10: 127428-127439, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 350, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 553, + 393 + ], + "type": "text", + "content": "[67] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):1964-1980, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 394, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 394, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 553, + 458 + ], + "type": "text", + "content": "[68] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8197-8207, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 553, + 514 + ], + "type": "text", + "content": "[69] Manuel Schwonberg, Joshua Niemeijer, Jan-Aike Termöhlen, Nico M. Schmidt, Hanno Gottschalk, and Tim Fingscheidt. Survey on unsupervised domain adaptation for semantic segmentation for visual perception in automated driving. IEEE Access, 11:54296-54336, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 320, + 515, + 553, + 558 + ], + "type": "text", + "content": "[70] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 12:51275-51306, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "type": "text", + "content": "[71] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 615, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 615, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 320, + 615, + 553, + 647 + ], + "type": "text", + "content": "[72] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. arXiv preprint arXiv:1708.07120, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 647, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 714 + ], + "type": "text", + "content": "[73] Bongki Son, Yunjae Suh, Sungho Kim, Heejae Jung, Jun-Seok Kim, Changwoo Shin, Keunju Park, Kyoobin Lee, Jinman Park, Jooyeon Woo, et al. 4.1 a " + }, + { + "bbox": [ + 320, + 647, + 553, + 714 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 320, + 647, + 553, + 714 + ], + "type": "text", + "content": " dynamic vision sensor with a " + }, + { + "bbox": [ + 320, + 647, + 553, + 714 + ], + "type": "inline_equation", + "content": "9\\mu \\mathrm{m}" + }, + { + "bbox": [ + 320, + 647, + 553, + 714 + ], + "type": "text", + "content": " pixel and 300meps address-event representation. In IEEE International Solid-State Circuits Conference, pages 66-67, 2017." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1482" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[74] Lea Steffen, Daniel Reichard, Jakob Weinland, Jacques Kaiser, Arne Roennau, and Rüdiger Dillmann. Neuromorphic stereo vision: A survey of bio-inspired sensors and algorithms. Frontiers in Neuroscience, 13:28, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "text", + "content": "[75] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In European Conference on Computer Vision, pages 341-357, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 164, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 164, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 164, + 294, + 217 + ], + "type": "text", + "content": "[76] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 195–1204, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 274 + ], + "type": "text", + "content": "[77] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 294, + 319 + ], + "type": "text", + "content": "[78] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshops, pages 350-357, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 322, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 294, + 376 + ], + "type": "text", + "content": "[79] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7472-7481, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 378, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 294, + 432 + ], + "type": "text", + "content": "[80] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate slam? combining events, images, and imu for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3 (2):994-1001, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 294, + 488 + ], + "type": "text", + "content": "[81] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517-2526, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "type": "text", + "content": "[82] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 294, + 600 + ], + "type": "text", + "content": "[83] Lin Wang, Yujeong Chae, and Kuk-Jin Yoon. Dual transfer learning for event-based end-task prediction via plug-gable event to image translation. In IEEE/CVF International Conference on Computer Vision, pages 2135-2145, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 294, + 666 + ], + "type": "text", + "content": "[84] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 608-619, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[85] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In IEEE/CVF Conference on" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 19248-19257, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 149 + ], + "type": "text", + "content": "[86] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyuan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 21572-21581, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 151, + 553, + 205 + ], + "type": "text", + "content": "[87] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9004-9021, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 206, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 206, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 553, + 238 + ], + "type": "text", + "content": "[88] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In IEEE/CVF International Conference on Computer Vision, pages 10699-10709, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 321, + 239, + 553, + 270 + ], + "type": "text", + "content": "[89] Zhen Yao and Mooi Choo Chuah. Event-guided low-light video semantic segmentation. arXiv preprint arXiv:2411.00639, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 273, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 273, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 321, + 273, + 553, + 315 + ], + "type": "text", + "content": "[90] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 317, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 317, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 321, + 317, + 553, + 370 + ], + "type": "text", + "content": "[91] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12):14679-14694, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 373, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 373, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 321, + 373, + 553, + 426 + ], + "type": "text", + "content": "[92] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12414-12424, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 427, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 427, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 321, + 427, + 553, + 524 + ], + "type": "text", + "content": "[93] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, and Sebastian Scherer. Subt-mrs dataset: Pushing slam towards all-weather environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 526, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 526, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 321, + 526, + 553, + 590 + ], + "type": "text", + "content": "[94] Xingchen Zhao, Niluthpol Chowdhury Mithun, Abhinav Rajvanshi, Han-Pang Chiu, and Supun Samarasekera. Unsupervised domain adaptation for semantic segmentation with pseudo label self-refinement. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2399-2409, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 321, + 593, + 553, + 646 + ], + "type": "text", + "content": "[95] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "text", + "content": "[96] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "text", + "content": "[97] Qianyu Zhou, Zhengyang Feng, Qiqi Gu, Jiangmiao Pang, Guangliang Cheng, Xuequan Lu, Jianping Shi, and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1483" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 294, + 487 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "text", + "content": "Lizhuang Ma. Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):804-817, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 62, + 107, + 294, + 162 + ], + "type": "text", + "content": "[98] Zhuyun Zhou, Zongwei Wu, Rémi Boutteau, Fan Yang, Cédric Demonceaux, and Dominique Gingac. Rgb-event fusion for moving object detection in autonomous driving. In IEEE International Conference on Robotics and Automation, pages 7808-7815, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 294, + 195 + ], + "type": "text", + "content": "[99] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Robotics: Science and Systems, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 59, + 197, + 294, + 251 + ], + "type": "text", + "content": "[100] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 58, + 252, + 294, + 306 + ], + "type": "text", + "content": "[101] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 294, + 351 + ], + "type": "text", + "content": "[102] Shifan Zhu, Zixun Xiong, and Donghyun Kim. Cear: Comprehensive event camera dataset for rapid perception of agile quadruped robots. IEEE Robotics and Automation Letters, 9(10):8999-9006, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 353, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 353, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 57, + 353, + 294, + 385 + ], + "type": "text", + "content": "[103] Rong Zou, Manasi Muglikar, Nico Messikommer, and Davide Scaramuzza. Seeing behind dynamic occlusions with event cameras. arXiv preprint arXiv:2307.15829, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 387, + 294, + 430 + ], + "type": "text", + "content": "[104] Yang Zou, Zhiding Yu, B. V. K. Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In European Conference on Computer Vision, pages 289-305, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 432, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 294, + 487 + ], + "type": "text", + "content": "[105] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, pages 12846-128567, 2023." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1484" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_content_list.json b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d751f60a6eddf95bdd6c149883e7bab2d534c322 --- /dev/null +++ b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_content_list.json @@ -0,0 +1,1919 @@ +[ + { + "type": "text", + "text": "EventGPT: Event Stream Understanding with Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 204, + 130, + 794, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaoyu Liu $^{1,2}$ , Jianing Li $^{2}$ , Guanghui Zhao $^{1*}$ , Yunjian Zhang $^{2}$ , Xin Meng $^{3}$ , Fei Richard Yu $^{4}$ , Xiangyang Ji $^{2*}$ , Ming Li $^{4*}$", + "bbox": [ + 218, + 202, + 800, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xidian University $^{2}$ Tsinghua University $^{3}$ Peking University $^{4}$ Guangdong Laboratory of Artificial Intelligence and Digital Economy(SZ)", + "bbox": [ + 210, + 255, + 808, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "23171110721@stu.xidian.edu.cn, lijianing@pku.edu.cn, ming.li@u.nus.edu, ghzhao@xidian.edu.cn", + "bbox": [ + 99, + 295, + 915, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras capture visual information as asynchronous pixel change streams, excelling in challenging lighting and high-dynamic scenarios. Existing multimodal large language models (MLLMs) concentrate on natural RGB images, failing in scenarios where event data fits better. In this paper, we introduce EventGPT, the first MLLM for event stream understanding, pioneering the integration of large language models (LLMs) with event-based vision. To bridge the huge domain gap, we propose a three-stage optimization paradigm to progressively equip a pre-trained LLM with event understanding. Our EventGPT consists of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Firstly, GPT-generated RGB image-text pairs warm up the linear projector, following LLaVA, as the gap between natural images and language is smaller. Secondly, we construct N-ImageNet-Chat, a large synthetic dataset of event data and corresponding texts to enable the use of the spatio-temporal aggregator and to train the event-language adapter, thereby aligning event features more closely with the language space. Finally, we gather an instruction dataset, EventChat, which contains extensive real-world data to fine-tune the entire model, further enhancing its generalization ability. We construct a comprehensive benchmark, and experiments show that EventGPT surpasses previous state-of-the-art MLLMs in generation quality, descriptive accuracy, and reasoning capability. Code: EventGPT", + "bbox": [ + 89, + 377, + 485, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 816, + 220, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras, with predominant advantages in capturing scene changes and temporal resolution, have great ap", + "bbox": [ + 89, + 843, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6719442a966cd1edfd8fd39e5a4de7f04a29dba2b600e35d10985b6641e0fe6e.jpg", + "image_caption": [ + "Figure 1. Our EventGPT is the first multimodal large language model tailored for event stream understanding, including scene summarization, reasoning, and question answering, with a great potential for downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 521, + 345, + 903, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plication potential in scenarios under imperfect lighting conditions or including high dynamic objects [7, 12, 57]. The relevant tasks on event data, e.g. object detection [8, 10, 16, 46], tracking [9, 31, 39, 53], and 3D reconstruction [14, 32, 35, 49], have been well-investigated.", + "bbox": [ + 511, + 609, + 906, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent multimodal large language models (MLLMs) have shown remarkable breakthroughs in conventional vision and language problems, demonstrating extraordinary capabilities of scene understanding and question answering (QA) [25, 26, 38, 47, 60]. Typically, a pioneering research Large Language and Vision Assistant (LLaVA) combines a vision encoder with an LLM, trained on GPT-4-generated multimodal data, to create a robust system that excels in visual understanding, instruction-following, and complex question answering [26]. These works mainly pay attention to natural RGB scenes and the resulting models degrade significantly when encountering the aforementioned scenarios. However, the relevant research on event data streams remains underexplored.", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29139", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Model (Vicuna v1.5)", + "text_level": 1, + "bbox": [ + 375, + 151, + 620, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/43512ca007ee5fe49f1b1bb13e80f993a2aca04578c53ecaedefcbfc99b57551.jpg", + "image_caption": [ + "Figure 2. Overview of our framework. The event encoder transforms raw event tensors into high-dimensional features, which are then aggregated by a spatio-temporal module. These representations are projected and aligned with the large language model, enabling nuanced understanding of event streams and supporting various downstream applications." + ], + "image_footnote": [], + "bbox": [ + 99, + 172, + 890, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some works have initialized the alignment between event data and textual information. For example, EventCLIP [45] adapts the pre-trained CLIP model [6, 34] for event-based object recognition, enabling zero-shot and few-shot classification. Similarly, EventBind [58] extends CLIP with an event encoder to align images, texts, and events in a unified representation space, realizing few-shot recognition and event retrieval. Nevertheless, both EventCLIP and EventBind lack the ability to leverage the embedded world knowledge in LLMs, limiting their capacity for nuanced scene understanding and dialog generation in complex scenarios. This gap highlights the necessity of a model, akin to LLaVA for RGB images, specifically designed to overcome these limitations by harnessing the power of LLMs for comprehensive event stream understanding.", + "bbox": [ + 88, + 500, + 483, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we pioneer the development of an MLLM for event stream understanding, termed EventGPT (see Fig 1), designed for the community to support scene summarization, reasoning, and question answering. To this end, several key challenges must be addressed. First of all, aligning event features with the high-level semantics of natural language is significantly challenging, as event streams are designed to capture asynchronous and sparse brightness changes within the field of view, inherently lacking the global scene context necessary for language-based understanding. Secondly, the fundamental difference in how", + "bbox": [ + 88, + 734, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "event and RGB cameras capture scene information complicates the straightforward adaptation of existing multimodal techniques, which are typically designed for natural images. Finally, the scarcity of large-scale, high-quality event-text datasets hinders the training of a scalable MLLM for comprehensive event scene understanding.", + "bbox": [ + 511, + 500, + 906, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a three-stage training paradigm that leverages an LLM as an anchor, given its rich world knowledge and a strong foundation for scene understanding, to progressively align the event modality with language. In general, our EventGPT is mainly composed of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Compared to event data, natural images are much closer to language in the embedding space. Therefore, in the first stage, we use the image-text dataset from LLaVA to warm up the linear projector, establishing preliminary scene-language associations. In the second stage, we focus on training the spatio-temporal aggregator and event-language adapter to map spatio-temporal information from event frames into language space. For this purpose, we construct a large-scale synthetic event-text dataset, NImageNet-Chat, to facilitate optimization. Finally, we conduct instructional fine-tuning of all model parameters on a self-collected, high-quality dataset, Event-Chat, which includes diverse scenes with unsatisfactory lighting condi", + "bbox": [ + 511, + 599, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "#", + "bbox": [ + 99, + 94, + 148, + 133 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "EventGPT: This image depicts a person puncturing a balloon with a needle, causing a sudden, dramatic eruption of water. The water splashes out in a forceful burst, creating intricate trails and droplets that scatter through the air in all directions ...", + "bbox": [ + 174, + 101, + 879, + 127 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29140", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions or high-speed object motions, enhancing the capability of our EventGPT to interpret complex real-world scenes.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a comprehensive evaluation benchmark inspired by Qwen2-72B-Instruct to assess the generation quality, descriptive accuracy, and reasoning capability of EventGPT. To facilitate a rigorous comparison, we select several state-of-the-art MLLMs excelling in natural scene understanding, including LLaVA [26], Qwen2-VL [43], and Deepseek-VL [28], as baseline models. Evaluation is conducted across three metrics, i.e. Detail Captioning (DC), Complex Reasoning (CR), and Visual Question Answering (VQA), in line with the existing literature. Experimental results highlight significant improvements achieved by EventGPT. Furthermore, our model demonstrates strong generalization ability on open-set downstream tasks, such as object detection and semantic segmentation.", + "bbox": [ + 88, + 121, + 482, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions in this work can be summarized as follows:", + "bbox": [ + 89, + 332, + 482, + 362 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present EventGPT, the first multimodal large language model tailored for event stream understanding, demonstrating capabilities in scene summarization, reasoning, and question answering. EventGPT establishes a solid baseline, paving the way for future research.", + "- We propose a systematic learning paradigm incorporating specialized training strategies and network modules to progressively bridge the substantial domain gap between event data and language, providing a valuable reference for future cross-modal learning efforts.", + "- We construct two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, containing 1 million synthetic samples for pre-training and 120,000 instruction samples for fine-tuning. These datasets will be publicly released to drive progress in multimodal learning within the event domain." + ], + "bbox": [ + 99, + 363, + 483, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 618, + 240, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. In deep learning era[19-23, 29, 41, 50, 55, 56], significant advancements have been made in various domains. The significant advancements in LLMs for natural language understanding have driven their integration into multimodal research domains, with a focus on enhancing vision-language alignment to expand their applicability [2-4, 24, 33, 38, 40, 42, 48, 52, 54]. Early contributions, such as Flamingo [1] and BLIP [17, 18], integrated visual and language modalities to enable zero-shot generalization across a broad range of visual-language tasks. Further advancements emerged with chat-oriented MLLMs like MiniGPT-4 [60] and LLaVA [26], which improved the interactive capabilities and adaptability of LLMs in multimodal applications, setting new benchmarks in performance. As MLLMs continue to evolve, the exploration of novel sensor modalities holds potential for expanding their functional scope. De", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0cae54f9b3797762b8bad9629f3d6081e584de57b5a474c711ba49b8a9a02cf0.jpg", + "image_caption": [ + "Figure 3. Data distribution across three progressive training stages in our EventGPT containing image-language alignment, event-language alignment, and instruction tuning." + ], + "image_footnote": [], + "bbox": [ + 519, + 95, + 668, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/771db84bbc6a2744a1c91e6738f97fe73dceb996cd4d3d6cababc94f153fc9ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 90, + 900, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "spite the distinct advantages of event cameras, such as high temporal resolution and robustness in challenging lighting conditions, the integration of event-based data with LLMs remains underexplored, particularly for applications in high-dynamic and low-light environments.", + "bbox": [ + 511, + 316, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-Driven Multimodal Models. The integration of event-based data into VLMs has attracted significant interest, demonstrating potential across various downstream tasks and highlighting the advantages of event-language synergy. [45, 58, 59]. Notably, frameworks such as ExACT have leveraged language-guided frameworks for event-based action recognition, effectively aligning action sequences with contextual textual cues to achieve strong performance [59]. Other approaches, such as EventCLIP [45] and EventBind [58], have also demonstrated strong zero-shot performance, underscoring the effectiveness of event-based embeddings when aligned with textual information. Despite these advances, developing chat-centric LLMs capable of incorporating event-based modalities presents significant research challenges. Addressing this gap could enhance models' contextual comprehension, enabling more effective processing of complex interactive tasks.", + "bbox": [ + 511, + 393, + 905, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Event-Text Data Generation", + "text_level": 1, + "bbox": [ + 511, + 662, + 774, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based datasets have been extensively studied in computer vision tasks. However, large-scale Event-Text datasets suitable for training MLLMs) are not yet widely available. Previous studies have successfully employed large models, such as GPT, to synthesize data for multimodal applications, yielding significant advancements in model capabilities [13, 44, 51, 61]. To address this gap, we construct an Event-Text dataset using a human-assisted, semi-automated pipeline that incorporates both synthetic and real-world event data. The distribution of the dataset is shown in Fig. 3.", + "bbox": [ + 511, + 686, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Synthetic Dataset", + "text_level": 1, + "bbox": [ + 511, + 847, + 684, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ImageNet serves as a foundational benchmark for visual recognition tasks. N-ImageNet [15] extends this founda", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29141", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instruction Examples", + "text_level": 1, + "bbox": [ + 228, + 95, + 361, + 109 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt:", + "text_level": 1, + "bbox": [ + 109, + 117, + 156, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Caption Prompt:", + "text_level": 1, + "bbox": [ + 107, + 131, + 200, + 141 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "You are an expert in image description and understanding. You are requested ...", + "bbox": [ + 107, + 142, + 375, + 164 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reasoning Prompt:", + "text_level": 1, + "bbox": [ + 109, + 165, + 217, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "You are an advanced multimodal model with the ability to perform complex reasoning ...", + "bbox": [ + 107, + 178, + 377, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual QA Prompt:", + "text_level": 1, + "bbox": [ + 107, + 202, + 214, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "You are an AI language model designed to generate detailed and insightful Visual Question Answering....", + "bbox": [ + 107, + 214, + 375, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/23a7c812e4c04efe921740bb56c657da6dac85aa1fce715ead44fcdd08ca80d3.jpg", + "image_caption": [ + "Figure 4. Instruction examples from the Event-Chat dataset for multimodal task prompts and responses in image captioning, visual reasoning, and visual question answering." + ], + "image_footnote": [], + "bbox": [ + 383, + 133, + 470, + 184 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e4a8e328bf52da19c140c8cc9b70dcc7cdaa140a6d1679f7e6bfd77628d4f13e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 186, + 470, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Results:", + "text_level": 1, + "bbox": [ + 109, + 244, + 153, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Caption:", + "text_level": 1, + "bbox": [ + 107, + 258, + 155, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query: Describe the buildings in the image.", + "bbox": [ + 107, + 268, + 326, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Answer: There are several multi-story buildings in the image...", + "bbox": [ + 107, + 280, + 416, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reasoning", + "text_level": 1, + "bbox": [ + 107, + 297, + 161, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query: If the construction barriers were removed, how would it impact the traffic flow at the intersection?", + "bbox": [ + 107, + 309, + 462, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Answer: Without the barriers, vehicles would have a direct path..", + "bbox": [ + 107, + 330, + 429, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual QA", + "text_level": 1, + "bbox": [ + 107, + 349, + 160, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query: How many cars are visible, and what directions are they moving?", + "bbox": [ + 107, + 359, + 467, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Answer: Three cars are visible. The one on the left is turning away..", + "bbox": [ + 107, + 371, + 444, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion by providing a synthetic event-based dataset generated through event simulators, aligned with ImageNet images. Using this alignment, we generate textual instructions for the images to enable cross-domain learning between visual and event-based data.", + "bbox": [ + 89, + 470, + 483, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "N-ImageNet-Chat. We present a large-scale dataset comprising over 1,000,000 instances with concise instructions, primarily designed to align event data with natural language. This dataset provides a foundation for the model's understanding and supports robust representation learning.", + "bbox": [ + 89, + 547, + 482, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "N-ImageNet-Instruction. We establish an event-based dataset of over 69,000 entries, designed to improve the model's responsiveness to complex, task-specific instructions. The dataset includes captions, visual question answering (VQA), and complex reasoning tasks aimed at finetuning the model's multimodal capabilities.", + "bbox": [ + 89, + 623, + 483, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Real-World Dataset", + "text_level": 1, + "bbox": [ + 89, + 724, + 279, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to synthetic datasets, we establish a real-world dataset to improve model generalization for practical applications. A human-assisted and semi-automated annotation process ensures high-quality labels for complex scenes.", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event Data for Event-Chat. We leverage real-world event datasets, such as DSEC [11] and E2VID [36], which are primarily designed for autonomous driving and provide synchronized image-event data under challenging conditions, including low-light environments and high-speed motion (e.g., entering/exitting tunnels, night scenes).", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instructions for Event-Chat. This dataset is designed to enhance the model's generalization capabilities, consisting of 59,000 high-quality annotated entries. Three distinct prompting strategies (i.e., captioning, VQA, and complex reasoning) form the basis of the Event-Chat dataset, providing diverse and comprehensive instructions. As shown in Fig. 4, examples from each prompting strategy illustrate the dataset's richness in contextual understanding. Combined with N-ImageNet-instruction, these annotated entries form the final Event-Chat dataset.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, our newly built event-text dataset contains over 1,000,000 entries from both synthetic and real-world data, offering a robust foundation for advancing multimodal understanding in both event-based and textual domains. These fundamental and challenging datasets are intended to foster progress in the development of MLLMs within the neuromorphic community.", + "bbox": [ + 511, + 242, + 905, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. EventGPT", + "text_level": 1, + "bbox": [ + 513, + 361, + 625, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 386, + 622, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EventGPT aims to serve as an event-based MLLM capable of understanding and generating responses based on spatio-temporal event data. As illustrated in Fig. 2, EventGPT formulates event data processing as an event-driven approach to question answering and descriptive generation. By leveraging the high temporal resolution and extended dynamic range inherent in event stream data, EventGPT enhances understanding in scenarios traditionally challenging for standard vision models, such as low-light or high-speed motion conditions. At the core of EventGPT lies an LLM that synthesizes complex spatio-temporal information from event streams, utilizing world knowledge to generate coherent, human-readable outputs.", + "bbox": [ + 511, + 409, + 905, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, an event bin can be mathematically described within a time window $T$ as $S = \\{(x_{j},y_{j},t_{j},p_{j})\\}_{j = 1}^{N}$ , where each event consists of spatial coordinates $(x_{j},y_{j})$ , timestamp $t_j$ , and polarity $p_j$ . Initially, high-dimensional event features $V$ are extracted from $S$ using an event encoder $f_{\\mathrm{enc}}$ , initialized with OpenCLIP ViT-L/14-336px to leverage its robust feature extraction capabilities. These features $V$ are then processed through a spatio-temporal encoder $f_{\\mathrm{ST}}$ to capture dynamic interactions, resulting in a comprehensive representation $E$ :", + "bbox": [ + 511, + 606, + 905, + 757 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV = f _ {\\text {e n c}} (S), \\quad E = f _ {\\mathrm {S T}} (V) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 771, + 903, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, a two-stage projector is employed to achieve cross-modal alignment. The first stage aligns the spatio-temporally encoded features $E$ with a textual representation, and the second stage adapts this representation specifically to the event modality, resulting in an event-aligned feature $F$ . Given a user query $Q$ , the large language model $f_{\\mathrm{llm}}$ performs reasoning over $E$ to generate a", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29142", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "response $A$", + "bbox": [ + 89, + 90, + 174, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF = f _ {\\text {p r o j}} \\left(f _ {\\text {a l i g n}} (E)\\right), \\quad A = f _ {\\mathrm {l l m}} (F \\mid Q) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 122, + 482, + 140 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EventGPT creates a strong representation of event streams and incorporates task-specific heads to support downstream tasks like object detection and instance segmentation. This flexibility enhances its applicability to a wide range of complex vision tasks.", + "bbox": [ + 89, + 146, + 483, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Spatio-Temporal Information Representation", + "text_level": 1, + "bbox": [ + 89, + 232, + 475, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The event stream provides rich spatiotemporal cues crucial for precise temporal and spatial modeling, especially in high-speed motion scenarios, to enhance a model's contextual understanding. The continuous event stream is typically segmented into discrete temporal bins for efficient processing. Each temporal bin $E_{t}$ is independently encoded, forming discrete spatiotemporal feature representations:", + "bbox": [ + 89, + 255, + 483, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} = \\left\\{f _ {t} ^ {e} \\right\\} _ {t = 1, 2, \\dots} ^ {T}, \\quad \\mathcal {Z} \\in \\mathbb {R} ^ {T \\times S \\times D} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 375, + 482, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S$ denotes the spatial resolution of the encoded features, and $D$ represents the feature dimensionality after encoding.", + "bbox": [ + 89, + 401, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To aggregate salient temporal and spatial information, we perform separate temporal and spatial encoding operations. Specifically, Temporal Encoding (TE) aggregates features along the temporal dimension via average pooling:", + "bbox": [ + 89, + 448, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} _ {T} ^ {a v g} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} f _ {t} ^ {e}, \\quad \\mathcal {Z} _ {T} ^ {a v g} \\in \\mathbb {R} ^ {S \\times D} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 520, + 482, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, Spatial Encoding (SE) aggregates features along the spatial dimension via average pooling:", + "bbox": [ + 89, + 568, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} _ {S} ^ {a v g} = \\frac {1}{S} \\sum_ {t = 1} ^ {S} f _ {t} ^ {e}, \\quad \\mathcal {Z} _ {S} ^ {a v g} \\in \\mathbb {R} ^ {T \\times D} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 611, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the temporally and spatially aggregated representations are concatenated along the feature dimension, forming a robust spatiotemporal embedding:", + "bbox": [ + 89, + 657, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathcal {Z}}} = \\operatorname {C o n c a t} \\left(\\mathcal {Z} _ {T} ^ {a v g}, \\mathcal {Z} _ {S} ^ {a v g}\\right), \\quad \\overline {{\\mathcal {Z}}} \\in \\mathbb {R} ^ {(T + S) \\times D} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 729, + 482, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The fused representation $\\overline{Z}$ is an end-to-end learned spatiotemporal feature from asynchronous events, forming a robust event embedding from various complex scenarios.", + "bbox": [ + 89, + 760, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training Pipelines", + "text_level": 1, + "bbox": [ + 89, + 816, + 267, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the huge domain gap between event and language, we include a visual-language pre-training to initialize our framework. As shown in Fig. 5, EventGPT adopts a three-phase training framework encompassing visual-language,", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6bba3818f32e079dce0ec1810d1ec93247808ffa3f61ba5eddb605f9773cf437.jpg", + "image_caption": [ + "Stage 1: Visual-Language Alignment Stage 2: Event-Language Alignment Stage 3: Instruction Finetuning", + "Figure 5. Our three-stage pipeline consists of (1) training a linear projector to achieve visual-language alignment, (2) incorporating a spatio-temporal aggregator and an event-language adapter to enable event-language alignment, and (3) performing full fine-tuning of all modules to facilitate cohesive multimodal learning." + ], + "image_footnote": [], + "bbox": [ + 531, + 103, + 903, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "event-language, and instruction tuning. This structured progression facilitates effective cross-domain alignment, enhancing EventGPT's capacity for multimodal reasoning and understanding.", + "bbox": [ + 511, + 380, + 905, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual-Language Adapter. In the visual-language alignment stage, we leverage the LLaVA-Pre-train dataset while keeping the event encoder and LLM modules frozen, focusing exclusively on training the linear projector. Visual-language alignment is prioritized due to the relatively small domain differences between image and text, compared to the distinct nature of event and textual data. This setup establishes initial cross-modal alignment between vision and language, enabling the LLM to interpret scene information and providing a foundational multimodal understanding.", + "bbox": [ + 511, + 453, + 906, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Event-Language Adapter. In the event-language alignment stage, we leverage our constructed N-ImageNet-Chat dataset, with the event encoder, linear projector, and LLM modules kept frozen, focusing exclusively on training the event-language adapter module. This stage aims to bridge the domain gap between event and language modalities. By aligning event embeddings with the LLM's representational space, the model's ability to interpret event streams and generate descriptive narratives is significantly enhanced.", + "bbox": [ + 511, + 616, + 905, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction Tuning. In the instruction tuning stage, all trainable parameters, including the event encoder, linear projector, event-language adapter, and LLM modules, are unfrozen to enable supervised fine-tuning. Trained on our high-quality Event-Chat question-answering dataset, this stage further enhances the model's understanding and generation capabilities for event stream data. This structured, three-phase training framework progressively narrows the domain gap between event and language modalities.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29143", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8f2333f3f84e76ad98d5bb8afac4445083f5382e8a1763c8f80f90022ede1038.jpg", + "table_caption": [ + "Table 1. Performance comparison between EventGPT and five state-of-the-art methods on the N-ImageNet-Chat and Event-Chat datasets across three evaluation metrics. EventGPT consistently demonstrates competent performance on all datasets." + ], + "table_footnote": [], + "table_body": "
ModelsLLM BackboneParamsEncoderN-ImageNet-ChatEvent-Chat
DCCRVQADCCRVQA
LLaVA-7B-v1.5 [26]Vicuna-v1.57BCLIP ViT-L/336PX1.541.071.882.204.043.26
Qwen2-VL-7B [43]Qwen27BOpenCLIP ViT-bigG1.741.461.912.384.022.91
Intern2VL-8B [5]InternLM2.58BInternViT-300M1.511.872.082.374.003.71
Deepseek-vl-7b [28]Deepseek-LLM7BSAM-B + SigLIP-L1.521.451.842.414.103.37
EventGPT-7BVicuna-v1.57BCLIP ViT-L/336PX2.392.572.233.524.094.29
LLaVA-13B-v1.5 [26]Vicuna-v1.513BCLIP ViT-L/336PX1.751.211.842.344.083.37
EventGPT-13BVicuna-v1.513BCLIP ViT-L/336PX2.412.812.403.404.134.26
", + "bbox": [ + 83, + 128, + 903, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/733cd18ab7858fb825dede73b4c1b10ae8c5e076fcff944aa35558875f528fb4.jpg", + "image_caption": [ + "Figure 6. Representative visualization comparison between EventGPT and other open-source MLLMs models under extreme lighting variations. EventGPT demonstrates superior accuracy in recognizing critical scene details, such as vehicle presence at tunnel exits, showcasing its robustness in challenging lighting conditions." + ], + "image_footnote": [], + "bbox": [ + 114, + 294, + 240, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4838196664db2922b319b7e8d5cf5650d4bf3eb4202ea15059d7af2d6661a6ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 295, + 364, + 368 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6adcdc8d7d4488bf49e7d9f70604a9cb67298a1344250fe8cd8c8a1a5cd8fed2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 295, + 493, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e05a64d6f27d68849795d41583c620a046097cb4e62ba2117c04c649e37283b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 295, + 625, + 368 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fa3ec0385e90f9974451546e8e224da59df1e4542302d2a0d315beefae0af59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 295, + 754, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8913e1d5c018c8b677747a514934989b6d59403d3c76667d50bc199657ea7193.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 295, + 885, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt: Describe the scene depicted in the picture and identify the dangers.", + "bbox": [ + 109, + 373, + 583, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "##LLaVA-7B: The scene depicted is a car driving on a highway, with a tunnel entrance ahead. The car is approaching the large tunnel opening which is currently blocked by a large boulder. The boulder is preventing the car from entering the tunnel and may potentially cause a dangerous situation, as the driver might not be aware of the situation or stop in time to avoid it...", + "bbox": [ + 107, + 390, + 888, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EventGPT -7B: This scene shows a vehicle traveling through a tunnel toward the open road outside. The dangers in this scene include the potential for oncoming traffic from the opposite lane, as well as the reduced visibility due to the tunnel's enclosed space. Drivers are advised to stay alert and prepared for any sudden appearances of vehicles or obstacles.", + "bbox": [ + 107, + 434, + 888, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prrs: Are there any vehicles driving at the tunnel exit?", + "bbox": [ + 107, + 483, + 475, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "```python\n>>> LLaVA-7B : No, there are no vehicles driving at the tunnel exit in the image.", + "bbox": [ + 107, + 502, + 566, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EventGPT: Yes, there is a vehicle driving at the tunnel exit.", + "bbox": [ + 109, + 515, + 480, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 609, + 215, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 89, + 633, + 282, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementations. We integrate the N-ImageNet[15], Dsec[11], and E2VID[36] datasets with synthesized multimodal conversational datasets, N-ImageNet-Chat and Event-Chat, using a GPT-assistant approach for three-stage model training. Our LLM backbone is Vicuna-v1.5, with a two-layer MLP as the linear projector and a linear layer as the event-language adapter. All models, including 7B and 13B versions, are trained on 8 NVIDIA RTX 6000 Ada GPUs. In the visual-language alignment stage, we use 558k data with a learning rate of $2 \\times 10^{-4}$ and a batch size of 64. For event-language alignment, we use 1000k data with the same learning rate and batch size. In instruction-supervised fine-tuning, we use 120k data, setting the learning rate to $2 \\times 10^{-5}$ and batch size to 32.", + "bbox": [ + 89, + 657, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following the evaluation strategies of LLaVA [26] and VideoGPT [30], we rigorously assess", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EventGPT's performance against state-of-the-art MLLMs using three core metrics within a zero-shot question-answering setting. Our benchmark, specifically designed for text generation from event streams, is based on the test sets of N-ImageNet-Chat and Event-Chat datasets, which contain question-answer pairs to evaluate the model's abilities in detailed description, complex reasoning, and visual question answering. Generated responses are quantitatively scored on a 1-5 scale across three dimensions: (i) Detailed Captioning (DC), assessing the accuracy and completeness of descriptions to capture key event details; (ii) Complex Reasoning (CR), evaluating the model's capacity for multi-step reasoning and logical inference; and (iii) Visual Question Answering (VQA), measuring precision and contextual understanding in responding to specific queries.", + "bbox": [ + 511, + 611, + 906, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-Art Models", + "text_level": 1, + "bbox": [ + 511, + 847, + 870, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. To assess the effectiveness of EventGPT, we report the quantitative evaluation results of its", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29144", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ee100de6a3848a5634026a637a4c9d4cde65b9c2496e34e01d525256a011fa4.jpg", + "image_caption": [ + "Figure 7. Representative comparison of EventGPT and other open-source multimodal language models (MLLMs) in high-speed motion scenarios, demonstrating EventGPT's enhanced capability to capture spatiotemporal details and interpret motion characteristics, including fragment trajectories and impact effects." + ], + "image_footnote": [], + "bbox": [ + 106, + 97, + 232, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6f04d867be761646038b6fe13b564360e87764d6332c34a056303e4e993c4bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 98, + 364, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/764473f41db5737dd2bd2e90d0a5d2c43578ad8b6992e62573272cbe208a36de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 98, + 496, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2438e4da493a505bc4ecbe0663adb1c98a260c9314bc5612d09e4d06e413b48d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 98, + 627, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f15b8c98d9db183636751f57cac1c5a388a3e47ff6c124a5b5ea406aaf4ef53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 98, + 759, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c119ba8c5fbd2b2b7a1cdeec7825d6c026ae03049f388e8e1c76d988db10965.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 98, + 893, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prrs: Plesse describe this scene and explain what scene or action it depicts.", + "bbox": [ + 102, + 179, + 583, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "##LLaVA-7B: The scene shows a broken piece of glass flying through the air in front of a white sheet. This glass could be from a broken window, creating an unexpected scene. The sheet of white fabric is likely designed to protect the area from the flying glass shards or to catch them if they land. The shards of glass are scattered around the scene, with some closer to the sheet and others further away, creating a sense of danger and urgency.", + "bbox": [ + 101, + 196, + 893, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qwen2-VL-7B: The scene depicts a scene where a ceramic object, possibly a bowl or a plate, is being shattered. The object is placed on a white surface, which appears to be a piece of paper or a cloth. The shattered pieces of the ceramic object are scattered around the surface, with some pieces still attached to the main body of the object...", + "bbox": [ + 101, + 253, + 893, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EventGPT : This scene depicts the process of a cup breaking. The moment of impact reveals fragments bursting outward, scattering in various directions. Each shard captures the force of the impact, some pieces spinning and tumbling through the air, while a faint dust cloud forms around the destruction...", + "bbox": [ + 101, + 297, + 893, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7B and 13B versions in Table 1. To maintain an end-to-end architecture and ensure fair comparison, we represent the event stream as an event image under identical input conditions. We compare our model with state-of-the-art MLLMs, including Qwen2-VL [43], DeepSeek-VL [28], Intern2VL [5], and LLaVA [26]. The results demonstrate that EventGPT exhibits strong capabilities in event understanding, which is mainly attributed to our three-stage training framework that gradually bridges the significant gap between events and language.", + "bbox": [ + 89, + 412, + 483, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis. In addition, we conduct qualitative assessments on selected challenging scenarios from the Event-Chat test set, specifically those involving extreme lighting conditions or high-speed object dynamics (see Fig. 6 and Fig. 7). The results illustrate that our proposed EventGPT consistently outperforms other models in these difficult settings, demonstrating robust event-driven scene understanding capabilities. For high dynamic range scenarios, such as the tunnel scene in Fig. 6, EventGPT successfully identifies the tunnel environment and issues relevant hazard warnings. Other models, however, frequently misinterpret the scene or fail to detect the tunnel due to the extreme light changes at the exit point. In such conditions, EventGPT demonstrates a unique advantage by effectively capturing high-exposure elements at the tunnel exit and generating timely alerts. For high-speed motion scenarios, such as a cup breaking in Fig. 7, EventGPT demonstrates advanced spatio-temporal understanding by capturing the dynamic rupture sequence of the object. In contrast, LLaVA-7B misinterprets the scene with an inaccurate description, while Qwen2-VL-7B lacks spatio-temporal comprehension, producing a static and incomplete depiction.", + "bbox": [ + 89, + 569, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 411, + 666, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contribution of Each Component. To explore the impact of each component, we first construct a baseline similar to LLaVA except adopting event frames in the second training stage. Two variants are constructed by incorporating our event-language adapter or spatio-temporal aggregator. As shown in Table 2, our proposed EventGPT achieves higher performance compared to the baseline and variants. Compared to the baseline, incorporating the spatio-temporal aggregator results in respective improvement of $3.24\\%$ , $2.02\\%$ , and $2.41\\%$ on the DC, CR, and VQA metrics, significantly enhancing the model's capability of capturing temporal dynamics and spatial relationships within event streams, which facilitates to better interpret complex motion patterns and scene changes. Additionally, the event-language adapter brings improvements of $2.35\\%$ , $1.26\\%$ , and $1.20\\%$ on the DC, CR, and VQA metrics respectively over the baseline. This highlights the superiority of our training pipeline. We extend the traditional two-stage paradigm to a three-stage fine-grained progress. By progressively bridging the large cross-modal gap, our model achieves more accurate event stream perception and contextual understanding.", + "bbox": [ + 511, + 441, + 906, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of Temporal Aggregation Size. To explore the effect of temporal aggregation size, we conduct an analysis to determine the optimal temporal window number $N_{w}$ within the spatio-temporal aggregator. $N_{w}$ represents the number of temporal bins used to divide a fixed observation period, with each bin aggregating events within its respective interval. Fewer windows may fail to capture fine-grained temporal details, while more windows can capture", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29145", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64b0dda9578c8976de17cc0b933eea42ac33c926e364c45e4347dbc7c5f51410.jpg", + "table_caption": [ + "Table 2. Performance impact of each component in EventGPT through ablation study. The results demonstrate that both the spatio-temporal aggregator and the event-language adapter contribute significant performance improvements over the baseline." + ], + "table_footnote": [], + "table_body": "
Mode TypeEvent LanguageSpatio TemporalEvent-Chat
DCCRVQA
Baseline××3.403.974.15
Variant A✓×3.48(+2.35%)4.02(+1.26%)4.20(+1.20%)
Variant B×✓3.51(+3.24%)4.05(+2.02%)4.25(+2.41%)
Ours✓✓3.52(+3.53%)4.09(+3.02%)4.29(+3.37%)
", + "bbox": [ + 94, + 155, + 480, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ca03b2eefc6cf1fc342a221b9e27a7e88c4618aedad881d95e71748545f39c64.jpg", + "table_caption": [ + "Table 3. Performance comparison of EventGPT-7B across different temporal window numbers $N_{w}$ . The results demonstrate that our framework is robust to a wide range of window numbers. However, both too few and too many windows adversely affect performance, with the optimal result achieved at $N_{w} = 5$ ." + ], + "table_footnote": [], + "table_body": "
NwEvent-Chat
35791113
DC3.473.523.563.413.363.32
CR4.034.094.053.963.973.91
VQA4.214.294.174.124.084.03
", + "bbox": [ + 112, + 337, + 464, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "richer temporal cues but lead to sparser event distribution per bin, making it challenging to model context information effectively. In other words, a small number of windows fail to capture sufficient scene dynamics, while an excessively large number of windows lead to the loss of scene context. The quantitative results in Table 3 reveal that our framework is robust to a wide range of temporal aggregation sizes. Furthermore, a temporal window number of 5 yields the best overall performance. Accordingly, we set $N_w = 5$ in the spatio-temporal aggregator by default.", + "bbox": [ + 88, + 445, + 483, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Downstream Applications", + "text_level": 1, + "bbox": [ + 89, + 604, + 326, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we assess EventGPT on downstream tasks, object detection, and instance segmentation.", + "bbox": [ + 89, + 628, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Detection. To evaluate the generalization ability of EventGPT on object-level understanding and reasoning, we integrate it with GroundingDINO [27], leveraging its open-set object detection capabilities. For each question, EventGPT generates textual descriptions of potential objects in the event streams, which are subsequently used by GroundingDINO to localize the corresponding objects within the data. As illustrated in Fig. 8, EventGPT demonstrates strong object-level reasoning ability, accurately inferring and localizing target objects in the scenario.", + "bbox": [ + 88, + 657, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instance Segmentation. To validate EventGPT's scalability on complex dense prediction tasks, we integrate it with GroundedSAM [37] for instance segmentation on event streams. EventGPT provides semantic cues to guide GroundedSAM in generating instance masks, and segmenting individual objects in the scene. The qualitative results", + "bbox": [ + 88, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f981dbd570fedb50d196455414127bc7461cc325dd57a9d4fbadd7d3ff350661.jpg", + "image_caption": [ + "Figure 8. Object detection [27] using events and text from our EventGPT. Note that, our EventGPT enables the generation of high-quality text for the open-set object detection task." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 903, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a43dd892dbf7d7be4c827f2ca6106103a5468fe2c2a69f0e0aee6be05bcfc73d.jpg", + "image_caption": [ + "Figure 9. Instance segmentation [37] leveraging events and text generated by our EventGPT. Our EventGPT is capable of producing high-quality text descriptions for open-set instance segmentation tasks." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 9 demonstrate our approach's efficacy in achieving segmentation under the challenging conditions inherent to event-based data. This affirms EventGPT's adaptability and potential in fine-grained scene understanding tasks.", + "bbox": [ + 511, + 498, + 906, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 571, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces EventGPT, the first multimodal large language model(MLLM) tailored for comprehensive event stream understanding. The model comprises an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. It employs a three-stage training pipeline to progressively bridge the significant domain gap between event streams and language for enhanced scene perception and description. Comprehensive evaluations consistently demonstrate the superiority of our method in event scene understanding. Additionally, we build two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, to advance cross-modal alignment research in the community.", + "bbox": [ + 511, + 597, + 908, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 806, + 671, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0102801, as well as the National Natural Science Foundation of China under Grant 61827804.", + "bbox": [ + 511, + 832, + 906, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29146", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NIPS, 35: 23716-23736, 2022. 3", + "[2] Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moratelli, Sara Sarto, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. The (r) evolution of multimodal large language models: A survey. arXiv preprint arXiv:2402.12451, 2024.3", + "[3] Tom Tongjia Chen, Hongshan Yu, Zhengeng Yang, Ming Li, Zechuan Li, Jingwen Wang, Wei Miao, Wei Sun, and Chen Chen. First place solution to the cvpr'2023 aqtc challenge: A function-interaction centric approach with spatiotemporal visual-language alignment. arXiv preprint arXiv:2306.13380, 2023.", + "[4] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In CVPR, pages 14239-14250, 2024. 3", + "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 6, 7", + "[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2", + "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 1", + "[8] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1", + "[9] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. *IJCV*, 128(3):601-618, 2020. 1", + "[10] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1", + "[11] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE RAL, 6(3):4947-4954, 2021. 4, 6", + "[12] Kunping Huang, Sen Zhang, Jing Zhang, and Dacheng Tao. Event-based simultaneous localization and mapping: A comprehensive survey. arXiv preprint arXiv:2304.09793, 2023. 1", + "[13] Qirui Jiao, Daoyuan Chen, Yilun Huang, Yaliang Li, and Ying Shen. Img-diff: Contrastive data synthesis for multimodal large language models. arXiv preprint arXiv:2408.04594, 2024. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, pages 349-364. Springer, 2016. 1", + "[15] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, pages 2146-2156, 2021. 3, 6", + "[16] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI, 45(11):14020-14037, 2023. 1", + "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3", + "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3", + "[19] Ming Li, Xinming Huang, and Ziming Zhang. Self-supervised geometric features discovery via interpretable attention for vehicle re-identification and beyond. In Proceedings of the IEEE/CVF international conference on computer vision, pages 194–204, 2021. 3", + "[20] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2021.", + "[21] Ming Li, Huazhu Fu, Shengfeng He, Hehe Fan, Jun Liu, Jussi Keppo, and Mike Zheng Shou. Dr-fer: Discriminative and robust representation learning for facial expression recognition. IEEE Transactions on Multimedia, 26:6297-6309, 2023.", + "[22] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. Stprivacy: Spatio-temporal privacy-preserving action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5106-5115, 2023.", + "[23] Ming Li, Pan Zhou, Jia-Wei Liu, Jussi Keppo, Min Lin, Shuicheng Yan, and Xiangyu Xu. Instant3d: instant text-to-3d generation. International Journal of Computer Vision, 132(10):4456-4472, 2024. 3", + "[24] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 3", + "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1", + "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 36, 2024. 1, 3, 6, 7", + "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8", + "[28] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29147", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6, 7", + "[29] Yecheng Lyu, Ming Li, Xinming Huang, Ulkuhan Guler, Patrick Schaumont, and Ziming Zhang. Treernn: Topology-preserving deep graph embedding and learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7493-7499. IEEE, 2021. 3", + "[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6", + "[31] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In CVPR, pages 5642-5651, 2023. 1", + "[32] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 3DV, pages 1165-1174. IEEE, 2021. 1", + "[33] Rohan Pandey, Rulin Shao, Paul Pu Liang, Ruslan Salakhutdinov, and Louis-Philippe Morency. Cross-modal attention congruence regularization for vision-language relation alignment. arXiv preprint arXiv:2212.10549, 2022. 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2", + "[35] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. *IJCV*, 126 (12):1394–1414, 2018. 1", + "[36] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. CVPR, 2019. 4, 6", + "[37] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024.8", + "[38] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In CVPR, pages 13193-13203, 2024. 1, 3", + "[39] Yichen Shen, Yijin Li, Shuo Chen, Guanglin Li, Zhaoyang Huang, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinktrack: Feature tracking over 100 fps via events and images. arXiv preprint arXiv:2409.17981, 2024. 1", + "[40] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, pages 15638-15650, 2022. 3", + "[41] Lulu Tian, Hongxun Yao, and Ming Li. Fakepei: A large-scale fake person of interest video detection benchmark and a strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6819-6831, 2023. 3", + "[42] Guanqun Wang, Jiaming Liu, Chenxuan Li, Yuan Zhang, Junpeng Ma, Xinyu Wei, Kevin Zhang, Maurice Chong," + ], + "bbox": [ + 91, + 90, + 482, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Renrui Zhang, Yijiang Liu, et al. Cloud-device collaborative learning for multimodal large language models. In CVPR, pages 12646-12655, 2024. 3", + "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6, 7", + "[44] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 3", + "[45] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2, 3", + "[46] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. In CVPR, pages 16933-16943, 2024. 1", + "[47] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024.1", + "[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3", + "[49] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In CVPR, pages 9602–9611, 2024. 1", + "[50] Yun Yue, Ming Li, Venkatesh Saligrama, and Ziming Zhang. Rnn training along locally optimal trajectories via frankwolfe algorithm. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 10532-10539. IEEE, 2021. 3", + "[51] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, et al. Anygpt: Unified multimodal llm with discrete sequence modeling. arXiv preprint arXiv:2402.12226, 2024.3", + "[52] Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, and Dong Yu. Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601, 2024. 3", + "[53] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, pages 8801-8810, 2022. 1", + "[54] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 3", + "[55] Yan Zhang, Chun Li, Zhaoxia Liu, and Ming Li. Semi-supervised disease classification based on limited medical image data. IEEE Journal of Biomedical and Health Informatics, 28(3):1575-1586, 2024. 3", + "[56] Ce Zheng, Yecheng Lyu, Ming Li, and Ziming Zhang. Lodonet: A deep neural network with 2d keypoint matching" + ], + "bbox": [ + 516, + 90, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29148", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for 3d lidar odometry estimation. In Proceedings of the 28th ACM international conference on multimedia, pages 2391-2399, 2020. 3", + "[57] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arxiv 2023. arXiv preprint arXiv:2302.08890, 2020. 1", + "[58] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. E-clip: Towards label-efficient event-based open-world understanding by clip. arXiv e-prints, pages arXiv-2308, 2023. 2, 3", + "[59] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In CVPR, pages 18633-18643, 2024. 3", + "[60] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3", + "[61] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 444 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29149", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_model.json b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8c7e21669f62bbef95c308c70340cb10f4f38f07 --- /dev/null +++ b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_model.json @@ -0,0 +1,2631 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.795, + 0.177 + ], + "angle": 0, + "content": "EventGPT: Event Stream Understanding with Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.203, + 0.802, + 0.241 + ], + "angle": 0, + "content": "Shaoyu Liu\\(^{1,2}\\), Jianing Li\\(^{2}\\), Guanghui Zhao\\(^{1*}\\), Yunjian Zhang\\(^{2}\\), Xin Meng\\(^{3}\\), Fei Richard Yu\\(^{4}\\), Xiangyang Ji\\(^{2*}\\), Ming Li\\(^{4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.256, + 0.81, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\)Xidian University \\(^{2}\\)Tsinghua University \\(^{3}\\)Peking University \\(^{4}\\)Guangdong Laboratory of Artificial Intelligence and Digital Economy(SZ)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.296, + 0.916, + 0.31 + ], + "angle": 0, + "content": "23171110721@stu.xidian.edu.cn, lijianing@pku.edu.cn, ming.li@u.nus.edu, ghzhao@xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.486, + 0.786 + ], + "angle": 0, + "content": "Event cameras capture visual information as asynchronous pixel change streams, excelling in challenging lighting and high-dynamic scenarios. Existing multimodal large language models (MLLMs) concentrate on natural RGB images, failing in scenarios where event data fits better. In this paper, we introduce EventGPT, the first MLLM for event stream understanding, pioneering the integration of large language models (LLMs) with event-based vision. To bridge the huge domain gap, we propose a three-stage optimization paradigm to progressively equip a pre-trained LLM with event understanding. Our EventGPT consists of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Firstly, GPT-generated RGB image-text pairs warm up the linear projector, following LLaVA, as the gap between natural images and language is smaller. Secondly, we construct N-ImageNet-Chat, a large synthetic dataset of event data and corresponding texts to enable the use of the spatio-temporal aggregator and to train the event-language adapter, thereby aligning event features more closely with the language space. Finally, we gather an instruction dataset, EventChat, which contains extensive real-world data to fine-tune the entire model, further enhancing its generalization ability. We construct a comprehensive benchmark, and experiments show that EventGPT surpasses previous state-of-the-art MLLMs in generation quality, descriptive accuracy, and reasoning capability. Code: EventGPT" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.817, + 0.222, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Event cameras, with predominant advantages in capturing scene changes and temporal resolution, have great ap" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.347, + 0.905, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.52, + 0.908, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Our EventGPT is the first multimodal large language model tailored for event stream understanding, including scene summarization, reasoning, and question answering, with a great potential for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.686 + ], + "angle": 0, + "content": "plication potential in scenarios under imperfect lighting conditions or including high dynamic objects [7, 12, 57]. The relevant tasks on event data, e.g. object detection [8, 10, 16, 46], tracking [9, 31, 39, 53], and 3D reconstruction [14, 32, 35, 49], have been well-investigated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Recent multimodal large language models (MLLMs) have shown remarkable breakthroughs in conventional vision and language problems, demonstrating extraordinary capabilities of scene understanding and question answering (QA) [25, 26, 38, 47, 60]. Typically, a pioneering research Large Language and Vision Assistant (LLaVA) combines a vision encoder with an LLM, trained on GPT-4-generated multimodal data, to create a robust system that excels in visual understanding, instruction-following, and complex question answering [26]. These works mainly pay attention to natural RGB scenes and the resulting models degrade significantly when encountering the aforementioned scenarios. However, the relevant research on event data streams remains underexplored." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29139" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.101, + 0.095, + 0.149, + 0.134 + ], + "angle": 0, + "content": "#" + }, + { + "type": "header", + "bbox": [ + 0.176, + 0.102, + 0.88, + 0.128 + ], + "angle": 0, + "content": "EventGPT: This image depicts a person puncturing a balloon with a needle, causing a sudden, dramatic eruption of water. The water splashes out in a forceful burst, creating intricate trails and droplets that scatter through the air in all directions ..." + }, + { + "type": "title", + "bbox": [ + 0.377, + 0.152, + 0.621, + 0.167 + ], + "angle": 0, + "content": "Large Language Model (Vicuna v1.5)" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.174, + 0.891, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.432, + 0.908, + 0.475 + ], + "angle": 0, + "content": "Figure 2. Overview of our framework. The event encoder transforms raw event tensors into high-dimensional features, which are then aggregated by a spatio-temporal module. These representations are projected and aligned with the large language model, enabling nuanced understanding of event streams and supporting various downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.501, + 0.484, + 0.728 + ], + "angle": 0, + "content": "Some works have initialized the alignment between event data and textual information. For example, EventCLIP [45] adapts the pre-trained CLIP model [6, 34] for event-based object recognition, enabling zero-shot and few-shot classification. Similarly, EventBind [58] extends CLIP with an event encoder to align images, texts, and events in a unified representation space, realizing few-shot recognition and event retrieval. Nevertheless, both EventCLIP and EventBind lack the ability to leverage the embedded world knowledge in LLMs, limiting their capacity for nuanced scene understanding and dialog generation in complex scenarios. This gap highlights the necessity of a model, akin to LLaVA for RGB images, specifically designed to overcome these limitations by harnessing the power of LLMs for comprehensive event stream understanding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this work, we pioneer the development of an MLLM for event stream understanding, termed EventGPT (see Fig 1), designed for the community to support scene summarization, reasoning, and question answering. To this end, several key challenges must be addressed. First of all, aligning event features with the high-level semantics of natural language is significantly challenging, as event streams are designed to capture asynchronous and sparse brightness changes within the field of view, inherently lacking the global scene context necessary for language-based understanding. Secondly, the fundamental difference in how" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.592 + ], + "angle": 0, + "content": "event and RGB cameras capture scene information complicates the straightforward adaptation of existing multimodal techniques, which are typically designed for natural images. Finally, the scarcity of large-scale, high-quality event-text datasets hinders the training of a scalable MLLM for comprehensive event scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address these challenges, we propose a three-stage training paradigm that leverages an LLM as an anchor, given its rich world knowledge and a strong foundation for scene understanding, to progressively align the event modality with language. In general, our EventGPT is mainly composed of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Compared to event data, natural images are much closer to language in the embedding space. Therefore, in the first stage, we use the image-text dataset from LLaVA to warm up the linear projector, establishing preliminary scene-language associations. In the second stage, we focus on training the spatio-temporal aggregator and event-language adapter to map spatio-temporal information from event frames into language space. For this purpose, we construct a large-scale synthetic event-text dataset, NImageNet-Chat, to facilitate optimization. Finally, we conduct instructional fine-tuning of all model parameters on a self-collected, high-quality dataset, Event-Chat, which includes diverse scenes with unsatisfactory lighting condi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29140" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "tions or high-speed object motions, enhancing the capability of our EventGPT to interpret complex real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.483, + 0.332 + ], + "angle": 0, + "content": "We introduce a comprehensive evaluation benchmark inspired by Qwen2-72B-Instruct to assess the generation quality, descriptive accuracy, and reasoning capability of EventGPT. To facilitate a rigorous comparison, we select several state-of-the-art MLLMs excelling in natural scene understanding, including LLaVA [26], Qwen2-VL [43], and Deepseek-VL [28], as baseline models. Evaluation is conducted across three metrics, i.e. Detail Captioning (DC), Complex Reasoning (CR), and Visual Question Answering (VQA), in line with the existing literature. Experimental results highlight significant improvements achieved by EventGPT. Furthermore, our model demonstrates strong generalization ability on open-set downstream tasks, such as object detection and semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.363 + ], + "angle": 0, + "content": "Our main contributions in this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.364, + 0.483, + 0.439 + ], + "angle": 0, + "content": "- We present EventGPT, the first multimodal large language model tailored for event stream understanding, demonstrating capabilities in scene summarization, reasoning, and question answering. EventGPT establishes a solid baseline, paving the way for future research." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.44, + 0.483, + 0.514 + ], + "angle": 0, + "content": "- We propose a systematic learning paradigm incorporating specialized training strategies and network modules to progressively bridge the substantial domain gap between event data and language, providing a valuable reference for future cross-modal learning efforts." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.515, + 0.484, + 0.605 + ], + "angle": 0, + "content": "- We construct two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, containing 1 million synthetic samples for pre-training and 120,000 instruction samples for fine-tuning. These datasets will be publicly released to drive progress in multimodal learning within the event domain." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.364, + 0.484, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.241, + 0.635 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Multimodal Large Language Models. In deep learning era[19-23, 29, 41, 50, 55, 56], significant advancements have been made in various domains. The significant advancements in LLMs for natural language understanding have driven their integration into multimodal research domains, with a focus on enhancing vision-language alignment to expand their applicability [2-4, 24, 33, 38, 40, 42, 48, 52, 54]. Early contributions, such as Flamingo [1] and BLIP [17, 18], integrated visual and language modalities to enable zero-shot generalization across a broad range of visual-language tasks. Further advancements emerged with chat-oriented MLLMs like MiniGPT-4 [60] and LLaVA [26], which improved the interactive capabilities and adaptability of LLMs in multimodal applications, setting new benchmarks in performance. As MLLMs continue to evolve, the exploration of novel sensor modalities holds potential for expanding their functional scope. De" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.096, + 0.669, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.092, + 0.901, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.252, + 0.905, + 0.295 + ], + "angle": 0, + "content": "Figure 3. Data distribution across three progressive training stages in our EventGPT containing image-language alignment, event-language alignment, and instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.394 + ], + "angle": 0, + "content": "spite the distinct advantages of event cameras, such as high temporal resolution and robustness in challenging lighting conditions, the integration of event-based data with LLMs remains underexplored, particularly for applications in high-dynamic and low-light environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.906, + 0.651 + ], + "angle": 0, + "content": "Event-Driven Multimodal Models. The integration of event-based data into VLMs has attracted significant interest, demonstrating potential across various downstream tasks and highlighting the advantages of event-language synergy. [45, 58, 59]. Notably, frameworks such as ExACT have leveraged language-guided frameworks for event-based action recognition, effectively aligning action sequences with contextual textual cues to achieve strong performance [59]. Other approaches, such as EventCLIP [45] and EventBind [58], have also demonstrated strong zero-shot performance, underscoring the effectiveness of event-based embeddings when aligned with textual information. Despite these advances, developing chat-centric LLMs capable of incorporating event-based modalities presents significant research challenges. Addressing this gap could enhance models' contextual comprehension, enabling more effective processing of complex interactive tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.776, + 0.68 + ], + "angle": 0, + "content": "3. Event-Text Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Event-based datasets have been extensively studied in computer vision tasks. However, large-scale Event-Text datasets suitable for training MLLMs) are not yet widely available. Previous studies have successfully employed large models, such as GPT, to synthesize data for multimodal applications, yielding significant advancements in model capabilities [13, 44, 51, 61]. To address this gap, we construct an Event-Text dataset using a human-assisted, semi-automated pipeline that incorporates both synthetic and real-world event data. The distribution of the dataset is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.686, + 0.864 + ], + "angle": 0, + "content": "3.1. Synthetic Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "ImageNet serves as a foundational benchmark for visual recognition tasks. N-ImageNet [15] extends this founda" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29141" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.097, + 0.362, + 0.111 + ], + "angle": 0, + "content": "Instruction Examples" + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.118, + 0.158, + 0.129 + ], + "angle": 0, + "content": "Prompt:" + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.132, + 0.201, + 0.142 + ], + "angle": 0, + "content": "Caption Prompt:" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.143, + 0.376, + 0.165 + ], + "angle": 0, + "content": "You are an expert in image description and understanding. You are requested ..." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.166, + 0.218, + 0.178 + ], + "angle": 0, + "content": "Reasoning Prompt:" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.179, + 0.378, + 0.201 + ], + "angle": 0, + "content": "You are an advanced multimodal model with the ability to perform complex reasoning ..." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.203, + 0.215, + 0.215 + ], + "angle": 0, + "content": "Visual QA Prompt:" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.215, + 0.376, + 0.237 + ], + "angle": 0, + "content": "You are an AI language model designed to generate detailed and insightful Visual Question Answering...." + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.134, + 0.471, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.187, + 0.471, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.245, + 0.155, + 0.255 + ], + "angle": 0, + "content": "Results:" + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.259, + 0.156, + 0.269 + ], + "angle": 0, + "content": "Caption:" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.27, + 0.327, + 0.28 + ], + "angle": 0, + "content": "Query: Describe the buildings in the image." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.281, + 0.418, + 0.292 + ], + "angle": 0, + "content": "Answer: There are several multi-story buildings in the image..." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.299, + 0.162, + 0.31 + ], + "angle": 0, + "content": "Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.31, + 0.463, + 0.331 + ], + "angle": 0, + "content": "Query: If the construction barriers were removed, how would it impact the traffic flow at the intersection?" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.332, + 0.43, + 0.343 + ], + "angle": 0, + "content": "Answer: Without the barriers, vehicles would have a direct path.." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.35, + 0.161, + 0.36 + ], + "angle": 0, + "content": "Visual QA" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.361, + 0.468, + 0.372 + ], + "angle": 0, + "content": "Query: How many cars are visible, and what directions are they moving?" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.372, + 0.445, + 0.384 + ], + "angle": 0, + "content": "Answer: Three cars are visible. The one on the left is turning away.." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.443 + ], + "angle": 0, + "content": "Figure 4. Instruction examples from the Event-Chat dataset for multimodal task prompts and responses in image captioning, visual reasoning, and visual question answering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.546 + ], + "angle": 0, + "content": "tion by providing a synthetic event-based dataset generated through event simulators, aligned with ImageNet images. Using this alignment, we generate textual instructions for the images to enable cross-domain learning between visual and event-based data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.623 + ], + "angle": 0, + "content": "N-ImageNet-Chat. We present a large-scale dataset comprising over 1,000,000 instances with concise instructions, primarily designed to align event data with natural language. This dataset provides a foundation for the model's understanding and supports robust representation learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.715 + ], + "angle": 0, + "content": "N-ImageNet-Instruction. We establish an event-based dataset of over 69,000 entries, designed to improve the model's responsiveness to complex, task-specific instructions. The dataset includes captions, visual question answering (VQA), and complex reasoning tasks aimed at finetuning the model's multimodal capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.28, + 0.74 + ], + "angle": 0, + "content": "3.2. Real-World Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In addition to synthetic datasets, we establish a real-world dataset to improve model generalization for practical applications. A human-assisted and semi-automated annotation process ensures high-quality labels for complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Event Data for Event-Chat. We leverage real-world event datasets, such as DSEC [11] and E2VID [36], which are primarily designed for autonomous driving and provide synchronized image-event data under challenging conditions, including low-light environments and high-speed motion (e.g., entering/exitting tunnels, night scenes)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "Instructions for Event-Chat. This dataset is designed to enhance the model's generalization capabilities, consisting of 59,000 high-quality annotated entries. Three distinct prompting strategies (i.e., captioning, VQA, and complex reasoning) form the basis of the Event-Chat dataset, providing diverse and comprehensive instructions. As shown in Fig. 4, examples from each prompting strategy illustrate the dataset's richness in contextual understanding. Combined with N-ImageNet-instruction, these annotated entries form the final Event-Chat dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.906, + 0.349 + ], + "angle": 0, + "content": "In summary, our newly built event-text dataset contains over 1,000,000 entries from both synthetic and real-world data, offering a robust foundation for advancing multimodal understanding in both event-based and textual domains. These fundamental and challenging datasets are intended to foster progress in the development of MLLMs within the neuromorphic community." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.627, + 0.377 + ], + "angle": 0, + "content": "4. EventGPT" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.624, + 0.401 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.606 + ], + "angle": 0, + "content": "EventGPT aims to serve as an event-based MLLM capable of understanding and generating responses based on spatio-temporal event data. As illustrated in Fig. 2, EventGPT formulates event data processing as an event-driven approach to question answering and descriptive generation. By leveraging the high temporal resolution and extended dynamic range inherent in event stream data, EventGPT enhances understanding in scenarios traditionally challenging for standard vision models, such as low-light or high-speed motion conditions. At the core of EventGPT lies an LLM that synthesizes complex spatio-temporal information from event streams, utilizing world knowledge to generate coherent, human-readable outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.758 + ], + "angle": 0, + "content": "In general, an event bin can be mathematically described within a time window \\( T \\) as \\( S = \\{(x_{j},y_{j},t_{j},p_{j})\\}_{j = 1}^{N} \\), where each event consists of spatial coordinates \\( (x_{j},y_{j}) \\), timestamp \\( t_j \\), and polarity \\( p_j \\). Initially, high-dimensional event features \\( V \\) are extracted from \\( S \\) using an event encoder \\( f_{\\mathrm{enc}} \\), initialized with OpenCLIP ViT-L/14-336px to leverage its robust feature extraction capabilities. These features \\( V \\) are then processed through a spatio-temporal encoder \\( f_{\\mathrm{ST}} \\) to capture dynamic interactions, resulting in a comprehensive representation \\( E \\):" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.772, + 0.905, + 0.789 + ], + "angle": 0, + "content": "\\[\nV = f _ {\\text {e n c}} (S), \\quad E = f _ {\\mathrm {S T}} (V) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Subsequently, a two-stage projector is employed to achieve cross-modal alignment. The first stage aligns the spatio-temporally encoded features \\( E \\) with a textual representation, and the second stage adapts this representation specifically to the event modality, resulting in an event-aligned feature \\( F \\). Given a user query \\( Q \\), the large language model \\( f_{\\mathrm{llm}} \\) performs reasoning over \\( E \\) to generate a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29142" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.175, + 0.107 + ], + "angle": 0, + "content": "response \\(A\\)" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.123, + 0.483, + 0.141 + ], + "angle": 0, + "content": "\\[\nF = f _ {\\text {p r o j}} \\left(f _ {\\text {a l i g n}} (E)\\right), \\quad A = f _ {\\mathrm {l l m}} (F \\mid Q) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.147, + 0.484, + 0.222 + ], + "angle": 0, + "content": "EventGPT creates a strong representation of event streams and incorporates task-specific heads to support downstream tasks like object detection and instance segmentation. This flexibility enhances its applicability to a wide range of complex vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.233, + 0.476, + 0.25 + ], + "angle": 0, + "content": "4.2. Spatio-Temporal Information Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.363 + ], + "angle": 0, + "content": "The event stream provides rich spatiotemporal cues crucial for precise temporal and spatial modeling, especially in high-speed motion scenarios, to enhance a model's contextual understanding. The continuous event stream is typically segmented into discrete temporal bins for efficient processing. Each temporal bin \\( E_{t} \\) is independently encoded, forming discrete spatiotemporal feature representations:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.376, + 0.483, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} = \\left\\{f _ {t} ^ {e} \\right\\} _ {t = 1, 2, \\dots} ^ {T}, \\quad \\mathcal {Z} \\in \\mathbb {R} ^ {T \\times S \\times D} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.447 + ], + "angle": 0, + "content": "where \\(S\\) denotes the spatial resolution of the encoded features, and \\(D\\) represents the feature dimensionality after encoding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.509 + ], + "angle": 0, + "content": "To aggregate salient temporal and spatial information, we perform separate temporal and spatial encoding operations. Specifically, Temporal Encoding (TE) aggregates features along the temporal dimension via average pooling:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.521, + 0.483, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} _ {T} ^ {a v g} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} f _ {t} ^ {e}, \\quad \\mathcal {Z} _ {T} ^ {a v g} \\in \\mathbb {R} ^ {S \\times D} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.6 + ], + "angle": 0, + "content": "Similarly, Spatial Encoding (SE) aggregates features along the spatial dimension via average pooling:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.612, + 0.483, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} _ {S} ^ {a v g} = \\frac {1}{S} \\sum_ {t = 1} ^ {S} f _ {t} ^ {e}, \\quad \\mathcal {Z} _ {S} ^ {a v g} \\in \\mathbb {R} ^ {T \\times D} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Finally, the temporally and spatially aggregated representations are concatenated along the feature dimension, forming a robust spatiotemporal embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.731, + 0.483, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathcal {Z}}} = \\operatorname {C o n c a t} \\left(\\mathcal {Z} _ {T} ^ {a v g}, \\mathcal {Z} _ {S} ^ {a v g}\\right), \\quad \\overline {{\\mathcal {Z}}} \\in \\mathbb {R} ^ {(T + S) \\times D} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.807 + ], + "angle": 0, + "content": "The fused representation \\(\\overline{Z}\\) is an end-to-end learned spatiotemporal feature from asynchronous events, forming a robust event embedding from various complex scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.268, + 0.833 + ], + "angle": 0, + "content": "4.3. Training Pipelines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Given the huge domain gap between event and language, we include a visual-language pre-training to initialize our framework. As shown in Fig. 5, EventGPT adopts a three-phase training framework encompassing visual-language," + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.103 + ], + "angle": 0, + "content": "Stage 1: Visual-Language Alignment Stage 2: Event-Language Alignment Stage 3: Instruction Finetuning" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.104, + 0.905, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.272, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 5. Our three-stage pipeline consists of (1) training a linear projector to achieve visual-language alignment, (2) incorporating a spatio-temporal aggregator and an event-language adapter to enable event-language alignment, and (3) performing full fine-tuning of all modules to facilitate cohesive multimodal learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.906, + 0.442 + ], + "angle": 0, + "content": "event-language, and instruction tuning. This structured progression facilitates effective cross-domain alignment, enhancing EventGPT's capacity for multimodal reasoning and understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Visual-Language Adapter. In the visual-language alignment stage, we leverage the LLaVA-Pre-train dataset while keeping the event encoder and LLM modules frozen, focusing exclusively on training the linear projector. Visual-language alignment is prioritized due to the relatively small domain differences between image and text, compared to the distinct nature of event and textual data. This setup establishes initial cross-modal alignment between vision and language, enabling the LLM to interpret scene information and providing a foundational multimodal understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.906, + 0.753 + ], + "angle": 0, + "content": "Event-Language Adapter. In the event-language alignment stage, we leverage our constructed N-ImageNet-Chat dataset, with the event encoder, linear projector, and LLM modules kept frozen, focusing exclusively on training the event-language adapter module. This stage aims to bridge the domain gap between event and language modalities. By aligning event embeddings with the LLM's representational space, the model's ability to interpret event streams and generate descriptive narratives is significantly enhanced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Instruction Tuning. In the instruction tuning stage, all trainable parameters, including the event encoder, linear projector, event-language adapter, and LLM modules, are unfrozen to enable supervised fine-tuning. Trained on our high-quality Event-Chat question-answering dataset, this stage further enhances the model's understanding and generation capabilities for event stream data. This structured, three-phase training framework progressively narrows the domain gap between event and language modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29143" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performance comparison between EventGPT and five state-of-the-art methods on the N-ImageNet-Chat and Event-Chat datasets across three evaluation metrics. EventGPT consistently demonstrates competent performance on all datasets." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.129, + 0.905, + 0.273 + ], + "angle": 0, + "content": "
ModelsLLM BackboneParamsEncoderN-ImageNet-ChatEvent-Chat
DCCRVQADCCRVQA
LLaVA-7B-v1.5 [26]Vicuna-v1.57BCLIP ViT-L/336PX1.541.071.882.204.043.26
Qwen2-VL-7B [43]Qwen27BOpenCLIP ViT-bigG1.741.461.912.384.022.91
Intern2VL-8B [5]InternLM2.58BInternViT-300M1.511.872.082.374.003.71
Deepseek-vl-7b [28]Deepseek-LLM7BSAM-B + SigLIP-L1.521.451.842.414.103.37
EventGPT-7BVicuna-v1.57BCLIP ViT-L/336PX2.392.572.233.524.094.29
LLaVA-13B-v1.5 [26]Vicuna-v1.513BCLIP ViT-L/336PX1.751.211.842.344.083.37
EventGPT-13BVicuna-v1.513BCLIP ViT-L/336PX2.412.812.403.404.134.26
" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.295, + 0.241, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.296, + 0.365, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.296, + 0.495, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.296, + 0.626, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.296, + 0.755, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.296, + 0.886, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.374, + 0.584, + 0.388 + ], + "angle": 0, + "content": "Prompt: Describe the scene depicted in the picture and identify the dangers." + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.391, + 0.89, + 0.432 + ], + "angle": 0, + "content": "##LLaVA-7B: The scene depicted is a car driving on a highway, with a tunnel entrance ahead. The car is approaching the large tunnel opening which is currently blocked by a large boulder. The boulder is preventing the car from entering the tunnel and may potentially cause a dangerous situation, as the driver might not be aware of the situation or stop in time to avoid it..." + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.435, + 0.89, + 0.475 + ], + "angle": 0, + "content": "EventGPT -7B: This scene shows a vehicle traveling through a tunnel toward the open road outside. The dangers in this scene include the potential for oncoming traffic from the opposite lane, as well as the reduced visibility due to the tunnel's enclosed space. Drivers are advised to stay alert and prepared for any sudden appearances of vehicles or obstacles." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.484, + 0.476, + 0.497 + ], + "angle": 0, + "content": "Prrs: Are there any vehicles driving at the tunnel exit?" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.503, + 0.568, + 0.516 + ], + "angle": 0, + "content": "```python\n>>> LLaVA-7B : No, there are no vehicles driving at the tunnel exit in the image." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.516, + 0.482, + 0.529 + ], + "angle": 0, + "content": "EventGPT: Yes, there is a vehicle driving at the tunnel exit." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.553, + 0.908, + 0.597 + ], + "angle": 0, + "content": "Figure 6. Representative visualization comparison between EventGPT and other open-source MLLMs models under extreme lighting variations. EventGPT demonstrates superior accuracy in recognizing critical scene details, such as vehicle presence at tunnel exits, showcasing its robustness in challenging lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.216, + 0.628 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.284, + 0.652 + ], + "angle": 0, + "content": "5.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Implementations. We integrate the N-ImageNet[15], Dsec[11], and E2VID[36] datasets with synthesized multimodal conversational datasets, N-ImageNet-Chat and Event-Chat, using a GPT-assistant approach for three-stage model training. Our LLM backbone is Vicuna-v1.5, with a two-layer MLP as the linear projector and a linear layer as the event-language adapter. All models, including 7B and 13B versions, are trained on 8 NVIDIA RTX 6000 Ada GPUs. In the visual-language alignment stage, we use 558k data with a learning rate of \\(2 \\times 10^{-4}\\) and a batch size of 64. For event-language alignment, we use 1000k data with the same learning rate and batch size. In instruction-supervised fine-tuning, we use 120k data, setting the learning rate to \\(2 \\times 10^{-5}\\) and batch size to 32." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. Following the evaluation strategies of LLaVA [26] and VideoGPT [30], we rigorously assess" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.84 + ], + "angle": 0, + "content": "EventGPT's performance against state-of-the-art MLLMs using three core metrics within a zero-shot question-answering setting. Our benchmark, specifically designed for text generation from event streams, is based on the test sets of N-ImageNet-Chat and Event-Chat datasets, which contain question-answer pairs to evaluate the model's abilities in detailed description, complex reasoning, and visual question answering. Generated responses are quantitatively scored on a 1-5 scale across three dimensions: (i) Detailed Captioning (DC), assessing the accuracy and completeness of descriptions to capture key event details; (ii) Complex Reasoning (CR), evaluating the model's capacity for multi-step reasoning and logical inference; and (iii) Visual Question Answering (VQA), measuring precision and contextual understanding in responding to specific queries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.871, + 0.864 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-Art Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. To assess the effectiveness of EventGPT, we report the quantitative evaluation results of its" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.098, + 0.233, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.099, + 0.365, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.099, + 0.498, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.099, + 0.628, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.099, + 0.761, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.099, + 0.894, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.18, + 0.584, + 0.194 + ], + "angle": 0, + "content": "Prrs: Plesse describe this scene and explain what scene or action it depicts." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.197, + 0.895, + 0.249 + ], + "angle": 0, + "content": "##LLaVA-7B: The scene shows a broken piece of glass flying through the air in front of a white sheet. This glass could be from a broken window, creating an unexpected scene. The sheet of white fabric is likely designed to protect the area from the flying glass shards or to catch them if they land. The shards of glass are scattered around the scene, with some closer to the sheet and others further away, creating a sense of danger and urgency." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.254, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Qwen2-VL-7B: The scene depicts a scene where a ceramic object, possibly a bowl or a plate, is being shattered. The object is placed on a white surface, which appears to be a piece of paper or a cloth. The shattered pieces of the ceramic object are scattered around the surface, with some pieces still attached to the main body of the object..." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.298, + 0.895, + 0.336 + ], + "angle": 0, + "content": "EventGPT : This scene depicts the process of a cup breaking. The moment of impact reveals fragments bursting outward, scattering in various directions. Each shard captures the force of the impact, some pieces spinning and tumbling through the air, while a faint dust cloud forms around the destruction..." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 7. Representative comparison of EventGPT and other open-source multimodal language models (MLLMs) in high-speed motion scenarios, demonstrating EventGPT's enhanced capability to capture spatiotemporal details and interpret motion characteristics, including fragment trajectories and impact effects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.565 + ], + "angle": 0, + "content": "7B and 13B versions in Table 1. To maintain an end-to-end architecture and ensure fair comparison, we represent the event stream as an event image under identical input conditions. We compare our model with state-of-the-art MLLMs, including Qwen2-VL [43], DeepSeek-VL [28], Intern2VL [5], and LLaVA [26]. The results demonstrate that EventGPT exhibits strong capabilities in event understanding, which is mainly attributed to our three-stage training framework that gradually bridges the significant gap between events and language." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative Analysis. In addition, we conduct qualitative assessments on selected challenging scenarios from the Event-Chat test set, specifically those involving extreme lighting conditions or high-speed object dynamics (see Fig. 6 and Fig. 7). The results illustrate that our proposed EventGPT consistently outperforms other models in these difficult settings, demonstrating robust event-driven scene understanding capabilities. For high dynamic range scenarios, such as the tunnel scene in Fig. 6, EventGPT successfully identifies the tunnel environment and issues relevant hazard warnings. Other models, however, frequently misinterpret the scene or fail to detect the tunnel due to the extreme light changes at the exit point. In such conditions, EventGPT demonstrates a unique advantage by effectively capturing high-exposure elements at the tunnel exit and generating timely alerts. For high-speed motion scenarios, such as a cup breaking in Fig. 7, EventGPT demonstrates advanced spatio-temporal understanding by capturing the dynamic rupture sequence of the object. In contrast, LLaVA-7B misinterprets the scene with an inaccurate description, while Qwen2-VL-7B lacks spatio-temporal comprehension, producing a static and incomplete depiction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.412, + 0.667, + 0.429 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Contribution of Each Component. To explore the impact of each component, we first construct a baseline similar to LLaVA except adopting event frames in the second training stage. Two variants are constructed by incorporating our event-language adapter or spatio-temporal aggregator. As shown in Table 2, our proposed EventGPT achieves higher performance compared to the baseline and variants. Compared to the baseline, incorporating the spatio-temporal aggregator results in respective improvement of \\(3.24\\%\\), \\(2.02\\%\\), and \\(2.41\\%\\) on the DC, CR, and VQA metrics, significantly enhancing the model's capability of capturing temporal dynamics and spatial relationships within event streams, which facilitates to better interpret complex motion patterns and scene changes. Additionally, the event-language adapter brings improvements of \\(2.35\\%\\), \\(1.26\\%\\), and \\(1.20\\%\\) on the DC, CR, and VQA metrics respectively over the baseline. This highlights the superiority of our training pipeline. We extend the traditional two-stage paradigm to a three-stage fine-grained progress. By progressively bridging the large cross-modal gap, our model achieves more accurate event stream perception and contextual understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Influence of Temporal Aggregation Size. To explore the effect of temporal aggregation size, we conduct an analysis to determine the optimal temporal window number \\( N_{w} \\) within the spatio-temporal aggregator. \\( N_{w} \\) represents the number of temporal bins used to divide a fixed observation period, with each bin aggregating events within its respective interval. Fewer windows may fail to capture fine-grained temporal details, while more windows can capture" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29145" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 2. Performance impact of each component in EventGPT through ablation study. The results demonstrate that both the spatio-temporal aggregator and the event-language adapter contribute significant performance improvements over the baseline." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.156, + 0.482, + 0.246 + ], + "angle": 0, + "content": "
Mode TypeEvent LanguageSpatio TemporalEvent-Chat
DCCRVQA
Baseline××3.403.974.15
Variant A✓×3.48(+2.35%)4.02(+1.26%)4.20(+1.20%)
Variant B×✓3.51(+3.24%)4.05(+2.02%)4.25(+2.41%)
Ours✓✓3.52(+3.53%)4.09(+3.02%)4.29(+3.37%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.257, + 0.485, + 0.33 + ], + "angle": 0, + "content": "Table 3. Performance comparison of EventGPT-7B across different temporal window numbers \\( N_{w} \\). The results demonstrate that our framework is robust to a wide range of window numbers. However, both too few and too many windows adversely affect performance, with the optimal result achieved at \\( N_{w} = 5 \\)." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.338, + 0.465, + 0.422 + ], + "angle": 0, + "content": "
NwEvent-Chat
35791113
DC3.473.523.563.413.363.32
CR4.034.094.053.963.973.91
VQA4.214.294.174.124.084.03
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.485, + 0.598 + ], + "angle": 0, + "content": "richer temporal cues but lead to sparser event distribution per bin, making it challenging to model context information effectively. In other words, a small number of windows fail to capture sufficient scene dynamics, while an excessively large number of windows lead to the loss of scene context. The quantitative results in Table 3 reveal that our framework is robust to a wide range of temporal aggregation sizes. Furthermore, a temporal window number of 5 yields the best overall performance. Accordingly, we set \\( N_w = 5 \\) in the spatio-temporal aggregator by default." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.328, + 0.623 + ], + "angle": 0, + "content": "5.4. Downstream Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.659 + ], + "angle": 0, + "content": "In this section, we assess EventGPT on downstream tasks, object detection, and instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Object Detection. To evaluate the generalization ability of EventGPT on object-level understanding and reasoning, we integrate it with GroundingDINO [27], leveraging its open-set object detection capabilities. For each question, EventGPT generates textual descriptions of potential objects in the event streams, which are subsequently used by GroundingDINO to localize the corresponding objects within the data. As illustrated in Fig. 8, EventGPT demonstrates strong object-level reasoning ability, accurately inferring and localizing target objects in the scenario." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Instance Segmentation. To validate EventGPT's scalability on complex dense prediction tasks, we integrate it with GroundedSAM [37] for instance segmentation on event streams. EventGPT provides semantic cues to guide GroundedSAM in generating instance masks, and segmenting individual objects in the scene. The qualitative results" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.905, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.233, + 0.908, + 0.277 + ], + "angle": 0, + "content": "Figure 8. Object detection [27] using events and text from our EventGPT. Note that, our EventGPT enables the generation of high-quality text for the open-set object detection task." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.49 + ], + "angle": 0, + "content": "Figure 9. Instance segmentation [37] leveraging events and text generated by our EventGPT. Our EventGPT is capable of producing high-quality text descriptions for open-set instance segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.561 + ], + "angle": 0, + "content": "in Fig. 9 demonstrate our approach's efficacy in achieving segmentation under the challenging conditions inherent to event-based data. This affirms EventGPT's adaptability and potential in fine-grained scene understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.634, + 0.589 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.797 + ], + "angle": 0, + "content": "This paper introduces EventGPT, the first multimodal large language model(MLLM) tailored for comprehensive event stream understanding. The model comprises an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. It employs a three-stage training pipeline to progressively bridge the significant domain gap between event streams and language for enhanced scene perception and description. Comprehensive evaluations consistently demonstrate the superiority of our method in event scene understanding. Additionally, we build two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, to advance cross-modal alignment research in the community." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.808, + 0.673, + 0.826 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.908, + 0.894 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0102801, as well as the National Natural Science Foundation of China under Grant 61827804." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29146" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NIPS, 35: 23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moratelli, Sara Sarto, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. The (r) evolution of multimodal large language models: A survey. arXiv preprint arXiv:2402.12451, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Tom Tongjia Chen, Hongshan Yu, Zhengeng Yang, Ming Li, Zechuan Li, Jingwen Wang, Wei Miao, Wei Sun, and Chen Chen. First place solution to the cvpr'2023 aqtc challenge: A function-interaction centric approach with spatiotemporal visual-language alignment. arXiv preprint arXiv:2306.13380, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.339, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In CVPR, pages 14239-14250, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[8] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[9] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. *IJCV*, 128(3):601-618, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[11] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE RAL, 6(3):4947-4954, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[12] Kunping Huang, Sen Zhang, Jing Zhang, and Dacheng Tao. Event-based simultaneous localization and mapping: A comprehensive survey. arXiv preprint arXiv:2304.09793, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Qirui Jiao, Daoyuan Chen, Yilun Huang, Yaliang Li, and Ying Shen. Img-diff: Contrastive data synthesis for multimodal large language models. arXiv preprint arXiv:2408.04594, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[14] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, pages 349-364. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[15] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, pages 2146-2156, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.236 + ], + "angle": 0, + "content": "[16] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI, 45(11):14020-14037, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.352 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[19] Ming Li, Xinming Huang, and Ziming Zhang. Self-supervised geometric features discovery via interpretable attention for vehicle re-identification and beyond. In Proceedings of the IEEE/CVF international conference on computer vision, pages 194–204, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[20] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.483, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[21] Ming Li, Huazhu Fu, Shengfeng He, Hehe Fan, Jun Liu, Jussi Keppo, and Mike Zheng Shou. Dr-fer: Discriminative and robust representation learning for facial expression recognition. IEEE Transactions on Multimedia, 26:6297-6309, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.624 + ], + "angle": 0, + "content": "[22] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. Stprivacy: Spatio-temporal privacy-preserving action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5106-5115, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.626, + 0.905, + 0.68 + ], + "angle": 0, + "content": "[23] Ming Li, Pan Zhou, Jia-Wei Liu, Jussi Keppo, Min Lin, Shuicheng Yan, and Xiangyu Xu. Instant3d: instant text-to-3d generation. International Journal of Computer Vision, 132(10):4456-4472, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.683, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[24] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.728, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.904, + 0.799 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 36, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29147" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[29] Yecheng Lyu, Ming Li, Xinming Huang, Ulkuhan Guler, Patrick Schaumont, and Ziming Zhang. Treernn: Topology-preserving deep graph embedding and learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7493-7499. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[31] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In CVPR, pages 5642-5651, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[32] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 3DV, pages 1165-1174. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[33] Rohan Pandey, Rulin Shao, Paul Pu Liang, Ruslan Salakhutdinov, and Louis-Philippe Morency. Cross-modal attention congruence regularization for vision-language relation alignment. arXiv preprint arXiv:2212.10549, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[35] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. *IJCV*, 126 (12):1394–1414, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[36] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. CVPR, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[37] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[38] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In CVPR, pages 13193-13203, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[39] Yichen Shen, Yijin Li, Shuo Chen, Guanglin Li, Zhaoyang Huang, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinktrack: Feature tracking over 100 fps via events and images. arXiv preprint arXiv:2409.17981, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[40] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, pages 15638-15650, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[41] Lulu Tian, Hongxun Yao, and Ming Li. Fakepei: A large-scale fake person of interest video detection benchmark and a strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6819-6831, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[42] Guanqun Wang, Jiaming Liu, Chenxuan Li, Yuan Zhang, Junpeng Ma, Xinyu Wei, Kevin Zhang, Maurice Chong," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Renrui Zhang, Yijiang Liu, et al. Cloud-device collaborative learning for multimodal large language models. In CVPR, pages 12646-12655, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[44] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[45] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[46] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. In CVPR, pages 16933-16943, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[47] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[49] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In CVPR, pages 9602–9611, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[50] Yun Yue, Ming Li, Venkatesh Saligrama, and Ziming Zhang. Rnn training along locally optimal trajectories via frankwolfe algorithm. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 10532-10539. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[51] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, et al. Anygpt: Unified multimodal llm with discrete sequence modeling. arXiv preprint arXiv:2402.12226, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[52] Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, and Dong Yu. Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[53] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, pages 8801-8810, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[54] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[55] Yan Zhang, Chun Li, Zhaoxia Liu, and Ming Li. Semi-supervised disease classification based on limited medical image data. IEEE Journal of Biomedical and Health Informatics, 28(3):1575-1586, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[56] Ce Zheng, Yecheng Lyu, Ming Li, and Ziming Zhang. Lodonet: A deep neural network with 2d keypoint matching" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29148" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "for 3d lidar odometry estimation. In Proceedings of the 28th ACM international conference on multimedia, pages 2391-2399, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[57] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arxiv 2023. arXiv preprint arXiv:2302.08890, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[58] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. E-clip: Towards label-efficient event-based open-world understanding by clip. arXiv e-prints, pages arXiv-2308, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[59] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In CVPR, pages 18633-18643, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[60] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[61] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29149" + } + ] +] \ No newline at end of file diff --git a/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_origin.pdf b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..724ddcb8c229f6f2d21eccd14f9d3ae196b9b566 --- /dev/null +++ b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/bb8d5233-3517-4797-9367-5fe773bd874a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57c7976c1d1b8a073f466eafac54862e22f161b49989a012929c5f165bee87bd +size 7474243 diff --git a/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/full.md b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f5d7fa4d5459c07f51f022b7aeda5da85c2e9e79 --- /dev/null +++ b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/full.md @@ -0,0 +1,371 @@ +# EventGPT: Event Stream Understanding with Multimodal Large Language Models + +Shaoyu Liu $^{1,2}$ , Jianing Li $^{2}$ , Guanghui Zhao $^{1*}$ , Yunjian Zhang $^{2}$ , Xin Meng $^{3}$ , Fei Richard Yu $^{4}$ , Xiangyang Ji $^{2*}$ , Ming Li $^{4*}$ + +$^{1}$ Xidian University $^{2}$ Tsinghua University $^{3}$ Peking University $^{4}$ Guangdong Laboratory of Artificial Intelligence and Digital Economy(SZ) + +23171110721@stu.xidian.edu.cn, lijianing@pku.edu.cn, ming.li@u.nus.edu, ghzhao@xidian.edu.cn + +# Abstract + +Event cameras capture visual information as asynchronous pixel change streams, excelling in challenging lighting and high-dynamic scenarios. Existing multimodal large language models (MLLMs) concentrate on natural RGB images, failing in scenarios where event data fits better. In this paper, we introduce EventGPT, the first MLLM for event stream understanding, pioneering the integration of large language models (LLMs) with event-based vision. To bridge the huge domain gap, we propose a three-stage optimization paradigm to progressively equip a pre-trained LLM with event understanding. Our EventGPT consists of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Firstly, GPT-generated RGB image-text pairs warm up the linear projector, following LLaVA, as the gap between natural images and language is smaller. Secondly, we construct N-ImageNet-Chat, a large synthetic dataset of event data and corresponding texts to enable the use of the spatio-temporal aggregator and to train the event-language adapter, thereby aligning event features more closely with the language space. Finally, we gather an instruction dataset, EventChat, which contains extensive real-world data to fine-tune the entire model, further enhancing its generalization ability. We construct a comprehensive benchmark, and experiments show that EventGPT surpasses previous state-of-the-art MLLMs in generation quality, descriptive accuracy, and reasoning capability. Code: EventGPT + +# 1. Introduction + +Event cameras, with predominant advantages in capturing scene changes and temporal resolution, have great ap + +![](images/6719442a966cd1edfd8fd39e5a4de7f04a29dba2b600e35d10985b6641e0fe6e.jpg) +Figure 1. Our EventGPT is the first multimodal large language model tailored for event stream understanding, including scene summarization, reasoning, and question answering, with a great potential for downstream tasks. + +plication potential in scenarios under imperfect lighting conditions or including high dynamic objects [7, 12, 57]. The relevant tasks on event data, e.g. object detection [8, 10, 16, 46], tracking [9, 31, 39, 53], and 3D reconstruction [14, 32, 35, 49], have been well-investigated. + +Recent multimodal large language models (MLLMs) have shown remarkable breakthroughs in conventional vision and language problems, demonstrating extraordinary capabilities of scene understanding and question answering (QA) [25, 26, 38, 47, 60]. Typically, a pioneering research Large Language and Vision Assistant (LLaVA) combines a vision encoder with an LLM, trained on GPT-4-generated multimodal data, to create a robust system that excels in visual understanding, instruction-following, and complex question answering [26]. These works mainly pay attention to natural RGB scenes and the resulting models degrade significantly when encountering the aforementioned scenarios. However, the relevant research on event data streams remains underexplored. + +# Large Language Model (Vicuna v1.5) + +![](images/43512ca007ee5fe49f1b1bb13e80f993a2aca04578c53ecaedefcbfc99b57551.jpg) +Figure 2. Overview of our framework. The event encoder transforms raw event tensors into high-dimensional features, which are then aggregated by a spatio-temporal module. These representations are projected and aligned with the large language model, enabling nuanced understanding of event streams and supporting various downstream applications. + +Some works have initialized the alignment between event data and textual information. For example, EventCLIP [45] adapts the pre-trained CLIP model [6, 34] for event-based object recognition, enabling zero-shot and few-shot classification. Similarly, EventBind [58] extends CLIP with an event encoder to align images, texts, and events in a unified representation space, realizing few-shot recognition and event retrieval. Nevertheless, both EventCLIP and EventBind lack the ability to leverage the embedded world knowledge in LLMs, limiting their capacity for nuanced scene understanding and dialog generation in complex scenarios. This gap highlights the necessity of a model, akin to LLaVA for RGB images, specifically designed to overcome these limitations by harnessing the power of LLMs for comprehensive event stream understanding. + +In this work, we pioneer the development of an MLLM for event stream understanding, termed EventGPT (see Fig 1), designed for the community to support scene summarization, reasoning, and question answering. To this end, several key challenges must be addressed. First of all, aligning event features with the high-level semantics of natural language is significantly challenging, as event streams are designed to capture asynchronous and sparse brightness changes within the field of view, inherently lacking the global scene context necessary for language-based understanding. Secondly, the fundamental difference in how + +event and RGB cameras capture scene information complicates the straightforward adaptation of existing multimodal techniques, which are typically designed for natural images. Finally, the scarcity of large-scale, high-quality event-text datasets hinders the training of a scalable MLLM for comprehensive event scene understanding. + +To address these challenges, we propose a three-stage training paradigm that leverages an LLM as an anchor, given its rich world knowledge and a strong foundation for scene understanding, to progressively align the event modality with language. In general, our EventGPT is mainly composed of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Compared to event data, natural images are much closer to language in the embedding space. Therefore, in the first stage, we use the image-text dataset from LLaVA to warm up the linear projector, establishing preliminary scene-language associations. In the second stage, we focus on training the spatio-temporal aggregator and event-language adapter to map spatio-temporal information from event frames into language space. For this purpose, we construct a large-scale synthetic event-text dataset, NImageNet-Chat, to facilitate optimization. Finally, we conduct instructional fine-tuning of all model parameters on a self-collected, high-quality dataset, Event-Chat, which includes diverse scenes with unsatisfactory lighting condi + +tions or high-speed object motions, enhancing the capability of our EventGPT to interpret complex real-world scenes. + +We introduce a comprehensive evaluation benchmark inspired by Qwen2-72B-Instruct to assess the generation quality, descriptive accuracy, and reasoning capability of EventGPT. To facilitate a rigorous comparison, we select several state-of-the-art MLLMs excelling in natural scene understanding, including LLaVA [26], Qwen2-VL [43], and Deepseek-VL [28], as baseline models. Evaluation is conducted across three metrics, i.e. Detail Captioning (DC), Complex Reasoning (CR), and Visual Question Answering (VQA), in line with the existing literature. Experimental results highlight significant improvements achieved by EventGPT. Furthermore, our model demonstrates strong generalization ability on open-set downstream tasks, such as object detection and semantic segmentation. + +Our main contributions in this work can be summarized as follows: + +- We present EventGPT, the first multimodal large language model tailored for event stream understanding, demonstrating capabilities in scene summarization, reasoning, and question answering. EventGPT establishes a solid baseline, paving the way for future research. +- We propose a systematic learning paradigm incorporating specialized training strategies and network modules to progressively bridge the substantial domain gap between event data and language, providing a valuable reference for future cross-modal learning efforts. +- We construct two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, containing 1 million synthetic samples for pre-training and 120,000 instruction samples for fine-tuning. These datasets will be publicly released to drive progress in multimodal learning within the event domain. + +# 2. Related Works + +Multimodal Large Language Models. In deep learning era[19-23, 29, 41, 50, 55, 56], significant advancements have been made in various domains. The significant advancements in LLMs for natural language understanding have driven their integration into multimodal research domains, with a focus on enhancing vision-language alignment to expand their applicability [2-4, 24, 33, 38, 40, 42, 48, 52, 54]. Early contributions, such as Flamingo [1] and BLIP [17, 18], integrated visual and language modalities to enable zero-shot generalization across a broad range of visual-language tasks. Further advancements emerged with chat-oriented MLLMs like MiniGPT-4 [60] and LLaVA [26], which improved the interactive capabilities and adaptability of LLMs in multimodal applications, setting new benchmarks in performance. As MLLMs continue to evolve, the exploration of novel sensor modalities holds potential for expanding their functional scope. De + +![](images/0cae54f9b3797762b8bad9629f3d6081e584de57b5a474c711ba49b8a9a02cf0.jpg) +Figure 3. Data distribution across three progressive training stages in our EventGPT containing image-language alignment, event-language alignment, and instruction tuning. + +![](images/771db84bbc6a2744a1c91e6738f97fe73dceb996cd4d3d6cababc94f153fc9ce.jpg) + +spite the distinct advantages of event cameras, such as high temporal resolution and robustness in challenging lighting conditions, the integration of event-based data with LLMs remains underexplored, particularly for applications in high-dynamic and low-light environments. + +Event-Driven Multimodal Models. The integration of event-based data into VLMs has attracted significant interest, demonstrating potential across various downstream tasks and highlighting the advantages of event-language synergy. [45, 58, 59]. Notably, frameworks such as ExACT have leveraged language-guided frameworks for event-based action recognition, effectively aligning action sequences with contextual textual cues to achieve strong performance [59]. Other approaches, such as EventCLIP [45] and EventBind [58], have also demonstrated strong zero-shot performance, underscoring the effectiveness of event-based embeddings when aligned with textual information. Despite these advances, developing chat-centric LLMs capable of incorporating event-based modalities presents significant research challenges. Addressing this gap could enhance models' contextual comprehension, enabling more effective processing of complex interactive tasks. + +# 3. Event-Text Data Generation + +Event-based datasets have been extensively studied in computer vision tasks. However, large-scale Event-Text datasets suitable for training MLLMs) are not yet widely available. Previous studies have successfully employed large models, such as GPT, to synthesize data for multimodal applications, yielding significant advancements in model capabilities [13, 44, 51, 61]. To address this gap, we construct an Event-Text dataset using a human-assisted, semi-automated pipeline that incorporates both synthetic and real-world event data. The distribution of the dataset is shown in Fig. 3. + +# 3.1. Synthetic Dataset + +ImageNet serves as a foundational benchmark for visual recognition tasks. N-ImageNet [15] extends this founda + +# Instruction Examples + +# Prompt: + +# Caption Prompt: + +You are an expert in image description and understanding. You are requested ... + +# Reasoning Prompt: + +You are an advanced multimodal model with the ability to perform complex reasoning ... + +# Visual QA Prompt: + +You are an AI language model designed to generate detailed and insightful Visual Question Answering.... + +![](images/23a7c812e4c04efe921740bb56c657da6dac85aa1fce715ead44fcdd08ca80d3.jpg) +Figure 4. Instruction examples from the Event-Chat dataset for multimodal task prompts and responses in image captioning, visual reasoning, and visual question answering. + +![](images/e4a8e328bf52da19c140c8cc9b70dcc7cdaa140a6d1679f7e6bfd77628d4f13e.jpg) + +# Results: + +# Caption: + +Query: Describe the buildings in the image. + +Answer: There are several multi-story buildings in the image... + +# Reasoning + +Query: If the construction barriers were removed, how would it impact the traffic flow at the intersection? + +Answer: Without the barriers, vehicles would have a direct path.. + +# Visual QA + +Query: How many cars are visible, and what directions are they moving? + +Answer: Three cars are visible. The one on the left is turning away.. + +tion by providing a synthetic event-based dataset generated through event simulators, aligned with ImageNet images. Using this alignment, we generate textual instructions for the images to enable cross-domain learning between visual and event-based data. + +N-ImageNet-Chat. We present a large-scale dataset comprising over 1,000,000 instances with concise instructions, primarily designed to align event data with natural language. This dataset provides a foundation for the model's understanding and supports robust representation learning. + +N-ImageNet-Instruction. We establish an event-based dataset of over 69,000 entries, designed to improve the model's responsiveness to complex, task-specific instructions. The dataset includes captions, visual question answering (VQA), and complex reasoning tasks aimed at finetuning the model's multimodal capabilities. + +# 3.2. Real-World Dataset + +In addition to synthetic datasets, we establish a real-world dataset to improve model generalization for practical applications. A human-assisted and semi-automated annotation process ensures high-quality labels for complex scenes. + +Event Data for Event-Chat. We leverage real-world event datasets, such as DSEC [11] and E2VID [36], which are primarily designed for autonomous driving and provide synchronized image-event data under challenging conditions, including low-light environments and high-speed motion (e.g., entering/exitting tunnels, night scenes). + +Instructions for Event-Chat. This dataset is designed to enhance the model's generalization capabilities, consisting of 59,000 high-quality annotated entries. Three distinct prompting strategies (i.e., captioning, VQA, and complex reasoning) form the basis of the Event-Chat dataset, providing diverse and comprehensive instructions. As shown in Fig. 4, examples from each prompting strategy illustrate the dataset's richness in contextual understanding. Combined with N-ImageNet-instruction, these annotated entries form the final Event-Chat dataset. + +In summary, our newly built event-text dataset contains over 1,000,000 entries from both synthetic and real-world data, offering a robust foundation for advancing multimodal understanding in both event-based and textual domains. These fundamental and challenging datasets are intended to foster progress in the development of MLLMs within the neuromorphic community. + +# 4. EventGPT + +# 4.1. Overview + +EventGPT aims to serve as an event-based MLLM capable of understanding and generating responses based on spatio-temporal event data. As illustrated in Fig. 2, EventGPT formulates event data processing as an event-driven approach to question answering and descriptive generation. By leveraging the high temporal resolution and extended dynamic range inherent in event stream data, EventGPT enhances understanding in scenarios traditionally challenging for standard vision models, such as low-light or high-speed motion conditions. At the core of EventGPT lies an LLM that synthesizes complex spatio-temporal information from event streams, utilizing world knowledge to generate coherent, human-readable outputs. + +In general, an event bin can be mathematically described within a time window $T$ as $S = \{(x_{j},y_{j},t_{j},p_{j})\}_{j = 1}^{N}$ , where each event consists of spatial coordinates $(x_{j},y_{j})$ , timestamp $t_j$ , and polarity $p_j$ . Initially, high-dimensional event features $V$ are extracted from $S$ using an event encoder $f_{\mathrm{enc}}$ , initialized with OpenCLIP ViT-L/14-336px to leverage its robust feature extraction capabilities. These features $V$ are then processed through a spatio-temporal encoder $f_{\mathrm{ST}}$ to capture dynamic interactions, resulting in a comprehensive representation $E$ : + +$$ +V = f _ {\text {e n c}} (S), \quad E = f _ {\mathrm {S T}} (V) \tag {1} +$$ + +Subsequently, a two-stage projector is employed to achieve cross-modal alignment. The first stage aligns the spatio-temporally encoded features $E$ with a textual representation, and the second stage adapts this representation specifically to the event modality, resulting in an event-aligned feature $F$ . Given a user query $Q$ , the large language model $f_{\mathrm{llm}}$ performs reasoning over $E$ to generate a + +response $A$ + +$$ +F = f _ {\text {p r o j}} \left(f _ {\text {a l i g n}} (E)\right), \quad A = f _ {\mathrm {l l m}} (F \mid Q) \tag {2} +$$ + +EventGPT creates a strong representation of event streams and incorporates task-specific heads to support downstream tasks like object detection and instance segmentation. This flexibility enhances its applicability to a wide range of complex vision tasks. + +# 4.2. Spatio-Temporal Information Representation + +The event stream provides rich spatiotemporal cues crucial for precise temporal and spatial modeling, especially in high-speed motion scenarios, to enhance a model's contextual understanding. The continuous event stream is typically segmented into discrete temporal bins for efficient processing. Each temporal bin $E_{t}$ is independently encoded, forming discrete spatiotemporal feature representations: + +$$ +\mathcal {Z} = \left\{f _ {t} ^ {e} \right\} _ {t = 1, 2, \dots} ^ {T}, \quad \mathcal {Z} \in \mathbb {R} ^ {T \times S \times D} \tag {3} +$$ + +where $S$ denotes the spatial resolution of the encoded features, and $D$ represents the feature dimensionality after encoding. + +To aggregate salient temporal and spatial information, we perform separate temporal and spatial encoding operations. Specifically, Temporal Encoding (TE) aggregates features along the temporal dimension via average pooling: + +$$ +\mathcal {Z} _ {T} ^ {a v g} = \frac {1}{T} \sum_ {t = 1} ^ {T} f _ {t} ^ {e}, \quad \mathcal {Z} _ {T} ^ {a v g} \in \mathbb {R} ^ {S \times D} \tag {4} +$$ + +Similarly, Spatial Encoding (SE) aggregates features along the spatial dimension via average pooling: + +$$ +\mathcal {Z} _ {S} ^ {a v g} = \frac {1}{S} \sum_ {t = 1} ^ {S} f _ {t} ^ {e}, \quad \mathcal {Z} _ {S} ^ {a v g} \in \mathbb {R} ^ {T \times D} \tag {5} +$$ + +Finally, the temporally and spatially aggregated representations are concatenated along the feature dimension, forming a robust spatiotemporal embedding: + +$$ +\overline {{\mathcal {Z}}} = \operatorname {C o n c a t} \left(\mathcal {Z} _ {T} ^ {a v g}, \mathcal {Z} _ {S} ^ {a v g}\right), \quad \overline {{\mathcal {Z}}} \in \mathbb {R} ^ {(T + S) \times D} \tag {6} +$$ + +The fused representation $\overline{Z}$ is an end-to-end learned spatiotemporal feature from asynchronous events, forming a robust event embedding from various complex scenarios. + +# 4.3. Training Pipelines + +Given the huge domain gap between event and language, we include a visual-language pre-training to initialize our framework. As shown in Fig. 5, EventGPT adopts a three-phase training framework encompassing visual-language, + +![](images/6bba3818f32e079dce0ec1810d1ec93247808ffa3f61ba5eddb605f9773cf437.jpg) +Stage 1: Visual-Language Alignment Stage 2: Event-Language Alignment Stage 3: Instruction Finetuning +Figure 5. Our three-stage pipeline consists of (1) training a linear projector to achieve visual-language alignment, (2) incorporating a spatio-temporal aggregator and an event-language adapter to enable event-language alignment, and (3) performing full fine-tuning of all modules to facilitate cohesive multimodal learning. + +event-language, and instruction tuning. This structured progression facilitates effective cross-domain alignment, enhancing EventGPT's capacity for multimodal reasoning and understanding. + +Visual-Language Adapter. In the visual-language alignment stage, we leverage the LLaVA-Pre-train dataset while keeping the event encoder and LLM modules frozen, focusing exclusively on training the linear projector. Visual-language alignment is prioritized due to the relatively small domain differences between image and text, compared to the distinct nature of event and textual data. This setup establishes initial cross-modal alignment between vision and language, enabling the LLM to interpret scene information and providing a foundational multimodal understanding. + +Event-Language Adapter. In the event-language alignment stage, we leverage our constructed N-ImageNet-Chat dataset, with the event encoder, linear projector, and LLM modules kept frozen, focusing exclusively on training the event-language adapter module. This stage aims to bridge the domain gap between event and language modalities. By aligning event embeddings with the LLM's representational space, the model's ability to interpret event streams and generate descriptive narratives is significantly enhanced. + +Instruction Tuning. In the instruction tuning stage, all trainable parameters, including the event encoder, linear projector, event-language adapter, and LLM modules, are unfrozen to enable supervised fine-tuning. Trained on our high-quality Event-Chat question-answering dataset, this stage further enhances the model's understanding and generation capabilities for event stream data. This structured, three-phase training framework progressively narrows the domain gap between event and language modalities. + +Table 1. Performance comparison between EventGPT and five state-of-the-art methods on the N-ImageNet-Chat and Event-Chat datasets across three evaluation metrics. EventGPT consistently demonstrates competent performance on all datasets. + +
ModelsLLM BackboneParamsEncoderN-ImageNet-ChatEvent-Chat
DCCRVQADCCRVQA
LLaVA-7B-v1.5 [26]Vicuna-v1.57BCLIP ViT-L/336PX1.541.071.882.204.043.26
Qwen2-VL-7B [43]Qwen27BOpenCLIP ViT-bigG1.741.461.912.384.022.91
Intern2VL-8B [5]InternLM2.58BInternViT-300M1.511.872.082.374.003.71
Deepseek-vl-7b [28]Deepseek-LLM7BSAM-B + SigLIP-L1.521.451.842.414.103.37
EventGPT-7BVicuna-v1.57BCLIP ViT-L/336PX2.392.572.233.524.094.29
LLaVA-13B-v1.5 [26]Vicuna-v1.513BCLIP ViT-L/336PX1.751.211.842.344.083.37
EventGPT-13BVicuna-v1.513BCLIP ViT-L/336PX2.412.812.403.404.134.26
+ +![](images/733cd18ab7858fb825dede73b4c1b10ae8c5e076fcff944aa35558875f528fb4.jpg) +Figure 6. Representative visualization comparison between EventGPT and other open-source MLLMs models under extreme lighting variations. EventGPT demonstrates superior accuracy in recognizing critical scene details, such as vehicle presence at tunnel exits, showcasing its robustness in challenging lighting conditions. + +![](images/4838196664db2922b319b7e8d5cf5650d4bf3eb4202ea15059d7af2d6661a6ce.jpg) + +![](images/6adcdc8d7d4488bf49e7d9f70604a9cb67298a1344250fe8cd8c8a1a5cd8fed2.jpg) + +![](images/e05a64d6f27d68849795d41583c620a046097cb4e62ba2117c04c649e37283b1.jpg) + +![](images/2fa3ec0385e90f9974451546e8e224da59df1e4542302d2a0d315beefae0af59.jpg) + +![](images/8913e1d5c018c8b677747a514934989b6d59403d3c76667d50bc199657ea7193.jpg) + +Prompt: Describe the scene depicted in the picture and identify the dangers. + +##LLaVA-7B: The scene depicted is a car driving on a highway, with a tunnel entrance ahead. The car is approaching the large tunnel opening which is currently blocked by a large boulder. The boulder is preventing the car from entering the tunnel and may potentially cause a dangerous situation, as the driver might not be aware of the situation or stop in time to avoid it... + +EventGPT -7B: This scene shows a vehicle traveling through a tunnel toward the open road outside. The dangers in this scene include the potential for oncoming traffic from the opposite lane, as well as the reduced visibility due to the tunnel's enclosed space. Drivers are advised to stay alert and prepared for any sudden appearances of vehicles or obstacles. + +Prrs: Are there any vehicles driving at the tunnel exit? + +```python +>>> LLaVA-7B : No, there are no vehicles driving at the tunnel exit in the image. + +EventGPT: Yes, there is a vehicle driving at the tunnel exit. + +# 5. Experiment + +# 5.1. Experiment Settings + +Implementations. We integrate the N-ImageNet[15], Dsec[11], and E2VID[36] datasets with synthesized multimodal conversational datasets, N-ImageNet-Chat and Event-Chat, using a GPT-assistant approach for three-stage model training. Our LLM backbone is Vicuna-v1.5, with a two-layer MLP as the linear projector and a linear layer as the event-language adapter. All models, including 7B and 13B versions, are trained on 8 NVIDIA RTX 6000 Ada GPUs. In the visual-language alignment stage, we use 558k data with a learning rate of $2 \times 10^{-4}$ and a batch size of 64. For event-language alignment, we use 1000k data with the same learning rate and batch size. In instruction-supervised fine-tuning, we use 120k data, setting the learning rate to $2 \times 10^{-5}$ and batch size to 32. + +Evaluation Metrics. Following the evaluation strategies of LLaVA [26] and VideoGPT [30], we rigorously assess + +EventGPT's performance against state-of-the-art MLLMs using three core metrics within a zero-shot question-answering setting. Our benchmark, specifically designed for text generation from event streams, is based on the test sets of N-ImageNet-Chat and Event-Chat datasets, which contain question-answer pairs to evaluate the model's abilities in detailed description, complex reasoning, and visual question answering. Generated responses are quantitatively scored on a 1-5 scale across three dimensions: (i) Detailed Captioning (DC), assessing the accuracy and completeness of descriptions to capture key event details; (ii) Complex Reasoning (CR), evaluating the model's capacity for multi-step reasoning and logical inference; and (iii) Visual Question Answering (VQA), measuring precision and contextual understanding in responding to specific queries. + +# 5.2. Comparison with State-of-the-Art Models + +Quantitative Results. To assess the effectiveness of EventGPT, we report the quantitative evaluation results of its + +![](images/4ee100de6a3848a5634026a637a4c9d4cde65b9c2496e34e01d525256a011fa4.jpg) +Figure 7. Representative comparison of EventGPT and other open-source multimodal language models (MLLMs) in high-speed motion scenarios, demonstrating EventGPT's enhanced capability to capture spatiotemporal details and interpret motion characteristics, including fragment trajectories and impact effects. + +![](images/b6f04d867be761646038b6fe13b564360e87764d6332c34a056303e4e993c4bb.jpg) + +![](images/764473f41db5737dd2bd2e90d0a5d2c43578ad8b6992e62573272cbe208a36de.jpg) + +![](images/2438e4da493a505bc4ecbe0663adb1c98a260c9314bc5612d09e4d06e413b48d.jpg) + +![](images/1f15b8c98d9db183636751f57cac1c5a388a3e47ff6c124a5b5ea406aaf4ef53.jpg) + +![](images/5c119ba8c5fbd2b2b7a1cdeec7825d6c026ae03049f388e8e1c76d988db10965.jpg) + +Prrs: Plesse describe this scene and explain what scene or action it depicts. + +##LLaVA-7B: The scene shows a broken piece of glass flying through the air in front of a white sheet. This glass could be from a broken window, creating an unexpected scene. The sheet of white fabric is likely designed to protect the area from the flying glass shards or to catch them if they land. The shards of glass are scattered around the scene, with some closer to the sheet and others further away, creating a sense of danger and urgency. + +Qwen2-VL-7B: The scene depicts a scene where a ceramic object, possibly a bowl or a plate, is being shattered. The object is placed on a white surface, which appears to be a piece of paper or a cloth. The shattered pieces of the ceramic object are scattered around the surface, with some pieces still attached to the main body of the object... + +EventGPT : This scene depicts the process of a cup breaking. The moment of impact reveals fragments bursting outward, scattering in various directions. Each shard captures the force of the impact, some pieces spinning and tumbling through the air, while a faint dust cloud forms around the destruction... + +7B and 13B versions in Table 1. To maintain an end-to-end architecture and ensure fair comparison, we represent the event stream as an event image under identical input conditions. We compare our model with state-of-the-art MLLMs, including Qwen2-VL [43], DeepSeek-VL [28], Intern2VL [5], and LLaVA [26]. The results demonstrate that EventGPT exhibits strong capabilities in event understanding, which is mainly attributed to our three-stage training framework that gradually bridges the significant gap between events and language. + +Qualitative Analysis. In addition, we conduct qualitative assessments on selected challenging scenarios from the Event-Chat test set, specifically those involving extreme lighting conditions or high-speed object dynamics (see Fig. 6 and Fig. 7). The results illustrate that our proposed EventGPT consistently outperforms other models in these difficult settings, demonstrating robust event-driven scene understanding capabilities. For high dynamic range scenarios, such as the tunnel scene in Fig. 6, EventGPT successfully identifies the tunnel environment and issues relevant hazard warnings. Other models, however, frequently misinterpret the scene or fail to detect the tunnel due to the extreme light changes at the exit point. In such conditions, EventGPT demonstrates a unique advantage by effectively capturing high-exposure elements at the tunnel exit and generating timely alerts. For high-speed motion scenarios, such as a cup breaking in Fig. 7, EventGPT demonstrates advanced spatio-temporal understanding by capturing the dynamic rupture sequence of the object. In contrast, LLaVA-7B misinterprets the scene with an inaccurate description, while Qwen2-VL-7B lacks spatio-temporal comprehension, producing a static and incomplete depiction. + +# 5.3. Ablation Study + +Contribution of Each Component. To explore the impact of each component, we first construct a baseline similar to LLaVA except adopting event frames in the second training stage. Two variants are constructed by incorporating our event-language adapter or spatio-temporal aggregator. As shown in Table 2, our proposed EventGPT achieves higher performance compared to the baseline and variants. Compared to the baseline, incorporating the spatio-temporal aggregator results in respective improvement of $3.24\%$ , $2.02\%$ , and $2.41\%$ on the DC, CR, and VQA metrics, significantly enhancing the model's capability of capturing temporal dynamics and spatial relationships within event streams, which facilitates to better interpret complex motion patterns and scene changes. Additionally, the event-language adapter brings improvements of $2.35\%$ , $1.26\%$ , and $1.20\%$ on the DC, CR, and VQA metrics respectively over the baseline. This highlights the superiority of our training pipeline. We extend the traditional two-stage paradigm to a three-stage fine-grained progress. By progressively bridging the large cross-modal gap, our model achieves more accurate event stream perception and contextual understanding. + +Influence of Temporal Aggregation Size. To explore the effect of temporal aggregation size, we conduct an analysis to determine the optimal temporal window number $N_{w}$ within the spatio-temporal aggregator. $N_{w}$ represents the number of temporal bins used to divide a fixed observation period, with each bin aggregating events within its respective interval. Fewer windows may fail to capture fine-grained temporal details, while more windows can capture + +Table 2. Performance impact of each component in EventGPT through ablation study. The results demonstrate that both the spatio-temporal aggregator and the event-language adapter contribute significant performance improvements over the baseline. + +
Mode TypeEvent LanguageSpatio TemporalEvent-Chat
DCCRVQA
Baseline××3.403.974.15
Variant A✓×3.48(+2.35%)4.02(+1.26%)4.20(+1.20%)
Variant B×✓3.51(+3.24%)4.05(+2.02%)4.25(+2.41%)
Ours✓✓3.52(+3.53%)4.09(+3.02%)4.29(+3.37%)
+ +Table 3. Performance comparison of EventGPT-7B across different temporal window numbers $N_{w}$ . The results demonstrate that our framework is robust to a wide range of window numbers. However, both too few and too many windows adversely affect performance, with the optimal result achieved at $N_{w} = 5$ . + +
NwEvent-Chat
35791113
DC3.473.523.563.413.363.32
CR4.034.094.053.963.973.91
VQA4.214.294.174.124.084.03
+ +richer temporal cues but lead to sparser event distribution per bin, making it challenging to model context information effectively. In other words, a small number of windows fail to capture sufficient scene dynamics, while an excessively large number of windows lead to the loss of scene context. The quantitative results in Table 3 reveal that our framework is robust to a wide range of temporal aggregation sizes. Furthermore, a temporal window number of 5 yields the best overall performance. Accordingly, we set $N_w = 5$ in the spatio-temporal aggregator by default. + +# 5.4. Downstream Applications + +In this section, we assess EventGPT on downstream tasks, object detection, and instance segmentation. + +Object Detection. To evaluate the generalization ability of EventGPT on object-level understanding and reasoning, we integrate it with GroundingDINO [27], leveraging its open-set object detection capabilities. For each question, EventGPT generates textual descriptions of potential objects in the event streams, which are subsequently used by GroundingDINO to localize the corresponding objects within the data. As illustrated in Fig. 8, EventGPT demonstrates strong object-level reasoning ability, accurately inferring and localizing target objects in the scenario. + +Instance Segmentation. To validate EventGPT's scalability on complex dense prediction tasks, we integrate it with GroundedSAM [37] for instance segmentation on event streams. EventGPT provides semantic cues to guide GroundedSAM in generating instance masks, and segmenting individual objects in the scene. The qualitative results + +![](images/f981dbd570fedb50d196455414127bc7461cc325dd57a9d4fbadd7d3ff350661.jpg) +Figure 8. Object detection [27] using events and text from our EventGPT. Note that, our EventGPT enables the generation of high-quality text for the open-set object detection task. + +![](images/a43dd892dbf7d7be4c827f2ca6106103a5468fe2c2a69f0e0aee6be05bcfc73d.jpg) +Figure 9. Instance segmentation [37] leveraging events and text generated by our EventGPT. Our EventGPT is capable of producing high-quality text descriptions for open-set instance segmentation tasks. + +in Fig. 9 demonstrate our approach's efficacy in achieving segmentation under the challenging conditions inherent to event-based data. This affirms EventGPT's adaptability and potential in fine-grained scene understanding tasks. + +# 6. Conclusion + +This paper introduces EventGPT, the first multimodal large language model(MLLM) tailored for comprehensive event stream understanding. The model comprises an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. It employs a three-stage training pipeline to progressively bridge the significant domain gap between event streams and language for enhanced scene perception and description. Comprehensive evaluations consistently demonstrate the superiority of our method in event scene understanding. Additionally, we build two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, to advance cross-modal alignment research in the community. + +# Acknowledgments + +This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0102801, as well as the National Natural Science Foundation of China under Grant 61827804. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NIPS, 35: 23716-23736, 2022. 3 +[2] Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moratelli, Sara Sarto, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. The (r) evolution of multimodal large language models: A survey. arXiv preprint arXiv:2402.12451, 2024.3 +[3] Tom Tongjia Chen, Hongshan Yu, Zhengeng Yang, Ming Li, Zechuan Li, Jingwen Wang, Wei Miao, Wei Sun, and Chen Chen. First place solution to the cvpr'2023 aqtc challenge: A function-interaction centric approach with spatiotemporal visual-language alignment. arXiv preprint arXiv:2306.13380, 2023. +[4] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In CVPR, pages 14239-14250, 2024. 3 +[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 6, 7 +[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2 +[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 1 +[8] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1 +[9] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. *IJCV*, 128(3):601-618, 2020. 1 +[10] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1 +[11] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE RAL, 6(3):4947-4954, 2021. 4, 6 +[12] Kunping Huang, Sen Zhang, Jing Zhang, and Dacheng Tao. Event-based simultaneous localization and mapping: A comprehensive survey. arXiv preprint arXiv:2304.09793, 2023. 1 +[13] Qirui Jiao, Daoyuan Chen, Yilun Huang, Yaliang Li, and Ying Shen. Img-diff: Contrastive data synthesis for multimodal large language models. arXiv preprint arXiv:2408.04594, 2024. 3 + +[14] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, pages 349-364. Springer, 2016. 1 +[15] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, pages 2146-2156, 2021. 3, 6 +[16] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI, 45(11):14020-14037, 2023. 1 +[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3 +[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3 +[19] Ming Li, Xinming Huang, and Ziming Zhang. Self-supervised geometric features discovery via interpretable attention for vehicle re-identification and beyond. In Proceedings of the IEEE/CVF international conference on computer vision, pages 194–204, 2021. 3 +[20] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2021. +[21] Ming Li, Huazhu Fu, Shengfeng He, Hehe Fan, Jun Liu, Jussi Keppo, and Mike Zheng Shou. Dr-fer: Discriminative and robust representation learning for facial expression recognition. IEEE Transactions on Multimedia, 26:6297-6309, 2023. +[22] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. Stprivacy: Spatio-temporal privacy-preserving action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5106-5115, 2023. +[23] Ming Li, Pan Zhou, Jia-Wei Liu, Jussi Keppo, Min Lin, Shuicheng Yan, and Xiangyu Xu. Instant3d: instant text-to-3d generation. International Journal of Computer Vision, 132(10):4456-4472, 2024. 3 +[24] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 3 +[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1 +[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 36, 2024. 1, 3, 6, 7 +[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8 +[28] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, + +Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6, 7 +[29] Yecheng Lyu, Ming Li, Xinming Huang, Ulkuhan Guler, Patrick Schaumont, and Ziming Zhang. Treernn: Topology-preserving deep graph embedding and learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7493-7499. IEEE, 2021. 3 +[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6 +[31] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In CVPR, pages 5642-5651, 2023. 1 +[32] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 3DV, pages 1165-1174. IEEE, 2021. 1 +[33] Rohan Pandey, Rulin Shao, Paul Pu Liang, Ruslan Salakhutdinov, and Louis-Philippe Morency. Cross-modal attention congruence regularization for vision-language relation alignment. arXiv preprint arXiv:2212.10549, 2022. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2 +[35] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. *IJCV*, 126 (12):1394–1414, 2018. 1 +[36] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. CVPR, 2019. 4, 6 +[37] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024.8 +[38] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In CVPR, pages 13193-13203, 2024. 1, 3 +[39] Yichen Shen, Yijin Li, Shuo Chen, Guanglin Li, Zhaoyang Huang, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinktrack: Feature tracking over 100 fps via events and images. arXiv preprint arXiv:2409.17981, 2024. 1 +[40] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, pages 15638-15650, 2022. 3 +[41] Lulu Tian, Hongxun Yao, and Ming Li. Fakepei: A large-scale fake person of interest video detection benchmark and a strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6819-6831, 2023. 3 +[42] Guanqun Wang, Jiaming Liu, Chenxuan Li, Yuan Zhang, Junpeng Ma, Xinyu Wei, Kevin Zhang, Maurice Chong, + +Renrui Zhang, Yijiang Liu, et al. Cloud-device collaborative learning for multimodal large language models. In CVPR, pages 12646-12655, 2024. 3 +[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6, 7 +[44] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 3 +[45] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2, 3 +[46] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. In CVPR, pages 16933-16943, 2024. 1 +[47] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024.1 +[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3 +[49] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In CVPR, pages 9602–9611, 2024. 1 +[50] Yun Yue, Ming Li, Venkatesh Saligrama, and Ziming Zhang. Rnn training along locally optimal trajectories via frankwolfe algorithm. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 10532-10539. IEEE, 2021. 3 +[51] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, et al. Anygpt: Unified multimodal llm with discrete sequence modeling. arXiv preprint arXiv:2402.12226, 2024.3 +[52] Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, and Dong Yu. Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601, 2024. 3 +[53] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, pages 8801-8810, 2022. 1 +[54] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 3 +[55] Yan Zhang, Chun Li, Zhaoxia Liu, and Ming Li. Semi-supervised disease classification based on limited medical image data. IEEE Journal of Biomedical and Health Informatics, 28(3):1575-1586, 2024. 3 +[56] Ce Zheng, Yecheng Lyu, Ming Li, and Ziming Zhang. Lodonet: A deep neural network with 2d keypoint matching + +for 3d lidar odometry estimation. In Proceedings of the 28th ACM international conference on multimedia, pages 2391-2399, 2020. 3 +[57] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arxiv 2023. arXiv preprint arXiv:2302.08890, 2020. 1 +[58] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. E-clip: Towards label-efficient event-based open-world understanding by clip. arXiv e-prints, pages arXiv-2308, 2023. 2, 3 +[59] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In CVPR, pages 18633-18643, 2024. 3 +[60] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3 +[61] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3 \ No newline at end of file diff --git a/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/images.zip b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9600174cb6889351541d372e7f8cf45c348eb126 --- /dev/null +++ b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:687a37cd0ea8466206d56ac00b75fe8474db14d9b2790282f359723abfbe2bf7 +size 549792 diff --git a/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/layout.json b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8aa644de82157a3bbd4c47f6d932450e8eb66d33 --- /dev/null +++ b/2025/EventGPT_ Event Stream Understanding with Multimodal Large Language Models/layout.json @@ -0,0 +1,9021 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 486, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 486, + 140 + ], + "type": "text", + "content": "EventGPT: Event Stream Understanding with Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": "Shaoyu Liu" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Jianing Li" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Guanghui Zhao" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Yunjian Zhang" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Xin Meng" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Fei Richard Yu" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Xiangyang Ji" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Ming Li" + }, + { + "bbox": [ + 134, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "spans": [ + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "text", + "content": "Xidian University " + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "text", + "content": "Peking University " + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 129, + 202, + 495, + 232 + ], + "type": "text", + "content": "Guangdong Laboratory of Artificial Intelligence and Digital Economy(SZ)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 234, + 560, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 234, + 560, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 234, + 560, + 245 + ], + "type": "text", + "content": "23171110721@stu.xidian.edu.cn, lijianing@pku.edu.cn, ming.li@u.nus.edu, ghzhao@xidian.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 299, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 297, + 622 + ], + "type": "text", + "content": "Event cameras capture visual information as asynchronous pixel change streams, excelling in challenging lighting and high-dynamic scenarios. Existing multimodal large language models (MLLMs) concentrate on natural RGB images, failing in scenarios where event data fits better. In this paper, we introduce EventGPT, the first MLLM for event stream understanding, pioneering the integration of large language models (LLMs) with event-based vision. To bridge the huge domain gap, we propose a three-stage optimization paradigm to progressively equip a pre-trained LLM with event understanding. Our EventGPT consists of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Firstly, GPT-generated RGB image-text pairs warm up the linear projector, following LLaVA, as the gap between natural images and language is smaller. Secondly, we construct N-ImageNet-Chat, a large synthetic dataset of event data and corresponding texts to enable the use of the spatio-temporal aggregator and to train the event-language adapter, thereby aligning event features more closely with the language space. Finally, we gather an instruction dataset, EventChat, which contains extensive real-world data to fine-tune the entire model, further enhancing its generalization ability. We construct a comprehensive benchmark, and experiments show that EventGPT surpasses previous state-of-the-art MLLMs in generation quality, descriptive accuracy, and reasoning capability. Code: EventGPT" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "type": "text", + "content": "Event cameras, with predominant advantages in capturing scene changes and temporal resolution, have great ap" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 274, + 553, + 401 + ], + "blocks": [ + { + "bbox": [ + 319, + 274, + 553, + 401 + ], + "lines": [ + { + "bbox": [ + 319, + 274, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 319, + 274, + 553, + 401 + ], + "type": "image", + "image_path": "6719442a966cd1edfd8fd39e5a4de7f04a29dba2b600e35d10985b6641e0fe6e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 455 + ], + "type": "text", + "content": "Figure 1. Our EventGPT is the first multimodal large language model tailored for event stream understanding, including scene summarization, reasoning, and question answering, with a great potential for downstream tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 543 + ], + "type": "text", + "content": "plication potential in scenarios under imperfect lighting conditions or including high dynamic objects [7, 12, 57]. The relevant tasks on event data, e.g. object detection [8, 10, 16, 46], tracking [9, 31, 39, 53], and 3D reconstruction [14, 32, 35, 49], have been well-investigated." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Recent multimodal large language models (MLLMs) have shown remarkable breakthroughs in conventional vision and language problems, demonstrating extraordinary capabilities of scene understanding and question answering (QA) [25, 26, 38, 47, 60]. Typically, a pioneering research Large Language and Vision Assistant (LLaVA) combines a vision encoder with an LLM, trained on GPT-4-generated multimodal data, to create a robust system that excels in visual understanding, instruction-following, and complex question answering [26]. These works mainly pay attention to natural RGB scenes and the resulting models degrade significantly when encountering the aforementioned scenarios. However, the relevant research on event data streams remains underexplored." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29139" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 230, + 120, + 380, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 120, + 380, + 132 + ], + "spans": [ + { + "bbox": [ + 230, + 120, + 380, + 132 + ], + "type": "text", + "content": "Large Language Model (Vicuna v1.5)" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 137, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 61, + 137, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 61, + 137, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 545, + 334 + ], + "type": "image", + "image_path": "43512ca007ee5fe49f1b1bb13e80f993a2aca04578c53ecaedefcbfc99b57551.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 342, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 54, + 342, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 555, + 376 + ], + "type": "text", + "content": "Figure 2. Overview of our framework. The event encoder transforms raw event tensors into high-dimensional features, which are then aggregated by a spatio-temporal module. These representations are projected and aligned with the large language model, enabling nuanced understanding of event streams and supporting various downstream applications." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 396, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 296, + 576 + ], + "type": "text", + "content": "Some works have initialized the alignment between event data and textual information. For example, EventCLIP [45] adapts the pre-trained CLIP model [6, 34] for event-based object recognition, enabling zero-shot and few-shot classification. Similarly, EventBind [58] extends CLIP with an event encoder to align images, texts, and events in a unified representation space, realizing few-shot recognition and event retrieval. Nevertheless, both EventCLIP and EventBind lack the ability to leverage the embedded world knowledge in LLMs, limiting their capacity for nuanced scene understanding and dialog generation in complex scenarios. This gap highlights the necessity of a model, akin to LLaVA for RGB images, specifically designed to overcome these limitations by harnessing the power of LLMs for comprehensive event stream understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 296, + 713 + ], + "type": "text", + "content": "In this work, we pioneer the development of an MLLM for event stream understanding, termed EventGPT (see Fig 1), designed for the community to support scene summarization, reasoning, and question answering. To this end, several key challenges must be addressed. First of all, aligning event features with the high-level semantics of natural language is significantly challenging, as event streams are designed to capture asynchronous and sparse brightness changes within the field of view, inherently lacking the global scene context necessary for language-based understanding. Secondly, the fundamental difference in how" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 396, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 468 + ], + "type": "text", + "content": "event and RGB cameras capture scene information complicates the straightforward adaptation of existing multimodal techniques, which are typically designed for natural images. Finally, the scarcity of large-scale, high-quality event-text datasets hinders the training of a scalable MLLM for comprehensive event scene understanding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "content": "To address these challenges, we propose a three-stage training paradigm that leverages an LLM as an anchor, given its rich world knowledge and a strong foundation for scene understanding, to progressively align the event modality with language. In general, our EventGPT is mainly composed of an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. Compared to event data, natural images are much closer to language in the embedding space. Therefore, in the first stage, we use the image-text dataset from LLaVA to warm up the linear projector, establishing preliminary scene-language associations. In the second stage, we focus on training the spatio-temporal aggregator and event-language adapter to map spatio-temporal information from event frames into language space. For this purpose, we construct a large-scale synthetic event-text dataset, NImageNet-Chat, to facilitate optimization. Finally, we conduct instructional fine-tuning of all model parameters on a self-collected, high-quality dataset, Event-Chat, which includes diverse scenes with unsatisfactory lighting condi" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 75, + 91, + 106 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 75, + 91, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 75, + 91, + 106 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 80, + 538, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 80, + 538, + 101 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 538, + 101 + ], + "type": "text", + "content": "EventGPT: This image depicts a person puncturing a balloon with a needle, causing a sudden, dramatic eruption of water. The water splashes out in a forceful burst, creating intricate trails and droplets that scatter through the air in all directions ..." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29140" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "tions or high-speed object motions, enhancing the capability of our EventGPT to interpret complex real-world scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 295, + 262 + ], + "type": "text", + "content": "We introduce a comprehensive evaluation benchmark inspired by Qwen2-72B-Instruct to assess the generation quality, descriptive accuracy, and reasoning capability of EventGPT. To facilitate a rigorous comparison, we select several state-of-the-art MLLMs excelling in natural scene understanding, including LLaVA [26], Qwen2-VL [43], and Deepseek-VL [28], as baseline models. Evaluation is conducted across three metrics, i.e. Detail Captioning (DC), Complex Reasoning (CR), and Visual Question Answering (VQA), in line with the existing literature. Experimental results highlight significant improvements achieved by EventGPT. Furthermore, our model demonstrates strong generalization ability on open-set downstream tasks, such as object detection and semantic segmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 287 + ], + "type": "text", + "content": "Our main contributions in this work can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 288, + 296, + 479 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 61, + 288, + 295, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 295, + 347 + ], + "type": "text", + "content": "- We present EventGPT, the first multimodal large language model tailored for event stream understanding, demonstrating capabilities in scene summarization, reasoning, and question answering. EventGPT establishes a solid baseline, paving the way for future research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 348, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 348, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 348, + 295, + 407 + ], + "type": "text", + "content": "- We propose a systematic learning paradigm incorporating specialized training strategies and network modules to progressively bridge the substantial domain gap between event data and language, providing a valuable reference for future cross-modal learning efforts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 407, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 296, + 479 + ], + "type": "text", + "content": "- We construct two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, containing 1 million synthetic samples for pre-training and 120,000 instruction samples for fine-tuning. These datasets will be publicly released to drive progress in multimodal learning within the event domain." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 490, + 147, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 147, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 147, + 502 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Multimodal Large Language Models. In deep learning era[19-23, 29, 41, 50, 55, 56], significant advancements have been made in various domains. The significant advancements in LLMs for natural language understanding have driven their integration into multimodal research domains, with a focus on enhancing vision-language alignment to expand their applicability [2-4, 24, 33, 38, 40, 42, 48, 52, 54]. Early contributions, such as Flamingo [1] and BLIP [17, 18], integrated visual and language modalities to enable zero-shot generalization across a broad range of visual-language tasks. Further advancements emerged with chat-oriented MLLMs like MiniGPT-4 [60] and LLaVA [26], which improved the interactive capabilities and adaptability of LLMs in multimodal applications, setting new benchmarks in performance. As MLLMs continue to evolve, the exploration of novel sensor modalities holds potential for expanding their functional scope. De" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 76, + 409, + 185 + ], + "blocks": [ + { + "bbox": [ + 318, + 76, + 409, + 185 + ], + "lines": [ + { + "bbox": [ + 318, + 76, + 409, + 185 + ], + "spans": [ + { + "bbox": [ + 318, + 76, + 409, + 185 + ], + "type": "image", + "image_path": "0cae54f9b3797762b8bad9629f3d6081e584de57b5a474c711ba49b8a9a02cf0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 199, + 553, + 233 + ], + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 233 + ], + "type": "text", + "content": "Figure 3. Data distribution across three progressive training stages in our EventGPT containing image-language alignment, event-language alignment, and instruction tuning." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 551, + 188 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 551, + 188 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 551, + 188 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 551, + 188 + ], + "type": "image", + "image_path": "771db84bbc6a2744a1c91e6738f97fe73dceb996cd4d3d6cababc94f153fc9ce.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 312 + ], + "type": "text", + "content": "spite the distinct advantages of event cameras, such as high temporal resolution and robustness in challenging lighting conditions, the integration of event-based data with LLMs remains underexplored, particularly for applications in high-dynamic and low-light environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 515 + ], + "type": "text", + "content": "Event-Driven Multimodal Models. The integration of event-based data into VLMs has attracted significant interest, demonstrating potential across various downstream tasks and highlighting the advantages of event-language synergy. [45, 58, 59]. Notably, frameworks such as ExACT have leveraged language-guided frameworks for event-based action recognition, effectively aligning action sequences with contextual textual cues to achieve strong performance [59]. Other approaches, such as EventCLIP [45] and EventBind [58], have also demonstrated strong zero-shot performance, underscoring the effectiveness of event-based embeddings when aligned with textual information. Despite these advances, developing chat-centric LLMs capable of incorporating event-based modalities presents significant research challenges. Addressing this gap could enhance models' contextual comprehension, enabling more effective processing of complex interactive tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 474, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 474, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 474, + 538 + ], + "type": "text", + "content": "3. Event-Text Data Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 665 + ], + "type": "text", + "content": "Event-based datasets have been extensively studied in computer vision tasks. However, large-scale Event-Text datasets suitable for training MLLMs) are not yet widely available. Previous studies have successfully employed large models, such as GPT, to synthesize data for multimodal applications, yielding significant advancements in model capabilities [13, 44, 51, 61]. To address this gap, we construct an Event-Text dataset using a human-assisted, semi-automated pipeline that incorporates both synthetic and real-world event data. The distribution of the dataset is shown in Fig. 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 671, + 419, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 419, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 419, + 684 + ], + "type": "text", + "content": "3.1. Synthetic Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "ImageNet serves as a foundational benchmark for visual recognition tasks. N-ImageNet [15] extends this founda" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29141" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 140, + 76, + 221, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 76, + 221, + 87 + ], + "spans": [ + { + "bbox": [ + 140, + 76, + 221, + 87 + ], + "type": "text", + "content": "Instruction Examples" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 93, + 96, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 93, + 96, + 102 + ], + "spans": [ + { + "bbox": [ + 67, + 93, + 96, + 102 + ], + "type": "text", + "content": "Prompt:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 104, + 123, + 112 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 104, + 123, + 112 + ], + "spans": [ + { + "bbox": [ + 66, + 104, + 123, + 112 + ], + "type": "text", + "content": "Caption Prompt:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 113, + 230, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 113, + 230, + 130 + ], + "spans": [ + { + "bbox": [ + 66, + 113, + 230, + 130 + ], + "type": "text", + "content": "You are an expert in image description and understanding. You are requested ..." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 131, + 133, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 131, + 133, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 131, + 133, + 140 + ], + "type": "text", + "content": "Reasoning Prompt:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 141, + 231, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 141, + 231, + 159 + ], + "spans": [ + { + "bbox": [ + 66, + 141, + 231, + 159 + ], + "type": "text", + "content": "You are an advanced multimodal model with the ability to perform complex reasoning ..." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 160, + 131, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 160, + 131, + 170 + ], + "spans": [ + { + "bbox": [ + 66, + 160, + 131, + 170 + ], + "type": "text", + "content": "Visual QA Prompt:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 170, + 230, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 170, + 230, + 187 + ], + "spans": [ + { + "bbox": [ + 66, + 170, + 230, + 187 + ], + "type": "text", + "content": "You are an AI language model designed to generate detailed and insightful Visual Question Answering...." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 235, + 106, + 288, + 146 + ], + "blocks": [ + { + "bbox": [ + 235, + 106, + 288, + 146 + ], + "lines": [ + { + "bbox": [ + 235, + 106, + 288, + 146 + ], + "spans": [ + { + "bbox": [ + 235, + 106, + 288, + 146 + ], + "type": "image", + "image_path": "23a7c812e4c04efe921740bb56c657da6dac85aa1fce715ead44fcdd08ca80d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 317, + 296, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 350 + ], + "type": "text", + "content": "Figure 4. Instruction examples from the Event-Chat dataset for multimodal task prompts and responses in image captioning, visual reasoning, and visual question answering." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 235, + 148, + 288, + 190 + ], + "blocks": [ + { + "bbox": [ + 235, + 148, + 288, + 190 + ], + "lines": [ + { + "bbox": [ + 235, + 148, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 235, + 148, + 288, + 190 + ], + "type": "image", + "image_path": "e4a8e328bf52da19c140c8cc9b70dcc7cdaa140a6d1679f7e6bfd77628d4f13e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 194, + 94, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 194, + 94, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 194, + 94, + 201 + ], + "type": "text", + "content": "Results:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 205, + 95, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 205, + 95, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 205, + 95, + 213 + ], + "type": "text", + "content": "Caption:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 213, + 200, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 213, + 200, + 221 + ], + "spans": [ + { + "bbox": [ + 66, + 213, + 200, + 221 + ], + "type": "text", + "content": "Query: Describe the buildings in the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 222, + 255, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 222, + 255, + 231 + ], + "spans": [ + { + "bbox": [ + 66, + 222, + 255, + 231 + ], + "type": "text", + "content": "Answer: There are several multi-story buildings in the image..." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 66, + 236, + 99, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 236, + 99, + 245 + ], + "spans": [ + { + "bbox": [ + 66, + 236, + 99, + 245 + ], + "type": "text", + "content": "Reasoning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 66, + 245, + 283, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 245, + 283, + 262 + ], + "spans": [ + { + "bbox": [ + 66, + 245, + 283, + 262 + ], + "type": "text", + "content": "Query: If the construction barriers were removed, how would it impact the traffic flow at the intersection?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 66, + 262, + 263, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 262, + 263, + 271 + ], + "spans": [ + { + "bbox": [ + 66, + 262, + 263, + 271 + ], + "type": "text", + "content": "Answer: Without the barriers, vehicles would have a direct path.." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 66, + 277, + 98, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 277, + 98, + 285 + ], + "spans": [ + { + "bbox": [ + 66, + 277, + 98, + 285 + ], + "type": "text", + "content": "Visual QA" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 66, + 285, + 286, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 285, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 66, + 285, + 286, + 294 + ], + "type": "text", + "content": "Query: How many cars are visible, and what directions are they moving?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 66, + 294, + 272, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 294, + 272, + 304 + ], + "spans": [ + { + "bbox": [ + 66, + 294, + 272, + 304 + ], + "type": "text", + "content": "Answer: Three cars are visible. The one on the left is turning away.." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 373, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 432 + ], + "type": "text", + "content": "tion by providing a synthetic event-based dataset generated through event simulators, aligned with ImageNet images. Using this alignment, we generate textual instructions for the images to enable cross-domain learning between visual and event-based data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 434, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 493 + ], + "type": "text", + "content": "N-ImageNet-Chat. We present a large-scale dataset comprising over 1,000,000 instances with concise instructions, primarily designed to align event data with natural language. This dataset provides a foundation for the model's understanding and supports robust representation learning." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "type": "text", + "content": "N-ImageNet-Instruction. We establish an event-based dataset of over 69,000 entries, designed to improve the model's responsiveness to complex, task-specific instructions. The dataset includes captions, visual question answering (VQA), and complex reasoning tasks aimed at finetuning the model's multimodal capabilities." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 574, + 171, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 171, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 171, + 586 + ], + "type": "text", + "content": "3.2. Real-World Dataset" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "In addition to synthetic datasets, we establish a real-world dataset to improve model generalization for practical applications. A human-assisted and semi-automated annotation process ensures high-quality labels for complex scenes." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Event Data for Event-Chat. We leverage real-world event datasets, such as DSEC [11] and E2VID [36], which are primarily designed for autonomous driving and provide synchronized image-event data under challenging conditions, including low-light environments and high-speed motion (e.g., entering/exitting tunnels, night scenes)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "Instructions for Event-Chat. This dataset is designed to enhance the model's generalization capabilities, consisting of 59,000 high-quality annotated entries. Three distinct prompting strategies (i.e., captioning, VQA, and complex reasoning) form the basis of the Event-Chat dataset, providing diverse and comprehensive instructions. As shown in Fig. 4, examples from each prompting strategy illustrate the dataset's richness in contextual understanding. Combined with N-ImageNet-instruction, these annotated entries form the final Event-Chat dataset." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 192, + 554, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 554, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 554, + 276 + ], + "type": "text", + "content": "In summary, our newly built event-text dataset contains over 1,000,000 entries from both synthetic and real-world data, offering a robust foundation for advancing multimodal understanding in both event-based and textual domains. These fundamental and challenging datasets are intended to foster progress in the development of MLLMs within the neuromorphic community." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 286, + 383, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 383, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 383, + 298 + ], + "type": "text", + "content": "4. EventGPT" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 306, + 381, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 381, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 381, + 317 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 324, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 479 + ], + "type": "text", + "content": "EventGPT aims to serve as an event-based MLLM capable of understanding and generating responses based on spatio-temporal event data. As illustrated in Fig. 2, EventGPT formulates event data processing as an event-driven approach to question answering and descriptive generation. By leveraging the high temporal resolution and extended dynamic range inherent in event stream data, EventGPT enhances understanding in scenarios traditionally challenging for standard vision models, such as low-light or high-speed motion conditions. At the core of EventGPT lies an LLM that synthesizes complex spatio-temporal information from event streams, utilizing world knowledge to generate coherent, human-readable outputs." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": "In general, an event bin can be mathematically described within a time window " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "S = \\{(x_{j},y_{j},t_{j},p_{j})\\}_{j = 1}^{N}" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": ", where each event consists of spatial coordinates " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "(x_{j},y_{j})" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": ", timestamp " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": ", and polarity " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": ". Initially, high-dimensional event features " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": " are extracted from " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": " using an event encoder " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": ", initialized with OpenCLIP ViT-L/14-336px to leverage its robust feature extraction capabilities. These features " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": " are then processed through a spatio-temporal encoder " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ST}}" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": " to capture dynamic interactions, resulting in a comprehensive representation " + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 480, + 554, + 600 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 375, + 611, + 553, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 611, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 375, + 611, + 553, + 624 + ], + "type": "interline_equation", + "content": "V = f _ {\\text {e n c}} (S), \\quad E = f _ {\\mathrm {S T}} (V) \\tag {1}", + "image_path": "8f60e3c2e82bd6acb89999f806c61d70b702f3dddb26363540ef03a10c38db18.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "Subsequently, a two-stage projector is employed to achieve cross-modal alignment. The first stage aligns the spatio-temporally encoded features " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " with a textual representation, and the second stage adapts this representation specifically to the event modality, resulting in an event-aligned feature " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ". Given a user query " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ", the large language model " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{llm}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " performs reasoning over " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " to generate a" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29142" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 107, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 107, + 84 + ], + "type": "text", + "content": "response " + }, + { + "bbox": [ + 55, + 72, + 107, + 84 + ], + "type": "inline_equation", + "content": "A" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 97, + 295, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 97, + 295, + 111 + ], + "spans": [ + { + "bbox": [ + 93, + 97, + 295, + 111 + ], + "type": "interline_equation", + "content": "F = f _ {\\text {p r o j}} \\left(f _ {\\text {a l i g n}} (E)\\right), \\quad A = f _ {\\mathrm {l l m}} (F \\mid Q) \\tag {2}", + "image_path": "608e32df0fda6829b588a58ea796de74a5f41a17b2c3ee9f87bf6c32100aa256.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 116, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 296, + 175 + ], + "type": "text", + "content": "EventGPT creates a strong representation of event streams and incorporates task-specific heads to support downstream tasks like object detection and instance segmentation. This flexibility enhances its applicability to a wide range of complex vision tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 184, + 291, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 291, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 291, + 198 + ], + "type": "text", + "content": "4.2. Spatio-Temporal Information Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 202, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 287 + ], + "type": "text", + "content": "The event stream provides rich spatiotemporal cues crucial for precise temporal and spatial modeling, especially in high-speed motion scenarios, to enhance a model's contextual understanding. The continuous event stream is typically segmented into discrete temporal bins for efficient processing. Each temporal bin " + }, + { + "bbox": [ + 55, + 202, + 296, + 287 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 55, + 202, + 296, + 287 + ], + "type": "text", + "content": " is independently encoded, forming discrete spatiotemporal feature representations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 297, + 295, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 297, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 102, + 297, + 295, + 314 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} = \\left\\{f _ {t} ^ {e} \\right\\} _ {t = 1, 2, \\dots} ^ {T}, \\quad \\mathcal {Z} \\in \\mathbb {R} ^ {T \\times S \\times D} \\tag {3}", + "image_path": "631ec50c6d5aaa4a79354058998cb317fe95d7c52378c543e3407a6a6612e9bb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "text", + "content": " denotes the spatial resolution of the encoded features, and " + }, + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "text", + "content": " represents the feature dimensionality after encoding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 355, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 403 + ], + "type": "text", + "content": "To aggregate salient temporal and spatial information, we perform separate temporal and spatial encoding operations. Specifically, Temporal Encoding (TE) aggregates features along the temporal dimension via average pooling:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 412, + 295, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 412, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 100, + 412, + 295, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} _ {T} ^ {a v g} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} f _ {t} ^ {e}, \\quad \\mathcal {Z} _ {T} ^ {a v g} \\in \\mathbb {R} ^ {S \\times D} \\tag {4}", + "image_path": "48e763c447ab1b55267cf6213492252f0da7a1f737df8fe3113d506a79eb56e9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 450, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 475 + ], + "type": "text", + "content": "Similarly, Spatial Encoding (SE) aggregates features along the spatial dimension via average pooling:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 484, + 295, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 484, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 100, + 484, + 295, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} _ {S} ^ {a v g} = \\frac {1}{S} \\sum_ {t = 1} ^ {S} f _ {t} ^ {e}, \\quad \\mathcal {Z} _ {S} ^ {a v g} \\in \\mathbb {R} ^ {T \\times D} \\tag {5}", + "image_path": "d0ca8e2f743a4be67c0e478d971e1ef7f206bd9c1566998b3525934524c9363c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 558 + ], + "type": "text", + "content": "Finally, the temporally and spatially aggregated representations are concatenated along the feature dimension, forming a robust spatiotemporal embedding:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 82, + 578, + 295, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 578, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 82, + 578, + 295, + 594 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathcal {Z}}} = \\operatorname {C o n c a t} \\left(\\mathcal {Z} _ {T} ^ {a v g}, \\mathcal {Z} _ {S} ^ {a v g}\\right), \\quad \\overline {{\\mathcal {Z}}} \\in \\mathbb {R} ^ {(T + S) \\times D} \\tag {6}", + "image_path": "92ff5f4a61bd1d98aec638afea9d3c2e4c8475ed578558f5a6a983e34c6c2dd5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "type": "text", + "content": "The fused representation " + }, + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\overline{Z}" + }, + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "type": "text", + "content": " is an end-to-end learned spatiotemporal feature from asynchronous events, forming a robust event embedding from various complex scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 647, + 164, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 164, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 164, + 659 + ], + "type": "text", + "content": "4.3. Training Pipelines" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Given the huge domain gap between event and language, we include a visual-language pre-training to initialize our framework. As shown in Fig. 5, EventGPT adopts a three-phase training framework encompassing visual-language," + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 325, + 82, + 553, + 206 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 81 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 81 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 81 + ], + "type": "text", + "content": "Stage 1: Visual-Language Alignment Stage 2: Event-Language Alignment Stage 3: Instruction Finetuning" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 82, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 325, + 82, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 325, + 82, + 553, + 206 + ], + "type": "image", + "image_path": "6bba3818f32e079dce0ec1810d1ec93247808ffa3f61ba5eddb605f9773cf437.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 215, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 270 + ], + "type": "text", + "content": "Figure 5. Our three-stage pipeline consists of (1) training a linear projector to achieve visual-language alignment, (2) incorporating a spatio-temporal aggregator and an event-language adapter to enable event-language alignment, and (3) performing full fine-tuning of all modules to facilitate cohesive multimodal learning." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "content": "event-language, and instruction tuning. This structured progression facilitates effective cross-domain alignment, enhancing EventGPT's capacity for multimodal reasoning and understanding." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 359, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 479 + ], + "type": "text", + "content": "Visual-Language Adapter. In the visual-language alignment stage, we leverage the LLaVA-Pre-train dataset while keeping the event encoder and LLM modules frozen, focusing exclusively on training the linear projector. Visual-language alignment is prioritized due to the relatively small domain differences between image and text, compared to the distinct nature of event and textual data. This setup establishes initial cross-modal alignment between vision and language, enabling the LLM to interpret scene information and providing a foundational multimodal understanding." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "text", + "content": "Event-Language Adapter. In the event-language alignment stage, we leverage our constructed N-ImageNet-Chat dataset, with the event encoder, linear projector, and LLM modules kept frozen, focusing exclusively on training the event-language adapter module. This stage aims to bridge the domain gap between event and language modalities. By aligning event embeddings with the LLM's representational space, the model's ability to interpret event streams and generate descriptive narratives is significantly enhanced." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Instruction Tuning. In the instruction tuning stage, all trainable parameters, including the event encoder, linear projector, event-language adapter, and LLM modules, are unfrozen to enable supervised fine-tuning. Trained on our high-quality Event-Chat question-answering dataset, this stage further enhances the model's understanding and generation capabilities for event stream data. This structured, three-phase training framework progressively narrows the domain gap between event and language modalities." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29143" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 102, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Performance comparison between EventGPT and five state-of-the-art methods on the N-ImageNet-Chat and Event-Chat datasets across three evaluation metrics. EventGPT consistently demonstrates competent performance on all datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 102, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 51, + 102, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 51, + 102, + 553, + 216 + ], + "type": "table", + "html": "
ModelsLLM BackboneParamsEncoderN-ImageNet-ChatEvent-Chat
DCCRVQADCCRVQA
LLaVA-7B-v1.5 [26]Vicuna-v1.57BCLIP ViT-L/336PX1.541.071.882.204.043.26
Qwen2-VL-7B [43]Qwen27BOpenCLIP ViT-bigG1.741.461.912.384.022.91
Intern2VL-8B [5]InternLM2.58BInternViT-300M1.511.872.082.374.003.71
Deepseek-vl-7b [28]Deepseek-LLM7BSAM-B + SigLIP-L1.521.451.842.414.103.37
EventGPT-7BVicuna-v1.57BCLIP ViT-L/336PX2.392.572.233.524.094.29
LLaVA-13B-v1.5 [26]Vicuna-v1.513BCLIP ViT-L/336PX1.751.211.842.344.083.37
EventGPT-13BVicuna-v1.513BCLIP ViT-L/336PX2.412.812.403.404.134.26
", + "image_path": "8f2333f3f84e76ad98d5bb8afac4445083f5382e8a1763c8f80f90022ede1038.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 233, + 147, + 293 + ], + "blocks": [ + { + "bbox": [ + 70, + 233, + 147, + 293 + ], + "lines": [ + { + "bbox": [ + 70, + 233, + 147, + 293 + ], + "spans": [ + { + "bbox": [ + 70, + 233, + 147, + 293 + ], + "type": "image", + "image_path": "733cd18ab7858fb825dede73b4c1b10ae8c5e076fcff944aa35558875f528fb4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 437, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 54, + 437, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 555, + 472 + ], + "type": "text", + "content": "Figure 6. Representative visualization comparison between EventGPT and other open-source MLLMs models under extreme lighting variations. EventGPT demonstrates superior accuracy in recognizing critical scene details, such as vehicle presence at tunnel exits, showcasing its robustness in challenging lighting conditions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 234, + 223, + 292 + ], + "blocks": [ + { + "bbox": [ + 149, + 234, + 223, + 292 + ], + "lines": [ + { + "bbox": [ + 149, + 234, + 223, + 292 + ], + "spans": [ + { + "bbox": [ + 149, + 234, + 223, + 292 + ], + "type": "image", + "image_path": "4838196664db2922b319b7e8d5cf5650d4bf3eb4202ea15059d7af2d6661a6ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 226, + 234, + 302, + 293 + ], + "blocks": [ + { + "bbox": [ + 226, + 234, + 302, + 293 + ], + "lines": [ + { + "bbox": [ + 226, + 234, + 302, + 293 + ], + "spans": [ + { + "bbox": [ + 226, + 234, + 302, + 293 + ], + "type": "image", + "image_path": "6adcdc8d7d4488bf49e7d9f70604a9cb67298a1344250fe8cd8c8a1a5cd8fed2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 234, + 383, + 292 + ], + "blocks": [ + { + "bbox": [ + 306, + 234, + 383, + 292 + ], + "lines": [ + { + "bbox": [ + 306, + 234, + 383, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 234, + 383, + 292 + ], + "type": "image", + "image_path": "e05a64d6f27d68849795d41583c620a046097cb4e62ba2117c04c649e37283b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 385, + 234, + 462, + 293 + ], + "blocks": [ + { + "bbox": [ + 385, + 234, + 462, + 293 + ], + "lines": [ + { + "bbox": [ + 385, + 234, + 462, + 293 + ], + "spans": [ + { + "bbox": [ + 385, + 234, + 462, + 293 + ], + "type": "image", + "image_path": "2fa3ec0385e90f9974451546e8e224da59df1e4542302d2a0d315beefae0af59.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 465, + 234, + 542, + 293 + ], + "blocks": [ + { + "bbox": [ + 465, + 234, + 542, + 293 + ], + "lines": [ + { + "bbox": [ + 465, + 234, + 542, + 293 + ], + "spans": [ + { + "bbox": [ + 465, + 234, + 542, + 293 + ], + "type": "image", + "image_path": "8913e1d5c018c8b677747a514934989b6d59403d3c76667d50bc199657ea7193.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 296, + 357, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 296, + 357, + 307 + ], + "spans": [ + { + "bbox": [ + 67, + 296, + 357, + 307 + ], + "type": "text", + "content": "Prompt: Describe the scene depicted in the picture and identify the dangers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 309, + 544, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 309, + 544, + 342 + ], + "spans": [ + { + "bbox": [ + 66, + 309, + 544, + 342 + ], + "type": "text", + "content": "##LLaVA-7B: The scene depicted is a car driving on a highway, with a tunnel entrance ahead. The car is approaching the large tunnel opening which is currently blocked by a large boulder. The boulder is preventing the car from entering the tunnel and may potentially cause a dangerous situation, as the driver might not be aware of the situation or stop in time to avoid it..." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 344, + 544, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 344, + 544, + 376 + ], + "spans": [ + { + "bbox": [ + 66, + 344, + 544, + 376 + ], + "type": "text", + "content": "EventGPT -7B: This scene shows a vehicle traveling through a tunnel toward the open road outside. The dangers in this scene include the potential for oncoming traffic from the opposite lane, as well as the reduced visibility due to the tunnel's enclosed space. Drivers are advised to stay alert and prepared for any sudden appearances of vehicles or obstacles." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 383, + 291, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 383, + 291, + 393 + ], + "spans": [ + { + "bbox": [ + 66, + 383, + 291, + 393 + ], + "type": "text", + "content": "Prrs: Are there any vehicles driving at the tunnel exit?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 398, + 347, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 398, + 347, + 408 + ], + "spans": [ + { + "bbox": [ + 66, + 398, + 347, + 408 + ], + "type": "text", + "content": "```python\n>>> LLaVA-7B : No, there are no vehicles driving at the tunnel exit in the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 408, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 408, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 67, + 408, + 294, + 418 + ], + "type": "text", + "content": "EventGPT: Yes, there is a vehicle driving at the tunnel exit." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 483, + 132, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 132, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 132, + 497 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 502, + 173, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 173, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 173, + 516 + ], + "type": "text", + "content": "5.1. Experiment Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "content": "Implementations. We integrate the N-ImageNet[15], Dsec[11], and E2VID[36] datasets with synthesized multimodal conversational datasets, N-ImageNet-Chat and Event-Chat, using a GPT-assistant approach for three-stage model training. Our LLM backbone is Vicuna-v1.5, with a two-layer MLP as the linear projector and a linear layer as the event-language adapter. All models, including 7B and 13B versions, are trained on 8 NVIDIA RTX 6000 Ada GPUs. In the visual-language alignment stage, we use 558k data with a learning rate of " + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "content": " and a batch size of 64. For event-language alignment, we use 1000k data with the same learning rate and batch size. In instruction-supervised fine-tuning, we use 120k data, setting the learning rate to " + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "content": " and batch size to 32." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. Following the evaluation strategies of LLaVA [26] and VideoGPT [30], we rigorously assess" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 484, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 665 + ], + "type": "text", + "content": "EventGPT's performance against state-of-the-art MLLMs using three core metrics within a zero-shot question-answering setting. Our benchmark, specifically designed for text generation from event streams, is based on the test sets of N-ImageNet-Chat and Event-Chat datasets, which contain question-answer pairs to evaluate the model's abilities in detailed description, complex reasoning, and visual question answering. Generated responses are quantitatively scored on a 1-5 scale across three dimensions: (i) Detailed Captioning (DC), assessing the accuracy and completeness of descriptions to capture key event details; (ii) Complex Reasoning (CR), evaluating the model's capacity for multi-step reasoning and logical inference; and (iii) Visual Question Answering (VQA), measuring precision and contextual understanding in responding to specific queries." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 671, + 533, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 533, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 533, + 684 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-Art Models" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Quantitative Results. To assess the effectiveness of EventGPT, we report the quantitative evaluation results of its" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29144" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 77, + 142, + 137 + ], + "blocks": [ + { + "bbox": [ + 65, + 77, + 142, + 137 + ], + "lines": [ + { + "bbox": [ + 65, + 77, + 142, + 137 + ], + "spans": [ + { + "bbox": [ + 65, + 77, + 142, + 137 + ], + "type": "image", + "image_path": "4ee100de6a3848a5634026a637a4c9d4cde65b9c2496e34e01d525256a011fa4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "type": "text", + "content": "Figure 7. Representative comparison of EventGPT and other open-source multimodal language models (MLLMs) in high-speed motion scenarios, demonstrating EventGPT's enhanced capability to capture spatiotemporal details and interpret motion characteristics, including fragment trajectories and impact effects." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 78, + 223, + 137 + ], + "blocks": [ + { + "bbox": [ + 145, + 78, + 223, + 137 + ], + "lines": [ + { + "bbox": [ + 145, + 78, + 223, + 137 + ], + "spans": [ + { + "bbox": [ + 145, + 78, + 223, + 137 + ], + "type": "image", + "image_path": "b6f04d867be761646038b6fe13b564360e87764d6332c34a056303e4e993c4bb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 78, + 304, + 137 + ], + "blocks": [ + { + "bbox": [ + 227, + 78, + 304, + 137 + ], + "lines": [ + { + "bbox": [ + 227, + 78, + 304, + 137 + ], + "spans": [ + { + "bbox": [ + 227, + 78, + 304, + 137 + ], + "type": "image", + "image_path": "764473f41db5737dd2bd2e90d0a5d2c43578ad8b6992e62573272cbe208a36de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 78, + 384, + 137 + ], + "blocks": [ + { + "bbox": [ + 309, + 78, + 384, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 78, + 384, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 78, + 384, + 137 + ], + "type": "image", + "image_path": "2438e4da493a505bc4ecbe0663adb1c98a260c9314bc5612d09e4d06e413b48d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 78, + 465, + 137 + ], + "blocks": [ + { + "bbox": [ + 389, + 78, + 465, + 137 + ], + "lines": [ + { + "bbox": [ + 389, + 78, + 465, + 137 + ], + "spans": [ + { + "bbox": [ + 389, + 78, + 465, + 137 + ], + "type": "image", + "image_path": "1f15b8c98d9db183636751f57cac1c5a388a3e47ff6c124a5b5ea406aaf4ef53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 470, + 78, + 547, + 137 + ], + "blocks": [ + { + "bbox": [ + 470, + 78, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 470, + 78, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 470, + 78, + 547, + 137 + ], + "type": "image", + "image_path": "5c119ba8c5fbd2b2b7a1cdeec7825d6c026ae03049f388e8e1c76d988db10965.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 142, + 357, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 142, + 357, + 153 + ], + "spans": [ + { + "bbox": [ + 63, + 142, + 357, + 153 + ], + "type": "text", + "content": "Prrs: Plesse describe this scene and explain what scene or action it depicts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 156, + 547, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 156, + 547, + 197 + ], + "spans": [ + { + "bbox": [ + 62, + 156, + 547, + 197 + ], + "type": "text", + "content": "##LLaVA-7B: The scene shows a broken piece of glass flying through the air in front of a white sheet. This glass could be from a broken window, creating an unexpected scene. The sheet of white fabric is likely designed to protect the area from the flying glass shards or to catch them if they land. The shards of glass are scattered around the scene, with some closer to the sheet and others further away, creating a sense of danger and urgency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 201, + 547, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 201, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 201, + 547, + 232 + ], + "type": "text", + "content": "Qwen2-VL-7B: The scene depicts a scene where a ceramic object, possibly a bowl or a plate, is being shattered. The object is placed on a white surface, which appears to be a piece of paper or a cloth. The shattered pieces of the ceramic object are scattered around the surface, with some pieces still attached to the main body of the object..." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 236, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 547, + 266 + ], + "type": "text", + "content": "EventGPT : This scene depicts the process of a cup breaking. The moment of impact reveals fragments bursting outward, scattering in various directions. Each shard captures the force of the impact, some pieces spinning and tumbling through the air, while a faint dust cloud forms around the destruction..." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 327, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 447 + ], + "type": "text", + "content": "7B and 13B versions in Table 1. To maintain an end-to-end architecture and ensure fair comparison, we represent the event stream as an event image under identical input conditions. We compare our model with state-of-the-art MLLMs, including Qwen2-VL [43], DeepSeek-VL [28], Intern2VL [5], and LLaVA [26]. The results demonstrate that EventGPT exhibits strong capabilities in event understanding, which is mainly attributed to our three-stage training framework that gradually bridges the significant gap between events and language." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "type": "text", + "content": "Qualitative Analysis. In addition, we conduct qualitative assessments on selected challenging scenarios from the Event-Chat test set, specifically those involving extreme lighting conditions or high-speed object dynamics (see Fig. 6 and Fig. 7). The results illustrate that our proposed EventGPT consistently outperforms other models in these difficult settings, demonstrating robust event-driven scene understanding capabilities. For high dynamic range scenarios, such as the tunnel scene in Fig. 6, EventGPT successfully identifies the tunnel environment and issues relevant hazard warnings. Other models, however, frequently misinterpret the scene or fail to detect the tunnel due to the extreme light changes at the exit point. In such conditions, EventGPT demonstrates a unique advantage by effectively capturing high-exposure elements at the tunnel exit and generating timely alerts. For high-speed motion scenarios, such as a cup breaking in Fig. 7, EventGPT demonstrates advanced spatio-temporal understanding by capturing the dynamic rupture sequence of the object. In contrast, LLaVA-7B misinterprets the scene with an inaccurate description, while Qwen2-VL-7B lacks spatio-temporal comprehension, producing a static and incomplete depiction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 326, + 408, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 326, + 408, + 339 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 408, + 339 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "content": "Contribution of Each Component. To explore the impact of each component, we first construct a baseline similar to LLaVA except adopting event frames in the second training stage. Two variants are constructed by incorporating our event-language adapter or spatio-temporal aggregator. As shown in Table 2, our proposed EventGPT achieves higher performance compared to the baseline and variants. Compared to the baseline, incorporating the spatio-temporal aggregator results in respective improvement of " + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "inline_equation", + "content": "3.24\\%" + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "inline_equation", + "content": "2.02\\%" + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "inline_equation", + "content": "2.41\\%" + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "content": " on the DC, CR, and VQA metrics, significantly enhancing the model's capability of capturing temporal dynamics and spatial relationships within event streams, which facilitates to better interpret complex motion patterns and scene changes. Additionally, the event-language adapter brings improvements of " + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "inline_equation", + "content": "2.35\\%" + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "inline_equation", + "content": "1.26\\%" + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "inline_equation", + "content": "1.20\\%" + }, + { + "bbox": [ + 313, + 350, + 555, + 613 + ], + "type": "text", + "content": " on the DC, CR, and VQA metrics respectively over the baseline. This highlights the superiority of our training pipeline. We extend the traditional two-stage paradigm to a three-stage fine-grained progress. By progressively bridging the large cross-modal gap, our model achieves more accurate event stream perception and contextual understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Influence of Temporal Aggregation Size. To explore the effect of temporal aggregation size, we conduct an analysis to determine the optimal temporal window number " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " within the spatio-temporal aggregator. " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " represents the number of temporal bins used to divide a fixed observation period, with each bin aggregating events within its respective interval. Fewer windows may fail to capture fine-grained temporal details, while more windows can capture" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29145" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 123, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 2. Performance impact of each component in EventGPT through ablation study. The results demonstrate that both the spatio-temporal aggregator and the event-language adapter contribute significant performance improvements over the baseline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 123, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 123, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 123, + 294, + 194 + ], + "type": "table", + "html": "
Mode TypeEvent LanguageSpatio TemporalEvent-Chat
DCCRVQA
Baseline××3.403.974.15
Variant A✓×3.48(+2.35%)4.02(+1.26%)4.20(+1.20%)
Variant B×✓3.51(+3.24%)4.05(+2.02%)4.25(+2.41%)
Ours✓✓3.52(+3.53%)4.09(+3.02%)4.29(+3.37%)
", + "image_path": "64b0dda9578c8976de17cc0b933eea42ac33c926e364c45e4347dbc7c5f51410.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 267, + 284, + 334 + ], + "blocks": [ + { + "bbox": [ + 55, + 203, + 296, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 296, + 261 + ], + "type": "text", + "content": "Table 3. Performance comparison of EventGPT-7B across different temporal window numbers " + }, + { + "bbox": [ + 55, + 203, + 296, + 261 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 55, + 203, + 296, + 261 + ], + "type": "text", + "content": ". The results demonstrate that our framework is robust to a wide range of window numbers. However, both too few and too many windows adversely affect performance, with the optimal result achieved at " + }, + { + "bbox": [ + 55, + 203, + 296, + 261 + ], + "type": "inline_equation", + "content": "N_{w} = 5" + }, + { + "bbox": [ + 55, + 203, + 296, + 261 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 267, + 284, + 334 + ], + "lines": [ + { + "bbox": [ + 69, + 267, + 284, + 334 + ], + "spans": [ + { + "bbox": [ + 69, + 267, + 284, + 334 + ], + "type": "table", + "html": "
NwEvent-Chat
35791113
DC3.473.523.563.413.363.32
CR4.034.094.053.963.973.91
VQA4.214.294.174.124.084.03
", + "image_path": "ca03b2eefc6cf1fc342a221b9e27a7e88c4618aedad881d95e71748545f39c64.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 353, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 296, + 473 + ], + "type": "text", + "content": "richer temporal cues but lead to sparser event distribution per bin, making it challenging to model context information effectively. In other words, a small number of windows fail to capture sufficient scene dynamics, while an excessively large number of windows lead to the loss of scene context. The quantitative results in Table 3 reveal that our framework is robust to a wide range of temporal aggregation sizes. Furthermore, a temporal window number of 5 yields the best overall performance. Accordingly, we set " + }, + { + "bbox": [ + 54, + 353, + 296, + 473 + ], + "type": "inline_equation", + "content": "N_w = 5" + }, + { + "bbox": [ + 54, + 353, + 296, + 473 + ], + "type": "text", + "content": " in the spatio-temporal aggregator by default." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 200, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 200, + 493 + ], + "type": "text", + "content": "5.4. Downstream Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "content": "In this section, we assess EventGPT on downstream tasks, object detection, and instance segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 521, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 641 + ], + "type": "text", + "content": "Object Detection. To evaluate the generalization ability of EventGPT on object-level understanding and reasoning, we integrate it with GroundingDINO [27], leveraging its open-set object detection capabilities. For each question, EventGPT generates textual descriptions of potential objects in the event streams, which are subsequently used by GroundingDINO to localize the corresponding objects within the data. As illustrated in Fig. 8, EventGPT demonstrates strong object-level reasoning ability, accurately inferring and localizing target objects in the scenario." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "type": "text", + "content": "Instance Segmentation. To validate EventGPT's scalability on complex dense prediction tasks, we integrate it with GroundedSAM [37] for instance segmentation on event streams. EventGPT provides semantic cues to guide GroundedSAM in generating instance masks, and segmenting individual objects in the scene. The qualitative results" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 553, + 175 + ], + "type": "image", + "image_path": "f981dbd570fedb50d196455414127bc7461cc325dd57a9d4fbadd7d3ff350661.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 184, + 555, + 219 + ], + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 219 + ], + "type": "text", + "content": "Figure 8. Object detection [27] using events and text from our EventGPT. Note that, our EventGPT enables the generation of high-quality text for the open-set object detection task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 553, + 335 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 553, + 335 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 335 + ], + "type": "image", + "image_path": "a43dd892dbf7d7be4c827f2ca6106103a5468fe2c2a69f0e0aee6be05bcfc73d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 343, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 388 + ], + "type": "text", + "content": "Figure 9. Instance segmentation [37] leveraging events and text generated by our EventGPT. Our EventGPT is capable of producing high-quality text descriptions for open-set instance segmentation tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 395, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 444 + ], + "type": "text", + "content": "in Fig. 9 demonstrate our approach's efficacy in achieving segmentation under the challenging conditions inherent to event-based data. This affirms EventGPT's adaptability and potential in fine-grained scene understanding tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 388, + 466 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 473, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 631 + ], + "type": "text", + "content": "This paper introduces EventGPT, the first multimodal large language model(MLLM) tailored for comprehensive event stream understanding. The model comprises an event encoder, a spatio-temporal aggregator, a linear projector, an event-language adapter, and an LLM. It employs a three-stage training pipeline to progressively bridge the significant domain gap between event streams and language for enhanced scene perception and description. Comprehensive evaluations consistently demonstrate the superiority of our method in event scene understanding. Additionally, we build two large-scale event-text datasets, N-ImageNet-Chat and Event-Chat, to advance cross-modal alignment research in the community." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 639, + 411, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 411, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 411, + 654 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 659, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 555, + 708 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0102801, as well as the National Natural Science Foundation of China under Grant 61827804." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29146" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NIPS, 35: 23716-23736, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moratelli, Sara Sarto, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. The (r) evolution of multimodal large language models: A survey. arXiv preprint arXiv:2402.12451, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "type": "text", + "content": "[3] Tom Tongjia Chen, Hongshan Yu, Zhengeng Yang, Ming Li, Zechuan Li, Jingwen Wang, Wei Miao, Wei Sun, and Chen Chen. First place solution to the cvpr'2023 aqtc challenge: A function-interaction centric approach with spatiotemporal visual-language alignment. arXiv preprint arXiv:2306.13380, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 268, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 295, + 312 + ], + "type": "text", + "content": "[4] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In CVPR, pages 14239-14250, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "type": "text", + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "text", + "content": "[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 478 + ], + "type": "text", + "content": "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 295, + 512 + ], + "type": "text", + "content": "[8] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "type": "text", + "content": "[9] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. *IJCV*, 128(3):601-618, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 557, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 590 + ], + "type": "text", + "content": "[10] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "text", + "content": "[11] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE RAL, 6(3):4947-4954, 2021. 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 666 + ], + "type": "text", + "content": "[12] Kunping Huang, Sen Zhang, Jing Zhang, and Dacheng Tao. Event-based simultaneous localization and mapping: A comprehensive survey. arXiv preprint arXiv:2304.09793, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "text", + "content": "[13] Qirui Jiao, Daoyuan Chen, Yilun Huang, Yaliang Li, and Ying Shen. Img-diff: Contrastive data synthesis for multimodal large language models. arXiv preprint arXiv:2408.04594, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[14] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, pages 349-364. Springer, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[15] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, pages 2146-2156, 2021. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 186 + ], + "type": "text", + "content": "[16] Dianze Li, Yonghong Tian, and Jianing Li. Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI, 45(11):14020-14037, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 234, + 553, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 278 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 280, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 334 + ], + "type": "text", + "content": "[19] Ming Li, Xinming Huang, and Ziming Zhang. Self-supervised geometric features discovery via interpretable attention for vehicle re-identification and beyond. In Proceedings of the IEEE/CVF international conference on computer vision, pages 194–204, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 335, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 553, + 380 + ], + "type": "text", + "content": "[20] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 382, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 382, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 382, + 553, + 437 + ], + "type": "text", + "content": "[21] Ming Li, Huazhu Fu, Shengfeng He, Hehe Fan, Jun Liu, Jussi Keppo, and Mike Zheng Shou. Dr-fer: Discriminative and robust representation learning for facial expression recognition. IEEE Transactions on Multimedia, 26:6297-6309, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 438, + 553, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 494 + ], + "type": "text", + "content": "[22] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. Stprivacy: Spatio-temporal privacy-preserving action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5106-5115, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 495, + 553, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 495, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 317, + 495, + 553, + 538 + ], + "type": "text", + "content": "[23] Ming Li, Pan Zhou, Jia-Wei Liu, Jussi Keppo, Min Lin, Shuicheng Yan, and Xiangyu Xu. Instant3d: instant text-to-3d generation. International Journal of Computer Vision, 132(10):4456-4472, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 540, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 540, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 317, + 540, + 553, + 574 + ], + "type": "text", + "content": "[24] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 576, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 576, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 576, + 553, + 609 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 611, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 553, + 632 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 36, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[28] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29147" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "text", + "content": "[29] Yecheng Lyu, Ming Li, Xinming Huang, Ulkuhan Guler, Patrick Schaumont, and Ziming Zhang. Treernn: Topology-preserving deep graph embedding and learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7493-7499. IEEE, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "type": "text", + "content": "[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 239 + ], + "type": "text", + "content": "[31] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In CVPR, pages 5642-5651, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 271 + ], + "type": "text", + "content": "[32] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. Esl: Event-based structured light. In 3DV, pages 1165-1174. IEEE, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "text", + "content": "[33] Rohan Pandey, Rulin Shao, Paul Pu Liang, Ruslan Salakhutdinov, and Louis-Philippe Morency. Cross-modal attention congruence regularization for vision-language relation alignment. arXiv preprint arXiv:2212.10549, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 371 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "text", + "content": "[35] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. *IJCV*, 126 (12):1394–1414, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 448 + ], + "type": "text", + "content": "[36] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. CVPR, 2019. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[37] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024.8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 503, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 557 + ], + "type": "text", + "content": "[38] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In CVPR, pages 13193-13203, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "text", + "content": "[39] Yichen Shen, Yijin Li, Shuo Chen, Guanglin Li, Zhaoyang Huang, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinktrack: Feature tracking over 100 fps via events and images. arXiv preprint arXiv:2409.17981, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 647 + ], + "type": "text", + "content": "[40] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, pages 15638-15650, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "text", + "content": "[41] Lulu Tian, Hongxun Yao, and Ming Li. Fakepei: A large-scale fake person of interest video detection benchmark and a strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6819-6831, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "text", + "content": "[42] Guanqun Wang, Jiaming Liu, Chenxuan Li, Yuan Zhang, Junpeng Ma, Xinyu Wei, Kevin Zhang, Maurice Chong," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Renrui Zhang, Yijiang Liu, et al. Cloud-device collaborative learning for multimodal large language models. In CVPR, pages 12646-12655, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 162, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 195 + ], + "type": "text", + "content": "[44] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "text", + "content": "[45] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 264 + ], + "type": "text", + "content": "[46] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. In CVPR, pages 16933-16943, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[47] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024.1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "text", + "content": "[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[49] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. Eventps: Real-time photometric stereo using an event camera. In CVPR, pages 9602–9611, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[50] Yun Yue, Ming Li, Venkatesh Saligrama, and Ziming Zhang. Rnn training along locally optimal trajectories via frankwolfe algorithm. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 10532-10539. IEEE, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "text", + "content": "[51] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, et al. Anygpt: Unified multimodal llm with discrete sequence modeling. arXiv preprint arXiv:2402.12226, 2024.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[52] Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, and Dong Yu. Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[53] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, pages 8801-8810, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "text", + "content": "[54] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[55] Yan Zhang, Chun Li, Zhaoxia Liu, and Ming Li. Semi-supervised disease classification based on limited medical image data. IEEE Journal of Biomedical and Health Informatics, 28(3):1575-1586, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "text", + "content": "[56] Ce Zheng, Yecheng Lyu, Ming Li, and Ziming Zhang. Lodonet: A deep neural network with 2d keypoint matching" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29148" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 352 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "for 3d lidar odometry estimation. In Proceedings of the 28th ACM international conference on multimedia, pages 2391-2399, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[57] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arxiv 2023. arXiv preprint arXiv:2302.08890, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 206 + ], + "type": "text", + "content": "[58] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. E-clip: Towards label-efficient event-based open-world understanding by clip. arXiv e-prints, pages arXiv-2308, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 252 + ], + "type": "text", + "content": "[59] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In CVPR, pages 18633-18643, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 297 + ], + "type": "text", + "content": "[60] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "type": "text", + "content": "[61] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29149" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_content_list.json b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5e515c027aba533739e2167d6fcc820decc218a --- /dev/null +++ b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_content_list.json @@ -0,0 +1,1644 @@ +[ + { + "type": "text", + "text": "EventPSR: Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera", + "text_level": 1, + "bbox": [ + 191, + 130, + 805, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bohan $\\mathrm{Yu}^{1,2\\dagger}$", + "bbox": [ + 245, + 202, + 352, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jin Han3,4", + "bbox": [ + 385, + 203, + 467, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boxin Shi $^{1,2*}$", + "bbox": [ + 506, + 203, + 612, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imari Sato3,4", + "bbox": [ + 647, + 203, + 750, + 219 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University", + "$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + ], + "bbox": [ + 165, + 223, + 828, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ The University of Tokyo $^{4}$ National Institute of Informatics", + "bbox": [ + 302, + 257, + 692, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ybh1998@pku.edu.cn, jinan@nii.ac.jp, shiboxin@pku.edu.cn, imarik@nii.ac.jp", + "bbox": [ + 161, + 277, + 825, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simultaneously acquisition of the surface normal and reflectance parameters is a crucial but challenging technique in the field of computer vision and graphics. It requires capturing multiple high dynamic range (HDR) images in existing methods using frame-based cameras. In this paper, we propose EventPSR, the first work to recover surface normal and reflectance parameters (e.g., metallic and roughness) simultaneously using an event camera. Compared with the existing methods based on photometric stereo or neural radiance fields, EventPSR is a robust and efficient approach that works consistently with different materials. Thanks to the extremely high temporal resolution and high dynamic range coverage of event cameras, EventPSR can recover accurate surface normal and reflectance of objects with various materials in 10 seconds. Extensive experiments on both synthetic data and real objects show that compared with existing methods using more than 100 HDR images, EventPSR recovers comparable surface normal and reflectance parameters with only about $30\\%$ of the data rate.", + "bbox": [ + 88, + 358, + 483, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 676, + 220, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The fast and accurate acquisition of the shape and appearance information of an object is essential for many tasks in computer vision and computer graphics. This technique enables re-rendering realistic digital twins of real-world objects, supporting a wide range of downstream applications, such as virtual reality, augmented reality (AR), and digital content creation. However, dynamic range is a big issue when the material is composed of diffuse and highly specular properties. The time-consuming capturing process and", + "bbox": [ + 89, + 702, + 483, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeef97a8dd3bf674f4849eff72787baa4e0d04db33c5ece325b3c5f38623a76e.jpg", + "image_caption": [ + "Figure 1. The EventPSR applies a sequence of light scanning patterns and an event camera to recover the accurate normal map and reflectance parameters (i.e., metallic and roughness) simultaneously. A two-stage pipeline consisting of a grid-matching and a gradient-tuning stage is proposed for EventPSR, which supports shape and reflectance recovery for different materials." + ], + "image_footnote": [], + "bbox": [ + 526, + 330, + 893, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "difficulty recovering highly specular surfaces are two major challenges in this area.", + "bbox": [ + 511, + 790, + 906, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional shape and reflectance estimation methods [1, 3, 10, 15, 28, 31, 32] rely on multiple high dynamic range (HDR) images captured under varying lighting conditions or viewing angles. Different from photometric stereo (PS) that estimates the surface normal of diffuse dominant ob", + "bbox": [ + 511, + 824, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Boxin Shi", + "bbox": [ + 89, + 849, + 277, + 862 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was conducted while the first author, Bohan Yu, was doing internship at National Institute of Informatics.", + "bbox": [ + 91, + 863, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code available: https://codeberg.org/ybh1998/EventPSR", + "bbox": [ + 91, + 888, + 475, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jects from several images, shape and appearance estimation methods [1, 15] require more images to ensure the accurate observation of various reflections, which is time-consuming especially for shiny materials recovery. It is difficult to fully cover the specular highlights in shiny and metallic (i.e., mirror-like) materials under directional light without complicated capturing setup [15] and multiple images to cover the dynamic range. Such a dilemma between speed and accuracy partly attributes to the insufficient capability of conventional frame-based cameras.", + "bbox": [ + 88, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event cameras characterized by extremely high temporal resolution and dynamic range, have shown promising advantages in many computer vision applications [9]. Researchers have explored applying event cameras in various photometric tasks [6, 19, 29, 33]. Particularly, EventPS [29] has demonstrated the superiority of event cameras for real-time surface normal estimation, showing the potential to accelerate and improve accuracy for shape and reflectance estimation tasks. However, it is still unclear how to extract material-related parameters, including roughness and metallic, from the event stream only.", + "bbox": [ + 89, + 242, + 482, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose EventPSR, a novel solution for simultaneous estimation of both surface normal and Reflectance properties from Photometric Stereo using an Event camera. As shown in Fig. 1, EventPSR can simultaneously recover an object's shape and reflectance properties $^{2}$ with designed light scanning patterns and a two-stage algorithm, which supports the recovery of different materials. Directly applying the setting and algorithm from EventPS [29] (i.e., capturing objects with a fixed event camera under a fast-moving point light source) is a non-trivial task. It is impossible to estimate the shape and reflectance at the same time without conquering the two challenges: 1) Design a suitable light scanning pattern that covers various materials and fully reveals the shape and reflectance properties of the objects. 2) Propose a new algorithm that converts the event signals to accurate normal and reflectance parameters simultaneously under the light scanning pattern.", + "bbox": [ + 91, + 409, + 482, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conquer the first challenge, we discover that there are three requirements that need to be satisfied for a suitable light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. Based on this, we design a pattern of \"1D strip light\" that covers the full sphere of the illumination direction space. It is designed not only for various materials ranging from ideal Lambertian to pure mirror, but also for efficient capture of event cameras with faster speed and low data rate. To conquer the second challenge, we propose a two-stage algorithm to estimate surface normal and reflectance parameters, consisting of a grid-matching stage and a gradient-tuning stage. We first render a database consisting of the intensity-changing curves, including the com", + "bbox": [ + 89, + 667, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "binations of various surface normal directions and materials. In the grid-matching stage, we match the event stream triggered in each pixel to a specific curve in the database and get a group of coarse reflectance parameters. In the gradient-tuning stage, we finetune the estimated parameters to a more accurate sub-grid location. To summarize, the contributions of this paper are as follows:", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EventPSR, the first solution to estimate surface normal and reflectance parameters simultaneously from continuous radiance differential event signals.", + "- We design a suitable light scanning pattern and an efficient two-stage algorithm with only $30\\%$ data rate compared with frame-based counterparts.", + "- We validate the effectiveness of EventPSR by a platform with multiple monitors. The results show that EventPSR can estimate accurate shape-related and material-related parameters of real objects." + ], + "bbox": [ + 511, + 198, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 362, + 661, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is related to PS works, and comprehensive surveys can be found in [14, 25]. This section focuses on works addressing PS and surface reflectometry using conventional frame-based cameras, as well as recent advancements in photometry using event cameras.", + "bbox": [ + 511, + 387, + 903, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Surface reflectometry. Measuring the detailed surface reflectance requires the use of controlled lighting to sample photometric stereo images. Romeiro et al. [23] inferred the bi-directional surface reflectance of materials using a light probe and a single HDR image of the known surface normal. Dupuy and Jakob [8] measured the reflectance spectra using a gonio-photometer to simultaneously handle BRDF acquisition, storage, and rendering. Learning-based BRDF acquisition and modeling approaches are summarized in the survey paper [17]. EventPSR differs from this category of research because we simultaneously recover the normal and reflectance parameters.", + "bbox": [ + 511, + 462, + 905, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Normal and reflectance estimation from conventional cameras. Alldrin et al. [1] proposed to employ novel bivariate approximations of isotropic reflectance functions to simultaneously recover the shape and reflectance of a surface from photometric stereo images. Chung et al. [7] optimized the geometry and spatially-varying BRDFs through a point-based rendering. Learning-based methods applied convolutional neural networks on single image [2, 21, 30] or multiple images [15, 24] to estimate the surface normal and reflectance parameters. These methods are limited in capturing speed due to the low dynamic range and high latency of conventional cameras.", + "bbox": [ + 511, + 643, + 905, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event camera-based photometry. Event cameras with high temporal resolution and high dynamic range properties have been applied to various photometry tasks in recent years. Muglikar et al. [19] put a linear polarizer rotating at high speeds in front of an event camera to reconstruct", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2In this work, we recover reflectance-related properties specifically with roughness and metallic.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11428", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relative intensities at multiple polarizer angles, which were used for surface normals estimation. Han et al. [11] computed the frequencies of the event triggering during the transient process of global lighting changes for depth sensing and object iso-contour reconstruction. Zhou et al. [33] proposed to use the differential signals captured from an event camera to separate the direct and global illumination components in RGB videos by applying a sweeping occluder. EventPS [29] leveraged an event camera and a fast-rotating light source to achieve real-time surface normal reconstruction, which demonstrated the superiority over the conventional cameras in speed and data bandwidth. EventPSR can further estimate surface normal and reflectance parameters with high robustness and efficiency by designing light scanning patterns and estimation algorithms.", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 330, + 181, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the overall workflow of EventPSR includes the design of suitable lighting scanning patterns (Sec. 3.2) and a two-stage algorithm that recovers surface normal and reflectance parameters (Sec. 3.3). In Sec. 3.1, we start from the mathematical formulation of event cameras, parameterized reflectance models, and target function.", + "bbox": [ + 89, + 354, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 454, + 294, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event formation model. Event cameras measure the scene brightness changes asynchronously on a logarithmic scale. An event $e = \\{t, x, y, p\\}$ is triggered when the changes of logarithmic brightness reach a triggering threshold $C$ , where $t$ is the timestamp, $(x, y)$ is the pixel coordinate, and $p \\in \\{-1, +1\\}$ is the polarity which represents the decrease or increase of brightness. Assuming there are totally $K$ events triggered at pixel $(x, y)$ during a short period of time, these events are represented as $\\mathcal{E}_{x,y} = \\{x, y, p_k, t_k\\}$ , where $k = \\{1, 2, \\dots, K\\}$ . The irradiance value $I$ in pixel $(x, y)$ changing from $t_{k-1}$ to $t_k$ is represented as:", + "bbox": [ + 89, + 477, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\log \\left(I _ {x, y} \\left(t _ {k}\\right) + \\epsilon\\right) = \\log \\left(I _ {x, y} \\left(t _ {k - 1} + \\eta\\right) + \\epsilon\\right) + p _ {k} C, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 652, + 482, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a small offset value to avoid taking the logarithm of zero, and $\\eta$ is the refractory time of the pixel [9].", + "bbox": [ + 89, + 679, + 482, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reflectance model and inverse problem. We use a simplified version of Disney Principled BRDF [4] model to represent the reflectance $(\\mathbf{f}_{r,x,y})$ of the surface. Given an opaque and non-emissive object with a distant light source, the irradiance value $I_{x,y}(t)$ at timestamp $t$ is as follows:", + "bbox": [ + 89, + 709, + 482, + 786 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {x, y} (t) = \\int_ {\\mathbf {w} _ {i} \\in \\Omega} \\left(\\mathbf {N} _ {x, y} \\cdot \\mathbf {w} _ {\\mathrm {i}}\\right) \\mathbf {f} _ {r, x, y} \\left(\\mathbf {w} _ {\\mathrm {i}}, \\mathbf {w} _ {\\mathrm {r}}\\right) \\mathbf {L} \\left(\\mathbf {w} _ {i}, t\\right) \\mathrm {d} \\mathbf {w} _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 794, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Omega$ is the whole sphere direction space, $\\mathbf{N}_{x,y}$ is the surface normal, $\\mathbf{L}$ is the light source function, $\\mathbf{w}_i$ and $\\mathbf{w}_r$ are the incident and reflection direction, respectively. The light source function $\\mathbf{L}(\\mathbf{w}_i,t)$ represents a distant light with", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the same brightness at an angle of $\\mathbf{w}_i$ . For each pixel $(x,y)$ , its reflectance $\\mathbf{f}_{r,x,y}$ is represented as:", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {r, x, y} \\left(\\mathbf {w} _ {\\mathrm {i}}, \\mathbf {w} _ {\\mathrm {r}}\\right) = \\frac {1 - M _ {x , y}}{\\pi} + \\frac {M _ {x , y} D G}{\\pi \\left(\\mathbf {N} _ {x , y} \\cdot \\mathbf {w} _ {\\mathrm {i}}\\right) \\left(\\mathbf {N} _ {x , y} \\cdot \\mathbf {w} _ {\\mathrm {r}}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 132, + 903, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D, G$ are parameters that describe specular reflection property. $^3$ The variation of diffuse albedo has been canceled by the log difference mechanism in event cameras [29], so we set it as 1 in our method. Other parameters and Fresnel term from Disney Principled BRDF [4] are removed for simplicity.", + "bbox": [ + 511, + 179, + 906, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Target function. To recover the surface properties, i.e., $\\mathbf{N}_{x,y}$ (Surface normal), $R_{x,y}$ (Roughness, in the supplementary material), and $M_{x,y}$ (Metallic), we design a light scanning pattern $\\mathbf{L}(\\mathbf{w}_i,t)$ and apply an event camera to solve the inverse reflectance estimation problem. We represent the shape and reflection estimation algorithm as $\\tilde{\\mathbf{N}} (\\mathcal{E})$ , $\\tilde{\\mathbf{R}} (\\mathcal{E})$ , and $\\tilde{\\mathbf{M}} (\\mathcal{E})$ . Our target is to design these algorithms and optimize light scanning patterns $\\mathbf{L}$ to estimate the surface normal and reflectance parameters:", + "bbox": [ + 511, + 290, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathbf {L}, \\tilde {\\mathbf {N}}, \\tilde {\\mathbf {R}}, \\tilde {\\mathbf {M}}} {\\operatorname {a r g m i n}} \\sum_ {\\mathbf {N}, M, R} | \\tilde {\\mathbf {N}} (\\mathcal {E}) - \\mathbf {N} | _ {2} + \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 436, + 903, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\tilde {\\mathbf {M}} (\\mathcal {E}) - M \\right| _ {2} + \\left| \\tilde {\\mathbf {R}} (\\mathcal {E}) - R \\right| _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 473, + 841, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the events $\\mathcal{E}$ are triggered from $\\mathbf{L},\\mathbf{N},R,M$", + "bbox": [ + 511, + 503, + 849, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Light Scanning Pattern Design", + "text_level": 1, + "bbox": [ + 511, + 530, + 785, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is necessary to design a suitable light scanning pattern to cover different materials and the shape of target objects. There are three requirements when designing the light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. In the following, we consider an effective use of a ring-shaped light source as the base scanning pattern. Taking a ring moving along the $Z$ axis as an example, the formula of $\\mathbf{L}(\\mathbf{w}_i,t)$ is:", + "bbox": [ + 511, + 551, + 906, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} \\left(\\mathbf {w} _ {i}, t\\right) = \\max _ {u} \\max (0, \\mathbf {w} _ {i} \\cdot \\mathbf {R} (u, t)) ^ {\\alpha}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 685, + 903, + 712 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} (u, t) = [ \\sin (t) \\sin (u), \\sin (t) \\cos (u), \\cos (t) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 710, + 864, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{R}(u,t)$ represent all light directions on a ring. $\\alpha$ is the width of the ring. In order to illustrate that our \"moving ring light\" pattern has better properties to overcome all these challenges, we compare it with two other light source patterns: \"point light\" and \"structured environment map\".", + "bbox": [ + 511, + 739, + 906, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the \"point light\" pattern, we combine two spiral movement trajectories to cover the whole sphere: the horizontal spiral and vertical spiral. In this way, we obtain a continuous moving light source pattern. We also add a", + "bbox": [ + 511, + 816, + 906, + 876 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3Detailed formulas are shown in the supplementary material.", + "bbox": [ + 513, + 886, + 834, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11429", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01d73d84d2c4bf9c9107e76ea05204d30cb3e36ae8cde7e1591567d04b495d14.jpg", + "image_caption": [ + "Figure 2. The design of light scanning pattern. (a) Illustration of specular coverage requirement. The first row shows a shiny ball under three lighting patterns. The second row shows where specular parameters can be recovered (the brighter the better). (b) Illustration of diffuse sensitivity requirement. The first row shows the rendered frames of a diffuse ball under three lighting patterns. The second row shows the profiles of intensity changes of two different points on the sphere over time. (c) Illustration of event efficiency requirement. The first row shows the event signals triggered under three lighting patterns. The second row plots the average (green) and max (red) number of events triggered over time when covering the whole specular reflection." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "smoothness term $\\varepsilon$ (the size of the point light), so the lighting will cover the whole sphere lighting direction gradually. For the horizontal spiral, the formula $\\mathbf{L}(\\mathbf{w}_i,t)$ is:", + "bbox": [ + 89, + 352, + 482, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {L} (\\mathbf {w} _ {i}, t) = (\\mathbf {R} (t) \\cdot \\mathbf {w} _ {i}) ^ {\\alpha}, \\\\ \\mathbf {R} (t) = [ \\begin{array}{l l l l l l} \\cdot & (\\partial_ {t}) & \\cdot & (\\partial_ {t}) & \\cdot & (\\partial_ {t}) \\end{array} ] \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 419, + 482, + 444 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} (t) = [ \\sin (\\beta t) \\cos (t), \\cos (\\beta t) \\cos (t), \\sin (t) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 439, + 441, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ determines the number of spirals. For the vertical spiral, the formula $\\mathbf{L}(\\mathbf{w}_i,t)$ is:", + "bbox": [ + 89, + 465, + 482, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {L} (\\mathbf {w} _ {i}, t) = (\\mathbf {R} (t) \\cdot \\mathbf {w} _ {i}) ^ {\\alpha}, \\\\ \\text {T r} (t) = [ (0) + (1) + (2) + (3) + (4) ] \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 518, + 482, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} (t) = [ \\cos (\\beta t) \\sin (t), \\cos (\\beta t) \\cos (t), \\sin (\\beta t) ].\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 537, + 436, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the \"structured environment map\" pattern, we project structured light along three axes alternatively. Taking structured light along the $Z$ axis as an example, the formula of this light source pattern $\\mathbf{L}(\\mathbf{w}_i,t)$ is:", + "bbox": [ + 89, + 563, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} (\\mathbf {w} _ {i}, t) = \\left\\lfloor \\mathrm {f m o d} (\\mathbf {R} (2 ^ {t} \\mathbf {w} _ {i}), 2) \\right\\rfloor ,\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 645, + 344, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} \\left(\\mathbf {w} _ {i}\\right) = \\frac {\\operatorname {a t a n 2} \\left(\\mathbf {w} _ {i} \\cdot [ 0 , 1 , 0 ] , \\mathbf {w} _ {i} \\cdot [ 1 , 0 , 0 ]\\right)}{\\pi} + 1, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 662, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where \"fmod\" is the modulo operator. $t$ should be integer and the \"structured environment map\" pattern is discrete.", + "bbox": [ + 89, + 704, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specular coverage requirement. Given metallic (i.e., without diffuse part) or even mirror-like objects, it is difficult to use single distant point light for surface normal estimation. Because the specular surface only responds to lighting from a small lobe. For the specular dielectric material (i.e., with a diffuse part), the surface normal can be estimated from the diffuse reflection. However, the material properties are still only recoverable in the small specular highlight part. We show rendered images of a shiny ball under all three lighting setups in the top row of Fig. 2 (a). The bottom row shows that the solvable regions for \"point", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "light” is a small circle around the lighting direction. In contrast, the “moving ring light” and “structured environment map” have larger solvable regions with a strip around the ring or structure edges, which is more efficient for specular coverage. The larger specular coverage results in faster data-capturing speed for challenging specular surface materials.", + "bbox": [ + 511, + 352, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffuse sensitivity requirement. Diffuse reflection has the opposite property from specular reflection: it spreads the input light to all the upper hemisphere directions instead of only a small highlight region. In order to estimate surface normal from the reflection, the light source pattern needs to contain high and diverse low-frequency parts to make it distinguishable from different surface normal. As shown in Fig. 2 (b), we render the images of a rough diffusive ball under all three lighting setups in the top row, and plot the reflection intensity of two pixels in the bottom row. The diffuse reflection of two pixels is shifted in time for \"point light\" and \"moving ring light\" patterns. Since event cameras have an extremely high temporal resolution, the time shift can be well-captured by the camera and converted to surface normal accurately. In contrast, the diffuse reflection intensity profiles for the two pixels are similar for \"structured environment map\", making it difficult to distinguish the surface normal of two points.", + "bbox": [ + 511, + 460, + 908, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event efficiency requirement. Since the event camera measures the brightness changes, the bandwidth consumption is mainly determined by the brightness changing magnitude in the scene. We show the synthetic event signals and the accumulated number of events from rendered a mirror-like ball in Fig. 2 (c). Both \"point light\" and \"moving ring light\" have a similar number of events triggered by different surface normal (observed by two pixels). The \"moving ring light\" pattern consumes fewer events for specular coverage than the \"point light\", indicating a better bandwidth efficiency. In contrast, the \"structured environment map\"", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11430", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "leads to a variable number of events across different surface normal. In the worst case, where there is a mirror plane facing at the worst surface normal with a maximum number of events generated, the whole scene is flickering, and the read-out will be congested.", + "bbox": [ + 89, + 90, + 480, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the comparisons and analysis above, the designed \"moving ring light\" achieves relatively larger coverage of various materials while preserving higher efficiency of event triggering.", + "bbox": [ + 89, + 167, + 480, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.EventPSR Algorithm", + "text_level": 1, + "bbox": [ + 89, + 234, + 290, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 1, we propose a two-stage solving pipeline, consisting of a grid-matching stage and a gradient-tuning stage. The proposed algorithm provides a global minimum result with high value resolution.", + "bbox": [ + 89, + 257, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage-1: Grid-matching. To estimate the surface normal and reflectance parameters from the captured event streams, we first render a database containing various intensity-time curves using a physically-based rendering engine. We assume that the rendered temporal resolution is high enough so that the intensity curves can be represented by linear interpolation of a series of frames. There are 4 parameters to be estimated for each pixel, including 2 parameters for surface normal, 1 parameter for roughness, and 1 parameter for metallic. The database is on a 5-dimension grid (1 additional dimension for time), which is defined as $\\mathcal{B} = \\{\\mathbf{N}, R, M, t\\}$ .", + "bbox": [ + 89, + 318, + 482, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a series of event signals $\\mathcal{E}_{x,y}$ at a pixel $(x,y)$ , the goal in this stage is to find the most matching curve in the database. Therefore, we define the loss function between the observed event stream and curve candidates as follows:", + "bbox": [ + 89, + 500, + 482, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {1} (x, y, \\mathbf {N}, R, M) = \\sum_ {k} | \\log (\\mathcal {B} (\\mathbf {N}, R, M, t _ {k}) + \\epsilon) \\tag {9} \\\\ - \\log (\\mathcal {B} (\\mathbf {N}, R, M, t _ {k - 1} + \\eta) + \\epsilon) - p _ {k} C | ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 566, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By designing the loss function on each pair of events, we prevent the event integral process, which has an integration drifting issue (More details about the drifting issue are shown in the supplementary material). Furthermore, our loss function averages errors of all events in the same pixel, implicitly suppressing event triggering noise.", + "bbox": [ + 89, + 628, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We enumerate 4D surface normal $\\mathbf{N}$ and material properties $R, M$ on a discrete grid and select the one with minimum loss. The results $\\tilde{\\mathbf{N}}, \\tilde{R}, \\tilde{M}$ at this stage are as follows:", + "bbox": [ + 89, + 719, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {N}} _ {x, y}, \\tilde {R} _ {x, y}, \\tilde {M} _ {x, y} = \\underset {\\mathbf {N}, R, M} {\\operatorname {a r g m i n}} \\mathcal {L} _ {1} (x, y, \\mathbf {N}, R, M). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 773, + 482, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage-2: Gradient-tuning. The grid-matching stage provides us with a coarse result close to the global minimum, but this result can be further improved for the following two reasons: 1) The resolution of grid-matching stage depends on the searching grid size. Due to the computational cost, this resolution is not high enough. 2) In real setup, there", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/41db8ea86cac9fdab44bb80aa8ac59f482a82782841ffb5e8ab23059cc814a71.jpg", + "image_caption": [ + "Data acquisition platform" + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 679, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd1d41468fa66b21ba52cd560a71e2b112a3bba29476fb5df48284287bb0f68e.jpg", + "image_caption": [ + "Light scanning patterns" + ], + "image_footnote": [], + "bbox": [ + 686, + 90, + 792, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/11377ca4609d62b4dfba6fd28c65da71b93135d613e3694e1eaed49251c86108.jpg", + "image_caption": [ + "Figure 3. Data acquisition platform for capturing real data. There are 5 commodity monitors surrounding the target object to illuminate the designed pattern from any incidence angle. We show the capture device setup on the left side. The target object is hung in the middle and an event camera is pointing at the target from the bottom. Three projected patterns are shown on the right." + ], + "image_footnote": [], + "bbox": [ + 795, + 90, + 901, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are some regions of the environment map that do not have a display. If the material is shiny, metallic, and the surface normal is reflecting to the dark region, there is no brightness change, which leads to ambiguity in these regions. More details are shown in the supplementary material.", + "bbox": [ + 511, + 335, + 903, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve the first problem, we calculate the gradient of surface normal $\\mathbf{N}$ and material properties $R, M$ , and interpolate the database on these parameters. In this way, we obtain an approximated rendered intensity and its gradient in a continuous parameter space. Therefore, we minimize the $\\mathcal{L}_1(x,y,\\mathbf{N},R,M)$ using gradient descent to get the optimized result.", + "bbox": [ + 511, + 412, + 905, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve the second problem, we introduce a small inter-pixel smoothness regularization term:", + "bbox": [ + 511, + 518, + 903, + 549 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {2} = \\beta \\sum_ {x, y} \\left(\\left| \\mathbf {N} _ {x, y} - \\mathbf {N} _ {x - 1, y} \\right| ^ {2} + \\left| \\mathbf {N} _ {x, y} - \\mathbf {N} _ {x, y - 1} \\right| ^ {2} + \\right. \\\\ \\left| R _ {x, y} - R _ {x - 1, y} \\right| ^ {2} + \\left| R _ {x, y} - R _ {x, y - 1} \\right| ^ {2} + \\tag {11} \\\\ \\left. \\left| M _ {x, y} - M _ {x - 1, y} \\right| ^ {2} + \\left| M _ {x, y} - M _ {x, y - 1} \\right| ^ {2}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 558, + 903, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, for shiny and metallic surface, the parameters in ambiguous regions are filled with interpolated values, while the diffusive or dielectric surfaces are not affected. The optimized parameters of this gradient-tuning are the final results of EventPSR.", + "bbox": [ + 511, + 643, + 905, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 733, + 645, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 757, + 730, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data acquisition platform. To illuminate the target object using our designed light scanning pattern, we build a data acquisition platform with multiple commodity monitors, as illustrated in Fig. 3. The target object is hanging at the center of the device and surrounded by several monitors. There is a Prophesee EVK4 event camera beneath the target object, pointing above at the object and capturing the event stream as the light scanning pattern changes.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7825355590b4d8a870f05629cc026aa99e0b929e08b8b8f67db4532786ae279f.jpg", + "image_caption": [ + "Figure 4. Quantitative comparison on synthetic data. We show the relative data consumption and per-object averaged error for normal, roughness, and metallic estimation. The color heatmap for each parameter (normal, roughness, and metallic) is normalized independently. The results from NeILF [31] are empty for some objects, which are marked as \"Failed\"." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 903, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c31fc2e87bf4cff58bb8bbe582728be0a83ae039ad4030b4078c02b8a63ef08b.jpg", + "image_caption": [ + "Figure 5. Quantitative comparison results on synthetic data. We show the estimated normal, roughness, metallic, and error maps of EventPSR and two comparing methods. The averaged error maps are labeled in red numbers." + ], + "image_footnote": [], + "bbox": [ + 93, + 287, + 483, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The light scanning pattern received by the target object is slightly different from the designed one because of some non-ideal factors. The uncovered area at the top and bottom will cause some missing illumination, and the interreflection of the monitors will cause some extra light from the wrong direction. To alleviate the influences caused by non-ideal light scanning pattern projection, we capture the received illumination by a mirror ball in the scene. The database $\\mathcal{B}(\\mathbf{N}, R, M, t)$ is rendered using the calibrated environment maps. In addition, we flash all panels to full white at the beginning and end of the illumination pattern for temporal alignment.", + "bbox": [ + 88, + 612, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm implementation. Both of the two stages are implemented using PyTorch [20]. For the grid-matching stage, the resolution of rendered database $\\mathcal{B}(\\mathbf{N}, R, M, t)$ are as follows: $64 \\times 64$ for normal resolution, 12 for roughness resolution, 8 for metallic resolution, and 600 for temporal resolution. According to Eq. (3), the reflectance model is a linear interpolation from pure diffuse to pure mirror. To", + "bbox": [ + 88, + 795, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reduce GPU VRAM consumption, we replace the metallic dimension using 1 pure diffuse component and 1 pure mirror component. Different material metallic properties are constructed on the GPU on the fly. Both of the grid-matching and gradient-tuning stages are conducted on a single NVIDIA GeForce RTX 3060 graphics card.", + "bbox": [ + 511, + 287, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dataset", + "text_level": 1, + "bbox": [ + 511, + 388, + 609, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic data. To evaluate the accuracy of our method compared with other methods, we render 16 objects using the geometries from the sculpture dataset [27]. The material parameters including albedo, roughness, and metallic, are set randomly per patch. We use a ray-tracing-based rendering engine, with 1024 samples per pixel and 3 maximum bounces. The rendered images are of $512 \\times 512$ resolution. An example is shown in the first column in Fig. 5. Each of the objects is rendered under 3 different setups: Our illumination with fixed camera position, DiLiGenT10² [22] light directions with fixed camera position (for the comparison with SDM-UniPS [12]), and fixed light direction with different camera positions (for the comparison with NeILF [31]). For the first setup, the event streams are simulated from high-frame-rate image sequences. The event triggering threshold noise is set to 0.01. For the other two setups, there are total 100 HDR images rendered per object. Besides, the independent Gaussian noise of 0.01 standard variance is added to each image.", + "bbox": [ + 511, + 410, + 906, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real data. We capture three types of real data using our data acquisition platform. The captured data are used for material roughness accuracy evaluation, surface normal accuracy evaluation, and complex object generalization evaluation, respectively.", + "bbox": [ + 511, + 698, + 906, + 773 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Surface normal evaluation data: We select several pairs of objects with almost identical geometry but extremely different materials (i.e., rough and shiny), as shown in the first row in Fig. 6. The ground truth of surface normal is captured by a structured light 3D scanner. $^4$", + "- Material roughness evaluation data: We create several" + ], + "bbox": [ + 511, + 773, + 908, + 864 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4https : / / www . shining3d . com / professional - solutions/Desktop- 3d-scanner/einscan-se-v2", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11432", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1831b2d7033b5eaeec6e46c8dac84ee06282778fec2d4fd5bb1b49a1f4cddf1e.jpg", + "image_caption": [ + "Figure 6. Top: Quantitative evaluation for surface normal estimation on real objects. The error value of each object is labeled in red number. Bottom: Relighting results based on the estimated surface parameters. The environment maps are labeled on the left. The albedo of rendered objects are set to uniform white." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 480, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "balls with varying surface roughness. A polished metal ball serves as the low roughness reference. For the high roughness surface, we apply a spray coating to achieve a diffuse surface. The ground truth values of roughness are measured using a frame-based camera with HDR exposure under ideal parallel light.", + "bbox": [ + 102, + 734, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Generalization evaluation data: We select multiple objects whose surface are composed of multiple different materials, as shown in the first row in Fig. 8. The complexity of the shape and materials will test the generalization ability of EventPSR on real-world objects.", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b3f6b9b980bcf2abb6e2c3ca3203cddb26f79e13241e196166c289b785abd5a.jpg", + "image_caption": [ + "Figure 7. Quantitative demonstration for robustness of material estimation on real objects. The first row is a metal ball with mirror-like surface material. The second row is a ball sprayed with a non-glossy coating." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on Synthetic Data", + "text_level": 1, + "bbox": [ + 511, + 328, + 774, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct quantitative evaluations of the surface normal and reflectance parameters estimation with the synthetic dataset. The comparison includes two frame-based methods [13, 31] that represent the state-of-the-art NeRF-based [31] and PS-based [13] normal and reflectance estimation approaches. For data rate usage comparison, we estimate the date consumption in the following way: For EventPSR, we assume that the input event streams employ 16-bit Prophesee EVT 3.0 format. For the other two comparison methods, we assume that the HDR input are captured by merging 3 multi-bracketing 8-bit grayscale images. We use mean angular error to measure the accuracy of recovered normal maps. For the recovered roughness and metallic, we first standardize the data to avoid representing differences between models, then we apply mean squared error (MSE) to measure the accuracy. More details are described in the supplementary material.", + "bbox": [ + 511, + 353, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The data consumption and parameters error of each object are shown in Fig. 4. On average, our method consumes about $30\\%$ of the data rate, achieving comparable accuracy of normal estimation and superior accuracy of roughness and metallic estimation. We visualize the result of one example from the synthetic dataset in Fig. 5. We can see that the error of our method concentrates mainly on the high geometry change area for all three parameter maps, where the inter-reflection and cast shadow are most severe. For the other two comparing methods [13, 31], the errors of normal estimation are more evenly distributed. However, the estimated roughness and metallic from the comparing methods [13, 31] can hardly reflect the high frequency of different material patches of the target object. More examples are shown in the supplementary material.", + "bbox": [ + 511, + 611, + 908, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the robustness of EventPSR in recovering various materials, we evaluate the performance of the three methods on a cow object with 6 different surface materials, from ideal Lambertion to pure mirror. More details", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7e0419f150cfb5fd7b07a6adc0606d410dd810a1e93885ca94b06817d25d4d2.jpg", + "image_caption": [ + "Figure 8. Qualitative demonstration of real-world objects. From left to right, the complexities of shape and material are increasing. EventPSR recovers parameter maps consistently, and preserves high-frequency details for simple and complex objects." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 472, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are shown in the supplementary material.", + "bbox": [ + 89, + 606, + 364, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Evaluation on Real Data", + "text_level": 1, + "bbox": [ + 89, + 628, + 316, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative evaluation on surface normal estimation. To demonstrate that EventPSR can estimate the surface normal of objects with various materials, we select some objects with identical shape but extremely different surface materials. As shown in the top part of Fig. 6, the surface normal estimation results across different materials remain consistent, with only minor errors. The bottom part of Fig. 6 are the relighting results rendered with the estimated normal and reflectance parameters under different lighting conditions. The results demonstrate that EventPSR effectively handles material variation and maintains accuracy across different surface types.", + "bbox": [ + 89, + 651, + 482, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative evaluation on reflectance estimation. To verify the accuracy of roughness estimation using EventPSR, we conduct experiments with two semi-spheres", + "bbox": [ + 89, + 833, + 482, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "featuring distinct roughness levels, as illustrated in Fig. 7. EventPSR achieves an average MSE of 0.0067 in roughness estimation, showing its high precision and robustness among different surface materials. The error maps highlight EventPSR's capability to differentiate subtle variations in material roughness, underscoring its robustness across diverse reflectance properties.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative evaluation on real objects. The results of surface normal and reflectance parameters estimation of three real-world complex objects are shown in Fig. 8. The nose model in the left column has spatially uniform specular material. The estimated surface normal is mostly unaffected by specular reflection. The estimated roughness and metallic are mostly uniform on the surface. In the middle column, there is a diffuse cat lying on a black and metallic (mirror-like) plate. From the estimated result, we can see the low roughness and high metallic on the outer rim. There is a splatoon model with complex materials in the right column. From the image captured by a frame-based camera, we can see that different parts are in different materials. The eyes, hair, and backpack have low roughness, while the rough skin has high roughness. The recovered roughness is consistent with the observation. Besides, the high metallic in the pipes of the ink gun is also recovered by EventPSR.", + "bbox": [ + 511, + 199, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 474, + 633, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose EventPSR, an efficient and robust approach for simultaneous estimation of surface normal and reflectance parameters using an event camera. EventPSR is the first to use continuous radiance differential signals for joint estimation of shape and material properties, reducing data rate by $30\\%$ compared to frame-based methods.", + "bbox": [ + 511, + 503, + 903, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The capturing speed is limited by the refresh rate of the current monitors. The accuracy of real object results is affected by cast shadow and inter-reflection. Despite the robustness of albedo variation, we cannot recover diffuse albedo. These factors can be improved by employing a faster device and designing more robust algorithms, which are described in the supplementary material.", + "bbox": [ + 511, + 595, + 905, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 513, + 722, + 671, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 624B2006, 62088102, Beijing Natural Science Foundation under Grant No. L233024, JST-Mirai Program under Grant No. JPMJM123G1, and JSPS Kakenhi under Grant No. 20H05953. We thank NHK (Japan Broadcasting Corporation) Science & Technology Research Laboratories (STRL) for supporting the illumination devices for real-world experiments. PKU-affiliated authors thank openbayes.com for providing computing resource.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "${}^{5}$ Objects are rendered with a uniform albedo of white.", + "bbox": [ + 91, + 886, + 377, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11434", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Neil Alldrin, Todd Zickler, and David Kriegman. Photometric stereo with non-parametric and spatially-varying reflectance. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2008. 1, 2", + "[2] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2", + "[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of Int. Conf. Comput. Vis., 2021. 1", + "[4] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, pages 1-7. vol. 2012, 2012. 3, 1", + "[5] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K Wong. Self-calibrating deep photometric stereo networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3", + "[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2", + "[7] Hoon-Gyu Chung, Seokjun Choi, and Seung-Hwan Baek. Differentiable point-based inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2", + "[8] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Trans. Graph., 37(6):1-14, 2018. 2", + "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2, 3", + "[10] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M. Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2010. 1", + "[11] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[12] Satoshi Ikehata. Universal photometric stereo network using global lighting contexts. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6", + "[13] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7, 2", + "[14] Yakun Ju, Kin-Man Lam, Wuyuan Xie, Huiyu Zhou, Junyu Dong, and Boxin Shi. Deep learning methods for calibrated photometric stereo and beyond. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3", + "[15] Kaizhang Kang, Cihui Xie, Chengan He, Mingqi Yi, Minyi Gu, Zimin Chen, Kun Zhou, and Hongzhi Wu. Learning efficient illumination multiplexing for joint capture of reflectance and shape. ACM Trans. Graph., 2019. 1, 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] M. Katzin. The scattering of electromagnetic waves from rough surfaces. Proceedings of the IEEE, 52(11):1389-1390, 1964. 1", + "[17] Behnaz Kavoosighafi, Saghi Hajisharif, Ehsan Miandi, Gabriel Baravdish, Wen Cao, and Jonas Unger. Deep SVBRDF acquisition and modelling: A survey. In Computer Graphics Forum, 2024. 2", + "[18] Zongrui Li, Qian Zheng, Boxin Shi, Gang Pan, and Xudong Jiang. DANI-Net: Uncalibrated photometric stereo by differentiable shadow handling, anisotropic reflectance modeling, and neural inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[19] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2", + "[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Proc. of Adv. Neural Inform. Process. Syst., 2019. 6", + "[21] Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Efstratios Gavves, and Tinne Tuytelaars. Deep reflectance maps. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2", + "[22] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. Diligent102: A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6", + "[23] Fabiano Romeiro, Yuriy Vasilyev, and Todd Zickler. Passive reflectometry. In Proc. of Eur. Conf. Comput. Vis., 2008. 2", + "[24] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep photometric stereo networks for determining surface normal and reflectances. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2", + "[25] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2", + "[26] K. E. Torrance and E. M. Sparrow. Theory for off-specular reflection from roughened surfaces*. J. Opt. Soc. Am., 57(9): 1105-1114, 1967. 1", + "[27] Olivia Wiles and Andrew Zisserman. SilNet: Single-and multi-view reconstruction by learning from silhouettes. In Proc. of Brit. Mach. Vis. Conf., 2017. 6", + "[28] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of Eur. Conf. Comput. Vis., 2022. 1", + "[29] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-time photometric stereo using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3", + "[30] Ye Yu and William AP Smith. InverseRenderNet: Learning single image inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11435", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In Proc. of Int. Conf. Comput. Vis., 2023. 1, 6, 7, 2", + "[32] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 1", + "[33] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided direct and global components separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3" + ], + "bbox": [ + 91, + 92, + 480, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11436", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_model.json b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..883e91a1f308aa9e8a3ba9f1b1c1e8fd40316e61 --- /dev/null +++ b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.806, + 0.177 + ], + "angle": 0, + "content": "EventPSR: Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.203, + 0.353, + 0.22 + ], + "angle": 0, + "content": "Bohan \\(\\mathrm{Yu}^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.204, + 0.468, + 0.22 + ], + "angle": 0, + "content": "Jin Han3,4" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.204, + 0.614, + 0.22 + ], + "angle": 0, + "content": "Boxin Shi\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.204, + 0.75, + 0.22 + ], + "angle": 0, + "content": "Imari Sato3,4" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.224, + 0.829, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.241, + 0.827, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "list", + "bbox": [ + 0.166, + 0.224, + 0.829, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.258, + 0.694, + 0.274 + ], + "angle": 0, + "content": "\\(^{3}\\) The University of Tokyo \\(^{4}\\) National Institute of Informatics" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.278, + 0.826, + 0.293 + ], + "angle": 0, + "content": "ybh1998@pku.edu.cn, jinan@nii.ac.jp, shiboxin@pku.edu.cn, imarik@nii.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.485, + 0.648 + ], + "angle": 0, + "content": "Simultaneously acquisition of the surface normal and reflectance parameters is a crucial but challenging technique in the field of computer vision and graphics. It requires capturing multiple high dynamic range (HDR) images in existing methods using frame-based cameras. In this paper, we propose EventPSR, the first work to recover surface normal and reflectance parameters (e.g., metallic and roughness) simultaneously using an event camera. Compared with the existing methods based on photometric stereo or neural radiance fields, EventPSR is a robust and efficient approach that works consistently with different materials. Thanks to the extremely high temporal resolution and high dynamic range coverage of event cameras, EventPSR can recover accurate surface normal and reflectance of objects with various materials in 10 seconds. Extensive experiments on both synthetic data and real objects show that compared with existing methods using more than 100 HDR images, EventPSR recovers comparable surface normal and reflectance parameters with only about \\(30\\%\\) of the data rate." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.677, + 0.222, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.84 + ], + "angle": 0, + "content": "The fast and accurate acquisition of the shape and appearance information of an object is essential for many tasks in computer vision and computer graphics. This technique enables re-rendering realistic digital twins of real-world objects, supporting a wide range of downstream applications, such as virtual reality, augmented reality (AR), and digital content creation. However, dynamic range is a big issue when the material is composed of diffuse and highly specular properties. The time-consuming capturing process and" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.332, + 0.895, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.767 + ], + "angle": 0, + "content": "Figure 1. The EventPSR applies a sequence of light scanning patterns and an event camera to recover the accurate normal map and reflectance parameters (i.e., metallic and roughness) simultaneously. A two-stage pipeline consisting of a grid-matching and a gradient-tuning stage is proposed for EventPSR, which supports shape and reflectance recovery for different materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.907, + 0.821 + ], + "angle": 0, + "content": "difficulty recovering highly specular surfaces are two major challenges in this area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Traditional shape and reflectance estimation methods [1, 3, 10, 15, 28, 31, 32] rely on multiple high dynamic range (HDR) images captured under varying lighting conditions or viewing angles. Different from photometric stereo (PS) that estimates the surface normal of diffuse dominant ob" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.85, + 0.279, + 0.863 + ], + "angle": 0, + "content": "*Corresponding author: Boxin Shi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.864, + 0.483, + 0.888 + ], + "angle": 0, + "content": "This work was conducted while the first author, Bohan Yu, was doing internship at National Institute of Informatics." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.889, + 0.476, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code available: https://codeberg.org/ybh1998/EventPSR" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.85, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11427" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "jects from several images, shape and appearance estimation methods [1, 15] require more images to ensure the accurate observation of various reflections, which is time-consuming especially for shiny materials recovery. It is difficult to fully cover the specular highlights in shiny and metallic (i.e., mirror-like) materials under directional light without complicated capturing setup [15] and multiple images to cover the dynamic range. Such a dilemma between speed and accuracy partly attributes to the insufficient capability of conventional frame-based cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Event cameras characterized by extremely high temporal resolution and dynamic range, have shown promising advantages in many computer vision applications [9]. Researchers have explored applying event cameras in various photometric tasks [6, 19, 29, 33]. Particularly, EventPS [29] has demonstrated the superiority of event cameras for real-time surface normal estimation, showing the potential to accelerate and improve accuracy for shape and reflectance estimation tasks. However, it is still unclear how to extract material-related parameters, including roughness and metallic, from the event stream only." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.41, + 0.483, + 0.667 + ], + "angle": 0, + "content": "In this paper, we propose EventPSR, a novel solution for simultaneous estimation of both surface normal and Reflectance properties from Photometric Stereo using an Event camera. As shown in Fig. 1, EventPSR can simultaneously recover an object's shape and reflectance properties\\(^{2}\\) with designed light scanning patterns and a two-stage algorithm, which supports the recovery of different materials. Directly applying the setting and algorithm from EventPS [29] (i.e., capturing objects with a fixed event camera under a fast-moving point light source) is a non-trivial task. It is impossible to estimate the shape and reflectance at the same time without conquering the two challenges: 1) Design a suitable light scanning pattern that covers various materials and fully reveals the shape and reflectance properties of the objects. 2) Propose a new algorithm that converts the event signals to accurate normal and reflectance parameters simultaneously under the light scanning pattern." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.864 + ], + "angle": 0, + "content": "To conquer the first challenge, we discover that there are three requirements that need to be satisfied for a suitable light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. Based on this, we design a pattern of \"1D strip light\" that covers the full sphere of the illumination direction space. It is designed not only for various materials ranging from ideal Lambertian to pure mirror, but also for efficient capture of event cameras with faster speed and low data rate. To conquer the second challenge, we propose a two-stage algorithm to estimate surface normal and reflectance parameters, consisting of a grid-matching stage and a gradient-tuning stage. We first render a database consisting of the intensity-changing curves, including the com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "binations of various surface normal directions and materials. In the grid-matching stage, we match the event stream triggered in each pixel to a specific curve in the database and get a group of coarse reflectance parameters. In the gradient-tuning stage, we finetune the estimated parameters to a more accurate sub-grid location. To summarize, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.245 + ], + "angle": 0, + "content": "- We propose EventPSR, the first solution to estimate surface normal and reflectance parameters simultaneously from continuous radiance differential event signals." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.245, + 0.905, + 0.289 + ], + "angle": 0, + "content": "- We design a suitable light scanning pattern and an efficient two-stage algorithm with only \\(30\\%\\) data rate compared with frame-based counterparts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.29, + 0.905, + 0.351 + ], + "angle": 0, + "content": "- We validate the effectiveness of EventPSR by a platform with multiple monitors. The results show that EventPSR can estimate accurate shape-related and material-related parameters of real objects." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.363, + 0.663, + 0.379 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.463 + ], + "angle": 0, + "content": "Our method is related to PS works, and comprehensive surveys can be found in [14, 25]. This section focuses on works addressing PS and surface reflectometry using conventional frame-based cameras, as well as recent advancements in photometry using event cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Surface reflectometry. Measuring the detailed surface reflectance requires the use of controlled lighting to sample photometric stereo images. Romeiro et al. [23] inferred the bi-directional surface reflectance of materials using a light probe and a single HDR image of the known surface normal. Dupuy and Jakob [8] measured the reflectance spectra using a gonio-photometer to simultaneously handle BRDF acquisition, storage, and rendering. Learning-based BRDF acquisition and modeling approaches are summarized in the survey paper [17]. EventPSR differs from this category of research because we simultaneously recover the normal and reflectance parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Normal and reflectance estimation from conventional cameras. Alldrin et al. [1] proposed to employ novel bivariate approximations of isotropic reflectance functions to simultaneously recover the shape and reflectance of a surface from photometric stereo images. Chung et al. [7] optimized the geometry and spatially-varying BRDFs through a point-based rendering. Learning-based methods applied convolutional neural networks on single image [2, 21, 30] or multiple images [15, 24] to estimate the surface normal and reflectance parameters. These methods are limited in capturing speed due to the low dynamic range and high latency of conventional cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Event camera-based photometry. Event cameras with high temporal resolution and high dynamic range properties have been applied to various photometry tasks in recent years. Muglikar et al. [19] put a linear polarizer rotating at high speeds in front of an event camera to reconstruct" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2In this work, we recover reflectance-related properties specifically with roughness and metallic." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11428" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "relative intensities at multiple polarizer angles, which were used for surface normals estimation. Han et al. [11] computed the frequencies of the event triggering during the transient process of global lighting changes for depth sensing and object iso-contour reconstruction. Zhou et al. [33] proposed to use the differential signals captured from an event camera to separate the direct and global illumination components in RGB videos by applying a sweeping occluder. EventPS [29] leveraged an event camera and a fast-rotating light source to achieve real-time surface normal reconstruction, which demonstrated the superiority over the conventional cameras in speed and data bandwidth. EventPSR can further estimate surface normal and reflectance parameters with high robustness and efficiency by designing light scanning patterns and estimation algorithms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.331, + 0.182, + 0.347 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.448 + ], + "angle": 0, + "content": "As shown in Fig. 1, the overall workflow of EventPSR includes the design of suitable lighting scanning patterns (Sec. 3.2) and a two-stage algorithm that recovers surface normal and reflectance parameters (Sec. 3.3). In Sec. 3.1, we start from the mathematical formulation of event cameras, parameterized reflectance models, and target function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.295, + 0.47 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.645 + ], + "angle": 0, + "content": "Event formation model. Event cameras measure the scene brightness changes asynchronously on a logarithmic scale. An event \\( e = \\{t, x, y, p\\} \\) is triggered when the changes of logarithmic brightness reach a triggering threshold \\( C \\), where \\( t \\) is the timestamp, \\( (x, y) \\) is the pixel coordinate, and \\( p \\in \\{-1, +1\\} \\) is the polarity which represents the decrease or increase of brightness. Assuming there are totally \\( K \\) events triggered at pixel \\( (x, y) \\) during a short period of time, these events are represented as \\( \\mathcal{E}_{x,y} = \\{x, y, p_k, t_k\\} \\), where \\( k = \\{1, 2, \\dots, K\\} \\). The irradiance value \\( I \\) in pixel \\( (x, y) \\) changing from \\( t_{k-1} \\) to \\( t_k \\) is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.654, + 0.483, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\log \\left(I _ {x, y} \\left(t _ {k}\\right) + \\epsilon\\right) = \\log \\left(I _ {x, y} \\left(t _ {k - 1} + \\eta\\right) + \\epsilon\\right) + p _ {k} C, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.71 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a small offset value to avoid taking the logarithm of zero, and \\(\\eta\\) is the refractory time of the pixel [9]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.483, + 0.787 + ], + "angle": 0, + "content": "Reflectance model and inverse problem. We use a simplified version of Disney Principled BRDF [4] model to represent the reflectance \\((\\mathbf{f}_{r,x,y})\\) of the surface. Given an opaque and non-emissive object with a distant light source, the irradiance value \\(I_{x,y}(t)\\) at timestamp \\(t\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.795, + 0.483, + 0.839 + ], + "angle": 0, + "content": "\\[\nI _ {x, y} (t) = \\int_ {\\mathbf {w} _ {i} \\in \\Omega} \\left(\\mathbf {N} _ {x, y} \\cdot \\mathbf {w} _ {\\mathrm {i}}\\right) \\mathbf {f} _ {r, x, y} \\left(\\mathbf {w} _ {\\mathrm {i}}, \\mathbf {w} _ {\\mathrm {r}}\\right) \\mathbf {L} \\left(\\mathbf {w} _ {i}, t\\right) \\mathrm {d} \\mathbf {w} _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Omega\\) is the whole sphere direction space, \\(\\mathbf{N}_{x,y}\\) is the surface normal, \\(\\mathbf{L}\\) is the light source function, \\(\\mathbf{w}_i\\) and \\(\\mathbf{w}_r\\) are the incident and reflection direction, respectively. The light source function \\(\\mathbf{L}(\\mathbf{w}_i,t)\\) represents a distant light with" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "the same brightness at an angle of \\(\\mathbf{w}_i\\). For each pixel \\((x,y)\\), its reflectance \\(\\mathbf{f}_{r,x,y}\\) is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.133, + 0.905, + 0.179 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {r, x, y} \\left(\\mathbf {w} _ {\\mathrm {i}}, \\mathbf {w} _ {\\mathrm {r}}\\right) = \\frac {1 - M _ {x , y}}{\\pi} + \\frac {M _ {x , y} D G}{\\pi \\left(\\mathbf {N} _ {x , y} \\cdot \\mathbf {w} _ {\\mathrm {i}}\\right) \\left(\\mathbf {N} _ {x , y} \\cdot \\mathbf {w} _ {\\mathrm {r}}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.18, + 0.907, + 0.271 + ], + "angle": 0, + "content": "where \\(D, G\\) are parameters that describe specular reflection property.\\(^3\\) The variation of diffuse albedo has been canceled by the log difference mechanism in event cameras [29], so we set it as 1 in our method. Other parameters and Fresnel term from Disney Principled BRDF [4] are removed for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.291, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Target function. To recover the surface properties, i.e., \\(\\mathbf{N}_{x,y}\\) (Surface normal), \\(R_{x,y}\\) (Roughness, in the supplementary material), and \\(M_{x,y}\\) (Metallic), we design a light scanning pattern \\(\\mathbf{L}(\\mathbf{w}_i,t)\\) and apply an event camera to solve the inverse reflectance estimation problem. We represent the shape and reflection estimation algorithm as \\(\\tilde{\\mathbf{N}} (\\mathcal{E})\\), \\(\\tilde{\\mathbf{R}} (\\mathcal{E})\\), and \\(\\tilde{\\mathbf{M}} (\\mathcal{E})\\). Our target is to design these algorithms and optimize light scanning patterns \\(\\mathbf{L}\\) to estimate the surface normal and reflectance parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.438, + 0.905, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathbf {L}, \\tilde {\\mathbf {N}}, \\tilde {\\mathbf {R}}, \\tilde {\\mathbf {M}}} {\\operatorname {a r g m i n}} \\sum_ {\\mathbf {N}, M, R} | \\tilde {\\mathbf {N}} (\\mathcal {E}) - \\mathbf {N} | _ {2} + \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.474, + 0.843, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\left| \\tilde {\\mathbf {M}} (\\mathcal {E}) - M \\right| _ {2} + \\left| \\tilde {\\mathbf {R}} (\\mathcal {E}) - R \\right| _ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.851, + 0.52 + ], + "angle": 0, + "content": "where the events \\(\\mathcal{E}\\) are triggered from \\(\\mathbf{L},\\mathbf{N},R,M\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.786, + 0.547 + ], + "angle": 0, + "content": "3.2. Light Scanning Pattern Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.675 + ], + "angle": 0, + "content": "It is necessary to design a suitable light scanning pattern to cover different materials and the shape of target objects. There are three requirements when designing the light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. In the following, we consider an effective use of a ring-shaped light source as the base scanning pattern. Taking a ring moving along the \\( Z \\) axis as an example, the formula of \\( \\mathbf{L}(\\mathbf{w}_i,t) \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.686, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} \\left(\\mathbf {w} _ {i}, t\\right) = \\max _ {u} \\max (0, \\mathbf {w} _ {i} \\cdot \\mathbf {R} (u, t)) ^ {\\alpha}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.712, + 0.866, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} (u, t) = [ \\sin (t) \\sin (u), \\sin (t) \\cos (u), \\cos (t) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.907, + 0.816 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}(u,t)\\) represent all light directions on a ring. \\(\\alpha\\) is the width of the ring. In order to illustrate that our \"moving ring light\" pattern has better properties to overcome all these challenges, we compare it with two other light source patterns: \"point light\" and \"structured environment map\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.907, + 0.877 + ], + "angle": 0, + "content": "For the \"point light\" pattern, we combine two spiral movement trajectories to cover the whole sphere: the horizontal spiral and vertical spiral. In this way, we obtain a continuous moving light source pattern. We also add a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.514, + 0.887, + 0.836, + 0.901 + ], + "angle": 0, + "content": "3Detailed formulas are shown in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11429" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.256, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. The design of light scanning pattern. (a) Illustration of specular coverage requirement. The first row shows a shiny ball under three lighting patterns. The second row shows where specular parameters can be recovered (the brighter the better). (b) Illustration of diffuse sensitivity requirement. The first row shows the rendered frames of a diffuse ball under three lighting patterns. The second row shows the profiles of intensity changes of two different points on the sphere over time. (c) Illustration of event efficiency requirement. The first row shows the event signals triggered under three lighting patterns. The second row plots the average (green) and max (red) number of events triggered over time when covering the whole specular reflection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.399 + ], + "angle": 0, + "content": "smoothness term \\(\\varepsilon\\) (the size of the point light), so the lighting will cover the whole sphere lighting direction gradually. For the horizontal spiral, the formula \\(\\mathbf{L}(\\mathbf{w}_i,t)\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.42, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {L} (\\mathbf {w} _ {i}, t) = (\\mathbf {R} (t) \\cdot \\mathbf {w} _ {i}) ^ {\\alpha}, \\\\ \\mathbf {R} (t) = [ \\begin{array}{l l l l l l} \\cdot & (\\partial_ {t}) & \\cdot & (\\partial_ {t}) & \\cdot & (\\partial_ {t}) \\end{array} ] \\end{array} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.44, + 0.442, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} (t) = [ \\sin (\\beta t) \\cos (t), \\cos (\\beta t) \\cos (t), \\sin (t) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.483, + 0.497 + ], + "angle": 0, + "content": "where \\(\\beta\\) determines the number of spirals. For the vertical spiral, the formula \\(\\mathbf{L}(\\mathbf{w}_i,t)\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.519, + 0.483, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {L} (\\mathbf {w} _ {i}, t) = (\\mathbf {R} (t) \\cdot \\mathbf {w} _ {i}) ^ {\\alpha}, \\\\ \\text {T r} (t) = [ (0) + (1) + (2) + (3) + (4) ] \\end{array} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.538, + 0.437, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} (t) = [ \\cos (\\beta t) \\sin (t), \\cos (\\beta t) \\cos (t), \\sin (\\beta t) ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.626 + ], + "angle": 0, + "content": "For the \"structured environment map\" pattern, we project structured light along three axes alternatively. Taking structured light along the \\(Z\\) axis as an example, the formula of this light source pattern \\(\\mathbf{L}(\\mathbf{w}_i,t)\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.646, + 0.345, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} (\\mathbf {w} _ {i}, t) = \\left\\lfloor \\mathrm {f m o d} (\\mathbf {R} (2 ^ {t} \\mathbf {w} _ {i}), 2) \\right\\rfloor ,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.664, + 0.483, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} \\left(\\mathbf {w} _ {i}\\right) = \\frac {\\operatorname {a t a n 2} \\left(\\mathbf {w} _ {i} \\cdot [ 0 , 1 , 0 ] , \\mathbf {w} _ {i} \\cdot [ 1 , 0 , 0 ]\\right)}{\\pi} + 1, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.734 + ], + "angle": 0, + "content": "where \"fmod\" is the modulo operator. \\(t\\) should be integer and the \"structured environment map\" pattern is discrete." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specular coverage requirement. Given metallic (i.e., without diffuse part) or even mirror-like objects, it is difficult to use single distant point light for surface normal estimation. Because the specular surface only responds to lighting from a small lobe. For the specular dielectric material (i.e., with a diffuse part), the surface normal can be estimated from the diffuse reflection. However, the material properties are still only recoverable in the small specular highlight part. We show rendered images of a shiny ball under all three lighting setups in the top row of Fig. 2 (a). The bottom row shows that the solvable regions for \"point" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.457 + ], + "angle": 0, + "content": "light” is a small circle around the lighting direction. In contrast, the “moving ring light” and “structured environment map” have larger solvable regions with a strip around the ring or structure edges, which is more efficient for specular coverage. The larger specular coverage results in faster data-capturing speed for challenging specular surface materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Diffuse sensitivity requirement. Diffuse reflection has the opposite property from specular reflection: it spreads the input light to all the upper hemisphere directions instead of only a small highlight region. In order to estimate surface normal from the reflection, the light source pattern needs to contain high and diverse low-frequency parts to make it distinguishable from different surface normal. As shown in Fig. 2 (b), we render the images of a rough diffusive ball under all three lighting setups in the top row, and plot the reflection intensity of two pixels in the bottom row. The diffuse reflection of two pixels is shifted in time for \"point light\" and \"moving ring light\" patterns. Since event cameras have an extremely high temporal resolution, the time shift can be well-captured by the camera and converted to surface normal accurately. In contrast, the diffuse reflection intensity profiles for the two pixels are similar for \"structured environment map\", making it difficult to distinguish the surface normal of two points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Event efficiency requirement. Since the event camera measures the brightness changes, the bandwidth consumption is mainly determined by the brightness changing magnitude in the scene. We show the synthetic event signals and the accumulated number of events from rendered a mirror-like ball in Fig. 2 (c). Both \"point light\" and \"moving ring light\" have a similar number of events triggered by different surface normal (observed by two pixels). The \"moving ring light\" pattern consumes fewer events for specular coverage than the \"point light\", indicating a better bandwidth efficiency. In contrast, the \"structured environment map\"" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11430" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "leads to a variable number of events across different surface normal. In the worst case, where there is a mirror plane facing at the worst surface normal with a maximum number of events generated, the whole scene is flickering, and the read-out will be congested." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.228 + ], + "angle": 0, + "content": "Based on the comparisons and analysis above, the designed \"moving ring light\" achieves relatively larger coverage of various materials while preserving higher efficiency of event triggering." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.236, + 0.291, + 0.252 + ], + "angle": 0, + "content": "3.3.EventPSR Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.318 + ], + "angle": 0, + "content": "As shown in Fig. 1, we propose a two-stage solving pipeline, consisting of a grid-matching stage and a gradient-tuning stage. The proposed algorithm provides a global minimum result with high value resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.5 + ], + "angle": 0, + "content": "Stage-1: Grid-matching. To estimate the surface normal and reflectance parameters from the captured event streams, we first render a database containing various intensity-time curves using a physically-based rendering engine. We assume that the rendered temporal resolution is high enough so that the intensity curves can be represented by linear interpolation of a series of frames. There are 4 parameters to be estimated for each pixel, including 2 parameters for surface normal, 1 parameter for roughness, and 1 parameter for metallic. The database is on a 5-dimension grid (1 additional dimension for time), which is defined as \\(\\mathcal{B} = \\{\\mathbf{N}, R, M, t\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Given a series of event signals \\(\\mathcal{E}_{x,y}\\) at a pixel \\((x,y)\\), the goal in this stage is to find the most matching curve in the database. Therefore, we define the loss function between the observed event stream and curve candidates as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.567, + 0.483, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {1} (x, y, \\mathbf {N}, R, M) = \\sum_ {k} | \\log (\\mathcal {B} (\\mathbf {N}, R, M, t _ {k}) + \\epsilon) \\tag {9} \\\\ - \\log (\\mathcal {B} (\\mathbf {N}, R, M, t _ {k - 1} + \\eta) + \\epsilon) - p _ {k} C | ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.719 + ], + "angle": 0, + "content": "By designing the loss function on each pair of events, we prevent the event integral process, which has an integration drifting issue (More details about the drifting issue are shown in the supplementary material). Furthermore, our loss function averages errors of all events in the same pixel, implicitly suppressing event triggering noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.765 + ], + "angle": 0, + "content": "We enumerate 4D surface normal \\(\\mathbf{N}\\) and material properties \\(R, M\\) on a discrete grid and select the one with minimum loss. The results \\(\\tilde{\\mathbf{N}}, \\tilde{R}, \\tilde{M}\\) at this stage are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.774, + 0.483, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {N}} _ {x, y}, \\tilde {R} _ {x, y}, \\tilde {M} _ {x, y} = \\underset {\\mathbf {N}, R, M} {\\operatorname {a r g m i n}} \\mathcal {L} _ {1} (x, y, \\mathbf {N}, R, M). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Stage-2: Gradient-tuning. The grid-matching stage provides us with a coarse result close to the global minimum, but this result can be further improved for the following two reasons: 1) The resolution of grid-matching stage depends on the searching grid size. Due to the computational cost, this resolution is not high enough. 2) In real setup, there" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.681, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.222, + 0.674, + 0.235 + ], + "angle": 0, + "content": "Data acquisition platform" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.091, + 0.794, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.222, + 0.866, + 0.236 + ], + "angle": 0, + "content": "Light scanning patterns" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.091, + 0.903, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.245, + 0.905, + 0.329 + ], + "angle": 0, + "content": "Figure 3. Data acquisition platform for capturing real data. There are 5 commodity monitors surrounding the target object to illuminate the designed pattern from any incidence angle. We show the capture device setup on the left side. The target object is hung in the middle and an event camera is pointing at the target from the bottom. Three projected patterns are shown on the right." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.413 + ], + "angle": 0, + "content": "are some regions of the environment map that do not have a display. If the material is shiny, metallic, and the surface normal is reflecting to the dark region, there is no brightness change, which leads to ambiguity in these regions. More details are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.906, + 0.517 + ], + "angle": 0, + "content": "To solve the first problem, we calculate the gradient of surface normal \\(\\mathbf{N}\\) and material properties \\(R, M\\), and interpolate the database on these parameters. In this way, we obtain an approximated rendered intensity and its gradient in a continuous parameter space. Therefore, we minimize the \\(\\mathcal{L}_1(x,y,\\mathbf{N},R,M)\\) using gradient descent to get the optimized result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.905, + 0.55 + ], + "angle": 0, + "content": "To solve the second problem, we introduce a small inter-pixel smoothness regularization term:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.559, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {2} = \\beta \\sum_ {x, y} \\left(\\left| \\mathbf {N} _ {x, y} - \\mathbf {N} _ {x - 1, y} \\right| ^ {2} + \\left| \\mathbf {N} _ {x, y} - \\mathbf {N} _ {x, y - 1} \\right| ^ {2} + \\right. \\\\ \\left| R _ {x, y} - R _ {x - 1, y} \\right| ^ {2} + \\left| R _ {x, y} - R _ {x, y - 1} \\right| ^ {2} + \\tag {11} \\\\ \\left. \\left| M _ {x, y} - M _ {x - 1, y} \\right| ^ {2} + \\left| M _ {x, y} - M _ {x, y - 1} \\right| ^ {2}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.718 + ], + "angle": 0, + "content": "In this way, for shiny and metallic surface, the parameters in ambiguous regions are filled with interpolated values, while the diffusive or dielectric surfaces are not affected. The optimized parameters of this gradient-tuning are the final results of EventPSR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.734, + 0.646, + 0.75 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.731, + 0.773 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Data acquisition platform. To illuminate the target object using our designed light scanning pattern, we build a data acquisition platform with multiple commodity monitors, as illustrated in Fig. 3. The target object is hanging at the center of the device and surrounded by several monitors. There is a Prophesee EVK4 event camera beneath the target object, pointing above at the object and capturing the event stream as the light scanning pattern changes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11431" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.091, + 0.905, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.237, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Figure 4. Quantitative comparison on synthetic data. We show the relative data consumption and per-object averaged error for normal, roughness, and metallic estimation. The color heatmap for each parameter (normal, roughness, and metallic) is normalized independently. The results from NeILF [31] are empty for some objects, which are marked as \"Failed\"." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.288, + 0.485, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Figure 5. Quantitative comparison results on synthetic data. We show the estimated normal, roughness, metallic, and error maps of EventPSR and two comparing methods. The averaged error maps are labeled in red numbers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.794 + ], + "angle": 0, + "content": "The light scanning pattern received by the target object is slightly different from the designed one because of some non-ideal factors. The uncovered area at the top and bottom will cause some missing illumination, and the interreflection of the monitors will cause some extra light from the wrong direction. To alleviate the influences caused by non-ideal light scanning pattern projection, we capture the received illumination by a mirror ball in the scene. The database \\(\\mathcal{B}(\\mathbf{N}, R, M, t)\\) is rendered using the calibrated environment maps. In addition, we flash all panels to full white at the beginning and end of the illumination pattern for temporal alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Algorithm implementation. Both of the two stages are implemented using PyTorch [20]. For the grid-matching stage, the resolution of rendered database \\(\\mathcal{B}(\\mathbf{N}, R, M, t)\\) are as follows: \\(64 \\times 64\\) for normal resolution, 12 for roughness resolution, 8 for metallic resolution, and 600 for temporal resolution. According to Eq. (3), the reflectance model is a linear interpolation from pure diffuse to pure mirror. To" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.38 + ], + "angle": 0, + "content": "reduce GPU VRAM consumption, we replace the metallic dimension using 1 pure diffuse component and 1 pure mirror component. Different material metallic properties are constructed on the GPU on the fly. Both of the grid-matching and gradient-tuning stages are conducted on a single NVIDIA GeForce RTX 3060 graphics card." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.61, + 0.403 + ], + "angle": 0, + "content": "4.2. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Synthetic data. To evaluate the accuracy of our method compared with other methods, we render 16 objects using the geometries from the sculpture dataset [27]. The material parameters including albedo, roughness, and metallic, are set randomly per patch. We use a ray-tracing-based rendering engine, with 1024 samples per pixel and 3 maximum bounces. The rendered images are of \\(512 \\times 512\\) resolution. An example is shown in the first column in Fig. 5. Each of the objects is rendered under 3 different setups: Our illumination with fixed camera position, DiLiGenT10² [22] light directions with fixed camera position (for the comparison with SDM-UniPS [12]), and fixed light direction with different camera positions (for the comparison with NeILF [31]). For the first setup, the event streams are simulated from high-frame-rate image sequences. The event triggering threshold noise is set to 0.01. For the other two setups, there are total 100 HDR images rendered per object. Besides, the independent Gaussian noise of 0.01 standard variance is added to each image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.907, + 0.774 + ], + "angle": 0, + "content": "Real data. We capture three types of real data using our data acquisition platform. The captured data are used for material roughness accuracy evaluation, surface normal accuracy evaluation, and complex object generalization evaluation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.775, + 0.909, + 0.849 + ], + "angle": 0, + "content": "- Surface normal evaluation data: We select several pairs of objects with almost identical geometry but extremely different materials (i.e., rough and shiny), as shown in the first row in Fig. 6. The ground truth of surface normal is captured by a structured light 3D scanner.\\(^4\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.85, + 0.905, + 0.865 + ], + "angle": 0, + "content": "- Material roughness evaluation data: We create several" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.775, + 0.909, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "4https : / / www . shining3d . com / professional - solutions/Desktop- 3d-scanner/einscan-se-v2" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11432" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.482, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.727 + ], + "angle": 0, + "content": "Figure 6. Top: Quantitative evaluation for surface normal estimation on real objects. The error value of each object is labeled in red number. Bottom: Relighting results based on the estimated surface parameters. The environment maps are labeled on the left. The albedo of rendered objects are set to uniform white." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.735, + 0.483, + 0.825 + ], + "angle": 0, + "content": "balls with varying surface roughness. A polished metal ball serves as the low roughness reference. For the high roughness surface, we apply a spray coating to achieve a diffuse surface. The ground truth values of roughness are measured using a frame-based camera with HDR exposure under ideal parallel light." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Generalization evaluation data: We select multiple objects whose surface are composed of multiple different materials, as shown in the first row in Fig. 8. The complexity of the shape and materials will test the generalization ability of EventPSR on real-world objects." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.261, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 7. Quantitative demonstration for robustness of material estimation on real objects. The first row is a metal ball with mirror-like surface material. The second row is a ball sprayed with a non-glossy coating." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.329, + 0.776, + 0.346 + ], + "angle": 0, + "content": "4.3. Evaluation on Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.61 + ], + "angle": 0, + "content": "We conduct quantitative evaluations of the surface normal and reflectance parameters estimation with the synthetic dataset. The comparison includes two frame-based methods [13, 31] that represent the state-of-the-art NeRF-based [31] and PS-based [13] normal and reflectance estimation approaches. For data rate usage comparison, we estimate the date consumption in the following way: For EventPSR, we assume that the input event streams employ 16-bit Prophesee EVT 3.0 format. For the other two comparison methods, we assume that the HDR input are captured by merging 3 multi-bracketing 8-bit grayscale images. We use mean angular error to measure the accuracy of recovered normal maps. For the recovered roughness and metallic, we first standardize the data to avoid representing differences between models, then we apply mean squared error (MSE) to measure the accuracy. More details are described in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.839 + ], + "angle": 0, + "content": "The data consumption and parameters error of each object are shown in Fig. 4. On average, our method consumes about \\(30\\%\\) of the data rate, achieving comparable accuracy of normal estimation and superior accuracy of roughness and metallic estimation. We visualize the result of one example from the synthetic dataset in Fig. 5. We can see that the error of our method concentrates mainly on the high geometry change area for all three parameter maps, where the inter-reflection and cast shadow are most severe. For the other two comparing methods [13, 31], the errors of normal estimation are more evenly distributed. However, the estimated roughness and metallic from the comparing methods [13, 31] can hardly reflect the high frequency of different material patches of the target object. More examples are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the robustness of EventPSR in recovering various materials, we evaluate the performance of the three methods on a cow object with 6 different surface materials, from ideal Lambertion to pure mirror. More details" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11433" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.473, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Figure 8. Qualitative demonstration of real-world objects. From left to right, the complexities of shape and material are increasing. EventPSR recovers parameter maps consistently, and preserves high-frequency details for simple and complex objects." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.607, + 0.365, + 0.621 + ], + "angle": 0, + "content": "are shown in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.63, + 0.317, + 0.645 + ], + "angle": 0, + "content": "4.4. Evaluation on Real Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Quantitative evaluation on surface normal estimation. To demonstrate that EventPSR can estimate the surface normal of objects with various materials, we select some objects with identical shape but extremely different surface materials. As shown in the top part of Fig. 6, the surface normal estimation results across different materials remain consistent, with only minor errors. The bottom part of Fig. 6 are the relighting results rendered with the estimated normal and reflectance parameters under different lighting conditions. The results demonstrate that EventPSR effectively handles material variation and maintains accuracy across different surface types." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.834, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Quantitative evaluation on reflectance estimation. To verify the accuracy of roughness estimation using EventPSR, we conduct experiments with two semi-spheres" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "featuring distinct roughness levels, as illustrated in Fig. 7. EventPSR achieves an average MSE of 0.0067 in roughness estimation, showing its high precision and robustness among different surface materials. The error maps highlight EventPSR's capability to differentiate subtle variations in material roughness, underscoring its robustness across diverse reflectance properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Qualitative evaluation on real objects. The results of surface normal and reflectance parameters estimation of three real-world complex objects are shown in Fig. 8. The nose model in the left column has spatially uniform specular material. The estimated surface normal is mostly unaffected by specular reflection. The estimated roughness and metallic are mostly uniform on the surface. In the middle column, there is a diffuse cat lying on a black and metallic (mirror-like) plate. From the estimated result, we can see the low roughness and high metallic on the outer rim. There is a splatoon model with complex materials in the right column. From the image captured by a frame-based camera, we can see that different parts are in different materials. The eyes, hair, and backpack have low roughness, while the rough skin has high roughness. The recovered roughness is consistent with the observation. Besides, the high metallic in the pipes of the ink gun is also recovered by EventPSR." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.476, + 0.634, + 0.491 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.905, + 0.594 + ], + "angle": 0, + "content": "We propose EventPSR, an efficient and robust approach for simultaneous estimation of surface normal and reflectance parameters using an event camera. EventPSR is the first to use continuous radiance differential signals for joint estimation of shape and material properties, reducing data rate by \\(30\\%\\) compared to frame-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.703 + ], + "angle": 0, + "content": "Limitations. The capturing speed is limited by the refresh rate of the current monitors. The accuracy of real object results is affected by cast shadow and inter-reflection. Despite the robustness of albedo variation, we cannot recover diffuse albedo. These factors can be improved by employing a faster device and designing more robust algorithms, which are described in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.723, + 0.673, + 0.74 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 624B2006, 62088102, Beijing Natural Science Foundation under Grant No. L233024, JST-Mirai Program under Grant No. JPMJM123G1, and JSPS Kakenhi under Grant No. 20H05953. We thank NHK (Japan Broadcasting Corporation) Science & Technology Research Laboratories (STRL) for supporting the illumination devices for real-world experiments. PKU-affiliated authors thank openbayes.com for providing computing resource." + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.887, + 0.378, + 0.9 + ], + "angle": 0, + "content": "\\( {}^{5} \\) Objects are rendered with a uniform albedo of white." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11434" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Neil Alldrin, Todd Zickler, and David Kriegman. Photometric stereo with non-parametric and spatially-varying reflectance. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of Int. Conf. Comput. Vis., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[4] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, pages 1-7. vol. 2012, 2012. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[5] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K Wong. Self-calibrating deep photometric stereo networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.429 + ], + "angle": 0, + "content": "[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[7] Hoon-Gyu Chung, Seokjun Choi, and Seung-Hwan Baek. Differentiable point-based inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[8] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Trans. Graph., 37(6):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[10] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M. Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[11] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[12] Satoshi Ikehata. Universal photometric stereo network using global lighting contexts. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[13] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[14] Yakun Ju, Kin-Man Lam, Wuyuan Xie, Huiyu Zhou, Junyu Dong, and Boxin Shi. Deep learning methods for calibrated photometric stereo and beyond. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Kaizhang Kang, Cihui Xie, Chengan He, Mingqi Yi, Minyi Gu, Zimin Chen, Kun Zhou, and Hongzhi Wu. Learning efficient illumination multiplexing for joint capture of reflectance and shape. ACM Trans. Graph., 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[16] M. Katzin. The scattering of electromagnetic waves from rough surfaces. Proceedings of the IEEE, 52(11):1389-1390, 1964. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[17] Behnaz Kavoosighafi, Saghi Hajisharif, Ehsan Miandi, Gabriel Baravdish, Wen Cao, and Jonas Unger. Deep SVBRDF acquisition and modelling: A survey. In Computer Graphics Forum, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[18] Zongrui Li, Qian Zheng, Boxin Shi, Gang Pan, and Xudong Jiang. DANI-Net: Uncalibrated photometric stereo by differentiable shadow handling, anisotropic reflectance modeling, and neural inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[19] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Proc. of Adv. Neural Inform. Process. Syst., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[21] Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Efstratios Gavves, and Tinne Tuytelaars. Deep reflectance maps. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[22] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. Diligent102: A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[23] Fabiano Romeiro, Yuriy Vasilyev, and Todd Zickler. Passive reflectometry. In Proc. of Eur. Conf. Comput. Vis., 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[24] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep photometric stereo networks for determining surface normal and reflectances. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[25] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[26] K. E. Torrance and E. M. Sparrow. Theory for off-specular reflection from roughened surfaces*. J. Opt. Soc. Am., 57(9): 1105-1114, 1967. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[27] Olivia Wiles and Andrew Zisserman. SilNet: Single-and multi-view reconstruction by learning from silhouettes. In Proc. of Brit. Mach. Vis. Conf., 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[28] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of Eur. Conf. Comput. Vis., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[29] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-time photometric stereo using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Ye Yu and William AP Smith. InverseRenderNet: Learning single image inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11435" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[31] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In Proc. of Int. Conf. Comput. Vis., 2023. 1, 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[32] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[33] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided direct and global components separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11436" + } + ] +] \ No newline at end of file diff --git a/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_origin.pdf b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a39476d5e7d2464e0e1fc316330f9abc66dae32 --- /dev/null +++ b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/133da766-89f1-4c63-bf52-44b0090134e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c64a1a26f2f2060faca7903328109e08f591277a955da24e6d62a22073d03685 +size 6374704 diff --git a/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/full.md b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/full.md new file mode 100644 index 0000000000000000000000000000000000000000..835fc6285b7744f98c72f643c28401a2c35f3350 --- /dev/null +++ b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/full.md @@ -0,0 +1,313 @@ +# EventPSR: Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera + +Bohan $\mathrm{Yu}^{1,2\dagger}$ + +Jin Han3,4 + +Boxin Shi $^{1,2*}$ + +Imari Sato3,4 + +$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University +$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +$^{3}$ The University of Tokyo $^{4}$ National Institute of Informatics + +ybh1998@pku.edu.cn, jinan@nii.ac.jp, shiboxin@pku.edu.cn, imarik@nii.ac.jp + +# Abstract + +Simultaneously acquisition of the surface normal and reflectance parameters is a crucial but challenging technique in the field of computer vision and graphics. It requires capturing multiple high dynamic range (HDR) images in existing methods using frame-based cameras. In this paper, we propose EventPSR, the first work to recover surface normal and reflectance parameters (e.g., metallic and roughness) simultaneously using an event camera. Compared with the existing methods based on photometric stereo or neural radiance fields, EventPSR is a robust and efficient approach that works consistently with different materials. Thanks to the extremely high temporal resolution and high dynamic range coverage of event cameras, EventPSR can recover accurate surface normal and reflectance of objects with various materials in 10 seconds. Extensive experiments on both synthetic data and real objects show that compared with existing methods using more than 100 HDR images, EventPSR recovers comparable surface normal and reflectance parameters with only about $30\%$ of the data rate. + +# 1. Introduction + +The fast and accurate acquisition of the shape and appearance information of an object is essential for many tasks in computer vision and computer graphics. This technique enables re-rendering realistic digital twins of real-world objects, supporting a wide range of downstream applications, such as virtual reality, augmented reality (AR), and digital content creation. However, dynamic range is a big issue when the material is composed of diffuse and highly specular properties. The time-consuming capturing process and + +![](images/eeef97a8dd3bf674f4849eff72787baa4e0d04db33c5ece325b3c5f38623a76e.jpg) +Figure 1. The EventPSR applies a sequence of light scanning patterns and an event camera to recover the accurate normal map and reflectance parameters (i.e., metallic and roughness) simultaneously. A two-stage pipeline consisting of a grid-matching and a gradient-tuning stage is proposed for EventPSR, which supports shape and reflectance recovery for different materials. + +difficulty recovering highly specular surfaces are two major challenges in this area. + +Traditional shape and reflectance estimation methods [1, 3, 10, 15, 28, 31, 32] rely on multiple high dynamic range (HDR) images captured under varying lighting conditions or viewing angles. Different from photometric stereo (PS) that estimates the surface normal of diffuse dominant ob + +jects from several images, shape and appearance estimation methods [1, 15] require more images to ensure the accurate observation of various reflections, which is time-consuming especially for shiny materials recovery. It is difficult to fully cover the specular highlights in shiny and metallic (i.e., mirror-like) materials under directional light without complicated capturing setup [15] and multiple images to cover the dynamic range. Such a dilemma between speed and accuracy partly attributes to the insufficient capability of conventional frame-based cameras. + +Event cameras characterized by extremely high temporal resolution and dynamic range, have shown promising advantages in many computer vision applications [9]. Researchers have explored applying event cameras in various photometric tasks [6, 19, 29, 33]. Particularly, EventPS [29] has demonstrated the superiority of event cameras for real-time surface normal estimation, showing the potential to accelerate and improve accuracy for shape and reflectance estimation tasks. However, it is still unclear how to extract material-related parameters, including roughness and metallic, from the event stream only. + +In this paper, we propose EventPSR, a novel solution for simultaneous estimation of both surface normal and Reflectance properties from Photometric Stereo using an Event camera. As shown in Fig. 1, EventPSR can simultaneously recover an object's shape and reflectance properties $^{2}$ with designed light scanning patterns and a two-stage algorithm, which supports the recovery of different materials. Directly applying the setting and algorithm from EventPS [29] (i.e., capturing objects with a fixed event camera under a fast-moving point light source) is a non-trivial task. It is impossible to estimate the shape and reflectance at the same time without conquering the two challenges: 1) Design a suitable light scanning pattern that covers various materials and fully reveals the shape and reflectance properties of the objects. 2) Propose a new algorithm that converts the event signals to accurate normal and reflectance parameters simultaneously under the light scanning pattern. + +To conquer the first challenge, we discover that there are three requirements that need to be satisfied for a suitable light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. Based on this, we design a pattern of "1D strip light" that covers the full sphere of the illumination direction space. It is designed not only for various materials ranging from ideal Lambertian to pure mirror, but also for efficient capture of event cameras with faster speed and low data rate. To conquer the second challenge, we propose a two-stage algorithm to estimate surface normal and reflectance parameters, consisting of a grid-matching stage and a gradient-tuning stage. We first render a database consisting of the intensity-changing curves, including the com + +binations of various surface normal directions and materials. In the grid-matching stage, we match the event stream triggered in each pixel to a specific curve in the database and get a group of coarse reflectance parameters. In the gradient-tuning stage, we finetune the estimated parameters to a more accurate sub-grid location. To summarize, the contributions of this paper are as follows: + +- We propose EventPSR, the first solution to estimate surface normal and reflectance parameters simultaneously from continuous radiance differential event signals. +- We design a suitable light scanning pattern and an efficient two-stage algorithm with only $30\%$ data rate compared with frame-based counterparts. +- We validate the effectiveness of EventPSR by a platform with multiple monitors. The results show that EventPSR can estimate accurate shape-related and material-related parameters of real objects. + +# 2. Related Works + +Our method is related to PS works, and comprehensive surveys can be found in [14, 25]. This section focuses on works addressing PS and surface reflectometry using conventional frame-based cameras, as well as recent advancements in photometry using event cameras. + +Surface reflectometry. Measuring the detailed surface reflectance requires the use of controlled lighting to sample photometric stereo images. Romeiro et al. [23] inferred the bi-directional surface reflectance of materials using a light probe and a single HDR image of the known surface normal. Dupuy and Jakob [8] measured the reflectance spectra using a gonio-photometer to simultaneously handle BRDF acquisition, storage, and rendering. Learning-based BRDF acquisition and modeling approaches are summarized in the survey paper [17]. EventPSR differs from this category of research because we simultaneously recover the normal and reflectance parameters. + +Normal and reflectance estimation from conventional cameras. Alldrin et al. [1] proposed to employ novel bivariate approximations of isotropic reflectance functions to simultaneously recover the shape and reflectance of a surface from photometric stereo images. Chung et al. [7] optimized the geometry and spatially-varying BRDFs through a point-based rendering. Learning-based methods applied convolutional neural networks on single image [2, 21, 30] or multiple images [15, 24] to estimate the surface normal and reflectance parameters. These methods are limited in capturing speed due to the low dynamic range and high latency of conventional cameras. + +Event camera-based photometry. Event cameras with high temporal resolution and high dynamic range properties have been applied to various photometry tasks in recent years. Muglikar et al. [19] put a linear polarizer rotating at high speeds in front of an event camera to reconstruct + +relative intensities at multiple polarizer angles, which were used for surface normals estimation. Han et al. [11] computed the frequencies of the event triggering during the transient process of global lighting changes for depth sensing and object iso-contour reconstruction. Zhou et al. [33] proposed to use the differential signals captured from an event camera to separate the direct and global illumination components in RGB videos by applying a sweeping occluder. EventPS [29] leveraged an event camera and a fast-rotating light source to achieve real-time surface normal reconstruction, which demonstrated the superiority over the conventional cameras in speed and data bandwidth. EventPSR can further estimate surface normal and reflectance parameters with high robustness and efficiency by designing light scanning patterns and estimation algorithms. + +# 3. Method + +As shown in Fig. 1, the overall workflow of EventPSR includes the design of suitable lighting scanning patterns (Sec. 3.2) and a two-stage algorithm that recovers surface normal and reflectance parameters (Sec. 3.3). In Sec. 3.1, we start from the mathematical formulation of event cameras, parameterized reflectance models, and target function. + +# 3.1. Problem Formulation + +Event formation model. Event cameras measure the scene brightness changes asynchronously on a logarithmic scale. An event $e = \{t, x, y, p\}$ is triggered when the changes of logarithmic brightness reach a triggering threshold $C$ , where $t$ is the timestamp, $(x, y)$ is the pixel coordinate, and $p \in \{-1, +1\}$ is the polarity which represents the decrease or increase of brightness. Assuming there are totally $K$ events triggered at pixel $(x, y)$ during a short period of time, these events are represented as $\mathcal{E}_{x,y} = \{x, y, p_k, t_k\}$ , where $k = \{1, 2, \dots, K\}$ . The irradiance value $I$ in pixel $(x, y)$ changing from $t_{k-1}$ to $t_k$ is represented as: + +$$ +\log \left(I _ {x, y} \left(t _ {k}\right) + \epsilon\right) = \log \left(I _ {x, y} \left(t _ {k - 1} + \eta\right) + \epsilon\right) + p _ {k} C, \tag {1} +$$ + +where $\epsilon$ is a small offset value to avoid taking the logarithm of zero, and $\eta$ is the refractory time of the pixel [9]. + +Reflectance model and inverse problem. We use a simplified version of Disney Principled BRDF [4] model to represent the reflectance $(\mathbf{f}_{r,x,y})$ of the surface. Given an opaque and non-emissive object with a distant light source, the irradiance value $I_{x,y}(t)$ at timestamp $t$ is as follows: + +$$ +I _ {x, y} (t) = \int_ {\mathbf {w} _ {i} \in \Omega} \left(\mathbf {N} _ {x, y} \cdot \mathbf {w} _ {\mathrm {i}}\right) \mathbf {f} _ {r, x, y} \left(\mathbf {w} _ {\mathrm {i}}, \mathbf {w} _ {\mathrm {r}}\right) \mathbf {L} \left(\mathbf {w} _ {i}, t\right) \mathrm {d} \mathbf {w} _ {i}, \tag {2} +$$ + +where $\Omega$ is the whole sphere direction space, $\mathbf{N}_{x,y}$ is the surface normal, $\mathbf{L}$ is the light source function, $\mathbf{w}_i$ and $\mathbf{w}_r$ are the incident and reflection direction, respectively. The light source function $\mathbf{L}(\mathbf{w}_i,t)$ represents a distant light with + +the same brightness at an angle of $\mathbf{w}_i$ . For each pixel $(x,y)$ , its reflectance $\mathbf{f}_{r,x,y}$ is represented as: + +$$ +\mathbf {f} _ {r, x, y} \left(\mathbf {w} _ {\mathrm {i}}, \mathbf {w} _ {\mathrm {r}}\right) = \frac {1 - M _ {x , y}}{\pi} + \frac {M _ {x , y} D G}{\pi \left(\mathbf {N} _ {x , y} \cdot \mathbf {w} _ {\mathrm {i}}\right) \left(\mathbf {N} _ {x , y} \cdot \mathbf {w} _ {\mathrm {r}}\right)}, \tag {3} +$$ + +where $D, G$ are parameters that describe specular reflection property. $^3$ The variation of diffuse albedo has been canceled by the log difference mechanism in event cameras [29], so we set it as 1 in our method. Other parameters and Fresnel term from Disney Principled BRDF [4] are removed for simplicity. + +Target function. To recover the surface properties, i.e., $\mathbf{N}_{x,y}$ (Surface normal), $R_{x,y}$ (Roughness, in the supplementary material), and $M_{x,y}$ (Metallic), we design a light scanning pattern $\mathbf{L}(\mathbf{w}_i,t)$ and apply an event camera to solve the inverse reflectance estimation problem. We represent the shape and reflection estimation algorithm as $\tilde{\mathbf{N}} (\mathcal{E})$ , $\tilde{\mathbf{R}} (\mathcal{E})$ , and $\tilde{\mathbf{M}} (\mathcal{E})$ . Our target is to design these algorithms and optimize light scanning patterns $\mathbf{L}$ to estimate the surface normal and reflectance parameters: + +$$ +\underset {\mathbf {L}, \tilde {\mathbf {N}}, \tilde {\mathbf {R}}, \tilde {\mathbf {M}}} {\operatorname {a r g m i n}} \sum_ {\mathbf {N}, M, R} | \tilde {\mathbf {N}} (\mathcal {E}) - \mathbf {N} | _ {2} + \tag {4} +$$ + +$$ +\left| \tilde {\mathbf {M}} (\mathcal {E}) - M \right| _ {2} + \left| \tilde {\mathbf {R}} (\mathcal {E}) - R \right| _ {2}, +$$ + +where the events $\mathcal{E}$ are triggered from $\mathbf{L},\mathbf{N},R,M$ + +# 3.2. Light Scanning Pattern Design + +It is necessary to design a suitable light scanning pattern to cover different materials and the shape of target objects. There are three requirements when designing the light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. In the following, we consider an effective use of a ring-shaped light source as the base scanning pattern. Taking a ring moving along the $Z$ axis as an example, the formula of $\mathbf{L}(\mathbf{w}_i,t)$ is: + +$$ +\mathbf {L} \left(\mathbf {w} _ {i}, t\right) = \max _ {u} \max (0, \mathbf {w} _ {i} \cdot \mathbf {R} (u, t)) ^ {\alpha}, \tag {5} +$$ + +$$ +\mathbf {R} (u, t) = [ \sin (t) \sin (u), \sin (t) \cos (u), \cos (t) ], +$$ + +where $\mathbf{R}(u,t)$ represent all light directions on a ring. $\alpha$ is the width of the ring. In order to illustrate that our "moving ring light" pattern has better properties to overcome all these challenges, we compare it with two other light source patterns: "point light" and "structured environment map". + +For the "point light" pattern, we combine two spiral movement trajectories to cover the whole sphere: the horizontal spiral and vertical spiral. In this way, we obtain a continuous moving light source pattern. We also add a + +![](images/01d73d84d2c4bf9c9107e76ea05204d30cb3e36ae8cde7e1591567d04b495d14.jpg) +Figure 2. The design of light scanning pattern. (a) Illustration of specular coverage requirement. The first row shows a shiny ball under three lighting patterns. The second row shows where specular parameters can be recovered (the brighter the better). (b) Illustration of diffuse sensitivity requirement. The first row shows the rendered frames of a diffuse ball under three lighting patterns. The second row shows the profiles of intensity changes of two different points on the sphere over time. (c) Illustration of event efficiency requirement. The first row shows the event signals triggered under three lighting patterns. The second row plots the average (green) and max (red) number of events triggered over time when covering the whole specular reflection. + +smoothness term $\varepsilon$ (the size of the point light), so the lighting will cover the whole sphere lighting direction gradually. For the horizontal spiral, the formula $\mathbf{L}(\mathbf{w}_i,t)$ is: + +$$ +\begin{array}{l} \mathbf {L} (\mathbf {w} _ {i}, t) = (\mathbf {R} (t) \cdot \mathbf {w} _ {i}) ^ {\alpha}, \\ \mathbf {R} (t) = [ \begin{array}{l l l l l l} \cdot & (\partial_ {t}) & \cdot & (\partial_ {t}) & \cdot & (\partial_ {t}) \end{array} ] \end{array} \tag {6} +$$ + +$$ +\mathbf {R} (t) = [ \sin (\beta t) \cos (t), \cos (\beta t) \cos (t), \sin (t) ], +$$ + +where $\beta$ determines the number of spirals. For the vertical spiral, the formula $\mathbf{L}(\mathbf{w}_i,t)$ is: + +$$ +\begin{array}{l} \mathbf {L} (\mathbf {w} _ {i}, t) = (\mathbf {R} (t) \cdot \mathbf {w} _ {i}) ^ {\alpha}, \\ \text {T r} (t) = [ (0) + (1) + (2) + (3) + (4) ] \end{array} \tag {7} +$$ + +$$ +\mathbf {R} (t) = [ \cos (\beta t) \sin (t), \cos (\beta t) \cos (t), \sin (\beta t) ]. +$$ + +For the "structured environment map" pattern, we project structured light along three axes alternatively. Taking structured light along the $Z$ axis as an example, the formula of this light source pattern $\mathbf{L}(\mathbf{w}_i,t)$ is: + +$$ +\mathbf {L} (\mathbf {w} _ {i}, t) = \left\lfloor \mathrm {f m o d} (\mathbf {R} (2 ^ {t} \mathbf {w} _ {i}), 2) \right\rfloor , +$$ + +$$ +\mathbf {R} \left(\mathbf {w} _ {i}\right) = \frac {\operatorname {a t a n 2} \left(\mathbf {w} _ {i} \cdot [ 0 , 1 , 0 ] , \mathbf {w} _ {i} \cdot [ 1 , 0 , 0 ]\right)}{\pi} + 1, \tag {8} +$$ + +where "fmod" is the modulo operator. $t$ should be integer and the "structured environment map" pattern is discrete. + +Specular coverage requirement. Given metallic (i.e., without diffuse part) or even mirror-like objects, it is difficult to use single distant point light for surface normal estimation. Because the specular surface only responds to lighting from a small lobe. For the specular dielectric material (i.e., with a diffuse part), the surface normal can be estimated from the diffuse reflection. However, the material properties are still only recoverable in the small specular highlight part. We show rendered images of a shiny ball under all three lighting setups in the top row of Fig. 2 (a). The bottom row shows that the solvable regions for "point + +light” is a small circle around the lighting direction. In contrast, the “moving ring light” and “structured environment map” have larger solvable regions with a strip around the ring or structure edges, which is more efficient for specular coverage. The larger specular coverage results in faster data-capturing speed for challenging specular surface materials. + +Diffuse sensitivity requirement. Diffuse reflection has the opposite property from specular reflection: it spreads the input light to all the upper hemisphere directions instead of only a small highlight region. In order to estimate surface normal from the reflection, the light source pattern needs to contain high and diverse low-frequency parts to make it distinguishable from different surface normal. As shown in Fig. 2 (b), we render the images of a rough diffusive ball under all three lighting setups in the top row, and plot the reflection intensity of two pixels in the bottom row. The diffuse reflection of two pixels is shifted in time for "point light" and "moving ring light" patterns. Since event cameras have an extremely high temporal resolution, the time shift can be well-captured by the camera and converted to surface normal accurately. In contrast, the diffuse reflection intensity profiles for the two pixels are similar for "structured environment map", making it difficult to distinguish the surface normal of two points. + +Event efficiency requirement. Since the event camera measures the brightness changes, the bandwidth consumption is mainly determined by the brightness changing magnitude in the scene. We show the synthetic event signals and the accumulated number of events from rendered a mirror-like ball in Fig. 2 (c). Both "point light" and "moving ring light" have a similar number of events triggered by different surface normal (observed by two pixels). The "moving ring light" pattern consumes fewer events for specular coverage than the "point light", indicating a better bandwidth efficiency. In contrast, the "structured environment map" + +leads to a variable number of events across different surface normal. In the worst case, where there is a mirror plane facing at the worst surface normal with a maximum number of events generated, the whole scene is flickering, and the read-out will be congested. + +Based on the comparisons and analysis above, the designed "moving ring light" achieves relatively larger coverage of various materials while preserving higher efficiency of event triggering. + +# 3.3.EventPSR Algorithm + +As shown in Fig. 1, we propose a two-stage solving pipeline, consisting of a grid-matching stage and a gradient-tuning stage. The proposed algorithm provides a global minimum result with high value resolution. + +Stage-1: Grid-matching. To estimate the surface normal and reflectance parameters from the captured event streams, we first render a database containing various intensity-time curves using a physically-based rendering engine. We assume that the rendered temporal resolution is high enough so that the intensity curves can be represented by linear interpolation of a series of frames. There are 4 parameters to be estimated for each pixel, including 2 parameters for surface normal, 1 parameter for roughness, and 1 parameter for metallic. The database is on a 5-dimension grid (1 additional dimension for time), which is defined as $\mathcal{B} = \{\mathbf{N}, R, M, t\}$ . + +Given a series of event signals $\mathcal{E}_{x,y}$ at a pixel $(x,y)$ , the goal in this stage is to find the most matching curve in the database. Therefore, we define the loss function between the observed event stream and curve candidates as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {1} (x, y, \mathbf {N}, R, M) = \sum_ {k} | \log (\mathcal {B} (\mathbf {N}, R, M, t _ {k}) + \epsilon) \tag {9} \\ - \log (\mathcal {B} (\mathbf {N}, R, M, t _ {k - 1} + \eta) + \epsilon) - p _ {k} C | ^ {2}. \\ \end{array} +$$ + +By designing the loss function on each pair of events, we prevent the event integral process, which has an integration drifting issue (More details about the drifting issue are shown in the supplementary material). Furthermore, our loss function averages errors of all events in the same pixel, implicitly suppressing event triggering noise. + +We enumerate 4D surface normal $\mathbf{N}$ and material properties $R, M$ on a discrete grid and select the one with minimum loss. The results $\tilde{\mathbf{N}}, \tilde{R}, \tilde{M}$ at this stage are as follows: + +$$ +\tilde {\mathbf {N}} _ {x, y}, \tilde {R} _ {x, y}, \tilde {M} _ {x, y} = \underset {\mathbf {N}, R, M} {\operatorname {a r g m i n}} \mathcal {L} _ {1} (x, y, \mathbf {N}, R, M). \tag {10} +$$ + +Stage-2: Gradient-tuning. The grid-matching stage provides us with a coarse result close to the global minimum, but this result can be further improved for the following two reasons: 1) The resolution of grid-matching stage depends on the searching grid size. Due to the computational cost, this resolution is not high enough. 2) In real setup, there + +![](images/41db8ea86cac9fdab44bb80aa8ac59f482a82782841ffb5e8ab23059cc814a71.jpg) +Data acquisition platform + +![](images/dd1d41468fa66b21ba52cd560a71e2b112a3bba29476fb5df48284287bb0f68e.jpg) +Light scanning patterns + +![](images/11377ca4609d62b4dfba6fd28c65da71b93135d613e3694e1eaed49251c86108.jpg) +Figure 3. Data acquisition platform for capturing real data. There are 5 commodity monitors surrounding the target object to illuminate the designed pattern from any incidence angle. We show the capture device setup on the left side. The target object is hung in the middle and an event camera is pointing at the target from the bottom. Three projected patterns are shown on the right. + +are some regions of the environment map that do not have a display. If the material is shiny, metallic, and the surface normal is reflecting to the dark region, there is no brightness change, which leads to ambiguity in these regions. More details are shown in the supplementary material. + +To solve the first problem, we calculate the gradient of surface normal $\mathbf{N}$ and material properties $R, M$ , and interpolate the database on these parameters. In this way, we obtain an approximated rendered intensity and its gradient in a continuous parameter space. Therefore, we minimize the $\mathcal{L}_1(x,y,\mathbf{N},R,M)$ using gradient descent to get the optimized result. + +To solve the second problem, we introduce a small inter-pixel smoothness regularization term: + +$$ +\begin{array}{l} \mathcal {L} _ {2} = \beta \sum_ {x, y} \left(\left| \mathbf {N} _ {x, y} - \mathbf {N} _ {x - 1, y} \right| ^ {2} + \left| \mathbf {N} _ {x, y} - \mathbf {N} _ {x, y - 1} \right| ^ {2} + \right. \\ \left| R _ {x, y} - R _ {x - 1, y} \right| ^ {2} + \left| R _ {x, y} - R _ {x, y - 1} \right| ^ {2} + \tag {11} \\ \left. \left| M _ {x, y} - M _ {x - 1, y} \right| ^ {2} + \left| M _ {x, y} - M _ {x, y - 1} \right| ^ {2}\right). \\ \end{array} +$$ + +In this way, for shiny and metallic surface, the parameters in ambiguous regions are filled with interpolated values, while the diffusive or dielectric surfaces are not affected. The optimized parameters of this gradient-tuning are the final results of EventPSR. + +# 4. Experiments + +# 4.1. Implementation Details + +Data acquisition platform. To illuminate the target object using our designed light scanning pattern, we build a data acquisition platform with multiple commodity monitors, as illustrated in Fig. 3. The target object is hanging at the center of the device and surrounded by several monitors. There is a Prophesee EVK4 event camera beneath the target object, pointing above at the object and capturing the event stream as the light scanning pattern changes. + +![](images/7825355590b4d8a870f05629cc026aa99e0b929e08b8b8f67db4532786ae279f.jpg) +Figure 4. Quantitative comparison on synthetic data. We show the relative data consumption and per-object averaged error for normal, roughness, and metallic estimation. The color heatmap for each parameter (normal, roughness, and metallic) is normalized independently. The results from NeILF [31] are empty for some objects, which are marked as "Failed". + +![](images/c31fc2e87bf4cff58bb8bbe582728be0a83ae039ad4030b4078c02b8a63ef08b.jpg) +Figure 5. Quantitative comparison results on synthetic data. We show the estimated normal, roughness, metallic, and error maps of EventPSR and two comparing methods. The averaged error maps are labeled in red numbers. + +The light scanning pattern received by the target object is slightly different from the designed one because of some non-ideal factors. The uncovered area at the top and bottom will cause some missing illumination, and the interreflection of the monitors will cause some extra light from the wrong direction. To alleviate the influences caused by non-ideal light scanning pattern projection, we capture the received illumination by a mirror ball in the scene. The database $\mathcal{B}(\mathbf{N}, R, M, t)$ is rendered using the calibrated environment maps. In addition, we flash all panels to full white at the beginning and end of the illumination pattern for temporal alignment. + +Algorithm implementation. Both of the two stages are implemented using PyTorch [20]. For the grid-matching stage, the resolution of rendered database $\mathcal{B}(\mathbf{N}, R, M, t)$ are as follows: $64 \times 64$ for normal resolution, 12 for roughness resolution, 8 for metallic resolution, and 600 for temporal resolution. According to Eq. (3), the reflectance model is a linear interpolation from pure diffuse to pure mirror. To + +reduce GPU VRAM consumption, we replace the metallic dimension using 1 pure diffuse component and 1 pure mirror component. Different material metallic properties are constructed on the GPU on the fly. Both of the grid-matching and gradient-tuning stages are conducted on a single NVIDIA GeForce RTX 3060 graphics card. + +# 4.2. Dataset + +Synthetic data. To evaluate the accuracy of our method compared with other methods, we render 16 objects using the geometries from the sculpture dataset [27]. The material parameters including albedo, roughness, and metallic, are set randomly per patch. We use a ray-tracing-based rendering engine, with 1024 samples per pixel and 3 maximum bounces. The rendered images are of $512 \times 512$ resolution. An example is shown in the first column in Fig. 5. Each of the objects is rendered under 3 different setups: Our illumination with fixed camera position, DiLiGenT10² [22] light directions with fixed camera position (for the comparison with SDM-UniPS [12]), and fixed light direction with different camera positions (for the comparison with NeILF [31]). For the first setup, the event streams are simulated from high-frame-rate image sequences. The event triggering threshold noise is set to 0.01. For the other two setups, there are total 100 HDR images rendered per object. Besides, the independent Gaussian noise of 0.01 standard variance is added to each image. + +Real data. We capture three types of real data using our data acquisition platform. The captured data are used for material roughness accuracy evaluation, surface normal accuracy evaluation, and complex object generalization evaluation, respectively. + +- Surface normal evaluation data: We select several pairs of objects with almost identical geometry but extremely different materials (i.e., rough and shiny), as shown in the first row in Fig. 6. The ground truth of surface normal is captured by a structured light 3D scanner. $^4$ +- Material roughness evaluation data: We create several + +![](images/1831b2d7033b5eaeec6e46c8dac84ee06282778fec2d4fd5bb1b49a1f4cddf1e.jpg) +Figure 6. Top: Quantitative evaluation for surface normal estimation on real objects. The error value of each object is labeled in red number. Bottom: Relighting results based on the estimated surface parameters. The environment maps are labeled on the left. The albedo of rendered objects are set to uniform white. + +balls with varying surface roughness. A polished metal ball serves as the low roughness reference. For the high roughness surface, we apply a spray coating to achieve a diffuse surface. The ground truth values of roughness are measured using a frame-based camera with HDR exposure under ideal parallel light. + +- Generalization evaluation data: We select multiple objects whose surface are composed of multiple different materials, as shown in the first row in Fig. 8. The complexity of the shape and materials will test the generalization ability of EventPSR on real-world objects. + +![](images/0b3f6b9b980bcf2abb6e2c3ca3203cddb26f79e13241e196166c289b785abd5a.jpg) +Figure 7. Quantitative demonstration for robustness of material estimation on real objects. The first row is a metal ball with mirror-like surface material. The second row is a ball sprayed with a non-glossy coating. + +# 4.3. Evaluation on Synthetic Data + +We conduct quantitative evaluations of the surface normal and reflectance parameters estimation with the synthetic dataset. The comparison includes two frame-based methods [13, 31] that represent the state-of-the-art NeRF-based [31] and PS-based [13] normal and reflectance estimation approaches. For data rate usage comparison, we estimate the date consumption in the following way: For EventPSR, we assume that the input event streams employ 16-bit Prophesee EVT 3.0 format. For the other two comparison methods, we assume that the HDR input are captured by merging 3 multi-bracketing 8-bit grayscale images. We use mean angular error to measure the accuracy of recovered normal maps. For the recovered roughness and metallic, we first standardize the data to avoid representing differences between models, then we apply mean squared error (MSE) to measure the accuracy. More details are described in the supplementary material. + +The data consumption and parameters error of each object are shown in Fig. 4. On average, our method consumes about $30\%$ of the data rate, achieving comparable accuracy of normal estimation and superior accuracy of roughness and metallic estimation. We visualize the result of one example from the synthetic dataset in Fig. 5. We can see that the error of our method concentrates mainly on the high geometry change area for all three parameter maps, where the inter-reflection and cast shadow are most severe. For the other two comparing methods [13, 31], the errors of normal estimation are more evenly distributed. However, the estimated roughness and metallic from the comparing methods [13, 31] can hardly reflect the high frequency of different material patches of the target object. More examples are shown in the supplementary material. + +To demonstrate the robustness of EventPSR in recovering various materials, we evaluate the performance of the three methods on a cow object with 6 different surface materials, from ideal Lambertion to pure mirror. More details + +![](images/b7e0419f150cfb5fd7b07a6adc0606d410dd810a1e93885ca94b06817d25d4d2.jpg) +Figure 8. Qualitative demonstration of real-world objects. From left to right, the complexities of shape and material are increasing. EventPSR recovers parameter maps consistently, and preserves high-frequency details for simple and complex objects. + +are shown in the supplementary material. + +# 4.4. Evaluation on Real Data + +Quantitative evaluation on surface normal estimation. To demonstrate that EventPSR can estimate the surface normal of objects with various materials, we select some objects with identical shape but extremely different surface materials. As shown in the top part of Fig. 6, the surface normal estimation results across different materials remain consistent, with only minor errors. The bottom part of Fig. 6 are the relighting results rendered with the estimated normal and reflectance parameters under different lighting conditions. The results demonstrate that EventPSR effectively handles material variation and maintains accuracy across different surface types. + +Quantitative evaluation on reflectance estimation. To verify the accuracy of roughness estimation using EventPSR, we conduct experiments with two semi-spheres + +featuring distinct roughness levels, as illustrated in Fig. 7. EventPSR achieves an average MSE of 0.0067 in roughness estimation, showing its high precision and robustness among different surface materials. The error maps highlight EventPSR's capability to differentiate subtle variations in material roughness, underscoring its robustness across diverse reflectance properties. + +Qualitative evaluation on real objects. The results of surface normal and reflectance parameters estimation of three real-world complex objects are shown in Fig. 8. The nose model in the left column has spatially uniform specular material. The estimated surface normal is mostly unaffected by specular reflection. The estimated roughness and metallic are mostly uniform on the surface. In the middle column, there is a diffuse cat lying on a black and metallic (mirror-like) plate. From the estimated result, we can see the low roughness and high metallic on the outer rim. There is a splatoon model with complex materials in the right column. From the image captured by a frame-based camera, we can see that different parts are in different materials. The eyes, hair, and backpack have low roughness, while the rough skin has high roughness. The recovered roughness is consistent with the observation. Besides, the high metallic in the pipes of the ink gun is also recovered by EventPSR. + +# 5. Conclusion + +We propose EventPSR, an efficient and robust approach for simultaneous estimation of surface normal and reflectance parameters using an event camera. EventPSR is the first to use continuous radiance differential signals for joint estimation of shape and material properties, reducing data rate by $30\%$ compared to frame-based methods. + +Limitations. The capturing speed is limited by the refresh rate of the current monitors. The accuracy of real object results is affected by cast shadow and inter-reflection. Despite the robustness of albedo variation, we cannot recover diffuse albedo. These factors can be improved by employing a faster device and designing more robust algorithms, which are described in the supplementary material. + +# Acknowledgement + +This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 624B2006, 62088102, Beijing Natural Science Foundation under Grant No. L233024, JST-Mirai Program under Grant No. JPMJM123G1, and JSPS Kakenhi under Grant No. 20H05953. We thank NHK (Japan Broadcasting Corporation) Science & Technology Research Laboratories (STRL) for supporting the illumination devices for real-world experiments. PKU-affiliated authors thank openbayes.com for providing computing resource. + +# References + +[1] Neil Alldrin, Todd Zickler, and David Kriegman. Photometric stereo with non-parametric and spatially-varying reflectance. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2008. 1, 2 +[2] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2 +[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of Int. Conf. Comput. Vis., 2021. 1 +[4] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, pages 1-7. vol. 2012, 2012. 3, 1 +[5] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K Wong. Self-calibrating deep photometric stereo networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3 +[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2 +[7] Hoon-Gyu Chung, Seokjun Choi, and Seung-Hwan Baek. Differentiable point-based inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2 +[8] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Trans. Graph., 37(6):1-14, 2018. 2 +[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2, 3 +[10] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M. Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2010. 1 +[11] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[12] Satoshi Ikehata. Universal photometric stereo network using global lighting contexts. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6 +[13] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7, 2 +[14] Yakun Ju, Kin-Man Lam, Wuyuan Xie, Huiyu Zhou, Junyu Dong, and Boxin Shi. Deep learning methods for calibrated photometric stereo and beyond. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3 +[15] Kaizhang Kang, Cihui Xie, Chengan He, Mingqi Yi, Minyi Gu, Zimin Chen, Kun Zhou, and Hongzhi Wu. Learning efficient illumination multiplexing for joint capture of reflectance and shape. ACM Trans. Graph., 2019. 1, 2 + +[16] M. Katzin. The scattering of electromagnetic waves from rough surfaces. Proceedings of the IEEE, 52(11):1389-1390, 1964. 1 +[17] Behnaz Kavoosighafi, Saghi Hajisharif, Ehsan Miandi, Gabriel Baravdish, Wen Cao, and Jonas Unger. Deep SVBRDF acquisition and modelling: A survey. In Computer Graphics Forum, 2024. 2 +[18] Zongrui Li, Qian Zheng, Boxin Shi, Gang Pan, and Xudong Jiang. DANI-Net: Uncalibrated photometric stereo by differentiable shadow handling, anisotropic reflectance modeling, and neural inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[19] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2 +[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Proc. of Adv. Neural Inform. Process. Syst., 2019. 6 +[21] Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Efstratios Gavves, and Tinne Tuytelaars. Deep reflectance maps. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2 +[22] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. Diligent102: A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6 +[23] Fabiano Romeiro, Yuriy Vasilyev, and Todd Zickler. Passive reflectometry. In Proc. of Eur. Conf. Comput. Vis., 2008. 2 +[24] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep photometric stereo networks for determining surface normal and reflectances. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2 +[25] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2 +[26] K. E. Torrance and E. M. Sparrow. Theory for off-specular reflection from roughened surfaces*. J. Opt. Soc. Am., 57(9): 1105-1114, 1967. 1 +[27] Olivia Wiles and Andrew Zisserman. SilNet: Single-and multi-view reconstruction by learning from silhouettes. In Proc. of Brit. Mach. Vis. Conf., 2017. 6 +[28] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of Eur. Conf. Comput. Vis., 2022. 1 +[29] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-time photometric stereo using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3 +[30] Ye Yu and William AP Smith. InverseRenderNet: Learning single image inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2 + +[31] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In Proc. of Int. Conf. Comput. Vis., 2023. 1, 6, 7, 2 +[32] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 1 +[33] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided direct and global components separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3 \ No newline at end of file diff --git a/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/images.zip b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4032a0633cc133ec0e640303088c5dbe599cd5c --- /dev/null +++ b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bcd1a3ee5e6c1f84bf2e59a2b26501ee2fd075b0bd7c847b293668c215e3cf6 +size 724040 diff --git a/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/layout.json b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5723d790791c85cd16bb8d2ac9ea0a18a71a7bc6 --- /dev/null +++ b/2025/EventPSR_ Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera/layout.json @@ -0,0 +1,7792 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 493, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 493, + 140 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 493, + 140 + ], + "type": "text", + "content": "EventPSR: Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 160, + 216, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 160, + 216, + 174 + ], + "spans": [ + { + "bbox": [ + 150, + 160, + 216, + 174 + ], + "type": "text", + "content": "Bohan " + }, + { + "bbox": [ + 150, + 160, + 216, + 174 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1,2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 161, + 286, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 161, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 236, + 161, + 286, + 174 + ], + "type": "text", + "content": "Jin Han3,4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 310, + 161, + 375, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 161, + 375, + 174 + ], + "spans": [ + { + "bbox": [ + 310, + 161, + 375, + 174 + ], + "type": "text", + "content": "Boxin Shi" + }, + { + "bbox": [ + 310, + 161, + 375, + 174 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 396, + 161, + 459, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 161, + 459, + 174 + ], + "spans": [ + { + "bbox": [ + 396, + 161, + 459, + 174 + ], + "type": "text", + "content": "Imari Sato3,4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 177, + 507, + 203 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 101, + 177, + 507, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 177, + 507, + 189 + ], + "spans": [ + { + "bbox": [ + 101, + 177, + 507, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 177, + 507, + 189 + ], + "type": "text", + "content": " State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 103, + 190, + 506, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 103, + 190, + 506, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 190, + 506, + 203 + ], + "type": "text", + "content": " National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 185, + 204, + 424, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 204, + 424, + 217 + ], + "spans": [ + { + "bbox": [ + 185, + 204, + 424, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 185, + 204, + 424, + 217 + ], + "type": "text", + "content": " The University of Tokyo " + }, + { + "bbox": [ + 185, + 204, + 424, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 185, + 204, + 424, + 217 + ], + "type": "text", + "content": " National Institute of Informatics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 220, + 505, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 99, + 220, + 505, + 232 + ], + "type": "text", + "content": "ybh1998@pku.edu.cn, jinan@nii.ac.jp, shiboxin@pku.edu.cn, imarik@nii.ac.jp" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 284, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 296, + 513 + ], + "type": "text", + "content": "Simultaneously acquisition of the surface normal and reflectance parameters is a crucial but challenging technique in the field of computer vision and graphics. It requires capturing multiple high dynamic range (HDR) images in existing methods using frame-based cameras. In this paper, we propose EventPSR, the first work to recover surface normal and reflectance parameters (e.g., metallic and roughness) simultaneously using an event camera. Compared with the existing methods based on photometric stereo or neural radiance fields, EventPSR is a robust and efficient approach that works consistently with different materials. Thanks to the extremely high temporal resolution and high dynamic range coverage of event cameras, EventPSR can recover accurate surface normal and reflectance of objects with various materials in 10 seconds. Extensive experiments on both synthetic data and real objects show that compared with existing methods using more than 100 HDR images, EventPSR recovers comparable surface normal and reflectance parameters with only about " + }, + { + "bbox": [ + 54, + 284, + 296, + 513 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 284, + 296, + 513 + ], + "type": "text", + "content": " of the data rate." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 556, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 665 + ], + "type": "text", + "content": "The fast and accurate acquisition of the shape and appearance information of an object is essential for many tasks in computer vision and computer graphics. This technique enables re-rendering realistic digital twins of real-world objects, supporting a wide range of downstream applications, such as virtual reality, augmented reality (AR), and digital content creation. However, dynamic range is a big issue when the material is composed of diffuse and highly specular properties. The time-consuming capturing process and" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 322, + 262, + 547, + 522 + ], + "blocks": [ + { + "bbox": [ + 322, + 262, + 547, + 522 + ], + "lines": [ + { + "bbox": [ + 322, + 262, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 322, + 262, + 547, + 522 + ], + "type": "image", + "image_path": "eeef97a8dd3bf674f4849eff72787baa4e0d04db33c5ece325b3c5f38623a76e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 540, + 555, + 607 + ], + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 607 + ], + "type": "text", + "content": "Figure 1. The EventPSR applies a sequence of light scanning patterns and an event camera to recover the accurate normal map and reflectance parameters (i.e., metallic and roughness) simultaneously. A two-stage pipeline consisting of a grid-matching and a gradient-tuning stage is proposed for EventPSR, which supports shape and reflectance recovery for different materials." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 626, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 555, + 650 + ], + "type": "text", + "content": "difficulty recovering highly specular surfaces are two major challenges in this area." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": "Traditional shape and reflectance estimation methods [1, 3, 10, 15, 28, 31, 32] rely on multiple high dynamic range (HDR) images captured under varying lighting conditions or viewing angles. Different from photometric stereo (PS) that estimates the surface normal of diffuse dominant ob" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 673, + 170, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 170, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 170, + 683 + ], + "type": "text", + "content": "*Corresponding author: Boxin Shi" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 56, + 684, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 684, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 684, + 295, + 703 + ], + "type": "text", + "content": "This work was conducted while the first author, Bohan Yu, was doing internship at National Institute of Informatics." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 56, + 704, + 291, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 704, + 291, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 704, + 291, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 704, + 291, + 712 + ], + "type": "text", + "content": "Code available: https://codeberg.org/ybh1998/EventPSR" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11427" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "content": "jects from several images, shape and appearance estimation methods [1, 15] require more images to ensure the accurate observation of various reflections, which is time-consuming especially for shiny materials recovery. It is difficult to fully cover the specular highlights in shiny and metallic (i.e., mirror-like) materials under directional light without complicated capturing setup [15] and multiple images to cover the dynamic range. Such a dilemma between speed and accuracy partly attributes to the insufficient capability of conventional frame-based cameras." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 324 + ], + "type": "text", + "content": "Event cameras characterized by extremely high temporal resolution and dynamic range, have shown promising advantages in many computer vision applications [9]. Researchers have explored applying event cameras in various photometric tasks [6, 19, 29, 33]. Particularly, EventPS [29] has demonstrated the superiority of event cameras for real-time surface normal estimation, showing the potential to accelerate and improve accuracy for shape and reflectance estimation tasks. However, it is still unclear how to extract material-related parameters, including roughness and metallic, from the event stream only." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 324, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 324, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 324, + 295, + 528 + ], + "type": "text", + "content": "In this paper, we propose EventPSR, a novel solution for simultaneous estimation of both surface normal and Reflectance properties from Photometric Stereo using an Event camera. As shown in Fig. 1, EventPSR can simultaneously recover an object's shape and reflectance properties" + }, + { + "bbox": [ + 56, + 324, + 295, + 528 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 324, + 295, + 528 + ], + "type": "text", + "content": " with designed light scanning patterns and a two-stage algorithm, which supports the recovery of different materials. Directly applying the setting and algorithm from EventPS [29] (i.e., capturing objects with a fixed event camera under a fast-moving point light source) is a non-trivial task. It is impossible to estimate the shape and reflectance at the same time without conquering the two challenges: 1) Design a suitable light scanning pattern that covers various materials and fully reveals the shape and reflectance properties of the objects. 2) Propose a new algorithm that converts the event signals to accurate normal and reflectance parameters simultaneously under the light scanning pattern." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 684 + ], + "type": "text", + "content": "To conquer the first challenge, we discover that there are three requirements that need to be satisfied for a suitable light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. Based on this, we design a pattern of \"1D strip light\" that covers the full sphere of the illumination direction space. It is designed not only for various materials ranging from ideal Lambertian to pure mirror, but also for efficient capture of event cameras with faster speed and low data rate. To conquer the second challenge, we propose a two-stage algorithm to estimate surface normal and reflectance parameters, consisting of a grid-matching stage and a gradient-tuning stage. We first render a database consisting of the intensity-changing curves, including the com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "binations of various surface normal directions and materials. In the grid-matching stage, we match the event stream triggered in each pixel to a specific curve in the database and get a group of coarse reflectance parameters. In the gradient-tuning stage, we finetune the estimated parameters to a more accurate sub-grid location. To summarize, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 553, + 277 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 157, + 553, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 194 + ], + "type": "text", + "content": "- We propose EventPSR, the first solution to estimate surface normal and reflectance parameters simultaneously from continuous radiance differential event signals." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "type": "text", + "content": "- We design a suitable light scanning pattern and an efficient two-stage algorithm with only " + }, + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "type": "text", + "content": " data rate compared with frame-based counterparts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 229, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 277 + ], + "type": "text", + "content": "- We validate the effectiveness of EventPSR by a platform with multiple monitors. The results show that EventPSR can estimate accurate shape-related and material-related parameters of real objects." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 287, + 405, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 405, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 405, + 300 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": "Our method is related to PS works, and comprehensive surveys can be found in [14, 25]. This section focuses on works addressing PS and surface reflectometry using conventional frame-based cameras, as well as recent advancements in photometry using event cameras." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "text", + "content": "Surface reflectometry. Measuring the detailed surface reflectance requires the use of controlled lighting to sample photometric stereo images. Romeiro et al. [23] inferred the bi-directional surface reflectance of materials using a light probe and a single HDR image of the known surface normal. Dupuy and Jakob [8] measured the reflectance spectra using a gonio-photometer to simultaneously handle BRDF acquisition, storage, and rendering. Learning-based BRDF acquisition and modeling approaches are summarized in the survey paper [17]. EventPSR differs from this category of research because we simultaneously recover the normal and reflectance parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "type": "text", + "content": "Normal and reflectance estimation from conventional cameras. Alldrin et al. [1] proposed to employ novel bivariate approximations of isotropic reflectance functions to simultaneously recover the shape and reflectance of a surface from photometric stereo images. Chung et al. [7] optimized the geometry and spatially-varying BRDFs through a point-based rendering. Learning-based methods applied convolutional neural networks on single image [2, 21, 30] or multiple images [15, 24] to estimate the surface normal and reflectance parameters. These methods are limited in capturing speed due to the low dynamic range and high latency of conventional cameras." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Event camera-based photometry. Event cameras with high temporal resolution and high dynamic range properties have been applied to various photometry tasks in recent years. Muglikar et al. [19] put a linear polarizer rotating at high speeds in front of an event camera to reconstruct" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2In this work, we recover reflectance-related properties specifically with roughness and metallic." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11428" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "relative intensities at multiple polarizer angles, which were used for surface normals estimation. Han et al. [11] computed the frequencies of the event triggering during the transient process of global lighting changes for depth sensing and object iso-contour reconstruction. Zhou et al. [33] proposed to use the differential signals captured from an event camera to separate the direct and global illumination components in RGB videos by applying a sweeping occluder. EventPS [29] leveraged an event camera and a fast-rotating light source to achieve real-time surface normal reconstruction, which demonstrated the superiority over the conventional cameras in speed and data bandwidth. EventPSR can further estimate surface normal and reflectance parameters with high robustness and efficiency by designing light scanning patterns and estimation algorithms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 111, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 111, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 111, + 274 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 354 + ], + "type": "text", + "content": "As shown in Fig. 1, the overall workflow of EventPSR includes the design of suitable lighting scanning patterns (Sec. 3.2) and a two-stage algorithm that recovers surface normal and reflectance parameters (Sec. 3.3). In Sec. 3.1, we start from the mathematical formulation of event cameras, parameterized reflectance models, and target function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 180, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 180, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 180, + 372 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": "Event formation model. Event cameras measure the scene brightness changes asynchronously on a logarithmic scale. An event " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "e = \\{t, x, y, p\\}" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " is triggered when the changes of logarithmic brightness reach a triggering threshold " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " is the timestamp, " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " is the pixel coordinate, and " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "p \\in \\{-1, +1\\}" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " is the polarity which represents the decrease or increase of brightness. Assuming there are totally " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " events triggered at pixel " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " during a short period of time, these events are represented as " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{x,y} = \\{x, y, p_k, t_k\\}" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "k = \\{1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": ". The irradiance value " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " in pixel " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " changing from " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "t_{k-1}" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 378, + 296, + 510 + ], + "type": "text", + "content": " is represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 517, + 295, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 517, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 517, + 295, + 532 + ], + "type": "interline_equation", + "content": "\\log \\left(I _ {x, y} \\left(t _ {k}\\right) + \\epsilon\\right) = \\log \\left(I _ {x, y} \\left(t _ {k - 1} + \\eta\\right) + \\epsilon\\right) + p _ {k} C, \\tag {1}", + "image_path": "4c27e5f403d2e598aaed37f258070181793ebd8b41e979ae88d45df83752152c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": " is a small offset value to avoid taking the logarithm of zero, and " + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": " is the refractory time of the pixel [9]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "text", + "content": "Reflectance model and inverse problem. We use a simplified version of Disney Principled BRDF [4] model to represent the reflectance " + }, + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{r,x,y})" + }, + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "text", + "content": " of the surface. Given an opaque and non-emissive object with a distant light source, the irradiance value " + }, + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "inline_equation", + "content": "I_{x,y}(t)" + }, + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "text", + "content": " at timestamp " + }, + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 562, + 295, + 623 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 629, + 295, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 629, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 61, + 629, + 295, + 664 + ], + "type": "interline_equation", + "content": "I _ {x, y} (t) = \\int_ {\\mathbf {w} _ {i} \\in \\Omega} \\left(\\mathbf {N} _ {x, y} \\cdot \\mathbf {w} _ {\\mathrm {i}}\\right) \\mathbf {f} _ {r, x, y} \\left(\\mathbf {w} _ {\\mathrm {i}}, \\mathbf {w} _ {\\mathrm {r}}\\right) \\mathbf {L} \\left(\\mathbf {w} _ {i}, t\\right) \\mathrm {d} \\mathbf {w} _ {i}, \\tag {2}", + "image_path": "07535ef97d483f9a2d4b3d76612268739474d629b2479c56effc92417f7b7afb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the whole sphere direction space, " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{x,y}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the surface normal, " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the light source function, " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_r" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " are the incident and reflection direction, respectively. The light source function " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\mathbf{w}_i,t)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represents a distant light with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "the same brightness at an angle of " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": ". For each pixel " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": ", its reflectance " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{r,x,y}" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": " is represented as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 105, + 553, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 105, + 553, + 141 + ], + "spans": [ + { + "bbox": [ + 318, + 105, + 553, + 141 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {r, x, y} \\left(\\mathbf {w} _ {\\mathrm {i}}, \\mathbf {w} _ {\\mathrm {r}}\\right) = \\frac {1 - M _ {x , y}}{\\pi} + \\frac {M _ {x , y} D G}{\\pi \\left(\\mathbf {N} _ {x , y} \\cdot \\mathbf {w} _ {\\mathrm {i}}\\right) \\left(\\mathbf {N} _ {x , y} \\cdot \\mathbf {w} _ {\\mathrm {r}}\\right)}, \\tag {3}", + "image_path": "394034dca0277bd8bb0ae48b9bc3d9becbf0b259cb39b89919ebaa083e018d2e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 142, + 555, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 555, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 142, + 555, + 214 + ], + "type": "inline_equation", + "content": "D, G" + }, + { + "bbox": [ + 313, + 142, + 555, + 214 + ], + "type": "text", + "content": " are parameters that describe specular reflection property." + }, + { + "bbox": [ + 313, + 142, + 555, + 214 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 313, + 142, + 555, + 214 + ], + "type": "text", + "content": " The variation of diffuse albedo has been canceled by the log difference mechanism in event cameras [29], so we set it as 1 in our method. Other parameters and Fresnel term from Disney Principled BRDF [4] are removed for simplicity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": "Target function. To recover the surface properties, i.e., " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{x,y}" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": " (Surface normal), " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "R_{x,y}" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": " (Roughness, in the supplementary material), and " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "M_{x,y}" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": " (Metallic), we design a light scanning pattern " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\mathbf{w}_i,t)" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": " and apply an event camera to solve the inverse reflectance estimation problem. We represent the shape and reflection estimation algorithm as " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{N}} (\\mathcal{E})" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{R}} (\\mathcal{E})" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}} (\\mathcal{E})" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": ". Our target is to design these algorithms and optimize light scanning patterns " + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 230, + 555, + 338 + ], + "type": "text", + "content": " to estimate the surface normal and reflectance parameters:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 346, + 553, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 346, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 352, + 346, + 553, + 374 + ], + "type": "interline_equation", + "content": "\\underset {\\mathbf {L}, \\tilde {\\mathbf {N}}, \\tilde {\\mathbf {R}}, \\tilde {\\mathbf {M}}} {\\operatorname {a r g m i n}} \\sum_ {\\mathbf {N}, M, R} | \\tilde {\\mathbf {N}} (\\mathcal {E}) - \\mathbf {N} | _ {2} + \\tag {4}", + "image_path": "0c06b7c0b3af3d87bc4b6c1ed1abcad3af82eb2c3c08075842265464ce71493a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 388, + 375, + 515, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 375, + 515, + 390 + ], + "spans": [ + { + "bbox": [ + 388, + 375, + 515, + 390 + ], + "type": "interline_equation", + "content": "\\left| \\tilde {\\mathbf {M}} (\\mathcal {E}) - M \\right| _ {2} + \\left| \\tilde {\\mathbf {R}} (\\mathcal {E}) - R \\right| _ {2},", + "image_path": "24a96e7f909b2c49110108988c0f62af2bda91099433e7ab931192127a242f58.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 399, + 520, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 520, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 520, + 411 + ], + "type": "text", + "content": "where the events " + }, + { + "bbox": [ + 313, + 399, + 520, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 399, + 520, + 411 + ], + "type": "text", + "content": " are triggered from " + }, + { + "bbox": [ + 313, + 399, + 520, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{L},\\mathbf{N},R,M" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 420, + 481, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 481, + 433 + ], + "type": "text", + "content": "3.2. Light Scanning Pattern Design" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "text", + "content": "It is necessary to design a suitable light scanning pattern to cover different materials and the shape of target objects. There are three requirements when designing the light scanning pattern: specular coverage, diffuse sensitivity, and event efficiency. In the following, we consider an effective use of a ring-shaped light source as the base scanning pattern. Taking a ring moving along the " + }, + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "text", + "content": " axis as an example, the formula of " + }, + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\mathbf{w}_i,t)" + }, + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 337, + 543, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 543, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 337, + 543, + 553, + 564 + ], + "type": "interline_equation", + "content": "\\mathbf {L} \\left(\\mathbf {w} _ {i}, t\\right) = \\max _ {u} \\max (0, \\mathbf {w} _ {i} \\cdot \\mathbf {R} (u, t)) ^ {\\alpha}, \\tag {5}", + "image_path": "6384a64aad93d695d5b92cb074293af38221be97ec29e36efb723a72c1cb04ee.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 340, + 563, + 529, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 563, + 529, + 576 + ], + "spans": [ + { + "bbox": [ + 340, + 563, + 529, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {R} (u, t) = [ \\sin (t) \\sin (u), \\sin (t) \\cos (u), \\cos (t) ],", + "image_path": "688f222f9ae2766da5e84a9af0c09837a7e6bd9ae9c06f7de65383f9f35ec81d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 586, + 555, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 586, + 555, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(u,t)" + }, + { + "bbox": [ + 313, + 586, + 555, + 646 + ], + "type": "text", + "content": " represent all light directions on a ring. " + }, + { + "bbox": [ + 313, + 586, + 555, + 646 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 586, + 555, + 646 + ], + "type": "text", + "content": " is the width of the ring. In order to illustrate that our \"moving ring light\" pattern has better properties to overcome all these challenges, we compare it with two other light source patterns: \"point light\" and \"structured environment map\"." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 647, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 694 + ], + "type": "text", + "content": "For the \"point light\" pattern, we combine two spiral movement trajectories to cover the whole sphere: the horizontal spiral and vertical spiral. In this way, we obtain a continuous moving light source pattern. We also add a" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 314, + 702, + 511, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 702, + 511, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 702, + 511, + 713 + ], + "type": "text", + "content": "3Detailed formulas are shown in the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11429" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 196 + ], + "type": "image", + "image_path": "01d73d84d2c4bf9c9107e76ea05204d30cb3e36ae8cde7e1591567d04b495d14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 202, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 202, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. The design of light scanning pattern. (a) Illustration of specular coverage requirement. The first row shows a shiny ball under three lighting patterns. The second row shows where specular parameters can be recovered (the brighter the better). (b) Illustration of diffuse sensitivity requirement. The first row shows the rendered frames of a diffuse ball under three lighting patterns. The second row shows the profiles of intensity changes of two different points on the sphere over time. (c) Illustration of event efficiency requirement. The first row shows the event signals triggered under three lighting patterns. The second row plots the average (green) and max (red) number of events triggered over time when covering the whole specular reflection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 316 + ], + "type": "text", + "content": "smoothness term " + }, + { + "bbox": [ + 55, + 279, + 295, + 316 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 279, + 295, + 316 + ], + "type": "text", + "content": " (the size of the point light), so the lighting will cover the whole sphere lighting direction gradually. For the horizontal spiral, the formula " + }, + { + "bbox": [ + 55, + 279, + 295, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\mathbf{w}_i,t)" + }, + { + "bbox": [ + 55, + 279, + 295, + 316 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 332, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 332, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 79, + 332, + 295, + 352 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {L} (\\mathbf {w} _ {i}, t) = (\\mathbf {R} (t) \\cdot \\mathbf {w} _ {i}) ^ {\\alpha}, \\\\ \\mathbf {R} (t) = [ \\begin{array}{l l l l l l} \\cdot & (\\partial_ {t}) & \\cdot & (\\partial_ {t}) & \\cdot & (\\partial_ {t}) \\end{array} ] \\end{array} \\tag {6}", + "image_path": "79382ed94a880092ec78c40141cfa0af75b1b50fc3f0b9281a57a4f870c5e10d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 348, + 270, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 348, + 270, + 361 + ], + "spans": [ + { + "bbox": [ + 82, + 348, + 270, + 361 + ], + "type": "interline_equation", + "content": "\\mathbf {R} (t) = [ \\sin (\\beta t) \\cos (t), \\cos (\\beta t) \\cos (t), \\sin (t) ],", + "image_path": "257bad0af5096a509c9738e1d20af41355bfb18ac5d3ff9fceb8994249afec39.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "text", + "content": " determines the number of spirals. For the vertical spiral, the formula " + }, + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\mathbf{w}_i,t)" + }, + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 411, + 295, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 411, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 70, + 411, + 295, + 429 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {L} (\\mathbf {w} _ {i}, t) = (\\mathbf {R} (t) \\cdot \\mathbf {w} _ {i}) ^ {\\alpha}, \\\\ \\text {T r} (t) = [ (0) + (1) + (2) + (3) + (4) ] \\end{array} \\tag {7}", + "image_path": "69ad7a83b8d78fcc1a65ab1b99548c7a46e4d1f45216378611eae72949152496.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 426, + 267, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 426, + 267, + 439 + ], + "spans": [ + { + "bbox": [ + 72, + 426, + 267, + 439 + ], + "type": "interline_equation", + "content": "\\mathbf {R} (t) = [ \\cos (\\beta t) \\sin (t), \\cos (\\beta t) \\cos (t), \\sin (\\beta t) ].", + "image_path": "0fecf44937283e6791352fc358282d5e3f8639a498ded4f501f4f9b49bace89b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 446, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 495 + ], + "type": "text", + "content": "For the \"structured environment map\" pattern, we project structured light along three axes alternatively. Taking structured light along the " + }, + { + "bbox": [ + 55, + 446, + 296, + 495 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 446, + 296, + 495 + ], + "type": "text", + "content": " axis as an example, the formula of this light source pattern " + }, + { + "bbox": [ + 55, + 446, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\mathbf{w}_i,t)" + }, + { + "bbox": [ + 55, + 446, + 296, + 495 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 511, + 211, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 511, + 211, + 526 + ], + "spans": [ + { + "bbox": [ + 69, + 511, + 211, + 526 + ], + "type": "interline_equation", + "content": "\\mathbf {L} (\\mathbf {w} _ {i}, t) = \\left\\lfloor \\mathrm {f m o d} (\\mathbf {R} (2 ^ {t} \\mathbf {w} _ {i}), 2) \\right\\rfloor ,", + "image_path": "451fa2eacb46933befe243f931869dffe038ed910c2a1d4f3f817bd89d2cf4b7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 525, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 525, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 70, + 525, + 295, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {R} \\left(\\mathbf {w} _ {i}\\right) = \\frac {\\operatorname {a t a n 2} \\left(\\mathbf {w} _ {i} \\cdot [ 0 , 1 , 0 ] , \\mathbf {w} _ {i} \\cdot [ 1 , 0 , 0 ]\\right)}{\\pi} + 1, \\tag {8}", + "image_path": "7c84ed17816c7cd09eceecaa7bfad0aab430ff35c607ef728abe8ba89c98fbf2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 581 + ], + "type": "text", + "content": "where \"fmod\" is the modulo operator. " + }, + { + "bbox": [ + 55, + 558, + 295, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 558, + 295, + 581 + ], + "type": "text", + "content": " should be integer and the \"structured environment map\" pattern is discrete." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Specular coverage requirement. Given metallic (i.e., without diffuse part) or even mirror-like objects, it is difficult to use single distant point light for surface normal estimation. Because the specular surface only responds to lighting from a small lobe. For the specular dielectric material (i.e., with a diffuse part), the surface normal can be estimated from the diffuse reflection. However, the material properties are still only recoverable in the small specular highlight part. We show rendered images of a shiny ball under all three lighting setups in the top row of Fig. 2 (a). The bottom row shows that the solvable regions for \"point" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 279, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 361 + ], + "type": "text", + "content": "light” is a small circle around the lighting direction. In contrast, the “moving ring light” and “structured environment map” have larger solvable regions with a strip around the ring or structure edges, which is more efficient for specular coverage. The larger specular coverage results in faster data-capturing speed for challenging specular surface materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "content": "Diffuse sensitivity requirement. Diffuse reflection has the opposite property from specular reflection: it spreads the input light to all the upper hemisphere directions instead of only a small highlight region. In order to estimate surface normal from the reflection, the light source pattern needs to contain high and diverse low-frequency parts to make it distinguishable from different surface normal. As shown in Fig. 2 (b), we render the images of a rough diffusive ball under all three lighting setups in the top row, and plot the reflection intensity of two pixels in the bottom row. The diffuse reflection of two pixels is shifted in time for \"point light\" and \"moving ring light\" patterns. Since event cameras have an extremely high temporal resolution, the time shift can be well-captured by the camera and converted to surface normal accurately. In contrast, the diffuse reflection intensity profiles for the two pixels are similar for \"structured environment map\", making it difficult to distinguish the surface normal of two points." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Event efficiency requirement. Since the event camera measures the brightness changes, the bandwidth consumption is mainly determined by the brightness changing magnitude in the scene. We show the synthetic event signals and the accumulated number of events from rendered a mirror-like ball in Fig. 2 (c). Both \"point light\" and \"moving ring light\" have a similar number of events triggered by different surface normal (observed by two pixels). The \"moving ring light\" pattern consumes fewer events for specular coverage than the \"point light\", indicating a better bandwidth efficiency. In contrast, the \"structured environment map\"" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11430" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": "leads to a variable number of events across different surface normal. In the worst case, where there is a mirror plane facing at the worst surface normal with a maximum number of events generated, the whole scene is flickering, and the read-out will be congested." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 180 + ], + "type": "text", + "content": "Based on the comparisons and analysis above, the designed \"moving ring light\" achieves relatively larger coverage of various materials while preserving higher efficiency of event triggering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 178, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 178, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 178, + 199 + ], + "type": "text", + "content": "3.3.EventPSR Algorithm" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "type": "text", + "content": "As shown in Fig. 1, we propose a two-stage solving pipeline, consisting of a grid-matching stage and a gradient-tuning stage. The proposed algorithm provides a global minimum result with high value resolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 252, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 396 + ], + "type": "text", + "content": "Stage-1: Grid-matching. To estimate the surface normal and reflectance parameters from the captured event streams, we first render a database containing various intensity-time curves using a physically-based rendering engine. We assume that the rendered temporal resolution is high enough so that the intensity curves can be represented by linear interpolation of a series of frames. There are 4 parameters to be estimated for each pixel, including 2 parameters for surface normal, 1 parameter for roughness, and 1 parameter for metallic. The database is on a 5-dimension grid (1 additional dimension for time), which is defined as " + }, + { + "bbox": [ + 55, + 252, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathbf{N}, R, M, t\\}" + }, + { + "bbox": [ + 55, + 252, + 295, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 443 + ], + "type": "text", + "content": "Given a series of event signals " + }, + { + "bbox": [ + 55, + 396, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{x,y}" + }, + { + "bbox": [ + 55, + 396, + 295, + 443 + ], + "type": "text", + "content": " at a pixel " + }, + { + "bbox": [ + 55, + 396, + 295, + 443 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 396, + 295, + 443 + ], + "type": "text", + "content": ", the goal in this stage is to find the most matching curve in the database. Therefore, we define the loss function between the observed event stream and curve candidates as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 449, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 449, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 63, + 449, + 295, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {1} (x, y, \\mathbf {N}, R, M) = \\sum_ {k} | \\log (\\mathcal {B} (\\mathbf {N}, R, M, t _ {k}) + \\epsilon) \\tag {9} \\\\ - \\log (\\mathcal {B} (\\mathbf {N}, R, M, t _ {k - 1} + \\eta) + \\epsilon) - p _ {k} C | ^ {2}. \\\\ \\end{array}", + "image_path": "8b2e2d58764b2dec518a9ee1243815bbc8f96f345c2967328b3ac5906c0b0af0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "text", + "content": "By designing the loss function on each pair of events, we prevent the event integral process, which has an integration drifting issue (More details about the drifting issue are shown in the supplementary material). Furthermore, our loss function averages errors of all events in the same pixel, implicitly suppressing event triggering noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": "We enumerate 4D surface normal " + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": " and material properties " + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "inline_equation", + "content": "R, M" + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": " on a discrete grid and select the one with minimum loss. The results " + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{N}}, \\tilde{R}, \\tilde{M}" + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": " at this stage are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 613, + 295, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 613, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 68, + 613, + 295, + 634 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {N}} _ {x, y}, \\tilde {R} _ {x, y}, \\tilde {M} _ {x, y} = \\underset {\\mathbf {N}, R, M} {\\operatorname {a r g m i n}} \\mathcal {L} _ {1} (x, y, \\mathbf {N}, R, M). \\tag {10}", + "image_path": "b747669af855727d4e661bd823dc37d29c36e8214e2317a53117a6e746f717ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Stage-2: Gradient-tuning. The grid-matching stage provides us with a coarse result close to the global minimum, but this result can be further improved for the following two reasons: 1) The resolution of grid-matching stage depends on the searching grid size. Due to the computational cost, this resolution is not high enough. 2) In real setup, there" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 416, + 174 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 416, + 174 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 416, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 416, + 174 + ], + "type": "image", + "image_path": "41db8ea86cac9fdab44bb80aa8ac59f482a82782841ffb5e8ab23059cc814a71.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 175, + 412, + 186 + ], + "lines": [ + { + "bbox": [ + 320, + 175, + 412, + 186 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 412, + 186 + ], + "type": "text", + "content": "Data acquisition platform" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 420, + 72, + 485, + 173 + ], + "blocks": [ + { + "bbox": [ + 420, + 72, + 485, + 173 + ], + "lines": [ + { + "bbox": [ + 420, + 72, + 485, + 173 + ], + "spans": [ + { + "bbox": [ + 420, + 72, + 485, + 173 + ], + "type": "image", + "image_path": "dd1d41468fa66b21ba52cd560a71e2b112a3bba29476fb5df48284287bb0f68e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 175, + 529, + 186 + ], + "lines": [ + { + "bbox": [ + 443, + 175, + 529, + 186 + ], + "spans": [ + { + "bbox": [ + 443, + 175, + 529, + 186 + ], + "type": "text", + "content": "Light scanning patterns" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 487, + 72, + 552, + 173 + ], + "blocks": [ + { + "bbox": [ + 487, + 72, + 552, + 173 + ], + "lines": [ + { + "bbox": [ + 487, + 72, + 552, + 173 + ], + "spans": [ + { + "bbox": [ + 487, + 72, + 552, + 173 + ], + "type": "image", + "image_path": "11377ca4609d62b4dfba6fd28c65da71b93135d613e3694e1eaed49251c86108.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 553, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 260 + ], + "type": "text", + "content": "Figure 3. Data acquisition platform for capturing real data. There are 5 commodity monitors surrounding the target object to illuminate the designed pattern from any incidence angle. We show the capture device setup on the left side. The target object is hung in the middle and an event camera is pointing at the target from the bottom. Three projected patterns are shown on the right." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 266, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 327 + ], + "type": "text", + "content": "are some regions of the environment map that do not have a display. If the material is shiny, metallic, and the surface normal is reflecting to the dark region, there is no brightness change, which leads to ambiguity in these regions. More details are shown in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "text", + "content": "To solve the first problem, we calculate the gradient of surface normal " + }, + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "text", + "content": " and material properties " + }, + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "inline_equation", + "content": "R, M" + }, + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "text", + "content": ", and interpolate the database on these parameters. In this way, we obtain an approximated rendered intensity and its gradient in a continuous parameter space. Therefore, we minimize the " + }, + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1(x,y,\\mathbf{N},R,M)" + }, + { + "bbox": [ + 313, + 327, + 554, + 409 + ], + "type": "text", + "content": " using gradient descent to get the optimized result." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 411, + 553, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 435 + ], + "type": "text", + "content": "To solve the second problem, we introduce a small inter-pixel smoothness regularization term:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 323, + 442, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 442, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 323, + 442, + 553, + 502 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {2} = \\beta \\sum_ {x, y} \\left(\\left| \\mathbf {N} _ {x, y} - \\mathbf {N} _ {x - 1, y} \\right| ^ {2} + \\left| \\mathbf {N} _ {x, y} - \\mathbf {N} _ {x, y - 1} \\right| ^ {2} + \\right. \\\\ \\left| R _ {x, y} - R _ {x - 1, y} \\right| ^ {2} + \\left| R _ {x, y} - R _ {x, y - 1} \\right| ^ {2} + \\tag {11} \\\\ \\left. \\left| M _ {x, y} - M _ {x - 1, y} \\right| ^ {2} + \\left| M _ {x, y} - M _ {x, y - 1} \\right| ^ {2}\\right). \\\\ \\end{array}", + "image_path": "0bc7140a9ba83d11b94be2f1fb3b7c142eb5cad0ac4b2d350e43df13c2c1ac3f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 510, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 568 + ], + "type": "text", + "content": "In this way, for shiny and metallic surface, the parameters in ambiguous regions are filled with interpolated values, while the diffusive or dielectric surfaces are not affected. The optimized parameters of this gradient-tuning are the final results of EventPSR." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 600, + 447, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 447, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 447, + 612 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Data acquisition platform. To illuminate the target object using our designed light scanning pattern, we build a data acquisition platform with multiple commodity monitors, as illustrated in Fig. 3. The target object is hanging at the center of the device and surrounded by several monitors. There is a Prophesee EVK4 event camera beneath the target object, pointing above at the object and capturing the event stream as the light scanning pattern changes." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11431" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 553, + 180 + ], + "type": "image", + "image_path": "7825355590b4d8a870f05629cc026aa99e0b929e08b8b8f67db4532786ae279f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 187, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 555, + 222 + ], + "type": "text", + "content": "Figure 4. Quantitative comparison on synthetic data. We show the relative data consumption and per-object averaged error for normal, roughness, and metallic estimation. The color heatmap for each parameter (normal, roughness, and metallic) is normalized independently. The results from NeILF [31] are empty for some objects, which are marked as \"Failed\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 228, + 296, + 427 + ], + "blocks": [ + { + "bbox": [ + 57, + 228, + 296, + 427 + ], + "lines": [ + { + "bbox": [ + 57, + 228, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 296, + 427 + ], + "type": "image", + "image_path": "c31fc2e87bf4cff58bb8bbe582728be0a83ae039ad4030b4078c02b8a63ef08b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 433, + 296, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 477 + ], + "type": "text", + "content": "Figure 5. Quantitative comparison results on synthetic data. We show the estimated normal, roughness, metallic, and error maps of EventPSR and two comparing methods. The averaged error maps are labeled in red numbers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "text", + "content": "The light scanning pattern received by the target object is slightly different from the designed one because of some non-ideal factors. The uncovered area at the top and bottom will cause some missing illumination, and the interreflection of the monitors will cause some extra light from the wrong direction. To alleviate the influences caused by non-ideal light scanning pattern projection, we capture the received illumination by a mirror ball in the scene. The database " + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\mathbf{N}, R, M, t)" + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "text", + "content": " is rendered using the calibrated environment maps. In addition, we flash all panels to full white at the beginning and end of the illumination pattern for temporal alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "content": "Algorithm implementation. Both of the two stages are implemented using PyTorch [20]. For the grid-matching stage, the resolution of rendered database " + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\mathbf{N}, R, M, t)" + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "content": " are as follows: " + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "content": " for normal resolution, 12 for roughness resolution, 8 for metallic resolution, and 600 for temporal resolution. According to Eq. (3), the reflectance model is a linear interpolation from pure diffuse to pure mirror. To" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": "reduce GPU VRAM consumption, we replace the metallic dimension using 1 pure diffuse component and 1 pure mirror component. Different material metallic properties are constructed on the GPU on the fly. Both of the grid-matching and gradient-tuning stages are conducted on a single NVIDIA GeForce RTX 3060 graphics card." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 308, + 373, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 373, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 373, + 319 + ], + "type": "text", + "content": "4.2. Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 325, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 553 + ], + "type": "text", + "content": "Synthetic data. To evaluate the accuracy of our method compared with other methods, we render 16 objects using the geometries from the sculpture dataset [27]. The material parameters including albedo, roughness, and metallic, are set randomly per patch. We use a ray-tracing-based rendering engine, with 1024 samples per pixel and 3 maximum bounces. The rendered images are of " + }, + { + "bbox": [ + 313, + 325, + 555, + 553 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 325, + 555, + 553 + ], + "type": "text", + "content": " resolution. An example is shown in the first column in Fig. 5. Each of the objects is rendered under 3 different setups: Our illumination with fixed camera position, DiLiGenT10² [22] light directions with fixed camera position (for the comparison with SDM-UniPS [12]), and fixed light direction with different camera positions (for the comparison with NeILF [31]). For the first setup, the event streams are simulated from high-frame-rate image sequences. The event triggering threshold noise is set to 0.01. For the other two setups, there are total 100 HDR images rendered per object. Besides, the independent Gaussian noise of 0.01 standard variance is added to each image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 553, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 613 + ], + "type": "text", + "content": "Real data. We capture three types of real data using our data acquisition platform. The captured data are used for material roughness accuracy evaluation, surface normal accuracy evaluation, and complex object generalization evaluation, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 613, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 556, + 672 + ], + "type": "text", + "content": "- Surface normal evaluation data: We select several pairs of objects with almost identical geometry but extremely different materials (i.e., rough and shiny), as shown in the first row in Fig. 6. The ground truth of surface normal is captured by a structured light 3D scanner." + }, + { + "bbox": [ + 313, + 613, + 556, + 672 + ], + "type": "inline_equation", + "content": "^4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 673, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 553, + 685 + ], + "type": "text", + "content": "- Material roughness evaluation data: We create several" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "4https : / / www . shining3d . com / professional - solutions/Desktop- 3d-scanner/einscan-se-v2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11432" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 294, + 507 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 294, + 507 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 294, + 507 + ], + "type": "image", + "image_path": "1831b2d7033b5eaeec6e46c8dac84ee06282778fec2d4fd5bb1b49a1f4cddf1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 519, + 295, + 575 + ], + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 575 + ], + "type": "text", + "content": "Figure 6. Top: Quantitative evaluation for surface normal estimation on real objects. The error value of each object is labeled in red number. Bottom: Relighting results based on the estimated surface parameters. The environment maps are labeled on the left. The albedo of rendered objects are set to uniform white." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 63, + 582, + 295, + 653 + ], + "type": "text", + "content": "balls with varying surface roughness. A polished metal ball serves as the low roughness reference. For the high roughness surface, we apply a spray coating to achieve a diffuse surface. The ground truth values of roughness are measured using a frame-based camera with HDR exposure under ideal parallel light." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "- Generalization evaluation data: We select multiple objects whose surface are composed of multiple different materials, as shown in the first row in Fig. 8. The complexity of the shape and materials will test the generalization ability of EventPSR on real-world objects." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 199 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 199 + ], + "type": "image", + "image_path": "0b3f6b9b980bcf2abb6e2c3ca3203cddb26f79e13241e196166c289b785abd5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 206, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 251 + ], + "type": "text", + "content": "Figure 7. Quantitative demonstration for robustness of material estimation on real objects. The first row is a metal ball with mirror-like surface material. The second row is a ball sprayed with a non-glossy coating." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 260, + 474, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 474, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 474, + 274 + ], + "type": "text", + "content": "4.3. Evaluation on Synthetic Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 280, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 483 + ], + "type": "text", + "content": "We conduct quantitative evaluations of the surface normal and reflectance parameters estimation with the synthetic dataset. The comparison includes two frame-based methods [13, 31] that represent the state-of-the-art NeRF-based [31] and PS-based [13] normal and reflectance estimation approaches. For data rate usage comparison, we estimate the date consumption in the following way: For EventPSR, we assume that the input event streams employ 16-bit Prophesee EVT 3.0 format. For the other two comparison methods, we assume that the HDR input are captured by merging 3 multi-bracketing 8-bit grayscale images. We use mean angular error to measure the accuracy of recovered normal maps. For the recovered roughness and metallic, we first standardize the data to avoid representing differences between models, then we apply mean squared error (MSE) to measure the accuracy. More details are described in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": "The data consumption and parameters error of each object are shown in Fig. 4. On average, our method consumes about " + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": " of the data rate, achieving comparable accuracy of normal estimation and superior accuracy of roughness and metallic estimation. We visualize the result of one example from the synthetic dataset in Fig. 5. We can see that the error of our method concentrates mainly on the high geometry change area for all three parameter maps, where the inter-reflection and cast shadow are most severe. For the other two comparing methods [13, 31], the errors of normal estimation are more evenly distributed. However, the estimated roughness and metallic from the comparing methods [13, 31] can hardly reflect the high frequency of different material patches of the target object. More examples are shown in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To demonstrate the robustness of EventPSR in recovering various materials, we evaluate the performance of the three methods on a cow object with 6 different surface materials, from ideal Lambertion to pure mirror. More details" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11433" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 289, + 425 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 289, + 425 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 289, + 425 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 289, + 425 + ], + "type": "image", + "image_path": "b7e0419f150cfb5fd7b07a6adc0606d410dd810a1e93885ca94b06817d25d4d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 429, + 295, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 473 + ], + "type": "text", + "content": "Figure 8. Qualitative demonstration of real-world objects. From left to right, the complexities of shape and material are increasing. EventPSR recovers parameter maps consistently, and preserves high-frequency details for simple and complex objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 480, + 223, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 223, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 223, + 491 + ], + "type": "text", + "content": "are shown in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 194, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 194, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 194, + 510 + ], + "type": "text", + "content": "4.4. Evaluation on Real Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 659 + ], + "type": "text", + "content": "Quantitative evaluation on surface normal estimation. To demonstrate that EventPSR can estimate the surface normal of objects with various materials, we select some objects with identical shape but extremely different surface materials. As shown in the top part of Fig. 6, the surface normal estimation results across different materials remain consistent, with only minor errors. The bottom part of Fig. 6 are the relighting results rendered with the estimated normal and reflectance parameters under different lighting conditions. The results demonstrate that EventPSR effectively handles material variation and maintains accuracy across different surface types." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 660, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 696 + ], + "type": "text", + "content": "Quantitative evaluation on reflectance estimation. To verify the accuracy of roughness estimation using EventPSR, we conduct experiments with two semi-spheres" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "featuring distinct roughness levels, as illustrated in Fig. 7. EventPSR achieves an average MSE of 0.0067 in roughness estimation, showing its high precision and robustness among different surface materials. The error maps highlight EventPSR's capability to differentiate subtle variations in material roughness, underscoring its robustness across diverse reflectance properties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 158, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 361 + ], + "type": "text", + "content": "Qualitative evaluation on real objects. The results of surface normal and reflectance parameters estimation of three real-world complex objects are shown in Fig. 8. The nose model in the left column has spatially uniform specular material. The estimated surface normal is mostly unaffected by specular reflection. The estimated roughness and metallic are mostly uniform on the surface. In the middle column, there is a diffuse cat lying on a black and metallic (mirror-like) plate. From the estimated result, we can see the low roughness and high metallic on the outer rim. There is a splatoon model with complex materials in the right column. From the image captured by a frame-based camera, we can see that different parts are in different materials. The eyes, hair, and backpack have low roughness, while the rough skin has high roughness. The recovered roughness is consistent with the observation. Besides, the high metallic in the pipes of the ink gun is also recovered by EventPSR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 376, + 388, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 376, + 388, + 388 + ], + "spans": [ + { + "bbox": [ + 314, + 376, + 388, + 388 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 399, + 553, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 470 + ], + "type": "text", + "content": "We propose EventPSR, an efficient and robust approach for simultaneous estimation of surface normal and reflectance parameters using an event camera. EventPSR is the first to use continuous radiance differential signals for joint estimation of shape and material properties, reducing data rate by " + }, + { + "bbox": [ + 313, + 399, + 553, + 470 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 399, + 553, + 470 + ], + "type": "text", + "content": " compared to frame-based methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "text", + "content": "Limitations. The capturing speed is limited by the refresh rate of the current monitors. The accuracy of real object results is affected by cast shadow and inter-reflection. Despite the robustness of albedo variation, we cannot recover diffuse albedo. These factors can be improved by employing a faster device and designing more robust algorithms, which are described in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 572, + 411, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 572, + 411, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 572, + 411, + 586 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 624B2006, 62088102, Beijing Natural Science Foundation under Grant No. L233024, JST-Mirai Program under Grant No. JPMJM123G1, and JSPS Kakenhi under Grant No. 20H05953. We thank NHK (Japan Broadcasting Corporation) Science & Technology Research Laboratories (STRL) for supporting the illumination devices for real-world experiments. PKU-affiliated authors thank openbayes.com for providing computing resource." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 56, + 702, + 231, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 231, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 231, + 712 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 56, + 702, + 231, + 712 + ], + "type": "text", + "content": " Objects are rendered with a uniform albedo of white." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11434" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Neil Alldrin, Todd Zickler, and David Kriegman. Photometric stereo with non-parametric and spatially-varying reflectance. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2008. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of Int. Conf. Comput. Vis., 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "type": "text", + "content": "[4] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, pages 1-7. vol. 2012, 2012. 3, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "type": "text", + "content": "[5] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K Wong. Self-calibrating deep photometric stereo networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 295, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 339 + ], + "type": "text", + "content": "[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor lighting estimation using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 373 + ], + "type": "text", + "content": "[7] Hoon-Gyu Chung, Seokjun Choi, and Seung-Hwan Baek. Differentiable point-based inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 375, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 407 + ], + "type": "text", + "content": "[8] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Trans. Graph., 37(6):1-14, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "type": "text", + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 465, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 295, + 508 + ], + "type": "text", + "content": "[10] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M. Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2010. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "text", + "content": "[11] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-fidelity event-radiance recovery via transient event frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 588 + ], + "type": "text", + "content": "[12] Satoshi Ikehata. Universal photometric stereo network using global lighting contexts. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "text", + "content": "[13] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "text", + "content": "[14] Yakun Ju, Kin-Man Lam, Wuyuan Xie, Huiyu Zhou, Junyu Dong, and Boxin Shi. Deep learning methods for calibrated photometric stereo and beyond. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[15] Kaizhang Kang, Cihui Xie, Chengan He, Mingqi Yi, Minyi Gu, Zimin Chen, Kun Zhou, and Hongzhi Wu. Learning efficient illumination multiplexing for joint capture of reflectance and shape. ACM Trans. Graph., 2019. 1, 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[16] M. Katzin. The scattering of electromagnetic waves from rough surfaces. Proceedings of the IEEE, 52(11):1389-1390, 1964. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "type": "text", + "content": "[17] Behnaz Kavoosighafi, Saghi Hajisharif, Ehsan Miandi, Gabriel Baravdish, Wen Cao, and Jonas Unger. Deep SVBRDF acquisition and modelling: A survey. In Computer Graphics Forum, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "text", + "content": "[18] Zongrui Li, Qian Zheng, Boxin Shi, Gang Pan, and Xudong Jiang. DANI-Net: Uncalibrated photometric stereo by differentiable shadow handling, anisotropic reflectance modeling, and neural inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[19] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 253, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 308 + ], + "type": "text", + "content": "[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Proc. of Adv. Neural Inform. Process. Syst., 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[21] Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Efstratios Gavves, and Tinne Tuytelaars. Deep reflectance maps. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "text", + "content": "[22] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. Diligent102: A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 421 + ], + "type": "text", + "content": "[23] Fabiano Romeiro, Yuriy Vasilyev, and Todd Zickler. Passive reflectometry. In Proc. of Eur. Conf. Comput. Vis., 2008. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 422, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 464 + ], + "type": "text", + "content": "[24] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep photometric stereo networks for determining surface normal and reflectances. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "text", + "content": "[25] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "text", + "content": "[26] K. E. Torrance and E. M. Sparrow. Theory for off-specular reflection from roughened surfaces*. J. Opt. Soc. Am., 57(9): 1105-1114, 1967. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[27] Olivia Wiles and Andrew Zisserman. SilNet: Single-and multi-view reconstruction by learning from silhouettes. In Proc. of Brit. Mach. Vis. Conf., 2017. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "text", + "content": "[28] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of Eur. Conf. Comput. Vis., 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[29] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-time photometric stereo using an event camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[30] Ye Yu and William AP Smith. InverseRenderNet: Learning single image inverse rendering. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11435" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[31] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In Proc. of Int. Conf. Comput. Vis., 2023. 1, 6, 7, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[32] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[33] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided direct and global components separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11436" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_content_list.json b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a4222b53c3a163394f87b1695c71f4dbdeb817b6 --- /dev/null +++ b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_content_list.json @@ -0,0 +1,1745 @@ +[ + { + "type": "text", + "text": "EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Toshiya Yura1,2", + "bbox": [ + 308, + 202, + 433, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "toshiya.yura@sony.com", + "bbox": [ + 277, + 224, + 467, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ashkan Mirzaei", + "bbox": [ + 557, + 203, + 692, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ashkan@cs.toronto.edu", + "bbox": [ + 531, + 224, + 718, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Igor Gilitschenski²", + "bbox": [ + 419, + 244, + 573, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gilitschenski@cs.toronto.edu", + "bbox": [ + 372, + 267, + 622, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Sony Semiconductor Solutions Corporation, 2University of Toronto", + "bbox": [ + 228, + 281, + 767, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 334, + 326, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a method for using event camera data in novel view synthesis via Gaussian Splatting. Event cameras offer exceptional temporal resolution and a high dynamic range. Leveraging these capabilities allows us to effectively address the novel view synthesis challenge in the presence of fast camera motion. For initialization of the optimization process, our approach uses prior knowledge encoded in an event-to-video model. We also use spline interpolation for obtaining high quality poses along the event camera trajectory. This enhances the reconstruction quality from fast-moving cameras while overcoming the computational limitations traditionally associated with event-based Neural Radiance Field (NeRF) methods. Our experimental evaluation demonstrates that our results achieve higher visual fidelity and better performance than existing event-based NeRF approaches while being an order of magnitude faster to render.", + "bbox": [ + 88, + 367, + 485, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 660, + 220, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras are a vision modality that draws strong inspiration from biological vision and neuromorphic design. In these sensors, each pixel operates independently by triggering events upon a log-intensity change above a pre-defined threshold [16, 37]. This results in an asynchronous data stream as opposed to capturing absolute intensity values at fixed intervals in traditional cameras. Event cameras offer low latency, high temporal resolution, and a high dynamic range while also ensuring minimal energy consumption [8, 38]. Consequently, they are particularly suitable for dynamic and challenging lighting conditions, as well as environments with rapid camera motions where classical cameras may exhibit strong motion blur [38, 63, 64, 78]. These advantages have motivated the adoption of event cameras in", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a wide range of applications covering most of the traditional computer vision tasks [13, 15, 27, 50, 51].", + "bbox": [ + 511, + 335, + 903, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, novel view synthesis has seen significant advances. This progress has been particularly driven by the emergence and development of Neural Radiance Fields (NeRF) [1-3, 33, 46, 74]. Recent works have led to NeRF improvements in terms of rendering speed, and adoption of NeRFs in 3D editing and other downstream applications [9, 43, 47-49, 84]. 3D Gaussian Splitting (3DGS) [26, 69] achieves real-time rendering by replacing the underlying representation with 3D Gaussians and using fast differentiable rasterization. The substantial speed improvement of 3DGS is achieved without compromising the visual quality.", + "bbox": [ + 511, + 367, + 906, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of NeRFs in view-synthesis has led to their adoption in applications with event-based cameras [25, 31, 42, 59]. Conventional NeRF approaches struggle with challenges, such as 3D reconstruction in environments with rapidly changing lighting conditions or high-speed objects due to RGB camera limitations. The integration of NeRFs with event-based cameras leverages their unique advantages and incorporates them into the NeRF framework to address the challenges that conventional NeRF approaches face.", + "bbox": [ + 511, + 534, + 905, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, one crucial area of research focuses on expediting the rendering processes to fully unlock the potential of event-based NeRFs. Leveraging the advantages of 3D Gaussian Splatting with events requires adapting the rasterization process to allow supervision with events or accumulated events rather than RGB images.", + "bbox": [ + 511, + 671, + 905, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In summary, this paper contributes a method for 3D Gaussian Splatting from event data by leveraging the following technical ideas:", + "bbox": [ + 511, + 763, + 903, + 808 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) An integration of event accumulation with the optimization process of 3DGS.", + "2) An event-to-video guided Structure from Motion (SfM) approach for initializing the 3DGS optimization process.", + "3) Use of cubic spline trajectory interpolation to assign camera poses to events at high rates." + ], + "bbox": [ + 514, + 810, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26876", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c99cff5ea35c3460ddcdfc4dde9e0e542feb19f53b56fad52f26f619ce808694.jpg", + "image_caption": [ + "Figure 1. EventSplat derives 3D representations of scenes in the form of 3D Gaussians from event data, enabling fast real-time renderings of scenes captured with event cameras via rasterization of 3D Gaussians. The use of event-based input is particularly advantageous when traditional RGB cameras fail due to various reasons including poor lighting conditions or motion blur from fast-moving cameras." + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 802, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our evaluations show that these design choices achieve state-of-the-art event-based novel view synthesis performance while benefiting from fast rendering speeds. Thus, our method represents an advancement in addressing the speed constraints associated with event-based NeRFs. It opens up new possibilities for real-time novel view synthesis for environments where traditional imaging techniques struggle.", + "bbox": [ + 88, + 329, + 486, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 467, + 233, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Event Cameras for Vision Applications", + "text_level": 1, + "bbox": [ + 89, + 493, + 426, + 510 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the low energy consumption, high dynamic range, and high asynchronous temporal resolution of event cameras, event-based vision has recently garnered significant attention [16]. These advantageous properties have led to promising advancements in downstream tasks such as object tracking, detection, and recognition [14, 18-21, 28, 34, 52, 61, 71], as well as optical flow estimation [4, 22, 83]. However, being a novel data format, event data necessitates specialized architectures and algorithms to be effectively processed and utilized in various applications. In this work, we present a method to exploit event data for real-time novel view synthesis, a scenario that has not been addressed previously.", + "bbox": [ + 89, + 516, + 485, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Volumetric Rendering for View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 726, + 444, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs)[46] have been the state-of-the-art in view synthesis, enabled by volumetric rendering and neural scene representations. Significant research has been conducted to improve NeRFs in terms of training and rendering speed [6, 9, 23, 32, 49, 58, 60, 76] and visual quality [1, 2, 12, 39, 65]. However, due to the expensive volumetric sampling process, NeRF-based methods have generally been unable to render high-quality views in real time. Recently, 3D Gaussian Splitting (3DGS) [7, 10, 26, 53, 85] proposed a rasterization-based approach for differentiably", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rendering scenes represented as sets of 3D Gaussian ellipsoids. This method achieves real-time renderings while preserving the quality of state-of-the-art NeRF-based methods. Since the introduction of 3DGS, real-time capabilities of 3DGS have resulted in rapid adoption of Gaussians in several downstream tasks including dynamic scene reconstruction [73], 3D asset generation [75], SLAM [45], and 4D object generation [40]. For view synthesis from event data, we leverage 3DGS and adapt it to be trainable using event sequences.", + "bbox": [ + 511, + 329, + 908, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Novel View Synthesis via Event Data", + "text_level": 1, + "bbox": [ + 511, + 488, + 831, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the unique benefits of event-based cameras, such as high dynamic range, low latency, and high temporal resolution, there has been significant research interest in adopting event data in applications like 3D reconstruction and novel view synthesis in cases where traditional RGB cameras underperform [16, 54, 81, 82]. Recent methods explore using NeRFs with event-vision supervision [25, 31, 42, 59] to reconstruct scenes and render them from novel views. While capable of generating high-quality renderings, NeRF-based methods are generally slow to render due to expensive sampling requirements [6, 17, 57, 62]. On the other hand, 3DGS requires RGB data for supervision, and adapting 3DGS for scenarios where the supervision signal comes from an event-based camera has been challenging [11, 66, 68, 70, 72, 77, 80]. In contrast to previous event-based view synthesis methods, we enable the use of event data to supervise 3DGS, allowing for real-time renderings and state-of-the-art quality using only event data, supported by event-based initial point clouds and effective camera trajectory interpolation.", + "bbox": [ + 511, + 510, + 908, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 511, + 823, + 640, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1.Event Camera Data", + "text_level": 1, + "bbox": [ + 511, + 848, + 702, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While traditional cameras output an intensity image $I$ , event cameras model changes in the log intensity $L = \\log (I)$", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26877", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of the current camera view. Similar to traditional cameras, they have a pixel array. However, in contrast to traditional cameras, each pixel acts as an independent sensor and triggers an event whenever the difference between the current log-intensity $l_{x,y}(t)$ and the log-intensity at the time of the most recent event $l_{x,y}(t_{\\mathrm{recent}})$ surpasses a given threshold $\\Delta > 0$ , i.e. when", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {x, y} (t) - l _ {x, y} \\left(t _ {\\text {r e c e n t}}\\right) > \\Delta . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 205, + 482, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This threshold is also known as contrast sensitivity [16]. It can be different for positive and negative polarity changes.", + "bbox": [ + 89, + 229, + 482, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, each event is a tuple $e := (t, x, y, p)$ where $t \\in \\mathbb{R}$ is a timestamp at which it was triggered, $(x, y) \\in \\mathbb{Z}^2$ are its pixel coordinates, and $p \\in \\{-1, 1\\}$ is the event's polarity indicating whether the log-intensity increased or decreased. Similar to traditional color cameras, a Bayer RGB pattern is used to obtain color information in color event cameras. In such sensors, each individual event provides information only for one color channel.", + "bbox": [ + 89, + 261, + 483, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 388, + 300, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable novel view synthesis, 3DGS learns a scene representation that consists of a set of ellipsoids with view-dependent color information encoded as spherical harmonics [7, 10, 26, 53, 85]. The essence of the method revolves around transforming optimized 3D Gaussians into 2D imagery based on predetermined camera positions. Subsequently the approach calculates pixel values accordingly by arranging and rasterization [7, 10, 26, 53, 85]. The transformation process involves mapping 3D Gaussian distributions, shaped like ellipsoids, onto a two-dimensional plane as ellipses for the purpose of rendering. This is achieved by applying a specific viewing transformation, denoted by $W$ , along with a 3D covariance matrix $\\Sigma$ . The outcome is the derivation of a 2D covariance matrix $\\Sigma'$ as follows", + "bbox": [ + 89, + 411, + 483, + 622 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {\\top} J ^ {\\top}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 628, + 482, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the Jacobian $J$ is used to linearly approximate projective transformations.", + "bbox": [ + 89, + 656, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pixel rendering procedure begins by pinpointing a pixel's location, represented as $p$ , within the image plane and then assess its closeness to the intersecting Gaussians. This step involves gauging the Gaussian's depth via applying a viewing transformation, denoted as $W$ [7]. This process forms an sorted sequence of Gaussians, labeled $N_{s}$ . Subsequently, a technique known as alpha compositing is employed to determine the pixel's color", + "bbox": [ + 89, + 686, + 483, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI = \\sum_ {i \\in N _ {s}} \\left(c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 814, + 482, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $c_{i}$ and $\\alpha_{i}^{\\prime}$ indicates the learned color and the opacity, respectively.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of our work is to obtain a Gaussian Splitting-based 3D representation of a static scene using a data stream from a moving event camera. Formally, we are given a sequence of $N$ events $e_k = (t_k, x_k, y_k, p_k)$ with corresponding camera poses $P_k \\in \\mathrm{SE}(3)$ denoted as $\\mathcal{E} := \\{(e_k, P_k)\\}_{k=1}^N$ . In practice, such dense pose information may not be available. However, it can be computed from sparse poses via interpolation. The interpolation to estimate event camera poses is explained in Sec. 5.1. From the posed event sequence, we aim to obtain a 3DGS-based scene representation $\\mathcal{G}$ that can be used for novel view synthesis of traditional rgb and grayscale images.", + "bbox": [ + 511, + 114, + 906, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key idea of our optimization approach is to model changes in logarithmic image intensity. That is, for a trajectory sequence over a given time interval $[a,b]$ and corresponding image intensity $I(t)\\in \\mathbb{R}^{H\\times W\\times 3}$ , our work models", + "bbox": [ + 511, + 297, + 906, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE (a, b) := \\int_ {a} ^ {b} \\log \\left(I ^ {\\prime} (t)\\right) \\mathrm {d} t. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 369, + 905, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For loss computation, we use both, event camera data and 3DGS to approximate this quantity. In the case of event cameras, this approximation can be obtained directly by creating an image of log intensity changes. Approximating Eq. (4) with 3DGS, however, requires additional modifications to the training process.", + "bbox": [ + 511, + 409, + 905, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Method Overview In its original formulation, 3D Gaussian Splitting uses a sequence of posed images as its training data. To make Gaussian Splitting applicable to event data, we introduce several modifications to the original method: First, we approximate Eq. (4) by accumulating events into an image representation $D \\in \\mathbb{R}^{H \\times W}$ using a random sub-trajectory of the given event stream $\\mathcal{E}$ (Sec. 4.1). This representation contains the sum of quantized log-intensity changes over the sub-trajectory. Second, we compute the corresponding log-intensity changes by rasterizing two views from the Gaussian scene representation (Sec. 4.2). For color cameras, this additionally requires a remosaicing step to account for the fact that the accumulated image representation of events only contains single color information for each pixel. To improve the 3DGS optimization process, we leverage an event-to-video guided structure from motion (Sec. 4.3). We also leverage cubic spline interpolation to more accurately estimate camera sub-trajectories as the rate at which the camera is tracked may be too coarse (Sec. 4.4). The full scene representation learning process is illustrated in Fig. 2.", + "bbox": [ + 511, + 520, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Event Accumulation", + "text_level": 1, + "bbox": [ + 511, + 848, + 707, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain an image of accumulated events, we first randomly sample start and endpoints $k_{\\mathrm{start}}, k_{\\mathrm{end}} \\in \\{1 \\dots N\\}$", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26878", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e4d59241728ca5cd360114d94235ead907adc107a7b1c2f034d3facfe022c479.jpg", + "image_caption": [ + "Figure 2. Overview of our 3D Gaussian Splitting training with moving event camera data. Event data streams from $t_{k_{start}}$ to $t_{k_{end}}$ are accumulated into $D$ , and distill the point priors from an event-to-video model. The log-difference image $\\hat{D}$ is obtained from 3D Gaussian Splitting and Rasterization at two camera view points. It is computed as in Eq. (6) and remosaicing is performed in case of a color event camera. The respective poses are estimated by cubic spline interpolation." + ], + "image_footnote": [], + "bbox": [ + 153, + 89, + 851, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from our trajectory sequence. The lengths of those sequences are chosen randomly with a maximum length ranging from $1\\%$ to $10\\%$ of the total number of events. This allows for considering scene information at different scales in order to have both global and local geometry information trained. Then, we obtain the entries of the accumulated log-intensity change image $D$ at pixel location $x,y$ as", + "bbox": [ + 88, + 488, + 480, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {x, y} := \\sum_ {\\substack {k \\in \\{k _ {\\text {start}}, \\dots , k _ {\\text {end}} \\} \\\\ x _ {k} = x, y _ {k} = y}} p _ {k} \\Delta . \\tag{5}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 603, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The resulting image serves as an approximation to Eq. (4), i.e. $D \\approx E(t_{\\mathrm{start}}, t_{\\mathrm{end}})$ . This method can easily be adapted to the case where the event-threshold $\\Delta$ is asymmetric, i.e., when it is different for positive and negative polarity events.", + "bbox": [ + 89, + 652, + 482, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Training View Generation and Remosaicing", + "text_level": 1, + "bbox": [ + 89, + 720, + 465, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate the log-difference image $\\hat{D} \\in \\mathbb{R}^{H \\times W}$ , we compute", + "bbox": [ + 89, + 742, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} := \\hat {L} _ {2} - \\hat {L} _ {1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 771, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which requires the remosaiced log images $L_{1}, L_{2} \\in \\mathbb{R}^{H \\times W}$ at the poses $P_{k_{\\mathrm{start}}}^{\\prime}$ and $P_{k_{\\mathrm{end}}}$ respectively. These poses are the same that were used for generating our target $D$ . To obtain $L_{1}$ and $L_{2}$ we first use the rasterizer to generate corresponding views $I_{1}, I_{2} \\in \\mathbb{R}^{H \\times W \\times 3}$ from the current Gaussian scene representation. Next, we perform a standard remosaicing operation because events are triggered per pixel", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "asynchronously, not allowing us to use conventional demosaicing methods [5, 29, 35, 36, 41, 44]", + "bbox": [ + 511, + 488, + 906, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {R e m o s a i c i n g}: \\mathbb {R} ^ {H \\times W \\times 3} \\rightarrow \\mathbb {R} ^ {H \\times W} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 525, + 903, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which reintroduces the Bayer RGB pattern. It consists of two steps: First, a set of color-channel specific matrices of size $2 \\times 2$ are Hadamard multiplied with each $2 \\times 2$ pixel block in the image. The channel specific matrices $R, G, B \\in \\mathbb{R}^{2 \\times 2}$ are given by", + "bbox": [ + 511, + 551, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR = \\left[ \\begin{array}{l l} 1 & 0 \\\\ 0 & 0 \\end{array} \\right], \\quad G = \\left[ \\begin{array}{l l} 0 & 1 \\\\ 1 & 0 \\end{array} \\right], \\quad B = \\left[ \\begin{array}{l l} 0 & 0 \\\\ 0 & 1 \\end{array} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 633, + 903, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The thus modified channels have non-zero entries at non-overlapping pixel locations. From this, the single-channel remosaiced image is obtained by addition across the channels. Finally, entry-wise logarithm computation yields the desired images $\\hat{L}_1, \\hat{L}_2$ and thus $\\hat{D}$ .", + "bbox": [ + 511, + 674, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Event-to-Video Guided Initialization", + "text_level": 1, + "bbox": [ + 511, + 757, + 830, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event cameras capture only the pixels that indicate changes in brightness. This results in sparse information that is often insufficient for effective supervision. Thus, reconstructing 3D information, including background details, solely from event data is exceedingly challenging. To simplify the optimization process we propose leveraging the prior knowledge encoded in an event-to-video model. We use this pretrained model to generate images from event streams that", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26879", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can subsequently be processed by Structure from Motion (SfM) techniques. Although the generated images often contain significant noise and do not accurately reproduce the underlying RGB image values, they retain substantial texture and background information. This information is sufficient to provide initial positions for the Gaussians, improve the 3DGS optimization process and lead to faster convergence compared to basic initialization with random initial points.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Event Camera Trajectory Interpolation", + "text_level": 1, + "bbox": [ + 89, + 234, + 429, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The camera poses, $P_{k_{\\mathrm{start}}}$ and $P_{k_{\\mathrm{end}}}$ , used to obtain the log images $\\hat{L}_1$ and $\\hat{L}_2$ in Eq. (6), are in most cases not given a priori. This is because event camera poses are typically obtained from a tracking or motion estimation system that operates at a fixed rate, which may not be aligned with the rate of event accumulation. Thus, to obtain an estimate of the complete trajectory of the event camera, we adopt an efficient trajectory interpolation method for camera poses. Our work utilizes Cubic Spline Interpolation and Spherical Cubic Spline Interpolation. These generate a smooth path by interpolating between discrete command points using third-degree polynomials. This interpolation technique closely approximates real-world camera motion by maintaining non-linear continuity in both velocity and acceleration of camera movement.", + "bbox": [ + 89, + 257, + 482, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 494, + 235, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Gaussian Splitting is updated using a loss that combines a $\\mathcal{L}_1$ reconstruction term and a perceptual term $\\mathcal{L}_{\\mathrm{SSIM}}$ that is based on the Structural Similarity Index Measure (SSIM) [67]", + "bbox": [ + 89, + 516, + 482, + 578 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {S S I M}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 588, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fully utilize real-world event camera data, undistortion operation is performed as the accumulated image $D$ is computed. We account for this by only incurring the loss where events were actually accumulated. This allows supervising the 3D representation more efficient.", + "bbox": [ + 89, + 614, + 482, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 702, + 220, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 727, + 192, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our evaluation of 3D Gaussian Splating from a moving event camera encompasses both synthetic and real-world scenes. Our method assumes that we can access the intrinsic camera matrix, its lens distortion parameters, and the camera's constant-rate poses, sampled at a high frequency to ensure precise interpolation for the event camera at any chosen time point. Estimation of camera poses at any given timestep is crucial for precise scene reconstruction from event data, as events can asynchronously occur at arbitrary timesteps between the constant-rate sampled camera poses.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1. Synthetic Scenes", + "text_level": 1, + "bbox": [ + 511, + 90, + 674, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The synthetic dataset used in our experiments consists of event data streams obtained from ESIM [55], initially introduced in [42], encompassing seven synthetic scenes with a white background. These simulated event data streams feature a diverse range of photorealistic objects with complex structures. The event camera in these synthetic scenes moves around the objects, providing diverse geometric and lighting conditions. This setup allows for a comprehensive evaluation of the robustness of our method against the baselines in a controlled environment.", + "bbox": [ + 511, + 109, + 903, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.2. Real-World Scenes", + "text_level": 1, + "bbox": [ + 511, + 270, + 691, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experiments on real scenes were conducted using subsets of the EDS [24], and TUM-VIE [30] datasets. From EDS, we utilize the sequences 03_rocket-earth DARK, 07_ziggy_and_fuzz hdr, 08_peanuts_runting, 11_all Characters and 13_airplane. From TUM-VIE, we utilize the sequences mocap-1d-trans and mocap-desk2. These sequences primarily capture various objects in 360-degree scenes and on a desk from a forward-facing perspective. These real sequences are relatively well-suited for novel view synthesis and were recorded using the high-resolution event sensor, the Prophesee Gen 3 in the EDS and the Prophesee Gen 4 in the TUM-VIE dataset, respectively.", + "bbox": [ + 511, + 290, + 903, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Baselines", + "text_level": 1, + "bbox": [ + 511, + 496, + 620, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We benchmarked our method against Robust-e-NeRF and E2VID [56] + 3D Gaussian Splatting (E2VID+3DGS). Robust-e-NeRF incorporates InstantNGP, which is significantly faster compared to traditional NeRF methods, but still falls short compared to 3DGS. E2VID generates RGB images from an event data stream, followed by feeding the images into original 3D Gaussian Splatting.", + "bbox": [ + 511, + 518, + 903, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Metrics", + "text_level": 1, + "bbox": [ + 511, + 636, + 609, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilized commonly used metrics to evaluate novel view synthesis, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [79]) with VGG16, to comprehensively assess the quality of rendered images. It's worth noting that SSIM and LPIPS are perceptual metrics introduced to align more closely with human preference, with LPIPS specifically tailored to mimic human perception. These metrics address issues associated with PSNR, such as favoring blurry images. Additionally, we measured and compared the rendering speeds of our method, Robust-e-NeRF, and E2VID+3DGS.", + "bbox": [ + 511, + 657, + 903, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since event cameras capture variations in log-radiance rather than absolute log-radiance values, the predicted intensity $I(t)$ from the 3D Gaussian Splatting has an unknown offset. To rectify this limitation, a linear color transforma", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26880", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff41e83e35d3082964b1d049335c78f10590a2c016dda0f0001fd765c3787db6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 186, + 88, + 346, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ea78d5915b7cfc9becea1d7f466e433b03771de3e1406ed67bd9baaabb9895a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 88, + 506, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebfbb3e6d9f319ac30322afd9e0c2b6485d5b7cbcd86723fb4600dd95d4f03bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 88, + 658, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c4708c3bf0d52f5e7b0c53cf19bbbf1d0d00339385252570f44fe1e5e99e64c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 89, + 812, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8df9fa988fa5ff3c9d93ac20a141a6bf908da482db1ef55ae34a8ed8c5d8d6bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 186, + 224, + 344, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b7a59d98c2daf99711123f734a20545a40404b012b4e2df05b4b091e80cf054.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 232, + 504, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/124daa59c57e4ac9db3e436aabbd67e938035323e55d3d7390fdbcfd154ba6ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 233, + 658, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/280db115ed57b244ec055245155afca539b8340f4350bb561af760d014db1403.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 233, + 812, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2cb8279c9c93689ae7441c7b8cf9afdf2b3db86f23518f94e3c02bf77048385f.jpg", + "image_caption": [ + "Figure 3. Generated Synthetic images comparing our work, event-based NeRF, and E2VID+3DGS qualitatively, with rendering times shown at the bottom of each image. Our work and Robust-e-NeRF outperform others in most scenes. In some cases, e.g., the drums scene, our work performs better and achieves significantly faster rendering times than conventional Robust-e-NeRF composed of InstantNGP." + ], + "image_footnote": [], + "bbox": [ + 207, + 345, + 331, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e1e0d057cc4fc9f9336589f445bbedeb41595c96c9d4d73ee9e428154f8922e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 345, + 503, + 465 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76b81c8e4856c97ec9051b464f57ea7c7df7ca5e91572577eb07ba04b643a458.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 345, + 658, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73c0744d31d88dad680de59f351631504f90c181dd6a77b373df1785bcad338b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 345, + 810, + 467 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d925b4fec6fe0113c2aef691e6249af3b68d560bdfcdabb7bc29d507418be5c4.jpg", + "table_caption": [ + "Table 1. Generated synthetic scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively. We supervised our model with randomly initialized points." + ], + "table_footnote": [], + "table_body": "
MetricMethodSynthetic SceneMean
chairdrumsficushotdoglegomaterialsmic
PSNR ↑E2VID + 3DGS21.3919.8619.9015.5518.1720.0820.1019.29
Robust e-NeRF30.2423.1530.7128.0727.3424.9832.8728.19
Ours28.6925.8129.9022.9129.2227.1633.2728.14
SSIM ↑E2VID + 3DGS0.9340.9150.9220.8970.8950.9010.9570.917
Robust e-NeRF0.9580.8970.9710.9530.9340.9230.9810.945
Ours0.9530.9470.9660.9400.9450.9360.9860.953
LPIPS ↓E2VID + 3DGS0.0760.0940.1080.2080.1450.1250.0690.118
Robust e-NeRF0.0400.0910.0220.0950.0740.0520.0290.057
Ours0.0470.0520.0280.0980.0550.0600.0150.051
", + "bbox": [ + 150, + 574, + 844, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion is designed to adjust our predictions in the logarithmic domain [42, 59]. This transformation is both necessary and adequately effective for aligning our predictions with the reference data. It ensures that the reconstructed intensity values are properly calibrated and aligned with the observed event data.", + "bbox": [ + 89, + 801, + 483, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Results and Evaluation", + "text_level": 1, + "bbox": [ + 513, + 800, + 728, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After completing the training of our model on each scene using event sequence data, the model acquires a set of 3D Gaussian ellipsoids representing the scene. By rasterizing these Gaussians, we are able to render various arbitrary views of each scene in real-time. The renderings of novel", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26881", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a96c47fa963139724dad03565260fb3b2b24182a37e492b97be4bfc4df3960a1.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons of the images generated by our method, event-based NeRF and E2VID+3DGS show that our technique appears to recover details in 5 real scenes. First three rows(07_ziggy_and_fuzz_hdr, 08_peanuts_runting and 11_all Characters) are EDS dataset and last two rows are TUM-VIE dataset(mocap-1d-trans and mocap-desk2)." + ], + "image_footnote": [], + "bbox": [ + 212, + 87, + 784, + 570 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f263ba7060235efb97aa7892dbed32c35ceb062e9f95c65b0f89cba0ce5e1d1.jpg", + "table_caption": [ + "Table 2. Real scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively." + ], + "table_footnote": [], + "table_body": "
MetricMethodReal SceneMean
0307081113
PSNR ↑E2VID + 3DGS15.6715.0514.0313.8318.9615.51
Robust e-NeRF19.1914.7814.7514.4318.1016.25
Ours20.7819.1417.5317.7919.0518.86
SSIM ↑E2VID + 3DGS0.7160.6890.6420.6910.7230.692
Robust e-NeRF0.8460.8150.7350.5690.7290.739
Ours0.8350.8160.7450.7890.7740.792
LPIPS ↓E2VID + 3DGS0.2660.3780.4020.4150.4150.375
Robust e-NeRF0.3240.4760.5670.7000.6500.543
Ours0.2390.3510.4240.3910.4070.363
", + "bbox": [ + 184, + 660, + 808, + 863 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26882", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "views, obtained from held-out views or test views provided by the datasets, are then used for evaluations.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.1. Synthetic Scenes", + "text_level": 1, + "bbox": [ + 89, + 128, + 253, + 143 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our methodology on a synthetic dataset using the metrics described in Sec. 5.3. Quantitative results are shown in Tab. 1, and visual samples and rendering times in Fig. 3. Rendering times are provided in milliseconds (ms). Our method achieves higher average SSIM and compared to previous works. A high SSIM value indicates structural similarity between generated and reference images, underscoring our model's accuracy in capturing essential scene geometry. Additionally, a low LPIPS score reflects perceptual quality, showing the produced images are both realistic and visually engaging. The elevated scores demonstrate that our novel view synthesis model faithfully reproduces the visual and perceptual attributes of the original scene. Moreover, the rendering speed of both our method and E2VID+3DGS significantly surpass that of Robust-e-NeRF. Therefore, our work maintains superior quality and speed compared to Robust-e-NeRF, highlighting its effectiveness in both fidelity and performance.", + "bbox": [ + 89, + 148, + 482, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the quality of rendered images remains high in scenes featuring drums and microphones, achieving novel view synthesis without sacrificing quality in other scenes as well.", + "bbox": [ + 89, + 420, + 482, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.2. Real Scenes", + "text_level": 1, + "bbox": [ + 89, + 487, + 218, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The EDS dataset is the only one available for quantifying the performance of novel view synthesis. Therefore, we primarily conducted quantitative and qualitative assessments, as shown Tab. 2 and Fig. 4. Our method successfully recovers fine details in all sequences and contrasts in the mocap-1d-trans image, particularly. This is particularly noticeable in the images containing the checkerboard and the cover of a book. Additionally, it effectively smooths uniform areas such as the checkerboard and the desk surface in the mocap-1d-trans and mocap-desk2. This smoothness likely results from the accumulation of event streams, which may have enhanced the signal-to-noise ratio due to the aggregation process. The visible artifacts near the borders of all synthesized views are attributed to the relatively narrow field of view of the event cameras.", + "bbox": [ + 89, + 506, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablations", + "text_level": 1, + "bbox": [ + 89, + 742, + 202, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of event-to-video guided initialization and camera trajectory estimation with cubic spline interpolation, we conducted ablation experiments by separately adding 1) event-to-video guided initialization and 2) both the initialization and cubic spline interpolation. As shown in Tab. 3, event-to-video guided initialization outperforms random initialization. Furthermore, Fig. 5 shows that our initialization approach results in better image quality (e.g. backgrounds) compared to random initialization.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7c1b53cc01b2c9a582b59909a01f73bb1645c23ca9fbba174908b95568e6c17c.jpg", + "table_caption": [ + "Table 3. Ablation results for the event-to-video guided initialization and cubic spline interpolation." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
random init18.020.7670.397
guided init18.750.7880.359
guided init + cubic18.860.7920.363
", + "bbox": [ + 529, + 128, + 887, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e7e67edf5eaab2048813d2c5381f29caadb59159351eacd977bf1d6e59f1367.jpg", + "image_caption": [ + "Random init" + ], + "image_footnote": [], + "bbox": [ + 516, + 227, + 707, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7826f7c045ab390dc83d260299516fa74f494d68ddef06c024bdce3cda7ae09.jpg", + "image_caption": [ + "Event-to-Video Guided init", + "Figure 5. Qualitative comparisons of the images generated by random initialization and event-to-video guided initialization are shown in 07_ziggy_and_fuzz hdr scene in EDS dataset." + ], + "image_footnote": [], + "bbox": [ + 707, + 227, + 901, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adding cubic spline interpolation further enhances performance.", + "bbox": [ + 511, + 421, + 903, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Limitations", + "text_level": 1, + "bbox": [ + 513, + 460, + 638, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method, though effective, has some limitations. It is designed exclusively for novel view synthesis in static scenes and cannot handle scenarios with dynamic moving objects, which we consider to be an exciting avenue for future research. Additionally, our approach trains 3D Gaussian Splatting on event-accumulated images between two viewpoints, representing relative intensity images. Consequently, similar to other event-based view synthesis methods, the model cannot directly estimate absolute intensity images and requires a linear transformation with evaluation data as a reference. This linear transformation is required only during inference and does not impact the training.", + "bbox": [ + 511, + 483, + 906, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 679, + 782, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a novel method for 3D Gaussian Splatting from moving event cameras. Our approach achieves the higher visual fidelity and better performance of novel view synthesis in static scenes. We distill geometric priors from the event-to-video model for initialization and leverage effective camera trajectory interpolation. Thus, EventSplat yields improved metrics compared to existing event-based NeRF approaches. Moreover, the incorporation of 3D Gaussian Splatting enables our model to render scenes at significantly higher frame rates. This enhances the practicality of real-time novel view synthesis with event cameras. Our work is thus particularly useful in scenarios where conventional imaging techniques are inadequate.", + "bbox": [ + 511, + 704, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26883", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work is in part supported by Sony. We would also like to thank Ziyi Wu, Yash Kant, and Umangi Jain, for valuable discussions and support.", + "bbox": [ + 89, + 114, + 483, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. ICCV, 2021. 1, 2", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2", + "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In ICCV, 2023. 1", + "[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. TNNLS, 2013. 2", + "[5] Hong Cao and Alex C Kot. Accurate detection of demosaicing regularity for digital image forensics. IEEE Trans. Inf. Forensics Secur, 2009. 4", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV. Springer, 2022. 2", + "[7] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 2, 3", + "[8] Jiaben Chen, Yichen Zhu, Dongze Lian, Jiaqi Yang, Yifu Wang, Renrui Zhang, Xinhang Liu, Shenhan Qian, Laurent Kneip, and Shenghua Gao. Revisiting Event-Based Video Frame Interpolation. In IROS, 2023. 1", + "[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In arXiv, 2022. 1, 2", + "[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3D using Gaussian Splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3", + "[11] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), 2024. 2", + "[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2", + "[13] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Sci. Robot, 2020. 1", + "[14] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. TPAMI, 2017. 2", + "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In CVPR, 2018. 1" + ], + "bbox": [ + 93, + 198, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. PAMI, 2022. 1, 2, 3", + "[17] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. FastNeRF: High-Fidelity Neural Rendering at 200FPS. In ICCV, 2021. 2", + "[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Asynchronous, photometric feature tracking using events and frames. In ECCV, 2018. 2", + "[19] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019.", + "[20] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. IJCV, 2020.", + "[21] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 2", + "[22] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 3DV, 2021. 2", + "[23] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2", + "[24] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5", + "[25] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-NeRF: Event based neural radiance field. In WACV, 2023. 1, 2", + "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ToG, 2023. 1, 2, 3, 12", + "[27] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In CVPR, 2016. 1", + "[28] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, 2021. 2", + "[29] Ron Kimmel. Demosaicing: Image reconstruction from color CCD samples. IEEE Trans. Image Process, 1999. 4", + "[30] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM Stereo Visual-Inertial Event Dataset. In IROS, 2021. 5", + "[31] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. RA-L, 2023. 1, 2", + "[32] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. AdaNeRF: Adaptive sampling for real-time rendering of neural radiance fields. In ECCV, 2022. 2", + "[33] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient Sphere-Based Neural Rendering. In CVPR, 2021. 1" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26884", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 2", + "[35] Xin Li. Demosaicing by successive approximation. IEEE Trans. Image Process, 2005. 4", + "[36] Xin Li, Bahadir Gunturk, and Lei Zhang. Image demosaic-ing: A systematic survey. In VCIP. SPIE, 2008. 4", + "[37] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 db $15\\mu s$ latency asynchronous temporal contrast vision sensor. JSSC, 2008. 1", + "[38] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning Event-Driven Video Deblurring and Interpolation. In ECCV, 2020. 1", + "[39] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. BACON: Band-limited coordinate networks for multiscale scene representation. In CVPR, 2022. 2", + "[40] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2", + "[41] Philippe Longere, Xuemei Zhang, Peter B Delahunt, and David H Brainard. Perceptual assessment of demosaicing algorithm performance. Proceedings of the IEEE, 2002. 4", + "[42] Weng Fei Low and Gim Hee Lee. Robust e-NeRF: NeRF from Sparse & Noisy Events under Non-Uniform Motion, 2023. 1, 2, 5, 6, 12", + "[43] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V. Sander. Deblur-NeRF: Neural Radiance Fields From Blurry Images. In CVPR, 2022. 1", + "[44] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, 2004. 4", + "[45] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2", + "[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: representing scenes as neural radiance fields for view synthesis. ECCV, 2021. 1, 2", + "[47] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. NeRF in the Dark: High Dynamic Range View Synthesis From Noisy Raw Images. In CVPR, 2022. 1", + "[48] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G. Derpanis, Jonathan Kelly, Marcus A. Brubaker, Igor Gilitschenski, and Alex Levinshtein. SPIn-NeRF: Multiview Segmentation and Perceptual Inpainting With Neural Radiance Fields. In CVPR, 2023.", + "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 1, 2", + "[50] Marc Osswald, Sio-Hoi Ieng, Ryad Benosman, and Giacomo Indiveri. A spiking neural network model of 3d perception for event-based neuromorphic stereo vision systems. Sci. Rep., 2017. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 1", + "[52] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurlPS, 2020. 2", + "[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion, 2022. 2, 3", + "[54] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. RA-L, 2016. 2", + "[55] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In CoRL. PMLR, 2018. 5", + "[56] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. PAMI, 2019. 5", + "[57] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In ICCV, 2021. 2", + "[58] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. In arXiv, 2023. 2", + "[59] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. EventNeRF: Neural radiance fields from a single colour event camera. In CVPR, 2023. 1, 2, 6, 12", + "[60] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[61] Amos Sironi, Manuele Brambilla, Nicolas Bourdis, Xavier Lagorce, and Ryad Benosman. HATS: Histograms of averaged time surfaces for robust event-based object classification. In CVPR, 2018. 2", + "[62] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In CVPR, 2021. 2", + "[63] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, 2021. 1", + "[64] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-Based Frame Interpolation With Parametric Non-Linear Flow and Multi-Scale Fusion. In CVPR, 2022. 1", + "[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26885", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2", + "[67] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale Structural Similarity for Image Quality Assessment. In ACSSC, 2003. 5", + "[68] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2", + "[69] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528, 2023. 1", + "[70] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. arXiv preprint arXiv:2407.11343, 2024. 2, 12, 14", + "[71] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. arXiv preprint arXiv:2311.17286, 2023. 2", + "[72] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 2", + "[73] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In ICLR, 2024. 2", + "[74] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In SIGGRAPH, 2023. 1", + "[75] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2", + "[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2", + "[77] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2", + "[78] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S. Ren. Training weakly supervised video frame interpolation with events. In ICCV, 2021. 1", + "[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In CVPR, 2018. 5", + "[80] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Yi Zhou, Guillermo Gallego, Henri Rebecq, Laurent Kneip, Hongdong Li, and Davide Scaramuzza. Semi-Dense 3D Reconstruction with a Stereo Event Camera. In ECCV, 2018. 2", + "[82] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. T-RO, 2021. 2", + "[83] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In RSS, 2018. 2", + "[84] Baozhou Zhu, Peter Hofstee, Johan Peltenburg, Jinho Lee, and Zaid Alars. AutoReCon: Neural Architecture Search-based Reconstruction for Data-free Compression, 2021. 1", + "[85] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. EWA volume splatting. In VIS, 2001. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 290 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26886", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_model.json b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f2eb9e694581a750ca9b318af95a92bea65f1f2 --- /dev/null +++ b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering" + }, + { + "type": "text", + "bbox": [ + 0.309, + 0.203, + 0.434, + 0.222 + ], + "angle": 0, + "content": "Toshiya Yura1,2" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.226, + 0.468, + 0.239 + ], + "angle": 0, + "content": "toshiya.yura@sony.com" + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.204, + 0.694, + 0.221 + ], + "angle": 0, + "content": "Ashkan Mirzaei" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.226, + 0.72, + 0.238 + ], + "angle": 0, + "content": "ashkan@cs.toronto.edu" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.246, + 0.575, + 0.265 + ], + "angle": 0, + "content": "Igor Gilitschenski²" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.268, + 0.624, + 0.281 + ], + "angle": 0, + "content": "gilitschenski@cs.toronto.edu" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.282, + 0.769, + 0.3 + ], + "angle": 0, + "content": "1Sony Semiconductor Solutions Corporation, 2University of Toronto" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.335, + 0.327, + 0.351 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.368, + 0.486, + 0.626 + ], + "angle": 0, + "content": "We introduce a method for using event camera data in novel view synthesis via Gaussian Splatting. Event cameras offer exceptional temporal resolution and a high dynamic range. Leveraging these capabilities allows us to effectively address the novel view synthesis challenge in the presence of fast camera motion. For initialization of the optimization process, our approach uses prior knowledge encoded in an event-to-video model. We also use spline interpolation for obtaining high quality poses along the event camera trajectory. This enhances the reconstruction quality from fast-moving cameras while overcoming the computational limitations traditionally associated with event-based Neural Radiance Field (NeRF) methods. Our experimental evaluation demonstrates that our results achieve higher visual fidelity and better performance than existing event-based NeRF approaches while being an order of magnitude faster to render." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.661, + 0.222, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Event cameras are a vision modality that draws strong inspiration from biological vision and neuromorphic design. In these sensors, each pixel operates independently by triggering events upon a log-intensity change above a pre-defined threshold [16, 37]. This results in an asynchronous data stream as opposed to capturing absolute intensity values at fixed intervals in traditional cameras. Event cameras offer low latency, high temporal resolution, and a high dynamic range while also ensuring minimal energy consumption [8, 38]. Consequently, they are particularly suitable for dynamic and challenging lighting conditions, as well as environments with rapid camera motions where classical cameras may exhibit strong motion blur [38, 63, 64, 78]. These advantages have motivated the adoption of event cameras in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.905, + 0.366 + ], + "angle": 0, + "content": "a wide range of applications covering most of the traditional computer vision tasks [13, 15, 27, 50, 51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.534 + ], + "angle": 0, + "content": "Recently, novel view synthesis has seen significant advances. This progress has been particularly driven by the emergence and development of Neural Radiance Fields (NeRF) [1-3, 33, 46, 74]. Recent works have led to NeRF improvements in terms of rendering speed, and adoption of NeRFs in 3D editing and other downstream applications [9, 43, 47-49, 84]. 3D Gaussian Splitting (3DGS) [26, 69] achieves real-time rendering by replacing the underlying representation with 3D Gaussians and using fast differentiable rasterization. The substantial speed improvement of 3DGS is achieved without compromising the visual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.672 + ], + "angle": 0, + "content": "The success of NeRFs in view-synthesis has led to their adoption in applications with event-based cameras [25, 31, 42, 59]. Conventional NeRF approaches struggle with challenges, such as 3D reconstruction in environments with rapidly changing lighting conditions or high-speed objects due to RGB camera limitations. The integration of NeRFs with event-based cameras leverages their unique advantages and incorporates them into the NeRF framework to address the challenges that conventional NeRF approaches face." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.763 + ], + "angle": 0, + "content": "However, one crucial area of research focuses on expediting the rendering processes to fully unlock the potential of event-based NeRFs. Leveraging the advantages of 3D Gaussian Splatting with events requires adapting the rasterization process to allow supervision with events or accumulated events rather than RGB images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.764, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In summary, this paper contributes a method for 3D Gaussian Splatting from event data by leveraging the following technical ideas:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.811, + 0.905, + 0.841 + ], + "angle": 0, + "content": "1) An integration of event accumulation with the optimization process of 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.841, + 0.905, + 0.871 + ], + "angle": 0, + "content": "2) An event-to-video guided Structure from Motion (SfM) approach for initializing the 3DGS optimization process." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3) Use of cubic spline trajectory interpolation to assign camera poses to events at high rates." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.089, + 0.803, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.261, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Figure 1. EventSplat derives 3D representations of scenes in the form of 3D Gaussians from event data, enabling fast real-time renderings of scenes captured with event cameras via rasterization of 3D Gaussians. The use of event-based input is particularly advantageous when traditional RGB cameras fail due to various reasons including poor lighting conditions or motion blur from fast-moving cameras." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.33, + 0.488, + 0.455 + ], + "angle": 0, + "content": "Our evaluations show that these design choices achieve state-of-the-art event-based novel view synthesis performance while benefiting from fast rendering speeds. Thus, our method represents an advancement in addressing the speed constraints associated with event-based NeRFs. It opens up new possibilities for real-time novel view synthesis for environments where traditional imaging techniques struggle." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.235, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.494, + 0.427, + 0.511 + ], + "angle": 0, + "content": "2.1. Event Cameras for Vision Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.486, + 0.716 + ], + "angle": 0, + "content": "Motivated by the low energy consumption, high dynamic range, and high asynchronous temporal resolution of event cameras, event-based vision has recently garnered significant attention [16]. These advantageous properties have led to promising advancements in downstream tasks such as object tracking, detection, and recognition [14, 18-21, 28, 34, 52, 61, 71], as well as optical flow estimation [4, 22, 83]. However, being a novel data format, event data necessitates specialized architectures and algorithms to be effectively processed and utilized in various applications. In this work, we present a method to exploit event data for real-time novel view synthesis, a scenario that has not been addressed previously." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.727, + 0.445, + 0.743 + ], + "angle": 0, + "content": "2.2. Volumetric Rendering for View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs)[46] have been the state-of-the-art in view synthesis, enabled by volumetric rendering and neural scene representations. Significant research has been conducted to improve NeRFs in terms of training and rendering speed [6, 9, 23, 32, 49, 58, 60, 76] and visual quality [1, 2, 12, 39, 65]. However, due to the expensive volumetric sampling process, NeRF-based methods have generally been unable to render high-quality views in real time. Recently, 3D Gaussian Splitting (3DGS) [7, 10, 26, 53, 85] proposed a rasterization-based approach for differentiably" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.909, + 0.483 + ], + "angle": 0, + "content": "rendering scenes represented as sets of 3D Gaussian ellipsoids. This method achieves real-time renderings while preserving the quality of state-of-the-art NeRF-based methods. Since the introduction of 3DGS, real-time capabilities of 3DGS have resulted in rapid adoption of Gaussians in several downstream tasks including dynamic scene reconstruction [73], 3D asset generation [75], SLAM [45], and 4D object generation [40]. For view synthesis from event data, we leverage 3DGS and adapt it to be trainable using event sequences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.489, + 0.833, + 0.505 + ], + "angle": 0, + "content": "2.3. Novel View Synthesis via Event Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.909, + 0.815 + ], + "angle": 0, + "content": "Motivated by the unique benefits of event-based cameras, such as high dynamic range, low latency, and high temporal resolution, there has been significant research interest in adopting event data in applications like 3D reconstruction and novel view synthesis in cases where traditional RGB cameras underperform [16, 54, 81, 82]. Recent methods explore using NeRFs with event-vision supervision [25, 31, 42, 59] to reconstruct scenes and render them from novel views. While capable of generating high-quality renderings, NeRF-based methods are generally slow to render due to expensive sampling requirements [6, 17, 57, 62]. On the other hand, 3DGS requires RGB data for supervision, and adapting 3DGS for scenarios where the supervision signal comes from an event-based camera has been challenging [11, 66, 68, 70, 72, 77, 80]. In contrast to previous event-based view synthesis methods, we enable the use of event data to supervise 3DGS, allowing for real-time renderings and state-of-the-art quality using only event data, supported by event-based initial point clouds and effective camera trajectory interpolation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.642, + 0.841 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.704, + 0.864 + ], + "angle": 0, + "content": "3.1.Event Camera Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "While traditional cameras output an intensity image \\(I\\), event cameras model changes in the log intensity \\(L = \\log (I)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26877" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "of the current camera view. Similar to traditional cameras, they have a pixel array. However, in contrast to traditional cameras, each pixel acts as an independent sensor and triggers an event whenever the difference between the current log-intensity \\( l_{x,y}(t) \\) and the log-intensity at the time of the most recent event \\( l_{x,y}(t_{\\mathrm{recent}}) \\) surpasses a given threshold \\( \\Delta > 0 \\), i.e. when" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.206, + 0.483, + 0.224 + ], + "angle": 0, + "content": "\\[\nl _ {x, y} (t) - l _ {x, y} \\left(t _ {\\text {r e c e n t}}\\right) > \\Delta . \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.261 + ], + "angle": 0, + "content": "This threshold is also known as contrast sensitivity [16]. It can be different for positive and negative polarity changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.382 + ], + "angle": 0, + "content": "Formally, each event is a tuple \\( e := (t, x, y, p) \\) where \\( t \\in \\mathbb{R} \\) is a timestamp at which it was triggered, \\( (x, y) \\in \\mathbb{Z}^2 \\) are its pixel coordinates, and \\( p \\in \\{-1, 1\\} \\) is the event's polarity indicating whether the log-intensity increased or decreased. Similar to traditional color cameras, a Bayer RGB pattern is used to obtain color information in color event cameras. In such sensors, each individual event provides information only for one color channel." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.301, + 0.406 + ], + "angle": 0, + "content": "3.2. 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.623 + ], + "angle": 0, + "content": "To enable novel view synthesis, 3DGS learns a scene representation that consists of a set of ellipsoids with view-dependent color information encoded as spherical harmonics [7, 10, 26, 53, 85]. The essence of the method revolves around transforming optimized 3D Gaussians into 2D imagery based on predetermined camera positions. Subsequently the approach calculates pixel values accordingly by arranging and rasterization [7, 10, 26, 53, 85]. The transformation process involves mapping 3D Gaussian distributions, shaped like ellipsoids, onto a two-dimensional plane as ellipses for the purpose of rendering. This is achieved by applying a specific viewing transformation, denoted by \\( W \\), along with a 3D covariance matrix \\( \\Sigma \\). The outcome is the derivation of a 2D covariance matrix \\( \\Sigma' \\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.63, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {\\top} J ^ {\\top}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.687 + ], + "angle": 0, + "content": "where the Jacobian \\( J \\) is used to linearly approximate projective transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.809 + ], + "angle": 0, + "content": "The pixel rendering procedure begins by pinpointing a pixel's location, represented as \\( p \\), within the image plane and then assess its closeness to the intersecting Gaussians. This step involves gauging the Gaussian's depth via applying a viewing transformation, denoted as \\( W \\) [7]. This process forms an sorted sequence of Gaussians, labeled \\( N_{s} \\). Subsequently, a technique known as alpha compositing is employed to determine the pixel's color" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.815, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\nI = \\sum_ {i \\in N _ {s}} \\left(c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Here, \\(c_{i}\\) and \\(\\alpha_{i}^{\\prime}\\) indicates the learned color and the opacity, respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.606, + 0.106 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.116, + 0.907, + 0.297 + ], + "angle": 0, + "content": "The goal of our work is to obtain a Gaussian Splitting-based 3D representation of a static scene using a data stream from a moving event camera. Formally, we are given a sequence of \\(N\\) events \\(e_k = (t_k, x_k, y_k, p_k)\\) with corresponding camera poses \\(P_k \\in \\mathrm{SE}(3)\\) denoted as \\(\\mathcal{E} := \\{(e_k, P_k)\\}_{k=1}^N\\). In practice, such dense pose information may not be available. However, it can be computed from sparse poses via interpolation. The interpolation to estimate event camera poses is explained in Sec. 5.1. From the posed event sequence, we aim to obtain a 3DGS-based scene representation \\(\\mathcal{G}\\) that can be used for novel view synthesis of traditional rgb and grayscale images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.37 + ], + "angle": 0, + "content": "The key idea of our optimization approach is to model changes in logarithmic image intensity. That is, for a trajectory sequence over a given time interval \\([a,b]\\) and corresponding image intensity \\(I(t)\\in \\mathbb{R}^{H\\times W\\times 3}\\), our work models" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.371, + 0.906, + 0.405 + ], + "angle": 0, + "content": "\\[\nE (a, b) := \\int_ {a} ^ {b} \\log \\left(I ^ {\\prime} (t)\\right) \\mathrm {d} t. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.502 + ], + "angle": 0, + "content": "For loss computation, we use both, event camera data and 3DGS to approximate this quantity. In the case of event cameras, this approximation can be obtained directly by creating an image of log intensity changes. Approximating Eq. (4) with 3DGS, however, requires additional modifications to the training process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Method Overview In its original formulation, 3D Gaussian Splitting uses a sequence of posed images as its training data. To make Gaussian Splitting applicable to event data, we introduce several modifications to the original method: First, we approximate Eq. (4) by accumulating events into an image representation \\( D \\in \\mathbb{R}^{H \\times W} \\) using a random sub-trajectory of the given event stream \\( \\mathcal{E} \\) (Sec. 4.1). This representation contains the sum of quantized log-intensity changes over the sub-trajectory. Second, we compute the corresponding log-intensity changes by rasterizing two views from the Gaussian scene representation (Sec. 4.2). For color cameras, this additionally requires a remosaicing step to account for the fact that the accumulated image representation of events only contains single color information for each pixel. To improve the 3DGS optimization process, we leverage an event-to-video guided structure from motion (Sec. 4.3). We also leverage cubic spline interpolation to more accurately estimate camera sub-trajectories as the rate at which the camera is tracked may be too coarse (Sec. 4.4). The full scene representation learning process is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.708, + 0.863 + ], + "angle": 0, + "content": "4.1. Event Accumulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "To obtain an image of accumulated events, we first randomly sample start and endpoints \\( k_{\\mathrm{start}}, k_{\\mathrm{end}} \\in \\{1 \\dots N\\} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26878" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.09, + 0.852, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.407, + 0.908, + 0.465 + ], + "angle": 0, + "content": "Figure 2. Overview of our 3D Gaussian Splitting training with moving event camera data. Event data streams from \\( t_{k_{start}} \\) to \\( t_{k_{end}} \\) are accumulated into \\( D \\), and distill the point priors from an event-to-video model. The log-difference image \\( \\hat{D} \\) is obtained from 3D Gaussian Splitting and Rasterization at two camera view points. It is computed as in Eq. (6) and remosaicing is performed in case of a color event camera. The respective poses are estimated by cubic spline interpolation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.489, + 0.482, + 0.595 + ], + "angle": 0, + "content": "from our trajectory sequence. The lengths of those sequences are chosen randomly with a maximum length ranging from \\(1\\%\\) to \\(10\\%\\) of the total number of events. This allows for considering scene information at different scales in order to have both global and local geometry information trained. Then, we obtain the entries of the accumulated log-intensity change image \\(D\\) at pixel location \\(x,y\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.604, + 0.483, + 0.645 + ], + "angle": 0, + "content": "\\[\nD _ {x, y} := \\sum_ {\\substack {k \\in \\{k _ {\\text {start}}, \\dots , k _ {\\text {end}} \\} \\\\ x _ {k} = x, y _ {k} = y}} p _ {k} \\Delta . \\tag{5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.714 + ], + "angle": 0, + "content": "The resulting image serves as an approximation to Eq. (4), i.e. \\( D \\approx E(t_{\\mathrm{start}}, t_{\\mathrm{end}}) \\). This method can easily be adapted to the case where the event-threshold \\( \\Delta \\) is asymmetric, i.e., when it is different for positive and negative polarity events." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.722, + 0.466, + 0.739 + ], + "angle": 0, + "content": "4.2. Training View Generation and Remosaicing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.773 + ], + "angle": 0, + "content": "To generate the log-difference image \\(\\hat{D} \\in \\mathbb{R}^{H \\times W}\\), we compute" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.772, + 0.483, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\hat {D} := \\hat {L} _ {2} - \\hat {L} _ {1} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "which requires the remosaiced log images \\(L_{1}, L_{2} \\in \\mathbb{R}^{H \\times W}\\) at the poses \\(P_{k_{\\mathrm{start}}}^{\\prime}\\) and \\(P_{k_{\\mathrm{end}}}\\) respectively. These poses are the same that were used for generating our target \\(D\\). To obtain \\(L_{1}\\) and \\(L_{2}\\) we first use the rasterizer to generate corresponding views \\(I_{1}, I_{2} \\in \\mathbb{R}^{H \\times W \\times 3}\\) from the current Gaussian scene representation. Next, we perform a standard remosaicing operation because events are triggered per pixel" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.489, + 0.907, + 0.52 + ], + "angle": 0, + "content": "asynchronously, not allowing us to use conventional demosaicing methods [5, 29, 35, 36, 41, 44]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.526, + 0.905, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\text {R e m o s a i c i n g}: \\mathbb {R} ^ {H \\times W \\times 3} \\rightarrow \\mathbb {R} ^ {H \\times W} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.629 + ], + "angle": 0, + "content": "which reintroduces the Bayer RGB pattern. It consists of two steps: First, a set of color-channel specific matrices of size \\(2 \\times 2\\) are Hadamard multiplied with each \\(2 \\times 2\\) pixel block in the image. The channel specific matrices \\(R, G, B \\in \\mathbb{R}^{2 \\times 2}\\) are given by" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.634, + 0.905, + 0.67 + ], + "angle": 0, + "content": "\\[\nR = \\left[ \\begin{array}{l l} 1 & 0 \\\\ 0 & 0 \\end{array} \\right], \\quad G = \\left[ \\begin{array}{l l} 0 & 1 \\\\ 1 & 0 \\end{array} \\right], \\quad B = \\left[ \\begin{array}{l l} 0 & 0 \\\\ 0 & 1 \\end{array} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.751 + ], + "angle": 0, + "content": "The thus modified channels have non-zero entries at non-overlapping pixel locations. From this, the single-channel remosaiced image is obtained by addition across the channels. Finally, entry-wise logarithm computation yields the desired images \\(\\hat{L}_1, \\hat{L}_2\\) and thus \\(\\hat{D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.831, + 0.773 + ], + "angle": 0, + "content": "4.3. Event-to-Video Guided Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Event cameras capture only the pixels that indicate changes in brightness. This results in sparse information that is often insufficient for effective supervision. Thus, reconstructing 3D information, including background details, solely from event data is exceedingly challenging. To simplify the optimization process we propose leveraging the prior knowledge encoded in an event-to-video model. We use this pretrained model to generate images from event streams that" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26879" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "can subsequently be processed by Structure from Motion (SfM) techniques. Although the generated images often contain significant noise and do not accurately reproduce the underlying RGB image values, they retain substantial texture and background information. This information is sufficient to provide initial positions for the Gaussians, improve the 3DGS optimization process and lead to faster convergence compared to basic initialization with random initial points." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.236, + 0.431, + 0.254 + ], + "angle": 0, + "content": "4.4. Event Camera Trajectory Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.486 + ], + "angle": 0, + "content": "The camera poses, \\( P_{k_{\\mathrm{start}}} \\) and \\( P_{k_{\\mathrm{end}}} \\), used to obtain the log images \\( \\hat{L}_1 \\) and \\( \\hat{L}_2 \\) in Eq. (6), are in most cases not given a priori. This is because event camera poses are typically obtained from a tracking or motion estimation system that operates at a fixed rate, which may not be aligned with the rate of event accumulation. Thus, to obtain an estimate of the complete trajectory of the event camera, we adopt an efficient trajectory interpolation method for camera poses. Our work utilizes Cubic Spline Interpolation and Spherical Cubic Spline Interpolation. These generate a smooth path by interpolating between discrete command points using third-degree polynomials. This interpolation technique closely approximates real-world camera motion by maintaining non-linear continuity in both velocity and acceleration of camera movement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.236, + 0.51 + ], + "angle": 0, + "content": "4.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.579 + ], + "angle": 0, + "content": "3D Gaussian Splitting is updated using a loss that combines a \\(\\mathcal{L}_1\\) reconstruction term and a perceptual term \\(\\mathcal{L}_{\\mathrm{SSIM}}\\) that is based on the Structural Similarity Index Measure (SSIM) [67]" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.589, + 0.483, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {S S I M}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.691 + ], + "angle": 0, + "content": "To fully utilize real-world event camera data, undistortion operation is performed as the accumulated image \\( D \\) is computed. We account for this by only incurring the loss where events were actually accumulated. This allows supervising the 3D representation more efficient." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.222, + 0.72 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.194, + 0.743 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our evaluation of 3D Gaussian Splating from a moving event camera encompasses both synthetic and real-world scenes. Our method assumes that we can access the intrinsic camera matrix, its lens distortion parameters, and the camera's constant-rate poses, sampled at a high frequency to ensure precise interpolation for the event camera at any chosen time point. Estimation of camera poses at any given timestep is crucial for precise scene reconstruction from event data, as events can asynchronously occur at arbitrary timesteps between the constant-rate sampled camera poses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.676, + 0.107 + ], + "angle": 0, + "content": "5.1.1. Synthetic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.111, + 0.905, + 0.261 + ], + "angle": 0, + "content": "The synthetic dataset used in our experiments consists of event data streams obtained from ESIM [55], initially introduced in [42], encompassing seven synthetic scenes with a white background. These simulated event data streams feature a diverse range of photorealistic objects with complex structures. The event camera in these synthetic scenes moves around the objects, providing diverse geometric and lighting conditions. This setup allows for a comprehensive evaluation of the robustness of our method against the baselines in a controlled environment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.271, + 0.692, + 0.285 + ], + "angle": 0, + "content": "5.1.2. Real-World Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.905, + 0.487 + ], + "angle": 0, + "content": "The experiments on real scenes were conducted using subsets of the EDS [24], and TUM-VIE [30] datasets. From EDS, we utilize the sequences 03_rocket-earth DARK, 07_ziggy_and_fuzz hdr, 08_peanuts_runting, 11_all Characters and 13_airplane. From TUM-VIE, we utilize the sequences mocap-1d-trans and mocap-desk2. These sequences primarily capture various objects in 360-degree scenes and on a desk from a forward-facing perspective. These real sequences are relatively well-suited for novel view synthesis and were recorded using the high-resolution event sensor, the Prophesee Gen 3 in the EDS and the Prophesee Gen 4 in the TUM-VIE dataset, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.622, + 0.512 + ], + "angle": 0, + "content": "5.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.905, + 0.627 + ], + "angle": 0, + "content": "We benchmarked our method against Robust-e-NeRF and E2VID [56] + 3D Gaussian Splatting (E2VID+3DGS). Robust-e-NeRF incorporates InstantNGP, which is significantly faster compared to traditional NeRF methods, but still falls short compared to 3DGS. E2VID generates RGB images from an event data stream, followed by feeding the images into original 3D Gaussian Splatting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.61, + 0.652 + ], + "angle": 0, + "content": "5.3. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.839 + ], + "angle": 0, + "content": "We utilized commonly used metrics to evaluate novel view synthesis, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [79]) with VGG16, to comprehensively assess the quality of rendered images. It's worth noting that SSIM and LPIPS are perceptual metrics introduced to align more closely with human preference, with LPIPS specifically tailored to mimic human perception. These metrics address issues associated with PSNR, such as favoring blurry images. Additionally, we measured and compared the rendering speeds of our method, Robust-e-NeRF, and E2VID+3DGS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Since event cameras capture variations in log-radiance rather than absolute log-radiance values, the predicted intensity \\( I(t) \\) from the 3D Gaussian Splatting has an unknown offset. To rectify this limitation, a linear color transforma" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26880" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.187, + 0.089, + 0.347, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.089, + 0.507, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.659, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.09, + 0.813, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.226, + 0.346, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.233, + 0.505, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.234, + 0.659, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.234, + 0.813, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.347, + 0.332, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.347, + 0.504, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.347, + 0.659, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.347, + 0.812, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.479, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Figure 3. Generated Synthetic images comparing our work, event-based NeRF, and E2VID+3DGS qualitatively, with rendering times shown at the bottom of each image. Our work and Robust-e-NeRF outperform others in most scenes. In some cases, e.g., the drums scene, our work performs better and achieves significantly faster rendering times than conventional Robust-e-NeRF composed of InstantNGP." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.535, + 0.908, + 0.566 + ], + "angle": 0, + "content": "Table 1. Generated synthetic scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively. We supervised our model with randomly initialized points." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.575, + 0.845, + 0.778 + ], + "angle": 0, + "content": "
MetricMethodSynthetic SceneMean
chairdrumsficushotdoglegomaterialsmic
PSNR ↑E2VID + 3DGS21.3919.8619.9015.5518.1720.0820.1019.29
Robust e-NeRF30.2423.1530.7128.0727.3424.9832.8728.19
Ours28.6925.8129.9022.9129.2227.1633.2728.14
SSIM ↑E2VID + 3DGS0.9340.9150.9220.8970.8950.9010.9570.917
Robust e-NeRF0.9580.8970.9710.9530.9340.9230.9810.945
Ours0.9530.9470.9660.9400.9450.9360.9860.953
LPIPS ↓E2VID + 3DGS0.0760.0940.1080.2080.1450.1250.0690.118
Robust e-NeRF0.0400.0910.0220.0950.0740.0520.0290.057
Ours0.0470.0520.0280.0980.0550.0600.0150.051
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.892 + ], + "angle": 0, + "content": "tion is designed to adjust our predictions in the logarithmic domain [42, 59]. This transformation is both necessary and adequately effective for aligning our predictions with the reference data. It ensures that the reconstructed intensity values are properly calibrated and aligned with the observed event data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.801, + 0.729, + 0.816 + ], + "angle": 0, + "content": "5.4. Results and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "After completing the training of our model on each scene using event sequence data, the model acquires a set of 3D Gaussian ellipsoids representing the scene. By rasterizing these Gaussians, we are able to render various arbitrary views of each scene in real-time. The renderings of novel" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26881" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.214, + 0.088, + 0.785, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.58, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons of the images generated by our method, event-based NeRF and E2VID+3DGS show that our technique appears to recover details in 5 real scenes. First three rows(07_ziggy_and_fuzz_hdr, 08_peanuts_runting and 11_all Characters) are EDS dataset and last two rows are TUM-VIE dataset(mocap-1d-trans and mocap-desk2)." + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.637, + 0.846, + 0.652 + ], + "angle": 0, + "content": "Table 2. Real scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively." + }, + { + "type": "table", + "bbox": [ + 0.186, + 0.661, + 0.809, + 0.864 + ], + "angle": 0, + "content": "
MetricMethodReal SceneMean
0307081113
PSNR ↑E2VID + 3DGS15.6715.0514.0313.8318.9615.51
Robust e-NeRF19.1914.7814.7514.4318.1016.25
Ours20.7819.1417.5317.7919.0518.86
SSIM ↑E2VID + 3DGS0.7160.6890.6420.6910.7230.692
Robust e-NeRF0.8460.8150.7350.5690.7290.739
Ours0.8350.8160.7450.7890.7740.792
LPIPS ↓E2VID + 3DGS0.2660.3780.4020.4150.4150.375
Robust e-NeRF0.3240.4760.5670.7000.6500.543
Ours0.2390.3510.4240.3910.4070.363
" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26882" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "views, obtained from held-out views or test views provided by the datasets, are then used for evaluations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.13, + 0.254, + 0.144 + ], + "angle": 0, + "content": "5.4.1. Synthetic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.483, + 0.42 + ], + "angle": 0, + "content": "We evaluate our methodology on a synthetic dataset using the metrics described in Sec. 5.3. Quantitative results are shown in Tab. 1, and visual samples and rendering times in Fig. 3. Rendering times are provided in milliseconds (ms). Our method achieves higher average SSIM and compared to previous works. A high SSIM value indicates structural similarity between generated and reference images, underscoring our model's accuracy in capturing essential scene geometry. Additionally, a low LPIPS score reflects perceptual quality, showing the produced images are both realistic and visually engaging. The elevated scores demonstrate that our novel view synthesis model faithfully reproduces the visual and perceptual attributes of the original scene. Moreover, the rendering speed of both our method and E2VID+3DGS significantly surpass that of Robust-e-NeRF. Therefore, our work maintains superior quality and speed compared to Robust-e-NeRF, highlighting its effectiveness in both fidelity and performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.483, + 0.48 + ], + "angle": 0, + "content": "As shown in Fig. 3, the quality of rendered images remains high in scenes featuring drums and microphones, achieving novel view synthesis without sacrificing quality in other scenes as well." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.488, + 0.22, + 0.502 + ], + "angle": 0, + "content": "5.4.2. Real Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.734 + ], + "angle": 0, + "content": "The EDS dataset is the only one available for quantifying the performance of novel view synthesis. Therefore, we primarily conducted quantitative and qualitative assessments, as shown Tab. 2 and Fig. 4. Our method successfully recovers fine details in all sequences and contrasts in the mocap-1d-trans image, particularly. This is particularly noticeable in the images containing the checkerboard and the cover of a book. Additionally, it effectively smooths uniform areas such as the checkerboard and the desk surface in the mocap-1d-trans and mocap-desk2. This smoothness likely results from the accumulation of event streams, which may have enhanced the signal-to-noise ratio due to the aggregation process. The visible artifacts near the borders of all synthesized views are attributed to the relatively narrow field of view of the event cameras." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.203, + 0.758 + ], + "angle": 0, + "content": "5.5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of event-to-video guided initialization and camera trajectory estimation with cubic spline interpolation, we conducted ablation experiments by separately adding 1) event-to-video guided initialization and 2) both the initialization and cubic spline interpolation. As shown in Tab. 3, event-to-video guided initialization outperforms random initialization. Furthermore, Fig. 5 shows that our initialization approach results in better image quality (e.g. backgrounds) compared to random initialization." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 3. Ablation results for the event-to-video guided initialization and cubic spline interpolation." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.13, + 0.888, + 0.201 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
random init18.020.7670.397
guided init18.750.7880.359
guided init + cubic18.860.7920.363
" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.215, + 0.659, + 0.227 + ], + "angle": 0, + "content": "Random init" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.228, + 0.708, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.214, + 0.902, + 0.227 + ], + "angle": 0, + "content": "Event-to-Video Guided init" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.228, + 0.903, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.354, + 0.906, + 0.396 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of the images generated by random initialization and event-to-video guided initialization are shown in 07_ziggy_and_fuzz hdr scene in EDS dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.422, + 0.905, + 0.452 + ], + "angle": 0, + "content": "Adding cubic spline interpolation further enhances performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.462, + 0.64, + 0.477 + ], + "angle": 0, + "content": "5.6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.667 + ], + "angle": 0, + "content": "Our method, though effective, has some limitations. It is designed exclusively for novel view synthesis in static scenes and cannot handle scenarios with dynamic moving objects, which we consider to be an exciting avenue for future research. Additionally, our approach trains 3D Gaussian Splatting on event-accumulated images between two viewpoints, representing relative intensity images. Consequently, similar to other event-based view synthesis methods, the model cannot directly estimate absolute intensity images and requires a linear transformation with evaluation data as a reference. This linear transformation is required only during inference and does not impact the training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.783, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we introduce a novel method for 3D Gaussian Splatting from moving event cameras. Our approach achieves the higher visual fidelity and better performance of novel view synthesis in static scenes. We distill geometric priors from the event-to-video model for initialization and leverage effective camera trajectory interpolation. Thus, EventSplat yields improved metrics compared to existing event-based NeRF approaches. Moreover, the incorporation of 3D Gaussian Splatting enables our model to render scenes at significantly higher frame rates. This enhances the practicality of real-time novel view synthesis with event cameras. Our work is thus particularly useful in scenarios where conventional imaging techniques are inadequate." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26883" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.162 + ], + "angle": 0, + "content": "The work is in part supported by Sony. We would also like to thank Ziyi Wu, Yash Kant, and Umangi Jain, for valuable discussions and support." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. TNNLS, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[5] Hong Cao and Alex C Kot. Accurate detection of demosaicing regularity for digital image forensics. IEEE Trans. Inf. Forensics Secur, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[7] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[8] Jiaben Chen, Yichen Zhu, Dongze Lian, Jiaqi Yang, Yifu Wang, Renrui Zhang, Xinhang Liu, Shenhan Qian, Laurent Kneip, and Shenghua Gao. Revisiting Event-Based Video Frame Interpolation. In IROS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In arXiv, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3D using Gaussian Splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[11] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[13] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Sci. Robot, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[14] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. TPAMI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In CVPR, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. PAMI, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[17] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. FastNeRF: High-Fidelity Neural Rendering at 200FPS. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Asynchronous, photometric feature tracking using events and frames. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[19] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[20] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. IJCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[21] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[22] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[23] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[24] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[25] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-NeRF: Event based neural radiance field. In WACV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ToG, 2023. 1, 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[27] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[28] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[29] Ron Kimmel. Demosaicing: Image reconstruction from color CCD samples. IEEE Trans. Image Process, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[30] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM Stereo Visual-Inertial Event Dataset. In IROS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[31] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. RA-L, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[32] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. AdaNeRF: Adaptive sampling for real-time rendering of neural radiance fields. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient Sphere-Based Neural Rendering. In CVPR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26884" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[34] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[35] Xin Li. Demosaicing by successive approximation. IEEE Trans. Image Process, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.206 + ], + "angle": 0, + "content": "[36] Xin Li, Bahadir Gunturk, and Lei Zhang. Image demosaic-ing: A systematic survey. In VCIP. SPIE, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[37] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 db \\(15\\mu s\\) latency asynchronous temporal contrast vision sensor. JSSC, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[38] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning Event-Driven Video Deblurring and Interpolation. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[39] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. BACON: Band-limited coordinate networks for multiscale scene representation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[40] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[41] Philippe Longere, Xuemei Zhang, Peter B Delahunt, and David H Brainard. Perceptual assessment of demosaicing algorithm performance. Proceedings of the IEEE, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[42] Weng Fei Low and Gim Hee Lee. Robust e-NeRF: NeRF from Sparse & Noisy Events under Non-Uniform Motion, 2023. 1, 2, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[43] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V. Sander. Deblur-NeRF: Neural Radiance Fields From Blurry Images. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[44] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[45] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: representing scenes as neural radiance fields for view synthesis. ECCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[47] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. NeRF in the Dark: High Dynamic Range View Synthesis From Noisy Raw Images. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[48] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G. Derpanis, Jonathan Kelly, Marcus A. Brubaker, Igor Gilitschenski, and Alex Levinshtein. SPIn-NeRF: Multiview Segmentation and Perceptual Inpainting With Neural Radiance Fields. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[50] Marc Osswald, Sio-Hoi Ieng, Ryad Benosman, and Giacomo Indiveri. A spiking neural network model of 3d perception for event-based neuromorphic stereo vision systems. Sci. Rep., 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[51] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[52] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurlPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.193, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[54] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. RA-L, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.292, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[55] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In CoRL. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.334, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[56] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. PAMI, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.378, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[57] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.421, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[58] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. In arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.491, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[59] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. EventNeRF: Neural radiance fields from a single colour event camera. In CVPR, 2023. 1, 2, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[60] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.591, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[61] Amos Sironi, Manuele Brambilla, Nicolas Bourdis, Xavier Lagorce, and Ryad Benosman. HATS: Histograms of averaged time surfaces for robust event-based object classification. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.648, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[62] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[63] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[64] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-Based Frame Interpolation With Parametric Non-Linear Flow and Multi-Scale Fusion. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26885" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[66] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[67] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale Structural Similarity for Image Quality Assessment. In ACSSC, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[68] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[69] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[70] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. arXiv preprint arXiv:2407.11343, 2024. 2, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[71] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. arXiv preprint arXiv:2311.17286, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[72] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[73] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[74] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In SIGGRAPH, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[75] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[77] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[78] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S. Ren. Training weakly supervised video frame interpolation with events. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[80] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[81] Yi Zhou, Guillermo Gallego, Henri Rebecq, Laurent Kneip, Hongdong Li, and Davide Scaramuzza. Semi-Dense 3D Reconstruction with a Stereo Event Camera. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[82] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. T-RO, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[83] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In RSS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[84] Baozhou Zhu, Peter Hofstee, Johan Peltenburg, Jinho Lee, and Zaid Alars. AutoReCon: Neural Architecture Search-based Reconstruction for Data-free Compression, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[85] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. EWA volume splatting. In VIS, 2001. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26886" + } + ] +] \ No newline at end of file diff --git a/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_origin.pdf b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..67774ac9aab7f007542db3d13106286c3506201d --- /dev/null +++ b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/5e6f3efa-e5cc-4708-8b09-713b537fa1a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:511b7f9663b8128710482473d419c7f23e0c785acae8546003b45277290634b4 +size 2082363 diff --git a/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/full.md b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fda4a61101d1c7710f7bf7b2dd1830f44e5e1334 --- /dev/null +++ b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/full.md @@ -0,0 +1,365 @@ +# EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering + +Toshiya Yura1,2 + +toshiya.yura@sony.com + +Ashkan Mirzaei + +ashkan@cs.toronto.edu + +Igor Gilitschenski² + +gilitschenski@cs.toronto.edu + +1Sony Semiconductor Solutions Corporation, 2University of Toronto + +# Abstract + +We introduce a method for using event camera data in novel view synthesis via Gaussian Splatting. Event cameras offer exceptional temporal resolution and a high dynamic range. Leveraging these capabilities allows us to effectively address the novel view synthesis challenge in the presence of fast camera motion. For initialization of the optimization process, our approach uses prior knowledge encoded in an event-to-video model. We also use spline interpolation for obtaining high quality poses along the event camera trajectory. This enhances the reconstruction quality from fast-moving cameras while overcoming the computational limitations traditionally associated with event-based Neural Radiance Field (NeRF) methods. Our experimental evaluation demonstrates that our results achieve higher visual fidelity and better performance than existing event-based NeRF approaches while being an order of magnitude faster to render. + +# 1. Introduction + +Event cameras are a vision modality that draws strong inspiration from biological vision and neuromorphic design. In these sensors, each pixel operates independently by triggering events upon a log-intensity change above a pre-defined threshold [16, 37]. This results in an asynchronous data stream as opposed to capturing absolute intensity values at fixed intervals in traditional cameras. Event cameras offer low latency, high temporal resolution, and a high dynamic range while also ensuring minimal energy consumption [8, 38]. Consequently, they are particularly suitable for dynamic and challenging lighting conditions, as well as environments with rapid camera motions where classical cameras may exhibit strong motion blur [38, 63, 64, 78]. These advantages have motivated the adoption of event cameras in + +a wide range of applications covering most of the traditional computer vision tasks [13, 15, 27, 50, 51]. + +Recently, novel view synthesis has seen significant advances. This progress has been particularly driven by the emergence and development of Neural Radiance Fields (NeRF) [1-3, 33, 46, 74]. Recent works have led to NeRF improvements in terms of rendering speed, and adoption of NeRFs in 3D editing and other downstream applications [9, 43, 47-49, 84]. 3D Gaussian Splitting (3DGS) [26, 69] achieves real-time rendering by replacing the underlying representation with 3D Gaussians and using fast differentiable rasterization. The substantial speed improvement of 3DGS is achieved without compromising the visual quality. + +The success of NeRFs in view-synthesis has led to their adoption in applications with event-based cameras [25, 31, 42, 59]. Conventional NeRF approaches struggle with challenges, such as 3D reconstruction in environments with rapidly changing lighting conditions or high-speed objects due to RGB camera limitations. The integration of NeRFs with event-based cameras leverages their unique advantages and incorporates them into the NeRF framework to address the challenges that conventional NeRF approaches face. + +However, one crucial area of research focuses on expediting the rendering processes to fully unlock the potential of event-based NeRFs. Leveraging the advantages of 3D Gaussian Splatting with events requires adapting the rasterization process to allow supervision with events or accumulated events rather than RGB images. + +In summary, this paper contributes a method for 3D Gaussian Splatting from event data by leveraging the following technical ideas: + +1) An integration of event accumulation with the optimization process of 3DGS. +2) An event-to-video guided Structure from Motion (SfM) approach for initializing the 3DGS optimization process. +3) Use of cubic spline trajectory interpolation to assign camera poses to events at high rates. + +![](images/c99cff5ea35c3460ddcdfc4dde9e0e542feb19f53b56fad52f26f619ce808694.jpg) +Figure 1. EventSplat derives 3D representations of scenes in the form of 3D Gaussians from event data, enabling fast real-time renderings of scenes captured with event cameras via rasterization of 3D Gaussians. The use of event-based input is particularly advantageous when traditional RGB cameras fail due to various reasons including poor lighting conditions or motion blur from fast-moving cameras. + +Our evaluations show that these design choices achieve state-of-the-art event-based novel view synthesis performance while benefiting from fast rendering speeds. Thus, our method represents an advancement in addressing the speed constraints associated with event-based NeRFs. It opens up new possibilities for real-time novel view synthesis for environments where traditional imaging techniques struggle. + +# 2. Related Work + +# 2.1. Event Cameras for Vision Applications + +Motivated by the low energy consumption, high dynamic range, and high asynchronous temporal resolution of event cameras, event-based vision has recently garnered significant attention [16]. These advantageous properties have led to promising advancements in downstream tasks such as object tracking, detection, and recognition [14, 18-21, 28, 34, 52, 61, 71], as well as optical flow estimation [4, 22, 83]. However, being a novel data format, event data necessitates specialized architectures and algorithms to be effectively processed and utilized in various applications. In this work, we present a method to exploit event data for real-time novel view synthesis, a scenario that has not been addressed previously. + +# 2.2. Volumetric Rendering for View Synthesis + +Neural Radiance Fields (NeRFs)[46] have been the state-of-the-art in view synthesis, enabled by volumetric rendering and neural scene representations. Significant research has been conducted to improve NeRFs in terms of training and rendering speed [6, 9, 23, 32, 49, 58, 60, 76] and visual quality [1, 2, 12, 39, 65]. However, due to the expensive volumetric sampling process, NeRF-based methods have generally been unable to render high-quality views in real time. Recently, 3D Gaussian Splitting (3DGS) [7, 10, 26, 53, 85] proposed a rasterization-based approach for differentiably + +rendering scenes represented as sets of 3D Gaussian ellipsoids. This method achieves real-time renderings while preserving the quality of state-of-the-art NeRF-based methods. Since the introduction of 3DGS, real-time capabilities of 3DGS have resulted in rapid adoption of Gaussians in several downstream tasks including dynamic scene reconstruction [73], 3D asset generation [75], SLAM [45], and 4D object generation [40]. For view synthesis from event data, we leverage 3DGS and adapt it to be trainable using event sequences. + +# 2.3. Novel View Synthesis via Event Data + +Motivated by the unique benefits of event-based cameras, such as high dynamic range, low latency, and high temporal resolution, there has been significant research interest in adopting event data in applications like 3D reconstruction and novel view synthesis in cases where traditional RGB cameras underperform [16, 54, 81, 82]. Recent methods explore using NeRFs with event-vision supervision [25, 31, 42, 59] to reconstruct scenes and render them from novel views. While capable of generating high-quality renderings, NeRF-based methods are generally slow to render due to expensive sampling requirements [6, 17, 57, 62]. On the other hand, 3DGS requires RGB data for supervision, and adapting 3DGS for scenarios where the supervision signal comes from an event-based camera has been challenging [11, 66, 68, 70, 72, 77, 80]. In contrast to previous event-based view synthesis methods, we enable the use of event data to supervise 3DGS, allowing for real-time renderings and state-of-the-art quality using only event data, supported by event-based initial point clouds and effective camera trajectory interpolation. + +# 3. Background + +# 3.1.Event Camera Data + +While traditional cameras output an intensity image $I$ , event cameras model changes in the log intensity $L = \log (I)$ + +of the current camera view. Similar to traditional cameras, they have a pixel array. However, in contrast to traditional cameras, each pixel acts as an independent sensor and triggers an event whenever the difference between the current log-intensity $l_{x,y}(t)$ and the log-intensity at the time of the most recent event $l_{x,y}(t_{\mathrm{recent}})$ surpasses a given threshold $\Delta > 0$ , i.e. when + +$$ +l _ {x, y} (t) - l _ {x, y} \left(t _ {\text {r e c e n t}}\right) > \Delta . \tag {1} +$$ + +This threshold is also known as contrast sensitivity [16]. It can be different for positive and negative polarity changes. + +Formally, each event is a tuple $e := (t, x, y, p)$ where $t \in \mathbb{R}$ is a timestamp at which it was triggered, $(x, y) \in \mathbb{Z}^2$ are its pixel coordinates, and $p \in \{-1, 1\}$ is the event's polarity indicating whether the log-intensity increased or decreased. Similar to traditional color cameras, a Bayer RGB pattern is used to obtain color information in color event cameras. In such sensors, each individual event provides information only for one color channel. + +# 3.2. 3D Gaussian Splitting + +To enable novel view synthesis, 3DGS learns a scene representation that consists of a set of ellipsoids with view-dependent color information encoded as spherical harmonics [7, 10, 26, 53, 85]. The essence of the method revolves around transforming optimized 3D Gaussians into 2D imagery based on predetermined camera positions. Subsequently the approach calculates pixel values accordingly by arranging and rasterization [7, 10, 26, 53, 85]. The transformation process involves mapping 3D Gaussian distributions, shaped like ellipsoids, onto a two-dimensional plane as ellipses for the purpose of rendering. This is achieved by applying a specific viewing transformation, denoted by $W$ , along with a 3D covariance matrix $\Sigma$ . The outcome is the derivation of a 2D covariance matrix $\Sigma'$ as follows + +$$ +\Sigma^ {\prime} = J W \Sigma W ^ {\top} J ^ {\top}, \tag {2} +$$ + +where the Jacobian $J$ is used to linearly approximate projective transformations. + +The pixel rendering procedure begins by pinpointing a pixel's location, represented as $p$ , within the image plane and then assess its closeness to the intersecting Gaussians. This step involves gauging the Gaussian's depth via applying a viewing transformation, denoted as $W$ [7]. This process forms an sorted sequence of Gaussians, labeled $N_{s}$ . Subsequently, a technique known as alpha compositing is employed to determine the pixel's color + +$$ +I = \sum_ {i \in N _ {s}} \left(c _ {i} \alpha_ {i} ^ {\prime} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j} ^ {\prime}\right)\right). \tag {3} +$$ + +Here, $c_{i}$ and $\alpha_{i}^{\prime}$ indicates the learned color and the opacity, respectively. + +# 4. Method + +The goal of our work is to obtain a Gaussian Splitting-based 3D representation of a static scene using a data stream from a moving event camera. Formally, we are given a sequence of $N$ events $e_k = (t_k, x_k, y_k, p_k)$ with corresponding camera poses $P_k \in \mathrm{SE}(3)$ denoted as $\mathcal{E} := \{(e_k, P_k)\}_{k=1}^N$ . In practice, such dense pose information may not be available. However, it can be computed from sparse poses via interpolation. The interpolation to estimate event camera poses is explained in Sec. 5.1. From the posed event sequence, we aim to obtain a 3DGS-based scene representation $\mathcal{G}$ that can be used for novel view synthesis of traditional rgb and grayscale images. + +The key idea of our optimization approach is to model changes in logarithmic image intensity. That is, for a trajectory sequence over a given time interval $[a,b]$ and corresponding image intensity $I(t)\in \mathbb{R}^{H\times W\times 3}$ , our work models + +$$ +E (a, b) := \int_ {a} ^ {b} \log \left(I ^ {\prime} (t)\right) \mathrm {d} t. \tag {4} +$$ + +For loss computation, we use both, event camera data and 3DGS to approximate this quantity. In the case of event cameras, this approximation can be obtained directly by creating an image of log intensity changes. Approximating Eq. (4) with 3DGS, however, requires additional modifications to the training process. + +Method Overview In its original formulation, 3D Gaussian Splitting uses a sequence of posed images as its training data. To make Gaussian Splitting applicable to event data, we introduce several modifications to the original method: First, we approximate Eq. (4) by accumulating events into an image representation $D \in \mathbb{R}^{H \times W}$ using a random sub-trajectory of the given event stream $\mathcal{E}$ (Sec. 4.1). This representation contains the sum of quantized log-intensity changes over the sub-trajectory. Second, we compute the corresponding log-intensity changes by rasterizing two views from the Gaussian scene representation (Sec. 4.2). For color cameras, this additionally requires a remosaicing step to account for the fact that the accumulated image representation of events only contains single color information for each pixel. To improve the 3DGS optimization process, we leverage an event-to-video guided structure from motion (Sec. 4.3). We also leverage cubic spline interpolation to more accurately estimate camera sub-trajectories as the rate at which the camera is tracked may be too coarse (Sec. 4.4). The full scene representation learning process is illustrated in Fig. 2. + +# 4.1. Event Accumulation + +To obtain an image of accumulated events, we first randomly sample start and endpoints $k_{\mathrm{start}}, k_{\mathrm{end}} \in \{1 \dots N\}$ + +![](images/e4d59241728ca5cd360114d94235ead907adc107a7b1c2f034d3facfe022c479.jpg) +Figure 2. Overview of our 3D Gaussian Splitting training with moving event camera data. Event data streams from $t_{k_{start}}$ to $t_{k_{end}}$ are accumulated into $D$ , and distill the point priors from an event-to-video model. The log-difference image $\hat{D}$ is obtained from 3D Gaussian Splitting and Rasterization at two camera view points. It is computed as in Eq. (6) and remosaicing is performed in case of a color event camera. The respective poses are estimated by cubic spline interpolation. + +from our trajectory sequence. The lengths of those sequences are chosen randomly with a maximum length ranging from $1\%$ to $10\%$ of the total number of events. This allows for considering scene information at different scales in order to have both global and local geometry information trained. Then, we obtain the entries of the accumulated log-intensity change image $D$ at pixel location $x,y$ as + +$$ +D _ {x, y} := \sum_ {\substack {k \in \{k _ {\text {start}}, \dots , k _ {\text {end}} \} \\ x _ {k} = x, y _ {k} = y}} p _ {k} \Delta . \tag{5} +$$ + +The resulting image serves as an approximation to Eq. (4), i.e. $D \approx E(t_{\mathrm{start}}, t_{\mathrm{end}})$ . This method can easily be adapted to the case where the event-threshold $\Delta$ is asymmetric, i.e., when it is different for positive and negative polarity events. + +# 4.2. Training View Generation and Remosaicing + +To generate the log-difference image $\hat{D} \in \mathbb{R}^{H \times W}$ , we compute + +$$ +\hat {D} := \hat {L} _ {2} - \hat {L} _ {1} \tag {6} +$$ + +which requires the remosaiced log images $L_{1}, L_{2} \in \mathbb{R}^{H \times W}$ at the poses $P_{k_{\mathrm{start}}}^{\prime}$ and $P_{k_{\mathrm{end}}}$ respectively. These poses are the same that were used for generating our target $D$ . To obtain $L_{1}$ and $L_{2}$ we first use the rasterizer to generate corresponding views $I_{1}, I_{2} \in \mathbb{R}^{H \times W \times 3}$ from the current Gaussian scene representation. Next, we perform a standard remosaicing operation because events are triggered per pixel + +asynchronously, not allowing us to use conventional demosaicing methods [5, 29, 35, 36, 41, 44] + +$$ +\text {R e m o s a i c i n g}: \mathbb {R} ^ {H \times W \times 3} \rightarrow \mathbb {R} ^ {H \times W} \tag {7} +$$ + +which reintroduces the Bayer RGB pattern. It consists of two steps: First, a set of color-channel specific matrices of size $2 \times 2$ are Hadamard multiplied with each $2 \times 2$ pixel block in the image. The channel specific matrices $R, G, B \in \mathbb{R}^{2 \times 2}$ are given by + +$$ +R = \left[ \begin{array}{l l} 1 & 0 \\ 0 & 0 \end{array} \right], \quad G = \left[ \begin{array}{l l} 0 & 1 \\ 1 & 0 \end{array} \right], \quad B = \left[ \begin{array}{l l} 0 & 0 \\ 0 & 1 \end{array} \right]. \tag {8} +$$ + +The thus modified channels have non-zero entries at non-overlapping pixel locations. From this, the single-channel remosaiced image is obtained by addition across the channels. Finally, entry-wise logarithm computation yields the desired images $\hat{L}_1, \hat{L}_2$ and thus $\hat{D}$ . + +# 4.3. Event-to-Video Guided Initialization + +Event cameras capture only the pixels that indicate changes in brightness. This results in sparse information that is often insufficient for effective supervision. Thus, reconstructing 3D information, including background details, solely from event data is exceedingly challenging. To simplify the optimization process we propose leveraging the prior knowledge encoded in an event-to-video model. We use this pretrained model to generate images from event streams that + +can subsequently be processed by Structure from Motion (SfM) techniques. Although the generated images often contain significant noise and do not accurately reproduce the underlying RGB image values, they retain substantial texture and background information. This information is sufficient to provide initial positions for the Gaussians, improve the 3DGS optimization process and lead to faster convergence compared to basic initialization with random initial points. + +# 4.4. Event Camera Trajectory Interpolation + +The camera poses, $P_{k_{\mathrm{start}}}$ and $P_{k_{\mathrm{end}}}$ , used to obtain the log images $\hat{L}_1$ and $\hat{L}_2$ in Eq. (6), are in most cases not given a priori. This is because event camera poses are typically obtained from a tracking or motion estimation system that operates at a fixed rate, which may not be aligned with the rate of event accumulation. Thus, to obtain an estimate of the complete trajectory of the event camera, we adopt an efficient trajectory interpolation method for camera poses. Our work utilizes Cubic Spline Interpolation and Spherical Cubic Spline Interpolation. These generate a smooth path by interpolating between discrete command points using third-degree polynomials. This interpolation technique closely approximates real-world camera motion by maintaining non-linear continuity in both velocity and acceleration of camera movement. + +# 4.5. Loss Function + +3D Gaussian Splitting is updated using a loss that combines a $\mathcal{L}_1$ reconstruction term and a perceptual term $\mathcal{L}_{\mathrm{SSIM}}$ that is based on the Structural Similarity Index Measure (SSIM) [67] + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {1} + \lambda \mathcal {L} _ {\mathrm {S S I M}}. \tag {9} +$$ + +To fully utilize real-world event camera data, undistortion operation is performed as the accumulated image $D$ is computed. We account for this by only incurring the loss where events were actually accumulated. This allows supervising the 3D representation more efficient. + +# 5. Experiments + +# 5.1. Datasets + +Our evaluation of 3D Gaussian Splating from a moving event camera encompasses both synthetic and real-world scenes. Our method assumes that we can access the intrinsic camera matrix, its lens distortion parameters, and the camera's constant-rate poses, sampled at a high frequency to ensure precise interpolation for the event camera at any chosen time point. Estimation of camera poses at any given timestep is crucial for precise scene reconstruction from event data, as events can asynchronously occur at arbitrary timesteps between the constant-rate sampled camera poses. + +# 5.1.1. Synthetic Scenes + +The synthetic dataset used in our experiments consists of event data streams obtained from ESIM [55], initially introduced in [42], encompassing seven synthetic scenes with a white background. These simulated event data streams feature a diverse range of photorealistic objects with complex structures. The event camera in these synthetic scenes moves around the objects, providing diverse geometric and lighting conditions. This setup allows for a comprehensive evaluation of the robustness of our method against the baselines in a controlled environment. + +# 5.1.2. Real-World Scenes + +The experiments on real scenes were conducted using subsets of the EDS [24], and TUM-VIE [30] datasets. From EDS, we utilize the sequences 03_rocket-earth DARK, 07_ziggy_and_fuzz hdr, 08_peanuts_runting, 11_all Characters and 13_airplane. From TUM-VIE, we utilize the sequences mocap-1d-trans and mocap-desk2. These sequences primarily capture various objects in 360-degree scenes and on a desk from a forward-facing perspective. These real sequences are relatively well-suited for novel view synthesis and were recorded using the high-resolution event sensor, the Prophesee Gen 3 in the EDS and the Prophesee Gen 4 in the TUM-VIE dataset, respectively. + +# 5.2. Baselines + +We benchmarked our method against Robust-e-NeRF and E2VID [56] + 3D Gaussian Splatting (E2VID+3DGS). Robust-e-NeRF incorporates InstantNGP, which is significantly faster compared to traditional NeRF methods, but still falls short compared to 3DGS. E2VID generates RGB images from an event data stream, followed by feeding the images into original 3D Gaussian Splatting. + +# 5.3. Metrics + +We utilized commonly used metrics to evaluate novel view synthesis, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [79]) with VGG16, to comprehensively assess the quality of rendered images. It's worth noting that SSIM and LPIPS are perceptual metrics introduced to align more closely with human preference, with LPIPS specifically tailored to mimic human perception. These metrics address issues associated with PSNR, such as favoring blurry images. Additionally, we measured and compared the rendering speeds of our method, Robust-e-NeRF, and E2VID+3DGS. + +Since event cameras capture variations in log-radiance rather than absolute log-radiance values, the predicted intensity $I(t)$ from the 3D Gaussian Splatting has an unknown offset. To rectify this limitation, a linear color transforma + +![](images/ff41e83e35d3082964b1d049335c78f10590a2c016dda0f0001fd765c3787db6.jpg) + +![](images/3ea78d5915b7cfc9becea1d7f466e433b03771de3e1406ed67bd9baaabb9895a.jpg) + +![](images/ebfbb3e6d9f319ac30322afd9e0c2b6485d5b7cbcd86723fb4600dd95d4f03bd.jpg) + +![](images/1c4708c3bf0d52f5e7b0c53cf19bbbf1d0d00339385252570f44fe1e5e99e64c.jpg) + +![](images/8df9fa988fa5ff3c9d93ac20a141a6bf908da482db1ef55ae34a8ed8c5d8d6bb.jpg) + +![](images/8b7a59d98c2daf99711123f734a20545a40404b012b4e2df05b4b091e80cf054.jpg) + +![](images/124daa59c57e4ac9db3e436aabbd67e938035323e55d3d7390fdbcfd154ba6ef.jpg) + +![](images/280db115ed57b244ec055245155afca539b8340f4350bb561af760d014db1403.jpg) + +![](images/2cb8279c9c93689ae7441c7b8cf9afdf2b3db86f23518f94e3c02bf77048385f.jpg) +Figure 3. Generated Synthetic images comparing our work, event-based NeRF, and E2VID+3DGS qualitatively, with rendering times shown at the bottom of each image. Our work and Robust-e-NeRF outperform others in most scenes. In some cases, e.g., the drums scene, our work performs better and achieves significantly faster rendering times than conventional Robust-e-NeRF composed of InstantNGP. + +![](images/0e1e0d057cc4fc9f9336589f445bbedeb41595c96c9d4d73ee9e428154f8922e.jpg) + +![](images/76b81c8e4856c97ec9051b464f57ea7c7df7ca5e91572577eb07ba04b643a458.jpg) + +![](images/73c0744d31d88dad680de59f351631504f90c181dd6a77b373df1785bcad338b.jpg) + +Table 1. Generated synthetic scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively. We supervised our model with randomly initialized points. + +
MetricMethodSynthetic SceneMean
chairdrumsficushotdoglegomaterialsmic
PSNR ↑E2VID + 3DGS21.3919.8619.9015.5518.1720.0820.1019.29
Robust e-NeRF30.2423.1530.7128.0727.3424.9832.8728.19
Ours28.6925.8129.9022.9129.2227.1633.2728.14
SSIM ↑E2VID + 3DGS0.9340.9150.9220.8970.8950.9010.9570.917
Robust e-NeRF0.9580.8970.9710.9530.9340.9230.9810.945
Ours0.9530.9470.9660.9400.9450.9360.9860.953
LPIPS ↓E2VID + 3DGS0.0760.0940.1080.2080.1450.1250.0690.118
Robust e-NeRF0.0400.0910.0220.0950.0740.0520.0290.057
Ours0.0470.0520.0280.0980.0550.0600.0150.051
+ +tion is designed to adjust our predictions in the logarithmic domain [42, 59]. This transformation is both necessary and adequately effective for aligning our predictions with the reference data. It ensures that the reconstructed intensity values are properly calibrated and aligned with the observed event data. + +# 5.4. Results and Evaluation + +After completing the training of our model on each scene using event sequence data, the model acquires a set of 3D Gaussian ellipsoids representing the scene. By rasterizing these Gaussians, we are able to render various arbitrary views of each scene in real-time. The renderings of novel + +![](images/a96c47fa963139724dad03565260fb3b2b24182a37e492b97be4bfc4df3960a1.jpg) +Figure 4. Qualitative comparisons of the images generated by our method, event-based NeRF and E2VID+3DGS show that our technique appears to recover details in 5 real scenes. First three rows(07_ziggy_and_fuzz_hdr, 08_peanuts_runting and 11_all Characters) are EDS dataset and last two rows are TUM-VIE dataset(mocap-1d-trans and mocap-desk2). + +Table 2. Real scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively. + +
MetricMethodReal SceneMean
0307081113
PSNR ↑E2VID + 3DGS15.6715.0514.0313.8318.9615.51
Robust e-NeRF19.1914.7814.7514.4318.1016.25
Ours20.7819.1417.5317.7919.0518.86
SSIM ↑E2VID + 3DGS0.7160.6890.6420.6910.7230.692
Robust e-NeRF0.8460.8150.7350.5690.7290.739
Ours0.8350.8160.7450.7890.7740.792
LPIPS ↓E2VID + 3DGS0.2660.3780.4020.4150.4150.375
Robust e-NeRF0.3240.4760.5670.7000.6500.543
Ours0.2390.3510.4240.3910.4070.363
+ +views, obtained from held-out views or test views provided by the datasets, are then used for evaluations. + +# 5.4.1. Synthetic Scenes + +We evaluate our methodology on a synthetic dataset using the metrics described in Sec. 5.3. Quantitative results are shown in Tab. 1, and visual samples and rendering times in Fig. 3. Rendering times are provided in milliseconds (ms). Our method achieves higher average SSIM and compared to previous works. A high SSIM value indicates structural similarity between generated and reference images, underscoring our model's accuracy in capturing essential scene geometry. Additionally, a low LPIPS score reflects perceptual quality, showing the produced images are both realistic and visually engaging. The elevated scores demonstrate that our novel view synthesis model faithfully reproduces the visual and perceptual attributes of the original scene. Moreover, the rendering speed of both our method and E2VID+3DGS significantly surpass that of Robust-e-NeRF. Therefore, our work maintains superior quality and speed compared to Robust-e-NeRF, highlighting its effectiveness in both fidelity and performance. + +As shown in Fig. 3, the quality of rendered images remains high in scenes featuring drums and microphones, achieving novel view synthesis without sacrificing quality in other scenes as well. + +# 5.4.2. Real Scenes + +The EDS dataset is the only one available for quantifying the performance of novel view synthesis. Therefore, we primarily conducted quantitative and qualitative assessments, as shown Tab. 2 and Fig. 4. Our method successfully recovers fine details in all sequences and contrasts in the mocap-1d-trans image, particularly. This is particularly noticeable in the images containing the checkerboard and the cover of a book. Additionally, it effectively smooths uniform areas such as the checkerboard and the desk surface in the mocap-1d-trans and mocap-desk2. This smoothness likely results from the accumulation of event streams, which may have enhanced the signal-to-noise ratio due to the aggregation process. The visible artifacts near the borders of all synthesized views are attributed to the relatively narrow field of view of the event cameras. + +# 5.5. Ablations + +To demonstrate the effectiveness of event-to-video guided initialization and camera trajectory estimation with cubic spline interpolation, we conducted ablation experiments by separately adding 1) event-to-video guided initialization and 2) both the initialization and cubic spline interpolation. As shown in Tab. 3, event-to-video guided initialization outperforms random initialization. Furthermore, Fig. 5 shows that our initialization approach results in better image quality (e.g. backgrounds) compared to random initialization. + +Table 3. Ablation results for the event-to-video guided initialization and cubic spline interpolation. + +
MethodPSNR↑SSIM↑LPIPS↓
random init18.020.7670.397
guided init18.750.7880.359
guided init + cubic18.860.7920.363
+ +![](images/1e7e67edf5eaab2048813d2c5381f29caadb59159351eacd977bf1d6e59f1367.jpg) +Random init + +![](images/c7826f7c045ab390dc83d260299516fa74f494d68ddef06c024bdce3cda7ae09.jpg) +Event-to-Video Guided init +Figure 5. Qualitative comparisons of the images generated by random initialization and event-to-video guided initialization are shown in 07_ziggy_and_fuzz hdr scene in EDS dataset. + +Adding cubic spline interpolation further enhances performance. + +# 5.6. Limitations + +Our method, though effective, has some limitations. It is designed exclusively for novel view synthesis in static scenes and cannot handle scenarios with dynamic moving objects, which we consider to be an exciting avenue for future research. Additionally, our approach trains 3D Gaussian Splatting on event-accumulated images between two viewpoints, representing relative intensity images. Consequently, similar to other event-based view synthesis methods, the model cannot directly estimate absolute intensity images and requires a linear transformation with evaluation data as a reference. This linear transformation is required only during inference and does not impact the training. + +# 6. Conclusion and Future Work + +In this work, we introduce a novel method for 3D Gaussian Splatting from moving event cameras. Our approach achieves the higher visual fidelity and better performance of novel view synthesis in static scenes. We distill geometric priors from the event-to-video model for initialization and leverage effective camera trajectory interpolation. Thus, EventSplat yields improved metrics compared to existing event-based NeRF approaches. Moreover, the incorporation of 3D Gaussian Splatting enables our model to render scenes at significantly higher frame rates. This enhances the practicality of real-time novel view synthesis with event cameras. Our work is thus particularly useful in scenarios where conventional imaging techniques are inadequate. + +# Acknowledgments + +The work is in part supported by Sony. We would also like to thank Ziyi Wu, Yash Kant, and Umangi Jain, for valuable discussions and support. + +# References + +[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. ICCV, 2021. 1, 2 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In ICCV, 2023. 1 +[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. TNNLS, 2013. 2 +[5] Hong Cao and Alex C Kot. Accurate detection of demosaicing regularity for digital image forensics. IEEE Trans. Inf. Forensics Secur, 2009. 4 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV. Springer, 2022. 2 +[7] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 2, 3 +[8] Jiaben Chen, Yichen Zhu, Dongze Lian, Jiaqi Yang, Yifu Wang, Renrui Zhang, Xinhang Liu, Shenhan Qian, Laurent Kneip, and Shenghua Gao. Revisiting Event-Based Video Frame Interpolation. In IROS, 2023. 1 +[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In arXiv, 2022. 1, 2 +[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3D using Gaussian Splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3 +[11] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), 2024. 2 +[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2 +[13] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Sci. Robot, 2020. 1 +[14] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. TPAMI, 2017. 2 +[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In CVPR, 2018. 1 + +[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. PAMI, 2022. 1, 2, 3 +[17] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. FastNeRF: High-Fidelity Neural Rendering at 200FPS. In ICCV, 2021. 2 +[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Asynchronous, photometric feature tracking using events and frames. In ECCV, 2018. 2 +[19] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019. +[20] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. IJCV, 2020. +[21] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 2 +[22] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 3DV, 2021. 2 +[23] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2 +[24] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5 +[25] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-NeRF: Event based neural radiance field. In WACV, 2023. 1, 2 +[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ToG, 2023. 1, 2, 3, 12 +[27] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In CVPR, 2016. 1 +[28] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, 2021. 2 +[29] Ron Kimmel. Demosaicing: Image reconstruction from color CCD samples. IEEE Trans. Image Process, 1999. 4 +[30] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM Stereo Visual-Inertial Event Dataset. In IROS, 2021. 5 +[31] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. RA-L, 2023. 1, 2 +[32] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. AdaNeRF: Adaptive sampling for real-time rendering of neural radiance fields. In ECCV, 2022. 2 +[33] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient Sphere-Based Neural Rendering. In CVPR, 2021. 1 + +[34] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 2 +[35] Xin Li. Demosaicing by successive approximation. IEEE Trans. Image Process, 2005. 4 +[36] Xin Li, Bahadir Gunturk, and Lei Zhang. Image demosaic-ing: A systematic survey. In VCIP. SPIE, 2008. 4 +[37] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 db $15\mu s$ latency asynchronous temporal contrast vision sensor. JSSC, 2008. 1 +[38] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning Event-Driven Video Deblurring and Interpolation. In ECCV, 2020. 1 +[39] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. BACON: Band-limited coordinate networks for multiscale scene representation. In CVPR, 2022. 2 +[40] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2 +[41] Philippe Longere, Xuemei Zhang, Peter B Delahunt, and David H Brainard. Perceptual assessment of demosaicing algorithm performance. Proceedings of the IEEE, 2002. 4 +[42] Weng Fei Low and Gim Hee Lee. Robust e-NeRF: NeRF from Sparse & Noisy Events under Non-Uniform Motion, 2023. 1, 2, 5, 6, 12 +[43] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V. Sander. Deblur-NeRF: Neural Radiance Fields From Blurry Images. In CVPR, 2022. 1 +[44] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, 2004. 4 +[45] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2 +[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: representing scenes as neural radiance fields for view synthesis. ECCV, 2021. 1, 2 +[47] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. NeRF in the Dark: High Dynamic Range View Synthesis From Noisy Raw Images. In CVPR, 2022. 1 +[48] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G. Derpanis, Jonathan Kelly, Marcus A. Brubaker, Igor Gilitschenski, and Alex Levinshtein. SPIn-NeRF: Multiview Segmentation and Perceptual Inpainting With Neural Radiance Fields. In CVPR, 2023. +[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 1, 2 +[50] Marc Osswald, Sio-Hoi Ieng, Ryad Benosman, and Giacomo Indiveri. A spiking neural network model of 3d perception for event-based neuromorphic stereo vision systems. Sci. Rep., 2017. 1 + +[51] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 1 +[52] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurlPS, 2020. 2 +[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion, 2022. 2, 3 +[54] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. RA-L, 2016. 2 +[55] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In CoRL. PMLR, 2018. 5 +[56] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. PAMI, 2019. 5 +[57] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In ICCV, 2021. 2 +[58] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. In arXiv, 2023. 2 +[59] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. EventNeRF: Neural radiance fields from a single colour event camera. In CVPR, 2023. 1, 2, 6, 12 +[60] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[61] Amos Sironi, Manuele Brambilla, Nicolas Bourdis, Xavier Lagorce, and Ryad Benosman. HATS: Histograms of averaged time surfaces for robust event-based object classification. In CVPR, 2018. 2 +[62] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In CVPR, 2021. 2 +[63] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, 2021. 1 +[64] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-Based Frame Interpolation With Parametric Non-Linear Flow and Multi-Scale Fusion. In CVPR, 2022. 1 +[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2 + +[66] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2 +[67] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale Structural Similarity for Image Quality Assessment. In ACSSC, 2003. 5 +[68] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2 +[69] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528, 2023. 1 +[70] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. arXiv preprint arXiv:2407.11343, 2024. 2, 12, 14 +[71] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. arXiv preprint arXiv:2311.17286, 2023. 2 +[72] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 2 +[73] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In ICLR, 2024. 2 +[74] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In SIGGRAPH, 2023. 1 +[75] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2 +[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2 +[77] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2 +[78] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S. Ren. Training weakly supervised video frame interpolation with events. In ICCV, 2021. 1 +[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In CVPR, 2018. 5 +[80] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 2 + +[81] Yi Zhou, Guillermo Gallego, Henri Rebecq, Laurent Kneip, Hongdong Li, and Davide Scaramuzza. Semi-Dense 3D Reconstruction with a Stereo Event Camera. In ECCV, 2018. 2 +[82] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. T-RO, 2021. 2 +[83] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In RSS, 2018. 2 +[84] Baozhou Zhu, Peter Hofstee, Johan Peltenburg, Jinho Lee, and Zaid Alars. AutoReCon: Neural Architecture Search-based Reconstruction for Data-free Compression, 2021. 1 +[85] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. EWA volume splatting. In VIS, 2001. 2, 3 \ No newline at end of file diff --git a/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/images.zip b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b48563217ee7cb1462b1247777cc0e2c78ff40f3 --- /dev/null +++ b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67097da5ec10f332f43341d8243944721d4d5be1cf6966a16803bd792a157ba4 +size 594696 diff --git a/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/layout.json b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6119bfdc90bf7053fbafdcba9f13050c3034bb90 --- /dev/null +++ b/2025/EventSplat_ 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering/layout.json @@ -0,0 +1,9300 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "text", + "content": "EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 160, + 265, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 160, + 265, + 175 + ], + "spans": [ + { + "bbox": [ + 189, + 160, + 265, + 175 + ], + "type": "text", + "content": "Toshiya Yura1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 178, + 286, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 178, + 286, + 189 + ], + "spans": [ + { + "bbox": [ + 170, + 178, + 286, + 189 + ], + "type": "text", + "content": "toshiya.yura@sony.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 341, + 161, + 424, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 161, + 424, + 175 + ], + "spans": [ + { + "bbox": [ + 341, + 161, + 424, + 175 + ], + "type": "text", + "content": "Ashkan Mirzaei" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 178, + 440, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 178, + 440, + 188 + ], + "spans": [ + { + "bbox": [ + 325, + 178, + 440, + 188 + ], + "type": "text", + "content": "ashkan@cs.toronto.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 257, + 194, + 351, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 194, + 351, + 209 + ], + "spans": [ + { + "bbox": [ + 257, + 194, + 351, + 209 + ], + "type": "text", + "content": "Igor Gilitschenski²" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 212, + 381, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 212, + 381, + 222 + ], + "spans": [ + { + "bbox": [ + 228, + 212, + 381, + 222 + ], + "type": "text", + "content": "gilitschenski@cs.toronto.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 223, + 470, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 223, + 470, + 237 + ], + "spans": [ + { + "bbox": [ + 140, + 223, + 470, + 237 + ], + "type": "text", + "content": "1Sony Semiconductor Solutions Corporation, 2University of Toronto" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 265, + 200, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 265, + 200, + 277 + ], + "spans": [ + { + "bbox": [ + 151, + 265, + 200, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 291, + 297, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 291, + 297, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 297, + 495 + ], + "type": "text", + "content": "We introduce a method for using event camera data in novel view synthesis via Gaussian Splatting. Event cameras offer exceptional temporal resolution and a high dynamic range. Leveraging these capabilities allows us to effectively address the novel view synthesis challenge in the presence of fast camera motion. For initialization of the optimization process, our approach uses prior knowledge encoded in an event-to-video model. We also use spline interpolation for obtaining high quality poses along the event camera trajectory. This enhances the reconstruction quality from fast-moving cameras while overcoming the computational limitations traditionally associated with event-based Neural Radiance Field (NeRF) methods. Our experimental evaluation demonstrates that our results achieve higher visual fidelity and better performance than existing event-based NeRF approaches while being an order of magnitude faster to render." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 523, + 135, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 135, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 135, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Event cameras are a vision modality that draws strong inspiration from biological vision and neuromorphic design. In these sensors, each pixel operates independently by triggering events upon a log-intensity change above a pre-defined threshold [16, 37]. This results in an asynchronous data stream as opposed to capturing absolute intensity values at fixed intervals in traditional cameras. Event cameras offer low latency, high temporal resolution, and a high dynamic range while also ensuring minimal energy consumption [8, 38]. Consequently, they are particularly suitable for dynamic and challenging lighting conditions, as well as environments with rapid camera motions where classical cameras may exhibit strong motion blur [38, 63, 64, 78]. These advantages have motivated the adoption of event cameras in" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "content": "a wide range of applications covering most of the traditional computer vision tasks [13, 15, 27, 50, 51]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 291, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 422 + ], + "type": "text", + "content": "Recently, novel view synthesis has seen significant advances. This progress has been particularly driven by the emergence and development of Neural Radiance Fields (NeRF) [1-3, 33, 46, 74]. Recent works have led to NeRF improvements in terms of rendering speed, and adoption of NeRFs in 3D editing and other downstream applications [9, 43, 47-49, 84]. 3D Gaussian Splitting (3DGS) [26, 69] achieves real-time rendering by replacing the underlying representation with 3D Gaussians and using fast differentiable rasterization. The substantial speed improvement of 3DGS is achieved without compromising the visual quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 423, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 532 + ], + "type": "text", + "content": "The success of NeRFs in view-synthesis has led to their adoption in applications with event-based cameras [25, 31, 42, 59]. Conventional NeRF approaches struggle with challenges, such as 3D reconstruction in environments with rapidly changing lighting conditions or high-speed objects due to RGB camera limitations. The integration of NeRFs with event-based cameras leverages their unique advantages and incorporates them into the NeRF framework to address the challenges that conventional NeRF approaches face." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 532, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 604 + ], + "type": "text", + "content": "However, one crucial area of research focuses on expediting the rendering processes to fully unlock the potential of event-based NeRFs. Leveraging the advantages of 3D Gaussian Splatting with events requires adapting the rasterization process to allow supervision with events or accumulated events rather than RGB images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": "In summary, this paper contributes a method for 3D Gaussian Splatting from event data by leveraging the following technical ideas:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 642, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 642, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 642, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 315, + 642, + 553, + 666 + ], + "type": "text", + "content": "1) An integration of event accumulation with the optimization process of 3DGS." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 666, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 666, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 666, + 553, + 689 + ], + "type": "text", + "content": "2) An event-to-video guided Structure from Motion (SfM) approach for initializing the 3DGS optimization process." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 689, + 553, + 713 + ], + "type": "text", + "content": "3) Use of cubic spline trajectory interpolation to assign camera poses to events at high rates." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26876" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 491, + 200 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 491, + 200 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 491, + 200 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 491, + 200 + ], + "type": "image", + "image_path": "c99cff5ea35c3460ddcdfc4dde9e0e542feb19f53b56fad52f26f619ce808694.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "text", + "content": "Figure 1. EventSplat derives 3D representations of scenes in the form of 3D Gaussians from event data, enabling fast real-time renderings of scenes captured with event cameras via rasterization of 3D Gaussians. The use of event-based input is particularly advantageous when traditional RGB cameras fail due to various reasons including poor lighting conditions or motion blur from fast-moving cameras." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 261, + 298, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 261, + 298, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 298, + 360 + ], + "type": "text", + "content": "Our evaluations show that these design choices achieve state-of-the-art event-based novel view synthesis performance while benefiting from fast rendering speeds. Thus, our method represents an advancement in addressing the speed constraints associated with event-based NeRFs. It opens up new possibilities for real-time novel view synthesis for environments where traditional imaging techniques struggle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 143, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 143, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 143, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 391, + 261, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 261, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 261, + 404 + ], + "type": "text", + "content": "2.1. Event Cameras for Vision Applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 409, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 297, + 567 + ], + "type": "text", + "content": "Motivated by the low energy consumption, high dynamic range, and high asynchronous temporal resolution of event cameras, event-based vision has recently garnered significant attention [16]. These advantageous properties have led to promising advancements in downstream tasks such as object tracking, detection, and recognition [14, 18-21, 28, 34, 52, 61, 71], as well as optical flow estimation [4, 22, 83]. However, being a novel data format, event data necessitates specialized architectures and algorithms to be effectively processed and utilized in various applications. In this work, we present a method to exploit event data for real-time novel view synthesis, a scenario that has not been addressed previously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 272, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 272, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 272, + 588 + ], + "type": "text", + "content": "2.2. Volumetric Rendering for View Synthesis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs)[46] have been the state-of-the-art in view synthesis, enabled by volumetric rendering and neural scene representations. Significant research has been conducted to improve NeRFs in terms of training and rendering speed [6, 9, 23, 32, 49, 58, 60, 76] and visual quality [1, 2, 12, 39, 65]. However, due to the expensive volumetric sampling process, NeRF-based methods have generally been unable to render high-quality views in real time. Recently, 3D Gaussian Splitting (3DGS) [7, 10, 26, 53, 85] proposed a rasterization-based approach for differentiably" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 261, + 556, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 556, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 556, + 382 + ], + "type": "text", + "content": "rendering scenes represented as sets of 3D Gaussian ellipsoids. This method achieves real-time renderings while preserving the quality of state-of-the-art NeRF-based methods. Since the introduction of 3DGS, real-time capabilities of 3DGS have resulted in rapid adoption of Gaussians in several downstream tasks including dynamic scene reconstruction [73], 3D asset generation [75], SLAM [45], and 4D object generation [40]. For view synthesis from event data, we leverage 3DGS and adapt it to be trainable using event sequences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 387, + 509, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 509, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 509, + 399 + ], + "type": "text", + "content": "2.3. Novel View Synthesis via Event Data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 404, + 556, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 556, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 556, + 645 + ], + "type": "text", + "content": "Motivated by the unique benefits of event-based cameras, such as high dynamic range, low latency, and high temporal resolution, there has been significant research interest in adopting event data in applications like 3D reconstruction and novel view synthesis in cases where traditional RGB cameras underperform [16, 54, 81, 82]. Recent methods explore using NeRFs with event-vision supervision [25, 31, 42, 59] to reconstruct scenes and render them from novel views. While capable of generating high-quality renderings, NeRF-based methods are generally slow to render due to expensive sampling requirements [6, 17, 57, 62]. On the other hand, 3DGS requires RGB data for supervision, and adapting 3DGS for scenarios where the supervision signal comes from an event-based camera has been challenging [11, 66, 68, 70, 72, 77, 80]. In contrast to previous event-based view synthesis methods, we enable the use of event data to supervise 3DGS, allowing for real-time renderings and state-of-the-art quality using only event data, supported by event-based initial point clouds and effective camera trajectory interpolation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 652, + 392, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 392, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 392, + 666 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 672, + 430, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 430, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 430, + 684 + ], + "type": "text", + "content": "3.1.Event Camera Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "While traditional cameras output an intensity image " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": ", event cameras model changes in the log intensity " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "L = \\log (I)" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26877" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "of the current camera view. Similar to traditional cameras, they have a pixel array. However, in contrast to traditional cameras, each pixel acts as an independent sensor and triggers an event whenever the difference between the current log-intensity " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "l_{x,y}(t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and the log-intensity at the time of the most recent event " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "l_{x,y}(t_{\\mathrm{recent}})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " surpasses a given threshold " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\Delta > 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", i.e. when" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 120, + 163, + 295, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 163, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 120, + 163, + 295, + 177 + ], + "type": "interline_equation", + "content": "l _ {x, y} (t) - l _ {x, y} \\left(t _ {\\text {r e c e n t}}\\right) > \\Delta . \\tag {1}", + "image_path": "c0125dda78bc60b5ce5e2b54949486c1d3c42ba02924500455b7ae67c4c2e96f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 295, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 206 + ], + "type": "text", + "content": "This threshold is also known as contrast sensitivity [16]. It can be different for positive and negative polarity changes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "text", + "content": "Formally, each event is a tuple " + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "inline_equation", + "content": "e := (t, x, y, p)" + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "text", + "content": " is a timestamp at which it was triggered, " + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "inline_equation", + "content": "(x, y) \\in \\mathbb{Z}^2" + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "text", + "content": " are its pixel coordinates, and " + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "inline_equation", + "content": "p \\in \\{-1, 1\\}" + }, + { + "bbox": [ + 55, + 207, + 296, + 302 + ], + "type": "text", + "content": " is the event's polarity indicating whether the log-intensity increased or decreased. Similar to traditional color cameras, a Bayer RGB pattern is used to obtain color information in color event cameras. In such sensors, each individual event provides information only for one color channel." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 184, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 184, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 184, + 321 + ], + "type": "text", + "content": "3.2. 3D Gaussian Splitting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "text", + "content": "To enable novel view synthesis, 3DGS learns a scene representation that consists of a set of ellipsoids with view-dependent color information encoded as spherical harmonics [7, 10, 26, 53, 85]. The essence of the method revolves around transforming optimized 3D Gaussians into 2D imagery based on predetermined camera positions. Subsequently the approach calculates pixel values accordingly by arranging and rasterization [7, 10, 26, 53, 85]. The transformation process involves mapping 3D Gaussian distributions, shaped like ellipsoids, onto a two-dimensional plane as ellipses for the purpose of rendering. This is achieved by applying a specific viewing transformation, denoted by " + }, + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "text", + "content": ", along with a 3D covariance matrix " + }, + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "text", + "content": ". The outcome is the derivation of a 2D covariance matrix " + }, + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\Sigma'" + }, + { + "bbox": [ + 55, + 326, + 296, + 493 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 498, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\Sigma^ {\\prime} = J W \\Sigma W ^ {\\top} J ^ {\\top}, \\tag {2}", + "image_path": "5e1b358c80ef01fec3b43dc468512725ec686c628978984916a0fce1fda6fdfa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "type": "text", + "content": "where the Jacobian " + }, + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "type": "text", + "content": " is used to linearly approximate projective transformations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "content": "The pixel rendering procedure begins by pinpointing a pixel's location, represented as " + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "content": ", within the image plane and then assess its closeness to the intersecting Gaussians. This step involves gauging the Gaussian's depth via applying a viewing transformation, denoted as " + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "content": " [7]. This process forms an sorted sequence of Gaussians, labeled " + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "content": ". Subsequently, a technique known as alpha compositing is employed to determine the pixel's color" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 645, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 645, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 108, + 645, + 295, + 684 + ], + "type": "interline_equation", + "content": "I = \\sum_ {i \\in N _ {s}} \\left(c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right)\\right). \\tag {3}", + "image_path": "dfa106155ec1f2307af7cc0c06f6d6bd4a9fa7a95c821c174c290da8747c814a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " indicates the learned color and the opacity, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "content": "The goal of our work is to obtain a Gaussian Splitting-based 3D representation of a static scene using a data stream from a moving event camera. Formally, we are given a sequence of " + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "content": " events " + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "inline_equation", + "content": "e_k = (t_k, x_k, y_k, p_k)" + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "content": " with corresponding camera poses " + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "inline_equation", + "content": "P_k \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{E} := \\{(e_k, P_k)\\}_{k=1}^N" + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "content": ". In practice, such dense pose information may not be available. However, it can be computed from sparse poses via interpolation. The interpolation to estimate event camera poses is explained in Sec. 5.1. From the posed event sequence, we aim to obtain a 3DGS-based scene representation " + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "content": " that can be used for novel view synthesis of traditional rgb and grayscale images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 236, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 293 + ], + "type": "text", + "content": "The key idea of our optimization approach is to model changes in logarithmic image intensity. That is, for a trajectory sequence over a given time interval " + }, + { + "bbox": [ + 313, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "[a,b]" + }, + { + "bbox": [ + 313, + 236, + 555, + 293 + ], + "type": "text", + "content": " and corresponding image intensity " + }, + { + "bbox": [ + 313, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "I(t)\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 313, + 236, + 555, + 293 + ], + "type": "text", + "content": ", our work models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 374, + 293, + 554, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 293, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 374, + 293, + 554, + 320 + ], + "type": "interline_equation", + "content": "E (a, b) := \\int_ {a} ^ {b} \\log \\left(I ^ {\\prime} (t)\\right) \\mathrm {d} t. \\tag {4}", + "image_path": "85ce5d9c077c0840d2d73259cf17f23aeee5a6ec334e8a6eb2e2a443104ab802.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 324, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 397 + ], + "type": "text", + "content": "For loss computation, we use both, event camera data and 3DGS to approximate this quantity. In the case of event cameras, this approximation can be obtained directly by creating an image of log intensity changes. Approximating Eq. (4) with 3DGS, however, requires additional modifications to the training process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 412, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 665 + ], + "type": "text", + "content": "Method Overview In its original formulation, 3D Gaussian Splitting uses a sequence of posed images as its training data. To make Gaussian Splitting applicable to event data, we introduce several modifications to the original method: First, we approximate Eq. (4) by accumulating events into an image representation " + }, + { + "bbox": [ + 313, + 412, + 555, + 665 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 412, + 555, + 665 + ], + "type": "text", + "content": " using a random sub-trajectory of the given event stream " + }, + { + "bbox": [ + 313, + 412, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 412, + 555, + 665 + ], + "type": "text", + "content": " (Sec. 4.1). This representation contains the sum of quantized log-intensity changes over the sub-trajectory. Second, we compute the corresponding log-intensity changes by rasterizing two views from the Gaussian scene representation (Sec. 4.2). For color cameras, this additionally requires a remosaicing step to account for the fact that the accumulated image representation of events only contains single color information for each pixel. To improve the 3DGS optimization process, we leverage an event-to-video guided structure from motion (Sec. 4.3). We also leverage cubic spline interpolation to more accurately estimate camera sub-trajectories as the rate at which the camera is tracked may be too coarse (Sec. 4.4). The full scene representation learning process is illustrated in Fig. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 672, + 433, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 433, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 433, + 683 + ], + "type": "text", + "content": "4.1. Event Accumulation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "To obtain an image of accumulated events, we first randomly sample start and endpoints " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{start}}, k_{\\mathrm{end}} \\in \\{1 \\dots N\\}" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26878" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 71, + 521, + 314 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 521, + 314 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 521, + 314 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 521, + 314 + ], + "type": "image", + "image_path": "e4d59241728ca5cd360114d94235ead907adc107a7b1c2f034d3facfe022c479.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "text", + "content": "Figure 2. Overview of our 3D Gaussian Splitting training with moving event camera data. Event data streams from " + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "inline_equation", + "content": "t_{k_{start}}" + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "inline_equation", + "content": "t_{k_{end}}" + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "text", + "content": " are accumulated into " + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "text", + "content": ", and distill the point priors from an event-to-video model. The log-difference image " + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 54, + 322, + 555, + 368 + ], + "type": "text", + "content": " is obtained from 3D Gaussian Splitting and Rasterization at two camera view points. It is computed as in Eq. (6) and remosaicing is performed in case of a color event camera. The respective poses are estimated by cubic spline interpolation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "text", + "content": "from our trajectory sequence. The lengths of those sequences are chosen randomly with a maximum length ranging from " + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "text", + "content": " of the total number of events. This allows for considering scene information at different scales in order to have both global and local geometry information trained. Then, we obtain the entries of the accumulated log-intensity change image " + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "text", + "content": " at pixel location " + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "inline_equation", + "content": "x,y" + }, + { + "bbox": [ + 54, + 387, + 294, + 471 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 478, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 478, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 115, + 478, + 295, + 510 + ], + "type": "interline_equation", + "content": "D _ {x, y} := \\sum_ {\\substack {k \\in \\{k _ {\\text {start}}, \\dots , k _ {\\text {end}} \\} \\\\ x _ {k} = x, y _ {k} = y}} p _ {k} \\Delta . \\tag{5}", + "image_path": "cd7d214f75617525c5eed0a190d4b4a1371ff4658e26c19f82af85ecd615317b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "content": "The resulting image serves as an approximation to Eq. (4), i.e. " + }, + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "inline_equation", + "content": "D \\approx E(t_{\\mathrm{start}}, t_{\\mathrm{end}})" + }, + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "content": ". This method can easily be adapted to the case where the event-threshold " + }, + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "content": " is asymmetric, i.e., when it is different for positive and negative polarity events." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 571, + 285, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 285, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 285, + 585 + ], + "type": "text", + "content": "4.2. Training View Generation and Remosaicing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "text", + "content": "To generate the log-difference image " + }, + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\hat{D} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "text", + "content": ", we compute" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 144, + 611, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 611, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 144, + 611, + 295, + 624 + ], + "type": "interline_equation", + "content": "\\hat {D} := \\hat {L} _ {2} - \\hat {L} _ {1} \\tag {6}", + "image_path": "a15bd9f8e5636ce2a2a181f54064aec5f1f234d00d1f5c27398ea683018f91e5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "which requires the remosaiced log images " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{1}, L_{2} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " at the poses " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "P_{k_{\\mathrm{start}}}^{\\prime}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "P_{k_{\\mathrm{end}}}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " respectively. These poses are the same that were used for generating our target " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ". To obtain " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " we first use the rasterizer to generate corresponding views " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{1}, I_{2} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " from the current Gaussian scene representation. Next, we perform a standard remosaicing operation because events are triggered per pixel" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 387, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 411 + ], + "type": "text", + "content": "asynchronously, not allowing us to use conventional demosaicing methods [5, 29, 35, 36, 41, 44]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 416, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 416, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 359, + 416, + 553, + 430 + ], + "type": "interline_equation", + "content": "\\text {R e m o s a i c i n g}: \\mathbb {R} ^ {H \\times W \\times 3} \\rightarrow \\mathbb {R} ^ {H \\times W} \\tag {7}", + "image_path": "4782e860e43f4247adcb45d1f5b73c289877c9397367f0f6dc80a380c473ca5b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": "which reintroduces the Bayer RGB pattern. It consists of two steps: First, a set of color-channel specific matrices of size " + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": " are Hadamard multiplied with each " + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": " pixel block in the image. The channel specific matrices " + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "inline_equation", + "content": "R, G, B \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": " are given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 502, + 553, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 502, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 331, + 502, + 553, + 530 + ], + "type": "interline_equation", + "content": "R = \\left[ \\begin{array}{l l} 1 & 0 \\\\ 0 & 0 \\end{array} \\right], \\quad G = \\left[ \\begin{array}{l l} 0 & 1 \\\\ 1 & 0 \\end{array} \\right], \\quad B = \\left[ \\begin{array}{l l} 0 & 0 \\\\ 0 & 1 \\end{array} \\right]. \\tag {8}", + "image_path": "f3ff0740148a611108097faca14df167ed12d6a0a1847ee4f736203993f98723.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "content": "The thus modified channels have non-zero entries at non-overlapping pixel locations. From this, the single-channel remosaiced image is obtained by addition across the channels. Finally, entry-wise logarithm computation yields the desired images " + }, + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\hat{L}_1, \\hat{L}_2" + }, + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "content": " and thus " + }, + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 600, + 508, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 508, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 508, + 612 + ], + "type": "text", + "content": "4.3. Event-to-Video Guided Initialization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "Event cameras capture only the pixels that indicate changes in brightness. This results in sparse information that is often insufficient for effective supervision. Thus, reconstructing 3D information, including background details, solely from event data is exceedingly challenging. To simplify the optimization process we propose leveraging the prior knowledge encoded in an event-to-video model. We use this pretrained model to generate images from event streams that" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26879" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "can subsequently be processed by Structure from Motion (SfM) techniques. Although the generated images often contain significant noise and do not accurately reproduce the underlying RGB image values, they retain substantial texture and background information. This information is sufficient to provide initial positions for the Gaussians, improve the 3DGS optimization process and lead to faster convergence compared to basic initialization with random initial points." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 263, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 263, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 263, + 201 + ], + "type": "text", + "content": "4.4. Event Camera Trajectory Interpolation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "text", + "content": "The camera poses, " + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "inline_equation", + "content": "P_{k_{\\mathrm{start}}}" + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "inline_equation", + "content": "P_{k_{\\mathrm{end}}}" + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "text", + "content": ", used to obtain the log images " + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "inline_equation", + "content": "\\hat{L}_1" + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "inline_equation", + "content": "\\hat{L}_2" + }, + { + "bbox": [ + 55, + 204, + 295, + 384 + ], + "type": "text", + "content": " in Eq. (6), are in most cases not given a priori. This is because event camera poses are typically obtained from a tracking or motion estimation system that operates at a fixed rate, which may not be aligned with the rate of event accumulation. Thus, to obtain an estimate of the complete trajectory of the event camera, we adopt an efficient trajectory interpolation method for camera poses. Our work utilizes Cubic Spline Interpolation and Spherical Cubic Spline Interpolation. These generate a smooth path by interpolating between discrete command points using third-degree polynomials. This interpolation technique closely approximates real-world camera motion by maintaining non-linear continuity in both velocity and acceleration of camera movement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 144, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 144, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 144, + 403 + ], + "type": "text", + "content": "4.5. Loss Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 409, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 458 + ], + "type": "text", + "content": "3D Gaussian Splitting is updated using a loss that combines a " + }, + { + "bbox": [ + 55, + 409, + 295, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 409, + 295, + 458 + ], + "type": "text", + "content": " reconstruction term and a perceptual term " + }, + { + "bbox": [ + 55, + 409, + 295, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 55, + 409, + 295, + 458 + ], + "type": "text", + "content": " that is based on the Structural Similarity Index Measure (SSIM) [67]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 466, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 466, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 119, + 466, + 295, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {S S I M}}. \\tag {9}", + "image_path": "905ea33124230ce2d8b56569efa001efed4464bdfe169047df53688c9b337c16.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "type": "text", + "content": "To fully utilize real-world event camera data, undistortion operation is performed as the accumulated image " + }, + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "type": "text", + "content": " is computed. We account for this by only incurring the loss where events were actually accumulated. This allows supervising the 3D representation more efficient." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 118, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 118, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 118, + 588 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Our evaluation of 3D Gaussian Splating from a moving event camera encompasses both synthetic and real-world scenes. Our method assumes that we can access the intrinsic camera matrix, its lens distortion parameters, and the camera's constant-rate poses, sampled at a high frequency to ensure precise interpolation for the event camera at any chosen time point. Estimation of camera poses at any given timestep is crucial for precise scene reconstruction from event data, as events can asynchronously occur at arbitrary timesteps between the constant-rate sampled camera poses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 413, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 413, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 413, + 84 + ], + "type": "text", + "content": "5.1.1. Synthetic Scenes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 87, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 553, + 206 + ], + "type": "text", + "content": "The synthetic dataset used in our experiments consists of event data streams obtained from ESIM [55], initially introduced in [42], encompassing seven synthetic scenes with a white background. These simulated event data streams feature a diverse range of photorealistic objects with complex structures. The event camera in these synthetic scenes moves around the objects, providing diverse geometric and lighting conditions. This setup allows for a comprehensive evaluation of the robustness of our method against the baselines in a controlled environment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 214, + 423, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 423, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 423, + 225 + ], + "type": "text", + "content": "5.1.2. Real-World Scenes" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 230, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 553, + 385 + ], + "type": "text", + "content": "The experiments on real scenes were conducted using subsets of the EDS [24], and TUM-VIE [30] datasets. From EDS, we utilize the sequences 03_rocket-earth DARK, 07_ziggy_and_fuzz hdr, 08_peanuts_runting, 11_all Characters and 13_airplane. From TUM-VIE, we utilize the sequences mocap-1d-trans and mocap-desk2. These sequences primarily capture various objects in 360-degree scenes and on a desk from a forward-facing perspective. These real sequences are relatively well-suited for novel view synthesis and were recorded using the high-resolution event sensor, the Prophesee Gen 3 in the EDS and the Prophesee Gen 4 in the TUM-VIE dataset, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 393, + 380, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 380, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 380, + 405 + ], + "type": "text", + "content": "5.2. Baselines" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 496 + ], + "type": "text", + "content": "We benchmarked our method against Robust-e-NeRF and E2VID [56] + 3D Gaussian Splatting (E2VID+3DGS). Robust-e-NeRF incorporates InstantNGP, which is significantly faster compared to traditional NeRF methods, but still falls short compared to 3DGS. E2VID generates RGB images from an event data stream, followed by feeding the images into original 3D Gaussian Splatting." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 504, + 373, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 373, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 373, + 516 + ], + "type": "text", + "content": "5.3. Metrics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 664 + ], + "type": "text", + "content": "We utilized commonly used metrics to evaluate novel view synthesis, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [79]) with VGG16, to comprehensively assess the quality of rendered images. It's worth noting that SSIM and LPIPS are perceptual metrics introduced to align more closely with human preference, with LPIPS specifically tailored to mimic human perception. These metrics address issues associated with PSNR, such as favoring blurry images. Additionally, we measured and compared the rendering speeds of our method, Robust-e-NeRF, and E2VID+3DGS." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Since event cameras capture variations in log-radiance rather than absolute log-radiance values, the predicted intensity " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " from the 3D Gaussian Splatting has an unknown offset. To rectify this limitation, a linear color transforma" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26880" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 70, + 212, + 178 + ], + "blocks": [ + { + "bbox": [ + 114, + 70, + 212, + 178 + ], + "lines": [ + { + "bbox": [ + 114, + 70, + 212, + 178 + ], + "spans": [ + { + "bbox": [ + 114, + 70, + 212, + 178 + ], + "type": "image", + "image_path": "ff41e83e35d3082964b1d049335c78f10590a2c016dda0f0001fd765c3787db6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 70, + 310, + 178 + ], + "blocks": [ + { + "bbox": [ + 224, + 70, + 310, + 178 + ], + "lines": [ + { + "bbox": [ + 224, + 70, + 310, + 178 + ], + "spans": [ + { + "bbox": [ + 224, + 70, + 310, + 178 + ], + "type": "image", + "image_path": "3ea78d5915b7cfc9becea1d7f466e433b03771de3e1406ed67bd9baaabb9895a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 403, + 177 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 403, + 177 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 403, + 177 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 403, + 177 + ], + "type": "image", + "image_path": "ebfbb3e6d9f319ac30322afd9e0c2b6485d5b7cbcd86723fb4600dd95d4f03bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 414, + 71, + 497, + 175 + ], + "blocks": [ + { + "bbox": [ + 414, + 71, + 497, + 175 + ], + "lines": [ + { + "bbox": [ + 414, + 71, + 497, + 175 + ], + "spans": [ + { + "bbox": [ + 414, + 71, + 497, + 175 + ], + "type": "image", + "image_path": "1c4708c3bf0d52f5e7b0c53cf19bbbf1d0d00339385252570f44fe1e5e99e64c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 114, + 178, + 211, + 273 + ], + "blocks": [ + { + "bbox": [ + 114, + 178, + 211, + 273 + ], + "lines": [ + { + "bbox": [ + 114, + 178, + 211, + 273 + ], + "spans": [ + { + "bbox": [ + 114, + 178, + 211, + 273 + ], + "type": "image", + "image_path": "8df9fa988fa5ff3c9d93ac20a141a6bf908da482db1ef55ae34a8ed8c5d8d6bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 184, + 309, + 273 + ], + "blocks": [ + { + "bbox": [ + 216, + 184, + 309, + 273 + ], + "lines": [ + { + "bbox": [ + 216, + 184, + 309, + 273 + ], + "spans": [ + { + "bbox": [ + 216, + 184, + 309, + 273 + ], + "type": "image", + "image_path": "8b7a59d98c2daf99711123f734a20545a40404b012b4e2df05b4b091e80cf054.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 185, + 403, + 273 + ], + "blocks": [ + { + "bbox": [ + 310, + 185, + 403, + 273 + ], + "lines": [ + { + "bbox": [ + 310, + 185, + 403, + 273 + ], + "spans": [ + { + "bbox": [ + 310, + 185, + 403, + 273 + ], + "type": "image", + "image_path": "124daa59c57e4ac9db3e436aabbd67e938035323e55d3d7390fdbcfd154ba6ef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 405, + 185, + 497, + 270 + ], + "blocks": [ + { + "bbox": [ + 405, + 185, + 497, + 270 + ], + "lines": [ + { + "bbox": [ + 405, + 185, + 497, + 270 + ], + "spans": [ + { + "bbox": [ + 405, + 185, + 497, + 270 + ], + "type": "image", + "image_path": "280db115ed57b244ec055245155afca539b8340f4350bb561af760d014db1403.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 127, + 274, + 203, + 370 + ], + "blocks": [ + { + "bbox": [ + 127, + 274, + 203, + 370 + ], + "lines": [ + { + "bbox": [ + 127, + 274, + 203, + 370 + ], + "spans": [ + { + "bbox": [ + 127, + 274, + 203, + 370 + ], + "type": "image", + "image_path": "2cb8279c9c93689ae7441c7b8cf9afdf2b3db86f23518f94e3c02bf77048385f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 379, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 54, + 379, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 555, + 413 + ], + "type": "text", + "content": "Figure 3. Generated Synthetic images comparing our work, event-based NeRF, and E2VID+3DGS qualitatively, with rendering times shown at the bottom of each image. Our work and Robust-e-NeRF outperform others in most scenes. In some cases, e.g., the drums scene, our work performs better and achieves significantly faster rendering times than conventional Robust-e-NeRF composed of InstantNGP." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 221, + 274, + 308, + 369 + ], + "blocks": [ + { + "bbox": [ + 221, + 274, + 308, + 369 + ], + "lines": [ + { + "bbox": [ + 221, + 274, + 308, + 369 + ], + "spans": [ + { + "bbox": [ + 221, + 274, + 308, + 369 + ], + "type": "image", + "image_path": "0e1e0d057cc4fc9f9336589f445bbedeb41595c96c9d4d73ee9e428154f8922e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 274, + 403, + 370 + ], + "blocks": [ + { + "bbox": [ + 317, + 274, + 403, + 370 + ], + "lines": [ + { + "bbox": [ + 317, + 274, + 403, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 403, + 370 + ], + "type": "image", + "image_path": "76b81c8e4856c97ec9051b464f57ea7c7df7ca5e91572577eb07ba04b643a458.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 411, + 274, + 496, + 370 + ], + "blocks": [ + { + "bbox": [ + 411, + 274, + 496, + 370 + ], + "lines": [ + { + "bbox": [ + 411, + 274, + 496, + 370 + ], + "spans": [ + { + "bbox": [ + 411, + 274, + 496, + 370 + ], + "type": "image", + "image_path": "73c0744d31d88dad680de59f351631504f90c181dd6a77b373df1785bcad338b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 92, + 455, + 517, + 616 + ], + "blocks": [ + { + "bbox": [ + 55, + 423, + 555, + 448 + ], + "lines": [ + { + "bbox": [ + 55, + 423, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 555, + 448 + ], + "type": "text", + "content": "Table 1. Generated synthetic scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively. We supervised our model with randomly initialized points." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 455, + 517, + 616 + ], + "lines": [ + { + "bbox": [ + 92, + 455, + 517, + 616 + ], + "spans": [ + { + "bbox": [ + 92, + 455, + 517, + 616 + ], + "type": "table", + "html": "
MetricMethodSynthetic SceneMean
chairdrumsficushotdoglegomaterialsmic
PSNR ↑E2VID + 3DGS21.3919.8619.9015.5518.1720.0820.1019.29
Robust e-NeRF30.2423.1530.7128.0727.3424.9832.8728.19
Ours28.6925.8129.9022.9129.2227.1633.2728.14
SSIM ↑E2VID + 3DGS0.9340.9150.9220.8970.8950.9010.9570.917
Robust e-NeRF0.9580.8970.9710.9530.9340.9230.9810.945
Ours0.9530.9470.9660.9400.9450.9360.9860.953
LPIPS ↓E2VID + 3DGS0.0760.0940.1080.2080.1450.1250.0690.118
Robust e-NeRF0.0400.0910.0220.0950.0740.0520.0290.057
Ours0.0470.0520.0280.0980.0550.0600.0150.051
", + "image_path": "d925b4fec6fe0113c2aef691e6249af3b68d560bdfcdabb7bc29d507418be5c4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 635, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 706 + ], + "type": "text", + "content": "tion is designed to adjust our predictions in the logarithmic domain [42, 59]. This transformation is both necessary and adequately effective for aligning our predictions with the reference data. It ensures that the reconstructed intensity values are properly calibrated and aligned with the observed event data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 634, + 446, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 634, + 446, + 646 + ], + "spans": [ + { + "bbox": [ + 314, + 634, + 446, + 646 + ], + "type": "text", + "content": "5.4. Results and Evaluation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "After completing the training of our model on each scene using event sequence data, the model acquires a set of 3D Gaussian ellipsoids representing the scene. By rasterizing these Gaussians, we are able to render various arbitrary views of each scene in real-time. The renderings of novel" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26881" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 69, + 480, + 452 + ], + "blocks": [ + { + "bbox": [ + 130, + 69, + 480, + 452 + ], + "lines": [ + { + "bbox": [ + 130, + 69, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 69, + 480, + 452 + ], + "type": "image", + "image_path": "a96c47fa963139724dad03565260fb3b2b24182a37e492b97be4bfc4df3960a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 459, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 459, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 555, + 492 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons of the images generated by our method, event-based NeRF and E2VID+3DGS show that our technique appears to recover details in 5 real scenes. First three rows(07_ziggy_and_fuzz_hdr, 08_peanuts_runting and 11_all Characters) are EDS dataset and last two rows are TUM-VIE dataset(mocap-1d-trans and mocap-desk2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 113, + 523, + 495, + 684 + ], + "blocks": [ + { + "bbox": [ + 93, + 504, + 517, + 516 + ], + "lines": [ + { + "bbox": [ + 93, + 504, + 517, + 516 + ], + "spans": [ + { + "bbox": [ + 93, + 504, + 517, + 516 + ], + "type": "text", + "content": "Table 2. Real scenes in comparison between our work, previous event-based NeRF and E2VID+3DGS quantitatively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 113, + 523, + 495, + 684 + ], + "lines": [ + { + "bbox": [ + 113, + 523, + 495, + 684 + ], + "spans": [ + { + "bbox": [ + 113, + 523, + 495, + 684 + ], + "type": "table", + "html": "
MetricMethodReal SceneMean
0307081113
PSNR ↑E2VID + 3DGS15.6715.0514.0313.8318.9615.51
Robust e-NeRF19.1914.7814.7514.4318.1016.25
Ours20.7819.1417.5317.7919.0518.86
SSIM ↑E2VID + 3DGS0.7160.6890.6420.6910.7230.692
Robust e-NeRF0.8460.8150.7350.5690.7290.739
Ours0.8350.8160.7450.7890.7740.792
LPIPS ↓E2VID + 3DGS0.2660.3780.4020.4150.4150.375
Robust e-NeRF0.3240.4760.5670.7000.6500.543
Ours0.2390.3510.4240.3910.4070.363
", + "image_path": "8f263ba7060235efb97aa7892dbed32c35ceb062e9f95c65b0f89cba0ce5e1d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26882" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "views, obtained from held-out views or test views provided by the datasets, are then used for evaluations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 155, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 155, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 155, + 114 + ], + "type": "text", + "content": "5.4.1. Synthetic Scenes" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 295, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 332 + ], + "type": "text", + "content": "We evaluate our methodology on a synthetic dataset using the metrics described in Sec. 5.3. Quantitative results are shown in Tab. 1, and visual samples and rendering times in Fig. 3. Rendering times are provided in milliseconds (ms). Our method achieves higher average SSIM and compared to previous works. A high SSIM value indicates structural similarity between generated and reference images, underscoring our model's accuracy in capturing essential scene geometry. Additionally, a low LPIPS score reflects perceptual quality, showing the produced images are both realistic and visually engaging. The elevated scores demonstrate that our novel view synthesis model faithfully reproduces the visual and perceptual attributes of the original scene. Moreover, the rendering speed of both our method and E2VID+3DGS significantly surpass that of Robust-e-NeRF. Therefore, our work maintains superior quality and speed compared to Robust-e-NeRF, highlighting its effectiveness in both fidelity and performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 333, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 295, + 380 + ], + "type": "text", + "content": "As shown in Fig. 3, the quality of rendered images remains high in scenes featuring drums and microphones, achieving novel view synthesis without sacrificing quality in other scenes as well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 386, + 134, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 134, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 134, + 397 + ], + "type": "text", + "content": "5.4.2. Real Scenes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 581 + ], + "type": "text", + "content": "The EDS dataset is the only one available for quantifying the performance of novel view synthesis. Therefore, we primarily conducted quantitative and qualitative assessments, as shown Tab. 2 and Fig. 4. Our method successfully recovers fine details in all sequences and contrasts in the mocap-1d-trans image, particularly. This is particularly noticeable in the images containing the checkerboard and the cover of a book. Additionally, it effectively smooths uniform areas such as the checkerboard and the desk surface in the mocap-1d-trans and mocap-desk2. This smoothness likely results from the accumulation of event streams, which may have enhanced the signal-to-noise ratio due to the aggregation process. The visible artifacts near the borders of all synthesized views are attributed to the relatively narrow field of view of the event cameras." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 124, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 124, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 124, + 600 + ], + "type": "text", + "content": "5.5. Ablations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "To demonstrate the effectiveness of event-to-video guided initialization and camera trajectory estimation with cubic spline interpolation, we conducted ablation experiments by separately adding 1) event-to-video guided initialization and 2) both the initialization and cubic spline interpolation. As shown in Tab. 3, event-to-video guided initialization outperforms random initialization. Furthermore, Fig. 5 shows that our initialization approach results in better image quality (e.g. backgrounds) compared to random initialization." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 324, + 102, + 543, + 159 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "text", + "content": "Table 3. Ablation results for the event-to-video guided initialization and cubic spline interpolation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 102, + 543, + 159 + ], + "lines": [ + { + "bbox": [ + 324, + 102, + 543, + 159 + ], + "spans": [ + { + "bbox": [ + 324, + 102, + 543, + 159 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
random init18.020.7670.397
guided init18.750.7880.359
guided init + cubic18.860.7920.363
", + "image_path": "7c1b53cc01b2c9a582b59909a01f73bb1645c23ca9fbba174908b95568e6c17c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 180, + 433, + 270 + ], + "blocks": [ + { + "bbox": [ + 347, + 170, + 403, + 179 + ], + "lines": [ + { + "bbox": [ + 347, + 170, + 403, + 179 + ], + "spans": [ + { + "bbox": [ + 347, + 170, + 403, + 179 + ], + "type": "text", + "content": "Random init" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 180, + 433, + 270 + ], + "lines": [ + { + "bbox": [ + 316, + 180, + 433, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 433, + 270 + ], + "type": "image", + "image_path": "1e7e67edf5eaab2048813d2c5381f29caadb59159351eacd977bf1d6e59f1367.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 433, + 180, + 552, + 270 + ], + "blocks": [ + { + "bbox": [ + 434, + 169, + 552, + 179 + ], + "lines": [ + { + "bbox": [ + 434, + 169, + 552, + 179 + ], + "spans": [ + { + "bbox": [ + 434, + 169, + 552, + 179 + ], + "type": "text", + "content": "Event-to-Video Guided init" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 180, + 552, + 270 + ], + "lines": [ + { + "bbox": [ + 433, + 180, + 552, + 270 + ], + "spans": [ + { + "bbox": [ + 433, + 180, + 552, + 270 + ], + "type": "image", + "image_path": "c7826f7c045ab390dc83d260299516fa74f494d68ddef06c024bdce3cda7ae09.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 280, + 554, + 313 + ], + "lines": [ + { + "bbox": [ + 313, + 280, + 554, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 554, + 313 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of the images generated by random initialization and event-to-video guided initialization are shown in 07_ziggy_and_fuzz hdr scene in EDS dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "text", + "content": "Adding cubic spline interpolation further enhances performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 365, + 391, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 365, + 391, + 377 + ], + "spans": [ + { + "bbox": [ + 314, + 365, + 391, + 377 + ], + "type": "text", + "content": "5.6. Limitations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 383, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 528 + ], + "type": "text", + "content": "Our method, though effective, has some limitations. It is designed exclusively for novel view synthesis in static scenes and cannot handle scenarios with dynamic moving objects, which we consider to be an exciting avenue for future research. Additionally, our approach trains 3D Gaussian Splatting on event-accumulated images between two viewpoints, representing relative intensity images. Consequently, similar to other event-based view synthesis methods, the model cannot directly estimate absolute intensity images and requires a linear transformation with evaluation data as a reference. This linear transformation is required only during inference and does not impact the training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 538, + 479, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 479, + 550 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": "In this work, we introduce a novel method for 3D Gaussian Splatting from moving event cameras. Our approach achieves the higher visual fidelity and better performance of novel view synthesis in static scenes. We distill geometric priors from the event-to-video model for initialization and leverage effective camera trajectory interpolation. Thus, EventSplat yields improved metrics compared to existing event-based NeRF approaches. Moreover, the incorporation of 3D Gaussian Splatting enables our model to render scenes at significantly higher frame rates. This enhances the practicality of real-time novel view synthesis with event cameras. Our work is thus particularly useful in scenarios where conventional imaging techniques are inadequate." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26883" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "content": "The work is in part supported by Sony. We would also like to thank Ziyi Wu, Yash Kant, and Umangi Jain, for valuable discussions and support." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 235 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 267 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In ICCV, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 269, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 295, + 300 + ], + "type": "text", + "content": "[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. TNNLS, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "type": "text", + "content": "[5] Hong Cao and Alex C Kot. Accurate detection of demosaicing regularity for digital image forensics. IEEE Trans. Inf. Forensics Secur, 2009. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 369, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 391 + ], + "type": "text", + "content": "[7] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "text", + "content": "[8] Jiaben Chen, Yichen Zhu, Dongze Lian, Jiaqi Yang, Yifu Wang, Renrui Zhang, Xinhang Liu, Shenhan Qian, Laurent Kneip, and Shenghua Gao. Revisiting Event-Based Video Frame Interpolation. In IROS, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 437, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 479 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In arXiv, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "text", + "content": "[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3D using Gaussian Splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "text", + "content": "[11] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "type": "text", + "content": "[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "text", + "content": "[13] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Sci. Robot, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[14] Guillermo Gallego, Jon EA Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-based, 6-dof camera tracking from photometric depth maps. TPAMI, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. PAMI, 2022. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "type": "text", + "content": "[17] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. FastNeRF: High-Fidelity Neural Rendering at 200FPS. In ICCV, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 163, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 195 + ], + "type": "text", + "content": "[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Asynchronous, photometric feature tracking using events and frames. In ECCV, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 197, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 554, + 239 + ], + "type": "text", + "content": "[19] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[20] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. IJCV, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "text", + "content": "[21] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "text", + "content": "[22] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 3DV, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[23] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 388, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 430 + ], + "type": "text", + "content": "[24] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 432, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 453 + ], + "type": "text", + "content": "[25] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-NeRF: Event based neural radiance field. In WACV, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "text", + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ToG, 2023. 1, 2, 3, 12" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 489, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 520 + ], + "type": "text", + "content": "[27] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In CVPR, 2016. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[28] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In ICCV, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "type": "text", + "content": "[29] Ron Kimmel. Demosaicing: Image reconstruction from color CCD samples. IEEE Trans. Image Process, 1999. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "text", + "content": "[30] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM Stereo Visual-Inertial Event Dataset. In IROS, 2021. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[31] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. RA-L, 2023. 1, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[32] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. AdaNeRF: Adaptive sampling for real-time rendering of neural radiance fields. In ECCV, 2022. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[33] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient Sphere-Based Neural Rendering. In CVPR, 2021. 1" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26884" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[34] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 139 + ], + "type": "text", + "content": "[35] Xin Li. Demosaicing by successive approximation. IEEE Trans. Image Process, 2005. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "type": "text", + "content": "[36] Xin Li, Bahadir Gunturk, and Lei Zhang. Image demosaic-ing: A systematic survey. In VCIP. SPIE, 2008. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "text", + "content": "[37] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. JSSC, 2008. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[38] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning Event-Driven Video Deblurring and Interpolation. In ECCV, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[39] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. BACON: Band-limited coordinate networks for multiscale scene representation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "type": "text", + "content": "[40] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "text", + "content": "[41] Philippe Longere, Xuemei Zhang, Peter B Delahunt, and David H Brainard. Perceptual assessment of demosaicing algorithm performance. Proceedings of the IEEE, 2002. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 355, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 388 + ], + "type": "text", + "content": "[42] Weng Fei Low and Gim Hee Lee. Robust e-NeRF: NeRF from Sparse & Noisy Events under Non-Uniform Motion, 2023. 1, 2, 5, 6, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "text", + "content": "[43] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V. Sander. Deblur-NeRF: Neural Radiance Fields From Blurry Images. In CVPR, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 423, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 456 + ], + "type": "text", + "content": "[44] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, 2004. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 456, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 489 + ], + "type": "text", + "content": "[45] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 491, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 533 + ], + "type": "text", + "content": "[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: representing scenes as neural radiance fields for view synthesis. ECCV, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "type": "text", + "content": "[47] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. NeRF in the Dark: High Dynamic Range View Synthesis From Noisy Raw Images. In CVPR, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "text", + "content": "[48] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G. Derpanis, Jonathan Kelly, Marcus A. Brubaker, Igor Gilitschenski, and Alex Levinshtein. SPIn-NeRF: Multiview Segmentation and Perceptual Inpainting With Neural Radiance Fields. In CVPR, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "text", + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[50] Marc Osswald, Sio-Hoi Ieng, Ryad Benosman, and Giacomo Indiveri. A spiking neural network model of 3d perception for event-based neuromorphic stereo vision systems. Sci. Rep., 2017. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[51] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[52] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurlPS, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "type": "text", + "content": "[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "type": "text", + "content": "[54] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. RA-L, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 231, + 553, + 262 + ], + "type": "text", + "content": "[55] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In CoRL. PMLR, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 553, + 297 + ], + "type": "text", + "content": "[56] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. PAMI, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 315, + 299, + 553, + 331 + ], + "type": "text", + "content": "[57] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 333, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 333, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 333, + 553, + 388 + ], + "type": "text", + "content": "[58] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. In arXiv, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 553, + 432 + ], + "type": "text", + "content": "[59] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. EventNeRF: Neural radiance fields from a single colour event camera. In CVPR, 2023. 1, 2, 6, 12" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 553, + 466 + ], + "type": "text", + "content": "[60] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 315, + 468, + 553, + 511 + ], + "type": "text", + "content": "[61] Amos Sironi, Manuele Brambilla, Nicolas Bourdis, Xavier Lagorce, and Ryad Benosman. HATS: Histograms of averaged time surfaces for robust event-based object classification. In CVPR, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 315, + 513, + 553, + 567 + ], + "type": "text", + "content": "[62] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 315, + 568, + 553, + 612 + ], + "type": "text", + "content": "[63] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, 2021. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 613, + 553, + 667 + ], + "type": "text", + "content": "[64] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-Based Frame Interpolation With Parametric Non-Linear Flow and Multi-Scale Fusion. In CVPR, 2022. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 669, + 553, + 713 + ], + "type": "text", + "content": "[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26885" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[66] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "text", + "content": "[67] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale Structural Similarity for Image Quality Assessment. In ACSSC, 2003. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[68] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 240 + ], + "type": "text", + "content": "[69] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[70] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. arXiv preprint arXiv:2407.11343, 2024. 2, 12, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "text", + "content": "[71] Ziyi Wu, Mathias Gehrig, Qing Lyu, Xudong Liu, and Igor Gilitschenski. Leod: Label-efficient object detection for event cameras. arXiv preprint arXiv:2311.17286, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "text", + "content": "[72] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "text", + "content": "[73] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In ICLR, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "text", + "content": "[74] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In SIGGRAPH, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 509 + ], + "type": "text", + "content": "[75] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 546, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 588 + ], + "type": "text", + "content": "[77] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 591, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 633 + ], + "type": "text", + "content": "[78] Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S. Ren. Training weakly supervised video frame interpolation with events. In ICCV, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "text", + "content": "[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[80] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 230 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[81] Yi Zhou, Guillermo Gallego, Henri Rebecq, Laurent Kneip, Hongdong Li, and Davide Scaramuzza. Semi-Dense 3D Reconstruction with a Stereo Event Camera. In ECCV, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 139 + ], + "type": "text", + "content": "[82] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. T-RO, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[83] Alex Zihao Zhu and Liangzhe Yuan. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In RSS, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[84] Baozhou Zhu, Peter Hofstee, Johan Peltenburg, Jinho Lee, and Zaid Alars. AutoReCon: Neural Architecture Search-based Reconstruction for Data-free Compression, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 230 + ], + "type": "text", + "content": "[85] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. EWA volume splatting. In VIS, 2001. 2, 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26886" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_content_list.json b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..638ff3ca180bcfb00c48d6cfc22a4b1cc655c79e --- /dev/null +++ b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_content_list.json @@ -0,0 +1,1412 @@ +[ + { + "type": "text", + "text": "Every SAM Drop Counts: Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond", + "text_level": 1, + "bbox": [ + 205, + 130, + 794, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guanyao Wu†, Haoyu Liu†, Hongming Fu†, Yichuan Peng†, Jinyuan Liu‡, Xin Fan†, Risheng Liu†*", + "bbox": [ + 112, + 202, + 888, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†School of Software Technology, Dalian University of Technology, China", + "bbox": [ + 210, + 220, + 787, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\ddagger}$ School of Mechanical Engineering, Dalian University of Technology, China", + "bbox": [ + 197, + 239, + 800, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rollingplainko@gmail.com, atlantis918@hotmail.com, {xin.fan, rsliu}@dlut.edu.cn", + "bbox": [ + 147, + 258, + 844, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 309, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modality image fusion, particularly infrared and visible, plays a crucial role in integrating diverse modalities to enhance scene understanding. Although early research prioritized visual quality, preserving fine details and adapting to downstream tasks remains challenging. Recent approaches attempt task-specific design but rarely achieve \"The Best of Both Worlds\" due to inconsistent optimization goals. To address these issues, we propose a novel method that leverages the semantic knowledge from the Segment Anything Model (SAM) to Grow the quality of fusion results and Enable downstream task adaptability, namely SAGE. Specifically, we design a Semantic Persistent Attention (SPA) Module that efficiently maintains source information via the persistent repository while extracting high-level semantic priors from SAM. More importantly, to eliminate the impractical dependence on SAM during inference, we introduce a bi-level optimization-driven distillation mechanism with triplet losses, which allow the student network to effectively extract knowledge. Extensive experiments show that our method achieves a balance between high-quality visual results and downstream task adaptability while maintaining practical deployment efficiency. The code is available at https://github.com/RollingPlain/SAGE_IVIF.", + "bbox": [ + 86, + 340, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 732, + 220, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In light of recent advancements in sensing technologies, multi-modality imaging [15] has gained significant attention, with applications in robotics [35], remote sensing [33], and autonomous driving [1]. Integrating infrared and visible sensors provides substantial advantages for intelligent processing [5], although both exhibit inherent limitations [16, 21]. Infrared images are robust against smoke, obstruction, and low light, while visible images excel in reso", + "bbox": [ + 89, + 756, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/032986408fbfc16d793221c9253fc95c0278746d72dd3613c50c98de82ac55c4.jpg", + "image_caption": [ + "Figure 1. Differences between the proposed method and existing mainstream comparative approaches: (a) Traditional and early DL-based methods focus on the fusion visual effect. (b) Task-specific methods (e.g., TarDAL [11] & SegMiF [13]) introduce task loss and features that lead to inconsistent optimization goals, causing a conflict between visual and task accuracy. (c) Our pipeline first leverages semantic priors from SAM within a large network and then distills the knowledge into a smaller sub-network achieving practical inference feasibility while ensuring \"the best of both worlds\" through SAM's inherent adaptability to these tasks." + ], + "image_footnote": [], + "bbox": [ + 517, + 308, + 905, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lution, contrast, and texture detail. These characteristics are highly complementary, motivating their fusion to generate a comprehensive fused image that balances visual quality and downstream task requirements. Consequently, achieving this fusion efficiently remains a pressing challenge [15].", + "bbox": [ + 511, + 640, + 906, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional Infrared and Visible Image Fusion (IVIF) methods, based on information theory [17, 37], aim to retain as much source image information as possible but often struggle with optimizing fused image quality, especially in managing redundancy and specific scenarios. Early deep learning-based methods [8, 12, 18, 22, 32, 39, 40] focused on visualization but faced issues like edge blurring and artifacts, failing to meet downstream perception needs. Current methods [11, 13, 20, 26] focus on specific tasks, coupling fusion with downstream processing, resulting in conflicting optimization objectives and challenges in balancing visual quality with task adaptability.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17882", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74df534b7eea6c9ad180d230068b72d93948a1cd34ac0e8750cda437264a6211.jpg", + "image_caption": [ + "Figure 2. Demonstration of SAM's robustness in MFNet normal scenes (top) and under challenging FMB conditions (bottom)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in large-scale visual models have significantly improved various visual analysis tasks [30]. Among these, the Segment Anything Model [6] stands out due to its exceptional and robust capability to provide rich semantic information, making it well-suited for IVIF, as illustrated in Figure 2. Also, its segmentation capabilities align naturally with the requirements of downstream tasks in IVIF field. However, current methods that integrate SAM for low-level tasks typically require full SAM during inference, resulting in excessive practical infeasibility.", + "bbox": [ + 88, + 339, + 482, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose the fusion method SAGE that fully integrates and distills semantic priors derived from SAM. As illustrated in Figure 1 III, one of our aim is to harness the advantages of SAM's semantic priors while reducing computational burden. The proposed method consists of two key components: Semantic Persistent Attention module and a bi-level distillation scheme. The former focuses on using a persistent repository throughout the process to retain source image information, while heuristically enabling the network to integrate semantic patches predicted by SAM. Furthermore, the distillation scheme is designed to efficiently exclude SAM from the inference process, thereby reducing computational complexity. By enhancing distilled information across different aspects, the triplet losses in the scheme significantly boost the performance of the fusion model. Consequently, efficient fusion is achieved using only the distilled model, delivering high-quality visual results and precise task performance without the direct involvement of SAM. In summary, our contributions can be summarized as follows:", + "bbox": [ + 91, + 489, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel fusion framework is proposed to leverage semantic priors from the SAM, effectively balancing visual quality and downstream task adaptability..", + "- The designed Semantic Persistent Attention (SPA) module utilizes a persistent repository to efficiently retain source information and simultaneously extract high-level semantic representation." + ], + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a bi-level optimization distillation scheme that transfers the information processed by SPA into the sub-network, effectively decoupling the fusion process from SAM during inference.", + "- Extensive experiments on multiple datasets validate the superior effectiveness of our proposed method." + ], + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 195, + 663, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning based IVIF Methods. In recent years, deep learning based multi-modality image fusion approaches [8, 12, 32, 39, 40, 44] achieved significant progress, primarily focusing on improving fusion quality, such as enhancing image details, maintaining structural consistency, and reducing noise. These approaches, including innovations in adversarial learning, Transformer models and diffusion network, have provided new insights. However, most of these methods have not addressed how the fused images can directly contribute to downstream tasks such as object detection, semantic segmentation, and other perception-related applications. With the growing demand for more integrated solutions, this gap has sparked increasing research interest in combining fusion with various downstream tasks.", + "bbox": [ + 511, + 220, + 906, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works began to bridge this gap, with TarDAL [11] pioneering the use of a dual-discriminator network for object detection optimization, which simultaneously considers modality differences and detection tasks during fusion. As the field has progressed, several methods [24, 38] have also focused on optimizing fusion for object detection. For semantic segmentation [36], SeAFusion [26] and SegMiF [13] have made early significant improvements, with the former enhancing fusion via task loss and the latter leveraging task-specific features. More recently, some approaches [3, 14, 19, 20, 25, 29] have integrated multi-task learning to jointly optimize detection and segmentation, improving performance on both tasks.", + "bbox": [ + 511, + 431, + 908, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SAM & Its Application. The Segment Anything Model (SAM) [6] is a large-scale pre-trained foundation model that excels in zero-shot generalization, enabling it to handle a wide range of segmentation and detection tasks without the need for task-specific training. SAM demonstrates impressive performance in low-level vision applications, such as image deblurring [9], dehazing [4], super-resolution [28], and low-light image enhancement [7]. Moreover, SAM's capabilities extend to high-level tasks like object detection [23], which further strengthens its versatility. This ability to handle both low-level and high-level tasks makes SAM inherently well-suited for the needs of the IVIF field, where fusion based segmentation and detection tasks are critical. However, its impressive performance is hindered by high computational complexity, posing challenges for practical deployment. Addressing these computational burdens, while fully leveraging the semantic priors of SAM, is a primary motivation for our work.", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17883", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d350ee2af8cb773fdc2ef3c1c3ba9e565d318ee92292168bdf0165f0bb39a7b5.jpg", + "image_caption": [ + "Figure 3. A overall workflow of our proposed method. (a) shows the flow structure of the main network, where the SPA module processes patches with semantic priors generated by SAM. (b) illustrates the detailed structure of the SPA module, where PR plays a key role in preserving the source and integrating the semantic information. (c) displays our distillation scheme formulation, with visualizations of the different components of the triplet loss. (d) provides a simple diagram of the sub-network, which is composed of stacked dense blocks." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 624, + 320 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/322fa75d65a95cc9cc44b6529ffb30411e2ee5c2916d7b3334f1dc08309dc097.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 88, + 901, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 412, + 302, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 438, + 294, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall workflow of our method is illustrated in Figure 3. Our aim is to fully exploit the semantic priors of SAM during the inference stage to enhance cross-modal fusion quality. However, directly deploying a large-scale SAM model often leads to prohibitive computational overhead. To mitigate this, we adopt a knowledge distillation strategy, transferring the information encoded by the SAM-driven main network to a lightweight sub-network, thereby significantly reducing inference costs while maintaining high-quality fusion. Nevertheless, the substantial capacity gap between the SAM-enhanced main network and the compact sub-network frequently results in incomplete semantic transfer or structural inconsistencies, which hinder ideal cross-modal fusion performance. To address this issue, we propose a bi-level optimization framework that jointly optimizes both networks as a unified system, aiming to bridge the distillation gap and maintain consistent fusion guided by SAM's semantic priors.", + "bbox": [ + 88, + 460, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{I}_{\\mathrm{s}}$ denote the source visible and infrared input images, i.e., $\\mathbf{I}_{\\mathrm{s}} = \\{\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}}\\}$ . Their corresponding semantic-prior patches generated under the guidance of SAM are denoted as $\\mathbf{P}_{\\mathrm{s}} = \\{\\mathbf{P}_{\\mathrm{vis}},\\mathbf{P}_{\\mathrm{ir}}\\}$ . Denote the main network by $\\mathcal{M}$ with parameters $\\omega_{\\mathbf{m}}$ , and the sub-network by $\\mathcal{S}$ with parameters $\\omega_{\\mathbf{s}}$ . We define the optimization process as:", + "bbox": [ + 89, + 734, + 482, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\omega} _ {s}} \\mathcal {L} _ {\\mathrm {s}} ^ {\\mathrm {d}} \\left(\\mathcal {S} \\left(\\mathbf {I} _ {s}, \\boldsymbol {\\omega} _ {s}\\right); \\boldsymbol {\\omega} _ {m} ^ {*}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 830, + 480, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {s . t .} \\quad \\boldsymbol {\\omega} _ {m} ^ {*} \\in \\arg \\min _ {\\boldsymbol {\\omega} _ {m}} \\mathcal {L} _ {\\mathrm {m}} ^ {\\mathrm {d}} \\big (\\mathcal {M} (\\mathbf {I} _ {s}, \\mathbf {P} _ {s}, \\boldsymbol {\\omega} _ {m}); \\boldsymbol {\\omega} _ {s} \\big),\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 859, + 437, + 880 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{m}}^{\\mathrm{d}}$ is the distillation loss that ensures high-quality fusion with meaningful semantic cues in the main network, and $\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}}$ is the distillation loss guiding the subnetwork to effectively mimic the main network's behavior. The fused reference image is given by: $\\mathbf{I}_{\\mathrm{ref}} = \\mathcal{M}\\big(\\{\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}}\\} ,\\{\\mathbf{P}_{\\mathrm{vis}},\\mathbf{P}_{\\mathrm{ir}}\\} ,\\boldsymbol {\\omega}_m\\big)$ , while the sub-network produces a fused image $\\mathbf{I}_{\\mathrm{f}} = S\\big(\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}},\\boldsymbol {\\omega}_s\\big)$ . Notably, both distillation objectives rely on outputs from the counterpart network, establishing a bidirectional dependency, which is a key feature of our distillation formulation.", + "bbox": [ + 511, + 412, + 906, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Within the overall framework, the main network integrates SAM-derived semantic priors into the fusion process, while the sub-network is optimized to align with its outputs under the bi-level formulation. The following sections elaborate on two key components that support this process: the Semantic Persistent Attention (SPA) module 3.2, which integrates SAM-derived priors with source features, and the triplet-based loss scheme 3.3, which facilitates structured knowledge transfer across networks.", + "bbox": [ + 511, + 566, + 908, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantic Persistent Attention Module", + "text_level": 1, + "bbox": [ + 511, + 712, + 841, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To fully exploit the semantic information provided by SAM, we propose the Semantic Persistent Attention (SPA) module, as illustrated in the flowchart in Figure 3 (b).", + "bbox": [ + 511, + 734, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key core of the SPA module is a Persistent Repository (PR), which serves as a static memory to store and maintain crucial contextual information for the fusion process. Specifically, PR stores the latent representation $(\\mathbf{Z})$ of the source features $\\mathbf{F}_{\\mathrm{src}}$ and the corresponding key-value pairs $(\\mathbf{K}_{\\mathrm{src}},\\mathbf{V}_{\\mathrm{src}})$ , which provide consistent contextual support during the cross-attention operation. The core insight behind this design is that PR acts as a stable, modality-", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17884", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/001ea98476575670761444f88549e1308937d98346e53e799a311ce46dfd5492.jpg", + "image_caption": [ + "Figure 4. Qualitative demonstrations of SOTA approaches across commonly used datasets, including TNO, RoadScene, $\\mathbf{M}^3\\mathrm{FD}$ and FMB." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 898, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "specific information source, guiding the cross-attention mechanism to fuse the semantic patches with rich contextual details from the original source.", + "bbox": [ + 89, + 435, + 482, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visible and infrared semantic patches $(\\mathbf{P}_{\\mathrm{vis}}$ and $\\mathbf{P}_{\\mathrm{ir}}$ extracted from the SAM are encoded into features $(\\mathbf{F}_{\\mathrm{pvis}}$ and $\\mathbf{F}_{\\mathrm{pir}})$ . These encoded semantic features represent limited portions of the scene and are processed through the cross-attention mechanism. By utilizing the key-value pairs $(\\mathbf{K}_{\\mathrm{src}},\\mathbf{V}_{\\mathrm{src}})$ stored in PR, the cross-attention mechanism enriches the semantic queries $(\\mathbf{Q}_{\\mathrm{pvis}}$ and $\\mathbf{Q}_{\\mathrm{pir}})$ with the full scene context from the source image, thereby addressing the inherent incomplete scene coverage in the patches. PR ensures the fusion process stays grounded in a stable, consistent context, enabling patches to be effectively enriched with modality-specific information. This stability maintains semantic coherence while allowing flexible feature refinement via attention mechanisms.", + "bbox": [ + 89, + 484, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, the SPA module uses PR to guide the cross-attention mechanism, ensuring that the semantic patches from SAM are enriched with consistent and detailed modality-specific information. This approach allows the fusion of infrared and visible image modalities by combining both the high-level semantic understanding from SAM and the detailed information from the source features. The final output, $\\mathbf{F}_{\\mathrm{SPA}}$ , represents a semantically enriched, structurally consistent feature set, rich in semantic priors of SAM, and ready for further processing.", + "bbox": [ + 89, + 700, + 482, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The main network heavily utilizes the SPA module, aiming to fully exploit the complex semantic priors provided by SAM, focus on capturing and retaining detailed semantic", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "knowledge, thereby facilitating a powerful representation that can later be distilled into a more efficient sub-network.", + "bbox": [ + 511, + 435, + 903, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Triplet Loss in Distillation Scheme", + "text_level": 1, + "bbox": [ + 511, + 474, + 815, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce a triplet loss-driven distillation scheme that facilitates semantic knowledge transfer from the main network $(\\mathcal{M})$ to the sub-network $(S)$ under a bi-level optimization framework. Specifically, we adopt a DARTS-style [10] training protocol, where the main network (teacher) and the sub-network (student) are updated in small alternating steps, approximating the inner-outer structure of bilevel optimization. This allows gradients to flow in a bidirectional manner: the sub-network learns from the distillation signals provided by the main network, while the main network also adapts its parameters based on the subnetwork's performance and a segmentation objective. Consequently, the two networks reach a cohesive transfer with mutual compromise.", + "bbox": [ + 511, + 496, + 906, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our distillation scheme is driven by three distinct losses, each focusing on a different aspect of the fused images. Let $\\mathbf{I}_{\\mathrm{ref}}$ be the fused reference image produced by the main network, and $\\mathbf{I}_{\\mathrm{f}}$ be the fused output from the sub-network. We first introduce a feature alignment term:", + "bbox": [ + 511, + 707, + 905, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {f e a}} = \\sum_ {m = 1} ^ {M} \\left(1 - \\frac {\\mathbf {F} _ {\\mathrm {D e n}} ^ {m} \\cdot \\mathbf {F} _ {\\mathrm {S P A}} ^ {m}}{\\| \\mathbf {F} _ {\\mathrm {D e n}} ^ {m} \\| _ {2} \\cdot \\| \\mathbf {F} _ {\\mathrm {S P A}} ^ {m} \\| _ {2}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 794, + 903, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{\\mathrm{Den}}^{m}$ and $\\mathbf{F}_{\\mathrm{SPA}}^{m}$ are the feature maps of the dense and SPA blocks at the same scale $m$ , and $\\| \\cdot \\|_2$ is the $\\ell_2$ -norm.", + "bbox": [ + 511, + 839, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second loss component operates at the context level, which consists of two parts: a gradient part to preserve", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17885", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49b38b64bc7d95d32ba55ba54ff636e88c9b316038392efcf7859efefae20075.jpg", + "image_caption": [ + "Figure 5. Quantitative comparison of fusion performance with other SOTA methods on $\\mathbf{M}^3\\mathrm{FD}$ and FMB benchmarks. The violin plots illustrate the distribution of metrics, in which the black lines and white triangles indicate the medium and mean values." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 500, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4e0c0959e472e289dd598aa7c93f897d3a5a6ffb075269bf86204727ad087a10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 903, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "structural consistency and an MSE loss for intensity consistency. They are defined as:", + "bbox": [ + 89, + 300, + 482, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g r a d}} = \\left\\| \\nabla \\mathbf {I} _ {\\mathrm {r e f}} - \\nabla \\mathbf {I} _ {\\mathrm {f u s}} \\right\\| _ {1}, \\quad \\mathcal {L} _ {\\mathrm {M S E}} = \\left\\| \\mathbf {I} _ {\\mathrm {r e f}} - \\mathbf {I} _ {\\mathrm {f u s}} \\right\\| _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 358, + 482, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla$ denotes the Sobel operator and $\\| \\cdot \\| _1$ is the $\\ell_1$ -norm. So the cont loss can be calculated as $\\mathcal{L}_{\\mathrm{context}} = \\mathcal{L}_{\\mathrm{grad}} + \\mathcal{L}_{\\mathrm{MSE}}$ . In practice, we apply these terms not only between $\\mathbf{I}_{\\mathrm{ref}}$ and $\\mathbf{I}_{\\mathrm{fus}}$ , but also between each fused output and the original source images to ensure the reconstruction fidelity. This prevents both networks from drifting away from the source image during the distillation process.", + "bbox": [ + 89, + 387, + 482, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The third loss component, contrastive semantic loss aims to ensure that the feature space of the fused image remains close to that of the reference image, while being distinct from the individual visible and infrared images archiving efficient distillation. We utilize the semantic feature space by encoder of SAM (defined as $S_{\\mathrm{E}}$ ) to construct positive and negative pairs through the element-wise multiplication of the binary masks $(\\mathbf{M}_{\\mathrm{vis}}$ and $\\mathbf{M}_{\\mathrm{ir}})$ from $\\mathbf{P}_{\\mathrm{vis}}$ and $\\mathbf{P}_{\\mathrm{ir}}$ . The contrastive semantic loss on each modality is defined as:", + "bbox": [ + 89, + 494, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c s}} ^ {\\mathrm {i r}} = \\sum_ {l = 1} ^ {L} \\frac {\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {f u s}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {r e f}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) \\right\\| _ {2}}{\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {x}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {i r}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) \\right\\| _ {2}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 656, + 482, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c s}} ^ {\\mathrm {v i s}} = \\sum_ {l = 1} ^ {L} \\frac {\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {f u s}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {r e f}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) \\right\\| _ {2}}{\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {x}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {v i s}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) \\right\\| _ {2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 726, + 482, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l$ indexes the layers of $\\mathcal{S}_{\\mathrm{E}}$ , and $\\mathbf{x} \\in \\{\\mathrm{ref}, \\mathrm{fus}\\}$ . So the total $\\mathcal{L}_{\\mathrm{cs}} = \\mathcal{L}_{\\mathrm{cs}}^{\\mathrm{ir}} + \\mathcal{L}_{\\mathrm{cs}}^{\\mathrm{vis}}$ can also be computed bidirectionally in the bi-level optimization scheme between two networks, ensuring that the feature spaces of both the $\\mathbf{I}_{\\mathrm{ref}}$ and the $\\mathbf{I}_{\\mathrm{f}}$ are easily aligned.", + "bbox": [ + 89, + 777, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total distillation loss for the sub-network, denoted as $\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}}$ in Eq. (1), is the sum of the feature loss, context loss, and contrastive semantic loss: $\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}} = \\mathcal{L}_{\\mathrm{fea}} + \\mathcal{L}_{\\mathrm{cont}} + \\mathcal{L}_{\\mathrm{cs}}$ .", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the main network, an additional segmentation cross-entropy loss $\\mathcal{L}_{\\mathrm{seg}}$ is incorporated. This loss is computed between the segmentation predictions generated by an open-vocabulary model $S_{\\mathrm{O}}$ [45, 46] and the ground truth segmentation labels. Instead of affecting the sub-network, this loss aims to prevent potential optimization conflicts. The segmentation loss $\\mathcal{L}_{\\mathrm{seg}}$ is defined as:", + "bbox": [ + 511, + 300, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e g}} = - \\sum_ {c} \\left[ \\mathbf {Y} _ {\\text {l a b e l}} ^ {c} \\log \\left(\\mathbf {Y} _ {\\text {p r e d}} ^ {c}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 417, + 903, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Y}_{\\mathrm{label}}$ represents the label segmentation map, $\\mathbf{Y}_{\\mathrm{pred}}$ is the predicted segmentation map, and $c$ is the class index. For now, the total distillation loss for the main network can be calculated as: $\\mathcal{L}_{\\mathrm{m}}^{\\mathrm{d}} = \\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}} + \\mathcal{L}_{\\mathrm{seg}}$ .", + "bbox": [ + 511, + 450, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 522, + 648, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings and Details", + "text_level": 1, + "bbox": [ + 511, + 547, + 702, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Five representative datasets, namely TNO [27], RoadScene [32], MFNet [2], FMB [13], and $\\mathbf{M}^3\\mathrm{FD}$ [11], are utilized for both training and evaluation. For segmentation, SegFormer [31] (B3 variant) is adopted as the backbone, and the models are trained for 100 epochs. The training and testing splits are strictly in accordance with the official dataset guidelines. Adam is adopted for training, with initial learning rates of $5\\times 10^{-4}$ and $2\\times 10^{-3}$ for the main and sub-networks, respectively. Cosine annealing decays both to $1\\times 10^{-5}$ . Prior to distillation, the models undergo a pre-training phase, followed by 5 epochs of distillation. The batch size is set to 4, and the images are randomly cropped and resized to $192\\times 256$ during training. The entire framework is implemented in PyTorch and executed on two NVIDIA GeForce RTX 4090 GPUs.", + "bbox": [ + 511, + 568, + 906, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results of Multi-modality Image Fusion", + "text_level": 1, + "bbox": [ + 511, + 801, + 856, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate our fusion quality by visualization and quantitative analysis with 9 other SOTA methods in recent years, including DDFM [40], U2Fusion [32], TarDAL [11], SegMiF [13], FILM [42], SHIP [43], MRFS [36], EMMA [41] and TIMFusion [19].", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17886", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/23001fdd093c50df54392f11416e66fb81ca0262c86cc3469b3297a57552b8b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FrameworkSegformer - B3 [31]X - Decoder [45]
MethodVenueRoadBldg.SignVeg.BusPer.CarmIoURoadBldg.SignVeg.BusPer.CarmIoU
Visible-84.881.064.284.746.861.177.149.179.267.818.273.252.056.982.050.1
Infrared-83.777.655.980.938.157.174.242.776.655.612.938.272.763.278.742.9
DDFMICCV'2386.283.069.584.862.561.080.858.276.165.910.173.372.066.682.549.3
U2FusionTPAMI'2286.782.468.584.650.361.880.457.079.770.622.272.775.765.583.350.4
TarDALCVPR'2285.881.868.484.749.561.379.555.774.163.123.267.574.866.481.148.7
SegMiFICCV'2386.382.268.085.063.760.780.057.379.668.614.073.874.365.883.250.8
FILMICML'2486.981.969.184.046.160.679.455.779.270.317.973.779.064.082.050.6
SHIPCVPR'2484.982.168.383.361.460.980.457.276.466.824.771.276.364.680.950.4
MRFSCVPR'2488.282.871.086.061.563.681.156.573.865.023.772.377.266.282.649.6
EMMACVPR'2485.281.266.883.359.058.979.955.878.668.415.074.377.164.582.250.7
TIMFusionTPAMI'2485.180.066.482.850.859.079.755.562.063.012.867.675.865.879.646.1
SAGEProposed89.083.971.486.366.764.181.761.281.070.417.075.277.764.082.851.1
", + "bbox": [ + 101, + 88, + 898, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative semantic segmentation results of SOTA approaches on the FMB dataset.", + "bbox": [ + 218, + 323, + 774, + 338 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/45efdd97dee50ffbedd6ace630d372671749efd86466b54db8f1f2ed245b9788.jpg", + "image_caption": [ + "Figure 6. Qualitative segmentation results of SOTA approaches on the FMB dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 352, + 901, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. Figure 4 presents a visual comparison of various methods. Overall, the proposed method has two main advantages. First, it effectively preserves the multi-modality information of the original images. In the TNO surveillance scene, the vegetation details from the visible image and the smoke from the chimney in the infrared image are both well retained (the image set in the top-left corner). In RoadScene, our method also achieves optimal leaf recovery. On the other hand, our approach demonstrates strong robustness to interference, as it accurately reconstructs the reflective pedestrian crossing lines on the ground and the distant building outlines in the dense fog at night (the green box in the second row). The integration of SAM enhances our method, giving it the ability to surpass other methods and achieve superior results.", + "bbox": [ + 88, + 569, + 482, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We also compare our numerical results with other competitors in Figure 5, based on 100 image pairs randomly selected from the FMB and $\\mathbf{M}^3\\mathrm{FD}$ datasets. We employ four fusion evaluation metrics [21], including the Entropy (EN), Standard Deviation (SD), Sum of Correlation Differences (SCD), and Multi-Scale Structural", + "bbox": [ + 88, + 809, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarity (MS-SSIM). Additionally, we introduce four no-reference image quality assessments [34] related to image quality: BRISQUE, NIQE, MUSIQ, and PaQ-2-PiQ.", + "bbox": [ + 511, + 569, + 903, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In terms of these metrics, our method demonstrates consistent superiority. High SCD and MS-SSIM values indicate that our method effectively preserves a significant amount of information from the source images, while EN reflects our rich detail and texture. The spatial, multi-scale, and patch evaluations of the NR-IQA confirm that our fusion approach has a better pixel distribution and aligns more closely with the human visual system perception.", + "bbox": [ + 511, + 614, + 905, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results of Multi-modality Segmentation", + "text_level": 1, + "bbox": [ + 511, + 742, + 854, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. We present the segmentation visualization results on the challenging new FMB dataset in Figure 6. In the daytime intersection scene, thanks to the powerful semantic priors provided by SAM, our method is the only one to completely distinguish between the truck and the bus. In the nighttime road scene, we successfully segment the sidewalk, achieving optimal performance. Additionally, we showcase the segmentation results on the MFNet dataset in Figure 7. Leveraging the advanced se", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17887", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de179d39cc08c18f9bc5ca1b2b4350100240603ef2fd05f5696322a31ad607ee.jpg", + "image_caption": [ + "Figure 7. Qualitative segmentation results of SOTA approaches on the MFNet dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 898, + 260 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/839f780a6ec62f6ed4a2ea2c3bc760e1d061e926c7b2a05b8a5dfd5bd993ce55.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FrameworkSegformer - B3 [31]
MethodCarPer.BikeCurveC.S.C.C.BumpmIoU
Visible83.354.759.67.4631.037.430.844.6
Infrared79.063.851.47.1025.234.141.844.4
DDFM86.762.263.430.835.545.741.151.7
U2Fusion87.365.960.833.436.046.140.952.0
TarDAL86.467.162.428.430.545.041.551.0
SegMiF85.966.364.434.034.743.048.052.4
FILM86.666.461.936.332.046.746.752.7
SHIP86.763.959.733.535.244.442.451.5
MRFS86.161.559.730.635.244.938.450.4
EMMA86.165.663.825.429.144.941.351.7
TIMFusion86.458.760.329.833.644.138.349.8
SAGE87.967.564.341.234.146.846.754.0
", + "bbox": [ + 91, + 304, + 488, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "matic information learned by SAM, our method successfully segments distant small pedestrian targets during the daytime and near-invisible lane curve at night.", + "bbox": [ + 89, + 563, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons. Table 1 presents a qualitative comparison of IoU results for two segmentation frameworks on the FMB dataset. The first framework involves re-training images generated by various fusion methods, while the second uses the output from an open-vocabulary segmentation network with prompt words, without retraining. In traditional comparisons, our method achieves a 3.0 mIoU improvement over the second-best approach, demonstrating its competitive edge across categories. Additionally, in the segmentation network that requires no training, our method also performs favorably, benefiting from the adaptability provided by SAM's semantic information. Table 2 further showcases the qualitative results on the MFNet dataset, which has lower resolution and fewer labels. Although our method shows some gaps in background categories, it achieves overall superiority in key perception categories and mean performance. Our method integrates SAM with the fusion network, achieving a breakthrough in performance for efficient segmentation.", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43ba012009b327bad72c6544741a67bb9ea660e959cfbdefbcbe5f8092d6c61b.jpg", + "table_caption": [ + "Table 2. Quantitative semantic segmentation results of SOTA approaches on the MFNet dataset." + ], + "table_footnote": [], + "table_body": "
VariantsmIoU\nMFNet / FMBFusion Metrics\nEN / SD/ SCD / MS-SSIM
I. The Impact of SAM
(a) w/o SAM50.7 / 56.56.713 / 36.93 / 1.612 / 1.127
(b) P from Ylabel53.2 / 56.46.239 / 33.52 / 1.412 / 1.102
(c) Swap SO53.4 / 55.56.312 / 34.79 / 1.316 / 1.097
II. Study on the SPA Module
(a) w/o Z52.3 / 57.26.333 / 32.81 / 1.384 / 1.002
(b) w/o k-v in PR52.1 / 56.16.417 / 37.19 / 1.261 / 1.063
(c) w/o PR53.3 / 57.26.556 / 35.48 / 1.576 / 1.046
III. Discussion on the Distillation Scheme
(a) w/o Lfea49.7 / 51.66.436 / 37.28 / 1.277 / 1.079
(b) w/o Lcont50.9 / 56.26.669 / 37.80 / 1.462 / 1.086
(c) w/o Lcs51.7 / 54.36.524 / 39.11 / 1.356 / 1.038
(d) Offline Dist.50.0 / 50.46.221 / 32.65 / 1.581 / 0.987
SAGE54.0 / 61.26.872 / 40.77 / 1.604 / 1.117
", + "bbox": [ + 514, + 304, + 905, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results on both fusion and segmentation of all variants in three ablation studies.", + "bbox": [ + 511, + 549, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 599, + 678, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Impact of SAM. The Segment Anything Module serves as a critical cornerstone in our method. To explore its impact, we derive three variants by isolating SAM from the core framework. Specifically, in variant (a), we replace the semantic patches with randomly cropped source image patches, thereby removing SAM from the main network $\\mathcal{M}$ . In variant (b), we assist the generation of semantic patches with segmentation labels. Additionally, we replace $S_{\\mathrm{O}}$ with a conventional segmentation network [31]. Variant (a) enhances the source image information, leading to the optimal similarity metrics, SCD and MS-SSIM, as shown in Table 3. Our method fully leverages the semantic priors provided by SAM, resulting in the best foggy building contours, as illustrated in the first row of Figure 8.", + "bbox": [ + 511, + 625, + 906, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Study on the SPA Module. The proposed Semantic Persistent Attention module plays a key role in integrating the semantic priors from SAM, effectively guiding the retention and enhancement of inherent modality features. To in-", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17888", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f58134d7d16e01c1c023ab1e6ab0fda3f671a7a23b64ebd6e5431af73e716367.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDDFMU2FusionTarDALSegMiFFILMSHIPMRFSEMMATIMFusionOurs
Time(ms)280K42.3116.57150.5183.327.9395.8625.7317.3910.47
FLOPs(G)1.34M518.0233.3500.4230.3401.5301.2106.4100.752.06
Params(M)552.70.6590.2970.6210.4910.525133.41.5160.1580.136
", + "bbox": [ + 99, + 89, + 898, + 164 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Efficiency comparison with other state-of-the-art methods on ${\\mathrm{M}}^{3}\\mathrm{{FD}}$ benchmarks.", + "bbox": [ + 228, + 172, + 764, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c11fbabedee1423a3eaafa8b881bb45daf4d0d667e6384d5606cf1fe845b8cd.jpg", + "image_caption": [ + "Figure 8. Visualization comparison of impact of SAM." + ], + "image_footnote": [], + "bbox": [ + 93, + 196, + 485, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed52bd26c737e908079d4aa8c62b6b71e5411f65e70743d2500a0b03cf2ed196.jpg", + "image_caption": [ + "Figure 9. Visualization comparison of ablation on SPA." + ], + "image_footnote": [], + "bbox": [ + 93, + 351, + 485, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vestigate its impact, we set up three variants: (a) $w/o$ latent representation $\\mathbf{Z}$ , (b) $w/o$ key-value pairs in the Persistent Repository, and (c) $w/o$ PR. The channel loss due to these modifications has been corrected. A visual comparison of these ablation variants is shown in Figure 9. It is evident that PR in the SPA module plays an essential role in maintaining critical information. The removal of any component leads to significant information loss. For instance, in (b), while the regions of interest are highlighted, the absence of source image information causes blurriness. Similarly, the absence of PR in the SPA module completely prevents the capture of beneficial semantic information, causing the network to lose focus on key areas and resulting in a low-contrast fusion results in (c).", + "bbox": [ + 88, + 536, + 482, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion on the Distillation Scheme. Furthermore, we discuss the distillation scheme. First, we conduct ablation studies on each component of the triple loss function, leading to variants (a)-(c) as shown in Table 3 III. Additionally, we replace the bi-level optimization distillation method with the offline distillation approach, forming variant (d). A visual comparison of these variants is shown in Figure 10. Notably, the results of the traditional online distillation are significantly inferior to our method, both in terms of visual quality and the gradient map generated by variant (b).", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c87e915aacd7d5bff0bf0a30aa9bb89481f01e842716e81a4f92da736f98afb.jpg", + "image_caption": [ + "Figure 10. Visualization comparison of ablation on Distillation." + ], + "image_footnote": [], + "bbox": [ + 516, + 198, + 908, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This confirms the effectiveness of the proposed distillation method when incorporating semantic information.", + "bbox": [ + 511, + 397, + 906, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Computational Efficiency Analysis", + "text_level": 1, + "bbox": [ + 511, + 440, + 815, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the $\\mathrm{M}^{3}\\mathrm{FD}$ benchmark, we compare our method with 9 other SOTA methods in terms of time, FLOPS, and parameters. As shown in Table 4, our method demonstrates significant advantages across all aspects, particularly in time and FLOPs. Specifically, due to our distillation scheme, our sub-network is able to adjust flexibly while maintaining high efficiency, resulting in reduced FLOPs. Compared to other methods, our approach achieves a processing time of $10.47\\mathrm{ms}$ and FLOPs of $52.06\\mathrm{G}$ , outperforming most of the existing methods, with a parameter count of only $0.136\\mathrm{M}$ , showcasing computational efficiency.", + "bbox": [ + 511, + 462, + 906, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method effectively alleviates the computational burden of SAM during inference, significantly reducing the computational overhead while preserving semantic information. This design allows our network to strike a balance between processing speed and computational resources, leading to superior efficiency in practical applications.", + "bbox": [ + 511, + 630, + 905, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 738, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a fusion method that utilizes semantic priors from the SAM for IVIF. We design the Semantic Persistent Attention module, which integrates semantic information while retaining source details. Additionally, we introduce a bi-level distillation scheme with triplet loss to reduce computational complexity by decoupling fusion from SAM during inference. Extensive experiments show that our method achieves superior fusion and task performance, outperforming across multiple datasets.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17889", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is partially supported by the National Natural Science Foundation of China (Nos. 62450072, U22B2052, 62302078), Central Guidance for Local Science and Technology Development Fund (Youth Science Fund Project, Category A, No. 2025JH6/101100001), the Distinguished Young Scholars Funds of Dalian (No. 2024RJ002), the China Postdoctoral Science Foundation (No. 2023M730741) and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 89, + 107, + 482, + 231 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 233, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1", + "[2] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multispectral scenes. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5108-5115. IEEE, 2017. 5", + "[3] Zhiying Jiang, Zengxi Zhang, Jinyuan Liu, Xin Fan, and Risheng Liu. Multispectral image stitching via global-aware quadrature pyramid regression. IEEE Transactions on Image Processing, pages 4288-4302, 2024. 2", + "[4] Zheyan Jin, Shiqi Chen, Yueting Chen, Zhihai Xu, and Huajun Feng. Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870, 2023. 2", + "[5] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of Computational Methods in Engineering, 28(7):4425-4447, 2021. 1", + "[6] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2", + "[7] Guanlin Li, Bin Zhao, and Xuelong Li. Low-light image enhancement with sam-based structure priors and guidance. IEEE Transactions on Multimedia, pages 10854-10866, 2024. 2", + "[8] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11040-11052, 2023. 1, 2", + "[9] Siwei Li, Mingxuan Liu, Yating Zhang, Shu Chen, Haoxiang Li, Zifei Dou, and Hong Chen. Sam-deblur: Let segment anything boost image deblurring. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2445-2449. IEEE, 2024. 2", + "[10] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In Proceedings of the International Conference on Learning Representations, 2019. 4", + "[11] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual" + ], + "bbox": [ + 93, + 258, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5", + "[12] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, pages 1748-1775, 2023. 1, 2", + "[13] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8115–8124, 2023. 1, 2, 5", + "[14] Jinyuan Liu, Guanyao Wu, Zhu Liu, Long Ma, Risheng Liu, and Xin Fan. Where elegance meets precision: Towards a compact, automatic, and flexible framework for multimodality image fusion and applications. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1110-1118, 2024. 2", + "[15] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, Xin Fan, and Risheng Liu. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(4):2349-2369, 2025. 1", + "[16] Risheng Liu, Shichao Cheng, Yi He, Xin Fan, Zhouchen Lin, and Zhongxuan Luo. On the convergence of learning-based iterative methods for nonconvex inverse problems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12):3027-3039, 2019. 1", + "[17] Risheng Liu, Jinyuan Liu, Zhiying Jiang, Xin Fan, and Zhongxuan Luo. A bilevel integrated model with data-driven layer ensemble for multi-modality image fusion. IEEE Transactions on Image Processing, 30:1261-1274, 2020. 1", + "[18] Risheng Liu, Zhiying Jiang, Shuzhou Yang, and Xin Fan. Twin adversarial contrastive learning for underwater image enhancement and beyond. IEEE Transactions on Image Processing, 31:4922-4936, 2022. 1", + "[19] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and meta-initialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6594-6609, 2024. 2, 5", + "[20] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1240-1248, 2023. 1, 2", + "[21] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1, 6", + "[22] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17890", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2", + "[24] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2", + "[25] Linfeng Tang, Yuxin Deng, Yong Ma, Jun Huang, and Jiayi Ma. Superfusion: A versatile image registration and fusion network with semantic awareness. IEEE/CAA Journal of Automatica Sinica, 9(12):2121-2137, 2022. 2", + "[26] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2", + "[27] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5", + "[28] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image superresolution. arXiv preprint arXiv:2402.17133, 2024. 2", + "[29] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 3508-3515, 2022. 2", + "[30] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, pages 1-36, 2023. 2", + "[31] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 5, 6, 7", + "[32] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5", + "[33] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 1", + "[34] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 6", + "[35] Hao Zhang, Han Xu, Xin Tian, Junjun Jiang, and Jiayi Ma. Image fusion meets deep learning: A survey and perspective. Information Fusion, 76:323-336, 2021. 1" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 2, 5", + "[37] Xingchen Zhang, Ping Ye, and Gang Xiao. Vifb: A visible and infrared image fusion benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 104-105, 2020. 1", + "[38] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 2", + "[39] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 2", + "[40] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 1, 2, 5", + "[41] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25912-25921, 2024. 5", + "[42] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 5", + "[43] Naishan Zheng, Man Zhou, Jie Huang, Junming Hou, Haoying Li, Yuan Xu, and Feng Zhao. Probing synergistic high-order interaction in infrared and visible image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26384-26395, 2024. 5", + "[44] Huabing Zhou, Jilei Hou, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Unified gradient-and intensity-discriminator generative adversarial network for image fusion. Information Fusion, 88:184-201, 2022. 2", + "[45] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 5, 6", + "[46] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36:19769-19782, 2023. 5" + ], + "bbox": [ + 516, + 90, + 906, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17891", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_model.json b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..88df8cf984e7494e4d65a577d2078bc95193bda2 --- /dev/null +++ b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.206, + 0.131, + 0.795, + 0.177 + ], + "angle": 0, + "content": "Every SAM Drop Counts: Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.203, + 0.89, + 0.222 + ], + "angle": 0, + "content": "Guanyao Wu†, Haoyu Liu†, Hongming Fu†, Yichuan Peng†, Jinyuan Liu‡, Xin Fan†, Risheng Liu†*" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.222, + 0.788, + 0.239 + ], + "angle": 0, + "content": "†School of Software Technology, Dalian University of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.24, + 0.802, + 0.258 + ], + "angle": 0, + "content": "\\(^{\\ddagger}\\)School of Mechanical Engineering, Dalian University of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.26, + 0.845, + 0.275 + ], + "angle": 0, + "content": "rollingplainko@gmail.com, atlantis918@hotmail.com, {xin.fan, rsliu}@dlut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.31, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.341, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Multi-modality image fusion, particularly infrared and visible, plays a crucial role in integrating diverse modalities to enhance scene understanding. Although early research prioritized visual quality, preserving fine details and adapting to downstream tasks remains challenging. Recent approaches attempt task-specific design but rarely achieve \"The Best of Both Worlds\" due to inconsistent optimization goals. To address these issues, we propose a novel method that leverages the semantic knowledge from the Segment Anything Model (SAM) to Grow the quality of fusion results and Enable downstream task adaptability, namely SAGE. Specifically, we design a Semantic Persistent Attention (SPA) Module that efficiently maintains source information via the persistent repository while extracting high-level semantic priors from SAM. More importantly, to eliminate the impractical dependence on SAM during inference, we introduce a bi-level optimization-driven distillation mechanism with triplet losses, which allow the student network to effectively extract knowledge. Extensive experiments show that our method achieves a balance between high-quality visual results and downstream task adaptability while maintaining practical deployment efficiency. The code is available at https://github.com/RollingPlain/SAGE_IVIF." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.733, + 0.222, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In light of recent advancements in sensing technologies, multi-modality imaging [15] has gained significant attention, with applications in robotics [35], remote sensing [33], and autonomous driving [1]. Integrating infrared and visible sensors provides substantial advantages for intelligent processing [5], although both exhibit inherent limitations [16, 21]. Infrared images are robust against smoke, obstruction, and low light, while visible images excel in reso" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.309, + 0.906, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Figure 1. Differences between the proposed method and existing mainstream comparative approaches: (a) Traditional and early DL-based methods focus on the fusion visual effect. (b) Task-specific methods (e.g., TarDAL [11] & SegMiF [13]) introduce task loss and features that lead to inconsistent optimization goals, causing a conflict between visual and task accuracy. (c) Our pipeline first leverages semantic priors from SAM within a large network and then distills the knowledge into a smaller sub-network achieving practical inference feasibility while ensuring \"the best of both worlds\" through SAM's inherent adaptability to these tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.718 + ], + "angle": 0, + "content": "lution, contrast, and texture detail. These characteristics are highly complementary, motivating their fusion to generate a comprehensive fused image that balances visual quality and downstream task requirements. Consequently, achieving this fusion efficiently remains a pressing challenge [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Traditional Infrared and Visible Image Fusion (IVIF) methods, based on information theory [17, 37], aim to retain as much source image information as possible but often struggle with optimizing fused image quality, especially in managing redundancy and specific scenarios. Early deep learning-based methods [8, 12, 18, 22, 32, 39, 40] focused on visualization but faced issues like edge blurring and artifacts, failing to meet downstream perception needs. Current methods [11, 13, 20, 26] focus on specific tasks, coupling fusion with downstream processing, resulting in conflicting optimization objectives and challenges in balancing visual quality with task adaptability." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.299, + 0.483, + 0.328 + ], + "angle": 0, + "content": "Figure 2. Demonstration of SAM's robustness in MFNet normal scenes (top) and under challenging FMB conditions (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Recent advancements in large-scale visual models have significantly improved various visual analysis tasks [30]. Among these, the Segment Anything Model [6] stands out due to its exceptional and robust capability to provide rich semantic information, making it well-suited for IVIF, as illustrated in Figure 2. Also, its segmentation capabilities align naturally with the requirements of downstream tasks in IVIF field. However, current methods that integrate SAM for low-level tasks typically require full SAM during inference, resulting in excessive practical infeasibility." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.793 + ], + "angle": 0, + "content": "To address these issues, we propose the fusion method SAGE that fully integrates and distills semantic priors derived from SAM. As illustrated in Figure 1 III, one of our aim is to harness the advantages of SAM's semantic priors while reducing computational burden. The proposed method consists of two key components: Semantic Persistent Attention module and a bi-level distillation scheme. The former focuses on using a persistent repository throughout the process to retain source image information, while heuristically enabling the network to integrate semantic patches predicted by SAM. Furthermore, the distillation scheme is designed to efficiently exclude SAM from the inference process, thereby reducing computational complexity. By enhancing distilled information across different aspects, the triplet losses in the scheme significantly boost the performance of the fusion model. Consequently, efficient fusion is achieved using only the distilled model, delivering high-quality visual results and precise task performance without the direct involvement of SAM. In summary, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.841 + ], + "angle": 0, + "content": "- A novel fusion framework is proposed to leverage semantic priors from the SAM, effectively balancing visual quality and downstream task adaptability.." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- The designed Semantic Persistent Attention (SPA) module utilizes a persistent repository to efficiently retain source information and simultaneously extract high-level semantic representation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "- We develop a bi-level optimization distillation scheme that transfers the information processed by SPA into the sub-network, effectively decoupling the fusion process from SAM during inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.152, + 0.905, + 0.183 + ], + "angle": 0, + "content": "- Extensive experiments on multiple datasets validate the superior effectiveness of our proposed method." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.196, + 0.664, + 0.212 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Learning based IVIF Methods. In recent years, deep learning based multi-modality image fusion approaches [8, 12, 32, 39, 40, 44] achieved significant progress, primarily focusing on improving fusion quality, such as enhancing image details, maintaining structural consistency, and reducing noise. These approaches, including innovations in adversarial learning, Transformer models and diffusion network, have provided new insights. However, most of these methods have not addressed how the fused images can directly contribute to downstream tasks such as object detection, semantic segmentation, and other perception-related applications. With the growing demand for more integrated solutions, this gap has sparked increasing research interest in combining fusion with various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.909, + 0.629 + ], + "angle": 0, + "content": "Early works began to bridge this gap, with TarDAL [11] pioneering the use of a dual-discriminator network for object detection optimization, which simultaneously considers modality differences and detection tasks during fusion. As the field has progressed, several methods [24, 38] have also focused on optimizing fusion for object detection. For semantic segmentation [36], SeAFusion [26] and SegMiF [13] have made early significant improvements, with the former enhancing fusion via task loss and the latter leveraging task-specific features. More recently, some approaches [3, 14, 19, 20, 25, 29] have integrated multi-task learning to jointly optimize detection and segmentation, improving performance on both tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "SAM & Its Application. The Segment Anything Model (SAM) [6] is a large-scale pre-trained foundation model that excels in zero-shot generalization, enabling it to handle a wide range of segmentation and detection tasks without the need for task-specific training. SAM demonstrates impressive performance in low-level vision applications, such as image deblurring [9], dehazing [4], super-resolution [28], and low-light image enhancement [7]. Moreover, SAM's capabilities extend to high-level tasks like object detection [23], which further strengthens its versatility. This ability to handle both low-level and high-level tasks makes SAM inherently well-suited for the needs of the IVIF field, where fusion based segmentation and detection tasks are critical. However, its impressive performance is hindered by high computational complexity, posing challenges for practical deployment. Addressing these computational burdens, while fully leveraging the semantic priors of SAM, is a primary motivation for our work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17883" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.625, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.089, + 0.902, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 3. A overall workflow of our proposed method. (a) shows the flow structure of the main network, where the SPA module processes patches with semantic priors generated by SAM. (b) illustrates the detailed structure of the SPA module, where PR plays a key role in preserving the source and integrating the semantic information. (c) displays our distillation scheme formulation, with visualizations of the different components of the triplet loss. (d) provides a simple diagram of the sub-network, which is composed of stacked dense blocks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.413, + 0.303, + 0.43 + ], + "angle": 0, + "content": "3. The Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.295, + 0.453 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.734 + ], + "angle": 0, + "content": "The overall workflow of our method is illustrated in Figure 3. Our aim is to fully exploit the semantic priors of SAM during the inference stage to enhance cross-modal fusion quality. However, directly deploying a large-scale SAM model often leads to prohibitive computational overhead. To mitigate this, we adopt a knowledge distillation strategy, transferring the information encoded by the SAM-driven main network to a lightweight sub-network, thereby significantly reducing inference costs while maintaining high-quality fusion. Nevertheless, the substantial capacity gap between the SAM-enhanced main network and the compact sub-network frequently results in incomplete semantic transfer or structural inconsistencies, which hinder ideal cross-modal fusion performance. To address this issue, we propose a bi-level optimization framework that jointly optimizes both networks as a unified system, aiming to bridge the distillation gap and maintain consistent fusion guided by SAM's semantic priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.826 + ], + "angle": 0, + "content": "Let \\(\\mathbf{I}_{\\mathrm{s}}\\) denote the source visible and infrared input images, i.e., \\(\\mathbf{I}_{\\mathrm{s}} = \\{\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}}\\}\\). Their corresponding semantic-prior patches generated under the guidance of SAM are denoted as \\(\\mathbf{P}_{\\mathrm{s}} = \\{\\mathbf{P}_{\\mathrm{vis}},\\mathbf{P}_{\\mathrm{ir}}\\}\\). Denote the main network by \\(\\mathcal{M}\\) with parameters \\(\\omega_{\\mathbf{m}}\\), and the sub-network by \\(\\mathcal{S}\\) with parameters \\(\\omega_{\\mathbf{s}}\\). We define the optimization process as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.832, + 0.482, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\omega} _ {s}} \\mathcal {L} _ {\\mathrm {s}} ^ {\\mathrm {d}} \\left(\\mathcal {S} \\left(\\mathbf {I} _ {s}, \\boldsymbol {\\omega} _ {s}\\right); \\boldsymbol {\\omega} _ {m} ^ {*}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.86, + 0.439, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\mathrm {s . t .} \\quad \\boldsymbol {\\omega} _ {m} ^ {*} \\in \\arg \\min _ {\\boldsymbol {\\omega} _ {m}} \\mathcal {L} _ {\\mathrm {m}} ^ {\\mathrm {d}} \\big (\\mathcal {M} (\\mathbf {I} _ {s}, \\mathbf {P} _ {s}, \\boldsymbol {\\omega} _ {m}); \\boldsymbol {\\omega} _ {s} \\big),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.566 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{m}}^{\\mathrm{d}}\\) is the distillation loss that ensures high-quality fusion with meaningful semantic cues in the main network, and \\(\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}}\\) is the distillation loss guiding the subnetwork to effectively mimic the main network's behavior. The fused reference image is given by: \\(\\mathbf{I}_{\\mathrm{ref}} = \\mathcal{M}\\big(\\{\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}}\\} ,\\{\\mathbf{P}_{\\mathrm{vis}},\\mathbf{P}_{\\mathrm{ir}}\\} ,\\boldsymbol {\\omega}_m\\big)\\), while the sub-network produces a fused image \\(\\mathbf{I}_{\\mathrm{f}} = S\\big(\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}},\\boldsymbol {\\omega}_s\\big)\\). Notably, both distillation objectives rely on outputs from the counterpart network, establishing a bidirectional dependency, which is a key feature of our distillation formulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.703 + ], + "angle": 0, + "content": "Within the overall framework, the main network integrates SAM-derived semantic priors into the fusion process, while the sub-network is optimized to align with its outputs under the bi-level formulation. The following sections elaborate on two key components that support this process: the Semantic Persistent Attention (SPA) module 3.2, which integrates SAM-derived priors with source features, and the triplet-based loss scheme 3.3, which facilitates structured knowledge transfer across networks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.842, + 0.727 + ], + "angle": 0, + "content": "3.2. Semantic Persistent Attention Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.78 + ], + "angle": 0, + "content": "To fully exploit the semantic information provided by SAM, we propose the Semantic Persistent Attention (SPA) module, as illustrated in the flowchart in Figure 3 (b)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The key core of the SPA module is a Persistent Repository (PR), which serves as a static memory to store and maintain crucial contextual information for the fusion process. Specifically, PR stores the latent representation \\((\\mathbf{Z})\\) of the source features \\(\\mathbf{F}_{\\mathrm{src}}\\) and the corresponding key-value pairs \\((\\mathbf{K}_{\\mathrm{src}},\\mathbf{V}_{\\mathrm{src}})\\), which provide consistent contextual support during the cross-attention operation. The core insight behind this design is that PR acts as a stable, modality-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17884" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.9, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.395, + 0.904, + 0.41 + ], + "angle": 0, + "content": "Figure 4. Qualitative demonstrations of SOTA approaches across commonly used datasets, including TNO, RoadScene, \\(\\mathbf{M}^3\\mathrm{FD}\\) and FMB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.483, + 0.481 + ], + "angle": 0, + "content": "specific information source, guiding the cross-attention mechanism to fuse the semantic patches with rich contextual details from the original source." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.697 + ], + "angle": 0, + "content": "The visible and infrared semantic patches \\((\\mathbf{P}_{\\mathrm{vis}}\\) and \\(\\mathbf{P}_{\\mathrm{ir}}\\) extracted from the SAM are encoded into features \\((\\mathbf{F}_{\\mathrm{pvis}}\\) and \\(\\mathbf{F}_{\\mathrm{pir}})\\). These encoded semantic features represent limited portions of the scene and are processed through the cross-attention mechanism. By utilizing the key-value pairs \\((\\mathbf{K}_{\\mathrm{src}},\\mathbf{V}_{\\mathrm{src}})\\) stored in PR, the cross-attention mechanism enriches the semantic queries \\((\\mathbf{Q}_{\\mathrm{pvis}}\\) and \\(\\mathbf{Q}_{\\mathrm{pir}})\\) with the full scene context from the source image, thereby addressing the inherent incomplete scene coverage in the patches. PR ensures the fusion process stays grounded in a stable, consistent context, enabling patches to be effectively enriched with modality-specific information. This stability maintains semantic coherence while allowing flexible feature refinement via attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.852 + ], + "angle": 0, + "content": "In summary, the SPA module uses PR to guide the cross-attention mechanism, ensuring that the semantic patches from SAM are enriched with consistent and detailed modality-specific information. This approach allows the fusion of infrared and visible image modalities by combining both the high-level semantic understanding from SAM and the detailed information from the source features. The final output, \\(\\mathbf{F}_{\\mathrm{SPA}}\\), represents a semantically enriched, structurally consistent feature set, rich in semantic priors of SAM, and ready for further processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The main network heavily utilizes the SPA module, aiming to fully exploit the complex semantic priors provided by SAM, focus on capturing and retaining detailed semantic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.905, + 0.465 + ], + "angle": 0, + "content": "knowledge, thereby facilitating a powerful representation that can later be distilled into a more efficient sub-network." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.475, + 0.816, + 0.491 + ], + "angle": 0, + "content": "3.3. Triplet Loss in Distillation Scheme" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.707 + ], + "angle": 0, + "content": "In this section, we introduce a triplet loss-driven distillation scheme that facilitates semantic knowledge transfer from the main network \\((\\mathcal{M})\\) to the sub-network \\((S)\\) under a bi-level optimization framework. Specifically, we adopt a DARTS-style [10] training protocol, where the main network (teacher) and the sub-network (student) are updated in small alternating steps, approximating the inner-outer structure of bilevel optimization. This allows gradients to flow in a bidirectional manner: the sub-network learns from the distillation signals provided by the main network, while the main network also adapts its parameters based on the subnetwork's performance and a segmentation objective. Consequently, the two networks reach a cohesive transfer with mutual compromise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.784 + ], + "angle": 0, + "content": "Our distillation scheme is driven by three distinct losses, each focusing on a different aspect of the fused images. Let \\(\\mathbf{I}_{\\mathrm{ref}}\\) be the fused reference image produced by the main network, and \\(\\mathbf{I}_{\\mathrm{f}}\\) be the fused output from the sub-network. We first introduce a feature alignment term:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.795, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {f e a}} = \\sum_ {m = 1} ^ {M} \\left(1 - \\frac {\\mathbf {F} _ {\\mathrm {D e n}} ^ {m} \\cdot \\mathbf {F} _ {\\mathrm {S P A}} ^ {m}}{\\| \\mathbf {F} _ {\\mathrm {D e n}} ^ {m} \\| _ {2} \\cdot \\| \\mathbf {F} _ {\\mathrm {S P A}} ^ {m} \\| _ {2}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.871 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{\\mathrm{Den}}^{m}\\) and \\(\\mathbf{F}_{\\mathrm{SPA}}^{m}\\) are the feature maps of the dense and SPA blocks at the same scale \\(m\\), and \\(\\| \\cdot \\|_2\\) is the \\(\\ell_2\\)-norm." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The second loss component operates at the context level, which consists of two parts: a gradient part to preserve" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17885" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.5, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.904, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.908, + 0.274 + ], + "angle": 0, + "content": "Figure 5. Quantitative comparison of fusion performance with other SOTA methods on \\(\\mathbf{M}^3\\mathrm{FD}\\) and FMB benchmarks. The violin plots illustrate the distribution of metrics, in which the black lines and white triangles indicate the medium and mean values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.483, + 0.332 + ], + "angle": 0, + "content": "structural consistency and an MSE loss for intensity consistency. They are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.359, + 0.483, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g r a d}} = \\left\\| \\nabla \\mathbf {I} _ {\\mathrm {r e f}} - \\nabla \\mathbf {I} _ {\\mathrm {f u s}} \\right\\| _ {1}, \\quad \\mathcal {L} _ {\\mathrm {M S E}} = \\left\\| \\mathbf {I} _ {\\mathrm {r e f}} - \\mathbf {I} _ {\\mathrm {f u s}} \\right\\| _ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.494 + ], + "angle": 0, + "content": "where \\(\\nabla\\) denotes the Sobel operator and \\(\\| \\cdot \\| _1\\) is the \\(\\ell_1\\) -norm. So the cont loss can be calculated as \\(\\mathcal{L}_{\\mathrm{context}} = \\mathcal{L}_{\\mathrm{grad}} + \\mathcal{L}_{\\mathrm{MSE}}\\). In practice, we apply these terms not only between \\(\\mathbf{I}_{\\mathrm{ref}}\\) and \\(\\mathbf{I}_{\\mathrm{fus}}\\), but also between each fused output and the original source images to ensure the reconstruction fidelity. This prevents both networks from drifting away from the source image during the distillation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.631 + ], + "angle": 0, + "content": "The third loss component, contrastive semantic loss aims to ensure that the feature space of the fused image remains close to that of the reference image, while being distinct from the individual visible and infrared images archiving efficient distillation. We utilize the semantic feature space by encoder of SAM (defined as \\( S_{\\mathrm{E}} \\)) to construct positive and negative pairs through the element-wise multiplication of the binary masks \\( (\\mathbf{M}_{\\mathrm{vis}} \\) and \\( \\mathbf{M}_{\\mathrm{ir}}) \\) from \\( \\mathbf{P}_{\\mathrm{vis}} \\) and \\( \\mathbf{P}_{\\mathrm{ir}} \\). The contrastive semantic loss on each modality is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.657, + 0.483, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c s}} ^ {\\mathrm {i r}} = \\sum_ {l = 1} ^ {L} \\frac {\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {f u s}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {r e f}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) \\right\\| _ {2}}{\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {x}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {i r}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) \\right\\| _ {2}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.727, + 0.483, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c s}} ^ {\\mathrm {v i s}} = \\sum_ {l = 1} ^ {L} \\frac {\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {f u s}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {r e f}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) \\right\\| _ {2}}{\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {x}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {v i s}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) \\right\\| _ {2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.854 + ], + "angle": 0, + "content": "where \\(l\\) indexes the layers of \\(\\mathcal{S}_{\\mathrm{E}}\\), and \\(\\mathbf{x} \\in \\{\\mathrm{ref}, \\mathrm{fus}\\}\\). So the total \\(\\mathcal{L}_{\\mathrm{cs}} = \\mathcal{L}_{\\mathrm{cs}}^{\\mathrm{ir}} + \\mathcal{L}_{\\mathrm{cs}}^{\\mathrm{vis}}\\) can also be computed bidirectionally in the bi-level optimization scheme between two networks, ensuring that the feature spaces of both the \\(\\mathbf{I}_{\\mathrm{ref}}\\) and the \\(\\mathbf{I}_{\\mathrm{f}}\\) are easily aligned." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The total distillation loss for the sub-network, denoted as \\(\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}}\\) in Eq. (1), is the sum of the feature loss, context loss, and contrastive semantic loss: \\(\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}} = \\mathcal{L}_{\\mathrm{fea}} + \\mathcal{L}_{\\mathrm{cont}} + \\mathcal{L}_{\\mathrm{cs}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.409 + ], + "angle": 0, + "content": "For the main network, an additional segmentation cross-entropy loss \\(\\mathcal{L}_{\\mathrm{seg}}\\) is incorporated. This loss is computed between the segmentation predictions generated by an open-vocabulary model \\(S_{\\mathrm{O}}\\) [45, 46] and the ground truth segmentation labels. Instead of affecting the sub-network, this loss aims to prevent potential optimization conflicts. The segmentation loss \\(\\mathcal{L}_{\\mathrm{seg}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.418, + 0.905, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e g}} = - \\sum_ {c} \\left[ \\mathbf {Y} _ {\\text {l a b e l}} ^ {c} \\log \\left(\\mathbf {Y} _ {\\text {p r e d}} ^ {c}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.514 + ], + "angle": 0, + "content": "where \\(\\mathbf{Y}_{\\mathrm{label}}\\) represents the label segmentation map, \\(\\mathbf{Y}_{\\mathrm{pred}}\\) is the predicted segmentation map, and \\(c\\) is the class index. For now, the total distillation loss for the main network can be calculated as: \\(\\mathcal{L}_{\\mathrm{m}}^{\\mathrm{d}} = \\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}} + \\mathcal{L}_{\\mathrm{seg}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.523, + 0.649, + 0.54 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.704, + 0.563 + ], + "angle": 0, + "content": "4.1. Settings and Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Five representative datasets, namely TNO [27], RoadScene [32], MFNet [2], FMB [13], and \\(\\mathbf{M}^3\\mathrm{FD}\\) [11], are utilized for both training and evaluation. For segmentation, SegFormer [31] (B3 variant) is adopted as the backbone, and the models are trained for 100 epochs. The training and testing splits are strictly in accordance with the official dataset guidelines. Adam is adopted for training, with initial learning rates of \\(5\\times 10^{-4}\\) and \\(2\\times 10^{-3}\\) for the main and sub-networks, respectively. Cosine annealing decays both to \\(1\\times 10^{-5}\\). Prior to distillation, the models undergo a pre-training phase, followed by 5 epochs of distillation. The batch size is set to 4, and the images are randomly cropped and resized to \\(192\\times 256\\) during training. The entire framework is implemented in PyTorch and executed on two NVIDIA GeForce RTX 4090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.857, + 0.82 + ], + "angle": 0, + "content": "4.2. Results of Multi-modality Image Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We demonstrate our fusion quality by visualization and quantitative analysis with 9 other SOTA methods in recent years, including DDFM [40], U2Fusion [32], TarDAL [11], SegMiF [13], FILM [42], SHIP [43], MRFS [36], EMMA [41] and TIMFusion [19]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17886" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.899, + 0.315 + ], + "angle": 0, + "content": "
FrameworkSegformer - B3 [31]X - Decoder [45]
MethodVenueRoadBldg.SignVeg.BusPer.CarmIoURoadBldg.SignVeg.BusPer.CarmIoU
Visible-84.881.064.284.746.861.177.149.179.267.818.273.252.056.982.050.1
Infrared-83.777.655.980.938.157.174.242.776.655.612.938.272.763.278.742.9
DDFMICCV'2386.283.069.584.862.561.080.858.276.165.910.173.372.066.682.549.3
U2FusionTPAMI'2286.782.468.584.650.361.880.457.079.770.622.272.775.765.583.350.4
TarDALCVPR'2285.881.868.484.749.561.379.555.774.163.123.267.574.866.481.148.7
SegMiFICCV'2386.382.268.085.063.760.780.057.379.668.614.073.874.365.883.250.8
FILMICML'2486.981.969.184.046.160.679.455.779.270.317.973.779.064.082.050.6
SHIPCVPR'2484.982.168.383.361.460.980.457.276.466.824.771.276.364.680.950.4
MRFSCVPR'2488.282.871.086.061.563.681.156.573.865.023.772.377.266.282.649.6
EMMACVPR'2485.281.266.883.359.058.979.955.878.668.415.074.377.164.582.250.7
TIMFusionTPAMI'2485.180.066.482.850.859.079.755.562.063.012.867.675.865.879.646.1
SAGEProposed89.083.971.486.366.764.181.761.281.070.417.075.277.764.082.851.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.324, + 0.776, + 0.339 + ], + "angle": 0, + "content": "Table 1. Quantitative semantic segmentation results of SOTA approaches on the FMB dataset." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.353, + 0.903, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.536, + 0.747, + 0.55 + ], + "angle": 0, + "content": "Figure 6. Qualitative segmentation results of SOTA approaches on the FMB dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.57, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Qualitative Comparisons. Figure 4 presents a visual comparison of various methods. Overall, the proposed method has two main advantages. First, it effectively preserves the multi-modality information of the original images. In the TNO surveillance scene, the vegetation details from the visible image and the smoke from the chimney in the infrared image are both well retained (the image set in the top-left corner). In RoadScene, our method also achieves optimal leaf recovery. On the other hand, our approach demonstrates strong robustness to interference, as it accurately reconstructs the reflective pedestrian crossing lines on the ground and the distant building outlines in the dense fog at night (the green box in the second row). The integration of SAM enhances our method, giving it the ability to surpass other methods and achieve superior results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Quantitative Comparisons. We also compare our numerical results with other competitors in Figure 5, based on 100 image pairs randomly selected from the FMB and \\(\\mathbf{M}^3\\mathrm{FD}\\) datasets. We employ four fusion evaluation metrics [21], including the Entropy (EN), Standard Deviation (SD), Sum of Correlation Differences (SCD), and Multi-Scale Structural" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.905, + 0.615 + ], + "angle": 0, + "content": "Similarity (MS-SSIM). Additionally, we introduce four no-reference image quality assessments [34] related to image quality: BRISQUE, NIQE, MUSIQ, and PaQ-2-PiQ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.736 + ], + "angle": 0, + "content": "In terms of these metrics, our method demonstrates consistent superiority. High SCD and MS-SSIM values indicate that our method effectively preserves a significant amount of information from the source images, while EN reflects our rich detail and texture. The spatial, multi-scale, and patch evaluations of the NR-IQA confirm that our fusion approach has a better pixel distribution and aligns more closely with the human visual system perception." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.856, + 0.759 + ], + "angle": 0, + "content": "4.3. Results of Multi-modality Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparisons. We present the segmentation visualization results on the challenging new FMB dataset in Figure 6. In the daytime intersection scene, thanks to the powerful semantic priors provided by SAM, our method is the only one to completely distinguish between the truck and the bus. In the nighttime road scene, we successfully segment the sidewalk, achieving optimal performance. Additionally, we showcase the segmentation results on the MFNet dataset in Figure 7. Leveraging the advanced se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17887" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.9, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.272, + 0.754, + 0.286 + ], + "angle": 0, + "content": "Figure 7. Qualitative segmentation results of SOTA approaches on the MFNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.305, + 0.489, + 0.51 + ], + "angle": 0, + "content": "
FrameworkSegformer - B3 [31]
MethodCarPer.BikeCurveC.S.C.C.BumpmIoU
Visible83.354.759.67.4631.037.430.844.6
Infrared79.063.851.47.1025.234.141.844.4
DDFM86.762.263.430.835.545.741.151.7
U2Fusion87.365.960.833.436.046.140.952.0
TarDAL86.467.162.428.430.545.041.551.0
SegMiF85.966.364.434.034.743.048.052.4
FILM86.666.461.936.332.046.746.752.7
SHIP86.763.959.733.535.244.442.451.5
MRFS86.161.559.730.635.244.938.450.4
EMMA86.165.663.825.429.144.941.351.7
TIMFusion86.458.760.329.833.644.138.349.8
SAGE87.967.564.341.234.146.846.754.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Table 2. Quantitative semantic segmentation results of SOTA approaches on the MFNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.61 + ], + "angle": 0, + "content": "matic information learned by SAM, our method successfully segments distant small pedestrian targets during the daytime and near-invisible lane curve at night." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparisons. Table 1 presents a qualitative comparison of IoU results for two segmentation frameworks on the FMB dataset. The first framework involves re-training images generated by various fusion methods, while the second uses the output from an open-vocabulary segmentation network with prompt words, without retraining. In traditional comparisons, our method achieves a 3.0 mIoU improvement over the second-best approach, demonstrating its competitive edge across categories. Additionally, in the segmentation network that requires no training, our method also performs favorably, benefiting from the adaptability provided by SAM's semantic information. Table 2 further showcases the qualitative results on the MFNet dataset, which has lower resolution and fewer labels. Although our method shows some gaps in background categories, it achieves overall superiority in key perception categories and mean performance. Our method integrates SAM with the fusion network, achieving a breakthrough in performance for efficient segmentation." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.305, + 0.906, + 0.538 + ], + "angle": 0, + "content": "
VariantsmIoU\nMFNet / FMBFusion Metrics\nEN / SD/ SCD / MS-SSIM
I. The Impact of SAM
(a) w/o SAM50.7 / 56.56.713 / 36.93 / 1.612 / 1.127
(b) P from Ylabel53.2 / 56.46.239 / 33.52 / 1.412 / 1.102
(c) Swap SO53.4 / 55.56.312 / 34.79 / 1.316 / 1.097
II. Study on the SPA Module
(a) w/o Z52.3 / 57.26.333 / 32.81 / 1.384 / 1.002
(b) w/o k-v in PR52.1 / 56.16.417 / 37.19 / 1.261 / 1.063
(c) w/o PR53.3 / 57.26.556 / 35.48 / 1.576 / 1.046
III. Discussion on the Distillation Scheme
(a) w/o Lfea49.7 / 51.66.436 / 37.28 / 1.277 / 1.079
(b) w/o Lcont50.9 / 56.26.669 / 37.80 / 1.462 / 1.086
(c) w/o Lcs51.7 / 54.36.524 / 39.11 / 1.356 / 1.038
(d) Offline Dist.50.0 / 50.46.221 / 32.65 / 1.581 / 0.987
SAGE54.0 / 61.26.872 / 40.77 / 1.604 / 1.117
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.55, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Table 3. Quantitative results on both fusion and segmentation of all variants in three ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.6, + 0.679, + 0.615 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.838 + ], + "angle": 0, + "content": "The Impact of SAM. The Segment Anything Module serves as a critical cornerstone in our method. To explore its impact, we derive three variants by isolating SAM from the core framework. Specifically, in variant (a), we replace the semantic patches with randomly cropped source image patches, thereby removing SAM from the main network \\(\\mathcal{M}\\). In variant (b), we assist the generation of semantic patches with segmentation labels. Additionally, we replace \\(S_{\\mathrm{O}}\\) with a conventional segmentation network [31]. Variant (a) enhances the source image information, leading to the optimal similarity metrics, SCD and MS-SSIM, as shown in Table 3. Our method fully leverages the semantic priors provided by SAM, resulting in the best foggy building contours, as illustrated in the first row of Figure 8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Study on the SPA Module. The proposed Semantic Persistent Attention module plays a key role in integrating the semantic priors from SAM, effectively guiding the retention and enhancement of inherent modality features. To in-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17888" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.09, + 0.899, + 0.165 + ], + "angle": 0, + "content": "
MethodDDFMU2FusionTarDALSegMiFFILMSHIPMRFSEMMATIMFusionOurs
Time(ms)280K42.3116.57150.5183.327.9395.8625.7317.3910.47
FLOPs(G)1.34M518.0233.3500.4230.3401.5301.2106.4100.752.06
Params(M)552.70.6590.2970.6210.4910.525133.41.5160.1580.136
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.174, + 0.766, + 0.188 + ], + "angle": 0, + "content": "Table 4. Efficiency comparison with other state-of-the-art methods on \\( {\\mathrm{M}}^{3}\\mathrm{{FD}} \\) benchmarks." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.198, + 0.486, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.327, + 0.451, + 0.341 + ], + "angle": 0, + "content": "Figure 8. Visualization comparison of impact of SAM." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.352, + 0.486, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.5, + 0.452, + 0.513 + ], + "angle": 0, + "content": "Figure 9. Visualization comparison of ablation on SPA." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.748 + ], + "angle": 0, + "content": "vestigate its impact, we set up three variants: (a) \\( w/o \\) latent representation \\( \\mathbf{Z} \\), (b) \\( w/o \\) key-value pairs in the Persistent Repository, and (c) \\( w/o \\) PR. The channel loss due to these modifications has been corrected. A visual comparison of these ablation variants is shown in Figure 9. It is evident that PR in the SPA module plays an essential role in maintaining critical information. The removal of any component leads to significant information loss. For instance, in (b), while the regions of interest are highlighted, the absence of source image information causes blurriness. Similarly, the absence of PR in the SPA module completely prevents the capture of beneficial semantic information, causing the network to lose focus on key areas and resulting in a low-contrast fusion results in (c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Discussion on the Distillation Scheme. Furthermore, we discuss the distillation scheme. First, we conduct ablation studies on each component of the triple loss function, leading to variants (a)-(c) as shown in Table 3 III. Additionally, we replace the bi-level optimization distillation method with the offline distillation approach, forming variant (d). A visual comparison of these variants is shown in Figure 10. Notably, the results of the traditional online distillation are significantly inferior to our method, both in terms of visual quality and the gradient map generated by variant (b)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.199, + 0.909, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.369, + 0.899, + 0.383 + ], + "angle": 0, + "content": "Figure 10. Visualization comparison of ablation on Distillation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.398, + 0.907, + 0.428 + ], + "angle": 0, + "content": "This confirms the effectiveness of the proposed distillation method when incorporating semantic information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.441, + 0.816, + 0.457 + ], + "angle": 0, + "content": "4.5. Computational Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.631 + ], + "angle": 0, + "content": "In the \\(\\mathrm{M}^{3}\\mathrm{FD}\\) benchmark, we compare our method with 9 other SOTA methods in terms of time, FLOPS, and parameters. As shown in Table 4, our method demonstrates significant advantages across all aspects, particularly in time and FLOPs. Specifically, due to our distillation scheme, our sub-network is able to adjust flexibly while maintaining high efficiency, resulting in reduced FLOPs. Compared to other methods, our approach achieves a processing time of \\(10.47\\mathrm{ms}\\) and FLOPs of \\(52.06\\mathrm{G}\\), outperforming most of the existing methods, with a parameter count of only \\(0.136\\mathrm{M}\\), showcasing computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.906, + 0.723 + ], + "angle": 0, + "content": "Our method effectively alleviates the computational burden of SAM during inference, significantly reducing the computational overhead while preserving semantic information. This design allows our network to strike a balance between processing speed and computational resources, leading to superior efficiency in practical applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.634, + 0.754 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose a fusion method that utilizes semantic priors from the SAM for IVIF. We design the Semantic Persistent Attention module, which integrates semantic information while retaining source details. Additionally, we introduce a bi-level distillation scheme with triplet loss to reduce computational complexity by decoupling fusion from SAM during inference. Extensive experiments show that our method achieves superior fusion and task performance, outperforming across multiple datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17889" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.483, + 0.232 + ], + "angle": 0, + "content": "This work is partially supported by the National Natural Science Foundation of China (Nos. 62450072, U22B2052, 62302078), Central Guidance for Local Science and Technology Development Fund (Youth Science Fund Project, Category A, No. 2025JH6/101100001), the Distinguished Young Scholars Funds of Dalian (No. 2024RJ002), the China Postdoctoral Science Foundation (No. 2023M730741) and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.234, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[1] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.342, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[2] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multispectral scenes. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5108-5115. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.426, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[3] Zhiying Jiang, Zengxi Zhang, Jinyuan Liu, Xin Fan, and Risheng Liu. Multispectral image stitching via global-aware quadrature pyramid regression. IEEE Transactions on Image Processing, pages 4288-4302, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.483, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[4] Zheyan Jin, Shiqi Chen, Yueting Chen, Zhihai Xu, and Huajun Feng. Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.565 + ], + "angle": 0, + "content": "[5] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of Computational Methods in Engineering, 28(7):4425-4447, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[6] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[7] Guanlin Li, Bin Zhao, and Xuelong Li. Low-light image enhancement with sam-based structure priors and guidance. IEEE Transactions on Multimedia, pages 10854-10866, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[8] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11040-11052, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[9] Siwei Li, Mingxuan Liu, Yating Zhang, Shu Chen, Haoxiang Li, Zifei Dou, and Hong Chen. Sam-deblur: Let segment anything boost image deblurring. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2445-2449. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[10] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In Proceedings of the International Conference on Learning Representations, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.259, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[12] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, pages 1748-1775, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[13] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8115–8124, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.316, + 0.907, + 0.398 + ], + "angle": 0, + "content": "[14] Jinyuan Liu, Guanyao Wu, Zhu Liu, Long Ma, Risheng Liu, and Xin Fan. Where elegance meets precision: Towards a compact, automatic, and flexible framework for multimodality image fusion and applications. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1110-1118, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[15] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, Xin Fan, and Risheng Liu. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(4):2349-2369, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[16] Risheng Liu, Shichao Cheng, Yi He, Xin Fan, Zhouchen Lin, and Zhongxuan Luo. On the convergence of learning-based iterative methods for nonconvex inverse problems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12):3027-3039, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[17] Risheng Liu, Jinyuan Liu, Zhiying Jiang, Xin Fan, and Zhongxuan Luo. A bilevel integrated model with data-driven layer ensemble for multi-modality image fusion. IEEE Transactions on Image Processing, 30:1261-1274, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[18] Risheng Liu, Zhiying Jiang, Shuzhou Yang, and Xin Fan. Twin adversarial contrastive learning for underwater image enhancement and beyond. IEEE Transactions on Image Processing, 31:4922-4936, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[19] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and meta-initialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6594-6609, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[20] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1240-1248, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[21] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[22] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17890" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[24] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.482, + 0.278 + ], + "angle": 0, + "content": "[25] Linfeng Tang, Yuxin Deng, Yong Ma, Jun Huang, and Jiayi Ma. Superfusion: A versatile image registration and fusion network with semantic awareness. IEEE/CAA Journal of Automatica Sinica, 9(12):2121-2137, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[26] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.367 + ], + "angle": 0, + "content": "[27] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.369, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[28] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image superresolution. arXiv preprint arXiv:2402.17133, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[29] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 3508-3515, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.5, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[30] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, pages 1-36, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[31] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[32] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[33] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[34] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Hao Zhang, Han Xu, Xin Tian, Junjun Jiang, and Jiayi Ma. Image fusion meets deep learning: A survey and perspective. Information Fusion, 76:323-336, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[36] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[37] Xingchen Zhang, Ping Ye, and Gang Xiao. Vifb: A visible and infrared image fusion benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 104-105, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[38] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.373 + ], + "angle": 0, + "content": "[39] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.906, + 0.457 + ], + "angle": 0, + "content": "[40] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.906, + 0.528 + ], + "angle": 0, + "content": "[41] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25912-25921, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.906, + 0.611 + ], + "angle": 0, + "content": "[42] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.613, + 0.906, + 0.683 + ], + "angle": 0, + "content": "[43] Naishan Zheng, Man Zhou, Jie Huang, Junming Hou, Haoying Li, Yuan Xu, and Feng Zhao. Probing synergistic high-order interaction in infrared and visible image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26384-26395, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.684, + 0.906, + 0.739 + ], + "angle": 0, + "content": "[44] Huabing Zhou, Jilei Hou, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Unified gradient-and intensity-discriminator generative adversarial network for image fusion. Information Fusion, 88:184-201, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.741, + 0.906, + 0.822 + ], + "angle": 0, + "content": "[45] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.825, + 0.906, + 0.893 + ], + "angle": 0, + "content": "[46] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36:19769-19782, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17891" + } + ] +] \ No newline at end of file diff --git a/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_origin.pdf b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab8f91133123ee7320d9c8d3cea1983a7d4c2773 --- /dev/null +++ b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/a7813684-e289-4dec-92c0-ee63fc512deb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8f899d909f052ef213da282036381fe1f2dbfdbe481311c955f352122e3aed0 +size 7754214 diff --git a/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/full.md b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95bcb579018baad24b4d6a7b6a36d1c0c39c1e7c --- /dev/null +++ b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/full.md @@ -0,0 +1,276 @@ +# Every SAM Drop Counts: Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond + +Guanyao Wu†, Haoyu Liu†, Hongming Fu†, Yichuan Peng†, Jinyuan Liu‡, Xin Fan†, Risheng Liu†* + +†School of Software Technology, Dalian University of Technology, China + +$^{\ddagger}$ School of Mechanical Engineering, Dalian University of Technology, China + +rollingplainko@gmail.com, atlantis918@hotmail.com, {xin.fan, rsliu}@dlut.edu.cn + +# Abstract + +Multi-modality image fusion, particularly infrared and visible, plays a crucial role in integrating diverse modalities to enhance scene understanding. Although early research prioritized visual quality, preserving fine details and adapting to downstream tasks remains challenging. Recent approaches attempt task-specific design but rarely achieve "The Best of Both Worlds" due to inconsistent optimization goals. To address these issues, we propose a novel method that leverages the semantic knowledge from the Segment Anything Model (SAM) to Grow the quality of fusion results and Enable downstream task adaptability, namely SAGE. Specifically, we design a Semantic Persistent Attention (SPA) Module that efficiently maintains source information via the persistent repository while extracting high-level semantic priors from SAM. More importantly, to eliminate the impractical dependence on SAM during inference, we introduce a bi-level optimization-driven distillation mechanism with triplet losses, which allow the student network to effectively extract knowledge. Extensive experiments show that our method achieves a balance between high-quality visual results and downstream task adaptability while maintaining practical deployment efficiency. The code is available at https://github.com/RollingPlain/SAGE_IVIF. + +# 1. Introduction + +In light of recent advancements in sensing technologies, multi-modality imaging [15] has gained significant attention, with applications in robotics [35], remote sensing [33], and autonomous driving [1]. Integrating infrared and visible sensors provides substantial advantages for intelligent processing [5], although both exhibit inherent limitations [16, 21]. Infrared images are robust against smoke, obstruction, and low light, while visible images excel in reso + +![](images/032986408fbfc16d793221c9253fc95c0278746d72dd3613c50c98de82ac55c4.jpg) +Figure 1. Differences between the proposed method and existing mainstream comparative approaches: (a) Traditional and early DL-based methods focus on the fusion visual effect. (b) Task-specific methods (e.g., TarDAL [11] & SegMiF [13]) introduce task loss and features that lead to inconsistent optimization goals, causing a conflict between visual and task accuracy. (c) Our pipeline first leverages semantic priors from SAM within a large network and then distills the knowledge into a smaller sub-network achieving practical inference feasibility while ensuring "the best of both worlds" through SAM's inherent adaptability to these tasks. + +lution, contrast, and texture detail. These characteristics are highly complementary, motivating their fusion to generate a comprehensive fused image that balances visual quality and downstream task requirements. Consequently, achieving this fusion efficiently remains a pressing challenge [15]. + +Traditional Infrared and Visible Image Fusion (IVIF) methods, based on information theory [17, 37], aim to retain as much source image information as possible but often struggle with optimizing fused image quality, especially in managing redundancy and specific scenarios. Early deep learning-based methods [8, 12, 18, 22, 32, 39, 40] focused on visualization but faced issues like edge blurring and artifacts, failing to meet downstream perception needs. Current methods [11, 13, 20, 26] focus on specific tasks, coupling fusion with downstream processing, resulting in conflicting optimization objectives and challenges in balancing visual quality with task adaptability. + +![](images/74df534b7eea6c9ad180d230068b72d93948a1cd34ac0e8750cda437264a6211.jpg) +Figure 2. Demonstration of SAM's robustness in MFNet normal scenes (top) and under challenging FMB conditions (bottom). + +Recent advancements in large-scale visual models have significantly improved various visual analysis tasks [30]. Among these, the Segment Anything Model [6] stands out due to its exceptional and robust capability to provide rich semantic information, making it well-suited for IVIF, as illustrated in Figure 2. Also, its segmentation capabilities align naturally with the requirements of downstream tasks in IVIF field. However, current methods that integrate SAM for low-level tasks typically require full SAM during inference, resulting in excessive practical infeasibility. + +To address these issues, we propose the fusion method SAGE that fully integrates and distills semantic priors derived from SAM. As illustrated in Figure 1 III, one of our aim is to harness the advantages of SAM's semantic priors while reducing computational burden. The proposed method consists of two key components: Semantic Persistent Attention module and a bi-level distillation scheme. The former focuses on using a persistent repository throughout the process to retain source image information, while heuristically enabling the network to integrate semantic patches predicted by SAM. Furthermore, the distillation scheme is designed to efficiently exclude SAM from the inference process, thereby reducing computational complexity. By enhancing distilled information across different aspects, the triplet losses in the scheme significantly boost the performance of the fusion model. Consequently, efficient fusion is achieved using only the distilled model, delivering high-quality visual results and precise task performance without the direct involvement of SAM. In summary, our contributions can be summarized as follows: + +- A novel fusion framework is proposed to leverage semantic priors from the SAM, effectively balancing visual quality and downstream task adaptability.. +- The designed Semantic Persistent Attention (SPA) module utilizes a persistent repository to efficiently retain source information and simultaneously extract high-level semantic representation. + +- We develop a bi-level optimization distillation scheme that transfers the information processed by SPA into the sub-network, effectively decoupling the fusion process from SAM during inference. +- Extensive experiments on multiple datasets validate the superior effectiveness of our proposed method. + +# 2. Related Works + +Learning based IVIF Methods. In recent years, deep learning based multi-modality image fusion approaches [8, 12, 32, 39, 40, 44] achieved significant progress, primarily focusing on improving fusion quality, such as enhancing image details, maintaining structural consistency, and reducing noise. These approaches, including innovations in adversarial learning, Transformer models and diffusion network, have provided new insights. However, most of these methods have not addressed how the fused images can directly contribute to downstream tasks such as object detection, semantic segmentation, and other perception-related applications. With the growing demand for more integrated solutions, this gap has sparked increasing research interest in combining fusion with various downstream tasks. + +Early works began to bridge this gap, with TarDAL [11] pioneering the use of a dual-discriminator network for object detection optimization, which simultaneously considers modality differences and detection tasks during fusion. As the field has progressed, several methods [24, 38] have also focused on optimizing fusion for object detection. For semantic segmentation [36], SeAFusion [26] and SegMiF [13] have made early significant improvements, with the former enhancing fusion via task loss and the latter leveraging task-specific features. More recently, some approaches [3, 14, 19, 20, 25, 29] have integrated multi-task learning to jointly optimize detection and segmentation, improving performance on both tasks. + +SAM & Its Application. The Segment Anything Model (SAM) [6] is a large-scale pre-trained foundation model that excels in zero-shot generalization, enabling it to handle a wide range of segmentation and detection tasks without the need for task-specific training. SAM demonstrates impressive performance in low-level vision applications, such as image deblurring [9], dehazing [4], super-resolution [28], and low-light image enhancement [7]. Moreover, SAM's capabilities extend to high-level tasks like object detection [23], which further strengthens its versatility. This ability to handle both low-level and high-level tasks makes SAM inherently well-suited for the needs of the IVIF field, where fusion based segmentation and detection tasks are critical. However, its impressive performance is hindered by high computational complexity, posing challenges for practical deployment. Addressing these computational burdens, while fully leveraging the semantic priors of SAM, is a primary motivation for our work. + +![](images/d350ee2af8cb773fdc2ef3c1c3ba9e565d318ee92292168bdf0165f0bb39a7b5.jpg) +Figure 3. A overall workflow of our proposed method. (a) shows the flow structure of the main network, where the SPA module processes patches with semantic priors generated by SAM. (b) illustrates the detailed structure of the SPA module, where PR plays a key role in preserving the source and integrating the semantic information. (c) displays our distillation scheme formulation, with visualizations of the different components of the triplet loss. (d) provides a simple diagram of the sub-network, which is composed of stacked dense blocks. + +![](images/322fa75d65a95cc9cc44b6529ffb30411e2ee5c2916d7b3334f1dc08309dc097.jpg) + +# 3. The Proposed Method + +# 3.1. Problem Formulation + +The overall workflow of our method is illustrated in Figure 3. Our aim is to fully exploit the semantic priors of SAM during the inference stage to enhance cross-modal fusion quality. However, directly deploying a large-scale SAM model often leads to prohibitive computational overhead. To mitigate this, we adopt a knowledge distillation strategy, transferring the information encoded by the SAM-driven main network to a lightweight sub-network, thereby significantly reducing inference costs while maintaining high-quality fusion. Nevertheless, the substantial capacity gap between the SAM-enhanced main network and the compact sub-network frequently results in incomplete semantic transfer or structural inconsistencies, which hinder ideal cross-modal fusion performance. To address this issue, we propose a bi-level optimization framework that jointly optimizes both networks as a unified system, aiming to bridge the distillation gap and maintain consistent fusion guided by SAM's semantic priors. + +Let $\mathbf{I}_{\mathrm{s}}$ denote the source visible and infrared input images, i.e., $\mathbf{I}_{\mathrm{s}} = \{\mathbf{I}_{\mathrm{vis}},\mathbf{I}_{\mathrm{ir}}\}$ . Their corresponding semantic-prior patches generated under the guidance of SAM are denoted as $\mathbf{P}_{\mathrm{s}} = \{\mathbf{P}_{\mathrm{vis}},\mathbf{P}_{\mathrm{ir}}\}$ . Denote the main network by $\mathcal{M}$ with parameters $\omega_{\mathbf{m}}$ , and the sub-network by $\mathcal{S}$ with parameters $\omega_{\mathbf{s}}$ . We define the optimization process as: + +$$ +\min _ {\boldsymbol {\omega} _ {s}} \mathcal {L} _ {\mathrm {s}} ^ {\mathrm {d}} \left(\mathcal {S} \left(\mathbf {I} _ {s}, \boldsymbol {\omega} _ {s}\right); \boldsymbol {\omega} _ {m} ^ {*}\right), \tag {1} +$$ + +$$ +\mathrm {s . t .} \quad \boldsymbol {\omega} _ {m} ^ {*} \in \arg \min _ {\boldsymbol {\omega} _ {m}} \mathcal {L} _ {\mathrm {m}} ^ {\mathrm {d}} \big (\mathcal {M} (\mathbf {I} _ {s}, \mathbf {P} _ {s}, \boldsymbol {\omega} _ {m}); \boldsymbol {\omega} _ {s} \big), +$$ + +where $\mathcal{L}_{\mathrm{m}}^{\mathrm{d}}$ is the distillation loss that ensures high-quality fusion with meaningful semantic cues in the main network, and $\mathcal{L}_{\mathrm{s}}^{\mathrm{d}}$ is the distillation loss guiding the subnetwork to effectively mimic the main network's behavior. The fused reference image is given by: $\mathbf{I}_{\mathrm{ref}} = \mathcal{M}\big(\{\mathbf{I}_{\mathrm{vis}},\mathbf{I}_{\mathrm{ir}}\} ,\{\mathbf{P}_{\mathrm{vis}},\mathbf{P}_{\mathrm{ir}}\} ,\boldsymbol {\omega}_m\big)$ , while the sub-network produces a fused image $\mathbf{I}_{\mathrm{f}} = S\big(\mathbf{I}_{\mathrm{vis}},\mathbf{I}_{\mathrm{ir}},\boldsymbol {\omega}_s\big)$ . Notably, both distillation objectives rely on outputs from the counterpart network, establishing a bidirectional dependency, which is a key feature of our distillation formulation. + +Within the overall framework, the main network integrates SAM-derived semantic priors into the fusion process, while the sub-network is optimized to align with its outputs under the bi-level formulation. The following sections elaborate on two key components that support this process: the Semantic Persistent Attention (SPA) module 3.2, which integrates SAM-derived priors with source features, and the triplet-based loss scheme 3.3, which facilitates structured knowledge transfer across networks. + +# 3.2. Semantic Persistent Attention Module + +To fully exploit the semantic information provided by SAM, we propose the Semantic Persistent Attention (SPA) module, as illustrated in the flowchart in Figure 3 (b). + +The key core of the SPA module is a Persistent Repository (PR), which serves as a static memory to store and maintain crucial contextual information for the fusion process. Specifically, PR stores the latent representation $(\mathbf{Z})$ of the source features $\mathbf{F}_{\mathrm{src}}$ and the corresponding key-value pairs $(\mathbf{K}_{\mathrm{src}},\mathbf{V}_{\mathrm{src}})$ , which provide consistent contextual support during the cross-attention operation. The core insight behind this design is that PR acts as a stable, modality- + +![](images/001ea98476575670761444f88549e1308937d98346e53e799a311ce46dfd5492.jpg) +Figure 4. Qualitative demonstrations of SOTA approaches across commonly used datasets, including TNO, RoadScene, $\mathbf{M}^3\mathrm{FD}$ and FMB. + +specific information source, guiding the cross-attention mechanism to fuse the semantic patches with rich contextual details from the original source. + +The visible and infrared semantic patches $(\mathbf{P}_{\mathrm{vis}}$ and $\mathbf{P}_{\mathrm{ir}}$ extracted from the SAM are encoded into features $(\mathbf{F}_{\mathrm{pvis}}$ and $\mathbf{F}_{\mathrm{pir}})$ . These encoded semantic features represent limited portions of the scene and are processed through the cross-attention mechanism. By utilizing the key-value pairs $(\mathbf{K}_{\mathrm{src}},\mathbf{V}_{\mathrm{src}})$ stored in PR, the cross-attention mechanism enriches the semantic queries $(\mathbf{Q}_{\mathrm{pvis}}$ and $\mathbf{Q}_{\mathrm{pir}})$ with the full scene context from the source image, thereby addressing the inherent incomplete scene coverage in the patches. PR ensures the fusion process stays grounded in a stable, consistent context, enabling patches to be effectively enriched with modality-specific information. This stability maintains semantic coherence while allowing flexible feature refinement via attention mechanisms. + +In summary, the SPA module uses PR to guide the cross-attention mechanism, ensuring that the semantic patches from SAM are enriched with consistent and detailed modality-specific information. This approach allows the fusion of infrared and visible image modalities by combining both the high-level semantic understanding from SAM and the detailed information from the source features. The final output, $\mathbf{F}_{\mathrm{SPA}}$ , represents a semantically enriched, structurally consistent feature set, rich in semantic priors of SAM, and ready for further processing. + +The main network heavily utilizes the SPA module, aiming to fully exploit the complex semantic priors provided by SAM, focus on capturing and retaining detailed semantic + +knowledge, thereby facilitating a powerful representation that can later be distilled into a more efficient sub-network. + +# 3.3. Triplet Loss in Distillation Scheme + +In this section, we introduce a triplet loss-driven distillation scheme that facilitates semantic knowledge transfer from the main network $(\mathcal{M})$ to the sub-network $(S)$ under a bi-level optimization framework. Specifically, we adopt a DARTS-style [10] training protocol, where the main network (teacher) and the sub-network (student) are updated in small alternating steps, approximating the inner-outer structure of bilevel optimization. This allows gradients to flow in a bidirectional manner: the sub-network learns from the distillation signals provided by the main network, while the main network also adapts its parameters based on the subnetwork's performance and a segmentation objective. Consequently, the two networks reach a cohesive transfer with mutual compromise. + +Our distillation scheme is driven by three distinct losses, each focusing on a different aspect of the fused images. Let $\mathbf{I}_{\mathrm{ref}}$ be the fused reference image produced by the main network, and $\mathbf{I}_{\mathrm{f}}$ be the fused output from the sub-network. We first introduce a feature alignment term: + +$$ +\mathcal {L} _ {\mathrm {f e a}} = \sum_ {m = 1} ^ {M} \left(1 - \frac {\mathbf {F} _ {\mathrm {D e n}} ^ {m} \cdot \mathbf {F} _ {\mathrm {S P A}} ^ {m}}{\| \mathbf {F} _ {\mathrm {D e n}} ^ {m} \| _ {2} \cdot \| \mathbf {F} _ {\mathrm {S P A}} ^ {m} \| _ {2}}\right), \tag {2} +$$ + +where $\mathbf{F}_{\mathrm{Den}}^{m}$ and $\mathbf{F}_{\mathrm{SPA}}^{m}$ are the feature maps of the dense and SPA blocks at the same scale $m$ , and $\| \cdot \|_2$ is the $\ell_2$ -norm. + +The second loss component operates at the context level, which consists of two parts: a gradient part to preserve + +![](images/49b38b64bc7d95d32ba55ba54ff636e88c9b316038392efcf7859efefae20075.jpg) +Figure 5. Quantitative comparison of fusion performance with other SOTA methods on $\mathbf{M}^3\mathrm{FD}$ and FMB benchmarks. The violin plots illustrate the distribution of metrics, in which the black lines and white triangles indicate the medium and mean values. + +![](images/4e0c0959e472e289dd598aa7c93f897d3a5a6ffb075269bf86204727ad087a10.jpg) + +structural consistency and an MSE loss for intensity consistency. They are defined as: + +$$ +\mathcal {L} _ {\mathrm {g r a d}} = \left\| \nabla \mathbf {I} _ {\mathrm {r e f}} - \nabla \mathbf {I} _ {\mathrm {f u s}} \right\| _ {1}, \quad \mathcal {L} _ {\mathrm {M S E}} = \left\| \mathbf {I} _ {\mathrm {r e f}} - \mathbf {I} _ {\mathrm {f u s}} \right\| _ {2}, \tag {3} +$$ + +where $\nabla$ denotes the Sobel operator and $\| \cdot \| _1$ is the $\ell_1$ -norm. So the cont loss can be calculated as $\mathcal{L}_{\mathrm{context}} = \mathcal{L}_{\mathrm{grad}} + \mathcal{L}_{\mathrm{MSE}}$ . In practice, we apply these terms not only between $\mathbf{I}_{\mathrm{ref}}$ and $\mathbf{I}_{\mathrm{fus}}$ , but also between each fused output and the original source images to ensure the reconstruction fidelity. This prevents both networks from drifting away from the source image during the distillation process. + +The third loss component, contrastive semantic loss aims to ensure that the feature space of the fused image remains close to that of the reference image, while being distinct from the individual visible and infrared images archiving efficient distillation. We utilize the semantic feature space by encoder of SAM (defined as $S_{\mathrm{E}}$ ) to construct positive and negative pairs through the element-wise multiplication of the binary masks $(\mathbf{M}_{\mathrm{vis}}$ and $\mathbf{M}_{\mathrm{ir}})$ from $\mathbf{P}_{\mathrm{vis}}$ and $\mathbf{P}_{\mathrm{ir}}$ . The contrastive semantic loss on each modality is defined as: + +$$ +\mathcal {L} _ {\mathrm {c s}} ^ {\mathrm {i r}} = \sum_ {l = 1} ^ {L} \frac {\left\| \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\text {f u s}} \odot \mathbf {M} _ {\mathrm {i r}}\right) - \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\text {r e f}} \odot \mathbf {M} _ {\mathrm {i r}}\right) \right\| _ {2}}{\left\| \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\mathrm {x}} \odot \mathbf {M} _ {\mathrm {i r}}\right) - \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\mathrm {i r}} \odot \mathbf {M} _ {\mathrm {i r}}\right) \right\| _ {2}}, \tag {4} +$$ + +$$ +\mathcal {L} _ {\mathrm {c s}} ^ {\mathrm {v i s}} = \sum_ {l = 1} ^ {L} \frac {\left\| \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\text {f u s}} \odot \mathbf {M} _ {\text {v i s}}\right) - \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\text {r e f}} \odot \mathbf {M} _ {\text {v i s}}\right) \right\| _ {2}}{\left\| \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\mathrm {x}} \odot \mathbf {M} _ {\text {v i s}}\right) - \mathcal {S} _ {\mathrm {E}} \left(\mathbf {I} _ {\text {v i s}} \odot \mathbf {M} _ {\text {v i s}}\right) \right\| _ {2}}, \tag {5} +$$ + +where $l$ indexes the layers of $\mathcal{S}_{\mathrm{E}}$ , and $\mathbf{x} \in \{\mathrm{ref}, \mathrm{fus}\}$ . So the total $\mathcal{L}_{\mathrm{cs}} = \mathcal{L}_{\mathrm{cs}}^{\mathrm{ir}} + \mathcal{L}_{\mathrm{cs}}^{\mathrm{vis}}$ can also be computed bidirectionally in the bi-level optimization scheme between two networks, ensuring that the feature spaces of both the $\mathbf{I}_{\mathrm{ref}}$ and the $\mathbf{I}_{\mathrm{f}}$ are easily aligned. + +The total distillation loss for the sub-network, denoted as $\mathcal{L}_{\mathrm{s}}^{\mathrm{d}}$ in Eq. (1), is the sum of the feature loss, context loss, and contrastive semantic loss: $\mathcal{L}_{\mathrm{s}}^{\mathrm{d}} = \mathcal{L}_{\mathrm{fea}} + \mathcal{L}_{\mathrm{cont}} + \mathcal{L}_{\mathrm{cs}}$ . + +For the main network, an additional segmentation cross-entropy loss $\mathcal{L}_{\mathrm{seg}}$ is incorporated. This loss is computed between the segmentation predictions generated by an open-vocabulary model $S_{\mathrm{O}}$ [45, 46] and the ground truth segmentation labels. Instead of affecting the sub-network, this loss aims to prevent potential optimization conflicts. The segmentation loss $\mathcal{L}_{\mathrm{seg}}$ is defined as: + +$$ +\mathcal {L} _ {\text {s e g}} = - \sum_ {c} \left[ \mathbf {Y} _ {\text {l a b e l}} ^ {c} \log \left(\mathbf {Y} _ {\text {p r e d}} ^ {c}\right) \right], \tag {6} +$$ + +where $\mathbf{Y}_{\mathrm{label}}$ represents the label segmentation map, $\mathbf{Y}_{\mathrm{pred}}$ is the predicted segmentation map, and $c$ is the class index. For now, the total distillation loss for the main network can be calculated as: $\mathcal{L}_{\mathrm{m}}^{\mathrm{d}} = \mathcal{L}_{\mathrm{s}}^{\mathrm{d}} + \mathcal{L}_{\mathrm{seg}}$ . + +# 4. Experiments + +# 4.1. Settings and Details + +Five representative datasets, namely TNO [27], RoadScene [32], MFNet [2], FMB [13], and $\mathbf{M}^3\mathrm{FD}$ [11], are utilized for both training and evaluation. For segmentation, SegFormer [31] (B3 variant) is adopted as the backbone, and the models are trained for 100 epochs. The training and testing splits are strictly in accordance with the official dataset guidelines. Adam is adopted for training, with initial learning rates of $5\times 10^{-4}$ and $2\times 10^{-3}$ for the main and sub-networks, respectively. Cosine annealing decays both to $1\times 10^{-5}$ . Prior to distillation, the models undergo a pre-training phase, followed by 5 epochs of distillation. The batch size is set to 4, and the images are randomly cropped and resized to $192\times 256$ during training. The entire framework is implemented in PyTorch and executed on two NVIDIA GeForce RTX 4090 GPUs. + +# 4.2. Results of Multi-modality Image Fusion + +We demonstrate our fusion quality by visualization and quantitative analysis with 9 other SOTA methods in recent years, including DDFM [40], U2Fusion [32], TarDAL [11], SegMiF [13], FILM [42], SHIP [43], MRFS [36], EMMA [41] and TIMFusion [19]. + +
FrameworkSegformer - B3 [31]X - Decoder [45]
MethodVenueRoadBldg.SignVeg.BusPer.CarmIoURoadBldg.SignVeg.BusPer.CarmIoU
Visible-84.881.064.284.746.861.177.149.179.267.818.273.252.056.982.050.1
Infrared-83.777.655.980.938.157.174.242.776.655.612.938.272.763.278.742.9
DDFMICCV'2386.283.069.584.862.561.080.858.276.165.910.173.372.066.682.549.3
U2FusionTPAMI'2286.782.468.584.650.361.880.457.079.770.622.272.775.765.583.350.4
TarDALCVPR'2285.881.868.484.749.561.379.555.774.163.123.267.574.866.481.148.7
SegMiFICCV'2386.382.268.085.063.760.780.057.379.668.614.073.874.365.883.250.8
FILMICML'2486.981.969.184.046.160.679.455.779.270.317.973.779.064.082.050.6
SHIPCVPR'2484.982.168.383.361.460.980.457.276.466.824.771.276.364.680.950.4
MRFSCVPR'2488.282.871.086.061.563.681.156.573.865.023.772.377.266.282.649.6
EMMACVPR'2485.281.266.883.359.058.979.955.878.668.415.074.377.164.582.250.7
TIMFusionTPAMI'2485.180.066.482.850.859.079.755.562.063.012.867.675.865.879.646.1
SAGEProposed89.083.971.486.366.764.181.761.281.070.417.075.277.764.082.851.1
+ +Table 1. Quantitative semantic segmentation results of SOTA approaches on the FMB dataset. + +![](images/45efdd97dee50ffbedd6ace630d372671749efd86466b54db8f1f2ed245b9788.jpg) +Figure 6. Qualitative segmentation results of SOTA approaches on the FMB dataset. + +Qualitative Comparisons. Figure 4 presents a visual comparison of various methods. Overall, the proposed method has two main advantages. First, it effectively preserves the multi-modality information of the original images. In the TNO surveillance scene, the vegetation details from the visible image and the smoke from the chimney in the infrared image are both well retained (the image set in the top-left corner). In RoadScene, our method also achieves optimal leaf recovery. On the other hand, our approach demonstrates strong robustness to interference, as it accurately reconstructs the reflective pedestrian crossing lines on the ground and the distant building outlines in the dense fog at night (the green box in the second row). The integration of SAM enhances our method, giving it the ability to surpass other methods and achieve superior results. + +Quantitative Comparisons. We also compare our numerical results with other competitors in Figure 5, based on 100 image pairs randomly selected from the FMB and $\mathbf{M}^3\mathrm{FD}$ datasets. We employ four fusion evaluation metrics [21], including the Entropy (EN), Standard Deviation (SD), Sum of Correlation Differences (SCD), and Multi-Scale Structural + +Similarity (MS-SSIM). Additionally, we introduce four no-reference image quality assessments [34] related to image quality: BRISQUE, NIQE, MUSIQ, and PaQ-2-PiQ. + +In terms of these metrics, our method demonstrates consistent superiority. High SCD and MS-SSIM values indicate that our method effectively preserves a significant amount of information from the source images, while EN reflects our rich detail and texture. The spatial, multi-scale, and patch evaluations of the NR-IQA confirm that our fusion approach has a better pixel distribution and aligns more closely with the human visual system perception. + +# 4.3. Results of Multi-modality Segmentation + +Qualitative Comparisons. We present the segmentation visualization results on the challenging new FMB dataset in Figure 6. In the daytime intersection scene, thanks to the powerful semantic priors provided by SAM, our method is the only one to completely distinguish between the truck and the bus. In the nighttime road scene, we successfully segment the sidewalk, achieving optimal performance. Additionally, we showcase the segmentation results on the MFNet dataset in Figure 7. Leveraging the advanced se + +![](images/de179d39cc08c18f9bc5ca1b2b4350100240603ef2fd05f5696322a31ad607ee.jpg) +Figure 7. Qualitative segmentation results of SOTA approaches on the MFNet dataset. + +
FrameworkSegformer - B3 [31]
MethodCarPer.BikeCurveC.S.C.C.BumpmIoU
Visible83.354.759.67.4631.037.430.844.6
Infrared79.063.851.47.1025.234.141.844.4
DDFM86.762.263.430.835.545.741.151.7
U2Fusion87.365.960.833.436.046.140.952.0
TarDAL86.467.162.428.430.545.041.551.0
SegMiF85.966.364.434.034.743.048.052.4
FILM86.666.461.936.332.046.746.752.7
SHIP86.763.959.733.535.244.442.451.5
MRFS86.161.559.730.635.244.938.450.4
EMMA86.165.663.825.429.144.941.351.7
TIMFusion86.458.760.329.833.644.138.349.8
SAGE87.967.564.341.234.146.846.754.0
+ +matic information learned by SAM, our method successfully segments distant small pedestrian targets during the daytime and near-invisible lane curve at night. + +Quantitative Comparisons. Table 1 presents a qualitative comparison of IoU results for two segmentation frameworks on the FMB dataset. The first framework involves re-training images generated by various fusion methods, while the second uses the output from an open-vocabulary segmentation network with prompt words, without retraining. In traditional comparisons, our method achieves a 3.0 mIoU improvement over the second-best approach, demonstrating its competitive edge across categories. Additionally, in the segmentation network that requires no training, our method also performs favorably, benefiting from the adaptability provided by SAM's semantic information. Table 2 further showcases the qualitative results on the MFNet dataset, which has lower resolution and fewer labels. Although our method shows some gaps in background categories, it achieves overall superiority in key perception categories and mean performance. Our method integrates SAM with the fusion network, achieving a breakthrough in performance for efficient segmentation. + +Table 2. Quantitative semantic segmentation results of SOTA approaches on the MFNet dataset. + +
VariantsmIoU +MFNet / FMBFusion Metrics +EN / SD/ SCD / MS-SSIM
I. The Impact of SAM
(a) w/o SAM50.7 / 56.56.713 / 36.93 / 1.612 / 1.127
(b) P from Ylabel53.2 / 56.46.239 / 33.52 / 1.412 / 1.102
(c) Swap SO53.4 / 55.56.312 / 34.79 / 1.316 / 1.097
II. Study on the SPA Module
(a) w/o Z52.3 / 57.26.333 / 32.81 / 1.384 / 1.002
(b) w/o k-v in PR52.1 / 56.16.417 / 37.19 / 1.261 / 1.063
(c) w/o PR53.3 / 57.26.556 / 35.48 / 1.576 / 1.046
III. Discussion on the Distillation Scheme
(a) w/o Lfea49.7 / 51.66.436 / 37.28 / 1.277 / 1.079
(b) w/o Lcont50.9 / 56.26.669 / 37.80 / 1.462 / 1.086
(c) w/o Lcs51.7 / 54.36.524 / 39.11 / 1.356 / 1.038
(d) Offline Dist.50.0 / 50.46.221 / 32.65 / 1.581 / 0.987
SAGE54.0 / 61.26.872 / 40.77 / 1.604 / 1.117
+ +Table 3. Quantitative results on both fusion and segmentation of all variants in three ablation studies. + +# 4.4. Ablation Studies + +The Impact of SAM. The Segment Anything Module serves as a critical cornerstone in our method. To explore its impact, we derive three variants by isolating SAM from the core framework. Specifically, in variant (a), we replace the semantic patches with randomly cropped source image patches, thereby removing SAM from the main network $\mathcal{M}$ . In variant (b), we assist the generation of semantic patches with segmentation labels. Additionally, we replace $S_{\mathrm{O}}$ with a conventional segmentation network [31]. Variant (a) enhances the source image information, leading to the optimal similarity metrics, SCD and MS-SSIM, as shown in Table 3. Our method fully leverages the semantic priors provided by SAM, resulting in the best foggy building contours, as illustrated in the first row of Figure 8. + +Study on the SPA Module. The proposed Semantic Persistent Attention module plays a key role in integrating the semantic priors from SAM, effectively guiding the retention and enhancement of inherent modality features. To in- + +
MethodDDFMU2FusionTarDALSegMiFFILMSHIPMRFSEMMATIMFusionOurs
Time(ms)280K42.3116.57150.5183.327.9395.8625.7317.3910.47
FLOPs(G)1.34M518.0233.3500.4230.3401.5301.2106.4100.752.06
Params(M)552.70.6590.2970.6210.4910.525133.41.5160.1580.136
+ +Table 4. Efficiency comparison with other state-of-the-art methods on ${\mathrm{M}}^{3}\mathrm{{FD}}$ benchmarks. + +![](images/3c11fbabedee1423a3eaafa8b881bb45daf4d0d667e6384d5606cf1fe845b8cd.jpg) +Figure 8. Visualization comparison of impact of SAM. + +![](images/ed52bd26c737e908079d4aa8c62b6b71e5411f65e70743d2500a0b03cf2ed196.jpg) +Figure 9. Visualization comparison of ablation on SPA. + +vestigate its impact, we set up three variants: (a) $w/o$ latent representation $\mathbf{Z}$ , (b) $w/o$ key-value pairs in the Persistent Repository, and (c) $w/o$ PR. The channel loss due to these modifications has been corrected. A visual comparison of these ablation variants is shown in Figure 9. It is evident that PR in the SPA module plays an essential role in maintaining critical information. The removal of any component leads to significant information loss. For instance, in (b), while the regions of interest are highlighted, the absence of source image information causes blurriness. Similarly, the absence of PR in the SPA module completely prevents the capture of beneficial semantic information, causing the network to lose focus on key areas and resulting in a low-contrast fusion results in (c). + +Discussion on the Distillation Scheme. Furthermore, we discuss the distillation scheme. First, we conduct ablation studies on each component of the triple loss function, leading to variants (a)-(c) as shown in Table 3 III. Additionally, we replace the bi-level optimization distillation method with the offline distillation approach, forming variant (d). A visual comparison of these variants is shown in Figure 10. Notably, the results of the traditional online distillation are significantly inferior to our method, both in terms of visual quality and the gradient map generated by variant (b). + +![](images/2c87e915aacd7d5bff0bf0a30aa9bb89481f01e842716e81a4f92da736f98afb.jpg) +Figure 10. Visualization comparison of ablation on Distillation. + +This confirms the effectiveness of the proposed distillation method when incorporating semantic information. + +# 4.5. Computational Efficiency Analysis + +In the $\mathrm{M}^{3}\mathrm{FD}$ benchmark, we compare our method with 9 other SOTA methods in terms of time, FLOPS, and parameters. As shown in Table 4, our method demonstrates significant advantages across all aspects, particularly in time and FLOPs. Specifically, due to our distillation scheme, our sub-network is able to adjust flexibly while maintaining high efficiency, resulting in reduced FLOPs. Compared to other methods, our approach achieves a processing time of $10.47\mathrm{ms}$ and FLOPs of $52.06\mathrm{G}$ , outperforming most of the existing methods, with a parameter count of only $0.136\mathrm{M}$ , showcasing computational efficiency. + +Our method effectively alleviates the computational burden of SAM during inference, significantly reducing the computational overhead while preserving semantic information. This design allows our network to strike a balance between processing speed and computational resources, leading to superior efficiency in practical applications. + +# 5. Conclusion + +In this work, we propose a fusion method that utilizes semantic priors from the SAM for IVIF. We design the Semantic Persistent Attention module, which integrates semantic information while retaining source details. Additionally, we introduce a bi-level distillation scheme with triplet loss to reduce computational complexity by decoupling fusion from SAM during inference. Extensive experiments show that our method achieves superior fusion and task performance, outperforming across multiple datasets. + +# Acknowledgments + +This work is partially supported by the National Natural Science Foundation of China (Nos. 62450072, U22B2052, 62302078), Central Guidance for Local Science and Technology Development Fund (Youth Science Fund Project, Category A, No. 2025JH6/101100001), the Distinguished Young Scholars Funds of Dalian (No. 2024RJ002), the China Postdoctoral Science Foundation (No. 2023M730741) and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1 +[2] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multispectral scenes. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5108-5115. IEEE, 2017. 5 +[3] Zhiying Jiang, Zengxi Zhang, Jinyuan Liu, Xin Fan, and Risheng Liu. Multispectral image stitching via global-aware quadrature pyramid regression. IEEE Transactions on Image Processing, pages 4288-4302, 2024. 2 +[4] Zheyan Jin, Shiqi Chen, Yueting Chen, Zhihai Xu, and Huajun Feng. Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870, 2023. 2 +[5] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of Computational Methods in Engineering, 28(7):4425-4447, 2021. 1 +[6] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2 +[7] Guanlin Li, Bin Zhao, and Xuelong Li. Low-light image enhancement with sam-based structure priors and guidance. IEEE Transactions on Multimedia, pages 10854-10866, 2024. 2 +[8] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11040-11052, 2023. 1, 2 +[9] Siwei Li, Mingxuan Liu, Yating Zhang, Shu Chen, Haoxiang Li, Zifei Dou, and Hong Chen. Sam-deblur: Let segment anything boost image deblurring. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2445-2449. IEEE, 2024. 2 +[10] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In Proceedings of the International Conference on Learning Representations, 2019. 4 +[11] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual + +adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5 +[12] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, pages 1748-1775, 2023. 1, 2 +[13] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8115–8124, 2023. 1, 2, 5 +[14] Jinyuan Liu, Guanyao Wu, Zhu Liu, Long Ma, Risheng Liu, and Xin Fan. Where elegance meets precision: Towards a compact, automatic, and flexible framework for multimodality image fusion and applications. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1110-1118, 2024. 2 +[15] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, Xin Fan, and Risheng Liu. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(4):2349-2369, 2025. 1 +[16] Risheng Liu, Shichao Cheng, Yi He, Xin Fan, Zhouchen Lin, and Zhongxuan Luo. On the convergence of learning-based iterative methods for nonconvex inverse problems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12):3027-3039, 2019. 1 +[17] Risheng Liu, Jinyuan Liu, Zhiying Jiang, Xin Fan, and Zhongxuan Luo. A bilevel integrated model with data-driven layer ensemble for multi-modality image fusion. IEEE Transactions on Image Processing, 30:1261-1274, 2020. 1 +[18] Risheng Liu, Zhiying Jiang, Shuzhou Yang, and Xin Fan. Twin adversarial contrastive learning for underwater image enhancement and beyond. IEEE Transactions on Image Processing, 31:4922-4936, 2022. 1 +[19] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and meta-initialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6594-6609, 2024. 2, 5 +[20] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1240-1248, 2023. 1, 2 +[21] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1, 6 +[22] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1 + +[23] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2 +[24] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2 +[25] Linfeng Tang, Yuxin Deng, Yong Ma, Jun Huang, and Jiayi Ma. Superfusion: A versatile image registration and fusion network with semantic awareness. IEEE/CAA Journal of Automatica Sinica, 9(12):2121-2137, 2022. 2 +[26] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2 +[27] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5 +[28] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image superresolution. arXiv preprint arXiv:2402.17133, 2024. 2 +[29] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 3508-3515, 2022. 2 +[30] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, pages 1-36, 2023. 2 +[31] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 5, 6, 7 +[32] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5 +[33] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 1 +[34] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 6 +[35] Hao Zhang, Han Xu, Xin Tian, Junjun Jiang, and Jiayi Ma. Image fusion meets deep learning: A survey and perspective. Information Fusion, 76:323-336, 2021. 1 + +[36] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 2, 5 +[37] Xingchen Zhang, Ping Ye, and Gang Xiao. Vifb: A visible and infrared image fusion benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 104-105, 2020. 1 +[38] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 2 +[39] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 2 +[40] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 1, 2, 5 +[41] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25912-25921, 2024. 5 +[42] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 5 +[43] Naishan Zheng, Man Zhou, Jie Huang, Junming Hou, Haoying Li, Yuan Xu, and Feng Zhao. Probing synergistic high-order interaction in infrared and visible image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26384-26395, 2024. 5 +[44] Huabing Zhou, Jilei Hou, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Unified gradient-and intensity-discriminator generative adversarial network for image fusion. Information Fusion, 88:184-201, 2022. 2 +[45] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 5, 6 +[46] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36:19769-19782, 2023. 5 \ No newline at end of file diff --git a/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/images.zip b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..545bc324b8c63a14f37992518af2efa070433e45 --- /dev/null +++ b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ae3fdee1a277385f42d4e13396bb3d4144d65e23753d698e03a4124a4bfb7dd +size 1145565 diff --git a/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/layout.json b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d7c1b504aed79cbb0887bdc97a6d38a18862ec3f --- /dev/null +++ b/2025/Every SAM Drop Counts_ Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond/layout.json @@ -0,0 +1,7576 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 486, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 486, + 140 + ], + "type": "text", + "content": "Every SAM Drop Counts: Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 160, + 544, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 160, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 544, + 175 + ], + "type": "text", + "content": "Guanyao Wu†, Haoyu Liu†, Hongming Fu†, Yichuan Peng†, Jinyuan Liu‡, Xin Fan†, Risheng Liu†*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 175, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 175, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 129, + 175, + 482, + 189 + ], + "type": "text", + "content": "†School of Software Technology, Dalian University of Technology, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 190, + 490, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 190, + 490, + 204 + ], + "spans": [ + { + "bbox": [ + 121, + 190, + 490, + 204 + ], + "type": "inline_equation", + "content": "^{\\ddagger}" + }, + { + "bbox": [ + 121, + 190, + 490, + 204 + ], + "type": "text", + "content": "School of Mechanical Engineering, Dalian University of Technology, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 205, + 517, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 205, + 517, + 217 + ], + "spans": [ + { + "bbox": [ + 90, + 205, + 517, + 217 + ], + "type": "text", + "content": "rollingplainko@gmail.com, atlantis918@hotmail.com, {xin.fan, rsliu}@dlut.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 270, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 296, + 558 + ], + "type": "text", + "content": "Multi-modality image fusion, particularly infrared and visible, plays a crucial role in integrating diverse modalities to enhance scene understanding. Although early research prioritized visual quality, preserving fine details and adapting to downstream tasks remains challenging. Recent approaches attempt task-specific design but rarely achieve \"The Best of Both Worlds\" due to inconsistent optimization goals. To address these issues, we propose a novel method that leverages the semantic knowledge from the Segment Anything Model (SAM) to Grow the quality of fusion results and Enable downstream task adaptability, namely SAGE. Specifically, we design a Semantic Persistent Attention (SPA) Module that efficiently maintains source information via the persistent repository while extracting high-level semantic priors from SAM. More importantly, to eliminate the impractical dependence on SAM during inference, we introduce a bi-level optimization-driven distillation mechanism with triplet losses, which allow the student network to effectively extract knowledge. Extensive experiments show that our method achieves a balance between high-quality visual results and downstream task adaptability while maintaining practical deployment efficiency. The code is available at https://github.com/RollingPlain/SAGE_IVIF." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "content": "In light of recent advancements in sensing technologies, multi-modality imaging [15] has gained significant attention, with applications in robotics [35], remote sensing [33], and autonomous driving [1]. Integrating infrared and visible sensors provides substantial advantages for intelligent processing [5], although both exhibit inherent limitations [16, 21]. Infrared images are robust against smoke, obstruction, and low light, while visible images excel in reso" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 244, + 554, + 376 + ], + "blocks": [ + { + "bbox": [ + 317, + 244, + 554, + 376 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 554, + 376 + ], + "type": "image", + "image_path": "032986408fbfc16d793221c9253fc95c0278746d72dd3613c50c98de82ac55c4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 386, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 496 + ], + "type": "text", + "content": "Figure 1. Differences between the proposed method and existing mainstream comparative approaches: (a) Traditional and early DL-based methods focus on the fusion visual effect. (b) Task-specific methods (e.g., TarDAL [11] & SegMiF [13]) introduce task loss and features that lead to inconsistent optimization goals, causing a conflict between visual and task accuracy. (c) Our pipeline first leverages semantic priors from SAM within a large network and then distills the knowledge into a smaller sub-network achieving practical inference feasibility while ensuring \"the best of both worlds\" through SAM's inherent adaptability to these tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 507, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 568 + ], + "type": "text", + "content": "lution, contrast, and texture detail. These characteristics are highly complementary, motivating their fusion to generate a comprehensive fused image that balances visual quality and downstream task requirements. Consequently, achieving this fusion efficiently remains a pressing challenge [15]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Traditional Infrared and Visible Image Fusion (IVIF) methods, based on information theory [17, 37], aim to retain as much source image information as possible but often struggle with optimizing fused image quality, especially in managing redundancy and specific scenarios. Early deep learning-based methods [8, 12, 18, 22, 32, 39, 40] focused on visualization but faced issues like edge blurring and artifacts, failing to meet downstream perception needs. Current methods [11, 13, 20, 26] focus on specific tasks, coupling fusion with downstream processing, resulting in conflicting optimization objectives and challenges in balancing visual quality with task adaptability." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17882" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 228 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 228 + ], + "type": "image", + "image_path": "74df534b7eea6c9ad180d230068b72d93948a1cd34ac0e8750cda437264a6211.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 295, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 259 + ], + "type": "text", + "content": "Figure 2. Demonstration of SAM's robustness in MFNet normal scenes (top) and under challenging FMB conditions (bottom)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 269, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 295, + 388 + ], + "type": "text", + "content": "Recent advancements in large-scale visual models have significantly improved various visual analysis tasks [30]. Among these, the Segment Anything Model [6] stands out due to its exceptional and robust capability to provide rich semantic information, making it well-suited for IVIF, as illustrated in Figure 2. Also, its segmentation capabilities align naturally with the requirements of downstream tasks in IVIF field. However, current methods that integrate SAM for low-level tasks typically require full SAM during inference, resulting in excessive practical infeasibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 388, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 628 + ], + "type": "text", + "content": "To address these issues, we propose the fusion method SAGE that fully integrates and distills semantic priors derived from SAM. As illustrated in Figure 1 III, one of our aim is to harness the advantages of SAM's semantic priors while reducing computational burden. The proposed method consists of two key components: Semantic Persistent Attention module and a bi-level distillation scheme. The former focuses on using a persistent repository throughout the process to retain source image information, while heuristically enabling the network to integrate semantic patches predicted by SAM. Furthermore, the distillation scheme is designed to efficiently exclude SAM from the inference process, thereby reducing computational complexity. By enhancing distilled information across different aspects, the triplet losses in the scheme significantly boost the performance of the fusion model. Consequently, efficient fusion is achieved using only the distilled model, delivering high-quality visual results and precise task performance without the direct involvement of SAM. In summary, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 630, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 666 + ], + "type": "text", + "content": "- A novel fusion framework is proposed to leverage semantic priors from the SAM, effectively balancing visual quality and downstream task adaptability.." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "- The designed Semantic Persistent Attention (SPA) module utilizes a persistent repository to efficiently retain source information and simultaneously extract high-level semantic representation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "type": "text", + "content": "- We develop a bi-level optimization distillation scheme that transfers the information processed by SPA into the sub-network, effectively decoupling the fusion process from SAM during inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "content": "- Extensive experiments on multiple datasets validate the superior effectiveness of our proposed method." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 155, + 406, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 406, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 406, + 167 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "text", + "content": "Learning based IVIF Methods. In recent years, deep learning based multi-modality image fusion approaches [8, 12, 32, 39, 40, 44] achieved significant progress, primarily focusing on improving fusion quality, such as enhancing image details, maintaining structural consistency, and reducing noise. These approaches, including innovations in adversarial learning, Transformer models and diffusion network, have provided new insights. However, most of these methods have not addressed how the fused images can directly contribute to downstream tasks such as object detection, semantic segmentation, and other perception-related applications. With the growing demand for more integrated solutions, this gap has sparked increasing research interest in combining fusion with various downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 498 + ], + "type": "text", + "content": "Early works began to bridge this gap, with TarDAL [11] pioneering the use of a dual-discriminator network for object detection optimization, which simultaneously considers modality differences and detection tasks during fusion. As the field has progressed, several methods [24, 38] have also focused on optimizing fusion for object detection. For semantic segmentation [36], SeAFusion [26] and SegMiF [13] have made early significant improvements, with the former enhancing fusion via task loss and the latter leveraging task-specific features. More recently, some approaches [3, 14, 19, 20, 25, 29] have integrated multi-task learning to jointly optimize detection and segmentation, improving performance on both tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "SAM & Its Application. The Segment Anything Model (SAM) [6] is a large-scale pre-trained foundation model that excels in zero-shot generalization, enabling it to handle a wide range of segmentation and detection tasks without the need for task-specific training. SAM demonstrates impressive performance in low-level vision applications, such as image deblurring [9], dehazing [4], super-resolution [28], and low-light image enhancement [7]. Moreover, SAM's capabilities extend to high-level tasks like object detection [23], which further strengthens its versatility. This ability to handle both low-level and high-level tasks makes SAM inherently well-suited for the needs of the IVIF field, where fusion based segmentation and detection tasks are critical. However, its impressive performance is hindered by high computational complexity, posing challenges for practical deployment. Addressing these computational burdens, while fully leveraging the semantic priors of SAM, is a primary motivation for our work." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17883" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 382, + 254 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 382, + 254 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 382, + 254 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 382, + 254 + ], + "type": "image", + "image_path": "d350ee2af8cb773fdc2ef3c1c3ba9e565d318ee92292168bdf0165f0bb39a7b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 307 + ], + "type": "text", + "content": "Figure 3. A overall workflow of our proposed method. (a) shows the flow structure of the main network, where the SPA module processes patches with semantic priors generated by SAM. (b) illustrates the detailed structure of the SPA module, where PR plays a key role in preserving the source and integrating the semantic information. (c) displays our distillation scheme formulation, with visualizations of the different components of the triplet loss. (d) provides a simple diagram of the sub-network, which is composed of stacked dense blocks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 385, + 70, + 552, + 254 + ], + "blocks": [ + { + "bbox": [ + 385, + 70, + 552, + 254 + ], + "lines": [ + { + "bbox": [ + 385, + 70, + 552, + 254 + ], + "spans": [ + { + "bbox": [ + 385, + 70, + 552, + 254 + ], + "type": "image", + "image_path": "322fa75d65a95cc9cc44b6529ffb30411e2ee5c2916d7b3334f1dc08309dc097.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 185, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 185, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 185, + 340 + ], + "type": "text", + "content": "3. The Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 180, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 180, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 180, + 358 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "type": "text", + "content": "The overall workflow of our method is illustrated in Figure 3. Our aim is to fully exploit the semantic priors of SAM during the inference stage to enhance cross-modal fusion quality. However, directly deploying a large-scale SAM model often leads to prohibitive computational overhead. To mitigate this, we adopt a knowledge distillation strategy, transferring the information encoded by the SAM-driven main network to a lightweight sub-network, thereby significantly reducing inference costs while maintaining high-quality fusion. Nevertheless, the substantial capacity gap between the SAM-enhanced main network and the compact sub-network frequently results in incomplete semantic transfer or structural inconsistencies, which hinder ideal cross-modal fusion performance. To address this issue, we propose a bi-level optimization framework that jointly optimizes both networks as a unified system, aiming to bridge the distillation gap and maintain consistent fusion guided by SAM's semantic priors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": " denote the source visible and infrared input images, i.e., " + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{s}} = \\{\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}}\\}" + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": ". Their corresponding semantic-prior patches generated under the guidance of SAM are denoted as " + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{s}} = \\{\\mathbf{P}_{\\mathrm{vis}},\\mathbf{P}_{\\mathrm{ir}}\\}" + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": ". Denote the main network by " + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathbf{m}}" + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": ", and the sub-network by " + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathbf{s}}" + }, + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": ". We define the optimization process as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 658, + 294, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 658, + 294, + 682 + ], + "spans": [ + { + "bbox": [ + 70, + 658, + 294, + 682 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\omega} _ {s}} \\mathcal {L} _ {\\mathrm {s}} ^ {\\mathrm {d}} \\left(\\mathcal {S} \\left(\\mathbf {I} _ {s}, \\boldsymbol {\\omega} _ {s}\\right); \\boldsymbol {\\omega} _ {m} ^ {*}\\right), \\tag {1}", + "image_path": "b9abfd7323ad369bff4ef356b2e7c796a0fdda01904fe315fd718e0db69a60f0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 681, + 268, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 681, + 268, + 697 + ], + "spans": [ + { + "bbox": [ + 72, + 681, + 268, + 697 + ], + "type": "interline_equation", + "content": "\\mathrm {s . t .} \\quad \\boldsymbol {\\omega} _ {m} ^ {*} \\in \\arg \\min _ {\\boldsymbol {\\omega} _ {m}} \\mathcal {L} _ {\\mathrm {m}} ^ {\\mathrm {d}} \\big (\\mathcal {M} (\\mathbf {I} _ {s}, \\mathbf {P} _ {s}, \\boldsymbol {\\omega} _ {m}); \\boldsymbol {\\omega} _ {s} \\big),", + "image_path": "e2470d23e48a28c587f09ecb58f34d2d6e26210b092d080707712535b16ee3ad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": " is the distillation loss that ensures high-quality fusion with meaningful semantic cues in the main network, and " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": " is the distillation loss guiding the subnetwork to effectively mimic the main network's behavior. The fused reference image is given by: " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}} = \\mathcal{M}\\big(\\{\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}}\\} ,\\{\\mathbf{P}_{\\mathrm{vis}},\\mathbf{P}_{\\mathrm{ir}}\\} ,\\boldsymbol {\\omega}_m\\big)" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": ", while the sub-network produces a fused image " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{f}} = S\\big(\\mathbf{I}_{\\mathrm{vis}},\\mathbf{I}_{\\mathrm{ir}},\\boldsymbol {\\omega}_s\\big)" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": ". Notably, both distillation objectives rely on outputs from the counterpart network, establishing a bidirectional dependency, which is a key feature of our distillation formulation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 556 + ], + "type": "text", + "content": "Within the overall framework, the main network integrates SAM-derived semantic priors into the fusion process, while the sub-network is optimized to align with its outputs under the bi-level formulation. The following sections elaborate on two key components that support this process: the Semantic Persistent Attention (SPA) module 3.2, which integrates SAM-derived priors with source features, and the triplet-based loss scheme 3.3, which facilitates structured knowledge transfer across networks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 515, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 515, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 515, + 575 + ], + "type": "text", + "content": "3.2. Semantic Persistent Attention Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "content": "To fully exploit the semantic information provided by SAM, we propose the Semantic Persistent Attention (SPA) module, as illustrated in the flowchart in Figure 3 (b)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "The key core of the SPA module is a Persistent Repository (PR), which serves as a static memory to store and maintain crucial contextual information for the fusion process. Specifically, PR stores the latent representation " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z})" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " of the source features " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " and the corresponding key-value pairs " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_{\\mathrm{src}},\\mathbf{V}_{\\mathrm{src}})" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": ", which provide consistent contextual support during the cross-attention operation. The core insight behind this design is that PR acts as a stable, modality-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17884" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 550, + 304 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 550, + 304 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 550, + 304 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 550, + 304 + ], + "type": "image", + "image_path": "001ea98476575670761444f88549e1308937d98346e53e799a311ce46dfd5492.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 312, + 553, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 553, + 324 + ], + "type": "text", + "content": "Figure 4. Qualitative demonstrations of SOTA approaches across commonly used datasets, including TNO, RoadScene, " + }, + { + "bbox": [ + 55, + 312, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 55, + 312, + 553, + 324 + ], + "type": "text", + "content": " and FMB." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 345, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 295, + 380 + ], + "type": "text", + "content": "specific information source, guiding the cross-attention mechanism to fuse the semantic patches with rich contextual details from the original source." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": "The visible and infrared semantic patches " + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "inline_equation", + "content": "(\\mathbf{P}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{ir}}" + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": " extracted from the SAM are encoded into features " + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_{\\mathrm{pvis}}" + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{pir}})" + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": ". These encoded semantic features represent limited portions of the scene and are processed through the cross-attention mechanism. By utilizing the key-value pairs " + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_{\\mathrm{src}},\\mathbf{V}_{\\mathrm{src}})" + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": " stored in PR, the cross-attention mechanism enriches the semantic queries " + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_{\\mathrm{pvis}}" + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{pir}})" + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": " with the full scene context from the source image, thereby addressing the inherent incomplete scene coverage in the patches. PR ensures the fusion process stays grounded in a stable, consistent context, enabling patches to be effectively enriched with modality-specific information. This stability maintains semantic coherence while allowing flexible feature refinement via attention mechanisms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 674 + ], + "type": "text", + "content": "In summary, the SPA module uses PR to guide the cross-attention mechanism, ensuring that the semantic patches from SAM are enriched with consistent and detailed modality-specific information. This approach allows the fusion of infrared and visible image modalities by combining both the high-level semantic understanding from SAM and the detailed information from the source features. The final output, " + }, + { + "bbox": [ + 55, + 555, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{SPA}}" + }, + { + "bbox": [ + 55, + 555, + 295, + 674 + ], + "type": "text", + "content": ", represents a semantically enriched, structurally consistent feature set, rich in semantic priors of SAM, and ready for further processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The main network heavily utilizes the SPA module, aiming to fully exploit the complex semantic priors provided by SAM, focus on capturing and retaining detailed semantic" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 345, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 553, + 368 + ], + "type": "text", + "content": "knowledge, thereby facilitating a powerful representation that can later be distilled into a more efficient sub-network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 376, + 499, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 499, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 499, + 388 + ], + "type": "text", + "content": "3.3. Triplet Loss in Distillation Scheme" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 393, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 559 + ], + "type": "text", + "content": "In this section, we introduce a triplet loss-driven distillation scheme that facilitates semantic knowledge transfer from the main network " + }, + { + "bbox": [ + 313, + 393, + 555, + 559 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 313, + 393, + 555, + 559 + ], + "type": "text", + "content": " to the sub-network " + }, + { + "bbox": [ + 313, + 393, + 555, + 559 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 313, + 393, + 555, + 559 + ], + "type": "text", + "content": " under a bi-level optimization framework. Specifically, we adopt a DARTS-style [10] training protocol, where the main network (teacher) and the sub-network (student) are updated in small alternating steps, approximating the inner-outer structure of bilevel optimization. This allows gradients to flow in a bidirectional manner: the sub-network learns from the distillation signals provided by the main network, while the main network also adapts its parameters based on the subnetwork's performance and a segmentation objective. Consequently, the two networks reach a cohesive transfer with mutual compromise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "content": "Our distillation scheme is driven by three distinct losses, each focusing on a different aspect of the fused images. Let " + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "content": " be the fused reference image produced by the main network, and " + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{f}}" + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "content": " be the fused output from the sub-network. We first introduce a feature alignment term:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 352, + 629, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 629, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 352, + 629, + 553, + 661 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {f e a}} = \\sum_ {m = 1} ^ {M} \\left(1 - \\frac {\\mathbf {F} _ {\\mathrm {D e n}} ^ {m} \\cdot \\mathbf {F} _ {\\mathrm {S P A}} ^ {m}}{\\| \\mathbf {F} _ {\\mathrm {D e n}} ^ {m} \\| _ {2} \\cdot \\| \\mathbf {F} _ {\\mathrm {S P A}} ^ {m} \\| _ {2}}\\right), \\tag {2}", + "image_path": "b56c1f1abba65e169b761a8268c56fc03acc0b354043502d78ba87bdb4ec0425.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{Den}}^{m}" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{SPA}}^{m}" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": " are the feature maps of the dense and SPA blocks at the same scale " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": "-norm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "The second loss component operates at the context level, which consists of two parts: a gradient part to preserve" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17885" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 306, + 187 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 306, + 187 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 306, + 187 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 306, + 187 + ], + "type": "image", + "image_path": "49b38b64bc7d95d32ba55ba54ff636e88c9b316038392efcf7859efefae20075.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "type": "text", + "content": "Figure 5. Quantitative comparison of fusion performance with other SOTA methods on " + }, + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "type": "text", + "content": " and FMB benchmarks. The violin plots illustrate the distribution of metrics, in which the black lines and white triangles indicate the medium and mean values." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 553, + 187 + ], + "type": "image", + "image_path": "4e0c0959e472e289dd598aa7c93f897d3a5a6ffb075269bf86204727ad087a10.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "text", + "content": "structural consistency and an MSE loss for intensity consistency. They are defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 284, + 295, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 284, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 63, + 284, + 295, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g r a d}} = \\left\\| \\nabla \\mathbf {I} _ {\\mathrm {r e f}} - \\nabla \\mathbf {I} _ {\\mathrm {f u s}} \\right\\| _ {1}, \\quad \\mathcal {L} _ {\\mathrm {M S E}} = \\left\\| \\mathbf {I} _ {\\mathrm {r e f}} - \\mathbf {I} _ {\\mathrm {f u s}} \\right\\| _ {2}, \\tag {3}", + "image_path": "6ce1123033c8fb43af2392729fa19e268def44acf993b92bbe7d6c9e9a979bfb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "content": " denotes the Sobel operator and " + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _1" + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "content": " -norm. So the cont loss can be calculated as " + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{context}} = \\mathcal{L}_{\\mathrm{grad}} + \\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "content": ". In practice, we apply these terms not only between " + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{fus}}" + }, + { + "bbox": [ + 55, + 307, + 295, + 391 + ], + "type": "text", + "content": ", but also between each fused output and the original source images to ensure the reconstruction fidelity. This prevents both networks from drifting away from the source image during the distillation process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": "The third loss component, contrastive semantic loss aims to ensure that the feature space of the fused image remains close to that of the reference image, while being distinct from the individual visible and infrared images archiving efficient distillation. We utilize the semantic feature space by encoder of SAM (defined as " + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": ") to construct positive and negative pairs through the element-wise multiplication of the binary masks " + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{ir}})" + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{ir}}" + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": ". The contrastive semantic loss on each modality is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 520, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 520, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 71, + 520, + 295, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c s}} ^ {\\mathrm {i r}} = \\sum_ {l = 1} ^ {L} \\frac {\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {f u s}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {r e f}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) \\right\\| _ {2}}{\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {x}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {i r}} \\odot \\mathbf {M} _ {\\mathrm {i r}}\\right) \\right\\| _ {2}}, \\tag {4}", + "image_path": "650cabca188d8ea861b2678c1cc41230911c7bb93f78d8ab1edb228c103cea4b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 575, + 295, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 575, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 66, + 575, + 295, + 608 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c s}} ^ {\\mathrm {v i s}} = \\sum_ {l = 1} ^ {L} \\frac {\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {f u s}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {r e f}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) \\right\\| _ {2}}{\\left\\| \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\mathrm {x}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) - \\mathcal {S} _ {\\mathrm {E}} \\left(\\mathbf {I} _ {\\text {v i s}} \\odot \\mathbf {M} _ {\\text {v i s}}\\right) \\right\\| _ {2}}, \\tag {5}", + "image_path": "853f9195b9bb6d09553b4db1d962cfb39dc5caadfef2ed509e87b8082e7993f8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": " indexes the layers of " + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\{\\mathrm{ref}, \\mathrm{fus}\\}" + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": ". So the total " + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cs}} = \\mathcal{L}_{\\mathrm{cs}}^{\\mathrm{ir}} + \\mathcal{L}_{\\mathrm{cs}}^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": " can also be computed bidirectionally in the bi-level optimization scheme between two networks, ensuring that the feature spaces of both the " + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": " are easily aligned." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "The total distillation loss for the sub-network, denoted as " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " in Eq. (1), is the sum of the feature loss, context loss, and contrastive semantic loss: " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}} = \\mathcal{L}_{\\mathrm{fea}} + \\mathcal{L}_{\\mathrm{cont}} + \\mathcal{L}_{\\mathrm{cs}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "text", + "content": "For the main network, an additional segmentation cross-entropy loss " + }, + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "text", + "content": " is incorporated. This loss is computed between the segmentation predictions generated by an open-vocabulary model " + }, + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{O}}" + }, + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "text", + "content": " [45, 46] and the ground truth segmentation labels. Instead of affecting the sub-network, this loss aims to prevent potential optimization conflicts. The segmentation loss " + }, + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 238, + 555, + 323 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 331, + 553, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 331, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 365, + 331, + 553, + 355 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e g}} = - \\sum_ {c} \\left[ \\mathbf {Y} _ {\\text {l a b e l}} ^ {c} \\log \\left(\\mathbf {Y} _ {\\text {p r e d}} ^ {c}\\right) \\right], \\tag {6}", + "image_path": "342afe7b2a4f6d1b728f40593e3d94b1fa9982e164f562b0f42b7337d6924c38.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathrm{label}}" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": " represents the label segmentation map, " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": " is the predicted segmentation map, and " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": " is the class index. For now, the total distillation loss for the main network can be calculated as: " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}^{\\mathrm{d}} = \\mathcal{L}_{\\mathrm{s}}^{\\mathrm{d}} + \\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 414, + 397, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 397, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 397, + 427 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 434, + 430, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 430, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 430, + 445 + ], + "type": "text", + "content": "4.1. Settings and Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "text", + "content": "Five representative datasets, namely TNO [27], RoadScene [32], MFNet [2], FMB [13], and " + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "text", + "content": " [11], are utilized for both training and evaluation. For segmentation, SegFormer [31] (B3 variant) is adopted as the backbone, and the models are trained for 100 epochs. The training and testing splits are strictly in accordance with the official dataset guidelines. Adam is adopted for training, with initial learning rates of " + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "inline_equation", + "content": "5\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "inline_equation", + "content": "2\\times 10^{-3}" + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "text", + "content": " for the main and sub-networks, respectively. Cosine annealing decays both to " + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "inline_equation", + "content": "1\\times 10^{-5}" + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "text", + "content": ". Prior to distillation, the models undergo a pre-training phase, followed by 5 epochs of distillation. The batch size is set to 4, and the images are randomly cropped and resized to " + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "inline_equation", + "content": "192\\times 256" + }, + { + "bbox": [ + 313, + 450, + 555, + 629 + ], + "type": "text", + "content": " during training. The entire framework is implemented in PyTorch and executed on two NVIDIA GeForce RTX 4090 GPUs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 524, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 524, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 524, + 649 + ], + "type": "text", + "content": "4.2. Results of Multi-modality Image Fusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "We demonstrate our fusion quality by visualization and quantitative analysis with 9 other SOTA methods in recent years, including DDFM [40], U2Fusion [32], TarDAL [11], SegMiF [13], FILM [42], SHIP [43], MRFS [36], EMMA [41] and TIMFusion [19]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17886" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 550, + 249 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 550, + 249 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 550, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 550, + 249 + ], + "type": "table", + "html": "
FrameworkSegformer - B3 [31]X - Decoder [45]
MethodVenueRoadBldg.SignVeg.BusPer.CarmIoURoadBldg.SignVeg.BusPer.CarmIoU
Visible-84.881.064.284.746.861.177.149.179.267.818.273.252.056.982.050.1
Infrared-83.777.655.980.938.157.174.242.776.655.612.938.272.763.278.742.9
DDFMICCV'2386.283.069.584.862.561.080.858.276.165.910.173.372.066.682.549.3
U2FusionTPAMI'2286.782.468.584.650.361.880.457.079.770.622.272.775.765.583.350.4
TarDALCVPR'2285.881.868.484.749.561.379.555.774.163.123.267.574.866.481.148.7
SegMiFICCV'2386.382.268.085.063.760.780.057.379.668.614.073.874.365.883.250.8
FILMICML'2486.981.969.184.046.160.679.455.779.270.317.973.779.064.082.050.6
SHIPCVPR'2484.982.168.383.361.460.980.457.276.466.824.771.276.364.680.950.4
MRFSCVPR'2488.282.871.086.061.563.681.156.573.865.023.772.377.266.282.649.6
EMMACVPR'2485.281.266.883.359.058.979.955.878.668.415.074.377.164.582.250.7
TIMFusionTPAMI'2485.180.066.482.850.859.079.755.562.063.012.867.675.865.879.646.1
SAGEProposed89.083.971.486.366.764.181.761.281.070.417.075.277.764.082.851.1
", + "image_path": "23001fdd093c50df54392f11416e66fb81ca0262c86cc3469b3297a57552b8b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 256, + 474, + 268 + ], + "lines": [ + { + "bbox": [ + 134, + 256, + 474, + 268 + ], + "spans": [ + { + "bbox": [ + 134, + 256, + 474, + 268 + ], + "type": "text", + "content": "Table 1. Quantitative semantic segmentation results of SOTA approaches on the FMB dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 279, + 552, + 415 + ], + "blocks": [ + { + "bbox": [ + 59, + 279, + 552, + 415 + ], + "lines": [ + { + "bbox": [ + 59, + 279, + 552, + 415 + ], + "spans": [ + { + "bbox": [ + 59, + 279, + 552, + 415 + ], + "type": "image", + "image_path": "45efdd97dee50ffbedd6ace630d372671749efd86466b54db8f1f2ed245b9788.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 424, + 457, + 435 + ], + "lines": [ + { + "bbox": [ + 151, + 424, + 457, + 435 + ], + "spans": [ + { + "bbox": [ + 151, + 424, + 457, + 435 + ], + "type": "text", + "content": "Figure 6. Qualitative segmentation results of SOTA approaches on the FMB dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 451, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 451, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 54, + 451, + 295, + 630 + ], + "type": "text", + "content": "Qualitative Comparisons. Figure 4 presents a visual comparison of various methods. Overall, the proposed method has two main advantages. First, it effectively preserves the multi-modality information of the original images. In the TNO surveillance scene, the vegetation details from the visible image and the smoke from the chimney in the infrared image are both well retained (the image set in the top-left corner). In RoadScene, our method also achieves optimal leaf recovery. On the other hand, our approach demonstrates strong robustness to interference, as it accurately reconstructs the reflective pedestrian crossing lines on the ground and the distant building outlines in the dense fog at night (the green box in the second row). The integration of SAM enhances our method, giving it the ability to surpass other methods and achieve superior results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 641, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 712 + ], + "type": "text", + "content": "Quantitative Comparisons. We also compare our numerical results with other competitors in Figure 5, based on 100 image pairs randomly selected from the FMB and " + }, + { + "bbox": [ + 54, + 641, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 54, + 641, + 295, + 712 + ], + "type": "text", + "content": " datasets. We employ four fusion evaluation metrics [21], including the Entropy (EN), Standard Deviation (SD), Sum of Correlation Differences (SCD), and Multi-Scale Structural" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 553, + 487 + ], + "type": "text", + "content": "Similarity (MS-SSIM). Additionally, we introduce four no-reference image quality assessments [34] related to image quality: BRISQUE, NIQE, MUSIQ, and PaQ-2-PiQ." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 487, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 582 + ], + "type": "text", + "content": "In terms of these metrics, our method demonstrates consistent superiority. High SCD and MS-SSIM values indicate that our method effectively preserves a significant amount of information from the source images, while EN reflects our rich detail and texture. The spatial, multi-scale, and patch evaluations of the NR-IQA confirm that our fusion approach has a better pixel distribution and aligns more closely with the human visual system perception." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 588, + 523, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 523, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 523, + 601 + ], + "type": "text", + "content": "4.3. Results of Multi-modality Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Qualitative Comparisons. We present the segmentation visualization results on the challenging new FMB dataset in Figure 6. In the daytime intersection scene, thanks to the powerful semantic priors provided by SAM, our method is the only one to completely distinguish between the truck and the bus. In the nighttime road scene, we successfully segment the sidewalk, achieving optimal performance. Additionally, we showcase the segmentation results on the MFNet dataset in Figure 7. Leveraging the advanced se" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17887" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 550, + 206 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 550, + 206 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 550, + 206 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 550, + 206 + ], + "type": "image", + "image_path": "de179d39cc08c18f9bc5ca1b2b4350100240603ef2fd05f5696322a31ad607ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 215, + 461, + 226 + ], + "lines": [ + { + "bbox": [ + 148, + 215, + 461, + 226 + ], + "spans": [ + { + "bbox": [ + 148, + 215, + 461, + 226 + ], + "type": "text", + "content": "Figure 7. Qualitative segmentation results of SOTA approaches on the MFNet dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 241, + 299, + 403 + ], + "blocks": [ + { + "bbox": [ + 56, + 241, + 299, + 403 + ], + "lines": [ + { + "bbox": [ + 56, + 241, + 299, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 299, + 403 + ], + "type": "table", + "html": "
FrameworkSegformer - B3 [31]
MethodCarPer.BikeCurveC.S.C.C.BumpmIoU
Visible83.354.759.67.4631.037.430.844.6
Infrared79.063.851.47.1025.234.141.844.4
DDFM86.762.263.430.835.545.741.151.7
U2Fusion87.365.960.833.436.046.140.952.0
TarDAL86.467.162.428.430.545.041.551.0
SegMiF85.966.364.434.034.743.048.052.4
FILM86.666.461.936.332.046.746.752.7
SHIP86.763.959.733.535.244.442.451.5
MRFS86.161.559.730.635.244.938.450.4
EMMA86.165.663.825.429.144.941.351.7
TIMFusion86.458.760.329.833.644.138.349.8
SAGE87.967.564.341.234.146.846.754.0
", + "image_path": "839f780a6ec62f6ed4a2ea2c3bc760e1d061e926c7b2a05b8a5dfd5bd993ce55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 483 + ], + "type": "text", + "content": "matic information learned by SAM, our method successfully segments distant small pedestrian targets during the daytime and near-invisible lane curve at night." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Comparisons. Table 1 presents a qualitative comparison of IoU results for two segmentation frameworks on the FMB dataset. The first framework involves re-training images generated by various fusion methods, while the second uses the output from an open-vocabulary segmentation network with prompt words, without retraining. In traditional comparisons, our method achieves a 3.0 mIoU improvement over the second-best approach, demonstrating its competitive edge across categories. Additionally, in the segmentation network that requires no training, our method also performs favorably, benefiting from the adaptability provided by SAM's semantic information. Table 2 further showcases the qualitative results on the MFNet dataset, which has lower resolution and fewer labels. Although our method shows some gaps in background categories, it achieves overall superiority in key perception categories and mean performance. Our method integrates SAM with the fusion network, achieving a breakthrough in performance for efficient segmentation." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 241, + 554, + 426 + ], + "blocks": [ + { + "bbox": [ + 55, + 412, + 295, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 434 + ], + "type": "text", + "content": "Table 2. Quantitative semantic segmentation results of SOTA approaches on the MFNet dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 241, + 554, + 426 + ], + "lines": [ + { + "bbox": [ + 315, + 241, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 554, + 426 + ], + "type": "table", + "html": "
VariantsmIoU\nMFNet / FMBFusion Metrics\nEN / SD/ SCD / MS-SSIM
I. The Impact of SAM
(a) w/o SAM50.7 / 56.56.713 / 36.93 / 1.612 / 1.127
(b) P from Ylabel53.2 / 56.46.239 / 33.52 / 1.412 / 1.102
(c) Swap SO53.4 / 55.56.312 / 34.79 / 1.316 / 1.097
II. Study on the SPA Module
(a) w/o Z52.3 / 57.26.333 / 32.81 / 1.384 / 1.002
(b) w/o k-v in PR52.1 / 56.16.417 / 37.19 / 1.261 / 1.063
(c) w/o PR53.3 / 57.26.556 / 35.48 / 1.576 / 1.046
III. Discussion on the Distillation Scheme
(a) w/o Lfea49.7 / 51.66.436 / 37.28 / 1.277 / 1.079
(b) w/o Lcont50.9 / 56.26.669 / 37.80 / 1.462 / 1.086
(c) w/o Lcs51.7 / 54.36.524 / 39.11 / 1.356 / 1.038
(d) Offline Dist.50.0 / 50.46.221 / 32.65 / 1.581 / 0.987
SAGE54.0 / 61.26.872 / 40.77 / 1.604 / 1.117
", + "image_path": "43ba012009b327bad72c6544741a67bb9ea660e959cfbdefbcbe5f8092d6c61b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "type": "text", + "content": "Table 3. Quantitative results on both fusion and segmentation of all variants in three ablation studies." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 475, + 415, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 415, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 415, + 487 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": "The Impact of SAM. The Segment Anything Module serves as a critical cornerstone in our method. To explore its impact, we derive three variants by isolating SAM from the core framework. Specifically, in variant (a), we replace the semantic patches with randomly cropped source image patches, thereby removing SAM from the main network " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": ". In variant (b), we assist the generation of semantic patches with segmentation labels. Additionally, we replace " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{O}}" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " with a conventional segmentation network [31]. Variant (a) enhances the source image information, leading to the optimal similarity metrics, SCD and MS-SSIM, as shown in Table 3. Our method fully leverages the semantic priors provided by SAM, resulting in the best foggy building contours, as illustrated in the first row of Figure 8." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Study on the SPA Module. The proposed Semantic Persistent Attention module plays a key role in integrating the semantic priors from SAM, effectively guiding the retention and enhancement of inherent modality features. To in-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17888" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 71, + 550, + 130 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 550, + 130 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 550, + 130 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 550, + 130 + ], + "type": "table", + "html": "
MethodDDFMU2FusionTarDALSegMiFFILMSHIPMRFSEMMATIMFusionOurs
Time(ms)280K42.3116.57150.5183.327.9395.8625.7317.3910.47
FLOPs(G)1.34M518.0233.3500.4230.3401.5301.2106.4100.752.06
Params(M)552.70.6590.2970.6210.4910.525133.41.5160.1580.136
", + "image_path": "f58134d7d16e01c1c023ab1e6ab0fda3f671a7a23b64ebd6e5431af73e716367.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 137, + 468, + 148 + ], + "lines": [ + { + "bbox": [ + 140, + 137, + 468, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 137, + 468, + 148 + ], + "type": "text", + "content": "Table 4. Efficiency comparison with other state-of-the-art methods on " + }, + { + "bbox": [ + 140, + 137, + 468, + 148 + ], + "type": "inline_equation", + "content": "{\\mathrm{M}}^{3}\\mathrm{{FD}}" + }, + { + "bbox": [ + 140, + 137, + 468, + 148 + ], + "type": "text", + "content": " benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 156, + 297, + 251 + ], + "blocks": [ + { + "bbox": [ + 57, + 156, + 297, + 251 + ], + "lines": [ + { + "bbox": [ + 57, + 156, + 297, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 156, + 297, + 251 + ], + "type": "image", + "image_path": "3c11fbabedee1423a3eaafa8b881bb45daf4d0d667e6384d5606cf1fe845b8cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 258, + 276, + 270 + ], + "lines": [ + { + "bbox": [ + 75, + 258, + 276, + 270 + ], + "spans": [ + { + "bbox": [ + 75, + 258, + 276, + 270 + ], + "type": "text", + "content": "Figure 8. Visualization comparison of impact of SAM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 278, + 297, + 388 + ], + "blocks": [ + { + "bbox": [ + 57, + 278, + 297, + 388 + ], + "lines": [ + { + "bbox": [ + 57, + 278, + 297, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 278, + 297, + 388 + ], + "type": "image", + "image_path": "ed52bd26c737e908079d4aa8c62b6b71e5411f65e70743d2500a0b03cf2ed196.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 396, + 276, + 406 + ], + "lines": [ + { + "bbox": [ + 74, + 396, + 276, + 406 + ], + "spans": [ + { + "bbox": [ + 74, + 396, + 276, + 406 + ], + "type": "text", + "content": "Figure 9. Visualization comparison of ablation on SPA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "text", + "content": "vestigate its impact, we set up three variants: (a) " + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "text", + "content": " latent representation " + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "text", + "content": ", (b) " + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "text", + "content": " key-value pairs in the Persistent Repository, and (c) " + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 54, + 425, + 295, + 592 + ], + "type": "text", + "content": " PR. The channel loss due to these modifications has been corrected. A visual comparison of these ablation variants is shown in Figure 9. It is evident that PR in the SPA module plays an essential role in maintaining critical information. The removal of any component leads to significant information loss. For instance, in (b), while the regions of interest are highlighted, the absence of source image information causes blurriness. Similarly, the absence of PR in the SPA module completely prevents the capture of beneficial semantic information, causing the network to lose focus on key areas and resulting in a low-contrast fusion results in (c)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Discussion on the Distillation Scheme. Furthermore, we discuss the distillation scheme. First, we conduct ablation studies on each component of the triple loss function, leading to variants (a)-(c) as shown in Table 3 III. Additionally, we replace the bi-level optimization distillation method with the offline distillation approach, forming variant (d). A visual comparison of these variants is shown in Figure 10. Notably, the results of the traditional online distillation are significantly inferior to our method, both in terms of visual quality and the gradient map generated by variant (b)." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 157, + 556, + 284 + ], + "blocks": [ + { + "bbox": [ + 316, + 157, + 556, + 284 + ], + "lines": [ + { + "bbox": [ + 316, + 157, + 556, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 157, + 556, + 284 + ], + "type": "image", + "image_path": "2c87e915aacd7d5bff0bf0a30aa9bb89481f01e842716e81a4f92da736f98afb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 292, + 550, + 303 + ], + "lines": [ + { + "bbox": [ + 318, + 292, + 550, + 303 + ], + "spans": [ + { + "bbox": [ + 318, + 292, + 550, + 303 + ], + "type": "text", + "content": "Figure 10. Visualization comparison of ablation on Distillation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 315, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 338 + ], + "type": "text", + "content": "This confirms the effectiveness of the proposed distillation method when incorporating semantic information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 349, + 499, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 499, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 499, + 361 + ], + "type": "text", + "content": "4.5. Computational Efficiency Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "text", + "content": "In the " + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^{3}\\mathrm{FD}" + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "text", + "content": " benchmark, we compare our method with 9 other SOTA methods in terms of time, FLOPS, and parameters. As shown in Table 4, our method demonstrates significant advantages across all aspects, particularly in time and FLOPs. Specifically, due to our distillation scheme, our sub-network is able to adjust flexibly while maintaining high efficiency, resulting in reduced FLOPs. Compared to other methods, our approach achieves a processing time of " + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "inline_equation", + "content": "10.47\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "text", + "content": " and FLOPs of " + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "inline_equation", + "content": "52.06\\mathrm{G}" + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "text", + "content": ", outperforming most of the existing methods, with a parameter count of only " + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "inline_equation", + "content": "0.136\\mathrm{M}" + }, + { + "bbox": [ + 313, + 366, + 555, + 499 + ], + "type": "text", + "content": ", showcasing computational efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 499, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 554, + 572 + ], + "type": "text", + "content": "Our method effectively alleviates the computational burden of SAM during inference, significantly reducing the computational overhead while preserving semantic information. This design allows our network to strike a balance between processing speed and computational resources, leading to superior efficiency in practical applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "In this work, we propose a fusion method that utilizes semantic priors from the SAM for IVIF. We design the Semantic Persistent Attention module, which integrates semantic information while retaining source details. Additionally, we introduce a bi-level distillation scheme with triplet loss to reduce computational complexity by decoupling fusion from SAM during inference. Extensive experiments show that our method achieves superior fusion and task performance, outperforming across multiple datasets." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17889" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 295, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 295, + 183 + ], + "type": "text", + "content": "This work is partially supported by the National Natural Science Foundation of China (Nos. 62450072, U22B2052, 62302078), Central Guidance for Local Science and Technology Development Fund (Youth Science Fund Project, Category A, No. 2025JH6/101100001), the Distinguished Young Scholars Funds of Dalian (No. 2024RJ002), the China Postdoctoral Science Foundation (No. 2023M730741) and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 295, + 270 + ], + "type": "text", + "content": "[1] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 270, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 335 + ], + "type": "text", + "content": "[2] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multispectral scenes. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5108-5115. IEEE, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 337, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 337, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 295, + 381 + ], + "type": "text", + "content": "[3] Zhiying Jiang, Zengxi Zhang, Jinyuan Liu, Xin Fan, and Risheng Liu. Multispectral image stitching via global-aware quadrature pyramid regression. IEEE Transactions on Image Processing, pages 4288-4302, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 382, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 382, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 295, + 414 + ], + "type": "text", + "content": "[4] Zheyan Jin, Shiqi Chen, Yueting Chen, Zhihai Xu, and Huajun Feng. Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 447 + ], + "type": "text", + "content": "[5] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of Computational Methods in Engineering, 28(7):4425-4447, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 295, + 491 + ], + "type": "text", + "content": "[6] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 535 + ], + "type": "text", + "content": "[7] Guanlin Li, Bin Zhao, and Xuelong Li. Low-light image enhancement with sam-based structure priors and guidance. IEEE Transactions on Multimedia, pages 10854-10866, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "text", + "content": "[8] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11040-11052, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 295, + 646 + ], + "type": "text", + "content": "[9] Siwei Li, Mingxuan Liu, Yating Zhang, Shu Chen, Haoxiang Li, Zifei Dou, and Hong Chen. Sam-deblur: Let segment anything boost image deblurring. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2445-2449. IEEE, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[10] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In Proceedings of the International Conference on Learning Representations, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[11] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 127 + ], + "type": "text", + "content": "adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[12] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, pages 1748-1775, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 249 + ], + "type": "text", + "content": "[13] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8115–8124, 2023. 1, 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 250, + 555, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 250, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 250, + 555, + 315 + ], + "type": "text", + "content": "[14] Jinyuan Liu, Guanyao Wu, Zhu Liu, Long Ma, Risheng Liu, and Xin Fan. Where elegance meets precision: Towards a compact, automatic, and flexible framework for multimodality image fusion and applications. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1110-1118, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "text", + "content": "[15] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, Xin Fan, and Risheng Liu. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(4):2349-2369, 2025. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 372, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 554, + 425 + ], + "type": "text", + "content": "[16] Risheng Liu, Shichao Cheng, Yi He, Xin Fan, Zhouchen Lin, and Zhongxuan Luo. On the convergence of learning-based iterative methods for nonconvex inverse problems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12):3027-3039, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "type": "text", + "content": "[17] Risheng Liu, Jinyuan Liu, Zhiying Jiang, Xin Fan, and Zhongxuan Luo. A bilevel integrated model with data-driven layer ensemble for multi-modality image fusion. IEEE Transactions on Image Processing, 30:1261-1274, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "text", + "content": "[18] Risheng Liu, Zhiying Jiang, Shuzhou Yang, and Xin Fan. Twin adversarial contrastive learning for underwater image enhancement and beyond. IEEE Transactions on Image Processing, 31:4922-4936, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 515, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 554, + 568 + ], + "type": "text", + "content": "[19] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and meta-initialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6594-6609, 2024. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 570, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 554, + 624 + ], + "type": "text", + "content": "[20] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1240-1248, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "type": "text", + "content": "[21] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "text", + "content": "[22] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17890" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[23] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 174 + ], + "type": "text", + "content": "[24] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 294, + 220 + ], + "type": "text", + "content": "[25] Linfeng Tang, Yuxin Deng, Yong Ma, Jun Huang, and Jiayi Ma. Superfusion: A versatile image registration and fusion network with semantic awareness. IEEE/CAA Journal of Automatica Sinica, 9(12):2121-2137, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "text", + "content": "[26] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 294, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 290 + ], + "type": "text", + "content": "[27] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 292, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 336 + ], + "type": "text", + "content": "[28] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image superresolution. arXiv preprint arXiv:2402.17133, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[29] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 3508-3515, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 396, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 450 + ], + "type": "text", + "content": "[30] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, pages 1-36, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "type": "text", + "content": "[31] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 553 + ], + "type": "text", + "content": "[32] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "text", + "content": "[33] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 677 + ], + "type": "text", + "content": "[34] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[35] Hao Zhang, Han Xu, Xin Tian, Junjun Jiang, and Jiayi Ma. Image fusion meets deep learning: A survey and perspective. Information Fusion, 76:323-336, 2021. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 707 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "text", + "content": "[36] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "type": "text", + "content": "[37] Xingchen Zhang, Ping Ye, and Gang Xiao. Vifb: A visible and infrared image fusion benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 104-105, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 228 + ], + "type": "text", + "content": "[38] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 554, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 295 + ], + "type": "text", + "content": "[39] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 554, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 361 + ], + "type": "text", + "content": "[40] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 363, + 554, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 554, + 418 + ], + "type": "text", + "content": "[41] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25912-25921, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 419, + 554, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 554, + 483 + ], + "type": "text", + "content": "[42] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 485, + 554, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 485, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 485, + 554, + 540 + ], + "type": "text", + "content": "[43] Naishan Zheng, Man Zhou, Jie Huang, Junming Hou, Haoying Li, Yuan Xu, and Feng Zhao. Probing synergistic high-order interaction in infrared and visible image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26384-26395, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 541, + 554, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 554, + 585 + ], + "type": "text", + "content": "[44] Huabing Zhou, Jilei Hou, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Unified gradient-and intensity-discriminator generative adversarial network for image fusion. Information Fusion, 88:184-201, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 586, + 554, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 586, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 316, + 586, + 554, + 651 + ], + "type": "text", + "content": "[45] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 653, + 554, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 653, + 554, + 707 + ], + "spans": [ + { + "bbox": [ + 316, + 653, + 554, + 707 + ], + "type": "text", + "content": "[46] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36:19769-19782, 2023. 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17891" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_content_list.json b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a753880187094c5cce41e6d667e8e26d5289767f --- /dev/null +++ b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_content_list.json @@ -0,0 +1,1898 @@ +[ + { + "type": "text", + "text": "Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiayi Guo $^{1,2}$ Junhao Zhao $^{1}$ Chaoqun Du $^{1}$ Yulin Wang $^{1}$ Chunjiang Ge $^{1}$ Zanlin Ni $^{1}$ Shiji Song $^{1}$ Humphrey Shi $^{2*}$ Gao Huang $^{1*}$ $^{1}$ Tsinghua University $^{2}$ SHI Labs @ Georgia Tech", + "bbox": [ + 153, + 200, + 841, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db4bdd40794b379802d7fddfa5d56b7684f4c15466580f40ff52b15f7e6ac36c.jpg", + "image_caption": [ + "Figure 1. Comparison of different test-time adaptation (TTA) frameworks. (a) Traditional TTA methods continuously adapt source model weights to fit target data batches. However, their performance is sensitive to the amount and order of target data streams, e.g., adapting the model with batches containing data from only a single category can lead to overfitting. (b) Diffusion-driven TTA methods project the target data back to the synthetic domain of diffusion models, which still maintains domain misalignment with the source domain. (c) We propose the Synthetic-domain Alignment (SDA) framework for TTA, which simultaneously aligns the domains of the source model and target data with the same synthetic domain for superior performance." + ], + "image_footnote": [], + "bbox": [ + 91, + 282, + 486, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f5d78e1469ebdce1b05f3eca09ee6086eab47bf020e7be00204b6cfa1b5bf98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 282, + 906, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 564, + 326, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test-time adaptation (TTA) aims to improve the performance of source-domain pre-trained models on previously unseen, shifted target domains. Traditional TTA methods primarily adapt model weights based on target data streams, making model performance sensitive to the amount and order of target data. The recently proposed diffusion-driven TTA methods mitigate this by adapting model inputs instead of weights, where an unconditional diffusion model, trained on the source domain, transforms target-domain data into a synthetic domain that is expected to approximate the source domain. However, in this paper, we reveal that although the synthetic data in diffusion-driven TTA seems indistinguishable from the source data, it is unaligned with, or even markedly different from the latter for deep networks. To address this issue, we propose a Synthetic-Domain Alignment (SDA) framework. Our key insight is to fine-tune the source model with synthetic data to ensure better alignment. Specifically, we first employ a conditional diffusion model to generate labeled", + "bbox": [ + 88, + 590, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "samples, creating a synthetic dataset. Subsequently, we use the aforementioned unconditional diffusion model to add noise to and denoise each sample before fine-tuning. This Mix of Diffusion (MoD) process mitigates the potential domain misalignment between the conditional and unconditional models. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLMs, e.g., LLaVA) demonstrate that SDA achieves superior domain alignment and consistently outperforms existing diffusion-driven TTA methods. Our code is available at https://github.com/SHI-Labs/Diffusion-Driven-Test-Time-Adaptation-via-Synthetic-Domain-Alignment.", + "bbox": [ + 511, + 566, + 908, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test-Time Adaptation (TTA) [15, 17, 31, 46, 61, 65, 66, 68, 71] is an emerging research field that tackles domain misalignment when source models are evaluated on shifted target data. Unlike traditional domain adaptation (DA) [16, 37, 57] and source-free adaptation (SFA) [28, 30, 33], TTA addresses more practical scenarios where neither source data nor complete target data are accessible. Instead, the adapta", + "bbox": [ + 511, + 796, + 908, + 905 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 107, + 888, + 241, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30503", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ac3c784ca0d59420a48b11792a781b7fc2e80d4ee86b7bc46eb41f7ea74b4b57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model Adaptation DirectionData Adaptation Direction
Traditional TTAExpected: Source→Target Risk: Imbalanced Target StreamsN/A
Diffusion-driven TTAN/AExpected: Target→Source Actual: Target→Synthetic
SDA (Ours)Source→SyntheticTarget→Synthetic
", + "bbox": [ + 91, + 88, + 483, + 170 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Adaptation directions of different TTA methods.", + "bbox": [ + 106, + 176, + 465, + 190 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion relies solely on streaming batches of target data.", + "bbox": [ + 89, + 199, + 439, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional TTA methods (Fig. 1a) [15, 31, 50, 56, 61, 66, 68, 71] typically employ a source-to-target model adaptation framework. These approaches continuously update the source model weights by processing target data batches. Without annotations, the adaptation process relies either on batch-wise updates of model statistics [31, 56, 66, 68], or unsupervised or self-supervised auxiliary tasks [15, 50, 61]. However, small or imbalanced batches may poorly represent the target domain, making these approaches sensitive to the amount and order of the data stream [12, 17, 66]. For instance, adapting the model with batches containing data from only a single category can lead to overfitting.", + "bbox": [ + 89, + 215, + 483, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the impressive generation capabilities of diffusion models [22, 25, 29, 48, 51, 53, 67] have sparked the development of diffusion-driven TTA methods (Fig. 1b) [17, 46, 65], leveraging a target-to-source framework. These approaches employ an unconditional diffusion model, pretrained on the source domain, aiming to project each target sample to the source domain independently. This enables the source model to make predictions without modifying its weights. As a preliminary work, DiffPure [46] addresses adversarial perturbations by first applying a forward diffusion process, introducing a small amount of noise to the target data, followed by a reverse diffusion process to restore a clean image to approach the source domain. Building on this concept, more recent studies [17, 65] tackle challenging domain shifts—such as severe data corruption—by incorporating additional structural guidance from the target data, helping preserve semantics and improve performance.", + "bbox": [ + 89, + 398, + 483, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we uncover that while diffusion-driven TTA methods aim to project target data back to the source domain, the projected target data remains confined within the synthetic domain of the unconditional diffusion model. As the synthetic-domain data are ultimately processed by the source-domain model, this domain misalignment limits the final performance. To address this issue, we propose Synthetic-Domain Alignment (SDA) (Fig. 1c), a new category of framework for TTA tasks which simultaneously aligns the domains of the source model and target data with the same synthetic domain of a diffusion model (Tab. 1).", + "bbox": [ + 89, + 657, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SDA distinguishes itself from existing diffusion-driven TTA methods [17, 46, 65] by introducing an additional source-to-synthetic model adaptation phase before testing on adapted target data (Fig. 2). Since the adapted target data aligns with the synthetic domain generated by an un", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae6a5cb9f9a020df47ff7598fa7d0efb0fecd225b2904f860a72db1472d2097d.jpg", + "image_caption": [ + "Figure 2. Enhanced domain alignment with our framework. Prior diffusion-driven TTA methods struggle with the domain misalignment between the source model and synthetic data, which we resolve by aligning the source model to the synthetic domain." + ], + "image_footnote": [], + "bbox": [ + 555, + 93, + 867, + 202 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conditional diffusion model, SDA aims to adapt the source model to this same synthetic domain. Specifically, SDA employs a Mix-of-Diffusion (MoD) technique to generate synthetic data for model adaptation. Given that the source data is inaccessible after pretraining, MoD first uses a conditional diffusion model to generate samples conditioned on domain-agnostic labels, creating a labeled synthetic dataset. Subsequently, the aforementioned unconditional diffusion model is leveraged by MoD add noise to and denoise these samples, addressing potential domain misalignment between the conditional and unconditional models. With a sufficiently large synthetic dataset, the fine-tuned model becomes highly effective at discriminating within the synthetic domain. Thus, the SDA framework transforms the cross-domain TTA task into an in-domain prediction task by aligning both the source model and target data with the same synthetic domain.", + "bbox": [ + 511, + 272, + 906, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SDA is a general framework, not limited to specific fine-tuning techniques or diffusion-driven data adaptation methods. This flexibility allows future advancements in these areas to further broaden its applicability. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLM, e.g., LLaVA) demonstrate that SDA consistently outperforms existing methods. Moreover, the effectiveness of our approach is reinforced through visualization analysis and ablation studies.", + "bbox": [ + 511, + 529, + 905, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 679, + 653, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time adaptation (TTA) is an emerging research area that addresses domain shifts by adapting either models [15, 31, 50, 56, 61, 66, 68, 71] or data [17, 46, 65] during evaluation on streaming target data batches. Early model adaptation methods update batch normalization statistics to match the target distribution [31, 56, 66, 71], while others leverage self-supervised tasks like rotation prediction [61] or image restoration [15, 50] to adjust model weights. However, these approaches rely heavily on continuous weight updates, making them sensitive to the amount, order, and diversity of target data. In contrast, diffusion-driven TTA methods [17, 46, 65] focus on data adaptation by projecting each target sample back into the source domain, achiev-", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30504", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f76f39f78b6dd9a714f6089a1acaac65d72590d4ea1936ffbb7ede0673e3b686.jpg", + "image_caption": [ + "Forward Diffusion (Add Noise t steps)", + "(a) Diffusion-driven data adaptation on source data" + ], + "image_footnote": [], + "bbox": [ + 104, + 88, + 176, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b411876b0eb926ae34ce1345cf1b496951c949f701391aaa467d6f2baa8f1b4f.jpg", + "image_caption": [ + "Reverse Diffusion (Denoise $t$ steps)" + ], + "image_footnote": [], + "bbox": [ + 251, + 88, + 321, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bf422646f3c3df0e2609b0f42bff7dc77d6c9ffd5d040179a2440eed609736c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 88, + 465, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9658ab4994ab010ae6a1feefd933a18dddb4d8263b4a3d554cc2580a18d47772.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 157, + 156, + 208 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6495ec9619f09d6e98a25298f008adf15adac7746c48e5972190c81c4d2eb86e.jpg", + "image_caption": [ + "$t = 600$" + ], + "image_footnote": [], + "bbox": [ + 158, + 157, + 199, + 207 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d135f7544f4ebca577ea4659c9ce5cd22e29611454d05ccf22891acc59179fcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 199, + 167, + 258, + 207 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa31cbe96d9044f80a73260319369fb2a08f96325220bf6571a1f1d93161edf4.jpg", + "image_caption": [ + "$t = 700$" + ], + "image_footnote": [], + "bbox": [ + 259, + 166, + 310, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3260e20ba09c92ec51441b3f25844ece57373735b7959a8ad4fbaeeb31474e1a.jpg", + "image_caption": [ + "$t = 800$", + "300 $t =$" + ], + "image_footnote": [], + "bbox": [ + 312, + 166, + 341, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/274cd19e60f21e2719277f5c30eed42ad7b08e2426279831547a19767f246940.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 166, + 393, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44cf3a6c80a96a086b40e25e25a62c1243aa91371f3421f53fdba6820940dc47.jpg", + "image_caption": [ + "00 $t =$" + ], + "image_footnote": [], + "bbox": [ + 395, + 166, + 436, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db2b1066657c2e9216f5b9fb14f32d83c52200a872653ac216afeb42d398bee4.jpg", + "image_caption": [ + "1000" + ], + "image_footnote": [], + "bbox": [ + 437, + 166, + 464, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/686858674df97e6cd33cd5264ade44d2bb3ad10798738558a56c35b5c032b0c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 209, + 156, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2dd879d31febf613cff71516778612c563b7a6ea75e51841c48b141d8e4043be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 209, + 256, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/849e9a314b2c444880088642bac7fd4ec50f8c30857b10afef7fef689f3feed1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 209, + 310, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e1f7349857c5cf5a9af7d546584782aa6ad105ed33ad38105942bf3c5a1f9cfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 209, + 362, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/362c0480fe2d9bd188d37e569a6ef77d7c47da79a7568c35f7b67ed812d9e0da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 209, + 405, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca30e1e485a0746f17f2c40b784d92643a06585bff81bb0e73a81fa665f420e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 209, + 462, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5e70fe42eed1b45b38083a2844b1abd2687e8daa9b83aeb3fe00ba98c1e8f255.jpg", + "image_caption": [ + "(b) Diffusion-driven data adaptation results (on source data) under different timesteps", + "Figure 3. (a) Illustration of diffusion-driven data adaptation on source data and (b) Adapted images across different timesteps. The results are obtained using DDA [17], with no noticeable visual degradation observed in the adapted images." + ], + "image_footnote": [], + "bbox": [ + 106, + 251, + 156, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a7b54d4cea9c478a4527c308180055254e4e9bd67412bbb4316b695ea651b1f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 251, + 210, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d696fe311078fd824df85682bdf04ecef309c7ad6190305ea057ec036dade8fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 251, + 258, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75ee21dd2d895a9ceffbc52e0f6d257916b6d043208ebfca8b13e9b266e32e94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 251, + 310, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b69451d7c92deb860cd419f12257cda74cbaf1956ebac352679bc18bbc9465f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 251, + 341, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/49651a04855085445a68e6b0df6b8bc9eb9a9726f9b050ef8a24e95a137ea152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 251, + 403, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e0b67c0d8ae1f9432ecf1f69a5cfcfac0b7face8dbea36153c01e8f470e897fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 251, + 439, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing stable performance without online model updates. DiffPure [46] purifies adversarial samples with diffusion models, while DDA [17] and GDA [65] use structural guidance to preserve image content under severe corruption. Building on this, our work investigates the issue of misalignment between domains of the source model and synthetic images in diffusion-driven TTA and proposes a new TTA framework.", + "bbox": [ + 89, + 382, + 483, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic data for discriminative tasks. Synthetic data, generated by models [20, 21, 42-44, 48, 52, 53] rather than collected from the real world, has shown significant potential in enhancing visual representations for various discriminative tasks [14, 62, 63]. It has been effectively applied in areas such as recognition [3, 63], object detection [49, 55], semantic segmentation [7, 54, 58], image assessment [19], autonomous driving [1], and robotics [39, 69]. In this work, we explore the potential of diffusion-generated synthetic data to enhance domain alignment in TTA.", + "bbox": [ + 89, + 489, + 483, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 652, + 225, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the background of diffusion-driven TTA methods and identify their source-synthetic domain misalignment issue in Sec. 3.1 and Sec. 3.2. To tackle this issue, we propose the synthetic-domain alignment (SDA) framework in Sec. 3.3 and introduce its key technique, the mix of diffusion (MoD) in Sec. 3.4.", + "bbox": [ + 89, + 676, + 483, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 89, + 776, + 220, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion process. Given a source data point $\\boldsymbol{x}_0^{\\mathrm{src}} \\sim p_0^{\\mathrm{src}}$ , diffusion models $\\epsilon_{\\theta}$ [25] gradually transform $p_0^{\\mathrm{src}}$ into a Gaussian noise distribution $N(\\mathbf{0}, \\mathbf{I})$ through a $T$ -step forward diffusion process. At each timestep $t \\in \\{1, 2, \\dots, T\\}$ , the intermediate state $\\boldsymbol{x}_t^{\\mathrm{src}} \\sim p_t^{\\mathrm{src}}$ is computed as:", + "bbox": [ + 89, + 797, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}} = \\sqrt {1 - \\beta_ {t}} \\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}} + \\sqrt {\\beta_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d6378c10976ab96ddb4e3a90c54eff733e065257e623f02d10688f87a2eebada.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Timestep t5006007008009001000
Swin-B (Source-Synthetic, Misaligned)61.657.855.752.645.038.5
Swin-B (Synthetic-Synthetic, Aligned)67.665.965.061.555.748.4
Δ+6.0+8.1+9.3+8.9+10.7+9.9
ConvNeXt-B (Source-Synthetic, Misaligned)65.161.360.357.250.041.5
ConvNeXt-B (Synthetic-Synthetic, Aligned)70.668.267.464.958.550.7
Δ+5.5+6.9+7.1+7.7+8.5+9.2
", + "bbox": [ + 514, + 77, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Source model accuracy across different timesteps of diffusion-driven data adaptation. For suitable timesteps for TTA $(t\\geq 500)$ [17], model accuracy shows a monotonically decreasing trend with the growth of timestep, indicating the increase of the misalignment of source domain $p_0^{\\mathrm{src}}$ and synthetic domain $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ . By aligning the source model to the synthetic domain, our methods (rows 3 & 6) significantly help alleviate the performance degradation. Results are evaluated on the ImageNet [9] validation set.", + "bbox": [ + 511, + 185, + 906, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{t}$ is random Gaussian noise and $\\beta_{t} \\in (0,1)$ represents the diffusion rate at step $t$ . By defining $\\alpha_{t} = 1 - \\beta_{t}$ , $\\overline{\\alpha_{t}} = \\prod_{t=1}^{T} \\alpha_{t}$ and $\\epsilon \\sim N(\\mathbf{0}, \\mathbf{I})$ , we obtain:", + "bbox": [ + 511, + 316, + 903, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} ^ {\\mathrm {s r c}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 371, + 903, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reverse diffusion process recovers a clean $\\widehat{\\boldsymbol{x}_0^{\\mathrm{src}}}$ by progressively removing noise from $\\boldsymbol{x}_T^{\\mathrm{src}}$ :", + "bbox": [ + 511, + 398, + 903, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\widehat {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(\\widehat {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 436, + 903, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma_{t}$ is the posterior noise variance [25].", + "bbox": [ + 511, + 477, + 812, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-driven data adaptation. Denote by $p_0^{\\mathrm{trg}}$ the target data distribution, from which each target data point $\\boldsymbol{x}_t^{\\mathrm{trg}}$ is sampled. Prior work [46] demonstrates that:", + "bbox": [ + 511, + 494, + 903, + 540 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\mathrm {K L}} \\left(p _ {t + 1} ^ {\\mathrm {s r c}} \\mid p _ {t + 1} ^ {\\mathrm {t r g}}\\right) - D _ {\\mathrm {K L}} \\left(p _ {t} ^ {\\mathrm {s r c}} \\mid p _ {t} ^ {\\mathrm {t r g}}\\right) \\leq 0, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 549, + 903, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{\\mathrm{KL}}$ is the KL divergence. Since $p_T^{\\mathrm{src}} = p_T^{\\mathrm{trg}} = N(\\mathbf{0}, \\mathbf{I})$ , for any arbitrarily small value $\\delta$ , there exists a minimum timestep $t^*$ such that $D_{\\mathrm{KL}}(p_{t^*}^{\\mathrm{src}}||p_{t^*}^{\\mathrm{trg}}) < \\delta$ . As discussed in [17], a diffusion process with $T = 1000$ requires $t^* \\geq 500$ to eliminate domain shifts. We empirically validate this choice of $t^*$ in the supplementary materials.", + "bbox": [ + 511, + 575, + 905, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initial diffusion-driven data adaptation in Diff-Pure [46] is carried out as follows: Given the minor divergence between $p_{t^*}^{\\mathrm{src}}$ and $p_{t^*}^{\\mathrm{trg}}$ , we recover each $\\boldsymbol{x}_{t^*}^{\\mathrm{trg}} \\sim p_{t^*}^{\\mathrm{trg}}$ to its corresponding $\\widehat{\\boldsymbol{x}_0^{\\mathrm{src}}} \\sim p_0^{\\mathrm{src}}$ by executing the reverse diffusion process (Eq. 3) $t^*$ times.", + "bbox": [ + 511, + 667, + 905, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the reverse process is stochastic, subsequent works [17, 65] further introduce additional structure guidance to ensure the content consistency between each $\\boldsymbol{x}_0^{\\mathrm{trg}}$ and its adapted $\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}$ . In this work, we adopt the same data adaptation process as DDA [17]:", + "bbox": [ + 511, + 744, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} = \\widehat {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} - \\boldsymbol {w} \\nabla_ {\\widetilde {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}}} \\left\\| \\phi \\left(\\boldsymbol {x} _ {0} ^ {\\mathrm {t r g}}\\right) - \\phi \\left(\\widetilde {\\boldsymbol {x} _ {0 , t} ^ {\\mathrm {s r c}}}\\right) \\right\\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 830, + 903, + 856 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\widehat{\\pmb{x}_{t - 1}^{\\mathrm{src}}}$ is computed as Eq. 3, $\\pmb{w}$ is the structure guidance scale, $\\phi$ is a structure extractor [17] and $\\widehat{\\pmb{x}_{0,t}^{\\mathrm{src}}}$ is an esti", + "bbox": [ + 511, + 864, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30505", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/80b7c7c4458d7d450c98575703dae29f2eeeb6ad025442a57c2e7eab58b13af3.jpg", + "image_caption": [ + "Figure 4. Overview of the Synthetic-Domain Alignment (SDA) framework. SDA is a novel TTA framework aligning both the domains of the source model and the target data with the synthetic domain. SDA involves three phases: (left): a source-domain model pretraining phase, where the source model is trained on source data prior to TTA; (middle): a source-to-synthetic model adaptation phase, where the source model is adapted to a synthetic-domain model using synthetic data generated via a Mix of Diffusion (MoD) technique; and (right): a target-to-synthetic data adaptation phase, where target data is adapted into synthetic data using an unconditional diffusion model. Finally, the adapted synthetic data is fed into the synthetic-domain model for test-time inference." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 903, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mate of $\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}$ at timestep $t$ [25]. Diffusion-driven data adaptation is typically performed with unconditional diffusion models [10] since the target data labels are unknown.", + "bbox": [ + 88, + 450, + 482, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Source-Synthetic Domain Misalignment", + "text_level": 1, + "bbox": [ + 89, + 508, + 431, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike a real source data point $\\boldsymbol{x}_0^{\\mathrm{src}}$ , the recovered version $\\widetilde{\\boldsymbol{x}_{0}^{\\mathrm{src}}},$ derived from $\\boldsymbol{x}_{t^{*}}^{\\mathrm{trg}}$ is synthetic. Specifically, $\\boldsymbol{x}_0^{\\mathrm{src}}$ follows the source domain $p_0^{\\mathrm{src}}$ , whereas $\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}$ follows the synthetic domain $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ of an unconditional diffusion model $\\epsilon_{\\theta}^{\\mathrm{u}}$ with parameters $\\theta$ . In this section, we empirically reveal the misalignment between the source domain $p_0^{\\mathrm{src}}$ and the synthetic domain $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ , and investigate how this misalignment impacts the performance of the source model. For simplicity and clarity, we will substitute $\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}$ with $\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}$ throughout the following discussion.", + "bbox": [ + 88, + 530, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above analysis, diffusion-driven TTA methods face two potential misalignments: (1) source-target domain misalignment arising from inherent data distribution shifts, which has been the primary focus of prior research, and (2) source-synthetic domain misalignment introduced by diffusion models, which we address as the main focus of this work, complementing existing efforts.", + "bbox": [ + 88, + 688, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To precisely examine the impact of the source-synthetic domain misalignment and isolate it from the influence of the source-target domain misalignment, we evaluate the performance of the source model $f$ on synthetic data $x_{0,\\mathrm{u}}^{\\mathrm{syn}}$ adapted by the diffusion model $\\epsilon_{\\theta}^{\\mathrm{u}}$ from source data $x_0^{\\mathrm{src}}$ (Fig. 3a). Specifically, we test ImageNet pretrained models [9] on the ImageNet validation set adapted by the popular diffusion", + "bbox": [ + 88, + 795, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "driven data adaptation method, DDA [17] across different timesteps $t$ . As indicated in Tab. 2, with an increase in $t$ , the accuracy of the model exhibits a monotonically decreasing trend. For ideal $t^* \\geq 500$ for TTA tasks, performance degradation of more than $18.8\\%$ for ConvNeXt [36] and $21.8\\%$ for Swin [35] is observed compared to their official results [8] on source data which are $83.4\\%$ and $83.9\\%$ . This indicates a significant domain misalignment between $p_0^{\\mathrm{src}}$ and $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ . By fine-tuning the source model on the synthetic data generated by our Mix of Diffusion process (introduced in Sec. 3.4), the aligned models $f'$ achieve performance improvements of $5.5\\%$ for ConvNeXt and $6.0\\%$ for Swin.", + "bbox": [ + 511, + 450, + 906, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 3b, we show that the diffusion synthetic data $\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}$ and source data $\\boldsymbol{x}_0^{\\mathrm{src}}$ exhibit no noticeable visual differences across different timesteps $t$ . This further suggests that the performance degradation is not due to the quality of diffusion-generated images but rather the implicit misalignment between the source and synthetic domains.", + "bbox": [ + 511, + 633, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Synthetic-Domain Alignment Framework", + "text_level": 1, + "bbox": [ + 511, + 739, + 867, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that the diffusion-adapted data $\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}$ aligns more closely with the synthetic domain $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ instead of the source domain $p_0^{\\mathrm{src}}$ , we propose simultaneously adapting the source model $f$ to the same synthetic domain $p_0^{\\mathrm{syn}}$ . By doing so, the alignment between the data and model within $p_0^{\\mathrm{syn}}$ can be effectively achieved. To this end, we introduce a novel TTA framework: Synthetic-Domain Alignment (SDA).", + "bbox": [ + 511, + 762, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 4, we present the complete diagram of SDA, which consists of three key phases: (1) a source-domain", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30506", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model pretraining phase, where the source model $f$ is trained on source data $\\pmb{x}_0^{\\mathrm{src}}$ prior to TTA; (2) a source-to-synthetic model adaptation phase, where $f$ is fine-tuned to a synthetic-domain model $f'$ ; and (3) a target-to-synthetic data adaptation phase, where target data $\\pmb{x}_0^{\\mathrm{trg}}$ is adapted into synthetic data $\\pmb{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}$ using an unconditional diffusion model $\\epsilon_{\\theta}^{\\mathrm{u}}$ following Eq. 5. Finally, the adapted synthetic data $\\pmb{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}$ is fed into the synthetic-domain model $f'$ for test-time inference. Consistent with the standard TTA protocol [66], the source data is accessible only during the source model pretraining phase and remains inaccessible during both the model and data adaptation phases.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The rationale behind SDA is straightforward: by aligning the model and data domains to the same synthetic domain, the original cross-domain TTA task is transformed into an easier in-domain prediction task, thus addressing the core challenge of TTA and improving performance.", + "bbox": [ + 89, + 272, + 480, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the pretraining phase follows standard supervised learning, and the data adaptation phase aligns with DDA [17], we omit detailed explanations of these phases. Instead, we focus on how we adapt the source model $f$ to an ideal synthetic domain model $f'$ using a novel technique called Mix-of-Diffusion (MoD), introduced next.", + "bbox": [ + 89, + 348, + 480, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Adaptation via Mix of Diffusion", + "text_level": 1, + "bbox": [ + 89, + 448, + 424, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4, MoD consists of two main processes: a data generation process powered by a conditional diffusion model $\\epsilon_{\\eta}^{\\mathrm{c}}$ with parameters $\\eta$ , and a data alignment process powered by the same unconditional diffusion model $\\epsilon_{\\theta}^{\\mathrm{u}}$ used in the target-to-synthetic data adaptation phase. It is worth noting that, both $\\epsilon_{\\eta}^{\\mathrm{c}}$ and $\\epsilon_{\\theta}^{\\mathrm{u}}$ are also pretrained on source data $\\mathbf{x}_0^{\\mathrm{src}}$ and have never been exposed to target data $\\mathbf{x}_0^{\\mathrm{trg}}$ , in accordance with the TTA setting.", + "bbox": [ + 89, + 470, + 480, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditional diffusion data generation. The generation process leverages the conditional generation capability of $\\epsilon_{\\eta}^{\\mathrm{c}}$ to synthesize conditional synthetic data $x_{0,c}^{\\mathrm{syn}} \\sim p_{0,c}^{\\mathrm{syn}}$ . In the context of TTA, the source and target domains share the same domain-agnostic label set $\\{y_i\\}_{i=1}^{K}$ . Utilizing $\\epsilon_{\\eta}^{\\mathrm{c}}$ , we uniformly generate samples for each class $y_i$ from Gaussian noise $(x_{T,c}^{\\mathrm{syn}})$ through a $T$ -step reverse diffusion process:", + "bbox": [ + 89, + 597, + 480, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t - 1, c} ^ {\\text {s y n}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t, c} ^ {\\text {s y n}} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\overline {{\\alpha_ {t}}}}} \\epsilon_ {\\eta} ^ {\\mathrm {c}} \\left(\\boldsymbol {x} _ {t, c} ^ {\\text {s y n}}, t, y _ {i}\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\epsilon}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 715, + 480, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generation capability of $\\epsilon_{\\eta}^{\\mathrm{c}}$ allows for the construction of a labeled synthetic-domain dataset $\\{x_{0,\\mathrm{c}}^{\\mathrm{syn}},y\\}^{N}$ of arbitrary size $N$ without any manual data collection. By fine-tuning the source model $f$ on this synthetic dataset, an adapted model $f_{\\mathrm{c}}^{\\prime}$ on domain $p_{0,\\mathrm{c}}^{\\mathrm{syn}}$ can be obtained.", + "bbox": [ + 89, + 758, + 480, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unconditional diffusion data alignment. However, since the test-time adapted data $x_{0,\\mathrm{u}}^{\\mathrm{syn}}$ is subject to domain $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ , we argue that there is still potential misalignment between different synthetic domains $p_{0,\\mathrm{c}}^{\\mathrm{syn}}$ and $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ . This is mainly", + "bbox": [ + 89, + 839, + 480, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "because of the differences in architectures and training schemes between $\\epsilon_{\\eta}^{\\mathrm{c}}$ and $\\epsilon_{\\theta}^{\\mathrm{u}}$ . We empirically validate this argument in Tab. 10, which indicates that the model $f_{\\mathrm{c}}'$ adapted to domain $p_{0,\\mathrm{c}}^{\\mathrm{syn}}$ performs worse than the model $f_{\\mathrm{u}}'$ adapted to domain $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ during testing.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the $f_{\\mathrm{u}}^{\\prime}$ on domain $p_{0,\\mathrm{u}}^{\\mathrm{syn}}$ we mirror the target-to-synthetic data adaptation phase as a conditional synthetic data $(\\boldsymbol{x}_{0,\\mathrm{c}}^{\\mathrm{syn}})$ to unconditional synthetic data $(\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}})$ adaptation process. In specific, we use the same $t^{*}$ in Sec. 3.1 with Eq. 2 to obtain $\\boldsymbol{x}_{t^{*},\\mathrm{c}}^{\\mathrm{syn}}$ . According to the analysis in Sec. 3.1, $\\boldsymbol{x}_{t^{*},\\mathrm{c}}^{\\mathrm{syn}}$ will be indistinguishable to its counterpart $\\boldsymbol{x}_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}$ on domain $p_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}$ . Therefore, we have:", + "bbox": [ + 511, + 167, + 903, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t ^ {*}, \\mathrm {u}} ^ {\\text {s y n}} \\approx \\boldsymbol {x} _ {t ^ {*}, \\mathrm {c}} ^ {\\text {s y n}} = \\sqrt {\\overline {{\\alpha_ {t}}}} \\boldsymbol {x} _ {0, c} ^ {\\text {s y n}} + \\sqrt {1 - \\overline {{\\alpha_ {t}}}} \\boldsymbol {\\epsilon}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 281, + 903, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, following Eq. 3 and Eq. 5, the noisy $\\boldsymbol{x}_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}$ can be gradually denoised to $\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}$ . Finally, the expected synthetic-domain model $f_{\\mathrm{u}}^{\\prime}$ is obtained by fine-tuning the source-domain model $f_{\\mathrm{u}}^{\\prime}$ on dataset $\\{\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}},y\\}^{N}$ .", + "bbox": [ + 511, + 306, + 903, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ensembling. As noted in previous diffusion-driven TTA methods [17], while diffusion models generally perform well for data adaptation, they may occasionally produce data points that are less recognizable than the original target data. To address this, prior approaches use an ensemble of model predictions on $\\boldsymbol{x}_{0}^{\\mathrm{trg}}$ and $\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}$ as the final output. Following this protocol, the final prediction in SDA is:", + "bbox": [ + 511, + 371, + 903, + 476 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\arg \\max _ {y} \\left(q \\left(y \\mid \\boldsymbol {x} _ {0} ^ {\\operatorname {t r g}}\\right) + q ^ {\\prime} \\left(y \\mid \\boldsymbol {x} _ {0, \\mathrm {u}} ^ {\\operatorname {s y n}}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 489, + 903, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q(\\cdot)$ and $q^{\\prime}(\\cdot)$ are output distributions of source model $f$ and synthetic-domain model $f_{\\mathrm{u}}^{\\prime}$ , respectively.", + "bbox": [ + 511, + 518, + 903, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 561, + 643, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first evaluate SDA on ImageNet classifiers with standard TTA benchmarks in Sec. 4.1. Next, we assess SDA's scalability across different dataset sizes, tasks, and model architectures in Sec. 4.2. In Sec. 4.3, we demonstrate SDA's advantages through Grad-CAM visualizations [59] and data stream sensitivity tests. Finally, ablation studies in Sec. 4.4 validate the design choices in our SDA framework.", + "bbox": [ + 511, + 585, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2e5e0a8b057d019c8f348c4e165633db0d7c9fc468d53cf249d44c817171d03f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSourceMEMODiffPureGDADDASDA (Ours)
ResNet-5018.724.716.831.829.732.5 (+2.8)
Swin-T33.529.524.842.240.042.5 (+2.5)
ConvNeXt-T39.337.828.844.844.247.0 (+2.8)
Swin-B40.537.028.9-44.547.4 (+2.9)
ConvNeXt-B45.645.832.7-49.451.9 (+2.5)
", + "bbox": [ + 514, + 715, + 903, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Comparison results on ImageNet-C. We compare SDA with source models, MEMO [71], DiffPure [46], GDA [65] and DDA [17]. Results are the average accuracy across 15 adaptation domains at severity level 5. SDA shows consistent performance improvements compared to baselines.", + "bbox": [ + 511, + 820, + 903, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30507", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/48ea1b07c8e1da0a1e26b16c9861b1402cb6a232a8cf9ad26196e0c2175b5736.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GaussianShotImpluseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source48.047.148.332.414.439.339.255.750.151.575.152.029.548.160.346.1
DiffPure42.642.040.916.822.219.123.039.029.19.162.53.034.346.860.732.7
DDA59.955.755.930.730.538.139.253.248.041.971.048.146.959.562.049.4
SDA (Ours)60.357.658.135.435.342.942.455.549.944.672.645.450.263.064.551.9 (+2.5)
", + "bbox": [ + 91, + 84, + 906, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results on ImageNet Classifiers", + "text_level": 1, + "bbox": [ + 89, + 247, + 413, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. We choose DDA [17] as our primary competitor since it is the best open-sourced method. SDA is also compared with DiffPure [46], MEMO [71], and the recent SOTA GDA [65] using their reported results. Source model performance is reported as \"Source\". DiT [48] and ADM [10] are adopted to generate and align 50K synthetic data for 15-epoch finetuning. For different source models, the synthetic data only needs to be generated once. Our results are tested on standard TTA benchmarks, ImageNet-C [24] (severity level 5) and ImageNet-W [32] using various models including ResNet [23], Swin [35] and ConvNeXt [36]. More implementation details are provided in the supplementary materials.", + "bbox": [ + 88, + 273, + 482, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison results on ImageNet-C. We begin by evaluating the performance of SDA on ImageNet-C. As reported in Tab. 3, our proposed SDA consistently outperforms all baseline methods across different model architectures and sizes. We emphasize the performance improvement over DDA, as we adopted DDA for target data adaptation in SDA. Compared to DDA, our SDA improves accuracy by $2.5\\% - 2.9\\%$ . This significant improvement indicates the misalignment between the source and synthetic domains, validating the effectiveness of our synthetic-domain alignment framework. Moreover, compared to the recent SOTA GDA, SDA also achieves an improvement of $2.2\\%$ with ConvNeXt-T. Notably, SDA focuses on synthetic domain alignment, an orthogonal research direction to existing diffusion-driven methods on better adapting the target data. Therefore, the performance of SDA could potentially be further enhanced with the release of more advanced codebases like GDA. Compared to the model adaptation method, MEMO, three diffusion-driven methods (SDA, DDA, and GDA) all demonstrate superior performance, highlighting the effectiveness of diffusion models in assisting TTA tasks. DiffPure presents worse results since it is primarily designed for adversarial attacks. Without the structural guidance introduced in DDA and GDA, DiffPure may not effectively recover images with severe domain shifts. In Tab. 4, we provide a detailed comparison of the results of SDA and baselines. SDA surpasses DiffPure in all 15 adaptation domains and outperforms DDA in 14 out of 15 domains, fur", + "bbox": [ + 91, + 477, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/13f00ce2921a8567b682040f674008f0c36e556e848461a9ac20848c39a2abf0.jpg", + "table_caption": [ + "Table 4. Detailed comparisons of SDA and baselines across 15 adaptation domains of ImageNet-C. SDA shows the best average accuracy. The results are tested with ConvNeXt-B. Comparisons using other models are deferred to the supplementary materials." + ], + "table_footnote": [], + "table_body": "
ModelSourceDiffPureDDASDA (Ours)
ResNet-5037.729.152.854.7 (+1.9)
Swin-T66.552.765.967.3 (+1.4)
ConvNeXt-T67.655.867.969.4 (+1.5)
Swin-B69.155.568.370.6 (+2.3)
ConvNeXt-B70.157.770.372.3 (+2.0)
", + "bbox": [ + 514, + 244, + 906, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Comparison results on ImageNet-W. SDA shows consistent performance improvements compared to baselines.", + "bbox": [ + 511, + 359, + 906, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ther affirming the superiority of SDA.", + "bbox": [ + 511, + 405, + 764, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison results on ImageNet-W. We extend our evaluation to ImageNet-W to assess SDA's performance under watermark-based domain shifts. As shown in Tab. 5, SDA consistently surpasses all baselines across different models. Compared to our primary baseline, DDA, SDA achieves accuracy gains ranging from $1.4\\%$ to $2.3\\%$ . Furthermore, the results in Tab. 5 reveal potential performance drops when DDA is applied to ImageNet-W with Swin-T and Swin-B models, suggesting that synthetic data may be less recognizable by the original source models. Although advancements in diffusion techniques could potentially improve outcomes, the consistent gains achieved by SDA indicate that aligning the source model with the synthetic domain offers a convenient and effective solution to enhance performance.", + "bbox": [ + 511, + 424, + 906, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Scalability to Other Datasets, Tasks and Models", + "text_level": 1, + "bbox": [ + 511, + 647, + 903, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to standard benchmarks, an effective TTA method should demonstrate its superiority across various dataset scales, task formats, and model architectures. In this section, we assess SDA from these aspects to validate its scalability. Implementation details of each experiment can be found in the supplementary materials.", + "bbox": [ + 511, + 669, + 905, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scaling to small datasets. We first evaluate the effectiveness of SDA on scenarios where both source and target domains are small-scale. Specifically, we test SDA on CIFAR-10-C [24] with ResNet-18 [23] as the source classifier. EDM [27] is used to generate synthetic data, and I-DDPM [45] is applied for data alignment. As shown in Tab. 6, SDA consistently outperforms both the source model and DDA, achieving an average accuracy improvement of $7.1\\%$ over DDA.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30508", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9dd9c63be1b3ccac0094480d4e651edf5f29c240df937f62552a137ba0667b71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source23.528.720.841.230.748.248.151.863.663.479.528.056.535.960.745.4
DDA79.480.376.841.356.649.048.977.473.562.680.561.561.257.673.265.3
SDA (Ours)82.683.579.559.661.465.364.676.578.471.588.953.475.464.480.772.4 (+7.1)
", + "bbox": [ + 93, + 89, + 903, + 183 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/32badcc958976904a05af786839795a85ea1827db2214a82aaa50ddc91347d8d.jpg", + "table_caption": [ + "Table 6. Comparisons of SDA and baselines with ResNet-18 on CIFAR-10-C. SDA shows the best average accuracy." + ], + "table_footnote": [], + "table_body": "
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPirelateJEPGAvg
Source9.813.210.022.75.720.422.136.728.453.966.69.724.316.043.825.6
DDA16.621.116.045.040.949.738.736.738.635.055.213.356.656.358.838.6
SDA (Ours)17.021.116.245.543.051.139.938.740.236.357.913.159.158.159.739.8 (+1.2)
", + "bbox": [ + 91, + 208, + 903, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Comparisons of SDA and baselines with DeepLabv3 on PASCAL VOC-C. SDA shows the best average mIOU.", + "bbox": [ + 130, + 306, + 862, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling to Semantic Segmentation Tasks. We extend our evaluation to dense prediction tasks by using PASCAL-VOC-C [13] as a standard semantic segmentation benchmark with DeepLabv3 [6] as the source segmenter. Dataset Diffusion [41] generates synthetic segmentation data, and FLUX Schnell [29] is used for data alignment. As shown in Tab. 7, SDA achieves the best performance, with an average mIOU improvement of $1.2\\%$ over DDA.", + "bbox": [ + 88, + 337, + 482, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling to multimodal large language models (MLLMs). As an emerging research direction, MLLMs [26, 34] present advanced visual question answering [2] capability. We design a language-based classification task format to test how SDA can help MLLMs on TTA tasks on ImageNet-C: Given an image, ask an MLLM (LLaVA 1.5-7b [34] in our experiments) to choose the correct image class from four provided options. In Tab. 8, the \"Source-Zero\" setting tests the zero-shot results of pretrained LLaVA. The \"Source\" and \"DDA\" settings evaluate a source-data fine-tuned LLaVA while the \"SDA\" setting tests a synthetic-data fine-tuned LLaVA. The fine-tuning task format is the same as that at test time. Pre-trained LLaVA already exhibits strong performance. While DDA improves results in some domains, it does not yield an overall gain compared to source-data fine-tuned LLaVA. In contrast, SDA aligns LLaVA with the synthetic domain, achieving the best accuracy with an improvement of $2.4\\%$ .", + "bbox": [ + 89, + 463, + 482, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 89, + 737, + 192, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization. To demonstrate how synthetic data fine-tuning in SDA enhances the performance of diffusion-driven TTA methods, we employ Gradient-weighted Class Activation Mapping (Grad-CAM) [59] to visualize the image regions that most influence classification scores across different images and models. As shown in Fig. 5, testing target images with the source model reveals distinct differences in activation maps and the occurrence of incorrect predictions compared to those from source images,", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d9145f445f0dc1c5dc179cde4f7eac7b94f6c5142bae6001f4f5fa5d6454627.jpg", + "image_caption": [ + "Figure 5. Grad-CAM visualization comparison. The first row shows activation maps for source and target images tested with the source model. The second row displays activation maps for diffusion synthetic images tested with the source model (DDA) and our synthetic-domain model (SDA). SDA aligns closely with the source model's response to source images." + ], + "image_footnote": [], + "bbox": [ + 516, + 338, + 900, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "underscoring the performance degradation due to domain shifts. Despite using adapted synthetic images, DDA still risks focusing on inappropriate regions and producing incorrect predictions. This highlights the domain misalignment of the synthetic data and source model. In contrast, SDA aligns both the data and model within the same synthetic domain, thereby producing activation maps and predictions that closely resemble those produced by the source model on source images.", + "bbox": [ + 511, + 564, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data stream sensitivity. In Tab. 9, we test data stream sensitivity using the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings closely aligned with diverse real-world TTA scenarios. In addition to diffusion-driven TTA methods, we include 13 popular traditional TTA methods [4, 5, 11, 12, 18, 38, 40, 47, 60, 66, 68, 70] for comparison. We report the average accuracy of 12 settings using ResNet-50. The results indicate that diffusion-driven TTA methods are preferred in these challenging settings since they are insensitive to different variants of data streams. SDA maintains this insensitivity and showcases the best performance. Detailed comparisons for each setting are provided in the supplementary materials.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30509", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/97a0f48091061b06980b81129d5f9eca1daa7ef5c45f5395e294f4aef4ed6333.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GaussianShotImpluseDerocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAve:
Source-Zero75.276.679.681.474.583.581.187.088.989.595.684.176.290.188.783.5
Source86.886.088.690.185.593.589.793.695.094.998.189.585.696.294.891.2
DDA92.491.592.588.690.192.288.593.093.189.697.875.889.696.497.091.2
SDA (Ours)94.994.194.691.092.994.091.593.695.292.298.583.692.897.397.693.6 (+2.4)
", + "bbox": [ + 91, + 88, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/368f9c354572416d7aad32827e240c07561a9e52b886d8eafbcaa84c714e8be2.jpg", + "table_caption": [ + "Table 8. Comparisons of SDA and baselines with LLaVA on ImageNet-C. SDA shows the best average accuracy." + ], + "table_footnote": [], + "table_body": "
SourceTENT [66]ROID [38]NOTE [18]CoTIA [68]TRIBE [60]BN [40]EATA [67]SANTA [5]RMT [11]UnMIX [64]RoTTA [70]LAME [14]UniTTA [12]DDA [17]SDA (Ours)
18.34.26.26.38.89.39.810.210.611.517.423.124.930.329.331.6 (+2.3)
", + "bbox": [ + 91, + 224, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a6baf7f944ce4e6048b64e70c9e58a1f92cee78fe861029a74a267c584b005af.jpg", + "table_caption": [ + "Table 9. Data stream sensitivity comparison. We additionally compare SDA with 13 traditional TTA methods on the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings. Results are reported as average accuracy across settings." + ], + "table_footnote": [], + "table_body": "
ComponentResNet-50Swin-TConvNeXt-T
DDA [17]29.740.044.2
+ Conditional Data Generation30.440.144.6
+ Unconditional Data Alignment32.542.547.0
", + "bbox": [ + 91, + 335, + 483, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 449, + 254, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Components. We examine the impact of two key components in our SDA framework's Mix of Diffusion technique in Tab. 10: (1) synthetic data generation using the conditional diffusion model and (2) synthetic data alignment using the unconditional diffusion model. Fine-tuning source models with synthetic data generated solely by the conditional diffusion model (+ Conditional Data Generation) yields only marginal gains. As discussed in Sec. 3.4, this limited gain arises from the domain misalignment between conditional and unconditional diffusion models. Specifically, using only conditional synthetic data results in models aligned with the conditional diffusion domain, whereas the test data belongs to the unconditional diffusion domain. Therefore, further aligning the synthetic data through the unconditional diffusion model (+ Unconditional Data Alignment) leads to significant performance gains over the baseline DDA. This demonstrates that bridging the misalignment of different diffusion domains is essential for the success of our SDA framework.", + "bbox": [ + 88, + 472, + 482, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of fine-tuning images. We examine the impact of different numbers of images $(N)$ used during synthetic-domain model fine-tuning in Tab. 11. Interestingly, even with only one image per class $(N = 1\\mathrm{K})$ , SDA still significantly outperforms DDA. This finding suggests a key attribute of the fine-tuning process: source models are primarily learning to adapt to the synthetic domain itself, rather than acquiring class-specific knowledge. Increasing the number of images helps the fine-tuning process capture the", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/38ab6e74e54ba495166171840608c00b4a8a14526a1810699a2d382709e0b877.jpg", + "table_caption": [ + "Table 10. Impact of different components in SDA. Results are evaluated on ImageNet-C." + ], + "table_footnote": [], + "table_body": "
MethodNumber (N)ResNet-50Swin-TConvNeXt-T
DDA [17]029.740.044.2
SDA1K31.942.345.5
10K31.942.546.6
50K (default)32.542.547.0
100K32.642.246.8
", + "bbox": [ + 532, + 335, + 887, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11. Impact of different numbers of fine-tuning images. Results are evaluated on ImageNet-C.", + "bbox": [ + 511, + 438, + 903, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "synthetic domain more accurately, thereby enhancing performance. Based on a balance between performance improvement and image generation resources, we select $N = 50\\mathrm{K}$ as our default experimental setting.", + "bbox": [ + 511, + 491, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 571, + 633, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed Synthetic-Domain Alignment (SDA), a novel test-time adaptation (TTA) framework that simultaneously aligns the domains of the source model and target data with the synthetic domain of a diffusion model. For the source model, SDA introduces a Mix of Diffusion (MoD) technique, which generates synthetic data to adapt the source model to a synthetic-domain model. MoD involves a conditional diffusion model for data generation and an unconditional diffusion model for data alignment. For the target data, SDA utilizes the aforementioned unconditional diffusion model to project the target data to synthetic data. As the domains of the model and data are aligned, SDA converts the cross-domain TTA task into an easier in-domain prediction task. Compared to existing diffusion-driven TTA methods, SDA significantly mitigates the source-synthetic domain misalignment issue. Compared to traditional TTA methods, SDA maintains insensitivity to different data streams. Extensive experiments across classifiers, segmenters, and MLLMs indicate that SDA achieves enhanced domain alignment and superior performance.", + "bbox": [ + 509, + 598, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30510", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant U24B20173.", + "bbox": [ + 89, + 114, + 485, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 189, + 187, + 205 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets computer vision: Efficient data generation for urban driving scenes. IJCV, 2018. 3", + "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 7", + "[3] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification. TMLR, 2023. 3", + "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, 2022. 7, 8", + "[5] Goirik Chakrabarty, Manogna Sreenivas, and Soma Biswas. Santa: Source anchoring network and target alignment for continual test time adaptation. TMLR, 2023. 7, 8", + "[6] Liang-Chieh Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7", + "[7] Yuhua Chen, Wen Li, Xiaoran Chen, and Luc Van Gool. Learning semantic segmentation from synthetic data: A geometrically guided input-output adaptation approach. In CVPR, 2019. 3", + "[8] MMPreTrain Contributors. Openmmlab's pre-training toolbox and benchmark. https://github.com/open-mmlab/mmpretrain, 2023.4", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3, 4", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4, 6", + "[11] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, 2023. 7, 8", + "[12] Chaoqun Du, Yulin Wang, Jiayi Guo, Yizeng Han, Jie Zhou, and Gao Huang. Unitta: Unified benchmark and versatile framework towards realistic test-time adaptation. arXiv preprint arXiv:2407.20080, 2024. 2, 7, 8", + "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7", + "[14] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 3", + "[15] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 1, 2" + ], + "bbox": [ + 93, + 214, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 1", + "[17] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8", + "[18] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. Note: Robust continual test-time adaptation against temporal correlation. In NeurIPS, 2022. 7, 8", + "[19] Jiayi Guo, Chaoqun Du, Jiangshan Wang, Huijuan Huang, Pengfei Wan, and Gao Huang. Assessing a single image in reference-guided image synthesis. In AAAI, 2022. 3", + "[20] Jiayi Guo, Hayk Manukyan, Chenyu Yang, Chaofei Wang, Levon Khachatryan, Shant Navasardyan, Shiji Song, Humphrey Shi, and Gao Huang. Faceclip: Facial image-to-video translation via a brief text description. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3", + "[21] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Humphrey Shi, Gao Huang, and Shiji Song. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, 2023. 3", + "[22] Jiayi Guo, Xingqian Xu, Yifan Pu, Zanlin Ni, Chaofei Wang, Manushree Vasu, Shiji Song, Gao Huang, and Humphrey Shi. Smooth diffusion: Crafting smooth latent spaces in diffusion models. In CVPR, 2024. 2", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[24] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 6", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4", + "[26] Gao Huang. Dynamic neural networks: advantages and challenges. National Science Review, 2024. 7", + "[27] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 6", + "[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 1", + "[29] Black Forest Labs. Flux. https:// / blackforestlabs.ai/, 2024. 2, 7", + "[30] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1", + "[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLR Workshops, 2017. 1, 2", + "[32] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, 2023. 6" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30511", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1", + "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 7", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4, 6", + "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 6", + "[37] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 1", + "[38] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, 2024. 7, 8", + "[39] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In CoRL, 2022. 3", + "[40] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 7, 8", + "[41] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2023. 7", + "[42] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 3", + "[43] Zanlin Ni, Yulin Wang, Renping Zhou, Yizeng Han, Jiayi Guo, Zhiyuan Liu, Yuan Yao, and Gao Huang. Enat: Rethinking spatial-temporal interactions in token-based image synthesis. In NeurIPS, 2024.", + "[44] Zanlin Ni, Yulin Wang, Renping Zhou, Rui Lu, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Yuan Yao, and Gao Huang. Adanat: Exploring adaptive policy for token-based image generation. In ECCV, 2024.3", + "[45] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 6", + "[46] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In ICML, 2022. 1, 2, 3, 5, 6", + "[47] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 7, 8", + "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 6", + "[49] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3d models. In ICCV, 2015. 3", + "[50] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-tta:" + ], + "bbox": [ + 89, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Test-time adaptation of discriminative models via generative feedback. In NeurIPS, 2023. 2", + "[51] Yifan Pu, Zhuofan Xia, Jiayi Guo, Dongchen Han, Qixiu Li, Duo Li, Yuhui Yuan, Ji Li, Yizeng Han, Shiji Song, et al. Efficient diffusion transformer with step-wise dynamic attention mediators. In ECCV, 2024. 2", + "[52] Yifan Pu, Yiming Zhao, Zhicong Tang, Ruihong Yin, Haoxing Ye, Yuhui Yuan, Dong Chen, Jianmin Bao, Sirui Zhang, Yanbin Wang, et al. Art: Anonymous region transformer for variable multi-layer transparent image generation. arXiv preprint arXiv:2502.18364, 2025. 3", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3", + "[54] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 3", + "[55] Artem Rozantsev, Vincent Lepetit, and Pascal Fua. On rendering synthetic images for training an object detector. Computer Vision and Image Understanding, 2015. 3", + "[56] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In ICLR, 2020. 2", + "[57] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1", + "[58] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In CVPR, 2018. 3", + "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5, 7", + "[60] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In AAAI, 2024. 7, 8", + "[61] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 1, 2", + "[62] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In CVPR, 2024. 3", + "[63] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. In NeurIPS, 2024. 3", + "[64] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozorgtabar. Un-mixing test-time normalization statistics: Combating label temporal correlation. arXiv preprint arXiv:2401.08328, 2024. 8", + "[65] Yun-Yun Tsai, Fu-Chen Chen, Albert YC Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. CVPR, 2024. 1, 2, 3, 5, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30512", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2, 5, 7, 8", + "[67] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2", + "[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 1, 2, 7, 8", + "[69] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. Nerf-supervision: Learning dense object descriptors from neural radiance fields. In ICRA, 2022. 3", + "[70] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In CVPR, 2023. 7, 8", + "[71] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 1, 2, 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "30513", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_model.json b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56c28bc1953e6b516641f4b7494a4c0228356cfc --- /dev/null +++ b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_model.json @@ -0,0 +1,2697 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.202, + 0.842, + 0.258 + ], + "angle": 0, + "content": "Jiayi Guo\\(^{1,2}\\) Junhao Zhao\\(^{1}\\) Chaoqun Du\\(^{1}\\) Yulin Wang\\(^{1}\\) Chunjiang Ge\\(^{1}\\) Zanlin Ni\\(^{1}\\) Shiji Song\\(^{1}\\) Humphrey Shi\\(^{2*}\\) Gao Huang\\(^{1*}\\) \\(^{1}\\)Tsinghua University \\(^{2}\\)SHI Labs @ Georgia Tech" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.283, + 0.487, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.283, + 0.907, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.468, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Figure 1. Comparison of different test-time adaptation (TTA) frameworks. (a) Traditional TTA methods continuously adapt source model weights to fit target data batches. However, their performance is sensitive to the amount and order of target data streams, e.g., adapting the model with batches containing data from only a single category can lead to overfitting. (b) Diffusion-driven TTA methods project the target data back to the synthetic domain of diffusion models, which still maintains domain misalignment with the source domain. (c) We propose the Synthetic-domain Alignment (SDA) framework for TTA, which simultaneously aligns the domains of the source model and target data with the same synthetic domain for superior performance." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.565, + 0.327, + 0.58 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.591, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Test-time adaptation (TTA) aims to improve the performance of source-domain pre-trained models on previously unseen, shifted target domains. Traditional TTA methods primarily adapt model weights based on target data streams, making model performance sensitive to the amount and order of target data. The recently proposed diffusion-driven TTA methods mitigate this by adapting model inputs instead of weights, where an unconditional diffusion model, trained on the source domain, transforms target-domain data into a synthetic domain that is expected to approximate the source domain. However, in this paper, we reveal that although the synthetic data in diffusion-driven TTA seems indistinguishable from the source data, it is unaligned with, or even markedly different from the latter for deep networks. To address this issue, we propose a Synthetic-Domain Alignment (SDA) framework. Our key insight is to fine-tune the source model with synthetic data to ensure better alignment. Specifically, we first employ a conditional diffusion model to generate labeled" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.75 + ], + "angle": 0, + "content": "samples, creating a synthetic dataset. Subsequently, we use the aforementioned unconditional diffusion model to add noise to and denoise each sample before fine-tuning. This Mix of Diffusion (MoD) process mitigates the potential domain misalignment between the conditional and unconditional models. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLMs, e.g., LLaVA) demonstrate that SDA achieves superior domain alignment and consistently outperforms existing diffusion-driven TTA methods. Our code is available at https://github.com/SHI-Labs/Diffusion-Driven-Test-Time-Adaptation-via-Synthetic-Domain-Alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.769, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.909, + 0.906 + ], + "angle": 0, + "content": "Test-Time Adaptation (TTA) [15, 17, 31, 46, 61, 65, 66, 68, 71] is an emerging research field that tackles domain misalignment when source models are evaluated on shifted target data. Unlike traditional domain adaptation (DA) [16, 37, 57] and source-free adaptation (SFA) [28, 30, 33], TTA addresses more practical scenarios where neither source data nor complete target data are accessible. Instead, the adapta" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.889, + 0.242, + 0.902 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30503" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.485, + 0.171 + ], + "angle": 0, + "content": "
Model Adaptation DirectionData Adaptation Direction
Traditional TTAExpected: Source→Target Risk: Imbalanced Target StreamsN/A
Diffusion-driven TTAN/AExpected: Target→Source Actual: Target→Synthetic
SDA (Ours)Source→SyntheticTarget→Synthetic
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.177, + 0.466, + 0.191 + ], + "angle": 0, + "content": "Table 1. Adaptation directions of different TTA methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.2, + 0.44, + 0.215 + ], + "angle": 0, + "content": "tion relies solely on streaming batches of target data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.399 + ], + "angle": 0, + "content": "Traditional TTA methods (Fig. 1a) [15, 31, 50, 56, 61, 66, 68, 71] typically employ a source-to-target model adaptation framework. These approaches continuously update the source model weights by processing target data batches. Without annotations, the adaptation process relies either on batch-wise updates of model statistics [31, 56, 66, 68], or unsupervised or self-supervised auxiliary tasks [15, 50, 61]. However, small or imbalanced batches may poorly represent the target domain, making these approaches sensitive to the amount and order of the data stream [12, 17, 66]. For instance, adapting the model with batches containing data from only a single category can lead to overfitting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Recently, the impressive generation capabilities of diffusion models [22, 25, 29, 48, 51, 53, 67] have sparked the development of diffusion-driven TTA methods (Fig. 1b) [17, 46, 65], leveraging a target-to-source framework. These approaches employ an unconditional diffusion model, pretrained on the source domain, aiming to project each target sample to the source domain independently. This enables the source model to make predictions without modifying its weights. As a preliminary work, DiffPure [46] addresses adversarial perturbations by first applying a forward diffusion process, introducing a small amount of noise to the target data, followed by a reverse diffusion process to restore a clean image to approach the source domain. Building on this concept, more recent studies [17, 65] tackle challenging domain shifts—such as severe data corruption—by incorporating additional structural guidance from the target data, helping preserve semantics and improve performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.824 + ], + "angle": 0, + "content": "In this paper, we uncover that while diffusion-driven TTA methods aim to project target data back to the source domain, the projected target data remains confined within the synthetic domain of the unconditional diffusion model. As the synthetic-domain data are ultimately processed by the source-domain model, this domain misalignment limits the final performance. To address this issue, we propose Synthetic-Domain Alignment (SDA) (Fig. 1c), a new category of framework for TTA tasks which simultaneously aligns the domains of the source model and target data with the same synthetic domain of a diffusion model (Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "SDA distinguishes itself from existing diffusion-driven TTA methods [17, 46, 65] by introducing an additional source-to-synthetic model adaptation phase before testing on adapted target data (Fig. 2). Since the adapted target data aligns with the synthetic domain generated by an un" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.094, + 0.868, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.207, + 0.906, + 0.263 + ], + "angle": 0, + "content": "Figure 2. Enhanced domain alignment with our framework. Prior diffusion-driven TTA methods struggle with the domain misalignment between the source model and synthetic data, which we resolve by aligning the source model to the synthetic domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.529 + ], + "angle": 0, + "content": "conditional diffusion model, SDA aims to adapt the source model to this same synthetic domain. Specifically, SDA employs a Mix-of-Diffusion (MoD) technique to generate synthetic data for model adaptation. Given that the source data is inaccessible after pretraining, MoD first uses a conditional diffusion model to generate samples conditioned on domain-agnostic labels, creating a labeled synthetic dataset. Subsequently, the aforementioned unconditional diffusion model is leveraged by MoD add noise to and denoise these samples, addressing potential domain misalignment between the conditional and unconditional models. With a sufficiently large synthetic dataset, the fine-tuned model becomes highly effective at discriminating within the synthetic domain. Thus, the SDA framework transforms the cross-domain TTA task into an in-domain prediction task by aligning both the source model and target data with the same synthetic domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.667 + ], + "angle": 0, + "content": "SDA is a general framework, not limited to specific fine-tuning techniques or diffusion-driven data adaptation methods. This flexibility allows future advancements in these areas to further broaden its applicability. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLM, e.g., LLaVA) demonstrate that SDA consistently outperforms existing methods. Moreover, the effectiveness of our approach is reinforced through visualization analysis and ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.655, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Test-time adaptation (TTA) is an emerging research area that addresses domain shifts by adapting either models [15, 31, 50, 56, 61, 66, 68, 71] or data [17, 46, 65] during evaluation on streaming target data batches. Early model adaptation methods update batch normalization statistics to match the target distribution [31, 56, 66, 71], while others leverage self-supervised tasks like rotation prediction [61] or image restoration [15, 50] to adjust model weights. However, these approaches rely heavily on continuous weight updates, making them sensitive to the amount, order, and diversity of target data. In contrast, diffusion-driven TTA methods [17, 46, 65] focus on data adaptation by projecting each target sample back into the source domain, achiev-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30504" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.106, + 0.089, + 0.177, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.103, + 0.252, + 0.123 + ], + "angle": 0, + "content": "Forward Diffusion (Add Noise t steps)" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.089, + 0.322, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.104, + 0.395, + 0.123 + ], + "angle": 0, + "content": "Reverse Diffusion (Denoise \\(t\\) steps)" + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.089, + 0.466, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.145, + 0.4, + 0.155 + ], + "angle": 0, + "content": "(a) Diffusion-driven data adaptation on source data" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.158, + 0.158, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.159, + 0.2, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.16, + 0.255, + 0.167 + ], + "angle": 0, + "content": "\\(t = 600\\)" + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.168, + 0.259, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.16, + 0.308, + 0.167 + ], + "angle": 0, + "content": "\\(t = 700\\)" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.167, + 0.312, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.16, + 0.359, + 0.167 + ], + "angle": 0, + "content": "\\(t = 800\\)" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.167, + 0.342, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.16, + 0.385, + 0.167 + ], + "angle": 0, + "content": "300 \\(t =\\)" + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.167, + 0.395, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.16, + 0.435, + 0.167 + ], + "angle": 0, + "content": "00 \\(t =\\)" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.167, + 0.437, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.16, + 0.461, + 0.167 + ], + "angle": 0, + "content": "1000" + }, + { + "type": "image", + "bbox": [ + 0.438, + 0.167, + 0.465, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.21, + 0.158, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.21, + 0.258, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.21, + 0.311, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.21, + 0.364, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.21, + 0.406, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.21, + 0.464, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.252, + 0.158, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.252, + 0.211, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.252, + 0.259, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.252, + 0.312, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.252, + 0.342, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.252, + 0.405, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.252, + 0.44, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.293, + 0.473, + 0.304 + ], + "angle": 0, + "content": "(b) Diffusion-driven data adaptation results (on source data) under different timesteps" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.362 + ], + "angle": 0, + "content": "Figure 3. (a) Illustration of diffusion-driven data adaptation on source data and (b) Adapted images across different timesteps. The results are obtained using DDA [17], with no noticeable visual degradation observed in the adapted images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.489 + ], + "angle": 0, + "content": "ing stable performance without online model updates. DiffPure [46] purifies adversarial samples with diffusion models, while DDA [17] and GDA [65] use structural guidance to preserve image content under severe corruption. Building on this, our work investigates the issue of misalignment between domains of the source model and synthetic images in diffusion-driven TTA and proposes a new TTA framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Synthetic data for discriminative tasks. Synthetic data, generated by models [20, 21, 42-44, 48, 52, 53] rather than collected from the real world, has shown significant potential in enhancing visual representations for various discriminative tasks [14, 62, 63]. It has been effectively applied in areas such as recognition [3, 63], object detection [49, 55], semantic segmentation [7, 54, 58], image assessment [19], autonomous driving [1], and robotics [39, 69]. In this work, we explore the potential of diffusion-generated synthetic data to enhance domain alignment in TTA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.226, + 0.67 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.484, + 0.768 + ], + "angle": 0, + "content": "In this section, we introduce the background of diffusion-driven TTA methods and identify their source-synthetic domain misalignment issue in Sec. 3.1 and Sec. 3.2. To tackle this issue, we propose the synthetic-domain alignment (SDA) framework in Sec. 3.3 and introduce its key technique, the mix of diffusion (MoD) in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.222, + 0.792 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Diffusion process. Given a source data point \\( \\boldsymbol{x}_0^{\\mathrm{src}} \\sim p_0^{\\mathrm{src}} \\), diffusion models \\( \\epsilon_{\\theta} \\) [25] gradually transform \\( p_0^{\\mathrm{src}} \\) into a Gaussian noise distribution \\( N(\\mathbf{0}, \\mathbf{I}) \\) through a \\( T \\)-step forward diffusion process. At each timestep \\( t \\in \\{1, 2, \\dots, T\\} \\), the intermediate state \\( \\boldsymbol{x}_t^{\\mathrm{src}} \\sim p_t^{\\mathrm{src}} \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}} = \\sqrt {1 - \\beta_ {t}} \\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}} + \\sqrt {\\beta_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {1}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.078, + 0.906, + 0.183 + ], + "angle": 0, + "content": "
Timestep t5006007008009001000
Swin-B (Source-Synthetic, Misaligned)61.657.855.752.645.038.5
Swin-B (Synthetic-Synthetic, Aligned)67.665.965.061.555.748.4
Δ+6.0+8.1+9.3+8.9+10.7+9.9
ConvNeXt-B (Source-Synthetic, Misaligned)65.161.360.357.250.041.5
ConvNeXt-B (Synthetic-Synthetic, Aligned)70.668.267.464.958.550.7
Δ+5.5+6.9+7.1+7.7+8.5+9.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.186, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Table 2. Source model accuracy across different timesteps of diffusion-driven data adaptation. For suitable timesteps for TTA \\((t\\geq 500)\\) [17], model accuracy shows a monotonically decreasing trend with the growth of timestep, indicating the increase of the misalignment of source domain \\(p_0^{\\mathrm{src}}\\) and synthetic domain \\(p_{0,\\mathrm{u}}^{\\mathrm{syn}}\\). By aligning the source model to the synthetic domain, our methods (rows 3 & 6) significantly help alleviate the performance degradation. Results are evaluated on the ImageNet [9] validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.366 + ], + "angle": 0, + "content": "where \\(\\epsilon_{t}\\) is random Gaussian noise and \\(\\beta_{t} \\in (0,1)\\) represents the diffusion rate at step \\(t\\). By defining \\(\\alpha_{t} = 1 - \\beta_{t}\\), \\(\\overline{\\alpha_{t}} = \\prod_{t=1}^{T} \\alpha_{t}\\) and \\(\\epsilon \\sim N(\\mathbf{0}, \\mathbf{I})\\), we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.372, + 0.905, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} ^ {\\mathrm {s r c}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.432 + ], + "angle": 0, + "content": "The reverse diffusion process recovers a clean \\(\\widehat{\\boldsymbol{x}_0^{\\mathrm{src}}}\\) by progressively removing noise from \\(\\boldsymbol{x}_T^{\\mathrm{src}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.438, + 0.905, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\widehat {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(\\widehat {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.813, + 0.493 + ], + "angle": 0, + "content": "where \\(\\sigma_{t}\\) is the posterior noise variance [25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.905, + 0.541 + ], + "angle": 0, + "content": "Diffusion-driven data adaptation. Denote by \\( p_0^{\\mathrm{trg}} \\) the target data distribution, from which each target data point \\( \\boldsymbol{x}_t^{\\mathrm{trg}} \\) is sampled. Prior work [46] demonstrates that:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.55, + 0.905, + 0.568 + ], + "angle": 0, + "content": "\\[\nD _ {\\mathrm {K L}} \\left(p _ {t + 1} ^ {\\mathrm {s r c}} \\mid p _ {t + 1} ^ {\\mathrm {t r g}}\\right) - D _ {\\mathrm {K L}} \\left(p _ {t} ^ {\\mathrm {s r c}} \\mid p _ {t} ^ {\\mathrm {t r g}}\\right) \\leq 0, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.906, + 0.669 + ], + "angle": 0, + "content": "where \\(D_{\\mathrm{KL}}\\) is the KL divergence. Since \\(p_T^{\\mathrm{src}} = p_T^{\\mathrm{trg}} = N(\\mathbf{0}, \\mathbf{I})\\), for any arbitrarily small value \\(\\delta\\), there exists a minimum timestep \\(t^*\\) such that \\(D_{\\mathrm{KL}}(p_{t^*}^{\\mathrm{src}}||p_{t^*}^{\\mathrm{trg}}) < \\delta\\). As discussed in [17], a diffusion process with \\(T = 1000\\) requires \\(t^* \\geq 500\\) to eliminate domain shifts. We empirically validate this choice of \\(t^*\\) in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.906, + 0.746 + ], + "angle": 0, + "content": "The initial diffusion-driven data adaptation in Diff-Pure [46] is carried out as follows: Given the minor divergence between \\( p_{t^*}^{\\mathrm{src}} \\) and \\( p_{t^*}^{\\mathrm{trg}} \\), we recover each \\( \\boldsymbol{x}_{t^*}^{\\mathrm{trg}} \\sim p_{t^*}^{\\mathrm{trg}} \\) to its corresponding \\( \\widehat{\\boldsymbol{x}_0^{\\mathrm{src}}} \\sim p_0^{\\mathrm{src}} \\) by executing the reverse diffusion process (Eq. 3) \\( t^* \\) times." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.823 + ], + "angle": 0, + "content": "Since the reverse process is stochastic, subsequent works [17, 65] further introduce additional structure guidance to ensure the content consistency between each \\( \\boldsymbol{x}_0^{\\mathrm{trg}} \\) and its adapted \\( \\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}} \\). In this work, we adopt the same data adaptation process as DDA [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.831, + 0.905, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} = \\widehat {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} - \\boldsymbol {w} \\nabla_ {\\widetilde {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}}} \\left\\| \\phi \\left(\\boldsymbol {x} _ {0} ^ {\\mathrm {t r g}}\\right) - \\phi \\left(\\widetilde {\\boldsymbol {x} _ {0 , t} ^ {\\mathrm {s r c}}}\\right) \\right\\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.866, + 0.905, + 0.903 + ], + "angle": 0, + "content": "where \\(\\widehat{\\pmb{x}_{t - 1}^{\\mathrm{src}}}\\) is computed as Eq. 3, \\(\\pmb{w}\\) is the structure guidance scale, \\(\\phi\\) is a structure extractor [17] and \\(\\widehat{\\pmb{x}_{0,t}^{\\mathrm{src}}}\\) is an esti" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30505" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.905, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.345, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 4. Overview of the Synthetic-Domain Alignment (SDA) framework. SDA is a novel TTA framework aligning both the domains of the source model and the target data with the synthetic domain. SDA involves three phases: (left): a source-domain model pretraining phase, where the source model is trained on source data prior to TTA; (middle): a source-to-synthetic model adaptation phase, where the source model is adapted to a synthetic-domain model using synthetic data generated via a Mix of Diffusion (MoD) technique; and (right): a target-to-synthetic data adaptation phase, where target data is adapted into synthetic data using an unconditional diffusion model. Finally, the adapted synthetic data is fed into the synthetic-domain model for test-time inference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.451, + 0.483, + 0.498 + ], + "angle": 0, + "content": "mate of \\(\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}\\) at timestep \\(t\\) [25]. Diffusion-driven data adaptation is typically performed with unconditional diffusion models [10] since the target data labels are unknown." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.509, + 0.433, + 0.526 + ], + "angle": 0, + "content": "3.2. Source-Synthetic Domain Misalignment" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.531, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Unlike a real source data point \\( \\boldsymbol{x}_0^{\\mathrm{src}} \\), the recovered version \\( \\widetilde{\\boldsymbol{x}_{0}^{\\mathrm{src}}}, \\) derived from \\( \\boldsymbol{x}_{t^{*}}^{\\mathrm{trg}} \\) is synthetic. Specifically, \\( \\boldsymbol{x}_0^{\\mathrm{src}} \\) follows the source domain \\( p_0^{\\mathrm{src}} \\), whereas \\( \\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}} \\) follows the synthetic domain \\( p_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) of an unconditional diffusion model \\( \\epsilon_{\\theta}^{\\mathrm{u}} \\) with parameters \\( \\theta \\). In this section, we empirically reveal the misalignment between the source domain \\( p_0^{\\mathrm{src}} \\) and the synthetic domain \\( p_{0,\\mathrm{u}}^{\\mathrm{syn}} \\), and investigate how this misalignment impacts the performance of the source model. For simplicity and clarity, we will substitute \\( \\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}} \\) with \\( \\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) throughout the following discussion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Based on the above analysis, diffusion-driven TTA methods face two potential misalignments: (1) source-target domain misalignment arising from inherent data distribution shifts, which has been the primary focus of prior research, and (2) source-synthetic domain misalignment introduced by diffusion models, which we address as the main focus of this work, complementing existing efforts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To precisely examine the impact of the source-synthetic domain misalignment and isolate it from the influence of the source-target domain misalignment, we evaluate the performance of the source model \\( f \\) on synthetic data \\( x_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) adapted by the diffusion model \\( \\epsilon_{\\theta}^{\\mathrm{u}} \\) from source data \\( x_0^{\\mathrm{src}} \\) (Fig. 3a). Specifically, we test ImageNet pretrained models [9] on the ImageNet validation set adapted by the popular diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.633 + ], + "angle": 0, + "content": "driven data adaptation method, DDA [17] across different timesteps \\(t\\). As indicated in Tab. 2, with an increase in \\(t\\), the accuracy of the model exhibits a monotonically decreasing trend. For ideal \\(t^* \\geq 500\\) for TTA tasks, performance degradation of more than \\(18.8\\%\\) for ConvNeXt [36] and \\(21.8\\%\\) for Swin [35] is observed compared to their official results [8] on source data which are \\(83.4\\%\\) and \\(83.9\\%\\). This indicates a significant domain misalignment between \\(p_0^{\\mathrm{src}}\\) and \\(p_{0,\\mathrm{u}}^{\\mathrm{syn}}\\). By fine-tuning the source model on the synthetic data generated by our Mix of Diffusion process (introduced in Sec. 3.4), the aligned models \\(f'\\) achieve performance improvements of \\(5.5\\%\\) for ConvNeXt and \\(6.0\\%\\) for Swin." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.726 + ], + "angle": 0, + "content": "In Fig. 3b, we show that the diffusion synthetic data \\( \\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) and source data \\( \\boldsymbol{x}_0^{\\mathrm{src}} \\) exhibit no noticeable visual differences across different timesteps \\( t \\). This further suggests that the performance degradation is not due to the quality of diffusion-generated images but rather the implicit misalignment between the source and synthetic domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.869, + 0.756 + ], + "angle": 0, + "content": "3.3. Synthetic-Domain Alignment Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.869 + ], + "angle": 0, + "content": "Given that the diffusion-adapted data \\( \\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) aligns more closely with the synthetic domain \\( p_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) instead of the source domain \\( p_0^{\\mathrm{src}} \\), we propose simultaneously adapting the source model \\( f \\) to the same synthetic domain \\( p_0^{\\mathrm{syn}} \\). By doing so, the alignment between the data and model within \\( p_0^{\\mathrm{syn}} \\) can be effectively achieved. To this end, we introduce a novel TTA framework: Synthetic-Domain Alignment (SDA)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Fig. 4, we present the complete diagram of SDA, which consists of three key phases: (1) a source-domain" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30506" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "model pretraining phase, where the source model \\(f\\) is trained on source data \\(\\pmb{x}_0^{\\mathrm{src}}\\) prior to TTA; (2) a source-to-synthetic model adaptation phase, where \\(f\\) is fine-tuned to a synthetic-domain model \\(f'\\); and (3) a target-to-synthetic data adaptation phase, where target data \\(\\pmb{x}_0^{\\mathrm{trg}}\\) is adapted into synthetic data \\(\\pmb{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}\\) using an unconditional diffusion model \\(\\epsilon_{\\theta}^{\\mathrm{u}}\\) following Eq. 5. Finally, the adapted synthetic data \\(\\pmb{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}\\) is fed into the synthetic-domain model \\(f'\\) for test-time inference. Consistent with the standard TTA protocol [66], the source data is accessible only during the source model pretraining phase and remains inaccessible during both the model and data adaptation phases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.348 + ], + "angle": 0, + "content": "The rationale behind SDA is straightforward: by aligning the model and data domains to the same synthetic domain, the original cross-domain TTA task is transformed into an easier in-domain prediction task, thus addressing the core challenge of TTA and improving performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.482, + 0.439 + ], + "angle": 0, + "content": "Since the pretraining phase follows standard supervised learning, and the data adaptation phase aligns with DDA [17], we omit detailed explanations of these phases. Instead, we focus on how we adapt the source model \\( f \\) to an ideal synthetic domain model \\( f' \\) using a novel technique called Mix-of-Diffusion (MoD), introduced next." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.449, + 0.425, + 0.465 + ], + "angle": 0, + "content": "3.4. Model Adaptation via Mix of Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.482, + 0.594 + ], + "angle": 0, + "content": "As shown in Fig. 4, MoD consists of two main processes: a data generation process powered by a conditional diffusion model \\(\\epsilon_{\\eta}^{\\mathrm{c}}\\) with parameters \\(\\eta\\), and a data alignment process powered by the same unconditional diffusion model \\(\\epsilon_{\\theta}^{\\mathrm{u}}\\) used in the target-to-synthetic data adaptation phase. It is worth noting that, both \\(\\epsilon_{\\eta}^{\\mathrm{c}}\\) and \\(\\epsilon_{\\theta}^{\\mathrm{u}}\\) are also pretrained on source data \\(\\mathbf{x}_0^{\\mathrm{src}}\\) and have never been exposed to target data \\(\\mathbf{x}_0^{\\mathrm{trg}}\\), in accordance with the TTA setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.482, + 0.705 + ], + "angle": 0, + "content": "Conditional diffusion data generation. The generation process leverages the conditional generation capability of \\(\\epsilon_{\\eta}^{\\mathrm{c}}\\) to synthesize conditional synthetic data \\(x_{0,c}^{\\mathrm{syn}} \\sim p_{0,c}^{\\mathrm{syn}}\\). In the context of TTA, the source and target domains share the same domain-agnostic label set \\(\\{y_i\\}_{i=1}^{K}\\). Utilizing \\(\\epsilon_{\\eta}^{\\mathrm{c}}\\), we uniformly generate samples for each class \\(y_i\\) from Gaussian noise \\((x_{T,c}^{\\mathrm{syn}})\\) through a \\(T\\)-step reverse diffusion process:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.716, + 0.482, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t - 1, c} ^ {\\text {s y n}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t, c} ^ {\\text {s y n}} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\overline {{\\alpha_ {t}}}}} \\epsilon_ {\\eta} ^ {\\mathrm {c}} \\left(\\boldsymbol {x} _ {t, c} ^ {\\text {s y n}}, t, y _ {i}\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\epsilon}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.482, + 0.838 + ], + "angle": 0, + "content": "The generation capability of \\(\\epsilon_{\\eta}^{\\mathrm{c}}\\) allows for the construction of a labeled synthetic-domain dataset \\(\\{x_{0,\\mathrm{c}}^{\\mathrm{syn}},y\\}^{N}\\) of arbitrary size \\(N\\) without any manual data collection. By fine-tuning the source model \\(f\\) on this synthetic dataset, an adapted model \\(f_{\\mathrm{c}}^{\\prime}\\) on domain \\(p_{0,\\mathrm{c}}^{\\mathrm{syn}}\\) can be obtained." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.482, + 0.903 + ], + "angle": 0, + "content": "Unconditional diffusion data alignment. However, since the test-time adapted data \\( x_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) is subject to domain \\( p_{0,\\mathrm{u}}^{\\mathrm{syn}} \\), we argue that there is still potential misalignment between different synthetic domains \\( p_{0,\\mathrm{c}}^{\\mathrm{syn}} \\) and \\( p_{0,\\mathrm{u}}^{\\mathrm{syn}} \\). This is mainly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "because of the differences in architectures and training schemes between \\(\\epsilon_{\\eta}^{\\mathrm{c}}\\) and \\(\\epsilon_{\\theta}^{\\mathrm{u}}\\). We empirically validate this argument in Tab. 10, which indicates that the model \\(f_{\\mathrm{c}}'\\) adapted to domain \\(p_{0,\\mathrm{c}}^{\\mathrm{syn}}\\) performs worse than the model \\(f_{\\mathrm{u}}'\\) adapted to domain \\(p_{0,\\mathrm{u}}^{\\mathrm{syn}}\\) during testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.274 + ], + "angle": 0, + "content": "To obtain the \\(f_{\\mathrm{u}}^{\\prime}\\) on domain \\(p_{0,\\mathrm{u}}^{\\mathrm{syn}}\\) we mirror the target-to-synthetic data adaptation phase as a conditional synthetic data \\((\\boldsymbol{x}_{0,\\mathrm{c}}^{\\mathrm{syn}})\\) to unconditional synthetic data \\((\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}})\\) adaptation process. In specific, we use the same \\(t^{*}\\) in Sec. 3.1 with Eq. 2 to obtain \\(\\boldsymbol{x}_{t^{*},\\mathrm{c}}^{\\mathrm{syn}}\\). According to the analysis in Sec. 3.1, \\(\\boldsymbol{x}_{t^{*},\\mathrm{c}}^{\\mathrm{syn}}\\) will be indistinguishable to its counterpart \\(\\boldsymbol{x}_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}\\) on domain \\(p_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}\\). Therefore, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.282, + 0.905, + 0.301 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t ^ {*}, \\mathrm {u}} ^ {\\text {s y n}} \\approx \\boldsymbol {x} _ {t ^ {*}, \\mathrm {c}} ^ {\\text {s y n}} = \\sqrt {\\overline {{\\alpha_ {t}}}} \\boldsymbol {x} _ {0, c} ^ {\\text {s y n}} + \\sqrt {1 - \\overline {{\\alpha_ {t}}}} \\boldsymbol {\\epsilon}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.905, + 0.37 + ], + "angle": 0, + "content": "Then, following Eq. 3 and Eq. 5, the noisy \\( \\boldsymbol{x}_{t^{*},\\mathrm{u}}^{\\mathrm{syn}} \\) can be gradually denoised to \\( \\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}} \\). Finally, the expected synthetic-domain model \\( f_{\\mathrm{u}}^{\\prime} \\) is obtained by fine-tuning the source-domain model \\( f_{\\mathrm{u}}^{\\prime} \\) on dataset \\( \\{\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}},y\\}^{N} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.477 + ], + "angle": 0, + "content": "Ensembling. As noted in previous diffusion-driven TTA methods [17], while diffusion models generally perform well for data adaptation, they may occasionally produce data points that are less recognizable than the original target data. To address this, prior approaches use an ensemble of model predictions on \\( \\boldsymbol{x}_{0}^{\\mathrm{trg}} \\) and \\( \\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}} \\) as the final output. Following this protocol, the final prediction in SDA is:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.491, + 0.905, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\arg \\max _ {y} \\left(q \\left(y \\mid \\boldsymbol {x} _ {0} ^ {\\operatorname {t r g}}\\right) + q ^ {\\prime} \\left(y \\mid \\boldsymbol {x} _ {0, \\mathrm {u}} ^ {\\operatorname {s y n}}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.519, + 0.905, + 0.551 + ], + "angle": 0, + "content": "where \\(q(\\cdot)\\) and \\(q^{\\prime}(\\cdot)\\) are output distributions of source model \\(f\\) and synthetic-domain model \\(f_{\\mathrm{u}}^{\\prime}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.645, + 0.579 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.905, + 0.707 + ], + "angle": 0, + "content": "In this section, we first evaluate SDA on ImageNet classifiers with standard TTA benchmarks in Sec. 4.1. Next, we assess SDA's scalability across different dataset sizes, tasks, and model architectures in Sec. 4.2. In Sec. 4.3, we demonstrate SDA's advantages through Grad-CAM visualizations [59] and data stream sensitivity tests. Finally, ablation studies in Sec. 4.4 validate the design choices in our SDA framework." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.716, + 0.905, + 0.808 + ], + "angle": 0, + "content": "
ModelSourceMEMODiffPureGDADDASDA (Ours)
ResNet-5018.724.716.831.829.732.5 (+2.8)
Swin-T33.529.524.842.240.042.5 (+2.5)
ConvNeXt-T39.337.828.844.844.247.0 (+2.8)
Swin-B40.537.028.9-44.547.4 (+2.9)
ConvNeXt-B45.645.832.7-49.451.9 (+2.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.821, + 0.905, + 0.89 + ], + "angle": 0, + "content": "Table 3. Comparison results on ImageNet-C. We compare SDA with source models, MEMO [71], DiffPure [46], GDA [65] and DDA [17]. Results are the average accuracy across 15 adaptation domains at severity level 5. SDA shows consistent performance improvements compared to baselines." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30507" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.085, + 0.907, + 0.194 + ], + "angle": 0, + "content": "
GaussianShotImpluseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source48.047.148.332.414.439.339.255.750.151.575.152.029.548.160.346.1
DiffPure42.642.040.916.822.219.123.039.029.19.162.53.034.346.860.732.7
DDA59.955.755.930.730.538.139.253.248.041.971.048.146.959.562.049.4
SDA (Ours)60.357.658.135.435.342.942.455.549.944.672.645.450.263.064.551.9 (+2.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.2, + 0.907, + 0.23 + ], + "angle": 0, + "content": "Table 4. Detailed comparisons of SDA and baselines across 15 adaptation domains of ImageNet-C. SDA shows the best average accuracy. The results are tested with ConvNeXt-B. Comparisons using other models are deferred to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.248, + 0.415, + 0.264 + ], + "angle": 0, + "content": "4.1. Main Results on ImageNet Classifiers" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.274, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Settings. We choose DDA [17] as our primary competitor since it is the best open-sourced method. SDA is also compared with DiffPure [46], MEMO [71], and the recent SOTA GDA [65] using their reported results. Source model performance is reported as \"Source\". DiT [48] and ADM [10] are adopted to generate and align 50K synthetic data for 15-epoch finetuning. For different source models, the synthetic data only needs to be generated once. Our results are tested on standard TTA benchmarks, ImageNet-C [24] (severity level 5) and ImageNet-W [32] using various models including ResNet [23], Swin [35] and ConvNeXt [36]. More implementation details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.478, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison results on ImageNet-C. We begin by evaluating the performance of SDA on ImageNet-C. As reported in Tab. 3, our proposed SDA consistently outperforms all baseline methods across different model architectures and sizes. We emphasize the performance improvement over DDA, as we adopted DDA for target data adaptation in SDA. Compared to DDA, our SDA improves accuracy by \\(2.5\\% - 2.9\\%\\). This significant improvement indicates the misalignment between the source and synthetic domains, validating the effectiveness of our synthetic-domain alignment framework. Moreover, compared to the recent SOTA GDA, SDA also achieves an improvement of \\(2.2\\%\\) with ConvNeXt-T. Notably, SDA focuses on synthetic domain alignment, an orthogonal research direction to existing diffusion-driven methods on better adapting the target data. Therefore, the performance of SDA could potentially be further enhanced with the release of more advanced codebases like GDA. Compared to the model adaptation method, MEMO, three diffusion-driven methods (SDA, DDA, and GDA) all demonstrate superior performance, highlighting the effectiveness of diffusion models in assisting TTA tasks. DiffPure presents worse results since it is primarily designed for adversarial attacks. Without the structural guidance introduced in DDA and GDA, DiffPure may not effectively recover images with severe domain shifts. In Tab. 4, we provide a detailed comparison of the results of SDA and baselines. SDA surpasses DiffPure in all 15 adaptation domains and outperforms DDA in 14 out of 15 domains, fur" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.246, + 0.907, + 0.35 + ], + "angle": 0, + "content": "
ModelSourceDiffPureDDASDA (Ours)
ResNet-5037.729.152.854.7 (+1.9)
Swin-T66.552.765.967.3 (+1.4)
ConvNeXt-T67.655.867.969.4 (+1.5)
Swin-B69.155.568.370.6 (+2.3)
ConvNeXt-B70.157.770.372.3 (+2.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.361, + 0.907, + 0.389 + ], + "angle": 0, + "content": "Table 5. Comparison results on ImageNet-W. SDA shows consistent performance improvements compared to baselines." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.766, + 0.421 + ], + "angle": 0, + "content": "ther affirming the superiority of SDA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Comparison results on ImageNet-W. We extend our evaluation to ImageNet-W to assess SDA's performance under watermark-based domain shifts. As shown in Tab. 5, SDA consistently surpasses all baselines across different models. Compared to our primary baseline, DDA, SDA achieves accuracy gains ranging from \\(1.4\\%\\) to \\(2.3\\%\\). Furthermore, the results in Tab. 5 reveal potential performance drops when DDA is applied to ImageNet-W with Swin-T and Swin-B models, suggesting that synthetic data may be less recognizable by the original source models. Although advancements in diffusion techniques could potentially improve outcomes, the consistent gains achieved by SDA indicate that aligning the source model with the synthetic domain offers a convenient and effective solution to enhance performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.905, + 0.664 + ], + "angle": 0, + "content": "4.2. Scalability to Other Datasets, Tasks and Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.906, + 0.761 + ], + "angle": 0, + "content": "In addition to standard benchmarks, an effective TTA method should demonstrate its superiority across various dataset scales, task formats, and model architectures. In this section, we assess SDA from these aspects to validate its scalability. Implementation details of each experiment can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Scaling to small datasets. We first evaluate the effectiveness of SDA on scenarios where both source and target domains are small-scale. Specifically, we test SDA on CIFAR-10-C [24] with ResNet-18 [23] as the source classifier. EDM [27] is used to generate synthetic data, and I-DDPM [45] is applied for data alignment. As shown in Tab. 6, SDA consistently outperforms both the source model and DDA, achieving an average accuracy improvement of \\(7.1\\%\\) over DDA." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30508" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.184 + ], + "angle": 0, + "content": "
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source23.528.720.841.230.748.248.151.863.663.479.528.056.535.960.745.4
DDA79.480.376.841.356.649.048.977.473.562.680.561.561.257.673.265.3
SDA (Ours)82.683.579.559.661.465.364.676.578.471.588.953.475.464.480.772.4 (+7.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.187, + 0.854, + 0.203 + ], + "angle": 0, + "content": "Table 6. Comparisons of SDA and baselines with ResNet-18 on CIFAR-10-C. SDA shows the best average accuracy." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.209, + 0.905, + 0.304 + ], + "angle": 0, + "content": "
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPirelateJEPGAvg
Source9.813.210.022.75.720.422.136.728.453.966.69.724.316.043.825.6
DDA16.621.116.045.040.949.738.736.738.635.055.213.356.656.358.838.6
SDA (Ours)17.021.116.245.543.051.139.938.740.236.357.913.159.158.159.739.8 (+1.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.307, + 0.863, + 0.322 + ], + "angle": 0, + "content": "Table 7. Comparisons of SDA and baselines with DeepLabv3 on PASCAL VOC-C. SDA shows the best average mIOU." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.483, + 0.459 + ], + "angle": 0, + "content": "Scaling to Semantic Segmentation Tasks. We extend our evaluation to dense prediction tasks by using PASCAL-VOC-C [13] as a standard semantic segmentation benchmark with DeepLabv3 [6] as the source segmenter. Dataset Diffusion [41] generates synthetic segmentation data, and FLUX Schnell [29] is used for data alignment. As shown in Tab. 7, SDA achieves the best performance, with an average mIOU improvement of \\(1.2\\%\\) over DDA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.722 + ], + "angle": 0, + "content": "Scaling to multimodal large language models (MLLMs). As an emerging research direction, MLLMs [26, 34] present advanced visual question answering [2] capability. We design a language-based classification task format to test how SDA can help MLLMs on TTA tasks on ImageNet-C: Given an image, ask an MLLM (LLaVA 1.5-7b [34] in our experiments) to choose the correct image class from four provided options. In Tab. 8, the \"Source-Zero\" setting tests the zero-shot results of pretrained LLaVA. The \"Source\" and \"DDA\" settings evaluate a source-data fine-tuned LLaVA while the \"SDA\" setting tests a synthetic-data fine-tuned LLaVA. The fine-tuning task format is the same as that at test time. Pre-trained LLaVA already exhibits strong performance. While DDA improves results in some domains, it does not yield an overall gain compared to source-data fine-tuned LLaVA. In contrast, SDA aligns LLaVA with the synthetic domain, achieving the best accuracy with an improvement of \\(2.4\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.194, + 0.753 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Visualization. To demonstrate how synthetic data fine-tuning in SDA enhances the performance of diffusion-driven TTA methods, we employ Gradient-weighted Class Activation Mapping (Grad-CAM) [59] to visualize the image regions that most influence classification scores across different images and models. As shown in Fig. 5, testing target images with the source model reveals distinct differences in activation maps and the occurrence of incorrect predictions compared to those from source images," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.339, + 0.901, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Figure 5. Grad-CAM visualization comparison. The first row shows activation maps for source and target images tested with the source model. The second row displays activation maps for diffusion synthetic images tested with the source model (DDA) and our synthetic-domain model (SDA). SDA aligns closely with the source model's response to source images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.701 + ], + "angle": 0, + "content": "underscoring the performance degradation due to domain shifts. Despite using adapted synthetic images, DDA still risks focusing on inappropriate regions and producing incorrect predictions. This highlights the domain misalignment of the synthetic data and source model. In contrast, SDA aligns both the data and model within the same synthetic domain, thereby producing activation maps and predictions that closely resemble those produced by the source model on source images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Data stream sensitivity. In Tab. 9, we test data stream sensitivity using the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings closely aligned with diverse real-world TTA scenarios. In addition to diffusion-driven TTA methods, we include 13 popular traditional TTA methods [4, 5, 11, 12, 18, 38, 40, 47, 60, 66, 68, 70] for comparison. We report the average accuracy of 12 settings using ResNet-50. The results indicate that diffusion-driven TTA methods are preferred in these challenging settings since they are insensitive to different variants of data streams. SDA maintains this insensitivity and showcases the best performance. Detailed comparisons for each setting are provided in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30509" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.198 + ], + "angle": 0, + "content": "
GaussianShotImpluseDerocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAve:
Source-Zero75.276.679.681.474.583.581.187.088.989.595.684.176.290.188.783.5
Source86.886.088.690.185.593.589.793.695.094.998.189.585.696.294.891.2
DDA92.491.592.588.690.192.288.593.093.189.697.875.889.696.497.091.2
SDA (Ours)94.994.194.691.092.994.091.593.695.292.298.583.692.897.397.693.6 (+2.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.201, + 0.842, + 0.216 + ], + "angle": 0, + "content": "Table 8. Comparisons of SDA and baselines with LLaVA on ImageNet-C. SDA shows the best average accuracy." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.226, + 0.905, + 0.288 + ], + "angle": 0, + "content": "
SourceTENT [66]ROID [38]NOTE [18]CoTIA [68]TRIBE [60]BN [40]EATA [67]SANTA [5]RMT [11]UnMIX [64]RoTTA [70]LAME [14]UniTTA [12]DDA [17]SDA (Ours)
18.34.26.26.38.89.39.810.210.611.517.423.124.930.329.331.6 (+2.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.291, + 0.905, + 0.321 + ], + "angle": 0, + "content": "Table 9. Data stream sensitivity comparison. We additionally compare SDA with 13 traditional TTA methods on the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings. Results are reported as average accuracy across settings." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.336, + 0.485, + 0.404 + ], + "angle": 0, + "content": "
ComponentResNet-50Swin-TConvNeXt-T
DDA [17]29.740.044.2
+ Conditional Data Generation30.440.144.6
+ Unconditional Data Alignment32.542.547.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.435 + ], + "angle": 0, + "content": "Table 10. Impact of different components in SDA. Results are evaluated on ImageNet-C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.255, + 0.465 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Components. We examine the impact of two key components in our SDA framework's Mix of Diffusion technique in Tab. 10: (1) synthetic data generation using the conditional diffusion model and (2) synthetic data alignment using the unconditional diffusion model. Fine-tuning source models with synthetic data generated solely by the conditional diffusion model (+ Conditional Data Generation) yields only marginal gains. As discussed in Sec. 3.4, this limited gain arises from the domain misalignment between conditional and unconditional diffusion models. Specifically, using only conditional synthetic data results in models aligned with the conditional diffusion domain, whereas the test data belongs to the unconditional diffusion domain. Therefore, further aligning the synthetic data through the unconditional diffusion model (+ Unconditional Data Alignment) leads to significant performance gains over the baseline DDA. This demonstrates that bridging the misalignment of different diffusion domains is essential for the success of our SDA framework." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Number of fine-tuning images. We examine the impact of different numbers of images \\((N)\\) used during synthetic-domain model fine-tuning in Tab. 11. Interestingly, even with only one image per class \\((N = 1\\mathrm{K})\\), SDA still significantly outperforms DDA. This finding suggests a key attribute of the fine-tuning process: source models are primarily learning to adapt to the synthetic domain itself, rather than acquiring class-specific knowledge. Increasing the number of images helps the fine-tuning process capture the" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.336, + 0.888, + 0.436 + ], + "angle": 0, + "content": "
MethodNumber (N)ResNet-50Swin-TConvNeXt-T
DDA [17]029.740.044.2
SDA1K31.942.345.5
10K31.942.546.6
50K (default)32.542.547.0
100K32.642.246.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.439, + 0.905, + 0.468 + ], + "angle": 0, + "content": "Table 11. Impact of different numbers of fine-tuning images. Results are evaluated on ImageNet-C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.553 + ], + "angle": 0, + "content": "synthetic domain more accurately, thereby enhancing performance. Based on a balance between performance improvement and image generation resources, we select \\( N = 50\\mathrm{K} \\) as our default experimental setting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.634, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we proposed Synthetic-Domain Alignment (SDA), a novel test-time adaptation (TTA) framework that simultaneously aligns the domains of the source model and target data with the synthetic domain of a diffusion model. For the source model, SDA introduces a Mix of Diffusion (MoD) technique, which generates synthetic data to adapt the source model to a synthetic-domain model. MoD involves a conditional diffusion model for data generation and an unconditional diffusion model for data alignment. For the target data, SDA utilizes the aforementioned unconditional diffusion model to project the target data to synthetic data. As the domains of the model and data are aligned, SDA converts the cross-domain TTA task into an easier in-domain prediction task. Compared to existing diffusion-driven TTA methods, SDA significantly mitigates the source-synthetic domain misalignment issue. Compared to traditional TTA methods, SDA maintains insensitivity to different data streams. Extensive experiments across classifiers, segmenters, and MLLMs indicate that SDA achieves enhanced domain alignment and superior performance." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30510" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.175 + ], + "angle": 0, + "content": "The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant U24B20173." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.19, + 0.188, + 0.207 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[1] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets computer vision: Efficient data generation for urban driving scenes. IJCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.273, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.316, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[3] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification. TMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[5] Goirik Chakrabarty, Manogna Sreenivas, and Soma Biswas. Santa: Source anchoring network and target alignment for continual test time adaptation. TMLR, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.499 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[7] Yuhua Chen, Wen Li, Xiaoran Chen, and Luc Van Gool. Learning semantic segmentation from synthetic data: A geometrically guided input-output adaptation approach. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[8] MMPreTrain Contributors. Openmmlab's pre-training toolbox and benchmark. https://github.com/open-mmlab/mmpretrain, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[11] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[12] Chaoqun Du, Yulin Wang, Jiayi Guo, Yizeng Han, Jie Zhou, and Gao Huang. Unitta: Unified benchmark and versatile framework towards realistic test-time adaptation. arXiv preprint arXiv:2407.20080, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[14] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[15] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.215, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.906, + 0.177 + ], + "angle": 0, + "content": "[17] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[18] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. Note: Robust continual test-time adaptation against temporal correlation. In NeurIPS, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.906, + 0.279 + ], + "angle": 0, + "content": "[19] Jiayi Guo, Chaoqun Du, Jiangshan Wang, Huijuan Huang, Pengfei Wan, and Gao Huang. Assessing a single image in reference-guided image synthesis. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.906, + 0.362 + ], + "angle": 0, + "content": "[20] Jiayi Guo, Hayk Manukyan, Chenyu Yang, Chaofei Wang, Levon Khachatryan, Shant Navasardyan, Shiji Song, Humphrey Shi, and Gao Huang. Faceclip: Facial image-to-video translation via a brief text description. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.366, + 0.906, + 0.422 + ], + "angle": 0, + "content": "[21] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Humphrey Shi, Gao Huang, and Shiji Song. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.906, + 0.478 + ], + "angle": 0, + "content": "[22] Jiayi Guo, Xingqian Xu, Yifan Pu, Zanlin Ni, Chaofei Wang, Manushree Vasu, Shiji Song, Gao Huang, and Humphrey Shi. Smooth diffusion: Crafting smooth latent spaces in diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.481, + 0.906, + 0.521 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.565 + ], + "angle": 0, + "content": "[24] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.906, + 0.596 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.906, + 0.625 + ], + "angle": 0, + "content": "[26] Gao Huang. Dynamic neural networks: advantages and challenges. National Science Review, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.906, + 0.668 + ], + "angle": 0, + "content": "[27] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.906, + 0.711 + ], + "angle": 0, + "content": "[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.906, + 0.742 + ], + "angle": 0, + "content": "[29] Black Forest Labs. Flux. https:// / blackforestlabs.ai/, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[30] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLR Workshops, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[32] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30511" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.482, + 0.164 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.165, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.221, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.264, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[37] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.305, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[38] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[39] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In CoRL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[40] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.461, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[41] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[42] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.56, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[43] Zanlin Ni, Yulin Wang, Renping Zhou, Yizeng Han, Jiayi Guo, Zhiyuan Liu, Yuan Yao, and Gao Huang. Enat: Rethinking spatial-temporal interactions in token-based image synthesis. In NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.617, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[44] Zanlin Ni, Yulin Wang, Renping Zhou, Rui Lu, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Yuan Yao, and Gao Huang. Adanat: Exploring adaptive policy for token-based image generation. In ECCV, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.673, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[45] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.702, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[46] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In ICML, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[47] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.801, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[49] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3d models. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.872, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[50] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-tta:" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Test-time adaptation of discriminative models via generative feedback. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[51] Yifan Pu, Zhuofan Xia, Jiayi Guo, Dongchen Han, Qixiu Li, Duo Li, Yuhui Yuan, Ji Li, Yizeng Han, Shiji Song, et al. Efficient diffusion transformer with step-wise dynamic attention mediators. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[52] Yifan Pu, Yiming Zhao, Zhicong Tang, Ruihong Yin, Haoxing Ye, Yuhui Yuan, Dong Chen, Jianmin Bao, Sirui Zhang, Yanbin Wang, et al. Art: Anonymous region transformer for variable multi-layer transparent image generation. arXiv preprint arXiv:2502.18364, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[54] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[55] Artem Rozantsev, Vincent Lepetit, and Pascal Fua. On rendering synthetic images for training an object detector. Computer Vision and Image Understanding, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[56] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[57] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[58] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[60] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In AAAI, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[61] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[62] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[63] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[64] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozorgtabar. Un-mixing test-time normalization statistics: Combating label temporal correlation. arXiv preprint arXiv:2401.08328, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[65] Yun-Yun Tsai, Fu-Chen Chen, Albert YC Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. CVPR, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30512" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[67] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[69] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. Nerf-supervision: Learning dense object descriptors from neural radiance fields. In ICRA, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[70] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In CVPR, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[71] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30513" + } + ] +] \ No newline at end of file diff --git a/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_origin.pdf b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5026cade517ae79a7ed163e5eb1e4485b02345f9 --- /dev/null +++ b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/fac30b3d-4109-4df5-93e5-c2263ebbddec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6f354bfd1f6632beda45b9a05eac406cf0e2c2607e617b5d2384b0415c89965 +size 3068624 diff --git a/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/full.md b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c53e6087ec047e1cb39da2f5aa46743a94a7d274 --- /dev/null +++ b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/full.md @@ -0,0 +1,381 @@ +# Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment + +Jiayi Guo $^{1,2}$ Junhao Zhao $^{1}$ Chaoqun Du $^{1}$ Yulin Wang $^{1}$ Chunjiang Ge $^{1}$ Zanlin Ni $^{1}$ Shiji Song $^{1}$ Humphrey Shi $^{2*}$ Gao Huang $^{1*}$ $^{1}$ Tsinghua University $^{2}$ SHI Labs @ Georgia Tech + +![](images/db4bdd40794b379802d7fddfa5d56b7684f4c15466580f40ff52b15f7e6ac36c.jpg) +Figure 1. Comparison of different test-time adaptation (TTA) frameworks. (a) Traditional TTA methods continuously adapt source model weights to fit target data batches. However, their performance is sensitive to the amount and order of target data streams, e.g., adapting the model with batches containing data from only a single category can lead to overfitting. (b) Diffusion-driven TTA methods project the target data back to the synthetic domain of diffusion models, which still maintains domain misalignment with the source domain. (c) We propose the Synthetic-domain Alignment (SDA) framework for TTA, which simultaneously aligns the domains of the source model and target data with the same synthetic domain for superior performance. + +![](images/0f5d78e1469ebdce1b05f3eca09ee6086eab47bf020e7be00204b6cfa1b5bf98.jpg) + +# Abstract + +Test-time adaptation (TTA) aims to improve the performance of source-domain pre-trained models on previously unseen, shifted target domains. Traditional TTA methods primarily adapt model weights based on target data streams, making model performance sensitive to the amount and order of target data. The recently proposed diffusion-driven TTA methods mitigate this by adapting model inputs instead of weights, where an unconditional diffusion model, trained on the source domain, transforms target-domain data into a synthetic domain that is expected to approximate the source domain. However, in this paper, we reveal that although the synthetic data in diffusion-driven TTA seems indistinguishable from the source data, it is unaligned with, or even markedly different from the latter for deep networks. To address this issue, we propose a Synthetic-Domain Alignment (SDA) framework. Our key insight is to fine-tune the source model with synthetic data to ensure better alignment. Specifically, we first employ a conditional diffusion model to generate labeled + +samples, creating a synthetic dataset. Subsequently, we use the aforementioned unconditional diffusion model to add noise to and denoise each sample before fine-tuning. This Mix of Diffusion (MoD) process mitigates the potential domain misalignment between the conditional and unconditional models. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLMs, e.g., LLaVA) demonstrate that SDA achieves superior domain alignment and consistently outperforms existing diffusion-driven TTA methods. Our code is available at https://github.com/SHI-Labs/Diffusion-Driven-Test-Time-Adaptation-via-Synthetic-Domain-Alignment. + +# 1. Introduction + +Test-Time Adaptation (TTA) [15, 17, 31, 46, 61, 65, 66, 68, 71] is an emerging research field that tackles domain misalignment when source models are evaluated on shifted target data. Unlike traditional domain adaptation (DA) [16, 37, 57] and source-free adaptation (SFA) [28, 30, 33], TTA addresses more practical scenarios where neither source data nor complete target data are accessible. Instead, the adapta + +
Model Adaptation DirectionData Adaptation Direction
Traditional TTAExpected: Source→Target Risk: Imbalanced Target StreamsN/A
Diffusion-driven TTAN/AExpected: Target→Source Actual: Target→Synthetic
SDA (Ours)Source→SyntheticTarget→Synthetic
+ +Table 1. Adaptation directions of different TTA methods. + +tion relies solely on streaming batches of target data. + +Traditional TTA methods (Fig. 1a) [15, 31, 50, 56, 61, 66, 68, 71] typically employ a source-to-target model adaptation framework. These approaches continuously update the source model weights by processing target data batches. Without annotations, the adaptation process relies either on batch-wise updates of model statistics [31, 56, 66, 68], or unsupervised or self-supervised auxiliary tasks [15, 50, 61]. However, small or imbalanced batches may poorly represent the target domain, making these approaches sensitive to the amount and order of the data stream [12, 17, 66]. For instance, adapting the model with batches containing data from only a single category can lead to overfitting. + +Recently, the impressive generation capabilities of diffusion models [22, 25, 29, 48, 51, 53, 67] have sparked the development of diffusion-driven TTA methods (Fig. 1b) [17, 46, 65], leveraging a target-to-source framework. These approaches employ an unconditional diffusion model, pretrained on the source domain, aiming to project each target sample to the source domain independently. This enables the source model to make predictions without modifying its weights. As a preliminary work, DiffPure [46] addresses adversarial perturbations by first applying a forward diffusion process, introducing a small amount of noise to the target data, followed by a reverse diffusion process to restore a clean image to approach the source domain. Building on this concept, more recent studies [17, 65] tackle challenging domain shifts—such as severe data corruption—by incorporating additional structural guidance from the target data, helping preserve semantics and improve performance. + +In this paper, we uncover that while diffusion-driven TTA methods aim to project target data back to the source domain, the projected target data remains confined within the synthetic domain of the unconditional diffusion model. As the synthetic-domain data are ultimately processed by the source-domain model, this domain misalignment limits the final performance. To address this issue, we propose Synthetic-Domain Alignment (SDA) (Fig. 1c), a new category of framework for TTA tasks which simultaneously aligns the domains of the source model and target data with the same synthetic domain of a diffusion model (Tab. 1). + +SDA distinguishes itself from existing diffusion-driven TTA methods [17, 46, 65] by introducing an additional source-to-synthetic model adaptation phase before testing on adapted target data (Fig. 2). Since the adapted target data aligns with the synthetic domain generated by an un + +![](images/ae6a5cb9f9a020df47ff7598fa7d0efb0fecd225b2904f860a72db1472d2097d.jpg) +Figure 2. Enhanced domain alignment with our framework. Prior diffusion-driven TTA methods struggle with the domain misalignment between the source model and synthetic data, which we resolve by aligning the source model to the synthetic domain. + +conditional diffusion model, SDA aims to adapt the source model to this same synthetic domain. Specifically, SDA employs a Mix-of-Diffusion (MoD) technique to generate synthetic data for model adaptation. Given that the source data is inaccessible after pretraining, MoD first uses a conditional diffusion model to generate samples conditioned on domain-agnostic labels, creating a labeled synthetic dataset. Subsequently, the aforementioned unconditional diffusion model is leveraged by MoD add noise to and denoise these samples, addressing potential domain misalignment between the conditional and unconditional models. With a sufficiently large synthetic dataset, the fine-tuned model becomes highly effective at discriminating within the synthetic domain. Thus, the SDA framework transforms the cross-domain TTA task into an in-domain prediction task by aligning both the source model and target data with the same synthetic domain. + +SDA is a general framework, not limited to specific fine-tuning techniques or diffusion-driven data adaptation methods. This flexibility allows future advancements in these areas to further broaden its applicability. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLM, e.g., LLaVA) demonstrate that SDA consistently outperforms existing methods. Moreover, the effectiveness of our approach is reinforced through visualization analysis and ablation studies. + +# 2. Related Work + +Test-time adaptation (TTA) is an emerging research area that addresses domain shifts by adapting either models [15, 31, 50, 56, 61, 66, 68, 71] or data [17, 46, 65] during evaluation on streaming target data batches. Early model adaptation methods update batch normalization statistics to match the target distribution [31, 56, 66, 71], while others leverage self-supervised tasks like rotation prediction [61] or image restoration [15, 50] to adjust model weights. However, these approaches rely heavily on continuous weight updates, making them sensitive to the amount, order, and diversity of target data. In contrast, diffusion-driven TTA methods [17, 46, 65] focus on data adaptation by projecting each target sample back into the source domain, achiev- + +![](images/f76f39f78b6dd9a714f6089a1acaac65d72590d4ea1936ffbb7ede0673e3b686.jpg) +Forward Diffusion (Add Noise t steps) +(a) Diffusion-driven data adaptation on source data + +![](images/b411876b0eb926ae34ce1345cf1b496951c949f701391aaa467d6f2baa8f1b4f.jpg) +Reverse Diffusion (Denoise $t$ steps) + +![](images/bf422646f3c3df0e2609b0f42bff7dc77d6c9ffd5d040179a2440eed609736c7.jpg) + +![](images/9658ab4994ab010ae6a1feefd933a18dddb4d8263b4a3d554cc2580a18d47772.jpg) + +![](images/6495ec9619f09d6e98a25298f008adf15adac7746c48e5972190c81c4d2eb86e.jpg) +$t = 600$ + +![](images/d135f7544f4ebca577ea4659c9ce5cd22e29611454d05ccf22891acc59179fcc.jpg) + +![](images/fa31cbe96d9044f80a73260319369fb2a08f96325220bf6571a1f1d93161edf4.jpg) +$t = 700$ + +![](images/3260e20ba09c92ec51441b3f25844ece57373735b7959a8ad4fbaeeb31474e1a.jpg) +$t = 800$ +300 $t =$ + +![](images/274cd19e60f21e2719277f5c30eed42ad7b08e2426279831547a19767f246940.jpg) + +![](images/44cf3a6c80a96a086b40e25e25a62c1243aa91371f3421f53fdba6820940dc47.jpg) +00 $t =$ + +![](images/db2b1066657c2e9216f5b9fb14f32d83c52200a872653ac216afeb42d398bee4.jpg) +1000 + +![](images/686858674df97e6cd33cd5264ade44d2bb3ad10798738558a56c35b5c032b0c1.jpg) + +![](images/2dd879d31febf613cff71516778612c563b7a6ea75e51841c48b141d8e4043be.jpg) + +![](images/849e9a314b2c444880088642bac7fd4ec50f8c30857b10afef7fef689f3feed1.jpg) + +![](images/e1f7349857c5cf5a9af7d546584782aa6ad105ed33ad38105942bf3c5a1f9cfd.jpg) + +![](images/362c0480fe2d9bd188d37e569a6ef77d7c47da79a7568c35f7b67ed812d9e0da.jpg) + +![](images/ca30e1e485a0746f17f2c40b784d92643a06585bff81bb0e73a81fa665f420e6.jpg) + +![](images/5e70fe42eed1b45b38083a2844b1abd2687e8daa9b83aeb3fe00ba98c1e8f255.jpg) +(b) Diffusion-driven data adaptation results (on source data) under different timesteps +Figure 3. (a) Illustration of diffusion-driven data adaptation on source data and (b) Adapted images across different timesteps. The results are obtained using DDA [17], with no noticeable visual degradation observed in the adapted images. + +![](images/a7b54d4cea9c478a4527c308180055254e4e9bd67412bbb4316b695ea651b1f1.jpg) + +![](images/d696fe311078fd824df85682bdf04ecef309c7ad6190305ea057ec036dade8fa.jpg) + +![](images/75ee21dd2d895a9ceffbc52e0f6d257916b6d043208ebfca8b13e9b266e32e94.jpg) + +![](images/b69451d7c92deb860cd419f12257cda74cbaf1956ebac352679bc18bbc9465f6.jpg) + +![](images/49651a04855085445a68e6b0df6b8bc9eb9a9726f9b050ef8a24e95a137ea152.jpg) + +![](images/e0b67c0d8ae1f9432ecf1f69a5cfcfac0b7face8dbea36153c01e8f470e897fe.jpg) + +ing stable performance without online model updates. DiffPure [46] purifies adversarial samples with diffusion models, while DDA [17] and GDA [65] use structural guidance to preserve image content under severe corruption. Building on this, our work investigates the issue of misalignment between domains of the source model and synthetic images in diffusion-driven TTA and proposes a new TTA framework. + +Synthetic data for discriminative tasks. Synthetic data, generated by models [20, 21, 42-44, 48, 52, 53] rather than collected from the real world, has shown significant potential in enhancing visual representations for various discriminative tasks [14, 62, 63]. It has been effectively applied in areas such as recognition [3, 63], object detection [49, 55], semantic segmentation [7, 54, 58], image assessment [19], autonomous driving [1], and robotics [39, 69]. In this work, we explore the potential of diffusion-generated synthetic data to enhance domain alignment in TTA. + +# 3. Methodology + +In this section, we introduce the background of diffusion-driven TTA methods and identify their source-synthetic domain misalignment issue in Sec. 3.1 and Sec. 3.2. To tackle this issue, we propose the synthetic-domain alignment (SDA) framework in Sec. 3.3 and introduce its key technique, the mix of diffusion (MoD) in Sec. 3.4. + +# 3.1. Background + +Diffusion process. Given a source data point $\boldsymbol{x}_0^{\mathrm{src}} \sim p_0^{\mathrm{src}}$ , diffusion models $\epsilon_{\theta}$ [25] gradually transform $p_0^{\mathrm{src}}$ into a Gaussian noise distribution $N(\mathbf{0}, \mathbf{I})$ through a $T$ -step forward diffusion process. At each timestep $t \in \{1, 2, \dots, T\}$ , the intermediate state $\boldsymbol{x}_t^{\mathrm{src}} \sim p_t^{\mathrm{src}}$ is computed as: + +$$ +\boldsymbol {x} _ {t} ^ {\mathrm {s r c}} = \sqrt {1 - \beta_ {t}} \boldsymbol {x} _ {t - 1} ^ {\mathrm {s r c}} + \sqrt {\beta_ {t}} \boldsymbol {\epsilon} _ {t}, \tag {1} +$$ + +
Timestep t5006007008009001000
Swin-B (Source-Synthetic, Misaligned)61.657.855.752.645.038.5
Swin-B (Synthetic-Synthetic, Aligned)67.665.965.061.555.748.4
Δ+6.0+8.1+9.3+8.9+10.7+9.9
ConvNeXt-B (Source-Synthetic, Misaligned)65.161.360.357.250.041.5
ConvNeXt-B (Synthetic-Synthetic, Aligned)70.668.267.464.958.550.7
Δ+5.5+6.9+7.1+7.7+8.5+9.2
+ +Table 2. Source model accuracy across different timesteps of diffusion-driven data adaptation. For suitable timesteps for TTA $(t\geq 500)$ [17], model accuracy shows a monotonically decreasing trend with the growth of timestep, indicating the increase of the misalignment of source domain $p_0^{\mathrm{src}}$ and synthetic domain $p_{0,\mathrm{u}}^{\mathrm{syn}}$ . By aligning the source model to the synthetic domain, our methods (rows 3 & 6) significantly help alleviate the performance degradation. Results are evaluated on the ImageNet [9] validation set. + +where $\epsilon_{t}$ is random Gaussian noise and $\beta_{t} \in (0,1)$ represents the diffusion rate at step $t$ . By defining $\alpha_{t} = 1 - \beta_{t}$ , $\overline{\alpha_{t}} = \prod_{t=1}^{T} \alpha_{t}$ and $\epsilon \sim N(\mathbf{0}, \mathbf{I})$ , we obtain: + +$$ +\boldsymbol {x} _ {t} ^ {\mathrm {s r c}} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0} ^ {\mathrm {s r c}} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}. \tag {2} +$$ + +The reverse diffusion process recovers a clean $\widehat{\boldsymbol{x}_0^{\mathrm{src}}}$ by progressively removing noise from $\boldsymbol{x}_T^{\mathrm{src}}$ : + +$$ +\widehat {\boldsymbol {x} _ {t - 1} ^ {\mathrm {s r c}}} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\widehat {\boldsymbol {x} _ {t} ^ {\mathrm {s r c}}} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(\widehat {\boldsymbol {x} _ {t} ^ {\mathrm {s r c}}}, t\right)\right) + \sigma_ {t} \boldsymbol {\epsilon}, \tag {3} +$$ + +where $\sigma_{t}$ is the posterior noise variance [25]. + +Diffusion-driven data adaptation. Denote by $p_0^{\mathrm{trg}}$ the target data distribution, from which each target data point $\boldsymbol{x}_t^{\mathrm{trg}}$ is sampled. Prior work [46] demonstrates that: + +$$ +D _ {\mathrm {K L}} \left(p _ {t + 1} ^ {\mathrm {s r c}} \mid p _ {t + 1} ^ {\mathrm {t r g}}\right) - D _ {\mathrm {K L}} \left(p _ {t} ^ {\mathrm {s r c}} \mid p _ {t} ^ {\mathrm {t r g}}\right) \leq 0, \tag {4} +$$ + +where $D_{\mathrm{KL}}$ is the KL divergence. Since $p_T^{\mathrm{src}} = p_T^{\mathrm{trg}} = N(\mathbf{0}, \mathbf{I})$ , for any arbitrarily small value $\delta$ , there exists a minimum timestep $t^*$ such that $D_{\mathrm{KL}}(p_{t^*}^{\mathrm{src}}||p_{t^*}^{\mathrm{trg}}) < \delta$ . As discussed in [17], a diffusion process with $T = 1000$ requires $t^* \geq 500$ to eliminate domain shifts. We empirically validate this choice of $t^*$ in the supplementary materials. + +The initial diffusion-driven data adaptation in Diff-Pure [46] is carried out as follows: Given the minor divergence between $p_{t^*}^{\mathrm{src}}$ and $p_{t^*}^{\mathrm{trg}}$ , we recover each $\boldsymbol{x}_{t^*}^{\mathrm{trg}} \sim p_{t^*}^{\mathrm{trg}}$ to its corresponding $\widehat{\boldsymbol{x}_0^{\mathrm{src}}} \sim p_0^{\mathrm{src}}$ by executing the reverse diffusion process (Eq. 3) $t^*$ times. + +Since the reverse process is stochastic, subsequent works [17, 65] further introduce additional structure guidance to ensure the content consistency between each $\boldsymbol{x}_0^{\mathrm{trg}}$ and its adapted $\widetilde{\boldsymbol{x}_0^{\mathrm{src}}}$ . In this work, we adopt the same data adaptation process as DDA [17]: + +$$ +\widetilde {\boldsymbol {x} _ {t - 1} ^ {\mathrm {s r c}}} = \widehat {\boldsymbol {x} _ {t - 1} ^ {\mathrm {s r c}}} - \boldsymbol {w} \nabla_ {\widetilde {\boldsymbol {x} _ {t} ^ {\mathrm {s r c}}}} \left\| \phi \left(\boldsymbol {x} _ {0} ^ {\mathrm {t r g}}\right) - \phi \left(\widetilde {\boldsymbol {x} _ {0 , t} ^ {\mathrm {s r c}}}\right) \right\| _ {2}, \tag {5} +$$ + +where $\widehat{\pmb{x}_{t - 1}^{\mathrm{src}}}$ is computed as Eq. 3, $\pmb{w}$ is the structure guidance scale, $\phi$ is a structure extractor [17] and $\widehat{\pmb{x}_{0,t}^{\mathrm{src}}}$ is an esti + +![](images/80b7c7c4458d7d450c98575703dae29f2eeeb6ad025442a57c2e7eab58b13af3.jpg) +Figure 4. Overview of the Synthetic-Domain Alignment (SDA) framework. SDA is a novel TTA framework aligning both the domains of the source model and the target data with the synthetic domain. SDA involves three phases: (left): a source-domain model pretraining phase, where the source model is trained on source data prior to TTA; (middle): a source-to-synthetic model adaptation phase, where the source model is adapted to a synthetic-domain model using synthetic data generated via a Mix of Diffusion (MoD) technique; and (right): a target-to-synthetic data adaptation phase, where target data is adapted into synthetic data using an unconditional diffusion model. Finally, the adapted synthetic data is fed into the synthetic-domain model for test-time inference. + +mate of $\widetilde{\boldsymbol{x}_0^{\mathrm{src}}}$ at timestep $t$ [25]. Diffusion-driven data adaptation is typically performed with unconditional diffusion models [10] since the target data labels are unknown. + +# 3.2. Source-Synthetic Domain Misalignment + +Unlike a real source data point $\boldsymbol{x}_0^{\mathrm{src}}$ , the recovered version $\widetilde{\boldsymbol{x}_{0}^{\mathrm{src}}},$ derived from $\boldsymbol{x}_{t^{*}}^{\mathrm{trg}}$ is synthetic. Specifically, $\boldsymbol{x}_0^{\mathrm{src}}$ follows the source domain $p_0^{\mathrm{src}}$ , whereas $\widetilde{\boldsymbol{x}_0^{\mathrm{src}}}$ follows the synthetic domain $p_{0,\mathrm{u}}^{\mathrm{syn}}$ of an unconditional diffusion model $\epsilon_{\theta}^{\mathrm{u}}$ with parameters $\theta$ . In this section, we empirically reveal the misalignment between the source domain $p_0^{\mathrm{src}}$ and the synthetic domain $p_{0,\mathrm{u}}^{\mathrm{syn}}$ , and investigate how this misalignment impacts the performance of the source model. For simplicity and clarity, we will substitute $\widetilde{\boldsymbol{x}_0^{\mathrm{src}}}$ with $\boldsymbol{x}_{0,\mathrm{u}}^{\mathrm{syn}}$ throughout the following discussion. + +Based on the above analysis, diffusion-driven TTA methods face two potential misalignments: (1) source-target domain misalignment arising from inherent data distribution shifts, which has been the primary focus of prior research, and (2) source-synthetic domain misalignment introduced by diffusion models, which we address as the main focus of this work, complementing existing efforts. + +To precisely examine the impact of the source-synthetic domain misalignment and isolate it from the influence of the source-target domain misalignment, we evaluate the performance of the source model $f$ on synthetic data $x_{0,\mathrm{u}}^{\mathrm{syn}}$ adapted by the diffusion model $\epsilon_{\theta}^{\mathrm{u}}$ from source data $x_0^{\mathrm{src}}$ (Fig. 3a). Specifically, we test ImageNet pretrained models [9] on the ImageNet validation set adapted by the popular diffusion + +driven data adaptation method, DDA [17] across different timesteps $t$ . As indicated in Tab. 2, with an increase in $t$ , the accuracy of the model exhibits a monotonically decreasing trend. For ideal $t^* \geq 500$ for TTA tasks, performance degradation of more than $18.8\%$ for ConvNeXt [36] and $21.8\%$ for Swin [35] is observed compared to their official results [8] on source data which are $83.4\%$ and $83.9\%$ . This indicates a significant domain misalignment between $p_0^{\mathrm{src}}$ and $p_{0,\mathrm{u}}^{\mathrm{syn}}$ . By fine-tuning the source model on the synthetic data generated by our Mix of Diffusion process (introduced in Sec. 3.4), the aligned models $f'$ achieve performance improvements of $5.5\%$ for ConvNeXt and $6.0\%$ for Swin. + +In Fig. 3b, we show that the diffusion synthetic data $\boldsymbol{x}_{0,\mathrm{u}}^{\mathrm{syn}}$ and source data $\boldsymbol{x}_0^{\mathrm{src}}$ exhibit no noticeable visual differences across different timesteps $t$ . This further suggests that the performance degradation is not due to the quality of diffusion-generated images but rather the implicit misalignment between the source and synthetic domains. + +# 3.3. Synthetic-Domain Alignment Framework + +Given that the diffusion-adapted data $\boldsymbol{x}_{0,\mathrm{u}}^{\mathrm{syn}}$ aligns more closely with the synthetic domain $p_{0,\mathrm{u}}^{\mathrm{syn}}$ instead of the source domain $p_0^{\mathrm{src}}$ , we propose simultaneously adapting the source model $f$ to the same synthetic domain $p_0^{\mathrm{syn}}$ . By doing so, the alignment between the data and model within $p_0^{\mathrm{syn}}$ can be effectively achieved. To this end, we introduce a novel TTA framework: Synthetic-Domain Alignment (SDA). + +In Fig. 4, we present the complete diagram of SDA, which consists of three key phases: (1) a source-domain + +model pretraining phase, where the source model $f$ is trained on source data $\pmb{x}_0^{\mathrm{src}}$ prior to TTA; (2) a source-to-synthetic model adaptation phase, where $f$ is fine-tuned to a synthetic-domain model $f'$ ; and (3) a target-to-synthetic data adaptation phase, where target data $\pmb{x}_0^{\mathrm{trg}}$ is adapted into synthetic data $\pmb{x}_{0,\mathrm{u}}^{\mathrm{syn}}$ using an unconditional diffusion model $\epsilon_{\theta}^{\mathrm{u}}$ following Eq. 5. Finally, the adapted synthetic data $\pmb{x}_{0,\mathrm{u}}^{\mathrm{syn}}$ is fed into the synthetic-domain model $f'$ for test-time inference. Consistent with the standard TTA protocol [66], the source data is accessible only during the source model pretraining phase and remains inaccessible during both the model and data adaptation phases. + +The rationale behind SDA is straightforward: by aligning the model and data domains to the same synthetic domain, the original cross-domain TTA task is transformed into an easier in-domain prediction task, thus addressing the core challenge of TTA and improving performance. + +Since the pretraining phase follows standard supervised learning, and the data adaptation phase aligns with DDA [17], we omit detailed explanations of these phases. Instead, we focus on how we adapt the source model $f$ to an ideal synthetic domain model $f'$ using a novel technique called Mix-of-Diffusion (MoD), introduced next. + +# 3.4. Model Adaptation via Mix of Diffusion + +As shown in Fig. 4, MoD consists of two main processes: a data generation process powered by a conditional diffusion model $\epsilon_{\eta}^{\mathrm{c}}$ with parameters $\eta$ , and a data alignment process powered by the same unconditional diffusion model $\epsilon_{\theta}^{\mathrm{u}}$ used in the target-to-synthetic data adaptation phase. It is worth noting that, both $\epsilon_{\eta}^{\mathrm{c}}$ and $\epsilon_{\theta}^{\mathrm{u}}$ are also pretrained on source data $\mathbf{x}_0^{\mathrm{src}}$ and have never been exposed to target data $\mathbf{x}_0^{\mathrm{trg}}$ , in accordance with the TTA setting. + +Conditional diffusion data generation. The generation process leverages the conditional generation capability of $\epsilon_{\eta}^{\mathrm{c}}$ to synthesize conditional synthetic data $x_{0,c}^{\mathrm{syn}} \sim p_{0,c}^{\mathrm{syn}}$ . In the context of TTA, the source and target domains share the same domain-agnostic label set $\{y_i\}_{i=1}^{K}$ . Utilizing $\epsilon_{\eta}^{\mathrm{c}}$ , we uniformly generate samples for each class $y_i$ from Gaussian noise $(x_{T,c}^{\mathrm{syn}})$ through a $T$ -step reverse diffusion process: + +$$ +\boldsymbol {x} _ {t - 1, c} ^ {\text {s y n}} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {x} _ {t, c} ^ {\text {s y n}} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \overline {{\alpha_ {t}}}}} \epsilon_ {\eta} ^ {\mathrm {c}} \left(\boldsymbol {x} _ {t, c} ^ {\text {s y n}}, t, y _ {i}\right)\right) + \sigma_ {t} \boldsymbol {\epsilon}. \tag {6} +$$ + +The generation capability of $\epsilon_{\eta}^{\mathrm{c}}$ allows for the construction of a labeled synthetic-domain dataset $\{x_{0,\mathrm{c}}^{\mathrm{syn}},y\}^{N}$ of arbitrary size $N$ without any manual data collection. By fine-tuning the source model $f$ on this synthetic dataset, an adapted model $f_{\mathrm{c}}^{\prime}$ on domain $p_{0,\mathrm{c}}^{\mathrm{syn}}$ can be obtained. + +Unconditional diffusion data alignment. However, since the test-time adapted data $x_{0,\mathrm{u}}^{\mathrm{syn}}$ is subject to domain $p_{0,\mathrm{u}}^{\mathrm{syn}}$ , we argue that there is still potential misalignment between different synthetic domains $p_{0,\mathrm{c}}^{\mathrm{syn}}$ and $p_{0,\mathrm{u}}^{\mathrm{syn}}$ . This is mainly + +because of the differences in architectures and training schemes between $\epsilon_{\eta}^{\mathrm{c}}$ and $\epsilon_{\theta}^{\mathrm{u}}$ . We empirically validate this argument in Tab. 10, which indicates that the model $f_{\mathrm{c}}'$ adapted to domain $p_{0,\mathrm{c}}^{\mathrm{syn}}$ performs worse than the model $f_{\mathrm{u}}'$ adapted to domain $p_{0,\mathrm{u}}^{\mathrm{syn}}$ during testing. + +To obtain the $f_{\mathrm{u}}^{\prime}$ on domain $p_{0,\mathrm{u}}^{\mathrm{syn}}$ we mirror the target-to-synthetic data adaptation phase as a conditional synthetic data $(\boldsymbol{x}_{0,\mathrm{c}}^{\mathrm{syn}})$ to unconditional synthetic data $(\boldsymbol{x}_{0,\mathrm{u}}^{\mathrm{syn}})$ adaptation process. In specific, we use the same $t^{*}$ in Sec. 3.1 with Eq. 2 to obtain $\boldsymbol{x}_{t^{*},\mathrm{c}}^{\mathrm{syn}}$ . According to the analysis in Sec. 3.1, $\boldsymbol{x}_{t^{*},\mathrm{c}}^{\mathrm{syn}}$ will be indistinguishable to its counterpart $\boldsymbol{x}_{t^{*},\mathrm{u}}^{\mathrm{syn}}$ on domain $p_{t^{*},\mathrm{u}}^{\mathrm{syn}}$ . Therefore, we have: + +$$ +\boldsymbol {x} _ {t ^ {*}, \mathrm {u}} ^ {\text {s y n}} \approx \boldsymbol {x} _ {t ^ {*}, \mathrm {c}} ^ {\text {s y n}} = \sqrt {\overline {{\alpha_ {t}}}} \boldsymbol {x} _ {0, c} ^ {\text {s y n}} + \sqrt {1 - \overline {{\alpha_ {t}}}} \boldsymbol {\epsilon}. \tag {7} +$$ + +Then, following Eq. 3 and Eq. 5, the noisy $\boldsymbol{x}_{t^{*},\mathrm{u}}^{\mathrm{syn}}$ can be gradually denoised to $\boldsymbol{x}_{0,\mathrm{u}}^{\mathrm{syn}}$ . Finally, the expected synthetic-domain model $f_{\mathrm{u}}^{\prime}$ is obtained by fine-tuning the source-domain model $f_{\mathrm{u}}^{\prime}$ on dataset $\{\boldsymbol{x}_{0,\mathrm{u}}^{\mathrm{syn}},y\}^{N}$ . + +Ensembling. As noted in previous diffusion-driven TTA methods [17], while diffusion models generally perform well for data adaptation, they may occasionally produce data points that are less recognizable than the original target data. To address this, prior approaches use an ensemble of model predictions on $\boldsymbol{x}_{0}^{\mathrm{trg}}$ and $\boldsymbol{x}_{0,\mathrm{u}}^{\mathrm{syn}}$ as the final output. Following this protocol, the final prediction in SDA is: + +$$ +\hat {y} = \arg \max _ {y} \left(q \left(y \mid \boldsymbol {x} _ {0} ^ {\operatorname {t r g}}\right) + q ^ {\prime} \left(y \mid \boldsymbol {x} _ {0, \mathrm {u}} ^ {\operatorname {s y n}}\right)\right), \tag {8} +$$ + +where $q(\cdot)$ and $q^{\prime}(\cdot)$ are output distributions of source model $f$ and synthetic-domain model $f_{\mathrm{u}}^{\prime}$ , respectively. + +# 4. Experiments + +In this section, we first evaluate SDA on ImageNet classifiers with standard TTA benchmarks in Sec. 4.1. Next, we assess SDA's scalability across different dataset sizes, tasks, and model architectures in Sec. 4.2. In Sec. 4.3, we demonstrate SDA's advantages through Grad-CAM visualizations [59] and data stream sensitivity tests. Finally, ablation studies in Sec. 4.4 validate the design choices in our SDA framework. + +
ModelSourceMEMODiffPureGDADDASDA (Ours)
ResNet-5018.724.716.831.829.732.5 (+2.8)
Swin-T33.529.524.842.240.042.5 (+2.5)
ConvNeXt-T39.337.828.844.844.247.0 (+2.8)
Swin-B40.537.028.9-44.547.4 (+2.9)
ConvNeXt-B45.645.832.7-49.451.9 (+2.5)
+ +Table 3. Comparison results on ImageNet-C. We compare SDA with source models, MEMO [71], DiffPure [46], GDA [65] and DDA [17]. Results are the average accuracy across 15 adaptation domains at severity level 5. SDA shows consistent performance improvements compared to baselines. + +
GaussianShotImpluseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source48.047.148.332.414.439.339.255.750.151.575.152.029.548.160.346.1
DiffPure42.642.040.916.822.219.123.039.029.19.162.53.034.346.860.732.7
DDA59.955.755.930.730.538.139.253.248.041.971.048.146.959.562.049.4
SDA (Ours)60.357.658.135.435.342.942.455.549.944.672.645.450.263.064.551.9 (+2.5)
+ +# 4.1. Main Results on ImageNet Classifiers + +Settings. We choose DDA [17] as our primary competitor since it is the best open-sourced method. SDA is also compared with DiffPure [46], MEMO [71], and the recent SOTA GDA [65] using their reported results. Source model performance is reported as "Source". DiT [48] and ADM [10] are adopted to generate and align 50K synthetic data for 15-epoch finetuning. For different source models, the synthetic data only needs to be generated once. Our results are tested on standard TTA benchmarks, ImageNet-C [24] (severity level 5) and ImageNet-W [32] using various models including ResNet [23], Swin [35] and ConvNeXt [36]. More implementation details are provided in the supplementary materials. + +Comparison results on ImageNet-C. We begin by evaluating the performance of SDA on ImageNet-C. As reported in Tab. 3, our proposed SDA consistently outperforms all baseline methods across different model architectures and sizes. We emphasize the performance improvement over DDA, as we adopted DDA for target data adaptation in SDA. Compared to DDA, our SDA improves accuracy by $2.5\% - 2.9\%$ . This significant improvement indicates the misalignment between the source and synthetic domains, validating the effectiveness of our synthetic-domain alignment framework. Moreover, compared to the recent SOTA GDA, SDA also achieves an improvement of $2.2\%$ with ConvNeXt-T. Notably, SDA focuses on synthetic domain alignment, an orthogonal research direction to existing diffusion-driven methods on better adapting the target data. Therefore, the performance of SDA could potentially be further enhanced with the release of more advanced codebases like GDA. Compared to the model adaptation method, MEMO, three diffusion-driven methods (SDA, DDA, and GDA) all demonstrate superior performance, highlighting the effectiveness of diffusion models in assisting TTA tasks. DiffPure presents worse results since it is primarily designed for adversarial attacks. Without the structural guidance introduced in DDA and GDA, DiffPure may not effectively recover images with severe domain shifts. In Tab. 4, we provide a detailed comparison of the results of SDA and baselines. SDA surpasses DiffPure in all 15 adaptation domains and outperforms DDA in 14 out of 15 domains, fur + +Table 4. Detailed comparisons of SDA and baselines across 15 adaptation domains of ImageNet-C. SDA shows the best average accuracy. The results are tested with ConvNeXt-B. Comparisons using other models are deferred to the supplementary materials. + +
ModelSourceDiffPureDDASDA (Ours)
ResNet-5037.729.152.854.7 (+1.9)
Swin-T66.552.765.967.3 (+1.4)
ConvNeXt-T67.655.867.969.4 (+1.5)
Swin-B69.155.568.370.6 (+2.3)
ConvNeXt-B70.157.770.372.3 (+2.0)
+ +Table 5. Comparison results on ImageNet-W. SDA shows consistent performance improvements compared to baselines. + +ther affirming the superiority of SDA. + +Comparison results on ImageNet-W. We extend our evaluation to ImageNet-W to assess SDA's performance under watermark-based domain shifts. As shown in Tab. 5, SDA consistently surpasses all baselines across different models. Compared to our primary baseline, DDA, SDA achieves accuracy gains ranging from $1.4\%$ to $2.3\%$ . Furthermore, the results in Tab. 5 reveal potential performance drops when DDA is applied to ImageNet-W with Swin-T and Swin-B models, suggesting that synthetic data may be less recognizable by the original source models. Although advancements in diffusion techniques could potentially improve outcomes, the consistent gains achieved by SDA indicate that aligning the source model with the synthetic domain offers a convenient and effective solution to enhance performance. + +# 4.2. Scalability to Other Datasets, Tasks and Models + +In addition to standard benchmarks, an effective TTA method should demonstrate its superiority across various dataset scales, task formats, and model architectures. In this section, we assess SDA from these aspects to validate its scalability. Implementation details of each experiment can be found in the supplementary materials. + +Scaling to small datasets. We first evaluate the effectiveness of SDA on scenarios where both source and target domains are small-scale. Specifically, we test SDA on CIFAR-10-C [24] with ResNet-18 [23] as the source classifier. EDM [27] is used to generate synthetic data, and I-DDPM [45] is applied for data alignment. As shown in Tab. 6, SDA consistently outperforms both the source model and DDA, achieving an average accuracy improvement of $7.1\%$ over DDA. + +
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source23.528.720.841.230.748.248.151.863.663.479.528.056.535.960.745.4
DDA79.480.376.841.356.649.048.977.473.562.680.561.561.257.673.265.3
SDA (Ours)82.683.579.559.661.465.364.676.578.471.588.953.475.464.480.772.4 (+7.1)
+ +Table 6. Comparisons of SDA and baselines with ResNet-18 on CIFAR-10-C. SDA shows the best average accuracy. + +
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPirelateJEPGAvg
Source9.813.210.022.75.720.422.136.728.453.966.69.724.316.043.825.6
DDA16.621.116.045.040.949.738.736.738.635.055.213.356.656.358.838.6
SDA (Ours)17.021.116.245.543.051.139.938.740.236.357.913.159.158.159.739.8 (+1.2)
+ +Table 7. Comparisons of SDA and baselines with DeepLabv3 on PASCAL VOC-C. SDA shows the best average mIOU. + +Scaling to Semantic Segmentation Tasks. We extend our evaluation to dense prediction tasks by using PASCAL-VOC-C [13] as a standard semantic segmentation benchmark with DeepLabv3 [6] as the source segmenter. Dataset Diffusion [41] generates synthetic segmentation data, and FLUX Schnell [29] is used for data alignment. As shown in Tab. 7, SDA achieves the best performance, with an average mIOU improvement of $1.2\%$ over DDA. + +Scaling to multimodal large language models (MLLMs). As an emerging research direction, MLLMs [26, 34] present advanced visual question answering [2] capability. We design a language-based classification task format to test how SDA can help MLLMs on TTA tasks on ImageNet-C: Given an image, ask an MLLM (LLaVA 1.5-7b [34] in our experiments) to choose the correct image class from four provided options. In Tab. 8, the "Source-Zero" setting tests the zero-shot results of pretrained LLaVA. The "Source" and "DDA" settings evaluate a source-data fine-tuned LLaVA while the "SDA" setting tests a synthetic-data fine-tuned LLaVA. The fine-tuning task format is the same as that at test time. Pre-trained LLaVA already exhibits strong performance. While DDA improves results in some domains, it does not yield an overall gain compared to source-data fine-tuned LLaVA. In contrast, SDA aligns LLaVA with the synthetic domain, achieving the best accuracy with an improvement of $2.4\%$ . + +# 4.3. Analysis + +Visualization. To demonstrate how synthetic data fine-tuning in SDA enhances the performance of diffusion-driven TTA methods, we employ Gradient-weighted Class Activation Mapping (Grad-CAM) [59] to visualize the image regions that most influence classification scores across different images and models. As shown in Fig. 5, testing target images with the source model reveals distinct differences in activation maps and the occurrence of incorrect predictions compared to those from source images, + +![](images/8d9145f445f0dc1c5dc179cde4f7eac7b94f6c5142bae6001f4f5fa5d6454627.jpg) +Figure 5. Grad-CAM visualization comparison. The first row shows activation maps for source and target images tested with the source model. The second row displays activation maps for diffusion synthetic images tested with the source model (DDA) and our synthetic-domain model (SDA). SDA aligns closely with the source model's response to source images. + +underscoring the performance degradation due to domain shifts. Despite using adapted synthetic images, DDA still risks focusing on inappropriate regions and producing incorrect predictions. This highlights the domain misalignment of the synthetic data and source model. In contrast, SDA aligns both the data and model within the same synthetic domain, thereby producing activation maps and predictions that closely resemble those produced by the source model on source images. + +Data stream sensitivity. In Tab. 9, we test data stream sensitivity using the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings closely aligned with diverse real-world TTA scenarios. In addition to diffusion-driven TTA methods, we include 13 popular traditional TTA methods [4, 5, 11, 12, 18, 38, 40, 47, 60, 66, 68, 70] for comparison. We report the average accuracy of 12 settings using ResNet-50. The results indicate that diffusion-driven TTA methods are preferred in these challenging settings since they are insensitive to different variants of data streams. SDA maintains this insensitivity and showcases the best performance. Detailed comparisons for each setting are provided in the supplementary materials. + +
GaussianShotImpluseDerocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAve:
Source-Zero75.276.679.681.474.583.581.187.088.989.595.684.176.290.188.783.5
Source86.886.088.690.185.593.589.793.695.094.998.189.585.696.294.891.2
DDA92.491.592.588.690.192.288.593.093.189.697.875.889.696.497.091.2
SDA (Ours)94.994.194.691.092.994.091.593.695.292.298.583.692.897.397.693.6 (+2.4)
+ +Table 8. Comparisons of SDA and baselines with LLaVA on ImageNet-C. SDA shows the best average accuracy. + +
SourceTENT [66]ROID [38]NOTE [18]CoTIA [68]TRIBE [60]BN [40]EATA [67]SANTA [5]RMT [11]UnMIX [64]RoTTA [70]LAME [14]UniTTA [12]DDA [17]SDA (Ours)
18.34.26.26.38.89.39.810.210.611.517.423.124.930.329.331.6 (+2.3)
+ +Table 9. Data stream sensitivity comparison. We additionally compare SDA with 13 traditional TTA methods on the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings. Results are reported as average accuracy across settings. + +
ComponentResNet-50Swin-TConvNeXt-T
DDA [17]29.740.044.2
+ Conditional Data Generation30.440.144.6
+ Unconditional Data Alignment32.542.547.0
+ +# 4.4. Ablation Studies + +Components. We examine the impact of two key components in our SDA framework's Mix of Diffusion technique in Tab. 10: (1) synthetic data generation using the conditional diffusion model and (2) synthetic data alignment using the unconditional diffusion model. Fine-tuning source models with synthetic data generated solely by the conditional diffusion model (+ Conditional Data Generation) yields only marginal gains. As discussed in Sec. 3.4, this limited gain arises from the domain misalignment between conditional and unconditional diffusion models. Specifically, using only conditional synthetic data results in models aligned with the conditional diffusion domain, whereas the test data belongs to the unconditional diffusion domain. Therefore, further aligning the synthetic data through the unconditional diffusion model (+ Unconditional Data Alignment) leads to significant performance gains over the baseline DDA. This demonstrates that bridging the misalignment of different diffusion domains is essential for the success of our SDA framework. + +Number of fine-tuning images. We examine the impact of different numbers of images $(N)$ used during synthetic-domain model fine-tuning in Tab. 11. Interestingly, even with only one image per class $(N = 1\mathrm{K})$ , SDA still significantly outperforms DDA. This finding suggests a key attribute of the fine-tuning process: source models are primarily learning to adapt to the synthetic domain itself, rather than acquiring class-specific knowledge. Increasing the number of images helps the fine-tuning process capture the + +Table 10. Impact of different components in SDA. Results are evaluated on ImageNet-C. + +
MethodNumber (N)ResNet-50Swin-TConvNeXt-T
DDA [17]029.740.044.2
SDA1K31.942.345.5
10K31.942.546.6
50K (default)32.542.547.0
100K32.642.246.8
+ +Table 11. Impact of different numbers of fine-tuning images. Results are evaluated on ImageNet-C. + +synthetic domain more accurately, thereby enhancing performance. Based on a balance between performance improvement and image generation resources, we select $N = 50\mathrm{K}$ as our default experimental setting. + +# 5. Conclusion + +In this paper, we proposed Synthetic-Domain Alignment (SDA), a novel test-time adaptation (TTA) framework that simultaneously aligns the domains of the source model and target data with the synthetic domain of a diffusion model. For the source model, SDA introduces a Mix of Diffusion (MoD) technique, which generates synthetic data to adapt the source model to a synthetic-domain model. MoD involves a conditional diffusion model for data generation and an unconditional diffusion model for data alignment. For the target data, SDA utilizes the aforementioned unconditional diffusion model to project the target data to synthetic data. As the domains of the model and data are aligned, SDA converts the cross-domain TTA task into an easier in-domain prediction task. Compared to existing diffusion-driven TTA methods, SDA significantly mitigates the source-synthetic domain misalignment issue. Compared to traditional TTA methods, SDA maintains insensitivity to different data streams. Extensive experiments across classifiers, segmenters, and MLLMs indicate that SDA achieves enhanced domain alignment and superior performance. + +# Acknowledgements + +The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant U24B20173. + +# References + +[1] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets computer vision: Efficient data generation for urban driving scenes. IJCV, 2018. 3 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 7 +[3] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification. TMLR, 2023. 3 +[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, 2022. 7, 8 +[5] Goirik Chakrabarty, Manogna Sreenivas, and Soma Biswas. Santa: Source anchoring network and target alignment for continual test time adaptation. TMLR, 2023. 7, 8 +[6] Liang-Chieh Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7 +[7] Yuhua Chen, Wen Li, Xiaoran Chen, and Luc Van Gool. Learning semantic segmentation from synthetic data: A geometrically guided input-output adaptation approach. In CVPR, 2019. 3 +[8] MMPreTrain Contributors. Openmmlab's pre-training toolbox and benchmark. https://github.com/open-mmlab/mmpretrain, 2023.4 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3, 4 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4, 6 +[11] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, 2023. 7, 8 +[12] Chaoqun Du, Yulin Wang, Jiayi Guo, Yizeng Han, Jie Zhou, and Gao Huang. Unitta: Unified benchmark and versatile framework towards realistic test-time adaptation. arXiv preprint arXiv:2407.20080, 2024. 2, 7, 8 +[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7 +[14] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 3 +[15] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 1, 2 + +[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 1 +[17] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8 +[18] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. Note: Robust continual test-time adaptation against temporal correlation. In NeurIPS, 2022. 7, 8 +[19] Jiayi Guo, Chaoqun Du, Jiangshan Wang, Huijuan Huang, Pengfei Wan, and Gao Huang. Assessing a single image in reference-guided image synthesis. In AAAI, 2022. 3 +[20] Jiayi Guo, Hayk Manukyan, Chenyu Yang, Chaofei Wang, Levon Khachatryan, Shant Navasardyan, Shiji Song, Humphrey Shi, and Gao Huang. Faceclip: Facial image-to-video translation via a brief text description. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3 +[21] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Humphrey Shi, Gao Huang, and Shiji Song. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, 2023. 3 +[22] Jiayi Guo, Xingqian Xu, Yifan Pu, Zanlin Ni, Chaofei Wang, Manushree Vasu, Shiji Song, Gao Huang, and Humphrey Shi. Smooth diffusion: Crafting smooth latent spaces in diffusion models. In CVPR, 2024. 2 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[24] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 6 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4 +[26] Gao Huang. Dynamic neural networks: advantages and challenges. National Science Review, 2024. 7 +[27] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 6 +[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 1 +[29] Black Forest Labs. Flux. https:// / blackforestlabs.ai/, 2024. 2, 7 +[30] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1 +[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLR Workshops, 2017. 1, 2 +[32] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, 2023. 6 + +[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 7 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4, 6 +[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 6 +[37] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 1 +[38] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, 2024. 7, 8 +[39] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In CoRL, 2022. 3 +[40] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 7, 8 +[41] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2023. 7 +[42] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 3 +[43] Zanlin Ni, Yulin Wang, Renping Zhou, Yizeng Han, Jiayi Guo, Zhiyuan Liu, Yuan Yao, and Gao Huang. Enat: Rethinking spatial-temporal interactions in token-based image synthesis. In NeurIPS, 2024. +[44] Zanlin Ni, Yulin Wang, Renping Zhou, Rui Lu, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Yuan Yao, and Gao Huang. Adanat: Exploring adaptive policy for token-based image generation. In ECCV, 2024.3 +[45] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 6 +[46] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In ICML, 2022. 1, 2, 3, 5, 6 +[47] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 7, 8 +[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 6 +[49] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3d models. In ICCV, 2015. 3 +[50] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-tta: + +Test-time adaptation of discriminative models via generative feedback. In NeurIPS, 2023. 2 +[51] Yifan Pu, Zhuofan Xia, Jiayi Guo, Dongchen Han, Qixiu Li, Duo Li, Yuhui Yuan, Ji Li, Yizeng Han, Shiji Song, et al. Efficient diffusion transformer with step-wise dynamic attention mediators. In ECCV, 2024. 2 +[52] Yifan Pu, Yiming Zhao, Zhicong Tang, Ruihong Yin, Haoxing Ye, Yuhui Yuan, Dong Chen, Jianmin Bao, Sirui Zhang, Yanbin Wang, et al. Art: Anonymous region transformer for variable multi-layer transparent image generation. arXiv preprint arXiv:2502.18364, 2025. 3 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3 +[54] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 3 +[55] Artem Rozantsev, Vincent Lepetit, and Pascal Fua. On rendering synthetic images for training an object detector. Computer Vision and Image Understanding, 2015. 3 +[56] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In ICLR, 2020. 2 +[57] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1 +[58] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In CVPR, 2018. 3 +[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5, 7 +[60] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In AAAI, 2024. 7, 8 +[61] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 1, 2 +[62] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In CVPR, 2024. 3 +[63] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. In NeurIPS, 2024. 3 +[64] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozorgtabar. Un-mixing test-time normalization statistics: Combating label temporal correlation. arXiv preprint arXiv:2401.08328, 2024. 8 +[65] Yun-Yun Tsai, Fu-Chen Chen, Albert YC Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. CVPR, 2024. 1, 2, 3, 5, 6 + +[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2, 5, 7, 8 +[67] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2 +[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 1, 2, 7, 8 +[69] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. Nerf-supervision: Learning dense object descriptors from neural radiance fields. In ICRA, 2022. 3 +[70] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In CVPR, 2023. 7, 8 +[71] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 1, 2, 5, 6 \ No newline at end of file diff --git a/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/images.zip b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e11d7edf1f5a65b414aebddb34e5d173d69b0c43 --- /dev/null +++ b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:750b54f19107f1b46a2396651e3c0060957858576f29de4a56097399ed4b3260 +size 678047 diff --git a/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/layout.json b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..972940e3f65b0d9ee20edfde2d75f305541bcae6 --- /dev/null +++ b/2025/Everything to the Synthetic_ Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment/layout.json @@ -0,0 +1,11820 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 140 + ], + "type": "text", + "content": "Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "spans": [ + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": "Jiayi Guo" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Junhao Zhao" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Chaoqun Du" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Yulin Wang" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Chunjiang Ge" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Zanlin Ni" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Shiji Song" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Humphrey Shi" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": " Gao Huang" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 159, + 515, + 204 + ], + "type": "text", + "content": "SHI Labs @ Georgia Tech" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 224, + 298, + 361 + ], + "blocks": [ + { + "bbox": [ + 56, + 224, + 298, + 361 + ], + "lines": [ + { + "bbox": [ + 56, + 224, + 298, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 298, + 361 + ], + "type": "image", + "image_path": "db4bdd40794b379802d7fddfa5d56b7684f4c15466580f40ff52b15f7e6ac36c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 370, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 54, + 370, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. Comparison of different test-time adaptation (TTA) frameworks. (a) Traditional TTA methods continuously adapt source model weights to fit target data batches. However, their performance is sensitive to the amount and order of target data streams, e.g., adapting the model with batches containing data from only a single category can lead to overfitting. (b) Diffusion-driven TTA methods project the target data back to the synthetic domain of diffusion models, which still maintains domain misalignment with the source domain. (c) We propose the Synthetic-domain Alignment (SDA) framework for TTA, which simultaneously aligns the domains of the source model and target data with the same synthetic domain for superior performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 224, + 555, + 361 + ], + "blocks": [ + { + "bbox": [ + 304, + 224, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 304, + 224, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 555, + 361 + ], + "type": "image", + "image_path": "0f5d78e1469ebdce1b05f3eca09ee6086eab47bf020e7be00204b6cfa1b5bf98.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 447, + 200, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 447, + 200, + 459 + ], + "spans": [ + { + "bbox": [ + 151, + 447, + 200, + 459 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 468, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 468, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 468, + 297, + 696 + ], + "type": "text", + "content": "Test-time adaptation (TTA) aims to improve the performance of source-domain pre-trained models on previously unseen, shifted target domains. Traditional TTA methods primarily adapt model weights based on target data streams, making model performance sensitive to the amount and order of target data. The recently proposed diffusion-driven TTA methods mitigate this by adapting model inputs instead of weights, where an unconditional diffusion model, trained on the source domain, transforms target-domain data into a synthetic domain that is expected to approximate the source domain. However, in this paper, we reveal that although the synthetic data in diffusion-driven TTA seems indistinguishable from the source data, it is unaligned with, or even markedly different from the latter for deep networks. To address this issue, we propose a Synthetic-Domain Alignment (SDA) framework. Our key insight is to fine-tune the source model with synthetic data to ensure better alignment. Specifically, we first employ a conditional diffusion model to generate labeled" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 594 + ], + "type": "text", + "content": "samples, creating a synthetic dataset. Subsequently, we use the aforementioned unconditional diffusion model to add noise to and denoise each sample before fine-tuning. This Mix of Diffusion (MoD) process mitigates the potential domain misalignment between the conditional and unconditional models. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLMs, e.g., LLaVA) demonstrate that SDA achieves superior domain alignment and consistently outperforms existing diffusion-driven TTA methods. Our code is available at https://github.com/SHI-Labs/Diffusion-Driven-Test-Time-Adaptation-via-Synthetic-Domain-Alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 631, + 556, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 556, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 556, + 717 + ], + "type": "text", + "content": "Test-Time Adaptation (TTA) [15, 17, 31, 46, 61, 65, 66, 68, 71] is an emerging research field that tackles domain misalignment when source models are evaluated on shifted target data. Unlike traditional domain adaptation (DA) [16, 37, 57] and source-free adaptation (SFA) [28, 30, 33], TTA addresses more practical scenarios where neither source data nor complete target data are accessible. Instead, the adapta" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 704, + 148, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 148, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 148, + 714 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30503" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 135 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 135 + ], + "type": "table", + "html": "
Model Adaptation DirectionData Adaptation Direction
Traditional TTAExpected: Source→Target Risk: Imbalanced Target StreamsN/A
Diffusion-driven TTAN/AExpected: Target→Source Actual: Target→Synthetic
SDA (Ours)Source→SyntheticTarget→Synthetic
", + "image_path": "ac3c784ca0d59420a48b11792a781b7fc2e80d4ee86b7bc46eb41f7ea74b4b57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 140, + 285, + 151 + ], + "lines": [ + { + "bbox": [ + 65, + 140, + 285, + 151 + ], + "spans": [ + { + "bbox": [ + 65, + 140, + 285, + 151 + ], + "type": "text", + "content": "Table 1. Adaptation directions of different TTA methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 158, + 269, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 269, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 269, + 170 + ], + "type": "text", + "content": "tion relies solely on streaming batches of target data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 316 + ], + "type": "text", + "content": "Traditional TTA methods (Fig. 1a) [15, 31, 50, 56, 61, 66, 68, 71] typically employ a source-to-target model adaptation framework. These approaches continuously update the source model weights by processing target data batches. Without annotations, the adaptation process relies either on batch-wise updates of model statistics [31, 56, 66, 68], or unsupervised or self-supervised auxiliary tasks [15, 50, 61]. However, small or imbalanced batches may poorly represent the target domain, making these approaches sensitive to the amount and order of the data stream [12, 17, 66]. For instance, adapting the model with batches containing data from only a single category can lead to overfitting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "text", + "content": "Recently, the impressive generation capabilities of diffusion models [22, 25, 29, 48, 51, 53, 67] have sparked the development of diffusion-driven TTA methods (Fig. 1b) [17, 46, 65], leveraging a target-to-source framework. These approaches employ an unconditional diffusion model, pretrained on the source domain, aiming to project each target sample to the source domain independently. This enables the source model to make predictions without modifying its weights. As a preliminary work, DiffPure [46] addresses adversarial perturbations by first applying a forward diffusion process, introducing a small amount of noise to the target data, followed by a reverse diffusion process to restore a clean image to approach the source domain. Building on this concept, more recent studies [17, 65] tackle challenging domain shifts—such as severe data corruption—by incorporating additional structural guidance from the target data, helping preserve semantics and improve performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 652 + ], + "type": "text", + "content": "In this paper, we uncover that while diffusion-driven TTA methods aim to project target data back to the source domain, the projected target data remains confined within the synthetic domain of the unconditional diffusion model. As the synthetic-domain data are ultimately processed by the source-domain model, this domain misalignment limits the final performance. To address this issue, we propose Synthetic-Domain Alignment (SDA) (Fig. 1c), a new category of framework for TTA tasks which simultaneously aligns the domains of the source model and target data with the same synthetic domain of a diffusion model (Tab. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "SDA distinguishes itself from existing diffusion-driven TTA methods [17, 46, 65] by introducing an additional source-to-synthetic model adaptation phase before testing on adapted target data (Fig. 2). Since the adapted target data aligns with the synthetic domain generated by an un" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 74, + 531, + 160 + ], + "blocks": [ + { + "bbox": [ + 340, + 74, + 531, + 160 + ], + "lines": [ + { + "bbox": [ + 340, + 74, + 531, + 160 + ], + "spans": [ + { + "bbox": [ + 340, + 74, + 531, + 160 + ], + "type": "image", + "image_path": "ae6a5cb9f9a020df47ff7598fa7d0efb0fecd225b2904f860a72db1472d2097d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 163, + 554, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 208 + ], + "type": "text", + "content": "Figure 2. Enhanced domain alignment with our framework. Prior diffusion-driven TTA methods struggle with the domain misalignment between the source model and synthetic data, which we resolve by aligning the source model to the synthetic domain." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 216, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 418 + ], + "type": "text", + "content": "conditional diffusion model, SDA aims to adapt the source model to this same synthetic domain. Specifically, SDA employs a Mix-of-Diffusion (MoD) technique to generate synthetic data for model adaptation. Given that the source data is inaccessible after pretraining, MoD first uses a conditional diffusion model to generate samples conditioned on domain-agnostic labels, creating a labeled synthetic dataset. Subsequently, the aforementioned unconditional diffusion model is leveraged by MoD add noise to and denoise these samples, addressing potential domain misalignment between the conditional and unconditional models. With a sufficiently large synthetic dataset, the fine-tuned model becomes highly effective at discriminating within the synthetic domain. Thus, the SDA framework transforms the cross-domain TTA task into an in-domain prediction task by aligning both the source model and target data with the same synthetic domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 419, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 528 + ], + "type": "text", + "content": "SDA is a general framework, not limited to specific fine-tuning techniques or diffusion-driven data adaptation methods. This flexibility allows future advancements in these areas to further broaden its applicability. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLM, e.g., LLaVA) demonstrate that SDA consistently outperforms existing methods. Moreover, the effectiveness of our approach is reinforced through visualization analysis and ablation studies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 538, + 400, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 400, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 400, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "Test-time adaptation (TTA) is an emerging research area that addresses domain shifts by adapting either models [15, 31, 50, 56, 61, 66, 68, 71] or data [17, 46, 65] during evaluation on streaming target data batches. Early model adaptation methods update batch normalization statistics to match the target distribution [31, 56, 66, 71], while others leverage self-supervised tasks like rotation prediction [61] or image restoration [15, 50] to adjust model weights. However, these approaches rely heavily on continuous weight updates, making them sensitive to the amount, order, and diversity of target data. In contrast, diffusion-driven TTA methods [17, 46, 65] focus on data adaptation by projecting each target sample back into the source domain, achiev-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30504" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 70, + 108, + 113 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 108, + 113 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 108, + 113 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 108, + 113 + ], + "type": "image", + "image_path": "f76f39f78b6dd9a714f6089a1acaac65d72590d4ea1936ffbb7ede0673e3b686.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 81, + 154, + 97 + ], + "lines": [ + { + "bbox": [ + 108, + 81, + 154, + 97 + ], + "spans": [ + { + "bbox": [ + 108, + 81, + 154, + 97 + ], + "type": "text", + "content": "Forward Diffusion (Add Noise t steps)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 114, + 244, + 122 + ], + "lines": [ + { + "bbox": [ + 107, + 114, + 244, + 122 + ], + "spans": [ + { + "bbox": [ + 107, + 114, + 244, + 122 + ], + "type": "text", + "content": "(a) Diffusion-driven data adaptation on source data" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 197, + 113 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 197, + 113 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 197, + 113 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 197, + 113 + ], + "type": "image", + "image_path": "b411876b0eb926ae34ce1345cf1b496951c949f701391aaa467d6f2baa8f1b4f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 82, + 241, + 97 + ], + "lines": [ + { + "bbox": [ + 198, + 82, + 241, + 97 + ], + "spans": [ + { + "bbox": [ + 198, + 82, + 241, + 97 + ], + "type": "text", + "content": "Reverse Diffusion (Denoise " + }, + { + "bbox": [ + 198, + 82, + 241, + 97 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 198, + 82, + 241, + 97 + ], + "type": "text", + "content": " steps)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 70, + 285, + 113 + ], + "blocks": [ + { + "bbox": [ + 244, + 70, + 285, + 113 + ], + "lines": [ + { + "bbox": [ + 244, + 70, + 285, + 113 + ], + "spans": [ + { + "bbox": [ + 244, + 70, + 285, + 113 + ], + "type": "image", + "image_path": "bf422646f3c3df0e2609b0f42bff7dc77d6c9ffd5d040179a2440eed609736c7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 64, + 125, + 96, + 165 + ], + "blocks": [ + { + "bbox": [ + 64, + 125, + 96, + 165 + ], + "lines": [ + { + "bbox": [ + 64, + 125, + 96, + 165 + ], + "spans": [ + { + "bbox": [ + 64, + 125, + 96, + 165 + ], + "type": "image", + "image_path": "9658ab4994ab010ae6a1feefd933a18dddb4d8263b4a3d554cc2580a18d47772.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 97, + 125, + 122, + 164 + ], + "blocks": [ + { + "bbox": [ + 97, + 125, + 122, + 164 + ], + "lines": [ + { + "bbox": [ + 97, + 125, + 122, + 164 + ], + "spans": [ + { + "bbox": [ + 97, + 125, + 122, + 164 + ], + "type": "image", + "image_path": "6495ec9619f09d6e98a25298f008adf15adac7746c48e5972190c81c4d2eb86e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 126, + 156, + 132 + ], + "lines": [ + { + "bbox": [ + 123, + 126, + 156, + 132 + ], + "spans": [ + { + "bbox": [ + 123, + 126, + 156, + 132 + ], + "type": "inline_equation", + "content": "t = 600" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 122, + 133, + 158, + 164 + ], + "blocks": [ + { + "bbox": [ + 122, + 133, + 158, + 164 + ], + "lines": [ + { + "bbox": [ + 122, + 133, + 158, + 164 + ], + "spans": [ + { + "bbox": [ + 122, + 133, + 158, + 164 + ], + "type": "image", + "image_path": "d135f7544f4ebca577ea4659c9ce5cd22e29611454d05ccf22891acc59179fcc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 159, + 132, + 190, + 163 + ], + "blocks": [ + { + "bbox": [ + 159, + 126, + 188, + 132 + ], + "lines": [ + { + "bbox": [ + 159, + 126, + 188, + 132 + ], + "spans": [ + { + "bbox": [ + 159, + 126, + 188, + 132 + ], + "type": "inline_equation", + "content": "t = 700" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 132, + 190, + 163 + ], + "lines": [ + { + "bbox": [ + 159, + 132, + 190, + 163 + ], + "spans": [ + { + "bbox": [ + 159, + 132, + 190, + 163 + ], + "type": "image", + "image_path": "fa31cbe96d9044f80a73260319369fb2a08f96325220bf6571a1f1d93161edf4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 191, + 132, + 209, + 163 + ], + "blocks": [ + { + "bbox": [ + 191, + 126, + 219, + 132 + ], + "lines": [ + { + "bbox": [ + 191, + 126, + 219, + 132 + ], + "spans": [ + { + "bbox": [ + 191, + 126, + 219, + 132 + ], + "type": "inline_equation", + "content": "t = 800" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 191, + 132, + 209, + 163 + ], + "lines": [ + { + "bbox": [ + 191, + 132, + 209, + 163 + ], + "spans": [ + { + "bbox": [ + 191, + 132, + 209, + 163 + ], + "type": "image", + "image_path": "3260e20ba09c92ec51441b3f25844ece57373735b7959a8ad4fbaeeb31474e1a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 126, + 235, + 132 + ], + "lines": [ + { + "bbox": [ + 209, + 126, + 235, + 132 + ], + "spans": [ + { + "bbox": [ + 209, + 126, + 235, + 132 + ], + "type": "text", + "content": "300 " + }, + { + "bbox": [ + 209, + 126, + 235, + 132 + ], + "type": "inline_equation", + "content": "t =" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 209, + 132, + 241, + 163 + ], + "blocks": [ + { + "bbox": [ + 209, + 132, + 241, + 163 + ], + "lines": [ + { + "bbox": [ + 209, + 132, + 241, + 163 + ], + "spans": [ + { + "bbox": [ + 209, + 132, + 241, + 163 + ], + "type": "image", + "image_path": "274cd19e60f21e2719277f5c30eed42ad7b08e2426279831547a19767f246940.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 242, + 132, + 267, + 163 + ], + "blocks": [ + { + "bbox": [ + 242, + 126, + 266, + 132 + ], + "lines": [ + { + "bbox": [ + 242, + 126, + 266, + 132 + ], + "spans": [ + { + "bbox": [ + 242, + 126, + 266, + 132 + ], + "type": "text", + "content": "00 " + }, + { + "bbox": [ + 242, + 126, + 266, + 132 + ], + "type": "inline_equation", + "content": "t =" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 132, + 267, + 163 + ], + "lines": [ + { + "bbox": [ + 242, + 132, + 267, + 163 + ], + "spans": [ + { + "bbox": [ + 242, + 132, + 267, + 163 + ], + "type": "image", + "image_path": "44cf3a6c80a96a086b40e25e25a62c1243aa91371f3421f53fdba6820940dc47.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 268, + 132, + 284, + 163 + ], + "blocks": [ + { + "bbox": [ + 268, + 126, + 282, + 132 + ], + "lines": [ + { + "bbox": [ + 268, + 126, + 282, + 132 + ], + "spans": [ + { + "bbox": [ + 268, + 126, + 282, + 132 + ], + "type": "text", + "content": "1000" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 132, + 284, + 163 + ], + "lines": [ + { + "bbox": [ + 268, + 132, + 284, + 163 + ], + "spans": [ + { + "bbox": [ + 268, + 132, + 284, + 163 + ], + "type": "image", + "image_path": "db2b1066657c2e9216f5b9fb14f32d83c52200a872653ac216afeb42d398bee4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 65, + 166, + 96, + 198 + ], + "blocks": [ + { + "bbox": [ + 65, + 166, + 96, + 198 + ], + "lines": [ + { + "bbox": [ + 65, + 166, + 96, + 198 + ], + "spans": [ + { + "bbox": [ + 65, + 166, + 96, + 198 + ], + "type": "image", + "image_path": "686858674df97e6cd33cd5264ade44d2bb3ad10798738558a56c35b5c032b0c1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 97, + 166, + 157, + 198 + ], + "blocks": [ + { + "bbox": [ + 97, + 166, + 157, + 198 + ], + "lines": [ + { + "bbox": [ + 97, + 166, + 157, + 198 + ], + "spans": [ + { + "bbox": [ + 97, + 166, + 157, + 198 + ], + "type": "image", + "image_path": "2dd879d31febf613cff71516778612c563b7a6ea75e51841c48b141d8e4043be.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 159, + 166, + 190, + 198 + ], + "blocks": [ + { + "bbox": [ + 159, + 166, + 190, + 198 + ], + "lines": [ + { + "bbox": [ + 159, + 166, + 190, + 198 + ], + "spans": [ + { + "bbox": [ + 159, + 166, + 190, + 198 + ], + "type": "image", + "image_path": "849e9a314b2c444880088642bac7fd4ec50f8c30857b10afef7fef689f3feed1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 190, + 166, + 222, + 198 + ], + "blocks": [ + { + "bbox": [ + 190, + 166, + 222, + 198 + ], + "lines": [ + { + "bbox": [ + 190, + 166, + 222, + 198 + ], + "spans": [ + { + "bbox": [ + 190, + 166, + 222, + 198 + ], + "type": "image", + "image_path": "e1f7349857c5cf5a9af7d546584782aa6ad105ed33ad38105942bf3c5a1f9cfd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 223, + 166, + 248, + 198 + ], + "blocks": [ + { + "bbox": [ + 223, + 166, + 248, + 198 + ], + "lines": [ + { + "bbox": [ + 223, + 166, + 248, + 198 + ], + "spans": [ + { + "bbox": [ + 223, + 166, + 248, + 198 + ], + "type": "image", + "image_path": "362c0480fe2d9bd188d37e569a6ef77d7c47da79a7568c35f7b67ed812d9e0da.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 249, + 166, + 283, + 198 + ], + "blocks": [ + { + "bbox": [ + 249, + 166, + 283, + 198 + ], + "lines": [ + { + "bbox": [ + 249, + 166, + 283, + 198 + ], + "spans": [ + { + "bbox": [ + 249, + 166, + 283, + 198 + ], + "type": "image", + "image_path": "ca30e1e485a0746f17f2c40b784d92643a06585bff81bb0e73a81fa665f420e6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 65, + 199, + 96, + 232 + ], + "blocks": [ + { + "bbox": [ + 65, + 199, + 96, + 232 + ], + "lines": [ + { + "bbox": [ + 65, + 199, + 96, + 232 + ], + "spans": [ + { + "bbox": [ + 65, + 199, + 96, + 232 + ], + "type": "image", + "image_path": "5e70fe42eed1b45b38083a2844b1abd2687e8daa9b83aeb3fe00ba98c1e8f255.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 232, + 289, + 240 + ], + "lines": [ + { + "bbox": [ + 59, + 232, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 59, + 232, + 289, + 240 + ], + "type": "text", + "content": "(b) Diffusion-driven data adaptation results (on source data) under different timesteps" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 243, + 295, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 286 + ], + "type": "text", + "content": "Figure 3. (a) Illustration of diffusion-driven data adaptation on source data and (b) Adapted images across different timesteps. The results are obtained using DDA [17], with no noticeable visual degradation observed in the adapted images." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 97, + 199, + 129, + 232 + ], + "blocks": [ + { + "bbox": [ + 97, + 199, + 129, + 232 + ], + "lines": [ + { + "bbox": [ + 97, + 199, + 129, + 232 + ], + "spans": [ + { + "bbox": [ + 97, + 199, + 129, + 232 + ], + "type": "image", + "image_path": "a7b54d4cea9c478a4527c308180055254e4e9bd67412bbb4316b695ea651b1f1.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 129, + 199, + 158, + 232 + ], + "blocks": [ + { + "bbox": [ + 129, + 199, + 158, + 232 + ], + "lines": [ + { + "bbox": [ + 129, + 199, + 158, + 232 + ], + "spans": [ + { + "bbox": [ + 129, + 199, + 158, + 232 + ], + "type": "image", + "image_path": "d696fe311078fd824df85682bdf04ecef309c7ad6190305ea057ec036dade8fa.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 159, + 199, + 190, + 232 + ], + "blocks": [ + { + "bbox": [ + 159, + 199, + 190, + 232 + ], + "lines": [ + { + "bbox": [ + 159, + 199, + 190, + 232 + ], + "spans": [ + { + "bbox": [ + 159, + 199, + 190, + 232 + ], + "type": "image", + "image_path": "75ee21dd2d895a9ceffbc52e0f6d257916b6d043208ebfca8b13e9b266e32e94.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 191, + 199, + 209, + 232 + ], + "blocks": [ + { + "bbox": [ + 191, + 199, + 209, + 232 + ], + "lines": [ + { + "bbox": [ + 191, + 199, + 209, + 232 + ], + "spans": [ + { + "bbox": [ + 191, + 199, + 209, + 232 + ], + "type": "image", + "image_path": "b69451d7c92deb860cd419f12257cda74cbaf1956ebac352679bc18bbc9465f6.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 209, + 199, + 247, + 232 + ], + "blocks": [ + { + "bbox": [ + 209, + 199, + 247, + 232 + ], + "lines": [ + { + "bbox": [ + 209, + 199, + 247, + 232 + ], + "spans": [ + { + "bbox": [ + 209, + 199, + 247, + 232 + ], + "type": "image", + "image_path": "49651a04855085445a68e6b0df6b8bc9eb9a9726f9b050ef8a24e95a137ea152.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 248, + 199, + 269, + 232 + ], + "blocks": [ + { + "bbox": [ + 248, + 199, + 269, + 232 + ], + "lines": [ + { + "bbox": [ + 248, + 199, + 269, + 232 + ], + "spans": [ + { + "bbox": [ + 248, + 199, + 269, + 232 + ], + "type": "image", + "image_path": "e0b67c0d8ae1f9432ecf1f69a5cfcfac0b7face8dbea36153c01e8f470e897fe.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 303, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 387 + ], + "type": "text", + "content": "ing stable performance without online model updates. DiffPure [46] purifies adversarial samples with diffusion models, while DDA [17] and GDA [65] use structural guidance to preserve image content under severe corruption. Building on this, our work investigates the issue of misalignment between domains of the source model and synthetic images in diffusion-driven TTA and proposes a new TTA framework." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 388, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 506 + ], + "type": "text", + "content": "Synthetic data for discriminative tasks. Synthetic data, generated by models [20, 21, 42-44, 48, 52, 53] rather than collected from the real world, has shown significant potential in enhancing visual representations for various discriminative tasks [14, 62, 63]. It has been effectively applied in areas such as recognition [3, 63], object detection [49, 55], semantic segmentation [7, 54, 58], image assessment [19], autonomous driving [1], and robotics [39, 69]. In this work, we explore the potential of diffusion-generated synthetic data to enhance domain alignment in TTA." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 517, + 138, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 138, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 138, + 530 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 536, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 608 + ], + "type": "text", + "content": "In this section, we introduce the background of diffusion-driven TTA methods and identify their source-synthetic domain misalignment issue in Sec. 3.1 and Sec. 3.2. To tackle this issue, we propose the synthetic-domain alignment (SDA) framework in Sec. 3.3 and introduce its key technique, the mix of diffusion (MoD) in Sec. 3.4." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 615, + 135, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 135, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 135, + 627 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "Diffusion process. Given a source data point " + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^{\\mathrm{src}} \\sim p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": ", diffusion models " + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": " [25] gradually transform " + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": " into a Gaussian noise distribution " + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "inline_equation", + "content": "N(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": " through a " + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "-step forward diffusion process. At each timestep " + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "inline_equation", + "content": "t \\in \\{1, 2, \\dots, T\\}" + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": ", the intermediate state " + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t^{\\mathrm{src}} \\sim p_t^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 108, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 108, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}} = \\sqrt {1 - \\beta_ {t}} \\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}} + \\sqrt {\\beta_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {1}", + "image_path": "f6e5b046f996d55e6fb4ebf834c341a3c84835f562eee80cedc256b243ec7e3a.jpg" + } + ] + } + ], + "index": 41 + }, + { + "type": "table", + "bbox": [ + 315, + 61, + 554, + 144 + ], + "blocks": [ + { + "bbox": [ + 315, + 61, + 554, + 144 + ], + "lines": [ + { + "bbox": [ + 315, + 61, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 315, + 61, + 554, + 144 + ], + "type": "table", + "html": "
Timestep t5006007008009001000
Swin-B (Source-Synthetic, Misaligned)61.657.855.752.645.038.5
Swin-B (Synthetic-Synthetic, Aligned)67.665.965.061.555.748.4
Δ+6.0+8.1+9.3+8.9+10.7+9.9
ConvNeXt-B (Source-Synthetic, Misaligned)65.161.360.357.250.041.5
ConvNeXt-B (Synthetic-Synthetic, Aligned)70.668.267.464.958.550.7
Δ+5.5+6.9+7.1+7.7+8.5+9.2
", + "image_path": "d6378c10976ab96ddb4e3a90c54eff733e065257e623f02d10688f87a2eebada.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "type": "text", + "content": "Table 2. Source model accuracy across different timesteps of diffusion-driven data adaptation. For suitable timesteps for TTA " + }, + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "type": "inline_equation", + "content": "(t\\geq 500)" + }, + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "type": "text", + "content": " [17], model accuracy shows a monotonically decreasing trend with the growth of timestep, indicating the increase of the misalignment of source domain " + }, + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "type": "text", + "content": " and synthetic domain " + }, + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 147, + 555, + 236 + ], + "type": "text", + "content": ". By aligning the source model to the synthetic domain, our methods (rows 3 & 6) significantly help alleviate the performance degradation. Results are evaluated on the ImageNet [9] validation set." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "content": " is random Gaussian noise and " + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\beta_{t} \\in (0,1)" + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "content": " represents the diffusion rate at step " + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "content": ". By defining " + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha_{t}} = \\prod_{t=1}^{T} \\alpha_{t}" + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 251, + 553, + 289 + ], + "type": "text", + "content": ", we obtain:" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 372, + 294, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 294, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 372, + 294, + 553, + 308 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} ^ {\\mathrm {s r c}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}. \\tag {2}", + "image_path": "7b21b5a92d3615cd99d6f9f8a495496050b70eadadda4647b7314d79e7b819db.jpg" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "content": "The reverse diffusion process recovers a clean " + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{x}_0^{\\mathrm{src}}}" + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "content": " by progressively removing noise from " + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_T^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 326, + 346, + 553, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 346, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 326, + 346, + 553, + 372 + ], + "type": "interline_equation", + "content": "\\widehat {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\widehat {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(\\widehat {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\tag {3}", + "image_path": "8e5776ae34c94d1cd48d39c385ceac09f346badb33f472961f8c5ce4fd9838e4.jpg" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 378, + 497, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 497, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 497, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 378, + 497, + 390 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 378, + 497, + 390 + ], + "type": "text", + "content": " is the posterior noise variance [25]." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 392, + 553, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 553, + 428 + ], + "type": "text", + "content": "Diffusion-driven data adaptation. Denote by " + }, + { + "bbox": [ + 313, + 392, + 553, + 428 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 392, + 553, + 428 + ], + "type": "text", + "content": " the target data distribution, from which each target data point " + }, + { + "bbox": [ + 313, + 392, + 553, + 428 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 392, + 553, + 428 + ], + "type": "text", + "content": " is sampled. Prior work [46] demonstrates that:" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 350, + 435, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 435, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 350, + 435, + 553, + 449 + ], + "type": "interline_equation", + "content": "D _ {\\mathrm {K L}} \\left(p _ {t + 1} ^ {\\mathrm {s r c}} \\mid p _ {t + 1} ^ {\\mathrm {t r g}}\\right) - D _ {\\mathrm {K L}} \\left(p _ {t} ^ {\\mathrm {s r c}} \\mid p _ {t} ^ {\\mathrm {t r g}}\\right) \\leq 0, \\tag {4}", + "image_path": "f60fbf191799c2c5f66bb0a1a671b9992e327cb5db8f5f167bad63c3858eff69.jpg" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": " is the KL divergence. Since " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "p_T^{\\mathrm{src}} = p_T^{\\mathrm{trg}} = N(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": ", for any arbitrarily small value " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": ", there exists a minimum timestep " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}(p_{t^*}^{\\mathrm{src}}||p_{t^*}^{\\mathrm{trg}}) < \\delta" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": ". As discussed in [17], a diffusion process with " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": " requires " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "t^* \\geq 500" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": " to eliminate domain shifts. We empirically validate this choice of " + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 313, + 456, + 554, + 529 + ], + "type": "text", + "content": " in the supplementary materials." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "text", + "content": "The initial diffusion-driven data adaptation in Diff-Pure [46] is carried out as follows: Given the minor divergence between " + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "inline_equation", + "content": "p_{t^*}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "inline_equation", + "content": "p_{t^*}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "text", + "content": ", we recover each " + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t^*}^{\\mathrm{trg}} \\sim p_{t^*}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "text", + "content": " to its corresponding " + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{x}_0^{\\mathrm{src}}} \\sim p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "text", + "content": " by executing the reverse diffusion process (Eq. 3) " + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 313, + 529, + 554, + 590 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": "Since the reverse process is stochastic, subsequent works [17, 65] further introduce additional structure guidance to ensure the content consistency between each " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": " and its adapted " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": ". In this work, we adopt the same data adaptation process as DDA [17]:" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 329, + 658, + 553, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 658, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 329, + 658, + 553, + 678 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} = \\widehat {\\boldsymbol {x} _ {t - 1} ^ {\\mathrm {s r c}}} - \\boldsymbol {w} \\nabla_ {\\widetilde {\\boldsymbol {x} _ {t} ^ {\\mathrm {s r c}}}} \\left\\| \\phi \\left(\\boldsymbol {x} _ {0} ^ {\\mathrm {t r g}}\\right) - \\phi \\left(\\widetilde {\\boldsymbol {x} _ {0 , t} ^ {\\mathrm {s r c}}}\\right) \\right\\| _ {2}, \\tag {5}", + "image_path": "bcd84b300b063cd8cb52be5ae503ac12d7e53d280b9efad1b5575025549ba464.jpg" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}_{t - 1}^{\\mathrm{src}}}" + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "text", + "content": " is computed as Eq. 3, " + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "text", + "content": " is the structure guidance scale, " + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "text", + "content": " is a structure extractor [17] and " + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}_{0,t}^{\\mathrm{src}}}" + }, + { + "bbox": [ + 313, + 685, + 553, + 715 + ], + "type": "text", + "content": " is an esti" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30505" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 553, + 270 + ], + "type": "image", + "image_path": "80b7c7c4458d7d450c98575703dae29f2eeeb6ad025442a57c2e7eab58b13af3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "type": "text", + "content": "Figure 4. Overview of the Synthetic-Domain Alignment (SDA) framework. SDA is a novel TTA framework aligning both the domains of the source model and the target data with the synthetic domain. SDA involves three phases: (left): a source-domain model pretraining phase, where the source model is trained on source data prior to TTA; (middle): a source-to-synthetic model adaptation phase, where the source model is adapted to a synthetic-domain model using synthetic data generated via a Mix of Diffusion (MoD) technique; and (right): a target-to-synthetic data adaptation phase, where target data is adapted into synthetic data using an unconditional diffusion model. Finally, the adapted synthetic data is fed into the synthetic-domain model for test-time inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 357, + 295, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 394 + ], + "type": "text", + "content": "mate of " + }, + { + "bbox": [ + 54, + 357, + 295, + 394 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}" + }, + { + "bbox": [ + 54, + 357, + 295, + 394 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 54, + 357, + 295, + 394 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 357, + 295, + 394 + ], + "type": "text", + "content": " [25]. Diffusion-driven data adaptation is typically performed with unconditional diffusion models [10] since the target data labels are unknown." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 264, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 264, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 264, + 416 + ], + "type": "text", + "content": "3.2. Source-Synthetic Domain Misalignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": "Unlike a real source data point " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": ", the recovered version " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{x}_{0}^{\\mathrm{src}}}," + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t^{*}}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " is synthetic. Specifically, " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " follows the source domain " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": ", whereas " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " follows the synthetic domain " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " of an unconditional diffusion model " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{u}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": ". In this section, we empirically reveal the misalignment between the source domain " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " and the synthetic domain " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": ", and investigate how this misalignment impacts the performance of the source model. For simplicity and clarity, we will substitute " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{x}_0^{\\mathrm{src}}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 54, + 420, + 295, + 544 + ], + "type": "text", + "content": " throughout the following discussion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 545, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 629 + ], + "type": "text", + "content": "Based on the above analysis, diffusion-driven TTA methods face two potential misalignments: (1) source-target domain misalignment arising from inherent data distribution shifts, which has been the primary focus of prior research, and (2) source-synthetic domain misalignment introduced by diffusion models, which we address as the main focus of this work, complementing existing efforts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": "To precisely examine the impact of the source-synthetic domain misalignment and isolate it from the influence of the source-target domain misalignment, we evaluate the performance of the source model " + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": " on synthetic data " + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": " adapted by the diffusion model " + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{u}}" + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": " from source data " + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": " (Fig. 3a). Specifically, we test ImageNet pretrained models [9] on the ImageNet validation set adapted by the popular diffusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": "driven data adaptation method, DDA [17] across different timesteps " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": ". As indicated in Tab. 2, with an increase in " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": ", the accuracy of the model exhibits a monotonically decreasing trend. For ideal " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "t^* \\geq 500" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " for TTA tasks, performance degradation of more than " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "18.8\\%" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " for ConvNeXt [36] and " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " for Swin [35] is observed compared to their official results [8] on source data which are " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "83.4\\%" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "83.9\\%" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": ". This indicates a significant domain misalignment between " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": ". By fine-tuning the source model on the synthetic data generated by our Mix of Diffusion process (introduced in Sec. 3.4), the aligned models " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " achieve performance improvements of " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " for ConvNeXt and " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " for Swin." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "content": "In Fig. 3b, we show that the diffusion synthetic data " + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "content": " and source data " + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "content": " exhibit no noticeable visual differences across different timesteps " + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "content": ". This further suggests that the performance degradation is not due to the quality of diffusion-generated images but rather the implicit misalignment between the source and synthetic domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 586, + 531, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 531, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 531, + 598 + ], + "type": "text", + "content": "3.3. Synthetic-Domain Alignment Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "content": "Given that the diffusion-adapted data " + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "content": " aligns more closely with the synthetic domain " + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "content": " instead of the source domain " + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "content": ", we propose simultaneously adapting the source model " + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "content": " to the same synthetic domain " + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "content": ". By doing so, the alignment between the data and model within " + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "inline_equation", + "content": "p_0^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 688 + ], + "type": "text", + "content": " can be effectively achieved. To this end, we introduce a novel TTA framework: Synthetic-Domain Alignment (SDA)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "In Fig. 4, we present the complete diagram of SDA, which consists of three key phases: (1) a source-domain" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30506" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "model pretraining phase, where the source model " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " is trained on source data " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " prior to TTA; (2) a source-to-synthetic model adaptation phase, where " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " is fine-tuned to a synthetic-domain model " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "; and (3) a target-to-synthetic data adaptation phase, where target data " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " is adapted into synthetic data " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " using an unconditional diffusion model " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{u}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " following Eq. 5. Finally, the adapted synthetic data " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " is fed into the synthetic-domain model " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " for test-time inference. Consistent with the standard TTA protocol [66], the source data is accessible only during the source model pretraining phase and remains inaccessible during both the model and data adaptation phases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 275 + ], + "type": "text", + "content": "The rationale behind SDA is straightforward: by aligning the model and data domains to the same synthetic domain, the original cross-domain TTA task is transformed into an easier in-domain prediction task, thus addressing the core challenge of TTA and improving performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 276, + 294, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 294, + 347 + ], + "type": "text", + "content": "Since the pretraining phase follows standard supervised learning, and the data adaptation phase aligns with DDA [17], we omit detailed explanations of these phases. Instead, we focus on how we adapt the source model " + }, + { + "bbox": [ + 55, + 276, + 294, + 347 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 276, + 294, + 347 + ], + "type": "text", + "content": " to an ideal synthetic domain model " + }, + { + "bbox": [ + 55, + 276, + 294, + 347 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 55, + 276, + 294, + 347 + ], + "type": "text", + "content": " using a novel technique called Mix-of-Diffusion (MoD), introduced next." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 260, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 260, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 260, + 368 + ], + "type": "text", + "content": "3.4. Model Adaptation via Mix of Diffusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": "As shown in Fig. 4, MoD consists of two main processes: a data generation process powered by a conditional diffusion model " + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\eta}^{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": ", and a data alignment process powered by the same unconditional diffusion model " + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{u}}" + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": " used in the target-to-synthetic data adaptation phase. It is worth noting that, both " + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\eta}^{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{u}}" + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": " are also pretrained on source data " + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": " and have never been exposed to target data " + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 373, + 294, + 470 + ], + "type": "text", + "content": ", in accordance with the TTA setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": "Conditional diffusion data generation. The generation process leverages the conditional generation capability of " + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\eta}^{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": " to synthesize conditional synthetic data " + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "inline_equation", + "content": "x_{0,c}^{\\mathrm{syn}} \\sim p_{0,c}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": ". In the context of TTA, the source and target domains share the same domain-agnostic label set " + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "inline_equation", + "content": "\\{y_i\\}_{i=1}^{K}" + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": ". Utilizing " + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\eta}^{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": ", we uniformly generate samples for each class " + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": " from Gaussian noise " + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "inline_equation", + "content": "(x_{T,c}^{\\mathrm{syn}})" + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": " through a " + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 473, + 294, + 558 + ], + "type": "text", + "content": "-step reverse diffusion process:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 567, + 294, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 567, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 61, + 567, + 294, + 594 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t - 1, c} ^ {\\text {s y n}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t, c} ^ {\\text {s y n}} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\overline {{\\alpha_ {t}}}}} \\epsilon_ {\\eta} ^ {\\mathrm {c}} \\left(\\boldsymbol {x} _ {t, c} ^ {\\text {s y n}}, t, y _ {i}\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\epsilon}. \\tag {6}", + "image_path": "de112dbcdb558e3228daca9e4122bc1f8cabf8674c28022fd043334d841ca334.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "content": "The generation capability of " + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\eta}^{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "content": " allows for the construction of a labeled synthetic-domain dataset " + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "inline_equation", + "content": "\\{x_{0,\\mathrm{c}}^{\\mathrm{syn}},y\\}^{N}" + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "content": " of arbitrary size " + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "content": " without any manual data collection. By fine-tuning the source model " + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "content": " on this synthetic dataset, an adapted model " + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{c}}^{\\prime}" + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "content": " on domain " + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{c}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 601, + 294, + 663 + ], + "type": "text", + "content": " can be obtained." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "text", + "content": "Unconditional diffusion data alignment. However, since the test-time adapted data " + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "inline_equation", + "content": "x_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "text", + "content": " is subject to domain " + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "text", + "content": ", we argue that there is still potential misalignment between different synthetic domains " + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{c}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 665, + 294, + 715 + ], + "type": "text", + "content": ". This is mainly" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "because of the differences in architectures and training schemes between " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\eta}^{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{u}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ". We empirically validate this argument in Tab. 10, which indicates that the model " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{c}}'" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " adapted to domain " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{c}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " performs worse than the model " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{u}}'" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " adapted to domain " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " during testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": "To obtain the " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{u}}^{\\prime}" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": " on domain " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "p_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": " we mirror the target-to-synthetic data adaptation phase as a conditional synthetic data " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{x}_{0,\\mathrm{c}}^{\\mathrm{syn}})" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": " to unconditional synthetic data " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}})" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": " adaptation process. In specific, we use the same " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": " in Sec. 3.1 with Eq. 2 to obtain " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t^{*},\\mathrm{c}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": ". According to the analysis in Sec. 3.1, " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t^{*},\\mathrm{c}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": " will be indistinguishable to its counterpart " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": " on domain " + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "inline_equation", + "content": "p_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": ". Therefore, we have:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 223, + 553, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 223, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 351, + 223, + 553, + 238 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t ^ {*}, \\mathrm {u}} ^ {\\text {s y n}} \\approx \\boldsymbol {x} _ {t ^ {*}, \\mathrm {c}} ^ {\\text {s y n}} = \\sqrt {\\overline {{\\alpha_ {t}}}} \\boldsymbol {x} _ {0, c} ^ {\\text {s y n}} + \\sqrt {1 - \\overline {{\\alpha_ {t}}}} \\boldsymbol {\\epsilon}. \\tag {7}", + "image_path": "fab73dc475ef1cb97647214f2e44f418907424430c46e4bd66adc1970ef15c18.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "text", + "content": "Then, following Eq. 3 and Eq. 5, the noisy " + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t^{*},\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "text", + "content": " can be gradually denoised to " + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "text", + "content": ". Finally, the expected synthetic-domain model " + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{u}}^{\\prime}" + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "text", + "content": " is obtained by fine-tuning the source-domain model " + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{u}}^{\\prime}" + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "text", + "content": " on dataset " + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}},y\\}^{N}" + }, + { + "bbox": [ + 313, + 243, + 553, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 294, + 553, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 377 + ], + "type": "text", + "content": "Ensembling. As noted in previous diffusion-driven TTA methods [17], while diffusion models generally perform well for data adaptation, they may occasionally produce data points that are less recognizable than the original target data. To address this, prior approaches use an ensemble of model predictions on " + }, + { + "bbox": [ + 313, + 294, + 553, + 377 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{0}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 294, + 553, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 294, + 553, + 377 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{0,\\mathrm{u}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 294, + 553, + 377 + ], + "type": "text", + "content": " as the final output. Following this protocol, the final prediction in SDA is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 388, + 553, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 388, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 354, + 388, + 553, + 407 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\arg \\max _ {y} \\left(q \\left(y \\mid \\boldsymbol {x} _ {0} ^ {\\operatorname {t r g}}\\right) + q ^ {\\prime} \\left(y \\mid \\boldsymbol {x} _ {0, \\mathrm {u}} ^ {\\operatorname {s y n}}\\right)\\right), \\tag {8}", + "image_path": "e1de5c550cbd6abf0d01362241f57ed1196fd41d1766caa949ae196fa5245954.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "inline_equation", + "content": "q^{\\prime}(\\cdot)" + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "text", + "content": " are output distributions of source model " + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "text", + "content": " and synthetic-domain model " + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{u}}^{\\prime}" + }, + { + "bbox": [ + 313, + 411, + 553, + 436 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 445, + 394, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 394, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 394, + 458 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 464, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 559 + ], + "type": "text", + "content": "In this section, we first evaluate SDA on ImageNet classifiers with standard TTA benchmarks in Sec. 4.1. Next, we assess SDA's scalability across different dataset sizes, tasks, and model architectures in Sec. 4.2. In Sec. 4.3, we demonstrate SDA's advantages through Grad-CAM visualizations [59] and data stream sensitivity tests. Finally, ablation studies in Sec. 4.4 validate the design choices in our SDA framework." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 315, + 567, + 553, + 639 + ], + "blocks": [ + { + "bbox": [ + 315, + 567, + 553, + 639 + ], + "lines": [ + { + "bbox": [ + 315, + 567, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 315, + 567, + 553, + 639 + ], + "type": "table", + "html": "
ModelSourceMEMODiffPureGDADDASDA (Ours)
ResNet-5018.724.716.831.829.732.5 (+2.8)
Swin-T33.529.524.842.240.042.5 (+2.5)
ConvNeXt-T39.337.828.844.844.247.0 (+2.8)
Swin-B40.537.028.9-44.547.4 (+2.9)
ConvNeXt-B45.645.832.7-49.451.9 (+2.5)
", + "image_path": "2e5e0a8b057d019c8f348c4e165633db0d7c9fc468d53cf249d44c817171d03f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 650, + 553, + 704 + ], + "lines": [ + { + "bbox": [ + 313, + 650, + 553, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 553, + 704 + ], + "type": "text", + "content": "Table 3. Comparison results on ImageNet-C. We compare SDA with source models, MEMO [71], DiffPure [46], GDA [65] and DDA [17]. Results are the average accuracy across 15 adaptation domains at severity level 5. SDA shows consistent performance improvements compared to baselines." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30507" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 67, + 555, + 153 + ], + "blocks": [ + { + "bbox": [ + 56, + 67, + 555, + 153 + ], + "lines": [ + { + "bbox": [ + 56, + 67, + 555, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 67, + 555, + 153 + ], + "type": "table", + "html": "
GaussianShotImpluseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source48.047.148.332.414.439.339.255.750.151.575.152.029.548.160.346.1
DiffPure42.642.040.916.822.219.123.039.029.19.162.53.034.346.860.732.7
DDA59.955.755.930.730.538.139.253.248.041.971.048.146.959.562.049.4
SDA (Ours)60.357.658.135.435.342.942.455.549.944.672.645.450.263.064.551.9 (+2.5)
", + "image_path": "48ea1b07c8e1da0a1e26b16c9861b1402cb6a232a8cf9ad26196e0c2175b5736.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 196, + 253, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 253, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 253, + 209 + ], + "type": "text", + "content": "4.1. Main Results on ImageNet Classifiers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 217, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 295, + 373 + ], + "type": "text", + "content": "Settings. We choose DDA [17] as our primary competitor since it is the best open-sourced method. SDA is also compared with DiffPure [46], MEMO [71], and the recent SOTA GDA [65] using their reported results. Source model performance is reported as \"Source\". DiT [48] and ADM [10] are adopted to generate and align 50K synthetic data for 15-epoch finetuning. For different source models, the synthetic data only needs to be generated once. Our results are tested on standard TTA benchmarks, ImageNet-C [24] (severity level 5) and ImageNet-W [32] using various models including ResNet [23], Swin [35] and ConvNeXt [36]. More implementation details are provided in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 378, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 295, + 713 + ], + "type": "text", + "content": "Comparison results on ImageNet-C. We begin by evaluating the performance of SDA on ImageNet-C. As reported in Tab. 3, our proposed SDA consistently outperforms all baseline methods across different model architectures and sizes. We emphasize the performance improvement over DDA, as we adopted DDA for target data adaptation in SDA. Compared to DDA, our SDA improves accuracy by " + }, + { + "bbox": [ + 56, + 378, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.5\\% - 2.9\\%" + }, + { + "bbox": [ + 56, + 378, + 295, + 713 + ], + "type": "text", + "content": ". This significant improvement indicates the misalignment between the source and synthetic domains, validating the effectiveness of our synthetic-domain alignment framework. Moreover, compared to the recent SOTA GDA, SDA also achieves an improvement of " + }, + { + "bbox": [ + 56, + 378, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 56, + 378, + 295, + 713 + ], + "type": "text", + "content": " with ConvNeXt-T. Notably, SDA focuses on synthetic domain alignment, an orthogonal research direction to existing diffusion-driven methods on better adapting the target data. Therefore, the performance of SDA could potentially be further enhanced with the release of more advanced codebases like GDA. Compared to the model adaptation method, MEMO, three diffusion-driven methods (SDA, DDA, and GDA) all demonstrate superior performance, highlighting the effectiveness of diffusion models in assisting TTA tasks. DiffPure presents worse results since it is primarily designed for adversarial attacks. Without the structural guidance introduced in DDA and GDA, DiffPure may not effectively recover images with severe domain shifts. In Tab. 4, we provide a detailed comparison of the results of SDA and baselines. SDA surpasses DiffPure in all 15 adaptation domains and outperforms DDA in 14 out of 15 domains, fur" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 194, + 555, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 158, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 555, + 182 + ], + "type": "text", + "content": "Table 4. Detailed comparisons of SDA and baselines across 15 adaptation domains of ImageNet-C. SDA shows the best average accuracy. The results are tested with ConvNeXt-B. Comparisons using other models are deferred to the supplementary materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 194, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 315, + 194, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 194, + 555, + 277 + ], + "type": "table", + "html": "
ModelSourceDiffPureDDASDA (Ours)
ResNet-5037.729.152.854.7 (+1.9)
Swin-T66.552.765.967.3 (+1.4)
ConvNeXt-T67.655.867.969.4 (+1.5)
Swin-B69.155.568.370.6 (+2.3)
ConvNeXt-B70.157.770.372.3 (+2.0)
", + "image_path": "13f00ce2921a8567b682040f674008f0c36e556e848461a9ac20848c39a2abf0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 285, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 308 + ], + "type": "text", + "content": "Table 5. Comparison results on ImageNet-W. SDA shows consistent performance improvements compared to baselines." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 321, + 468, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 468, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 468, + 333 + ], + "type": "text", + "content": "ther affirming the superiority of SDA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 336, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 504 + ], + "type": "text", + "content": "Comparison results on ImageNet-W. We extend our evaluation to ImageNet-W to assess SDA's performance under watermark-based domain shifts. As shown in Tab. 5, SDA consistently surpasses all baselines across different models. Compared to our primary baseline, DDA, SDA achieves accuracy gains ranging from " + }, + { + "bbox": [ + 313, + 336, + 555, + 504 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 313, + 336, + 555, + 504 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 336, + 555, + 504 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 313, + 336, + 555, + 504 + ], + "type": "text", + "content": ". Furthermore, the results in Tab. 5 reveal potential performance drops when DDA is applied to ImageNet-W with Swin-T and Swin-B models, suggesting that synthetic data may be less recognizable by the original source models. Although advancements in diffusion techniques could potentially improve outcomes, the consistent gains achieved by SDA indicate that aligning the source model with the synthetic domain offers a convenient and effective solution to enhance performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 513, + 553, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 553, + 525 + ], + "type": "text", + "content": "4.2. Scalability to Other Datasets, Tasks and Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 530, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 602 + ], + "type": "text", + "content": "In addition to standard benchmarks, an effective TTA method should demonstrate its superiority across various dataset scales, task formats, and model architectures. In this section, we assess SDA from these aspects to validate its scalability. Implementation details of each experiment can be found in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Scaling to small datasets. We first evaluate the effectiveness of SDA on scenarios where both source and target domains are small-scale. Specifically, we test SDA on CIFAR-10-C [24] with ResNet-18 [23] as the source classifier. EDM [27] is used to generate synthetic data, and I-DDPM [45] is applied for data alignment. As shown in Tab. 6, SDA consistently outperforms both the source model and DDA, achieving an average accuracy improvement of " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " over DDA." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30508" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 553, + 145 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 145 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 145 + ], + "type": "table", + "html": "
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAvg.
Source23.528.720.841.230.748.248.151.863.663.479.528.056.535.960.745.4
DDA79.480.376.841.356.649.048.977.473.562.680.561.561.257.673.265.3
SDA (Ours)82.683.579.559.661.465.364.676.578.471.588.953.475.464.480.772.4 (+7.1)
", + "image_path": "9dd9c63be1b3ccac0094480d4e651edf5f29c240df937f62552a137ba0667b71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 165, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 87, + 148, + 522, + 160 + ], + "lines": [ + { + "bbox": [ + 87, + 148, + 522, + 160 + ], + "spans": [ + { + "bbox": [ + 87, + 148, + 522, + 160 + ], + "type": "text", + "content": "Table 6. Comparisons of SDA and baselines with ResNet-18 on CIFAR-10-C. SDA shows the best average accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 165, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 165, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 553, + 240 + ], + "type": "table", + "html": "
GaussianShotImpulseDefocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPirelateJEPGAvg
Source9.813.210.022.75.720.422.136.728.453.966.69.724.316.043.825.6
DDA16.621.116.045.040.949.738.736.738.635.055.213.356.656.358.838.6
SDA (Ours)17.021.116.245.543.051.139.938.740.236.357.913.159.158.159.739.8 (+1.2)
", + "image_path": "32badcc958976904a05af786839795a85ea1827db2214a82aaa50ddc91347d8d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 243, + 528, + 255 + ], + "lines": [ + { + "bbox": [ + 80, + 243, + 528, + 255 + ], + "spans": [ + { + "bbox": [ + 80, + 243, + 528, + 255 + ], + "type": "text", + "content": "Table 7. Comparisons of SDA and baselines with DeepLabv3 on PASCAL VOC-C. SDA shows the best average mIOU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 267, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 295, + 363 + ], + "type": "text", + "content": "Scaling to Semantic Segmentation Tasks. We extend our evaluation to dense prediction tasks by using PASCAL-VOC-C [13] as a standard semantic segmentation benchmark with DeepLabv3 [6] as the source segmenter. Dataset Diffusion [41] generates synthetic segmentation data, and FLUX Schnell [29] is used for data alignment. As shown in Tab. 7, SDA achieves the best performance, with an average mIOU improvement of " + }, + { + "bbox": [ + 54, + 267, + 295, + 363 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 54, + 267, + 295, + 363 + ], + "type": "text", + "content": " over DDA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 571 + ], + "type": "text", + "content": "Scaling to multimodal large language models (MLLMs). As an emerging research direction, MLLMs [26, 34] present advanced visual question answering [2] capability. We design a language-based classification task format to test how SDA can help MLLMs on TTA tasks on ImageNet-C: Given an image, ask an MLLM (LLaVA 1.5-7b [34] in our experiments) to choose the correct image class from four provided options. In Tab. 8, the \"Source-Zero\" setting tests the zero-shot results of pretrained LLaVA. The \"Source\" and \"DDA\" settings evaluate a source-data fine-tuned LLaVA while the \"SDA\" setting tests a synthetic-data fine-tuned LLaVA. The fine-tuning task format is the same as that at test time. Pre-trained LLaVA already exhibits strong performance. While DDA improves results in some domains, it does not yield an overall gain compared to source-data fine-tuned LLaVA. In contrast, SDA aligns LLaVA with the synthetic domain, achieving the best accuracy with an improvement of " + }, + { + "bbox": [ + 55, + 367, + 295, + 571 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 55, + 367, + 295, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 584, + 118, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 118, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 118, + 596 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Visualization. To demonstrate how synthetic data fine-tuning in SDA enhances the performance of diffusion-driven TTA methods, we employ Gradient-weighted Class Activation Mapping (Grad-CAM) [59] to visualize the image regions that most influence classification scores across different images and models. As shown in Fig. 5, testing target images with the source model reveals distinct differences in activation maps and the occurrence of incorrect predictions compared to those from source images," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 268, + 551, + 365 + ], + "blocks": [ + { + "bbox": [ + 316, + 268, + 551, + 365 + ], + "lines": [ + { + "bbox": [ + 316, + 268, + 551, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 551, + 365 + ], + "type": "image", + "image_path": "8d9145f445f0dc1c5dc179cde4f7eac7b94f6c5142bae6001f4f5fa5d6454627.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 436 + ], + "type": "text", + "content": "Figure 5. Grad-CAM visualization comparison. The first row shows activation maps for source and target images tested with the source model. The second row displays activation maps for diffusion synthetic images tested with the source model (DDA) and our synthetic-domain model (SDA). SDA aligns closely with the source model's response to source images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 555 + ], + "type": "text", + "content": "underscoring the performance degradation due to domain shifts. Despite using adapted synthetic images, DDA still risks focusing on inappropriate regions and producing incorrect predictions. This highlights the domain misalignment of the synthetic data and source model. In contrast, SDA aligns both the data and model within the same synthetic domain, thereby producing activation maps and predictions that closely resemble those produced by the source model on source images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Data stream sensitivity. In Tab. 9, we test data stream sensitivity using the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings closely aligned with diverse real-world TTA scenarios. In addition to diffusion-driven TTA methods, we include 13 popular traditional TTA methods [4, 5, 11, 12, 18, 38, 40, 47, 60, 66, 68, 70] for comparison. We report the average accuracy of 12 settings using ResNet-50. The results indicate that diffusion-driven TTA methods are preferred in these challenging settings since they are insensitive to different variants of data streams. SDA maintains this insensitivity and showcases the best performance. Detailed comparisons for each setting are provided in the supplementary materials." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30509" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 156 + ], + "type": "table", + "html": "
GaussianShotImpluseDerocusGlassMotionZoomFrostSnowFogBrightnessContrastElasticPixelateJEPGAve:
Source-Zero75.276.679.681.474.583.581.187.088.989.595.684.176.290.188.783.5
Source86.886.088.690.185.593.589.793.695.094.998.189.585.696.294.891.2
DDA92.491.592.588.690.192.288.593.093.189.697.875.889.696.497.091.2
SDA (Ours)94.994.194.691.092.994.091.593.695.292.298.583.692.897.397.693.6 (+2.4)
", + "image_path": "97a0f48091061b06980b81129d5f9eca1daa7ef5c45f5395e294f4aef4ed6333.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 178, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 94, + 159, + 515, + 171 + ], + "lines": [ + { + "bbox": [ + 94, + 159, + 515, + 171 + ], + "spans": [ + { + "bbox": [ + 94, + 159, + 515, + 171 + ], + "type": "text", + "content": "Table 8. Comparisons of SDA and baselines with LLaVA on ImageNet-C. SDA shows the best average accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 178, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 178, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 553, + 228 + ], + "type": "table", + "html": "
SourceTENT [66]ROID [38]NOTE [18]CoTIA [68]TRIBE [60]BN [40]EATA [67]SANTA [5]RMT [11]UnMIX [64]RoTTA [70]LAME [14]UniTTA [12]DDA [17]SDA (Ours)
18.34.26.26.38.89.39.810.210.611.517.423.124.930.329.331.6 (+2.3)
", + "image_path": "368f9c354572416d7aad32827e240c07561a9e52b886d8eafbcaa84c714e8be2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 266, + 296, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 230, + 553, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 553, + 254 + ], + "type": "text", + "content": "Table 9. Data stream sensitivity comparison. We additionally compare SDA with 13 traditional TTA methods on the UniTTA [12] benchmark which contains 12 class/domain balance/imbalance settings. Results are reported as average accuracy across settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 266, + 296, + 319 + ], + "lines": [ + { + "bbox": [ + 56, + 266, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 296, + 319 + ], + "type": "table", + "html": "
ComponentResNet-50Swin-TConvNeXt-T
DDA [17]29.740.044.2
+ Conditional Data Generation30.440.144.6
+ Unconditional Data Alignment32.542.547.0
", + "image_path": "a6baf7f944ce4e6048b64e70c9e58a1f92cee78fe861029a74a267c584b005af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 356, + 156, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 156, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 156, + 368 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 374, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 601 + ], + "type": "text", + "content": "Components. We examine the impact of two key components in our SDA framework's Mix of Diffusion technique in Tab. 10: (1) synthetic data generation using the conditional diffusion model and (2) synthetic data alignment using the unconditional diffusion model. Fine-tuning source models with synthetic data generated solely by the conditional diffusion model (+ Conditional Data Generation) yields only marginal gains. As discussed in Sec. 3.4, this limited gain arises from the domain misalignment between conditional and unconditional diffusion models. Specifically, using only conditional synthetic data results in models aligned with the conditional diffusion domain, whereas the test data belongs to the unconditional diffusion domain. Therefore, further aligning the synthetic data through the unconditional diffusion model (+ Unconditional Data Alignment) leads to significant performance gains over the baseline DDA. This demonstrates that bridging the misalignment of different diffusion domains is essential for the success of our SDA framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "Number of fine-tuning images. We examine the impact of different numbers of images " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " used during synthetic-domain model fine-tuning in Tab. 11. Interestingly, even with only one image per class " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "(N = 1\\mathrm{K})" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": ", SDA still significantly outperforms DDA. This finding suggests a key attribute of the fine-tuning process: source models are primarily learning to adapt to the synthetic domain itself, rather than acquiring class-specific knowledge. Increasing the number of images helps the fine-tuning process capture the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 266, + 543, + 345 + ], + "blocks": [ + { + "bbox": [ + 55, + 322, + 296, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 344 + ], + "type": "text", + "content": "Table 10. Impact of different components in SDA. Results are evaluated on ImageNet-C." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 266, + 543, + 345 + ], + "lines": [ + { + "bbox": [ + 326, + 266, + 543, + 345 + ], + "spans": [ + { + "bbox": [ + 326, + 266, + 543, + 345 + ], + "type": "table", + "html": "
MethodNumber (N)ResNet-50Swin-TConvNeXt-T
DDA [17]029.740.044.2
SDA1K31.942.345.5
10K31.942.546.6
50K (default)32.542.547.0
100K32.642.246.8
", + "image_path": "38ab6e74e54ba495166171840608c00b4a8a14526a1810699a2d382709e0b877.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 347, + 553, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 347, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 553, + 370 + ], + "type": "text", + "content": "Table 11. Impact of different numbers of fine-tuning images. Results are evaluated on ImageNet-C." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "content": "synthetic domain more accurately, thereby enhancing performance. Based on a balance between performance improvement and image generation resources, we select " + }, + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "inline_equation", + "content": "N = 50\\mathrm{K}" + }, + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "content": " as our default experimental setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 453, + 388, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 388, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 388, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we proposed Synthetic-Domain Alignment (SDA), a novel test-time adaptation (TTA) framework that simultaneously aligns the domains of the source model and target data with the synthetic domain of a diffusion model. For the source model, SDA introduces a Mix of Diffusion (MoD) technique, which generates synthetic data to adapt the source model to a synthetic-domain model. MoD involves a conditional diffusion model for data generation and an unconditional diffusion model for data alignment. For the target data, SDA utilizes the aforementioned unconditional diffusion model to project the target data to synthetic data. As the domains of the model and data are aligned, SDA converts the cross-domain TTA task into an easier in-domain prediction task. Compared to existing diffusion-driven TTA methods, SDA significantly mitigates the source-synthetic domain misalignment issue. Compared to traditional TTA methods, SDA maintains insensitivity to different data streams. Extensive experiments across classifiers, segmenters, and MLLMs indicate that SDA achieves enhanced domain alignment and superior performance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30510" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "content": "The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant U24B20173." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 115, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 115, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 115, + 163 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 170, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 170, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 296, + 214 + ], + "type": "text", + "content": "[1] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets computer vision: Efficient data generation for urban driving scenes. IJCV, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 216, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 248 + ], + "type": "text", + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 250, + 296, + 293 + ], + "type": "text", + "content": "[3] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification. TMLR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 295, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 327 + ], + "type": "text", + "content": "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, 2022. 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "type": "text", + "content": "[5] Goirik Chakrabarty, Manogna Sreenivas, and Soma Biswas. Santa: Source anchoring network and target alignment for continual test time adaptation. TMLR, 2023. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 363, + 296, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 395 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 397, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 440 + ], + "type": "text", + "content": "[7] Yuhua Chen, Wen Li, Xiaoran Chen, and Luc Van Gool. Learning semantic segmentation from synthetic data: A geometrically guided input-output adaptation approach. In CVPR, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 442, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 296, + 474 + ], + "type": "text", + "content": "[8] MMPreTrain Contributors. Openmmlab's pre-training toolbox and benchmark. https://github.com/open-mmlab/mmpretrain, 2023.4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 476, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 508 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 510, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 532 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 533, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 296, + 565 + ], + "type": "text", + "content": "[11] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, 2023. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 296, + 610 + ], + "type": "text", + "content": "[12] Chaoqun Du, Yulin Wang, Jiayi Guo, Yizeng Han, Jie Zhou, and Gao Huang. Unitta: Unified benchmark and versatile framework towards realistic test-time adaptation. arXiv preprint arXiv:2407.20080, 2024. 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 613, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 645 + ], + "type": "text", + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "text", + "content": "[14] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "text", + "content": "[15] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 98, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 554, + 140 + ], + "type": "text", + "content": "[17] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 143, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 554, + 185 + ], + "type": "text", + "content": "[18] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. Note: Robust continual test-time adaptation against temporal correlation. In NeurIPS, 2022. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 188, + 554, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 554, + 220 + ], + "type": "text", + "content": "[19] Jiayi Guo, Chaoqun Du, Jiangshan Wang, Huijuan Huang, Pengfei Wan, and Gao Huang. Assessing a single image in reference-guided image synthesis. In AAAI, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 222, + 554, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 554, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 554, + 286 + ], + "type": "text", + "content": "[20] Jiayi Guo, Hayk Manukyan, Chenyu Yang, Chaofei Wang, Levon Khachatryan, Shant Navasardyan, Shiji Song, Humphrey Shi, and Gao Huang. Faceclip: Facial image-to-video translation via a brief text description. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 289, + 554, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 554, + 334 + ], + "type": "text", + "content": "[21] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Humphrey Shi, Gao Huang, and Shiji Song. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 335, + 554, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 554, + 378 + ], + "type": "text", + "content": "[22] Jiayi Guo, Xingqian Xu, Yifan Pu, Zanlin Ni, Chaofei Wang, Manushree Vasu, Shiji Song, Gao Huang, and Humphrey Shi. Smooth diffusion: Crafting smooth latent spaces in diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 380, + 554, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 380, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 316, + 380, + 554, + 412 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 415, + 554, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 447 + ], + "type": "text", + "content": "[24] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 449, + 554, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 472 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 474, + 554, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 554, + 495 + ], + "type": "text", + "content": "[26] Gao Huang. Dynamic neural networks: advantages and challenges. National Science Review, 2024. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 497, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 554, + 529 + ], + "type": "text", + "content": "[27] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 532, + 554, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 554, + 563 + ], + "type": "text", + "content": "[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 566, + 554, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 554, + 587 + ], + "type": "text", + "content": "[29] Black Forest Labs. Flux. https:// / blackforestlabs.ai/, 2024. 2, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 590, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 621 + ], + "type": "text", + "content": "[30] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 624, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 656 + ], + "type": "text", + "content": "[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLR Workshops, 2017. 1, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[32] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, 2023. 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30511" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 294, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 294, + 129 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 173 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "text", + "content": "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "text", + "content": "[37] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 241, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 275 + ], + "type": "text", + "content": "[38] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, 2024. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[39] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In CoRL, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[40] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "text", + "content": "[41] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[42] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "text", + "content": "[43] Zanlin Ni, Yulin Wang, Renping Zhou, Yizeng Han, Jiayi Guo, Zhiyuan Liu, Yuan Yao, and Gao Huang. Enat: Rethinking spatial-temporal interactions in token-based image synthesis. In NeurIPS, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 488, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 532 + ], + "type": "text", + "content": "[44] Zanlin Ni, Yulin Wang, Renping Zhou, Rui Lu, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Yuan Yao, and Gao Huang. Adanat: Exploring adaptive policy for token-based image generation. In ECCV, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 533, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 555 + ], + "type": "text", + "content": "[45] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 555, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 589 + ], + "type": "text", + "content": "[46] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In ICML, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[47] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 634, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 294, + 656 + ], + "type": "text", + "content": "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[49] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3d models. In ICCV, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "text", + "content": "[50] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-tta:" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Test-time adaptation of discriminative models via generative feedback. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[51] Yifan Pu, Zhuofan Xia, Jiayi Guo, Dongchen Han, Qixiu Li, Duo Li, Yuhui Yuan, Ji Li, Yizeng Han, Shiji Song, et al. Efficient diffusion transformer with step-wise dynamic attention mediators. In ECCV, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[52] Yifan Pu, Yiming Zhao, Zhicong Tang, Ruihong Yin, Haoxing Ye, Yuhui Yuan, Dong Chen, Jianmin Bao, Sirui Zhang, Yanbin Wang, et al. Art: Anonymous region transformer for variable multi-layer transparent image generation. arXiv preprint arXiv:2502.18364, 2025. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[54] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "text", + "content": "[55] Artem Rozantsev, Vincent Lepetit, and Pascal Fua. On rendering synthetic images for training an object detector. Computer Vision and Image Understanding, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "text", + "content": "[56] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In ICLR, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "text", + "content": "[57] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[58] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "type": "text", + "content": "[60] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In AAAI, 2024. 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "text", + "content": "[61] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "text", + "content": "[62] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In CVPR, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[63] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "text", + "content": "[64] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozorgtabar. Un-mixing test-time normalization statistics: Combating label temporal correlation. arXiv preprint arXiv:2401.08328, 2024. 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[65] Yun-Yun Tsai, Fu-Chen Chen, Albert YC Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. CVPR, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30512" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 285 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[67] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[69] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. Nerf-supervision: Learning dense object descriptors from neural radiance fields. In ICRA, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 252 + ], + "type": "text", + "content": "[70] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In CVPR, 2023. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[71] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30513" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_content_list.json b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..572982dde1d774f4ef67e2ed8981d6ab7c6a3bd7 --- /dev/null +++ b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_content_list.json @@ -0,0 +1,1510 @@ +[ + { + "type": "text", + "text": "Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "You Shen $^{1\\dagger}$ , Zhipeng Zhang $^{2\\dagger}$ , Xinyang Li $^{1}$ , Yansong Qu $^{1}$ , Yu Lin $^{1}$ , Shengchuan Zhang $^{1}$ , Liujuan Cao $^{1*}$ $^{1}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University \n $^{2}$ School of Artificial Intelligence, Shanghai Jiao Tong University", + "bbox": [ + 99, + 202, + 915, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Representing 3D scenes from multiview images is a core challenge in computer vision and graphics, which requires both precise rendering and accurate reconstruction. Recently, 3D Gaussian Splating (3DGS) has garnered significant attention for its high-quality rendering and fast inference speed. Yet, due to the unstructured and irregular nature of Gaussian point clouds, ensuring accurate geometry reconstruction remains difficult. Existing methods primarily focus on geometry regularization, with common approaches including primitive-based and dual-model frameworks. However, the former suffers from inherent conflicts between rendering and reconstruction, while the latter is computationally and storage-intensive. To address these challenges, we propose CarGS, a unified model leveraging Contribution-adaptive regularization to achieve simultaneous, high-quality rendering and surface reconstruction. The essence of our framework is learning adaptive contribution for Gaussian primitives by squeezing the knowledge from geometry regularization into a compact MLP. Additionally, we introduce a geometry-guided densification strategy with clues from both normals and Signed Distance Fields (SDF) to improve the capability of capturing high-frequency details. Our design improves the mutual learning of the two tasks, meanwhile its unified structure doesn't require separate models as in dual-model based approaches, guaranteeing efficiency. Extensive experiments demonstrate CarGS's ability to achieve state-of-the-art (SOTA) results in both rendering fidelity and reconstruction accuracy while maintaining real-time speed and minimal storage size.", + "bbox": [ + 88, + 340, + 485, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 808, + 223, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) and surface reconstruction from multi-view images are both challenging and critical", + "bbox": [ + 89, + 834, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c957c27f9c1ce164332089e34e54cbfc552ea924e3a6143e1a1d5cda2cba9fa9.jpg", + "image_caption": [ + "Figure 1. The performance of CarGS in TnT datasets, achieving high-quality rendering and reconstruction in a unified model. The storage size of the model is represented by the radius of the circle." + ], + "image_footnote": [], + "bbox": [ + 514, + 306, + 906, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks in the fields of computer vision and graphics. Both academia and industry are actively pursuing unified scene representations that enable high-quality rendering and accurate geometry reconstruction concurrently to address the growing demands of applications in areas such as augmented reality (AR), virtual reality (VR), robotics, and autonomous driving. However, simultaneously achieving SOTA performance in both rendering and reconstruction with a unified model is non-trivial, as an inherent conflict exists between the two tasks [3, 7, 9].", + "bbox": [ + 509, + 566, + 908, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Revisiting recent approaches, we identify two solutions aimed at mitigating the conflict. The first branch, which we refer to as \"primitive-based\", introduces additional geometric regularization to 3D Gaussian primitives to improve surface reconstruction. For example, Sugar [7] regularizes 3D Gaussians to align more closely with surfaces. Although this approach ensures parameter efficiency, it leads to a misaligned prioritization between appearance and geometry, eventually degrading rendering performance. The second branch, termed \"dual-model based\", seeks to guide the learning of one task through information derived from the other. For instance, GSDF [30] introduces a dual-model", + "bbox": [ + 509, + 719, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 107, + 875, + 238, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal Contribution.", + "bbox": [ + 109, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16346", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "architecture that jointly supervises the 3D Gaussian Splating model for rendering and the SDF model for reconstruction. However, this strategy necessitates separate models for training and inference, resulting in much larger computational and storage costs. Consequently, a natural question arises: Is it possible to develop an efficient framework (fast speed with small storage) that is also effective (SOTA performances on both tasks)?", + "bbox": [ + 88, + 90, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before answering this question, we first present an intriguing observation that motivates us to pursue the \"unified goal\" from the perspective of learning adaptive contribution of Gaussian primitives. Specifically, with the baseline model PGSR [3] trained on the \"Barn\" scene from the Tanks and Temples dataset [11], we first calculate the Chamfer Distance (CD) from the center of each Gaussian to the ground truth surface. Then, as shown in Fig. 2, when retaining $50\\%$ of the Gaussian points by discarding those with larger CD values, we find that reconstruction quality (F1) is surprisingly maintained, while rendering quality (PSNR) experiences a sharp decline. Conversely, when additional Gaussian points with smaller CDs are dropped, reconstruction quality begins to deteriorate, while rendering quality remains relatively stable. This observation suggests that, 1) despite geometric regularization encouraging Gaussian primitives to closely adhere to the surface [3, 9], a significant number of floaters persist in the scene. 2) These floaters contribute less to reconstruction but are essential for rendering quality. More precisely, each Gaussian primitive contributes differently to each task.", + "bbox": [ + 91, + 215, + 486, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, an intuitive idea is to build a unified model with adaptive contributions for each Gaussian primitive. This \"unified\" structure contrasts with dual-model designs, significantly reducing storage and computational demands from the outset. The \"adaptiveness\" mitigates conflicts in Gaussian primitive contributions when jointly optimizing for rendering and reconstruction tasks. To this end, we propose CarGS: a unified Gaussian Splitting model based on Contribution-adaptive regularization. Firstly, we squeeze geometry regularization knowledge into a lightweight residual structure to adaptively adjust each Gaussian primitive's contribution for rendering and reconstruction. Then, to further capture geometric details, we introduce a geometry-guided densification strategy that controls Gaussian expansion using clues from both normals and SDF. Experimental results show that our method achieves state-of-the-art reconstruction performance within a unified model, without compromising rendering quality. Remarkably, our model consumes only $9\\%$ of the storage and $40\\%$ of the training time required by GSDF [30] (dual-model based) while running $100\\times$ faster. Moreover, its rendering and reconstruction performance surpasses both GSDF and PGSR [3] (primitive-based). We believe our approach establishes a strong baseline for unified rendering and reconstruction.", + "bbox": [ + 91, + 537, + 486, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/197d7f31c35dd377593f5c548dd8e11e294c31fff8e4ad6455d0c059edf4cdbd.jpg", + "image_caption": [ + "Figure 2. The performance of the Baseline model [3] across various $\\mathrm{CD\\%}$ thresholds demonstrates that each Gaussian primitive contributes differently to specific tasks." + ], + "image_footnote": [], + "bbox": [ + 516, + 85, + 903, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows: 1) We identify the key bottleneck in developing a unified, efficient, and effective framework for both rendering and reconstruction. 2) We propose to learn the adaptive contribution of each Gaussian primitive, addressing the inherent conflict in creating a unified framework for rendering and reconstruction. 3) We introduce a geometry-guided densification strategy to enhance the model's ability to capture fine details. 4) Our unified framework simultaneously achieves high-quality rendering and reconstruction while consuming significantly less storage than previous methods (see Fig. 1).", + "bbox": [ + 511, + 306, + 906, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 488, + 651, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Novel View Synthesis", + "text_level": 1, + "bbox": [ + 511, + 513, + 714, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis aims to generate new perspectives of a scene using a limited collection of existing views.", + "bbox": [ + 511, + 536, + 906, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NVS based on Neural Radiance Fields. A pivotal advancement in this field was the introduction of Neural Radiance Fields (NeRF) [16]. NeRF employs MLPs to map 3D spatial locations along rays to color and density values, which are subsequently aggregated through neural volumetric rendering to produce high-quality images. Subsequent research has focused on enhancing the capabilities of NeRF models through feature-grid representations [5, 12, 20, 24] and improving rendering speed using baking techniques [8, 22, 29]. Recently, NeRF has been adapted to address challenges such as anti-aliasing [1] and the modeling of unbounded scenes [2, 34].", + "bbox": [ + 511, + 566, + 908, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NVS based on 3D Gaussian Splatting. Recently, 3D Gaussian Splatting [10] has emerged as a transformative approach in neural rendering by utilizing anisotropic 3D Gaussians as primitives. This method sorts Gaussians by depth and rasterizes them onto a 2D screen through alpha blending, resulting in impressive rendering quality and real-time speed. Subsequent research has enhanced 3D Gaussian Splatting by improving rendering quality [4, 21, 31] and expanding its applicability to large-scale scenes [15, 26, 27, 33]. For example, Mip-splatting [31] introduces", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16347", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "anti-aliasing techniques to mitigate pixelation and blurring in high-resolution renders, thereby enhancing visual sharpness and consistency across viewpoints. Scaffold-GS [15] employs anchor points to manage local 3D Gaussians and dynamically predicts their attributes based on viewing direction and distance, thereby optimizing the adaptive rendering process.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these NVS methods excel in view synthesis, they inherently overlook the accurate reconstruction of scene geometry, resulting in blurred volumetric density fields that impede the extraction of high-quality surfaces.", + "bbox": [ + 89, + 198, + 480, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Neural Surface Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 271, + 362, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The success of neural rendering has sparked significant interest in neural surface reconstruction [6, 13, 19, 25, 28]. By leveraging coordinate-based neural networks, these methods encode scene geometry using occupancy fields or SDF values, facilitating a smooth and differentiable representation of 3D structures. Though NeRF-based methods with volume rendering yield smooth, complete surfaces, they struggle with fine detail capture and slow optimization. To address this, recent work [13] incorporates multi-resolution hashed grids from Instant-NGP [17], enhancing representation and achieving SOTA results. Yet, large-scale scene reconstruction remains inefficient; e.g., Neuralangelo [13] takes 128 GPU hours to reconstruct a single Tanks and Temples scene [11].", + "bbox": [ + 89, + 294, + 482, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the impressive efficiency of 3DGS in NVS, recent research has investigated the use of anisotropic 3D Gaussians as foundational primitives for scene geometry representation. SuGaR [24] aligns 3D Gaussians with potential surfaces by approximating them as 2D planar primitives with binary opacity. 2D Gaussian Splatting [9] replaces 3D Gaussians with 2D counterparts, resulting in more accurate ray-splat intersections and regularizing depth maps to achieve a denser distribution of Gaussian primitives. PGSR [3] employs both single-view and multi-view geometric regularization to attain globally consistent geometry. Despite these advancements in neural surface reconstruction, a fidelity gap persists, as the explicit regularization of Gaussian primitives constrains rendering quality. To address this limitation, GSDF [30] introduces a dual-model framework that separates rendering (3DGS) and geometry (SDF) tasks, utilizing joint supervision during training to enhance both reconstruction and rendering. However, this design incurs additional training and storage costs due to the requirement of maintaining two independent models. Moreover, the geometry reconstruction branch is not capable of real-time inference.", + "bbox": [ + 89, + 506, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from existing approaches, our paper presents a unified model designed to simultaneously achieve real-time and robust rendering and reconstruction tasks without compromising geometry accuracy and rendering quality.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section details the proposed framework. Particularly, Sec. 3.1 presents preliminary studies on 3D Gaussian Splating and its representative anchor-based paradigm. Then in Sec. 3.2, we analyze the problem of contribution conflicts to identify the most significant attributes for geometry reconstruction. Sec. 3.3 introduces Lite-Geo, a lightweight residual module that adapts the geometry contribution of Gaussian primitives. Based on Lite-Geo, we further propose a geometry-guided densification strategy in Sec. 3.4 that leverages geometric cues to align Gaussians with the surface, enhancing both structural fidelity and representational capacity. Finally, in Sec. 3.5, we introduce geometric regularization to guide the optimization of our modules, which improves the quality of geometry reconstruction.", + "bbox": [ + 511, + 114, + 906, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 333, + 643, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. 3DGS models the scene with a set of explicit Gaussians that inherit the differentiable properties of volumetric representations, allowing efficient rendering via tile-based rasterization. Initially, each point generated by the Structure-from-Motion (SfM) serves as the position (mean) $\\mu$ of a corresponding 3D Gaussian:", + "bbox": [ + 511, + 356, + 905, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 452, + 903, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x$ denotes an arbitrary position in the 3D scene, while $\\Sigma$ represents the covariance matrix of the 3D Gaussian. $\\Sigma$ is constructed using a scaling matrix $S$ and a rotation matrix $R$ to maintain positive semi-definiteness.", + "bbox": [ + 511, + 479, + 905, + 540 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma = R S S ^ {T} R ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 546, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each 3D Gaussian is assigned with an opacity value $\\alpha$ , which is multiplied with $G(x)$ to adjust its contribution weight to rendering and reconstruction during the blending process. 3DGS renders scenes efficiently through tile-based rasterization instead of traditional ray marching. Each 3D Gaussian $G(x)$ is projected as a 2D Gaussian onto the image plane, after which a tile-based rasterizer sorts and blends the 2D Gaussians using $\\alpha$ -blending:", + "bbox": [ + 511, + 571, + 905, + 693 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF \\left(x ^ {\\prime}\\right) = \\sum_ {i \\in N} f _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} (x) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 700, + 903, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x'$ is the queried pixel location in the image plane and $N$ denotes the number of sorted 2D Gaussians associated with the queried pixel. For a Gaussian with 3D center $(x_i, y_i, z_i)$ and color $c_i$ , $f_i$ is assigned to $c_i$ in the rendering task, and to $z_i$ for depth blending in the reconstruction task. Anchor-based Gasussian Splatting. Our goal is to simultaneously achieve high-quality rendering and reconstruction while maintaining efficiency. In our work, we adopt anchor-based GS modeling as our baseline due to its balanced efficiency and effectiveness. A representative work in this", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16348", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/822141261f4c3ed97ed83e84d9c01cd1382bfe0b6bb8c7a0f8a883f036b1d68c.jpg", + "image_caption": [ + "Figure 3. The illustration of contribution conflict between rendering and reconstruction tasks." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 475, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "paradigm, Scaffold-GS [15], enhances original 3DGS by improving scene structure fidelity and incorporating view clues. Specifically, it constructs a hierarchical 3D Gaussian framework with anchor points that encode local scene details and generate local neural Gaussians,", + "bbox": [ + 89, + 227, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mu_ {0}, \\dots , \\mu_ {k - 1} \\right\\} = x + \\left\\{o _ {0}, \\dots , o _ {k - 1} \\right\\} \\cdot \\gamma , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 313, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $o_k \\in \\mathbb{R}^3$ denotes the learned offsets, and $\\gamma$ represents the scaling factor. For each anchor, the optimized attributes (color, center, variance, and opacity) of the $k$ neural Gaussians are decoded through an MLP, $M$ , based on the anchor feature $\\mathbf{f}$ , as well as the distance $\\delta$ and direction $\\theta$ from the camera to the anchor point. Specifically, the opacity values for neural Gaussians generated from an anchor point are,", + "bbox": [ + 89, + 339, + 483, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {0}, \\dots , \\alpha_ {k - 1} = M (\\mathbf {f}, \\delta , \\theta), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 455, + 482, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Analysis", + "text_level": 1, + "bbox": [ + 89, + 503, + 192, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we analyze the underlying reasons for the rendering performance degradation when geometry regularization is applied to 3DGS primitives. Then, we conduct experiments to analyze the effect of Gaussian's internal attributes to geometry reconstruction.", + "bbox": [ + 89, + 526, + 483, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis of Contribution Conflicts. In Fig 3, we consider a simple demonstration of alpha-blending along a ray intersecting three Gaussian primitives. As shown, $G_{2}$ is close to the real surface, $G_{1}$ and $G_{3}$ are farther away. According to Eq. 3, the color $C(x')$ and depth $D(x')$ corresponding to the queried pixel along the ray is calculated as follows:", + "bbox": [ + 89, + 601, + 483, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC \\left(x ^ {\\prime}\\right) = c _ {1} w _ {1} + c _ {2} w _ {2} + c _ {3} w _ {3}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 700, + 482, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD \\left(x ^ {\\prime}\\right) = z _ {1} w _ {1} + z _ {2} w _ {2} + z _ {3} w _ {3} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 727, + 482, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} (x) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 750, + 482, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ represents the contribution of the Gaussian and serves as the weighting factor for ray integration. When only applying rendering regularization, $w_{1}$ and $w_{3}$ tend to increase to capture richer textures, as proved in Fig. 2. Conversely, when geometry regularization is applied, the values of $w_{1}$ and $w_{3}$ tend to decrease due to $z_{1}$ and $z_{3}$ deviating from the SDF 0-level set, while $w_{2}$ increases to provide", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/27e4aee1effbdb9c39e043372ef91abf1e7decb08b999f10790f096ed1d005b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CovOpacityBaseline
F1↑0.460.640.65
PSNR↑23.8428.7228.76
", + "bbox": [ + 516, + 89, + 905, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. The importance of different attributes in Gaussian's contribution for the geometry reconstruction task.", + "bbox": [ + 511, + 161, + 906, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more consistent depth information. This shift in the contribution weights $w_{i}$ creates conflicts, which can lead to a decline in rendering performance when geometry regularization is simply added when training the rendering task [3, 9]. Analysis of Attributes. As discussed, the degradation in rendering performance after using geometry regularization stems from contribution conflicts. To further investigate this issue, we analyze which internal attributes of the Gaussian's contribution are most susceptible to the effects of geometry regularization. As shown in Eq. 8, Gaussian's contribution $w_{i}$ is defined by its opacity $\\alpha$ and Gaussian sampling value $G(x)$ . Lu et al. [15] emphasize that stable Gaussian positions are crucial for maintaining scene structure, suggesting that adjustments to $\\mu$ in Eq. 1 remain relatively minor during optimization. Consequently, changes in $G(x)$ are more likely to arise from covariance in the covariance $\\Sigma$ in Eq. 1.", + "bbox": [ + 511, + 215, + 906, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, to investigate the roles of $\\alpha$ (opacity) and $\\Sigma$ (covariance) within the geometry regularization process, we conduct analysis experiments in Tab. 1. Specifically, with the baseline model PGSR [3] on TnT's \"Barn\" scene, we respectively detach pacity and covariance from the geometry supervision to stop their gradigents and then assess their independent effect on the performance of geometry reconstruction. As shown in Tab. 1, excluding the gradient of opacity from geometry optimization shows minimal impact on reconstruction quality, while excluding covariance's gradient severely impairs geometry learning. Hence, covariance plays the essential role in regulating the contributions of Gaussian primitives, making it important to tailor covariance for both high-quality rendering and reconstruction.", + "bbox": [ + 511, + 459, + 908, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 3DGS with Adaptive Contribution", + "text_level": 1, + "bbox": [ + 511, + 681, + 815, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 4, existing approaches for realizing rendering and reconstruction are categorized into primitive based [3, 9] and dual-model based [30] frameworks. The former supervises Gaussian primitives with both rendering and geometry losses, leading to conflicting contributions and, consequently, performance degradation. In contrast, dual-model based methods train separate models for each task, i.e., a 3DGS model for rendering and an SDF model for reconstruction. While enhancing both tasks, it requires substantial computational and storage resources (see Tab. 4). Building on the above analysis, we are motivated to propose a unified and efficient contribution-adaptive 3DGS architecture to simultaneously achieve state-of-the-art per", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16349", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f3419967cdbd7fbb8ca1a8e085d1ca851b00fd85c8e5fb7b362ed98f33a1d8f6.jpg", + "image_caption": [ + "(a) Primitive based regularization" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 477, + 178 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c8502415539db21926705ad4e71172cad425cd7fedf06e7db4563b06dcfb470b.jpg", + "image_caption": [ + "(b) Dual-model based regularization" + ], + "image_footnote": [], + "bbox": [ + 96, + 200, + 478, + 321 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2a1b85d16916cd4cbd73b2ed9f04c6b076ed2339d32bc1481f1469a00967b491.jpg", + "image_caption": [ + "(c) Contribution-adaptive regularization", + "Figure 4. The illustration of the geometry regularization GS framework: (a) Primitive based regularization paradigm. (b) Dual-model-based regularization paradigm. (c) the proposed contribution adaptive regularization paradigm." + ], + "image_footnote": [], + "bbox": [ + 96, + 348, + 478, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formances by alleviating the adverse effects on rendering when using geometry regularization.", + "bbox": [ + 89, + 549, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geo-MLP. In Sec. 3.2, we have identified covariance as the primary factor contributing to conflicts in geometry and rendering. Thus, we propose decoupling Gaussian contributions by predicting an additional covariance parameter dedicated to geometry reconstruction, while retaining the original covariance for rendering. As outlined in Eq. 2, covariance is decomposed into scaling and rotation matrices. Therefore, we first attempt to introduce an auxiliary MLP that takes the anchor feature $\\mathbf{f}$ and view direction as inputs, predicting the scaling and rotation matrix. The predicted geometry-related scaling and rotation are supervised by reconstruction loss, as shown in Fig. 4 (c). However, nothing is as simple as it seems. As visualized in Fig. 5 (top), this straightforward design results in overfitting in geometry estimation. This challenge arises because conventional geometric regularization primarily enforces continuity and consistency, but lacks accurate supervision for depth, which is predominantly informed by rendering-based supervision.", + "bbox": [ + 89, + 580, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lite-Geo. To address this challenge, we introduce Lite-Geo, designed to inherit the Gaussian primitive's contribution from the rendering task, which encodes inherent", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "depth information after rendering optimization. This implicit depth information helps mitigate overfitting brought by the sole geometry reconstruction loss applied to GeoMLP. More precisely, we structure the learning of covariance for geometry reconstruction as a residual update, where it is obtained by adding the offset predicted by GeoMLP $(M_{\\Sigma}^{geo})$ to the rendering covariance.", + "bbox": [ + 511, + 90, + 903, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny = M _ {\\Sigma} ^ {r g b} \\left(\\mathbf {f}, \\theta ; \\phi_ {1}\\right) + \\triangle y, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 210, + 903, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\triangle y = M _ {\\Sigma} ^ {g e o} (\\mathbf {f}, \\theta ; \\phi_ {2}), \\phi_ {2} = \\lambda \\phi_ {1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 250, + 903, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{1}$ and $\\phi_{2}$ denote the parameters of $M_{\\Sigma}^{rgb}$ and $M_{\\Sigma}^{geo}$ , respectively. $y$ and $\\triangle y$ are 7-dimensional vectors, with the first three dimensions representing scale and the remaining four representing rotation through quaternions. This approach allows the depth information in rendering covariance to effectively inform the depth estimation in geometry reconstruction. To further reduce noise and enhance convergence, we initialize $\\phi_{1}$ based on $\\phi_{2}$ using a controllable scaling factor $\\lambda$ . Specifically, the rendering task is pre-trained independently for several epochs before incorporating the Lite-Geo module.", + "bbox": [ + 511, + 273, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 5, this residual structure effectively suppresses geometry overfitting. It also constrains outliers with high Chamfer Distance values that negatively impact rendering and reconstruction quality.", + "bbox": [ + 511, + 441, + 905, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Geometry-guided Densification", + "text_level": 1, + "bbox": [ + 511, + 512, + 790, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further enhance the distribution of Gaussian primitives through a geometry-guided densification strategy. Our motivation stems from the observation that previous geometry-based densification methods [30] rely solely on SDF values to control Gaussian growing, which limits their ability to capture fine details in high-frequency regions and often results in blurred rendering and overly smooth reconstructions. As shown in Fig. 6, to better accommodate high-frequency details, we introduce normal vector clues alongside the original SDF information to guide Gaussian growing. Specifically, for a Gaussian anchor located at position $x$ and the SDF value within a certain range $\\theta$ , we evaluate its normal difference with the normal map by $n = N_d(x) - N(x)$ . The SDF is obtained from the depth map by $s = D(x) - Z(x)$ . Detailed calculations of normals and SDF are provided in the supplementary. We then define the criteria for Gaussian growing as follows:", + "bbox": [ + 511, + 534, + 906, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {g} = \\nabla_ {g} + \\omega_ {g} (\\zeta (s) \\cdot (\\omega_ {n} n \\text {i f} \\zeta (s) < \\theta \\text {e l s e} 1)), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 803, + 906, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta (s) = \\exp \\left(- \\frac {s ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 830, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla_g$ is the averaged gradient of the Gaussian primitives accumulated over several training iterations. $\\zeta (s)$", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16350", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ed157a0b46cfb3cda6752412c430330136b34ef908d9023151d169b495e29da.jpg", + "image_caption": [ + "Figure 5. The illustration of geometry overfitting. If the Geo-MLP is solely associated with geometry reconstruction, notable overfitting is observed. By leveraging the approximate depth knowledge provided by rendering, Lite-Geo can effectively reduce overfitting and suppress noise where the Chamfer Distance exceeds the threshold." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 903, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c8d625293b8bff845f906d89a8f9b628aa546c578f3574b0ab12b0bd5167404.jpg", + "image_caption": [ + "Figure 6. The illustration of high-frequency details, along with the proposed densification strategy which facilitates Gaussian growing with clues from both normals and SDF." + ], + "image_footnote": [], + "bbox": [ + 112, + 313, + 460, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is a Gaussian function that transforms the SDF to a positive number, which monotonically decreases as the distance from the zero level set increases. The parameter $\\omega_{g}$ controls the influence of geometric guidance. New Gaussian primitives are generated when $\\epsilon_{g}$ exceeds a predefined threshold.", + "bbox": [ + 89, + 473, + 483, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Optimization", + "text_level": 1, + "bbox": [ + 89, + 556, + 230, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relying solely on photometric loss to optimize 3DGS introduces noise into the output, as multi-view 3D reconstruction lacks sufficient constraints [2, 34] for geometric continuity. To improve the recovery of shape detail, we adopt the geometry regularization from PGSR [3], incorporating additional plane constraint and cross-view constraint. To ensure the paper is self-contained, we provide a detailed explanation of these two regularization terms.", + "bbox": [ + 89, + 579, + 482, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Plane Constraint. To ensure local depth-normal consistency, each pixel and its neighbors should define a relatively smooth plane. Using the depths of a pixel and its neighboring pixels, we can generate a normal map, which is then optimized to align with the rendered normal map, thereby enforcing geometric consistency. The loss based on the plane constraint is defined as follows:", + "bbox": [ + 89, + 700, + 482, + 805 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p l a n e}} = \\frac {1}{| \\mathcal {W} |} \\sum_ {x \\in \\mathcal {W}} \\| N _ {d} (x) - N (x) \\| _ {1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 810, + 482, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$N_{d}(x)$ and $N(x)$ represent the normal derived from the depth gradient and the normal of the splats, respectively. $\\mathcal{W}$ denotes the image pixels.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-view Constraint. Plane constraint ensures local depth-normal consistency but lacks cross-view consistency. Hence, cross-view constraintsd should be introduced to improve global geometric consistency. Specifically, a pixel $x_{r}$ in the reference frame, associated with a normal $n_r$ and distance $d_{r}$ , can be transformed to a corresponding pixel $x_{n}$ in a neighboring frame with the homography matrix $H_{rn}$ . After forward projection from the reference frame to the neighboring frame, a backward projection $H_{nr}$ is applied. Then the consistency is ensured by minimizing the forward-backward projection error:", + "bbox": [ + 511, + 316, + 906, + 482 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c r o s s} = \\frac {1}{| \\mathcal {W} |} \\sum_ {x _ {r} \\in \\mathcal {W}} \\| x _ {r} - H _ {n r} H _ {r n} x _ {r} \\|. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 493, + 906, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Final Loss. Finally, we optimize our model by minimizing,", + "bbox": [ + 511, + 541, + 905, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\alpha \\mathcal {L} _ {d} + \\beta \\mathcal {L} _ {n} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 569, + 903, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c$ is the RGB reconstruction loss that combines L1 loss with D-SSIM from [10], while $\\mathcal{L}_{\\mathrm{plane}}$ and $\\mathcal{L}_{\\mathrm{cross}}$ represent the geometry regularization terms.", + "bbox": [ + 511, + 597, + 905, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 657, + 648, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 681, + 702, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To validate the effectiveness of our method, we conduct experiments on diverse real-world datasets, including both indoor and outdoor environments. With the camera poses provided by these datasets, we employ COLMAP [23] to generate an initial sparse point cloud for each scene. We select large and complex scenes from the TnT dataset [11] to evaluate the quality of surface reconstruction. The widely used Mip-NeRF360 dataset [2] is used to assess novel view synthesis performance.", + "bbox": [ + 511, + 703, + 905, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocols. For fair comparisons with previous methods, we select three widely used image evaluation metrics to validate novel view synthesis: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM),", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16351", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/03a70c51a27fffe933904b6498f7e040cf3598da5d7055543a3517d34d9a5310.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
F1-Score↑PSNR↑
COLMAPNeuSNeurlangelo2DGSGOFPGSRCarGSNeuS2DGSRaDe-GSGOFPGSRCarGS
Barn0.550.290.700.360.510.660.6026.3628.7926.4526.7128.8028.91
Caterpillar0.010.290.360.230.410.410.4325.2124.2322.4522.9023.0425.16
Courthouse0.110.170.280.130.280.210.1723.5523.5119.7321.4421.8123.54
Ignatius0.220.830.890.440.680.800.7823.2723.8221.5722.2022.2124.77
Meetingroom0.190.240.320.160.280.290.3525.3826.1525.5225.7224.7227.27
Truck0.190.450.480.260.580.600.6523.7126.8523.7224.3226.1226.41
Mean0.210.380.500.300.460.500.5024.5825.5623.2423.8824.4526.01
", + "bbox": [ + 107, + 89, + 893, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e81ff514c379b3b5288354b2ffa60a734575bbcffd72862f299557564b29d60f.jpg", + "table_caption": [ + "Table 2. Quantitative result on Tanks and Temples dataset. CarGS achieves the SOTA performance on surface reconstruction and image synthesis. The color green indicates the best result and blue indicates the second best one." + ], + "table_footnote": [], + "table_body": "
Indoor scenesOutdoor scenesAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ImplicitNeRF26.840.7900.37021.460.4580.51524.150.6240.443
Deep Blending26.400.8440.26121.540.5240.36423.970.6840.313
INGP29.150.8800.21622.900.5660.37126.030.7230.294
NeuS25.100.7890.31921.930.6290.60023.740.7200.439
Explicit3DGS30.990.9260.19924.240.7050.28327.620.8030.246
Scaffold-GS31.240.9250.18525.360.7390.20328.300.8320.194
2DGS30.390.9230.18324.330.7090.28427.030.8040.239
PGSR30.410.9300.16724.450.7300.22427.430.8300.193
CarGS30.850.9280.17124.510.7410.24627.680.8340.209
", + "bbox": [ + 93, + 316, + 903, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative result of rendering quality on Mip-NeRF360 dataset. The proposed CarGS achieves the best average performance on novel view synthesis compared with the representative implicit Nerf-based and explicit GS-based methods.", + "bbox": [ + 89, + 523, + 906, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and learned perceptual image patch similarity (LPIPS) [35]. We employ the F1-score to assess the surface quality on the TnT dataset. Apart from the commonly used metrics, we additionally report the storage size (MB) for model compactness and the rendering speed (FPS) for efficiency.", + "bbox": [ + 88, + 565, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. Most of our training strategies and hyper-parameters are consistent with those used in 3DGS [10]. Specifically, we adopt a similar approach to halt densification at 15,000 iterations and optimize all models for 30,000 iterations. All MLPs in our method are 2-layer networks with ReLU activation function, and each hidden layer contains 32 units. The loss weights $\\alpha$ and $\\beta$ in Eq. 15 are set to 0.01 and 0.2, respectively. We begin by rendering depth for each training view, then apply the TSDF Fusion algorithm [18] to create the corresponding TSDF field, from which we subsequently extract the mesh [14]. All experiments were conducted on an Nvidia RTX 3090 GPU.", + "bbox": [ + 88, + 641, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Result Analysis", + "text_level": 1, + "bbox": [ + 89, + 832, + 246, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rendering Comparison. As shown in Tab. 2 and Tab. 3, our method achieves rendering quality comparable to recent state-of-the-art 3DGS and NeRF-based methods. Unlike", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prior primitive-based approaches that apply shared Gaussian primitives for both rendering and reconstruction (e.g., 2DGS [9], PGSR [3]), our approach adaptively adjusts contributions to meet specific task requirements, allowing high-quality geometry reconstruction without sacrificing rendering fidelity. Our method also diverges from dual-model approaches (e.g., GSDF [30]) by supporting joint training within a unified network for both tasks. As shown in Fig. 7, our approach shows better quality in rendering texture-less regions, where conventional 3D Gaussians struggle due to low accumulated gradients.", + "bbox": [ + 511, + 565, + 906, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction Comparison. We compare our method, CarGS, with current state-of-the-art neural surface reconstruction approaches, including NeuS [25] and NeuralAngelo [13], as well as recent 3DGS-based methods like 2DGS [9], GOF [32], and PGSR [3]. Reconstruction performance is quantitatively evaluated on the TNT dataset, where CarGS achieves an F1 score comparable to NeuralAngelo and PGSR, outperforming other existing methods, as shown in Tab.2. Fig.7 illustrates visual results from various Gaussian splatting approaches, with CarGS generating smoother and more precise shapes.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16352", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10f53ffa674da7687e85fb27c29015ea1e021814eb14b5cb2d70850c5c5fa235.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on Tanks and Temples dataset. In terms of rendering, our approach surpasses the baseline methods in effectively suppressing blurry regions. For reconstruction, we assess surface quality by visualizing the normal map generated from the reconstructed mesh. CarGS demonstrates advancements in capturing details of the scene compared to other approaches." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 903, + 296 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/44a95be93c9dd7b3bed7565ef5e024a41c948daea7fdfdbbac478e6227405a04.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR↑F1↑Time↓Storage↓FPS↑
NeuS23.710.45>24h1.4GB<1
Neurlangelo25.430.48>24h4.2GB<1
GSDF26.050.463h1.7GB<1
Scaffold-GS25.840.2326.5m0.10GB101
PGSR26.120.601.2h0.42GB103
CarGS26.410.651.2h0.16GB90
", + "bbox": [ + 109, + 369, + 467, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency Comparison. In terms of efficiency, we compare the training time, rendering speed, and storage requirements of different methods on the \"Truck\" scene from the TnT dataset, as shown in Tab. 4. FPS includes both depth (reconstruction) and RGB generation (rendering). Our method significantly reduces storage requirements, effectively inheriting the compact characteristics of anchor-based models. Compared to the dual-model approach GSDF, our method substantially reduces training time and achieves real-time rendering. In comparison to the Gaussian primitive regularization method PGSR, our model maintains high rendering and reconstruction quality with only a small FPS decrease. Notably, our method only consumes $38\\%$ storage than PGSR.", + "bbox": [ + 89, + 597, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 818, + 254, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of each individual module using the 'Truck' scenes from the TnT dataset. As shown in Tab. 5 (a), directly applying the geometry loss proposed in [3] to the baseline Scaffold-GS [15] reveals significant limi", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c38f6aa76367aec8d6c0a92f2c169c1b757ce0aa241d18234430968621e8e645.jpg", + "table_caption": [ + "Table 4. Comparison of efficiency. Our CarGS maintains high rendering and reconstruction quality while achieving notable computational and storage efficiency." + ], + "table_footnote": [], + "table_body": "
MethodsGeo-MLPLite-GeoGeo-DensifyF1↑PSNR↑
(a)0.3825.72
(b)✓0.4425.75
(c)✓✓0.6226.13
(d)✓✓✓0.6526.41
", + "bbox": [ + 514, + 371, + 908, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study on the Truck scene of TnT dataset.", + "bbox": [ + 524, + 479, + 892, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tations in reconstruction quality. (b) While employing GeoMLP to adjust the contribution of primitives leads to noticeable improvements in reconstruction quality, overfitting to the geometry loss indicates there is still room for further enhancement. (c) By incorporating the proposed Lite-Geo, the model effectively utilizes depth information from rendering supervision, resulting in a substantial performance boost for both tasks. (d) Finally, the inclusion of geometry-guided densification enables the Gaussian primitives to capture finer details, further enhancing overall performance.", + "bbox": [ + 511, + 520, + 906, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes CarGS, a unified model that leverages Contribution-adaptive Regularization to effectively address the conflicting requirements of high-quality rendering and accurate surface reconstruction. By introducing a compact MLP for adaptive contribution learning and a geometry-guided densification strategy, CarGS enhances the mutual learning between rendering and reconstruction tasks. Extensive experiments validate that CarGS achieves state-of-the-art results in both rendering fidelity and reconstruction accuracy, while maintaining real-time speed and minimal storage requirements. We believe our approach establishes a strong baseline for unified rendering and reconstruction.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16353", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Science Fund for Distinguished Young Scholars (No. 62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001).", + "bbox": [ + 89, + 114, + 485, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 247, + 187, + 263 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2, 6", + "[3] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 3, 4, 6, 7, 8", + "[4] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2", + "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2", + "[6] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 3", + "[7] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 1", + "[8] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5875-5884, 2021. 2", + "[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 7", + "[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + ], + "bbox": [ + 93, + 273, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 6, 7", + "[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 3, 6", + "[12] Jonas Kulhanek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18458-18469, 2023. 2", + "[13] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3, 7", + "[14] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 7", + "[15] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2, 3, 4, 8", + "[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[17] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3", + "[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. Ieee, 2011. 7", + "[19] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3", + "[20] Yansong Qu, Yuze Wang, and Yue Qi. Sg-nerf: Semantic-guided point-based neural radiance fields. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 570-575. IEEE, 2023. 2", + "[21] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 2", + "[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16354", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2", + "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6", + "[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2, 3", + "[25] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3, 7", + "[26] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 2", + "[27] Yuze Wang, Junyi Wang, Ruicheng Gao, Yansong Qu, Wantong Duan, Shuo Yang, and Yue Qi. Look at the sky: Sky-aware efficient 3d gaussian splattering in the wild. IEEE Transactions on Visualization and Computer Graphics, 2025. 2", + "[28] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3", + "[29] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 2", + "[30] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2, 3, 4, 5, 7", + "[31] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2", + "[32] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 7", + "[33] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2", + "[34] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6", + "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + ], + "bbox": [ + 91, + 90, + 483, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16355", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_model.json b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_model.json new file mode 100644 index 0000000000000000000000000000000000000000..12d15451db01ad35bd41e005d8d26fc9b6b6cca4 --- /dev/null +++ b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.916, + 0.278 + ], + "angle": 0, + "content": "You Shen\\(^{1\\dagger}\\), Zhipeng Zhang\\(^{2\\dagger}\\), Xinyang Li\\(^{1}\\), Yansong Qu\\(^{1}\\), Yu Lin\\(^{1}\\), Shengchuan Zhang\\(^{1}\\), Liujuan Cao\\(^{1*}\\) \n\\(^{1}\\)Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University \n\\(^{2}\\)School of Artificial Intelligence, Shanghai Jiao Tong University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.782 + ], + "angle": 0, + "content": "Representing 3D scenes from multiview images is a core challenge in computer vision and graphics, which requires both precise rendering and accurate reconstruction. Recently, 3D Gaussian Splating (3DGS) has garnered significant attention for its high-quality rendering and fast inference speed. Yet, due to the unstructured and irregular nature of Gaussian point clouds, ensuring accurate geometry reconstruction remains difficult. Existing methods primarily focus on geometry regularization, with common approaches including primitive-based and dual-model frameworks. However, the former suffers from inherent conflicts between rendering and reconstruction, while the latter is computationally and storage-intensive. To address these challenges, we propose CarGS, a unified model leveraging Contribution-adaptive regularization to achieve simultaneous, high-quality rendering and surface reconstruction. The essence of our framework is learning adaptive contribution for Gaussian primitives by squeezing the knowledge from geometry regularization into a compact MLP. Additionally, we introduce a geometry-guided densification strategy with clues from both normals and Signed Distance Fields (SDF) to improve the capability of capturing high-frequency details. Our design improves the mutual learning of the two tasks, meanwhile its unified structure doesn't require separate models as in dual-model based approaches, guaranteeing efficiency. Extensive experiments demonstrate CarGS's ability to achieve state-of-the-art (SOTA) results in both rendering fidelity and reconstruction accuracy while maintaining real-time speed and minimal storage size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.224, + 0.825 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) and surface reconstruction from multi-view images are both challenging and critical" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.308, + 0.907, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Figure 1. The performance of CarGS in TnT datasets, achieving high-quality rendering and reconstruction in a unified model. The storage size of the model is represented by the radius of the circle." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.567, + 0.909, + 0.718 + ], + "angle": 0, + "content": "tasks in the fields of computer vision and graphics. Both academia and industry are actively pursuing unified scene representations that enable high-quality rendering and accurate geometry reconstruction concurrently to address the growing demands of applications in areas such as augmented reality (AR), virtual reality (VR), robotics, and autonomous driving. However, simultaneously achieving SOTA performance in both rendering and reconstruction with a unified model is non-trivial, as an inherent conflict exists between the two tasks [3, 7, 9]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Revisiting recent approaches, we identify two solutions aimed at mitigating the conflict. The first branch, which we refer to as \"primitive-based\", introduces additional geometric regularization to 3D Gaussian primitives to improve surface reconstruction. For example, Sugar [7] regularizes 3D Gaussians to align more closely with surfaces. Although this approach ensures parameter efficiency, it leads to a misaligned prioritization between appearance and geometry, eventually degrading rendering performance. The second branch, termed \"dual-model based\", seeks to guide the learning of one task through information derived from the other. For instance, GSDF [30] introduces a dual-model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.24, + 0.888 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "Equal Contribution." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.24, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16346" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "architecture that jointly supervises the 3D Gaussian Splating model for rendering and the SDF model for reconstruction. However, this strategy necessitates separate models for training and inference, resulting in much larger computational and storage costs. Consequently, a natural question arises: Is it possible to develop an efficient framework (fast speed with small storage) that is also effective (SOTA performances on both tasks)?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.217, + 0.487, + 0.534 + ], + "angle": 0, + "content": "Before answering this question, we first present an intriguing observation that motivates us to pursue the \"unified goal\" from the perspective of learning adaptive contribution of Gaussian primitives. Specifically, with the baseline model PGSR [3] trained on the \"Barn\" scene from the Tanks and Temples dataset [11], we first calculate the Chamfer Distance (CD) from the center of each Gaussian to the ground truth surface. Then, as shown in Fig. 2, when retaining \\(50\\%\\) of the Gaussian points by discarding those with larger CD values, we find that reconstruction quality (F1) is surprisingly maintained, while rendering quality (PSNR) experiences a sharp decline. Conversely, when additional Gaussian points with smaller CDs are dropped, reconstruction quality begins to deteriorate, while rendering quality remains relatively stable. This observation suggests that, 1) despite geometric regularization encouraging Gaussian primitives to closely adhere to the surface [3, 9], a significant number of floaters persist in the scene. 2) These floaters contribute less to reconstruction but are essential for rendering quality. More precisely, each Gaussian primitive contributes differently to each task." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.539, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Therefore, an intuitive idea is to build a unified model with adaptive contributions for each Gaussian primitive. This \"unified\" structure contrasts with dual-model designs, significantly reducing storage and computational demands from the outset. The \"adaptiveness\" mitigates conflicts in Gaussian primitive contributions when jointly optimizing for rendering and reconstruction tasks. To this end, we propose CarGS: a unified Gaussian Splitting model based on Contribution-adaptive regularization. Firstly, we squeeze geometry regularization knowledge into a lightweight residual structure to adaptively adjust each Gaussian primitive's contribution for rendering and reconstruction. Then, to further capture geometric details, we introduce a geometry-guided densification strategy that controls Gaussian expansion using clues from both normals and SDF. Experimental results show that our method achieves state-of-the-art reconstruction performance within a unified model, without compromising rendering quality. Remarkably, our model consumes only \\(9\\%\\) of the storage and \\(40\\%\\) of the training time required by GSDF [30] (dual-model based) while running \\(100\\times\\) faster. Moreover, its rendering and reconstruction performance surpasses both GSDF and PGSR [3] (primitive-based). We believe our approach establishes a strong baseline for unified rendering and reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.087, + 0.905, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.246, + 0.908, + 0.29 + ], + "angle": 0, + "content": "Figure 2. The performance of the Baseline model [3] across various \\(\\mathrm{CD\\%}\\) thresholds demonstrates that each Gaussian primitive contributes differently to specific tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.908, + 0.473 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows: 1) We identify the key bottleneck in developing a unified, efficient, and effective framework for both rendering and reconstruction. 2) We propose to learn the adaptive contribution of each Gaussian primitive, addressing the inherent conflict in creating a unified framework for rendering and reconstruction. 3) We introduce a geometry-guided densification strategy to enhance the model's ability to capture fine details. 4) Our unified framework simultaneously achieves high-quality rendering and reconstruction while consuming significantly less storage than previous methods (see Fig. 1)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.489, + 0.652, + 0.505 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.514, + 0.715, + 0.531 + ], + "angle": 0, + "content": "2.1. Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Novel View Synthesis aims to generate new perspectives of a scene using a limited collection of existing views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.749 + ], + "angle": 0, + "content": "NVS based on Neural Radiance Fields. A pivotal advancement in this field was the introduction of Neural Radiance Fields (NeRF) [16]. NeRF employs MLPs to map 3D spatial locations along rays to color and density values, which are subsequently aggregated through neural volumetric rendering to produce high-quality images. Subsequent research has focused on enhancing the capabilities of NeRF models through feature-grid representations [5, 12, 20, 24] and improving rendering speed using baking techniques [8, 22, 29]. Recently, NeRF has been adapted to address challenges such as anti-aliasing [1] and the modeling of unbounded scenes [2, 34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "NVS based on 3D Gaussian Splatting. Recently, 3D Gaussian Splatting [10] has emerged as a transformative approach in neural rendering by utilizing anisotropic 3D Gaussians as primitives. This method sorts Gaussians by depth and rasterizes them onto a 2D screen through alpha blending, resulting in impressive rendering quality and real-time speed. Subsequent research has enhanced 3D Gaussian Splatting by improving rendering quality [4, 21, 31] and expanding its applicability to large-scale scenes [15, 26, 27, 33]. For example, Mip-splatting [31] introduces" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16347" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "anti-aliasing techniques to mitigate pixelation and blurring in high-resolution renders, thereby enhancing visual sharpness and consistency across viewpoints. Scaffold-GS [15] employs anchor points to manage local 3D Gaussians and dynamically predicts their attributes based on viewing direction and distance, thereby optimizing the adaptive rendering process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.259 + ], + "angle": 0, + "content": "While these NVS methods excel in view synthesis, they inherently overlook the accurate reconstruction of scene geometry, resulting in blurred volumetric density fields that impede the extraction of high-quality surfaces." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.272, + 0.364, + 0.286 + ], + "angle": 0, + "content": "2.2. Neural Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.506 + ], + "angle": 0, + "content": "The success of neural rendering has sparked significant interest in neural surface reconstruction [6, 13, 19, 25, 28]. By leveraging coordinate-based neural networks, these methods encode scene geometry using occupancy fields or SDF values, facilitating a smooth and differentiable representation of 3D structures. Though NeRF-based methods with volume rendering yield smooth, complete surfaces, they struggle with fine detail capture and slow optimization. To address this, recent work [13] incorporates multi-resolution hashed grids from Instant-NGP [17], enhancing representation and achieving SOTA results. Yet, large-scale scene reconstruction remains inefficient; e.g., Neuralangelo [13] takes 128 GPU hours to reconstruct a single Tanks and Temples scene [11]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Motivated by the impressive efficiency of 3DGS in NVS, recent research has investigated the use of anisotropic 3D Gaussians as foundational primitives for scene geometry representation. SuGaR [24] aligns 3D Gaussians with potential surfaces by approximating them as 2D planar primitives with binary opacity. 2D Gaussian Splatting [9] replaces 3D Gaussians with 2D counterparts, resulting in more accurate ray-splat intersections and regularizing depth maps to achieve a denser distribution of Gaussian primitives. PGSR [3] employs both single-view and multi-view geometric regularization to attain globally consistent geometry. Despite these advancements in neural surface reconstruction, a fidelity gap persists, as the explicit regularization of Gaussian primitives constrains rendering quality. To address this limitation, GSDF [30] introduces a dual-model framework that separates rendering (3DGS) and geometry (SDF) tasks, utilizing joint supervision during training to enhance both reconstruction and rendering. However, this design incurs additional training and storage costs due to the requirement of maintaining two independent models. Moreover, the geometry reconstruction branch is not capable of real-time inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Different from existing approaches, our paper presents a unified model designed to simultaneously achieve real-time and robust rendering and reconstruction tasks without compromising geometry accuracy and rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.328 + ], + "angle": 0, + "content": "This section details the proposed framework. Particularly, Sec. 3.1 presents preliminary studies on 3D Gaussian Splating and its representative anchor-based paradigm. Then in Sec. 3.2, we analyze the problem of contribution conflicts to identify the most significant attributes for geometry reconstruction. Sec. 3.3 introduces Lite-Geo, a lightweight residual module that adapts the geometry contribution of Gaussian primitives. Based on Lite-Geo, we further propose a geometry-guided densification strategy in Sec. 3.4 that leverages geometric cues to align Gaussians with the surface, enhancing both structural fidelity and representational capacity. Finally, in Sec. 3.5, we introduce geometric regularization to guide the optimization of our modules, which improves the quality of geometry reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.334, + 0.644, + 0.35 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.448 + ], + "angle": 0, + "content": "3D Gaussian Splatting. 3DGS models the scene with a set of explicit Gaussians that inherit the differentiable properties of volumetric representations, allowing efficient rendering via tile-based rasterization. Initially, each point generated by the Structure-from-Motion (SfM) serves as the position (mean) \\(\\mu\\) of a corresponding 3D Gaussian:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.453, + 0.905, + 0.473 + ], + "angle": 0, + "content": "\\[\nG (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.481, + 0.906, + 0.541 + ], + "angle": 0, + "content": "where \\( x \\) denotes an arbitrary position in the 3D scene, while \\( \\Sigma \\) represents the covariance matrix of the 3D Gaussian. \\( \\Sigma \\) is constructed using a scaling matrix \\( S \\) and a rotation matrix \\( R \\) to maintain positive semi-definiteness." + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.547, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\Sigma = R S S ^ {T} R ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.906, + 0.694 + ], + "angle": 0, + "content": "Each 3D Gaussian is assigned with an opacity value \\(\\alpha\\), which is multiplied with \\(G(x)\\) to adjust its contribution weight to rendering and reconstruction during the blending process. 3DGS renders scenes efficiently through tile-based rasterization instead of traditional ray marching. Each 3D Gaussian \\(G(x)\\) is projected as a 2D Gaussian onto the image plane, after which a tile-based rasterizer sorts and blends the 2D Gaussians using \\(\\alpha\\)-blending:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.701, + 0.905, + 0.743 + ], + "angle": 0, + "content": "\\[\nF \\left(x ^ {\\prime}\\right) = \\sum_ {i \\in N} f _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} (x) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( x' \\) is the queried pixel location in the image plane and \\( N \\) denotes the number of sorted 2D Gaussians associated with the queried pixel. For a Gaussian with 3D center \\( (x_i, y_i, z_i) \\) and color \\( c_i \\), \\( f_i \\) is assigned to \\( c_i \\) in the rendering task, and to \\( z_i \\) for depth blending in the reconstruction task. Anchor-based Gasussian Splatting. Our goal is to simultaneously achieve high-quality rendering and reconstruction while maintaining efficiency. In our work, we adopt anchor-based GS modeling as our baseline due to its balanced efficiency and effectiveness. A representative work in this" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16348" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.476, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.187, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Figure 3. The illustration of contribution conflict between rendering and reconstruction tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.304 + ], + "angle": 0, + "content": "paradigm, Scaffold-GS [15], enhances original 3DGS by improving scene structure fidelity and incorporating view clues. Specifically, it constructs a hierarchical 3D Gaussian framework with anchor points that encode local scene details and generate local neural Gaussians," + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.314, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mu_ {0}, \\dots , \\mu_ {k - 1} \\right\\} = x + \\left\\{o _ {0}, \\dots , o _ {k - 1} \\right\\} \\cdot \\gamma , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.446 + ], + "angle": 0, + "content": "where \\(o_k \\in \\mathbb{R}^3\\) denotes the learned offsets, and \\(\\gamma\\) represents the scaling factor. For each anchor, the optimized attributes (color, center, variance, and opacity) of the \\(k\\) neural Gaussians are decoded through an MLP, \\(M\\), based on the anchor feature \\(\\mathbf{f}\\), as well as the distance \\(\\delta\\) and direction \\(\\theta\\) from the camera to the anchor point. Specifically, the opacity values for neural Gaussians generated from an anchor point are," + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.456, + 0.483, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {0}, \\dots , \\alpha_ {k - 1} = M (\\mathbf {f}, \\delta , \\theta), \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.505, + 0.194, + 0.52 + ], + "angle": 0, + "content": "3.2. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.601 + ], + "angle": 0, + "content": "In this section, we analyze the underlying reasons for the rendering performance degradation when geometry regularization is applied to 3DGS primitives. Then, we conduct experiments to analyze the effect of Gaussian's internal attributes to geometry reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.484, + 0.693 + ], + "angle": 0, + "content": "Analysis of Contribution Conflicts. In Fig 3, we consider a simple demonstration of alpha-blending along a ray intersecting three Gaussian primitives. As shown, \\( G_{2} \\) is close to the real surface, \\( G_{1} \\) and \\( G_{3} \\) are farther away. According to Eq. 3, the color \\( C(x') \\) and depth \\( D(x') \\) corresponding to the queried pixel along the ray is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.702, + 0.483, + 0.72 + ], + "angle": 0, + "content": "\\[\nC \\left(x ^ {\\prime}\\right) = c _ {1} w _ {1} + c _ {2} w _ {2} + c _ {3} w _ {3}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.728, + 0.483, + 0.745 + ], + "angle": 0, + "content": "\\[\nD \\left(x ^ {\\prime}\\right) = z _ {1} w _ {1} + z _ {2} w _ {2} + z _ {3} w _ {3} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.75, + 0.483, + 0.791 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} (x) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(w_{i}\\) represents the contribution of the Gaussian and serves as the weighting factor for ray integration. When only applying rendering regularization, \\(w_{1}\\) and \\(w_{3}\\) tend to increase to capture richer textures, as proved in Fig. 2. Conversely, when geometry regularization is applied, the values of \\(w_{1}\\) and \\(w_{3}\\) tend to decrease due to \\(z_{1}\\) and \\(z_{3}\\) deviating from the SDF 0-level set, while \\(w_{2}\\) increases to provide" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.906, + 0.151 + ], + "angle": 0, + "content": "
CovOpacityBaseline
F1↑0.460.640.65
PSNR↑23.8428.7228.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.162, + 0.907, + 0.19 + ], + "angle": 0, + "content": "Table 1. The importance of different attributes in Gaussian's contribution for the geometry reconstruction task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.907, + 0.459 + ], + "angle": 0, + "content": "more consistent depth information. This shift in the contribution weights \\( w_{i} \\) creates conflicts, which can lead to a decline in rendering performance when geometry regularization is simply added when training the rendering task [3, 9]. Analysis of Attributes. As discussed, the degradation in rendering performance after using geometry regularization stems from contribution conflicts. To further investigate this issue, we analyze which internal attributes of the Gaussian's contribution are most susceptible to the effects of geometry regularization. As shown in Eq. 8, Gaussian's contribution \\( w_{i} \\) is defined by its opacity \\( \\alpha \\) and Gaussian sampling value \\( G(x) \\). Lu et al. [15] emphasize that stable Gaussian positions are crucial for maintaining scene structure, suggesting that adjustments to \\( \\mu \\) in Eq. 1 remain relatively minor during optimization. Consequently, changes in \\( G(x) \\) are more likely to arise from covariance in the covariance \\( \\Sigma \\) in Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Then, to investigate the roles of \\(\\alpha\\) (opacity) and \\(\\Sigma\\) (covariance) within the geometry regularization process, we conduct analysis experiments in Tab. 1. Specifically, with the baseline model PGSR [3] on TnT's \"Barn\" scene, we respectively detach pacity and covariance from the geometry supervision to stop their gradigents and then assess their independent effect on the performance of geometry reconstruction. As shown in Tab. 1, excluding the gradient of opacity from geometry optimization shows minimal impact on reconstruction quality, while excluding covariance's gradient severely impairs geometry learning. Hence, covariance plays the essential role in regulating the contributions of Gaussian primitives, making it important to tailor covariance for both high-quality rendering and reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.816, + 0.699 + ], + "angle": 0, + "content": "3.3. 3DGS with Adaptive Contribution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 4, existing approaches for realizing rendering and reconstruction are categorized into primitive based [3, 9] and dual-model based [30] frameworks. The former supervises Gaussian primitives with both rendering and geometry losses, leading to conflicting contributions and, consequently, performance degradation. In contrast, dual-model based methods train separate models for each task, i.e., a 3DGS model for rendering and an SDF model for reconstruction. While enhancing both tasks, it requires substantial computational and storage resources (see Tab. 4). Building on the above analysis, we are motivated to propose a unified and efficient contribution-adaptive 3DGS architecture to simultaneously achieve state-of-the-art per" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16349" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.478, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.18, + 0.383, + 0.193 + ], + "angle": 0, + "content": "(a) Primitive based regularization" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.201, + 0.48, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.324, + 0.392, + 0.337 + ], + "angle": 0, + "content": "(b) Dual-model based regularization" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.349, + 0.48, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.451, + 0.413, + 0.464 + ], + "angle": 0, + "content": "(c) Contribution-adaptive regularization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Figure 4. The illustration of the geometry regularization GS framework: (a) Primitive based regularization paradigm. (b) Dual-model-based regularization paradigm. (c) the proposed contribution adaptive regularization paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.58 + ], + "angle": 0, + "content": "formances by alleviating the adverse effects on rendering when using geometry regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Geo-MLP. In Sec. 3.2, we have identified covariance as the primary factor contributing to conflicts in geometry and rendering. Thus, we propose decoupling Gaussian contributions by predicting an additional covariance parameter dedicated to geometry reconstruction, while retaining the original covariance for rendering. As outlined in Eq. 2, covariance is decomposed into scaling and rotation matrices. Therefore, we first attempt to introduce an auxiliary MLP that takes the anchor feature \\( \\mathbf{f} \\) and view direction as inputs, predicting the scaling and rotation matrix. The predicted geometry-related scaling and rotation are supervised by reconstruction loss, as shown in Fig. 4 (c). However, nothing is as simple as it seems. As visualized in Fig. 5 (top), this straightforward design results in overfitting in geometry estimation. This challenge arises because conventional geometric regularization primarily enforces continuity and consistency, but lacks accurate supervision for depth, which is predominantly informed by rendering-based supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Lite-Geo. To address this challenge, we introduce Lite-Geo, designed to inherit the Gaussian primitive's contribution from the rendering task, which encodes inherent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.199 + ], + "angle": 0, + "content": "depth information after rendering optimization. This implicit depth information helps mitigate overfitting brought by the sole geometry reconstruction loss applied to GeoMLP. More precisely, we structure the learning of covariance for geometry reconstruction as a residual update, where it is obtained by adding the offset predicted by GeoMLP \\((M_{\\Sigma}^{geo})\\) to the rendering covariance." + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.211, + 0.905, + 0.231 + ], + "angle": 0, + "content": "\\[\ny = M _ {\\Sigma} ^ {r g b} \\left(\\mathbf {f}, \\theta ; \\phi_ {1}\\right) + \\triangle y, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.25, + 0.905, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\triangle y = M _ {\\Sigma} ^ {g e o} (\\mathbf {f}, \\theta ; \\phi_ {2}), \\phi_ {2} = \\lambda \\phi_ {1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.442 + ], + "angle": 0, + "content": "where \\(\\phi_{1}\\) and \\(\\phi_{2}\\) denote the parameters of \\(M_{\\Sigma}^{rgb}\\) and \\(M_{\\Sigma}^{geo}\\), respectively. \\(y\\) and \\(\\triangle y\\) are 7-dimensional vectors, with the first three dimensions representing scale and the remaining four representing rotation through quaternions. This approach allows the depth information in rendering covariance to effectively inform the depth estimation in geometry reconstruction. To further reduce noise and enhance convergence, we initialize \\(\\phi_{1}\\) based on \\(\\phi_{2}\\) using a controllable scaling factor \\(\\lambda\\). Specifically, the rendering task is pre-trained independently for several epochs before incorporating the Lite-Geo module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.906, + 0.503 + ], + "angle": 0, + "content": "As shown in Fig. 5, this residual structure effectively suppresses geometry overfitting. It also constrains outliers with high Chamfer Distance values that negatively impact rendering and reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.792, + 0.529 + ], + "angle": 0, + "content": "3.4. Geometry-guided Densification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.792 + ], + "angle": 0, + "content": "We further enhance the distribution of Gaussian primitives through a geometry-guided densification strategy. Our motivation stems from the observation that previous geometry-based densification methods [30] rely solely on SDF values to control Gaussian growing, which limits their ability to capture fine details in high-frequency regions and often results in blurred rendering and overly smooth reconstructions. As shown in Fig. 6, to better accommodate high-frequency details, we introduce normal vector clues alongside the original SDF information to guide Gaussian growing. Specifically, for a Gaussian anchor located at position \\( x \\) and the SDF value within a certain range \\( \\theta \\), we evaluate its normal difference with the normal map by \\( n = N_d(x) - N(x) \\). The SDF is obtained from the depth map by \\( s = D(x) - Z(x) \\). Detailed calculations of normals and SDF are provided in the supplementary. We then define the criteria for Gaussian growing as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.804, + 0.907, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {g} = \\nabla_ {g} + \\omega_ {g} (\\zeta (s) \\cdot (\\omega_ {n} n \\text {i f} \\zeta (s) < \\theta \\text {e l s e} 1)), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.832, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\zeta (s) = \\exp \\left(- \\frac {s ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\nabla_g\\) is the averaged gradient of the Gaussian primitives accumulated over several training iterations. \\(\\zeta (s)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16350" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Figure 5. The illustration of geometry overfitting. If the Geo-MLP is solely associated with geometry reconstruction, notable overfitting is observed. By leveraging the approximate depth knowledge provided by rendering, Lite-Geo can effectively reduce overfitting and suppress noise where the Chamfer Distance exceeds the threshold." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.314, + 0.462, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.483, + 0.462 + ], + "angle": 0, + "content": "Figure 6. The illustration of high-frequency details, along with the proposed densification strategy which facilitates Gaussian growing with clues from both normals and SDF." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.552 + ], + "angle": 0, + "content": "is a Gaussian function that transforms the SDF to a positive number, which monotonically decreases as the distance from the zero level set increases. The parameter \\(\\omega_{g}\\) controls the influence of geometric guidance. New Gaussian primitives are generated when \\(\\epsilon_{g}\\) exceeds a predefined threshold." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.232, + 0.574 + ], + "angle": 0, + "content": "3.5. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.7 + ], + "angle": 0, + "content": "Relying solely on photometric loss to optimize 3DGS introduces noise into the output, as multi-view 3D reconstruction lacks sufficient constraints [2, 34] for geometric continuity. To improve the recovery of shape detail, we adopt the geometry regularization from PGSR [3], incorporating additional plane constraint and cross-view constraint. To ensure the paper is self-contained, we provide a detailed explanation of these two regularization terms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Plane Constraint. To ensure local depth-normal consistency, each pixel and its neighbors should define a relatively smooth plane. Using the depths of a pixel and its neighboring pixels, we can generate a normal map, which is then optimized to align with the rendered normal map, thereby enforcing geometric consistency. The loss based on the plane constraint is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.811, + 0.483, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p l a n e}} = \\frac {1}{| \\mathcal {W} |} \\sum_ {x \\in \\mathcal {W}} \\| N _ {d} (x) - N (x) \\| _ {1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\(N_{d}(x)\\) and \\(N(x)\\) represent the normal derived from the depth gradient and the normal of the splats, respectively. \\(\\mathcal{W}\\) denotes the image pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Cross-view Constraint. Plane constraint ensures local depth-normal consistency but lacks cross-view consistency. Hence, cross-view constraintsd should be introduced to improve global geometric consistency. Specifically, a pixel \\( x_{r} \\) in the reference frame, associated with a normal \\( n_r \\) and distance \\( d_{r} \\), can be transformed to a corresponding pixel \\( x_{n} \\) in a neighboring frame with the homography matrix \\( H_{rn} \\). After forward projection from the reference frame to the neighboring frame, a backward projection \\( H_{nr} \\) is applied. Then the consistency is ensured by minimizing the forward-backward projection error:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.494, + 0.907, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c r o s s} = \\frac {1}{| \\mathcal {W} |} \\sum_ {x _ {r} \\in \\mathcal {W}} \\| x _ {r} - H _ {n r} H _ {r n} x _ {r} \\|. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.542, + 0.906, + 0.558 + ], + "angle": 0, + "content": "Final Loss. Finally, we optimize our model by minimizing," + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.57, + 0.905, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\alpha \\mathcal {L} _ {d} + \\beta \\mathcal {L} _ {n} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.644 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c\\) is the RGB reconstruction loss that combines L1 loss with D-SSIM from [10], while \\(\\mathcal{L}_{\\mathrm{plane}}\\) and \\(\\mathcal{L}_{\\mathrm{cross}}\\) represent the geometry regularization terms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.649, + 0.674 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.704, + 0.699 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Datasets. To validate the effectiveness of our method, we conduct experiments on diverse real-world datasets, including both indoor and outdoor environments. With the camera poses provided by these datasets, we employ COLMAP [23] to generate an initial sparse point cloud for each scene. We select large and complex scenes from the TnT dataset [11] to evaluate the quality of surface reconstruction. The widely used Mip-NeRF360 dataset [2] is used to assess novel view synthesis performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Evaluation protocols. For fair comparisons with previous methods, we select three widely used image evaluation metrics to validate novel view synthesis: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16351" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.09, + 0.895, + 0.272 + ], + "angle": 0, + "content": "
F1-Score↑PSNR↑
COLMAPNeuSNeurlangelo2DGSGOFPGSRCarGSNeuS2DGSRaDe-GSGOFPGSRCarGS
Barn0.550.290.700.360.510.660.6026.3628.7926.4526.7128.8028.91
Caterpillar0.010.290.360.230.410.410.4325.2124.2322.4522.9023.0425.16
Courthouse0.110.170.280.130.280.210.1723.5523.5119.7321.4421.8123.54
Ignatius0.220.830.890.440.680.800.7823.2723.8221.5722.2022.2124.77
Meetingroom0.190.240.320.160.280.290.3525.3826.1525.5225.7224.7227.27
Truck0.190.450.480.260.580.600.6523.7126.8523.7224.3226.1226.41
Mean0.210.380.500.300.460.500.5024.5825.5623.2423.8824.4526.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.276, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Table 2. Quantitative result on Tanks and Temples dataset. CarGS achieves the SOTA performance on surface reconstruction and image synthesis. The color green indicates the best result and blue indicates the second best one." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.317, + 0.905, + 0.522 + ], + "angle": 0, + "content": "
Indoor scenesOutdoor scenesAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ImplicitNeRF26.840.7900.37021.460.4580.51524.150.6240.443
Deep Blending26.400.8440.26121.540.5240.36423.970.6840.313
INGP29.150.8800.21622.900.5660.37126.030.7230.294
NeuS25.100.7890.31921.930.6290.60023.740.7200.439
Explicit3DGS30.990.9260.19924.240.7050.28327.620.8030.246
Scaffold-GS31.240.9250.18525.360.7390.20328.300.8320.194
2DGS30.390.9230.18324.330.7090.28427.030.8040.239
PGSR30.410.9300.16724.450.7300.22427.430.8300.193
CarGS30.850.9280.17124.510.7410.24627.680.8340.209
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.525, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Table 3. Quantitative result of rendering quality on Mip-NeRF360 dataset. The proposed CarGS achieves the best average performance on novel view synthesis compared with the representative implicit Nerf-based and explicit GS-based methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.483, + 0.642 + ], + "angle": 0, + "content": "and learned perceptual image patch similarity (LPIPS) [35]. We employ the F1-score to assess the surface quality on the TnT dataset. Apart from the commonly used metrics, we additionally report the storage size (MB) for model compactness and the rendering speed (FPS) for efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Implementation Details. Most of our training strategies and hyper-parameters are consistent with those used in 3DGS [10]. Specifically, we adopt a similar approach to halt densification at 15,000 iterations and optimize all models for 30,000 iterations. All MLPs in our method are 2-layer networks with ReLU activation function, and each hidden layer contains 32 units. The loss weights \\(\\alpha\\) and \\(\\beta\\) in Eq. 15 are set to 0.01 and 0.2, respectively. We begin by rendering depth for each training view, then apply the TSDF Fusion algorithm [18] to create the corresponding TSDF field, from which we subsequently extract the mesh [14]. All experiments were conducted on an Nvidia RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.248, + 0.85 + ], + "angle": 0, + "content": "4.2. Result Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Rendering Comparison. As shown in Tab. 2 and Tab. 3, our method achieves rendering quality comparable to recent state-of-the-art 3DGS and NeRF-based methods. Unlike" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.732 + ], + "angle": 0, + "content": "prior primitive-based approaches that apply shared Gaussian primitives for both rendering and reconstruction (e.g., 2DGS [9], PGSR [3]), our approach adaptively adjusts contributions to meet specific task requirements, allowing high-quality geometry reconstruction without sacrificing rendering fidelity. Our method also diverges from dual-model approaches (e.g., GSDF [30]) by supporting joint training within a unified network for both tasks. As shown in Fig. 7, our approach shows better quality in rendering texture-less regions, where conventional 3D Gaussians struggle due to low accumulated gradients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Reconstruction Comparison. We compare our method, CarGS, with current state-of-the-art neural surface reconstruction approaches, including NeuS [25] and NeuralAngelo [13], as well as recent 3DGS-based methods like 2DGS [9], GOF [32], and PGSR [3]. Reconstruction performance is quantitatively evaluated on the TNT dataset, where CarGS achieves an F1 score comparable to NeuralAngelo and PGSR, outperforming other existing methods, as shown in Tab.2. Fig.7 illustrates visual results from various Gaussian splatting approaches, with CarGS generating smoother and more precise shapes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16352" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.905, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.307, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on Tanks and Temples dataset. In terms of rendering, our approach surpasses the baseline methods in effectively suppressing blurry regions. For reconstruction, we assess surface quality by visualizing the normal map generated from the reconstructed mesh. CarGS demonstrates advancements in capturing details of the scene compared to other approaches." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.37, + 0.468, + 0.524 + ], + "angle": 0, + "content": "
PSNR↑F1↑Time↓Storage↓FPS↑
NeuS23.710.45>24h1.4GB<1
Neurlangelo25.430.48>24h4.2GB<1
GSDF26.050.463h1.7GB<1
Scaffold-GS25.840.2326.5m0.10GB101
PGSR26.120.601.2h0.42GB103
CarGS26.410.651.2h0.16GB90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Table 4. Comparison of efficiency. Our CarGS maintains high rendering and reconstruction quality while achieving notable computational and storage efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Efficiency Comparison. In terms of efficiency, we compare the training time, rendering speed, and storage requirements of different methods on the \"Truck\" scene from the TnT dataset, as shown in Tab. 4. FPS includes both depth (reconstruction) and RGB generation (rendering). Our method significantly reduces storage requirements, effectively inheriting the compact characteristics of anchor-based models. Compared to the dual-model approach GSDF, our method substantially reduces training time and achieves real-time rendering. In comparison to the Gaussian primitive regularization method PGSR, our model maintains high rendering and reconstruction quality with only a small FPS decrease. Notably, our method only consumes \\(38\\%\\) storage than PGSR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.255, + 0.833 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We evaluate the effectiveness of each individual module using the 'Truck' scenes from the TnT dataset. As shown in Tab. 5 (a), directly applying the geometry loss proposed in [3] to the baseline Scaffold-GS [15] reveals significant limi" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.372, + 0.909, + 0.47 + ], + "angle": 0, + "content": "
MethodsGeo-MLPLite-GeoGeo-DensifyF1↑PSNR↑
(a)0.3825.72
(b)✓0.4425.75
(c)✓✓0.6226.13
(d)✓✓✓0.6526.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.48, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Table 5. Ablation study on the Truck scene of TnT dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.673 + ], + "angle": 0, + "content": "tations in reconstruction quality. (b) While employing GeoMLP to adjust the contribution of primitives leads to noticeable improvements in reconstruction quality, overfitting to the geometry loss indicates there is still room for further enhancement. (c) By incorporating the proposed Lite-Geo, the model effectively utilizes depth information from rendering supervision, resulting in a substantial performance boost for both tasks. (d) Finally, the inclusion of geometry-guided densification enables the Gaussian primitives to capture finer details, further enhancing overall performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.634, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work proposes CarGS, a unified model that leverages Contribution-adaptive Regularization to effectively address the conflicting requirements of high-quality rendering and accurate surface reconstruction. By introducing a compact MLP for adaptive contribution learning and a geometry-guided densification strategy, CarGS enhances the mutual learning between rendering and reconstruction tasks. Extensive experiments validate that CarGS achieves state-of-the-art results in both rendering fidelity and reconstruction accuracy, while maintaining real-time speed and minimal storage requirements. We believe our approach establishes a strong baseline for unified rendering and reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16353" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.237 + ], + "angle": 0, + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars (No. 62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.248, + 0.188, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.274, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.357, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.427, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[3] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[4] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[6] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[7] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[8] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5875-5884, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.274, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[12] Jonas Kulhanek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18458-18469, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[13] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[14] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[15] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[17] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. Ieee, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[19] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[20] Yansong Qu, Yuze Wang, and Yue Qi. Sg-nerf: Semantic-guided point-based neural radiance fields. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 570-575. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[21] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16354" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[25] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[26] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[27] Yuze Wang, Junyi Wang, Ruicheng Gao, Yansong Qu, Wantong Duan, Shuo Yang, and Yue Qi. Look at the sky: Sky-aware efficient 3d gaussian splattering in the wild. IEEE Transactions on Visualization and Computer Graphics, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[28] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[29] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[30] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[31] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.484, + 0.727 + ], + "angle": 0, + "content": "[32] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.484, + 0.784 + ], + "angle": 0, + "content": "[33] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.786, + 0.484, + 0.827 + ], + "angle": 0, + "content": "[34] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.829, + 0.484, + 0.898 + ], + "angle": 0, + "content": "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16355" + } + ] +] \ No newline at end of file diff --git a/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_origin.pdf b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1301e1b659eaf2570850170bb32931a2c9ffcf1b --- /dev/null +++ b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/00b4d7e3-bd83-465f-afc0-737dc88f6946_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:096a0138b6f05317607ce70390cf3ff44ecb9c68b3eee28a01e5ff4237d1c08b +size 3730247 diff --git a/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/full.md b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..29474b4267eda463382bbef1f717289117bdfc54 --- /dev/null +++ b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/full.md @@ -0,0 +1,300 @@ +# Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization + +You Shen $^{1\dagger}$ , Zhipeng Zhang $^{2\dagger}$ , Xinyang Li $^{1}$ , Yansong Qu $^{1}$ , Yu Lin $^{1}$ , Shengchuan Zhang $^{1}$ , Liujuan Cao $^{1*}$ $^{1}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University + $^{2}$ School of Artificial Intelligence, Shanghai Jiao Tong University + +# Abstract + +Representing 3D scenes from multiview images is a core challenge in computer vision and graphics, which requires both precise rendering and accurate reconstruction. Recently, 3D Gaussian Splating (3DGS) has garnered significant attention for its high-quality rendering and fast inference speed. Yet, due to the unstructured and irregular nature of Gaussian point clouds, ensuring accurate geometry reconstruction remains difficult. Existing methods primarily focus on geometry regularization, with common approaches including primitive-based and dual-model frameworks. However, the former suffers from inherent conflicts between rendering and reconstruction, while the latter is computationally and storage-intensive. To address these challenges, we propose CarGS, a unified model leveraging Contribution-adaptive regularization to achieve simultaneous, high-quality rendering and surface reconstruction. The essence of our framework is learning adaptive contribution for Gaussian primitives by squeezing the knowledge from geometry regularization into a compact MLP. Additionally, we introduce a geometry-guided densification strategy with clues from both normals and Signed Distance Fields (SDF) to improve the capability of capturing high-frequency details. Our design improves the mutual learning of the two tasks, meanwhile its unified structure doesn't require separate models as in dual-model based approaches, guaranteeing efficiency. Extensive experiments demonstrate CarGS's ability to achieve state-of-the-art (SOTA) results in both rendering fidelity and reconstruction accuracy while maintaining real-time speed and minimal storage size. + +# 1. Introduction + +Novel view synthesis (NVS) and surface reconstruction from multi-view images are both challenging and critical + +![](images/c957c27f9c1ce164332089e34e54cbfc552ea924e3a6143e1a1d5cda2cba9fa9.jpg) +Figure 1. The performance of CarGS in TnT datasets, achieving high-quality rendering and reconstruction in a unified model. The storage size of the model is represented by the radius of the circle. + +tasks in the fields of computer vision and graphics. Both academia and industry are actively pursuing unified scene representations that enable high-quality rendering and accurate geometry reconstruction concurrently to address the growing demands of applications in areas such as augmented reality (AR), virtual reality (VR), robotics, and autonomous driving. However, simultaneously achieving SOTA performance in both rendering and reconstruction with a unified model is non-trivial, as an inherent conflict exists between the two tasks [3, 7, 9]. + +Revisiting recent approaches, we identify two solutions aimed at mitigating the conflict. The first branch, which we refer to as "primitive-based", introduces additional geometric regularization to 3D Gaussian primitives to improve surface reconstruction. For example, Sugar [7] regularizes 3D Gaussians to align more closely with surfaces. Although this approach ensures parameter efficiency, it leads to a misaligned prioritization between appearance and geometry, eventually degrading rendering performance. The second branch, termed "dual-model based", seeks to guide the learning of one task through information derived from the other. For instance, GSDF [30] introduces a dual-model + +architecture that jointly supervises the 3D Gaussian Splating model for rendering and the SDF model for reconstruction. However, this strategy necessitates separate models for training and inference, resulting in much larger computational and storage costs. Consequently, a natural question arises: Is it possible to develop an efficient framework (fast speed with small storage) that is also effective (SOTA performances on both tasks)? + +Before answering this question, we first present an intriguing observation that motivates us to pursue the "unified goal" from the perspective of learning adaptive contribution of Gaussian primitives. Specifically, with the baseline model PGSR [3] trained on the "Barn" scene from the Tanks and Temples dataset [11], we first calculate the Chamfer Distance (CD) from the center of each Gaussian to the ground truth surface. Then, as shown in Fig. 2, when retaining $50\%$ of the Gaussian points by discarding those with larger CD values, we find that reconstruction quality (F1) is surprisingly maintained, while rendering quality (PSNR) experiences a sharp decline. Conversely, when additional Gaussian points with smaller CDs are dropped, reconstruction quality begins to deteriorate, while rendering quality remains relatively stable. This observation suggests that, 1) despite geometric regularization encouraging Gaussian primitives to closely adhere to the surface [3, 9], a significant number of floaters persist in the scene. 2) These floaters contribute less to reconstruction but are essential for rendering quality. More precisely, each Gaussian primitive contributes differently to each task. + +Therefore, an intuitive idea is to build a unified model with adaptive contributions for each Gaussian primitive. This "unified" structure contrasts with dual-model designs, significantly reducing storage and computational demands from the outset. The "adaptiveness" mitigates conflicts in Gaussian primitive contributions when jointly optimizing for rendering and reconstruction tasks. To this end, we propose CarGS: a unified Gaussian Splitting model based on Contribution-adaptive regularization. Firstly, we squeeze geometry regularization knowledge into a lightweight residual structure to adaptively adjust each Gaussian primitive's contribution for rendering and reconstruction. Then, to further capture geometric details, we introduce a geometry-guided densification strategy that controls Gaussian expansion using clues from both normals and SDF. Experimental results show that our method achieves state-of-the-art reconstruction performance within a unified model, without compromising rendering quality. Remarkably, our model consumes only $9\%$ of the storage and $40\%$ of the training time required by GSDF [30] (dual-model based) while running $100\times$ faster. Moreover, its rendering and reconstruction performance surpasses both GSDF and PGSR [3] (primitive-based). We believe our approach establishes a strong baseline for unified rendering and reconstruction. + +![](images/197d7f31c35dd377593f5c548dd8e11e294c31fff8e4ad6455d0c059edf4cdbd.jpg) +Figure 2. The performance of the Baseline model [3] across various $\mathrm{CD\%}$ thresholds demonstrates that each Gaussian primitive contributes differently to specific tasks. + +In summary, our main contributions are as follows: 1) We identify the key bottleneck in developing a unified, efficient, and effective framework for both rendering and reconstruction. 2) We propose to learn the adaptive contribution of each Gaussian primitive, addressing the inherent conflict in creating a unified framework for rendering and reconstruction. 3) We introduce a geometry-guided densification strategy to enhance the model's ability to capture fine details. 4) Our unified framework simultaneously achieves high-quality rendering and reconstruction while consuming significantly less storage than previous methods (see Fig. 1). + +# 2. Related work + +# 2.1. Novel View Synthesis + +Novel View Synthesis aims to generate new perspectives of a scene using a limited collection of existing views. + +NVS based on Neural Radiance Fields. A pivotal advancement in this field was the introduction of Neural Radiance Fields (NeRF) [16]. NeRF employs MLPs to map 3D spatial locations along rays to color and density values, which are subsequently aggregated through neural volumetric rendering to produce high-quality images. Subsequent research has focused on enhancing the capabilities of NeRF models through feature-grid representations [5, 12, 20, 24] and improving rendering speed using baking techniques [8, 22, 29]. Recently, NeRF has been adapted to address challenges such as anti-aliasing [1] and the modeling of unbounded scenes [2, 34]. + +NVS based on 3D Gaussian Splatting. Recently, 3D Gaussian Splatting [10] has emerged as a transformative approach in neural rendering by utilizing anisotropic 3D Gaussians as primitives. This method sorts Gaussians by depth and rasterizes them onto a 2D screen through alpha blending, resulting in impressive rendering quality and real-time speed. Subsequent research has enhanced 3D Gaussian Splatting by improving rendering quality [4, 21, 31] and expanding its applicability to large-scale scenes [15, 26, 27, 33]. For example, Mip-splatting [31] introduces + +anti-aliasing techniques to mitigate pixelation and blurring in high-resolution renders, thereby enhancing visual sharpness and consistency across viewpoints. Scaffold-GS [15] employs anchor points to manage local 3D Gaussians and dynamically predicts their attributes based on viewing direction and distance, thereby optimizing the adaptive rendering process. + +While these NVS methods excel in view synthesis, they inherently overlook the accurate reconstruction of scene geometry, resulting in blurred volumetric density fields that impede the extraction of high-quality surfaces. + +# 2.2. Neural Surface Reconstruction + +The success of neural rendering has sparked significant interest in neural surface reconstruction [6, 13, 19, 25, 28]. By leveraging coordinate-based neural networks, these methods encode scene geometry using occupancy fields or SDF values, facilitating a smooth and differentiable representation of 3D structures. Though NeRF-based methods with volume rendering yield smooth, complete surfaces, they struggle with fine detail capture and slow optimization. To address this, recent work [13] incorporates multi-resolution hashed grids from Instant-NGP [17], enhancing representation and achieving SOTA results. Yet, large-scale scene reconstruction remains inefficient; e.g., Neuralangelo [13] takes 128 GPU hours to reconstruct a single Tanks and Temples scene [11]. + +Motivated by the impressive efficiency of 3DGS in NVS, recent research has investigated the use of anisotropic 3D Gaussians as foundational primitives for scene geometry representation. SuGaR [24] aligns 3D Gaussians with potential surfaces by approximating them as 2D planar primitives with binary opacity. 2D Gaussian Splatting [9] replaces 3D Gaussians with 2D counterparts, resulting in more accurate ray-splat intersections and regularizing depth maps to achieve a denser distribution of Gaussian primitives. PGSR [3] employs both single-view and multi-view geometric regularization to attain globally consistent geometry. Despite these advancements in neural surface reconstruction, a fidelity gap persists, as the explicit regularization of Gaussian primitives constrains rendering quality. To address this limitation, GSDF [30] introduces a dual-model framework that separates rendering (3DGS) and geometry (SDF) tasks, utilizing joint supervision during training to enhance both reconstruction and rendering. However, this design incurs additional training and storage costs due to the requirement of maintaining two independent models. Moreover, the geometry reconstruction branch is not capable of real-time inference. + +Different from existing approaches, our paper presents a unified model designed to simultaneously achieve real-time and robust rendering and reconstruction tasks without compromising geometry accuracy and rendering quality. + +# 3. Method + +This section details the proposed framework. Particularly, Sec. 3.1 presents preliminary studies on 3D Gaussian Splating and its representative anchor-based paradigm. Then in Sec. 3.2, we analyze the problem of contribution conflicts to identify the most significant attributes for geometry reconstruction. Sec. 3.3 introduces Lite-Geo, a lightweight residual module that adapts the geometry contribution of Gaussian primitives. Based on Lite-Geo, we further propose a geometry-guided densification strategy in Sec. 3.4 that leverages geometric cues to align Gaussians with the surface, enhancing both structural fidelity and representational capacity. Finally, in Sec. 3.5, we introduce geometric regularization to guide the optimization of our modules, which improves the quality of geometry reconstruction. + +# 3.1. Preliminary + +3D Gaussian Splatting. 3DGS models the scene with a set of explicit Gaussians that inherit the differentiable properties of volumetric representations, allowing efficient rendering via tile-based rasterization. Initially, each point generated by the Structure-from-Motion (SfM) serves as the position (mean) $\mu$ of a corresponding 3D Gaussian: + +$$ +G (x) = e ^ {- \frac {1}{2} (x - \mu) ^ {T} \Sigma^ {- 1} (x - \mu)}, \tag {1} +$$ + +where $x$ denotes an arbitrary position in the 3D scene, while $\Sigma$ represents the covariance matrix of the 3D Gaussian. $\Sigma$ is constructed using a scaling matrix $S$ and a rotation matrix $R$ to maintain positive semi-definiteness. + +$$ +\Sigma = R S S ^ {T} R ^ {T}, \tag {2} +$$ + +Each 3D Gaussian is assigned with an opacity value $\alpha$ , which is multiplied with $G(x)$ to adjust its contribution weight to rendering and reconstruction during the blending process. 3DGS renders scenes efficiently through tile-based rasterization instead of traditional ray marching. Each 3D Gaussian $G(x)$ is projected as a 2D Gaussian onto the image plane, after which a tile-based rasterizer sorts and blends the 2D Gaussians using $\alpha$ -blending: + +$$ +F \left(x ^ {\prime}\right) = \sum_ {i \in N} f _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \sigma_ {j}\right), \quad \sigma_ {i} = \alpha_ {i} G _ {i} (x) \tag {3} +$$ + +where $x'$ is the queried pixel location in the image plane and $N$ denotes the number of sorted 2D Gaussians associated with the queried pixel. For a Gaussian with 3D center $(x_i, y_i, z_i)$ and color $c_i$ , $f_i$ is assigned to $c_i$ in the rendering task, and to $z_i$ for depth blending in the reconstruction task. Anchor-based Gasussian Splatting. Our goal is to simultaneously achieve high-quality rendering and reconstruction while maintaining efficiency. In our work, we adopt anchor-based GS modeling as our baseline due to its balanced efficiency and effectiveness. A representative work in this + +![](images/822141261f4c3ed97ed83e84d9c01cd1382bfe0b6bb8c7a0f8a883f036b1d68c.jpg) +Figure 3. The illustration of contribution conflict between rendering and reconstruction tasks. + +paradigm, Scaffold-GS [15], enhances original 3DGS by improving scene structure fidelity and incorporating view clues. Specifically, it constructs a hierarchical 3D Gaussian framework with anchor points that encode local scene details and generate local neural Gaussians, + +$$ +\left\{\mu_ {0}, \dots , \mu_ {k - 1} \right\} = x + \left\{o _ {0}, \dots , o _ {k - 1} \right\} \cdot \gamma , \tag {4} +$$ + +where $o_k \in \mathbb{R}^3$ denotes the learned offsets, and $\gamma$ represents the scaling factor. For each anchor, the optimized attributes (color, center, variance, and opacity) of the $k$ neural Gaussians are decoded through an MLP, $M$ , based on the anchor feature $\mathbf{f}$ , as well as the distance $\delta$ and direction $\theta$ from the camera to the anchor point. Specifically, the opacity values for neural Gaussians generated from an anchor point are, + +$$ +\alpha_ {0}, \dots , \alpha_ {k - 1} = M (\mathbf {f}, \delta , \theta), \tag {5} +$$ + +# 3.2. Analysis + +In this section, we analyze the underlying reasons for the rendering performance degradation when geometry regularization is applied to 3DGS primitives. Then, we conduct experiments to analyze the effect of Gaussian's internal attributes to geometry reconstruction. + +Analysis of Contribution Conflicts. In Fig 3, we consider a simple demonstration of alpha-blending along a ray intersecting three Gaussian primitives. As shown, $G_{2}$ is close to the real surface, $G_{1}$ and $G_{3}$ are farther away. According to Eq. 3, the color $C(x')$ and depth $D(x')$ corresponding to the queried pixel along the ray is calculated as follows: + +$$ +C \left(x ^ {\prime}\right) = c _ {1} w _ {1} + c _ {2} w _ {2} + c _ {3} w _ {3}, \tag {6} +$$ + +$$ +D \left(x ^ {\prime}\right) = z _ {1} w _ {1} + z _ {2} w _ {2} + z _ {3} w _ {3} \tag {7} +$$ + +$$ +w _ {i} = \sigma_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \sigma_ {j}), \quad \sigma_ {i} = \alpha_ {i} G _ {i} (x) \tag {8} +$$ + +where $w_{i}$ represents the contribution of the Gaussian and serves as the weighting factor for ray integration. When only applying rendering regularization, $w_{1}$ and $w_{3}$ tend to increase to capture richer textures, as proved in Fig. 2. Conversely, when geometry regularization is applied, the values of $w_{1}$ and $w_{3}$ tend to decrease due to $z_{1}$ and $z_{3}$ deviating from the SDF 0-level set, while $w_{2}$ increases to provide + +
CovOpacityBaseline
F1↑0.460.640.65
PSNR↑23.8428.7228.76
+ +Table 1. The importance of different attributes in Gaussian's contribution for the geometry reconstruction task. + +more consistent depth information. This shift in the contribution weights $w_{i}$ creates conflicts, which can lead to a decline in rendering performance when geometry regularization is simply added when training the rendering task [3, 9]. Analysis of Attributes. As discussed, the degradation in rendering performance after using geometry regularization stems from contribution conflicts. To further investigate this issue, we analyze which internal attributes of the Gaussian's contribution are most susceptible to the effects of geometry regularization. As shown in Eq. 8, Gaussian's contribution $w_{i}$ is defined by its opacity $\alpha$ and Gaussian sampling value $G(x)$ . Lu et al. [15] emphasize that stable Gaussian positions are crucial for maintaining scene structure, suggesting that adjustments to $\mu$ in Eq. 1 remain relatively minor during optimization. Consequently, changes in $G(x)$ are more likely to arise from covariance in the covariance $\Sigma$ in Eq. 1. + +Then, to investigate the roles of $\alpha$ (opacity) and $\Sigma$ (covariance) within the geometry regularization process, we conduct analysis experiments in Tab. 1. Specifically, with the baseline model PGSR [3] on TnT's "Barn" scene, we respectively detach pacity and covariance from the geometry supervision to stop their gradigents and then assess their independent effect on the performance of geometry reconstruction. As shown in Tab. 1, excluding the gradient of opacity from geometry optimization shows minimal impact on reconstruction quality, while excluding covariance's gradient severely impairs geometry learning. Hence, covariance plays the essential role in regulating the contributions of Gaussian primitives, making it important to tailor covariance for both high-quality rendering and reconstruction. + +# 3.3. 3DGS with Adaptive Contribution + +As shown in Fig. 4, existing approaches for realizing rendering and reconstruction are categorized into primitive based [3, 9] and dual-model based [30] frameworks. The former supervises Gaussian primitives with both rendering and geometry losses, leading to conflicting contributions and, consequently, performance degradation. In contrast, dual-model based methods train separate models for each task, i.e., a 3DGS model for rendering and an SDF model for reconstruction. While enhancing both tasks, it requires substantial computational and storage resources (see Tab. 4). Building on the above analysis, we are motivated to propose a unified and efficient contribution-adaptive 3DGS architecture to simultaneously achieve state-of-the-art per + +![](images/f3419967cdbd7fbb8ca1a8e085d1ca851b00fd85c8e5fb7b362ed98f33a1d8f6.jpg) +(a) Primitive based regularization + +![](images/c8502415539db21926705ad4e71172cad425cd7fedf06e7db4563b06dcfb470b.jpg) +(b) Dual-model based regularization + +![](images/2a1b85d16916cd4cbd73b2ed9f04c6b076ed2339d32bc1481f1469a00967b491.jpg) +(c) Contribution-adaptive regularization +Figure 4. The illustration of the geometry regularization GS framework: (a) Primitive based regularization paradigm. (b) Dual-model-based regularization paradigm. (c) the proposed contribution adaptive regularization paradigm. + +formances by alleviating the adverse effects on rendering when using geometry regularization. + +Geo-MLP. In Sec. 3.2, we have identified covariance as the primary factor contributing to conflicts in geometry and rendering. Thus, we propose decoupling Gaussian contributions by predicting an additional covariance parameter dedicated to geometry reconstruction, while retaining the original covariance for rendering. As outlined in Eq. 2, covariance is decomposed into scaling and rotation matrices. Therefore, we first attempt to introduce an auxiliary MLP that takes the anchor feature $\mathbf{f}$ and view direction as inputs, predicting the scaling and rotation matrix. The predicted geometry-related scaling and rotation are supervised by reconstruction loss, as shown in Fig. 4 (c). However, nothing is as simple as it seems. As visualized in Fig. 5 (top), this straightforward design results in overfitting in geometry estimation. This challenge arises because conventional geometric regularization primarily enforces continuity and consistency, but lacks accurate supervision for depth, which is predominantly informed by rendering-based supervision. + +Lite-Geo. To address this challenge, we introduce Lite-Geo, designed to inherit the Gaussian primitive's contribution from the rendering task, which encodes inherent + +depth information after rendering optimization. This implicit depth information helps mitigate overfitting brought by the sole geometry reconstruction loss applied to GeoMLP. More precisely, we structure the learning of covariance for geometry reconstruction as a residual update, where it is obtained by adding the offset predicted by GeoMLP $(M_{\Sigma}^{geo})$ to the rendering covariance. + +$$ +y = M _ {\Sigma} ^ {r g b} \left(\mathbf {f}, \theta ; \phi_ {1}\right) + \triangle y, \tag {9} +$$ + +$$ +\triangle y = M _ {\Sigma} ^ {g e o} (\mathbf {f}, \theta ; \phi_ {2}), \phi_ {2} = \lambda \phi_ {1} \tag {10} +$$ + +where $\phi_{1}$ and $\phi_{2}$ denote the parameters of $M_{\Sigma}^{rgb}$ and $M_{\Sigma}^{geo}$ , respectively. $y$ and $\triangle y$ are 7-dimensional vectors, with the first three dimensions representing scale and the remaining four representing rotation through quaternions. This approach allows the depth information in rendering covariance to effectively inform the depth estimation in geometry reconstruction. To further reduce noise and enhance convergence, we initialize $\phi_{1}$ based on $\phi_{2}$ using a controllable scaling factor $\lambda$ . Specifically, the rendering task is pre-trained independently for several epochs before incorporating the Lite-Geo module. + +As shown in Fig. 5, this residual structure effectively suppresses geometry overfitting. It also constrains outliers with high Chamfer Distance values that negatively impact rendering and reconstruction quality. + +# 3.4. Geometry-guided Densification + +We further enhance the distribution of Gaussian primitives through a geometry-guided densification strategy. Our motivation stems from the observation that previous geometry-based densification methods [30] rely solely on SDF values to control Gaussian growing, which limits their ability to capture fine details in high-frequency regions and often results in blurred rendering and overly smooth reconstructions. As shown in Fig. 6, to better accommodate high-frequency details, we introduce normal vector clues alongside the original SDF information to guide Gaussian growing. Specifically, for a Gaussian anchor located at position $x$ and the SDF value within a certain range $\theta$ , we evaluate its normal difference with the normal map by $n = N_d(x) - N(x)$ . The SDF is obtained from the depth map by $s = D(x) - Z(x)$ . Detailed calculations of normals and SDF are provided in the supplementary. We then define the criteria for Gaussian growing as follows: + +$$ +\epsilon_ {g} = \nabla_ {g} + \omega_ {g} (\zeta (s) \cdot (\omega_ {n} n \text {i f} \zeta (s) < \theta \text {e l s e} 1)), \tag {11} +$$ + +$$ +\zeta (s) = \exp \left(- \frac {s ^ {2}}{2 \sigma^ {2}}\right), \tag {12} +$$ + +where $\nabla_g$ is the averaged gradient of the Gaussian primitives accumulated over several training iterations. $\zeta (s)$ + +![](images/4ed157a0b46cfb3cda6752412c430330136b34ef908d9023151d169b495e29da.jpg) +Figure 5. The illustration of geometry overfitting. If the Geo-MLP is solely associated with geometry reconstruction, notable overfitting is observed. By leveraging the approximate depth knowledge provided by rendering, Lite-Geo can effectively reduce overfitting and suppress noise where the Chamfer Distance exceeds the threshold. + +![](images/2c8d625293b8bff845f906d89a8f9b628aa546c578f3574b0ab12b0bd5167404.jpg) +Figure 6. The illustration of high-frequency details, along with the proposed densification strategy which facilitates Gaussian growing with clues from both normals and SDF. + +is a Gaussian function that transforms the SDF to a positive number, which monotonically decreases as the distance from the zero level set increases. The parameter $\omega_{g}$ controls the influence of geometric guidance. New Gaussian primitives are generated when $\epsilon_{g}$ exceeds a predefined threshold. + +# 3.5. Optimization + +Relying solely on photometric loss to optimize 3DGS introduces noise into the output, as multi-view 3D reconstruction lacks sufficient constraints [2, 34] for geometric continuity. To improve the recovery of shape detail, we adopt the geometry regularization from PGSR [3], incorporating additional plane constraint and cross-view constraint. To ensure the paper is self-contained, we provide a detailed explanation of these two regularization terms. + +Plane Constraint. To ensure local depth-normal consistency, each pixel and its neighbors should define a relatively smooth plane. Using the depths of a pixel and its neighboring pixels, we can generate a normal map, which is then optimized to align with the rendered normal map, thereby enforcing geometric consistency. The loss based on the plane constraint is defined as follows: + +$$ +\mathcal {L} _ {\text {p l a n e}} = \frac {1}{| \mathcal {W} |} \sum_ {x \in \mathcal {W}} \| N _ {d} (x) - N (x) \| _ {1}, \tag {13} +$$ + +$N_{d}(x)$ and $N(x)$ represent the normal derived from the depth gradient and the normal of the splats, respectively. $\mathcal{W}$ denotes the image pixels. + +Cross-view Constraint. Plane constraint ensures local depth-normal consistency but lacks cross-view consistency. Hence, cross-view constraintsd should be introduced to improve global geometric consistency. Specifically, a pixel $x_{r}$ in the reference frame, associated with a normal $n_r$ and distance $d_{r}$ , can be transformed to a corresponding pixel $x_{n}$ in a neighboring frame with the homography matrix $H_{rn}$ . After forward projection from the reference frame to the neighboring frame, a backward projection $H_{nr}$ is applied. Then the consistency is ensured by minimizing the forward-backward projection error: + +$$ +\mathcal {L} _ {c r o s s} = \frac {1}{| \mathcal {W} |} \sum_ {x _ {r} \in \mathcal {W}} \| x _ {r} - H _ {n r} H _ {r n} x _ {r} \|. \tag {14} +$$ + +Final Loss. Finally, we optimize our model by minimizing, + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \alpha \mathcal {L} _ {d} + \beta \mathcal {L} _ {n} \tag {15} +$$ + +where $\mathcal{L}_c$ is the RGB reconstruction loss that combines L1 loss with D-SSIM from [10], while $\mathcal{L}_{\mathrm{plane}}$ and $\mathcal{L}_{\mathrm{cross}}$ represent the geometry regularization terms. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. To validate the effectiveness of our method, we conduct experiments on diverse real-world datasets, including both indoor and outdoor environments. With the camera poses provided by these datasets, we employ COLMAP [23] to generate an initial sparse point cloud for each scene. We select large and complex scenes from the TnT dataset [11] to evaluate the quality of surface reconstruction. The widely used Mip-NeRF360 dataset [2] is used to assess novel view synthesis performance. + +Evaluation protocols. For fair comparisons with previous methods, we select three widely used image evaluation metrics to validate novel view synthesis: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), + +
F1-Score↑PSNR↑
COLMAPNeuSNeurlangelo2DGSGOFPGSRCarGSNeuS2DGSRaDe-GSGOFPGSRCarGS
Barn0.550.290.700.360.510.660.6026.3628.7926.4526.7128.8028.91
Caterpillar0.010.290.360.230.410.410.4325.2124.2322.4522.9023.0425.16
Courthouse0.110.170.280.130.280.210.1723.5523.5119.7321.4421.8123.54
Ignatius0.220.830.890.440.680.800.7823.2723.8221.5722.2022.2124.77
Meetingroom0.190.240.320.160.280.290.3525.3826.1525.5225.7224.7227.27
Truck0.190.450.480.260.580.600.6523.7126.8523.7224.3226.1226.41
Mean0.210.380.500.300.460.500.5024.5825.5623.2423.8824.4526.01
+ +Table 2. Quantitative result on Tanks and Temples dataset. CarGS achieves the SOTA performance on surface reconstruction and image synthesis. The color green indicates the best result and blue indicates the second best one. + +
Indoor scenesOutdoor scenesAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ImplicitNeRF26.840.7900.37021.460.4580.51524.150.6240.443
Deep Blending26.400.8440.26121.540.5240.36423.970.6840.313
INGP29.150.8800.21622.900.5660.37126.030.7230.294
NeuS25.100.7890.31921.930.6290.60023.740.7200.439
Explicit3DGS30.990.9260.19924.240.7050.28327.620.8030.246
Scaffold-GS31.240.9250.18525.360.7390.20328.300.8320.194
2DGS30.390.9230.18324.330.7090.28427.030.8040.239
PGSR30.410.9300.16724.450.7300.22427.430.8300.193
CarGS30.850.9280.17124.510.7410.24627.680.8340.209
+ +Table 3. Quantitative result of rendering quality on Mip-NeRF360 dataset. The proposed CarGS achieves the best average performance on novel view synthesis compared with the representative implicit Nerf-based and explicit GS-based methods. + +and learned perceptual image patch similarity (LPIPS) [35]. We employ the F1-score to assess the surface quality on the TnT dataset. Apart from the commonly used metrics, we additionally report the storage size (MB) for model compactness and the rendering speed (FPS) for efficiency. + +Implementation Details. Most of our training strategies and hyper-parameters are consistent with those used in 3DGS [10]. Specifically, we adopt a similar approach to halt densification at 15,000 iterations and optimize all models for 30,000 iterations. All MLPs in our method are 2-layer networks with ReLU activation function, and each hidden layer contains 32 units. The loss weights $\alpha$ and $\beta$ in Eq. 15 are set to 0.01 and 0.2, respectively. We begin by rendering depth for each training view, then apply the TSDF Fusion algorithm [18] to create the corresponding TSDF field, from which we subsequently extract the mesh [14]. All experiments were conducted on an Nvidia RTX 3090 GPU. + +# 4.2. Result Analysis + +Rendering Comparison. As shown in Tab. 2 and Tab. 3, our method achieves rendering quality comparable to recent state-of-the-art 3DGS and NeRF-based methods. Unlike + +prior primitive-based approaches that apply shared Gaussian primitives for both rendering and reconstruction (e.g., 2DGS [9], PGSR [3]), our approach adaptively adjusts contributions to meet specific task requirements, allowing high-quality geometry reconstruction without sacrificing rendering fidelity. Our method also diverges from dual-model approaches (e.g., GSDF [30]) by supporting joint training within a unified network for both tasks. As shown in Fig. 7, our approach shows better quality in rendering texture-less regions, where conventional 3D Gaussians struggle due to low accumulated gradients. + +Reconstruction Comparison. We compare our method, CarGS, with current state-of-the-art neural surface reconstruction approaches, including NeuS [25] and NeuralAngelo [13], as well as recent 3DGS-based methods like 2DGS [9], GOF [32], and PGSR [3]. Reconstruction performance is quantitatively evaluated on the TNT dataset, where CarGS achieves an F1 score comparable to NeuralAngelo and PGSR, outperforming other existing methods, as shown in Tab.2. Fig.7 illustrates visual results from various Gaussian splatting approaches, with CarGS generating smoother and more precise shapes. + +![](images/10f53ffa674da7687e85fb27c29015ea1e021814eb14b5cb2d70850c5c5fa235.jpg) +Figure 7. Qualitative comparison on Tanks and Temples dataset. In terms of rendering, our approach surpasses the baseline methods in effectively suppressing blurry regions. For reconstruction, we assess surface quality by visualizing the normal map generated from the reconstructed mesh. CarGS demonstrates advancements in capturing details of the scene compared to other approaches. + +
PSNR↑F1↑Time↓Storage↓FPS↑
NeuS23.710.45>24h1.4GB<1
Neurlangelo25.430.48>24h4.2GB<1
GSDF26.050.463h1.7GB<1
Scaffold-GS25.840.2326.5m0.10GB101
PGSR26.120.601.2h0.42GB103
CarGS26.410.651.2h0.16GB90
+ +Efficiency Comparison. In terms of efficiency, we compare the training time, rendering speed, and storage requirements of different methods on the "Truck" scene from the TnT dataset, as shown in Tab. 4. FPS includes both depth (reconstruction) and RGB generation (rendering). Our method significantly reduces storage requirements, effectively inheriting the compact characteristics of anchor-based models. Compared to the dual-model approach GSDF, our method substantially reduces training time and achieves real-time rendering. In comparison to the Gaussian primitive regularization method PGSR, our model maintains high rendering and reconstruction quality with only a small FPS decrease. Notably, our method only consumes $38\%$ storage than PGSR. + +# 4.3. Ablation Studies + +We evaluate the effectiveness of each individual module using the 'Truck' scenes from the TnT dataset. As shown in Tab. 5 (a), directly applying the geometry loss proposed in [3] to the baseline Scaffold-GS [15] reveals significant limi + +Table 4. Comparison of efficiency. Our CarGS maintains high rendering and reconstruction quality while achieving notable computational and storage efficiency. + +
MethodsGeo-MLPLite-GeoGeo-DensifyF1↑PSNR↑
(a)0.3825.72
(b)✓0.4425.75
(c)✓✓0.6226.13
(d)✓✓✓0.6526.41
+ +Table 5. Ablation study on the Truck scene of TnT dataset. + +tations in reconstruction quality. (b) While employing GeoMLP to adjust the contribution of primitives leads to noticeable improvements in reconstruction quality, overfitting to the geometry loss indicates there is still room for further enhancement. (c) By incorporating the proposed Lite-Geo, the model effectively utilizes depth information from rendering supervision, resulting in a substantial performance boost for both tasks. (d) Finally, the inclusion of geometry-guided densification enables the Gaussian primitives to capture finer details, further enhancing overall performance. + +# 5. Conclusion + +This work proposes CarGS, a unified model that leverages Contribution-adaptive Regularization to effectively address the conflicting requirements of high-quality rendering and accurate surface reconstruction. By introducing a compact MLP for adaptive contribution learning and a geometry-guided densification strategy, CarGS enhances the mutual learning between rendering and reconstruction tasks. Extensive experiments validate that CarGS achieves state-of-the-art results in both rendering fidelity and reconstruction accuracy, while maintaining real-time speed and minimal storage requirements. We believe our approach establishes a strong baseline for unified rendering and reconstruction. + +# 6. Acknowledgements + +This work was supported by the National Science Fund for Distinguished Young Scholars (No. 62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001). + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2, 6 +[3] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 3, 4, 6, 7, 8 +[4] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2 +[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2 +[6] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 3 +[7] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 1 +[8] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5875-5884, 2021. 2 +[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 7 +[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time + +radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 6, 7 +[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 3, 6 +[12] Jonas Kulhanek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18458-18469, 2023. 2 +[13] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3, 7 +[14] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 7 +[15] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2, 3, 4, 8 +[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[17] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3 +[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. Ieee, 2011. 7 +[19] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3 +[20] Yansong Qu, Yuze Wang, and Yue Qi. Sg-nerf: Semantic-guided point-based neural radiance fields. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 570-575. IEEE, 2023. 2 +[21] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 2 +[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with + +thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2 +[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6 +[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2, 3 +[25] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3, 7 +[26] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 2 +[27] Yuze Wang, Junyi Wang, Ruicheng Gao, Yansong Qu, Wantong Duan, Shuo Yang, and Yue Qi. Look at the sky: Sky-aware efficient 3d gaussian splattering in the wild. IEEE Transactions on Visualization and Computer Graphics, 2025. 2 +[28] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3 +[29] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 2 +[30] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2, 3, 4, 5, 7 +[31] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2 +[32] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 7 +[33] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2 +[34] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6 +[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 \ No newline at end of file diff --git a/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/images.zip b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a1b9d3e3204b498c498f9a82ccd21c580a0a07e5 --- /dev/null +++ b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:109b7e859ebe8e20541a2444a2b304592bf03d36625afa8d141ffb3db527592a +size 639387 diff --git a/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/layout.json b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f60121e9a8a0d61c1983e9b729013ab2952a2a89 --- /dev/null +++ b/2025/Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization/layout.json @@ -0,0 +1,7839 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "text", + "content": "Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": "You Shen" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": ", Zhipeng Zhang" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": ", Xinyang Li" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": ", Yansong Qu" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": ", Yu Lin" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": ", Shengchuan Zhang" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": ", Liujuan Cao" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": "Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University \n" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 560, + 220 + ], + "type": "text", + "content": "School of Artificial Intelligence, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 297, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 619 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 619 + ], + "type": "text", + "content": "Representing 3D scenes from multiview images is a core challenge in computer vision and graphics, which requires both precise rendering and accurate reconstruction. Recently, 3D Gaussian Splating (3DGS) has garnered significant attention for its high-quality rendering and fast inference speed. Yet, due to the unstructured and irregular nature of Gaussian point clouds, ensuring accurate geometry reconstruction remains difficult. Existing methods primarily focus on geometry regularization, with common approaches including primitive-based and dual-model frameworks. However, the former suffers from inherent conflicts between rendering and reconstruction, while the latter is computationally and storage-intensive. To address these challenges, we propose CarGS, a unified model leveraging Contribution-adaptive regularization to achieve simultaneous, high-quality rendering and surface reconstruction. The essence of our framework is learning adaptive contribution for Gaussian primitives by squeezing the knowledge from geometry regularization into a compact MLP. Additionally, we introduce a geometry-guided densification strategy with clues from both normals and Signed Distance Fields (SDF) to improve the capability of capturing high-frequency details. Our design improves the mutual learning of the two tasks, meanwhile its unified structure doesn't require separate models as in dual-model based approaches, guaranteeing efficiency. Extensive experiments demonstrate CarGS's ability to achieve state-of-the-art (SOTA) results in both rendering fidelity and reconstruction accuracy while maintaining real-time speed and minimal storage size." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 640, + 137, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 137, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 137, + 653 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 661, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 296, + 685 + ], + "type": "text", + "content": "Novel view synthesis (NVS) and surface reconstruction from multi-view images are both challenging and critical" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 243, + 555, + 397 + ], + "blocks": [ + { + "bbox": [ + 315, + 243, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 555, + 397 + ], + "type": "image", + "image_path": "c957c27f9c1ce164332089e34e54cbfc552ea924e3a6143e1a1d5cda2cba9fa9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 402, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 436 + ], + "type": "text", + "content": "Figure 1. The performance of CarGS in TnT datasets, achieving high-quality rendering and reconstruction in a unified model. The storage size of the model is represented by the radius of the circle." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 449, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 556, + 568 + ], + "type": "text", + "content": "tasks in the fields of computer vision and graphics. Both academia and industry are actively pursuing unified scene representations that enable high-quality rendering and accurate geometry reconstruction concurrently to address the growing demands of applications in areas such as augmented reality (AR), virtual reality (VR), robotics, and autonomous driving. However, simultaneously achieving SOTA performance in both rendering and reconstruction with a unified model is non-trivial, as an inherent conflict exists between the two tasks [3, 7, 9]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 570, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 557, + 714 + ], + "type": "text", + "content": "Revisiting recent approaches, we identify two solutions aimed at mitigating the conflict. The first branch, which we refer to as \"primitive-based\", introduces additional geometric regularization to 3D Gaussian primitives to improve surface reconstruction. For example, Sugar [7] regularizes 3D Gaussians to align more closely with surfaces. Although this approach ensures parameter efficiency, it leads to a misaligned prioritization between appearance and geometry, eventually degrading rendering performance. The second branch, termed \"dual-model based\", seeks to guide the learning of one task through information derived from the other. For instance, GSDF [30] introduces a dual-model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 146, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 146, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 146, + 703 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 136, + 712 + ], + "type": "text", + "content": "Equal Contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16346" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 168 + ], + "type": "text", + "content": "architecture that jointly supervises the 3D Gaussian Splating model for rendering and the SDF model for reconstruction. However, this strategy necessitates separate models for training and inference, resulting in much larger computational and storage costs. Consequently, a natural question arises: Is it possible to develop an efficient framework (fast speed with small storage) that is also effective (SOTA performances on both tasks)?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 298, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 298, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 298, + 422 + ], + "type": "text", + "content": "Before answering this question, we first present an intriguing observation that motivates us to pursue the \"unified goal\" from the perspective of learning adaptive contribution of Gaussian primitives. Specifically, with the baseline model PGSR [3] trained on the \"Barn\" scene from the Tanks and Temples dataset [11], we first calculate the Chamfer Distance (CD) from the center of each Gaussian to the ground truth surface. Then, as shown in Fig. 2, when retaining " + }, + { + "bbox": [ + 56, + 171, + 298, + 422 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 171, + 298, + 422 + ], + "type": "text", + "content": " of the Gaussian points by discarding those with larger CD values, we find that reconstruction quality (F1) is surprisingly maintained, while rendering quality (PSNR) experiences a sharp decline. Conversely, when additional Gaussian points with smaller CDs are dropped, reconstruction quality begins to deteriorate, while rendering quality remains relatively stable. This observation suggests that, 1) despite geometric regularization encouraging Gaussian primitives to closely adhere to the surface [3, 9], a significant number of floaters persist in the scene. 2) These floaters contribute less to reconstruction but are essential for rendering quality. More precisely, each Gaussian primitive contributes differently to each task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "text", + "content": "Therefore, an intuitive idea is to build a unified model with adaptive contributions for each Gaussian primitive. This \"unified\" structure contrasts with dual-model designs, significantly reducing storage and computational demands from the outset. The \"adaptiveness\" mitigates conflicts in Gaussian primitive contributions when jointly optimizing for rendering and reconstruction tasks. To this end, we propose CarGS: a unified Gaussian Splitting model based on Contribution-adaptive regularization. Firstly, we squeeze geometry regularization knowledge into a lightweight residual structure to adaptively adjust each Gaussian primitive's contribution for rendering and reconstruction. Then, to further capture geometric details, we introduce a geometry-guided densification strategy that controls Gaussian expansion using clues from both normals and SDF. Experimental results show that our method achieves state-of-the-art reconstruction performance within a unified model, without compromising rendering quality. Remarkably, our model consumes only " + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "text", + "content": " of the storage and " + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "text", + "content": " of the training time required by GSDF [30] (dual-model based) while running " + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "text", + "content": " faster. Moreover, its rendering and reconstruction performance surpasses both GSDF and PGSR [3] (primitive-based). We believe our approach establishes a strong baseline for unified rendering and reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 68, + 553, + 188 + ], + "blocks": [ + { + "bbox": [ + 316, + 68, + 553, + 188 + ], + "lines": [ + { + "bbox": [ + 316, + 68, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 316, + 68, + 553, + 188 + ], + "type": "image", + "image_path": "197d7f31c35dd377593f5c548dd8e11e294c31fff8e4ad6455d0c059edf4cdbd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 229 + ], + "type": "text", + "content": "Figure 2. The performance of the Baseline model [3] across various " + }, + { + "bbox": [ + 313, + 194, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\mathrm{CD\\%}" + }, + { + "bbox": [ + 313, + 194, + 555, + 229 + ], + "type": "text", + "content": " thresholds demonstrates that each Gaussian primitive contributes differently to specific tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 243, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 374 + ], + "type": "text", + "content": "In summary, our main contributions are as follows: 1) We identify the key bottleneck in developing a unified, efficient, and effective framework for both rendering and reconstruction. 2) We propose to learn the adaptive contribution of each Gaussian primitive, addressing the inherent conflict in creating a unified framework for rendering and reconstruction. 3) We introduce a geometry-guided densification strategy to enhance the model's ability to capture fine details. 4) Our unified framework simultaneously achieves high-quality rendering and reconstruction while consuming significantly less storage than previous methods (see Fig. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 387, + 399, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 387, + 399, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 387, + 399, + 399 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 407, + 437, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 437, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 437, + 420 + ], + "type": "text", + "content": "2.1. Novel View Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 425, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 449 + ], + "type": "text", + "content": "Novel View Synthesis aims to generate new perspectives of a scene using a limited collection of existing views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 593 + ], + "type": "text", + "content": "NVS based on Neural Radiance Fields. A pivotal advancement in this field was the introduction of Neural Radiance Fields (NeRF) [16]. NeRF employs MLPs to map 3D spatial locations along rays to color and density values, which are subsequently aggregated through neural volumetric rendering to produce high-quality images. Subsequent research has focused on enhancing the capabilities of NeRF models through feature-grid representations [5, 12, 20, 24] and improving rendering speed using baking techniques [8, 22, 29]. Recently, NeRF has been adapted to address challenges such as anti-aliasing [1] and the modeling of unbounded scenes [2, 34]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "NVS based on 3D Gaussian Splatting. Recently, 3D Gaussian Splatting [10] has emerged as a transformative approach in neural rendering by utilizing anisotropic 3D Gaussians as primitives. This method sorts Gaussians by depth and rasterizes them onto a 2D screen through alpha blending, resulting in impressive rendering quality and real-time speed. Subsequent research has enhanced 3D Gaussian Splatting by improving rendering quality [4, 21, 31] and expanding its applicability to large-scale scenes [15, 26, 27, 33]. For example, Mip-splatting [31] introduces" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16347" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "anti-aliasing techniques to mitigate pixelation and blurring in high-resolution renders, thereby enhancing visual sharpness and consistency across viewpoints. Scaffold-GS [15] employs anchor points to manage local 3D Gaussians and dynamically predicts their attributes based on viewing direction and distance, thereby optimizing the adaptive rendering process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 205 + ], + "type": "text", + "content": "While these NVS methods excel in view synthesis, they inherently overlook the accurate reconstruction of scene geometry, resulting in blurred volumetric density fields that impede the extraction of high-quality surfaces." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 222, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 222, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 222, + 226 + ], + "type": "text", + "content": "2.2. Neural Surface Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 233, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 400 + ], + "type": "text", + "content": "The success of neural rendering has sparked significant interest in neural surface reconstruction [6, 13, 19, 25, 28]. By leveraging coordinate-based neural networks, these methods encode scene geometry using occupancy fields or SDF values, facilitating a smooth and differentiable representation of 3D structures. Though NeRF-based methods with volume rendering yield smooth, complete surfaces, they struggle with fine detail capture and slow optimization. To address this, recent work [13] incorporates multi-resolution hashed grids from Instant-NGP [17], enhancing representation and achieving SOTA results. Yet, large-scale scene reconstruction remains inefficient; e.g., Neuralangelo [13] takes 128 GPU hours to reconstruct a single Tanks and Temples scene [11]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 664 + ], + "type": "text", + "content": "Motivated by the impressive efficiency of 3DGS in NVS, recent research has investigated the use of anisotropic 3D Gaussians as foundational primitives for scene geometry representation. SuGaR [24] aligns 3D Gaussians with potential surfaces by approximating them as 2D planar primitives with binary opacity. 2D Gaussian Splatting [9] replaces 3D Gaussians with 2D counterparts, resulting in more accurate ray-splat intersections and regularizing depth maps to achieve a denser distribution of Gaussian primitives. PGSR [3] employs both single-view and multi-view geometric regularization to attain globally consistent geometry. Despite these advancements in neural surface reconstruction, a fidelity gap persists, as the explicit regularization of Gaussian primitives constrains rendering quality. To address this limitation, GSDF [30] introduces a dual-model framework that separates rendering (3DGS) and geometry (SDF) tasks, utilizing joint supervision during training to enhance both reconstruction and rendering. However, this design incurs additional training and storage costs due to the requirement of maintaining two independent models. Moreover, the geometry reconstruction branch is not capable of real-time inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Different from existing approaches, our paper presents a unified model designed to simultaneously achieve real-time and robust rendering and reconstruction tasks without compromising geometry accuracy and rendering quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": "This section details the proposed framework. Particularly, Sec. 3.1 presents preliminary studies on 3D Gaussian Splating and its representative anchor-based paradigm. Then in Sec. 3.2, we analyze the problem of contribution conflicts to identify the most significant attributes for geometry reconstruction. Sec. 3.3 introduces Lite-Geo, a lightweight residual module that adapts the geometry contribution of Gaussian primitives. Based on Lite-Geo, we further propose a geometry-guided densification strategy in Sec. 3.4 that leverages geometric cues to align Gaussians with the surface, enhancing both structural fidelity and representational capacity. Finally, in Sec. 3.5, we introduce geometric regularization to guide the optimization of our modules, which improves the quality of geometry reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 264, + 394, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 394, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 394, + 277 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "type": "text", + "content": "3D Gaussian Splatting. 3DGS models the scene with a set of explicit Gaussians that inherit the differentiable properties of volumetric representations, allowing efficient rendering via tile-based rasterization. Initially, each point generated by the Structure-from-Motion (SfM) serves as the position (mean) " + }, + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "type": "text", + "content": " of a corresponding 3D Gaussian:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 358, + 553, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 358, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 373, + 358, + 553, + 374 + ], + "type": "interline_equation", + "content": "G (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)}, \\tag {1}", + "image_path": "96ee4ad35792c8e85a71d617084e9feeafdc1a2445033f00500a73db7e5ccf67.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "text", + "content": " denotes an arbitrary position in the 3D scene, while " + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "text", + "content": " represents the covariance matrix of the 3D Gaussian. " + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "text", + "content": " is constructed using a scaling matrix " + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 380, + 554, + 428 + ], + "type": "text", + "content": " to maintain positive semi-definiteness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 400, + 433, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 433, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 400, + 433, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\Sigma = R S S ^ {T} R ^ {T}, \\tag {2}", + "image_path": "bbb8c9f19b2e1556bcf7ba7b12858461b2fd353f39beb850baf32fcc8acedf6a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "text", + "content": "Each 3D Gaussian is assigned with an opacity value " + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "text", + "content": ", which is multiplied with " + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "text", + "content": " to adjust its contribution weight to rendering and reconstruction during the blending process. 3DGS renders scenes efficiently through tile-based rasterization instead of traditional ray marching. Each 3D Gaussian " + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "text", + "content": " is projected as a 2D Gaussian onto the image plane, after which a tile-based rasterizer sorts and blends the 2D Gaussians using " + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 453, + 554, + 549 + ], + "type": "text", + "content": "-blending:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 555, + 553, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 555, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 330, + 555, + 553, + 588 + ], + "type": "interline_equation", + "content": "F \\left(x ^ {\\prime}\\right) = \\sum_ {i \\in N} f _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} (x) \\tag {3}", + "image_path": "e59073812eacec717b27c919aeb134153382a259a22b0d7c78abd99ed34fb4f5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " is the queried pixel location in the image plane and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " denotes the number of sorted 2D Gaussians associated with the queried pixel. For a Gaussian with 3D center " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "(x_i, y_i, z_i)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " is assigned to " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " in the rendering task, and to " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " for depth blending in the reconstruction task. Anchor-based Gasussian Splatting. Our goal is to simultaneously achieve high-quality rendering and reconstruction while maintaining efficiency. In our work, we adopt anchor-based GS modeling as our baseline due to its balanced efficiency and effectiveness. A representative work in this" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16348" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 291, + 140 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 291, + 140 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 291, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 291, + 140 + ], + "type": "image", + "image_path": "822141261f4c3ed97ed83e84d9c01cd1382bfe0b6bb8c7a0f8a883f036b1d68c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 148, + 295, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 169 + ], + "type": "text", + "content": "Figure 3. The illustration of contribution conflict between rendering and reconstruction tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "content": "paradigm, Scaffold-GS [15], enhances original 3DGS by improving scene structure fidelity and incorporating view clues. Specifically, it constructs a hierarchical 3D Gaussian framework with anchor points that encode local scene details and generate local neural Gaussians," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 248, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 248, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 88, + 248, + 295, + 262 + ], + "type": "interline_equation", + "content": "\\left\\{\\mu_ {0}, \\dots , \\mu_ {k - 1} \\right\\} = x + \\left\\{o _ {0}, \\dots , o _ {k - 1} \\right\\} \\cdot \\gamma , \\tag {4}", + "image_path": "bea0a7ee633f7ca61247b8c1b08c105bf15533e8f2be7fb390f7bed38504c2c8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "o_k \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " denotes the learned offsets, and " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " represents the scaling factor. For each anchor, the optimized attributes (color, center, variance, and opacity) of the " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " neural Gaussians are decoded through an MLP, " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": ", based on the anchor feature " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": ", as well as the distance " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " from the camera to the anchor point. Specifically, the opacity values for neural Gaussians generated from an anchor point are," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 361, + 295, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 361, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 117, + 361, + 295, + 373 + ], + "type": "interline_equation", + "content": "\\alpha_ {0}, \\dots , \\alpha_ {k - 1} = M (\\mathbf {f}, \\delta , \\theta), \\tag {5}", + "image_path": "6ccc7a421beb99387d8cf06b06e68829ee9044ff855d7bfe761402325ca90174.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 118, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 118, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 118, + 411 + ], + "type": "text", + "content": "3.2. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 417, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 475 + ], + "type": "text", + "content": "In this section, we analyze the underlying reasons for the rendering performance degradation when geometry regularization is applied to 3DGS primitives. Then, we conduct experiments to analyze the effect of Gaussian's internal attributes to geometry reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "text", + "content": "Analysis of Contribution Conflicts. In Fig 3, we consider a simple demonstration of alpha-blending along a ray intersecting three Gaussian primitives. As shown, " + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "text", + "content": " is close to the real surface, " + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "inline_equation", + "content": "G_{1}" + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "inline_equation", + "content": "G_{3}" + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "text", + "content": " are farther away. According to Eq. 3, the color " + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "inline_equation", + "content": "C(x')" + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "inline_equation", + "content": "D(x')" + }, + { + "bbox": [ + 55, + 476, + 296, + 548 + ], + "type": "text", + "content": " corresponding to the queried pixel along the ray is calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 555, + 295, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 555, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 110, + 555, + 295, + 570 + ], + "type": "interline_equation", + "content": "C \\left(x ^ {\\prime}\\right) = c _ {1} w _ {1} + c _ {2} w _ {2} + c _ {3} w _ {3}, \\tag {6}", + "image_path": "73a2986d578bf884cebac3f7e9c4e5f01cb079be6eb03e8eee4893f1c6cbc150.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 112, + 576, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 576, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 112, + 576, + 295, + 590 + ], + "type": "interline_equation", + "content": "D \\left(x ^ {\\prime}\\right) = z _ {1} w _ {1} + z _ {2} w _ {2} + z _ {3} w _ {3} \\tag {7}", + "image_path": "5f9d3fe98a2b9142fffd3e9352370a8e5ff75ab2d9a819bb043cfa0bf237dda2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 96, + 594, + 295, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 594, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 96, + 594, + 295, + 626 + ], + "type": "interline_equation", + "content": "w _ {i} = \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} (x) \\tag {8}", + "image_path": "3a432da2caf5a43cb2d0b18841b98440f05159597c9d83c73701bfc7344f8965.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " represents the contribution of the Gaussian and serves as the weighting factor for ray integration. When only applying rendering regularization, " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{3}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " tend to increase to capture richer textures, as proved in Fig. 2. Conversely, when geometry regularization is applied, the values of " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{3}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " tend to decrease due to " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "z_{1}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "z_{3}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " deviating from the SDF 0-level set, while " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{2}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " increases to provide" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 554, + 119 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 554, + 119 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 554, + 119 + ], + "type": "table", + "html": "
CovOpacityBaseline
F1↑0.460.640.65
PSNR↑23.8428.7228.76
", + "image_path": "27e4aee1effbdb9c39e043372ef91abf1e7decb08b999f10790f096ed1d005b9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 128, + 555, + 150 + ], + "lines": [ + { + "bbox": [ + 313, + 128, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 555, + 150 + ], + "type": "text", + "content": "Table 1. The importance of different attributes in Gaussian's contribution for the geometry reconstruction task." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": "more consistent depth information. This shift in the contribution weights " + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": " creates conflicts, which can lead to a decline in rendering performance when geometry regularization is simply added when training the rendering task [3, 9]. Analysis of Attributes. As discussed, the degradation in rendering performance after using geometry regularization stems from contribution conflicts. To further investigate this issue, we analyze which internal attributes of the Gaussian's contribution are most susceptible to the effects of geometry regularization. As shown in Eq. 8, Gaussian's contribution " + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": " is defined by its opacity " + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": " and Gaussian sampling value " + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": ". Lu et al. [15] emphasize that stable Gaussian positions are crucial for maintaining scene structure, suggesting that adjustments to " + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": " in Eq. 1 remain relatively minor during optimization. Consequently, changes in " + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": " are more likely to arise from covariance in the covariance " + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 171, + 555, + 363 + ], + "type": "text", + "content": " in Eq. 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "text", + "content": "Then, to investigate the roles of " + }, + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "text", + "content": " (opacity) and " + }, + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "text", + "content": " (covariance) within the geometry regularization process, we conduct analysis experiments in Tab. 1. Specifically, with the baseline model PGSR [3] on TnT's \"Barn\" scene, we respectively detach pacity and covariance from the geometry supervision to stop their gradigents and then assess their independent effect on the performance of geometry reconstruction. As shown in Tab. 1, excluding the gradient of opacity from geometry optimization shows minimal impact on reconstruction quality, while excluding covariance's gradient severely impairs geometry learning. Hence, covariance plays the essential role in regulating the contributions of Gaussian primitives, making it important to tailor covariance for both high-quality rendering and reconstruction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 540, + 499, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 499, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 499, + 553 + ], + "type": "text", + "content": "3.3. 3DGS with Adaptive Contribution" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "As shown in Fig. 4, existing approaches for realizing rendering and reconstruction are categorized into primitive based [3, 9] and dual-model based [30] frameworks. The former supervises Gaussian primitives with both rendering and geometry losses, leading to conflicting contributions and, consequently, performance degradation. In contrast, dual-model based methods train separate models for each task, i.e., a 3DGS model for rendering and an SDF model for reconstruction. While enhancing both tasks, it requires substantial computational and storage resources (see Tab. 4). Building on the above analysis, we are motivated to propose a unified and efficient contribution-adaptive 3DGS architecture to simultaneously achieve state-of-the-art per" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16349" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 292, + 141 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 292, + 141 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 292, + 141 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 292, + 141 + ], + "type": "image", + "image_path": "f3419967cdbd7fbb8ca1a8e085d1ca851b00fd85c8e5fb7b362ed98f33a1d8f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 142, + 234, + 152 + ], + "lines": [ + { + "bbox": [ + 113, + 142, + 234, + 152 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 234, + 152 + ], + "type": "text", + "content": "(a) Primitive based regularization" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 159, + 293, + 255 + ], + "blocks": [ + { + "bbox": [ + 59, + 159, + 293, + 255 + ], + "lines": [ + { + "bbox": [ + 59, + 159, + 293, + 255 + ], + "spans": [ + { + "bbox": [ + 59, + 159, + 293, + 255 + ], + "type": "image", + "image_path": "c8502415539db21926705ad4e71172cad425cd7fedf06e7db4563b06dcfb470b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 256, + 239, + 266 + ], + "lines": [ + { + "bbox": [ + 109, + 256, + 239, + 266 + ], + "spans": [ + { + "bbox": [ + 109, + 256, + 239, + 266 + ], + "type": "text", + "content": "(b) Dual-model based regularization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 276, + 293, + 354 + ], + "blocks": [ + { + "bbox": [ + 59, + 276, + 293, + 354 + ], + "lines": [ + { + "bbox": [ + 59, + 276, + 293, + 354 + ], + "spans": [ + { + "bbox": [ + 59, + 276, + 293, + 354 + ], + "type": "image", + "image_path": "2a1b85d16916cd4cbd73b2ed9f04c6b076ed2339d32bc1481f1469a00967b491.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 357, + 252, + 367 + ], + "lines": [ + { + "bbox": [ + 110, + 357, + 252, + 367 + ], + "spans": [ + { + "bbox": [ + 110, + 357, + 252, + 367 + ], + "type": "text", + "content": "(c) Contribution-adaptive regularization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 375, + 295, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 420 + ], + "type": "text", + "content": "Figure 4. The illustration of the geometry regularization GS framework: (a) Primitive based regularization paradigm. (b) Dual-model-based regularization paradigm. (c) the proposed contribution adaptive regularization paradigm." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "text", + "content": "formances by alleviating the adverse effects on rendering when using geometry regularization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "content": "Geo-MLP. In Sec. 3.2, we have identified covariance as the primary factor contributing to conflicts in geometry and rendering. Thus, we propose decoupling Gaussian contributions by predicting an additional covariance parameter dedicated to geometry reconstruction, while retaining the original covariance for rendering. As outlined in Eq. 2, covariance is decomposed into scaling and rotation matrices. Therefore, we first attempt to introduce an auxiliary MLP that takes the anchor feature " + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "content": " and view direction as inputs, predicting the scaling and rotation matrix. The predicted geometry-related scaling and rotation are supervised by reconstruction loss, as shown in Fig. 4 (c). However, nothing is as simple as it seems. As visualized in Fig. 5 (top), this straightforward design results in overfitting in geometry estimation. This challenge arises because conventional geometric regularization primarily enforces continuity and consistency, but lacks accurate supervision for depth, which is predominantly informed by rendering-based supervision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Lite-Geo. To address this challenge, we introduce Lite-Geo, designed to inherit the Gaussian primitive's contribution from the rendering task, which encodes inherent" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": "depth information after rendering optimization. This implicit depth information helps mitigate overfitting brought by the sole geometry reconstruction loss applied to GeoMLP. More precisely, we structure the learning of covariance for geometry reconstruction as a residual update, where it is obtained by adding the offset predicted by GeoMLP " + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "inline_equation", + "content": "(M_{\\Sigma}^{geo})" + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": " to the rendering covariance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 167, + 553, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 167, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 378, + 167, + 553, + 182 + ], + "type": "interline_equation", + "content": "y = M _ {\\Sigma} ^ {r g b} \\left(\\mathbf {f}, \\theta ; \\phi_ {1}\\right) + \\triangle y, \\tag {9}", + "image_path": "50deeadded06a71939b71f616b79acc415f1eb101e79b3f5fb9af9b8867176cf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 198, + 553, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 198, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 367, + 198, + 553, + 212 + ], + "type": "interline_equation", + "content": "\\triangle y = M _ {\\Sigma} ^ {g e o} (\\mathbf {f}, \\theta ; \\phi_ {2}), \\phi_ {2} = \\lambda \\phi_ {1} \\tag {10}", + "image_path": "39e51f37486991482b62e6fe40f81fe3b34241b628b7b3011e89c217443991fc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": " denote the parameters of " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "M_{\\Sigma}^{rgb}" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "M_{\\Sigma}^{geo}" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\triangle y" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": " are 7-dimensional vectors, with the first three dimensions representing scale and the remaining four representing rotation through quaternions. This approach allows the depth information in rendering covariance to effectively inform the depth estimation in geometry reconstruction. To further reduce noise and enhance convergence, we initialize " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": " using a controllable scaling factor " + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 217, + 555, + 350 + ], + "type": "text", + "content": ". Specifically, the rendering task is pre-trained independently for several epochs before incorporating the Lite-Geo module." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 350, + 554, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 398 + ], + "type": "text", + "content": "As shown in Fig. 5, this residual structure effectively suppresses geometry overfitting. It also constrains outliers with high Chamfer Distance values that negatively impact rendering and reconstruction quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 406, + 484, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 484, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 484, + 418 + ], + "type": "text", + "content": "3.4. Geometry-guided Densification" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "text", + "content": "We further enhance the distribution of Gaussian primitives through a geometry-guided densification strategy. Our motivation stems from the observation that previous geometry-based densification methods [30] rely solely on SDF values to control Gaussian growing, which limits their ability to capture fine details in high-frequency regions and often results in blurred rendering and overly smooth reconstructions. As shown in Fig. 6, to better accommodate high-frequency details, we introduce normal vector clues alongside the original SDF information to guide Gaussian growing. Specifically, for a Gaussian anchor located at position " + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "text", + "content": " and the SDF value within a certain range " + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "text", + "content": ", we evaluate its normal difference with the normal map by " + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "inline_equation", + "content": "n = N_d(x) - N(x)" + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "text", + "content": ". The SDF is obtained from the depth map by " + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "inline_equation", + "content": "s = D(x) - Z(x)" + }, + { + "bbox": [ + 313, + 423, + 555, + 627 + ], + "type": "text", + "content": ". Detailed calculations of normals and SDF are provided in the supplementary. We then define the criteria for Gaussian growing as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 636, + 555, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 636, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 326, + 636, + 555, + 651 + ], + "type": "interline_equation", + "content": "\\epsilon_ {g} = \\nabla_ {g} + \\omega_ {g} (\\zeta (s) \\cdot (\\omega_ {n} n \\text {i f} \\zeta (s) < \\theta \\text {e l s e} 1)), \\tag {11}", + "image_path": "aaa40d3d5555325b9eab83259467d6c61c45b4685a1a697a902e04169b5a1e40.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 658, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 658, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 386, + 658, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\zeta (s) = \\exp \\left(- \\frac {s ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {12}", + "image_path": "1fb44820488b12b033d8bd8f3024a5af68fb2c572d5560d6084dd9db78474546.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\nabla_g" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " is the averaged gradient of the Gaussian primitives accumulated over several training iterations. " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\zeta (s)" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16350" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 553, + 205 + ], + "type": "image", + "image_path": "4ed157a0b46cfb3cda6752412c430330136b34ef908d9023151d169b495e29da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 555, + 242 + ], + "type": "text", + "content": "Figure 5. The illustration of geometry overfitting. If the Geo-MLP is solely associated with geometry reconstruction, notable overfitting is observed. By leveraging the approximate depth knowledge provided by rendering, Lite-Geo can effectively reduce overfitting and suppress noise where the Chamfer Distance exceeds the threshold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 248, + 282, + 323 + ], + "blocks": [ + { + "bbox": [ + 69, + 248, + 282, + 323 + ], + "lines": [ + { + "bbox": [ + 69, + 248, + 282, + 323 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 282, + 323 + ], + "type": "image", + "image_path": "2c8d625293b8bff845f906d89a8f9b628aa546c578f3574b0ab12b0bd5167404.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 295, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 365 + ], + "type": "text", + "content": "Figure 6. The illustration of high-frequency details, along with the proposed densification strategy which facilitates Gaussian growing with clues from both normals and SDF." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 375, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 437 + ], + "type": "text", + "content": "is a Gaussian function that transforms the SDF to a positive number, which monotonically decreases as the distance from the zero level set increases. The parameter " + }, + { + "bbox": [ + 55, + 375, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\omega_{g}" + }, + { + "bbox": [ + 55, + 375, + 296, + 437 + ], + "type": "text", + "content": " controls the influence of geometric guidance. New Gaussian primitives are generated when " + }, + { + "bbox": [ + 55, + 375, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon_{g}" + }, + { + "bbox": [ + 55, + 375, + 296, + 437 + ], + "type": "text", + "content": " exceeds a predefined threshold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 141, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 141, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 141, + 454 + ], + "type": "text", + "content": "3.5. Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 459, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 554 + ], + "type": "text", + "content": "Relying solely on photometric loss to optimize 3DGS introduces noise into the output, as multi-view 3D reconstruction lacks sufficient constraints [2, 34] for geometric continuity. To improve the recovery of shape detail, we adopt the geometry regularization from PGSR [3], incorporating additional plane constraint and cross-view constraint. To ensure the paper is self-contained, we provide a detailed explanation of these two regularization terms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 638 + ], + "type": "text", + "content": "Plane Constraint. To ensure local depth-normal consistency, each pixel and its neighbors should define a relatively smooth plane. Using the depths of a pixel and its neighboring pixels, we can generate a normal map, which is then optimized to align with the rendered normal map, thereby enforcing geometric consistency. The loss based on the plane constraint is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 642, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 642, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 96, + 642, + 295, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p l a n e}} = \\frac {1}{| \\mathcal {W} |} \\sum_ {x \\in \\mathcal {W}} \\| N _ {d} (x) - N (x) \\| _ {1}, \\tag {13}", + "image_path": "7f34bb81aa551c300cd715ea7c1f4d7bba856e063a2e1025810258c3c8397463.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{d}(x)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "N(x)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " represent the normal derived from the depth gradient and the normal of the splats, respectively. " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denotes the image pixels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "content": "Cross-view Constraint. Plane constraint ensures local depth-normal consistency but lacks cross-view consistency. Hence, cross-view constraintsd should be introduced to improve global geometric consistency. Specifically, a pixel " + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "content": " in the reference frame, associated with a normal " + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "inline_equation", + "content": "n_r" + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "content": " and distance " + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "inline_equation", + "content": "d_{r}" + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "content": ", can be transformed to a corresponding pixel " + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "content": " in a neighboring frame with the homography matrix " + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "inline_equation", + "content": "H_{rn}" + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "content": ". After forward projection from the reference frame to the neighboring frame, a backward projection " + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "inline_equation", + "content": "H_{nr}" + }, + { + "bbox": [ + 313, + 251, + 555, + 382 + ], + "type": "text", + "content": " is applied. Then the consistency is ensured by minimizing the forward-backward projection error:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 391, + 555, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 391, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 349, + 391, + 555, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c r o s s} = \\frac {1}{| \\mathcal {W} |} \\sum_ {x _ {r} \\in \\mathcal {W}} \\| x _ {r} - H _ {n r} H _ {r n} x _ {r} \\|. \\tag {14}", + "image_path": "d120037c12daac0290365acfed417c12583b9ebe1e0e3fe88a3289da2dc7f7cf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 429, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 441 + ], + "type": "text", + "content": "Final Loss. Finally, we optimize our model by minimizing," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 451, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 451, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 387, + 451, + 553, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\alpha \\mathcal {L} _ {d} + \\beta \\mathcal {L} _ {n} \\tag {15}", + "image_path": "4c318937a475726d81b924baaa5bf28650507dad83edab233a7f3298b56ee471.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "text", + "content": " is the RGB reconstruction loss that combines L1 loss with D-SSIM from [10], while " + }, + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{plane}}" + }, + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cross}}" + }, + { + "bbox": [ + 313, + 473, + 554, + 510 + ], + "type": "text", + "content": " represent the geometry regularization terms." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 397, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 397, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 397, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "content": "Datasets. To validate the effectiveness of our method, we conduct experiments on diverse real-world datasets, including both indoor and outdoor environments. With the camera poses provided by these datasets, we employ COLMAP [23] to generate an initial sparse point cloud for each scene. We select large and complex scenes from the TnT dataset [11] to evaluate the quality of surface reconstruction. The widely used Mip-NeRF360 dataset [2] is used to assess novel view synthesis performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Evaluation protocols. For fair comparisons with previous methods, we select three widely used image evaluation metrics to validate novel view synthesis: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM)," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16351" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 71, + 547, + 215 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 547, + 215 + ], + "type": "table", + "html": "
F1-Score↑PSNR↑
COLMAPNeuSNeurlangelo2DGSGOFPGSRCarGSNeuS2DGSRaDe-GSGOFPGSRCarGS
Barn0.550.290.700.360.510.660.6026.3628.7926.4526.7128.8028.91
Caterpillar0.010.290.360.230.410.410.4325.2124.2322.4522.9023.0425.16
Courthouse0.110.170.280.130.280.210.1723.5523.5119.7321.4421.8123.54
Ignatius0.220.830.890.440.680.800.7823.2723.8221.5722.2022.2124.77
Meetingroom0.190.240.320.160.280.290.3525.3826.1525.5225.7224.7227.27
Truck0.190.450.480.260.580.600.6523.7126.8523.7224.3226.1226.41
Mean0.210.380.500.300.460.500.5024.5825.5623.2423.8824.4526.01
", + "image_path": "03a70c51a27fffe933904b6498f7e040cf3598da5d7055543a3517d34d9a5310.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 251, + 553, + 413 + ], + "blocks": [ + { + "bbox": [ + 55, + 218, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 555, + 240 + ], + "type": "text", + "content": "Table 2. Quantitative result on Tanks and Temples dataset. CarGS achieves the SOTA performance on surface reconstruction and image synthesis. The color green indicates the best result and blue indicates the second best one." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 251, + 553, + 413 + ], + "lines": [ + { + "bbox": [ + 57, + 251, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 57, + 251, + 553, + 413 + ], + "type": "table", + "html": "
Indoor scenesOutdoor scenesAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ImplicitNeRF26.840.7900.37021.460.4580.51524.150.6240.443
Deep Blending26.400.8440.26121.540.5240.36423.970.6840.313
INGP29.150.8800.21622.900.5660.37126.030.7230.294
NeuS25.100.7890.31921.930.6290.60023.740.7200.439
Explicit3DGS30.990.9260.19924.240.7050.28327.620.8030.246
Scaffold-GS31.240.9250.18525.360.7390.20328.300.8320.194
2DGS30.390.9230.18324.330.7090.28427.030.8040.239
PGSR30.410.9300.16724.450.7300.22427.430.8300.193
CarGS30.850.9280.17124.510.7410.24627.680.8340.209
", + "image_path": "e81ff514c379b3b5288354b2ffa60a734575bbcffd72862f299557564b29d60f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 555, + 440 + ], + "type": "text", + "content": "Table 3. Quantitative result of rendering quality on Mip-NeRF360 dataset. The proposed CarGS achieves the best average performance on novel view synthesis compared with the representative implicit Nerf-based and explicit GS-based methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 448, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 295, + 508 + ], + "type": "text", + "content": "and learned perceptual image patch similarity (LPIPS) [35]. We employ the F1-score to assess the surface quality on the TnT dataset. Apart from the commonly used metrics, we additionally report the storage size (MB) for model compactness and the rendering speed (FPS) for efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "text", + "content": "Implementation Details. Most of our training strategies and hyper-parameters are consistent with those used in 3DGS [10]. Specifically, we adopt a similar approach to halt densification at 15,000 iterations and optimize all models for 30,000 iterations. All MLPs in our method are 2-layer networks with ReLU activation function, and each hidden layer contains 32 units. The loss weights " + }, + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "text", + "content": " in Eq. 15 are set to 0.01 and 0.2, respectively. We begin by rendering depth for each training view, then apply the TSDF Fusion algorithm [18] to create the corresponding TSDF field, from which we subsequently extract the mesh [14]. All experiments were conducted on an Nvidia RTX 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 151, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 151, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 151, + 673 + ], + "type": "text", + "content": "4.2. Result Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Rendering Comparison. As shown in Tab. 2 and Tab. 3, our method achieves rendering quality comparable to recent state-of-the-art 3DGS and NeRF-based methods. Unlike" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 448, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 579 + ], + "type": "text", + "content": "prior primitive-based approaches that apply shared Gaussian primitives for both rendering and reconstruction (e.g., 2DGS [9], PGSR [3]), our approach adaptively adjusts contributions to meet specific task requirements, allowing high-quality geometry reconstruction without sacrificing rendering fidelity. Our method also diverges from dual-model approaches (e.g., GSDF [30]) by supporting joint training within a unified network for both tasks. As shown in Fig. 7, our approach shows better quality in rendering texture-less regions, where conventional 3D Gaussians struggle due to low accumulated gradients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Reconstruction Comparison. We compare our method, CarGS, with current state-of-the-art neural surface reconstruction approaches, including NeuS [25] and NeuralAngelo [13], as well as recent 3DGS-based methods like 2DGS [9], GOF [32], and PGSR [3]. Reconstruction performance is quantitatively evaluated on the TNT dataset, where CarGS achieves an F1 score comparable to NeuralAngelo and PGSR, outperforming other existing methods, as shown in Tab.2. Fig.7 illustrates visual results from various Gaussian splatting approaches, with CarGS generating smoother and more precise shapes." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16352" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 235 + ], + "type": "image", + "image_path": "10f53ffa674da7687e85fb27c29015ea1e021814eb14b5cb2d70850c5c5fa235.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on Tanks and Temples dataset. In terms of rendering, our approach surpasses the baseline methods in effectively suppressing blurry regions. For reconstruction, we assess surface quality by visualizing the normal map generated from the reconstructed mesh. CarGS demonstrates advancements in capturing details of the scene compared to other approaches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 293, + 286, + 415 + ], + "blocks": [ + { + "bbox": [ + 67, + 293, + 286, + 415 + ], + "lines": [ + { + "bbox": [ + 67, + 293, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 286, + 415 + ], + "type": "table", + "html": "
PSNR↑F1↑Time↓Storage↓FPS↑
NeuS23.710.45>24h1.4GB<1
Neurlangelo25.430.48>24h4.2GB<1
GSDF26.050.463h1.7GB<1
Scaffold-GS25.840.2326.5m0.10GB101
PGSR26.120.601.2h0.42GB103
CarGS26.410.651.2h0.16GB90
", + "image_path": "44a95be93c9dd7b3bed7565ef5e024a41c948daea7fdfdbbac478e6227405a04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 473, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 641 + ], + "type": "text", + "content": "Efficiency Comparison. In terms of efficiency, we compare the training time, rendering speed, and storage requirements of different methods on the \"Truck\" scene from the TnT dataset, as shown in Tab. 4. FPS includes both depth (reconstruction) and RGB generation (rendering). Our method significantly reduces storage requirements, effectively inheriting the compact characteristics of anchor-based models. Compared to the dual-model approach GSDF, our method substantially reduces training time and achieves real-time rendering. In comparison to the Gaussian primitive regularization method PGSR, our model maintains high rendering and reconstruction quality with only a small FPS decrease. Notably, our method only consumes " + }, + { + "bbox": [ + 55, + 473, + 295, + 641 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 55, + 473, + 295, + 641 + ], + "type": "text", + "content": " storage than PGSR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 648, + 156, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 156, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 156, + 659 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We evaluate the effectiveness of each individual module using the 'Truck' scenes from the TnT dataset. As shown in Tab. 5 (a), directly applying the geometry loss proposed in [3] to the baseline Scaffold-GS [15] reveals significant limi" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 294, + 556, + 372 + ], + "blocks": [ + { + "bbox": [ + 55, + 423, + 295, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 456 + ], + "type": "text", + "content": "Table 4. Comparison of efficiency. Our CarGS maintains high rendering and reconstruction quality while achieving notable computational and storage efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 294, + 556, + 372 + ], + "lines": [ + { + "bbox": [ + 315, + 294, + 556, + 372 + ], + "spans": [ + { + "bbox": [ + 315, + 294, + 556, + 372 + ], + "type": "table", + "html": "
MethodsGeo-MLPLite-GeoGeo-DensifyF1↑PSNR↑
(a)0.3825.72
(b)✓0.4425.75
(c)✓✓0.6226.13
(d)✓✓✓0.6526.41
", + "image_path": "c38f6aa76367aec8d6c0a92f2c169c1b757ce0aa241d18234430968621e8e645.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 380, + 546, + 390 + ], + "lines": [ + { + "bbox": [ + 321, + 380, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 321, + 380, + 546, + 390 + ], + "type": "text", + "content": "Table 5. Ablation study on the Truck scene of TnT dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 412, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 533 + ], + "type": "text", + "content": "tations in reconstruction quality. (b) While employing GeoMLP to adjust the contribution of primitives leads to noticeable improvements in reconstruction quality, overfitting to the geometry loss indicates there is still room for further enhancement. (c) By incorporating the proposed Lite-Geo, the model effectively utilizes depth information from rendering supervision, resulting in a substantial performance boost for both tasks. (d) Finally, the inclusion of geometry-guided densification enables the Gaussian primitives to capture finer details, further enhancing overall performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "This work proposes CarGS, a unified model that leverages Contribution-adaptive Regularization to effectively address the conflicting requirements of high-quality rendering and accurate surface reconstruction. By introducing a compact MLP for adaptive contribution learning and a geometry-guided densification strategy, CarGS enhances the mutual learning between rendering and reconstruction tasks. Extensive experiments validate that CarGS achieves state-of-the-art results in both rendering fidelity and reconstruction accuracy, while maintaining real-time speed and minimal storage requirements. We believe our approach establishes a strong baseline for unified rendering and reconstruction." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16353" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars (No. 62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 217, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 282, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 296, + 337 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "type": "text", + "content": "[3] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 393, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 296, + 437 + ], + "type": "text", + "content": "[4] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 491 + ], + "type": "text", + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 492, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 296, + 536 + ], + "type": "text", + "content": "[6] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 591 + ], + "type": "text", + "content": "[7] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 592, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 646 + ], + "type": "text", + "content": "[8] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5875-5884, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "type": "text", + "content": "[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[12] Jonas Kulhanek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18458-18469, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "text", + "content": "[13] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[14] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 354 + ], + "type": "text", + "content": "[15] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 408 + ], + "type": "text", + "content": "[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[17] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 521 + ], + "type": "text", + "content": "[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. Ieee, 2011. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "text", + "content": "[19] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[20] Yansong Qu, Yuze Wang, and Yue Qi. Sg-nerf: Semantic-guided point-based neural radiance fields. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 570-575. IEEE, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "text", + "content": "[21] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16354" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "type": "text", + "content": "[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "text", + "content": "[25] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "text", + "content": "[26] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "type": "text", + "content": "[27] Yuze Wang, Junyi Wang, Ruicheng Gao, Yansong Qu, Wantong Duan, Shuo Yang, and Yue Qi. Look at the sky: Sky-aware efficient 3d gaussian splattering in the wild. IEEE Transactions on Visualization and Computer Graphics, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 374 + ], + "type": "text", + "content": "[28] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 431 + ], + "type": "text", + "content": "[29] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 475 + ], + "type": "text", + "content": "[30] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 476, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 296, + 529 + ], + "type": "text", + "content": "[31] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 533, + 296, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 296, + 575 + ], + "type": "text", + "content": "[32] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 577, + 296, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 296, + 620 + ], + "type": "text", + "content": "[33] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 622, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 296, + 654 + ], + "type": "text", + "content": "[34] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 656, + 296, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 656, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 656, + 296, + 711 + ], + "type": "text", + "content": "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16355" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_content_list.json b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2fbcc188a2880bc38f97c3159e8e16b1e3612cef --- /dev/null +++ b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_content_list.json @@ -0,0 +1,1699 @@ +[ + { + "type": "text", + "text": "Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation", + "text_level": 1, + "bbox": [ + 145, + 128, + 851, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Zhu $^{1}$ , Yan Zhu $^{1}$ , Jiayu Xiao $^{1}$ , Tianxiang Xiao $^{1}$ , Yike Ma $^{1}$ , Yucheng Zhang $^{1}$ , Feng Dai $^{1\\dagger}$ $^{1}$ Institute of Computing Technology, Chinese Academy of Sciences", + "bbox": [ + 135, + 202, + 866, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhuhao22z，fdai}@ict.ac.cn", + "bbox": [ + 377, + 241, + 614, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automated crop mapping through Satellite Image Time Series (SITS) has emerged as a crucial avenue for agricultural monitoring and management. However, due to the low resolution and unclear parcel boundaries, annotating pixel-level masks is exceptionally complex and time-consuming in SITS. This paper embraces the weakly supervised paradigm (i.e., only image-level categories available) to liberate the crop mapping task from the exhaustive annotation burden. The unique characteristics of SITS give rise to several challenges in weakly supervised learning: (1) noise perturbation from spatially neighboring regions, and (2) erroneous semantic bias from anomalous temporal periods. To address the above difficulties, we propose a novel method, termed exploring space-time perceptive clues (Exact). First, we introduce a set of spatial clues to explicitly capture the representative patterns of different crops from the most class-relative regions. Besides, we leverage the temporal-to-class interaction of the model to emphasize the contributions of pivotal clips, thereby enhancing the model perception for crop regions. Building upon the space-time perceptive clues, we derive the clue-based CAMs to effectively supervise the SITS segmentation network. Our method demonstrates impressive performance on various SITS benchmarks. Remarkably, the segmentation network trained on Exact-generated masks achieves $95\\%$ of its fully supervised performance, showing the bright promise of weakly supervised paradigm in crop mapping scenario. Our code will be publicly available here.", + "bbox": [ + 89, + 323, + 485, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 760, + 220, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The launch of numerous public and commercial satellites provides broader opportunities to record, analyze, and predict the evolution of crop land [5, 17, 47, 50]. In this context, Satellite Image Time Series (SITS) with $10\\mathrm{m}$ resolution offered by the high-frequency Sentinel-2 (S2) satellites serve as a valuable data source for automated crop map", + "bbox": [ + 89, + 785, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c797dc2c5c990e075a7496532915e26ea68afc19ef2deac339d839bb9aaee48.jpg", + "image_caption": [ + "(a) Spatial Noise Perturbation" + ], + "image_footnote": [], + "bbox": [ + 521, + 305, + 901, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80f40b0adaf35652df8aba154f9a24be2af06926319a47aece727dd2a1e919a6.jpg", + "image_caption": [ + "(b) Temporal Erroneous Semantic Bias", + "Figure 1. Illustration of the two inherent issues arises from spatial and temporal perspectives in SITS. (a) shows noise perturbation from the spatial perspective. We visualize the high-level feature manifold of Dog (natural image) and Barley (SITS) to reveal the distinct spatial properties. The feature dimensions are reduced by t-SNE [48]. (b) shows the erroneous semantic bias induced by anomalous temporal clips. We denote the parcel regions with $\\star$ . The white circles refer to false positive activation regions in CAMs." + ], + "image_footnote": [], + "bbox": [ + 521, + 501, + 901, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ping [6, 14, 22, 28, 35, 56]. The core of crop mapping lies in the semantic segmentation of crop parcels. Recently, many efforts are devoted to exploiting the versatile and powerful relation modeling capability of deep neural networks for this task [20, 21, 34, 46]. While such methods have shown a significant progress, they rely heavily on pixel-level manual annotation, which is notoriously complex and time-consuming [49]. The low resolution of satellite images and the indistinct boundaries between crop parcels complicate the annotation process. Even worse, the varying phenologi", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 91, + 886, + 217, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14036", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal cycles among crops require annotators to meticulously select appropriate acquisition times for annotation.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, one promising solution is weakly supervised semantic segmentation (WSSS), which relies solely on time-efficient annotation form, i.e., image-level categories. The image-level WSSS has commonly studied in natural image domain. Existing methods primarily extract Class Activation Map (CAM) [57] to generate pseudo labels for training semantic segmentation models. The intention of CAM is to identify the regions that highly contribute to the prediction of each class, revealing the shared patterns among the images within the same category. In context of natural images, CAM tends to highlight local discriminative parts of the target object. The reason is that the dispersed intraclass distribution enforces the model to learn a extremely sharp decision boundary, thereby the classifier weights tend to interact with feature representations within more discriminative object regions, as shown in Fig. 1a left. Thus the primary focus of researchers falls into expanding the CAM to identify the entire object semantics for natural images.", + "bbox": [ + 88, + 125, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from WSSS for natural image, crop mapping for SITS mainly faces with two challenges: (1) from the spatial perspective, parcel objects within the same category share uniform appearance and color, exhibiting strong neighboring consistency. The intra-class compactness of local patterns looses the tolerance of the classifier to noise perturbation, resulting in pronounced over-activation phenomena in CAM, as shown in Fig. 1a right. The disparate characteristic lead to the advances in natural image domain cannot directly benefit the SITS crop segmentation. (2) From the temporal perspective, although different crops show distinct phenological cycles and varying characteristics, they may display similar appearance in some specific periods. This confusion imbues wrong semantic bias to the learning process of the model, thereby activating some undesired semantic regions in CAM. An intuitive illustration is shown in Fig. 1b, the temporal clips that deviate from the pivotal semantic affect the perceptual ability of CAM to correct crop regions.", + "bbox": [ + 91, + 398, + 483, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present exploring space-time perceptive clues (Exact), a tailored WSSS framework for crop mapping, to cope with the challenges arisen from the spatial and temporal aspects respectively. Firstly, we introduce a set of spatial clues to explicitly capture the patterns of different crops. Leveraging the filtered CAM as an indicator, we update representative clues by conducting spatial clustering from the most class-relative regions. These clues are then used to regularize the feature space via optimizing the contrastive objective, thereby sharpening the decision boundary of the model and mitigating the perturbations from illusory patterns. Secondly, to cope with the erroneous semantic bias caused by anomalous temporal periods, we propose temporal-aware affinity propagation to emphasize the contributions of pivotal clips for crop perception. In detail, we extract the temporal-", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to-class attention from the model to reweight the temporal sequence embeddings. The modulated representations can be used to model temporal-aware pairwise affinity for propagation on the raw CAM, thus effectively suppressing the undesired semantic regions in a self-supervised manner.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike existing WSSS methods that rely on classifier weights to generate CAMs, we exploit the well-updated space-time perceptive clues to derive the clue-based CAMs (CB-CAMs) as pseudo labels for segmentation. Compared to the raw CAM, the CB-CAMs (1) remarkably suppress the perturbations from spatial and temporal aspects, and (2) delineate the crop regions more precisely, thereby providing more reliable supervision for the subsequent segmentation.", + "bbox": [ + 511, + 167, + 903, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 531, + 290, + 893, + 304 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the WSSS paradigm to SITS crop mapping task to tackle the daunting annotation challenge. To the best of our knowledge, this is the first work that relies solely on image-level categories for crop segmentation.", + "- To overcome the drawbacks arising from the spatial and temporal aspects of SITS, we propose Exact that explores space-time perceptive clues to reduce the noise perturbation and rectify the wrong semantic bias, ultimately providing reliable supervision for SITS segmentation.", + "- We experimentally show that Exact achieves impressive performance on common benchmarks. Using Exact-generated labels for training, SITS segmentation model attains up to $95\\%$ of its fully supervised performance. Our results significantly advance the upper bound of image-level WSSS technique compared to the other domains." + ], + "bbox": [ + 513, + 306, + 903, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 556, + 653, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic segmentation on SITS. Automated crop monitoring through Satellite Image Time Series (SITS) has attracted great interest among researchers and demonstrated considerable social impact [14, 22, 23, 33, 36, 43, 44]. One of the challenging tasks is the semantic segmentation of agricultural parcels. The goal of the network is to learn a mapping function that assigns each pixel in the SITS to the corresponding crop type or background. Some works process SITS inputs by first extracting spatial information and then compressing the temporal dimension. For example, U-ConvLSTM [32] relied on U-Net [37] architecture to encode the spatial dimension, followed by a ConvLSTM [42] for the temporal dimension. Similarly, the FPN-ConvLSTM [34] replaced the U-Net with a Feature Pyramid Network [29] as the spatial encoder. U-TAE [20] compressed the temporal dimension through the outstanding temporal attention mechanism [21]. The other option to encode SITS is the temporal-spatio scheme, which first processes the temporal dimension and then extracts the spatial information. Rußwurm et al. [40] employed bidirectional LSTM to extract temporal features and then used CNNs to integrate spatial informa", + "bbox": [ + 511, + 583, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14037", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c36968eef2f675745a7a03a0e92cc89f0c041882122ae6cd3e3aa4e848f74bbd.jpg", + "image_caption": [ + "(a) Training Pipeline of the Exact" + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 898, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f3199a724496cdbe687d2cd3c9f2eb9f9feeb4fcdff1d5ee32d7950e8317c4f.jpg", + "image_caption": [ + "(b) Explore Spatial Perceptive Clues (§3.2)", + "Figure 2. (a) The training pipeline of Exact. We adopt the Temporal-Spatio scheme to handle the SITS input, which contains two transformer encoders. The first temporal encoder models interactions between acquisition times, then the followed spatial encoder discards the temporal dimension and models interactions between spatial positions. To overcome the difficulties arises from spatial and temporal aspects, we propose two novel technologies in temporal embedding space: (b) Explore Spatial Perceptive Clues to mitigate the noise perturbation (see §3.2) and (c) Temporal-Aware Affinity Propagation to rectify the wrong semantic bias(see §3.3)." + ], + "image_footnote": [], + "bbox": [ + 101, + 282, + 500, + 419 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe2ace1cb5469efb5d32887a328b8b9516c0c4e997e41ac09eed5b8c3b8c0324.jpg", + "image_caption": [ + "(c) Temporal-Aware Affinity Propagation (§3.3)" + ], + "image_footnote": [], + "bbox": [ + 535, + 282, + 897, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion. Recently, TSViT [46] proposed to borrow the powerful dependency modeling capability of the Vision Transformer (ViT) [16] to handle SITS, achieving state-of-the-art performance at lower computational cost. The TSViT also comprehensively illustrated the superiority of the temporal-spatio scheme. Taking a holistic view, we adopt this scheme to process SITS, aiming to achieve an optimal trade-off for automated crop mapping.", + "bbox": [ + 89, + 527, + 485, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly supervised semantic segmentation. Weakly supervised semantic segmentation (WSSS) with image-level labels (i.e. ground-truth object categories) has shown significant success in natural images [13, 18, 52, 53, 59]. Most advanced WSSS methods follow the three-steps pipeline of: 1) training a classification network with image-level labels, 2) obtaining class activation map (CAM) [9, 41, 57] from the well-optimized classification network as pixel-level coarse labels, 3) replacing ground-truth masks with the pseudo labels to train a off-the-shelf segmentation model. The objective of our work focuses on the first and second steps, i.e., how to generate accurate pseudo labels. Numerous works has been proposed to address the under- and over-activation issues of the initial CAM [3, 10, 11, 26, 39, 55] in natural images. To derive the final pseudo labels, CAM requires cumbersome post-processing, including random walks [1, 2] and dense", + "bbox": [ + 89, + 659, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CRF [25] refinement. Thanks to these excellent works, the performance of image-level WSSS in natural images currently achieves $90\\%$ of pixel-level supervision. However, the WSSS networks designed for natural images require significant adaptation to be applied to SITS, and the results are still unsatisfactory due to the distinct data characteristics. In this work, we incorporate WSSS technique into the SITS and overcome the difficulties caused by inherent data properties.", + "bbox": [ + 511, + 527, + 908, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 662, + 604, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first look more closely at the temporal-spatio scheme and class activation map technique (§3.1). Next, we introduce how to explore spatial perceptive clues (§3.2) and conduct temporal-aware affinity propagation (§3.3). Finally, we show the overall objective function and the clue-based CAM generation strategy of Exact (§3.4). The training pipeline of Exact is shown in Fig. 2.", + "bbox": [ + 511, + 686, + 908, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 803, + 653, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The temporal-spatio scheme that first processes the temporal dimension and then extracting the spatial information has shown significant superiority when dealing with SITS data. Following the TSViT [46], we use the variant ViT [16] as the backbone for both temporal and spatial encoders. We", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14038", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "consider a SITS input $\\mathbf{X} \\in \\mathbb{R}^{T \\times C \\times H \\times W}$ with $T$ the length of time series, $C$ the number of channels and $H \\times W$ the spatial dimensions. The input $\\mathbf{X}$ is mapped into a sequence of patch tokens and reshaped to $\\mathbf{Z} \\in \\mathbb{R}^{N_h \\cdot N_w \\times T \\times d}$ , where $N_h = \\left[\\frac{H}{h}\\right]$ , $N_w = \\left[\\frac{W}{w}\\right]$ and $h \\times w$ is the spatial extent for each patch. We then add the temporal position embeddings $\\mathbf{P}_T \\in \\mathbb{R}^{T \\times d}$ and concatenate the temporal multi-class tokens $\\mathbf{Z}_T^{\\mathrm{cls}} \\in \\mathbb{R}^{K \\times d}$ to obtain the input of temporal encoder:", + "bbox": [ + 89, + 89, + 483, + 213 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {T} ^ {\\mathrm {i n}} = \\left[ \\mathbf {Z} _ {T} ^ {\\mathrm {c l s}}, \\mathbf {Z} + \\mathbf {P} _ {T} \\right], \\mathbf {Z} _ {T} ^ {\\mathrm {i n}} \\in \\mathbb {R} ^ {N _ {h} \\cdot N _ {w} \\times (K + T) \\times d} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 237, + 483, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here $K$ denotes the number of categories, $\\mathbf{Z}_T^{\\mathrm{cls}}$ and $\\mathbf{P}_T$ are repeated $N_{h}\\cdot N_{w}$ times to match the spatial shape. With the output feature maps $\\mathbf{Z}_T^{\\mathrm{out}} = [\\mathbf{Z}_T^{\\mathrm{dense}}|\\mathbf{Z}_T^{\\mathrm{seq}}]$ from the temporal encoder, we extract the first $K$ tokens $\\mathbf{Z}_T^{\\mathrm{dense}}\\in \\mathbb{R}^{N_h\\cdot N_w\\times K\\times d}$ and permute the first and second dimensions to serve as input patch tokens $\\mathbf{Z}_S\\in \\mathbb{R}^{K\\times N_h\\cdot N_w\\times d}$ for spatial encoder. These inputs are then combined with spatial multi-class tokens $\\mathbf{Z}_{S}^{\\mathrm{cls}}\\in \\mathbb{R}^{K\\times 1\\times d}$ and spatial position embeddings $\\mathbf{P}_S\\in \\mathbb{R}^{N_h\\cdot N_w\\times d}$ at all $K$ spatial representations:", + "bbox": [ + 89, + 267, + 483, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {S} ^ {\\mathrm {i n}} = \\left[ \\mathbf {Z} _ {S} ^ {\\mathrm {c l s}}, \\mathbf {Z} _ {S} + \\mathbf {P} _ {S} \\right], \\mathbf {Z} _ {S} ^ {\\mathrm {i n}} \\in \\mathbb {R} ^ {K \\times (1 + N _ {h} \\cdot N _ {w}) \\times d} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 412, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the spatial encoder phase, we separate the output features $\\mathbf{Z}_S^{\\mathrm{out}} = [\\mathbf{Z}_S^{\\mathrm{global}}|\\mathbf{Z}_S^{\\mathrm{dense}}]$ to align with different downstream tasks. For classification task, we feed the global tokens $\\mathbf{Z}_S^{\\mathrm{global}}\\in \\mathbb{R}^{K\\times 1\\times d}$ into the classifier to obtain classification logits. For segmentation task, to derive the dense prediction mask, the tokens $\\mathbf{Z}_S^{\\mathrm{dense}}\\in \\mathbb{R}^{K\\times N_h\\cdot N_w\\times d}$ are fed into the segmentation decoder.", + "bbox": [ + 89, + 443, + 483, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Class activation map (CAM) [57] is widely used in WSSS to provide weak annotations that rely on image-level labels. Given a natural image, its feature maps $\\mathbf{F} \\in \\mathbb{R}^{H' \\times W' \\times D}$ are extracted by a classification backbone. To derive the classification score, the feature maps are average pooled and multiplied by the classifier weights $\\boldsymbol{w} \\in \\mathbb{R}^{K \\times D}$ . CAM is generated by weighting and summing each channel in the feature maps with the classifier weights, as follows:", + "bbox": [ + 89, + 549, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} ^ {k} = \\operatorname {R e L U} \\left(\\sum_ {i} \\boldsymbol {w} _ {i} ^ {k} \\cdot \\mathbf {F} _ {:,: i}\\right), \\quad \\forall k \\in K. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 680, + 483, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following most WSSS methods, we normalize $\\mathcal{M}^k$ to the range [0,1] and apply a global threshold to filter out background pixels to obtain final pseudo labels. In the temporal-spatio network, we feed the dense tokens $\\mathbf{Z}_T^{\\mathrm{dense}}$ and $\\mathbf{Z}_S^{\\mathrm{dense}}$ into the classifier and compute the additional classification loss $\\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{aux}}$ to supervise the fused raw CAM for SITS. We provide more details in supplement.", + "bbox": [ + 89, + 732, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Explore Spatial Perceptive Clues", + "text_level": 1, + "bbox": [ + 89, + 848, + 377, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CAM filtering. Given an input SITS $\\mathbf{X}$ and its image-level label $y\\in [0,1]^K$ , we first compute its normalized fused", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CAM $\\mathcal{M} \\in \\mathbb{R}^{N_h \\cdot N_w \\times K}$ by the classifier weights and the output dense tokens of temporal and spatial encoder. We use two threshold scores $\\mu_l$ and $\\mu_h$ to filter out the reliable foreground, background and uncertain regions:", + "bbox": [ + 511, + 89, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {M}} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\mathcal {M} \\leq \\mu_ {l}, \\\\ 1, & \\text {i f} \\mathcal {M} \\geq \\mu_ {h}, \\\\ \\text {i g n o r e ,} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 159, + 906, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clues clustering. To capture the compact patterns of different crops, we establish a group of class-wise representative prototypes, which later serve as perceptive clues to generate high-quality pseudo labels. In crop segmentation task, temporal features often provide more information than spatial context [46], so we choose to perform spatial clustering on temporal dense embeddings $\\mathbf{Z}_T^{\\mathrm{dense}}$ over the whole dataset. Specifically, we build the class-wise positive prototype set $\\mathcal{P}_{\\mathrm{pos}} = \\{\\pmb{p}_{\\mathrm{pos}}^k \\in \\mathbb{R}^{N_p \\times d}\\}_{k=1}^K$ and negative prototype set $\\mathcal{P}_{\\mathrm{neg}} = \\{\\pmb{p}_{\\mathrm{neg}}^k \\in \\mathbb{R}^{N_p \\times d}\\}_{k=1}^K$ , where $N_p$ denotes the number of prototypes. If $k$ -th class appears in a training batch, we update the prototypes $\\pmb{p}^k$ via solving the optimal transport problem [8, 58]. Given a mapping matrix $\\mathbf{C}^k$ that represents the assignment between each pixel and its prototype, it can be referred as an element of the transportation polytope [4]:", + "bbox": [ + 511, + 223, + 906, + 452 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {C} = \\left\\{\\mathbf {C} ^ {k} \\in \\mathbb {R} _ {+} ^ {N _ {p} \\times N _ {k}} \\mid \\mathbf {C} ^ {k} \\mathbf {1} = \\boldsymbol {u}, \\mathbf {C} ^ {k \\top} \\mathbf {1} = \\boldsymbol {r} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 458, + 906, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{k}$ denotes the number of pixels belonging to class $k$ . 1 represents the vectors of ones in appropriate dimensions, $\\mathbf{u}$ and $\\pmb{r}$ are the marginal projections onto the rows and columns of $\\mathbf{C}^k$ . We can maximize the following objective function to optimize the mapping matrix:", + "bbox": [ + 511, + 486, + 906, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathbf {C} ^ {k} \\in \\mathbb {C}} {\\text {m a x i m i z e}} \\operatorname {T r} \\left(\\mathbf {C} ^ {k \\top} \\boldsymbol {p} ^ {k} \\mathbf {Z} ^ {k \\top}\\right) - \\eta \\sum_ {n _ {p}, n _ {k} \\in N _ {p}, N _ {k}} \\mathbf {C} _ {n _ {p} n _ {k}} ^ {k} \\log \\mathbf {C} _ {n _ {p} n _ {k}} ^ {k}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 569, + 906, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here $\\mathbf{Z}^k\\in \\mathbb{R}^{N_k\\times d}$ is the corresponding temporal dense embeddings belonging to class $k$ , $\\eta$ controls the smoothness of entropy regularization term. The continuous approximate solution of Eq. (6) can be obtained through iteratively applying the Sinkhorn-Knopp algorithm [15]. Subsequently, we momentum update the $n_p$ -th prototype of class $k$ according to the assignment matrix and the embeddings:", + "bbox": [ + 511, + 608, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {n _ {p}} ^ {k} = \\alpha \\boldsymbol {p} _ {n _ {p}} ^ {k} + \\frac {1 - \\alpha}{\\left| \\left| \\mathbf {C} _ {n _ {p}} ^ {k} \\right| \\right| _ {1}} \\cdot \\left(\\mathbf {C} _ {n _ {p}} ^ {k} \\mathbf {Z} ^ {k}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 722, + 906, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here $\\alpha \\in [0,1]$ is the momentum coefficient. We leverage the fused CAM $\\hat{\\mathcal{M}}$ as the pseudo labels to update the class-wise positive prototype set $\\mathcal{P}_{\\mathrm{pos}}$ and $1 - \\hat{\\mathcal{M}}$ for the negative set $\\mathcal{P}_{\\mathrm{neg}}$ . Notably, each prototype is not involved in gradient backpropagation to avoid the noise from the classifier.", + "bbox": [ + 511, + 763, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clue-based contrastive learning. Based on the spatial perceptive clues, we introduce the clue-based contrastive learning [7, 12] to regularize the embedding space. More specifically, for each temporal dense embedding $\\mathbf{z}_{n_k}^k$ and its most", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14039", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2738a8636e98134d3e37dceb2686efeda35de10884ee6a1141df6e8631b7f205.jpg", + "image_caption": [ + "Figure 3. Visualization of the clue-based CAM generation. This process is performed after training the classification network." + ], + "image_footnote": [], + "bbox": [ + 107, + 94, + 472, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "relative prototype $\\pmb{p}_{n_p}^k$ , we adopt the cosine distance to measure their similarity [24]:", + "bbox": [ + 89, + 268, + 483, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p} _ {n _ {p}} ^ {k}\\right) = \\frac {\\boldsymbol {z} _ {n _ {k}} ^ {k} \\boldsymbol {p} _ {n _ {p}} ^ {k \\top}}{\\left\\| \\boldsymbol {z} _ {n _ {k}} ^ {k} \\right\\| \\cdot \\left\\| \\boldsymbol {p} _ {n _ {p}} ^ {k} \\right\\| / \\tau}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 308, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ indicates the temperature parameter. Subsequently, we enforce the pixel embedding to closely match its prototype and be distinct from other prototypes:", + "bbox": [ + 89, + 358, + 485, + 405 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c b l}} = \\sum_ {k} \\sum_ {n _ {k}} \\mathbb {1} (\\log (\\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {-}} \\exp \\mathrm {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p}\\right)) - \\mathrm {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p} _ {n _ {p}} ^ {k}\\right)), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 414, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}^{-} = \\{\\mathcal{P}_{\\mathrm{pos}} \\cup \\mathcal{P}_{\\mathrm{neg}} \\setminus \\boldsymbol{p}_{n_p}^k\\}$ and $\\mathbb{1}(\\cdot)$ is an indicator function, being 1 if class $k$ appears in image-level labels and 0 otherwise. Minimizing the above objective can pull pixel embedding closely to the semantic center and push it away from other illusory patterns, thereby sharpening the model decision boundary and mitigating the noise perturbation.", + "bbox": [ + 89, + 462, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Temporal-Aware Affinity Propagation", + "text_level": 1, + "bbox": [ + 89, + 561, + 419, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we propose the temporal-aware affinity propagation to cope with the wrong semantic bias arising from anomalous temporal clips.", + "bbox": [ + 89, + 585, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal-aware affinity mining. In the temporal encoder, the input tokens are normalized and projected into query matrix $\\mathbf{Q} \\in \\mathbb{R}^{(K + T) \\times d}$ and key matrix $\\mathbf{K} \\in \\mathbb{R}^{(K + T) \\times d}$ respectively. Then the self-attention $\\mathcal{A} \\in \\mathbb{R}^{(K + T) \\times (K + T)}$ with respect to $N_h \\cdot N_w$ sequences is computed as below:", + "bbox": [ + 89, + 630, + 483, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d}}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 715, + 483, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without additional computations nor supervision, we can explicitly extract the temporal-to-class attention $\\tilde{\\mathcal{A}}\\in \\mathbb{R}^{T\\times K}$ from the self-attention $\\mathcal{A}$ , which represents the contributions of high-level representations to crop recognition at different temporal clips. We then reweight the temporal sequence embeddings $\\mathbf{Z}_T^{\\mathrm{seq}}$ based on the normalized attention $\\tilde{\\mathcal{A}}$ :", + "bbox": [ + 89, + 760, + 483, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} ^ {k} = \\sum_ {t = 1} ^ {T} \\tilde {\\mathcal {A}} _ {t} ^ {k} \\cdot \\left(\\mathbf {Z} _ {T} ^ {\\text {s e q}}\\right) _ {:, t,:}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 862, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the modulated high-level representation $\\mathcal{V}^k\\in$ $\\mathbb{R}^{N_h\\cdot N_w\\times d}$ enlarges the variation among different semantic crops, thereby modeling more precise pixel relations.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Affinity propagation and guidance. After obtaining the temporal-aware representation $\\mathcal{V}$ , we conduct affinity propagation to suppress erroneous semantic regions on the raw CAM. Particularly, the temporal-aware pairwise affinity of the pixel at position $i$ can be estimated as follows:", + "bbox": [ + 511, + 136, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A f f} _ {j \\in \\mathcal {N} _ {i}} \\left(\\boldsymbol {v} _ {i} ^ {k}, \\boldsymbol {v} _ {j} ^ {k}\\right) = \\exp \\left(\\frac {\\mathcal {S} \\left(\\boldsymbol {v} _ {i} ^ {k} , \\boldsymbol {v} _ {j} ^ {k}\\right)}{\\sigma_ {i} ^ {k}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 219, + 906, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here $\\sigma_{i}^{k}$ is the standard deviation of $\\pmb{v}_{i}^{k}$ and $\\mathcal{N}_i$ indicates the local receptive fields (e.g., 8-way local neighbors [27]). We denoise the raw CAM via iteratively propagating the pairwise affinity from the temporal-aware representation $\\nu$ :", + "bbox": [ + 511, + 268, + 906, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {M}} _ {i} ^ {k} = \\delta_ {i} ^ {k} \\sum_ {j \\in \\mathcal {N} _ {i}} \\operatorname {A f f} \\left(\\boldsymbol {v} _ {i} ^ {k}, \\boldsymbol {v} _ {j} ^ {k}\\right) \\cdot \\mathcal {M} _ {j} ^ {k}, \\quad \\text {w h e r e} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 337, + 905, + 376 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {i} ^ {k} = \\frac {1}{\\sum_ {j \\in \\mathcal {N} _ {i}} \\mathsf {A f f} (\\boldsymbol {v} _ {i} ^ {k} , \\boldsymbol {v} _ {j} ^ {k})}.\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 371, + 754, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the improved activation map, we align the raw CAM with it to effectively guide the learning process:", + "bbox": [ + 511, + 409, + 905, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t a p}} = \\sum_ {k} \\mathbb {1} \\left(\\left| \\tilde {\\mathcal {M}} ^ {k} - \\mathcal {M} ^ {k} \\right|\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 446, + 906, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By minimizing $\\mathcal{L}_{\\mathrm{tap}}$ , we can rectify erroneous activations in the raw CAM and incorporate temporal-aware priors into the embedding space, ultimately benefiting the perceptive clues.", + "bbox": [ + 511, + 488, + 906, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Network Optimization and Clue-based CAM Generation Strategy", + "text_level": 1, + "bbox": [ + 511, + 541, + 905, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the whole objective function for training the classification network consists of four components:", + "bbox": [ + 511, + 578, + 905, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {c l s}} ^ {\\text {a u x}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {c b l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {t a p}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 618, + 906, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here $\\mathcal{L}_{\\mathrm{cls}}$ and $\\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{aux}}$ are the conventional binary cross entropy loss and $\\lambda_{i}$ denotes the weight to rescale the loss terms.", + "bbox": [ + 511, + 643, + 905, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training process, we perform per-pixel perception based on the well-updated space-time clues in temporal dense embedding space to generate clue-based CAMs (CB-CAMs). For each embedding $z_{i}$ in temporal dense embeddings $\\mathbf{Z}_T^{\\mathrm{dense}}$ , we measure the maximum similarity with the positive prototypes and minus the misguide activations with the negative prototypes:", + "bbox": [ + 511, + 674, + 906, + 780 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Y} _ {i} ^ {k} = \\operatorname {R e L u} \\left(\\max _ {\\boldsymbol {p} ^ {+} \\in \\boldsymbol {p} _ {k} ^ {\\text {p o s}}} \\mathrm {S} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {p} ^ {+}\\right) - \\max _ {\\boldsymbol {p} ^ {-} \\in \\boldsymbol {p} _ {k} ^ {\\text {n e g}}} \\mathrm {S} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {p} ^ {-}\\right)\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 787, + 906, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If the class $k$ does not appear in a training image, we set $\\mathcal{Y}^k$ to all zeros. Fig. 3 gives a visualization of this generation process. Finally, the CB-CAMs $\\mathcal{Y}$ are filtered using a global background score to generate the pseudo labels, which are then used to train the SITS semantic segmentation network.", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14040", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e9d2c2347ac2effcf5159ca51cb640e953245389f2791381af07f1be7419d8e0.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Pseudo labels performance of different methods on Germany and PASTIS train set." + ], + "table_body": "
MethodTypeGermany [40]PASTIS [20]
OAmIoUOAmIoU
MCTFormercVPR'22 [52]ViT-PCMECCV'22 [38]LP-CAMCVPR'23 [13]TSCDAAI'23 [54]DuPLCVPR'24 [51]SeCOCVPR'24 [55]RGB70.656.366.749.6
66.852.269.353.2
64.749.167.050.1
68.452.767.251.3
66.250.665.548.7
64.548.363.646.1
baseline+PAMRCVPR'20 [3]+TS-CAMICCV'21 [19]+SIPECVPR'22 [11]+FPRICCV'23 [10]SITS82.574.181.269.5
83.976.382.071.2
81.170.580.467.6
82.073.281.570.1
82.475.681.771.0
+ours-Exact88.3 ↑ 5.880.6 ↑ 6.584.1 ↑ 2.975.6 ↑ 6.1
", + "bbox": [ + 89, + 75, + 545, + 299 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9bc6ccccfa8eba94f83db8868fcd562401c43f7e4f741bac9ec312372b7f4cf8.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Segmentation performance trained with pseudo labels generated by different methods on PASTIS test set." + ], + "table_body": "
MethodSup.OAmIoUratio
ConvLSTM [42]78.250.1
BiCGRU [40]80.556.2
FPN-ConvLSTM [34]81.959.5
Unet-3D [32]G82.360.4-
Unet-3Df [45]82.160.2
U-TAE [20]82.962.4
TSViT [46]83.465.1100%
baseline77.157.788%
+PAMRCVPR'20 [3]78.558.790%
+TS-CAMICCV'21 [19]P76.856.587%
+SIPECVPR'22 [11]77.158.089%
+FPRICCV'23 [10]78.258.490%
+ours-Exact80.3 ↑ 3.261.8 ↑ 4.195%
", + "bbox": [ + 558, + 77, + 903, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparisons with existing WSSS methods in OA(%) and mIoU(%) . RGB: the networks designed for processing natural images. SITs: the networks designed for processing SITS (include four adapted modules from natural image domain). $\\mathcal{G}$ : various SITS segmentation networks supervised by ground-truth labels. $\\mathcal{P}$ : the TSViT segmentation model supervised by different pseudo labels from the methods in SITs. The ratio refers to the proportion of mIoU between fully supervised and weakly supervised of TSViT segmentation.", + "bbox": [ + 88, + 338, + 906, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 414, + 223, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 439, + 281, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets & evaluation metric. We conducted comprehensive experiments on two widely used Satellite-2 time series crop recognition datasets to validate the effectiveness of our approach quantitatively. The PASTIS dataset [20] contains 2433 multi-spectral time series of size $128 \\times 128$ , each series including 10 bands and 33 to 61 temporal observations. It consists of 18 crop types and a background category. Following the settings of [46], we used the fold-1 among the five folds provided in PASTIS. We partitioned each sample into multiple patches and assigned category labels according to the mask annotations (see the supplement for details). The Germany dataset [40] comprises $137k$ field parcels of time series imagery with 13 spectral bands. Each series contains 36 observations and is labeled with 17 crop types. We employed the mean Intersection over Union (mIoU) and pixel-wise overall accuracy (OA) as evaluation metrics, both widely used to measure segmentation performance.", + "bbox": [ + 88, + 458, + 485, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The classification network was trained on 8 NVIDIA RTX 3090 GPUs with batch size 8 for 15k iterations. During the training stage, we used the AdamW optimizer [30] with an initial learning rate of 1e-3 and cosine weight decay [31]. As the raw CAM were unreliable in early iterations, the gradients of $\\mathcal{L}_{\\mathrm{cbl}}$ were backpropagated after the 4k iteration. For the raw CAM filtering, the threshold score $(\\mu_l,\\mu_h)$ was set to (0.2, 0.4). Due to the compact intra-class patterns, we set the number of class-specific prototypes $N_{p}$ to 2, and the momentum used to update the prototypes was set to 0.999. Moreover, the other hyper-parameters $\\eta$ , $\\lambda_1$ , $\\lambda_2$ and $\\tau$ were empirically set to 0.05, 0.01, 0.015 and", + "bbox": [ + 88, + 719, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "0.1, respectively. We employed the original TSViT with a segmentation decoder as the semantic segmentation model for SITS. The training details of the segmentation model exactly followed the settings in [46] without any modifications. Please refer to supplementary materials for more details.", + "bbox": [ + 511, + 415, + 908, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 503, + 718, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Base Architectures. In the evaluation stage, we fuse the raw CAMs generated by the spatial dense embeddings $\\mathbf{Z}_S^{\\mathrm{dense}}$ and temporal dense embeddings $\\mathbf{Z}_T^{\\mathrm{dense}}$ as our baseline.", + "bbox": [ + 511, + 523, + 906, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Comparison with other WSSS Methods on CAMs.", + "text_level": 1, + "bbox": [ + 511, + 575, + 906, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The key of the weakly supervised learning is to provide reliable training pseudo labels for segmentation network. Thus, we conducted comparative experiments on the train splits to validate the quality of pseudo labels.", + "bbox": [ + 511, + 595, + 908, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competing methods adaptation. We reimplemented six well-performing transformer-based WSSS methods designed for natural images as competing methods. To accommodate the temporal dimension, we transformed the input dimensions of SITS into a 3D format $(T \\times C, H, W)$ for these methods. In addition, to enable a fairer comparison between Exact and existing WSSS works, we carefully selected four off-the-shelf modules from other WSSS methods in the natural image domain and adapted them to the temporal-spatio framework. These modules were also reinplemented within the temporal dense embedding space to align with our approach. We attempted various adaptation strategies, more details and analysis available in supplementary materials.", + "bbox": [ + 511, + 657, + 910, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Tab. 1a presents the mIoU and OA of the pseudo labels generated by other WSSS methods and our Exact. As can be seen, the WSSS methods designed for", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14041", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2f69f1662379c00d5dfb8f0f3b684f6ed3590af011cc5d6bef06b6a44aff1bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method\\( \\mathcal{L}_{\\mathrm{cbl}} \\)\\( \\mathcal{L}_{\\mathrm{tap}} \\)CB-CAMRecallOAmIoU
baseline81.181.269.5
✓81.882.472.3
✓✓82.583.073.4
ours-Exact✓✓✓84.884.175.6
", + "bbox": [ + 94, + 88, + 480, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ab1e07de61c4d1bdf3ef1cc2f697ee9a3b0f54d58bfb2a23e657c66da804261e.jpg", + "table_caption": [ + "Table 2. Ablation study results of different components. The baseline is a setting with only image-level classification loss." + ], + "table_footnote": [], + "table_body": "
MethodSpa.Tem.RecallPrecisionOAmIoU
✓81.978.480.566.7
✓80.881.681.068.3
baseline CAM✓✓81.183.581.269.5
", + "bbox": [ + 96, + 241, + 478, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f23413f7fe8c158218f3adf5fb59eb751b03e6404fc821ce6b89860b190d3e4.jpg", + "table_caption": [ + "Table 3. Baseline CAMs from different embedding spaces. Overall, the fused CAM shows the best performance." + ], + "table_footnote": [], + "table_body": "
MethodLow.Ta.RecallPrecisionOAmIoU
baseline81.183.581.269.5
✓80.984.081.169.9
ours-Exact✓81.885.682.472.3
✓✓81.785.282.071.8
", + "bbox": [ + 94, + 377, + 480, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Impact of modeling affinity from different sources. Low and $Ta$ refer to low-level cues and temporal-aware affinity. The temporal-aware affinity remarkably reduces noise.", + "bbox": [ + 89, + 488, + 483, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "natural images only achieve a maximum $53.2\\%$ mIoU (refer to PASTIS) with the limited encoding capability for SITS. Furthermore, due to the distinct data characteristics, the off-the-shelf modules struggle to perform well on SITS, and may even disturb the learning process. These phenomena indicate that the advancements in natural image cannot be directly translated to benefit the SITS domain. By contrast, our Exact overcome the intrinsic challenges from SITS, delivering the best performance compared to others.", + "bbox": [ + 89, + 556, + 483, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Segmentation results trained by pseudo labels.", + "text_level": 1, + "bbox": [ + 89, + 700, + 457, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After obtaining the pseudo labels, we use them as replacement for ground truth to train the SITS semantic segmentation network. We employ TSViT followed by a segmentation head as our segmentation model, which is the top-performing approach in fully supervised oracles.", + "bbox": [ + 89, + 719, + 483, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. In Tab. 1b, we show that the semantic segmentation model can achieve $95\\%$ to its fully supervised performance (refer to mIoU) using pseudo labels generated by Exact. Our best results are $80.3\\%$ OA and $61.8\\%$ mIoU on PASTIS test set, which surpasses the baseline by $3.2\\%$ and $4.1\\%$ , respectively. The experiment results quantitatively imply that the labels quality outperforms other methods by a", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2450f385f3c5704297cf92af160c1ebe02576b1239492bcbac2c4ab13ae829d.jpg", + "image_caption": [ + "(a) baseline" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 709, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3285f440ee71c71d5c8242afe9236af849a91e671f628177f6b9dddce8df9e86.jpg", + "image_caption": [ + "(b)ours-Exact" + ], + "image_footnote": [], + "bbox": [ + 712, + 88, + 903, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f2a286cd46ec9feffd2128e76a6407afdca99f41c2d6906d219bdcef2a581c6.jpg", + "image_caption": [ + "Figure 4. Visualization of temporal feature spaces on PASTIS train set. The feature dimensions are reduced by t-SNE [48].", + "(a) $N_{p}$", + "Figure 5. Effect of the hyper-parameters. (a) the number of class-specific prototypes $N_{p}$ . (b) the temperature of similarity $\\tau$ ." + ], + "image_footnote": [], + "bbox": [ + 514, + 290, + 705, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c3504aaee52cc9b4562328c6fbef05f93393f166ee5a07f0e61ba3bd6535f4b.jpg", + "image_caption": [ + "(b) $\\pmb{\\tau}$" + ], + "image_footnote": [], + "bbox": [ + 705, + 290, + 903, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "large margin. In contrast, WSSS networks in natural image have undergone a series of evolutions to barely achieve $90\\%$ of fully supervised performance. This suggests that SITS WSSS techniques offers greater potential in SITS scenario.", + "bbox": [ + 511, + 465, + 905, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 532, + 676, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, our primary purpose is to demonstrate the collective effectiveness of all components within our approach. We also choose the fused raw CAM as the baseline, and the CAM comparison results reported on PASTIS train set.", + "bbox": [ + 511, + 551, + 906, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All the components matter. Our Exact consists of several components, including spatial perceptive clues exploration, temporal-aware affinity propagation and clue-based CAM generation strategy. We validate the contribution of each module, the results are presented in Tab. 2. We can see that the clue-based contrastive learning $\\mathcal{L}_{\\mathrm{cbl}}$ and the temporal-aware affinity propagation $\\mathcal{L}_{\\mathrm{tap}}$ improves the performance by $1.8\\%$ OA and $3.9\\%$ mIoU, respectively. This suggests that these two modules complement each other, regularizing the entire embedding space and ultimately sharpening the decision boundary. We provide an intuitive comparison in Fig. 4, after introducing both objectives, the intra-class features become more compact while the inter-class features are more separated. Besides, using the well-updated space-time clues to generate CAM (i.e., CB-CAM) within the temporal dense embedding space brings a further improvement of $2.3\\%$ recall, $1.1\\%$ OA and $2.2\\%$ mIoU.", + "bbox": [ + 511, + 613, + 908, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline CAM from different embedding spaces. In Tab. 3, we investigate the performance of the raw CAMs derived", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14042", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81fea31a89be1ab6db254c719ca06134f0075e92638b38685c81730d3e4d49aa.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons. Left: CAMs comparisons on PASTIS train set. The stars represent the corresponding activation category. Right: Outputs of segmentation models trained with different CAMs on the PASTIS test set." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from different dense embedding spaces. The raw CAM obtained from the spatial dense embeddings $\\mathbf{Z}_S^{\\mathrm{dense}}$ falls significantly behind that obtained from the temporal dense embeddings $\\mathbf{Z}_T^{\\mathrm{dense}}$ in terms of precision, OA and mIoU. It demonstrates that the temporal correlations contain more critical information compared to the spatial context in this task. We combine the CAMs from both embedding spaces as our baseline (the last row in Tab. 3) and utilize the resulting pseudo labels to guide spatial clustering.", + "bbox": [ + 88, + 468, + 482, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the temporal-aware affinity. Mining low-level cues affinity (e.g., color and intensity) as additional guidance is prevalent in natural images. We also attempted to incorporate the low-level cues to model pairwise affinity in SITS, the results can be found in Tab. 4. Compared to our proposed temporal-aware affinity, modeling low-level cues offers only limited improvement. This is due to the presence of substantial noise interference within the low-level cues of SITS (e.g., cloud cover and shadow). In contrast, our approach highlights the contributions of pivotal temporal clips within a high-level embedding space, effectively mitigating the wrong semantic bias.", + "bbox": [ + 88, + 609, + 482, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the hyper-parameters. Fig. 5 shows the effects of the prototype number $N_{p}$ and the temperature $\\tau$ . Notably, utilizing only a minimal number of prototypes (i.e., $N_{p} = 2$ ) yields optimal results, whereas increasing the number of prototypes beyond this point brings negative impact on the model. This indicates that the excessive number may enforce prototypes to focus on local discriminative patterns", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "within the class, leading to under-activation issue (the major difficulty in natural image). More analysis of parameters is specifically discussed in supplementary materials.", + "bbox": [ + 511, + 468, + 906, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 518, + 694, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 6 presents the visual comparison between the baseline TSViT-CAMs and the proposed CB-CAMs generated by Exact on PASTIS dataset. The first five columns show the visualization of the CAMs. As we can see, our method remarkably suppress the erroneous regions, whose shape is closer to ground truth masks than baseline. Our high-quality pseudo labels subsequently enhance the segmentation performance compared to the baseline, as shown in last three columns. However, Exact still exhibits limitations in edge processing, which will be the focus in our future work.", + "bbox": [ + 511, + 537, + 908, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Broader Impact", + "text_level": 1, + "bbox": [ + 511, + 696, + 808, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a tailored WSSS approach $\\text{Exact}$ to alleviate the daunting annotation challenge in SITS crop mapping task. $\\text{Exact}$ explores space-time perceptive clues to capture the essential patterns of different crop types, thereby overcoming the issues of spatial noise perturbation and wrong temporal semantic bias. Extensive experiments show the superiority of $\\text{Exact}$ -generated masks both quantitatively and qualitatively. We believe that our method marks a pioneering step in effectively applying WSSS technologies to SITS. If follow-up work can find and resolve the limitations under our framework, there is great potential that only image-level labels are needed for crop mapping in the future.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14043", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by National Key R&D Program of China (2022YFD2001601) and National Natural Science Foundation of China (62372433).", + "bbox": [ + 89, + 114, + 485, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 175, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, 2018. 3", + "[2] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 3", + "[3] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 3, 6", + "[4] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 4", + "[5] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation. In ECCV, 2024. 1", + "[6] Vitus Benson, Claire Robin, Christian Requena-Mesa, Lazaro Alonso, Nuno Carvalhais, José Cortés, Zhihan Gao, Nora Linscheid, Mélanie Weynants, and Markus Reichstein. Multimodal learning for geospatial vegetation forecasting. In CVPR, 2024. 1", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4", + "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4", + "[9] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In WACV, 2018. 3", + "[10] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, 2023. 3, 6", + "[11] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6", + "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4", + "[13] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, 2023. 3, 6", + "[14] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 1, 2" + ], + "bbox": [ + 93, + 200, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 4", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[17] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: Esa's optical high-resolution mission for gmes operational services. Remote Sensing of Environment, 2012. 1", + "[18] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, 2022. 3", + "[19] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In CVPR, 2021. 6", + "[20] Vivien Sainte Fare Garnot and Loic Landrieu. Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In ICCV, 2021. 1, 2, 6", + "[21] Vivien Sainte Fare Garnot, Loic Landrieu, Sebastien Giordano, and Nesrine Chehata. Satellite image time series classification with pixel-set encoders and temporal self-attention. In CVPR, 2020. 1, 2", + "[22] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In CVPR, 2024. 1, 2", + "[23] Pengyu Hao, Yulin Zhan, Li Wang, Zheng Niu, and Muhammad Shakir. Feature selection of time series modis data for early crop classification using random forest: A case study in kansas, usa. Remote Sensing, 2015. 2", + "[24] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, 2023. 5", + "[25] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, 2011. 3", + "[26] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3", + "[27] Wentong Li, Yuqian Yuan, Song Wang, Wenyu Liu, Dongqi Tang, Jianke Zhu, Lei Zhang, et al. Label-efficient segmentation via affinity propagation. In NeurIPS, 2023. 5", + "[28] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In CVPR, 2024. 1", + "[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2", + "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2022. 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14044", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In CVPR Workshops, 2019. 2, 6", + "[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 2024. 2", + "[34] Jorge Andres Chamorro Martinez, Laura Elena Cué La Rosa, Raul Queiroz Feitosa, Ieda Del'Arco Sanches, and Patrick Nigri Happ. Fully convolutional recurrent networks for multideate crop recognition from multitemporal image sequences. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 1, 2, 6", + "[35] Sina Mohammadi, Mariana Belgiu, and Alfred Stein. Improvement in crop mapping from satellite image time series by effectively supervising deep neural networks. ISPRS Journal of Photogrammetry and Remote Sensing, 2023. 1", + "[36] Charlotte Pelletier, Silvia Valero, Jordi Inglada, Nicolas Champion, and Gérard Dedieu. Assessing the robustness of random forests to map land cover with high resolution satellite image time series over large areas. Remote Sensing of Environment, 2016. 2", + "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2", + "[38] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, 2022. 6", + "[39] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In CVPR, 2023. 3", + "[40] Marc Rußwurm and Marco Körner. Multi-temporal land cover classification with sequential recurrent encoders. ISPRS International Journal of Geo-Information, 2018. 2, 6", + "[41] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3", + "[42] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, 2015. 2, 6", + "[43] Sofia Siachalou, Giorgos Mallinis, and Maria Tsakiri-Strati. A hidden markov models approach for crop classification: Linking crop phenology to time series of multi-sensor remote sensing data. Remote Sensing, 2015. 2", + "[44] Haopeng Sun, Yingwei Zhang, Lumin Xu, Sheng Jin, and Yiqiang Chen. Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer. arXiv preprint arXiv:2412.10181, 2024. 2", + "[45] Michail Tarasiou, Riza Alp Güler, and Stefanos Zafeiriou. Context-self contrastive pretraining for crop type semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6", + "[46] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In CVPR, 2023. 1, 3, 4, 6" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Ramon Torres, Paul Snoeij, Dirk Geudtner, David Bibby, Malcolm Davidson, Evert Attema, Pierre Potin, Björn Rommen, Nicolas Floury, Mike Brown, et al. Gmes sentinel-1 mission. Remote Sensing of Environment, 2012. 1", + "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008. 1, 7", + "[49] Sherrie Wang, William Chen, Sang Michael Xie, George Azzari, and David B Lobell. Weakly supervised deep learning for segmentation of remote sensing imagery. Remote Sensing, 2020. 1", + "[50] Curtis E Woodcock, Richard Allen, Martha Anderson, Alan Belward, Robert Bindschadler, Warren Cohen, Feng Gao, Samuel N Goward, Dennis Helder, Eileen Helmer, et al. Free access to landsat imagery. Science, 2008. 1", + "[51] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, 2024. 6", + "[52] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6", + "[53] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[54] Rongtao Xu, Changwei Wang, Jiaxi Sun, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Self correspondence distillation for end-to-end weakly-supervised semantic segmentation. In AAAI, 2023. 6", + "[55] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, 2024. 3, 6", + "[56] Liheng Zhong, Lina Hu, and Hang Zhou. Deep learning based multi-temporal crop classification. Remote Sensing of Environment, 2019. 1", + "[57] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2, 3, 4", + "[58] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 4", + "[59] Hao Zhu, Yan Zhu, Jiayu Xiao, Yike Ma, Yucheng Zhang, Jintao Li, and Feng Dai. Misa: mining saliency-aware semantic prior for box supervised instance segmentation. In IJCAI, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14045", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_model.json b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e9b7c485c4316c11c69e5275074ed4b5c36a473f --- /dev/null +++ b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.13, + 0.852, + 0.177 + ], + "angle": 0, + "content": "Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.203, + 0.867, + 0.239 + ], + "angle": 0, + "content": "Hao Zhu\\(^{1}\\), Yan Zhu\\(^{1}\\), Jiayu Xiao\\(^{1}\\), Tianxiang Xiao\\(^{1}\\), Yike Ma\\(^{1}\\), Yucheng Zhang\\(^{1}\\), Feng Dai\\(^{1\\dagger}\\) \n\\(^{1}\\)Institute of Computing Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.242, + 0.615, + 0.256 + ], + "angle": 0, + "content": "{zhuhao22z，fdai}@ict.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Automated crop mapping through Satellite Image Time Series (SITS) has emerged as a crucial avenue for agricultural monitoring and management. However, due to the low resolution and unclear parcel boundaries, annotating pixel-level masks is exceptionally complex and time-consuming in SITS. This paper embraces the weakly supervised paradigm (i.e., only image-level categories available) to liberate the crop mapping task from the exhaustive annotation burden. The unique characteristics of SITS give rise to several challenges in weakly supervised learning: (1) noise perturbation from spatially neighboring regions, and (2) erroneous semantic bias from anomalous temporal periods. To address the above difficulties, we propose a novel method, termed exploring space-time perceptive clues (Exact). First, we introduce a set of spatial clues to explicitly capture the representative patterns of different crops from the most class-relative regions. Besides, we leverage the temporal-to-class interaction of the model to emphasize the contributions of pivotal clips, thereby enhancing the model perception for crop regions. Building upon the space-time perceptive clues, we derive the clue-based CAMs to effectively supervise the SITS segmentation network. Our method demonstrates impressive performance on various SITS benchmarks. Remarkably, the segmentation network trained on Exact-generated masks achieves \\(95\\%\\) of its fully supervised performance, showing the bright promise of weakly supervised paradigm in crop mapping scenario. Our code will be publicly available here." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.761, + 0.222, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.486, + 0.879 + ], + "angle": 0, + "content": "The launch of numerous public and commercial satellites provides broader opportunities to record, analyze, and predict the evolution of crop land [5, 17, 47, 50]. In this context, Satellite Image Time Series (SITS) with \\(10\\mathrm{m}\\) resolution offered by the high-frequency Sentinel-2 (S2) satellites serve as a valuable data source for automated crop map" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.291, + 0.725, + 0.305 + ], + "angle": 0, + "content": "(a) Spatial Noise Perturbation" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.306, + 0.902, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.488, + 0.779, + 0.501 + ], + "angle": 0, + "content": "(b) Temporal Erroneous Semantic Bias" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.502, + 0.902, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.611, + 0.908, + 0.724 + ], + "angle": 0, + "content": "Figure 1. Illustration of the two inherent issues arises from spatial and temporal perspectives in SITS. (a) shows noise perturbation from the spatial perspective. We visualize the high-level feature manifold of Dog (natural image) and Barley (SITS) to reveal the distinct spatial properties. The feature dimensions are reduced by t-SNE [48]. (b) shows the erroneous semantic bias induced by anomalous temporal clips. We denote the parcel regions with \\(\\star\\). The white circles refer to false positive activation regions in CAMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "ping [6, 14, 22, 28, 35, 56]. The core of crop mapping lies in the semantic segmentation of crop parcels. Recently, many efforts are devoted to exploiting the versatile and powerful relation modeling capability of deep neural networks for this task [20, 21, 34, 46]. While such methods have shown a significant progress, they rely heavily on pixel-level manual annotation, which is notoriously complex and time-consuming [49]. The low resolution of satellite images and the indistinct boundaries between crop parcels complicate the annotation process. Even worse, the varying phenologi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.887, + 0.218, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14036" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "cal cycles among crops require annotators to meticulously select appropriate acquisition times for annotation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.125, + 0.483, + 0.397 + ], + "angle": 0, + "content": "To address this, one promising solution is weakly supervised semantic segmentation (WSSS), which relies solely on time-efficient annotation form, i.e., image-level categories. The image-level WSSS has commonly studied in natural image domain. Existing methods primarily extract Class Activation Map (CAM) [57] to generate pseudo labels for training semantic segmentation models. The intention of CAM is to identify the regions that highly contribute to the prediction of each class, revealing the shared patterns among the images within the same category. In context of natural images, CAM tends to highlight local discriminative parts of the target object. The reason is that the dispersed intraclass distribution enforces the model to learn a extremely sharp decision boundary, thereby the classifier weights tend to interact with feature representations within more discriminative object regions, as shown in Fig. 1a left. Thus the primary focus of researchers falls into expanding the CAM to identify the entire object semantics for natural images." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.4, + 0.485, + 0.672 + ], + "angle": 0, + "content": "Different from WSSS for natural image, crop mapping for SITS mainly faces with two challenges: (1) from the spatial perspective, parcel objects within the same category share uniform appearance and color, exhibiting strong neighboring consistency. The intra-class compactness of local patterns looses the tolerance of the classifier to noise perturbation, resulting in pronounced over-activation phenomena in CAM, as shown in Fig. 1a right. The disparate characteristic lead to the advances in natural image domain cannot directly benefit the SITS crop segmentation. (2) From the temporal perspective, although different crops show distinct phenological cycles and varying characteristics, they may display similar appearance in some specific periods. This confusion imbues wrong semantic bias to the learning process of the model, thereby activating some undesired semantic regions in CAM. An intuitive illustration is shown in Fig. 1b, the temporal clips that deviate from the pivotal semantic affect the perceptual ability of CAM to correct crop regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we present exploring space-time perceptive clues (Exact), a tailored WSSS framework for crop mapping, to cope with the challenges arisen from the spatial and temporal aspects respectively. Firstly, we introduce a set of spatial clues to explicitly capture the patterns of different crops. Leveraging the filtered CAM as an indicator, we update representative clues by conducting spatial clustering from the most class-relative regions. These clues are then used to regularize the feature space via optimizing the contrastive objective, thereby sharpening the decision boundary of the model and mitigating the perturbations from illusory patterns. Secondly, to cope with the erroneous semantic bias caused by anomalous temporal periods, we propose temporal-aware affinity propagation to emphasize the contributions of pivotal clips for crop perception. In detail, we extract the temporal-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "to-class attention from the model to reweight the temporal sequence embeddings. The modulated representations can be used to model temporal-aware pairwise affinity for propagation on the raw CAM, thus effectively suppressing the undesired semantic regions in a self-supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.905, + 0.29 + ], + "angle": 0, + "content": "Unlike existing WSSS methods that rely on classifier weights to generate CAMs, we exploit the well-updated space-time perceptive clues to derive the clue-based CAMs (CB-CAMs) as pseudo labels for segmentation. Compared to the raw CAM, the CB-CAMs (1) remarkably suppress the perturbations from spatial and temporal aspects, and (2) delineate the crop regions more precisely, thereby providing more reliable supervision for the subsequent segmentation." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.291, + 0.895, + 0.305 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.308, + 0.904, + 0.368 + ], + "angle": 0, + "content": "- We introduce the WSSS paradigm to SITS crop mapping task to tackle the daunting annotation challenge. To the best of our knowledge, this is the first work that relies solely on image-level categories for crop segmentation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.371, + 0.905, + 0.446 + ], + "angle": 0, + "content": "- To overcome the drawbacks arising from the spatial and temporal aspects of SITS, we propose Exact that explores space-time perceptive clues to reduce the noise perturbation and rectify the wrong semantic bias, ultimately providing reliable supervision for SITS segmentation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.449, + 0.905, + 0.539 + ], + "angle": 0, + "content": "- We experimentally show that Exact achieves impressive performance on common benchmarks. Using Exact-generated labels for training, SITS segmentation model attains up to \\(95\\%\\) of its fully supervised performance. Our results significantly advance the upper bound of image-level WSSS technique compared to the other domains." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.308, + 0.905, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.655, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Semantic segmentation on SITS. Automated crop monitoring through Satellite Image Time Series (SITS) has attracted great interest among researchers and demonstrated considerable social impact [14, 22, 23, 33, 36, 43, 44]. One of the challenging tasks is the semantic segmentation of agricultural parcels. The goal of the network is to learn a mapping function that assigns each pixel in the SITS to the corresponding crop type or background. Some works process SITS inputs by first extracting spatial information and then compressing the temporal dimension. For example, U-ConvLSTM [32] relied on U-Net [37] architecture to encode the spatial dimension, followed by a ConvLSTM [42] for the temporal dimension. Similarly, the FPN-ConvLSTM [34] replaced the U-Net with a Feature Pyramid Network [29] as the spatial encoder. U-TAE [20] compressed the temporal dimension through the outstanding temporal attention mechanism [21]. The other option to encode SITS is the temporal-spatio scheme, which first processes the temporal dimension and then extracts the spatial information. Rußwurm et al. [40] employed bidirectional LSTM to extract temporal features and then used CNNs to integrate spatial informa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14037" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.09, + 0.308, + 0.105 + ], + "angle": 0, + "content": "(a) Training Pipeline of the Exact" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.9, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.268, + 0.36, + 0.283 + ], + "angle": 0, + "content": "(b) Explore Spatial Perceptive Clues (§3.2)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.283, + 0.5, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.268, + 0.816, + 0.283 + ], + "angle": 0, + "content": "(c) Temporal-Aware Affinity Propagation (§3.3)" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.283, + 0.898, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.433, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 2. (a) The training pipeline of Exact. We adopt the Temporal-Spatio scheme to handle the SITS input, which contains two transformer encoders. The first temporal encoder models interactions between acquisition times, then the followed spatial encoder discards the temporal dimension and models interactions between spatial positions. To overcome the difficulties arises from spatial and temporal aspects, we propose two novel technologies in temporal embedding space: (b) Explore Spatial Perceptive Clues to mitigate the noise perturbation (see §3.2) and (c) Temporal-Aware Affinity Propagation to rectify the wrong semantic bias(see §3.3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.486, + 0.651 + ], + "angle": 0, + "content": "tion. Recently, TSViT [46] proposed to borrow the powerful dependency modeling capability of the Vision Transformer (ViT) [16] to handle SITS, achieving state-of-the-art performance at lower computational cost. The TSViT also comprehensively illustrated the superiority of the temporal-spatio scheme. Taking a holistic view, we adopt this scheme to process SITS, aiming to achieve an optimal trade-off for automated crop mapping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Weakly supervised semantic segmentation. Weakly supervised semantic segmentation (WSSS) with image-level labels (i.e. ground-truth object categories) has shown significant success in natural images [13, 18, 52, 53, 59]. Most advanced WSSS methods follow the three-steps pipeline of: 1) training a classification network with image-level labels, 2) obtaining class activation map (CAM) [9, 41, 57] from the well-optimized classification network as pixel-level coarse labels, 3) replacing ground-truth masks with the pseudo labels to train a off-the-shelf segmentation model. The objective of our work focuses on the first and second steps, i.e., how to generate accurate pseudo labels. Numerous works has been proposed to address the under- and over-activation issues of the initial CAM [3, 10, 11, 26, 39, 55] in natural images. To derive the final pseudo labels, CAM requires cumbersome post-processing, including random walks [1, 2] and dense" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.909, + 0.651 + ], + "angle": 0, + "content": "CRF [25] refinement. Thanks to these excellent works, the performance of image-level WSSS in natural images currently achieves \\(90\\%\\) of pixel-level supervision. However, the WSSS networks designed for natural images require significant adaptation to be applied to SITS, and the results are still unsatisfactory due to the distinct data characteristics. In this work, we incorporate WSSS technique into the SITS and overcome the difficulties caused by inherent data properties." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.606, + 0.679 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.795 + ], + "angle": 0, + "content": "In this section, we first look more closely at the temporal-spatio scheme and class activation map technique (§3.1). Next, we introduce how to explore spatial perceptive clues (§3.2) and conduct temporal-aware affinity propagation (§3.3). Finally, we show the overall objective function and the clue-based CAM generation strategy of Exact (§3.4). The training pipeline of Exact is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.655, + 0.818 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.908, + 0.901 + ], + "angle": 0, + "content": "The temporal-spatio scheme that first processes the temporal dimension and then extracting the spatial information has shown significant superiority when dealing with SITS data. Following the TSViT [46], we use the variant ViT [16] as the backbone for both temporal and spatial encoders. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14038" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.214 + ], + "angle": 0, + "content": "consider a SITS input \\(\\mathbf{X} \\in \\mathbb{R}^{T \\times C \\times H \\times W}\\) with \\(T\\) the length of time series, \\(C\\) the number of channels and \\(H \\times W\\) the spatial dimensions. The input \\(\\mathbf{X}\\) is mapped into a sequence of patch tokens and reshaped to \\(\\mathbf{Z} \\in \\mathbb{R}^{N_h \\cdot N_w \\times T \\times d}\\), where \\(N_h = \\left[\\frac{H}{h}\\right]\\), \\(N_w = \\left[\\frac{W}{w}\\right]\\) and \\(h \\times w\\) is the spatial extent for each patch. We then add the temporal position embeddings \\(\\mathbf{P}_T \\in \\mathbb{R}^{T \\times d}\\) and concatenate the temporal multi-class tokens \\(\\mathbf{Z}_T^{\\mathrm{cls}} \\in \\mathbb{R}^{K \\times d}\\) to obtain the input of temporal encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.238, + 0.484, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {T} ^ {\\mathrm {i n}} = \\left[ \\mathbf {Z} _ {T} ^ {\\mathrm {c l s}}, \\mathbf {Z} + \\mathbf {P} _ {T} \\right], \\mathbf {Z} _ {T} ^ {\\mathrm {i n}} \\in \\mathbb {R} ^ {N _ {h} \\cdot N _ {w} \\times (K + T) \\times d} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.485, + 0.405 + ], + "angle": 0, + "content": "here \\(K\\) denotes the number of categories, \\(\\mathbf{Z}_T^{\\mathrm{cls}}\\) and \\(\\mathbf{P}_T\\) are repeated \\(N_{h}\\cdot N_{w}\\) times to match the spatial shape. With the output feature maps \\(\\mathbf{Z}_T^{\\mathrm{out}} = [\\mathbf{Z}_T^{\\mathrm{dense}}|\\mathbf{Z}_T^{\\mathrm{seq}}]\\) from the temporal encoder, we extract the first \\(K\\) tokens \\(\\mathbf{Z}_T^{\\mathrm{dense}}\\in \\mathbb{R}^{N_h\\cdot N_w\\times K\\times d}\\) and permute the first and second dimensions to serve as input patch tokens \\(\\mathbf{Z}_S\\in \\mathbb{R}^{K\\times N_h\\cdot N_w\\times d}\\) for spatial encoder. These inputs are then combined with spatial multi-class tokens \\(\\mathbf{Z}_{S}^{\\mathrm{cls}}\\in \\mathbb{R}^{K\\times 1\\times d}\\) and spatial position embeddings \\(\\mathbf{P}_S\\in \\mathbb{R}^{N_h\\cdot N_w\\times d}\\) at all \\(K\\) spatial representations:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.414, + 0.484, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {S} ^ {\\mathrm {i n}} = \\left[ \\mathbf {Z} _ {S} ^ {\\mathrm {c l s}}, \\mathbf {Z} _ {S} + \\mathbf {P} _ {S} \\right], \\mathbf {Z} _ {S} ^ {\\mathrm {i n}} \\in \\mathbb {R} ^ {K \\times (1 + N _ {h} \\cdot N _ {w}) \\times d} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.548 + ], + "angle": 0, + "content": "After the spatial encoder phase, we separate the output features \\(\\mathbf{Z}_S^{\\mathrm{out}} = [\\mathbf{Z}_S^{\\mathrm{global}}|\\mathbf{Z}_S^{\\mathrm{dense}}]\\) to align with different downstream tasks. For classification task, we feed the global tokens \\(\\mathbf{Z}_S^{\\mathrm{global}}\\in \\mathbb{R}^{K\\times 1\\times d}\\) into the classifier to obtain classification logits. For segmentation task, to derive the dense prediction mask, the tokens \\(\\mathbf{Z}_S^{\\mathrm{dense}}\\in \\mathbb{R}^{K\\times N_h\\cdot N_w\\times d}\\) are fed into the segmentation decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Class activation map (CAM) [57] is widely used in WSSS to provide weak annotations that rely on image-level labels. Given a natural image, its feature maps \\(\\mathbf{F} \\in \\mathbb{R}^{H' \\times W' \\times D}\\) are extracted by a classification backbone. To derive the classification score, the feature maps are average pooled and multiplied by the classifier weights \\(\\boldsymbol{w} \\in \\mathbb{R}^{K \\times D}\\). CAM is generated by weighting and summing each channel in the feature maps with the classifier weights, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.681, + 0.484, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} ^ {k} = \\operatorname {R e L U} \\left(\\sum_ {i} \\boldsymbol {w} _ {i} ^ {k} \\cdot \\mathbf {F} _ {:,: i}\\right), \\quad \\forall k \\in K. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Following most WSSS methods, we normalize \\(\\mathcal{M}^k\\) to the range [0,1] and apply a global threshold to filter out background pixels to obtain final pseudo labels. In the temporal-spatio network, we feed the dense tokens \\(\\mathbf{Z}_T^{\\mathrm{dense}}\\) and \\(\\mathbf{Z}_S^{\\mathrm{dense}}\\) into the classifier and compute the additional classification loss \\(\\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{aux}}\\) to supervise the fused raw CAM for SITS. We provide more details in supplement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.379, + 0.866 + ], + "angle": 0, + "content": "3.2. Explore Spatial Perceptive Clues" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "CAM filtering. Given an input SITS \\(\\mathbf{X}\\) and its image-level label \\(y\\in [0,1]^K\\) , we first compute its normalized fused" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.906, + 0.152 + ], + "angle": 0, + "content": "CAM \\(\\mathcal{M} \\in \\mathbb{R}^{N_h \\cdot N_w \\times K}\\) by the classifier weights and the output dense tokens of temporal and spatial encoder. We use two threshold scores \\(\\mu_l\\) and \\(\\mu_h\\) to filter out the reliable foreground, background and uncertain regions:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.16, + 0.907, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {M}} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\mathcal {M} \\leq \\mu_ {l}, \\\\ 1, & \\text {i f} \\mathcal {M} \\geq \\mu_ {h}, \\\\ \\text {i g n o r e ,} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.907, + 0.453 + ], + "angle": 0, + "content": "Clues clustering. To capture the compact patterns of different crops, we establish a group of class-wise representative prototypes, which later serve as perceptive clues to generate high-quality pseudo labels. In crop segmentation task, temporal features often provide more information than spatial context [46], so we choose to perform spatial clustering on temporal dense embeddings \\(\\mathbf{Z}_T^{\\mathrm{dense}}\\) over the whole dataset. Specifically, we build the class-wise positive prototype set \\(\\mathcal{P}_{\\mathrm{pos}} = \\{\\pmb{p}_{\\mathrm{pos}}^k \\in \\mathbb{R}^{N_p \\times d}\\}_{k=1}^K\\) and negative prototype set \\(\\mathcal{P}_{\\mathrm{neg}} = \\{\\pmb{p}_{\\mathrm{neg}}^k \\in \\mathbb{R}^{N_p \\times d}\\}_{k=1}^K\\), where \\(N_p\\) denotes the number of prototypes. If \\(k\\)-th class appears in a training batch, we update the prototypes \\(\\pmb{p}^k\\) via solving the optimal transport problem [8, 58]. Given a mapping matrix \\(\\mathbf{C}^k\\) that represents the assignment between each pixel and its prototype, it can be referred as an element of the transportation polytope [4]:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.459, + 0.907, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\mathbb {C} = \\left\\{\\mathbf {C} ^ {k} \\in \\mathbb {R} _ {+} ^ {N _ {p} \\times N _ {k}} \\mid \\mathbf {C} ^ {k} \\mathbf {1} = \\boldsymbol {u}, \\mathbf {C} ^ {k \\top} \\mathbf {1} = \\boldsymbol {r} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.563 + ], + "angle": 0, + "content": "where \\(N_{k}\\) denotes the number of pixels belonging to class \\(k\\). 1 represents the vectors of ones in appropriate dimensions, \\(\\mathbf{u}\\) and \\(\\pmb{r}\\) are the marginal projections onto the rows and columns of \\(\\mathbf{C}^k\\). We can maximize the following objective function to optimize the mapping matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.57, + 0.907, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathbf {C} ^ {k} \\in \\mathbb {C}} {\\text {m a x i m i z e}} \\operatorname {T r} \\left(\\mathbf {C} ^ {k \\top} \\boldsymbol {p} ^ {k} \\mathbf {Z} ^ {k \\top}\\right) - \\eta \\sum_ {n _ {p}, n _ {k} \\in N _ {p}, N _ {k}} \\mathbf {C} _ {n _ {p} n _ {k}} ^ {k} \\log \\mathbf {C} _ {n _ {p} n _ {k}} ^ {k}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.716 + ], + "angle": 0, + "content": "here \\(\\mathbf{Z}^k\\in \\mathbb{R}^{N_k\\times d}\\) is the corresponding temporal dense embeddings belonging to class \\(k\\), \\(\\eta\\) controls the smoothness of entropy regularization term. The continuous approximate solution of Eq. (6) can be obtained through iteratively applying the Sinkhorn-Knopp algorithm [15]. Subsequently, we momentum update the \\(n_p\\)-th prototype of class \\(k\\) according to the assignment matrix and the embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.723, + 0.907, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {n _ {p}} ^ {k} = \\alpha \\boldsymbol {p} _ {n _ {p}} ^ {k} + \\frac {1 - \\alpha}{\\left| \\left| \\mathbf {C} _ {n _ {p}} ^ {k} \\right| \\right| _ {1}} \\cdot \\left(\\mathbf {C} _ {n _ {p}} ^ {k} \\mathbf {Z} ^ {k}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.84 + ], + "angle": 0, + "content": "here \\(\\alpha \\in [0,1]\\) is the momentum coefficient. We leverage the fused CAM \\(\\hat{\\mathcal{M}}\\) as the pseudo labels to update the class-wise positive prototype set \\(\\mathcal{P}_{\\mathrm{pos}}\\) and \\(1 - \\hat{\\mathcal{M}}\\) for the negative set \\(\\mathcal{P}_{\\mathrm{neg}}\\). Notably, each prototype is not involved in gradient backpropagation to avoid the noise from the classifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Clue-based contrastive learning. Based on the spatial perceptive clues, we introduce the clue-based contrastive learning [7, 12] to regularize the embedding space. More specifically, for each temporal dense embedding \\( \\mathbf{z}_{n_k}^k \\) and its most" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14039" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.095, + 0.473, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.214, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Figure 3. Visualization of the clue-based CAM generation. This process is performed after training the classification network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.485, + 0.3 + ], + "angle": 0, + "content": "relative prototype \\(\\pmb{p}_{n_p}^k\\), we adopt the cosine distance to measure their similarity [24]:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.309, + 0.484, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p} _ {n _ {p}} ^ {k}\\right) = \\frac {\\boldsymbol {z} _ {n _ {k}} ^ {k} \\boldsymbol {p} _ {n _ {p}} ^ {k \\top}}{\\left\\| \\boldsymbol {z} _ {n _ {k}} ^ {k} \\right\\| \\cdot \\left\\| \\boldsymbol {p} _ {n _ {p}} ^ {k} \\right\\| / \\tau}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.406 + ], + "angle": 0, + "content": "where \\(\\tau\\) indicates the temperature parameter. Subsequently, we enforce the pixel embedding to closely match its prototype and be distinct from other prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.415, + 0.484, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c b l}} = \\sum_ {k} \\sum_ {n _ {k}} \\mathbb {1} (\\log (\\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {-}} \\exp \\mathrm {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p}\\right)) - \\mathrm {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p} _ {n _ {p}} ^ {k}\\right)), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.554 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}^{-} = \\{\\mathcal{P}_{\\mathrm{pos}} \\cup \\mathcal{P}_{\\mathrm{neg}} \\setminus \\boldsymbol{p}_{n_p}^k\\}\\) and \\(\\mathbb{1}(\\cdot)\\) is an indicator function, being 1 if class \\(k\\) appears in image-level labels and 0 otherwise. Minimizing the above objective can pull pixel embedding closely to the semantic center and push it away from other illusory patterns, thereby sharpening the model decision boundary and mitigating the noise perturbation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.421, + 0.581 + ], + "angle": 0, + "content": "3.3. Temporal-Aware Affinity Propagation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.484, + 0.631 + ], + "angle": 0, + "content": "In this section, we propose the temporal-aware affinity propagation to cope with the wrong semantic bias arising from anomalous temporal clips." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.707 + ], + "angle": 0, + "content": "Temporal-aware affinity mining. In the temporal encoder, the input tokens are normalized and projected into query matrix \\(\\mathbf{Q} \\in \\mathbb{R}^{(K + T) \\times d}\\) and key matrix \\(\\mathbf{K} \\in \\mathbb{R}^{(K + T) \\times d}\\) respectively. Then the self-attention \\(\\mathcal{A} \\in \\mathbb{R}^{(K + T) \\times (K + T)}\\) with respect to \\(N_h \\cdot N_w\\) sequences is computed as below:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.716, + 0.484, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d}}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Without additional computations nor supervision, we can explicitly extract the temporal-to-class attention \\(\\tilde{\\mathcal{A}}\\in \\mathbb{R}^{T\\times K}\\) from the self-attention \\(\\mathcal{A}\\), which represents the contributions of high-level representations to crop recognition at different temporal clips. We then reweight the temporal sequence embeddings \\(\\mathbf{Z}_T^{\\mathrm{seq}}\\) based on the normalized attention \\(\\tilde{\\mathcal{A}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.863, + 0.484, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} ^ {k} = \\sum_ {t = 1} ^ {T} \\tilde {\\mathcal {A}} _ {t} ^ {k} \\cdot \\left(\\mathbf {Z} _ {T} ^ {\\text {s e q}}\\right) _ {:, t,:}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "Intuitively, the modulated high-level representation \\(\\mathcal{V}^k\\in\\) \\(\\mathbb{R}^{N_h\\cdot N_w\\times d}\\) enlarges the variation among different semantic crops, thereby modeling more precise pixel relations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.213 + ], + "angle": 0, + "content": "Affinity propagation and guidance. After obtaining the temporal-aware representation \\(\\mathcal{V}\\), we conduct affinity propagation to suppress erroneous semantic regions on the raw CAM. Particularly, the temporal-aware pairwise affinity of the pixel at position \\(i\\) can be estimated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.22, + 0.907, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A f f} _ {j \\in \\mathcal {N} _ {i}} \\left(\\boldsymbol {v} _ {i} ^ {k}, \\boldsymbol {v} _ {j} ^ {k}\\right) = \\exp \\left(\\frac {\\mathcal {S} \\left(\\boldsymbol {v} _ {i} ^ {k} , \\boldsymbol {v} _ {j} ^ {k}\\right)}{\\sigma_ {i} ^ {k}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.332 + ], + "angle": 0, + "content": "here \\(\\sigma_{i}^{k}\\) is the standard deviation of \\(\\pmb{v}_{i}^{k}\\) and \\(\\mathcal{N}_i\\) indicates the local receptive fields (e.g., 8-way local neighbors [27]). We denoise the raw CAM via iteratively propagating the pairwise affinity from the temporal-aware representation \\(\\nu\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.338, + 0.906, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {M}} _ {i} ^ {k} = \\delta_ {i} ^ {k} \\sum_ {j \\in \\mathcal {N} _ {i}} \\operatorname {A f f} \\left(\\boldsymbol {v} _ {i} ^ {k}, \\boldsymbol {v} _ {j} ^ {k}\\right) \\cdot \\mathcal {M} _ {j} ^ {k}, \\quad \\text {w h e r e} \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.372, + 0.756, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\delta_ {i} ^ {k} = \\frac {1}{\\sum_ {j \\in \\mathcal {N} _ {i}} \\mathsf {A f f} (\\boldsymbol {v} _ {i} ^ {k} , \\boldsymbol {v} _ {j} ^ {k})}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.906, + 0.44 + ], + "angle": 0, + "content": "After obtaining the improved activation map, we align the raw CAM with it to effectively guide the learning process:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.448, + 0.907, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t a p}} = \\sum_ {k} \\mathbb {1} \\left(\\left| \\tilde {\\mathcal {M}} ^ {k} - \\mathcal {M} ^ {k} \\right|\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.489, + 0.907, + 0.534 + ], + "angle": 0, + "content": "By minimizing \\(\\mathcal{L}_{\\mathrm{tap}}\\), we can rectify erroneous activations in the raw CAM and incorporate temporal-aware priors into the embedding space, ultimately benefiting the perceptive clues." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.906, + 0.573 + ], + "angle": 0, + "content": "3.4. Network Optimization and Clue-based CAM Generation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.579, + 0.906, + 0.61 + ], + "angle": 0, + "content": "As shown in Fig. 2, the whole objective function for training the classification network consists of four components:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.619, + 0.907, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {c l s}} ^ {\\text {a u x}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {c b l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {t a p}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.674 + ], + "angle": 0, + "content": "here \\(\\mathcal{L}_{\\mathrm{cls}}\\) and \\(\\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{aux}}\\) are the conventional binary cross entropy loss and \\(\\lambda_{i}\\) denotes the weight to rescale the loss terms." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.907, + 0.781 + ], + "angle": 0, + "content": "After training process, we perform per-pixel perception based on the well-updated space-time clues in temporal dense embedding space to generate clue-based CAMs (CB-CAMs). For each embedding \\( z_{i} \\) in temporal dense embeddings \\( \\mathbf{Z}_T^{\\mathrm{dense}} \\), we measure the maximum similarity with the positive prototypes and minus the misguide activations with the negative prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.788, + 0.907, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Y} _ {i} ^ {k} = \\operatorname {R e L u} \\left(\\max _ {\\boldsymbol {p} ^ {+} \\in \\boldsymbol {p} _ {k} ^ {\\text {p o s}}} \\mathrm {S} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {p} ^ {+}\\right) - \\max _ {\\boldsymbol {p} ^ {-} \\in \\boldsymbol {p} _ {k} ^ {\\text {n e g}}} \\mathrm {S} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {p} ^ {-}\\right)\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "If the class \\( k \\) does not appear in a training image, we set \\( \\mathcal{Y}^k \\) to all zeros. Fig. 3 gives a visualization of this generation process. Finally, the CB-CAMs \\( \\mathcal{Y} \\) are filtered using a global background score to generate the pseudo labels, which are then used to train the SITS semantic segmentation network." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14040" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.077, + 0.546, + 0.3 + ], + "angle": 0, + "content": "
MethodTypeGermany [40]PASTIS [20]
OAmIoUOAmIoU
MCTFormercVPR'22 [52]ViT-PCMECCV'22 [38]LP-CAMCVPR'23 [13]TSCDAAI'23 [54]DuPLCVPR'24 [51]SeCOCVPR'24 [55]RGB70.656.366.749.6
66.852.269.353.2
64.749.167.050.1
68.452.767.251.3
66.250.665.548.7
64.548.363.646.1
baseline+PAMRCVPR'20 [3]+TS-CAMICCV'21 [19]+SIPECVPR'22 [11]+FPRICCV'23 [10]SITS82.574.181.269.5
83.976.382.071.2
81.170.580.467.6
82.073.281.570.1
82.475.681.771.0
+ours-Exact88.3 ↑ 5.880.6 ↑ 6.584.1 ↑ 2.975.6 ↑ 6.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.092, + 0.302, + 0.546, + 0.315 + ], + "angle": 0, + "content": "(a) Pseudo labels performance of different methods on Germany and PASTIS train set." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.078, + 0.905, + 0.3 + ], + "angle": 0, + "content": "
MethodSup.OAmIoUratio
ConvLSTM [42]78.250.1
BiCGRU [40]80.556.2
FPN-ConvLSTM [34]81.959.5
Unet-3D [32]G82.360.4-
Unet-3Df [45]82.160.2
U-TAE [20]82.962.4
TSViT [46]83.465.1100%
baseline77.157.788%
+PAMRCVPR'20 [3]78.558.790%
+TS-CAMICCV'21 [19]P76.856.587%
+SIPECVPR'22 [11]77.158.089%
+FPRICCV'23 [10]78.258.490%
+ours-Exact80.3 ↑ 3.261.8 ↑ 4.195%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.564, + 0.303, + 0.9, + 0.327 + ], + "angle": 0, + "content": "(b) Segmentation performance trained with pseudo labels generated by different methods on PASTIS test set." + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.339, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Table 1. Comparisons with existing WSSS methods in OA(%) and mIoU(%) . RGB: the networks designed for processing natural images. SITs: the networks designed for processing SITS (include four adapted modules from natural image domain). \\(\\mathcal{G}\\) : various SITS segmentation networks supervised by ground-truth labels. \\(\\mathcal{P}\\) : the TSViT segmentation model supervised by different pseudo labels from the methods in SITs. The ratio refers to the proportion of mIoU between fully supervised and weakly supervised of TSViT segmentation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.415, + 0.224, + 0.433 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.44, + 0.282, + 0.457 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.486, + 0.717 + ], + "angle": 0, + "content": "Datasets & evaluation metric. We conducted comprehensive experiments on two widely used Satellite-2 time series crop recognition datasets to validate the effectiveness of our approach quantitatively. The PASTIS dataset [20] contains 2433 multi-spectral time series of size \\(128 \\times 128\\), each series including 10 bands and 33 to 61 temporal observations. It consists of 18 crop types and a background category. Following the settings of [46], we used the fold-1 among the five folds provided in PASTIS. We partitioned each sample into multiple patches and assigned category labels according to the mask annotations (see the supplement for details). The Germany dataset [40] comprises \\(137k\\) field parcels of time series imagery with 13 spectral bands. Each series contains 36 observations and is labeled with 17 crop types. We employed the mean Intersection over Union (mIoU) and pixel-wise overall accuracy (OA) as evaluation metrics, both widely used to measure segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Implementation details. The classification network was trained on 8 NVIDIA RTX 3090 GPUs with batch size 8 for 15k iterations. During the training stage, we used the AdamW optimizer [30] with an initial learning rate of 1e-3 and cosine weight decay [31]. As the raw CAM were unreliable in early iterations, the gradients of \\(\\mathcal{L}_{\\mathrm{cbl}}\\) were backpropagated after the 4k iteration. For the raw CAM filtering, the threshold score \\((\\mu_l,\\mu_h)\\) was set to (0.2, 0.4). Due to the compact intra-class patterns, we set the number of class-specific prototypes \\(N_{p}\\) to 2, and the momentum used to update the prototypes was set to 0.999. Moreover, the other hyper-parameters \\(\\eta\\), \\(\\lambda_1\\), \\(\\lambda_2\\) and \\(\\tau\\) were empirically set to 0.05, 0.01, 0.015 and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.909, + 0.493 + ], + "angle": 0, + "content": "0.1, respectively. We employed the original TSViT with a segmentation decoder as the semantic segmentation model for SITS. The training details of the segmentation model exactly followed the settings in [46] without any modifications. Please refer to supplementary materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.505, + 0.719, + 0.521 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Base Architectures. In the evaluation stage, we fuse the raw CAMs generated by the spatial dense embeddings \\(\\mathbf{Z}_S^{\\mathrm{dense}}\\) and temporal dense embeddings \\(\\mathbf{Z}_T^{\\mathrm{dense}}\\) as our baseline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.907, + 0.593 + ], + "angle": 0, + "content": "4.2.1. Comparison with other WSSS Methods on CAMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.658 + ], + "angle": 0, + "content": "The key of the weakly supervised learning is to provide reliable training pseudo labels for segmentation network. Thus, we conducted comparative experiments on the train splits to validate the quality of pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.911, + 0.854 + ], + "angle": 0, + "content": "Competing methods adaptation. We reimplemented six well-performing transformer-based WSSS methods designed for natural images as competing methods. To accommodate the temporal dimension, we transformed the input dimensions of SITS into a 3D format \\((T \\times C, H, W)\\) for these methods. In addition, to enable a fairer comparison between Exact and existing WSSS works, we carefully selected four off-the-shelf modules from other WSSS methods in the natural image domain and adapted them to the temporal-spatio framework. These modules were also reinplemented within the temporal dense embedding space to align with our approach. We attempted various adaptation strategies, more details and analysis available in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. Tab. 1a presents the mIoU and OA of the pseudo labels generated by other WSSS methods and our Exact. As can be seen, the WSSS methods designed for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14041" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.189 + ], + "angle": 0, + "content": "
Method\\( \\mathcal{L}_{\\mathrm{cbl}} \\)\\( \\mathcal{L}_{\\mathrm{tap}} \\)CB-CAMRecallOAmIoU
baseline81.181.269.5
✓81.882.472.3
✓✓82.583.073.4
ours-Exact✓✓✓84.884.175.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.2, + 0.483, + 0.228 + ], + "angle": 0, + "content": "Table 2. Ablation study results of different components. The baseline is a setting with only image-level classification loss." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.242, + 0.48, + 0.326 + ], + "angle": 0, + "content": "
MethodSpa.Tem.RecallPrecisionOAmIoU
✓81.978.480.566.7
✓80.881.681.068.3
baseline CAM✓✓81.183.581.269.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.336, + 0.485, + 0.365 + ], + "angle": 0, + "content": "Table 3. Baseline CAMs from different embedding spaces. Overall, the fused CAM shows the best performance." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.378, + 0.481, + 0.478 + ], + "angle": 0, + "content": "
MethodLow.Ta.RecallPrecisionOAmIoU
baseline81.183.581.269.5
✓80.984.081.169.9
ours-Exact✓81.885.682.472.3
✓✓81.785.282.071.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.531 + ], + "angle": 0, + "content": "Table 4. Impact of modeling affinity from different sources. Low and \\( Ta \\) refer to low-level cues and temporal-aware affinity. The temporal-aware affinity remarkably reduces noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.693 + ], + "angle": 0, + "content": "natural images only achieve a maximum \\(53.2\\%\\) mIoU (refer to PASTIS) with the limited encoding capability for SITS. Furthermore, due to the distinct data characteristics, the off-the-shelf modules struggle to perform well on SITS, and may even disturb the learning process. These phenomena indicate that the advancements in natural image cannot be directly translated to benefit the SITS domain. By contrast, our Exact overcome the intrinsic challenges from SITS, delivering the best performance compared to others." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.458, + 0.717 + ], + "angle": 0, + "content": "4.2.2. Segmentation results trained by pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.796 + ], + "angle": 0, + "content": "After obtaining the pseudo labels, we use them as replacement for ground truth to train the SITS semantic segmentation network. We employ TSViT followed by a segmentation head as our segmentation model, which is the top-performing approach in fully supervised oracles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. In Tab. 1b, we show that the semantic segmentation model can achieve \\(95\\%\\) to its fully supervised performance (refer to mIoU) using pseudo labels generated by Exact. Our best results are \\(80.3\\%\\) OA and \\(61.8\\%\\) mIoU on PASTIS test set, which surpasses the baseline by \\(3.2\\%\\) and \\(4.1\\%\\), respectively. The experiment results quantitatively imply that the labels quality outperforms other methods by a" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.71, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.225, + 0.645, + 0.236 + ], + "angle": 0, + "content": "(a) baseline" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.089, + 0.905, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.225, + 0.854, + 0.236 + ], + "angle": 0, + "content": "(b)ours-Exact" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.247, + 0.906, + 0.276 + ], + "angle": 0, + "content": "Figure 4. Visualization of temporal feature spaces on PASTIS train set. The feature dimensions are reduced by t-SNE [48]." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.291, + 0.707, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.397, + 0.648, + 0.411 + ], + "angle": 0, + "content": "(a) \\(N_{p}\\)" + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.291, + 0.904, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.397, + 0.835, + 0.409 + ], + "angle": 0, + "content": "(b) \\(\\pmb{\\tau}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.422, + 0.908, + 0.452 + ], + "angle": 0, + "content": "Figure 5. Effect of the hyper-parameters. (a) the number of class-specific prototypes \\( N_{p} \\). (b) the temperature of similarity \\( \\tau \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.906, + 0.528 + ], + "angle": 0, + "content": "large margin. In contrast, WSSS networks in natural image have undergone a series of evolutions to barely achieve \\(90\\%\\) of fully supervised performance. This suggests that SITS WSSS techniques offers greater potential in SITS scenario." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.677, + 0.548 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.908, + 0.613 + ], + "angle": 0, + "content": "In this section, our primary purpose is to demonstrate the collective effectiveness of all components within our approach. We also choose the fused raw CAM as the baseline, and the CAM comparison results reported on PASTIS train set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.869 + ], + "angle": 0, + "content": "All the components matter. Our Exact consists of several components, including spatial perceptive clues exploration, temporal-aware affinity propagation and clue-based CAM generation strategy. We validate the contribution of each module, the results are presented in Tab. 2. We can see that the clue-based contrastive learning \\(\\mathcal{L}_{\\mathrm{cbl}}\\) and the temporal-aware affinity propagation \\(\\mathcal{L}_{\\mathrm{tap}}\\) improves the performance by \\(1.8\\%\\) OA and \\(3.9\\%\\) mIoU, respectively. This suggests that these two modules complement each other, regularizing the entire embedding space and ultimately sharpening the decision boundary. We provide an intuitive comparison in Fig. 4, after introducing both objectives, the intra-class features become more compact while the inter-class features are more separated. Besides, using the well-updated space-time clues to generate CAM (i.e., CB-CAM) within the temporal dense embedding space brings a further improvement of \\(2.3\\%\\) recall, \\(1.1\\%\\) OA and \\(2.2\\%\\) mIoU." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Baseline CAM from different embedding spaces. In Tab. 3, we investigate the performance of the raw CAMs derived" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14042" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.907, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.908, + 0.444 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons. Left: CAMs comparisons on PASTIS train set. The stars represent the corresponding activation category. Right: Outputs of segmentation models trained with different CAMs on the PASTIS test set." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.469, + 0.483, + 0.606 + ], + "angle": 0, + "content": "from different dense embedding spaces. The raw CAM obtained from the spatial dense embeddings \\(\\mathbf{Z}_S^{\\mathrm{dense}}\\) falls significantly behind that obtained from the temporal dense embeddings \\(\\mathbf{Z}_T^{\\mathrm{dense}}\\) in terms of precision, OA and mIoU. It demonstrates that the temporal correlations contain more critical information compared to the spatial context in this task. We combine the CAMs from both embedding spaces as our baseline (the last row in Tab. 3) and utilize the resulting pseudo labels to guide spatial clustering." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.61, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Effect of the temporal-aware affinity. Mining low-level cues affinity (e.g., color and intensity) as additional guidance is prevalent in natural images. We also attempted to incorporate the low-level cues to model pairwise affinity in SITS, the results can be found in Tab. 4. Compared to our proposed temporal-aware affinity, modeling low-level cues offers only limited improvement. This is due to the presence of substantial noise interference within the low-level cues of SITS (e.g., cloud cover and shadow). In contrast, our approach highlights the contributions of pivotal temporal clips within a high-level embedding space, effectively mitigating the wrong semantic bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Effect of the hyper-parameters. Fig. 5 shows the effects of the prototype number \\( N_{p} \\) and the temperature \\( \\tau \\). Notably, utilizing only a minimal number of prototypes (i.e., \\( N_{p} = 2 \\)) yields optimal results, whereas increasing the number of prototypes beyond this point brings negative impact on the model. This indicates that the excessive number may enforce prototypes to focus on local discriminative patterns" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.515 + ], + "angle": 0, + "content": "within the class, leading to under-activation issue (the major difficulty in natural image). More analysis of parameters is specifically discussed in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.696, + 0.535 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.69 + ], + "angle": 0, + "content": "Fig. 6 presents the visual comparison between the baseline TSViT-CAMs and the proposed CB-CAMs generated by Exact on PASTIS dataset. The first five columns show the visualization of the CAMs. As we can see, our method remarkably suppress the erroneous regions, whose shape is closer to ground truth masks than baseline. Our high-quality pseudo labels subsequently enhance the segmentation performance compared to the baseline, as shown in last three columns. However, Exact still exhibits limitations in edge processing, which will be the focus in our future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.809, + 0.716 + ], + "angle": 0, + "content": "5. Conclusion and Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a tailored WSSS approach \\( \\text{Exact} \\) to alleviate the daunting annotation challenge in SITS crop mapping task. \\( \\text{Exact} \\) explores space-time perceptive clues to capture the essential patterns of different crop types, thereby overcoming the issues of spatial noise perturbation and wrong temporal semantic bias. Extensive experiments show the superiority of \\( \\text{Exact} \\)-generated masks both quantitatively and qualitatively. We believe that our method marks a pioneering step in effectively applying WSSS technologies to SITS. If follow-up work can find and resolve the limitations under our framework, there is great potential that only image-level labels are needed for crop mapping in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14043" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.161 + ], + "angle": 0, + "content": "This work is supported by National Key R&D Program of China (2022YFD2001601) and National Natural Science Foundation of China (62372433)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.176, + 0.188, + 0.192 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.244 + ], + "angle": 0, + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.245, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[2] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[3] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[4] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[5] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[6] Vitus Benson, Claire Robin, Christian Requena-Mesa, Lazaro Alonso, Nuno Carvalhais, José Cortés, Zhihan Gao, Nora Linscheid, Mélanie Weynants, and Markus Reichstein. Multimodal learning for geospatial vegetation forecasting. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[9] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In WACV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[10] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[11] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[13] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.201, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "[15] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.908, + 0.189 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.908, + 0.272 + ], + "angle": 0, + "content": "[17] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: Esa's optical high-resolution mission for gmes operational services. Remote Sensing of Environment, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[18] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[19] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[20] Vivien Sainte Fare Garnot and Loic Landrieu. Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In ICCV, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.467 + ], + "angle": 0, + "content": "[21] Vivien Sainte Fare Garnot, Loic Landrieu, Sebastien Giordano, and Nesrine Chehata. Satellite image time series classification with pixel-set encoders and temporal self-attention. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.907, + 0.537 + ], + "angle": 0, + "content": "[22] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[23] Pengyu Hao, Yulin Zhan, Li Wang, Zheng Niu, and Muhammad Shakir. Feature selection of time series modis data for early crop classification using random forest: A case study in kansas, usa. Remote Sensing, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.907, + 0.635 + ], + "angle": 0, + "content": "[24] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.906, + 0.676 + ], + "angle": 0, + "content": "[25] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[26] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[27] Wentong Li, Yuqian Yuan, Song Wang, Wenyu Liu, Dongqi Tang, Jianke Zhu, Lei Zhang, et al. Label-efficient segmentation via affinity propagation. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[28] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14044" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[32] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In CVPR Workshops, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[34] Jorge Andres Chamorro Martinez, Laura Elena Cué La Rosa, Raul Queiroz Feitosa, Ieda Del'Arco Sanches, and Patrick Nigri Happ. Fully convolutional recurrent networks for multideate crop recognition from multitemporal image sequences. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[35] Sina Mohammadi, Mariana Belgiu, and Alfred Stein. Improvement in crop mapping from satellite image time series by effectively supervising deep neural networks. ISPRS Journal of Photogrammetry and Remote Sensing, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[36] Charlotte Pelletier, Silvia Valero, Jordi Inglada, Nicolas Champion, and Gérard Dedieu. Assessing the robustness of random forests to map land cover with high resolution satellite image time series over large areas. Remote Sensing of Environment, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.439 + ], + "angle": 0, + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[38] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.484, + 0.536 + ], + "angle": 0, + "content": "[39] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[40] Marc Rußwurm and Marco Körner. Multi-temporal land cover classification with sequential recurrent encoders. ISPRS International Journal of Geo-Information, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[41] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[42] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[43] Sofia Siachalou, Giorgos Mallinis, and Maria Tsakiri-Strati. A hidden markov models approach for crop classification: Linking crop phenology to time series of multi-sensor remote sensing data. Remote Sensing, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[44] Haopeng Sun, Yingwei Zhang, Lumin Xu, Sheng Jin, and Yiqiang Chen. Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer. arXiv preprint arXiv:2412.10181, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[45] Michail Tarasiou, Riza Alp Güler, and Stefanos Zafeiriou. Context-self contrastive pretraining for crop type semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[46] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In CVPR, 2023. 1, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[47] Ramon Torres, Paul Snoeij, Dirk Geudtner, David Bibby, Malcolm Davidson, Evert Attema, Pierre Potin, Björn Rommen, Nicolas Floury, Mike Brown, et al. Gmes sentinel-1 mission. Remote Sensing of Environment, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.246 + ], + "angle": 0, + "content": "[49] Sherrie Wang, William Chen, Sang Michael Xie, George Azzari, and David B Lobell. Weakly supervised deep learning for segmentation of remote sensing imagery. Remote Sensing, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[50] Curtis E Woodcock, Richard Allen, Martha Anderson, Alan Belward, Robert Bindschadler, Warren Cohen, Feng Gao, Samuel N Goward, Dennis Helder, Eileen Helmer, et al. Free access to landsat imagery. Science, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.908, + 0.359 + ], + "angle": 0, + "content": "[51] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.415 + ], + "angle": 0, + "content": "[52] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.907, + 0.487 + ], + "angle": 0, + "content": "[53] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.907, + 0.543 + ], + "angle": 0, + "content": "[54] Rongtao Xu, Changwei Wang, Jiaxi Sun, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Self correspondence distillation for end-to-end weakly-supervised semantic segmentation. In AAAI, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.907, + 0.613 + ], + "angle": 0, + "content": "[55] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.908, + 0.656 + ], + "angle": 0, + "content": "[56] Liheng Zhong, Lina Hu, and Hang Zhou. Deep learning based multi-temporal crop classification. Remote Sensing of Environment, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.907, + 0.699 + ], + "angle": 0, + "content": "[57] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.907, + 0.742 + ], + "angle": 0, + "content": "[58] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.908, + 0.798 + ], + "angle": 0, + "content": "[59] Hao Zhu, Yan Zhu, Jiayu Xiao, Yike Ma, Yucheng Zhang, Jintao Li, and Feng Dai. Misa: mining saliency-aware semantic prior for box supervised instance segmentation. In IJCAI, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14045" + } + ] +] \ No newline at end of file diff --git a/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_origin.pdf b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc420ea20adcf835e0a010913007db5a6036305c --- /dev/null +++ b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/e1d2758f-cf15-4719-87c6-3d5649f01379_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eed4b01351d0eda62000646a370d8b45f9453479877e0b5aa9892e7d6e56d71e +size 2003017 diff --git a/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/full.md b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..447e845ce8142286a7c602200be6fc7319cf2be3 --- /dev/null +++ b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/full.md @@ -0,0 +1,363 @@ +# Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation + +Hao Zhu $^{1}$ , Yan Zhu $^{1}$ , Jiayu Xiao $^{1}$ , Tianxiang Xiao $^{1}$ , Yike Ma $^{1}$ , Yucheng Zhang $^{1}$ , Feng Dai $^{1\dagger}$ $^{1}$ Institute of Computing Technology, Chinese Academy of Sciences + +{zhuhao22z，fdai}@ict.ac.cn + +# Abstract + +Automated crop mapping through Satellite Image Time Series (SITS) has emerged as a crucial avenue for agricultural monitoring and management. However, due to the low resolution and unclear parcel boundaries, annotating pixel-level masks is exceptionally complex and time-consuming in SITS. This paper embraces the weakly supervised paradigm (i.e., only image-level categories available) to liberate the crop mapping task from the exhaustive annotation burden. The unique characteristics of SITS give rise to several challenges in weakly supervised learning: (1) noise perturbation from spatially neighboring regions, and (2) erroneous semantic bias from anomalous temporal periods. To address the above difficulties, we propose a novel method, termed exploring space-time perceptive clues (Exact). First, we introduce a set of spatial clues to explicitly capture the representative patterns of different crops from the most class-relative regions. Besides, we leverage the temporal-to-class interaction of the model to emphasize the contributions of pivotal clips, thereby enhancing the model perception for crop regions. Building upon the space-time perceptive clues, we derive the clue-based CAMs to effectively supervise the SITS segmentation network. Our method demonstrates impressive performance on various SITS benchmarks. Remarkably, the segmentation network trained on Exact-generated masks achieves $95\%$ of its fully supervised performance, showing the bright promise of weakly supervised paradigm in crop mapping scenario. Our code will be publicly available here. + +# 1. Introduction + +The launch of numerous public and commercial satellites provides broader opportunities to record, analyze, and predict the evolution of crop land [5, 17, 47, 50]. In this context, Satellite Image Time Series (SITS) with $10\mathrm{m}$ resolution offered by the high-frequency Sentinel-2 (S2) satellites serve as a valuable data source for automated crop map + +![](images/5c797dc2c5c990e075a7496532915e26ea68afc19ef2deac339d839bb9aaee48.jpg) +(a) Spatial Noise Perturbation + +![](images/80f40b0adaf35652df8aba154f9a24be2af06926319a47aece727dd2a1e919a6.jpg) +(b) Temporal Erroneous Semantic Bias +Figure 1. Illustration of the two inherent issues arises from spatial and temporal perspectives in SITS. (a) shows noise perturbation from the spatial perspective. We visualize the high-level feature manifold of Dog (natural image) and Barley (SITS) to reveal the distinct spatial properties. The feature dimensions are reduced by t-SNE [48]. (b) shows the erroneous semantic bias induced by anomalous temporal clips. We denote the parcel regions with $\star$ . The white circles refer to false positive activation regions in CAMs. + +ping [6, 14, 22, 28, 35, 56]. The core of crop mapping lies in the semantic segmentation of crop parcels. Recently, many efforts are devoted to exploiting the versatile and powerful relation modeling capability of deep neural networks for this task [20, 21, 34, 46]. While such methods have shown a significant progress, they rely heavily on pixel-level manual annotation, which is notoriously complex and time-consuming [49]. The low resolution of satellite images and the indistinct boundaries between crop parcels complicate the annotation process. Even worse, the varying phenologi + +cal cycles among crops require annotators to meticulously select appropriate acquisition times for annotation. + +To address this, one promising solution is weakly supervised semantic segmentation (WSSS), which relies solely on time-efficient annotation form, i.e., image-level categories. The image-level WSSS has commonly studied in natural image domain. Existing methods primarily extract Class Activation Map (CAM) [57] to generate pseudo labels for training semantic segmentation models. The intention of CAM is to identify the regions that highly contribute to the prediction of each class, revealing the shared patterns among the images within the same category. In context of natural images, CAM tends to highlight local discriminative parts of the target object. The reason is that the dispersed intraclass distribution enforces the model to learn a extremely sharp decision boundary, thereby the classifier weights tend to interact with feature representations within more discriminative object regions, as shown in Fig. 1a left. Thus the primary focus of researchers falls into expanding the CAM to identify the entire object semantics for natural images. + +Different from WSSS for natural image, crop mapping for SITS mainly faces with two challenges: (1) from the spatial perspective, parcel objects within the same category share uniform appearance and color, exhibiting strong neighboring consistency. The intra-class compactness of local patterns looses the tolerance of the classifier to noise perturbation, resulting in pronounced over-activation phenomena in CAM, as shown in Fig. 1a right. The disparate characteristic lead to the advances in natural image domain cannot directly benefit the SITS crop segmentation. (2) From the temporal perspective, although different crops show distinct phenological cycles and varying characteristics, they may display similar appearance in some specific periods. This confusion imbues wrong semantic bias to the learning process of the model, thereby activating some undesired semantic regions in CAM. An intuitive illustration is shown in Fig. 1b, the temporal clips that deviate from the pivotal semantic affect the perceptual ability of CAM to correct crop regions. + +In this work, we present exploring space-time perceptive clues (Exact), a tailored WSSS framework for crop mapping, to cope with the challenges arisen from the spatial and temporal aspects respectively. Firstly, we introduce a set of spatial clues to explicitly capture the patterns of different crops. Leveraging the filtered CAM as an indicator, we update representative clues by conducting spatial clustering from the most class-relative regions. These clues are then used to regularize the feature space via optimizing the contrastive objective, thereby sharpening the decision boundary of the model and mitigating the perturbations from illusory patterns. Secondly, to cope with the erroneous semantic bias caused by anomalous temporal periods, we propose temporal-aware affinity propagation to emphasize the contributions of pivotal clips for crop perception. In detail, we extract the temporal- + +to-class attention from the model to reweight the temporal sequence embeddings. The modulated representations can be used to model temporal-aware pairwise affinity for propagation on the raw CAM, thus effectively suppressing the undesired semantic regions in a self-supervised manner. + +Unlike existing WSSS methods that rely on classifier weights to generate CAMs, we exploit the well-updated space-time perceptive clues to derive the clue-based CAMs (CB-CAMs) as pseudo labels for segmentation. Compared to the raw CAM, the CB-CAMs (1) remarkably suppress the perturbations from spatial and temporal aspects, and (2) delineate the crop regions more precisely, thereby providing more reliable supervision for the subsequent segmentation. + +Our main contributions can be summarized as follows: + +- We introduce the WSSS paradigm to SITS crop mapping task to tackle the daunting annotation challenge. To the best of our knowledge, this is the first work that relies solely on image-level categories for crop segmentation. +- To overcome the drawbacks arising from the spatial and temporal aspects of SITS, we propose Exact that explores space-time perceptive clues to reduce the noise perturbation and rectify the wrong semantic bias, ultimately providing reliable supervision for SITS segmentation. +- We experimentally show that Exact achieves impressive performance on common benchmarks. Using Exact-generated labels for training, SITS segmentation model attains up to $95\%$ of its fully supervised performance. Our results significantly advance the upper bound of image-level WSSS technique compared to the other domains. + +# 2. Related Work + +Semantic segmentation on SITS. Automated crop monitoring through Satellite Image Time Series (SITS) has attracted great interest among researchers and demonstrated considerable social impact [14, 22, 23, 33, 36, 43, 44]. One of the challenging tasks is the semantic segmentation of agricultural parcels. The goal of the network is to learn a mapping function that assigns each pixel in the SITS to the corresponding crop type or background. Some works process SITS inputs by first extracting spatial information and then compressing the temporal dimension. For example, U-ConvLSTM [32] relied on U-Net [37] architecture to encode the spatial dimension, followed by a ConvLSTM [42] for the temporal dimension. Similarly, the FPN-ConvLSTM [34] replaced the U-Net with a Feature Pyramid Network [29] as the spatial encoder. U-TAE [20] compressed the temporal dimension through the outstanding temporal attention mechanism [21]. The other option to encode SITS is the temporal-spatio scheme, which first processes the temporal dimension and then extracts the spatial information. Rußwurm et al. [40] employed bidirectional LSTM to extract temporal features and then used CNNs to integrate spatial informa + +![](images/c36968eef2f675745a7a03a0e92cc89f0c041882122ae6cd3e3aa4e848f74bbd.jpg) +(a) Training Pipeline of the Exact + +![](images/5f3199a724496cdbe687d2cd3c9f2eb9f9feeb4fcdff1d5ee32d7950e8317c4f.jpg) +(b) Explore Spatial Perceptive Clues (§3.2) +Figure 2. (a) The training pipeline of Exact. We adopt the Temporal-Spatio scheme to handle the SITS input, which contains two transformer encoders. The first temporal encoder models interactions between acquisition times, then the followed spatial encoder discards the temporal dimension and models interactions between spatial positions. To overcome the difficulties arises from spatial and temporal aspects, we propose two novel technologies in temporal embedding space: (b) Explore Spatial Perceptive Clues to mitigate the noise perturbation (see §3.2) and (c) Temporal-Aware Affinity Propagation to rectify the wrong semantic bias(see §3.3). + +![](images/fe2ace1cb5469efb5d32887a328b8b9516c0c4e997e41ac09eed5b8c3b8c0324.jpg) +(c) Temporal-Aware Affinity Propagation (§3.3) + +tion. Recently, TSViT [46] proposed to borrow the powerful dependency modeling capability of the Vision Transformer (ViT) [16] to handle SITS, achieving state-of-the-art performance at lower computational cost. The TSViT also comprehensively illustrated the superiority of the temporal-spatio scheme. Taking a holistic view, we adopt this scheme to process SITS, aiming to achieve an optimal trade-off for automated crop mapping. + +Weakly supervised semantic segmentation. Weakly supervised semantic segmentation (WSSS) with image-level labels (i.e. ground-truth object categories) has shown significant success in natural images [13, 18, 52, 53, 59]. Most advanced WSSS methods follow the three-steps pipeline of: 1) training a classification network with image-level labels, 2) obtaining class activation map (CAM) [9, 41, 57] from the well-optimized classification network as pixel-level coarse labels, 3) replacing ground-truth masks with the pseudo labels to train a off-the-shelf segmentation model. The objective of our work focuses on the first and second steps, i.e., how to generate accurate pseudo labels. Numerous works has been proposed to address the under- and over-activation issues of the initial CAM [3, 10, 11, 26, 39, 55] in natural images. To derive the final pseudo labels, CAM requires cumbersome post-processing, including random walks [1, 2] and dense + +CRF [25] refinement. Thanks to these excellent works, the performance of image-level WSSS in natural images currently achieves $90\%$ of pixel-level supervision. However, the WSSS networks designed for natural images require significant adaptation to be applied to SITS, and the results are still unsatisfactory due to the distinct data characteristics. In this work, we incorporate WSSS technique into the SITS and overcome the difficulties caused by inherent data properties. + +# 3. Method + +In this section, we first look more closely at the temporal-spatio scheme and class activation map technique (§3.1). Next, we introduce how to explore spatial perceptive clues (§3.2) and conduct temporal-aware affinity propagation (§3.3). Finally, we show the overall objective function and the clue-based CAM generation strategy of Exact (§3.4). The training pipeline of Exact is shown in Fig. 2. + +# 3.1. Preliminaries + +The temporal-spatio scheme that first processes the temporal dimension and then extracting the spatial information has shown significant superiority when dealing with SITS data. Following the TSViT [46], we use the variant ViT [16] as the backbone for both temporal and spatial encoders. We + +consider a SITS input $\mathbf{X} \in \mathbb{R}^{T \times C \times H \times W}$ with $T$ the length of time series, $C$ the number of channels and $H \times W$ the spatial dimensions. The input $\mathbf{X}$ is mapped into a sequence of patch tokens and reshaped to $\mathbf{Z} \in \mathbb{R}^{N_h \cdot N_w \times T \times d}$ , where $N_h = \left[\frac{H}{h}\right]$ , $N_w = \left[\frac{W}{w}\right]$ and $h \times w$ is the spatial extent for each patch. We then add the temporal position embeddings $\mathbf{P}_T \in \mathbb{R}^{T \times d}$ and concatenate the temporal multi-class tokens $\mathbf{Z}_T^{\mathrm{cls}} \in \mathbb{R}^{K \times d}$ to obtain the input of temporal encoder: + +$$ +\mathbf {Z} _ {T} ^ {\mathrm {i n}} = \left[ \mathbf {Z} _ {T} ^ {\mathrm {c l s}}, \mathbf {Z} + \mathbf {P} _ {T} \right], \mathbf {Z} _ {T} ^ {\mathrm {i n}} \in \mathbb {R} ^ {N _ {h} \cdot N _ {w} \times (K + T) \times d} \tag {1} +$$ + +here $K$ denotes the number of categories, $\mathbf{Z}_T^{\mathrm{cls}}$ and $\mathbf{P}_T$ are repeated $N_{h}\cdot N_{w}$ times to match the spatial shape. With the output feature maps $\mathbf{Z}_T^{\mathrm{out}} = [\mathbf{Z}_T^{\mathrm{dense}}|\mathbf{Z}_T^{\mathrm{seq}}]$ from the temporal encoder, we extract the first $K$ tokens $\mathbf{Z}_T^{\mathrm{dense}}\in \mathbb{R}^{N_h\cdot N_w\times K\times d}$ and permute the first and second dimensions to serve as input patch tokens $\mathbf{Z}_S\in \mathbb{R}^{K\times N_h\cdot N_w\times d}$ for spatial encoder. These inputs are then combined with spatial multi-class tokens $\mathbf{Z}_{S}^{\mathrm{cls}}\in \mathbb{R}^{K\times 1\times d}$ and spatial position embeddings $\mathbf{P}_S\in \mathbb{R}^{N_h\cdot N_w\times d}$ at all $K$ spatial representations: + +$$ +\mathbf {Z} _ {S} ^ {\mathrm {i n}} = \left[ \mathbf {Z} _ {S} ^ {\mathrm {c l s}}, \mathbf {Z} _ {S} + \mathbf {P} _ {S} \right], \mathbf {Z} _ {S} ^ {\mathrm {i n}} \in \mathbb {R} ^ {K \times (1 + N _ {h} \cdot N _ {w}) \times d} \tag {2} +$$ + +After the spatial encoder phase, we separate the output features $\mathbf{Z}_S^{\mathrm{out}} = [\mathbf{Z}_S^{\mathrm{global}}|\mathbf{Z}_S^{\mathrm{dense}}]$ to align with different downstream tasks. For classification task, we feed the global tokens $\mathbf{Z}_S^{\mathrm{global}}\in \mathbb{R}^{K\times 1\times d}$ into the classifier to obtain classification logits. For segmentation task, to derive the dense prediction mask, the tokens $\mathbf{Z}_S^{\mathrm{dense}}\in \mathbb{R}^{K\times N_h\cdot N_w\times d}$ are fed into the segmentation decoder. + +Class activation map (CAM) [57] is widely used in WSSS to provide weak annotations that rely on image-level labels. Given a natural image, its feature maps $\mathbf{F} \in \mathbb{R}^{H' \times W' \times D}$ are extracted by a classification backbone. To derive the classification score, the feature maps are average pooled and multiplied by the classifier weights $\boldsymbol{w} \in \mathbb{R}^{K \times D}$ . CAM is generated by weighting and summing each channel in the feature maps with the classifier weights, as follows: + +$$ +\mathcal {M} ^ {k} = \operatorname {R e L U} \left(\sum_ {i} \boldsymbol {w} _ {i} ^ {k} \cdot \mathbf {F} _ {:,: i}\right), \quad \forall k \in K. \tag {3} +$$ + +Following most WSSS methods, we normalize $\mathcal{M}^k$ to the range [0,1] and apply a global threshold to filter out background pixels to obtain final pseudo labels. In the temporal-spatio network, we feed the dense tokens $\mathbf{Z}_T^{\mathrm{dense}}$ and $\mathbf{Z}_S^{\mathrm{dense}}$ into the classifier and compute the additional classification loss $\mathcal{L}_{\mathrm{cls}}^{\mathrm{aux}}$ to supervise the fused raw CAM for SITS. We provide more details in supplement. + +# 3.2. Explore Spatial Perceptive Clues + +CAM filtering. Given an input SITS $\mathbf{X}$ and its image-level label $y\in [0,1]^K$ , we first compute its normalized fused + +CAM $\mathcal{M} \in \mathbb{R}^{N_h \cdot N_w \times K}$ by the classifier weights and the output dense tokens of temporal and spatial encoder. We use two threshold scores $\mu_l$ and $\mu_h$ to filter out the reliable foreground, background and uncertain regions: + +$$ +\hat {\mathcal {M}} = \left\{ \begin{array}{l l} 0, & \text {i f} \mathcal {M} \leq \mu_ {l}, \\ 1, & \text {i f} \mathcal {M} \geq \mu_ {h}, \\ \text {i g n o r e ,} & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +Clues clustering. To capture the compact patterns of different crops, we establish a group of class-wise representative prototypes, which later serve as perceptive clues to generate high-quality pseudo labels. In crop segmentation task, temporal features often provide more information than spatial context [46], so we choose to perform spatial clustering on temporal dense embeddings $\mathbf{Z}_T^{\mathrm{dense}}$ over the whole dataset. Specifically, we build the class-wise positive prototype set $\mathcal{P}_{\mathrm{pos}} = \{\pmb{p}_{\mathrm{pos}}^k \in \mathbb{R}^{N_p \times d}\}_{k=1}^K$ and negative prototype set $\mathcal{P}_{\mathrm{neg}} = \{\pmb{p}_{\mathrm{neg}}^k \in \mathbb{R}^{N_p \times d}\}_{k=1}^K$ , where $N_p$ denotes the number of prototypes. If $k$ -th class appears in a training batch, we update the prototypes $\pmb{p}^k$ via solving the optimal transport problem [8, 58]. Given a mapping matrix $\mathbf{C}^k$ that represents the assignment between each pixel and its prototype, it can be referred as an element of the transportation polytope [4]: + +$$ +\mathbb {C} = \left\{\mathbf {C} ^ {k} \in \mathbb {R} _ {+} ^ {N _ {p} \times N _ {k}} \mid \mathbf {C} ^ {k} \mathbf {1} = \boldsymbol {u}, \mathbf {C} ^ {k \top} \mathbf {1} = \boldsymbol {r} \right\}, \tag {5} +$$ + +where $N_{k}$ denotes the number of pixels belonging to class $k$ . 1 represents the vectors of ones in appropriate dimensions, $\mathbf{u}$ and $\pmb{r}$ are the marginal projections onto the rows and columns of $\mathbf{C}^k$ . We can maximize the following objective function to optimize the mapping matrix: + +$$ +\underset {\mathbf {C} ^ {k} \in \mathbb {C}} {\text {m a x i m i z e}} \operatorname {T r} \left(\mathbf {C} ^ {k \top} \boldsymbol {p} ^ {k} \mathbf {Z} ^ {k \top}\right) - \eta \sum_ {n _ {p}, n _ {k} \in N _ {p}, N _ {k}} \mathbf {C} _ {n _ {p} n _ {k}} ^ {k} \log \mathbf {C} _ {n _ {p} n _ {k}} ^ {k}, \tag {6} +$$ + +here $\mathbf{Z}^k\in \mathbb{R}^{N_k\times d}$ is the corresponding temporal dense embeddings belonging to class $k$ , $\eta$ controls the smoothness of entropy regularization term. The continuous approximate solution of Eq. (6) can be obtained through iteratively applying the Sinkhorn-Knopp algorithm [15]. Subsequently, we momentum update the $n_p$ -th prototype of class $k$ according to the assignment matrix and the embeddings: + +$$ +\boldsymbol {p} _ {n _ {p}} ^ {k} = \alpha \boldsymbol {p} _ {n _ {p}} ^ {k} + \frac {1 - \alpha}{\left| \left| \mathbf {C} _ {n _ {p}} ^ {k} \right| \right| _ {1}} \cdot \left(\mathbf {C} _ {n _ {p}} ^ {k} \mathbf {Z} ^ {k}\right), \tag {7} +$$ + +here $\alpha \in [0,1]$ is the momentum coefficient. We leverage the fused CAM $\hat{\mathcal{M}}$ as the pseudo labels to update the class-wise positive prototype set $\mathcal{P}_{\mathrm{pos}}$ and $1 - \hat{\mathcal{M}}$ for the negative set $\mathcal{P}_{\mathrm{neg}}$ . Notably, each prototype is not involved in gradient backpropagation to avoid the noise from the classifier. + +Clue-based contrastive learning. Based on the spatial perceptive clues, we introduce the clue-based contrastive learning [7, 12] to regularize the embedding space. More specifically, for each temporal dense embedding $\mathbf{z}_{n_k}^k$ and its most + +![](images/2738a8636e98134d3e37dceb2686efeda35de10884ee6a1141df6e8631b7f205.jpg) +Figure 3. Visualization of the clue-based CAM generation. This process is performed after training the classification network. + +relative prototype $\pmb{p}_{n_p}^k$ , we adopt the cosine distance to measure their similarity [24]: + +$$ +\mathbf {S} \left(\boldsymbol {z} _ {n _ {k}} ^ {k}, \boldsymbol {p} _ {n _ {p}} ^ {k}\right) = \frac {\boldsymbol {z} _ {n _ {k}} ^ {k} \boldsymbol {p} _ {n _ {p}} ^ {k \top}}{\left\| \boldsymbol {z} _ {n _ {k}} ^ {k} \right\| \cdot \left\| \boldsymbol {p} _ {n _ {p}} ^ {k} \right\| / \tau}, \tag {8} +$$ + +where $\tau$ indicates the temperature parameter. Subsequently, we enforce the pixel embedding to closely match its prototype and be distinct from other prototypes: + +$$ +\mathcal {L} _ {\mathrm {c b l}} = \sum_ {k} \sum_ {n _ {k}} \mathbb {1} (\log (\sum_ {\boldsymbol {p} \in \mathcal {P} ^ {-}} \exp \mathrm {S} \left(\boldsymbol {z} _ {n _ {k}} ^ {k}, \boldsymbol {p}\right)) - \mathrm {S} \left(\boldsymbol {z} _ {n _ {k}} ^ {k}, \boldsymbol {p} _ {n _ {p}} ^ {k}\right)), \tag {9} +$$ + +where $\mathcal{P}^{-} = \{\mathcal{P}_{\mathrm{pos}} \cup \mathcal{P}_{\mathrm{neg}} \setminus \boldsymbol{p}_{n_p}^k\}$ and $\mathbb{1}(\cdot)$ is an indicator function, being 1 if class $k$ appears in image-level labels and 0 otherwise. Minimizing the above objective can pull pixel embedding closely to the semantic center and push it away from other illusory patterns, thereby sharpening the model decision boundary and mitigating the noise perturbation. + +# 3.3. Temporal-Aware Affinity Propagation + +In this section, we propose the temporal-aware affinity propagation to cope with the wrong semantic bias arising from anomalous temporal clips. + +Temporal-aware affinity mining. In the temporal encoder, the input tokens are normalized and projected into query matrix $\mathbf{Q} \in \mathbb{R}^{(K + T) \times d}$ and key matrix $\mathbf{K} \in \mathbb{R}^{(K + T) \times d}$ respectively. Then the self-attention $\mathcal{A} \in \mathbb{R}^{(K + T) \times (K + T)}$ with respect to $N_h \cdot N_w$ sequences is computed as below: + +$$ +\mathcal {A} = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q K} ^ {\top}}{\sqrt {d}}\right). \tag {10} +$$ + +Without additional computations nor supervision, we can explicitly extract the temporal-to-class attention $\tilde{\mathcal{A}}\in \mathbb{R}^{T\times K}$ from the self-attention $\mathcal{A}$ , which represents the contributions of high-level representations to crop recognition at different temporal clips. We then reweight the temporal sequence embeddings $\mathbf{Z}_T^{\mathrm{seq}}$ based on the normalized attention $\tilde{\mathcal{A}}$ : + +$$ +\mathcal {V} ^ {k} = \sum_ {t = 1} ^ {T} \tilde {\mathcal {A}} _ {t} ^ {k} \cdot \left(\mathbf {Z} _ {T} ^ {\text {s e q}}\right) _ {:, t,:}. \tag {11} +$$ + +Intuitively, the modulated high-level representation $\mathcal{V}^k\in$ $\mathbb{R}^{N_h\cdot N_w\times d}$ enlarges the variation among different semantic crops, thereby modeling more precise pixel relations. + +Affinity propagation and guidance. After obtaining the temporal-aware representation $\mathcal{V}$ , we conduct affinity propagation to suppress erroneous semantic regions on the raw CAM. Particularly, the temporal-aware pairwise affinity of the pixel at position $i$ can be estimated as follows: + +$$ +\operatorname {A f f} _ {j \in \mathcal {N} _ {i}} \left(\boldsymbol {v} _ {i} ^ {k}, \boldsymbol {v} _ {j} ^ {k}\right) = \exp \left(\frac {\mathcal {S} \left(\boldsymbol {v} _ {i} ^ {k} , \boldsymbol {v} _ {j} ^ {k}\right)}{\sigma_ {i} ^ {k}}\right), \tag {12} +$$ + +here $\sigma_{i}^{k}$ is the standard deviation of $\pmb{v}_{i}^{k}$ and $\mathcal{N}_i$ indicates the local receptive fields (e.g., 8-way local neighbors [27]). We denoise the raw CAM via iteratively propagating the pairwise affinity from the temporal-aware representation $\nu$ : + +$$ +\tilde {\mathcal {M}} _ {i} ^ {k} = \delta_ {i} ^ {k} \sum_ {j \in \mathcal {N} _ {i}} \operatorname {A f f} \left(\boldsymbol {v} _ {i} ^ {k}, \boldsymbol {v} _ {j} ^ {k}\right) \cdot \mathcal {M} _ {j} ^ {k}, \quad \text {w h e r e} \tag {13} +$$ + +$$ +\delta_ {i} ^ {k} = \frac {1}{\sum_ {j \in \mathcal {N} _ {i}} \mathsf {A f f} (\boldsymbol {v} _ {i} ^ {k} , \boldsymbol {v} _ {j} ^ {k})}. +$$ + +After obtaining the improved activation map, we align the raw CAM with it to effectively guide the learning process: + +$$ +\mathcal {L} _ {\text {t a p}} = \sum_ {k} \mathbb {1} \left(\left| \tilde {\mathcal {M}} ^ {k} - \mathcal {M} ^ {k} \right|\right). \tag {14} +$$ + +By minimizing $\mathcal{L}_{\mathrm{tap}}$ , we can rectify erroneous activations in the raw CAM and incorporate temporal-aware priors into the embedding space, ultimately benefiting the perceptive clues. + +# 3.4. Network Optimization and Clue-based CAM Generation Strategy + +As shown in Fig. 2, the whole objective function for training the classification network consists of four components: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {c l s}} + \mathcal {L} _ {\mathrm {c l s}} ^ {\text {a u x}} + \lambda_ {1} \mathcal {L} _ {\mathrm {c b l}} + \lambda_ {2} \mathcal {L} _ {\text {t a p}}, \tag {15} +$$ + +here $\mathcal{L}_{\mathrm{cls}}$ and $\mathcal{L}_{\mathrm{cls}}^{\mathrm{aux}}$ are the conventional binary cross entropy loss and $\lambda_{i}$ denotes the weight to rescale the loss terms. + +After training process, we perform per-pixel perception based on the well-updated space-time clues in temporal dense embedding space to generate clue-based CAMs (CB-CAMs). For each embedding $z_{i}$ in temporal dense embeddings $\mathbf{Z}_T^{\mathrm{dense}}$ , we measure the maximum similarity with the positive prototypes and minus the misguide activations with the negative prototypes: + +$$ +\mathcal {Y} _ {i} ^ {k} = \operatorname {R e L u} \left(\max _ {\boldsymbol {p} ^ {+} \in \boldsymbol {p} _ {k} ^ {\text {p o s}}} \mathrm {S} \left(\boldsymbol {z} _ {i}, \boldsymbol {p} ^ {+}\right) - \max _ {\boldsymbol {p} ^ {-} \in \boldsymbol {p} _ {k} ^ {\text {n e g}}} \mathrm {S} \left(\boldsymbol {z} _ {i}, \boldsymbol {p} ^ {-}\right)\right). \tag {16} +$$ + +If the class $k$ does not appear in a training image, we set $\mathcal{Y}^k$ to all zeros. Fig. 3 gives a visualization of this generation process. Finally, the CB-CAMs $\mathcal{Y}$ are filtered using a global background score to generate the pseudo labels, which are then used to train the SITS semantic segmentation network. + +
MethodTypeGermany [40]PASTIS [20]
OAmIoUOAmIoU
MCTFormercVPR'22 [52]ViT-PCMECCV'22 [38]LP-CAMCVPR'23 [13]TSCDAAI'23 [54]DuPLCVPR'24 [51]SeCOCVPR'24 [55]RGB70.656.366.749.6
66.852.269.353.2
64.749.167.050.1
68.452.767.251.3
66.250.665.548.7
64.548.363.646.1
baseline+PAMRCVPR'20 [3]+TS-CAMICCV'21 [19]+SIPECVPR'22 [11]+FPRICCV'23 [10]SITS82.574.181.269.5
83.976.382.071.2
81.170.580.467.6
82.073.281.570.1
82.475.681.771.0
+ours-Exact88.3 ↑ 5.880.6 ↑ 6.584.1 ↑ 2.975.6 ↑ 6.1
+ +(a) Pseudo labels performance of different methods on Germany and PASTIS train set. + +
MethodSup.OAmIoUratio
ConvLSTM [42]78.250.1
BiCGRU [40]80.556.2
FPN-ConvLSTM [34]81.959.5
Unet-3D [32]G82.360.4-
Unet-3Df [45]82.160.2
U-TAE [20]82.962.4
TSViT [46]83.465.1100%
baseline77.157.788%
+PAMRCVPR'20 [3]78.558.790%
+TS-CAMICCV'21 [19]P76.856.587%
+SIPECVPR'22 [11]77.158.089%
+FPRICCV'23 [10]78.258.490%
+ours-Exact80.3 ↑ 3.261.8 ↑ 4.195%
+ +(b) Segmentation performance trained with pseudo labels generated by different methods on PASTIS test set. + +Table 1. Comparisons with existing WSSS methods in OA(%) and mIoU(%) . RGB: the networks designed for processing natural images. SITs: the networks designed for processing SITS (include four adapted modules from natural image domain). $\mathcal{G}$ : various SITS segmentation networks supervised by ground-truth labels. $\mathcal{P}$ : the TSViT segmentation model supervised by different pseudo labels from the methods in SITs. The ratio refers to the proportion of mIoU between fully supervised and weakly supervised of TSViT segmentation. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets & evaluation metric. We conducted comprehensive experiments on two widely used Satellite-2 time series crop recognition datasets to validate the effectiveness of our approach quantitatively. The PASTIS dataset [20] contains 2433 multi-spectral time series of size $128 \times 128$ , each series including 10 bands and 33 to 61 temporal observations. It consists of 18 crop types and a background category. Following the settings of [46], we used the fold-1 among the five folds provided in PASTIS. We partitioned each sample into multiple patches and assigned category labels according to the mask annotations (see the supplement for details). The Germany dataset [40] comprises $137k$ field parcels of time series imagery with 13 spectral bands. Each series contains 36 observations and is labeled with 17 crop types. We employed the mean Intersection over Union (mIoU) and pixel-wise overall accuracy (OA) as evaluation metrics, both widely used to measure segmentation performance. + +Implementation details. The classification network was trained on 8 NVIDIA RTX 3090 GPUs with batch size 8 for 15k iterations. During the training stage, we used the AdamW optimizer [30] with an initial learning rate of 1e-3 and cosine weight decay [31]. As the raw CAM were unreliable in early iterations, the gradients of $\mathcal{L}_{\mathrm{cbl}}$ were backpropagated after the 4k iteration. For the raw CAM filtering, the threshold score $(\mu_l,\mu_h)$ was set to (0.2, 0.4). Due to the compact intra-class patterns, we set the number of class-specific prototypes $N_{p}$ to 2, and the momentum used to update the prototypes was set to 0.999. Moreover, the other hyper-parameters $\eta$ , $\lambda_1$ , $\lambda_2$ and $\tau$ were empirically set to 0.05, 0.01, 0.015 and + +0.1, respectively. We employed the original TSViT with a segmentation decoder as the semantic segmentation model for SITS. The training details of the segmentation model exactly followed the settings in [46] without any modifications. Please refer to supplementary materials for more details. + +# 4.2. Experimental Results + +Base Architectures. In the evaluation stage, we fuse the raw CAMs generated by the spatial dense embeddings $\mathbf{Z}_S^{\mathrm{dense}}$ and temporal dense embeddings $\mathbf{Z}_T^{\mathrm{dense}}$ as our baseline. + +# 4.2.1. Comparison with other WSSS Methods on CAMs. + +The key of the weakly supervised learning is to provide reliable training pseudo labels for segmentation network. Thus, we conducted comparative experiments on the train splits to validate the quality of pseudo labels. + +Competing methods adaptation. We reimplemented six well-performing transformer-based WSSS methods designed for natural images as competing methods. To accommodate the temporal dimension, we transformed the input dimensions of SITS into a 3D format $(T \times C, H, W)$ for these methods. In addition, to enable a fairer comparison between Exact and existing WSSS works, we carefully selected four off-the-shelf modules from other WSSS methods in the natural image domain and adapted them to the temporal-spatio framework. These modules were also reinplemented within the temporal dense embedding space to align with our approach. We attempted various adaptation strategies, more details and analysis available in supplementary materials. + +Quantitative results. Tab. 1a presents the mIoU and OA of the pseudo labels generated by other WSSS methods and our Exact. As can be seen, the WSSS methods designed for + +
Method\( \mathcal{L}_{\mathrm{cbl}} \)\( \mathcal{L}_{\mathrm{tap}} \)CB-CAMRecallOAmIoU
baseline81.181.269.5
✓81.882.472.3
✓✓82.583.073.4
ours-Exact✓✓✓84.884.175.6
+ +Table 2. Ablation study results of different components. The baseline is a setting with only image-level classification loss. + +
MethodSpa.Tem.RecallPrecisionOAmIoU
✓81.978.480.566.7
✓80.881.681.068.3
baseline CAM✓✓81.183.581.269.5
+ +Table 3. Baseline CAMs from different embedding spaces. Overall, the fused CAM shows the best performance. + +
MethodLow.Ta.RecallPrecisionOAmIoU
baseline81.183.581.269.5
✓80.984.081.169.9
ours-Exact✓81.885.682.472.3
✓✓81.785.282.071.8
+ +Table 4. Impact of modeling affinity from different sources. Low and $Ta$ refer to low-level cues and temporal-aware affinity. The temporal-aware affinity remarkably reduces noise. + +natural images only achieve a maximum $53.2\%$ mIoU (refer to PASTIS) with the limited encoding capability for SITS. Furthermore, due to the distinct data characteristics, the off-the-shelf modules struggle to perform well on SITS, and may even disturb the learning process. These phenomena indicate that the advancements in natural image cannot be directly translated to benefit the SITS domain. By contrast, our Exact overcome the intrinsic challenges from SITS, delivering the best performance compared to others. + +# 4.2.2. Segmentation results trained by pseudo labels. + +After obtaining the pseudo labels, we use them as replacement for ground truth to train the SITS semantic segmentation network. We employ TSViT followed by a segmentation head as our segmentation model, which is the top-performing approach in fully supervised oracles. + +Quantitative results. In Tab. 1b, we show that the semantic segmentation model can achieve $95\%$ to its fully supervised performance (refer to mIoU) using pseudo labels generated by Exact. Our best results are $80.3\%$ OA and $61.8\%$ mIoU on PASTIS test set, which surpasses the baseline by $3.2\%$ and $4.1\%$ , respectively. The experiment results quantitatively imply that the labels quality outperforms other methods by a + +![](images/f2450f385f3c5704297cf92af160c1ebe02576b1239492bcbac2c4ab13ae829d.jpg) +(a) baseline + +![](images/3285f440ee71c71d5c8242afe9236af849a91e671f628177f6b9dddce8df9e86.jpg) +(b)ours-Exact + +![](images/1f2a286cd46ec9feffd2128e76a6407afdca99f41c2d6906d219bdcef2a581c6.jpg) +Figure 4. Visualization of temporal feature spaces on PASTIS train set. The feature dimensions are reduced by t-SNE [48]. +(a) $N_{p}$ +Figure 5. Effect of the hyper-parameters. (a) the number of class-specific prototypes $N_{p}$ . (b) the temperature of similarity $\tau$ . + +![](images/0c3504aaee52cc9b4562328c6fbef05f93393f166ee5a07f0e61ba3bd6535f4b.jpg) +(b) $\pmb{\tau}$ + +large margin. In contrast, WSSS networks in natural image have undergone a series of evolutions to barely achieve $90\%$ of fully supervised performance. This suggests that SITS WSSS techniques offers greater potential in SITS scenario. + +# 4.3. Ablation Studies + +In this section, our primary purpose is to demonstrate the collective effectiveness of all components within our approach. We also choose the fused raw CAM as the baseline, and the CAM comparison results reported on PASTIS train set. + +All the components matter. Our Exact consists of several components, including spatial perceptive clues exploration, temporal-aware affinity propagation and clue-based CAM generation strategy. We validate the contribution of each module, the results are presented in Tab. 2. We can see that the clue-based contrastive learning $\mathcal{L}_{\mathrm{cbl}}$ and the temporal-aware affinity propagation $\mathcal{L}_{\mathrm{tap}}$ improves the performance by $1.8\%$ OA and $3.9\%$ mIoU, respectively. This suggests that these two modules complement each other, regularizing the entire embedding space and ultimately sharpening the decision boundary. We provide an intuitive comparison in Fig. 4, after introducing both objectives, the intra-class features become more compact while the inter-class features are more separated. Besides, using the well-updated space-time clues to generate CAM (i.e., CB-CAM) within the temporal dense embedding space brings a further improvement of $2.3\%$ recall, $1.1\%$ OA and $2.2\%$ mIoU. + +Baseline CAM from different embedding spaces. In Tab. 3, we investigate the performance of the raw CAMs derived + +![](images/81fea31a89be1ab6db254c719ca06134f0075e92638b38685c81730d3e4d49aa.jpg) +Figure 6. Qualitative comparisons. Left: CAMs comparisons on PASTIS train set. The stars represent the corresponding activation category. Right: Outputs of segmentation models trained with different CAMs on the PASTIS test set. + +from different dense embedding spaces. The raw CAM obtained from the spatial dense embeddings $\mathbf{Z}_S^{\mathrm{dense}}$ falls significantly behind that obtained from the temporal dense embeddings $\mathbf{Z}_T^{\mathrm{dense}}$ in terms of precision, OA and mIoU. It demonstrates that the temporal correlations contain more critical information compared to the spatial context in this task. We combine the CAMs from both embedding spaces as our baseline (the last row in Tab. 3) and utilize the resulting pseudo labels to guide spatial clustering. + +Effect of the temporal-aware affinity. Mining low-level cues affinity (e.g., color and intensity) as additional guidance is prevalent in natural images. We also attempted to incorporate the low-level cues to model pairwise affinity in SITS, the results can be found in Tab. 4. Compared to our proposed temporal-aware affinity, modeling low-level cues offers only limited improvement. This is due to the presence of substantial noise interference within the low-level cues of SITS (e.g., cloud cover and shadow). In contrast, our approach highlights the contributions of pivotal temporal clips within a high-level embedding space, effectively mitigating the wrong semantic bias. + +Effect of the hyper-parameters. Fig. 5 shows the effects of the prototype number $N_{p}$ and the temperature $\tau$ . Notably, utilizing only a minimal number of prototypes (i.e., $N_{p} = 2$ ) yields optimal results, whereas increasing the number of prototypes beyond this point brings negative impact on the model. This indicates that the excessive number may enforce prototypes to focus on local discriminative patterns + +within the class, leading to under-activation issue (the major difficulty in natural image). More analysis of parameters is specifically discussed in supplementary materials. + +# 4.4. Qualitative Results + +Fig. 6 presents the visual comparison between the baseline TSViT-CAMs and the proposed CB-CAMs generated by Exact on PASTIS dataset. The first five columns show the visualization of the CAMs. As we can see, our method remarkably suppress the erroneous regions, whose shape is closer to ground truth masks than baseline. Our high-quality pseudo labels subsequently enhance the segmentation performance compared to the baseline, as shown in last three columns. However, Exact still exhibits limitations in edge processing, which will be the focus in our future work. + +# 5. Conclusion and Broader Impact + +In this paper, we propose a tailored WSSS approach $\text{Exact}$ to alleviate the daunting annotation challenge in SITS crop mapping task. $\text{Exact}$ explores space-time perceptive clues to capture the essential patterns of different crop types, thereby overcoming the issues of spatial noise perturbation and wrong temporal semantic bias. Extensive experiments show the superiority of $\text{Exact}$ -generated masks both quantitatively and qualitatively. We believe that our method marks a pioneering step in effectively applying WSSS technologies to SITS. If follow-up work can find and resolve the limitations under our framework, there is great potential that only image-level labels are needed for crop mapping in the future. + +# Acknowledgements + +This work is supported by National Key R&D Program of China (2022YFD2001601) and National Natural Science Foundation of China (62372433). + +# References + +[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, 2018. 3 +[2] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 3 +[3] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 3, 6 +[4] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 4 +[5] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation. In ECCV, 2024. 1 +[6] Vitus Benson, Claire Robin, Christian Requena-Mesa, Lazaro Alonso, Nuno Carvalhais, José Cortés, Zhihan Gao, Nora Linscheid, Mélanie Weynants, and Markus Reichstein. Multimodal learning for geospatial vegetation forecasting. In CVPR, 2024. 1 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4 +[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4 +[9] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In WACV, 2018. 3 +[10] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, 2023. 3, 6 +[11] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6 +[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4 +[13] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, 2023. 3, 6 +[14] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 1, 2 + +[15] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 4 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[17] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: Esa's optical high-resolution mission for gmes operational services. Remote Sensing of Environment, 2012. 1 +[18] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, 2022. 3 +[19] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In CVPR, 2021. 6 +[20] Vivien Sainte Fare Garnot and Loic Landrieu. Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In ICCV, 2021. 1, 2, 6 +[21] Vivien Sainte Fare Garnot, Loic Landrieu, Sebastien Giordano, and Nesrine Chehata. Satellite image time series classification with pixel-set encoders and temporal self-attention. In CVPR, 2020. 1, 2 +[22] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In CVPR, 2024. 1, 2 +[23] Pengyu Hao, Yulin Zhan, Li Wang, Zheng Niu, and Muhammad Shakir. Feature selection of time series modis data for early crop classification using random forest: A case study in kansas, usa. Remote Sensing, 2015. 2 +[24] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, 2023. 5 +[25] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, 2011. 3 +[26] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3 +[27] Wentong Li, Yuqian Yuan, Song Wang, Wenyu Liu, Dongqi Tang, Jianke Zhu, Lei Zhang, et al. Label-efficient segmentation via affinity propagation. In NeurIPS, 2023. 5 +[28] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In CVPR, 2024. 1 +[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2 +[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2022. 6 + +[32] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In CVPR Workshops, 2019. 2, 6 +[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 2024. 2 +[34] Jorge Andres Chamorro Martinez, Laura Elena Cué La Rosa, Raul Queiroz Feitosa, Ieda Del'Arco Sanches, and Patrick Nigri Happ. Fully convolutional recurrent networks for multideate crop recognition from multitemporal image sequences. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 1, 2, 6 +[35] Sina Mohammadi, Mariana Belgiu, and Alfred Stein. Improvement in crop mapping from satellite image time series by effectively supervising deep neural networks. ISPRS Journal of Photogrammetry and Remote Sensing, 2023. 1 +[36] Charlotte Pelletier, Silvia Valero, Jordi Inglada, Nicolas Champion, and Gérard Dedieu. Assessing the robustness of random forests to map land cover with high resolution satellite image time series over large areas. Remote Sensing of Environment, 2016. 2 +[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2 +[38] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, 2022. 6 +[39] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In CVPR, 2023. 3 +[40] Marc Rußwurm and Marco Körner. Multi-temporal land cover classification with sequential recurrent encoders. ISPRS International Journal of Geo-Information, 2018. 2, 6 +[41] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3 +[42] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, 2015. 2, 6 +[43] Sofia Siachalou, Giorgos Mallinis, and Maria Tsakiri-Strati. A hidden markov models approach for crop classification: Linking crop phenology to time series of multi-sensor remote sensing data. Remote Sensing, 2015. 2 +[44] Haopeng Sun, Yingwei Zhang, Lumin Xu, Sheng Jin, and Yiqiang Chen. Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer. arXiv preprint arXiv:2412.10181, 2024. 2 +[45] Michail Tarasiou, Riza Alp Güler, and Stefanos Zafeiriou. Context-self contrastive pretraining for crop type semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6 +[46] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In CVPR, 2023. 1, 3, 4, 6 + +[47] Ramon Torres, Paul Snoeij, Dirk Geudtner, David Bibby, Malcolm Davidson, Evert Attema, Pierre Potin, Björn Rommen, Nicolas Floury, Mike Brown, et al. Gmes sentinel-1 mission. Remote Sensing of Environment, 2012. 1 +[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008. 1, 7 +[49] Sherrie Wang, William Chen, Sang Michael Xie, George Azzari, and David B Lobell. Weakly supervised deep learning for segmentation of remote sensing imagery. Remote Sensing, 2020. 1 +[50] Curtis E Woodcock, Richard Allen, Martha Anderson, Alan Belward, Robert Bindschadler, Warren Cohen, Feng Gao, Samuel N Goward, Dennis Helder, Eileen Helmer, et al. Free access to landsat imagery. Science, 2008. 1 +[51] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, 2024. 6 +[52] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6 +[53] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[54] Rongtao Xu, Changwei Wang, Jiaxi Sun, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Self correspondence distillation for end-to-end weakly-supervised semantic segmentation. In AAAI, 2023. 6 +[55] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, 2024. 3, 6 +[56] Liheng Zhong, Lina Hu, and Hang Zhou. Deep learning based multi-temporal crop classification. Remote Sensing of Environment, 2019. 1 +[57] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2, 3, 4 +[58] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 4 +[59] Hao Zhu, Yan Zhu, Jiayu Xiao, Yike Ma, Yucheng Zhang, Jintao Li, and Feng Dai. Misa: mining saliency-aware semantic prior for box supervised instance segmentation. In IJCAI, 2024. 3 \ No newline at end of file diff --git a/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/images.zip b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1fc322556117fa1593b4031802c99d2c832ee42 --- /dev/null +++ b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd100bfbe785f2eb6835c658c4d16c269d48a7cc4b8f299b1f9a3563327f986b +size 694389 diff --git a/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/layout.json b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ae24e35992b75b1bb6bd9a03086033e11880be9 --- /dev/null +++ b/2025/Exact_ Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation/layout.json @@ -0,0 +1,10282 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 102, + 521, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 102, + 521, + 140 + ], + "spans": [ + { + "bbox": [ + 89, + 102, + 521, + 140 + ], + "type": "text", + "content": "Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "spans": [ + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": "Hao Zhu" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": ", Yan Zhu" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": ", Jiayu Xiao" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": ", Tianxiang Xiao" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": ", Yike Ma" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": ", Yucheng Zhang" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": ", Feng Dai" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 530, + 189 + ], + "type": "text", + "content": "Institute of Computing Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 191, + 376, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 191, + 376, + 202 + ], + "spans": [ + { + "bbox": [ + 231, + 191, + 376, + 202 + ], + "type": "text", + "content": "{zhuhao22z，fdai}@ict.ac.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "content": "Automated crop mapping through Satellite Image Time Series (SITS) has emerged as a crucial avenue for agricultural monitoring and management. However, due to the low resolution and unclear parcel boundaries, annotating pixel-level masks is exceptionally complex and time-consuming in SITS. This paper embraces the weakly supervised paradigm (i.e., only image-level categories available) to liberate the crop mapping task from the exhaustive annotation burden. The unique characteristics of SITS give rise to several challenges in weakly supervised learning: (1) noise perturbation from spatially neighboring regions, and (2) erroneous semantic bias from anomalous temporal periods. To address the above difficulties, we propose a novel method, termed exploring space-time perceptive clues (Exact). First, we introduce a set of spatial clues to explicitly capture the representative patterns of different crops from the most class-relative regions. Besides, we leverage the temporal-to-class interaction of the model to emphasize the contributions of pivotal clips, thereby enhancing the model perception for crop regions. Building upon the space-time perceptive clues, we derive the clue-based CAMs to effectively supervise the SITS segmentation network. Our method demonstrates impressive performance on various SITS benchmarks. Remarkably, the segmentation network trained on Exact-generated masks achieves " + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "content": " of its fully supervised performance, showing the bright promise of weakly supervised paradigm in crop mapping scenario. Our code will be publicly available here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "type": "text", + "content": "The launch of numerous public and commercial satellites provides broader opportunities to record, analyze, and predict the evolution of crop land [5, 17, 47, 50]. In this context, Satellite Image Time Series (SITS) with " + }, + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "type": "text", + "content": " resolution offered by the high-frequency Sentinel-2 (S2) satellites serve as a valuable data source for automated crop map" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 242, + 552, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 443, + 241 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 443, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 443, + 241 + ], + "type": "text", + "content": "(a) Spatial Noise Perturbation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 242, + 552, + 384 + ], + "lines": [ + { + "bbox": [ + 319, + 242, + 552, + 384 + ], + "spans": [ + { + "bbox": [ + 319, + 242, + 552, + 384 + ], + "type": "image", + "image_path": "5c797dc2c5c990e075a7496532915e26ea68afc19ef2deac339d839bb9aaee48.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 397, + 552, + 475 + ], + "blocks": [ + { + "bbox": [ + 316, + 386, + 476, + 396 + ], + "lines": [ + { + "bbox": [ + 316, + 386, + 476, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 476, + 396 + ], + "type": "text", + "content": "(b) Temporal Erroneous Semantic Bias" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 397, + 552, + 475 + ], + "lines": [ + { + "bbox": [ + 319, + 397, + 552, + 475 + ], + "spans": [ + { + "bbox": [ + 319, + 397, + 552, + 475 + ], + "type": "image", + "image_path": "80f40b0adaf35652df8aba154f9a24be2af06926319a47aece727dd2a1e919a6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 483, + 555, + 573 + ], + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 573 + ], + "type": "text", + "content": "Figure 1. Illustration of the two inherent issues arises from spatial and temporal perspectives in SITS. (a) shows noise perturbation from the spatial perspective. We visualize the high-level feature manifold of Dog (natural image) and Barley (SITS) to reveal the distinct spatial properties. The feature dimensions are reduced by t-SNE [48]. (b) shows the erroneous semantic bias induced by anomalous temporal clips. We denote the parcel regions with " + }, + { + "bbox": [ + 313, + 483, + 555, + 573 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 313, + 483, + 555, + 573 + ], + "type": "text", + "content": ". The white circles refer to false positive activation regions in CAMs." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "ping [6, 14, 22, 28, 35, 56]. The core of crop mapping lies in the semantic segmentation of crop parcels. Recently, many efforts are devoted to exploiting the versatile and powerful relation modeling capability of deep neural networks for this task [20, 21, 34, 46]. While such methods have shown a significant progress, they rely heavily on pixel-level manual annotation, which is notoriously complex and time-consuming [49]. The low resolution of satellite images and the indistinct boundaries between crop parcels complicate the annotation process. Even worse, the varying phenologi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 702, + 133, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 133, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 133, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14036" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "cal cycles among crops require annotators to meticulously select appropriate acquisition times for annotation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 99, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 99, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 99, + 295, + 314 + ], + "type": "text", + "content": "To address this, one promising solution is weakly supervised semantic segmentation (WSSS), which relies solely on time-efficient annotation form, i.e., image-level categories. The image-level WSSS has commonly studied in natural image domain. Existing methods primarily extract Class Activation Map (CAM) [57] to generate pseudo labels for training semantic segmentation models. The intention of CAM is to identify the regions that highly contribute to the prediction of each class, revealing the shared patterns among the images within the same category. In context of natural images, CAM tends to highlight local discriminative parts of the target object. The reason is that the dispersed intraclass distribution enforces the model to learn a extremely sharp decision boundary, thereby the classifier weights tend to interact with feature representations within more discriminative object regions, as shown in Fig. 1a left. Thus the primary focus of researchers falls into expanding the CAM to identify the entire object semantics for natural images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 316, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 532 + ], + "type": "text", + "content": "Different from WSSS for natural image, crop mapping for SITS mainly faces with two challenges: (1) from the spatial perspective, parcel objects within the same category share uniform appearance and color, exhibiting strong neighboring consistency. The intra-class compactness of local patterns looses the tolerance of the classifier to noise perturbation, resulting in pronounced over-activation phenomena in CAM, as shown in Fig. 1a right. The disparate characteristic lead to the advances in natural image domain cannot directly benefit the SITS crop segmentation. (2) From the temporal perspective, although different crops show distinct phenological cycles and varying characteristics, they may display similar appearance in some specific periods. This confusion imbues wrong semantic bias to the learning process of the model, thereby activating some undesired semantic regions in CAM. An intuitive illustration is shown in Fig. 1b, the temporal clips that deviate from the pivotal semantic affect the perceptual ability of CAM to correct crop regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "In this work, we present exploring space-time perceptive clues (Exact), a tailored WSSS framework for crop mapping, to cope with the challenges arisen from the spatial and temporal aspects respectively. Firstly, we introduce a set of spatial clues to explicitly capture the patterns of different crops. Leveraging the filtered CAM as an indicator, we update representative clues by conducting spatial clustering from the most class-relative regions. These clues are then used to regularize the feature space via optimizing the contrastive objective, thereby sharpening the decision boundary of the model and mitigating the perturbations from illusory patterns. Secondly, to cope with the erroneous semantic bias caused by anomalous temporal periods, we propose temporal-aware affinity propagation to emphasize the contributions of pivotal clips for crop perception. In detail, we extract the temporal-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "to-class attention from the model to reweight the temporal sequence embeddings. The modulated representations can be used to model temporal-aware pairwise affinity for propagation on the raw CAM, thus effectively suppressing the undesired semantic regions in a self-supervised manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 133, + 553, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 229 + ], + "type": "text", + "content": "Unlike existing WSSS methods that rely on classifier weights to generate CAMs, we exploit the well-updated space-time perceptive clues to derive the clue-based CAMs (CB-CAMs) as pseudo labels for segmentation. Compared to the raw CAM, the CB-CAMs (1) remarkably suppress the perturbations from spatial and temporal aspects, and (2) delineate the crop regions more precisely, thereby providing more reliable supervision for the subsequent segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 230, + 547, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 230, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 325, + 230, + 547, + 241 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 243, + 553, + 426 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 243, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 553, + 291 + ], + "type": "text", + "content": "- We introduce the WSSS paradigm to SITS crop mapping task to tackle the daunting annotation challenge. To the best of our knowledge, this is the first work that relies solely on image-level categories for crop segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 293, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 293, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 314, + 293, + 553, + 353 + ], + "type": "text", + "content": "- To overcome the drawbacks arising from the spatial and temporal aspects of SITS, we propose Exact that explores space-time perceptive clues to reduce the noise perturbation and rectify the wrong semantic bias, ultimately providing reliable supervision for SITS segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 355, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 355, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 355, + 553, + 426 + ], + "type": "text", + "content": "- We experimentally show that Exact achieves impressive performance on common benchmarks. Using Exact-generated labels for training, SITS segmentation model attains up to " + }, + { + "bbox": [ + 314, + 355, + 553, + 426 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 314, + 355, + 553, + 426 + ], + "type": "text", + "content": " of its fully supervised performance. Our results significantly advance the upper bound of image-level WSSS technique compared to the other domains." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 441, + 400, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 400, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 400, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 714 + ], + "type": "text", + "content": "Semantic segmentation on SITS. Automated crop monitoring through Satellite Image Time Series (SITS) has attracted great interest among researchers and demonstrated considerable social impact [14, 22, 23, 33, 36, 43, 44]. One of the challenging tasks is the semantic segmentation of agricultural parcels. The goal of the network is to learn a mapping function that assigns each pixel in the SITS to the corresponding crop type or background. Some works process SITS inputs by first extracting spatial information and then compressing the temporal dimension. For example, U-ConvLSTM [32] relied on U-Net [37] architecture to encode the spatial dimension, followed by a ConvLSTM [42] for the temporal dimension. Similarly, the FPN-ConvLSTM [34] replaced the U-Net with a Feature Pyramid Network [29] as the spatial encoder. U-TAE [20] compressed the temporal dimension through the outstanding temporal attention mechanism [21]. The other option to encode SITS is the temporal-spatio scheme, which first processes the temporal dimension and then extracts the spatial information. Rußwurm et al. [40] employed bidirectional LSTM to extract temporal features and then used CNNs to integrate spatial informa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14037" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 550, + 205 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 188, + 83 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 188, + 83 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 188, + 83 + ], + "type": "text", + "content": "(a) Training Pipeline of the Exact" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 74, + 550, + 205 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 550, + 205 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 550, + 205 + ], + "type": "image", + "image_path": "c36968eef2f675745a7a03a0e92cc89f0c041882122ae6cd3e3aa4e848f74bbd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 224, + 306, + 332 + ], + "blocks": [ + { + "bbox": [ + 72, + 212, + 220, + 224 + ], + "lines": [ + { + "bbox": [ + 72, + 212, + 220, + 224 + ], + "spans": [ + { + "bbox": [ + 72, + 212, + 220, + 224 + ], + "type": "text", + "content": "(b) Explore Spatial Perceptive Clues (§3.2)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 224, + 306, + 332 + ], + "lines": [ + { + "bbox": [ + 62, + 224, + 306, + 332 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 306, + 332 + ], + "type": "image", + "image_path": "5f3199a724496cdbe687d2cd3c9f2eb9f9feeb4fcdff1d5ee32d7950e8317c4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 342, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 399 + ], + "type": "text", + "content": "Figure 2. (a) The training pipeline of Exact. We adopt the Temporal-Spatio scheme to handle the SITS input, which contains two transformer encoders. The first temporal encoder models interactions between acquisition times, then the followed spatial encoder discards the temporal dimension and models interactions between spatial positions. To overcome the difficulties arises from spatial and temporal aspects, we propose two novel technologies in temporal embedding space: (b) Explore Spatial Perceptive Clues to mitigate the noise perturbation (see §3.2) and (c) Temporal-Aware Affinity Propagation to rectify the wrong semantic bias(see §3.3)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 224, + 549, + 332 + ], + "blocks": [ + { + "bbox": [ + 337, + 212, + 499, + 224 + ], + "lines": [ + { + "bbox": [ + 337, + 212, + 499, + 224 + ], + "spans": [ + { + "bbox": [ + 337, + 212, + 499, + 224 + ], + "type": "text", + "content": "(c) Temporal-Aware Affinity Propagation (§3.3)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 224, + 549, + 332 + ], + "lines": [ + { + "bbox": [ + 328, + 224, + 549, + 332 + ], + "spans": [ + { + "bbox": [ + 328, + 224, + 549, + 332 + ], + "type": "image", + "image_path": "fe2ace1cb5469efb5d32887a328b8b9516c0c4e997e41ac09eed5b8c3b8c0324.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 297, + 515 + ], + "type": "text", + "content": "tion. Recently, TSViT [46] proposed to borrow the powerful dependency modeling capability of the Vision Transformer (ViT) [16] to handle SITS, achieving state-of-the-art performance at lower computational cost. The TSViT also comprehensively illustrated the superiority of the temporal-spatio scheme. Taking a holistic view, we adopt this scheme to process SITS, aiming to achieve an optimal trade-off for automated crop mapping." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": "Weakly supervised semantic segmentation. Weakly supervised semantic segmentation (WSSS) with image-level labels (i.e. ground-truth object categories) has shown significant success in natural images [13, 18, 52, 53, 59]. Most advanced WSSS methods follow the three-steps pipeline of: 1) training a classification network with image-level labels, 2) obtaining class activation map (CAM) [9, 41, 57] from the well-optimized classification network as pixel-level coarse labels, 3) replacing ground-truth masks with the pseudo labels to train a off-the-shelf segmentation model. The objective of our work focuses on the first and second steps, i.e., how to generate accurate pseudo labels. Numerous works has been proposed to address the under- and over-activation issues of the initial CAM [3, 10, 11, 26, 39, 55] in natural images. To derive the final pseudo labels, CAM requires cumbersome post-processing, including random walks [1, 2] and dense" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 418, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 515 + ], + "type": "text", + "content": "CRF [25] refinement. Thanks to these excellent works, the performance of image-level WSSS in natural images currently achieves " + }, + { + "bbox": [ + 313, + 418, + 556, + 515 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 418, + 556, + 515 + ], + "type": "text", + "content": " of pixel-level supervision. However, the WSSS networks designed for natural images require significant adaptation to be applied to SITS, and the results are still unsatisfactory due to the distinct data characteristics. In this work, we incorporate WSSS technique into the SITS and overcome the difficulties caused by inherent data properties." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 525, + 370, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 370, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 370, + 537 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 544, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 629 + ], + "type": "text", + "content": "In this section, we first look more closely at the temporal-spatio scheme and class activation map technique (§3.1). Next, we introduce how to explore spatial perceptive clues (§3.2) and conduct temporal-aware affinity propagation (§3.3). Finally, we show the overall objective function and the clue-based CAM generation strategy of Exact (§3.4). The training pipeline of Exact is shown in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 636, + 400, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 400, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 400, + 647 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "The temporal-spatio scheme that first processes the temporal dimension and then extracting the spatial information has shown significant superiority when dealing with SITS data. Following the TSViT [46], we use the variant ViT [16] as the backbone for both temporal and spatial encoders. We" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14038" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": "consider a SITS input " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " the length of time series, " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " the number of channels and " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " the spatial dimensions. The input " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " is mapped into a sequence of patch tokens and reshaped to " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{N_h \\cdot N_w \\times T \\times d}" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "N_h = \\left[\\frac{H}{h}\\right]" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "N_w = \\left[\\frac{W}{w}\\right]" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " is the spatial extent for each patch. We then add the temporal position embeddings " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_T \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " and concatenate the temporal multi-class tokens " + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{cls}} \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 55, + 71, + 296, + 169 + ], + "type": "text", + "content": " to obtain the input of temporal encoder:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 188, + 296, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 188, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 78, + 188, + 296, + 204 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {T} ^ {\\mathrm {i n}} = \\left[ \\mathbf {Z} _ {T} ^ {\\mathrm {c l s}}, \\mathbf {Z} + \\mathbf {P} _ {T} \\right], \\mathbf {Z} _ {T} ^ {\\mathrm {i n}} \\in \\mathbb {R} ^ {N _ {h} \\cdot N _ {w} \\times (K + T) \\times d} \\tag {1}", + "image_path": "c07fd6c61e29ca475a466b6989e82a712777c6d9190370c1bffea96f6a26c5f3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " denotes the number of categories, " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_T" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " are repeated " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "N_{h}\\cdot N_{w}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " times to match the spatial shape. With the output feature maps " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{out}} = [\\mathbf{Z}_T^{\\mathrm{dense}}|\\mathbf{Z}_T^{\\mathrm{seq}}]" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " from the temporal encoder, we extract the first " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{dense}}\\in \\mathbb{R}^{N_h\\cdot N_w\\times K\\times d}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " and permute the first and second dimensions to serve as input patch tokens " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_S\\in \\mathbb{R}^{K\\times N_h\\cdot N_w\\times d}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " for spatial encoder. These inputs are then combined with spatial multi-class tokens " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{S}^{\\mathrm{cls}}\\in \\mathbb{R}^{K\\times 1\\times d}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " and spatial position embeddings " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_S\\in \\mathbb{R}^{N_h\\cdot N_w\\times d}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " at all " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " spatial representations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 327, + 296, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 296, + 342 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {S} ^ {\\mathrm {i n}} = \\left[ \\mathbf {Z} _ {S} ^ {\\mathrm {c l s}}, \\mathbf {Z} _ {S} + \\mathbf {P} _ {S} \\right], \\mathbf {Z} _ {S} ^ {\\mathrm {i n}} \\in \\mathbb {R} ^ {K \\times (1 + N _ {h} \\cdot N _ {w}) \\times d} \\tag {2}", + "image_path": "2c4344b8aef0cddb1c6800093207e88d0f5f5fc5637240716c240d97d593d8da.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "text", + "content": "After the spatial encoder phase, we separate the output features " + }, + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_S^{\\mathrm{out}} = [\\mathbf{Z}_S^{\\mathrm{global}}|\\mathbf{Z}_S^{\\mathrm{dense}}]" + }, + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "text", + "content": " to align with different downstream tasks. For classification task, we feed the global tokens " + }, + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_S^{\\mathrm{global}}\\in \\mathbb{R}^{K\\times 1\\times d}" + }, + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "text", + "content": " into the classifier to obtain classification logits. For segmentation task, to derive the dense prediction mask, the tokens " + }, + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_S^{\\mathrm{dense}}\\in \\mathbb{R}^{K\\times N_h\\cdot N_w\\times d}" + }, + { + "bbox": [ + 55, + 351, + 296, + 434 + ], + "type": "text", + "content": " are fed into the segmentation decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 531 + ], + "type": "text", + "content": "Class activation map (CAM) [57] is widely used in WSSS to provide weak annotations that rely on image-level labels. Given a natural image, its feature maps " + }, + { + "bbox": [ + 55, + 435, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{H' \\times W' \\times D}" + }, + { + "bbox": [ + 55, + 435, + 296, + 531 + ], + "type": "text", + "content": " are extracted by a classification backbone. To derive the classification score, the feature maps are average pooled and multiplied by the classifier weights " + }, + { + "bbox": [ + 55, + 435, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 55, + 435, + 296, + 531 + ], + "type": "text", + "content": ". CAM is generated by weighting and summing each channel in the feature maps with the classifier weights, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 539, + 296, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 539, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 85, + 539, + 296, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {M} ^ {k} = \\operatorname {R e L U} \\left(\\sum_ {i} \\boldsymbol {w} _ {i} ^ {k} \\cdot \\mathbf {F} _ {:,: i}\\right), \\quad \\forall k \\in K. \\tag {3}", + "image_path": "d10b4313c06c84f7a476398ac80fe453401d2e003b7394cc9884ae44223943fc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "text", + "content": "Following most WSSS methods, we normalize " + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^k" + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "text", + "content": " to the range [0,1] and apply a global threshold to filter out background pixels to obtain final pseudo labels. In the temporal-spatio network, we feed the dense tokens " + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{dense}}" + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_S^{\\mathrm{dense}}" + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "text", + "content": " into the classifier and compute the additional classification loss " + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 55, + 580, + 296, + 665 + ], + "type": "text", + "content": " to supervise the fused raw CAM for SITS. We provide more details in supplement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 231, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 231, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 231, + 685 + ], + "type": "text", + "content": "3.2. Explore Spatial Perceptive Clues" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "CAM filtering. Given an input SITS " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and its image-level label " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "y\\in [0,1]^K" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " , we first compute its normalized fused" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": "CAM " + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N_h \\cdot N_w \\times K}" + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": " by the classifier weights and the output dense tokens of temporal and spatial encoder. We use two threshold scores " + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mu_l" + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mu_h" + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": " to filter out the reliable foreground, background and uncertain regions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 370, + 126, + 555, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 126, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 370, + 126, + 555, + 171 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {M}} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\mathcal {M} \\leq \\mu_ {l}, \\\\ 1, & \\text {i f} \\mathcal {M} \\geq \\mu_ {h}, \\\\ \\text {i g n o r e ,} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "d664b63930f445fba89d47fc26c52ed720da5574f400803a00ae4326152588b2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": "Clues clustering. To capture the compact patterns of different crops, we establish a group of class-wise representative prototypes, which later serve as perceptive clues to generate high-quality pseudo labels. In crop segmentation task, temporal features often provide more information than spatial context [46], so we choose to perform spatial clustering on temporal dense embeddings " + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{dense}}" + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": " over the whole dataset. Specifically, we build the class-wise positive prototype set " + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pos}} = \\{\\pmb{p}_{\\mathrm{pos}}^k \\in \\mathbb{R}^{N_p \\times d}\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": " and negative prototype set " + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{neg}} = \\{\\pmb{p}_{\\mathrm{neg}}^k \\in \\mathbb{R}^{N_p \\times d}\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": " denotes the number of prototypes. If " + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": "-th class appears in a training batch, we update the prototypes " + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{p}^k" + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": " via solving the optimal transport problem [8, 58]. Given a mapping matrix " + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^k" + }, + { + "bbox": [ + 313, + 177, + 555, + 358 + ], + "type": "text", + "content": " that represents the assignment between each pixel and its prototype, it can be referred as an element of the transportation polytope [4]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 363, + 555, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 363, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 340, + 363, + 555, + 380 + ], + "type": "interline_equation", + "content": "\\mathbb {C} = \\left\\{\\mathbf {C} ^ {k} \\in \\mathbb {R} _ {+} ^ {N _ {p} \\times N _ {k}} \\mid \\mathbf {C} ^ {k} \\mathbf {1} = \\boldsymbol {u}, \\mathbf {C} ^ {k \\top} \\mathbf {1} = \\boldsymbol {r} \\right\\}, \\tag {5}", + "image_path": "f46a52db4fdc47b6632fda7ce808741407ed7f8660fc83df364ceaf0b13ec380.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": " denotes the number of pixels belonging to class " + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": ". 1 represents the vectors of ones in appropriate dimensions, " + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": " are the marginal projections onto the rows and columns of " + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^k" + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": ". We can maximize the following objective function to optimize the mapping matrix:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 451, + 555, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 451, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 324, + 451, + 555, + 476 + ], + "type": "interline_equation", + "content": "\\underset {\\mathbf {C} ^ {k} \\in \\mathbb {C}} {\\text {m a x i m i z e}} \\operatorname {T r} \\left(\\mathbf {C} ^ {k \\top} \\boldsymbol {p} ^ {k} \\mathbf {Z} ^ {k \\top}\\right) - \\eta \\sum_ {n _ {p}, n _ {k} \\in N _ {p}, N _ {k}} \\mathbf {C} _ {n _ {p} n _ {k}} ^ {k} \\log \\mathbf {C} _ {n _ {p} n _ {k}} ^ {k}, \\tag {6}", + "image_path": "0e80714100e7d556857cb1256b5e1d11fd276386cbe07d4d21814a2eae1d4a43.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^k\\in \\mathbb{R}^{N_k\\times d}" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": " is the corresponding temporal dense embeddings belonging to class " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": " controls the smoothness of entropy regularization term. The continuous approximate solution of Eq. (6) can be obtained through iteratively applying the Sinkhorn-Knopp algorithm [15]. Subsequently, we momentum update the " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": "-th prototype of class " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": " according to the assignment matrix and the embeddings:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 358, + 572, + 555, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 572, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 358, + 572, + 555, + 600 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {n _ {p}} ^ {k} = \\alpha \\boldsymbol {p} _ {n _ {p}} ^ {k} + \\frac {1 - \\alpha}{\\left| \\left| \\mathbf {C} _ {n _ {p}} ^ {k} \\right| \\right| _ {1}} \\cdot \\left(\\mathbf {C} _ {n _ {p}} ^ {k} \\mathbf {Z} ^ {k}\\right), \\tag {7}", + "image_path": "0c108b372c260406d828b288eeb19f92009ce571101978c9032bd74d5e621465.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " is the momentum coefficient. We leverage the fused CAM " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " as the pseudo labels to update the class-wise positive prototype set " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "1 - \\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " for the negative set " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": ". Notably, each prototype is not involved in gradient backpropagation to avoid the noise from the classifier." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "Clue-based contrastive learning. Based on the spatial perceptive clues, we introduce the clue-based contrastive learning [7, 12] to regularize the embedding space. More specifically, for each temporal dense embedding " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{n_k}^k" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " and its most" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14039" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 75, + 289, + 157 + ], + "blocks": [ + { + "bbox": [ + 66, + 75, + 289, + 157 + ], + "lines": [ + { + "bbox": [ + 66, + 75, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 66, + 75, + 289, + 157 + ], + "type": "image", + "image_path": "2738a8636e98134d3e37dceb2686efeda35de10884ee6a1141df6e8631b7f205.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 169, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 295, + 194 + ], + "type": "text", + "content": "Figure 3. Visualization of the clue-based CAM generation. This process is performed after training the classification network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 213, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 237 + ], + "type": "text", + "content": "relative prototype " + }, + { + "bbox": [ + 55, + 213, + 296, + 237 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{n_p}^k" + }, + { + "bbox": [ + 55, + 213, + 296, + 237 + ], + "type": "text", + "content": ", we adopt the cosine distance to measure their similarity [24]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 244, + 296, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 244, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 244, + 296, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p} _ {n _ {p}} ^ {k}\\right) = \\frac {\\boldsymbol {z} _ {n _ {k}} ^ {k} \\boldsymbol {p} _ {n _ {p}} ^ {k \\top}}{\\left\\| \\boldsymbol {z} _ {n _ {k}} ^ {k} \\right\\| \\cdot \\left\\| \\boldsymbol {p} _ {n _ {p}} ^ {k} \\right\\| / \\tau}, \\tag {8}", + "image_path": "54bd156249af1d832a61fa9a05d693b20a75d208cab342a9241671e8715077b1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 284, + 297, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 284, + 297, + 321 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 284, + 297, + 321 + ], + "type": "text", + "content": " indicates the temperature parameter. Subsequently, we enforce the pixel embedding to closely match its prototype and be distinct from other prototypes:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 328, + 296, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 328, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 328, + 296, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c b l}} = \\sum_ {k} \\sum_ {n _ {k}} \\mathbb {1} (\\log (\\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {-}} \\exp \\mathrm {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p}\\right)) - \\mathrm {S} \\left(\\boldsymbol {z} _ {n _ {k}} ^ {k}, \\boldsymbol {p} _ {n _ {p}} ^ {k}\\right)), \\tag {9}", + "image_path": "312cfb94a566474c0a6d635872181d01e93dc653f7f501350b97396b43d4002c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{-} = \\{\\mathcal{P}_{\\mathrm{pos}} \\cup \\mathcal{P}_{\\mathrm{neg}} \\setminus \\boldsymbol{p}_{n_p}^k\\}" + }, + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "text", + "content": " is an indicator function, being 1 if class " + }, + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 366, + 296, + 438 + ], + "type": "text", + "content": " appears in image-level labels and 0 otherwise. Minimizing the above objective can pull pixel embedding closely to the semantic center and push it away from other illusory patterns, thereby sharpening the model decision boundary and mitigating the noise perturbation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 257, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 257, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 257, + 460 + ], + "type": "text", + "content": "3.3. Temporal-Aware Affinity Propagation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 464, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 499 + ], + "type": "text", + "content": "In this section, we propose the temporal-aware affinity propagation to cope with the wrong semantic bias arising from anomalous temporal clips." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "text", + "content": "Temporal-aware affinity mining. In the temporal encoder, the input tokens are normalized and projected into query matrix " + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{(K + T) \\times d}" + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "text", + "content": " and key matrix " + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{(K + T) \\times d}" + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "text", + "content": " respectively. Then the self-attention " + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{(K + T) \\times (K + T)}" + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "inline_equation", + "content": "N_h \\cdot N_w" + }, + { + "bbox": [ + 55, + 499, + 296, + 559 + ], + "type": "text", + "content": " sequences is computed as below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 567, + 296, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 567, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 126, + 567, + 296, + 595 + ], + "type": "interline_equation", + "content": "\\mathcal {A} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d}}\\right). \\tag {10}", + "image_path": "6bf98bd1cf644f73ea2372b34ff358ed5525f80facada4ba6d3129b0c5a3656e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": "Without additional computations nor supervision, we can explicitly extract the temporal-to-class attention " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{A}}\\in \\mathbb{R}^{T\\times K}" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": " from the self-attention " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": ", which represents the contributions of high-level representations to crop recognition at different temporal clips. We then reweight the temporal sequence embeddings " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{seq}}" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": " based on the normalized attention " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{A}}" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 123, + 683, + 296, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 683, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 123, + 683, + 296, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {V} ^ {k} = \\sum_ {t = 1} ^ {T} \\tilde {\\mathcal {A}} _ {t} ^ {k} \\cdot \\left(\\mathbf {Z} _ {T} ^ {\\text {s e q}}\\right) _ {:, t,:}. \\tag {11}", + "image_path": "ac46e2f294b8c48e759c9e1ddef7e38431447f10ea7eda2ad37b43c20ffe64d3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "Intuitively, the modulated high-level representation " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^k\\in" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N_h\\cdot N_w\\times d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " enlarges the variation among different semantic crops, thereby modeling more precise pixel relations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "text", + "content": "Affinity propagation and guidance. After obtaining the temporal-aware representation " + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "text", + "content": ", we conduct affinity propagation to suppress erroneous semantic regions on the raw CAM. Particularly, the temporal-aware pairwise affinity of the pixel at position " + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "text", + "content": " can be estimated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 366, + 174, + 555, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 174, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 366, + 174, + 555, + 206 + ], + "type": "interline_equation", + "content": "\\operatorname {A f f} _ {j \\in \\mathcal {N} _ {i}} \\left(\\boldsymbol {v} _ {i} ^ {k}, \\boldsymbol {v} _ {j} ^ {k}\\right) = \\exp \\left(\\frac {\\mathcal {S} \\left(\\boldsymbol {v} _ {i} ^ {k} , \\boldsymbol {v} _ {j} ^ {k}\\right)}{\\sigma_ {i} ^ {k}}\\right), \\tag {12}", + "image_path": "92ef736858c8d1e7eccaec67c76613f391f74f03d3bcb3955a635436ed239d1d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\sigma_{i}^{k}" + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "text", + "content": " is the standard deviation of " + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{i}^{k}" + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "text", + "content": " indicates the local receptive fields (e.g., 8-way local neighbors [27]). We denoise the raw CAM via iteratively propagating the pairwise affinity from the temporal-aware representation " + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 213, + 555, + 262 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 267, + 554, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 267, + 554, + 298 + ], + "spans": [ + { + "bbox": [ + 353, + 267, + 554, + 298 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {M}} _ {i} ^ {k} = \\delta_ {i} ^ {k} \\sum_ {j \\in \\mathcal {N} _ {i}} \\operatorname {A f f} \\left(\\boldsymbol {v} _ {i} ^ {k}, \\boldsymbol {v} _ {j} ^ {k}\\right) \\cdot \\mathcal {M} _ {j} ^ {k}, \\quad \\text {w h e r e} \\tag {13}", + "image_path": "d008225a703811b6ed529f33396ebabf16cad3d32ea9162f8904835450981166.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 294, + 462, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 294, + 462, + 319 + ], + "spans": [ + { + "bbox": [ + 361, + 294, + 462, + 319 + ], + "type": "interline_equation", + "content": "\\delta_ {i} ^ {k} = \\frac {1}{\\sum_ {j \\in \\mathcal {N} _ {i}} \\mathsf {A f f} (\\boldsymbol {v} _ {i} ^ {k} , \\boldsymbol {v} _ {j} ^ {k})}.", + "image_path": "968797a6e42eb6e6b97cee64e6c09c8caa4865827204f7bcfde5466bfad516c0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 324, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 348 + ], + "type": "text", + "content": "After obtaining the improved activation map, we align the raw CAM with it to effectively guide the learning process:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 376, + 354, + 555, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 354, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 376, + 354, + 555, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t a p}} = \\sum_ {k} \\mathbb {1} \\left(\\left| \\tilde {\\mathcal {M}} ^ {k} - \\mathcal {M} ^ {k} \\right|\\right). \\tag {14}", + "image_path": "aef22cb32ffc0301224000c795624e89901052bd8eb0efd95dc195dd8ba1f464.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 387, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 422 + ], + "type": "text", + "content": "By minimizing " + }, + { + "bbox": [ + 313, + 387, + 555, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tap}}" + }, + { + "bbox": [ + 313, + 387, + 555, + 422 + ], + "type": "text", + "content": ", we can rectify erroneous activations in the raw CAM and incorporate temporal-aware priors into the embedding space, ultimately benefiting the perceptive clues." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 429, + 554, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 453 + ], + "type": "text", + "content": "3.4. Network Optimization and Clue-based CAM Generation Strategy" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 458, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 554, + 483 + ], + "type": "text", + "content": "As shown in Fig. 2, the whole objective function for training the classification network consists of four components:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 359, + 490, + 555, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 490, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 359, + 490, + 555, + 504 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {c l s}} ^ {\\text {a u x}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {c b l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {t a p}}, \\tag {15}", + "image_path": "77b8889063be48f398e15cfeefb0b2c26ed447a23ed203df56f9915c5deba5dc.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "text", + "content": " are the conventional binary cross entropy loss and " + }, + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 313, + 510, + 554, + 533 + ], + "type": "text", + "content": " denotes the weight to rescale the loss terms." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 534, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 618 + ], + "type": "text", + "content": "After training process, we perform per-pixel perception based on the well-updated space-time clues in temporal dense embedding space to generate clue-based CAMs (CB-CAMs). For each embedding " + }, + { + "bbox": [ + 313, + 534, + 555, + 618 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 534, + 555, + 618 + ], + "type": "text", + "content": " in temporal dense embeddings " + }, + { + "bbox": [ + 313, + 534, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{dense}}" + }, + { + "bbox": [ + 313, + 534, + 555, + 618 + ], + "type": "text", + "content": ", we measure the maximum similarity with the positive prototypes and minus the misguide activations with the negative prototypes:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 624, + 555, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 555, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {Y} _ {i} ^ {k} = \\operatorname {R e L u} \\left(\\max _ {\\boldsymbol {p} ^ {+} \\in \\boldsymbol {p} _ {k} ^ {\\text {p o s}}} \\mathrm {S} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {p} ^ {+}\\right) - \\max _ {\\boldsymbol {p} ^ {-} \\in \\boldsymbol {p} _ {k} ^ {\\text {n e g}}} \\mathrm {S} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {p} ^ {-}\\right)\\right). \\tag {16}", + "image_path": "ac48425585469b75744bf36bc5fbd6ecf8b80d162af6586efa90abcc04b0d959.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "If the class " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " does not appear in a training image, we set " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^k" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " to all zeros. Fig. 3 gives a visualization of this generation process. Finally, the CB-CAMs " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " are filtered using a global background score to generate the pseudo labels, which are then used to train the SITS semantic segmentation network." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14040" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 60, + 334, + 237 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 334, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 334, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 334, + 237 + ], + "type": "table", + "html": "
MethodTypeGermany [40]PASTIS [20]
OAmIoUOAmIoU
MCTFormercVPR'22 [52]ViT-PCMECCV'22 [38]LP-CAMCVPR'23 [13]TSCDAAI'23 [54]DuPLCVPR'24 [51]SeCOCVPR'24 [55]RGB70.656.366.749.6
66.852.269.353.2
64.749.167.050.1
68.452.767.251.3
66.250.665.548.7
64.548.363.646.1
baseline+PAMRCVPR'20 [3]+TS-CAMICCV'21 [19]+SIPECVPR'22 [11]+FPRICCV'23 [10]SITS82.574.181.269.5
83.976.382.071.2
81.170.580.467.6
82.073.281.570.1
82.475.681.771.0
+ours-Exact88.3 ↑ 5.880.6 ↑ 6.584.1 ↑ 2.975.6 ↑ 6.1
", + "image_path": "e9d2c2347ac2effcf5159ca51cb640e953245389f2791381af07f1be7419d8e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 239, + 334, + 249 + ], + "lines": [ + { + "bbox": [ + 56, + 239, + 334, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 334, + 249 + ], + "type": "text", + "content": "(a) Pseudo labels performance of different methods on Germany and PASTIS train set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 342, + 61, + 553, + 237 + ], + "blocks": [ + { + "bbox": [ + 342, + 61, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 342, + 61, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 342, + 61, + 553, + 237 + ], + "type": "table", + "html": "
MethodSup.OAmIoUratio
ConvLSTM [42]78.250.1
BiCGRU [40]80.556.2
FPN-ConvLSTM [34]81.959.5
Unet-3D [32]G82.360.4-
Unet-3Df [45]82.160.2
U-TAE [20]82.962.4
TSViT [46]83.465.1100%
baseline77.157.788%
+PAMRCVPR'20 [3]78.558.790%
+TS-CAMICCV'21 [19]P76.856.587%
+SIPECVPR'22 [11]77.158.089%
+FPRICCV'23 [10]78.258.490%
+ours-Exact80.3 ↑ 3.261.8 ↑ 4.195%
", + "image_path": "9bc6ccccfa8eba94f83db8868fcd562401c43f7e4f741bac9ec312372b7f4cf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 345, + 239, + 550, + 258 + ], + "lines": [ + { + "bbox": [ + 345, + 239, + 550, + 258 + ], + "spans": [ + { + "bbox": [ + 345, + 239, + 550, + 258 + ], + "type": "text", + "content": "(b) Segmentation performance trained with pseudo labels generated by different methods on PASTIS test set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 268, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 313 + ], + "type": "text", + "content": "Table 1. Comparisons with existing WSSS methods in OA(%) and mIoU(%) . RGB: the networks designed for processing natural images. SITs: the networks designed for processing SITS (include four adapted modules from natural image domain). " + }, + { + "bbox": [ + 54, + 268, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 54, + 268, + 555, + 313 + ], + "type": "text", + "content": " : various SITS segmentation networks supervised by ground-truth labels. " + }, + { + "bbox": [ + 54, + 268, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 54, + 268, + 555, + 313 + ], + "type": "text", + "content": " : the TSViT segmentation model supervised by different pseudo labels from the methods in SITs. The ratio refers to the proportion of mIoU between fully supervised and weakly supervised of TSViT segmentation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 328, + 137, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 137, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 137, + 342 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 348, + 172, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 172, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 172, + 361 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 363, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 297, + 567 + ], + "type": "text", + "content": "Datasets & evaluation metric. We conducted comprehensive experiments on two widely used Satellite-2 time series crop recognition datasets to validate the effectiveness of our approach quantitatively. The PASTIS dataset [20] contains 2433 multi-spectral time series of size " + }, + { + "bbox": [ + 54, + 363, + 297, + 567 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 54, + 363, + 297, + 567 + ], + "type": "text", + "content": ", each series including 10 bands and 33 to 61 temporal observations. It consists of 18 crop types and a background category. Following the settings of [46], we used the fold-1 among the five folds provided in PASTIS. We partitioned each sample into multiple patches and assigned category labels according to the mask annotations (see the supplement for details). The Germany dataset [40] comprises " + }, + { + "bbox": [ + 54, + 363, + 297, + 567 + ], + "type": "inline_equation", + "content": "137k" + }, + { + "bbox": [ + 54, + 363, + 297, + 567 + ], + "type": "text", + "content": " field parcels of time series imagery with 13 spectral bands. Each series contains 36 observations and is labeled with 17 crop types. We employed the mean Intersection over Union (mIoU) and pixel-wise overall accuracy (OA) as evaluation metrics, both widely used to measure segmentation performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": "Implementation details. The classification network was trained on 8 NVIDIA RTX 3090 GPUs with batch size 8 for 15k iterations. During the training stage, we used the AdamW optimizer [30] with an initial learning rate of 1e-3 and cosine weight decay [31]. As the raw CAM were unreliable in early iterations, the gradients of " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cbl}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " were backpropagated after the 4k iteration. For the raw CAM filtering, the threshold score " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "(\\mu_l,\\mu_h)" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " was set to (0.2, 0.4). Due to the compact intra-class patterns, we set the number of class-specific prototypes " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " to 2, and the momentum used to update the prototypes was set to 0.999. Moreover, the other hyper-parameters " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " were empirically set to 0.05, 0.01, 0.015 and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "content": "0.1, respectively. We employed the original TSViT with a segmentation decoder as the semantic segmentation model for SITS. The training details of the segmentation model exactly followed the settings in [46] without any modifications. Please refer to supplementary materials for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 440, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 440, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 440, + 412 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "text", + "content": "Base Architectures. In the evaluation stage, we fuse the raw CAMs generated by the spatial dense embeddings " + }, + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_S^{\\mathrm{dense}}" + }, + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "text", + "content": " and temporal dense embeddings " + }, + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{dense}}" + }, + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "text", + "content": " as our baseline." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 456, + 555, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 469 + ], + "type": "text", + "content": "4.2.1. Comparison with other WSSS Methods on CAMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 521 + ], + "type": "text", + "content": "The key of the weakly supervised learning is to provide reliable training pseudo labels for segmentation network. Thus, we conducted comparative experiments on the train splits to validate the quality of pseudo labels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 557, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 557, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 557, + 676 + ], + "type": "text", + "content": "Competing methods adaptation. We reimplemented six well-performing transformer-based WSSS methods designed for natural images as competing methods. To accommodate the temporal dimension, we transformed the input dimensions of SITS into a 3D format " + }, + { + "bbox": [ + 313, + 521, + 557, + 676 + ], + "type": "inline_equation", + "content": "(T \\times C, H, W)" + }, + { + "bbox": [ + 313, + 521, + 557, + 676 + ], + "type": "text", + "content": " for these methods. In addition, to enable a fairer comparison between Exact and existing WSSS works, we carefully selected four off-the-shelf modules from other WSSS methods in the natural image domain and adapted them to the temporal-spatio framework. These modules were also reinplemented within the temporal dense embedding space to align with our approach. We attempted various adaptation strategies, more details and analysis available in supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Quantitative results. Tab. 1a presents the mIoU and OA of the pseudo labels generated by other WSSS methods and our Exact. As can be seen, the WSSS methods designed for" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14041" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 149 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 149 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 149 + ], + "type": "table", + "html": "
Method\\( \\mathcal{L}_{\\mathrm{cbl}} \\)\\( \\mathcal{L}_{\\mathrm{tap}} \\)CB-CAMRecallOAmIoU
baseline81.181.269.5
✓81.882.472.3
✓✓82.583.073.4
ours-Exact✓✓✓84.884.175.6
", + "image_path": "a2f69f1662379c00d5dfb8f0f3b684f6ed3590af011cc5d6bef06b6a44aff1bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 191, + 293, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "type": "text", + "content": "Table 2. Ablation study results of different components. The baseline is a setting with only image-level classification loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 191, + 293, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 191, + 293, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 191, + 293, + 258 + ], + "type": "table", + "html": "
MethodSpa.Tem.RecallPrecisionOAmIoU
✓81.978.480.566.7
✓80.881.681.068.3
baseline CAM✓✓81.183.581.269.5
", + "image_path": "ab1e07de61c4d1bdf3ef1cc2f697ee9a3b0f54d58bfb2a23e657c66da804261e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 299, + 294, + 378 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 296, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 289 + ], + "type": "text", + "content": "Table 3. Baseline CAMs from different embedding spaces. Overall, the fused CAM shows the best performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 299, + 294, + 378 + ], + "lines": [ + { + "bbox": [ + 58, + 299, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 299, + 294, + 378 + ], + "type": "table", + "html": "
MethodLow.Ta.RecallPrecisionOAmIoU
baseline81.183.581.269.5
✓80.984.081.169.9
ours-Exact✓81.885.682.472.3
✓✓81.785.282.071.8
", + "image_path": "9f23413f7fe8c158218f3adf5fb59eb751b03e6404fc821ce6b89860b190d3e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 387, + 296, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 420 + ], + "type": "text", + "content": "Table 4. Impact of modeling affinity from different sources. Low and " + }, + { + "bbox": [ + 55, + 387, + 296, + 420 + ], + "type": "inline_equation", + "content": "Ta" + }, + { + "bbox": [ + 55, + 387, + 296, + 420 + ], + "type": "text", + "content": " refer to low-level cues and temporal-aware affinity. The temporal-aware affinity remarkably reduces noise." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": "natural images only achieve a maximum " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "53.2\\%" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " mIoU (refer to PASTIS) with the limited encoding capability for SITS. Furthermore, due to the distinct data characteristics, the off-the-shelf modules struggle to perform well on SITS, and may even disturb the learning process. These phenomena indicate that the advancements in natural image cannot be directly translated to benefit the SITS domain. By contrast, our Exact overcome the intrinsic challenges from SITS, delivering the best performance compared to others." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 280, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 280, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 280, + 567 + ], + "type": "text", + "content": "4.2.2. Segmentation results trained by pseudo labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "text", + "content": "After obtaining the pseudo labels, we use them as replacement for ground truth to train the SITS semantic segmentation network. We employ TSViT followed by a segmentation head as our segmentation model, which is the top-performing approach in fully supervised oracles." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Quantitative results. In Tab. 1b, we show that the semantic segmentation model can achieve " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " to its fully supervised performance (refer to mIoU) using pseudo labels generated by Exact. Our best results are " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "80.3\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " OA and " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "61.8\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " mIoU on PASTIS test set, which surpasses the baseline by " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", respectively. The experiment results quantitatively imply that the labels quality outperforms other methods by a" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 434, + 178 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 434, + 178 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 434, + 178 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 434, + 178 + ], + "type": "image", + "image_path": "f2450f385f3c5704297cf92af160c1ebe02576b1239492bcbac2c4ab13ae829d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 178, + 394, + 186 + ], + "lines": [ + { + "bbox": [ + 352, + 178, + 394, + 186 + ], + "spans": [ + { + "bbox": [ + 352, + 178, + 394, + 186 + ], + "type": "text", + "content": "(a) baseline" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 436, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 436, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 436, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "3285f440ee71c71d5c8242afe9236af849a91e671f628177f6b9dddce8df9e86.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 178, + 522, + 186 + ], + "lines": [ + { + "bbox": [ + 471, + 178, + 522, + 186 + ], + "spans": [ + { + "bbox": [ + 471, + 178, + 522, + 186 + ], + "type": "text", + "content": "(b)ours-Exact" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 230, + 432, + 314 + ], + "blocks": [ + { + "bbox": [ + 313, + 195, + 554, + 218 + ], + "lines": [ + { + "bbox": [ + 313, + 195, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 554, + 218 + ], + "type": "text", + "content": "Figure 4. Visualization of temporal feature spaces on PASTIS train set. The feature dimensions are reduced by t-SNE [48]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 230, + 432, + 314 + ], + "lines": [ + { + "bbox": [ + 315, + 230, + 432, + 314 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 432, + 314 + ], + "type": "image", + "image_path": "1f2a286cd46ec9feffd2128e76a6407afdca99f41c2d6906d219bdcef2a581c6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 314, + 396, + 325 + ], + "lines": [ + { + "bbox": [ + 374, + 314, + 396, + 325 + ], + "spans": [ + { + "bbox": [ + 374, + 314, + 396, + 325 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 374, + 314, + 396, + 325 + ], + "type": "inline_equation", + "content": "N_{p}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "text", + "content": "Figure 5. Effect of the hyper-parameters. (a) the number of class-specific prototypes " + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "text", + "content": ". (b) the temperature of similarity " + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 432, + 230, + 553, + 314 + ], + "blocks": [ + { + "bbox": [ + 432, + 230, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 432, + 230, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 432, + 230, + 553, + 314 + ], + "type": "image", + "image_path": "0c3504aaee52cc9b4562328c6fbef05f93393f166ee5a07f0e61ba3bd6535f4b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 314, + 511, + 323 + ], + "lines": [ + { + "bbox": [ + 490, + 314, + 511, + 323 + ], + "spans": [ + { + "bbox": [ + 490, + 314, + 511, + 323 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 490, + 314, + 511, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "type": "text", + "content": "large margin. In contrast, WSSS networks in natural image have undergone a series of evolutions to barely achieve " + }, + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "type": "text", + "content": " of fully supervised performance. This suggests that SITS WSSS techniques offers greater potential in SITS scenario." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 422, + 414, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 414, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 414, + 434 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "text", + "content": "In this section, our primary purpose is to demonstrate the collective effectiveness of all components within our approach. We also choose the fused raw CAM as the baseline, and the CAM comparison results reported on PASTIS train set." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": "All the components matter. Our Exact consists of several components, including spatial perceptive clues exploration, temporal-aware affinity propagation and clue-based CAM generation strategy. We validate the contribution of each module, the results are presented in Tab. 2. We can see that the clue-based contrastive learning " + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cbl}}" + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": " and the temporal-aware affinity propagation " + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tap}}" + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": " improves the performance by " + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": " OA and " + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": " mIoU, respectively. This suggests that these two modules complement each other, regularizing the entire embedding space and ultimately sharpening the decision boundary. We provide an intuitive comparison in Fig. 4, after introducing both objectives, the intra-class features become more compact while the inter-class features are more separated. Besides, using the well-updated space-time clues to generate CAM (i.e., CB-CAM) within the temporal dense embedding space brings a further improvement of " + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": " recall, " + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": " OA and " + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 688 + ], + "type": "text", + "content": " mIoU." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Baseline CAM from different embedding spaces. In Tab. 3, we investigate the performance of the raw CAMs derived" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14042" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 555, + 319 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 555, + 319 + ], + "type": "image", + "image_path": "81fea31a89be1ab6db254c719ca06134f0075e92638b38685c81730d3e4d49aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 351 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons. Left: CAMs comparisons on PASTIS train set. The stars represent the corresponding activation category. Right: Outputs of segmentation models trained with different CAMs on the PASTIS test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 371, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 371, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 295, + 479 + ], + "type": "text", + "content": "from different dense embedding spaces. The raw CAM obtained from the spatial dense embeddings " + }, + { + "bbox": [ + 54, + 371, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_S^{\\mathrm{dense}}" + }, + { + "bbox": [ + 54, + 371, + 295, + 479 + ], + "type": "text", + "content": " falls significantly behind that obtained from the temporal dense embeddings " + }, + { + "bbox": [ + 54, + 371, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T^{\\mathrm{dense}}" + }, + { + "bbox": [ + 54, + 371, + 295, + 479 + ], + "type": "text", + "content": " in terms of precision, OA and mIoU. It demonstrates that the temporal correlations contain more critical information compared to the spatial context in this task. We combine the CAMs from both embedding spaces as our baseline (the last row in Tab. 3) and utilize the resulting pseudo labels to guide spatial clustering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 483, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 295, + 626 + ], + "type": "text", + "content": "Effect of the temporal-aware affinity. Mining low-level cues affinity (e.g., color and intensity) as additional guidance is prevalent in natural images. We also attempted to incorporate the low-level cues to model pairwise affinity in SITS, the results can be found in Tab. 4. Compared to our proposed temporal-aware affinity, modeling low-level cues offers only limited improvement. This is due to the presence of substantial noise interference within the low-level cues of SITS (e.g., cloud cover and shadow). In contrast, our approach highlights the contributions of pivotal temporal clips within a high-level embedding space, effectively mitigating the wrong semantic bias." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Effect of the hyper-parameters. Fig. 5 shows the effects of the prototype number " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": " and the temperature " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": ". Notably, utilizing only a minimal number of prototypes (i.e., " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "N_{p} = 2" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": ") yields optimal results, whereas increasing the number of prototypes beyond this point brings negative impact on the model. This indicates that the excessive number may enforce prototypes to focus on local discriminative patterns" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 371, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 407 + ], + "type": "text", + "content": "within the class, leading to under-activation issue (the major difficulty in natural image). More analysis of parameters is specifically discussed in supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 425, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 425, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 425, + 423 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 426, + 556, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 546 + ], + "type": "text", + "content": "Fig. 6 presents the visual comparison between the baseline TSViT-CAMs and the proposed CB-CAMs generated by Exact on PASTIS dataset. The first five columns show the visualization of the CAMs. As we can see, our method remarkably suppress the erroneous regions, whose shape is closer to ground truth masks than baseline. Our high-quality pseudo labels subsequently enhance the segmentation performance compared to the baseline, as shown in last three columns. However, Exact still exhibits limitations in edge processing, which will be the focus in our future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 552, + 495, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 495, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 495, + 567 + ], + "type": "text", + "content": "5. Conclusion and Broader Impact" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we propose a tailored WSSS approach " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\text{Exact}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " to alleviate the daunting annotation challenge in SITS crop mapping task. " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\text{Exact}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " explores space-time perceptive clues to capture the essential patterns of different crop types, thereby overcoming the issues of spatial noise perturbation and wrong temporal semantic bias. Extensive experiments show the superiority of " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\text{Exact}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "-generated masks both quantitatively and qualitatively. We believe that our method marks a pioneering step in effectively applying WSSS technologies to SITS. If follow-up work can find and resolve the limitations under our framework, there is great potential that only image-level labels are needed for crop mapping in the future." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14043" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "This work is supported by National Key R&D Program of China (2022YFD2001601) and National Natural Science Foundation of China (62372433)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 139, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 193 + ], + "type": "text", + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 225 + ], + "type": "text", + "content": "[2] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 228, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 295, + 250 + ], + "type": "text", + "content": "[3] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 251, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 283 + ], + "type": "text", + "content": "[4] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "text", + "content": "[5] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation. In ECCV, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "type": "text", + "content": "[6] Vitus Benson, Claire Robin, Christian Requena-Mesa, Lazaro Alonso, Nuno Carvalhais, José Cortés, Zhihan Gao, Nora Linscheid, Mélanie Weynants, and Markus Reichstein. Multimodal learning for geospatial vegetation forecasting. In CVPR, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 376, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 418 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 421, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 296, + 463 + ], + "type": "text", + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "text", + "content": "[9] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In WACV, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 511, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 296, + 553 + ], + "type": "text", + "content": "[10] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, 2023. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 556, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 296, + 599 + ], + "type": "text", + "content": "[11] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 601, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 296, + 634 + ], + "type": "text", + "content": "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 636, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 667 + ], + "type": "text", + "content": "[13] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, 2023. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[14] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[15] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 151, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 215 + ], + "type": "text", + "content": "[17] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: Esa's optical high-resolution mission for gmes operational services. Remote Sensing of Environment, 2012. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "type": "text", + "content": "[18] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "type": "text", + "content": "[19] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In CVPR, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 294, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 326 + ], + "type": "text", + "content": "[20] Vivien Sainte Fare Garnot and Loic Landrieu. Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In ICCV, 2021. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "text", + "content": "[21] Vivien Sainte Fare Garnot, Loic Landrieu, Sebastien Giordano, and Nesrine Chehata. Satellite image time series classification with pixel-set encoders and temporal self-attention. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 372, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 555, + 425 + ], + "type": "text", + "content": "[22] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 426, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 469 + ], + "type": "text", + "content": "[23] Pengyu Hao, Yulin Zhan, Li Wang, Zheng Niu, and Muhammad Shakir. Feature selection of time series modis data for early crop classification using random forest: A case study in kansas, usa. Remote Sensing, 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 471, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 555, + 502 + ], + "type": "text", + "content": "[24] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, 2023. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 504, + 554, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 554, + 535 + ], + "type": "text", + "content": "[25] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, 2011. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 537, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 568 + ], + "type": "text", + "content": "[26] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 571, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 555, + 602 + ], + "type": "text", + "content": "[27] Wentong Li, Yuqian Yuan, Song Wang, Wenyu Liu, Dongqi Tang, Jianke Zhu, Lei Zhang, et al. Label-efficient segmentation via affinity propagation. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 604, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 554, + 635 + ], + "type": "text", + "content": "[28] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In CVPR, 2024. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 636, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 668 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 691 + ], + "type": "text", + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 692, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 712 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2022. 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14044" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[32] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In CVPR Workshops, 2019. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 149 + ], + "type": "text", + "content": "[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 296, + 215 + ], + "type": "text", + "content": "[34] Jorge Andres Chamorro Martinez, Laura Elena Cué La Rosa, Raul Queiroz Feitosa, Ieda Del'Arco Sanches, and Patrick Nigri Happ. Fully convolutional recurrent networks for multideate crop recognition from multitemporal image sequences. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "type": "text", + "content": "[35] Sina Mohammadi, Mariana Belgiu, and Alfred Stein. Improvement in crop mapping from satellite image time series by effectively supervising deep neural networks. ISPRS Journal of Photogrammetry and Remote Sensing, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 315 + ], + "type": "text", + "content": "[36] Charlotte Pelletier, Silvia Valero, Jordi Inglada, Nicolas Champion, and Gérard Dedieu. Assessing the robustness of random forests to map land cover with high resolution satellite image time series over large areas. Remote Sensing of Environment, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 296, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 347 + ], + "type": "text", + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "type": "text", + "content": "[38] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 296, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 424 + ], + "type": "text", + "content": "[39] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "text", + "content": "[40] Marc Rußwurm and Marco Körner. Multi-temporal land cover classification with sequential recurrent encoders. ISPRS International Journal of Geo-Information, 2018. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "type": "text", + "content": "[41] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "type": "text", + "content": "[42] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, 2015. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "text", + "content": "[43] Sofia Siachalou, Giorgos Mallinis, and Maria Tsakiri-Strati. A hidden markov models approach for crop classification: Linking crop phenology to time series of multi-sensor remote sensing data. Remote Sensing, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 296, + 635 + ], + "type": "text", + "content": "[44] Haopeng Sun, Yingwei Zhang, Lumin Xu, Sheng Jin, and Yiqiang Chen. Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer. arXiv preprint arXiv:2412.10181, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 679 + ], + "type": "text", + "content": "[45] Michail Tarasiou, Riza Alp Güler, and Stefanos Zafeiriou. Context-self contrastive pretraining for crop type semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "type": "text", + "content": "[46] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In CVPR, 2023. 1, 3, 4, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 632 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[47] Ramon Torres, Paul Snoeij, Dirk Geudtner, David Bibby, Malcolm Davidson, Evert Attema, Pierre Potin, Björn Rommen, Nicolas Floury, Mike Brown, et al. Gmes sentinel-1 mission. Remote Sensing of Environment, 2012. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 150 + ], + "type": "text", + "content": "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008. 1, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "type": "text", + "content": "[49] Sherrie Wang, William Chen, Sang Michael Xie, George Azzari, and David B Lobell. Weakly supervised deep learning for segmentation of remote sensing imagery. Remote Sensing, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "text", + "content": "[50] Curtis E Woodcock, Richard Allen, Martha Anderson, Alan Belward, Robert Bindschadler, Warren Cohen, Feng Gao, Samuel N Goward, Dennis Helder, Eileen Helmer, et al. Free access to landsat imagery. Science, 2008. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 284 + ], + "type": "text", + "content": "[51] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 555, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 328 + ], + "type": "text", + "content": "[52] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In CVPR, 2022. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 555, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 385 + ], + "type": "text", + "content": "[53] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 387, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 555, + 430 + ], + "type": "text", + "content": "[54] Rongtao Xu, Changwei Wang, Jiaxi Sun, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Self correspondence distillation for end-to-end weakly-supervised semantic segmentation. In AAAI, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 432, + 555, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 555, + 485 + ], + "type": "text", + "content": "[55] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, 2024. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 487, + 555, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 555, + 519 + ], + "type": "text", + "content": "[56] Liheng Zhong, Lina Hu, and Hang Zhou. Deep learning based multi-temporal crop classification. Remote Sensing of Environment, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 521, + 555, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 555, + 553 + ], + "type": "text", + "content": "[57] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 555, + 555, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 555, + 587 + ], + "type": "text", + "content": "[58] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 589, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 555, + 632 + ], + "type": "text", + "content": "[59] Hao Zhu, Yan Zhu, Jiayu Xiao, Yike Ma, Yucheng Zhang, Jintao Li, and Feng Dai. Misa: mining saliency-aware semantic prior for box supervised instance segmentation. In IJCAI, 2024. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14045" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_content_list.json b/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..21fa52f4f67106460c8b721caa63b8ffe832d5cf --- /dev/null +++ b/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_content_list.json @@ -0,0 +1,1921 @@ +[ + { + "type": "text", + "text": "ExpertAF: Expert Actionable Feedback from Video", + "text_level": 1, + "bbox": [ + 236, + 130, + 759, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kumar Ashutosh $^{1,2}$ , Tushar Nagarajan $^{2}$ , Georgios Pavlakos $^{1}$ , Kris Kitani $^{2,3}$ , Kristen Grauman $^{1,2}$ $^{1}$ University of Texas at Austin, $^{2}$ FAIR Meta, $^{3}$ Carnegie Mellon University", + "bbox": [ + 120, + 179, + 875, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 251, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feedback is essential for learning a new skill or improving one's current skill-level. However, current methods for skill-assessment from video only provide scores or compare demonstrations, leaving the burden of knowing what to do differently on the user. We introduce a novel method to generate actionable feedback (AF) from video of a person doing a physical activity, such as basketball or soccer. Our method takes a video demonstration and its accompanying 3D body pose and generates (1) free-form expert commentary describing what the person is doing well and what they could improve, and (2) a visual expert demonstration that incorporates the required corrections. We show how to leverage Ego-Exo4D's [29] videos of skilled activity and expert commentary together with a strong language model to create a weakly-supervised training dataset for this task, and we devise a multimodal video-language model to infer coaching feedback. Our method is able to reason across multi-modal input combinations to output full-spectrum, actionable coaching—expert commentary, expert video retrieval, and expert pose generation—outperforming strong vision-language models on both established metrics and human preference studies.", + "bbox": [ + 89, + 282, + 483, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 642, + 220, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An abundance of instructional \"how-to\" videos on the internet enables skill-learning by observing expert demonstrations. Instructional videos cover many skills one might want to learn—cooking, DIY, sports, crafts, and hobbies. Recent work leverages such videos to assist human [5, 6, 53] and robot [56] skill learning. Acquiring skills from video is especially appealing for equitable learning: access to video is generally less costly and more widely available than face-to-face access to an expert coach.", + "bbox": [ + 89, + 667, + 482, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the myriad of videos and tutorials, people prefer to learn in a feedback loop: getting to know their mistakes, finding the improvements they need to do and correcting those. For that reason, coaching-based iterative training is more effective than self-learning [48]. A good coach has", + "bbox": [ + 89, + 804, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learner demonstration (input)", + "bbox": [ + 516, + 250, + 687, + 261 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e3b994c58c5bc89cce539628bb1eff839d9ae97b08bdb44e825cd5fb66abfed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 263, + 727, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6038d690b7c7e225445bee77186e627ab34e311406aef7fba9c5b97de97afd6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 364, + 733, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Expert demonstration (output)", + "bbox": [ + 733, + 250, + 901, + 261 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edfaf44dfd1543af4a8afffc07c108f5412bac965a4bf6da4551e53bf82cd72a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 263, + 921, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8bc25e3a2b530b15ffb32fbb177680dc5304d424b8a3fb10dd20934ec3e4ede3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 362, + 895, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Expert commentary (output)", + "bbox": [ + 521, + 438, + 681, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The player is taking big steps to control the ball, but lacks body control, and needs to slow down to maintain a better control.", + "bbox": [ + 532, + 457, + 901, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. An example of expert feedback. When a player is dribbling the ball fast, they tend to lose control (top left). Our proposed method provides an expert commentary to the learner suggesting improvements (bottom). The method also provides an expert demonstration that shows the desired correction, where the player is maintaining smaller steps and body control (top right).", + "bbox": [ + 511, + 497, + 906, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "three components: finding mistakes, providing verbal corrections and finally, showing visual demonstrations. For example, suppose a person learning to dribble a soccer ball takes big steps that decrease body control. The coach should identify this mistake, provide verbal feedback, and show the correct dribbling technique. See Figure 1.", + "bbox": [ + 511, + 594, + 905, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At present, a person watching a how-to video has to identify mistakes themselves and attempt to correct them. Pinpointing the exact mistake is itself challenging for beginners—particularly when learning physical skills, like sports and dance, where subtle movements and positioning of the body are critical to success.", + "bbox": [ + 511, + 686, + 905, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What role could AI play in overcoming this gap? Current work on proficiency understanding from video focuses on scoring a demonstration [9, 28, 29, 78], thus addressing only one component of coaching. There is no method that can provide actionable verbal and visual feedback on a learner's demonstration. Doing so is technically challenging since it requires the method to (a) understand the user's activity, (b) detect mistakes in the execution by comparing", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://vision.cs.utexas.edu/projects/ExpertAF/", + "bbox": [ + 112, + 887, + 434, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13582", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the user's activity with correct examples and finally, (c) suggest edits with respect to the better way of doing the activity. In short, the output should be coach-like—specific actionable feedback that will improve a learner's proficiency.", + "bbox": [ + 89, + 90, + 480, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose ExpertAF: a novel method to provide expert actionable feedback on a user's physical activity. ExpertAF's actionable feedback has two forms: language-based commentary and visual demonstration. The freeform expert commentary output describes the mistake and what should be corrected, e.g., \"take smaller steps and slow down to maintain control\", while the expert visual demonstration output shows the correct way of doing it.", + "bbox": [ + 89, + 161, + 482, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both components introduce unique technical challenges. For the former, the task of generating free-form text feedback is distinct from standard image and video captioning, as the model needs to identify the mistake and actions that would correct it—rather than simply describe the observed pose [16] or activity [32, 38, 57, 75, 81]. Similarly, for the latter, the task of retrieving (or generating) a demonstration that corrects a specific mistake is distinct from similarity-based retrieval [3, 40, 54, 77, 79] or open-ended generation [12, 19]. We hypothesize—and experimentally validate—that personalized, video-conditioned coaching is more accurate than simply returning a global expert execution because there are many right ways of doing an activity, e.g., penalty shots in soccer, and hence different ways to fix what is wrong in a given execution. Overall, the model must understand both what is being done and what actionable tweaks would make it better.", + "bbox": [ + 91, + 292, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We are the first to address video-based coaching with actionable feedback. We develop the first model to generate free-form text feedback on a user's activity (captured in video and 3D pose sequences) as well as a video demonstrating how to improve, by building on a recent vision-language model [45]. There is currently no dataset for this challenging task. Therefore, we augment Ego-Exo4D [29], an existing dataset with video, commentary, and poses, and show how to use a strong large language model to create a weakly-supervised training set consisting of paired learner-expert demonstrations, along with expert textual commentary that relates the two. We also obtain a gold-standard manually labeled test set for rigorous evaluation.", + "bbox": [ + 89, + 559, + 482, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate ExpertAF on three diverse physical scenarios: soccer, basketball, and rock climbing. Our results show a significant improvement over strong baselines, establishing the first method to provide actionable feedback, including a novel technique for expert pose generation. Alongside consistent gains in quantitative metrics against the ground truth, we also show our model's promise via direct human evaluation, where ExpertAF outperforms off-the-shelf video models by as much as $3 \\times$ .", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 90, + 651, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skill learning from videos. Instructional video datasets like HowTo100M [53], COIN [70], and CrossTask [87] enable procedural understanding through step recognition [5, 42, 85], procedure planning [11, 13, 84], task-graph discovery [5, 18, 29, 85], action anticipation [1, 21, 22, 26, 51], and alignment detection [4, 30]. Capitalizing on their instructional nature, recent work learns robot policies from these videos [50, 69], uses external knowledge-bases like WikiHow to ground the steps of a procedural task [42, 52, 85], and explores new ways to navigate between multiple demonstrations [6]. Despite their scale, how-to videos in these datasets [53, 70, 87] are often created by domain experts and hence lack mistakes or suboptimal executions. In contrast, Ego-Exo4D [29] contains multiple executions of various common scenarios by people with varied skill levels—from novice to late expert. Moreover, it contains experts' feedback on those actions. We show how to augment this data via large language models to support our new idea for video-based coaching.", + "bbox": [ + 511, + 114, + 903, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D body pose for activity understanding. 3D body pose is crucial for human activity understanding, capturing a person's stance and motion. Supported by valuable datasets like Humans3.6M [33] and NTU RGB-D [46, 66], recent methods improve body pose understanding [25, 34, 67, 72, 83, 86]. Beyond action classification, pose is also crucial for markerless motion capture of sports [8, 15] and interacting in augmented reality [24, 31]. While our ExpertAF leverages pose as an important signal of human activity, unlike prior work we interface video, pose, and free-form text to generate expert feedback and an expert demonstration video. Our work is also different from generating [20] or modifying [17] body pose from text (e.g., \"raise your left arm\"), as our task is to understand a potentially incorrect pose and provide feedback. Understanding these minor differences between incorrect and correct poses offers unique challenges that are addressed by our proposed learning scheme. Our work is orthogonal to methods that improve pose estimation itself; any future improvements on that front would only benefit our approach.", + "bbox": [ + 511, + 402, + 903, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skill assessment and coaching. Prior work explores skill assessment for a variety of tasks, particularly sports. Most methods pose skill-assessment as a score-prediction task, i.e., for skating [78], gymnastics [28], basketball [9], or multiple sports [61]. Since absolute scoring can be ambiguous and error-prone, recent work [71] explores uncertainty-aware score distributions. Instead of explicitly scoring a demonstration, other work determines which of two basketball players are better [9], uses group-aware contrastive regression to learn the relative quality of demonstrations [80], or generates a full ideal trajectory from the first frame [10]. Fitness-AQA [63] and Action Quality [64] provide outputs like knees inward error, shallow squat error,", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13583", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or an arrow highlight for more localized feedback. There is also research on skill assessment in non-sports domains, like surgery [43, 76] and piano [62]. All the prior work assumes a fixed taxonomy of errors, and the taxonomy is designed separately for each activity. Furthermore, unlike our work, none of the prior work provides feedback akin to a personal coach—which requires not only telling what is wrong, but also expressing how to correct the mistake.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 223, + 181, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce the problem statement in Sec. 3.1, the dataset creation strategy in Sec. 3.2, the training design in Sec. 3.3, and implementation details in Sec. 3.4.", + "bbox": [ + 89, + 248, + 482, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem statement", + "text_level": 1, + "bbox": [ + 89, + 304, + 272, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a dataset $\\mathcal{E} = \\{(\\mathcal{V}, T, \\bar{\\mathcal{V}})\\}$ where each $\\mathcal{V}_i$ is a video demonstration, and $T_i$ is a free-form text commentary critiquing the activity in the video, e.g., \"take smaller steps and slow down to maintain control.\" $\\bar{\\mathcal{V}}$ is a related video demonstration but without the error mentioned in $T$ , e.g., the player takes smaller steps with better control. See Fig. 1. Each demonstration $\\mathcal{V} = \\{V, P, S\\}$ consists of three parts: $V$ is the RGB video, and $P$ is the 3D pose sequence of the participant with skill-level $S$ in the video. The pose representation $P \\in \\mathbb{R}^{n \\times d \\times 3}$ contains $n$ frames of 3D positions for $d = 17$ body joints, consistent with MS-COCO [41]. $S$ is the participant's skill-level for an activity, i.e., novice, early expert, intermediate expert, or late expert, and will be used later in creating the dataset. We assume there is one active person doing a physical activity in any given video who is the subject of ExpertAF.", + "bbox": [ + 89, + 325, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal in this work is to provide feedback on a given learner's video containing a physical scenario. The feedback is an expert demonstration $\\bar{\\mathcal{V}}$ and an expert commentary $T$ , tailored to $\\mathcal{V}$ . Formally, we want to learn a mapping $\\mathcal{V} \\rightarrow (\\bar{\\mathcal{V}}, T)$ . While it is possible to learn these mappings separately (i.e., $\\mathcal{V} \\rightarrow \\bar{\\mathcal{V}}$ and $\\mathcal{V} \\rightarrow T$ ), the expert commentary and demonstration are tightly related and provide important contextual information for generating each other. We therefore treat this as an autoregressive generation problem where during training, we use both $\\mathcal{V}$ and $T$ to generate $\\bar{\\mathcal{V}}$ (or $\\mathcal{V}$ and $\\bar{\\mathcal{V}}$ to generate $T$ ), while during testing, we drop the extra context information to generate $\\bar{\\mathcal{V}}$ and $T$ directly.", + "bbox": [ + 89, + 568, + 482, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mathematically, consider $\\mathcal{F}:(\\mathcal{V},\\bar{\\mathcal{V}},T)\\to \\mathbb{R}$ as the autoregressive function that jointly learns from the learner demonstration $\\nu$ , the expert demonstration $\\tilde{\\nu}$ , and the expert commentary $T$ . The output $\\in \\mathbb{R}$ is the loss that we aim to minimize. We use the same unified $\\mathcal{F}$ for the joint training (detailed in Sec. 3.3) and inference, as follows. We use $\\mathcal{F}'$ when the autoregressive model $\\mathcal{F}$ is used for output token generation, i.e., text or pose tokens. $\\mathcal{F}'$ can take in any input used during training and generate the remaining one—typical for autoregressive models [74].", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expert commentary generation. At inference time, to generate an expert commentary $T$ , we only have a learner demonstration, thus we mask out the expert demonstration and output the expert commentary:", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {T} = \\mathcal {F} _ {t} (\\mathcal {V}) = \\mathcal {F} ^ {\\prime} (\\mathcal {V}, \\varnothing) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 159, + 903, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_t$ denotes using the model to generate commentary.", + "bbox": [ + 511, + 185, + 903, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expert demonstration generation. Next, we obtain the expert demonstration in two forms—retrieving a correct execution (video and pose) and generating a pose sequence. These two output formats offer complementary information to understand the actionable feedback. While a video exemplar is helpful for many learner mistakes, pose generation is useful in the absence of a correct expert demonstration in the retrieval set—allowing our model to generalize its coaching beyond the set of discrete expert videos. We leave expert video generation as a future work.", + "bbox": [ + 511, + 200, + 903, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoting $\\mathcal{F}_r$ and $\\mathcal{F}_g$ as the expert demonstration retrieval and expert pose generation functions, respectively, we have:", + "bbox": [ + 511, + 351, + 903, + 395 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathcal {V}} = \\mathcal {F} _ {r} (\\mathcal {V}) = \\underset {\\forall \\mathcal {V} ^ {\\prime} \\in \\mathcal {E}} {\\operatorname {a r g m i n}} \\mathcal {F} (\\mathcal {V}, \\mathcal {V} ^ {\\prime}, \\hat {T}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 404, + 903, + 430 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP = \\mathcal {F} _ {g} (\\mathcal {V}) = \\mathcal {F} ^ {\\prime} (\\mathcal {V}, \\hat {T}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 434, + 903, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{T}$ is the output from Eq. 1 above, and $\\mathcal{F}_r$ and $\\mathcal{F}_g$ denote using the model to retrieve or generate the expert demonstration, respectively.", + "bbox": [ + 511, + 460, + 903, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the training step $\\mathcal{F}$ uses tuples $\\nu$ , $\\bar{\\nu}$ , $T$ in a unified way, described in Sec. 3.3, whereas, at inference, all the three functions $\\mathcal{F}_t$ , $\\mathcal{F}_r$ and $\\mathcal{F}_g$ use only $\\nu$ .", + "bbox": [ + 511, + 508, + 903, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Forming the expert feedback dataset", + "text_level": 1, + "bbox": [ + 511, + 561, + 830, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn the desired functions, we need to obtain pairs of videos $(\\mathcal{V},\\bar{\\mathcal{V}})$ where there is an error in the demonstration in the first video that is corrected in the second video, along with the expert commentary $T$ about $\\mathcal{V}$ . Ego-Exo4D [29] offers a great starting point for our setup. It contains ego-exo videos, extracted 3D pose sequences, and time-stamped commentary by experts (e.g., professional soccer coaches) on the demonstrations in the video. The experts watched the entire video and stopped each time they saw something to critique or compliment, offering free-form spoken commentary. In total, this led to 117,812 sentences across 221 hours of video, with most videos commented by 2-5 unique experts [29]. See examples in Figure 2 (left) and details in Sec. 3.4 and Supp.", + "bbox": [ + 511, + 583, + 903, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose to automatically augment this data in two ways to enable coaching. First, we seek (pseudo-)annotations of whether a given commentary statement describes a needed improvement or applauds a correct execution. Second, we localize each piece of feedback on a body region, e.g., incorrect hand stretch vs. wrong legs movement. These distinctions are crucial to generate feedback", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13584", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c9480bfaf5667ee33ef0064ff330a334988e4ec43cc52d550c0e5681e4548c23.jpg", + "image_caption": [ + "Figure 2. Overview of the dataset creation. We first summarize the human-provided expert commentary [29] in one sentence using an LLM, and then map it to a body region and correct (green) or incorrect (red) execution label. We then choose incorrect-correct pairs for the same body region to obtain $\\mathcal{C}$ . Finally, we choose pairs with minimum temporal alignment loss to obtain the training data. Best in zoom." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 897, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and show corrections. To this end, we create a weakly-supervised training dataset consisting of tuples $(\\mathcal{V}, T, \\bar{\\mathcal{V}})$ from Ego-Exo4D [29], as follows.", + "bbox": [ + 89, + 362, + 482, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Expert commentary classification and body localization. The commentary in [29] is obtained from voice recordings of the experts converted to text with ASR. Most of the samples contain extra comments like \"oh, I will give this a five out of ten\" and \"that's how I would do it too\". Additionally, as discussed above, they lack positive and negative labels and do not explicitly indicate the body region involved (e.g., legs, arms). Thus, we preprocess the expert commentary for three things—making the commentary concise to extract the improvable feedback, marking which body region the feedback is about, and marking whether the feedback states the need for improvement or not.", + "bbox": [ + 88, + 410, + 482, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since these are all addressable with text reasoning, we use a large language model (LLM) to provide the desired answers. We use Llama3-70B [2], a recent open-sourced model that performs well in current benchmarks. In essence, given a commentary $T$ , the language model $\\mathcal{L}$ responds to the prompt: \"Given an expert's commentary, summarize the feedback into a single sentence and also provide which body regions need improvement or are correctly executed...\" (see Supp.). Formally, this yields $\\mathcal{L}(T) = (T', (b^1, c^j), \\ldots, (b^s, c^s))$ where $T'$ is the concise summary and $b^i$ is a body region like head or arms and $c^i \\in \\{0, 1, 2\\}$ are labels representing needs improvement, correct execution, and no mention, respectively. We group the skeleton joints into six pre-defined body regions $b^i$ (details in Supp.). See red and green boxes in Fig. 2.", + "bbox": [ + 89, + 595, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pairing incorrect and correct executions. Next, we use the above information to mine pairs of incorrect and correct execution in the dataset. Ego-Exo4D also contains metadata about the skill-level $S$ of the demonstrator, broken into four categories (1-4 in increasing expertise) starting from", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "novices who have not performed the activity before to late experts who have performed the same activity, e.g., basketball for $10+$ years. We sample incorrect learner demonstrations from beginners and correct demonstrations from experts from the same activity in Ego-Exo4D, e.g., penalty kicks in soccer or reverse layup in basketball. Even though there can be incorrect executions by experts and correct ones by beginners, errors by experts are likely incomparable to beginners' due to the skill gap. We use the mapping of body regions to find (in)correct executions referring to the same body region. This results in a collection of video pairs with negative and positive feedback about the same body region, e.g., legs in Fig. 2. Formally, the collection $\\mathcal{C}$ is curated as", + "bbox": [ + 511, + 362, + 906, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\left(\\mathcal {V} _ {1}, T, \\mathcal {V} _ {2}\\right) \\mid S _ {1} \\in \\mathcal {S} ^ {n}, S _ {2} \\in \\mathcal {S} ^ {e}, \\exists j s. t. c _ {1} ^ {j} = 0, c _ {2} ^ {j} = 1 \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 584, + 895, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S^n = \\{\\text{novice, early expert}\\}$ and $S^e = \\{\\text{late expert, intermediate expert}\\}$ . Using this matching, we obtain a collection of video pairs of incorrect and correct execution (Fig. 2, yellow box).", + "bbox": [ + 511, + 613, + 903, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal alignment and filtering demonstrations. The commentary annotation process in [29] allows the experts to pause at any instant and provide their feedback. Thus, the clip $\\mathcal{V}_1$ could have the start of a basketball jump shot, whereas the clip $\\mathcal{V}_2$ could also have content before the shot and the follow-through. Thus, we temporally align the learner and the expert video in the collection $\\mathcal{C}$ to ensure we are capturing the same step of a demonstration.", + "bbox": [ + 511, + 674, + 903, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For all pairs of clips in the collection $\\mathcal{C}$ , we compare the two poses $P$ to quantify their alignment (similarity) using Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) [27], a translation-invariant and body shape-invariant measure. Note that the video component of the input (Sec. 4) is complementary, as it does capture the person's overall movement in the space.", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13585", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d03e703cd2fa192b4df0b944e99b3fed49db8bd6efe7c67f23b4572a3e62ce5c.jpg", + "image_caption": [ + "Figure 3. Model overview. We tokenize individual modalities using a modality-specific architecture (top). Once all the modalities are encoded as tokens, we use a large language model to learn expert commentary generation, demonstration retrieval, and pose generation. At inference, the model only takes the learner demonstration video $\\nu$ . See text for details." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 369, + 297 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c2c3a868ade3c1644286079acdb2d28ce8eb640a9af282e0f5943bb611567e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 85, + 632, + 296 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/712ed670a0f8e50099c78ee6cf9d695f486b2e7c3c81027d8ae106bacb0a3300.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 85, + 906, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denote the commentary timestamp as $t_1$ and $t_2$ for the two videos. We first choose a fixed window around $t_1$ , say $[t_1^a, t_1^b]$ such that $t_1^a < t_1 < t_1^b$ . Next, we find the corresponding time window $[t_2^a, t_2^b]$ with $t_2^a < t_2 < t_2^b$ in $\\mathcal{V}_2$ such that the PA-MPJPE error is the minimum. Specifically,", + "bbox": [ + 89, + 357, + 483, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ t _ {2} ^ {a}, t _ {2} ^ {b} ] = \\operatorname * {a r g m i n} _ {[ t _ {2} ^ {i}, t _ {2} ^ {j} ] \\in \\mathcal {V} _ {2}} \\mathrm {P A - M P J P E} \\left(\\mathcal {V} _ {1} [ t _ {1} ^ {a}, t _ {1} ^ {b} ], \\mathcal {V} _ {2} [ t _ {2} ^ {i}, t _ {2} ^ {j} ]\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 441, + 460, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "such that $|t_2^j - t_2^i| = |t_1^b - t_1^a|$ .", + "bbox": [ + 254, + 478, + 462, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We obtain a subset from $\\mathcal{C}$ that contains the aligned video segments. There can be some video pairs where the demonstration can be very different and hence no appropriate match exists. Therefore, we only keep the top-k pairs with maximum alignment for every incorrect execution.", + "bbox": [ + 89, + 507, + 483, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, we obtain a novel dataset $\\mathcal{D}$ that contains pairs of videos where the first video has an incorrect execution (on some body region) and the second video corrects it, thus obtaining the desired tuple $(\\mathcal{V}, T, \\bar{\\mathcal{V}})$ for training and testing. See Fig. 2 (right) and examples in Supp. To ensure fair evaluation, we separately establish a clean gold standard test set, free of potential noise from LLM inference. The test set is manually verified (see Supp for details).", + "bbox": [ + 89, + 583, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Architecture and training design", + "text_level": 1, + "bbox": [ + 89, + 712, + 377, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we discuss the architecture that encodes the videos $\\mathcal{V}$ , poses $P$ , and text $T$ and enables training the auto-regressive function $\\mathcal{F}$ . The overall idea is to encode all the representations into a text embedding space and use the strong capabilities of recent language models to obtain output text and pose tokens. This approach has been recently used in vision-conditioned language models [6, 35, 45, 55] for image and video captioning. This setup also allows for a unified architecture for various input and output combinations, as opposed to different input streams for individual modalities. Fig. 3 shows a schematic diagram of the architecture,", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and each part is explained below.", + "bbox": [ + 511, + 357, + 733, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Encoding video as tokens. Each video $\\mathcal{V}$ is an ego-exo clip pair. $^1$ The input thus allows the model to see both close-up hand-object interactions (more visible in the learner's ego view) as well as full-body poses in the scene context (more visible in the observer's exo view). We use a pretrained video model to extract spatio-temporal features from both videos' frames, followed by a mapper to convert the video features to video tokens. Formally, $\\mathbf{v} = f_{vm}(f_V(\\mathcal{V}))$ where $f_{V}$ is a standard feature extractor (we use InternVideo2 [77]) and $f_{vm}$ is the visual mapper. See Fig. 3 (top left). The mapper is typically a low-parameter model that is trainable, whereas the high-parameter feature extractor is kept frozen. The mapper helps transform the visual mapper to be akin to text tokens—a popular strategy in visual instruction tuning models [35, 45].", + "bbox": [ + 511, + 373, + 906, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Encoding pose sequences as tokens. To encode a pose sequence, we use a series of linear layers and MLP-mixer [25] to convert a single pose $P \\in \\mathbb{R}^{d \\times 3}$ to an embedding that can be discretized using a codebook, i.e., $\\mathbf{p}' = f_P(P)$ . The architecture in [25] also contains a decoder that converts the embeddings back to human poses, i.e., $P = f_{P}^{-1}(\\mathbf{p}')$ , which we use in $\\mathcal{F}_g$ to generate corrected poses. Similar to the above, we use a pose mapper to convert the tokens to embeddings. Formally, $\\mathbf{p} = f_{pm}(f_P(P)) = f_{pm}(\\mathbf{p}')$ is the pose token where $f_{pm}$ is the pose mapper. See Fig. 3 (top center). We concatenate embeddings from every frame to obtain the representation for the whole sequence. Having $f_{pm}$ allows training with fewer parameters and thus, we use $\\mathbf{p}$ for learning $\\mathcal{F}_t$ and $\\mathcal{F}_r$ . However, using $f_{pm}$ for generation $\\mathcal{F}_g$ would require adding an inverting function for $f_{pm}$ . Thus, we directly add $\\mathbf{p}'$ to LLM tokens for pose generation $\\mathcal{F}_g$ . An innovative as", + "bbox": [ + 511, + 601, + 908, + 859 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1The exo view is the one annotated in Ego-Exo4D as having the maximum subject visibility.", + "bbox": [ + 511, + 875, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13586", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pect of our approach is to encode pose as multimodal tokens and train with LLMs, unlike other work regressing the pose parameters from a special pose embedding [20] or using a dedicated pose transformer [49, 82].", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoding text as tokens. Text tokenization is the standard process before inputting a text sentence to the LLM [59, 73], i.e., $\\mathbf{t} = f_t(T)$ where $f_{t}$ is the tokenization function. See Fig. 3 (top right).", + "bbox": [ + 89, + 152, + 482, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-modal sequence prediction. The previous steps yield multimodal tokens $\\mathbf{v}$ , $\\mathbf{p}$ , and $\\mathbf{t}$ for video, pose, and text, respectively. Next, we use the strong sequence prediction capabilities of large language models to obtain the desired output tokens based on the sequence of multi-modal inputs [6, 35, 45, 55].", + "bbox": [ + 89, + 214, + 480, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training $\\mathcal{F}$ for expert commentary generation, we provide a sequence of learner video and pose tokens (v and p) and the corrected pose tokens ( $\\bar{\\mathbf{v}}$ and $\\bar{\\mathbf{p}}$ corresponding to $\\bar{\\nu}$ ). We ask the model to predict the expert commentary token sequence. For the sequence prediction language model $\\mathcal{L}_s$ , we wish to obtain $\\mathbf{t} = \\mathcal{L}_s(\\mathbf{v},\\mathbf{p},\\bar{\\mathbf{v}},\\bar{\\mathbf{p}})$ . Consequently, the training objective is the standard cross-entropy loss, $\\min_{\\theta}\\{-\\log (\\mathbf{t}|\\mathbf{v},\\mathbf{p},\\bar{\\mathbf{v}},\\bar{\\mathbf{p}};\\theta)\\}$ , where $\\theta$ are the parameters of the model. See Fig. 3 (bottom left). For consistency with the training of these language models, the sequence is formatted to be conversational in nature, e.g., \"provide an expert's commentary based on this pose sequence:\". Similarly, for expert demonstration retrieval, we wish to obtain a retrieval candidate $\\bar{\\mathbf{p}} = \\mathcal{L}_s(\\mathbf{v},\\mathbf{p},\\mathbf{t})$ . Likewise, the training objective with parameters $\\gamma$ is $\\min_{\\gamma}\\{-\\log (\\bar{\\mathbf{p}} |\\mathbf{v},\\mathbf{p},\\mathbf{t};\\gamma)\\}$ , which we also use as the relevance score for retrieval during inference. Expert pose generation is trained with the same objective, except we use $\\mathbf{p}'$ instead of $\\mathbf{p}$ (and obtain $\\bar{\\mathbf{p}}'$ ). See Fig. 3 (bottom). Generating pose further requires converting back the pose tokens $\\bar{\\mathbf{p}}'$ to 3D joints using the pose decoder $f_P^{-1}$ .", + "bbox": [ + 91, + 305, + 483, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At inference time, as introduced in Sec. 3.1, we only use $\\mathcal{V}$ as the input: $\\mathcal{F}_t$ drops the expert demonstration and predicts the expert commentary $\\hat{T}$ . We use this predicted commentary as input for $\\mathcal{F}_r$ and $\\mathcal{F}_g$ .", + "bbox": [ + 89, + 623, + 482, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 696, + 303, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and statistics. Ego-Exo4D [29] contains 5,035 videos of participants doing activities across eight scenarios. We focus on three physical scenarios—basketball, soccer, and rock climbing—though our model is in principle generalizable to other physical skills, without any task-specific design, unlike [63, 78]. We use physical scenarios since the coaching feedback is groundable in the body regions, as opposed to procedural tasks like cooking, where the suggestions can be alternate ingredients or steps that are not visually present. Details and statistics are in Supp. Following the dataset creation process outlined in Sec. 3.2, we obtain a dataset of 25,505 training and 1,272 testing tuples", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of $(\\mathcal{V}, T, \\bar{\\mathcal{V}})$ . We choose $k = 5$ (train) and $k = 1$ (test) for min-k choice of correct demonstrations per incorrect execution. To reiterate, we manually examine the test set for correct commentary summary and body region labeling.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network architecture. The video model $f_{V}$ is an InternVideo2 [77] encoder that provides strong visual representations. The pose encoder $f_{P}$ and decoder $f_{P}^{-1}$ are adapted from PCT [25], which learns compositional tokens from human poses. Since the original training in [25] uses 2D poses, we adapt it to use 3-dimensions and retrain on the human poses in [29]. Both the visual and pose mappers are low-parameter MLP layers [44]. Finally, we use Llama 3-8B [2] as the LLM multimodal encoder for sequence prediction. We finetune $\\mathcal{F}_t$ and $\\mathcal{F}_r$ for 5 epochs with a learning rate of $5 \\times 10^{-5}$ . Next, we modify the token dimension in $\\mathcal{F}_g$ to accommodate pose tokens $\\mathbf{p}'$ , and hence, we fine-tune $\\mathcal{L}_s$ when learning $\\mathcal{F}_g$ with a learning rate of $5 \\times 10^{-6}$ for 5 epochs. Video $(f_{V})$ and pose models $(f_{P}, f_{P}^{-1})$ are kept frozen. All the models are trained on 8 V100 32GB GPUs.", + "bbox": [ + 511, + 152, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and results", + "text_level": 1, + "bbox": [ + 511, + 393, + 741, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first discuss the baselines and ablations, followed by the evaluation setup and results for the three outputs—expert commentary generation $(\\mathcal{F}_t)$ , expert demonstration retrieval $(\\mathcal{F}_r)$ , and pose demonstration generation $(\\mathcal{F}_g)$ . We also show qualitative examples and discuss the limitations.", + "bbox": [ + 511, + 420, + 903, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare with the following baselines:", + "bbox": [ + 511, + 496, + 864, + 511 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- InternVideo2-NN, InternVideo2-FT [77]: Given a query video $\\mathcal{V}$ , the nearest neighbor baseline (NN) finds the most similar video by InternVideo2 [77] feature similarity in the training data and returns the corresponding commentary or demonstration, for all tasks. The FT version finetunes the model to contrastively match the learner demonstration with the expert commentary/demonstration.", + "- VideoChat2 [36, 37], LLaVA [45]: In these methods, we prompt SOTA video and image captioning models to generate commentary for an input demonstration. We use the log-likelihood loss to find the retrieved expert demonstration. These baselines evaluate if existing video captioning methods can provide expert feedback or retrieve expert demonstrations. We use the ego and \"best exo\" frames, same as for our method. Note that neither of these baselines are applicable for pose generation $\\mathcal{F}_g$ .", + "- LLaVA-FT [45], LLaVA-FT w/ pose [45]: These baselines are based on the SOTA visual-language method LLaVA [45] but trained on our dataset with the same text model, i.e., Llama 3-8B [2], for an apples-to-apples comparison. The \"w/ pose\" variant also takes the 3D pose coordinates in text form as input.", + "- PoseScript [16], PoseFix [17]: These two works enable pose-to-text and text-to-pose reasoning. The text generated or used for pose generation contains detail about the" + ], + "bbox": [ + 508, + 512, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13587", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/63e969fd713711291835bef2c4260157be928605317fc24c03dfdc29c2a9508a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCommentary Gen.Demo Ret.Pose Gen.
B@4MR-LRmedR↓P↓
InternVideo2-NN [77]42.146.949.313.5198161
InternVideo2-FT [77]42.947.650.014.1190157
VideoChat2 [37]27.844.341.914.9183—
LLaVA [45]28.544.144.215.0183—
LLaVA-FT [45]43.548.551.517.8177—
LLaVA-FT w/ pose [45]43.648.551.718.0172150
PoseScript/Fix [16, 17]24.144.546.315.9182182
ExpertAF44.949.654.619.1158135
ExpertAF w/o video44.649.454.218.7161139
ExpertAF w/o pose44.249.354.218.6163—
w/o alignment42.048.651.516.9180153
w/ global pose43.047.952.616.5184150
ExpertAF w/ full-sup45.850.955.722.5146131
", + "bbox": [ + 106, + 93, + 607, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/157726add4d6f8b06b60e90cf3b0ed9405c16461026d5e6be8d4d15c5aa473c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 89, + 903, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Results on automatic metrics (left) and human evaluation (right). We break down results for the three outputs—expert commentary generation, expert demonstration retrieval, and expert pose generation. Our method outperforms all baselines and prior work on all tasks. The last row \"w/ full-sup\" uses privileged input (the demo video $\\bar{\\nu}$ ) at inference. (B@4: BLEU-4, M: Meteor, R-L: ROUGE-L F1, R: recall@50, medR: median rank, P: PA-MPJPE). For all metrics higher is better, except medR and PA-MPJPE ( $\\downarrow$ ). Our method is also rated higher by human raters on a Likert scale (min:1, max:4), compared to all the other methods (right). See text for details.", + "bbox": [ + 88, + 316, + 906, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "location of body parts, e.g., the hands are raised, as opposed to expert commentary. Hence, they let us evaluate if pose description is adequate for expert feedback. For pose generation, we evaluate if providing an expert commentary helps generate the desired expert demonstration. Both methods use SMPL [47] pose and hence we convert 3D pose to SMPL and vice versa [39], as needed.", + "bbox": [ + 98, + 393, + 480, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations. In addition to the strong baselines, we also compare the performance against ablations and variants of our design choices. ExpertAF w/o pose and ExpertAF w/o video evaluate the performance when only one modality is used. ExpertAF w/o alignment quantifies the need for temporal alignment (Sec. 3.2), while ExpertAF w/ global pose evaluates if just providing the model with one correct pose per activity chosen based on expert commentary is enough (vs. finding the closest correct pose sequence). ExpertAF w/ full-sup is a stronger variant with privileged input for inference— $\\mathcal{F}_t$ uses $\\mathcal{V}$ , $\\bar{\\mathcal{V}}$ to predict $T$ and $\\mathcal{F}_r$ and $\\mathcal{F}_p$ uses both $\\mathcal{V}$ and $T$ . See Supp. for additional ablations of the choice of LLM $\\mathcal{L}_s$ , contribution of ego and exo videos, and joint training.", + "bbox": [ + 89, + 500, + 482, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expert commentary generation. To evaluate the text commentary, we use standard metrics BLEU-4 [60], METEOR [7] and ROUGE-L F1 [65], following prior work in evaluating text generation [14, 58] (higher is better).", + "bbox": [ + 89, + 715, + 482, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 (left) shows the results. Our method outperforms all methods on all captioning metrics, with gains up to $3\\%$ over the best baseline trained on the same dataset. All the gains are statistically significant using paired t-test with $p < 0.05$ . Fig. 4 (top) shows commentary generation by various methods. Firstly, SOTA methods LLaVA [45] and VideoChat2 [37], being captioning methods, yield a verbose description of the activity, rather than a critique.", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-tuning LLaVA with our method improves the performance, but still it misses our use of expert demonstrations and pose sequences. Even using pose information as text only marginally improves \"LLaVA-FT w/ pose\", showing the effectiveness of our explicit pose encoding. Moreover, PoseScript [16] generates pose descriptions like \"head is turned to the left\" which is inadequate for feedback. Our results over the baselines show the advantage of temporal alignment and a diverse expert demonstration set.", + "bbox": [ + 511, + 393, + 906, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our stronger variant \"w/ full-sup\" achieves an even better result by taking $\\mathcal{V}$ and $\\bar{\\mathcal{V}}$ as inputs during inference. It is useful in cases where a learner has access to the expert demonstration but wants actionable feedback in languageform. Finally, Fig. 4 (bottom row) shows a failure case of commentary generation, showcasing the difficulty in pinpointing the exact mistake.", + "bbox": [ + 511, + 532, + 908, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expert demonstration retrieval. For retrieval, we use standard retrieval metrics: recall@k and median rank. We set k=50 to account for the fact that multiple expert demonstrations could suitably correct the given learner demonstration. The retrieval set contains 1,272 samples, the same as the test set size. Higher recall and lower ranks are better.", + "bbox": [ + 511, + 640, + 908, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 (middle) shows the results. Our method improves the median rank by 14 positions compared to the best baseline and a significantly better recall@50. Our method and the baselines show a similar trend as with the expert commentary generation above, the stronger baseline w/ \"full-sup\" ( $>4\\%$ better recall and 26 rank improvement) being crucial in cases where a learner has the commentary but wants a video exemplar to learn from.", + "bbox": [ + 511, + 732, + 908, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expert pose generation. Tab. 1 (right column) shows the expert pose generation results, with error measured by PA-MPJPE [27] (mm). Our novel pose generation method", + "bbox": [ + 511, + 854, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13588", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69cdc67194d5ff823d450d701bf899b2ddf19993421584b8a7a9b21f62352b24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 90, + 215, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground truth", + "text_level": 1, + "bbox": [ + 245, + 90, + 325, + 103 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The climber is exhausted and struggling to make a move, with poor hand and arm positioning.", + "bbox": [ + 225, + 116, + 344, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours", + "text_level": 1, + "bbox": [ + 403, + 90, + 437, + 103 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The climber is struggling to maintain the position and is unable to make the next move.", + "bbox": [ + 359, + 116, + 478, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LLaVA-FT", + "text_level": 1, + "bbox": [ + 522, + 90, + 584, + 102 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The climber is wasting energy by pulling away, which is slowing down reaching the next handhold", + "bbox": [ + 491, + 116, + 616, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VideoChat2", + "text_level": 1, + "bbox": [ + 651, + 90, + 725, + 103 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The person can improve by practicing their technique, warming up and stretching.", + "bbox": [ + 629, + 116, + 746, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PoseScript", + "text_level": 1, + "bbox": [ + 787, + 90, + 857, + 103 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The person is standing on their left leg with the right leg lifted up and bent at the knee. The torso is leaning to the left. The head is turned to the left.", + "bbox": [ + 761, + 107, + 883, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e45b0af9cb986232e56d87bf322ba561c167774408777408ef1a405e3ea98f90.jpg", + "image_caption": [ + "Learner demonstration" + ], + "image_footnote": [], + "bbox": [ + 99, + 214, + 243, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91bbee65fbc312ff8144dc85b7f6bbca3b34431f76fd0d93c3c533035c6fb734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 292, + 243, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47b2c8773f2af44cb3acb144c8785a4f23033f6d73e01871f269410cf0265e59.jpg", + "image_caption": [ + "Failure case", + "Figure 4. Qualitative results. (Top) Comparison of expert commentary generated by various baselines. (Second and third row) Examples of expert commentary generation, demonstration retrieval, and pose generation by our method. Notice the expert demonstration and pose generation corrects the mistake pointed out in the commentary, i.e., one hand is used to throw and the other to guide, and body position is improved to control better control and power. Colored text and marks (red for mistake, green for correction) are shown only for visualization. (Bottom) Failure cases. See Supp. for video results." + ], + "image_footnote": [], + "bbox": [ + 94, + 392, + 243, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Expert comment. - Ours", + "text_level": 1, + "bbox": [ + 261, + 203, + 408, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The shooter's stance is correct, but they need to throw the ball with one hand and guide it with the other.", + "bbox": [ + 256, + 231, + 411, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The player's stance is too straight, which prevents them from generating the required power and control on the ball.", + "bbox": [ + 254, + 306, + 415, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The shooter's left elbow is bent and his wrist is not fully extended, affecting the ball's rotation and accuracy.", + "bbox": [ + 256, + 392, + 413, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground truth commentary", + "text_level": 1, + "bbox": [ + 421, + 203, + 578, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The shooter is using two hands to shoot the ball, instead of using one shooting hand and arm with the other at a 90-degree angle.", + "bbox": [ + 419, + 226, + 580, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The player's foot position and body lean are affecting the quality of their ball contact, leading to a loss of power and control.", + "bbox": [ + 419, + 301, + 583, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "He does a good job using his left hand on the left side, but needs to improve his posture and weight distribution.", + "bbox": [ + 421, + 392, + 581, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d3a9e59a065beb6e33d6a2f25173115e85cd9a1cadf8fa06833e863f479c951.jpg", + "image_caption": [ + "Expert demo. retrieval" + ], + "image_footnote": [], + "bbox": [ + 589, + 214, + 736, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1565ada75121f6b2de9c9aa1170d1b34c52cad644ca1380f5f6f78771bdcffc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 291, + 736, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2129b9d6136ad20fd3630a97e35df46c6750cb1ae7d1aa44ecaf0f3d371a4a77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 376, + 736, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01db810c9d6d4ad3df6c95bb51a86d8a84de8f7480d187b60710c63e9f860d0e.jpg", + "image_caption": [ + "Pose generation" + ], + "image_footnote": [], + "bbox": [ + 745, + 215, + 893, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4720fc1e800611ef2157eec73a90bbfda8e653bbe71df5edced359ff33f9bd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 291, + 893, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e50ab803604f6adbb398f2cbdfe71aefb09dfff0d31a964c3fa47e28de3ff05c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 376, + 893, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performs better than learning 3D position using text in \"LLaVA-FT w/ pose\". Furthermore, PoseFix [17] is designed to generate SMPL parameters based on a coarse modification text description. Our use of generation to output poses (vs. retrieval) is advantageous when the candidate retrieval set does not contain the correct demonstration; see Supp. See Fig. 4 and Supp. for qualitative examples including failure cases. The 3D pose sequences in the dataset are often auto-generated and have some noisy samples, discussed in Supp. Across the three tasks, our method is superior to its ablations.", + "bbox": [ + 89, + 544, + 482, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human evaluation. For all the tasks, we also solicit human evaluation (Tab. 1 right) to rate the quality of the generated text descriptions. Five raters (per scenario) uninvolved with this project rate the quality of each generated output by scoring on a Likert scale from 1 to 4 (higher is better); details in Supp. We ensure that the raters have a basic knowledge of the scenario they are rating (basketball, soccer, rock climbing).", + "bbox": [ + 89, + 714, + 482, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Across all three tasks, human raters score our method the highest. A significant gap of up to 0.7 over the possible range $(1 - 4)$ , as well as gains up to $3\\times$ , showcase ExpertAF's expert feedback quality. That said, there is nat", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "urally room for improvement on this new task. We find that while humans say our model excels on cases where the feedback is visually groundable, e.g. incorrect hand angle, it tends to fall short when the feedback is not directly visible, e.g. the climber looks fatigued. The human evaluation complements the automatic metrics above, overcoming the limitation that a \"ground truth\" commentary or a visual demonstration may capture only one of multiple possible errors in the learner's demonstration [23, 68] (e.g., for a basketball shot, both the hand placement and the jump could be incorrect, but only one may be mentioned in the ground truth).", + "bbox": [ + 511, + 544, + 906, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 734, + 633, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a novel task and method to generate expert commentary and demonstrations from a learner's video. We develop a weakly-supervised training approach and benchmark for this problem. Our novel method fusing multimodal inputs from learner and expert demonstrations together with expert commentary results in state-of-the-art performance in actionable feedback, and lays the groundwork for accessible, affordable, and actionable AI coaching applications in the future.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13589", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. UT Austin is supported in part by the IFML NSF AI Institute. Thanks to Fu-Jen Chu, Jing Huang and Xitong Yang for help with the Exo-Ego4D [29] pose extraction pipeline.", + "bbox": [ + 89, + 90, + 485, + 148 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 174, + 187, + 189 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yazan Abu Farha, Alexander Richard, and Juergen Gall. When will you do what?-anticipating temporal occurrences of activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5343-5352, 2018. 2", + "[2] AI@Meta. Llama 3 model card, 2024. 4, 6, 1", + "[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23066-23078, 2023. 2", + "[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. What you say is what you show: Visual narration detection in instructional videos. arXiv preprint arXiv:2301.02307, 2023. 2", + "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[6] Kumar Ashutosh, Zihui Xue, Tushar Nagarajan, and Kristen Grauman. Detours for navigating instructional videos. In CVPR, 2024. 1, 2, 5, 6", + "[7] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7", + "[8] Tobias Baumgartner and Stefanie Klatt. Monocular 3d human pose estimation for sports broadcasts using partial sports field registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5108-5117, 2023. 2", + "[9] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the IEEE international conference on computer vision, pages 2177-2185, 2017. 1, 2", + "[10] Gedas Bertasius, Aaron Chan, and Jianbo Shi. Egocentric basketball motion planning from a single first-person image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[11] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 2", + "[12] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya" + ], + "bbox": [ + 93, + 199, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ramesh. Video generation models as world simulators, 2024. 2", + "[13] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 2", + "[14] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. arXiv preprint arXiv:2305.18500, 2023. 7", + "[15] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos. International Journal of Computer Vision, 129:2846-2864, 2021. 2", + "[16] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022. 2, 6, 7", + "[17] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2, 6, 7, 8", + "[18] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 319-335. Springer, 2022. 2", + "[19] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Posegpt: Chatting about 3d human pose. arXiv preprint arXiv:2311.18836, 2023. 2", + "[20] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. In CVPR, 2024. 2, 6", + "[21] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence, 43(11):4021-4036, 2020. 2", + "[22] Jiyang Gao, Zhenheng Yang, and Ram Nevatia. Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818, 2017. 2", + "[23] Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal, Pawan Sasanka Ammanamanchi, Anuoluwapo Aremu, Antoine Bosselut, Khyathi Raghavi Chandu, Miruna-Adriana Clinciu, Dipanjan Das, Kaustubh Dhole, Wanyu Du, Esin Durmus, Ondrej Dušek, Chris Chinenye Emezue, Varun Gangal, Cristina Garbacea, Tatsunori Hashimoto, Yufang Hou, Yacine Jernite, Harsh Jhamtani, Yangfeng Ji, Shailza Jolly, Mhir Kale, Dhruv Kumar, Faisal Ladhak, Aman Madaan, Mounica Maddela, Khyati Mahajan, Saad Mahamood, Bodhisattwa Prasad Majumder, Pedro Henrique Martins, Angelina McMillan-Major, Simon Mille, Emiel van Miltenburg, Moin Nadeem, Shashi Narayan, Vitaly Nikolaev, Andre Niyongabo Rubungo, Salomey Osei, Ankur" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13590", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Parikh, Laura Perez-Beltrachini, Niranjan Ramesh Rao, Vikas Raunak, Juan Diego Rodriguez, Sashank Santhanam, João Sedoc, Thibault Sellam, Samira Shaikh, Anastasia Shimorina, Marco Antonio Sobrevilla Cabezudo, Hendrik Strobelt, Nishant Subramani, Wei Xu, Diyi Yang, Akhila Yerukola, and Jiawei Zhou. The GEM benchmark: Natural language generation, its evaluation and metrics. In Proceedings of the 1st Workshop on Natural Language Generation, Evaluation, and Metrics (GEM 2021), pages 96-120, Online, 2021. Association for Computational Linguistics. 8", + "[24] Adelaide Genay, Anatole Lécuyer, and Martin Hachet. Being an avatar \"for real\": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 28(12):5071-5090, 2021. 2", + "[25] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 660-671, 2023. 2, 5, 6", + "[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 2", + "[27] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 4, 7", + "[28] Andrew S Gordon. Automated video assessment of human performance. In Proceedings of AI-ED, page 10, 1995. 1, 2", + "[29] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In CVPR, 2024. 1, 2, 3, 4, 6, 9", + "[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2", + "[31] William Hoff and Tyrone Vincent. Analysis of head pose accuracy in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 6(4):319-334, 2000. 2", + "[32] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2", + "[33] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2", + "[34] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7718-7727, 2019. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 5, 6", + "[36] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6", + "[37] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark, 2023. 6, 7", + "[38] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 2", + "[39] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yukun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv preprint arXiv:2304.00553, 2023. 7", + "[40] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In NeurIPS, 2022. 2", + "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3", + "[42] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2", + "[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9522-9531, 2021. 3", + "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6", + "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 5, 6, 7", + "[46] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2", + "[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph., 34(6), 2015. 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13591", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Sabine Losch, Eva Traut-Mattausch, Maximilian D Muhlberger, and Eva Jonas. Comparing the effectiveness of individual coaching, self-coaching, and group training: How leadership makes the difference. Frontiers in psychology, 7: 175595, 2016. 1", + "[49] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 6", + "[50] Priyanka Mandikal and Kristen Grauman. Dexvip: Learning dexterous grasping with human hand pose priors from video. In Conference on Robot Learning, pages 651-661. PMLR, 2022. 2", + "[51] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action forecasting@ ego4d challenge 2022. arXiv preprint arXiv:2207.12080, 2022. 2", + "[52] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations, 2023. 2", + "[53] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2", + "[54] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2", + "[55] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. In CVPR, 2024. 5, 6", + "[56] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 1", + "[57] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. arXiv preprint arXiv:2207.09666, 2022. 2", + "[58] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. In European Conference on Computer Vision, pages 167–184. Springer, 2022. 7", + "[59] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fo" + ], + "bbox": [ + 89, + 92, + 486, + 900 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mély, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O’Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael. Pokorny. Michelle Pokrass. Vitchyr H. Pong Tolly Powell Alethea Power Boris Power Elizabeth Proehl Raul Puri Alec Radford Jack Rae Aditya Ramesh Cameron Raymond Francis Real Kendra Rimbach Carl Ross Bob Rotsted Henri Roussez Nick Ryder Mario Saltarelli Ted Sanders Shibani Santurkar Girish Sastry. Heather Schmidt David Schnurr John Schulman Daniel Selsam. Kyla Sheppard Toki Sherbakov Jessica Shieh Sarah Shoker Pranav Shyam Szymon Sidor Eric Sigler Maddie Simens Jordan Sitkin Katarina Slama Ian Sohl Benjamin Sokolowsky Yang Song Natalie Staudacher Felipe Petroski Such Natalie Summers Ilya Sutskever Jie Tang. Nikolas Tezak Madeleine B. Thompson Phil Tillet. Amin Tootoonchian Elizabeth Tseng Preston Tuggle Nick Turley Jerry Tworek Juan Felipe Cerón Uribe Andrea Vallone Arun Vijayvergiya Chelsea Voss Carroll Wainwright Justin Jay Wang Alvin Wang Ben Wang Jonathan Ward Jason Wei CJ Weinmann Akila Welihinda Peter Welin", + "bbox": [ + 545, + 92, + 906, + 897 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13592", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "der, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report, 2024. 6", + "[60] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7", + "[61] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[62] Paritosh Parmar, Jaiden Reddy, and Brendan Morris. Piano skills assessment. In 2021 IEEE 23rd international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2021. 3", + "[63] Paritosh Parmar, Amol Gharat, and Helge Rhodin. Domain knowledge-informed self-supervised representations for workout form assessment. In European Conference on Computer Vision, pages 105-123. Springer, 2022. 2, 6", + "[64] Hamed Pirsiavash, Carl Vondrick, and Antonio Torralba. Assessing the quality of actions. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 556-571. Springer, 2014. 2", + "[65] Lin CY Rouge. A package for automatic evaluation of summaries. In Proceedings of Workshop on Text Summarization of ACL, Spain, 2004. 7", + "[66] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2", + "[67] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2", + "[68] Chenhui Shen, Liying Cheng, Xuan-Phi Nguyen, Yang You, and Lidong Bing. Large language models are not yet human-level evaluators for abstractive summarization. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 4215–4233, 2023. 8", + "[69] Sumedh Sontakke, Jesse Zhang, Séb Arnold, Karl Pertsch, Erdem Biryik, Dorsa Sadigh, Chelsea Finn, and Laurent Itti. Roboclip: One demonstration is enough to learn robot policies. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[70] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2", + "[71] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9839-9848, 2020. 2", + "[72] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4790-4799, 2023. 2", + "[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 6", + "[74] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[75] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2", + "[76] Tianyu Wang, Yijie Wang, and Mian Li. Towards accurate and interpretable surgical skill assessment: A video-based method incorporating recognized surgical gestures and skill levels. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 668-678. Springer, 2020. 3", + "[77] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 5, 6, 7", + "[78] Chengming Xu, Yanwei Fu, Bing Zhang, Zitian Chen, YuGang Jiang, and Xiangyang Xue. Learning to score figure skating sport videos. IEEE Transactions on Circuits and Systems for Video Technology, 30(12):4578-4590, 2020. 1, 2, 6", + "[79] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2", + "[80] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 2", + "[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2", + "[82] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 6", + "[83] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129:703-718, 2021. 2", + "[84] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 2", + "[85] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2", + "[86] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2", + "[87] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2" + ], + "bbox": [ + 91, + 92, + 482, + 642 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13594", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_model.json b/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8902cfe0af87474ad23cafcf5ccfc40e8b263d3c --- /dev/null +++ b/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_model.json @@ -0,0 +1,2888 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.238, + 0.131, + 0.761, + 0.154 + ], + "angle": 0, + "content": "ExpertAF: Expert Actionable Feedback from Video" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.18, + 0.877, + 0.219 + ], + "angle": 0, + "content": "Kumar Ashutosh\\(^{1,2}\\), Tushar Nagarajan\\(^{2}\\), Georgios Pavlakos\\(^{1}\\), Kris Kitani\\(^{2,3}\\), Kristen Grauman\\(^{1,2}\\) \n\\(^{1}\\) University of Texas at Austin, \\(^{2}\\) FAIR Meta, \\(^{3}\\) Carnegie Mellon University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.252, + 0.327, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.485, + 0.617 + ], + "angle": 0, + "content": "Feedback is essential for learning a new skill or improving one's current skill-level. However, current methods for skill-assessment from video only provide scores or compare demonstrations, leaving the burden of knowing what to do differently on the user. We introduce a novel method to generate actionable feedback (AF) from video of a person doing a physical activity, such as basketball or soccer. Our method takes a video demonstration and its accompanying 3D body pose and generates (1) free-form expert commentary describing what the person is doing well and what they could improve, and (2) a visual expert demonstration that incorporates the required corrections. We show how to leverage Ego-Exo4D's [29] videos of skilled activity and expert commentary together with a strong language model to create a weakly-supervised training dataset for this task, and we devise a multimodal video-language model to infer coaching feedback. Our method is able to reason across multi-modal input combinations to output full-spectrum, actionable coaching—expert commentary, expert video retrieval, and expert pose generation—outperforming strong vision-language models on both established metrics and human preference studies." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.643, + 0.222, + 0.658 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.804 + ], + "angle": 0, + "content": "An abundance of instructional \"how-to\" videos on the internet enables skill-learning by observing expert demonstrations. Instructional videos cover many skills one might want to learn—cooking, DIY, sports, crafts, and hobbies. Recent work leverages such videos to assist human [5, 6, 53] and robot [56] skill learning. Acquiring skills from video is especially appealing for equitable learning: access to video is generally less costly and more widely available than face-to-face access to an expert coach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Despite the myriad of videos and tutorials, people prefer to learn in a feedback loop: getting to know their mistakes, finding the improvements they need to do and correcting those. For that reason, coaching-based iterative training is more effective than self-learning [48]. A good coach has" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.25, + 0.689, + 0.262 + ], + "angle": 0, + "content": "Learner demonstration (input)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.264, + 0.728, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.365, + 0.734, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.734, + 0.25, + 0.903, + 0.262 + ], + "angle": 0, + "content": "Expert demonstration (output)" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.264, + 0.923, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.363, + 0.896, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.439, + 0.682, + 0.452 + ], + "angle": 0, + "content": "Expert commentary (output)" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.458, + 0.903, + 0.485 + ], + "angle": 0, + "content": "The player is taking big steps to control the ball, but lacks body control, and needs to slow down to maintain a better control." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 1. An example of expert feedback. When a player is dribbling the ball fast, they tend to lose control (top left). Our proposed method provides an expert commentary to the learner suggesting improvements (bottom). The method also provides an expert demonstration that shows the desired correction, where the player is maintaining smaller steps and body control (top right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.685 + ], + "angle": 0, + "content": "three components: finding mistakes, providing verbal corrections and finally, showing visual demonstrations. For example, suppose a person learning to dribble a soccer ball takes big steps that decrease body control. The coach should identify this mistake, provide verbal feedback, and show the correct dribbling technique. See Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.778 + ], + "angle": 0, + "content": "At present, a person watching a how-to video has to identify mistakes themselves and attempt to correct them. Pinpointing the exact mistake is itself challenging for beginners—particularly when learning physical skills, like sports and dance, where subtle movements and positioning of the body are critical to success." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "What role could AI play in overcoming this gap? Current work on proficiency understanding from video focuses on scoring a demonstration [9, 28, 29, 78], thus addressing only one component of coaching. There is no method that can provide actionable verbal and visual feedback on a learner's demonstration. Doing so is technically challenging since it requires the method to (a) understand the user's activity, (b) detect mistakes in the execution by comparing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.436, + 0.901 + ], + "angle": 0, + "content": "Project page: https://vision.cs.utexas.edu/projects/ExpertAF/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.155 + ], + "angle": 0, + "content": "the user's activity with correct examples and finally, (c) suggest edits with respect to the better way of doing the activity. In short, the output should be coach-like—specific actionable feedback that will improve a learner's proficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.162, + 0.483, + 0.284 + ], + "angle": 0, + "content": "We propose ExpertAF: a novel method to provide expert actionable feedback on a user's physical activity. ExpertAF's actionable feedback has two forms: language-based commentary and visual demonstration. The freeform expert commentary output describes the mistake and what should be corrected, e.g., \"take smaller steps and slow down to maintain control\", while the expert visual demonstration output shows the correct way of doing it." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.485, + 0.549 + ], + "angle": 0, + "content": "Both components introduce unique technical challenges. For the former, the task of generating free-form text feedback is distinct from standard image and video captioning, as the model needs to identify the mistake and actions that would correct it—rather than simply describe the observed pose [16] or activity [32, 38, 57, 75, 81]. Similarly, for the latter, the task of retrieving (or generating) a demonstration that corrects a specific mistake is distinct from similarity-based retrieval [3, 40, 54, 77, 79] or open-ended generation [12, 19]. We hypothesize—and experimentally validate—that personalized, video-conditioned coaching is more accurate than simply returning a global expert execution because there are many right ways of doing an activity, e.g., penalty shots in soccer, and hence different ways to fix what is wrong in a given execution. Overall, the model must understand both what is being done and what actionable tweaks would make it better." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.756 + ], + "angle": 0, + "content": "We are the first to address video-based coaching with actionable feedback. We develop the first model to generate free-form text feedback on a user's activity (captured in video and 3D pose sequences) as well as a video demonstrating how to improve, by building on a recent vision-language model [45]. There is currently no dataset for this challenging task. Therefore, we augment Ego-Exo4D [29], an existing dataset with video, commentary, and poses, and show how to use a strong large language model to create a weakly-supervised training set consisting of paired learner-expert demonstrations, along with expert textual commentary that relates the two. We also obtain a gold-standard manually labeled test set for rigorous evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We validate ExpertAF on three diverse physical scenarios: soccer, basketball, and rock climbing. Our results show a significant improvement over strong baselines, establishing the first method to provide actionable feedback, including a novel technique for expert pose generation. Alongside consistent gains in quantitative metrics against the ground truth, we also show our model's promise via direct human evaluation, where ExpertAF outperforms off-the-shelf video models by as much as \\(3 \\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.652, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.401 + ], + "angle": 0, + "content": "Skill learning from videos. Instructional video datasets like HowTo100M [53], COIN [70], and CrossTask [87] enable procedural understanding through step recognition [5, 42, 85], procedure planning [11, 13, 84], task-graph discovery [5, 18, 29, 85], action anticipation [1, 21, 22, 26, 51], and alignment detection [4, 30]. Capitalizing on their instructional nature, recent work learns robot policies from these videos [50, 69], uses external knowledge-bases like WikiHow to ground the steps of a procedural task [42, 52, 85], and explores new ways to navigate between multiple demonstrations [6]. Despite their scale, how-to videos in these datasets [53, 70, 87] are often created by domain experts and hence lack mistakes or suboptimal executions. In contrast, Ego-Exo4D [29] contains multiple executions of various common scenarios by people with varied skill levels—from novice to late expert. Moreover, it contains experts' feedback on those actions. We show how to augment this data via large language models to support our new idea for video-based coaching." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.905, + 0.704 + ], + "angle": 0, + "content": "3D body pose for activity understanding. 3D body pose is crucial for human activity understanding, capturing a person's stance and motion. Supported by valuable datasets like Humans3.6M [33] and NTU RGB-D [46, 66], recent methods improve body pose understanding [25, 34, 67, 72, 83, 86]. Beyond action classification, pose is also crucial for markerless motion capture of sports [8, 15] and interacting in augmented reality [24, 31]. While our ExpertAF leverages pose as an important signal of human activity, unlike prior work we interface video, pose, and free-form text to generate expert feedback and an expert demonstration video. Our work is also different from generating [20] or modifying [17] body pose from text (e.g., \"raise your left arm\"), as our task is to understand a potentially incorrect pose and provide feedback. Understanding these minor differences between incorrect and correct poses offers unique challenges that are addressed by our proposed learning scheme. Our work is orthogonal to methods that improve pose estimation itself; any future improvements on that front would only benefit our approach." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Skill assessment and coaching. Prior work explores skill assessment for a variety of tasks, particularly sports. Most methods pose skill-assessment as a score-prediction task, i.e., for skating [78], gymnastics [28], basketball [9], or multiple sports [61]. Since absolute scoring can be ambiguous and error-prone, recent work [71] explores uncertainty-aware score distributions. Instead of explicitly scoring a demonstration, other work determines which of two basketball players are better [9], uses group-aware contrastive regression to learn the relative quality of demonstrations [80], or generates a full ideal trajectory from the first frame [10]. Fitness-AQA [63] and Action Quality [64] provide outputs like knees inward error, shallow squat error," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "or an arrow highlight for more localized feedback. There is also research on skill assessment in non-sports domains, like surgery [43, 76] and piano [62]. All the prior work assumes a fixed taxonomy of errors, and the taxonomy is designed separately for each activity. Furthermore, unlike our work, none of the prior work provides feedback akin to a personal coach—which requires not only telling what is wrong, but also expressing how to correct the mistake." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.224, + 0.182, + 0.239 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.294 + ], + "angle": 0, + "content": "We introduce the problem statement in Sec. 3.1, the dataset creation strategy in Sec. 3.2, the training design in Sec. 3.3, and implementation details in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.305, + 0.273, + 0.319 + ], + "angle": 0, + "content": "3.1. Problem statement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Consider a dataset \\(\\mathcal{E} = \\{(\\mathcal{V}, T, \\bar{\\mathcal{V}})\\}\\) where each \\(\\mathcal{V}_i\\) is a video demonstration, and \\(T_i\\) is a free-form text commentary critiquing the activity in the video, e.g., \"take smaller steps and slow down to maintain control.\" \\(\\bar{\\mathcal{V}}\\) is a related video demonstration but without the error mentioned in \\(T\\), e.g., the player takes smaller steps with better control. See Fig. 1. Each demonstration \\(\\mathcal{V} = \\{V, P, S\\}\\) consists of three parts: \\(V\\) is the RGB video, and \\(P\\) is the 3D pose sequence of the participant with skill-level \\(S\\) in the video. The pose representation \\(P \\in \\mathbb{R}^{n \\times d \\times 3}\\) contains \\(n\\) frames of 3D positions for \\(d = 17\\) body joints, consistent with MS-COCO [41]. \\(S\\) is the participant's skill-level for an activity, i.e., novice, early expert, intermediate expert, or late expert, and will be used later in creating the dataset. We assume there is one active person doing a physical activity in any given video who is the subject of ExpertAF." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.749 + ], + "angle": 0, + "content": "The goal in this work is to provide feedback on a given learner's video containing a physical scenario. The feedback is an expert demonstration \\(\\bar{\\mathcal{V}}\\) and an expert commentary \\(T\\), tailored to \\(\\mathcal{V}\\). Formally, we want to learn a mapping \\(\\mathcal{V} \\rightarrow (\\bar{\\mathcal{V}}, T)\\). While it is possible to learn these mappings separately (i.e., \\(\\mathcal{V} \\rightarrow \\bar{\\mathcal{V}}\\) and \\(\\mathcal{V} \\rightarrow T\\)), the expert commentary and demonstration are tightly related and provide important contextual information for generating each other. We therefore treat this as an autoregressive generation problem where during training, we use both \\(\\mathcal{V}\\) and \\(T\\) to generate \\(\\bar{\\mathcal{V}}\\) (or \\(\\mathcal{V}\\) and \\(\\bar{\\mathcal{V}}\\) to generate \\(T\\)), while during testing, we drop the extra context information to generate \\(\\bar{\\mathcal{V}}\\) and \\(T\\) directly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Mathematically, consider \\(\\mathcal{F}:(\\mathcal{V},\\bar{\\mathcal{V}},T)\\to \\mathbb{R}\\) as the autoregressive function that jointly learns from the learner demonstration \\(\\nu\\) , the expert demonstration \\(\\tilde{\\nu}\\) , and the expert commentary \\(T\\) . The output \\(\\in \\mathbb{R}\\) is the loss that we aim to minimize. We use the same unified \\(\\mathcal{F}\\) for the joint training (detailed in Sec. 3.3) and inference, as follows. We use \\(\\mathcal{F}'\\) when the autoregressive model \\(\\mathcal{F}\\) is used for output token generation, i.e., text or pose tokens. \\(\\mathcal{F}'\\) can take in any input used during training and generate the remaining one—typical for autoregressive models [74]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "Expert commentary generation. At inference time, to generate an expert commentary \\( T \\), we only have a learner demonstration, thus we mask out the expert demonstration and output the expert commentary:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.16, + 0.905, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\hat {T} = \\mathcal {F} _ {t} (\\mathcal {V}) = \\mathcal {F} ^ {\\prime} (\\mathcal {V}, \\varnothing) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.905, + 0.201 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_t\\) denotes using the model to generate commentary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.905, + 0.351 + ], + "angle": 0, + "content": "Expert demonstration generation. Next, we obtain the expert demonstration in two forms—retrieving a correct execution (video and pose) and generating a pose sequence. These two output formats offer complementary information to understand the actionable feedback. While a video exemplar is helpful for many learner mistakes, pose generation is useful in the absence of a correct expert demonstration in the retrieval set—allowing our model to generalize its coaching beyond the set of discrete expert videos. We leave expert video generation as a future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.905, + 0.396 + ], + "angle": 0, + "content": "Denoting \\(\\mathcal{F}_r\\) and \\(\\mathcal{F}_g\\) as the expert demonstration retrieval and expert pose generation functions, respectively, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.405, + 0.905, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathcal {V}} = \\mathcal {F} _ {r} (\\mathcal {V}) = \\underset {\\forall \\mathcal {V} ^ {\\prime} \\in \\mathcal {E}} {\\operatorname {a r g m i n}} \\mathcal {F} (\\mathcal {V}, \\mathcal {V} ^ {\\prime}, \\hat {T}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.435, + 0.905, + 0.453 + ], + "angle": 0, + "content": "\\[\nP = \\mathcal {F} _ {g} (\\mathcal {V}) = \\mathcal {F} ^ {\\prime} (\\mathcal {V}, \\hat {T}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.905, + 0.508 + ], + "angle": 0, + "content": "where \\(\\hat{T}\\) is the output from Eq. 1 above, and \\(\\mathcal{F}_r\\) and \\(\\mathcal{F}_g\\) denote using the model to retrieve or generate the expert demonstration, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.905, + 0.554 + ], + "angle": 0, + "content": "In summary, the training step \\(\\mathcal{F}\\) uses tuples \\(\\nu\\), \\(\\bar{\\nu}\\), \\(T\\) in a unified way, described in Sec. 3.3, whereas, at inference, all the three functions \\(\\mathcal{F}_t\\), \\(\\mathcal{F}_r\\) and \\(\\mathcal{F}_g\\) use only \\(\\nu\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.831, + 0.578 + ], + "angle": 0, + "content": "3.2. Forming the expert feedback dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.795 + ], + "angle": 0, + "content": "To learn the desired functions, we need to obtain pairs of videos \\((\\mathcal{V},\\bar{\\mathcal{V}})\\) where there is an error in the demonstration in the first video that is corrected in the second video, along with the expert commentary \\(T\\) about \\(\\mathcal{V}\\). Ego-Exo4D [29] offers a great starting point for our setup. It contains ego-exo videos, extracted 3D pose sequences, and time-stamped commentary by experts (e.g., professional soccer coaches) on the demonstrations in the video. The experts watched the entire video and stopped each time they saw something to critique or compliment, offering free-form spoken commentary. In total, this led to 117,812 sentences across 221 hours of video, with most videos commented by 2-5 unique experts [29]. See examples in Figure 2 (left) and details in Sec. 3.4 and Supp." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We propose to automatically augment this data in two ways to enable coaching. First, we seek (pseudo-)annotations of whether a given commentary statement describes a needed improvement or applauds a correct execution. Second, we localize each piece of feedback on a body region, e.g., incorrect hand stretch vs. wrong legs movement. These distinctions are crucial to generate feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.898, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.312, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Overview of the dataset creation. We first summarize the human-provided expert commentary [29] in one sentence using an LLM, and then map it to a body region and correct (green) or incorrect (red) execution label. We then choose incorrect-correct pairs for the same body region to obtain \\(\\mathcal{C}\\). Finally, we choose pairs with minimum temporal alignment loss to obtain the training data. Best in zoom." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.483, + 0.409 + ], + "angle": 0, + "content": "and show corrections. To this end, we create a weakly-supervised training dataset consisting of tuples \\((\\mathcal{V}, T, \\bar{\\mathcal{V}})\\) from Ego-Exo4D [29], as follows." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.411, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Expert commentary classification and body localization. The commentary in [29] is obtained from voice recordings of the experts converted to text with ASR. Most of the samples contain extra comments like \"oh, I will give this a five out of ten\" and \"that's how I would do it too\". Additionally, as discussed above, they lack positive and negative labels and do not explicitly indicate the body region involved (e.g., legs, arms). Thus, we preprocess the expert commentary for three things—making the commentary concise to extract the improvable feedback, marking which body region the feedback is about, and marking whether the feedback states the need for improvement or not." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Since these are all addressable with text reasoning, we use a large language model (LLM) to provide the desired answers. We use Llama3-70B [2], a recent open-sourced model that performs well in current benchmarks. In essence, given a commentary \\( T \\), the language model \\( \\mathcal{L} \\) responds to the prompt: \"Given an expert's commentary, summarize the feedback into a single sentence and also provide which body regions need improvement or are correctly executed...\" (see Supp.). Formally, this yields \\( \\mathcal{L}(T) = (T', (b^1, c^j), \\ldots, (b^s, c^s)) \\) where \\( T' \\) is the concise summary and \\( b^i \\) is a body region like head or arms and \\( c^i \\in \\{0, 1, 2\\} \\) are labels representing needs improvement, correct execution, and no mention, respectively. We group the skeleton joints into six pre-defined body regions \\( b^i \\) (details in Supp.). See red and green boxes in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Pairing incorrect and correct executions. Next, we use the above information to mine pairs of incorrect and correct execution in the dataset. Ego-Exo4D also contains metadata about the skill-level \\( S \\) of the demonstrator, broken into four categories (1-4 in increasing expertise) starting from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.573 + ], + "angle": 0, + "content": "novices who have not performed the activity before to late experts who have performed the same activity, e.g., basketball for \\(10+\\) years. We sample incorrect learner demonstrations from beginners and correct demonstrations from experts from the same activity in Ego-Exo4D, e.g., penalty kicks in soccer or reverse layup in basketball. Even though there can be incorrect executions by experts and correct ones by beginners, errors by experts are likely incomparable to beginners' due to the skill gap. We use the mapping of body regions to find (in)correct executions referring to the same body region. This results in a collection of video pairs with negative and positive feedback about the same body region, e.g., legs in Fig. 2. Formally, the collection \\(\\mathcal{C}\\) is curated as" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.585, + 0.897, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\left(\\mathcal {V} _ {1}, T, \\mathcal {V} _ {2}\\right) \\mid S _ {1} \\in \\mathcal {S} ^ {n}, S _ {2} \\in \\mathcal {S} ^ {e}, \\exists j s. t. c _ {1} ^ {j} = 0, c _ {2} ^ {j} = 1 \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.674 + ], + "angle": 0, + "content": "where \\( S^n = \\{\\text{novice, early expert}\\} \\) and \\( S^e = \\{\\text{late expert, intermediate expert}\\} \\). Using this matching, we obtain a collection of video pairs of incorrect and correct execution (Fig. 2, yellow box)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.795 + ], + "angle": 0, + "content": "Temporal alignment and filtering demonstrations. The commentary annotation process in [29] allows the experts to pause at any instant and provide their feedback. Thus, the clip \\(\\mathcal{V}_1\\) could have the start of a basketball jump shot, whereas the clip \\(\\mathcal{V}_2\\) could also have content before the shot and the follow-through. Thus, we temporally align the learner and the expert video in the collection \\(\\mathcal{C}\\) to ensure we are capturing the same step of a demonstration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "For all pairs of clips in the collection \\( \\mathcal{C} \\), we compare the two poses \\( P \\) to quantify their alignment (similarity) using Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) [27], a translation-invariant and body shape-invariant measure. Note that the video component of the input (Sec. 4) is complementary, as it does capture the person's overall movement in the space." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13585" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.37, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.087, + 0.633, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.087, + 0.907, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.306, + 0.908, + 0.349 + ], + "angle": 0, + "content": "Figure 3. Model overview. We tokenize individual modalities using a modality-specific architecture (top). Once all the modalities are encoded as tokens, we use a large language model to learn expert commentary generation, demonstration retrieval, and pose generation. At inference, the model only takes the learner demonstration video \\(\\nu\\). See text for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.435 + ], + "angle": 0, + "content": "Denote the commentary timestamp as \\( t_1 \\) and \\( t_2 \\) for the two videos. We first choose a fixed window around \\( t_1 \\), say \\( [t_1^a, t_1^b] \\) such that \\( t_1^a < t_1 < t_1^b \\). Next, we find the corresponding time window \\( [t_2^a, t_2^b] \\) with \\( t_2^a < t_2 < t_2^b \\) in \\( \\mathcal{V}_2 \\) such that the PA-MPJPE error is the minimum. Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.443, + 0.462, + 0.477 + ], + "angle": 0, + "content": "\\[\n[ t _ {2} ^ {a}, t _ {2} ^ {b} ] = \\operatorname * {a r g m i n} _ {[ t _ {2} ^ {i}, t _ {2} ^ {j} ] \\in \\mathcal {V} _ {2}} \\mathrm {P A - M P J P E} \\left(\\mathcal {V} _ {1} [ t _ {1} ^ {a}, t _ {1} ^ {b} ], \\mathcal {V} _ {2} [ t _ {2} ^ {i}, t _ {2} ^ {j} ]\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.479, + 0.463, + 0.499 + ], + "angle": 0, + "content": "such that \\( |t_2^j - t_2^i| = |t_1^b - t_1^a| \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.583 + ], + "angle": 0, + "content": "We obtain a subset from \\(\\mathcal{C}\\) that contains the aligned video segments. There can be some video pairs where the demonstration can be very different and hence no appropriate match exists. Therefore, we only keep the top-k pairs with maximum alignment for every incorrect execution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Overall, we obtain a novel dataset \\(\\mathcal{D}\\) that contains pairs of videos where the first video has an incorrect execution (on some body region) and the second video corrects it, thus obtaining the desired tuple \\((\\mathcal{V}, T, \\bar{\\mathcal{V}})\\) for training and testing. See Fig. 2 (right) and examples in Supp. To ensure fair evaluation, we separately establish a clean gold standard test set, free of potential noise from LLM inference. The test set is manually verified (see Supp for details)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.379, + 0.73 + ], + "angle": 0, + "content": "3.3. Architecture and training design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Next, we discuss the architecture that encodes the videos \\(\\mathcal{V}\\), poses \\(P\\), and text \\(T\\) and enables training the auto-regressive function \\(\\mathcal{F}\\). The overall idea is to encode all the representations into a text embedding space and use the strong capabilities of recent language models to obtain output text and pose tokens. This approach has been recently used in vision-conditioned language models [6, 35, 45, 55] for image and video captioning. This setup also allows for a unified architecture for various input and output combinations, as opposed to different input streams for individual modalities. Fig. 3 shows a schematic diagram of the architecture," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.358, + 0.734, + 0.373 + ], + "angle": 0, + "content": "and each part is explained below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Encoding video as tokens. Each video \\(\\mathcal{V}\\) is an ego-exo clip pair.\\(^1\\) The input thus allows the model to see both close-up hand-object interactions (more visible in the learner's ego view) as well as full-body poses in the scene context (more visible in the observer's exo view). We use a pretrained video model to extract spatio-temporal features from both videos' frames, followed by a mapper to convert the video features to video tokens. Formally, \\(\\mathbf{v} = f_{vm}(f_V(\\mathcal{V}))\\) where \\(f_{V}\\) is a standard feature extractor (we use InternVideo2 [77]) and \\(f_{vm}\\) is the visual mapper. See Fig. 3 (top left). The mapper is typically a low-parameter model that is trainable, whereas the high-parameter feature extractor is kept frozen. The mapper helps transform the visual mapper to be akin to text tokens—a popular strategy in visual instruction tuning models [35, 45]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.909, + 0.86 + ], + "angle": 0, + "content": "Encoding pose sequences as tokens. To encode a pose sequence, we use a series of linear layers and MLP-mixer [25] to convert a single pose \\(P \\in \\mathbb{R}^{d \\times 3}\\) to an embedding that can be discretized using a codebook, i.e., \\(\\mathbf{p}' = f_P(P)\\). The architecture in [25] also contains a decoder that converts the embeddings back to human poses, i.e., \\(P = f_{P}^{-1}(\\mathbf{p}')\\), which we use in \\(\\mathcal{F}_g\\) to generate corrected poses. Similar to the above, we use a pose mapper to convert the tokens to embeddings. Formally, \\(\\mathbf{p} = f_{pm}(f_P(P)) = f_{pm}(\\mathbf{p}')\\) is the pose token where \\(f_{pm}\\) is the pose mapper. See Fig. 3 (top center). We concatenate embeddings from every frame to obtain the representation for the whole sequence. Having \\(f_{pm}\\) allows training with fewer parameters and thus, we use \\(\\mathbf{p}\\) for learning \\(\\mathcal{F}_t\\) and \\(\\mathcal{F}_r\\). However, using \\(f_{pm}\\) for generation \\(\\mathcal{F}_g\\) would require adding an inverting function for \\(f_{pm}\\). Thus, we directly add \\(\\mathbf{p}'\\) to LLM tokens for pose generation \\(\\mathcal{F}_g\\). An innovative as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.902 + ], + "angle": 0, + "content": "1The exo view is the one annotated in Ego-Exo4D as having the maximum subject visibility." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13586" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "pect of our approach is to encode pose as multimodal tokens and train with LLMs, unlike other work regressing the pose parameters from a special pose embedding [20] or using a dedicated pose transformer [49, 82]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Encoding text as tokens. Text tokenization is the standard process before inputting a text sentence to the LLM [59, 73], i.e., \\(\\mathbf{t} = f_t(T)\\) where \\(f_{t}\\) is the tokenization function. See Fig. 3 (top right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.482, + 0.305 + ], + "angle": 0, + "content": "Multi-modal sequence prediction. The previous steps yield multimodal tokens \\(\\mathbf{v}\\), \\(\\mathbf{p}\\), and \\(\\mathbf{t}\\) for video, pose, and text, respectively. Next, we use the strong sequence prediction capabilities of large language models to obtain the desired output tokens based on the sequence of multi-modal inputs [6, 35, 45, 55]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.306, + 0.485, + 0.624 + ], + "angle": 0, + "content": "For training \\(\\mathcal{F}\\) for expert commentary generation, we provide a sequence of learner video and pose tokens (v and p) and the corrected pose tokens (\\(\\bar{\\mathbf{v}}\\) and \\(\\bar{\\mathbf{p}}\\) corresponding to \\(\\bar{\\nu}\\)). We ask the model to predict the expert commentary token sequence. For the sequence prediction language model \\(\\mathcal{L}_s\\), we wish to obtain \\(\\mathbf{t} = \\mathcal{L}_s(\\mathbf{v},\\mathbf{p},\\bar{\\mathbf{v}},\\bar{\\mathbf{p}})\\). Consequently, the training objective is the standard cross-entropy loss, \\(\\min_{\\theta}\\{-\\log (\\mathbf{t}|\\mathbf{v},\\mathbf{p},\\bar{\\mathbf{v}},\\bar{\\mathbf{p}};\\theta)\\}\\), where \\(\\theta\\) are the parameters of the model. See Fig. 3 (bottom left). For consistency with the training of these language models, the sequence is formatted to be conversational in nature, e.g., \"provide an expert's commentary based on this pose sequence:\". Similarly, for expert demonstration retrieval, we wish to obtain a retrieval candidate \\(\\bar{\\mathbf{p}} = \\mathcal{L}_s(\\mathbf{v},\\mathbf{p},\\mathbf{t})\\). Likewise, the training objective with parameters \\(\\gamma\\) is \\(\\min_{\\gamma}\\{-\\log (\\bar{\\mathbf{p}} |\\mathbf{v},\\mathbf{p},\\mathbf{t};\\gamma)\\}\\), which we also use as the relevance score for retrieval during inference. Expert pose generation is trained with the same objective, except we use \\(\\mathbf{p}'\\) instead of \\(\\mathbf{p}\\) (and obtain \\(\\bar{\\mathbf{p}}'\\)). See Fig. 3 (bottom). Generating pose further requires converting back the pose tokens \\(\\bar{\\mathbf{p}}'\\) to 3D joints using the pose decoder \\(f_P^{-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.686 + ], + "angle": 0, + "content": "At inference time, as introduced in Sec. 3.1, we only use \\(\\mathcal{V}\\) as the input: \\(\\mathcal{F}_t\\) drops the expert demonstration and predicts the expert commentary \\(\\hat{T}\\). We use this predicted commentary as input for \\(\\mathcal{F}_r\\) and \\(\\mathcal{F}_g\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.305, + 0.713 + ], + "angle": 0, + "content": "3.4. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Dataset and statistics. Ego-Exo4D [29] contains 5,035 videos of participants doing activities across eight scenarios. We focus on three physical scenarios—basketball, soccer, and rock climbing—though our model is in principle generalizable to other physical skills, without any task-specific design, unlike [63, 78]. We use physical scenarios since the coaching feedback is groundable in the body regions, as opposed to procedural tasks like cooking, where the suggestions can be alternate ingredients or steps that are not visually present. Details and statistics are in Supp. Following the dataset creation process outlined in Sec. 3.2, we obtain a dataset of 25,505 training and 1,272 testing tuples" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.152 + ], + "angle": 0, + "content": "of \\((\\mathcal{V}, T, \\bar{\\mathcal{V}})\\). We choose \\(k = 5\\) (train) and \\(k = 1\\) (test) for min-k choice of correct demonstrations per incorrect execution. To reiterate, we manually examine the test set for correct commentary summary and body region labeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Network architecture. The video model \\( f_{V} \\) is an InternVideo2 [77] encoder that provides strong visual representations. The pose encoder \\( f_{P} \\) and decoder \\( f_{P}^{-1} \\) are adapted from PCT [25], which learns compositional tokens from human poses. Since the original training in [25] uses 2D poses, we adapt it to use 3-dimensions and retrain on the human poses in [29]. Both the visual and pose mappers are low-parameter MLP layers [44]. Finally, we use Llama 3-8B [2] as the LLM multimodal encoder for sequence prediction. We finetune \\( \\mathcal{F}_t \\) and \\( \\mathcal{F}_r \\) for 5 epochs with a learning rate of \\( 5 \\times 10^{-5} \\). Next, we modify the token dimension in \\( \\mathcal{F}_g \\) to accommodate pose tokens \\( \\mathbf{p}' \\), and hence, we fine-tune \\( \\mathcal{L}_s \\) when learning \\( \\mathcal{F}_g \\) with a learning rate of \\( 5 \\times 10^{-6} \\) for 5 epochs. Video \\( (f_{V}) \\) and pose models \\( (f_{P}, f_{P}^{-1}) \\) are kept frozen. All the models are trained on 8 V100 32GB GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.742, + 0.412 + ], + "angle": 0, + "content": "4. Experiments and results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.905, + 0.496 + ], + "angle": 0, + "content": "We first discuss the baselines and ablations, followed by the evaluation setup and results for the three outputs—expert commentary generation \\((\\mathcal{F}_t)\\), expert demonstration retrieval \\((\\mathcal{F}_r)\\), and pose demonstration generation \\((\\mathcal{F}_g)\\). We also show qualitative examples and discuss the limitations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.497, + 0.866, + 0.512 + ], + "angle": 0, + "content": "Baselines. We compare with the following baselines:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.513, + 0.905, + 0.619 + ], + "angle": 0, + "content": "- InternVideo2-NN, InternVideo2-FT [77]: Given a query video \\(\\mathcal{V}\\), the nearest neighbor baseline (NN) finds the most similar video by InternVideo2 [77] feature similarity in the training data and returns the corresponding commentary or demonstration, for all tasks. The FT version finetunes the model to contrastively match the learner demonstration with the expert commentary/demonstration." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.622, + 0.905, + 0.759 + ], + "angle": 0, + "content": "- VideoChat2 [36, 37], LLaVA [45]: In these methods, we prompt SOTA video and image captioning models to generate commentary for an input demonstration. We use the log-likelihood loss to find the retrieved expert demonstration. These baselines evaluate if existing video captioning methods can provide expert feedback or retrieve expert demonstrations. We use the ego and \"best exo\" frames, same as for our method. Note that neither of these baselines are applicable for pose generation \\(\\mathcal{F}_g\\)." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.761, + 0.905, + 0.852 + ], + "angle": 0, + "content": "- LLaVA-FT [45], LLaVA-FT w/ pose [45]: These baselines are based on the SOTA visual-language method LLaVA [45] but trained on our dataset with the same text model, i.e., Llama 3-8B [2], for an apples-to-apples comparison. The \"w/ pose\" variant also takes the 3D pose coordinates in text form as input." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- PoseScript [16], PoseFix [17]: These two works enable pose-to-text and text-to-pose reasoning. The text generated or used for pose generation contains detail about the" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.513, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13587" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.094, + 0.609, + 0.309 + ], + "angle": 0, + "content": "
MethodCommentary Gen.Demo Ret.Pose Gen.
B@4MR-LRmedR↓P↓
InternVideo2-NN [77]42.146.949.313.5198161
InternVideo2-FT [77]42.947.650.014.1190157
VideoChat2 [37]27.844.341.914.9183—
LLaVA [45]28.544.144.215.0183—
LLaVA-FT [45]43.548.551.517.8177—
LLaVA-FT w/ pose [45]43.648.551.718.0172150
PoseScript/Fix [16, 17]24.144.546.315.9182182
ExpertAF44.949.654.619.1158135
ExpertAF w/o video44.649.454.218.7161139
ExpertAF w/o pose44.249.354.218.6163—
w/o alignment42.048.651.516.9180153
w/ global pose43.047.952.616.5184150
ExpertAF w/ full-sup45.850.955.722.5146131
" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.09, + 0.905, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.317, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Table 1. Results on automatic metrics (left) and human evaluation (right). We break down results for the three outputs—expert commentary generation, expert demonstration retrieval, and expert pose generation. Our method outperforms all baselines and prior work on all tasks. The last row \"w/ full-sup\" uses privileged input (the demo video \\(\\bar{\\nu}\\)) at inference. (B@4: BLEU-4, M: Meteor, R-L: ROUGE-L F1, R: recall@50, medR: median rank, P: PA-MPJPE). For all metrics higher is better, except medR and PA-MPJPE (\\(\\downarrow\\)). Our method is also rated higher by human raters on a Likert scale (min:1, max:4), compared to all the other methods (right). See text for details." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.395, + 0.482, + 0.501 + ], + "angle": 0, + "content": "location of body parts, e.g., the hands are raised, as opposed to expert commentary. Hence, they let us evaluate if pose description is adequate for expert feedback. For pose generation, we evaluate if providing an expert commentary helps generate the desired expert demonstration. Both methods use SMPL [47] pose and hence we convert 3D pose to SMPL and vice versa [39], as needed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Ablations. In addition to the strong baselines, we also compare the performance against ablations and variants of our design choices. ExpertAF w/o pose and ExpertAF w/o video evaluate the performance when only one modality is used. ExpertAF w/o alignment quantifies the need for temporal alignment (Sec. 3.2), while ExpertAF w/ global pose evaluates if just providing the model with one correct pose per activity chosen based on expert commentary is enough (vs. finding the closest correct pose sequence). ExpertAF w/ full-sup is a stronger variant with privileged input for inference—\\(\\mathcal{F}_t\\) uses \\(\\mathcal{V}\\), \\(\\bar{\\mathcal{V}}\\) to predict \\(T\\) and \\(\\mathcal{F}_r\\) and \\(\\mathcal{F}_p\\) uses both \\(\\mathcal{V}\\) and \\(T\\). See Supp. for additional ablations of the choice of LLM \\(\\mathcal{L}_s\\), contribution of ego and exo videos, and joint training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Expert commentary generation. To evaluate the text commentary, we use standard metrics BLEU-4 [60], METEOR [7] and ROUGE-L F1 [65], following prior work in evaluating text generation [14, 58] (higher is better)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tab. 1 (left) shows the results. Our method outperforms all methods on all captioning metrics, with gains up to \\(3\\%\\) over the best baseline trained on the same dataset. All the gains are statistically significant using paired t-test with \\(p < 0.05\\). Fig. 4 (top) shows commentary generation by various methods. Firstly, SOTA methods LLaVA [45] and VideoChat2 [37], being captioning methods, yield a verbose description of the activity, rather than a critique." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Fine-tuning LLaVA with our method improves the performance, but still it misses our use of expert demonstrations and pose sequences. Even using pose information as text only marginally improves \"LLaVA-FT w/ pose\", showing the effectiveness of our explicit pose encoding. Moreover, PoseScript [16] generates pose descriptions like \"head is turned to the left\" which is inadequate for feedback. Our results over the baselines show the advantage of temporal alignment and a diverse expert demonstration set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.909, + 0.639 + ], + "angle": 0, + "content": "Our stronger variant \"w/ full-sup\" achieves an even better result by taking \\(\\mathcal{V}\\) and \\(\\bar{\\mathcal{V}}\\) as inputs during inference. It is useful in cases where a learner has access to the expert demonstration but wants actionable feedback in languageform. Finally, Fig. 4 (bottom row) shows a failure case of commentary generation, showcasing the difficulty in pinpointing the exact mistake." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Expert demonstration retrieval. For retrieval, we use standard retrieval metrics: recall@k and median rank. We set k=50 to account for the fact that multiple expert demonstrations could suitably correct the given learner demonstration. The retrieval set contains 1,272 samples, the same as the test set size. Higher recall and lower ranks are better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Tab. 1 (middle) shows the results. Our method improves the median rank by 14 positions compared to the best baseline and a significantly better recall@50. Our method and the baselines show a similar trend as with the expert commentary generation above, the stronger baseline w/ \"full-sup\" (\\(>4\\%\\) better recall and 26 rank improvement) being crucial in cases where a learner has the commentary but wants a video exemplar to learn from." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Expert pose generation. Tab. 1 (right column) shows the expert pose generation results, with error measured by PA-MPJPE [27] (mm). Our novel pose generation method" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.092, + 0.216, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.246, + 0.092, + 0.326, + 0.104 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.117, + 0.346, + 0.175 + ], + "angle": 0, + "content": "The climber is exhausted and struggling to make a move, with poor hand and arm positioning." + }, + { + "type": "title", + "bbox": [ + 0.404, + 0.092, + 0.438, + 0.104 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.117, + 0.48, + 0.174 + ], + "angle": 0, + "content": "The climber is struggling to maintain the position and is unable to make the next move." + }, + { + "type": "title", + "bbox": [ + 0.523, + 0.092, + 0.586, + 0.103 + ], + "angle": 0, + "content": "LLaVA-FT" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.117, + 0.617, + 0.173 + ], + "angle": 0, + "content": "The climber is wasting energy by pulling away, which is slowing down reaching the next handhold" + }, + { + "type": "title", + "bbox": [ + 0.652, + 0.092, + 0.726, + 0.104 + ], + "angle": 0, + "content": "VideoChat2" + }, + { + "type": "text", + "bbox": [ + 0.63, + 0.117, + 0.747, + 0.174 + ], + "angle": 0, + "content": "The person can improve by practicing their technique, warming up and stretching." + }, + { + "type": "title", + "bbox": [ + 0.789, + 0.092, + 0.858, + 0.104 + ], + "angle": 0, + "content": "PoseScript" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.108, + 0.885, + 0.182 + ], + "angle": 0, + "content": "The person is standing on their left leg with the right leg lifted up and bent at the knee. The torso is leaning to the left. The head is turned to the left." + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.204, + 0.242, + 0.215 + ], + "angle": 0, + "content": "Learner demonstration" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.215, + 0.245, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.293, + 0.245, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.381, + 0.207, + 0.392 + ], + "angle": 0, + "content": "Failure case" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.393, + 0.244, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.262, + 0.204, + 0.409, + 0.216 + ], + "angle": 0, + "content": "Expert comment. - Ours" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.232, + 0.413, + 0.277 + ], + "angle": 0, + "content": "The shooter's stance is correct, but they need to throw the ball with one hand and guide it with the other." + }, + { + "type": "text", + "bbox": [ + 0.256, + 0.308, + 0.416, + 0.354 + ], + "angle": 0, + "content": "The player's stance is too straight, which prevents them from generating the required power and control on the ball." + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.393, + 0.414, + 0.439 + ], + "angle": 0, + "content": "The shooter's left elbow is bent and his wrist is not fully extended, affecting the ball's rotation and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.422, + 0.204, + 0.58, + 0.216 + ], + "angle": 0, + "content": "Ground truth commentary" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.227, + 0.581, + 0.283 + ], + "angle": 0, + "content": "The shooter is using two hands to shoot the ball, instead of using one shooting hand and arm with the other at a 90-degree angle." + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.302, + 0.584, + 0.358 + ], + "angle": 0, + "content": "The player's foot position and body lean are affecting the quality of their ball contact, leading to a loss of power and control." + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.393, + 0.582, + 0.439 + ], + "angle": 0, + "content": "He does a good job using his left hand on the left side, but needs to improve his posture and weight distribution." + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.204, + 0.734, + 0.216 + ], + "angle": 0, + "content": "Expert demo. retrieval" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.215, + 0.738, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.292, + 0.738, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.377, + 0.738, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.204, + 0.869, + 0.216 + ], + "angle": 0, + "content": "Pose generation" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.216, + 0.894, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.292, + 0.894, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.377, + 0.894, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 4. Qualitative results. (Top) Comparison of expert commentary generated by various baselines. (Second and third row) Examples of expert commentary generation, demonstration retrieval, and pose generation by our method. Notice the expert demonstration and pose generation corrects the mistake pointed out in the commentary, i.e., one hand is used to throw and the other to guide, and body position is improved to control better control and power. Colored text and marks (red for mistake, green for correction) are shown only for visualization. (Bottom) Failure cases. See Supp. for video results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.711 + ], + "angle": 0, + "content": "performs better than learning 3D position using text in \"LLaVA-FT w/ pose\". Furthermore, PoseFix [17] is designed to generate SMPL parameters based on a coarse modification text description. Our use of generation to output poses (vs. retrieval) is advantageous when the candidate retrieval set does not contain the correct demonstration; see Supp. See Fig. 4 and Supp. for qualitative examples including failure cases. The 3D pose sequences in the dataset are often auto-generated and have some noisy samples, discussed in Supp. Across the three tasks, our method is superior to its ablations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Human evaluation. For all the tasks, we also solicit human evaluation (Tab. 1 right) to rate the quality of the generated text descriptions. Five raters (per scenario) uninvolved with this project rate the quality of each generated output by scoring on a Likert scale from 1 to 4 (higher is better); details in Supp. We ensure that the raters have a basic knowledge of the scenario they are rating (basketball, soccer, rock climbing)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Across all three tasks, human raters score our method the highest. A significant gap of up to 0.7 over the possible range \\((1 - 4)\\), as well as gains up to \\(3\\times\\), showcase ExpertAF's expert feedback quality. That said, there is nat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.713 + ], + "angle": 0, + "content": "urally room for improvement on this new task. We find that while humans say our model excels on cases where the feedback is visually groundable, e.g. incorrect hand angle, it tends to fall short when the feedback is not directly visible, e.g. the climber looks fatigued. The human evaluation complements the automatic metrics above, overcoming the limitation that a \"ground truth\" commentary or a visual demonstration may capture only one of multiple possible errors in the learner's demonstration [23, 68] (e.g., for a basketball shot, both the hand placement and the jump could be incorrect, but only one may be mentioned in the ground truth)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.736, + 0.634, + 0.752 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We proposed a novel task and method to generate expert commentary and demonstrations from a learner's video. We develop a weakly-supervised training approach and benchmark for this problem. Our novel method fusing multimodal inputs from learner and expert demonstrations together with expert commentary results in state-of-the-art performance in actionable feedback, and lays the groundwork for accessible, affordable, and actionable AI coaching applications in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13589" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.15 + ], + "angle": 0, + "content": "Acknowledgement. UT Austin is supported in part by the IFML NSF AI Institute. Thanks to Fu-Jen Chu, Jing Huang and Xitong Yang for help with the Exo-Ego4D [29] pose extraction pipeline." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.175, + 0.188, + 0.19 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.482, + 0.268 + ], + "angle": 0, + "content": "[1] Yazan Abu Farha, Alexander Richard, and Juergen Gall. When will you do what?-anticipating temporal occurrences of activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5343-5352, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.395, + 0.284 + ], + "angle": 0, + "content": "[2] AI@Meta. Llama 3 model card, 2024. 4, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.286, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23066-23078, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. What you say is what you show: Visual narration detection in instructional videos. arXiv preprint arXiv:2301.02307, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Kumar Ashutosh, Zihui Xue, Tushar Nagarajan, and Kristen Grauman. Detours for navigating instructional videos. In CVPR, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[7] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[8] Tobias Baumgartner and Stefanie Klatt. Monocular 3d human pose estimation for sports broadcasts using partial sports field registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5108-5117, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the IEEE international conference on computer vision, pages 2177-2185, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Gedas Bertasius, Aaron Chan, and Jianbo Shi. Egocentric basketball motion planning from a single first-person image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.2, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Ramesh. Video generation models as world simulators, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. arXiv preprint arXiv:2305.18500, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[15] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos. International Journal of Computer Vision, 129:2846-2864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[16] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[17] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[18] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 319-335. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[19] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Posegpt: Chatting about 3d human pose. arXiv preprint arXiv:2311.18836, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[20] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[21] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence, 43(11):4021-4036, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.68, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[22] Jiyang Gao, Zhenheng Yang, and Ram Nevatia. Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.722, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal, Pawan Sasanka Ammanamanchi, Anuoluwapo Aremu, Antoine Bosselut, Khyathi Raghavi Chandu, Miruna-Adriana Clinciu, Dipanjan Das, Kaustubh Dhole, Wanyu Du, Esin Durmus, Ondrej Dušek, Chris Chinenye Emezue, Varun Gangal, Cristina Garbacea, Tatsunori Hashimoto, Yufang Hou, Yacine Jernite, Harsh Jhamtani, Yangfeng Ji, Shailza Jolly, Mhir Kale, Dhruv Kumar, Faisal Ladhak, Aman Madaan, Mounica Maddela, Khyati Mahajan, Saad Mahamood, Bodhisattwa Prasad Majumder, Pedro Henrique Martins, Angelina McMillan-Major, Simon Mille, Emiel van Miltenburg, Moin Nadeem, Shashi Narayan, Vitaly Nikolaev, Andre Niyongabo Rubungo, Salomey Osei, Ankur" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13590" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.232 + ], + "angle": 0, + "content": "Parikh, Laura Perez-Beltrachini, Niranjan Ramesh Rao, Vikas Raunak, Juan Diego Rodriguez, Sashank Santhanam, João Sedoc, Thibault Sellam, Samira Shaikh, Anastasia Shimorina, Marco Antonio Sobrevilla Cabezudo, Hendrik Strobelt, Nishant Subramani, Wei Xu, Diyi Yang, Akhila Yerukola, and Jiawei Zhou. The GEM benchmark: Natural language generation, its evaluation and metrics. In Proceedings of the 1st Workshop on Natural Language Generation, Evaluation, and Metrics (GEM 2021), pages 96-120, Online, 2021. Association for Computational Linguistics. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[24] Adelaide Genay, Anatole Lécuyer, and Martin Hachet. Being an avatar \"for real\": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 28(12):5071-5090, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[25] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 660-671, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.414 + ], + "angle": 0, + "content": "[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[27] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[28] Andrew S Gordon. Automated video assessment of human performance. In Proceedings of AI-ED, page 10, 1995. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[29] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In CVPR, 2024. 1, 2, 3, 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[31] William Hoff and Tyrone Vincent. Analysis of head pose accuracy in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 6(4):319-334, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[32] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[33] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7718-7727, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[35] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[36] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[37] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[38] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[39] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yukun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv preprint arXiv:2304.00553, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[40] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[42] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9522-9531, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[46] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph., 34(6), 2015. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13591" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.487, + 0.161 + ], + "angle": 0, + "content": "[48] Sabine Losch, Eva Traut-Mattausch, Maximilian D Muhlberger, and Eva Jonas. Comparing the effectiveness of individual coaching, self-coaching, and group training: How leadership makes the difference. Frontiers in psychology, 7: 175595, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.164, + 0.486, + 0.22 + ], + "angle": 0, + "content": "[49] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.221, + 0.485, + 0.275 + ], + "angle": 0, + "content": "[50] Priyanka Mandikal and Kristen Grauman. Dexvip: Learning dexterous grasping with human hand pose priors from video. In Conference on Robot Learning, pages 651-661. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.278, + 0.485, + 0.332 + ], + "angle": 0, + "content": "[51] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action forecasting@ ego4d challenge 2022. arXiv preprint arXiv:2207.12080, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.334, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[52] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.378, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[53] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.463, + 0.485, + 0.545 + ], + "angle": 0, + "content": "[54] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.548, + 0.484, + 0.574 + ], + "angle": 0, + "content": "[55] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. In CVPR, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.577, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[56] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.634, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[57] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. arXiv preprint arXiv:2207.09666, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.691, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[58] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. In European Conference on Computer Vision, pages 167–184. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.749, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[59] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fo" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.487, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.546, + 0.093, + 0.908, + 0.898 + ], + "angle": 0, + "content": "tis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mély, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O’Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael. Pokorny. Michelle Pokrass. Vitchyr H. Pong Tolly Powell Alethea Power Boris Power Elizabeth Proehl Raul Puri Alec Radford Jack Rae Aditya Ramesh Cameron Raymond Francis Real Kendra Rimbach Carl Ross Bob Rotsted Henri Roussez Nick Ryder Mario Saltarelli Ted Sanders Shibani Santurkar Girish Sastry. Heather Schmidt David Schnurr John Schulman Daniel Selsam. Kyla Sheppard Toki Sherbakov Jessica Shieh Sarah Shoker Pranav Shyam Szymon Sidor Eric Sigler Maddie Simens Jordan Sitkin Katarina Slama Ian Sohl Benjamin Sokolowsky Yang Song Natalie Staudacher Felipe Petroski Such Natalie Summers Ilya Sutskever Jie Tang. Nikolas Tezak Madeleine B. Thompson Phil Tillet. Amin Tootoonchian Elizabeth Tseng Preston Tuggle Nick Turley Jerry Tworek Juan Felipe Cerón Uribe Andrea Vallone Arun Vijayvergiya Chelsea Voss Carroll Wainwright Justin Jay Wang Alvin Wang Ben Wang Jonathan Ward Jason Wei CJ Weinmann Akila Welihinda Peter Welin" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13592" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.189 + ], + "angle": 0, + "content": "der, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[60] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[61] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[62] Paritosh Parmar, Jaiden Reddy, and Brendan Morris. Piano skills assessment. In 2021 IEEE 23rd international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[63] Paritosh Parmar, Amol Gharat, and Helge Rhodin. Domain knowledge-informed self-supervised representations for workout form assessment. In European Conference on Computer Vision, pages 105-123. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[64] Hamed Pirsiavash, Carl Vondrick, and Antonio Torralba. Assessing the quality of actions. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 556-571. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[65] Lin CY Rouge. A package for automatic evaluation of summaries. In Proceedings of Workshop on Text Summarization of ACL, Spain, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[66] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[67] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[68] Chenhui Shen, Liying Cheng, Xuan-Phi Nguyen, Yang You, and Lidong Bing. Large language models are not yet human-level evaluators for abstractive summarization. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 4215–4233, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[69] Sumedh Sontakke, Jesse Zhang, Séb Arnold, Karl Pertsch, Erdem Biryik, Dorsa Sadigh, Chelsea Finn, and Laurent Itti. Roboclip: One demonstration is enough to learn robot policies. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[70] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[71] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9839-9848, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[72] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4790-4799, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[74] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[75] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[76] Tianyu Wang, Yijie Wang, and Mian Li. Towards accurate and interpretable surgical skill assessment: A video-based method incorporating recognized surgical gestures and skill levels. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 668-678. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[77] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[78] Chengming Xu, Yanwei Fu, Bing Zhang, Zitian Chen, YuGang Jiang, and Xiangyang Xue. Learning to score figure skating sport videos. IEEE Transactions on Circuits and Systems for Video Technology, 30(12):4578-4590, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[79] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13593" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[80] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[82] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[83] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129:703-718, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[84] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[85] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[86] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[87] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13594" + } + ] +] \ No newline at end of file diff --git a/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_origin.pdf b/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1b58d9bc222da8d16333576aedbcdc891baefbdf --- /dev/null +++ b/2025/ExpertAF_ Expert Actionable Feedback from Video/ccac2a98-a341-4697-825f-e838745145a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f44abb8bbb05364ccea032674e71c6f5047c54ca18a3c12ce2573c1f332c6bb +size 3038893 diff --git a/2025/ExpertAF_ Expert Actionable Feedback from Video/full.md b/2025/ExpertAF_ Expert Actionable Feedback from Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..620c27446ff6ef3e1f055813f4c2e2d475353a1f --- /dev/null +++ b/2025/ExpertAF_ Expert Actionable Feedback from Video/full.md @@ -0,0 +1,383 @@ +# ExpertAF: Expert Actionable Feedback from Video + +Kumar Ashutosh $^{1,2}$ , Tushar Nagarajan $^{2}$ , Georgios Pavlakos $^{1}$ , Kris Kitani $^{2,3}$ , Kristen Grauman $^{1,2}$ $^{1}$ University of Texas at Austin, $^{2}$ FAIR Meta, $^{3}$ Carnegie Mellon University + +# Abstract + +Feedback is essential for learning a new skill or improving one's current skill-level. However, current methods for skill-assessment from video only provide scores or compare demonstrations, leaving the burden of knowing what to do differently on the user. We introduce a novel method to generate actionable feedback (AF) from video of a person doing a physical activity, such as basketball or soccer. Our method takes a video demonstration and its accompanying 3D body pose and generates (1) free-form expert commentary describing what the person is doing well and what they could improve, and (2) a visual expert demonstration that incorporates the required corrections. We show how to leverage Ego-Exo4D's [29] videos of skilled activity and expert commentary together with a strong language model to create a weakly-supervised training dataset for this task, and we devise a multimodal video-language model to infer coaching feedback. Our method is able to reason across multi-modal input combinations to output full-spectrum, actionable coaching—expert commentary, expert video retrieval, and expert pose generation—outperforming strong vision-language models on both established metrics and human preference studies. + +# 1. Introduction + +An abundance of instructional "how-to" videos on the internet enables skill-learning by observing expert demonstrations. Instructional videos cover many skills one might want to learn—cooking, DIY, sports, crafts, and hobbies. Recent work leverages such videos to assist human [5, 6, 53] and robot [56] skill learning. Acquiring skills from video is especially appealing for equitable learning: access to video is generally less costly and more widely available than face-to-face access to an expert coach. + +Despite the myriad of videos and tutorials, people prefer to learn in a feedback loop: getting to know their mistakes, finding the improvements they need to do and correcting those. For that reason, coaching-based iterative training is more effective than self-learning [48]. A good coach has + +Learner demonstration (input) + +![](images/1e3b994c58c5bc89cce539628bb1eff839d9ae97b08bdb44e825cd5fb66abfed.jpg) + +![](images/6038d690b7c7e225445bee77186e627ab34e311406aef7fba9c5b97de97afd6b.jpg) + +Expert demonstration (output) + +![](images/edfaf44dfd1543af4a8afffc07c108f5412bac965a4bf6da4551e53bf82cd72a.jpg) + +![](images/8bc25e3a2b530b15ffb32fbb177680dc5304d424b8a3fb10dd20934ec3e4ede3.jpg) + +Expert commentary (output) + +The player is taking big steps to control the ball, but lacks body control, and needs to slow down to maintain a better control. + +Figure 1. An example of expert feedback. When a player is dribbling the ball fast, they tend to lose control (top left). Our proposed method provides an expert commentary to the learner suggesting improvements (bottom). The method also provides an expert demonstration that shows the desired correction, where the player is maintaining smaller steps and body control (top right). + +three components: finding mistakes, providing verbal corrections and finally, showing visual demonstrations. For example, suppose a person learning to dribble a soccer ball takes big steps that decrease body control. The coach should identify this mistake, provide verbal feedback, and show the correct dribbling technique. See Figure 1. + +At present, a person watching a how-to video has to identify mistakes themselves and attempt to correct them. Pinpointing the exact mistake is itself challenging for beginners—particularly when learning physical skills, like sports and dance, where subtle movements and positioning of the body are critical to success. + +What role could AI play in overcoming this gap? Current work on proficiency understanding from video focuses on scoring a demonstration [9, 28, 29, 78], thus addressing only one component of coaching. There is no method that can provide actionable verbal and visual feedback on a learner's demonstration. Doing so is technically challenging since it requires the method to (a) understand the user's activity, (b) detect mistakes in the execution by comparing + +the user's activity with correct examples and finally, (c) suggest edits with respect to the better way of doing the activity. In short, the output should be coach-like—specific actionable feedback that will improve a learner's proficiency. + +We propose ExpertAF: a novel method to provide expert actionable feedback on a user's physical activity. ExpertAF's actionable feedback has two forms: language-based commentary and visual demonstration. The freeform expert commentary output describes the mistake and what should be corrected, e.g., "take smaller steps and slow down to maintain control", while the expert visual demonstration output shows the correct way of doing it. + +Both components introduce unique technical challenges. For the former, the task of generating free-form text feedback is distinct from standard image and video captioning, as the model needs to identify the mistake and actions that would correct it—rather than simply describe the observed pose [16] or activity [32, 38, 57, 75, 81]. Similarly, for the latter, the task of retrieving (or generating) a demonstration that corrects a specific mistake is distinct from similarity-based retrieval [3, 40, 54, 77, 79] or open-ended generation [12, 19]. We hypothesize—and experimentally validate—that personalized, video-conditioned coaching is more accurate than simply returning a global expert execution because there are many right ways of doing an activity, e.g., penalty shots in soccer, and hence different ways to fix what is wrong in a given execution. Overall, the model must understand both what is being done and what actionable tweaks would make it better. + +We are the first to address video-based coaching with actionable feedback. We develop the first model to generate free-form text feedback on a user's activity (captured in video and 3D pose sequences) as well as a video demonstrating how to improve, by building on a recent vision-language model [45]. There is currently no dataset for this challenging task. Therefore, we augment Ego-Exo4D [29], an existing dataset with video, commentary, and poses, and show how to use a strong large language model to create a weakly-supervised training set consisting of paired learner-expert demonstrations, along with expert textual commentary that relates the two. We also obtain a gold-standard manually labeled test set for rigorous evaluation. + +We validate ExpertAF on three diverse physical scenarios: soccer, basketball, and rock climbing. Our results show a significant improvement over strong baselines, establishing the first method to provide actionable feedback, including a novel technique for expert pose generation. Alongside consistent gains in quantitative metrics against the ground truth, we also show our model's promise via direct human evaluation, where ExpertAF outperforms off-the-shelf video models by as much as $3 \times$ . + +# 2. Related work + +Skill learning from videos. Instructional video datasets like HowTo100M [53], COIN [70], and CrossTask [87] enable procedural understanding through step recognition [5, 42, 85], procedure planning [11, 13, 84], task-graph discovery [5, 18, 29, 85], action anticipation [1, 21, 22, 26, 51], and alignment detection [4, 30]. Capitalizing on their instructional nature, recent work learns robot policies from these videos [50, 69], uses external knowledge-bases like WikiHow to ground the steps of a procedural task [42, 52, 85], and explores new ways to navigate between multiple demonstrations [6]. Despite their scale, how-to videos in these datasets [53, 70, 87] are often created by domain experts and hence lack mistakes or suboptimal executions. In contrast, Ego-Exo4D [29] contains multiple executions of various common scenarios by people with varied skill levels—from novice to late expert. Moreover, it contains experts' feedback on those actions. We show how to augment this data via large language models to support our new idea for video-based coaching. + +3D body pose for activity understanding. 3D body pose is crucial for human activity understanding, capturing a person's stance and motion. Supported by valuable datasets like Humans3.6M [33] and NTU RGB-D [46, 66], recent methods improve body pose understanding [25, 34, 67, 72, 83, 86]. Beyond action classification, pose is also crucial for markerless motion capture of sports [8, 15] and interacting in augmented reality [24, 31]. While our ExpertAF leverages pose as an important signal of human activity, unlike prior work we interface video, pose, and free-form text to generate expert feedback and an expert demonstration video. Our work is also different from generating [20] or modifying [17] body pose from text (e.g., "raise your left arm"), as our task is to understand a potentially incorrect pose and provide feedback. Understanding these minor differences between incorrect and correct poses offers unique challenges that are addressed by our proposed learning scheme. Our work is orthogonal to methods that improve pose estimation itself; any future improvements on that front would only benefit our approach. + +Skill assessment and coaching. Prior work explores skill assessment for a variety of tasks, particularly sports. Most methods pose skill-assessment as a score-prediction task, i.e., for skating [78], gymnastics [28], basketball [9], or multiple sports [61]. Since absolute scoring can be ambiguous and error-prone, recent work [71] explores uncertainty-aware score distributions. Instead of explicitly scoring a demonstration, other work determines which of two basketball players are better [9], uses group-aware contrastive regression to learn the relative quality of demonstrations [80], or generates a full ideal trajectory from the first frame [10]. Fitness-AQA [63] and Action Quality [64] provide outputs like knees inward error, shallow squat error, + +or an arrow highlight for more localized feedback. There is also research on skill assessment in non-sports domains, like surgery [43, 76] and piano [62]. All the prior work assumes a fixed taxonomy of errors, and the taxonomy is designed separately for each activity. Furthermore, unlike our work, none of the prior work provides feedback akin to a personal coach—which requires not only telling what is wrong, but also expressing how to correct the mistake. + +# 3. Method + +We introduce the problem statement in Sec. 3.1, the dataset creation strategy in Sec. 3.2, the training design in Sec. 3.3, and implementation details in Sec. 3.4. + +# 3.1. Problem statement + +Consider a dataset $\mathcal{E} = \{(\mathcal{V}, T, \bar{\mathcal{V}})\}$ where each $\mathcal{V}_i$ is a video demonstration, and $T_i$ is a free-form text commentary critiquing the activity in the video, e.g., "take smaller steps and slow down to maintain control." $\bar{\mathcal{V}}$ is a related video demonstration but without the error mentioned in $T$ , e.g., the player takes smaller steps with better control. See Fig. 1. Each demonstration $\mathcal{V} = \{V, P, S\}$ consists of three parts: $V$ is the RGB video, and $P$ is the 3D pose sequence of the participant with skill-level $S$ in the video. The pose representation $P \in \mathbb{R}^{n \times d \times 3}$ contains $n$ frames of 3D positions for $d = 17$ body joints, consistent with MS-COCO [41]. $S$ is the participant's skill-level for an activity, i.e., novice, early expert, intermediate expert, or late expert, and will be used later in creating the dataset. We assume there is one active person doing a physical activity in any given video who is the subject of ExpertAF. + +The goal in this work is to provide feedback on a given learner's video containing a physical scenario. The feedback is an expert demonstration $\bar{\mathcal{V}}$ and an expert commentary $T$ , tailored to $\mathcal{V}$ . Formally, we want to learn a mapping $\mathcal{V} \rightarrow (\bar{\mathcal{V}}, T)$ . While it is possible to learn these mappings separately (i.e., $\mathcal{V} \rightarrow \bar{\mathcal{V}}$ and $\mathcal{V} \rightarrow T$ ), the expert commentary and demonstration are tightly related and provide important contextual information for generating each other. We therefore treat this as an autoregressive generation problem where during training, we use both $\mathcal{V}$ and $T$ to generate $\bar{\mathcal{V}}$ (or $\mathcal{V}$ and $\bar{\mathcal{V}}$ to generate $T$ ), while during testing, we drop the extra context information to generate $\bar{\mathcal{V}}$ and $T$ directly. + +Mathematically, consider $\mathcal{F}:(\mathcal{V},\bar{\mathcal{V}},T)\to \mathbb{R}$ as the autoregressive function that jointly learns from the learner demonstration $\nu$ , the expert demonstration $\tilde{\nu}$ , and the expert commentary $T$ . The output $\in \mathbb{R}$ is the loss that we aim to minimize. We use the same unified $\mathcal{F}$ for the joint training (detailed in Sec. 3.3) and inference, as follows. We use $\mathcal{F}'$ when the autoregressive model $\mathcal{F}$ is used for output token generation, i.e., text or pose tokens. $\mathcal{F}'$ can take in any input used during training and generate the remaining one—typical for autoregressive models [74]. + +Expert commentary generation. At inference time, to generate an expert commentary $T$ , we only have a learner demonstration, thus we mask out the expert demonstration and output the expert commentary: + +$$ +\hat {T} = \mathcal {F} _ {t} (\mathcal {V}) = \mathcal {F} ^ {\prime} (\mathcal {V}, \varnothing) \tag {1} +$$ + +where $\mathcal{F}_t$ denotes using the model to generate commentary. + +Expert demonstration generation. Next, we obtain the expert demonstration in two forms—retrieving a correct execution (video and pose) and generating a pose sequence. These two output formats offer complementary information to understand the actionable feedback. While a video exemplar is helpful for many learner mistakes, pose generation is useful in the absence of a correct expert demonstration in the retrieval set—allowing our model to generalize its coaching beyond the set of discrete expert videos. We leave expert video generation as a future work. + +Denoting $\mathcal{F}_r$ and $\mathcal{F}_g$ as the expert demonstration retrieval and expert pose generation functions, respectively, we have: + +$$ +\bar {\mathcal {V}} = \mathcal {F} _ {r} (\mathcal {V}) = \underset {\forall \mathcal {V} ^ {\prime} \in \mathcal {E}} {\operatorname {a r g m i n}} \mathcal {F} (\mathcal {V}, \mathcal {V} ^ {\prime}, \hat {T}) \tag {2} +$$ + +$$ +P = \mathcal {F} _ {g} (\mathcal {V}) = \mathcal {F} ^ {\prime} (\mathcal {V}, \hat {T}) \tag {3} +$$ + +where $\hat{T}$ is the output from Eq. 1 above, and $\mathcal{F}_r$ and $\mathcal{F}_g$ denote using the model to retrieve or generate the expert demonstration, respectively. + +In summary, the training step $\mathcal{F}$ uses tuples $\nu$ , $\bar{\nu}$ , $T$ in a unified way, described in Sec. 3.3, whereas, at inference, all the three functions $\mathcal{F}_t$ , $\mathcal{F}_r$ and $\mathcal{F}_g$ use only $\nu$ . + +# 3.2. Forming the expert feedback dataset + +To learn the desired functions, we need to obtain pairs of videos $(\mathcal{V},\bar{\mathcal{V}})$ where there is an error in the demonstration in the first video that is corrected in the second video, along with the expert commentary $T$ about $\mathcal{V}$ . Ego-Exo4D [29] offers a great starting point for our setup. It contains ego-exo videos, extracted 3D pose sequences, and time-stamped commentary by experts (e.g., professional soccer coaches) on the demonstrations in the video. The experts watched the entire video and stopped each time they saw something to critique or compliment, offering free-form spoken commentary. In total, this led to 117,812 sentences across 221 hours of video, with most videos commented by 2-5 unique experts [29]. See examples in Figure 2 (left) and details in Sec. 3.4 and Supp. + +We propose to automatically augment this data in two ways to enable coaching. First, we seek (pseudo-)annotations of whether a given commentary statement describes a needed improvement or applauds a correct execution. Second, we localize each piece of feedback on a body region, e.g., incorrect hand stretch vs. wrong legs movement. These distinctions are crucial to generate feedback + +![](images/c9480bfaf5667ee33ef0064ff330a334988e4ec43cc52d550c0e5681e4548c23.jpg) +Figure 2. Overview of the dataset creation. We first summarize the human-provided expert commentary [29] in one sentence using an LLM, and then map it to a body region and correct (green) or incorrect (red) execution label. We then choose incorrect-correct pairs for the same body region to obtain $\mathcal{C}$ . Finally, we choose pairs with minimum temporal alignment loss to obtain the training data. Best in zoom. + +and show corrections. To this end, we create a weakly-supervised training dataset consisting of tuples $(\mathcal{V}, T, \bar{\mathcal{V}})$ from Ego-Exo4D [29], as follows. + +Expert commentary classification and body localization. The commentary in [29] is obtained from voice recordings of the experts converted to text with ASR. Most of the samples contain extra comments like "oh, I will give this a five out of ten" and "that's how I would do it too". Additionally, as discussed above, they lack positive and negative labels and do not explicitly indicate the body region involved (e.g., legs, arms). Thus, we preprocess the expert commentary for three things—making the commentary concise to extract the improvable feedback, marking which body region the feedback is about, and marking whether the feedback states the need for improvement or not. + +Since these are all addressable with text reasoning, we use a large language model (LLM) to provide the desired answers. We use Llama3-70B [2], a recent open-sourced model that performs well in current benchmarks. In essence, given a commentary $T$ , the language model $\mathcal{L}$ responds to the prompt: "Given an expert's commentary, summarize the feedback into a single sentence and also provide which body regions need improvement or are correctly executed..." (see Supp.). Formally, this yields $\mathcal{L}(T) = (T', (b^1, c^j), \ldots, (b^s, c^s))$ where $T'$ is the concise summary and $b^i$ is a body region like head or arms and $c^i \in \{0, 1, 2\}$ are labels representing needs improvement, correct execution, and no mention, respectively. We group the skeleton joints into six pre-defined body regions $b^i$ (details in Supp.). See red and green boxes in Fig. 2. + +Pairing incorrect and correct executions. Next, we use the above information to mine pairs of incorrect and correct execution in the dataset. Ego-Exo4D also contains metadata about the skill-level $S$ of the demonstrator, broken into four categories (1-4 in increasing expertise) starting from + +novices who have not performed the activity before to late experts who have performed the same activity, e.g., basketball for $10+$ years. We sample incorrect learner demonstrations from beginners and correct demonstrations from experts from the same activity in Ego-Exo4D, e.g., penalty kicks in soccer or reverse layup in basketball. Even though there can be incorrect executions by experts and correct ones by beginners, errors by experts are likely incomparable to beginners' due to the skill gap. We use the mapping of body regions to find (in)correct executions referring to the same body region. This results in a collection of video pairs with negative and positive feedback about the same body region, e.g., legs in Fig. 2. Formally, the collection $\mathcal{C}$ is curated as + +$$ +\left\{\left(\mathcal {V} _ {1}, T, \mathcal {V} _ {2}\right) \mid S _ {1} \in \mathcal {S} ^ {n}, S _ {2} \in \mathcal {S} ^ {e}, \exists j s. t. c _ {1} ^ {j} = 0, c _ {2} ^ {j} = 1 \right\} +$$ + +where $S^n = \{\text{novice, early expert}\}$ and $S^e = \{\text{late expert, intermediate expert}\}$ . Using this matching, we obtain a collection of video pairs of incorrect and correct execution (Fig. 2, yellow box). + +Temporal alignment and filtering demonstrations. The commentary annotation process in [29] allows the experts to pause at any instant and provide their feedback. Thus, the clip $\mathcal{V}_1$ could have the start of a basketball jump shot, whereas the clip $\mathcal{V}_2$ could also have content before the shot and the follow-through. Thus, we temporally align the learner and the expert video in the collection $\mathcal{C}$ to ensure we are capturing the same step of a demonstration. + +For all pairs of clips in the collection $\mathcal{C}$ , we compare the two poses $P$ to quantify their alignment (similarity) using Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) [27], a translation-invariant and body shape-invariant measure. Note that the video component of the input (Sec. 4) is complementary, as it does capture the person's overall movement in the space. + +![](images/d03e703cd2fa192b4df0b944e99b3fed49db8bd6efe7c67f23b4572a3e62ce5c.jpg) +Figure 3. Model overview. We tokenize individual modalities using a modality-specific architecture (top). Once all the modalities are encoded as tokens, we use a large language model to learn expert commentary generation, demonstration retrieval, and pose generation. At inference, the model only takes the learner demonstration video $\nu$ . See text for details. + +![](images/5c2c3a868ade3c1644286079acdb2d28ce8eb640a9af282e0f5943bb611567e7.jpg) + +![](images/712ed670a0f8e50099c78ee6cf9d695f486b2e7c3c81027d8ae106bacb0a3300.jpg) + +Denote the commentary timestamp as $t_1$ and $t_2$ for the two videos. We first choose a fixed window around $t_1$ , say $[t_1^a, t_1^b]$ such that $t_1^a < t_1 < t_1^b$ . Next, we find the corresponding time window $[t_2^a, t_2^b]$ with $t_2^a < t_2 < t_2^b$ in $\mathcal{V}_2$ such that the PA-MPJPE error is the minimum. Specifically, + +$$ +[ t _ {2} ^ {a}, t _ {2} ^ {b} ] = \operatorname * {a r g m i n} _ {[ t _ {2} ^ {i}, t _ {2} ^ {j} ] \in \mathcal {V} _ {2}} \mathrm {P A - M P J P E} \left(\mathcal {V} _ {1} [ t _ {1} ^ {a}, t _ {1} ^ {b} ], \mathcal {V} _ {2} [ t _ {2} ^ {i}, t _ {2} ^ {j} ]\right) +$$ + +such that $|t_2^j - t_2^i| = |t_1^b - t_1^a|$ . + +We obtain a subset from $\mathcal{C}$ that contains the aligned video segments. There can be some video pairs where the demonstration can be very different and hence no appropriate match exists. Therefore, we only keep the top-k pairs with maximum alignment for every incorrect execution. + +Overall, we obtain a novel dataset $\mathcal{D}$ that contains pairs of videos where the first video has an incorrect execution (on some body region) and the second video corrects it, thus obtaining the desired tuple $(\mathcal{V}, T, \bar{\mathcal{V}})$ for training and testing. See Fig. 2 (right) and examples in Supp. To ensure fair evaluation, we separately establish a clean gold standard test set, free of potential noise from LLM inference. The test set is manually verified (see Supp for details). + +# 3.3. Architecture and training design + +Next, we discuss the architecture that encodes the videos $\mathcal{V}$ , poses $P$ , and text $T$ and enables training the auto-regressive function $\mathcal{F}$ . The overall idea is to encode all the representations into a text embedding space and use the strong capabilities of recent language models to obtain output text and pose tokens. This approach has been recently used in vision-conditioned language models [6, 35, 45, 55] for image and video captioning. This setup also allows for a unified architecture for various input and output combinations, as opposed to different input streams for individual modalities. Fig. 3 shows a schematic diagram of the architecture, + +and each part is explained below. + +Encoding video as tokens. Each video $\mathcal{V}$ is an ego-exo clip pair. $^1$ The input thus allows the model to see both close-up hand-object interactions (more visible in the learner's ego view) as well as full-body poses in the scene context (more visible in the observer's exo view). We use a pretrained video model to extract spatio-temporal features from both videos' frames, followed by a mapper to convert the video features to video tokens. Formally, $\mathbf{v} = f_{vm}(f_V(\mathcal{V}))$ where $f_{V}$ is a standard feature extractor (we use InternVideo2 [77]) and $f_{vm}$ is the visual mapper. See Fig. 3 (top left). The mapper is typically a low-parameter model that is trainable, whereas the high-parameter feature extractor is kept frozen. The mapper helps transform the visual mapper to be akin to text tokens—a popular strategy in visual instruction tuning models [35, 45]. + +Encoding pose sequences as tokens. To encode a pose sequence, we use a series of linear layers and MLP-mixer [25] to convert a single pose $P \in \mathbb{R}^{d \times 3}$ to an embedding that can be discretized using a codebook, i.e., $\mathbf{p}' = f_P(P)$ . The architecture in [25] also contains a decoder that converts the embeddings back to human poses, i.e., $P = f_{P}^{-1}(\mathbf{p}')$ , which we use in $\mathcal{F}_g$ to generate corrected poses. Similar to the above, we use a pose mapper to convert the tokens to embeddings. Formally, $\mathbf{p} = f_{pm}(f_P(P)) = f_{pm}(\mathbf{p}')$ is the pose token where $f_{pm}$ is the pose mapper. See Fig. 3 (top center). We concatenate embeddings from every frame to obtain the representation for the whole sequence. Having $f_{pm}$ allows training with fewer parameters and thus, we use $\mathbf{p}$ for learning $\mathcal{F}_t$ and $\mathcal{F}_r$ . However, using $f_{pm}$ for generation $\mathcal{F}_g$ would require adding an inverting function for $f_{pm}$ . Thus, we directly add $\mathbf{p}'$ to LLM tokens for pose generation $\mathcal{F}_g$ . An innovative as + +pect of our approach is to encode pose as multimodal tokens and train with LLMs, unlike other work regressing the pose parameters from a special pose embedding [20] or using a dedicated pose transformer [49, 82]. + +Encoding text as tokens. Text tokenization is the standard process before inputting a text sentence to the LLM [59, 73], i.e., $\mathbf{t} = f_t(T)$ where $f_{t}$ is the tokenization function. See Fig. 3 (top right). + +Multi-modal sequence prediction. The previous steps yield multimodal tokens $\mathbf{v}$ , $\mathbf{p}$ , and $\mathbf{t}$ for video, pose, and text, respectively. Next, we use the strong sequence prediction capabilities of large language models to obtain the desired output tokens based on the sequence of multi-modal inputs [6, 35, 45, 55]. + +For training $\mathcal{F}$ for expert commentary generation, we provide a sequence of learner video and pose tokens (v and p) and the corrected pose tokens ( $\bar{\mathbf{v}}$ and $\bar{\mathbf{p}}$ corresponding to $\bar{\nu}$ ). We ask the model to predict the expert commentary token sequence. For the sequence prediction language model $\mathcal{L}_s$ , we wish to obtain $\mathbf{t} = \mathcal{L}_s(\mathbf{v},\mathbf{p},\bar{\mathbf{v}},\bar{\mathbf{p}})$ . Consequently, the training objective is the standard cross-entropy loss, $\min_{\theta}\{-\log (\mathbf{t}|\mathbf{v},\mathbf{p},\bar{\mathbf{v}},\bar{\mathbf{p}};\theta)\}$ , where $\theta$ are the parameters of the model. See Fig. 3 (bottom left). For consistency with the training of these language models, the sequence is formatted to be conversational in nature, e.g., "provide an expert's commentary based on this pose sequence:". Similarly, for expert demonstration retrieval, we wish to obtain a retrieval candidate $\bar{\mathbf{p}} = \mathcal{L}_s(\mathbf{v},\mathbf{p},\mathbf{t})$ . Likewise, the training objective with parameters $\gamma$ is $\min_{\gamma}\{-\log (\bar{\mathbf{p}} |\mathbf{v},\mathbf{p},\mathbf{t};\gamma)\}$ , which we also use as the relevance score for retrieval during inference. Expert pose generation is trained with the same objective, except we use $\mathbf{p}'$ instead of $\mathbf{p}$ (and obtain $\bar{\mathbf{p}}'$ ). See Fig. 3 (bottom). Generating pose further requires converting back the pose tokens $\bar{\mathbf{p}}'$ to 3D joints using the pose decoder $f_P^{-1}$ . + +At inference time, as introduced in Sec. 3.1, we only use $\mathcal{V}$ as the input: $\mathcal{F}_t$ drops the expert demonstration and predicts the expert commentary $\hat{T}$ . We use this predicted commentary as input for $\mathcal{F}_r$ and $\mathcal{F}_g$ . + +# 3.4. Implementation details + +Dataset and statistics. Ego-Exo4D [29] contains 5,035 videos of participants doing activities across eight scenarios. We focus on three physical scenarios—basketball, soccer, and rock climbing—though our model is in principle generalizable to other physical skills, without any task-specific design, unlike [63, 78]. We use physical scenarios since the coaching feedback is groundable in the body regions, as opposed to procedural tasks like cooking, where the suggestions can be alternate ingredients or steps that are not visually present. Details and statistics are in Supp. Following the dataset creation process outlined in Sec. 3.2, we obtain a dataset of 25,505 training and 1,272 testing tuples + +of $(\mathcal{V}, T, \bar{\mathcal{V}})$ . We choose $k = 5$ (train) and $k = 1$ (test) for min-k choice of correct demonstrations per incorrect execution. To reiterate, we manually examine the test set for correct commentary summary and body region labeling. + +Network architecture. The video model $f_{V}$ is an InternVideo2 [77] encoder that provides strong visual representations. The pose encoder $f_{P}$ and decoder $f_{P}^{-1}$ are adapted from PCT [25], which learns compositional tokens from human poses. Since the original training in [25] uses 2D poses, we adapt it to use 3-dimensions and retrain on the human poses in [29]. Both the visual and pose mappers are low-parameter MLP layers [44]. Finally, we use Llama 3-8B [2] as the LLM multimodal encoder for sequence prediction. We finetune $\mathcal{F}_t$ and $\mathcal{F}_r$ for 5 epochs with a learning rate of $5 \times 10^{-5}$ . Next, we modify the token dimension in $\mathcal{F}_g$ to accommodate pose tokens $\mathbf{p}'$ , and hence, we fine-tune $\mathcal{L}_s$ when learning $\mathcal{F}_g$ with a learning rate of $5 \times 10^{-6}$ for 5 epochs. Video $(f_{V})$ and pose models $(f_{P}, f_{P}^{-1})$ are kept frozen. All the models are trained on 8 V100 32GB GPUs. + +# 4. Experiments and results + +We first discuss the baselines and ablations, followed by the evaluation setup and results for the three outputs—expert commentary generation $(\mathcal{F}_t)$ , expert demonstration retrieval $(\mathcal{F}_r)$ , and pose demonstration generation $(\mathcal{F}_g)$ . We also show qualitative examples and discuss the limitations. + +Baselines. We compare with the following baselines: + +- InternVideo2-NN, InternVideo2-FT [77]: Given a query video $\mathcal{V}$ , the nearest neighbor baseline (NN) finds the most similar video by InternVideo2 [77] feature similarity in the training data and returns the corresponding commentary or demonstration, for all tasks. The FT version finetunes the model to contrastively match the learner demonstration with the expert commentary/demonstration. +- VideoChat2 [36, 37], LLaVA [45]: In these methods, we prompt SOTA video and image captioning models to generate commentary for an input demonstration. We use the log-likelihood loss to find the retrieved expert demonstration. These baselines evaluate if existing video captioning methods can provide expert feedback or retrieve expert demonstrations. We use the ego and "best exo" frames, same as for our method. Note that neither of these baselines are applicable for pose generation $\mathcal{F}_g$ . +- LLaVA-FT [45], LLaVA-FT w/ pose [45]: These baselines are based on the SOTA visual-language method LLaVA [45] but trained on our dataset with the same text model, i.e., Llama 3-8B [2], for an apples-to-apples comparison. The "w/ pose" variant also takes the 3D pose coordinates in text form as input. +- PoseScript [16], PoseFix [17]: These two works enable pose-to-text and text-to-pose reasoning. The text generated or used for pose generation contains detail about the + +
MethodCommentary Gen.Demo Ret.Pose Gen.
B@4MR-LRmedR↓P↓
InternVideo2-NN [77]42.146.949.313.5198161
InternVideo2-FT [77]42.947.650.014.1190157
VideoChat2 [37]27.844.341.914.9183—
LLaVA [45]28.544.144.215.0183—
LLaVA-FT [45]43.548.551.517.8177—
LLaVA-FT w/ pose [45]43.648.551.718.0172150
PoseScript/Fix [16, 17]24.144.546.315.9182182
ExpertAF44.949.654.619.1158135
ExpertAF w/o video44.649.454.218.7161139
ExpertAF w/o pose44.249.354.218.6163—
w/o alignment42.048.651.516.9180153
w/ global pose43.047.952.616.5184150
ExpertAF w/ full-sup45.850.955.722.5146131
+ +![](images/157726add4d6f8b06b60e90cf3b0ed9405c16461026d5e6be8d4d15c5aa473c0.jpg) + +Table 1. Results on automatic metrics (left) and human evaluation (right). We break down results for the three outputs—expert commentary generation, expert demonstration retrieval, and expert pose generation. Our method outperforms all baselines and prior work on all tasks. The last row "w/ full-sup" uses privileged input (the demo video $\bar{\nu}$ ) at inference. (B@4: BLEU-4, M: Meteor, R-L: ROUGE-L F1, R: recall@50, medR: median rank, P: PA-MPJPE). For all metrics higher is better, except medR and PA-MPJPE ( $\downarrow$ ). Our method is also rated higher by human raters on a Likert scale (min:1, max:4), compared to all the other methods (right). See text for details. + +location of body parts, e.g., the hands are raised, as opposed to expert commentary. Hence, they let us evaluate if pose description is adequate for expert feedback. For pose generation, we evaluate if providing an expert commentary helps generate the desired expert demonstration. Both methods use SMPL [47] pose and hence we convert 3D pose to SMPL and vice versa [39], as needed. + +Ablations. In addition to the strong baselines, we also compare the performance against ablations and variants of our design choices. ExpertAF w/o pose and ExpertAF w/o video evaluate the performance when only one modality is used. ExpertAF w/o alignment quantifies the need for temporal alignment (Sec. 3.2), while ExpertAF w/ global pose evaluates if just providing the model with one correct pose per activity chosen based on expert commentary is enough (vs. finding the closest correct pose sequence). ExpertAF w/ full-sup is a stronger variant with privileged input for inference— $\mathcal{F}_t$ uses $\mathcal{V}$ , $\bar{\mathcal{V}}$ to predict $T$ and $\mathcal{F}_r$ and $\mathcal{F}_p$ uses both $\mathcal{V}$ and $T$ . See Supp. for additional ablations of the choice of LLM $\mathcal{L}_s$ , contribution of ego and exo videos, and joint training. + +Expert commentary generation. To evaluate the text commentary, we use standard metrics BLEU-4 [60], METEOR [7] and ROUGE-L F1 [65], following prior work in evaluating text generation [14, 58] (higher is better). + +Tab. 1 (left) shows the results. Our method outperforms all methods on all captioning metrics, with gains up to $3\%$ over the best baseline trained on the same dataset. All the gains are statistically significant using paired t-test with $p < 0.05$ . Fig. 4 (top) shows commentary generation by various methods. Firstly, SOTA methods LLaVA [45] and VideoChat2 [37], being captioning methods, yield a verbose description of the activity, rather than a critique. + +Fine-tuning LLaVA with our method improves the performance, but still it misses our use of expert demonstrations and pose sequences. Even using pose information as text only marginally improves "LLaVA-FT w/ pose", showing the effectiveness of our explicit pose encoding. Moreover, PoseScript [16] generates pose descriptions like "head is turned to the left" which is inadequate for feedback. Our results over the baselines show the advantage of temporal alignment and a diverse expert demonstration set. + +Our stronger variant "w/ full-sup" achieves an even better result by taking $\mathcal{V}$ and $\bar{\mathcal{V}}$ as inputs during inference. It is useful in cases where a learner has access to the expert demonstration but wants actionable feedback in languageform. Finally, Fig. 4 (bottom row) shows a failure case of commentary generation, showcasing the difficulty in pinpointing the exact mistake. + +Expert demonstration retrieval. For retrieval, we use standard retrieval metrics: recall@k and median rank. We set k=50 to account for the fact that multiple expert demonstrations could suitably correct the given learner demonstration. The retrieval set contains 1,272 samples, the same as the test set size. Higher recall and lower ranks are better. + +Tab. 1 (middle) shows the results. Our method improves the median rank by 14 positions compared to the best baseline and a significantly better recall@50. Our method and the baselines show a similar trend as with the expert commentary generation above, the stronger baseline w/ "full-sup" ( $>4\%$ better recall and 26 rank improvement) being crucial in cases where a learner has the commentary but wants a video exemplar to learn from. + +Expert pose generation. Tab. 1 (right column) shows the expert pose generation results, with error measured by PA-MPJPE [27] (mm). Our novel pose generation method + +![](images/69cdc67194d5ff823d450d701bf899b2ddf19993421584b8a7a9b21f62352b24.jpg) + +# Ground truth + +The climber is exhausted and struggling to make a move, with poor hand and arm positioning. + +# Ours + +The climber is struggling to maintain the position and is unable to make the next move. + +# LLaVA-FT + +The climber is wasting energy by pulling away, which is slowing down reaching the next handhold + +# VideoChat2 + +The person can improve by practicing their technique, warming up and stretching. + +# PoseScript + +The person is standing on their left leg with the right leg lifted up and bent at the knee. The torso is leaning to the left. The head is turned to the left. + +![](images/e45b0af9cb986232e56d87bf322ba561c167774408777408ef1a405e3ea98f90.jpg) +Learner demonstration + +![](images/91bbee65fbc312ff8144dc85b7f6bbca3b34431f76fd0d93c3c533035c6fb734.jpg) + +![](images/47b2c8773f2af44cb3acb144c8785a4f23033f6d73e01871f269410cf0265e59.jpg) +Failure case +Figure 4. Qualitative results. (Top) Comparison of expert commentary generated by various baselines. (Second and third row) Examples of expert commentary generation, demonstration retrieval, and pose generation by our method. Notice the expert demonstration and pose generation corrects the mistake pointed out in the commentary, i.e., one hand is used to throw and the other to guide, and body position is improved to control better control and power. Colored text and marks (red for mistake, green for correction) are shown only for visualization. (Bottom) Failure cases. See Supp. for video results. + +# Expert comment. - Ours + +The shooter's stance is correct, but they need to throw the ball with one hand and guide it with the other. + +The player's stance is too straight, which prevents them from generating the required power and control on the ball. + +The shooter's left elbow is bent and his wrist is not fully extended, affecting the ball's rotation and accuracy. + +# Ground truth commentary + +The shooter is using two hands to shoot the ball, instead of using one shooting hand and arm with the other at a 90-degree angle. + +The player's foot position and body lean are affecting the quality of their ball contact, leading to a loss of power and control. + +He does a good job using his left hand on the left side, but needs to improve his posture and weight distribution. + +![](images/7d3a9e59a065beb6e33d6a2f25173115e85cd9a1cadf8fa06833e863f479c951.jpg) +Expert demo. retrieval + +![](images/1565ada75121f6b2de9c9aa1170d1b34c52cad644ca1380f5f6f78771bdcffc1.jpg) + +![](images/2129b9d6136ad20fd3630a97e35df46c6750cb1ae7d1aa44ecaf0f3d371a4a77.jpg) + +![](images/01db810c9d6d4ad3df6c95bb51a86d8a84de8f7480d187b60710c63e9f860d0e.jpg) +Pose generation + +![](images/b4720fc1e800611ef2157eec73a90bbfda8e653bbe71df5edced359ff33f9bd8.jpg) + +![](images/e50ab803604f6adbb398f2cbdfe71aefb09dfff0d31a964c3fa47e28de3ff05c.jpg) + +performs better than learning 3D position using text in "LLaVA-FT w/ pose". Furthermore, PoseFix [17] is designed to generate SMPL parameters based on a coarse modification text description. Our use of generation to output poses (vs. retrieval) is advantageous when the candidate retrieval set does not contain the correct demonstration; see Supp. See Fig. 4 and Supp. for qualitative examples including failure cases. The 3D pose sequences in the dataset are often auto-generated and have some noisy samples, discussed in Supp. Across the three tasks, our method is superior to its ablations. + +Human evaluation. For all the tasks, we also solicit human evaluation (Tab. 1 right) to rate the quality of the generated text descriptions. Five raters (per scenario) uninvolved with this project rate the quality of each generated output by scoring on a Likert scale from 1 to 4 (higher is better); details in Supp. We ensure that the raters have a basic knowledge of the scenario they are rating (basketball, soccer, rock climbing). + +Across all three tasks, human raters score our method the highest. A significant gap of up to 0.7 over the possible range $(1 - 4)$ , as well as gains up to $3\times$ , showcase ExpertAF's expert feedback quality. That said, there is nat + +urally room for improvement on this new task. We find that while humans say our model excels on cases where the feedback is visually groundable, e.g. incorrect hand angle, it tends to fall short when the feedback is not directly visible, e.g. the climber looks fatigued. The human evaluation complements the automatic metrics above, overcoming the limitation that a "ground truth" commentary or a visual demonstration may capture only one of multiple possible errors in the learner's demonstration [23, 68] (e.g., for a basketball shot, both the hand placement and the jump could be incorrect, but only one may be mentioned in the ground truth). + +# 5. Conclusion + +We proposed a novel task and method to generate expert commentary and demonstrations from a learner's video. We develop a weakly-supervised training approach and benchmark for this problem. Our novel method fusing multimodal inputs from learner and expert demonstrations together with expert commentary results in state-of-the-art performance in actionable feedback, and lays the groundwork for accessible, affordable, and actionable AI coaching applications in the future. + +Acknowledgement. UT Austin is supported in part by the IFML NSF AI Institute. Thanks to Fu-Jen Chu, Jing Huang and Xitong Yang for help with the Exo-Ego4D [29] pose extraction pipeline. + +# References + +[1] Yazan Abu Farha, Alexander Richard, and Juergen Gall. When will you do what?-anticipating temporal occurrences of activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5343-5352, 2018. 2 +[2] AI@Meta. Llama 3 model card, 2024. 4, 6, 1 +[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23066-23078, 2023. 2 +[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. What you say is what you show: Visual narration detection in instructional videos. arXiv preprint arXiv:2301.02307, 2023. 2 +[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[6] Kumar Ashutosh, Zihui Xue, Tushar Nagarajan, and Kristen Grauman. Detours for navigating instructional videos. In CVPR, 2024. 1, 2, 5, 6 +[7] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7 +[8] Tobias Baumgartner and Stefanie Klatt. Monocular 3d human pose estimation for sports broadcasts using partial sports field registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5108-5117, 2023. 2 +[9] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the IEEE international conference on computer vision, pages 2177-2185, 2017. 1, 2 +[10] Gedas Bertasius, Aaron Chan, and Jianbo Shi. Egocentric basketball motion planning from a single first-person image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[11] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 2 +[12] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya + +Ramesh. Video generation models as world simulators, 2024. 2 +[13] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 2 +[14] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. arXiv preprint arXiv:2305.18500, 2023. 7 +[15] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos. International Journal of Computer Vision, 129:2846-2864, 2021. 2 +[16] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022. 2, 6, 7 +[17] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2, 6, 7, 8 +[18] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 319-335. Springer, 2022. 2 +[19] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Posegpt: Chatting about 3d human pose. arXiv preprint arXiv:2311.18836, 2023. 2 +[20] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. In CVPR, 2024. 2, 6 +[21] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence, 43(11):4021-4036, 2020. 2 +[22] Jiyang Gao, Zhenheng Yang, and Ram Nevatia. Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818, 2017. 2 +[23] Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal, Pawan Sasanka Ammanamanchi, Anuoluwapo Aremu, Antoine Bosselut, Khyathi Raghavi Chandu, Miruna-Adriana Clinciu, Dipanjan Das, Kaustubh Dhole, Wanyu Du, Esin Durmus, Ondrej Dušek, Chris Chinenye Emezue, Varun Gangal, Cristina Garbacea, Tatsunori Hashimoto, Yufang Hou, Yacine Jernite, Harsh Jhamtani, Yangfeng Ji, Shailza Jolly, Mhir Kale, Dhruv Kumar, Faisal Ladhak, Aman Madaan, Mounica Maddela, Khyati Mahajan, Saad Mahamood, Bodhisattwa Prasad Majumder, Pedro Henrique Martins, Angelina McMillan-Major, Simon Mille, Emiel van Miltenburg, Moin Nadeem, Shashi Narayan, Vitaly Nikolaev, Andre Niyongabo Rubungo, Salomey Osei, Ankur + +Parikh, Laura Perez-Beltrachini, Niranjan Ramesh Rao, Vikas Raunak, Juan Diego Rodriguez, Sashank Santhanam, João Sedoc, Thibault Sellam, Samira Shaikh, Anastasia Shimorina, Marco Antonio Sobrevilla Cabezudo, Hendrik Strobelt, Nishant Subramani, Wei Xu, Diyi Yang, Akhila Yerukola, and Jiawei Zhou. The GEM benchmark: Natural language generation, its evaluation and metrics. In Proceedings of the 1st Workshop on Natural Language Generation, Evaluation, and Metrics (GEM 2021), pages 96-120, Online, 2021. Association for Computational Linguistics. 8 +[24] Adelaide Genay, Anatole Lécuyer, and Martin Hachet. Being an avatar "for real": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 28(12):5071-5090, 2021. 2 +[25] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 660-671, 2023. 2, 5, 6 +[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 2 +[27] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 4, 7 +[28] Andrew S Gordon. Automated video assessment of human performance. In Proceedings of AI-ED, page 10, 1995. 1, 2 +[29] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In CVPR, 2024. 1, 2, 3, 4, 6, 9 +[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2 +[31] William Hoff and Tyrone Vincent. Analysis of head pose accuracy in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 6(4):319-334, 2000. 2 +[32] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2 +[33] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2 +[34] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7718-7727, 2019. 2 + +[35] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 5, 6 +[36] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6 +[37] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark, 2023. 6, 7 +[38] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 2 +[39] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yukun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv preprint arXiv:2304.00553, 2023. 7 +[40] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In NeurIPS, 2022. 2 +[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3 +[42] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2 +[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9522-9531, 2021. 3 +[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6 +[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 5, 6, 7 +[46] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2 +[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph., 34(6), 2015. 7 + +[48] Sabine Losch, Eva Traut-Mattausch, Maximilian D Muhlberger, and Eva Jonas. Comparing the effectiveness of individual coaching, self-coaching, and group training: How leadership makes the difference. Frontiers in psychology, 7: 175595, 2016. 1 +[49] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 6 +[50] Priyanka Mandikal and Kristen Grauman. Dexvip: Learning dexterous grasping with human hand pose priors from video. In Conference on Robot Learning, pages 651-661. PMLR, 2022. 2 +[51] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action forecasting@ ego4d challenge 2022. arXiv preprint arXiv:2207.12080, 2022. 2 +[52] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations, 2023. 2 +[53] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2 +[54] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2 +[55] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. In CVPR, 2024. 5, 6 +[56] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 1 +[57] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. arXiv preprint arXiv:2207.09666, 2022. 2 +[58] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. In European Conference on Computer Vision, pages 167–184. Springer, 2022. 7 +[59] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fo + +tis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mély, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O’Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael. Pokorny. Michelle Pokrass. Vitchyr H. Pong Tolly Powell Alethea Power Boris Power Elizabeth Proehl Raul Puri Alec Radford Jack Rae Aditya Ramesh Cameron Raymond Francis Real Kendra Rimbach Carl Ross Bob Rotsted Henri Roussez Nick Ryder Mario Saltarelli Ted Sanders Shibani Santurkar Girish Sastry. Heather Schmidt David Schnurr John Schulman Daniel Selsam. Kyla Sheppard Toki Sherbakov Jessica Shieh Sarah Shoker Pranav Shyam Szymon Sidor Eric Sigler Maddie Simens Jordan Sitkin Katarina Slama Ian Sohl Benjamin Sokolowsky Yang Song Natalie Staudacher Felipe Petroski Such Natalie Summers Ilya Sutskever Jie Tang. Nikolas Tezak Madeleine B. Thompson Phil Tillet. Amin Tootoonchian Elizabeth Tseng Preston Tuggle Nick Turley Jerry Tworek Juan Felipe Cerón Uribe Andrea Vallone Arun Vijayvergiya Chelsea Voss Carroll Wainwright Justin Jay Wang Alvin Wang Ben Wang Jonathan Ward Jason Wei CJ Weinmann Akila Welihinda Peter Welin + +der, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report, 2024. 6 +[60] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7 +[61] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[62] Paritosh Parmar, Jaiden Reddy, and Brendan Morris. Piano skills assessment. In 2021 IEEE 23rd international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2021. 3 +[63] Paritosh Parmar, Amol Gharat, and Helge Rhodin. Domain knowledge-informed self-supervised representations for workout form assessment. In European Conference on Computer Vision, pages 105-123. Springer, 2022. 2, 6 +[64] Hamed Pirsiavash, Carl Vondrick, and Antonio Torralba. Assessing the quality of actions. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 556-571. Springer, 2014. 2 +[65] Lin CY Rouge. A package for automatic evaluation of summaries. In Proceedings of Workshop on Text Summarization of ACL, Spain, 2004. 7 +[66] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2 +[67] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2 +[68] Chenhui Shen, Liying Cheng, Xuan-Phi Nguyen, Yang You, and Lidong Bing. Large language models are not yet human-level evaluators for abstractive summarization. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 4215–4233, 2023. 8 +[69] Sumedh Sontakke, Jesse Zhang, Séb Arnold, Karl Pertsch, Erdem Biryik, Dorsa Sadigh, Chelsea Finn, and Laurent Itti. Roboclip: One demonstration is enough to learn robot policies. Advances in Neural Information Processing Systems, 36, 2024. 2 +[70] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2 +[71] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9839-9848, 2020. 2 +[72] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4790-4799, 2023. 2 +[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 6 +[74] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[75] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2 +[76] Tianyu Wang, Yijie Wang, and Mian Li. Towards accurate and interpretable surgical skill assessment: A video-based method incorporating recognized surgical gestures and skill levels. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 668-678. Springer, 2020. 3 +[77] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 5, 6, 7 +[78] Chengming Xu, Yanwei Fu, Bing Zhang, Zitian Chen, YuGang Jiang, and Xiangyang Xue. Learning to score figure skating sport videos. IEEE Transactions on Circuits and Systems for Video Technology, 30(12):4578-4590, 2020. 1, 2, 6 +[79] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training + +for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2 +[80] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 2 +[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2 +[82] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 6 +[83] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129:703-718, 2021. 2 +[84] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 2 +[85] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2 +[86] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2 +[87] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2 \ No newline at end of file diff --git a/2025/ExpertAF_ Expert Actionable Feedback from Video/images.zip b/2025/ExpertAF_ Expert Actionable Feedback from Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2845fc1c4a1b868fb934dc529b72f5286b6359c0 --- /dev/null +++ b/2025/ExpertAF_ Expert Actionable Feedback from Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a59dd05e9d5c10e62678ffa98ce80673310181e5bb91103aeb7650c193217f3 +size 418075 diff --git a/2025/ExpertAF_ Expert Actionable Feedback from Video/layout.json b/2025/ExpertAF_ Expert Actionable Feedback from Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9c4211c9fc2516977a0a97cbfd94d29a9ee2707c --- /dev/null +++ b/2025/ExpertAF_ Expert Actionable Feedback from Video/layout.json @@ -0,0 +1,12474 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 103, + 465, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 103, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 145, + 103, + 465, + 121 + ], + "type": "text", + "content": "ExpertAF: Expert Actionable Feedback from Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "spans": [ + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": "Kumar Ashutosh" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": ", Tushar Nagarajan" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": ", Georgios Pavlakos" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": ", Kris Kitani" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": ", Kristen Grauman" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": " University of Texas at Austin, " + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": " FAIR Meta, " + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 74, + 142, + 536, + 173 + ], + "type": "text", + "content": " Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 224, + 296, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 488 + ], + "type": "text", + "content": "Feedback is essential for learning a new skill or improving one's current skill-level. However, current methods for skill-assessment from video only provide scores or compare demonstrations, leaving the burden of knowing what to do differently on the user. We introduce a novel method to generate actionable feedback (AF) from video of a person doing a physical activity, such as basketball or soccer. Our method takes a video demonstration and its accompanying 3D body pose and generates (1) free-form expert commentary describing what the person is doing well and what they could improve, and (2) a visual expert demonstration that incorporates the required corrections. We show how to leverage Ego-Exo4D's [29] videos of skilled activity and expert commentary together with a strong language model to create a weakly-supervised training dataset for this task, and we devise a multimodal video-language model to infer coaching feedback. Our method is able to reason across multi-modal input combinations to output full-spectrum, actionable coaching—expert commentary, expert video retrieval, and expert pose generation—outperforming strong vision-language models on both established metrics and human preference studies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 509, + 135, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 135, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 135, + 521 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": "An abundance of instructional \"how-to\" videos on the internet enables skill-learning by observing expert demonstrations. Instructional videos cover many skills one might want to learn—cooking, DIY, sports, crafts, and hobbies. Recent work leverages such videos to assist human [5, 6, 53] and robot [56] skill learning. Acquiring skills from video is especially appealing for equitable learning: access to video is generally less costly and more widely available than face-to-face access to an expert coach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 637, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 696 + ], + "type": "text", + "content": "Despite the myriad of videos and tutorials, people prefer to learn in a feedback loop: getting to know their mistakes, finding the improvements they need to do and correcting those. For that reason, coaching-based iterative training is more effective than self-learning [48]. A good coach has" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 198, + 421, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 421, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 421, + 207 + ], + "type": "text", + "content": "Learner demonstration (input)" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 209, + 445, + 286 + ], + "blocks": [ + { + "bbox": [ + 316, + 209, + 445, + 286 + ], + "lines": [ + { + "bbox": [ + 316, + 209, + 445, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 445, + 286 + ], + "type": "image", + "image_path": "1e3b994c58c5bc89cce539628bb1eff839d9ae97b08bdb44e825cd5fb66abfed.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 335, + 289, + 449, + 344 + ], + "blocks": [ + { + "bbox": [ + 335, + 289, + 449, + 344 + ], + "lines": [ + { + "bbox": [ + 335, + 289, + 449, + 344 + ], + "spans": [ + { + "bbox": [ + 335, + 289, + 449, + 344 + ], + "type": "image", + "image_path": "6038d690b7c7e225445bee77186e627ab34e311406aef7fba9c5b97de97afd6b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 449, + 198, + 552, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 449, + 198, + 552, + 207 + ], + "spans": [ + { + "bbox": [ + 449, + 198, + 552, + 207 + ], + "type": "text", + "content": "Expert demonstration (output)" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 449, + 209, + 564, + 286 + ], + "blocks": [ + { + "bbox": [ + 449, + 209, + 564, + 286 + ], + "lines": [ + { + "bbox": [ + 449, + 209, + 564, + 286 + ], + "spans": [ + { + "bbox": [ + 449, + 209, + 564, + 286 + ], + "type": "image", + "image_path": "edfaf44dfd1543af4a8afffc07c108f5412bac965a4bf6da4551e53bf82cd72a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 457, + 287, + 548, + 344 + ], + "blocks": [ + { + "bbox": [ + 457, + 287, + 548, + 344 + ], + "lines": [ + { + "bbox": [ + 457, + 287, + 548, + 344 + ], + "spans": [ + { + "bbox": [ + 457, + 287, + 548, + 344 + ], + "type": "image", + "image_path": "8bc25e3a2b530b15ffb32fbb177680dc5304d424b8a3fb10dd20934ec3e4ede3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 347, + 417, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 347, + 417, + 357 + ], + "spans": [ + { + "bbox": [ + 319, + 347, + 417, + 357 + ], + "type": "text", + "content": "Expert commentary (output)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 362, + 552, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 362, + 552, + 384 + ], + "spans": [ + { + "bbox": [ + 326, + 362, + 552, + 384 + ], + "type": "text", + "content": "The player is taking big steps to control the ball, but lacks body control, and needs to slow down to maintain a better control." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. An example of expert feedback. When a player is dribbling the ball fast, they tend to lose control (top left). Our proposed method provides an expert commentary to the learner suggesting improvements (bottom). The method also provides an expert demonstration that shows the desired correction, where the player is maintaining smaller steps and body control (top right)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 471, + 554, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 542 + ], + "type": "text", + "content": "three components: finding mistakes, providing verbal corrections and finally, showing visual demonstrations. For example, suppose a person learning to dribble a soccer ball takes big steps that decrease body control. The coach should identify this mistake, provide verbal feedback, and show the correct dribbling technique. See Figure 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": "At present, a person watching a how-to video has to identify mistakes themselves and attempt to correct them. Pinpointing the exact mistake is itself challenging for beginners—particularly when learning physical skills, like sports and dance, where subtle movements and positioning of the body are critical to success." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "What role could AI play in overcoming this gap? Current work on proficiency understanding from video focuses on scoring a demonstration [9, 28, 29, 78], thus addressing only one component of coaching. There is no method that can provide actionable verbal and visual feedback on a learner's demonstration. Doing so is technically challenging since it requires the method to (a) understand the user's activity, (b) detect mistakes in the execution by comparing" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 703, + 266, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 266, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 266, + 713 + ], + "type": "text", + "content": "Project page: https://vision.cs.utexas.edu/projects/ExpertAF/" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13582" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": "the user's activity with correct examples and finally, (c) suggest edits with respect to the better way of doing the activity. In short, the output should be coach-like—specific actionable feedback that will improve a learner's proficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 128, + 295, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 224 + ], + "type": "text", + "content": "We propose ExpertAF: a novel method to provide expert actionable feedback on a user's physical activity. ExpertAF's actionable feedback has two forms: language-based commentary and visual demonstration. The freeform expert commentary output describes the mistake and what should be corrected, e.g., \"take smaller steps and slow down to maintain control\", while the expert visual demonstration output shows the correct way of doing it." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 434 + ], + "type": "text", + "content": "Both components introduce unique technical challenges. For the former, the task of generating free-form text feedback is distinct from standard image and video captioning, as the model needs to identify the mistake and actions that would correct it—rather than simply describe the observed pose [16] or activity [32, 38, 57, 75, 81]. Similarly, for the latter, the task of retrieving (or generating) a demonstration that corrects a specific mistake is distinct from similarity-based retrieval [3, 40, 54, 77, 79] or open-ended generation [12, 19]. We hypothesize—and experimentally validate—that personalized, video-conditioned coaching is more accurate than simply returning a global expert execution because there are many right ways of doing an activity, e.g., penalty shots in soccer, and hence different ways to fix what is wrong in a given execution. Overall, the model must understand both what is being done and what actionable tweaks would make it better." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 598 + ], + "type": "text", + "content": "We are the first to address video-based coaching with actionable feedback. We develop the first model to generate free-form text feedback on a user's activity (captured in video and 3D pose sequences) as well as a video demonstrating how to improve, by building on a recent vision-language model [45]. There is currently no dataset for this challenging task. Therefore, we augment Ego-Exo4D [29], an existing dataset with video, commentary, and poses, and show how to use a strong large language model to create a weakly-supervised training set consisting of paired learner-expert demonstrations, along with expert textual commentary that relates the two. We also obtain a gold-standard manually labeled test set for rigorous evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "We validate ExpertAF on three diverse physical scenarios: soccer, basketball, and rock climbing. Our results show a significant improvement over strong baselines, establishing the first method to provide actionable feedback, including a novel technique for expert pose generation. Alongside consistent gains in quantitative metrics against the ground truth, we also show our model's promise via direct human evaluation, where ExpertAF outperforms off-the-shelf video models by as much as " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 317 + ], + "type": "text", + "content": "Skill learning from videos. Instructional video datasets like HowTo100M [53], COIN [70], and CrossTask [87] enable procedural understanding through step recognition [5, 42, 85], procedure planning [11, 13, 84], task-graph discovery [5, 18, 29, 85], action anticipation [1, 21, 22, 26, 51], and alignment detection [4, 30]. Capitalizing on their instructional nature, recent work learns robot policies from these videos [50, 69], uses external knowledge-bases like WikiHow to ground the steps of a procedural task [42, 52, 85], and explores new ways to navigate between multiple demonstrations [6]. Despite their scale, how-to videos in these datasets [53, 70, 87] are often created by domain experts and hence lack mistakes or suboptimal executions. In contrast, Ego-Exo4D [29] contains multiple executions of various common scenarios by people with varied skill levels—from novice to late expert. Moreover, it contains experts' feedback on those actions. We show how to augment this data via large language models to support our new idea for video-based coaching." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 319, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 557 + ], + "type": "text", + "content": "3D body pose for activity understanding. 3D body pose is crucial for human activity understanding, capturing a person's stance and motion. Supported by valuable datasets like Humans3.6M [33] and NTU RGB-D [46, 66], recent methods improve body pose understanding [25, 34, 67, 72, 83, 86]. Beyond action classification, pose is also crucial for markerless motion capture of sports [8, 15] and interacting in augmented reality [24, 31]. While our ExpertAF leverages pose as an important signal of human activity, unlike prior work we interface video, pose, and free-form text to generate expert feedback and an expert demonstration video. Our work is also different from generating [20] or modifying [17] body pose from text (e.g., \"raise your left arm\"), as our task is to understand a potentially incorrect pose and provide feedback. Understanding these minor differences between incorrect and correct poses offers unique challenges that are addressed by our proposed learning scheme. Our work is orthogonal to methods that improve pose estimation itself; any future improvements on that front would only benefit our approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Skill assessment and coaching. Prior work explores skill assessment for a variety of tasks, particularly sports. Most methods pose skill-assessment as a score-prediction task, i.e., for skating [78], gymnastics [28], basketball [9], or multiple sports [61]. Since absolute scoring can be ambiguous and error-prone, recent work [71] explores uncertainty-aware score distributions. Instead of explicitly scoring a demonstration, other work determines which of two basketball players are better [9], uses group-aware contrastive regression to learn the relative quality of demonstrations [80], or generates a full ideal trajectory from the first frame [10]. Fitness-AQA [63] and Action Quality [64] provide outputs like knees inward error, shallow squat error," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13583" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "or an arrow highlight for more localized feedback. There is also research on skill assessment in non-sports domains, like surgery [43, 76] and piano [62]. All the prior work assumes a fixed taxonomy of errors, and the taxonomy is designed separately for each activity. Furthermore, unlike our work, none of the prior work provides feedback akin to a personal coach—which requires not only telling what is wrong, but also expressing how to correct the mistake." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 111, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 111, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 111, + 189 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 197, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 232 + ], + "type": "text", + "content": "We introduce the problem statement in Sec. 3.1, the dataset creation strategy in Sec. 3.2, the training design in Sec. 3.3, and implementation details in Sec. 3.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 241, + 167, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 167, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 167, + 252 + ], + "type": "text", + "content": "3.1. Problem statement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": "Consider a dataset " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{(\\mathcal{V}, T, \\bar{\\mathcal{V}})\\}" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " where each " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " is a video demonstration, and " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " is a free-form text commentary critiquing the activity in the video, e.g., \"take smaller steps and slow down to maintain control.\" " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " is a related video demonstration but without the error mentioned in " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": ", e.g., the player takes smaller steps with better control. See Fig. 1. Each demonstration " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{V, P, S\\}" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " consists of three parts: " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " is the RGB video, and " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " is the 3D pose sequence of the participant with skill-level " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " in the video. The pose representation " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{n \\times d \\times 3}" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " frames of 3D positions for " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "d = 17" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " body joints, consistent with MS-COCO [41]. " + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 258, + 296, + 449 + ], + "type": "text", + "content": " is the participant's skill-level for an activity, i.e., novice, early expert, intermediate expert, or late expert, and will be used later in creating the dataset. We assume there is one active person doing a physical activity in any given video who is the subject of ExpertAF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": "The goal in this work is to provide feedback on a given learner's video containing a physical scenario. The feedback is an expert demonstration " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " and an expert commentary " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": ", tailored to " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": ". Formally, we want to learn a mapping " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\rightarrow (\\bar{\\mathcal{V}}, T)" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": ". While it is possible to learn these mappings separately (i.e., " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\rightarrow \\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\rightarrow T" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": "), the expert commentary and demonstration are tightly related and provide important contextual information for generating each other. We therefore treat this as an autoregressive generation problem where during training, we use both " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": "), while during testing, we drop the extra context information to generate " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 450, + 295, + 593 + ], + "type": "text", + "content": " directly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Mathematically, consider " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}:(\\mathcal{V},\\bar{\\mathcal{V}},T)\\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " as the autoregressive function that jointly learns from the learner demonstration " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " , the expert demonstration " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\nu}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " , and the expert commentary " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " . The output " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " is the loss that we aim to minimize. We use the same unified " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " for the joint training (detailed in Sec. 3.3) and inference, as follows. We use " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " when the autoregressive model " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " is used for output token generation, i.e., text or pose tokens. " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " can take in any input used during training and generate the remaining one—typical for autoregressive models [74]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "Expert commentary generation. At inference time, to generate an expert commentary " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ", we only have a learner demonstration, thus we mask out the expert demonstration and output the expert commentary:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 383, + 126, + 553, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 126, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 383, + 126, + 553, + 140 + ], + "type": "interline_equation", + "content": "\\hat {T} = \\mathcal {F} _ {t} (\\mathcal {V}) = \\mathcal {F} ^ {\\prime} (\\mathcal {V}, \\varnothing) \\tag {1}", + "image_path": "67d624576381361391600e55d0f6f42dbaa1865427215d9d8e9f486baf723aac.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 147, + 553, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 159 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 147, + 553, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 147, + 553, + 159 + ], + "type": "text", + "content": " denotes using the model to generate commentary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 159, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 277 + ], + "type": "text", + "content": "Expert demonstration generation. Next, we obtain the expert demonstration in two forms—retrieving a correct execution (video and pose) and generating a pose sequence. These two output formats offer complementary information to understand the actionable feedback. While a video exemplar is helpful for many learner mistakes, pose generation is useful in the absence of a correct expert demonstration in the retrieval set—allowing our model to generalize its coaching beyond the set of discrete expert videos. We leave expert video generation as a future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "text", + "content": "Denoting " + }, + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r" + }, + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "text", + "content": " as the expert demonstration retrieval and expert pose generation functions, respectively, we have:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 362, + 320, + 553, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 320, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 362, + 320, + 553, + 341 + ], + "type": "interline_equation", + "content": "\\bar {\\mathcal {V}} = \\mathcal {F} _ {r} (\\mathcal {V}) = \\underset {\\forall \\mathcal {V} ^ {\\prime} \\in \\mathcal {E}} {\\operatorname {a r g m i n}} \\mathcal {F} (\\mathcal {V}, \\mathcal {V} ^ {\\prime}, \\hat {T}) \\tag {2}", + "image_path": "2d0a4eadf2c8166ad8a640620dde27921bb9bc0443907737bd93117226723017.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 344, + 553, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 344, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 362, + 344, + 553, + 358 + ], + "type": "interline_equation", + "content": "P = \\mathcal {F} _ {g} (\\mathcal {V}) = \\mathcal {F} ^ {\\prime} (\\mathcal {V}, \\hat {T}) \\tag {3}", + "image_path": "983c477c9bbb445e52b914d5fb87bafc1838fd714e291c8e6845152f95393b6a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": " is the output from Eq. 1 above, and " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": " denote using the model to retrieve or generate the expert demonstration, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": "In summary, the training step " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": " uses tuples " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\bar{\\nu}" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": " in a unified way, described in Sec. 3.3, whereas, at inference, all the three functions " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": " use only " + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 403, + 553, + 438 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 445, + 508, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 508, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 508, + 457 + ], + "type": "text", + "content": "3.2. Forming the expert feedback dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "text", + "content": "To learn the desired functions, we need to obtain pairs of videos " + }, + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "inline_equation", + "content": "(\\mathcal{V},\\bar{\\mathcal{V}})" + }, + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "text", + "content": " where there is an error in the demonstration in the first video that is corrected in the second video, along with the expert commentary " + }, + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "text", + "content": " about " + }, + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 462, + 553, + 629 + ], + "type": "text", + "content": ". Ego-Exo4D [29] offers a great starting point for our setup. It contains ego-exo videos, extracted 3D pose sequences, and time-stamped commentary by experts (e.g., professional soccer coaches) on the demonstrations in the video. The experts watched the entire video and stopped each time they saw something to critique or compliment, offering free-form spoken commentary. In total, this led to 117,812 sentences across 221 hours of video, with most videos commented by 2-5 unique experts [29]. See examples in Figure 2 (left) and details in Sec. 3.4 and Supp." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "We propose to automatically augment this data in two ways to enable coaching. First, we seek (pseudo-)annotations of whether a given commentary statement describes a needed improvement or applauds a correct execution. Second, we localize each piece of feedback on a body region, e.g., incorrect hand stretch vs. wrong legs movement. These distinctions are crucial to generate feedback" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13584" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 549, + 238 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 549, + 238 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 549, + 238 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 549, + 238 + ], + "type": "image", + "image_path": "c9480bfaf5667ee33ef0064ff330a334988e4ec43cc52d550c0e5681e4548c23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "type": "text", + "content": "Figure 2. Overview of the dataset creation. We first summarize the human-provided expert commentary [29] in one sentence using an LLM, and then map it to a body region and correct (green) or incorrect (red) execution label. We then choose incorrect-correct pairs for the same body region to obtain " + }, + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "type": "text", + "content": ". Finally, we choose pairs with minimum temporal alignment loss to obtain the training data. Best in zoom." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 287, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 295, + 323 + ], + "type": "text", + "content": "and show corrections. To this end, we create a weakly-supervised training dataset consisting of tuples " + }, + { + "bbox": [ + 55, + 287, + 295, + 323 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}, T, \\bar{\\mathcal{V}})" + }, + { + "bbox": [ + 55, + 287, + 295, + 323 + ], + "type": "text", + "content": " from Ego-Exo4D [29], as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 325, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 325, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 295, + 469 + ], + "type": "text", + "content": "Expert commentary classification and body localization. The commentary in [29] is obtained from voice recordings of the experts converted to text with ASR. Most of the samples contain extra comments like \"oh, I will give this a five out of ten\" and \"that's how I would do it too\". Additionally, as discussed above, they lack positive and negative labels and do not explicitly indicate the body region involved (e.g., legs, arms). Thus, we preprocess the expert commentary for three things—making the commentary concise to extract the improvable feedback, marking which body region the feedback is about, and marking whether the feedback states the need for improvement or not." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": "Since these are all addressable with text reasoning, we use a large language model (LLM) to provide the desired answers. We use Llama3-70B [2], a recent open-sourced model that performs well in current benchmarks. In essence, given a commentary " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": ", the language model " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": " responds to the prompt: \"Given an expert's commentary, summarize the feedback into a single sentence and also provide which body regions need improvement or are correctly executed...\" (see Supp.). Formally, this yields " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(T) = (T', (b^1, c^j), \\ldots, (b^s, c^s))" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": " is the concise summary and " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "b^i" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": " is a body region like head or arms and " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "c^i \\in \\{0, 1, 2\\}" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": " are labels representing needs improvement, correct execution, and no mention, respectively. We group the skeleton joints into six pre-defined body regions " + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "inline_equation", + "content": "b^i" + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": " (details in Supp.). See red and green boxes in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Pairing incorrect and correct executions. Next, we use the above information to mine pairs of incorrect and correct execution in the dataset. Ego-Exo4D also contains metadata about the skill-level " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": " of the demonstrator, broken into four categories (1-4 in increasing expertise) starting from" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 287, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 453 + ], + "type": "text", + "content": "novices who have not performed the activity before to late experts who have performed the same activity, e.g., basketball for " + }, + { + "bbox": [ + 313, + 287, + 555, + 453 + ], + "type": "inline_equation", + "content": "10+" + }, + { + "bbox": [ + 313, + 287, + 555, + 453 + ], + "type": "text", + "content": " years. We sample incorrect learner demonstrations from beginners and correct demonstrations from experts from the same activity in Ego-Exo4D, e.g., penalty kicks in soccer or reverse layup in basketball. Even though there can be incorrect executions by experts and correct ones by beginners, errors by experts are likely incomparable to beginners' due to the skill gap. We use the mapping of body regions to find (in)correct executions referring to the same body region. This results in a collection of video pairs with negative and positive feedback about the same body region, e.g., legs in Fig. 2. Formally, the collection " + }, + { + "bbox": [ + 313, + 287, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 287, + 555, + 453 + ], + "type": "text", + "content": " is curated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 319, + 463, + 548, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 463, + 548, + 478 + ], + "spans": [ + { + "bbox": [ + 319, + 463, + 548, + 478 + ], + "type": "interline_equation", + "content": "\\left\\{\\left(\\mathcal {V} _ {1}, T, \\mathcal {V} _ {2}\\right) \\mid S _ {1} \\in \\mathcal {S} ^ {n}, S _ {2} \\in \\mathcal {S} ^ {e}, \\exists j s. t. c _ {1} ^ {j} = 0, c _ {2} ^ {j} = 1 \\right\\}", + "image_path": "10f023a9f2cfe4c9a32b98178e4b03781665bac43b6e1e9f38ab7c17d32c566f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "inline_equation", + "content": "S^n = \\{\\text{novice, early expert}\\}" + }, + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "inline_equation", + "content": "S^e = \\{\\text{late expert, intermediate expert}\\}" + }, + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "text", + "content": ". Using this matching, we obtain a collection of video pairs of incorrect and correct execution (Fig. 2, yellow box)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "text", + "content": "Temporal alignment and filtering demonstrations. The commentary annotation process in [29] allows the experts to pause at any instant and provide their feedback. Thus, the clip " + }, + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_1" + }, + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "text", + "content": " could have the start of a basketball jump shot, whereas the clip " + }, + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_2" + }, + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "text", + "content": " could also have content before the shot and the follow-through. Thus, we temporally align the learner and the expert video in the collection " + }, + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 534, + 553, + 629 + ], + "type": "text", + "content": " to ensure we are capturing the same step of a demonstration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "For all pairs of clips in the collection " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ", we compare the two poses " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " to quantify their alignment (similarity) using Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) [27], a translation-invariant and body shape-invariant measure. Note that the video component of the input (Sec. 4) is complementary, as it does capture the person's overall movement in the space." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13585" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 226, + 236 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 226, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 226, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 226, + 236 + ], + "type": "image", + "image_path": "d03e703cd2fa192b4df0b944e99b3fed49db8bd6efe7c67f23b4572a3e62ce5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "type": "text", + "content": "Figure 3. Model overview. We tokenize individual modalities using a modality-specific architecture (top). Once all the modalities are encoded as tokens, we use a large language model to learn expert commentary generation, demonstration retrieval, and pose generation. At inference, the model only takes the learner demonstration video " + }, + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "type": "text", + "content": ". See text for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 68, + 387, + 235 + ], + "blocks": [ + { + "bbox": [ + 230, + 68, + 387, + 235 + ], + "lines": [ + { + "bbox": [ + 230, + 68, + 387, + 235 + ], + "spans": [ + { + "bbox": [ + 230, + 68, + 387, + 235 + ], + "type": "image", + "image_path": "5c2c3a868ade3c1644286079acdb2d28ce8eb640a9af282e0f5943bb611567e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 385, + 68, + 555, + 235 + ], + "blocks": [ + { + "bbox": [ + 385, + 68, + 555, + 235 + ], + "lines": [ + { + "bbox": [ + 385, + 68, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 385, + 68, + 555, + 235 + ], + "type": "image", + "image_path": "712ed670a0f8e50099c78ee6cf9d695f486b2e7c3c81027d8ae106bacb0a3300.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": "Denote the commentary timestamp as " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": " for the two videos. We first choose a fixed window around " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": ", say " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "[t_1^a, t_1^b]" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "t_1^a < t_1 < t_1^b" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": ". Next, we find the corresponding time window " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "[t_2^a, t_2^b]" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "t_2^a < t_2 < t_2^b" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_2" + }, + { + "bbox": [ + 55, + 283, + 296, + 344 + ], + "type": "text", + "content": " such that the PA-MPJPE error is the minimum. Specifically," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 350, + 282, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 350, + 282, + 377 + ], + "spans": [ + { + "bbox": [ + 67, + 350, + 282, + 377 + ], + "type": "interline_equation", + "content": "[ t _ {2} ^ {a}, t _ {2} ^ {b} ] = \\operatorname * {a r g m i n} _ {[ t _ {2} ^ {i}, t _ {2} ^ {j} ] \\in \\mathcal {V} _ {2}} \\mathrm {P A - M P J P E} \\left(\\mathcal {V} _ {1} [ t _ {1} ^ {a}, t _ {1} ^ {b} ], \\mathcal {V} _ {2} [ t _ {2} ^ {i}, t _ {2} ^ {j} ]\\right)", + "image_path": "14ed5421daef61a0566f2ba4c114144b898d6ff05934f603439a766f2ddebbd5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 379, + 283, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 379, + 283, + 395 + ], + "spans": [ + { + "bbox": [ + 156, + 379, + 283, + 395 + ], + "type": "text", + "content": "such that " + }, + { + "bbox": [ + 156, + 379, + 283, + 395 + ], + "type": "inline_equation", + "content": "|t_2^j - t_2^i| = |t_1^b - t_1^a|" + }, + { + "bbox": [ + 156, + 379, + 283, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "text", + "content": "We obtain a subset from " + }, + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "text", + "content": " that contains the aligned video segments. There can be some video pairs where the demonstration can be very different and hence no appropriate match exists. Therefore, we only keep the top-k pairs with maximum alignment for every incorrect execution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 558 + ], + "type": "text", + "content": "Overall, we obtain a novel dataset " + }, + { + "bbox": [ + 55, + 462, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 462, + 296, + 558 + ], + "type": "text", + "content": " that contains pairs of videos where the first video has an incorrect execution (on some body region) and the second video corrects it, thus obtaining the desired tuple " + }, + { + "bbox": [ + 55, + 462, + 296, + 558 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}, T, \\bar{\\mathcal{V}})" + }, + { + "bbox": [ + 55, + 462, + 296, + 558 + ], + "type": "text", + "content": " for training and testing. See Fig. 2 (right) and examples in Supp. To ensure fair evaluation, we separately establish a clean gold standard test set, free of potential noise from LLM inference. The test set is manually verified (see Supp for details)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 564, + 231, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 231, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 231, + 578 + ], + "type": "text", + "content": "3.3. Architecture and training design" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Next, we discuss the architecture that encodes the videos " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", poses " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", and text " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " and enables training the auto-regressive function " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". The overall idea is to encode all the representations into a text embedding space and use the strong capabilities of recent language models to obtain output text and pose tokens. This approach has been recently used in vision-conditioned language models [6, 35, 45, 55] for image and video captioning. This setup also allows for a unified architecture for various input and output combinations, as opposed to different input streams for individual modalities. Fig. 3 shows a schematic diagram of the architecture," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 283, + 449, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 449, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 449, + 295 + ], + "type": "text", + "content": "and each part is explained below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "text", + "content": "Encoding video as tokens. Each video " + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "text", + "content": " is an ego-exo clip pair." + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "text", + "content": " The input thus allows the model to see both close-up hand-object interactions (more visible in the learner's ego view) as well as full-body poses in the scene context (more visible in the observer's exo view). We use a pretrained video model to extract spatio-temporal features from both videos' frames, followed by a mapper to convert the video features to video tokens. Formally, " + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = f_{vm}(f_V(\\mathcal{V}))" + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "inline_equation", + "content": "f_{V}" + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "text", + "content": " is a standard feature extractor (we use InternVideo2 [77]) and " + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "inline_equation", + "content": "f_{vm}" + }, + { + "bbox": [ + 313, + 296, + 555, + 475 + ], + "type": "text", + "content": " is the visual mapper. See Fig. 3 (top left). The mapper is typically a low-parameter model that is trainable, whereas the high-parameter feature extractor is kept frozen. The mapper helps transform the visual mapper to be akin to text tokens—a popular strategy in visual instruction tuning models [35, 45]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": "Encoding pose sequences as tokens. To encode a pose sequence, we use a series of linear layers and MLP-mixer [25] to convert a single pose " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{d \\times 3}" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " to an embedding that can be discretized using a codebook, i.e., " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' = f_P(P)" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": ". The architecture in [25] also contains a decoder that converts the embeddings back to human poses, i.e., " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "P = f_{P}^{-1}(\\mathbf{p}')" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": ", which we use in " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " to generate corrected poses. Similar to the above, we use a pose mapper to convert the tokens to embeddings. Formally, " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = f_{pm}(f_P(P)) = f_{pm}(\\mathbf{p}')" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " is the pose token where " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "f_{pm}" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " is the pose mapper. See Fig. 3 (top center). We concatenate embeddings from every frame to obtain the representation for the whole sequence. Having " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "f_{pm}" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " allows training with fewer parameters and thus, we use " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " for learning " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": ". However, using " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "f_{pm}" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " for generation " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " would require adding an inverting function for " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "f_{pm}" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": ". Thus, we directly add " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": " to LLM tokens for pose generation " + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 476, + 556, + 681 + ], + "type": "text", + "content": ". An innovative as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "text", + "content": "1The exo view is the one annotated in Ego-Exo4D as having the maximum subject visibility." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13586" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "pect of our approach is to encode pose as multimodal tokens and train with LLMs, unlike other work regressing the pose parameters from a special pose embedding [20] or using a dedicated pose transformer [49, 82]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 169 + ], + "type": "text", + "content": "Encoding text as tokens. Text tokenization is the standard process before inputting a text sentence to the LLM [59, 73], i.e., " + }, + { + "bbox": [ + 55, + 121, + 295, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{t} = f_t(T)" + }, + { + "bbox": [ + 55, + 121, + 295, + 169 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 121, + 295, + 169 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 121, + 295, + 169 + ], + "type": "text", + "content": " is the tokenization function. See Fig. 3 (top right)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "text", + "content": "Multi-modal sequence prediction. The previous steps yield multimodal tokens " + }, + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 55, + 170, + 294, + 241 + ], + "type": "text", + "content": " for video, pose, and text, respectively. Next, we use the strong sequence prediction capabilities of large language models to obtain the desired output tokens based on the sequence of multi-modal inputs [6, 35, 45, 55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": "For training " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " for expert commentary generation, we provide a sequence of learner video and pose tokens (v and p) and the corrected pose tokens (" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\nu}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": "). We ask the model to predict the expert commentary token sequence. For the sequence prediction language model " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": ", we wish to obtain " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{t} = \\mathcal{L}_s(\\mathbf{v},\\mathbf{p},\\bar{\\mathbf{v}},\\bar{\\mathbf{p}})" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": ". Consequently, the training objective is the standard cross-entropy loss, " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\min_{\\theta}\\{-\\log (\\mathbf{t}|\\mathbf{v},\\mathbf{p},\\bar{\\mathbf{v}},\\bar{\\mathbf{p}};\\theta)\\}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " are the parameters of the model. See Fig. 3 (bottom left). For consistency with the training of these language models, the sequence is formatted to be conversational in nature, e.g., \"provide an expert's commentary based on this pose sequence:\". Similarly, for expert demonstration retrieval, we wish to obtain a retrieval candidate " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}} = \\mathcal{L}_s(\\mathbf{v},\\mathbf{p},\\mathbf{t})" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": ". Likewise, the training objective with parameters " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\min_{\\gamma}\\{-\\log (\\bar{\\mathbf{p}} |\\mathbf{v},\\mathbf{p},\\mathbf{t};\\gamma)\\}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": ", which we also use as the relevance score for retrieval during inference. Expert pose generation is trained with the same objective, except we use " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " (and obtain " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}}'" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": "). See Fig. 3 (bottom). Generating pose further requires converting back the pose tokens " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}}'" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": " to 3D joints using the pose decoder " + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "inline_equation", + "content": "f_P^{-1}" + }, + { + "bbox": [ + 56, + 242, + 296, + 494 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "text", + "content": "At inference time, as introduced in Sec. 3.1, we only use " + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "text", + "content": " as the input: " + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "text", + "content": " drops the expert demonstration and predicts the expert commentary " + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "text", + "content": ". We use this predicted commentary as input for " + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r" + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 55, + 494, + 295, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 186, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 186, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 186, + 564 + ], + "type": "text", + "content": "3.4. Implementation details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Dataset and statistics. Ego-Exo4D [29] contains 5,035 videos of participants doing activities across eight scenarios. We focus on three physical scenarios—basketball, soccer, and rock climbing—though our model is in principle generalizable to other physical skills, without any task-specific design, unlike [63, 78]. We use physical scenarios since the coaching feedback is groundable in the body regions, as opposed to procedural tasks like cooking, where the suggestions can be alternate ingredients or steps that are not visually present. Details and statistics are in Supp. Following the dataset creation process outlined in Sec. 3.2, we obtain a dataset of 25,505 training and 1,272 testing tuples" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}, T, \\bar{\\mathcal{V}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ". We choose " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " (train) and " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " (test) for min-k choice of correct demonstrations per incorrect execution. To reiterate, we manually examine the test set for correct commentary summary and body region labeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": "Network architecture. The video model " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "f_{V}" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " is an InternVideo2 [77] encoder that provides strong visual representations. The pose encoder " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "f_{P}" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "f_{P}^{-1}" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " are adapted from PCT [25], which learns compositional tokens from human poses. Since the original training in [25] uses 2D poses, we adapt it to use 3-dimensions and retrain on the human poses in [29]. Both the visual and pose mappers are low-parameter MLP layers [44]. Finally, we use Llama 3-8B [2] as the LLM multimodal encoder for sequence prediction. We finetune " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " for 5 epochs with a learning rate of " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": ". Next, we modify the token dimension in " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " to accommodate pose tokens " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": ", and hence, we fine-tune " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " when learning " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " with a learning rate of " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " for 5 epochs. Video " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "(f_{V})" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " and pose models " + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "inline_equation", + "content": "(f_{P}, f_{P}^{-1})" + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "text", + "content": " are kept frozen. All the models are trained on 8 V100 32GB GPUs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 312, + 454, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 454, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 454, + 326 + ], + "type": "text", + "content": "4. Experiments and results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "text", + "content": "We first discuss the baselines and ablations, followed by the evaluation setup and results for the three outputs—expert commentary generation " + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_t)" + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "text", + "content": ", expert demonstration retrieval " + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_r)" + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "text", + "content": ", and pose demonstration generation " + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_g)" + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "text", + "content": ". We also show qualitative examples and discuss the limitations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 393, + 529, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 529, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 529, + 405 + ], + "type": "text", + "content": "Baselines. We compare with the following baselines:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 406, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 311, + 406, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 406, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 311, + 406, + 553, + 490 + ], + "type": "text", + "content": "- InternVideo2-NN, InternVideo2-FT [77]: Given a query video " + }, + { + "bbox": [ + 311, + 406, + 553, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 311, + 406, + 553, + 490 + ], + "type": "text", + "content": ", the nearest neighbor baseline (NN) finds the most similar video by InternVideo2 [77] feature similarity in the training data and returns the corresponding commentary or demonstration, for all tasks. The FT version finetunes the model to contrastively match the learner demonstration with the expert commentary/demonstration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 492, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 492, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 311, + 492, + 553, + 601 + ], + "type": "text", + "content": "- VideoChat2 [36, 37], LLaVA [45]: In these methods, we prompt SOTA video and image captioning models to generate commentary for an input demonstration. We use the log-likelihood loss to find the retrieved expert demonstration. These baselines evaluate if existing video captioning methods can provide expert feedback or retrieve expert demonstrations. We use the ego and \"best exo\" frames, same as for our method. Note that neither of these baselines are applicable for pose generation " + }, + { + "bbox": [ + 311, + 492, + 553, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 311, + 492, + 553, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 602, + 553, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 602, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 311, + 602, + 553, + 674 + ], + "type": "text", + "content": "- LLaVA-FT [45], LLaVA-FT w/ pose [45]: These baselines are based on the SOTA visual-language method LLaVA [45] but trained on our dataset with the same text model, i.e., Llama 3-8B [2], for an apples-to-apples comparison. The \"w/ pose\" variant also takes the 3D pose coordinates in text form as input." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 677, + 553, + 713 + ], + "type": "text", + "content": "- PoseScript [16], PoseFix [17]: These two works enable pose-to-text and text-to-pose reasoning. The text generated or used for pose generation contains detail about the" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13587" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 74, + 372, + 244 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 372, + 244 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 372, + 244 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 372, + 244 + ], + "type": "table", + "html": "
MethodCommentary Gen.Demo Ret.Pose Gen.
B@4MR-LRmedR↓P↓
InternVideo2-NN [77]42.146.949.313.5198161
InternVideo2-FT [77]42.947.650.014.1190157
VideoChat2 [37]27.844.341.914.9183—
LLaVA [45]28.544.144.215.0183—
LLaVA-FT [45]43.548.551.517.8177—
LLaVA-FT w/ pose [45]43.648.551.718.0172150
PoseScript/Fix [16, 17]24.144.546.315.9182182
ExpertAF44.949.654.619.1158135
ExpertAF w/o video44.649.454.218.7161139
ExpertAF w/o pose44.249.354.218.6163—
w/o alignment42.048.651.516.9180153
w/ global pose43.047.952.616.5184150
ExpertAF w/ full-sup45.850.955.722.5146131
", + "image_path": "63e969fd713711291835bef2c4260157be928605317fc24c03dfdc29c2a9508a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 378, + 71, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 378, + 71, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 378, + 71, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 378, + 71, + 553, + 247 + ], + "type": "image", + "image_path": "157726add4d6f8b06b60e90cf3b0ed9405c16461026d5e6be8d4d15c5aa473c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": "Table 1. Results on automatic metrics (left) and human evaluation (right). We break down results for the three outputs—expert commentary generation, expert demonstration retrieval, and expert pose generation. Our method outperforms all baselines and prior work on all tasks. The last row \"w/ full-sup\" uses privileged input (the demo video " + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\bar{\\nu}" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": ") at inference. (B@4: BLEU-4, M: Meteor, R-L: ROUGE-L F1, R: recall@50, medR: median rank, P: PA-MPJPE). For all metrics higher is better, except medR and PA-MPJPE (" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": "). Our method is also rated higher by human raters on a Likert scale (min:1, max:4), compared to all the other methods (right). See text for details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 60, + 312, + 294, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 312, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 60, + 312, + 294, + 396 + ], + "type": "text", + "content": "location of body parts, e.g., the hands are raised, as opposed to expert commentary. Hence, they let us evaluate if pose description is adequate for expert feedback. For pose generation, we evaluate if providing an expert commentary helps generate the desired expert demonstration. Both methods use SMPL [47] pose and hence we convert 3D pose to SMPL and vice versa [39], as needed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": "Ablations. In addition to the strong baselines, we also compare the performance against ablations and variants of our design choices. ExpertAF w/o pose and ExpertAF w/o video evaluate the performance when only one modality is used. ExpertAF w/o alignment quantifies the need for temporal alignment (Sec. 3.2), while ExpertAF w/ global pose evaluates if just providing the model with one correct pose per activity chosen based on expert commentary is enough (vs. finding the closest correct pose sequence). ExpertAF w/ full-sup is a stronger variant with privileged input for inference—" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": " uses " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": " uses both " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": ". See Supp. for additional ablations of the choice of LLM " + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 55, + 396, + 295, + 564 + ], + "type": "text", + "content": ", contribution of ego and exo videos, and joint training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 567, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 615 + ], + "type": "text", + "content": "Expert commentary generation. To evaluate the text commentary, we use standard metrics BLEU-4 [60], METEOR [7] and ROUGE-L F1 [65], following prior work in evaluating text generation [14, 58] (higher is better)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Tab. 1 (left) shows the results. Our method outperforms all methods on all captioning metrics, with gains up to " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " over the best baseline trained on the same dataset. All the gains are statistically significant using paired t-test with " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "p < 0.05" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Fig. 4 (top) shows commentary generation by various methods. Firstly, SOTA methods LLaVA [45] and VideoChat2 [37], being captioning methods, yield a verbose description of the activity, rather than a critique." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 421 + ], + "type": "text", + "content": "Fine-tuning LLaVA with our method improves the performance, but still it misses our use of expert demonstrations and pose sequences. Even using pose information as text only marginally improves \"LLaVA-FT w/ pose\", showing the effectiveness of our explicit pose encoding. Moreover, PoseScript [16] generates pose descriptions like \"head is turned to the left\" which is inadequate for feedback. Our results over the baselines show the advantage of temporal alignment and a diverse expert demonstration set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 422, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 506 + ], + "type": "text", + "content": "Our stronger variant \"w/ full-sup\" achieves an even better result by taking " + }, + { + "bbox": [ + 313, + 422, + 556, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 422, + 556, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 556, + 506 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}" + }, + { + "bbox": [ + 313, + 422, + 556, + 506 + ], + "type": "text", + "content": " as inputs during inference. It is useful in cases where a learner has access to the expert demonstration but wants actionable feedback in languageform. Finally, Fig. 4 (bottom row) shows a failure case of commentary generation, showcasing the difficulty in pinpointing the exact mistake." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 507, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 556, + 579 + ], + "type": "text", + "content": "Expert demonstration retrieval. For retrieval, we use standard retrieval metrics: recall@k and median rank. We set k=50 to account for the fact that multiple expert demonstrations could suitably correct the given learner demonstration. The retrieval set contains 1,272 samples, the same as the test set size. Higher recall and lower ranks are better." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 580, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 676 + ], + "type": "text", + "content": "Tab. 1 (middle) shows the results. Our method improves the median rank by 14 positions compared to the best baseline and a significantly better recall@50. Our method and the baselines show a similar trend as with the expert commentary generation above, the stronger baseline w/ \"full-sup\" (" + }, + { + "bbox": [ + 313, + 580, + 556, + 676 + ], + "type": "inline_equation", + "content": ">4\\%" + }, + { + "bbox": [ + 313, + 580, + 556, + 676 + ], + "type": "text", + "content": " better recall and 26 rank improvement) being crucial in cases where a learner has the commentary but wants a video exemplar to learn from." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "content": "Expert pose generation. Tab. 1 (right column) shows the expert pose generation results, with error measured by PA-MPJPE [27] (mm). Our novel pose generation method" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13588" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 132, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 132, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 132, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 132, + 148 + ], + "type": "image", + "image_path": "69cdc67194d5ff823d450d701bf899b2ddf19993421584b8a7a9b21f62352b24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 72, + 199, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 72, + 199, + 82 + ], + "spans": [ + { + "bbox": [ + 150, + 72, + 199, + 82 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 92, + 211, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 92, + 211, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 92, + 211, + 138 + ], + "type": "text", + "content": "The climber is exhausted and struggling to make a move, with poor hand and arm positioning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 247, + 72, + 268, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 72, + 268, + 82 + ], + "spans": [ + { + "bbox": [ + 247, + 72, + 268, + 82 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 92, + 293, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 92, + 293, + 137 + ], + "spans": [ + { + "bbox": [ + 220, + 92, + 293, + 137 + ], + "type": "text", + "content": "The climber is struggling to maintain the position and is unable to make the next move." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 320, + 72, + 358, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 72, + 358, + 81 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 358, + 81 + ], + "type": "text", + "content": "LLaVA-FT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 301, + 92, + 377, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 92, + 377, + 137 + ], + "spans": [ + { + "bbox": [ + 301, + 92, + 377, + 137 + ], + "type": "text", + "content": "The climber is wasting energy by pulling away, which is slowing down reaching the next handhold" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 399, + 72, + 444, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 72, + 444, + 82 + ], + "spans": [ + { + "bbox": [ + 399, + 72, + 444, + 82 + ], + "type": "text", + "content": "VideoChat2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 385, + 92, + 457, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 92, + 457, + 137 + ], + "spans": [ + { + "bbox": [ + 385, + 92, + 457, + 137 + ], + "type": "text", + "content": "The person can improve by practicing their technique, warming up and stretching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 482, + 72, + 525, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 482, + 72, + 525, + 82 + ], + "spans": [ + { + "bbox": [ + 482, + 72, + 525, + 82 + ], + "type": "text", + "content": "PoseScript" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 466, + 85, + 541, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 85, + 541, + 144 + ], + "spans": [ + { + "bbox": [ + 466, + 85, + 541, + 144 + ], + "type": "text", + "content": "The person is standing on their left leg with the right leg lifted up and bent at the knee. The torso is leaning to the left. The head is turned to the left." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 61, + 170, + 149, + 231 + ], + "blocks": [ + { + "bbox": [ + 62, + 161, + 148, + 170 + ], + "lines": [ + { + "bbox": [ + 62, + 161, + 148, + 170 + ], + "spans": [ + { + "bbox": [ + 62, + 161, + 148, + 170 + ], + "type": "text", + "content": "Learner demonstration" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 170, + 149, + 231 + ], + "lines": [ + { + "bbox": [ + 61, + 170, + 149, + 231 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 149, + 231 + ], + "type": "image", + "image_path": "e45b0af9cb986232e56d87bf322ba561c167774408777408ef1a405e3ea98f90.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 60, + 232, + 149, + 291 + ], + "blocks": [ + { + "bbox": [ + 60, + 232, + 149, + 291 + ], + "lines": [ + { + "bbox": [ + 60, + 232, + 149, + 291 + ], + "spans": [ + { + "bbox": [ + 60, + 232, + 149, + 291 + ], + "type": "image", + "image_path": "91bbee65fbc312ff8144dc85b7f6bbca3b34431f76fd0d93c3c533035c6fb734.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 58, + 311, + 149, + 358 + ], + "blocks": [ + { + "bbox": [ + 80, + 301, + 126, + 310 + ], + "lines": [ + { + "bbox": [ + 80, + 301, + 126, + 310 + ], + "spans": [ + { + "bbox": [ + 80, + 301, + 126, + 310 + ], + "type": "text", + "content": "Failure case" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 311, + 149, + 358 + ], + "lines": [ + { + "bbox": [ + 58, + 311, + 149, + 358 + ], + "spans": [ + { + "bbox": [ + 58, + 311, + 149, + 358 + ], + "type": "image", + "image_path": "47b2c8773f2af44cb3acb144c8785a4f23033f6d73e01871f269410cf0265e59.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 425 + ], + "type": "text", + "content": "Figure 4. Qualitative results. (Top) Comparison of expert commentary generated by various baselines. (Second and third row) Examples of expert commentary generation, demonstration retrieval, and pose generation by our method. Notice the expert demonstration and pose generation corrects the mistake pointed out in the commentary, i.e., one hand is used to throw and the other to guide, and body position is improved to control better control and power. Colored text and marks (red for mistake, green for correction) are shown only for visualization. (Bottom) Failure cases. See Supp. for video results." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 160, + 161, + 250, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 161, + 250, + 171 + ], + "spans": [ + { + "bbox": [ + 160, + 161, + 250, + 171 + ], + "type": "text", + "content": "Expert comment. - Ours" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 157, + 183, + 252, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 183, + 252, + 219 + ], + "spans": [ + { + "bbox": [ + 157, + 183, + 252, + 219 + ], + "type": "text", + "content": "The shooter's stance is correct, but they need to throw the ball with one hand and guide it with the other." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 156, + 243, + 254, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 243, + 254, + 280 + ], + "spans": [ + { + "bbox": [ + 156, + 243, + 254, + 280 + ], + "type": "text", + "content": "The player's stance is too straight, which prevents them from generating the required power and control on the ball." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 157, + 311, + 253, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 311, + 253, + 347 + ], + "spans": [ + { + "bbox": [ + 157, + 311, + 253, + 347 + ], + "type": "text", + "content": "The shooter's left elbow is bent and his wrist is not fully extended, affecting the ball's rotation and accuracy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 258, + 161, + 354, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 161, + 354, + 171 + ], + "spans": [ + { + "bbox": [ + 258, + 161, + 354, + 171 + ], + "type": "text", + "content": "Ground truth commentary" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 257, + 179, + 355, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 179, + 355, + 224 + ], + "spans": [ + { + "bbox": [ + 257, + 179, + 355, + 224 + ], + "type": "text", + "content": "The shooter is using two hands to shoot the ball, instead of using one shooting hand and arm with the other at a 90-degree angle." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 257, + 239, + 357, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 239, + 357, + 283 + ], + "spans": [ + { + "bbox": [ + 257, + 239, + 357, + 283 + ], + "type": "text", + "content": "The player's foot position and body lean are affecting the quality of their ball contact, leading to a loss of power and control." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 258, + 311, + 356, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 311, + 356, + 347 + ], + "spans": [ + { + "bbox": [ + 258, + 311, + 356, + 347 + ], + "type": "text", + "content": "He does a good job using his left hand on the left side, but needs to improve his posture and weight distribution." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 361, + 170, + 451, + 230 + ], + "blocks": [ + { + "bbox": [ + 364, + 161, + 449, + 171 + ], + "lines": [ + { + "bbox": [ + 364, + 161, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 364, + 161, + 449, + 171 + ], + "type": "text", + "content": "Expert demo. retrieval" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 170, + 451, + 230 + ], + "lines": [ + { + "bbox": [ + 361, + 170, + 451, + 230 + ], + "spans": [ + { + "bbox": [ + 361, + 170, + 451, + 230 + ], + "type": "image", + "image_path": "7d3a9e59a065beb6e33d6a2f25173115e85cd9a1cadf8fa06833e863f479c951.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 361, + 231, + 451, + 291 + ], + "blocks": [ + { + "bbox": [ + 361, + 231, + 451, + 291 + ], + "lines": [ + { + "bbox": [ + 361, + 231, + 451, + 291 + ], + "spans": [ + { + "bbox": [ + 361, + 231, + 451, + 291 + ], + "type": "image", + "image_path": "1565ada75121f6b2de9c9aa1170d1b34c52cad644ca1380f5f6f78771bdcffc1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 360, + 298, + 451, + 358 + ], + "blocks": [ + { + "bbox": [ + 360, + 298, + 451, + 358 + ], + "lines": [ + { + "bbox": [ + 360, + 298, + 451, + 358 + ], + "spans": [ + { + "bbox": [ + 360, + 298, + 451, + 358 + ], + "type": "image", + "image_path": "2129b9d6136ad20fd3630a97e35df46c6750cb1ae7d1aa44ecaf0f3d371a4a77.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 456, + 171, + 547, + 230 + ], + "blocks": [ + { + "bbox": [ + 470, + 161, + 531, + 171 + ], + "lines": [ + { + "bbox": [ + 470, + 161, + 531, + 171 + ], + "spans": [ + { + "bbox": [ + 470, + 161, + 531, + 171 + ], + "type": "text", + "content": "Pose generation" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 456, + 171, + 547, + 230 + ], + "lines": [ + { + "bbox": [ + 456, + 171, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 456, + 171, + 547, + 230 + ], + "type": "image", + "image_path": "01db810c9d6d4ad3df6c95bb51a86d8a84de8f7480d187b60710c63e9f860d0e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 456, + 231, + 547, + 291 + ], + "blocks": [ + { + "bbox": [ + 456, + 231, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 456, + 231, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 456, + 231, + 547, + 291 + ], + "type": "image", + "image_path": "b4720fc1e800611ef2157eec73a90bbfda8e653bbe71df5edced359ff33f9bd8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 455, + 298, + 547, + 358 + ], + "blocks": [ + { + "bbox": [ + 455, + 298, + 547, + 358 + ], + "lines": [ + { + "bbox": [ + 455, + 298, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 455, + 298, + 547, + 358 + ], + "type": "image", + "image_path": "e50ab803604f6adbb398f2cbdfe71aefb09dfff0d31a964c3fa47e28de3ff05c.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 431, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 563 + ], + "type": "text", + "content": "performs better than learning 3D position using text in \"LLaVA-FT w/ pose\". Furthermore, PoseFix [17] is designed to generate SMPL parameters based on a coarse modification text description. Our use of generation to output poses (vs. retrieval) is advantageous when the candidate retrieval set does not contain the correct demonstration; see Supp. See Fig. 4 and Supp. for qualitative examples including failure cases. The 3D pose sequences in the dataset are often auto-generated and have some noisy samples, discussed in Supp. Across the three tasks, our method is superior to its ablations." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 566, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 662 + ], + "type": "text", + "content": "Human evaluation. For all the tasks, we also solicit human evaluation (Tab. 1 right) to rate the quality of the generated text descriptions. Five raters (per scenario) uninvolved with this project rate the quality of each generated output by scoring on a Likert scale from 1 to 4 (higher is better); details in Supp. We ensure that the raters have a basic knowledge of the scenario they are rating (basketball, soccer, rock climbing)." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Across all three tasks, human raters score our method the highest. A significant gap of up to 0.7 over the possible range " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(1 - 4)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", as well as gains up to " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", showcase ExpertAF's expert feedback quality. That said, there is nat" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 431, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 564 + ], + "type": "text", + "content": "urally room for improvement on this new task. We find that while humans say our model excels on cases where the feedback is visually groundable, e.g. incorrect hand angle, it tends to fall short when the feedback is not directly visible, e.g. the climber looks fatigued. The human evaluation complements the automatic metrics above, overcoming the limitation that a \"ground truth\" commentary or a visual demonstration may capture only one of multiple possible errors in the learner's demonstration [23, 68] (e.g., for a basketball shot, both the hand placement and the jump could be incorrect, but only one may be mentioned in the ground truth)." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 314, + 582, + 388, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 582, + 388, + 595 + ], + "spans": [ + { + "bbox": [ + 314, + 582, + 388, + 595 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "We proposed a novel task and method to generate expert commentary and demonstrations from a learner's video. We develop a weakly-supervised training approach and benchmark for this problem. Our novel method fusing multimodal inputs from learner and expert demonstrations together with expert commentary results in state-of-the-art performance in actionable feedback, and lays the groundwork for accessible, affordable, and actionable AI coaching applications in the future." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13589" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "type": "text", + "content": "Acknowledgement. UT Austin is supported in part by the IFML NSF AI Institute. Thanks to Fu-Jen Chu, Jing Huang and Xitong Yang for help with the Exo-Ego4D [29] pose extraction pipeline." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 158, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 158, + 294, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 294, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 294, + 212 + ], + "type": "text", + "content": "[1] Yazan Abu Farha, Alexander Richard, and Juergen Gall. When will you do what?-anticipating temporal occurrences of activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5343-5352, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 241, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 241, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 241, + 224 + ], + "type": "text", + "content": "[2] AI@Meta. Llama 3 model card, 2024. 4, 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 226, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 226, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 226, + 295, + 280 + ], + "type": "text", + "content": "[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23066-23078, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "type": "text", + "content": "[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. What you say is what you show: Visual narration detection in instructional videos. arXiv preprint arXiv:2301.02307, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 379 + ], + "type": "text", + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "text", + "content": "[6] Kumar Ashutosh, Zihui Xue, Tushar Nagarajan, and Kristen Grauman. Detours for navigating instructional videos. In CVPR, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "text", + "content": "[7] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 295, + 524 + ], + "type": "text", + "content": "[8] Tobias Baumgartner and Stefanie Klatt. Monocular 3d human pose estimation for sports broadcasts using partial sports field registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5108-5117, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "type": "text", + "content": "[9] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the IEEE international conference on computer vision, pages 2177-2185, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "text", + "content": "[10] Gedas Bertasius, Aaron Chan, and Jianbo Shi. Egocentric basketball motion planning from a single first-person image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "text", + "content": "[11] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "text", + "content": "[12] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "Ramesh. Video generation models as world simulators, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[13] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "text", + "content": "[14] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. arXiv preprint arXiv:2305.18500, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "text", + "content": "[15] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos. International Journal of Computer Vision, 129:2846-2864, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "text", + "content": "[16] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "text", + "content": "[17] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 361, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 426 + ], + "type": "text", + "content": "[18] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 319-335. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "type": "text", + "content": "[19] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Posegpt: Chatting about 3d human pose. arXiv preprint arXiv:2311.18836, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[20] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "text", + "content": "[21] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence, 43(11):4021-4036, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 538, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 538, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 538, + 553, + 570 + ], + "type": "text", + "content": "[22] Jiyang Gao, Zhenheng Yang, and Ram Nevatia. Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 571, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 713 + ], + "type": "text", + "content": "[23] Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal, Pawan Sasanka Ammanamanchi, Anuoluwapo Aremu, Antoine Bosselut, Khyathi Raghavi Chandu, Miruna-Adriana Clinciu, Dipanjan Das, Kaustubh Dhole, Wanyu Du, Esin Durmus, Ondrej Dušek, Chris Chinenye Emezue, Varun Gangal, Cristina Garbacea, Tatsunori Hashimoto, Yufang Hou, Yacine Jernite, Harsh Jhamtani, Yangfeng Ji, Shailza Jolly, Mhir Kale, Dhruv Kumar, Faisal Ladhak, Aman Madaan, Mounica Maddela, Khyati Mahajan, Saad Mahamood, Bodhisattwa Prasad Majumder, Pedro Henrique Martins, Angelina McMillan-Major, Simon Mille, Emiel van Miltenburg, Moin Nadeem, Shashi Narayan, Vitaly Nikolaev, Andre Niyongabo Rubungo, Salomey Osei, Ankur" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13590" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 183 + ], + "type": "text", + "content": "Parikh, Laura Perez-Beltrachini, Niranjan Ramesh Rao, Vikas Raunak, Juan Diego Rodriguez, Sashank Santhanam, João Sedoc, Thibault Sellam, Samira Shaikh, Anastasia Shimorina, Marco Antonio Sobrevilla Cabezudo, Hendrik Strobelt, Nishant Subramani, Wei Xu, Diyi Yang, Akhila Yerukola, and Jiawei Zhou. The GEM benchmark: Natural language generation, its evaluation and metrics. In Proceedings of the 1st Workshop on Natural Language Generation, Evaluation, and Metrics (GEM 2021), pages 96-120, Online, 2021. Association for Computational Linguistics. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 184, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 228 + ], + "type": "text", + "content": "[24] Adelaide Genay, Anatole Lécuyer, and Martin Hachet. Being an avatar \"for real\": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 28(12):5071-5090, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 229, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 295, + 282 + ], + "type": "text", + "content": "[25] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 660-671, 2023. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 285, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 327 + ], + "type": "text", + "content": "[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 331, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 385 + ], + "type": "text", + "content": "[27] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 387, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 409 + ], + "type": "text", + "content": "[28] Andrew S Gordon. Automated video assessment of human performance. In Proceedings of AI-ED, page 10, 1995. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "type": "text", + "content": "[29] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In CVPR, 2024. 1, 2, 3, 4, 6, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "text", + "content": "[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "type": "text", + "content": "[31] William Hoff and Tyrone Vincent. Analysis of head pose accuracy in augmented reality. IEEE Transactions on Visualization and Computer Graphics, 6(4):319-334, 2000. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "text", + "content": "[32] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 667 + ], + "type": "text", + "content": "[33] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[34] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7718-7727, 2019. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[35] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[36] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[37] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark, 2023. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[38] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "type": "text", + "content": "[39] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yukun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv preprint arXiv:2304.00553, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "text", + "content": "[40] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 442 + ], + "type": "text", + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 445, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 507 + ], + "type": "text", + "content": "[42] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 511, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 564 + ], + "type": "text", + "content": "[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9522-9531, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 567, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 588 + ], + "type": "text", + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "text", + "content": "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[46] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph., 34(6), 2015. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13591" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 298, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "text", + "content": "[48] Sabine Losch, Eva Traut-Mattausch, Maximilian D Muhlberger, and Eva Jonas. Comparing the effectiveness of individual coaching, self-coaching, and group training: How leadership makes the difference. Frontiers in psychology, 7: 175595, 2016. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 297, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 297, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 297, + 174 + ], + "type": "text", + "content": "[49] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "type": "text", + "content": "[50] Priyanka Mandikal and Kristen Grauman. Dexvip: Learning dexterous grasping with human hand pose priors from video. In Conference on Robot Learning, pages 651-661. PMLR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "text", + "content": "[51] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action forecasting@ ego4d challenge 2022. arXiv preprint arXiv:2207.12080, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "text", + "content": "[52] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 365 + ], + "type": "text", + "content": "[53] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 296, + 431 + ], + "type": "text", + "content": "[54] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 454 + ], + "type": "text", + "content": "[55] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. In CVPR, 2024. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 500 + ], + "type": "text", + "content": "[56] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 502, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 296, + 545 + ], + "type": "text", + "content": "[57] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. arXiv preprint arXiv:2207.09666, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 547, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 296, + 591 + ], + "type": "text", + "content": "[58] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning transformer using dual visual features. In European Conference on Computer Vision, pages 167–184. Springer, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 593, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 296, + 713 + ], + "type": "text", + "content": "[59] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fo" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 73, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 555, + 711 + ], + "type": "text", + "content": "tis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mély, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O’Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael. Pokorny. Michelle Pokrass. Vitchyr H. Pong Tolly Powell Alethea Power Boris Power Elizabeth Proehl Raul Puri Alec Radford Jack Rae Aditya Ramesh Cameron Raymond Francis Real Kendra Rimbach Carl Ross Bob Rotsted Henri Roussez Nick Ryder Mario Saltarelli Ted Sanders Shibani Santurkar Girish Sastry. Heather Schmidt David Schnurr John Schulman Daniel Selsam. Kyla Sheppard Toki Sherbakov Jessica Shieh Sarah Shoker Pranav Shyam Szymon Sidor Eric Sigler Maddie Simens Jordan Sitkin Katarina Slama Ian Sohl Benjamin Sokolowsky Yang Song Natalie Staudacher Felipe Petroski Such Natalie Summers Ilya Sutskever Jie Tang. Nikolas Tezak Madeleine B. Thompson Phil Tillet. Amin Tootoonchian Elizabeth Tseng Preston Tuggle Nick Turley Jerry Tworek Juan Felipe Cerón Uribe Andrea Vallone Arun Vijayvergiya Chelsea Voss Carroll Wainwright Justin Jay Wang Alvin Wang Ben Wang Jonathan Ward Jason Wei CJ Weinmann Akila Welihinda Peter Welin" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13592" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 149 + ], + "type": "text", + "content": "der, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "type": "text", + "content": "[60] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "text", + "content": "[61] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "text", + "content": "[62] Paritosh Parmar, Jaiden Reddy, and Brendan Morris. Piano skills assessment. In 2021 IEEE 23rd international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 353 + ], + "type": "text", + "content": "[63] Paritosh Parmar, Amol Gharat, and Helge Rhodin. Domain knowledge-informed self-supervised representations for workout form assessment. In European Conference on Computer Vision, pages 105-123. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[64] Hamed Pirsiavash, Carl Vondrick, and Antonio Torralba. Assessing the quality of actions. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 556-571. Springer, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[65] Lin CY Rouge. A package for automatic evaluation of summaries. In Proceedings of Workshop on Text Summarization of ACL, Spain, 2004. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "text", + "content": "[66] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 554 + ], + "type": "text", + "content": "[67] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761-14771, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "type": "text", + "content": "[68] Chenhui Shen, Liying Cheng, Xuan-Phi Nguyen, Yang You, and Lidong Bing. Large language models are not yet human-level evaluators for abstractive summarization. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 4215–4233, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "type": "text", + "content": "[69] Sumedh Sontakke, Jesse Zhang, Séb Arnold, Karl Pertsch, Erdem Biryik, Dorsa Sadigh, Chelsea Finn, and Laurent Itti. Roboclip: One demonstration is enough to learn robot policies. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[70] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[71] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9839-9848, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[72] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4790-4799, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 435 + ], + "type": "text", + "content": "[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 437, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 458 + ], + "type": "text", + "content": "[74] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[75] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 504, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 580 + ], + "type": "text", + "content": "[76] Tianyu Wang, Yijie Wang, and Mian Li. Towards accurate and interpretable surgical skill assessment: A video-based method incorporating recognized surgical gestures and skill levels. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 668-678. Springer, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "text", + "content": "[77] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[78] Chengming Xu, Yanwei Fu, Bing Zhang, Zitian Chen, YuGang Jiang, and Xiangyang Xue. Learning to score figure skating sport videos. IEEE Transactions on Circuits and Systems for Video Technology, 30(12):4578-4590, 2020. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[79] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13593" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 509 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[80] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "text", + "content": "[82] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "type": "text", + "content": "[83] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129:703-718, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 342 + ], + "type": "text", + "content": "[84] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "type": "text", + "content": "[85] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 295, + 453 + ], + "type": "text", + "content": "[86] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 295, + 509 + ], + "type": "text", + "content": "[87] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13594" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_content_list.json b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..90077aff7eced07af2a9db9c1681b4490086533b --- /dev/null +++ b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_content_list.json @@ -0,0 +1,2042 @@ +[ + { + "type": "text", + "text": "Explainable Saliency: Articulating Reasoning with Contextual Prioritization", + "text_level": 1, + "bbox": [ + 99, + 130, + 869, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nuo Chen", + "bbox": [ + 323, + 181, + 406, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Jiang", + "bbox": [ + 446, + 181, + 537, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Zhao", + "bbox": [ + 578, + 181, + 647, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science and Engineering University of Minnesota", + "bbox": [ + 285, + 199, + 683, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chen9029, mjiang, qzhao}@umn.edu", + "bbox": [ + 336, + 237, + 629, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep saliency models, which predict what parts of an image capture our attention, are often like black boxes. This limits their use, especially in areas where understanding why a model makes a decision is crucial. Our research tackles this challenge by developing an explainable saliency (XSal) model that not only identifies what is important in an image, but also explains its choices in a way that makes sense to humans. We achieve this by using vision-language models to reason about images and by focusing the model's attention on the most crucial information using a contextual prioritization mechanism. Unlike prior approaches that rely on fixation descriptions or soft-attention based semantic aggregation, our method directly models the reasoning steps involved in saliency prediction, generating selectively prioritized explanations clarify why specific regions are prioritized. Comprehensive evaluations demonstrate the effectiveness of our model in generating high-quality saliency maps and coherent, contextually relevant explanations. This research is a step towards more transparent and trustworthy AI systems that can help us understand and navigate the world around us.", + "bbox": [ + 76, + 316, + 473, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans have this amazing ability to look at a scene and instantly know what is important. Saliency models, the cornerstone of computer vision, aim to mimic human gaze, identifying the most captivating regions within an image. These models have found wide application in image and video quality assessment [4, 54], virtual reality [42, 45], autonomous systems [8, 20], image captioning [13], and clinical diagnostics [11, 21]. However, while current saliency models [26, 36, 52] excel at replicating generic gaze patterns, they often operate as enigmatic black boxes. This lack of transparency hinders their application in scenarios where understanding why a region is deemed salient is as crucial as the prediction itself.", + "bbox": [ + 75, + 688, + 468, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The development of explainable models that mimic hu", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/041e108dddd326591628d520158dfe384eabc9456149c6beeb95834f716b7ffd.jpg", + "image_caption": [ + "Figure 1. Our proposed XSal model provides step-by-step natural language explanations, semantic prototype visualization, as well as saliency weights to explain its prediction." + ], + "image_footnote": [], + "bbox": [ + 509, + 284, + 870, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "man attention remains an underexplored research frontier. While some models have attempted to introduce explainability, they often rely on massive predefined vocabularies of semantics [10, 11] or simple descriptions of eye fixations without deep reasoning [15]. These methods, while offering some insights, lack true interpretability due to the lack of reasoning and prioritization within the context. This limitation is rooted in the inherent complexity of human attention. Human attention is a multifaceted phenomenon, influenced by a confluence of factors including visual cues, semantic understanding, individual preferences, and even cultural context. To generate meaningful explanations that unravel this interplay, a deep understanding of human cognition and perception is essential. Moreover, creating a model that can articulate its reasoning in a clear and concise manner poses a significant technical challenge.", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9601", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge the gap between saliency prediction and interpretability, we develop a saliency model that not only accurately predicts salient regions but also provides clear, human-understandable explanations. As shown in Fig. 1, our goal is to factorize saliency prediction with a weighted combination of spatial feature maps, namely semantic prototypes. Each semantic prototype represents the grounding of a semantic proposal (e.g., frisbee, fence, and sheep) relevant to the visual task, and its relevance is explained with a natural language description. In this example, \"frisbee\" and \"sheep\" have higher weights in saliency prediction, because \"to find the frisbee's location\" and \"to check if there is a sheep to the right of the fence\" are essential reasoning steps to answer the question \"Is there a sheep to the right of the fence that is to the right of the frisbee?\"", + "bbox": [ + 80, + 90, + 467, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we propose an explainable saliency (XSal) model that advances the state-of-the-art in saliency prediction and explanation. First, it integrates a vision-language model (VLM) to perform explicit reasoning for saliency prediction, generating step-by-step explanations and delivering natural language interpretations that clarify why specific regions are salient. Second, we introduce a contextual prioritization mechanism designed to handle the challenges of selecting truly important ones from a potentially large set of semantic explanations. This mechanism allows the model to focus on key elements, achieving interpretability without overcomplicating the representation. By integrating explicit reasoning with targeted selection, we create a cohesive framework that not only identifies and explains salient regions but also adapts dynamically to scene complexity.", + "bbox": [ + 80, + 318, + 467, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this paper makes three key contributions to the field of saliency prediction and explainable AI:", + "bbox": [ + 80, + 559, + 467, + 589 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. This paper presents the first saliency model to incorporate a vision-language model for explicit reasoning. This integration allows the model to not only identify salient regions but also articulate its reasoning process, providing clear and human-understandable explanations.", + "2. We introduce a novel contextual prioritization mechanism that prioritizes the most relevant information within the context of saliency prediction and explanation. This results in a more interpretable model by focusing on key relevant semantics and simplifying the representation.", + "3. We conduct extensive evaluations of our model, demonstrating its effectiveness in generating high-quality saliency maps while providing coherent, contextually relevant explanations. This comprehensive assessment highlights the model's ability to effectively bridge the gap between prediction accuracy and interpretability." + ], + "bbox": [ + 78, + 590, + 467, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 845, + 215, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Saliency Prediction. Saliency prediction models have traditionally focused on bottom-up approaches, relying on", + "bbox": [ + 80, + 871, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "low-level visual features in free-viewing conditions [5, 25, 27, 29]. However, with the rapid progress of deep learning methods, deep saliency models have emerged, achieving promising accuracy in predicting both task-free bottom-up saliency [26, 34, 52] and task-driven top-down attention [1, 30, 48]. A key strength of deep saliency models is their ability to extract high-level semantic features. Numerous deep neural network-based methods have boosted saliency prediction performance by implicitly encoding semantics using different approaches [46, 52]. However, such models often lack transparency and trustworthiness due to explicitly learned representations. While there has been growing interest in understanding saliency model mechanisms, existing approaches fall short of providing comprehensive explanations. For instance, Chen et al. [10, 11] analyze model behaviors using semantic prototypes and soft attention weights, but these approaches offer limited insight into the model's decision-making process, particularly when dealing with thousands of prototypes. Similarly, GazeXplain [15] generates plain natural language descriptions of fixation points, lacking a reasoning or prioritization mechanism to identify crucial factors driving its predictions. These limitations highlight the need for a more robust methodology that can uncover the underlying rationale behind saliency model behavior.", + "bbox": [ + 501, + 90, + 888, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed method addresses these limitations by introducing a novel framework that combines explicit vision-language reasoning with a hard attention mechanism. This ensures that only the most critical elements are selected for saliency prediction and explanation, which not only improves the accuracy of saliency predictions but also provides a more transparent and insightful understanding of the model's behavior.", + "bbox": [ + 503, + 472, + 888, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Language Models. Our work builds upon the growing field of explainable AI (XAI) within vision-language models (VLMs). Recent advancements in VLMs [13, 14, 17, 22, 31, 35, 47, 50, 51] have revolutionized multimodal understanding, enabling these models to seamlessly integrate visual and textual information. This progress has opened new avenues for developing XAI systems capable of providing interpretable insights into complex decision-making processes. For instance, researchers have explored using VLMs to generate natural language explanations for tasks such as image classification [23, 38], abusive language detection [19], and autonomous driving [7], allowing users to understand the rationale behind model predictions. Additionally, VLMs have been employed to enhance the interpretability of visual scanpath models, describing image features driving human attention to a sequence of fixation points [15].", + "bbox": [ + 503, + 595, + 888, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from these studies, by integrating textual context and introducing reasoning and prioritization mechanisms, we aim to provide deeper insights into the factors", + "bbox": [ + 503, + 856, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9602", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3d07ba1ed8a2d68840c8a8265484df918cc42fcb823737b58d6f0357720aaa9.jpg", + "image_caption": [ + "Figure 2. The proposed XSal architecture begins with a vision-language model generating semantic proposals from the input image, and uses a segmentation model to generate their corresponding semantic prototypes. A contextual prioritization module then selects the top-K relevant semantic prototypes, which are processed by a final saliency head to predict a focused saliency map." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 872, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "driving saliency predictions, improving the transparency and interpretability of saliency models.", + "bbox": [ + 75, + 420, + 470, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 465, + 168, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The human visual attention process is a captivating blend of explicit reasoning and selective focus, guiding us to prioritize relevant elements while navigating complex scenes. While existing saliency models have made strides, they often fall short of capturing this nuanced behavior, relying instead on dense representations or lacking a mechanism for explaining their choices. This incongruence between model behavior and human attention motivates us to introduce a novel approach that leverages VLMs to perform explicit reasoning and prioritizing the reasoning outcomes based on contextual relevance for saliency prediction and explanation.", + "bbox": [ + 75, + 492, + 470, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Explicit Reasoning", + "text_level": 1, + "bbox": [ + 76, + 684, + 258, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the heart of our approach lies a vision-language model (VLM), specifically MiniCPM-V 2.6 [53], which enables the model to identify and explain salient elements across task-driven or task-free contexts. Specifically, the VLM produces a set of semantic proposals $\\mathbf{C}$ and associated explanations $\\mathbf{R}$ :", + "bbox": [ + 75, + 708, + 468, + 799 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C}, \\mathbf {R} = \\operatorname {V L M} (\\mathbf {I}, \\mathbf {Q}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 815, + 468, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathbf{I}$ is the input image, and $\\mathbf{Q}$ is the question guiding attention. $\\mathbf{C} = \\{P_1,P_2,\\dots ,P_n\\}$ represents semantic proposals, each focusing on specific objects within the scene, while $\\mathbf{R}$ provides rationale for why each element was", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/628d3d12c74b2398fee286c3fcafdac4d6f799a9d2cffe8de5508e86eb4ce8c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 420, + 542, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Can you explain step by step how you analyzed the image to answer the question?", + "bbox": [ + 576, + 430, + 864, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Question: What is the drink in the cup made of plastic called?", + "bbox": [ + 568, + 452, + 869, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image:", + "bbox": [ + 578, + 464, + 614, + 474 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89abd22ae4177de0aef4df90d1c9717ab592f40342b4a9cfbfa76ab1a410d7b5.jpg", + "image_caption": [ + "Figure 3. The two-stage prompting method to obtain semantic proposals and explanations from the VLM." + ], + "image_footnote": [], + "bbox": [ + 683, + 465, + 867, + 573 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. I identified the drink in the cup.", + "2. I examined the cup's appearance to determine its material.", + "3. I confirmed the cup was plastic based on its transparency.", + "4. Finally, I considered common drinks served in plastic cups." + ], + "bbox": [ + 522, + 594, + 810, + 638 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/31010921043e78619354a5adfff5e6f2f4520b3607e990efa4cfabca4ce96c43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 849, + 585, + 888, + 616 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65b634e2207a108f8c6cd0ee3fe59aa8779ebfeac403771edff103d1fc3389bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 652, + 540, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In each step, which objects did you focus on, and why?", + "bbox": [ + 576, + 661, + 836, + 672 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. I focused on the drink in the cup, because it was necessary to identify the main object in question.", + "2. I focused on the cup to determine its material by examining its appearance and texture.", + "3. I focused on the plastic to confirm the material based on its transparency and common usage.", + "4. I focused on the drink in the plastic cup to the specific type of drinks served in such cups." + ], + "bbox": [ + 521, + 696, + 813, + 782 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2568075b9e3a8baa464cc452ec6ec8ffd7d59c3e3ab4d2daac9d942554f8263f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 849, + 688, + 888, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "selected, ensuring interpretability through alignment with human attention patterns. This integration supports a rea", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9603", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "soning process that mirrors human problem-solving, where decisions are informed by both visual cues and a structured understanding of why certain elements are relevant. To obtain the semantic proposals and explanations from the VLM, as shown in Fig. 3, we employ a two-stage prompting method:", + "bbox": [ + 75, + 90, + 468, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the initial planning stage, we ask the VLM to describe \"how you analyzed the image to answer the question\" (for top-down task-driven cases) or \"important objects in the image\" (for bottom-up free-viewing cases). The model responds with a coherent sequence of actions of reasoning steps to follow when answering the question. For example, to answer the question \"What is the drink in the cup made of plastic called?\", the reasoning steps include identifying the drink in the cup, examining the cup's appearance to determine its material, confirming the cup was plastic based on its transparency, and considering common drinks served in plastic cups.", + "bbox": [ + 75, + 181, + 470, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following this planning, the VLM is asked to explain what specific objects it focused on in each step, and explain the reason. It then proceeds to generate the set of semantic proposals $\\mathbf{C}$ (such as drink, cup, and plastic) and their corresponding explanations $\\mathbf{R}$ . Finally, to ensure precision in the final output, the VLM is also prompted to verify that the identified semantic proposals are concrete and relevant, addressing potential limitations in model consistency.", + "bbox": [ + 75, + 364, + 468, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This layered approach replicates human selective attention and explanatory reasoning, ensuring the model's focus and reasoning align closely with human attention patterns.", + "bbox": [ + 76, + 486, + 468, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After identifying semantic proposals through the VLM, we utilize an open-vocabulary segmentation model, ClipSeg [37], to extract the visual grounding of these proposals on the image. ClipSeg's ability to handle open-vocabulary concepts is crucial as the language model may describe elements that don't perfectly align with traditional segmentation categories. This step bridges the semantic reasoning process with visual representations, allowing for visual validation of the model's reasoning. While ClipSeg effectively outlines entire object contours, human attention typically focuses on smaller, more localized regions. Therefore, instead of using the segmentation results directly for saliency prediction, we use them as binary masks to filter the image features, and apply a convolutional layer $\\mathcal{F}$ to obtain fine-grained localization. This results in a sequence of focused spatial maps, namely semantic prototypes, where each prototype $S_{i}$ aligns with a semantic proposal $P_{i}$ , localizing its representations in the image. This process pinpoints focused regions that naturally attract human attention, associating them with natural language explanations.", + "bbox": [ + 75, + 532, + 470, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Contextual Prioritization Mechanism", + "text_level": 1, + "bbox": [ + 76, + 845, + 398, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach utilizes a contextual prioritization mechanism to extract the most relevant semantic prototypes for saliency", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prediction. Different from existing saliency models that often integrate hundreds to thousands of spatial feature maps with linear weights, humans naturally focus on a limited number of features, while other elements, though present, are not prioritized to the same extent. Our proposed contextual prioritization is a highly selective approach, which only allows attention to be directed by the most relevant features. It introduces a dynamic approach that adapts to varying numbers of semantic prototypes. This flexibility allows the model to handle scenes of different complexities effectively.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our contextual prioritization mechanism adjusts importance by applying global average pooling on the extracted semantic prototypes, followed by a linear layer to generate initial weights. These weights are then sorted, and only the top $K$ and their corresponding features are selected. A softmax operation is applied to normalize the selected weights, yielding a refined representation:", + "bbox": [ + 496, + 257, + 890, + 363 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} = \\operatorname {L i n e a r} (\\operatorname {G A P} (\\mathbf {S})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 377, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {\\text {t o p K}} = \\operatorname {S o f t m a x} (\\operatorname {T o p K} (\\mathbf {w})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 414, + 890, + 430 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {O} = \\sum_ {i = 1} ^ {K} \\mathbf {w} _ {\\text {t o p K}, i} \\cdot \\mathbf {S} _ {\\text {t o p K}, i} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 446, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{S}$ represents a collection of semantic prototypes, $\\{S_1,S_2,\\dots ,S_n\\}$ . The global average pooling (GAP) and subsequent linear transformation extract initial weights $\\mathbf{w}$ for each prototype, allowing the model to assign varying importance to each semantic element. By selecting and normalizing the top $K$ weights, we obtain $\\mathbf{w}_{\\mathrm{topK}}$ and the corresponding $\\mathbf{S}_{\\mathrm{topK}}$ . Here, $\\mathbf{S}_{\\mathrm{topK}}$ corresponds to the semantic prototypes associated with the top $K$ weights in $\\mathbf{w}$ , determined by sorting. This ensures that the model prioritizes the most significant semantic elements, aligning its focus with the most contextually important regions, and discarding less relevant elements. Finally, the aggregated saliency feature O, the weighted combination of the top $K$ semantic prototypes, is processed by a saliency head to produce the final saliency prediction. This saliency head consists of only a single linear layer (with bias set to 0) that simply compresses the channel dimension, acting as a weighted sum without introducing any complex black-box behavior.", + "bbox": [ + 496, + 493, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure adequate feature optimization and prevent limited feature selection from hindering convergence, we introduce a Progressive Top-K Reduction strategy. Rather than starting with a fixed, small set of key features, this strategy initially preserves a larger number of semantic elements, enabling comprehensive feature optimization in early training. As training progresses, the $K$ value is gradually reduced, guiding the model to focus on the most significant features. The progression of $K$ across epochs is defined as", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9604", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {t} = K _ {\\text {i n i t}} - \\frac {\\left(K _ {\\text {i n i t}} - K _ {\\text {f i n a l}}\\right) \\times t}{T - 1} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 101, + 468, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K_{t}$ is the $K$ at epoch $t$ , $K_{\\mathrm{init}}$ and $K_{\\mathrm{final}}$ are the initial and final values of $K$ , and $T$ is the total number of epochs. This progressive reduction allows the model to gradually shift from broad feature optimization to a selective focus on the most critical elements. With this strategy, the contextual prioritization mimics the adaptive prioritization seen in human attention, flexibly weighing elements based on their relevance and adapting to scenes with different numbers of semantic prototypes.", + "bbox": [ + 75, + 136, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 285, + 209, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents a comprehensive evaluation of our proposed model, examining its performance and explainability across both task-driven and task-free eye-tracking datasets. We conduct an ablation study to further investigate the contributions of individual components, and present qualitative examples to illustrate the model's capabilities in generating interpretable saliency maps and explanations. Additional experimental results and analyses are provided in the supplementary material for further reference.", + "bbox": [ + 75, + 310, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 454, + 266, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Our experiments are conducted on two eye-tracking datasets: The AiR dataset [9], derived from the GQA dataset, comprises images and questions paired with eye-tracking data from 20 participants answering those questions. This setup provides a realistic scenario for evaluating saliency models, as it captures human attention in the context of a specific task (question answering). The OSIE [49] dataset is a free-viewing dataset featuring multiple salient objects competing for attention in the same image context, suitable for developing and evaluating saliency models that consider object- and semantic-level information. In our experiments, we split both datasets into training, validation, and test sets following [12, 16].", + "bbox": [ + 75, + 478, + 468, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared Models. We compare our model's performance against several state-of-the-art saliency prediction models including SAM [18], SALICON [26], DINet [52], as well as explainable models such as Chen et al. [10] and GazeXplain [15]. Note that for GazeXplain [15] we aggregate fixation points from its predicted scanpaths into saliency maps for comparison. By comparing our model with these diverse methods, we aim to demonstrate its performance in various aspects, including general saliency prediction and explainability.", + "bbox": [ + 75, + 674, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To comprehensively evaluate our model, we utilize a combination of metrics that assess saliency prediction performance and explainability. 1) Saliency Prediction Metrics: We employ standard metrics to evaluate the alignment between our model's predicted", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "saliency maps and human attention patterns. These metrics are Area Under the ROC Curve (AUC) [3], Normalized Scanpath Saliency (NSS) [41], and Correlation Coefficient (CC) [39], which provide quantitative insights into the accuracy and relevance of the generated saliency maps. 2) Explainability Metrics: We evaluate the interpretability of saliency predictions adapting three XAI metrics: AUC [2] assesses model stability by calculating the area under the curve that represents changes in model performance as different proportions of semantic prototypes are progressively perturbed. Note that this AUC differs from the AUC used in saliency evaluation: here, it measures the model's sensitivity to perturbations, while saliency AUC measures alignment with human attention data. To facilitate distinction, we refer to this interpretability metric as AUC-E throughout this paper. Area Over the Perturbation Curve (AOPC) [6, 40] quantifies the sensitivity of the model to perturbations by measuring the change in NSS as semantic prototypes are sequentially altered. Log-odds Score (LOdds) [43, 44] evaluates the change in NSS as semantic prototypes are perturbed, focusing on the degree of reduction in NSS in response to prototype alterations. These metrics provide a comprehensive framework for assessing the quality of our model's explanations, enabling us to gauge not only where the model focuses, but also the robustness of its interpretative reasoning under perturbations.", + "bbox": [ + 496, + 90, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. To align the model's saliency predictions effectively with human attention patterns, we employ a combination of loss functions, including NSS, CC, and Kullback-Leibler Divergence (KLD) [33]. Our implementation follows [10]. These loss functions provide complementary signals that guide the model to capture nuanced attention details, helping to achieve balanced performance across various saliency prediction metrics.", + "bbox": [ + 496, + 488, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our model employs a dilated ResNet-50 [24] backbone, following the setting used in DINet [52]. This architecture incorporates dilated convolutions to expand the receptive field without additional downsampling, which enhances the model's capacity to capture high-resolution spatial details essential for accurate saliency prediction. The model is optimized using the Adam optimizer [32] with an initial learning rate of 4e-4 and a weight decay of 1e-7. To ensure stable convergence, the learning rate is adaptively reduced by a decay factor of 0.1 every two epochs, while gradient clipping is applied with a threshold of 0.1 to prevent gradient explosion. Training is conducted over 10 epochs with a batch size of 10. To ensure stable training and prevent convergence issues, we employ the proposed Progressive Top-K Reduction strategy, which begins with a broader feature selection and gradually narrows the focus to the most significant features. Specifically, we set the initial $K$ value to 20, which progressively reduces to a final value of 3 over 3 epochs, enabling adaptive and selec", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9605", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fbc66cdfd3b44deabd0477d599009efe3e484c5b0929f746ee85a4c4265a4cf1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AiROSIE
NSSCCNSSCC
SAM [18]1.650.622.700.65
SALICON [28]1.700.632.750.63
DINet [52]1.770.632.880.63
Chen et al. [10]1.760.632.910.64
GazeXplain [15]1.850.662.530.74
XSal1.940.702.960.72
", + "bbox": [ + 109, + 88, + 436, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparative results of saliency prediction across the AiR and OSIE datasets.", + "bbox": [ + 76, + 273, + 467, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tive feature prioritization.", + "bbox": [ + 76, + 330, + 246, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Saliency Prediction Results", + "text_level": 1, + "bbox": [ + 76, + 357, + 320, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents a quantitative comparison of saliency prediction performance across the AiR (task-driven) and OSIE (task-free) saliency datasets for our proposed model and several state-of-the-art methods. As shown in the table, our proposed model consistently outperforms other methods on the AiR dataset, achieving the highest NSS score of 1.94 and the highest CC score of 0.70. This suggests that our model effectively captures top-down human attention patterns in task-driven scenarios. Our model also achieves the highest NSS score of 2.96 on the OSIE dataset. Though our CC score (0.72) is slightly lower than GazeXplain (0.74), it still demonstrates strong performance in capturing the overall distribution of human attention in task-free image-viewing. The consistently high NSS scores across both datasets indicates that our model accurately predicts the locations where humans are mostly likely to fixate their gaze, demonstrating a strong accuracy of predicting the most salient regions. However, our model might be slightly less effective at capturing the overall distribution of human attention in images, because image regions with low saliency (low agreements across human observers) are less explainable and likely to be discarded by the contextual prioritization mechanism. Overall, the quantitative results demonstrate that our proposed method significantly improves saliency prediction performance compared to existing approaches. The incorporation of explicit reasoning and selective attention mechanisms contributes to more accurate and robust models that effectively capture the complexity of human visual attention.", + "bbox": [ + 76, + 381, + 468, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Saliency Explanation Results", + "text_level": 1, + "bbox": [ + 76, + 830, + 334, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 presents the quantitative results based on explainability metrics, comparing our model's performance with Chen et al. [10] on both the AiR and OSIE datasets. Notably,", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our model demonstrates significantly lower AUC-E, LOdds, and higher AOPC scores compared to Chen et al. Specifically, on the AiR dataset, our model significantly outperforms Chen et al. with a lower AUC-E (0.713 vs. 0.899) and a higher AOPC (0.748 vs. 0.241). The substantial improvement in LOdds (-5.157 vs. -0.143) also reflects the model's ability to focus on the most relevant elements, minimizing irrelevant attention. On the OSIE dataset, a similar trend is observed. Our model achieves a better AUC-E (0.685 vs. 0.810) and significantly higher AOPC (1.115 vs. 0.586) compared to Chen et al. Additionally, the considerable decrease in LOdds (-4.169 vs. -0.250) further supports our model's capability to highlight key semantic concepts more effectively, thereby improving interpretability in a task-free viewing context.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These results indicate that our model is more sensitive to changes in semantic prototypes, highlighting the stronger reliance on these prototypes for generating accurate saliency predictions. The lower AUC-E and LOdds scores suggest that perturbing prototypes leads to a more substantial decrease in performance for our model, highlighting the critical role of semantic understanding in our approach. Conversely, the higher AOPC score signifies a greater sensitivity to changes in prototypes, revealing a stronger correlation between the model's performance and the identified semantic elements. This reinforces the notion that our model's interpretability is closely tied to the semantic prototypes it extracts, providing a more transparent and interpretable reasoning process compared to methods that rely on complex aggregation of numerous features. Overall, these results emphasize the importance of explicit reasoning and selective attention in achieving a higher degree of interpretability and understanding in saliency prediction.", + "bbox": [ + 496, + 318, + 890, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 603, + 651, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our ablation study, conducted on both the AiR and OSIE datasets, reveals the significant contributions of each component to our model's performance and explainability. As shown in Tab. 3, a simple baseline model, using only a dilated ResNet-50 backbone [24] with a saliency head, achieves reasonable performance in both settings, yielding an NSS of 1.777 on AiR and 2.835 on OSIE. The baseline model lacks explainability, with a LOdds of -0.131 and -0.246 on these datasets, respectively.", + "bbox": [ + 496, + 627, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By incorporating the VLM for reasoning, our model links saliency with task-relevant semantic elements and significantly improves both accuracy and interpretability across both datasets. For instance, on the AiR dataset, the NSS increases to 1.917 and the CC score improves to 0.703, while the AUC remains relatively stable at 0.858. This highlights the crucial role of task-oriented alignment and semantic understanding in generating clear and contextual explanations for saliency predictions. This is further evidenced", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9606", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b1c3c0bce96afc34cbd2fcc51ae6df50b37ff53146b61d89e44ebd8f7007b742.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AiROSIE
AUC-E↓AOPC↑LOdds↓AUC-E↓AOPC↑LOdds↓
Chen et al. [52]0.8990.241-0.1430.8100.586-0.250
XSal0.7130.748-5.1570.6851.115-4.169
", + "bbox": [ + 202, + 88, + 767, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/01aea2a12d57658fa9e1746fcdb29df7616a8946612fbcd8843845bd525f2cc3.jpg", + "table_caption": [ + "Table 2. Comparative results of explainability metrics across the AiR and OSIE datasets." + ], + "table_footnote": [], + "table_body": "
DatasetBaselineReasoningContPriNSSCCAUCAUC-E↓AOPC ↑LOdds ↓
AiR✓1.7770.6400.8490.9100.225-0.131
✓✓1.9170.7030.8580.7000.749-5.077
✓✓✓1.9420.7010.8580.7130.748-5.157
OSIE✓2.8350.6810.8850.8230.567-0.246
✓✓2.9040.7170.8970.7650.876-1.738
✓✓✓2.9650.7260.8950.6851.115-4.169
", + "bbox": [ + 133, + 210, + 838, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on the impacts of Baseline, Reasoning, and Contextual Prioritization (ContPri).", + "bbox": [ + 178, + 349, + 790, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by the significant improvements in interpretability metrics – the AUC-E drops to 0.700 and the AOPC increases to 0.749, suggesting our model's ability to provide more coherent explanations for task-driven saliency.", + "bbox": [ + 75, + 391, + 468, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, the contextual prioritization mechanism exhibits a more pronounced impact, especially in the OSIE task-free setting. For example, on OSIE, the NSS increases to 2.965, the CC improves to 0.726, and the AUC remains relatively stable at 0.895. This suggests that prioritizing the most salient elements plays a crucial role in handling complex scenes with multiple potential salient objects, ultimately leading to more concise and clear attention maps. This is further evidenced by the significant improvements in interpretability metrics – the AUC-E decreases to 0.685, the AOPC increases to 1.115, and the LOdds decreases to -4.169. This greater impact on the OSIE dataset could be attributed to the presence of a large number of potential semantic prototypes in task-free settings, making selective focus particularly valuable for distilling relevant information and improving interpretability.", + "bbox": [ + 75, + 452, + 468, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These findings demonstrate that the combination of explicit reasoning and contextual prioritization mechanisms, implemented in our model, is essential for achieving superior performance and explainability in both task-driven and task-free saliency prediction.", + "bbox": [ + 75, + 696, + 468, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 785, + 269, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative analysis of our model reveals its ability to effectively integrate task-driven guidance and contextual prioritization to generate accurate saliency predictions and explanations. As illustrated in Fig. 4, the model aligns its attention with the goals of each question, selectively focusing on essential objects while ignoring irrelevant elements. For ex", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ample, in questions about the position of a skateboard and a red cone (row 1), the model prioritizes these objects over traditionally salient regions like the skateboarder, demonstrating the advantage of explicit reasoning. The model also adapts well to abstract concepts, such as interpreting \"modern\" style in a question about a faucet (row 2). Finally, the model excels in context prioritization by isolating key elements in complex scenes, highlighting the person, bag, and cars in a question about cars (row 3) while filtering out other visual elements. These qualitative findings demonstrate the model's ability to produce explanations that are closely aligned with the model's internal reasoning, ensuring outputs that are both interpretable and reliable.", + "bbox": [ + 496, + 391, + 890, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond its interpretability, our model shows significant performance advantages in accuracy and task relevance. In Fig. 5, our model achieves the highest accuracy among state-of-the-art methods, aligning closely with ground truth. For example, in the question \"What is the drink in the cup made of plastic called?\", it focuses on the plastic cup, unlike other models that are distracted by irrelevant food items (row 1), showcasing the strength of our explicit reasoning approach. Additionally, our model demonstrates strong context prioritization, effectively concentrating on relevant elements even in complex scenes with multiple distractions. For instance, in the question \"Is the fire hydrant to the right or to the left of the garbage can?\", our model isolates the hydrant and garbage can, while other models diffuse attention across the scene (row 2), highlighting its ability to prioritize contextually important elements. Furthermore, in dynamic scenes, our model reliably focuses on salient objects. In the question \"Are both the helmet and the bat the same color?\", it highlights the helmet and bat specifically, unlike other models that diffuse attention (row 3), demonstrating", + "bbox": [ + 496, + 598, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9607", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25e90289825018cfd898e673aa37ddf6e2c7ba241d06207868706693dc8ec4b5.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 120, + 104, + 199, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38e540eadb7bf36144315c8250cc891ea9f43e943773b62d2dea4a3125fc4d7e.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 202, + 104, + 284, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/225f8ddc761de2c3e62a6392ab736cf40885eb6ccd38ed749b2080a6ba68c021.jpg", + "image_caption": [ + "Prediction" + ], + "image_footnote": [], + "bbox": [ + 287, + 104, + 366, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84c1bc806cd7f4b63d1e486e192c211fd462c20982c68026a56e728178eaf74b.jpg", + "image_caption": [ + "Prototype 1", + "$\\mathbf{B}\\mathbf{a}\\mathbf{g}(w_1 = 0.378)$" + ], + "image_footnote": [], + "bbox": [ + 444, + 104, + 526, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The question specifies a bag, so it is relevant to locate its position in the image and use it as a reference point for determining the direction of 'to the right'.", + "bbox": [ + 370, + 162, + 521, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f38d6140f5c8ab2af3136b689b8a84370b2dbc79eb9edca1feee4652f9a3b24.jpg", + "image_caption": [ + "Prototype 2", + "Person $(w_{2} = 0.330)$" + ], + "image_footnote": [], + "bbox": [ + 596, + 104, + 676, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To identify the reference point for 'to the right' in relation to the bag and determine where cars are located relative to this person.", + "bbox": [ + 527, + 162, + 673, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/481dd8e1266e3d3564130bd7611455502be11aafc81faea7241bfe0db5c111ea.jpg", + "image_caption": [ + "Prototype 3", + "Cars $(w_{3} = 0.291)$" + ], + "image_footnote": [], + "bbox": [ + 761, + 104, + 841, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cars are mentioned in the question and need to be identified in their location relative to the person and bag.", + "bbox": [ + 705, + 162, + 839, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: Are there any cars to the right of the person that is to the right of the bag?", + "bbox": [ + 124, + 152, + 192, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c6a07fc5f102915ce0553d72bbb60219b49817fd259f19f8d7a5b1b19c49fde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 213, + 199, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35844a742d6a68031532d5eaa8f3db51734d31fc7f525bb354163bd6931f77e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 213, + 282, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c01f2fb14723aa66248dc1b97865973157f76fe7349766966c28685229b1c61e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 213, + 366, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: Does the faucet to the left of the vase look modern and silver?", + "bbox": [ + 122, + 261, + 196, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f8e80c2a4d00e6ac5f1a5cff738bc20b70738db238aaef42aed6535b095a643.jpg", + "image_caption": [ + "Vase $(w_{1} = 0.357)$" + ], + "image_footnote": [], + "bbox": [ + 444, + 213, + 524, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The vase is mentioned in the question and serves as a reference point for the faucet's location.", + "bbox": [ + 372, + 271, + 521, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfa26a18d07460e3f87cf10c9722e5d224670706158e2546bdb85409fa49e108.jpg", + "image_caption": [ + "Faucets $(w_{2} = 0.319)$" + ], + "image_footnote": [], + "bbox": [ + 596, + 213, + 676, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The question specifically asks about the faucet's appearance, so it was a key element in answering.", + "bbox": [ + 529, + 271, + 671, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73418ff3002998be5af46fd654c9e29e9090b747eda550ab2292fe827469eecf.jpg", + "image_caption": [ + "Modern design $(w_{3} = 0.329)$" + ], + "image_footnote": [], + "bbox": [ + 761, + 213, + 843, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The question refers to the faucets style, which can be described as modern.", + "bbox": [ + 694, + 271, + 839, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9a5f490c053d17f027d35927a1e85c927106887d14d56fa75b9da74e4f451cb.jpg", + "image_caption": [ + "Figure 4. Qualitative examples from our methods. Each semantic prototype is shown along with their saliency weights and explanations." + ], + "image_footnote": [], + "bbox": [ + 120, + 311, + 199, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e9996dc591349d633107f6234428d5fc005a96de2d0d05bde846f4c5842d3ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 311, + 282, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f51a130be62d71c429bb45d861dc8933dba5f406f22f614ab9a56030c973c35e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 311, + 366, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ae8d40964d00dd12a4c791ee3ae85d2a2589bdc384616b6872fd50eb1f71313.jpg", + "image_caption": [ + "Skateboard $(w_{1} = 0.371)$" + ], + "image_footnote": [], + "bbox": [ + 444, + 311, + 524, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The skateboard is the object being asked about in terms of its position relative to the red cone.", + "bbox": [ + 370, + 369, + 521, + 398 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e03a9a852250e436a2c56ec46ddeb3a5baa887a71598426a9a6974ac475c9ad8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 311, + 676, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Red cone $(w_{2} = 0.344)$ red cone is the reference point for determining the position of the skateboard relative to the question.", + "bbox": [ + 571, + 359, + 671, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d32b077db8140c35298b7196607f4e2f92b2363dd10baba7698c8f6cef0bda9f.jpg", + "image_caption": [ + "Skateboarder $(w_{3} = 0.283)$" + ], + "image_footnote": [], + "bbox": [ + 761, + 311, + 841, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The skateboarder position and orientation are crucial in understanding where the skateboard is in relation to the red cone.", + "bbox": [ + 683, + 369, + 839, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: What is the drink in the cup made of plastic called?", + "bbox": [ + 124, + 467, + 199, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: Is the fire hydrant to the right or to the left of the garbage can?", + "bbox": [ + 124, + 526, + 215, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: Are both the helmet and the bat the same color?", + "bbox": [ + 125, + 585, + 210, + 617 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43436310f7f66fd499635549f286c6794cd96306c2a0f62fd1d30b9df64910a2.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 223, + 465, + 326, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ebe140a887c7385c263615a87b8b035762f39e2a3d09592ed45c81032bf215d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 523, + 326, + 582 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7defecdca0b81525a3aa8753daa761923d9cc49d9538aa2abfff39a2ae4d18b4.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of saliency predictions across models (XSal, DINet, SALICON, SAM). Each row shows an input question and image, with the corresponding ground truth attention map and model outputs." + ], + "image_footnote": [], + "bbox": [ + 223, + 584, + 326, + 643 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/393c224a27810f79eef0c9f6866d0cc7f158f608acec605dd7db468893378402.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 330, + 465, + 431, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cd7c79544fa4b95386325409697857c67173c1b9c00ffcf9732a2e5d1fa5fdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 523, + 431, + 582 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/647d6486bc775446863a22177768988618e0e8271d5f3cac66345ca48d287d93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 584, + 431, + 643 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acd996e56cfd054df6d94413eca7c2d58e3bcc16ec18890d0d7a909263870ea0.jpg", + "image_caption": [ + "XSal" + ], + "image_footnote": [], + "bbox": [ + 433, + 465, + 535, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/567b83f83a3be5a57484d6d3b72d975a82fb4ed93db6394b88d5e2dd0a0b1ff9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 523, + 535, + 582 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bea1ecdd7b565fbe3d857a8d75e4e93094c7be2e06bd607b021ad988b4f0145f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 584, + 535, + 643 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cfa01521c156af9fe089fb56433eb3a05a687386c3a95660e3b7a020cf6619c.jpg", + "image_caption": [ + "DINet" + ], + "image_footnote": [], + "bbox": [ + 537, + 465, + 640, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da89bbf62cbc4c925612f7e68cfab91804fe08a3ff6018c72e461d86e2c9fbd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 523, + 640, + 582 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/380397bdde725dadd0d89fa07d72d1677a73186cac69984515e9c84c33ffb081.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 583, + 640, + 643 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3645347d2168522f59ac6de9d09ed894b51a33a810691b007f0f82b3cef13605.jpg", + "image_caption": [ + "SALICON" + ], + "image_footnote": [], + "bbox": [ + 640, + 465, + 743, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c01d014c3b0f577e5d13bae3ca7346cffb1b845adbc8368694eb8743df0f1ae4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 523, + 743, + 582 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdcd09de8683348dadcc95ad5ff4790c51e5e8d716cce47cd8d969dbd96b67e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 583, + 743, + 643 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f074d7b10c467ab73ca4737ebe44ad4f427a0567c03e7e14b214ba3817b0984.jpg", + "image_caption": [ + "SAM" + ], + "image_footnote": [], + "bbox": [ + 746, + 465, + 849, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba6375fd8a3649f61e058544dfb38db17c80f92316777a8aad0c64c4400f8a73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 523, + 849, + 582 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec76b46edcd49346e4109185e1b90ff4d8888ddfe40cbfe8334064739b556055.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 583, + 849, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "its adaptability to diverse scenarios. These examples highlight our model's superior performance in generating accurate and contextually relevant saliency predictions.", + "bbox": [ + 75, + 710, + 468, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 768, + 194, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper has presented a novel approach for saliency prediction that combines explicit reasoning and contextual prioritization, generating both accurate and explainable saliency predictions. We have demonstrated that our method, powered by a VLM, outperforms existing models in terms of accuracy and explainability. This is achieved by leveraging an explicit reasoning process that allows the", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model to provide clear explanations for its predictions and a contextual prioritization mechanism that focuses on a few truly important elements, mimicking human attention behavior. Our approach not only achieves state-of-the-art saliency prediction accuracy but also offers valuable insights into the decision-making processes of saliency models. This advancement contributes to the development of more trustworthy and reliable saliency models.", + "bbox": [ + 496, + 710, + 893, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 845, + 658, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by NSF Grant 2143197 and Fasikl Incorporated.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9608", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2", + "[2] Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 5", + "[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5", + "[4] Patrick Le Callet and Ernst Niebur. Visual attention and applications in multimedia technologies. Proceedings of the Institution of Electrical Engineers, 2013. 1", + "[5] Moran Cerf, E Paxon Frady, and Christof Koch. Faces and text attract gaze independent of the task: Experimental data and computer model. Journal of vision, 9(12):10-10, 2009. 2", + "[6] Hanjie Chen, Guangtao Zheng, and Yangfeng Ji. Generating hierarchical explanations on text classification via feature interaction detection. In Proceedings of the 2020 Conference on Association for Computational Linguistics (ACL), 2020. 5", + "[7] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2", + "[8] Nuo Chen, Jin Xie, Jing Nie, Jiale Cao, Zhuang Shao, and Yanwei Pang. Attentive alignment network for multispectral pedestrian detection. In Proceedings of the 31st ACM international conference on multimedia, pages 3787-3795, 2023. 1", + "[9] Shi Chen, Ming Jiang, Jinhui Yang, and Qi Zhao. AiR: Attention with reasoning capability. In European Conference on Computer Vision (ECCV), 2020. 5", + "[10] Shi Chen, Ming Jiang, and Qi Zhao. What do deep saliency models learn about visual attention? In Advances in Neural Information Processing Systems, pages 9543-9555. Curran Associates, Inc., 2023. 1, 2, 5, 6", + "[11] Shi Chen, Ming Jiang, and Qi Zhao. Deep learning to interpret autism spectrum disorder behind the camera. IEEE Transactions on Cognitive and Developmental Systems, pages 1-12, 2024. 1, 2", + "[12] Xianyu Chen, Ming Jiang, and Qi Zhao. Predicting human scanpaths in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10876-10885, 2021. 5", + "[13] Xianyu Chen, Ming Jiang, and Qi Zhao. Leveraging human attention in novel object captioning. In Proceedings of the" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Joint Conference on Artificial Intelligence (IJ-CAI), 2021. 1, 2", + "[14] Xianyu Chen, Ming Jiang, and Qi Zhao. Self-distillation for few-shot image captioning. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2", + "[15] Xianyu Chen, Ming Jiang, and Qi Zhao. Gazexplain: Learning to predict natural language explanations of visual scanpaths. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2, 5, 6", + "[16] Xianyu Chen, Ming Jiang, and Qi Zhao. Beyond average: Individualized visual scanpath prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5", + "[17] Xianyu Chen, Jinhui Yang, Shi Chen, Louis Wang, Ming Jiang, and Qi Zhao. Every problem, every step, all in focus: Learning to solve real-world problems with integrated attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI), 2024. 2", + "[18] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting Human Eye Fixations via an LSTM-based Saliency Attentive Model. TIP, 27(10):5142-5154, 2018. 5, 6", + "[19] Chiara Di Bonaventura, Lucia Siciliani, Pierpaolo Basile, Albert Merono Penuela, and Barbara McGillivray. Is explanation all you need? an expert survey on llm-generated explanations for abusive language detection. In Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), 2024. 2", + "[20] Ning Ding, Ce Zhang, and Azim Eskandarian. Saliendet: A saliency-based feature enhancement algorithm for object detection for autonomous driving. IEEE Transactions on Intelligent Vehicles, 9(1):2624-2635, 2023. 1", + "[21] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Zhaohui Che, Yi Fang, Xiaokang Yang, Jesús Gutierrez, and Patrick Le Callet. A dataset of eye movements for the children with autism spectrum disorder. In ACM Multimedia Systems Conference (MMSys), 2019. 1", + "[22] Yang Feng, Lin Ma, Wei Liu, and Jiebo Luo. Unsupervised image captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[23] Songhao Han, Le Zhuo, Yue Liao, and Si Liu. Llms as visual explainers: Advancing image classification with evolving visual descriptions. arXiv preprint arXiv:2311.11904, 2023. 2", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5, 6", + "[25] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2", + "[26] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In ICCV, pages 262-270, 2015. 1, 2, 5", + "[27] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9609", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Pattern Analysis and Machine Intelligence, 20(11):1254-1259, 1998. 2", + "[28] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Silicon: Saliency in context. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6", + "[29] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In 2009 IEEE 12th international conference on computer vision, pages 2106-2113. IEEE, 2009. 2", + "[30] Christopher Kanan, Mathew H Tong, Lingyun Zhang, and Garrison W Cottrell. Sun: Top-down saliency using natural statistics. Visual cognition, 17(6-7):979-1003, 2009. 2", + "[31] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Image captioning with very scarce supervised data: Adversarial semi-supervised learning approach. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 2", + "[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[33] Solomon Kullback. Information Theory and Statistics. Wiley, 1959. 5", + "[34] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. In ICLR workshop, 2015. 2", + "[35] Iro Laina, Christian Rupprecht, and Nassir Navab. Towards unsupervised image captioning with shared multimodal embeddings. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2", + "[36] Jianxun Lou, Hanhe Lin, David Marshall, Dietmar Saupe, and Hantao Liu. Transalnet: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 1", + "[37] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7086-7096, 2022. 4", + "[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2", + "[39] Olivier Le Meur, Patrick Le Callet, and Dominique Barba. Predicting visual fixations on video based on low-level visual features. Vision Research, 47(19):2483 - 2498, 2007. 5", + "[40] Dong Nguyen. Comparing automatic and human evaluation of local explanations for text classification. In Proceedings of the 2018 Conference on North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 5", + "[41] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397 - 2416, 2005. 5", + "[42] Thammathip Piumsomboon, Gun Lee, Robert W. Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In IEEE Symposium on 3D User Interfaces (3DUI), 2017. 1", + "[43] Yao Qiang, Deng Pan, Chengyin Li, Rhongho Jang, and Dongxiao Zhu. Attcat: Explaining transformers via attentive" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "class activation tokens. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2022. 5", + "[44] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning (ICML), 2017. 5", + "[45] Vincent Sitzmann, Ana Serrano, Amy Pavel, Maneesh Agrawala, Diego Gutierrez, Belen Masia, and Gordon Wetzstein. Saliency in vr: How do people explore virtual environments? IEEE Transactions on Visualization and Computer Graphics, 24(4):1633-1642, 2018. 1", + "[46] Yubao Sun, Mengyang Zhao, Kai Hu, and Shaojing Fan. Visual saliency prediction using multi-scale attention gated network. Multimedia Systems, 28(1):131-139, 2022. 2", + "[47] Hamed R. Tavakoli, Rakshith Shetty, Ali Borji, and Jorma Laaksonen. Paying attention to descriptions generated by image captioning models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[48] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question answering. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 451-466. Springer, 2016. 2", + "[49] Juan Xu, Ming Jiang, Shuo Wang, Mohan S Kankanhalli, and Qi Zhao. Predicting human gaze beyond pixels. Journal of vision, 14(1):28-28, 2014. 5", + "[50] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2", + "[51] Jinhui Yang, Xianyu Chen, Ming Jiang, Shi Chen, Louis Wang, and Qi Zhao. VisualHow: Multimodal problem solving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[52] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. TMM, 22(8):2163-2176, 2020. 1, 2, 5, 6, 7", + "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3", + "[54] Wei Zhang and Hantao Liu. Study of saliency in objective video quality assessment. IEEE Transactions on Image Processing, 26(3):1275-1288, 2017. 1" + ], + "bbox": [ + 501, + 92, + 890, + 770 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9610", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_model.json b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e80a9fa9f37d2e460dce0e0066eb654a0c7be5dc --- /dev/null +++ b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_model.json @@ -0,0 +1,2849 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.87, + 0.152 + ], + "angle": 0, + "content": "Explainable Saliency: Articulating Reasoning with Contextual Prioritization" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.182, + 0.407, + 0.199 + ], + "angle": 0, + "content": "Nuo Chen" + }, + { + "type": "text", + "bbox": [ + 0.447, + 0.182, + 0.539, + 0.199 + ], + "angle": 0, + "content": "Ming Jiang" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.182, + 0.648, + 0.199 + ], + "angle": 0, + "content": "Qi Zhao" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.2, + 0.684, + 0.234 + ], + "angle": 0, + "content": "Department of Computer Science and Engineering University of Minnesota" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.238, + 0.63, + 0.251 + ], + "angle": 0, + "content": "{chen9029, mjiang, qzhao}@umn.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.636 + ], + "angle": 0, + "content": "Deep saliency models, which predict what parts of an image capture our attention, are often like black boxes. This limits their use, especially in areas where understanding why a model makes a decision is crucial. Our research tackles this challenge by developing an explainable saliency (XSal) model that not only identifies what is important in an image, but also explains its choices in a way that makes sense to humans. We achieve this by using vision-language models to reason about images and by focusing the model's attention on the most crucial information using a contextual prioritization mechanism. Unlike prior approaches that rely on fixation descriptions or soft-attention based semantic aggregation, our method directly models the reasoning steps involved in saliency prediction, generating selectively prioritized explanations clarify why specific regions are prioritized. Comprehensive evaluations demonstrate the effectiveness of our model in generating high-quality saliency maps and coherent, contextually relevant explanations. This research is a step towards more transparent and trustworthy AI systems that can help us understand and navigate the world around us." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.21, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Humans have this amazing ability to look at a scene and instantly know what is important. Saliency models, the cornerstone of computer vision, aim to mimic human gaze, identifying the most captivating regions within an image. These models have found wide application in image and video quality assessment [4, 54], virtual reality [42, 45], autonomous systems [8, 20], image captioning [13], and clinical diagnostics [11, 21]. However, while current saliency models [26, 36, 52] excel at replicating generic gaze patterns, they often operate as enigmatic black boxes. This lack of transparency hinders their application in scenarios where understanding why a region is deemed salient is as crucial as the prediction itself." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The development of explainable models that mimic hu" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.285, + 0.871, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "Figure 1. Our proposed XSal model provides step-by-step natural language explanations, semantic prototype visualization, as well as saliency weights to explain its prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "man attention remains an underexplored research frontier. While some models have attempted to introduce explainability, they often rely on massive predefined vocabularies of semantics [10, 11] or simple descriptions of eye fixations without deep reasoning [15]. These methods, while offering some insights, lack true interpretability due to the lack of reasoning and prioritization within the context. This limitation is rooted in the inherent complexity of human attention. Human attention is a multifaceted phenomenon, influenced by a confluence of factors including visual cues, semantic understanding, individual preferences, and even cultural context. To generate meaningful explanations that unravel this interplay, a deep understanding of human cognition and perception is essential. Moreover, creating a model that can articulate its reasoning in a clear and concise manner poses a significant technical challenge." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.317 + ], + "angle": 0, + "content": "To bridge the gap between saliency prediction and interpretability, we develop a saliency model that not only accurately predicts salient regions but also provides clear, human-understandable explanations. As shown in Fig. 1, our goal is to factorize saliency prediction with a weighted combination of spatial feature maps, namely semantic prototypes. Each semantic prototype represents the grounding of a semantic proposal (e.g., frisbee, fence, and sheep) relevant to the visual task, and its relevance is explained with a natural language description. In this example, \"frisbee\" and \"sheep\" have higher weights in saliency prediction, because \"to find the frisbee's location\" and \"to check if there is a sheep to the right of the fence\" are essential reasoning steps to answer the question \"Is there a sheep to the right of the fence that is to the right of the frisbee?\"" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.319, + 0.468, + 0.559 + ], + "angle": 0, + "content": "To achieve this, we propose an explainable saliency (XSal) model that advances the state-of-the-art in saliency prediction and explanation. First, it integrates a vision-language model (VLM) to perform explicit reasoning for saliency prediction, generating step-by-step explanations and delivering natural language interpretations that clarify why specific regions are salient. Second, we introduce a contextual prioritization mechanism designed to handle the challenges of selecting truly important ones from a potentially large set of semantic explanations. This mechanism allows the model to focus on key elements, achieving interpretability without overcomplicating the representation. By integrating explicit reasoning with targeted selection, we create a cohesive framework that not only identifies and explains salient regions but also adapts dynamically to scene complexity." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.56, + 0.468, + 0.59 + ], + "angle": 0, + "content": "In summary, this paper makes three key contributions to the field of saliency prediction and explainable AI:" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.592, + 0.468, + 0.667 + ], + "angle": 0, + "content": "1. This paper presents the first saliency model to incorporate a vision-language model for explicit reasoning. This integration allows the model to not only identify salient regions but also articulate its reasoning process, providing clear and human-understandable explanations." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.667, + 0.468, + 0.741 + ], + "angle": 0, + "content": "2. We introduce a novel contextual prioritization mechanism that prioritizes the most relevant information within the context of saliency prediction and explanation. This results in a more interpretable model by focusing on key relevant semantics and simplifying the representation." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.742, + 0.468, + 0.833 + ], + "angle": 0, + "content": "3. We conduct extensive evaluations of our model, demonstrating its effectiveness in generating high-quality saliency maps while providing coherent, contextually relevant explanations. This comprehensive assessment highlights the model's ability to effectively bridge the gap between prediction accuracy and interpretability." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.592, + 0.468, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.846, + 0.217, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.872, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Saliency Prediction. Saliency prediction models have traditionally focused on bottom-up approaches, relying on" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.89, + 0.469 + ], + "angle": 0, + "content": "low-level visual features in free-viewing conditions [5, 25, 27, 29]. However, with the rapid progress of deep learning methods, deep saliency models have emerged, achieving promising accuracy in predicting both task-free bottom-up saliency [26, 34, 52] and task-driven top-down attention [1, 30, 48]. A key strength of deep saliency models is their ability to extract high-level semantic features. Numerous deep neural network-based methods have boosted saliency prediction performance by implicitly encoding semantics using different approaches [46, 52]. However, such models often lack transparency and trustworthiness due to explicitly learned representations. While there has been growing interest in understanding saliency model mechanisms, existing approaches fall short of providing comprehensive explanations. For instance, Chen et al. [10, 11] analyze model behaviors using semantic prototypes and soft attention weights, but these approaches offer limited insight into the model's decision-making process, particularly when dealing with thousands of prototypes. Similarly, GazeXplain [15] generates plain natural language descriptions of fixation points, lacking a reasoning or prioritization mechanism to identify crucial factors driving its predictions. These limitations highlight the need for a more robust methodology that can uncover the underlying rationale behind saliency model behavior." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.473, + 0.89, + 0.592 + ], + "angle": 0, + "content": "Our proposed method addresses these limitations by introducing a novel framework that combines explicit vision-language reasoning with a hard attention mechanism. This ensures that only the most critical elements are selected for saliency prediction and explanation, which not only improves the accuracy of saliency predictions but also provides a more transparent and insightful understanding of the model's behavior." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.597, + 0.89, + 0.852 + ], + "angle": 0, + "content": "Vision Language Models. Our work builds upon the growing field of explainable AI (XAI) within vision-language models (VLMs). Recent advancements in VLMs [13, 14, 17, 22, 31, 35, 47, 50, 51] have revolutionized multimodal understanding, enabling these models to seamlessly integrate visual and textual information. This progress has opened new avenues for developing XAI systems capable of providing interpretable insights into complex decision-making processes. For instance, researchers have explored using VLMs to generate natural language explanations for tasks such as image classification [23, 38], abusive language detection [19], and autonomous driving [7], allowing users to understand the rationale behind model predictions. Additionally, VLMs have been employed to enhance the interpretability of visual scanpath models, describing image features driving human attention to a sequence of fixation points [15]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.857, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Different from these studies, by integrating textual context and introducing reasoning and prioritization mechanisms, we aim to provide deeper insights into the factors" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9602" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.086, + 0.873, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Figure 2. The proposed XSal architecture begins with a vision-language model generating semantic proposals from the input image, and uses a segmentation model to generate their corresponding semantic prototypes. A contextual prioritization module then selects the top-K relevant semantic prototypes, which are processed by a final saliency head to predict a focused saliency map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.471, + 0.453 + ], + "angle": 0, + "content": "driving saliency predictions, improving the transparency and interpretability of saliency models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.467, + 0.169, + 0.482 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.471, + 0.674 + ], + "angle": 0, + "content": "The human visual attention process is a captivating blend of explicit reasoning and selective focus, guiding us to prioritize relevant elements while navigating complex scenes. While existing saliency models have made strides, they often fall short of capturing this nuanced behavior, relying instead on dense representations or lacking a mechanism for explaining their choices. This incongruence between model behavior and human attention motivates us to introduce a novel approach that leverages VLMs to perform explicit reasoning and prioritizing the reasoning outcomes based on contextual relevance for saliency prediction and explanation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.259, + 0.701 + ], + "angle": 0, + "content": "3.1. Explicit Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.47, + 0.8 + ], + "angle": 0, + "content": "At the heart of our approach lies a vision-language model (VLM), specifically MiniCPM-V 2.6 [53], which enables the model to identify and explain salient elements across task-driven or task-free contexts. Specifically, the VLM produces a set of semantic proposals \\(\\mathbf{C}\\) and associated explanations \\(\\mathbf{R}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.816, + 0.47, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C}, \\mathbf {R} = \\operatorname {V L M} (\\mathbf {I}, \\mathbf {Q}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{I}\\) is the input image, and \\(\\mathbf{Q}\\) is the question guiding attention. \\(\\mathbf{C} = \\{P_1,P_2,\\dots ,P_n\\}\\) represents semantic proposals, each focusing on specific objects within the scene, while \\(\\mathbf{R}\\) provides rationale for why each element was" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.421, + 0.543, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.578, + 0.431, + 0.865, + 0.453 + ], + "angle": 0, + "content": "Can you explain step by step how you analyzed the image to answer the question?" + }, + { + "type": "text", + "bbox": [ + 0.569, + 0.453, + 0.87, + 0.464 + ], + "angle": 0, + "content": "Question: What is the drink in the cup made of plastic called?" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.465, + 0.616, + 0.475 + ], + "angle": 0, + "content": "Image:" + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.466, + 0.868, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.595, + 0.685, + 0.606 + ], + "angle": 0, + "content": "1. I identified the drink in the cup." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.607, + 0.806, + 0.617 + ], + "angle": 0, + "content": "2. I examined the cup's appearance to determine its material." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.617, + 0.805, + 0.628 + ], + "angle": 0, + "content": "3. I confirmed the cup was plastic based on its transparency." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.629, + 0.812, + 0.64 + ], + "angle": 0, + "content": "4. Finally, I considered common drinks served in plastic cups." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.595, + 0.812, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.851, + 0.587, + 0.89, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.653, + 0.542, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.577, + 0.662, + 0.837, + 0.674 + ], + "angle": 0, + "content": "In each step, which objects did you focus on, and why?" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.698, + 0.812, + 0.719 + ], + "angle": 0, + "content": "1. I focused on the drink in the cup, because it was necessary to identify the main object in question." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.72, + 0.815, + 0.741 + ], + "angle": 0, + "content": "2. I focused on the cup to determine its material by examining its appearance and texture." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.741, + 0.814, + 0.763 + ], + "angle": 0, + "content": "3. I focused on the plastic to confirm the material based on its transparency and common usage." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.763, + 0.815, + 0.784 + ], + "angle": 0, + "content": "4. I focused on the drink in the plastic cup to the specific type of drinks served in such cups." + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.698, + 0.815, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.85, + 0.689, + 0.89, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.806, + 0.892, + 0.834 + ], + "angle": 0, + "content": "Figure 3. The two-stage prompting method to obtain semantic proposals and explanations from the VLM." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "selected, ensuring interpretability through alignment with human attention patterns. This integration supports a rea" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.181 + ], + "angle": 0, + "content": "soning process that mirrors human problem-solving, where decisions are informed by both visual cues and a structured understanding of why certain elements are relevant. To obtain the semantic proposals and explanations from the VLM, as shown in Fig. 3, we employ a two-stage prompting method:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.365 + ], + "angle": 0, + "content": "In the initial planning stage, we ask the VLM to describe \"how you analyzed the image to answer the question\" (for top-down task-driven cases) or \"important objects in the image\" (for bottom-up free-viewing cases). The model responds with a coherent sequence of actions of reasoning steps to follow when answering the question. For example, to answer the question \"What is the drink in the cup made of plastic called?\", the reasoning steps include identifying the drink in the cup, examining the cup's appearance to determine its material, confirming the cup was plastic based on its transparency, and considering common drinks served in plastic cups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.487 + ], + "angle": 0, + "content": "Following this planning, the VLM is asked to explain what specific objects it focused on in each step, and explain the reason. It then proceeds to generate the set of semantic proposals \\(\\mathbf{C}\\) (such as drink, cup, and plastic) and their corresponding explanations \\(\\mathbf{R}\\). Finally, to ensure precision in the final output, the VLM is also prompted to verify that the identified semantic proposals are concrete and relevant, addressing potential limitations in model consistency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.469, + 0.532 + ], + "angle": 0, + "content": "This layered approach replicates human selective attention and explanatory reasoning, ensuring the model's focus and reasoning align closely with human attention patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.471, + 0.835 + ], + "angle": 0, + "content": "After identifying semantic proposals through the VLM, we utilize an open-vocabulary segmentation model, ClipSeg [37], to extract the visual grounding of these proposals on the image. ClipSeg's ability to handle open-vocabulary concepts is crucial as the language model may describe elements that don't perfectly align with traditional segmentation categories. This step bridges the semantic reasoning process with visual representations, allowing for visual validation of the model's reasoning. While ClipSeg effectively outlines entire object contours, human attention typically focuses on smaller, more localized regions. Therefore, instead of using the segmentation results directly for saliency prediction, we use them as binary masks to filter the image features, and apply a convolutional layer \\(\\mathcal{F}\\) to obtain fine-grained localization. This results in a sequence of focused spatial maps, namely semantic prototypes, where each prototype \\(S_{i}\\) aligns with a semantic proposal \\(P_{i}\\), localizing its representations in the image. This process pinpoints focused regions that naturally attract human attention, associating them with natural language explanations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.4, + 0.862 + ], + "angle": 0, + "content": "3.2. Contextual Prioritization Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Our approach utilizes a contextual prioritization mechanism to extract the most relevant semantic prototypes for saliency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "prediction. Different from existing saliency models that often integrate hundreds to thousands of spatial feature maps with linear weights, humans naturally focus on a limited number of features, while other elements, though present, are not prioritized to the same extent. Our proposed contextual prioritization is a highly selective approach, which only allows attention to be directed by the most relevant features. It introduces a dynamic approach that adapts to varying numbers of semantic prototypes. This flexibility allows the model to handle scenes of different complexities effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Our contextual prioritization mechanism adjusts importance by applying global average pooling on the extracted semantic prototypes, followed by a linear layer to generate initial weights. These weights are then sorted, and only the top \\( K \\) and their corresponding features are selected. A softmax operation is applied to normalize the selected weights, yielding a refined representation:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.378, + 0.892, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} = \\operatorname {L i n e a r} (\\operatorname {G A P} (\\mathbf {S})), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.415, + 0.892, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {\\text {t o p K}} = \\operatorname {S o f t m a x} (\\operatorname {T o p K} (\\mathbf {w})), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathrm {O} = \\sum_ {i = 1} ^ {K} \\mathbf {w} _ {\\text {t o p K}, i} \\cdot \\mathbf {S} _ {\\text {t o p K}, i} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.765 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}\\) represents a collection of semantic prototypes, \\(\\{S_1,S_2,\\dots ,S_n\\}\\). The global average pooling (GAP) and subsequent linear transformation extract initial weights \\(\\mathbf{w}\\) for each prototype, allowing the model to assign varying importance to each semantic element. By selecting and normalizing the top \\(K\\) weights, we obtain \\(\\mathbf{w}_{\\mathrm{topK}}\\) and the corresponding \\(\\mathbf{S}_{\\mathrm{topK}}\\). Here, \\(\\mathbf{S}_{\\mathrm{topK}}\\) corresponds to the semantic prototypes associated with the top \\(K\\) weights in \\(\\mathbf{w}\\), determined by sorting. This ensures that the model prioritizes the most significant semantic elements, aligning its focus with the most contextually important regions, and discarding less relevant elements. Finally, the aggregated saliency feature O, the weighted combination of the top \\(K\\) semantic prototypes, is processed by a saliency head to produce the final saliency prediction. This saliency head consists of only a single linear layer (with bias set to 0) that simply compresses the channel dimension, acting as a weighted sum without introducing any complex black-box behavior." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To ensure adequate feature optimization and prevent limited feature selection from hindering convergence, we introduce a Progressive Top-K Reduction strategy. Rather than starting with a fixed, small set of key features, this strategy initially preserves a larger number of semantic elements, enabling comprehensive feature optimization in early training. As training progresses, the \\( K \\) value is gradually reduced, guiding the model to focus on the most significant features. The progression of \\( K \\) across epochs is defined as" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9604" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.162, + 0.102, + 0.469, + 0.133 + ], + "angle": 0, + "content": "\\[\nK _ {t} = K _ {\\text {i n i t}} - \\frac {\\left(K _ {\\text {i n i t}} - K _ {\\text {f i n a l}}\\right) \\times t}{T - 1} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.473, + 0.273 + ], + "angle": 0, + "content": "where \\( K_{t} \\) is the \\( K \\) at epoch \\( t \\), \\( K_{\\mathrm{init}} \\) and \\( K_{\\mathrm{final}} \\) are the initial and final values of \\( K \\), and \\( T \\) is the total number of epochs. This progressive reduction allows the model to gradually shift from broad feature optimization to a selective focus on the most critical elements. With this strategy, the contextual prioritization mimics the adaptive prioritization seen in human attention, flexibly weighing elements based on their relevance and adapting to scenes with different numbers of semantic prototypes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.286, + 0.21, + 0.303 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.448 + ], + "angle": 0, + "content": "This section presents a comprehensive evaluation of our proposed model, examining its performance and explainability across both task-driven and task-free eye-tracking datasets. We conduct an ablation study to further investigate the contributions of individual components, and present qualitative examples to illustrate the model's capabilities in generating interpretable saliency maps and explanations. Additional experimental results and analyses are provided in the supplementary material for further reference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.267, + 0.472 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.675 + ], + "angle": 0, + "content": "Datasets. Our experiments are conducted on two eye-tracking datasets: The AiR dataset [9], derived from the GQA dataset, comprises images and questions paired with eye-tracking data from 20 participants answering those questions. This setup provides a realistic scenario for evaluating saliency models, as it captures human attention in the context of a specific task (question answering). The OSIE [49] dataset is a free-viewing dataset featuring multiple salient objects competing for attention in the same image context, suitable for developing and evaluating saliency models that consider object- and semantic-level information. In our experiments, we split both datasets into training, validation, and test sets following [12, 16]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Compared Models. We compare our model's performance against several state-of-the-art saliency prediction models including SAM [18], SALICON [26], DINet [52], as well as explainable models such as Chen et al. [10] and GazeXplain [15]. Note that for GazeXplain [15] we aggregate fixation points from its predicted scanpaths into saliency maps for comparison. By comparing our model with these diverse methods, we aim to demonstrate its performance in various aspects, including general saliency prediction and explainability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. To comprehensively evaluate our model, we utilize a combination of metrics that assess saliency prediction performance and explainability. 1) Saliency Prediction Metrics: We employ standard metrics to evaluate the alignment between our model's predicted" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.485 + ], + "angle": 0, + "content": "saliency maps and human attention patterns. These metrics are Area Under the ROC Curve (AUC) [3], Normalized Scanpath Saliency (NSS) [41], and Correlation Coefficient (CC) [39], which provide quantitative insights into the accuracy and relevance of the generated saliency maps. 2) Explainability Metrics: We evaluate the interpretability of saliency predictions adapting three XAI metrics: AUC [2] assesses model stability by calculating the area under the curve that represents changes in model performance as different proportions of semantic prototypes are progressively perturbed. Note that this AUC differs from the AUC used in saliency evaluation: here, it measures the model's sensitivity to perturbations, while saliency AUC measures alignment with human attention data. To facilitate distinction, we refer to this interpretability metric as AUC-E throughout this paper. Area Over the Perturbation Curve (AOPC) [6, 40] quantifies the sensitivity of the model to perturbations by measuring the change in NSS as semantic prototypes are sequentially altered. Log-odds Score (LOdds) [43, 44] evaluates the change in NSS as semantic prototypes are perturbed, focusing on the degree of reduction in NSS in response to prototype alterations. These metrics provide a comprehensive framework for assessing the quality of our model's explanations, enabling us to gauge not only where the model focuses, but also the robustness of its interpretative reasoning under perturbations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Loss Function. To align the model's saliency predictions effectively with human attention patterns, we employ a combination of loss functions, including NSS, CC, and Kullback-Leibler Divergence (KLD) [33]. Our implementation follows [10]. These loss functions provide complementary signals that guide the model to capture nuanced attention details, helping to achieve balanced performance across various saliency prediction metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Our model employs a dilated ResNet-50 [24] backbone, following the setting used in DINet [52]. This architecture incorporates dilated convolutions to expand the receptive field without additional downsampling, which enhances the model's capacity to capture high-resolution spatial details essential for accurate saliency prediction. The model is optimized using the Adam optimizer [32] with an initial learning rate of 4e-4 and a weight decay of 1e-7. To ensure stable convergence, the learning rate is adaptively reduced by a decay factor of 0.1 every two epochs, while gradient clipping is applied with a threshold of 0.1 to prevent gradient explosion. Training is conducted over 10 epochs with a batch size of 10. To ensure stable training and prevent convergence issues, we employ the proposed Progressive Top-K Reduction strategy, which begins with a broader feature selection and gradually narrows the focus to the most significant features. Specifically, we set the initial \\(K\\) value to 20, which progressively reduces to a final value of 3 over 3 epochs, enabling adaptive and selec" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9605" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.437, + 0.263 + ], + "angle": 0, + "content": "
AiROSIE
NSSCCNSSCC
SAM [18]1.650.622.700.65
SALICON [28]1.700.632.750.63
DINet [52]1.770.632.880.63
Chen et al. [10]1.760.632.910.64
GazeXplain [15]1.850.662.530.74
XSal1.940.702.960.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.274, + 0.468, + 0.303 + ], + "angle": 0, + "content": "Table 1. Comparative results of saliency prediction across the AiR and OSIE datasets." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.331, + 0.248, + 0.346 + ], + "angle": 0, + "content": "tive feature prioritization." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.358, + 0.321, + 0.373 + ], + "angle": 0, + "content": "4.2. Saliency Prediction Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.382, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Tab. 1 presents a quantitative comparison of saliency prediction performance across the AiR (task-driven) and OSIE (task-free) saliency datasets for our proposed model and several state-of-the-art methods. As shown in the table, our proposed model consistently outperforms other methods on the AiR dataset, achieving the highest NSS score of 1.94 and the highest CC score of 0.70. This suggests that our model effectively captures top-down human attention patterns in task-driven scenarios. Our model also achieves the highest NSS score of 2.96 on the OSIE dataset. Though our CC score (0.72) is slightly lower than GazeXplain (0.74), it still demonstrates strong performance in capturing the overall distribution of human attention in task-free image-viewing. The consistently high NSS scores across both datasets indicates that our model accurately predicts the locations where humans are mostly likely to fixate their gaze, demonstrating a strong accuracy of predicting the most salient regions. However, our model might be slightly less effective at capturing the overall distribution of human attention in images, because image regions with low saliency (low agreements across human observers) are less explainable and likely to be discarded by the contextual prioritization mechanism. Overall, the quantitative results demonstrate that our proposed method significantly improves saliency prediction performance compared to existing approaches. The incorporation of explicit reasoning and selective attention mechanisms contributes to more accurate and robust models that effectively capture the complexity of human visual attention." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.336, + 0.848 + ], + "angle": 0, + "content": "4.3. Saliency Explanation Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 presents the quantitative results based on explainability metrics, comparing our model's performance with Chen et al. [10] on both the AiR and OSIE datasets. Notably," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "our model demonstrates significantly lower AUC-E, LOdds, and higher AOPC scores compared to Chen et al. Specifically, on the AiR dataset, our model significantly outperforms Chen et al. with a lower AUC-E (0.713 vs. 0.899) and a higher AOPC (0.748 vs. 0.241). The substantial improvement in LOdds (-5.157 vs. -0.143) also reflects the model's ability to focus on the most relevant elements, minimizing irrelevant attention. On the OSIE dataset, a similar trend is observed. Our model achieves a better AUC-E (0.685 vs. 0.810) and significantly higher AOPC (1.115 vs. 0.586) compared to Chen et al. Additionally, the considerable decrease in LOdds (-4.169 vs. -0.250) further supports our model's capability to highlight key semantic concepts more effectively, thereby improving interpretability in a task-free viewing context." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.591 + ], + "angle": 0, + "content": "These results indicate that our model is more sensitive to changes in semantic prototypes, highlighting the stronger reliance on these prototypes for generating accurate saliency predictions. The lower AUC-E and LOdds scores suggest that perturbing prototypes leads to a more substantial decrease in performance for our model, highlighting the critical role of semantic understanding in our approach. Conversely, the higher AOPC score signifies a greater sensitivity to changes in prototypes, revealing a stronger correlation between the model's performance and the identified semantic elements. This reinforces the notion that our model's interpretability is closely tied to the semantic prototypes it extracts, providing a more transparent and interpretable reasoning process compared to methods that rely on complex aggregation of numerous features. Overall, these results emphasize the importance of explicit reasoning and selective attention in achieving a higher degree of interpretability and understanding in saliency prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.653, + 0.62 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Our ablation study, conducted on both the AiR and OSIE datasets, reveals the significant contributions of each component to our model's performance and explainability. As shown in Tab. 3, a simple baseline model, using only a dilated ResNet-50 backbone [24] with a saliency head, achieves reasonable performance in both settings, yielding an NSS of 1.777 on AiR and 2.835 on OSIE. The baseline model lacks explainability, with a LOdds of -0.131 and -0.246 on these datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "By incorporating the VLM for reasoning, our model links saliency with task-relevant semantic elements and significantly improves both accuracy and interpretability across both datasets. For instance, on the AiR dataset, the NSS increases to 1.917 and the CC score improves to 0.703, while the AUC remains relatively stable at 0.858. This highlights the crucial role of task-oriented alignment and semantic understanding in generating clear and contextual explanations for saliency predictions. This is further evidenced" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9606" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.204, + 0.089, + 0.768, + 0.173 + ], + "angle": 0, + "content": "
AiROSIE
AUC-E↓AOPC↑LOdds↓AUC-E↓AOPC↑LOdds↓
Chen et al. [52]0.8990.241-0.1430.8100.586-0.250
XSal0.7130.748-5.1570.6851.115-4.169
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.183, + 0.747, + 0.197 + ], + "angle": 0, + "content": "Table 2. Comparative results of explainability metrics across the AiR and OSIE datasets." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.211, + 0.839, + 0.34 + ], + "angle": 0, + "content": "
DatasetBaselineReasoningContPriNSSCCAUCAUC-E↓AOPC ↑LOdds ↓
AiR✓1.7770.6400.8490.9100.225-0.131
✓✓1.9170.7030.8580.7000.749-5.077
✓✓✓1.9420.7010.8580.7130.748-5.157
OSIE✓2.8350.6810.8850.8230.567-0.246
✓✓2.9040.7170.8970.7650.876-1.738
✓✓✓2.9650.7260.8950.6851.115-4.169
" + }, + { + "type": "table_caption", + "bbox": [ + 0.179, + 0.35, + 0.792, + 0.364 + ], + "angle": 0, + "content": "Table 3. Ablation study on the impacts of Baseline, Reasoning, and Contextual Prioritization (ContPri)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.452 + ], + "angle": 0, + "content": "by the significant improvements in interpretability metrics – the AUC-E drops to 0.700 and the AOPC increases to 0.749, suggesting our model's ability to provide more coherent explanations for task-driven saliency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.695 + ], + "angle": 0, + "content": "Further, the contextual prioritization mechanism exhibits a more pronounced impact, especially in the OSIE task-free setting. For example, on OSIE, the NSS increases to 2.965, the CC improves to 0.726, and the AUC remains relatively stable at 0.895. This suggests that prioritizing the most salient elements plays a crucial role in handling complex scenes with multiple potential salient objects, ultimately leading to more concise and clear attention maps. This is further evidenced by the significant improvements in interpretability metrics – the AUC-E decreases to 0.685, the AOPC increases to 1.115, and the LOdds decreases to -4.169. This greater impact on the OSIE dataset could be attributed to the presence of a large number of potential semantic prototypes in task-free settings, making selective focus particularly valuable for distilling relevant information and improving interpretability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.773 + ], + "angle": 0, + "content": "These findings demonstrate that the combination of explicit reasoning and contextual prioritization mechanisms, implemented in our model, is essential for achieving superior performance and explainability in both task-driven and task-free saliency prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.27, + 0.803 + ], + "angle": 0, + "content": "4.5. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Qualitative analysis of our model reveals its ability to effectively integrate task-driven guidance and contextual prioritization to generate accurate saliency predictions and explanations. As illustrated in Fig. 4, the model aligns its attention with the goals of each question, selectively focusing on essential objects while ignoring irrelevant elements. For ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.589 + ], + "angle": 0, + "content": "ample, in questions about the position of a skateboard and a red cone (row 1), the model prioritizes these objects over traditionally salient regions like the skateboarder, demonstrating the advantage of explicit reasoning. The model also adapts well to abstract concepts, such as interpreting \"modern\" style in a question about a faucet (row 2). Finally, the model excels in context prioritization by isolating key elements in complex scenes, highlighting the person, bag, and cars in a question about cars (row 3) while filtering out other visual elements. These qualitative findings demonstrate the model's ability to produce explanations that are closely aligned with the model's internal reasoning, ensuring outputs that are both interpretable and reliable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Beyond its interpretability, our model shows significant performance advantages in accuracy and task relevance. In Fig. 5, our model achieves the highest accuracy among state-of-the-art methods, aligning closely with ground truth. For example, in the question \"What is the drink in the cup made of plastic called?\", it focuses on the plastic cup, unlike other models that are distracted by irrelevant food items (row 1), showcasing the strength of our explicit reasoning approach. Additionally, our model demonstrates strong context prioritization, effectively concentrating on relevant elements even in complex scenes with multiple distractions. For instance, in the question \"Is the fire hydrant to the right or to the left of the garbage can?\", our model isolates the hydrant and garbage can, while other models diffuse attention across the scene (row 2), highlighting its ability to prioritize contextually important elements. Furthermore, in dynamic scenes, our model reliably focuses on salient objects. In the question \"Are both the helmet and the bat the same color?\", it highlights the helmet and bat specifically, unlike other models that diffuse attention (row 3), demonstrating" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9607" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.094, + 0.176, + 0.103 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.105, + 0.2, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.094, + 0.277, + 0.103 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.105, + 0.285, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.094, + 0.351, + 0.103 + ], + "angle": 0, + "content": "Prediction" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.105, + 0.367, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.094, + 0.514, + 0.103 + ], + "angle": 0, + "content": "Prototype 1" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.105, + 0.527, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.153, + 0.522, + 0.163 + ], + "angle": 0, + "content": "\\(\\mathbf{B}\\mathbf{a}\\mathbf{g}(w_1 = 0.378)\\)" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.164, + 0.522, + 0.212 + ], + "angle": 0, + "content": "The question specifies a bag, so it is relevant to locate its position in the image and use it as a reference point for determining the direction of 'to the right'." + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.094, + 0.665, + 0.103 + ], + "angle": 0, + "content": "Prototype 2" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.105, + 0.677, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.153, + 0.672, + 0.163 + ], + "angle": 0, + "content": "Person \\((w_{2} = 0.330)\\)" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.164, + 0.674, + 0.203 + ], + "angle": 0, + "content": "To identify the reference point for 'to the right' in relation to the bag and determine where cars are located relative to this person." + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.094, + 0.831, + 0.103 + ], + "angle": 0, + "content": "Prototype 3" + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.105, + 0.843, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.153, + 0.84, + 0.163 + ], + "angle": 0, + "content": "Cars \\((w_{3} = 0.291)\\)" + }, + { + "type": "text", + "bbox": [ + 0.706, + 0.164, + 0.84, + 0.203 + ], + "angle": 0, + "content": "Cars are mentioned in the question and need to be identified in their location relative to the person and bag." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.154, + 0.194, + 0.212 + ], + "angle": 0, + "content": "Q: Are there any cars to the right of the person that is to the right of the bag?" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.214, + 0.2, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.214, + 0.284, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.214, + 0.367, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.262, + 0.197, + 0.31 + ], + "angle": 0, + "content": "Q: Does the faucet to the left of the vase look modern and silver?" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.214, + 0.526, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.262, + 0.522, + 0.272 + ], + "angle": 0, + "content": "Vase \\((w_{1} = 0.357)\\)" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.272, + 0.522, + 0.302 + ], + "angle": 0, + "content": "The vase is mentioned in the question and serves as a reference point for the faucet's location." + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.214, + 0.677, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.262, + 0.672, + 0.271 + ], + "angle": 0, + "content": "Faucets \\((w_{2} = 0.319)\\)" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.272, + 0.673, + 0.311 + ], + "angle": 0, + "content": "The question specifically asks about the faucet's appearance, so it was a key element in answering." + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.214, + 0.844, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.262, + 0.84, + 0.271 + ], + "angle": 0, + "content": "Modern design \\((w_{3} = 0.329)\\)" + }, + { + "type": "text", + "bbox": [ + 0.695, + 0.272, + 0.84, + 0.3 + ], + "angle": 0, + "content": "The question refers to the faucets style, which can be described as modern." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.313, + 0.2, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.313, + 0.284, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.313, + 0.367, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.313, + 0.526, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.361, + 0.521, + 0.37 + ], + "angle": 0, + "content": "Skateboard \\((w_{1} = 0.371)\\)" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.371, + 0.522, + 0.399 + ], + "angle": 0, + "content": "The skateboard is the object being asked about in terms of its position relative to the red cone." + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.313, + 0.677, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.361, + 0.673, + 0.411 + ], + "angle": 0, + "content": "Red cone \\((w_{2} = 0.344)\\) red cone is the reference point for determining the position of the skateboard relative to the question." + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.313, + 0.843, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.361, + 0.84, + 0.37 + ], + "angle": 0, + "content": "Skateboarder \\((w_{3} = 0.283)\\)" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.371, + 0.84, + 0.41 + ], + "angle": 0, + "content": "The skateboarder position and orientation are crucial in understanding where the skateboard is in relation to the red cone." + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.42, + 0.887, + 0.435 + ], + "angle": 0, + "content": "Figure 4. Qualitative examples from our methods. Each semantic prototype is shown along with their saliency weights and explanations." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.468, + 0.2, + 0.508 + ], + "angle": 0, + "content": "Q: What is the drink in the cup made of plastic called?" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.527, + 0.216, + 0.57 + ], + "angle": 0, + "content": "Q: Is the fire hydrant to the right or to the left of the garbage can?" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.587, + 0.212, + 0.618 + ], + "angle": 0, + "content": "Q: Are both the helmet and the bat the same color?" + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.454, + 0.294, + 0.464 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.466, + 0.328, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.525, + 0.328, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.585, + 0.328, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.454, + 0.42, + 0.464 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.466, + 0.432, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.525, + 0.432, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.585, + 0.432, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.454, + 0.5, + 0.464 + ], + "angle": 0, + "content": "XSal" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.466, + 0.536, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.525, + 0.536, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.585, + 0.537, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.454, + 0.608, + 0.464 + ], + "angle": 0, + "content": "DINet" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.466, + 0.641, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.525, + 0.641, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.584, + 0.641, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.454, + 0.724, + 0.464 + ], + "angle": 0, + "content": "SALICON" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.466, + 0.745, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.525, + 0.745, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.584, + 0.745, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.454, + 0.815, + 0.464 + ], + "angle": 0, + "content": "SAM" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.466, + 0.851, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.525, + 0.851, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.584, + 0.851, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.656, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of saliency predictions across models (XSal, DINet, SALICON, SAM). Each row shows an input question and image, with the corresponding ground truth attention map and model outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.469, + 0.757 + ], + "angle": 0, + "content": "its adaptability to diverse scenarios. These examples highlight our model's superior performance in generating accurate and contextually relevant saliency predictions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.196, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "This paper has presented a novel approach for saliency prediction that combines explicit reasoning and contextual prioritization, generating both accurate and explainable saliency predictions. We have demonstrated that our method, powered by a VLM, outperforms existing models in terms of accuracy and explainability. This is achieved by leveraging an explicit reasoning process that allows the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.894, + 0.832 + ], + "angle": 0, + "content": "model to provide clear explanations for its predictions and a contextual prioritization mechanism that focuses on a few truly important elements, mimicking human attention behavior. Our approach not only achieves state-of-the-art saliency prediction accuracy but also offers valuable insights into the decision-making processes of saliency models. This advancement contributes to the development of more trustworthy and reliable saliency models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.66, + 0.863 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by NSF Grant 2143197 and Fasikl Incorporated." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9608" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.27 + ], + "angle": 0, + "content": "[2] Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[4] Patrick Le Callet and Ernst Niebur. Visual attention and applications in multimedia technologies. Proceedings of the Institution of Electrical Engineers, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[5] Moran Cerf, E Paxon Frady, and Christof Koch. Faces and text attract gaze independent of the task: Experimental data and computer model. Journal of vision, 9(12):10-10, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.497 + ], + "angle": 0, + "content": "[6] Hanjie Chen, Guangtao Zheng, and Yangfeng Ji. Generating hierarchical explanations on text classification via feature interaction detection. In Proceedings of the 2020 Conference on Association for Computational Linguistics (ACL), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.584 + ], + "angle": 0, + "content": "[7] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.47, + 0.653 + ], + "angle": 0, + "content": "[8] Nuo Chen, Jin Xie, Jing Nie, Jiale Cao, Zhuang Shao, and Yanwei Pang. Attentive alignment network for multispectral pedestrian detection. In Proceedings of the 31st ACM international conference on multimedia, pages 3787-3795, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.657, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[9] Shi Chen, Ming Jiang, Jinhui Yang, and Qi Zhao. AiR: Attention with reasoning capability. In European Conference on Computer Vision (ECCV), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.701, + 0.47, + 0.756 + ], + "angle": 0, + "content": "[10] Shi Chen, Ming Jiang, and Qi Zhao. What do deep saliency models learn about visual attention? In Advances in Neural Information Processing Systems, pages 9543-9555. Curran Associates, Inc., 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.758, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[11] Shi Chen, Ming Jiang, and Qi Zhao. Deep learning to interpret autism spectrum disorder behind the camera. IEEE Transactions on Cognitive and Developmental Systems, pages 1-12, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Xianyu Chen, Ming Jiang, and Qi Zhao. Predicting human scanpaths in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10876-10885, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Xianyu Chen, Ming Jiang, and Qi Zhao. Leveraging human attention in novel object captioning. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "International Joint Conference on Artificial Intelligence (IJ-CAI), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[14] Xianyu Chen, Ming Jiang, and Qi Zhao. Self-distillation for few-shot image captioning. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[15] Xianyu Chen, Ming Jiang, and Qi Zhao. Gazexplain: Learning to predict natural language explanations of visual scanpaths. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[16] Xianyu Chen, Ming Jiang, and Qi Zhao. Beyond average: Individualized visual scanpath prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[17] Xianyu Chen, Jinhui Yang, Shi Chen, Louis Wang, Ming Jiang, and Qi Zhao. Every problem, every step, all in focus: Learning to solve real-world problems with integrated attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[18] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting Human Eye Fixations via an LSTM-based Saliency Attentive Model. TIP, 27(10):5142-5154, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.894, + 0.495 + ], + "angle": 0, + "content": "[19] Chiara Di Bonaventura, Lucia Siciliani, Pierpaolo Basile, Albert Merono Penuela, and Barbara McGillivray. Is explanation all you need? an expert survey on llm-generated explanations for abusive language detection. In Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[20] Ning Ding, Ce Zhang, and Azim Eskandarian. Saliendet: A saliency-based feature enhancement algorithm for object detection for autonomous driving. IEEE Transactions on Intelligent Vehicles, 9(1):2624-2635, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.894, + 0.621 + ], + "angle": 0, + "content": "[21] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Zhaohui Che, Yi Fang, Xiaokang Yang, Jesús Gutierrez, and Patrick Le Callet. A dataset of eye movements for the children with autism spectrum disorder. In ACM Multimedia Systems Conference (MMSys), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.894, + 0.676 + ], + "angle": 0, + "content": "[22] Yang Feng, Lin Ma, Wei Liu, and Jiebo Luo. Unsupervised image captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.678, + 0.894, + 0.719 + ], + "angle": 0, + "content": "[23] Songhao Han, Le Zhuo, Yue Liao, and Si Liu. Llms as visual explainers: Advancing image classification with evolving visual descriptions. arXiv preprint arXiv:2311.11904, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[25] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[26] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In ICCV, pages 262-270, 2015. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[27] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9609" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "on Pattern Analysis and Machine Intelligence, 20(11):1254-1259, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[28] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Silicon: Saliency in context. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[29] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In 2009 IEEE 12th international conference on computer vision, pages 2106-2113. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[30] Christopher Kanan, Mathew H Tong, Lingyun Zhang, and Garrison W Cottrell. Sun: Top-down saliency using natural statistics. Visual cognition, 17(6-7):979-1003, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[31] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Image captioning with very scarce supervised data: Adversarial semi-supervised learning approach. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[33] Solomon Kullback. Information Theory and Statistics. Wiley, 1959. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.47, + 0.447 + ], + "angle": 0, + "content": "[34] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. In ICLR workshop, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[35] Iro Laina, Christian Rupprecht, and Nassir Navab. Towards unsupervised image captioning with shared multimodal embeddings. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[36] Jianxun Lou, Hanhe Lin, David Marshall, Dietmar Saupe, and Hantao Liu. Transalnet: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[37] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7086-7096, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[39] Olivier Le Meur, Patrick Le Callet, and Dominique Barba. Predicting visual fixations on video based on low-level visual features. Vision Research, 47(19):2483 - 2498, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[40] Dong Nguyen. Comparing automatic and human evaluation of local explanations for text classification. In Proceedings of the 2018 Conference on North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[41] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397 - 2416, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[42] Thammathip Piumsomboon, Gun Lee, Robert W. Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In IEEE Symposium on 3D User Interfaces (3DUI), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[43] Yao Qiang, Deng Pan, Chengyin Li, Rhongho Jang, and Dongxiao Zhu. Attcat: Explaining transformers via attentive" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "class activation tokens. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[44] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning (ICML), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[45] Vincent Sitzmann, Ana Serrano, Amy Pavel, Maneesh Agrawala, Diego Gutierrez, Belen Masia, and Gordon Wetzstein. Saliency in vr: How do people explore virtual environments? IEEE Transactions on Visualization and Computer Graphics, 24(4):1633-1642, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[46] Yubao Sun, Mengyang Zhao, Kai Hu, and Shaojing Fan. Visual saliency prediction using multi-scale attention gated network. Multimedia Systems, 28(1):131-139, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[47] Hamed R. Tavakoli, Rakshith Shetty, Ali Borji, and Jorma Laaksonen. Paying attention to descriptions generated by image captioning models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[48] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question answering. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 451-466. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[49] Juan Xu, Ming Jiang, Shuo Wang, Mohan S Kankanhalli, and Qi Zhao. Predicting human gaze beyond pixels. Journal of vision, 14(1):28-28, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[50] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[51] Jinhui Yang, Xianyu Chen, Ming Jiang, Shi Chen, Louis Wang, and Qi Zhao. VisualHow: Multimodal problem solving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[52] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. TMM, 22(8):2163-2176, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[54] Wei Zhang and Hantao Liu. Study of saliency in objective video quality assessment. IEEE Transactions on Image Processing, 26(3):1275-1288, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9610" + } + ] +] \ No newline at end of file diff --git a/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_origin.pdf b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..66636195e08e6e7004966fc51b2730cb0c132af4 --- /dev/null +++ b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/5a9d4569-3627-48db-87ca-f6f74c038d6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de8fa3e9fe51516703ccf1296bbe541939b3d30f98b041f3b97d1e49bd60975d +size 1544724 diff --git a/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/full.md b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b39097b9c18b374b66ca0f9f24e370176b6673cc --- /dev/null +++ b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/full.md @@ -0,0 +1,399 @@ +# Explainable Saliency: Articulating Reasoning with Contextual Prioritization + +Nuo Chen + +Ming Jiang + +Qi Zhao + +Department of Computer Science and Engineering University of Minnesota + +{chen9029, mjiang, qzhao}@umn.edu + +# Abstract + +Deep saliency models, which predict what parts of an image capture our attention, are often like black boxes. This limits their use, especially in areas where understanding why a model makes a decision is crucial. Our research tackles this challenge by developing an explainable saliency (XSal) model that not only identifies what is important in an image, but also explains its choices in a way that makes sense to humans. We achieve this by using vision-language models to reason about images and by focusing the model's attention on the most crucial information using a contextual prioritization mechanism. Unlike prior approaches that rely on fixation descriptions or soft-attention based semantic aggregation, our method directly models the reasoning steps involved in saliency prediction, generating selectively prioritized explanations clarify why specific regions are prioritized. Comprehensive evaluations demonstrate the effectiveness of our model in generating high-quality saliency maps and coherent, contextually relevant explanations. This research is a step towards more transparent and trustworthy AI systems that can help us understand and navigate the world around us. + +# 1. Introduction + +Humans have this amazing ability to look at a scene and instantly know what is important. Saliency models, the cornerstone of computer vision, aim to mimic human gaze, identifying the most captivating regions within an image. These models have found wide application in image and video quality assessment [4, 54], virtual reality [42, 45], autonomous systems [8, 20], image captioning [13], and clinical diagnostics [11, 21]. However, while current saliency models [26, 36, 52] excel at replicating generic gaze patterns, they often operate as enigmatic black boxes. This lack of transparency hinders their application in scenarios where understanding why a region is deemed salient is as crucial as the prediction itself. + +The development of explainable models that mimic hu + +![](images/041e108dddd326591628d520158dfe384eabc9456149c6beeb95834f716b7ffd.jpg) +Figure 1. Our proposed XSal model provides step-by-step natural language explanations, semantic prototype visualization, as well as saliency weights to explain its prediction. + +man attention remains an underexplored research frontier. While some models have attempted to introduce explainability, they often rely on massive predefined vocabularies of semantics [10, 11] or simple descriptions of eye fixations without deep reasoning [15]. These methods, while offering some insights, lack true interpretability due to the lack of reasoning and prioritization within the context. This limitation is rooted in the inherent complexity of human attention. Human attention is a multifaceted phenomenon, influenced by a confluence of factors including visual cues, semantic understanding, individual preferences, and even cultural context. To generate meaningful explanations that unravel this interplay, a deep understanding of human cognition and perception is essential. Moreover, creating a model that can articulate its reasoning in a clear and concise manner poses a significant technical challenge. + +To bridge the gap between saliency prediction and interpretability, we develop a saliency model that not only accurately predicts salient regions but also provides clear, human-understandable explanations. As shown in Fig. 1, our goal is to factorize saliency prediction with a weighted combination of spatial feature maps, namely semantic prototypes. Each semantic prototype represents the grounding of a semantic proposal (e.g., frisbee, fence, and sheep) relevant to the visual task, and its relevance is explained with a natural language description. In this example, "frisbee" and "sheep" have higher weights in saliency prediction, because "to find the frisbee's location" and "to check if there is a sheep to the right of the fence" are essential reasoning steps to answer the question "Is there a sheep to the right of the fence that is to the right of the frisbee?" + +To achieve this, we propose an explainable saliency (XSal) model that advances the state-of-the-art in saliency prediction and explanation. First, it integrates a vision-language model (VLM) to perform explicit reasoning for saliency prediction, generating step-by-step explanations and delivering natural language interpretations that clarify why specific regions are salient. Second, we introduce a contextual prioritization mechanism designed to handle the challenges of selecting truly important ones from a potentially large set of semantic explanations. This mechanism allows the model to focus on key elements, achieving interpretability without overcomplicating the representation. By integrating explicit reasoning with targeted selection, we create a cohesive framework that not only identifies and explains salient regions but also adapts dynamically to scene complexity. + +In summary, this paper makes three key contributions to the field of saliency prediction and explainable AI: + +1. This paper presents the first saliency model to incorporate a vision-language model for explicit reasoning. This integration allows the model to not only identify salient regions but also articulate its reasoning process, providing clear and human-understandable explanations. +2. We introduce a novel contextual prioritization mechanism that prioritizes the most relevant information within the context of saliency prediction and explanation. This results in a more interpretable model by focusing on key relevant semantics and simplifying the representation. +3. We conduct extensive evaluations of our model, demonstrating its effectiveness in generating high-quality saliency maps while providing coherent, contextually relevant explanations. This comprehensive assessment highlights the model's ability to effectively bridge the gap between prediction accuracy and interpretability. + +# 2. Related Work + +Saliency Prediction. Saliency prediction models have traditionally focused on bottom-up approaches, relying on + +low-level visual features in free-viewing conditions [5, 25, 27, 29]. However, with the rapid progress of deep learning methods, deep saliency models have emerged, achieving promising accuracy in predicting both task-free bottom-up saliency [26, 34, 52] and task-driven top-down attention [1, 30, 48]. A key strength of deep saliency models is their ability to extract high-level semantic features. Numerous deep neural network-based methods have boosted saliency prediction performance by implicitly encoding semantics using different approaches [46, 52]. However, such models often lack transparency and trustworthiness due to explicitly learned representations. While there has been growing interest in understanding saliency model mechanisms, existing approaches fall short of providing comprehensive explanations. For instance, Chen et al. [10, 11] analyze model behaviors using semantic prototypes and soft attention weights, but these approaches offer limited insight into the model's decision-making process, particularly when dealing with thousands of prototypes. Similarly, GazeXplain [15] generates plain natural language descriptions of fixation points, lacking a reasoning or prioritization mechanism to identify crucial factors driving its predictions. These limitations highlight the need for a more robust methodology that can uncover the underlying rationale behind saliency model behavior. + +Our proposed method addresses these limitations by introducing a novel framework that combines explicit vision-language reasoning with a hard attention mechanism. This ensures that only the most critical elements are selected for saliency prediction and explanation, which not only improves the accuracy of saliency predictions but also provides a more transparent and insightful understanding of the model's behavior. + +Vision Language Models. Our work builds upon the growing field of explainable AI (XAI) within vision-language models (VLMs). Recent advancements in VLMs [13, 14, 17, 22, 31, 35, 47, 50, 51] have revolutionized multimodal understanding, enabling these models to seamlessly integrate visual and textual information. This progress has opened new avenues for developing XAI systems capable of providing interpretable insights into complex decision-making processes. For instance, researchers have explored using VLMs to generate natural language explanations for tasks such as image classification [23, 38], abusive language detection [19], and autonomous driving [7], allowing users to understand the rationale behind model predictions. Additionally, VLMs have been employed to enhance the interpretability of visual scanpath models, describing image features driving human attention to a sequence of fixation points [15]. + +Different from these studies, by integrating textual context and introducing reasoning and prioritization mechanisms, we aim to provide deeper insights into the factors + +![](images/c3d07ba1ed8a2d68840c8a8265484df918cc42fcb823737b58d6f0357720aaa9.jpg) +Figure 2. The proposed XSal architecture begins with a vision-language model generating semantic proposals from the input image, and uses a segmentation model to generate their corresponding semantic prototypes. A contextual prioritization module then selects the top-K relevant semantic prototypes, which are processed by a final saliency head to predict a focused saliency map. + +driving saliency predictions, improving the transparency and interpretability of saliency models. + +# 3. Method + +The human visual attention process is a captivating blend of explicit reasoning and selective focus, guiding us to prioritize relevant elements while navigating complex scenes. While existing saliency models have made strides, they often fall short of capturing this nuanced behavior, relying instead on dense representations or lacking a mechanism for explaining their choices. This incongruence between model behavior and human attention motivates us to introduce a novel approach that leverages VLMs to perform explicit reasoning and prioritizing the reasoning outcomes based on contextual relevance for saliency prediction and explanation. + +# 3.1. Explicit Reasoning + +At the heart of our approach lies a vision-language model (VLM), specifically MiniCPM-V 2.6 [53], which enables the model to identify and explain salient elements across task-driven or task-free contexts. Specifically, the VLM produces a set of semantic proposals $\mathbf{C}$ and associated explanations $\mathbf{R}$ : + +$$ +\mathbf {C}, \mathbf {R} = \operatorname {V L M} (\mathbf {I}, \mathbf {Q}) \tag {1} +$$ + +Here, $\mathbf{I}$ is the input image, and $\mathbf{Q}$ is the question guiding attention. $\mathbf{C} = \{P_1,P_2,\dots ,P_n\}$ represents semantic proposals, each focusing on specific objects within the scene, while $\mathbf{R}$ provides rationale for why each element was + +![](images/628d3d12c74b2398fee286c3fcafdac4d6f799a9d2cffe8de5508e86eb4ce8c3.jpg) + +Can you explain step by step how you analyzed the image to answer the question? + +Question: What is the drink in the cup made of plastic called? + +Image: + +![](images/89abd22ae4177de0aef4df90d1c9717ab592f40342b4a9cfbfa76ab1a410d7b5.jpg) +Figure 3. The two-stage prompting method to obtain semantic proposals and explanations from the VLM. + +1. I identified the drink in the cup. +2. I examined the cup's appearance to determine its material. +3. I confirmed the cup was plastic based on its transparency. +4. Finally, I considered common drinks served in plastic cups. + +![](images/31010921043e78619354a5adfff5e6f2f4520b3607e990efa4cfabca4ce96c43.jpg) + +![](images/65b634e2207a108f8c6cd0ee3fe59aa8779ebfeac403771edff103d1fc3389bd.jpg) + +In each step, which objects did you focus on, and why? + +1. I focused on the drink in the cup, because it was necessary to identify the main object in question. +2. I focused on the cup to determine its material by examining its appearance and texture. +3. I focused on the plastic to confirm the material based on its transparency and common usage. +4. I focused on the drink in the plastic cup to the specific type of drinks served in such cups. + +![](images/2568075b9e3a8baa464cc452ec6ec8ffd7d59c3e3ab4d2daac9d942554f8263f.jpg) + +selected, ensuring interpretability through alignment with human attention patterns. This integration supports a rea + +soning process that mirrors human problem-solving, where decisions are informed by both visual cues and a structured understanding of why certain elements are relevant. To obtain the semantic proposals and explanations from the VLM, as shown in Fig. 3, we employ a two-stage prompting method: + +In the initial planning stage, we ask the VLM to describe "how you analyzed the image to answer the question" (for top-down task-driven cases) or "important objects in the image" (for bottom-up free-viewing cases). The model responds with a coherent sequence of actions of reasoning steps to follow when answering the question. For example, to answer the question "What is the drink in the cup made of plastic called?", the reasoning steps include identifying the drink in the cup, examining the cup's appearance to determine its material, confirming the cup was plastic based on its transparency, and considering common drinks served in plastic cups. + +Following this planning, the VLM is asked to explain what specific objects it focused on in each step, and explain the reason. It then proceeds to generate the set of semantic proposals $\mathbf{C}$ (such as drink, cup, and plastic) and their corresponding explanations $\mathbf{R}$ . Finally, to ensure precision in the final output, the VLM is also prompted to verify that the identified semantic proposals are concrete and relevant, addressing potential limitations in model consistency. + +This layered approach replicates human selective attention and explanatory reasoning, ensuring the model's focus and reasoning align closely with human attention patterns. + +After identifying semantic proposals through the VLM, we utilize an open-vocabulary segmentation model, ClipSeg [37], to extract the visual grounding of these proposals on the image. ClipSeg's ability to handle open-vocabulary concepts is crucial as the language model may describe elements that don't perfectly align with traditional segmentation categories. This step bridges the semantic reasoning process with visual representations, allowing for visual validation of the model's reasoning. While ClipSeg effectively outlines entire object contours, human attention typically focuses on smaller, more localized regions. Therefore, instead of using the segmentation results directly for saliency prediction, we use them as binary masks to filter the image features, and apply a convolutional layer $\mathcal{F}$ to obtain fine-grained localization. This results in a sequence of focused spatial maps, namely semantic prototypes, where each prototype $S_{i}$ aligns with a semantic proposal $P_{i}$ , localizing its representations in the image. This process pinpoints focused regions that naturally attract human attention, associating them with natural language explanations. + +# 3.2. Contextual Prioritization Mechanism + +Our approach utilizes a contextual prioritization mechanism to extract the most relevant semantic prototypes for saliency + +prediction. Different from existing saliency models that often integrate hundreds to thousands of spatial feature maps with linear weights, humans naturally focus on a limited number of features, while other elements, though present, are not prioritized to the same extent. Our proposed contextual prioritization is a highly selective approach, which only allows attention to be directed by the most relevant features. It introduces a dynamic approach that adapts to varying numbers of semantic prototypes. This flexibility allows the model to handle scenes of different complexities effectively. + +Our contextual prioritization mechanism adjusts importance by applying global average pooling on the extracted semantic prototypes, followed by a linear layer to generate initial weights. These weights are then sorted, and only the top $K$ and their corresponding features are selected. A softmax operation is applied to normalize the selected weights, yielding a refined representation: + +$$ +\mathbf {w} = \operatorname {L i n e a r} (\operatorname {G A P} (\mathbf {S})), \tag {2} +$$ + +$$ +\mathbf {w} _ {\text {t o p K}} = \operatorname {S o f t m a x} (\operatorname {T o p K} (\mathbf {w})), \tag {3} +$$ + +$$ +\mathrm {O} = \sum_ {i = 1} ^ {K} \mathbf {w} _ {\text {t o p K}, i} \cdot \mathbf {S} _ {\text {t o p K}, i} \tag {4} +$$ + +where $\mathbf{S}$ represents a collection of semantic prototypes, $\{S_1,S_2,\dots ,S_n\}$ . The global average pooling (GAP) and subsequent linear transformation extract initial weights $\mathbf{w}$ for each prototype, allowing the model to assign varying importance to each semantic element. By selecting and normalizing the top $K$ weights, we obtain $\mathbf{w}_{\mathrm{topK}}$ and the corresponding $\mathbf{S}_{\mathrm{topK}}$ . Here, $\mathbf{S}_{\mathrm{topK}}$ corresponds to the semantic prototypes associated with the top $K$ weights in $\mathbf{w}$ , determined by sorting. This ensures that the model prioritizes the most significant semantic elements, aligning its focus with the most contextually important regions, and discarding less relevant elements. Finally, the aggregated saliency feature O, the weighted combination of the top $K$ semantic prototypes, is processed by a saliency head to produce the final saliency prediction. This saliency head consists of only a single linear layer (with bias set to 0) that simply compresses the channel dimension, acting as a weighted sum without introducing any complex black-box behavior. + +To ensure adequate feature optimization and prevent limited feature selection from hindering convergence, we introduce a Progressive Top-K Reduction strategy. Rather than starting with a fixed, small set of key features, this strategy initially preserves a larger number of semantic elements, enabling comprehensive feature optimization in early training. As training progresses, the $K$ value is gradually reduced, guiding the model to focus on the most significant features. The progression of $K$ across epochs is defined as + +$$ +K _ {t} = K _ {\text {i n i t}} - \frac {\left(K _ {\text {i n i t}} - K _ {\text {f i n a l}}\right) \times t}{T - 1} \tag {5} +$$ + +where $K_{t}$ is the $K$ at epoch $t$ , $K_{\mathrm{init}}$ and $K_{\mathrm{final}}$ are the initial and final values of $K$ , and $T$ is the total number of epochs. This progressive reduction allows the model to gradually shift from broad feature optimization to a selective focus on the most critical elements. With this strategy, the contextual prioritization mimics the adaptive prioritization seen in human attention, flexibly weighing elements based on their relevance and adapting to scenes with different numbers of semantic prototypes. + +# 4. Experiments + +This section presents a comprehensive evaluation of our proposed model, examining its performance and explainability across both task-driven and task-free eye-tracking datasets. We conduct an ablation study to further investigate the contributions of individual components, and present qualitative examples to illustrate the model's capabilities in generating interpretable saliency maps and explanations. Additional experimental results and analyses are provided in the supplementary material for further reference. + +# 4.1. Experimental Setup + +Datasets. Our experiments are conducted on two eye-tracking datasets: The AiR dataset [9], derived from the GQA dataset, comprises images and questions paired with eye-tracking data from 20 participants answering those questions. This setup provides a realistic scenario for evaluating saliency models, as it captures human attention in the context of a specific task (question answering). The OSIE [49] dataset is a free-viewing dataset featuring multiple salient objects competing for attention in the same image context, suitable for developing and evaluating saliency models that consider object- and semantic-level information. In our experiments, we split both datasets into training, validation, and test sets following [12, 16]. + +Compared Models. We compare our model's performance against several state-of-the-art saliency prediction models including SAM [18], SALICON [26], DINet [52], as well as explainable models such as Chen et al. [10] and GazeXplain [15]. Note that for GazeXplain [15] we aggregate fixation points from its predicted scanpaths into saliency maps for comparison. By comparing our model with these diverse methods, we aim to demonstrate its performance in various aspects, including general saliency prediction and explainability. + +Evaluation Metrics. To comprehensively evaluate our model, we utilize a combination of metrics that assess saliency prediction performance and explainability. 1) Saliency Prediction Metrics: We employ standard metrics to evaluate the alignment between our model's predicted + +saliency maps and human attention patterns. These metrics are Area Under the ROC Curve (AUC) [3], Normalized Scanpath Saliency (NSS) [41], and Correlation Coefficient (CC) [39], which provide quantitative insights into the accuracy and relevance of the generated saliency maps. 2) Explainability Metrics: We evaluate the interpretability of saliency predictions adapting three XAI metrics: AUC [2] assesses model stability by calculating the area under the curve that represents changes in model performance as different proportions of semantic prototypes are progressively perturbed. Note that this AUC differs from the AUC used in saliency evaluation: here, it measures the model's sensitivity to perturbations, while saliency AUC measures alignment with human attention data. To facilitate distinction, we refer to this interpretability metric as AUC-E throughout this paper. Area Over the Perturbation Curve (AOPC) [6, 40] quantifies the sensitivity of the model to perturbations by measuring the change in NSS as semantic prototypes are sequentially altered. Log-odds Score (LOdds) [43, 44] evaluates the change in NSS as semantic prototypes are perturbed, focusing on the degree of reduction in NSS in response to prototype alterations. These metrics provide a comprehensive framework for assessing the quality of our model's explanations, enabling us to gauge not only where the model focuses, but also the robustness of its interpretative reasoning under perturbations. + +Loss Function. To align the model's saliency predictions effectively with human attention patterns, we employ a combination of loss functions, including NSS, CC, and Kullback-Leibler Divergence (KLD) [33]. Our implementation follows [10]. These loss functions provide complementary signals that guide the model to capture nuanced attention details, helping to achieve balanced performance across various saliency prediction metrics. + +Implementation Details. Our model employs a dilated ResNet-50 [24] backbone, following the setting used in DINet [52]. This architecture incorporates dilated convolutions to expand the receptive field without additional downsampling, which enhances the model's capacity to capture high-resolution spatial details essential for accurate saliency prediction. The model is optimized using the Adam optimizer [32] with an initial learning rate of 4e-4 and a weight decay of 1e-7. To ensure stable convergence, the learning rate is adaptively reduced by a decay factor of 0.1 every two epochs, while gradient clipping is applied with a threshold of 0.1 to prevent gradient explosion. Training is conducted over 10 epochs with a batch size of 10. To ensure stable training and prevent convergence issues, we employ the proposed Progressive Top-K Reduction strategy, which begins with a broader feature selection and gradually narrows the focus to the most significant features. Specifically, we set the initial $K$ value to 20, which progressively reduces to a final value of 3 over 3 epochs, enabling adaptive and selec + +
AiROSIE
NSSCCNSSCC
SAM [18]1.650.622.700.65
SALICON [28]1.700.632.750.63
DINet [52]1.770.632.880.63
Chen et al. [10]1.760.632.910.64
GazeXplain [15]1.850.662.530.74
XSal1.940.702.960.72
+ +Table 1. Comparative results of saliency prediction across the AiR and OSIE datasets. + +tive feature prioritization. + +# 4.2. Saliency Prediction Results + +Tab. 1 presents a quantitative comparison of saliency prediction performance across the AiR (task-driven) and OSIE (task-free) saliency datasets for our proposed model and several state-of-the-art methods. As shown in the table, our proposed model consistently outperforms other methods on the AiR dataset, achieving the highest NSS score of 1.94 and the highest CC score of 0.70. This suggests that our model effectively captures top-down human attention patterns in task-driven scenarios. Our model also achieves the highest NSS score of 2.96 on the OSIE dataset. Though our CC score (0.72) is slightly lower than GazeXplain (0.74), it still demonstrates strong performance in capturing the overall distribution of human attention in task-free image-viewing. The consistently high NSS scores across both datasets indicates that our model accurately predicts the locations where humans are mostly likely to fixate their gaze, demonstrating a strong accuracy of predicting the most salient regions. However, our model might be slightly less effective at capturing the overall distribution of human attention in images, because image regions with low saliency (low agreements across human observers) are less explainable and likely to be discarded by the contextual prioritization mechanism. Overall, the quantitative results demonstrate that our proposed method significantly improves saliency prediction performance compared to existing approaches. The incorporation of explicit reasoning and selective attention mechanisms contributes to more accurate and robust models that effectively capture the complexity of human visual attention. + +# 4.3. Saliency Explanation Results + +Tab. 2 presents the quantitative results based on explainability metrics, comparing our model's performance with Chen et al. [10] on both the AiR and OSIE datasets. Notably, + +our model demonstrates significantly lower AUC-E, LOdds, and higher AOPC scores compared to Chen et al. Specifically, on the AiR dataset, our model significantly outperforms Chen et al. with a lower AUC-E (0.713 vs. 0.899) and a higher AOPC (0.748 vs. 0.241). The substantial improvement in LOdds (-5.157 vs. -0.143) also reflects the model's ability to focus on the most relevant elements, minimizing irrelevant attention. On the OSIE dataset, a similar trend is observed. Our model achieves a better AUC-E (0.685 vs. 0.810) and significantly higher AOPC (1.115 vs. 0.586) compared to Chen et al. Additionally, the considerable decrease in LOdds (-4.169 vs. -0.250) further supports our model's capability to highlight key semantic concepts more effectively, thereby improving interpretability in a task-free viewing context. + +These results indicate that our model is more sensitive to changes in semantic prototypes, highlighting the stronger reliance on these prototypes for generating accurate saliency predictions. The lower AUC-E and LOdds scores suggest that perturbing prototypes leads to a more substantial decrease in performance for our model, highlighting the critical role of semantic understanding in our approach. Conversely, the higher AOPC score signifies a greater sensitivity to changes in prototypes, revealing a stronger correlation between the model's performance and the identified semantic elements. This reinforces the notion that our model's interpretability is closely tied to the semantic prototypes it extracts, providing a more transparent and interpretable reasoning process compared to methods that rely on complex aggregation of numerous features. Overall, these results emphasize the importance of explicit reasoning and selective attention in achieving a higher degree of interpretability and understanding in saliency prediction. + +# 4.4. Ablation Study + +Our ablation study, conducted on both the AiR and OSIE datasets, reveals the significant contributions of each component to our model's performance and explainability. As shown in Tab. 3, a simple baseline model, using only a dilated ResNet-50 backbone [24] with a saliency head, achieves reasonable performance in both settings, yielding an NSS of 1.777 on AiR and 2.835 on OSIE. The baseline model lacks explainability, with a LOdds of -0.131 and -0.246 on these datasets, respectively. + +By incorporating the VLM for reasoning, our model links saliency with task-relevant semantic elements and significantly improves both accuracy and interpretability across both datasets. For instance, on the AiR dataset, the NSS increases to 1.917 and the CC score improves to 0.703, while the AUC remains relatively stable at 0.858. This highlights the crucial role of task-oriented alignment and semantic understanding in generating clear and contextual explanations for saliency predictions. This is further evidenced + +
AiROSIE
AUC-E↓AOPC↑LOdds↓AUC-E↓AOPC↑LOdds↓
Chen et al. [52]0.8990.241-0.1430.8100.586-0.250
XSal0.7130.748-5.1570.6851.115-4.169
+ +Table 2. Comparative results of explainability metrics across the AiR and OSIE datasets. + +
DatasetBaselineReasoningContPriNSSCCAUCAUC-E↓AOPC ↑LOdds ↓
AiR✓1.7770.6400.8490.9100.225-0.131
✓✓1.9170.7030.8580.7000.749-5.077
✓✓✓1.9420.7010.8580.7130.748-5.157
OSIE✓2.8350.6810.8850.8230.567-0.246
✓✓2.9040.7170.8970.7650.876-1.738
✓✓✓2.9650.7260.8950.6851.115-4.169
+ +Table 3. Ablation study on the impacts of Baseline, Reasoning, and Contextual Prioritization (ContPri). + +by the significant improvements in interpretability metrics – the AUC-E drops to 0.700 and the AOPC increases to 0.749, suggesting our model's ability to provide more coherent explanations for task-driven saliency. + +Further, the contextual prioritization mechanism exhibits a more pronounced impact, especially in the OSIE task-free setting. For example, on OSIE, the NSS increases to 2.965, the CC improves to 0.726, and the AUC remains relatively stable at 0.895. This suggests that prioritizing the most salient elements plays a crucial role in handling complex scenes with multiple potential salient objects, ultimately leading to more concise and clear attention maps. This is further evidenced by the significant improvements in interpretability metrics – the AUC-E decreases to 0.685, the AOPC increases to 1.115, and the LOdds decreases to -4.169. This greater impact on the OSIE dataset could be attributed to the presence of a large number of potential semantic prototypes in task-free settings, making selective focus particularly valuable for distilling relevant information and improving interpretability. + +These findings demonstrate that the combination of explicit reasoning and contextual prioritization mechanisms, implemented in our model, is essential for achieving superior performance and explainability in both task-driven and task-free saliency prediction. + +# 4.5. Qualitative Analysis + +Qualitative analysis of our model reveals its ability to effectively integrate task-driven guidance and contextual prioritization to generate accurate saliency predictions and explanations. As illustrated in Fig. 4, the model aligns its attention with the goals of each question, selectively focusing on essential objects while ignoring irrelevant elements. For ex + +ample, in questions about the position of a skateboard and a red cone (row 1), the model prioritizes these objects over traditionally salient regions like the skateboarder, demonstrating the advantage of explicit reasoning. The model also adapts well to abstract concepts, such as interpreting "modern" style in a question about a faucet (row 2). Finally, the model excels in context prioritization by isolating key elements in complex scenes, highlighting the person, bag, and cars in a question about cars (row 3) while filtering out other visual elements. These qualitative findings demonstrate the model's ability to produce explanations that are closely aligned with the model's internal reasoning, ensuring outputs that are both interpretable and reliable. + +Beyond its interpretability, our model shows significant performance advantages in accuracy and task relevance. In Fig. 5, our model achieves the highest accuracy among state-of-the-art methods, aligning closely with ground truth. For example, in the question "What is the drink in the cup made of plastic called?", it focuses on the plastic cup, unlike other models that are distracted by irrelevant food items (row 1), showcasing the strength of our explicit reasoning approach. Additionally, our model demonstrates strong context prioritization, effectively concentrating on relevant elements even in complex scenes with multiple distractions. For instance, in the question "Is the fire hydrant to the right or to the left of the garbage can?", our model isolates the hydrant and garbage can, while other models diffuse attention across the scene (row 2), highlighting its ability to prioritize contextually important elements. Furthermore, in dynamic scenes, our model reliably focuses on salient objects. In the question "Are both the helmet and the bat the same color?", it highlights the helmet and bat specifically, unlike other models that diffuse attention (row 3), demonstrating + +![](images/25e90289825018cfd898e673aa37ddf6e2c7ba241d06207868706693dc8ec4b5.jpg) +Image + +![](images/38e540eadb7bf36144315c8250cc891ea9f43e943773b62d2dea4a3125fc4d7e.jpg) +Ground Truth + +![](images/225f8ddc761de2c3e62a6392ab736cf40885eb6ccd38ed749b2080a6ba68c021.jpg) +Prediction + +![](images/84c1bc806cd7f4b63d1e486e192c211fd462c20982c68026a56e728178eaf74b.jpg) +Prototype 1 +$\mathbf{B}\mathbf{a}\mathbf{g}(w_1 = 0.378)$ + +The question specifies a bag, so it is relevant to locate its position in the image and use it as a reference point for determining the direction of 'to the right'. + +![](images/1f38d6140f5c8ab2af3136b689b8a84370b2dbc79eb9edca1feee4652f9a3b24.jpg) +Prototype 2 +Person $(w_{2} = 0.330)$ + +To identify the reference point for 'to the right' in relation to the bag and determine where cars are located relative to this person. + +![](images/481dd8e1266e3d3564130bd7611455502be11aafc81faea7241bfe0db5c111ea.jpg) +Prototype 3 +Cars $(w_{3} = 0.291)$ + +Cars are mentioned in the question and need to be identified in their location relative to the person and bag. + +Q: Are there any cars to the right of the person that is to the right of the bag? + +![](images/7c6a07fc5f102915ce0553d72bbb60219b49817fd259f19f8d7a5b1b19c49fde.jpg) + +![](images/35844a742d6a68031532d5eaa8f3db51734d31fc7f525bb354163bd6931f77e4.jpg) + +![](images/c01f2fb14723aa66248dc1b97865973157f76fe7349766966c28685229b1c61e.jpg) + +Q: Does the faucet to the left of the vase look modern and silver? + +![](images/3f8e80c2a4d00e6ac5f1a5cff738bc20b70738db238aaef42aed6535b095a643.jpg) +Vase $(w_{1} = 0.357)$ + +The vase is mentioned in the question and serves as a reference point for the faucet's location. + +![](images/cfa26a18d07460e3f87cf10c9722e5d224670706158e2546bdb85409fa49e108.jpg) +Faucets $(w_{2} = 0.319)$ + +The question specifically asks about the faucet's appearance, so it was a key element in answering. + +![](images/73418ff3002998be5af46fd654c9e29e9090b747eda550ab2292fe827469eecf.jpg) +Modern design $(w_{3} = 0.329)$ + +The question refers to the faucets style, which can be described as modern. + +![](images/f9a5f490c053d17f027d35927a1e85c927106887d14d56fa75b9da74e4f451cb.jpg) +Figure 4. Qualitative examples from our methods. Each semantic prototype is shown along with their saliency weights and explanations. + +![](images/4e9996dc591349d633107f6234428d5fc005a96de2d0d05bde846f4c5842d3ca.jpg) + +![](images/f51a130be62d71c429bb45d861dc8933dba5f406f22f614ab9a56030c973c35e.jpg) + +![](images/5ae8d40964d00dd12a4c791ee3ae85d2a2589bdc384616b6872fd50eb1f71313.jpg) +Skateboard $(w_{1} = 0.371)$ + +The skateboard is the object being asked about in terms of its position relative to the red cone. + +![](images/e03a9a852250e436a2c56ec46ddeb3a5baa887a71598426a9a6974ac475c9ad8.jpg) + +Red cone $(w_{2} = 0.344)$ red cone is the reference point for determining the position of the skateboard relative to the question. + +![](images/d32b077db8140c35298b7196607f4e2f92b2363dd10baba7698c8f6cef0bda9f.jpg) +Skateboarder $(w_{3} = 0.283)$ + +The skateboarder position and orientation are crucial in understanding where the skateboard is in relation to the red cone. + +Q: What is the drink in the cup made of plastic called? + +Q: Is the fire hydrant to the right or to the left of the garbage can? + +Q: Are both the helmet and the bat the same color? + +![](images/43436310f7f66fd499635549f286c6794cd96306c2a0f62fd1d30b9df64910a2.jpg) +Image + +![](images/7ebe140a887c7385c263615a87b8b035762f39e2a3d09592ed45c81032bf215d.jpg) + +![](images/7defecdca0b81525a3aa8753daa761923d9cc49d9538aa2abfff39a2ae4d18b4.jpg) +Figure 5. Qualitative comparison of saliency predictions across models (XSal, DINet, SALICON, SAM). Each row shows an input question and image, with the corresponding ground truth attention map and model outputs. + +![](images/393c224a27810f79eef0c9f6866d0cc7f158f608acec605dd7db468893378402.jpg) +Ground Truth + +![](images/9cd7c79544fa4b95386325409697857c67173c1b9c00ffcf9732a2e5d1fa5fdd.jpg) + +![](images/647d6486bc775446863a22177768988618e0e8271d5f3cac66345ca48d287d93.jpg) + +![](images/acd996e56cfd054df6d94413eca7c2d58e3bcc16ec18890d0d7a909263870ea0.jpg) +XSal + +![](images/567b83f83a3be5a57484d6d3b72d975a82fb4ed93db6394b88d5e2dd0a0b1ff9.jpg) + +![](images/bea1ecdd7b565fbe3d857a8d75e4e93094c7be2e06bd607b021ad988b4f0145f.jpg) + +![](images/9cfa01521c156af9fe089fb56433eb3a05a687386c3a95660e3b7a020cf6619c.jpg) +DINet + +![](images/da89bbf62cbc4c925612f7e68cfab91804fe08a3ff6018c72e461d86e2c9fbd5.jpg) + +![](images/380397bdde725dadd0d89fa07d72d1677a73186cac69984515e9c84c33ffb081.jpg) + +![](images/3645347d2168522f59ac6de9d09ed894b51a33a810691b007f0f82b3cef13605.jpg) +SALICON + +![](images/c01d014c3b0f577e5d13bae3ca7346cffb1b845adbc8368694eb8743df0f1ae4.jpg) + +![](images/cdcd09de8683348dadcc95ad5ff4790c51e5e8d716cce47cd8d969dbd96b67e5.jpg) + +![](images/5f074d7b10c467ab73ca4737ebe44ad4f427a0567c03e7e14b214ba3817b0984.jpg) +SAM + +![](images/ba6375fd8a3649f61e058544dfb38db17c80f92316777a8aad0c64c4400f8a73.jpg) + +![](images/ec76b46edcd49346e4109185e1b90ff4d8888ddfe40cbfe8334064739b556055.jpg) + +its adaptability to diverse scenarios. These examples highlight our model's superior performance in generating accurate and contextually relevant saliency predictions. + +# 5. Conclusion + +This paper has presented a novel approach for saliency prediction that combines explicit reasoning and contextual prioritization, generating both accurate and explainable saliency predictions. We have demonstrated that our method, powered by a VLM, outperforms existing models in terms of accuracy and explainability. This is achieved by leveraging an explicit reasoning process that allows the + +model to provide clear explanations for its predictions and a contextual prioritization mechanism that focuses on a few truly important elements, mimicking human attention behavior. Our approach not only achieves state-of-the-art saliency prediction accuracy but also offers valuable insights into the decision-making processes of saliency models. This advancement contributes to the development of more trustworthy and reliable saliency models. + +# Acknowledgments + +This work is supported by NSF Grant 2143197 and Fasikl Incorporated. + +# References + +[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2 +[2] Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 5 +[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5 +[4] Patrick Le Callet and Ernst Niebur. Visual attention and applications in multimedia technologies. Proceedings of the Institution of Electrical Engineers, 2013. 1 +[5] Moran Cerf, E Paxon Frady, and Christof Koch. Faces and text attract gaze independent of the task: Experimental data and computer model. Journal of vision, 9(12):10-10, 2009. 2 +[6] Hanjie Chen, Guangtao Zheng, and Yangfeng Ji. Generating hierarchical explanations on text classification via feature interaction detection. In Proceedings of the 2020 Conference on Association for Computational Linguistics (ACL), 2020. 5 +[7] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2 +[8] Nuo Chen, Jin Xie, Jing Nie, Jiale Cao, Zhuang Shao, and Yanwei Pang. Attentive alignment network for multispectral pedestrian detection. In Proceedings of the 31st ACM international conference on multimedia, pages 3787-3795, 2023. 1 +[9] Shi Chen, Ming Jiang, Jinhui Yang, and Qi Zhao. AiR: Attention with reasoning capability. In European Conference on Computer Vision (ECCV), 2020. 5 +[10] Shi Chen, Ming Jiang, and Qi Zhao. What do deep saliency models learn about visual attention? In Advances in Neural Information Processing Systems, pages 9543-9555. Curran Associates, Inc., 2023. 1, 2, 5, 6 +[11] Shi Chen, Ming Jiang, and Qi Zhao. Deep learning to interpret autism spectrum disorder behind the camera. IEEE Transactions on Cognitive and Developmental Systems, pages 1-12, 2024. 1, 2 +[12] Xianyu Chen, Ming Jiang, and Qi Zhao. Predicting human scanpaths in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10876-10885, 2021. 5 +[13] Xianyu Chen, Ming Jiang, and Qi Zhao. Leveraging human attention in novel object captioning. In Proceedings of the + +International Joint Conference on Artificial Intelligence (IJ-CAI), 2021. 1, 2 +[14] Xianyu Chen, Ming Jiang, and Qi Zhao. Self-distillation for few-shot image captioning. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2 +[15] Xianyu Chen, Ming Jiang, and Qi Zhao. Gazexplain: Learning to predict natural language explanations of visual scanpaths. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2, 5, 6 +[16] Xianyu Chen, Ming Jiang, and Qi Zhao. Beyond average: Individualized visual scanpath prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5 +[17] Xianyu Chen, Jinhui Yang, Shi Chen, Louis Wang, Ming Jiang, and Qi Zhao. Every problem, every step, all in focus: Learning to solve real-world problems with integrated attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI), 2024. 2 +[18] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting Human Eye Fixations via an LSTM-based Saliency Attentive Model. TIP, 27(10):5142-5154, 2018. 5, 6 +[19] Chiara Di Bonaventura, Lucia Siciliani, Pierpaolo Basile, Albert Merono Penuela, and Barbara McGillivray. Is explanation all you need? an expert survey on llm-generated explanations for abusive language detection. In Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), 2024. 2 +[20] Ning Ding, Ce Zhang, and Azim Eskandarian. Saliendet: A saliency-based feature enhancement algorithm for object detection for autonomous driving. IEEE Transactions on Intelligent Vehicles, 9(1):2624-2635, 2023. 1 +[21] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Zhaohui Che, Yi Fang, Xiaokang Yang, Jesús Gutierrez, and Patrick Le Callet. A dataset of eye movements for the children with autism spectrum disorder. In ACM Multimedia Systems Conference (MMSys), 2019. 1 +[22] Yang Feng, Lin Ma, Wei Liu, and Jiebo Luo. Unsupervised image captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[23] Songhao Han, Le Zhuo, Yue Liao, and Si Liu. Llms as visual explainers: Advancing image classification with evolving visual descriptions. arXiv preprint arXiv:2311.11904, 2023. 2 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5, 6 +[25] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2 +[26] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In ICCV, pages 262-270, 2015. 1, 2, 5 +[27] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions + +on Pattern Analysis and Machine Intelligence, 20(11):1254-1259, 1998. 2 +[28] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Silicon: Saliency in context. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6 +[29] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In 2009 IEEE 12th international conference on computer vision, pages 2106-2113. IEEE, 2009. 2 +[30] Christopher Kanan, Mathew H Tong, Lingyun Zhang, and Garrison W Cottrell. Sun: Top-down saliency using natural statistics. Visual cognition, 17(6-7):979-1003, 2009. 2 +[31] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Image captioning with very scarce supervised data: Adversarial semi-supervised learning approach. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 2 +[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[33] Solomon Kullback. Information Theory and Statistics. Wiley, 1959. 5 +[34] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. In ICLR workshop, 2015. 2 +[35] Iro Laina, Christian Rupprecht, and Nassir Navab. Towards unsupervised image captioning with shared multimodal embeddings. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2 +[36] Jianxun Lou, Hanhe Lin, David Marshall, Dietmar Saupe, and Hantao Liu. Transalnet: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 1 +[37] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7086-7096, 2022. 4 +[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2 +[39] Olivier Le Meur, Patrick Le Callet, and Dominique Barba. Predicting visual fixations on video based on low-level visual features. Vision Research, 47(19):2483 - 2498, 2007. 5 +[40] Dong Nguyen. Comparing automatic and human evaluation of local explanations for text classification. In Proceedings of the 2018 Conference on North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 5 +[41] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397 - 2416, 2005. 5 +[42] Thammathip Piumsomboon, Gun Lee, Robert W. Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In IEEE Symposium on 3D User Interfaces (3DUI), 2017. 1 +[43] Yao Qiang, Deng Pan, Chengyin Li, Rhongho Jang, and Dongxiao Zhu. Attcat: Explaining transformers via attentive + +class activation tokens. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2022. 5 +[44] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning (ICML), 2017. 5 +[45] Vincent Sitzmann, Ana Serrano, Amy Pavel, Maneesh Agrawala, Diego Gutierrez, Belen Masia, and Gordon Wetzstein. Saliency in vr: How do people explore virtual environments? IEEE Transactions on Visualization and Computer Graphics, 24(4):1633-1642, 2018. 1 +[46] Yubao Sun, Mengyang Zhao, Kai Hu, and Shaojing Fan. Visual saliency prediction using multi-scale attention gated network. Multimedia Systems, 28(1):131-139, 2022. 2 +[47] Hamed R. Tavakoli, Rakshith Shetty, Ali Borji, and Jorma Laaksonen. Paying attention to descriptions generated by image captioning models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[48] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question answering. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 451-466. Springer, 2016. 2 +[49] Juan Xu, Ming Jiang, Shuo Wang, Mohan S Kankanhalli, and Qi Zhao. Predicting human gaze beyond pixels. Journal of vision, 14(1):28-28, 2014. 5 +[50] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2 +[51] Jinhui Yang, Xianyu Chen, Ming Jiang, Shi Chen, Louis Wang, and Qi Zhao. VisualHow: Multimodal problem solving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[52] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. TMM, 22(8):2163-2176, 2020. 1, 2, 5, 6, 7 +[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3 +[54] Wei Zhang and Hantao Liu. Study of saliency in objective video quality assessment. IEEE Transactions on Image Processing, 26(3):1275-1288, 2017. 1 \ No newline at end of file diff --git a/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/images.zip b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..240dc3fa460483df37e2daa268f9665a284944a8 --- /dev/null +++ b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:594cd42a97e89d244c014ccf000199881cf61a3489c5ff26b70396b054c16b51 +size 443677 diff --git a/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/layout.json b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..31edc15478be2cca15bbb718f2482041f127ba2c --- /dev/null +++ b/2025/Explainable Saliency_ Articulating Reasoning with Contextual Prioritization/layout.json @@ -0,0 +1,9754 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "text", + "content": "Explainable Saliency: Articulating Reasoning with Contextual Prioritization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 144, + 249, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 144, + 249, + 157 + ], + "spans": [ + { + "bbox": [ + 198, + 144, + 249, + 157 + ], + "type": "text", + "content": "Nuo Chen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 273, + 144, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 144, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 273, + 144, + 329, + 157 + ], + "type": "text", + "content": "Ming Jiang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 354, + 144, + 396, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 144, + 396, + 157 + ], + "spans": [ + { + "bbox": [ + 354, + 144, + 396, + 157 + ], + "type": "text", + "content": "Qi Zhao" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 175, + 158, + 418, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 158, + 418, + 185 + ], + "spans": [ + { + "bbox": [ + 175, + 158, + 418, + 185 + ], + "type": "text", + "content": "Department of Computer Science and Engineering University of Minnesota" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 206, + 188, + 385, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 188, + 385, + 198 + ], + "spans": [ + { + "bbox": [ + 206, + 188, + 385, + 198 + ], + "type": "text", + "content": "{chen9029, mjiang, qzhao}@umn.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 251, + 290, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 503 + ], + "type": "text", + "content": "Deep saliency models, which predict what parts of an image capture our attention, are often like black boxes. This limits their use, especially in areas where understanding why a model makes a decision is crucial. Our research tackles this challenge by developing an explainable saliency (XSal) model that not only identifies what is important in an image, but also explains its choices in a way that makes sense to humans. We achieve this by using vision-language models to reason about images and by focusing the model's attention on the most crucial information using a contextual prioritization mechanism. Unlike prior approaches that rely on fixation descriptions or soft-attention based semantic aggregation, our method directly models the reasoning steps involved in saliency prediction, generating selectively prioritized explanations clarify why specific regions are prioritized. Comprehensive evaluations demonstrate the effectiveness of our model in generating high-quality saliency maps and coherent, contextually relevant explanations. This research is a step towards more transparent and trustworthy AI systems that can help us understand and navigate the world around us." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 545, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 701 + ], + "type": "text", + "content": "Humans have this amazing ability to look at a scene and instantly know what is important. Saliency models, the cornerstone of computer vision, aim to mimic human gaze, identifying the most captivating regions within an image. These models have found wide application in image and video quality assessment [4, 54], virtual reality [42, 45], autonomous systems [8, 20], image captioning [13], and clinical diagnostics [11, 21]. However, while current saliency models [26, 36, 52] excel at replicating generic gaze patterns, they often operate as enigmatic black boxes. This lack of transparency hinders their application in scenarios where understanding why a region is deemed salient is as crucial as the prediction itself." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "The development of explainable models that mimic hu" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 225, + 533, + 467 + ], + "blocks": [ + { + "bbox": [ + 312, + 225, + 533, + 467 + ], + "lines": [ + { + "bbox": [ + 312, + 225, + 533, + 467 + ], + "spans": [ + { + "bbox": [ + 312, + 225, + 533, + 467 + ], + "type": "image", + "image_path": "041e108dddd326591628d520158dfe384eabc9456149c6beeb95834f716b7ffd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 478, + 545, + 511 + ], + "lines": [ + { + "bbox": [ + 305, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 478, + 545, + 511 + ], + "type": "text", + "content": "Figure 1. Our proposed XSal model provides step-by-step natural language explanations, semantic prototype visualization, as well as saliency weights to explain its prediction." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "man attention remains an underexplored research frontier. While some models have attempted to introduce explainability, they often rely on massive predefined vocabularies of semantics [10, 11] or simple descriptions of eye fixations without deep reasoning [15]. These methods, while offering some insights, lack true interpretability due to the lack of reasoning and prioritization within the context. This limitation is rooted in the inherent complexity of human attention. Human attention is a multifaceted phenomenon, influenced by a confluence of factors including visual cues, semantic understanding, individual preferences, and even cultural context. To generate meaningful explanations that unravel this interplay, a deep understanding of human cognition and perception is essential. Moreover, creating a model that can articulate its reasoning in a clear and concise manner poses a significant technical challenge." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9601" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "type": "text", + "content": "To bridge the gap between saliency prediction and interpretability, we develop a saliency model that not only accurately predicts salient regions but also provides clear, human-understandable explanations. As shown in Fig. 1, our goal is to factorize saliency prediction with a weighted combination of spatial feature maps, namely semantic prototypes. Each semantic prototype represents the grounding of a semantic proposal (e.g., frisbee, fence, and sheep) relevant to the visual task, and its relevance is explained with a natural language description. In this example, \"frisbee\" and \"sheep\" have higher weights in saliency prediction, because \"to find the frisbee's location\" and \"to check if there is a sheep to the right of the fence\" are essential reasoning steps to answer the question \"Is there a sheep to the right of the fence that is to the right of the frisbee?\"" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 252, + 286, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 252, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 286, + 442 + ], + "type": "text", + "content": "To achieve this, we propose an explainable saliency (XSal) model that advances the state-of-the-art in saliency prediction and explanation. First, it integrates a vision-language model (VLM) to perform explicit reasoning for saliency prediction, generating step-by-step explanations and delivering natural language interpretations that clarify why specific regions are salient. Second, we introduce a contextual prioritization mechanism designed to handle the challenges of selecting truly important ones from a potentially large set of semantic explanations. This mechanism allows the model to focus on key elements, achieving interpretability without overcomplicating the representation. By integrating explicit reasoning with targeted selection, we create a cohesive framework that not only identifies and explains salient regions but also adapts dynamically to scene complexity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 443, + 286, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 443, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 49, + 443, + 286, + 467 + ], + "type": "text", + "content": "In summary, this paper makes three key contributions to the field of saliency prediction and explainable AI:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 468, + 286, + 659 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 468, + 286, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 286, + 528 + ], + "type": "text", + "content": "1. This paper presents the first saliency model to incorporate a vision-language model for explicit reasoning. This integration allows the model to not only identify salient regions but also articulate its reasoning process, providing clear and human-understandable explanations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 528, + 286, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 528, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 528, + 286, + 586 + ], + "type": "text", + "content": "2. We introduce a novel contextual prioritization mechanism that prioritizes the most relevant information within the context of saliency prediction and explanation. This results in a more interpretable model by focusing on key relevant semantics and simplifying the representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 587, + 286, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 286, + 659 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 286, + 659 + ], + "type": "text", + "content": "3. We conduct extensive evaluations of our model, demonstrating its effectiveness in generating high-quality saliency maps while providing coherent, contextually relevant explanations. This comprehensive assessment highlights the model's ability to effectively bridge the gap between prediction accuracy and interpretability." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 690, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 690, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 690, + 286, + 713 + ], + "type": "text", + "content": "Saliency Prediction. Saliency prediction models have traditionally focused on bottom-up approaches, relying on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 72, + 544, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 544, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 544, + 371 + ], + "type": "text", + "content": "low-level visual features in free-viewing conditions [5, 25, 27, 29]. However, with the rapid progress of deep learning methods, deep saliency models have emerged, achieving promising accuracy in predicting both task-free bottom-up saliency [26, 34, 52] and task-driven top-down attention [1, 30, 48]. A key strength of deep saliency models is their ability to extract high-level semantic features. Numerous deep neural network-based methods have boosted saliency prediction performance by implicitly encoding semantics using different approaches [46, 52]. However, such models often lack transparency and trustworthiness due to explicitly learned representations. While there has been growing interest in understanding saliency model mechanisms, existing approaches fall short of providing comprehensive explanations. For instance, Chen et al. [10, 11] analyze model behaviors using semantic prototypes and soft attention weights, but these approaches offer limited insight into the model's decision-making process, particularly when dealing with thousands of prototypes. Similarly, GazeXplain [15] generates plain natural language descriptions of fixation points, lacking a reasoning or prioritization mechanism to identify crucial factors driving its predictions. These limitations highlight the need for a more robust methodology that can uncover the underlying rationale behind saliency model behavior." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 374, + 544, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 374, + 544, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 374, + 544, + 468 + ], + "type": "text", + "content": "Our proposed method addresses these limitations by introducing a novel framework that combines explicit vision-language reasoning with a hard attention mechanism. This ensures that only the most critical elements are selected for saliency prediction and explanation, which not only improves the accuracy of saliency predictions but also provides a more transparent and insightful understanding of the model's behavior." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 472, + 544, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 544, + 674 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 544, + 674 + ], + "type": "text", + "content": "Vision Language Models. Our work builds upon the growing field of explainable AI (XAI) within vision-language models (VLMs). Recent advancements in VLMs [13, 14, 17, 22, 31, 35, 47, 50, 51] have revolutionized multimodal understanding, enabling these models to seamlessly integrate visual and textual information. This progress has opened new avenues for developing XAI systems capable of providing interpretable insights into complex decision-making processes. For instance, researchers have explored using VLMs to generate natural language explanations for tasks such as image classification [23, 38], abusive language detection [19], and autonomous driving [7], allowing users to understand the rationale behind model predictions. Additionally, VLMs have been employed to enhance the interpretability of visual scanpath models, describing image features driving human attention to a sequence of fixation points [15]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 678, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 678, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 678, + 544, + 712 + ], + "type": "text", + "content": "Different from these studies, by integrating textual context and introducing reasoning and prioritization mechanisms, we aim to provide deeper insights into the factors" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "9602" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 534, + 291 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 534, + 291 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 534, + 291 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 534, + 291 + ], + "type": "image", + "image_path": "c3d07ba1ed8a2d68840c8a8265484df918cc42fcb823737b58d6f0357720aaa9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "text", + "content": "Figure 2. The proposed XSal architecture begins with a vision-language model generating semantic proposals from the input image, and uses a segmentation model to generate their corresponding semantic prototypes. A contextual prioritization module then selects the top-K relevant semantic prototypes, which are processed by a final saliency head to predict a focused saliency map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "text", + "content": "driving saliency predictions, improving the transparency and interpretability of saliency models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 103, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 103, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 103, + 381 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 288, + 533 + ], + "type": "text", + "content": "The human visual attention process is a captivating blend of explicit reasoning and selective focus, guiding us to prioritize relevant elements while navigating complex scenes. While existing saliency models have made strides, they often fall short of capturing this nuanced behavior, relying instead on dense representations or lacking a mechanism for explaining their choices. This incongruence between model behavior and human attention motivates us to introduce a novel approach that leverages VLMs to perform explicit reasoning and prioritizing the reasoning outcomes based on contextual relevance for saliency prediction and explanation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 542, + 158, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 158, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 158, + 555 + ], + "type": "text", + "content": "3.1. Explicit Reasoning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 561, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 287, + 633 + ], + "type": "text", + "content": "At the heart of our approach lies a vision-language model (VLM), specifically MiniCPM-V 2.6 [53], which enables the model to identify and explain salient elements across task-driven or task-free contexts. Specifically, the VLM produces a set of semantic proposals " + }, + { + "bbox": [ + 46, + 561, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 561, + 287, + 633 + ], + "type": "text", + "content": " and associated explanations " + }, + { + "bbox": [ + 46, + 561, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 561, + 287, + 633 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 646, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 646, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 123, + 646, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {C}, \\mathbf {R} = \\operatorname {V L M} (\\mathbf {I}, \\mathbf {Q}) \\tag {1}", + "image_path": "a71f4393512b5f936658966d53092d36afad30c80d71afe0e31846ab9506b8c8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " is the input image, and " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " is the question guiding attention. " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{P_1,P_2,\\dots ,P_n\\}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " represents semantic proposals, each focusing on specific objects within the scene, while " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " provides rationale for why each element was" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 333, + 332, + 356 + ], + "blocks": [ + { + "bbox": [ + 310, + 333, + 332, + 356 + ], + "lines": [ + { + "bbox": [ + 310, + 333, + 332, + 356 + ], + "spans": [ + { + "bbox": [ + 310, + 333, + 332, + 356 + ], + "type": "image", + "image_path": "628d3d12c74b2398fee286c3fcafdac4d6f799a9d2cffe8de5508e86eb4ce8c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 341, + 529, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 341, + 529, + 358 + ], + "spans": [ + { + "bbox": [ + 353, + 341, + 529, + 358 + ], + "type": "text", + "content": "Can you explain step by step how you analyzed the image to answer the question?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 358, + 532, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 358, + 532, + 367 + ], + "spans": [ + { + "bbox": [ + 348, + 358, + 532, + 367 + ], + "type": "text", + "content": "Question: What is the drink in the cup made of plastic called?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 368, + 376, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 368, + 376, + 376 + ], + "spans": [ + { + "bbox": [ + 354, + 368, + 376, + 376 + ], + "type": "text", + "content": "Image:" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 418, + 369, + 531, + 454 + ], + "blocks": [ + { + "bbox": [ + 418, + 369, + 531, + 454 + ], + "lines": [ + { + "bbox": [ + 418, + 369, + 531, + 454 + ], + "spans": [ + { + "bbox": [ + 418, + 369, + 531, + 454 + ], + "type": "image", + "image_path": "89abd22ae4177de0aef4df90d1c9717ab592f40342b4a9cfbfa76ab1a410d7b5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 638, + 545, + 660 + ], + "lines": [ + { + "bbox": [ + 306, + 638, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 545, + 660 + ], + "type": "text", + "content": "Figure 3. The two-stage prompting method to obtain semantic proposals and explanations from the VLM." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 471, + 496, + 506 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 320, + 471, + 419, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 471, + 419, + 479 + ], + "spans": [ + { + "bbox": [ + 320, + 471, + 419, + 479 + ], + "type": "text", + "content": "1. I identified the drink in the cup." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 480, + 493, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 480, + 493, + 488 + ], + "spans": [ + { + "bbox": [ + 320, + 480, + 493, + 488 + ], + "type": "text", + "content": "2. I examined the cup's appearance to determine its material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 488, + 492, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 488, + 492, + 497 + ], + "spans": [ + { + "bbox": [ + 320, + 488, + 492, + 497 + ], + "type": "text", + "content": "3. I confirmed the cup was plastic based on its transparency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 498, + 496, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 498, + 496, + 506 + ], + "spans": [ + { + "bbox": [ + 320, + 498, + 496, + 506 + ], + "type": "text", + "content": "4. Finally, I considered common drinks served in plastic cups." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 520, + 464, + 544, + 488 + ], + "blocks": [ + { + "bbox": [ + 520, + 464, + 544, + 488 + ], + "lines": [ + { + "bbox": [ + 520, + 464, + 544, + 488 + ], + "spans": [ + { + "bbox": [ + 520, + 464, + 544, + 488 + ], + "type": "image", + "image_path": "31010921043e78619354a5adfff5e6f2f4520b3607e990efa4cfabca4ce96c43.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 310, + 517, + 331, + 539 + ], + "blocks": [ + { + "bbox": [ + 310, + 517, + 331, + 539 + ], + "lines": [ + { + "bbox": [ + 310, + 517, + 331, + 539 + ], + "spans": [ + { + "bbox": [ + 310, + 517, + 331, + 539 + ], + "type": "image", + "image_path": "65b634e2207a108f8c6cd0ee3fe59aa8779ebfeac403771edff103d1fc3389bd.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 353, + 524, + 512, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 524, + 512, + 533 + ], + "spans": [ + { + "bbox": [ + 353, + 524, + 512, + 533 + ], + "type": "text", + "content": "In each step, which objects did you focus on, and why?" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 319, + 552, + 498, + 620 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 319, + 552, + 496, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 552, + 496, + 569 + ], + "spans": [ + { + "bbox": [ + 319, + 552, + 496, + 569 + ], + "type": "text", + "content": "1. I focused on the drink in the cup, because it was necessary to identify the main object in question." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 319, + 570, + 498, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 570, + 498, + 586 + ], + "spans": [ + { + "bbox": [ + 319, + 570, + 498, + 586 + ], + "type": "text", + "content": "2. I focused on the cup to determine its material by examining its appearance and texture." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 319, + 586, + 498, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 586, + 498, + 604 + ], + "spans": [ + { + "bbox": [ + 319, + 586, + 498, + 604 + ], + "type": "text", + "content": "3. I focused on the plastic to confirm the material based on its transparency and common usage." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 319, + 604, + 498, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 604, + 498, + 620 + ], + "spans": [ + { + "bbox": [ + 319, + 604, + 498, + 620 + ], + "type": "text", + "content": "4. I focused on the drink in the plastic cup to the specific type of drinks served in such cups." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 520, + 545, + 544, + 568 + ], + "blocks": [ + { + "bbox": [ + 520, + 545, + 544, + 568 + ], + "lines": [ + { + "bbox": [ + 520, + 545, + 544, + 568 + ], + "spans": [ + { + "bbox": [ + 520, + 545, + 544, + 568 + ], + "type": "image", + "image_path": "2568075b9e3a8baa464cc452ec6ec8ffd7d59c3e3ab4d2daac9d942554f8263f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "selected, ensuring interpretability through alignment with human attention patterns. This integration supports a rea" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9603" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "content": "soning process that mirrors human problem-solving, where decisions are informed by both visual cues and a structured understanding of why certain elements are relevant. To obtain the semantic proposals and explanations from the VLM, as shown in Fig. 3, we employ a two-stage prompting method:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 289 + ], + "type": "text", + "content": "In the initial planning stage, we ask the VLM to describe \"how you analyzed the image to answer the question\" (for top-down task-driven cases) or \"important objects in the image\" (for bottom-up free-viewing cases). The model responds with a coherent sequence of actions of reasoning steps to follow when answering the question. For example, to answer the question \"What is the drink in the cup made of plastic called?\", the reasoning steps include identifying the drink in the cup, examining the cup's appearance to determine its material, confirming the cup was plastic based on its transparency, and considering common drinks served in plastic cups." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 385 + ], + "type": "text", + "content": "Following this planning, the VLM is asked to explain what specific objects it focused on in each step, and explain the reason. It then proceeds to generate the set of semantic proposals " + }, + { + "bbox": [ + 46, + 289, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 289, + 287, + 385 + ], + "type": "text", + "content": " (such as drink, cup, and plastic) and their corresponding explanations " + }, + { + "bbox": [ + 46, + 289, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 289, + 287, + 385 + ], + "type": "text", + "content": ". Finally, to ensure precision in the final output, the VLM is also prompted to verify that the identified semantic proposals are concrete and relevant, addressing potential limitations in model consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 385, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 287, + 421 + ], + "type": "text", + "content": "This layered approach replicates human selective attention and explanatory reasoning, ensuring the model's focus and reasoning align closely with human attention patterns." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "text", + "content": "After identifying semantic proposals through the VLM, we utilize an open-vocabulary segmentation model, ClipSeg [37], to extract the visual grounding of these proposals on the image. ClipSeg's ability to handle open-vocabulary concepts is crucial as the language model may describe elements that don't perfectly align with traditional segmentation categories. This step bridges the semantic reasoning process with visual representations, allowing for visual validation of the model's reasoning. While ClipSeg effectively outlines entire object contours, human attention typically focuses on smaller, more localized regions. Therefore, instead of using the segmentation results directly for saliency prediction, we use them as binary masks to filter the image features, and apply a convolutional layer " + }, + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "text", + "content": " to obtain fine-grained localization. This results in a sequence of focused spatial maps, namely semantic prototypes, where each prototype " + }, + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "text", + "content": " aligns with a semantic proposal " + }, + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 46, + 422, + 288, + 661 + ], + "type": "text", + "content": ", localizing its representations in the image. This process pinpoints focused regions that naturally attract human attention, associating them with natural language explanations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 244, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 244, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 244, + 682 + ], + "type": "text", + "content": "3.2. Contextual Prioritization Mechanism" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Our approach utilizes a contextual prioritization mechanism to extract the most relevant semantic prototypes for saliency" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "prediction. Different from existing saliency models that often integrate hundreds to thousands of spatial feature maps with linear weights, humans naturally focus on a limited number of features, while other elements, though present, are not prioritized to the same extent. Our proposed contextual prioritization is a highly selective approach, which only allows attention to be directed by the most relevant features. It introduces a dynamic approach that adapts to varying numbers of semantic prototypes. This flexibility allows the model to handle scenes of different complexities effectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 204, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 288 + ], + "type": "text", + "content": "Our contextual prioritization mechanism adjusts importance by applying global average pooling on the extracted semantic prototypes, followed by a linear layer to generate initial weights. These weights are then sorted, and only the top " + }, + { + "bbox": [ + 304, + 204, + 545, + 288 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 204, + 545, + 288 + ], + "type": "text", + "content": " and their corresponding features are selected. A softmax operation is applied to normalize the selected weights, yielding a refined representation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 375, + 299, + 545, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 299, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 375, + 299, + 545, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {w} = \\operatorname {L i n e a r} (\\operatorname {G A P} (\\mathbf {S})), \\tag {2}", + "image_path": "5c66e0a30bc51f17212a3606dd5d129cff60c4c026fbd95cf0166e1e90648f7a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 328, + 545, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 328, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 361, + 328, + 545, + 341 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {\\text {t o p K}} = \\operatorname {S o f t m a x} (\\operatorname {T o p K} (\\mathbf {w})), \\tag {3}", + "image_path": "841bfd9f2e75d8f60d9718fdab8721313d0feb61767b6c4980cb6f690e7f11a7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 371, + 354, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 371, + 354, + 545, + 387 + ], + "type": "interline_equation", + "content": "\\mathrm {O} = \\sum_ {i = 1} ^ {K} \\mathbf {w} _ {\\text {t o p K}, i} \\cdot \\mathbf {S} _ {\\text {t o p K}, i} \\tag {4}", + "image_path": "1099e54edd7718c0bea63b8be9e4a4f9667cf2e85dcd4714c20fa9ab3dbc9297.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": " represents a collection of semantic prototypes, " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\{S_1,S_2,\\dots ,S_n\\}" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": ". The global average pooling (GAP) and subsequent linear transformation extract initial weights " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": " for each prototype, allowing the model to assign varying importance to each semantic element. By selecting and normalizing the top " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": " weights, we obtain " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{topK}}" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": " and the corresponding " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{topK}}" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{topK}}" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": " corresponds to the semantic prototypes associated with the top " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": " weights in " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": ", determined by sorting. This ensures that the model prioritizes the most significant semantic elements, aligning its focus with the most contextually important regions, and discarding less relevant elements. Finally, the aggregated saliency feature O, the weighted combination of the top " + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 391, + 545, + 605 + ], + "type": "text", + "content": " semantic prototypes, is processed by a saliency head to produce the final saliency prediction. This saliency head consists of only a single linear layer (with bias set to 0) that simply compresses the channel dimension, acting as a weighted sum without introducing any complex black-box behavior." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "To ensure adequate feature optimization and prevent limited feature selection from hindering convergence, we introduce a Progressive Top-K Reduction strategy. Rather than starting with a fixed, small set of key features, this strategy initially preserves a larger number of semantic elements, enabling comprehensive feature optimization in early training. As training progresses, the " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " value is gradually reduced, guiding the model to focus on the most significant features. The progression of " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " across epochs is defined as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9604" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 99, + 80, + 287, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 80, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 99, + 80, + 287, + 105 + ], + "type": "interline_equation", + "content": "K _ {t} = K _ {\\text {i n i t}} - \\frac {\\left(K _ {\\text {i n i t}} - K _ {\\text {f i n a l}}\\right) \\times t}{T - 1} \\tag {5}", + "image_path": "dec9364261bd1ff031fc892bda7b97a5333e46a2f911182d16bd32412395dcc5.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": " at epoch " + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{init}}" + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{final}}" + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": " are the initial and final values of " + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 108, + 289, + 216 + ], + "type": "text", + "content": " is the total number of epochs. This progressive reduction allows the model to gradually shift from broad feature optimization to a selective focus on the most critical elements. With this strategy, the contextual prioritization mimics the adaptive prioritization seen in human attention, flexibly weighing elements based on their relevance and adapting to scenes with different numbers of semantic prototypes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 226, + 128, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 226, + 128, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 128, + 239 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 246, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 354 + ], + "type": "text", + "content": "This section presents a comprehensive evaluation of our proposed model, examining its performance and explainability across both task-driven and task-free eye-tracking datasets. We conduct an ablation study to further investigate the contributions of individual components, and present qualitative examples to illustrate the model's capabilities in generating interpretable saliency maps and explanations. Additional experimental results and analyses are provided in the supplementary material for further reference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 360, + 163, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 163, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 163, + 373 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 379, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 534 + ], + "type": "text", + "content": "Datasets. Our experiments are conducted on two eye-tracking datasets: The AiR dataset [9], derived from the GQA dataset, comprises images and questions paired with eye-tracking data from 20 participants answering those questions. This setup provides a realistic scenario for evaluating saliency models, as it captures human attention in the context of a specific task (question answering). The OSIE [49] dataset is a free-viewing dataset featuring multiple salient objects competing for attention in the same image context, suitable for developing and evaluating saliency models that consider object- and semantic-level information. In our experiments, we split both datasets into training, validation, and test sets following [12, 16]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "type": "text", + "content": "Compared Models. We compare our model's performance against several state-of-the-art saliency prediction models including SAM [18], SALICON [26], DINet [52], as well as explainable models such as Chen et al. [10] and GazeXplain [15]. Note that for GazeXplain [15] we aggregate fixation points from its predicted scanpaths into saliency maps for comparison. By comparing our model with these diverse methods, we aim to demonstrate its performance in various aspects, including general saliency prediction and explainability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. To comprehensively evaluate our model, we utilize a combination of metrics that assess saliency prediction performance and explainability. 1) Saliency Prediction Metrics: We employ standard metrics to evaluate the alignment between our model's predicted" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 384 + ], + "type": "text", + "content": "saliency maps and human attention patterns. These metrics are Area Under the ROC Curve (AUC) [3], Normalized Scanpath Saliency (NSS) [41], and Correlation Coefficient (CC) [39], which provide quantitative insights into the accuracy and relevance of the generated saliency maps. 2) Explainability Metrics: We evaluate the interpretability of saliency predictions adapting three XAI metrics: AUC [2] assesses model stability by calculating the area under the curve that represents changes in model performance as different proportions of semantic prototypes are progressively perturbed. Note that this AUC differs from the AUC used in saliency evaluation: here, it measures the model's sensitivity to perturbations, while saliency AUC measures alignment with human attention data. To facilitate distinction, we refer to this interpretability metric as AUC-E throughout this paper. Area Over the Perturbation Curve (AOPC) [6, 40] quantifies the sensitivity of the model to perturbations by measuring the change in NSS as semantic prototypes are sequentially altered. Log-odds Score (LOdds) [43, 44] evaluates the change in NSS as semantic prototypes are perturbed, focusing on the degree of reduction in NSS in response to prototype alterations. These metrics provide a comprehensive framework for assessing the quality of our model's explanations, enabling us to gauge not only where the model focuses, but also the robustness of its interpretative reasoning under perturbations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": "Loss Function. To align the model's saliency predictions effectively with human attention patterns, we employ a combination of loss functions, including NSS, CC, and Kullback-Leibler Divergence (KLD) [33]. Our implementation follows [10]. These loss functions provide complementary signals that guide the model to capture nuanced attention details, helping to achieve balanced performance across various saliency prediction metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. Our model employs a dilated ResNet-50 [24] backbone, following the setting used in DINet [52]. This architecture incorporates dilated convolutions to expand the receptive field without additional downsampling, which enhances the model's capacity to capture high-resolution spatial details essential for accurate saliency prediction. The model is optimized using the Adam optimizer [32] with an initial learning rate of 4e-4 and a weight decay of 1e-7. To ensure stable convergence, the learning rate is adaptively reduced by a decay factor of 0.1 every two epochs, while gradient clipping is applied with a threshold of 0.1 to prevent gradient explosion. Training is conducted over 10 epochs with a batch size of 10. To ensure stable training and prevent convergence issues, we employ the proposed Progressive Top-K Reduction strategy, which begins with a broader feature selection and gradually narrows the focus to the most significant features. Specifically, we set the initial " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " value to 20, which progressively reduces to a final value of 3 over 3 epochs, enabling adaptive and selec" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9605" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 267, + 208 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 267, + 208 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 267, + 208 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 267, + 208 + ], + "type": "table", + "html": "
AiROSIE
NSSCCNSSCC
SAM [18]1.650.622.700.65
SALICON [28]1.700.632.750.63
DINet [52]1.770.632.880.63
Chen et al. [10]1.760.632.910.64
GazeXplain [15]1.850.662.530.74
XSal1.940.702.960.72
", + "image_path": "fbc66cdfd3b44deabd0477d599009efe3e484c5b0929f746ee85a4c4265a4cf1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 217, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 286, + 239 + ], + "type": "text", + "content": "Table 1. Comparative results of saliency prediction across the AiR and OSIE datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 262, + 151, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 151, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 151, + 274 + ], + "type": "text", + "content": "tive feature prioritization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 283, + 196, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 196, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 196, + 295 + ], + "type": "text", + "content": "4.2. Saliency Prediction Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 302, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 287, + 648 + ], + "type": "text", + "content": "Tab. 1 presents a quantitative comparison of saliency prediction performance across the AiR (task-driven) and OSIE (task-free) saliency datasets for our proposed model and several state-of-the-art methods. As shown in the table, our proposed model consistently outperforms other methods on the AiR dataset, achieving the highest NSS score of 1.94 and the highest CC score of 0.70. This suggests that our model effectively captures top-down human attention patterns in task-driven scenarios. Our model also achieves the highest NSS score of 2.96 on the OSIE dataset. Though our CC score (0.72) is slightly lower than GazeXplain (0.74), it still demonstrates strong performance in capturing the overall distribution of human attention in task-free image-viewing. The consistently high NSS scores across both datasets indicates that our model accurately predicts the locations where humans are mostly likely to fixate their gaze, demonstrating a strong accuracy of predicting the most salient regions. However, our model might be slightly less effective at capturing the overall distribution of human attention in images, because image regions with low saliency (low agreements across human observers) are less explainable and likely to be discarded by the contextual prioritization mechanism. Overall, the quantitative results demonstrate that our proposed method significantly improves saliency prediction performance compared to existing approaches. The incorporation of explicit reasoning and selective attention mechanisms contributes to more accurate and robust models that effectively capture the complexity of human visual attention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 205, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 205, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 205, + 671 + ], + "type": "text", + "content": "4.3. Saliency Explanation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Tab. 2 presents the quantitative results based on explainability metrics, comparing our model's performance with Chen et al. [10] on both the AiR and OSIE datasets. Notably," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "our model demonstrates significantly lower AUC-E, LOdds, and higher AOPC scores compared to Chen et al. Specifically, on the AiR dataset, our model significantly outperforms Chen et al. with a lower AUC-E (0.713 vs. 0.899) and a higher AOPC (0.748 vs. 0.241). The substantial improvement in LOdds (-5.157 vs. -0.143) also reflects the model's ability to focus on the most relevant elements, minimizing irrelevant attention. On the OSIE dataset, a similar trend is observed. Our model achieves a better AUC-E (0.685 vs. 0.810) and significantly higher AOPC (1.115 vs. 0.586) compared to Chen et al. Additionally, the considerable decrease in LOdds (-4.169 vs. -0.250) further supports our model's capability to highlight key semantic concepts more effectively, thereby improving interpretability in a task-free viewing context." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 252, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 468 + ], + "type": "text", + "content": "These results indicate that our model is more sensitive to changes in semantic prototypes, highlighting the stronger reliance on these prototypes for generating accurate saliency predictions. The lower AUC-E and LOdds scores suggest that perturbing prototypes leads to a more substantial decrease in performance for our model, highlighting the critical role of semantic understanding in our approach. Conversely, the higher AOPC score signifies a greater sensitivity to changes in prototypes, revealing a stronger correlation between the model's performance and the identified semantic elements. This reinforces the notion that our model's interpretability is closely tied to the semantic prototypes it extracts, providing a more transparent and interpretable reasoning process compared to methods that rely on complex aggregation of numerous features. Overall, these results emphasize the importance of explicit reasoning and selective attention in achieving a higher degree of interpretability and understanding in saliency prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 478, + 399, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 399, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 399, + 491 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "type": "text", + "content": "Our ablation study, conducted on both the AiR and OSIE datasets, reveals the significant contributions of each component to our model's performance and explainability. As shown in Tab. 3, a simple baseline model, using only a dilated ResNet-50 backbone [24] with a saliency head, achieves reasonable performance in both settings, yielding an NSS of 1.777 on AiR and 2.835 on OSIE. The baseline model lacks explainability, with a LOdds of -0.131 and -0.246 on these datasets, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "By incorporating the VLM for reasoning, our model links saliency with task-relevant semantic elements and significantly improves both accuracy and interpretability across both datasets. For instance, on the AiR dataset, the NSS increases to 1.917 and the CC score improves to 0.703, while the AUC remains relatively stable at 0.858. This highlights the crucial role of task-oriented alignment and semantic understanding in generating clear and contextual explanations for saliency predictions. This is further evidenced" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9606" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 70, + 470, + 137 + ], + "blocks": [ + { + "bbox": [ + 124, + 70, + 470, + 137 + ], + "lines": [ + { + "bbox": [ + 124, + 70, + 470, + 137 + ], + "spans": [ + { + "bbox": [ + 124, + 70, + 470, + 137 + ], + "type": "table", + "html": "
AiROSIE
AUC-E↓AOPC↑LOdds↓AUC-E↓AOPC↑LOdds↓
Chen et al. [52]0.8990.241-0.1430.8100.586-0.250
XSal0.7130.748-5.1570.6851.115-4.169
", + "image_path": "b1c3c0bce96afc34cbd2fcc51ae6df50b37ff53146b61d89e44ebd8f7007b742.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 167, + 513, + 269 + ], + "blocks": [ + { + "bbox": [ + 135, + 144, + 457, + 156 + ], + "lines": [ + { + "bbox": [ + 135, + 144, + 457, + 156 + ], + "spans": [ + { + "bbox": [ + 135, + 144, + 457, + 156 + ], + "type": "text", + "content": "Table 2. Comparative results of explainability metrics across the AiR and OSIE datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 167, + 513, + 269 + ], + "lines": [ + { + "bbox": [ + 82, + 167, + 513, + 269 + ], + "spans": [ + { + "bbox": [ + 82, + 167, + 513, + 269 + ], + "type": "table", + "html": "
DatasetBaselineReasoningContPriNSSCCAUCAUC-E↓AOPC ↑LOdds ↓
AiR✓1.7770.6400.8490.9100.225-0.131
✓✓1.9170.7030.8580.7000.749-5.077
✓✓✓1.9420.7010.8580.7130.748-5.157
OSIE✓2.8350.6810.8850.8230.567-0.246
✓✓2.9040.7170.8970.7650.876-1.738
✓✓✓2.9650.7260.8950.6851.115-4.169
", + "image_path": "01aea2a12d57658fa9e1746fcdb29df7616a8946612fbcd8843845bd525f2cc3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 277, + 484, + 288 + ], + "lines": [ + { + "bbox": [ + 109, + 277, + 484, + 288 + ], + "spans": [ + { + "bbox": [ + 109, + 277, + 484, + 288 + ], + "type": "text", + "content": "Table 3. Ablation study on the impacts of Baseline, Reasoning, and Contextual Prioritization (ContPri)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 310, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 357 + ], + "type": "text", + "content": "by the significant improvements in interpretability metrics – the AUC-E drops to 0.700 and the AOPC increases to 0.749, suggesting our model's ability to provide more coherent explanations for task-driven saliency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 358, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 550 + ], + "type": "text", + "content": "Further, the contextual prioritization mechanism exhibits a more pronounced impact, especially in the OSIE task-free setting. For example, on OSIE, the NSS increases to 2.965, the CC improves to 0.726, and the AUC remains relatively stable at 0.895. This suggests that prioritizing the most salient elements plays a crucial role in handling complex scenes with multiple potential salient objects, ultimately leading to more concise and clear attention maps. This is further evidenced by the significant improvements in interpretability metrics – the AUC-E decreases to 0.685, the AOPC increases to 1.115, and the LOdds decreases to -4.169. This greater impact on the OSIE dataset could be attributed to the presence of a large number of potential semantic prototypes in task-free settings, making selective focus particularly valuable for distilling relevant information and improving interpretability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "text", + "content": "These findings demonstrate that the combination of explicit reasoning and contextual prioritization mechanisms, implemented in our model, is essential for achieving superior performance and explainability in both task-driven and task-free saliency prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 622, + 165, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 165, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 165, + 635 + ], + "type": "text", + "content": "4.5. Qualitative Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Qualitative analysis of our model reveals its ability to effectively integrate task-driven guidance and contextual prioritization to generate accurate saliency predictions and explanations. As illustrated in Fig. 4, the model aligns its attention with the goals of each question, selectively focusing on essential objects while ignoring irrelevant elements. For ex" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 310, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 466 + ], + "type": "text", + "content": "ample, in questions about the position of a skateboard and a red cone (row 1), the model prioritizes these objects over traditionally salient regions like the skateboarder, demonstrating the advantage of explicit reasoning. The model also adapts well to abstract concepts, such as interpreting \"modern\" style in a question about a faucet (row 2). Finally, the model excels in context prioritization by isolating key elements in complex scenes, highlighting the person, bag, and cars in a question about cars (row 3) while filtering out other visual elements. These qualitative findings demonstrate the model's ability to produce explanations that are closely aligned with the model's internal reasoning, ensuring outputs that are both interpretable and reliable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": "Beyond its interpretability, our model shows significant performance advantages in accuracy and task relevance. In Fig. 5, our model achieves the highest accuracy among state-of-the-art methods, aligning closely with ground truth. For example, in the question \"What is the drink in the cup made of plastic called?\", it focuses on the plastic cup, unlike other models that are distracted by irrelevant food items (row 1), showcasing the strength of our explicit reasoning approach. Additionally, our model demonstrates strong context prioritization, effectively concentrating on relevant elements even in complex scenes with multiple distractions. For instance, in the question \"Is the fire hydrant to the right or to the left of the garbage can?\", our model isolates the hydrant and garbage can, while other models diffuse attention across the scene (row 2), highlighting its ability to prioritize contextually important elements. Furthermore, in dynamic scenes, our model reliably focuses on salient objects. In the question \"Are both the helmet and the bat the same color?\", it highlights the helmet and bat specifically, unlike other models that diffuse attention (row 3), demonstrating" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9607" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 83, + 122, + 120 + ], + "blocks": [ + { + "bbox": [ + 88, + 74, + 107, + 81 + ], + "lines": [ + { + "bbox": [ + 88, + 74, + 107, + 81 + ], + "spans": [ + { + "bbox": [ + 88, + 74, + 107, + 81 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 83, + 122, + 120 + ], + "lines": [ + { + "bbox": [ + 74, + 83, + 122, + 120 + ], + "spans": [ + { + "bbox": [ + 74, + 83, + 122, + 120 + ], + "type": "image", + "image_path": "25e90289825018cfd898e673aa37ddf6e2c7ba241d06207868706693dc8ec4b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 124, + 83, + 174, + 119 + ], + "blocks": [ + { + "bbox": [ + 128, + 74, + 169, + 81 + ], + "lines": [ + { + "bbox": [ + 128, + 74, + 169, + 81 + ], + "spans": [ + { + "bbox": [ + 128, + 74, + 169, + 81 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 83, + 174, + 119 + ], + "lines": [ + { + "bbox": [ + 124, + 83, + 174, + 119 + ], + "spans": [ + { + "bbox": [ + 124, + 83, + 174, + 119 + ], + "type": "image", + "image_path": "38e540eadb7bf36144315c8250cc891ea9f43e943773b62d2dea4a3125fc4d7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 176, + 83, + 224, + 119 + ], + "blocks": [ + { + "bbox": [ + 185, + 74, + 214, + 81 + ], + "lines": [ + { + "bbox": [ + 185, + 74, + 214, + 81 + ], + "spans": [ + { + "bbox": [ + 185, + 74, + 214, + 81 + ], + "type": "text", + "content": "Prediction" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 83, + 224, + 119 + ], + "lines": [ + { + "bbox": [ + 176, + 83, + 224, + 119 + ], + "spans": [ + { + "bbox": [ + 176, + 83, + 224, + 119 + ], + "type": "image", + "image_path": "225f8ddc761de2c3e62a6392ab736cf40885eb6ccd38ed749b2080a6ba68c021.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 272, + 83, + 322, + 120 + ], + "blocks": [ + { + "bbox": [ + 281, + 74, + 314, + 81 + ], + "lines": [ + { + "bbox": [ + 281, + 74, + 314, + 81 + ], + "spans": [ + { + "bbox": [ + 281, + 74, + 314, + 81 + ], + "type": "text", + "content": "Prototype 1" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 83, + 322, + 120 + ], + "lines": [ + { + "bbox": [ + 272, + 83, + 322, + 120 + ], + "spans": [ + { + "bbox": [ + 272, + 83, + 322, + 120 + ], + "type": "image", + "image_path": "84c1bc806cd7f4b63d1e486e192c211fd462c20982c68026a56e728178eaf74b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 121, + 319, + 129 + ], + "lines": [ + { + "bbox": [ + 272, + 121, + 319, + 129 + ], + "spans": [ + { + "bbox": [ + 272, + 121, + 319, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{B}\\mathbf{a}\\mathbf{g}(w_1 = 0.378)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 129, + 319, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 129, + 319, + 167 + ], + "spans": [ + { + "bbox": [ + 227, + 129, + 319, + 167 + ], + "type": "text", + "content": "The question specifies a bag, so it is relevant to locate its position in the image and use it as a reference point for determining the direction of 'to the right'." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 83, + 414, + 120 + ], + "blocks": [ + { + "bbox": [ + 372, + 74, + 406, + 81 + ], + "lines": [ + { + "bbox": [ + 372, + 74, + 406, + 81 + ], + "spans": [ + { + "bbox": [ + 372, + 74, + 406, + 81 + ], + "type": "text", + "content": "Prototype 2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 83, + 414, + 120 + ], + "lines": [ + { + "bbox": [ + 365, + 83, + 414, + 120 + ], + "spans": [ + { + "bbox": [ + 365, + 83, + 414, + 120 + ], + "type": "image", + "image_path": "1f38d6140f5c8ab2af3136b689b8a84370b2dbc79eb9edca1feee4652f9a3b24.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 121, + 411, + 129 + ], + "lines": [ + { + "bbox": [ + 356, + 121, + 411, + 129 + ], + "spans": [ + { + "bbox": [ + 356, + 121, + 411, + 129 + ], + "type": "text", + "content": "Person " + }, + { + "bbox": [ + 356, + 121, + 411, + 129 + ], + "type": "inline_equation", + "content": "(w_{2} = 0.330)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 129, + 412, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 129, + 412, + 160 + ], + "spans": [ + { + "bbox": [ + 323, + 129, + 412, + 160 + ], + "type": "text", + "content": "To identify the reference point for 'to the right' in relation to the bag and determine where cars are located relative to this person." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 466, + 83, + 515, + 120 + ], + "blocks": [ + { + "bbox": [ + 474, + 74, + 508, + 81 + ], + "lines": [ + { + "bbox": [ + 474, + 74, + 508, + 81 + ], + "spans": [ + { + "bbox": [ + 474, + 74, + 508, + 81 + ], + "type": "text", + "content": "Prototype 3" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 83, + 515, + 120 + ], + "lines": [ + { + "bbox": [ + 466, + 83, + 515, + 120 + ], + "spans": [ + { + "bbox": [ + 466, + 83, + 515, + 120 + ], + "type": "image", + "image_path": "481dd8e1266e3d3564130bd7611455502be11aafc81faea7241bfe0db5c111ea.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 121, + 514, + 129 + ], + "lines": [ + { + "bbox": [ + 463, + 121, + 514, + 129 + ], + "spans": [ + { + "bbox": [ + 463, + 121, + 514, + 129 + ], + "type": "text", + "content": "Cars " + }, + { + "bbox": [ + 463, + 121, + 514, + 129 + ], + "type": "inline_equation", + "content": "(w_{3} = 0.291)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 432, + 129, + 514, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 129, + 514, + 160 + ], + "spans": [ + { + "bbox": [ + 432, + 129, + 514, + 160 + ], + "type": "text", + "content": "Cars are mentioned in the question and need to be identified in their location relative to the person and bag." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 121, + 118, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 118, + 167 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 118, + 167 + ], + "type": "text", + "content": "Q: Are there any cars to the right of the person that is to the right of the bag?" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 74, + 169, + 122, + 206 + ], + "blocks": [ + { + "bbox": [ + 74, + 169, + 122, + 206 + ], + "lines": [ + { + "bbox": [ + 74, + 169, + 122, + 206 + ], + "spans": [ + { + "bbox": [ + 74, + 169, + 122, + 206 + ], + "type": "image", + "image_path": "7c6a07fc5f102915ce0553d72bbb60219b49817fd259f19f8d7a5b1b19c49fde.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 124, + 169, + 173, + 205 + ], + "blocks": [ + { + "bbox": [ + 124, + 169, + 173, + 205 + ], + "lines": [ + { + "bbox": [ + 124, + 169, + 173, + 205 + ], + "spans": [ + { + "bbox": [ + 124, + 169, + 173, + 205 + ], + "type": "image", + "image_path": "35844a742d6a68031532d5eaa8f3db51734d31fc7f525bb354163bd6931f77e4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 176, + 169, + 224, + 205 + ], + "blocks": [ + { + "bbox": [ + 176, + 169, + 224, + 205 + ], + "lines": [ + { + "bbox": [ + 176, + 169, + 224, + 205 + ], + "spans": [ + { + "bbox": [ + 176, + 169, + 224, + 205 + ], + "type": "image", + "image_path": "c01f2fb14723aa66248dc1b97865973157f76fe7349766966c28685229b1c61e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 75, + 207, + 120, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 207, + 120, + 245 + ], + "spans": [ + { + "bbox": [ + 75, + 207, + 120, + 245 + ], + "type": "text", + "content": "Q: Does the faucet to the left of the vase look modern and silver?" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 272, + 169, + 321, + 206 + ], + "blocks": [ + { + "bbox": [ + 272, + 169, + 321, + 206 + ], + "lines": [ + { + "bbox": [ + 272, + 169, + 321, + 206 + ], + "spans": [ + { + "bbox": [ + 272, + 169, + 321, + 206 + ], + "type": "image", + "image_path": "3f8e80c2a4d00e6ac5f1a5cff738bc20b70738db238aaef42aed6535b095a643.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 207, + 319, + 215 + ], + "lines": [ + { + "bbox": [ + 270, + 207, + 319, + 215 + ], + "spans": [ + { + "bbox": [ + 270, + 207, + 319, + 215 + ], + "type": "text", + "content": "Vase " + }, + { + "bbox": [ + 270, + 207, + 319, + 215 + ], + "type": "inline_equation", + "content": "(w_{1} = 0.357)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 228, + 215, + 319, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 215, + 319, + 239 + ], + "spans": [ + { + "bbox": [ + 228, + 215, + 319, + 239 + ], + "type": "text", + "content": "The vase is mentioned in the question and serves as a reference point for the faucet's location." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 365, + 169, + 414, + 206 + ], + "blocks": [ + { + "bbox": [ + 365, + 169, + 414, + 206 + ], + "lines": [ + { + "bbox": [ + 365, + 169, + 414, + 206 + ], + "spans": [ + { + "bbox": [ + 365, + 169, + 414, + 206 + ], + "type": "image", + "image_path": "cfa26a18d07460e3f87cf10c9722e5d224670706158e2546bdb85409fa49e108.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 207, + 411, + 214 + ], + "lines": [ + { + "bbox": [ + 354, + 207, + 411, + 214 + ], + "spans": [ + { + "bbox": [ + 354, + 207, + 411, + 214 + ], + "type": "text", + "content": "Faucets " + }, + { + "bbox": [ + 354, + 207, + 411, + 214 + ], + "type": "inline_equation", + "content": "(w_{2} = 0.319)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 324, + 215, + 411, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 215, + 411, + 246 + ], + "spans": [ + { + "bbox": [ + 324, + 215, + 411, + 246 + ], + "type": "text", + "content": "The question specifically asks about the faucet's appearance, so it was a key element in answering." + } + ] + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 466, + 169, + 516, + 206 + ], + "blocks": [ + { + "bbox": [ + 466, + 169, + 516, + 206 + ], + "lines": [ + { + "bbox": [ + 466, + 169, + 516, + 206 + ], + "spans": [ + { + "bbox": [ + 466, + 169, + 516, + 206 + ], + "type": "image", + "image_path": "73418ff3002998be5af46fd654c9e29e9090b747eda550ab2292fe827469eecf.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 207, + 514, + 214 + ], + "lines": [ + { + "bbox": [ + 435, + 207, + 514, + 214 + ], + "spans": [ + { + "bbox": [ + 435, + 207, + 514, + 214 + ], + "type": "text", + "content": "Modern design " + }, + { + "bbox": [ + 435, + 207, + 514, + 214 + ], + "type": "inline_equation", + "content": "(w_{3} = 0.329)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 425, + 215, + 514, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 215, + 514, + 237 + ], + "spans": [ + { + "bbox": [ + 425, + 215, + 514, + 237 + ], + "type": "text", + "content": "The question refers to the faucets style, which can be described as modern." + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 74, + 247, + 122, + 284 + ], + "blocks": [ + { + "bbox": [ + 74, + 247, + 122, + 284 + ], + "lines": [ + { + "bbox": [ + 74, + 247, + 122, + 284 + ], + "spans": [ + { + "bbox": [ + 74, + 247, + 122, + 284 + ], + "type": "image", + "image_path": "f9a5f490c053d17f027d35927a1e85c927106887d14d56fa75b9da74e4f451cb.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 332, + 542, + 344 + ], + "lines": [ + { + "bbox": [ + 50, + 332, + 542, + 344 + ], + "spans": [ + { + "bbox": [ + 50, + 332, + 542, + 344 + ], + "type": "text", + "content": "Figure 4. Qualitative examples from our methods. Each semantic prototype is shown along with their saliency weights and explanations." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 124, + 247, + 173, + 284 + ], + "blocks": [ + { + "bbox": [ + 124, + 247, + 173, + 284 + ], + "lines": [ + { + "bbox": [ + 124, + 247, + 173, + 284 + ], + "spans": [ + { + "bbox": [ + 124, + 247, + 173, + 284 + ], + "type": "image", + "image_path": "4e9996dc591349d633107f6234428d5fc005a96de2d0d05bde846f4c5842d3ca.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 176, + 247, + 224, + 284 + ], + "blocks": [ + { + "bbox": [ + 176, + 247, + 224, + 284 + ], + "lines": [ + { + "bbox": [ + 176, + 247, + 224, + 284 + ], + "spans": [ + { + "bbox": [ + 176, + 247, + 224, + 284 + ], + "type": "image", + "image_path": "f51a130be62d71c429bb45d861dc8933dba5f406f22f614ab9a56030c973c35e.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 272, + 247, + 321, + 284 + ], + "blocks": [ + { + "bbox": [ + 272, + 247, + 321, + 284 + ], + "lines": [ + { + "bbox": [ + 272, + 247, + 321, + 284 + ], + "spans": [ + { + "bbox": [ + 272, + 247, + 321, + 284 + ], + "type": "image", + "image_path": "5ae8d40964d00dd12a4c791ee3ae85d2a2589bdc384616b6872fd50eb1f71313.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 285, + 318, + 293 + ], + "lines": [ + { + "bbox": [ + 250, + 285, + 318, + 293 + ], + "spans": [ + { + "bbox": [ + 250, + 285, + 318, + 293 + ], + "type": "text", + "content": "Skateboard " + }, + { + "bbox": [ + 250, + 285, + 318, + 293 + ], + "type": "inline_equation", + "content": "(w_{1} = 0.371)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 227, + 293, + 319, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 293, + 319, + 316 + ], + "spans": [ + { + "bbox": [ + 227, + 293, + 319, + 316 + ], + "type": "text", + "content": "The skateboard is the object being asked about in terms of its position relative to the red cone." + } + ] + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 365, + 247, + 414, + 284 + ], + "blocks": [ + { + "bbox": [ + 365, + 247, + 414, + 284 + ], + "lines": [ + { + "bbox": [ + 365, + 247, + 414, + 284 + ], + "spans": [ + { + "bbox": [ + 365, + 247, + 414, + 284 + ], + "type": "image", + "image_path": "e03a9a852250e436a2c56ec46ddeb3a5baa887a71598426a9a6974ac475c9ad8.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 350, + 285, + 411, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 285, + 411, + 325 + ], + "spans": [ + { + "bbox": [ + 350, + 285, + 411, + 325 + ], + "type": "text", + "content": "Red cone " + }, + { + "bbox": [ + 350, + 285, + 411, + 325 + ], + "type": "inline_equation", + "content": "(w_{2} = 0.344)" + }, + { + "bbox": [ + 350, + 285, + 411, + 325 + ], + "type": "text", + "content": " red cone is the reference point for determining the position of the skateboard relative to the question." + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 466, + 247, + 515, + 284 + ], + "blocks": [ + { + "bbox": [ + 466, + 247, + 515, + 284 + ], + "lines": [ + { + "bbox": [ + 466, + 247, + 515, + 284 + ], + "spans": [ + { + "bbox": [ + 466, + 247, + 515, + 284 + ], + "type": "image", + "image_path": "d32b077db8140c35298b7196607f4e2f92b2363dd10baba7698c8f6cef0bda9f.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 285, + 514, + 293 + ], + "lines": [ + { + "bbox": [ + 438, + 285, + 514, + 293 + ], + "spans": [ + { + "bbox": [ + 438, + 285, + 514, + 293 + ], + "type": "text", + "content": "Skateboarder " + }, + { + "bbox": [ + 438, + 285, + 514, + 293 + ], + "type": "inline_equation", + "content": "(w_{3} = 0.283)" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 418, + 293, + 514, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 293, + 514, + 324 + ], + "spans": [ + { + "bbox": [ + 418, + 293, + 514, + 324 + ], + "type": "text", + "content": "The skateboarder position and orientation are crucial in understanding where the skateboard is in relation to the red cone." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 76, + 370, + 122, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 370, + 122, + 402 + ], + "spans": [ + { + "bbox": [ + 76, + 370, + 122, + 402 + ], + "type": "text", + "content": "Q: What is the drink in the cup made of plastic called?" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 76, + 417, + 132, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 417, + 132, + 451 + ], + "spans": [ + { + "bbox": [ + 76, + 417, + 132, + 451 + ], + "type": "text", + "content": "Q: Is the fire hydrant to the right or to the left of the garbage can?" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 77, + 464, + 129, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 464, + 129, + 489 + ], + "spans": [ + { + "bbox": [ + 77, + 464, + 129, + 489 + ], + "type": "text", + "content": "Q: Are both the helmet and the bat the same color?" + } + ] + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 137, + 369, + 200, + 415 + ], + "blocks": [ + { + "bbox": [ + 157, + 359, + 179, + 367 + ], + "lines": [ + { + "bbox": [ + 157, + 359, + 179, + 367 + ], + "spans": [ + { + "bbox": [ + 157, + 359, + 179, + 367 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 369, + 200, + 415 + ], + "lines": [ + { + "bbox": [ + 137, + 369, + 200, + 415 + ], + "spans": [ + { + "bbox": [ + 137, + 369, + 200, + 415 + ], + "type": "image", + "image_path": "43436310f7f66fd499635549f286c6794cd96306c2a0f62fd1d30b9df64910a2.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 137, + 415, + 200, + 461 + ], + "blocks": [ + { + "bbox": [ + 137, + 415, + 200, + 461 + ], + "lines": [ + { + "bbox": [ + 137, + 415, + 200, + 461 + ], + "spans": [ + { + "bbox": [ + 137, + 415, + 200, + 461 + ], + "type": "image", + "image_path": "7ebe140a887c7385c263615a87b8b035762f39e2a3d09592ed45c81032bf215d.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 137, + 463, + 200, + 510 + ], + "blocks": [ + { + "bbox": [ + 137, + 463, + 200, + 510 + ], + "lines": [ + { + "bbox": [ + 137, + 463, + 200, + 510 + ], + "spans": [ + { + "bbox": [ + 137, + 463, + 200, + 510 + ], + "type": "image", + "image_path": "7defecdca0b81525a3aa8753daa761923d9cc49d9538aa2abfff39a2ae4d18b4.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 519, + 545, + 542 + ], + "lines": [ + { + "bbox": [ + 46, + 519, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 545, + 542 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of saliency predictions across models (XSal, DINet, SALICON, SAM). Each row shows an input question and image, with the corresponding ground truth attention map and model outputs." + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 202, + 369, + 264, + 415 + ], + "blocks": [ + { + "bbox": [ + 209, + 359, + 257, + 367 + ], + "lines": [ + { + "bbox": [ + 209, + 359, + 257, + 367 + ], + "spans": [ + { + "bbox": [ + 209, + 359, + 257, + 367 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 369, + 264, + 415 + ], + "lines": [ + { + "bbox": [ + 202, + 369, + 264, + 415 + ], + "spans": [ + { + "bbox": [ + 202, + 369, + 264, + 415 + ], + "type": "image", + "image_path": "393c224a27810f79eef0c9f6866d0cc7f158f608acec605dd7db468893378402.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 202, + 415, + 264, + 461 + ], + "blocks": [ + { + "bbox": [ + 202, + 415, + 264, + 461 + ], + "lines": [ + { + "bbox": [ + 202, + 415, + 264, + 461 + ], + "spans": [ + { + "bbox": [ + 202, + 415, + 264, + 461 + ], + "type": "image", + "image_path": "9cd7c79544fa4b95386325409697857c67173c1b9c00ffcf9732a2e5d1fa5fdd.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 202, + 463, + 264, + 510 + ], + "blocks": [ + { + "bbox": [ + 202, + 463, + 264, + 510 + ], + "lines": [ + { + "bbox": [ + 202, + 463, + 264, + 510 + ], + "spans": [ + { + "bbox": [ + 202, + 463, + 264, + 510 + ], + "type": "image", + "image_path": "647d6486bc775446863a22177768988618e0e8271d5f3cac66345ca48d287d93.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 265, + 369, + 328, + 415 + ], + "blocks": [ + { + "bbox": [ + 288, + 359, + 306, + 367 + ], + "lines": [ + { + "bbox": [ + 288, + 359, + 306, + 367 + ], + "spans": [ + { + "bbox": [ + 288, + 359, + 306, + 367 + ], + "type": "text", + "content": "XSal" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 265, + 369, + 328, + 415 + ], + "lines": [ + { + "bbox": [ + 265, + 369, + 328, + 415 + ], + "spans": [ + { + "bbox": [ + 265, + 369, + 328, + 415 + ], + "type": "image", + "image_path": "acd996e56cfd054df6d94413eca7c2d58e3bcc16ec18890d0d7a909263870ea0.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 266, + 415, + 328, + 461 + ], + "blocks": [ + { + "bbox": [ + 266, + 415, + 328, + 461 + ], + "lines": [ + { + "bbox": [ + 266, + 415, + 328, + 461 + ], + "spans": [ + { + "bbox": [ + 266, + 415, + 328, + 461 + ], + "type": "image", + "image_path": "567b83f83a3be5a57484d6d3b72d975a82fb4ed93db6394b88d5e2dd0a0b1ff9.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 266, + 463, + 328, + 510 + ], + "blocks": [ + { + "bbox": [ + 266, + 463, + 328, + 510 + ], + "lines": [ + { + "bbox": [ + 266, + 463, + 328, + 510 + ], + "spans": [ + { + "bbox": [ + 266, + 463, + 328, + 510 + ], + "type": "image", + "image_path": "bea1ecdd7b565fbe3d857a8d75e4e93094c7be2e06bd607b021ad988b4f0145f.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 329, + 369, + 392, + 415 + ], + "blocks": [ + { + "bbox": [ + 350, + 359, + 372, + 367 + ], + "lines": [ + { + "bbox": [ + 350, + 359, + 372, + 367 + ], + "spans": [ + { + "bbox": [ + 350, + 359, + 372, + 367 + ], + "type": "text", + "content": "DINet" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 369, + 392, + 415 + ], + "lines": [ + { + "bbox": [ + 329, + 369, + 392, + 415 + ], + "spans": [ + { + "bbox": [ + 329, + 369, + 392, + 415 + ], + "type": "image", + "image_path": "9cfa01521c156af9fe089fb56433eb3a05a687386c3a95660e3b7a020cf6619c.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 329, + 415, + 392, + 461 + ], + "blocks": [ + { + "bbox": [ + 329, + 415, + 392, + 461 + ], + "lines": [ + { + "bbox": [ + 329, + 415, + 392, + 461 + ], + "spans": [ + { + "bbox": [ + 329, + 415, + 392, + 461 + ], + "type": "image", + "image_path": "da89bbf62cbc4c925612f7e68cfab91804fe08a3ff6018c72e461d86e2c9fbd5.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 329, + 462, + 392, + 510 + ], + "blocks": [ + { + "bbox": [ + 329, + 462, + 392, + 510 + ], + "lines": [ + { + "bbox": [ + 329, + 462, + 392, + 510 + ], + "spans": [ + { + "bbox": [ + 329, + 462, + 392, + 510 + ], + "type": "image", + "image_path": "380397bdde725dadd0d89fa07d72d1677a73186cac69984515e9c84c33ffb081.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 392, + 369, + 455, + 415 + ], + "blocks": [ + { + "bbox": [ + 406, + 359, + 443, + 367 + ], + "lines": [ + { + "bbox": [ + 406, + 359, + 443, + 367 + ], + "spans": [ + { + "bbox": [ + 406, + 359, + 443, + 367 + ], + "type": "text", + "content": "SALICON" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 369, + 455, + 415 + ], + "lines": [ + { + "bbox": [ + 392, + 369, + 455, + 415 + ], + "spans": [ + { + "bbox": [ + 392, + 369, + 455, + 415 + ], + "type": "image", + "image_path": "3645347d2168522f59ac6de9d09ed894b51a33a810691b007f0f82b3cef13605.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 393, + 415, + 455, + 461 + ], + "blocks": [ + { + "bbox": [ + 393, + 415, + 455, + 461 + ], + "lines": [ + { + "bbox": [ + 393, + 415, + 455, + 461 + ], + "spans": [ + { + "bbox": [ + 393, + 415, + 455, + 461 + ], + "type": "image", + "image_path": "c01d014c3b0f577e5d13bae3ca7346cffb1b845adbc8368694eb8743df0f1ae4.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 393, + 462, + 455, + 510 + ], + "blocks": [ + { + "bbox": [ + 393, + 462, + 455, + 510 + ], + "lines": [ + { + "bbox": [ + 393, + 462, + 455, + 510 + ], + "spans": [ + { + "bbox": [ + 393, + 462, + 455, + 510 + ], + "type": "image", + "image_path": "cdcd09de8683348dadcc95ad5ff4790c51e5e8d716cce47cd8d969dbd96b67e5.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 457, + 369, + 520, + 415 + ], + "blocks": [ + { + "bbox": [ + 479, + 359, + 498, + 367 + ], + "lines": [ + { + "bbox": [ + 479, + 359, + 498, + 367 + ], + "spans": [ + { + "bbox": [ + 479, + 359, + 498, + 367 + ], + "type": "text", + "content": "SAM" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 369, + 520, + 415 + ], + "lines": [ + { + "bbox": [ + 457, + 369, + 520, + 415 + ], + "spans": [ + { + "bbox": [ + 457, + 369, + 520, + 415 + ], + "type": "image", + "image_path": "5f074d7b10c467ab73ca4737ebe44ad4f427a0567c03e7e14b214ba3817b0984.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 457, + 415, + 520, + 461 + ], + "blocks": [ + { + "bbox": [ + 457, + 415, + 520, + 461 + ], + "lines": [ + { + "bbox": [ + 457, + 415, + 520, + 461 + ], + "spans": [ + { + "bbox": [ + 457, + 415, + 520, + 461 + ], + "type": "image", + "image_path": "ba6375fd8a3649f61e058544dfb38db17c80f92316777a8aad0c64c4400f8a73.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 457, + 462, + 520, + 510 + ], + "blocks": [ + { + "bbox": [ + 457, + 462, + 520, + 510 + ], + "lines": [ + { + "bbox": [ + 457, + 462, + 520, + 510 + ], + "spans": [ + { + "bbox": [ + 457, + 462, + 520, + 510 + ], + "type": "image", + "image_path": "ec76b46edcd49346e4109185e1b90ff4d8888ddfe40cbfe8334064739b556055.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "bbox": [ + 46, + 563, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 599 + ], + "type": "text", + "content": "its adaptability to diverse scenarios. These examples highlight our model's superior performance in generating accurate and contextually relevant saliency predictions." + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "This paper has presented a novel approach for saliency prediction that combines explicit reasoning and contextual prioritization, generating both accurate and explainable saliency predictions. We have demonstrated that our method, powered by a VLM, outperforms existing models in terms of accuracy and explainability. This is achieved by leveraging an explicit reasoning process that allows the" + } + ] + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 563, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 547, + 658 + ], + "type": "text", + "content": "model to provide clear explanations for its predictions and a contextual prioritization mechanism that focuses on a few truly important elements, mimicking human attention behavior. Our approach not only achieves state-of-the-art saliency prediction accuracy but also offers valuable insights into the decision-making processes of saliency models. This advancement contributes to the development of more trustworthy and reliable saliency models." + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 306, + 670, + 403, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 403, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 403, + 683 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by NSF Grant 2143197 and Fasikl Incorporated." + } + ] + } + ], + "index": 77 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9608" + } + ] + } + ], + "index": 78 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "text", + "content": "[2] Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 258 + ], + "type": "text", + "content": "[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 287, + 293 + ], + "type": "text", + "content": "[4] Patrick Le Callet and Ernst Niebur. Visual attention and applications in multimedia technologies. Proceedings of the Institution of Electrical Engineers, 2013. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 337 + ], + "type": "text", + "content": "[5] Moran Cerf, E Paxon Frady, and Christof Koch. Faces and text attract gaze independent of the task: Experimental data and computer model. Journal of vision, 9(12):10-10, 2009. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 393 + ], + "type": "text", + "content": "[6] Hanjie Chen, Guangtao Zheng, and Yangfeng Ji. Generating hierarchical explanations on text classification via feature interaction detection. In Proceedings of the 2020 Conference on Association for Computational Linguistics (ACL), 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 462 + ], + "type": "text", + "content": "[7] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 517 + ], + "type": "text", + "content": "[8] Nuo Chen, Jin Xie, Jing Nie, Jiale Cao, Zhuang Shao, and Yanwei Pang. Attentive alignment network for multispectral pedestrian detection. In Proceedings of the 31st ACM international conference on multimedia, pages 3787-3795, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 520, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 520, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 520, + 287, + 553 + ], + "type": "text", + "content": "[9] Shi Chen, Ming Jiang, Jinhui Yang, and Qi Zhao. AiR: Attention with reasoning capability. In European Conference on Computer Vision (ECCV), 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "text", + "content": "[10] Shi Chen, Ming Jiang, and Qi Zhao. What do deep saliency models learn about visual attention? In Advances in Neural Information Processing Systems, pages 9543-9555. Curran Associates, Inc., 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[11] Shi Chen, Ming Jiang, and Qi Zhao. Deep learning to interpret autism spectrum disorder behind the camera. IEEE Transactions on Cognitive and Developmental Systems, pages 1-12, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[12] Xianyu Chen, Ming Jiang, and Qi Zhao. Predicting human scanpaths in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10876-10885, 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[13] Xianyu Chen, Ming Jiang, and Qi Zhao. Leveraging human attention in novel object captioning. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "International Joint Conference on Artificial Intelligence (IJ-CAI), 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 138 + ], + "type": "text", + "content": "[14] Xianyu Chen, Ming Jiang, and Qi Zhao. Self-distillation for few-shot image captioning. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "type": "text", + "content": "[15] Xianyu Chen, Ming Jiang, and Qi Zhao. Gazexplain: Learning to predict natural language explanations of visual scanpaths. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "type": "text", + "content": "[16] Xianyu Chen, Ming Jiang, and Qi Zhao. Beyond average: Individualized visual scanpath prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 282 + ], + "type": "text", + "content": "[17] Xianyu Chen, Jinhui Yang, Shi Chen, Louis Wang, Ming Jiang, and Qi Zhao. Every problem, every step, all in focus: Learning to solve real-world problems with integrated attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI), 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 283, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 547, + 326 + ], + "type": "text", + "content": "[18] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting Human Eye Fixations via an LSTM-based Saliency Attentive Model. TIP, 27(10):5142-5154, 2018. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 547, + 392 + ], + "type": "text", + "content": "[19] Chiara Di Bonaventura, Lucia Siciliani, Pierpaolo Basile, Albert Merono Penuela, and Barbara McGillivray. Is explanation all you need? an expert survey on llm-generated explanations for abusive language detection. In Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 547, + 437 + ], + "type": "text", + "content": "[20] Ning Ding, Ce Zhang, and Azim Eskandarian. Saliendet: A saliency-based feature enhancement algorithm for object detection for autonomous driving. IEEE Transactions on Intelligent Vehicles, 9(1):2624-2635, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 547, + 491 + ], + "type": "text", + "content": "[21] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Zhaohui Che, Yi Fang, Xiaokang Yang, Jesús Gutierrez, and Patrick Le Callet. A dataset of eye movements for the children with autism spectrum disorder. In ACM Multimedia Systems Conference (MMSys), 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 492, + 547, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 547, + 535 + ], + "type": "text", + "content": "[22] Yang Feng, Lin Ma, Wei Liu, and Jiebo Luo. Unsupervised image captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 536, + 547, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 547, + 569 + ], + "type": "text", + "content": "[23] Songhao Han, Le Zhuo, Yue Liao, and Si Liu. Llms as visual explainers: Advancing image classification with evolving visual descriptions. arXiv preprint arXiv:2311.11904, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 547, + 602 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 547, + 645 + ], + "type": "text", + "content": "[25] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "text", + "content": "[26] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In ICCV, pages 262-270, 2015. 1, 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "text", + "content": "[27] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9609" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "on Pattern Analysis and Machine Intelligence, 20(11):1254-1259, 1998. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[28] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Silicon: Saliency in context. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[29] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In 2009 IEEE 12th international conference on computer vision, pages 2106-2113. IEEE, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "text", + "content": "[30] Christopher Kanan, Mathew H Tong, Lingyun Zhang, and Garrison W Cottrell. Sun: Top-down saliency using natural statistics. Visual cognition, 17(6-7):979-1003, 2009. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "text", + "content": "[31] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Image captioning with very scarce supervised data: Adversarial semi-supervised learning approach. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 275, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 297 + ], + "type": "text", + "content": "[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 319 + ], + "type": "text", + "content": "[33] Solomon Kullback. Information Theory and Statistics. Wiley, 1959. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "text", + "content": "[34] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. In ICLR workshop, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "text", + "content": "[35] Iro Laina, Christian Rupprecht, and Nassir Navab. Towards unsupervised image captioning with shared multimodal embeddings. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "text", + "content": "[36] Jianxun Lou, Hanhe Lin, David Marshall, Dietmar Saupe, and Hantao Liu. Transalnet: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[37] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7086-7096, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "type": "text", + "content": "[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[39] Olivier Le Meur, Patrick Le Callet, and Dominique Barba. Predicting visual fixations on video based on low-level visual features. Vision Research, 47(19):2483 - 2498, 2007. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "text", + "content": "[40] Dong Nguyen. Comparing automatic and human evaluation of local explanations for text classification. In Proceedings of the 2018 Conference on North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "text", + "content": "[41] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397 - 2416, 2005. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[42] Thammathip Piumsomboon, Gun Lee, Robert W. Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In IEEE Symposium on 3D User Interfaces (3DUI), 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[43] Yao Qiang, Deng Pan, Chengyin Li, Rhongho Jang, and Dongxiao Zhu. Attcat: Explaining transformers via attentive" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 610 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "class activation tokens. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[44] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning (ICML), 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[45] Vincent Sitzmann, Ana Serrano, Amy Pavel, Maneesh Agrawala, Diego Gutierrez, Belen Masia, and Gordon Wetzstein. Saliency in vr: How do people explore virtual environments? IEEE Transactions on Visualization and Computer Graphics, 24(4):1633-1642, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[46] Yubao Sun, Mengyang Zhao, Kai Hu, and Shaojing Fan. Visual saliency prediction using multi-scale attention gated network. Multimedia Systems, 28(1):131-139, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "type": "text", + "content": "[47] Hamed R. Tavakoli, Rakshith Shetty, Ali Borji, and Jorma Laaksonen. Paying attention to descriptions generated by image captioning models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 286, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 351 + ], + "type": "text", + "content": "[48] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question answering. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 451-466. Springer, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 353, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 385 + ], + "type": "text", + "content": "[49] Juan Xu, Ming Jiang, Shuo Wang, Mohan S Kankanhalli, and Qi Zhao. Predicting human gaze beyond pixels. Journal of vision, 14(1):28-28, 2014. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 387, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 452 + ], + "type": "text", + "content": "[50] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "text", + "content": "[51] Jinhui Yang, Xianyu Chen, Ming Jiang, Shi Chen, Louis Wang, and Qi Zhao. VisualHow: Multimodal problem solving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 498, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 530 + ], + "type": "text", + "content": "[52] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. TMM, 22(8):2163-2176, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "type": "text", + "content": "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "text", + "content": "[54] Wei Zhang and Hantao Liu. Study of saliency in objective video quality assessment. IEEE Transactions on Image Processing, 26(3):1275-1288, 2017. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9610" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_content_list.json b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5b38e788a360a41fdace55770c79df1c242f1cc --- /dev/null +++ b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_content_list.json @@ -0,0 +1,1695 @@ +[ + { + "type": "text", + "text": "Explaining Domain Shifts in Language: Concept erasing for Interpretable Image Classification", + "text_level": 1, + "bbox": [ + 196, + 130, + 802, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zequn Zeng $^{1*}$ , Yudi Su $^{1*}$ , Jianqiao Sun $^{1}$ , Tiansheng Wen $^{1}$ , Hao Zhang $^{1}$ , Zhengjue Wang $^{2}$ , Bo Chen $^{1\\dagger}$ , Hongwei Liu $^{1\\dagger}$ and Jiawei Ma $^{3}$", + "bbox": [ + 210, + 202, + 785, + 238 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China.", + "$^{2}$ State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China." + ], + "bbox": [ + 117, + 239, + 877, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ City University of Hong Kong, Hong Kong SAR.", + "bbox": [ + 299, + 273, + 696, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zqzeng_1@stu.xidian.edu.cn, bchen@mail.xidian.edu.cn", + "bbox": [ + 261, + 295, + 718, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concept-based models can map black-box representations to human-understandable concepts, which makes the decision-making process more transparent and then allows users to understand the reason behind predictions. However, domain-specific concepts often impact the final predictions, which subsequently undermine the model generalization capabilities, and prevent the model from being used in high-stake applications. In this paper, we propose a novel Language-guided Concept-Erasing (LanCE) framework. In particular, we empirically demonstrate that pre-trained vision-language models (VLMs) can approximate distinct visual domain shifts via domain descriptors while prompting large Language Models (LLMs) can easily simulate a wide range of descriptors of unseen visual domains. Then, we introduce a novel plug-in domain descriptor orthogonality (DDO) regularizer to mitigate the impact of these domain-specific concepts on the final predictions. Notably, the DDO regularizer is agnostic to the design of concept-based models and we integrate it into several prevailing models. Through evaluation of domain generalization on four standard benchmarks and three newly introduced benchmarks, we demonstrate that DDO can significantly improve the out-of-distribution (OOD) generalization over the previous state-of-the-art concept-based models. Our code is available at https://github.com/joeyz0z/LanCE.", + "bbox": [ + 89, + 377, + 483, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concept-based models [11, 27, 42] are prominent approaches for achieving model interpretability, which leverage human-understandable concepts to explain the black-box image representation. Specifically, these models first", + "bbox": [ + 89, + 811, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f5a04b89c74ccf53b4b64acc28fdc8f2ae3680c3508afbab689b7ae319a69ea.jpg", + "image_caption": [ + "Photo ( train )", + "Predict:Apple" + ], + "image_footnote": [], + "bbox": [ + 575, + 357, + 638, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f691512ec1b34f8b94bd8503f154cf3ec0b170d2576bdc3c206069f90951d98.jpg", + "image_caption": [ + "Confidence: $98\\%$" + ], + "image_footnote": [], + "bbox": [ + 557, + 436, + 630, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b1c74964fb0523b19cad0b2d40f0e0d9275c1c19f46ef15c278ebcdeab3fc6e.jpg", + "image_caption": [ + "Clipart ( test )", + "Predict:Apple" + ], + "image_footnote": [], + "bbox": [ + 679, + 357, + 745, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f76669c91a4b991a4dcaa88e97cdc032d91fdc96d4c7641a86fc10173edda871.jpg", + "image_caption": [ + "Confidence: $52\\%$", + "(a) Concept-based predictions on three domains." + ], + "image_footnote": [], + "bbox": [ + 665, + 436, + 740, + 472 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb5f2075e5070feb9afe5349637f0d76fc797739e2e3840c133b2c5abdf3b255.jpg", + "image_caption": [ + "Sketch (test)", + "Predict: Pear X" + ], + "image_footnote": [], + "bbox": [ + 782, + 357, + 851, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/337d7025f18ded1eb7875fa186073edfc564b2ee06d8bd1f3225b1225fe0e0be.jpg", + "image_caption": [ + "Confidence: $37\\%$" + ], + "image_footnote": [], + "bbox": [ + 772, + 436, + 848, + 472 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca7c2452e649077e4e30bf2dd73d205167f62ab5cc2266b873b3e044f0ef2ac4.jpg", + "image_caption": [ + "Concept: Red color", + "(b) Distribution shifts of domain-specific and domain-shared concepts.", + "Figure 1. Domain shifts in concept space. (a) Given the images of an apple in different visual domains, the prediction confidence of a concept-based model, trained on the photo domain, degrades due to the missing of concepts. (b) Distribution comparison of concept activation value (image-concept similarity computed via CLIP [26]) between photo domain and sketch domain, for two concepts, i.e., \"red color\" and \"round shape\", respectively. JS divergence indicates the distance between two distributions and tends to be larger for domain-specific concepts (e.g. \"red color\")." + ], + "image_footnote": [], + "bbox": [ + 516, + 498, + 707, + 621 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/085228f4148871c7fc09b0719f54b6fb53848a44bcccf3e3293627f950a3c67b.jpg", + "image_caption": [ + "Concept: Round shape" + ], + "image_footnote": [], + "bbox": [ + 712, + 498, + 903, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "map the image feature to a concept activation vector in a concept space (each dimension corresponds to an interpretable concept) and then use the concept activation vector to predict the final output. Recently, equipped with pretrained vision-language models (VLMs) [8, 26], concept-based models [41, 42] can obtain the concept activation value by calculating the similarity between image embed", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding authors", + "bbox": [ + 109, + 886, + 370, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9517", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dings and textual concept embeddings.", + "bbox": [ + 89, + 90, + 346, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, current concept-based models still struggle to handle domain shifts encountered during inference. The distinct distribution of visual concepts across different domains presents a significant generalization challenge for these models. As shown in Fig. 1a, the concept-based models trained on the photographic images usually excel at associating the discriminative visual clues – such as red color, waxy texture, and round shape – with specific classes like “apple”, akin to human beings. However, these models experience substantial performance degradation when applied to images in unseen visual domains, such as clipart or sketches, due to the absence of domain-specific texture and color concepts. These domain-specific concepts (e.g. “red color”) generally exhibit a larger distributional difference between visual domains than domain-shared concepts (e.g. “round shape”), as illustrated in Fig. 1b. Consequently, the concept-based models trained on a single domain are inevitably biased to associate those domain-specific concepts with the final predictions, thereby limiting their out-of-distribution (OOD) generalization capabilities.", + "bbox": [ + 89, + 109, + 485, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the issue of domain shifts in the concept space, a straightforward solution is to involve human experts in filtering out the domain-specific concepts and retaining only domain-shared concepts for classification tasks, which is nevertheless labor-intensive. Another possible solution is to diversify the training domain [1, 3] to mitigate the impact of the domain-specific concepts. However, it is impractical to cover all possible unseen domains during training exhaustively. Additionally, current domain generalization approaches [49] can not be directly applied to concept-based models due to the lack of interpretability of semantics of the learned representation [34]. Therefore, automating the identification and elimination of domain-specific concepts remains an under-explored area.", + "bbox": [ + 89, + 414, + 483, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a Language-guided ConceptErasing framework, namely LanCE, to enhance the OOD generalization capabilities of current concept-based models. Specifically, we empirically demonstrate that pretrained VLMs, such as CLIP, can interpret the domain shifts through language, i.e., domain shifts can be approximated by a set of domain descriptors generated by large language models (LLMs). To alleviate the biased association between final predictions and domain-specific concepts, we introduce a plug-in domain descriptor orthogonality (DDO) loss that reduces the fluctuation of concept activation caused by these language-guided domain shifts. We evaluate our method on seven benchmarks, including four common benchmarks (CUB-Painting [35], PACS [13], OfficeHome [31] and DomainNet [24]), and three new proposed benchmarks (AwA2-clipart, LADA-Sculpture and LADV3D) based on several existing visual classification datasets with concept annotations [39, 47]. These newly-collected", + "bbox": [ + 89, + 628, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datasets are gathered from the web and manually filtered. Extensive experiments demonstrate that the proposed approach can preserve the In-distribution (ID) classification capabilities and significantly improve the OOD generalization capabilities of prevailing concept-based models. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide a concept-level explanation for the domain shifts and empirically reveal that VLMs can represent visual domain shifts into language space.", + "- Based on this observation, we propose a plug-in domain descriptor orthogonality regularizer to learn a domain-shared concept-based model without changing the model architecture and increasing training data.", + "- We introduce three new concept-based OOD generalization benchmarks including more challenging scenarios such as natural photos $\\rightarrow$ 3D-renderers and real animals $\\rightarrow$ scluptures. Extensive experiments on four common and three new benchmarks demonstrate that our proposed method significantly improves the OOD generalization capabilities of current concept-based models." + ], + "bbox": [ + 511, + 181, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 407, + 653, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept-based models are one of the mainstream interpretable approaches that aim to understand models by associating black box features with meaningful concepts. Current concept-based models can broadly be divided into two categories, concept bottleneck models (CBMs) [11] which map the image features into an intermediate concept bottleneck layer whose neurons indicate concept activation values of pre-defined concepts, and concept activation vectors (CAVs) [9] which represent concepts as normal vectors of decision boundaries that distinguish positive and negative samples of a concept. Vanilla CBM requires fine-grained and precise concept annotation on the concept bottleneck layer and CAVs need positive and negative samples of each concept to learn meaningful concept representations. Many recent approaches are proposed to tackle the above concerns from two perspectives. On the one hand, many methods generate the concepts derived from ConceptNet [42], LLMs [21, 41] or multimodal datasets [28, 43] to automate the construction of the concept bottleneck layer. On the other hand, to remove the reliance on human-annotated concept labels, these methods deploy pre-trained VLMs such as CLIP to map concepts to text embeddings and compute image-concept similarity to serve as concept activation annotations. However, concepts of the above approaches are typically domain-sensitive and the corresponding concept activations have large distributional differences across various domains. They often fail to handle domain shifts when applied to unseen domains.", + "bbox": [ + 511, + 433, + 906, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain adaptation & generalization aims to handle the domain shifts between training and test data. In contrast to traditional deep learning based on in-distribution assump-", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9518", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/107be90093f88dd5cd39df0c8266f606815b36a9781746acb40b2ff80fe28d5d.jpg", + "image_caption": [ + "(a) Interpret the visual domain shifts into language." + ], + "image_footnote": [], + "bbox": [ + 98, + 94, + 575, + 299 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/20f8d5d28f97dde4756e151b9ae81b85edd82bfc39670c46ebdcb660375525d5.jpg", + "image_caption": [ + "A sketch of an apple A photo of an apple" + ], + "image_footnote": [], + "bbox": [ + 599, + 112, + 901, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ae7b0b6976a6ebfe07f1fed017f8f45baa696c0683682127b453bf8889497840.jpg", + "image_caption": [ + "A bronze sculpture of a cat A photo of a cat", + "(b) Discover the domain-specific concepts.", + "Figure 2. (a) We empirically demonstrate that the visual domain shift can be interpreted in language. For each class, we obtain the caption from the embedding difference of images from two domains. Then, by aggregating the captions across the classes, the keywords regarding the domain shift can be highlighted. At the same time, (b) the domain-specific concepts can be discovered from language descriptions of the different domains. In detail, they have higher similarities with the difference of domain-related class descriptions, i.e., textural domain shift, in the CLIP embedding space. More analyses are shown in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 596, + 219, + 898, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion, domain adaptation methods [4, 20] target to improve the performance on the target domain when only a few samples [14] or unlabeled data [40] from target domains are available. Instead, domain generalization methods [49] focused on more difficult scenarios in which no target domain data are available during the training phase. The setting of this paper belongs to the single domain generalization scenario, i.e. only one single domain is available during training [19, 36, 48]. While the domain generalization area is well studied, there is still researcher doubt about the interpretability of domain generalization methods [34]. Although disentanglement-based domain generalization methods [25, 33] decompose a feature into domain-shared and domain-specific parts, there is still a lack of deep understanding of the semantics of the learned features in domain generalization models. Our proposed method provides a concept-level explanation of domain shifts and takes steps to interpretable domain generation.", + "bbox": [ + 91, + 417, + 483, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Pretraining aims to bridge the gap between image and text representations. Compared with vision-only pretraining methods [5, 6, 17], vision-language pertaining like CLIP [26] achieved remarkable success and showed superior performance on some high-level vision understanding tasks [16, 44, 46] and multi-modality tasks [29, 38, 45]. Our empirical findings further validate that CLIP can interpret a variety of domain shifts into a set of domain descriptors.", + "bbox": [ + 89, + 690, + 483, + 827 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Empirical Observations", + "text_level": 1, + "bbox": [ + 89, + 844, + 316, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section demonstrates that the pre-trained VLMs can approximate the visual domain shifts via domain descrip", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tors. Meanwhile, these language-guided domain shifts can effectively distinguish the domain-specific concepts as illustrated in Fig. 2. It serves as the key insight for our proposed LanCE method.", + "bbox": [ + 511, + 417, + 906, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-text alignment. CLIP trained on multi-modal contrastive learning, can effectively bridge the gap between image representation and text representation. The alignment facilitates a new approach to manipulate images using language, applicable for various vision tasks, e.g. image editing [7, 23], style transfer [12, 37], data augmentation [2, 3, 18]. The CLIP image-text similarity can be computed as:", + "bbox": [ + 511, + 479, + 908, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L I P}} (I, T) = \\operatorname {s i m} \\left(E _ {I} (I), E _ {T} (T)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 611, + 903, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{I}$ and $E_{T}$ mean the image encoder and text encoder of CLIP, respectively. sim indicate the cosine similarity.", + "bbox": [ + 511, + 642, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interpreting the visual domain shifts. Inspired by previous works [2, 23], we assume that the visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space. To validate this hypothesis, we employ a CLIP-based zero-shot captioner [44] to translate the visual embedding differences into language, as illustrated in Fig. 2(a). Let $\\{I_c^{src}\\}$ and $\\{I_c^{tgt}\\}$ denote the images of the same class $c$ from the source domain and the target domain, respectively. The class-level visual domain gap $\\Delta d_{c}$ and corresponding description $t_c$ can be generated by solving the following maximization problem:", + "bbox": [ + 511, + 674, + 908, + 840 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {t _ {c}} \\mathcal {L} _ {\\text {C L I P}} \\left(\\Delta d _ {c}, t _ {c}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 854, + 903, + 877 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {d} _ {c} = \\mathbb {E} \\left(E _ {I} \\left(I _ {c} ^ {t g t}\\right)\\right) - \\mathbb {E} \\left(E _ {I} \\left(I _ {c} ^ {s r c}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 878, + 903, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9519", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7adb23ec0613b26e3cc1008ae85f6ff553500a5e384e47db3762bf23283ae603.jpg", + "image_caption": [ + "Figure 3. Overview of the LanCE. Blue part is the data flow of vanilla CLIP-CBMs (Sec. 4.1). To provide concept-level explanations, we first construct a human-written or LLMs-generated concept set $\\mathcal{C}$ and extract the concept embeddings via the frozen CLIP text encoder. Given an image, we can extract the image embeddings via the frozen CLIP image encoder. The concept activations are the cosine similarity between image embeddings and concept embeddings. A learnable linear layer $W_{F}$ is fitted on top of the concept activation vector and is responsible for predicting the final class and is optimized via cross-entropy loss. Yellow part is the data flow of our proposed DDO regularizer (Sec. 4.2). Similarly, we first construct a domain descriptor set (Sec. 4.3) to obtain the language-guided domain shifts and then simulate the domain-specific concept activations. To erase the effect of domain-specific concepts, the DDO regularizer encourages the orthogonality between the class-concept correlation matrix $W_{F}$ (i.e. the final linear weight) and domain-specific concept activation $\\widehat{\\pmb{a}_{\\mathrm{sp}}}$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2(a), we produce high-matching (CLIP-Score $>0.9$ ) descriptions $t_c$ that effectively interpret the class-level visual domain gap $\\Delta d_c$ between sketch images, sculpture images and photographic images. We observe that these class-level domain descriptors are typically descriptions of the style difference of the two domains, regarding class-level domain descriptors. To further get the global difference between the two domains, we aggregate all class-level domain descriptors and transfer them into word cloud format, where the word size is proportional to the word frequency in the generated domain description corpus. As a result, we empirically find that global visual domain descriptors $\\Delta d$ primarily consist of style words such as \"sketch and sculpture\" and background words e.g. \"white\".", + "bbox": [ + 88, + 469, + 480, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discover the domain-specific concepts. Since we have revealed that visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space, we can synthesize the visual domain shifts via domain descriptors. Subsequently, we can leverage these language-guided domain shifts to discover domain-specific concepts. Specifically, we compute the similarity between textual domain shift embeddings (e.g. \"A sketch of an apple\" - \"A photo of an apple\") and all candidate concept embeddings extracted by CLIP text encoder. The results in Fig. 2(b) show that domain-specific concepts like \"shiny, waxy surface texture\" generally exhibit a higher similarity score due to their substantial distribution shift between those two domains while those domain-shared concepts such as \"Leaf around", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the stem\" have relatively lower similarity score. Consequently, we can distinguish the domain-specific concepts via language-guided domain shifts.", + "bbox": [ + 511, + 469, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 526, + 604, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the generalization capabilities of prevailing concept-based models, as shown in Fig. 3, we propose a novel LanCE framework. In this section, we first formulate the problem and introduce Concept bottleneck models (CBMs), which can effectively map the black-box visual features to human-understandable concept space to interpret the final outputs (Sec. 4.1). However, domain-specific concepts will undermine the generalization capabilities. Then, to alleviate the biased association between domain-specific concepts and class predictions within the learnable linear classifier, we propose a domain descriptor orthogonality loss to improve the OOD classification performances of concept-based models (Sec. 4.2). Furthermore, considering all possible visual domains, we prompt LLMs to generate a bunch of domain descriptors to simulate numerous unseen visual domains (Sec. 4.3).", + "bbox": [ + 509, + 551, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 801, + 718, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a training domain data $\\mathcal{D}_{train} = \\{(\\pmb{x}_i, y_i)\\}$ , where $\\pmb{x}_i \\in \\mathcal{X}$ are training images and $y_i \\in \\mathcal{Y}$ are corresponding labels, the concept-based models are trained only on $\\mathcal{D}_{train}$ , however evaluated on both $\\mathcal{D}_{train}$ and $k$ unseen domains $\\{\\mathcal{D}_{\\text{unseen}}^i\\}_{i=1}^k$ . To provide a concept-level explanation, we", + "bbox": [ + 511, + 824, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9520", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "should construct a concept set $\\mathcal{C} = \\{c_i\\}_{i=1}^M$ including $M$ interpretable concepts written by humans or LLMs. To evaluate the effectiveness of our method, we choose the well-studied concept bottleneck models (CBMs) family as the baselines of our methods. As one of the main branches of concept-based models, CBMs first project the image feature to an intermediate concept bottleneck layer to obtain the concept activation vector. Each activation value indicates the presence of the corresponding concept in the input. The forward pipeline of vanilla CBM [11] can be formulated as:", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} = f _ {C} \\left(f _ {I} (\\boldsymbol {x})\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 267, + 482, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = f _ {F} (\\boldsymbol {a}) = W _ {F} \\cdot \\boldsymbol {a}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 285, + 480, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{a} \\in \\mathbb{R}^{M}$ is the concept activation vector, $f_{I}: \\mathbb{R}^{h \\times w} \\to \\mathbb{R}^{d}$ is the image feature extractor maps the input $\\mathbf{x}$ into feature space, $f_{C}: \\mathbb{R}^{d} \\to \\mathbb{R}^{M}$ is the concept projection layer that projects the image feature into concept activation vector, and $f_{F}: \\mathbb{R}^{M} \\to \\mathbb{R}^{N_{y}}$ is the final linear classifier $W_{F}$ maps the concept activation into a final prediction.", + "bbox": [ + 89, + 311, + 483, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLIP-CBM. Vanilla CBM requires labor-intensive concept activation annotations to supervise the learning of concept activation. Contrastively, as illustrated in the blue part of Fig. 3, recent CLIP-based CBMs [21, 41, 42] employ the pre-trained CLIP model to extract the image embeddings $E_{I}(x)$ and concept embeddings $E_{T}(c)$ , respectively. Subsequently, CLIP-CBM obtains the concept activations $a$ by computing the similarity between image and concept embeddings, formulated as:", + "bbox": [ + 89, + 402, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na _ {i} = \\operatorname {s i m} \\left(E _ {I} (x), E _ {T} \\left(c _ {i}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 547, + 482, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} = \\left[ a _ {1}, a _ {2}, \\dots , a _ {M} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 566, + 480, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where sim indicates the cosine similarity.", + "bbox": [ + 89, + 593, + 364, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective of CLIP-CBM is:", + "bbox": [ + 109, + 607, + 323, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} (\\hat {y}, y) = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D} _ {t r a i n}} \\left[ \\mathcal {L} \\left(W _ {F} \\cdot \\boldsymbol {a}, y\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 632, + 482, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{CE}}(\\hat{y},y)$ is the cross-entropy loss.", + "bbox": [ + 89, + 657, + 370, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Domain Descriptors Orthogonality Loss", + "text_level": 1, + "bbox": [ + 89, + 681, + 434, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1, CLIP-CBMs suffer from the OOD generalization problem due to the negative impact of domain-specific concepts. To erase the negative influence of domain-specific concepts, we encourage the learnable class-concept correlation matrix (i.e. the weight of the final linear classification layer), to be orthogonal to the language-guided synthesized domain-specific concept activations.", + "bbox": [ + 89, + 704, + 483, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous works [15, 25], we decompose the concept activation $\\mathbf{a}$ into two components, the domain-specific concept activation vector $\\mathbf{a}_{sp}$ and the domain-shared concept vector $\\mathbf{a}_{sh}$ , represented as:", + "bbox": [ + 89, + 810, + 483, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} = \\boldsymbol {a} _ {s p} + \\boldsymbol {a} _ {s h}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 882, + 482, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to erase the impact of domain-specific concepts to the final predictions, formulated as:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {F} \\cdot \\boldsymbol {a} _ {s p} = 0. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 133, + 903, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the observations in Sec. 3, we can simulate the $\\mathbf{a}_{sp}$ via computing the similarity between language-guided domain shifts and the concepts. Specifically, the language-guided class-level domain shifts $\\Delta t_{py}$ are the subtraction between text embeddings of prompts of possible unseen domains and the training domain (e.g. \"A sketch of an apple\" - \"A photo of an apple\").", + "bbox": [ + 511, + 161, + 906, + 267 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta t \\left(p _ {i}, y\\right) = E _ {T} \\left(\\left[ p _ {i}, y \\right]\\right) - E _ {T} \\left(\\left[ p _ {\\text {t r a i n}}, y \\right]\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 277, + 903, + 295 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\boldsymbol {a} _ {\\mathrm {s p}} \\left(p _ {i} , y\\right)} = E _ {T} (c) \\cdot \\Delta t \\left(p _ {i}, y\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 297, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widehat{\\pmb{a}_{\\mathrm{sp}}(p_i,y)}$ is simulated domain-specific concept activation and $p_i,p_{train},y$ are the unseen domain descriptor, training domain descriptor and class, respectively.", + "bbox": [ + 511, + 330, + 903, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For all possible unseen domains $\\mathcal{P}$ and candidate classes $\\mathcal{Y}$ , based on Eq. (10) and Eq. (12), we derive our proposed domain descriptor orthogonality (DDO) loss:", + "bbox": [ + 511, + 378, + 905, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D D O}} = \\mathbb {E} _ {(p _ {i}, y) \\sim \\mathcal {P} \\times \\mathcal {Y}} \\left[ \\left| W _ {F} \\cdot \\widehat {\\boldsymbol {a} _ {\\mathrm {s p}} (p _ {i} , y)} \\right| \\right] \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 434, + 903, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The forward process of DDO is illustrated in the yellow part of Fig. 3. Note that the input to the DDO loss is independent of the current samples and only acts as a regularization term applied to the $W_{F}$ . Therefore, it can be plugged into various CBMs. The final loss is the combination of classification loss and DDO loss, stated as:", + "bbox": [ + 511, + 470, + 905, + 560 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} + \\lambda \\mathcal {L} _ {\\mathrm {D D O}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 571, + 903, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Generate Domain Descriptors", + "text_level": 1, + "bbox": [ + 511, + 599, + 781, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To simulate various unseen visual domains as much as possible, we need to write numerous domain descriptors. However, hand-writing these domain descriptors can be costly, and does not scale to large numbers of unseen domains. We can automatically construct this domain descriptor set $\\mathcal{P} = \\{p_i\\}_{i=1}^{N_p}$ by prompting a large language model, such as GPT-3.5 [22], to list possible unseen visual domains except for the training domain. We prompt the large language model with the input:", + "bbox": [ + 511, + 622, + 906, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Please list visual domains in short phrases as much as possible.", + "bbox": [ + 544, + 770, + 872, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: Here are some visual domains in short phrases: real-world photography, clipart illustrations, 3D renders...", + "bbox": [ + 544, + 800, + 872, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generated $N_p$ domain descriptors comprise the domain descriptor set $\\mathcal{P}$ . The full generation list and further implementation details can be found in the Appendix.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9521", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/36ebfe753fa27ff87d2c8cce652addffb1832aa49771b6d9f89940170a5bb395.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Performance on four single unseen domain benchmarks. For comparison, we list the performance of prevailing CBMs as baselines and report the results of integrating our proposed DDO regularizer into these baselines. $\\dagger$ indicates re-implemented with CLIP backbone. $\\ddagger$ means visual concepts are obtained with our re-implemented LLM prompts. ID is the performance on photo domains, and OOD is the generalization performance on other domains." + ], + "table_body": "
ModelConceptMethodCUB-PaintingAwA2-clipartLADA-SculptureLADV-3D
IDOODIDOODIDOODIDOOD
CLIP ZS [26]XX62.2152.7795.7090.2691.2682.0571.8266.29
CLIP LP [26]82.0061.4097.1186.7596.8174.4093.6863.81
CLIP-CBMhumanbaseline78.5150.5495.6981.9196.6670.4492.2160.64
+DDO78.7055.5395.7183.7296.7775.7692.5963.51
PCBM† [42]ConceptNetbaseline75.8554.4197.1784.7797.6076.6994.7165.88
+DDO76.4857.5097.1986.5897.6479.7494.8268.33
LaBO‡ [41]LLMbaseline81.9156.2497.1484.1597.4174.5699.9063.17
+DDO82.3459.6097.2687.6698.1280.0099.9368.01
", + "bbox": [ + 133, + 88, + 864, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 335, + 223, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Baselines", + "text_level": 1, + "bbox": [ + 89, + 359, + 302, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on seven domain adaptation benchmarks. Notably, CUB-Painting, PACS, OfficeHome and DomainNet are the previous common benchmarks, while AwA2-clipart, LADA-Sculpture, and LADV-3D are newly proposed in the paper.", + "bbox": [ + 89, + 383, + 482, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single unseen domain benchmarks. There are some classical image classification datasets with attribute annotations, which are widely applied for the evaluation of concept-based models. For instance, CUB-Painting is a fine-grained bird classification dataset that consists of two visual domains, the photo domain [32] and the painting domain [35]. Each image has an attribute annotation vector. Following CUB-Painting, we collect images with various other styles for photographic images from Animals with Attributes 2 (AwA2) [39], LAD-animal [47] and LAD-vehicle [47]. Specifically, we introduce three new benchmarks, namely AwA2-clipart, LADA-Sculpture and LADV-3D, which focused on photo $\\rightarrow$ clipart, real $\\rightarrow$ sculpture and real $\\rightarrow$ 3D model, respectively.", + "bbox": [ + 89, + 460, + 482, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multiple unseen domain benchmarks. There are also some classical domain adaptation benchmarks with multiple visual domains. PACS [13] is a domain adaptation benchmark with 7 classes and 4 visual styles, including art, cartoon, photo, and sketch. OfficeHome [31] contains 4 domains, i.e. art, clipart, product, realworld, where each domain consists of 65 categories. DomainNet [24] is a domain adaptation dataset containing 345 common classes from six visual domains, i.e. real, clipart, infographic, painting, quickdraw, and sketch. Following previous single domain generalization approaches [19, 48], we train on one source domain and evaluate on the other target domains. For each source domain, we report the average generalization accuracy on all target domains.", + "bbox": [ + 89, + 672, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our proposed method with the fol-", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lowing baseline approaches. 1) Black-box classifier: $CLIP-ZS$ [26] uses cosine similarity between class textual representation and image representation in CLIP shared space. $CLIP-LP$ [26] fits a linear classifier on top of the frozen CLIP image encoder. 2) CLIP-based CBMs: $CLIP-CBM$ leverages human-written attributes to serve as a concept bank and utilizes CLIP to compute concept activation label. $PCBM$ [42] uses ConceptNet [30] to retrieve concepts relevant to these classes and automatically construct the concept bank. Original PCBM employs different visual backbones on different datasets. For a fair comparison, we re-implement the PCBM with the CLIP backbone. $LaBO$ [41] applies the LLM to generate abundant visual concepts in contrast to PCBM, and we employ a two-layer classifier. For each CBM baseline, we keep their original objectives to solve the optimization.", + "bbox": [ + 511, + 335, + 906, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 589, + 732, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We prompt GPT3.5-turbo to generate 200 domain descriptors. These domain descriptors are used across all datasets and $\\lambda$ is set as 1. We pre-process the domain descriptors into text embeddings by CLIP text encoder. The DDO loss is added at the final classification layer of the CBMs in a plug-in manner. Following LaBO [41], we utilize the official CLIP ViT-L/141 by OpenAI as the default vision backbone to extract image embeddings. We train the linear function $W_{F}$ using the Adam [10] optimizer. All experiments are conducted on a single RTX3090 GPU. More implementation details, hyperparameters and experiments with other vision backbones are listed in the Appendix.", + "bbox": [ + 511, + 613, + 906, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Generalization to single unseen domain", + "text_level": 1, + "bbox": [ + 511, + 806, + 852, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate our model on single unseen domain benchmarks which focused on photo $\\rightarrow$ others generalization. Table. 1 shows in-distribution (ID) and out-of-distribution", + "bbox": [ + 511, + 830, + 905, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/openai/CLIP", + "bbox": [ + 531, + 886, + 777, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9522", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c88137de33cb773cd60f48116a070e87842cb969e4ab05dc82a3a6d82f722381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 361, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5419fc55f43a079b733ad88b205c3e15027c6a44018ab3448d5fec6e57dd7fde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 88, + 633, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5076c64d7d33fda148f009e1d3fc6d509aca76feff83c3e3593968e7e41d2dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 97, + 905, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/089571b2646149e42f7a64ffd1acd1f5c22fbfd6b0b105a38f7857226ad41955.jpg", + "image_caption": [ + "Figure 4. OOD performance on three multiple unseen domain generalization benchmarks, PACS, OfficeHome and DomainNet." + ], + "image_footnote": [], + "bbox": [ + 91, + 270, + 290, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bae3f9c2605426e23d0ba1a34ae6297c7c4e7998a760a556e58a94bdde2e24f9.jpg", + "image_caption": [ + "Figure 5. Ablation studies for effect of the number of domain descriptors. For each quantity, we randomly selected domain descriptors from a total of 200 domain descriptors and averaged the results over five random selections. Results of DomainNet are shown in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 295, + 270, + 493, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e22162dda3b3b7eb3cc10541a849d38a5d5c01b2a08377105bf0a6ec5663a637.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 270, + 699, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b57abaa88d90d1d60f95d6985cf4e9af426168a6979631e0f04870a1980eafe0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 270, + 903, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(OOD) accuracy on CUB-Painting, AwA2-clipart, LADA-Sculpture and LADV-3D. For all benchmarks, DDO loss is able to preserve or slightly improve the ID accuracy of each CBM while significantly improving their OOD accuracy (average 3 points improvements). We attribute the slight improvements in ID performance to the minor distribution differences between the training and test sets, even in the same visual domain. Moreover, LaBo with DDO loss can achieve competitive and superior performance on both ID and OOD performance with black-box CLIP-LP and narrow the gap with CLIP-ZS on OOD accuracy. These results demonstrate that our proposed LanCE design can effectively improve the generalization capabilities without sacrificing interpretability or classification accuracy. Besides, we notice that all models' generalization performance to sculpture and 3D model images on the LADA-Sculpture and LADV-3D benchmark is limited, indicating generalization from 2D to 3D remains challenging.", + "bbox": [ + 88, + 444, + 485, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Generalization to multiple unseen domains", + "text_level": 1, + "bbox": [ + 89, + 726, + 455, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the ability to generalize to multiple unseen domains and the generalization capability when different domains act as the source domain, we conduct experiments on PACS, OfficeHome, and DomainNet.", + "bbox": [ + 89, + 750, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with only using domain-shared concepts. In the absence of a pre-defined human-written concept set, we employ LLM to generate a concept set $\\mathcal{C}_{sp + in}$ for each dataset following LaBO [41]. This set contains both domain-specific concepts and domain-shared concepts. The main motivation of this paper is to mitigate the negative", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f92ef7d6b0db5aa35b4486103fd25eaa2a70672ec555a9252a2efa2b2168c705.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMethodCUB-PAwA2-cLADA-SLADV-3D
LanCEbaseline56.2484.1574.5663.17
+DDO(IR)57.6085.5077.7065.46
+DDO59.6087.6680.0068.01
", + "bbox": [ + 526, + 441, + 893, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation studies for the effect of relevance of the domain descriptors. +DDO(IR) only use the domain-irrelevant descriptors while +DDO use all domain descriptors.", + "bbox": [ + 511, + 502, + 906, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "impact of domain-specific concepts. A natural question that arises is why not simply query the LLM to generate only domain-shared concepts. To explore the differences between this domain-shared approach and our DDO loss, we introduce an additional baseline, where the LLM is queried exclusively to generate a domain-shared concept set $\\mathcal{C}_{sh}$ . Specifically, we adopt the single-domain generalization evaluation setting where one source domain is used for training, and the OOD performance is averaged across other domains. As shown in Fig. 4, only using $\\mathcal{C}_{sh}$ performs well only in certain cases, such as when the source domains in the PACS dataset are photo or sketch. We attribute this to the fact that the manifestation of domain-shared concepts varies across different domains. For instance, deer antlers in photo (realistic) and cartoon (simplified or exaggerated) domains differ significantly. Moreover, we empirically observe that prompt-based methods are highly sensitive to the wording of human-written prompts. Different prompts lead to significant performance variations, yet they share a common limitation: none consistently improve performance across diverse source domains. In contrast, DDO is more flexible. It adaptively erases the influence of domain-specific concepts while suppressing domain-specific infer", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9523", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bce1fdd097aeb73ee654f4e1196a1a009fb411fa195bacff3386a945a855d0df.jpg", + "image_caption": [ + "Figure 6. Qualitative examples about the top-5 concepts, ranking by their weights in the final linear layer $W_{F}$ . Baseline indicates the results of the original LaBo. Domain-specific concepts are highlighted in red. More results are listed in Appendix." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 901, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mation in domain-shared concept activations, leading to stable improvements across a variety of source domains.", + "bbox": [ + 89, + 311, + 480, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 348, + 254, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To explore the impact of domain descriptors on final performance, we conduct comprehensive ablation studies, with LaBO as the baseline. Specifically, we compare results with different numbers of domain descriptors and investigate the role of distinct domain descriptors.", + "bbox": [ + 89, + 372, + 482, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of the numbers of domain descriptors. We evaluate our method with regard to the number of domain descriptors $N_{p}$ as shown in Fig. 5. The ID accuracy remains almost constant as the number of domain descriptors increases, while the OOD accuracy gradually improves with an increasing number of domain descriptors, though with diminishing returns. Specifically, when the number of domain descriptors is below 20, each additional descriptor noticeably enhances OOD accuracy. However, as the number of domain descriptors continues to increase beyond this point, the improvement in OOD accuracy slows progressively. The results show that 100 domain descriptors are enough to achieve decent performance.", + "bbox": [ + 89, + 448, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of the relevance of domain descriptors. To investigate which parts of 200 domain descriptors contribute to the most to OOD accuracy improvements, we manually split all domain descriptors into two components, domain-relevant descriptors and domain-irrelevant descriptors based on their relevance with the test unseen domains. For example, for the LADA-Sculpture dataset, domain-relevant domain descriptors are \"sculpture\", \"statue\", and \"furniture\", referring to the generated word cloud mentioned in Fig. 2 and domain-irrelevant parts are the remaining ones. Detailed splits for each dataset are listed in the Appendix. Results are shown in the Table. 2. DDO(IR) indicates that we only use the domain-irrelevant domain descriptors to compute the DDO loss. This evaluation shows that domain-irrelevant domain descriptors can also slightly prompt the OOD accuracy on the unseen domains, although relatively fewer than the domain-relevant counterparts. It", + "bbox": [ + 89, + 645, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "indicates those domain descriptors are not fully orthonormal, also demonstrating that the DDO loss has the potential to generalize to pure unseen visual domains that are not mentioned in the domain descriptors set $\\mathcal{P}$ .", + "bbox": [ + 511, + 311, + 903, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 388, + 694, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed LanCE method, we further visualize the Top-5 concepts highly related to randomly selected classes within different unseen domains as illustrated in Fig. 6. The qualitative results demonstrate that our proposed method can significantly decrease the association between the final prediction and the domain-specific concepts, in other words, erasing the impact of domain-specific concepts on the final output. More qualitative results are shown in the Appendix.", + "bbox": [ + 511, + 414, + 903, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 570, + 632, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a language-guided concept-erasing (LanCE) framework that can effectively mitigate the biased association between domain-specific concepts and the final output. We first empirically demonstrate that the pre-trained CLIP can interpret the visual domain shifts into language. And these language-guided domain shifts can distinguish the domain-specific concepts via similarity within the CLIP embedding space. Based on these observations, we introduce a plug-in domain descriptor orthogonality (DDO) loss to erase the negative impact of domain-specific concepts, and that can significantly improve the generalization capabilities of prevailing concept bottleneck models (CBMs). We prompt the large language models (LLMs) to generate a bunch of domain descriptors to simulate numerous possible unseen domains. Moreover, considering that the current domain generalization community focuses more on generalizing to artistic unseen domains, we collect three new benchmarks that includes more difficult scenarios, i.e. 2D→3D, real animals→sculptures. Extensive experiments demonstrate the efficacy of our LanCE.", + "bbox": [ + 511, + 599, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9524", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407).", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 292, + 187, + 308 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2023. 2", + "[2] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. International Conference on Learning Representations (ICLR), 2023. 3", + "[3] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. Advances in neural information processing systems, 36:79024-79034, 2023. 2, 3", + "[4] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3", + "[5] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3", + "[6] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3", + "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3", + "[8] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1", + "[9] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept" + ], + "bbox": [ + 99, + 316, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2", + "[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 5", + "[12] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3", + "[13] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 2, 6", + "[14] Jichang Li, Guanbin Li, Yemin Shi, and Yizhou Yu. Cross-domain adaptive clustering for semi-supervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2505-2514, 2021. 3", + "[15] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 5", + "[16] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3", + "[17] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 3", + "[18] Jiawei Ma, Yulei Niu, Shiyuan Huang, Guangxing Han, and Shih-Fu Chang. Widin: Wording image for domain-invariant representation in single-source domain generalization. arXiv preprint arXiv:2405.18405, 2024. 3", + "[19] Lianbo Ma, Haidong Kang, Guo Yu, Qing Li, and Qiang He. Single-domain generalized predictor for neural architecture search system. IEEE Transactions on Computers, 2024. 3, 6", + "[20] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In Proceedings of the IEEE international conference on computer vision, pages 5715-5725, 2017. 3", + "[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2, 5", + "[22] OpenAI. Chatgpt. https://www.openai.com/chatgpt, 2023.5", + "[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9525", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1406-1415, 2019. 2, 6", + "[25] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International conference on machine learning, pages 7728-7738. PMLR, 2020. 3, 5", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 6", + "[27] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1", + "[28] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2", + "[29] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 3", + "[30] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6", + "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 2, 6", + "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6", + "[33] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 3", + "[34] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 2, 3", + "[35] Sinan Wang, Xinyang Chen, Yunbo Wang, Mingsheng Long, and Jianmin Wang. Progressive adversarial networks for fine-grained domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9213-9222, 2020. 2, 6", + "[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 3" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3", + "[38] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 3", + "[39] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 2, 6", + "[40] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3", + "[41] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 2, 5, 6, 7", + "[42] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 1, 2, 5, 6", + "[43] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2", + "[44] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23465-23476, 2023. 3", + "[45] Zequn Zeng, Jianqiao Sun, Hao Zhang, Tiansheng Wen, Yudi Su, Yan Xie, Zhengjue Wang, and Bo Chen. Hiclescore: A hierarchical metric for image captioning evaluation. arXiv preprint arXiv:2407.18589, 2024. 3", + "[46] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero-shot image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14100-14110, 2024. 3", + "[47] Bo Zhao, Yanwei Fu, Rui Liang, Jiahong Wu, Yonggang Wang, and Yizhou Wang. A large-scale attribute dataset for zero-shot learning. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 6", + "[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI conference on artificial intelligence, pages 21832-21840, 2024. 3, 6", + "[49] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 2, 3" + ], + "bbox": [ + 516, + 90, + 905, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9526", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_model.json b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fda538af24f4e1d2d0ca38566e53e58e262eca6f --- /dev/null +++ b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.803, + 0.177 + ], + "angle": 0, + "content": "Explaining Domain Shifts in Language: Concept erasing for Interpretable Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.203, + 0.786, + 0.239 + ], + "angle": 0, + "content": "Zequn Zeng\\(^{1*}\\), Yudi Su\\(^{1*}\\), Jianqiao Sun\\(^{1}\\), Tiansheng Wen\\(^{1}\\), Hao Zhang\\(^{1}\\), Zhengjue Wang\\(^{2}\\), Bo Chen\\(^{1\\dagger}\\), Hongwei Liu\\(^{1\\dagger}\\) and Jiawei Ma\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.24, + 0.877, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China." + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.257, + 0.878, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China." + }, + { + "type": "list", + "bbox": [ + 0.119, + 0.24, + 0.878, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.275, + 0.697, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\)City University of Hong Kong, Hong Kong SAR." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.296, + 0.72, + 0.309 + ], + "angle": 0, + "content": "zqzeng_1@stu.xidian.edu.cn, bchen@mail.xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.485, + 0.756 + ], + "angle": 0, + "content": "Concept-based models can map black-box representations to human-understandable concepts, which makes the decision-making process more transparent and then allows users to understand the reason behind predictions. However, domain-specific concepts often impact the final predictions, which subsequently undermine the model generalization capabilities, and prevent the model from being used in high-stake applications. In this paper, we propose a novel Language-guided Concept-Erasing (LanCE) framework. In particular, we empirically demonstrate that pre-trained vision-language models (VLMs) can approximate distinct visual domain shifts via domain descriptors while prompting large Language Models (LLMs) can easily simulate a wide range of descriptors of unseen visual domains. Then, we introduce a novel plug-in domain descriptor orthogonality (DDO) regularizer to mitigate the impact of these domain-specific concepts on the final predictions. Notably, the DDO regularizer is agnostic to the design of concept-based models and we integrate it into several prevailing models. Through evaluation of domain generalization on four standard benchmarks and three newly introduced benchmarks, we demonstrate that DDO can significantly improve the out-of-distribution (OOD) generalization over the previous state-of-the-art concept-based models. Our code is available at https://github.com/joeyz0z/LanCE." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Concept-based models [11, 27, 42] are prominent approaches for achieving model interpretability, which leverage human-understandable concepts to explain the black-box image representation. Specifically, these models first" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.344, + 0.639, + 0.357 + ], + "angle": 0, + "content": "Photo ( train )" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.358, + 0.64, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.415, + 0.65, + 0.425 + ], + "angle": 0, + "content": "Predict:Apple" + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.426, + 0.65, + 0.436 + ], + "angle": 0, + "content": "Confidence: \\(98\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.437, + 0.632, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.344, + 0.751, + 0.357 + ], + "angle": 0, + "content": "Clipart ( test )" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.358, + 0.746, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.414, + 0.757, + 0.425 + ], + "angle": 0, + "content": "Predict:Apple" + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.426, + 0.756, + 0.436 + ], + "angle": 0, + "content": "Confidence: \\(52\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.437, + 0.741, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.344, + 0.858, + 0.357 + ], + "angle": 0, + "content": "Sketch (test)" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.358, + 0.852, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.415, + 0.864, + 0.425 + ], + "angle": 0, + "content": "Predict: Pear X" + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.426, + 0.865, + 0.436 + ], + "angle": 0, + "content": "Confidence: \\(37\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.437, + 0.849, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.476, + 0.836, + 0.487 + ], + "angle": 0, + "content": "(a) Concept-based predictions on three domains." + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.489, + 0.667, + 0.499 + ], + "angle": 0, + "content": "Concept: Red color" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.499, + 0.709, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.489, + 0.869, + 0.499 + ], + "angle": 0, + "content": "Concept: Round shape" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.499, + 0.904, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.625, + 0.895, + 0.637 + ], + "angle": 0, + "content": "(b) Distribution shifts of domain-specific and domain-shared concepts." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Figure 1. Domain shifts in concept space. (a) Given the images of an apple in different visual domains, the prediction confidence of a concept-based model, trained on the photo domain, degrades due to the missing of concepts. (b) Distribution comparison of concept activation value (image-concept similarity computed via CLIP [26]) between photo domain and sketch domain, for two concepts, i.e., \"red color\" and \"round shape\", respectively. JS divergence indicates the distance between two distributions and tends to be larger for domain-specific concepts (e.g. \"red color\")." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "map the image feature to a concept activation vector in a concept space (each dimension corresponds to an interpretable concept) and then use the concept activation vector to predict the final output. Recently, equipped with pretrained vision-language models (VLMs) [8, 26], concept-based models [41, 42] can obtain the concept activation value by calculating the similarity between image embed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.372, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.348, + 0.108 + ], + "angle": 0, + "content": "dings and textual concept embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.486, + 0.413 + ], + "angle": 0, + "content": "Nevertheless, current concept-based models still struggle to handle domain shifts encountered during inference. The distinct distribution of visual concepts across different domains presents a significant generalization challenge for these models. As shown in Fig. 1a, the concept-based models trained on the photographic images usually excel at associating the discriminative visual clues – such as red color, waxy texture, and round shape – with specific classes like “apple”, akin to human beings. However, these models experience substantial performance degradation when applied to images in unseen visual domains, such as clipart or sketches, due to the absence of domain-specific texture and color concepts. These domain-specific concepts (e.g. “red color”) generally exhibit a larger distributional difference between visual domains than domain-shared concepts (e.g. “round shape”), as illustrated in Fig. 1b. Consequently, the concept-based models trained on a single domain are inevitably biased to associate those domain-specific concepts with the final predictions, thereby limiting their out-of-distribution (OOD) generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.485, + 0.627 + ], + "angle": 0, + "content": "To mitigate the issue of domain shifts in the concept space, a straightforward solution is to involve human experts in filtering out the domain-specific concepts and retaining only domain-shared concepts for classification tasks, which is nevertheless labor-intensive. Another possible solution is to diversify the training domain [1, 3] to mitigate the impact of the domain-specific concepts. However, it is impractical to cover all possible unseen domains during training exhaustively. Additionally, current domain generalization approaches [49] can not be directly applied to concept-based models due to the lack of interpretability of semantics of the learned representation [34]. Therefore, automating the identification and elimination of domain-specific concepts remains an under-explored area." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a Language-guided ConceptErasing framework, namely LanCE, to enhance the OOD generalization capabilities of current concept-based models. Specifically, we empirically demonstrate that pretrained VLMs, such as CLIP, can interpret the domain shifts through language, i.e., domain shifts can be approximated by a set of domain descriptors generated by large language models (LLMs). To alleviate the biased association between final predictions and domain-specific concepts, we introduce a plug-in domain descriptor orthogonality (DDO) loss that reduces the fluctuation of concept activation caused by these language-guided domain shifts. We evaluate our method on seven benchmarks, including four common benchmarks (CUB-Painting [35], PACS [13], OfficeHome [31] and DomainNet [24]), and three new proposed benchmarks (AwA2-clipart, LADA-Sculpture and LADV3D) based on several existing visual classification datasets with concept annotations [39, 47]. These newly-collected" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "datasets are gathered from the web and manually filtered. Extensive experiments demonstrate that the proposed approach can preserve the In-distribution (ID) classification capabilities and significantly improve the OOD generalization capabilities of prevailing concept-based models. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- We provide a concept-level explanation for the domain shifts and empirically reveal that VLMs can represent visual domain shifts into language space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.288 + ], + "angle": 0, + "content": "- Based on this observation, we propose a plug-in domain descriptor orthogonality regularizer to learn a domain-shared concept-based model without changing the model architecture and increasing training data." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.394 + ], + "angle": 0, + "content": "- We introduce three new concept-based OOD generalization benchmarks including more challenging scenarios such as natural photos \\(\\rightarrow\\) 3D-renderers and real animals \\(\\rightarrow\\) scluptures. Extensive experiments on four common and three new benchmarks demonstrate that our proposed method significantly improves the OOD generalization capabilities of current concept-based models." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.655, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.434, + 0.908, + 0.856 + ], + "angle": 0, + "content": "Concept-based models are one of the mainstream interpretable approaches that aim to understand models by associating black box features with meaningful concepts. Current concept-based models can broadly be divided into two categories, concept bottleneck models (CBMs) [11] which map the image features into an intermediate concept bottleneck layer whose neurons indicate concept activation values of pre-defined concepts, and concept activation vectors (CAVs) [9] which represent concepts as normal vectors of decision boundaries that distinguish positive and negative samples of a concept. Vanilla CBM requires fine-grained and precise concept annotation on the concept bottleneck layer and CAVs need positive and negative samples of each concept to learn meaningful concept representations. Many recent approaches are proposed to tackle the above concerns from two perspectives. On the one hand, many methods generate the concepts derived from ConceptNet [42], LLMs [21, 41] or multimodal datasets [28, 43] to automate the construction of the concept bottleneck layer. On the other hand, to remove the reliance on human-annotated concept labels, these methods deploy pre-trained VLMs such as CLIP to map concepts to text embeddings and compute image-concept similarity to serve as concept activation annotations. However, concepts of the above approaches are typically domain-sensitive and the corresponding concept activations have large distributional differences across various domains. They often fail to handle domain shifts when applied to unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Domain adaptation & generalization aims to handle the domain shifts between training and test data. In contrast to traditional deep learning based on in-distribution assump-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9518" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.095, + 0.576, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.309, + 0.483, + 0.324 + ], + "angle": 0, + "content": "(a) Interpret the visual domain shifts into language." + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.096, + 0.884, + 0.109 + ], + "angle": 0, + "content": "A sketch of an apple A photo of an apple" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.113, + 0.902, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.204, + 0.884, + 0.217 + ], + "angle": 0, + "content": "A bronze sculpture of a cat A photo of a cat" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.22, + 0.9, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.309, + 0.873, + 0.324 + ], + "angle": 0, + "content": "(b) Discover the domain-specific concepts." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.329, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 2. (a) We empirically demonstrate that the visual domain shift can be interpreted in language. For each class, we obtain the caption from the embedding difference of images from two domains. Then, by aggregating the captions across the classes, the keywords regarding the domain shift can be highlighted. At the same time, (b) the domain-specific concepts can be discovered from language descriptions of the different domains. In detail, they have higher similarities with the difference of domain-related class descriptions, i.e., textural domain shift, in the CLIP embedding space. More analyses are shown in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.418, + 0.485, + 0.69 + ], + "angle": 0, + "content": "tion, domain adaptation methods [4, 20] target to improve the performance on the target domain when only a few samples [14] or unlabeled data [40] from target domains are available. Instead, domain generalization methods [49] focused on more difficult scenarios in which no target domain data are available during the training phase. The setting of this paper belongs to the single domain generalization scenario, i.e. only one single domain is available during training [19, 36, 48]. While the domain generalization area is well studied, there is still researcher doubt about the interpretability of domain generalization methods [34]. Although disentanglement-based domain generalization methods [25, 33] decompose a feature into domain-shared and domain-specific parts, there is still a lack of deep understanding of the semantics of the learned features in domain generalization models. Our proposed method provides a concept-level explanation of domain shifts and takes steps to interpretable domain generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.484, + 0.828 + ], + "angle": 0, + "content": "Vision-Language Pretraining aims to bridge the gap between image and text representations. Compared with vision-only pretraining methods [5, 6, 17], vision-language pertaining like CLIP [26] achieved remarkable success and showed superior performance on some high-level vision understanding tasks [16, 44, 46] and multi-modality tasks [29, 38, 45]. Our empirical findings further validate that CLIP can interpret a variety of domain shifts into a set of domain descriptors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.317, + 0.862 + ], + "angle": 0, + "content": "3. Empirical Observations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "This section demonstrates that the pre-trained VLMs can approximate the visual domain shifts via domain descrip" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.478 + ], + "angle": 0, + "content": "tors. Meanwhile, these language-guided domain shifts can effectively distinguish the domain-specific concepts as illustrated in Fig. 2. It serves as the key insight for our proposed LanCE method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.601 + ], + "angle": 0, + "content": "Image-text alignment. CLIP trained on multi-modal contrastive learning, can effectively bridge the gap between image representation and text representation. The alignment facilitates a new approach to manipulate images using language, applicable for various vision tasks, e.g. image editing [7, 23], style transfer [12, 37], data augmentation [2, 3, 18]. The CLIP image-text similarity can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.612, + 0.905, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L I P}} (I, T) = \\operatorname {s i m} \\left(E _ {I} (I), E _ {T} (T)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.673 + ], + "angle": 0, + "content": "where \\( E_{I} \\) and \\( E_{T} \\) mean the image encoder and text encoder of CLIP, respectively. sim indicate the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.842 + ], + "angle": 0, + "content": "Interpreting the visual domain shifts. Inspired by previous works [2, 23], we assume that the visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space. To validate this hypothesis, we employ a CLIP-based zero-shot captioner [44] to translate the visual embedding differences into language, as illustrated in Fig. 2(a). Let \\(\\{I_c^{src}\\}\\) and \\(\\{I_c^{tgt}\\}\\) denote the images of the same class \\(c\\) from the source domain and the target domain, respectively. The class-level visual domain gap \\(\\Delta d_{c}\\) and corresponding description \\(t_c\\) can be generated by solving the following maximization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.855, + 0.905, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\max _ {t _ {c}} \\mathcal {L} _ {\\text {C L I P}} \\left(\\Delta d _ {c}, t _ {c}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.88, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {d} _ {c} = \\mathbb {E} \\left(E _ {I} \\left(I _ {c} ^ {t g t}\\right)\\right) - \\mathbb {E} \\left(E _ {I} \\left(I _ {c} ^ {s r c}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Figure 3. Overview of the LanCE. Blue part is the data flow of vanilla CLIP-CBMs (Sec. 4.1). To provide concept-level explanations, we first construct a human-written or LLMs-generated concept set \\(\\mathcal{C}\\) and extract the concept embeddings via the frozen CLIP text encoder. Given an image, we can extract the image embeddings via the frozen CLIP image encoder. The concept activations are the cosine similarity between image embeddings and concept embeddings. A learnable linear layer \\(W_{F}\\) is fitted on top of the concept activation vector and is responsible for predicting the final class and is optimized via cross-entropy loss. Yellow part is the data flow of our proposed DDO regularizer (Sec. 4.2). Similarly, we first construct a domain descriptor set (Sec. 4.3) to obtain the language-guided domain shifts and then simulate the domain-specific concept activations. To erase the effect of domain-specific concepts, the DDO regularizer encourages the orthogonality between the class-concept correlation matrix \\(W_{F}\\) (i.e. the final linear weight) and domain-specific concept activation \\(\\widehat{\\pmb{a}_{\\mathrm{sp}}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.482, + 0.681 + ], + "angle": 0, + "content": "As shown in Fig. 2(a), we produce high-matching (CLIP-Score \\(>0.9\\)) descriptions \\(t_c\\) that effectively interpret the class-level visual domain gap \\(\\Delta d_c\\) between sketch images, sculpture images and photographic images. We observe that these class-level domain descriptors are typically descriptions of the style difference of the two domains, regarding class-level domain descriptors. To further get the global difference between the two domains, we aggregate all class-level domain descriptors and transfer them into word cloud format, where the word size is proportional to the word frequency in the generated domain description corpus. As a result, we empirically find that global visual domain descriptors \\(\\Delta d\\) primarily consist of style words such as \"sketch and sculpture\" and background words e.g. \"white\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Discover the domain-specific concepts. Since we have revealed that visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space, we can synthesize the visual domain shifts via domain descriptors. Subsequently, we can leverage these language-guided domain shifts to discover domain-specific concepts. Specifically, we compute the similarity between textual domain shift embeddings (e.g. \"A sketch of an apple\" - \"A photo of an apple\") and all candidate concept embeddings extracted by CLIP text encoder. The results in Fig. 2(b) show that domain-specific concepts like \"shiny, waxy surface texture\" generally exhibit a higher similarity score due to their substantial distribution shift between those two domains while those domain-shared concepts such as \"Leaf around" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.515 + ], + "angle": 0, + "content": "the stem\" have relatively lower similarity score. Consequently, we can distinguish the domain-specific concepts via language-guided domain shifts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.605, + 0.543 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.553, + 0.907, + 0.793 + ], + "angle": 0, + "content": "To improve the generalization capabilities of prevailing concept-based models, as shown in Fig. 3, we propose a novel LanCE framework. In this section, we first formulate the problem and introduce Concept bottleneck models (CBMs), which can effectively map the black-box visual features to human-understandable concept space to interpret the final outputs (Sec. 4.1). However, domain-specific concepts will undermine the generalization capabilities. Then, to alleviate the biased association between domain-specific concepts and class predictions within the learnable linear classifier, we propose a domain descriptor orthogonality loss to improve the OOD classification performances of concept-based models (Sec. 4.2). Furthermore, considering all possible visual domains, we prompt LLMs to generate a bunch of domain descriptors to simulate numerous unseen visual domains (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.719, + 0.817 + ], + "angle": 0, + "content": "4.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Given a training domain data \\(\\mathcal{D}_{train} = \\{(\\pmb{x}_i, y_i)\\}\\), where \\(\\pmb{x}_i \\in \\mathcal{X}\\) are training images and \\(y_i \\in \\mathcal{Y}\\) are corresponding labels, the concept-based models are trained only on \\(\\mathcal{D}_{train}\\), however evaluated on both \\(\\mathcal{D}_{train}\\) and \\(k\\) unseen domains \\(\\{\\mathcal{D}_{\\text{unseen}}^i\\}_{i=1}^k\\). To provide a concept-level explanation, we" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9520" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.244 + ], + "angle": 0, + "content": "should construct a concept set \\(\\mathcal{C} = \\{c_i\\}_{i=1}^M\\) including \\(M\\) interpretable concepts written by humans or LLMs. To evaluate the effectiveness of our method, we choose the well-studied concept bottleneck models (CBMs) family as the baselines of our methods. As one of the main branches of concept-based models, CBMs first project the image feature to an intermediate concept bottleneck layer to obtain the concept activation vector. Each activation value indicates the presence of the corresponding concept in the input. The forward pipeline of vanilla CBM [11] can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.268, + 0.483, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} = f _ {C} \\left(f _ {I} (\\boldsymbol {x})\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.286, + 0.482, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = f _ {F} (\\boldsymbol {a}) = W _ {F} \\cdot \\boldsymbol {a}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.404 + ], + "angle": 0, + "content": "where \\(\\mathbf{a} \\in \\mathbb{R}^{M}\\) is the concept activation vector, \\(f_{I}: \\mathbb{R}^{h \\times w} \\to \\mathbb{R}^{d}\\) is the image feature extractor maps the input \\(\\mathbf{x}\\) into feature space, \\(f_{C}: \\mathbb{R}^{d} \\to \\mathbb{R}^{M}\\) is the concept projection layer that projects the image feature into concept activation vector, and \\(f_{F}: \\mathbb{R}^{M} \\to \\mathbb{R}^{N_{y}}\\) is the final linear classifier \\(W_{F}\\) maps the concept activation into a final prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.539 + ], + "angle": 0, + "content": "CLIP-CBM. Vanilla CBM requires labor-intensive concept activation annotations to supervise the learning of concept activation. Contrastively, as illustrated in the blue part of Fig. 3, recent CLIP-based CBMs [21, 41, 42] employ the pre-trained CLIP model to extract the image embeddings \\( E_{I}(x) \\) and concept embeddings \\( E_{T}(c) \\), respectively. Subsequently, CLIP-CBM obtains the concept activations \\( a \\) by computing the similarity between image and concept embeddings, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.549, + 0.483, + 0.566 + ], + "angle": 0, + "content": "\\[\na _ {i} = \\operatorname {s i m} \\left(E _ {I} (x), E _ {T} \\left(c _ {i}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.568, + 0.482, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} = \\left[ a _ {1}, a _ {2}, \\dots , a _ {M} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.594, + 0.366, + 0.608 + ], + "angle": 0, + "content": "where sim indicates the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.608, + 0.324, + 0.623 + ], + "angle": 0, + "content": "The objective of CLIP-CBM is:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.633, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} (\\hat {y}, y) = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D} _ {t r a i n}} \\left[ \\mathcal {L} \\left(W _ {F} \\cdot \\boldsymbol {a}, y\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.659, + 0.371, + 0.675 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{CE}}(\\hat{y},y)\\) is the cross-entropy loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.436, + 0.699 + ], + "angle": 0, + "content": "4.2. Domain Descriptors Orthogonality Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.811 + ], + "angle": 0, + "content": "As mentioned in Sec. 1, CLIP-CBMs suffer from the OOD generalization problem due to the negative impact of domain-specific concepts. To erase the negative influence of domain-specific concepts, we encourage the learnable class-concept correlation matrix (i.e. the weight of the final linear classification layer), to be orthogonal to the language-guided synthesized domain-specific concept activations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.872 + ], + "angle": 0, + "content": "Following previous works [15, 25], we decompose the concept activation \\( \\mathbf{a} \\) into two components, the domain-specific concept activation vector \\( \\mathbf{a}_{sp} \\) and the domain-shared concept vector \\( \\mathbf{a}_{sh} \\), represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.883, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} = \\boldsymbol {a} _ {s p} + \\boldsymbol {a} _ {s h}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "Our goal is to erase the impact of domain-specific concepts to the final predictions, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.134, + 0.905, + 0.15 + ], + "angle": 0, + "content": "\\[\nW _ {F} \\cdot \\boldsymbol {a} _ {s p} = 0. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.162, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Based on the observations in Sec. 3, we can simulate the \\( \\mathbf{a}_{sp} \\) via computing the similarity between language-guided domain shifts and the concepts. Specifically, the language-guided class-level domain shifts \\( \\Delta t_{py} \\) are the subtraction between text embeddings of prompts of possible unseen domains and the training domain (e.g. \"A sketch of an apple\" - \"A photo of an apple\")." + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.278, + 0.905, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\Delta t \\left(p _ {i}, y\\right) = E _ {T} \\left(\\left[ p _ {i}, y \\right]\\right) - E _ {T} \\left(\\left[ p _ {\\text {t r a i n}}, y \\right]\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.299, + 0.905, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\boldsymbol {a} _ {\\mathrm {s p}} \\left(p _ {i} , y\\right)} = E _ {T} (c) \\cdot \\Delta t \\left(p _ {i}, y\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.905, + 0.379 + ], + "angle": 0, + "content": "where \\(\\widehat{\\pmb{a}_{\\mathrm{sp}}(p_i,y)}\\) is simulated domain-specific concept activation and \\(p_i,p_{train},y\\) are the unseen domain descriptor, training domain descriptor and class, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.906, + 0.425 + ], + "angle": 0, + "content": "For all possible unseen domains \\(\\mathcal{P}\\) and candidate classes \\(\\mathcal{Y}\\), based on Eq. (10) and Eq. (12), we derive our proposed domain descriptor orthogonality (DDO) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.435, + 0.905, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D D O}} = \\mathbb {E} _ {(p _ {i}, y) \\sim \\mathcal {P} \\times \\mathcal {Y}} \\left[ \\left| W _ {F} \\cdot \\widehat {\\boldsymbol {a} _ {\\mathrm {s p}} (p _ {i} , y)} \\right| \\right] \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.561 + ], + "angle": 0, + "content": "The forward process of DDO is illustrated in the yellow part of Fig. 3. Note that the input to the DDO loss is independent of the current samples and only acts as a regularization term applied to the \\( W_{F} \\). Therefore, it can be plugged into various CBMs. The final loss is the combination of classification loss and DDO loss, stated as:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.573, + 0.905, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} + \\lambda \\mathcal {L} _ {\\mathrm {D D O}} \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.782, + 0.616 + ], + "angle": 0, + "content": "4.3. Generate Domain Descriptors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.762 + ], + "angle": 0, + "content": "To simulate various unseen visual domains as much as possible, we need to write numerous domain descriptors. However, hand-writing these domain descriptors can be costly, and does not scale to large numbers of unseen domains. We can automatically construct this domain descriptor set \\(\\mathcal{P} = \\{p_i\\}_{i=1}^{N_p}\\) by prompting a large language model, such as GPT-3.5 [22], to list possible unseen visual domains except for the training domain. We prompt the large language model with the input:" + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.771, + 0.874, + 0.8 + ], + "angle": 0, + "content": "Q: Please list visual domains in short phrases as much as possible." + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.801, + 0.874, + 0.845 + ], + "angle": 0, + "content": "A: Here are some visual domains in short phrases: real-world photography, clipart illustrations, 3D renders..." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The generated \\(N_p\\) domain descriptors comprise the domain descriptor set \\(\\mathcal{P}\\). The full generation list and further implementation details can be found in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9521" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.865, + 0.268 + ], + "angle": 0, + "content": "
ModelConceptMethodCUB-PaintingAwA2-clipartLADA-SculptureLADV-3D
IDOODIDOODIDOODIDOOD
CLIP ZS [26]XX62.2152.7795.7090.2691.2682.0571.8266.29
CLIP LP [26]82.0061.4097.1186.7596.8174.4093.6863.81
CLIP-CBMhumanbaseline78.5150.5495.6981.9196.6670.4492.2160.64
+DDO78.7055.5395.7183.7296.7775.7692.5963.51
PCBM† [42]ConceptNetbaseline75.8554.4197.1784.7797.6076.6994.7165.88
+DDO76.4857.5097.1986.5897.6479.7494.8268.33
LaBO‡ [41]LLMbaseline81.9156.2497.1484.1597.4174.5699.9063.17
+DDO82.3459.6097.2687.6698.1280.0099.9368.01
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.269, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Table 1. Performance on four single unseen domain benchmarks. For comparison, we list the performance of prevailing CBMs as baselines and report the results of integrating our proposed DDO regularizer into these baselines. \\(\\dagger\\) indicates re-implemented with CLIP backbone. \\(\\ddagger\\) means visual concepts are obtained with our re-implemented LLM prompts. ID is the performance on photo domains, and OOD is the generalization performance on other domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.224, + 0.354 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.361, + 0.303, + 0.377 + ], + "angle": 0, + "content": "5.1. Datasets and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.461 + ], + "angle": 0, + "content": "We conduct experiments on seven domain adaptation benchmarks. Notably, CUB-Painting, PACS, OfficeHome and DomainNet are the previous common benchmarks, while AwA2-clipart, LADA-Sculpture, and LADV-3D are newly proposed in the paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Single unseen domain benchmarks. There are some classical image classification datasets with attribute annotations, which are widely applied for the evaluation of concept-based models. For instance, CUB-Painting is a fine-grained bird classification dataset that consists of two visual domains, the photo domain [32] and the painting domain [35]. Each image has an attribute annotation vector. Following CUB-Painting, we collect images with various other styles for photographic images from Animals with Attributes 2 (AwA2) [39], LAD-animal [47] and LAD-vehicle [47]. Specifically, we introduce three new benchmarks, namely AwA2-clipart, LADA-Sculpture and LADV-3D, which focused on photo \\(\\rightarrow\\) clipart, real \\(\\rightarrow\\) sculpture and real \\(\\rightarrow\\) 3D model, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Multiple unseen domain benchmarks. There are also some classical domain adaptation benchmarks with multiple visual domains. PACS [13] is a domain adaptation benchmark with 7 classes and 4 visual styles, including art, cartoon, photo, and sketch. OfficeHome [31] contains 4 domains, i.e. art, clipart, product, realworld, where each domain consists of 65 categories. DomainNet [24] is a domain adaptation dataset containing 345 common classes from six visual domains, i.e. real, clipart, infographic, painting, quickdraw, and sketch. Following previous single domain generalization approaches [19, 48], we train on one source domain and evaluate on the other target domains. For each source domain, we report the average generalization accuracy on all target domains." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our proposed method with the fol-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.581 + ], + "angle": 0, + "content": "lowing baseline approaches. 1) Black-box classifier: \\(CLIP-ZS\\) [26] uses cosine similarity between class textual representation and image representation in CLIP shared space. \\(CLIP-LP\\) [26] fits a linear classifier on top of the frozen CLIP image encoder. 2) CLIP-based CBMs: \\(CLIP-CBM\\) leverages human-written attributes to serve as a concept bank and utilizes CLIP to compute concept activation label. \\(PCBM\\) [42] uses ConceptNet [30] to retrieve concepts relevant to these classes and automatically construct the concept bank. Original PCBM employs different visual backbones on different datasets. For a fair comparison, we re-implement the PCBM with the CLIP backbone. \\(LaBO\\) [41] applies the LLM to generate abundant visual concepts in contrast to PCBM, and we employ a two-layer classifier. For each CBM baseline, we keep their original objectives to solve the optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.733, + 0.607 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.796 + ], + "angle": 0, + "content": "We prompt GPT3.5-turbo to generate 200 domain descriptors. These domain descriptors are used across all datasets and \\(\\lambda\\) is set as 1. We pre-process the domain descriptors into text embeddings by CLIP text encoder. The DDO loss is added at the final classification layer of the CBMs in a plug-in manner. Following LaBO [41], we utilize the official CLIP ViT-L/141 by OpenAI as the default vision backbone to extract image embeddings. We train the linear function \\(W_{F}\\) using the Adam [10] optimizer. All experiments are conducted on a single RTX3090 GPU. More implementation details, hyperparameters and experiments with other vision backbones are listed in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.853, + 0.823 + ], + "angle": 0, + "content": "5.3. Generalization to single unseen domain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.906, + 0.876 + ], + "angle": 0, + "content": "We first evaluate our model on single unseen domain benchmarks which focused on photo \\(\\rightarrow\\) others generalization. Table. 1 shows in-distribution (ID) and out-of-distribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.778, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/openai/CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9522" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.362, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.089, + 0.635, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.098, + 0.906, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.255, + 0.873, + 0.269 + ], + "angle": 0, + "content": "Figure 4. OOD performance on three multiple unseen domain generalization benchmarks, PACS, OfficeHome and DomainNet." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.271, + 0.291, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.271, + 0.495, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.271, + 0.7, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.271, + 0.904, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 5. Ablation studies for effect of the number of domain descriptors. For each quantity, we randomly selected domain descriptors from a total of 200 domain descriptors and averaged the results over five random selections. Results of DomainNet are shown in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.486, + 0.718 + ], + "angle": 0, + "content": "(OOD) accuracy on CUB-Painting, AwA2-clipart, LADA-Sculpture and LADV-3D. For all benchmarks, DDO loss is able to preserve or slightly improve the ID accuracy of each CBM while significantly improving their OOD accuracy (average 3 points improvements). We attribute the slight improvements in ID performance to the minor distribution differences between the training and test sets, even in the same visual domain. Moreover, LaBo with DDO loss can achieve competitive and superior performance on both ID and OOD performance with black-box CLIP-LP and narrow the gap with CLIP-ZS on OOD accuracy. These results demonstrate that our proposed LanCE design can effectively improve the generalization capabilities without sacrificing interpretability or classification accuracy. Besides, we notice that all models' generalization performance to sculpture and 3D model images on the LADA-Sculpture and LADV-3D benchmark is limited, indicating generalization from 2D to 3D remains challenging." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.456, + 0.743 + ], + "angle": 0, + "content": "5.4. Generalization to multiple unseen domains" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.809 + ], + "angle": 0, + "content": "To assess the ability to generalize to multiple unseen domains and the generalization capability when different domains act as the source domain, we conduct experiments on PACS, OfficeHome, and DomainNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparison with only using domain-shared concepts. In the absence of a pre-defined human-written concept set, we employ LLM to generate a concept set \\(\\mathcal{C}_{sp + in}\\) for each dataset following LaBO [41]. This set contains both domain-specific concepts and domain-shared concepts. The main motivation of this paper is to mitigate the negative" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.442, + 0.895, + 0.502 + ], + "angle": 0, + "content": "
ModelMethodCUB-PAwA2-cLADA-SLADV-3D
LanCEbaseline56.2484.1574.5663.17
+DDO(IR)57.6085.5077.7065.46
+DDO59.6087.6680.0068.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.503, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Table 2. Ablation studies for the effect of relevance of the domain descriptors. +DDO(IR) only use the domain-irrelevant descriptors while +DDO use all domain descriptors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "impact of domain-specific concepts. A natural question that arises is why not simply query the LLM to generate only domain-shared concepts. To explore the differences between this domain-shared approach and our DDO loss, we introduce an additional baseline, where the LLM is queried exclusively to generate a domain-shared concept set \\(\\mathcal{C}_{sh}\\). Specifically, we adopt the single-domain generalization evaluation setting where one source domain is used for training, and the OOD performance is averaged across other domains. As shown in Fig. 4, only using \\(\\mathcal{C}_{sh}\\) performs well only in certain cases, such as when the source domains in the PACS dataset are photo or sketch. We attribute this to the fact that the manifestation of domain-shared concepts varies across different domains. For instance, deer antlers in photo (realistic) and cartoon (simplified or exaggerated) domains differ significantly. Moreover, we empirically observe that prompt-based methods are highly sensitive to the wording of human-written prompts. Different prompts lead to significant performance variations, yet they share a common limitation: none consistently improve performance across diverse source domains. In contrast, DDO is more flexible. It adaptively erases the influence of domain-specific concepts while suppressing domain-specific infer" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9523" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.903, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.905, + 0.3 + ], + "angle": 0, + "content": "Figure 6. Qualitative examples about the top-5 concepts, ranking by their weights in the final linear layer \\( W_{F} \\). Baseline indicates the results of the original LaBo. Domain-specific concepts are highlighted in red. More results are listed in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.482, + 0.342 + ], + "angle": 0, + "content": "mation in domain-shared concept activations, leading to stable improvements across a variety of source domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.349, + 0.255, + 0.364 + ], + "angle": 0, + "content": "5.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.448 + ], + "angle": 0, + "content": "To explore the impact of domain descriptors on final performance, we conduct comprehensive ablation studies, with LaBO as the baseline. Specifically, we compare results with different numbers of domain descriptors and investigate the role of distinct domain descriptors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.644 + ], + "angle": 0, + "content": "The impact of the numbers of domain descriptors. We evaluate our method with regard to the number of domain descriptors \\( N_{p} \\) as shown in Fig. 5. The ID accuracy remains almost constant as the number of domain descriptors increases, while the OOD accuracy gradually improves with an increasing number of domain descriptors, though with diminishing returns. Specifically, when the number of domain descriptors is below 20, each additional descriptor noticeably enhances OOD accuracy. However, as the number of domain descriptors continues to increase beyond this point, the improvement in OOD accuracy slows progressively. The results show that 100 domain descriptors are enough to achieve decent performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The impact of the relevance of domain descriptors. To investigate which parts of 200 domain descriptors contribute to the most to OOD accuracy improvements, we manually split all domain descriptors into two components, domain-relevant descriptors and domain-irrelevant descriptors based on their relevance with the test unseen domains. For example, for the LADA-Sculpture dataset, domain-relevant domain descriptors are \"sculpture\", \"statue\", and \"furniture\", referring to the generated word cloud mentioned in Fig. 2 and domain-irrelevant parts are the remaining ones. Detailed splits for each dataset are listed in the Appendix. Results are shown in the Table. 2. DDO(IR) indicates that we only use the domain-irrelevant domain descriptors to compute the DDO loss. This evaluation shows that domain-irrelevant domain descriptors can also slightly prompt the OOD accuracy on the unseen domains, although relatively fewer than the domain-relevant counterparts. It" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.373 + ], + "angle": 0, + "content": "indicates those domain descriptors are not fully orthonormal, also demonstrating that the DDO loss has the potential to generalize to pure unseen visual domains that are not mentioned in the domain descriptors set \\(\\mathcal{P}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.696, + 0.405 + ], + "angle": 0, + "content": "5.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.905, + 0.551 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed LanCE method, we further visualize the Top-5 concepts highly related to randomly selected classes within different unseen domains as illustrated in Fig. 6. The qualitative results demonstrate that our proposed method can significantly decrease the association between the final prediction and the domain-specific concepts, in other words, erasing the impact of domain-specific concepts on the final output. More qualitative results are shown in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.633, + 0.587 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a language-guided concept-erasing (LanCE) framework that can effectively mitigate the biased association between domain-specific concepts and the final output. We first empirically demonstrate that the pre-trained CLIP can interpret the visual domain shifts into language. And these language-guided domain shifts can distinguish the domain-specific concepts via similarity within the CLIP embedding space. Based on these observations, we introduce a plug-in domain descriptor orthogonality (DDO) loss to erase the negative impact of domain-specific concepts, and that can significantly improve the generalization capabilities of prevailing concept bottleneck models (CBMs). We prompt the large language models (LLMs) to generate a bunch of domain descriptors to simulate numerous possible unseen domains. Moreover, considering that the current domain generalization community focuses more on generalizing to artistic unseen domains, we collect three new benchmarks that includes more difficult scenarios, i.e. 2D→3D, real animals→sculptures. Extensive experiments demonstrate the efficacy of our LanCE." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9524" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.283 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.293, + 0.188, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.486, + 0.374 + ], + "angle": 0, + "content": "[1] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.374, + 0.484, + 0.443 + ], + "angle": 0, + "content": "[2] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[3] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. Advances in neural information processing systems, 36:79024-79034, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[4] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[5] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.484, + 0.72 + ], + "angle": 0, + "content": "[6] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[8] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[9] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.318, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.15 + ], + "angle": 0, + "content": "[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.906, + 0.207 + ], + "angle": 0, + "content": "[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.908, + 0.263 + ], + "angle": 0, + "content": "[12] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.32 + ], + "angle": 0, + "content": "[13] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[14] Jichang Li, Guanbin Li, Yemin Shi, and Yizhou Yu. Cross-domain adaptive clustering for semi-supervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2505-2514, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[15] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[16] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[17] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[18] Jiawei Ma, Yulei Niu, Shiyuan Huang, Guangxing Han, and Shih-Fu Chang. Widin: Wording image for domain-invariant representation in single-source domain generalization. arXiv preprint arXiv:2405.18405, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.906, + 0.689 + ], + "angle": 0, + "content": "[19] Lianbo Ma, Haidong Kang, Guo Yu, Qing Li, and Qiang He. Single-domain generalized predictor for neural architecture search system. IEEE Transactions on Computers, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[20] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In Proceedings of the IEEE international conference on computer vision, pages 5715-5725, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[22] OpenAI. Chatgpt. https://www.openai.com/chatgpt, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9525" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1406-1415, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[25] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International conference on machine learning, pages 7728-7738. PMLR, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.485, + 0.3 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.485, + 0.329 + ], + "angle": 0, + "content": "[27] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.485, + 0.399 + ], + "angle": 0, + "content": "[28] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.485, + 0.455 + ], + "angle": 0, + "content": "[29] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.485, + 0.511 + ], + "angle": 0, + "content": "[30] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.485, + 0.58 + ], + "angle": 0, + "content": "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.485, + 0.622 + ], + "angle": 0, + "content": "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.485, + 0.692 + ], + "angle": 0, + "content": "[33] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.485, + 0.762 + ], + "angle": 0, + "content": "[34] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.485, + 0.831 + ], + "angle": 0, + "content": "[35] Sinan Wang, Xinyang Chen, Yunbo Wang, Mingsheng Long, and Jianmin Wang. Progressive adversarial networks for fine-grained domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9213-9222, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.485, + 0.9 + ], + "angle": 0, + "content": "[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[38] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[39] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.345 + ], + "angle": 0, + "content": "[40] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.906, + 0.43 + ], + "angle": 0, + "content": "[41] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.906, + 0.472 + ], + "angle": 0, + "content": "[42] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.472, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[43] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.906, + 0.586 + ], + "angle": 0, + "content": "[44] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23465-23476, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[45] Zequn Zeng, Jianqiao Sun, Hao Zhang, Tiansheng Wen, Yudi Su, Yan Xie, Zhengjue Wang, and Bo Chen. Hiclescore: A hierarchical metric for image captioning evaluation. arXiv preprint arXiv:2407.18589, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.906, + 0.713 + ], + "angle": 0, + "content": "[46] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero-shot image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14100-14110, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.713, + 0.906, + 0.784 + ], + "angle": 0, + "content": "[47] Bo Zhao, Yanwei Fu, Rui Liang, Jiahong Wu, Yonggang Wang, and Yizhou Wang. A large-scale attribute dataset for zero-shot learning. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.784, + 0.906, + 0.841 + ], + "angle": 0, + "content": "[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI conference on artificial intelligence, pages 21832-21840, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.841, + 0.906, + 0.897 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9526" + } + ] +] \ No newline at end of file diff --git a/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_origin.pdf b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b5ffb98cad4f5859669599f2e97a3ebf38bf4103 --- /dev/null +++ b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/5b5f023c-5e83-4e83-95d9-ae7c139631a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:509cfa1332b7ddc82d4981ae7748b14e5147c2f3ef346cbfe2380a2574e95388 +size 2810099 diff --git a/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/full.md b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1de2158d8cfac5c662bea55eb6ae047139043e29 --- /dev/null +++ b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/full.md @@ -0,0 +1,347 @@ +# Explaining Domain Shifts in Language: Concept erasing for Interpretable Image Classification + +Zequn Zeng $^{1*}$ , Yudi Su $^{1*}$ , Jianqiao Sun $^{1}$ , Tiansheng Wen $^{1}$ , Hao Zhang $^{1}$ , Zhengjue Wang $^{2}$ , Bo Chen $^{1\dagger}$ , Hongwei Liu $^{1\dagger}$ and Jiawei Ma $^{3}$ + +$^{1}$ National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China. +$^{2}$ State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China. + +$^{3}$ City University of Hong Kong, Hong Kong SAR. + +zqzeng_1@stu.xidian.edu.cn, bchen@mail.xidian.edu.cn + +# Abstract + +Concept-based models can map black-box representations to human-understandable concepts, which makes the decision-making process more transparent and then allows users to understand the reason behind predictions. However, domain-specific concepts often impact the final predictions, which subsequently undermine the model generalization capabilities, and prevent the model from being used in high-stake applications. In this paper, we propose a novel Language-guided Concept-Erasing (LanCE) framework. In particular, we empirically demonstrate that pre-trained vision-language models (VLMs) can approximate distinct visual domain shifts via domain descriptors while prompting large Language Models (LLMs) can easily simulate a wide range of descriptors of unseen visual domains. Then, we introduce a novel plug-in domain descriptor orthogonality (DDO) regularizer to mitigate the impact of these domain-specific concepts on the final predictions. Notably, the DDO regularizer is agnostic to the design of concept-based models and we integrate it into several prevailing models. Through evaluation of domain generalization on four standard benchmarks and three newly introduced benchmarks, we demonstrate that DDO can significantly improve the out-of-distribution (OOD) generalization over the previous state-of-the-art concept-based models. Our code is available at https://github.com/joeyz0z/LanCE. + +# 1. Introduction + +Concept-based models [11, 27, 42] are prominent approaches for achieving model interpretability, which leverage human-understandable concepts to explain the black-box image representation. Specifically, these models first + +![](images/7f5a04b89c74ccf53b4b64acc28fdc8f2ae3680c3508afbab689b7ae319a69ea.jpg) +Photo ( train ) +Predict:Apple + +![](images/6f691512ec1b34f8b94bd8503f154cf3ec0b170d2576bdc3c206069f90951d98.jpg) +Confidence: $98\%$ + +![](images/3b1c74964fb0523b19cad0b2d40f0e0d9275c1c19f46ef15c278ebcdeab3fc6e.jpg) +Clipart ( test ) +Predict:Apple + +![](images/f76669c91a4b991a4dcaa88e97cdc032d91fdc96d4c7641a86fc10173edda871.jpg) +Confidence: $52\%$ +(a) Concept-based predictions on three domains. + +![](images/cb5f2075e5070feb9afe5349637f0d76fc797739e2e3840c133b2c5abdf3b255.jpg) +Sketch (test) +Predict: Pear X + +![](images/337d7025f18ded1eb7875fa186073edfc564b2ee06d8bd1f3225b1225fe0e0be.jpg) +Confidence: $37\%$ + +![](images/ca7c2452e649077e4e30bf2dd73d205167f62ab5cc2266b873b3e044f0ef2ac4.jpg) +Concept: Red color +(b) Distribution shifts of domain-specific and domain-shared concepts. +Figure 1. Domain shifts in concept space. (a) Given the images of an apple in different visual domains, the prediction confidence of a concept-based model, trained on the photo domain, degrades due to the missing of concepts. (b) Distribution comparison of concept activation value (image-concept similarity computed via CLIP [26]) between photo domain and sketch domain, for two concepts, i.e., "red color" and "round shape", respectively. JS divergence indicates the distance between two distributions and tends to be larger for domain-specific concepts (e.g. "red color"). + +![](images/085228f4148871c7fc09b0719f54b6fb53848a44bcccf3e3293627f950a3c67b.jpg) +Concept: Round shape + +map the image feature to a concept activation vector in a concept space (each dimension corresponds to an interpretable concept) and then use the concept activation vector to predict the final output. Recently, equipped with pretrained vision-language models (VLMs) [8, 26], concept-based models [41, 42] can obtain the concept activation value by calculating the similarity between image embed + +dings and textual concept embeddings. + +Nevertheless, current concept-based models still struggle to handle domain shifts encountered during inference. The distinct distribution of visual concepts across different domains presents a significant generalization challenge for these models. As shown in Fig. 1a, the concept-based models trained on the photographic images usually excel at associating the discriminative visual clues – such as red color, waxy texture, and round shape – with specific classes like “apple”, akin to human beings. However, these models experience substantial performance degradation when applied to images in unseen visual domains, such as clipart or sketches, due to the absence of domain-specific texture and color concepts. These domain-specific concepts (e.g. “red color”) generally exhibit a larger distributional difference between visual domains than domain-shared concepts (e.g. “round shape”), as illustrated in Fig. 1b. Consequently, the concept-based models trained on a single domain are inevitably biased to associate those domain-specific concepts with the final predictions, thereby limiting their out-of-distribution (OOD) generalization capabilities. + +To mitigate the issue of domain shifts in the concept space, a straightforward solution is to involve human experts in filtering out the domain-specific concepts and retaining only domain-shared concepts for classification tasks, which is nevertheless labor-intensive. Another possible solution is to diversify the training domain [1, 3] to mitigate the impact of the domain-specific concepts. However, it is impractical to cover all possible unseen domains during training exhaustively. Additionally, current domain generalization approaches [49] can not be directly applied to concept-based models due to the lack of interpretability of semantics of the learned representation [34]. Therefore, automating the identification and elimination of domain-specific concepts remains an under-explored area. + +In this paper, we propose a Language-guided ConceptErasing framework, namely LanCE, to enhance the OOD generalization capabilities of current concept-based models. Specifically, we empirically demonstrate that pretrained VLMs, such as CLIP, can interpret the domain shifts through language, i.e., domain shifts can be approximated by a set of domain descriptors generated by large language models (LLMs). To alleviate the biased association between final predictions and domain-specific concepts, we introduce a plug-in domain descriptor orthogonality (DDO) loss that reduces the fluctuation of concept activation caused by these language-guided domain shifts. We evaluate our method on seven benchmarks, including four common benchmarks (CUB-Painting [35], PACS [13], OfficeHome [31] and DomainNet [24]), and three new proposed benchmarks (AwA2-clipart, LADA-Sculpture and LADV3D) based on several existing visual classification datasets with concept annotations [39, 47]. These newly-collected + +datasets are gathered from the web and manually filtered. Extensive experiments demonstrate that the proposed approach can preserve the In-distribution (ID) classification capabilities and significantly improve the OOD generalization capabilities of prevailing concept-based models. Our contributions can be summarized as follows: + +- We provide a concept-level explanation for the domain shifts and empirically reveal that VLMs can represent visual domain shifts into language space. +- Based on this observation, we propose a plug-in domain descriptor orthogonality regularizer to learn a domain-shared concept-based model without changing the model architecture and increasing training data. +- We introduce three new concept-based OOD generalization benchmarks including more challenging scenarios such as natural photos $\rightarrow$ 3D-renderers and real animals $\rightarrow$ scluptures. Extensive experiments on four common and three new benchmarks demonstrate that our proposed method significantly improves the OOD generalization capabilities of current concept-based models. + +# 2. Related Work + +Concept-based models are one of the mainstream interpretable approaches that aim to understand models by associating black box features with meaningful concepts. Current concept-based models can broadly be divided into two categories, concept bottleneck models (CBMs) [11] which map the image features into an intermediate concept bottleneck layer whose neurons indicate concept activation values of pre-defined concepts, and concept activation vectors (CAVs) [9] which represent concepts as normal vectors of decision boundaries that distinguish positive and negative samples of a concept. Vanilla CBM requires fine-grained and precise concept annotation on the concept bottleneck layer and CAVs need positive and negative samples of each concept to learn meaningful concept representations. Many recent approaches are proposed to tackle the above concerns from two perspectives. On the one hand, many methods generate the concepts derived from ConceptNet [42], LLMs [21, 41] or multimodal datasets [28, 43] to automate the construction of the concept bottleneck layer. On the other hand, to remove the reliance on human-annotated concept labels, these methods deploy pre-trained VLMs such as CLIP to map concepts to text embeddings and compute image-concept similarity to serve as concept activation annotations. However, concepts of the above approaches are typically domain-sensitive and the corresponding concept activations have large distributional differences across various domains. They often fail to handle domain shifts when applied to unseen domains. + +Domain adaptation & generalization aims to handle the domain shifts between training and test data. In contrast to traditional deep learning based on in-distribution assump- + +![](images/107be90093f88dd5cd39df0c8266f606815b36a9781746acb40b2ff80fe28d5d.jpg) +(a) Interpret the visual domain shifts into language. + +![](images/20f8d5d28f97dde4756e151b9ae81b85edd82bfc39670c46ebdcb660375525d5.jpg) +A sketch of an apple A photo of an apple + +![](images/ae7b0b6976a6ebfe07f1fed017f8f45baa696c0683682127b453bf8889497840.jpg) +A bronze sculpture of a cat A photo of a cat +(b) Discover the domain-specific concepts. +Figure 2. (a) We empirically demonstrate that the visual domain shift can be interpreted in language. For each class, we obtain the caption from the embedding difference of images from two domains. Then, by aggregating the captions across the classes, the keywords regarding the domain shift can be highlighted. At the same time, (b) the domain-specific concepts can be discovered from language descriptions of the different domains. In detail, they have higher similarities with the difference of domain-related class descriptions, i.e., textural domain shift, in the CLIP embedding space. More analyses are shown in the Appendix. + +tion, domain adaptation methods [4, 20] target to improve the performance on the target domain when only a few samples [14] or unlabeled data [40] from target domains are available. Instead, domain generalization methods [49] focused on more difficult scenarios in which no target domain data are available during the training phase. The setting of this paper belongs to the single domain generalization scenario, i.e. only one single domain is available during training [19, 36, 48]. While the domain generalization area is well studied, there is still researcher doubt about the interpretability of domain generalization methods [34]. Although disentanglement-based domain generalization methods [25, 33] decompose a feature into domain-shared and domain-specific parts, there is still a lack of deep understanding of the semantics of the learned features in domain generalization models. Our proposed method provides a concept-level explanation of domain shifts and takes steps to interpretable domain generation. + +Vision-Language Pretraining aims to bridge the gap between image and text representations. Compared with vision-only pretraining methods [5, 6, 17], vision-language pertaining like CLIP [26] achieved remarkable success and showed superior performance on some high-level vision understanding tasks [16, 44, 46] and multi-modality tasks [29, 38, 45]. Our empirical findings further validate that CLIP can interpret a variety of domain shifts into a set of domain descriptors. + +# 3. Empirical Observations + +This section demonstrates that the pre-trained VLMs can approximate the visual domain shifts via domain descrip + +tors. Meanwhile, these language-guided domain shifts can effectively distinguish the domain-specific concepts as illustrated in Fig. 2. It serves as the key insight for our proposed LanCE method. + +Image-text alignment. CLIP trained on multi-modal contrastive learning, can effectively bridge the gap between image representation and text representation. The alignment facilitates a new approach to manipulate images using language, applicable for various vision tasks, e.g. image editing [7, 23], style transfer [12, 37], data augmentation [2, 3, 18]. The CLIP image-text similarity can be computed as: + +$$ +\mathcal {L} _ {\mathrm {C L I P}} (I, T) = \operatorname {s i m} \left(E _ {I} (I), E _ {T} (T)\right) \tag {1} +$$ + +where $E_{I}$ and $E_{T}$ mean the image encoder and text encoder of CLIP, respectively. sim indicate the cosine similarity. + +Interpreting the visual domain shifts. Inspired by previous works [2, 23], we assume that the visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space. To validate this hypothesis, we employ a CLIP-based zero-shot captioner [44] to translate the visual embedding differences into language, as illustrated in Fig. 2(a). Let $\{I_c^{src}\}$ and $\{I_c^{tgt}\}$ denote the images of the same class $c$ from the source domain and the target domain, respectively. The class-level visual domain gap $\Delta d_{c}$ and corresponding description $t_c$ can be generated by solving the following maximization problem: + +$$ +\max _ {t _ {c}} \mathcal {L} _ {\text {C L I P}} \left(\Delta d _ {c}, t _ {c}\right), \tag {2} +$$ + +$$ +\Delta \boldsymbol {d} _ {c} = \mathbb {E} \left(E _ {I} \left(I _ {c} ^ {t g t}\right)\right) - \mathbb {E} \left(E _ {I} \left(I _ {c} ^ {s r c}\right)\right), \tag {3} +$$ + +![](images/7adb23ec0613b26e3cc1008ae85f6ff553500a5e384e47db3762bf23283ae603.jpg) +Figure 3. Overview of the LanCE. Blue part is the data flow of vanilla CLIP-CBMs (Sec. 4.1). To provide concept-level explanations, we first construct a human-written or LLMs-generated concept set $\mathcal{C}$ and extract the concept embeddings via the frozen CLIP text encoder. Given an image, we can extract the image embeddings via the frozen CLIP image encoder. The concept activations are the cosine similarity between image embeddings and concept embeddings. A learnable linear layer $W_{F}$ is fitted on top of the concept activation vector and is responsible for predicting the final class and is optimized via cross-entropy loss. Yellow part is the data flow of our proposed DDO regularizer (Sec. 4.2). Similarly, we first construct a domain descriptor set (Sec. 4.3) to obtain the language-guided domain shifts and then simulate the domain-specific concept activations. To erase the effect of domain-specific concepts, the DDO regularizer encourages the orthogonality between the class-concept correlation matrix $W_{F}$ (i.e. the final linear weight) and domain-specific concept activation $\widehat{\pmb{a}_{\mathrm{sp}}}$ . + +As shown in Fig. 2(a), we produce high-matching (CLIP-Score $>0.9$ ) descriptions $t_c$ that effectively interpret the class-level visual domain gap $\Delta d_c$ between sketch images, sculpture images and photographic images. We observe that these class-level domain descriptors are typically descriptions of the style difference of the two domains, regarding class-level domain descriptors. To further get the global difference between the two domains, we aggregate all class-level domain descriptors and transfer them into word cloud format, where the word size is proportional to the word frequency in the generated domain description corpus. As a result, we empirically find that global visual domain descriptors $\Delta d$ primarily consist of style words such as "sketch and sculpture" and background words e.g. "white". + +Discover the domain-specific concepts. Since we have revealed that visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space, we can synthesize the visual domain shifts via domain descriptors. Subsequently, we can leverage these language-guided domain shifts to discover domain-specific concepts. Specifically, we compute the similarity between textual domain shift embeddings (e.g. "A sketch of an apple" - "A photo of an apple") and all candidate concept embeddings extracted by CLIP text encoder. The results in Fig. 2(b) show that domain-specific concepts like "shiny, waxy surface texture" generally exhibit a higher similarity score due to their substantial distribution shift between those two domains while those domain-shared concepts such as "Leaf around + +the stem" have relatively lower similarity score. Consequently, we can distinguish the domain-specific concepts via language-guided domain shifts. + +# 4. Method + +To improve the generalization capabilities of prevailing concept-based models, as shown in Fig. 3, we propose a novel LanCE framework. In this section, we first formulate the problem and introduce Concept bottleneck models (CBMs), which can effectively map the black-box visual features to human-understandable concept space to interpret the final outputs (Sec. 4.1). However, domain-specific concepts will undermine the generalization capabilities. Then, to alleviate the biased association between domain-specific concepts and class predictions within the learnable linear classifier, we propose a domain descriptor orthogonality loss to improve the OOD classification performances of concept-based models (Sec. 4.2). Furthermore, considering all possible visual domains, we prompt LLMs to generate a bunch of domain descriptors to simulate numerous unseen visual domains (Sec. 4.3). + +# 4.1. Problem Formulation + +Given a training domain data $\mathcal{D}_{train} = \{(\pmb{x}_i, y_i)\}$ , where $\pmb{x}_i \in \mathcal{X}$ are training images and $y_i \in \mathcal{Y}$ are corresponding labels, the concept-based models are trained only on $\mathcal{D}_{train}$ , however evaluated on both $\mathcal{D}_{train}$ and $k$ unseen domains $\{\mathcal{D}_{\text{unseen}}^i\}_{i=1}^k$ . To provide a concept-level explanation, we + +should construct a concept set $\mathcal{C} = \{c_i\}_{i=1}^M$ including $M$ interpretable concepts written by humans or LLMs. To evaluate the effectiveness of our method, we choose the well-studied concept bottleneck models (CBMs) family as the baselines of our methods. As one of the main branches of concept-based models, CBMs first project the image feature to an intermediate concept bottleneck layer to obtain the concept activation vector. Each activation value indicates the presence of the corresponding concept in the input. The forward pipeline of vanilla CBM [11] can be formulated as: + +$$ +\boldsymbol {a} = f _ {C} \left(f _ {I} (\boldsymbol {x})\right), \tag {4} +$$ + +$$ +\hat {y} = f _ {F} (\boldsymbol {a}) = W _ {F} \cdot \boldsymbol {a}, \tag {5} +$$ + +where $\mathbf{a} \in \mathbb{R}^{M}$ is the concept activation vector, $f_{I}: \mathbb{R}^{h \times w} \to \mathbb{R}^{d}$ is the image feature extractor maps the input $\mathbf{x}$ into feature space, $f_{C}: \mathbb{R}^{d} \to \mathbb{R}^{M}$ is the concept projection layer that projects the image feature into concept activation vector, and $f_{F}: \mathbb{R}^{M} \to \mathbb{R}^{N_{y}}$ is the final linear classifier $W_{F}$ maps the concept activation into a final prediction. + +CLIP-CBM. Vanilla CBM requires labor-intensive concept activation annotations to supervise the learning of concept activation. Contrastively, as illustrated in the blue part of Fig. 3, recent CLIP-based CBMs [21, 41, 42] employ the pre-trained CLIP model to extract the image embeddings $E_{I}(x)$ and concept embeddings $E_{T}(c)$ , respectively. Subsequently, CLIP-CBM obtains the concept activations $a$ by computing the similarity between image and concept embeddings, formulated as: + +$$ +a _ {i} = \operatorname {s i m} \left(E _ {I} (x), E _ {T} \left(c _ {i}\right)\right), \tag {6} +$$ + +$$ +\boldsymbol {a} = \left[ a _ {1}, a _ {2}, \dots , a _ {M} \right], \tag {7} +$$ + +where sim indicates the cosine similarity. + +The objective of CLIP-CBM is: + +$$ +\mathcal {L} _ {\mathrm {C E}} (\hat {y}, y) = \mathbb {E} _ {(\boldsymbol {x}, y) \sim \mathcal {D} _ {t r a i n}} \left[ \mathcal {L} \left(W _ {F} \cdot \boldsymbol {a}, y\right) \right], \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{CE}}(\hat{y},y)$ is the cross-entropy loss. + +# 4.2. Domain Descriptors Orthogonality Loss + +As mentioned in Sec. 1, CLIP-CBMs suffer from the OOD generalization problem due to the negative impact of domain-specific concepts. To erase the negative influence of domain-specific concepts, we encourage the learnable class-concept correlation matrix (i.e. the weight of the final linear classification layer), to be orthogonal to the language-guided synthesized domain-specific concept activations. + +Following previous works [15, 25], we decompose the concept activation $\mathbf{a}$ into two components, the domain-specific concept activation vector $\mathbf{a}_{sp}$ and the domain-shared concept vector $\mathbf{a}_{sh}$ , represented as: + +$$ +\boldsymbol {a} = \boldsymbol {a} _ {s p} + \boldsymbol {a} _ {s h}, \tag {9} +$$ + +Our goal is to erase the impact of domain-specific concepts to the final predictions, formulated as: + +$$ +W _ {F} \cdot \boldsymbol {a} _ {s p} = 0. \tag {10} +$$ + +Based on the observations in Sec. 3, we can simulate the $\mathbf{a}_{sp}$ via computing the similarity between language-guided domain shifts and the concepts. Specifically, the language-guided class-level domain shifts $\Delta t_{py}$ are the subtraction between text embeddings of prompts of possible unseen domains and the training domain (e.g. "A sketch of an apple" - "A photo of an apple"). + +$$ +\Delta t \left(p _ {i}, y\right) = E _ {T} \left(\left[ p _ {i}, y \right]\right) - E _ {T} \left(\left[ p _ {\text {t r a i n}}, y \right]\right), \tag {11} +$$ + +$$ +\widehat {\boldsymbol {a} _ {\mathrm {s p}} \left(p _ {i} , y\right)} = E _ {T} (c) \cdot \Delta t \left(p _ {i}, y\right) \tag {12} +$$ + +where $\widehat{\pmb{a}_{\mathrm{sp}}(p_i,y)}$ is simulated domain-specific concept activation and $p_i,p_{train},y$ are the unseen domain descriptor, training domain descriptor and class, respectively. + +For all possible unseen domains $\mathcal{P}$ and candidate classes $\mathcal{Y}$ , based on Eq. (10) and Eq. (12), we derive our proposed domain descriptor orthogonality (DDO) loss: + +$$ +\mathcal {L} _ {\mathrm {D D O}} = \mathbb {E} _ {(p _ {i}, y) \sim \mathcal {P} \times \mathcal {Y}} \left[ \left| W _ {F} \cdot \widehat {\boldsymbol {a} _ {\mathrm {s p}} (p _ {i} , y)} \right| \right] \tag {13} +$$ + +The forward process of DDO is illustrated in the yellow part of Fig. 3. Note that the input to the DDO loss is independent of the current samples and only acts as a regularization term applied to the $W_{F}$ . Therefore, it can be plugged into various CBMs. The final loss is the combination of classification loss and DDO loss, stated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C E}} + \lambda \mathcal {L} _ {\mathrm {D D O}} \tag {14} +$$ + +# 4.3. Generate Domain Descriptors + +To simulate various unseen visual domains as much as possible, we need to write numerous domain descriptors. However, hand-writing these domain descriptors can be costly, and does not scale to large numbers of unseen domains. We can automatically construct this domain descriptor set $\mathcal{P} = \{p_i\}_{i=1}^{N_p}$ by prompting a large language model, such as GPT-3.5 [22], to list possible unseen visual domains except for the training domain. We prompt the large language model with the input: + +Q: Please list visual domains in short phrases as much as possible. + +A: Here are some visual domains in short phrases: real-world photography, clipart illustrations, 3D renders... + +The generated $N_p$ domain descriptors comprise the domain descriptor set $\mathcal{P}$ . The full generation list and further implementation details can be found in the Appendix. + +
ModelConceptMethodCUB-PaintingAwA2-clipartLADA-SculptureLADV-3D
IDOODIDOODIDOODIDOOD
CLIP ZS [26]XX62.2152.7795.7090.2691.2682.0571.8266.29
CLIP LP [26]82.0061.4097.1186.7596.8174.4093.6863.81
CLIP-CBMhumanbaseline78.5150.5495.6981.9196.6670.4492.2160.64
+DDO78.7055.5395.7183.7296.7775.7692.5963.51
PCBM† [42]ConceptNetbaseline75.8554.4197.1784.7797.6076.6994.7165.88
+DDO76.4857.5097.1986.5897.6479.7494.8268.33
LaBO‡ [41]LLMbaseline81.9156.2497.1484.1597.4174.5699.9063.17
+DDO82.3459.6097.2687.6698.1280.0099.9368.01
+ +Table 1. Performance on four single unseen domain benchmarks. For comparison, we list the performance of prevailing CBMs as baselines and report the results of integrating our proposed DDO regularizer into these baselines. $\dagger$ indicates re-implemented with CLIP backbone. $\ddagger$ means visual concepts are obtained with our re-implemented LLM prompts. ID is the performance on photo domains, and OOD is the generalization performance on other domains. + +# 5. Experiments + +# 5.1. Datasets and Baselines + +We conduct experiments on seven domain adaptation benchmarks. Notably, CUB-Painting, PACS, OfficeHome and DomainNet are the previous common benchmarks, while AwA2-clipart, LADA-Sculpture, and LADV-3D are newly proposed in the paper. + +Single unseen domain benchmarks. There are some classical image classification datasets with attribute annotations, which are widely applied for the evaluation of concept-based models. For instance, CUB-Painting is a fine-grained bird classification dataset that consists of two visual domains, the photo domain [32] and the painting domain [35]. Each image has an attribute annotation vector. Following CUB-Painting, we collect images with various other styles for photographic images from Animals with Attributes 2 (AwA2) [39], LAD-animal [47] and LAD-vehicle [47]. Specifically, we introduce three new benchmarks, namely AwA2-clipart, LADA-Sculpture and LADV-3D, which focused on photo $\rightarrow$ clipart, real $\rightarrow$ sculpture and real $\rightarrow$ 3D model, respectively. + +Multiple unseen domain benchmarks. There are also some classical domain adaptation benchmarks with multiple visual domains. PACS [13] is a domain adaptation benchmark with 7 classes and 4 visual styles, including art, cartoon, photo, and sketch. OfficeHome [31] contains 4 domains, i.e. art, clipart, product, realworld, where each domain consists of 65 categories. DomainNet [24] is a domain adaptation dataset containing 345 common classes from six visual domains, i.e. real, clipart, infographic, painting, quickdraw, and sketch. Following previous single domain generalization approaches [19, 48], we train on one source domain and evaluate on the other target domains. For each source domain, we report the average generalization accuracy on all target domains. + +Baselines. We compare our proposed method with the fol- + +lowing baseline approaches. 1) Black-box classifier: $CLIP-ZS$ [26] uses cosine similarity between class textual representation and image representation in CLIP shared space. $CLIP-LP$ [26] fits a linear classifier on top of the frozen CLIP image encoder. 2) CLIP-based CBMs: $CLIP-CBM$ leverages human-written attributes to serve as a concept bank and utilizes CLIP to compute concept activation label. $PCBM$ [42] uses ConceptNet [30] to retrieve concepts relevant to these classes and automatically construct the concept bank. Original PCBM employs different visual backbones on different datasets. For a fair comparison, we re-implement the PCBM with the CLIP backbone. $LaBO$ [41] applies the LLM to generate abundant visual concepts in contrast to PCBM, and we employ a two-layer classifier. For each CBM baseline, we keep their original objectives to solve the optimization. + +# 5.2. Implementation Details + +We prompt GPT3.5-turbo to generate 200 domain descriptors. These domain descriptors are used across all datasets and $\lambda$ is set as 1. We pre-process the domain descriptors into text embeddings by CLIP text encoder. The DDO loss is added at the final classification layer of the CBMs in a plug-in manner. Following LaBO [41], we utilize the official CLIP ViT-L/141 by OpenAI as the default vision backbone to extract image embeddings. We train the linear function $W_{F}$ using the Adam [10] optimizer. All experiments are conducted on a single RTX3090 GPU. More implementation details, hyperparameters and experiments with other vision backbones are listed in the Appendix. + +# 5.3. Generalization to single unseen domain + +We first evaluate our model on single unseen domain benchmarks which focused on photo $\rightarrow$ others generalization. Table. 1 shows in-distribution (ID) and out-of-distribution + +![](images/c88137de33cb773cd60f48116a070e87842cb969e4ab05dc82a3a6d82f722381.jpg) + +![](images/5419fc55f43a079b733ad88b205c3e15027c6a44018ab3448d5fec6e57dd7fde.jpg) + +![](images/a5076c64d7d33fda148f009e1d3fc6d509aca76feff83c3e3593968e7e41d2dd.jpg) + +![](images/089571b2646149e42f7a64ffd1acd1f5c22fbfd6b0b105a38f7857226ad41955.jpg) +Figure 4. OOD performance on three multiple unseen domain generalization benchmarks, PACS, OfficeHome and DomainNet. + +![](images/bae3f9c2605426e23d0ba1a34ae6297c7c4e7998a760a556e58a94bdde2e24f9.jpg) +Figure 5. Ablation studies for effect of the number of domain descriptors. For each quantity, we randomly selected domain descriptors from a total of 200 domain descriptors and averaged the results over five random selections. Results of DomainNet are shown in the Appendix. + +![](images/e22162dda3b3b7eb3cc10541a849d38a5d5c01b2a08377105bf0a6ec5663a637.jpg) + +![](images/b57abaa88d90d1d60f95d6985cf4e9af426168a6979631e0f04870a1980eafe0.jpg) + +(OOD) accuracy on CUB-Painting, AwA2-clipart, LADA-Sculpture and LADV-3D. For all benchmarks, DDO loss is able to preserve or slightly improve the ID accuracy of each CBM while significantly improving their OOD accuracy (average 3 points improvements). We attribute the slight improvements in ID performance to the minor distribution differences between the training and test sets, even in the same visual domain. Moreover, LaBo with DDO loss can achieve competitive and superior performance on both ID and OOD performance with black-box CLIP-LP and narrow the gap with CLIP-ZS on OOD accuracy. These results demonstrate that our proposed LanCE design can effectively improve the generalization capabilities without sacrificing interpretability or classification accuracy. Besides, we notice that all models' generalization performance to sculpture and 3D model images on the LADA-Sculpture and LADV-3D benchmark is limited, indicating generalization from 2D to 3D remains challenging. + +# 5.4. Generalization to multiple unseen domains + +To assess the ability to generalize to multiple unseen domains and the generalization capability when different domains act as the source domain, we conduct experiments on PACS, OfficeHome, and DomainNet. + +Comparison with only using domain-shared concepts. In the absence of a pre-defined human-written concept set, we employ LLM to generate a concept set $\mathcal{C}_{sp + in}$ for each dataset following LaBO [41]. This set contains both domain-specific concepts and domain-shared concepts. The main motivation of this paper is to mitigate the negative + +
ModelMethodCUB-PAwA2-cLADA-SLADV-3D
LanCEbaseline56.2484.1574.5663.17
+DDO(IR)57.6085.5077.7065.46
+DDO59.6087.6680.0068.01
+ +Table 2. Ablation studies for the effect of relevance of the domain descriptors. +DDO(IR) only use the domain-irrelevant descriptors while +DDO use all domain descriptors. + +impact of domain-specific concepts. A natural question that arises is why not simply query the LLM to generate only domain-shared concepts. To explore the differences between this domain-shared approach and our DDO loss, we introduce an additional baseline, where the LLM is queried exclusively to generate a domain-shared concept set $\mathcal{C}_{sh}$ . Specifically, we adopt the single-domain generalization evaluation setting where one source domain is used for training, and the OOD performance is averaged across other domains. As shown in Fig. 4, only using $\mathcal{C}_{sh}$ performs well only in certain cases, such as when the source domains in the PACS dataset are photo or sketch. We attribute this to the fact that the manifestation of domain-shared concepts varies across different domains. For instance, deer antlers in photo (realistic) and cartoon (simplified or exaggerated) domains differ significantly. Moreover, we empirically observe that prompt-based methods are highly sensitive to the wording of human-written prompts. Different prompts lead to significant performance variations, yet they share a common limitation: none consistently improve performance across diverse source domains. In contrast, DDO is more flexible. It adaptively erases the influence of domain-specific concepts while suppressing domain-specific infer + +![](images/bce1fdd097aeb73ee654f4e1196a1a009fb411fa195bacff3386a945a855d0df.jpg) +Figure 6. Qualitative examples about the top-5 concepts, ranking by their weights in the final linear layer $W_{F}$ . Baseline indicates the results of the original LaBo. Domain-specific concepts are highlighted in red. More results are listed in Appendix. + +mation in domain-shared concept activations, leading to stable improvements across a variety of source domains. + +# 5.5. Ablation Studies + +To explore the impact of domain descriptors on final performance, we conduct comprehensive ablation studies, with LaBO as the baseline. Specifically, we compare results with different numbers of domain descriptors and investigate the role of distinct domain descriptors. + +The impact of the numbers of domain descriptors. We evaluate our method with regard to the number of domain descriptors $N_{p}$ as shown in Fig. 5. The ID accuracy remains almost constant as the number of domain descriptors increases, while the OOD accuracy gradually improves with an increasing number of domain descriptors, though with diminishing returns. Specifically, when the number of domain descriptors is below 20, each additional descriptor noticeably enhances OOD accuracy. However, as the number of domain descriptors continues to increase beyond this point, the improvement in OOD accuracy slows progressively. The results show that 100 domain descriptors are enough to achieve decent performance. + +The impact of the relevance of domain descriptors. To investigate which parts of 200 domain descriptors contribute to the most to OOD accuracy improvements, we manually split all domain descriptors into two components, domain-relevant descriptors and domain-irrelevant descriptors based on their relevance with the test unseen domains. For example, for the LADA-Sculpture dataset, domain-relevant domain descriptors are "sculpture", "statue", and "furniture", referring to the generated word cloud mentioned in Fig. 2 and domain-irrelevant parts are the remaining ones. Detailed splits for each dataset are listed in the Appendix. Results are shown in the Table. 2. DDO(IR) indicates that we only use the domain-irrelevant domain descriptors to compute the DDO loss. This evaluation shows that domain-irrelevant domain descriptors can also slightly prompt the OOD accuracy on the unseen domains, although relatively fewer than the domain-relevant counterparts. It + +indicates those domain descriptors are not fully orthonormal, also demonstrating that the DDO loss has the potential to generalize to pure unseen visual domains that are not mentioned in the domain descriptors set $\mathcal{P}$ . + +# 5.6. Qualitative Results + +To evaluate the effectiveness of our proposed LanCE method, we further visualize the Top-5 concepts highly related to randomly selected classes within different unseen domains as illustrated in Fig. 6. The qualitative results demonstrate that our proposed method can significantly decrease the association between the final prediction and the domain-specific concepts, in other words, erasing the impact of domain-specific concepts on the final output. More qualitative results are shown in the Appendix. + +# 6. Conclusion + +In this paper, we propose a language-guided concept-erasing (LanCE) framework that can effectively mitigate the biased association between domain-specific concepts and the final output. We first empirically demonstrate that the pre-trained CLIP can interpret the visual domain shifts into language. And these language-guided domain shifts can distinguish the domain-specific concepts via similarity within the CLIP embedding space. Based on these observations, we introduce a plug-in domain descriptor orthogonality (DDO) loss to erase the negative impact of domain-specific concepts, and that can significantly improve the generalization capabilities of prevailing concept bottleneck models (CBMs). We prompt the large language models (LLMs) to generate a bunch of domain descriptors to simulate numerous possible unseen domains. Moreover, considering that the current domain generalization community focuses more on generalizing to artistic unseen domains, we collect three new benchmarks that includes more difficult scenarios, i.e. 2D→3D, real animals→sculptures. Extensive experiments demonstrate the efficacy of our LanCE. + +# 7. Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407). + +# References + +[1] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2023. 2 +[2] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. International Conference on Learning Representations (ICLR), 2023. 3 +[3] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. Advances in neural information processing systems, 36:79024-79034, 2023. 2, 3 +[4] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3 +[5] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3 +[6] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3 +[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3 +[8] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1 +[9] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept + +activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2 +[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 5 +[12] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3 +[13] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 2, 6 +[14] Jichang Li, Guanbin Li, Yemin Shi, and Yizhou Yu. Cross-domain adaptive clustering for semi-supervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2505-2514, 2021. 3 +[15] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 5 +[16] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3 +[17] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 3 +[18] Jiawei Ma, Yulei Niu, Shiyuan Huang, Guangxing Han, and Shih-Fu Chang. Widin: Wording image for domain-invariant representation in single-source domain generalization. arXiv preprint arXiv:2405.18405, 2024. 3 +[19] Lianbo Ma, Haidong Kang, Guo Yu, Qing Li, and Qiang He. Single-domain generalized predictor for neural architecture search system. IEEE Transactions on Computers, 2024. 3, 6 +[20] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In Proceedings of the IEEE international conference on computer vision, pages 5715-5725, 2017. 3 +[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2, 5 +[22] OpenAI. Chatgpt. https://www.openai.com/chatgpt, 2023.5 +[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 3 + +[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1406-1415, 2019. 2, 6 +[25] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International conference on machine learning, pages 7728-7738. PMLR, 2020. 3, 5 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 6 +[27] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1 +[28] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2 +[29] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 3 +[30] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6 +[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 2, 6 +[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6 +[33] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 3 +[34] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 2, 3 +[35] Sinan Wang, Xinyang Chen, Yunbo Wang, Mingsheng Long, and Jianmin Wang. Progressive adversarial networks for fine-grained domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9213-9222, 2020. 2, 6 +[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 3 + +[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3 +[38] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 3 +[39] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 2, 6 +[40] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3 +[41] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 2, 5, 6, 7 +[42] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 1, 2, 5, 6 +[43] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2 +[44] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23465-23476, 2023. 3 +[45] Zequn Zeng, Jianqiao Sun, Hao Zhang, Tiansheng Wen, Yudi Su, Yan Xie, Zhengjue Wang, and Bo Chen. Hiclescore: A hierarchical metric for image captioning evaluation. arXiv preprint arXiv:2407.18589, 2024. 3 +[46] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero-shot image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14100-14110, 2024. 3 +[47] Bo Zhao, Yanwei Fu, Rui Liang, Jiahong Wu, Yonggang Wang, and Yizhou Wang. A large-scale attribute dataset for zero-shot learning. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 6 +[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI conference on artificial intelligence, pages 21832-21840, 2024. 3, 6 +[49] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 2, 3 \ No newline at end of file diff --git a/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/images.zip b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ce65c57ef8645c98cb5fc7c2ecbb4e6dbe27db8 --- /dev/null +++ b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:873f6633cf500916286c30195652220a6f19cfe08b82bd83cc8b697a80bb9b2d +size 578896 diff --git a/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/layout.json b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b909fd6656f4aa436a1690701ae5334649e81aa --- /dev/null +++ b/2025/Explaining Domain Shifts in Language_ Concept Erasing for Interpretable Image Classification/layout.json @@ -0,0 +1,9000 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 491, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 491, + 140 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 491, + 140 + ], + "type": "text", + "content": "Explaining Domain Shifts in Language: Concept erasing for Interpretable Image Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": "Zequn Zeng" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Yudi Su" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Jianqiao Sun" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Tiansheng Wen" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Hao Zhang" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Zhengjue Wang" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Bo Chen" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Hongwei Liu" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "text", + "content": " and Jiawei Ma" + }, + { + "bbox": [ + 129, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 190, + 537, + 217 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 72, + 190, + 536, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 190, + 536, + 203 + ], + "spans": [ + { + "bbox": [ + 72, + 190, + 536, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 190, + 536, + 203 + ], + "type": "text", + "content": "National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 203, + 537, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 203, + 537, + 217 + ], + "spans": [ + { + "bbox": [ + 72, + 203, + 537, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 203, + 537, + 217 + ], + "type": "text", + "content": "State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 183, + 217, + 426, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 217, + 426, + 232 + ], + "spans": [ + { + "bbox": [ + 183, + 217, + 426, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 183, + 217, + 426, + 232 + ], + "type": "text", + "content": "City University of Hong Kong, Hong Kong SAR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 234, + 440, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 234, + 440, + 244 + ], + "spans": [ + { + "bbox": [ + 160, + 234, + 440, + 244 + ], + "type": "text", + "content": "zqzeng_1@stu.xidian.edu.cn, bchen@mail.xidian.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 299, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 598 + ], + "type": "text", + "content": "Concept-based models can map black-box representations to human-understandable concepts, which makes the decision-making process more transparent and then allows users to understand the reason behind predictions. However, domain-specific concepts often impact the final predictions, which subsequently undermine the model generalization capabilities, and prevent the model from being used in high-stake applications. In this paper, we propose a novel Language-guided Concept-Erasing (LanCE) framework. In particular, we empirically demonstrate that pre-trained vision-language models (VLMs) can approximate distinct visual domain shifts via domain descriptors while prompting large Language Models (LLMs) can easily simulate a wide range of descriptors of unseen visual domains. Then, we introduce a novel plug-in domain descriptor orthogonality (DDO) regularizer to mitigate the impact of these domain-specific concepts on the final predictions. Notably, the DDO regularizer is agnostic to the design of concept-based models and we integrate it into several prevailing models. Through evaluation of domain generalization on four standard benchmarks and three newly introduced benchmarks, we demonstrate that DDO can significantly improve the out-of-distribution (OOD) generalization over the previous state-of-the-art concept-based models. Our code is available at https://github.com/joeyz0z/LanCE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "content": "Concept-based models [11, 27, 42] are prominent approaches for achieving model interpretability, which leverage human-understandable concepts to explain the black-box image representation. Specifically, these models first" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 352, + 283, + 391, + 327 + ], + "blocks": [ + { + "bbox": [ + 343, + 272, + 391, + 282 + ], + "lines": [ + { + "bbox": [ + 343, + 272, + 391, + 282 + ], + "spans": [ + { + "bbox": [ + 343, + 272, + 391, + 282 + ], + "type": "text", + "content": "Photo ( train )" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 283, + 391, + 327 + ], + "lines": [ + { + "bbox": [ + 352, + 283, + 391, + 327 + ], + "spans": [ + { + "bbox": [ + 352, + 283, + 391, + 327 + ], + "type": "image", + "image_path": "7f5a04b89c74ccf53b4b64acc28fdc8f2ae3680c3508afbab689b7ae319a69ea.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 328, + 397, + 336 + ], + "lines": [ + { + "bbox": [ + 340, + 328, + 397, + 336 + ], + "spans": [ + { + "bbox": [ + 340, + 328, + 397, + 336 + ], + "type": "text", + "content": "Predict:Apple" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 341, + 346, + 386, + 371 + ], + "blocks": [ + { + "bbox": [ + 340, + 337, + 397, + 345 + ], + "lines": [ + { + "bbox": [ + 340, + 337, + 397, + 345 + ], + "spans": [ + { + "bbox": [ + 340, + 337, + 397, + 345 + ], + "type": "text", + "content": "Confidence: " + }, + { + "bbox": [ + 340, + 337, + 397, + 345 + ], + "type": "inline_equation", + "content": "98\\%" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 346, + 386, + 371 + ], + "lines": [ + { + "bbox": [ + 341, + 346, + 386, + 371 + ], + "spans": [ + { + "bbox": [ + 341, + 346, + 386, + 371 + ], + "type": "image", + "image_path": "6f691512ec1b34f8b94bd8503f154cf3ec0b170d2576bdc3c206069f90951d98.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 416, + 283, + 456, + 326 + ], + "blocks": [ + { + "bbox": [ + 413, + 272, + 459, + 282 + ], + "lines": [ + { + "bbox": [ + 413, + 272, + 459, + 282 + ], + "spans": [ + { + "bbox": [ + 413, + 272, + 459, + 282 + ], + "type": "text", + "content": "Clipart ( test )" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 283, + 456, + 326 + ], + "lines": [ + { + "bbox": [ + 416, + 283, + 456, + 326 + ], + "spans": [ + { + "bbox": [ + 416, + 283, + 456, + 326 + ], + "type": "image", + "image_path": "3b1c74964fb0523b19cad0b2d40f0e0d9275c1c19f46ef15c278ebcdeab3fc6e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 327, + 463, + 336 + ], + "lines": [ + { + "bbox": [ + 406, + 327, + 463, + 336 + ], + "spans": [ + { + "bbox": [ + 406, + 327, + 463, + 336 + ], + "type": "text", + "content": "Predict:Apple" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 407, + 346, + 453, + 374 + ], + "blocks": [ + { + "bbox": [ + 407, + 337, + 462, + 345 + ], + "lines": [ + { + "bbox": [ + 407, + 337, + 462, + 345 + ], + "spans": [ + { + "bbox": [ + 407, + 337, + 462, + 345 + ], + "type": "text", + "content": "Confidence: " + }, + { + "bbox": [ + 407, + 337, + 462, + 345 + ], + "type": "inline_equation", + "content": "52\\%" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 407, + 346, + 453, + 374 + ], + "lines": [ + { + "bbox": [ + 407, + 346, + 453, + 374 + ], + "spans": [ + { + "bbox": [ + 407, + 346, + 453, + 374 + ], + "type": "image", + "image_path": "f76669c91a4b991a4dcaa88e97cdc032d91fdc96d4c7641a86fc10173edda871.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 376, + 511, + 385 + ], + "lines": [ + { + "bbox": [ + 355, + 376, + 511, + 385 + ], + "spans": [ + { + "bbox": [ + 355, + 376, + 511, + 385 + ], + "type": "text", + "content": "(a) Concept-based predictions on three domains." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 479, + 283, + 521, + 327 + ], + "blocks": [ + { + "bbox": [ + 480, + 272, + 525, + 282 + ], + "lines": [ + { + "bbox": [ + 480, + 272, + 525, + 282 + ], + "spans": [ + { + "bbox": [ + 480, + 272, + 525, + 282 + ], + "type": "text", + "content": "Sketch (test)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 479, + 283, + 521, + 327 + ], + "lines": [ + { + "bbox": [ + 479, + 283, + 521, + 327 + ], + "spans": [ + { + "bbox": [ + 479, + 283, + 521, + 327 + ], + "type": "image", + "image_path": "cb5f2075e5070feb9afe5349637f0d76fc797739e2e3840c133b2c5abdf3b255.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 328, + 528, + 336 + ], + "lines": [ + { + "bbox": [ + 473, + 328, + 528, + 336 + ], + "spans": [ + { + "bbox": [ + 473, + 328, + 528, + 336 + ], + "type": "text", + "content": "Predict: Pear X" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 473, + 346, + 519, + 374 + ], + "blocks": [ + { + "bbox": [ + 473, + 337, + 529, + 345 + ], + "lines": [ + { + "bbox": [ + 473, + 337, + 529, + 345 + ], + "spans": [ + { + "bbox": [ + 473, + 337, + 529, + 345 + ], + "type": "text", + "content": "Confidence: " + }, + { + "bbox": [ + 473, + 337, + 529, + 345 + ], + "type": "inline_equation", + "content": "37\\%" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 346, + 519, + 374 + ], + "lines": [ + { + "bbox": [ + 473, + 346, + 519, + 374 + ], + "spans": [ + { + "bbox": [ + 473, + 346, + 519, + 374 + ], + "type": "image", + "image_path": "337d7025f18ded1eb7875fa186073edfc564b2ee06d8bd1f3225b1225fe0e0be.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 316, + 395, + 433, + 492 + ], + "blocks": [ + { + "bbox": [ + 355, + 387, + 408, + 395 + ], + "lines": [ + { + "bbox": [ + 355, + 387, + 408, + 395 + ], + "spans": [ + { + "bbox": [ + 355, + 387, + 408, + 395 + ], + "type": "text", + "content": "Concept: Red color" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 395, + 433, + 492 + ], + "lines": [ + { + "bbox": [ + 316, + 395, + 433, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 433, + 492 + ], + "type": "image", + "image_path": "ca7c2452e649077e4e30bf2dd73d205167f62ab5cc2266b873b3e044f0ef2ac4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 495, + 547, + 504 + ], + "lines": [ + { + "bbox": [ + 320, + 495, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 320, + 495, + 547, + 504 + ], + "type": "text", + "content": "(b) Distribution shifts of domain-specific and domain-shared concepts." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 506, + 555, + 605 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 605 + ], + "type": "text", + "content": "Figure 1. Domain shifts in concept space. (a) Given the images of an apple in different visual domains, the prediction confidence of a concept-based model, trained on the photo domain, degrades due to the missing of concepts. (b) Distribution comparison of concept activation value (image-concept similarity computed via CLIP [26]) between photo domain and sketch domain, for two concepts, i.e., \"red color\" and \"round shape\", respectively. JS divergence indicates the distance between two distributions and tends to be larger for domain-specific concepts (e.g. \"red color\")." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 436, + 395, + 553, + 492 + ], + "blocks": [ + { + "bbox": [ + 470, + 387, + 531, + 395 + ], + "lines": [ + { + "bbox": [ + 470, + 387, + 531, + 395 + ], + "spans": [ + { + "bbox": [ + 470, + 387, + 531, + 395 + ], + "type": "text", + "content": "Concept: Round shape" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 395, + 553, + 492 + ], + "lines": [ + { + "bbox": [ + 436, + 395, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 436, + 395, + 553, + 492 + ], + "type": "image", + "image_path": "085228f4148871c7fc09b0719f54b6fb53848a44bcccf3e3293627f950a3c67b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "map the image feature to a concept activation vector in a concept space (each dimension corresponds to an interpretable concept) and then use the concept activation vector to predict the final output. Recently, equipped with pretrained vision-language models (VLMs) [8, 26], concept-based models [41, 42] can obtain the concept activation value by calculating the similarity between image embed" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 227, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 227, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 227, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding authors" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9517" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "type": "text", + "content": "dings and textual concept embeddings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 327 + ], + "type": "text", + "content": "Nevertheless, current concept-based models still struggle to handle domain shifts encountered during inference. The distinct distribution of visual concepts across different domains presents a significant generalization challenge for these models. As shown in Fig. 1a, the concept-based models trained on the photographic images usually excel at associating the discriminative visual clues – such as red color, waxy texture, and round shape – with specific classes like “apple”, akin to human beings. However, these models experience substantial performance degradation when applied to images in unseen visual domains, such as clipart or sketches, due to the absence of domain-specific texture and color concepts. These domain-specific concepts (e.g. “red color”) generally exhibit a larger distributional difference between visual domains than domain-shared concepts (e.g. “round shape”), as illustrated in Fig. 1b. Consequently, the concept-based models trained on a single domain are inevitably biased to associate those domain-specific concepts with the final predictions, thereby limiting their out-of-distribution (OOD) generalization capabilities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 496 + ], + "type": "text", + "content": "To mitigate the issue of domain shifts in the concept space, a straightforward solution is to involve human experts in filtering out the domain-specific concepts and retaining only domain-shared concepts for classification tasks, which is nevertheless labor-intensive. Another possible solution is to diversify the training domain [1, 3] to mitigate the impact of the domain-specific concepts. However, it is impractical to cover all possible unseen domains during training exhaustively. Additionally, current domain generalization approaches [49] can not be directly applied to concept-based models due to the lack of interpretability of semantics of the learned representation [34]. Therefore, automating the identification and elimination of domain-specific concepts remains an under-explored area." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": "In this paper, we propose a Language-guided ConceptErasing framework, namely LanCE, to enhance the OOD generalization capabilities of current concept-based models. Specifically, we empirically demonstrate that pretrained VLMs, such as CLIP, can interpret the domain shifts through language, i.e., domain shifts can be approximated by a set of domain descriptors generated by large language models (LLMs). To alleviate the biased association between final predictions and domain-specific concepts, we introduce a plug-in domain descriptor orthogonality (DDO) loss that reduces the fluctuation of concept activation caused by these language-guided domain shifts. We evaluate our method on seven benchmarks, including four common benchmarks (CUB-Painting [35], PACS [13], OfficeHome [31] and DomainNet [24]), and three new proposed benchmarks (AwA2-clipart, LADA-Sculpture and LADV3D) based on several existing visual classification datasets with concept annotations [39, 47]. These newly-collected" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "datasets are gathered from the web and manually filtered. Extensive experiments demonstrate that the proposed approach can preserve the In-distribution (ID) classification capabilities and significantly improve the OOD generalization capabilities of prevailing concept-based models. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 553, + 312 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "type": "text", + "content": "- We provide a concept-level explanation for the domain shifts and empirically reveal that VLMs can represent visual domain shifts into language space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "text", + "content": "- Based on this observation, we propose a plug-in domain descriptor orthogonality regularizer to learn a domain-shared concept-based model without changing the model architecture and increasing training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 228, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 312 + ], + "type": "text", + "content": "- We introduce three new concept-based OOD generalization benchmarks including more challenging scenarios such as natural photos " + }, + { + "bbox": [ + 314, + 228, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 314, + 228, + 553, + 312 + ], + "type": "text", + "content": " 3D-renderers and real animals " + }, + { + "bbox": [ + 314, + 228, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 314, + 228, + 553, + 312 + ], + "type": "text", + "content": " scluptures. Extensive experiments on four common and three new benchmarks demonstrate that our proposed method significantly improves the OOD generalization capabilities of current concept-based models." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 343, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 677 + ], + "type": "text", + "content": "Concept-based models are one of the mainstream interpretable approaches that aim to understand models by associating black box features with meaningful concepts. Current concept-based models can broadly be divided into two categories, concept bottleneck models (CBMs) [11] which map the image features into an intermediate concept bottleneck layer whose neurons indicate concept activation values of pre-defined concepts, and concept activation vectors (CAVs) [9] which represent concepts as normal vectors of decision boundaries that distinguish positive and negative samples of a concept. Vanilla CBM requires fine-grained and precise concept annotation on the concept bottleneck layer and CAVs need positive and negative samples of each concept to learn meaningful concept representations. Many recent approaches are proposed to tackle the above concerns from two perspectives. On the one hand, many methods generate the concepts derived from ConceptNet [42], LLMs [21, 41] or multimodal datasets [28, 43] to automate the construction of the concept bottleneck layer. On the other hand, to remove the reliance on human-annotated concept labels, these methods deploy pre-trained VLMs such as CLIP to map concepts to text embeddings and compute image-concept similarity to serve as concept activation annotations. However, concepts of the above approaches are typically domain-sensitive and the corresponding concept activations have large distributional differences across various domains. They often fail to handle domain shifts when applied to unseen domains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Domain adaptation & generalization aims to handle the domain shifts between training and test data. In contrast to traditional deep learning based on in-distribution assump-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9518" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 75, + 352, + 237 + ], + "blocks": [ + { + "bbox": [ + 60, + 75, + 352, + 237 + ], + "lines": [ + { + "bbox": [ + 60, + 75, + 352, + 237 + ], + "spans": [ + { + "bbox": [ + 60, + 75, + 352, + 237 + ], + "type": "image", + "image_path": "107be90093f88dd5cd39df0c8266f606815b36a9781746acb40b2ff80fe28d5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 244, + 295, + 256 + ], + "lines": [ + { + "bbox": [ + 92, + 244, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 92, + 244, + 295, + 256 + ], + "type": "text", + "content": "(a) Interpret the visual domain shifts into language." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 367, + 89, + 552, + 155 + ], + "blocks": [ + { + "bbox": [ + 372, + 76, + 541, + 86 + ], + "lines": [ + { + "bbox": [ + 372, + 76, + 541, + 86 + ], + "spans": [ + { + "bbox": [ + 372, + 76, + 541, + 86 + ], + "type": "text", + "content": "A sketch of an apple A photo of an apple" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 89, + 552, + 155 + ], + "lines": [ + { + "bbox": [ + 367, + 89, + 552, + 155 + ], + "spans": [ + { + "bbox": [ + 367, + 89, + 552, + 155 + ], + "type": "image", + "image_path": "20f8d5d28f97dde4756e151b9ae81b85edd82bfc39670c46ebdcb660375525d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 365, + 174, + 550, + 240 + ], + "blocks": [ + { + "bbox": [ + 372, + 161, + 541, + 171 + ], + "lines": [ + { + "bbox": [ + 372, + 161, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 372, + 161, + 541, + 171 + ], + "type": "text", + "content": "A bronze sculpture of a cat A photo of a cat" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 174, + 550, + 240 + ], + "lines": [ + { + "bbox": [ + 365, + 174, + 550, + 240 + ], + "spans": [ + { + "bbox": [ + 365, + 174, + 550, + 240 + ], + "type": "image", + "image_path": "ae7b0b6976a6ebfe07f1fed017f8f45baa696c0683682127b453bf8889497840.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 244, + 534, + 256 + ], + "lines": [ + { + "bbox": [ + 365, + 244, + 534, + 256 + ], + "spans": [ + { + "bbox": [ + 365, + 244, + 534, + 256 + ], + "type": "text", + "content": "(b) Discover the domain-specific concepts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 260, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 54, + 260, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 260, + 555, + 316 + ], + "type": "text", + "content": "Figure 2. (a) We empirically demonstrate that the visual domain shift can be interpreted in language. For each class, we obtain the caption from the embedding difference of images from two domains. Then, by aggregating the captions across the classes, the keywords regarding the domain shift can be highlighted. At the same time, (b) the domain-specific concepts can be discovered from language descriptions of the different domains. In detail, they have higher similarities with the difference of domain-related class descriptions, i.e., textural domain shift, in the CLIP embedding space. More analyses are shown in the Appendix." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 546 + ], + "type": "text", + "content": "tion, domain adaptation methods [4, 20] target to improve the performance on the target domain when only a few samples [14] or unlabeled data [40] from target domains are available. Instead, domain generalization methods [49] focused on more difficult scenarios in which no target domain data are available during the training phase. The setting of this paper belongs to the single domain generalization scenario, i.e. only one single domain is available during training [19, 36, 48]. While the domain generalization area is well studied, there is still researcher doubt about the interpretability of domain generalization methods [34]. Although disentanglement-based domain generalization methods [25, 33] decompose a feature into domain-shared and domain-specific parts, there is still a lack of deep understanding of the semantics of the learned features in domain generalization models. Our proposed method provides a concept-level explanation of domain shifts and takes steps to interpretable domain generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 547, + 296, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 296, + 655 + ], + "type": "text", + "content": "Vision-Language Pretraining aims to bridge the gap between image and text representations. Compared with vision-only pretraining methods [5, 6, 17], vision-language pertaining like CLIP [26] achieved remarkable success and showed superior performance on some high-level vision understanding tasks [16, 44, 46] and multi-modality tasks [29, 38, 45]. Our empirical findings further validate that CLIP can interpret a variety of domain shifts into a set of domain descriptors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 669, + 194, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 194, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 194, + 682 + ], + "type": "text", + "content": "3. Empirical Observations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "This section demonstrates that the pre-trained VLMs can approximate the visual domain shifts via domain descrip" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 378 + ], + "type": "text", + "content": "tors. Meanwhile, these language-guided domain shifts can effectively distinguish the domain-specific concepts as illustrated in Fig. 2. It serves as the key insight for our proposed LanCE method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 380, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 475 + ], + "type": "text", + "content": "Image-text alignment. CLIP trained on multi-modal contrastive learning, can effectively bridge the gap between image representation and text representation. The alignment facilitates a new approach to manipulate images using language, applicable for various vision tasks, e.g. image editing [7, 23], style transfer [12, 37], data augmentation [2, 3, 18]. The CLIP image-text similarity can be computed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 484, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 484, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 362, + 484, + 553, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L I P}} (I, T) = \\operatorname {s i m} \\left(E _ {I} (I), E _ {T} (T)\\right) \\tag {1}", + "image_path": "ed64935953cd7366fc49ccf2f2930873b101d012e6ebdd3ee3efaf299ec4cdb6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "text", + "content": " mean the image encoder and text encoder of CLIP, respectively. sim indicate the cosine similarity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "text", + "content": "Interpreting the visual domain shifts. Inspired by previous works [2, 23], we assume that the visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space. To validate this hypothesis, we employ a CLIP-based zero-shot captioner [44] to translate the visual embedding differences into language, as illustrated in Fig. 2(a). Let " + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\{I_c^{src}\\}" + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\{I_c^{tgt}\\}" + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "text", + "content": " denote the images of the same class " + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "text", + "content": " from the source domain and the target domain, respectively. The class-level visual domain gap " + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\Delta d_{c}" + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "text", + "content": " and corresponding description " + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 313, + 534, + 556, + 666 + ], + "type": "text", + "content": " can be generated by solving the following maximization problem:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 677, + 553, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 677, + 553, + 695 + ], + "spans": [ + { + "bbox": [ + 369, + 677, + 553, + 695 + ], + "type": "interline_equation", + "content": "\\max _ {t _ {c}} \\mathcal {L} _ {\\text {C L I P}} \\left(\\Delta d _ {c}, t _ {c}\\right), \\tag {2}", + "image_path": "8a40e0f86a83210f63471d3ee4c0ec5abc82d0190d8af295c5dc82a4988f146b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 696, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 696, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 361, + 696, + 553, + 711 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {d} _ {c} = \\mathbb {E} \\left(E _ {I} \\left(I _ {c} ^ {t g t}\\right)\\right) - \\mathbb {E} \\left(E _ {I} \\left(I _ {c} ^ {s r c}\\right)\\right), \\tag {3}", + "image_path": "27f142d2d6aca621273f19cbd5c48c00b46854969f1e61d883b9c754adece63e.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9519" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 261 + ], + "type": "image", + "image_path": "7adb23ec0613b26e3cc1008ae85f6ff553500a5e384e47db3762bf23283ae603.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "text", + "content": "Figure 3. Overview of the LanCE. Blue part is the data flow of vanilla CLIP-CBMs (Sec. 4.1). To provide concept-level explanations, we first construct a human-written or LLMs-generated concept set " + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "text", + "content": " and extract the concept embeddings via the frozen CLIP text encoder. Given an image, we can extract the image embeddings via the frozen CLIP image encoder. The concept activations are the cosine similarity between image embeddings and concept embeddings. A learnable linear layer " + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "inline_equation", + "content": "W_{F}" + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "text", + "content": " is fitted on top of the concept activation vector and is responsible for predicting the final class and is optimized via cross-entropy loss. Yellow part is the data flow of our proposed DDO regularizer (Sec. 4.2). Similarly, we first construct a domain descriptor set (Sec. 4.3) to obtain the language-guided domain shifts and then simulate the domain-specific concept activations. To erase the effect of domain-specific concepts, the DDO regularizer encourages the orthogonality between the class-concept correlation matrix " + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "inline_equation", + "content": "W_{F}" + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "text", + "content": " (i.e. the final linear weight) and domain-specific concept activation " + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{a}_{\\mathrm{sp}}}" + }, + { + "bbox": [ + 54, + 268, + 555, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "text", + "content": "As shown in Fig. 2(a), we produce high-matching (CLIP-Score " + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "inline_equation", + "content": ">0.9" + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "text", + "content": ") descriptions " + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "text", + "content": " that effectively interpret the class-level visual domain gap " + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "inline_equation", + "content": "\\Delta d_c" + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "text", + "content": " between sketch images, sculpture images and photographic images. We observe that these class-level domain descriptors are typically descriptions of the style difference of the two domains, regarding class-level domain descriptors. To further get the global difference between the two domains, we aggregate all class-level domain descriptors and transfer them into word cloud format, where the word size is proportional to the word frequency in the generated domain description corpus. As a result, we empirically find that global visual domain descriptors " + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "inline_equation", + "content": "\\Delta d" + }, + { + "bbox": [ + 54, + 372, + 294, + 539 + ], + "type": "text", + "content": " primarily consist of style words such as \"sketch and sculpture\" and background words e.g. \"white\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Discover the domain-specific concepts. Since we have revealed that visual domain shifts and descriptions of domain changes are aligned within the CLIP embedding space, we can synthesize the visual domain shifts via domain descriptors. Subsequently, we can leverage these language-guided domain shifts to discover domain-specific concepts. Specifically, we compute the similarity between textual domain shift embeddings (e.g. \"A sketch of an apple\" - \"A photo of an apple\") and all candidate concept embeddings extracted by CLIP text encoder. The results in Fig. 2(b) show that domain-specific concepts like \"shiny, waxy surface texture\" generally exhibit a higher similarity score due to their substantial distribution shift between those two domains while those domain-shared concepts such as \"Leaf around" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 372, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 407 + ], + "type": "text", + "content": "the stem\" have relatively lower similarity score. Consequently, we can distinguish the domain-specific concepts via language-guided domain shifts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 417, + 370, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 370, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 370, + 430 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 437, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 555, + 628 + ], + "type": "text", + "content": "To improve the generalization capabilities of prevailing concept-based models, as shown in Fig. 3, we propose a novel LanCE framework. In this section, we first formulate the problem and introduce Concept bottleneck models (CBMs), which can effectively map the black-box visual features to human-understandable concept space to interpret the final outputs (Sec. 4.1). However, domain-specific concepts will undermine the generalization capabilities. Then, to alleviate the biased association between domain-specific concepts and class predictions within the learnable linear classifier, we propose a domain descriptor orthogonality loss to improve the OOD classification performances of concept-based models (Sec. 4.2). Furthermore, considering all possible visual domains, we prompt LLMs to generate a bunch of domain descriptors to simulate numerous unseen visual domains (Sec. 4.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 635, + 440, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 440, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 440, + 647 + ], + "type": "text", + "content": "4.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": "Given a training domain data " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train} = \\{(\\pmb{x}_i, y_i)\\}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " are training images and " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " are corresponding labels, the concept-based models are trained only on " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": ", however evaluated on both " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " unseen domains " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{\\text{unseen}}^i\\}_{i=1}^k" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": ". To provide a concept-level explanation, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9520" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "should construct a concept set " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_i\\}_{i=1}^M" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " including " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " interpretable concepts written by humans or LLMs. To evaluate the effectiveness of our method, we choose the well-studied concept bottleneck models (CBMs) family as the baselines of our methods. As one of the main branches of concept-based models, CBMs first project the image feature to an intermediate concept bottleneck layer to obtain the concept activation vector. Each activation value indicates the presence of the corresponding concept in the input. The forward pipeline of vanilla CBM [11] can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 212, + 295, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 212, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 129, + 212, + 295, + 224 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} = f _ {C} \\left(f _ {I} (\\boldsymbol {x})\\right), \\tag {4}", + "image_path": "c5aaac51d66a02b51cb6bdc0a5e52b62db7703b3824892345d4208b4c4397230.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 226, + 294, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 294, + 239 + ], + "type": "interline_equation", + "content": "\\hat {y} = f _ {F} (\\boldsymbol {a}) = W _ {F} \\cdot \\boldsymbol {a}, \\tag {5}", + "image_path": "d5bd0080cfac6c35e70361c0abd50bbd40dabe1e6e9da3006117af079b57a412.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": " is the concept activation vector, " + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "inline_equation", + "content": "f_{I}: \\mathbb{R}^{h \\times w} \\to \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": " is the image feature extractor maps the input " + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": " into feature space, " + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "inline_equation", + "content": "f_{C}: \\mathbb{R}^{d} \\to \\mathbb{R}^{M}" + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": " is the concept projection layer that projects the image feature into concept activation vector, and " + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "inline_equation", + "content": "f_{F}: \\mathbb{R}^{M} \\to \\mathbb{R}^{N_{y}}" + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": " is the final linear classifier " + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "inline_equation", + "content": "W_{F}" + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": " maps the concept activation into a final prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": "CLIP-CBM. Vanilla CBM requires labor-intensive concept activation annotations to supervise the learning of concept activation. Contrastively, as illustrated in the blue part of Fig. 3, recent CLIP-based CBMs [21, 41, 42] employ the pre-trained CLIP model to extract the image embeddings " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "E_{I}(x)" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": " and concept embeddings " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "E_{T}(c)" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": ", respectively. Subsequently, CLIP-CBM obtains the concept activations " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": " by computing the similarity between image and concept embeddings, formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 122, + 434, + 295, + 448 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 434, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 122, + 434, + 295, + 448 + ], + "type": "interline_equation", + "content": "a _ {i} = \\operatorname {s i m} \\left(E _ {I} (x), E _ {T} \\left(c _ {i}\\right)\\right), \\tag {6}", + "image_path": "16b6c622b7b88ffac5a4d366e94c797ad4d08d42970dfe3cc3350100327cf15c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 449, + 294, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 449, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 125, + 449, + 294, + 463 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} = \\left[ a _ {1}, a _ {2}, \\dots , a _ {M} \\right], \\tag {7}", + "image_path": "af967c48ed5fb6f59f1ebb186ccf9195164a3df82716baf2335546be44aa6eb7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 470, + 223, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 223, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 223, + 481 + ], + "type": "text", + "content": "where sim indicates the cosine similarity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 481, + 198, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 481, + 198, + 493 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 198, + 493 + ], + "type": "text", + "content": "The objective of CLIP-CBM is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 501, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 501, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 86, + 501, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} (\\hat {y}, y) = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D} _ {t r a i n}} \\left[ \\mathcal {L} \\left(W _ {F} \\cdot \\boldsymbol {a}, y\\right) \\right], \\tag {8}", + "image_path": "36f7850af75dc94f99f0166119e7da265688f888a3e2488ed33a5e2994495604.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 227, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 227, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 227, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 227, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}(\\hat{y},y)" + }, + { + "bbox": [ + 55, + 521, + 227, + 534 + ], + "type": "text", + "content": " is the cross-entropy loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 540, + 266, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 266, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 266, + 553 + ], + "type": "text", + "content": "4.2. Domain Descriptors Orthogonality Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 558, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 642 + ], + "type": "text", + "content": "As mentioned in Sec. 1, CLIP-CBMs suffer from the OOD generalization problem due to the negative impact of domain-specific concepts. To erase the negative influence of domain-specific concepts, we encourage the learnable class-concept correlation matrix (i.e. the weight of the final linear classification layer), to be orthogonal to the language-guided synthesized domain-specific concept activations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "text", + "content": "Following previous works [15, 25], we decompose the concept activation " + }, + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "text", + "content": " into two components, the domain-specific concept activation vector " + }, + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{sp}" + }, + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "text", + "content": " and the domain-shared concept vector " + }, + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{sh}" + }, + { + "bbox": [ + 55, + 642, + 296, + 690 + ], + "type": "text", + "content": ", represented as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 699, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 699, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 141, + 699, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} = \\boldsymbol {a} _ {s p} + \\boldsymbol {a} _ {s h}, \\tag {9}", + "image_path": "ed8789cf4b8b782aca8a9993a5876d84d42533da29d379b7b32c4f485a023f46.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "Our goal is to erase the impact of domain-specific concepts to the final predictions, formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 403, + 106, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 106, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 403, + 106, + 553, + 118 + ], + "type": "interline_equation", + "content": "W _ {F} \\cdot \\boldsymbol {a} _ {s p} = 0. \\tag {10}", + "image_path": "a719bea191eb836f13790ed53129aa9371c555d5f14ca0de837dee2e783a027d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 128, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 555, + 212 + ], + "type": "text", + "content": "Based on the observations in Sec. 3, we can simulate the " + }, + { + "bbox": [ + 313, + 128, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{sp}" + }, + { + "bbox": [ + 313, + 128, + 555, + 212 + ], + "type": "text", + "content": " via computing the similarity between language-guided domain shifts and the concepts. Specifically, the language-guided class-level domain shifts " + }, + { + "bbox": [ + 313, + 128, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\Delta t_{py}" + }, + { + "bbox": [ + 313, + 128, + 555, + 212 + ], + "type": "text", + "content": " are the subtraction between text embeddings of prompts of possible unseen domains and the training domain (e.g. \"A sketch of an apple\" - \"A photo of an apple\")." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 220, + 553, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 220, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 339, + 220, + 553, + 234 + ], + "type": "interline_equation", + "content": "\\Delta t \\left(p _ {i}, y\\right) = E _ {T} \\left(\\left[ p _ {i}, y \\right]\\right) - E _ {T} \\left(\\left[ p _ {\\text {t r a i n}}, y \\right]\\right), \\tag {11}", + "image_path": "32d91949b31599843ae41484238d8c7817536047ed66d19bb771dd56c19adbba.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 387, + 236, + 553, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 236, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 387, + 236, + 553, + 253 + ], + "type": "interline_equation", + "content": "\\widehat {\\boldsymbol {a} _ {\\mathrm {s p}} \\left(p _ {i} , y\\right)} = E _ {T} (c) \\cdot \\Delta t \\left(p _ {i}, y\\right) \\tag {12}", + "image_path": "4b5ec93c12128fa2f3171719e0a26a78cc96a59929ec6f0a2490309ad032fb80.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 262, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 262, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{a}_{\\mathrm{sp}}(p_i,y)}" + }, + { + "bbox": [ + 313, + 262, + 553, + 300 + ], + "type": "text", + "content": " is simulated domain-specific concept activation and " + }, + { + "bbox": [ + 313, + 262, + 553, + 300 + ], + "type": "inline_equation", + "content": "p_i,p_{train},y" + }, + { + "bbox": [ + 313, + 262, + 553, + 300 + ], + "type": "text", + "content": " are the unseen domain descriptor, training domain descriptor and class, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 300, + 554, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 336 + ], + "type": "text", + "content": "For all possible unseen domains " + }, + { + "bbox": [ + 313, + 300, + 554, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 300, + 554, + 336 + ], + "type": "text", + "content": " and candidate classes " + }, + { + "bbox": [ + 313, + 300, + 554, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 300, + 554, + 336 + ], + "type": "text", + "content": ", based on Eq. (10) and Eq. (12), we derive our proposed domain descriptor orthogonality (DDO) loss:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 348, + 344, + 553, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 344, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 348, + 344, + 553, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D D O}} = \\mathbb {E} _ {(p _ {i}, y) \\sim \\mathcal {P} \\times \\mathcal {Y}} \\left[ \\left| W _ {F} \\cdot \\widehat {\\boldsymbol {a} _ {\\mathrm {s p}} (p _ {i} , y)} \\right| \\right] \\tag {13}", + "image_path": "436883080f0cac1b296d0622f009864f6b79884d89a5ae1d35747ba150dee7ab.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 373, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 444 + ], + "type": "text", + "content": "The forward process of DDO is illustrated in the yellow part of Fig. 3. Note that the input to the DDO loss is independent of the current samples and only acts as a regularization term applied to the " + }, + { + "bbox": [ + 313, + 373, + 554, + 444 + ], + "type": "inline_equation", + "content": "W_{F}" + }, + { + "bbox": [ + 313, + 373, + 554, + 444 + ], + "type": "text", + "content": ". Therefore, it can be plugged into various CBMs. The final loss is the combination of classification loss and DDO loss, stated as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 394, + 453, + 553, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 453, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 394, + 453, + 553, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} + \\lambda \\mathcal {L} _ {\\mathrm {D D O}} \\tag {14}", + "image_path": "c33a6aa1f2fca67feee331d4f18f9dc9ba581976385230656473b3994dfb3400.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 475, + 478, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 478, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 478, + 487 + ], + "type": "text", + "content": "4.3. Generate Domain Descriptors" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 493, + 555, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 603 + ], + "type": "text", + "content": "To simulate various unseen visual domains as much as possible, we need to write numerous domain descriptors. However, hand-writing these domain descriptors can be costly, and does not scale to large numbers of unseen domains. We can automatically construct this domain descriptor set " + }, + { + "bbox": [ + 313, + 493, + 555, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_i\\}_{i=1}^{N_p}" + }, + { + "bbox": [ + 313, + 493, + 555, + 603 + ], + "type": "text", + "content": " by prompting a large language model, such as GPT-3.5 [22], to list possible unseen visual domains except for the training domain. We prompt the large language model with the input:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 333, + 610, + 534, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 610, + 534, + 633 + ], + "spans": [ + { + "bbox": [ + 333, + 610, + 534, + 633 + ], + "type": "text", + "content": "Q: Please list visual domains in short phrases as much as possible." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 333, + 634, + 534, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 634, + 534, + 669 + ], + "spans": [ + { + "bbox": [ + 333, + 634, + 534, + 669 + ], + "type": "text", + "content": "A: Here are some visual domains in short phrases: real-world photography, clipart illustrations, 3D renders..." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "The generated " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " domain descriptors comprise the domain descriptor set " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ". The full generation list and further implementation details can be found in the Appendix." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9521" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 529, + 212 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 529, + 212 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 529, + 212 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 529, + 212 + ], + "type": "table", + "html": "
ModelConceptMethodCUB-PaintingAwA2-clipartLADA-SculptureLADV-3D
IDOODIDOODIDOODIDOOD
CLIP ZS [26]XX62.2152.7795.7090.2691.2682.0571.8266.29
CLIP LP [26]82.0061.4097.1186.7596.8174.4093.6863.81
CLIP-CBMhumanbaseline78.5150.5495.6981.9196.6670.4492.2160.64
+DDO78.7055.5395.7183.7296.7775.7692.5963.51
PCBM† [42]ConceptNetbaseline75.8554.4197.1784.7797.6076.6994.7165.88
+DDO76.4857.5097.1986.5897.6479.7494.8268.33
LaBO‡ [41]LLMbaseline81.9156.2497.1484.1597.4174.5699.9063.17
+DDO82.3459.6097.2687.6698.1280.0099.9368.01
", + "image_path": "36ebfe753fa27ff87d2c8cce652addffb1832aa49771b6d9f89940170a5bb395.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 213, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 257 + ], + "type": "text", + "content": "Table 1. Performance on four single unseen domain benchmarks. For comparison, we list the performance of prevailing CBMs as baselines and report the results of integrating our proposed DDO regularizer into these baselines. " + }, + { + "bbox": [ + 55, + 213, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 213, + 555, + 257 + ], + "type": "text", + "content": " indicates re-implemented with CLIP backbone. " + }, + { + "bbox": [ + 55, + 213, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 213, + 555, + 257 + ], + "type": "text", + "content": " means visual concepts are obtained with our re-implemented LLM prompts. ID is the performance on photo domains, and OOD is the generalization performance on other domains." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 137, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 137, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 137, + 280 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 185, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 185, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 185, + 298 + ], + "type": "text", + "content": "5.1. Datasets and Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 365 + ], + "type": "text", + "content": "We conduct experiments on seven domain adaptation benchmarks. Notably, CUB-Painting, PACS, OfficeHome and DomainNet are the previous common benchmarks, while AwA2-clipart, LADA-Sculpture, and LADV-3D are newly proposed in the paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "text", + "content": "Single unseen domain benchmarks. There are some classical image classification datasets with attribute annotations, which are widely applied for the evaluation of concept-based models. For instance, CUB-Painting is a fine-grained bird classification dataset that consists of two visual domains, the photo domain [32] and the painting domain [35]. Each image has an attribute annotation vector. Following CUB-Painting, we collect images with various other styles for photographic images from Animals with Attributes 2 (AwA2) [39], LAD-animal [47] and LAD-vehicle [47]. Specifically, we introduce three new benchmarks, namely AwA2-clipart, LADA-Sculpture and LADV-3D, which focused on photo " + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "text", + "content": " clipart, real " + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "text", + "content": " sculpture and real " + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "text", + "content": " 3D model, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "content": "Multiple unseen domain benchmarks. There are also some classical domain adaptation benchmarks with multiple visual domains. PACS [13] is a domain adaptation benchmark with 7 classes and 4 visual styles, including art, cartoon, photo, and sketch. OfficeHome [31] contains 4 domains, i.e. art, clipart, product, realworld, where each domain consists of 65 categories. DomainNet [24] is a domain adaptation dataset containing 345 common classes from six visual domains, i.e. real, clipart, infographic, painting, quickdraw, and sketch. Following previous single domain generalization approaches [19, 48], we train on one source domain and evaluate on the other target domains. For each source domain, we report the average generalization accuracy on all target domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Baselines. We compare our proposed method with the fol-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "content": "lowing baseline approaches. 1) Black-box classifier: " + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "inline_equation", + "content": "CLIP-ZS" + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "content": " [26] uses cosine similarity between class textual representation and image representation in CLIP shared space. " + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "inline_equation", + "content": "CLIP-LP" + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "content": " [26] fits a linear classifier on top of the frozen CLIP image encoder. 2) CLIP-based CBMs: " + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "inline_equation", + "content": "CLIP-CBM" + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "content": " leverages human-written attributes to serve as a concept bank and utilizes CLIP to compute concept activation label. " + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "inline_equation", + "content": "PCBM" + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "content": " [42] uses ConceptNet [30] to retrieve concepts relevant to these classes and automatically construct the concept bank. Original PCBM employs different visual backbones on different datasets. For a fair comparison, we re-implement the PCBM with the CLIP backbone. " + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "inline_equation", + "content": "LaBO" + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "content": " [41] applies the LLM to generate abundant visual concepts in contrast to PCBM, and we employ a two-layer classifier. For each CBM baseline, we keep their original objectives to solve the optimization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 467, + 448, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 448, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 448, + 480 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 630 + ], + "type": "text", + "content": "We prompt GPT3.5-turbo to generate 200 domain descriptors. These domain descriptors are used across all datasets and " + }, + { + "bbox": [ + 313, + 486, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 486, + 555, + 630 + ], + "type": "text", + "content": " is set as 1. We pre-process the domain descriptors into text embeddings by CLIP text encoder. The DDO loss is added at the final classification layer of the CBMs in a plug-in manner. Following LaBO [41], we utilize the official CLIP ViT-L/141 by OpenAI as the default vision backbone to extract image embeddings. We train the linear function " + }, + { + "bbox": [ + 313, + 486, + 555, + 630 + ], + "type": "inline_equation", + "content": "W_{F}" + }, + { + "bbox": [ + 313, + 486, + 555, + 630 + ], + "type": "text", + "content": " using the Adam [10] optimizer. All experiments are conducted on a single RTX3090 GPU. More implementation details, hyperparameters and experiments with other vision backbones are listed in the Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 639, + 522, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 522, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 522, + 651 + ], + "type": "text", + "content": "5.3. Generalization to single unseen domain" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "type": "text", + "content": "We first evaluate our model on single unseen domain benchmarks which focused on photo " + }, + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "type": "text", + "content": " others generalization. Table. 1 shows in-distribution (ID) and out-of-distribution" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 476, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 476, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 476, + 713 + ], + "type": "text", + "content": "1https://github.com/openai/CLIP" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9522" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 221, + 195 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 221, + 195 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 221, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 221, + 195 + ], + "type": "image", + "image_path": "c88137de33cb773cd60f48116a070e87842cb969e4ab05dc82a3a6d82f722381.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 388, + 196 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 388, + 196 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 388, + 196 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 388, + 196 + ], + "type": "image", + "image_path": "5419fc55f43a079b733ad88b205c3e15027c6a44018ab3448d5fec6e57dd7fde.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 77, + 554, + 195 + ], + "blocks": [ + { + "bbox": [ + 391, + 77, + 554, + 195 + ], + "lines": [ + { + "bbox": [ + 391, + 77, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 391, + 77, + 554, + 195 + ], + "type": "image", + "image_path": "a5076c64d7d33fda148f009e1d3fc6d509aca76feff83c3e3593968e7e41d2dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 214, + 178, + 310 + ], + "blocks": [ + { + "bbox": [ + 74, + 201, + 534, + 213 + ], + "lines": [ + { + "bbox": [ + 74, + 201, + 534, + 213 + ], + "spans": [ + { + "bbox": [ + 74, + 201, + 534, + 213 + ], + "type": "text", + "content": "Figure 4. OOD performance on three multiple unseen domain generalization benchmarks, PACS, OfficeHome and DomainNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 214, + 178, + 310 + ], + "lines": [ + { + "bbox": [ + 56, + 214, + 178, + 310 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 178, + 310 + ], + "type": "image", + "image_path": "089571b2646149e42f7a64ffd1acd1f5c22fbfd6b0b105a38f7857226ad41955.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 214, + 302, + 310 + ], + "blocks": [ + { + "bbox": [ + 181, + 214, + 302, + 310 + ], + "lines": [ + { + "bbox": [ + 181, + 214, + 302, + 310 + ], + "spans": [ + { + "bbox": [ + 181, + 214, + 302, + 310 + ], + "type": "image", + "image_path": "bae3f9c2605426e23d0ba1a34ae6297c7c4e7998a760a556e58a94bdde2e24f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "text", + "content": "Figure 5. Ablation studies for effect of the number of domain descriptors. For each quantity, we randomly selected domain descriptors from a total of 200 domain descriptors and averaged the results over five random selections. Results of DomainNet are shown in the Appendix." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 214, + 428, + 310 + ], + "blocks": [ + { + "bbox": [ + 307, + 214, + 428, + 310 + ], + "lines": [ + { + "bbox": [ + 307, + 214, + 428, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 214, + 428, + 310 + ], + "type": "image", + "image_path": "e22162dda3b3b7eb3cc10541a849d38a5d5c01b2a08377105bf0a6ec5663a637.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 432, + 214, + 553, + 310 + ], + "blocks": [ + { + "bbox": [ + 432, + 214, + 553, + 310 + ], + "lines": [ + { + "bbox": [ + 432, + 214, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 432, + 214, + 553, + 310 + ], + "type": "image", + "image_path": "b57abaa88d90d1d60f95d6985cf4e9af426168a6979631e0f04870a1980eafe0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 352, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 297, + 568 + ], + "type": "text", + "content": "(OOD) accuracy on CUB-Painting, AwA2-clipart, LADA-Sculpture and LADV-3D. For all benchmarks, DDO loss is able to preserve or slightly improve the ID accuracy of each CBM while significantly improving their OOD accuracy (average 3 points improvements). We attribute the slight improvements in ID performance to the minor distribution differences between the training and test sets, even in the same visual domain. Moreover, LaBo with DDO loss can achieve competitive and superior performance on both ID and OOD performance with black-box CLIP-LP and narrow the gap with CLIP-ZS on OOD accuracy. These results demonstrate that our proposed LanCE design can effectively improve the generalization capabilities without sacrificing interpretability or classification accuracy. Besides, we notice that all models' generalization performance to sculpture and 3D model images on the LADA-Sculpture and LADV-3D benchmark is limited, indicating generalization from 2D to 3D remains challenging." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 575, + 279, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 279, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 279, + 588 + ], + "type": "text", + "content": "5.4. Generalization to multiple unseen domains" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "content": "To assess the ability to generalize to multiple unseen domains and the generalization capability when different domains act as the source domain, we conduct experiments on PACS, OfficeHome, and DomainNet." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Comparison with only using domain-shared concepts. In the absence of a pre-defined human-written concept set, we employ LLM to generate a concept set " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sp + in}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " for each dataset following LaBO [41]. This set contains both domain-specific concepts and domain-shared concepts. The main motivation of this paper is to mitigate the negative" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 322, + 350, + 547, + 397 + ], + "blocks": [ + { + "bbox": [ + 322, + 350, + 547, + 397 + ], + "lines": [ + { + "bbox": [ + 322, + 350, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 322, + 350, + 547, + 397 + ], + "type": "table", + "html": "
ModelMethodCUB-PAwA2-cLADA-SLADV-3D
LanCEbaseline56.2484.1574.5663.17
+DDO(IR)57.6085.5077.7065.46
+DDO59.6087.6680.0068.01
", + "image_path": "f92ef7d6b0db5aa35b4486103fd25eaa2a70672ec555a9252a2efa2b2168c705.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 398, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 431 + ], + "type": "text", + "content": "Table 2. Ablation studies for the effect of relevance of the domain descriptors. +DDO(IR) only use the domain-irrelevant descriptors while +DDO use all domain descriptors." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "impact of domain-specific concepts. A natural question that arises is why not simply query the LLM to generate only domain-shared concepts. To explore the differences between this domain-shared approach and our DDO loss, we introduce an additional baseline, where the LLM is queried exclusively to generate a domain-shared concept set " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sh}" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": ". Specifically, we adopt the single-domain generalization evaluation setting where one source domain is used for training, and the OOD performance is averaged across other domains. As shown in Fig. 4, only using " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sh}" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": " performs well only in certain cases, such as when the source domains in the PACS dataset are photo or sketch. We attribute this to the fact that the manifestation of domain-shared concepts varies across different domains. For instance, deer antlers in photo (realistic) and cartoon (simplified or exaggerated) domains differ significantly. Moreover, we empirically observe that prompt-based methods are highly sensitive to the wording of human-written prompts. Different prompts lead to significant performance variations, yet they share a common limitation: none consistently improve performance across diverse source domains. In contrast, DDO is more flexible. It adaptively erases the influence of domain-specific concepts while suppressing domain-specific infer" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9523" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 552, + 213 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 552, + 213 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 552, + 213 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 552, + 213 + ], + "type": "image", + "image_path": "bce1fdd097aeb73ee654f4e1196a1a009fb411fa195bacff3386a945a855d0df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 553, + 237 + ], + "type": "text", + "content": "Figure 6. Qualitative examples about the top-5 concepts, ranking by their weights in the final linear layer " + }, + { + "bbox": [ + 55, + 215, + 553, + 237 + ], + "type": "inline_equation", + "content": "W_{F}" + }, + { + "bbox": [ + 55, + 215, + 553, + 237 + ], + "type": "text", + "content": ". Baseline indicates the results of the original LaBo. Domain-specific concepts are highlighted in red. More results are listed in Appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 294, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 294, + 270 + ], + "type": "text", + "content": "mation in domain-shared concept activations, leading to stable improvements across a variety of source domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 276, + 156, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 156, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 156, + 288 + ], + "type": "text", + "content": "5.5. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 354 + ], + "type": "text", + "content": "To explore the impact of domain descriptors on final performance, we conduct comprehensive ablation studies, with LaBO as the baseline. Specifically, we compare results with different numbers of domain descriptors and investigate the role of distinct domain descriptors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 355, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 510 + ], + "type": "text", + "content": "The impact of the numbers of domain descriptors. We evaluate our method with regard to the number of domain descriptors " + }, + { + "bbox": [ + 55, + 355, + 295, + 510 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 355, + 295, + 510 + ], + "type": "text", + "content": " as shown in Fig. 5. The ID accuracy remains almost constant as the number of domain descriptors increases, while the OOD accuracy gradually improves with an increasing number of domain descriptors, though with diminishing returns. Specifically, when the number of domain descriptors is below 20, each additional descriptor noticeably enhances OOD accuracy. However, as the number of domain descriptors continues to increase beyond this point, the improvement in OOD accuracy slows progressively. The results show that 100 domain descriptors are enough to achieve decent performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 511, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 713 + ], + "type": "text", + "content": "The impact of the relevance of domain descriptors. To investigate which parts of 200 domain descriptors contribute to the most to OOD accuracy improvements, we manually split all domain descriptors into two components, domain-relevant descriptors and domain-irrelevant descriptors based on their relevance with the test unseen domains. For example, for the LADA-Sculpture dataset, domain-relevant domain descriptors are \"sculpture\", \"statue\", and \"furniture\", referring to the generated word cloud mentioned in Fig. 2 and domain-irrelevant parts are the remaining ones. Detailed splits for each dataset are listed in the Appendix. Results are shown in the Table. 2. DDO(IR) indicates that we only use the domain-irrelevant domain descriptors to compute the DDO loss. This evaluation shows that domain-irrelevant domain descriptors can also slightly prompt the OOD accuracy on the unseen domains, although relatively fewer than the domain-relevant counterparts. It" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 247, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 295 + ], + "type": "text", + "content": "indicates those domain descriptors are not fully orthonormal, also demonstrating that the DDO loss has the potential to generalize to pure unseen visual domains that are not mentioned in the domain descriptors set " + }, + { + "bbox": [ + 313, + 247, + 553, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 247, + 553, + 295 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 308, + 425, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 425, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 425, + 320 + ], + "type": "text", + "content": "5.6. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 328, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 553, + 436 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed LanCE method, we further visualize the Top-5 concepts highly related to randomly selected classes within different unseen domains as illustrated in Fig. 6. The qualitative results demonstrate that our proposed method can significantly decrease the association between the final prediction and the domain-specific concepts, in other words, erasing the impact of domain-specific concepts on the final output. More qualitative results are shown in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 452, + 387, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 387, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 387, + 464 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a language-guided concept-erasing (LanCE) framework that can effectively mitigate the biased association between domain-specific concepts and the final output. We first empirically demonstrate that the pre-trained CLIP can interpret the visual domain shifts into language. And these language-guided domain shifts can distinguish the domain-specific concepts via similarity within the CLIP embedding space. Based on these observations, we introduce a plug-in domain descriptor orthogonality (DDO) loss to erase the negative impact of domain-specific concepts, and that can significantly improve the generalization capabilities of prevailing concept bottleneck models (CBMs). We prompt the large language models (LLMs) to generate a bunch of domain descriptors to simulate numerous possible unseen domains. Moreover, considering that the current domain generalization community focuses more on generalizing to artistic unseen domains, we collect three new benchmarks that includes more difficult scenarios, i.e. 2D→3D, real animals→sculptures. Extensive experiments demonstrate the efficacy of our LanCE." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9524" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 251, + 297, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 297, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 297, + 296 + ], + "type": "text", + "content": "[1] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "type": "text", + "content": "[2] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 351, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 405 + ], + "type": "text", + "content": "[3] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. Advances in neural information processing systems, 36:79024-79034, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 406, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 459 + ], + "type": "text", + "content": "[4] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "type": "text", + "content": "[5] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 515, + 296, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 296, + 570 + ], + "type": "text", + "content": "[6] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 571, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 296, + 613 + ], + "type": "text", + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 679 + ], + "type": "text", + "content": "[8] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 714 + ], + "type": "text", + "content": "[9] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 118 + ], + "type": "text", + "content": "[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 119, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 554, + 163 + ], + "type": "text", + "content": "[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "type": "text", + "content": "[12] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 554, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 253 + ], + "type": "text", + "content": "[13] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "type": "text", + "content": "[14] Jichang Li, Guanbin Li, Yemin Shi, and Yizhou Yu. Cross-domain adaptive clustering for semi-supervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2505-2514, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "text", + "content": "[15] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "text", + "content": "[16] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 410, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 466 + ], + "type": "text", + "content": "[17] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 554, + 511 + ], + "type": "text", + "content": "[18] Jiawei Ma, Yulei Niu, Shiyuan Huang, Guangxing Han, and Shih-Fu Chang. Widin: Wording image for domain-invariant representation in single-source domain generalization. arXiv preprint arXiv:2405.18405, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 511, + 554, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 554, + 545 + ], + "type": "text", + "content": "[19] Lianbo Ma, Haidong Kang, Guo Yu, Qing Li, and Qiang He. Single-domain generalized predictor for neural architecture search system. IEEE Transactions on Computers, 2024. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 600 + ], + "type": "text", + "content": "[20] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In Proceedings of the IEEE international conference on computer vision, pages 5715-5725, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 635 + ], + "type": "text", + "content": "[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 657 + ], + "type": "text", + "content": "[22] OpenAI. Chatgpt. https://www.openai.com/chatgpt, 2023.5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9525" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1406-1415, 2019. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "type": "text", + "content": "[25] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International conference on machine learning, pages 7728-7738. PMLR, 2020. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 237 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 296, + 260 + ], + "type": "text", + "content": "[27] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "text", + "content": "[28] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "text", + "content": "[29] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 404 + ], + "type": "text", + "content": "[30] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "text", + "content": "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 492 + ], + "type": "text", + "content": "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 548 + ], + "type": "text", + "content": "[33] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 603 + ], + "type": "text", + "content": "[34] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 658 + ], + "type": "text", + "content": "[35] Sinan Wang, Xinyang Chen, Yunbo Wang, Mingsheng Long, and Jianmin Wang. Progressive adversarial networks for fine-grained domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9213-9222, 2020. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 710 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "text", + "content": "[38] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 554, + 228 + ], + "type": "text", + "content": "[39] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 554, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 273 + ], + "type": "text", + "content": "[40] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 554, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 340 + ], + "type": "text", + "content": "[41] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 341, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 554, + 373 + ], + "type": "text", + "content": "[42] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 373, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 554, + 408 + ], + "type": "text", + "content": "[43] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 408, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 554, + 464 + ], + "type": "text", + "content": "[44] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23465-23476, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 554, + 509 + ], + "type": "text", + "content": "[45] Zequn Zeng, Jianqiao Sun, Hao Zhang, Tiansheng Wen, Yudi Su, Yan Xie, Zhengjue Wang, and Bo Chen. Hiclescore: A hierarchical metric for image captioning evaluation. arXiv preprint arXiv:2407.18589, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 509, + 554, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 554, + 564 + ], + "type": "text", + "content": "[46] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero-shot image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14100-14110, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 564, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 564, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 564, + 554, + 620 + ], + "type": "text", + "content": "[47] Bo Zhao, Yanwei Fu, Rui Liang, Jiahong Wu, Yonggang Wang, and Yizhou Wang. A large-scale attribute dataset for zero-shot learning. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 620, + 554, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 620, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 620, + 554, + 666 + ], + "type": "text", + "content": "[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI conference on artificial intelligence, pages 21832-21840, 2024. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 666, + 554, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 666, + 554, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 666, + 554, + 710 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "9526" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_content_list.json b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b34eb36781a53ca2151b9a1478f3e657b25e8c74 --- /dev/null +++ b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_content_list.json @@ -0,0 +1,1343 @@ +[ + { + "type": "text", + "text": "Explaining in Diffusion: Explaining a Classifier with Diffusion Semantics", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tahira Kazimi† Virginia Tech", + "bbox": [ + 251, + 180, + 372, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tahirakazimi@vt.edu", + "bbox": [ + 227, + 220, + 397, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ritika Allada† Virginia Tech", + "bbox": [ + 465, + 180, + 578, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ritika88@vt.edu", + "bbox": [ + 455, + 220, + 589, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pinar Yanardag Virginia Tech", + "bbox": [ + 647, + 181, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": " pineary@vt.edu", + "bbox": [ + 648, + 220, + 766, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://explain-in-diffusion.github.io", + "bbox": [ + 326, + 244, + 663, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8af0c4f0105c182e2920e6af801a517e6576fe80fed90eb68f759d23e71f9468.jpg", + "image_caption": [ + "Figure 1. DiffEx explains the decisions of domain-specific classifiers by identifying the most influential semantics affecting their predictions. Classifier scores for each example are displayed in the top-left corner, demonstrating how classifier predictions change in response to the manipulation of different semantics (original images are shown with red borders). Our approach is capable of explaining classifiers that concentrate on individual concepts such as faces or animals (top row) as well as those that manage complex scenes involving multiple objects, such as a formal/casual fit in a fashion context (bottom row)." + ], + "image_footnote": [], + "bbox": [ + 93, + 282, + 289, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9775f157ca494465e96aadd14f7d0626789655281c8a0d1f16f98687cb4cbac7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 282, + 486, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7dbc047a4efdfabff0916122a8a1ba42b9fd46a217f1ad98b2c8003689fe873b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 282, + 694, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32ca24e47a20d40bf513c8582bf56588f3e9dbd49d33e16733add3bb12bdff51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 282, + 893, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 571, + 326, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Classifiers are important components in many computer vision tasks, serving as the foundational backbone of a wide variety of models employed across diverse applications. However, understanding the decision-making process of classifiers remains a significant challenge. We propose DiffEx, a novel method that leverages the capabilities of text-to-image diffusion models to explain classifier decisions. Unlike traditional GAN-based explainability models, which are limited to simple, single-concept analyses and typically require training a new model for each classifier, our approach can explain classifiers that focus on single concepts (such as faces or animals) as well as those that handle complex scenes involving multiple concepts. DiffEx employs vision-language models to create a hierarchical list of semantics, allowing users to identify not only the overarching semantic influences on classifiers (e.g., the 'beard' semantic in a facial classifier) but also their sub-types, such as 'goatee' or 'Balbo' beard. Our experiments demonstrate", + "bbox": [ + 88, + 604, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that DiffEx is able to cover a significantly broader spectrum of semantics compared to its GAN counterparts, providing a hierarchical tool that delivers a more detailed and fine-grained understanding of classifier decisions.", + "bbox": [ + 511, + 573, + 906, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 647, + 643, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Classifiers are fundamental to computer vision tasks, forming the backbone of many models used in a broad spectrum of applications [15, 21, 30, 32, 52]. Their ability to generalize across tasks and adapt to new domains with minimal retraining makes them highly transferable, and thus they are employed extensively in fields such as healthcare [37, 39, 42, 69], finance [38, 60, 62], security [2, 29, 33], and autonomous systems [5, 59, 65]. Despite their versatility and widespread utility, understanding the decision-making process of classifiers remains a significant challenge [3, 44, 50, 74, 76]. This challenge comes mainly from their \"black box\" nature. As images move through the deep layers of the network, the features used for classification become more abstract and harder to understand. The lack of interpretable features in such mod-", + "bbox": [ + 509, + 674, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contributions.", + "bbox": [ + 109, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14799", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els raises critical concerns, particularly in high-stakes environments such as medical diagnosis, where understanding the reasoning behind a model's prediction is crucial for ensuring trust, accountability, and informed decision-making [10, 34, 47, 57, 77]. Explaining classifier decisions is crucial for enhancing the transparency and reliability of these models. Prior research [31] has used generative adversarial networks (GANs) [18] to interpret classifier decisions by generating counterfactual examples that manipulate GAN latent semantics. These manipulations illustrate how changes in specific attributes, like the addition of the eyeglasses semantic, impact classifier outputs. However, GANs are often limited to single domains, such as facial images, and typically require training a new model per classifier, which is resource-intensive and time-consuming. Moreover, in GAN-based methods, understanding which latent semantics affect classifier decisions often requires manual intervention to identify and interpret relevant features, such as recognizing that a discovered semantic controls the eyeglasses attribute. This manual process is not only time-consuming but also less feasible in specialized fields like medicine, where identifying intricate attributes requires substantial domain expertise, making the approach impractical in critical scenarios.", + "bbox": [ + 89, + 90, + 480, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This limitation highlights the need for more automated and versatile approaches to interpret classifier decisions. Text-to-Image (T2I) diffusion models [45] emerge as a compelling alternative, widely recognized for their ability to generate high-quality images across various domains, which makes them a promising tool for explaining classifier decisions. These models offer the potential for a richer and more diverse set of semantic features compared to existing methods. However, their ability to interpret and utilize latent space semantics remains limited in the context of diffusion models. Existing techniques for identifying meaningful semantics rely largely on supervised approaches [7, 8], which require users to craft detailed text prompts to specify particular features for editing, such as a mustache. This process requires domain expertise to craft appropriate prompts; thus, a broad semantic corpus is essential to improve diffusion models for classifier explanations while reducing manual effort. In this paper, we first employ Vision-Language Models (VLMs) [36] to extract a largescale corpus covering domain-specific hierarchical semantics (see Fig. 2). Then, we introduce a training-free method, DiffEx, which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics. Our method provides explanations for both coarse and fine-grained semantics. For example, it can recognize a 'beard' as a coarse semantic influencing age classification scores and also demonstrate how specific beard types (such as 'Balbo' or 'Anchor' beards) impact the classifier's scores.", + "bbox": [ + 91, + 463, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae1b7fea0847d7d6c1e91ff692825fc889f7d64066085020bd93e34f40ee6ea5.jpg", + "image_caption": [ + "Figure 2. Hierarchical List of Attributes for the Bird Domain. We use VLMs to extract a hierarchical corpus of semantics within a given domain. This representation helps illustrate how different attributes are grouped within a broader domain, facilitating a better understanding of how each semantic contributes to the overall decision-making process of a classifier." + ], + "image_footnote": [], + "bbox": [ + 545, + 90, + 874, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This hierarchical strategy provides users with an overview of the most significant semantics for a classifier, allowing them to dig deeper into particular fine-grained semantics that are essential for understanding classifier behavior. Our qualitative and quantitative experiments reveal that DiffEx offers considerably richer and more comprehensive explanations for binary and multi-class classifiers across various domains, including facial features, retinal health, and plant pathology. Our contributions are as follows:", + "bbox": [ + 511, + 323, + 906, + 459 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose DiffEx, a training-free approach using VLMs and T2I diffusion models to explain classifier decisions. To the best of our knowledge, this is the first hierarchical approach that explains classifier decisions.", + "- Our method employs a VLM to develop a comprehensive semantic corpus that spans multiple domains, with our source code publicly available to support future research.", + "- Unlike GAN-based methods, our approach can address classifiers that focus on single concepts (such as an 'age' classifier analyzing a headshot of a person) and also extend to classifiers for complex scenes (such as a 'modern/traditional architecture' classifier).", + "- Our method offers more comprehensible explanations for classifiers in applications ranging from facial recognition to retinal health compared to prior approaches and is adaptable to both binary and multiclassifiers." + ], + "bbox": [ + 513, + 464, + 906, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 723, + 653, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional research focuses mainly on heat maps and patch-based extractions to explain classifier decisions. Specifically, class activation and saliency maps attempt to emulate human visual strategies by focusing on the image regions most relevant to a specific class [43, 49, 53, 56, 72, 75, 78]. While these maps highlight the object or part of an image that most influences the classifier's decision, they fail to reveal which specific, fine-grained object attributes (e.g., color, pattern, or texture) impact the classification. Other approaches try to explain classifier outputs by an-", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14800", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ab4b3263592897b9a8e04f126f817f3c82542289ab72056892a53664655ceda3.jpg", + "image_caption": [ + "Figure 3. An Overview of DiffEx. Our pipeline processes a set of sample domain-specific images and a text prompt using a VLM to generate a hierarchical semantic corpus of attributes relevant to a specific domain. Based on this corpus, DiffEx identifies and ranks the most influential features affecting the classifier's decisions, sorting them from most to least impactful (rightmost image). The hierarchical explanation of semantics (like a \"beard\" and its subtypes) provides a fine-grained understanding of which features drive classifier outputs." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "alyzing extracted image patches [16, 73]; however, these methods can only reveal spatially localized attributes. Previous studies have leveraged generative models such as variational autoencoders (VAEs) [19] or GANs [17, 48, 54, 55]. The most similar research to ours, StylEx [31], introduces a GAN-based approach to identify various attributes that influence a classifier's decisions. However, this method has several limitations. Firstly, it requires training a new GAN for each classifier, which can be resource-intensive and time-consuming. Secondly, each identified attribute requires manual labeling, which can require domain expertise. Lastly, StylEx only uncovers a limited range of semantic attributes, potentially overlooking others that might significantly affect a classifier's scores. Additionally, a notable limitation of the StylEx method and similar GAN models is their focus on single concepts, such as human or animal faces, or individual objects like leaves, rather than entire scenes. In contrast, diffusion models, known for their robust capability to generate complex and detailed entire scenes, offer a significant advantage. Our approach leverages the power of diffusion models to cover a wider spectrum of visual contexts, and enhances the versatility and applicability of our method across various classifiers that assess not only individual elements but also the interaction and composition of entire scenes.", + "bbox": [ + 91, + 311, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent approaches have begun using diffusion models to generate counterfactual examples. One method utilizes shortcut learning to generate counterfactual images but fails to make semantically meaningful edits for certain attributes [68]. Another study explores modifying the diffusion process via adaptive parametrization and cone regularization to produce realistic counterfactual images; however, this approach depends on a robust model, which can be difficult to train [4]. [26] explored counterfactual image generation, however their approach is computationally demanding and uses DDPM [22] models trained on single domains. As a result, it does not take advantage of large-scale latent diffusion models like Stable Diffusion [46], which can handle more complex scenes. Furthermore, even though some", + "bbox": [ + 89, + 688, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "studies have leveraged diffusion models to generate realistic counterfactuals in high-stakes domains [24], such as the medical field, they use a less efficient base model for the image generation process and focus predominantly on single-attribute modifications. Furthermore, to the best of our knowledge, there is no existing research that explores a hierarchical explanation of classifiers. Such an approach would systematically unpack the layers of influence that different semantic levels have on a classifier's scoring mechanism. This gap highlights a significant opportunity to enhance understanding by detailing how various semantic categories and their sub-types contribute to the decisions made by classifiers.", + "bbox": [ + 511, + 311, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 513, + 648, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first discuss the background on identifying attributes that influence classifier decisions through changes in logits. Next, we introduce our method, which includes curating hierarchical attributes using VLMs and a novel algorithm inspired by beam search to pinpoint attributes that affect classifier scores. Our pipeline is detailed in Fig. 3.", + "bbox": [ + 511, + 537, + 906, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 513, + 637, + 643, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StylEx [31] identifies semantics that meaningfully influence classifier decisions by ranking each attribute based on its impact on the classifier's logit outputs. This ranking process aims to identify and select attributes that best explain the classifier's behavior in a given context, such as understanding the factors influencing age classification.", + "bbox": [ + 511, + 657, + 906, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a semantic corpus $S$ and a set of $N$ images, counterfactual images are generated to analyze the influence of semantic attributes on classifier decisions. For each original image $x_{i}$ , an edited version $x^{\\prime}_{i} = g(x_{i}, s)$ is generated by applying a semantic attribute $s \\in S$ through a transformation function $g$ . This transformation highlights the effect of each semantic attribute on a classifier's output. Moreover, the logit difference for each attribute measures how the classifier's score changes due to the presence of $s$ .", + "bbox": [ + 511, + 750, + 908, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The influence of each attribute $s$ is quantified as the aver", + "bbox": [ + 511, + 886, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1289057b7271758f9b7ce2a009e1f676c7f6d04d49b9c4c356386448aaef76c3.jpg", + "image_caption": [ + "Figure 4. Top-7 Facial Attributes for Age Classification. DiffEx identifies key attributes and their top subtypes for the age classifier. Edited images are arranged hierarchically, with scores for the \"young\" label shown at the top-left of each image." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 475, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "age logit difference between the original and edited images across a set of sample images, defined as:", + "bbox": [ + 89, + 316, + 483, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI (s) = \\left| f \\left(X ^ {\\prime}, s, y\\right) - \\frac {1}{N} \\sum_ {i = 1} ^ {N} C \\left(x _ {i}, y\\right) \\right| \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 359, + 482, + 400 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (X, s, y) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} C \\left(g \\left(x _ {i}, s\\right), y\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 414, + 482, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f(X,s,y)$ denotes the average classifier's logit score for target class $y$ on image set $X$ edited with attribute $s$ . Here, $y$ represents the unedited images' target labels, such as the \"young\" label for an age classifier. This influence score $I(s)$ captures the average impact of each semantic attribute on classifier decisions by reflecting the logit score change due to attribute manipulation.", + "bbox": [ + 89, + 462, + 483, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Our Method", + "text_level": 1, + "bbox": [ + 89, + 578, + 225, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first employ Vision-Language Models (VLMs) [36] to extract a large-scale corpus covering domain-specific hierarchical semantics. Then, we introduce a training-free method, $DiffEx$ , which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics.", + "bbox": [ + 89, + 599, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. VLM-Based Semantic Space", + "text_level": 1, + "bbox": [ + 89, + 715, + 336, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the StylEx method outlined in Section 3.1 employs a logit-based approach to identify semantics, it requires manually labeling each attribute extracted from the trained GAN model. Moreover, this method does not support the explanation of hierarchical attributes. Therefore, we first compile a large-scale set of semantics using VLMs. Given a domain $d$ , such as the 'facial domain,' our objective is to identify a comprehensive set of domain-specific attributes from a collection of domain-specific images, denoted as $N_{d}$ . To accomplish this, we utilize a Vision-Language Model (VLM) [36] to extract a range of relevant features, represented by", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 DiffEx", + "text_level": 1, + "bbox": [ + 514, + 90, + 651, + 104 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Require: Hierarchical structure $\\mathcal{H}$ with semantic groups and features, beam width $B$ , classifier or scoring function $f$ , scoring threshold $\\delta$", + "Ensure: Optimal semantics maximizing $I$", + "1: Initialize $S \\gets$ root-level groups in $\\mathcal{H}$ {Initial candidate set at top-level groups}", + "2: Initialize beam $\\mathcal{B}\\gets \\emptyset$", + "3: Score each candidate $s \\in S$ using the scoring function $f(X, s, y)$ for $y$ class label", + "4: Select top $B$ candidates with $I(b) \\geq \\delta$ and store in beam $\\mathcal{B}$ and $S$ {Apply thresholding to filter relevant candidates}", + "5: while $\\mathcal{B} \\neq \\emptyset$ do", + "6: Initialize $S_{\\mathrm{next}} \\gets \\emptyset$", + "7: for each candidate $b \\in \\mathcal{B}$ do", + "8: Expand $b$ by adding sub-features from its next level in $\\mathcal{H}$ to form new candidates", + "9: for each new combination $b'$ expanded from $b$ to 10: if $I(b') > I(b)$ then", + "11: Add $b'$ to $S_{\\mathrm{next}}$ .", + "12: end if", + "13: end for", + "14: end for", + "15: Set $\\mathcal{B}\\gets S_{\\mathrm{next}}$", + "16: Append $S_{\\mathrm{next}}$ to $S$", + "17: end while", + "18: Return highest-scoring combination from final $S$ as the optimal joint semantic combination" + ], + "bbox": [ + 514, + 109, + 906, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{H}$ . Employing in-context learning [9, 58], we prompt the VLM with a small set of images $N_{d}$ , along with a detailed task description and examples of desired outputs. This process allows us to generate a substantial semantic corpus of keywords, $\\mathcal{H}$ , that captures the fine-grained attributes relevant to the domain. The resulting corpus, $\\mathcal{H}$ , comprises a comprehensive collection of keywords that covers full spectrum of the domain's attributes. We use this dataset to hierarchically explain how different attributes influence classifier decisions.", + "bbox": [ + 511, + 566, + 906, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2.DiffEx", + "text_level": 1, + "bbox": [ + 513, + 729, + 606, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the extensive number of semantics identified using a VLM, it is computationally expensive to evaluate every possible semantic or combination of attributes. We introduce DiffEx, an efficient approach inspired by beam search to explain classifier decisions (see Algorithm 1). This method leverages our hierarchical semantic corpus to streamline the process. Our approach iteratively refines candidate attributes by expanding only the most impactful semantic paths at each hierarchical level, with each path's relevance guided by a scoring function that assesses the clas", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14802", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5ce3a010250f4351b7efccb2c7f1d7f7fb8bdc399c9d602f28c6880c42a67a04.jpg", + "image_caption": [ + "Figure 5. Top-7 Discovered Attributes Across Different Animal Domains. Our method successfully identifies key attributes for multiple domains, such as bird, wildcat, and pet species. The original images are depicted with red borders while the edited images are depicted with black borders. For the pet species domain, we used a binary classifier and for the bird and wildcat species domains, we used a multi-classifier." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 460, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sifier's response to each semantic feature. Formally, the scoring function $f$ calculates the average classifier $C$ logit scores across $N$ images generated using each semantic attribute $s$ , as defined in Eq. (2), where $g$ represents the generative diffusion model applied to each sample image.", + "bbox": [ + 89, + 462, + 482, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin with an initial candidate set $S$ , which includes high-level groups from the semantic hierarchy (e.g., broader categories like 'mouth features' or 'eye features'). Each candidate in $S$ is evaluated by the scoring function $f(X, s, y)$ , which quantifies the influence of the candidate on the classifier's output. Only the top $B$ candidates that surpass a predefined score threshold $\\delta$ are retained, setting a beam width that limits the search to the most impactful candidates. For each candidate in the beam, the algorithm proceeds by expanding to the next hierarchical level, incorporating more specific sub-features (e.g., for \"mouth features,\" sub-features such as \"beard\" and \"mustache\" are included). For each expanded candidate $b'$ , the scoring function in Eq. (2) is re-evaluated. Only candidates with a score exceeding that of their parent $I(b') > I(b)$ are retained in the next candidate set $S_{next}$ , ensuring that only those additions that yield significant incremental impact are added. This process of expansion, scoring, and filtering continues iteratively, moving from general to more specific attributes at each hierarchical level. By dynamically adjusting the candidate set based on beam width $B$ and scoring threshold $\\delta$ , our method remains computationally efficient, focusing on high-impact combinations rather than exhaustively evaluating all possible attribute pairings. For generating counter", + "bbox": [ + 89, + 537, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d94ce4c3d27b7b8a02852f2c2442bc68196beb52106c4d6ec7b4cd825c16615.jpg", + "image_caption": [ + "Figure 6. Visual Comparison of Key Attributes Identified by StylEx and DiffEx in the Plant Health and Retinal Disease Domains. This figure illustrates the enhanced capability of our method in identifying a broader set of significant attributes compared to StylEx within the plant health and retinal disease domains. DiffEx uncovers more detailed and diagnostically relevant features, such as \"leaf vein color\" and \"macular hole,\" which provide deeper insights into leaf and retina health." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "factual images, we utilize an off-the-shelf diffusion-based editing tool, Ledits++ [8]. However, our approach is versatile enough to accommodate any editing method.", + "bbox": [ + 511, + 349, + 906, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 407, + 645, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setup. Our experiments utilize Ledits++ [8] and Stable Diffusion XL (SDXL) [40] for generating counterfactual images. Twenty-five time steps are omitted to boost computational efficiency while preserving the quality of edits. The edit threshold, a hyperparameter that dictates the global application scope of edits, is adjusted based on each domain. We test our method across diverse classifiers to evaluate its effectiveness in explaining model behavior through semantic influence. Specifically, we utilize classifiers trained on facial attributes data (i.e. age and gender classifiers) [28] as well as plant health [23], retinal disease [13], bird species [64], wildcat/pet species [11], fashion [35], places [79], and food data [6]. The experiments for the face and plant health domains utilize a CLIP-based classifier [41] to evaluate and interpret edits, while experiments within the bird, wildcat, pet, food, fashion, and places domains use CNN-based classifiers. These CNN classifiers were built on the EfficientNet [61] architecture and achieved an accuracy of over 95 percent on their test sets. For the retinal disease domain, we utilized the FLAIR model [51], which is based on a pre-trained vision-language model, to classify various retina scans.", + "bbox": [ + 511, + 431, + 908, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Qualitative Experiments", + "text_level": 1, + "bbox": [ + 511, + 773, + 738, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Explaining Classifiers with Single Attributes. We used various classifiers to analyze fine-grained semantic features for age categorization, species identification, leaf health assessment, retinal disease detection, etc. In the figures illustrating the visual results, the images with red borders indicate the original, unedited versions, while those with black borders are the edited versions, unless noted otherwise.", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14803", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fb91ef1216e1942487ac46fe08f6bdc48e0411e2b72bd6df62edd46ca4bdf81a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Face (Age)Bird (Species)Leaves (Health)Retina Scans (Disease)Wildcat (Species)Pet (Cat/Dog)
StylExOursStylExOursStylExOursStylExOursStylExOursStylExOurs
Skin PigmentationEyebrowBelly ColorUpperparts ColorBase Leaf ColorBase ColorExudatesGlaucomaSpotsStripesOpen MouthNose Color
Eyebrow ThicknessMakeupUpperparts ColorHead ColorApex ColorVein ColorCotton Wool SpotsSubretinal HemorrhageBlack Tear MarkOpen EyesClosed MouthWhiskers
EyeglassesMustache TypeWing PatternBeak ShapeSpotsApex ColorHemorrhagesIntraretinal HemorrhageEye Shape + SizeNose ColorEye ShapeEye Color
Hair ColorTeethBeak ColorBeak ColorBlightSpots PresenceClustered ExudatesMacular HoleXSpotsDropped EarsClosed Mouth
Lip Thickness + PositionLip VolumeHead ColorWing PatternHalosDisease PresenceXHard ExudatesXFur ColorPointed EarsOpen Mouth
BangsLip ColorBreast ColorEye ColorXBlight SizeXBlackened MaculaXManeEye CircumferenceSpots
Eye MakeupEyelashXThroat ColorXLeaf TextureXSoft ExudatesXEye ColorXCollar
Facial Hair ColorBeard TypeXWing ColorXSpots ColorXRetinal DruenXTongueXPointed Ears
XFacewearXCrest PresenceXDiscolorationXOptic Disc HemorrhageXPupil SizeXMouth Color
XHeadwearXFeather TextureXLeaf OrientationXCataractXWhiskersXFur Pattern
", + "bbox": [ + 93, + 88, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2acaa466a2c465eae8e736be7a37d602048c6630a2080de741790207d06fafff.jpg", + "table_caption": [ + "Table 1. Comparison of Top Attributes Across Different Domains and Classifiers. The table above contains a list of the top attributes discovered by DiffEx (Ours) vs. StylEx. The $\\pmb{x}$ in the table indicates attributes that were not mentioned in StylEx. It is also important to note that \"cotton wool spots\" and \"soft exudates\" refer to the same condition within the retinal disease domain." + ], + "table_footnote": [], + "table_body": "
MethodCrest PresenceBeak ShapeThroat ColorFeather TextureEye ColorBeak ColorHead ColorUpperparts ColorAvg. Correct Response
Grad-CAM36%50%56%35%47%65%59%76%53%
StylEx68%85%79%82%74%68%91%65%76.5%
DiffEx (Ours)88%91%88%91%82%82%97%88%88.4%
", + "bbox": [ + 93, + 232, + 903, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with Other Explainability Methods. The table above displays the percentage of correct attribute selections for the bird class, as chosen by users when viewing outputs from different explainability methods. It also includes the average percentage of correct responses across all attributes for each method. As shown, for each attribute presented, the majority of users identified the correct attribute when viewing the output generated by DiffEx.", + "bbox": [ + 89, + 289, + 906, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face Domain: In the facial domain, as illustrated in Fig. 4, DiffEx identifies and ranks key attributes impacting age classification. For example, features such as \"makeup styles,\" \"lip volume,\" and \"accessories\" (e.g. \"hairbands\") are associated with perceived youthfulness, whereas attributes like \"facial hair\" and \"eyeglasses\" are linked with perceived older age. Notably, the eyeglasses attribute consistently reduces the classifier's score for the \"young\" label, reflecting its association with older demographics. Additionally, DiffEx uncovers hierarchical attribute structures, demonstrating how subcategories within a feature can have varying effects on classifier outcomes. For instance, as shown in Fig. 4, different beard styles impact the perceived age differently (e.g. a \"Balbo beard\" significantly increases the age classification score more than a \"full beard\"). Additional examples of hierarchical explanations of age classifiers are detailed in the appendix (S10).", + "bbox": [ + 88, + 349, + 480, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Animal Domain: Furthermore, DiffEx explains classifiers for a variety of animal types. Fig. 5 highlights the top-7 most influential attributes in the bird, wildcat, and pet domains where our method was able to identify fine-grained semantics to explain classifier behavior. For example, in the wildcat classifier, attributes like stripes and manes are critical in distinguishing wildcats from other species. Explaining these classifiers also reveals potential biases. For instance, the presence of a collar significantly increases the likelihood of an animal being classified as a dog. This bias may stem from the training dataset where images of dogs more frequently featured collars compared to cats.", + "bbox": [ + 88, + 609, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Medical and Plant Health Domains: Fig. 6 demonstrates the results for retinal disease and plant health domains. For the retinal disease domain, we use the FLAIR model to classify images of retinal fundus scans. This model utilizes detailed domain expert knowledge descriptions, such as \"no hemorrhages, microaneurysms, or exudates\" compared to general descriptions like \"no diabetic retinopathy\" to aid in", + "bbox": [ + 88, + 795, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "its classification. For the plant health domain, the CLIP classifier we use looks at features such as the presence of spots, fungus, etc. to make its decision.", + "bbox": [ + 511, + 349, + 906, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Object Domains involving Complex Scenes: Unlike traditional GAN-based methods that primarily focus on single-object scenarios like a cropped face, DiffEx extends its utility by providing a list of relevant features for domains encompassing multiple objects, such as places, food, and fashion. Classifiers within these domains often evaluate multiple elements simultaneously. For example, a fashion classifier might determine the 'formal or casual fit' of an outfit by considering various components such as the hairstyle, top, and bottom. Similarly, an architectural classifier assessing whether a building appears 'urban or rural' may base its evaluation on not just the structure itself but also its exteriors and surroundings. Explaining the decision-making process of such classifiers is crucial, as they analyze multiple objects within a single image, adding complexity to their interpretative frameworks. Fig. 7 illustrates the top-7 attributes identified by DiffEx for the food, place, and fashion domains. Moreover, we demonstrate DiffEx on various diverse domains. For example, for the food domain, DiffEx was able to discover that removing \"caviar\" or a \"plate decoration\" from the image made the food appear to be more perceived as fast food compared to fine dining. For the places domain, DiffEx was able to find that adding a \"tractor\" to an image or a \"dirt road\" made the place seem more \"rural\" than \"urban.\" In the fashion domain, the style of a neckline significantly influences whether an outfit was classified as \"formal\" or \"casual.\" For instance, a V-neckline was often associated with a more casual look, whereas a classic boat neckline was generally perceived as more formal.", + "bbox": [ + 511, + 398, + 906, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Explaining Classifiers with Joint Attributes. Our method goes beyond analyzing individual attributes by identifying attribute combinations that collectively improve classifier", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14804", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6955f2fc9dd51787fc2889e50d11b978631d8cab0a2f1bf8c528a8ec670b4eeb.jpg", + "image_caption": [ + "Figure 7. Top-7 Attributes Discovered by DiffEx for Multi-Object Domains involving Complex Scenes. Compared to existing methods, DiffEx was used to identify meaningful attributes for multi-object domains, such as \"food,\" \"places,\" and \"fashion.\" By identifying semantic features that make food appear \"healthier\" or part of \"fine dining,\" as well as attributes that give a place a more \"rural\" or \"modern\" feel, DiffEx can serve as a powerful tool for understanding perceptions through targeted edits." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 472, + 510 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c29e58c95029574ca35e093448f6dfbac4b39a5f3ac6b98fa3451fea08385b8f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RatingBird DomainFace Domain
Edit Quality3.386 ± 0.2233.659 ± 0.248
Disentanglement3.163 ± 0.1973.204 ± 0.213
", + "bbox": [ + 99, + 633, + 475, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Edit Quality and Disentanglement Ratings. These scores across various domains from User Study 1 are shown above.", + "bbox": [ + 89, + 694, + 482, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "interpretation. While single attributes may have a minimal impact on logits, their combined effect can substantially influence the classifier's output, uncovering subtle interactions that shape decision-making. For example, as shown in Fig. 8, Section C, individual changes in \"lip color\" or \"eye makeup\" result in minimal score changes, but when modified together, they make the subject appear significantly younger. This analysis shows how classifiers may respond more robustly to specific attribute combinations. Additional examples are provided in the appendix (S7, S8, S9).", + "bbox": [ + 89, + 734, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extending DiffEx for Multi-Classifier Analysis. We", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adapt DiffEx for multi-classifier applications to uncover semantic attributes essential for tasks like retinal disease classification (Fig. 6) and bird and wildcat species identification (Fig. 5). This approach highlights key features such as \"upperparts color\" and \"beak shape\" for bird species, \"stripes\" and \"nose color\" for wildcat species, and types of hemorrhages and exudates for retinal conditions. The results in Fig. 6 for the retinal diseases domain and Fig. 5 for the bird/wildcat species domain demonstrate that modifying these attributes significantly alters classifier scores, impacting assessments for species identification and disease likelihood.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons. In Fig. 6, we present a visual comparison of the top attributes and classification scores identified by DiffEx against those identified by StylEx for the plant health and retinal disease domains. As illustrated, DiffEx successfully uncovers a broader set of semantically meaningful attributes compared to StylEx. For instance, DiffEx identifies detailed attributes such as \"leaf vein color\" and \"spots color,\" in addition to the more general attributes found by StylEx, like \"leaf base color\" and \"apex color.\" Table 1 provides a comprehensive comparison of the top features identified by StylEx and our method, highlighting DiffEx's superior ability to uncover semantics that are crucial for the classifier's decision-making process. In particular, the table presents an extended list of features across domains such as faces, bird species, plant health, retina scans, wildcat species, and pet types, all of which influence a classifier's score—further emphasizing DiffEx's ability to uncover fine-grained and contextually rich features. Additionally, in Figure 8 in Sections A and B, we visually compare DiffEx with other diffusion-based, counterfactual generation methods, such as FastDiME [68], DiME [26], and ACE [27] in the face and medical domains. As seen in Figure 8, Section B, DiffEx is able to preserve important features, such as a person's eyes, glasses, etc., in various cases, such as removing a smile or making an older person look younger. Furthermore, Figure 8, Section A shows DiffEx's ability to generate counterfactuals in various medical subdomains, such as skin lesions and chest x-rays. Specifically, for the skin lesion domain, DiffEx was able to remove ruler markings in images from the ISIC dataset [12, 63], while in the chest x-rays domain, DiffEx was able to remove chest drains and pacemakers from images from the NIH [67] dataset and CheXpert [25] dataset. As seen in Figure 8, Section A, DiffEx achieved similar results to FastDiME (Med) [68], a diffusion-based counterfactual model that was trained on medical images.", + "bbox": [ + 511, + 272, + 906, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Experiments", + "text_level": 1, + "bbox": [ + 511, + 830, + 750, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. To quantitatively evaluate the effectiveness of DiffEx compared to other explainability methods, such as StylEx [31] and Grad-CAM [49], we conducted a series of", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9f65854f319e7a13091cdd841d25ad0355c1a9365ba433faa7321199ec01cf2.jpg", + "image_caption": [ + "Figure 8. More Comparisons and Exploring Joint Attributes. (a) We demonstrate DiffEx's ability to produce counterfactuals comparable to FastDiME-2+ in skin lesion and chest X-ray domains. (b) compares DiffEx with ACE, DiME, FastDiME, and FastDiME-2+, showing its capacity to add or remove attributes in a disentangled manner. (c) highlights DiffEx's effectiveness in explaining classifier decisions by analyzing the influence of joint attributes, with perceived age scores (e.g., \"younger\" label) displayed to show how attribute pairings affect classifier outputs." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comprehensive user studies to assess how easily participants could identify the relevant features extracted by our method. User Study 1: Visual Quality and Disentanglement. To evaluate the visual quality and disentanglement of edited images for the face and bird domains, we conducted a user study with 50 participants on Prolific.com. For each domain, we showed pairs of unedited and edited images for ten attributes and asked users to assess whether the edited image contained the desired attribute and was disentangled. For each pair of images, we asked users to rate the edit and disentanglement from one to five, with five representing the highest score. Our results (see Table 3) indicate that our edits reflected the intended attributes while minimizing unrelated changes (see S6.1 in the appendix for more details).", + "bbox": [ + 91, + 311, + 483, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study 2: Comparison with Grad-CAM and StylEx. To evaluate how effectively our method, DiffEx, explains various semantics within a specific domain compared to other explainability methods (Grad-CAM and StylEx), we conducted a user study with 35 participants on Prolific. We focused on images from the bird domain, generating three sets of three images per attribute: one original image, one edited image, and one image illustrating the explainability method. For Grad-CAM, participants were shown a heatmap overlay on the edited image. For StylEx, we displayed the edited image alone, as this method requires users to manually label the edits. For DiffEx, participants viewed the edited image and the attribute automatically assigned by the VLM. We then asked users to select the attribute (e.g., \"beak color,\" \"crest presence\") that best explained the edited image from four choices. Results indicate that DiffEx significantly outperforms Grad-CAM and StylEx in explaining edited attributes in images (see Table 2). Please refer to S6.2 in the appendix for details.", + "bbox": [ + 91, + 522, + 483, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study 3: Disentanglement in Images with Multiple Attributes. We evaluated DiffEx's disentanglement ability when editing multiple attributes simultaneously via a user study. Users rated images on a scale of 1 to 5 (5 = very disentangled, 1 = entangled). DiffEx achieved an average score", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of 4.6 for single-attribute edits and 4.3 for edits involving four or more attributes, demonstrating effective handling of multiple edits (see S6.3 in the appendix for more details).", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies. To assess the robustness of our method, we tested various Vision-Language Models (VLMs) and Diffusion Models. For VLMs, we probed Llama 3.2 [20], Qwen2-VL [66], and Claude 3.5 Haiku [1] within the face domain, measuring semantic consistency via IoU scores against GPT-4. The results were 0.94, 0.87, and 0.93, respectively, which indicate substantial semantic overlap. For diffusion models, we compared top attributes from Stable Diffusion (SD) 1.5 and SD 2.0 [46] against SDXL [40] using Normalized Discounted Cumulative Gain (nDCG) to measure the attributes' overlap, achieving scores of 0.75 and 0.76, respectively. These results confirm the stability of our method across different models.", + "bbox": [ + 511, + 136, + 908, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 345, + 627, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While our approach provides valuable insights into classifier behavior through semantic edits, it has limitations. Relying on VLM-curated semantics restricts us to the quality and scope of the initial corpus, which may miss relevant features. Additionally, using off-the-shelf editing models like Ledits++ can result in entangled edits (a common issue in image editing algorithms [71]) or fail to capture fine-grained details like earlobes or nose rings (see Appendix S1). Nonetheless, our framework is flexible and can be enhanced with domain-specific adjustments, such as task-specific RAGs [70] or editing methods [7, 14]. RAGs can automatically retrieve semantics in specialized domains (e.g. medical) to improve generalization.", + "bbox": [ + 511, + 371, + 908, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this work, we introduce DiffEx, a novel approach for explaining classifier decisions by utilizing semantic edits within diffusion models. By harnessing the power of vision language models, we curate a comprehensive, hierarchical semantic corpus across various domains and propose a novel algorithm inspired by beam search to filter and rank the most impactful features. DiffEx ranks these semantic features based on their influence on classifier logits, capturing both individual and joint attribute effects, which are crucial for understanding complex classifier behaviors. Through experiments conducted on a wide range of domains—including face, bird, and medical classifiers—we showcase the robustness and adaptability of our approach. Our work provides a powerful tool for interpreting model decisions across diverse applications and promoting transparency in AI-driven classification systems.", + "bbox": [ + 511, + 583, + 908, + 825 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14806", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] The claude 3 model family: Opus, sonnet, haiku. 8", + "[2] Mostofa Ahsan, Rahul Gomes, Md. Minhaz Chowdhury, and Kendall E. Nygard. Enhancing machine learning prediction in cybersecurity using dynamic feature selector. Journal of Cybersecurity and Privacy, 1(1):199-218, 2021. 1", + "[3] Leila Amgoud. Explaining black-box classifiers: Properties and functions. International Journal of Approximate Reasoning, 155:40-65, 2023. 1", + "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. Advances in Neural Information Processing Systems, 35:364-377, 2022. 3", + "[5] Ouahiba Azouaoui and Amine Chohra. Soft computing based pattern classifiers for the obstacle avoidance behavior of intelligent autonomous vehicles (iav). Applied intelligence, 16:249-272, 2002. 1", + "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5", + "[7] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. SEGA: Instructing text-to-image models using semantic guidance. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 8", + "[8] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2, 5", + "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4", + "[10] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2", + "[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. CoRR, abs/1912.01865, 2019. 5", + "[12] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 7", + "[13] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. arXiv preprint arXiv:2312.05390, 2023. 8", + "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[16] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 3", + "[17] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5744-5753, 2019. 3", + "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[19] Yash Goyal, Amir Feder, Uri Shalit, and Been Kim. Explaining classifiers with causal concept effect (cace). arXiv preprint arXiv:1907.07165, 2019. 3", + "[20] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. (arXiv:1512.03385), 2015. arXiv:1512.03385. 1", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[23] David P. Hughes and Marcel Salathe. An open access repository of images on plant health to enable the development of mobile disease diagnostics. (arXiv:1511.08060), 2016. arXiv:1511.08060 [cs]. 5", + "[24] Indu Ilanchezian, Valentyn Boreiko, Laura Kuhlewein, Zwei Huang, Murat Seckin Ayhan, Matthias Hein, Lisa Koch, and Philipp Berens. Generating realistic counterfactuals for retinal fundus and oct images using diffusion models. arXiv preprint arXiv:2311.11629, 2023. 3", + "[25] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7", + "[26] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. In Proceedings of the Asian Conference on Computer Vision, pages 858-876, 2022. 3, 7", + "[27] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Adversarial counterfactual visual explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16425-16435, 2023. 7" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14807", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 4396-4405, Long Beach, CA, USA, 2019. IEEE. 5", + "[29] Ziv Katzir and Yuval Elovici. Quantifying the resilience of machine learning classifiers used for cyber security. Expert Systems with Applications, 92:419-429, 2018. 1", + "[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1", + "[31] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2, 3, 7", + "[32] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1", + "[33] Joffrey L. Leevy, John Hancock, Richard Zuech, and Taghi M. Khoshgoftaar. Detecting cybersecurity attacks across different network features and learners. Journal of Big Data, 8(1):38, 2021. 1", + "[34] Xuhong Li, Haoyi Xiong, Xingjian Li, Xuanyu Wu, Xiao Zhang, Ji Liu, Jiang Bian, and Dejing Dou. Interpretable deep learning: Interpretation, interpretability, trustworthiness, and beyond. Knowledge and Information Systems, 64 (12):3197-3234, 2022. 2", + "[35] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 5", + "[36] OpenAI. Gpt-4 technical report, 2024. 2, 4", + "[37] Akin Ozcift and Arif Gulten. Classifier ensemble construction with rotation forest to improve medical diagnosis performance of machine learning algorithms. Computer Methods and Programs in Biomedicine, 104(3):443-451, 2011. 1", + "[38] Yi Peng, Guoxun Wang, Gang Kou, and Yong Shi. An empirical study of classification algorithm evaluation for financial risk prediction. Applied Soft Computing, 11(2):2906-2915, 2011. 1", + "[39] Emmanuel Pintelas, Meletis Liaskos, Ioannis E Livieris, Sotiris Kotsiantis, and Panagiotis Pintelas. Explainable machine learning framework for image classification problems: case study on glioma cancer prediction. Journal of imaging, 6(6):37, 2020. 1", + "[40] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 8", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 91, + 90, + 482, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[42] R. Joshua Samuel Raj, S. Jeya Shobana, Irina Valeryevna Pustokhina, Denis Alexandrovich Pustokhin, Deepak Gupta, and K. Shankar. Optimal feature selection-based medical image classification using deep learning model in internet of medical things. IEEE Access, 8:58006-58017, 2020. 1", + "[43] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8848, 2020. 2", + "[44] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1135-1144, New York, NY, USA, 2016. Association for Computing Machinery. 1", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 8", + "[47] Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. Interpretable machine learning: Fundamental principles and 10 grand challenges. Statistic Surveys, 16:1-85, 2022. 2", + "[48] Axel Sauer and Andreas Geiger. Counterfactual generative networks. (arXiv:2101.06046), 2021. arXiv:2101.06046 [cs]. 3", + "[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 7", + "[50] Andy Shih, Arthur Choi, and Adnan Darwiche. A symbolic approach to explaining bayesian network classifiers. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, page 5103-5111. AAAI Press, 2018. 1", + "[51] Julio Silva-Rodriguez, Hadi Chakor, Riadh Kobbi, Jose Dolz, and Ismail Ben Ayed. A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision. Medical Image Analysis, 99:103357, 2025. 5", + "[52] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 2014. 1", + "[53] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14808", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. (arXiv:1911.00483), 2020. arXiv:1911.00483 [cs]. 3", + "[55] Sumedha Singla, Motahhare Eslami, Brian Pollack, Stephen Wallace, and Kayhan Batmanghelich. Explaining the black-box smoothly- a counterfactual approach. (arXiv:2101.04230), 2022. arXiv:2101.04230 [cs, eess]. 3", + "[56] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. Advances in neural information processing systems, 32, 2019. 2", + "[57] Gregor Stiglic, Primoz Kocbek, Nino Fijacko, Marinka Zitnik, Katrien Verbert, and Leona Cilar. Interpretability of machine learning-based prediction models in healthcare. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 10(5):e1379, 2020. 2", + "[58] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, et al. Selective annotation makes language models better few-shot learners. arXiv preprint arXiv:2209.01975, 2022. 4", + "[59] Prabu Subramani, Khalid Sattar, Rocio De Prado, Balasubramanian Girirajan, and Marcin Wozniak. Multi-classifier feature fusion-based road detection for connected autonomous vehicles. Applied Sciences, 11(17):7984, 2021. 1", + "[60] Jie Sun and Hui Li. Financial distress prediction based on serial combination of multiple classifiers. Expert Systems with Applications, 36(4):8659-8666, 2009. 1", + "[61] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2020. 5", + "[62] Manoj Thakur and Deepak Kumar. A hybrid financial trading support system using multi-category classifiers and random forest. Applied Soft Computing, 67:337-349, 2018. 1", + "[63] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 7", + "[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[65] Wahyono, Laksono Kurnianggoro, Joko Hariyono, and Kang-Hyun Jo. Traffic sign recognition system for autonomous vehicle using cascadesvm classifier. In IECON 2014 - 40th Annual Conference of the IEEE Industrial Electronics Society, pages 4081-4086, 2014. 1", + "[66] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8", + "[67] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 7", + "[68] Nina Weng, Paraskevas Pegios, Eike Petersen, Aasa Feragen, and Siavash Bigdeli. Fast diffusion-based counterfactuals for" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shortcut removal and generation. In European Conference on Computer Vision, pages 338-357. Springer, 2025. 3, 7", + "[69] M. Wiggins, A. Saad, B. Litt, and G. Vachtsevanos. Evolving a bayesian classifier for ecg-based age classification in medical applications. Applied Soft Computing, 8(1):599-608, 2008. 1", + "[70] Junde Wu, Jiayuan Zhu, and Yunli Qi. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187, 2024. 8", + "[71] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 8", + "[72] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. (arXiv:2004.03383), 2020. arXiv:2004.03383 [cs]. 2", + "[73] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. Advances in neural information processing systems, 33: 20554-20565, 2020. 3", + "[74] Muhammad Bilal Zafar, Michele Donini, Dylan Slack, Cédric Archambeau, Sanjiv Das, and Krishnamaram Kenthapadi. On the lack of robust interpretability of neural text classifiers. arXiv preprint arXiv:2106.04631, 2021. 1", + "[75] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2", + "[76] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1", + "[77] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2", + "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2", + "[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5" + ], + "bbox": [ + 516, + 92, + 903, + 770 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14809", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_model.json b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39baa49c1d27343e36c959576bf8d739fa692eaa --- /dev/null +++ b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.868, + 0.153 + ], + "angle": 0, + "content": "Explaining in Diffusion: Explaining a Classifier with Diffusion Semantics" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.181, + 0.374, + 0.217 + ], + "angle": 0, + "content": "Tahira Kazimi† Virginia Tech" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.221, + 0.398, + 0.233 + ], + "angle": 0, + "content": "tahirakazimi@vt.edu" + }, + { + "type": "text", + "bbox": [ + 0.466, + 0.181, + 0.58, + 0.217 + ], + "angle": 0, + "content": "Ritika Allada† Virginia Tech" + }, + { + "type": "text", + "bbox": [ + 0.456, + 0.221, + 0.59, + 0.233 + ], + "angle": 0, + "content": "ritika88@vt.edu" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.182, + 0.771, + 0.217 + ], + "angle": 0, + "content": "Pinar Yanardag Virginia Tech" + }, + { + "type": "text", + "bbox": [ + 0.65, + 0.221, + 0.767, + 0.234 + ], + "angle": 0, + "content": " pineary@vt.edu" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.245, + 0.665, + 0.259 + ], + "angle": 0, + "content": "https://explain-in-diffusion.github.io" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.283, + 0.29, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.283, + 0.488, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.283, + 0.695, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.283, + 0.895, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.49, + 0.906, + 0.56 + ], + "angle": 0, + "content": "Figure 1. DiffEx explains the decisions of domain-specific classifiers by identifying the most influential semantics affecting their predictions. Classifier scores for each example are displayed in the top-left corner, demonstrating how classifier predictions change in response to the manipulation of different semantics (original images are shown with red borders). Our approach is capable of explaining classifiers that concentrate on individual concepts such as faces or animals (top row) as well as those that manage complex scenes involving multiple objects, such as a formal/casual fit in a fashion context (bottom row)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.573, + 0.327, + 0.589 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.605, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Classifiers are important components in many computer vision tasks, serving as the foundational backbone of a wide variety of models employed across diverse applications. However, understanding the decision-making process of classifiers remains a significant challenge. We propose DiffEx, a novel method that leverages the capabilities of text-to-image diffusion models to explain classifier decisions. Unlike traditional GAN-based explainability models, which are limited to simple, single-concept analyses and typically require training a new model for each classifier, our approach can explain classifiers that focus on single concepts (such as faces or animals) as well as those that handle complex scenes involving multiple concepts. DiffEx employs vision-language models to create a hierarchical list of semantics, allowing users to identify not only the overarching semantic influences on classifiers (e.g., the 'beard' semantic in a facial classifier) but also their sub-types, such as 'goatee' or 'Balbo' beard. Our experiments demonstrate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.635 + ], + "angle": 0, + "content": "that DiffEx is able to cover a significantly broader spectrum of semantics compared to its GAN counterparts, providing a hierarchical tool that delivers a more detailed and fine-grained understanding of classifier decisions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.648, + 0.645, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Classifiers are fundamental to computer vision tasks, forming the backbone of many models used in a broad spectrum of applications [15, 21, 30, 32, 52]. Their ability to generalize across tasks and adapt to new domains with minimal retraining makes them highly transferable, and thus they are employed extensively in fields such as healthcare [37, 39, 42, 69], finance [38, 60, 62], security [2, 29, 33], and autonomous systems [5, 59, 65]. Despite their versatility and widespread utility, understanding the decision-making process of classifiers remains a significant challenge [3, 44, 50, 74, 76]. This challenge comes mainly from their \"black box\" nature. As images move through the deep layers of the network, the features used for classification become more abstract and harder to understand. The lack of interpretable features in such mod-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contributions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.455 + ], + "angle": 0, + "content": "els raises critical concerns, particularly in high-stakes environments such as medical diagnosis, where understanding the reasoning behind a model's prediction is crucial for ensuring trust, accountability, and informed decision-making [10, 34, 47, 57, 77]. Explaining classifier decisions is crucial for enhancing the transparency and reliability of these models. Prior research [31] has used generative adversarial networks (GANs) [18] to interpret classifier decisions by generating counterfactual examples that manipulate GAN latent semantics. These manipulations illustrate how changes in specific attributes, like the addition of the eyeglasses semantic, impact classifier outputs. However, GANs are often limited to single domains, such as facial images, and typically require training a new model per classifier, which is resource-intensive and time-consuming. Moreover, in GAN-based methods, understanding which latent semantics affect classifier decisions often requires manual intervention to identify and interpret relevant features, such as recognizing that a discovered semantic controls the eyeglasses attribute. This manual process is not only time-consuming but also less feasible in specialized fields like medicine, where identifying intricate attributes requires substantial domain expertise, making the approach impractical in critical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.464, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This limitation highlights the need for more automated and versatile approaches to interpret classifier decisions. Text-to-Image (T2I) diffusion models [45] emerge as a compelling alternative, widely recognized for their ability to generate high-quality images across various domains, which makes them a promising tool for explaining classifier decisions. These models offer the potential for a richer and more diverse set of semantic features compared to existing methods. However, their ability to interpret and utilize latent space semantics remains limited in the context of diffusion models. Existing techniques for identifying meaningful semantics rely largely on supervised approaches [7, 8], which require users to craft detailed text prompts to specify particular features for editing, such as a mustache. This process requires domain expertise to craft appropriate prompts; thus, a broad semantic corpus is essential to improve diffusion models for classifier explanations while reducing manual effort. In this paper, we first employ Vision-Language Models (VLMs) [36] to extract a largescale corpus covering domain-specific hierarchical semantics (see Fig. 2). Then, we introduce a training-free method, DiffEx, which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics. Our method provides explanations for both coarse and fine-grained semantics. For example, it can recognize a 'beard' as a coarse semantic influencing age classification scores and also demonstrate how specific beard types (such as 'Balbo' or 'Anchor' beards) impact the classifier's scores." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.091, + 0.875, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.231, + 0.907, + 0.314 + ], + "angle": 0, + "content": "Figure 2. Hierarchical List of Attributes for the Bird Domain. We use VLMs to extract a hierarchical corpus of semantics within a given domain. This representation helps illustrate how different attributes are grouped within a broader domain, facilitating a better understanding of how each semantic contributes to the overall decision-making process of a classifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.46 + ], + "angle": 0, + "content": "This hierarchical strategy provides users with an overview of the most significant semantics for a classifier, allowing them to dig deeper into particular fine-grained semantics that are essential for understanding classifier behavior. Our qualitative and quantitative experiments reveal that DiffEx offers considerably richer and more comprehensive explanations for binary and multi-class classifiers across various domains, including facial features, retinal health, and plant pathology. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.465, + 0.907, + 0.525 + ], + "angle": 0, + "content": "- We propose DiffEx, a training-free approach using VLMs and T2I diffusion models to explain classifier decisions. To the best of our knowledge, this is the first hierarchical approach that explains classifier decisions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.526, + 0.906, + 0.569 + ], + "angle": 0, + "content": "- Our method employs a VLM to develop a comprehensive semantic corpus that spans multiple domains, with our source code publicly available to support future research." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.57, + 0.905, + 0.645 + ], + "angle": 0, + "content": "- Unlike GAN-based methods, our approach can address classifiers that focus on single concepts (such as an 'age' classifier analyzing a headshot of a person) and also extend to classifiers for complex scenes (such as a 'modern/traditional architecture' classifier)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.646, + 0.905, + 0.706 + ], + "angle": 0, + "content": "- Our method offers more comprehensible explanations for classifiers in applications ranging from facial recognition to retinal health compared to prior approaches and is adaptable to both binary and multiclassifiers." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.465, + 0.907, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.724, + 0.655, + 0.739 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Traditional research focuses mainly on heat maps and patch-based extractions to explain classifier decisions. Specifically, class activation and saliency maps attempt to emulate human visual strategies by focusing on the image regions most relevant to a specific class [43, 49, 53, 56, 72, 75, 78]. While these maps highlight the object or part of an image that most influences the classifier's decision, they fail to reveal which specific, fine-grained object attributes (e.g., color, pattern, or texture) impact the classification. Other approaches try to explain classifier outputs by an-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14800" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 3. An Overview of DiffEx. Our pipeline processes a set of sample domain-specific images and a text prompt using a VLM to generate a hierarchical semantic corpus of attributes relevant to a specific domain. Based on this corpus, DiffEx identifies and ranks the most influential features affecting the classifier's decisions, sorting them from most to least impactful (rightmost image). The hierarchical explanation of semantics (like a \"beard\" and its subtypes) provides a fine-grained understanding of which features drive classifier outputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.312, + 0.485, + 0.687 + ], + "angle": 0, + "content": "alyzing extracted image patches [16, 73]; however, these methods can only reveal spatially localized attributes. Previous studies have leveraged generative models such as variational autoencoders (VAEs) [19] or GANs [17, 48, 54, 55]. The most similar research to ours, StylEx [31], introduces a GAN-based approach to identify various attributes that influence a classifier's decisions. However, this method has several limitations. Firstly, it requires training a new GAN for each classifier, which can be resource-intensive and time-consuming. Secondly, each identified attribute requires manual labeling, which can require domain expertise. Lastly, StylEx only uncovers a limited range of semantic attributes, potentially overlooking others that might significantly affect a classifier's scores. Additionally, a notable limitation of the StylEx method and similar GAN models is their focus on single concepts, such as human or animal faces, or individual objects like leaves, rather than entire scenes. In contrast, diffusion models, known for their robust capability to generate complex and detailed entire scenes, offer a significant advantage. Our approach leverages the power of diffusion models to cover a wider spectrum of visual contexts, and enhances the versatility and applicability of our method across various classifiers that assess not only individual elements but also the interaction and composition of entire scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Recent approaches have begun using diffusion models to generate counterfactual examples. One method utilizes shortcut learning to generate counterfactual images but fails to make semantically meaningful edits for certain attributes [68]. Another study explores modifying the diffusion process via adaptive parametrization and cone regularization to produce realistic counterfactual images; however, this approach depends on a robust model, which can be difficult to train [4]. [26] explored counterfactual image generation, however their approach is computationally demanding and uses DDPM [22] models trained on single domains. As a result, it does not take advantage of large-scale latent diffusion models like Stable Diffusion [46], which can handle more complex scenes. Furthermore, even though some" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.908, + 0.508 + ], + "angle": 0, + "content": "studies have leveraged diffusion models to generate realistic counterfactuals in high-stakes domains [24], such as the medical field, they use a less efficient base model for the image generation process and focus predominantly on single-attribute modifications. Furthermore, to the best of our knowledge, there is no existing research that explores a hierarchical explanation of classifiers. Such an approach would systematically unpack the layers of influence that different semantic levels have on a classifier's scoring mechanism. This gap highlights a significant opportunity to enhance understanding by detailing how various semantic categories and their sub-types contribute to the decisions made by classifiers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.514, + 0.649, + 0.531 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.631 + ], + "angle": 0, + "content": "We first discuss the background on identifying attributes that influence classifier decisions through changes in logits. Next, we introduce our method, which includes curating hierarchical attributes using VLMs and a novel algorithm inspired by beam search to pinpoint attributes that affect classifier scores. Our pipeline is detailed in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.638, + 0.645, + 0.654 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.749 + ], + "angle": 0, + "content": "StylEx [31] identifies semantics that meaningfully influence classifier decisions by ranking each attribute based on its impact on the classifier's logit outputs. This ranking process aims to identify and select attributes that best explain the classifier's behavior in a given context, such as understanding the factors influencing age classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.909, + 0.886 + ], + "angle": 0, + "content": "Given a semantic corpus \\( S \\) and a set of \\( N \\) images, counterfactual images are generated to analyze the influence of semantic attributes on classifier decisions. For each original image \\( x_{i} \\), an edited version \\( x^{\\prime}_{i} = g(x_{i}, s) \\) is generated by applying a semantic attribute \\( s \\in S \\) through a transformation function \\( g \\). This transformation highlights the effect of each semantic attribute on a classifier's output. Moreover, the logit difference for each attribute measures how the classifier's score changes due to the presence of \\( s \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "The influence of each attribute \\( s \\) is quantified as the aver" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.476, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.483, + 0.314 + ], + "angle": 0, + "content": "Figure 4. Top-7 Facial Attributes for Age Classification. DiffEx identifies key attributes and their top subtypes for the age classifier. Edited images are arranged hierarchically, with scores for the \"young\" label shown at the top-left of each image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.349 + ], + "angle": 0, + "content": "age logit difference between the original and edited images across a set of sample images, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.36, + 0.483, + 0.401 + ], + "angle": 0, + "content": "\\[\nI (s) = \\left| f \\left(X ^ {\\prime}, s, y\\right) - \\frac {1}{N} \\sum_ {i = 1} ^ {N} C \\left(x _ {i}, y\\right) \\right| \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.415, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\nf (X, s, y) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} C \\left(g \\left(x _ {i}, s\\right), y\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.569 + ], + "angle": 0, + "content": "where \\( f(X,s,y) \\) denotes the average classifier's logit score for target class \\( y \\) on image set \\( X \\) edited with attribute \\( s \\). Here, \\( y \\) represents the unedited images' target labels, such as the \"young\" label for an age classifier. This influence score \\( I(s) \\) captures the average impact of each semantic attribute on classifier decisions by reflecting the logit score change due to attribute manipulation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.579, + 0.226, + 0.594 + ], + "angle": 0, + "content": "3.2. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.707 + ], + "angle": 0, + "content": "We first employ Vision-Language Models (VLMs) [36] to extract a large-scale corpus covering domain-specific hierarchical semantics. Then, we introduce a training-free method, \\(DiffEx\\), which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.337, + 0.731 + ], + "angle": 0, + "content": "3.2.1. VLM-Based Semantic Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While the StylEx method outlined in Section 3.1 employs a logit-based approach to identify semantics, it requires manually labeling each attribute extracted from the trained GAN model. Moreover, this method does not support the explanation of hierarchical attributes. Therefore, we first compile a large-scale set of semantics using VLMs. Given a domain \\(d\\), such as the 'facial domain,' our objective is to identify a comprehensive set of domain-specific attributes from a collection of domain-specific images, denoted as \\(N_{d}\\). To accomplish this, we utilize a Vision-Language Model (VLM) [36] to extract a range of relevant features, represented by" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.652, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 DiffEx" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.111, + 0.907, + 0.156 + ], + "angle": 0, + "content": "Require: Hierarchical structure \\(\\mathcal{H}\\) with semantic groups and features, beam width \\(B\\), classifier or scoring function \\(f\\), scoring threshold \\(\\delta\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.157, + 0.8, + 0.171 + ], + "angle": 0, + "content": "Ensure: Optimal semantics maximizing \\( I \\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.172, + 0.905, + 0.201 + ], + "angle": 0, + "content": "1: Initialize \\(S \\gets\\) root-level groups in \\(\\mathcal{H}\\) {Initial candidate set at top-level groups}" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.203, + 0.693, + 0.215 + ], + "angle": 0, + "content": "2: Initialize beam \\(\\mathcal{B}\\gets \\emptyset\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.526, + 0.217, + 0.905, + 0.247 + ], + "angle": 0, + "content": "3: Score each candidate \\( s \\in S \\) using the scoring function \\( f(X, s, y) \\) for \\( y \\) class label" + }, + { + "type": "ref_text", + "bbox": [ + 0.526, + 0.248, + 0.905, + 0.291 + ], + "angle": 0, + "content": "4: Select top \\( B \\) candidates with \\( I(b) \\geq \\delta \\) and store in beam \\( \\mathcal{B} \\) and \\( S \\) {Apply thresholding to filter relevant candidates}" + }, + { + "type": "ref_text", + "bbox": [ + 0.526, + 0.293, + 0.651, + 0.306 + ], + "angle": 0, + "content": "5: while \\(\\mathcal{B} \\neq \\emptyset\\) do" + }, + { + "type": "ref_text", + "bbox": [ + 0.526, + 0.308, + 0.689, + 0.321 + ], + "angle": 0, + "content": "6: Initialize \\(S_{\\mathrm{next}} \\gets \\emptyset\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.526, + 0.323, + 0.747, + 0.335 + ], + "angle": 0, + "content": "7: for each candidate \\( b \\in \\mathcal{B} \\) do" + }, + { + "type": "ref_text", + "bbox": [ + 0.526, + 0.338, + 0.905, + 0.366 + ], + "angle": 0, + "content": "8: Expand \\( b \\) by adding sub-features from its next level in \\( \\mathcal{H} \\) to form new candidates" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.369, + 0.9, + 0.396 + ], + "angle": 0, + "content": "9: for each new combination \\( b' \\) expanded from \\( b \\) to 10: if \\( I(b') > I(b) \\) then" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.399, + 0.705, + 0.412 + ], + "angle": 0, + "content": "11: Add \\( b' \\) to \\( S_{\\mathrm{next}} \\)." + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.414, + 0.636, + 0.426 + ], + "angle": 0, + "content": "12: end if" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.429, + 0.629, + 0.441 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.444, + 0.613, + 0.456 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.459, + 0.654, + 0.472 + ], + "angle": 0, + "content": "15: Set \\(\\mathcal{B}\\gets S_{\\mathrm{next}}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.474, + 0.68, + 0.488 + ], + "angle": 0, + "content": "16: Append \\(S_{\\mathrm{next}}\\) to \\(S\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.489, + 0.613, + 0.501 + ], + "angle": 0, + "content": "17: end while" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.504, + 0.904, + 0.533 + ], + "angle": 0, + "content": "18: Return highest-scoring combination from final \\( S \\) as the optimal joint semantic combination" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.111, + 0.907, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.717 + ], + "angle": 0, + "content": "\\(\\mathcal{H}\\). Employing in-context learning [9, 58], we prompt the VLM with a small set of images \\(N_{d}\\), along with a detailed task description and examples of desired outputs. This process allows us to generate a substantial semantic corpus of keywords, \\(\\mathcal{H}\\), that captures the fine-grained attributes relevant to the domain. The resulting corpus, \\(\\mathcal{H}\\), comprises a comprehensive collection of keywords that covers full spectrum of the domain's attributes. We use this dataset to hierarchically explain how different attributes influence classifier decisions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.73, + 0.607, + 0.743 + ], + "angle": 0, + "content": "3.2.2.DiffEx" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Considering the extensive number of semantics identified using a VLM, it is computationally expensive to evaluate every possible semantic or combination of attributes. We introduce DiffEx, an efficient approach inspired by beam search to explain classifier decisions (see Algorithm 1). This method leverages our hierarchical semantic corpus to streamline the process. Our approach iteratively refines candidate attributes by expanding only the most impactful semantic paths at each hierarchical level, with each path's relevance guided by a scoring function that assesses the clas" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14802" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.462, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.442 + ], + "angle": 0, + "content": "Figure 5. Top-7 Discovered Attributes Across Different Animal Domains. Our method successfully identifies key attributes for multiple domains, such as bird, wildcat, and pet species. The original images are depicted with red borders while the edited images are depicted with black borders. For the pet species domain, we used a binary classifier and for the bird and wildcat species domains, we used a multi-classifier." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.538 + ], + "angle": 0, + "content": "sifier's response to each semantic feature. Formally, the scoring function \\( f \\) calculates the average classifier \\( C \\) logit scores across \\( N \\) images generated using each semantic attribute \\( s \\), as defined in Eq. (2), where \\( g \\) represents the generative diffusion model applied to each sample image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We begin with an initial candidate set \\( S \\), which includes high-level groups from the semantic hierarchy (e.g., broader categories like 'mouth features' or 'eye features'). Each candidate in \\( S \\) is evaluated by the scoring function \\( f(X, s, y) \\), which quantifies the influence of the candidate on the classifier's output. Only the top \\( B \\) candidates that surpass a predefined score threshold \\( \\delta \\) are retained, setting a beam width that limits the search to the most impactful candidates. For each candidate in the beam, the algorithm proceeds by expanding to the next hierarchical level, incorporating more specific sub-features (e.g., for \"mouth features,\" sub-features such as \"beard\" and \"mustache\" are included). For each expanded candidate \\( b' \\), the scoring function in Eq. (2) is re-evaluated. Only candidates with a score exceeding that of their parent \\( I(b') > I(b) \\) are retained in the next candidate set \\( S_{next} \\), ensuring that only those additions that yield significant incremental impact are added. This process of expansion, scoring, and filtering continues iteratively, moving from general to more specific attributes at each hierarchical level. By dynamically adjusting the candidate set based on beam width \\( B \\) and scoring threshold \\( \\delta \\), our method remains computationally efficient, focusing on high-impact combinations rather than exhaustively evaluating all possible attribute pairings. For generating counter" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.235, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Figure 6. Visual Comparison of Key Attributes Identified by StylEx and DiffEx in the Plant Health and Retinal Disease Domains. This figure illustrates the enhanced capability of our method in identifying a broader set of significant attributes compared to StylEx within the plant health and retinal disease domains. DiffEx uncovers more detailed and diagnostically relevant features, such as \"leaf vein color\" and \"macular hole,\" which provide deeper insights into leaf and retina health." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.35, + 0.907, + 0.396 + ], + "angle": 0, + "content": "factual images, we utilize an off-the-shelf diffusion-based editing tool, Ledits++ [8]. However, our approach is versatile enough to accommodate any editing method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.646, + 0.424 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.909, + 0.766 + ], + "angle": 0, + "content": "Experimental Setup. Our experiments utilize Ledits++ [8] and Stable Diffusion XL (SDXL) [40] for generating counterfactual images. Twenty-five time steps are omitted to boost computational efficiency while preserving the quality of edits. The edit threshold, a hyperparameter that dictates the global application scope of edits, is adjusted based on each domain. We test our method across diverse classifiers to evaluate its effectiveness in explaining model behavior through semantic influence. Specifically, we utilize classifiers trained on facial attributes data (i.e. age and gender classifiers) [28] as well as plant health [23], retinal disease [13], bird species [64], wildcat/pet species [11], fashion [35], places [79], and food data [6]. The experiments for the face and plant health domains utilize a CLIP-based classifier [41] to evaluate and interpret edits, while experiments within the bird, wildcat, pet, food, fashion, and places domains use CNN-based classifiers. These CNN classifiers were built on the EfficientNet [61] architecture and achieved an accuracy of over 95 percent on their test sets. For the retinal disease domain, we utilized the FLAIR model [51], which is based on a pre-trained vision-language model, to classify various retina scans." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.74, + 0.79 + ], + "angle": 0, + "content": "4.1. Qualitative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Explaining Classifiers with Single Attributes. We used various classifiers to analyze fine-grained semantic features for age categorization, species identification, leaf health assessment, retinal disease detection, etc. In the figures illustrating the visual results, the images with red borders indicate the original, unedited versions, while those with black borders are the edited versions, unless noted otherwise." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14803" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.183 + ], + "angle": 0, + "content": "
Face (Age)Bird (Species)Leaves (Health)Retina Scans (Disease)Wildcat (Species)Pet (Cat/Dog)
StylExOursStylExOursStylExOursStylExOursStylExOursStylExOurs
Skin PigmentationEyebrowBelly ColorUpperparts ColorBase Leaf ColorBase ColorExudatesGlaucomaSpotsStripesOpen MouthNose Color
Eyebrow ThicknessMakeupUpperparts ColorHead ColorApex ColorVein ColorCotton Wool SpotsSubretinal HemorrhageBlack Tear MarkOpen EyesClosed MouthWhiskers
EyeglassesMustache TypeWing PatternBeak ShapeSpotsApex ColorHemorrhagesIntraretinal HemorrhageEye Shape + SizeNose ColorEye ShapeEye Color
Hair ColorTeethBeak ColorBeak ColorBlightSpots PresenceClustered ExudatesMacular HoleXSpotsDropped EarsClosed Mouth
Lip Thickness + PositionLip VolumeHead ColorWing PatternHalosDisease PresenceXHard ExudatesXFur ColorPointed EarsOpen Mouth
BangsLip ColorBreast ColorEye ColorXBlight SizeXBlackened MaculaXManeEye CircumferenceSpots
Eye MakeupEyelashXThroat ColorXLeaf TextureXSoft ExudatesXEye ColorXCollar
Facial Hair ColorBeard TypeXWing ColorXSpots ColorXRetinal DruenXTongueXPointed Ears
XFacewearXCrest PresenceXDiscolorationXOptic Disc HemorrhageXPupil SizeXMouth Color
XHeadwearXFeather TextureXLeaf OrientationXCataractXWhiskersXFur Pattern
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.192, + 0.907, + 0.233 + ], + "angle": 0, + "content": "Table 1. Comparison of Top Attributes Across Different Domains and Classifiers. The table above contains a list of the top attributes discovered by DiffEx (Ours) vs. StylEx. The \\(\\pmb{x}\\) in the table indicates attributes that were not mentioned in StylEx. It is also important to note that \"cotton wool spots\" and \"soft exudates\" refer to the same condition within the retinal disease domain." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.233, + 0.905, + 0.281 + ], + "angle": 0, + "content": "
MethodCrest PresenceBeak ShapeThroat ColorFeather TextureEye ColorBeak ColorHead ColorUpperparts ColorAvg. Correct Response
Grad-CAM36%50%56%35%47%65%59%76%53%
StylEx68%85%79%82%74%68%91%65%76.5%
DiffEx (Ours)88%91%88%91%82%82%97%88%88.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.29, + 0.908, + 0.346 + ], + "angle": 0, + "content": "Table 2. Comparison with Other Explainability Methods. The table above displays the percentage of correct attribute selections for the bird class, as chosen by users when viewing outputs from different explainability methods. It also includes the average percentage of correct responses across all attributes for each method. As shown, for each attribute presented, the majority of users identified the correct attribute when viewing the output generated by DiffEx." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.482, + 0.608 + ], + "angle": 0, + "content": "Face Domain: In the facial domain, as illustrated in Fig. 4, DiffEx identifies and ranks key attributes impacting age classification. For example, features such as \"makeup styles,\" \"lip volume,\" and \"accessories\" (e.g. \"hairbands\") are associated with perceived youthfulness, whereas attributes like \"facial hair\" and \"eyeglasses\" are linked with perceived older age. Notably, the eyeglasses attribute consistently reduces the classifier's score for the \"young\" label, reflecting its association with older demographics. Additionally, DiffEx uncovers hierarchical attribute structures, demonstrating how subcategories within a feature can have varying effects on classifier outcomes. For instance, as shown in Fig. 4, different beard styles impact the perceived age differently (e.g. a \"Balbo beard\" significantly increases the age classification score more than a \"full beard\"). Additional examples of hierarchical explanations of age classifiers are detailed in the appendix (S10)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.611, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Animal Domain: Furthermore, DiffEx explains classifiers for a variety of animal types. Fig. 5 highlights the top-7 most influential attributes in the bird, wildcat, and pet domains where our method was able to identify fine-grained semantics to explain classifier behavior. For example, in the wildcat classifier, attributes like stripes and manes are critical in distinguishing wildcats from other species. Explaining these classifiers also reveals potential biases. For instance, the presence of a collar significantly increases the likelihood of an animal being classified as a dog. This bias may stem from the training dataset where images of dogs more frequently featured collars compared to cats." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Medical and Plant Health Domains: Fig. 6 demonstrates the results for retinal disease and plant health domains. For the retinal disease domain, we use the FLAIR model to classify images of retinal fundus scans. This model utilizes detailed domain expert knowledge descriptions, such as \"no hemorrhages, microaneurysms, or exudates\" compared to general descriptions like \"no diabetic retinopathy\" to aid in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.397 + ], + "angle": 0, + "content": "its classification. For the plant health domain, the CLIP classifier we use looks at features such as the presence of spots, fungus, etc. to make its decision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.908, + 0.852 + ], + "angle": 0, + "content": "Multi-Object Domains involving Complex Scenes: Unlike traditional GAN-based methods that primarily focus on single-object scenarios like a cropped face, DiffEx extends its utility by providing a list of relevant features for domains encompassing multiple objects, such as places, food, and fashion. Classifiers within these domains often evaluate multiple elements simultaneously. For example, a fashion classifier might determine the 'formal or casual fit' of an outfit by considering various components such as the hairstyle, top, and bottom. Similarly, an architectural classifier assessing whether a building appears 'urban or rural' may base its evaluation on not just the structure itself but also its exteriors and surroundings. Explaining the decision-making process of such classifiers is crucial, as they analyze multiple objects within a single image, adding complexity to their interpretative frameworks. Fig. 7 illustrates the top-7 attributes identified by DiffEx for the food, place, and fashion domains. Moreover, we demonstrate DiffEx on various diverse domains. For example, for the food domain, DiffEx was able to discover that removing \"caviar\" or a \"plate decoration\" from the image made the food appear to be more perceived as fast food compared to fine dining. For the places domain, DiffEx was able to find that adding a \"tractor\" to an image or a \"dirt road\" made the place seem more \"rural\" than \"urban.\" In the fashion domain, the style of a neckline significantly influences whether an outfit was classified as \"formal\" or \"casual.\" For instance, a V-neckline was often associated with a more casual look, whereas a classic boat neckline was generally perceived as more formal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Explaining Classifiers with Joint Attributes. Our method goes beyond analyzing individual attributes by identifying attribute combinations that collectively improve classifier" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14804" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.473, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Figure 7. Top-7 Attributes Discovered by DiffEx for Multi-Object Domains involving Complex Scenes. Compared to existing methods, DiffEx was used to identify meaningful attributes for multi-object domains, such as \"food,\" \"places,\" and \"fashion.\" By identifying semantic features that make food appear \"healthier\" or part of \"fine dining,\" as well as attributes that give a place a more \"rural\" or \"modern\" feel, DiffEx can serve as a powerful tool for understanding perceptions through targeted edits." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.634, + 0.477, + 0.686 + ], + "angle": 0, + "content": "
RatingBird DomainFace Domain
Edit Quality3.386 ± 0.2233.659 ± 0.248
Disentanglement3.163 ± 0.1973.204 ± 0.213
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Table 3. Edit Quality and Disentanglement Ratings. These scores across various domains from User Study 1 are shown above." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.885 + ], + "angle": 0, + "content": "interpretation. While single attributes may have a minimal impact on logits, their combined effect can substantially influence the classifier's output, uncovering subtle interactions that shape decision-making. For example, as shown in Fig. 8, Section C, individual changes in \"lip color\" or \"eye makeup\" result in minimal score changes, but when modified together, they make the subject appear significantly younger. This analysis shows how classifiers may respond more robustly to specific attribute combinations. Additional examples are provided in the appendix (S7, S8, S9)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Extending DiffEx for Multi-Classifier Analysis. We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "adapt DiffEx for multi-classifier applications to uncover semantic attributes essential for tasks like retinal disease classification (Fig. 6) and bird and wildcat species identification (Fig. 5). This approach highlights key features such as \"upperparts color\" and \"beak shape\" for bird species, \"stripes\" and \"nose color\" for wildcat species, and types of hemorrhages and exudates for retinal conditions. The results in Fig. 6 for the retinal diseases domain and Fig. 5 for the bird/wildcat species domain demonstrate that modifying these attributes significantly alters classifier scores, impacting assessments for species identification and disease likelihood." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.908, + 0.818 + ], + "angle": 0, + "content": "Qualitative Comparisons. In Fig. 6, we present a visual comparison of the top attributes and classification scores identified by DiffEx against those identified by StylEx for the plant health and retinal disease domains. As illustrated, DiffEx successfully uncovers a broader set of semantically meaningful attributes compared to StylEx. For instance, DiffEx identifies detailed attributes such as \"leaf vein color\" and \"spots color,\" in addition to the more general attributes found by StylEx, like \"leaf base color\" and \"apex color.\" Table 1 provides a comprehensive comparison of the top features identified by StylEx and our method, highlighting DiffEx's superior ability to uncover semantics that are crucial for the classifier's decision-making process. In particular, the table presents an extended list of features across domains such as faces, bird species, plant health, retina scans, wildcat species, and pet types, all of which influence a classifier's score—further emphasizing DiffEx's ability to uncover fine-grained and contextually rich features. Additionally, in Figure 8 in Sections A and B, we visually compare DiffEx with other diffusion-based, counterfactual generation methods, such as FastDiME [68], DiME [26], and ACE [27] in the face and medical domains. As seen in Figure 8, Section B, DiffEx is able to preserve important features, such as a person's eyes, glasses, etc., in various cases, such as removing a smile or making an older person look younger. Furthermore, Figure 8, Section A shows DiffEx's ability to generate counterfactuals in various medical subdomains, such as skin lesions and chest x-rays. Specifically, for the skin lesion domain, DiffEx was able to remove ruler markings in images from the ISIC dataset [12, 63], while in the chest x-rays domain, DiffEx was able to remove chest drains and pacemakers from images from the NIH [67] dataset and CheXpert [25] dataset. As seen in Figure 8, Section A, DiffEx achieved similar results to FastDiME (Med) [68], a diffusion-based counterfactual model that was trained on medical images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.75, + 0.848 + ], + "angle": 0, + "content": "4.2. Quantitative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Baselines. To quantitatively evaluate the effectiveness of DiffEx compared to other explainability methods, such as StylEx [31] and Grad-CAM [49], we conducted a series of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14805" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.179, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 8. More Comparisons and Exploring Joint Attributes. (a) We demonstrate DiffEx's ability to produce counterfactuals comparable to FastDiME-2+ in skin lesion and chest X-ray domains. (b) compares DiffEx with ACE, DiME, FastDiME, and FastDiME-2+, showing its capacity to add or remove attributes in a disentangled manner. (c) highlights DiffEx's effectiveness in explaining classifier decisions by analyzing the influence of joint attributes, with perceived age scores (e.g., \"younger\" label) displayed to show how attribute pairings affect classifier outputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.312, + 0.484, + 0.522 + ], + "angle": 0, + "content": "comprehensive user studies to assess how easily participants could identify the relevant features extracted by our method. User Study 1: Visual Quality and Disentanglement. To evaluate the visual quality and disentanglement of edited images for the face and bird domains, we conducted a user study with 50 participants on Prolific.com. For each domain, we showed pairs of unedited and edited images for ten attributes and asked users to assess whether the edited image contained the desired attribute and was disentangled. For each pair of images, we asked users to rate the edit and disentanglement from one to five, with five representing the highest score. Our results (see Table 3) indicate that our edits reflected the intended attributes while minimizing unrelated changes (see S6.1 in the appendix for more details)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.523, + 0.484, + 0.81 + ], + "angle": 0, + "content": "User Study 2: Comparison with Grad-CAM and StylEx. To evaluate how effectively our method, DiffEx, explains various semantics within a specific domain compared to other explainability methods (Grad-CAM and StylEx), we conducted a user study with 35 participants on Prolific. We focused on images from the bird domain, generating three sets of three images per attribute: one original image, one edited image, and one image illustrating the explainability method. For Grad-CAM, participants were shown a heatmap overlay on the edited image. For StylEx, we displayed the edited image alone, as this method requires users to manually label the edits. For DiffEx, participants viewed the edited image and the attribute automatically assigned by the VLM. We then asked users to select the attribute (e.g., \"beak color,\" \"crest presence\") that best explained the edited image from four choices. Results indicate that DiffEx significantly outperforms Grad-CAM and StylEx in explaining edited attributes in images (see Table 2). Please refer to S6.2 in the appendix for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "User Study 3: Disentanglement in Images with Multiple Attributes. We evaluated DiffEx's disentanglement ability when editing multiple attributes simultaneously via a user study. Users rated images on a scale of 1 to 5 (5 = very disentangled, 1 = entangled). DiffEx achieved an average score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "of 4.6 for single-attribute edits and 4.3 for edits involving four or more attributes, demonstrating effective handling of multiple edits (see S6.3 in the appendix for more details)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.909, + 0.333 + ], + "angle": 0, + "content": "Ablation Studies. To assess the robustness of our method, we tested various Vision-Language Models (VLMs) and Diffusion Models. For VLMs, we probed Llama 3.2 [20], Qwen2-VL [66], and Claude 3.5 Haiku [1] within the face domain, measuring semantic consistency via IoU scores against GPT-4. The results were 0.94, 0.87, and 0.93, respectively, which indicate substantial semantic overlap. For diffusion models, we compared top attributes from Stable Diffusion (SD) 1.5 and SD 2.0 [46] against SDXL [40] using Normalized Discounted Cumulative Gain (nDCG) to measure the attributes' overlap, achieving scores of 0.75 and 0.76, respectively. These results confirm the stability of our method across different models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.628, + 0.362 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.909, + 0.569 + ], + "angle": 0, + "content": "Limitations. While our approach provides valuable insights into classifier behavior through semantic edits, it has limitations. Relying on VLM-curated semantics restricts us to the quality and scope of the initial corpus, which may miss relevant features. Additionally, using off-the-shelf editing models like Ledits++ can result in entangled edits (a common issue in image editing algorithms [71]) or fail to capture fine-grained details like earlobes or nose rings (see Appendix S1). Nonetheless, our framework is flexible and can be enhanced with domain-specific adjustments, such as task-specific RAGs [70] or editing methods [7, 14]. RAGs can automatically retrieve semantics in specialized domains (e.g. medical) to improve generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.826 + ], + "angle": 0, + "content": "Conclusion. In this work, we introduce DiffEx, a novel approach for explaining classifier decisions by utilizing semantic edits within diffusion models. By harnessing the power of vision language models, we curate a comprehensive, hierarchical semantic corpus across various domains and propose a novel algorithm inspired by beam search to filter and rank the most impactful features. DiffEx ranks these semantic features based on their influence on classifier logits, capturing both individual and joint attribute effects, which are crucial for understanding complex classifier behaviors. Through experiments conducted on a wide range of domains—including face, bird, and medical classifiers—we showcase the robustness and adaptability of our approach. Our work provides a powerful tool for interpreting model decisions across diverse applications and promoting transparency in AI-driven classification systems." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14806" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.428, + 0.129 + ], + "angle": 0, + "content": "[1] The claude 3 model family: Opus, sonnet, haiku. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.131, + 0.484, + 0.186 + ], + "angle": 0, + "content": "[2] Mostofa Ahsan, Rahul Gomes, Md. Minhaz Chowdhury, and Kendall E. Nygard. Enhancing machine learning prediction in cybersecurity using dynamic feature selector. Journal of Cybersecurity and Privacy, 1(1):199-218, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.189, + 0.482, + 0.23 + ], + "angle": 0, + "content": "[3] Leila Amgoud. Explaining black-box classifiers: Properties and functions. International Journal of Approximate Reasoning, 155:40-65, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.232, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. Advances in Neural Information Processing Systems, 35:364-377, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.29, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[5] Ouahiba Azouaoui and Amine Chohra. Soft computing based pattern classifiers for the obstacle avoidance behavior of intelligent autonomous vehicles (iav). Applied intelligence, 16:249-272, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[7] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. SEGA: Instructing text-to-image models using semantic guidance. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[8] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[10] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. CoRR, abs/1912.01865, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.747, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[12] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.847, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[13] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[14] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. arXiv preprint arXiv:2312.05390, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.906, + 0.246 + ], + "angle": 0, + "content": "[16] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[17] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5744-5753, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[19] Yash Goyal, Amir Feder, Uri Shalit, and Been Kim. Explaining classifiers with causal concept effect (cace). arXiv preprint arXiv:1907.07165, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[20] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. (arXiv:1512.03385), 2015. arXiv:1512.03385. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.535, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.578, + 0.905, + 0.632 + ], + "angle": 0, + "content": "[23] David P. Hughes and Marcel Salathe. An open access repository of images on plant health to enable the development of mobile disease diagnostics. (arXiv:1511.08060), 2016. arXiv:1511.08060 [cs]. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[24] Indu Ilanchezian, Valentyn Boreiko, Laura Kuhlewein, Zwei Huang, Murat Seckin Ayhan, Matthias Hein, Lisa Koch, and Philipp Berens. Generating realistic counterfactuals for retinal fundus and oct images using diffusion models. arXiv preprint arXiv:2311.11629, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[25] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. In Proceedings of the Asian Conference on Computer Vision, pages 858-876, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Adversarial counterfactual visual explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16425-16435, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14807" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 4396-4405, Long Beach, CA, USA, 2019. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[29] Ziv Katzir and Yuval Elovici. Quantifying the resilience of machine learning classifiers used for cyber security. Expert Systems with Applications, 92:419-429, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[31] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[32] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[33] Joffrey L. Leevy, John Hancock, Richard Zuech, and Taghi M. Khoshgoftaar. Detecting cybersecurity attacks across different network features and learners. Journal of Big Data, 8(1):38, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[34] Xuhong Li, Haoyi Xiong, Xingjian Li, Xuanyu Wu, Xiao Zhang, Ji Liu, Jiang Bian, and Dejing Dou. Interpretable deep learning: Interpretation, interpretability, trustworthiness, and beyond. Knowledge and Information Systems, 64 (12):3197-3234, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[35] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.383, + 0.618 + ], + "angle": 0, + "content": "[36] OpenAI. Gpt-4 technical report, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[37] Akin Ozcift and Arif Gulten. Classifier ensemble construction with rotation forest to improve medical diagnosis performance of machine learning algorithms. Computer Methods and Programs in Biomedicine, 104(3):443-451, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[38] Yi Peng, Guoxun Wang, Gang Kou, and Yong Shi. An empirical study of classification algorithm evaluation for financial risk prediction. Applied Soft Computing, 11(2):2906-2915, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[39] Emmanuel Pintelas, Meletis Liaskos, Ioannis E Livieris, Sotiris Kotsiantis, and Panagiotis Pintelas. Explainable machine learning framework for image classification problems: case study on glioma cancer prediction. Journal of imaging, 6(6):37, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[40] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[42] R. Joshua Samuel Raj, S. Jeya Shobana, Irina Valeryevna Pustokhina, Denis Alexandrovich Pustokhin, Deepak Gupta, and K. Shankar. Optimal feature selection-based medical image classification using deep learning model in internet of medical things. IEEE Access, 8:58006-58017, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[43] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8848, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[44] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1135-1144, New York, NY, USA, 2016. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[47] Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. Interpretable machine learning: Fundamental principles and 10 grand challenges. Statistic Surveys, 16:1-85, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[48] Axel Sauer and Andreas Geiger. Counterfactual generative networks. (arXiv:2101.06046), 2021. arXiv:2101.06046 [cs]. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[50] Andy Shih, Arthur Choi, and Adnan Darwiche. A symbolic approach to explaining bayesian network classifiers. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, page 5103-5111. AAAI Press, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[51] Julio Silva-Rodriguez, Hadi Chakor, Riadh Kobbi, Jose Dolz, and Ismail Ben Ayed. A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision. Medical Image Analysis, 99:103357, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[52] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14808" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[54] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. (arXiv:1911.00483), 2020. arXiv:1911.00483 [cs]. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[55] Sumedha Singla, Motahhare Eslami, Brian Pollack, Stephen Wallace, and Kayhan Batmanghelich. Explaining the black-box smoothly- a counterfactual approach. (arXiv:2101.04230), 2022. arXiv:2101.04230 [cs, eess]. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[56] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[57] Gregor Stiglic, Primoz Kocbek, Nino Fijacko, Marinka Zitnik, Katrien Verbert, and Leona Cilar. Interpretability of machine learning-based prediction models in healthcare. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 10(5):e1379, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[58] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, et al. Selective annotation makes language models better few-shot learners. arXiv preprint arXiv:2209.01975, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[59] Prabu Subramani, Khalid Sattar, Rocio De Prado, Balasubramanian Girirajan, and Marcin Wozniak. Multi-classifier feature fusion-based road detection for connected autonomous vehicles. Applied Sciences, 11(17):7984, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[60] Jie Sun and Hui Li. Financial distress prediction based on serial combination of multiple classifiers. Expert Systems with Applications, 36(4):8659-8666, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[61] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[62] Manoj Thakur and Deepak Kumar. A hybrid financial trading support system using multi-category classifiers and random forest. Applied Soft Computing, 67:337-349, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[63] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.707 + ], + "angle": 0, + "content": "[65] Wahyono, Laksono Kurnianggoro, Joko Hariyono, and Kang-Hyun Jo. Traffic sign recognition system for autonomous vehicle using cascadesvm classifier. In IECON 2014 - 40th Annual Conference of the IEEE Industrial Electronics Society, pages 4081-4086, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[66] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[67] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[68] Nina Weng, Paraskevas Pegios, Eike Petersen, Aasa Feragen, and Siavash Bigdeli. Fast diffusion-based counterfactuals for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "shortcut removal and generation. In European Conference on Computer Vision, pages 338-357. Springer, 2025. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[69] M. Wiggins, A. Saad, B. Litt, and G. Vachtsevanos. Evolving a bayesian classifier for ecg-based age classification in medical applications. Applied Soft Computing, 8(1):599-608, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[70] Junde Wu, Jiayuan Zhu, and Yunli Qi. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[71] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[72] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. (arXiv:2004.03383), 2020. arXiv:2004.03383 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[73] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. Advances in neural information processing systems, 33: 20554-20565, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[74] Muhammad Bilal Zafar, Michele Donini, Dylan Slack, Cédric Archambeau, Sanjiv Das, and Krishnamaram Kenthapadi. On the lack of robust interpretability of neural text classifiers. arXiv preprint arXiv:2106.04631, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[75] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[76] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[77] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.713, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14809" + } + ] +] \ No newline at end of file diff --git a/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_origin.pdf b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d8dd41962cdff5511cfe77e55502377a68d098b5 --- /dev/null +++ b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/e12e2fee-c2c0-4bbc-99f8-f5d2cb6dbe06_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8428f7e6ad1420a3c881b92bdddd6fda305a73217eedda27a64e62a8d07d6f8d +size 9547863 diff --git a/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/full.md b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c97cc23143627dd5d19f1fc4c30c9fb106785d9 --- /dev/null +++ b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/full.md @@ -0,0 +1,291 @@ +# Explaining in Diffusion: Explaining a Classifier with Diffusion Semantics + +Tahira Kazimi† Virginia Tech + +tahirakazimi@vt.edu + +Ritika Allada† Virginia Tech + +ritika88@vt.edu + +Pinar Yanardag Virginia Tech + +pineary@vt.edu + +https://explain-in-diffusion.github.io + +![](images/8af0c4f0105c182e2920e6af801a517e6576fe80fed90eb68f759d23e71f9468.jpg) +Figure 1. DiffEx explains the decisions of domain-specific classifiers by identifying the most influential semantics affecting their predictions. Classifier scores for each example are displayed in the top-left corner, demonstrating how classifier predictions change in response to the manipulation of different semantics (original images are shown with red borders). Our approach is capable of explaining classifiers that concentrate on individual concepts such as faces or animals (top row) as well as those that manage complex scenes involving multiple objects, such as a formal/casual fit in a fashion context (bottom row). + +![](images/9775f157ca494465e96aadd14f7d0626789655281c8a0d1f16f98687cb4cbac7.jpg) + +![](images/7dbc047a4efdfabff0916122a8a1ba42b9fd46a217f1ad98b2c8003689fe873b.jpg) + +![](images/32ca24e47a20d40bf513c8582bf56588f3e9dbd49d33e16733add3bb12bdff51.jpg) + +# Abstract + +Classifiers are important components in many computer vision tasks, serving as the foundational backbone of a wide variety of models employed across diverse applications. However, understanding the decision-making process of classifiers remains a significant challenge. We propose DiffEx, a novel method that leverages the capabilities of text-to-image diffusion models to explain classifier decisions. Unlike traditional GAN-based explainability models, which are limited to simple, single-concept analyses and typically require training a new model for each classifier, our approach can explain classifiers that focus on single concepts (such as faces or animals) as well as those that handle complex scenes involving multiple concepts. DiffEx employs vision-language models to create a hierarchical list of semantics, allowing users to identify not only the overarching semantic influences on classifiers (e.g., the 'beard' semantic in a facial classifier) but also their sub-types, such as 'goatee' or 'Balbo' beard. Our experiments demonstrate + +that DiffEx is able to cover a significantly broader spectrum of semantics compared to its GAN counterparts, providing a hierarchical tool that delivers a more detailed and fine-grained understanding of classifier decisions. + +# 1. Introduction + +Classifiers are fundamental to computer vision tasks, forming the backbone of many models used in a broad spectrum of applications [15, 21, 30, 32, 52]. Their ability to generalize across tasks and adapt to new domains with minimal retraining makes them highly transferable, and thus they are employed extensively in fields such as healthcare [37, 39, 42, 69], finance [38, 60, 62], security [2, 29, 33], and autonomous systems [5, 59, 65]. Despite their versatility and widespread utility, understanding the decision-making process of classifiers remains a significant challenge [3, 44, 50, 74, 76]. This challenge comes mainly from their "black box" nature. As images move through the deep layers of the network, the features used for classification become more abstract and harder to understand. The lack of interpretable features in such mod- + +els raises critical concerns, particularly in high-stakes environments such as medical diagnosis, where understanding the reasoning behind a model's prediction is crucial for ensuring trust, accountability, and informed decision-making [10, 34, 47, 57, 77]. Explaining classifier decisions is crucial for enhancing the transparency and reliability of these models. Prior research [31] has used generative adversarial networks (GANs) [18] to interpret classifier decisions by generating counterfactual examples that manipulate GAN latent semantics. These manipulations illustrate how changes in specific attributes, like the addition of the eyeglasses semantic, impact classifier outputs. However, GANs are often limited to single domains, such as facial images, and typically require training a new model per classifier, which is resource-intensive and time-consuming. Moreover, in GAN-based methods, understanding which latent semantics affect classifier decisions often requires manual intervention to identify and interpret relevant features, such as recognizing that a discovered semantic controls the eyeglasses attribute. This manual process is not only time-consuming but also less feasible in specialized fields like medicine, where identifying intricate attributes requires substantial domain expertise, making the approach impractical in critical scenarios. + +This limitation highlights the need for more automated and versatile approaches to interpret classifier decisions. Text-to-Image (T2I) diffusion models [45] emerge as a compelling alternative, widely recognized for their ability to generate high-quality images across various domains, which makes them a promising tool for explaining classifier decisions. These models offer the potential for a richer and more diverse set of semantic features compared to existing methods. However, their ability to interpret and utilize latent space semantics remains limited in the context of diffusion models. Existing techniques for identifying meaningful semantics rely largely on supervised approaches [7, 8], which require users to craft detailed text prompts to specify particular features for editing, such as a mustache. This process requires domain expertise to craft appropriate prompts; thus, a broad semantic corpus is essential to improve diffusion models for classifier explanations while reducing manual effort. In this paper, we first employ Vision-Language Models (VLMs) [36] to extract a largescale corpus covering domain-specific hierarchical semantics (see Fig. 2). Then, we introduce a training-free method, DiffEx, which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics. Our method provides explanations for both coarse and fine-grained semantics. For example, it can recognize a 'beard' as a coarse semantic influencing age classification scores and also demonstrate how specific beard types (such as 'Balbo' or 'Anchor' beards) impact the classifier's scores. + +![](images/ae1b7fea0847d7d6c1e91ff692825fc889f7d64066085020bd93e34f40ee6ea5.jpg) +Figure 2. Hierarchical List of Attributes for the Bird Domain. We use VLMs to extract a hierarchical corpus of semantics within a given domain. This representation helps illustrate how different attributes are grouped within a broader domain, facilitating a better understanding of how each semantic contributes to the overall decision-making process of a classifier. + +This hierarchical strategy provides users with an overview of the most significant semantics for a classifier, allowing them to dig deeper into particular fine-grained semantics that are essential for understanding classifier behavior. Our qualitative and quantitative experiments reveal that DiffEx offers considerably richer and more comprehensive explanations for binary and multi-class classifiers across various domains, including facial features, retinal health, and plant pathology. Our contributions are as follows: + +- We propose DiffEx, a training-free approach using VLMs and T2I diffusion models to explain classifier decisions. To the best of our knowledge, this is the first hierarchical approach that explains classifier decisions. +- Our method employs a VLM to develop a comprehensive semantic corpus that spans multiple domains, with our source code publicly available to support future research. +- Unlike GAN-based methods, our approach can address classifiers that focus on single concepts (such as an 'age' classifier analyzing a headshot of a person) and also extend to classifiers for complex scenes (such as a 'modern/traditional architecture' classifier). +- Our method offers more comprehensible explanations for classifiers in applications ranging from facial recognition to retinal health compared to prior approaches and is adaptable to both binary and multiclassifiers. + +# 2. Related Work + +Traditional research focuses mainly on heat maps and patch-based extractions to explain classifier decisions. Specifically, class activation and saliency maps attempt to emulate human visual strategies by focusing on the image regions most relevant to a specific class [43, 49, 53, 56, 72, 75, 78]. While these maps highlight the object or part of an image that most influences the classifier's decision, they fail to reveal which specific, fine-grained object attributes (e.g., color, pattern, or texture) impact the classification. Other approaches try to explain classifier outputs by an- + +![](images/ab4b3263592897b9a8e04f126f817f3c82542289ab72056892a53664655ceda3.jpg) +Figure 3. An Overview of DiffEx. Our pipeline processes a set of sample domain-specific images and a text prompt using a VLM to generate a hierarchical semantic corpus of attributes relevant to a specific domain. Based on this corpus, DiffEx identifies and ranks the most influential features affecting the classifier's decisions, sorting them from most to least impactful (rightmost image). The hierarchical explanation of semantics (like a "beard" and its subtypes) provides a fine-grained understanding of which features drive classifier outputs. + +alyzing extracted image patches [16, 73]; however, these methods can only reveal spatially localized attributes. Previous studies have leveraged generative models such as variational autoencoders (VAEs) [19] or GANs [17, 48, 54, 55]. The most similar research to ours, StylEx [31], introduces a GAN-based approach to identify various attributes that influence a classifier's decisions. However, this method has several limitations. Firstly, it requires training a new GAN for each classifier, which can be resource-intensive and time-consuming. Secondly, each identified attribute requires manual labeling, which can require domain expertise. Lastly, StylEx only uncovers a limited range of semantic attributes, potentially overlooking others that might significantly affect a classifier's scores. Additionally, a notable limitation of the StylEx method and similar GAN models is their focus on single concepts, such as human or animal faces, or individual objects like leaves, rather than entire scenes. In contrast, diffusion models, known for their robust capability to generate complex and detailed entire scenes, offer a significant advantage. Our approach leverages the power of diffusion models to cover a wider spectrum of visual contexts, and enhances the versatility and applicability of our method across various classifiers that assess not only individual elements but also the interaction and composition of entire scenes. + +Recent approaches have begun using diffusion models to generate counterfactual examples. One method utilizes shortcut learning to generate counterfactual images but fails to make semantically meaningful edits for certain attributes [68]. Another study explores modifying the diffusion process via adaptive parametrization and cone regularization to produce realistic counterfactual images; however, this approach depends on a robust model, which can be difficult to train [4]. [26] explored counterfactual image generation, however their approach is computationally demanding and uses DDPM [22] models trained on single domains. As a result, it does not take advantage of large-scale latent diffusion models like Stable Diffusion [46], which can handle more complex scenes. Furthermore, even though some + +studies have leveraged diffusion models to generate realistic counterfactuals in high-stakes domains [24], such as the medical field, they use a less efficient base model for the image generation process and focus predominantly on single-attribute modifications. Furthermore, to the best of our knowledge, there is no existing research that explores a hierarchical explanation of classifiers. Such an approach would systematically unpack the layers of influence that different semantic levels have on a classifier's scoring mechanism. This gap highlights a significant opportunity to enhance understanding by detailing how various semantic categories and their sub-types contribute to the decisions made by classifiers. + +# 3. Methodology + +We first discuss the background on identifying attributes that influence classifier decisions through changes in logits. Next, we introduce our method, which includes curating hierarchical attributes using VLMs and a novel algorithm inspired by beam search to pinpoint attributes that affect classifier scores. Our pipeline is detailed in Fig. 3. + +# 3.1. Background + +StylEx [31] identifies semantics that meaningfully influence classifier decisions by ranking each attribute based on its impact on the classifier's logit outputs. This ranking process aims to identify and select attributes that best explain the classifier's behavior in a given context, such as understanding the factors influencing age classification. + +Given a semantic corpus $S$ and a set of $N$ images, counterfactual images are generated to analyze the influence of semantic attributes on classifier decisions. For each original image $x_{i}$ , an edited version $x^{\prime}_{i} = g(x_{i}, s)$ is generated by applying a semantic attribute $s \in S$ through a transformation function $g$ . This transformation highlights the effect of each semantic attribute on a classifier's output. Moreover, the logit difference for each attribute measures how the classifier's score changes due to the presence of $s$ . + +The influence of each attribute $s$ is quantified as the aver + +![](images/1289057b7271758f9b7ce2a009e1f676c7f6d04d49b9c4c356386448aaef76c3.jpg) +Figure 4. Top-7 Facial Attributes for Age Classification. DiffEx identifies key attributes and their top subtypes for the age classifier. Edited images are arranged hierarchically, with scores for the "young" label shown at the top-left of each image. + +age logit difference between the original and edited images across a set of sample images, defined as: + +$$ +I (s) = \left| f \left(X ^ {\prime}, s, y\right) - \frac {1}{N} \sum_ {i = 1} ^ {N} C \left(x _ {i}, y\right) \right| \tag {1} +$$ + +$$ +f (X, s, y) = \frac {1}{N} \sum_ {i = 1} ^ {N} C \left(g \left(x _ {i}, s\right), y\right) \tag {2} +$$ + +where $f(X,s,y)$ denotes the average classifier's logit score for target class $y$ on image set $X$ edited with attribute $s$ . Here, $y$ represents the unedited images' target labels, such as the "young" label for an age classifier. This influence score $I(s)$ captures the average impact of each semantic attribute on classifier decisions by reflecting the logit score change due to attribute manipulation. + +# 3.2. Our Method + +We first employ Vision-Language Models (VLMs) [36] to extract a large-scale corpus covering domain-specific hierarchical semantics. Then, we introduce a training-free method, $DiffEx$ , which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics. + +# 3.2.1. VLM-Based Semantic Space + +While the StylEx method outlined in Section 3.1 employs a logit-based approach to identify semantics, it requires manually labeling each attribute extracted from the trained GAN model. Moreover, this method does not support the explanation of hierarchical attributes. Therefore, we first compile a large-scale set of semantics using VLMs. Given a domain $d$ , such as the 'facial domain,' our objective is to identify a comprehensive set of domain-specific attributes from a collection of domain-specific images, denoted as $N_{d}$ . To accomplish this, we utilize a Vision-Language Model (VLM) [36] to extract a range of relevant features, represented by + +# Algorithm 1 DiffEx + +Require: Hierarchical structure $\mathcal{H}$ with semantic groups and features, beam width $B$ , classifier or scoring function $f$ , scoring threshold $\delta$ +Ensure: Optimal semantics maximizing $I$ +1: Initialize $S \gets$ root-level groups in $\mathcal{H}$ {Initial candidate set at top-level groups} +2: Initialize beam $\mathcal{B}\gets \emptyset$ +3: Score each candidate $s \in S$ using the scoring function $f(X, s, y)$ for $y$ class label +4: Select top $B$ candidates with $I(b) \geq \delta$ and store in beam $\mathcal{B}$ and $S$ {Apply thresholding to filter relevant candidates} +5: while $\mathcal{B} \neq \emptyset$ do +6: Initialize $S_{\mathrm{next}} \gets \emptyset$ +7: for each candidate $b \in \mathcal{B}$ do +8: Expand $b$ by adding sub-features from its next level in $\mathcal{H}$ to form new candidates +9: for each new combination $b'$ expanded from $b$ to 10: if $I(b') > I(b)$ then +11: Add $b'$ to $S_{\mathrm{next}}$ . +12: end if +13: end for +14: end for +15: Set $\mathcal{B}\gets S_{\mathrm{next}}$ +16: Append $S_{\mathrm{next}}$ to $S$ +17: end while +18: Return highest-scoring combination from final $S$ as the optimal joint semantic combination + +$\mathcal{H}$ . Employing in-context learning [9, 58], we prompt the VLM with a small set of images $N_{d}$ , along with a detailed task description and examples of desired outputs. This process allows us to generate a substantial semantic corpus of keywords, $\mathcal{H}$ , that captures the fine-grained attributes relevant to the domain. The resulting corpus, $\mathcal{H}$ , comprises a comprehensive collection of keywords that covers full spectrum of the domain's attributes. We use this dataset to hierarchically explain how different attributes influence classifier decisions. + +# 3.2.2.DiffEx + +Considering the extensive number of semantics identified using a VLM, it is computationally expensive to evaluate every possible semantic or combination of attributes. We introduce DiffEx, an efficient approach inspired by beam search to explain classifier decisions (see Algorithm 1). This method leverages our hierarchical semantic corpus to streamline the process. Our approach iteratively refines candidate attributes by expanding only the most impactful semantic paths at each hierarchical level, with each path's relevance guided by a scoring function that assesses the clas + +![](images/5ce3a010250f4351b7efccb2c7f1d7f7fb8bdc399c9d602f28c6880c42a67a04.jpg) +Figure 5. Top-7 Discovered Attributes Across Different Animal Domains. Our method successfully identifies key attributes for multiple domains, such as bird, wildcat, and pet species. The original images are depicted with red borders while the edited images are depicted with black borders. For the pet species domain, we used a binary classifier and for the bird and wildcat species domains, we used a multi-classifier. + +sifier's response to each semantic feature. Formally, the scoring function $f$ calculates the average classifier $C$ logit scores across $N$ images generated using each semantic attribute $s$ , as defined in Eq. (2), where $g$ represents the generative diffusion model applied to each sample image. + +We begin with an initial candidate set $S$ , which includes high-level groups from the semantic hierarchy (e.g., broader categories like 'mouth features' or 'eye features'). Each candidate in $S$ is evaluated by the scoring function $f(X, s, y)$ , which quantifies the influence of the candidate on the classifier's output. Only the top $B$ candidates that surpass a predefined score threshold $\delta$ are retained, setting a beam width that limits the search to the most impactful candidates. For each candidate in the beam, the algorithm proceeds by expanding to the next hierarchical level, incorporating more specific sub-features (e.g., for "mouth features," sub-features such as "beard" and "mustache" are included). For each expanded candidate $b'$ , the scoring function in Eq. (2) is re-evaluated. Only candidates with a score exceeding that of their parent $I(b') > I(b)$ are retained in the next candidate set $S_{next}$ , ensuring that only those additions that yield significant incremental impact are added. This process of expansion, scoring, and filtering continues iteratively, moving from general to more specific attributes at each hierarchical level. By dynamically adjusting the candidate set based on beam width $B$ and scoring threshold $\delta$ , our method remains computationally efficient, focusing on high-impact combinations rather than exhaustively evaluating all possible attribute pairings. For generating counter + +![](images/8d94ce4c3d27b7b8a02852f2c2442bc68196beb52106c4d6ec7b4cd825c16615.jpg) +Figure 6. Visual Comparison of Key Attributes Identified by StylEx and DiffEx in the Plant Health and Retinal Disease Domains. This figure illustrates the enhanced capability of our method in identifying a broader set of significant attributes compared to StylEx within the plant health and retinal disease domains. DiffEx uncovers more detailed and diagnostically relevant features, such as "leaf vein color" and "macular hole," which provide deeper insights into leaf and retina health. + +factual images, we utilize an off-the-shelf diffusion-based editing tool, Ledits++ [8]. However, our approach is versatile enough to accommodate any editing method. + +# 4. Experiments + +Experimental Setup. Our experiments utilize Ledits++ [8] and Stable Diffusion XL (SDXL) [40] for generating counterfactual images. Twenty-five time steps are omitted to boost computational efficiency while preserving the quality of edits. The edit threshold, a hyperparameter that dictates the global application scope of edits, is adjusted based on each domain. We test our method across diverse classifiers to evaluate its effectiveness in explaining model behavior through semantic influence. Specifically, we utilize classifiers trained on facial attributes data (i.e. age and gender classifiers) [28] as well as plant health [23], retinal disease [13], bird species [64], wildcat/pet species [11], fashion [35], places [79], and food data [6]. The experiments for the face and plant health domains utilize a CLIP-based classifier [41] to evaluate and interpret edits, while experiments within the bird, wildcat, pet, food, fashion, and places domains use CNN-based classifiers. These CNN classifiers were built on the EfficientNet [61] architecture and achieved an accuracy of over 95 percent on their test sets. For the retinal disease domain, we utilized the FLAIR model [51], which is based on a pre-trained vision-language model, to classify various retina scans. + +# 4.1. Qualitative Experiments + +Explaining Classifiers with Single Attributes. We used various classifiers to analyze fine-grained semantic features for age categorization, species identification, leaf health assessment, retinal disease detection, etc. In the figures illustrating the visual results, the images with red borders indicate the original, unedited versions, while those with black borders are the edited versions, unless noted otherwise. + +
Face (Age)Bird (Species)Leaves (Health)Retina Scans (Disease)Wildcat (Species)Pet (Cat/Dog)
StylExOursStylExOursStylExOursStylExOursStylExOursStylExOurs
Skin PigmentationEyebrowBelly ColorUpperparts ColorBase Leaf ColorBase ColorExudatesGlaucomaSpotsStripesOpen MouthNose Color
Eyebrow ThicknessMakeupUpperparts ColorHead ColorApex ColorVein ColorCotton Wool SpotsSubretinal HemorrhageBlack Tear MarkOpen EyesClosed MouthWhiskers
EyeglassesMustache TypeWing PatternBeak ShapeSpotsApex ColorHemorrhagesIntraretinal HemorrhageEye Shape + SizeNose ColorEye ShapeEye Color
Hair ColorTeethBeak ColorBeak ColorBlightSpots PresenceClustered ExudatesMacular HoleXSpotsDropped EarsClosed Mouth
Lip Thickness + PositionLip VolumeHead ColorWing PatternHalosDisease PresenceXHard ExudatesXFur ColorPointed EarsOpen Mouth
BangsLip ColorBreast ColorEye ColorXBlight SizeXBlackened MaculaXManeEye CircumferenceSpots
Eye MakeupEyelashXThroat ColorXLeaf TextureXSoft ExudatesXEye ColorXCollar
Facial Hair ColorBeard TypeXWing ColorXSpots ColorXRetinal DruenXTongueXPointed Ears
XFacewearXCrest PresenceXDiscolorationXOptic Disc HemorrhageXPupil SizeXMouth Color
XHeadwearXFeather TextureXLeaf OrientationXCataractXWhiskersXFur Pattern
+ +Table 1. Comparison of Top Attributes Across Different Domains and Classifiers. The table above contains a list of the top attributes discovered by DiffEx (Ours) vs. StylEx. The $\pmb{x}$ in the table indicates attributes that were not mentioned in StylEx. It is also important to note that "cotton wool spots" and "soft exudates" refer to the same condition within the retinal disease domain. + +
MethodCrest PresenceBeak ShapeThroat ColorFeather TextureEye ColorBeak ColorHead ColorUpperparts ColorAvg. Correct Response
Grad-CAM36%50%56%35%47%65%59%76%53%
StylEx68%85%79%82%74%68%91%65%76.5%
DiffEx (Ours)88%91%88%91%82%82%97%88%88.4%
+ +Table 2. Comparison with Other Explainability Methods. The table above displays the percentage of correct attribute selections for the bird class, as chosen by users when viewing outputs from different explainability methods. It also includes the average percentage of correct responses across all attributes for each method. As shown, for each attribute presented, the majority of users identified the correct attribute when viewing the output generated by DiffEx. + +Face Domain: In the facial domain, as illustrated in Fig. 4, DiffEx identifies and ranks key attributes impacting age classification. For example, features such as "makeup styles," "lip volume," and "accessories" (e.g. "hairbands") are associated with perceived youthfulness, whereas attributes like "facial hair" and "eyeglasses" are linked with perceived older age. Notably, the eyeglasses attribute consistently reduces the classifier's score for the "young" label, reflecting its association with older demographics. Additionally, DiffEx uncovers hierarchical attribute structures, demonstrating how subcategories within a feature can have varying effects on classifier outcomes. For instance, as shown in Fig. 4, different beard styles impact the perceived age differently (e.g. a "Balbo beard" significantly increases the age classification score more than a "full beard"). Additional examples of hierarchical explanations of age classifiers are detailed in the appendix (S10). + +Animal Domain: Furthermore, DiffEx explains classifiers for a variety of animal types. Fig. 5 highlights the top-7 most influential attributes in the bird, wildcat, and pet domains where our method was able to identify fine-grained semantics to explain classifier behavior. For example, in the wildcat classifier, attributes like stripes and manes are critical in distinguishing wildcats from other species. Explaining these classifiers also reveals potential biases. For instance, the presence of a collar significantly increases the likelihood of an animal being classified as a dog. This bias may stem from the training dataset where images of dogs more frequently featured collars compared to cats. + +Medical and Plant Health Domains: Fig. 6 demonstrates the results for retinal disease and plant health domains. For the retinal disease domain, we use the FLAIR model to classify images of retinal fundus scans. This model utilizes detailed domain expert knowledge descriptions, such as "no hemorrhages, microaneurysms, or exudates" compared to general descriptions like "no diabetic retinopathy" to aid in + +its classification. For the plant health domain, the CLIP classifier we use looks at features such as the presence of spots, fungus, etc. to make its decision. + +Multi-Object Domains involving Complex Scenes: Unlike traditional GAN-based methods that primarily focus on single-object scenarios like a cropped face, DiffEx extends its utility by providing a list of relevant features for domains encompassing multiple objects, such as places, food, and fashion. Classifiers within these domains often evaluate multiple elements simultaneously. For example, a fashion classifier might determine the 'formal or casual fit' of an outfit by considering various components such as the hairstyle, top, and bottom. Similarly, an architectural classifier assessing whether a building appears 'urban or rural' may base its evaluation on not just the structure itself but also its exteriors and surroundings. Explaining the decision-making process of such classifiers is crucial, as they analyze multiple objects within a single image, adding complexity to their interpretative frameworks. Fig. 7 illustrates the top-7 attributes identified by DiffEx for the food, place, and fashion domains. Moreover, we demonstrate DiffEx on various diverse domains. For example, for the food domain, DiffEx was able to discover that removing "caviar" or a "plate decoration" from the image made the food appear to be more perceived as fast food compared to fine dining. For the places domain, DiffEx was able to find that adding a "tractor" to an image or a "dirt road" made the place seem more "rural" than "urban." In the fashion domain, the style of a neckline significantly influences whether an outfit was classified as "formal" or "casual." For instance, a V-neckline was often associated with a more casual look, whereas a classic boat neckline was generally perceived as more formal. + +Explaining Classifiers with Joint Attributes. Our method goes beyond analyzing individual attributes by identifying attribute combinations that collectively improve classifier + +![](images/6955f2fc9dd51787fc2889e50d11b978631d8cab0a2f1bf8c528a8ec670b4eeb.jpg) +Figure 7. Top-7 Attributes Discovered by DiffEx for Multi-Object Domains involving Complex Scenes. Compared to existing methods, DiffEx was used to identify meaningful attributes for multi-object domains, such as "food," "places," and "fashion." By identifying semantic features that make food appear "healthier" or part of "fine dining," as well as attributes that give a place a more "rural" or "modern" feel, DiffEx can serve as a powerful tool for understanding perceptions through targeted edits. + +
RatingBird DomainFace Domain
Edit Quality3.386 ± 0.2233.659 ± 0.248
Disentanglement3.163 ± 0.1973.204 ± 0.213
+ +Table 3. Edit Quality and Disentanglement Ratings. These scores across various domains from User Study 1 are shown above. + +interpretation. While single attributes may have a minimal impact on logits, their combined effect can substantially influence the classifier's output, uncovering subtle interactions that shape decision-making. For example, as shown in Fig. 8, Section C, individual changes in "lip color" or "eye makeup" result in minimal score changes, but when modified together, they make the subject appear significantly younger. This analysis shows how classifiers may respond more robustly to specific attribute combinations. Additional examples are provided in the appendix (S7, S8, S9). + +Extending DiffEx for Multi-Classifier Analysis. We + +adapt DiffEx for multi-classifier applications to uncover semantic attributes essential for tasks like retinal disease classification (Fig. 6) and bird and wildcat species identification (Fig. 5). This approach highlights key features such as "upperparts color" and "beak shape" for bird species, "stripes" and "nose color" for wildcat species, and types of hemorrhages and exudates for retinal conditions. The results in Fig. 6 for the retinal diseases domain and Fig. 5 for the bird/wildcat species domain demonstrate that modifying these attributes significantly alters classifier scores, impacting assessments for species identification and disease likelihood. + +Qualitative Comparisons. In Fig. 6, we present a visual comparison of the top attributes and classification scores identified by DiffEx against those identified by StylEx for the plant health and retinal disease domains. As illustrated, DiffEx successfully uncovers a broader set of semantically meaningful attributes compared to StylEx. For instance, DiffEx identifies detailed attributes such as "leaf vein color" and "spots color," in addition to the more general attributes found by StylEx, like "leaf base color" and "apex color." Table 1 provides a comprehensive comparison of the top features identified by StylEx and our method, highlighting DiffEx's superior ability to uncover semantics that are crucial for the classifier's decision-making process. In particular, the table presents an extended list of features across domains such as faces, bird species, plant health, retina scans, wildcat species, and pet types, all of which influence a classifier's score—further emphasizing DiffEx's ability to uncover fine-grained and contextually rich features. Additionally, in Figure 8 in Sections A and B, we visually compare DiffEx with other diffusion-based, counterfactual generation methods, such as FastDiME [68], DiME [26], and ACE [27] in the face and medical domains. As seen in Figure 8, Section B, DiffEx is able to preserve important features, such as a person's eyes, glasses, etc., in various cases, such as removing a smile or making an older person look younger. Furthermore, Figure 8, Section A shows DiffEx's ability to generate counterfactuals in various medical subdomains, such as skin lesions and chest x-rays. Specifically, for the skin lesion domain, DiffEx was able to remove ruler markings in images from the ISIC dataset [12, 63], while in the chest x-rays domain, DiffEx was able to remove chest drains and pacemakers from images from the NIH [67] dataset and CheXpert [25] dataset. As seen in Figure 8, Section A, DiffEx achieved similar results to FastDiME (Med) [68], a diffusion-based counterfactual model that was trained on medical images. + +# 4.2. Quantitative Experiments + +Baselines. To quantitatively evaluate the effectiveness of DiffEx compared to other explainability methods, such as StylEx [31] and Grad-CAM [49], we conducted a series of + +![](images/c9f65854f319e7a13091cdd841d25ad0355c1a9365ba433faa7321199ec01cf2.jpg) +Figure 8. More Comparisons and Exploring Joint Attributes. (a) We demonstrate DiffEx's ability to produce counterfactuals comparable to FastDiME-2+ in skin lesion and chest X-ray domains. (b) compares DiffEx with ACE, DiME, FastDiME, and FastDiME-2+, showing its capacity to add or remove attributes in a disentangled manner. (c) highlights DiffEx's effectiveness in explaining classifier decisions by analyzing the influence of joint attributes, with perceived age scores (e.g., "younger" label) displayed to show how attribute pairings affect classifier outputs. + +comprehensive user studies to assess how easily participants could identify the relevant features extracted by our method. User Study 1: Visual Quality and Disentanglement. To evaluate the visual quality and disentanglement of edited images for the face and bird domains, we conducted a user study with 50 participants on Prolific.com. For each domain, we showed pairs of unedited and edited images for ten attributes and asked users to assess whether the edited image contained the desired attribute and was disentangled. For each pair of images, we asked users to rate the edit and disentanglement from one to five, with five representing the highest score. Our results (see Table 3) indicate that our edits reflected the intended attributes while minimizing unrelated changes (see S6.1 in the appendix for more details). + +User Study 2: Comparison with Grad-CAM and StylEx. To evaluate how effectively our method, DiffEx, explains various semantics within a specific domain compared to other explainability methods (Grad-CAM and StylEx), we conducted a user study with 35 participants on Prolific. We focused on images from the bird domain, generating three sets of three images per attribute: one original image, one edited image, and one image illustrating the explainability method. For Grad-CAM, participants were shown a heatmap overlay on the edited image. For StylEx, we displayed the edited image alone, as this method requires users to manually label the edits. For DiffEx, participants viewed the edited image and the attribute automatically assigned by the VLM. We then asked users to select the attribute (e.g., "beak color," "crest presence") that best explained the edited image from four choices. Results indicate that DiffEx significantly outperforms Grad-CAM and StylEx in explaining edited attributes in images (see Table 2). Please refer to S6.2 in the appendix for details. + +User Study 3: Disentanglement in Images with Multiple Attributes. We evaluated DiffEx's disentanglement ability when editing multiple attributes simultaneously via a user study. Users rated images on a scale of 1 to 5 (5 = very disentangled, 1 = entangled). DiffEx achieved an average score + +of 4.6 for single-attribute edits and 4.3 for edits involving four or more attributes, demonstrating effective handling of multiple edits (see S6.3 in the appendix for more details). + +Ablation Studies. To assess the robustness of our method, we tested various Vision-Language Models (VLMs) and Diffusion Models. For VLMs, we probed Llama 3.2 [20], Qwen2-VL [66], and Claude 3.5 Haiku [1] within the face domain, measuring semantic consistency via IoU scores against GPT-4. The results were 0.94, 0.87, and 0.93, respectively, which indicate substantial semantic overlap. For diffusion models, we compared top attributes from Stable Diffusion (SD) 1.5 and SD 2.0 [46] against SDXL [40] using Normalized Discounted Cumulative Gain (nDCG) to measure the attributes' overlap, achieving scores of 0.75 and 0.76, respectively. These results confirm the stability of our method across different models. + +# 5. Discussion + +Limitations. While our approach provides valuable insights into classifier behavior through semantic edits, it has limitations. Relying on VLM-curated semantics restricts us to the quality and scope of the initial corpus, which may miss relevant features. Additionally, using off-the-shelf editing models like Ledits++ can result in entangled edits (a common issue in image editing algorithms [71]) or fail to capture fine-grained details like earlobes or nose rings (see Appendix S1). Nonetheless, our framework is flexible and can be enhanced with domain-specific adjustments, such as task-specific RAGs [70] or editing methods [7, 14]. RAGs can automatically retrieve semantics in specialized domains (e.g. medical) to improve generalization. + +Conclusion. In this work, we introduce DiffEx, a novel approach for explaining classifier decisions by utilizing semantic edits within diffusion models. By harnessing the power of vision language models, we curate a comprehensive, hierarchical semantic corpus across various domains and propose a novel algorithm inspired by beam search to filter and rank the most impactful features. DiffEx ranks these semantic features based on their influence on classifier logits, capturing both individual and joint attribute effects, which are crucial for understanding complex classifier behaviors. Through experiments conducted on a wide range of domains—including face, bird, and medical classifiers—we showcase the robustness and adaptability of our approach. Our work provides a powerful tool for interpreting model decisions across diverse applications and promoting transparency in AI-driven classification systems. + +# References + +[1] The claude 3 model family: Opus, sonnet, haiku. 8 +[2] Mostofa Ahsan, Rahul Gomes, Md. Minhaz Chowdhury, and Kendall E. Nygard. Enhancing machine learning prediction in cybersecurity using dynamic feature selector. Journal of Cybersecurity and Privacy, 1(1):199-218, 2021. 1 +[3] Leila Amgoud. Explaining black-box classifiers: Properties and functions. International Journal of Approximate Reasoning, 155:40-65, 2023. 1 +[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. Advances in Neural Information Processing Systems, 35:364-377, 2022. 3 +[5] Ouahiba Azouaoui and Amine Chohra. Soft computing based pattern classifiers for the obstacle avoidance behavior of intelligent autonomous vehicles (iav). Applied intelligence, 16:249-272, 2002. 1 +[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5 +[7] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. SEGA: Instructing text-to-image models using semantic guidance. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 8 +[8] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2, 5 +[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4 +[10] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2 +[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. CoRR, abs/1912.01865, 2019. 5 +[12] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 7 +[13] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. 5 + +[14] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. arXiv preprint arXiv:2312.05390, 2023. 8 +[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[16] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 3 +[17] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5744-5753, 2019. 3 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[19] Yash Goyal, Amir Feder, Uri Shalit, and Been Kim. Explaining classifiers with causal concept effect (cace). arXiv preprint arXiv:1907.07165, 2019. 3 +[20] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. (arXiv:1512.03385), 2015. arXiv:1512.03385. 1 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[23] David P. Hughes and Marcel Salathe. An open access repository of images on plant health to enable the development of mobile disease diagnostics. (arXiv:1511.08060), 2016. arXiv:1511.08060 [cs]. 5 +[24] Indu Ilanchezian, Valentyn Boreiko, Laura Kuhlewein, Zwei Huang, Murat Seckin Ayhan, Matthias Hein, Lisa Koch, and Philipp Berens. Generating realistic counterfactuals for retinal fundus and oct images using diffusion models. arXiv preprint arXiv:2311.11629, 2023. 3 +[25] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7 +[26] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. In Proceedings of the Asian Conference on Computer Vision, pages 858-876, 2022. 3, 7 +[27] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Adversarial counterfactual visual explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16425-16435, 2023. 7 + +[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 4396-4405, Long Beach, CA, USA, 2019. IEEE. 5 +[29] Ziv Katzir and Yuval Elovici. Quantifying the resilience of machine learning classifiers used for cyber security. Expert Systems with Applications, 92:419-429, 2018. 1 +[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1 +[31] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2, 3, 7 +[32] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1 +[33] Joffrey L. Leevy, John Hancock, Richard Zuech, and Taghi M. Khoshgoftaar. Detecting cybersecurity attacks across different network features and learners. Journal of Big Data, 8(1):38, 2021. 1 +[34] Xuhong Li, Haoyi Xiong, Xingjian Li, Xuanyu Wu, Xiao Zhang, Ji Liu, Jiang Bian, and Dejing Dou. Interpretable deep learning: Interpretation, interpretability, trustworthiness, and beyond. Knowledge and Information Systems, 64 (12):3197-3234, 2022. 2 +[35] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 5 +[36] OpenAI. Gpt-4 technical report, 2024. 2, 4 +[37] Akin Ozcift and Arif Gulten. Classifier ensemble construction with rotation forest to improve medical diagnosis performance of machine learning algorithms. Computer Methods and Programs in Biomedicine, 104(3):443-451, 2011. 1 +[38] Yi Peng, Guoxun Wang, Gang Kou, and Yong Shi. An empirical study of classification algorithm evaluation for financial risk prediction. Applied Soft Computing, 11(2):2906-2915, 2011. 1 +[39] Emmanuel Pintelas, Meletis Liaskos, Ioannis E Livieris, Sotiris Kotsiantis, and Panagiotis Pintelas. Explainable machine learning framework for image classification problems: case study on glioma cancer prediction. Journal of imaging, 6(6):37, 2020. 1 +[40] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 8 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[42] R. Joshua Samuel Raj, S. Jeya Shobana, Irina Valeryevna Pustokhina, Denis Alexandrovich Pustokhin, Deepak Gupta, and K. Shankar. Optimal feature selection-based medical image classification using deep learning model in internet of medical things. IEEE Access, 8:58006-58017, 2020. 1 +[43] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8848, 2020. 2 +[44] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1135-1144, New York, NY, USA, 2016. Association for Computing Machinery. 1 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 8 +[47] Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. Interpretable machine learning: Fundamental principles and 10 grand challenges. Statistic Surveys, 16:1-85, 2022. 2 +[48] Axel Sauer and Andreas Geiger. Counterfactual generative networks. (arXiv:2101.06046), 2021. arXiv:2101.06046 [cs]. 3 +[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 7 +[50] Andy Shih, Arthur Choi, and Adnan Darwiche. A symbolic approach to explaining bayesian network classifiers. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, page 5103-5111. AAAI Press, 2018. 1 +[51] Julio Silva-Rodriguez, Hadi Chakor, Riadh Kobbi, Jose Dolz, and Ismail Ben Ayed. A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision. Medical Image Analysis, 99:103357, 2025. 5 +[52] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 2014. 1 +[53] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2 + +[54] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. (arXiv:1911.00483), 2020. arXiv:1911.00483 [cs]. 3 +[55] Sumedha Singla, Motahhare Eslami, Brian Pollack, Stephen Wallace, and Kayhan Batmanghelich. Explaining the black-box smoothly- a counterfactual approach. (arXiv:2101.04230), 2022. arXiv:2101.04230 [cs, eess]. 3 +[56] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. Advances in neural information processing systems, 32, 2019. 2 +[57] Gregor Stiglic, Primoz Kocbek, Nino Fijacko, Marinka Zitnik, Katrien Verbert, and Leona Cilar. Interpretability of machine learning-based prediction models in healthcare. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 10(5):e1379, 2020. 2 +[58] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, et al. Selective annotation makes language models better few-shot learners. arXiv preprint arXiv:2209.01975, 2022. 4 +[59] Prabu Subramani, Khalid Sattar, Rocio De Prado, Balasubramanian Girirajan, and Marcin Wozniak. Multi-classifier feature fusion-based road detection for connected autonomous vehicles. Applied Sciences, 11(17):7984, 2021. 1 +[60] Jie Sun and Hui Li. Financial distress prediction based on serial combination of multiple classifiers. Expert Systems with Applications, 36(4):8659-8666, 2009. 1 +[61] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2020. 5 +[62] Manoj Thakur and Deepak Kumar. A hybrid financial trading support system using multi-category classifiers and random forest. Applied Soft Computing, 67:337-349, 2018. 1 +[63] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 7 +[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[65] Wahyono, Laksono Kurnianggoro, Joko Hariyono, and Kang-Hyun Jo. Traffic sign recognition system for autonomous vehicle using cascadesvm classifier. In IECON 2014 - 40th Annual Conference of the IEEE Industrial Electronics Society, pages 4081-4086, 2014. 1 +[66] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8 +[67] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 7 +[68] Nina Weng, Paraskevas Pegios, Eike Petersen, Aasa Feragen, and Siavash Bigdeli. Fast diffusion-based counterfactuals for + +shortcut removal and generation. In European Conference on Computer Vision, pages 338-357. Springer, 2025. 3, 7 +[69] M. Wiggins, A. Saad, B. Litt, and G. Vachtsevanos. Evolving a bayesian classifier for ecg-based age classification in medical applications. Applied Soft Computing, 8(1):599-608, 2008. 1 +[70] Junde Wu, Jiayuan Zhu, and Yunli Qi. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187, 2024. 8 +[71] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 8 +[72] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. (arXiv:2004.03383), 2020. arXiv:2004.03383 [cs]. 2 +[73] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. Advances in neural information processing systems, 33: 20554-20565, 2020. 3 +[74] Muhammad Bilal Zafar, Michele Donini, Dylan Slack, Cédric Archambeau, Sanjiv Das, and Krishnamaram Kenthapadi. On the lack of robust interpretability of neural text classifiers. arXiv preprint arXiv:2106.04631, 2021. 1 +[75] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2 +[76] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1 +[77] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2 +[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2 +[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5 \ No newline at end of file diff --git a/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/images.zip b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..909286c4cce1d45e30891d6d335dba1104823305 --- /dev/null +++ b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a9a0581cc42f2a0c2a12cd3bb91a65c6de7640581b27b9f7df2bd2dadfe587b +size 704836 diff --git a/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/layout.json b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63487bc4131ed5712bcc7d5f67f8bcb0b8b9d678 --- /dev/null +++ b/2025/Explaining in Diffusion_ Explaining a Classifier with Diffusion Semantics/layout.json @@ -0,0 +1,8602 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "text", + "content": "Explaining in Diffusion: Explaining a Classifier with Diffusion Semantics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 143, + 228, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 143, + 228, + 171 + ], + "spans": [ + { + "bbox": [ + 154, + 143, + 228, + 171 + ], + "type": "text", + "content": "Tahira Kazimi† Virginia Tech" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 175, + 243, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 175, + 243, + 184 + ], + "spans": [ + { + "bbox": [ + 139, + 175, + 243, + 184 + ], + "type": "text", + "content": "tahirakazimi@vt.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 285, + 143, + 354, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 143, + 354, + 171 + ], + "spans": [ + { + "bbox": [ + 285, + 143, + 354, + 171 + ], + "type": "text", + "content": "Ritika Allada† Virginia Tech" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 279, + 175, + 361, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 175, + 361, + 184 + ], + "spans": [ + { + "bbox": [ + 279, + 175, + 361, + 184 + ], + "type": "text", + "content": "ritika88@vt.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 396, + 144, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 144, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 396, + 144, + 471, + 171 + ], + "type": "text", + "content": "Pinar Yanardag Virginia Tech" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 397, + 175, + 469, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 175, + 469, + 185 + ], + "spans": [ + { + "bbox": [ + 397, + 175, + 469, + 185 + ], + "type": "text", + "content": " pineary@vt.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 194, + 406, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 194, + 406, + 205 + ], + "spans": [ + { + "bbox": [ + 200, + 194, + 406, + 205 + ], + "type": "text", + "content": "https://explain-in-diffusion.github.io" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 57, + 224, + 177, + 376 + ], + "blocks": [ + { + "bbox": [ + 57, + 224, + 177, + 376 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 177, + 376 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 177, + 376 + ], + "type": "image", + "image_path": "8af0c4f0105c182e2920e6af801a517e6576fe80fed90eb68f759d23e71f9468.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 554, + 443 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 554, + 443 + ], + "type": "text", + "content": "Figure 1. DiffEx explains the decisions of domain-specific classifiers by identifying the most influential semantics affecting their predictions. Classifier scores for each example are displayed in the top-left corner, demonstrating how classifier predictions change in response to the manipulation of different semantics (original images are shown with red borders). Our approach is capable of explaining classifiers that concentrate on individual concepts such as faces or animals (top row) as well as those that manage complex scenes involving multiple objects, such as a formal/casual fit in a fashion context (bottom row)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 181, + 224, + 298, + 375 + ], + "blocks": [ + { + "bbox": [ + 181, + 224, + 298, + 375 + ], + "lines": [ + { + "bbox": [ + 181, + 224, + 298, + 375 + ], + "spans": [ + { + "bbox": [ + 181, + 224, + 298, + 375 + ], + "type": "image", + "image_path": "9775f157ca494465e96aadd14f7d0626789655281c8a0d1f16f98687cb4cbac7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 302, + 224, + 425, + 375 + ], + "blocks": [ + { + "bbox": [ + 302, + 224, + 425, + 375 + ], + "lines": [ + { + "bbox": [ + 302, + 224, + 425, + 375 + ], + "spans": [ + { + "bbox": [ + 302, + 224, + 425, + 375 + ], + "type": "image", + "image_path": "7dbc047a4efdfabff0916122a8a1ba42b9fd46a217f1ad98b2c8003689fe873b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 428, + 224, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 428, + 224, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 428, + 224, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 428, + 224, + 547, + 376 + ], + "type": "image", + "image_path": "32ca24e47a20d40bf513c8582bf56588f3e9dbd49d33e16733add3bb12bdff51.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 453, + 200, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 453, + 200, + 466 + ], + "spans": [ + { + "bbox": [ + 151, + 453, + 200, + 466 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 479, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 297, + 694 + ], + "type": "text", + "content": "Classifiers are important components in many computer vision tasks, serving as the foundational backbone of a wide variety of models employed across diverse applications. However, understanding the decision-making process of classifiers remains a significant challenge. We propose DiffEx, a novel method that leverages the capabilities of text-to-image diffusion models to explain classifier decisions. Unlike traditional GAN-based explainability models, which are limited to simple, single-concept analyses and typically require training a new model for each classifier, our approach can explain classifiers that focus on single concepts (such as faces or animals) as well as those that handle complex scenes involving multiple concepts. DiffEx employs vision-language models to create a hierarchical list of semantics, allowing users to identify not only the overarching semantic influences on classifiers (e.g., the 'beard' semantic in a facial classifier) but also their sub-types, such as 'goatee' or 'Balbo' beard. Our experiments demonstrate" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 454, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 502 + ], + "type": "text", + "content": "that DiffEx is able to cover a significantly broader spectrum of semantics compared to its GAN counterparts, providing a hierarchical tool that delivers a more detailed and fine-grained understanding of classifier decisions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 513, + 394, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 513, + 394, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 513, + 394, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "type": "text", + "content": "Classifiers are fundamental to computer vision tasks, forming the backbone of many models used in a broad spectrum of applications [15, 21, 30, 32, 52]. Their ability to generalize across tasks and adapt to new domains with minimal retraining makes them highly transferable, and thus they are employed extensively in fields such as healthcare [37, 39, 42, 69], finance [38, 60, 62], security [2, 29, 33], and autonomous systems [5, 59, 65]. Despite their versatility and widespread utility, understanding the decision-making process of classifiers remains a significant challenge [3, 44, 50, 74, 76]. This challenge comes mainly from their \"black box\" nature. As images move through the deep layers of the network, the features used for classification become more abstract and harder to understand. The lack of interpretable features in such mod-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 137, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 702, + 137, + 712 + ], + "type": "text", + "content": " Equal contributions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14799" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 360 + ], + "type": "text", + "content": "els raises critical concerns, particularly in high-stakes environments such as medical diagnosis, where understanding the reasoning behind a model's prediction is crucial for ensuring trust, accountability, and informed decision-making [10, 34, 47, 57, 77]. Explaining classifier decisions is crucial for enhancing the transparency and reliability of these models. Prior research [31] has used generative adversarial networks (GANs) [18] to interpret classifier decisions by generating counterfactual examples that manipulate GAN latent semantics. These manipulations illustrate how changes in specific attributes, like the addition of the eyeglasses semantic, impact classifier outputs. However, GANs are often limited to single domains, such as facial images, and typically require training a new model per classifier, which is resource-intensive and time-consuming. Moreover, in GAN-based methods, understanding which latent semantics affect classifier decisions often requires manual intervention to identify and interpret relevant features, such as recognizing that a discovered semantic controls the eyeglasses attribute. This manual process is not only time-consuming but also less feasible in specialized fields like medicine, where identifying intricate attributes requires substantial domain expertise, making the approach impractical in critical scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 367, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 295, + 713 + ], + "type": "text", + "content": "This limitation highlights the need for more automated and versatile approaches to interpret classifier decisions. Text-to-Image (T2I) diffusion models [45] emerge as a compelling alternative, widely recognized for their ability to generate high-quality images across various domains, which makes them a promising tool for explaining classifier decisions. These models offer the potential for a richer and more diverse set of semantic features compared to existing methods. However, their ability to interpret and utilize latent space semantics remains limited in the context of diffusion models. Existing techniques for identifying meaningful semantics rely largely on supervised approaches [7, 8], which require users to craft detailed text prompts to specify particular features for editing, such as a mustache. This process requires domain expertise to craft appropriate prompts; thus, a broad semantic corpus is essential to improve diffusion models for classifier explanations while reducing manual effort. In this paper, we first employ Vision-Language Models (VLMs) [36] to extract a largescale corpus covering domain-specific hierarchical semantics (see Fig. 2). Then, we introduce a training-free method, DiffEx, which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics. Our method provides explanations for both coarse and fine-grained semantics. For example, it can recognize a 'beard' as a coarse semantic influencing age classification scores and also demonstrate how specific beard types (such as 'Balbo' or 'Anchor' beards) impact the classifier's scores." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 334, + 72, + 535, + 173 + ], + "blocks": [ + { + "bbox": [ + 334, + 72, + 535, + 173 + ], + "lines": [ + { + "bbox": [ + 334, + 72, + 535, + 173 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 535, + 173 + ], + "type": "image", + "image_path": "ae1b7fea0847d7d6c1e91ff692825fc889f7d64066085020bd93e34f40ee6ea5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 248 + ], + "type": "text", + "content": "Figure 2. Hierarchical List of Attributes for the Bird Domain. We use VLMs to extract a hierarchical corpus of semantics within a given domain. This representation helps illustrate how different attributes are grouped within a broader domain, facilitating a better understanding of how each semantic contributes to the overall decision-making process of a classifier." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": "This hierarchical strategy provides users with an overview of the most significant semantics for a classifier, allowing them to dig deeper into particular fine-grained semantics that are essential for understanding classifier behavior. Our qualitative and quantitative experiments reveal that DiffEx offers considerably richer and more comprehensive explanations for binary and multi-class classifiers across various domains, including facial features, retinal health, and plant pathology. Our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 368, + 555, + 559 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 368, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 368, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 368, + 555, + 415 + ], + "type": "text", + "content": "- We propose DiffEx, a training-free approach using VLMs and T2I diffusion models to explain classifier decisions. To the best of our knowledge, this is the first hierarchical approach that explains classifier decisions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 416, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 416, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 314, + 416, + 554, + 450 + ], + "type": "text", + "content": "- Our method employs a VLM to develop a comprehensive semantic corpus that spans multiple domains, with our source code publicly available to support future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 451, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 451, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 451, + 553, + 510 + ], + "type": "text", + "content": "- Unlike GAN-based methods, our approach can address classifiers that focus on single concepts (such as an 'age' classifier analyzing a headshot of a person) and also extend to classifiers for complex scenes (such as a 'modern/traditional architecture' classifier)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 511, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 553, + 559 + ], + "type": "text", + "content": "- Our method offers more comprehensible explanations for classifiers in applications ranging from facial recognition to retinal health compared to prior approaches and is adaptable to both binary and multiclassifiers." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 573, + 400, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 573, + 400, + 585 + ], + "spans": [ + { + "bbox": [ + 314, + 573, + 400, + 585 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Traditional research focuses mainly on heat maps and patch-based extractions to explain classifier decisions. Specifically, class activation and saliency maps attempt to emulate human visual strategies by focusing on the image regions most relevant to a specific class [43, 49, 53, 56, 72, 75, 78]. While these maps highlight the object or part of an image that most influences the classifier's decision, they fail to reveal which specific, fine-grained object attributes (e.g., color, pattern, or texture) impact the classification. Other approaches try to explain classifier outputs by an-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14800" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 182 + ], + "type": "image", + "image_path": "ab4b3263592897b9a8e04f126f817f3c82542289ab72056892a53664655ceda3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 237 + ], + "type": "text", + "content": "Figure 3. An Overview of DiffEx. Our pipeline processes a set of sample domain-specific images and a text prompt using a VLM to generate a hierarchical semantic corpus of attributes relevant to a specific domain. Based on this corpus, DiffEx identifies and ranks the most influential features affecting the classifier's decisions, sorting them from most to least impactful (rightmost image). The hierarchical explanation of semantics (like a \"beard\" and its subtypes) provides a fine-grained understanding of which features drive classifier outputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 247, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 296, + 544 + ], + "type": "text", + "content": "alyzing extracted image patches [16, 73]; however, these methods can only reveal spatially localized attributes. Previous studies have leveraged generative models such as variational autoencoders (VAEs) [19] or GANs [17, 48, 54, 55]. The most similar research to ours, StylEx [31], introduces a GAN-based approach to identify various attributes that influence a classifier's decisions. However, this method has several limitations. Firstly, it requires training a new GAN for each classifier, which can be resource-intensive and time-consuming. Secondly, each identified attribute requires manual labeling, which can require domain expertise. Lastly, StylEx only uncovers a limited range of semantic attributes, potentially overlooking others that might significantly affect a classifier's scores. Additionally, a notable limitation of the StylEx method and similar GAN models is their focus on single concepts, such as human or animal faces, or individual objects like leaves, rather than entire scenes. In contrast, diffusion models, known for their robust capability to generate complex and detailed entire scenes, offer a significant advantage. Our approach leverages the power of diffusion models to cover a wider spectrum of visual contexts, and enhances the versatility and applicability of our method across various classifiers that assess not only individual elements but also the interaction and composition of entire scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 712 + ], + "type": "text", + "content": "Recent approaches have begun using diffusion models to generate counterfactual examples. One method utilizes shortcut learning to generate counterfactual images but fails to make semantically meaningful edits for certain attributes [68]. Another study explores modifying the diffusion process via adaptive parametrization and cone regularization to produce realistic counterfactual images; however, this approach depends on a robust model, which can be difficult to train [4]. [26] explored counterfactual image generation, however their approach is computationally demanding and uses DDPM [22] models trained on single domains. As a result, it does not take advantage of large-scale latent diffusion models like Stable Diffusion [46], which can handle more complex scenes. Furthermore, even though some" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 247, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 402 + ], + "type": "text", + "content": "studies have leveraged diffusion models to generate realistic counterfactuals in high-stakes domains [24], such as the medical field, they use a less efficient base model for the image generation process and focus predominantly on single-attribute modifications. Furthermore, to the best of our knowledge, there is no existing research that explores a hierarchical explanation of classifiers. Such an approach would systematically unpack the layers of influence that different semantic levels have on a classifier's scoring mechanism. This gap highlights a significant opportunity to enhance understanding by detailing how various semantic categories and their sub-types contribute to the decisions made by classifiers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 407, + 397, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 397, + 420 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 397, + 420 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 426, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 499 + ], + "type": "text", + "content": "We first discuss the background on identifying attributes that influence classifier decisions through changes in logits. Next, we introduce our method, which includes curating hierarchical attributes using VLMs and a novel algorithm inspired by beam search to pinpoint attributes that affect classifier scores. Our pipeline is detailed in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 505, + 394, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 505, + 394, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 505, + 394, + 517 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 593 + ], + "type": "text", + "content": "StylEx [31] identifies semantics that meaningfully influence classifier decisions by ranking each attribute based on its impact on the classifier's logit outputs. This ranking process aims to identify and select attributes that best explain the classifier's behavior in a given context, such as understanding the factors influencing age classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": "Given a semantic corpus " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": " images, counterfactual images are generated to analyze the influence of semantic attributes on classifier decisions. For each original image " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ", an edited version " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "x^{\\prime}_{i} = g(x_{i}, s)" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": " is generated by applying a semantic attribute " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": " through a transformation function " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ". This transformation highlights the effect of each semantic attribute on a classifier's output. Moreover, the logit difference for each attribute measures how the classifier's score changes due to the presence of " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "content": "The influence of each attribute " + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "content": " is quantified as the aver" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14801" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 291, + 191 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 291, + 191 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 291, + 191 + ], + "type": "image", + "image_path": "1289057b7271758f9b7ce2a009e1f676c7f6d04d49b9c4c356386448aaef76c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 203, + 295, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 295, + 248 + ], + "type": "text", + "content": "Figure 4. Top-7 Facial Attributes for Age Classification. DiffEx identifies key attributes and their top subtypes for the age classifier. Edited images are arranged hierarchically, with scores for the \"young\" label shown at the top-left of each image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "type": "text", + "content": "age logit difference between the original and edited images across a set of sample images, defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 285, + 295, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 285, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 95, + 285, + 295, + 317 + ], + "type": "interline_equation", + "content": "I (s) = \\left| f \\left(X ^ {\\prime}, s, y\\right) - \\frac {1}{N} \\sum_ {i = 1} ^ {N} C \\left(x _ {i}, y\\right) \\right| \\tag {1}", + "image_path": "7cc1c4f8baf56a9b60adf3902d9025112266f248ff3b630b14cb01f98df7475d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 328, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 328, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 328, + 295, + 360 + ], + "type": "interline_equation", + "content": "f (X, s, y) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} C \\left(g \\left(x _ {i}, s\\right), y\\right) \\tag {2}", + "image_path": "c0eacc9d6faa8fe0ee02f8807638e965e6b85a3a459bcc600668cbe6d5a74f48.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "inline_equation", + "content": "f(X,s,y)" + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "content": " denotes the average classifier's logit score for target class " + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "content": " on image set " + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "content": " edited with attribute " + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "content": " represents the unedited images' target labels, such as the \"young\" label for an age classifier. This influence score " + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "inline_equation", + "content": "I(s)" + }, + { + "bbox": [ + 55, + 366, + 296, + 450 + ], + "type": "text", + "content": " captures the average impact of each semantic attribute on classifier decisions by reflecting the logit score change due to attribute manipulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 458, + 138, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 138, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 138, + 470 + ], + "type": "text", + "content": "3.2. Our Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 559 + ], + "type": "text", + "content": "We first employ Vision-Language Models (VLMs) [36] to extract a large-scale corpus covering domain-specific hierarchical semantics. Then, we introduce a training-free method, " + }, + { + "bbox": [ + 55, + 475, + 296, + 559 + ], + "type": "inline_equation", + "content": "DiffEx" + }, + { + "bbox": [ + 55, + 475, + 296, + 559 + ], + "type": "text", + "content": ", which leverages this hierarchical corpus and text-to-image diffusion models to explain the decision-making process of classifiers by identifying the most influential semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 206, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 206, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 206, + 578 + ], + "type": "text", + "content": "3.2.1. VLM-Based Semantic Space" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "While the StylEx method outlined in Section 3.1 employs a logit-based approach to identify semantics, it requires manually labeling each attribute extracted from the trained GAN model. Moreover, this method does not support the explanation of hierarchical attributes. Therefore, we first compile a large-scale set of semantics using VLMs. Given a domain " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", such as the 'facial domain,' our objective is to identify a comprehensive set of domain-specific attributes from a collection of domain-specific images, denoted as " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". To accomplish this, we utilize a Vision-Language Model (VLM) [36] to extract a range of relevant features, represented by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 72, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 399, + 83 + ], + "type": "text", + "content": "Algorithm 1 DiffEx" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 87, + 555, + 422 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": "Require: Hierarchical structure " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": " with semantic groups and features, beam width " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", classifier or scoring function " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", scoring threshold " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "\\delta" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 124, + 489, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 124, + 489, + 135 + ], + "spans": [ + { + "bbox": [ + 315, + 124, + 489, + 135 + ], + "type": "text", + "content": "Ensure: Optimal semantics maximizing " + }, + { + "bbox": [ + 315, + 124, + 489, + 135 + ], + "type": "inline_equation", + "content": "I" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 136, + 553, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 136, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 320, + 136, + 553, + 159 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 320, + 136, + 553, + 159 + ], + "type": "inline_equation", + "content": "S \\gets" + }, + { + "bbox": [ + 320, + 136, + 553, + 159 + ], + "type": "text", + "content": " root-level groups in " + }, + { + "bbox": [ + 320, + 136, + 553, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 320, + 136, + 553, + 159 + ], + "type": "text", + "content": " {Initial candidate set at top-level groups}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 160, + 424, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 160, + 424, + 170 + ], + "spans": [ + { + "bbox": [ + 321, + 160, + 424, + 170 + ], + "type": "text", + "content": "2: Initialize beam " + }, + { + "bbox": [ + 321, + 160, + 424, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{B}\\gets \\emptyset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "text", + "content": "3: Score each candidate " + }, + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "text", + "content": " using the scoring function " + }, + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "inline_equation", + "content": "f(X, s, y)" + }, + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 321, + 171, + 553, + 195 + ], + "type": "text", + "content": " class label" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "text", + "content": "4: Select top " + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "text", + "content": " candidates with " + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "inline_equation", + "content": "I(b) \\geq \\delta" + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "text", + "content": " and store in beam " + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 321, + 196, + 553, + 230 + ], + "type": "text", + "content": " {Apply thresholding to filter relevant candidates}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 232, + 398, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 232, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 398, + 242 + ], + "type": "text", + "content": "5: while " + }, + { + "bbox": [ + 321, + 232, + 398, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\neq \\emptyset" + }, + { + "bbox": [ + 321, + 232, + 398, + 242 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 243, + 421, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 243, + 421, + 254 + ], + "spans": [ + { + "bbox": [ + 321, + 243, + 421, + 254 + ], + "type": "text", + "content": "6: Initialize " + }, + { + "bbox": [ + 321, + 243, + 421, + 254 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{next}} \\gets \\emptyset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 255, + 457, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 255, + 457, + 265 + ], + "spans": [ + { + "bbox": [ + 321, + 255, + 457, + 265 + ], + "type": "text", + "content": "7: for each candidate " + }, + { + "bbox": [ + 321, + 255, + 457, + 265 + ], + "type": "inline_equation", + "content": "b \\in \\mathcal{B}" + }, + { + "bbox": [ + 321, + 255, + 457, + 265 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 267, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 267, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 321, + 267, + 553, + 289 + ], + "type": "text", + "content": "8: Expand " + }, + { + "bbox": [ + 321, + 267, + 553, + 289 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 321, + 267, + 553, + 289 + ], + "type": "text", + "content": " by adding sub-features from its next level in " + }, + { + "bbox": [ + 321, + 267, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 321, + 267, + 553, + 289 + ], + "type": "text", + "content": " to form new candidates" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "spans": [ + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "text", + "content": "9: for each new combination " + }, + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "inline_equation", + "content": "b'" + }, + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "text", + "content": " expanded from " + }, + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "text", + "content": " to 10: if " + }, + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "inline_equation", + "content": "I(b') > I(b)" + }, + { + "bbox": [ + 321, + 292, + 550, + 313 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 319, + 316, + 431, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 316, + 431, + 326 + ], + "spans": [ + { + "bbox": [ + 319, + 316, + 431, + 326 + ], + "type": "text", + "content": "11: Add " + }, + { + "bbox": [ + 319, + 316, + 431, + 326 + ], + "type": "inline_equation", + "content": "b'" + }, + { + "bbox": [ + 319, + 316, + 431, + 326 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 319, + 316, + 431, + 326 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{next}}" + }, + { + "bbox": [ + 319, + 316, + 431, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 327, + 389, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 327, + 389, + 337 + ], + "spans": [ + { + "bbox": [ + 318, + 327, + 389, + 337 + ], + "type": "text", + "content": "12: end if" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 339, + 384, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 339, + 384, + 349 + ], + "spans": [ + { + "bbox": [ + 318, + 339, + 384, + 349 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 351, + 375, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 351, + 375, + 361 + ], + "spans": [ + { + "bbox": [ + 318, + 351, + 375, + 361 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 363, + 400, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 363, + 400, + 373 + ], + "spans": [ + { + "bbox": [ + 318, + 363, + 400, + 373 + ], + "type": "text", + "content": "15: Set " + }, + { + "bbox": [ + 318, + 363, + 400, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{B}\\gets S_{\\mathrm{next}}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 375, + 416, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 375, + 416, + 386 + ], + "spans": [ + { + "bbox": [ + 318, + 375, + 416, + 386 + ], + "type": "text", + "content": "16: Append " + }, + { + "bbox": [ + 318, + 375, + 416, + 386 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{next}}" + }, + { + "bbox": [ + 318, + 375, + 416, + 386 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 318, + 375, + 416, + 386 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 387, + 375, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 387, + 375, + 396 + ], + "spans": [ + { + "bbox": [ + 318, + 387, + 375, + 396 + ], + "type": "text", + "content": "17: end while" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 399, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 399, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 318, + 399, + 553, + 422 + ], + "type": "text", + "content": "18: Return highest-scoring combination from final " + }, + { + "bbox": [ + 318, + 399, + 553, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 318, + 399, + 553, + 422 + ], + "type": "text", + "content": " as the optimal joint semantic combination" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "text", + "content": ". Employing in-context learning [9, 58], we prompt the VLM with a small set of images " + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "text", + "content": ", along with a detailed task description and examples of desired outputs. This process allows us to generate a substantial semantic corpus of keywords, " + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "text", + "content": ", that captures the fine-grained attributes relevant to the domain. The resulting corpus, " + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "text", + "content": ", comprises a comprehensive collection of keywords that covers full spectrum of the domain's attributes. We use this dataset to hierarchically explain how different attributes influence classifier decisions." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 578, + 371, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 578, + 371, + 588 + ], + "spans": [ + { + "bbox": [ + 314, + 578, + 371, + 588 + ], + "type": "text", + "content": "3.2.2.DiffEx" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Considering the extensive number of semantics identified using a VLM, it is computationally expensive to evaluate every possible semantic or combination of attributes. We introduce DiffEx, an efficient approach inspired by beam search to explain classifier decisions (see Algorithm 1). This method leverages our hierarchical semantic corpus to streamline the process. Our approach iteratively refines candidate attributes by expanding only the most impactful semantic paths at each hierarchical level, with each path's relevance guided by a scoring function that assesses the clas" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14802" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 282, + 262 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 282, + 262 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 282, + 262 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 282, + 262 + ], + "type": "image", + "image_path": "5ce3a010250f4351b7efccb2c7f1d7f7fb8bdc399c9d602f28c6880c42a67a04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 296, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 350 + ], + "type": "text", + "content": "Figure 5. Top-7 Discovered Attributes Across Different Animal Domains. Our method successfully identifies key attributes for multiple domains, such as bird, wildcat, and pet species. The original images are depicted with red borders while the edited images are depicted with black borders. For the pet species domain, we used a binary classifier and for the bird and wildcat species domains, we used a multi-classifier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "text", + "content": "sifier's response to each semantic feature. Formally, the scoring function " + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "text", + "content": " calculates the average classifier " + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "text", + "content": " logit scores across " + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "text", + "content": " images generated using each semantic attribute " + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "text", + "content": ", as defined in Eq. (2), where " + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 366, + 295, + 426 + ], + "type": "text", + "content": " represents the generative diffusion model applied to each sample image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": "We begin with an initial candidate set " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": ", which includes high-level groups from the semantic hierarchy (e.g., broader categories like 'mouth features' or 'eye features'). Each candidate in " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": " is evaluated by the scoring function " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "f(X, s, y)" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": ", which quantifies the influence of the candidate on the classifier's output. Only the top " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": " candidates that surpass a predefined score threshold " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": " are retained, setting a beam width that limits the search to the most impactful candidates. For each candidate in the beam, the algorithm proceeds by expanding to the next hierarchical level, incorporating more specific sub-features (e.g., for \"mouth features,\" sub-features such as \"beard\" and \"mustache\" are included). For each expanded candidate " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "b'" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": ", the scoring function in Eq. (2) is re-evaluated. Only candidates with a score exceeding that of their parent " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "I(b') > I(b)" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": " are retained in the next candidate set " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "S_{next}" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": ", ensuring that only those additions that yield significant incremental impact are added. This process of expansion, scoring, and filtering continues iteratively, moving from general to more specific attributes at each hierarchical level. By dynamically adjusting the candidate set based on beam width " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": " and scoring threshold " + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": ", our method remains computationally efficient, focusing on high-impact combinations rather than exhaustively evaluating all possible attribute pairings. For generating counter" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 177 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 177 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 177 + ], + "type": "image", + "image_path": "8d94ce4c3d27b7b8a02852f2c2442bc68196beb52106c4d6ec7b4cd825c16615.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 186, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 274 + ], + "type": "text", + "content": "Figure 6. Visual Comparison of Key Attributes Identified by StylEx and DiffEx in the Plant Health and Retinal Disease Domains. This figure illustrates the enhanced capability of our method in identifying a broader set of significant attributes compared to StylEx within the plant health and retinal disease domains. DiffEx uncovers more detailed and diagnostically relevant features, such as \"leaf vein color\" and \"macular hole,\" which provide deeper insights into leaf and retina health." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "content": "factual images, we utilize an off-the-shelf diffusion-based editing tool, Ledits++ [8]. However, our approach is versatile enough to accommodate any editing method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 323, + 395, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 395, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 395, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 342, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 606 + ], + "type": "text", + "content": "Experimental Setup. Our experiments utilize Ledits++ [8] and Stable Diffusion XL (SDXL) [40] for generating counterfactual images. Twenty-five time steps are omitted to boost computational efficiency while preserving the quality of edits. The edit threshold, a hyperparameter that dictates the global application scope of edits, is adjusted based on each domain. We test our method across diverse classifiers to evaluate its effectiveness in explaining model behavior through semantic influence. Specifically, we utilize classifiers trained on facial attributes data (i.e. age and gender classifiers) [28] as well as plant health [23], retinal disease [13], bird species [64], wildcat/pet species [11], fashion [35], places [79], and food data [6]. The experiments for the face and plant health domains utilize a CLIP-based classifier [41] to evaluate and interpret edits, while experiments within the bird, wildcat, pet, food, fashion, and places domains use CNN-based classifiers. These CNN classifiers were built on the EfficientNet [61] architecture and achieved an accuracy of over 95 percent on their test sets. For the retinal disease domain, we utilized the FLAIR model [51], which is based on a pre-trained vision-language model, to classify various retina scans." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 613, + 452, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 452, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 452, + 625 + ], + "type": "text", + "content": "4.1. Qualitative Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Explaining Classifiers with Single Attributes. We used various classifiers to analyze fine-grained semantic features for age categorization, species identification, leaf health assessment, retinal disease detection, etc. In the figures illustrating the visual results, the images with red borders indicate the original, unedited versions, while those with black borders are the edited versions, unless noted otherwise." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14803" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 144 + ], + "type": "table", + "html": "
Face (Age)Bird (Species)Leaves (Health)Retina Scans (Disease)Wildcat (Species)Pet (Cat/Dog)
StylExOursStylExOursStylExOursStylExOursStylExOursStylExOurs
Skin PigmentationEyebrowBelly ColorUpperparts ColorBase Leaf ColorBase ColorExudatesGlaucomaSpotsStripesOpen MouthNose Color
Eyebrow ThicknessMakeupUpperparts ColorHead ColorApex ColorVein ColorCotton Wool SpotsSubretinal HemorrhageBlack Tear MarkOpen EyesClosed MouthWhiskers
EyeglassesMustache TypeWing PatternBeak ShapeSpotsApex ColorHemorrhagesIntraretinal HemorrhageEye Shape + SizeNose ColorEye ShapeEye Color
Hair ColorTeethBeak ColorBeak ColorBlightSpots PresenceClustered ExudatesMacular HoleXSpotsDropped EarsClosed Mouth
Lip Thickness + PositionLip VolumeHead ColorWing PatternHalosDisease PresenceXHard ExudatesXFur ColorPointed EarsOpen Mouth
BangsLip ColorBreast ColorEye ColorXBlight SizeXBlackened MaculaXManeEye CircumferenceSpots
Eye MakeupEyelashXThroat ColorXLeaf TextureXSoft ExudatesXEye ColorXCollar
Facial Hair ColorBeard TypeXWing ColorXSpots ColorXRetinal DruenXTongueXPointed Ears
XFacewearXCrest PresenceXDiscolorationXOptic Disc HemorrhageXPupil SizeXMouth Color
XHeadwearXFeather TextureXLeaf OrientationXCataractXWhiskersXFur Pattern
", + "image_path": "fb91ef1216e1942487ac46fe08f6bdc48e0411e2b72bd6df62edd46ca4bdf81a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 184, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 55, + 152, + 555, + 184 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 555, + 184 + ], + "type": "text", + "content": "Table 1. Comparison of Top Attributes Across Different Domains and Classifiers. The table above contains a list of the top attributes discovered by DiffEx (Ours) vs. StylEx. The " + }, + { + "bbox": [ + 55, + 152, + 555, + 184 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 152, + 555, + 184 + ], + "type": "text", + "content": " in the table indicates attributes that were not mentioned in StylEx. It is also important to note that \"cotton wool spots\" and \"soft exudates\" refer to the same condition within the retinal disease domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 184, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 57, + 184, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 57, + 184, + 553, + 222 + ], + "type": "table", + "html": "
MethodCrest PresenceBeak ShapeThroat ColorFeather TextureEye ColorBeak ColorHead ColorUpperparts ColorAvg. Correct Response
Grad-CAM36%50%56%35%47%65%59%76%53%
StylEx68%85%79%82%74%68%91%65%76.5%
DiffEx (Ours)88%91%88%91%82%82%97%88%88.4%
", + "image_path": "2acaa466a2c465eae8e736be7a37d602048c6630a2080de741790207d06fafff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 229, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 555, + 274 + ], + "type": "text", + "content": "Table 2. Comparison with Other Explainability Methods. The table above displays the percentage of correct attribute selections for the bird class, as chosen by users when viewing outputs from different explainability methods. It also includes the average percentage of correct responses across all attributes for each method. As shown, for each attribute presented, the majority of users identified the correct attribute when viewing the output generated by DiffEx." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 277, + 294, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 294, + 481 + ], + "type": "text", + "content": "Face Domain: In the facial domain, as illustrated in Fig. 4, DiffEx identifies and ranks key attributes impacting age classification. For example, features such as \"makeup styles,\" \"lip volume,\" and \"accessories\" (e.g. \"hairbands\") are associated with perceived youthfulness, whereas attributes like \"facial hair\" and \"eyeglasses\" are linked with perceived older age. Notably, the eyeglasses attribute consistently reduces the classifier's score for the \"young\" label, reflecting its association with older demographics. Additionally, DiffEx uncovers hierarchical attribute structures, demonstrating how subcategories within a feature can have varying effects on classifier outcomes. For instance, as shown in Fig. 4, different beard styles impact the perceived age differently (e.g. a \"Balbo beard\" significantly increases the age classification score more than a \"full beard\"). Additional examples of hierarchical explanations of age classifiers are detailed in the appendix (S10)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 483, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 295, + 628 + ], + "type": "text", + "content": "Animal Domain: Furthermore, DiffEx explains classifiers for a variety of animal types. Fig. 5 highlights the top-7 most influential attributes in the bird, wildcat, and pet domains where our method was able to identify fine-grained semantics to explain classifier behavior. For example, in the wildcat classifier, attributes like stripes and manes are critical in distinguishing wildcats from other species. Explaining these classifiers also reveals potential biases. For instance, the presence of a collar significantly increases the likelihood of an animal being classified as a dog. This bias may stem from the training dataset where images of dogs more frequently featured collars compared to cats." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "content": "Medical and Plant Health Domains: Fig. 6 demonstrates the results for retinal disease and plant health domains. For the retinal disease domain, we use the FLAIR model to classify images of retinal fundus scans. This model utilizes detailed domain expert knowledge descriptions, such as \"no hemorrhages, microaneurysms, or exudates\" compared to general descriptions like \"no diabetic retinopathy\" to aid in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "type": "text", + "content": "its classification. For the plant health domain, the CLIP classifier we use looks at features such as the presence of spots, fungus, etc. to make its decision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 316, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 674 + ], + "type": "text", + "content": "Multi-Object Domains involving Complex Scenes: Unlike traditional GAN-based methods that primarily focus on single-object scenarios like a cropped face, DiffEx extends its utility by providing a list of relevant features for domains encompassing multiple objects, such as places, food, and fashion. Classifiers within these domains often evaluate multiple elements simultaneously. For example, a fashion classifier might determine the 'formal or casual fit' of an outfit by considering various components such as the hairstyle, top, and bottom. Similarly, an architectural classifier assessing whether a building appears 'urban or rural' may base its evaluation on not just the structure itself but also its exteriors and surroundings. Explaining the decision-making process of such classifiers is crucial, as they analyze multiple objects within a single image, adding complexity to their interpretative frameworks. Fig. 7 illustrates the top-7 attributes identified by DiffEx for the food, place, and fashion domains. Moreover, we demonstrate DiffEx on various diverse domains. For example, for the food domain, DiffEx was able to discover that removing \"caviar\" or a \"plate decoration\" from the image made the food appear to be more perceived as fast food compared to fine dining. For the places domain, DiffEx was able to find that adding a \"tractor\" to an image or a \"dirt road\" made the place seem more \"rural\" than \"urban.\" In the fashion domain, the style of a neckline significantly influences whether an outfit was classified as \"formal\" or \"casual.\" For instance, a V-neckline was often associated with a more casual look, whereas a classic boat neckline was generally perceived as more formal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Explaining Classifiers with Joint Attributes. Our method goes beyond analyzing individual attributes by identifying attribute combinations that collectively improve classifier" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14804" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 289, + 404 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 404 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 404 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 404 + ], + "type": "image", + "image_path": "6955f2fc9dd51787fc2889e50d11b978631d8cab0a2f1bf8c528a8ec670b4eeb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 413, + 295, + 501 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 501 + ], + "type": "text", + "content": "Figure 7. Top-7 Attributes Discovered by DiffEx for Multi-Object Domains involving Complex Scenes. Compared to existing methods, DiffEx was used to identify meaningful attributes for multi-object domains, such as \"food,\" \"places,\" and \"fashion.\" By identifying semantic features that make food appear \"healthier\" or part of \"fine dining,\" as well as attributes that give a place a more \"rural\" or \"modern\" feel, DiffEx can serve as a powerful tool for understanding perceptions through targeted edits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 502, + 291, + 543 + ], + "blocks": [ + { + "bbox": [ + 61, + 502, + 291, + 543 + ], + "lines": [ + { + "bbox": [ + 61, + 502, + 291, + 543 + ], + "spans": [ + { + "bbox": [ + 61, + 502, + 291, + 543 + ], + "type": "table", + "html": "
RatingBird DomainFace Domain
Edit Quality3.386 ± 0.2233.659 ± 0.248
Disentanglement3.163 ± 0.1973.204 ± 0.213
", + "image_path": "c29e58c95029574ca35e093448f6dfbac4b39a5f3ac6b98fa3451fea08385b8f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 550, + 295, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 572 + ], + "type": "text", + "content": "Table 3. Edit Quality and Disentanglement Ratings. These scores across various domains from User Study 1 are shown above." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "content": "interpretation. While single attributes may have a minimal impact on logits, their combined effect can substantially influence the classifier's output, uncovering subtle interactions that shape decision-making. For example, as shown in Fig. 8, Section C, individual changes in \"lip color\" or \"eye makeup\" result in minimal score changes, but when modified together, they make the subject appear significantly younger. This analysis shows how classifiers may respond more robustly to specific attribute combinations. Additional examples are provided in the appendix (S7, S8, S9)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Extending DiffEx for Multi-Classifier Analysis. We" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "adapt DiffEx for multi-classifier applications to uncover semantic attributes essential for tasks like retinal disease classification (Fig. 6) and bird and wildcat species identification (Fig. 5). This approach highlights key features such as \"upperparts color\" and \"beak shape\" for bird species, \"stripes\" and \"nose color\" for wildcat species, and types of hemorrhages and exudates for retinal conditions. The results in Fig. 6 for the retinal diseases domain and Fig. 5 for the bird/wildcat species domain demonstrate that modifying these attributes significantly alters classifier scores, impacting assessments for species identification and disease likelihood." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 647 + ], + "type": "text", + "content": "Qualitative Comparisons. In Fig. 6, we present a visual comparison of the top attributes and classification scores identified by DiffEx against those identified by StylEx for the plant health and retinal disease domains. As illustrated, DiffEx successfully uncovers a broader set of semantically meaningful attributes compared to StylEx. For instance, DiffEx identifies detailed attributes such as \"leaf vein color\" and \"spots color,\" in addition to the more general attributes found by StylEx, like \"leaf base color\" and \"apex color.\" Table 1 provides a comprehensive comparison of the top features identified by StylEx and our method, highlighting DiffEx's superior ability to uncover semantics that are crucial for the classifier's decision-making process. In particular, the table presents an extended list of features across domains such as faces, bird species, plant health, retina scans, wildcat species, and pet types, all of which influence a classifier's score—further emphasizing DiffEx's ability to uncover fine-grained and contextually rich features. Additionally, in Figure 8 in Sections A and B, we visually compare DiffEx with other diffusion-based, counterfactual generation methods, such as FastDiME [68], DiME [26], and ACE [27] in the face and medical domains. As seen in Figure 8, Section B, DiffEx is able to preserve important features, such as a person's eyes, glasses, etc., in various cases, such as removing a smile or making an older person look younger. Furthermore, Figure 8, Section A shows DiffEx's ability to generate counterfactuals in various medical subdomains, such as skin lesions and chest x-rays. Specifically, for the skin lesion domain, DiffEx was able to remove ruler markings in images from the ISIC dataset [12, 63], while in the chest x-rays domain, DiffEx was able to remove chest drains and pacemakers from images from the NIH [67] dataset and CheXpert [25] dataset. As seen in Figure 8, Section A, DiffEx achieved similar results to FastDiME (Med) [68], a diffusion-based counterfactual model that was trained on medical images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 658, + 459, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 459, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 459, + 671 + ], + "type": "text", + "content": "4.2. Quantitative Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Baselines. To quantitatively evaluate the effectiveness of DiffEx compared to other explainability methods, such as StylEx [31] and Grad-CAM [49], we conducted a series of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14805" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 133 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 133 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 133 + ], + "type": "image", + "image_path": "c9f65854f319e7a13091cdd841d25ad0355c1a9365ba433faa7321199ec01cf2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 141, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 141, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 295, + 240 + ], + "type": "text", + "content": "Figure 8. More Comparisons and Exploring Joint Attributes. (a) We demonstrate DiffEx's ability to produce counterfactuals comparable to FastDiME-2+ in skin lesion and chest X-ray domains. (b) compares DiffEx with ACE, DiME, FastDiME, and FastDiME-2+, showing its capacity to add or remove attributes in a disentangled manner. (c) highlights DiffEx's effectiveness in explaining classifier decisions by analyzing the influence of joint attributes, with perceived age scores (e.g., \"younger\" label) displayed to show how attribute pairings affect classifier outputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 247, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 296, + 413 + ], + "type": "text", + "content": "comprehensive user studies to assess how easily participants could identify the relevant features extracted by our method. User Study 1: Visual Quality and Disentanglement. To evaluate the visual quality and disentanglement of edited images for the face and bird domains, we conducted a user study with 50 participants on Prolific.com. For each domain, we showed pairs of unedited and edited images for ten attributes and asked users to assess whether the edited image contained the desired attribute and was disentangled. For each pair of images, we asked users to rate the edit and disentanglement from one to five, with five representing the highest score. Our results (see Table 3) indicate that our edits reflected the intended attributes while minimizing unrelated changes (see S6.1 in the appendix for more details)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 414, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 414, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 56, + 414, + 296, + 641 + ], + "type": "text", + "content": "User Study 2: Comparison with Grad-CAM and StylEx. To evaluate how effectively our method, DiffEx, explains various semantics within a specific domain compared to other explainability methods (Grad-CAM and StylEx), we conducted a user study with 35 participants on Prolific. We focused on images from the bird domain, generating three sets of three images per attribute: one original image, one edited image, and one image illustrating the explainability method. For Grad-CAM, participants were shown a heatmap overlay on the edited image. For StylEx, we displayed the edited image alone, as this method requires users to manually label the edits. For DiffEx, participants viewed the edited image and the attribute automatically assigned by the VLM. We then asked users to select the attribute (e.g., \"beak color,\" \"crest presence\") that best explained the edited image from four choices. Results indicate that DiffEx significantly outperforms Grad-CAM and StylEx in explaining edited attributes in images (see Table 2). Please refer to S6.2 in the appendix for details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "User Study 3: Disentanglement in Images with Multiple Attributes. We evaluated DiffEx's disentanglement ability when editing multiple attributes simultaneously via a user study. Users rated images on a scale of 1 to 5 (5 = very disentangled, 1 = entangled). DiffEx achieved an average score" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "of 4.6 for single-attribute edits and 4.3 for edits involving four or more attributes, demonstrating effective handling of multiple edits (see S6.3 in the appendix for more details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 556, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 556, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 556, + 263 + ], + "type": "text", + "content": "Ablation Studies. To assess the robustness of our method, we tested various Vision-Language Models (VLMs) and Diffusion Models. For VLMs, we probed Llama 3.2 [20], Qwen2-VL [66], and Claude 3.5 Haiku [1] within the face domain, measuring semantic consistency via IoU scores against GPT-4. The results were 0.94, 0.87, and 0.93, respectively, which indicate substantial semantic overlap. For diffusion models, we compared top attributes from Stable Diffusion (SD) 1.5 and SD 2.0 [46] against SDXL [40] using Normalized Discounted Cumulative Gain (nDCG) to measure the attributes' overlap, achieving scores of 0.75 and 0.76, respectively. These results confirm the stability of our method across different models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 274, + 384, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 384, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 384, + 286 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 294, + 556, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 556, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 556, + 450 + ], + "type": "text", + "content": "Limitations. While our approach provides valuable insights into classifier behavior through semantic edits, it has limitations. Relying on VLM-curated semantics restricts us to the quality and scope of the initial corpus, which may miss relevant features. Additionally, using off-the-shelf editing models like Ledits++ can result in entangled edits (a common issue in image editing algorithms [71]) or fail to capture fine-grained details like earlobes or nose rings (see Appendix S1). Nonetheless, our framework is flexible and can be enhanced with domain-specific adjustments, such as task-specific RAGs [70] or editing methods [7, 14]. RAGs can automatically retrieve semantics in specialized domains (e.g. medical) to improve generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": "Conclusion. In this work, we introduce DiffEx, a novel approach for explaining classifier decisions by utilizing semantic edits within diffusion models. By harnessing the power of vision language models, we curate a comprehensive, hierarchical semantic corpus across various domains and propose a novel algorithm inspired by beam search to filter and rank the most impactful features. DiffEx ranks these semantic features based on their influence on classifier logits, capturing both individual and joint attribute effects, which are crucial for understanding complex classifier behaviors. Through experiments conducted on a wide range of domains—including face, bird, and medical classifiers—we showcase the robustness and adaptability of our approach. Our work provides a powerful tool for interpreting model decisions across diverse applications and promoting transparency in AI-driven classification systems." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14806" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 261, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 261, + 102 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 261, + 102 + ], + "type": "text", + "content": "[1] The claude 3 model family: Opus, sonnet, haiku. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 103, + 296, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 296, + 147 + ], + "type": "text", + "content": "[2] Mostofa Ahsan, Rahul Gomes, Md. Minhaz Chowdhury, and Kendall E. Nygard. Enhancing machine learning prediction in cybersecurity using dynamic feature selector. Journal of Cybersecurity and Privacy, 1(1):199-218, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 149, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 149, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 149, + 294, + 182 + ], + "type": "text", + "content": "[3] Leila Amgoud. Explaining black-box classifiers: Properties and functions. International Journal of Approximate Reasoning, 155:40-65, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 183, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 183, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 295, + 227 + ], + "type": "text", + "content": "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. Advances in Neural Information Processing Systems, 35:364-377, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 229, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 229, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 229, + 294, + 272 + ], + "type": "text", + "content": "[5] Ouahiba Azouaoui and Amine Chohra. Soft computing based pattern classifiers for the obstacle avoidance behavior of intelligent autonomous vehicles (iav). Applied intelligence, 16:249-272, 2002. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 329 + ], + "type": "text", + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "type": "text", + "content": "[7] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. SEGA: Instructing text-to-image models using semantic guidance. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 453 + ], + "type": "text", + "content": "[8] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 510 + ], + "type": "text", + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 555 + ], + "type": "text", + "content": "[10] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 589 + ], + "type": "text", + "content": "[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. CoRR, abs/1912.01865, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 591, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 591, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 58, + 591, + 295, + 667 + ], + "type": "text", + "content": "[12] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[14] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. arXiv preprint arXiv:2312.05390, 2023. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 150 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 554, + 194 + ], + "type": "text", + "content": "[16] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[17] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5744-5753, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "text", + "content": "[19] Yash Goyal, Amir Feder, Uri Shalit, and Been Kim. Explaining classifiers with causal concept effect (cace). arXiv preprint arXiv:1907.07165, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 333, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 554, + 386 + ], + "type": "text", + "content": "[20] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. (arXiv:1512.03385), 2015. arXiv:1512.03385. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 423, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 423, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 423, + 553, + 455 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 457, + 553, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 457, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 457, + 553, + 500 + ], + "type": "text", + "content": "[23] David P. Hughes and Marcel Salathe. An open access repository of images on plant health to enable the development of mobile disease diagnostics. (arXiv:1511.08060), 2016. arXiv:1511.08060 [cs]. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 502, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 553, + 555 + ], + "type": "text", + "content": "[24] Indu Ilanchezian, Valentyn Boreiko, Laura Kuhlewein, Zwei Huang, Murat Seckin Ayhan, Matthias Hein, Lisa Koch, and Philipp Berens. Generating realistic counterfactuals for retinal fundus and oct images using diffusion models. arXiv preprint arXiv:2311.11629, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 558, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 554, + 623 + ], + "type": "text", + "content": "[25] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "text", + "content": "[26] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. In Proceedings of the Asian Conference on Computer Vision, pages 858-876, 2022. 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Adversarial counterfactual visual explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16425-16435, 2023. 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14807" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 4396-4405, Long Beach, CA, USA, 2019. IEEE. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 163 + ], + "type": "text", + "content": "[29] Ziv Katzir and Yuval Elovici. Quantifying the resilience of machine learning classifiers used for cyber security. Expert Systems with Applications, 92:419-429, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "type": "text", + "content": "[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "text", + "content": "[31] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[32] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 365 + ], + "type": "text", + "content": "[33] Joffrey L. Leevy, John Hancock, Richard Zuech, and Taghi M. Khoshgoftaar. Detecting cybersecurity attacks across different network features and learners. Journal of Big Data, 8(1):38, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "text", + "content": "[34] Xuhong Li, Haoyi Xiong, Xingjian Li, Xuanyu Wu, Xiao Zhang, Ji Liu, Jiang Bian, and Dejing Dou. Interpretable deep learning: Interpretation, interpretability, trustworthiness, and beyond. Knowledge and Information Systems, 64 (12):3197-3234, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 476 + ], + "type": "text", + "content": "[35] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 234, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 234, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 234, + 489 + ], + "type": "text", + "content": "[36] OpenAI. Gpt-4 technical report, 2024. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[37] Akin Ozcift and Arif Gulten. Classifier ensemble construction with rotation forest to improve medical diagnosis performance of machine learning algorithms. Computer Methods and Programs in Biomedicine, 104(3):443-451, 2011. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[38] Yi Peng, Guoxun Wang, Gang Kou, and Yong Shi. An empirical study of classification algorithm evaluation for financial risk prediction. Applied Soft Computing, 11(2):2906-2915, 2011. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[39] Emmanuel Pintelas, Meletis Liaskos, Ioannis E Livieris, Sotiris Kotsiantis, and Panagiotis Pintelas. Explainable machine learning framework for image classification problems: case study on glioma cancer prediction. Journal of imaging, 6(6):37, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[40] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[42] R. Joshua Samuel Raj, S. Jeya Shobana, Irina Valeryevna Pustokhina, Denis Alexandrovich Pustokhin, Deepak Gupta, and K. Shankar. Optimal feature selection-based medical image classification using deep learning model in internet of medical things. IEEE Access, 8:58006-58017, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[43] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8848, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 297 + ], + "type": "text", + "content": "[44] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1135-1144, New York, NY, USA, 2016. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 330 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[47] Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. Interpretable machine learning: Fundamental principles and 10 grand challenges. Statistic Surveys, 16:1-85, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "text", + "content": "[48] Axel Sauer and Andreas Geiger. Counterfactual generative networks. (arXiv:2101.06046), 2021. arXiv:2101.06046 [cs]. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "text", + "content": "[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "text", + "content": "[50] Andy Shih, Arthur Choi, and Adnan Darwiche. A symbolic approach to explaining bayesian network classifiers. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, page 5103-5111. AAAI Press, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[51] Julio Silva-Rodriguez, Hadi Chakor, Riadh Kobbi, Jose Dolz, and Ismail Ben Ayed. A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision. Medical Image Analysis, 99:103357, 2025. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[52] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 2014. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[53] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14808" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[54] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. (arXiv:1911.00483), 2020. arXiv:1911.00483 [cs]. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "text", + "content": "[55] Sumedha Singla, Motahhare Eslami, Brian Pollack, Stephen Wallace, and Kayhan Batmanghelich. Explaining the black-box smoothly- a counterfactual approach. (arXiv:2101.04230), 2022. arXiv:2101.04230 [cs, eess]. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[56] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "text", + "content": "[57] Gregor Stiglic, Primoz Kocbek, Nino Fijacko, Marinka Zitnik, Katrien Verbert, and Leona Cilar. Interpretability of machine learning-based prediction models in healthcare. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 10(5):e1379, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "text", + "content": "[58] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, et al. Selective annotation makes language models better few-shot learners. arXiv preprint arXiv:2209.01975, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "type": "text", + "content": "[59] Prabu Subramani, Khalid Sattar, Rocio De Prado, Balasubramanian Girirajan, and Marcin Wozniak. Multi-classifier feature fusion-based road detection for connected autonomous vehicles. Applied Sciences, 11(17):7984, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[60] Jie Sun and Hui Li. Financial distress prediction based on serial combination of multiple classifiers. Expert Systems with Applications, 36(4):8659-8666, 2009. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "type": "text", + "content": "[61] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[62] Manoj Thakur and Deepak Kumar. A hybrid financial trading support system using multi-category classifiers and random forest. Applied Soft Computing, 67:337-349, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[63] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 503, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 559 + ], + "type": "text", + "content": "[65] Wahyono, Laksono Kurnianggoro, Joko Hariyono, and Kang-Hyun Jo. Traffic sign recognition system for autonomous vehicle using cascadesvm classifier. In IECON 2014 - 40th Annual Conference of the IEEE Industrial Electronics Society, pages 4081-4086, 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[66] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 689 + ], + "type": "text", + "content": "[67] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 689, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 294, + 713 + ], + "type": "text", + "content": "[68] Nina Weng, Paraskevas Pegios, Eike Petersen, Aasa Feragen, and Siavash Bigdeli. Fast diffusion-based counterfactuals for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 610 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "shortcut removal and generation. In European Conference on Computer Vision, pages 338-357. Springer, 2025. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[69] M. Wiggins, A. Saad, B. Litt, and G. Vachtsevanos. Evolving a bayesian classifier for ecg-based age classification in medical applications. Applied Soft Computing, 8(1):599-608, 2008. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "text", + "content": "[70] Junde Wu, Jiayuan Zhu, and Yunli Qi. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 185, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 252 + ], + "type": "text", + "content": "[71] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[72] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. (arXiv:2004.03383), 2020. arXiv:2004.03383 [cs]. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "type": "text", + "content": "[73] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. Advances in neural information processing systems, 33: 20554-20565, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 341, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 553, + 387 + ], + "type": "text", + "content": "[74] Muhammad Bilal Zafar, Michele Donini, Dylan Slack, Cédric Archambeau, Sanjiv Das, and Krishnamaram Kenthapadi. On the lack of robust interpretability of neural text classifiers. arXiv preprint arXiv:2106.04631, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "type": "text", + "content": "[75] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "text", + "content": "[76] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "type": "text", + "content": "[77] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 509, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 564 + ], + "type": "text", + "content": "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 564, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 564, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 564, + 553, + 610 + ], + "type": "text", + "content": "[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14809" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_content_list.json b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f8e4bf4410f9c4c220d55039bbfa3d3c8b41c852 --- /dev/null +++ b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_content_list.json @@ -0,0 +1,1781 @@ +[ + { + "type": "text", + "text": "Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves", + "text_level": 1, + "bbox": [ + 98, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zaoming Yan", + "bbox": [ + 107, + 200, + 232, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengcheng Lei $^{1*}$", + "bbox": [ + 269, + 202, + 408, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tingting Wang1", + "bbox": [ + 437, + 203, + 565, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Faming Fang $^{1\\dagger}$", + "bbox": [ + 602, + 202, + 725, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junkang Zhang1", + "bbox": [ + 754, + 203, + 883, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yaomin Huang", + "bbox": [ + 433, + 228, + 560, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haichuan Song $^{1\\dagger}$", + "bbox": [ + 433, + 253, + 570, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, and the KLATASDS Shanghai Key Laboratory of MIP, East China Normal University, Shanghai, China.", + "bbox": [ + 169, + 271, + 823, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{$ yan.zaoming,pengchenglei1995\\} $@$ 163.com", + "bbox": [ + 321, + 309, + 661, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ymhuang@stu.ecnu.edu.cn $\\mathbf{\\Pi}^{1}$ fmfang@cs.ecnu.edu.cn", + "bbox": [ + 282, + 325, + 709, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 377, + 326, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Blurry video frame interpolation (BVFI), which aims to generate high-frame-rate clear videos from low-frame-rate blurry inputs, is a challenging yet significant task in computer vision. Current state-of-the-art approaches typically rely on linear or quadratic models to estimate intermediate motion. However, these methods often overlook depth variations that occur during fast object motion, leading to changes in object size and hindering interpolation performance. This paper proposes the Differential Curves-guided Blur Video Frame Interpolation (DC-BVFI) framework, which leverages the differential curves theory to analyze and mitigate the effects of depth variations caused by object motion. Specifically, DC-BVFI consists of UBNet and MPNet. Unlike prior approaches that rely on optical flow for frame interpolation, MPNet is designed to estimate the 3D scene flow, which facilitates a more precise awareness of depth and velocity variations. Since scene flow cannot be directly inferred in the 2D frame space, UBNet is introduced to transform them into 3D point maps. Extensive experiments demonstrate that the proposed DC-BVFI framework surpasses state-of-the-art performance in simulated and real-world datasets.", + "bbox": [ + 88, + 411, + 485, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 758, + 220, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Natural videos captured by consumer cameras often suffer from real-world movement complexity. Blur Video Frame Interpolation (BVFI) enhances low frame rate and motion-blurred videos into high-quality, sharp outputs between consecutive input frames by reconstructing motion,", + "bbox": [ + 89, + 784, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf69111e34782c861c9fae4961e03207c4aadd5a853c7c743b3eab07d8149535.jpg", + "image_caption": [ + "Linear methods" + ], + "image_footnote": [], + "bbox": [ + 566, + 375, + 705, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f53272ebed1e5ad6eb4cd4cea286a7a7acb9f96b2b7bc93ac878cd95a57fa5aa.jpg", + "image_caption": [ + "Quadratic methods" + ], + "image_footnote": [], + "bbox": [ + 714, + 375, + 856, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20d70d3e27f5afc714bf6009a54feb60ab37795b1471b51075490a4eb10e465f.jpg", + "image_caption": [ + "Differential curves", + "Figure 1. Challenge. Linear methods: Synthesizing intermediate frames through linear approximation. Quadratic methods: Synthesizing intermediate frames by fitting velocity variations. Differential curves: Exploring the depth variations." + ], + "image_footnote": [], + "bbox": [ + 560, + 507, + 700, + 609 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c3765191ef6b11033abd73dd13a1422a93bc4488ba370b61d95891b49c4c397.jpg", + "image_caption": [ + "Real-world motion" + ], + "image_footnote": [], + "bbox": [ + 712, + 506, + 852, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thereby improving perception for the human visual system [22, 23]. Therefore, it is widely used for diverse applications[20, 34, 49], and it has become a significant research area in computer vision.", + "bbox": [ + 511, + 700, + 906, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "BVFI methods can be broadly divided into two categories: linear methods and quadratic methods. Linear methods [3, 12, 29, 30, 32] typically employ existing motion estimators [17, 43, 45] to predict optical flow between input frames. The intermediate optical flow is obtained by linearly approximating the optical flow between input frames to synthesize the intermediate frame. However, linear methods often overlook variations in the velocity of object motion. To more accurately reconstruct complex motion, un", + "bbox": [ + 509, + 763, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1994", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "like linear methods, quadratic methods [11, 28, 50, 51, 56] leverage optical flow to explore velocity variations in videos for fitting complex motions.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, upon scrutinizing and experimenting on the released implementations of quadratic methods[11, 27, 28, 51], we observed that they overlook the depth variations and size changes caused by the car's motion, making them ineffective for video frame interpolation. We conjecture that the reasons are twofolds. Limitations of the Quadratic Theory. The quadratic theory primarily focuses on velocity variations in the horizontal and vertical directions, overlooking the object size variations caused by depth variations in the real world. Specifically, the quadratic theory is based on two linearly independent vectors and lacks a linearly independent vector to represent depth variations. As a result, when an object's size changes, the quadratic theory fails to describe it. Limitations of the Optical Flow. The implementation of the quadratic theory relies on optical flow; however, optical flow struggles to capture motion-induced depth variations. Specifically, an intrinsic limitation of optical flow is that a nearby object that moves slowly will generate the same optical flow as a faraway object that moves fast [5, 52].", + "bbox": [ + 91, + 138, + 482, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above analysis, to overcome the limitations of the quadratic theory in handling depth variations, we propose leveraging the differential curves theory to unify the modeling of depth and velocity variations in motion. Specifically, we map pixels onto slight curve segments in 3D space, with each curve defined by a unit basis and motion parameters. The unit basis represents spatial coordinates in 3D space, while the motion parameters describe the 3D displacement of the unit basis for the next time step. Differential curves provide an intuitive core idea for describing 3D spatial motion; however, its implementation presents certain challenges. The primary challenges include converting 2D video frames into 3D point map representations and accurately estimating the motion parameters between point maps.", + "bbox": [ + 89, + 443, + 482, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To convert 2D video frames into 3D point maps, we propose UBNet. Although the idea of converting video frames into 3D camera space point maps is intuitive and sensible, the task is challenging because it requires estimating camera parameters and monocular depth maps from video frames. To address this issue, UBNet leverages a general transformer architecture [46], which [37, 46, 47, 59] can estimate camera intrinsics, depth maps, and camera-space point maps from unposed frames.", + "bbox": [ + 89, + 671, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To estimate motion parameters between point maps, we introduce MPNet. MPNet estimates motion parameters with depth awareness, but this task is challenging because it requires estimating scene flow from the desired point map to the source point map, which existing methods struggle to achieve. To address this issue, MPNet combines a method", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that upgrades the 2D optical flow obtained from the video interpolation model (IFnet [14]) to produce the camera-space scene flow. By leveraging these modules, we build a Differential Curves-guided Blurry Video Frame Interpolation (DC-BVFI) framework.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper presents three key contributions: (1) To the best of our knowledge, this is the first work that introduces differential curves theory to alleviate the challenge of depth variations in blurry video multi-frame interpolation. (2) This paper proposes an explicit depth-aware video interpolation framework to overcome depth variation and synthesize intermediate video frames more effectively. The UB-Net converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. (3) DC-VFI outperforms state-of-the-art results across four benchmark datasets.", + "bbox": [ + 511, + 167, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video multi-frame interpolation. Recently, advancements in deep learning have led to various methods for video frame interpolation. These methods can be broadly categorized into two main paradigms: reconstruction-based and generation-based.", + "bbox": [ + 511, + 371, + 903, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstruction-based methods are mainly categorized into two types: optical flow-based methods and kernel-based methods. Initially, optical flow-based methods introduced a linear prior[3, 12, 29, 30, 32], which were subsequently expanded to quadratic flow prior [11, 28, 51, 55] and even cubic [6] trajectory models to represent the local motion of pixels accurately. To obtain the optical flow from the intermediate frame to the input frames, [11, 13] proposed distillation strategies to obtain the optical flow from the intermediate frame to the input frames. These Kernel-based methods are implemented as separable convolutions [35], deformable convolutions [4, 25, 26, 40], and trilinear sampling kernels[30].", + "bbox": [ + 511, + 448, + 905, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods based on Denoising Diffusion Probabilistic Models (DDPM)[10] leverage generative techniques like DDPM to fill occlusions caused by motion. These DDPM-based approaches are implemented as score-based diffusion [8, 19, 48], motion-aware diffusion [15], and Brownian bridge diffusion [24, 31]. However, most DDPM-based methods are significantly time-consuming and challenging for real-time inference.", + "bbox": [ + 511, + 643, + 905, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Blurry video multi-frame interpolation. When the original reference frames degrade, existing VFI methods face significant challenges in recovering sharp intermediate frames and reconstructing motion trajectories. Early methods primarily employed a cascade of deblurring and interpolation techniques [9, 21]. [21, 38, 39] extract frames, which are then subsequently used to generate intermediate clear frames by jointly optimizing. Oh et al. [36] proposed a DeMFI for joint deblurring and multi-frame interpolation,", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1995", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "incorporating a flow-guided attentive-correlation-based feature bolstering module and recursive boosting. The critical contribution of Zhong et al. [58] is contributing a real-world blurry video with a low frame rate dataset.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 164, + 181, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary of Differential Curves", + "text_level": 1, + "bbox": [ + 89, + 189, + 390, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to the definition of differential curves [44], the unit basis can be written as:", + "bbox": [ + 89, + 210, + 483, + 239 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} \\alpha (s) = r ^ {\\prime} (s), \\\\ \\beta (s) = r ^ {\\prime \\prime} (s), \\\\ \\gamma (s) = \\alpha (s) \\wedge \\beta (s), \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 250, + 482, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where ${}^{\\prime}$ is the derivative symbol of $r(s)$ , ${}^{\\prime\\prime}$ is the second derivative symbol of $r(s)$ , and $\\wedge$ is the symbol of the outer product. The $r(s)$ is a slight segment curve of the point $s$ . The $\\alpha(s)$ , $\\beta(s)$ , and $\\gamma(s)$ together represent the spatial position of $r(s)$ , termed the unit basis.", + "bbox": [ + 89, + 319, + 483, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to the definition of differential curves [44], the derivative of the unit basis can be written as:", + "bbox": [ + 89, + 396, + 483, + 424 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} \\alpha^ {\\prime} (s) \\\\ \\beta^ {\\prime} (s) \\\\ \\gamma^ {\\prime} (s) \\end{array} \\right] = \\underbrace {\\left[ \\begin{array}{c c c} 0 & \\kappa (s) & 0 \\\\ - \\kappa (s) & 0 & \\tau (s) \\\\ 0 & - \\tau (s) & 0 \\end{array} \\right]} _ {\\text {m o t i o n p a r a m e t e r s}} \\left[ \\begin{array}{l} \\alpha (s) \\\\ \\beta (s) \\\\ \\gamma (s) \\end{array} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 431, + 482, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\kappa(s)$ and $\\tau(s)$ is the curvature and torsion. The $\\kappa(s)$ and $\\tau(s)$ together represent the spatial displacement of $r(s)$ , termed the motion parameters. Essentially, the motion parameters depict the spatial displacement of $r(s)$ , similar to optical flow. We simplify the notation $(\\alpha(s), \\beta(s), \\gamma(s), \\kappa(s)$ and $\\tau(s))$ to $(\\alpha, \\beta, \\gamma, \\kappa$ and $\\tau$ ) for convenience expression.", + "bbox": [ + 89, + 513, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem formulation", + "text_level": 1, + "bbox": [ + 89, + 628, + 287, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Blurry Video Frame Interpolation. In the blurry video frame interpolation task, it can be written as", + "bbox": [ + 89, + 650, + 483, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} = \\mathbf {F} \\left(B _ {i}, B _ {i + \\eta}\\right), t \\in (i, i + 1, \\dots , i + \\eta), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 691, + 482, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B_{i}$ and $B_{i + \\eta}$ are input frames. $i$ is a current time index and $t$ is an interpolated time index between $i$ and $i + \\eta$ . The goal is to synthesize a frame $I_{t}$ by inputting frames $B_{i}, B_{i + \\eta}$ and an optimization function $(\\mathbf{F}(\\cdot))$ .", + "bbox": [ + 89, + 719, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Blurry Video Frame Interpolation under Differential Curves. Most failures in video frame interpolation occur due to object motion, which inevitably leads to depth changes (object scale changes) between adjacent frames. To mitigate the impact of depth variation on video interpolation quality, we redefine the video interpolation problem under differential curves theory, allowing for explicit exploration of depth changes.", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/da97134077be8d4cdb2d2743b9204778cf0026f1454d2b66f245a4104ec1977e.jpg", + "image_caption": [ + "Figure 2. Video Frame under differential curves. From the perspective of differential curves, each pixel $s_n$ can be viewed as a light segment curve $r(s_n)$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 901, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the convenience of expression, we write a time sequence $B_{i}\\rightarrow I_{t}$ . Each frame $B_{i}$ contains $H\\times W$ pixels $(s_n,n = 1,\\dots ,H\\times W)$ . From the perspective of differential curves, each pixel $s_n$ can be viewed as a slight segment curve $r(s_{n})$ , where its spatial position denotes as $[\\alpha (s_n),\\beta (s_n),\\gamma (s_n)]\\in \\mathbb{R}^{1\\times 3}$ . Then, the frame $B_{i}$ can be presented as $H\\times W$ slight segment curves, denoted as $[\\alpha ,\\beta ,\\gamma ]_i\\in \\mathbb{R}^{H\\times W\\times 3}$ . Subsequently, the video frame interpolation process $(\\mathbf{F}(B_i)\\to I_t)$ can be reformulated as $(\\mathbf{F}([\\alpha ,\\beta ,\\gamma ]_i)\\to [\\alpha ,\\beta ,\\gamma ]_t)$ . We reformulate the blurry video frame interpolation problem under the differential curves theory and write as follows:", + "bbox": [ + 511, + 287, + 906, + 469 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {t} = \\mathbf {F} \\left(\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {i}\\right), t \\in (i, i + 1, \\dots , i + \\eta) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 474, + 903, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is $F([\\alpha, \\beta, \\gamma]_i) \\to [\\alpha, \\beta, \\gamma]_t$ . $\\mathbf{F}(\\cdot)$ is an optimization function that predicts the position at time $t$ from the position at time $i$ through optimization algorithm.", + "bbox": [ + 511, + 529, + 903, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Optimization algorithm", + "text_level": 1, + "bbox": [ + 511, + 582, + 732, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A multi-stage strategy ensures the temporal continuity of motion [27]; however, current multi-stage strategies[27, 36, 38, 39] lack algorithmic convergence, making it challenging for them to converge quickly. To swiftly and directly approach the target, we propose an optimization algorithm. Following Eq. (4), the minimized objective function can be written as", + "bbox": [ + 511, + 604, + 905, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\left\\| \\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {t} - \\mathbf {F} \\left(\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {i}\\right) \\right\\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 712, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To solve the minimization Eq. (5), we utilize the gradient descent algorithm [1].", + "bbox": [ + 511, + 766, + 903, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging the gradient descent algorithm [1], the Eq. (5) can be optimized:", + "bbox": [ + 511, + 797, + 903, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {2} = \\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {1} + \\underbrace {\\left[\\begin{array}{c c c}0&\\kappa&0\\\\- \\kappa&0&\\tau\\\\0&- \\tau&0\\end{array}\\right] _ {i} ^ {1} \\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {1}} _ {\\text {g r a d i e n t} (S F _ {i \\rightarrow t})}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 830, + 903, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1996", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d08a29bb02c176fedb50a7341e66ee2cf5eec772b665fbfe60b07b1bca64030.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 85, + 885, + 176 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34ec66a6da5808b5993c12f89c21a389ea7dc04e28586835374d0c5395c46270.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 181, + 790, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50c7f933084fda398affecbaa3a9ba60d21938ae8e6e2d57fa76470479d87db4.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed DC-BVFI. UBNet primarily converts video frame $I_{i}$ into camera space point map $[\\alpha, \\beta, \\gamma]_{i} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where 3 represents the 3D coordinate, and $H \\times W$ indicates the total number of point. MPNet estimates depth-aware scene flow $SF_{i \\rightarrow t} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where 3 represents the 3D coordinate offset between two point maps." + ], + "image_footnote": [], + "bbox": [ + 94, + 356, + 517, + 522 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3eb326dd75e3a1cb8a34044643e9ca3da395446166a8de567b1b43f70ef63d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 356, + 785, + 498 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b0b43f99a91f4d85facba09a2986b5f19ce90c558770345542c747998fc84812.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 186, + 898, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$[\\alpha, \\beta, \\gamma]_i^1$ denotes the spatial position of $H \\times W$ slight segment curves at the $1^{st}$ iteration. And $[\\alpha, \\beta, \\gamma]_i^2$ is the spatial position of the same slight segment curves at the $2^{nd}$ iteration. $[\\alpha, \\beta, \\gamma]_i^1$ is updated by adding the gradient of $[\\alpha, \\beta, \\gamma]_i^1$ . Figure 2 shows the iterative optimization process from $1^{st}$ iteration to $2^{nd}$ iteration (Eq. (6)).", + "bbox": [ + 89, + 590, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Differential Curves for Blurry Video Frame Interpolation", + "text_level": 1, + "bbox": [ + 89, + 705, + 482, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although using differential curves theory to describe spatial motion is reasonable, its implementation is challenging: 1) How to convert a video frame $I_{i}$ into a camera space point map $[\\alpha, \\beta, \\gamma]_{i}$ and represent the spatial position. 2) How to estimate the relative displacement $SF$ from an unknown point map to a source point map. And existing methods struggle to obtain the relative displacement. To this end, UBNet and MPNet are designed in DC-BVFI to address these issues. The pipeline of the proposed DC-BVFI is outlined in Fig. 3.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1. UBNet", + "text_level": 1, + "bbox": [ + 513, + 590, + 607, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Movement inevitably causes motion blur [38] and depth change. To convert the blurred video frame $B_{i}$ into a 3D camera space point map $[\\alpha, \\beta, \\gamma]_i$ , we designed a UBNet. More specifically, the UBNet includes two submodules:", + "bbox": [ + 511, + 609, + 905, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setup 1. The function of this submodule is to deblur. Traditional methods often simply apply Direct Stacking of video frames for deblurring. However, motion usually causes the content of frames to be different, resulting in misalignment between frames. Directly Stacking (DS) misalignment frames can hinder intermediate frame restoration, which may not facilitate video deblurring. Therefore, we introduce a Discriminative Deblur (DD) strategy, as shown in Fig. 3, to mitigate the adverse effects of blurring. The operations can be expressed as follows: $\\{B_i\\} \\xrightarrow{\\mathrm{DD}} \\{I_i\\}$ .", + "bbox": [ + 511, + 671, + 905, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setup 2. Real-world motion is a three-dimensional movement involving both depth and velocity variations. However, existing video frame interpolation methods typically focus only on velocity variations, which limits the quality of interpolation. To effectively explore three-dimensional", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1997", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "motion, it is essential to convert the two-dimensional pixel space into the three-dimensional camera space. To convert the two-dimensional pixel space into the three-dimensional camera space, we employ a general 3D reconstruction paradigm [46], which can estimate camera intrinsics, depth maps, and point maps in the camera space from video frames. Specifically, the coordinate of the frame is mapped to a three-dimensional coordinate $[\\alpha, \\beta, \\gamma] \\in \\mathbb{R}^{H \\times W \\times 3}$ , where 3 represents the 3D coordinate, and $H \\times W$ indicates the total number of point. The operations can be expressed as follows: $\\{I_i\\} \\xrightarrow{\\mathrm{ViT}} \\{[\\alpha, \\beta, \\gamma]_i\\}$ .", + "bbox": [ + 89, + 90, + 483, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. MPNet", + "text_level": 1, + "bbox": [ + 89, + 271, + 187, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In optical flow-based video interpolation methods, it is often necessary to estimate the optical flow $F_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 2}$ from the frame $I_{i}$ to the frame $I_{t}$ for interpolation [13, 28, 51]. Similarly, in 3D camera space, it is also necessary to estimate the scene flow from the point map $[\\alpha ,\\beta ,\\gamma ]_i$ to the point map $[\\alpha ,\\beta ,\\gamma ]_t$ . However, because the $[\\alpha ,\\beta ,\\gamma ]_t$ is unknown, existing methods cannot directly obtain the scene flow. To this end, we propose MPNet to upgrade the optical flow $F_{i\\rightarrow t}$ to the scene flow $SF_{i\\rightarrow t}$ .", + "bbox": [ + 89, + 291, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 1. Given the frames $\\{I_i\\}$ , flow-based methods [13, 51, 54] require approximating the intermediate flows $F_{i\\rightarrow t}$ from the perspective of the synthesized frame $I_{t}$ . Considering real-time performance, we employ IFNet [13] to estimate the optical flow $F_{i\\rightarrow t}$ .", + "bbox": [ + 89, + 428, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 2. To express positional changes between point maps, we upgrade the optical flow $F_{i\\rightarrow t}$ to camera space scene flow $SF_{i\\rightarrow t}$ to aware depth variations. The process is written as the following formula:", + "bbox": [ + 89, + 505, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS F _ {i \\rightarrow t} = \\left(F _ {i \\rightarrow t} \\mid D _ {i \\rightarrow t}\\right) \\times K ^ {- 1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 579, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 1} = D_i - D_t$ is the depth maps and the camera intrinsics $K\\in \\mathbb{R}^{3\\times 3}$ . We set $D_{t} = (1 - t)D_{i} + tD_{i + \\eta}$ , where $D_{i}$ and $D_{i + \\eta}$ are the depth maps of two adjacent frames. $t\\in \\{i,i + \\eta \\}$ is the time index for interpolation. The $\\times$ denotes matrix multiplication, while | represents concatenation. The scene flow $SF_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 3}$ , where 3 represents the 3D coordinate offset between two point maps. We obtain the coarse scene flow $SF_{i\\rightarrow t}$ .", + "bbox": [ + 89, + 609, + 483, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 3. Blurry frames lead to rough optical flow, leading to coarse scene flow. Therefore, it is necessary to refine the coarse scene flow. To achieve this, we introduce the Local Self-Attention (LSA) and Residual Block (RB) strategies, which perform fine-grained adjustments at the pixel level to obtain refining scene flow.", + "bbox": [ + 89, + 748, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3.Warp", + "text_level": 1, + "bbox": [ + 89, + 849, + 176, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A brief explanation of the point map warping process by scene flow. Specifically, given a scene flow $SF_{i\\rightarrow t}$ and a", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point map $[\\alpha, \\beta, \\gamma]_i$ . The warp process can be expressed as:", + "bbox": [ + 511, + 90, + 903, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ \\alpha , \\beta , \\gamma ] _ {t} = [ \\alpha , \\beta , \\gamma ] _ {i} + S F _ {i \\rightarrow t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 114, + 903, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $+$ denotes element-wise addition and $[\\alpha, \\beta, \\gamma]_t$ represents the warped point map.", + "bbox": [ + 511, + 140, + 903, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To Frame. It is important to note that the warped point map represents 3D coordinates. We need to project the point map back into the 2D coordinate system. The pseudocode for warping and projecting onto the 2D frame is as follows:", + "bbox": [ + 511, + 170, + 905, + 231 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Pseudocode." + ], + "code_body": "Warp pt_t = pt + sf # point map is projected to frame b, _, h, w = pt.size() pt_t_ = torch/matmul(K, pt_t.view(b, 3, -1)) coord_t = pt_t_.div(pt_t_[:, 2:3, :])[:, 0:2, :] coord_t = coord_t.view(b, 2, h, w) grid = coord_t.transpose(1, 2).transpose(2, 3) I_t = torch.grid_sample(I_i, grid)", + "bbox": [ + 527, + 265, + 955, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function.", + "text_level": 1, + "bbox": [ + 513, + 422, + 663, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the mean square error (MSE) as the training loss function:", + "bbox": [ + 511, + 445, + 905, + 473 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| I _ {i} - I _ {i} ^ {g t} \\right\\| ^ {2} + \\left\\| I _ {t} - I _ {t} ^ {g t} \\right\\| ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 482, + 903, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The variable $I_{t}^{gt}$ denotes the ground truth frame. $I_{t}$ denotes the output frame of the model.", + "bbox": [ + 511, + 508, + 905, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 551, + 645, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 575, + 609, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adobe. The Adobe dataset [42] consists of 120 videos at 240 fps with a resolution of $1280 \\times 720$ . The blurry frames are 30 fps and downsized to $640 \\times 352$ as done in [36, 38]. We follow [36, 39] to use the same 8 videos of the Adobe dataset for evaluation and other videos of the Adobe dataset for training. RBI. The real-world blurry frame interpolation [58] dataset comprises 55 videos. We use the same videos as the test set and the others as the training set [58]. The blurred image has an exposure time of $18~\\mathrm{ms}$ , while the sharp image has an exposure time of nearly $2\\mathrm{ms}$ [58]. This means that there are 9 sharp frames corresponding to one blurred frame, and 11 sharp frames exist in the readout dead-time between adjacent blurred frames [58]. GoPro. We use the same 11 videos of the GoPro dataset [33] to test the model [38]. The test datasets are also temporally downsampled to 30 fps with the blurring as [36]. YouTube. As a test dataset, the YouTube [38] consists of 59 video clips with different scenes and 12590 frames in total [38]. Please note that the evaluation datasets are also temporally downsampled to 30 fps with the blurring as [36].", + "bbox": [ + 511, + 598, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1998", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c46330881359bdd860a066e3bbeeff2286a1f73ee36652227bbec6af3485e1e7.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on Adobe and RBI. The best and the second best results are denoted by pink and yellow." + ], + "table_footnote": [], + "table_body": "
MethodAdobeRBI
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]28.730.86528.660.86428.670.86427.040.84026.920.83827.130.840
TNTT [21]29.400.89329.450.87629.450.87628.880.85727.360.85428.110.855
DAIN [2]--29.300.877----28.380.860--
BIN [38]31.760.89628.360.84729.950.87829.010.86027.580.86128.300.861
ALANET [9]33.710.93232.980.93633.340.93528.810.86828.880.86828.840.868
UTI-VFI [56]28.730.86528.660.86428.670.86427.340.84427.120.84727.230.844
EQVI [28]--29.690.889----27.790.840--
ANIB [57]31.300.89032.830.89932.520.89428.830.88828.990.87228.900.870
DeMFI [36]33.830.93733.790.94133.790.94029.690.89929.000.89029.350.893
BiT [58]33.600.93633.860.93833.730.93729.580.89329.840.89729.750.894
JNMR [27]31.330.90731.680.93531.430.93928.520.87328.710.87928.630.875
IQ-VFI [11]33.830.94033.800.94133.790.94029.950.89429.920.89729.910.894
Ours34.170.94234.100.94234.140.94231.020.90430.890.91531.030.913
", + "bbox": [ + 166, + 109, + 826, + 348 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/74ae59ae8a1440e203843d28c4767408b08cfb2b9ea7fe0f1d9166ef91c47036.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on GoPro and YouTube. The best and the second best results are denoted by pink and yellow." + ], + "table_footnote": [], + "table_body": "
MethodYouTubeGoPro
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]29.270.88429.980.89029.990.89027.200.81526.620.810--
TNTT [21]29.590.88929.770.89029.750.89026.480.80826.680.81526.650.814
DAIN [2]--32.020.901----25.540.859--
UTI-VFI [56]28.610.87928.640.87328.640.87725.660.80825.630.84525.640.814
EQVI [28]--29.110.893----29.680.881--
DeMFI [36]32.900.92532.790.92632.800.92630.820.89330.780.90030.750.900
BiT [58]32.370.92031.880.91831.500.91531.270.90631.200.90528.950.885
JNMR [27]--32.250.922----31.020.900--
IQ-VFI [11]33.030.92932.900.93032.940.92831.320.90631.260.90431.280.904
Ours33.400.93233.320.93333.350.93232.140.91232.050.91032.080.910
", + "bbox": [ + 165, + 382, + 823, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details.", + "text_level": 1, + "bbox": [ + 89, + 593, + 310, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We optimize the loss using Adam in the PyTorch framework. The cosine scheduler schedules the learning rate from $1e - 4$ to $1e - 6$ . Standard data augmentation techniques, such as flipping, rotation, and cropping, are applied to the data with a size of $256 \\times 256$ . We train our model on the training datasets with a batch size of 4 for 900 epochs on 4 NVIDIA 3090ti GPUs.", + "bbox": [ + 89, + 618, + 483, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with Previous Methods", + "text_level": 1, + "bbox": [ + 89, + 739, + 401, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, a total of 12 outperformers are compared, which are fine-tuned on the Adobe and RBI datasets.", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on the RBI. Our approach achieves a 1.12dB improvement over the current state-of-the-art. The RBI dataset contains a large number of depth-varying scenes, such as vehicle motion and camera movement, resulting in a high average motion magnitude [58]. DC-BVFI performs well on the RBI dataset, demonstrating that the explicit depth-aware approach is advantageous in challenging scenarios,", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "such as vehicle motion, where many existing methods tend to fail. The proposed DC-BVFI consistently outperforms existing methods, achieving a PSNR of 34.14 on Adobe.", + "bbox": [ + 511, + 594, + 903, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the generalization capability, we conducted cross-dataset experiments using YouTube and GoPro as test sets, as shown in Tab. 2. Results on the GoPro. We observed that our method achieved an impressive performance of 32.08 PSNR on the GoPro test set. The GoPro dataset contains a large number of dynamic scenes, resulting in a high average motion magnitude, making it a challenging benchmark for video frame interpolation. Our method performs well on the GoPro dataset, suggesting that DC-BVFI exhibits strong generalization across various motion scenarios. We also observed the final performance of 33.35 PSNR on the YouTube test set.", + "bbox": [ + 511, + 640, + 906, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison of Visual Results.", + "text_level": 1, + "bbox": [ + 511, + 832, + 781, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further compare the visualization results in depth-varying motion scenarios, such as car motion and camera motion. As shown in Fig. 4 our method's predictions are", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1999", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/340315f5ee2f3dcf609005d75cc80d76c1a9f891e9fb83cecef6cf7006f4d47a.jpg", + "image_caption": [ + "Figure 4. Visual comparison with different methods. The comparison primarily focuses on depth change scenarios, such as car motion and camera movement." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2367541d7718110ebb4a26fdfa7db0877bf03450b71f2186e4b0b762b95b7cbd.jpg", + "image_caption": [ + "Figure 5. Comparison of Motion-in-Depth." + ], + "image_footnote": [], + "bbox": [ + 112, + 349, + 464, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "closer to the ground truth, with more precise details on both car motion and camera motion. Methods like QVI [51], EQVI [28], and IQ-VFI [11] explore inter-frame motion through estimated acceleration to enhance motion reconstruction. However, they struggle to express changes in object depth during motion, often leading to suboptimal outcomes.", + "bbox": [ + 89, + 608, + 482, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Motion-in-Depth. Moreover, to quantitatively measure depth variations during motion. We employ the MiD metric [18] to evaluate the depth variations. Since the video interpolation dataset lacks ground truth depth maps, we use the depth maps estimated from clear video frames [53] as the ground truth labels. Fig. 5 reports the performance of quadratic methods on multi-frame interpolation in Adobe, RBI, YouTube, and GoPro. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of MiD. This indicates that our method effectively explores the depth variations cue to enhance video frame", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36d03ef72aa0b1895d518ca52e42780da7e3fa1af4beeabf0eb2d7ca989e5030.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 348, + 691, + 457 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f17612969cad5b6966540cf8451e926a7dbc13e1e78d4837f3e197dd7d59351.jpg", + "image_caption": [ + "Figure 6. Comparison of temporal consistency performance under tOF metric in Adobe, RBI, YouTube, and GoPro datasets." + ], + "image_footnote": [], + "bbox": [ + 516, + 459, + 691, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a69dc98d0ad741ae7c4855c414cf549f719d6bc6f9c6cda3de77e8d8eaecb9d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 348, + 867, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39c8376a3ee967108d799b9b896a63dca7d612a0a2f1c56168edfd12a950eeb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 459, + 867, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "interpolation.", + "bbox": [ + 513, + 612, + 606, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Temporal Consistency. We employ the tOF metric [7, 16, 41] to evaluate the temporal consistency. Figure 6 reports the quality of motion reconstruction across several existing models on the Adobe, RBI, GoPro, and YouTube test sets. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of tOF. The outstanding temporal consistency is due to DC-BVFI's explicit depth modeling of rapid object motion, which greatly improves motion temporal consistency.", + "bbox": [ + 511, + 627, + 906, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 787, + 679, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation experiments to evaluate the contribution of the 3D architecture and the impact of each submodule in our method, namely the DD and the LSA. Additionally, we investigated the influence of scene flow and optical flow on video frame interpolation. Furthermore, we measured the parameter count and computational complex", + "bbox": [ + 511, + 810, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2000", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45c9e0650b262a73cf51a7e00bc963f1f02ed9a71a28650217f37b33b6180086.jpg", + "table_caption": [ + "Table 3. Ablation studies of architecture on the RBI dataset. DS stands for stacking video frames along the channel axis, while DD denotes the discriminative deblur strategy. LSA and RB represent Local Self-Attention and Residual Block, respectively." + ], + "table_footnote": [], + "table_body": "
ArchitectureBackBoneSettingUBNetMPNetRBI
DDDSLSARBPSNRMiD(↓)
3DViT-Base(a)✓✓31.0376
(b)✓✓30.9883
(c)✓✓30.6681
2DViT-Base(d)✓✓28.91137
", + "bbox": [ + 94, + 156, + 480, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/435732997026487a3b6bd0cc13f36f975e19bd890c0700f84df1c9572914ff79.jpg", + "image_caption": [ + "Figure 7. Visualization of the architecture ablation experiments. In the scene with car depth variations, our architecture achieves promising results." + ], + "image_footnote": [], + "bbox": [ + 91, + 262, + 483, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ity of various models.", + "bbox": [ + 89, + 460, + 238, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture. We empirically evaluated the proposed 3D architecture's effectiveness in guiding video frame interpolation in scenes with depth variations. For a fair comparison, both methods use the same backbone but differ in their output prediction heads. PSNR and MiD are presented in Tab. 3. The results of Tab. 3 and Fig. 7 indicate that the 3D architecture significantly outperforms the 2D architecture in scenes with depth variations. The results of setting (b) of Tab. 3 and Fig. 7 highlight the importance of refining scene flow. The results of setting (c) of Tab. 3 and Fig. 7 show that directly stacking misaligned frames hinders intermediate frame restoration, and may not facilitate video deblurring.", + "bbox": [ + 88, + 477, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Flow. We analyzed the effects of scene flow and optical flow on video frame interpolation performance. For a fair evaluation, in the 2D architecture, we replace the depth estimation in the 3D architecture with a ViT-Base model that has the same parameter size. The results in Tab. 4 and Fig. 8 indicate that scene flow outperforms optical flow in scenarios with depth variations. This is in line with our intuitions, as scene flow integrates depth information on top of optical flow, providing a more comprehensive consideration of useful cues. In other words, when motion doesn't cause depth variations, the scene flow in the depth spatial dimension is deactivated, degrading scene flow into the optical flow. Essentially, scene flow encompasses optical flow. These additional insights contribute to improved video frame interpolation performance in scenes with depth variations.", + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1129cd8c3be4eaac729f829f08276a84e90c59ecc952e7da2745ead0fe1883d.jpg", + "table_caption": [ + "Table 4. Ablation studies of Flow on the RBI dataset." + ], + "table_footnote": [], + "table_body": "
ArchitectureFlowSettingFlowNetDepthRBI
IFNetRAFTMeGeDust3RPSNRMiD(↓)
3DScene Flow(a)✓✓31.0376
(b)✓✓30.9982
(c)✓✓31.0079
(d)✓✓30.8685
ArchitectureFlowIFNetRAFTViT-Base 2DPSNRMiD(↓)
2DOptical Flow(e)✓✓28.80143
(f)✓✓28.78147
", + "bbox": [ + 517, + 114, + 903, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58ba4783ba13072ebff9ae21e54e6b27d6755ff0444587d084b17957a3ab7510.jpg", + "image_caption": [ + "Figure 8. The results of video multi-frame interpolation are presented. The results of the 3D-based architecture (a) are shown in the top rows, and the 2D-based architecture (e) in the bottom rows." + ], + "image_footnote": [], + "bbox": [ + 516, + 262, + 898, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 614, + 633, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This is the first work to focus on solving the depth variations in video frame interpolation. We introduce the differential curves to analyze the issue of depth variations in video frame interpolation. Based on the analysis, this paper builds an explicit depth-aware video interpolation framework to overcome the issue. The UBNet converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. DC-BVFI has achieved state-of-the-art performance on four benchmark datasets: Adobe, RBI, YouTube, and GoPro.", + "bbox": [ + 511, + 648, + 906, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the National Key R&D Program of China (2022ZD0161800), the National Natural Science Foundation of China under Grant 62271203, AI-Empowered Research Paradigm Reform and Discipline Leap Plan under Grant 2024AI01012 and the Open Research Fund of KLATASDS-MOE, ECNU.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2001", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing, 5(4-5):185-196, 1993. 3", + "[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 6", + "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):933-948, 2019. 1, 2", + "[4] Xianhang Cheng and Zhenzhong Chen. Multiple video frame interpolation via enhanced deformable separable convolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7029-7045, 2021. 2", + "[5] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 2", + "[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 2", + "[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 39(4):75-1, 2020. 7", + "[8] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 2", + "[9] Akash Gupta, Abhishek Aich, and Amit K Roy-Chowdhury. Alanet: Adaptive latent attention network for joint video deblurring and interpolation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 256-264, 2020. 2, 6", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2", + "[11] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 2, 6, 7", + "[12] Ping Hu, Simon Niklaus, Stan Sclaroff, and Kate Saenko. Many-to-many splatting for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3553-3562, 2022. 1, 2" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2, 5", + "[14] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[15] Zhilin Huang, Yijie Yu, Ling Yang, Chujun Qin, Bing Zheng, Xiawu Zheng, Zikun Zhou, Yaowei Wang, and Wenming Yang. Motion-aware latent diffusion models for video frame interpolation. arXiv preprint arXiv:2404.13534, 2024. 2", + "[16] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 169-184. Springer, 2020. 7", + "[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 1", + "[18] Junhwa Hur and Stefan Roth. Self-supervised multi-frame monocular scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2684-2694, 2021. 7", + "[19] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2", + "[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 1", + "[21] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8112-8121, 2019. 2, 6", + "[22] Yoshihiko Kuroki, Tomohiro Nishi, Seiji Kobayashi, Hideki Oyaizu, and Shinichi Yoshimura. A psychophysical study of improvements in motion-image quality by using high frame rates. Journal of the Society for Information Display, 15(1): 61-68, 2007. 1", + "[23] Yoshihiko Kuroki, Haruo Takahashi, Masahiro Kusakabe, and Ken-ichi Yamakoshi. Effects of motion image stimuli with normal and high frame rates on eeg power spectra: comparison with continuous motion image stimuli. Journal of the Society for Information Display, 22(4):191-198, 2014. 1", + "[24] Danyeong Lee, Dohoon Lee, Dongmin Bang, and Sun Kim. Disco: Diffusion schrödinger bridge for molecular conformer optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13365-13373, 2024. 2", + "[25] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive col" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2002", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "laboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2", + "[26] Pengcheng Lei, Zaoming Yan, Tingting Wang, Faming Fang, and Guixu Zhang. Three-stage temporal deformable network for blurry video frame interpolation. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2", + "[27] Meiqin Liu, Chenming Xu, Chao Yao, Chunyu Lin, and Yao Zhao. Jnmr: Joint non-linear motion regression for video frame interpolation. IEEE Transactions on Image Processing, 2023. 2, 3, 6", + "[28] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 41-56. Springer, 2020. 2, 5, 6, 7", + "[29] Yu-Lun Liu, Yi-Tung Liao, Yen-Yu Lin, and Yung-Yu Chuang. Deep video frame interpolation using cyclic frame generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8794-8802, 2019. 1, 2", + "[30] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 1, 2", + "[31] Zonglin Lyu, Ming Li, Jianbo Jiao, and Chen Chen. Frame interpolation with consecutive brownian bridge diffusion. arXiv preprint arXiv:2405.05953, 2024. 2", + "[32] Simone Meyer, Oliver Wang, Henning Zimmer, Max Grosse, and Alexander Sorkine-Hornung. Phase-based frame interpolation for video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1410-1418, 2015. 1, 2", + "[33] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5", + "[34] Simon Niklaus and Feng Liu. Context-aware synthesis for video frame interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2018. 1", + "[35] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 261-270, 2017. 2", + "[36] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, ECCV 2022, pages 1-17. European Conference on Computer Vision, 2022. 2, 3, 5, 6", + "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[38] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 2, 3, 4, 5, 6", + "[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing, 30:277-292, 2020. 2, 3, 5", + "[40] Zhihao Shi, Xiaohong Liu, Kangdi Shi, Linhui Dai, and Jun Chen. Video frame interpolation via generalized deformable convolution. IEEE Transactions on Multimedia, 24:426-439, 2021. 2", + "[41] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVfi: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14489-14498, 2021. 7", + "[42] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 5", + "[43] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1", + "[44] Kristopher Tapp. Differential geometry of curves and surfaces. Springer, 2016. 3", + "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1", + "[46] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2, 5", + "[47] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2", + "[48] Wen Wang, Qiuyu Wang, Kecheng Zheng, Hao Ouyang, Zhekai Chen, Biao Gong, Hao Chen, Yujun Shen, and Chunhua Shen. Framer: Interactive frame interpolation. arXiv preprint arXiv:2410.18978, 2024. 2", + "[49] Jiyan Wu, Chau Yuen, Ngai-Man Cheung, Junliang Chen, and Chang Wen Chen. Modeling and optimization of high frame rate video transmission over wireless networks. IEEE Transactions on Wireless Communications, 15(4): 2713-2726, 2015. 1", + "[50] Jinbo Xing, Wenbo Hu, Yuechen Zhang, and Tien-Tsin Wong. Flow-aware synthesis: A generic motion model for video frame interpolation. Computational Visual Media, 7: 393-405, 2021. 2", + "[51] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 2, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2003", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Gengshan Yang and Deva Ramanan. Upgrading optical flow to 3d scene flow through optical expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1343, 2020. 2", + "[53] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. 2024. 7", + "[54] Zhiyang Yu, Yu Zhang, Xujie Xiang, Dongqing Zou, Xijun Chen, and Jimmy S Ren. Deep bayesian video frame interpolation. In European Conference on Computer Vision, pages 144-160. Springer, 2022. 5", + "[55] Zhiyang Yu, Yu Zhang, Dongqing Zou, Xijun Chen, Jimmy S Ren, and Shunqing Ren. Range-nullspace video frame interpolation with focalized motion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22159-22168, 2023. 2", + "[56] Youjian Zhang, Chaoyue Wang, and Dacheng Tao. Video frame interpolation without temporal priors. Advances in Neural Information Processing Systems, 33:13308-13318, 2020. 2, 6", + "[57] Zhihang Zhong, Xiao Sun, Zhirong Wu, Yinqiang Zheng, Stephen Lin, and Imari Sato. Animation from blur: Multimodal blur decomposition with motion guidance. In European Conference on Computer Vision, pages 599-615. Springer, 2022. 6", + "[58] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 3, 5, 6", + "[59] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2004", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_model.json b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b322f9558fba68035342fffeb017a1016a704312 --- /dev/null +++ b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.202, + 0.233, + 0.222 + ], + "angle": 0, + "content": "Zaoming Yan" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.203, + 0.41, + 0.222 + ], + "angle": 0, + "content": "Pengcheng Lei \\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.204, + 0.566, + 0.223 + ], + "angle": 0, + "content": "Tingting Wang1" + }, + { + "type": "text", + "bbox": [ + 0.604, + 0.203, + 0.726, + 0.223 + ], + "angle": 0, + "content": "Faming Fang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.755, + 0.204, + 0.885, + 0.223 + ], + "angle": 0, + "content": "Junkang Zhang1" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.229, + 0.562, + 0.248 + ], + "angle": 0, + "content": "Yaomin Huang" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.254, + 0.571, + 0.272 + ], + "angle": 0, + "content": "Haichuan Song\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.272, + 0.825, + 0.308 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Technology, and the KLATASDS Shanghai Key Laboratory of MIP, East China Normal University, Shanghai, China." + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.31, + 0.662, + 0.325 + ], + "angle": 0, + "content": "\\(^{1}\\{\\) yan.zaoming,pengchenglei1995\\} \\(@\\) 163.com" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.327, + 0.71, + 0.342 + ], + "angle": 0, + "content": "\\(^{1}\\)ymhuang@stu.ecnu.edu.cn \\(\\mathbf{\\Pi}^{1}\\)fmfang@cs.ecnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.378, + 0.327, + 0.394 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.412, + 0.486, + 0.744 + ], + "angle": 0, + "content": "Blurry video frame interpolation (BVFI), which aims to generate high-frame-rate clear videos from low-frame-rate blurry inputs, is a challenging yet significant task in computer vision. Current state-of-the-art approaches typically rely on linear or quadratic models to estimate intermediate motion. However, these methods often overlook depth variations that occur during fast object motion, leading to changes in object size and hindering interpolation performance. This paper proposes the Differential Curves-guided Blur Video Frame Interpolation (DC-BVFI) framework, which leverages the differential curves theory to analyze and mitigate the effects of depth variations caused by object motion. Specifically, DC-BVFI consists of UBNet and MPNet. Unlike prior approaches that rely on optical flow for frame interpolation, MPNet is designed to estimate the 3D scene flow, which facilitates a more precise awareness of depth and velocity variations. Since scene flow cannot be directly inferred in the 2D frame space, UBNet is introduced to transform them into 3D point maps. Extensive experiments demonstrate that the proposed DC-BVFI framework surpasses state-of-the-art performance in simulated and real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.759, + 0.222, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Natural videos captured by consumer cameras often suffer from real-world movement complexity. Blur Video Frame Interpolation (BVFI) enhances low frame rate and motion-blurred videos into high-quality, sharp outputs between consecutive input frames by reconstructing motion," + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.376, + 0.707, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.481, + 0.68, + 0.493 + ], + "angle": 0, + "content": "Linear methods" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.376, + 0.857, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.479, + 0.85, + 0.492 + ], + "angle": 0, + "content": "Quadratic methods" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.508, + 0.702, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.611, + 0.687, + 0.623 + ], + "angle": 0, + "content": "Differential curves" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.507, + 0.853, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.613, + 0.844, + 0.625 + ], + "angle": 0, + "content": "Real-world motion" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.684 + ], + "angle": 0, + "content": "Figure 1. Challenge. Linear methods: Synthesizing intermediate frames through linear approximation. Quadratic methods: Synthesizing intermediate frames by fitting velocity variations. Differential curves: Exploring the depth variations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.763 + ], + "angle": 0, + "content": "thereby improving perception for the human visual system [22, 23]. Therefore, it is widely used for diverse applications[20, 34, 49], and it has become a significant research area in computer vision." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "BVFI methods can be broadly divided into two categories: linear methods and quadratic methods. Linear methods [3, 12, 29, 30, 32] typically employ existing motion estimators [17, 43, 45] to predict optical flow between input frames. The intermediate optical flow is obtained by linearly approximating the optical flow between input frames to synthesize the intermediate frame. However, linear methods often overlook variations in the velocity of object motion. To more accurately reconstruct complex motion, un" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.237, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.237, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1994" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "like linear methods, quadratic methods [11, 28, 50, 51, 56] leverage optical flow to explore velocity variations in videos for fitting complex motions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.139, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Unfortunately, upon scrutinizing and experimenting on the released implementations of quadratic methods[11, 27, 28, 51], we observed that they overlook the depth variations and size changes caused by the car's motion, making them ineffective for video frame interpolation. We conjecture that the reasons are twofolds. Limitations of the Quadratic Theory. The quadratic theory primarily focuses on velocity variations in the horizontal and vertical directions, overlooking the object size variations caused by depth variations in the real world. Specifically, the quadratic theory is based on two linearly independent vectors and lacks a linearly independent vector to represent depth variations. As a result, when an object's size changes, the quadratic theory fails to describe it. Limitations of the Optical Flow. The implementation of the quadratic theory relies on optical flow; however, optical flow struggles to capture motion-induced depth variations. Specifically, an intrinsic limitation of optical flow is that a nearby object that moves slowly will generate the same optical flow as a faraway object that moves fast [5, 52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.483, + 0.67 + ], + "angle": 0, + "content": "Based on the above analysis, to overcome the limitations of the quadratic theory in handling depth variations, we propose leveraging the differential curves theory to unify the modeling of depth and velocity variations in motion. Specifically, we map pixels onto slight curve segments in 3D space, with each curve defined by a unit basis and motion parameters. The unit basis represents spatial coordinates in 3D space, while the motion parameters describe the 3D displacement of the unit basis for the next time step. Differential curves provide an intuitive core idea for describing 3D spatial motion; however, its implementation presents certain challenges. The primary challenges include converting 2D video frames into 3D point map representations and accurately estimating the motion parameters between point maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.809 + ], + "angle": 0, + "content": "To convert 2D video frames into 3D point maps, we propose UBNet. Although the idea of converting video frames into 3D camera space point maps is intuitive and sensible, the task is challenging because it requires estimating camera parameters and monocular depth maps from video frames. To address this issue, UBNet leverages a general transformer architecture [46], which [37, 46, 47, 59] can estimate camera intrinsics, depth maps, and camera-space point maps from unposed frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To estimate motion parameters between point maps, we introduce MPNet. MPNet estimates motion parameters with depth awareness, but this task is challenging because it requires estimating scene flow from the desired point map to the source point map, which existing methods struggle to achieve. To address this issue, MPNet combines a method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "that upgrades the 2D optical flow obtained from the video interpolation model (IFnet [14]) to produce the camera-space scene flow. By leveraging these modules, we build a Differential Curves-guided Blurry Video Frame Interpolation (DC-BVFI) framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.333 + ], + "angle": 0, + "content": "This paper presents three key contributions: (1) To the best of our knowledge, this is the first work that introduces differential curves theory to alleviate the challenge of depth variations in blurry video multi-frame interpolation. (2) This paper proposes an explicit depth-aware video interpolation framework to overcome depth variation and synthesize intermediate video frames more effectively. The UB-Net converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. (3) DC-VFI outperforms state-of-the-art results across four benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.655, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Video multi-frame interpolation. Recently, advancements in deep learning have led to various methods for video frame interpolation. These methods can be broadly categorized into two main paradigms: reconstruction-based and generation-based." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Reconstruction-based methods are mainly categorized into two types: optical flow-based methods and kernel-based methods. Initially, optical flow-based methods introduced a linear prior[3, 12, 29, 30, 32], which were subsequently expanded to quadratic flow prior [11, 28, 51, 55] and even cubic [6] trajectory models to represent the local motion of pixels accurately. To obtain the optical flow from the intermediate frame to the input frames, [11, 13] proposed distillation strategies to obtain the optical flow from the intermediate frame to the input frames. These Kernel-based methods are implemented as separable convolutions [35], deformable convolutions [4, 25, 26, 40], and trilinear sampling kernels[30]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Methods based on Denoising Diffusion Probabilistic Models (DDPM)[10] leverage generative techniques like DDPM to fill occlusions caused by motion. These DDPM-based approaches are implemented as score-based diffusion [8, 19, 48], motion-aware diffusion [15], and Brownian bridge diffusion [24, 31]. However, most DDPM-based methods are significantly time-consuming and challenging for real-time inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Blurry video multi-frame interpolation. When the original reference frames degrade, existing VFI methods face significant challenges in recovering sharp intermediate frames and reconstructing motion trajectories. Early methods primarily employed a cascade of deblurring and interpolation techniques [9, 21]. [21, 38, 39] extract frames, which are then subsequently used to generate intermediate clear frames by jointly optimizing. Oh et al. [36] proposed a DeMFI for joint deblurring and multi-frame interpolation," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1995" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "incorporating a flow-guided attentive-correlation-based feature bolstering module and recursive boosting. The critical contribution of Zhong et al. [58] is contributing a real-world blurry video with a low frame rate dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.165, + 0.182, + 0.18 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.19, + 0.391, + 0.206 + ], + "angle": 0, + "content": "3.1. Preliminary of Differential Curves" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.484, + 0.241 + ], + "angle": 0, + "content": "According to the definition of differential curves [44], the unit basis can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.25, + 0.483, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} \\alpha (s) = r ^ {\\prime} (s), \\\\ \\beta (s) = r ^ {\\prime \\prime} (s), \\\\ \\gamma (s) = \\alpha (s) \\wedge \\beta (s), \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.396 + ], + "angle": 0, + "content": "where \\( {}^{\\prime} \\) is the derivative symbol of \\( r(s) \\), \\( {}^{\\prime\\prime} \\) is the second derivative symbol of \\( r(s) \\), and \\( \\wedge \\) is the symbol of the outer product. The \\( r(s) \\) is a slight segment curve of the point \\( s \\). The \\( \\alpha(s) \\), \\( \\beta(s) \\), and \\( \\gamma(s) \\) together represent the spatial position of \\( r(s) \\), termed the unit basis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.425 + ], + "angle": 0, + "content": "According to the definition of differential curves [44], the derivative of the unit basis can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.433, + 0.483, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} \\alpha^ {\\prime} (s) \\\\ \\beta^ {\\prime} (s) \\\\ \\gamma^ {\\prime} (s) \\end{array} \\right] = \\underbrace {\\left[ \\begin{array}{c c c} 0 & \\kappa (s) & 0 \\\\ - \\kappa (s) & 0 & \\tau (s) \\\\ 0 & - \\tau (s) & 0 \\end{array} \\right]} _ {\\text {m o t i o n p a r a m e t e r s}} \\left[ \\begin{array}{l} \\alpha (s) \\\\ \\beta (s) \\\\ \\gamma (s) \\end{array} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.62 + ], + "angle": 0, + "content": "\\(\\kappa(s)\\) and \\(\\tau(s)\\) is the curvature and torsion. The \\(\\kappa(s)\\) and \\(\\tau(s)\\) together represent the spatial displacement of \\(r(s)\\), termed the motion parameters. Essentially, the motion parameters depict the spatial displacement of \\(r(s)\\), similar to optical flow. We simplify the notation \\((\\alpha(s), \\beta(s), \\gamma(s), \\kappa(s)\\) and \\(\\tau(s))\\) to \\((\\alpha, \\beta, \\gamma, \\kappa\\) and \\(\\tau\\)) for convenience expression." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.629, + 0.289, + 0.643 + ], + "angle": 0, + "content": "3.2. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Blurry Video Frame Interpolation. In the blurry video frame interpolation task, it can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.693, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\nI _ {t} = \\mathbf {F} \\left(B _ {i}, B _ {i + \\eta}\\right), t \\in (i, i + 1, \\dots , i + \\eta), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.78 + ], + "angle": 0, + "content": "where \\( B_{i} \\) and \\( B_{i + \\eta} \\) are input frames. \\( i \\) is a current time index and \\( t \\) is an interpolated time index between \\( i \\) and \\( i + \\eta \\). The goal is to synthesize a frame \\( I_{t} \\) by inputting frames \\( B_{i}, B_{i + \\eta} \\) and an optimization function \\( (\\mathbf{F}(\\cdot)) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Blurry Video Frame Interpolation under Differential Curves. Most failures in video frame interpolation occur due to object motion, which inevitably leads to depth changes (object scale changes) between adjacent frames. To mitigate the impact of depth variation on video interpolation quality, we redefine the video interpolation problem under differential curves theory, allowing for explicit exploration of depth changes." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.902, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.286 + ], + "angle": 0, + "content": "Figure 2. Video Frame under differential curves. From the perspective of differential curves, each pixel \\( s_n \\) can be viewed as a light segment curve \\( r(s_n) \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.907, + 0.47 + ], + "angle": 0, + "content": "For the convenience of expression, we write a time sequence \\(B_{i}\\rightarrow I_{t}\\). Each frame \\(B_{i}\\) contains \\(H\\times W\\) pixels \\((s_n,n = 1,\\dots ,H\\times W)\\). From the perspective of differential curves, each pixel \\(s_n\\) can be viewed as a slight segment curve \\(r(s_{n})\\), where its spatial position denotes as \\([\\alpha (s_n),\\beta (s_n),\\gamma (s_n)]\\in \\mathbb{R}^{1\\times 3}\\). Then, the frame \\(B_{i}\\) can be presented as \\(H\\times W\\) slight segment curves, denoted as \\([\\alpha ,\\beta ,\\gamma ]_i\\in \\mathbb{R}^{H\\times W\\times 3}\\). Subsequently, the video frame interpolation process \\((\\mathbf{F}(B_i)\\to I_t)\\) can be reformulated as \\((\\mathbf{F}([\\alpha ,\\beta ,\\gamma ]_i)\\to [\\alpha ,\\beta ,\\gamma ]_t)\\). We reformulate the blurry video frame interpolation problem under the differential curves theory and write as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.476, + 0.905, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {t} = \\mathbf {F} \\left(\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {i}\\right), t \\in (i, i + 1, \\dots , i + \\eta) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.905, + 0.577 + ], + "angle": 0, + "content": "Our goal is \\( F([\\alpha, \\beta, \\gamma]_i) \\to [\\alpha, \\beta, \\gamma]_t \\). \\( \\mathbf{F}(\\cdot) \\) is an optimization function that predicts the position at time \\( t \\) from the position at time \\( i \\) through optimization algorithm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.733, + 0.599 + ], + "angle": 0, + "content": "3.3. Optimization algorithm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.906, + 0.709 + ], + "angle": 0, + "content": "A multi-stage strategy ensures the temporal continuity of motion [27]; however, current multi-stage strategies[27, 36, 38, 39] lack algorithmic convergence, making it challenging for them to converge quickly. To swiftly and directly approach the target, we propose an optimization algorithm. Following Eq. (4), the minimized objective function can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.713, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\min \\left\\| \\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {t} - \\mathbf {F} \\left(\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {i}\\right) \\right\\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.905, + 0.797 + ], + "angle": 0, + "content": "To solve the minimization Eq. (5), we utilize the gradient descent algorithm [1]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.798, + 0.905, + 0.827 + ], + "angle": 0, + "content": "Leveraging the gradient descent algorithm [1], the Eq. (5) can be optimized:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.832, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {2} = \\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {1} + \\underbrace {\\left[\\begin{array}{c c c}0&\\kappa&0\\\\- \\kappa&0&\\tau\\\\0&- \\tau&0\\end{array}\\right] _ {i} ^ {1} \\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {1}} _ {\\text {g r a d i e n t} (S F _ {i \\rightarrow t})}. \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1996" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.087, + 0.887, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.183, + 0.791, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.357, + 0.518, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.357, + 0.787, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.187, + 0.899, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.537, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed DC-BVFI. UBNet primarily converts video frame \\(I_{i}\\) into camera space point map \\([\\alpha, \\beta, \\gamma]_{i} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where 3 represents the 3D coordinate, and \\(H \\times W\\) indicates the total number of point. MPNet estimates depth-aware scene flow \\(SF_{i \\rightarrow t} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where 3 represents the 3D coordinate offset between two point maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.682 + ], + "angle": 0, + "content": "\\([\\alpha, \\beta, \\gamma]_i^1\\) denotes the spatial position of \\(H \\times W\\) slight segment curves at the \\(1^{st}\\) iteration. And \\([\\alpha, \\beta, \\gamma]_i^2\\) is the spatial position of the same slight segment curves at the \\(2^{nd}\\) iteration. \\([\\alpha, \\beta, \\gamma]_i^1\\) is updated by adding the gradient of \\([\\alpha, \\beta, \\gamma]_i^1\\). Figure 2 shows the iterative optimization process from \\(1^{st}\\) iteration to \\(2^{nd}\\) iteration (Eq. (6))." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.739 + ], + "angle": 0, + "content": "3.4. Differential Curves for Blurry Video Frame Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Although using differential curves theory to describe spatial motion is reasonable, its implementation is challenging: 1) How to convert a video frame \\( I_{i} \\) into a camera space point map \\( [\\alpha, \\beta, \\gamma]_{i} \\) and represent the spatial position. 2) How to estimate the relative displacement \\( SF \\) from an unknown point map to a source point map. And existing methods struggle to obtain the relative displacement. To this end, UBNet and MPNet are designed in DC-BVFI to address these issues. The pipeline of the proposed DC-BVFI is outlined in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.591, + 0.608, + 0.605 + ], + "angle": 0, + "content": "3.4.1. UBNet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.671 + ], + "angle": 0, + "content": "Movement inevitably causes motion blur [38] and depth change. To convert the blurred video frame \\( B_{i} \\) into a 3D camera space point map \\( [\\alpha, \\beta, \\gamma]_i \\), we designed a UBNet. More specifically, the UBNet includes two submodules:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.827 + ], + "angle": 0, + "content": "Setup 1. The function of this submodule is to deblur. Traditional methods often simply apply Direct Stacking of video frames for deblurring. However, motion usually causes the content of frames to be different, resulting in misalignment between frames. Directly Stacking (DS) misalignment frames can hinder intermediate frame restoration, which may not facilitate video deblurring. Therefore, we introduce a Discriminative Deblur (DD) strategy, as shown in Fig. 3, to mitigate the adverse effects of blurring. The operations can be expressed as follows: \\(\\{B_i\\} \\xrightarrow{\\mathrm{DD}} \\{I_i\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Setup 2. Real-world motion is a three-dimensional movement involving both depth and velocity variations. However, existing video frame interpolation methods typically focus only on velocity variations, which limits the quality of interpolation. To effectively explore three-dimensional" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1997" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.262 + ], + "angle": 0, + "content": "motion, it is essential to convert the two-dimensional pixel space into the three-dimensional camera space. To convert the two-dimensional pixel space into the three-dimensional camera space, we employ a general 3D reconstruction paradigm [46], which can estimate camera intrinsics, depth maps, and point maps in the camera space from video frames. Specifically, the coordinate of the frame is mapped to a three-dimensional coordinate \\([\\alpha, \\beta, \\gamma] \\in \\mathbb{R}^{H \\times W \\times 3}\\), where 3 represents the 3D coordinate, and \\(H \\times W\\) indicates the total number of point. The operations can be expressed as follows: \\(\\{I_i\\} \\xrightarrow{\\mathrm{ViT}} \\{[\\alpha, \\beta, \\gamma]_i\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.272, + 0.188, + 0.286 + ], + "angle": 0, + "content": "3.4.2. MPNet" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.428 + ], + "angle": 0, + "content": "In optical flow-based video interpolation methods, it is often necessary to estimate the optical flow \\(F_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 2}\\) from the frame \\(I_{i}\\) to the frame \\(I_{t}\\) for interpolation [13, 28, 51]. Similarly, in 3D camera space, it is also necessary to estimate the scene flow from the point map \\([\\alpha ,\\beta ,\\gamma ]_i\\) to the point map \\([\\alpha ,\\beta ,\\gamma ]_t\\). However, because the \\([\\alpha ,\\beta ,\\gamma ]_t\\) is unknown, existing methods cannot directly obtain the scene flow. To this end, we propose MPNet to upgrade the optical flow \\(F_{i\\rightarrow t}\\) to the scene flow \\(SF_{i\\rightarrow t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.505 + ], + "angle": 0, + "content": "Step 1. Given the frames \\(\\{I_i\\}\\), flow-based methods [13, 51, 54] require approximating the intermediate flows \\(F_{i\\rightarrow t}\\) from the perspective of the synthesized frame \\(I_{t}\\). Considering real-time performance, we employ IFNet [13] to estimate the optical flow \\(F_{i\\rightarrow t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.567 + ], + "angle": 0, + "content": "Step 2. To express positional changes between point maps, we upgrade the optical flow \\( F_{i\\rightarrow t} \\) to camera space scene flow \\( SF_{i\\rightarrow t} \\) to aware depth variations. The process is written as the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.58, + 0.483, + 0.599 + ], + "angle": 0, + "content": "\\[\nS F _ {i \\rightarrow t} = \\left(F _ {i \\rightarrow t} \\mid D _ {i \\rightarrow t}\\right) \\times K ^ {- 1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.748 + ], + "angle": 0, + "content": "where \\(D_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 1} = D_i - D_t\\) is the depth maps and the camera intrinsics \\(K\\in \\mathbb{R}^{3\\times 3}\\). We set \\(D_{t} = (1 - t)D_{i} + tD_{i + \\eta}\\), where \\(D_{i}\\) and \\(D_{i + \\eta}\\) are the depth maps of two adjacent frames. \\(t\\in \\{i,i + \\eta \\}\\) is the time index for interpolation. The \\(\\times\\) denotes matrix multiplication, while | represents concatenation. The scene flow \\(SF_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 3}\\), where 3 represents the 3D coordinate offset between two point maps. We obtain the coarse scene flow \\(SF_{i\\rightarrow t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Step 3. Blurry frames lead to rough optical flow, leading to coarse scene flow. Therefore, it is necessary to refine the coarse scene flow. To achieve this, we introduce the Local Self-Attention (LSA) and Residual Block (RB) strategies, which perform fine-grained adjustments at the pixel level to obtain refining scene flow." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.178, + 0.866 + ], + "angle": 0, + "content": "3.4.3.Warp" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "A brief explanation of the point map warping process by scene flow. Specifically, given a scene flow \\( SF_{i\\rightarrow t} \\) and a" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.108 + ], + "angle": 0, + "content": "point map \\([\\alpha, \\beta, \\gamma]_i\\). The warp process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.116, + 0.905, + 0.133 + ], + "angle": 0, + "content": "\\[\n[ \\alpha , \\beta , \\gamma ] _ {t} = [ \\alpha , \\beta , \\gamma ] _ {i} + S F _ {i \\rightarrow t}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.905, + 0.171 + ], + "angle": 0, + "content": "where \\(+\\) denotes element-wise addition and \\([\\alpha, \\beta, \\gamma]_t\\) represents the warped point map." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.906, + 0.232 + ], + "angle": 0, + "content": "To Frame. It is important to note that the warped point map represents 3D coordinates. We need to project the point map back into the 2D coordinate system. The pseudocode for warping and projecting onto the 2D frame is as follows:" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.239, + 0.691, + 0.254 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode." + }, + { + "type": "algorithm", + "bbox": [ + 0.529, + 0.266, + 0.956, + 0.407 + ], + "angle": 0, + "content": "Warp pt_t = pt + sf # point map is projected to frame b, _, h, w = pt.size() pt_t_ = torch/matmul(K, pt_t.view(b, 3, -1)) coord_t = pt_t_.div(pt_t_[:, 2:3, :])[:, 0:2, :] coord_t = coord_t.view(b, 2, h, w) grid = coord_t.transpose(1, 2).transpose(2, 3) I_t = torch.grid_sample(I_i, grid)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.424, + 0.664, + 0.438 + ], + "angle": 0, + "content": "3.5. Loss Function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.906, + 0.474 + ], + "angle": 0, + "content": "We use the mean square error (MSE) as the training loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.483, + 0.905, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| I _ {i} - I _ {i} ^ {g t} \\right\\| ^ {2} + \\left\\| I _ {t} - I _ {t} ^ {g t} \\right\\| ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.54 + ], + "angle": 0, + "content": "The variable \\( I_{t}^{gt} \\) denotes the ground truth frame. \\( I_{t} \\) denotes the output frame of the model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.552, + 0.646, + 0.569 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.61, + 0.592 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Adobe. The Adobe dataset [42] consists of 120 videos at 240 fps with a resolution of \\(1280 \\times 720\\). The blurry frames are 30 fps and downsized to \\(640 \\times 352\\) as done in [36, 38]. We follow [36, 39] to use the same 8 videos of the Adobe dataset for evaluation and other videos of the Adobe dataset for training. RBI. The real-world blurry frame interpolation [58] dataset comprises 55 videos. We use the same videos as the test set and the others as the training set [58]. The blurred image has an exposure time of \\(18~\\mathrm{ms}\\), while the sharp image has an exposure time of nearly \\(2\\mathrm{ms}\\) [58]. This means that there are 9 sharp frames corresponding to one blurred frame, and 11 sharp frames exist in the readout dead-time between adjacent blurred frames [58]. GoPro. We use the same 11 videos of the GoPro dataset [33] to test the model [38]. The test datasets are also temporally downsampled to 30 fps with the blurring as [36]. YouTube. As a test dataset, the YouTube [38] consists of 59 video clips with different scenes and 12590 frames in total [38]. Please note that the evaluation datasets are also temporally downsampled to 30 fps with the blurring as [36]." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1998" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.089, + 0.861, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on Adobe and RBI. The best and the second best results are denoted by pink and yellow." + }, + { + "type": "table", + "bbox": [ + 0.167, + 0.11, + 0.828, + 0.349 + ], + "angle": 0, + "content": "
MethodAdobeRBI
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]28.730.86528.660.86428.670.86427.040.84026.920.83827.130.840
TNTT [21]29.400.89329.450.87629.450.87628.880.85727.360.85428.110.855
DAIN [2]--29.300.877----28.380.860--
BIN [38]31.760.89628.360.84729.950.87829.010.86027.580.86128.300.861
ALANET [9]33.710.93232.980.93633.340.93528.810.86828.880.86828.840.868
UTI-VFI [56]28.730.86528.660.86428.670.86427.340.84427.120.84727.230.844
EQVI [28]--29.690.889----27.790.840--
ANIB [57]31.300.89032.830.89932.520.89428.830.88828.990.87228.900.870
DeMFI [36]33.830.93733.790.94133.790.94029.690.89929.000.89029.350.893
BiT [58]33.600.93633.860.93833.730.93729.580.89329.840.89729.750.894
JNMR [27]31.330.90731.680.93531.430.93928.520.87328.710.87928.630.875
IQ-VFI [11]33.830.94033.800.94133.790.94029.950.89429.920.89729.910.894
Ours34.170.94234.100.94234.140.94231.020.90430.890.91531.030.913
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.362, + 0.875, + 0.377 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on GoPro and YouTube. The best and the second best results are denoted by pink and yellow." + }, + { + "type": "table", + "bbox": [ + 0.166, + 0.383, + 0.825, + 0.58 + ], + "angle": 0, + "content": "
MethodYouTubeGoPro
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]29.270.88429.980.89029.990.89027.200.81526.620.810--
TNTT [21]29.590.88929.770.89029.750.89026.480.80826.680.81526.650.814
DAIN [2]--32.020.901----25.540.859--
UTI-VFI [56]28.610.87928.640.87328.640.87725.660.80825.630.84525.640.814
EQVI [28]--29.110.893----29.680.881--
DeMFI [36]32.900.92532.790.92632.800.92630.820.89330.780.90030.750.900
BiT [58]32.370.92031.880.91831.500.91531.270.90631.200.90528.950.885
JNMR [27]--32.250.922----31.020.900--
IQ-VFI [11]33.030.92932.900.93032.940.92831.320.90631.260.90431.280.904
Ours33.400.93233.320.93333.350.93232.140.91232.050.91032.080.910
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.594, + 0.312, + 0.61 + ], + "angle": 0, + "content": "4.2. Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.724 + ], + "angle": 0, + "content": "We optimize the loss using Adam in the PyTorch framework. The cosine scheduler schedules the learning rate from \\(1e - 4\\) to \\(1e - 6\\). Standard data augmentation techniques, such as flipping, rotation, and cropping, are applied to the data with a size of \\(256 \\times 256\\). We train our model on the training datasets with a batch size of 4 for 900 epochs on 4 NVIDIA 3090ti GPUs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.741, + 0.403, + 0.756 + ], + "angle": 0, + "content": "4.3. Comparison with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.794 + ], + "angle": 0, + "content": "As shown in Tab. 1, a total of 12 outperformers are compared, which are fine-tuned on the Adobe and RBI datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Results on the RBI. Our approach achieves a 1.12dB improvement over the current state-of-the-art. The RBI dataset contains a large number of depth-varying scenes, such as vehicle motion and camera movement, resulting in a high average motion magnitude [58]. DC-BVFI performs well on the RBI dataset, demonstrating that the explicit depth-aware approach is advantageous in challenging scenarios," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.905, + 0.64 + ], + "angle": 0, + "content": "such as vehicle motion, where many existing methods tend to fail. The proposed DC-BVFI consistently outperforms existing methods, achieving a PSNR of 34.14 on Adobe." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.821 + ], + "angle": 0, + "content": "To verify the generalization capability, we conducted cross-dataset experiments using YouTube and GoPro as test sets, as shown in Tab. 2. Results on the GoPro. We observed that our method achieved an impressive performance of 32.08 PSNR on the GoPro test set. The GoPro dataset contains a large number of dynamic scenes, resulting in a high average motion magnitude, making it a challenging benchmark for video frame interpolation. Our method performs well on the GoPro dataset, suggesting that DC-BVFI exhibits strong generalization across various motion scenarios. We also observed the final performance of 33.35 PSNR on the YouTube test set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.782, + 0.849 + ], + "angle": 0, + "content": "4.4. Comparison of Visual Results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We further compare the visualization results in depth-varying motion scenarios, such as car motion and camera motion. As shown in Fig. 4 our method's predictions are" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1999" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Visual comparison with different methods. The comparison primarily focuses on depth change scenarios, such as car motion and camera movement." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.35, + 0.465, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.58, + 0.416, + 0.595 + ], + "angle": 0, + "content": "Figure 5. Comparison of Motion-in-Depth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.715 + ], + "angle": 0, + "content": "closer to the ground truth, with more precise details on both car motion and camera motion. Methods like QVI [51], EQVI [28], and IQ-VFI [11] explore inter-frame motion through estimated acceleration to enhance motion reconstruction. However, they struggle to express changes in object depth during motion, often leading to suboptimal outcomes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison of Motion-in-Depth. Moreover, to quantitatively measure depth variations during motion. We employ the MiD metric [18] to evaluate the depth variations. Since the video interpolation dataset lacks ground truth depth maps, we use the depth maps estimated from clear video frames [53] as the ground truth labels. Fig. 5 reports the performance of quadratic methods on multi-frame interpolation in Adobe, RBI, YouTube, and GoPro. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of MiD. This indicates that our method effectively explores the depth variations cue to enhance video frame" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.349, + 0.692, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.46, + 0.692, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.349, + 0.868, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.46, + 0.868, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.579, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Figure 6. Comparison of temporal consistency performance under tOF metric in Adobe, RBI, YouTube, and GoPro datasets." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.613, + 0.607, + 0.627 + ], + "angle": 0, + "content": "interpolation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Comparison of Temporal Consistency. We employ the tOF metric [7, 16, 41] to evaluate the temporal consistency. Figure 6 reports the quality of motion reconstruction across several existing models on the Adobe, RBI, GoPro, and YouTube test sets. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of tOF. The outstanding temporal consistency is due to DC-BVFI's explicit depth modeling of rapid object motion, which greatly improves motion temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.68, + 0.803 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We conducted ablation experiments to evaluate the contribution of the 3D architecture and the impact of each submodule in our method, namely the DD and the LSA. Additionally, we investigated the influence of scene flow and optical flow on video frame interpolation. Furthermore, we measured the parameter count and computational complex" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2000" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 3. Ablation studies of architecture on the RBI dataset. DS stands for stacking video frames along the channel axis, while DD denotes the discriminative deblur strategy. LSA and RB represent Local Self-Attention and Residual Block, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.157, + 0.482, + 0.251 + ], + "angle": 0, + "content": "
ArchitectureBackBoneSettingUBNetMPNetRBI
DDDSLSARBPSNRMiD(↓)
3DViT-Base(a)✓✓31.0376
(b)✓✓30.9883
(c)✓✓30.6681
2DViT-Base(d)✓✓28.91137
" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.263, + 0.484, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.485, + 0.445 + ], + "angle": 0, + "content": "Figure 7. Visualization of the architecture ablation experiments. In the scene with car depth variations, our architecture achieves promising results." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.462, + 0.24, + 0.476 + ], + "angle": 0, + "content": "ity of various models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Architecture. We empirically evaluated the proposed 3D architecture's effectiveness in guiding video frame interpolation in scenes with depth variations. For a fair comparison, both methods use the same backbone but differ in their output prediction heads. PSNR and MiD are presented in Tab. 3. The results of Tab. 3 and Fig. 7 indicate that the 3D architecture significantly outperforms the 2D architecture in scenes with depth variations. The results of setting (b) of Tab. 3 and Fig. 7 highlight the importance of refining scene flow. The results of setting (c) of Tab. 3 and Fig. 7 show that directly stacking misaligned frames hinders intermediate frame restoration, and may not facilitate video deblurring." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Flow. We analyzed the effects of scene flow and optical flow on video frame interpolation performance. For a fair evaluation, in the 2D architecture, we replace the depth estimation in the 3D architecture with a ViT-Base model that has the same parameter size. The results in Tab. 4 and Fig. 8 indicate that scene flow outperforms optical flow in scenarios with depth variations. This is in line with our intuitions, as scene flow integrates depth information on top of optical flow, providing a more comprehensive consideration of useful cues. In other words, when motion doesn't cause depth variations, the scene flow in the depth spatial dimension is deactivated, degrading scene flow into the optical flow. Essentially, scene flow encompasses optical flow. These additional insights contribute to improved video frame interpolation performance in scenes with depth variations." + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.09, + 0.868, + 0.103 + ], + "angle": 0, + "content": "Table 4. Ablation studies of Flow on the RBI dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.116, + 0.905, + 0.234 + ], + "angle": 0, + "content": "
ArchitectureFlowSettingFlowNetDepthRBI
IFNetRAFTMeGeDust3RPSNRMiD(↓)
3DScene Flow(a)✓✓31.0376
(b)✓✓30.9982
(c)✓✓31.0079
(d)✓✓30.8685
ArchitectureFlowIFNetRAFTViT-Base 2DPSNRMiD(↓)
2DOptical Flow(e)✓✓28.80143
(f)✓✓28.78147
" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.263, + 0.9, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.554, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 8. The results of video multi-frame interpolation are presented. The results of the 3D-based architecture (a) are shown in the top rows, and the 2D-based architecture (e) in the bottom rows." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.634, + 0.631 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.801 + ], + "angle": 0, + "content": "This is the first work to focus on solving the depth variations in video frame interpolation. We introduce the differential curves to analyze the issue of depth variations in video frame interpolation. Based on the analysis, this paper builds an explicit depth-aware video interpolation framework to overcome the issue. The UBNet converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. DC-BVFI has achieved state-of-the-art performance on four benchmark datasets: Adobe, RBI, YouTube, and GoPro." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the National Key R&D Program of China (2022ZD0161800), the National Natural Science Foundation of China under Grant 62271203, AI-Empowered Research Paradigm Reform and Discipline Leap Plan under Grant 2024AI01012 and the Open Research Fund of KLATASDS-MOE, ECNU." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "2001" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing, 5(4-5):185-196, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.146, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.482, + 0.298 + ], + "angle": 0, + "content": "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):933-948, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[4] Xianhang Cheng and Zhenzhong Chen. Multiple video frame interpolation via enhanced deformable separable convolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7029-7045, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[5] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 39(4):75-1, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[8] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[9] Akash Gupta, Abhishek Aich, and Amit K Roy-Chowdhury. Alanet: Adaptive latent attention network for joint video deblurring and interpolation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 256-264, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[11] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[12] Ping Hu, Simon Niklaus, Stan Sclaroff, and Kate Saenko. Many-to-many splatting for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3553-3562, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[13] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[15] Zhilin Huang, Yijie Yu, Ling Yang, Chujun Qin, Bing Zheng, Xiawu Zheng, Zikun Zhou, Yaowei Wang, and Wenming Yang. Motion-aware latent diffusion models for video frame interpolation. arXiv preprint arXiv:2404.13534, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[16] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 169-184. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.335, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[18] Junhwa Hur and Stefan Roth. Self-supervised multi-frame monocular scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2684-2694, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[19] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[21] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8112-8121, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[22] Yoshihiko Kuroki, Tomohiro Nishi, Seiji Kobayashi, Hideki Oyaizu, and Shinichi Yoshimura. A psychophysical study of improvements in motion-image quality by using high frame rates. Journal of the Society for Information Display, 15(1): 61-68, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[23] Yoshihiko Kuroki, Haruo Takahashi, Masahiro Kusakabe, and Ken-ichi Yamakoshi. Effects of motion image stimuli with normal and high frame rates on eeg power spectra: comparison with continuous motion image stimuli. Journal of the Society for Information Display, 22(4):191-198, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Danyeong Lee, Dohoon Lee, Dongmin Bang, and Sun Kim. Disco: Diffusion schrödinger bridge for molecular conformer optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13365-13373, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive col" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2002" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "laboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[26] Pengcheng Lei, Zaoming Yan, Tingting Wang, Faming Fang, and Guixu Zhang. Three-stage temporal deformable network for blurry video frame interpolation. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[27] Meiqin Liu, Chenming Xu, Chao Yao, Chunyu Lin, and Yao Zhao. Jnmr: Joint non-linear motion regression for video frame interpolation. IEEE Transactions on Image Processing, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[28] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 41-56. Springer, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.331, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[29] Yu-Lun Liu, Yi-Tung Liao, Yen-Yu Lin, and Yung-Yu Chuang. Deep video frame interpolation using cyclic frame generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8794-8802, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.387, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[30] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.443, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[31] Zonglin Lyu, Ming Li, Jianbo Jiao, and Chen Chen. Frame interpolation with consecutive brownian bridge diffusion. arXiv preprint arXiv:2405.05953, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[32] Simone Meyer, Oliver Wang, Henning Zimmer, Max Grosse, and Alexander Sorkine-Hornung. Phase-based frame interpolation for video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1410-1418, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[33] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[34] Simon Niklaus and Feng Liu. Context-aware synthesis for video frame interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[35] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 261-270, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[36] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, ECCV 2022, pages 1-17. European Conference on Computer Vision, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing, 30:277-292, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[40] Zhihao Shi, Xiaohong Liu, Kangdi Shi, Linhui Dai, and Jun Chen. Video frame interpolation via generalized deformable convolution. IEEE Transactions on Multimedia, 24:426-439, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[41] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVfi: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14489-14498, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[42] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[43] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[44] Kristopher Tapp. Differential geometry of curves and surfaces. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[46] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[47] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[48] Wen Wang, Qiuyu Wang, Kecheng Zheng, Hao Ouyang, Zhekai Chen, Biao Gong, Hao Chen, Yujun Shen, and Chunhua Shen. Framer: Interactive frame interpolation. arXiv preprint arXiv:2410.18978, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[49] Jiyan Wu, Chau Yuen, Ngai-Man Cheung, Junliang Chen, and Chang Wen Chen. Modeling and optimization of high frame rate video transmission over wireless networks. IEEE Transactions on Wireless Communications, 15(4): 2713-2726, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[50] Jinbo Xing, Wenbo Hu, Yuechen Zhang, and Tien-Tsin Wong. Flow-aware synthesis: A generic motion model for video frame interpolation. Computational Visual Media, 7: 393-405, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2003" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[52] Gengshan Yang and Deva Ramanan. Upgrading optical flow to 3d scene flow through optical expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1343, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[53] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[54] Zhiyang Yu, Yu Zhang, Xujie Xiang, Dongqing Zou, Xijun Chen, and Jimmy S Ren. Deep bayesian video frame interpolation. In European Conference on Computer Vision, pages 144-160. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[55] Zhiyang Yu, Yu Zhang, Dongqing Zou, Xijun Chen, Jimmy S Ren, and Shunqing Ren. Range-nullspace video frame interpolation with focalized motion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22159-22168, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[56] Youjian Zhang, Chaoyue Wang, and Dacheng Tao. Video frame interpolation without temporal priors. Advances in Neural Information Processing Systems, 33:13308-13318, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[57] Zhihang Zhong, Xiao Sun, Zhirong Wu, Yinqiang Zheng, Stephen Lin, and Imari Sato. Animation from blur: Multimodal blur decomposition with motion guidance. In European Conference on Computer Vision, pages 599-615. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[58] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[59] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2004" + } + ] +] \ No newline at end of file diff --git a/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_origin.pdf b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..200051ece236b648e974f6a2f712773a7c96dc47 --- /dev/null +++ b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/335d79cd-ff57-4d73-aab7-e080ed86c5a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95ff6493567396524132e5f82420890e6ef7377eae3c1e62311d560be6f46db0 +size 631082 diff --git a/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/full.md b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1ce64829595ab6726b3ce7fd8f8fd6d517bdf7b8 --- /dev/null +++ b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/full.md @@ -0,0 +1,354 @@ +# Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves + +Zaoming Yan + +Pengcheng Lei $^{1*}$ + +Tingting Wang1 + +Faming Fang $^{1\dagger}$ + +Junkang Zhang1 + +Yaomin Huang + +Haichuan Song $^{1\dagger}$ + +$^{1}$ School of Computer Science and Technology, and the KLATASDS Shanghai Key Laboratory of MIP, East China Normal University, Shanghai, China. + +$^{1}\{$ yan.zaoming,pengchenglei1995\} $@$ 163.com + +$^{1}$ ymhuang@stu.ecnu.edu.cn $\mathbf{\Pi}^{1}$ fmfang@cs.ecnu.edu.cn + +# Abstract + +Blurry video frame interpolation (BVFI), which aims to generate high-frame-rate clear videos from low-frame-rate blurry inputs, is a challenging yet significant task in computer vision. Current state-of-the-art approaches typically rely on linear or quadratic models to estimate intermediate motion. However, these methods often overlook depth variations that occur during fast object motion, leading to changes in object size and hindering interpolation performance. This paper proposes the Differential Curves-guided Blur Video Frame Interpolation (DC-BVFI) framework, which leverages the differential curves theory to analyze and mitigate the effects of depth variations caused by object motion. Specifically, DC-BVFI consists of UBNet and MPNet. Unlike prior approaches that rely on optical flow for frame interpolation, MPNet is designed to estimate the 3D scene flow, which facilitates a more precise awareness of depth and velocity variations. Since scene flow cannot be directly inferred in the 2D frame space, UBNet is introduced to transform them into 3D point maps. Extensive experiments demonstrate that the proposed DC-BVFI framework surpasses state-of-the-art performance in simulated and real-world datasets. + +# 1. Introduction + +Natural videos captured by consumer cameras often suffer from real-world movement complexity. Blur Video Frame Interpolation (BVFI) enhances low frame rate and motion-blurred videos into high-quality, sharp outputs between consecutive input frames by reconstructing motion, + +![](images/bf69111e34782c861c9fae4961e03207c4aadd5a853c7c743b3eab07d8149535.jpg) +Linear methods + +![](images/f53272ebed1e5ad6eb4cd4cea286a7a7acb9f96b2b7bc93ac878cd95a57fa5aa.jpg) +Quadratic methods + +![](images/20d70d3e27f5afc714bf6009a54feb60ab37795b1471b51075490a4eb10e465f.jpg) +Differential curves +Figure 1. Challenge. Linear methods: Synthesizing intermediate frames through linear approximation. Quadratic methods: Synthesizing intermediate frames by fitting velocity variations. Differential curves: Exploring the depth variations. + +![](images/1c3765191ef6b11033abd73dd13a1422a93bc4488ba370b61d95891b49c4c397.jpg) +Real-world motion + +thereby improving perception for the human visual system [22, 23]. Therefore, it is widely used for diverse applications[20, 34, 49], and it has become a significant research area in computer vision. + +BVFI methods can be broadly divided into two categories: linear methods and quadratic methods. Linear methods [3, 12, 29, 30, 32] typically employ existing motion estimators [17, 43, 45] to predict optical flow between input frames. The intermediate optical flow is obtained by linearly approximating the optical flow between input frames to synthesize the intermediate frame. However, linear methods often overlook variations in the velocity of object motion. To more accurately reconstruct complex motion, un + +like linear methods, quadratic methods [11, 28, 50, 51, 56] leverage optical flow to explore velocity variations in videos for fitting complex motions. + +Unfortunately, upon scrutinizing and experimenting on the released implementations of quadratic methods[11, 27, 28, 51], we observed that they overlook the depth variations and size changes caused by the car's motion, making them ineffective for video frame interpolation. We conjecture that the reasons are twofolds. Limitations of the Quadratic Theory. The quadratic theory primarily focuses on velocity variations in the horizontal and vertical directions, overlooking the object size variations caused by depth variations in the real world. Specifically, the quadratic theory is based on two linearly independent vectors and lacks a linearly independent vector to represent depth variations. As a result, when an object's size changes, the quadratic theory fails to describe it. Limitations of the Optical Flow. The implementation of the quadratic theory relies on optical flow; however, optical flow struggles to capture motion-induced depth variations. Specifically, an intrinsic limitation of optical flow is that a nearby object that moves slowly will generate the same optical flow as a faraway object that moves fast [5, 52]. + +Based on the above analysis, to overcome the limitations of the quadratic theory in handling depth variations, we propose leveraging the differential curves theory to unify the modeling of depth and velocity variations in motion. Specifically, we map pixels onto slight curve segments in 3D space, with each curve defined by a unit basis and motion parameters. The unit basis represents spatial coordinates in 3D space, while the motion parameters describe the 3D displacement of the unit basis for the next time step. Differential curves provide an intuitive core idea for describing 3D spatial motion; however, its implementation presents certain challenges. The primary challenges include converting 2D video frames into 3D point map representations and accurately estimating the motion parameters between point maps. + +To convert 2D video frames into 3D point maps, we propose UBNet. Although the idea of converting video frames into 3D camera space point maps is intuitive and sensible, the task is challenging because it requires estimating camera parameters and monocular depth maps from video frames. To address this issue, UBNet leverages a general transformer architecture [46], which [37, 46, 47, 59] can estimate camera intrinsics, depth maps, and camera-space point maps from unposed frames. + +To estimate motion parameters between point maps, we introduce MPNet. MPNet estimates motion parameters with depth awareness, but this task is challenging because it requires estimating scene flow from the desired point map to the source point map, which existing methods struggle to achieve. To address this issue, MPNet combines a method + +that upgrades the 2D optical flow obtained from the video interpolation model (IFnet [14]) to produce the camera-space scene flow. By leveraging these modules, we build a Differential Curves-guided Blurry Video Frame Interpolation (DC-BVFI) framework. + +This paper presents three key contributions: (1) To the best of our knowledge, this is the first work that introduces differential curves theory to alleviate the challenge of depth variations in blurry video multi-frame interpolation. (2) This paper proposes an explicit depth-aware video interpolation framework to overcome depth variation and synthesize intermediate video frames more effectively. The UB-Net converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. (3) DC-VFI outperforms state-of-the-art results across four benchmark datasets. + +# 2. Related Work + +Video multi-frame interpolation. Recently, advancements in deep learning have led to various methods for video frame interpolation. These methods can be broadly categorized into two main paradigms: reconstruction-based and generation-based. + +Reconstruction-based methods are mainly categorized into two types: optical flow-based methods and kernel-based methods. Initially, optical flow-based methods introduced a linear prior[3, 12, 29, 30, 32], which were subsequently expanded to quadratic flow prior [11, 28, 51, 55] and even cubic [6] trajectory models to represent the local motion of pixels accurately. To obtain the optical flow from the intermediate frame to the input frames, [11, 13] proposed distillation strategies to obtain the optical flow from the intermediate frame to the input frames. These Kernel-based methods are implemented as separable convolutions [35], deformable convolutions [4, 25, 26, 40], and trilinear sampling kernels[30]. + +Methods based on Denoising Diffusion Probabilistic Models (DDPM)[10] leverage generative techniques like DDPM to fill occlusions caused by motion. These DDPM-based approaches are implemented as score-based diffusion [8, 19, 48], motion-aware diffusion [15], and Brownian bridge diffusion [24, 31]. However, most DDPM-based methods are significantly time-consuming and challenging for real-time inference. + +Blurry video multi-frame interpolation. When the original reference frames degrade, existing VFI methods face significant challenges in recovering sharp intermediate frames and reconstructing motion trajectories. Early methods primarily employed a cascade of deblurring and interpolation techniques [9, 21]. [21, 38, 39] extract frames, which are then subsequently used to generate intermediate clear frames by jointly optimizing. Oh et al. [36] proposed a DeMFI for joint deblurring and multi-frame interpolation, + +incorporating a flow-guided attentive-correlation-based feature bolstering module and recursive boosting. The critical contribution of Zhong et al. [58] is contributing a real-world blurry video with a low frame rate dataset. + +# 3. Method + +# 3.1. Preliminary of Differential Curves + +According to the definition of differential curves [44], the unit basis can be written as: + +$$ +\left\{ \begin{array}{c} \alpha (s) = r ^ {\prime} (s), \\ \beta (s) = r ^ {\prime \prime} (s), \\ \gamma (s) = \alpha (s) \wedge \beta (s), \end{array} \right. \tag {1} +$$ + +where ${}^{\prime}$ is the derivative symbol of $r(s)$ , ${}^{\prime\prime}$ is the second derivative symbol of $r(s)$ , and $\wedge$ is the symbol of the outer product. The $r(s)$ is a slight segment curve of the point $s$ . The $\alpha(s)$ , $\beta(s)$ , and $\gamma(s)$ together represent the spatial position of $r(s)$ , termed the unit basis. + +According to the definition of differential curves [44], the derivative of the unit basis can be written as: + +$$ +\left[ \begin{array}{l} \alpha^ {\prime} (s) \\ \beta^ {\prime} (s) \\ \gamma^ {\prime} (s) \end{array} \right] = \underbrace {\left[ \begin{array}{c c c} 0 & \kappa (s) & 0 \\ - \kappa (s) & 0 & \tau (s) \\ 0 & - \tau (s) & 0 \end{array} \right]} _ {\text {m o t i o n p a r a m e t e r s}} \left[ \begin{array}{l} \alpha (s) \\ \beta (s) \\ \gamma (s) \end{array} \right]. \tag {2} +$$ + +$\kappa(s)$ and $\tau(s)$ is the curvature and torsion. The $\kappa(s)$ and $\tau(s)$ together represent the spatial displacement of $r(s)$ , termed the motion parameters. Essentially, the motion parameters depict the spatial displacement of $r(s)$ , similar to optical flow. We simplify the notation $(\alpha(s), \beta(s), \gamma(s), \kappa(s)$ and $\tau(s))$ to $(\alpha, \beta, \gamma, \kappa$ and $\tau$ ) for convenience expression. + +# 3.2. Problem formulation + +Blurry Video Frame Interpolation. In the blurry video frame interpolation task, it can be written as + +$$ +I _ {t} = \mathbf {F} \left(B _ {i}, B _ {i + \eta}\right), t \in (i, i + 1, \dots , i + \eta), \tag {3} +$$ + +where $B_{i}$ and $B_{i + \eta}$ are input frames. $i$ is a current time index and $t$ is an interpolated time index between $i$ and $i + \eta$ . The goal is to synthesize a frame $I_{t}$ by inputting frames $B_{i}, B_{i + \eta}$ and an optimization function $(\mathbf{F}(\cdot))$ . + +Blurry Video Frame Interpolation under Differential Curves. Most failures in video frame interpolation occur due to object motion, which inevitably leads to depth changes (object scale changes) between adjacent frames. To mitigate the impact of depth variation on video interpolation quality, we redefine the video interpolation problem under differential curves theory, allowing for explicit exploration of depth changes. + +![](images/da97134077be8d4cdb2d2743b9204778cf0026f1454d2b66f245a4104ec1977e.jpg) +Figure 2. Video Frame under differential curves. From the perspective of differential curves, each pixel $s_n$ can be viewed as a light segment curve $r(s_n)$ . + +For the convenience of expression, we write a time sequence $B_{i}\rightarrow I_{t}$ . Each frame $B_{i}$ contains $H\times W$ pixels $(s_n,n = 1,\dots ,H\times W)$ . From the perspective of differential curves, each pixel $s_n$ can be viewed as a slight segment curve $r(s_{n})$ , where its spatial position denotes as $[\alpha (s_n),\beta (s_n),\gamma (s_n)]\in \mathbb{R}^{1\times 3}$ . Then, the frame $B_{i}$ can be presented as $H\times W$ slight segment curves, denoted as $[\alpha ,\beta ,\gamma ]_i\in \mathbb{R}^{H\times W\times 3}$ . Subsequently, the video frame interpolation process $(\mathbf{F}(B_i)\to I_t)$ can be reformulated as $(\mathbf{F}([\alpha ,\beta ,\gamma ]_i)\to [\alpha ,\beta ,\gamma ]_t)$ . We reformulate the blurry video frame interpolation problem under the differential curves theory and write as follows: + +$$ +\left[ \begin{array}{l} \alpha \\ \beta \\ \gamma \end{array} \right] _ {t} = \mathbf {F} \left(\left[ \begin{array}{l} \alpha \\ \beta \\ \gamma \end{array} \right] _ {i}\right), t \in (i, i + 1, \dots , i + \eta) \tag {4} +$$ + +Our goal is $F([\alpha, \beta, \gamma]_i) \to [\alpha, \beta, \gamma]_t$ . $\mathbf{F}(\cdot)$ is an optimization function that predicts the position at time $t$ from the position at time $i$ through optimization algorithm. + +# 3.3. Optimization algorithm + +A multi-stage strategy ensures the temporal continuity of motion [27]; however, current multi-stage strategies[27, 36, 38, 39] lack algorithmic convergence, making it challenging for them to converge quickly. To swiftly and directly approach the target, we propose an optimization algorithm. Following Eq. (4), the minimized objective function can be written as + +$$ +\min \left\| \left[ \begin{array}{l} \alpha \\ \beta \\ \gamma \end{array} \right] _ {t} - \mathbf {F} \left(\left[ \begin{array}{l} \alpha \\ \beta \\ \gamma \end{array} \right] _ {i}\right) \right\| _ {2} ^ {2}. \tag {5} +$$ + +To solve the minimization Eq. (5), we utilize the gradient descent algorithm [1]. + +Leveraging the gradient descent algorithm [1], the Eq. (5) can be optimized: + +$$ +\left[\begin{array}{l}\alpha\\\beta\\\gamma\end{array}\right] _ {i} ^ {2} = \left[\begin{array}{l}\alpha\\\beta\\\gamma\end{array}\right] _ {i} ^ {1} + \underbrace {\left[\begin{array}{c c c}0&\kappa&0\\- \kappa&0&\tau\\0&- \tau&0\end{array}\right] _ {i} ^ {1} \left[\begin{array}{l}\alpha\\\beta\\\gamma\end{array}\right] _ {i} ^ {1}} _ {\text {g r a d i e n t} (S F _ {i \rightarrow t})}. \tag {6} +$$ + +![](images/0d08a29bb02c176fedb50a7341e66ee2cf5eec772b665fbfe60b07b1bca64030.jpg) + +![](images/34ec66a6da5808b5993c12f89c21a389ea7dc04e28586835374d0c5395c46270.jpg) + +![](images/50c7f933084fda398affecbaa3a9ba60d21938ae8e6e2d57fa76470479d87db4.jpg) +Figure 3. Overview of the proposed DC-BVFI. UBNet primarily converts video frame $I_{i}$ into camera space point map $[\alpha, \beta, \gamma]_{i} \in \mathbb{R}^{H \times W \times 3}$ , where 3 represents the 3D coordinate, and $H \times W$ indicates the total number of point. MPNet estimates depth-aware scene flow $SF_{i \rightarrow t} \in \mathbb{R}^{H \times W \times 3}$ , where 3 represents the 3D coordinate offset between two point maps. + +![](images/b3eb326dd75e3a1cb8a34044643e9ca3da395446166a8de567b1b43f70ef63d5.jpg) + +![](images/b0b43f99a91f4d85facba09a2986b5f19ce90c558770345542c747998fc84812.jpg) + +$[\alpha, \beta, \gamma]_i^1$ denotes the spatial position of $H \times W$ slight segment curves at the $1^{st}$ iteration. And $[\alpha, \beta, \gamma]_i^2$ is the spatial position of the same slight segment curves at the $2^{nd}$ iteration. $[\alpha, \beta, \gamma]_i^1$ is updated by adding the gradient of $[\alpha, \beta, \gamma]_i^1$ . Figure 2 shows the iterative optimization process from $1^{st}$ iteration to $2^{nd}$ iteration (Eq. (6)). + +# 3.4. Differential Curves for Blurry Video Frame Interpolation + +Although using differential curves theory to describe spatial motion is reasonable, its implementation is challenging: 1) How to convert a video frame $I_{i}$ into a camera space point map $[\alpha, \beta, \gamma]_{i}$ and represent the spatial position. 2) How to estimate the relative displacement $SF$ from an unknown point map to a source point map. And existing methods struggle to obtain the relative displacement. To this end, UBNet and MPNet are designed in DC-BVFI to address these issues. The pipeline of the proposed DC-BVFI is outlined in Fig. 3. + +# 3.4.1. UBNet + +Movement inevitably causes motion blur [38] and depth change. To convert the blurred video frame $B_{i}$ into a 3D camera space point map $[\alpha, \beta, \gamma]_i$ , we designed a UBNet. More specifically, the UBNet includes two submodules: + +Setup 1. The function of this submodule is to deblur. Traditional methods often simply apply Direct Stacking of video frames for deblurring. However, motion usually causes the content of frames to be different, resulting in misalignment between frames. Directly Stacking (DS) misalignment frames can hinder intermediate frame restoration, which may not facilitate video deblurring. Therefore, we introduce a Discriminative Deblur (DD) strategy, as shown in Fig. 3, to mitigate the adverse effects of blurring. The operations can be expressed as follows: $\{B_i\} \xrightarrow{\mathrm{DD}} \{I_i\}$ . + +Setup 2. Real-world motion is a three-dimensional movement involving both depth and velocity variations. However, existing video frame interpolation methods typically focus only on velocity variations, which limits the quality of interpolation. To effectively explore three-dimensional + +motion, it is essential to convert the two-dimensional pixel space into the three-dimensional camera space. To convert the two-dimensional pixel space into the three-dimensional camera space, we employ a general 3D reconstruction paradigm [46], which can estimate camera intrinsics, depth maps, and point maps in the camera space from video frames. Specifically, the coordinate of the frame is mapped to a three-dimensional coordinate $[\alpha, \beta, \gamma] \in \mathbb{R}^{H \times W \times 3}$ , where 3 represents the 3D coordinate, and $H \times W$ indicates the total number of point. The operations can be expressed as follows: $\{I_i\} \xrightarrow{\mathrm{ViT}} \{[\alpha, \beta, \gamma]_i\}$ . + +# 3.4.2. MPNet + +In optical flow-based video interpolation methods, it is often necessary to estimate the optical flow $F_{i\rightarrow t}\in \mathbb{R}^{H\times W\times 2}$ from the frame $I_{i}$ to the frame $I_{t}$ for interpolation [13, 28, 51]. Similarly, in 3D camera space, it is also necessary to estimate the scene flow from the point map $[\alpha ,\beta ,\gamma ]_i$ to the point map $[\alpha ,\beta ,\gamma ]_t$ . However, because the $[\alpha ,\beta ,\gamma ]_t$ is unknown, existing methods cannot directly obtain the scene flow. To this end, we propose MPNet to upgrade the optical flow $F_{i\rightarrow t}$ to the scene flow $SF_{i\rightarrow t}$ . + +Step 1. Given the frames $\{I_i\}$ , flow-based methods [13, 51, 54] require approximating the intermediate flows $F_{i\rightarrow t}$ from the perspective of the synthesized frame $I_{t}$ . Considering real-time performance, we employ IFNet [13] to estimate the optical flow $F_{i\rightarrow t}$ . + +Step 2. To express positional changes between point maps, we upgrade the optical flow $F_{i\rightarrow t}$ to camera space scene flow $SF_{i\rightarrow t}$ to aware depth variations. The process is written as the following formula: + +$$ +S F _ {i \rightarrow t} = \left(F _ {i \rightarrow t} \mid D _ {i \rightarrow t}\right) \times K ^ {- 1}, \tag {7} +$$ + +where $D_{i\rightarrow t}\in \mathbb{R}^{H\times W\times 1} = D_i - D_t$ is the depth maps and the camera intrinsics $K\in \mathbb{R}^{3\times 3}$ . We set $D_{t} = (1 - t)D_{i} + tD_{i + \eta}$ , where $D_{i}$ and $D_{i + \eta}$ are the depth maps of two adjacent frames. $t\in \{i,i + \eta \}$ is the time index for interpolation. The $\times$ denotes matrix multiplication, while | represents concatenation. The scene flow $SF_{i\rightarrow t}\in \mathbb{R}^{H\times W\times 3}$ , where 3 represents the 3D coordinate offset between two point maps. We obtain the coarse scene flow $SF_{i\rightarrow t}$ . + +Step 3. Blurry frames lead to rough optical flow, leading to coarse scene flow. Therefore, it is necessary to refine the coarse scene flow. To achieve this, we introduce the Local Self-Attention (LSA) and Residual Block (RB) strategies, which perform fine-grained adjustments at the pixel level to obtain refining scene flow. + +# 3.4.3.Warp + +A brief explanation of the point map warping process by scene flow. Specifically, given a scene flow $SF_{i\rightarrow t}$ and a + +point map $[\alpha, \beta, \gamma]_i$ . The warp process can be expressed as: + +$$ +[ \alpha , \beta , \gamma ] _ {t} = [ \alpha , \beta , \gamma ] _ {i} + S F _ {i \rightarrow t}, \tag {8} +$$ + +where $+$ denotes element-wise addition and $[\alpha, \beta, \gamma]_t$ represents the warped point map. + +To Frame. It is important to note that the warped point map represents 3D coordinates. We need to project the point map back into the 2D coordinate system. The pseudocode for warping and projecting onto the 2D frame is as follows: + +Algorithm 1 Pseudocode. +Warp pt_t = pt + sf # point map is projected to frame b, _, h, w = pt.size() pt_t_ = torch/matmul(K, pt_t.view(b, 3, -1)) coord_t = pt_t_.div(pt_t_[:, 2:3, :])[:, 0:2, :] coord_t = coord_t.view(b, 2, h, w) grid = coord_t.transpose(1, 2).transpose(2, 3) I_t = torch.grid_sample(I_i, grid) + +# 3.5. Loss Function. + +We use the mean square error (MSE) as the training loss function: + +$$ +\mathcal {L} = \left\| I _ {i} - I _ {i} ^ {g t} \right\| ^ {2} + \left\| I _ {t} - I _ {t} ^ {g t} \right\| ^ {2} \tag {9} +$$ + +The variable $I_{t}^{gt}$ denotes the ground truth frame. $I_{t}$ denotes the output frame of the model. + +# 4. Experiments + +# 4.1. Dataset + +Adobe. The Adobe dataset [42] consists of 120 videos at 240 fps with a resolution of $1280 \times 720$ . The blurry frames are 30 fps and downsized to $640 \times 352$ as done in [36, 38]. We follow [36, 39] to use the same 8 videos of the Adobe dataset for evaluation and other videos of the Adobe dataset for training. RBI. The real-world blurry frame interpolation [58] dataset comprises 55 videos. We use the same videos as the test set and the others as the training set [58]. The blurred image has an exposure time of $18~\mathrm{ms}$ , while the sharp image has an exposure time of nearly $2\mathrm{ms}$ [58]. This means that there are 9 sharp frames corresponding to one blurred frame, and 11 sharp frames exist in the readout dead-time between adjacent blurred frames [58]. GoPro. We use the same 11 videos of the GoPro dataset [33] to test the model [38]. The test datasets are also temporally downsampled to 30 fps with the blurring as [36]. YouTube. As a test dataset, the YouTube [38] consists of 59 video clips with different scenes and 12590 frames in total [38]. Please note that the evaluation datasets are also temporally downsampled to 30 fps with the blurring as [36]. + +Table 1. Quantitative comparisons on Adobe and RBI. The best and the second best results are denoted by pink and yellow. + +
MethodAdobeRBI
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]28.730.86528.660.86428.670.86427.040.84026.920.83827.130.840
TNTT [21]29.400.89329.450.87629.450.87628.880.85727.360.85428.110.855
DAIN [2]--29.300.877----28.380.860--
BIN [38]31.760.89628.360.84729.950.87829.010.86027.580.86128.300.861
ALANET [9]33.710.93232.980.93633.340.93528.810.86828.880.86828.840.868
UTI-VFI [56]28.730.86528.660.86428.670.86427.340.84427.120.84727.230.844
EQVI [28]--29.690.889----27.790.840--
ANIB [57]31.300.89032.830.89932.520.89428.830.88828.990.87228.900.870
DeMFI [36]33.830.93733.790.94133.790.94029.690.89929.000.89029.350.893
BiT [58]33.600.93633.860.93833.730.93729.580.89329.840.89729.750.894
JNMR [27]31.330.90731.680.93531.430.93928.520.87328.710.87928.630.875
IQ-VFI [11]33.830.94033.800.94133.790.94029.950.89429.920.89729.910.894
Ours34.170.94234.100.94234.140.94231.020.90430.890.91531.030.913
+ +Table 2. Quantitative comparisons on GoPro and YouTube. The best and the second best results are denoted by pink and yellow. + +
MethodYouTubeGoPro
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]29.270.88429.980.89029.990.89027.200.81526.620.810--
TNTT [21]29.590.88929.770.89029.750.89026.480.80826.680.81526.650.814
DAIN [2]--32.020.901----25.540.859--
UTI-VFI [56]28.610.87928.640.87328.640.87725.660.80825.630.84525.640.814
EQVI [28]--29.110.893----29.680.881--
DeMFI [36]32.900.92532.790.92632.800.92630.820.89330.780.90030.750.900
BiT [58]32.370.92031.880.91831.500.91531.270.90631.200.90528.950.885
JNMR [27]--32.250.922----31.020.900--
IQ-VFI [11]33.030.92932.900.93032.940.92831.320.90631.260.90431.280.904
Ours33.400.93233.320.93333.350.93232.140.91232.050.91032.080.910
+ +# 4.2. Implementation Details. + +We optimize the loss using Adam in the PyTorch framework. The cosine scheduler schedules the learning rate from $1e - 4$ to $1e - 6$ . Standard data augmentation techniques, such as flipping, rotation, and cropping, are applied to the data with a size of $256 \times 256$ . We train our model on the training datasets with a batch size of 4 for 900 epochs on 4 NVIDIA 3090ti GPUs. + +# 4.3. Comparison with Previous Methods + +As shown in Tab. 1, a total of 12 outperformers are compared, which are fine-tuned on the Adobe and RBI datasets. + +Results on the RBI. Our approach achieves a 1.12dB improvement over the current state-of-the-art. The RBI dataset contains a large number of depth-varying scenes, such as vehicle motion and camera movement, resulting in a high average motion magnitude [58]. DC-BVFI performs well on the RBI dataset, demonstrating that the explicit depth-aware approach is advantageous in challenging scenarios, + +such as vehicle motion, where many existing methods tend to fail. The proposed DC-BVFI consistently outperforms existing methods, achieving a PSNR of 34.14 on Adobe. + +To verify the generalization capability, we conducted cross-dataset experiments using YouTube and GoPro as test sets, as shown in Tab. 2. Results on the GoPro. We observed that our method achieved an impressive performance of 32.08 PSNR on the GoPro test set. The GoPro dataset contains a large number of dynamic scenes, resulting in a high average motion magnitude, making it a challenging benchmark for video frame interpolation. Our method performs well on the GoPro dataset, suggesting that DC-BVFI exhibits strong generalization across various motion scenarios. We also observed the final performance of 33.35 PSNR on the YouTube test set. + +# 4.4. Comparison of Visual Results. + +We further compare the visualization results in depth-varying motion scenarios, such as car motion and camera motion. As shown in Fig. 4 our method's predictions are + +![](images/340315f5ee2f3dcf609005d75cc80d76c1a9f891e9fb83cecef6cf7006f4d47a.jpg) +Figure 4. Visual comparison with different methods. The comparison primarily focuses on depth change scenarios, such as car motion and camera movement. + +![](images/2367541d7718110ebb4a26fdfa7db0877bf03450b71f2186e4b0b762b95b7cbd.jpg) +Figure 5. Comparison of Motion-in-Depth. + +closer to the ground truth, with more precise details on both car motion and camera motion. Methods like QVI [51], EQVI [28], and IQ-VFI [11] explore inter-frame motion through estimated acceleration to enhance motion reconstruction. However, they struggle to express changes in object depth during motion, often leading to suboptimal outcomes. + +Comparison of Motion-in-Depth. Moreover, to quantitatively measure depth variations during motion. We employ the MiD metric [18] to evaluate the depth variations. Since the video interpolation dataset lacks ground truth depth maps, we use the depth maps estimated from clear video frames [53] as the ground truth labels. Fig. 5 reports the performance of quadratic methods on multi-frame interpolation in Adobe, RBI, YouTube, and GoPro. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of MiD. This indicates that our method effectively explores the depth variations cue to enhance video frame + +![](images/36d03ef72aa0b1895d518ca52e42780da7e3fa1af4beeabf0eb2d7ca989e5030.jpg) + +![](images/6f17612969cad5b6966540cf8451e926a7dbc13e1e78d4837f3e197dd7d59351.jpg) +Figure 6. Comparison of temporal consistency performance under tOF metric in Adobe, RBI, YouTube, and GoPro datasets. + +![](images/a69dc98d0ad741ae7c4855c414cf549f719d6bc6f9c6cda3de77e8d8eaecb9d7.jpg) + +![](images/39c8376a3ee967108d799b9b896a63dca7d612a0a2f1c56168edfd12a950eeb2.jpg) + +interpolation. + +Comparison of Temporal Consistency. We employ the tOF metric [7, 16, 41] to evaluate the temporal consistency. Figure 6 reports the quality of motion reconstruction across several existing models on the Adobe, RBI, GoPro, and YouTube test sets. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of tOF. The outstanding temporal consistency is due to DC-BVFI's explicit depth modeling of rapid object motion, which greatly improves motion temporal consistency. + +# 4.5. Ablation Studies + +We conducted ablation experiments to evaluate the contribution of the 3D architecture and the impact of each submodule in our method, namely the DD and the LSA. Additionally, we investigated the influence of scene flow and optical flow on video frame interpolation. Furthermore, we measured the parameter count and computational complex + +Table 3. Ablation studies of architecture on the RBI dataset. DS stands for stacking video frames along the channel axis, while DD denotes the discriminative deblur strategy. LSA and RB represent Local Self-Attention and Residual Block, respectively. + +
ArchitectureBackBoneSettingUBNetMPNetRBI
DDDSLSARBPSNRMiD(↓)
3DViT-Base(a)✓✓31.0376
(b)✓✓30.9883
(c)✓✓30.6681
2DViT-Base(d)✓✓28.91137
+ +![](images/435732997026487a3b6bd0cc13f36f975e19bd890c0700f84df1c9572914ff79.jpg) +Figure 7. Visualization of the architecture ablation experiments. In the scene with car depth variations, our architecture achieves promising results. + +ity of various models. + +Architecture. We empirically evaluated the proposed 3D architecture's effectiveness in guiding video frame interpolation in scenes with depth variations. For a fair comparison, both methods use the same backbone but differ in their output prediction heads. PSNR and MiD are presented in Tab. 3. The results of Tab. 3 and Fig. 7 indicate that the 3D architecture significantly outperforms the 2D architecture in scenes with depth variations. The results of setting (b) of Tab. 3 and Fig. 7 highlight the importance of refining scene flow. The results of setting (c) of Tab. 3 and Fig. 7 show that directly stacking misaligned frames hinders intermediate frame restoration, and may not facilitate video deblurring. + +Flow. We analyzed the effects of scene flow and optical flow on video frame interpolation performance. For a fair evaluation, in the 2D architecture, we replace the depth estimation in the 3D architecture with a ViT-Base model that has the same parameter size. The results in Tab. 4 and Fig. 8 indicate that scene flow outperforms optical flow in scenarios with depth variations. This is in line with our intuitions, as scene flow integrates depth information on top of optical flow, providing a more comprehensive consideration of useful cues. In other words, when motion doesn't cause depth variations, the scene flow in the depth spatial dimension is deactivated, degrading scene flow into the optical flow. Essentially, scene flow encompasses optical flow. These additional insights contribute to improved video frame interpolation performance in scenes with depth variations. + +Table 4. Ablation studies of Flow on the RBI dataset. + +
ArchitectureFlowSettingFlowNetDepthRBI
IFNetRAFTMeGeDust3RPSNRMiD(↓)
3DScene Flow(a)✓✓31.0376
(b)✓✓30.9982
(c)✓✓31.0079
(d)✓✓30.8685
ArchitectureFlowIFNetRAFTViT-Base 2DPSNRMiD(↓)
2DOptical Flow(e)✓✓28.80143
(f)✓✓28.78147
+ +![](images/58ba4783ba13072ebff9ae21e54e6b27d6755ff0444587d084b17957a3ab7510.jpg) +Figure 8. The results of video multi-frame interpolation are presented. The results of the 3D-based architecture (a) are shown in the top rows, and the 2D-based architecture (e) in the bottom rows. + +# 5. Conclusion + +This is the first work to focus on solving the depth variations in video frame interpolation. We introduce the differential curves to analyze the issue of depth variations in video frame interpolation. Based on the analysis, this paper builds an explicit depth-aware video interpolation framework to overcome the issue. The UBNet converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. DC-BVFI has achieved state-of-the-art performance on four benchmark datasets: Adobe, RBI, YouTube, and GoPro. + +Acknowledgment. This work was supported by the National Key R&D Program of China (2022ZD0161800), the National Natural Science Foundation of China under Grant 62271203, AI-Empowered Research Paradigm Reform and Discipline Leap Plan under Grant 2024AI01012 and the Open Research Fund of KLATASDS-MOE, ECNU. + +# References + +[1] Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing, 5(4-5):185-196, 1993. 3 +[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 6 +[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):933-948, 2019. 1, 2 +[4] Xianhang Cheng and Zhenzhong Chen. Multiple video frame interpolation via enhanced deformable separable convolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7029-7045, 2021. 2 +[5] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 2 +[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 2 +[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 39(4):75-1, 2020. 7 +[8] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 2 +[9] Akash Gupta, Abhishek Aich, and Amit K Roy-Chowdhury. Alanet: Adaptive latent attention network for joint video deblurring and interpolation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 256-264, 2020. 2, 6 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2 +[11] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 2, 6, 7 +[12] Ping Hu, Simon Niklaus, Stan Sclaroff, and Kate Saenko. Many-to-many splatting for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3553-3562, 2022. 1, 2 + +[13] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2, 5 +[14] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[15] Zhilin Huang, Yijie Yu, Ling Yang, Chujun Qin, Bing Zheng, Xiawu Zheng, Zikun Zhou, Yaowei Wang, and Wenming Yang. Motion-aware latent diffusion models for video frame interpolation. arXiv preprint arXiv:2404.13534, 2024. 2 +[16] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 169-184. Springer, 2020. 7 +[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 1 +[18] Junhwa Hur and Stefan Roth. Self-supervised multi-frame monocular scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2684-2694, 2021. 7 +[19] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2 +[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 1 +[21] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8112-8121, 2019. 2, 6 +[22] Yoshihiko Kuroki, Tomohiro Nishi, Seiji Kobayashi, Hideki Oyaizu, and Shinichi Yoshimura. A psychophysical study of improvements in motion-image quality by using high frame rates. Journal of the Society for Information Display, 15(1): 61-68, 2007. 1 +[23] Yoshihiko Kuroki, Haruo Takahashi, Masahiro Kusakabe, and Ken-ichi Yamakoshi. Effects of motion image stimuli with normal and high frame rates on eeg power spectra: comparison with continuous motion image stimuli. Journal of the Society for Information Display, 22(4):191-198, 2014. 1 +[24] Danyeong Lee, Dohoon Lee, Dongmin Bang, and Sun Kim. Disco: Diffusion schrödinger bridge for molecular conformer optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13365-13373, 2024. 2 +[25] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive col + +laboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2 +[26] Pengcheng Lei, Zaoming Yan, Tingting Wang, Faming Fang, and Guixu Zhang. Three-stage temporal deformable network for blurry video frame interpolation. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2 +[27] Meiqin Liu, Chenming Xu, Chao Yao, Chunyu Lin, and Yao Zhao. Jnmr: Joint non-linear motion regression for video frame interpolation. IEEE Transactions on Image Processing, 2023. 2, 3, 6 +[28] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 41-56. Springer, 2020. 2, 5, 6, 7 +[29] Yu-Lun Liu, Yi-Tung Liao, Yen-Yu Lin, and Yung-Yu Chuang. Deep video frame interpolation using cyclic frame generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8794-8802, 2019. 1, 2 +[30] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 1, 2 +[31] Zonglin Lyu, Ming Li, Jianbo Jiao, and Chen Chen. Frame interpolation with consecutive brownian bridge diffusion. arXiv preprint arXiv:2405.05953, 2024. 2 +[32] Simone Meyer, Oliver Wang, Henning Zimmer, Max Grosse, and Alexander Sorkine-Hornung. Phase-based frame interpolation for video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1410-1418, 2015. 1, 2 +[33] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5 +[34] Simon Niklaus and Feng Liu. Context-aware synthesis for video frame interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2018. 1 +[35] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 261-270, 2017. 2 +[36] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, ECCV 2022, pages 1-17. European Conference on Computer Vision, 2022. 2, 3, 5, 6 +[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[38] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 2, 3, 4, 5, 6 +[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing, 30:277-292, 2020. 2, 3, 5 +[40] Zhihao Shi, Xiaohong Liu, Kangdi Shi, Linhui Dai, and Jun Chen. Video frame interpolation via generalized deformable convolution. IEEE Transactions on Multimedia, 24:426-439, 2021. 2 +[41] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVfi: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14489-14498, 2021. 7 +[42] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 5 +[43] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1 +[44] Kristopher Tapp. Differential geometry of curves and surfaces. Springer, 2016. 3 +[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1 +[46] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2, 5 +[47] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2 +[48] Wen Wang, Qiuyu Wang, Kecheng Zheng, Hao Ouyang, Zhekai Chen, Biao Gong, Hao Chen, Yujun Shen, and Chunhua Shen. Framer: Interactive frame interpolation. arXiv preprint arXiv:2410.18978, 2024. 2 +[49] Jiyan Wu, Chau Yuen, Ngai-Man Cheung, Junliang Chen, and Chang Wen Chen. Modeling and optimization of high frame rate video transmission over wireless networks. IEEE Transactions on Wireless Communications, 15(4): 2713-2726, 2015. 1 +[50] Jinbo Xing, Wenbo Hu, Yuechen Zhang, and Tien-Tsin Wong. Flow-aware synthesis: A generic motion model for video frame interpolation. Computational Visual Media, 7: 393-405, 2021. 2 +[51] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 2, 5, 6, 7 + +[52] Gengshan Yang and Deva Ramanan. Upgrading optical flow to 3d scene flow through optical expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1343, 2020. 2 +[53] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. 2024. 7 +[54] Zhiyang Yu, Yu Zhang, Xujie Xiang, Dongqing Zou, Xijun Chen, and Jimmy S Ren. Deep bayesian video frame interpolation. In European Conference on Computer Vision, pages 144-160. Springer, 2022. 5 +[55] Zhiyang Yu, Yu Zhang, Dongqing Zou, Xijun Chen, Jimmy S Ren, and Shunqing Ren. Range-nullspace video frame interpolation with focalized motion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22159-22168, 2023. 2 +[56] Youjian Zhang, Chaoyue Wang, and Dacheng Tao. Video frame interpolation without temporal priors. Advances in Neural Information Processing Systems, 33:13308-13318, 2020. 2, 6 +[57] Zhihang Zhong, Xiao Sun, Zhirong Wu, Yinqiang Zheng, Stephen Lin, and Imari Sato. Animation from blur: Multimodal blur decomposition with motion guidance. In European Conference on Computer Vision, pages 599-615. Springer, 2022. 6 +[58] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 3, 5, 6 +[59] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 2 \ No newline at end of file diff --git a/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/images.zip b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c86f02ae777db3b898f6a4b8d568d73be98522c0 --- /dev/null +++ b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16b316f44d9a65af2226d9310f290beefb4c442655fbff00e86b599fd4ed1e74 +size 715033 diff --git a/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/layout.json b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..73ac621d7877308cdbf0337f1a2878c94ebdd36e --- /dev/null +++ b/2025/Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves/layout.json @@ -0,0 +1,9881 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "type": "text", + "content": "Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 159, + 142, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 142, + 175 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 142, + 175 + ], + "type": "text", + "content": "Zaoming Yan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 165, + 160, + 250, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 160, + 250, + 175 + ], + "spans": [ + { + "bbox": [ + 165, + 160, + 250, + 175 + ], + "type": "text", + "content": "Pengcheng Lei " + }, + { + "bbox": [ + 165, + 160, + 250, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 268, + 161, + 346, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 161, + 346, + 176 + ], + "spans": [ + { + "bbox": [ + 268, + 161, + 346, + 176 + ], + "type": "text", + "content": "Tingting Wang1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 369, + 160, + 444, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 160, + 444, + 176 + ], + "spans": [ + { + "bbox": [ + 369, + 160, + 444, + 176 + ], + "type": "text", + "content": "Faming Fang" + }, + { + "bbox": [ + 369, + 160, + 444, + 176 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 462, + 161, + 541, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 462, + 161, + 541, + 176 + ], + "spans": [ + { + "bbox": [ + 462, + 161, + 541, + 176 + ], + "type": "text", + "content": "Junkang Zhang1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 265, + 181, + 343, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 181, + 343, + 196 + ], + "spans": [ + { + "bbox": [ + 265, + 181, + 343, + 196 + ], + "type": "text", + "content": "Yaomin Huang" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 265, + 201, + 349, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 201, + 349, + 215 + ], + "spans": [ + { + "bbox": [ + 265, + 201, + 349, + 215 + ], + "type": "text", + "content": "Haichuan Song" + }, + { + "bbox": [ + 265, + 201, + 349, + 215 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 215, + 504, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 215, + 504, + 243 + ], + "spans": [ + { + "bbox": [ + 104, + 215, + 504, + 243 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 215, + 504, + 243 + ], + "type": "text", + "content": " School of Computer Science and Technology, and the KLATASDS Shanghai Key Laboratory of MIP, East China Normal University, Shanghai, China." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 197, + 245, + 405, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 245, + 405, + 257 + ], + "spans": [ + { + "bbox": [ + 197, + 245, + 405, + 257 + ], + "type": "inline_equation", + "content": "^{1}\\{" + }, + { + "bbox": [ + 197, + 245, + 405, + 257 + ], + "type": "text", + "content": " yan.zaoming,pengchenglei1995\\} " + }, + { + "bbox": [ + 197, + 245, + 405, + 257 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 197, + 245, + 405, + 257 + ], + "type": "text", + "content": " 163.com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 173, + 258, + 434, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 258, + 434, + 270 + ], + "spans": [ + { + "bbox": [ + 173, + 258, + 434, + 270 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 258, + 434, + 270 + ], + "type": "text", + "content": "ymhuang@stu.ecnu.edu.cn " + }, + { + "bbox": [ + 173, + 258, + 434, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{\\Pi}^{1}" + }, + { + "bbox": [ + 173, + 258, + 434, + 270 + ], + "type": "text", + "content": "fmfang@cs.ecnu.edu.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "spans": [ + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 326, + 297, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 326, + 297, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 297, + 589 + ], + "type": "text", + "content": "Blurry video frame interpolation (BVFI), which aims to generate high-frame-rate clear videos from low-frame-rate blurry inputs, is a challenging yet significant task in computer vision. Current state-of-the-art approaches typically rely on linear or quadratic models to estimate intermediate motion. However, these methods often overlook depth variations that occur during fast object motion, leading to changes in object size and hindering interpolation performance. This paper proposes the Differential Curves-guided Blur Video Frame Interpolation (DC-BVFI) framework, which leverages the differential curves theory to analyze and mitigate the effects of depth variations caused by object motion. Specifically, DC-BVFI consists of UBNet and MPNet. Unlike prior approaches that rely on optical flow for frame interpolation, MPNet is designed to estimate the 3D scene flow, which facilitates a more precise awareness of depth and velocity variations. Since scene flow cannot be directly inferred in the 2D frame space, UBNet is introduced to transform them into 3D point maps. Extensive experiments demonstrate that the proposed DC-BVFI framework surpasses state-of-the-art performance in simulated and real-world datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 621, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 296, + 682 + ], + "type": "text", + "content": "Natural videos captured by consumer cameras often suffer from real-world movement complexity. Blur Video Frame Interpolation (BVFI) enhances low frame rate and motion-blurred videos into high-quality, sharp outputs between consecutive input frames by reconstructing motion," + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 347, + 297, + 432, + 380 + ], + "blocks": [ + { + "bbox": [ + 347, + 297, + 432, + 380 + ], + "lines": [ + { + "bbox": [ + 347, + 297, + 432, + 380 + ], + "spans": [ + { + "bbox": [ + 347, + 297, + 432, + 380 + ], + "type": "image", + "image_path": "bf69111e34782c861c9fae4961e03207c4aadd5a853c7c743b3eab07d8149535.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 380, + 416, + 390 + ], + "lines": [ + { + "bbox": [ + 355, + 380, + 416, + 390 + ], + "spans": [ + { + "bbox": [ + 355, + 380, + 416, + 390 + ], + "type": "text", + "content": "Linear methods" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 437, + 297, + 524, + 378 + ], + "blocks": [ + { + "bbox": [ + 437, + 297, + 524, + 378 + ], + "lines": [ + { + "bbox": [ + 437, + 297, + 524, + 378 + ], + "spans": [ + { + "bbox": [ + 437, + 297, + 524, + 378 + ], + "type": "image", + "image_path": "f53272ebed1e5ad6eb4cd4cea286a7a7acb9f96b2b7bc93ac878cd95a57fa5aa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 379, + 520, + 389 + ], + "lines": [ + { + "bbox": [ + 446, + 379, + 520, + 389 + ], + "spans": [ + { + "bbox": [ + 446, + 379, + 520, + 389 + ], + "type": "text", + "content": "Quadratic methods" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 343, + 402, + 429, + 483 + ], + "blocks": [ + { + "bbox": [ + 343, + 402, + 429, + 483 + ], + "lines": [ + { + "bbox": [ + 343, + 402, + 429, + 483 + ], + "spans": [ + { + "bbox": [ + 343, + 402, + 429, + 483 + ], + "type": "image", + "image_path": "20d70d3e27f5afc714bf6009a54feb60ab37795b1471b51075490a4eb10e465f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 483, + 420, + 493 + ], + "lines": [ + { + "bbox": [ + 348, + 483, + 420, + 493 + ], + "spans": [ + { + "bbox": [ + 348, + 483, + 420, + 493 + ], + "type": "text", + "content": "Differential curves" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 497, + 555, + 541 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 541 + ], + "type": "text", + "content": "Figure 1. Challenge. Linear methods: Synthesizing intermediate frames through linear approximation. Quadratic methods: Synthesizing intermediate frames by fitting velocity variations. Differential curves: Exploring the depth variations." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 436, + 401, + 522, + 483 + ], + "blocks": [ + { + "bbox": [ + 436, + 401, + 522, + 483 + ], + "lines": [ + { + "bbox": [ + 436, + 401, + 522, + 483 + ], + "spans": [ + { + "bbox": [ + 436, + 401, + 522, + 483 + ], + "type": "image", + "image_path": "1c3765191ef6b11033abd73dd13a1422a93bc4488ba370b61d95891b49c4c397.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 485, + 516, + 495 + ], + "lines": [ + { + "bbox": [ + 443, + 485, + 516, + 495 + ], + "spans": [ + { + "bbox": [ + 443, + 485, + 516, + 495 + ], + "type": "text", + "content": "Real-world motion" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 555, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 604 + ], + "type": "text", + "content": "thereby improving perception for the human visual system [22, 23]. Therefore, it is widely used for diverse applications[20, 34, 49], and it has become a significant research area in computer vision." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 714 + ], + "type": "text", + "content": "BVFI methods can be broadly divided into two categories: linear methods and quadratic methods. Linear methods [3, 12, 29, 30, 32] typically employ existing motion estimators [17, 43, 45] to predict optical flow between input frames. The intermediate optical flow is obtained by linearly approximating the optical flow between input frames to synthesize the intermediate frame. However, linear methods often overlook variations in the velocity of object motion. To more accurately reconstruct complex motion, un" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1994" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "like linear methods, quadratic methods [11, 28, 50, 51, 56] leverage optical flow to explore velocity variations in videos for fitting complex motions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 295, + 350 + ], + "type": "text", + "content": "Unfortunately, upon scrutinizing and experimenting on the released implementations of quadratic methods[11, 27, 28, 51], we observed that they overlook the depth variations and size changes caused by the car's motion, making them ineffective for video frame interpolation. We conjecture that the reasons are twofolds. Limitations of the Quadratic Theory. The quadratic theory primarily focuses on velocity variations in the horizontal and vertical directions, overlooking the object size variations caused by depth variations in the real world. Specifically, the quadratic theory is based on two linearly independent vectors and lacks a linearly independent vector to represent depth variations. As a result, when an object's size changes, the quadratic theory fails to describe it. Limitations of the Optical Flow. The implementation of the quadratic theory relies on optical flow; however, optical flow struggles to capture motion-induced depth variations. Specifically, an intrinsic limitation of optical flow is that a nearby object that moves slowly will generate the same optical flow as a faraway object that moves fast [5, 52]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 351, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 295, + 530 + ], + "type": "text", + "content": "Based on the above analysis, to overcome the limitations of the quadratic theory in handling depth variations, we propose leveraging the differential curves theory to unify the modeling of depth and velocity variations in motion. Specifically, we map pixels onto slight curve segments in 3D space, with each curve defined by a unit basis and motion parameters. The unit basis represents spatial coordinates in 3D space, while the motion parameters describe the 3D displacement of the unit basis for the next time step. Differential curves provide an intuitive core idea for describing 3D spatial motion; however, its implementation presents certain challenges. The primary challenges include converting 2D video frames into 3D point map representations and accurately estimating the motion parameters between point maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 532, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 640 + ], + "type": "text", + "content": "To convert 2D video frames into 3D point maps, we propose UBNet. Although the idea of converting video frames into 3D camera space point maps is intuitive and sensible, the task is challenging because it requires estimating camera parameters and monocular depth maps from video frames. To address this issue, UBNet leverages a general transformer architecture [46], which [37, 46, 47, 59] can estimate camera intrinsics, depth maps, and camera-space point maps from unposed frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "To estimate motion parameters between point maps, we introduce MPNet. MPNet estimates motion parameters with depth awareness, but this task is challenging because it requires estimating scene flow from the desired point map to the source point map, which existing methods struggle to achieve. To address this issue, MPNet combines a method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "that upgrades the 2D optical flow obtained from the video interpolation model (IFnet [14]) to produce the camera-space scene flow. By leveraging these modules, we build a Differential Curves-guided Blurry Video Frame Interpolation (DC-BVFI) framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 133, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 263 + ], + "type": "text", + "content": "This paper presents three key contributions: (1) To the best of our knowledge, this is the first work that introduces differential curves theory to alleviate the challenge of depth variations in blurry video multi-frame interpolation. (2) This paper proposes an explicit depth-aware video interpolation framework to overcome depth variation and synthesize intermediate video frames more effectively. The UB-Net converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. (3) DC-VFI outperforms state-of-the-art results across four benchmark datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 294, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 353 + ], + "type": "text", + "content": "Video multi-frame interpolation. Recently, advancements in deep learning have led to various methods for video frame interpolation. These methods can be broadly categorized into two main paradigms: reconstruction-based and generation-based." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 355, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 510 + ], + "type": "text", + "content": "Reconstruction-based methods are mainly categorized into two types: optical flow-based methods and kernel-based methods. Initially, optical flow-based methods introduced a linear prior[3, 12, 29, 30, 32], which were subsequently expanded to quadratic flow prior [11, 28, 51, 55] and even cubic [6] trajectory models to represent the local motion of pixels accurately. To obtain the optical flow from the intermediate frame to the input frames, [11, 13] proposed distillation strategies to obtain the optical flow from the intermediate frame to the input frames. These Kernel-based methods are implemented as separable convolutions [35], deformable convolutions [4, 25, 26, 40], and trilinear sampling kernels[30]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "content": "Methods based on Denoising Diffusion Probabilistic Models (DDPM)[10] leverage generative techniques like DDPM to fill occlusions caused by motion. These DDPM-based approaches are implemented as score-based diffusion [8, 19, 48], motion-aware diffusion [15], and Brownian bridge diffusion [24, 31]. However, most DDPM-based methods are significantly time-consuming and challenging for real-time inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Blurry video multi-frame interpolation. When the original reference frames degrade, existing VFI methods face significant challenges in recovering sharp intermediate frames and reconstructing motion trajectories. Early methods primarily employed a cascade of deblurring and interpolation techniques [9, 21]. [21, 38, 39] extract frames, which are then subsequently used to generate intermediate clear frames by jointly optimizing. Oh et al. [36] proposed a DeMFI for joint deblurring and multi-frame interpolation," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1995" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "incorporating a flow-guided attentive-correlation-based feature bolstering module and recursive boosting. The critical contribution of Zhong et al. [58] is contributing a real-world blurry video with a low frame rate dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 130, + 111, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 130, + 111, + 142 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 111, + 142 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 150, + 239, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 239, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 239, + 163 + ], + "type": "text", + "content": "3.1. Preliminary of Differential Curves" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "type": "text", + "content": "According to the definition of differential curves [44], the unit basis can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 198, + 295, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 198, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 128, + 198, + 295, + 245 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} \\alpha (s) = r ^ {\\prime} (s), \\\\ \\beta (s) = r ^ {\\prime \\prime} (s), \\\\ \\gamma (s) = \\alpha (s) \\wedge \\beta (s), \\end{array} \\right. \\tag {1}", + "image_path": "654c283c451199ffa588d7561d651fd68a8ccfd03af69fad4a7310d2fa297123.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "{}^{\\prime}" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": " is the derivative symbol of " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "r(s)" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "{}^{\\prime\\prime}" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": " is the second derivative symbol of " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "r(s)" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\wedge" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": " is the symbol of the outer product. The " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "r(s)" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": " is a slight segment curve of the point " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\alpha(s)" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\beta(s)" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\gamma(s)" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": " together represent the spatial position of " + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "inline_equation", + "content": "r(s)" + }, + { + "bbox": [ + 55, + 253, + 296, + 313 + ], + "type": "text", + "content": ", termed the unit basis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 314, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 336 + ], + "type": "text", + "content": "According to the definition of differential curves [44], the derivative of the unit basis can be written as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 342, + 295, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 342, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 80, + 342, + 295, + 398 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} \\alpha^ {\\prime} (s) \\\\ \\beta^ {\\prime} (s) \\\\ \\gamma^ {\\prime} (s) \\end{array} \\right] = \\underbrace {\\left[ \\begin{array}{c c c} 0 & \\kappa (s) & 0 \\\\ - \\kappa (s) & 0 & \\tau (s) \\\\ 0 & - \\tau (s) & 0 \\end{array} \\right]} _ {\\text {m o t i o n p a r a m e t e r s}} \\left[ \\begin{array}{l} \\alpha (s) \\\\ \\beta (s) \\\\ \\gamma (s) \\end{array} \\right]. \\tag {2}", + "image_path": "a4c8658bcf1648a96854226b281aaceddf8e39e6140874278fe0b1f5a5ce484a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\kappa(s)" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\tau(s)" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": " is the curvature and torsion. The " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\kappa(s)" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\tau(s)" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": " together represent the spatial displacement of " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "r(s)" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": ", termed the motion parameters. Essentially, the motion parameters depict the spatial displacement of " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "r(s)" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": ", similar to optical flow. We simplify the notation " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "(\\alpha(s), \\beta(s), \\gamma(s), \\kappa(s)" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\tau(s))" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta, \\gamma, \\kappa" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 407, + 296, + 491 + ], + "type": "text", + "content": ") for convenience expression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 176, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 176, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 176, + 509 + ], + "type": "text", + "content": "3.2. Problem formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 515, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 296, + 539 + ], + "type": "text", + "content": "Blurry Video Frame Interpolation. In the blurry video frame interpolation task, it can be written as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 548, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 548, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 86, + 548, + 295, + 562 + ], + "type": "interline_equation", + "content": "I _ {t} = \\mathbf {F} \\left(B _ {i}, B _ {i + \\eta}\\right), t \\in (i, i + 1, \\dots , i + \\eta), \\tag {3}", + "image_path": "7cc0608e952fb201682dd5da452112873f1132865b578ebebf8a032199a1432a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "B_{i + \\eta}" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " are input frames. " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " is a current time index and " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " is an interpolated time index between " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "i + \\eta" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": ". The goal is to synthesize a frame " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " by inputting frames " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "B_{i}, B_{i + \\eta}" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " and an optimization function " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}(\\cdot))" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Blurry Video Frame Interpolation under Differential Curves. Most failures in video frame interpolation occur due to object motion, which inevitably leads to depth changes (object scale changes) between adjacent frames. To mitigate the impact of depth variation on video interpolation quality, we redefine the video interpolation problem under differential curves theory, allowing for explicit exploration of depth changes." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 552, + 185 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 552, + 185 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 552, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 552, + 185 + ], + "type": "image", + "image_path": "da97134077be8d4cdb2d2743b9204778cf0026f1454d2b66f245a4104ec1977e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 192, + 553, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 226 + ], + "type": "text", + "content": "Figure 2. Video Frame under differential curves. From the perspective of differential curves, each pixel " + }, + { + "bbox": [ + 313, + 192, + 553, + 226 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 313, + 192, + 553, + 226 + ], + "type": "text", + "content": " can be viewed as a light segment curve " + }, + { + "bbox": [ + 313, + 192, + 553, + 226 + ], + "type": "inline_equation", + "content": "r(s_n)" + }, + { + "bbox": [ + 313, + 192, + 553, + 226 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": "For the convenience of expression, we write a time sequence " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "B_{i}\\rightarrow I_{t}" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": ". Each frame " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": " pixels " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "(s_n,n = 1,\\dots ,H\\times W)" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": ". From the perspective of differential curves, each pixel " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": " can be viewed as a slight segment curve " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "r(s_{n})" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": ", where its spatial position denotes as " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "[\\alpha (s_n),\\beta (s_n),\\gamma (s_n)]\\in \\mathbb{R}^{1\\times 3}" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": ". Then, the frame " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": " can be presented as " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": " slight segment curves, denoted as " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "[\\alpha ,\\beta ,\\gamma ]_i\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": ". Subsequently, the video frame interpolation process " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}(B_i)\\to I_t)" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": " can be reformulated as " + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}([\\alpha ,\\beta ,\\gamma ]_i)\\to [\\alpha ,\\beta ,\\gamma ]_t)" + }, + { + "bbox": [ + 313, + 228, + 555, + 372 + ], + "type": "text", + "content": ". We reformulate the blurry video frame interpolation problem under the differential curves theory and write as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 340, + 376, + 553, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 376, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 340, + 376, + 553, + 416 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {t} = \\mathbf {F} \\left(\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {i}\\right), t \\in (i, i + 1, \\dots , i + \\eta) \\tag {4}", + "image_path": "39f3966a734848d6b6dc85a71a8ab7615ad30512de2dc31f3c4de3847526808b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": "Our goal is " + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "inline_equation", + "content": "F([\\alpha, \\beta, \\gamma]_i) \\to [\\alpha, \\beta, \\gamma]_t" + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\cdot)" + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": " is an optimization function that predicts the position at time " + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": " from the position at time " + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": " through optimization algorithm." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 461, + 448, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 448, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 448, + 474 + ], + "type": "text", + "content": "3.3. Optimization algorithm" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 479, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 561 + ], + "type": "text", + "content": "A multi-stage strategy ensures the temporal continuity of motion [27]; however, current multi-stage strategies[27, 36, 38, 39] lack algorithmic convergence, making it challenging for them to converge quickly. To swiftly and directly approach the target, we propose an optimization algorithm. Following Eq. (4), the minimized objective function can be written as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 370, + 564, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 564, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 370, + 564, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\min \\left\\| \\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {t} - \\mathbf {F} \\left(\\left[ \\begin{array}{l} \\alpha \\\\ \\beta \\\\ \\gamma \\end{array} \\right] _ {i}\\right) \\right\\| _ {2} ^ {2}. \\tag {5}", + "image_path": "112b9ffb408d73528657d97b4afb6d342e69d899924ebacb53660af4cb555cf7.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 607, + 553, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 553, + 631 + ], + "type": "text", + "content": "To solve the minimization Eq. (5), we utilize the gradient descent algorithm [1]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 632, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 553, + 654 + ], + "type": "text", + "content": "Leveraging the gradient descent algorithm [1], the Eq. (5) can be optimized:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 347, + 658, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 658, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 347, + 658, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {2} = \\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {1} + \\underbrace {\\left[\\begin{array}{c c c}0&\\kappa&0\\\\- \\kappa&0&\\tau\\\\0&- \\tau&0\\end{array}\\right] _ {i} ^ {1} \\left[\\begin{array}{l}\\alpha\\\\\\beta\\\\\\gamma\\end{array}\\right] _ {i} ^ {1}} _ {\\text {g r a d i e n t} (S F _ {i \\rightarrow t})}. \\tag {6}", + "image_path": "8917143aaec26a6c09195bc562a45f60fa49184d290f722cf2779ec259b001a8.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1996" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 68, + 542, + 140 + ], + "blocks": [ + { + "bbox": [ + 64, + 68, + 542, + 140 + ], + "lines": [ + { + "bbox": [ + 64, + 68, + 542, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 68, + 542, + 140 + ], + "type": "image", + "image_path": "0d08a29bb02c176fedb50a7341e66ee2cf5eec772b665fbfe60b07b1bca64030.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 144, + 484, + 281 + ], + "blocks": [ + { + "bbox": [ + 59, + 144, + 484, + 281 + ], + "lines": [ + { + "bbox": [ + 59, + 144, + 484, + 281 + ], + "spans": [ + { + "bbox": [ + 59, + 144, + 484, + 281 + ], + "type": "image", + "image_path": "34ec66a6da5808b5993c12f89c21a389ea7dc04e28586835374d0c5395c46270.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 282, + 317, + 414 + ], + "blocks": [ + { + "bbox": [ + 58, + 282, + 317, + 414 + ], + "lines": [ + { + "bbox": [ + 58, + 282, + 317, + 414 + ], + "spans": [ + { + "bbox": [ + 58, + 282, + 317, + 414 + ], + "type": "image", + "image_path": "50c7f933084fda398affecbaa3a9ba60d21938ae8e6e2d57fa76470479d87db4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed DC-BVFI. UBNet primarily converts video frame " + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "text", + "content": " into camera space point map " + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_{i} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "text", + "content": ", where 3 represents the 3D coordinate, and " + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "text", + "content": " indicates the total number of point. MPNet estimates depth-aware scene flow " + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "inline_equation", + "content": "SF_{i \\rightarrow t} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 425, + 555, + 460 + ], + "type": "text", + "content": ", where 3 represents the 3D coordinate offset between two point maps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 282, + 481, + 395 + ], + "blocks": [ + { + "bbox": [ + 318, + 282, + 481, + 395 + ], + "lines": [ + { + "bbox": [ + 318, + 282, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 318, + 282, + 481, + 395 + ], + "type": "image", + "image_path": "b3eb326dd75e3a1cb8a34044643e9ca3da395446166a8de567b1b43f70ef63d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 486, + 148, + 550, + 399 + ], + "blocks": [ + { + "bbox": [ + 486, + 148, + 550, + 399 + ], + "lines": [ + { + "bbox": [ + 486, + 148, + 550, + 399 + ], + "spans": [ + { + "bbox": [ + 486, + 148, + 550, + 399 + ], + "type": "image", + "image_path": "b0b43f99a91f4d85facba09a2986b5f19ce90c558770345542c747998fc84812.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_i^1" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " denotes the spatial position of " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " slight segment curves at the " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " iteration. And " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_i^2" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " is the spatial position of the same slight segment curves at the " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " iteration. " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_i^1" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " is updated by adding the gradient of " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_i^1" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": ". Figure 2 shows the iterative optimization process from " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " iteration to " + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 55, + 468, + 295, + 540 + ], + "type": "text", + "content": " iteration (Eq. (6))." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 559, + 295, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 585 + ], + "type": "text", + "content": "3.4. Differential Curves for Blurry Video Frame Interpolation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Although using differential curves theory to describe spatial motion is reasonable, its implementation is challenging: 1) How to convert a video frame " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " into a camera space point map " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_{i}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " and represent the spatial position. 2) How to estimate the relative displacement " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "SF" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " from an unknown point map to a source point map. And existing methods struggle to obtain the relative displacement. To this end, UBNet and MPNet are designed in DC-BVFI to address these issues. The pipeline of the proposed DC-BVFI is outlined in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 468, + 372, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 372, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 372, + 479 + ], + "type": "text", + "content": "3.4.1. UBNet" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 554, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 531 + ], + "type": "text", + "content": "Movement inevitably causes motion blur [38] and depth change. To convert the blurred video frame " + }, + { + "bbox": [ + 313, + 483, + 554, + 531 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 313, + 483, + 554, + 531 + ], + "type": "text", + "content": " into a 3D camera space point map " + }, + { + "bbox": [ + 313, + 483, + 554, + 531 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_i" + }, + { + "bbox": [ + 313, + 483, + 554, + 531 + ], + "type": "text", + "content": ", we designed a UBNet. More specifically, the UBNet includes two submodules:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 532, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 654 + ], + "type": "text", + "content": "Setup 1. The function of this submodule is to deblur. Traditional methods often simply apply Direct Stacking of video frames for deblurring. However, motion usually causes the content of frames to be different, resulting in misalignment between frames. Directly Stacking (DS) misalignment frames can hinder intermediate frame restoration, which may not facilitate video deblurring. Therefore, we introduce a Discriminative Deblur (DD) strategy, as shown in Fig. 3, to mitigate the adverse effects of blurring. The operations can be expressed as follows: " + }, + { + "bbox": [ + 313, + 532, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\{B_i\\} \\xrightarrow{\\mathrm{DD}} \\{I_i\\}" + }, + { + "bbox": [ + 313, + 532, + 554, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Setup 2. Real-world motion is a three-dimensional movement involving both depth and velocity variations. However, existing video frame interpolation methods typically focus only on velocity variations, which limits the quality of interpolation. To effectively explore three-dimensional" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1997" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "text", + "content": "motion, it is essential to convert the two-dimensional pixel space into the three-dimensional camera space. To convert the two-dimensional pixel space into the three-dimensional camera space, we employ a general 3D reconstruction paradigm [46], which can estimate camera intrinsics, depth maps, and point maps in the camera space from video frames. Specifically, the coordinate of the frame is mapped to a three-dimensional coordinate " + }, + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma] \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "text", + "content": ", where 3 represents the 3D coordinate, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "text", + "content": " indicates the total number of point. The operations can be expressed as follows: " + }, + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "inline_equation", + "content": "\\{I_i\\} \\xrightarrow{\\mathrm{ViT}} \\{[\\alpha, \\beta, \\gamma]_i\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 207 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 215, + 115, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 115, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 115, + 226 + ], + "type": "text", + "content": "3.4.2. MPNet" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": "In optical flow-based video interpolation methods, it is often necessary to estimate the optical flow " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "F_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 2}" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": " from the frame " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": " to the frame " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": " for interpolation [13, 28, 51]. Similarly, in 3D camera space, it is also necessary to estimate the scene flow from the point map " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "[\\alpha ,\\beta ,\\gamma ]_i" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": " to the point map " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "[\\alpha ,\\beta ,\\gamma ]_t" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": ". However, because the " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "[\\alpha ,\\beta ,\\gamma ]_t" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": " is unknown, existing methods cannot directly obtain the scene flow. To this end, we propose MPNet to upgrade the optical flow " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "F_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": " to the scene flow " + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "inline_equation", + "content": "SF_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 231, + 296, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": "Step 1. Given the frames " + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\{I_i\\}" + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": ", flow-based methods [13, 51, 54] require approximating the intermediate flows " + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "inline_equation", + "content": "F_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": " from the perspective of the synthesized frame " + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": ". Considering real-time performance, we employ IFNet [13] to estimate the optical flow " + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "inline_equation", + "content": "F_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 449 + ], + "type": "text", + "content": "Step 2. To express positional changes between point maps, we upgrade the optical flow " + }, + { + "bbox": [ + 55, + 400, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 400, + 296, + 449 + ], + "type": "text", + "content": " to camera space scene flow " + }, + { + "bbox": [ + 55, + 400, + 296, + 449 + ], + "type": "inline_equation", + "content": "SF_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 400, + 296, + 449 + ], + "type": "text", + "content": " to aware depth variations. The process is written as the following formula:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 459, + 295, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 459, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 108, + 459, + 295, + 474 + ], + "type": "interline_equation", + "content": "S F _ {i \\rightarrow t} = \\left(F _ {i \\rightarrow t} \\mid D _ {i \\rightarrow t}\\right) \\times K ^ {- 1}, \\tag {7}", + "image_path": "551088f37b9a2d54a3cec55fd649a3f9b88c959e8dba2e502cd40d068ef44aff.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "D_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 1} = D_i - D_t" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": " is the depth maps and the camera intrinsics " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "K\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "D_{t} = (1 - t)D_{i} + tD_{i + \\eta}" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "D_{i + \\eta}" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": " are the depth maps of two adjacent frames. " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "t\\in \\{i,i + \\eta \\}" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": " is the time index for interpolation. The " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": " denotes matrix multiplication, while | represents concatenation. The scene flow " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "SF_{i\\rightarrow t}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": ", where 3 represents the 3D coordinate offset between two point maps. We obtain the coarse scene flow " + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "inline_equation", + "content": "SF_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 483, + 296, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 593, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 665 + ], + "type": "text", + "content": "Step 3. Blurry frames lead to rough optical flow, leading to coarse scene flow. Therefore, it is necessary to refine the coarse scene flow. To achieve this, we introduce the Local Self-Attention (LSA) and Residual Block (RB) strategies, which perform fine-grained adjustments at the pixel level to obtain refining scene flow." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 673, + 108, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 108, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 108, + 685 + ], + "type": "text", + "content": "3.4.3.Warp" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "A brief explanation of the point map warping process by scene flow. Specifically, given a scene flow " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "SF_{i\\rightarrow t}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "text", + "content": "point map " + }, + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "text", + "content": ". The warp process can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 368, + 91, + 553, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 91, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 368, + 91, + 553, + 105 + ], + "type": "interline_equation", + "content": "[ \\alpha , \\beta , \\gamma ] _ {t} = [ \\alpha , \\beta , \\gamma ] _ {i} + S F _ {i \\rightarrow t}, \\tag {8}", + "image_path": "2ea2f847f0d483783d432daba308d515b1c3767ab5c9b434510f61a8c0abad93.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "text", + "content": " denotes element-wise addition and " + }, + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "inline_equation", + "content": "[\\alpha, \\beta, \\gamma]_t" + }, + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "text", + "content": " represents the warped point map." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 135, + 554, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 554, + 183 + ], + "type": "text", + "content": "To Frame. It is important to note that the warped point map represents 3D coordinates. We need to project the point map back into the 2D coordinate system. The pseudocode for warping and projecting onto the 2D frame is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "type": "code", + "bbox": [ + 323, + 210, + 585, + 322 + ], + "blocks": [ + { + "bbox": [ + 315, + 189, + 422, + 201 + ], + "lines": [ + { + "bbox": [ + 315, + 189, + 422, + 201 + ], + "spans": [ + { + "bbox": [ + 315, + 189, + 422, + 201 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 323, + 210, + 585, + 322 + ], + "lines": [ + { + "bbox": [ + 323, + 210, + 585, + 322 + ], + "spans": [ + { + "bbox": [ + 323, + 210, + 585, + 322 + ], + "type": "text", + "content": "Warp pt_t = pt + sf # point map is projected to frame b, _, h, w = pt.size() pt_t_ = torch/matmul(K, pt_t.view(b, 3, -1)) coord_t = pt_t_.div(pt_t_[:, 2:3, :])[:, 0:2, :] coord_t = coord_t.view(b, 2, h, w) grid = coord_t.transpose(1, 2).transpose(2, 3) I_t = torch.grid_sample(I_i, grid)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_body" + } + ], + "index": 15, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 335, + 406, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 406, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 406, + 346 + ], + "type": "text", + "content": "3.5. Loss Function." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 353, + 554, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 554, + 375 + ], + "type": "text", + "content": "We use the mean square error (MSE) as the training loss function:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 369, + 382, + 553, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 382, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 369, + 382, + 553, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| I _ {i} - I _ {i} ^ {g t} \\right\\| ^ {2} + \\left\\| I _ {t} - I _ {t} ^ {g t} \\right\\| ^ {2} \\tag {9}", + "image_path": "10cd6ebd433c7a027b8db35d83ad97b52c8664d2233ac92f0618046647e99517.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "text", + "content": "The variable " + }, + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "inline_equation", + "content": "I_{t}^{gt}" + }, + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "text", + "content": " denotes the ground truth frame. " + }, + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "text", + "content": " denotes the output frame of the model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 437, + 395, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 395, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 395, + 450 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 456, + 373, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 373, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 373, + 468 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": "Adobe. The Adobe dataset [42] consists of 120 videos at 240 fps with a resolution of " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": ". The blurry frames are 30 fps and downsized to " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "640 \\times 352" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " as done in [36, 38]. We follow [36, 39] to use the same 8 videos of the Adobe dataset for evaluation and other videos of the Adobe dataset for training. RBI. The real-world blurry frame interpolation [58] dataset comprises 55 videos. We use the same videos as the test set and the others as the training set [58]. The blurred image has an exposure time of " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "18~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": ", while the sharp image has an exposure time of nearly " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "2\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " [58]. This means that there are 9 sharp frames corresponding to one blurred frame, and 11 sharp frames exist in the readout dead-time between adjacent blurred frames [58]. GoPro. We use the same 11 videos of the GoPro dataset [33] to test the model [38]. The test datasets are also temporally downsampled to 30 fps with the blurring as [36]. YouTube. As a test dataset, the YouTube [38] consists of 59 video clips with different scenes and 12590 frames in total [38]. Please note that the evaluation datasets are also temporally downsampled to 30 fps with the blurring as [36]." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1998" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 87, + 506, + 276 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on Adobe and RBI. The best and the second best results are denoted by pink and yellow." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 87, + 506, + 276 + ], + "lines": [ + { + "bbox": [ + 102, + 87, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 102, + 87, + 506, + 276 + ], + "type": "table", + "html": "
MethodAdobeRBI
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]28.730.86528.660.86428.670.86427.040.84026.920.83827.130.840
TNTT [21]29.400.89329.450.87629.450.87628.880.85727.360.85428.110.855
DAIN [2]--29.300.877----28.380.860--
BIN [38]31.760.89628.360.84729.950.87829.010.86027.580.86128.300.861
ALANET [9]33.710.93232.980.93633.340.93528.810.86828.880.86828.840.868
UTI-VFI [56]28.730.86528.660.86428.670.86427.340.84427.120.84727.230.844
EQVI [28]--29.690.889----27.790.840--
ANIB [57]31.300.89032.830.89932.520.89428.830.88828.990.87228.900.870
DeMFI [36]33.830.93733.790.94133.790.94029.690.89929.000.89029.350.893
BiT [58]33.600.93633.860.93833.730.93729.580.89329.840.89729.750.894
JNMR [27]31.330.90731.680.93531.430.93928.520.87328.710.87928.630.875
IQ-VFI [11]33.830.94033.800.94133.790.94029.950.89429.920.89729.910.894
Ours34.170.94234.100.94234.140.94231.020.90430.890.91531.030.913
", + "image_path": "c46330881359bdd860a066e3bbeeff2286a1f73ee36652227bbec6af3485e1e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 101, + 303, + 504, + 459 + ], + "blocks": [ + { + "bbox": [ + 73, + 286, + 535, + 298 + ], + "lines": [ + { + "bbox": [ + 73, + 286, + 535, + 298 + ], + "spans": [ + { + "bbox": [ + 73, + 286, + 535, + 298 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on GoPro and YouTube. The best and the second best results are denoted by pink and yellow." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 101, + 303, + 504, + 459 + ], + "lines": [ + { + "bbox": [ + 101, + 303, + 504, + 459 + ], + "spans": [ + { + "bbox": [ + 101, + 303, + 504, + 459 + ], + "type": "table", + "html": "
MethodYouTubeGoPro
DeblurringMFIAvgDeblurringMFIAvg
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
QVI [51]29.270.88429.980.89029.990.89027.200.81526.620.810--
TNTT [21]29.590.88929.770.89029.750.89026.480.80826.680.81526.650.814
DAIN [2]--32.020.901----25.540.859--
UTI-VFI [56]28.610.87928.640.87328.640.87725.660.80825.630.84525.640.814
EQVI [28]--29.110.893----29.680.881--
DeMFI [36]32.900.92532.790.92632.800.92630.820.89330.780.90030.750.900
BiT [58]32.370.92031.880.91831.500.91531.270.90631.200.90528.950.885
JNMR [27]--32.250.922----31.020.900--
IQ-VFI [11]33.030.92932.900.93032.940.92831.320.90631.260.90431.280.904
Ours33.400.93233.320.93333.350.93232.140.91232.050.91032.080.910
", + "image_path": "74ae59ae8a1440e203843d28c4767408b08cfb2b9ea7fe0f1d9166ef91c47036.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 190, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 190, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 190, + 483 + ], + "type": "text", + "content": "4.2. Implementation Details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "text", + "content": "We optimize the loss using Adam in the PyTorch framework. The cosine scheduler schedules the learning rate from " + }, + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "inline_equation", + "content": "1e - 6" + }, + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "text", + "content": ". Standard data augmentation techniques, such as flipping, rotation, and cropping, are applied to the data with a size of " + }, + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 490, + 296, + 573 + ], + "type": "text", + "content": ". We train our model on the training datasets with a batch size of 4 for 900 epochs on 4 NVIDIA 3090ti GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 246, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 246, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 246, + 598 + ], + "type": "text", + "content": "4.3. Comparison with Previous Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "As shown in Tab. 1, a total of 12 outperformers are compared, which are fine-tuned on the Adobe and RBI datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Results on the RBI. Our approach achieves a 1.12dB improvement over the current state-of-the-art. The RBI dataset contains a large number of depth-varying scenes, such as vehicle motion and camera movement, resulting in a high average motion magnitude [58]. DC-BVFI performs well on the RBI dataset, demonstrating that the explicit depth-aware approach is advantageous in challenging scenarios," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 471, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 506 + ], + "type": "text", + "content": "such as vehicle motion, where many existing methods tend to fail. The proposed DC-BVFI consistently outperforms existing methods, achieving a PSNR of 34.14 on Adobe." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 507, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 650 + ], + "type": "text", + "content": "To verify the generalization capability, we conducted cross-dataset experiments using YouTube and GoPro as test sets, as shown in Tab. 2. Results on the GoPro. We observed that our method achieved an impressive performance of 32.08 PSNR on the GoPro test set. The GoPro dataset contains a large number of dynamic scenes, resulting in a high average motion magnitude, making it a challenging benchmark for video frame interpolation. Our method performs well on the GoPro dataset, suggesting that DC-BVFI exhibits strong generalization across various motion scenarios. We also observed the final performance of 33.35 PSNR on the YouTube test set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 659, + 478, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 478, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 478, + 672 + ], + "type": "text", + "content": "4.4. Comparison of Visual Results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We further compare the visualization results in depth-varying motion scenarios, such as car motion and camera motion. As shown in Fig. 4 our method's predictions are" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1999" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 242 + ], + "type": "image", + "image_path": "340315f5ee2f3dcf609005d75cc80d76c1a9f891e9fb83cecef6cf7006f4d47a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 270 + ], + "type": "text", + "content": "Figure 4. Visual comparison with different methods. The comparison primarily focuses on depth change scenarios, such as car motion and camera movement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 277, + 284, + 451 + ], + "blocks": [ + { + "bbox": [ + 69, + 277, + 284, + 451 + ], + "lines": [ + { + "bbox": [ + 69, + 277, + 284, + 451 + ], + "spans": [ + { + "bbox": [ + 69, + 277, + 284, + 451 + ], + "type": "image", + "image_path": "2367541d7718110ebb4a26fdfa7db0877bf03450b71f2186e4b0b762b95b7cbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 459, + 254, + 471 + ], + "lines": [ + { + "bbox": [ + 96, + 459, + 254, + 471 + ], + "spans": [ + { + "bbox": [ + 96, + 459, + 254, + 471 + ], + "type": "text", + "content": "Figure 5. Comparison of Motion-in-Depth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 482, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 566 + ], + "type": "text", + "content": "closer to the ground truth, with more precise details on both car motion and camera motion. Methods like QVI [51], EQVI [28], and IQ-VFI [11] explore inter-frame motion through estimated acceleration to enhance motion reconstruction. However, they struggle to express changes in object depth during motion, often leading to suboptimal outcomes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Comparison of Motion-in-Depth. Moreover, to quantitatively measure depth variations during motion. We employ the MiD metric [18] to evaluate the depth variations. Since the video interpolation dataset lacks ground truth depth maps, we use the depth maps estimated from clear video frames [53] as the ground truth labels. Fig. 5 reports the performance of quadratic methods on multi-frame interpolation in Adobe, RBI, YouTube, and GoPro. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of MiD. This indicates that our method effectively explores the depth variations cue to enhance video frame" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 276, + 423, + 362 + ], + "blocks": [ + { + "bbox": [ + 316, + 276, + 423, + 362 + ], + "lines": [ + { + "bbox": [ + 316, + 276, + 423, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 423, + 362 + ], + "type": "image", + "image_path": "36d03ef72aa0b1895d518ca52e42780da7e3fa1af4beeabf0eb2d7ca989e5030.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 364, + 423, + 450 + ], + "blocks": [ + { + "bbox": [ + 316, + 364, + 423, + 450 + ], + "lines": [ + { + "bbox": [ + 316, + 364, + 423, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 423, + 450 + ], + "type": "image", + "image_path": "6f17612969cad5b6966540cf8451e926a7dbc13e1e78d4837f3e197dd7d59351.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 458, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 479 + ], + "type": "text", + "content": "Figure 6. Comparison of temporal consistency performance under tOF metric in Adobe, RBI, YouTube, and GoPro datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 425, + 276, + 531, + 361 + ], + "blocks": [ + { + "bbox": [ + 425, + 276, + 531, + 361 + ], + "lines": [ + { + "bbox": [ + 425, + 276, + 531, + 361 + ], + "spans": [ + { + "bbox": [ + 425, + 276, + 531, + 361 + ], + "type": "image", + "image_path": "a69dc98d0ad741ae7c4855c414cf549f719d6bc6f9c6cda3de77e8d8eaecb9d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 364, + 531, + 449 + ], + "blocks": [ + { + "bbox": [ + 425, + 364, + 531, + 449 + ], + "lines": [ + { + "bbox": [ + 425, + 364, + 531, + 449 + ], + "spans": [ + { + "bbox": [ + 425, + 364, + 531, + 449 + ], + "type": "image", + "image_path": "39c8376a3ee967108d799b9b896a63dca7d612a0a2f1c56168edfd12a950eeb2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 485, + 371, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 485, + 371, + 496 + ], + "spans": [ + { + "bbox": [ + 314, + 485, + 371, + 496 + ], + "type": "text", + "content": "interpolation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "content": "Comparison of Temporal Consistency. We employ the tOF metric [7, 16, 41] to evaluate the temporal consistency. Figure 6 reports the quality of motion reconstruction across several existing models on the Adobe, RBI, GoPro, and YouTube test sets. We can clearly observe that the proposed DC-BVFI method consistently outperforms existing methods in both the mean and variance of tOF. The outstanding temporal consistency is due to DC-BVFI's explicit depth modeling of rapid object motion, which greatly improves motion temporal consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 416, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 416, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 416, + 635 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": "We conducted ablation experiments to evaluate the contribution of the 3D architecture and the impact of each submodule in our method, namely the DD and the LSA. Additionally, we investigated the influence of scene flow and optical flow on video frame interpolation. Furthermore, we measured the parameter count and computational complex" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2000" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 124, + 294, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 3. Ablation studies of architecture on the RBI dataset. DS stands for stacking video frames along the channel axis, while DD denotes the discriminative deblur strategy. LSA and RB represent Local Self-Attention and Residual Block, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 124, + 294, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 124, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 124, + 294, + 198 + ], + "type": "table", + "html": "
ArchitectureBackBoneSettingUBNetMPNetRBI
DDDSLSARBPSNRMiD(↓)
3DViT-Base(a)✓✓31.0376
(b)✓✓30.9883
(c)✓✓30.6681
2DViT-Base(d)✓✓28.91137
", + "image_path": "45c9e0650b262a73cf51a7e00bc963f1f02ed9a71a28650217f37b33b6180086.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 208, + 296, + 310 + ], + "blocks": [ + { + "bbox": [ + 56, + 208, + 296, + 310 + ], + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 310 + ], + "type": "image", + "image_path": "435732997026487a3b6bd0cc13f36f975e19bd890c0700f84df1c9572914ff79.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 296, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 352 + ], + "type": "text", + "content": "Figure 7. Visualization of the architecture ablation experiments. In the scene with car depth variations, our architecture achieves promising results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 146, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 146, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 146, + 376 + ], + "type": "text", + "content": "ity of various models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 378, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 533 + ], + "type": "text", + "content": "Architecture. We empirically evaluated the proposed 3D architecture's effectiveness in guiding video frame interpolation in scenes with depth variations. For a fair comparison, both methods use the same backbone but differ in their output prediction heads. PSNR and MiD are presented in Tab. 3. The results of Tab. 3 and Fig. 7 indicate that the 3D architecture significantly outperforms the 2D architecture in scenes with depth variations. The results of setting (b) of Tab. 3 and Fig. 7 highlight the importance of refining scene flow. The results of setting (c) of Tab. 3 and Fig. 7 show that directly stacking misaligned frames hinders intermediate frame restoration, and may not facilitate video deblurring." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": "Flow. We analyzed the effects of scene flow and optical flow on video frame interpolation performance. For a fair evaluation, in the 2D architecture, we replace the depth estimation in the 3D architecture with a ViT-Base model that has the same parameter size. The results in Tab. 4 and Fig. 8 indicate that scene flow outperforms optical flow in scenarios with depth variations. This is in line with our intuitions, as scene flow integrates depth information on top of optical flow, providing a more comprehensive consideration of useful cues. In other words, when motion doesn't cause depth variations, the scene flow in the depth spatial dimension is deactivated, degrading scene flow into the optical flow. Essentially, scene flow encompasses optical flow. These additional insights contribute to improved video frame interpolation performance in scenes with depth variations." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 91, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 337, + 71, + 531, + 81 + ], + "lines": [ + { + "bbox": [ + 337, + 71, + 531, + 81 + ], + "spans": [ + { + "bbox": [ + 337, + 71, + 531, + 81 + ], + "type": "text", + "content": "Table 4. Ablation studies of Flow on the RBI dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 91, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 317, + 91, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 553, + 185 + ], + "type": "table", + "html": "
ArchitectureFlowSettingFlowNetDepthRBI
IFNetRAFTMeGeDust3RPSNRMiD(↓)
3DScene Flow(a)✓✓31.0376
(b)✓✓30.9982
(c)✓✓31.0079
(d)✓✓30.8685
ArchitectureFlowIFNetRAFTViT-Base 2DPSNRMiD(↓)
2DOptical Flow(e)✓✓28.80143
(f)✓✓28.78147
", + "image_path": "f1129cd8c3be4eaac729f829f08276a84e90c59ecc952e7da2745ead0fe1883d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 208, + 550, + 430 + ], + "blocks": [ + { + "bbox": [ + 316, + 208, + 550, + 430 + ], + "lines": [ + { + "bbox": [ + 316, + 208, + 550, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 550, + 430 + ], + "type": "image", + "image_path": "58ba4783ba13072ebff9ae21e54e6b27d6755ff0444587d084b17957a3ab7510.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 438, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 472 + ], + "type": "text", + "content": "Figure 8. The results of video multi-frame interpolation are presented. The results of the 3D-based architecture (a) are shown in the top rows, and the 2D-based architecture (e) in the bottom rows." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "content": "This is the first work to focus on solving the depth variations in video frame interpolation. We introduce the differential curves to analyze the issue of depth variations in video frame interpolation. Based on the analysis, this paper builds an explicit depth-aware video interpolation framework to overcome the issue. The UBNet converts video frames into explicit 3D point maps in camera space, and the MPNet estimates the scene flow between these point maps. DC-BVFI has achieved state-of-the-art performance on four benchmark datasets: Adobe, RBI, YouTube, and GoPro." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the National Key R&D Program of China (2022ZD0161800), the National Natural Science Foundation of China under Grant 62271203, AI-Empowered Research Paradigm Reform and Discipline Leap Plan under Grant 2024AI01012 and the Open Research Fund of KLATASDS-MOE, ECNU." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "2001" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing, 5(4-5):185-196, 1993. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 115, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 115, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 115, + 295, + 168 + ], + "type": "text", + "content": "[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 294, + 236 + ], + "type": "text", + "content": "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):933-948, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "text", + "content": "[4] Xianhang Cheng and Zhenzhong Chen. Multiple video frame interpolation via enhanced deformable separable convolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7029-7045, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "type": "text", + "content": "[5] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 417 + ], + "type": "text", + "content": "[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 107-123. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "text", + "content": "[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 39(4):75-1, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 294, + 508 + ], + "type": "text", + "content": "[8] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 294, + 564 + ], + "type": "text", + "content": "[9] Akash Gupta, Abhishek Aich, and Amit K Roy-Chowdhury. Alanet: Adaptive latent attention network for joint video deblurring and interpolation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 256-264, 2020. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "text", + "content": "[11] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[12] Ping Hu, Simon Niklaus, Stan Sclaroff, and Kate Saenko. Many-to-many splatting for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3553-3562, 2022. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[13] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[14] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "text", + "content": "[15] Zhilin Huang, Yijie Yu, Ling Yang, Chujun Qin, Bing Zheng, Xiawu Zheng, Zikun Zhou, Yaowei Wang, and Wenming Yang. Motion-aware latent diffusion models for video frame interpolation. arXiv preprint arXiv:2404.13534, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[16] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 169-184. Springer, 2020. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 553, + 319 + ], + "type": "text", + "content": "[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 363 + ], + "type": "text", + "content": "[18] Junhwa Hur and Stefan Roth. Self-supervised multi-frame monocular scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2684-2694, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "type": "text", + "content": "[19] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "type": "text", + "content": "[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 555, + 533 + ], + "type": "text", + "content": "[21] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8112-8121, 2019. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[22] Yoshihiko Kuroki, Tomohiro Nishi, Seiji Kobayashi, Hideki Oyaizu, and Shinichi Yoshimura. A psychophysical study of improvements in motion-image quality by using high frame rates. Journal of the Society for Information Display, 15(1): 61-68, 2007. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "text", + "content": "[23] Yoshihiko Kuroki, Haruo Takahashi, Masahiro Kusakabe, and Ken-ichi Yamakoshi. Effects of motion image stimuli with normal and high frame rates on eeg power spectra: comparison with continuous motion image stimuli. Journal of the Society for Information Display, 22(4):191-198, 2014. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Danyeong Lee, Dohoon Lee, Dongmin Bang, and Sun Kim. Disco: Diffusion schrödinger bridge for molecular conformer optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13365-13373, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive col" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2002" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "laboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 160 + ], + "type": "text", + "content": "[26] Pengcheng Lei, Zaoming Yan, Tingting Wang, Faming Fang, and Guixu Zhang. Three-stage temporal deformable network for blurry video frame interpolation. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[27] Meiqin Liu, Chenming Xu, Chao Yao, Chunyu Lin, and Yao Zhao. Jnmr: Joint non-linear motion regression for video frame interpolation. IEEE Transactions on Image Processing, 2023. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "text", + "content": "[28] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 41-56. Springer, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 262, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 304 + ], + "type": "text", + "content": "[29] Yu-Lun Liu, Yi-Tung Liao, Yen-Yu Lin, and Yung-Yu Chuang. Deep video frame interpolation using cyclic frame generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8794-8802, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 306, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 294, + 349 + ], + "type": "text", + "content": "[30] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "type": "text", + "content": "[31] Zonglin Lyu, Ming Li, Jianbo Jiao, and Chen Chen. Frame interpolation with consecutive brownian bridge diffusion. arXiv preprint arXiv:2405.05953, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "text", + "content": "[32] Simone Meyer, Oliver Wang, Henning Zimmer, Max Grosse, and Alexander Sorkine-Hornung. Phase-based frame interpolation for video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1410-1418, 2015. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "type": "text", + "content": "[33] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[34] Simon Niklaus and Feng Liu. Context-aware synthesis for video frame interpolation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[35] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive separable convolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 261-270, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[36] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, ECCV 2022, pages 1-17. European Conference on Computer Vision, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "text", + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[38] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing, 30:277-292, 2020. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "text", + "content": "[40] Zhihao Shi, Xiaohong Liu, Kangdi Shi, Linhui Dai, and Jun Chen. Video frame interpolation via generalized deformable convolution. IEEE Transactions on Multimedia, 24:426-439, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "text", + "content": "[41] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVfi: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14489-14498, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[42] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 348 + ], + "type": "text", + "content": "[43] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "text", + "content": "[44] Kristopher Tapp. Differential geometry of curves and surfaces. Springer, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 425 + ], + "type": "text", + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[46] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "text", + "content": "[47] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[48] Wen Wang, Qiuyu Wang, Kecheng Zheng, Hao Ouyang, Zhekai Chen, Biao Gong, Hao Chen, Yujun Shen, and Chunhua Shen. Framer: Interactive frame interpolation. arXiv preprint arXiv:2410.18978, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[49] Jiyan Wu, Chau Yuen, Ngai-Man Cheung, Junliang Chen, and Chang Wen Chen. Modeling and optimization of high frame rate video transmission over wireless networks. IEEE Transactions on Wireless Communications, 15(4): 2713-2726, 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[50] Jinbo Xing, Wenbo Hu, Yuechen Zhang, and Tien-Tsin Wong. Flow-aware synthesis: A generic motion model for video frame interpolation. Computational Visual Media, 7: 393-405, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[51] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2003" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 441 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[52] Gengshan Yang and Deva Ramanan. Upgrading optical flow to 3d scene flow through optical expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1343, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[53] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[54] Zhiyang Yu, Yu Zhang, Xujie Xiang, Dongqing Zou, Xijun Chen, and Jimmy S Ren. Deep bayesian video frame interpolation. In European Conference on Computer Vision, pages 144-160. Springer, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[55] Zhiyang Yu, Yu Zhang, Dongqing Zou, Xijun Chen, Jimmy S Ren, and Shunqing Ren. Range-nullspace video frame interpolation with focalized motion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22159-22168, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 295 + ], + "type": "text", + "content": "[56] Youjian Zhang, Chaoyue Wang, and Dacheng Tao. Video frame interpolation without temporal priors. Advances in Neural Information Processing Systems, 33:13308-13318, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "type": "text", + "content": "[57] Zhihang Zhong, Xiao Sun, Zhirong Wu, Yinqiang Zheng, Stephen Lin, and Imari Sato. Animation from blur: Multimodal blur decomposition with motion guidance. In European Conference on Computer Vision, pages 599-615. Springer, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 407 + ], + "type": "text", + "content": "[58] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "type": "text", + "content": "[59] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2004" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_content_list.json b/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0459b96690a8e014a80d6bb02f506f3fc423c0c7 --- /dev/null +++ b/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_content_list.json @@ -0,0 +1,1541 @@ +[ + { + "type": "text", + "text": "Exploiting Deblurring Networks for Radiance Fields", + "text_level": 1, + "bbox": [ + 233, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haeyun Choi $^{1,\\dagger}$", + "bbox": [ + 173, + 179, + 300, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heemin Yang", + "bbox": [ + 346, + 179, + 467, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Janghyeok Han²", + "bbox": [ + 513, + 179, + 647, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sunghyun Cho²", + "bbox": [ + 694, + 179, + 823, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "${}^{1}\\mathrm{KT}$", + "bbox": [ + 401, + 215, + 439, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ POSTECH", + "bbox": [ + 491, + 215, + 591, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 270, + 328, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose DeepDeblurRF, a novel radiance field deblurring approach that can synthesize high-quality novel views from blurred training views with significantly reduced training time. DeepDeblurRF leverages deep neural network (DNN)-based deblurring modules to enjoy their deblurring performance and computational efficiency. To effectively combine DNN-based deblurring and radiance field construction, we propose a novel radiance field (RF)-guided deblurring and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Moreover, DeepDeblurRF is compatible with various scene representations, such as voxel grids and 3D Gaussians, expanding its applicability. We also present BlurRF-Synth, the first large-scale synthetic dataset for training radiance field deblurring frameworks. We conduct extensive experiments on both camera motion blur and defocus blur, demonstrating that DeepDeblurRF achieves state-of-the-art novel-view synthesis quality with significantly reduced training time.", + "bbox": [ + 86, + 303, + 485, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 621, + 220, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel-view synthesis has seen significant advancements in recent years, leading to impressive improvements in photorealistic rendering quality. One key development in this field is the Neural Radiance Field (NeRF) [25], which leverages neural networks to generate highly detailed images from novel viewpoints. Building on this, various works have focused on reducing training time and enhancing rendering quality [3, 12, 15, 27]. More recently, 3D Gaussian Splatting (3DGS) [17] has emerged, utilizing Gaussians as an explicit 3D representation. This technique enables high-quality scene reconstruction and real-time rendering through a differentiable rasterization method.", + "bbox": [ + 89, + 646, + 483, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, synthesizing sharp novel views from degraded training views remains a significant challenge. Capturing images in real-world conditions often leads to various degra", + "bbox": [ + 89, + 829, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ditions such as blur and noise. Among them, blur makes it difficult to aggregate accurate 3D information from the training views, resulting in a blurry radiance field. To address blur, several works [20-22, 24, 31, 32, 39] have been proposed to reconstruct sharp radiance fields from training views with either camera motion blur or defocus blur. These radiance field deblurring methods jointly optimize blur kernels and the radiance field, and have shown promising results.", + "bbox": [ + 511, + 272, + 906, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, these methods still possess several limitations that hinder their effectiveness. First, existing approaches rely on linear blur models, describing blurred pixels in training views as linear combinations of sharp pixels. In contrast, real-world blurred images often suffer from non-linear outliers such as saturated pixels due to sensor clipping, noise from various sources, and non-linear in-camera processing [8, 35]. These violate the linear blur model, leading to deblurring artifacts and severely degrading the performance of current radiance field deblurring techniques. Second, existing approaches do not utilize priors on sharp images, relying solely on complementary information from different views to model blur kernels and estimate a sharp radiance field. Consequently, they tend to produce radiance fields with residual blur and may completely fail when all input views exhibit similar blur directions [24], akin to classical multi-frame deblurring approaches [5, 7, 33], due to the lack of sharp image priors. Third, existing ray-based methods [21, 22, 24, 31, 39] employ multiple ray samples per pixel to depict a blurry image, resulting in a substantial increase in rendering time for a single view from a radiance field, and consequently, a significant computation time for radiance field construction.", + "bbox": [ + 511, + 397, + 908, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Apart from the radiance field deblurring methods, image deblurring has been extensively studied for decades [6, 14, 30, 40, 43]. Recently, a number of deep neural network (DNN)-based single-image deblurring methods have been proposed [4, 10, 19, 28, 38, 42]. These methods do not rely on linear blur models, but learn image deblurring from large-scale datasets. Thanks to this, they can effectively remove blur from a single image without the need for complementary information from other images, and also handle nonlinearities such as saturated pixels [34]. Moreover, their feedforward approach allows markedly reduced processing times,", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†This work was done at POSTECH.", + "bbox": [ + 114, + 886, + 307, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6012", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a825a29597ef2e0bd15045d45c3ae4227429d68f93ee73ba2d4621ad91b7c305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 88, + 236, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e1184483bea44f28c0011d0159c45c94a87202883b1c033064a9953d7b7c464.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 88, + 372, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08e266c21c2f529abb7cedc616145f69139cd6419e720b72d79b0f1f9cdb6498.jpg", + "image_caption": [ + "PDRF-10 (4 hrs)" + ], + "image_footnote": [], + "bbox": [ + 374, + 89, + 509, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/85118b729b553c03f49b821a13cf8a2cb5f822a2cefc749f3a6243282e60fc97.jpg", + "image_caption": [ + "Blurry Input View" + ], + "image_footnote": [], + "bbox": [ + 101, + 200, + 236, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b5b69971ac04608946190baa0c2a5f3055062a8c2164b24ce446b9766c861e3b.jpg", + "image_caption": [ + "BAGS (75 mins)" + ], + "image_footnote": [], + "bbox": [ + 101, + 272, + 236, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f815a23f1db8d555d761d5df8f860e310c6c91ee6b81c856439da47df3a30f08.jpg", + "image_caption": [ + "Deblur-NeRF (31 hrs)" + ], + "image_footnote": [], + "bbox": [ + 236, + 200, + 372, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/931d0f2dda88774247c89e18d5121fe3f74ddc3ded13d6aa058e5c674b8f6b47.jpg", + "image_caption": [ + "DeepDeblurRF-G (17 mins)", + "Figure 1. Given a set of multi-view blurry images even with non-linear outliers such as saturated pixels and noise, DeepDeblurRF performs high-quality novel-view synthesis with highly efficient training. DeepDeblurRF-P and DeepDeblurRF-G are the results of our framework, where radiance fields are constructed using Plenoxels [12] and 3D Gaussian Splatting [17], respectively. Note that DP-NeRF [21] was trained with two GPUs due to its memory demands, whereas other models were trained on a single NVIDIA TITAN RTX GPU." + ], + "image_footnote": [], + "bbox": [ + 236, + 272, + 362, + 300 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c9378b73223feba652c9da77710696079b5361a87d55422ce647b5b13c418dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 200, + 511, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a4784dad801f62c0407f3326134d5c099c54a297e80806f086d917b017fe967e.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 374, + 272, + 470, + 300 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5105211b7eaee71448d070b74a4160e4022c3a0c8ed6c2219403c94a626d5a1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 93, + 897, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unlike classical iterative optimization-based approaches.", + "bbox": [ + 89, + 397, + 464, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the effectiveness of recent DNN-based deblurring approaches, a promising direction to construct a sharp radiance field from blurry images would be to combine a deblurring network and radiance field construction. However, a naive combination of a deblurring network and radiance field construction, which performs single-image deblurring to each input blurry image and trains a radiance field using the deblurred images, results in unsatisfactory results as reported by Ma et al. [24]. This is because of the limited performance of single-image deblurring, which stems from the insufficient information available in a single image.", + "bbox": [ + 88, + 414, + 483, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose DeepDeblurRF, a novel radiance field deblurring framework that enables highly efficient training and high-quality novel-view synthesis from images blurred by motion blur or defocus blur, even in the presence of noise and saturated pixels, as shown in Fig. 1. Unlike existing radiance field deblurring approaches, our method leverages DNN-based deblurring modules to achieve superior performance and computational efficiency. Additionally, our approach is versatile, capable of handling various 3D representations such as voxel grids [12] and 3D Gaussians [17].", + "bbox": [ + 88, + 580, + 483, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, solely relying on deblurring networks to remove blur from input blurry images may yield unsatisfactory results as discussed above. To effectively combine DNN-based deblurring and radiance field construction, our framework adopts a novel radiance field (RF)-guided deblurring scheme, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Specifically, our framework first deblurs input images, and then trains a radiance field using the deblurred images. Then, at the next iteration, it renders images corresponding to the input views from the trained radiance field. While the de", + "bbox": [ + 88, + 734, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "blurred images may have limited qualities with inaccurately restored details, the rendered images provide higher-quality details as they are rendered from aggregated information from multiple deblurred images. Using the rendered images as guidance, our framework then performs RF-guided deblurring to the input blurred images. Thanks to the guidance, we can obtain higher-quality deblurred images, which we use for radiance field construction again. Finally, our framework iterates RF-guided deblurring and radiance field construction to gradually enhance the quality of the radiance field.", + "bbox": [ + 511, + 397, + 906, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training the deblurring modules of our framework requires a large-scale dataset of blurred images paired with ground-truth sharp images for radiance field reconstruction. However, no such datasets are available as existing radiance field deblurring approaches adopt linear blur model-based approaches, which do not require large-scale training datasets [20-22, 24, 31, 32, 39]. Consequently, the datasets provided by previous works, such as those from DeblurNeRF [24] and ExBluRF [22], are small and primarily designed for evaluation rather than training. Furthermore, these datasets often overlook crucial factors like camera noise, in-camera processing, and other non-linear outliers, which are essential for realistic deblurring tasks.", + "bbox": [ + 511, + 550, + 908, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these gaps, we propose BlurRF-Synth, a large-scale dataset for radiance field deblurring approaches. It includes 4,350 blurred-sharp image pairs across 150 scenes, with 2,175 pairs for each of 75 scenes, encompassing both camera motion blur and defocus blur. These images are carefully synthesized to reflect real-world camera degradations such as noise, saturated pixels, and in-camera processing pipelines. Additionally, we introduce a real-world dataset called BlurRF-Real for evaluation in non-ideal conditions. Unlike the Deblur-NeRF dataset, which has less noise and", + "bbox": [ + 511, + 750, + 910, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6013", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adequate lighting, BlurRF-Real consists of five low-light indoor scenes captured with a machine vision camera.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We validate DeepDeblurRF on both synthetic and real-world datasets with two types of blur: camera motion blur and defocus blur. Experimental results demonstrate that our method achieves state-of-the-art novel-view synthesis performance with highly reduced processing times. Furthermore, we demonstrate the extensibility of our framework by constructing radiance fields using different scene representations, such as voxel grids and Gaussian Splatting. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 121, + 483, + 256 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose DeepDeblurRF, a novel radiance field deblurring framework, which is the first approach that leverages DNN-based deblurring modules to overcome the limitations of the linear blur model.", + "- To this end, we present RF-guided deblurring, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner to gradually enhance the quality of the radiance field.", + "- We also present the BlurRF-Synth dataset, the first large-scale dataset for training and evaluation of novel-view synthesis from blurry images." + ], + "bbox": [ + 89, + 257, + 482, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 434, + 232, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Radiance field deblurring Several works have recently been proposed to tackle radiance field deblurring [21, 22, 24, 31, 39]. Ma et al. presented Deblur-NeRF, the first framework to train a sharp radiance field from blurry training views [24]. Deblur-NeRF adopts an additional MLP to model blur kernels and jointly optimizes them with a radiance field. DP-NeRF [21] models blur kernels based on physical priors for rigid motions, while PDRF [31] uses a two-stage kernel modeling scheme where the first stage coarsely aggregates kernel points to reduce computational cost. All these methods support both camera motion blur and defocus blur by modeling them using multiple rays. On the other hand, BAD-NeRF [39] and ExBluRF [22] propose radiance field deblurring for camera motion blur, modeling camera motion blur using camera trajectories. Nevertheless, all these methods rely on linear blur models, and require multiple ray-tracing steps, resulting in limited deblurring quality and slow training speeds.", + "bbox": [ + 89, + 462, + 483, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, with the emergence of 3D Gaussian Splating [17], 3D Gaussian-based radiance field deblurring methods have been proposed. Deblurring-3DGS [20] handles camera motion blur with small offsets in 3D Gaussian positions and defocus blur by adjusting the geometry of Gaussians. BAGS [32] models both types of blur using dense per-pixel blur kernels optimized in a coarse-to-fine manner. However, BAGS requires substantial computation time to optimize dense blur kernels. Moreover, these methods also rely on linear blur models and are limited in handling real-world blurred images.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image deblurring Image deblurring has been widely studied to address blur that degrades image quality. Traditional methods, relying on linear blur models, often struggle with real-world blurred images containing non-linear outliers [6, 7, 9, 14, 30, 33, 40, 43]. Their iterative optimization processes are also computationally intensive, limiting their practical use. Recent advancements in deep learning have introduced DNN-based approaches, including single-image [4, 10, 19, 28, 38, 42] and multi-frame/video deblurring [2, 11, 13, 16, 29, 37, 41, 44, 45]. These methods can handle more complex blurs and reduce processing time with a feed-forward structure.", + "bbox": [ + 511, + 90, + 906, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While applying DNN-based deblurring approaches might be a potential solution for constructing radiance fields from blurry images, naively using single-image deblurring techniques yields unsatisfactory results as discussed in Sec. 1. Another possible direction is to adopt multi-frame and video deblurring approaches that leverage complementary information from multiple images for higher-quality results. However, these methods assume that input images are captured from nearly the same viewpoints, a condition that does not hold for training views in radiance field construction.", + "bbox": [ + 511, + 272, + 908, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. DeepDeblurRF", + "text_level": 1, + "bbox": [ + 511, + 436, + 666, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2-(a) illustrates the overall framework of DeepDeblurRF. It takes a set of $M$ blurred images $\\mathbf{B} = \\{B_1, \\dots, B_M\\}$ of a scene and estimates a sharp radiance field that can synthesize a sharp novel view given an arbitrary pose. We assume that each blurred image $B_m$ is obtained from its latent sharp image $L_m$ through an unknown degradation process including blur and non-linear outliers such as saturated pixels and nonlinear in-camera processing. Based on this assumption, DeepDeblurRF first performs initial deblurring to the input blurred images, and obtains initial deblurred images. Then, our method iteratively performs radiance field construction using deblurred images and RF-guided deblurring to gradually enhance the quality of the radiance field and the deblurred images. At the last iteration, we perform only the radiance field construction step and obtain a final radiance field, from which we can synthesize sharp novel views. In the following, each step of DeepDeblurRF is described in more detail.", + "bbox": [ + 511, + 460, + 908, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Initial Deblurring", + "text_level": 1, + "bbox": [ + 511, + 742, + 687, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initial deblurring step removes blur from each blurry training view $B_{m} \\in \\mathbf{B}$ so that the following radiance field construction step can estimate the pose of each view more accurately, and more effectively aggregate information from different views. To this end, we adopt an off-the-shelf single-image deblurring network. Specifically, we adopt NAFNet [4], a state-of-the-art deblurring network, for its computational efficiency and performance. We denote the deblurred image from the initial deblurring step correspond", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6014", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/55a8984932fcb43a456eba91f49ff6402ed0cdbd887d1b49742c8883647c855d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 93, + 594, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c12843d879350de223aaf310b195504d14519ef8617e4b6d98aa207be10e89e8.jpg", + "image_caption": [ + "(b) Blurry training view" + ], + "image_footnote": [], + "bbox": [ + 606, + 93, + 750, + 169 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f05916773816b1a04174bd5f2044e7e83f6f7dfe1aadee05b54d15151fa94a0.jpg", + "image_caption": [ + "(d) $1^{\\mathrm{st}}$ iteration rendering" + ], + "image_footnote": [], + "bbox": [ + 606, + 181, + 751, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/45ff9b2c1b4ffbc3f616f370fc1673e8f69232b3834eea0d0a70a98420ba1714.jpg", + "image_caption": [ + "(c) Initial deblurring" + ], + "image_footnote": [], + "bbox": [ + 751, + 94, + 895, + 169 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/44a242c3f6921bb1d55e7b5f7ebb56dbe8eb0c15fe2ed17922a0fa4fa62a76db.jpg", + "image_caption": [ + "(e) $1^{\\mathrm{st}}$ RF-guided deblurring" + ], + "image_footnote": [], + "bbox": [ + 751, + 181, + 895, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/504f36ad21fa11bec55ef5980635dd1c99c4032b86a58040cef44c631928afc9.jpg", + "image_caption": [ + "(a) $1^{\\mathrm{st}}$ iteration rendering" + ], + "image_footnote": [], + "bbox": [ + 101, + 297, + 295, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5cdaface6f9b95374b79a2ca75a695e4791d0d4d7a21f30f27d7ed417b0d8d58.jpg", + "image_caption": [ + "Figure 2. Overall framework and intermediate result of each step of DeepDeblurRF.", + "(b) $2^{\\mathrm{nd}}$ iteration rendering" + ], + "image_footnote": [], + "bbox": [ + 302, + 297, + 495, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/800ecb96940b15ebc0924d2cf61839dd9b1e78d39b3921c2e8121d0df988db7e.jpg", + "image_caption": [ + "(c) $3^{\\mathrm{rd}}$ iteration rendering", + "Figure 3. As iterations progress, the rendered images contain increasingly high-quality scene information, which subsequently improves the RF-guided deblurring network's performance in the next iteration." + ], + "image_footnote": [], + "bbox": [ + 503, + 297, + 696, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32a8a4ba4397870008a4bba02cde3e91e2098621bce9d94fc7ad87e5ac495807.jpg", + "image_caption": [ + "(d) $4^{\\mathrm{th}}$ iteration rendering" + ], + "image_footnote": [], + "bbox": [ + 702, + 297, + 897, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing to $B_{m}$ as $D_{m}^{0}$ , and the set of the deblurred images as $\\mathbf{D}^0 = \\{D_1^0,\\dots,D_M^0\\}$ .", + "bbox": [ + 89, + 470, + 482, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Radiance Field Construction", + "text_level": 1, + "bbox": [ + 89, + 513, + 346, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the radiance field construction step at the $i$ -th iteration, we first estimate the camera poses of the deblurred images from the previous step, denoted as $\\mathbf{D}^{i - 1} = \\{D_1^{i - 1},\\dots,D_M^{i - 1}\\}$ using COLMAP [36], where $i\\in [1,N]$ is the index of our iterative process. Using the estimated poses and deblurred images, we aggregate information about the target 3D scene and construct a radiance field $\\mathbf{V}^i$ . As discussed in Sec. 1, our framework can adopt diverse radiance field representations. To demonstrate the extensibility of our framework, we adopt two different representations, Plenoxels [12] and 3D Gaussians [17], in our experiments. We refer to our framework using Plenoxels and 3D Gaussians as DeepDeblurRF-P and DeepDeblurRF-G, respectively.", + "bbox": [ + 88, + 537, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that deblurred images $\\mathbf{D}^{i-1}$ from the previous step may still contain residual blur and deblurring artifacts, which can introduce corrupted information into the aggregation process. Nevertheless, since the input images have overlapping regions, other deblurred images can provide information for those regions that appear corrupted in some images. By aggregating such information from different deblurred images on the same regions, the radiance field construction step can effectively suppress corrupted information and obtain a higher-quality radiance field, from which we can render higher-quality images compared to the input deblurred im", + "bbox": [ + 88, + 734, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ages at this step.", + "bbox": [ + 513, + 470, + 624, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. RF-guided Deblurring", + "text_level": 1, + "bbox": [ + 513, + 498, + 723, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The RF-guided deblurring step deblurs the input views $\\mathbf{B}$ using the aggregated information in $\\mathbf{V}^i$ . Specifically, for each input view $B_{m}$ , we render $\\mathbf{V}^i$ to obtain a rendered image $R_{m}^{i}$ . The rendered image $R_{m}^{i}$ shares the same content as its corresponding deblurred image $D_{m}^{i - 1}$ , but contains fewer artifacts and finer details. Motivated by this, our RF-guided deblurring leverages $R_{m}^{i}$ to obtain an updated deblurred image $D_{m}^{i}$ from $B_{m}$ .", + "bbox": [ + 511, + 521, + 906, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To perform RF-guided deblurring, we employ a novel RF-guided deblurring network that takes both the rendered image $R_{m}^{i}$ and the input blurred image $B_{m}$ to guide the deblurring process with the aggregated information in $R_{m}^{i}$ . For the deblurring network, we adopt NAFNet [4] and modify its first layer to take a concatenation of the two images. Once we obtain the updated deblurred images $\\mathbf{D}^{i}$ , we return to the radiance field construction step and carry out the process for the $(i + 1)$ -th iteration, gradually enhancing the quality of the radiance field and the deblurred images.", + "bbox": [ + 511, + 643, + 908, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intermediate results of each step are visualized in Fig. 2-(b) to Fig. 2-(e). Due to the severe blur in the input blurred image, its initial deblurring result retains some residual blur. In contrast, thanks to the information aggregation occurring in the radiance field construction step, the rendered image contains more accurate details. Finally, thanks to the guidance of the rendered image, the RF-guided deblurring", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6015", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d269df9f5fcd7783dff492ad8c61ea7b71bc6ff0273d31af1d12141f1fe5fa9c.jpg", + "image_caption": [ + "Figure 4. Examples of the BlurRF-Synth and BlurRF-Real datasets. The examples show blurred views in BlurRF-Real, while the top and bottom rows in BlurRF-Synth include blurred views and their corresponding sharp views." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 620, + 262 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7794864bf3b5f9ed8280b9cfc307af75cd6e75a968ab0f2f60441b0ad3514e46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 901, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "result shows sharper and more precise details that cannot be observed in the intermediate results of previous steps.", + "bbox": [ + 89, + 319, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 visualizes rendered images at different iterations. Although the rendered image at the first iteration Fig. 3-(a) contains blurry details, they are gradually refined at each iteration (Fig. 3 (b)-(d)) due to our iterative process, which incrementally improves the quality of the radiance fields and the deblurred images. The effectiveness of our iterative process is also demonstrated by the increasing PSNR score as iterations progress.", + "bbox": [ + 89, + 349, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. BlurRF Dataset", + "text_level": 1, + "bbox": [ + 89, + 482, + 246, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. BlurRF-Synth", + "text_level": 1, + "bbox": [ + 89, + 507, + 236, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose the BlurRF-Synth dataset, the first large-scale dataset for training and evaluating novel-view synthesis from blurred images. The dataset comprises a train set and a test set for each type of blur: camera motion blur and defocus blur. The train set includes 65 scenes, each containing 29 pairs of synthetically blurred images and their corresponding sharp images, captured from different viewpoints. The test set includes 10 scenes, each with 29 blurred-sharp image pairs and five additional sharp images from different viewpoints for evaluating novel-view synthesis quality. Examples of blurred images and ground-truth sharp images are shown in Fig. 4. We synthetically generated the dataset using Blender models, collecting 95 models from Blendswap1 under Creative Commons licenses and 5 models from the synthetic dataset of Deblur-NeRF [24]. For each model, we sampled 29 camera poses to capture different viewpoints.", + "bbox": [ + 89, + 529, + 483, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera motion blur We simulated camera shakes by randomly sampling a camera trajectory for each pose during the exposure duration. To this end, we adopted Bezier interpolation to model 6-DOF camera motion and densely sampled 51 intermediate poses along this camera trajectory. We then rendered the scene using Blender at each intermediate camera pose to produce a series of sharp images, which we then", + "bbox": [ + 89, + 773, + 483, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "averaged to create a blurred image. All rendering and averaging were performed in the linear sRGB color space to accurately replicate the real-world image formation process. Among the sharp images, we sampled the temporally central image (i.e., the 26th image) as the ground-truth sharp image.", + "bbox": [ + 511, + 319, + 906, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Defocus blur We also generated images with defocus blur using Blender. We adjusted the camera's depth-of-field (DoF) in Blender to create defocus blur. We controlled the aperture size and randomly set the blade count between 7 and 9 to simulate realistic defocus effects. For each camera pose, we sampled a random focal distance within a predefined range based on the scene scale, shifting the focal plane to introduce variation in the defocus blur.", + "bbox": [ + 511, + 397, + 908, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Realistic blur synthesis To reflect real-world image degradation, we adopted the blur synthesis pipeline of RSBur [35]. Specifically, we generated camera motion blur by averaging sharp images in the linear sRGB space. Then, we synthesized saturated pixels to the blurred images and converted them to the camera RAW space. In the camera RAW space, we synthesized shot and read noise, and converted the images back to the camera sRGB space. For defocus blur, we rendered blurred images in the linear sRGB space and added noise in the same manner. For a more detailed pipeline, we refer the readers to [35]. The blur synthesis pipeline requires camera-specific noise parameters and a color correction matrix. We used a Sony A7R3 camera to estimate them. For more details about the BlurRF-Synth dataset, we refer the readers to the supplementary material.", + "bbox": [ + 511, + 520, + 908, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.BlurRF-Real", + "text_level": 1, + "bbox": [ + 511, + 757, + 651, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation under more realistic and challenging conditions, we propose BlurRF-Real, a real-world low-light camera motion blur dataset. The existing real datasets [22, 24] are captured in well-lit environments with minimal noise, making them less representative of camera motion blur in low-light conditions. Our dataset addresses this gap by providing data captured under challenging low-light scenarios. We collected five indoor scenes using a machine vision camera,", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://blendswap.com/", + "bbox": [ + 107, + 886, + 290, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6016", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e5ee1d3e301b50dd75e5f5b47959b522f6dfebfd7e036167bf3b270b5940f791.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
3D RepresentationModelCamera Motion BlurDefocus BlurComputation Time (Hr.)
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MLPDeblur-NeRF [24]27.670.83400.145030.030.87270.113731.33
BAD-NeRF [39]21.740.52980.3969---24.68
DP-NeRF [21]28.030.84120.126730.150.87630.099130.00*
Voxel gridExBluRF [22]26.560.78230.1955---8.92
PDRF-10 [31]28.330.84350.149530.030.87500.12254.26
DeepDeblurRF-P29.810.86680.114232.510.90580.09611.14
3D GaussiansDeblurring-3DGS [20]26.300.77290.172829.370.85450.14700.33
BAGS [32]27.410.81080.138229.900.86380.11521.25
DeepDeblurRF-G29.940.86810.105932.580.90600.07740.28
", + "bbox": [ + 94, + 88, + 924, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of novel-view synthesis on BlurRF-Synth test scenes. We highlight the best metrics and the second best metrics. Note that the computation time of DP-NeRF [21] was measured with two GPUs due to its memory demands, whereas those of the other models were measured on a single GPU.", + "bbox": [ + 89, + 260, + 906, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8257f61519d0fa50b13e87bcc9982ff97e69d54c59ee1c6ca76000630c7976b2.jpg", + "image_caption": [ + "Figure 5. Qualitative results of novel-view synthesis on BlurRF-Synth test scenes." + ], + "image_footnote": [], + "bbox": [ + 91, + 316, + 906, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each providing 20-40 multi-view blurry images, including 3-5 images for novel-view synthesis evaluation. Examples of blurred training views are shown in Fig. 4.", + "bbox": [ + 89, + 542, + 482, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 608, + 220, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets The deblurring networks of DeepDeblurRF are trained with the train set of BlurRF-Synth. To evaluate DeepDeblurRF and other radiance field deblurring methods, we use the test set of BlurRF-Synth, the BlurRF-Real dataset, and the benchmark datasets from Deblur-NeRF [24]. The benchmark datasets include both synthetic and real-world scenes. The synthetic dataset consists of five scenes, while the real-world dataset contains ten scenes for each blur type. The real scenes of Deblur-NeRF were captured with a Canon EOS RP in manual exposure mode. The camera poses are estimated using COLMAP [36].", + "bbox": [ + 88, + 638, + 482, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details We set the number of iterations $N$ to 5. Thus, as detailed in Sec. 3, DeepDeblurRF includes one single-image deblurring network and four RF-guided deblurring networks for each blur type: camera motion blur and defocus blur. Note that our approach requires an additional training phase for the deblurring networks, unlike", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "previous radiance field deblurring methods. Once trained, our framework can construct radiance fields for arbitrary scenes and render novel views, similar to conventional approaches. To train the deblurring networks, we use the Adam optimizer [18] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.9$ . The learning rate is initially set to $10^{-3}$ and gradually reduced to $10^{-7}$ using the cosine annealing scheduler [23].", + "bbox": [ + 511, + 542, + 906, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For constructing radiance fields in DeepDeblurRF-P, we follow the forward-facing experimental setup in Plenoxels, applying pruning and upsampling to the voxel grids every 38,400 iterations. For DeepDeblurRF-G, we adopt the original Gaussian Splatting method, incorporating the depth-based pruning and additional points strategy from DeblurGS [20] to address the sparse point clouds obtained from COLMAP on blurred images.", + "bbox": [ + 511, + 652, + 908, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We measured the computation time for radiance field deblurring using the camera motion blur test set of BlurRF-Synth on a single NVIDIA Titan RTX, except for DP-NeRF, which requires two GPUs due to its memory demands. Additional implementation details, including network configurations and the training settings for each iteration of DeepDeblurRF-P and DeepDeblurRF-G, are provided in the supplementary material.", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6017", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d73d4bdf292e3f39a5ceefcb869857eec8d9830f01e28cbbf5ff234fb5693e4.jpg", + "image_caption": [ + "Figure 6. Qualitative results of novel-view synthesis on real-world datasets." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 397 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e248613e7f5667956334566827f3bea8df7fe67fc175849a19ba2c04d81abc6.jpg", + "image_caption": [ + "Figure 7. Qualitative results of novel-view synthesis on BlurRF-SB dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 426, + 906, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison using BlurRF-Synth We compare DeepDeblurRF with other radiance field deblurring methods [20-22, 24, 31, 32, 39]. Tab. 1 shows the quantitative results of DeepDeblurRF and the other methods on the test set of BlurRF-Synth. The table demonstrates that our approach outperforms the existing methods in terms of quality, regardless of the radiance field representation. Moreover, it achieves substantially shorter computation times. Specifically, DeepDeblurRF-P shows the shortest computation time among the approaches based on MLPs and voxel grids, while DeepDeblurRF-G is the fastest among all the methods. This superior performance can be attributed to its DNN-based deblurring modules, which handle blur efficiently and effectively, in contrast to the other methods constrained by linear blur models.", + "bbox": [ + 88, + 633, + 485, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The superior performance of our approach is further demonstrated in Fig. 5, which shows that our method", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves sharper results on the BlurRF-Synth dataset. Specifically, on the camera motion blur scene, all the other methods produce noisy results, whereas both DeepDeblurRF-P and DeepDeblurRF-G effectively remove the blur even in challenging conditions with noise and pixel saturation, achieving high-quality novel-view synthesis. Similarly, in the defocus blur scene, our methods render sharper novel views by minimizing residual blur than the other methods.", + "bbox": [ + 511, + 633, + 908, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on real datasets We compare the novel-view synthesis performance of DeepDeblurRF against other approaches using both BlurRF-Real and the real scenes from Deblur-NeRF [24]. As detailed in Sec. 4.2, BlurRF-Real provides an evaluation of how well each method handles camera motion blur with non-linear outliers and noise in real-world low-light conditions. As shown in Tab. 2 and Fig. 6, our method not only outperforms the others in non-reference metrics [1, 26] but also achieves high-quality qualitative", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6018", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/639e075c27d5729651d07f95b68b2d3d025edd18a782aa4c771063713d7fc462.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBlurRF-RealBlurRF-SB
NIQE (↓)ARNIQA (↑)PSNR (↑)SSIM (↑)LPIPS (↓)
MLP
Deblur-NeRF [24]6.3410.27925.430.72640.2319
BAD-NeRF [39]9.9080.23721.970.52310.4731
DP-NeRF [21]6.4980.26627.070.79400.1795
Voxel grid
ExBluRF [22]7.4790.27924.160.66490.2659
PDRF-10 [31]6.2430.26626.660.77300.2124
DeepDeblurRF-P5.8290.32329.450.84720.1677
3D Gaussians
Deblurring-3DGS [20]6.0860.26223.870.64590.2475
BAGS [32]5.8370.29024.490.67480.2391
DeepDeblurRF-G5.4230.32929.590.85480.1399
", + "bbox": [ + 94, + 88, + 495, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results, clearly surpassing the competing methods. It also shows superior performance on real scenes from Deblur-NeRF. Specifically, while other methods struggle with saturated pixels in the camera motion blur scene, our method effectively handles it. Likewise, in the defocus blur scene, it renders sharper novel views without residual blur.", + "bbox": [ + 89, + 311, + 483, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While the real-world scenes of Deblur-NeRF provide blur-free reference images, they suffer from misalignment and exposure differences [24, 31, 32]. Similarly, BlurRF-Real also includes blur-free reference images, but they have not only similar limitations but also severe noise as they are captured under severe low-light conditions. Thus, we report quantitative evaluations against the reference images in the supplementary material. More experimental results are also provided in the supplementary material.", + "bbox": [ + 89, + 405, + 483, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input views with same blur directions As discussed in Sec. 1, unlike previous approaches, our method does not rely on the assumption that input blurred images contain complementary information from different blur directions. To verify this, we generated an additional synthetic test set, named BlurRF-SB, where the blurred images of the same scene share the same blur directions, while their blur magnitudes are different. BlurRF-SB was generated using five scenes from the BlurRF-Synth test set, following the same generation process.", + "bbox": [ + 89, + 546, + 483, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 7 shows a qualitative comparison on BlurRF-SB. Due to the lack of complementary information in different training views, the other methods fail to restore sharp details. Conversely, our methods successfully restore sharp details. The quantitative comparison in Tab. 2 also demonstrates that our approach significantly outperforms the others for blurred images with the same blur directions. These results verify the benefit of using prior knowledge on sharp images from pre-trained deblurring networks.", + "bbox": [ + 89, + 699, + 483, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of iterations We investigate the impact of the hyperparameter $N$ , which sets the number of iterations in our framework. Tab. 3 shows that DeepDeblurRF-G achieves improved novel-view synthesis as $N$ increases, highlighting", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/932a83894cf7cf6cfb528c83c30d0068480e46f4f45f46ad62d09d86396aab64.jpg", + "table_caption": [ + "Table 2. Quantitative results of novel-view synthesis on BlurRF-Real and BlurRF-SB datasets. We highlight the best metrics and the second best metrics." + ], + "table_footnote": [], + "table_body": "
#Iter.Camera motionDefocus
PSNRSSIMLPIPSPSNRSSIMLPIPS
N=128.400.81750.170730.740.88620.1297
N=229.230.84620.128131.220.89050.0955
N=329.650.85880.115232.120.90030.0899
N=429.830.86320.110832.300.90380.0860
N=529.940.86810.105932.580.90600.0774
N=629.970.86920.103432.750.90870.0736
", + "bbox": [ + 519, + 89, + 901, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative results of ablation study on the number of iterations $N$ using DeepDeblurRF-G on the test sets of BlurRF-Synth.", + "bbox": [ + 511, + 210, + 906, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d3e2eca8f11b9c8612d42165ee06a9c75fb908632ecc922d8f3ba4d49a429e1.jpg", + "image_caption": [ + "Figure 8. Qualitative results of ablation study on the number of iterations $N$ using DeepDeblurRF-G on the real-world scenes [24]." + ], + "image_footnote": [], + "bbox": [ + 516, + 257, + 906, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the effectiveness of our iterative approach. However, gains diminish beyond $N = 5$ , suggesting that $N = 5$ is sufficient. Fig. 8 further supports this with qualitative results on real scenes [24] (camera motion blur and defocus blur), showing similar trends. Similar results were observed for DeepDeblurRF-P, as detailed in the supplementary material.", + "bbox": [ + 511, + 465, + 908, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 570, + 633, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed DeepDeblurRF, a novel radiance field deblurring approach that leverages DNN-based deblurring to overcome the limitations of the linear blur model of previous methods. To integrate DNN-based deblurring with radiance field construction, we presented a novel RF-guided deblurring scheme and the framework that alternates between RF-guided deblurring and radiance field training. We also presented BlurRF-Synth, the first large-scale dataset for training and evaluating radiance field deblurring frameworks. Experimental results showed that our method outperforms others, handling both camera motion blur and defocus blur effectively with highly reduced computation time.", + "bbox": [ + 511, + 595, + 908, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work Our method leverages deblurring networks to remove blur from training views, but restoring images with more severe blur than in the training data remains challenging. Our RF-guided deblurring uses only rendered images from each input view for guidance. Imposing additional priors on other views and incorporating additional information, e.g., depth maps rendered from radiance fields, may further improve the deblurring performance.", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6019", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government(MSIT) (RS-2024-00457882, AI Research Hub Project, RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH), RS-2024-00437866, ITRC(Information Technology Research Center)).", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 208, + 187, + 223 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lorenzo Agnolucci et al. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 7", + "[2] Miika Aittala and Fredo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 3", + "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1", + "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 1, 3, 4", + "[5] Wei-Ge Chen, N Nandhakumar, and Worthy N Martin. Image motion estimation from motion smear-a new computational model. IEEE transactions on pattern analysis and machine intelligence, 18(4):412-425, 1996. 1", + "[6] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Trans. Graph., 28(5):145:1-145:8, 2009. 1, 3", + "[7] Sunghyun Cho, Yasuyuki Matsushita, and Seungyong Lee. Removing non-uniform motion blur from images. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 1, 3", + "[8] Sunghyun Cho, Jue Wang, and Seungyong Lee. Handling outliers in non-blind image deconvolution. In 2011 international conference on computer vision, pages 495-502. IEEE, 2011. 1", + "[9] Sunghyun Cho, Jue Wang, and Seungyong Lee. Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3", + "[10] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 3", + "[11] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 3", + "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2, 4" + ], + "bbox": [ + 93, + 233, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Chunzhi Gu, Xuequan Lu, Ying He, and Chao Zhang. Blur removal via blurred-noisy image pair. IEEE Transactions on Image Processing, 30:345-359, 2020. 3", + "[14] Michael Hirsch, Christian J Schuler, Stefan Harmeling, and Bernhard Schölkopf. Fast removal of non-uniform camera shake. In 2011 International Conference on Computer Vision, pages 463-470. IEEE, 2011. 1, 3", + "[15] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 1", + "[16] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE international conference on computer vision, pages 4038-4047, 2017. 3", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 1, 2, 3, 4", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[19] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 1, 3", + "[20] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 1, 2, 3, 6, 7, 8", + "[21] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 1, 2, 3, 6, 8", + "[22] Dongwoo Lee, Jeongtaek Oh, Jaesung Rim, Sunghyun Cho, and Kyoung Mu Lee. Exblurf: Efficient radiance fields for extreme motion blurred images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17639-17648, 2023. 1, 2, 3, 5, 6, 7, 8", + "[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[24] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1, 2, 3, 5, 6, 7, 8", + "[25] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1", + "[26] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 7" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6020", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1", + "[28] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3", + "[29] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8102-8111, 2019. 3", + "[30] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2901–2908, 2014. 1, 3", + "[31] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 1, 2, 3, 6, 7, 8", + "[32] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling, 2024. 1, 2, 3, 6, 7, 8", + "[33] Alex Rav-Acha and Shmuel Peleg. Two motion-blurred images are better than one. Pattern recognition letters, 26(3): 311-317, 2005. 1, 3", + "[34] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1", + "[35] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 5", + "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6", + "[37] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3", + "[38] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3", + "[39] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 1, 2, 3, 6, 7, 8", + "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. In 2010" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 491-498, 2010. 1, 3", + "[41] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3", + "[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 3", + "[43] Haichao Zhang and David Wipf. Non-uniform camera shake removal using a spatially-adaptive sparse penalty. Advances in Neural Information Processing Systems, 26, 2013. 1, 3", + "[44] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2482-2491, 2019. 3", + "[45] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, Haozhe Xie, Jinshan Pan, and Jimmy S Ren. Davanet: Stereo deblurring with view aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10996-11005, 2019. 3" + ], + "bbox": [ + 516, + 92, + 906, + 430 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6021", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_model.json b/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_model.json new file mode 100644 index 0000000000000000000000000000000000000000..edf74e9ae8ca3a14cde1792498d3ae6908cfd887 --- /dev/null +++ b/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "Exploiting Deblurring Networks for Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.18, + 0.301, + 0.2 + ], + "angle": 0, + "content": "Haeyun Choi\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.18, + 0.468, + 0.2 + ], + "angle": 0, + "content": "Heemin Yang" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.18, + 0.648, + 0.2 + ], + "angle": 0, + "content": "Janghyeok Han²" + }, + { + "type": "text", + "bbox": [ + 0.695, + 0.18, + 0.824, + 0.2 + ], + "angle": 0, + "content": "Sunghyun Cho²" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.216, + 0.441, + 0.233 + ], + "angle": 0, + "content": "\\({}^{1}\\mathrm{KT}\\)" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.216, + 0.592, + 0.233 + ], + "angle": 0, + "content": "\\(^{2}\\)POSTECH" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.271, + 0.329, + 0.287 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.304, + 0.486, + 0.591 + ], + "angle": 0, + "content": "In this paper, we propose DeepDeblurRF, a novel radiance field deblurring approach that can synthesize high-quality novel views from blurred training views with significantly reduced training time. DeepDeblurRF leverages deep neural network (DNN)-based deblurring modules to enjoy their deblurring performance and computational efficiency. To effectively combine DNN-based deblurring and radiance field construction, we propose a novel radiance field (RF)-guided deblurring and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Moreover, DeepDeblurRF is compatible with various scene representations, such as voxel grids and 3D Gaussians, expanding its applicability. We also present BlurRF-Synth, the first large-scale synthetic dataset for training radiance field deblurring frameworks. We conduct extensive experiments on both camera motion blur and defocus blur, demonstrating that DeepDeblurRF achieves state-of-the-art novel-view synthesis quality with significantly reduced training time." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.622, + 0.222, + 0.637 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.485, + 0.828 + ], + "angle": 0, + "content": "Novel-view synthesis has seen significant advancements in recent years, leading to impressive improvements in photorealistic rendering quality. One key development in this field is the Neural Radiance Field (NeRF) [25], which leverages neural networks to generate highly detailed images from novel viewpoints. Building on this, various works have focused on reducing training time and enhancing rendering quality [3, 12, 15, 27]. More recently, 3D Gaussian Splatting (3DGS) [17] has emerged, utilizing Gaussians as an explicit 3D representation. This technique enables high-quality scene reconstruction and real-time rendering through a differentiable rasterization method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.485, + 0.875 + ], + "angle": 0, + "content": "However, synthesizing sharp novel views from degraded training views remains a significant challenge. Capturing images in real-world conditions often leads to various degra" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.908, + 0.395 + ], + "angle": 0, + "content": "ditions such as blur and noise. Among them, blur makes it difficult to aggregate accurate 3D information from the training views, resulting in a blurry radiance field. To address blur, several works [20-22, 24, 31, 32, 39] have been proposed to reconstruct sharp radiance fields from training views with either camera motion blur or defocus blur. These radiance field deblurring methods jointly optimize blur kernels and the radiance field, and have shown promising results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.909, + 0.732 + ], + "angle": 0, + "content": "However, these methods still possess several limitations that hinder their effectiveness. First, existing approaches rely on linear blur models, describing blurred pixels in training views as linear combinations of sharp pixels. In contrast, real-world blurred images often suffer from non-linear outliers such as saturated pixels due to sensor clipping, noise from various sources, and non-linear in-camera processing [8, 35]. These violate the linear blur model, leading to deblurring artifacts and severely degrading the performance of current radiance field deblurring techniques. Second, existing approaches do not utilize priors on sharp images, relying solely on complementary information from different views to model blur kernels and estimate a sharp radiance field. Consequently, they tend to produce radiance fields with residual blur and may completely fail when all input views exhibit similar blur directions [24], akin to classical multi-frame deblurring approaches [5, 7, 33], due to the lack of sharp image priors. Third, existing ray-based methods [21, 22, 24, 31, 39] employ multiple ray samples per pixel to depict a blurry image, resulting in a substantial increase in rendering time for a single view from a radiance field, and consequently, a significant computation time for radiance field construction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Apart from the radiance field deblurring methods, image deblurring has been extensively studied for decades [6, 14, 30, 40, 43]. Recently, a number of deep neural network (DNN)-based single-image deblurring methods have been proposed [4, 10, 19, 28, 38, 42]. These methods do not rely on linear blur models, but learn image deblurring from large-scale datasets. Thanks to this, they can effectively remove blur from a single image without the need for complementary information from other images, and also handle nonlinearities such as saturated pixels [34]. Moreover, their feedforward approach allows markedly reduced processing times," + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.887, + 0.308, + 0.9 + ], + "angle": 0, + "content": "†This work was done at POSTECH." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6012" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.237, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.191, + 0.215, + 0.201 + ], + "angle": 0, + "content": "Blurry Input View" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.089, + 0.373, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.191, + 0.361, + 0.201 + ], + "angle": 0, + "content": "Deblur-NeRF (31 hrs)" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.09, + 0.511, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.191, + 0.485, + 0.2 + ], + "angle": 0, + "content": "PDRF-10 (4 hrs)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.202, + 0.237, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.273, + 0.237, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.303, + 0.212, + 0.313 + ], + "angle": 0, + "content": "BAGS (75 mins)" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.202, + 0.374, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.273, + 0.364, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.302, + 0.375, + 0.313 + ], + "angle": 0, + "content": "DeepDeblurRF-G (17 mins)" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.202, + 0.512, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.273, + 0.471, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.303, + 0.468, + 0.313 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.094, + 0.898, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.326, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 1. Given a set of multi-view blurry images even with non-linear outliers such as saturated pixels and noise, DeepDeblurRF performs high-quality novel-view synthesis with highly efficient training. DeepDeblurRF-P and DeepDeblurRF-G are the results of our framework, where radiance fields are constructed using Plenoxels [12] and 3D Gaussian Splatting [17], respectively. Note that DP-NeRF [21] was trained with two GPUs due to its memory demands, whereas other models were trained on a single NVIDIA TITAN RTX GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.465, + 0.413 + ], + "angle": 0, + "content": "unlike classical iterative optimization-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Given the effectiveness of recent DNN-based deblurring approaches, a promising direction to construct a sharp radiance field from blurry images would be to combine a deblurring network and radiance field construction. However, a naive combination of a deblurring network and radiance field construction, which performs single-image deblurring to each input blurry image and trains a radiance field using the deblurred images, results in unsatisfactory results as reported by Ma et al. [24]. This is because of the limited performance of single-image deblurring, which stems from the insufficient information available in a single image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.484, + 0.734 + ], + "angle": 0, + "content": "In this paper, we propose DeepDeblurRF, a novel radiance field deblurring framework that enables highly efficient training and high-quality novel-view synthesis from images blurred by motion blur or defocus blur, even in the presence of noise and saturated pixels, as shown in Fig. 1. Unlike existing radiance field deblurring approaches, our method leverages DNN-based deblurring modules to achieve superior performance and computational efficiency. Additionally, our approach is versatile, capable of handling various 3D representations such as voxel grids [12] and 3D Gaussians [17]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "However, solely relying on deblurring networks to remove blur from input blurry images may yield unsatisfactory results as discussed above. To effectively combine DNN-based deblurring and radiance field construction, our framework adopts a novel radiance field (RF)-guided deblurring scheme, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Specifically, our framework first deblurs input images, and then trains a radiance field using the deblurred images. Then, at the next iteration, it renders images corresponding to the input views from the trained radiance field. While the de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.549 + ], + "angle": 0, + "content": "blurred images may have limited qualities with inaccurately restored details, the rendered images provide higher-quality details as they are rendered from aggregated information from multiple deblurred images. Using the rendered images as guidance, our framework then performs RF-guided deblurring to the input blurred images. Thanks to the guidance, we can obtain higher-quality deblurred images, which we use for radiance field construction again. Finally, our framework iterates RF-guided deblurring and radiance field construction to gradually enhance the quality of the radiance field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Training the deblurring modules of our framework requires a large-scale dataset of blurred images paired with ground-truth sharp images for radiance field reconstruction. However, no such datasets are available as existing radiance field deblurring approaches adopt linear blur model-based approaches, which do not require large-scale training datasets [20-22, 24, 31, 32, 39]. Consequently, the datasets provided by previous works, such as those from DeblurNeRF [24] and ExBluRF [22], are small and primarily designed for evaluation rather than training. Furthermore, these datasets often overlook crucial factors like camera noise, in-camera processing, and other non-linear outliers, which are essential for realistic deblurring tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.901 + ], + "angle": 0, + "content": "To address these gaps, we propose BlurRF-Synth, a large-scale dataset for radiance field deblurring approaches. It includes 4,350 blurred-sharp image pairs across 150 scenes, with 2,175 pairs for each of 75 scenes, encompassing both camera motion blur and defocus blur. These images are carefully synthesized to reflect real-world camera degradations such as noise, saturated pixels, and in-camera processing pipelines. Additionally, we introduce a real-world dataset called BlurRF-Real for evaluation in non-ideal conditions. Unlike the Deblur-NeRF dataset, which has less noise and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6013" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "adequate lighting, BlurRF-Real consists of five low-light indoor scenes captured with a machine vision camera." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.257 + ], + "angle": 0, + "content": "We validate DeepDeblurRF on both synthetic and real-world datasets with two types of blur: camera motion blur and defocus blur. Experimental results demonstrate that our method achieves state-of-the-art novel-view synthesis performance with highly reduced processing times. Furthermore, we demonstrate the extensibility of our framework by constructing radiance fields using different scene representations, such as voxel grids and Gaussian Splatting. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.315 + ], + "angle": 0, + "content": "- We propose DeepDeblurRF, a novel radiance field deblurring framework, which is the first approach that leverages DNN-based deblurring modules to overcome the limitations of the linear blur model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.318, + 0.481, + 0.376 + ], + "angle": 0, + "content": "- To this end, we present RF-guided deblurring, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner to gradually enhance the quality of the radiance field." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.379, + 0.482, + 0.422 + ], + "angle": 0, + "content": "- We also present the BlurRF-Synth dataset, the first large-scale dataset for training and evaluation of novel-view synthesis from blurry images." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.233, + 0.451 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Radiance field deblurring Several works have recently been proposed to tackle radiance field deblurring [21, 22, 24, 31, 39]. Ma et al. presented Deblur-NeRF, the first framework to train a sharp radiance field from blurry training views [24]. Deblur-NeRF adopts an additional MLP to model blur kernels and jointly optimizes them with a radiance field. DP-NeRF [21] models blur kernels based on physical priors for rigid motions, while PDRF [31] uses a two-stage kernel modeling scheme where the first stage coarsely aggregates kernel points to reduce computational cost. All these methods support both camera motion blur and defocus blur by modeling them using multiple rays. On the other hand, BAD-NeRF [39] and ExBluRF [22] propose radiance field deblurring for camera motion blur, modeling camera motion blur using camera trajectories. Nevertheless, all these methods rely on linear blur models, and require multiple ray-tracing steps, resulting in limited deblurring quality and slow training speeds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "More recently, with the emergence of 3D Gaussian Splating [17], 3D Gaussian-based radiance field deblurring methods have been proposed. Deblurring-3DGS [20] handles camera motion blur with small offsets in 3D Gaussian positions and defocus blur by adjusting the geometry of Gaussians. BAGS [32] models both types of blur using dense per-pixel blur kernels optimized in a coarse-to-fine manner. However, BAGS requires substantial computation time to optimize dense blur kernels. Moreover, these methods also rely on linear blur models and are limited in handling real-world blurred images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.272 + ], + "angle": 0, + "content": "Image deblurring Image deblurring has been widely studied to address blur that degrades image quality. Traditional methods, relying on linear blur models, often struggle with real-world blurred images containing non-linear outliers [6, 7, 9, 14, 30, 33, 40, 43]. Their iterative optimization processes are also computationally intensive, limiting their practical use. Recent advancements in deep learning have introduced DNN-based approaches, including single-image [4, 10, 19, 28, 38, 42] and multi-frame/video deblurring [2, 11, 13, 16, 29, 37, 41, 44, 45]. These methods can handle more complex blurs and reduce processing time with a feed-forward structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.909, + 0.424 + ], + "angle": 0, + "content": "While applying DNN-based deblurring approaches might be a potential solution for constructing radiance fields from blurry images, naively using single-image deblurring techniques yields unsatisfactory results as discussed in Sec. 1. Another possible direction is to adopt multi-frame and video deblurring approaches that leverage complementary information from multiple images for higher-quality results. However, these methods assume that input images are captured from nearly the same viewpoints, a condition that does not hold for training views in radiance field construction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.667, + 0.454 + ], + "angle": 0, + "content": "3. DeepDeblurRF" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Fig. 2-(a) illustrates the overall framework of DeepDeblurRF. It takes a set of \\(M\\) blurred images \\(\\mathbf{B} = \\{B_1, \\dots, B_M\\}\\) of a scene and estimates a sharp radiance field that can synthesize a sharp novel view given an arbitrary pose. We assume that each blurred image \\(B_m\\) is obtained from its latent sharp image \\(L_m\\) through an unknown degradation process including blur and non-linear outliers such as saturated pixels and nonlinear in-camera processing. Based on this assumption, DeepDeblurRF first performs initial deblurring to the input blurred images, and obtains initial deblurred images. Then, our method iteratively performs radiance field construction using deblurred images and RF-guided deblurring to gradually enhance the quality of the radiance field and the deblurred images. At the last iteration, we perform only the radiance field construction step and obtain a final radiance field, from which we can synthesize sharp novel views. In the following, each step of DeepDeblurRF is described in more detail." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.688, + 0.759 + ], + "angle": 0, + "content": "3.1. Initial Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The initial deblurring step removes blur from each blurry training view \\( B_{m} \\in \\mathbf{B} \\) so that the following radiance field construction step can estimate the pose of each view more accurately, and more effectively aggregate information from different views. To this end, we adopt an off-the-shelf single-image deblurring network. Specifically, we adopt NAFNet [4], a state-of-the-art deblurring network, for its computational efficiency and performance. We denote the deblurred image from the initial deblurring step correspond" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6014" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.094, + 0.596, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.094, + 0.751, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.17, + 0.738, + 0.181 + ], + "angle": 0, + "content": "(b) Blurry training view" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.182, + 0.752, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.255, + 0.741, + 0.267 + ], + "angle": 0, + "content": "(d) \\(1^{\\mathrm{st}}\\) iteration rendering" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.095, + 0.897, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.17, + 0.874, + 0.181 + ], + "angle": 0, + "content": "(c) Initial deblurring" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.182, + 0.897, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.255, + 0.892, + 0.267 + ], + "angle": 0, + "content": "(e) \\(1^{\\mathrm{st}}\\) RF-guided deblurring" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.28, + 0.747, + 0.295 + ], + "angle": 0, + "content": "Figure 2. Overall framework and intermediate result of each step of DeepDeblurRF." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.298, + 0.296, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.402, + 0.276, + 0.416 + ], + "angle": 0, + "content": "(a) \\(1^{\\mathrm{st}}\\) iteration rendering" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.298, + 0.496, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.401, + 0.479, + 0.416 + ], + "angle": 0, + "content": "(b) \\(2^{\\mathrm{nd}}\\) iteration rendering" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.298, + 0.697, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.401, + 0.678, + 0.416 + ], + "angle": 0, + "content": "(c) \\(3^{\\mathrm{rd}}\\) iteration rendering" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.298, + 0.898, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.401, + 0.878, + 0.416 + ], + "angle": 0, + "content": "(d) \\(4^{\\mathrm{th}}\\) iteration rendering" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.427, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 3. As iterations progress, the rendered images contain increasingly high-quality scene information, which subsequently improves the RF-guided deblurring network's performance in the next iteration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.503 + ], + "angle": 0, + "content": "ing to \\(B_{m}\\) as \\(D_{m}^{0}\\), and the set of the deblurred images as \\(\\mathbf{D}^0 = \\{D_1^0,\\dots,D_M^0\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.514, + 0.348, + 0.529 + ], + "angle": 0, + "content": "3.2. Radiance Field Construction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.485, + 0.734 + ], + "angle": 0, + "content": "In the radiance field construction step at the \\(i\\)-th iteration, we first estimate the camera poses of the deblurred images from the previous step, denoted as \\(\\mathbf{D}^{i - 1} = \\{D_1^{i - 1},\\dots,D_M^{i - 1}\\}\\) using COLMAP [36], where \\(i\\in [1,N]\\) is the index of our iterative process. Using the estimated poses and deblurred images, we aggregate information about the target 3D scene and construct a radiance field \\(\\mathbf{V}^i\\). As discussed in Sec. 1, our framework can adopt diverse radiance field representations. To demonstrate the extensibility of our framework, we adopt two different representations, Plenoxels [12] and 3D Gaussians [17], in our experiments. We refer to our framework using Plenoxels and 3D Gaussians as DeepDeblurRF-P and DeepDeblurRF-G, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Note that deblurred images \\(\\mathbf{D}^{i-1}\\) from the previous step may still contain residual blur and deblurring artifacts, which can introduce corrupted information into the aggregation process. Nevertheless, since the input images have overlapping regions, other deblurred images can provide information for those regions that appear corrupted in some images. By aggregating such information from different deblurred images on the same regions, the radiance field construction step can effectively suppress corrupted information and obtain a higher-quality radiance field, from which we can render higher-quality images compared to the input deblurred im" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.472, + 0.625, + 0.487 + ], + "angle": 0, + "content": "ages at this step." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.499, + 0.724, + 0.515 + ], + "angle": 0, + "content": "3.3. RF-guided Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.644 + ], + "angle": 0, + "content": "The RF-guided deblurring step deblurs the input views \\(\\mathbf{B}\\) using the aggregated information in \\(\\mathbf{V}^i\\). Specifically, for each input view \\(B_{m}\\), we render \\(\\mathbf{V}^i\\) to obtain a rendered image \\(R_{m}^{i}\\). The rendered image \\(R_{m}^{i}\\) shares the same content as its corresponding deblurred image \\(D_{m}^{i - 1}\\), but contains fewer artifacts and finer details. Motivated by this, our RF-guided deblurring leverages \\(R_{m}^{i}\\) to obtain an updated deblurred image \\(D_{m}^{i}\\) from \\(B_{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.795 + ], + "angle": 0, + "content": "To perform RF-guided deblurring, we employ a novel RF-guided deblurring network that takes both the rendered image \\( R_{m}^{i} \\) and the input blurred image \\( B_{m} \\) to guide the deblurring process with the aggregated information in \\( R_{m}^{i} \\). For the deblurring network, we adopt NAFNet [4] and modify its first layer to take a concatenation of the two images. Once we obtain the updated deblurred images \\( \\mathbf{D}^{i} \\), we return to the radiance field construction step and carry out the process for the \\( (i + 1) \\)-th iteration, gradually enhancing the quality of the radiance field and the deblurred images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The intermediate results of each step are visualized in Fig. 2-(b) to Fig. 2-(e). Due to the severe blur in the input blurred image, its initial deblurring result retains some residual blur. In contrast, thanks to the information aggregation occurring in the radiance field construction step, the rendered image contains more accurate details. Finally, thanks to the guidance of the rendered image, the RF-guided deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6015" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.621, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.902, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Examples of the BlurRF-Synth and BlurRF-Real datasets. The examples show blurred views in BlurRF-Real, while the top and bottom rows in BlurRF-Synth include blurred views and their corresponding sharp views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.35 + ], + "angle": 0, + "content": "result shows sharper and more precise details that cannot be observed in the intermediate results of previous steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.484, + 0.471 + ], + "angle": 0, + "content": "Fig. 3 visualizes rendered images at different iterations. Although the rendered image at the first iteration Fig. 3-(a) contains blurry details, they are gradually refined at each iteration (Fig. 3 (b)-(d)) due to our iterative process, which incrementally improves the quality of the radiance fields and the deblurred images. The effectiveness of our iterative process is also demonstrated by the increasing PSNR score as iterations progress." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.248, + 0.499 + ], + "angle": 0, + "content": "4. BlurRF Dataset" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.508, + 0.238, + 0.525 + ], + "angle": 0, + "content": "4.1. BlurRF-Synth" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.773 + ], + "angle": 0, + "content": "We propose the BlurRF-Synth dataset, the first large-scale dataset for training and evaluating novel-view synthesis from blurred images. The dataset comprises a train set and a test set for each type of blur: camera motion blur and defocus blur. The train set includes 65 scenes, each containing 29 pairs of synthetically blurred images and their corresponding sharp images, captured from different viewpoints. The test set includes 10 scenes, each with 29 blurred-sharp image pairs and five additional sharp images from different viewpoints for evaluating novel-view synthesis quality. Examples of blurred images and ground-truth sharp images are shown in Fig. 4. We synthetically generated the dataset using Blender models, collecting 95 models from Blendswap1 under Creative Commons licenses and 5 models from the synthetic dataset of Deblur-NeRF [24]. For each model, we sampled 29 camera poses to capture different viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Camera motion blur We simulated camera shakes by randomly sampling a camera trajectory for each pose during the exposure duration. To this end, we adopted Bezier interpolation to model 6-DOF camera motion and densely sampled 51 intermediate poses along this camera trajectory. We then rendered the scene using Blender at each intermediate camera pose to produce a series of sharp images, which we then" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.907, + 0.396 + ], + "angle": 0, + "content": "averaged to create a blurred image. All rendering and averaging were performed in the linear sRGB color space to accurately replicate the real-world image formation process. Among the sharp images, we sampled the temporally central image (i.e., the 26th image) as the ground-truth sharp image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.909, + 0.518 + ], + "angle": 0, + "content": "Defocus blur We also generated images with defocus blur using Blender. We adjusted the camera's depth-of-field (DoF) in Blender to create defocus blur. We controlled the aperture size and randomly set the blade count between 7 and 9 to simulate realistic defocus effects. For each camera pose, we sampled a random focal distance within a predefined range based on the scene scale, shifting the focal plane to introduce variation in the defocus blur." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.909, + 0.749 + ], + "angle": 0, + "content": "Realistic blur synthesis To reflect real-world image degradation, we adopted the blur synthesis pipeline of RSBur [35]. Specifically, we generated camera motion blur by averaging sharp images in the linear sRGB space. Then, we synthesized saturated pixels to the blurred images and converted them to the camera RAW space. In the camera RAW space, we synthesized shot and read noise, and converted the images back to the camera sRGB space. For defocus blur, we rendered blurred images in the linear sRGB space and added noise in the same manner. For a more detailed pipeline, we refer the readers to [35]. The blur synthesis pipeline requires camera-specific noise parameters and a color correction matrix. We used a Sony A7R3 camera to estimate them. For more details about the BlurRF-Synth dataset, we refer the readers to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.652, + 0.773 + ], + "angle": 0, + "content": "4.2.BlurRF-Real" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For evaluation under more realistic and challenging conditions, we propose BlurRF-Real, a real-world low-light camera motion blur dataset. The existing real datasets [22, 24] are captured in well-lit environments with minimal noise, making them less representative of camera motion blur in low-light conditions. Our dataset addresses this gap by providing data captured under challenging low-light scenarios. We collected five indoor scenes using a machine vision camera," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.292, + 0.901 + ], + "angle": 0, + "content": "1https://blendswap.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6016" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.925, + 0.25 + ], + "angle": 0, + "content": "
3D RepresentationModelCamera Motion BlurDefocus BlurComputation Time (Hr.)
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MLPDeblur-NeRF [24]27.670.83400.145030.030.87270.113731.33
BAD-NeRF [39]21.740.52980.3969---24.68
DP-NeRF [21]28.030.84120.126730.150.87630.099130.00*
Voxel gridExBluRF [22]26.560.78230.1955---8.92
PDRF-10 [31]28.330.84350.149530.030.87500.12254.26
DeepDeblurRF-P29.810.86680.114232.510.90580.09611.14
3D GaussiansDeblurring-3DGS [20]26.300.77290.172829.370.85450.14700.33
BAGS [32]27.410.81080.138229.900.86380.11521.25
DeepDeblurRF-G29.940.86810.105932.580.90600.07740.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Table 1. Quantitative results of novel-view synthesis on BlurRF-Synth test scenes. We highlight the best metrics and the second best metrics. Note that the computation time of DP-NeRF [21] was measured with two GPUs due to its memory demands, whereas those of the other models were measured on a single GPU." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.318, + 0.907, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.502, + 0.741, + 0.516 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of novel-view synthesis on BlurRF-Synth test scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.483, + 0.589 + ], + "angle": 0, + "content": "each providing 20-40 multi-view blurry images, including 3-5 images for novel-view synthesis evaluation. Examples of blurred training views are shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.609, + 0.222, + 0.627 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.639, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Datasets The deblurring networks of DeepDeblurRF are trained with the train set of BlurRF-Synth. To evaluate DeepDeblurRF and other radiance field deblurring methods, we use the test set of BlurRF-Synth, the BlurRF-Real dataset, and the benchmark datasets from Deblur-NeRF [24]. The benchmark datasets include both synthetic and real-world scenes. The synthetic dataset consists of five scenes, while the real-world dataset contains ten scenes for each blur type. The real scenes of Deblur-NeRF were captured with a Canon EOS RP in manual exposure mode. The camera poses are estimated using COLMAP [36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation details We set the number of iterations \\(N\\) to 5. Thus, as detailed in Sec. 3, DeepDeblurRF includes one single-image deblurring network and four RF-guided deblurring networks for each blur type: camera motion blur and defocus blur. Note that our approach requires an additional training phase for the deblurring networks, unlike" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.65 + ], + "angle": 0, + "content": "previous radiance field deblurring methods. Once trained, our framework can construct radiance fields for arbitrary scenes and render novel views, similar to conventional approaches. To train the deblurring networks, we use the Adam optimizer [18] with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.9\\). The learning rate is initially set to \\(10^{-3}\\) and gradually reduced to \\(10^{-7}\\) using the cosine annealing scheduler [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.909, + 0.776 + ], + "angle": 0, + "content": "For constructing radiance fields in DeepDeblurRF-P, we follow the forward-facing experimental setup in Plenoxels, applying pruning and upsampling to the voxel grids every 38,400 iterations. For DeepDeblurRF-G, we adopt the original Gaussian Splatting method, incorporating the depth-based pruning and additional points strategy from DeblurGS [20] to address the sparse point clouds obtained from COLMAP on blurred images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "We measured the computation time for radiance field deblurring using the camera motion blur test set of BlurRF-Synth on a single NVIDIA Titan RTX, except for DP-NeRF, which requires two GPUs due to its memory demands. Additional implementation details, including network configurations and the training settings for each iteration of DeepDeblurRF-P and DeepDeblurRF-G, are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6017" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.905, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.408, + 0.722, + 0.421 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of novel-view synthesis on real-world datasets." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.428, + 0.907, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.602, + 0.722, + 0.616 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of novel-view synthesis on BlurRF-SB dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.635, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Comparison using BlurRF-Synth We compare DeepDeblurRF with other radiance field deblurring methods [20-22, 24, 31, 32, 39]. Tab. 1 shows the quantitative results of DeepDeblurRF and the other methods on the test set of BlurRF-Synth. The table demonstrates that our approach outperforms the existing methods in terms of quality, regardless of the radiance field representation. Moreover, it achieves substantially shorter computation times. Specifically, DeepDeblurRF-P shows the shortest computation time among the approaches based on MLPs and voxel grids, while DeepDeblurRF-G is the fastest among all the methods. This superior performance can be attributed to its DNN-based deblurring modules, which handle blur efficiently and effectively, in contrast to the other methods constrained by linear blur models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The superior performance of our approach is further demonstrated in Fig. 5, which shows that our method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.909, + 0.756 + ], + "angle": 0, + "content": "achieves sharper results on the BlurRF-Synth dataset. Specifically, on the camera motion blur scene, all the other methods produce noisy results, whereas both DeepDeblurRF-P and DeepDeblurRF-G effectively remove the blur even in challenging conditions with noise and pixel saturation, achieving high-quality novel-view synthesis. Similarly, in the defocus blur scene, our methods render sharper novel views by minimizing residual blur than the other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Comparison on real datasets We compare the novel-view synthesis performance of DeepDeblurRF against other approaches using both BlurRF-Real and the real scenes from Deblur-NeRF [24]. As detailed in Sec. 4.2, BlurRF-Real provides an evaluation of how well each method handles camera motion blur with non-linear outliers and noise in real-world low-light conditions. As shown in Tab. 2 and Fig. 6, our method not only outperforms the others in non-reference metrics [1, 26] but also achieves high-quality qualitative" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6018" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.496, + 0.236 + ], + "angle": 0, + "content": "
ModelBlurRF-RealBlurRF-SB
NIQE (↓)ARNIQA (↑)PSNR (↑)SSIM (↑)LPIPS (↓)
MLP
Deblur-NeRF [24]6.3410.27925.430.72640.2319
BAD-NeRF [39]9.9080.23721.970.52310.4731
DP-NeRF [21]6.4980.26627.070.79400.1795
Voxel grid
ExBluRF [22]7.4790.27924.160.66490.2659
PDRF-10 [31]6.2430.26626.660.77300.2124
DeepDeblurRF-P5.8290.32329.450.84720.1677
3D Gaussians
Deblurring-3DGS [20]6.0860.26223.870.64590.2475
BAGS [32]5.8370.29024.490.67480.2391
DeepDeblurRF-G5.4230.32929.590.85480.1399
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.246, + 0.486, + 0.293 + ], + "angle": 0, + "content": "Table 2. Quantitative results of novel-view synthesis on BlurRF-Real and BlurRF-SB datasets. We highlight the best metrics and the second best metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.403 + ], + "angle": 0, + "content": "results, clearly surpassing the competing methods. It also shows superior performance on real scenes from Deblur-NeRF. Specifically, while other methods struggle with saturated pixels in the camera motion blur scene, our method effectively handles it. Likewise, in the defocus blur scene, it renders sharper novel views without residual blur." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.542 + ], + "angle": 0, + "content": "While the real-world scenes of Deblur-NeRF provide blur-free reference images, they suffer from misalignment and exposure differences [24, 31, 32]. Similarly, BlurRF-Real also includes blur-free reference images, but they have not only similar limitations but also severe noise as they are captured under severe low-light conditions. Thus, we report quantitative evaluations against the reference images in the supplementary material. More experimental results are also provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.698 + ], + "angle": 0, + "content": "Input views with same blur directions As discussed in Sec. 1, unlike previous approaches, our method does not rely on the assumption that input blurred images contain complementary information from different blur directions. To verify this, we generated an additional synthetic test set, named BlurRF-SB, where the blurred images of the same scene share the same blur directions, while their blur magnitudes are different. BlurRF-SB was generated using five scenes from the BlurRF-Synth test set, following the same generation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Fig. 7 shows a qualitative comparison on BlurRF-SB. Due to the lack of complementary information in different training views, the other methods fail to restore sharp details. Conversely, our methods successfully restore sharp details. The quantitative comparison in Tab. 2 also demonstrates that our approach significantly outperforms the others for blurred images with the same blur directions. These results verify the benefit of using prior knowledge on sharp images from pre-trained deblurring networks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Number of iterations We investigate the impact of the hyperparameter \\(N\\), which sets the number of iterations in our framework. Tab. 3 shows that DeepDeblurRF-G achieves improved novel-view synthesis as \\(N\\) increases, highlighting" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.09, + 0.903, + 0.201 + ], + "angle": 0, + "content": "
#Iter.Camera motionDefocus
PSNRSSIMLPIPSPSNRSSIMLPIPS
N=128.400.81750.170730.740.88620.1297
N=229.230.84620.128131.220.89050.0955
N=329.650.85880.115232.120.90030.0899
N=429.830.86320.110832.300.90380.0860
N=529.940.86810.105932.580.90600.0774
N=629.970.86920.103432.750.90870.0736
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.211, + 0.908, + 0.253 + ], + "angle": 0, + "content": "Table 3. Quantitative results of ablation study on the number of iterations \\(N\\) using DeepDeblurRF-G on the test sets of BlurRF-Synth." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.258, + 0.907, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.422, + 0.908, + 0.451 + ], + "angle": 0, + "content": "Figure 8. Qualitative results of ablation study on the number of iterations \\(N\\) using DeepDeblurRF-G on the real-world scenes [24]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.909, + 0.557 + ], + "angle": 0, + "content": "the effectiveness of our iterative approach. However, gains diminish beyond \\( N = 5 \\), suggesting that \\( N = 5 \\) is sufficient. Fig. 8 further supports this with qualitative results on real scenes [24] (camera motion blur and defocus blur), showing similar trends. Similar results were observed for DeepDeblurRF-P, as detailed in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.634, + 0.587 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.778 + ], + "angle": 0, + "content": "In this paper, we proposed DeepDeblurRF, a novel radiance field deblurring approach that leverages DNN-based deblurring to overcome the limitations of the linear blur model of previous methods. To integrate DNN-based deblurring with radiance field construction, we presented a novel RF-guided deblurring scheme and the framework that alternates between RF-guided deblurring and radiance field training. We also presented BlurRF-Synth, the first large-scale dataset for training and evaluating radiance field deblurring frameworks. Experimental results showed that our method outperforms others, handling both camera motion blur and defocus blur effectively with highly reduced computation time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Limitations and future work Our method leverages deblurring networks to remove blur from training views, but restoring images with more severe blur than in the training data remains challenging. Our RF-guided deblurring uses only rendered images from each input view for guidance. Imposing additional priors on other views and incorporating additional information, e.g., depth maps rendered from radiance fields, may further improve the deblurring performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government(MSIT) (RS-2024-00457882, AI Research Hub Project, RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH), RS-2024-00437866, ITRC(Information Technology Research Center))." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.209, + 0.188, + 0.224 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[1] Lorenzo Agnolucci et al. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.29, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[2] Miika Aittala and Fredo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.485, + 0.455 + ], + "angle": 0, + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[5] Wei-Ge Chen, N Nandhakumar, and Worthy N Martin. Image motion estimation from motion smear-a new computational model. IEEE transactions on pattern analysis and machine intelligence, 18(4):412-425, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[6] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Trans. Graph., 28(5):145:1-145:8, 2009. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[7] Sunghyun Cho, Yasuyuki Matsushita, and Seungyong Lee. Removing non-uniform motion blur from images. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[8] Sunghyun Cho, Jue Wang, and Seungyong Lee. Handling outliers in non-blind image deconvolution. In 2011 international conference on computer vision, pages 495-502. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[9] Sunghyun Cho, Jue Wang, and Seungyong Lee. Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.693, + 0.485, + 0.761 + ], + "angle": 0, + "content": "[10] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[11] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.234, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[13] Chunzhi Gu, Xuequan Lu, Ying He, and Chao Zhang. Blur removal via blurred-noisy image pair. IEEE Transactions on Image Processing, 30:345-359, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[14] Michael Hirsch, Christian J Schuler, Stefan Harmeling, and Bernhard Schölkopf. Fast removal of non-uniform camera shake. In 2011 International Conference on Computer Vision, pages 463-470. IEEE, 2011. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[15] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[16] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE international conference on computer vision, pages 4038-4047, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.388 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[19] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.544 + ], + "angle": 0, + "content": "[20] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[21] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[22] Dongwoo Lee, Jeongtaek Oh, Jaesung Rim, Sunghyun Cho, and Kyoung Mu Lee. Exblurf: Efficient radiance fields for extreme motion blurred images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17639-17648, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.907, + 0.729 + ], + "angle": 0, + "content": "[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[24] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[25] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[26] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6020" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[28] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[29] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8102-8111, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[30] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2901–2908, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[31] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[32] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[33] Alex Rav-Acha and Shmuel Peleg. Two motion-blurred images are better than one. Pattern recognition letters, 26(3): 311-317, 2005. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.484, + 0.574 + ], + "angle": 0, + "content": "[34] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[35] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[37] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[38] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[39] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. In 2010" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 491-498, 2010. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[41] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.291 + ], + "angle": 0, + "content": "[43] Haichao Zhang and David Wipf. Non-uniform camera shake removal using a spatially-adaptive sparse penalty. Advances in Neural Information Processing Systems, 26, 2013. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[44] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2482-2491, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[45] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, Haozhe Xie, Jinshan Pan, and Jimmy S Ren. Davanet: Stereo deblurring with view aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10996-11005, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6021" + } + ] +] \ No newline at end of file diff --git a/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_origin.pdf b/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c7c7b5ee91af830163eace775c4beb25b7c65c11 --- /dev/null +++ b/2025/Exploiting Deblurring Networks for Radiance Fields/2b1465fe-6939-4acc-92a7-0a9426b9c956_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c89485921e4ec1e1d87caa2b54208591eed4b78476ed6027655a02ad5b6d2481 +size 2069070 diff --git a/2025/Exploiting Deblurring Networks for Radiance Fields/full.md b/2025/Exploiting Deblurring Networks for Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4489ec811cf5673395d229ab751e1ee874a2e3fb --- /dev/null +++ b/2025/Exploiting Deblurring Networks for Radiance Fields/full.md @@ -0,0 +1,287 @@ +# Exploiting Deblurring Networks for Radiance Fields + +Haeyun Choi $^{1,\dagger}$ + +Heemin Yang + +Janghyeok Han² + +Sunghyun Cho² + +${}^{1}\mathrm{KT}$ + +$^{2}$ POSTECH + +# Abstract + +In this paper, we propose DeepDeblurRF, a novel radiance field deblurring approach that can synthesize high-quality novel views from blurred training views with significantly reduced training time. DeepDeblurRF leverages deep neural network (DNN)-based deblurring modules to enjoy their deblurring performance and computational efficiency. To effectively combine DNN-based deblurring and radiance field construction, we propose a novel radiance field (RF)-guided deblurring and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Moreover, DeepDeblurRF is compatible with various scene representations, such as voxel grids and 3D Gaussians, expanding its applicability. We also present BlurRF-Synth, the first large-scale synthetic dataset for training radiance field deblurring frameworks. We conduct extensive experiments on both camera motion blur and defocus blur, demonstrating that DeepDeblurRF achieves state-of-the-art novel-view synthesis quality with significantly reduced training time. + +# 1. Introduction + +Novel-view synthesis has seen significant advancements in recent years, leading to impressive improvements in photorealistic rendering quality. One key development in this field is the Neural Radiance Field (NeRF) [25], which leverages neural networks to generate highly detailed images from novel viewpoints. Building on this, various works have focused on reducing training time and enhancing rendering quality [3, 12, 15, 27]. More recently, 3D Gaussian Splatting (3DGS) [17] has emerged, utilizing Gaussians as an explicit 3D representation. This technique enables high-quality scene reconstruction and real-time rendering through a differentiable rasterization method. + +However, synthesizing sharp novel views from degraded training views remains a significant challenge. Capturing images in real-world conditions often leads to various degra + +ditions such as blur and noise. Among them, blur makes it difficult to aggregate accurate 3D information from the training views, resulting in a blurry radiance field. To address blur, several works [20-22, 24, 31, 32, 39] have been proposed to reconstruct sharp radiance fields from training views with either camera motion blur or defocus blur. These radiance field deblurring methods jointly optimize blur kernels and the radiance field, and have shown promising results. + +However, these methods still possess several limitations that hinder their effectiveness. First, existing approaches rely on linear blur models, describing blurred pixels in training views as linear combinations of sharp pixels. In contrast, real-world blurred images often suffer from non-linear outliers such as saturated pixels due to sensor clipping, noise from various sources, and non-linear in-camera processing [8, 35]. These violate the linear blur model, leading to deblurring artifacts and severely degrading the performance of current radiance field deblurring techniques. Second, existing approaches do not utilize priors on sharp images, relying solely on complementary information from different views to model blur kernels and estimate a sharp radiance field. Consequently, they tend to produce radiance fields with residual blur and may completely fail when all input views exhibit similar blur directions [24], akin to classical multi-frame deblurring approaches [5, 7, 33], due to the lack of sharp image priors. Third, existing ray-based methods [21, 22, 24, 31, 39] employ multiple ray samples per pixel to depict a blurry image, resulting in a substantial increase in rendering time for a single view from a radiance field, and consequently, a significant computation time for radiance field construction. + +Apart from the radiance field deblurring methods, image deblurring has been extensively studied for decades [6, 14, 30, 40, 43]. Recently, a number of deep neural network (DNN)-based single-image deblurring methods have been proposed [4, 10, 19, 28, 38, 42]. These methods do not rely on linear blur models, but learn image deblurring from large-scale datasets. Thanks to this, they can effectively remove blur from a single image without the need for complementary information from other images, and also handle nonlinearities such as saturated pixels [34]. Moreover, their feedforward approach allows markedly reduced processing times, + +![](images/a825a29597ef2e0bd15045d45c3ae4227429d68f93ee73ba2d4621ad91b7c305.jpg) + +![](images/8e1184483bea44f28c0011d0159c45c94a87202883b1c033064a9953d7b7c464.jpg) + +![](images/08e266c21c2f529abb7cedc616145f69139cd6419e720b72d79b0f1f9cdb6498.jpg) +PDRF-10 (4 hrs) + +![](images/85118b729b553c03f49b821a13cf8a2cb5f822a2cefc749f3a6243282e60fc97.jpg) +Blurry Input View + +![](images/b5b69971ac04608946190baa0c2a5f3055062a8c2164b24ce446b9766c861e3b.jpg) +BAGS (75 mins) + +![](images/f815a23f1db8d555d761d5df8f860e310c6c91ee6b81c856439da47df3a30f08.jpg) +Deblur-NeRF (31 hrs) + +![](images/931d0f2dda88774247c89e18d5121fe3f74ddc3ded13d6aa058e5c674b8f6b47.jpg) +DeepDeblurRF-G (17 mins) +Figure 1. Given a set of multi-view blurry images even with non-linear outliers such as saturated pixels and noise, DeepDeblurRF performs high-quality novel-view synthesis with highly efficient training. DeepDeblurRF-P and DeepDeblurRF-G are the results of our framework, where radiance fields are constructed using Plenoxels [12] and 3D Gaussian Splatting [17], respectively. Note that DP-NeRF [21] was trained with two GPUs due to its memory demands, whereas other models were trained on a single NVIDIA TITAN RTX GPU. + +![](images/2c9378b73223feba652c9da77710696079b5361a87d55422ce647b5b13c418dd.jpg) + +![](images/a4784dad801f62c0407f3326134d5c099c54a297e80806f086d917b017fe967e.jpg) +Reference + +![](images/5105211b7eaee71448d070b74a4160e4022c3a0c8ed6c2219403c94a626d5a1c.jpg) + +unlike classical iterative optimization-based approaches. + +Given the effectiveness of recent DNN-based deblurring approaches, a promising direction to construct a sharp radiance field from blurry images would be to combine a deblurring network and radiance field construction. However, a naive combination of a deblurring network and radiance field construction, which performs single-image deblurring to each input blurry image and trains a radiance field using the deblurred images, results in unsatisfactory results as reported by Ma et al. [24]. This is because of the limited performance of single-image deblurring, which stems from the insufficient information available in a single image. + +In this paper, we propose DeepDeblurRF, a novel radiance field deblurring framework that enables highly efficient training and high-quality novel-view synthesis from images blurred by motion blur or defocus blur, even in the presence of noise and saturated pixels, as shown in Fig. 1. Unlike existing radiance field deblurring approaches, our method leverages DNN-based deblurring modules to achieve superior performance and computational efficiency. Additionally, our approach is versatile, capable of handling various 3D representations such as voxel grids [12] and 3D Gaussians [17]. + +However, solely relying on deblurring networks to remove blur from input blurry images may yield unsatisfactory results as discussed above. To effectively combine DNN-based deblurring and radiance field construction, our framework adopts a novel radiance field (RF)-guided deblurring scheme, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Specifically, our framework first deblurs input images, and then trains a radiance field using the deblurred images. Then, at the next iteration, it renders images corresponding to the input views from the trained radiance field. While the de + +blurred images may have limited qualities with inaccurately restored details, the rendered images provide higher-quality details as they are rendered from aggregated information from multiple deblurred images. Using the rendered images as guidance, our framework then performs RF-guided deblurring to the input blurred images. Thanks to the guidance, we can obtain higher-quality deblurred images, which we use for radiance field construction again. Finally, our framework iterates RF-guided deblurring and radiance field construction to gradually enhance the quality of the radiance field. + +Training the deblurring modules of our framework requires a large-scale dataset of blurred images paired with ground-truth sharp images for radiance field reconstruction. However, no such datasets are available as existing radiance field deblurring approaches adopt linear blur model-based approaches, which do not require large-scale training datasets [20-22, 24, 31, 32, 39]. Consequently, the datasets provided by previous works, such as those from DeblurNeRF [24] and ExBluRF [22], are small and primarily designed for evaluation rather than training. Furthermore, these datasets often overlook crucial factors like camera noise, in-camera processing, and other non-linear outliers, which are essential for realistic deblurring tasks. + +To address these gaps, we propose BlurRF-Synth, a large-scale dataset for radiance field deblurring approaches. It includes 4,350 blurred-sharp image pairs across 150 scenes, with 2,175 pairs for each of 75 scenes, encompassing both camera motion blur and defocus blur. These images are carefully synthesized to reflect real-world camera degradations such as noise, saturated pixels, and in-camera processing pipelines. Additionally, we introduce a real-world dataset called BlurRF-Real for evaluation in non-ideal conditions. Unlike the Deblur-NeRF dataset, which has less noise and + +adequate lighting, BlurRF-Real consists of five low-light indoor scenes captured with a machine vision camera. + +We validate DeepDeblurRF on both synthetic and real-world datasets with two types of blur: camera motion blur and defocus blur. Experimental results demonstrate that our method achieves state-of-the-art novel-view synthesis performance with highly reduced processing times. Furthermore, we demonstrate the extensibility of our framework by constructing radiance fields using different scene representations, such as voxel grids and Gaussian Splatting. Our contributions can be summarized as follows: + +- We propose DeepDeblurRF, a novel radiance field deblurring framework, which is the first approach that leverages DNN-based deblurring modules to overcome the limitations of the linear blur model. +- To this end, we present RF-guided deblurring, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner to gradually enhance the quality of the radiance field. +- We also present the BlurRF-Synth dataset, the first large-scale dataset for training and evaluation of novel-view synthesis from blurry images. + +# 2. Related Work + +Radiance field deblurring Several works have recently been proposed to tackle radiance field deblurring [21, 22, 24, 31, 39]. Ma et al. presented Deblur-NeRF, the first framework to train a sharp radiance field from blurry training views [24]. Deblur-NeRF adopts an additional MLP to model blur kernels and jointly optimizes them with a radiance field. DP-NeRF [21] models blur kernels based on physical priors for rigid motions, while PDRF [31] uses a two-stage kernel modeling scheme where the first stage coarsely aggregates kernel points to reduce computational cost. All these methods support both camera motion blur and defocus blur by modeling them using multiple rays. On the other hand, BAD-NeRF [39] and ExBluRF [22] propose radiance field deblurring for camera motion blur, modeling camera motion blur using camera trajectories. Nevertheless, all these methods rely on linear blur models, and require multiple ray-tracing steps, resulting in limited deblurring quality and slow training speeds. + +More recently, with the emergence of 3D Gaussian Splating [17], 3D Gaussian-based radiance field deblurring methods have been proposed. Deblurring-3DGS [20] handles camera motion blur with small offsets in 3D Gaussian positions and defocus blur by adjusting the geometry of Gaussians. BAGS [32] models both types of blur using dense per-pixel blur kernels optimized in a coarse-to-fine manner. However, BAGS requires substantial computation time to optimize dense blur kernels. Moreover, these methods also rely on linear blur models and are limited in handling real-world blurred images. + +Image deblurring Image deblurring has been widely studied to address blur that degrades image quality. Traditional methods, relying on linear blur models, often struggle with real-world blurred images containing non-linear outliers [6, 7, 9, 14, 30, 33, 40, 43]. Their iterative optimization processes are also computationally intensive, limiting their practical use. Recent advancements in deep learning have introduced DNN-based approaches, including single-image [4, 10, 19, 28, 38, 42] and multi-frame/video deblurring [2, 11, 13, 16, 29, 37, 41, 44, 45]. These methods can handle more complex blurs and reduce processing time with a feed-forward structure. + +While applying DNN-based deblurring approaches might be a potential solution for constructing radiance fields from blurry images, naively using single-image deblurring techniques yields unsatisfactory results as discussed in Sec. 1. Another possible direction is to adopt multi-frame and video deblurring approaches that leverage complementary information from multiple images for higher-quality results. However, these methods assume that input images are captured from nearly the same viewpoints, a condition that does not hold for training views in radiance field construction. + +# 3. DeepDeblurRF + +Fig. 2-(a) illustrates the overall framework of DeepDeblurRF. It takes a set of $M$ blurred images $\mathbf{B} = \{B_1, \dots, B_M\}$ of a scene and estimates a sharp radiance field that can synthesize a sharp novel view given an arbitrary pose. We assume that each blurred image $B_m$ is obtained from its latent sharp image $L_m$ through an unknown degradation process including blur and non-linear outliers such as saturated pixels and nonlinear in-camera processing. Based on this assumption, DeepDeblurRF first performs initial deblurring to the input blurred images, and obtains initial deblurred images. Then, our method iteratively performs radiance field construction using deblurred images and RF-guided deblurring to gradually enhance the quality of the radiance field and the deblurred images. At the last iteration, we perform only the radiance field construction step and obtain a final radiance field, from which we can synthesize sharp novel views. In the following, each step of DeepDeblurRF is described in more detail. + +# 3.1. Initial Deblurring + +The initial deblurring step removes blur from each blurry training view $B_{m} \in \mathbf{B}$ so that the following radiance field construction step can estimate the pose of each view more accurately, and more effectively aggregate information from different views. To this end, we adopt an off-the-shelf single-image deblurring network. Specifically, we adopt NAFNet [4], a state-of-the-art deblurring network, for its computational efficiency and performance. We denote the deblurred image from the initial deblurring step correspond + +![](images/55a8984932fcb43a456eba91f49ff6402ed0cdbd887d1b49742c8883647c855d.jpg) + +![](images/c12843d879350de223aaf310b195504d14519ef8617e4b6d98aa207be10e89e8.jpg) +(b) Blurry training view + +![](images/9f05916773816b1a04174bd5f2044e7e83f6f7dfe1aadee05b54d15151fa94a0.jpg) +(d) $1^{\mathrm{st}}$ iteration rendering + +![](images/45ff9b2c1b4ffbc3f616f370fc1673e8f69232b3834eea0d0a70a98420ba1714.jpg) +(c) Initial deblurring + +![](images/44a242c3f6921bb1d55e7b5f7ebb56dbe8eb0c15fe2ed17922a0fa4fa62a76db.jpg) +(e) $1^{\mathrm{st}}$ RF-guided deblurring + +![](images/504f36ad21fa11bec55ef5980635dd1c99c4032b86a58040cef44c631928afc9.jpg) +(a) $1^{\mathrm{st}}$ iteration rendering + +![](images/5cdaface6f9b95374b79a2ca75a695e4791d0d4d7a21f30f27d7ed417b0d8d58.jpg) +Figure 2. Overall framework and intermediate result of each step of DeepDeblurRF. +(b) $2^{\mathrm{nd}}$ iteration rendering + +![](images/800ecb96940b15ebc0924d2cf61839dd9b1e78d39b3921c2e8121d0df988db7e.jpg) +(c) $3^{\mathrm{rd}}$ iteration rendering +Figure 3. As iterations progress, the rendered images contain increasingly high-quality scene information, which subsequently improves the RF-guided deblurring network's performance in the next iteration. + +![](images/32a8a4ba4397870008a4bba02cde3e91e2098621bce9d94fc7ad87e5ac495807.jpg) +(d) $4^{\mathrm{th}}$ iteration rendering + +ing to $B_{m}$ as $D_{m}^{0}$ , and the set of the deblurred images as $\mathbf{D}^0 = \{D_1^0,\dots,D_M^0\}$ . + +# 3.2. Radiance Field Construction + +In the radiance field construction step at the $i$ -th iteration, we first estimate the camera poses of the deblurred images from the previous step, denoted as $\mathbf{D}^{i - 1} = \{D_1^{i - 1},\dots,D_M^{i - 1}\}$ using COLMAP [36], where $i\in [1,N]$ is the index of our iterative process. Using the estimated poses and deblurred images, we aggregate information about the target 3D scene and construct a radiance field $\mathbf{V}^i$ . As discussed in Sec. 1, our framework can adopt diverse radiance field representations. To demonstrate the extensibility of our framework, we adopt two different representations, Plenoxels [12] and 3D Gaussians [17], in our experiments. We refer to our framework using Plenoxels and 3D Gaussians as DeepDeblurRF-P and DeepDeblurRF-G, respectively. + +Note that deblurred images $\mathbf{D}^{i-1}$ from the previous step may still contain residual blur and deblurring artifacts, which can introduce corrupted information into the aggregation process. Nevertheless, since the input images have overlapping regions, other deblurred images can provide information for those regions that appear corrupted in some images. By aggregating such information from different deblurred images on the same regions, the radiance field construction step can effectively suppress corrupted information and obtain a higher-quality radiance field, from which we can render higher-quality images compared to the input deblurred im + +ages at this step. + +# 3.3. RF-guided Deblurring + +The RF-guided deblurring step deblurs the input views $\mathbf{B}$ using the aggregated information in $\mathbf{V}^i$ . Specifically, for each input view $B_{m}$ , we render $\mathbf{V}^i$ to obtain a rendered image $R_{m}^{i}$ . The rendered image $R_{m}^{i}$ shares the same content as its corresponding deblurred image $D_{m}^{i - 1}$ , but contains fewer artifacts and finer details. Motivated by this, our RF-guided deblurring leverages $R_{m}^{i}$ to obtain an updated deblurred image $D_{m}^{i}$ from $B_{m}$ . + +To perform RF-guided deblurring, we employ a novel RF-guided deblurring network that takes both the rendered image $R_{m}^{i}$ and the input blurred image $B_{m}$ to guide the deblurring process with the aggregated information in $R_{m}^{i}$ . For the deblurring network, we adopt NAFNet [4] and modify its first layer to take a concatenation of the two images. Once we obtain the updated deblurred images $\mathbf{D}^{i}$ , we return to the radiance field construction step and carry out the process for the $(i + 1)$ -th iteration, gradually enhancing the quality of the radiance field and the deblurred images. + +The intermediate results of each step are visualized in Fig. 2-(b) to Fig. 2-(e). Due to the severe blur in the input blurred image, its initial deblurring result retains some residual blur. In contrast, thanks to the information aggregation occurring in the radiance field construction step, the rendered image contains more accurate details. Finally, thanks to the guidance of the rendered image, the RF-guided deblurring + +![](images/d269df9f5fcd7783dff492ad8c61ea7b71bc6ff0273d31af1d12141f1fe5fa9c.jpg) +Figure 4. Examples of the BlurRF-Synth and BlurRF-Real datasets. The examples show blurred views in BlurRF-Real, while the top and bottom rows in BlurRF-Synth include blurred views and their corresponding sharp views. + +![](images/7794864bf3b5f9ed8280b9cfc307af75cd6e75a968ab0f2f60441b0ad3514e46.jpg) + +result shows sharper and more precise details that cannot be observed in the intermediate results of previous steps. + +Fig. 3 visualizes rendered images at different iterations. Although the rendered image at the first iteration Fig. 3-(a) contains blurry details, they are gradually refined at each iteration (Fig. 3 (b)-(d)) due to our iterative process, which incrementally improves the quality of the radiance fields and the deblurred images. The effectiveness of our iterative process is also demonstrated by the increasing PSNR score as iterations progress. + +# 4. BlurRF Dataset + +# 4.1. BlurRF-Synth + +We propose the BlurRF-Synth dataset, the first large-scale dataset for training and evaluating novel-view synthesis from blurred images. The dataset comprises a train set and a test set for each type of blur: camera motion blur and defocus blur. The train set includes 65 scenes, each containing 29 pairs of synthetically blurred images and their corresponding sharp images, captured from different viewpoints. The test set includes 10 scenes, each with 29 blurred-sharp image pairs and five additional sharp images from different viewpoints for evaluating novel-view synthesis quality. Examples of blurred images and ground-truth sharp images are shown in Fig. 4. We synthetically generated the dataset using Blender models, collecting 95 models from Blendswap1 under Creative Commons licenses and 5 models from the synthetic dataset of Deblur-NeRF [24]. For each model, we sampled 29 camera poses to capture different viewpoints. + +Camera motion blur We simulated camera shakes by randomly sampling a camera trajectory for each pose during the exposure duration. To this end, we adopted Bezier interpolation to model 6-DOF camera motion and densely sampled 51 intermediate poses along this camera trajectory. We then rendered the scene using Blender at each intermediate camera pose to produce a series of sharp images, which we then + +averaged to create a blurred image. All rendering and averaging were performed in the linear sRGB color space to accurately replicate the real-world image formation process. Among the sharp images, we sampled the temporally central image (i.e., the 26th image) as the ground-truth sharp image. + +Defocus blur We also generated images with defocus blur using Blender. We adjusted the camera's depth-of-field (DoF) in Blender to create defocus blur. We controlled the aperture size and randomly set the blade count between 7 and 9 to simulate realistic defocus effects. For each camera pose, we sampled a random focal distance within a predefined range based on the scene scale, shifting the focal plane to introduce variation in the defocus blur. + +Realistic blur synthesis To reflect real-world image degradation, we adopted the blur synthesis pipeline of RSBur [35]. Specifically, we generated camera motion blur by averaging sharp images in the linear sRGB space. Then, we synthesized saturated pixels to the blurred images and converted them to the camera RAW space. In the camera RAW space, we synthesized shot and read noise, and converted the images back to the camera sRGB space. For defocus blur, we rendered blurred images in the linear sRGB space and added noise in the same manner. For a more detailed pipeline, we refer the readers to [35]. The blur synthesis pipeline requires camera-specific noise parameters and a color correction matrix. We used a Sony A7R3 camera to estimate them. For more details about the BlurRF-Synth dataset, we refer the readers to the supplementary material. + +# 4.2.BlurRF-Real + +For evaluation under more realistic and challenging conditions, we propose BlurRF-Real, a real-world low-light camera motion blur dataset. The existing real datasets [22, 24] are captured in well-lit environments with minimal noise, making them less representative of camera motion blur in low-light conditions. Our dataset addresses this gap by providing data captured under challenging low-light scenarios. We collected five indoor scenes using a machine vision camera, + +
3D RepresentationModelCamera Motion BlurDefocus BlurComputation Time (Hr.)
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MLPDeblur-NeRF [24]27.670.83400.145030.030.87270.113731.33
BAD-NeRF [39]21.740.52980.3969---24.68
DP-NeRF [21]28.030.84120.126730.150.87630.099130.00*
Voxel gridExBluRF [22]26.560.78230.1955---8.92
PDRF-10 [31]28.330.84350.149530.030.87500.12254.26
DeepDeblurRF-P29.810.86680.114232.510.90580.09611.14
3D GaussiansDeblurring-3DGS [20]26.300.77290.172829.370.85450.14700.33
BAGS [32]27.410.81080.138229.900.86380.11521.25
DeepDeblurRF-G29.940.86810.105932.580.90600.07740.28
+ +Table 1. Quantitative results of novel-view synthesis on BlurRF-Synth test scenes. We highlight the best metrics and the second best metrics. Note that the computation time of DP-NeRF [21] was measured with two GPUs due to its memory demands, whereas those of the other models were measured on a single GPU. + +![](images/8257f61519d0fa50b13e87bcc9982ff97e69d54c59ee1c6ca76000630c7976b2.jpg) +Figure 5. Qualitative results of novel-view synthesis on BlurRF-Synth test scenes. + +each providing 20-40 multi-view blurry images, including 3-5 images for novel-view synthesis evaluation. Examples of blurred training views are shown in Fig. 4. + +# 5. Experiments + +Datasets The deblurring networks of DeepDeblurRF are trained with the train set of BlurRF-Synth. To evaluate DeepDeblurRF and other radiance field deblurring methods, we use the test set of BlurRF-Synth, the BlurRF-Real dataset, and the benchmark datasets from Deblur-NeRF [24]. The benchmark datasets include both synthetic and real-world scenes. The synthetic dataset consists of five scenes, while the real-world dataset contains ten scenes for each blur type. The real scenes of Deblur-NeRF were captured with a Canon EOS RP in manual exposure mode. The camera poses are estimated using COLMAP [36]. + +Implementation details We set the number of iterations $N$ to 5. Thus, as detailed in Sec. 3, DeepDeblurRF includes one single-image deblurring network and four RF-guided deblurring networks for each blur type: camera motion blur and defocus blur. Note that our approach requires an additional training phase for the deblurring networks, unlike + +previous radiance field deblurring methods. Once trained, our framework can construct radiance fields for arbitrary scenes and render novel views, similar to conventional approaches. To train the deblurring networks, we use the Adam optimizer [18] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.9$ . The learning rate is initially set to $10^{-3}$ and gradually reduced to $10^{-7}$ using the cosine annealing scheduler [23]. + +For constructing radiance fields in DeepDeblurRF-P, we follow the forward-facing experimental setup in Plenoxels, applying pruning and upsampling to the voxel grids every 38,400 iterations. For DeepDeblurRF-G, we adopt the original Gaussian Splatting method, incorporating the depth-based pruning and additional points strategy from DeblurGS [20] to address the sparse point clouds obtained from COLMAP on blurred images. + +We measured the computation time for radiance field deblurring using the camera motion blur test set of BlurRF-Synth on a single NVIDIA Titan RTX, except for DP-NeRF, which requires two GPUs due to its memory demands. Additional implementation details, including network configurations and the training settings for each iteration of DeepDeblurRF-P and DeepDeblurRF-G, are provided in the supplementary material. + +![](images/7d73d4bdf292e3f39a5ceefcb869857eec8d9830f01e28cbbf5ff234fb5693e4.jpg) +Figure 6. Qualitative results of novel-view synthesis on real-world datasets. + +![](images/3e248613e7f5667956334566827f3bea8df7fe67fc175849a19ba2c04d81abc6.jpg) +Figure 7. Qualitative results of novel-view synthesis on BlurRF-SB dataset. + +Comparison using BlurRF-Synth We compare DeepDeblurRF with other radiance field deblurring methods [20-22, 24, 31, 32, 39]. Tab. 1 shows the quantitative results of DeepDeblurRF and the other methods on the test set of BlurRF-Synth. The table demonstrates that our approach outperforms the existing methods in terms of quality, regardless of the radiance field representation. Moreover, it achieves substantially shorter computation times. Specifically, DeepDeblurRF-P shows the shortest computation time among the approaches based on MLPs and voxel grids, while DeepDeblurRF-G is the fastest among all the methods. This superior performance can be attributed to its DNN-based deblurring modules, which handle blur efficiently and effectively, in contrast to the other methods constrained by linear blur models. + +The superior performance of our approach is further demonstrated in Fig. 5, which shows that our method + +achieves sharper results on the BlurRF-Synth dataset. Specifically, on the camera motion blur scene, all the other methods produce noisy results, whereas both DeepDeblurRF-P and DeepDeblurRF-G effectively remove the blur even in challenging conditions with noise and pixel saturation, achieving high-quality novel-view synthesis. Similarly, in the defocus blur scene, our methods render sharper novel views by minimizing residual blur than the other methods. + +Comparison on real datasets We compare the novel-view synthesis performance of DeepDeblurRF against other approaches using both BlurRF-Real and the real scenes from Deblur-NeRF [24]. As detailed in Sec. 4.2, BlurRF-Real provides an evaluation of how well each method handles camera motion blur with non-linear outliers and noise in real-world low-light conditions. As shown in Tab. 2 and Fig. 6, our method not only outperforms the others in non-reference metrics [1, 26] but also achieves high-quality qualitative + +
ModelBlurRF-RealBlurRF-SB
NIQE (↓)ARNIQA (↑)PSNR (↑)SSIM (↑)LPIPS (↓)
MLP
Deblur-NeRF [24]6.3410.27925.430.72640.2319
BAD-NeRF [39]9.9080.23721.970.52310.4731
DP-NeRF [21]6.4980.26627.070.79400.1795
Voxel grid
ExBluRF [22]7.4790.27924.160.66490.2659
PDRF-10 [31]6.2430.26626.660.77300.2124
DeepDeblurRF-P5.8290.32329.450.84720.1677
3D Gaussians
Deblurring-3DGS [20]6.0860.26223.870.64590.2475
BAGS [32]5.8370.29024.490.67480.2391
DeepDeblurRF-G5.4230.32929.590.85480.1399
+ +results, clearly surpassing the competing methods. It also shows superior performance on real scenes from Deblur-NeRF. Specifically, while other methods struggle with saturated pixels in the camera motion blur scene, our method effectively handles it. Likewise, in the defocus blur scene, it renders sharper novel views without residual blur. + +While the real-world scenes of Deblur-NeRF provide blur-free reference images, they suffer from misalignment and exposure differences [24, 31, 32]. Similarly, BlurRF-Real also includes blur-free reference images, but they have not only similar limitations but also severe noise as they are captured under severe low-light conditions. Thus, we report quantitative evaluations against the reference images in the supplementary material. More experimental results are also provided in the supplementary material. + +Input views with same blur directions As discussed in Sec. 1, unlike previous approaches, our method does not rely on the assumption that input blurred images contain complementary information from different blur directions. To verify this, we generated an additional synthetic test set, named BlurRF-SB, where the blurred images of the same scene share the same blur directions, while their blur magnitudes are different. BlurRF-SB was generated using five scenes from the BlurRF-Synth test set, following the same generation process. + +Fig. 7 shows a qualitative comparison on BlurRF-SB. Due to the lack of complementary information in different training views, the other methods fail to restore sharp details. Conversely, our methods successfully restore sharp details. The quantitative comparison in Tab. 2 also demonstrates that our approach significantly outperforms the others for blurred images with the same blur directions. These results verify the benefit of using prior knowledge on sharp images from pre-trained deblurring networks. + +Number of iterations We investigate the impact of the hyperparameter $N$ , which sets the number of iterations in our framework. Tab. 3 shows that DeepDeblurRF-G achieves improved novel-view synthesis as $N$ increases, highlighting + +Table 2. Quantitative results of novel-view synthesis on BlurRF-Real and BlurRF-SB datasets. We highlight the best metrics and the second best metrics. + +
#Iter.Camera motionDefocus
PSNRSSIMLPIPSPSNRSSIMLPIPS
N=128.400.81750.170730.740.88620.1297
N=229.230.84620.128131.220.89050.0955
N=329.650.85880.115232.120.90030.0899
N=429.830.86320.110832.300.90380.0860
N=529.940.86810.105932.580.90600.0774
N=629.970.86920.103432.750.90870.0736
+ +Table 3. Quantitative results of ablation study on the number of iterations $N$ using DeepDeblurRF-G on the test sets of BlurRF-Synth. + +![](images/4d3e2eca8f11b9c8612d42165ee06a9c75fb908632ecc922d8f3ba4d49a429e1.jpg) +Figure 8. Qualitative results of ablation study on the number of iterations $N$ using DeepDeblurRF-G on the real-world scenes [24]. + +the effectiveness of our iterative approach. However, gains diminish beyond $N = 5$ , suggesting that $N = 5$ is sufficient. Fig. 8 further supports this with qualitative results on real scenes [24] (camera motion blur and defocus blur), showing similar trends. Similar results were observed for DeepDeblurRF-P, as detailed in the supplementary material. + +# 6. Conclusion + +In this paper, we proposed DeepDeblurRF, a novel radiance field deblurring approach that leverages DNN-based deblurring to overcome the limitations of the linear blur model of previous methods. To integrate DNN-based deblurring with radiance field construction, we presented a novel RF-guided deblurring scheme and the framework that alternates between RF-guided deblurring and radiance field training. We also presented BlurRF-Synth, the first large-scale dataset for training and evaluating radiance field deblurring frameworks. Experimental results showed that our method outperforms others, handling both camera motion blur and defocus blur effectively with highly reduced computation time. + +Limitations and future work Our method leverages deblurring networks to remove blur from training views, but restoring images with more severe blur than in the training data remains challenging. Our RF-guided deblurring uses only rendered images from each input view for guidance. Imposing additional priors on other views and incorporating additional information, e.g., depth maps rendered from radiance fields, may further improve the deblurring performance. + +Acknowledgements This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government(MSIT) (RS-2024-00457882, AI Research Hub Project, RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH), RS-2024-00437866, ITRC(Information Technology Research Center)). + +# References + +[1] Lorenzo Agnolucci et al. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 7 +[2] Miika Aittala and Fredo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 3 +[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1 +[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 1, 3, 4 +[5] Wei-Ge Chen, N Nandhakumar, and Worthy N Martin. Image motion estimation from motion smear-a new computational model. IEEE transactions on pattern analysis and machine intelligence, 18(4):412-425, 1996. 1 +[6] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Trans. Graph., 28(5):145:1-145:8, 2009. 1, 3 +[7] Sunghyun Cho, Yasuyuki Matsushita, and Seungyong Lee. Removing non-uniform motion blur from images. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 1, 3 +[8] Sunghyun Cho, Jue Wang, and Seungyong Lee. Handling outliers in non-blind image deconvolution. In 2011 international conference on computer vision, pages 495-502. IEEE, 2011. 1 +[9] Sunghyun Cho, Jue Wang, and Seungyong Lee. Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3 +[10] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 3 +[11] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 3 +[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2, 4 + +[13] Chunzhi Gu, Xuequan Lu, Ying He, and Chao Zhang. Blur removal via blurred-noisy image pair. IEEE Transactions on Image Processing, 30:345-359, 2020. 3 +[14] Michael Hirsch, Christian J Schuler, Stefan Harmeling, and Bernhard Schölkopf. Fast removal of non-uniform camera shake. In 2011 International Conference on Computer Vision, pages 463-470. IEEE, 2011. 1, 3 +[15] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 1 +[16] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE international conference on computer vision, pages 4038-4047, 2017. 3 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 1, 2, 3, 4 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[19] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 1, 3 +[20] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 1, 2, 3, 6, 7, 8 +[21] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 1, 2, 3, 6, 8 +[22] Dongwoo Lee, Jeongtaek Oh, Jaesung Rim, Sunghyun Cho, and Kyoung Mu Lee. Exblurf: Efficient radiance fields for extreme motion blurred images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17639-17648, 2023. 1, 2, 3, 5, 6, 7, 8 +[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[24] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1, 2, 3, 5, 6, 7, 8 +[25] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1 +[26] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 7 + +[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1 +[28] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3 +[29] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8102-8111, 2019. 3 +[30] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2901–2908, 2014. 1, 3 +[31] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 1, 2, 3, 6, 7, 8 +[32] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling, 2024. 1, 2, 3, 6, 7, 8 +[33] Alex Rav-Acha and Shmuel Peleg. Two motion-blurred images are better than one. Pattern recognition letters, 26(3): 311-317, 2005. 1, 3 +[34] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1 +[35] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 5 +[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6 +[37] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3 +[38] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3 +[39] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 1, 2, 3, 6, 7, 8 +[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. In 2010 + +IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 491-498, 2010. 1, 3 +[41] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3 +[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 3 +[43] Haichao Zhang and David Wipf. Non-uniform camera shake removal using a spatially-adaptive sparse penalty. Advances in Neural Information Processing Systems, 26, 2013. 1, 3 +[44] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2482-2491, 2019. 3 +[45] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, Haozhe Xie, Jinshan Pan, and Jimmy S Ren. Davanet: Stereo deblurring with view aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10996-11005, 2019. 3 \ No newline at end of file diff --git a/2025/Exploiting Deblurring Networks for Radiance Fields/images.zip b/2025/Exploiting Deblurring Networks for Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..31c649130dad63b5e07107ae6b03acab0eb3de1e --- /dev/null +++ b/2025/Exploiting Deblurring Networks for Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c7bab62316422fe305ebfe2a90f49b24a95df8bf87b6fd57816c2f5870cda26 +size 765908 diff --git a/2025/Exploiting Deblurring Networks for Radiance Fields/layout.json b/2025/Exploiting Deblurring Networks for Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4274f9bb7f18a5e3472c03305bb6bdb2c7f060bf --- /dev/null +++ b/2025/Exploiting Deblurring Networks for Radiance Fields/layout.json @@ -0,0 +1,7748 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 470, + 121 + ], + "type": "text", + "content": "Exploiting Deblurring Networks for Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 184, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 142, + 184, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 184, + 158 + ], + "type": "text", + "content": "Haeyun Choi" + }, + { + "bbox": [ + 106, + 142, + 184, + 158 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 142, + 286, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 142, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 212, + 142, + 286, + 158 + ], + "type": "text", + "content": "Heemin Yang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 142, + 396, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 142, + 396, + 158 + ], + "spans": [ + { + "bbox": [ + 314, + 142, + 396, + 158 + ], + "type": "text", + "content": "Janghyeok Han²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 425, + 142, + 504, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 142, + 504, + 158 + ], + "spans": [ + { + "bbox": [ + 425, + 142, + 504, + 158 + ], + "type": "text", + "content": "Sunghyun Cho²" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 171, + 269, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 171, + 269, + 184 + ], + "spans": [ + { + "bbox": [ + 246, + 171, + 269, + 184 + ], + "type": "inline_equation", + "content": "{}^{1}\\mathrm{KT}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 301, + 171, + 362, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 171, + 362, + 184 + ], + "spans": [ + { + "bbox": [ + 301, + 171, + 362, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 301, + 171, + 362, + 184 + ], + "type": "text", + "content": "POSTECH" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 214, + 201, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 214, + 201, + 227 + ], + "spans": [ + { + "bbox": [ + 151, + 214, + 201, + 227 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 240, + 297, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 297, + 468 + ], + "type": "text", + "content": "In this paper, we propose DeepDeblurRF, a novel radiance field deblurring approach that can synthesize high-quality novel views from blurred training views with significantly reduced training time. DeepDeblurRF leverages deep neural network (DNN)-based deblurring modules to enjoy their deblurring performance and computational efficiency. To effectively combine DNN-based deblurring and radiance field construction, we propose a novel radiance field (RF)-guided deblurring and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Moreover, DeepDeblurRF is compatible with various scene representations, such as voxel grids and 3D Gaussians, expanding its applicability. We also present BlurRF-Synth, the first large-scale synthetic dataset for training radiance field deblurring frameworks. We conduct extensive experiments on both camera motion blur and defocus blur, demonstrating that DeepDeblurRF achieves state-of-the-art novel-view synthesis quality with significantly reduced training time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 492, + 135, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 135, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 135, + 504 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 512, + 296, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 655 + ], + "type": "text", + "content": "Novel-view synthesis has seen significant advancements in recent years, leading to impressive improvements in photorealistic rendering quality. One key development in this field is the Neural Radiance Field (NeRF) [25], which leverages neural networks to generate highly detailed images from novel viewpoints. Building on this, various works have focused on reducing training time and enhancing rendering quality [3, 12, 15, 27]. More recently, 3D Gaussian Splatting (3DGS) [17] has emerged, utilizing Gaussians as an explicit 3D representation. This technique enables high-quality scene reconstruction and real-time rendering through a differentiable rasterization method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "content": "However, synthesizing sharp novel views from degraded training views remains a significant challenge. Capturing images in real-world conditions often leads to various degra" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "content": "ditions such as blur and noise. Among them, blur makes it difficult to aggregate accurate 3D information from the training views, resulting in a blurry radiance field. To address blur, several works [20-22, 24, 31, 32, 39] have been proposed to reconstruct sharp radiance fields from training views with either camera motion blur or defocus blur. These radiance field deblurring methods jointly optimize blur kernels and the radiance field, and have shown promising results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 315, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 556, + 579 + ], + "type": "text", + "content": "However, these methods still possess several limitations that hinder their effectiveness. First, existing approaches rely on linear blur models, describing blurred pixels in training views as linear combinations of sharp pixels. In contrast, real-world blurred images often suffer from non-linear outliers such as saturated pixels due to sensor clipping, noise from various sources, and non-linear in-camera processing [8, 35]. These violate the linear blur model, leading to deblurring artifacts and severely degrading the performance of current radiance field deblurring techniques. Second, existing approaches do not utilize priors on sharp images, relying solely on complementary information from different views to model blur kernels and estimate a sharp radiance field. Consequently, they tend to produce radiance fields with residual blur and may completely fail when all input views exhibit similar blur directions [24], akin to classical multi-frame deblurring approaches [5, 7, 33], due to the lack of sharp image priors. Third, existing ray-based methods [21, 22, 24, 31, 39] employ multiple ray samples per pixel to depict a blurry image, resulting in a substantial increase in rendering time for a single view from a radiance field, and consequently, a significant computation time for radiance field construction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Apart from the radiance field deblurring methods, image deblurring has been extensively studied for decades [6, 14, 30, 40, 43]. Recently, a number of deep neural network (DNN)-based single-image deblurring methods have been proposed [4, 10, 19, 28, 38, 42]. These methods do not rely on linear blur models, but learn image deblurring from large-scale datasets. Thanks to this, they can effectively remove blur from a single image without the need for complementary information from other images, and also handle nonlinearities such as saturated pixels [34]. Moreover, their feedforward approach allows markedly reduced processing times," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 188, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 188, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 188, + 712 + ], + "type": "text", + "content": "†This work was done at POSTECH." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6012" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 145, + 150 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 145, + 150 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 145, + 150 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 145, + 150 + ], + "type": "image", + "image_path": "a825a29597ef2e0bd15045d45c3ae4227429d68f93ee73ba2d4621ad91b7c305.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 70, + 228, + 150 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 228, + 150 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 228, + 150 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 228, + 150 + ], + "type": "image", + "image_path": "8e1184483bea44f28c0011d0159c45c94a87202883b1c033064a9953d7b7c464.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 229, + 71, + 312, + 150 + ], + "blocks": [ + { + "bbox": [ + 229, + 71, + 312, + 150 + ], + "lines": [ + { + "bbox": [ + 229, + 71, + 312, + 150 + ], + "spans": [ + { + "bbox": [ + 229, + 71, + 312, + 150 + ], + "type": "image", + "image_path": "08e266c21c2f529abb7cedc616145f69139cd6419e720b72d79b0f1f9cdb6498.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 245, + 151, + 296, + 158 + ], + "lines": [ + { + "bbox": [ + 245, + 151, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 245, + 151, + 296, + 158 + ], + "type": "text", + "content": "PDRF-10 (4 hrs)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 159, + 145, + 216 + ], + "blocks": [ + { + "bbox": [ + 76, + 151, + 131, + 159 + ], + "lines": [ + { + "bbox": [ + 76, + 151, + 131, + 159 + ], + "spans": [ + { + "bbox": [ + 76, + 151, + 131, + 159 + ], + "type": "text", + "content": "Blurry Input View" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 159, + 145, + 216 + ], + "lines": [ + { + "bbox": [ + 62, + 159, + 145, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 145, + 216 + ], + "type": "image", + "image_path": "85118b729b553c03f49b821a13cf8a2cb5f822a2cefc749f3a6243282e60fc97.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 216, + 145, + 237 + ], + "blocks": [ + { + "bbox": [ + 62, + 216, + 145, + 237 + ], + "lines": [ + { + "bbox": [ + 62, + 216, + 145, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 145, + 237 + ], + "type": "image", + "image_path": "b5b69971ac04608946190baa0c2a5f3055062a8c2164b24ce446b9766c861e3b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 239, + 129, + 247 + ], + "lines": [ + { + "bbox": [ + 78, + 239, + 129, + 247 + ], + "spans": [ + { + "bbox": [ + 78, + 239, + 129, + 247 + ], + "type": "text", + "content": "BAGS (75 mins)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 145, + 159, + 228, + 216 + ], + "blocks": [ + { + "bbox": [ + 153, + 151, + 220, + 159 + ], + "lines": [ + { + "bbox": [ + 153, + 151, + 220, + 159 + ], + "spans": [ + { + "bbox": [ + 153, + 151, + 220, + 159 + ], + "type": "text", + "content": "Deblur-NeRF (31 hrs)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 159, + 228, + 216 + ], + "lines": [ + { + "bbox": [ + 145, + 159, + 228, + 216 + ], + "spans": [ + { + "bbox": [ + 145, + 159, + 228, + 216 + ], + "type": "image", + "image_path": "f815a23f1db8d555d761d5df8f860e310c6c91ee6b81c856439da47df3a30f08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 145, + 216, + 222, + 238 + ], + "blocks": [ + { + "bbox": [ + 145, + 216, + 222, + 238 + ], + "lines": [ + { + "bbox": [ + 145, + 216, + 222, + 238 + ], + "spans": [ + { + "bbox": [ + 145, + 216, + 222, + 238 + ], + "type": "image", + "image_path": "931d0f2dda88774247c89e18d5121fe3f74ddc3ded13d6aa058e5c674b8f6b47.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 239, + 229, + 247 + ], + "lines": [ + { + "bbox": [ + 145, + 239, + 229, + 247 + ], + "spans": [ + { + "bbox": [ + 145, + 239, + 229, + 247 + ], + "type": "text", + "content": "DeepDeblurRF-G (17 mins)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 258, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 302 + ], + "type": "text", + "content": "Figure 1. Given a set of multi-view blurry images even with non-linear outliers such as saturated pixels and noise, DeepDeblurRF performs high-quality novel-view synthesis with highly efficient training. DeepDeblurRF-P and DeepDeblurRF-G are the results of our framework, where radiance fields are constructed using Plenoxels [12] and 3D Gaussian Splatting [17], respectively. Note that DP-NeRF [21] was trained with two GPUs due to its memory demands, whereas other models were trained on a single NVIDIA TITAN RTX GPU." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 229, + 159, + 313, + 216 + ], + "blocks": [ + { + "bbox": [ + 229, + 159, + 313, + 216 + ], + "lines": [ + { + "bbox": [ + 229, + 159, + 313, + 216 + ], + "spans": [ + { + "bbox": [ + 229, + 159, + 313, + 216 + ], + "type": "image", + "image_path": "2c9378b73223feba652c9da77710696079b5361a87d55422ce647b5b13c418dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 229, + 216, + 288, + 238 + ], + "blocks": [ + { + "bbox": [ + 229, + 216, + 288, + 238 + ], + "lines": [ + { + "bbox": [ + 229, + 216, + 288, + 238 + ], + "spans": [ + { + "bbox": [ + 229, + 216, + 288, + 238 + ], + "type": "image", + "image_path": "a4784dad801f62c0407f3326134d5c099c54a297e80806f086d917b017fe967e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 239, + 286, + 247 + ], + "lines": [ + { + "bbox": [ + 254, + 239, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 254, + 239, + 286, + 247 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 74, + 549, + 236 + ], + "blocks": [ + { + "bbox": [ + 313, + 74, + 549, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 74, + 549, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 74, + 549, + 236 + ], + "type": "image", + "image_path": "5105211b7eaee71448d070b74a4160e4022c3a0c8ed6c2219403c94a626d5a1c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 315, + 284, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 284, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 284, + 327 + ], + "type": "text", + "content": "unlike classical iterative optimization-based approaches." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 328, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 296, + 460 + ], + "type": "text", + "content": "Given the effectiveness of recent DNN-based deblurring approaches, a promising direction to construct a sharp radiance field from blurry images would be to combine a deblurring network and radiance field construction. However, a naive combination of a deblurring network and radiance field construction, which performs single-image deblurring to each input blurry image and trains a radiance field using the deblurred images, results in unsatisfactory results as reported by Ma et al. [24]. This is because of the limited performance of single-image deblurring, which stems from the insufficient information available in a single image." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 460, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 296, + 581 + ], + "type": "text", + "content": "In this paper, we propose DeepDeblurRF, a novel radiance field deblurring framework that enables highly efficient training and high-quality novel-view synthesis from images blurred by motion blur or defocus blur, even in the presence of noise and saturated pixels, as shown in Fig. 1. Unlike existing radiance field deblurring approaches, our method leverages DNN-based deblurring modules to achieve superior performance and computational efficiency. Additionally, our approach is versatile, capable of handling various 3D representations such as voxel grids [12] and 3D Gaussians [17]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 713 + ], + "type": "text", + "content": "However, solely relying on deblurring networks to remove blur from input blurry images may yield unsatisfactory results as discussed above. To effectively combine DNN-based deblurring and radiance field construction, our framework adopts a novel radiance field (RF)-guided deblurring scheme, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Specifically, our framework first deblurs input images, and then trains a radiance field using the deblurred images. Then, at the next iteration, it renders images corresponding to the input views from the trained radiance field. While the de" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "text", + "content": "blurred images may have limited qualities with inaccurately restored details, the rendered images provide higher-quality details as they are rendered from aggregated information from multiple deblurred images. Using the rendered images as guidance, our framework then performs RF-guided deblurring to the input blurred images. Thanks to the guidance, we can obtain higher-quality deblurred images, which we use for radiance field construction again. Finally, our framework iterates RF-guided deblurring and radiance field construction to gradually enhance the quality of the radiance field." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 436, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 592 + ], + "type": "text", + "content": "Training the deblurring modules of our framework requires a large-scale dataset of blurred images paired with ground-truth sharp images for radiance field reconstruction. However, no such datasets are available as existing radiance field deblurring approaches adopt linear blur model-based approaches, which do not require large-scale training datasets [20-22, 24, 31, 32, 39]. Consequently, the datasets provided by previous works, such as those from DeblurNeRF [24] and ExBluRF [22], are small and primarily designed for evaluation rather than training. Furthermore, these datasets often overlook crucial factors like camera noise, in-camera processing, and other non-linear outliers, which are essential for realistic deblurring tasks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "type": "text", + "content": "To address these gaps, we propose BlurRF-Synth, a large-scale dataset for radiance field deblurring approaches. It includes 4,350 blurred-sharp image pairs across 150 scenes, with 2,175 pairs for each of 75 scenes, encompassing both camera motion blur and defocus blur. These images are carefully synthesized to reflect real-world camera degradations such as noise, saturated pixels, and in-camera processing pipelines. Additionally, we introduce a real-world dataset called BlurRF-Real for evaluation in non-ideal conditions. Unlike the Deblur-NeRF dataset, which has less noise and" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6013" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "adequate lighting, BlurRF-Real consists of five low-light indoor scenes captured with a machine vision camera." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 203 + ], + "type": "text", + "content": "We validate DeepDeblurRF on both synthetic and real-world datasets with two types of blur: camera motion blur and defocus blur. Experimental results demonstrate that our method achieves state-of-the-art novel-view synthesis performance with highly reduced processing times. Furthermore, we demonstrate the extensibility of our framework by constructing radiance fields using different scene representations, such as voxel grids and Gaussian Splatting. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 295, + 334 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 204, + 295, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 249 + ], + "type": "text", + "content": "- We propose DeepDeblurRF, a novel radiance field deblurring framework, which is the first approach that leverages DNN-based deblurring modules to overcome the limitations of the linear blur model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 294, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 297 + ], + "type": "text", + "content": "- To this end, we present RF-guided deblurring, and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner to gradually enhance the quality of the radiance field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 334 + ], + "type": "text", + "content": "- We also present the BlurRF-Synth dataset, the first large-scale dataset for training and evaluation of novel-view synthesis from blurry images." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 344, + 142, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 142, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 142, + 357 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 366, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 581 + ], + "type": "text", + "content": "Radiance field deblurring Several works have recently been proposed to tackle radiance field deblurring [21, 22, 24, 31, 39]. Ma et al. presented Deblur-NeRF, the first framework to train a sharp radiance field from blurry training views [24]. Deblur-NeRF adopts an additional MLP to model blur kernels and jointly optimizes them with a radiance field. DP-NeRF [21] models blur kernels based on physical priors for rigid motions, while PDRF [31] uses a two-stage kernel modeling scheme where the first stage coarsely aggregates kernel points to reduce computational cost. All these methods support both camera motion blur and defocus blur by modeling them using multiple rays. On the other hand, BAD-NeRF [39] and ExBluRF [22] propose radiance field deblurring for camera motion blur, modeling camera motion blur using camera trajectories. Nevertheless, all these methods rely on linear blur models, and require multiple ray-tracing steps, resulting in limited deblurring quality and slow training speeds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "More recently, with the emergence of 3D Gaussian Splating [17], 3D Gaussian-based radiance field deblurring methods have been proposed. Deblurring-3DGS [20] handles camera motion blur with small offsets in 3D Gaussian positions and defocus blur by adjusting the geometry of Gaussians. BAGS [32] models both types of blur using dense per-pixel blur kernels optimized in a coarse-to-fine manner. However, BAGS requires substantial computation time to optimize dense blur kernels. Moreover, these methods also rely on linear blur models and are limited in handling real-world blurred images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 215 + ], + "type": "text", + "content": "Image deblurring Image deblurring has been widely studied to address blur that degrades image quality. Traditional methods, relying on linear blur models, often struggle with real-world blurred images containing non-linear outliers [6, 7, 9, 14, 30, 33, 40, 43]. Their iterative optimization processes are also computationally intensive, limiting their practical use. Recent advancements in deep learning have introduced DNN-based approaches, including single-image [4, 10, 19, 28, 38, 42] and multi-frame/video deblurring [2, 11, 13, 16, 29, 37, 41, 44, 45]. These methods can handle more complex blurs and reduce processing time with a feed-forward structure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 556, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 556, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 556, + 335 + ], + "type": "text", + "content": "While applying DNN-based deblurring approaches might be a potential solution for constructing radiance fields from blurry images, naively using single-image deblurring techniques yields unsatisfactory results as discussed in Sec. 1. Another possible direction is to adopt multi-frame and video deblurring approaches that leverage complementary information from multiple images for higher-quality results. However, these methods assume that input images are captured from nearly the same viewpoints, a condition that does not hold for training views in radiance field construction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 346, + 408, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 408, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 408, + 359 + ], + "type": "text", + "content": "3. DeepDeblurRF" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "content": "Fig. 2-(a) illustrates the overall framework of DeepDeblurRF. It takes a set of " + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "content": " blurred images " + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = \\{B_1, \\dots, B_M\\}" + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "content": " of a scene and estimates a sharp radiance field that can synthesize a sharp novel view given an arbitrary pose. We assume that each blurred image " + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "inline_equation", + "content": "B_m" + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "content": " is obtained from its latent sharp image " + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "inline_equation", + "content": "L_m" + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "content": " through an unknown degradation process including blur and non-linear outliers such as saturated pixels and nonlinear in-camera processing. Based on this assumption, DeepDeblurRF first performs initial deblurring to the input blurred images, and obtains initial deblurred images. Then, our method iteratively performs radiance field construction using deblurred images and RF-guided deblurring to gradually enhance the quality of the radiance field and the deblurred images. At the last iteration, we perform only the radiance field construction step and obtain a final radiance field, from which we can synthesize sharp novel views. In the following, each step of DeepDeblurRF is described in more detail." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 588, + 421, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 421, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 421, + 601 + ], + "type": "text", + "content": "3.1. Initial Deblurring" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "The initial deblurring step removes blur from each blurry training view " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "B_{m} \\in \\mathbf{B}" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " so that the following radiance field construction step can estimate the pose of each view more accurately, and more effectively aggregate information from different views. To this end, we adopt an off-the-shelf single-image deblurring network. Specifically, we adopt NAFNet [4], a state-of-the-art deblurring network, for its computational efficiency and performance. We denote the deblurred image from the initial deblurring step correspond" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6014" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 74, + 364, + 210 + ], + "blocks": [ + { + "bbox": [ + 62, + 74, + 364, + 210 + ], + "lines": [ + { + "bbox": [ + 62, + 74, + 364, + 210 + ], + "spans": [ + { + "bbox": [ + 62, + 74, + 364, + 210 + ], + "type": "image", + "image_path": "55a8984932fcb43a456eba91f49ff6402ed0cdbd887d1b49742c8883647c855d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 371, + 74, + 459, + 134 + ], + "blocks": [ + { + "bbox": [ + 371, + 74, + 459, + 134 + ], + "lines": [ + { + "bbox": [ + 371, + 74, + 459, + 134 + ], + "spans": [ + { + "bbox": [ + 371, + 74, + 459, + 134 + ], + "type": "image", + "image_path": "c12843d879350de223aaf310b195504d14519ef8617e4b6d98aa207be10e89e8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 134, + 451, + 143 + ], + "lines": [ + { + "bbox": [ + 379, + 134, + 451, + 143 + ], + "spans": [ + { + "bbox": [ + 379, + 134, + 451, + 143 + ], + "type": "text", + "content": "(b) Blurry training view" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 371, + 144, + 460, + 201 + ], + "blocks": [ + { + "bbox": [ + 371, + 144, + 460, + 201 + ], + "lines": [ + { + "bbox": [ + 371, + 144, + 460, + 201 + ], + "spans": [ + { + "bbox": [ + 371, + 144, + 460, + 201 + ], + "type": "image", + "image_path": "9f05916773816b1a04174bd5f2044e7e83f6f7dfe1aadee05b54d15151fa94a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 201, + 453, + 211 + ], + "lines": [ + { + "bbox": [ + 378, + 201, + 453, + 211 + ], + "spans": [ + { + "bbox": [ + 378, + 201, + 453, + 211 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 378, + 201, + 453, + 211 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 378, + 201, + 453, + 211 + ], + "type": "text", + "content": " iteration rendering" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 460, + 75, + 548, + 134 + ], + "blocks": [ + { + "bbox": [ + 460, + 75, + 548, + 134 + ], + "lines": [ + { + "bbox": [ + 460, + 75, + 548, + 134 + ], + "spans": [ + { + "bbox": [ + 460, + 75, + 548, + 134 + ], + "type": "image", + "image_path": "45ff9b2c1b4ffbc3f616f370fc1673e8f69232b3834eea0d0a70a98420ba1714.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 134, + 534, + 143 + ], + "lines": [ + { + "bbox": [ + 473, + 134, + 534, + 143 + ], + "spans": [ + { + "bbox": [ + 473, + 134, + 534, + 143 + ], + "type": "text", + "content": "(c) Initial deblurring" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 460, + 144, + 548, + 201 + ], + "blocks": [ + { + "bbox": [ + 460, + 144, + 548, + 201 + ], + "lines": [ + { + "bbox": [ + 460, + 144, + 548, + 201 + ], + "spans": [ + { + "bbox": [ + 460, + 144, + 548, + 201 + ], + "type": "image", + "image_path": "44a242c3f6921bb1d55e7b5f7ebb56dbe8eb0c15fe2ed17922a0fa4fa62a76db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 201, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 462, + 201, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 462, + 201, + 545, + 211 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 462, + 201, + 545, + 211 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 462, + 201, + 545, + 211 + ], + "type": "text", + "content": " RF-guided deblurring" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 236, + 181, + 316 + ], + "blocks": [ + { + "bbox": [ + 62, + 236, + 181, + 316 + ], + "lines": [ + { + "bbox": [ + 62, + 236, + 181, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 181, + 316 + ], + "type": "image", + "image_path": "504f36ad21fa11bec55ef5980635dd1c99c4032b86a58040cef44c631928afc9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 318, + 168, + 329 + ], + "lines": [ + { + "bbox": [ + 74, + 318, + 168, + 329 + ], + "spans": [ + { + "bbox": [ + 74, + 318, + 168, + 329 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 74, + 318, + 168, + 329 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 74, + 318, + 168, + 329 + ], + "type": "text", + "content": " iteration rendering" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 185, + 236, + 303, + 316 + ], + "blocks": [ + { + "bbox": [ + 151, + 221, + 457, + 233 + ], + "lines": [ + { + "bbox": [ + 151, + 221, + 457, + 233 + ], + "spans": [ + { + "bbox": [ + 151, + 221, + 457, + 233 + ], + "type": "text", + "content": "Figure 2. Overall framework and intermediate result of each step of DeepDeblurRF." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 236, + 303, + 316 + ], + "lines": [ + { + "bbox": [ + 185, + 236, + 303, + 316 + ], + "spans": [ + { + "bbox": [ + 185, + 236, + 303, + 316 + ], + "type": "image", + "image_path": "5cdaface6f9b95374b79a2ca75a695e4791d0d4d7a21f30f27d7ed417b0d8d58.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 317, + 293, + 329 + ], + "lines": [ + { + "bbox": [ + 195, + 317, + 293, + 329 + ], + "spans": [ + { + "bbox": [ + 195, + 317, + 293, + 329 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 195, + 317, + 293, + 329 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 195, + 317, + 293, + 329 + ], + "type": "text", + "content": " iteration rendering" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 236, + 426, + 316 + ], + "blocks": [ + { + "bbox": [ + 308, + 236, + 426, + 316 + ], + "lines": [ + { + "bbox": [ + 308, + 236, + 426, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 236, + 426, + 316 + ], + "type": "image", + "image_path": "800ecb96940b15ebc0924d2cf61839dd9b1e78d39b3921c2e8121d0df988db7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 317, + 414, + 329 + ], + "lines": [ + { + "bbox": [ + 318, + 317, + 414, + 329 + ], + "spans": [ + { + "bbox": [ + 318, + 317, + 414, + 329 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 318, + 317, + 414, + 329 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 318, + 317, + 414, + 329 + ], + "type": "text", + "content": " iteration rendering" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": "Figure 3. As iterations progress, the rendered images contain increasingly high-quality scene information, which subsequently improves the RF-guided deblurring network's performance in the next iteration." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 430, + 236, + 549, + 316 + ], + "blocks": [ + { + "bbox": [ + 430, + 236, + 549, + 316 + ], + "lines": [ + { + "bbox": [ + 430, + 236, + 549, + 316 + ], + "spans": [ + { + "bbox": [ + 430, + 236, + 549, + 316 + ], + "type": "image", + "image_path": "32a8a4ba4397870008a4bba02cde3e91e2098621bce9d94fc7ad87e5ac495807.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 317, + 537, + 329 + ], + "lines": [ + { + "bbox": [ + 441, + 317, + 537, + 329 + ], + "spans": [ + { + "bbox": [ + 441, + 317, + 537, + 329 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 441, + 317, + 537, + 329 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 441, + 317, + 537, + 329 + ], + "type": "text", + "content": " iteration rendering" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "text", + "content": "ing to " + }, + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "inline_equation", + "content": "B_{m}" + }, + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "inline_equation", + "content": "D_{m}^{0}" + }, + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "text", + "content": ", and the set of the deblurred images as " + }, + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^0 = \\{D_1^0,\\dots,D_M^0\\}" + }, + { + "bbox": [ + 55, + 373, + 295, + 398 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 407, + 212, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 212, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 212, + 418 + ], + "type": "text", + "content": "3.2. Radiance Field Construction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "content": "In the radiance field construction step at the " + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "content": "-th iteration, we first estimate the camera poses of the deblurred images from the previous step, denoted as " + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{i - 1} = \\{D_1^{i - 1},\\dots,D_M^{i - 1}\\}" + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "content": " using COLMAP [36], where " + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "inline_equation", + "content": "i\\in [1,N]" + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "content": " is the index of our iterative process. Using the estimated poses and deblurred images, we aggregate information about the target 3D scene and construct a radiance field " + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i" + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "content": ". As discussed in Sec. 1, our framework can adopt diverse radiance field representations. To demonstrate the extensibility of our framework, we adopt two different representations, Plenoxels [12] and 3D Gaussians [17], in our experiments. We refer to our framework using Plenoxels and 3D Gaussians as DeepDeblurRF-P and DeepDeblurRF-G, respectively." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": "Note that deblurred images " + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{i-1}" + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": " from the previous step may still contain residual blur and deblurring artifacts, which can introduce corrupted information into the aggregation process. Nevertheless, since the input images have overlapping regions, other deblurred images can provide information for those regions that appear corrupted in some images. By aggregating such information from different deblurred images on the same regions, the radiance field construction step can effectively suppress corrupted information and obtain a higher-quality radiance field, from which we can render higher-quality images compared to the input deblurred im" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 373, + 382, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 382, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 382, + 385 + ], + "type": "text", + "content": "ages at this step." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 395, + 443, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 395, + 443, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 395, + 443, + 407 + ], + "type": "text", + "content": "3.3. RF-guided Deblurring" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": "The RF-guided deblurring step deblurs the input views " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": " using the aggregated information in " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": ". Specifically, for each input view " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "B_{m}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": ", we render " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": " to obtain a rendered image " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "R_{m}^{i}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": ". The rendered image " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "R_{m}^{i}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": " shares the same content as its corresponding deblurred image " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "D_{m}^{i - 1}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": ", but contains fewer artifacts and finer details. Motivated by this, our RF-guided deblurring leverages " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "R_{m}^{i}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": " to obtain an updated deblurred image " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "D_{m}^{i}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "inline_equation", + "content": "B_{m}" + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": "To perform RF-guided deblurring, we employ a novel RF-guided deblurring network that takes both the rendered image " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "R_{m}^{i}" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": " and the input blurred image " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "B_{m}" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": " to guide the deblurring process with the aggregated information in " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "R_{m}^{i}" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": ". For the deblurring network, we adopt NAFNet [4] and modify its first layer to take a concatenation of the two images. Once we obtain the updated deblurred images " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{i}" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": ", we return to the radiance field construction step and carry out the process for the " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": "-th iteration, gradually enhancing the quality of the radiance field and the deblurred images." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "The intermediate results of each step are visualized in Fig. 2-(b) to Fig. 2-(e). Due to the severe blur in the input blurred image, its initial deblurring result retains some residual blur. In contrast, thanks to the information aggregation occurring in the radiance field construction step, the rendered image contains more accurate details. Finally, thanks to the guidance of the rendered image, the RF-guided deblurring" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6015" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 380, + 208 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 380, + 208 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 380, + 208 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 380, + 208 + ], + "type": "image", + "image_path": "d269df9f5fcd7783dff492ad8c61ea7b71bc6ff0273d31af1d12141f1fe5fa9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 218, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 555, + 240 + ], + "type": "text", + "content": "Figure 4. Examples of the BlurRF-Synth and BlurRF-Real datasets. The examples show blurred views in BlurRF-Real, while the top and bottom rows in BlurRF-Synth include blurred views and their corresponding sharp views." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 552, + 208 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 552, + 208 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 552, + 208 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 552, + 208 + ], + "type": "image", + "image_path": "7794864bf3b5f9ed8280b9cfc307af75cd6e75a968ab0f2f60441b0ad3514e46.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "type": "text", + "content": "result shows sharper and more precise details that cannot be observed in the intermediate results of previous steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": "Fig. 3 visualizes rendered images at different iterations. Although the rendered image at the first iteration Fig. 3-(a) contains blurry details, they are gradually refined at each iteration (Fig. 3 (b)-(d)) due to our iterative process, which incrementally improves the quality of the radiance fields and the deblurred images. The effectiveness of our iterative process is also demonstrated by the increasing PSNR score as iterations progress." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 382, + 151, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 151, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 151, + 395 + ], + "type": "text", + "content": "4. BlurRF Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 402, + 145, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 145, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 145, + 415 + ], + "type": "text", + "content": "4.1. BlurRF-Synth" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 419, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 612 + ], + "type": "text", + "content": "We propose the BlurRF-Synth dataset, the first large-scale dataset for training and evaluating novel-view synthesis from blurred images. The dataset comprises a train set and a test set for each type of blur: camera motion blur and defocus blur. The train set includes 65 scenes, each containing 29 pairs of synthetically blurred images and their corresponding sharp images, captured from different viewpoints. The test set includes 10 scenes, each with 29 blurred-sharp image pairs and five additional sharp images from different viewpoints for evaluating novel-view synthesis quality. Examples of blurred images and ground-truth sharp images are shown in Fig. 4. We synthetically generated the dataset using Blender models, collecting 95 models from Blendswap1 under Creative Commons licenses and 5 models from the synthetic dataset of Deblur-NeRF [24]. For each model, we sampled 29 camera poses to capture different viewpoints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "type": "text", + "content": "Camera motion blur We simulated camera shakes by randomly sampling a camera trajectory for each pose during the exposure duration. To this end, we adopted Bezier interpolation to model 6-DOF camera motion and densely sampled 51 intermediate poses along this camera trajectory. We then rendered the scene using Blender at each intermediate camera pose to produce a series of sharp images, which we then" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 253, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 313 + ], + "type": "text", + "content": "averaged to create a blurred image. All rendering and averaging were performed in the linear sRGB color space to accurately replicate the real-world image formation process. Among the sharp images, we sampled the temporally central image (i.e., the 26th image) as the ground-truth sharp image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 315, + 556, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 556, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 556, + 410 + ], + "type": "text", + "content": "Defocus blur We also generated images with defocus blur using Blender. We adjusted the camera's depth-of-field (DoF) in Blender to create defocus blur. We controlled the aperture size and randomly set the blade count between 7 and 9 to simulate realistic defocus effects. For each camera pose, we sampled a random focal distance within a predefined range based on the scene scale, shifting the focal plane to introduce variation in the defocus blur." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 412, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 593 + ], + "type": "text", + "content": "Realistic blur synthesis To reflect real-world image degradation, we adopted the blur synthesis pipeline of RSBur [35]. Specifically, we generated camera motion blur by averaging sharp images in the linear sRGB space. Then, we synthesized saturated pixels to the blurred images and converted them to the camera RAW space. In the camera RAW space, we synthesized shot and read noise, and converted the images back to the camera sRGB space. For defocus blur, we rendered blurred images in the linear sRGB space and added noise in the same manner. For a more detailed pipeline, we refer the readers to [35]. The blur synthesis pipeline requires camera-specific noise parameters and a color correction matrix. We used a Sony A7R3 camera to estimate them. For more details about the BlurRF-Synth dataset, we refer the readers to the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 399, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 399, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 399, + 612 + ], + "type": "text", + "content": "4.2.BlurRF-Real" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "For evaluation under more realistic and challenging conditions, we propose BlurRF-Real, a real-world low-light camera motion blur dataset. The existing real datasets [22, 24] are captured in well-lit environments with minimal noise, making them less representative of camera motion blur in low-light conditions. Our dataset addresses this gap by providing data captured under challenging low-light scenarios. We collected five indoor scenes using a machine vision camera," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 178, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 178, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 178, + 713 + ], + "type": "text", + "content": "1https://blendswap.com/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6016" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 566, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 566, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 566, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 566, + 198 + ], + "type": "table", + "html": "
3D RepresentationModelCamera Motion BlurDefocus BlurComputation Time (Hr.)
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MLPDeblur-NeRF [24]27.670.83400.145030.030.87270.113731.33
BAD-NeRF [39]21.740.52980.3969---24.68
DP-NeRF [21]28.030.84120.126730.150.87630.099130.00*
Voxel gridExBluRF [22]26.560.78230.1955---8.92
PDRF-10 [31]28.330.84350.149530.030.87500.12254.26
DeepDeblurRF-P29.810.86680.114232.510.90580.09611.14
3D GaussiansDeblurring-3DGS [20]26.300.77290.172829.370.85450.14700.33
BAGS [32]27.410.81080.138229.900.86380.11521.25
DeepDeblurRF-G29.940.86810.105932.580.90600.07740.28
", + "image_path": "e5ee1d3e301b50dd75e5f5b47959b522f6dfebfd7e036167bf3b270b5940f791.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 243 + ], + "type": "text", + "content": "Table 1. Quantitative results of novel-view synthesis on BlurRF-Synth test scenes. We highlight the best metrics and the second best metrics. Note that the computation time of DP-NeRF [21] was measured with two GPUs due to its memory demands, whereas those of the other models were measured on a single GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 251, + 555, + 388 + ], + "blocks": [ + { + "bbox": [ + 56, + 251, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 555, + 388 + ], + "type": "image", + "image_path": "8257f61519d0fa50b13e87bcc9982ff97e69d54c59ee1c6ca76000630c7976b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 397, + 453, + 408 + ], + "lines": [ + { + "bbox": [ + 156, + 397, + 453, + 408 + ], + "spans": [ + { + "bbox": [ + 156, + 397, + 453, + 408 + ], + "type": "text", + "content": "Figure 5. Qualitative results of novel-view synthesis on BlurRF-Synth test scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 295, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 466 + ], + "type": "text", + "content": "each providing 20-40 multi-view blurry images, including 3-5 images for novel-view synthesis evaluation. Examples of blurred training views are shown in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 482, + 135, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 135, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 135, + 496 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 506, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 295, + 638 + ], + "type": "text", + "content": "Datasets The deblurring networks of DeepDeblurRF are trained with the train set of BlurRF-Synth. To evaluate DeepDeblurRF and other radiance field deblurring methods, we use the test set of BlurRF-Synth, the BlurRF-Real dataset, and the benchmark datasets from Deblur-NeRF [24]. The benchmark datasets include both synthetic and real-world scenes. The synthetic dataset consists of five scenes, while the real-world dataset contains ten scenes for each blur type. The real scenes of Deblur-NeRF were captured with a Canon EOS RP in manual exposure mode. The camera poses are estimated using COLMAP [36]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Implementation details We set the number of iterations " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " to 5. Thus, as detailed in Sec. 3, DeepDeblurRF includes one single-image deblurring network and four RF-guided deblurring networks for each blur type: camera motion blur and defocus blur. Note that our approach requires an additional training phase for the deblurring networks, unlike" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "text", + "content": "previous radiance field deblurring methods. Once trained, our framework can construct radiance fields for arbitrary scenes and render novel views, similar to conventional approaches. To train the deblurring networks, we use the Adam optimizer [18] with " + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.9" + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "text", + "content": ". The learning rate is initially set to " + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "text", + "content": " and gradually reduced to " + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "inline_equation", + "content": "10^{-7}" + }, + { + "bbox": [ + 313, + 430, + 555, + 514 + ], + "type": "text", + "content": " using the cosine annealing scheduler [23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 517, + 556, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 614 + ], + "type": "text", + "content": "For constructing radiance fields in DeepDeblurRF-P, we follow the forward-facing experimental setup in Plenoxels, applying pruning and upsampling to the voxel grids every 38,400 iterations. For DeepDeblurRF-G, we adopt the original Gaussian Splatting method, incorporating the depth-based pruning and additional points strategy from DeblurGS [20] to address the sparse point clouds obtained from COLMAP on blurred images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "We measured the computation time for radiance field deblurring using the camera motion blur test set of BlurRF-Synth on a single NVIDIA Titan RTX, except for DP-NeRF, which requires two GPUs due to its memory demands. Additional implementation details, including network configurations and the training settings for each iteration of DeepDeblurRF-P and DeepDeblurRF-G, are provided in the supplementary material." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6017" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 315 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 315 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 315 + ], + "type": "image", + "image_path": "7d73d4bdf292e3f39a5ceefcb869857eec8d9830f01e28cbbf5ff234fb5693e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 323, + 441, + 333 + ], + "lines": [ + { + "bbox": [ + 168, + 323, + 441, + 333 + ], + "spans": [ + { + "bbox": [ + 168, + 323, + 441, + 333 + ], + "type": "text", + "content": "Figure 6. Qualitative results of novel-view synthesis on real-world datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 338, + 555, + 468 + ], + "blocks": [ + { + "bbox": [ + 58, + 338, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 58, + 338, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 58, + 338, + 555, + 468 + ], + "type": "image", + "image_path": "3e248613e7f5667956334566827f3bea8df7fe67fc175849a19ba2c04d81abc6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 476, + 441, + 487 + ], + "lines": [ + { + "bbox": [ + 167, + 476, + 441, + 487 + ], + "spans": [ + { + "bbox": [ + 167, + 476, + 441, + 487 + ], + "type": "text", + "content": "Figure 7. Qualitative results of novel-view synthesis on BlurRF-SB dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 502, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 502, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 502, + 297, + 682 + ], + "type": "text", + "content": "Comparison using BlurRF-Synth We compare DeepDeblurRF with other radiance field deblurring methods [20-22, 24, 31, 32, 39]. Tab. 1 shows the quantitative results of DeepDeblurRF and the other methods on the test set of BlurRF-Synth. The table demonstrates that our approach outperforms the existing methods in terms of quality, regardless of the radiance field representation. Moreover, it achieves substantially shorter computation times. Specifically, DeepDeblurRF-P shows the shortest computation time among the approaches based on MLPs and voxel grids, while DeepDeblurRF-G is the fastest among all the methods. This superior performance can be attributed to its DNN-based deblurring modules, which handle blur efficiently and effectively, in contrast to the other methods constrained by linear blur models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The superior performance of our approach is further demonstrated in Fig. 5, which shows that our method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 502, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 556, + 598 + ], + "type": "text", + "content": "achieves sharper results on the BlurRF-Synth dataset. Specifically, on the camera motion blur scene, all the other methods produce noisy results, whereas both DeepDeblurRF-P and DeepDeblurRF-G effectively remove the blur even in challenging conditions with noise and pixel saturation, achieving high-quality novel-view synthesis. Similarly, in the defocus blur scene, our methods render sharper novel views by minimizing residual blur than the other methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "Comparison on real datasets We compare the novel-view synthesis performance of DeepDeblurRF against other approaches using both BlurRF-Real and the real scenes from Deblur-NeRF [24]. As detailed in Sec. 4.2, BlurRF-Real provides an evaluation of how well each method handles camera motion blur with non-linear outliers and noise in real-world low-light conditions. As shown in Tab. 2 and Fig. 6, our method not only outperforms the others in non-reference metrics [1, 26] but also achieves high-quality qualitative" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6018" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 303, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 303, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 303, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 303, + 186 + ], + "type": "table", + "html": "
ModelBlurRF-RealBlurRF-SB
NIQE (↓)ARNIQA (↑)PSNR (↑)SSIM (↑)LPIPS (↓)
MLP
Deblur-NeRF [24]6.3410.27925.430.72640.2319
BAD-NeRF [39]9.9080.23721.970.52310.4731
DP-NeRF [21]6.4980.26627.070.79400.1795
Voxel grid
ExBluRF [22]7.4790.27924.160.66490.2659
PDRF-10 [31]6.2430.26626.660.77300.2124
DeepDeblurRF-P5.8290.32329.450.84720.1677
3D Gaussians
Deblurring-3DGS [20]6.0860.26223.870.64590.2475
BAGS [32]5.8370.29024.490.67480.2391
DeepDeblurRF-G5.4230.32929.590.85480.1399
", + "image_path": "639e075c27d5729651d07f95b68b2d3d025edd18a782aa4c771063713d7fc462.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 319 + ], + "type": "text", + "content": "results, clearly surpassing the competing methods. It also shows superior performance on real scenes from Deblur-NeRF. Specifically, while other methods struggle with saturated pixels in the camera motion blur scene, our method effectively handles it. Likewise, in the defocus blur scene, it renders sharper novel views without residual blur." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 429 + ], + "type": "text", + "content": "While the real-world scenes of Deblur-NeRF provide blur-free reference images, they suffer from misalignment and exposure differences [24, 31, 32]. Similarly, BlurRF-Real also includes blur-free reference images, but they have not only similar limitations but also severe noise as they are captured under severe low-light conditions. Thus, we report quantitative evaluations against the reference images in the supplementary material. More experimental results are also provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "text", + "content": "Input views with same blur directions As discussed in Sec. 1, unlike previous approaches, our method does not rely on the assumption that input blurred images contain complementary information from different blur directions. To verify this, we generated an additional synthetic test set, named BlurRF-SB, where the blurred images of the same scene share the same blur directions, while their blur magnitudes are different. BlurRF-SB was generated using five scenes from the BlurRF-Synth test set, following the same generation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 662 + ], + "type": "text", + "content": "Fig. 7 shows a qualitative comparison on BlurRF-SB. Due to the lack of complementary information in different training views, the other methods fail to restore sharp details. Conversely, our methods successfully restore sharp details. The quantitative comparison in Tab. 2 also demonstrates that our approach significantly outperforms the others for blurred images with the same blur directions. These results verify the benefit of using prior knowledge on sharp images from pre-trained deblurring networks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Number of iterations We investigate the impact of the hyperparameter " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", which sets the number of iterations in our framework. Tab. 3 shows that DeepDeblurRF-G achieves improved novel-view synthesis as " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " increases, highlighting" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 552, + 159 + ], + "blocks": [ + { + "bbox": [ + 55, + 194, + 297, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 297, + 232 + ], + "type": "text", + "content": "Table 2. Quantitative results of novel-view synthesis on BlurRF-Real and BlurRF-SB datasets. We highlight the best metrics and the second best metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 71, + 552, + 159 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 552, + 159 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 552, + 159 + ], + "type": "table", + "html": "
#Iter.Camera motionDefocus
PSNRSSIMLPIPSPSNRSSIMLPIPS
N=128.400.81750.170730.740.88620.1297
N=229.230.84620.128131.220.89050.0955
N=329.650.85880.115232.120.90030.0899
N=429.830.86320.110832.300.90380.0860
N=529.940.86810.105932.580.90600.0774
N=629.970.86920.103432.750.90870.0736
", + "image_path": "932a83894cf7cf6cfb528c83c30d0068480e46f4f45f46ad62d09d86396aab64.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 167, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 200 + ], + "type": "text", + "content": "Table 3. Quantitative results of ablation study on the number of iterations " + }, + { + "bbox": [ + 313, + 167, + 555, + 200 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 167, + 555, + 200 + ], + "type": "text", + "content": " using DeepDeblurRF-G on the test sets of BlurRF-Synth." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 204, + 555, + 325 + ], + "blocks": [ + { + "bbox": [ + 316, + 204, + 555, + 325 + ], + "lines": [ + { + "bbox": [ + 316, + 204, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 204, + 555, + 325 + ], + "type": "image", + "image_path": "4d3e2eca8f11b9c8612d42165ee06a9c75fb908632ecc922d8f3ba4d49a429e1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "text", + "content": "Figure 8. Qualitative results of ablation study on the number of iterations " + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "text", + "content": " using DeepDeblurRF-G on the real-world scenes [24]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 369, + 556, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 441 + ], + "type": "text", + "content": "the effectiveness of our iterative approach. However, gains diminish beyond " + }, + { + "bbox": [ + 313, + 369, + 556, + 441 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 313, + 369, + 556, + 441 + ], + "type": "text", + "content": ", suggesting that " + }, + { + "bbox": [ + 313, + 369, + 556, + 441 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 313, + 369, + 556, + 441 + ], + "type": "text", + "content": " is sufficient. Fig. 8 further supports this with qualitative results on real scenes [24] (camera motion blur and defocus blur), showing similar trends. Similar results were observed for DeepDeblurRF-P, as detailed in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 616 + ], + "type": "text", + "content": "In this paper, we proposed DeepDeblurRF, a novel radiance field deblurring approach that leverages DNN-based deblurring to overcome the limitations of the linear blur model of previous methods. To integrate DNN-based deblurring with radiance field construction, we presented a novel RF-guided deblurring scheme and the framework that alternates between RF-guided deblurring and radiance field training. We also presented BlurRF-Synth, the first large-scale dataset for training and evaluating radiance field deblurring frameworks. Experimental results showed that our method outperforms others, handling both camera motion blur and defocus blur effectively with highly reduced computation time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "Limitations and future work Our method leverages deblurring networks to remove blur from training views, but restoring images with more severe blur than in the training data remains challenging. Our RF-guided deblurring uses only rendered images from each input view for guidance. Imposing additional priors on other views and incorporating additional information, e.g., depth maps rendered from radiance fields, may further improve the deblurring performance." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6019" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "Acknowledgements This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government(MSIT) (RS-2024-00457882, AI Research Hub Project, RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH), RS-2024-00437866, ITRC(Information Technology Research Center))." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 165, + 115, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 115, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 115, + 177 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 185, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 185, + 296, + 228 + ], + "type": "text", + "content": "[1] Lorenzo Agnolucci et al. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 229, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 273 + ], + "type": "text", + "content": "[2] Miika Aittala and Fredo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 296, + 316 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "text", + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 1, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 405 + ], + "type": "text", + "content": "[5] Wei-Ge Chen, N Nandhakumar, and Worthy N Martin. Image motion estimation from motion smear-a new computational model. IEEE transactions on pattern analysis and machine intelligence, 18(4):412-425, 1996. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 406, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 426 + ], + "type": "text", + "content": "[6] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Trans. Graph., 28(5):145:1-145:8, 2009. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 296, + 471 + ], + "type": "text", + "content": "[7] Sunghyun Cho, Yasuyuki Matsushita, and Seungyong Lee. Removing non-uniform motion blur from images. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 472, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 514 + ], + "type": "text", + "content": "[8] Sunghyun Cho, Jue Wang, and Seungyong Lee. Handling outliers in non-blind image deconvolution. In 2011 international conference on computer vision, pages 495-502. IEEE, 2011. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 516, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 296, + 548 + ], + "type": "text", + "content": "[9] Sunghyun Cho, Jue Wang, and Seungyong Lee. Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 548, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 296, + 602 + ], + "type": "text", + "content": "[10] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "type": "text", + "content": "[11] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[13] Chunzhi Gu, Xuequan Lu, Ying He, and Chao Zhang. Blur removal via blurred-noisy image pair. IEEE Transactions on Image Processing, 30:345-359, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "type": "text", + "content": "[14] Michael Hirsch, Christian J Schuler, Stefan Harmeling, and Bernhard Schölkopf. Fast removal of non-uniform camera shake. In 2011 International Conference on Computer Vision, pages 463-470. IEEE, 2011. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 206 + ], + "type": "text", + "content": "[15] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "text", + "content": "[16] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE international conference on computer vision, pages 4038-4047, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 554, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 307 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 341 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "text", + "content": "[19] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 430 + ], + "type": "text", + "content": "[20] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "type": "text", + "content": "[21] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 490, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 555, + 544 + ], + "type": "text", + "content": "[22] Dongwoo Lee, Jeongtaek Oh, Jaesung Rim, Sunghyun Cho, and Kyoung Mu Lee. Exblurf: Efficient radiance fields for extreme motion blurred images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17639-17648, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 555, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 555, + 577 + ], + "type": "text", + "content": "[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 580, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 633 + ], + "type": "text", + "content": "[24] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "text", + "content": "[25] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "text", + "content": "[26] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6020" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "text", + "content": "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "text", + "content": "[28] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "type": "text", + "content": "[29] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8102-8111, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "text", + "content": "[30] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2901–2908, 2014. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "text", + "content": "[31] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 363 + ], + "type": "text", + "content": "[32] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "text", + "content": "[33] Alex Rav-Acha and Shmuel Peleg. Two motion-blurred images are better than one. Pattern recognition letters, 26(3): 311-317, 2005. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 454 + ], + "type": "text", + "content": "[34] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "type": "text", + "content": "[35] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 533 + ], + "type": "text", + "content": "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 588 + ], + "type": "text", + "content": "[37] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "text", + "content": "[38] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "text", + "content": "[39] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. In 2010" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 341 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 491-498, 2010. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[41] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "text", + "content": "[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 554, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 230 + ], + "type": "text", + "content": "[43] Haichao Zhang and David Wipf. Non-uniform camera shake removal using a spatially-adaptive sparse penalty. Advances in Neural Information Processing Systems, 26, 2013. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "type": "text", + "content": "[44] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2482-2491, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "text", + "content": "[45] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, Haozhe Xie, Jinshan Pan, and Jimmy S Ren. Davanet: Stereo deblurring with view aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10996-11005, 2019. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6021" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_content_list.json b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c97d52105a4f0ebb4b3ae25da08e6d1c69e9d20 --- /dev/null +++ b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_content_list.json @@ -0,0 +1,1464 @@ +[ + { + "type": "text", + "text": "Exploiting Temporal State Space Sharing for Video Semantic Segmentation", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Syed Ariff Syed Hesham $^{1,2}$ , Yun Liu $^{3*}$ , Guolei Sun $^{4*}$ , Henghui Ding $^{5}$ , Jing Yang $^{6}$ , Ender Konukoglu $^{4}$ , Xue Geng $^{2}$ , Xudong Jiang $^{1}$", + "bbox": [ + 178, + 179, + 818, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanyang Technological University $^{2}$ Institute for Infocomm Research, A*STAR $^{3}$ VCIP, CS, Nankai University $^{4}$ CVL, ETH Zurich $^{5}$ Fudan University $^{6}$ Guizhou University", + "bbox": [ + 99, + 215, + 888, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video semantic segmentation (VSS) plays a vital role in understanding the temporal evolution of scenes. Traditional methods often segment videos frame-by-frame or in a short temporal window, leading to limited temporal context, redundant computations, and heavy memory requirements. To this end, we introduce a Temporal Video State Space Sharing (TV3S) architecture to leverage Mamba state space models for temporal feature sharing. Our model features a selective gating mechanism that efficiently propagates relevant information across video frames, eliminating the need for a memory-heavy feature pool. By processing spatial patches independently and incorporating shifted operation, TV3S supports highly parallel computation in both training and inference stages, which reduces the delay in sequential state space processing and improves the scalability for long video sequences. Moreover, TV3S incorporates information from prior frames during inference, achieving long-range temporal coherence and superior adaptability to extended sequences. Evaluations on the VSPW and Cityscapes datasets reveal that our approach outperforms current state-of-the-art methods, establishing a new standard for VSS with consistent results across long video sequences. By achieving a good balance between accuracy and efficiency, TV3S shows a significant advancement in spatiotemporal modeling, paving the way for efficient video analysis. The code is publicly available at https://github.com/Ashesham/TV3S.git.", + "bbox": [ + 88, + 318, + 485, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 757, + 220, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic Segmentation has achieved substantial progress through the introduction of convolutional neural networks [5, 40, 48, 49, 72] and, more recently with vision transformers [4, 14, 39] due to their ability to capture spatial patterns and contextual information. However, these techniques are mainly designed for static images and do not leverage the", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f112e19c8150db5ce5ff5d3ceb033e872524d5425a453123f1fd8c04ad3b187.jpg", + "image_caption": [ + "Figure 1. Comparison of the proposed TV3S with baseline models for VSS. By enhancing temporal information, our TV3S demonstrates superior performance over the baselines." + ], + "image_footnote": [], + "bbox": [ + 516, + 286, + 903, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "temporal dynamics present in videos [2, 11, 12, 23, 57, 67].", + "bbox": [ + 511, + 597, + 903, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, research attention has increasingly shifted toward video semantic segmentation (VSS) [36, 53, 54, 61], due to its potential to benefit a wide range of practical applications. Unlike static image segmentation, VSS requires models to track motion, adapt to changes in appearance, and handle interactions among objects across consecutive frames. These challenges necessitate sophisticated techniques to accurately interpret dynamic environments where objects may occlude one another, shift positions, or exhibit varying states of motion and stillness [7, 47].", + "bbox": [ + 509, + 613, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The contemporary approaches of VSS fall into two main categories. The first leverages optical flow, modeling pixel movement between consecutive frames to align features and support object tracking, thereby enhancing temporal coherence. These methods effectively propagate temporal information, making them useful for fine-grained temporal alignment [8, 13, 17, 50, 67]. However, they have significant drawbacks, including high computational costs due to the complexity of estimating accurate flow fields, especially in", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Yun Liu and Guolei Sun", + "bbox": [ + 107, + 886, + 374, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24211", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scenes with occlusions or sudden changes. Additionally, optical flow relies on precise motion estimation, where inaccuracies in dynamic scenes can propagate and lead to reduced accuracy in video segmentation.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second category involves feature aggregation, where information from multiple frames are combined to improve segmentation accuracy [28, 70]. The researchers relied on recurrent neural networks (RNNs) particularly Convolutional Long Short-Term Memory (ConvLSTM) [52] to introduce temporal structures to the models [15, 43, 46, 62, 64]. Recently, VM-RNN [60] has advanced this area by applying an efficient LSTM model to work with Vision Mamba [74] to capture spatiotemporal dynamics. While these methods can capture both spatial and temporal information, they face challenges when scaling to handle long video sequences, largely due to the high computational and memory demands of recurrent networks. Furthermore, the sensitivity of LSTMs to sequence length often leads to training instability, requiring careful tuning and optimization [46, 52, 64]. These limitations underscore the need for more scalable architectures.", + "bbox": [ + 89, + 154, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to such challenges in past feature aggregation methods, the focus in this domain has shifted towards the usage of transformer-based models [25, 29-31, 57-59, 65] [63]. However, these approaches still face challenges, including high memory costs and limited scalability, particularly with long video sequences or high-resolution frames. Efficient information sharing across numerous frames continues to be a challenge, making real-time applications difficult.", + "bbox": [ + 89, + 414, + 482, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges of integrating spatial and temporal information in VSS, we present Temporal State Space Sharing (TV3S), designed to overcome the limitations of existing models. Unlike traditional optical flow methods, which struggle with inaccuracies in dynamic scenes, and RNN- and Transformer-based architectures, which are constrained by expensive feature pooling and high computational costs, TV3S takes a more efficient approach by processing spatial patches in parallel while dynamically sharing temporal information across video frames. Specifically, a spatially encoded input frame is split into discrete patches, which are processed independently through a series of TV3S blocks, each containing Temporal State Space (TSS) modules with a selective gating mechanism that effectively integrates and propagates spatiotemporal information across frames, ensuring minimal computational overhead. To further enhance motion handling at the edges of patches, we introduce a shifted window-based approach that works with un-shifted and shifted encoded features, enabling the model to capture movements near the boundaries while still maintaining efficient temporal information sharing. These components work jointly to integrate both local and long-range temporal features, significantly improving VSS performance", + "bbox": [ + 89, + 537, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "while avoiding excessive computational costs and improving overall efficiency as evident in Fig. 1.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of our paper include:", + "bbox": [ + 532, + 121, + 872, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the Temporal Video State Space Sharing (TV3S) architecture, a novel framework that shares and propagates temporal information across video frames efficiently and effectively.", + "- We process spatial patches independently with a selective gating mechanism efficiently, thus enabling parallel computation during both training and inference and supporting scalability for long, high-resolution video sequences.", + "- We design a shifted window-based approach within the TV3S block, enhancing temporal state space sharing and capturing long-range spatial context effectively." + ], + "bbox": [ + 511, + 136, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments on the VSPW [42] and Cityscapes [9] datasets, we demonstrate that our approach surpasses existing state-of-the-art methods, establishing new benchmarks for efficiency and accuracy in VSS.", + "bbox": [ + 511, + 303, + 903, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 375, + 653, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic segmentation began with natural images, with the introduction of fully convolutional networks (FCNs) [40] pioneering an end-to-end pixel-wise classification framework. Building on this foundation, subsequent works enhanced segmentation accuracy by adopting atrous convolutional layers [5, 6], employing pyramid architectures [5, 6, 72], leveraging encoder-decoder architectures [10, 45, 48], and incorporating attention mechanisms [16, 24, 66]. More recently, transformer-based architectures like SegFormer [69], Segmenter [56], SETR [73], and Mask2Former [7], have further advanced the field by learning global dependencies. While these developments have significantly improved image semantic segmentation, transitioning them to VSS introduces new challenges related to temporal coherence and efficient exploitation of temporal information.", + "bbox": [ + 511, + 400, + 906, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. VSS Based on Recurrent Neural Networks", + "text_level": 1, + "bbox": [ + 511, + 635, + 874, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works in this field explored recurrent neural networks (RNNs) for temporal modeling. Valipour et al. [62] incorporated a recurrent unit between the encoder and decoder, significantly improving video segmentation performance. Evaluations by [15] on different RNN structures such as ConvRNN [55], ConvGRU [3], and ConvLSTM [52] on the KITTI dataset [18] demonstrated the superiority of ConvLSTM in handling video sequences. Further explorations have combined ConvGRU with optical flow to represent pixel displacements and maintain temporal continuity [44], while bidirectional ConvLSTM has been applied to merge temporally adjacent features, enhancing stability across frames [43]. Despite these advancements, the high computational and memory cost of RNNs for processing longer video sequences poses notable challenges, highlighting the need for more efficient models capable of handling", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24212", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03216e2b2bafeeaa9d1f4325fab384de104fa39baa4fcb2d2b20743361ff4acd.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed TV3S architecture, illustrating the encoder-decoder framework that employs state space models and our TSS module based on Mamba [19] for independent spatial and temporal processing." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 903, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "long-range temporal dependencies [46, 52].", + "bbox": [ + 89, + 439, + 380, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. VSS Based on CNNs and Transformers", + "text_level": 1, + "bbox": [ + 89, + 465, + 426, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address RNN limitations, CNN-based methods were developed for temporal modeling. DFF [75] and Accel [26] utilized optical flow to propagate features between frames, reducing redundancy. ClockNet [51] and LLVS [35] introduced adaptive feature reuse to exploit semantic similarity across frames for efficiency and temporal coherence.", + "bbox": [ + 89, + 491, + 482, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recognizing the limitations of optical flow and CNNs for long-range temporal modeling (see Sec. 1), recent advancements [1, 31, 33, 34] have seen a shift towards using transformers [63] in VSS. Transformers, with their self-attention mechanism, can capture global dependencies across frames, making them suited for temporal feature aggregation. Among notable approaches, MPVSS [67] proposes a memory-augmented transformer framework to capture multi-frame dependencies, enabling efficient temporal aggregation across longer video sequences. Similarly, CFFM [57] and MRCFA [58] employ multi-resolution cross-frame attention to handle temporal variations by disentangling static and dynamic contexts within video frames, allowing for a refined segmentation process that distinguishes between stationary and moving elements.", + "bbox": [ + 89, + 583, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite advancements, transformer-based methods face high computational costs due to the quadratic complexity of self-attention, especially with high-resolution frames or long videos [63]. Additionally, they are usually designed to learn temporal information in a short video sequence due to the high complexity. This gap highlights the need for more", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "efficient and holistic models that can simultaneously manage computational costs while effectively modeling long-range temporal information in VSS.", + "bbox": [ + 511, + 439, + 906, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. State Space Models", + "text_level": 1, + "bbox": [ + 513, + 498, + 697, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State space models (SSMs) present a promising alternative for temporal modeling, addressing the shortcomings of RNNs and transformers. Unlike RNNs, which suffer from scaling challenges with sequence length, SSMs like S4 [21, 22] exhibit with linear complexity by imposing diagonal structures on state matrices, making them more efficient for long data sequences. Enhanced through HiPPO [20] initialization, SSMs handle extensive dependencies while requiring less memory, making them well-suited for tasks requiring to store long temporal context. Recently, Mamba [19] introduced a selective-scan mechanism to process temporal data efficiently. This advancement has spurred adaptations in vision-specific models, such as Vision Mamba [74] and VMamba [38], which incorporate Mamba blocks in a hierarchical structure to overcome directional sensitivities and maintain scalability across high-resolution inputs. Additionally, Vim [74] refined the scanning techniques to prevent overfitting. Furthermore, U-Mamba [41] has explored a hybrid network architecture that combines SSMs with convolutional layers. VideoMamba [32] represents one of the early frameworks to leverage SSM-based modules for video understanding, focusing on video clip classification. However, it is limited by its offline processing, which increases memory and computational demands, restricts dense semantic segmentation, and hinders its ability", + "bbox": [ + 511, + 523, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24213", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b0e0a5e3e62b6931fef56838a88dd5b9e22fa32c6c0208648cbdd91c4e930599.jpg", + "image_caption": [ + "Figure 3. Internal structure of the TV3S block, illustrating the flow of internal operations, and the propagation of hidden states for efficient spatiotemporal integration." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 486, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to capture the global context of sequences.", + "bbox": [ + 89, + 470, + 375, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, despite these advancements, current vision-specific models largely rely on hierarchical designs fail to fully exploit varying temporal resolutions and context lengths, highlighting a critical need for improved mechanisms for temporal state sharing. To this end, our proposed TV3S architecture integrates parameter-efficient SSMs to enhance effective temporal state sharing, thus proving advantageous for VSS. This integration not only underscores the computational efficiency and scalability of SSMs in VSS but also emphasizes their suitability for tasks requiring long-range temporal coherence, representing a significant novelty in the field.", + "bbox": [ + 88, + 487, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 685, + 227, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the proposed TV3S architecture, which effectively captures and integrates temporal dependencies within video sequences using a state space modeling framework. It aims to leverage both temporal and spatial information to achieve accurate and consistent segmentation results.", + "bbox": [ + 89, + 710, + 485, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 800, + 287, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture processes a sequence of video frames denoted as $\\{I_{t - l},\\dots,I_{t - k},\\dots,I_{t - 1},I_t\\}$ , where $I_{t}$ represents the frame at time $t$ , and $\\{l,\\dots,k,\\dots,1\\}$ are temporal offsets capturing frames from the past relative to $t$ . These frames are fed into an image encoder such as MiT [69] and", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Swin [39]. The encoder produces a corresponding set of encoded feature $\\{E_{t - l},\\dots ,E_{t - k},\\dots ,E_{t - 1},E_t\\}$ , capturing rich spatial context.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The TV3S architecture processes these feature maps sequentially through a series of $N$ TV3S blocks (with a default value of $N = 4$ ), each containing two TSS modules that independently process spatial patches of the feature map, facilitating parallelization. Layer normalization is applied before each TSS module, and residual connections are included after each module to stabilize training. Each TSS module consists of operations including a linear projection, a 1-dimension depthwise convolution (DWConv1), a SiLU activation feeding into the state space model (S6), and finally ending with another linear projection to project back to the input dimension, as detailed in Fig. 3.", + "bbox": [ + 511, + 136, + 908, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Temporal Video State Space Sharing Block", + "text_level": 1, + "bbox": [ + 511, + 325, + 879, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The TV3S block is the core unit of the TV3S architecture, which enables seamless integration of temporal information across frames. Each block consists of two TSS modules that employ a state space model to capture and propagate temporal dependencies within frames. The encoded feature map $\\pmb{E}_{t}$ is partitioned into non-overlapping patches of size $w \\times w$ . Each patch $P_{t}^{i,j}$ is then flattened, transforming the spatial dimensions into a linear sequence suitable for parallel processing, with $i$ and $j$ representing the patch indices. Mathematically, $P_{t}^{i,j}$ can be expressed as", + "bbox": [ + 511, + 348, + 908, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {P} _ {t} ^ {i, j} = \\boldsymbol {E} _ {t} [ i: i + w, j: j + w,: ] \\in \\mathbb {R} ^ {C _ {E} \\times w \\times w}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 506, + 903, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {P}} _ {t} ^ {i, j} = \\operatorname {F l a t t e n} \\left(\\boldsymbol {P} _ {t} ^ {i, j}\\right) \\in \\mathbb {R} ^ {C _ {E} \\times (w \\times w)},\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 527, + 797, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C_E$ denotes the number of channels in $E_t$ after encoding $I_t$ . In contrast to traditional Mamba variants [38, 41, 74], which process sequences sequentially, our approach parallelizes the handling of each window or patch, offering two main advantages. First, the encoder effectively encodes complex spatial information across channels, eliminating for further processing of spatial relationships. Second, minimal changes between consecutive frames allow window-based processing to efficiently capture temporal dynamics, enabling a highly parallelized and computationally efficient model.", + "bbox": [ + 511, + 553, + 908, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State space model for temporal aggregation.", + "text_level": 1, + "bbox": [ + 511, + 724, + 823, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each TSS module within a TV3S block utilizes a state space model to capture temporal dependencies between frames. Given encoded feature map $\\pmb{E}_t$ , the hidden state $\\pmb{H}_t$ is updated using the current flattened patch $\\hat{\\pmb{P}}_t^{i,j}$ and the previous hidden state $\\pmb{H}_{t-1}$ :", + "bbox": [ + 511, + 739, + 906, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {H} _ {t} ^ {i, j} = f _ {A} (\\Delta , \\boldsymbol {A} _ {s}) \\boldsymbol {H} _ {t - 1} ^ {i, j} + f _ {B} (\\Delta , \\boldsymbol {A} _ {s}, \\boldsymbol {B} _ {s}) \\hat {\\boldsymbol {P}} _ {t} ^ {i, j}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 821, + 903, + 847 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {t} ^ {i, j} = \\boldsymbol {C} _ {s} \\boldsymbol {H} _ {t} ^ {i, j}.\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 843, + 656, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\pmb{A}_s \\in \\mathbb{R}^{N_s \\times N_s}$ , $\\pmb{B}_s \\in \\mathbb{R}^{N_s \\times C_E}$ , and $\\pmb{C}_s \\in \\mathbb{R}^{C_E \\times N_s}$ are learnable state space matrices. As seen in Eq. (2),", + "bbox": [ + 511, + 868, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24214", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the matrices $\\mathbf{A}$ and $\\mathbf{B}$ are discretized using the time scale parameter $\\Delta_{s}$ with fixed discretization functions $f_{A}(.)$ & $f_{B}(.)$ , referred as discretization rules, as described in [19] due to their many advantages. Here, $N_{s}$ denotes the hidden states dimensionality of the state space model and the aggregated feature output is denoted as $F_{t}^{i,j}$ . This mechanism enables each patch to independently capture and propagate its temporal dependencies across two dimensions $(i,j)$ . Consequently, temporal data storage consists of a total of $\\frac{W}{w} \\times \\frac{H}{w}$ hidden states, promoting effective temporal consistency and high parallelization.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shifting and edge handling. To enhance spatial context sharing, the reshaped feature maps undergo a shifting operation inspired by Swin Transformer [39]. This shift is parameterized as $s \\times s$ , where $s < w$ (typically $s = w / 2$ ). The shifting rearranges the blocks so that edge blocks receive information from adjacent regions, thereby enriching spatial-temporal representation. The shifted feature maps are then re-partitioned into patches, with those on the right and bottom edges further subdivided into smaller sub-blocks of dimensions $2 \\times w \\times \\frac{w}{2}$ , $2 \\times \\frac{w}{2} \\times w$ , and $4 \\times \\frac{w}{2} \\times \\frac{w}{2}$ , as visualized in Fig. 2. These re-partitioned patches are processed through a second TSS module in the same TV3S block, integrating the newly acquired spatial context with temporal information.", + "bbox": [ + 89, + 261, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the proposed architecture, each TV3S block consists of a pair of TSS modules, with one module processing unshifted blocks and the other handling shifted blocks. By stacking $N$ such TV3S blocks, the architecture forms a deep hierarchical structure termed TV3S architecture, which is capable of capturing intricate temporal dependencies across frames. The final aggregated feature representation, $\\pmb{F}^{N}$ , is passed through a linear projection layer to map it to $C$ classes, followed by an interpolation to produce the final segmentation output.", + "bbox": [ + 89, + 473, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 635, + 264, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training strategy for the TV3S architecture focuses on optimizing both temporal feature aggregation and robust spatial feature extraction to enable efficient learning of short- and long-term dependencies within video sequences, facilitating high VSS performance.", + "bbox": [ + 89, + 657, + 483, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, sequential input frames are taken at the intervals $\\{I_{t-9}, I_{t-6}, I_{t-3}, I_t\\}$ the same as [42]. Each of the frames in set $\\{I_{t-k}\\}$ with $k = \\{9, 6, 3, 0\\}$ , is passed through the encoder to generate encoded features $E_{t-k}$ encapsulating rich spatial information. An intermediate feature $\\hat{O}_{t-k}$ is extracted straight from $E_{t-k}$ with the use of linear projection that aligns the channel dimensions of $E_{t-k}$ with the number of classes $C$ in the dataset. Concurrently, the encoded features $E_{t-k}$ are fed into the TV3S blocks to integrate temporal information, producing aggregated features $F_{t-k}^N$ .", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Inference Procedure for TV3S Architecture" + ], + "code_body": "1: Input: Frame sequence $\\{I_t\\}$ \n2: Output: Segmentation map $O_t$ \n3: Initialize: Hidden states $\\pmb{H}$ for all TV3S blocks \n4: for each frame $\\pmb{I}_t$ in the sequence do \n5: $\\pmb{E}_t \\gets \\text{Encoder}(\\pmb{I}_t) \\quad //$ Extract features \n6: for each patch $(i,j)$ in $\\pmb{E}_t$ do \n7: $\\hat{\\pmb{P}}^{i,j} \\gets \\text{Flatten}(\\pmb{E}_t[i:i+w,j:j+w,:])$ \n8: for each TV3S block $n$ from 1 to $N$ do \n9: //Update Features and Hidden State. \n10: $\\pmb{F}_t^{n,i,j}, \\pmb{H}_t^{n,i,j} \\gets \\text{TV3S}^n(\\hat{\\pmb{P}}^{i,j}, \\pmb{H}^{n,i,j})$ \n11: end for \n12: end for \n13: $O_t \\gets \\text{Interpolate}(\\text{Linear}(\\pmb{F}_t^N)) \\quad //$ Segmentation \n14: $\\pmb{H} \\gets \\{\\pmb{H}_t\\} \\quad //$ Store for next frame \n15: end for", + "bbox": [ + 521, + 109, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With both output predictions extracted, a weighted cross-entropy between these output predictions $O$ and the ground-truth segmentation masks $M$ is computed as training loss for all frames in the input sequence. The training loss is computed as:", + "bbox": [ + 511, + 345, + 906, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\sum_ {k = \\{9, 6, 3, 0 \\}} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\hat {\\boldsymbol {O}} _ {t - k}, \\boldsymbol {M} _ {t - k}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {O} _ {t}, \\boldsymbol {M} _ {t}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 431, + 906, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss formulation presented in Eq. (3) employs a dual-loss strategy to ensure accurate segmentation of the final frame $O_{t}$ while preserving spatial relationships in the intermediate features $\\hat{O}_{t - k}$ from the encoder. By applying cross-entropy losses $\\mathcal{L}_{\\mathrm{CE}}$ , with intermediate losses weighted at $\\lambda = 0.5$ , the model effectively balances spatial feature learning and temporal coherence, enhancing its capability to extract temporally consistent spatial information for improved segmentation across consecutive frames.", + "bbox": [ + 511, + 474, + 906, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference Procedure", + "text_level": 1, + "bbox": [ + 513, + 619, + 705, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, the TV3S architecture processes each frame sequentially while leveraging stored hidden states to maintain temporal coherence across the video sequence, as detailed in Algorithm 1. For each frame $\\pmb{I}_t$ , the encoder first extracts the encoded features $\\pmb{E}_t$ , which are then partitioned into non-overlapping spatial patches indexed by $(i,j)$ and flattened into $\\pmb{P}_{i,j}$ . Each flattened patch is processed through the TV3S blocks, updating the current hidden states $\\pmb{H}_t$ using the previous states $\\pmb{H}_{t-1}$ . After processing all patches, the aggregated feature map $\\pmb{F}_t^N$ is reconstructed from the updated hidden states. This feature map is then passed through a linear projection layer and interpolated to match the original image resolution, producing the final segmentation map $\\pmb{O}_t$ . The updated hidden states $\\pmb{H}_t$ are then stored back in $\\pmb{H}$ for use with subsequent frames, ensuring continuous temporal integration and consistent segmentation results.", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24215", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f7ba27d6f3f93443f3109aaabd1282f3c2bf3cc9c213220fe152814e3ae41bc1.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative comparison of our model with existing methods on the VSPW dataset [42]. Our model achieves a strong balance between accuracy, model complexity, and operational speed. FPS and FLOPs are calculated with an input resolution of $480 \\times 853$ . (†Frame-by-Frame processing)" + ], + "table_body": "
MethodsBackbonesmIoU↑mVC8↑mVC16↑GFLOPs↓Params(M)↓FPS↑
Segformer† [69]MiT-B136.584.779.926.613.858.7
CFFM [57]MiT-B138.588.684.1-15.529.8
MRCFA [58]MiT-B138.988.884.4-16.240.1
TV3S (Ours)MiT-B140.090.787.036.917.324.7
Segformer [69]MiT-B243.986.081.2100.824.816.2
CFFM [57]MiT-B244.989.885.8143.226.510.1
MRCFA [58]MiT-B245.390.386.2127.927.310.7
TV3S (Ours)MiT-B246.391.588.3553.928.321.9
Mask2Former† [7]R5038.581.376.4110.644.019.4
MPVSS [67]R5037.584.177.238.984.133.9
Mask2Former† [7]R10139.382.577.6141.363.016.9
MPVSS [67]R10138.884.879.645.1103.132.3
DeepLabv3+† [6]R10134.783.278.2379.062.79.2
UperNet† [68]R10136.582.676.1403.683.216.0
PSPNet† [72]R10136.584.279.6401.870.513.8
OCRNet† [71]R10136.784.079.0361.758.114.3
TCB† [42]R10137.887.984.01692--
ETC† [37]OCRNet37.584.179.1361.7--
Segformer [69]MiT-B548.987.883.7185.082.19.4
CFFM [57]MiT-B549.390.887.1413.585.54.5
MRCFA [58]MiT-B549.990.987.4373.084.55.0
TV3S (Ours)MiT-B549.891.788.7137.085.614.0
Mask2Former† [7]Swin-T41.284.580.0114.447.417.1
MPVSS [67]Swin-T39.985.980.439.7114.032.8
TV3S (Ours)Swin-T44.988.083.557.331.722.9
Mask2Former† [7]Swin-S42.184.779.3152.268.914.5
MPVSS [67]Swin-S40.486.080.747.3108.030.6
TV3S (Ours)Swin-S50.689.685.894.153.119.5
", + "bbox": [ + 91, + 88, + 903, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 569, + 223, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 599, + 279, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The implementation of our approach is based on MMSegmentation codebase and all the experiments including training and inference were conducted with 2 A100 NVIDIA GPUs. The main experiments are done with the backbone same as the SegFormer (Variants with MiT and Swin) which are pre-trained with ImageNet. While the model is aimed to work with any number of frame sequences, during training, the model is trained with just three reference frames, $\\{k_1,k_2,k_3\\} = \\{-9, -6, -3\\}$ . More information on training is found in Sec. 3.3. To improve receptive field when processing the features from the backbone, the window size $w$ and the shift $s$ are set to 20 and 10 with number of TV3S blocks $N$ set to 4 (more information on these at Sec. 4.3) and the decoder made with mamba following its default parameters following [19] with the input embedding dimension of 256 matching the SegFormer implementation. The images from VSPW dataset [42] are cropped down to $480 \\times 480$ and", + "bbox": [ + 89, + 628, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are augmented using various augmentation methods during training, including cropping, resizing, flipping and addition of photometric distortion during training. Optimization of the model is done with the use AdamW optimizer and a \"poly\" learning rate schedule initializing the learning rate at $6e - 5$ . Testing is performed using the context of full video with the frame receiving the context of all the past frames within the video through the hidden states $\\pmb{H}$ due to its high efficiency & effectiveness following Sec. 3.4. Note that for all cases there was no use of post-processing on the obtained output like in [27].", + "bbox": [ + 511, + 571, + 906, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The experiments were mainly conducted with the use of VSPW dataset [42] which stands as one of the largest VSS benchmark. The dataset consists of training, validation and test subsets containing 2,806 clips (198,244 frames), 343 clips (24,502 frames) and 387 clips (28,887 frames) accordingly. The dataset consisting of a rich 124 categories with a dense annotation of frame rate of 15fps contrasts itself from the previously available datasets, which had very sparse annotation with just one frame being annotated every 10s of frames. Furthermore, the dataset cover", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24216", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/520d899f14f566d4d8fcff58f7c55a1ded872aa5d780f73402f1fd5e73ca5531.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBackbonesmIoU↑GFLOPs↓Params(M)↓FPS↑
ETC† [37]R1871.1434.1--
SegFormer† [69]MiT-B071.9-3.758.5
CFFM [57]MiT-B074.080.74.615.8
MRCFA [58]MiT-B072.877.54.216.6
SegFormer† [69]MiT-B174.1-13.846.8
CFFM [57]MiT-B175.1158.715.411.7
MRCFA [58]MiT-B175.1145.014.913.0
TV3S (ours)MiT-B175.683.617.325.1
", + "bbox": [ + 91, + 88, + 483, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bda2ca1de0f64045d88ad5882c602828c4e1206e339eaf38fb0a3e1c257462f9.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of our method with efficient alternative approaches on the Cityscapes dataset [9], using a resolution of $512 \\times 1024$ . (†Frame-by-Frame processing)" + ], + "table_footnote": [], + "table_body": "
ModelsEvaluation (mIoU)
AddConcatDirect
1 TSS (No Shift)37.637.838.0
TV3S (No Shift)38.138.438.9
TV3S (Shift)38.539.339.5
", + "bbox": [ + 91, + 276, + 483, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing various different scenarios of both outdoor and indoor scenes makes it suitable for training models to well verify the adaptability of the performance standing as the best benchmark in the field of VSS. While most experiments and training were done with the VSPW dataset [42], the proposed method was also evaluated on the Cityscapes dataset [9] which annotates one frame out of every 30 frames for benchmarking the results.", + "bbox": [ + 89, + 412, + 483, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We use mean IoU (mIoU) and mean video consistency (mVC) as key metrics. The metric mVC evaluates the smoothness of predicted segmentation maps over time, assessing performance in the temporal domain. More formally, given a video clip $\\{I_c\\}_{c=0}^{C_v}$ with ground truth masks $\\{M_c\\}_{c=1}^{C_v}$ and predicted outputs $\\{O_c\\}_{c=1}^{C_v}$ , $VC_n$ is computed as follows:", + "bbox": [ + 89, + 539, + 483, + 645 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V C} _ {n} = \\frac {1}{C _ {v} - n + 1} \\sum_ {i = 1} ^ {C _ {v} - n + 1} \\frac {\\left(\\cap_ {i} ^ {i + n - 1} \\boldsymbol {M} _ {i}\\right) \\cap \\left(\\cap_ {i} ^ {i + n - 1} \\boldsymbol {O} _ {i}\\right)}{\\cap_ {i} ^ {i + n - 1} \\boldsymbol {M} _ {i}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 657, + 482, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $C_{\\mathrm{v}} \\geq n$ . Once the $VC_{n}$ of all the videos are computed, their mean is computed to obtain the $\\mathrm{mVC}_{\\mathrm{n}}$ . Eq. (4) shows that $\\mathrm{mVC}_{\\mathrm{n}}$ finds the common areas of the predicted masks among frames which indicates the level of consistency of the prediction masks across time. More information on the metric can be found at [42].", + "bbox": [ + 89, + 712, + 483, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Models", + "text_level": 1, + "bbox": [ + 89, + 815, + 444, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model is compared against state-of-the-art models on the VSPW dataset [42], as shown in Tab. 1. The table is divided into three groups based on model size and backbone, offering insights at varying scales. In the first group", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5410ceb1170bd3edbaf85cf17bcb4735b92e8b63b761d0842cf69513f0fead9.jpg", + "table_caption": [ + "Table 3. Evaluation on the implication of the shifted mechanism and the output method." + ], + "table_footnote": [], + "table_body": "
Window sizesmIoUmVC8mVC16
438.589.084.6
638.689.284.7
1239.389.385.0
1639.289.385.0
2040.090.787.0
2840.088.884.2
3639.989.184.9
", + "bbox": [ + 522, + 88, + 893, + 212 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0d182467ffde0b7b1fd0ab5096b869dc796d95d1202c9b00fcffcecb6cc86b5a.jpg", + "table_caption": [ + "Table 4. Impact of the window size on model performance." + ], + "table_footnote": [], + "table_body": "
BackbonesTV3S BlocksmIoUmVC8mVC16
MiT-B1136.288.183.6
237.488.684.3
337.688.583.5
440.090.787.0
", + "bbox": [ + 514, + 234, + 905, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Performance metrics based on the number of TV3S blocks in the model.", + "bbox": [ + 511, + 323, + 905, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with small models (<30M parameters), our method outperforms the baselines, demonstrating efficiency even with limited model capacity. In the second group with larger models (>30M parameters), TV3S achieves near-state-of-the-art performance by effectively capturing rich contextual information. The third group focuses on the Swin Transformer backbone, where our model excels with over 8 mIoU ahead of the next best, highlighting its ability to preserve temporal correlations and ensure spatial accuracy. Overall, our method demonstrates superior visual consistency across all groups, further illustrated by Fig. 4 showcasing temporally consistent segmentation.", + "bbox": [ + 511, + 359, + 906, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also present results on the Cityscapes dataset [9] in Tab. 2, using smaller model variants at an input resolution of $1024 \\times 512$ . Due to the dataset's annotation structure, metrics such as $\\mathrm{mVC}_{8}$ and $\\mathrm{mVC}_{16}$ are not applicable, but our model still achieves top performance in mIoU, underscoring its strong generalizability across datasets.", + "bbox": [ + 511, + 542, + 905, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 643, + 666, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All ablation studies were conducted on the VSPW dataset [42] using the MiT-B1 backbone and following the same training and inference strategies as previously described.", + "bbox": [ + 511, + 667, + 905, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of shifted representations and feature integration. We evaluated the effectiveness of different TSS/TV3S decoder (see Tab. 3), including one TSS layer, unshifted TV3S, and shifted TV3S. Despite the only difference between unshifted and shifted TV3S being the feature representation, the results clearly show that shifted representation significantly outperforms its unshifted counterparts, highlighting their effectiveness. We further tested adding feature maps through addition and concatenation before the final prediction, but both methods reduced performance, suggesting that reintroducing spatial information after temporal modeling disrupts temporal coherence.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24217", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd706f4a7dc08bf41e498e59bffa00dfcc4e95279c405830fd08b4339586d0e2.jpg", + "image_caption": [ + "Figure 4. Qualitative example of our TV3S architecture compared with the current baseline. This displays both improved performance in performing spatial predictions and utilizing the temporal information to produce temporally consistent segmentation results." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 885, + 323 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b92327d9210825984e13a94a685def021ff89f3c17b1d100c9e19e6e8fffef69.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Number of FramesmIoUmVC8mVC16
137.984.579.1
239.286.981.8
439.588.984.2
839.790.485.9
1639.791.287.2
3239.791.587.8
", + "bbox": [ + 91, + 358, + 483, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Performance based on the amount of temporal information available during inference", + "bbox": [ + 89, + 474, + 482, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of window size. TV3S establishes temporal relationships with window size controlling the spatial context during processing. Larger windows help summarize stable scenes but may overload the model in dynamic settings, while smaller windows enhance understanding in fast-changing environments but may miss correlations in slower scenes. As shown in Tab. 4, the optimal performance is achieved with a $20 \\times 20$ window size, striking a balance between spatial context and model efficiency.", + "bbox": [ + 89, + 507, + 483, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect on the number of TV3S blocks. In our experiments, we assessed the impact of varying the number of TV3S blocks, as shown in Tab. 5. Increasing the number of blocks improved mIoU from 36.18 with one block to 40.00 with four blocks, along with enhanced temporal consistency, as shown by the $\\mathrm{mVC}_{8}$ and $\\mathrm{mVC}_{16}$ scores. This upward trend suggests that stacking TV3S blocks better aggregates temporal features, capturing complex relationships without introducing excessive redundancy.", + "bbox": [ + 89, + 648, + 483, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Temporal Context", + "text_level": 1, + "bbox": [ + 89, + 800, + 266, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of long temporal context. Given Mamba's RNN-like structure, we conducted an evaluation of its performance on the VSPW dataset [42] with varying temporal context sizes that ranged from 1 to 32 frames. In our approach, the videos that were shorter than 32 frames were", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "excluded from the and at the same time the first 32 frames were omitted during inference to ensure a fair and unbiased evaluation of the model's capabilities. The findings presented in Tab. 6 indicate that there is a notable improvement in performance as the temporal context sizes increase. However, this enhancement reaches a plateau once the temporal information saturates. This saturation effect is likely attributed to the limited changes that can be captured within smaller temporal windows, which causes diminishing returns in performance improvements beyond a certain threshold.", + "bbox": [ + 511, + 361, + 906, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 537, + 633, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work proposes a TV3S architecture that addresses the challenges of VSS by capturing temporal dynamics. The proposed structure is designed to achieve good computational efficiency and scalability. By leveraging independent processing of spatial blocks through state space models enhanced with Mamba [19], our approach enables parallel computation during both training and inference. This design mitigates the time delay and high memory demand typically associated with sequential processing in traditional state space and recurrent models, making it highly suitable for long, high-resolution video sequences. Extensive experiments on the VSPW [42] and Cityscapes [9] datasets show that the TV3S architecture not only surpasses existing methods in segmentation accuracy but also significantly improves temporal prediction consistency.", + "bbox": [ + 511, + 564, + 906, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future works. Our model enhances the encoder's focus on key temporal features, integrating spatial and temporal information for improved video frame segmentation. Its adaptability extends to tasks like object detection and action recognition, while its multi-modal data integration offers new research opportunities in audio-visual learning, emphasizing temporal synchronization.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24218", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Ministry of Education Singapore Tier 1 project funding grant No. MOE Tier 1 RG98/24 and by Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021).", + "bbox": [ + 89, + 114, + 480, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 186, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. In BMVC, 2023. 3", + "[2] Gowtham Atluri, Anuj Karpatne, and Vipin Kumar. Spatiotemporal data mining: A survey of problems and methods. ACM Computing Surveys, 51(4):1-41, 2018. 1", + "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 2", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4): 834-848, 2017. 1, 2", + "[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 2, 6", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 2, 6", + "[8] Jingchun Cheng, Yi-Hsuan Tsai, Shengjin Wang, and Ming-Hsuan Yang. SegFlow: Joint learning for video object segmentation and optical flow. In ICCV, pages 686-695, 2017. 1", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 2, 7, 8", + "[10] Henghui Ding, Xudong Jiang, Bing Shuai, Ai Qun Liu, and Gang Wang. Context contrasted feature and gated multiscale aggregation for scene segmentation. In CVPR, pages 2393-2402, 2018. 2", + "[11] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1", + "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1" + ], + "bbox": [ + 93, + 231, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In AAAI, pages 10713-10720, 2020. 1", + "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1", + "[15] Ekrem Emre Yurdakul and Yucel Yemez. Semantic segmentation of rgbd videos with recurrent fully convolutional neural networks. In ICCV workshops, pages 367-374, 2017. 2", + "[16] Jun Fu, Jing Liu, Hajie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 2", + "[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, pages 641–649, 2023. 1", + "[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2", + "[19] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 5, 6, 8", + "[20] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. HiPPO: Recurrent memory with optimal polynomial projections. In NeurIPS, pages 1474–1487, 2020. 3", + "[21] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021. 3", + "[22] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. In NeurIPS, pages 572-585, 2021. 3", + "[23] Shunan Guo, Zhuochen Jin, David Gotz, Fan Du, Hongyuan Zha, and Nan Cao. Visual progression analysis of event sequence data. IEEE TVCG, 25(1):417-426, 2018. 1", + "[24] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2", + "[25] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In CVPR, pages 8818-8827, 2020. 2", + "[26] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In CVPR, pages 8866-8875, 2019. 3", + "[27] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6", + "[28] Abhijit Kundu, Vibhav Vineet, and Vladlen Koltun. Feature space optimization for semantic video segmentation. In CVPR, pages 3168-3175, 2016. 2", + "[29] Jiangwei Lao, Weixiang Hong, Xin Guo, Yingying Zhang, Jian Wang, Jingdong Chen, and Wei Chu. Simultaneously short-and long-term temporal modeling for semi-supervised video semantic segmentation. In CVPR, pages 14763-14772, 2023. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24219", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Jiangyun Li, Yikai Zhao, Jun Fu, Jiajia Wu, and Jing Liu. Attention-guided network for semantic video segmentation. IEEE Access, 7:140680-140689, 2019.", + "[31] Jiangtong Li, Wentao Wang, Junjie Chen, Li Niu, Jianlou Si, Chen Qian, and Liqing Zhang. Video semantic segmentation via sparse temporal transformer. In ACM MM, pages 59-68, 2021. 2, 3", + "[32] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. In ECCV, pages 237-255, 2025. 3", + "[33] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video K-Net: A simple, strong, and unified baseline for video segmentation. In CVPR, pages 18847-18857, 2022. 3", + "[34] Xiangtai Li, Haobo Yuan, Wenwei Zhang, Guangliang Cheng, Jiangmiao Pang, and Chen Change Loy. Tube-Link: A flexible cross tube framework for universal video segmentation. In ICCV, pages 13923-13933, 2023. 3", + "[35] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In CVPR, pages 5997-6005, 2018. 3", + "[36] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60-88, 2017. 1", + "[37] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In ECCV, pages 352-368, 2020. 6, 7", + "[38] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 3, 4", + "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 4, 5", + "[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2", + "[41] Jun Ma, Feifei Li, and Bo Wang. U-Mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3, 4", + "[42] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. VSPW: A large-scale dataset for video scene parsing in the wild. In CVPR, pages 4133-4143, 2021. 2, 5, 6, 7, 8", + "[43] Seyed Shahabeddin Nabavi, Mrigank Rochan, and Yang Wang. Future semantic segmentation with convolutional LSTM. CoRR, 2018. 2", + "[44] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In CVPR, pages 6819-6828, 2018. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Pedro O O Pinheiro, Ronan Collobert, and Piotr Dólár. Learning to segment object candidates. In NeurIPS, pages 1441-1447, 2015. 2", + "[46] Andreas Pfeuffer, Karina Schulz, and Klaus Dietmayer. Semantic segmentation of video sequences with convolutional LSTMs. In IEEE Intelligent Vehicles Symposium, pages 1441-1447, 2019. 2, 3", + "[47] Rui Qian, Weiyao Lin, John See, and Dian Li. Controllable augmentations for video representation learning. Visual Intelligence, 2(1):1, 2024. 1", + "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015. 1, 2", + "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 1", + "[50] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In CVPR, pages 3889-3898, 2016. 1", + "[51] Evan Shelhamer, Kate Rakelly, Judy Hoffman, and Trevor Darrell. Clockwork convnets for video semantic segmentation. In ECCV, pages 852-868, 2016. 3", + "[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, pages 802-810, 2015. 2, 3", + "[53] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPR Workshop, pages 587-597, 2018. 1", + "[54] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1", + "[55] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 2", + "[56] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2", + "[57] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In CVPR, pages 3126-3137, 2022. 1, 2, 3, 6, 7", + "[58] Guolei Sun, Yun Liu, Hao Tang, Ajad Chhatkuli, Le Zhang, and Luc Van Gool. Mining relations among cross-frame affinities for video semantic segmentation. In ECCV, pages 522-539, 2022. 3, 6, 7", + "[59] Guolei Sun, Yun Liu, Henghui Ding, Min Wu, and Luc Van Gool. Learning local and global temporal contexts for video semantic segmentation. IEEE TPAMI, 46(10):6919-6934, 2024. 2", + "[60] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. VMRNN: Integrating vision mamba and" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24220", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "LSTM for efficient and accurate spatiotemporal forecasting. In CVPR Workshop, pages 5663-5673, 2024. 2", + "[61] Vassilios Tsakanikas and Tasos Dagiuklas. Video surveillance systems-current status and future trends. Computers & Electrical Engineering, 70:736-753, 2018. 1", + "[62] Sepehr Valipour, Mennatullah Siam, Martin Jagersand, and Nilanjan Ray. Recurrent fully convolutional networks for video segmentation. In IEEE Winter Conference on Applications of Computer Vision, pages 29-36, 2017. 2", + "[63] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 3", + "[64] Bowen Wang, Liangzhi Li, Yuta Nakashima, Ryo Kawasaki, Hajime Nagahara, and Yasushi Yagi. Noisy-LSTM: Improving temporal awareness for video semantic segmentation. IEEE Access, 9:46810-46820, 2021. 2", + "[65] Hao Wang, Weining Wang, and Jing Liu. Temporal memory attention for video semantic segmentation. In ICIP, pages 2254-2258, 2021. 2", + "[66] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 2", + "[67] Yuetian Weng, Mingfei Han, Haoyu He, Mingjie Li, Lina Yao, Xiaojun Chang, and Bohan Zhuang. Mask propagation for efficient video semantic segmentation. In NeurIPS, pages 7170-7183, 2024. 1, 3, 6", + "[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018. 6", + "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 2, 4, 6, 7", + "[70] Chenliang Xu, Caiming Xiong, and Jason J Corso. Streaming hierarchical video segmentation. In ECCV, pages 626-639, 2012. 2", + "[71] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, pages 173-190, 2020. 6", + "[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 1, 2, 6", + "[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2", + "[74] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3, 4", + "[75] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In CVPR, pages 2349-2358, 2017. 3" + ], + "bbox": [ + 91, + 90, + 482, + 872 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24221", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_model.json b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..09cbe7ddc3714cd1bbd76056abb54c6ff92fe174 --- /dev/null +++ b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "Exploiting Temporal State Space Sharing for Video Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.18, + 0.819, + 0.217 + ], + "angle": 0, + "content": "Syed Ariff Syed Hesham\\(^{1,2}\\), Yun Liu\\(^{3*}\\), Guolei Sun\\(^{4*}\\), Henghui Ding\\(^{5}\\), Jing Yang\\(^{6}\\), Ender Konukoglu\\(^{4}\\), Xue Geng\\(^{2}\\), Xudong Jiang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.217, + 0.89, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\) Nanyang Technological University \\(^{2}\\) Institute for Infocomm Research, A*STAR \\(^{3}\\) VCIP, CS, Nankai University \\(^{4}\\) CVL, ETH Zurich \\(^{5}\\) Fudan University \\(^{6}\\) Guizhou University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.486, + 0.728 + ], + "angle": 0, + "content": "Video semantic segmentation (VSS) plays a vital role in understanding the temporal evolution of scenes. Traditional methods often segment videos frame-by-frame or in a short temporal window, leading to limited temporal context, redundant computations, and heavy memory requirements. To this end, we introduce a Temporal Video State Space Sharing (TV3S) architecture to leverage Mamba state space models for temporal feature sharing. Our model features a selective gating mechanism that efficiently propagates relevant information across video frames, eliminating the need for a memory-heavy feature pool. By processing spatial patches independently and incorporating shifted operation, TV3S supports highly parallel computation in both training and inference stages, which reduces the delay in sequential state space processing and improves the scalability for long video sequences. Moreover, TV3S incorporates information from prior frames during inference, achieving long-range temporal coherence and superior adaptability to extended sequences. Evaluations on the VSPW and Cityscapes datasets reveal that our approach outperforms current state-of-the-art methods, establishing a new standard for VSS with consistent results across long video sequences. By achieving a good balance between accuracy and efficiency, TV3S shows a significant advancement in spatiotemporal modeling, paving the way for efficient video analysis. The code is publicly available at https://github.com/Ashesham/TV3S.git." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.758, + 0.222, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Semantic Segmentation has achieved substantial progress through the introduction of convolutional neural networks [5, 40, 48, 49, 72] and, more recently with vision transformers [4, 14, 39] due to their ability to capture spatial patterns and contextual information. However, these techniques are mainly designed for static images and do not leverage the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.287, + 0.905, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.545, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 1. Comparison of the proposed TV3S with baseline models for VSS. By enhancing temporal information, our TV3S demonstrates superior performance over the baselines." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.598, + 0.905, + 0.613 + ], + "angle": 0, + "content": "temporal dynamics present in videos [2, 11, 12, 23, 57, 67]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.614, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Recently, research attention has increasingly shifted toward video semantic segmentation (VSS) [36, 53, 54, 61], due to its potential to benefit a wide range of practical applications. Unlike static image segmentation, VSS requires models to track motion, adapt to changes in appearance, and handle interactions among objects across consecutive frames. These challenges necessitate sophisticated techniques to accurately interpret dynamic environments where objects may occlude one another, shift positions, or exhibit varying states of motion and stillness [7, 47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The contemporary approaches of VSS fall into two main categories. The first leverages optical flow, modeling pixel movement between consecutive frames to align features and support object tracking, thereby enhancing temporal coherence. These methods effectively propagate temporal information, making them useful for fine-grained temporal alignment [8, 13, 17, 50, 67]. However, they have significant drawbacks, including high computational costs due to the complexity of estimating accurate flow fields, especially in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.375, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors: Yun Liu and Guolei Sun" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24211" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "scenes with occlusions or sudden changes. Additionally, optical flow relies on precise motion estimation, where inaccuracies in dynamic scenes can propagate and lead to reduced accuracy in video segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.412 + ], + "angle": 0, + "content": "The second category involves feature aggregation, where information from multiple frames are combined to improve segmentation accuracy [28, 70]. The researchers relied on recurrent neural networks (RNNs) particularly Convolutional Long Short-Term Memory (ConvLSTM) [52] to introduce temporal structures to the models [15, 43, 46, 62, 64]. Recently, VM-RNN [60] has advanced this area by applying an efficient LSTM model to work with Vision Mamba [74] to capture spatiotemporal dynamics. While these methods can capture both spatial and temporal information, they face challenges when scaling to handle long video sequences, largely due to the high computational and memory demands of recurrent networks. Furthermore, the sensitivity of LSTMs to sequence length often leads to training instability, requiring careful tuning and optimization [46, 52, 64]. These limitations underscore the need for more scalable architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Due to such challenges in past feature aggregation methods, the focus in this domain has shifted towards the usage of transformer-based models [25, 29-31, 57-59, 65] [63]. However, these approaches still face challenges, including high memory costs and limited scalability, particularly with long video sequences or high-resolution frames. Efficient information sharing across numerous frames continues to be a challenge, making real-time applications difficult." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To address the challenges of integrating spatial and temporal information in VSS, we present Temporal State Space Sharing (TV3S), designed to overcome the limitations of existing models. Unlike traditional optical flow methods, which struggle with inaccuracies in dynamic scenes, and RNN- and Transformer-based architectures, which are constrained by expensive feature pooling and high computational costs, TV3S takes a more efficient approach by processing spatial patches in parallel while dynamically sharing temporal information across video frames. Specifically, a spatially encoded input frame is split into discrete patches, which are processed independently through a series of TV3S blocks, each containing Temporal State Space (TSS) modules with a selective gating mechanism that effectively integrates and propagates spatiotemporal information across frames, ensuring minimal computational overhead. To further enhance motion handling at the edges of patches, we introduce a shifted window-based approach that works with un-shifted and shifted encoded features, enabling the model to capture movements near the boundaries while still maintaining efficient temporal information sharing. These components work jointly to integrate both local and long-range temporal features, significantly improving VSS performance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "while avoiding excessive computational costs and improving overall efficiency as evident in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.122, + 0.874, + 0.137 + ], + "angle": 0, + "content": "In summary, the contributions of our paper include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.197 + ], + "angle": 0, + "content": "- We present the Temporal Video State Space Sharing (TV3S) architecture, a novel framework that shares and propagates temporal information across video frames efficiently and effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- We process spatial patches independently with a selective gating mechanism efficiently, thus enabling parallel computation during both training and inference and supporting scalability for long, high-resolution video sequences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- We design a shifted window-based approach within the TV3S block, enhancing temporal state space sharing and capturing long-range spatial context effectively." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Through extensive experiments on the VSPW [42] and Cityscapes [9] datasets, we demonstrate that our approach surpasses existing state-of-the-art methods, establishing new benchmarks for efficiency and accuracy in VSS." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.376, + 0.655, + 0.391 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.628 + ], + "angle": 0, + "content": "Semantic segmentation began with natural images, with the introduction of fully convolutional networks (FCNs) [40] pioneering an end-to-end pixel-wise classification framework. Building on this foundation, subsequent works enhanced segmentation accuracy by adopting atrous convolutional layers [5, 6], employing pyramid architectures [5, 6, 72], leveraging encoder-decoder architectures [10, 45, 48], and incorporating attention mechanisms [16, 24, 66]. More recently, transformer-based architectures like SegFormer [69], Segmenter [56], SETR [73], and Mask2Former [7], have further advanced the field by learning global dependencies. While these developments have significantly improved image semantic segmentation, transitioning them to VSS introduces new challenges related to temporal coherence and efficient exploitation of temporal information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.875, + 0.651 + ], + "angle": 0, + "content": "2.1. VSS Based on Recurrent Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Early works in this field explored recurrent neural networks (RNNs) for temporal modeling. Valipour et al. [62] incorporated a recurrent unit between the encoder and decoder, significantly improving video segmentation performance. Evaluations by [15] on different RNN structures such as ConvRNN [55], ConvGRU [3], and ConvLSTM [52] on the KITTI dataset [18] demonstrated the superiority of ConvLSTM in handling video sequences. Further explorations have combined ConvGRU with optical flow to represent pixel displacements and maintain temporal continuity [44], while bidirectional ConvLSTM has been applied to merge temporally adjacent features, enhancing stability across frames [43]. Despite these advancements, the high computational and memory cost of RNNs for processing longer video sequences poses notable challenges, highlighting the need for more efficient models capable of handling" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24212" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.087, + 0.904, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed TV3S architecture, illustrating the encoder-decoder framework that employs state space models and our TSS module based on Mamba [19] for independent spatial and temporal processing." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.44, + 0.381, + 0.456 + ], + "angle": 0, + "content": "long-range temporal dependencies [46, 52]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.467, + 0.428, + 0.483 + ], + "angle": 0, + "content": "2.2. VSS Based on CNNs and Transformers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.581 + ], + "angle": 0, + "content": "To address RNN limitations, CNN-based methods were developed for temporal modeling. DFF [75] and Accel [26] utilized optical flow to propagate features between frames, reducing redundancy. ClockNet [51] and LLVS [35] introduced adaptive feature reuse to exploit semantic similarity across frames for efficiency and temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Recognizing the limitations of optical flow and CNNs for long-range temporal modeling (see Sec. 1), recent advancements [1, 31, 33, 34] have seen a shift towards using transformers [63] in VSS. Transformers, with their self-attention mechanism, can capture global dependencies across frames, making them suited for temporal feature aggregation. Among notable approaches, MPVSS [67] proposes a memory-augmented transformer framework to capture multi-frame dependencies, enabling efficient temporal aggregation across longer video sequences. Similarly, CFFM [57] and MRCFA [58] employ multi-resolution cross-frame attention to handle temporal variations by disentangling static and dynamic contexts within video frames, allowing for a refined segmentation process that distinguishes between stationary and moving elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Despite advancements, transformer-based methods face high computational costs due to the quadratic complexity of self-attention, especially with high-resolution frames or long videos [63]. Additionally, they are usually designed to learn temporal information in a short video sequence due to the high complexity. This gap highlights the need for more" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.907, + 0.486 + ], + "angle": 0, + "content": "efficient and holistic models that can simultaneously manage computational costs while effectively modeling long-range temporal information in VSS." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.499, + 0.698, + 0.515 + ], + "angle": 0, + "content": "2.3. State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "State space models (SSMs) present a promising alternative for temporal modeling, addressing the shortcomings of RNNs and transformers. Unlike RNNs, which suffer from scaling challenges with sequence length, SSMs like S4 [21, 22] exhibit with linear complexity by imposing diagonal structures on state matrices, making them more efficient for long data sequences. Enhanced through HiPPO [20] initialization, SSMs handle extensive dependencies while requiring less memory, making them well-suited for tasks requiring to store long temporal context. Recently, Mamba [19] introduced a selective-scan mechanism to process temporal data efficiently. This advancement has spurred adaptations in vision-specific models, such as Vision Mamba [74] and VMamba [38], which incorporate Mamba blocks in a hierarchical structure to overcome directional sensitivities and maintain scalability across high-resolution inputs. Additionally, Vim [74] refined the scanning techniques to prevent overfitting. Furthermore, U-Mamba [41] has explored a hybrid network architecture that combines SSMs with convolutional layers. VideoMamba [32] represents one of the early frameworks to leverage SSM-based modules for video understanding, focusing on video clip classification. However, it is limited by its offline processing, which increases memory and computational demands, restricts dense semantic segmentation, and hinders its ability" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24213" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.488, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.486, + 0.464 + ], + "angle": 0, + "content": "Figure 3. Internal structure of the TV3S block, illustrating the flow of internal operations, and the propagation of hidden states for efficient spatiotemporal integration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.377, + 0.487 + ], + "angle": 0, + "content": "to capture the global context of sequences." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.488, + 0.485, + 0.671 + ], + "angle": 0, + "content": "However, despite these advancements, current vision-specific models largely rely on hierarchical designs fail to fully exploit varying temporal resolutions and context lengths, highlighting a critical need for improved mechanisms for temporal state sharing. To this end, our proposed TV3S architecture integrates parameter-efficient SSMs to enhance effective temporal state sharing, thus proving advantageous for VSS. This integration not only underscores the computational efficiency and scalability of SSMs in VSS but also emphasizes their suitability for tasks requiring long-range temporal coherence, representing a significant novelty in the field." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.686, + 0.228, + 0.705 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.486, + 0.79 + ], + "angle": 0, + "content": "Fig. 2 illustrates the proposed TV3S architecture, which effectively captures and integrates temporal dependencies within video sequences using a state space modeling framework. It aims to leverage both temporal and spatial information to achieve accurate and consistent segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.288, + 0.817 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The architecture processes a sequence of video frames denoted as \\(\\{I_{t - l},\\dots,I_{t - k},\\dots,I_{t - 1},I_t\\}\\), where \\(I_{t}\\) represents the frame at time \\(t\\), and \\(\\{l,\\dots,k,\\dots,1\\}\\) are temporal offsets capturing frames from the past relative to \\(t\\). These frames are fed into an image encoder such as MiT [69] and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "Swin [39]. The encoder produces a corresponding set of encoded feature \\(\\{E_{t - l},\\dots ,E_{t - k},\\dots ,E_{t - 1},E_t\\}\\), capturing rich spatial context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.909, + 0.319 + ], + "angle": 0, + "content": "The TV3S architecture processes these feature maps sequentially through a series of \\( N \\) TV3S blocks (with a default value of \\( N = 4 \\)), each containing two TSS modules that independently process spatial patches of the feature map, facilitating parallelization. Layer normalization is applied before each TSS module, and residual connections are included after each module to stabilize training. Each TSS module consists of operations including a linear projection, a 1-dimension depthwise convolution (DWConv1), a SiLU activation feeding into the state space model (S6), and finally ending with another linear projection to project back to the input dimension, as detailed in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.326, + 0.88, + 0.342 + ], + "angle": 0, + "content": "3.2. Temporal Video State Space Sharing Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.909, + 0.502 + ], + "angle": 0, + "content": "The TV3S block is the core unit of the TV3S architecture, which enables seamless integration of temporal information across frames. Each block consists of two TSS modules that employ a state space model to capture and propagate temporal dependencies within frames. The encoded feature map \\( \\pmb{E}_{t} \\) is partitioned into non-overlapping patches of size \\( w \\times w \\). Each patch \\( P_{t}^{i,j} \\) is then flattened, transforming the spatial dimensions into a linear sequence suitable for parallel processing, with \\( i \\) and \\( j \\) representing the patch indices. Mathematically, \\( P_{t}^{i,j} \\) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.507, + 0.905, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {P} _ {t} ^ {i, j} = \\boldsymbol {E} _ {t} [ i: i + w, j: j + w,: ] \\in \\mathbb {R} ^ {C _ {E} \\times w \\times w}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.529, + 0.799, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {P}} _ {t} ^ {i, j} = \\operatorname {F l a t t e n} \\left(\\boldsymbol {P} _ {t} ^ {i, j}\\right) \\in \\mathbb {R} ^ {C _ {E} \\times (w \\times w)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.72 + ], + "angle": 0, + "content": "where \\( C_E \\) denotes the number of channels in \\( E_t \\) after encoding \\( I_t \\). In contrast to traditional Mamba variants [38, 41, 74], which process sequences sequentially, our approach parallelizes the handling of each window or patch, offering two main advantages. First, the encoder effectively encodes complex spatial information across channels, eliminating for further processing of spatial relationships. Second, minimal changes between consecutive frames allow window-based processing to efficiently capture temporal dynamics, enabling a highly parallelized and computationally efficient model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.825, + 0.74 + ], + "angle": 0, + "content": "State space model for temporal aggregation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.908, + 0.817 + ], + "angle": 0, + "content": "Each TSS module within a TV3S block utilizes a state space model to capture temporal dependencies between frames. Given encoded feature map \\( \\pmb{E}_t \\), the hidden state \\( \\pmb{H}_t \\) is updated using the current flattened patch \\( \\hat{\\pmb{P}}_t^{i,j} \\) and the previous hidden state \\( \\pmb{H}_{t-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.822, + 0.905, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {H} _ {t} ^ {i, j} = f _ {A} (\\Delta , \\boldsymbol {A} _ {s}) \\boldsymbol {H} _ {t - 1} ^ {i, j} + f _ {B} (\\Delta , \\boldsymbol {A} _ {s}, \\boldsymbol {B} _ {s}) \\hat {\\boldsymbol {P}} _ {t} ^ {i, j}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.844, + 0.658, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {t} ^ {i, j} = \\boldsymbol {C} _ {s} \\boldsymbol {H} _ {t} ^ {i, j}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.869, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Here, \\( \\pmb{A}_s \\in \\mathbb{R}^{N_s \\times N_s} \\), \\( \\pmb{B}_s \\in \\mathbb{R}^{N_s \\times C_E} \\), and \\( \\pmb{C}_s \\in \\mathbb{R}^{C_E \\times N_s} \\) are learnable state space matrices. As seen in Eq. (2)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24214" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "the matrices \\( \\mathbf{A} \\) and \\( \\mathbf{B} \\) are discretized using the time scale parameter \\( \\Delta_{s} \\) with fixed discretization functions \\( f_{A}(.) \\) & \\( f_{B}(.) \\), referred as discretization rules, as described in [19] due to their many advantages. Here, \\( N_{s} \\) denotes the hidden states dimensionality of the state space model and the aggregated feature output is denoted as \\( F_{t}^{i,j} \\). This mechanism enables each patch to independently capture and propagate its temporal dependencies across two dimensions \\( (i,j) \\). Consequently, temporal data storage consists of a total of \\( \\frac{W}{w} \\times \\frac{H}{w} \\) hidden states, promoting effective temporal consistency and high parallelization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.472 + ], + "angle": 0, + "content": "Shifting and edge handling. To enhance spatial context sharing, the reshaped feature maps undergo a shifting operation inspired by Swin Transformer [39]. This shift is parameterized as \\( s \\times s \\), where \\( s < w \\) (typically \\( s = w / 2 \\)). The shifting rearranges the blocks so that edge blocks receive information from adjacent regions, thereby enriching spatial-temporal representation. The shifted feature maps are then re-partitioned into patches, with those on the right and bottom edges further subdivided into smaller sub-blocks of dimensions \\( 2 \\times w \\times \\frac{w}{2} \\), \\( 2 \\times \\frac{w}{2} \\times w \\), and \\( 4 \\times \\frac{w}{2} \\times \\frac{w}{2} \\), as visualized in Fig. 2. These re-partitioned patches are processed through a second TSS module in the same TV3S block, integrating the newly acquired spatial context with temporal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.626 + ], + "angle": 0, + "content": "In the proposed architecture, each TV3S block consists of a pair of TSS modules, with one module processing unshifted blocks and the other handling shifted blocks. By stacking \\(N\\) such TV3S blocks, the architecture forms a deep hierarchical structure termed TV3S architecture, which is capable of capturing intricate temporal dependencies across frames. The final aggregated feature representation, \\(\\pmb{F}^{N}\\), is passed through a linear projection layer to map it to \\(C\\) classes, followed by an interpolation to produce the final segmentation output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.265, + 0.652 + ], + "angle": 0, + "content": "3.3. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.735 + ], + "angle": 0, + "content": "The training strategy for the TV3S architecture focuses on optimizing both temporal feature aggregation and robust spatial feature extraction to enable efficient learning of short- and long-term dependencies within video sequences, facilitating high VSS performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.904 + ], + "angle": 0, + "content": "During training, sequential input frames are taken at the intervals \\(\\{I_{t-9}, I_{t-6}, I_{t-3}, I_t\\}\\) the same as [42]. Each of the frames in set \\(\\{I_{t-k}\\}\\) with \\(k = \\{9, 6, 3, 0\\}\\), is passed through the encoder to generate encoded features \\(E_{t-k}\\) encapsulating rich spatial information. An intermediate feature \\(\\hat{O}_{t-k}\\) is extracted straight from \\(E_{t-k}\\) with the use of linear projection that aligns the channel dimensions of \\(E_{t-k}\\) with the number of classes \\(C\\) in the dataset. Concurrently, the encoded features \\(E_{t-k}\\) are fed into the TV3S blocks to integrate temporal information, producing aggregated features \\(F_{t-k}^N\\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.091, + 0.892, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Inference Procedure for TV3S Architecture" + }, + { + "type": "algorithm", + "bbox": [ + 0.522, + 0.111, + 0.892, + 0.337 + ], + "angle": 0, + "content": "1: Input: Frame sequence \\(\\{I_t\\}\\) \n2: Output: Segmentation map \\(O_t\\) \n3: Initialize: Hidden states \\(\\pmb{H}\\) for all TV3S blocks \n4: for each frame \\(\\pmb{I}_t\\) in the sequence do \n5: \\(\\pmb{E}_t \\gets \\text{Encoder}(\\pmb{I}_t) \\quad //\\) Extract features \n6: for each patch \\((i,j)\\) in \\(\\pmb{E}_t\\) do \n7: \\(\\hat{\\pmb{P}}^{i,j} \\gets \\text{Flatten}(\\pmb{E}_t[i:i+w,j:j+w,:])\\) \n8: for each TV3S block \\(n\\) from 1 to \\(N\\) do \n9: //Update Features and Hidden State. \n10: \\(\\pmb{F}_t^{n,i,j}, \\pmb{H}_t^{n,i,j} \\gets \\text{TV3S}^n(\\hat{\\pmb{P}}^{i,j}, \\pmb{H}^{n,i,j})\\) \n11: end for \n12: end for \n13: \\(O_t \\gets \\text{Interpolate}(\\text{Linear}(\\pmb{F}_t^N)) \\quad //\\) Segmentation \n14: \\(\\pmb{H} \\gets \\{\\pmb{H}_t\\} \\quad //\\) Store for next frame \n15: end for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.424 + ], + "angle": 0, + "content": "With both output predictions extracted, a weighted cross-entropy between these output predictions \\( O \\) and the ground-truth segmentation masks \\( M \\) is computed as training loss for all frames in the input sequence. The training loss is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.433, + 0.907, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\sum_ {k = \\{9, 6, 3, 0 \\}} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\hat {\\boldsymbol {O}} _ {t - k}, \\boldsymbol {M} _ {t - k}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {O} _ {t}, \\boldsymbol {M} _ {t}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Loss formulation presented in Eq. (3) employs a dual-loss strategy to ensure accurate segmentation of the final frame \\( O_{t} \\) while preserving spatial relationships in the intermediate features \\( \\hat{O}_{t - k} \\) from the encoder. By applying cross-entropy losses \\( \\mathcal{L}_{\\mathrm{CE}} \\), with intermediate losses weighted at \\( \\lambda = 0.5 \\), the model effectively balances spatial feature learning and temporal coherence, enhancing its capability to extract temporally consistent spatial information for improved segmentation across consecutive frames." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.707, + 0.636 + ], + "angle": 0, + "content": "3.4. Inference Procedure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "During inference, the TV3S architecture processes each frame sequentially while leveraging stored hidden states to maintain temporal coherence across the video sequence, as detailed in Algorithm 1. For each frame \\( \\pmb{I}_t \\), the encoder first extracts the encoded features \\( \\pmb{E}_t \\), which are then partitioned into non-overlapping spatial patches indexed by \\( (i,j) \\) and flattened into \\( \\pmb{P}_{i,j} \\). Each flattened patch is processed through the TV3S blocks, updating the current hidden states \\( \\pmb{H}_t \\) using the previous states \\( \\pmb{H}_{t-1} \\). After processing all patches, the aggregated feature map \\( \\pmb{F}_t^N \\) is reconstructed from the updated hidden states. This feature map is then passed through a linear projection layer and interpolated to match the original image resolution, producing the final segmentation map \\( \\pmb{O}_t \\). The updated hidden states \\( \\pmb{H}_t \\) are then stored back in \\( \\pmb{H} \\) for use with subsequent frames, ensuring continuous temporal integration and consistent segmentation results." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24215" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.516 + ], + "angle": 0, + "content": "
MethodsBackbonesmIoU↑mVC8↑mVC16↑GFLOPs↓Params(M)↓FPS↑
Segformer† [69]MiT-B136.584.779.926.613.858.7
CFFM [57]MiT-B138.588.684.1-15.529.8
MRCFA [58]MiT-B138.988.884.4-16.240.1
TV3S (Ours)MiT-B140.090.787.036.917.324.7
Segformer [69]MiT-B243.986.081.2100.824.816.2
CFFM [57]MiT-B244.989.885.8143.226.510.1
MRCFA [58]MiT-B245.390.386.2127.927.310.7
TV3S (Ours)MiT-B246.391.588.3553.928.321.9
Mask2Former† [7]R5038.581.376.4110.644.019.4
MPVSS [67]R5037.584.177.238.984.133.9
Mask2Former† [7]R10139.382.577.6141.363.016.9
MPVSS [67]R10138.884.879.645.1103.132.3
DeepLabv3+† [6]R10134.783.278.2379.062.79.2
UperNet† [68]R10136.582.676.1403.683.216.0
PSPNet† [72]R10136.584.279.6401.870.513.8
OCRNet† [71]R10136.784.079.0361.758.114.3
TCB† [42]R10137.887.984.01692--
ETC† [37]OCRNet37.584.179.1361.7--
Segformer [69]MiT-B548.987.883.7185.082.19.4
CFFM [57]MiT-B549.390.887.1413.585.54.5
MRCFA [58]MiT-B549.990.987.4373.084.55.0
TV3S (Ours)MiT-B549.891.788.7137.085.614.0
Mask2Former† [7]Swin-T41.284.580.0114.447.417.1
MPVSS [67]Swin-T39.985.980.439.7114.032.8
TV3S (Ours)Swin-T44.988.083.557.331.722.9
Mask2Former† [7]Swin-S42.184.779.3152.268.914.5
MPVSS [67]Swin-S40.486.080.747.3108.030.6
TV3S (Ours)Swin-S50.689.685.894.153.119.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.522, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of our model with existing methods on the VSPW dataset [42]. Our model achieves a strong balance between accuracy, model complexity, and operational speed. FPS and FLOPs are calculated with an input resolution of \\(480 \\times 853\\). (†Frame-by-Frame processing)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.57, + 0.224, + 0.587 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.281, + 0.617 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Implementation details. The implementation of our approach is based on MMSegmentation codebase and all the experiments including training and inference were conducted with 2 A100 NVIDIA GPUs. The main experiments are done with the backbone same as the SegFormer (Variants with MiT and Swin) which are pre-trained with ImageNet. While the model is aimed to work with any number of frame sequences, during training, the model is trained with just three reference frames, \\(\\{k_1,k_2,k_3\\} = \\{-9, -6, -3\\}\\). More information on training is found in Sec. 3.3. To improve receptive field when processing the features from the backbone, the window size \\(w\\) and the shift \\(s\\) are set to 20 and 10 with number of TV3S blocks \\(N\\) set to 4 (more information on these at Sec. 4.3) and the decoder made with mamba following its default parameters following [19] with the input embedding dimension of 256 matching the SegFormer implementation. The images from VSPW dataset [42] are cropped down to \\(480 \\times 480\\) and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.739 + ], + "angle": 0, + "content": "are augmented using various augmentation methods during training, including cropping, resizing, flipping and addition of photometric distortion during training. Optimization of the model is done with the use AdamW optimizer and a \"poly\" learning rate schedule initializing the learning rate at \\(6e - 5\\). Testing is performed using the context of full video with the frame receiving the context of all the past frames within the video through the hidden states \\(\\pmb{H}\\) due to its high efficiency & effectiveness following Sec. 3.4. Note that for all cases there was no use of post-processing on the obtained output like in [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Datasets. The experiments were mainly conducted with the use of VSPW dataset [42] which stands as one of the largest VSS benchmark. The dataset consists of training, validation and test subsets containing 2,806 clips (198,244 frames), 343 clips (24,502 frames) and 387 clips (28,887 frames) accordingly. The dataset consisting of a rich 124 categories with a dense annotation of frame rate of 15fps contrasts itself from the previously available datasets, which had very sparse annotation with just one frame being annotated every 10s of frames. Furthermore, the dataset cover" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24216" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.485, + 0.216 + ], + "angle": 0, + "content": "
MethodsBackbonesmIoU↑GFLOPs↓Params(M)↓FPS↑
ETC† [37]R1871.1434.1--
SegFormer† [69]MiT-B071.9-3.758.5
CFFM [57]MiT-B074.080.74.615.8
MRCFA [58]MiT-B072.877.54.216.6
SegFormer† [69]MiT-B174.1-13.846.8
CFFM [57]MiT-B175.1158.715.411.7
MRCFA [58]MiT-B175.1145.014.913.0
TV3S (ours)MiT-B175.683.617.325.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.262 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of our method with efficient alternative approaches on the Cityscapes dataset [9], using a resolution of \\(512 \\times 1024\\). (†Frame-by-Frame processing)" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.37 + ], + "angle": 0, + "content": "
ModelsEvaluation (mIoU)
AddConcatDirect
1 TSS (No Shift)37.637.838.0
TV3S (No Shift)38.138.438.9
TV3S (Shift)38.539.339.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.483, + 0.404 + ], + "angle": 0, + "content": "Table 3. Evaluation on the implication of the shifted mechanism and the output method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.534 + ], + "angle": 0, + "content": "ing various different scenarios of both outdoor and indoor scenes makes it suitable for training models to well verify the adaptability of the performance standing as the best benchmark in the field of VSS. While most experiments and training were done with the VSPW dataset [42], the proposed method was also evaluated on the Cityscapes dataset [9] which annotates one frame out of every 30 frames for benchmarking the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.646 + ], + "angle": 0, + "content": "Evaluation metrics. We use mean IoU (mIoU) and mean video consistency (mVC) as key metrics. The metric mVC evaluates the smoothness of predicted segmentation maps over time, assessing performance in the temporal domain. More formally, given a video clip \\(\\{I_c\\}_{c=0}^{C_v}\\) with ground truth masks \\(\\{M_c\\}_{c=1}^{C_v}\\) and predicted outputs \\(\\{O_c\\}_{c=1}^{C_v}\\), \\(VC_n\\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V C} _ {n} = \\frac {1}{C _ {v} - n + 1} \\sum_ {i = 1} ^ {C _ {v} - n + 1} \\frac {\\left(\\cap_ {i} ^ {i + n - 1} \\boldsymbol {M} _ {i}\\right) \\cap \\left(\\cap_ {i} ^ {i + n - 1} \\boldsymbol {O} _ {i}\\right)}{\\cap_ {i} ^ {i + n - 1} \\boldsymbol {M} _ {i}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.805 + ], + "angle": 0, + "content": "where \\(C_{\\mathrm{v}} \\geq n\\). Once the \\(VC_{n}\\) of all the videos are computed, their mean is computed to obtain the \\(\\mathrm{mVC}_{\\mathrm{n}}\\). Eq. (4) shows that \\(\\mathrm{mVC}_{\\mathrm{n}}\\) finds the common areas of the predicted masks among frames which indicates the level of consistency of the prediction masks across time. More information on the metric can be found at [42]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.816, + 0.445, + 0.833 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our model is compared against state-of-the-art models on the VSPW dataset [42], as shown in Tab. 1. The table is divided into three groups based on model size and backbone, offering insights at varying scales. In the first group" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.089, + 0.895, + 0.213 + ], + "angle": 0, + "content": "
Window sizesmIoUmVC8mVC16
438.589.084.6
638.689.284.7
1239.389.385.0
1639.289.385.0
2040.090.787.0
2840.088.884.2
3639.989.184.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.216, + 0.885, + 0.23 + ], + "angle": 0, + "content": "Table 4. Impact of the window size on model performance." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.236, + 0.906, + 0.318 + ], + "angle": 0, + "content": "
BackbonesTV3S BlocksmIoUmVC8mVC16
MiT-B1136.288.183.6
237.488.684.3
337.688.583.5
440.090.787.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.324, + 0.906, + 0.351 + ], + "angle": 0, + "content": "Table 5. Performance metrics based on the number of TV3S blocks in the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.542 + ], + "angle": 0, + "content": "with small models (<30M parameters), our method outperforms the baselines, demonstrating efficiency even with limited model capacity. In the second group with larger models (>30M parameters), TV3S achieves near-state-of-the-art performance by effectively capturing rich contextual information. The third group focuses on the Swin Transformer backbone, where our model excels with over 8 mIoU ahead of the next best, highlighting its ability to preserve temporal correlations and ensure spatial accuracy. Overall, our method demonstrates superior visual consistency across all groups, further illustrated by Fig. 4 showcasing temporally consistent segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.906, + 0.635 + ], + "angle": 0, + "content": "We also present results on the Cityscapes dataset [9] in Tab. 2, using smaller model variants at an input resolution of \\(1024 \\times 512\\). Due to the dataset's annotation structure, metrics such as \\(\\mathrm{mVC}_{8}\\) and \\(\\mathrm{mVC}_{16}\\) are not applicable, but our model still achieves top performance in mIoU, underscoring its strong generalizability across datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.667, + 0.661 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.906, + 0.716 + ], + "angle": 0, + "content": "All ablation studies were conducted on the VSPW dataset [42] using the MiT-B1 backbone and following the same training and inference strategies as previously described." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Impact of shifted representations and feature integration. We evaluated the effectiveness of different TSS/TV3S decoder (see Tab. 3), including one TSS layer, unshifted TV3S, and shifted TV3S. Despite the only difference between unshifted and shifted TV3S being the feature representation, the results clearly show that shifted representation significantly outperforms its unshifted counterparts, highlighting their effectiveness. We further tested adding feature maps through addition and concatenation before the final prediction, but both methods reduced performance, suggesting that reintroducing spatial information after temporal modeling disrupts temporal coherence." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24217" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.887, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.324, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 4. Qualitative example of our TV3S architecture compared with the current baseline. This displays both improved performance in performing spatial predictions and utilizing the temporal information to produce temporally consistent segmentation results." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.359, + 0.485, + 0.472 + ], + "angle": 0, + "content": "
Number of FramesmIoUmVC8mVC16
137.984.579.1
239.286.981.8
439.588.984.2
839.790.485.9
1639.791.287.2
3239.791.587.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.503 + ], + "angle": 0, + "content": "Table 6. Performance based on the amount of temporal information available during inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.645 + ], + "angle": 0, + "content": "Impact of window size. TV3S establishes temporal relationships with window size controlling the spatial context during processing. Larger windows help summarize stable scenes but may overload the model in dynamic settings, while smaller windows enhance understanding in fast-changing environments but may miss correlations in slower scenes. As shown in Tab. 4, the optimal performance is achieved with a \\(20 \\times 20\\) window size, striking a balance between spatial context and model efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.788 + ], + "angle": 0, + "content": "Effect on the number of TV3S blocks. In our experiments, we assessed the impact of varying the number of TV3S blocks, as shown in Tab. 5. Increasing the number of blocks improved mIoU from 36.18 with one block to 40.00 with four blocks, along with enhanced temporal consistency, as shown by the \\(\\mathrm{mVC}_{8}\\) and \\(\\mathrm{mVC}_{16}\\) scores. This upward trend suggests that stacking TV3S blocks better aggregates temporal features, capturing complex relationships without introducing excessive redundancy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.267, + 0.817 + ], + "angle": 0, + "content": "4.4. Temporal Context" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Importance of long temporal context. Given Mamba's RNN-like structure, we conducted an evaluation of its performance on the VSPW dataset [42] with varying temporal context sizes that ranged from 1 to 32 frames. In our approach, the videos that were shorter than 32 frames were" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.907, + 0.527 + ], + "angle": 0, + "content": "excluded from the and at the same time the first 32 frames were omitted during inference to ensure a fair and unbiased evaluation of the model's capabilities. The findings presented in Tab. 6 indicate that there is a notable improvement in performance as the temporal context sizes increase. However, this enhancement reaches a plateau once the temporal information saturates. This saturation effect is likely attributed to the limited changes that can be captured within smaller temporal windows, which causes diminishing returns in performance improvements beyond a certain threshold." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.634, + 0.555 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.791 + ], + "angle": 0, + "content": "The work proposes a TV3S architecture that addresses the challenges of VSS by capturing temporal dynamics. The proposed structure is designed to achieve good computational efficiency and scalability. By leveraging independent processing of spatial blocks through state space models enhanced with Mamba [19], our approach enables parallel computation during both training and inference. This design mitigates the time delay and high memory demand typically associated with sequential processing in traditional state space and recurrent models, making it highly suitable for long, high-resolution video sequences. Extensive experiments on the VSPW [42] and Cityscapes [9] datasets show that the TV3S architecture not only surpasses existing methods in segmentation accuracy but also significantly improves temporal prediction consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Future works. Our model enhances the encoder's focus on key temporal features, integrating spatial and temporal information for improved video frame segmentation. Its adaptability extends to tasks like object detection and action recognition, while its multi-modal data integration offers new research opportunities in audio-visual learning, emphasizing temporal synchronization." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24218" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.482, + 0.191 + ], + "angle": 0, + "content": "This work is supported by the Ministry of Education Singapore Tier 1 project funding grant No. MOE Tier 1 RG98/24 and by Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.187, + 0.222 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[1] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. In BMVC, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[2] Gowtham Atluri, Anuj Karpatne, and Vipin Kumar. Spatiotemporal data mining: A survey of problems and methods. ACM Computing Surveys, 51(4):1-41, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.483, + 0.358 + ], + "angle": 0, + "content": "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4): 834-848, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[8] Jingchun Cheng, Yi-Hsuan Tsai, Shengjin Wang, and Ming-Hsuan Yang. SegFlow: Joint learning for video object segmentation and optical flow. In ICCV, pages 686-695, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[10] Henghui Ding, Xudong Jiang, Bing Shuai, Ai Qun Liu, and Gang Wang. Context contrasted feature and gated multiscale aggregation for scene segmentation. In CVPR, pages 2393-2402, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[11] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.232, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[13] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In AAAI, pages 10713-10720, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[15] Ekrem Emre Yurdakul and Yucel Yemez. Semantic segmentation of rgbd videos with recurrent fully convolutional neural networks. In ICCV workshops, pages 367-374, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[16] Jun Fu, Jing Liu, Hajie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.268, + 0.905, + 0.321 + ], + "angle": 0, + "content": "[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, pages 641–649, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.325, + 0.905, + 0.379 + ], + "angle": 0, + "content": "[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[19] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.904, + 0.467 + ], + "angle": 0, + "content": "[20] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. HiPPO: Recurrent memory with optimal polynomial projections. In NeurIPS, pages 1474–1487, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.509 + ], + "angle": 0, + "content": "[21] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.513, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[22] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. In NeurIPS, pages 572-585, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.571, + 0.905, + 0.611 + ], + "angle": 0, + "content": "[23] Shunan Guo, Zhuochen Jin, David Gotz, Fan Du, Hongyuan Zha, and Nan Cao. Visual progression analysis of event sequence data. IEEE TVCG, 25(1):417-426, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.614, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[24] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.644, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[25] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In CVPR, pages 8818-8827, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[26] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In CVPR, pages 8866-8875, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.745, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[27] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[28] Abhijit Kundu, Vibhav Vineet, and Vladlen Koltun. Feature space optimization for semantic video segmentation. In CVPR, pages 3168-3175, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[29] Jiangwei Lao, Weixiang Hong, Xin Guo, Yingying Zhang, Jian Wang, Jingdong Chen, and Wei Chu. Simultaneously short-and long-term temporal modeling for semi-supervised video semantic segmentation. In CVPR, pages 14763-14772, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24219" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Jiangyun Li, Yikai Zhao, Jun Fu, Jiajia Wu, and Jing Liu. Attention-guided network for semantic video segmentation. IEEE Access, 7:140680-140689, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[31] Jiangtong Li, Wentao Wang, Junjie Chen, Li Niu, Jianlou Si, Chen Qian, and Liqing Zhang. Video semantic segmentation via sparse temporal transformer. In ACM MM, pages 59-68, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[32] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. In ECCV, pages 237-255, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[33] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video K-Net: A simple, strong, and unified baseline for video segmentation. In CVPR, pages 18847-18857, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.482, + 0.379 + ], + "angle": 0, + "content": "[34] Xiangtai Li, Haobo Yuan, Wenwei Zhang, Guangliang Cheng, Jiangmiao Pang, and Chen Change Loy. Tube-Link: A flexible cross tube framework for universal video segmentation. In ICCV, pages 13923-13933, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.381, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[35] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In CVPR, pages 5997-6005, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.424, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[36] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60-88, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.51, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[37] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In ECCV, pages 352-368, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[38] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.612, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.67, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[41] Jun Ma, Feifei Li, and Bo Wang. U-Mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[42] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. VSPW: A large-scale dataset for video scene parsing in the wild. In CVPR, pages 4133-4143, 2021. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[43] Seyed Shahabeddin Nabavi, Mrigank Rochan, and Yang Wang. Future semantic segmentation with convolutional LSTM. CoRR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In CVPR, pages 6819-6828, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[45] Pedro O O Pinheiro, Ronan Collobert, and Piotr Dólár. Learning to segment object candidates. In NeurIPS, pages 1441-1447, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[46] Andreas Pfeuffer, Karina Schulz, and Klaus Dietmayer. Semantic segmentation of video sequences with convolutional LSTMs. In IEEE Intelligent Vehicles Symposium, pages 1441-1447, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[47] Rui Qian, Weiyao Lin, John See, and Dian Li. Controllable augmentations for video representation learning. Visual Intelligence, 2(1):1, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.236, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.305, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[50] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In CVPR, pages 3889-3898, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.405, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[51] Evan Shelhamer, Kate Rakelly, Judy Hoffman, and Trevor Darrell. Clockwork convnets for video semantic segmentation. In ECCV, pages 852-868, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, pages 802-810, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[53] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPR Workshop, pages 587-597, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.575, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[54] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.617, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[55] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[56] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[57] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In CVPR, pages 3126-3137, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[58] Guolei Sun, Yun Liu, Hao Tang, Ajad Chhatkuli, Le Zhang, and Luc Van Gool. Mining relations among cross-frame affinities for video semantic segmentation. In ECCV, pages 522-539, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[59] Guolei Sun, Yun Liu, Henghui Ding, Min Wu, and Luc Van Gool. Learning local and global temporal contexts for video semantic segmentation. IEEE TPAMI, 46(10):6919-6934, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[60] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. VMRNN: Integrating vision mamba and" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24220" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "LSTM for efficient and accurate spatiotemporal forecasting. In CVPR Workshop, pages 5663-5673, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[61] Vassilios Tsakanikas and Tasos Dagiuklas. Video surveillance systems-current status and future trends. Computers & Electrical Engineering, 70:736-753, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[62] Sepehr Valipour, Mennatullah Siam, Martin Jagersand, and Nilanjan Ray. Recurrent fully convolutional networks for video segmentation. In IEEE Winter Conference on Applications of Computer Vision, pages 29-36, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[63] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[64] Bowen Wang, Liangzhi Li, Yuta Nakashima, Ryo Kawasaki, Hajime Nagahara, and Yasushi Yagi. Noisy-LSTM: Improving temporal awareness for video semantic segmentation. IEEE Access, 9:46810-46820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[65] Hao Wang, Weining Wang, and Jing Liu. Temporal memory attention for video semantic segmentation. In ICIP, pages 2254-2258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[66] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[67] Yuetian Weng, Mingfei Han, Haoyu He, Mingjie Li, Lina Yao, Xiaojun Chang, and Bohan Zhuang. Mask propagation for efficient video semantic segmentation. In NeurIPS, pages 7170-7183, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[70] Chenliang Xu, Caiming Xiong, and Jason J Corso. Streaming hierarchical video segmentation. In ECCV, pages 626-639, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[71] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, pages 173-190, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[74] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[75] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In CVPR, pages 2349-2358, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.873 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24221" + } + ] +] \ No newline at end of file diff --git a/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_origin.pdf b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76e6cc39d0524d2c6c38435bc5050be6820a52a9 --- /dev/null +++ b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/205ea7bc-4216-405a-b9ee-ef9433534ad4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf70c19be8c8cf64afa2947c1605b9d050affa31745dd84dbe0087a757892c20 +size 1109833 diff --git a/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/full.md b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb0fed2345070108fdfd7e8cd4d3e22d5a165275 --- /dev/null +++ b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/full.md @@ -0,0 +1,319 @@ +# Exploiting Temporal State Space Sharing for Video Semantic Segmentation + +Syed Ariff Syed Hesham $^{1,2}$ , Yun Liu $^{3*}$ , Guolei Sun $^{4*}$ , Henghui Ding $^{5}$ , Jing Yang $^{6}$ , Ender Konukoglu $^{4}$ , Xue Geng $^{2}$ , Xudong Jiang $^{1}$ + +$^{1}$ Nanyang Technological University $^{2}$ Institute for Infocomm Research, A*STAR $^{3}$ VCIP, CS, Nankai University $^{4}$ CVL, ETH Zurich $^{5}$ Fudan University $^{6}$ Guizhou University + +# Abstract + +Video semantic segmentation (VSS) plays a vital role in understanding the temporal evolution of scenes. Traditional methods often segment videos frame-by-frame or in a short temporal window, leading to limited temporal context, redundant computations, and heavy memory requirements. To this end, we introduce a Temporal Video State Space Sharing (TV3S) architecture to leverage Mamba state space models for temporal feature sharing. Our model features a selective gating mechanism that efficiently propagates relevant information across video frames, eliminating the need for a memory-heavy feature pool. By processing spatial patches independently and incorporating shifted operation, TV3S supports highly parallel computation in both training and inference stages, which reduces the delay in sequential state space processing and improves the scalability for long video sequences. Moreover, TV3S incorporates information from prior frames during inference, achieving long-range temporal coherence and superior adaptability to extended sequences. Evaluations on the VSPW and Cityscapes datasets reveal that our approach outperforms current state-of-the-art methods, establishing a new standard for VSS with consistent results across long video sequences. By achieving a good balance between accuracy and efficiency, TV3S shows a significant advancement in spatiotemporal modeling, paving the way for efficient video analysis. The code is publicly available at https://github.com/Ashesham/TV3S.git. + +# 1. Introduction + +Semantic Segmentation has achieved substantial progress through the introduction of convolutional neural networks [5, 40, 48, 49, 72] and, more recently with vision transformers [4, 14, 39] due to their ability to capture spatial patterns and contextual information. However, these techniques are mainly designed for static images and do not leverage the + +![](images/6f112e19c8150db5ce5ff5d3ceb033e872524d5425a453123f1fd8c04ad3b187.jpg) +Figure 1. Comparison of the proposed TV3S with baseline models for VSS. By enhancing temporal information, our TV3S demonstrates superior performance over the baselines. + +temporal dynamics present in videos [2, 11, 12, 23, 57, 67]. + +Recently, research attention has increasingly shifted toward video semantic segmentation (VSS) [36, 53, 54, 61], due to its potential to benefit a wide range of practical applications. Unlike static image segmentation, VSS requires models to track motion, adapt to changes in appearance, and handle interactions among objects across consecutive frames. These challenges necessitate sophisticated techniques to accurately interpret dynamic environments where objects may occlude one another, shift positions, or exhibit varying states of motion and stillness [7, 47]. + +The contemporary approaches of VSS fall into two main categories. The first leverages optical flow, modeling pixel movement between consecutive frames to align features and support object tracking, thereby enhancing temporal coherence. These methods effectively propagate temporal information, making them useful for fine-grained temporal alignment [8, 13, 17, 50, 67]. However, they have significant drawbacks, including high computational costs due to the complexity of estimating accurate flow fields, especially in + +scenes with occlusions or sudden changes. Additionally, optical flow relies on precise motion estimation, where inaccuracies in dynamic scenes can propagate and lead to reduced accuracy in video segmentation. + +The second category involves feature aggregation, where information from multiple frames are combined to improve segmentation accuracy [28, 70]. The researchers relied on recurrent neural networks (RNNs) particularly Convolutional Long Short-Term Memory (ConvLSTM) [52] to introduce temporal structures to the models [15, 43, 46, 62, 64]. Recently, VM-RNN [60] has advanced this area by applying an efficient LSTM model to work with Vision Mamba [74] to capture spatiotemporal dynamics. While these methods can capture both spatial and temporal information, they face challenges when scaling to handle long video sequences, largely due to the high computational and memory demands of recurrent networks. Furthermore, the sensitivity of LSTMs to sequence length often leads to training instability, requiring careful tuning and optimization [46, 52, 64]. These limitations underscore the need for more scalable architectures. + +Due to such challenges in past feature aggregation methods, the focus in this domain has shifted towards the usage of transformer-based models [25, 29-31, 57-59, 65] [63]. However, these approaches still face challenges, including high memory costs and limited scalability, particularly with long video sequences or high-resolution frames. Efficient information sharing across numerous frames continues to be a challenge, making real-time applications difficult. + +To address the challenges of integrating spatial and temporal information in VSS, we present Temporal State Space Sharing (TV3S), designed to overcome the limitations of existing models. Unlike traditional optical flow methods, which struggle with inaccuracies in dynamic scenes, and RNN- and Transformer-based architectures, which are constrained by expensive feature pooling and high computational costs, TV3S takes a more efficient approach by processing spatial patches in parallel while dynamically sharing temporal information across video frames. Specifically, a spatially encoded input frame is split into discrete patches, which are processed independently through a series of TV3S blocks, each containing Temporal State Space (TSS) modules with a selective gating mechanism that effectively integrates and propagates spatiotemporal information across frames, ensuring minimal computational overhead. To further enhance motion handling at the edges of patches, we introduce a shifted window-based approach that works with un-shifted and shifted encoded features, enabling the model to capture movements near the boundaries while still maintaining efficient temporal information sharing. These components work jointly to integrate both local and long-range temporal features, significantly improving VSS performance + +while avoiding excessive computational costs and improving overall efficiency as evident in Fig. 1. + +In summary, the contributions of our paper include: + +- We present the Temporal Video State Space Sharing (TV3S) architecture, a novel framework that shares and propagates temporal information across video frames efficiently and effectively. +- We process spatial patches independently with a selective gating mechanism efficiently, thus enabling parallel computation during both training and inference and supporting scalability for long, high-resolution video sequences. +- We design a shifted window-based approach within the TV3S block, enhancing temporal state space sharing and capturing long-range spatial context effectively. + +Through extensive experiments on the VSPW [42] and Cityscapes [9] datasets, we demonstrate that our approach surpasses existing state-of-the-art methods, establishing new benchmarks for efficiency and accuracy in VSS. + +# 2. Related Work + +Semantic segmentation began with natural images, with the introduction of fully convolutional networks (FCNs) [40] pioneering an end-to-end pixel-wise classification framework. Building on this foundation, subsequent works enhanced segmentation accuracy by adopting atrous convolutional layers [5, 6], employing pyramid architectures [5, 6, 72], leveraging encoder-decoder architectures [10, 45, 48], and incorporating attention mechanisms [16, 24, 66]. More recently, transformer-based architectures like SegFormer [69], Segmenter [56], SETR [73], and Mask2Former [7], have further advanced the field by learning global dependencies. While these developments have significantly improved image semantic segmentation, transitioning them to VSS introduces new challenges related to temporal coherence and efficient exploitation of temporal information. + +# 2.1. VSS Based on Recurrent Neural Networks + +Early works in this field explored recurrent neural networks (RNNs) for temporal modeling. Valipour et al. [62] incorporated a recurrent unit between the encoder and decoder, significantly improving video segmentation performance. Evaluations by [15] on different RNN structures such as ConvRNN [55], ConvGRU [3], and ConvLSTM [52] on the KITTI dataset [18] demonstrated the superiority of ConvLSTM in handling video sequences. Further explorations have combined ConvGRU with optical flow to represent pixel displacements and maintain temporal continuity [44], while bidirectional ConvLSTM has been applied to merge temporally adjacent features, enhancing stability across frames [43]. Despite these advancements, the high computational and memory cost of RNNs for processing longer video sequences poses notable challenges, highlighting the need for more efficient models capable of handling + +![](images/03216e2b2bafeeaa9d1f4325fab384de104fa39baa4fcb2d2b20743361ff4acd.jpg) +Figure 2. Overview of the proposed TV3S architecture, illustrating the encoder-decoder framework that employs state space models and our TSS module based on Mamba [19] for independent spatial and temporal processing. + +long-range temporal dependencies [46, 52]. + +# 2.2. VSS Based on CNNs and Transformers + +To address RNN limitations, CNN-based methods were developed for temporal modeling. DFF [75] and Accel [26] utilized optical flow to propagate features between frames, reducing redundancy. ClockNet [51] and LLVS [35] introduced adaptive feature reuse to exploit semantic similarity across frames for efficiency and temporal coherence. + +Recognizing the limitations of optical flow and CNNs for long-range temporal modeling (see Sec. 1), recent advancements [1, 31, 33, 34] have seen a shift towards using transformers [63] in VSS. Transformers, with their self-attention mechanism, can capture global dependencies across frames, making them suited for temporal feature aggregation. Among notable approaches, MPVSS [67] proposes a memory-augmented transformer framework to capture multi-frame dependencies, enabling efficient temporal aggregation across longer video sequences. Similarly, CFFM [57] and MRCFA [58] employ multi-resolution cross-frame attention to handle temporal variations by disentangling static and dynamic contexts within video frames, allowing for a refined segmentation process that distinguishes between stationary and moving elements. + +Despite advancements, transformer-based methods face high computational costs due to the quadratic complexity of self-attention, especially with high-resolution frames or long videos [63]. Additionally, they are usually designed to learn temporal information in a short video sequence due to the high complexity. This gap highlights the need for more + +efficient and holistic models that can simultaneously manage computational costs while effectively modeling long-range temporal information in VSS. + +# 2.3. State Space Models + +State space models (SSMs) present a promising alternative for temporal modeling, addressing the shortcomings of RNNs and transformers. Unlike RNNs, which suffer from scaling challenges with sequence length, SSMs like S4 [21, 22] exhibit with linear complexity by imposing diagonal structures on state matrices, making them more efficient for long data sequences. Enhanced through HiPPO [20] initialization, SSMs handle extensive dependencies while requiring less memory, making them well-suited for tasks requiring to store long temporal context. Recently, Mamba [19] introduced a selective-scan mechanism to process temporal data efficiently. This advancement has spurred adaptations in vision-specific models, such as Vision Mamba [74] and VMamba [38], which incorporate Mamba blocks in a hierarchical structure to overcome directional sensitivities and maintain scalability across high-resolution inputs. Additionally, Vim [74] refined the scanning techniques to prevent overfitting. Furthermore, U-Mamba [41] has explored a hybrid network architecture that combines SSMs with convolutional layers. VideoMamba [32] represents one of the early frameworks to leverage SSM-based modules for video understanding, focusing on video clip classification. However, it is limited by its offline processing, which increases memory and computational demands, restricts dense semantic segmentation, and hinders its ability + +![](images/b0e0a5e3e62b6931fef56838a88dd5b9e22fa32c6c0208648cbdd91c4e930599.jpg) +Figure 3. Internal structure of the TV3S block, illustrating the flow of internal operations, and the propagation of hidden states for efficient spatiotemporal integration. + +to capture the global context of sequences. + +However, despite these advancements, current vision-specific models largely rely on hierarchical designs fail to fully exploit varying temporal resolutions and context lengths, highlighting a critical need for improved mechanisms for temporal state sharing. To this end, our proposed TV3S architecture integrates parameter-efficient SSMs to enhance effective temporal state sharing, thus proving advantageous for VSS. This integration not only underscores the computational efficiency and scalability of SSMs in VSS but also emphasizes their suitability for tasks requiring long-range temporal coherence, representing a significant novelty in the field. + +# 3. Methodology + +Fig. 2 illustrates the proposed TV3S architecture, which effectively captures and integrates temporal dependencies within video sequences using a state space modeling framework. It aims to leverage both temporal and spatial information to achieve accurate and consistent segmentation results. + +# 3.1. Overall Architecture + +The architecture processes a sequence of video frames denoted as $\{I_{t - l},\dots,I_{t - k},\dots,I_{t - 1},I_t\}$ , where $I_{t}$ represents the frame at time $t$ , and $\{l,\dots,k,\dots,1\}$ are temporal offsets capturing frames from the past relative to $t$ . These frames are fed into an image encoder such as MiT [69] and + +Swin [39]. The encoder produces a corresponding set of encoded feature $\{E_{t - l},\dots ,E_{t - k},\dots ,E_{t - 1},E_t\}$ , capturing rich spatial context. + +The TV3S architecture processes these feature maps sequentially through a series of $N$ TV3S blocks (with a default value of $N = 4$ ), each containing two TSS modules that independently process spatial patches of the feature map, facilitating parallelization. Layer normalization is applied before each TSS module, and residual connections are included after each module to stabilize training. Each TSS module consists of operations including a linear projection, a 1-dimension depthwise convolution (DWConv1), a SiLU activation feeding into the state space model (S6), and finally ending with another linear projection to project back to the input dimension, as detailed in Fig. 3. + +# 3.2. Temporal Video State Space Sharing Block + +The TV3S block is the core unit of the TV3S architecture, which enables seamless integration of temporal information across frames. Each block consists of two TSS modules that employ a state space model to capture and propagate temporal dependencies within frames. The encoded feature map $\pmb{E}_{t}$ is partitioned into non-overlapping patches of size $w \times w$ . Each patch $P_{t}^{i,j}$ is then flattened, transforming the spatial dimensions into a linear sequence suitable for parallel processing, with $i$ and $j$ representing the patch indices. Mathematically, $P_{t}^{i,j}$ can be expressed as + +$$ +\boldsymbol {P} _ {t} ^ {i, j} = \boldsymbol {E} _ {t} [ i: i + w, j: j + w,: ] \in \mathbb {R} ^ {C _ {E} \times w \times w}, \tag {1} +$$ + +$$ +\hat {\boldsymbol {P}} _ {t} ^ {i, j} = \operatorname {F l a t t e n} \left(\boldsymbol {P} _ {t} ^ {i, j}\right) \in \mathbb {R} ^ {C _ {E} \times (w \times w)}, +$$ + +where $C_E$ denotes the number of channels in $E_t$ after encoding $I_t$ . In contrast to traditional Mamba variants [38, 41, 74], which process sequences sequentially, our approach parallelizes the handling of each window or patch, offering two main advantages. First, the encoder effectively encodes complex spatial information across channels, eliminating for further processing of spatial relationships. Second, minimal changes between consecutive frames allow window-based processing to efficiently capture temporal dynamics, enabling a highly parallelized and computationally efficient model. + +# State space model for temporal aggregation. + +Each TSS module within a TV3S block utilizes a state space model to capture temporal dependencies between frames. Given encoded feature map $\pmb{E}_t$ , the hidden state $\pmb{H}_t$ is updated using the current flattened patch $\hat{\pmb{P}}_t^{i,j}$ and the previous hidden state $\pmb{H}_{t-1}$ : + +$$ +\boldsymbol {H} _ {t} ^ {i, j} = f _ {A} (\Delta , \boldsymbol {A} _ {s}) \boldsymbol {H} _ {t - 1} ^ {i, j} + f _ {B} (\Delta , \boldsymbol {A} _ {s}, \boldsymbol {B} _ {s}) \hat {\boldsymbol {P}} _ {t} ^ {i, j}, \tag {2} +$$ + +$$ +\boldsymbol {F} _ {t} ^ {i, j} = \boldsymbol {C} _ {s} \boldsymbol {H} _ {t} ^ {i, j}. +$$ + +Here, $\pmb{A}_s \in \mathbb{R}^{N_s \times N_s}$ , $\pmb{B}_s \in \mathbb{R}^{N_s \times C_E}$ , and $\pmb{C}_s \in \mathbb{R}^{C_E \times N_s}$ are learnable state space matrices. As seen in Eq. (2), + +the matrices $\mathbf{A}$ and $\mathbf{B}$ are discretized using the time scale parameter $\Delta_{s}$ with fixed discretization functions $f_{A}(.)$ & $f_{B}(.)$ , referred as discretization rules, as described in [19] due to their many advantages. Here, $N_{s}$ denotes the hidden states dimensionality of the state space model and the aggregated feature output is denoted as $F_{t}^{i,j}$ . This mechanism enables each patch to independently capture and propagate its temporal dependencies across two dimensions $(i,j)$ . Consequently, temporal data storage consists of a total of $\frac{W}{w} \times \frac{H}{w}$ hidden states, promoting effective temporal consistency and high parallelization. + +Shifting and edge handling. To enhance spatial context sharing, the reshaped feature maps undergo a shifting operation inspired by Swin Transformer [39]. This shift is parameterized as $s \times s$ , where $s < w$ (typically $s = w / 2$ ). The shifting rearranges the blocks so that edge blocks receive information from adjacent regions, thereby enriching spatial-temporal representation. The shifted feature maps are then re-partitioned into patches, with those on the right and bottom edges further subdivided into smaller sub-blocks of dimensions $2 \times w \times \frac{w}{2}$ , $2 \times \frac{w}{2} \times w$ , and $4 \times \frac{w}{2} \times \frac{w}{2}$ , as visualized in Fig. 2. These re-partitioned patches are processed through a second TSS module in the same TV3S block, integrating the newly acquired spatial context with temporal information. + +In the proposed architecture, each TV3S block consists of a pair of TSS modules, with one module processing unshifted blocks and the other handling shifted blocks. By stacking $N$ such TV3S blocks, the architecture forms a deep hierarchical structure termed TV3S architecture, which is capable of capturing intricate temporal dependencies across frames. The final aggregated feature representation, $\pmb{F}^{N}$ , is passed through a linear projection layer to map it to $C$ classes, followed by an interpolation to produce the final segmentation output. + +# 3.3. Training Strategy + +The training strategy for the TV3S architecture focuses on optimizing both temporal feature aggregation and robust spatial feature extraction to enable efficient learning of short- and long-term dependencies within video sequences, facilitating high VSS performance. + +During training, sequential input frames are taken at the intervals $\{I_{t-9}, I_{t-6}, I_{t-3}, I_t\}$ the same as [42]. Each of the frames in set $\{I_{t-k}\}$ with $k = \{9, 6, 3, 0\}$ , is passed through the encoder to generate encoded features $E_{t-k}$ encapsulating rich spatial information. An intermediate feature $\hat{O}_{t-k}$ is extracted straight from $E_{t-k}$ with the use of linear projection that aligns the channel dimensions of $E_{t-k}$ with the number of classes $C$ in the dataset. Concurrently, the encoded features $E_{t-k}$ are fed into the TV3S blocks to integrate temporal information, producing aggregated features $F_{t-k}^N$ . + +Algorithm 1 Inference Procedure for TV3S Architecture +1: Input: Frame sequence $\{I_t\}$ +2: Output: Segmentation map $O_t$ +3: Initialize: Hidden states $\pmb{H}$ for all TV3S blocks +4: for each frame $\pmb{I}_t$ in the sequence do +5: $\pmb{E}_t \gets \text{Encoder}(\pmb{I}_t) \quad //$ Extract features +6: for each patch $(i,j)$ in $\pmb{E}_t$ do +7: $\hat{\pmb{P}}^{i,j} \gets \text{Flatten}(\pmb{E}_t[i:i+w,j:j+w,:])$ +8: for each TV3S block $n$ from 1 to $N$ do +9: //Update Features and Hidden State. +10: $\pmb{F}_t^{n,i,j}, \pmb{H}_t^{n,i,j} \gets \text{TV3S}^n(\hat{\pmb{P}}^{i,j}, \pmb{H}^{n,i,j})$ +11: end for +12: end for +13: $O_t \gets \text{Interpolate}(\text{Linear}(\pmb{F}_t^N)) \quad //$ Segmentation +14: $\pmb{H} \gets \{\pmb{H}_t\} \quad //$ Store for next frame +15: end for + +With both output predictions extracted, a weighted cross-entropy between these output predictions $O$ and the ground-truth segmentation masks $M$ is computed as training loss for all frames in the input sequence. The training loss is computed as: + +$$ +\mathcal {L} = \lambda \sum_ {k = \{9, 6, 3, 0 \}} \mathcal {L} _ {\mathrm {C E}} \left(\hat {\boldsymbol {O}} _ {t - k}, \boldsymbol {M} _ {t - k}\right) + \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {O} _ {t}, \boldsymbol {M} _ {t}\right). \tag {3} +$$ + +Loss formulation presented in Eq. (3) employs a dual-loss strategy to ensure accurate segmentation of the final frame $O_{t}$ while preserving spatial relationships in the intermediate features $\hat{O}_{t - k}$ from the encoder. By applying cross-entropy losses $\mathcal{L}_{\mathrm{CE}}$ , with intermediate losses weighted at $\lambda = 0.5$ , the model effectively balances spatial feature learning and temporal coherence, enhancing its capability to extract temporally consistent spatial information for improved segmentation across consecutive frames. + +# 3.4. Inference Procedure + +During inference, the TV3S architecture processes each frame sequentially while leveraging stored hidden states to maintain temporal coherence across the video sequence, as detailed in Algorithm 1. For each frame $\pmb{I}_t$ , the encoder first extracts the encoded features $\pmb{E}_t$ , which are then partitioned into non-overlapping spatial patches indexed by $(i,j)$ and flattened into $\pmb{P}_{i,j}$ . Each flattened patch is processed through the TV3S blocks, updating the current hidden states $\pmb{H}_t$ using the previous states $\pmb{H}_{t-1}$ . After processing all patches, the aggregated feature map $\pmb{F}_t^N$ is reconstructed from the updated hidden states. This feature map is then passed through a linear projection layer and interpolated to match the original image resolution, producing the final segmentation map $\pmb{O}_t$ . The updated hidden states $\pmb{H}_t$ are then stored back in $\pmb{H}$ for use with subsequent frames, ensuring continuous temporal integration and consistent segmentation results. + +
MethodsBackbonesmIoU↑mVC8↑mVC16↑GFLOPs↓Params(M)↓FPS↑
Segformer† [69]MiT-B136.584.779.926.613.858.7
CFFM [57]MiT-B138.588.684.1-15.529.8
MRCFA [58]MiT-B138.988.884.4-16.240.1
TV3S (Ours)MiT-B140.090.787.036.917.324.7
Segformer [69]MiT-B243.986.081.2100.824.816.2
CFFM [57]MiT-B244.989.885.8143.226.510.1
MRCFA [58]MiT-B245.390.386.2127.927.310.7
TV3S (Ours)MiT-B246.391.588.3553.928.321.9
Mask2Former† [7]R5038.581.376.4110.644.019.4
MPVSS [67]R5037.584.177.238.984.133.9
Mask2Former† [7]R10139.382.577.6141.363.016.9
MPVSS [67]R10138.884.879.645.1103.132.3
DeepLabv3+† [6]R10134.783.278.2379.062.79.2
UperNet† [68]R10136.582.676.1403.683.216.0
PSPNet† [72]R10136.584.279.6401.870.513.8
OCRNet† [71]R10136.784.079.0361.758.114.3
TCB† [42]R10137.887.984.01692--
ETC† [37]OCRNet37.584.179.1361.7--
Segformer [69]MiT-B548.987.883.7185.082.19.4
CFFM [57]MiT-B549.390.887.1413.585.54.5
MRCFA [58]MiT-B549.990.987.4373.084.55.0
TV3S (Ours)MiT-B549.891.788.7137.085.614.0
Mask2Former† [7]Swin-T41.284.580.0114.447.417.1
MPVSS [67]Swin-T39.985.980.439.7114.032.8
TV3S (Ours)Swin-T44.988.083.557.331.722.9
Mask2Former† [7]Swin-S42.184.779.3152.268.914.5
MPVSS [67]Swin-S40.486.080.747.3108.030.6
TV3S (Ours)Swin-S50.689.685.894.153.119.5
+ +Table 1. Quantitative comparison of our model with existing methods on the VSPW dataset [42]. Our model achieves a strong balance between accuracy, model complexity, and operational speed. FPS and FLOPs are calculated with an input resolution of $480 \times 853$ . (†Frame-by-Frame processing) + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation details. The implementation of our approach is based on MMSegmentation codebase and all the experiments including training and inference were conducted with 2 A100 NVIDIA GPUs. The main experiments are done with the backbone same as the SegFormer (Variants with MiT and Swin) which are pre-trained with ImageNet. While the model is aimed to work with any number of frame sequences, during training, the model is trained with just three reference frames, $\{k_1,k_2,k_3\} = \{-9, -6, -3\}$ . More information on training is found in Sec. 3.3. To improve receptive field when processing the features from the backbone, the window size $w$ and the shift $s$ are set to 20 and 10 with number of TV3S blocks $N$ set to 4 (more information on these at Sec. 4.3) and the decoder made with mamba following its default parameters following [19] with the input embedding dimension of 256 matching the SegFormer implementation. The images from VSPW dataset [42] are cropped down to $480 \times 480$ and + +are augmented using various augmentation methods during training, including cropping, resizing, flipping and addition of photometric distortion during training. Optimization of the model is done with the use AdamW optimizer and a "poly" learning rate schedule initializing the learning rate at $6e - 5$ . Testing is performed using the context of full video with the frame receiving the context of all the past frames within the video through the hidden states $\pmb{H}$ due to its high efficiency & effectiveness following Sec. 3.4. Note that for all cases there was no use of post-processing on the obtained output like in [27]. + +Datasets. The experiments were mainly conducted with the use of VSPW dataset [42] which stands as one of the largest VSS benchmark. The dataset consists of training, validation and test subsets containing 2,806 clips (198,244 frames), 343 clips (24,502 frames) and 387 clips (28,887 frames) accordingly. The dataset consisting of a rich 124 categories with a dense annotation of frame rate of 15fps contrasts itself from the previously available datasets, which had very sparse annotation with just one frame being annotated every 10s of frames. Furthermore, the dataset cover + +
MethodsBackbonesmIoU↑GFLOPs↓Params(M)↓FPS↑
ETC† [37]R1871.1434.1--
SegFormer† [69]MiT-B071.9-3.758.5
CFFM [57]MiT-B074.080.74.615.8
MRCFA [58]MiT-B072.877.54.216.6
SegFormer† [69]MiT-B174.1-13.846.8
CFFM [57]MiT-B175.1158.715.411.7
MRCFA [58]MiT-B175.1145.014.913.0
TV3S (ours)MiT-B175.683.617.325.1
+ +Table 2. Quantitative comparison of our method with efficient alternative approaches on the Cityscapes dataset [9], using a resolution of $512 \times 1024$ . (†Frame-by-Frame processing) + +
ModelsEvaluation (mIoU)
AddConcatDirect
1 TSS (No Shift)37.637.838.0
TV3S (No Shift)38.138.438.9
TV3S (Shift)38.539.339.5
+ +ing various different scenarios of both outdoor and indoor scenes makes it suitable for training models to well verify the adaptability of the performance standing as the best benchmark in the field of VSS. While most experiments and training were done with the VSPW dataset [42], the proposed method was also evaluated on the Cityscapes dataset [9] which annotates one frame out of every 30 frames for benchmarking the results. + +Evaluation metrics. We use mean IoU (mIoU) and mean video consistency (mVC) as key metrics. The metric mVC evaluates the smoothness of predicted segmentation maps over time, assessing performance in the temporal domain. More formally, given a video clip $\{I_c\}_{c=0}^{C_v}$ with ground truth masks $\{M_c\}_{c=1}^{C_v}$ and predicted outputs $\{O_c\}_{c=1}^{C_v}$ , $VC_n$ is computed as follows: + +$$ +\mathrm {V C} _ {n} = \frac {1}{C _ {v} - n + 1} \sum_ {i = 1} ^ {C _ {v} - n + 1} \frac {\left(\cap_ {i} ^ {i + n - 1} \boldsymbol {M} _ {i}\right) \cap \left(\cap_ {i} ^ {i + n - 1} \boldsymbol {O} _ {i}\right)}{\cap_ {i} ^ {i + n - 1} \boldsymbol {M} _ {i}}, \tag {4} +$$ + +where $C_{\mathrm{v}} \geq n$ . Once the $VC_{n}$ of all the videos are computed, their mean is computed to obtain the $\mathrm{mVC}_{\mathrm{n}}$ . Eq. (4) shows that $\mathrm{mVC}_{\mathrm{n}}$ finds the common areas of the predicted masks among frames which indicates the level of consistency of the prediction masks across time. More information on the metric can be found at [42]. + +# 4.2. Comparison with State-of-the-art Models + +Our model is compared against state-of-the-art models on the VSPW dataset [42], as shown in Tab. 1. The table is divided into three groups based on model size and backbone, offering insights at varying scales. In the first group + +Table 3. Evaluation on the implication of the shifted mechanism and the output method. + +
Window sizesmIoUmVC8mVC16
438.589.084.6
638.689.284.7
1239.389.385.0
1639.289.385.0
2040.090.787.0
2840.088.884.2
3639.989.184.9
+ +Table 4. Impact of the window size on model performance. + +
BackbonesTV3S BlocksmIoUmVC8mVC16
MiT-B1136.288.183.6
237.488.684.3
337.688.583.5
440.090.787.0
+ +Table 5. Performance metrics based on the number of TV3S blocks in the model. + +with small models (<30M parameters), our method outperforms the baselines, demonstrating efficiency even with limited model capacity. In the second group with larger models (>30M parameters), TV3S achieves near-state-of-the-art performance by effectively capturing rich contextual information. The third group focuses on the Swin Transformer backbone, where our model excels with over 8 mIoU ahead of the next best, highlighting its ability to preserve temporal correlations and ensure spatial accuracy. Overall, our method demonstrates superior visual consistency across all groups, further illustrated by Fig. 4 showcasing temporally consistent segmentation. + +We also present results on the Cityscapes dataset [9] in Tab. 2, using smaller model variants at an input resolution of $1024 \times 512$ . Due to the dataset's annotation structure, metrics such as $\mathrm{mVC}_{8}$ and $\mathrm{mVC}_{16}$ are not applicable, but our model still achieves top performance in mIoU, underscoring its strong generalizability across datasets. + +# 4.3. Ablation Study + +All ablation studies were conducted on the VSPW dataset [42] using the MiT-B1 backbone and following the same training and inference strategies as previously described. + +Impact of shifted representations and feature integration. We evaluated the effectiveness of different TSS/TV3S decoder (see Tab. 3), including one TSS layer, unshifted TV3S, and shifted TV3S. Despite the only difference between unshifted and shifted TV3S being the feature representation, the results clearly show that shifted representation significantly outperforms its unshifted counterparts, highlighting their effectiveness. We further tested adding feature maps through addition and concatenation before the final prediction, but both methods reduced performance, suggesting that reintroducing spatial information after temporal modeling disrupts temporal coherence. + +![](images/dd706f4a7dc08bf41e498e59bffa00dfcc4e95279c405830fd08b4339586d0e2.jpg) +Figure 4. Qualitative example of our TV3S architecture compared with the current baseline. This displays both improved performance in performing spatial predictions and utilizing the temporal information to produce temporally consistent segmentation results. + +
Number of FramesmIoUmVC8mVC16
137.984.579.1
239.286.981.8
439.588.984.2
839.790.485.9
1639.791.287.2
3239.791.587.8
+ +Table 6. Performance based on the amount of temporal information available during inference + +Impact of window size. TV3S establishes temporal relationships with window size controlling the spatial context during processing. Larger windows help summarize stable scenes but may overload the model in dynamic settings, while smaller windows enhance understanding in fast-changing environments but may miss correlations in slower scenes. As shown in Tab. 4, the optimal performance is achieved with a $20 \times 20$ window size, striking a balance between spatial context and model efficiency. + +Effect on the number of TV3S blocks. In our experiments, we assessed the impact of varying the number of TV3S blocks, as shown in Tab. 5. Increasing the number of blocks improved mIoU from 36.18 with one block to 40.00 with four blocks, along with enhanced temporal consistency, as shown by the $\mathrm{mVC}_{8}$ and $\mathrm{mVC}_{16}$ scores. This upward trend suggests that stacking TV3S blocks better aggregates temporal features, capturing complex relationships without introducing excessive redundancy. + +# 4.4. Temporal Context + +Importance of long temporal context. Given Mamba's RNN-like structure, we conducted an evaluation of its performance on the VSPW dataset [42] with varying temporal context sizes that ranged from 1 to 32 frames. In our approach, the videos that were shorter than 32 frames were + +excluded from the and at the same time the first 32 frames were omitted during inference to ensure a fair and unbiased evaluation of the model's capabilities. The findings presented in Tab. 6 indicate that there is a notable improvement in performance as the temporal context sizes increase. However, this enhancement reaches a plateau once the temporal information saturates. This saturation effect is likely attributed to the limited changes that can be captured within smaller temporal windows, which causes diminishing returns in performance improvements beyond a certain threshold. + +# 5. Conclusion + +The work proposes a TV3S architecture that addresses the challenges of VSS by capturing temporal dynamics. The proposed structure is designed to achieve good computational efficiency and scalability. By leveraging independent processing of spatial blocks through state space models enhanced with Mamba [19], our approach enables parallel computation during both training and inference. This design mitigates the time delay and high memory demand typically associated with sequential processing in traditional state space and recurrent models, making it highly suitable for long, high-resolution video sequences. Extensive experiments on the VSPW [42] and Cityscapes [9] datasets show that the TV3S architecture not only surpasses existing methods in segmentation accuracy but also significantly improves temporal prediction consistency. + +Future works. Our model enhances the encoder's focus on key temporal features, integrating spatial and temporal information for improved video frame segmentation. Its adaptability extends to tasks like object detection and action recognition, while its multi-modal data integration offers new research opportunities in audio-visual learning, emphasizing temporal synchronization. + +# Acknowledgment + +This work is supported by the Ministry of Education Singapore Tier 1 project funding grant No. MOE Tier 1 RG98/24 and by Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). + +# References + +[1] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. In BMVC, 2023. 3 +[2] Gowtham Atluri, Anuj Karpatne, and Vipin Kumar. Spatiotemporal data mining: A survey of problems and methods. ACM Computing Surveys, 51(4):1-41, 2018. 1 +[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 2 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4): 834-848, 2017. 1, 2 +[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 2, 6 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 2, 6 +[8] Jingchun Cheng, Yi-Hsuan Tsai, Shengjin Wang, and Ming-Hsuan Yang. SegFlow: Joint learning for video object segmentation and optical flow. In ICCV, pages 686-695, 2017. 1 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 2, 7, 8 +[10] Henghui Ding, Xudong Jiang, Bing Shuai, Ai Qun Liu, and Gang Wang. Context contrasted feature and gated multiscale aggregation for scene segmentation. In CVPR, pages 2393-2402, 2018. 2 +[11] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1 +[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1 + +[13] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In AAAI, pages 10713-10720, 2020. 1 +[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1 +[15] Ekrem Emre Yurdakul and Yucel Yemez. Semantic segmentation of rgbd videos with recurrent fully convolutional neural networks. In ICCV workshops, pages 367-374, 2017. 2 +[16] Jun Fu, Jing Liu, Hajie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 2 +[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, pages 641–649, 2023. 1 +[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2 +[19] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 5, 6, 8 +[20] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. HiPPO: Recurrent memory with optimal polynomial projections. In NeurIPS, pages 1474–1487, 2020. 3 +[21] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021. 3 +[22] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. In NeurIPS, pages 572-585, 2021. 3 +[23] Shunan Guo, Zhuochen Jin, David Gotz, Fan Du, Hongyuan Zha, and Nan Cao. Visual progression analysis of event sequence data. IEEE TVCG, 25(1):417-426, 2018. 1 +[24] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2 +[25] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In CVPR, pages 8818-8827, 2020. 2 +[26] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In CVPR, pages 8866-8875, 2019. 3 +[27] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6 +[28] Abhijit Kundu, Vibhav Vineet, and Vladlen Koltun. Feature space optimization for semantic video segmentation. In CVPR, pages 3168-3175, 2016. 2 +[29] Jiangwei Lao, Weixiang Hong, Xin Guo, Yingying Zhang, Jian Wang, Jingdong Chen, and Wei Chu. Simultaneously short-and long-term temporal modeling for semi-supervised video semantic segmentation. In CVPR, pages 14763-14772, 2023. 2 + +[30] Jiangyun Li, Yikai Zhao, Jun Fu, Jiajia Wu, and Jing Liu. Attention-guided network for semantic video segmentation. IEEE Access, 7:140680-140689, 2019. +[31] Jiangtong Li, Wentao Wang, Junjie Chen, Li Niu, Jianlou Si, Chen Qian, and Liqing Zhang. Video semantic segmentation via sparse temporal transformer. In ACM MM, pages 59-68, 2021. 2, 3 +[32] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. In ECCV, pages 237-255, 2025. 3 +[33] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video K-Net: A simple, strong, and unified baseline for video segmentation. In CVPR, pages 18847-18857, 2022. 3 +[34] Xiangtai Li, Haobo Yuan, Wenwei Zhang, Guangliang Cheng, Jiangmiao Pang, and Chen Change Loy. Tube-Link: A flexible cross tube framework for universal video segmentation. In ICCV, pages 13923-13933, 2023. 3 +[35] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In CVPR, pages 5997-6005, 2018. 3 +[36] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60-88, 2017. 1 +[37] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In ECCV, pages 352-368, 2020. 6, 7 +[38] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 3, 4 +[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 4, 5 +[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2 +[41] Jun Ma, Feifei Li, and Bo Wang. U-Mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3, 4 +[42] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. VSPW: A large-scale dataset for video scene parsing in the wild. In CVPR, pages 4133-4143, 2021. 2, 5, 6, 7, 8 +[43] Seyed Shahabeddin Nabavi, Mrigank Rochan, and Yang Wang. Future semantic segmentation with convolutional LSTM. CoRR, 2018. 2 +[44] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In CVPR, pages 6819-6828, 2018. 2 + +[45] Pedro O O Pinheiro, Ronan Collobert, and Piotr Dólár. Learning to segment object candidates. In NeurIPS, pages 1441-1447, 2015. 2 +[46] Andreas Pfeuffer, Karina Schulz, and Klaus Dietmayer. Semantic segmentation of video sequences with convolutional LSTMs. In IEEE Intelligent Vehicles Symposium, pages 1441-1447, 2019. 2, 3 +[47] Rui Qian, Weiyao Lin, John See, and Dian Li. Controllable augmentations for video representation learning. Visual Intelligence, 2(1):1, 2024. 1 +[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015. 1, 2 +[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 1 +[50] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In CVPR, pages 3889-3898, 2016. 1 +[51] Evan Shelhamer, Kate Rakelly, Judy Hoffman, and Trevor Darrell. Clockwork convnets for video semantic segmentation. In ECCV, pages 852-868, 2016. 3 +[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, pages 802-810, 2015. 2, 3 +[53] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPR Workshop, pages 587-597, 2018. 1 +[54] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1 +[55] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 2 +[56] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2 +[57] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In CVPR, pages 3126-3137, 2022. 1, 2, 3, 6, 7 +[58] Guolei Sun, Yun Liu, Hao Tang, Ajad Chhatkuli, Le Zhang, and Luc Van Gool. Mining relations among cross-frame affinities for video semantic segmentation. In ECCV, pages 522-539, 2022. 3, 6, 7 +[59] Guolei Sun, Yun Liu, Henghui Ding, Min Wu, and Luc Van Gool. Learning local and global temporal contexts for video semantic segmentation. IEEE TPAMI, 46(10):6919-6934, 2024. 2 +[60] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. VMRNN: Integrating vision mamba and + +LSTM for efficient and accurate spatiotemporal forecasting. In CVPR Workshop, pages 5663-5673, 2024. 2 +[61] Vassilios Tsakanikas and Tasos Dagiuklas. Video surveillance systems-current status and future trends. Computers & Electrical Engineering, 70:736-753, 2018. 1 +[62] Sepehr Valipour, Mennatullah Siam, Martin Jagersand, and Nilanjan Ray. Recurrent fully convolutional networks for video segmentation. In IEEE Winter Conference on Applications of Computer Vision, pages 29-36, 2017. 2 +[63] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 3 +[64] Bowen Wang, Liangzhi Li, Yuta Nakashima, Ryo Kawasaki, Hajime Nagahara, and Yasushi Yagi. Noisy-LSTM: Improving temporal awareness for video semantic segmentation. IEEE Access, 9:46810-46820, 2021. 2 +[65] Hao Wang, Weining Wang, and Jing Liu. Temporal memory attention for video semantic segmentation. In ICIP, pages 2254-2258, 2021. 2 +[66] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 2 +[67] Yuetian Weng, Mingfei Han, Haoyu He, Mingjie Li, Lina Yao, Xiaojun Chang, and Bohan Zhuang. Mask propagation for efficient video semantic segmentation. In NeurIPS, pages 7170-7183, 2024. 1, 3, 6 +[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018. 6 +[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 2, 4, 6, 7 +[70] Chenliang Xu, Caiming Xiong, and Jason J Corso. Streaming hierarchical video segmentation. In ECCV, pages 626-639, 2012. 2 +[71] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, pages 173-190, 2020. 6 +[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 1, 2, 6 +[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2 +[74] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3, 4 +[75] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In CVPR, pages 2349-2358, 2017. 3 \ No newline at end of file diff --git a/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/images.zip b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eff85194e80c4c01a16b9080ce3612ecb0f5c735 --- /dev/null +++ b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ddced8fbc3a298ad3c4d2a21f2567ddc8168877c4a96b7aa35623ebeee85c2b +size 600085 diff --git a/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/layout.json b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..146593353451cd40afa65f4f2be1c535a7d1f89c --- /dev/null +++ b/2025/Exploiting Temporal State Space Sharing for Video Semantic Segmentation/layout.json @@ -0,0 +1,9296 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "text", + "content": "Exploiting Temporal State Space Sharing for Video Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "spans": [ + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": "Syed Ariff Syed Hesham" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": ", Yun Liu" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": ", Guolei Sun" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": ", Henghui Ding" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": ", Jing Yang" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": ", Ender Konukoglu" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": ", Xue Geng" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "text", + "content": ", Xudong Jiang" + }, + { + "bbox": [ + 109, + 142, + 501, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "text", + "content": " Nanyang Technological University " + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "text", + "content": " Institute for Infocomm Research, A*STAR " + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "text", + "content": " VCIP, CS, Nankai University " + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "text", + "content": " CVL, ETH Zurich " + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "text", + "content": " Fudan University " + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 61, + 171, + 544, + 200 + ], + "type": "text", + "content": " Guizhou University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 252, + 297, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 297, + 576 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 297, + 576 + ], + "type": "text", + "content": "Video semantic segmentation (VSS) plays a vital role in understanding the temporal evolution of scenes. Traditional methods often segment videos frame-by-frame or in a short temporal window, leading to limited temporal context, redundant computations, and heavy memory requirements. To this end, we introduce a Temporal Video State Space Sharing (TV3S) architecture to leverage Mamba state space models for temporal feature sharing. Our model features a selective gating mechanism that efficiently propagates relevant information across video frames, eliminating the need for a memory-heavy feature pool. By processing spatial patches independently and incorporating shifted operation, TV3S supports highly parallel computation in both training and inference stages, which reduces the delay in sequential state space processing and improves the scalability for long video sequences. Moreover, TV3S incorporates information from prior frames during inference, achieving long-range temporal coherence and superior adaptability to extended sequences. Evaluations on the VSPW and Cityscapes datasets reveal that our approach outperforms current state-of-the-art methods, establishing a new standard for VSS with consistent results across long video sequences. By achieving a good balance between accuracy and efficiency, TV3S shows a significant advancement in spatiotemporal modeling, paving the way for efficient video analysis. The code is publicly available at https://github.com/Ashesham/TV3S.git." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 600, + 135, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 135, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 135, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "Semantic Segmentation has achieved substantial progress through the introduction of convolutional neural networks [5, 40, 48, 49, 72] and, more recently with vision transformers [4, 14, 39] due to their ability to capture spatial patterns and contextual information. However, these techniques are mainly designed for static images and do not leverage the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 227, + 553, + 430 + ], + "blocks": [ + { + "bbox": [ + 316, + 227, + 553, + 430 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 553, + 430 + ], + "type": "image", + "image_path": "6f112e19c8150db5ce5ff5d3ceb033e872524d5425a453123f1fd8c04ad3b187.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 431, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Comparison of the proposed TV3S with baseline models for VSS. By enhancing temporal information, our TV3S demonstrates superior performance over the baselines." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 553, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 485 + ], + "type": "text", + "content": "temporal dynamics present in videos [2, 11, 12, 23, 57, 67]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 486, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 555, + 605 + ], + "type": "text", + "content": "Recently, research attention has increasingly shifted toward video semantic segmentation (VSS) [36, 53, 54, 61], due to its potential to benefit a wide range of practical applications. Unlike static image segmentation, VSS requires models to track motion, adapt to changes in appearance, and handle interactions among objects across consecutive frames. These challenges necessitate sophisticated techniques to accurately interpret dynamic environments where objects may occlude one another, shift positions, or exhibit varying states of motion and stillness [7, 47]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "The contemporary approaches of VSS fall into two main categories. The first leverages optical flow, modeling pixel movement between consecutive frames to align features and support object tracking, thereby enhancing temporal coherence. These methods effectively propagate temporal information, making them useful for fine-grained temporal alignment [8, 13, 17, 50, 67]. However, they have significant drawbacks, including high computational costs due to the complexity of estimating accurate flow fields, especially in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 229, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 229, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 229, + 712 + ], + "type": "text", + "content": "*Corresponding authors: Yun Liu and Guolei Sun" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24211" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "scenes with occlusions or sudden changes. Additionally, optical flow relies on precise motion estimation, where inaccuracies in dynamic scenes can propagate and lead to reduced accuracy in video segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "type": "text", + "content": "The second category involves feature aggregation, where information from multiple frames are combined to improve segmentation accuracy [28, 70]. The researchers relied on recurrent neural networks (RNNs) particularly Convolutional Long Short-Term Memory (ConvLSTM) [52] to introduce temporal structures to the models [15, 43, 46, 62, 64]. Recently, VM-RNN [60] has advanced this area by applying an efficient LSTM model to work with Vision Mamba [74] to capture spatiotemporal dynamics. While these methods can capture both spatial and temporal information, they face challenges when scaling to handle long video sequences, largely due to the high computational and memory demands of recurrent networks. Furthermore, the sensitivity of LSTMs to sequence length often leads to training instability, requiring careful tuning and optimization [46, 52, 64]. These limitations underscore the need for more scalable architectures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 425 + ], + "type": "text", + "content": "Due to such challenges in past feature aggregation methods, the focus in this domain has shifted towards the usage of transformer-based models [25, 29-31, 57-59, 65] [63]. However, these approaches still face challenges, including high memory costs and limited scalability, particularly with long video sequences or high-resolution frames. Efficient information sharing across numerous frames continues to be a challenge, making real-time applications difficult." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 715 + ], + "type": "text", + "content": "To address the challenges of integrating spatial and temporal information in VSS, we present Temporal State Space Sharing (TV3S), designed to overcome the limitations of existing models. Unlike traditional optical flow methods, which struggle with inaccuracies in dynamic scenes, and RNN- and Transformer-based architectures, which are constrained by expensive feature pooling and high computational costs, TV3S takes a more efficient approach by processing spatial patches in parallel while dynamically sharing temporal information across video frames. Specifically, a spatially encoded input frame is split into discrete patches, which are processed independently through a series of TV3S blocks, each containing Temporal State Space (TSS) modules with a selective gating mechanism that effectively integrates and propagates spatiotemporal information across frames, ensuring minimal computational overhead. To further enhance motion handling at the edges of patches, we introduce a shifted window-based approach that works with un-shifted and shifted encoded features, enabling the model to capture movements near the boundaries while still maintaining efficient temporal information sharing. These components work jointly to integrate both local and long-range temporal features, significantly improving VSS performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "while avoiding excessive computational costs and improving overall efficiency as evident in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 96, + 534, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 96, + 534, + 108 + ], + "spans": [ + { + "bbox": [ + 326, + 96, + 534, + 108 + ], + "type": "text", + "content": "In summary, the contributions of our paper include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 156 + ], + "type": "text", + "content": "- We present the Temporal Video State Space Sharing (TV3S) architecture, a novel framework that shares and propagates temporal information across video frames efficiently and effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "type": "text", + "content": "- We process spatial patches independently with a selective gating mechanism efficiently, thus enabling parallel computation during both training and inference and supporting scalability for long, high-resolution video sequences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 204, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 239 + ], + "type": "text", + "content": "- We design a shifted window-based approach within the TV3S block, enhancing temporal state space sharing and capturing long-range spatial context effectively." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 240, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 288 + ], + "type": "text", + "content": "Through extensive experiments on the VSPW [42] and Cityscapes [9] datasets, we demonstrate that our approach surpasses existing state-of-the-art methods, establishing new benchmarks for efficiency and accuracy in VSS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 297, + 400, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 297, + 400, + 309 + ], + "spans": [ + { + "bbox": [ + 314, + 297, + 400, + 309 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 317, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 497 + ], + "type": "text", + "content": "Semantic segmentation began with natural images, with the introduction of fully convolutional networks (FCNs) [40] pioneering an end-to-end pixel-wise classification framework. Building on this foundation, subsequent works enhanced segmentation accuracy by adopting atrous convolutional layers [5, 6], employing pyramid architectures [5, 6, 72], leveraging encoder-decoder architectures [10, 45, 48], and incorporating attention mechanisms [16, 24, 66]. More recently, transformer-based architectures like SegFormer [69], Segmenter [56], SETR [73], and Mask2Former [7], have further advanced the field by learning global dependencies. While these developments have significantly improved image semantic segmentation, transitioning them to VSS introduces new challenges related to temporal coherence and efficient exploitation of temporal information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 503, + 535, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 535, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 535, + 515 + ], + "type": "text", + "content": "2.1. VSS Based on Recurrent Neural Networks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "Early works in this field explored recurrent neural networks (RNNs) for temporal modeling. Valipour et al. [62] incorporated a recurrent unit between the encoder and decoder, significantly improving video segmentation performance. Evaluations by [15] on different RNN structures such as ConvRNN [55], ConvGRU [3], and ConvLSTM [52] on the KITTI dataset [18] demonstrated the superiority of ConvLSTM in handling video sequences. Further explorations have combined ConvGRU with optical flow to represent pixel displacements and maintain temporal continuity [44], while bidirectional ConvLSTM has been applied to merge temporally adjacent features, enhancing stability across frames [43]. Despite these advancements, the high computational and memory cost of RNNs for processing longer video sequences poses notable challenges, highlighting the need for more efficient models capable of handling" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24212" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 553, + 316 + ], + "type": "image", + "image_path": "03216e2b2bafeeaa9d1f4325fab384de104fa39baa4fcb2d2b20743361ff4acd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed TV3S architecture, illustrating the encoder-decoder framework that employs state space models and our TSS module based on Mamba [19] for independent spatial and temporal processing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 348, + 233, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 233, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 233, + 361 + ], + "type": "text", + "content": "long-range temporal dependencies [46, 52]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 261, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 261, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 261, + 382 + ], + "type": "text", + "content": "2.2. VSS Based on CNNs and Transformers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 460 + ], + "type": "text", + "content": "To address RNN limitations, CNN-based methods were developed for temporal modeling. DFF [75] and Accel [26] utilized optical flow to propagate features between frames, reducing redundancy. ClockNet [51] and LLVS [35] introduced adaptive feature reuse to exploit semantic similarity across frames for efficiency and temporal coherence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 640 + ], + "type": "text", + "content": "Recognizing the limitations of optical flow and CNNs for long-range temporal modeling (see Sec. 1), recent advancements [1, 31, 33, 34] have seen a shift towards using transformers [63] in VSS. Transformers, with their self-attention mechanism, can capture global dependencies across frames, making them suited for temporal feature aggregation. Among notable approaches, MPVSS [67] proposes a memory-augmented transformer framework to capture multi-frame dependencies, enabling efficient temporal aggregation across longer video sequences. Similarly, CFFM [57] and MRCFA [58] employ multi-resolution cross-frame attention to handle temporal variations by disentangling static and dynamic contexts within video frames, allowing for a refined segmentation process that distinguishes between stationary and moving elements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Despite advancements, transformer-based methods face high computational costs due to the quadratic complexity of self-attention, especially with high-resolution frames or long videos [63]. Additionally, they are usually designed to learn temporal information in a short video sequence due to the high complexity. This gap highlights the need for more" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "content": "efficient and holistic models that can simultaneously manage computational costs while effectively modeling long-range temporal information in VSS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 395, + 427, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 395, + 427, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 395, + 427, + 407 + ], + "type": "text", + "content": "2.3. State Space Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": "State space models (SSMs) present a promising alternative for temporal modeling, addressing the shortcomings of RNNs and transformers. Unlike RNNs, which suffer from scaling challenges with sequence length, SSMs like S4 [21, 22] exhibit with linear complexity by imposing diagonal structures on state matrices, making them more efficient for long data sequences. Enhanced through HiPPO [20] initialization, SSMs handle extensive dependencies while requiring less memory, making them well-suited for tasks requiring to store long temporal context. Recently, Mamba [19] introduced a selective-scan mechanism to process temporal data efficiently. This advancement has spurred adaptations in vision-specific models, such as Vision Mamba [74] and VMamba [38], which incorporate Mamba blocks in a hierarchical structure to overcome directional sensitivities and maintain scalability across high-resolution inputs. Additionally, Vim [74] refined the scanning techniques to prevent overfitting. Furthermore, U-Mamba [41] has explored a hybrid network architecture that combines SSMs with convolutional layers. VideoMamba [32] represents one of the early frameworks to leverage SSM-based modules for video understanding, focusing on video clip classification. However, it is limited by its offline processing, which increases memory and computational demands, restricts dense semantic segmentation, and hinders its ability" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24213" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 298, + 323 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 298, + 323 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 298, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 298, + 323 + ], + "type": "image", + "image_path": "b0e0a5e3e62b6931fef56838a88dd5b9e22fa32c6c0208648cbdd91c4e930599.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 297, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 297, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 297, + 367 + ], + "type": "text", + "content": "Figure 3. Internal structure of the TV3S block, illustrating the flow of internal operations, and the propagation of hidden states for efficient spatiotemporal integration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 373, + 230, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 230, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 230, + 385 + ], + "type": "text", + "content": "to capture the global context of sequences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 386, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 386, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 296, + 531 + ], + "type": "text", + "content": "However, despite these advancements, current vision-specific models largely rely on hierarchical designs fail to fully exploit varying temporal resolutions and context lengths, highlighting a critical need for improved mechanisms for temporal state sharing. To this end, our proposed TV3S architecture integrates parameter-efficient SSMs to enhance effective temporal state sharing, thus proving advantageous for VSS. This integration not only underscores the computational efficiency and scalability of SSMs in VSS but also emphasizes their suitability for tasks requiring long-range temporal coherence, representing a significant novelty in the field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 139, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 139, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 139, + 558 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 297, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 297, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 297, + 625 + ], + "type": "text", + "content": "Fig. 2 illustrates the proposed TV3S architecture, which effectively captures and integrates temporal dependencies within video sequences using a state space modeling framework. It aims to leverage both temporal and spatial information to achieve accurate and consistent segmentation results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 634, + 176, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 176, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 176, + 647 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "The architecture processes a sequence of video frames denoted as " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{I_{t - l},\\dots,I_{t - k},\\dots,I_{t - 1},I_t\\}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " represents the frame at time " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{l,\\dots,k,\\dots,1\\}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " are temporal offsets capturing frames from the past relative to " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ". These frames are fed into an image encoder such as MiT [69] and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "Swin [39]. The encoder produces a corresponding set of encoded feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "\\{E_{t - l},\\dots ,E_{t - k},\\dots ,E_{t - 1},E_t\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": ", capturing rich spatial context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 108, + 556, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 556, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 556, + 252 + ], + "type": "text", + "content": "The TV3S architecture processes these feature maps sequentially through a series of " + }, + { + "bbox": [ + 313, + 108, + 556, + 252 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 108, + 556, + 252 + ], + "type": "text", + "content": " TV3S blocks (with a default value of " + }, + { + "bbox": [ + 313, + 108, + 556, + 252 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 313, + 108, + 556, + 252 + ], + "type": "text", + "content": "), each containing two TSS modules that independently process spatial patches of the feature map, facilitating parallelization. Layer normalization is applied before each TSS module, and residual connections are included after each module to stabilize training. Each TSS module consists of operations including a linear projection, a 1-dimension depthwise convolution (DWConv1), a SiLU activation feeding into the state space model (S6), and finally ending with another linear projection to project back to the input dimension, as detailed in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 538, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 538, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 538, + 270 + ], + "type": "text", + "content": "3.2. Temporal Video State Space Sharing Block" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "content": "The TV3S block is the core unit of the TV3S architecture, which enables seamless integration of temporal information across frames. Each block consists of two TSS modules that employ a state space model to capture and propagate temporal dependencies within frames. The encoded feature map " + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{t}" + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "content": " is partitioned into non-overlapping patches of size " + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "inline_equation", + "content": "w \\times w" + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "content": ". Each patch " + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "inline_equation", + "content": "P_{t}^{i,j}" + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "content": " is then flattened, transforming the spatial dimensions into a linear sequence suitable for parallel processing, with " + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "content": " representing the patch indices. Mathematically, " + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "inline_equation", + "content": "P_{t}^{i,j}" + }, + { + "bbox": [ + 313, + 276, + 556, + 397 + ], + "type": "text", + "content": " can be expressed as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 401, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 401, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 331, + 401, + 553, + 422 + ], + "type": "interline_equation", + "content": "\\boldsymbol {P} _ {t} ^ {i, j} = \\boldsymbol {E} _ {t} [ i: i + w, j: j + w,: ] \\in \\mathbb {R} ^ {C _ {E} \\times w \\times w}, \\tag {1}", + "image_path": "e608edf68442e1431b16f4fe16b3862bf420a5c5d22c95140f050ca97fe6ef66.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 333, + 418, + 488, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 418, + 488, + 433 + ], + "spans": [ + { + "bbox": [ + 333, + 418, + 488, + 433 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {P}} _ {t} ^ {i, j} = \\operatorname {F l a t t e n} \\left(\\boldsymbol {P} _ {t} ^ {i, j}\\right) \\in \\mathbb {R} ^ {C _ {E} \\times (w \\times w)},", + "image_path": "7d7b597d3774ad1eac31c9ff4f27d8bfa9d787d3ca3c150d598cf10302aef17d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "inline_equation", + "content": "C_E" + }, + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "text", + "content": " denotes the number of channels in " + }, + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "inline_equation", + "content": "E_t" + }, + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "text", + "content": " after encoding " + }, + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 313, + 438, + 556, + 570 + ], + "type": "text", + "content": ". In contrast to traditional Mamba variants [38, 41, 74], which process sequences sequentially, our approach parallelizes the handling of each window or patch, offering two main advantages. First, the encoder effectively encodes complex spatial information across channels, eliminating for further processing of spatial relationships. Second, minimal changes between consecutive frames allow window-based processing to efficiently capture temporal dynamics, enabling a highly parallelized and computationally efficient model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 574, + 504, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 504, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 504, + 586 + ], + "type": "text", + "content": "State space model for temporal aggregation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "text", + "content": "Each TSS module within a TV3S block utilizes a state space model to capture temporal dependencies between frames. Given encoded feature map " + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{E}_t" + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "text", + "content": ", the hidden state " + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{H}_t" + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "text", + "content": " is updated using the current flattened patch " + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{P}}_t^{i,j}" + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "text", + "content": " and the previous hidden state " + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{H}_{t-1}" + }, + { + "bbox": [ + 313, + 586, + 555, + 647 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 328, + 651, + 553, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 651, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 328, + 651, + 553, + 671 + ], + "type": "interline_equation", + "content": "\\boldsymbol {H} _ {t} ^ {i, j} = f _ {A} (\\Delta , \\boldsymbol {A} _ {s}) \\boldsymbol {H} _ {t - 1} ^ {i, j} + f _ {B} (\\Delta , \\boldsymbol {A} _ {s}, \\boldsymbol {B} _ {s}) \\hat {\\boldsymbol {P}} _ {t} ^ {i, j}, \\tag {2}", + "image_path": "e787c400142cfa2f8b5a8a6d428ca62c0286a94d6f0684542de63695893e981d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 327, + 668, + 402, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 668, + 402, + 682 + ], + "spans": [ + { + "bbox": [ + 327, + 668, + 402, + 682 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {t} ^ {i, j} = \\boldsymbol {C} _ {s} \\boldsymbol {H} _ {t} ^ {i, j}.", + "image_path": "8fac47bf4ad7f6f6327efd66941ee26d43f3bd59eb679b8c92feb4139d6bfd01.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{A}_s \\in \\mathbb{R}^{N_s \\times N_s}" + }, + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{B}_s \\in \\mathbb{R}^{N_s \\times C_E}" + }, + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{C}_s \\in \\mathbb{R}^{C_E \\times N_s}" + }, + { + "bbox": [ + 313, + 688, + 555, + 715 + ], + "type": "text", + "content": " are learnable state space matrices. As seen in Eq. (2)," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24214" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "the matrices " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " are discretized using the time scale parameter " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\Delta_{s}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " with fixed discretization functions " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "f_{A}(.)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "f_{B}(.)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", referred as discretization rules, as described in [19] due to their many advantages. Here, " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " denotes the hidden states dimensionality of the state space model and the aggregated feature output is denoted as " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "F_{t}^{i,j}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". This mechanism enables each patch to independently capture and propagate its temporal dependencies across two dimensions " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". Consequently, temporal data storage consists of a total of " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\frac{W}{w} \\times \\frac{H}{w}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " hidden states, promoting effective temporal consistency and high parallelization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "content": "Shifting and edge handling. To enhance spatial context sharing, the reshaped feature maps undergo a shifting operation inspired by Swin Transformer [39]. This shift is parameterized as " + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "inline_equation", + "content": "s < w" + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "content": " (typically " + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "inline_equation", + "content": "s = w / 2" + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "content": "). The shifting rearranges the blocks so that edge blocks receive information from adjacent regions, thereby enriching spatial-temporal representation. The shifted feature maps are then re-partitioned into patches, with those on the right and bottom edges further subdivided into smaller sub-blocks of dimensions " + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "inline_equation", + "content": "2 \\times w \\times \\frac{w}{2}" + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "inline_equation", + "content": "2 \\times \\frac{w}{2} \\times w" + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "inline_equation", + "content": "4 \\times \\frac{w}{2} \\times \\frac{w}{2}" + }, + { + "bbox": [ + 55, + 207, + 296, + 373 + ], + "type": "text", + "content": ", as visualized in Fig. 2. These re-partitioned patches are processed through a second TSS module in the same TV3S block, integrating the newly acquired spatial context with temporal information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "text", + "content": "In the proposed architecture, each TV3S block consists of a pair of TSS modules, with one module processing unshifted blocks and the other handling shifted blocks. By stacking " + }, + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "text", + "content": " such TV3S blocks, the architecture forms a deep hierarchical structure termed TV3S architecture, which is capable of capturing intricate temporal dependencies across frames. The final aggregated feature representation, " + }, + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{N}" + }, + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "text", + "content": ", is passed through a linear projection layer to map it to " + }, + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 375, + 296, + 495 + ], + "type": "text", + "content": " classes, followed by an interpolation to produce the final segmentation output." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 503, + 162, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 162, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 162, + 516 + ], + "type": "text", + "content": "3.3. Training Strategy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 582 + ], + "type": "text", + "content": "The training strategy for the TV3S architecture focuses on optimizing both temporal feature aggregation and robust spatial feature extraction to enable efficient learning of short- and long-term dependencies within video sequences, facilitating high VSS performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "During training, sequential input frames are taken at the intervals " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{I_{t-9}, I_{t-6}, I_{t-3}, I_t\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " the same as [42]. Each of the frames in set " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{I_{t-k}\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "k = \\{9, 6, 3, 0\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", is passed through the encoder to generate encoded features " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{t-k}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " encapsulating rich spatial information. An intermediate feature " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{O}_{t-k}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " is extracted straight from " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{t-k}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " with the use of linear projection that aligns the channel dimensions of " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{t-k}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " with the number of classes " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " in the dataset. Concurrently, the encoded features " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{t-k}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " are fed into the TV3S blocks to integrate temporal information, producing aggregated features " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{t-k}^N" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 319, + 87, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 545, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 545, + 84 + ], + "type": "text", + "content": "Algorithm 1 Inference Procedure for TV3S Architecture" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": "1: Input: Frame sequence " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\{I_t\\}" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " \n2: Output: Segmentation map " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "O_t" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " \n3: Initialize: Hidden states " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{H}" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " for all TV3S blocks \n4: for each frame " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{I}_t" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " in the sequence do \n5: " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{E}_t \\gets \\text{Encoder}(\\pmb{I}_t) \\quad //" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " Extract features \n6: for each patch " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{E}_t" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " do \n7: " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{P}}^{i,j} \\gets \\text{Flatten}(\\pmb{E}_t[i:i+w,j:j+w,:])" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " \n8: for each TV3S block " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " do \n9: //Update Features and Hidden State. \n10: " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t^{n,i,j}, \\pmb{H}_t^{n,i,j} \\gets \\text{TV3S}^n(\\hat{\\pmb{P}}^{i,j}, \\pmb{H}^{n,i,j})" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " \n11: end for \n12: end for \n13: " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "O_t \\gets \\text{Interpolate}(\\text{Linear}(\\pmb{F}_t^N)) \\quad //" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " Segmentation \n14: " + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{H} \\gets \\{\\pmb{H}_t\\} \\quad //" + }, + { + "bbox": [ + 319, + 87, + 545, + 266 + ], + "type": "text", + "content": " Store for next frame \n15: end for" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 274, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 335 + ], + "type": "text", + "content": "With both output predictions extracted, a weighted cross-entropy between these output predictions " + }, + { + "bbox": [ + 313, + 274, + 555, + 335 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 274, + 555, + 335 + ], + "type": "text", + "content": " and the ground-truth segmentation masks " + }, + { + "bbox": [ + 313, + 274, + 555, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 274, + 555, + 335 + ], + "type": "text", + "content": " is computed as training loss for all frames in the input sequence. The training loss is computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 342, + 555, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 342, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 342, + 555, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\sum_ {k = \\{9, 6, 3, 0 \\}} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\hat {\\boldsymbol {O}} _ {t - k}, \\boldsymbol {M} _ {t - k}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {O} _ {t}, \\boldsymbol {M} _ {t}\\right). \\tag {3}", + "image_path": "f11454cb42ddc07d2773a69acdcf7888f49bdb285335b27d19215479bacf55bb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "content": "Loss formulation presented in Eq. (3) employs a dual-loss strategy to ensure accurate segmentation of the final frame " + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "inline_equation", + "content": "O_{t}" + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "content": " while preserving spatial relationships in the intermediate features " + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "inline_equation", + "content": "\\hat{O}_{t - k}" + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "content": " from the encoder. By applying cross-entropy losses " + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "content": ", with intermediate losses weighted at " + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "content": ", the model effectively balances spatial feature learning and temporal coherence, enhancing its capability to extract temporally consistent spatial information for improved segmentation across consecutive frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 491, + 432, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 432, + 503 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 432, + 503 + ], + "type": "text", + "content": "3.4. Inference Procedure" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "During inference, the TV3S architecture processes each frame sequentially while leveraging stored hidden states to maintain temporal coherence across the video sequence, as detailed in Algorithm 1. For each frame " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{I}_t" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", the encoder first extracts the encoded features " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{E}_t" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", which are then partitioned into non-overlapping spatial patches indexed by " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " and flattened into " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{P}_{i,j}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ". Each flattened patch is processed through the TV3S blocks, updating the current hidden states " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{H}_t" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " using the previous states " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{H}_{t-1}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ". After processing all patches, the aggregated feature map " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t^N" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " is reconstructed from the updated hidden states. This feature map is then passed through a linear projection layer and interpolated to match the original image resolution, producing the final segmentation map " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{O}_t" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ". The updated hidden states " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{H}_t" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " are then stored back in " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{H}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " for use with subsequent frames, ensuring continuous temporal integration and consistent segmentation results." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24215" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 408 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 408 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 408 + ], + "type": "table", + "html": "
MethodsBackbonesmIoU↑mVC8↑mVC16↑GFLOPs↓Params(M)↓FPS↑
Segformer† [69]MiT-B136.584.779.926.613.858.7
CFFM [57]MiT-B138.588.684.1-15.529.8
MRCFA [58]MiT-B138.988.884.4-16.240.1
TV3S (Ours)MiT-B140.090.787.036.917.324.7
Segformer [69]MiT-B243.986.081.2100.824.816.2
CFFM [57]MiT-B244.989.885.8143.226.510.1
MRCFA [58]MiT-B245.390.386.2127.927.310.7
TV3S (Ours)MiT-B246.391.588.3553.928.321.9
Mask2Former† [7]R5038.581.376.4110.644.019.4
MPVSS [67]R5037.584.177.238.984.133.9
Mask2Former† [7]R10139.382.577.6141.363.016.9
MPVSS [67]R10138.884.879.645.1103.132.3
DeepLabv3+† [6]R10134.783.278.2379.062.79.2
UperNet† [68]R10136.582.676.1403.683.216.0
PSPNet† [72]R10136.584.279.6401.870.513.8
OCRNet† [71]R10136.784.079.0361.758.114.3
TCB† [42]R10137.887.984.01692--
ETC† [37]OCRNet37.584.179.1361.7--
Segformer [69]MiT-B548.987.883.7185.082.19.4
CFFM [57]MiT-B549.390.887.1413.585.54.5
MRCFA [58]MiT-B549.990.987.4373.084.55.0
TV3S (Ours)MiT-B549.891.788.7137.085.614.0
Mask2Former† [7]Swin-T41.284.580.0114.447.417.1
MPVSS [67]Swin-T39.985.980.439.7114.032.8
TV3S (Ours)Swin-T44.988.083.557.331.722.9
Mask2Former† [7]Swin-S42.184.779.3152.268.914.5
MPVSS [67]Swin-S40.486.080.747.3108.030.6
TV3S (Ours)Swin-S50.689.685.894.153.119.5
", + "image_path": "f7ba27d6f3f93443f3109aaabd1282f3c2bf3cc9c213220fe152814e3ae41bc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 413, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 555, + 444 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of our model with existing methods on the VSPW dataset [42]. Our model achieves a strong balance between accuracy, model complexity, and operational speed. FPS and FLOPs are calculated with an input resolution of " + }, + { + "bbox": [ + 55, + 413, + 555, + 444 + ], + "type": "inline_equation", + "content": "480 \\times 853" + }, + { + "bbox": [ + 55, + 413, + 555, + 444 + ], + "type": "text", + "content": ". (†Frame-by-Frame processing)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 451, + 137, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 137, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 137, + 464 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 475, + 171, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 171, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 171, + 488 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": "Implementation details. The implementation of our approach is based on MMSegmentation codebase and all the experiments including training and inference were conducted with 2 A100 NVIDIA GPUs. The main experiments are done with the backbone same as the SegFormer (Variants with MiT and Swin) which are pre-trained with ImageNet. While the model is aimed to work with any number of frame sequences, during training, the model is trained with just three reference frames, " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\{k_1,k_2,k_3\\} = \\{-9, -6, -3\\}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": ". More information on training is found in Sec. 3.3. To improve receptive field when processing the features from the backbone, the window size " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " and the shift " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " are set to 20 and 10 with number of TV3S blocks " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " set to 4 (more information on these at Sec. 4.3) and the decoder made with mamba following its default parameters following [19] with the input embedding dimension of 256 matching the SegFormer implementation. The images from VSPW dataset [42] are cropped down to " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": "are augmented using various augmentation methods during training, including cropping, resizing, flipping and addition of photometric distortion during training. Optimization of the model is done with the use AdamW optimizer and a \"poly\" learning rate schedule initializing the learning rate at " + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "inline_equation", + "content": "6e - 5" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": ". Testing is performed using the context of full video with the frame receiving the context of all the past frames within the video through the hidden states " + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "inline_equation", + "content": "\\pmb{H}" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": " due to its high efficiency & effectiveness following Sec. 3.4. Note that for all cases there was no use of post-processing on the obtained output like in [27]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Datasets. The experiments were mainly conducted with the use of VSPW dataset [42] which stands as one of the largest VSS benchmark. The dataset consists of training, validation and test subsets containing 2,806 clips (198,244 frames), 343 clips (24,502 frames) and 387 clips (28,887 frames) accordingly. The dataset consisting of a rich 124 categories with a dense annotation of frame rate of 15fps contrasts itself from the previously available datasets, which had very sparse annotation with just one frame being annotated every 10s of frames. Furthermore, the dataset cover" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24216" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 171 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 171 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 171 + ], + "type": "table", + "html": "
MethodsBackbonesmIoU↑GFLOPs↓Params(M)↓FPS↑
ETC† [37]R1871.1434.1--
SegFormer† [69]MiT-B071.9-3.758.5
CFFM [57]MiT-B074.080.74.615.8
MRCFA [58]MiT-B072.877.54.216.6
SegFormer† [69]MiT-B174.1-13.846.8
CFFM [57]MiT-B175.1158.715.411.7
MRCFA [58]MiT-B175.1145.014.913.0
TV3S (ours)MiT-B175.683.617.325.1
", + "image_path": "520d899f14f566d4d8fcff58f7c55a1ded872aa5d780f73402f1fd5e73ca5531.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 219, + 296, + 293 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 296, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 207 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of our method with efficient alternative approaches on the Cityscapes dataset [9], using a resolution of " + }, + { + "bbox": [ + 55, + 175, + 296, + 207 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 55, + 175, + 296, + 207 + ], + "type": "text", + "content": ". (†Frame-by-Frame processing)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 219, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 293 + ], + "type": "table", + "html": "
ModelsEvaluation (mIoU)
AddConcatDirect
1 TSS (No Shift)37.637.838.0
TV3S (No Shift)38.138.438.9
TV3S (Shift)38.539.339.5
", + "image_path": "bda2ca1de0f64045d88ad5882c602828c4e1206e339eaf38fb0a3e1c257462f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "content": "ing various different scenarios of both outdoor and indoor scenes makes it suitable for training models to well verify the adaptability of the performance standing as the best benchmark in the field of VSS. While most experiments and training were done with the VSPW dataset [42], the proposed method was also evaluated on the Cityscapes dataset [9] which annotates one frame out of every 30 frames for benchmarking the results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "text", + "content": "Evaluation metrics. We use mean IoU (mIoU) and mean video consistency (mVC) as key metrics. The metric mVC evaluates the smoothness of predicted segmentation maps over time, assessing performance in the temporal domain. More formally, given a video clip " + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\{I_c\\}_{c=0}^{C_v}" + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "text", + "content": " with ground truth masks " + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\{M_c\\}_{c=1}^{C_v}" + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "text", + "content": " and predicted outputs " + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\{O_c\\}_{c=1}^{C_v}" + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "inline_equation", + "content": "VC_n" + }, + { + "bbox": [ + 55, + 427, + 296, + 511 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 564 + ], + "type": "interline_equation", + "content": "\\mathrm {V C} _ {n} = \\frac {1}{C _ {v} - n + 1} \\sum_ {i = 1} ^ {C _ {v} - n + 1} \\frac {\\left(\\cap_ {i} ^ {i + n - 1} \\boldsymbol {M} _ {i}\\right) \\cap \\left(\\cap_ {i} ^ {i + n - 1} \\boldsymbol {O} _ {i}\\right)}{\\cap_ {i} ^ {i + n - 1} \\boldsymbol {M} _ {i}}, \\tag {4}", + "image_path": "42f2dd064c26abd78da3626a646353ad8d63d352fa3544761308ae8d2271278b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{v}} \\geq n" + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "text", + "content": ". Once the " + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "inline_equation", + "content": "VC_{n}" + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "text", + "content": " of all the videos are computed, their mean is computed to obtain the " + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{mVC}_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "text", + "content": ". Eq. (4) shows that " + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{mVC}_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 564, + 296, + 637 + ], + "type": "text", + "content": " finds the common areas of the predicted masks among frames which indicates the level of consistency of the prediction masks across time. More information on the metric can be found at [42]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 646, + 272, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 272, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 272, + 659 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Our model is compared against state-of-the-art models on the VSPW dataset [42], as shown in Tab. 1. The table is divided into three groups based on model size and backbone, offering insights at varying scales. In the first group" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 320, + 70, + 547, + 168 + ], + "blocks": [ + { + "bbox": [ + 55, + 297, + 295, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 319 + ], + "type": "text", + "content": "Table 3. Evaluation on the implication of the shifted mechanism and the output method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 70, + 547, + 168 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 547, + 168 + ], + "type": "table", + "html": "
Window sizesmIoUmVC8mVC16
438.589.084.6
638.689.284.7
1239.389.385.0
1639.289.385.0
2040.090.787.0
2840.088.884.2
3639.989.184.9
", + "image_path": "f5410ceb1170bd3edbaf85cf17bcb4735b92e8b63b761d0842cf69513f0fead9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 186, + 554, + 251 + ], + "blocks": [ + { + "bbox": [ + 326, + 171, + 541, + 182 + ], + "lines": [ + { + "bbox": [ + 326, + 171, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 326, + 171, + 541, + 182 + ], + "type": "text", + "content": "Table 4. Impact of the window size on model performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 186, + 554, + 251 + ], + "lines": [ + { + "bbox": [ + 315, + 186, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 554, + 251 + ], + "type": "table", + "html": "
BackbonesTV3S BlocksmIoUmVC8mVC16
MiT-B1136.288.183.6
237.488.684.3
337.688.583.5
440.090.787.0
", + "image_path": "0d182467ffde0b7b1fd0ab5096b869dc796d95d1202c9b00fcffcecb6cc86b5a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 256, + 554, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 277 + ], + "type": "text", + "content": "Table 5. Performance metrics based on the number of TV3S blocks in the model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "type": "text", + "content": "with small models (<30M parameters), our method outperforms the baselines, demonstrating efficiency even with limited model capacity. In the second group with larger models (>30M parameters), TV3S achieves near-state-of-the-art performance by effectively capturing rich contextual information. The third group focuses on the Swin Transformer backbone, where our model excels with over 8 mIoU ahead of the next best, highlighting its ability to preserve temporal correlations and ensure spatial accuracy. Overall, our method demonstrates superior visual consistency across all groups, further illustrated by Fig. 4 showcasing temporally consistent segmentation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "content": "We also present results on the Cityscapes dataset [9] in Tab. 2, using smaller model variants at an input resolution of " + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "inline_equation", + "content": "1024 \\times 512" + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "content": ". Due to the dataset's annotation structure, metrics such as " + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{mVC}_{8}" + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{mVC}_{16}" + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "content": " are not applicable, but our model still achieves top performance in mIoU, underscoring its strong generalizability across datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 408, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 408, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 408, + 523 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 529, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 567 + ], + "type": "text", + "content": "All ablation studies were conducted on the VSPW dataset [42] using the MiT-B1 backbone and following the same training and inference strategies as previously described." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Impact of shifted representations and feature integration. We evaluated the effectiveness of different TSS/TV3S decoder (see Tab. 3), including one TSS layer, unshifted TV3S, and shifted TV3S. Despite the only difference between unshifted and shifted TV3S being the feature representation, the results clearly show that shifted representation significantly outperforms its unshifted counterparts, highlighting their effectiveness. We further tested adding feature maps through addition and concatenation before the final prediction, but both methods reduced performance, suggesting that reintroducing spatial information after temporal modeling disrupts temporal coherence." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24217" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 542, + 256 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 542, + 256 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 542, + 256 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 542, + 256 + ], + "type": "image", + "image_path": "dd706f4a7dc08bf41e498e59bffa00dfcc4e95279c405830fd08b4339586d0e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 256, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 555, + 278 + ], + "type": "text", + "content": "Figure 4. Qualitative example of our TV3S architecture compared with the current baseline. This displays both improved performance in performing spatial predictions and utilizing the temporal information to produce temporally consistent segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 284, + 296, + 373 + ], + "blocks": [ + { + "bbox": [ + 56, + 284, + 296, + 373 + ], + "lines": [ + { + "bbox": [ + 56, + 284, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 296, + 373 + ], + "type": "table", + "html": "
Number of FramesmIoUmVC8mVC16
137.984.579.1
239.286.981.8
439.588.984.2
839.790.485.9
1639.791.287.2
3239.791.587.8
", + "image_path": "b92327d9210825984e13a94a685def021ff89f3c17b1d100c9e19e6e8fffef69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 398 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 398 + ], + "type": "text", + "content": "Table 6. Performance based on the amount of temporal information available during inference" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 402, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 510 + ], + "type": "text", + "content": "Impact of window size. TV3S establishes temporal relationships with window size controlling the spatial context during processing. Larger windows help summarize stable scenes but may overload the model in dynamic settings, while smaller windows enhance understanding in fast-changing environments but may miss correlations in slower scenes. As shown in Tab. 4, the optimal performance is achieved with a " + }, + { + "bbox": [ + 55, + 402, + 296, + 510 + ], + "type": "inline_equation", + "content": "20 \\times 20" + }, + { + "bbox": [ + 55, + 402, + 296, + 510 + ], + "type": "text", + "content": " window size, striking a balance between spatial context and model efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 624 + ], + "type": "text", + "content": "Effect on the number of TV3S blocks. In our experiments, we assessed the impact of varying the number of TV3S blocks, as shown in Tab. 5. Increasing the number of blocks improved mIoU from 36.18 with one block to 40.00 with four blocks, along with enhanced temporal consistency, as shown by the " + }, + { + "bbox": [ + 55, + 514, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{mVC}_{8}" + }, + { + "bbox": [ + 55, + 514, + 296, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 514, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{mVC}_{16}" + }, + { + "bbox": [ + 55, + 514, + 296, + 624 + ], + "type": "text", + "content": " scores. This upward trend suggests that stacking TV3S blocks better aggregates temporal features, capturing complex relationships without introducing excessive redundancy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 634, + 163, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 163, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 163, + 647 + ], + "type": "text", + "content": "4.4. Temporal Context" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Importance of long temporal context. Given Mamba's RNN-like structure, we conducted an evaluation of its performance on the VSPW dataset [42] with varying temporal context sizes that ranged from 1 to 32 frames. In our approach, the videos that were shorter than 32 frames were" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 286, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 417 + ], + "type": "text", + "content": "excluded from the and at the same time the first 32 frames were omitted during inference to ensure a fair and unbiased evaluation of the model's capabilities. The findings presented in Tab. 6 indicate that there is a notable improvement in performance as the temporal context sizes increase. However, this enhancement reaches a plateau once the temporal information saturates. This saturation effect is likely attributed to the limited changes that can be captured within smaller temporal windows, which causes diminishing returns in performance improvements beyond a certain threshold." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "content": "The work proposes a TV3S architecture that addresses the challenges of VSS by capturing temporal dynamics. The proposed structure is designed to achieve good computational efficiency and scalability. By leveraging independent processing of spatial blocks through state space models enhanced with Mamba [19], our approach enables parallel computation during both training and inference. This design mitigates the time delay and high memory demand typically associated with sequential processing in traditional state space and recurrent models, making it highly suitable for long, high-resolution video sequences. Extensive experiments on the VSPW [42] and Cityscapes [9] datasets show that the TV3S architecture not only surpasses existing methods in segmentation accuracy but also significantly improves temporal prediction consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Future works. Our model enhances the encoder's focus on key temporal features, integrating spatial and temporal information for improved video frame segmentation. Its adaptability extends to tasks like object detection and action recognition, while its multi-modal data integration offers new research opportunities in audio-visual learning, emphasizing temporal synchronization." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24218" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "type": "text", + "content": "This work is supported by the Ministry of Education Singapore Tier 1 project funding grant No. MOE Tier 1 RG98/24 and by Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 183, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 183, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 295, + 216 + ], + "type": "text", + "content": "[1] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. In BMVC, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 218, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 295, + 250 + ], + "type": "text", + "content": "[2] Gowtham Atluri, Anuj Karpatne, and Vipin Kumar. Spatiotemporal data mining: A survey of problems and methods. ACM Computing Surveys, 51(4):1-41, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 283 + ], + "type": "text", + "content": "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 286, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 294, + 328 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4): 834-848, 2017. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 433, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 294, + 476 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 478, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 294, + 520 + ], + "type": "text", + "content": "[8] Jingchun Cheng, Yi-Hsuan Tsai, Shengjin Wang, and Ming-Hsuan Yang. SegFlow: Joint learning for video object segmentation and optical flow. In ICCV, pages 686-695, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 523, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 294, + 577 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 621 + ], + "type": "text", + "content": "[10] Henghui Ding, Xudong Jiang, Bing Shuai, Ai Qun Liu, and Gang Wang. Context contrasted feature and gated multiscale aggregation for scene segmentation. In CVPR, pages 2393-2402, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "text", + "content": "[11] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[13] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In AAAI, pages 10713-10720, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 143, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 175 + ], + "type": "text", + "content": "[15] Ekrem Emre Yurdakul and Yucel Yemez. Semantic segmentation of rgbd videos with recurrent fully convolutional neural networks. In ICCV workshops, pages 367-374, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 177, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 553, + 209 + ], + "type": "text", + "content": "[16] Jun Fu, Jing Liu, Hajie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 212, + 553, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 212, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 553, + 254 + ], + "type": "text", + "content": "[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, pages 641–649, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 257, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 257, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 257, + 553, + 300 + ], + "type": "text", + "content": "[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 303, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 553, + 334 + ], + "type": "text", + "content": "[19] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 337, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 369 + ], + "type": "text", + "content": "[20] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. HiPPO: Recurrent memory with optimal polynomial projections. In NeurIPS, pages 1474–1487, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[21] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 406, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 553, + 449 + ], + "type": "text", + "content": "[22] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. In NeurIPS, pages 572-585, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 452, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 452, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 452, + 553, + 483 + ], + "type": "text", + "content": "[23] Shunan Guo, Zhuochen Jin, David Gotz, Fan Du, Hongyuan Zha, and Nan Cao. Visual progression analysis of event sequence data. IEEE TVCG, 25(1):417-426, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 486, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 486, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 486, + 553, + 507 + ], + "type": "text", + "content": "[24] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 510, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 553, + 552 + ], + "type": "text", + "content": "[25] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In CVPR, pages 8818-8827, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 555, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 587 + ], + "type": "text", + "content": "[26] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In CVPR, pages 8866-8875, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "type": "text", + "content": "[27] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 656 + ], + "type": "text", + "content": "[28] Abhijit Kundu, Vibhav Vineet, and Vladlen Koltun. Feature space optimization for semantic video segmentation. In CVPR, pages 3168-3175, 2016. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[29] Jiangwei Lao, Weixiang Hong, Xin Guo, Yingying Zhang, Jian Wang, Jingdong Chen, and Wei Chu. Simultaneously short-and long-term temporal modeling for semi-supervised video semantic segmentation. In CVPR, pages 14763-14772, 2023. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24219" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[30] Jiangyun Li, Yikai Zhao, Jun Fu, Jiajia Wu, and Jing Liu. Attention-guided network for semantic video segmentation. IEEE Access, 7:140680-140689, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "text", + "content": "[31] Jiangtong Li, Wentao Wang, Junjie Chen, Li Niu, Jianlou Si, Chen Qian, and Liqing Zhang. Video semantic segmentation via sparse temporal transformer. In ACM MM, pages 59-68, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "text", + "content": "[32] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. In ECCV, pages 237-255, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 253 + ], + "type": "text", + "content": "[33] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video K-Net: A simple, strong, and unified baseline for video segmentation. In CVPR, pages 18847-18857, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 300 + ], + "type": "text", + "content": "[34] Xiangtai Li, Haobo Yuan, Wenwei Zhang, Guangliang Cheng, Jiangmiao Pang, and Chen Change Loy. Tube-Link: A flexible cross tube framework for universal video segmentation. In ICCV, pages 13923-13933, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 301, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 294, + 334 + ], + "type": "text", + "content": "[35] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In CVPR, pages 5997-6005, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 335, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 294, + 401 + ], + "type": "text", + "content": "[36] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60-88, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 403, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 403, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 403, + 294, + 437 + ], + "type": "text", + "content": "[37] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In ECCV, pages 352-368, 2020. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 482 + ], + "type": "text", + "content": "[38] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 484, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 484, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 484, + 294, + 529 + ], + "type": "text", + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 530, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 530, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 530, + 294, + 563 + ], + "type": "text", + "content": "[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 565, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 294, + 597 + ], + "type": "text", + "content": "[41] Jun Ma, Feifei Li, and Bo Wang. U-Mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "text", + "content": "[42] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. VSPW: A large-scale dataset for video scene parsing in the wild. In CVPR, pages 4133-4143, 2021. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "type": "text", + "content": "[43] Seyed Shahabeddin Nabavi, Mrigank Rochan, and Yang Wang. Future semantic segmentation with convolutional LSTM. CoRR, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[44] David Nilsson and Cristian Sminchisescu. Semantic video segmentation by gated recurrent flow propagation. In CVPR, pages 6819-6828, 2018. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[45] Pedro O O Pinheiro, Ronan Collobert, and Piotr Dólár. Learning to segment object candidates. In NeurIPS, pages 1441-1447, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "text", + "content": "[46] Andreas Pfeuffer, Karina Schulz, and Klaus Dietmayer. Semantic segmentation of video sequences with convolutional LSTMs. In IEEE Intelligent Vehicles Symposium, pages 1441-1447, 2019. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "text", + "content": "[47] Rui Qian, Weiyao Lin, John See, and Dian Li. Controllable augmentations for video representation learning. Visual Intelligence, 2(1):1, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 239 + ], + "type": "text", + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 241, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 553, + 284 + ], + "type": "text", + "content": "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 286, + 553, + 319 + ], + "type": "text", + "content": "[50] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In CVPR, pages 3889-3898, 2016. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 320, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 320, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 315, + 320, + 553, + 353 + ], + "type": "text", + "content": "[51] Evan Shelhamer, Kate Rakelly, Judy Hoffman, and Trevor Darrell. Clockwork convnets for video semantic segmentation. In ECCV, pages 852-868, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 553, + 398 + ], + "type": "text", + "content": "[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In NeurIPS, pages 802-810, 2015. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 553, + 453 + ], + "type": "text", + "content": "[53] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPR Workshop, pages 587-597, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 315, + 455, + 553, + 487 + ], + "type": "text", + "content": "[54] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 488, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 488, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 315, + 488, + 553, + 521 + ], + "type": "text", + "content": "[55] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 523, + 553, + 555 + ], + "type": "text", + "content": "[56] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 315, + 557, + 553, + 600 + ], + "type": "text", + "content": "[57] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In CVPR, pages 3126-3137, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 553, + 644 + ], + "type": "text", + "content": "[58] Guolei Sun, Yun Liu, Hao Tang, Ajad Chhatkuli, Le Zhang, and Luc Van Gool. Mining relations among cross-frame affinities for video semantic segmentation. In ECCV, pages 522-539, 2022. 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 689 + ], + "type": "text", + "content": "[59] Guolei Sun, Yun Liu, Henghui Ding, Min Wu, and Luc Van Gool. Learning local and global temporal contexts for video semantic segmentation. IEEE TPAMI, 46(10):6919-6934, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "text", + "content": "[60] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. VMRNN: Integrating vision mamba and" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24220" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 691 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "LSTM for efficient and accurate spatiotemporal forecasting. In CVPR Workshop, pages 5663-5673, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[61] Vassilios Tsakanikas and Tasos Dagiuklas. Video surveillance systems-current status and future trends. Computers & Electrical Engineering, 70:736-753, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[62] Sepehr Valipour, Mennatullah Siam, Martin Jagersand, and Nilanjan Ray. Recurrent fully convolutional networks for video segmentation. In IEEE Winter Conference on Applications of Computer Vision, pages 29-36, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[63] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[64] Bowen Wang, Liangzhi Li, Yuta Nakashima, Ryo Kawasaki, Hajime Nagahara, and Yasushi Yagi. Noisy-LSTM: Improving temporal awareness for video semantic segmentation. IEEE Access, 9:46810-46820, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[65] Hao Wang, Weining Wang, and Jing Liu. Temporal memory attention for video semantic segmentation. In ICIP, pages 2254-2258, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "text", + "content": "[66] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[67] Yuetian Weng, Mingfei Han, Haoyu He, Mingjie Li, Lina Yao, Xiaojun Chang, and Bohan Zhuang. Mask propagation for efficient video semantic segmentation. In NeurIPS, pages 7170-7183, 2024. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "text", + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 2, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "text", + "content": "[70] Chenliang Xu, Caiming Xiong, and Jason J Corso. Streaming hierarchical video segmentation. In ECCV, pages 626-639, 2012. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 489, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 522 + ], + "type": "text", + "content": "[71] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, pages 173-190, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "type": "text", + "content": "[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "text", + "content": "[74] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[75] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In CVPR, pages 2349-2358, 2017. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24221" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_content_list.json b/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..499bf770070733942c04a8d1610e03c213eaef2c --- /dev/null +++ b/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_content_list.json @@ -0,0 +1,1376 @@ +[ + { + "type": "text", + "text": "Exploration-Driven Generative Interactive Environments", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/523a425dc0fbd2aff438c9045131f881214fbf172b51a5b1b61e150d5269b881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 150, + 178, + 846, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 282, + 326, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern world models require costly and time-consuming collection of large video datasets with action demonstrations by people or by environment-specific agents. To simplify training, we focus on using many virtual environments for inexpensive, automatically collected interaction data. Genie [5], a recent multi-environment world model, demonstrates simulation abilities of many environments with shared behavior. Unfortunately, training their model requires expensive demonstrations. Therefore, we propose a training framework merely using a random agent in virtual environments. While the model trained in this manner exhibits good controls, it is limited by the random exploration possibilities. To address this limitation, we propose AutoExplore Agent - an exploration agent that entirely relies on the uncertainty of the world model, delivering diverse data from which it can learn the best. Our agent is fully independent of environment-specific rewards and thus adapts easily to new environments. With this approach, the pretrained multi-environment model can quickly adapt to new environments achieving video fidelity and controllability improvement.", + "bbox": [ + 89, + 316, + 483, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to obtain automatically large-scale interaction datasets for pretraining, we group environments with similar behavior and controls. To this end, we annotate the behavior and controls of 974 virtual environments - a dataset that we name RetroAct. For building our model, we first create an open implementation of Genie - GenieRedux and apply enhancements and adaptations in our version GenieRedux-G. Our code and data are available at https://github.com/insait-institute/GenieRedux.", + "bbox": [ + 88, + 635, + 483, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning from interactive environments allows us to understand and represent the rules, the possible actions, and the consequences that govern them. As an alternative to labori", + "bbox": [ + 89, + 829, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d92641e2d123750a2b97de7fb1b311ddc86da408832ae75ce35a1d568f602a70.jpg", + "image_caption": [ + "Figure 1. Our proposed world model training framework. It consists of a pretrained multi-environment world model on random agent data, and a new AutoExplore Agent that explores an environment and delivers diverse data for fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 906, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ously hand-coded synthetic simulators, world models have emerged as deep learning tools for realistic environment modeling entirely from observations, commonly images of the observed environment [1, 5, 37, 65].", + "bbox": [ + 511, + 609, + 906, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous work such as [19, 23, 33] uses light world models to support goal-driven agents with goal-specific state representations. The focus is on coarse future predictions, not on their high visual quality. In contrast, the objective of recent world models is to achieve high-quality future predictions given past observations and actions. Such recent models are able to offer realistic action execution and even real-time interaction with people [1, 58]. This has become possible with the rise of diffusion, transformers [14, 59], and state space models [17], and by borrowing architectural choices from video generation pipelines [51, 60]. Typically, these generative models are designed to closely match a single selected environment. One of the state-of-the-art models, Genie, distinguishes itself by being trained on many visually diverse environments with similar dynamics, thus", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. firstname(secondname@insait.ai", + "bbox": [ + 114, + 887, + 415, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "demonstrating generalization across new visuals.", + "bbox": [ + 89, + 90, + 413, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building these high-quality statistical simulators requires diverse observations of the environment as well as of the actions to simulate. Some obtain this data by costly video dataset collection and curation with human demonstrations of the actions [1, 5, 63]. If actions are unavailable, an extra component is designed to predict them, which can introduce uncertainty compared to ground truth labels [5, 37]. Extension to new environments with new types of actions in this setting is difficult as it requires again an expensive data collection process. Others, such as [58] have explored retrieving data with an environment-specific agent, in their case - the game Doom.", + "bbox": [ + 89, + 109, + 483, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a framework for accessible and effort-free training of world models in multiple environments. To this end, we first build RetroAct- an annotated and curated large dataset of retro game environments (based on the environments of Stable Retro [48]). We group them based on behavior labels and control descriptions. This grouping allows us to generate large-scale interaction datasets across environments with similar behaviors. Next, we pretrain a multi-environment world model GenieRedux - our open implementation of Genie [5], using a random agent. Unlike [66], which reports the agent's improved behavior from pretraining, we aim to improve the world model. For this, we adapt GenieRedux to virtual environments and implement architectural and training procedure enhancements, resulting in the GenieRedux-G model. We observe that just by training GenieRedux-G on random interactions from subsets of 200 environments and 50 environments with mapped controls, automatically collected from RetroAct, we are able to obtain control behavior (0.450 $\\Delta$ PSNR in 50 environments) and reasonable visual fidelity (26.36 PSNR in 50 environments).", + "bbox": [ + 89, + 292, + 483, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As random actions are limited in their ability to explore the environment, we develop a method to obtain more diverse interaction data to improve the control behavior and visual fidelity of our model. To this end, inspired by [53], we develop our own environment-independent reward function, allowing an agent to explore different environments, entirely without relying on predefined environment rewards. While they aim at a high-performing goal-driven agent, we base our design on improving the underlying large world model for the simulation of environments in terms of higher visual fidelity and improved controllability. For graphical illustrations, see Fig. 1. The objective of our exploration-driven agent is to maximize the world model's uncertainty, estimated by the classification entropy available in the observation prediction stage of GenieRedux-G. Once the diverse data is obtained, we fine-tune GenieRedux-G. We show that this method leads to significant visual (up to 7.4 PSNR) and control (up to $1.4\\Delta$ PSNR) improvements, compared to random agent pretraining.", + "bbox": [ + 89, + 613, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 532, + 90, + 751, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A framework for training world models with cheap data collection - by training an exploration agent based on our world's model uncertainty.", + "- The implementation and release of GenieRedux and GenieRedux-G - open Pytorch models based on [5].", + "- Architectural and loss changes to the model leading to fidelity improvements, based on our tokenizer representation study.", + "- Preparing a large scale environment dataset for multi-environment world model training." + ], + "bbox": [ + 511, + 106, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 271, + 653, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "World models. Initially built as rough imagination models assisting reinforcement learning (RL) agents [10, 19, 21, 24, 53], world models have evolved into independent realistic video generation models conditioned on actions [9, 39, 50, 64]. They facilitate task-specific agent training by providing predictive representations of the environment. Inspired by [20], Ha and Schmidhuber [18] use a VAE to encode visual observations into latent states, with an MDNRNN predicting future states based on prior states, actions, and VAE outputs to facilitate policy learning. DreamerV2 [22] introduce an RL agent, achieving human-level performance in Atari. It encodes images with a CNN and computes posterior and prior stochastic states using recurrent states. Unlike our work, though, it does not assess the agent's impact on world model improvement nor generalize task rewards across different environments.", + "bbox": [ + 511, + 296, + 906, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "World models also aim to generate realistic video conditioned on actions [27, 37, 65]. Genie [5] trains a video tokenizer and a Latent Action Model (LAM) for dynamic next-frame generation. GAIA-1 [27] tackle autonomous driving in unstructured settings by encoding multi-modal inputs into a unified representation and predicting image tokens based on prior inputs, using an autoregressive transformer. Menapace et al. [37] employ an encoder-decoder architecture in which the predicted action labels act as a bottleneck, allowing a user to control the generated video by a discrete action. The key gap in these works is automatic data collection, which is addressed in our approach.", + "bbox": [ + 511, + 537, + 905, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient exploration. The importance of efficient exploration in RL is highlighted by [28]. Early methods enhanced exploration by adding noise [16, 34] or using entropy regularization [40], but they have action space limitations and often fail with complex dynamics, where varied actions do not always drive meaningful exploration. A more direct approach uses heterogeneous actors [26, 29, 52] with diverse exploration strategies to enhance environment exploration. Bayesian methods [54, 57] have also been introduced to create acquisition functions for uncertainty-driven exploration [2, 38, 41-44], but often struggle to generalize to high-dimensional inputs like images.", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27598", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3138283f2fbff07dd930742fb664a31a4e28a7464d6e32665e8d1ed31f672fd3.jpg", + "image_caption": [ + "Figure 2. Method Overview. We propose an alternative to costly human interaction data collection - by exploring environments with an agent. The reward is solely based on the classification uncertainty of our model." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 908, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent exploration methods emphasize state novelty [3, 7, 11, 35, 36, 45, 55, 67], focusing on encouraging agents to assess novelty only after visiting states. In contrast, our approach, inspired by [7, 47, 53], uses model disagreement to proactively guide agents to states with the highest potential without environment target-driven reward. [6, 46] propose exploration agents driven by uncertainty in state transition and simple feature extractors. Instead, we propose an exploration agent designed to improve a world model that does not model states. Plan2Explore [53] enables agents to seek novel states using a reward that maximizes the state entropy of an RSSM model. While Plan2Explore improves goal-driven agents with their framework, we improve a modern transformer world model with a novel exploration-based reward using token uncertainty.", + "bbox": [ + 88, + 388, + 482, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rather than relying on world models, EX2 [15] learns a classifier to distinguish visited states, providing intrinsic rewards for states that are difficult for the classifier to differentiate. KL-divergence-based approaches [30-32], guide exploration by comparing distributions. For example, SMM [32] computes the KL divergence between the policy-induced state distribution and a uniform target. Tao et al. [56] propose an intrinsic reward based on the distance between a state and its nearest neighbors in a low-dimensional feature space. However, low-dimensionality leads to information loss, restricting full state space exploration — an issue we address by the use of a world model.", + "bbox": [ + 89, + 616, + 483, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. RetroAct Dataset", + "text_level": 1, + "bbox": [ + 89, + 814, + 261, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first tackle the problem of accessible training of multi-environment world models by building a framework for cheaply acquiring multi-environment interaction data. In particular, we aim to collect interactions of similar actions", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/58fd854ba9e7c60078c9858b2e37bf48605ad44bc91f70352ec1ee9f5e65c861.jpg", + "image_caption": [ + "Figure 3. RetroAct Annotation. Description of environments in RetroAct by annotated attribute. Better viewed zoomed." + ], + "image_footnote": [], + "bbox": [ + 514, + 383, + 910, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in many environments. Instead of relying on expensive human interaction, we obtain and curate a collection of virtual environments. As a source, we use the Stable Retro framework by [48], which is a collection of retro games across multiple platforms, with an accompanying starting state. We make no use of the defined rewards. We obtain almost all the supported games (974).", + "bbox": [ + 511, + 537, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This raw set contains an environment mix of very different visuals and behaviors. However, in our setting of learning from similar dynamics, it is required to establish correspondence between the environments' behaviors. We perform annotation where for each environment three aspects are classified. The motion style classifies the general style of what and how is moved by the controls, closely relating to game genre; the camera viewpoint; the control axis describing in which direction the player can be moved. The label distributions are shown in Fig. 3. In Tab. 1 we compare our RetroAct with other related datasets. RetroAct distinguishes itself by providing behavior and control annotations, while maintaining a high number of environments.", + "bbox": [ + 511, + 643, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is discovered that the most prevalent type of environment in our set is the platformer - 483 titles. As the largest subset, we filter only these games for further use, as it is required to have many environments exhibiting similar con", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27599", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e3ad0f9fa00c7f7c26ef69ffe4244af79e28f929e5ccd055a37f7005d3c12112.jpg", + "table_caption": [ + "Table 1. Comparison of RetroAct dataset to others." + ], + "table_footnote": [], + "table_body": "
DatasetType#EnvironmentsDiverse BehaviorsOpenBehavior AnnotationControl Annotation
Coinrun [13]Environments1X✓XX
ALE [4]Environments57✓✓XX
Stable Retro [48]Environments1003✓✓XX
Platformers [5]VideosUnknown✓XXX
RetroAct(Ours)Environments974✓✓✓✓
", + "bbox": [ + 120, + 103, + 874, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Controls. Five motion actions are defined for our model - moving left, right, up, down and jump. Each game has its own mapping of buttons to actions. Therefore, we generate a short clip of each of the 5 selected actions for each of the 483 titles and build an annotation tool to observe and annotate the executed action. Eventually, we annotated 2,925 behavior and 2,898 control labels.", + "bbox": [ + 88, + 218, + 480, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After experimenting, we observed that models require more training with a higher number of environments, so we defined two subsets of our large set to handle computational cost: a subset consisting of the first 200 games of 483 behavior-filtered games for pretraining, and another subset of 50 randomly selected action-consistent games using RetroAct's action labels for fine-tuning.", + "bbox": [ + 88, + 324, + 482, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect a large scale dataset by launching a random agent in all of the environments, collecting actions and observations. From the 200-game set we build Platformers-200 - a dataset with 10,000 episodes (50 episodes per game) with 500 frames each at most, resulting in 4.6mln images. From the 50-game set we obtain Platformers-50 - 5000 episodes (100 per game) of length at most 1000, resulting in 4.8mln images. In our protocol, we take $1\\%$ of the sessions of each environment as a validation set. We show that using a random agent is already sufficient to learn a level of controllability and later build on top with an exploration agent of our design.", + "bbox": [ + 88, + 430, + 482, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To validate our GenieRedux implementation, we implement the CoinRun case study in [5]. Using the protocol from above, we obtain a dataset of 10k episodes with a maximum length of 500, resulting in 4mln images.", + "bbox": [ + 88, + 612, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Multi-Environment World Model", + "text_level": 1, + "bbox": [ + 89, + 686, + 395, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given virtual environments, our first goal is to automatically obtain a dataset of image sequences $I_{1}, \\ldots, I_{N}$ and corresponding actions $a_{1}, \\ldots, a_{N-1}$ . Given a sequence $I_{1}, \\ldots, I_{N}$ and past and future actions $a_{1}, \\ldots, a_{N+T-1}$ , our world model aims to predict the future $T$ frames $I_{N+1}, \\ldots, I_{T}$ , corresponding to the actions executed.", + "bbox": [ + 88, + 713, + 482, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GenieRedux. As Genie [5] is not made available by the authors, we create an open implementation and call it GenieRedux. We validate our implementation quantitatively and qualitatively in Sec. 5 and Sup.Mat F. It consists of three components. A video", + "bbox": [ + 88, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenizer encodes input frame sequences into spatiotemporal tokens: $e_1, \\ldots, e_N = T_{enc}(I_1, \\ldots, I_N)$ , and decodes back to images: $I_1, \\ldots, I_N = T_{dec}(e_1, \\ldots, e_N)$ . A Latent Action Model encodes input frame sequences into spatio-temporal tokens: $a_1, \\ldots, a_{N-1} = LAM_{enc}(I_1, \\ldots, I_{N-1})$ , and decodes them to reconstruct future prediction $I_2, \\ldots, I_N = LAM_{dec}(a_1, \\ldots, a_{N-1})$ . A Dynamics module predicts the next frames based on partially masked frame tokens and actions: $I_2, \\ldots, I_{N+T-1} = D(e_1, \\ldots, e_N, \\ldots, e_{N+T-1}; a_1, \\ldots, a_{N+T-1})$ , where in inference $e_N, \\ldots, e_{N+T-1}$ are masked. We adhere closely to Genie's specifications for implementing these components.", + "bbox": [ + 511, + 218, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All components use the causal Spatial Temporal Transformer (STTN) [62]. We use Position Encoding Generator (PEG) [12] for spatial and temporal attention, and Attention with Linear Biases (ALiBi) [49] for temporal attention.", + "bbox": [ + 511, + 400, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train our models with a sequence size of 16 frames and resolution of $64 \\times 64$ to address computational limitations. We train a U-Net-based superresolution network on $50 \\mathrm{~K}$ data samples to upscale the output to $256 \\times 256$ . (Sup.Mat. B)", + "bbox": [ + 511, + 460, + 903, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GenieRedux-G. Building upon the base model, we offer a variant - GenieRedux-G, which is adapted to virtual environments and contains architectural and training improvements. While GenieRedux uses an indispensable LAM model to obtain the actions, we discard it, as ground truth actions are available from our agent. Instead, the one-hot actions are concatenated to each layer of the Dynamics module for conditioning. In this way, we avoid the uncertainty of a prediction.", + "bbox": [ + 511, + 561, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Dynamics module consists of an ST-ViViT encoder, followed by a MaskGIT architecture [8], which predicts indices from the tokenizer's codebook for randomly masked input tokens during training, according to a schedule. As standard cross-entropy is used, token classification has the drawback to penalize equally any prediction different from the ground truth. However, close tokens in the codebook result in significantly fewer changes than far tokens, as also shown in Sec. 5. To enable this concept of a distance between tokens in the classification of $N_{E}$ tokens, we design a Token Distance Cross-Entropy (TDCE) Loss:", + "bbox": [ + 511, + 698, + 905, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT D C E (x, y) = \\left(y ^ {T} K\\right) \\cdot \\text {s o f t m a x} (x) + C E (x, y) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 878, + 903, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27600", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $x \\in \\mathcal{R}^{N_E}$ is the prediction logits, $y \\in \\mathcal{R}^{N_E}$ is the ground truth one-hot class. $K \\in \\mathcal{R}^{N_E \\times N_E}$ is a precomputed table at the start of training of the cosine distances between all tokens; $CE(.)$ denotes standard Cross-Entropy Loss. When an incorrect token class is given probability, it is penalized based on its distance to the ground truth class.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MaskGIT's design is to take as input learnable embeddings, indexed by the tokens predicted by the Tokenizer. They are randomly initialized, and therefore contain none of the content of the tokens. Given that the encoding itself and the distance between tokens can contribute to Dynamic module's performance, we add a skip connection by adding the embedding to the token itself, which improves visual fidelity and controllability of the model.", + "bbox": [ + 89, + 181, + 480, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AutoExplore Agent We extend our framework with an exploration agent that obtains data by going deeper into the environments. We name it AutoExplore Agent. The reward of the agent is entirely based on the world model performance and operates without any environment rewards. Therefore, it can be trained in various environments without tuning to their specifics or relying on a reward definition.", + "bbox": [ + 89, + 323, + 480, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The design of our reward is based on the fact that GenieRedux-G employs classification for token prediction. Each token is predicted by sampling from a categorical distribution over the codebook. We first obtain all $N_T$ token prediction distributions by running GenieRedux-G-50 5 steps back from the current observation $I_c$ for which we want to estimate the reward. We provide 2 images $I_{c - 4}, I_{c - 3}$ , predict 3 images $-I_{c - 2}, \\dots, I_c$ , and take the distributions of the predicted tokens of $I_c$ to obtain $x = [x_1, \\dots, x_t, \\dots, x_{N_T}]$ . We evaluate the uncertainty per predicted token $u_t$ by calculating the entropy over the categorical distribution and normalize it in the range [0, 2]:", + "bbox": [ + 89, + 429, + 480, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nu _ {t} = \\frac {2 \\cdot \\sum_ {i} ^ {N _ {T}} x _ {i} \\cdot \\log \\left(x _ {i}\\right)}{N _ {e}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 619, + 480, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Studying the properties of the Tokenizer representation, we find that a prevalent token is learned representing static parts of the environment. Only the changing parts generate high uncertainty and, therefore, we take the subset $S_{top}$ of $25\\%$ highest uncertainties of the entire set of uncertainties $S = \\{u_t\\}$ . The reward, shown in Eq. 3, establishes the agent's goal to collect data that maximizes uncertainty of the world model.", + "bbox": [ + 89, + 666, + 482, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {25 \\%} = \\{u \\in S \\mid u \\geq Q _ {75} (S) \\} \\tag{3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 801, + 480, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR \\left(I _ {c}\\right) = \\frac {1}{\\left| S _ {25 \\%} \\right|} \\sum_ {u \\in S _ {25 \\%}} u \\tag{4}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 821, + 480, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our agent is an actor-critic, trained with the Policy Gradient method. For the agent architecture, we follow [39]. It", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "consists of a CNN encoder followed by an LSTM. As standard in RL, 4 frames are stacked, max-pooled, and the result is the input to the agent for a single time step.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exploration-driven World Model Training. We initially pretrain GenieRedux-G on Platformers-200 and fine-tune on Platformers-50 to obtain the model GenieRedux-G-50. Then, we train AutoExplore Agent by using GenieRedux-G-50, using it as a source of reward. The details of training the agent are presented in Sup.Mat A.3.", + "bbox": [ + 511, + 152, + 903, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Running the trained exploratory agent on a selected environment, we obtain a new diverse dataset with action demonstrations under unseen scenes. We first fine-tune the decoder of the Tokenizer for 1,000 iterations to adapt to the new unseen scenes. The Dynamic module of GenieRedux-G is then fine-tuned on the new data to achieve greater visual fidelity and controllability under new conditions. In order to build test sets to evaluate our approach, we train an Agent-57 model for each of the environments we explored, using the available environment rewards. More details on the test setup are provided in Sup.Mat A.2.", + "bbox": [ + 511, + 244, + 903, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For visual fidelity evaluation, we use FID (Frechet inception distance) Heusel et al. [25], PSNR (signal-to-noise ratio) and SSIM (structural similarity index measure) Wang et al. [61]. To evaluate controllability, we use the recently proposed $\\Delta_t$ PSNR metric [5], which compares the visual effect of the ground truth action $(\\hat{x}_t)$ versus a random action $(\\hat{x}_t')$ : $\\Delta_t$ PSNR = PSNR $(x_t, \\hat{x}_t) - \\mathrm{PSNR}(x_t, \\hat{x}_t')$ , where $x_t$ is the ground truth frame at time $t$ . A higher $\\Delta_t$ PSNR indicates a higher level of controllability. As in Bruce et al. [5], for all experiments we report $\\Delta_t$ PSNR with $t = 4$ .", + "bbox": [ + 511, + 410, + 903, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 573, + 643, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing GenieRedux and GenieRedux-G. We implement the original CoinRun case study with a random agent, as advised by [5], in order to validate and compare GenieRedux with LAM, and GenieRedux-G which uses agent-provided actions instead. In this study, the presence of LAM is the only difference between the models. We first train on a dataset, collected by a random agent. Visual fidelity results are in Tab. 2. Our GenieRedux implementation exhibits high visual quality and matches all seven CoinRun environment actions, as well as progressing environment motions (demonstrated in Sup.Mat. F). However, as demonstrated by the metrics, GenieRedux-G shows superior visual fidelity and controllability (more in Sup.Mat. F), as it avoids the uncertainty of LAM prediction. This study demonstrates that even using a random agent can result in action performance abilities in the world model.", + "bbox": [ + 511, + 598, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we train an actor-critic agent with PPO on the environment reward, following [13] to collect data and train GenieRedux-TA and GenieRedux-G-TA. Tab. 3 shows evaluation on a test set collected by a trained agent.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27601", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1421fccd8406eabb69f6e540cce18ec6a7d912649df0cf3697cee326057cc86c.jpg", + "table_caption": [ + "Table 2. Comparison of GenieRedux and GenieRedux-G on Basic Test Set. Performed on a test set, collected from the Coinrun environment with randomly sampled actions." + ], + "table_footnote": [], + "table_body": "
ModelBasic Test Set
FID↓PSNR↑SSIM↑
Tokenizer18.1438.250.96
LAM37.0133.970.92
GenieRedux21.8825.510.77
GenieRedux-G18.8833.410.92
", + "bbox": [ + 125, + 131, + 444, + 246 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ed3328ed038c23c673b790d1060122cf1de55bdddddf66ba9d8c236fb92fc74c.jpg", + "table_caption": [ + "Table 3. Comparison of GenieRedux and GenieRedux-G on Diverse Test Set. The models are trained with data collected by random agent and trained agent (-TA), and tested on data collected by a trained agent from the Coinrun environment." + ], + "table_footnote": [], + "table_body": "
ModelDiverse Test Set
FID↓PSNR↑SSIM↑
Tokenizer19.1335.850.94
Tokenizer-TA11.6340.620.97
GenieRedux23.9723.820.73
GenieRedux-G19.5131.660.90
GenieRedux-TA12.5731.970.90
GenieRedux-G-TA12.4034.440.92
", + "bbox": [ + 114, + 316, + 455, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GenieRedux-G outperforms GenieRedux on all settings. Furthermore, models trained on diverse agent-collected data are visually superior to those trained on random agents. The higher $\\Delta$ PSNR of 1.89 for GenieRedux-G-TA compared to 0.70 for GenieRedux-G shows the superiority of diverse data training in controllability. (more in Sup.Mat. F)", + "bbox": [ + 89, + 463, + 483, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Environment Models. Here, we evaluate the models we initially train on many environments from RetroAct. GenieRedux-G-200 is pretrained on the Platformers-200 dataset for 180k iterations. On the validation set, we obtain 23.32 PSNR and 17.12 FID. Using this model as a base, GenieRedux-G-50 is trained on Platformers-50. Its quantitative evaluation on a test set of 10k sessions separately generated from the selected 50 environments is at the start of Tab. 4. As the 50 environments are selected with corresponding action controls between each other, we see a boost in the quality of prediction. Fig. 4 demonstrates that the instructed action is executed successfully by GenieRedux-G. As the up action is rarely used, it serves more as a no-operation action. (more in Sup.Mat C.1)", + "bbox": [ + 89, + 575, + 483, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Study. In this experiment we evaluate the additive gain of each proposed improvement in GenieRedux-G - the additive token input and training with the Token Distance Cross-Entropy Loss. The ablation is performed on a generated test set of 10k sessions, each 500 frames long.", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6d60a36887e6dd6f6b1d683d1e06205bb6ae41ef329c986fbe366e4f9606b45b.jpg", + "table_caption": [ + "Table 4. Ablation study on improvements in GenieRedux-G." + ], + "table_footnote": [], + "table_body": "
ModelFID↓PSNR↑SSIM↑
GenieRedux-G-20022.3125.110.80
GenieRedux-G-5023.8026.360.84
+ Token Input22.9626.650.84
+ TDCE Loss22.9527.060.85
Autoregressive22.1128.070.88
", + "bbox": [ + 535, + 107, + 880, + 202 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/929d3064ff054f87b1bde004904d29cbd7e9311016ee6b0c408afc2702b211fc.jpg", + "image_caption": [ + "Figure 4. Control Of GenieRedux-G-50. Demonstrating all controls of our multi-environment model on multiple games." + ], + "image_footnote": [], + "bbox": [ + 524, + 209, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The data is collected using a random action policy from the environments in Platformers-50. Visual fidelity evaluation is provided in Tab. 4. It can be seen that each component gives our model a benefit in terms of visual fidelity. Finally, we perform an autoregressive evaluation of the best model to achieve our highest score.", + "bbox": [ + 511, + 405, + 906, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tokenizer Representation Study. This experiment provides insights into the inner workings of GenieRedux-G to motivate our proposed changes. As the Dynamics module operates entirely on the token representation, we examine it closely. Fig. 5 shows the reconstructions of an input sequence (first row) and the visualized token representation (last row), where each predicted token index is assigned a different color. The visual features of the first frame are captured by various tokens. Starting with the second frame, the representation drastically changes - a token is specialized in representing the static frame regions compared to the past, while all motion regions are updated with new content. Observing that visually similar patches predict identical or similar tokens, we replace each predicted token with its closest in the codebook. We only keep the special background token unchanged. In the second row of Fig. 5 we show the resulting reconstruction - while some blurriness appears, the image remains largely the same. Conversely, replacing each token with its furthest in the codebook (third row) results in a significantly different image. This property - closer tokens having more similar appearance - motivates our Token Distance Cross-Entropy Loss, which penalizes predicting tokens further away from the ground truth.", + "bbox": [ + 511, + 521, + 906, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 6 visualizes the uncertainty of GenieRedux-G-50 for each predicted token of its Dynamics module given a", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27602", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ead44dfc4c7119736676acb777043924648f12ebfefac020552d1864115d73b7.jpg", + "table_caption": [ + "Table 5. Quantitative Results on 3 environments. We evaluate the benefit of the data from the propose AutoExplore Agent to our models. GenieRedux-G-50 is our pretrained world model on 50 environments. GenieRedux-G-50-ft are fine-tuned models using data from a random agent or AutoExplore (Exploration). GenieRedux-G denotes a non-fine-tuned model, trained with the exploration data." + ], + "table_footnote": [], + "table_body": "
EnvironmentStrategyModelFID↓PSNR↑SSIM↑ΔPSNR↑
Adventure Island IIRandomGenieRedux-G-5041.9926.320.810.83
GenieRedux-G-50-ft42.3427.040.811.19
ExplorationTokenizer-ft11.0138.950.98-
GenieRedux-G11.9428.330.880.37
GenieRedux-G-50-ft12.7730.600.901.47
Random AutoregressiveGenieRedux-G-50-ft41.5527.820.831.24
Exploration AutoregressiveGenieRedux-G-50-ft11.3333.610.942.09
Super Mario BrosRandomGenieRedux-G-5029.8334.240.940.56
GenieRedux-G-50-ft30.1334.540.940.54
ExplorationTokenizer8.0942.000.99-
GenieRedux-G9.5634.000.950.09
GenieRedux-G-50-ft9.5536.130.970.57
Random AutoregressiveGenieRedux-G-50-ft30.8434.850.950.57
Exploration AutoregressiveGenieRedux-G-50-ft9.3337.770.970.76
SmurfsRandomGenieRedux-G-5079.5121.470.690.47
GenieRedux-G-50-ft80.6121.830.700.65
ExplorationTokenizer17.8635.610.98-
GenieRedux-G20.4335.420.800.85
GenieRedux-G-50-ft20.0127.450.851.55
Random AutoregressiveGenieRedux-G-50-ft80.1622.160.710.69
Exploration AutoregressiveGenieRedux-G-50-ft18.9729.530.902.06
", + "bbox": [ + 106, + 135, + 890, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/940b81b9ca8148245440c2acb56b6d04da92ce4710442b5c8cca7d6dff176dff.jpg", + "image_caption": [ + "Figure 5. Tokenizer Representation. Reconstruction images from the tokenizer, and the effect of replacing each token with its closest and furthest in the codebook. Lastly, we visualize the indices of the predicted tokens." + ], + "image_footnote": [], + "bbox": [ + 101, + 484, + 470, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sequence. The uncertainty metric is the entropy of the classification over 1024 codebook tokens. Tokens corresponding to motion have the highest uncertainty; other regions are mostly classified as the \"static\" token. Thus, minimal character movement yields low uncertainty, while forward motion increases it. This motivates us to build AutoExplore Agent's reward based on this uncertainty.", + "bbox": [ + 89, + 714, + 482, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Exploration-based training. We demonstrate our exploration-based training of GenieRedux-G. We perform the procedure on 3 environments - AdventureIslandII, which provides an easy set-", + "bbox": [ + 89, + 840, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/060b83ac3139b750e8665ce11bd220606dc67a266aa2f8636117757232268f03.jpg", + "image_caption": [ + "Figure 6. Dynamics Uncertainty. Shown is the uncertainty per token predicted for each image of an example sequence. Uncertainty is generated in the regions of motion." + ], + "image_footnote": [], + "bbox": [ + 522, + 484, + 893, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ting for the agent to learn (single platform with no enemies at the start), SuperMario Bros provides an enemy and obstacles soon after the start and Smurfs provides a more complex background imagery and different action dynamics. For each of the environments, we train an AutoExplorer Agent. We observe that the agent learns to move forward and navigate obstacles to maximize reward. (more in Sup.Mat. D)", + "bbox": [ + 511, + 626, + 906, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use our pretrained GenieRedux-G-50 model as a baseline and fine-tune it for each environment in two settings - a dataset collected on the selected environment by a random agent and by our AutoExplorer Agent. Each dataset consists of 10k sessions, each 700 frames long. We fine-tune (GenieRedux-G-50-ft) for 10k iterations and pick the best performing model. In our comparison, we also include a GenieRedux-G model trained from scratch on the diverse exploration datasets for 15k iterations to show the effect of pretraining. We perform single-pass generation for all mod", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27603", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a54e18eb289388484a0bdf631c9d312baeb3d6c92dbea5f64de9c04751bbe3fb.jpg", + "image_caption": [ + "Figure 7. AutoExplore Agent vs Random Agent Qualitative Comparison. We show that AutoExplore exhibits better visual quality and avoids losing track of the agent." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/673bbf611cae3d740045fa8cc952e8ed184f5ee50dcdf73c6d7d76853e8986da.jpg", + "table_caption": [ + "Table 6. Comparison of AutoExplore Agent with others." + ], + "table_footnote": [], + "table_body": "
AgentSuperMarioBrosAdventureIslandII
PSNR↓SSIM↓ΔPSNR↓PSNR↓SSIM↓ΔPSNR↓
RF28.580.940.18124.820.780.44
VAE24.400.860.08716.570.50.072
Ours23.810.850.06515.200.410.070
", + "bbox": [ + 94, + 273, + 478, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "els and the more computationally heavy autoregressive evaluation for the fine-tuned models on data from random and AutoExplore Agent's datasets. Tab. 5 shows visual fidelity and controllability metrics for each environment, confirming the effectiveness of our exploration method. The model fine-tuned on AutoExplore Agent's data consistently outperforms the models trained on random actions in terms of visual fidelity. Exploration-based fine-tuning also improves controllability. Environments with small characters and uniform backgrounds can be more challenging for all models to learn. However, the gain in controllability in this case remains noticeable during autoregressive evaluation. Fig. 7 demonstrates the superior quality of our method. In addition, we observe that the multienvironment pretraining leads to significant gains in both studied aspects compared to the nonpretrained model. (more in Sup.Mat. C)", + "bbox": [ + 88, + 364, + 482, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoExplore Agent Evaluation. We compare AutoExplore Agent with exploration-based methods in [6]. We train agents based on SSE of RF and VAE features on top of GenieRedux and compare with ours on Tab. 6. AutoExplore Agent's reward results in maximum world model visual and controllability errors (on 1k episodes of agent actions), fulfilling its intended role in our framework.", + "bbox": [ + 89, + 625, + 483, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Studies. To validate the quality of our final results, we perform a user study in which we ask people to rate from 1 to 5 the quality of samples produced respectively by GenieRedux-G trained on random agent's data and on AutoExplore Agent's data. Each sample in our study consists of two 16-frame clips playing in a synchronized manner - the ground truth clip and our GenieRedux-G-50-ft reconstruction, given two initial frames and generating the rest autoregressively. We provide a total of 120 samples to the users - 40 samples per model and 40 samples of two ground", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd93a57c19be2f18949958e60622afee40ba84b0c64ccfd6c1b60bfa5e8ae4df.jpg", + "image_caption": [ + "Figure 8. User study results. Our user study on two games shows that our model trained with AutoExplore Agent's data is consistently rated higher." + ], + "image_footnote": [], + "bbox": [ + 519, + 95, + 702, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db754866630a6436c6985947509e3e762a1a58281c217489525370380681fb47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 95, + 897, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "truth samples, to establish scale. We give 20 samples from each of the two selected games - SuperMario Bros and AdventureIslandII. We get reviews from 19 participants. The results are shown in Fig. 8. The model, trained on data from AutoExplore Agent is clearly rated closer to the ground truth, establishing the quality of our method.", + "bbox": [ + 511, + 285, + 906, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With a second user study, we evaluate the action accuracy of the generated frames. We use ambiguous single input cases (character starting mid-air) and generate 60 clips with 3 actions on AdventureIslandII. Users prefer our exploration-trained model, rating it $\\mathbf{0.75} \\pm \\mathbf{0.019}$ on a scale from 0 (random preferred) to 1 (exploration preferred). (more in Sup.Mat. E.2)", + "bbox": [ + 511, + 381, + 906, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 517, + 633, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As world models have developed into large models with impressive simulation properties, they require large interaction datasets, complete with diverse observations and actions. Genie [5] demonstrates impressive abilities by training on multiple environments, however, requiring the collection of a large video dataset and a model to infer actions.", + "bbox": [ + 511, + 547, + 906, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we address the heavy burden of data collection and curation by building a new framework for training large world models by collecting interaction data from a large number of virtual environments. We first build an open implementation of Genie - GenieRedux and enhance it into its version GenieRedux-G. We obtain models exhibiting control by pretraining on a large set of virtual environments. We address the overfitting limitations of random data collection policy by proposing AutoExplore Agent, an agent entirely independent of the environment reward, maximizing the uncertainty of GenieRedux-G. After finetuning on the explored environment, our model is able to improve its visual fidelity and controllability much better than training solely on random agent's data. Demonstrating this on multiple environments, we show the potential of our framework to make training of next-generation world models more accessible, cost-effective, and effort-free.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27604", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "INSAIT, Sofia University \"St. Kliment Ohridski\". Partially funded by the Ministry of Education and Science of Bulgaria's support for INSAIT as part of the Bulgarian National Roadmap for Research Infrastructure. This project was supported with computational resources provided by Google Cloud Platform (GCP).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37: 58757-58791, 2024.", + "[2] Kamyar Azizzadenesheli, Emma Brunskill, and Animashree Anandkumar. Efficient exploration through bayesian deep q-networks. In 2018 Information Theory and Applications Workshop (ITA), pages 1-9. IEEE, 2018.", + "[3] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. Advances in neural information processing systems, 29, 2016.", + "[4] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253-279, 2013.", + "[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024.", + "[6] Yuri Burda, Harri Edwards, Deepak Pathak, Amos Storkey, Trevor Darrell, and Alexei A Efros. Large-scale study of curiosity-driven learning. In International Conference on Learning Representations, 2019.", + "[7] Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation. In International Conference on Learning Representations, 2019.", + "[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022.", + "[9] Chang Chen, Yi-Fu Wu, Jaesik Yoon, and Sungjin Ahn. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481, 2022.", + "[10] Silvia Chiappa, Sébastien Racanière, Daan Wierstra, and Shakir Mohamed. Recurrent environment simulators. In International Conference on Learning Representations, 2017.", + "[11] Leshem Choshen, Lior Fox, and Yonatan Loewenstein. Dora the explorer: Directed outreach reinforcement action-selection. arXiv preprint arXiv:1804.04012, 2018.", + "[12] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In The Eleventh International Conference on Learning Representations, 2023." + ], + "bbox": [ + 93, + 246, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In International conference on machine learning, pages 1282-1289. PMLR, 2019.", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020.", + "[15] Justin Fu, John Co-Reyes, and Sergey Levine. Ex2: Exploration with exemplar models for deep reinforcement learning. Advances in neural information processing systems, 30, 2017.", + "[16] Scott Fujimoto, Herke Hoof, and David Meger. Addressing function approximation error in actor-critic methods. In International conference on machine learning, pages 1587-1596. PMLR, 2018.", + "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023.", + "[18] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018.", + "[19] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018.", + "[20] David Ha, Jonas Jongejan, and Ian Johnson. Draw together with a neural network. Retrieved Oct, 5:2021, 2017.", + "[21] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019.", + "[22] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021.", + "[23] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021.", + "[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023.", + "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017.", + "[26] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay. In International Conference on Learning Representations, 2018.", + "[27] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023." + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27605", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In Proceedings of the Nineteenth International Conference on Machine Learning, pages 267-274, 2002.", + "[29] Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent experience replay in distributed reinforcement learning. In International conference on learning representations, 2018.", + "[30] Youngjin Kim, Wontae Nam, Hyunwoo Kim, Ji-Hoon Kim, and Gunhee Kim. Curiosity-bottleneck: Exploration by distilling task-specific novelty. In International conference on machine learning, pages 3379-3388. PMLR, 2019.", + "[31] Martin Klissarov, Riashat Islam, Khimya Khetarpal, and Doina Precup. Variational state encoding as intrinsic motivation in reinforcement learning. In *Task-Agnostic Reinforcement Learning Workshop at Proceedings of the International Conference on Learning Representations*, pages 16–32, 2019.", + "[32] Lisa Lee, Benjamin Eysenbach, Emilio Parisotto, Eric Xing, Sergey Levine, and Ruslan Salakhutdinov. Efficient exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019.", + "[33] Ian Lenz, Ross A Knepper, and Ashutosh Saxena. Deepmpc: Learning deep latent features for model predictive control. In Robotics: Science and Systems, page 25. Rome, Italy, 2015.", + "[34] Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. Advances in neural information processing systems, 30, 2017.", + "[35] Marlos C Machado, Marc G Bellemare, and Michael Bowling. Count-based exploration with the successor representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5125-5133, 2020.", + "[36] Jarryd Martin, Suraj Narayanan Sasikumar, Tom Everitt, and Marcus Hutter. Count-based exploration in feature space for reinforcement learning. arXiv preprint arXiv:1706.08090, 2017.", + "[37] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021.", + "[38] Alberto Maria Metelli, Amarildo Likmeta, and Marcello Restelli. Propagating uncertainty in reinforcement learning via wasserstein barycenters. Advances in Neural Information Processing Systems, 32, 2019.", + "[39] Vincent Micheli, Eloi Alonso, and François Fleuret. Transformers are sample-efficient world models. In The Eleventh International Conference on Learning Representations, 2023.", + "[40] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016.", + "[41] Ian Osband and Benjamin Van Roy. Bootstrapped thompson sampling and deep exploration. arXiv preprint arXiv:1507.00300, 2015." + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Ian Osband, Charles Blundell, Alexander Pritzel, and Benjamin Van Roy. Deep exploration via bootstrapped dqn. Advances in neural information processing systems, 29, 2016.", + "[43] Ian Osband, Benjamin Van Roy, and Zheng Wen. Generalization and exploration via randomized value functions. In International Conference on Machine Learning, pages 2377-2386. PMLR, 2016.", + "[44] Ian Osband, John Aslanides, and Albin Cassirer. Randomized prior functions for deep reinforcement learning. Advances in Neural Information Processing Systems, 31, 2018.", + "[45] Georg Ostrovski, Marc G Bellemare, Aäron Oord, and Rémi Munos. Count-based exploration with neural density models. In International conference on machine learning, pages 2721-2730. PMLR, 2017.", + "[46] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International conference on machine learning, pages 2778-2787. PMLR, 2017.", + "[47] Deepak Pathak, Dhiraj Gandhi, and Abhinav Gupta. Self-supervised exploration via disagreement. In International conference on machine learning, pages 5062-5071. PMLR, 2019.", + "[48] Mathieu Poliquin. Stable retro, a maintained fork of openai's gym-retro. https://github.com/Farama-Foundation/stable-retro, 2024.", + "[49] Ofir Press, Noah Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. In International Conference on Learning Representations, 2022.", + "[50] Jan Robine, Marc Höftmann, Tobias Uelwer, and Stefan Harmeling. Transformer-based world models are happy with 100k interactions. In The Eleventh International Conference on Learning Representations, 2023.", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.", + "[52] Tom Schaul. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015.", + "[53] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International conference on machine learning, pages 8583-8592. PMLR, 2020.", + "[54] Niranjan Srinivas, Andreas Krause, Sham Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: no regret and experimental design. In Proceedings of the 27th International Conference on International Conference on Machine Learning, pages 1015-1022, 2010.", + "[55] Haoran Tang, Rein Houthooft, Davis Foote, Adam Stooke, OpenAI Xi Chen, Yan Duan, John Schulman, Filip DeTurck, and Pieter Abbeel. # exploration: A study of count-based exploration for deep reinforcement learning. Advances in neural information processing systems, 30, 2017.", + "[56] Ruo Yu Tao, Vincent François-Lavet, and Joelle Pineau. Novelty search in representational space for sample efficient" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27606", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "exploration. Advances in Neural Information Processing Systems, 33:8114-8126, 2020.", + "[57] William R Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3-4):285-294, 1933.", + "[58] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024.", + "[59] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017.", + "[60] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022.", + "[61] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004.", + "[62] Mingxing Xu, Wenrui Dai, Chunmiao Liu, Xing Gao, Weiyao Lin, Guo-Jun Qi, and Hongkai Xiong. Spatial-temporal transformer networks for traffic flow forecasting. arXiv preprint arXiv:2001.02908, 2020.", + "[63] Sherry Yang, Yilun Du, Seyed Kamyar Seyed Ghasemipour, Jonathan Tompson, Leslie Pack Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. In The Twelfth International Conference on Learning Representations, 2023.", + "[64] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024.", + "[65] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023.", + "[66] Lixuan Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Prelar: World model pre-training with learnable action representation. In European Conference on Computer Vision, pages 185-201, 2024.", + "[67] Tianjun Zhang, Paria Rashidinejad, Jiantao Jiao, Yuandong Tian, Joseph E Gonzalez, and Stuart Russell. Made: Exploration via maximizing deviation from explored regions. Advances in Neural Information Processing Systems, 34:9663-9680, 2021." + ], + "bbox": [ + 91, + 92, + 482, + 756 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27607", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_model.json b/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8af788a86484f2b0e9366c85ca1e9bdda6c06587 --- /dev/null +++ b/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_model.json @@ -0,0 +1,2180 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Exploration-Driven Generative Interactive Environments" + }, + { + "type": "image", + "bbox": [ + 0.151, + 0.179, + 0.848, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.284, + 0.327, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.635 + ], + "angle": 0, + "content": "Modern world models require costly and time-consuming collection of large video datasets with action demonstrations by people or by environment-specific agents. To simplify training, we focus on using many virtual environments for inexpensive, automatically collected interaction data. Genie [5], a recent multi-environment world model, demonstrates simulation abilities of many environments with shared behavior. Unfortunately, training their model requires expensive demonstrations. Therefore, we propose a training framework merely using a random agent in virtual environments. While the model trained in this manner exhibits good controls, it is limited by the random exploration possibilities. To address this limitation, we propose AutoExplore Agent - an exploration agent that entirely relies on the uncertainty of the world model, delivering diverse data from which it can learn the best. Our agent is fully independent of environment-specific rewards and thus adapts easily to new environments. With this approach, the pretrained multi-environment model can quickly adapt to new environments achieving video fidelity and controllability improvement." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.636, + 0.484, + 0.773 + ], + "angle": 0, + "content": "In order to obtain automatically large-scale interaction datasets for pretraining, we group environments with similar behavior and controls. To this end, we annotate the behavior and controls of 974 virtual environments - a dataset that we name RetroAct. For building our model, we first create an open implementation of Genie - GenieRedux and apply enhancements and adaptations in our version GenieRedux-G. Our code and data are available at https://github.com/insait-institute/GenieRedux." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Learning from interactive environments allows us to understand and represent the rules, the possible actions, and the consequences that govern them. As an alternative to labori" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.907, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.541, + 0.908, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Our proposed world model training framework. It consists of a pretrained multi-environment world model on random agent data, and a new AutoExplore Agent that explores an environment and delivers diverse data for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.672 + ], + "angle": 0, + "content": "ously hand-coded synthetic simulators, world models have emerged as deep learning tools for realistic environment modeling entirely from observations, commonly images of the observed environment [1, 5, 37, 65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Previous work such as [19, 23, 33] uses light world models to support goal-driven agents with goal-specific state representations. The focus is on coarse future predictions, not on their high visual quality. In contrast, the objective of recent world models is to achieve high-quality future predictions given past observations and actions. Such recent models are able to offer realistic action execution and even real-time interaction with people [1, 58]. This has become possible with the rise of diffusion, transformers [14, 59], and state space models [17], and by borrowing architectural choices from video generation pipelines [51, 60]. Typically, these generative models are designed to closely match a single selected environment. One of the state-of-the-art models, Genie, distinguishes itself by being trained on many visually diverse environments with similar dynamics, thus" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.416, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author. firstname(secondname@insait.ai" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27597" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.415, + 0.106 + ], + "angle": 0, + "content": "demonstrating generalization across new visuals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.291 + ], + "angle": 0, + "content": "Building these high-quality statistical simulators requires diverse observations of the environment as well as of the actions to simulate. Some obtain this data by costly video dataset collection and curation with human demonstrations of the actions [1, 5, 63]. If actions are unavailable, an extra component is designed to predict them, which can introduce uncertainty compared to ground truth labels [5, 37]. Extension to new environments with new types of actions in this setting is difficult as it requires again an expensive data collection process. Others, such as [58] have explored retrieving data with an environment-specific agent, in their case - the game Doom." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.484, + 0.61 + ], + "angle": 0, + "content": "In this work, we propose a framework for accessible and effort-free training of world models in multiple environments. To this end, we first build RetroAct- an annotated and curated large dataset of retro game environments (based on the environments of Stable Retro [48]). We group them based on behavior labels and control descriptions. This grouping allows us to generate large-scale interaction datasets across environments with similar behaviors. Next, we pretrain a multi-environment world model GenieRedux - our open implementation of Genie [5], using a random agent. Unlike [66], which reports the agent's improved behavior from pretraining, we aim to improve the world model. For this, we adapt GenieRedux to virtual environments and implement architectural and training procedure enhancements, resulting in the GenieRedux-G model. We observe that just by training GenieRedux-G on random interactions from subsets of 200 environments and 50 environments with mapped controls, automatically collected from RetroAct, we are able to obtain control behavior (0.450 \\(\\Delta\\)PSNR in 50 environments) and reasonable visual fidelity (26.36 PSNR in 50 environments)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As random actions are limited in their ability to explore the environment, we develop a method to obtain more diverse interaction data to improve the control behavior and visual fidelity of our model. To this end, inspired by [53], we develop our own environment-independent reward function, allowing an agent to explore different environments, entirely without relying on predefined environment rewards. While they aim at a high-performing goal-driven agent, we base our design on improving the underlying large world model for the simulation of environments in terms of higher visual fidelity and improved controllability. For graphical illustrations, see Fig. 1. The objective of our exploration-driven agent is to maximize the world model's uncertainty, estimated by the classification entropy available in the observation prediction stage of GenieRedux-G. Once the diverse data is obtained, we fine-tune GenieRedux-G. We show that this method leads to significant visual (up to 7.4 PSNR) and control (up to \\(1.4\\Delta\\) PSNR) improvements, compared to random agent pretraining." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.753, + 0.106 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.151 + ], + "angle": 0, + "content": "- A framework for training world models with cheap data collection - by training an exploration agent based on our world's model uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- The implementation and release of GenieRedux and GenieRedux-G - open Pytorch models based on [5]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- Architectural and loss changes to the model leading to fidelity improvements, based on our tokenizer representation study." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.904, + 0.258 + ], + "angle": 0, + "content": "- Preparing a large scale environment dataset for multi-environment world model training." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.272, + 0.655, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.537 + ], + "angle": 0, + "content": "World models. Initially built as rough imagination models assisting reinforcement learning (RL) agents [10, 19, 21, 24, 53], world models have evolved into independent realistic video generation models conditioned on actions [9, 39, 50, 64]. They facilitate task-specific agent training by providing predictive representations of the environment. Inspired by [20], Ha and Schmidhuber [18] use a VAE to encode visual observations into latent states, with an MDNRNN predicting future states based on prior states, actions, and VAE outputs to facilitate policy learning. DreamerV2 [22] introduce an RL agent, achieving human-level performance in Atari. It encodes images with a CNN and computes posterior and prior stochastic states using recurrent states. Unlike our work, though, it does not assess the agent's impact on world model improvement nor generalize task rewards across different environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.719 + ], + "angle": 0, + "content": "World models also aim to generate realistic video conditioned on actions [27, 37, 65]. Genie [5] trains a video tokenizer and a Latent Action Model (LAM) for dynamic next-frame generation. GAIA-1 [27] tackle autonomous driving in unstructured settings by encoding multi-modal inputs into a unified representation and predicting image tokens based on prior inputs, using an autoregressive transformer. Menapace et al. [37] employ an encoder-decoder architecture in which the predicted action labels act as a bottleneck, allowing a user to control the generated video by a discrete action. The key gap in these works is automatic data collection, which is addressed in our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Efficient exploration. The importance of efficient exploration in RL is highlighted by [28]. Early methods enhanced exploration by adding noise [16, 34] or using entropy regularization [40], but they have action space limitations and often fail with complex dynamics, where varied actions do not always drive meaningful exploration. A more direct approach uses heterogeneous actors [26, 29, 52] with diverse exploration strategies to enhance environment exploration. Bayesian methods [54, 57] have also been introduced to create acquisition functions for uncertainty-driven exploration [2, 38, 41-44], but often struggle to generalize to high-dimensional inputs like images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27598" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.91, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Method Overview. We propose an alternative to costly human interaction data collection - by exploring environments with an agent. The reward is solely based on the classification uncertainty of our model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.39, + 0.483, + 0.617 + ], + "angle": 0, + "content": "Recent exploration methods emphasize state novelty [3, 7, 11, 35, 36, 45, 55, 67], focusing on encouraging agents to assess novelty only after visiting states. In contrast, our approach, inspired by [7, 47, 53], uses model disagreement to proactively guide agents to states with the highest potential without environment target-driven reward. [6, 46] propose exploration agents driven by uncertainty in state transition and simple feature extractors. Instead, we propose an exploration agent designed to improve a world model that does not model states. Plan2Explore [53] enables agents to seek novel states using a reward that maximizes the state entropy of an RSSM model. While Plan2Explore improves goal-driven agents with their framework, we improve a modern transformer world model with a novel exploration-based reward using token uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.798 + ], + "angle": 0, + "content": "Rather than relying on world models, EX2 [15] learns a classifier to distinguish visited states, providing intrinsic rewards for states that are difficult for the classifier to differentiate. KL-divergence-based approaches [30-32], guide exploration by comparing distributions. For example, SMM [32] computes the KL divergence between the policy-induced state distribution and a uniform target. Tao et al. [56] propose an intrinsic reward based on the distance between a state and its nearest neighbors in a low-dimensional feature space. However, low-dimensionality leads to information loss, restricting full state space exploration — an issue we address by the use of a world model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.262, + 0.83 + ], + "angle": 0, + "content": "3. RetroAct Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We first tackle the problem of accessible training of multi-environment world models by building a framework for cheaply acquiring multi-environment interaction data. In particular, we aim to collect interactions of similar actions" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.385, + 0.911, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.533 + ], + "angle": 0, + "content": "Figure 3. RetroAct Annotation. Description of environments in RetroAct by annotated attribute. Better viewed zoomed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.644 + ], + "angle": 0, + "content": "in many environments. Instead of relying on expensive human interaction, we obtain and curate a collection of virtual environments. As a source, we use the Stable Retro framework by [48], which is a collection of retro games across multiple platforms, with an accompanying starting state. We make no use of the defined rewards. We obtain almost all the supported games (974)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.84 + ], + "angle": 0, + "content": "This raw set contains an environment mix of very different visuals and behaviors. However, in our setting of learning from similar dynamics, it is required to establish correspondence between the environments' behaviors. We perform annotation where for each environment three aspects are classified. The motion style classifies the general style of what and how is moved by the controls, closely relating to game genre; the camera viewpoint; the control axis describing in which direction the player can be moved. The label distributions are shown in Fig. 3. In Tab. 1 we compare our RetroAct with other related datasets. RetroAct distinguishes itself by providing behavior and control annotations, while maintaining a high number of environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "It is discovered that the most prevalent type of environment in our set is the platformer - 483 titles. As the largest subset, we filter only these games for further use, as it is required to have many environments exhibiting similar con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27599" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.331, + 0.09, + 0.667, + 0.103 + ], + "angle": 0, + "content": "Table 1. Comparison of RetroAct dataset to others." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.104, + 0.875, + 0.212 + ], + "angle": 0, + "content": "
DatasetType#EnvironmentsDiverse BehaviorsOpenBehavior AnnotationControl Annotation
Coinrun [13]Environments1X✓XX
ALE [4]Environments57✓✓XX
Stable Retro [48]Environments1003✓✓XX
Platformers [5]VideosUnknown✓XXX
RetroAct(Ours)Environments974✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.219, + 0.482, + 0.323 + ], + "angle": 0, + "content": "Controls. Five motion actions are defined for our model - moving left, right, up, down and jump. Each game has its own mapping of buttons to actions. Therefore, we generate a short clip of each of the 5 selected actions for each of the 483 titles and build an annotation tool to observe and annotate the executed action. Eventually, we annotated 2,925 behavior and 2,898 control labels." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.483, + 0.431 + ], + "angle": 0, + "content": "After experimenting, we observed that models require more training with a higher number of environments, so we defined two subsets of our large set to handle computational cost: a subset consisting of the first 200 games of 483 behavior-filtered games for pretraining, and another subset of 50 randomly selected action-consistent games using RetroAct's action labels for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.483, + 0.612 + ], + "angle": 0, + "content": "We collect a large scale dataset by launching a random agent in all of the environments, collecting actions and observations. From the 200-game set we build Platformers-200 - a dataset with 10,000 episodes (50 episodes per game) with 500 frames each at most, resulting in 4.6mln images. From the 50-game set we obtain Platformers-50 - 5000 episodes (100 per game) of length at most 1000, resulting in 4.8mln images. In our protocol, we take \\(1\\%\\) of the sessions of each environment as a validation set. We show that using a random agent is already sufficient to learn a level of controllability and later build on top with an exploration agent of our design." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.673 + ], + "angle": 0, + "content": "To validate our GenieRedux implementation, we implement the CoinRun case study in [5]. Using the protocol from above, we obtain a dataset of 10k episodes with a maximum length of 500, resulting in 4mln images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.688, + 0.396, + 0.705 + ], + "angle": 0, + "content": "4. Multi-Environment World Model" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.714, + 0.483, + 0.805 + ], + "angle": 0, + "content": "Given virtual environments, our first goal is to automatically obtain a dataset of image sequences \\( I_{1}, \\ldots, I_{N} \\) and corresponding actions \\( a_{1}, \\ldots, a_{N-1} \\). Given a sequence \\( I_{1}, \\ldots, I_{N} \\) and past and future actions \\( a_{1}, \\ldots, a_{N+T-1} \\), our world model aims to predict the future \\( T \\) frames \\( I_{N+1}, \\ldots, I_{T} \\), corresponding to the actions executed." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "GenieRedux. As Genie [5] is not made available by the authors, we create an open implementation and call it GenieRedux. We validate our implementation quantitatively and qualitatively in Sec. 5 and Sup.Mat F. It consists of three components. A video" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.905, + 0.4 + ], + "angle": 0, + "content": "Tokenizer encodes input frame sequences into spatiotemporal tokens: \\( e_1, \\ldots, e_N = T_{enc}(I_1, \\ldots, I_N) \\), and decodes back to images: \\( I_1, \\ldots, I_N = T_{dec}(e_1, \\ldots, e_N) \\). A Latent Action Model encodes input frame sequences into spatio-temporal tokens: \\( a_1, \\ldots, a_{N-1} = LAM_{enc}(I_1, \\ldots, I_{N-1}) \\), and decodes them to reconstruct future prediction \\( I_2, \\ldots, I_N = LAM_{dec}(a_1, \\ldots, a_{N-1}) \\). A Dynamics module predicts the next frames based on partially masked frame tokens and actions: \\( I_2, \\ldots, I_{N+T-1} = D(e_1, \\ldots, e_N, \\ldots, e_{N+T-1}; a_1, \\ldots, a_{N+T-1}) \\), where in inference \\( e_N, \\ldots, e_{N+T-1} \\) are masked. We adhere closely to Genie's specifications for implementing these components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.905, + 0.462 + ], + "angle": 0, + "content": "All components use the causal Spatial Temporal Transformer (STTN) [62]. We use Position Encoding Generator (PEG) [12] for spatial and temporal attention, and Attention with Linear Biases (ALiBi) [49] for temporal attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.905, + 0.538 + ], + "angle": 0, + "content": "We train our models with a sequence size of 16 frames and resolution of \\(64 \\times 64\\) to address computational limitations. We train a U-Net-based superresolution network on \\(50 \\mathrm{~K}\\) data samples to upscale the output to \\(256 \\times 256\\). (Sup.Mat. B)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.906, + 0.698 + ], + "angle": 0, + "content": "GenieRedux-G. Building upon the base model, we offer a variant - GenieRedux-G, which is adapted to virtual environments and contains architectural and training improvements. While GenieRedux uses an indispensable LAM model to obtain the actions, we discard it, as ground truth actions are available from our agent. Instead, the one-hot actions are concatenated to each layer of the Dynamics module for conditioning. In this way, we avoid the uncertainty of a prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.906, + 0.865 + ], + "angle": 0, + "content": "The Dynamics module consists of an ST-ViViT encoder, followed by a MaskGIT architecture [8], which predicts indices from the tokenizer's codebook for randomly masked input tokens during training, according to a schedule. As standard cross-entropy is used, token classification has the drawback to penalize equally any prediction different from the ground truth. However, close tokens in the codebook result in significantly fewer changes than far tokens, as also shown in Sec. 5. To enable this concept of a distance between tokens in the classification of \\( N_{E} \\) tokens, we design a Token Distance Cross-Entropy (TDCE) Loss:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.88, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\nT D C E (x, y) = \\left(y ^ {T} K\\right) \\cdot \\text {s o f t m a x} (x) + C E (x, y) \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27600" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.182 + ], + "angle": 0, + "content": "Here \\( x \\in \\mathcal{R}^{N_E} \\) is the prediction logits, \\( y \\in \\mathcal{R}^{N_E} \\) is the ground truth one-hot class. \\( K \\in \\mathcal{R}^{N_E \\times N_E} \\) is a precomputed table at the start of training of the cosine distances between all tokens; \\( CE(.) \\) denotes standard Cross-Entropy Loss. When an incorrect token class is given probability, it is penalized based on its distance to the ground truth class." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.304 + ], + "angle": 0, + "content": "MaskGIT's design is to take as input learnable embeddings, indexed by the tokens predicted by the Tokenizer. They are randomly initialized, and therefore contain none of the content of the tokens. Given that the encoding itself and the distance between tokens can contribute to Dynamic module's performance, we add a skip connection by adding the embedding to the token itself, which improves visual fidelity and controllability of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.482, + 0.429 + ], + "angle": 0, + "content": "AutoExplore Agent We extend our framework with an exploration agent that obtains data by going deeper into the environments. We name it AutoExplore Agent. The reward of the agent is entirely based on the world model performance and operates without any environment rewards. Therefore, it can be trained in various environments without tuning to their specifics or relying on a reward definition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.482, + 0.611 + ], + "angle": 0, + "content": "The design of our reward is based on the fact that GenieRedux-G employs classification for token prediction. Each token is predicted by sampling from a categorical distribution over the codebook. We first obtain all \\( N_T \\) token prediction distributions by running GenieRedux-G-50 5 steps back from the current observation \\( I_c \\) for which we want to estimate the reward. We provide 2 images \\( I_{c - 4}, I_{c - 3} \\), predict 3 images \\( -I_{c - 2}, \\dots, I_c \\), and take the distributions of the predicted tokens of \\( I_c \\) to obtain \\( x = [x_1, \\dots, x_t, \\dots, x_{N_T}] \\). We evaluate the uncertainty per predicted token \\( u_t \\) by calculating the entropy over the categorical distribution and normalize it in the range [0, 2]:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.621, + 0.482, + 0.656 + ], + "angle": 0, + "content": "\\[\nu _ {t} = \\frac {2 \\cdot \\sum_ {i} ^ {N _ {T}} x _ {i} \\cdot \\log \\left(x _ {i}\\right)}{N _ {e}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.787 + ], + "angle": 0, + "content": "Studying the properties of the Tokenizer representation, we find that a prevalent token is learned representing static parts of the environment. Only the changing parts generate high uncertainty and, therefore, we take the subset \\( S_{top} \\) of \\( 25\\% \\) highest uncertainties of the entire set of uncertainties \\( S = \\{u_t\\} \\). The reward, shown in Eq. 3, establishes the agent's goal to collect data that maximizes uncertainty of the world model." + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.802, + 0.482, + 0.819 + ], + "angle": 0, + "content": "\\[\nS _ {25 \\%} = \\{u \\in S \\mid u \\geq Q _ {75} (S) \\} \\tag{3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.822, + 0.482, + 0.859 + ], + "angle": 0, + "content": "\\[\nR \\left(I _ {c}\\right) = \\frac {1}{\\left| S _ {25 \\%} \\right|} \\sum_ {u \\in S _ {25 \\%}} u \\tag{4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our agent is an actor-critic, trained with the Policy Gradient method. For the agent architecture, we follow [39]. It" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "consists of a CNN encoder followed by an LSTM. As standard in RL, 4 frames are stacked, max-pooled, and the result is the input to the agent for a single time step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.905, + 0.244 + ], + "angle": 0, + "content": "Exploration-driven World Model Training. We initially pretrain GenieRedux-G on Platformers-200 and fine-tune on Platformers-50 to obtain the model GenieRedux-G-50. Then, we train AutoExplore Agent by using GenieRedux-G-50, using it as a source of reward. The details of training the agent are presented in Sup.Mat A.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.905, + 0.41 + ], + "angle": 0, + "content": "Running the trained exploratory agent on a selected environment, we obtain a new diverse dataset with action demonstrations under unseen scenes. We first fine-tune the decoder of the Tokenizer for 1,000 iterations to adapt to the new unseen scenes. The Dynamic module of GenieRedux-G is then fine-tuned on the new data to achieve greater visual fidelity and controllability under new conditions. In order to build test sets to evaluate our approach, we train an Agent-57 model for each of the environments we explored, using the available environment rewards. More details on the test setup are provided in Sup.Mat A.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.904, + 0.562 + ], + "angle": 0, + "content": "For visual fidelity evaluation, we use FID (Frechet inception distance) Heusel et al. [25], PSNR (signal-to-noise ratio) and SSIM (structural similarity index measure) Wang et al. [61]. To evaluate controllability, we use the recently proposed \\(\\Delta_t\\) PSNR metric [5], which compares the visual effect of the ground truth action \\((\\hat{x}_t)\\) versus a random action \\((\\hat{x}_t')\\): \\(\\Delta_t\\) PSNR = PSNR \\((x_t, \\hat{x}_t) - \\mathrm{PSNR}(x_t, \\hat{x}_t')\\), where \\(x_t\\) is the ground truth frame at time \\(t\\). A higher \\(\\Delta_t\\) PSNR indicates a higher level of controllability. As in Bruce et al. [5], for all experiments we report \\(\\Delta_t\\) PSNR with \\(t = 4\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.645, + 0.59 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Comparing GenieRedux and GenieRedux-G. We implement the original CoinRun case study with a random agent, as advised by [5], in order to validate and compare GenieRedux with LAM, and GenieRedux-G which uses agent-provided actions instead. In this study, the presence of LAM is the only difference between the models. We first train on a dataset, collected by a random agent. Visual fidelity results are in Tab. 2. Our GenieRedux implementation exhibits high visual quality and matches all seven CoinRun environment actions, as well as progressing environment motions (demonstrated in Sup.Mat. F). However, as demonstrated by the metrics, GenieRedux-G shows superior visual fidelity and controllability (more in Sup.Mat. F), as it avoids the uncertainty of LAM prediction. This study demonstrates that even using a random agent can result in action performance abilities in the world model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Next, we train an actor-critic agent with PPO on the environment reward, following [13] to collect data and train GenieRedux-TA and GenieRedux-G-TA. Tab. 3 shows evaluation on a test set collected by a trained agent." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27601" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.131 + ], + "angle": 0, + "content": "Table 2. Comparison of GenieRedux and GenieRedux-G on Basic Test Set. Performed on a test set, collected from the Coinrun environment with randomly sampled actions." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.132, + 0.445, + 0.247 + ], + "angle": 0, + "content": "
ModelBasic Test Set
FID↓PSNR↑SSIM↑
Tokenizer18.1438.250.96
LAM37.0133.970.92
GenieRedux21.8825.510.77
GenieRedux-G18.8833.410.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.316 + ], + "angle": 0, + "content": "Table 3. Comparison of GenieRedux and GenieRedux-G on Diverse Test Set. The models are trained with data collected by random agent and trained agent (-TA), and tested on data collected by a trained agent from the Coinrun environment." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.318, + 0.457, + 0.455 + ], + "angle": 0, + "content": "
ModelDiverse Test Set
FID↓PSNR↑SSIM↑
Tokenizer19.1335.850.94
Tokenizer-TA11.6340.620.97
GenieRedux23.9723.820.73
GenieRedux-G19.5131.660.90
GenieRedux-TA12.5731.970.90
GenieRedux-G-TA12.4034.440.92
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.555 + ], + "angle": 0, + "content": "GenieRedux-G outperforms GenieRedux on all settings. Furthermore, models trained on diverse agent-collected data are visually superior to those trained on random agents. The higher \\(\\Delta\\)PSNR of 1.89 for GenieRedux-G-TA compared to 0.70 for GenieRedux-G shows the superiority of diverse data training in controllability. (more in Sup.Mat. F)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.803 + ], + "angle": 0, + "content": "Multi-Environment Models. Here, we evaluate the models we initially train on many environments from RetroAct. GenieRedux-G-200 is pretrained on the Platformers-200 dataset for 180k iterations. On the validation set, we obtain 23.32 PSNR and 17.12 FID. Using this model as a base, GenieRedux-G-50 is trained on Platformers-50. Its quantitative evaluation on a test set of 10k sessions separately generated from the selected 50 environments is at the start of Tab. 4. As the 50 environments are selected with corresponding action controls between each other, we see a boost in the quality of prediction. Fig. 4 demonstrates that the instructed action is executed successfully by GenieRedux-G. As the up action is rarely used, it serves more as a no-operation action. (more in Sup.Mat C.1)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Ablation Study. In this experiment we evaluate the additive gain of each proposed improvement in GenieRedux-G - the additive token input and training with the Token Distance Cross-Entropy Loss. The ablation is performed on a generated test set of 10k sessions, each 500 frames long." + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.089, + 0.898, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation study on improvements in GenieRedux-G." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.108, + 0.882, + 0.203 + ], + "angle": 0, + "content": "
ModelFID↓PSNR↑SSIM↑
GenieRedux-G-20022.3125.110.80
GenieRedux-G-5023.8026.360.84
+ Token Input22.9626.650.84
+ TDCE Loss22.9527.060.85
Autoregressive22.1128.070.88
" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.21, + 0.892, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.368, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Figure 4. Control Of GenieRedux-G-50. Demonstrating all controls of our multi-environment model on multiple games." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.907, + 0.497 + ], + "angle": 0, + "content": "The data is collected using a random action policy from the environments in Platformers-50. Visual fidelity evaluation is provided in Tab. 4. It can be seen that each component gives our model a benefit in terms of visual fidelity. Finally, we perform an autoregressive evaluation of the best model to achieve our highest score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Tokenizer Representation Study. This experiment provides insights into the inner workings of GenieRedux-G to motivate our proposed changes. As the Dynamics module operates entirely on the token representation, we examine it closely. Fig. 5 shows the reconstructions of an input sequence (first row) and the visualized token representation (last row), where each predicted token index is assigned a different color. The visual features of the first frame are captured by various tokens. Starting with the second frame, the representation drastically changes - a token is specialized in representing the static frame regions compared to the past, while all motion regions are updated with new content. Observing that visually similar patches predict identical or similar tokens, we replace each predicted token with its closest in the codebook. We only keep the special background token unchanged. In the second row of Fig. 5 we show the resulting reconstruction - while some blurriness appears, the image remains largely the same. Conversely, replacing each token with its furthest in the codebook (third row) results in a significantly different image. This property - closer tokens having more similar appearance - motivates our Token Distance Cross-Entropy Loss, which penalizes predicting tokens further away from the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Fig. 6 visualizes the uncertainty of GenieRedux-G-50 for each predicted token of its Dynamics module given a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27602" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 5. Quantitative Results on 3 environments. We evaluate the benefit of the data from the propose AutoExplore Agent to our models. GenieRedux-G-50 is our pretrained world model on 50 environments. GenieRedux-G-50-ft are fine-tuned models using data from a random agent or AutoExplore (Exploration). GenieRedux-G denotes a non-fine-tuned model, trained with the exploration data." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.136, + 0.891, + 0.475 + ], + "angle": 0, + "content": "
EnvironmentStrategyModelFID↓PSNR↑SSIM↑ΔPSNR↑
Adventure Island IIRandomGenieRedux-G-5041.9926.320.810.83
GenieRedux-G-50-ft42.3427.040.811.19
ExplorationTokenizer-ft11.0138.950.98-
GenieRedux-G11.9428.330.880.37
GenieRedux-G-50-ft12.7730.600.901.47
Random AutoregressiveGenieRedux-G-50-ft41.5527.820.831.24
Exploration AutoregressiveGenieRedux-G-50-ft11.3333.610.942.09
Super Mario BrosRandomGenieRedux-G-5029.8334.240.940.56
GenieRedux-G-50-ft30.1334.540.940.54
ExplorationTokenizer8.0942.000.99-
GenieRedux-G9.5634.000.950.09
GenieRedux-G-50-ft9.5536.130.970.57
Random AutoregressiveGenieRedux-G-50-ft30.8434.850.950.57
Exploration AutoregressiveGenieRedux-G-50-ft9.3337.770.970.76
SmurfsRandomGenieRedux-G-5079.5121.470.690.47
GenieRedux-G-50-ft80.6121.830.700.65
ExplorationTokenizer17.8635.610.98-
GenieRedux-G20.4335.420.800.85
GenieRedux-G-50-ft20.0127.450.851.55
Random AutoregressiveGenieRedux-G-50-ft80.1622.160.710.69
Exploration AutoregressiveGenieRedux-G-50-ft18.9729.530.902.06
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.485, + 0.472, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Figure 5. Tokenizer Representation. Reconstruction images from the tokenizer, and the effect of replacing each token with its closest and furthest in the codebook. Lastly, we visualize the indices of the predicted tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.821 + ], + "angle": 0, + "content": "sequence. The uncertainty metric is the entropy of the classification over 1024 codebook tokens. Tokens corresponding to motion have the highest uncertainty; other regions are mostly classified as the \"static\" token. Thus, minimal character movement yields low uncertainty, while forward motion increases it. This motivates us to build AutoExplore Agent's reward based on this uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Exploration-based training. We demonstrate our exploration-based training of GenieRedux-G. We perform the procedure on 3 environments - AdventureIslandII, which provides an easy set-" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.485, + 0.895, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.572, + 0.907, + 0.614 + ], + "angle": 0, + "content": "Figure 6. Dynamics Uncertainty. Shown is the uncertainty per token predicted for each image of an example sequence. Uncertainty is generated in the regions of motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.748 + ], + "angle": 0, + "content": "ting for the agent to learn (single platform with no enemies at the start), SuperMario Bros provides an enemy and obstacles soon after the start and Smurfs provides a more complex background imagery and different action dynamics. For each of the environments, we train an AutoExplorer Agent. We observe that the agent learns to move forward and navigate obstacles to maximize reward. (more in Sup.Mat. D)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We use our pretrained GenieRedux-G-50 model as a baseline and fine-tune it for each environment in two settings - a dataset collected on the selected environment by a random agent and by our AutoExplorer Agent. Each dataset consists of 10k sessions, each 700 frames long. We fine-tune (GenieRedux-G-50-ft) for 10k iterations and pick the best performing model. In our comparison, we also include a GenieRedux-G model trained from scratch on the diverse exploration datasets for 15k iterations to show the effect of pretraining. We perform single-pass generation for all mod" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.482, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.21, + 0.483, + 0.252 + ], + "angle": 0, + "content": "Figure 7. AutoExplore Agent vs Random Agent Qualitative Comparison. We show that AutoExplore exhibits better visual quality and avoids losing track of the agent." + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.26, + 0.463, + 0.273 + ], + "angle": 0, + "content": "Table 6. Comparison of AutoExplore Agent with others." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.274, + 0.48, + 0.356 + ], + "angle": 0, + "content": "
AgentSuperMarioBrosAdventureIslandII
PSNR↓SSIM↓ΔPSNR↓PSNR↓SSIM↓ΔPSNR↓
RF28.580.940.18124.820.780.44
VAE24.400.860.08716.570.50.072
Ours23.810.850.06515.200.410.070
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.365, + 0.483, + 0.606 + ], + "angle": 0, + "content": "els and the more computationally heavy autoregressive evaluation for the fine-tuned models on data from random and AutoExplore Agent's datasets. Tab. 5 shows visual fidelity and controllability metrics for each environment, confirming the effectiveness of our exploration method. The model fine-tuned on AutoExplore Agent's data consistently outperforms the models trained on random actions in terms of visual fidelity. Exploration-based fine-tuning also improves controllability. Environments with small characters and uniform backgrounds can be more challenging for all models to learn. However, the gain in controllability in this case remains noticeable during autoregressive evaluation. Fig. 7 demonstrates the superior quality of our method. In addition, we observe that the multienvironment pretraining leads to significant gains in both studied aspects compared to the nonpretrained model. (more in Sup.Mat. C)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.732 + ], + "angle": 0, + "content": "AutoExplore Agent Evaluation. We compare AutoExplore Agent with exploration-based methods in [6]. We train agents based on SSE of RF and VAE features on top of GenieRedux and compare with ours on Tab. 6. AutoExplore Agent's reward results in maximum world model visual and controllability errors (on 1k episodes of agent actions), fulfilling its intended role in our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "User Studies. To validate the quality of our final results, we perform a user study in which we ask people to rate from 1 to 5 the quality of samples produced respectively by GenieRedux-G trained on random agent's data and on AutoExplore Agent's data. Each sample in our study consists of two 16-frame clips playing in a synchronized manner - the ground truth clip and our GenieRedux-G-50-ft reconstruction, given two initial frames and generating the rest autoregressively. We provide a total of 120 samples to the users - 40 samples per model and 40 samples of two ground" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.096, + 0.703, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.096, + 0.898, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.224, + 0.925, + 0.266 + ], + "angle": 0, + "content": "Figure 8. User study results. Our user study on two games shows that our model trained with AutoExplore Agent's data is consistently rated higher." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.378 + ], + "angle": 0, + "content": "truth samples, to establish scale. We give 20 samples from each of the two selected games - SuperMario Bros and AdventureIslandII. We get reviews from 19 participants. The results are shown in Fig. 8. The model, trained on data from AutoExplore Agent is clearly rated closer to the ground truth, establishing the quality of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.489 + ], + "angle": 0, + "content": "With a second user study, we evaluate the action accuracy of the generated frames. We use ambiguous single input cases (character starting mid-air) and generate 60 clips with 3 actions on AdventureIslandII. Users prefer our exploration-trained model, rating it \\( \\mathbf{0.75} \\pm \\mathbf{0.019} \\) on a scale from 0 (random preferred) to 1 (exploration preferred). (more in Sup.Mat. E.2)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.518, + 0.634, + 0.533 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.64 + ], + "angle": 0, + "content": "As world models have developed into large models with impressive simulation properties, they require large interaction datasets, complete with diverse observations and actions. Genie [5] demonstrates impressive abilities by training on multiple environments, however, requiring the collection of a large video dataset and a model to infer actions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we address the heavy burden of data collection and curation by building a new framework for training large world models by collecting interaction data from a large number of virtual environments. We first build an open implementation of Genie - GenieRedux and enhance it into its version GenieRedux-G. We obtain models exhibiting control by pretraining on a large set of virtual environments. We address the overfitting limitations of random data collection policy by proposing AutoExplore Agent, an agent entirely independent of the environment reward, maximizing the uncertainty of GenieRedux-G. After finetuning on the explored environment, our model is able to improve its visual fidelity and controllability much better than training solely on random agent's data. Demonstrating this on multiple environments, we show the potential of our framework to make training of next-generation world models more accessible, cost-effective, and effort-free." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27604" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "INSAIT, Sofia University \"St. Kliment Ohridski\". Partially funded by the Ministry of Education and Science of Bulgaria's support for INSAIT as part of the Bulgarian National Roadmap for Research Infrastructure. This project was supported with computational resources provided by Google Cloud Platform (GCP)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[1] Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37: 58757-58791, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Kamyar Azizzadenesheli, Emma Brunskill, and Animashree Anandkumar. Efficient exploration through bayesian deep q-networks. In 2018 Information Theory and Applications Workshop (ITA), pages 1-9. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.483, + 0.43 + ], + "angle": 0, + "content": "[3] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. Advances in neural information processing systems, 29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[4] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253-279, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[6] Yuri Burda, Harri Edwards, Deepak Pathak, Amos Storkey, Trevor Darrell, and Alexei A Efros. Large-scale study of curiosity-driven learning. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[7] Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[9] Chang Chen, Yi-Fu Wu, Jaesik Yoon, and Sungjin Ahn. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Silvia Chiappa, Sébastien Racanière, Daan Wierstra, and Shakir Mohamed. Recurrent environment simulators. In International Conference on Learning Representations, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Leshem Choshen, Lior Fox, and Yonatan Loewenstein. Dora the explorer: Directed outreach reinforcement action-selection. arXiv preprint arXiv:1804.04012, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In The Eleventh International Conference on Learning Representations, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[13] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In International conference on machine learning, pages 1282-1289. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[15] Justin Fu, John Co-Reyes, and Sergey Levine. Ex2: Exploration with exemplar models for deep reinforcement learning. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[16] Scott Fujimoto, Herke Hoof, and David Meger. Addressing function approximation error in actor-critic methods. In International conference on machine learning, pages 1587-1596. PMLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[18] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[19] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[20] David Ha, Jonas Jongejan, and Ian Johnson. Draw together with a neural network. Retrieved Oct, 5:2021, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[21] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[22] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[23] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[26] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27605" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In Proceedings of the Nineteenth International Conference on Machine Learning, pages 267-274, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[29] Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent experience replay in distributed reinforcement learning. In International conference on learning representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[30] Youngjin Kim, Wontae Nam, Hyunwoo Kim, Ji-Hoon Kim, and Gunhee Kim. Curiosity-bottleneck: Exploration by distilling task-specific novelty. In International conference on machine learning, pages 3379-3388. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[31] Martin Klissarov, Riashat Islam, Khimya Khetarpal, and Doina Precup. Variational state encoding as intrinsic motivation in reinforcement learning. In *Task-Agnostic Reinforcement Learning Workshop at Proceedings of the International Conference on Learning Representations*, pages 16–32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[32] Lisa Lee, Benjamin Eysenbach, Emilio Parisotto, Eric Xing, Sergey Levine, and Ruslan Salakhutdinov. Efficient exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[33] Ian Lenz, Ross A Knepper, and Ashutosh Saxena. Deepmpc: Learning deep latent features for model predictive control. In Robotics: Science and Systems, page 25. Rome, Italy, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[34] Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[35] Marlos C Machado, Marc G Bellemare, and Michael Bowling. Count-based exploration with the successor representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5125-5133, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[36] Jarryd Martin, Suraj Narayanan Sasikumar, Tom Everitt, and Marcus Hutter. Count-based exploration in feature space for reinforcement learning. arXiv preprint arXiv:1706.08090, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[37] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[38] Alberto Maria Metelli, Amarildo Likmeta, and Marcello Restelli. Propagating uncertainty in reinforcement learning via wasserstein barycenters. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[39] Vincent Micheli, Eloi Alonso, and François Fleuret. Transformers are sample-efficient world models. In The Eleventh International Conference on Learning Representations, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[40] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Ian Osband and Benjamin Van Roy. Bootstrapped thompson sampling and deep exploration. arXiv preprint arXiv:1507.00300, 2015." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[42] Ian Osband, Charles Blundell, Alexander Pritzel, and Benjamin Van Roy. Deep exploration via bootstrapped dqn. Advances in neural information processing systems, 29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[43] Ian Osband, Benjamin Van Roy, and Zheng Wen. Generalization and exploration via randomized value functions. In International Conference on Machine Learning, pages 2377-2386. PMLR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[44] Ian Osband, John Aslanides, and Albin Cassirer. Randomized prior functions for deep reinforcement learning. Advances in Neural Information Processing Systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[45] Georg Ostrovski, Marc G Bellemare, Aäron Oord, and Rémi Munos. Count-based exploration with neural density models. In International conference on machine learning, pages 2721-2730. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[46] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International conference on machine learning, pages 2778-2787. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[47] Deepak Pathak, Dhiraj Gandhi, and Abhinav Gupta. Self-supervised exploration via disagreement. In International conference on machine learning, pages 5062-5071. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[48] Mathieu Poliquin. Stable retro, a maintained fork of openai's gym-retro. https://github.com/Farama-Foundation/stable-retro, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[49] Ofir Press, Noah Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[50] Jan Robine, Marc Höftmann, Tobias Uelwer, and Stefan Harmeling. Transformer-based world models are happy with 100k interactions. In The Eleventh International Conference on Learning Representations, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[52] Tom Schaul. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[53] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International conference on machine learning, pages 8583-8592. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[54] Niranjan Srinivas, Andreas Krause, Sham Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: no regret and experimental design. In Proceedings of the 27th International Conference on International Conference on Machine Learning, pages 1015-1022, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[55] Haoran Tang, Rein Houthooft, Davis Foote, Adam Stooke, OpenAI Xi Chen, Yan Duan, John Schulman, Filip DeTurck, and Pieter Abbeel. # exploration: A study of count-based exploration for deep reinforcement learning. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Ruo Yu Tao, Vincent François-Lavet, and Joelle Pineau. Novelty search in representational space for sample efficient" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27606" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "exploration. Advances in Neural Information Processing Systems, 33:8114-8126, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[57] William R Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3-4):285-294, 1933." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[58] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.235 + ], + "angle": 0, + "content": "[59] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[60] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[61] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[62] Mingxing Xu, Wenrui Dai, Chunmiao Liu, Xing Gao, Weiyao Lin, Guo-Jun Qi, and Hongkai Xiong. Spatial-temporal transformer networks for traffic flow forecasting. arXiv preprint arXiv:2001.02908, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[63] Sherry Yang, Yilun Du, Seyed Kamyar Seyed Ghasemipour, Jonathan Tompson, Leslie Pack Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. In The Twelfth International Conference on Learning Representations, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[64] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[65] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[66] Lixuan Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Prelar: World model pre-training with learnable action representation. In European Conference on Computer Vision, pages 185-201, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[67] Tianjun Zhang, Paria Rashidinejad, Jiantao Jiao, Yuandong Tian, Joseph E Gonzalez, and Stuart Russell. Made: Exploration via maximizing deviation from explored regions. Advances in Neural Information Processing Systems, 34:9663-9680, 2021." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27607" + } + ] +] \ No newline at end of file diff --git a/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_origin.pdf b/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7d18d8b2f9aa9f46aac5282626019a20475357d4 --- /dev/null +++ b/2025/Exploration-Driven Generative Interactive Environments/4c28c31d-7f28-4419-9de8-e3891fd9cbcc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a845d765766f2862166be95ecd8ecac3f31691c5598bb844a5c8844da17957fd +size 7748028 diff --git a/2025/Exploration-Driven Generative Interactive Environments/full.md b/2025/Exploration-Driven Generative Interactive Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..632d7e37365b801d094f364a00f4f94c0a2381ad --- /dev/null +++ b/2025/Exploration-Driven Generative Interactive Environments/full.md @@ -0,0 +1,285 @@ +# Exploration-Driven Generative Interactive Environments + +![](images/523a425dc0fbd2aff438c9045131f881214fbf172b51a5b1b61e150d5269b881.jpg) + +# Abstract + +Modern world models require costly and time-consuming collection of large video datasets with action demonstrations by people or by environment-specific agents. To simplify training, we focus on using many virtual environments for inexpensive, automatically collected interaction data. Genie [5], a recent multi-environment world model, demonstrates simulation abilities of many environments with shared behavior. Unfortunately, training their model requires expensive demonstrations. Therefore, we propose a training framework merely using a random agent in virtual environments. While the model trained in this manner exhibits good controls, it is limited by the random exploration possibilities. To address this limitation, we propose AutoExplore Agent - an exploration agent that entirely relies on the uncertainty of the world model, delivering diverse data from which it can learn the best. Our agent is fully independent of environment-specific rewards and thus adapts easily to new environments. With this approach, the pretrained multi-environment model can quickly adapt to new environments achieving video fidelity and controllability improvement. + +In order to obtain automatically large-scale interaction datasets for pretraining, we group environments with similar behavior and controls. To this end, we annotate the behavior and controls of 974 virtual environments - a dataset that we name RetroAct. For building our model, we first create an open implementation of Genie - GenieRedux and apply enhancements and adaptations in our version GenieRedux-G. Our code and data are available at https://github.com/insait-institute/GenieRedux. + +# 1. Introduction + +Learning from interactive environments allows us to understand and represent the rules, the possible actions, and the consequences that govern them. As an alternative to labori + +![](images/d92641e2d123750a2b97de7fb1b311ddc86da408832ae75ce35a1d568f602a70.jpg) +Figure 1. Our proposed world model training framework. It consists of a pretrained multi-environment world model on random agent data, and a new AutoExplore Agent that explores an environment and delivers diverse data for fine-tuning. + +ously hand-coded synthetic simulators, world models have emerged as deep learning tools for realistic environment modeling entirely from observations, commonly images of the observed environment [1, 5, 37, 65]. + +Previous work such as [19, 23, 33] uses light world models to support goal-driven agents with goal-specific state representations. The focus is on coarse future predictions, not on their high visual quality. In contrast, the objective of recent world models is to achieve high-quality future predictions given past observations and actions. Such recent models are able to offer realistic action execution and even real-time interaction with people [1, 58]. This has become possible with the rise of diffusion, transformers [14, 59], and state space models [17], and by borrowing architectural choices from video generation pipelines [51, 60]. Typically, these generative models are designed to closely match a single selected environment. One of the state-of-the-art models, Genie, distinguishes itself by being trained on many visually diverse environments with similar dynamics, thus + +demonstrating generalization across new visuals. + +Building these high-quality statistical simulators requires diverse observations of the environment as well as of the actions to simulate. Some obtain this data by costly video dataset collection and curation with human demonstrations of the actions [1, 5, 63]. If actions are unavailable, an extra component is designed to predict them, which can introduce uncertainty compared to ground truth labels [5, 37]. Extension to new environments with new types of actions in this setting is difficult as it requires again an expensive data collection process. Others, such as [58] have explored retrieving data with an environment-specific agent, in their case - the game Doom. + +In this work, we propose a framework for accessible and effort-free training of world models in multiple environments. To this end, we first build RetroAct- an annotated and curated large dataset of retro game environments (based on the environments of Stable Retro [48]). We group them based on behavior labels and control descriptions. This grouping allows us to generate large-scale interaction datasets across environments with similar behaviors. Next, we pretrain a multi-environment world model GenieRedux - our open implementation of Genie [5], using a random agent. Unlike [66], which reports the agent's improved behavior from pretraining, we aim to improve the world model. For this, we adapt GenieRedux to virtual environments and implement architectural and training procedure enhancements, resulting in the GenieRedux-G model. We observe that just by training GenieRedux-G on random interactions from subsets of 200 environments and 50 environments with mapped controls, automatically collected from RetroAct, we are able to obtain control behavior (0.450 $\Delta$ PSNR in 50 environments) and reasonable visual fidelity (26.36 PSNR in 50 environments). + +As random actions are limited in their ability to explore the environment, we develop a method to obtain more diverse interaction data to improve the control behavior and visual fidelity of our model. To this end, inspired by [53], we develop our own environment-independent reward function, allowing an agent to explore different environments, entirely without relying on predefined environment rewards. While they aim at a high-performing goal-driven agent, we base our design on improving the underlying large world model for the simulation of environments in terms of higher visual fidelity and improved controllability. For graphical illustrations, see Fig. 1. The objective of our exploration-driven agent is to maximize the world model's uncertainty, estimated by the classification entropy available in the observation prediction stage of GenieRedux-G. Once the diverse data is obtained, we fine-tune GenieRedux-G. We show that this method leads to significant visual (up to 7.4 PSNR) and control (up to $1.4\Delta$ PSNR) improvements, compared to random agent pretraining. + +Our contributions are as follows: + +- A framework for training world models with cheap data collection - by training an exploration agent based on our world's model uncertainty. +- The implementation and release of GenieRedux and GenieRedux-G - open Pytorch models based on [5]. +- Architectural and loss changes to the model leading to fidelity improvements, based on our tokenizer representation study. +- Preparing a large scale environment dataset for multi-environment world model training. + +# 2. Related Work + +World models. Initially built as rough imagination models assisting reinforcement learning (RL) agents [10, 19, 21, 24, 53], world models have evolved into independent realistic video generation models conditioned on actions [9, 39, 50, 64]. They facilitate task-specific agent training by providing predictive representations of the environment. Inspired by [20], Ha and Schmidhuber [18] use a VAE to encode visual observations into latent states, with an MDNRNN predicting future states based on prior states, actions, and VAE outputs to facilitate policy learning. DreamerV2 [22] introduce an RL agent, achieving human-level performance in Atari. It encodes images with a CNN and computes posterior and prior stochastic states using recurrent states. Unlike our work, though, it does not assess the agent's impact on world model improvement nor generalize task rewards across different environments. + +World models also aim to generate realistic video conditioned on actions [27, 37, 65]. Genie [5] trains a video tokenizer and a Latent Action Model (LAM) for dynamic next-frame generation. GAIA-1 [27] tackle autonomous driving in unstructured settings by encoding multi-modal inputs into a unified representation and predicting image tokens based on prior inputs, using an autoregressive transformer. Menapace et al. [37] employ an encoder-decoder architecture in which the predicted action labels act as a bottleneck, allowing a user to control the generated video by a discrete action. The key gap in these works is automatic data collection, which is addressed in our approach. + +Efficient exploration. The importance of efficient exploration in RL is highlighted by [28]. Early methods enhanced exploration by adding noise [16, 34] or using entropy regularization [40], but they have action space limitations and often fail with complex dynamics, where varied actions do not always drive meaningful exploration. A more direct approach uses heterogeneous actors [26, 29, 52] with diverse exploration strategies to enhance environment exploration. Bayesian methods [54, 57] have also been introduced to create acquisition functions for uncertainty-driven exploration [2, 38, 41-44], but often struggle to generalize to high-dimensional inputs like images. + +![](images/3138283f2fbff07dd930742fb664a31a4e28a7464d6e32665e8d1ed31f672fd3.jpg) +Figure 2. Method Overview. We propose an alternative to costly human interaction data collection - by exploring environments with an agent. The reward is solely based on the classification uncertainty of our model. + +Recent exploration methods emphasize state novelty [3, 7, 11, 35, 36, 45, 55, 67], focusing on encouraging agents to assess novelty only after visiting states. In contrast, our approach, inspired by [7, 47, 53], uses model disagreement to proactively guide agents to states with the highest potential without environment target-driven reward. [6, 46] propose exploration agents driven by uncertainty in state transition and simple feature extractors. Instead, we propose an exploration agent designed to improve a world model that does not model states. Plan2Explore [53] enables agents to seek novel states using a reward that maximizes the state entropy of an RSSM model. While Plan2Explore improves goal-driven agents with their framework, we improve a modern transformer world model with a novel exploration-based reward using token uncertainty. + +Rather than relying on world models, EX2 [15] learns a classifier to distinguish visited states, providing intrinsic rewards for states that are difficult for the classifier to differentiate. KL-divergence-based approaches [30-32], guide exploration by comparing distributions. For example, SMM [32] computes the KL divergence between the policy-induced state distribution and a uniform target. Tao et al. [56] propose an intrinsic reward based on the distance between a state and its nearest neighbors in a low-dimensional feature space. However, low-dimensionality leads to information loss, restricting full state space exploration — an issue we address by the use of a world model. + +# 3. RetroAct Dataset + +We first tackle the problem of accessible training of multi-environment world models by building a framework for cheaply acquiring multi-environment interaction data. In particular, we aim to collect interactions of similar actions + +![](images/58fd854ba9e7c60078c9858b2e37bf48605ad44bc91f70352ec1ee9f5e65c861.jpg) +Figure 3. RetroAct Annotation. Description of environments in RetroAct by annotated attribute. Better viewed zoomed. + +in many environments. Instead of relying on expensive human interaction, we obtain and curate a collection of virtual environments. As a source, we use the Stable Retro framework by [48], which is a collection of retro games across multiple platforms, with an accompanying starting state. We make no use of the defined rewards. We obtain almost all the supported games (974). + +This raw set contains an environment mix of very different visuals and behaviors. However, in our setting of learning from similar dynamics, it is required to establish correspondence between the environments' behaviors. We perform annotation where for each environment three aspects are classified. The motion style classifies the general style of what and how is moved by the controls, closely relating to game genre; the camera viewpoint; the control axis describing in which direction the player can be moved. The label distributions are shown in Fig. 3. In Tab. 1 we compare our RetroAct with other related datasets. RetroAct distinguishes itself by providing behavior and control annotations, while maintaining a high number of environments. + +It is discovered that the most prevalent type of environment in our set is the platformer - 483 titles. As the largest subset, we filter only these games for further use, as it is required to have many environments exhibiting similar con + +Table 1. Comparison of RetroAct dataset to others. + +
DatasetType#EnvironmentsDiverse BehaviorsOpenBehavior AnnotationControl Annotation
Coinrun [13]Environments1X✓XX
ALE [4]Environments57✓✓XX
Stable Retro [48]Environments1003✓✓XX
Platformers [5]VideosUnknown✓XXX
RetroAct(Ours)Environments974✓✓✓✓
+ +Controls. Five motion actions are defined for our model - moving left, right, up, down and jump. Each game has its own mapping of buttons to actions. Therefore, we generate a short clip of each of the 5 selected actions for each of the 483 titles and build an annotation tool to observe and annotate the executed action. Eventually, we annotated 2,925 behavior and 2,898 control labels. + +After experimenting, we observed that models require more training with a higher number of environments, so we defined two subsets of our large set to handle computational cost: a subset consisting of the first 200 games of 483 behavior-filtered games for pretraining, and another subset of 50 randomly selected action-consistent games using RetroAct's action labels for fine-tuning. + +We collect a large scale dataset by launching a random agent in all of the environments, collecting actions and observations. From the 200-game set we build Platformers-200 - a dataset with 10,000 episodes (50 episodes per game) with 500 frames each at most, resulting in 4.6mln images. From the 50-game set we obtain Platformers-50 - 5000 episodes (100 per game) of length at most 1000, resulting in 4.8mln images. In our protocol, we take $1\%$ of the sessions of each environment as a validation set. We show that using a random agent is already sufficient to learn a level of controllability and later build on top with an exploration agent of our design. + +To validate our GenieRedux implementation, we implement the CoinRun case study in [5]. Using the protocol from above, we obtain a dataset of 10k episodes with a maximum length of 500, resulting in 4mln images. + +# 4. Multi-Environment World Model + +Given virtual environments, our first goal is to automatically obtain a dataset of image sequences $I_{1}, \ldots, I_{N}$ and corresponding actions $a_{1}, \ldots, a_{N-1}$ . Given a sequence $I_{1}, \ldots, I_{N}$ and past and future actions $a_{1}, \ldots, a_{N+T-1}$ , our world model aims to predict the future $T$ frames $I_{N+1}, \ldots, I_{T}$ , corresponding to the actions executed. + +GenieRedux. As Genie [5] is not made available by the authors, we create an open implementation and call it GenieRedux. We validate our implementation quantitatively and qualitatively in Sec. 5 and Sup.Mat F. It consists of three components. A video + +Tokenizer encodes input frame sequences into spatiotemporal tokens: $e_1, \ldots, e_N = T_{enc}(I_1, \ldots, I_N)$ , and decodes back to images: $I_1, \ldots, I_N = T_{dec}(e_1, \ldots, e_N)$ . A Latent Action Model encodes input frame sequences into spatio-temporal tokens: $a_1, \ldots, a_{N-1} = LAM_{enc}(I_1, \ldots, I_{N-1})$ , and decodes them to reconstruct future prediction $I_2, \ldots, I_N = LAM_{dec}(a_1, \ldots, a_{N-1})$ . A Dynamics module predicts the next frames based on partially masked frame tokens and actions: $I_2, \ldots, I_{N+T-1} = D(e_1, \ldots, e_N, \ldots, e_{N+T-1}; a_1, \ldots, a_{N+T-1})$ , where in inference $e_N, \ldots, e_{N+T-1}$ are masked. We adhere closely to Genie's specifications for implementing these components. + +All components use the causal Spatial Temporal Transformer (STTN) [62]. We use Position Encoding Generator (PEG) [12] for spatial and temporal attention, and Attention with Linear Biases (ALiBi) [49] for temporal attention. + +We train our models with a sequence size of 16 frames and resolution of $64 \times 64$ to address computational limitations. We train a U-Net-based superresolution network on $50 \mathrm{~K}$ data samples to upscale the output to $256 \times 256$ . (Sup.Mat. B) + +GenieRedux-G. Building upon the base model, we offer a variant - GenieRedux-G, which is adapted to virtual environments and contains architectural and training improvements. While GenieRedux uses an indispensable LAM model to obtain the actions, we discard it, as ground truth actions are available from our agent. Instead, the one-hot actions are concatenated to each layer of the Dynamics module for conditioning. In this way, we avoid the uncertainty of a prediction. + +The Dynamics module consists of an ST-ViViT encoder, followed by a MaskGIT architecture [8], which predicts indices from the tokenizer's codebook for randomly masked input tokens during training, according to a schedule. As standard cross-entropy is used, token classification has the drawback to penalize equally any prediction different from the ground truth. However, close tokens in the codebook result in significantly fewer changes than far tokens, as also shown in Sec. 5. To enable this concept of a distance between tokens in the classification of $N_{E}$ tokens, we design a Token Distance Cross-Entropy (TDCE) Loss: + +$$ +T D C E (x, y) = \left(y ^ {T} K\right) \cdot \text {s o f t m a x} (x) + C E (x, y) \tag {1} +$$ + +Here $x \in \mathcal{R}^{N_E}$ is the prediction logits, $y \in \mathcal{R}^{N_E}$ is the ground truth one-hot class. $K \in \mathcal{R}^{N_E \times N_E}$ is a precomputed table at the start of training of the cosine distances between all tokens; $CE(.)$ denotes standard Cross-Entropy Loss. When an incorrect token class is given probability, it is penalized based on its distance to the ground truth class. + +MaskGIT's design is to take as input learnable embeddings, indexed by the tokens predicted by the Tokenizer. They are randomly initialized, and therefore contain none of the content of the tokens. Given that the encoding itself and the distance between tokens can contribute to Dynamic module's performance, we add a skip connection by adding the embedding to the token itself, which improves visual fidelity and controllability of the model. + +AutoExplore Agent We extend our framework with an exploration agent that obtains data by going deeper into the environments. We name it AutoExplore Agent. The reward of the agent is entirely based on the world model performance and operates without any environment rewards. Therefore, it can be trained in various environments without tuning to their specifics or relying on a reward definition. + +The design of our reward is based on the fact that GenieRedux-G employs classification for token prediction. Each token is predicted by sampling from a categorical distribution over the codebook. We first obtain all $N_T$ token prediction distributions by running GenieRedux-G-50 5 steps back from the current observation $I_c$ for which we want to estimate the reward. We provide 2 images $I_{c - 4}, I_{c - 3}$ , predict 3 images $-I_{c - 2}, \dots, I_c$ , and take the distributions of the predicted tokens of $I_c$ to obtain $x = [x_1, \dots, x_t, \dots, x_{N_T}]$ . We evaluate the uncertainty per predicted token $u_t$ by calculating the entropy over the categorical distribution and normalize it in the range [0, 2]: + +$$ +u _ {t} = \frac {2 \cdot \sum_ {i} ^ {N _ {T}} x _ {i} \cdot \log \left(x _ {i}\right)}{N _ {e}} \tag {2} +$$ + +Studying the properties of the Tokenizer representation, we find that a prevalent token is learned representing static parts of the environment. Only the changing parts generate high uncertainty and, therefore, we take the subset $S_{top}$ of $25\%$ highest uncertainties of the entire set of uncertainties $S = \{u_t\}$ . The reward, shown in Eq. 3, establishes the agent's goal to collect data that maximizes uncertainty of the world model. + +$$ +S _ {25 \%} = \{u \in S \mid u \geq Q _ {75} (S) \} \tag{3} +$$ + +$$ +R \left(I _ {c}\right) = \frac {1}{\left| S _ {25 \%} \right|} \sum_ {u \in S _ {25 \%}} u \tag{4} +$$ + +Our agent is an actor-critic, trained with the Policy Gradient method. For the agent architecture, we follow [39]. It + +consists of a CNN encoder followed by an LSTM. As standard in RL, 4 frames are stacked, max-pooled, and the result is the input to the agent for a single time step. + +Exploration-driven World Model Training. We initially pretrain GenieRedux-G on Platformers-200 and fine-tune on Platformers-50 to obtain the model GenieRedux-G-50. Then, we train AutoExplore Agent by using GenieRedux-G-50, using it as a source of reward. The details of training the agent are presented in Sup.Mat A.3. + +Running the trained exploratory agent on a selected environment, we obtain a new diverse dataset with action demonstrations under unseen scenes. We first fine-tune the decoder of the Tokenizer for 1,000 iterations to adapt to the new unseen scenes. The Dynamic module of GenieRedux-G is then fine-tuned on the new data to achieve greater visual fidelity and controllability under new conditions. In order to build test sets to evaluate our approach, we train an Agent-57 model for each of the environments we explored, using the available environment rewards. More details on the test setup are provided in Sup.Mat A.2. + +For visual fidelity evaluation, we use FID (Frechet inception distance) Heusel et al. [25], PSNR (signal-to-noise ratio) and SSIM (structural similarity index measure) Wang et al. [61]. To evaluate controllability, we use the recently proposed $\Delta_t$ PSNR metric [5], which compares the visual effect of the ground truth action $(\hat{x}_t)$ versus a random action $(\hat{x}_t')$ : $\Delta_t$ PSNR = PSNR $(x_t, \hat{x}_t) - \mathrm{PSNR}(x_t, \hat{x}_t')$ , where $x_t$ is the ground truth frame at time $t$ . A higher $\Delta_t$ PSNR indicates a higher level of controllability. As in Bruce et al. [5], for all experiments we report $\Delta_t$ PSNR with $t = 4$ . + +# 5. Experiments + +Comparing GenieRedux and GenieRedux-G. We implement the original CoinRun case study with a random agent, as advised by [5], in order to validate and compare GenieRedux with LAM, and GenieRedux-G which uses agent-provided actions instead. In this study, the presence of LAM is the only difference between the models. We first train on a dataset, collected by a random agent. Visual fidelity results are in Tab. 2. Our GenieRedux implementation exhibits high visual quality and matches all seven CoinRun environment actions, as well as progressing environment motions (demonstrated in Sup.Mat. F). However, as demonstrated by the metrics, GenieRedux-G shows superior visual fidelity and controllability (more in Sup.Mat. F), as it avoids the uncertainty of LAM prediction. This study demonstrates that even using a random agent can result in action performance abilities in the world model. + +Next, we train an actor-critic agent with PPO on the environment reward, following [13] to collect data and train GenieRedux-TA and GenieRedux-G-TA. Tab. 3 shows evaluation on a test set collected by a trained agent. + +Table 2. Comparison of GenieRedux and GenieRedux-G on Basic Test Set. Performed on a test set, collected from the Coinrun environment with randomly sampled actions. + +
ModelBasic Test Set
FID↓PSNR↑SSIM↑
Tokenizer18.1438.250.96
LAM37.0133.970.92
GenieRedux21.8825.510.77
GenieRedux-G18.8833.410.92
+ +Table 3. Comparison of GenieRedux and GenieRedux-G on Diverse Test Set. The models are trained with data collected by random agent and trained agent (-TA), and tested on data collected by a trained agent from the Coinrun environment. + +
ModelDiverse Test Set
FID↓PSNR↑SSIM↑
Tokenizer19.1335.850.94
Tokenizer-TA11.6340.620.97
GenieRedux23.9723.820.73
GenieRedux-G19.5131.660.90
GenieRedux-TA12.5731.970.90
GenieRedux-G-TA12.4034.440.92
+ +GenieRedux-G outperforms GenieRedux on all settings. Furthermore, models trained on diverse agent-collected data are visually superior to those trained on random agents. The higher $\Delta$ PSNR of 1.89 for GenieRedux-G-TA compared to 0.70 for GenieRedux-G shows the superiority of diverse data training in controllability. (more in Sup.Mat. F) + +Multi-Environment Models. Here, we evaluate the models we initially train on many environments from RetroAct. GenieRedux-G-200 is pretrained on the Platformers-200 dataset for 180k iterations. On the validation set, we obtain 23.32 PSNR and 17.12 FID. Using this model as a base, GenieRedux-G-50 is trained on Platformers-50. Its quantitative evaluation on a test set of 10k sessions separately generated from the selected 50 environments is at the start of Tab. 4. As the 50 environments are selected with corresponding action controls between each other, we see a boost in the quality of prediction. Fig. 4 demonstrates that the instructed action is executed successfully by GenieRedux-G. As the up action is rarely used, it serves more as a no-operation action. (more in Sup.Mat C.1) + +Ablation Study. In this experiment we evaluate the additive gain of each proposed improvement in GenieRedux-G - the additive token input and training with the Token Distance Cross-Entropy Loss. The ablation is performed on a generated test set of 10k sessions, each 500 frames long. + +Table 4. Ablation study on improvements in GenieRedux-G. + +
ModelFID↓PSNR↑SSIM↑
GenieRedux-G-20022.3125.110.80
GenieRedux-G-5023.8026.360.84
+ Token Input22.9626.650.84
+ TDCE Loss22.9527.060.85
Autoregressive22.1128.070.88
+ +![](images/929d3064ff054f87b1bde004904d29cbd7e9311016ee6b0c408afc2702b211fc.jpg) +Figure 4. Control Of GenieRedux-G-50. Demonstrating all controls of our multi-environment model on multiple games. + +The data is collected using a random action policy from the environments in Platformers-50. Visual fidelity evaluation is provided in Tab. 4. It can be seen that each component gives our model a benefit in terms of visual fidelity. Finally, we perform an autoregressive evaluation of the best model to achieve our highest score. + +Tokenizer Representation Study. This experiment provides insights into the inner workings of GenieRedux-G to motivate our proposed changes. As the Dynamics module operates entirely on the token representation, we examine it closely. Fig. 5 shows the reconstructions of an input sequence (first row) and the visualized token representation (last row), where each predicted token index is assigned a different color. The visual features of the first frame are captured by various tokens. Starting with the second frame, the representation drastically changes - a token is specialized in representing the static frame regions compared to the past, while all motion regions are updated with new content. Observing that visually similar patches predict identical or similar tokens, we replace each predicted token with its closest in the codebook. We only keep the special background token unchanged. In the second row of Fig. 5 we show the resulting reconstruction - while some blurriness appears, the image remains largely the same. Conversely, replacing each token with its furthest in the codebook (third row) results in a significantly different image. This property - closer tokens having more similar appearance - motivates our Token Distance Cross-Entropy Loss, which penalizes predicting tokens further away from the ground truth. + +Fig. 6 visualizes the uncertainty of GenieRedux-G-50 for each predicted token of its Dynamics module given a + +Table 5. Quantitative Results on 3 environments. We evaluate the benefit of the data from the propose AutoExplore Agent to our models. GenieRedux-G-50 is our pretrained world model on 50 environments. GenieRedux-G-50-ft are fine-tuned models using data from a random agent or AutoExplore (Exploration). GenieRedux-G denotes a non-fine-tuned model, trained with the exploration data. + +
EnvironmentStrategyModelFID↓PSNR↑SSIM↑ΔPSNR↑
Adventure Island IIRandomGenieRedux-G-5041.9926.320.810.83
GenieRedux-G-50-ft42.3427.040.811.19
ExplorationTokenizer-ft11.0138.950.98-
GenieRedux-G11.9428.330.880.37
GenieRedux-G-50-ft12.7730.600.901.47
Random AutoregressiveGenieRedux-G-50-ft41.5527.820.831.24
Exploration AutoregressiveGenieRedux-G-50-ft11.3333.610.942.09
Super Mario BrosRandomGenieRedux-G-5029.8334.240.940.56
GenieRedux-G-50-ft30.1334.540.940.54
ExplorationTokenizer8.0942.000.99-
GenieRedux-G9.5634.000.950.09
GenieRedux-G-50-ft9.5536.130.970.57
Random AutoregressiveGenieRedux-G-50-ft30.8434.850.950.57
Exploration AutoregressiveGenieRedux-G-50-ft9.3337.770.970.76
SmurfsRandomGenieRedux-G-5079.5121.470.690.47
GenieRedux-G-50-ft80.6121.830.700.65
ExplorationTokenizer17.8635.610.98-
GenieRedux-G20.4335.420.800.85
GenieRedux-G-50-ft20.0127.450.851.55
Random AutoregressiveGenieRedux-G-50-ft80.1622.160.710.69
Exploration AutoregressiveGenieRedux-G-50-ft18.9729.530.902.06
+ +![](images/940b81b9ca8148245440c2acb56b6d04da92ce4710442b5c8cca7d6dff176dff.jpg) +Figure 5. Tokenizer Representation. Reconstruction images from the tokenizer, and the effect of replacing each token with its closest and furthest in the codebook. Lastly, we visualize the indices of the predicted tokens. + +sequence. The uncertainty metric is the entropy of the classification over 1024 codebook tokens. Tokens corresponding to motion have the highest uncertainty; other regions are mostly classified as the "static" token. Thus, minimal character movement yields low uncertainty, while forward motion increases it. This motivates us to build AutoExplore Agent's reward based on this uncertainty. + +Exploration-based training. We demonstrate our exploration-based training of GenieRedux-G. We perform the procedure on 3 environments - AdventureIslandII, which provides an easy set- + +![](images/060b83ac3139b750e8665ce11bd220606dc67a266aa2f8636117757232268f03.jpg) +Figure 6. Dynamics Uncertainty. Shown is the uncertainty per token predicted for each image of an example sequence. Uncertainty is generated in the regions of motion. + +ting for the agent to learn (single platform with no enemies at the start), SuperMario Bros provides an enemy and obstacles soon after the start and Smurfs provides a more complex background imagery and different action dynamics. For each of the environments, we train an AutoExplorer Agent. We observe that the agent learns to move forward and navigate obstacles to maximize reward. (more in Sup.Mat. D) + +We use our pretrained GenieRedux-G-50 model as a baseline and fine-tune it for each environment in two settings - a dataset collected on the selected environment by a random agent and by our AutoExplorer Agent. Each dataset consists of 10k sessions, each 700 frames long. We fine-tune (GenieRedux-G-50-ft) for 10k iterations and pick the best performing model. In our comparison, we also include a GenieRedux-G model trained from scratch on the diverse exploration datasets for 15k iterations to show the effect of pretraining. We perform single-pass generation for all mod + +![](images/a54e18eb289388484a0bdf631c9d312baeb3d6c92dbea5f64de9c04751bbe3fb.jpg) +Figure 7. AutoExplore Agent vs Random Agent Qualitative Comparison. We show that AutoExplore exhibits better visual quality and avoids losing track of the agent. + +Table 6. Comparison of AutoExplore Agent with others. + +
AgentSuperMarioBrosAdventureIslandII
PSNR↓SSIM↓ΔPSNR↓PSNR↓SSIM↓ΔPSNR↓
RF28.580.940.18124.820.780.44
VAE24.400.860.08716.570.50.072
Ours23.810.850.06515.200.410.070
+ +els and the more computationally heavy autoregressive evaluation for the fine-tuned models on data from random and AutoExplore Agent's datasets. Tab. 5 shows visual fidelity and controllability metrics for each environment, confirming the effectiveness of our exploration method. The model fine-tuned on AutoExplore Agent's data consistently outperforms the models trained on random actions in terms of visual fidelity. Exploration-based fine-tuning also improves controllability. Environments with small characters and uniform backgrounds can be more challenging for all models to learn. However, the gain in controllability in this case remains noticeable during autoregressive evaluation. Fig. 7 demonstrates the superior quality of our method. In addition, we observe that the multienvironment pretraining leads to significant gains in both studied aspects compared to the nonpretrained model. (more in Sup.Mat. C) + +AutoExplore Agent Evaluation. We compare AutoExplore Agent with exploration-based methods in [6]. We train agents based on SSE of RF and VAE features on top of GenieRedux and compare with ours on Tab. 6. AutoExplore Agent's reward results in maximum world model visual and controllability errors (on 1k episodes of agent actions), fulfilling its intended role in our framework. + +User Studies. To validate the quality of our final results, we perform a user study in which we ask people to rate from 1 to 5 the quality of samples produced respectively by GenieRedux-G trained on random agent's data and on AutoExplore Agent's data. Each sample in our study consists of two 16-frame clips playing in a synchronized manner - the ground truth clip and our GenieRedux-G-50-ft reconstruction, given two initial frames and generating the rest autoregressively. We provide a total of 120 samples to the users - 40 samples per model and 40 samples of two ground + +![](images/cd93a57c19be2f18949958e60622afee40ba84b0c64ccfd6c1b60bfa5e8ae4df.jpg) +Figure 8. User study results. Our user study on two games shows that our model trained with AutoExplore Agent's data is consistently rated higher. + +![](images/db754866630a6436c6985947509e3e762a1a58281c217489525370380681fb47.jpg) + +truth samples, to establish scale. We give 20 samples from each of the two selected games - SuperMario Bros and AdventureIslandII. We get reviews from 19 participants. The results are shown in Fig. 8. The model, trained on data from AutoExplore Agent is clearly rated closer to the ground truth, establishing the quality of our method. + +With a second user study, we evaluate the action accuracy of the generated frames. We use ambiguous single input cases (character starting mid-air) and generate 60 clips with 3 actions on AdventureIslandII. Users prefer our exploration-trained model, rating it $\mathbf{0.75} \pm \mathbf{0.019}$ on a scale from 0 (random preferred) to 1 (exploration preferred). (more in Sup.Mat. E.2) + +# 6. Conclusion + +As world models have developed into large models with impressive simulation properties, they require large interaction datasets, complete with diverse observations and actions. Genie [5] demonstrates impressive abilities by training on multiple environments, however, requiring the collection of a large video dataset and a model to infer actions. + +In this work, we address the heavy burden of data collection and curation by building a new framework for training large world models by collecting interaction data from a large number of virtual environments. We first build an open implementation of Genie - GenieRedux and enhance it into its version GenieRedux-G. We obtain models exhibiting control by pretraining on a large set of virtual environments. We address the overfitting limitations of random data collection policy by proposing AutoExplore Agent, an agent entirely independent of the environment reward, maximizing the uncertainty of GenieRedux-G. After finetuning on the explored environment, our model is able to improve its visual fidelity and controllability much better than training solely on random agent's data. Demonstrating this on multiple environments, we show the potential of our framework to make training of next-generation world models more accessible, cost-effective, and effort-free. + +# 7. Acknowledgments + +INSAIT, Sofia University "St. Kliment Ohridski". Partially funded by the Ministry of Education and Science of Bulgaria's support for INSAIT as part of the Bulgarian National Roadmap for Research Infrastructure. This project was supported with computational resources provided by Google Cloud Platform (GCP). + +# References + +[1] Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37: 58757-58791, 2024. +[2] Kamyar Azizzadenesheli, Emma Brunskill, and Animashree Anandkumar. Efficient exploration through bayesian deep q-networks. In 2018 Information Theory and Applications Workshop (ITA), pages 1-9. IEEE, 2018. +[3] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. Advances in neural information processing systems, 29, 2016. +[4] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253-279, 2013. +[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. +[6] Yuri Burda, Harri Edwards, Deepak Pathak, Amos Storkey, Trevor Darrell, and Alexei A Efros. Large-scale study of curiosity-driven learning. In International Conference on Learning Representations, 2019. +[7] Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation. In International Conference on Learning Representations, 2019. +[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. +[9] Chang Chen, Yi-Fu Wu, Jaesik Yoon, and Sungjin Ahn. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481, 2022. +[10] Silvia Chiappa, Sébastien Racanière, Daan Wierstra, and Shakir Mohamed. Recurrent environment simulators. In International Conference on Learning Representations, 2017. +[11] Leshem Choshen, Lior Fox, and Yonatan Loewenstein. Dora the explorer: Directed outreach reinforcement action-selection. arXiv preprint arXiv:1804.04012, 2018. +[12] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In The Eleventh International Conference on Learning Representations, 2023. + +[13] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In International conference on machine learning, pages 1282-1289. PMLR, 2019. +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. +[15] Justin Fu, John Co-Reyes, and Sergey Levine. Ex2: Exploration with exemplar models for deep reinforcement learning. Advances in neural information processing systems, 30, 2017. +[16] Scott Fujimoto, Herke Hoof, and David Meger. Addressing function approximation error in actor-critic methods. In International conference on machine learning, pages 1587-1596. PMLR, 2018. +[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. +[18] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. +[19] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. +[20] David Ha, Jonas Jongejan, and Ian Johnson. Draw together with a neural network. Retrieved Oct, 5:2021, 2017. +[21] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019. +[22] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021. +[23] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021. +[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. +[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. +[26] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay. In International Conference on Learning Representations, 2018. +[27] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. + +[28] Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In Proceedings of the Nineteenth International Conference on Machine Learning, pages 267-274, 2002. +[29] Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent experience replay in distributed reinforcement learning. In International conference on learning representations, 2018. +[30] Youngjin Kim, Wontae Nam, Hyunwoo Kim, Ji-Hoon Kim, and Gunhee Kim. Curiosity-bottleneck: Exploration by distilling task-specific novelty. In International conference on machine learning, pages 3379-3388. PMLR, 2019. +[31] Martin Klissarov, Riashat Islam, Khimya Khetarpal, and Doina Precup. Variational state encoding as intrinsic motivation in reinforcement learning. In *Task-Agnostic Reinforcement Learning Workshop at Proceedings of the International Conference on Learning Representations*, pages 16–32, 2019. +[32] Lisa Lee, Benjamin Eysenbach, Emilio Parisotto, Eric Xing, Sergey Levine, and Ruslan Salakhutdinov. Efficient exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019. +[33] Ian Lenz, Ross A Knepper, and Ashutosh Saxena. Deepmpc: Learning deep latent features for model predictive control. In Robotics: Science and Systems, page 25. Rome, Italy, 2015. +[34] Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. Advances in neural information processing systems, 30, 2017. +[35] Marlos C Machado, Marc G Bellemare, and Michael Bowling. Count-based exploration with the successor representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5125-5133, 2020. +[36] Jarryd Martin, Suraj Narayanan Sasikumar, Tom Everitt, and Marcus Hutter. Count-based exploration in feature space for reinforcement learning. arXiv preprint arXiv:1706.08090, 2017. +[37] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021. +[38] Alberto Maria Metelli, Amarildo Likmeta, and Marcello Restelli. Propagating uncertainty in reinforcement learning via wasserstein barycenters. Advances in Neural Information Processing Systems, 32, 2019. +[39] Vincent Micheli, Eloi Alonso, and François Fleuret. Transformers are sample-efficient world models. In The Eleventh International Conference on Learning Representations, 2023. +[40] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016. +[41] Ian Osband and Benjamin Van Roy. Bootstrapped thompson sampling and deep exploration. arXiv preprint arXiv:1507.00300, 2015. + +[42] Ian Osband, Charles Blundell, Alexander Pritzel, and Benjamin Van Roy. Deep exploration via bootstrapped dqn. Advances in neural information processing systems, 29, 2016. +[43] Ian Osband, Benjamin Van Roy, and Zheng Wen. Generalization and exploration via randomized value functions. In International Conference on Machine Learning, pages 2377-2386. PMLR, 2016. +[44] Ian Osband, John Aslanides, and Albin Cassirer. Randomized prior functions for deep reinforcement learning. Advances in Neural Information Processing Systems, 31, 2018. +[45] Georg Ostrovski, Marc G Bellemare, Aäron Oord, and Rémi Munos. Count-based exploration with neural density models. In International conference on machine learning, pages 2721-2730. PMLR, 2017. +[46] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International conference on machine learning, pages 2778-2787. PMLR, 2017. +[47] Deepak Pathak, Dhiraj Gandhi, and Abhinav Gupta. Self-supervised exploration via disagreement. In International conference on machine learning, pages 5062-5071. PMLR, 2019. +[48] Mathieu Poliquin. Stable retro, a maintained fork of openai's gym-retro. https://github.com/Farama-Foundation/stable-retro, 2024. +[49] Ofir Press, Noah Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. In International Conference on Learning Representations, 2022. +[50] Jan Robine, Marc Höftmann, Tobias Uelwer, and Stefan Harmeling. Transformer-based world models are happy with 100k interactions. In The Eleventh International Conference on Learning Representations, 2023. +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[52] Tom Schaul. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. +[53] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International conference on machine learning, pages 8583-8592. PMLR, 2020. +[54] Niranjan Srinivas, Andreas Krause, Sham Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: no regret and experimental design. In Proceedings of the 27th International Conference on International Conference on Machine Learning, pages 1015-1022, 2010. +[55] Haoran Tang, Rein Houthooft, Davis Foote, Adam Stooke, OpenAI Xi Chen, Yan Duan, John Schulman, Filip DeTurck, and Pieter Abbeel. # exploration: A study of count-based exploration for deep reinforcement learning. Advances in neural information processing systems, 30, 2017. +[56] Ruo Yu Tao, Vincent François-Lavet, and Joelle Pineau. Novelty search in representational space for sample efficient + +exploration. Advances in Neural Information Processing Systems, 33:8114-8126, 2020. +[57] William R Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3-4):285-294, 1933. +[58] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024. +[59] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. +[60] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. +[61] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. +[62] Mingxing Xu, Wenrui Dai, Chunmiao Liu, Xing Gao, Weiyao Lin, Guo-Jun Qi, and Hongkai Xiong. Spatial-temporal transformer networks for traffic flow forecasting. arXiv preprint arXiv:2001.02908, 2020. +[63] Sherry Yang, Yilun Du, Seyed Kamyar Seyed Ghasemipour, Jonathan Tompson, Leslie Pack Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. In The Twelfth International Conference on Learning Representations, 2023. +[64] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. +[65] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. +[66] Lixuan Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Prelar: World model pre-training with learnable action representation. In European Conference on Computer Vision, pages 185-201, 2024. +[67] Tianjun Zhang, Paria Rashidinejad, Jiantao Jiao, Yuandong Tian, Joseph E Gonzalez, and Stuart Russell. Made: Exploration via maximizing deviation from explored regions. Advances in Neural Information Processing Systems, 34:9663-9680, 2021. \ No newline at end of file diff --git a/2025/Exploration-Driven Generative Interactive Environments/images.zip b/2025/Exploration-Driven Generative Interactive Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab19bbbb0c8756d9ffc5accd19276ef91366bb93 --- /dev/null +++ b/2025/Exploration-Driven Generative Interactive Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:682f8d0c85868bb7ed86b4d43ef239599bfba3dd7a8e2e7d7d1e39cc509a790d +size 691447 diff --git a/2025/Exploration-Driven Generative Interactive Environments/layout.json b/2025/Exploration-Driven Generative Interactive Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64cbb7856bb060569454264e65ac3fd41017b84a --- /dev/null +++ b/2025/Exploration-Driven Generative Interactive Environments/layout.json @@ -0,0 +1,7583 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "Exploration-Driven Generative Interactive Environments" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 92, + 141, + 518, + 186 + ], + "blocks": [ + { + "bbox": [ + 92, + 141, + 518, + 186 + ], + "lines": [ + { + "bbox": [ + 92, + 141, + 518, + 186 + ], + "spans": [ + { + "bbox": [ + 92, + 141, + 518, + 186 + ], + "type": "image", + "image_path": "523a425dc0fbd2aff438c9045131f881214fbf172b51a5b1b61e150d5269b881.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 224, + 200, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 224, + 200, + 238 + ], + "spans": [ + { + "bbox": [ + 152, + 224, + 200, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 502 + ], + "type": "text", + "content": "Modern world models require costly and time-consuming collection of large video datasets with action demonstrations by people or by environment-specific agents. To simplify training, we focus on using many virtual environments for inexpensive, automatically collected interaction data. Genie [5], a recent multi-environment world model, demonstrates simulation abilities of many environments with shared behavior. Unfortunately, training their model requires expensive demonstrations. Therefore, we propose a training framework merely using a random agent in virtual environments. While the model trained in this manner exhibits good controls, it is limited by the random exploration possibilities. To address this limitation, we propose AutoExplore Agent - an exploration agent that entirely relies on the uncertainty of the world model, delivering diverse data from which it can learn the best. Our agent is fully independent of environment-specific rewards and thus adapts easily to new environments. With this approach, the pretrained multi-environment model can quickly adapt to new environments achieving video fidelity and controllability improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 503, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 503, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 503, + 296, + 612 + ], + "type": "text", + "content": "In order to obtain automatically large-scale interaction datasets for pretraining, we group environments with similar behavior and controls. To this end, we annotate the behavior and controls of 974 virtual environments - a dataset that we name RetroAct. For building our model, we first create an open implementation of Genie - GenieRedux and apply enhancements and adaptations in our version GenieRedux-G. Our code and data are available at https://github.com/insait-institute/GenieRedux." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "content": "Learning from interactive environments allows us to understand and represent the rules, the possible actions, and the consequences that govern them. As an alternative to labori" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 555, + 414 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 555, + 414 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 555, + 414 + ], + "type": "image", + "image_path": "d92641e2d123750a2b97de7fb1b311ddc86da408832ae75ce35a1d568f602a70.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 428, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. Our proposed world model training framework. It consists of a pretrained multi-environment world model on random agent data, and a new AutoExplore Agent that explores an environment and delivers diverse data for fine-tuning." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "text", + "content": "ously hand-coded synthetic simulators, world models have emerged as deep learning tools for realistic environment modeling entirely from observations, commonly images of the observed environment [1, 5, 37, 65]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Previous work such as [19, 23, 33] uses light world models to support goal-driven agents with goal-specific state representations. The focus is on coarse future predictions, not on their high visual quality. In contrast, the objective of recent world models is to achieve high-quality future predictions given past observations and actions. Such recent models are able to offer realistic action execution and even real-time interaction with people [1, 58]. This has become possible with the rise of diffusion, transformers [14, 59], and state space models [17], and by borrowing architectural choices from video generation pipelines [51, 60]. Typically, these generative models are designed to closely match a single selected environment. One of the state-of-the-art models, Genie, distinguishes itself by being trained on many visually diverse environments with similar dynamics, thus" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 254, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 254, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 254, + 713 + ], + "type": "text", + "content": "†Corresponding author. firstname(secondname@insait.ai" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27597" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 253, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 253, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 253, + 83 + ], + "type": "text", + "content": "demonstrating generalization across new visuals." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 230 + ], + "type": "text", + "content": "Building these high-quality statistical simulators requires diverse observations of the environment as well as of the actions to simulate. Some obtain this data by costly video dataset collection and curation with human demonstrations of the actions [1, 5, 63]. If actions are unavailable, an extra component is designed to predict them, which can introduce uncertainty compared to ground truth labels [5, 37]. Extension to new environments with new types of actions in this setting is difficult as it requires again an expensive data collection process. Others, such as [58] have explored retrieving data with an environment-specific agent, in their case - the game Doom." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 483 + ], + "type": "text", + "content": "In this work, we propose a framework for accessible and effort-free training of world models in multiple environments. To this end, we first build RetroAct- an annotated and curated large dataset of retro game environments (based on the environments of Stable Retro [48]). We group them based on behavior labels and control descriptions. This grouping allows us to generate large-scale interaction datasets across environments with similar behaviors. Next, we pretrain a multi-environment world model GenieRedux - our open implementation of Genie [5], using a random agent. Unlike [66], which reports the agent's improved behavior from pretraining, we aim to improve the world model. For this, we adapt GenieRedux to virtual environments and implement architectural and training procedure enhancements, resulting in the GenieRedux-G model. We observe that just by training GenieRedux-G on random interactions from subsets of 200 environments and 50 environments with mapped controls, automatically collected from RetroAct, we are able to obtain control behavior (0.450 " + }, + { + "bbox": [ + 55, + 232, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 232, + 296, + 483 + ], + "type": "text", + "content": "PSNR in 50 environments) and reasonable visual fidelity (26.36 PSNR in 50 environments)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "content": "As random actions are limited in their ability to explore the environment, we develop a method to obtain more diverse interaction data to improve the control behavior and visual fidelity of our model. To this end, inspired by [53], we develop our own environment-independent reward function, allowing an agent to explore different environments, entirely without relying on predefined environment rewards. While they aim at a high-performing goal-driven agent, we base our design on improving the underlying large world model for the simulation of environments in terms of higher visual fidelity and improved controllability. For graphical illustrations, see Fig. 1. The objective of our exploration-driven agent is to maximize the world model's uncertainty, estimated by the classification entropy available in the observation prediction stage of GenieRedux-G. Once the diverse data is obtained, we fine-tune GenieRedux-G. We show that this method leads to significant visual (up to 7.4 PSNR) and control (up to " + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "inline_equation", + "content": "1.4\\Delta" + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "content": " PSNR) improvements, compared to random agent pretraining." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 72, + 460, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 460, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 460, + 83 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 84, + 553, + 204 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "type": "text", + "content": "- A framework for training world models with cheap data collection - by training an exploration agent based on our world's model uncertainty." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 144 + ], + "type": "text", + "content": "- The implementation and release of GenieRedux and GenieRedux-G - open Pytorch models based on [5]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "content": "- Architectural and loss changes to the model leading to fidelity improvements, based on our tokenizer representation study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "content": "- Preparing a large scale environment dataset for multi-environment world model training." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 215, + 400, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 400, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 400, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 235, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 425 + ], + "type": "text", + "content": "World models. Initially built as rough imagination models assisting reinforcement learning (RL) agents [10, 19, 21, 24, 53], world models have evolved into independent realistic video generation models conditioned on actions [9, 39, 50, 64]. They facilitate task-specific agent training by providing predictive representations of the environment. Inspired by [20], Ha and Schmidhuber [18] use a VAE to encode visual observations into latent states, with an MDNRNN predicting future states based on prior states, actions, and VAE outputs to facilitate policy learning. DreamerV2 [22] introduce an RL agent, achieving human-level performance in Atari. It encodes images with a CNN and computes posterior and prior stochastic states using recurrent states. Unlike our work, though, it does not assess the agent's impact on world model improvement nor generalize task rewards across different environments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": "World models also aim to generate realistic video conditioned on actions [27, 37, 65]. Genie [5] trains a video tokenizer and a Latent Action Model (LAM) for dynamic next-frame generation. GAIA-1 [27] tackle autonomous driving in unstructured settings by encoding multi-modal inputs into a unified representation and predicting image tokens based on prior inputs, using an autoregressive transformer. Menapace et al. [37] employ an encoder-decoder architecture in which the predicted action labels act as a bottleneck, allowing a user to control the generated video by a discrete action. The key gap in these works is automatic data collection, which is addressed in our approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "Efficient exploration. The importance of efficient exploration in RL is highlighted by [28]. Early methods enhanced exploration by adding noise [16, 34] or using entropy regularization [40], but they have action space limitations and often fail with complex dynamics, where varied actions do not always drive meaningful exploration. A more direct approach uses heterogeneous actors [26, 29, 52] with diverse exploration strategies to enhance environment exploration. Bayesian methods [54, 57] have also been introduced to create acquisition functions for uncertainty-driven exploration [2, 38, 41-44], but often struggle to generalize to high-dimensional inputs like images." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27598" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 556, + 274 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 274 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 274 + ], + "type": "image", + "image_path": "3138283f2fbff07dd930742fb664a31a4e28a7464d6e32665e8d1ed31f672fd3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. Method Overview. We propose an alternative to costly human interaction data collection - by exploring environments with an agent. The reward is solely based on the classification uncertainty of our model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 308, + 295, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 295, + 488 + ], + "type": "text", + "content": "Recent exploration methods emphasize state novelty [3, 7, 11, 35, 36, 45, 55, 67], focusing on encouraging agents to assess novelty only after visiting states. In contrast, our approach, inspired by [7, 47, 53], uses model disagreement to proactively guide agents to states with the highest potential without environment target-driven reward. [6, 46] propose exploration agents driven by uncertainty in state transition and simple feature extractors. Instead, we propose an exploration agent designed to improve a world model that does not model states. Plan2Explore [53] enables agents to seek novel states using a reward that maximizes the state entropy of an RSSM model. While Plan2Explore improves goal-driven agents with their framework, we improve a modern transformer world model with a novel exploration-based reward using token uncertainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 488, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 632 + ], + "type": "text", + "content": "Rather than relying on world models, EX2 [15] learns a classifier to distinguish visited states, providing intrinsic rewards for states that are difficult for the classifier to differentiate. KL-divergence-based approaches [30-32], guide exploration by comparing distributions. For example, SMM [32] computes the KL divergence between the policy-induced state distribution and a uniform target. Tao et al. [56] propose an intrinsic reward based on the distance between a state and its nearest neighbors in a low-dimensional feature space. However, low-dimensionality leads to information loss, restricting full state space exploration — an issue we address by the use of a world model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 645, + 160, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 160, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 160, + 657 + ], + "type": "text", + "content": "3. RetroAct Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "We first tackle the problem of accessible training of multi-environment world models by building a framework for cheaply acquiring multi-environment interaction data. In particular, we aim to collect interactions of similar actions" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 304, + 557, + 392 + ], + "blocks": [ + { + "bbox": [ + 315, + 304, + 557, + 392 + ], + "lines": [ + { + "bbox": [ + 315, + 304, + 557, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 304, + 557, + 392 + ], + "type": "image", + "image_path": "58fd854ba9e7c60078c9858b2e37bf48605ad44bc91f70352ec1ee9f5e65c861.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "type": "text", + "content": "Figure 3. RetroAct Annotation. Description of environments in RetroAct by annotated attribute. Better viewed zoomed." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 426, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 510 + ], + "type": "text", + "content": "in many environments. Instead of relying on expensive human interaction, we obtain and curate a collection of virtual environments. As a source, we use the Stable Retro framework by [48], which is a collection of retro games across multiple platforms, with an accompanying starting state. We make no use of the defined rewards. We obtain almost all the supported games (974)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 665 + ], + "type": "text", + "content": "This raw set contains an environment mix of very different visuals and behaviors. However, in our setting of learning from similar dynamics, it is required to establish correspondence between the environments' behaviors. We perform annotation where for each environment three aspects are classified. The motion style classifies the general style of what and how is moved by the controls, closely relating to game genre; the camera viewpoint; the control axis describing in which direction the player can be moved. The label distributions are shown in Fig. 3. In Tab. 1 we compare our RetroAct with other related datasets. RetroAct distinguishes itself by providing behavior and control annotations, while maintaining a high number of environments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "It is discovered that the most prevalent type of environment in our set is the platformer - 483 titles. As the largest subset, we filter only these games for further use, as it is required to have many environments exhibiting similar con" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27599" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 82, + 535, + 167 + ], + "blocks": [ + { + "bbox": [ + 202, + 71, + 408, + 81 + ], + "lines": [ + { + "bbox": [ + 202, + 71, + 408, + 81 + ], + "spans": [ + { + "bbox": [ + 202, + 71, + 408, + 81 + ], + "type": "text", + "content": "Table 1. Comparison of RetroAct dataset to others." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 82, + 535, + 167 + ], + "lines": [ + { + "bbox": [ + 74, + 82, + 535, + 167 + ], + "spans": [ + { + "bbox": [ + 74, + 82, + 535, + 167 + ], + "type": "table", + "html": "
DatasetType#EnvironmentsDiverse BehaviorsOpenBehavior AnnotationControl Annotation
Coinrun [13]Environments1X✓XX
ALE [4]Environments57✓✓XX
Stable Retro [48]Environments1003✓✓XX
Platformers [5]VideosUnknown✓XXX
RetroAct(Ours)Environments974✓✓✓✓
", + "image_path": "e3ad0f9fa00c7f7c26ef69ffe4244af79e28f929e5ccd055a37f7005d3c12112.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 173, + 294, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 173, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 54, + 173, + 294, + 255 + ], + "type": "text", + "content": "Controls. Five motion actions are defined for our model - moving left, right, up, down and jump. Each game has its own mapping of buttons to actions. Therefore, we generate a short clip of each of the 5 selected actions for each of the 483 titles and build an annotation tool to observe and annotate the executed action. Eventually, we annotated 2,925 behavior and 2,898 control labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 257, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 295, + 341 + ], + "type": "text", + "content": "After experimenting, we observed that models require more training with a higher number of environments, so we defined two subsets of our large set to handle computational cost: a subset consisting of the first 200 games of 483 behavior-filtered games for pretraining, and another subset of 50 randomly selected action-consistent games using RetroAct's action labels for fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": "We collect a large scale dataset by launching a random agent in all of the environments, collecting actions and observations. From the 200-game set we build Platformers-200 - a dataset with 10,000 episodes (50 episodes per game) with 500 frames each at most, resulting in 4.6mln images. From the 50-game set we obtain Platformers-50 - 5000 episodes (100 per game) of length at most 1000, resulting in 4.8mln images. In our protocol, we take " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " of the sessions of each environment as a validation set. We show that using a random agent is already sufficient to learn a level of controllability and later build on top with an exploration agent of our design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 485, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 533 + ], + "type": "text", + "content": "To validate our GenieRedux implementation, we implement the CoinRun case study in [5]. Using the protocol from above, we obtain a dataset of 10k episodes with a maximum length of 500, resulting in 4mln images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 242, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 242, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 242, + 558 + ], + "type": "text", + "content": "4. Multi-Environment World Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "content": "Given virtual environments, our first goal is to automatically obtain a dataset of image sequences " + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "inline_equation", + "content": "I_{1}, \\ldots, I_{N}" + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "content": " and corresponding actions " + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "inline_equation", + "content": "a_{1}, \\ldots, a_{N-1}" + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "content": ". Given a sequence " + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "inline_equation", + "content": "I_{1}, \\ldots, I_{N}" + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "content": " and past and future actions " + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "inline_equation", + "content": "a_{1}, \\ldots, a_{N+T-1}" + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "content": ", our world model aims to predict the future " + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "inline_equation", + "content": "I_{N+1}, \\ldots, I_{T}" + }, + { + "bbox": [ + 54, + 565, + 295, + 637 + ], + "type": "text", + "content": ", corresponding to the actions executed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "content": "GenieRedux. As Genie [5] is not made available by the authors, we create an open implementation and call it GenieRedux. We validate our implementation quantitatively and qualitatively in Sec. 5 and Sup.Mat F. It consists of three components. A video" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "content": "Tokenizer encodes input frame sequences into spatiotemporal tokens: " + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "inline_equation", + "content": "e_1, \\ldots, e_N = T_{enc}(I_1, \\ldots, I_N)" + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "content": ", and decodes back to images: " + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "inline_equation", + "content": "I_1, \\ldots, I_N = T_{dec}(e_1, \\ldots, e_N)" + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "content": ". A Latent Action Model encodes input frame sequences into spatio-temporal tokens: " + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "inline_equation", + "content": "a_1, \\ldots, a_{N-1} = LAM_{enc}(I_1, \\ldots, I_{N-1})" + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "content": ", and decodes them to reconstruct future prediction " + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "inline_equation", + "content": "I_2, \\ldots, I_N = LAM_{dec}(a_1, \\ldots, a_{N-1})" + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "content": ". A Dynamics module predicts the next frames based on partially masked frame tokens and actions: " + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "inline_equation", + "content": "I_2, \\ldots, I_{N+T-1} = D(e_1, \\ldots, e_N, \\ldots, e_{N+T-1}; a_1, \\ldots, a_{N+T-1})" + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "content": ", where in inference " + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "inline_equation", + "content": "e_N, \\ldots, e_{N+T-1}" + }, + { + "bbox": [ + 313, + 173, + 553, + 316 + ], + "type": "text", + "content": " are masked. We adhere closely to Genie's specifications for implementing these components." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 317, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 553, + 365 + ], + "type": "text", + "content": "All components use the causal Spatial Temporal Transformer (STTN) [62]. We use Position Encoding Generator (PEG) [12] for spatial and temporal attention, and Attention with Linear Biases (ALiBi) [49] for temporal attention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "text", + "content": "We train our models with a sequence size of 16 frames and resolution of " + }, + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "text", + "content": " to address computational limitations. We train a U-Net-based superresolution network on " + }, + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "inline_equation", + "content": "50 \\mathrm{~K}" + }, + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "text", + "content": " data samples to upscale the output to " + }, + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 365, + 553, + 426 + ], + "type": "text", + "content": ". (Sup.Mat. B)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "content": "GenieRedux-G. Building upon the base model, we offer a variant - GenieRedux-G, which is adapted to virtual environments and contains architectural and training improvements. While GenieRedux uses an indispensable LAM model to obtain the actions, we discard it, as ground truth actions are available from our agent. Instead, the one-hot actions are concatenated to each layer of the Dynamics module for conditioning. In this way, we avoid the uncertainty of a prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 553, + 554, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 554, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 554, + 685 + ], + "type": "text", + "content": "The Dynamics module consists of an ST-ViViT encoder, followed by a MaskGIT architecture [8], which predicts indices from the tokenizer's codebook for randomly masked input tokens during training, according to a schedule. As standard cross-entropy is used, token classification has the drawback to penalize equally any prediction different from the ground truth. However, close tokens in the codebook result in significantly fewer changes than far tokens, as also shown in Sec. 5. To enable this concept of a distance between tokens in the classification of " + }, + { + "bbox": [ + 313, + 553, + 554, + 685 + ], + "type": "inline_equation", + "content": "N_{E}" + }, + { + "bbox": [ + 313, + 553, + 554, + 685 + ], + "type": "text", + "content": " tokens, we design a Token Distance Cross-Entropy (TDCE) Loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 696, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 696, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 323, + 696, + 553, + 711 + ], + "type": "interline_equation", + "content": "T D C E (x, y) = \\left(y ^ {T} K\\right) \\cdot \\text {s o f t m a x} (x) + C E (x, y) \\tag {1}", + "image_path": "bf8ebe4dc6ccf70960737298ff20cc227707e4b14a2886ea4ddace9e86ae52ba.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27600" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{R}^{N_E}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is the prediction logits, " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{R}^{N_E}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is the ground truth one-hot class. " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "K \\in \\mathcal{R}^{N_E \\times N_E}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is a precomputed table at the start of training of the cosine distances between all tokens; " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "CE(.)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " denotes standard Cross-Entropy Loss. When an incorrect token class is given probability, it is penalized based on its distance to the ground truth class." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 240 + ], + "type": "text", + "content": "MaskGIT's design is to take as input learnable embeddings, indexed by the tokens predicted by the Tokenizer. They are randomly initialized, and therefore contain none of the content of the tokens. Given that the encoding itself and the distance between tokens can contribute to Dynamic module's performance, we add a skip connection by adding the embedding to the token itself, which improves visual fidelity and controllability of the model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 256, + 294, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 294, + 339 + ], + "type": "text", + "content": "AutoExplore Agent We extend our framework with an exploration agent that obtains data by going deeper into the environments. We name it AutoExplore Agent. The reward of the agent is entirely based on the world model performance and operates without any environment rewards. Therefore, it can be trained in various environments without tuning to their specifics or relying on a reward definition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": "The design of our reward is based on the fact that GenieRedux-G employs classification for token prediction. Each token is predicted by sampling from a categorical distribution over the codebook. We first obtain all " + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "inline_equation", + "content": "N_T" + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": " token prediction distributions by running GenieRedux-G-50 5 steps back from the current observation " + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "inline_equation", + "content": "I_c" + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": " for which we want to estimate the reward. We provide 2 images " + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "inline_equation", + "content": "I_{c - 4}, I_{c - 3}" + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": ", predict 3 images " + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "inline_equation", + "content": "-I_{c - 2}, \\dots, I_c" + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": ", and take the distributions of the predicted tokens of " + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "inline_equation", + "content": "I_c" + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "inline_equation", + "content": "x = [x_1, \\dots, x_t, \\dots, x_{N_T}]" + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": ". We evaluate the uncertainty per predicted token " + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 55, + 340, + 294, + 483 + ], + "type": "text", + "content": " by calculating the entropy over the categorical distribution and normalize it in the range [0, 2]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 491, + 294, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 491, + 294, + 519 + ], + "spans": [ + { + "bbox": [ + 119, + 491, + 294, + 519 + ], + "type": "interline_equation", + "content": "u _ {t} = \\frac {2 \\cdot \\sum_ {i} ^ {N _ {T}} x _ {i} \\cdot \\log \\left(x _ {i}\\right)}{N _ {e}} \\tag {2}", + "image_path": "1949ec11896f9f46502d32be94d0940ac1bc449edf6ed416c51fff1b05dcfaad.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "text", + "content": "Studying the properties of the Tokenizer representation, we find that a prevalent token is learned representing static parts of the environment. Only the changing parts generate high uncertainty and, therefore, we take the subset " + }, + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "inline_equation", + "content": "S_{top}" + }, + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "text", + "content": " highest uncertainties of the entire set of uncertainties " + }, + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "inline_equation", + "content": "S = \\{u_t\\}" + }, + { + "bbox": [ + 55, + 528, + 295, + 623 + ], + "type": "text", + "content": ". The reward, shown in Eq. 3, establishes the agent's goal to collect data that maximizes uncertainty of the world model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 635, + 294, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 635, + 294, + 648 + ], + "spans": [ + { + "bbox": [ + 110, + 635, + 294, + 648 + ], + "type": "interline_equation", + "content": "S _ {25 \\%} = \\{u \\in S \\mid u \\geq Q _ {75} (S) \\} \\tag{3}", + "image_path": "00488890a70a93729fcc0e973ddfc182044073c33f9ee47e83614f6b21328e6b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 651, + 294, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 651, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 134, + 651, + 294, + 680 + ], + "type": "interline_equation", + "content": "R \\left(I _ {c}\\right) = \\frac {1}{\\left| S _ {25 \\%} \\right|} \\sum_ {u \\in S _ {25 \\%}} u \\tag{4}", + "image_path": "510cb409475c141a3a87c3c1eb71178861c6ffcfd1293e3dfcbc261c004cd4e1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our agent is an actor-critic, trained with the Policy Gradient method. For the agent architecture, we follow [39]. It" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "consists of a CNN encoder followed by an LSTM. As standard in RL, 4 frames are stacked, max-pooled, and the result is the input to the agent for a single time step." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 193 + ], + "type": "text", + "content": "Exploration-driven World Model Training. We initially pretrain GenieRedux-G on Platformers-200 and fine-tune on Platformers-50 to obtain the model GenieRedux-G-50. Then, we train AutoExplore Agent by using GenieRedux-G-50, using it as a source of reward. The details of training the agent are presented in Sup.Mat A.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 194, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 324 + ], + "type": "text", + "content": "Running the trained exploratory agent on a selected environment, we obtain a new diverse dataset with action demonstrations under unseen scenes. We first fine-tune the decoder of the Tokenizer for 1,000 iterations to adapt to the new unseen scenes. The Dynamic module of GenieRedux-G is then fine-tuned on the new data to achieve greater visual fidelity and controllability under new conditions. In order to build test sets to evaluate our approach, we train an Agent-57 model for each of the environments we explored, using the available environment rewards. More details on the test setup are provided in Sup.Mat A.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": "For visual fidelity evaluation, we use FID (Frechet inception distance) Heusel et al. [25], PSNR (signal-to-noise ratio) and SSIM (structural similarity index measure) Wang et al. [61]. To evaluate controllability, we use the recently proposed " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\Delta_t" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": " PSNR metric [5], which compares the visual effect of the ground truth action " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "(\\hat{x}_t)" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": " versus a random action " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "(\\hat{x}_t')" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\Delta_t" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": " PSNR = PSNR " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "(x_t, \\hat{x}_t) - \\mathrm{PSNR}(x_t, \\hat{x}_t')" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": " is the ground truth frame at time " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": ". A higher " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\Delta_t" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": " PSNR indicates a higher level of controllability. As in Bruce et al. [5], for all experiments we report " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\Delta_t" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": " PSNR with " + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "inline_equation", + "content": "t = 4" + }, + { + "bbox": [ + 313, + 325, + 553, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 454, + 394, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 394, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 394, + 467 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 665 + ], + "type": "text", + "content": "Comparing GenieRedux and GenieRedux-G. We implement the original CoinRun case study with a random agent, as advised by [5], in order to validate and compare GenieRedux with LAM, and GenieRedux-G which uses agent-provided actions instead. In this study, the presence of LAM is the only difference between the models. We first train on a dataset, collected by a random agent. Visual fidelity results are in Tab. 2. Our GenieRedux implementation exhibits high visual quality and matches all seven CoinRun environment actions, as well as progressing environment motions (demonstrated in Sup.Mat. F). However, as demonstrated by the metrics, GenieRedux-G shows superior visual fidelity and controllability (more in Sup.Mat. F), as it avoids the uncertainty of LAM prediction. This study demonstrates that even using a random agent can result in action performance abilities in the world model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Next, we train an actor-critic agent with PPO on the environment reward, following [13] to collect data and train GenieRedux-TA and GenieRedux-G-TA. Tab. 3 shows evaluation on a test set collected by a trained agent." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27601" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 104, + 272, + 195 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 103 + ], + "type": "text", + "content": "Table 2. Comparison of GenieRedux and GenieRedux-G on Basic Test Set. Performed on a test set, collected from the Coinrun environment with randomly sampled actions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 104, + 272, + 195 + ], + "lines": [ + { + "bbox": [ + 77, + 104, + 272, + 195 + ], + "spans": [ + { + "bbox": [ + 77, + 104, + 272, + 195 + ], + "type": "table", + "html": "
ModelBasic Test Set
FID↓PSNR↑SSIM↑
Tokenizer18.1438.250.96
LAM37.0133.970.92
GenieRedux21.8825.510.77
GenieRedux-G18.8833.410.92
", + "image_path": "1421fccd8406eabb69f6e540cce18ec6a7d912649df0cf3697cee326057cc86c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 251, + 279, + 360 + ], + "blocks": [ + { + "bbox": [ + 55, + 206, + 295, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 250 + ], + "type": "text", + "content": "Table 3. Comparison of GenieRedux and GenieRedux-G on Diverse Test Set. The models are trained with data collected by random agent and trained agent (-TA), and tested on data collected by a trained agent from the Coinrun environment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 251, + 279, + 360 + ], + "lines": [ + { + "bbox": [ + 70, + 251, + 279, + 360 + ], + "spans": [ + { + "bbox": [ + 70, + 251, + 279, + 360 + ], + "type": "table", + "html": "
ModelDiverse Test Set
FID↓PSNR↑SSIM↑
Tokenizer19.1335.850.94
Tokenizer-TA11.6340.620.97
GenieRedux23.9723.820.73
GenieRedux-G19.5131.660.90
GenieRedux-TA12.5731.970.90
GenieRedux-G-TA12.4034.440.92
", + "image_path": "ed3328ed038c23c673b790d1060122cf1de55bdddddf66ba9d8c236fb92fc74c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 367, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 439 + ], + "type": "text", + "content": "GenieRedux-G outperforms GenieRedux on all settings. Furthermore, models trained on diverse agent-collected data are visually superior to those trained on random agents. The higher " + }, + { + "bbox": [ + 55, + 367, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 367, + 296, + 439 + ], + "type": "text", + "content": "PSNR of 1.89 for GenieRedux-G-TA compared to 0.70 for GenieRedux-G shows the superiority of diverse data training in controllability. (more in Sup.Mat. F)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 635 + ], + "type": "text", + "content": "Multi-Environment Models. Here, we evaluate the models we initially train on many environments from RetroAct. GenieRedux-G-200 is pretrained on the Platformers-200 dataset for 180k iterations. On the validation set, we obtain 23.32 PSNR and 17.12 FID. Using this model as a base, GenieRedux-G-50 is trained on Platformers-50. Its quantitative evaluation on a test set of 10k sessions separately generated from the selected 50 environments is at the start of Tab. 4. As the 50 environments are selected with corresponding action controls between each other, we see a boost in the quality of prediction. Fig. 4 demonstrates that the instructed action is executed successfully by GenieRedux-G. As the up action is rarely used, it serves more as a no-operation action. (more in Sup.Mat C.1)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Ablation Study. In this experiment we evaluate the additive gain of each proposed improvement in GenieRedux-G - the additive token input and training with the Token Distance Cross-Entropy Loss. The ablation is performed on a generated test set of 10k sessions, each 500 frames long." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 85, + 539, + 160 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 549, + 82 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 549, + 82 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 549, + 82 + ], + "type": "text", + "content": "Table 4. Ablation study on improvements in GenieRedux-G." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 85, + 539, + 160 + ], + "lines": [ + { + "bbox": [ + 328, + 85, + 539, + 160 + ], + "spans": [ + { + "bbox": [ + 328, + 85, + 539, + 160 + ], + "type": "table", + "html": "
ModelFID↓PSNR↑SSIM↑
GenieRedux-G-20022.3125.110.80
GenieRedux-G-5023.8026.360.84
+ Token Input22.9626.650.84
+ TDCE Loss22.9527.060.85
Autoregressive22.1128.070.88
", + "image_path": "6d60a36887e6dd6f6b1d683d1e06205bb6ae41ef329c986fbe366e4f9606b45b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 166, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 321, + 166, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 321, + 166, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 321, + 166, + 545, + 289 + ], + "type": "image", + "image_path": "929d3064ff054f87b1bde004904d29cbd7e9311016ee6b0c408afc2702b211fc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 291, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 314, + 291, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 291, + 553, + 312 + ], + "type": "text", + "content": "Figure 4. Control Of GenieRedux-G-50. Demonstrating all controls of our multi-environment model on multiple games." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 321, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 393 + ], + "type": "text", + "content": "The data is collected using a random action policy from the environments in Platformers-50. Visual fidelity evaluation is provided in Tab. 4. It can be seen that each component gives our model a benefit in terms of visual fidelity. Finally, we perform an autoregressive evaluation of the best model to achieve our highest score." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 688 + ], + "type": "text", + "content": "Tokenizer Representation Study. This experiment provides insights into the inner workings of GenieRedux-G to motivate our proposed changes. As the Dynamics module operates entirely on the token representation, we examine it closely. Fig. 5 shows the reconstructions of an input sequence (first row) and the visualized token representation (last row), where each predicted token index is assigned a different color. The visual features of the first frame are captured by various tokens. Starting with the second frame, the representation drastically changes - a token is specialized in representing the static frame regions compared to the past, while all motion regions are updated with new content. Observing that visually similar patches predict identical or similar tokens, we replace each predicted token with its closest in the codebook. We only keep the special background token unchanged. In the second row of Fig. 5 we show the resulting reconstruction - while some blurriness appears, the image remains largely the same. Conversely, replacing each token with its furthest in the codebook (third row) results in a significantly different image. This property - closer tokens having more similar appearance - motivates our Token Distance Cross-Entropy Loss, which penalizes predicting tokens further away from the ground truth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Fig. 6 visualizes the uncertainty of GenieRedux-G-50 for each predicted token of its Dynamics module given a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27602" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 107, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 5. Quantitative Results on 3 environments. We evaluate the benefit of the data from the propose AutoExplore Agent to our models. GenieRedux-G-50 is our pretrained world model on 50 environments. GenieRedux-G-50-ft are fine-tuned models using data from a random agent or AutoExplore (Exploration). GenieRedux-G denotes a non-fine-tuned model, trained with the exploration data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 107, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 65, + 107, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 65, + 107, + 545, + 376 + ], + "type": "table", + "html": "
EnvironmentStrategyModelFID↓PSNR↑SSIM↑ΔPSNR↑
Adventure Island IIRandomGenieRedux-G-5041.9926.320.810.83
GenieRedux-G-50-ft42.3427.040.811.19
ExplorationTokenizer-ft11.0138.950.98-
GenieRedux-G11.9428.330.880.37
GenieRedux-G-50-ft12.7730.600.901.47
Random AutoregressiveGenieRedux-G-50-ft41.5527.820.831.24
Exploration AutoregressiveGenieRedux-G-50-ft11.3333.610.942.09
Super Mario BrosRandomGenieRedux-G-5029.8334.240.940.56
GenieRedux-G-50-ft30.1334.540.940.54
ExplorationTokenizer8.0942.000.99-
GenieRedux-G9.5634.000.950.09
GenieRedux-G-50-ft9.5536.130.970.57
Random AutoregressiveGenieRedux-G-50-ft30.8434.850.950.57
Exploration AutoregressiveGenieRedux-G-50-ft9.3337.770.970.76
SmurfsRandomGenieRedux-G-5079.5121.470.690.47
GenieRedux-G-50-ft80.6121.830.700.65
ExplorationTokenizer17.8635.610.98-
GenieRedux-G20.4335.420.800.85
GenieRedux-G-50-ft20.0127.450.851.55
Random AutoregressiveGenieRedux-G-50-ft80.1622.160.710.69
Exploration AutoregressiveGenieRedux-G-50-ft18.9729.530.902.06
", + "image_path": "ead44dfc4c7119736676acb777043924648f12ebfefac020552d1864115d73b7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 384, + 288, + 511 + ], + "blocks": [ + { + "bbox": [ + 62, + 384, + 288, + 511 + ], + "lines": [ + { + "bbox": [ + 62, + 384, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 288, + 511 + ], + "type": "image", + "image_path": "940b81b9ca8148245440c2acb56b6d04da92ce4710442b5c8cca7d6dff176dff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 514, + 295, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 559 + ], + "type": "text", + "content": "Figure 5. Tokenizer Representation. Reconstruction images from the tokenizer, and the effect of replacing each token with its closest and furthest in the codebook. Lastly, we visualize the indices of the predicted tokens." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 566, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 650 + ], + "type": "text", + "content": "sequence. The uncertainty metric is the entropy of the classification over 1024 codebook tokens. Tokens corresponding to motion have the highest uncertainty; other regions are mostly classified as the \"static\" token. Thus, minimal character movement yields low uncertainty, while forward motion increases it. This motivates us to build AutoExplore Agent's reward based on this uncertainty." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 715 + ], + "type": "text", + "content": "Exploration-based training. We demonstrate our exploration-based training of GenieRedux-G. We perform the procedure on 3 environments - AdventureIslandII, which provides an easy set-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 384, + 547, + 450 + ], + "blocks": [ + { + "bbox": [ + 320, + 384, + 547, + 450 + ], + "lines": [ + { + "bbox": [ + 320, + 384, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 320, + 384, + 547, + 450 + ], + "type": "image", + "image_path": "060b83ac3139b750e8665ce11bd220606dc67a266aa2f8636117757232268f03.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 453, + 555, + 486 + ], + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 486 + ], + "type": "text", + "content": "Figure 6. Dynamics Uncertainty. Shown is the uncertainty per token predicted for each image of an example sequence. Uncertainty is generated in the regions of motion." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 496, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 592 + ], + "type": "text", + "content": "ting for the agent to learn (single platform with no enemies at the start), SuperMario Bros provides an enemy and obstacles soon after the start and Smurfs provides a more complex background imagery and different action dynamics. For each of the environments, we train an AutoExplorer Agent. We observe that the agent learns to move forward and navigate obstacles to maximize reward. (more in Sup.Mat. D)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "We use our pretrained GenieRedux-G-50 model as a baseline and fine-tune it for each environment in two settings - a dataset collected on the selected environment by a random agent and by our AutoExplorer Agent. Each dataset consists of 10k sessions, each 700 frames long. We fine-tune (GenieRedux-G-50-ft) for 10k iterations and pick the best performing model. In our comparison, we also include a GenieRedux-G model trained from scratch on the diverse exploration datasets for 15k iterations to show the effect of pretraining. We perform single-pass generation for all mod" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27603" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 159 + ], + "type": "image", + "image_path": "a54e18eb289388484a0bdf631c9d312baeb3d6c92dbea5f64de9c04751bbe3fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 166, + 295, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 295, + 199 + ], + "type": "text", + "content": "Figure 7. AutoExplore Agent vs Random Agent Qualitative Comparison. We show that AutoExplore exhibits better visual quality and avoids losing track of the agent." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 217, + 293, + 281 + ], + "blocks": [ + { + "bbox": [ + 67, + 205, + 283, + 216 + ], + "lines": [ + { + "bbox": [ + 67, + 205, + 283, + 216 + ], + "spans": [ + { + "bbox": [ + 67, + 205, + 283, + 216 + ], + "type": "text", + "content": "Table 6. Comparison of AutoExplore Agent with others." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 217, + 293, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 293, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 293, + 281 + ], + "type": "table", + "html": "
AgentSuperMarioBrosAdventureIslandII
PSNR↓SSIM↓ΔPSNR↓PSNR↓SSIM↓ΔPSNR↓
RF28.580.940.18124.820.780.44
VAE24.400.860.08716.570.50.072
Ours23.810.850.06515.200.410.070
", + "image_path": "673bbf611cae3d740045fa8cc952e8ed184f5ee50dcdf73c6d7d76853e8986da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 289, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 479 + ], + "type": "text", + "content": "els and the more computationally heavy autoregressive evaluation for the fine-tuned models on data from random and AutoExplore Agent's datasets. Tab. 5 shows visual fidelity and controllability metrics for each environment, confirming the effectiveness of our exploration method. The model fine-tuned on AutoExplore Agent's data consistently outperforms the models trained on random actions in terms of visual fidelity. Exploration-based fine-tuning also improves controllability. Environments with small characters and uniform backgrounds can be more challenging for all models to learn. However, the gain in controllability in this case remains noticeable during autoregressive evaluation. Fig. 7 demonstrates the superior quality of our method. In addition, we observe that the multienvironment pretraining leads to significant gains in both studied aspects compared to the nonpretrained model. (more in Sup.Mat. C)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 579 + ], + "type": "text", + "content": "AutoExplore Agent Evaluation. We compare AutoExplore Agent with exploration-based methods in [6]. We train agents based on SSE of RF and VAE features on top of GenieRedux and compare with ours on Tab. 6. AutoExplore Agent's reward results in maximum world model visual and controllability errors (on 1k episodes of agent actions), fulfilling its intended role in our framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "User Studies. To validate the quality of our final results, we perform a user study in which we ask people to rate from 1 to 5 the quality of samples produced respectively by GenieRedux-G trained on random agent's data and on AutoExplore Agent's data. Each sample in our study consists of two 16-frame clips playing in a synchronized manner - the ground truth clip and our GenieRedux-G-50-ft reconstruction, given two initial frames and generating the rest autoregressively. We provide a total of 120 samples to the users - 40 samples per model and 40 samples of two ground" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 76, + 430, + 176 + ], + "blocks": [ + { + "bbox": [ + 318, + 76, + 430, + 176 + ], + "lines": [ + { + "bbox": [ + 318, + 76, + 430, + 176 + ], + "spans": [ + { + "bbox": [ + 318, + 76, + 430, + 176 + ], + "type": "image", + "image_path": "cd93a57c19be2f18949958e60622afee40ba84b0c64ccfd6c1b60bfa5e8ae4df.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 177, + 566, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 566, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 566, + 210 + ], + "type": "text", + "content": "Figure 8. User study results. Our user study on two games shows that our model trained with AutoExplore Agent's data is consistently rated higher." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 439, + 76, + 549, + 176 + ], + "blocks": [ + { + "bbox": [ + 439, + 76, + 549, + 176 + ], + "lines": [ + { + "bbox": [ + 439, + 76, + 549, + 176 + ], + "spans": [ + { + "bbox": [ + 439, + 76, + 549, + 176 + ], + "type": "image", + "image_path": "db754866630a6436c6985947509e3e762a1a58281c217489525370380681fb47.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 226, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 299 + ], + "type": "text", + "content": "truth samples, to establish scale. We give 20 samples from each of the two selected games - SuperMario Bros and AdventureIslandII. We get reviews from 19 participants. The results are shown in Fig. 8. The model, trained on data from AutoExplore Agent is clearly rated closer to the ground truth, establishing the quality of our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "type": "text", + "content": "With a second user study, we evaluate the action accuracy of the generated frames. We use ambiguous single input cases (character starting mid-air) and generate 60 clips with 3 actions on AdventureIslandII. Users prefer our exploration-trained model, rating it " + }, + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{0.75} \\pm \\mathbf{0.019}" + }, + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "type": "text", + "content": " on a scale from 0 (random preferred) to 1 (exploration preferred). (more in Sup.Mat. E.2)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 410, + 388, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 410, + 388, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 410, + 388, + 422 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 506 + ], + "type": "text", + "content": "As world models have developed into large models with impressive simulation properties, they require large interaction datasets, complete with diverse observations and actions. Genie [5] demonstrates impressive abilities by training on multiple environments, however, requiring the collection of a large video dataset and a model to infer actions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "In this work, we address the heavy burden of data collection and curation by building a new framework for training large world models by collecting interaction data from a large number of virtual environments. We first build an open implementation of Genie - GenieRedux and enhance it into its version GenieRedux-G. We obtain models exhibiting control by pretraining on a large set of virtual environments. We address the overfitting limitations of random data collection policy by proposing AutoExplore Agent, an agent entirely independent of the environment reward, maximizing the uncertainty of GenieRedux-G. After finetuning on the explored environment, our model is able to improve its visual fidelity and controllability much better than training solely on random agent's data. Demonstrating this on multiple environments, we show the potential of our framework to make training of next-generation world models more accessible, cost-effective, and effort-free." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27604" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "INSAIT, Sofia University \"St. Kliment Ohridski\". Partially funded by the Ministry of Education and Science of Bulgaria's support for INSAIT as part of the Bulgarian National Roadmap for Research Infrastructure. This project was supported with computational resources provided by Google Cloud Platform (GCP)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "type": "text", + "content": "[1] Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37: 58757-58791, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "text", + "content": "[2] Kamyar Azizzadenesheli, Emma Brunskill, and Animashree Anandkumar. Efficient exploration through bayesian deep q-networks. In 2018 Information Theory and Applications Workshop (ITA), pages 1-9. IEEE, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 295, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 340 + ], + "type": "text", + "content": "[3] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. Advances in neural information processing systems, 29, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 384 + ], + "type": "text", + "content": "[4] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253-279, 2013." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "type": "text", + "content": "[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "type": "text", + "content": "[6] Yuri Burda, Harri Edwards, Deepak Pathak, Amos Storkey, Trevor Darrell, and Alexei A Efros. Large-scale study of curiosity-driven learning. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 520 + ], + "type": "text", + "content": "[7] Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 295, + 566 + ], + "type": "text", + "content": "[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 295, + 600 + ], + "type": "text", + "content": "[9] Chang Chen, Yi-Fu Wu, Jaesik Yoon, and Sungjin Ahn. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "type": "text", + "content": "[10] Silvia Chiappa, Sébastien Racanière, Daan Wierstra, and Shakir Mohamed. Recurrent environment simulators. In International Conference on Learning Representations, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "text", + "content": "[11] Leshem Choshen, Lior Fox, and Yonatan Loewenstein. Dora the explorer: Directed outreach reinforcement action-selection. arXiv preprint arXiv:1804.04012, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In The Eleventh International Conference on Learning Representations, 2023." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[13] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In International conference on machine learning, pages 1282-1289. PMLR, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "text", + "content": "[15] Justin Fu, John Co-Reyes, and Sergey Levine. Ex2: Exploration with exemplar models for deep reinforcement learning. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "text", + "content": "[16] Scott Fujimoto, Herke Hoof, and David Meger. Addressing function approximation error in actor-critic methods. In International conference on machine learning, pages 1587-1596. PMLR, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 275, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 553, + 307 + ], + "type": "text", + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[18] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "type": "text", + "content": "[19] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "type": "text", + "content": "[20] David Ha, Jonas Jongejan, and Ian Johnson. Draw together with a neural network. Retrieved Oct, 5:2021, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 389, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 553, + 442 + ], + "type": "text", + "content": "[21] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[22] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "text", + "content": "[23] Danijar Hafner, Timothy P Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 535, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 567 + ], + "type": "text", + "content": "[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "text", + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "text", + "content": "[26] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27605" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In Proceedings of the Nineteenth International Conference on Machine Learning, pages 267-274, 2002." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[29] Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent experience replay in distributed reinforcement learning. In International conference on learning representations, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[30] Youngjin Kim, Wontae Nam, Hyunwoo Kim, Ji-Hoon Kim, and Gunhee Kim. Curiosity-bottleneck: Exploration by distilling task-specific novelty. In International conference on machine learning, pages 3379-3388. PMLR, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "type": "text", + "content": "[31] Martin Klissarov, Riashat Islam, Khimya Khetarpal, and Doina Precup. Variational state encoding as intrinsic motivation in reinforcement learning. In *Task-Agnostic Reinforcement Learning Workshop at Proceedings of the International Conference on Learning Representations*, pages 16–32, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "text", + "content": "[32] Lisa Lee, Benjamin Eysenbach, Emilio Parisotto, Eric Xing, Sergey Levine, and Ruslan Salakhutdinov. Efficient exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "text", + "content": "[33] Ian Lenz, Ross A Knepper, and Ashutosh Saxena. Deepmpc: Learning deep latent features for model predictive control. In Robotics: Science and Systems, page 25. Rome, Italy, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[34] Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[35] Marlos C Machado, Marc G Bellemare, and Michael Bowling. Count-based exploration with the successor representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5125-5133, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 480 + ], + "type": "text", + "content": "[36] Jarryd Martin, Suraj Narayanan Sasikumar, Tom Everitt, and Marcus Hutter. Count-based exploration in feature space for reinforcement learning. arXiv preprint arXiv:1706.08090, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "type": "text", + "content": "[37] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "text", + "content": "[38] Alberto Maria Metelli, Amarildo Likmeta, and Marcello Restelli. Propagating uncertainty in reinforcement learning via wasserstein barycenters. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[39] Vincent Micheli, Eloi Alonso, and François Fleuret. Transformers are sample-efficient world models. In The Eleventh International Conference on Learning Representations, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "text", + "content": "[40] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[41] Ian Osband and Benjamin Van Roy. Bootstrapped thompson sampling and deep exploration. arXiv preprint arXiv:1507.00300, 2015." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[42] Ian Osband, Charles Blundell, Alexander Pritzel, and Benjamin Van Roy. Deep exploration via bootstrapped dqn. Advances in neural information processing systems, 29, 2016." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[43] Ian Osband, Benjamin Van Roy, and Zheng Wen. Generalization and exploration via randomized value functions. In International Conference on Machine Learning, pages 2377-2386. PMLR, 2016." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "text", + "content": "[44] Ian Osband, John Aslanides, and Albin Cassirer. Randomized prior functions for deep reinforcement learning. Advances in Neural Information Processing Systems, 31, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "text", + "content": "[45] Georg Ostrovski, Marc G Bellemare, Aäron Oord, and Rémi Munos. Count-based exploration with neural density models. In International conference on machine learning, pages 2721-2730. PMLR, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[46] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International conference on machine learning, pages 2778-2787. PMLR, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[47] Deepak Pathak, Dhiraj Gandhi, and Abhinav Gupta. Self-supervised exploration via disagreement. In International conference on machine learning, pages 5062-5071. PMLR, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "type": "text", + "content": "[48] Mathieu Poliquin. Stable retro, a maintained fork of openai's gym-retro. https://github.com/Farama-Foundation/stable-retro, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "type": "text", + "content": "[49] Ofir Press, Noah Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[50] Jan Robine, Marc Höftmann, Tobias Uelwer, and Stefan Harmeling. Transformer-based world models are happy with 100k interactions. In The Eleventh International Conference on Learning Representations, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 521 + ], + "type": "text", + "content": "[52] Tom Schaul. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "text", + "content": "[53] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International conference on machine learning, pages 8583-8592. PMLR, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 580, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 634 + ], + "type": "text", + "content": "[54] Niranjan Srinivas, Andreas Krause, Sham Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: no regret and experimental design. In Proceedings of the 27th International Conference on International Conference on Machine Learning, pages 1015-1022, 2010." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[55] Haoran Tang, Rein Houthooft, Davis Foote, Adam Stooke, OpenAI Xi Chen, Yan Duan, John Schulman, Filip DeTurck, and Pieter Abbeel. # exploration: A study of count-based exploration for deep reinforcement learning. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[56] Ruo Yu Tao, Vincent François-Lavet, and Joelle Pineau. Novelty search in representational space for sample efficient" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27606" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 599 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "exploration. Advances in Neural Information Processing Systems, 33:8114-8126, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[57] William R Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3-4):285-294, 1933." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[58] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 186 + ], + "type": "text", + "content": "[59] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 186, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 252 + ], + "type": "text", + "content": "[60] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "text", + "content": "[61] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "text", + "content": "[62] Mingxing Xu, Wenrui Dai, Chunmiao Liu, Xing Gao, Weiyao Lin, Guo-Jun Qi, and Hongkai Xiong. Spatial-temporal transformer networks for traffic flow forecasting. arXiv preprint arXiv:2001.02908, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "text", + "content": "[63] Sherry Yang, Yilun Du, Seyed Kamyar Seyed Ghasemipour, Jonathan Tompson, Leslie Pack Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. In The Twelfth International Conference on Learning Representations, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[64] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "text", + "content": "[65] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 499, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 295, + 544 + ], + "type": "text", + "content": "[66] Lixuan Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Prelar: World model pre-training with learnable action representation. In European Conference on Computer Vision, pages 185-201, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "text", + "content": "[67] Tianjun Zhang, Paria Rashidinejad, Jiantao Jiao, Yuandong Tian, Joseph E Gonzalez, and Stuart Russell. Made: Exploration via maximizing deviation from explored regions. Advances in Neural Information Processing Systems, 34:9663-9680, 2021." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27607" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_content_list.json b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd49b63d49cac3943db6657fb3c7b19dcc247de2 --- /dev/null +++ b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_content_list.json @@ -0,0 +1,2035 @@ +[ + { + "type": "text", + "text": "Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiwei Yang $^{1,2}$ Yucong Meng $^{2,3}$ \nKexue Fu $^{4}$ Feilong Tang $^{1}$ Shuo Wang $^{2,3^{*}}$ Zhijian Song $^{1,2,3^{*}}$", + "bbox": [ + 220, + 200, + 782, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Academy for Engineering and Technology, Fudan University, Shanghai 200433, China \n $^{2}$ Shanghai Key Laboratory of Medical Image Computing and Computer Assisted Intervention \n $^{3}$ Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, China \n $^{4}$ Shandong Computer Science Center (National Supercomputer Center in Jinan)", + "bbox": [ + 125, + 241, + 870, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 347, + 326, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly Supervised Semantic Segmentation (WSSS) with image-level labels aims to achieve pixel-level predictions using Class Activation Maps (CAMs). Recently, Contrastive Language-Image Pre-training (CLIP) has been introduced in WSSS. However, recent methods primarily focus on image-text alignment for CAM generation, while CLIP's potential in patch-text alignment remains unexplored. In this work, we propose ExCEL to explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. Specifically, we propose Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules to improve the dense alignment across both text and vision modalities. To make text embeddings semantically informative, our TSE module applies Large Language Models (LLMs) to build a dataset-wide knowledge base and enriches the text representations with an implicit attribute-hunting process. To mine fine-grained knowledge from visual features, our VC module first proposes Static Visual Calibration (SVC) to propagate fine-grained knowledge in a non-parametric manner. Then Learnable Visual Calibration (LVC) is further proposed to dynamically shift the frozen features towards distributions with diverse semantics. With these enhancements, ExCEL not only retains CLIP's training-free advantages but also significantly outperforms other state-of-the-art methods with much less training cost on PASCAL VOC and MS COCO. Code is available at https://github.com/zwyang6/ExCEL.", + "bbox": [ + 88, + 378, + 483, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 804, + 220, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly Supervised Semantic Segmentation (WSSS) intends to generate pixel-level predictions using weak annotations like points [2], scribbles [18, 32], bounding boxes [8,", + "bbox": [ + 89, + 830, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7b90d3683d00401b2759d0fc96acc05aa2121d9bcfb5f1ccee7270597c57c67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 345, + 906, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d8ebe21c87dcf4ca0fe5bf33f1f13a18e5088fb87af9751159439d68c40fa7a.jpg", + "image_caption": [ + "Figure 1. Our motivation. (a) Previous methods leverage CLIP to generate CAMs with global image-text alignment, leaving CLIP's dense knowledge unexplored. (b) The proposed ExCEL explores CLIP's dense knowledge via a novel patch-text alignment paradigm, which generates better CAMs with less training cost." + ], + "image_footnote": [], + "bbox": [ + 517, + 436, + 906, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "16], or image-level labels [1, 23, 40]. It significantly reduces the annotating cost of fully supervised methods and has attracted increasing attention in the community. Among all cheap annotation types, most WSSS approaches leverage image-level labels to provide dense localization cues, linking visual concepts to specific pixel regions [5, 33]. In this work, we focus on WSSS with image-level labels as well.", + "bbox": [ + 511, + 628, + 906, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Commonly, the WSSS pipeline involves three stages: generating Class Activation Maps (CAMs) [46] by training a classification network, refining CAMs into pseudo labels [28], and using these labels to train a segmentation model [7]. However, due to the minimal semantic information from image-level labels, CAMs intend to highlight the most distinctive object parts, significantly limiting WSSS performance. Recently, Contrastive Language-Image Pre-training (CLIP) [24] has been introduced in WSSS. CLIMs [36] applies image-text pairs to regularize visual relations among different semantics. CLIP-ES [20]", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors. Email:{shuowang, zjsong}@fudan.edu.cn.", + "bbox": [ + 107, + 886, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20223", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leverages image-text alignment for gradient and produces high-quality GradCAM [30]. WeCLIP [43] further streamlines this process by using CLIP's visual encoder for segmentation. Despite these advancements, current methods primarily focus on CLIP's global image-text alignment, as shown in Fig. 1 (a). CLIP's dense knowledge with patch-text alignment still remains under-explored in WSSS.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose ExCEL to explore CLIP's dense knowledge via a patch-text alignment paradigm for WSSS, i.e., generating CAMs by calculating patch-wise similarity between text and individual patch tokens, as shown in Fig. 1 (b). We identify two key challenges: (1) Semantic sparsity in textual prompts, where the template 'a photo of [CLASS]' only indicates object presence but lacks knowledge for localization, and (2) Fine-grained insufficiency in visual features, as CLIP prioritizes global representation due to its image-text pairing nature. To address these, ExCEL enhances CLIP's dense alignment with Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules, unlocking its potential across text and vision modalities.", + "bbox": [ + 89, + 199, + 480, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate semantically rich text representations, we propose TSE through an implicit attribute feature space. Instead of relying on explicit text templates, TSE module implicitly constructs text embeddings using universal attributes across the dataset. We first employ Large Language Models (LLMs) to generate detailed descriptions for each class, which are then processed by CLIP's text encoder to build a dataset-wide knowledge base. Rather than directly fusing these class-specific descriptions for text prompting, we focus on clustering the descriptive embeddings into generalized attributes, which effectively capture complementary knowledge from other classes and supplement missing information for the target class. With this implicit feature space, we enhance the text embeddings by hunting for its most relevant attribute features and aggregate them into the final class-specific text representation. This approach enables TSE to generate more informative text embeddings, providing a strong foundation for visual recognition.", + "bbox": [ + 89, + 398, + 482, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mine fine-grained knowledge from visual features, we propose VC to calibrate CLIP in both non-trainable and efficient-learnable ways. Our findings suggest that CLIP's q-k attention loses fine-grained details. Therefore, we first propose a Static Visual Calibration (SVC) module to replace the suboptimal q-k attention with a straightforward Intra-correlation operation. It focuses on extracting fine-grained details from intermediate layers, which progressively propagates fine-grained visual knowledge. Without any retraining, SVC generates CAMs comparable to training-required WSSS methods. Building on this, we further propose a Learnable Visual Calibration (LVC) module to dynamically calibrate CLIP's frozen features. LVC extracts spatial correlations from SVC's static CAMs. These correlations further supervise a lightweight adapter to learn", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the dynamic shift, pushing frozen features towards spatial-aware distributions. LVC and SVC complement each other, enabling precise patch-text alignment for CAM generation.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work are listed as follows:", + "bbox": [ + 535, + 137, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. The proposed ExCEL generates better pseudo labels in both training-free and efficient learning manners, revealing the dense capabilities of CLIP for efficient CAM generation.", + "- To enhance patch-text alignment, we propose the Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules. TSE applies LLMs to build a dataset-wide knowledge base and treats text prompting as an implicit attribute-hunting process, making text embeddings more informative. VC propagates the fine-grained visual knowledge in a non-parametric manner and further dynamically calibrates the frozen features with a lightweight adapter. TSE and VC work across two modalities, generating better dense alignment and pseudo labels.", + "- Extensive experiments on PASCAL VOC and MS COCO demonstrate that ExCEL significantly outperforms recent state-of-the-art methods, while reducing training cost with only 3.2 GB of GPU memory and $6\\%$ of the training time required by recent methods." + ], + "bbox": [ + 511, + 152, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 470, + 663, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Weakly Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 496, + 880, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly supervised semantic segmentation with image-level labels typically relies on CAMs to provide dense supervision for segmentation [42, 44]. However, CAMs usually highlight the most discriminative parts of objects [34]. To address this issue, considerable efforts have been made from many intriguing insights. MCTformer [37] incorporates multiple class tokens in Vision Transformer and proposes generating CAMs from class-patch attention. ToCo [29] proposes token contrast learning and generates more precise CAMs. SeCo [39] designs a separate and conquer scheme and succeeds in tackling co-occurrence. Despite these advancements, prior methods commonly require retraining the entire classification network for CAM generation. In this work, our ExCEL directly generates CAMs from frozen CLIP, and further boosts its quality via a lightweight adapter, significantly reducing the training cost.", + "bbox": [ + 511, + 518, + 906, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision-Language Pre-training", + "text_level": 1, + "bbox": [ + 511, + 787, + 779, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Language Image Pre-training (CLIP) [24], known for pretraining on billion-scale image-text pairs, has demonstrated remarkable transferability in many downstream tasks. CoOP [48] and CLIP-Adapter [12] incorporate lightweight trainable parameters into CLIP and succeed in few-shot classification. DenseCLIP [25] and", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20224", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e36c53b2f5353223a5823b875eba234946566a301e81563cf57415f6f9f75e0.jpg", + "image_caption": [ + "Figure 2. ExCEL Architecture. We explore CLIP's dense knowledge with Text Semantic Enrichment (TSE) and Visual Calibration (VC). (a) TSE uses LLMs to build a knowledge base and clusters it into an implicit attribute space. The final text representation $T_{c}$ is enhanced by hunting for relevant attributes. For vision modality, (b) we introduce Static Visual Calibration (SVC) to calibrate visual features using the Inter-correlation operation across $N$ intermediate layers. It generates static CAMs with $T_{c}$ and calibrated features $P_{s}$ . (c) Learnable Visual Calibration (LVC) designs a learnable adapter to add a dynamic shift $R$ to SVC. It generates optimized features $P_{d}$ based on static CAMs guidance, creating dynamic CAMs from $P_{d}$ and $T_{c}$ . Dynamic CAMs are refined for segmentation supervision. Details are in Sec. 3.1." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 897, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MaskCLIP [47] leverage the alignment between text and vision modalities for the dense segmentation task. Recently, some studies have introduced CLIP into WSSS. CLIMS [36] treats image-text pairing of CLIP as regularization and leverages it to regularize the visual concepts. CLIP-ES [20] finds that the image-text alignment of CLIP generates class gradient and leverages it for GradCAM generation. WeCLIP [43] further streamlines this process and directly leverages CLIP's visual encoder for segmentation. However, these methods mainly focus on a global image-text alignment while ignoring the dense capabilities of CLIP. In contrast, our ExCEL explores CLIP's dense knowledge via a patch-text alignment paradigm for WSSS.", + "bbox": [ + 88, + 444, + 485, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 652, + 225, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 678, + 230, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patch-text CAM Generation. CLIP uses image and text encoders to project image and text into the same feature space, enabling robust vision-language alignment. In this work, we utilize this property to generate CAM in a patch-text alignment paradigm. Given the encoded text embeddings $T \\in \\mathbb{R}^{D \\times C}$ and visual features $P \\in \\mathbb{R}^{h \\times w \\times D}$ , where $D$ is the feature channel, $h$ and $w$ are the spatial sizes, $C$ is the number of classes, we generate CAM by calculating the patch-wise similarities between text and visual features:", + "bbox": [ + 89, + 700, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C A M} = \\operatorname {N o r m} (\\cos (P, T), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 845, + 482, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{Norm}(\\cdot)$ is the min-max normalization and $\\cos (\\cdot)$ is the cosine similarity calculation. However, due to", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the image-level pairing nature, CLIP suffers from textual semantic sparsity and visual fine-grained insufficiency. Therefore, our ExCEL incorporates TSE and VC (SVC and LVC) into CLIP to further explore its dense potential.", + "bbox": [ + 511, + 444, + 906, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Framework Overview. Our ExCEL generates CAMs in both training-free and efficient learning manners, as shown in Fig. 2. Its training pipeline is generalized as follows:", + "bbox": [ + 511, + 505, + 906, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Enriching textual semantics via TSE. We first use GPT-4 to generate descriptions for each class, which are encoded into a dataset-wide knowledge base with CLIP's text encoder. We cluster this knowledge into class-agnostic attributes and use the global text prompt to hunt for its most relevant ones. They are then aggregated into the final text representation. (2) Static CAM generation via SVC. We replace CLIP's q-k self-attention with our Intra-correlation operation from intermediate layers. Then the calibrated visual features and enhanced text embeddings are used for static CAMs via Eq. (1). (3) Dynamic CAM generation via LVC. A lightweight adapter is designed to learn dynamic token relations from static CAMs. The relations are added to SVC and serve as a distribution shift to make the visual features more diverse. The dynamic CAMs are generated with the enhanced text embeddings and LVC features via Eq. (1). (4) Segmentation training. Dynamic CAMs are refined to pseudo labels for segmentation supervision.", + "bbox": [ + 511, + 551, + 908, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Text Semantic Enrichment", + "text_level": 1, + "bbox": [ + 511, + 832, + 756, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Base Construction. The global text template 'a clean origami of [CLASS]' only indicates the presence of objects while limited providing dense knowledge for", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20225", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "patch-wise visual recognition. To enrich the text representation, we first adapt LLMs, such as GPT-4 to generate detailed class descriptions, as shown in Fig. 2 (a). Specifically, given the global template $E_{c}$ from class label space $Y \\in \\{1,2,\\dots,C\\}$ , where $c$ is the class index and $C$ is the number of categories, we carefully construct instructions for GPT: \"List $n$ descriptions with key properties to describe the [CLASS] in terms of appearance, color, shape, size, or material, etc. These descriptions will help visually distinguish the [CLASS] from other classes in the dataset. Each description should follow the format: 'a clean origami [CLASS]. it + descriptive contexts.' With this instruction, we ask GPT to generate $n$ detailed descriptions for each class, which are subsequently encoded into a dataset-wide knowledge base with CLIP's text encoder. The knowledge base is denoted as $\\mathcal{T} = \\{\\Phi(e_i)\\}_{i=1}^{n \\times C}$ . $e_i$ is the description from GPT, $\\Phi(\\cdot)$ is CLIP's text encoder and $n \\times C$ is the number of all descriptions. This knowledge base gathers descriptive properties for the whole dataset, building a strong foundation for the textual category representation.", + "bbox": [ + 89, + 90, + 483, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implicit Attribute Hunting. Instead of explicitly merging class-related knowledge into a single text embedding, we cluster this knowledge into generalized attributes and treat text prompting as an implicit attribute-hunting process. This has two main benefits: (1) $n$ explicit descriptions may still be limited in covering all characteristics of the class. The clustered attributes efficiently capture shared contextual knowledge from other categories, supplementing missing information for target class recognition. (2) The generated descriptions inevitably contain redundant or noisy content. The use of attributes makes the knowledge more compact and representative, leading to precise text prompting.", + "bbox": [ + 89, + 393, + 483, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we leverage a clustering algorithm to generate multiple centroids based on the knowledge base. Each cluster centroid is viewed as the implicit attribute that represents a group of descriptions sharing similar properties:", + "bbox": [ + 89, + 574, + 483, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {K m e a n s} (\\mathcal {T}, B) = \\left\\{a _ {i} \\right\\} _ {i = 1} ^ {B}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 643, + 480, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ is the number of centroids and kmeans algorithm [21] is used for clustering for simplicity. $a_{i}\\in \\mathbb{R}^{D\\times 1}$ represents the cluster centroid, i.e., the implicit attribute.", + "bbox": [ + 89, + 674, + 482, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the attribute feature space $A$ , we first send the global template $E_{c}$ into the text encoder of CLIP to generate global text embedding $t_c \\in \\mathbb{R}^{D \\times 1}$ . $D$ is the channel dimension. Then $t_c$ is leveraged to search for its most relevant attributes in the implicit attribute space. To exclude irrelevant attributes, we further propose to select TOPK attribute neighbors based on the similarity scores with $t_c$ :", + "bbox": [ + 89, + 720, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {c} = \\left\\{a _ {j}: j \\in \\operatorname {a r g m a x} _ {\\mathrm {T O P K}} \\left\\{t _ {c} ^ {T} a _ {j} \\right\\} _ {j = 1} ^ {B} \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 835, + 482, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we gently aggregate the implicit attribute neighbors according to the corresponding similarity weights to", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$t_c$ and take the aggregated features as the complementary knowledge for textual semantic enrichment. The final text representation $T_{c}\\in \\mathbb{R}^{D\\times 1}$ is denoted as:", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {c} = t _ {c} + \\lambda \\sum_ {j = 1} ^ {K} \\operatorname {s o f t m a x} \\left(t _ {c} ^ {T} A _ {c}\\right) a _ {j}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 143, + 906, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is the factor to balance the attribute information.", + "bbox": [ + 511, + 191, + 890, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Visual Calibrations", + "text_level": 1, + "bbox": [ + 511, + 214, + 699, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Visual Calibration. Due to the image-text pairing nature of CLIP, the visual features lack fine-grained information, leading to unreasonable localization maps via patch-text alignment. To delve into this, let us review the self-attention mechanism of the original CLIP first. As shown in Fig. 2 (b), given the input image $X \\in \\mathbb{R}^{3 \\times \\mathcal{H} \\times \\mathcal{W}}$ we send it to the image encoder of CLIP, which includes 12 Attention layers in this work. $\\mathcal{H} \\times \\mathcal{W}$ is the image size. For the feature $F_{l} \\in \\mathbb{R}^{D_{s} \\times hw}$ from $l$ -th layer of CLIP, it is first projected into three different spaces $\\{q, k, v\\}$ , named query, key, and value, respectively. $q, k$ and $v$ have the same shape of $D_{s} \\times hw$ , where $D_{s}$ and $hw$ represents the channel dimension and sequence length. Then the attention map between $q$ and $k$ is calculated by measuring the similarity:", + "bbox": [ + 511, + 237, + 906, + 449 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} (q, k) = \\operatorname {s o f m a x} \\left(q ^ {T} k / \\sqrt {D _ {s}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 455, + 906, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{SA}(\\cdot)$ is the calculation of attention. Then the output features $F_{l + 1}$ are generated by aggregating the tokens of $v$ according to similarity weights in the attention map.", + "bbox": [ + 511, + 486, + 905, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, due to the inherent image-text alignment of CLIP, the original q-k attention produces overly uniform attention maps, homogenizing diverse tokens from $v$ to capture broad semantics for global image representation (see discussions in Sec. 4.4). It leads to inaccurate object recognition. MaskCLIP [47] holds a similar observation, supporting this claim by removing the final q-k attention layer and using $v$ from the last layer as the visual output to preserve diversity. In our work, we choose to replace the suboptimal q-k attention with a straightforward Intra-correlation operation and focus on extracting fine-grained details from intermediate layers. This non-parametric approach effectively mines spatial semantics in intermediate $\\{q, k, v\\}$ and avoids the smoothing effect of q-k attention, resulting in more consistent attention maps and improved object localization.", + "bbox": [ + 511, + 532, + 905, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, instead of generating q-k correlation, Intra-correlation calculates the attention within each space of $\\{q,k,v\\}$ across intermediate layers. The attention map $S_{attn}^{l}\\in \\mathbb{R}^{hw\\times hw}$ from $l$ -th SVC layer is generated by:", + "bbox": [ + 511, + 758, + 906, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {a t t n} ^ {l} = \\sum w _ {i} \\operatorname {S A} \\left(O _ {i} ^ {l}, O _ {i} ^ {l}\\right), O _ {i} ^ {l} \\in \\left\\{q ^ {l}, k ^ {l}, v ^ {l} \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 827, + 906, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ is the contribution weight for different correlation maps. $l\\in \\{12 - N,\\dots,12\\}$ and $N$ is the number of intermediate layers for this operation. Then $S_{attn}^{l}$ and $v^{l}$ are used to", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20226", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generate the output features. Finally, the calibrated features $P_{s} \\in \\mathbb{R}^{D \\times h \\times w}$ from the last layer is used to generate static CAM $CAM_{s}$ with text embedding $T_{c}$ via Eq. (1).", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learnable Visual Calibration. Although ExCEL generates comparable CAMs without training, its performance is still limited by the fixed features in CLIP. To further unleash the dense potential of CLIP, we design a lightweight adapter to dynamically calibrate the visual features with diverse details. This adapter only incorporates a distribution shift to calibrate the fixed features without changing CLIP's pre-trained weights, thereby retaining CLIP's transferability and enhancing its dense performance for WSSS.", + "bbox": [ + 88, + 136, + 482, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 2 (c), frozen features $F_{l}$ from 1-12th layer of CLIP are extracted to learn a dynamic feature $F_{d}$ via the adapter. The process is expressed as:", + "bbox": [ + 89, + 272, + 483, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {d} = \\operatorname {C o n v} (\\operatorname {C o n c a t e} [ \\delta_ {l} (F _ {l}) ] _ {l = 1} ^ {1 2}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 325, + 482, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{d}\\in \\mathbb{R}^{D_{d}\\times hw}$ $D_{d}$ is the channel dimension. $\\mathrm{Conv}(\\cdot)$ is the convolution layer, $\\mathrm{Concat}[\\cdot ]$ is the concatenate operation that connects all features along channel dimension, and $\\delta_l(\\cdot)$ is the individual MLP layer for each $F_{l}$ Then $F_{d}$ is used to generate dynamic token relations by:", + "bbox": [ + 89, + 354, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr = \\alpha (\\cos (F _ {d}, F _ {d}) - \\beta \\overline {{\\cos (F _ {d} , F _ {d})}}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 440, + 482, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r \\in \\mathbb{R}^{hw \\times hw}$ , $\\alpha$ and $\\beta$ are the scaling and shifting factors to adjust the relations, respectively. $\\overline{\\cos(F_d, F_d)}$ means the mean value of similarity scores of $F_d$ . It is designed to remove the irrelevant relations in low values by:", + "bbox": [ + 89, + 467, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i j} = \\left\\{ \\begin{array}{l l} r _ {i j}, & \\text {i f} r _ {i j} \\geq 0 \\\\ - i n f, & \\text {e l s e} \\end{array} . \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 539, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the dynamic relation $R \\in \\mathbb{R}^{hw \\times hw}$ , we add it as a distribution bias to the static attention map $S_{attn}^{l}$ , dynamically grouping the frozen tokens within related semantics and shifting the features towards denser distribution. The optimized attention map $L_{attn}^{l} \\in \\mathbb{R}^{hw \\times hw}$ is denoted as:", + "bbox": [ + 89, + 589, + 483, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a t t n} ^ {l} = S _ {a t t n} ^ {l} + \\operatorname {s o f t m a x} (R). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 674, + 482, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we extract the dynamically calibrated features $P_{d} \\in \\mathbb{R}^{D \\times h \\times w}$ from the last layer of LVC and generate dynamic CAMs with Eq. (1), which are then refined to final pseudo labels $M_{d}$ for the segmentation.", + "bbox": [ + 89, + 703, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Objectives", + "text_level": 1, + "bbox": [ + 89, + 772, + 279, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We formulate a diversity loss to supervise the learning of $F_{d}$ in our LVC module. We first measure token correlations of $F_{d}$ by calculating the self similarity: $\\hat{\\mathcal{R}} = \\mathrm{sigmoid}(\\cos (F_d,F_d))$ , where sigmoid(·) is the activation function and $\\hat{\\mathcal{R}}\\in \\mathbb{R}^{hw\\times hw}$ . Then we refine $CAM_{s}$ from SVC into static pseudo labels $M_{s}$ and leverage its pixelwise affinity to guide the diversifying of $\\hat{\\mathcal{R}}$ . Specifically, if", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the pixel with coordinate $(i,j)$ shows the same pixel value as the pixel in $(\\varepsilon,\\eta)$ , the token pair with the same coordinates on $F_{d}$ is semantically related and its corresponding correlation logit on $\\hat{\\mathcal{R}}$ should be maximized, and vice versa. The diversity loss can be formulated as:", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i v}} = \\frac {1}{N ^ {+}} \\sum_ {u ^ {+} \\in \\hat {\\mathcal {R}} ^ {+}} (1 - u ^ {+}) + \\frac {1}{N ^ {-}} \\sum_ {u ^ {-} \\in \\hat {\\mathcal {R}} ^ {-}} u ^ {-}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 176, + 903, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N^{+} / N^{-}$ are the number of positive and negative pairs, $u^{+} / u^{-}$ are the positive and negative relation logits, and $\\hat{\\mathcal{R}}^{+} / \\hat{\\mathcal{R}}^{-}$ are the positive and negative sets of logit on $\\hat{\\mathcal{R}}$ , respectively. The diversity loss groups tokens with similar semantics and suppresses irrelevant ones, enhancing fine-grained details of visual features for precise text response.", + "bbox": [ + 511, + 226, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, ExCEL is streamlined as a single-stage method. We adopt a lightweight Transformer-based segmentation head [43] and directly take the frozen visual encoder of CLIP for segmentation. The dynamic pseudo labels are used as supervision with a cross-entropy loss $\\mathcal{L}_{seg}$ . The loss objectives of our ExCEL are formulated as:", + "bbox": [ + 511, + 316, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {E X C E L}} = \\mathcal {L} _ {s e g} + \\gamma \\mathcal {L} _ {\\text {d i v}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 420, + 903, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the weight factor. By efficiently training the adapter and a segmentation head, ExCEL achieves strong WSSS performance and significantly reduces training cost.", + "bbox": [ + 511, + 446, + 903, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 511, + 506, + 748, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 531, + 720, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics. The proposed ExCEL is evaluated on PASCAL VOC 2012 [11] and MS COCO 2014 [19]. VOC contains 21 categories (1 for background). Following prior methods [10, 17, 39], the augmented dataset with 10,582, 1,449, and 1,456 images are used for training, validating, and testing, respectively. COCO includes 81 classes, in which 82,081 images are used for training and 40,137 images are for validating. Mean Intersection-Over-Union (mIoU) is used as the main evaluation metric.", + "bbox": [ + 511, + 551, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. CLIP model with ViT-B [9] is used as ExCEL's encoder, which is frozen during the training. For TSE module, we generate $n = 20$ descriptions from GPT-4 for each category. The number of attribute embeddings $B$ is set to 112 and 224 for PASCAL VOC and MS COCO, respectively. The SVC module is conducted in the last $N = 5$ Transformer layers. Our decoder adopts a simple Transformer-based head [43]. Features $F_{l}$ from each layer of CLIP are sent to it for the segmentation predictions. The scaling and shifting factors in Eq. (8) are set as 3.0 and 1.0, respectively. The loss weight $\\gamma$ is set as 0.1. Following [29, 35, 39], the AdamW optimizer is used for training the adapter and decoder. The learning rate is 1e-4 with a weight decay of 1e-2. The training iteration is set", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20227", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5301420fc8865447c8c1f658f64629bb88710e6c0b272f98be0b0cf98101ae9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 101, + 181, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3cf356d6bcb85be39af554e85aa9ae60da491f5ae8199b38fed976f53559b185.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 101, + 267, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8341b6c0a8a3382c0fb8af763250a9d559aec1198a864dccbdcd7b63a33a9513.jpg", + "image_caption": [ + "PASCAL VOC" + ], + "image_footnote": [], + "bbox": [ + 272, + 101, + 356, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22f512584c0be30c1ffaf594f4585075fc84962a1609bf59e9051c09a36a6978.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 101, + 444, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/851349b2259265cdcc9d4f7cb4c9d72faa4887a1d9dbcaa8204e7a4789403436.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 101, + 534, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/316efb51d73ee1432a28e419b77b6f213e78dc03a679cd6105c8eeffdb57b608.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 146, + 179, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f74f7342c9bd38e85ccccf988a630ed6df648ccd4e7c37874fc3e2ed420d5b86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 146, + 267, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46ba46418105780174965328e844f1039f7cf6ba6bb845e56f6e9bdef5d4e93c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 146, + 356, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2bb1f5eea6c8a12e3505f943f800f9431b6414b16d8c31bb778c1dccb2be748.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 146, + 444, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5114db16ca06ec6772bada96533feaa21babdba9f61e9360cb97d41fa12b449.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 146, + 534, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0820de22073960a29f5d52de03d08656c62463f095e9caf7cb7da0c246c620d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 191, + 179, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/388bc4a2bd3c0cf2444b90089f1a7164064a2a0ed4260df95d41bf99e22ac2ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 191, + 267, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bd7a6d018eedd9f55a445e37adfe8ba060d1495a084df2b77b4e9b12e5fa72f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 191, + 356, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3599e883cd521f0c40f4618ba859159582be2d9bb94e3ce37c7ed5c27da298ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 191, + 444, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a099ccafe16b9d68dd8d4398051f24a7d45a477a1015a2f1c8922a759d0bec7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 191, + 534, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/853857e3c93474f209433edaa13ebad39e31788efb3e014e2004141d2bb751af.jpg", + "image_caption": [ + "Figure 3. Segmentation visualizations of SeCo [39], WeCLIP [43] and ours on VOC and COCO. ExCEL segments objects more precisely." + ], + "image_footnote": [], + "bbox": [ + 96, + 234, + 179, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9c64ba882624693b3d4e3de98f783de003f855a3c4275e4711672fb1fe2215f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 234, + 267, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0c6051f088c68437c8d0d4e19df66cf6194ca4704b147958ca5c96eae17bf64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 234, + 356, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/407257d2cc35efd54090f414c835ffd10fd347c1c94e2696287b9da91c60fe8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 234, + 444, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cb732dedb5053b10e25b17de0312daee744fe05fc297cd9680fa2b5c269fdff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 234, + 534, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c8a2dde2dc5e2bb3062dc8d1fc52dc7e11c6f4da1b8a2a94e557f48afb7cf0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 89, + 633, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/615e9b4bfb1296f453373dbb8da005a45ce6215b90b367eb64649975ad83987e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 89, + 723, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c18013dc17f10be6b7fc5d01c0817a31a5a6306335896bafe651c90f39a8bd69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 101, + 813, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9a592c946fab0c5e254996fbb1027530458ab6932689dc33850d902fafe20365.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 101, + 900, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0c07d1dd506f1791c29c140bcee28232b00df3ebd8b1fbfe3a347a3d6a5507a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 146, + 635, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c826c87e4fd50bf112a593962731b2eaf9d2af5fb9b2a326ef96ea612e58e3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 146, + 723, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1abe07791c3f893850d03b10d98dc161b7e98d42d36f411662827af52d8210ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 146, + 813, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb5bb2657738c3517c6dada1ca94ff908721c365d9d43527dd77033d7b93b5b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 146, + 900, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f66e0660334b6e3fd8f15cdd3314761408886b96fe8ff5b0cdfdfac188c1fe8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 191, + 635, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a5931862b2a698f9ebdd3ba5c5b52658585c531040dbc909fa7994f6d456f47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 191, + 723, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f610f88abaee40af22c521424154630cbceb42d45376f19195aa278bc495fc71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 191, + 813, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc82a5c0ee1e47840032587580b809eff34cfb26483364aad4df08cb97eaef3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 191, + 900, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4740cf6ccbaabd038a40da72d990c0341534e1a41b2dc24a286ebeac55b86433.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 234, + 635, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a70299b32eb9c9d4b781b775f287a4ed3775196df9a81909637a218f483f6eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 234, + 723, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6c7020eaf13999ff40c738a94b1b660907037ea78564be246790d95fe8eb66a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 234, + 813, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66e6889670bb1e0db1b5f0b19616a18f0e8840e66bf66004b44553c71ff4802c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 234, + 900, + 279 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/28a7a25060e08e8f4dca1015bbfe51829f5c02bb76c80736e0320827a63f6999.jpg", + "table_caption": [ + "Table 1. Segmentation comparisons on VOC and COCO. Net. is the backbone for segmentation. Sup. is the supervision type. $\\mathcal{I}$ image-level labels. SA: saliency maps. $\\mathcal{L}$ : language." + ], + "table_footnote": [], + "table_body": "
MethodSup.Net.VOCCOCO
ValTestVal
Multi-stage WSSS methods.
L2G [14] CVPR'2022I + SARN10172.171.744.2
RCA [49] CVPR'2023I + SARN3872.272.836.8
OCR [7] CVPR'2023IRN3872.772.042.5
BECO [26] CVPR'2023IRN10173.773.545.1
MCTformer+ [38] TPAMI'2024IRN3874.073.645.2
CTI [41] CVPR'2024IRN10174.173.245.4
CLIMS [36] CVPR'2022I + LRN10170.470.0-
CLIP-ES [20] CVPR'2023I + LRN10172.272.845.4
PSDPM [45] CVPR'2024I + LRN10174.174.947.2
CPAL [31] CVPR'2024I + LRN10174.574.746.8
Single-stage WSSS methods.
AFA [28] CVPR'2022IMiT-B166.066.338.9
ViT-PCM [27] ECCV'2022IViT-B70.370.9-
ToCo [29] CVPR'2023IViT-B71.172.242.3
DuPL [35] CVPR'2024IViT-B73.372.844.6
SeCo [39] CVPR'2024IViT-B74.073.846.7
DIAL [13] ECCV'2024I + LViT-B74.574.944.4
WeCLIP [43] CVPR'2024I + LViT-B76.477.247.1
ExCEL(w/o CRF)I + LViT-B77.277.349.3
ExCEL (Ours)I + LViT-B78.478.550.3
", + "bbox": [ + 94, + 357, + 486, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as 30,000 for VOC and 100,000 for COCO. Please refer to Supplementary Materials for more details.", + "bbox": [ + 89, + 670, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 712, + 462, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance of Semantic Segmentation. Tab. 1 shows segmentation comparisons between our ExCEL and recent methods on VOC and COCO. The single-stage ExCEL achieves $78.4\\%$ and $78.5\\%$ mIoU on VOC val set and test set, which even significantly outperforms the sophisticated multi-stage methods by at least $3.9\\%$ and $3.6\\%$ mIoU, respectively. For more complicated benchmark COCO, ExCEL achieves $50.3\\%$ mIoU on val set, which brings a noticeable $3.2\\%$ increase over the CLIP-based state-of-the-art (SOTA) WeCLIP. In addition, without time-consuming post-processing techniques, such as CRF [15], ExCEL still", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/79e2d5d12d4ed2e1c483c616f844662fbea26252cc1e980b488925d82dec3ac6.jpg", + "table_caption": [ + "Table 2. CAM seed comparisons on VOC train set. $\\mathcal{M}$ : multistage methods. $S$ : single-stage methods. †: our reproduction following official codes. ExCEL*: ExCEL in a training-free manner." + ], + "table_footnote": [], + "table_body": "
MethodTypeSup.Net.VOC
Train
Training-free WSSS methods.
CLIP-ES [20] CVPR'2023MI + LViT-B70.8
ExCEL* (Ours)SI + LViT-B74.6
Training-required WSSS methods.
ReCAM [6] CVPR'2022MIRN10154.8
FPR [3] CVPR'2023MIRN10163.8
LPCAM [5] CVPR'2023MIRN5065.3
MCTformer+ [38] TPAMI'2024MIRN3868.8
SFC [44] AAAI'2024MIRN10164.7
CTI [41] CVPR'2024MIRN10169.5
AFA [28] CVPR'2022SIMiT-B165.0
ViT-PCM [27] ECCV'2022SIViT-B67.7
†ToCo [29] CVPR'2023SIViT-B71.6
†DuPL [35] CVPR'2024SIViT-B75.0
SeCo [39] CVPR'2024SIViT-B74.8
CLIMS [36] CVPR'2022MI + LRN10156.6
POLE [22] WACV'2023MI + LRN5059.0
CPAL [31] CVPR'2024MI + LRN10171.9
DIAL [13] ECCV'2024SI + LViT-B75.2
†WeCLIP [43] CVPR'2024SI + LViT-B75.4
ExCEL (Ours)SI + LViT-B78.0
", + "bbox": [ + 519, + 358, + 903, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "maintains consistent superiority over SOTAs with CRF.", + "bbox": [ + 511, + 672, + 879, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative comparisons on VOC and COCO are shown in Fig. 3. By densely matching the patches and texts, ExCEL consistently demonstrates more precise object segmentation than recent methods in an image-text paradigm.", + "bbox": [ + 511, + 688, + 903, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation of CAM Seeds. Tab. 2 reports the quality of raw CAM seeds on VOC train set. Compared with recent methods, ExCEL achieves $74.6\\%$ mIoU in a training-free setup, outperforming CLIP-ES by $3.8\\%$ and performing comparably to most training-required methods. With the optimized LVC module, ExCEL further boosts CAM quality to $78.0\\%$ , surpassing SOTAs in the image-text paradigm by at least $2.6\\%$ . In addition, visual comparisons in Fig. 4 (e-h) also plainly illustrates that ExCEL generates better CAMs with the designed patch-text alignment paradigm.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20228", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6fd9378d29824e2ce862dab5fc4b6daaa14aea99033ac822c2f0c8a6a7b0851.jpg", + "image_caption": [ + "Figure 4. CAM visualizations on VOC train set. (a) Image. (b-e) Ablative visualizations of proposed modules. (e-h) Qualitative comparisons of (e) ExCEL and recent CLIP-based methods, i.e., (f) WeCLIP [43], (g) CLIP-ES [20] and (h) MaskCLIP [47]. (i) Ground truth." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 903, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/97dcac90ecc71ccd31aca4fd613d8bfd71403ef11362dbe99990eadfd4152a14.jpg", + "table_caption": [ + "Table 3. Ablation study of ExCEL on VOC val set." + ], + "table_footnote": [], + "table_body": "
ConditionsSVCTSELVCPrecisionRecallmIoU
Baseline (CLIP)18.821.312.1
w/ SVC✓81.286.272.5
w/o LVC✓✓80.789.874.7
w/o TSE✓✓83.786.375.1
ExCEL✓✓✓85.088.477.2
", + "bbox": [ + 91, + 354, + 485, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23376d17613377901da151701a378f90d8d2d38996a8919fc1e5565bcd17c967.jpg", + "table_caption": [ + "Table 4. Ablation study of attribute number $B$ on VOC val set." + ], + "table_footnote": [], + "table_body": "
Number of AttrNone3264112144196
mIoU75.175.876.277.277.076.5
", + "bbox": [ + 93, + 467, + 485, + 507 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8ec2218118d0fa0916c231c388efb0bf1dcdf9e721bae7ce6c9ad95a22d135e.jpg", + "table_caption": [ + "Table 5. Ablation study of VC module on VOC train set." + ], + "table_footnote": [], + "table_body": "
Conditionsq-kvI.C.M.C.LVCPrecisionRecallmIoU
Baseline (CLIP)✓18.021.811.2
MaskCLIP✓77.180.965.8
w/ I.C.✓79.184.769.7
SVC✓✓82.288.274.6
ExCEL✓✓✓86.687.978.0
", + "bbox": [ + 91, + 532, + 485, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 633, + 254, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficacy of Key Components. Quantitative ablative experiments of ExCEL are reported in Tab. 3. The baseline is the vanilla CLIP using our training settings, which only achieves $12.1\\%$ mIoU for segmentation. Our SVC module replaces the q-k attention with Intra-correlation from the intermediate layers. The performance increases to $72.5\\%$ . TSE module enriches the semantics of text representation for robust visual recognition. Introducing TSE brings a $3.6\\%$ recall increase compared to the original text templates. LVC module provides a dynamic shift to diversify the features. It further benefits SVC's segmentation performance to $75.1\\%$ mIoU. With all these enhancements, ExCEL generates $77.2\\%$ mIoU for segmentation.", + "bbox": [ + 88, + 657, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative ablation results are further illustrated in Fig. 4 (b-e) to evaluate the efficacy of our modules. In Fig. 4 (b), the CLIP baseline produces inaccurate CAMs with mis", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "localized activation. SVC corrects token relations and preserves fine-grained details, effectively suppressing false activations, as seen in Fig. 4 (c). TSE incorporates comprehensive textual attributes into the text representation, enhancing patch-text matching and producing more complete CAMs, shown in Fig. 4 (d). LVC dynamically optimizes attention maps, further improving CAM accuracy and completeness, as illustrated in Fig. 4 (e). Both quantitative and qualitative results confirm the effectiveness of our modules.", + "bbox": [ + 511, + 340, + 903, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Implicit Attributes. Tab. 4 analyzes the effect of varying the number of clustering attributes. 'None' means no clustering and we explicitly fuse the $n$ description embeddings for each class. It shows that the performance drops from $77.2\\%$ to $75.1\\%$ , which validates the efficacy of implicit attributes and its superiority over explicit descriptions. With this operation, we can expand the representation of 20 classes up to 196 attributes or more, greatly enhancing text semantics. It reports that ExCEL achieves more favorable performance when $B$ is set to 112.", + "bbox": [ + 511, + 478, + 905, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Visual Calibrations. Tab. 5 compares different strategies in VC module for CAM generation. I.C. refers to Intra-correlation in the last layer, and M.C. (Intermediate Calibration) applies I.C. across intermediate layers. Vanilla q-k attention in CLIP loses diversity and cannot generate reasonable CAMs. $v$ contains fine-grained knowledge and MaskCLIP improves CAMs to $65.8\\%$ by using $v$ from the last layer. In contrast, our I.C. and M.C. focus on mining diverse knowledge from intermediate layers and boost the performance to $69.7\\%$ and $74.6\\%$ . In addition, introducing LVC module raises the final performance to $78.0\\%$ . Results in Tab. 5 and corresponding visualizations in Fig. 4 clearly highlight the efficacy of our components.", + "bbox": [ + 511, + 628, + 906, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 833, + 681, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyper-parameter Analysis. Hyper-parameters, such as TOP-K, scaling factors $\\alpha$ and $\\beta$ , and the number of SVC layers $N$ , etc., are discussed in Supplementary Materials.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20229", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6f1b183998189eb28bdabdbe4728e64bce83267cd36a9814988a3261b38caec.jpg", + "table_caption": [ + "Table 6. Comparisons with the fully-supervised counterparts on VOC val set. $\\mathcal{F}$ :fully-supervised. ViT-B*: pretrained from CLIP." + ], + "table_footnote": [], + "table_body": "
MethodsTypeSup.Net.ValRatio
DeepLabV2 [4] TPAMI'2017-FRN10177.7-
DeepLabV2 [4] TPAMI'2017-FViT-B82.3-
WeCLIP-Full [43] CVPR'2024-FViT-B*81.6-
CLIMS [36] CVPR'2022MI + LRN10170.490.6%
CLIP-ES [20] CVPR'2023MI + LRN10172.292.9%
CPAL [31] CVPR'2024MI + LRN10174.595.9%
ToCo [29] CVPR'2024SIViT-B71.186.4%
DuPL [35] CVPR'2024SIViT-B73.389.1%
SeCo [43] CVPR'2024SIViT-B74.089.9%
DIAL [13] ECCV'2024SI + LViT-B74.590.5%
WeCLIP [43] CVPR'2024SI + LViT-B*76.493.6%
ExCEL (Ours)SI + LViT-B*78.496.1%
", + "bbox": [ + 91, + 116, + 486, + 308 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df209429a16cb59305b2099d05e746d69b0062b25292e73bee4e669831ae8efc.jpg", + "table_caption": [ + "Table 7. Training efficiency comparisons on VOC train (CAM) and val set (Seg). All experiments are conducted on RTX 3090." + ], + "table_footnote": [], + "table_body": "
MethodTypeTraining TimeGPUCAMSeg
CLIMS [36] CVPR'2022M1068 mins18.0 G56.670.4
CLIP-ES [20] CVPR'2023M420 mins12.0 G70.872.2
MCTformer+ [38] TPAMI'2024M1496 mins18.0 G68.874.0
ToCo [29] CVPR'2023S506 mins17.9 G71.671.1
DuPL [35] CVPR'2024S508 mins14.9 G75.073.3
SeCo [39] CVPR'2024S407 mins17.6 G74.874.0
WeCLIP [43] CVPR'2024S270 mins6.2 G75.476.4
ExCEL* (Training-free)S-2.9 G74.6-
ExCEL (Ours)S90 mins3.2 G78.078.4
", + "bbox": [ + 91, + 343, + 485, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fully-supervised Counterparts. Tab. 6 presents a fairness comparison between WSSS methods and their fully-supervised counterparts using the same segmentation backbone. With CLIP's visual encoder as the backbone, ExCEL achieves $78.4\\%$ mIoU, reaching $96.1\\%$ of the fully-supervised performance. It significantly outperforms CLIP-based WeCLIP by $2.5\\%$ and demonstrates ExCEL's advantage over other multi-stage CLIP-based methods as well.", + "bbox": [ + 88, + 507, + 482, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Efficiency Analysis. Our method only trains the adapter and decoder in a single-stage paradigm. Tab. 7 compares the training efficiency between ExCEL and recent methods. Without training, ExCEL requires just 2.9 GB of GPU memory and generates comparable CAMs to recent SOTAs. When training is included, the entire pipeline only takes 90 minutes and 3.2 GB of memory for SOTA performance. ExCEL just requires $6.0\\%$ training time of multi-stage MCTformer+ and $33.3\\%$ of single-stage WeCLIP, highlighting ExCEL's remarkable training efficiency.", + "bbox": [ + 88, + 628, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attribute Response Analysis. We treat text prompting as an implicit attribute-hunting process to comprehensively enrich text representation semantics. To evaluate if the clustered attributes capture distinct object characteristics, we visualize 5 implicit attributes based on similarity scores. As shown in Fig. 5, given instances of $\\{\\text{aeroplane}\\}$ and $\\{\\text{train}\\}$ , our attributes highlight different object parts, which clearly validates that our TSE module enhances inte", + "bbox": [ + 88, + 780, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/178f8d7f60bfa66fe2b0fd0d2b2000aaf26998844024d8e16377bf35b7bb58d7.jpg", + "image_caption": [ + "Figure 5. Implicit attribute responses. Based on the TOPK similarity scores, 5 attributes are sampled for visualizations." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 900, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d9e85d1175b51ab2d9ff7dd9fe40e8f79ab87d68d4b4f0c327585ba9676213e.jpg", + "image_caption": [ + "Figure 6. Comparisons of attention maps from the last visual encoder layer. (a) Attention features from the query patches (marked by red stars). (b) Token relations measured by cosine similarity." + ], + "image_footnote": [], + "bbox": [ + 516, + 205, + 908, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gral visual responses by gathering relevant semantics.", + "bbox": [ + 511, + 364, + 867, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Representation Analysis. CLIP lacks fine-grained details, leading to inaccurate patch-text responses. To explore further, we visualize the self-attention features in Fig. 6 (a). Given the query patch (red star), CLIP's q-k attention falls short in generating diverse features, supporting our claim in Sec. 3.3. MaskCLIP observes that $v$ keeps diversity and takes it from the last layer for visual response. We visualize it by calculating $v - v$ attention. Although effective, MaskCLIP still misses fine granularity. Instead, SVC calculates attention within each space and implements it from intermediate layers. LVC further diversifies the features with a dynamic adapter, both of which effectively generate features with clear boundaries and spatial details.", + "bbox": [ + 511, + 378, + 903, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we explore the pairwise token relations in Fig. 6 (b). Unlike the smoother attention maps of CLIP or MaskCLIP, our approach distinctly groups tokens with similar semantics, aligning pairwise similarities with corresponding semantics. This validates that ExCEL successfully enhances the frozen features of CLIP by calibrating it towards distributions with more diverse spatial information.", + "bbox": [ + 511, + 575, + 905, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 632, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose ExCEL, a novel patch-text alignment method to explore CLIP's dense knowledge for WSSS, which provides a different insight to generate better pseudo labels based on CLIP. To this end, Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules are designed to improve the dense alignment across text and vision modalities. In addition, ExCEL generates CAMs in both training-free and efficient training modes, calibrating CLIP without altering its pre-trained weights. It retains CLIP's transferability while significantly reducing training cost. We believe ExCEL can inspire more future research to unlock CLIP's dense capabilities in the WSSS field.", + "bbox": [ + 509, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20230", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, pages 4981-4990, 2018. 1", + "[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 549-565. Springer, 2016. 1", + "[3] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, pages 1108-1118, 2023. 6", + "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8", + "[5] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, pages 3135-3144, 2023. 1, 6", + "[6] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, pages 969–978, 2022. 6", + "[7] Zesen Cheng, Pengchong Qiao, Kehan Li, Siheng Li, Pengxu Wei, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Out-of-candidate rectification for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23673-23684, 2023. 1, 6", + "[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, pages 1635-1643, 2015. 1", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, pages 4320–4329, 2022. 5", + "[11] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The" + ], + "bbox": [ + 93, + 258, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the National Natural Science Foundation of China under Grant No.82372097, Shanghai Sailing Program under Grant 22YF1409300, International Science and Technology Cooperation Program under the 2023 Shanghai Action Plan for Science under Grant 23410710400, Taishan Scholars Program under Grant NO.tsqn202408245.", + "pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 5", + "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 2", + "[13] Soojin Jang, Jungmin Yun, Junehyoung Kwon, Eunju Lee, and Youngbin Kim. Dial: Dense image-text alignment for weakly supervised semantic segmentation. arXiv preprint arXiv:2409.15801, 2024. 6, 8", + "[14] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 6", + "[15] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 24, 2011. 6", + "[16] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1", + "[17] Xueyi Li, Tianfei Zhou, Jianwu Li, Yi Zhou, and Zhaoxiang Zhang. Group-wise semantic mining for weakly supervised semantic segmentation. In AAAI, pages 1984-1992, 2021. 5", + "[18] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In CVPR, pages 3159-3167, 2016. 1", + "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[20] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. arXiv preprint arXiv:2212.09506, 2022. 1, 3, 6, 7, 8", + "[21] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 4", + "[22] Balamurali Murugesan, Rukhshanda Hussain, Rajarshi Bhattacharya, Ismail Ben Ayed, and Jose Dolz. Prompting classes: exploring the power of prompt class learning in weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 291-302, 2024. 6", + "[23] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In CVPR, pages 1713-1721, 2015. 1", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20231", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 2", + "[26] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 6", + "[27] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, pages 446-463. Springer, 2022. 6", + "[28] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: end-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1, 6", + "[29] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. arXiv preprint arXiv:2303.01267, 2023. 2, 5, 6, 8", + "[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2", + "[31] Feilong Tang, Zhongxing Xu, Zhaojun Qu, Wei Feng, Xingjian Jiang, and Zongyuan Ge. Hunting attributes: Context prototype-aware learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3324-3334, 2024. 6, 8", + "[32] Paul Vernaza and Manmohan Chandraker. Learning random-walk label propagation for weakly-supervised semantic segmentation. In CVPR, pages 7158-7166, 2017. 1", + "[33] Xiang Wang, Shaodi You, Xi Li, and Huimin Ma. Weakly-supervised semantic segmentation by iteratively mining common object features. In CVPR, pages 1354-1362, 2018. 1", + "[34] Yuanchen Wu, Xiaogiang Li, Jide Li, Pinpin Zhu, Shaohua Zhang, et al. Dino is also a semantic guider: Exploiting class-aware affinity for weakly supervised semantic segmentation. In ACM Multimedia 2024. 2", + "[35] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, pages 3534-3543, 2024. 5, 6, 8", + "[36] Jinheng Xie, Xianxu Hou, Kai Ye, and Linlin Shen. Clims: cross language image matching for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4483-4492, 2022. 1, 3, 6, 8" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2", + "[38] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 2024. 6, 8", + "[39] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, pages 3606–3615, 2024. 2, 5, 6, 8", + "[40] Zhiwei Yang, Yuong Meng, Kexue Fu, Shuo Wang, and Zhijian Song. Tackling ambiguity from perspective of uncertainty inference and affinity diversification for weakly supervised semantic segmentation. arXiv preprint arXiv:2404.08195, 2024. 1", + "[41] Sung-Hoon Yoon, Hoyong Kwon, Hyeonseong Kim, and Kuk-Jin Yoon. Class tokens infusion for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3595-3605, 2024. 6", + "[42] Bingfeng Zhang, Xuru Gao, Siyue Yu, and Weifeng Liu. Enhanced online cam: Single-stage weakly supervised semantic segmentation via collaborative guidance. Pattern Recognition, 156:110787, 2024. 2", + "[43] Bingfeng Zhang, Siyue Yu, Yunchao Wei, Yao Zhao, and Jimin Xiao. Frozen clip: A strong backbone for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3796-3806, 2024. 2, 3, 5, 6, 7, 8", + "[44] Xinqiao Zhao, Feilong Tang, Xiaoyang Wang, and Jimin Xiao. Sfc: Shared feature calibration in weakly supervised semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7525-7533, 2024. 2, 6", + "[45] Xinqiao Zhao, Ziqian Yang, Tianhong Dai, Bingfeng Zhang, and Jimin Xiao. Psdpm: Prototype-based secondary discriminative pixels mining for weakly supervised semantic segmentation. In CVPR, pages 3437-3446, 2024. 6", + "[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, 2016. 1", + "[47] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 4, 7", + "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2", + "[49] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In CVPR, pages 4299-4309, 2022. 6" + ], + "bbox": [ + 516, + 92, + 903, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20232", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_model.json b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6280951456e7303778736dd41587701bb96edaf6 --- /dev/null +++ b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_model.json @@ -0,0 +1,2585 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.13, + 0.85, + 0.178 + ], + "angle": 0, + "content": "Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.202, + 0.783, + 0.242 + ], + "angle": 0, + "content": "Zhiwei Yang\\(^{1,2}\\) Yucong Meng\\(^{2,3}\\) \nKexue Fu\\(^{4}\\) Feilong Tang\\(^{1}\\) Shuo Wang\\(^{2,3^{*}}\\) Zhijian Song\\(^{1,2,3^{*}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.242, + 0.872, + 0.314 + ], + "angle": 0, + "content": "\\(^{1}\\)Academy for Engineering and Technology, Fudan University, Shanghai 200433, China \n\\(^{2}\\)Shanghai Key Laboratory of Medical Image Computing and Computer Assisted Intervention \n\\(^{3}\\)Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, China \n\\(^{4}\\)Shandong Computer Science Center (National Supercomputer Center in Jinan)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.348, + 0.327, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.485, + 0.789 + ], + "angle": 0, + "content": "Weakly Supervised Semantic Segmentation (WSSS) with image-level labels aims to achieve pixel-level predictions using Class Activation Maps (CAMs). Recently, Contrastive Language-Image Pre-training (CLIP) has been introduced in WSSS. However, recent methods primarily focus on image-text alignment for CAM generation, while CLIP's potential in patch-text alignment remains unexplored. In this work, we propose ExCEL to explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. Specifically, we propose Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules to improve the dense alignment across both text and vision modalities. To make text embeddings semantically informative, our TSE module applies Large Language Models (LLMs) to build a dataset-wide knowledge base and enriches the text representations with an implicit attribute-hunting process. To mine fine-grained knowledge from visual features, our VC module first proposes Static Visual Calibration (SVC) to propagate fine-grained knowledge in a non-parametric manner. Then Learnable Visual Calibration (LVC) is further proposed to dynamically shift the frozen features towards distributions with diverse semantics. With these enhancements, ExCEL not only retains CLIP's training-free advantages but also significantly outperforms other state-of-the-art methods with much less training cost on PASCAL VOC and MS COCO. Code is available at https://github.com/zwyang6/ExCEL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.805, + 0.222, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.831, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Weakly Supervised Semantic Segmentation (WSSS) intends to generate pixel-level predictions using weak annotations like points [2], scribbles [18, 32], bounding boxes [8," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.346, + 0.907, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.437, + 0.907, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Our motivation. (a) Previous methods leverage CLIP to generate CAMs with global image-text alignment, leaving CLIP's dense knowledge unexplored. (b) The proposed ExCEL explores CLIP's dense knowledge via a novel patch-text alignment paradigm, which generates better CAMs with less training cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.735 + ], + "angle": 0, + "content": "16], or image-level labels [1, 23, 40]. It significantly reduces the annotating cost of fully supervised methods and has attracted increasing attention in the community. Among all cheap annotation types, most WSSS approaches leverage image-level labels to provide dense localization cues, linking visual concepts to specific pixel regions [5, 33]. In this work, we focus on WSSS with image-level labels as well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Commonly, the WSSS pipeline involves three stages: generating Class Activation Maps (CAMs) [46] by training a classification network, refining CAMs into pseudo labels [28], and using these labels to train a segmentation model [7]. However, due to the minimal semantic information from image-level labels, CAMs intend to highlight the most distinctive object parts, significantly limiting WSSS performance. Recently, Contrastive Language-Image Pre-training (CLIP) [24] has been introduced in WSSS. CLIMs [36] applies image-text pairs to regularize visual relations among different semantics. CLIP-ES [20]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.471, + 0.902 + ], + "angle": 0, + "content": "*Corresponding authors. Email:{shuowang, zjsong}@fudan.edu.cn." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20223" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "leverages image-text alignment for gradient and produces high-quality GradCAM [30]. WeCLIP [43] further streamlines this process by using CLIP's visual encoder for segmentation. Despite these advancements, current methods primarily focus on CLIP's global image-text alignment, as shown in Fig. 1 (a). CLIP's dense knowledge with patch-text alignment still remains under-explored in WSSS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.482, + 0.397 + ], + "angle": 0, + "content": "In this work, we propose ExCEL to explore CLIP's dense knowledge via a patch-text alignment paradigm for WSSS, i.e., generating CAMs by calculating patch-wise similarity between text and individual patch tokens, as shown in Fig. 1 (b). We identify two key challenges: (1) Semantic sparsity in textual prompts, where the template 'a photo of [CLASS]' only indicates object presence but lacks knowledge for localization, and (2) Fine-grained insufficiency in visual features, as CLIP prioritizes global representation due to its image-text pairing nature. To address these, ExCEL enhances CLIP's dense alignment with Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules, unlocking its potential across text and vision modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.672 + ], + "angle": 0, + "content": "To generate semantically rich text representations, we propose TSE through an implicit attribute feature space. Instead of relying on explicit text templates, TSE module implicitly constructs text embeddings using universal attributes across the dataset. We first employ Large Language Models (LLMs) to generate detailed descriptions for each class, which are then processed by CLIP's text encoder to build a dataset-wide knowledge base. Rather than directly fusing these class-specific descriptions for text prompting, we focus on clustering the descriptive embeddings into generalized attributes, which effectively capture complementary knowledge from other classes and supplement missing information for the target class. With this implicit feature space, we enhance the text embeddings by hunting for its most relevant attribute features and aggregate them into the final class-specific text representation. This approach enables TSE to generate more informative text embeddings, providing a strong foundation for visual recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To mine fine-grained knowledge from visual features, we propose VC to calibrate CLIP in both non-trainable and efficient-learnable ways. Our findings suggest that CLIP's q-k attention loses fine-grained details. Therefore, we first propose a Static Visual Calibration (SVC) module to replace the suboptimal q-k attention with a straightforward Intra-correlation operation. It focuses on extracting fine-grained details from intermediate layers, which progressively propagates fine-grained visual knowledge. Without any retraining, SVC generates CAMs comparable to training-required WSSS methods. Building on this, we further propose a Learnable Visual Calibration (LVC) module to dynamically calibrate CLIP's frozen features. LVC extracts spatial correlations from SVC's static CAMs. These correlations further supervise a lightweight adapter to learn" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "the dynamic shift, pushing frozen features towards spatial-aware distributions. LVC and SVC complement each other, enabling precise patch-text alignment for CAM generation." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.138, + 0.905, + 0.152 + ], + "angle": 0, + "content": "The main contributions of our work are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.229 + ], + "angle": 0, + "content": "- We explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. The proposed ExCEL generates better pseudo labels in both training-free and efficient learning manners, revealing the dense capabilities of CLIP for efficient CAM generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- To enhance patch-text alignment, we propose the Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules. TSE applies LLMs to build a dataset-wide knowledge base and treats text prompting as an implicit attribute-hunting process, making text embeddings more informative. VC propagates the fine-grained visual knowledge in a non-parametric manner and further dynamically calibrates the frozen features with a lightweight adapter. TSE and VC work across two modalities, generating better dense alignment and pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.38, + 0.905, + 0.456 + ], + "angle": 0, + "content": "- Extensive experiments on PASCAL VOC and MS COCO demonstrate that ExCEL significantly outperforms recent state-of-the-art methods, while reducing training cost with only 3.2 GB of GPU memory and \\(6\\%\\) of the training time required by recent methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.471, + 0.664, + 0.487 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.881, + 0.513 + ], + "angle": 0, + "content": "2.1. Weakly Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Weakly supervised semantic segmentation with image-level labels typically relies on CAMs to provide dense supervision for segmentation [42, 44]. However, CAMs usually highlight the most discriminative parts of objects [34]. To address this issue, considerable efforts have been made from many intriguing insights. MCTformer [37] incorporates multiple class tokens in Vision Transformer and proposes generating CAMs from class-patch attention. ToCo [29] proposes token contrast learning and generates more precise CAMs. SeCo [39] designs a separate and conquer scheme and succeeds in tackling co-occurrence. Despite these advancements, prior methods commonly require retraining the entire classification network for CAM generation. In this work, our ExCEL directly generates CAMs from frozen CLIP, and further boosts its quality via a lightweight adapter, significantly reducing the training cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.78, + 0.804 + ], + "angle": 0, + "content": "2.2. Vision-Language Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Contrastive Language Image Pre-training (CLIP) [24], known for pretraining on billion-scale image-text pairs, has demonstrated remarkable transferability in many downstream tasks. CoOP [48] and CLIP-Adapter [12] incorporate lightweight trainable parameters into CLIP and succeed in few-shot classification. DenseCLIP [25] and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20224" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.898, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.348, + 0.908, + 0.432 + ], + "angle": 0, + "content": "Figure 2. ExCEL Architecture. We explore CLIP's dense knowledge with Text Semantic Enrichment (TSE) and Visual Calibration (VC). (a) TSE uses LLMs to build a knowledge base and clusters it into an implicit attribute space. The final text representation \\( T_{c} \\) is enhanced by hunting for relevant attributes. For vision modality, (b) we introduce Static Visual Calibration (SVC) to calibrate visual features using the Inter-correlation operation across \\( N \\) intermediate layers. It generates static CAMs with \\( T_{c} \\) and calibrated features \\( P_{s} \\). (c) Learnable Visual Calibration (LVC) designs a learnable adapter to add a dynamic shift \\( R \\) to SVC. It generates optimized features \\( P_{d} \\) based on static CAMs guidance, creating dynamic CAMs from \\( P_{d} \\) and \\( T_{c} \\). Dynamic CAMs are refined for segmentation supervision. Details are in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.486, + 0.643 + ], + "angle": 0, + "content": "MaskCLIP [47] leverage the alignment between text and vision modalities for the dense segmentation task. Recently, some studies have introduced CLIP into WSSS. CLIMS [36] treats image-text pairing of CLIP as regularization and leverages it to regularize the visual concepts. CLIP-ES [20] finds that the image-text alignment of CLIP generates class gradient and leverages it for GradCAM generation. WeCLIP [43] further streamlines this process and directly leverages CLIP's visual encoder for segmentation. However, these methods mainly focus on a global image-text alignment while ignoring the dense capabilities of CLIP. In contrast, our ExCEL explores CLIP's dense knowledge via a patch-text alignment paradigm for WSSS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.654, + 0.226, + 0.672 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.231, + 0.694 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Patch-text CAM Generation. CLIP uses image and text encoders to project image and text into the same feature space, enabling robust vision-language alignment. In this work, we utilize this property to generate CAM in a patch-text alignment paradigm. Given the encoded text embeddings \\( T \\in \\mathbb{R}^{D \\times C} \\) and visual features \\( P \\in \\mathbb{R}^{h \\times w \\times D} \\), where \\( D \\) is the feature channel, \\( h \\) and \\( w \\) are the spatial sizes, \\( C \\) is the number of classes, we generate CAM by calculating the patch-wise similarities between text and visual features:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.846, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C A M} = \\operatorname {N o r m} (\\cos (P, T), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathrm{Norm}(\\cdot)\\) is the min-max normalization and \\(\\cos (\\cdot)\\) is the cosine similarity calculation. However, due to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.506 + ], + "angle": 0, + "content": "the image-level pairing nature, CLIP suffers from textual semantic sparsity and visual fine-grained insufficiency. Therefore, our ExCEL incorporates TSE and VC (SVC and LVC) into CLIP to further explore its dense potential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Framework Overview. Our ExCEL generates CAMs in both training-free and efficient learning manners, as shown in Fig. 2. Its training pipeline is generalized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.909, + 0.825 + ], + "angle": 0, + "content": "(1) Enriching textual semantics via TSE. We first use GPT-4 to generate descriptions for each class, which are encoded into a dataset-wide knowledge base with CLIP's text encoder. We cluster this knowledge into class-agnostic attributes and use the global text prompt to hunt for its most relevant ones. They are then aggregated into the final text representation. (2) Static CAM generation via SVC. We replace CLIP's q-k self-attention with our Intra-correlation operation from intermediate layers. Then the calibrated visual features and enhanced text embeddings are used for static CAMs via Eq. (1). (3) Dynamic CAM generation via LVC. A lightweight adapter is designed to learn dynamic token relations from static CAMs. The relations are added to SVC and serve as a distribution shift to make the visual features more diverse. The dynamic CAMs are generated with the enhanced text embeddings and LVC features via Eq. (1). (4) Segmentation training. Dynamic CAMs are refined to pseudo labels for segmentation supervision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.757, + 0.849 + ], + "angle": 0, + "content": "3.2. Text Semantic Enrichment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Knowledge Base Construction. The global text template 'a clean origami of [CLASS]' only indicates the presence of objects while limited providing dense knowledge for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20225" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.394 + ], + "angle": 0, + "content": "patch-wise visual recognition. To enrich the text representation, we first adapt LLMs, such as GPT-4 to generate detailed class descriptions, as shown in Fig. 2 (a). Specifically, given the global template \\( E_{c} \\) from class label space \\( Y \\in \\{1,2,\\dots,C\\} \\), where \\( c \\) is the class index and \\( C \\) is the number of categories, we carefully construct instructions for GPT: \"List \\( n \\) descriptions with key properties to describe the [CLASS] in terms of appearance, color, shape, size, or material, etc. These descriptions will help visually distinguish the [CLASS] from other classes in the dataset. Each description should follow the format: 'a clean origami [CLASS]. it + descriptive contexts.' With this instruction, we ask GPT to generate \\( n \\) detailed descriptions for each class, which are subsequently encoded into a dataset-wide knowledge base with CLIP's text encoder. The knowledge base is denoted as \\( \\mathcal{T} = \\{\\Phi(e_i)\\}_{i=1}^{n \\times C} \\). \\( e_i \\) is the description from GPT, \\( \\Phi(\\cdot) \\) is CLIP's text encoder and \\( n \\times C \\) is the number of all descriptions. This knowledge base gathers descriptive properties for the whole dataset, building a strong foundation for the textual category representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Implicit Attribute Hunting. Instead of explicitly merging class-related knowledge into a single text embedding, we cluster this knowledge into generalized attributes and treat text prompting as an implicit attribute-hunting process. This has two main benefits: (1) \\( n \\) explicit descriptions may still be limited in covering all characteristics of the class. The clustered attributes efficiently capture shared contextual knowledge from other categories, supplementing missing information for target class recognition. (2) The generated descriptions inevitably contain redundant or noisy content. The use of attributes makes the knowledge more compact and representative, leading to precise text prompting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.637 + ], + "angle": 0, + "content": "To this end, we leverage a clustering algorithm to generate multiple centroids based on the knowledge base. Each cluster centroid is viewed as the implicit attribute that represents a group of descriptions sharing similar properties:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.645, + 0.482, + 0.665 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {K m e a n s} (\\mathcal {T}, B) = \\left\\{a _ {i} \\right\\} _ {i = 1} ^ {B}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.72 + ], + "angle": 0, + "content": "where \\(B\\) is the number of centroids and kmeans algorithm [21] is used for clustering for simplicity. \\(a_{i}\\in \\mathbb{R}^{D\\times 1}\\) represents the cluster centroid, i.e., the implicit attribute." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.827 + ], + "angle": 0, + "content": "With the attribute feature space \\(A\\), we first send the global template \\(E_{c}\\) into the text encoder of CLIP to generate global text embedding \\(t_c \\in \\mathbb{R}^{D \\times 1}\\). \\(D\\) is the channel dimension. Then \\(t_c\\) is leveraged to search for its most relevant attributes in the implicit attribute space. To exclude irrelevant attributes, we further propose to select TOPK attribute neighbors based on the similarity scores with \\(t_c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.837, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nA _ {c} = \\left\\{a _ {j}: j \\in \\operatorname {a r g m a x} _ {\\mathrm {T O P K}} \\left\\{t _ {c} ^ {T} a _ {j} \\right\\} _ {j = 1} ^ {B} \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Finally, we gently aggregate the implicit attribute neighbors according to the corresponding similarity weights to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "\\(t_c\\) and take the aggregated features as the complementary knowledge for textual semantic enrichment. The final text representation \\(T_{c}\\in \\mathbb{R}^{D\\times 1}\\) is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.144, + 0.907, + 0.186 + ], + "angle": 0, + "content": "\\[\nT _ {c} = t _ {c} + \\lambda \\sum_ {j = 1} ^ {K} \\operatorname {s o f t m a x} \\left(t _ {c} ^ {T} A _ {c}\\right) a _ {j}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.193, + 0.892, + 0.208 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the factor to balance the attribute information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.215, + 0.7, + 0.23 + ], + "angle": 0, + "content": "3.3. Visual Calibrations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.45 + ], + "angle": 0, + "content": "Static Visual Calibration. Due to the image-text pairing nature of CLIP, the visual features lack fine-grained information, leading to unreasonable localization maps via patch-text alignment. To delve into this, let us review the self-attention mechanism of the original CLIP first. As shown in Fig. 2 (b), given the input image \\( X \\in \\mathbb{R}^{3 \\times \\mathcal{H} \\times \\mathcal{W}} \\) we send it to the image encoder of CLIP, which includes 12 Attention layers in this work. \\( \\mathcal{H} \\times \\mathcal{W} \\) is the image size. For the feature \\( F_{l} \\in \\mathbb{R}^{D_{s} \\times hw} \\) from \\( l \\)-th layer of CLIP, it is first projected into three different spaces \\( \\{q, k, v\\} \\), named query, key, and value, respectively. \\( q, k \\) and \\( v \\) have the same shape of \\( D_{s} \\times hw \\), where \\( D_{s} \\) and \\( hw \\) represents the channel dimension and sequence length. Then the attention map between \\( q \\) and \\( k \\) is calculated by measuring the similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.456, + 0.907, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} (q, k) = \\operatorname {s o f m a x} \\left(q ^ {T} k / \\sqrt {D _ {s}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.533 + ], + "angle": 0, + "content": "where \\(\\mathrm{SA}(\\cdot)\\) is the calculation of attention. Then the output features \\(F_{l + 1}\\) are generated by aggregating the tokens of \\(v\\) according to similarity weights in the attention map." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.906, + 0.759 + ], + "angle": 0, + "content": "However, due to the inherent image-text alignment of CLIP, the original q-k attention produces overly uniform attention maps, homogenizing diverse tokens from \\( v \\) to capture broad semantics for global image representation (see discussions in Sec. 4.4). It leads to inaccurate object recognition. MaskCLIP [47] holds a similar observation, supporting this claim by removing the final q-k attention layer and using \\( v \\) from the last layer as the visual output to preserve diversity. In our work, we choose to replace the suboptimal q-k attention with a straightforward Intra-correlation operation and focus on extracting fine-grained details from intermediate layers. This non-parametric approach effectively mines spatial semantics in intermediate \\( \\{q, k, v\\} \\) and avoids the smoothing effect of q-k attention, resulting in more consistent attention maps and improved object localization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.76, + 0.907, + 0.822 + ], + "angle": 0, + "content": "Specifically, instead of generating q-k correlation, Intra-correlation calculates the attention within each space of \\(\\{q,k,v\\}\\) across intermediate layers. The attention map \\(S_{attn}^{l}\\in \\mathbb{R}^{hw\\times hw}\\) from \\(l\\) -th SVC layer is generated by:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.828, + 0.907, + 0.851 + ], + "angle": 0, + "content": "\\[\nS _ {a t t n} ^ {l} = \\sum w _ {i} \\operatorname {S A} \\left(O _ {i} ^ {l}, O _ {i} ^ {l}\\right), O _ {i} ^ {l} \\in \\left\\{q ^ {l}, k ^ {l}, v ^ {l} \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(w_{i}\\) is the contribution weight for different correlation maps. \\(l\\in \\{12 - N,\\dots,12\\}\\) and \\(N\\) is the number of intermediate layers for this operation. Then \\(S_{attn}^{l}\\) and \\(v^{l}\\) are used to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20226" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "generate the output features. Finally, the calibrated features \\( P_{s} \\in \\mathbb{R}^{D \\times h \\times w} \\) from the last layer is used to generate static CAM \\( CAM_{s} \\) with text embedding \\( T_{c} \\) via Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.137, + 0.483, + 0.272 + ], + "angle": 0, + "content": "Learnable Visual Calibration. Although ExCEL generates comparable CAMs without training, its performance is still limited by the fixed features in CLIP. To further unleash the dense potential of CLIP, we design a lightweight adapter to dynamically calibrate the visual features with diverse details. This adapter only incorporates a distribution shift to calibrate the fixed features without changing CLIP's pre-trained weights, thereby retaining CLIP's transferability and enhancing its dense performance for WSSS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 2 (c), frozen features \\( F_{l} \\) from 1-12th layer of CLIP are extracted to learn a dynamic feature \\( F_{d} \\) via the adapter. The process is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.327, + 0.483, + 0.348 + ], + "angle": 0, + "content": "\\[\nF _ {d} = \\operatorname {C o n v} (\\operatorname {C o n c a t e} [ \\delta_ {l} (F _ {l}) ] _ {l = 1} ^ {1 2}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.432 + ], + "angle": 0, + "content": "where \\(F_{d}\\in \\mathbb{R}^{D_{d}\\times hw}\\) \\(D_{d}\\) is the channel dimension. \\(\\mathrm{Conv}(\\cdot)\\) is the convolution layer, \\(\\mathrm{Concat}[\\cdot ]\\) is the concatenate operation that connects all features along channel dimension, and \\(\\delta_l(\\cdot)\\) is the individual MLP layer for each \\(F_{l}\\) Then \\(F_{d}\\) is used to generate dynamic token relations by:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.441, + 0.483, + 0.46 + ], + "angle": 0, + "content": "\\[\nr = \\alpha (\\cos (F _ {d}, F _ {d}) - \\beta \\overline {{\\cos (F _ {d} , F _ {d})}}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.53 + ], + "angle": 0, + "content": "where \\(r \\in \\mathbb{R}^{hw \\times hw}\\), \\(\\alpha\\) and \\(\\beta\\) are the scaling and shifting factors to adjust the relations, respectively. \\(\\overline{\\cos(F_d, F_d)}\\) means the mean value of similarity scores of \\(F_d\\). It is designed to remove the irrelevant relations in low values by:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.54, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\nR _ {i j} = \\left\\{ \\begin{array}{l l} r _ {i j}, & \\text {i f} r _ {i j} \\geq 0 \\\\ - i n f, & \\text {e l s e} \\end{array} . \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.667 + ], + "angle": 0, + "content": "With the dynamic relation \\(R \\in \\mathbb{R}^{hw \\times hw}\\), we add it as a distribution bias to the static attention map \\(S_{attn}^{l}\\), dynamically grouping the frozen tokens within related semantics and shifting the features towards denser distribution. The optimized attention map \\(L_{attn}^{l} \\in \\mathbb{R}^{hw \\times hw}\\) is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.675, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\nL _ {a t t n} ^ {l} = S _ {a t t n} ^ {l} + \\operatorname {s o f t m a x} (R). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.765 + ], + "angle": 0, + "content": "Subsequently, we extract the dynamically calibrated features \\( P_{d} \\in \\mathbb{R}^{D \\times h \\times w} \\) from the last layer of LVC and generate dynamic CAMs with Eq. (1), which are then refined to final pseudo labels \\( M_{d} \\) for the segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.281, + 0.79 + ], + "angle": 0, + "content": "3.4. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We formulate a diversity loss to supervise the learning of \\( F_{d} \\) in our LVC module. We first measure token correlations of \\( F_{d} \\) by calculating the self similarity: \\( \\hat{\\mathcal{R}} = \\mathrm{sigmoid}(\\cos (F_d,F_d)) \\), where sigmoid(·) is the activation function and \\( \\hat{\\mathcal{R}}\\in \\mathbb{R}^{hw\\times hw} \\). Then we refine \\( CAM_{s} \\) from SVC into static pseudo labels \\( M_{s} \\) and leverage its pixelwise affinity to guide the diversifying of \\( \\hat{\\mathcal{R}} \\). Specifically, if" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "the pixel with coordinate \\((i,j)\\) shows the same pixel value as the pixel in \\((\\varepsilon,\\eta)\\), the token pair with the same coordinates on \\(F_{d}\\) is semantically related and its corresponding correlation logit on \\(\\hat{\\mathcal{R}}\\) should be maximized, and vice versa. The diversity loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.177, + 0.905, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i v}} = \\frac {1}{N ^ {+}} \\sum_ {u ^ {+} \\in \\hat {\\mathcal {R}} ^ {+}} (1 - u ^ {+}) + \\frac {1}{N ^ {-}} \\sum_ {u ^ {-} \\in \\hat {\\mathcal {R}} ^ {-}} u ^ {-}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.905, + 0.318 + ], + "angle": 0, + "content": "where \\( N^{+} / N^{-} \\) are the number of positive and negative pairs, \\( u^{+} / u^{-} \\) are the positive and negative relation logits, and \\( \\hat{\\mathcal{R}}^{+} / \\hat{\\mathcal{R}}^{-} \\) are the positive and negative sets of logit on \\( \\hat{\\mathcal{R}} \\), respectively. The diversity loss groups tokens with similar semantics and suppresses irrelevant ones, enhancing fine-grained details of visual features for precise text response." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.408 + ], + "angle": 0, + "content": "In addition, ExCEL is streamlined as a single-stage method. We adopt a lightweight Transformer-based segmentation head [43] and directly take the frozen visual encoder of CLIP for segmentation. The dynamic pseudo labels are used as supervision with a cross-entropy loss \\(\\mathcal{L}_{seg}\\). The loss objectives of our ExCEL are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.421, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {E X C E L}} = \\mathcal {L} _ {s e g} + \\gamma \\mathcal {L} _ {\\text {d i v}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.905, + 0.494 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the weight factor. By efficiently training the adapter and a segmentation head, ExCEL achieves strong WSSS performance and significantly reduces training cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.749, + 0.524 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.722, + 0.548 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Datasets and Metrics. The proposed ExCEL is evaluated on PASCAL VOC 2012 [11] and MS COCO 2014 [19]. VOC contains 21 categories (1 for background). Following prior methods [10, 17, 39], the augmented dataset with 10,582, 1,449, and 1,456 images are used for training, validating, and testing, respectively. COCO includes 81 classes, in which 82,081 images are used for training and 40,137 images are for validating. Mean Intersection-Over-Union (mIoU) is used as the main evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. CLIP model with ViT-B [9] is used as ExCEL's encoder, which is frozen during the training. For TSE module, we generate \\( n = 20 \\) descriptions from GPT-4 for each category. The number of attribute embeddings \\( B \\) is set to 112 and 224 for PASCAL VOC and MS COCO, respectively. The SVC module is conducted in the last \\( N = 5 \\) Transformer layers. Our decoder adopts a simple Transformer-based head [43]. Features \\( F_{l} \\) from each layer of CLIP are sent to it for the segmentation predictions. The scaling and shifting factors in Eq. (8) are set as 3.0 and 1.0, respectively. The loss weight \\( \\gamma \\) is set as 0.1. Following [29, 35, 39], the AdamW optimizer is used for training the adapter and decoder. The learning rate is 1e-4 with a weight decay of 1e-2. The training iteration is set" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20227" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.09, + 0.354, + 0.1 + ], + "angle": 0, + "content": "PASCAL VOC" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.102, + 0.182, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.102, + 0.268, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.102, + 0.357, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.102, + 0.446, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.102, + 0.535, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.147, + 0.181, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.147, + 0.268, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.147, + 0.357, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.147, + 0.446, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.147, + 0.535, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.192, + 0.181, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.192, + 0.268, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.192, + 0.357, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.192, + 0.446, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.192, + 0.535, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.236, + 0.181, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.236, + 0.268, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.236, + 0.357, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.236, + 0.446, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.236, + 0.535, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.09, + 0.635, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.09, + 0.725, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.102, + 0.814, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.102, + 0.901, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.147, + 0.636, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.147, + 0.725, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.147, + 0.814, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.147, + 0.901, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.192, + 0.636, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.192, + 0.725, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.192, + 0.814, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.192, + 0.901, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.236, + 0.636, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.236, + 0.725, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.236, + 0.814, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.236, + 0.901, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.288, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Segmentation visualizations of SeCo [39], WeCLIP [43] and ours on VOC and COCO. ExCEL segments objects more precisely." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.315, + 0.484, + 0.357 + ], + "angle": 0, + "content": "Table 1. Segmentation comparisons on VOC and COCO. Net. is the backbone for segmentation. Sup. is the supervision type. \\(\\mathcal{I}\\) image-level labels. SA: saliency maps. \\(\\mathcal{L}\\) : language." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.358, + 0.487, + 0.655 + ], + "angle": 0, + "content": "
MethodSup.Net.VOCCOCO
ValTestVal
Multi-stage WSSS methods.
L2G [14] CVPR'2022I + SARN10172.171.744.2
RCA [49] CVPR'2023I + SARN3872.272.836.8
OCR [7] CVPR'2023IRN3872.772.042.5
BECO [26] CVPR'2023IRN10173.773.545.1
MCTformer+ [38] TPAMI'2024IRN3874.073.645.2
CTI [41] CVPR'2024IRN10174.173.245.4
CLIMS [36] CVPR'2022I + LRN10170.470.0-
CLIP-ES [20] CVPR'2023I + LRN10172.272.845.4
PSDPM [45] CVPR'2024I + LRN10174.174.947.2
CPAL [31] CVPR'2024I + LRN10174.574.746.8
Single-stage WSSS methods.
AFA [28] CVPR'2022IMiT-B166.066.338.9
ViT-PCM [27] ECCV'2022IViT-B70.370.9-
ToCo [29] CVPR'2023IViT-B71.172.242.3
DuPL [35] CVPR'2024IViT-B73.372.844.6
SeCo [39] CVPR'2024IViT-B74.073.846.7
DIAL [13] ECCV'2024I + LViT-B74.574.944.4
WeCLIP [43] CVPR'2024I + LViT-B76.477.247.1
ExCEL(w/o CRF)I + LViT-B77.277.349.3
ExCEL (Ours)I + LViT-B78.478.550.3
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.702 + ], + "angle": 0, + "content": "as 30,000 for VOC and 100,000 for COCO. Please refer to Supplementary Materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.713, + 0.463, + 0.729 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Performance of Semantic Segmentation. Tab. 1 shows segmentation comparisons between our ExCEL and recent methods on VOC and COCO. The single-stage ExCEL achieves \\(78.4\\%\\) and \\(78.5\\%\\) mIoU on VOC val set and test set, which even significantly outperforms the sophisticated multi-stage methods by at least \\(3.9\\%\\) and \\(3.6\\%\\) mIoU, respectively. For more complicated benchmark COCO, ExCEL achieves \\(50.3\\%\\) mIoU on val set, which brings a noticeable \\(3.2\\%\\) increase over the CLIP-based state-of-the-art (SOTA) WeCLIP. In addition, without time-consuming post-processing techniques, such as CRF [15], ExCEL still" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.315, + 0.905, + 0.357 + ], + "angle": 0, + "content": "Table 2. CAM seed comparisons on VOC train set. \\(\\mathcal{M}\\): multistage methods. \\(S\\): single-stage methods. †: our reproduction following official codes. ExCEL*: ExCEL in a training-free manner." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.359, + 0.905, + 0.653 + ], + "angle": 0, + "content": "
MethodTypeSup.Net.VOC
Train
Training-free WSSS methods.
CLIP-ES [20] CVPR'2023MI + LViT-B70.8
ExCEL* (Ours)SI + LViT-B74.6
Training-required WSSS methods.
ReCAM [6] CVPR'2022MIRN10154.8
FPR [3] CVPR'2023MIRN10163.8
LPCAM [5] CVPR'2023MIRN5065.3
MCTformer+ [38] TPAMI'2024MIRN3868.8
SFC [44] AAAI'2024MIRN10164.7
CTI [41] CVPR'2024MIRN10169.5
AFA [28] CVPR'2022SIMiT-B165.0
ViT-PCM [27] ECCV'2022SIViT-B67.7
†ToCo [29] CVPR'2023SIViT-B71.6
†DuPL [35] CVPR'2024SIViT-B75.0
SeCo [39] CVPR'2024SIViT-B74.8
CLIMS [36] CVPR'2022MI + LRN10156.6
POLE [22] WACV'2023MI + LRN5059.0
CPAL [31] CVPR'2024MI + LRN10171.9
DIAL [13] ECCV'2024SI + LViT-B75.2
†WeCLIP [43] CVPR'2024SI + LViT-B75.4
ExCEL (Ours)SI + LViT-B78.0
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.673, + 0.88, + 0.687 + ], + "angle": 0, + "content": "maintains consistent superiority over SOTAs with CRF." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.749 + ], + "angle": 0, + "content": "The qualitative comparisons on VOC and COCO are shown in Fig. 3. By densely matching the patches and texts, ExCEL consistently demonstrates more precise object segmentation than recent methods in an image-text paradigm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Evaluation of CAM Seeds. Tab. 2 reports the quality of raw CAM seeds on VOC train set. Compared with recent methods, ExCEL achieves \\(74.6\\%\\) mIoU in a training-free setup, outperforming CLIP-ES by \\(3.8\\%\\) and performing comparably to most training-required methods. With the optimized LVC module, ExCEL further boosts CAM quality to \\(78.0\\%\\), surpassing SOTAs in the image-text paradigm by at least \\(2.6\\%\\). In addition, visual comparisons in Fig. 4 (e-h) also plainly illustrates that ExCEL generates better CAMs with the designed patch-text alignment paradigm." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20228" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.33 + ], + "angle": 0, + "content": "Figure 4. CAM visualizations on VOC train set. (a) Image. (b-e) Ablative visualizations of proposed modules. (e-h) Qualitative comparisons of (e) ExCEL and recent CLIP-based methods, i.e., (f) WeCLIP [43], (g) CLIP-ES [20] and (h) MaskCLIP [47]. (i) Ground truth." + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.341, + 0.439, + 0.354 + ], + "angle": 0, + "content": "Table 3. Ablation study of ExCEL on VOC val set." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.355, + 0.486, + 0.442 + ], + "angle": 0, + "content": "
ConditionsSVCTSELVCPrecisionRecallmIoU
Baseline (CLIP)18.821.312.1
w/ SVC✓81.286.272.5
w/o LVC✓✓80.789.874.7
w/o TSE✓✓83.786.375.1
ExCEL✓✓✓85.088.477.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.454, + 0.473, + 0.467 + ], + "angle": 0, + "content": "Table 4. Ablation study of attribute number \\( B \\) on VOC val set." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.468, + 0.486, + 0.508 + ], + "angle": 0, + "content": "
Number of AttrNone3264112144196
mIoU75.175.876.277.277.076.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.52, + 0.455, + 0.533 + ], + "angle": 0, + "content": "Table 5. Ablation study of VC module on VOC train set." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.534, + 0.486, + 0.621 + ], + "angle": 0, + "content": "
Conditionsq-kvI.C.M.C.LVCPrecisionRecallmIoU
Baseline (CLIP)✓18.021.811.2
MaskCLIP✓77.180.965.8
w/ I.C.✓79.184.769.7
SVC✓✓82.288.274.6
ExCEL✓✓✓86.687.978.0
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.255, + 0.649 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.658, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Efficacy of Key Components. Quantitative ablative experiments of ExCEL are reported in Tab. 3. The baseline is the vanilla CLIP using our training settings, which only achieves \\(12.1\\%\\) mIoU for segmentation. Our SVC module replaces the q-k attention with Intra-correlation from the intermediate layers. The performance increases to \\(72.5\\%\\). TSE module enriches the semantics of text representation for robust visual recognition. Introducing TSE brings a \\(3.6\\%\\) recall increase compared to the original text templates. LVC module provides a dynamic shift to diversify the features. It further benefits SVC's segmentation performance to \\(75.1\\%\\) mIoU. With all these enhancements, ExCEL generates \\(77.2\\%\\) mIoU for segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Qualitative ablation results are further illustrated in Fig. 4 (b-e) to evaluate the efficacy of our modules. In Fig. 4 (b), the CLIP baseline produces inaccurate CAMs with mis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.478 + ], + "angle": 0, + "content": "localized activation. SVC corrects token relations and preserves fine-grained details, effectively suppressing false activations, as seen in Fig. 4 (c). TSE incorporates comprehensive textual attributes into the text representation, enhancing patch-text matching and producing more complete CAMs, shown in Fig. 4 (d). LVC dynamically optimizes attention maps, further improving CAM accuracy and completeness, as illustrated in Fig. 4 (e). Both quantitative and qualitative results confirm the effectiveness of our modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Effectiveness of Implicit Attributes. Tab. 4 analyzes the effect of varying the number of clustering attributes. 'None' means no clustering and we explicitly fuse the \\( n \\) description embeddings for each class. It shows that the performance drops from \\( 77.2\\% \\) to \\( 75.1\\% \\), which validates the efficacy of implicit attributes and its superiority over explicit descriptions. With this operation, we can expand the representation of 20 classes up to 196 attributes or more, greatly enhancing text semantics. It reports that ExCEL achieves more favorable performance when \\( B \\) is set to 112." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.826 + ], + "angle": 0, + "content": "Effectiveness of Visual Calibrations. Tab. 5 compares different strategies in VC module for CAM generation. I.C. refers to Intra-correlation in the last layer, and M.C. (Intermediate Calibration) applies I.C. across intermediate layers. Vanilla q-k attention in CLIP loses diversity and cannot generate reasonable CAMs. \\( v \\) contains fine-grained knowledge and MaskCLIP improves CAMs to \\( 65.8\\% \\) by using \\( v \\) from the last layer. In contrast, our I.C. and M.C. focus on mining diverse knowledge from intermediate layers and boost the performance to \\( 69.7\\% \\) and \\( 74.6\\% \\). In addition, introducing LVC module raises the final performance to \\( 78.0\\% \\). Results in Tab. 5 and corresponding visualizations in Fig. 4 clearly highlight the efficacy of our components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.682, + 0.85 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Hyper-parameter Analysis. Hyper-parameters, such as TOP-K, scaling factors \\(\\alpha\\) and \\(\\beta\\), and the number of SVC layers \\(N\\), etc., are discussed in Supplementary Materials." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20229" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 6. Comparisons with the fully-supervised counterparts on VOC val set. \\( \\mathcal{F} \\) :fully-supervised. ViT-B*: pretrained from CLIP." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.117, + 0.487, + 0.309 + ], + "angle": 0, + "content": "
MethodsTypeSup.Net.ValRatio
DeepLabV2 [4] TPAMI'2017-FRN10177.7-
DeepLabV2 [4] TPAMI'2017-FViT-B82.3-
WeCLIP-Full [43] CVPR'2024-FViT-B*81.6-
CLIMS [36] CVPR'2022MI + LRN10170.490.6%
CLIP-ES [20] CVPR'2023MI + LRN10172.292.9%
CPAL [31] CVPR'2024MI + LRN10174.595.9%
ToCo [29] CVPR'2024SIViT-B71.186.4%
DuPL [35] CVPR'2024SIViT-B73.389.1%
SeCo [43] CVPR'2024SIViT-B74.089.9%
DIAL [13] ECCV'2024SI + LViT-B74.590.5%
WeCLIP [43] CVPR'2024SI + LViT-B*76.493.6%
ExCEL (Ours)SI + LViT-B*78.496.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.316, + 0.483, + 0.343 + ], + "angle": 0, + "content": "Table 7. Training efficiency comparisons on VOC train (CAM) and val set (Seg). All experiments are conducted on RTX 3090." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.344, + 0.486, + 0.495 + ], + "angle": 0, + "content": "
MethodTypeTraining TimeGPUCAMSeg
CLIMS [36] CVPR'2022M1068 mins18.0 G56.670.4
CLIP-ES [20] CVPR'2023M420 mins12.0 G70.872.2
MCTformer+ [38] TPAMI'2024M1496 mins18.0 G68.874.0
ToCo [29] CVPR'2023S506 mins17.9 G71.671.1
DuPL [35] CVPR'2024S508 mins14.9 G75.073.3
SeCo [39] CVPR'2024S407 mins17.6 G74.874.0
WeCLIP [43] CVPR'2024S270 mins6.2 G75.476.4
ExCEL* (Training-free)S-2.9 G74.6-
ExCEL (Ours)S90 mins3.2 G78.078.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.628 + ], + "angle": 0, + "content": "Fully-supervised Counterparts. Tab. 6 presents a fairness comparison between WSSS methods and their fully-supervised counterparts using the same segmentation backbone. With CLIP's visual encoder as the backbone, ExCEL achieves \\(78.4\\%\\) mIoU, reaching \\(96.1\\%\\) of the fully-supervised performance. It significantly outperforms CLIP-based WeCLIP by \\(2.5\\%\\) and demonstrates ExCEL's advantage over other multi-stage CLIP-based methods as well." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.63, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Training Efficiency Analysis. Our method only trains the adapter and decoder in a single-stage paradigm. Tab. 7 compares the training efficiency between ExCEL and recent methods. Without training, ExCEL requires just 2.9 GB of GPU memory and generates comparable CAMs to recent SOTAs. When training is included, the entire pipeline only takes 90 minutes and 3.2 GB of memory for SOTA performance. ExCEL just requires \\(6.0\\%\\) training time of multi-stage MCTformer+ and \\(33.3\\%\\) of single-stage WeCLIP, highlighting ExCEL's remarkable training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Attribute Response Analysis. We treat text prompting as an implicit attribute-hunting process to comprehensively enrich text representation semantics. To evaluate if the clustered attributes capture distinct object characteristics, we visualize 5 implicit attributes based on similarity scores. As shown in Fig. 5, given instances of \\(\\{\\text{aeroplane}\\}\\) and \\(\\{\\text{train}\\}\\), our attributes highlight different object parts, which clearly validates that our TSE module enhances inte" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.901, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.175, + 0.905, + 0.202 + ], + "angle": 0, + "content": "Figure 5. Implicit attribute responses. Based on the TOPK similarity scores, 5 attributes are sampled for visualizations." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.206, + 0.909, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.316, + 0.905, + 0.358 + ], + "angle": 0, + "content": "Figure 6. Comparisons of attention maps from the last visual encoder layer. (a) Attention features from the query patches (marked by red stars). (b) Token relations measured by cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.868, + 0.379 + ], + "angle": 0, + "content": "gral visual responses by gathering relevant semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Feature Representation Analysis. CLIP lacks fine-grained details, leading to inaccurate patch-text responses. To explore further, we visualize the self-attention features in Fig. 6 (a). Given the query patch (red star), CLIP's q-k attention falls short in generating diverse features, supporting our claim in Sec. 3.3. MaskCLIP observes that \\( v \\) keeps diversity and takes it from the last layer for visual response. We visualize it by calculating \\( v - v \\) attention. Although effective, MaskCLIP still misses fine granularity. Instead, SVC calculates attention within each space and implements it from intermediate layers. LVC further diversifies the features with a dynamic adapter, both of which effectively generate features with clear boundaries and spatial details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.682 + ], + "angle": 0, + "content": "Additionally, we explore the pairwise token relations in Fig. 6 (b). Unlike the smoother attention maps of CLIP or MaskCLIP, our approach distinctly groups tokens with similar semantics, aligning pairwise similarities with corresponding semantics. This validates that ExCEL successfully enhances the frozen features of CLIP by calibrating it towards distributions with more diverse spatial information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.633, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose ExCEL, a novel patch-text alignment method to explore CLIP's dense knowledge for WSSS, which provides a different insight to generate better pseudo labels based on CLIP. To this end, Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules are designed to improve the dense alignment across text and vision modalities. In addition, ExCEL generates CAMs in both training-free and efficient training modes, calibrating CLIP without altering its pre-trained weights. It retains CLIP's transferability while significantly reducing training cost. We believe ExCEL can inspire more future research to unlock CLIP's dense capabilities in the WSSS field." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20230" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant No.82372097, Shanghai Sailing Program under Grant 22YF1409300, International Science and Technology Cooperation Program under the 2023 Shanghai Action Plan for Science under Grant 23410710400, Taishan Scholars Program under Grant NO.tsqn202408245." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, pages 4981-4990, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.316, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 549-565. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[3] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, pages 1108-1118, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.565 + ], + "angle": 0, + "content": "[5] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, pages 3135-3144, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[6] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, pages 969–978, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[7] Zesen Cheng, Pengchong Qiao, Kehan Li, Siheng Li, Pengxu Wei, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Out-of-candidate rectification for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23673-23684, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, pages 1635-1643, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, pages 4320–4329, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.259, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[13] Soojin Jang, Jungmin Yun, Junehyoung Kwon, Eunju Lee, and Youngbin Kim. Dial: Dense image-text alignment for weakly supervised semantic segmentation. arXiv preprint arXiv:2409.15801, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[14] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[15] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 24, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[16] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[17] Xueyi Li, Tianfei Zhou, Jianwu Li, Yi Zhou, and Zhaoxiang Zhang. Group-wise semantic mining for weakly supervised semantic segmentation. In AAAI, pages 1984-1992, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.478, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In CVPR, pages 3159-3167, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[20] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. arXiv preprint arXiv:2212.09506, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[21] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[22] Balamurali Murugesan, Rukhshanda Hussain, Rajarshi Bhattacharya, Ismail Ben Ayed, and Jose Dolz. Prompting classes: exploring the power of prompt class learning in weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 291-302, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[23] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In CVPR, pages 1713-1721, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20231" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[26] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[27] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, pages 446-463. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[28] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: end-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[29] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. arXiv preprint arXiv:2303.01267, 2023. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[31] Feilong Tang, Zhongxing Xu, Zhaojun Qu, Wei Feng, Xingjian Jiang, and Zongyuan Ge. Hunting attributes: Context prototype-aware learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3324-3334, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[32] Paul Vernaza and Manmohan Chandraker. Learning random-walk label propagation for weakly-supervised semantic segmentation. In CVPR, pages 7158-7166, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[33] Xiang Wang, Shaodi You, Xi Li, and Huimin Ma. Weakly-supervised semantic segmentation by iteratively mining common object features. In CVPR, pages 1354-1362, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[34] Yuanchen Wu, Xiaogiang Li, Jide Li, Pinpin Zhu, Shaohua Zhang, et al. Dino is also a semantic guider: Exploiting class-aware affinity for weakly supervised semantic segmentation. In ACM Multimedia 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, pages 3534-3543, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Jinheng Xie, Xianxu Hou, Kai Ye, and Linlin Shen. Clims: cross language image matching for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4483-4492, 2022. 1, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[37] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[39] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, pages 3606–3615, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[40] Zhiwei Yang, Yuong Meng, Kexue Fu, Shuo Wang, and Zhijian Song. Tackling ambiguity from perspective of uncertainty inference and affinity diversification for weakly supervised semantic segmentation. arXiv preprint arXiv:2404.08195, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[41] Sung-Hoon Yoon, Hoyong Kwon, Hyeonseong Kim, and Kuk-Jin Yoon. Class tokens infusion for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3595-3605, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[42] Bingfeng Zhang, Xuru Gao, Siyue Yu, and Weifeng Liu. Enhanced online cam: Single-stage weakly supervised semantic segmentation via collaborative guidance. Pattern Recognition, 156:110787, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[43] Bingfeng Zhang, Siyue Yu, Yunchao Wei, Yao Zhao, and Jimin Xiao. Frozen clip: A strong backbone for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3796-3806, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[44] Xinqiao Zhao, Feilong Tang, Xiaoyang Wang, and Jimin Xiao. Sfc: Shared feature calibration in weakly supervised semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7525-7533, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[45] Xinqiao Zhao, Ziqian Yang, Tianhong Dai, Bingfeng Zhang, and Jimin Xiao. Psdpm: Prototype-based secondary discriminative pixels mining for weakly supervised semantic segmentation. In CVPR, pages 3437-3446, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[47] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.839 + ], + "angle": 0, + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.841, + 0.905, + 0.895 + ], + "angle": 0, + "content": "[49] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In CVPR, pages 4299-4309, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20232" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_origin.pdf b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..00501e0a5e97e403814263af27c8f8328896bfee --- /dev/null +++ b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/051b2c9f-5a13-4131-b880-e276059baaff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16f12494170ff3a699270fe63188997de460919f3e61888e134ae87a45ca3256 +size 3949858 diff --git a/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/full.md b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..51321c3de343a2656b3203200f927b69cb8cf71d --- /dev/null +++ b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/full.md @@ -0,0 +1,400 @@ +# Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation + +Zhiwei Yang $^{1,2}$ Yucong Meng $^{2,3}$ +Kexue Fu $^{4}$ Feilong Tang $^{1}$ Shuo Wang $^{2,3^{*}}$ Zhijian Song $^{1,2,3^{*}}$ + +$^{1}$ Academy for Engineering and Technology, Fudan University, Shanghai 200433, China + $^{2}$ Shanghai Key Laboratory of Medical Image Computing and Computer Assisted Intervention + $^{3}$ Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, China + $^{4}$ Shandong Computer Science Center (National Supercomputer Center in Jinan) + +# Abstract + +Weakly Supervised Semantic Segmentation (WSSS) with image-level labels aims to achieve pixel-level predictions using Class Activation Maps (CAMs). Recently, Contrastive Language-Image Pre-training (CLIP) has been introduced in WSSS. However, recent methods primarily focus on image-text alignment for CAM generation, while CLIP's potential in patch-text alignment remains unexplored. In this work, we propose ExCEL to explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. Specifically, we propose Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules to improve the dense alignment across both text and vision modalities. To make text embeddings semantically informative, our TSE module applies Large Language Models (LLMs) to build a dataset-wide knowledge base and enriches the text representations with an implicit attribute-hunting process. To mine fine-grained knowledge from visual features, our VC module first proposes Static Visual Calibration (SVC) to propagate fine-grained knowledge in a non-parametric manner. Then Learnable Visual Calibration (LVC) is further proposed to dynamically shift the frozen features towards distributions with diverse semantics. With these enhancements, ExCEL not only retains CLIP's training-free advantages but also significantly outperforms other state-of-the-art methods with much less training cost on PASCAL VOC and MS COCO. Code is available at https://github.com/zwyang6/ExCEL. + +# 1. Introduction + +Weakly Supervised Semantic Segmentation (WSSS) intends to generate pixel-level predictions using weak annotations like points [2], scribbles [18, 32], bounding boxes [8, + +![](images/c7b90d3683d00401b2759d0fc96acc05aa2121d9bcfb5f1ccee7270597c57c67.jpg) + +![](images/6d8ebe21c87dcf4ca0fe5bf33f1f13a18e5088fb87af9751159439d68c40fa7a.jpg) +Figure 1. Our motivation. (a) Previous methods leverage CLIP to generate CAMs with global image-text alignment, leaving CLIP's dense knowledge unexplored. (b) The proposed ExCEL explores CLIP's dense knowledge via a novel patch-text alignment paradigm, which generates better CAMs with less training cost. + +16], or image-level labels [1, 23, 40]. It significantly reduces the annotating cost of fully supervised methods and has attracted increasing attention in the community. Among all cheap annotation types, most WSSS approaches leverage image-level labels to provide dense localization cues, linking visual concepts to specific pixel regions [5, 33]. In this work, we focus on WSSS with image-level labels as well. + +Commonly, the WSSS pipeline involves three stages: generating Class Activation Maps (CAMs) [46] by training a classification network, refining CAMs into pseudo labels [28], and using these labels to train a segmentation model [7]. However, due to the minimal semantic information from image-level labels, CAMs intend to highlight the most distinctive object parts, significantly limiting WSSS performance. Recently, Contrastive Language-Image Pre-training (CLIP) [24] has been introduced in WSSS. CLIMs [36] applies image-text pairs to regularize visual relations among different semantics. CLIP-ES [20] + +leverages image-text alignment for gradient and produces high-quality GradCAM [30]. WeCLIP [43] further streamlines this process by using CLIP's visual encoder for segmentation. Despite these advancements, current methods primarily focus on CLIP's global image-text alignment, as shown in Fig. 1 (a). CLIP's dense knowledge with patch-text alignment still remains under-explored in WSSS. + +In this work, we propose ExCEL to explore CLIP's dense knowledge via a patch-text alignment paradigm for WSSS, i.e., generating CAMs by calculating patch-wise similarity between text and individual patch tokens, as shown in Fig. 1 (b). We identify two key challenges: (1) Semantic sparsity in textual prompts, where the template 'a photo of [CLASS]' only indicates object presence but lacks knowledge for localization, and (2) Fine-grained insufficiency in visual features, as CLIP prioritizes global representation due to its image-text pairing nature. To address these, ExCEL enhances CLIP's dense alignment with Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules, unlocking its potential across text and vision modalities. + +To generate semantically rich text representations, we propose TSE through an implicit attribute feature space. Instead of relying on explicit text templates, TSE module implicitly constructs text embeddings using universal attributes across the dataset. We first employ Large Language Models (LLMs) to generate detailed descriptions for each class, which are then processed by CLIP's text encoder to build a dataset-wide knowledge base. Rather than directly fusing these class-specific descriptions for text prompting, we focus on clustering the descriptive embeddings into generalized attributes, which effectively capture complementary knowledge from other classes and supplement missing information for the target class. With this implicit feature space, we enhance the text embeddings by hunting for its most relevant attribute features and aggregate them into the final class-specific text representation. This approach enables TSE to generate more informative text embeddings, providing a strong foundation for visual recognition. + +To mine fine-grained knowledge from visual features, we propose VC to calibrate CLIP in both non-trainable and efficient-learnable ways. Our findings suggest that CLIP's q-k attention loses fine-grained details. Therefore, we first propose a Static Visual Calibration (SVC) module to replace the suboptimal q-k attention with a straightforward Intra-correlation operation. It focuses on extracting fine-grained details from intermediate layers, which progressively propagates fine-grained visual knowledge. Without any retraining, SVC generates CAMs comparable to training-required WSSS methods. Building on this, we further propose a Learnable Visual Calibration (LVC) module to dynamically calibrate CLIP's frozen features. LVC extracts spatial correlations from SVC's static CAMs. These correlations further supervise a lightweight adapter to learn + +the dynamic shift, pushing frozen features towards spatial-aware distributions. LVC and SVC complement each other, enabling precise patch-text alignment for CAM generation. + +The main contributions of our work are listed as follows: + +- We explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. The proposed ExCEL generates better pseudo labels in both training-free and efficient learning manners, revealing the dense capabilities of CLIP for efficient CAM generation. +- To enhance patch-text alignment, we propose the Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules. TSE applies LLMs to build a dataset-wide knowledge base and treats text prompting as an implicit attribute-hunting process, making text embeddings more informative. VC propagates the fine-grained visual knowledge in a non-parametric manner and further dynamically calibrates the frozen features with a lightweight adapter. TSE and VC work across two modalities, generating better dense alignment and pseudo labels. +- Extensive experiments on PASCAL VOC and MS COCO demonstrate that ExCEL significantly outperforms recent state-of-the-art methods, while reducing training cost with only 3.2 GB of GPU memory and $6\%$ of the training time required by recent methods. + +# 2. Related Works + +# 2.1. Weakly Supervised Semantic Segmentation + +Weakly supervised semantic segmentation with image-level labels typically relies on CAMs to provide dense supervision for segmentation [42, 44]. However, CAMs usually highlight the most discriminative parts of objects [34]. To address this issue, considerable efforts have been made from many intriguing insights. MCTformer [37] incorporates multiple class tokens in Vision Transformer and proposes generating CAMs from class-patch attention. ToCo [29] proposes token contrast learning and generates more precise CAMs. SeCo [39] designs a separate and conquer scheme and succeeds in tackling co-occurrence. Despite these advancements, prior methods commonly require retraining the entire classification network for CAM generation. In this work, our ExCEL directly generates CAMs from frozen CLIP, and further boosts its quality via a lightweight adapter, significantly reducing the training cost. + +# 2.2. Vision-Language Pre-training + +Contrastive Language Image Pre-training (CLIP) [24], known for pretraining on billion-scale image-text pairs, has demonstrated remarkable transferability in many downstream tasks. CoOP [48] and CLIP-Adapter [12] incorporate lightweight trainable parameters into CLIP and succeed in few-shot classification. DenseCLIP [25] and + +![](images/2e36c53b2f5353223a5823b875eba234946566a301e81563cf57415f6f9f75e0.jpg) +Figure 2. ExCEL Architecture. We explore CLIP's dense knowledge with Text Semantic Enrichment (TSE) and Visual Calibration (VC). (a) TSE uses LLMs to build a knowledge base and clusters it into an implicit attribute space. The final text representation $T_{c}$ is enhanced by hunting for relevant attributes. For vision modality, (b) we introduce Static Visual Calibration (SVC) to calibrate visual features using the Inter-correlation operation across $N$ intermediate layers. It generates static CAMs with $T_{c}$ and calibrated features $P_{s}$ . (c) Learnable Visual Calibration (LVC) designs a learnable adapter to add a dynamic shift $R$ to SVC. It generates optimized features $P_{d}$ based on static CAMs guidance, creating dynamic CAMs from $P_{d}$ and $T_{c}$ . Dynamic CAMs are refined for segmentation supervision. Details are in Sec. 3.1. + +MaskCLIP [47] leverage the alignment between text and vision modalities for the dense segmentation task. Recently, some studies have introduced CLIP into WSSS. CLIMS [36] treats image-text pairing of CLIP as regularization and leverages it to regularize the visual concepts. CLIP-ES [20] finds that the image-text alignment of CLIP generates class gradient and leverages it for GradCAM generation. WeCLIP [43] further streamlines this process and directly leverages CLIP's visual encoder for segmentation. However, these methods mainly focus on a global image-text alignment while ignoring the dense capabilities of CLIP. In contrast, our ExCEL explores CLIP's dense knowledge via a patch-text alignment paradigm for WSSS. + +# 3. Methodology + +# 3.1. Preliminaries + +Patch-text CAM Generation. CLIP uses image and text encoders to project image and text into the same feature space, enabling robust vision-language alignment. In this work, we utilize this property to generate CAM in a patch-text alignment paradigm. Given the encoded text embeddings $T \in \mathbb{R}^{D \times C}$ and visual features $P \in \mathbb{R}^{h \times w \times D}$ , where $D$ is the feature channel, $h$ and $w$ are the spatial sizes, $C$ is the number of classes, we generate CAM by calculating the patch-wise similarities between text and visual features: + +$$ +\mathrm {C A M} = \operatorname {N o r m} (\cos (P, T), \tag {1} +$$ + +where $\mathrm{Norm}(\cdot)$ is the min-max normalization and $\cos (\cdot)$ is the cosine similarity calculation. However, due to + +the image-level pairing nature, CLIP suffers from textual semantic sparsity and visual fine-grained insufficiency. Therefore, our ExCEL incorporates TSE and VC (SVC and LVC) into CLIP to further explore its dense potential. + +Framework Overview. Our ExCEL generates CAMs in both training-free and efficient learning manners, as shown in Fig. 2. Its training pipeline is generalized as follows: + +(1) Enriching textual semantics via TSE. We first use GPT-4 to generate descriptions for each class, which are encoded into a dataset-wide knowledge base with CLIP's text encoder. We cluster this knowledge into class-agnostic attributes and use the global text prompt to hunt for its most relevant ones. They are then aggregated into the final text representation. (2) Static CAM generation via SVC. We replace CLIP's q-k self-attention with our Intra-correlation operation from intermediate layers. Then the calibrated visual features and enhanced text embeddings are used for static CAMs via Eq. (1). (3) Dynamic CAM generation via LVC. A lightweight adapter is designed to learn dynamic token relations from static CAMs. The relations are added to SVC and serve as a distribution shift to make the visual features more diverse. The dynamic CAMs are generated with the enhanced text embeddings and LVC features via Eq. (1). (4) Segmentation training. Dynamic CAMs are refined to pseudo labels for segmentation supervision. + +# 3.2. Text Semantic Enrichment + +Knowledge Base Construction. The global text template 'a clean origami of [CLASS]' only indicates the presence of objects while limited providing dense knowledge for + +patch-wise visual recognition. To enrich the text representation, we first adapt LLMs, such as GPT-4 to generate detailed class descriptions, as shown in Fig. 2 (a). Specifically, given the global template $E_{c}$ from class label space $Y \in \{1,2,\dots,C\}$ , where $c$ is the class index and $C$ is the number of categories, we carefully construct instructions for GPT: "List $n$ descriptions with key properties to describe the [CLASS] in terms of appearance, color, shape, size, or material, etc. These descriptions will help visually distinguish the [CLASS] from other classes in the dataset. Each description should follow the format: 'a clean origami [CLASS]. it + descriptive contexts.' With this instruction, we ask GPT to generate $n$ detailed descriptions for each class, which are subsequently encoded into a dataset-wide knowledge base with CLIP's text encoder. The knowledge base is denoted as $\mathcal{T} = \{\Phi(e_i)\}_{i=1}^{n \times C}$ . $e_i$ is the description from GPT, $\Phi(\cdot)$ is CLIP's text encoder and $n \times C$ is the number of all descriptions. This knowledge base gathers descriptive properties for the whole dataset, building a strong foundation for the textual category representation. + +Implicit Attribute Hunting. Instead of explicitly merging class-related knowledge into a single text embedding, we cluster this knowledge into generalized attributes and treat text prompting as an implicit attribute-hunting process. This has two main benefits: (1) $n$ explicit descriptions may still be limited in covering all characteristics of the class. The clustered attributes efficiently capture shared contextual knowledge from other categories, supplementing missing information for target class recognition. (2) The generated descriptions inevitably contain redundant or noisy content. The use of attributes makes the knowledge more compact and representative, leading to precise text prompting. + +To this end, we leverage a clustering algorithm to generate multiple centroids based on the knowledge base. Each cluster centroid is viewed as the implicit attribute that represents a group of descriptions sharing similar properties: + +$$ +A = \operatorname {K m e a n s} (\mathcal {T}, B) = \left\{a _ {i} \right\} _ {i = 1} ^ {B}, \tag {2} +$$ + +where $B$ is the number of centroids and kmeans algorithm [21] is used for clustering for simplicity. $a_{i}\in \mathbb{R}^{D\times 1}$ represents the cluster centroid, i.e., the implicit attribute. + +With the attribute feature space $A$ , we first send the global template $E_{c}$ into the text encoder of CLIP to generate global text embedding $t_c \in \mathbb{R}^{D \times 1}$ . $D$ is the channel dimension. Then $t_c$ is leveraged to search for its most relevant attributes in the implicit attribute space. To exclude irrelevant attributes, we further propose to select TOPK attribute neighbors based on the similarity scores with $t_c$ : + +$$ +A _ {c} = \left\{a _ {j}: j \in \operatorname {a r g m a x} _ {\mathrm {T O P K}} \left\{t _ {c} ^ {T} a _ {j} \right\} _ {j = 1} ^ {B} \right\}. \tag {3} +$$ + +Finally, we gently aggregate the implicit attribute neighbors according to the corresponding similarity weights to + +$t_c$ and take the aggregated features as the complementary knowledge for textual semantic enrichment. The final text representation $T_{c}\in \mathbb{R}^{D\times 1}$ is denoted as: + +$$ +T _ {c} = t _ {c} + \lambda \sum_ {j = 1} ^ {K} \operatorname {s o f t m a x} \left(t _ {c} ^ {T} A _ {c}\right) a _ {j}, \tag {4} +$$ + +where $\lambda$ is the factor to balance the attribute information. + +# 3.3. Visual Calibrations + +Static Visual Calibration. Due to the image-text pairing nature of CLIP, the visual features lack fine-grained information, leading to unreasonable localization maps via patch-text alignment. To delve into this, let us review the self-attention mechanism of the original CLIP first. As shown in Fig. 2 (b), given the input image $X \in \mathbb{R}^{3 \times \mathcal{H} \times \mathcal{W}}$ we send it to the image encoder of CLIP, which includes 12 Attention layers in this work. $\mathcal{H} \times \mathcal{W}$ is the image size. For the feature $F_{l} \in \mathbb{R}^{D_{s} \times hw}$ from $l$ -th layer of CLIP, it is first projected into three different spaces $\{q, k, v\}$ , named query, key, and value, respectively. $q, k$ and $v$ have the same shape of $D_{s} \times hw$ , where $D_{s}$ and $hw$ represents the channel dimension and sequence length. Then the attention map between $q$ and $k$ is calculated by measuring the similarity: + +$$ +\operatorname {S A} (q, k) = \operatorname {s o f m a x} \left(q ^ {T} k / \sqrt {D _ {s}}\right), \tag {5} +$$ + +where $\mathrm{SA}(\cdot)$ is the calculation of attention. Then the output features $F_{l + 1}$ are generated by aggregating the tokens of $v$ according to similarity weights in the attention map. + +However, due to the inherent image-text alignment of CLIP, the original q-k attention produces overly uniform attention maps, homogenizing diverse tokens from $v$ to capture broad semantics for global image representation (see discussions in Sec. 4.4). It leads to inaccurate object recognition. MaskCLIP [47] holds a similar observation, supporting this claim by removing the final q-k attention layer and using $v$ from the last layer as the visual output to preserve diversity. In our work, we choose to replace the suboptimal q-k attention with a straightforward Intra-correlation operation and focus on extracting fine-grained details from intermediate layers. This non-parametric approach effectively mines spatial semantics in intermediate $\{q, k, v\}$ and avoids the smoothing effect of q-k attention, resulting in more consistent attention maps and improved object localization. + +Specifically, instead of generating q-k correlation, Intra-correlation calculates the attention within each space of $\{q,k,v\}$ across intermediate layers. The attention map $S_{attn}^{l}\in \mathbb{R}^{hw\times hw}$ from $l$ -th SVC layer is generated by: + +$$ +S _ {a t t n} ^ {l} = \sum w _ {i} \operatorname {S A} \left(O _ {i} ^ {l}, O _ {i} ^ {l}\right), O _ {i} ^ {l} \in \left\{q ^ {l}, k ^ {l}, v ^ {l} \right\}, \tag {6} +$$ + +where $w_{i}$ is the contribution weight for different correlation maps. $l\in \{12 - N,\dots,12\}$ and $N$ is the number of intermediate layers for this operation. Then $S_{attn}^{l}$ and $v^{l}$ are used to + +generate the output features. Finally, the calibrated features $P_{s} \in \mathbb{R}^{D \times h \times w}$ from the last layer is used to generate static CAM $CAM_{s}$ with text embedding $T_{c}$ via Eq. (1). + +Learnable Visual Calibration. Although ExCEL generates comparable CAMs without training, its performance is still limited by the fixed features in CLIP. To further unleash the dense potential of CLIP, we design a lightweight adapter to dynamically calibrate the visual features with diverse details. This adapter only incorporates a distribution shift to calibrate the fixed features without changing CLIP's pre-trained weights, thereby retaining CLIP's transferability and enhancing its dense performance for WSSS. + +Specifically, as shown in Fig. 2 (c), frozen features $F_{l}$ from 1-12th layer of CLIP are extracted to learn a dynamic feature $F_{d}$ via the adapter. The process is expressed as: + +$$ +F _ {d} = \operatorname {C o n v} (\operatorname {C o n c a t e} [ \delta_ {l} (F _ {l}) ] _ {l = 1} ^ {1 2}), \tag {7} +$$ + +where $F_{d}\in \mathbb{R}^{D_{d}\times hw}$ $D_{d}$ is the channel dimension. $\mathrm{Conv}(\cdot)$ is the convolution layer, $\mathrm{Concat}[\cdot ]$ is the concatenate operation that connects all features along channel dimension, and $\delta_l(\cdot)$ is the individual MLP layer for each $F_{l}$ Then $F_{d}$ is used to generate dynamic token relations by: + +$$ +r = \alpha (\cos (F _ {d}, F _ {d}) - \beta \overline {{\cos (F _ {d} , F _ {d})}}), \tag {8} +$$ + +where $r \in \mathbb{R}^{hw \times hw}$ , $\alpha$ and $\beta$ are the scaling and shifting factors to adjust the relations, respectively. $\overline{\cos(F_d, F_d)}$ means the mean value of similarity scores of $F_d$ . It is designed to remove the irrelevant relations in low values by: + +$$ +R _ {i j} = \left\{ \begin{array}{l l} r _ {i j}, & \text {i f} r _ {i j} \geq 0 \\ - i n f, & \text {e l s e} \end{array} . \right. \tag {9} +$$ + +With the dynamic relation $R \in \mathbb{R}^{hw \times hw}$ , we add it as a distribution bias to the static attention map $S_{attn}^{l}$ , dynamically grouping the frozen tokens within related semantics and shifting the features towards denser distribution. The optimized attention map $L_{attn}^{l} \in \mathbb{R}^{hw \times hw}$ is denoted as: + +$$ +L _ {a t t n} ^ {l} = S _ {a t t n} ^ {l} + \operatorname {s o f t m a x} (R). \tag {10} +$$ + +Subsequently, we extract the dynamically calibrated features $P_{d} \in \mathbb{R}^{D \times h \times w}$ from the last layer of LVC and generate dynamic CAMs with Eq. (1), which are then refined to final pseudo labels $M_{d}$ for the segmentation. + +# 3.4. Training Objectives + +We formulate a diversity loss to supervise the learning of $F_{d}$ in our LVC module. We first measure token correlations of $F_{d}$ by calculating the self similarity: $\hat{\mathcal{R}} = \mathrm{sigmoid}(\cos (F_d,F_d))$ , where sigmoid(·) is the activation function and $\hat{\mathcal{R}}\in \mathbb{R}^{hw\times hw}$ . Then we refine $CAM_{s}$ from SVC into static pseudo labels $M_{s}$ and leverage its pixelwise affinity to guide the diversifying of $\hat{\mathcal{R}}$ . Specifically, if + +the pixel with coordinate $(i,j)$ shows the same pixel value as the pixel in $(\varepsilon,\eta)$ , the token pair with the same coordinates on $F_{d}$ is semantically related and its corresponding correlation logit on $\hat{\mathcal{R}}$ should be maximized, and vice versa. The diversity loss can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {d i v}} = \frac {1}{N ^ {+}} \sum_ {u ^ {+} \in \hat {\mathcal {R}} ^ {+}} (1 - u ^ {+}) + \frac {1}{N ^ {-}} \sum_ {u ^ {-} \in \hat {\mathcal {R}} ^ {-}} u ^ {-}, \tag {11} +$$ + +where $N^{+} / N^{-}$ are the number of positive and negative pairs, $u^{+} / u^{-}$ are the positive and negative relation logits, and $\hat{\mathcal{R}}^{+} / \hat{\mathcal{R}}^{-}$ are the positive and negative sets of logit on $\hat{\mathcal{R}}$ , respectively. The diversity loss groups tokens with similar semantics and suppresses irrelevant ones, enhancing fine-grained details of visual features for precise text response. + +In addition, ExCEL is streamlined as a single-stage method. We adopt a lightweight Transformer-based segmentation head [43] and directly take the frozen visual encoder of CLIP for segmentation. The dynamic pseudo labels are used as supervision with a cross-entropy loss $\mathcal{L}_{seg}$ . The loss objectives of our ExCEL are formulated as: + +$$ +\mathcal {L} _ {\text {E X C E L}} = \mathcal {L} _ {s e g} + \gamma \mathcal {L} _ {\text {d i v}}, \tag {12} +$$ + +where $\gamma$ is the weight factor. By efficiently training the adapter and a segmentation head, ExCEL achieves strong WSSS performance and significantly reduces training cost. + +# 4. Experiments and Results + +# 4.1. Experimental Settings + +Datasets and Metrics. The proposed ExCEL is evaluated on PASCAL VOC 2012 [11] and MS COCO 2014 [19]. VOC contains 21 categories (1 for background). Following prior methods [10, 17, 39], the augmented dataset with 10,582, 1,449, and 1,456 images are used for training, validating, and testing, respectively. COCO includes 81 classes, in which 82,081 images are used for training and 40,137 images are for validating. Mean Intersection-Over-Union (mIoU) is used as the main evaluation metric. + +Implementation Details. CLIP model with ViT-B [9] is used as ExCEL's encoder, which is frozen during the training. For TSE module, we generate $n = 20$ descriptions from GPT-4 for each category. The number of attribute embeddings $B$ is set to 112 and 224 for PASCAL VOC and MS COCO, respectively. The SVC module is conducted in the last $N = 5$ Transformer layers. Our decoder adopts a simple Transformer-based head [43]. Features $F_{l}$ from each layer of CLIP are sent to it for the segmentation predictions. The scaling and shifting factors in Eq. (8) are set as 3.0 and 1.0, respectively. The loss weight $\gamma$ is set as 0.1. Following [29, 35, 39], the AdamW optimizer is used for training the adapter and decoder. The learning rate is 1e-4 with a weight decay of 1e-2. The training iteration is set + +![](images/c5301420fc8865447c8c1f658f64629bb88710e6c0b272f98be0b0cf98101ae9.jpg) + +![](images/3cf356d6bcb85be39af554e85aa9ae60da491f5ae8199b38fed976f53559b185.jpg) + +![](images/8341b6c0a8a3382c0fb8af763250a9d559aec1198a864dccbdcd7b63a33a9513.jpg) +PASCAL VOC + +![](images/22f512584c0be30c1ffaf594f4585075fc84962a1609bf59e9051c09a36a6978.jpg) + +![](images/851349b2259265cdcc9d4f7cb4c9d72faa4887a1d9dbcaa8204e7a4789403436.jpg) + +![](images/316efb51d73ee1432a28e419b77b6f213e78dc03a679cd6105c8eeffdb57b608.jpg) + +![](images/f74f7342c9bd38e85ccccf988a630ed6df648ccd4e7c37874fc3e2ed420d5b86.jpg) + +![](images/46ba46418105780174965328e844f1039f7cf6ba6bb845e56f6e9bdef5d4e93c.jpg) + +![](images/e2bb1f5eea6c8a12e3505f943f800f9431b6414b16d8c31bb778c1dccb2be748.jpg) + +![](images/a5114db16ca06ec6772bada96533feaa21babdba9f61e9360cb97d41fa12b449.jpg) + +![](images/0820de22073960a29f5d52de03d08656c62463f095e9caf7cb7da0c246c620d2.jpg) + +![](images/388bc4a2bd3c0cf2444b90089f1a7164064a2a0ed4260df95d41bf99e22ac2ca.jpg) + +![](images/bd7a6d018eedd9f55a445e37adfe8ba060d1495a084df2b77b4e9b12e5fa72f0.jpg) + +![](images/3599e883cd521f0c40f4618ba859159582be2d9bb94e3ce37c7ed5c27da298ac.jpg) + +![](images/2a099ccafe16b9d68dd8d4398051f24a7d45a477a1015a2f1c8922a759d0bec7.jpg) + +![](images/853857e3c93474f209433edaa13ebad39e31788efb3e014e2004141d2bb751af.jpg) +Figure 3. Segmentation visualizations of SeCo [39], WeCLIP [43] and ours on VOC and COCO. ExCEL segments objects more precisely. + +![](images/9c64ba882624693b3d4e3de98f783de003f855a3c4275e4711672fb1fe2215f7.jpg) + +![](images/a0c6051f088c68437c8d0d4e19df66cf6194ca4704b147958ca5c96eae17bf64.jpg) + +![](images/407257d2cc35efd54090f414c835ffd10fd347c1c94e2696287b9da91c60fe8e.jpg) + +![](images/4cb732dedb5053b10e25b17de0312daee744fe05fc297cd9680fa2b5c269fdff.jpg) + +![](images/5c8a2dde2dc5e2bb3062dc8d1fc52dc7e11c6f4da1b8a2a94e557f48afb7cf0a.jpg) + +![](images/615e9b4bfb1296f453373dbb8da005a45ce6215b90b367eb64649975ad83987e.jpg) + +![](images/c18013dc17f10be6b7fc5d01c0817a31a5a6306335896bafe651c90f39a8bd69.jpg) + +![](images/9a592c946fab0c5e254996fbb1027530458ab6932689dc33850d902fafe20365.jpg) + +![](images/d0c07d1dd506f1791c29c140bcee28232b00df3ebd8b1fbfe3a347a3d6a5507a.jpg) + +![](images/1c826c87e4fd50bf112a593962731b2eaf9d2af5fb9b2a326ef96ea612e58e3b.jpg) + +![](images/1abe07791c3f893850d03b10d98dc161b7e98d42d36f411662827af52d8210ae.jpg) + +![](images/fb5bb2657738c3517c6dada1ca94ff908721c365d9d43527dd77033d7b93b5b4.jpg) + +![](images/4f66e0660334b6e3fd8f15cdd3314761408886b96fe8ff5b0cdfdfac188c1fe8.jpg) + +![](images/5a5931862b2a698f9ebdd3ba5c5b52658585c531040dbc909fa7994f6d456f47.jpg) + +![](images/f610f88abaee40af22c521424154630cbceb42d45376f19195aa278bc495fc71.jpg) + +![](images/fc82a5c0ee1e47840032587580b809eff34cfb26483364aad4df08cb97eaef3f.jpg) + +![](images/4740cf6ccbaabd038a40da72d990c0341534e1a41b2dc24a286ebeac55b86433.jpg) + +![](images/1a70299b32eb9c9d4b781b775f287a4ed3775196df9a81909637a218f483f6eb.jpg) + +![](images/d6c7020eaf13999ff40c738a94b1b660907037ea78564be246790d95fe8eb66a.jpg) + +![](images/66e6889670bb1e0db1b5f0b19616a18f0e8840e66bf66004b44553c71ff4802c.jpg) + +Table 1. Segmentation comparisons on VOC and COCO. Net. is the backbone for segmentation. Sup. is the supervision type. $\mathcal{I}$ image-level labels. SA: saliency maps. $\mathcal{L}$ : language. + +
MethodSup.Net.VOCCOCO
ValTestVal
Multi-stage WSSS methods.
L2G [14] CVPR'2022I + SARN10172.171.744.2
RCA [49] CVPR'2023I + SARN3872.272.836.8
OCR [7] CVPR'2023IRN3872.772.042.5
BECO [26] CVPR'2023IRN10173.773.545.1
MCTformer+ [38] TPAMI'2024IRN3874.073.645.2
CTI [41] CVPR'2024IRN10174.173.245.4
CLIMS [36] CVPR'2022I + LRN10170.470.0-
CLIP-ES [20] CVPR'2023I + LRN10172.272.845.4
PSDPM [45] CVPR'2024I + LRN10174.174.947.2
CPAL [31] CVPR'2024I + LRN10174.574.746.8
Single-stage WSSS methods.
AFA [28] CVPR'2022IMiT-B166.066.338.9
ViT-PCM [27] ECCV'2022IViT-B70.370.9-
ToCo [29] CVPR'2023IViT-B71.172.242.3
DuPL [35] CVPR'2024IViT-B73.372.844.6
SeCo [39] CVPR'2024IViT-B74.073.846.7
DIAL [13] ECCV'2024I + LViT-B74.574.944.4
WeCLIP [43] CVPR'2024I + LViT-B76.477.247.1
ExCEL(w/o CRF)I + LViT-B77.277.349.3
ExCEL (Ours)I + LViT-B78.478.550.3
+ +as 30,000 for VOC and 100,000 for COCO. Please refer to Supplementary Materials for more details. + +# 4.2. Comparisons with State-of-the-art Methods + +Performance of Semantic Segmentation. Tab. 1 shows segmentation comparisons between our ExCEL and recent methods on VOC and COCO. The single-stage ExCEL achieves $78.4\%$ and $78.5\%$ mIoU on VOC val set and test set, which even significantly outperforms the sophisticated multi-stage methods by at least $3.9\%$ and $3.6\%$ mIoU, respectively. For more complicated benchmark COCO, ExCEL achieves $50.3\%$ mIoU on val set, which brings a noticeable $3.2\%$ increase over the CLIP-based state-of-the-art (SOTA) WeCLIP. In addition, without time-consuming post-processing techniques, such as CRF [15], ExCEL still + +Table 2. CAM seed comparisons on VOC train set. $\mathcal{M}$ : multistage methods. $S$ : single-stage methods. †: our reproduction following official codes. ExCEL*: ExCEL in a training-free manner. + +
MethodTypeSup.Net.VOC
Train
Training-free WSSS methods.
CLIP-ES [20] CVPR'2023MI + LViT-B70.8
ExCEL* (Ours)SI + LViT-B74.6
Training-required WSSS methods.
ReCAM [6] CVPR'2022MIRN10154.8
FPR [3] CVPR'2023MIRN10163.8
LPCAM [5] CVPR'2023MIRN5065.3
MCTformer+ [38] TPAMI'2024MIRN3868.8
SFC [44] AAAI'2024MIRN10164.7
CTI [41] CVPR'2024MIRN10169.5
AFA [28] CVPR'2022SIMiT-B165.0
ViT-PCM [27] ECCV'2022SIViT-B67.7
†ToCo [29] CVPR'2023SIViT-B71.6
†DuPL [35] CVPR'2024SIViT-B75.0
SeCo [39] CVPR'2024SIViT-B74.8
CLIMS [36] CVPR'2022MI + LRN10156.6
POLE [22] WACV'2023MI + LRN5059.0
CPAL [31] CVPR'2024MI + LRN10171.9
DIAL [13] ECCV'2024SI + LViT-B75.2
†WeCLIP [43] CVPR'2024SI + LViT-B75.4
ExCEL (Ours)SI + LViT-B78.0
+ +maintains consistent superiority over SOTAs with CRF. + +The qualitative comparisons on VOC and COCO are shown in Fig. 3. By densely matching the patches and texts, ExCEL consistently demonstrates more precise object segmentation than recent methods in an image-text paradigm. + +Evaluation of CAM Seeds. Tab. 2 reports the quality of raw CAM seeds on VOC train set. Compared with recent methods, ExCEL achieves $74.6\%$ mIoU in a training-free setup, outperforming CLIP-ES by $3.8\%$ and performing comparably to most training-required methods. With the optimized LVC module, ExCEL further boosts CAM quality to $78.0\%$ , surpassing SOTAs in the image-text paradigm by at least $2.6\%$ . In addition, visual comparisons in Fig. 4 (e-h) also plainly illustrates that ExCEL generates better CAMs with the designed patch-text alignment paradigm. + +![](images/f6fd9378d29824e2ce862dab5fc4b6daaa14aea99033ac822c2f0c8a6a7b0851.jpg) +Figure 4. CAM visualizations on VOC train set. (a) Image. (b-e) Ablative visualizations of proposed modules. (e-h) Qualitative comparisons of (e) ExCEL and recent CLIP-based methods, i.e., (f) WeCLIP [43], (g) CLIP-ES [20] and (h) MaskCLIP [47]. (i) Ground truth. + +Table 3. Ablation study of ExCEL on VOC val set. + +
ConditionsSVCTSELVCPrecisionRecallmIoU
Baseline (CLIP)18.821.312.1
w/ SVC✓81.286.272.5
w/o LVC✓✓80.789.874.7
w/o TSE✓✓83.786.375.1
ExCEL✓✓✓85.088.477.2
+ +Table 4. Ablation study of attribute number $B$ on VOC val set. + +
Number of AttrNone3264112144196
mIoU75.175.876.277.277.076.5
+ +Table 5. Ablation study of VC module on VOC train set. + +
Conditionsq-kvI.C.M.C.LVCPrecisionRecallmIoU
Baseline (CLIP)✓18.021.811.2
MaskCLIP✓77.180.965.8
w/ I.C.✓79.184.769.7
SVC✓✓82.288.274.6
ExCEL✓✓✓86.687.978.0
+ +# 4.3. Ablation Studies + +Efficacy of Key Components. Quantitative ablative experiments of ExCEL are reported in Tab. 3. The baseline is the vanilla CLIP using our training settings, which only achieves $12.1\%$ mIoU for segmentation. Our SVC module replaces the q-k attention with Intra-correlation from the intermediate layers. The performance increases to $72.5\%$ . TSE module enriches the semantics of text representation for robust visual recognition. Introducing TSE brings a $3.6\%$ recall increase compared to the original text templates. LVC module provides a dynamic shift to diversify the features. It further benefits SVC's segmentation performance to $75.1\%$ mIoU. With all these enhancements, ExCEL generates $77.2\%$ mIoU for segmentation. + +Qualitative ablation results are further illustrated in Fig. 4 (b-e) to evaluate the efficacy of our modules. In Fig. 4 (b), the CLIP baseline produces inaccurate CAMs with mis + +localized activation. SVC corrects token relations and preserves fine-grained details, effectively suppressing false activations, as seen in Fig. 4 (c). TSE incorporates comprehensive textual attributes into the text representation, enhancing patch-text matching and producing more complete CAMs, shown in Fig. 4 (d). LVC dynamically optimizes attention maps, further improving CAM accuracy and completeness, as illustrated in Fig. 4 (e). Both quantitative and qualitative results confirm the effectiveness of our modules. + +Effectiveness of Implicit Attributes. Tab. 4 analyzes the effect of varying the number of clustering attributes. 'None' means no clustering and we explicitly fuse the $n$ description embeddings for each class. It shows that the performance drops from $77.2\%$ to $75.1\%$ , which validates the efficacy of implicit attributes and its superiority over explicit descriptions. With this operation, we can expand the representation of 20 classes up to 196 attributes or more, greatly enhancing text semantics. It reports that ExCEL achieves more favorable performance when $B$ is set to 112. + +Effectiveness of Visual Calibrations. Tab. 5 compares different strategies in VC module for CAM generation. I.C. refers to Intra-correlation in the last layer, and M.C. (Intermediate Calibration) applies I.C. across intermediate layers. Vanilla q-k attention in CLIP loses diversity and cannot generate reasonable CAMs. $v$ contains fine-grained knowledge and MaskCLIP improves CAMs to $65.8\%$ by using $v$ from the last layer. In contrast, our I.C. and M.C. focus on mining diverse knowledge from intermediate layers and boost the performance to $69.7\%$ and $74.6\%$ . In addition, introducing LVC module raises the final performance to $78.0\%$ . Results in Tab. 5 and corresponding visualizations in Fig. 4 clearly highlight the efficacy of our components. + +# 4.4. Further Analysis + +Hyper-parameter Analysis. Hyper-parameters, such as TOP-K, scaling factors $\alpha$ and $\beta$ , and the number of SVC layers $N$ , etc., are discussed in Supplementary Materials. + +Table 6. Comparisons with the fully-supervised counterparts on VOC val set. $\mathcal{F}$ :fully-supervised. ViT-B*: pretrained from CLIP. + +
MethodsTypeSup.Net.ValRatio
DeepLabV2 [4] TPAMI'2017-FRN10177.7-
DeepLabV2 [4] TPAMI'2017-FViT-B82.3-
WeCLIP-Full [43] CVPR'2024-FViT-B*81.6-
CLIMS [36] CVPR'2022MI + LRN10170.490.6%
CLIP-ES [20] CVPR'2023MI + LRN10172.292.9%
CPAL [31] CVPR'2024MI + LRN10174.595.9%
ToCo [29] CVPR'2024SIViT-B71.186.4%
DuPL [35] CVPR'2024SIViT-B73.389.1%
SeCo [43] CVPR'2024SIViT-B74.089.9%
DIAL [13] ECCV'2024SI + LViT-B74.590.5%
WeCLIP [43] CVPR'2024SI + LViT-B*76.493.6%
ExCEL (Ours)SI + LViT-B*78.496.1%
+ +Table 7. Training efficiency comparisons on VOC train (CAM) and val set (Seg). All experiments are conducted on RTX 3090. + +
MethodTypeTraining TimeGPUCAMSeg
CLIMS [36] CVPR'2022M1068 mins18.0 G56.670.4
CLIP-ES [20] CVPR'2023M420 mins12.0 G70.872.2
MCTformer+ [38] TPAMI'2024M1496 mins18.0 G68.874.0
ToCo [29] CVPR'2023S506 mins17.9 G71.671.1
DuPL [35] CVPR'2024S508 mins14.9 G75.073.3
SeCo [39] CVPR'2024S407 mins17.6 G74.874.0
WeCLIP [43] CVPR'2024S270 mins6.2 G75.476.4
ExCEL* (Training-free)S-2.9 G74.6-
ExCEL (Ours)S90 mins3.2 G78.078.4
+ +Fully-supervised Counterparts. Tab. 6 presents a fairness comparison between WSSS methods and their fully-supervised counterparts using the same segmentation backbone. With CLIP's visual encoder as the backbone, ExCEL achieves $78.4\%$ mIoU, reaching $96.1\%$ of the fully-supervised performance. It significantly outperforms CLIP-based WeCLIP by $2.5\%$ and demonstrates ExCEL's advantage over other multi-stage CLIP-based methods as well. + +Training Efficiency Analysis. Our method only trains the adapter and decoder in a single-stage paradigm. Tab. 7 compares the training efficiency between ExCEL and recent methods. Without training, ExCEL requires just 2.9 GB of GPU memory and generates comparable CAMs to recent SOTAs. When training is included, the entire pipeline only takes 90 minutes and 3.2 GB of memory for SOTA performance. ExCEL just requires $6.0\%$ training time of multi-stage MCTformer+ and $33.3\%$ of single-stage WeCLIP, highlighting ExCEL's remarkable training efficiency. + +Attribute Response Analysis. We treat text prompting as an implicit attribute-hunting process to comprehensively enrich text representation semantics. To evaluate if the clustered attributes capture distinct object characteristics, we visualize 5 implicit attributes based on similarity scores. As shown in Fig. 5, given instances of $\{\text{aeroplane}\}$ and $\{\text{train}\}$ , our attributes highlight different object parts, which clearly validates that our TSE module enhances inte + +![](images/178f8d7f60bfa66fe2b0fd0d2b2000aaf26998844024d8e16377bf35b7bb58d7.jpg) +Figure 5. Implicit attribute responses. Based on the TOPK similarity scores, 5 attributes are sampled for visualizations. + +![](images/5d9e85d1175b51ab2d9ff7dd9fe40e8f79ab87d68d4b4f0c327585ba9676213e.jpg) +Figure 6. Comparisons of attention maps from the last visual encoder layer. (a) Attention features from the query patches (marked by red stars). (b) Token relations measured by cosine similarity. + +gral visual responses by gathering relevant semantics. + +Feature Representation Analysis. CLIP lacks fine-grained details, leading to inaccurate patch-text responses. To explore further, we visualize the self-attention features in Fig. 6 (a). Given the query patch (red star), CLIP's q-k attention falls short in generating diverse features, supporting our claim in Sec. 3.3. MaskCLIP observes that $v$ keeps diversity and takes it from the last layer for visual response. We visualize it by calculating $v - v$ attention. Although effective, MaskCLIP still misses fine granularity. Instead, SVC calculates attention within each space and implements it from intermediate layers. LVC further diversifies the features with a dynamic adapter, both of which effectively generate features with clear boundaries and spatial details. + +Additionally, we explore the pairwise token relations in Fig. 6 (b). Unlike the smoother attention maps of CLIP or MaskCLIP, our approach distinctly groups tokens with similar semantics, aligning pairwise similarities with corresponding semantics. This validates that ExCEL successfully enhances the frozen features of CLIP by calibrating it towards distributions with more diverse spatial information. + +# 5. Conclusion + +In this paper, we propose ExCEL, a novel patch-text alignment method to explore CLIP's dense knowledge for WSSS, which provides a different insight to generate better pseudo labels based on CLIP. To this end, Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules are designed to improve the dense alignment across text and vision modalities. In addition, ExCEL generates CAMs in both training-free and efficient training modes, calibrating CLIP without altering its pre-trained weights. It retains CLIP's transferability while significantly reducing training cost. We believe ExCEL can inspire more future research to unlock CLIP's dense capabilities in the WSSS field. + +# 6. Acknowledgments + +# References + +[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, pages 4981-4990, 2018. 1 +[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 549-565. Springer, 2016. 1 +[3] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, pages 1108-1118, 2023. 6 +[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8 +[5] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, pages 3135-3144, 2023. 1, 6 +[6] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, pages 969–978, 2022. 6 +[7] Zesen Cheng, Pengchong Qiao, Kehan Li, Siheng Li, Pengxu Wei, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Out-of-candidate rectification for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23673-23684, 2023. 1, 6 +[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, pages 1635-1643, 2015. 1 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, pages 4320–4329, 2022. 5 +[11] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The + +This work was supported by the National Natural Science Foundation of China under Grant No.82372097, Shanghai Sailing Program under Grant 22YF1409300, International Science and Technology Cooperation Program under the 2023 Shanghai Action Plan for Science under Grant 23410710400, Taishan Scholars Program under Grant NO.tsqn202408245. +pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 5 +[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 2 +[13] Soojin Jang, Jungmin Yun, Junehyoung Kwon, Eunju Lee, and Youngbin Kim. Dial: Dense image-text alignment for weakly supervised semantic segmentation. arXiv preprint arXiv:2409.15801, 2024. 6, 8 +[14] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 6 +[15] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 24, 2011. 6 +[16] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1 +[17] Xueyi Li, Tianfei Zhou, Jianwu Li, Yi Zhou, and Zhaoxiang Zhang. Group-wise semantic mining for weakly supervised semantic segmentation. In AAAI, pages 1984-1992, 2021. 5 +[18] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In CVPR, pages 3159-3167, 2016. 1 +[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[20] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. arXiv preprint arXiv:2212.09506, 2022. 1, 3, 6, 7, 8 +[21] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 4 +[22] Balamurali Murugesan, Rukhshanda Hussain, Rajarshi Bhattacharya, Ismail Ben Ayed, and Jose Dolz. Prompting classes: exploring the power of prompt class learning in weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 291-302, 2024. 6 +[23] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In CVPR, pages 1713-1721, 2015. 1 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi + +sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 2 +[26] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 6 +[27] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, pages 446-463. Springer, 2022. 6 +[28] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: end-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1, 6 +[29] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. arXiv preprint arXiv:2303.01267, 2023. 2, 5, 6, 8 +[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2 +[31] Feilong Tang, Zhongxing Xu, Zhaojun Qu, Wei Feng, Xingjian Jiang, and Zongyuan Ge. Hunting attributes: Context prototype-aware learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3324-3334, 2024. 6, 8 +[32] Paul Vernaza and Manmohan Chandraker. Learning random-walk label propagation for weakly-supervised semantic segmentation. In CVPR, pages 7158-7166, 2017. 1 +[33] Xiang Wang, Shaodi You, Xi Li, and Huimin Ma. Weakly-supervised semantic segmentation by iteratively mining common object features. In CVPR, pages 1354-1362, 2018. 1 +[34] Yuanchen Wu, Xiaogiang Li, Jide Li, Pinpin Zhu, Shaohua Zhang, et al. Dino is also a semantic guider: Exploiting class-aware affinity for weakly supervised semantic segmentation. In ACM Multimedia 2024. 2 +[35] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, pages 3534-3543, 2024. 5, 6, 8 +[36] Jinheng Xie, Xianxu Hou, Kai Ye, and Linlin Shen. Clims: cross language image matching for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4483-4492, 2022. 1, 3, 6, 8 + +[37] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2 +[38] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 2024. 6, 8 +[39] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, pages 3606–3615, 2024. 2, 5, 6, 8 +[40] Zhiwei Yang, Yuong Meng, Kexue Fu, Shuo Wang, and Zhijian Song. Tackling ambiguity from perspective of uncertainty inference and affinity diversification for weakly supervised semantic segmentation. arXiv preprint arXiv:2404.08195, 2024. 1 +[41] Sung-Hoon Yoon, Hoyong Kwon, Hyeonseong Kim, and Kuk-Jin Yoon. Class tokens infusion for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3595-3605, 2024. 6 +[42] Bingfeng Zhang, Xuru Gao, Siyue Yu, and Weifeng Liu. Enhanced online cam: Single-stage weakly supervised semantic segmentation via collaborative guidance. Pattern Recognition, 156:110787, 2024. 2 +[43] Bingfeng Zhang, Siyue Yu, Yunchao Wei, Yao Zhao, and Jimin Xiao. Frozen clip: A strong backbone for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3796-3806, 2024. 2, 3, 5, 6, 7, 8 +[44] Xinqiao Zhao, Feilong Tang, Xiaoyang Wang, and Jimin Xiao. Sfc: Shared feature calibration in weakly supervised semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7525-7533, 2024. 2, 6 +[45] Xinqiao Zhao, Ziqian Yang, Tianhong Dai, Bingfeng Zhang, and Jimin Xiao. Psdpm: Prototype-based secondary discriminative pixels mining for weakly supervised semantic segmentation. In CVPR, pages 3437-3446, 2024. 6 +[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, 2016. 1 +[47] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 4, 7 +[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2 +[49] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In CVPR, pages 4299-4309, 2022. 6 \ No newline at end of file diff --git a/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/images.zip b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb8e40535bef3e8b91275ad373b22ae04c75a437 --- /dev/null +++ b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:272f8c7ad3121f27edf55656a077065824633581e96c2a2f463fb98262bcbe1a +size 821238 diff --git a/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/layout.json b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2e86f66de7a6433326b95936cbbfacc60f34ffef --- /dev/null +++ b/2025/Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation/layout.json @@ -0,0 +1,11517 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "text", + "content": "Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "spans": [ + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "text", + "content": "Zhiwei Yang" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "text", + "content": " Yucong Meng" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "text", + "content": " \nKexue Fu" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "text", + "content": " Feilong Tang" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "text", + "content": " Shuo Wang" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "inline_equation", + "content": "^{2,3^{*}}" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "text", + "content": " Zhijian Song" + }, + { + "bbox": [ + 135, + 159, + 479, + 191 + ], + "type": "inline_equation", + "content": "^{1,2,3^{*}}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "spans": [ + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "text", + "content": "Academy for Engineering and Technology, Fudan University, Shanghai 200433, China \n" + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "text", + "content": "Shanghai Key Laboratory of Medical Image Computing and Computer Assisted Intervention \n" + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "text", + "content": "Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, China \n" + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 191, + 533, + 248 + ], + "type": "text", + "content": "Shandong Computer Science Center (National Supercomputer Center in Jinan)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "spans": [ + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 300, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 296, + 624 + ], + "type": "text", + "content": "Weakly Supervised Semantic Segmentation (WSSS) with image-level labels aims to achieve pixel-level predictions using Class Activation Maps (CAMs). Recently, Contrastive Language-Image Pre-training (CLIP) has been introduced in WSSS. However, recent methods primarily focus on image-text alignment for CAM generation, while CLIP's potential in patch-text alignment remains unexplored. In this work, we propose ExCEL to explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. Specifically, we propose Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules to improve the dense alignment across both text and vision modalities. To make text embeddings semantically informative, our TSE module applies Large Language Models (LLMs) to build a dataset-wide knowledge base and enriches the text representations with an implicit attribute-hunting process. To mine fine-grained knowledge from visual features, our VC module first proposes Static Visual Calibration (SVC) to propagate fine-grained knowledge in a non-parametric manner. Then Learnable Visual Calibration (LVC) is further proposed to dynamically shift the frozen features towards distributions with diverse semantics. With these enhancements, ExCEL not only retains CLIP's training-free advantages but also significantly outperforms other state-of-the-art methods with much less training cost on PASCAL VOC and MS COCO. Code is available at https://github.com/zwyang6/ExCEL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 637, + 135, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 135, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 135, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 658, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 696 + ], + "type": "text", + "content": "Weakly Supervised Semantic Segmentation (WSSS) intends to generate pixel-level predictions using weak annotations like points [2], scribbles [18, 32], bounding boxes [8," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 274, + 555, + 335 + ], + "blocks": [ + { + "bbox": [ + 317, + 274, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 317, + 274, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 555, + 335 + ], + "type": "image", + "image_path": "c7b90d3683d00401b2759d0fc96acc05aa2121d9bcfb5f1ccee7270597c57c67.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 346, + 555, + 425 + ], + "blocks": [ + { + "bbox": [ + 317, + 346, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 317, + 346, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 346, + 555, + 425 + ], + "type": "image", + "image_path": "6d8ebe21c87dcf4ca0fe5bf33f1f13a18e5088fb87af9751159439d68c40fa7a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 433, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 489 + ], + "type": "text", + "content": "Figure 1. Our motivation. (a) Previous methods leverage CLIP to generate CAMs with global image-text alignment, leaving CLIP's dense knowledge unexplored. (b) The proposed ExCEL explores CLIP's dense knowledge via a novel patch-text alignment paradigm, which generates better CAMs with less training cost." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "content": "16], or image-level labels [1, 23, 40]. It significantly reduces the annotating cost of fully supervised methods and has attracted increasing attention in the community. Among all cheap annotation types, most WSSS approaches leverage image-level labels to provide dense localization cues, linking visual concepts to specific pixel regions [5, 33]. In this work, we focus on WSSS with image-level labels as well." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Commonly, the WSSS pipeline involves three stages: generating Class Activation Maps (CAMs) [46] by training a classification network, refining CAMs into pseudo labels [28], and using these labels to train a segmentation model [7]. However, due to the minimal semantic information from image-level labels, CAMs intend to highlight the most distinctive object parts, significantly limiting WSSS performance. Recently, Contrastive Language-Image Pre-training (CLIP) [24] has been introduced in WSSS. CLIMs [36] applies image-text pairs to regularize visual relations among different semantics. CLIP-ES [20]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 288, + 714 + ], + "type": "text", + "content": "*Corresponding authors. Email:{shuowang, zjsong}@fudan.edu.cn." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20223" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "leverages image-text alignment for gradient and produces high-quality GradCAM [30]. WeCLIP [43] further streamlines this process by using CLIP's visual encoder for segmentation. Despite these advancements, current methods primarily focus on CLIP's global image-text alignment, as shown in Fig. 1 (a). CLIP's dense knowledge with patch-text alignment still remains under-explored in WSSS." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 294, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 294, + 314 + ], + "type": "text", + "content": "In this work, we propose ExCEL to explore CLIP's dense knowledge via a patch-text alignment paradigm for WSSS, i.e., generating CAMs by calculating patch-wise similarity between text and individual patch tokens, as shown in Fig. 1 (b). We identify two key challenges: (1) Semantic sparsity in textual prompts, where the template 'a photo of [CLASS]' only indicates object presence but lacks knowledge for localization, and (2) Fine-grained insufficiency in visual features, as CLIP prioritizes global representation due to its image-text pairing nature. To address these, ExCEL enhances CLIP's dense alignment with Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules, unlocking its potential across text and vision modalities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 532 + ], + "type": "text", + "content": "To generate semantically rich text representations, we propose TSE through an implicit attribute feature space. Instead of relying on explicit text templates, TSE module implicitly constructs text embeddings using universal attributes across the dataset. We first employ Large Language Models (LLMs) to generate detailed descriptions for each class, which are then processed by CLIP's text encoder to build a dataset-wide knowledge base. Rather than directly fusing these class-specific descriptions for text prompting, we focus on clustering the descriptive embeddings into generalized attributes, which effectively capture complementary knowledge from other classes and supplement missing information for the target class. With this implicit feature space, we enhance the text embeddings by hunting for its most relevant attribute features and aggregate them into the final class-specific text representation. This approach enables TSE to generate more informative text embeddings, providing a strong foundation for visual recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "To mine fine-grained knowledge from visual features, we propose VC to calibrate CLIP in both non-trainable and efficient-learnable ways. Our findings suggest that CLIP's q-k attention loses fine-grained details. Therefore, we first propose a Static Visual Calibration (SVC) module to replace the suboptimal q-k attention with a straightforward Intra-correlation operation. It focuses on extracting fine-grained details from intermediate layers, which progressively propagates fine-grained visual knowledge. Without any retraining, SVC generates CAMs comparable to training-required WSSS methods. Building on this, we further propose a Learnable Visual Calibration (LVC) module to dynamically calibrate CLIP's frozen features. LVC extracts spatial correlations from SVC's static CAMs. These correlations further supervise a lightweight adapter to learn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "the dynamic shift, pushing frozen features towards spatial-aware distributions. LVC and SVC complement each other, enabling precise patch-text alignment for CAM generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 328, + 109, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 109, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 328, + 109, + 553, + 120 + ], + "type": "text", + "content": "The main contributions of our work are listed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 361 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 181 + ], + "type": "text", + "content": "- We explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. The proposed ExCEL generates better pseudo labels in both training-free and efficient learning manners, revealing the dense capabilities of CLIP for efficient CAM generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 182, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 300 + ], + "type": "text", + "content": "- To enhance patch-text alignment, we propose the Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules. TSE applies LLMs to build a dataset-wide knowledge base and treats text prompting as an implicit attribute-hunting process, making text embeddings more informative. VC propagates the fine-grained visual knowledge in a non-parametric manner and further dynamically calibrates the frozen features with a lightweight adapter. TSE and VC work across two modalities, generating better dense alignment and pseudo labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 300, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 553, + 361 + ], + "type": "text", + "content": "- Extensive experiments on PASCAL VOC and MS COCO demonstrate that ExCEL significantly outperforms recent state-of-the-art methods, while reducing training cost with only 3.2 GB of GPU memory and " + }, + { + "bbox": [ + 314, + 300, + 553, + 361 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 314, + 300, + 553, + 361 + ], + "type": "text", + "content": " of the training time required by recent methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 373, + 406, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 406, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 406, + 385 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 393, + 539, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 539, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 539, + 406 + ], + "type": "text", + "content": "2.1. Weakly Supervised Semantic Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 604 + ], + "type": "text", + "content": "Weakly supervised semantic segmentation with image-level labels typically relies on CAMs to provide dense supervision for segmentation [42, 44]. However, CAMs usually highlight the most discriminative parts of objects [34]. To address this issue, considerable efforts have been made from many intriguing insights. MCTformer [37] incorporates multiple class tokens in Vision Transformer and proposes generating CAMs from class-patch attention. ToCo [29] proposes token contrast learning and generates more precise CAMs. SeCo [39] designs a separate and conquer scheme and succeeds in tackling co-occurrence. Despite these advancements, prior methods commonly require retraining the entire classification network for CAM generation. In this work, our ExCEL directly generates CAMs from frozen CLIP, and further boosts its quality via a lightweight adapter, significantly reducing the training cost." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 477, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 477, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 477, + 636 + ], + "type": "text", + "content": "2.2. Vision-Language Pre-training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "Contrastive Language Image Pre-training (CLIP) [24], known for pretraining on billion-scale image-text pairs, has demonstrated remarkable transferability in many downstream tasks. CoOP [48] and CLIP-Adapter [12] incorporate lightweight trainable parameters into CLIP and succeed in few-shot classification. DenseCLIP [25] and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20224" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 549, + 266 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 549, + 266 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 549, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 549, + 266 + ], + "type": "image", + "image_path": "2e36c53b2f5353223a5823b875eba234946566a301e81563cf57415f6f9f75e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": "Figure 2. ExCEL Architecture. We explore CLIP's dense knowledge with Text Semantic Enrichment (TSE) and Visual Calibration (VC). (a) TSE uses LLMs to build a knowledge base and clusters it into an implicit attribute space. The final text representation " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": " is enhanced by hunting for relevant attributes. For vision modality, (b) we introduce Static Visual Calibration (SVC) to calibrate visual features using the Inter-correlation operation across " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": " intermediate layers. It generates static CAMs with " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": " and calibrated features " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": ". (c) Learnable Visual Calibration (LVC) designs a learnable adapter to add a dynamic shift " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": " to SVC. It generates optimized features " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": " based on static CAMs guidance, creating dynamic CAMs from " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 275, + 555, + 342 + ], + "type": "text", + "content": ". Dynamic CAMs are refined for segmentation supervision. Details are in Sec. 3.1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 352, + 297, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 297, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 297, + 509 + ], + "type": "text", + "content": "MaskCLIP [47] leverage the alignment between text and vision modalities for the dense segmentation task. Recently, some studies have introduced CLIP into WSSS. CLIMS [36] treats image-text pairing of CLIP as regularization and leverages it to regularize the visual concepts. CLIP-ES [20] finds that the image-text alignment of CLIP generates class gradient and leverages it for GradCAM generation. WeCLIP [43] further streamlines this process and directly leverages CLIP's visual encoder for segmentation. However, these methods mainly focus on a global image-text alignment while ignoring the dense capabilities of CLIP. In contrast, our ExCEL explores CLIP's dense knowledge via a patch-text alignment paradigm for WSSS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 138, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 138, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 138, + 532 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 537, + 141, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 141, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 141, + 549 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": "Patch-text CAM Generation. CLIP uses image and text encoders to project image and text into the same feature space, enabling robust vision-language alignment. In this work, we utilize this property to generate CAM in a patch-text alignment paradigm. Given the encoded text embeddings " + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{D \\times C}" + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": " and visual features " + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{h \\times w \\times D}" + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": " is the feature channel, " + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": " are the spatial sizes, " + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": " is the number of classes, we generate CAM by calculating the patch-wise similarities between text and visual features:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 670, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 670, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 295, + 683 + ], + "type": "interline_equation", + "content": "\\mathrm {C A M} = \\operatorname {N o r m} (\\cos (P, T), \\tag {1}", + "image_path": "8f45ae2291e167c2a9e1b2adc9a067991f9ad8fdfea51be2e258a23d0af1366c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Norm}(\\cdot)" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is the min-max normalization and " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is the cosine similarity calculation. However, due to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 352, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 400 + ], + "type": "text", + "content": "the image-level pairing nature, CLIP suffers from textual semantic sparsity and visual fine-grained insufficiency. Therefore, our ExCEL incorporates TSE and VC (SVC and LVC) into CLIP to further explore its dense potential." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 400, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 436 + ], + "type": "text", + "content": "Framework Overview. Our ExCEL generates CAMs in both training-free and efficient learning manners, as shown in Fig. 2. Its training pipeline is generalized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 653 + ], + "type": "text", + "content": "(1) Enriching textual semantics via TSE. We first use GPT-4 to generate descriptions for each class, which are encoded into a dataset-wide knowledge base with CLIP's text encoder. We cluster this knowledge into class-agnostic attributes and use the global text prompt to hunt for its most relevant ones. They are then aggregated into the final text representation. (2) Static CAM generation via SVC. We replace CLIP's q-k self-attention with our Intra-correlation operation from intermediate layers. Then the calibrated visual features and enhanced text embeddings are used for static CAMs via Eq. (1). (3) Dynamic CAM generation via LVC. A lightweight adapter is designed to learn dynamic token relations from static CAMs. The relations are added to SVC and serve as a distribution shift to make the visual features more diverse. The dynamic CAMs are generated with the enhanced text embeddings and LVC features via Eq. (1). (4) Segmentation training. Dynamic CAMs are refined to pseudo labels for segmentation supervision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 659, + 463, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 463, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 463, + 672 + ], + "type": "text", + "content": "3.2. Text Semantic Enrichment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Knowledge Base Construction. The global text template 'a clean origami of [CLASS]' only indicates the presence of objects while limited providing dense knowledge for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20225" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": "patch-wise visual recognition. To enrich the text representation, we first adapt LLMs, such as GPT-4 to generate detailed class descriptions, as shown in Fig. 2 (a). Specifically, given the global template " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " from class label space " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "Y \\in \\{1,2,\\dots,C\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " is the class index and " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " is the number of categories, we carefully construct instructions for GPT: \"List " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " descriptions with key properties to describe the [CLASS] in terms of appearance, color, shape, size, or material, etc. These descriptions will help visually distinguish the [CLASS] from other classes in the dataset. Each description should follow the format: 'a clean origami [CLASS]. it + descriptive contexts.' With this instruction, we ask GPT to generate " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " detailed descriptions for each class, which are subsequently encoded into a dataset-wide knowledge base with CLIP's text encoder. The knowledge base is denoted as " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\Phi(e_i)\\}_{i=1}^{n \\times C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " is the description from GPT, " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " is CLIP's text encoder and " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "n \\times C" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " is the number of all descriptions. This knowledge base gathers descriptive properties for the whole dataset, building a strong foundation for the textual category representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "type": "text", + "content": "Implicit Attribute Hunting. Instead of explicitly merging class-related knowledge into a single text embedding, we cluster this knowledge into generalized attributes and treat text prompting as an implicit attribute-hunting process. This has two main benefits: (1) " + }, + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "type": "text", + "content": " explicit descriptions may still be limited in covering all characteristics of the class. The clustered attributes efficiently capture shared contextual knowledge from other categories, supplementing missing information for target class recognition. (2) The generated descriptions inevitably contain redundant or noisy content. The use of attributes makes the knowledge more compact and representative, leading to precise text prompting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "content": "To this end, we leverage a clustering algorithm to generate multiple centroids based on the knowledge base. Each cluster centroid is viewed as the implicit attribute that represents a group of descriptions sharing similar properties:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 510, + 294, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 510, + 294, + 526 + ], + "spans": [ + { + "bbox": [ + 107, + 510, + 294, + 526 + ], + "type": "interline_equation", + "content": "A = \\operatorname {K m e a n s} (\\mathcal {T}, B) = \\left\\{a _ {i} \\right\\} _ {i = 1} ^ {B}, \\tag {2}", + "image_path": "744b5e615c40950ccfcb1755e478f736743249701fe452618d80f2ef497465de.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "content": " is the number of centroids and kmeans algorithm [21] is used for clustering for simplicity. " + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "inline_equation", + "content": "a_{i}\\in \\mathbb{R}^{D\\times 1}" + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "content": " represents the cluster centroid, i.e., the implicit attribute." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "content": "With the attribute feature space " + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "content": ", we first send the global template " + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "content": " into the text encoder of CLIP to generate global text embedding " + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "inline_equation", + "content": "t_c \\in \\mathbb{R}^{D \\times 1}" + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "content": " is the channel dimension. Then " + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "content": " is leveraged to search for its most relevant attributes in the implicit attribute space. To exclude irrelevant attributes, we further propose to select TOPK attribute neighbors based on the similarity scores with " + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 55, + 571, + 296, + 654 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 662, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 662, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 88, + 662, + 295, + 682 + ], + "type": "interline_equation", + "content": "A _ {c} = \\left\\{a _ {j}: j \\in \\operatorname {a r g m a x} _ {\\mathrm {T O P K}} \\left\\{t _ {c} ^ {T} a _ {j} \\right\\} _ {j = 1} ^ {B} \\right\\}. \\tag {3}", + "image_path": "99612a1ae367c46820a9b0882dbbb024903bdbf3be389f3ed4b5017fe7b1e585.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Finally, we gently aggregate the implicit attribute neighbors according to the corresponding similarity weights to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " and take the aggregated features as the complementary knowledge for textual semantic enrichment. The final text representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "T_{c}\\in \\mathbb{R}^{D\\times 1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " is denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 357, + 114, + 555, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 114, + 555, + 147 + ], + "spans": [ + { + "bbox": [ + 357, + 114, + 555, + 147 + ], + "type": "interline_equation", + "content": "T _ {c} = t _ {c} + \\lambda \\sum_ {j = 1} ^ {K} \\operatorname {s o f t m a x} \\left(t _ {c} ^ {T} A _ {c}\\right) a _ {j}, \\tag {4}", + "image_path": "e88e49785e98ce2551cb5c7abc2b94ca5aea3f60534476824154fe682de7a201.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 152, + 545, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 545, + 164 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 152, + 545, + 164 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 152, + 545, + 164 + ], + "type": "text", + "content": " is the factor to balance the attribute information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 170, + 428, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 428, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 428, + 182 + ], + "type": "text", + "content": "3.3. Visual Calibrations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": "Static Visual Calibration. Due to the image-text pairing nature of CLIP, the visual features lack fine-grained information, leading to unreasonable localization maps via patch-text alignment. To delve into this, let us review the self-attention mechanism of the original CLIP first. As shown in Fig. 2 (b), given the input image " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{3 \\times \\mathcal{H} \\times \\mathcal{W}}" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " we send it to the image encoder of CLIP, which includes 12 Attention layers in this work. " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\times \\mathcal{W}" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " is the image size. For the feature " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "F_{l} \\in \\mathbb{R}^{D_{s} \\times hw}" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": "-th layer of CLIP, it is first projected into three different spaces " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\{q, k, v\\}" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": ", named query, key, and value, respectively. " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "q, k" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " have the same shape of " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "D_{s} \\times hw" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "hw" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " represents the channel dimension and sequence length. Then the attention map between " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 188, + 555, + 356 + ], + "type": "text", + "content": " is calculated by measuring the similarity:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 361, + 555, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 361, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 361, + 361, + 555, + 380 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} (q, k) = \\operatorname {s o f m a x} \\left(q ^ {T} k / \\sqrt {D _ {s}}\\right), \\tag {5}", + "image_path": "f8495dd7746d46545e19c09d7364f6544bc7f5061fb42f0832ff140196b4781f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{SA}(\\cdot)" + }, + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "text", + "content": " is the calculation of attention. Then the output features " + }, + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "inline_equation", + "content": "F_{l + 1}" + }, + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "text", + "content": " are generated by aggregating the tokens of " + }, + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 385, + 554, + 422 + ], + "type": "text", + "content": " according to similarity weights in the attention map." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "text", + "content": "However, due to the inherent image-text alignment of CLIP, the original q-k attention produces overly uniform attention maps, homogenizing diverse tokens from " + }, + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "text", + "content": " to capture broad semantics for global image representation (see discussions in Sec. 4.4). It leads to inaccurate object recognition. MaskCLIP [47] holds a similar observation, supporting this claim by removing the final q-k attention layer and using " + }, + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "text", + "content": " from the last layer as the visual output to preserve diversity. In our work, we choose to replace the suboptimal q-k attention with a straightforward Intra-correlation operation and focus on extracting fine-grained details from intermediate layers. This non-parametric approach effectively mines spatial semantics in intermediate " + }, + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\{q, k, v\\}" + }, + { + "bbox": [ + 313, + 422, + 554, + 601 + ], + "type": "text", + "content": " and avoids the smoothing effect of q-k attention, resulting in more consistent attention maps and improved object localization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "text", + "content": "Specifically, instead of generating q-k correlation, Intra-correlation calculates the attention within each space of " + }, + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\{q,k,v\\}" + }, + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "text", + "content": " across intermediate layers. The attention map " + }, + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "inline_equation", + "content": "S_{attn}^{l}\\in \\mathbb{R}^{hw\\times hw}" + }, + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 601, + 555, + 651 + ], + "type": "text", + "content": " -th SVC layer is generated by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 655, + 555, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 655, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 330, + 655, + 555, + 673 + ], + "type": "interline_equation", + "content": "S _ {a t t n} ^ {l} = \\sum w _ {i} \\operatorname {S A} \\left(O _ {i} ^ {l}, O _ {i} ^ {l}\\right), O _ {i} ^ {l} \\in \\left\\{q ^ {l}, k ^ {l}, v ^ {l} \\right\\}, \\tag {6}", + "image_path": "62b7ed2cbc38ce027f9baf98d74d3f86eccf30abef7889f451515f8323ec8087.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " is the contribution weight for different correlation maps. " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "l\\in \\{12 - N,\\dots,12\\}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " is the number of intermediate layers for this operation. Then " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "S_{attn}^{l}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "v^{l}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " are used to" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20226" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "generate the output features. Finally, the calibrated features " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "P_{s} \\in \\mathbb{R}^{D \\times h \\times w}" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " from the last layer is used to generate static CAM " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "CAM_{s}" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " with text embedding " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " via Eq. (1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 108, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 108, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 108, + 295, + 215 + ], + "type": "text", + "content": "Learnable Visual Calibration. Although ExCEL generates comparable CAMs without training, its performance is still limited by the fixed features in CLIP. To further unleash the dense potential of CLIP, we design a lightweight adapter to dynamically calibrate the visual features with diverse details. This adapter only incorporates a distribution shift to calibrate the fixed features without changing CLIP's pre-trained weights, thereby retaining CLIP's transferability and enhancing its dense performance for WSSS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 251 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 2 (c), frozen features " + }, + { + "bbox": [ + 55, + 216, + 296, + 251 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 55, + 216, + 296, + 251 + ], + "type": "text", + "content": " from 1-12th layer of CLIP are extracted to learn a dynamic feature " + }, + { + "bbox": [ + 55, + 216, + 296, + 251 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 55, + 216, + 296, + 251 + ], + "type": "text", + "content": " via the adapter. The process is expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 258, + 295, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 258, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 258, + 295, + 275 + ], + "type": "interline_equation", + "content": "F _ {d} = \\operatorname {C o n v} (\\operatorname {C o n c a t e} [ \\delta_ {l} (F _ {l}) ] _ {l = 1} ^ {1 2}), \\tag {7}", + "image_path": "5c826cdbcb901c02b5c182fcd025fea7cf81f2a5f66f40d45067de5eb9bac9a7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "F_{d}\\in \\mathbb{R}^{D_{d}\\times hw}" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "D_{d}" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": " is the channel dimension. " + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(\\cdot)" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": " is the convolution layer, " + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}[\\cdot ]" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": " is the concatenate operation that connects all features along channel dimension, and " + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\delta_l(\\cdot)" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": " is the individual MLP layer for each " + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": " Then " + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": " is used to generate dynamic token relations by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 349, + 295, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 349, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 97, + 349, + 295, + 364 + ], + "type": "interline_equation", + "content": "r = \\alpha (\\cos (F _ {d}, F _ {d}) - \\beta \\overline {{\\cos (F _ {d} , F _ {d})}}), \\tag {8}", + "image_path": "4486af896c5a661782434a5d1134a2338ddb66dce0faf5bcdeba390faca5e28e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^{hw \\times hw}" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": " are the scaling and shifting factors to adjust the relations, respectively. " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\overline{\\cos(F_d, F_d)}" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": " means the mean value of similarity scores of " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "F_d" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": ". It is designed to remove the irrelevant relations in low values by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 427, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 427, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 115, + 427, + 295, + 459 + ], + "type": "interline_equation", + "content": "R _ {i j} = \\left\\{ \\begin{array}{l l} r _ {i j}, & \\text {i f} r _ {i j} \\geq 0 \\\\ - i n f, & \\text {e l s e} \\end{array} . \\right. \\tag {9}", + "image_path": "36d43c9cc74a523665cf17e687e296162407f1f46f703c7792b23d880ebf1cad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": "With the dynamic relation " + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{hw \\times hw}" + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": ", we add it as a distribution bias to the static attention map " + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "inline_equation", + "content": "S_{attn}^{l}" + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": ", dynamically grouping the frozen tokens within related semantics and shifting the features towards denser distribution. The optimized attention map " + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "inline_equation", + "content": "L_{attn}^{l} \\in \\mathbb{R}^{hw \\times hw}" + }, + { + "bbox": [ + 55, + 467, + 296, + 528 + ], + "type": "text", + "content": " is denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 112, + 534, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 534, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 112, + 534, + 295, + 550 + ], + "type": "interline_equation", + "content": "L _ {a t t n} ^ {l} = S _ {a t t n} ^ {l} + \\operatorname {s o f t m a x} (R). \\tag {10}", + "image_path": "6fe283ac1ae8150e4dcb58b8af5b5fdb63dbe01a72c36694b99d8180f66476d0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": "Subsequently, we extract the dynamically calibrated features " + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "inline_equation", + "content": "P_{d} \\in \\mathbb{R}^{D \\times h \\times w}" + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": " from the last layer of LVC and generate dynamic CAMs with Eq. (1), which are then refined to final pseudo labels " + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "inline_equation", + "content": "M_{d}" + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": " for the segmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 612, + 171, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 171, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 171, + 625 + ], + "type": "text", + "content": "3.4. Training Objectives" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "We formulate a diversity loss to supervise the learning of " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in our LVC module. We first measure token correlations of " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " by calculating the self similarity: " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}} = \\mathrm{sigmoid}(\\cos (F_d,F_d))" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", where sigmoid(·) is the activation function and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}\\in \\mathbb{R}^{hw\\times hw}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". Then we refine " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "CAM_{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " from SVC into static pseudo labels " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and leverage its pixelwise affinity to guide the diversifying of " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". Specifically, if" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "the pixel with coordinate " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " shows the same pixel value as the pixel in " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "(\\varepsilon,\\eta)" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": ", the token pair with the same coordinates on " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " is semantically related and its corresponding correlation logit on " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " should be maximized, and vice versa. The diversity loss can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 140, + 553, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 324, + 140, + 553, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i v}} = \\frac {1}{N ^ {+}} \\sum_ {u ^ {+} \\in \\hat {\\mathcal {R}} ^ {+}} (1 - u ^ {+}) + \\frac {1}{N ^ {-}} \\sum_ {u ^ {-} \\in \\hat {\\mathcal {R}} ^ {-}} u ^ {-}, \\tag {11}", + "image_path": "e21a99797f429330f04e1a4a8e24e384a5c458577d2bd5ea6ab5ff577bdcb737.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "inline_equation", + "content": "N^{+} / N^{-}" + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "content": " are the number of positive and negative pairs, " + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "inline_equation", + "content": "u^{+} / u^{-}" + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "content": " are the positive and negative relation logits, and " + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}^{+} / \\hat{\\mathcal{R}}^{-}" + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "content": " are the positive and negative sets of logit on " + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}" + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "content": ", respectively. The diversity loss groups tokens with similar semantics and suppresses irrelevant ones, enhancing fine-grained details of visual features for precise text response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "type": "text", + "content": "In addition, ExCEL is streamlined as a single-stage method. We adopt a lightweight Transformer-based segmentation head [43] and directly take the frozen visual encoder of CLIP for segmentation. The dynamic pseudo labels are used as supervision with a cross-entropy loss " + }, + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "type": "text", + "content": ". The loss objectives of our ExCEL are formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 333, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 333, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 383, + 333, + 553, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {E X C E L}} = \\mathcal {L} _ {s e g} + \\gamma \\mathcal {L} _ {\\text {d i v}}, \\tag {12}", + "image_path": "00c184f5e2aaddf4a92a9f54b358ee04b9acdd1e7b1d5716cb310499deb1d98a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 354, + 553, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 391 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 354, + 553, + 391 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 354, + 553, + 391 + ], + "type": "text", + "content": " is the weight factor. By efficiently training the adapter and a segmentation head, ExCEL achieves strong WSSS performance and significantly reduces training cost." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 401, + 458, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 458, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 458, + 415 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 421, + 441, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 441, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 441, + 434 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "content": "Datasets and Metrics. The proposed ExCEL is evaluated on PASCAL VOC 2012 [11] and MS COCO 2014 [19]. VOC contains 21 categories (1 for background). Following prior methods [10, 17, 39], the augmented dataset with 10,582, 1,449, and 1,456 images are used for training, validating, and testing, respectively. COCO includes 81 classes, in which 82,081 images are used for training and 40,137 images are for validating. Mean Intersection-Over-Union (mIoU) is used as the main evaluation metric." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Implementation Details. CLIP model with ViT-B [9] is used as ExCEL's encoder, which is frozen during the training. For TSE module, we generate " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "n = 20" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " descriptions from GPT-4 for each category. The number of attribute embeddings " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " is set to 112 and 224 for PASCAL VOC and MS COCO, respectively. The SVC module is conducted in the last " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " Transformer layers. Our decoder adopts a simple Transformer-based head [43]. Features " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " from each layer of CLIP are sent to it for the segmentation predictions. The scaling and shifting factors in Eq. (8) are set as 3.0 and 1.0, respectively. The loss weight " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " is set as 0.1. Following [29, 35, 39], the AdamW optimizer is used for training the adapter and decoder. The learning rate is 1e-4 with a weight decay of 1e-2. The training iteration is set" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20227" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 80, + 111, + 114 + ], + "blocks": [ + { + "bbox": [ + 59, + 80, + 111, + 114 + ], + "lines": [ + { + "bbox": [ + 59, + 80, + 111, + 114 + ], + "spans": [ + { + "bbox": [ + 59, + 80, + 111, + 114 + ], + "type": "image", + "image_path": "c5301420fc8865447c8c1f658f64629bb88710e6c0b272f98be0b0cf98101ae9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 112, + 80, + 164, + 114 + ], + "blocks": [ + { + "bbox": [ + 112, + 80, + 164, + 114 + ], + "lines": [ + { + "bbox": [ + 112, + 80, + 164, + 114 + ], + "spans": [ + { + "bbox": [ + 112, + 80, + 164, + 114 + ], + "type": "image", + "image_path": "3cf356d6bcb85be39af554e85aa9ae60da491f5ae8199b38fed976f53559b185.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 80, + 218, + 114 + ], + "blocks": [ + { + "bbox": [ + 170, + 71, + 216, + 79 + ], + "lines": [ + { + "bbox": [ + 170, + 71, + 216, + 79 + ], + "spans": [ + { + "bbox": [ + 170, + 71, + 216, + 79 + ], + "type": "text", + "content": "PASCAL VOC" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 80, + 218, + 114 + ], + "lines": [ + { + "bbox": [ + 167, + 80, + 218, + 114 + ], + "spans": [ + { + "bbox": [ + 167, + 80, + 218, + 114 + ], + "type": "image", + "image_path": "8341b6c0a8a3382c0fb8af763250a9d559aec1198a864dccbdcd7b63a33a9513.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 80, + 272, + 114 + ], + "blocks": [ + { + "bbox": [ + 221, + 80, + 272, + 114 + ], + "lines": [ + { + "bbox": [ + 221, + 80, + 272, + 114 + ], + "spans": [ + { + "bbox": [ + 221, + 80, + 272, + 114 + ], + "type": "image", + "image_path": "22f512584c0be30c1ffaf594f4585075fc84962a1609bf59e9051c09a36a6978.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 80, + 327, + 114 + ], + "blocks": [ + { + "bbox": [ + 276, + 80, + 327, + 114 + ], + "lines": [ + { + "bbox": [ + 276, + 80, + 327, + 114 + ], + "spans": [ + { + "bbox": [ + 276, + 80, + 327, + 114 + ], + "type": "image", + "image_path": "851349b2259265cdcc9d4f7cb4c9d72faa4887a1d9dbcaa8204e7a4789403436.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 116, + 110, + 150 + ], + "blocks": [ + { + "bbox": [ + 59, + 116, + 110, + 150 + ], + "lines": [ + { + "bbox": [ + 59, + 116, + 110, + 150 + ], + "spans": [ + { + "bbox": [ + 59, + 116, + 110, + 150 + ], + "type": "image", + "image_path": "316efb51d73ee1432a28e419b77b6f213e78dc03a679cd6105c8eeffdb57b608.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 112, + 116, + 164, + 150 + ], + "blocks": [ + { + "bbox": [ + 112, + 116, + 164, + 150 + ], + "lines": [ + { + "bbox": [ + 112, + 116, + 164, + 150 + ], + "spans": [ + { + "bbox": [ + 112, + 116, + 164, + 150 + ], + "type": "image", + "image_path": "f74f7342c9bd38e85ccccf988a630ed6df648ccd4e7c37874fc3e2ed420d5b86.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 167, + 116, + 218, + 150 + ], + "blocks": [ + { + "bbox": [ + 167, + 116, + 218, + 150 + ], + "lines": [ + { + "bbox": [ + 167, + 116, + 218, + 150 + ], + "spans": [ + { + "bbox": [ + 167, + 116, + 218, + 150 + ], + "type": "image", + "image_path": "46ba46418105780174965328e844f1039f7cf6ba6bb845e56f6e9bdef5d4e93c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 221, + 116, + 272, + 150 + ], + "blocks": [ + { + "bbox": [ + 221, + 116, + 272, + 150 + ], + "lines": [ + { + "bbox": [ + 221, + 116, + 272, + 150 + ], + "spans": [ + { + "bbox": [ + 221, + 116, + 272, + 150 + ], + "type": "image", + "image_path": "e2bb1f5eea6c8a12e3505f943f800f9431b6414b16d8c31bb778c1dccb2be748.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 276, + 116, + 327, + 150 + ], + "blocks": [ + { + "bbox": [ + 276, + 116, + 327, + 150 + ], + "lines": [ + { + "bbox": [ + 276, + 116, + 327, + 150 + ], + "spans": [ + { + "bbox": [ + 276, + 116, + 327, + 150 + ], + "type": "image", + "image_path": "a5114db16ca06ec6772bada96533feaa21babdba9f61e9360cb97d41fa12b449.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 59, + 152, + 110, + 185 + ], + "blocks": [ + { + "bbox": [ + 59, + 152, + 110, + 185 + ], + "lines": [ + { + "bbox": [ + 59, + 152, + 110, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 152, + 110, + 185 + ], + "type": "image", + "image_path": "0820de22073960a29f5d52de03d08656c62463f095e9caf7cb7da0c246c620d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 112, + 152, + 164, + 185 + ], + "blocks": [ + { + "bbox": [ + 112, + 152, + 164, + 185 + ], + "lines": [ + { + "bbox": [ + 112, + 152, + 164, + 185 + ], + "spans": [ + { + "bbox": [ + 112, + 152, + 164, + 185 + ], + "type": "image", + "image_path": "388bc4a2bd3c0cf2444b90089f1a7164064a2a0ed4260df95d41bf99e22ac2ca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 167, + 152, + 218, + 185 + ], + "blocks": [ + { + "bbox": [ + 167, + 152, + 218, + 185 + ], + "lines": [ + { + "bbox": [ + 167, + 152, + 218, + 185 + ], + "spans": [ + { + "bbox": [ + 167, + 152, + 218, + 185 + ], + "type": "image", + "image_path": "bd7a6d018eedd9f55a445e37adfe8ba060d1495a084df2b77b4e9b12e5fa72f0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 221, + 152, + 272, + 185 + ], + "blocks": [ + { + "bbox": [ + 221, + 152, + 272, + 185 + ], + "lines": [ + { + "bbox": [ + 221, + 152, + 272, + 185 + ], + "spans": [ + { + "bbox": [ + 221, + 152, + 272, + 185 + ], + "type": "image", + "image_path": "3599e883cd521f0c40f4618ba859159582be2d9bb94e3ce37c7ed5c27da298ac.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 276, + 152, + 327, + 185 + ], + "blocks": [ + { + "bbox": [ + 276, + 152, + 327, + 185 + ], + "lines": [ + { + "bbox": [ + 276, + 152, + 327, + 185 + ], + "spans": [ + { + "bbox": [ + 276, + 152, + 327, + 185 + ], + "type": "image", + "image_path": "2a099ccafe16b9d68dd8d4398051f24a7d45a477a1015a2f1c8922a759d0bec7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 186, + 110, + 221 + ], + "blocks": [ + { + "bbox": [ + 59, + 186, + 110, + 221 + ], + "lines": [ + { + "bbox": [ + 59, + 186, + 110, + 221 + ], + "spans": [ + { + "bbox": [ + 59, + 186, + 110, + 221 + ], + "type": "image", + "image_path": "853857e3c93474f209433edaa13ebad39e31788efb3e014e2004141d2bb751af.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 553, + 240 + ], + "type": "text", + "content": "Figure 3. Segmentation visualizations of SeCo [39], WeCLIP [43] and ours on VOC and COCO. ExCEL segments objects more precisely." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 112, + 186, + 164, + 221 + ], + "blocks": [ + { + "bbox": [ + 112, + 186, + 164, + 221 + ], + "lines": [ + { + "bbox": [ + 112, + 186, + 164, + 221 + ], + "spans": [ + { + "bbox": [ + 112, + 186, + 164, + 221 + ], + "type": "image", + "image_path": "9c64ba882624693b3d4e3de98f783de003f855a3c4275e4711672fb1fe2215f7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 167, + 186, + 218, + 221 + ], + "blocks": [ + { + "bbox": [ + 167, + 186, + 218, + 221 + ], + "lines": [ + { + "bbox": [ + 167, + 186, + 218, + 221 + ], + "spans": [ + { + "bbox": [ + 167, + 186, + 218, + 221 + ], + "type": "image", + "image_path": "a0c6051f088c68437c8d0d4e19df66cf6194ca4704b147958ca5c96eae17bf64.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 221, + 186, + 272, + 221 + ], + "blocks": [ + { + "bbox": [ + 221, + 186, + 272, + 221 + ], + "lines": [ + { + "bbox": [ + 221, + 186, + 272, + 221 + ], + "spans": [ + { + "bbox": [ + 221, + 186, + 272, + 221 + ], + "type": "image", + "image_path": "407257d2cc35efd54090f414c835ffd10fd347c1c94e2696287b9da91c60fe8e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 276, + 186, + 327, + 221 + ], + "blocks": [ + { + "bbox": [ + 276, + 186, + 327, + 221 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 327, + 221 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 327, + 221 + ], + "type": "image", + "image_path": "4cb732dedb5053b10e25b17de0312daee744fe05fc297cd9680fa2b5c269fdff.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 337, + 71, + 388, + 114 + ], + "blocks": [ + { + "bbox": [ + 337, + 71, + 388, + 114 + ], + "lines": [ + { + "bbox": [ + 337, + 71, + 388, + 114 + ], + "spans": [ + { + "bbox": [ + 337, + 71, + 388, + 114 + ], + "type": "image", + "image_path": "5c8a2dde2dc5e2bb3062dc8d1fc52dc7e11c6f4da1b8a2a94e557f48afb7cf0a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 443, + 114 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 443, + 114 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 443, + 114 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 443, + 114 + ], + "type": "image", + "image_path": "615e9b4bfb1296f453373dbb8da005a45ce6215b90b367eb64649975ad83987e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 446, + 80, + 498, + 114 + ], + "blocks": [ + { + "bbox": [ + 446, + 80, + 498, + 114 + ], + "lines": [ + { + "bbox": [ + 446, + 80, + 498, + 114 + ], + "spans": [ + { + "bbox": [ + 446, + 80, + 498, + 114 + ], + "type": "image", + "image_path": "c18013dc17f10be6b7fc5d01c0817a31a5a6306335896bafe651c90f39a8bd69.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 500, + 80, + 551, + 114 + ], + "blocks": [ + { + "bbox": [ + 500, + 80, + 551, + 114 + ], + "lines": [ + { + "bbox": [ + 500, + 80, + 551, + 114 + ], + "spans": [ + { + "bbox": [ + 500, + 80, + 551, + 114 + ], + "type": "image", + "image_path": "9a592c946fab0c5e254996fbb1027530458ab6932689dc33850d902fafe20365.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 337, + 116, + 389, + 150 + ], + "blocks": [ + { + "bbox": [ + 337, + 116, + 389, + 150 + ], + "lines": [ + { + "bbox": [ + 337, + 116, + 389, + 150 + ], + "spans": [ + { + "bbox": [ + 337, + 116, + 389, + 150 + ], + "type": "image", + "image_path": "d0c07d1dd506f1791c29c140bcee28232b00df3ebd8b1fbfe3a347a3d6a5507a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 392, + 116, + 443, + 150 + ], + "blocks": [ + { + "bbox": [ + 392, + 116, + 443, + 150 + ], + "lines": [ + { + "bbox": [ + 392, + 116, + 443, + 150 + ], + "spans": [ + { + "bbox": [ + 392, + 116, + 443, + 150 + ], + "type": "image", + "image_path": "1c826c87e4fd50bf112a593962731b2eaf9d2af5fb9b2a326ef96ea612e58e3b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 446, + 116, + 498, + 150 + ], + "blocks": [ + { + "bbox": [ + 446, + 116, + 498, + 150 + ], + "lines": [ + { + "bbox": [ + 446, + 116, + 498, + 150 + ], + "spans": [ + { + "bbox": [ + 446, + 116, + 498, + 150 + ], + "type": "image", + "image_path": "1abe07791c3f893850d03b10d98dc161b7e98d42d36f411662827af52d8210ae.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 500, + 116, + 551, + 150 + ], + "blocks": [ + { + "bbox": [ + 500, + 116, + 551, + 150 + ], + "lines": [ + { + "bbox": [ + 500, + 116, + 551, + 150 + ], + "spans": [ + { + "bbox": [ + 500, + 116, + 551, + 150 + ], + "type": "image", + "image_path": "fb5bb2657738c3517c6dada1ca94ff908721c365d9d43527dd77033d7b93b5b4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 337, + 152, + 389, + 185 + ], + "blocks": [ + { + "bbox": [ + 337, + 152, + 389, + 185 + ], + "lines": [ + { + "bbox": [ + 337, + 152, + 389, + 185 + ], + "spans": [ + { + "bbox": [ + 337, + 152, + 389, + 185 + ], + "type": "image", + "image_path": "4f66e0660334b6e3fd8f15cdd3314761408886b96fe8ff5b0cdfdfac188c1fe8.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 392, + 152, + 443, + 185 + ], + "blocks": [ + { + "bbox": [ + 392, + 152, + 443, + 185 + ], + "lines": [ + { + "bbox": [ + 392, + 152, + 443, + 185 + ], + "spans": [ + { + "bbox": [ + 392, + 152, + 443, + 185 + ], + "type": "image", + "image_path": "5a5931862b2a698f9ebdd3ba5c5b52658585c531040dbc909fa7994f6d456f47.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 446, + 152, + 498, + 185 + ], + "blocks": [ + { + "bbox": [ + 446, + 152, + 498, + 185 + ], + "lines": [ + { + "bbox": [ + 446, + 152, + 498, + 185 + ], + "spans": [ + { + "bbox": [ + 446, + 152, + 498, + 185 + ], + "type": "image", + "image_path": "f610f88abaee40af22c521424154630cbceb42d45376f19195aa278bc495fc71.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 500, + 152, + 551, + 185 + ], + "blocks": [ + { + "bbox": [ + 500, + 152, + 551, + 185 + ], + "lines": [ + { + "bbox": [ + 500, + 152, + 551, + 185 + ], + "spans": [ + { + "bbox": [ + 500, + 152, + 551, + 185 + ], + "type": "image", + "image_path": "fc82a5c0ee1e47840032587580b809eff34cfb26483364aad4df08cb97eaef3f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 337, + 186, + 389, + 221 + ], + "blocks": [ + { + "bbox": [ + 337, + 186, + 389, + 221 + ], + "lines": [ + { + "bbox": [ + 337, + 186, + 389, + 221 + ], + "spans": [ + { + "bbox": [ + 337, + 186, + 389, + 221 + ], + "type": "image", + "image_path": "4740cf6ccbaabd038a40da72d990c0341534e1a41b2dc24a286ebeac55b86433.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 392, + 186, + 443, + 221 + ], + "blocks": [ + { + "bbox": [ + 392, + 186, + 443, + 221 + ], + "lines": [ + { + "bbox": [ + 392, + 186, + 443, + 221 + ], + "spans": [ + { + "bbox": [ + 392, + 186, + 443, + 221 + ], + "type": "image", + "image_path": "1a70299b32eb9c9d4b781b775f287a4ed3775196df9a81909637a218f483f6eb.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 446, + 186, + 498, + 221 + ], + "blocks": [ + { + "bbox": [ + 446, + 186, + 498, + 221 + ], + "lines": [ + { + "bbox": [ + 446, + 186, + 498, + 221 + ], + "spans": [ + { + "bbox": [ + 446, + 186, + 498, + 221 + ], + "type": "image", + "image_path": "d6c7020eaf13999ff40c738a94b1b660907037ea78564be246790d95fe8eb66a.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 500, + 186, + 551, + 221 + ], + "blocks": [ + { + "bbox": [ + 500, + 186, + 551, + 221 + ], + "lines": [ + { + "bbox": [ + 500, + 186, + 551, + 221 + ], + "spans": [ + { + "bbox": [ + 500, + 186, + 551, + 221 + ], + "type": "image", + "image_path": "66e6889670bb1e0db1b5f0b19616a18f0e8840e66bf66004b44553c71ff4802c.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 58, + 283, + 298, + 518 + ], + "blocks": [ + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "type": "text", + "content": "Table 1. Segmentation comparisons on VOC and COCO. Net. is the backbone for segmentation. Sup. is the supervision type. " + }, + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "type": "text", + "content": " image-level labels. SA: saliency maps. " + }, + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "type": "text", + "content": " : language." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 283, + 298, + 518 + ], + "lines": [ + { + "bbox": [ + 58, + 283, + 298, + 518 + ], + "spans": [ + { + "bbox": [ + 58, + 283, + 298, + 518 + ], + "type": "table", + "html": "
MethodSup.Net.VOCCOCO
ValTestVal
Multi-stage WSSS methods.
L2G [14] CVPR'2022I + SARN10172.171.744.2
RCA [49] CVPR'2023I + SARN3872.272.836.8
OCR [7] CVPR'2023IRN3872.772.042.5
BECO [26] CVPR'2023IRN10173.773.545.1
MCTformer+ [38] TPAMI'2024IRN3874.073.645.2
CTI [41] CVPR'2024IRN10174.173.245.4
CLIMS [36] CVPR'2022I + LRN10170.470.0-
CLIP-ES [20] CVPR'2023I + LRN10172.272.845.4
PSDPM [45] CVPR'2024I + LRN10174.174.947.2
CPAL [31] CVPR'2024I + LRN10174.574.746.8
Single-stage WSSS methods.
AFA [28] CVPR'2022IMiT-B166.066.338.9
ViT-PCM [27] ECCV'2022IViT-B70.370.9-
ToCo [29] CVPR'2023IViT-B71.172.242.3
DuPL [35] CVPR'2024IViT-B73.372.844.6
SeCo [39] CVPR'2024IViT-B74.073.846.7
DIAL [13] ECCV'2024I + LViT-B74.574.944.4
WeCLIP [43] CVPR'2024I + LViT-B76.477.247.1
ExCEL(w/o CRF)I + LViT-B77.277.349.3
ExCEL (Ours)I + LViT-B78.478.550.3
", + "image_path": "28a7a25060e08e8f4dca1015bbfe51829f5c02bb76c80736e0320827a63f6999.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 531, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 555 + ], + "type": "text", + "content": "as 30,000 for VOC and 100,000 for COCO. Please refer to Supplementary Materials for more details." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 564, + 283, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 283, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 283, + 577 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Performance of Semantic Segmentation. Tab. 1 shows segmentation comparisons between our ExCEL and recent methods on VOC and COCO. The single-stage ExCEL achieves " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "78.4\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " mIoU on VOC val set and test set, which even significantly outperforms the sophisticated multi-stage methods by at least " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " mIoU, respectively. For more complicated benchmark COCO, ExCEL achieves " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "50.3\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " mIoU on val set, which brings a noticeable " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " increase over the CLIP-based state-of-the-art (SOTA) WeCLIP. In addition, without time-consuming post-processing techniques, such as CRF [15], ExCEL still" + } + ] + } + ], + "index": 42 + }, + { + "type": "table", + "bbox": [ + 318, + 284, + 553, + 517 + ], + "blocks": [ + { + "bbox": [ + 313, + 249, + 553, + 282 + ], + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 282 + ], + "type": "text", + "content": "Table 2. CAM seed comparisons on VOC train set. " + }, + { + "bbox": [ + 313, + 249, + 553, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 249, + 553, + 282 + ], + "type": "text", + "content": ": multistage methods. " + }, + { + "bbox": [ + 313, + 249, + 553, + 282 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 249, + 553, + 282 + ], + "type": "text", + "content": ": single-stage methods. †: our reproduction following official codes. ExCEL*: ExCEL in a training-free manner." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 284, + 553, + 517 + ], + "lines": [ + { + "bbox": [ + 318, + 284, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 318, + 284, + 553, + 517 + ], + "type": "table", + "html": "
MethodTypeSup.Net.VOC
Train
Training-free WSSS methods.
CLIP-ES [20] CVPR'2023MI + LViT-B70.8
ExCEL* (Ours)SI + LViT-B74.6
Training-required WSSS methods.
ReCAM [6] CVPR'2022MIRN10154.8
FPR [3] CVPR'2023MIRN10163.8
LPCAM [5] CVPR'2023MIRN5065.3
MCTformer+ [38] TPAMI'2024MIRN3868.8
SFC [44] AAAI'2024MIRN10164.7
CTI [41] CVPR'2024MIRN10169.5
AFA [28] CVPR'2022SIMiT-B165.0
ViT-PCM [27] ECCV'2022SIViT-B67.7
†ToCo [29] CVPR'2023SIViT-B71.6
†DuPL [35] CVPR'2024SIViT-B75.0
SeCo [39] CVPR'2024SIViT-B74.8
CLIMS [36] CVPR'2022MI + LRN10156.6
POLE [22] WACV'2023MI + LRN5059.0
CPAL [31] CVPR'2024MI + LRN10171.9
DIAL [13] ECCV'2024SI + LViT-B75.2
†WeCLIP [43] CVPR'2024SI + LViT-B75.4
ExCEL (Ours)SI + LViT-B78.0
", + "image_path": "79e2d5d12d4ed2e1c483c616f844662fbea26252cc1e980b488925d82dec3ac6.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "table_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 533, + 538, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 538, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 538, + 544 + ], + "type": "text", + "content": "maintains consistent superiority over SOTAs with CRF." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 545, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 593 + ], + "type": "text", + "content": "The qualitative comparisons on VOC and COCO are shown in Fig. 3. By densely matching the patches and texts, ExCEL consistently demonstrates more precise object segmentation than recent methods in an image-text paradigm." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Evaluation of CAM Seeds. Tab. 2 reports the quality of raw CAM seeds on VOC train set. Compared with recent methods, ExCEL achieves " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "74.6\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " mIoU in a training-free setup, outperforming CLIP-ES by " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and performing comparably to most training-required methods. With the optimized LVC module, ExCEL further boosts CAM quality to " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "78.0\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", surpassing SOTAs in the image-text paradigm by at least " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". In addition, visual comparisons in Fig. 4 (e-h) also plainly illustrates that ExCEL generates better CAMs with the designed patch-text alignment paradigm." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20228" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 228 + ], + "type": "image", + "image_path": "f6fd9378d29824e2ce862dab5fc4b6daaa14aea99033ac822c2f0c8a6a7b0851.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 261 + ], + "type": "text", + "content": "Figure 4. CAM visualizations on VOC train set. (a) Image. (b-e) Ablative visualizations of proposed modules. (e-h) Qualitative comparisons of (e) ExCEL and recent CLIP-based methods, i.e., (f) WeCLIP [43], (g) CLIP-ES [20] and (h) MaskCLIP [47]. (i) Ground truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 281, + 297, + 350 + ], + "blocks": [ + { + "bbox": [ + 82, + 270, + 268, + 280 + ], + "lines": [ + { + "bbox": [ + 82, + 270, + 268, + 280 + ], + "spans": [ + { + "bbox": [ + 82, + 270, + 268, + 280 + ], + "type": "text", + "content": "Table 3. Ablation study of ExCEL on VOC val set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 281, + 297, + 350 + ], + "lines": [ + { + "bbox": [ + 56, + 281, + 297, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 297, + 350 + ], + "type": "table", + "html": "
ConditionsSVCTSELVCPrecisionRecallmIoU
Baseline (CLIP)18.821.312.1
w/ SVC✓81.286.272.5
w/o LVC✓✓80.789.874.7
w/o TSE✓✓83.786.375.1
ExCEL✓✓✓85.088.477.2
", + "image_path": "97dcac90ecc71ccd31aca4fd613d8bfd71403ef11362dbe99990eadfd4152a14.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 370, + 297, + 402 + ], + "blocks": [ + { + "bbox": [ + 61, + 359, + 289, + 369 + ], + "lines": [ + { + "bbox": [ + 61, + 359, + 289, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 359, + 289, + 369 + ], + "type": "text", + "content": "Table 4. Ablation study of attribute number " + }, + { + "bbox": [ + 61, + 359, + 289, + 369 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 61, + 359, + 289, + 369 + ], + "type": "text", + "content": " on VOC val set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 370, + 297, + 402 + ], + "lines": [ + { + "bbox": [ + 57, + 370, + 297, + 402 + ], + "spans": [ + { + "bbox": [ + 57, + 370, + 297, + 402 + ], + "type": "table", + "html": "
Number of AttrNone3264112144196
mIoU75.175.876.277.277.076.5
", + "image_path": "23376d17613377901da151701a378f90d8d2d38996a8919fc1e5565bcd17c967.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 422, + 297, + 491 + ], + "blocks": [ + { + "bbox": [ + 72, + 411, + 278, + 422 + ], + "lines": [ + { + "bbox": [ + 72, + 411, + 278, + 422 + ], + "spans": [ + { + "bbox": [ + 72, + 411, + 278, + 422 + ], + "type": "text", + "content": "Table 5. Ablation study of VC module on VOC train set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 422, + 297, + 491 + ], + "lines": [ + { + "bbox": [ + 56, + 422, + 297, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 297, + 491 + ], + "type": "table", + "html": "
Conditionsq-kvI.C.M.C.LVCPrecisionRecallmIoU
Baseline (CLIP)✓18.021.811.2
MaskCLIP✓77.180.965.8
w/ I.C.✓79.184.769.7
SVC✓✓82.288.274.6
ExCEL✓✓✓86.687.978.0
", + "image_path": "d8ec2218118d0fa0916c231c388efb0bf1dcdf9e721bae7ce6c9ad95a22d135e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 156, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 156, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 156, + 514 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "text", + "content": "Efficacy of Key Components. Quantitative ablative experiments of ExCEL are reported in Tab. 3. The baseline is the vanilla CLIP using our training settings, which only achieves " + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "inline_equation", + "content": "12.1\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "text", + "content": " mIoU for segmentation. Our SVC module replaces the q-k attention with Intra-correlation from the intermediate layers. The performance increases to " + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "inline_equation", + "content": "72.5\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "text", + "content": ". TSE module enriches the semantics of text representation for robust visual recognition. Introducing TSE brings a " + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "text", + "content": " recall increase compared to the original text templates. LVC module provides a dynamic shift to diversify the features. It further benefits SVC's segmentation performance to " + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "text", + "content": " mIoU. With all these enhancements, ExCEL generates " + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 676 + ], + "type": "text", + "content": " mIoU for segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Qualitative ablation results are further illustrated in Fig. 4 (b-e) to evaluate the efficacy of our modules. In Fig. 4 (b), the CLIP baseline produces inaccurate CAMs with mis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 378 + ], + "type": "text", + "content": "localized activation. SVC corrects token relations and preserves fine-grained details, effectively suppressing false activations, as seen in Fig. 4 (c). TSE incorporates comprehensive textual attributes into the text representation, enhancing patch-text matching and producing more complete CAMs, shown in Fig. 4 (d). LVC dynamically optimizes attention maps, further improving CAM accuracy and completeness, as illustrated in Fig. 4 (e). Both quantitative and qualitative results confirm the effectiveness of our modules." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "text", + "content": "Effectiveness of Implicit Attributes. Tab. 4 analyzes the effect of varying the number of clustering attributes. 'None' means no clustering and we explicitly fuse the " + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "text", + "content": " description embeddings for each class. It shows that the performance drops from " + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "text", + "content": ", which validates the efficacy of implicit attributes and its superiority over explicit descriptions. With this operation, we can expand the representation of 20 classes up to 196 attributes or more, greatly enhancing text semantics. It reports that ExCEL achieves more favorable performance when " + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 379, + 554, + 498 + ], + "type": "text", + "content": " is set to 112." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": "Effectiveness of Visual Calibrations. Tab. 5 compares different strategies in VC module for CAM generation. I.C. refers to Intra-correlation in the last layer, and M.C. (Intermediate Calibration) applies I.C. across intermediate layers. Vanilla q-k attention in CLIP loses diversity and cannot generate reasonable CAMs. " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": " contains fine-grained knowledge and MaskCLIP improves CAMs to " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "65.8\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": " by using " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": " from the last layer. In contrast, our I.C. and M.C. focus on mining diverse knowledge from intermediate layers and boost the performance to " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "69.7\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "74.6\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": ". In addition, introducing LVC module raises the final performance to " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "78.0\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": ". Results in Tab. 5 and corresponding visualizations in Fig. 4 clearly highlight the efficacy of our components." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 660, + 417, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 417, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 417, + 673 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Hyper-parameter Analysis. Hyper-parameters, such as TOP-K, scaling factors " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": ", and the number of SVC layers " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": ", etc., are discussed in Supplementary Materials." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20229" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 92, + 298, + 244 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 6. Comparisons with the fully-supervised counterparts on VOC val set. " + }, + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "text", + "content": " :fully-supervised. ViT-B*: pretrained from CLIP." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 92, + 298, + 244 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 298, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 298, + 244 + ], + "type": "table", + "html": "
MethodsTypeSup.Net.ValRatio
DeepLabV2 [4] TPAMI'2017-FRN10177.7-
DeepLabV2 [4] TPAMI'2017-FViT-B82.3-
WeCLIP-Full [43] CVPR'2024-FViT-B*81.6-
CLIMS [36] CVPR'2022MI + LRN10170.490.6%
CLIP-ES [20] CVPR'2023MI + LRN10172.292.9%
CPAL [31] CVPR'2024MI + LRN10174.595.9%
ToCo [29] CVPR'2024SIViT-B71.186.4%
DuPL [35] CVPR'2024SIViT-B73.389.1%
SeCo [43] CVPR'2024SIViT-B74.089.9%
DIAL [13] ECCV'2024SI + LViT-B74.590.5%
WeCLIP [43] CVPR'2024SI + LViT-B*76.493.6%
ExCEL (Ours)SI + LViT-B*78.496.1%
", + "image_path": "c6f1b183998189eb28bdabdbe4728e64bce83267cd36a9814988a3261b38caec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 272, + 297, + 392 + ], + "blocks": [ + { + "bbox": [ + 56, + 250, + 295, + 271 + ], + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 271 + ], + "type": "text", + "content": "Table 7. Training efficiency comparisons on VOC train (CAM) and val set (Seg). All experiments are conducted on RTX 3090." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 272, + 297, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 272, + 297, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 297, + 392 + ], + "type": "table", + "html": "
MethodTypeTraining TimeGPUCAMSeg
CLIMS [36] CVPR'2022M1068 mins18.0 G56.670.4
CLIP-ES [20] CVPR'2023M420 mins12.0 G70.872.2
MCTformer+ [38] TPAMI'2024M1496 mins18.0 G68.874.0
ToCo [29] CVPR'2023S506 mins17.9 G71.671.1
DuPL [35] CVPR'2024S508 mins14.9 G75.073.3
SeCo [39] CVPR'2024S407 mins17.6 G74.874.0
WeCLIP [43] CVPR'2024S270 mins6.2 G75.476.4
ExCEL* (Training-free)S-2.9 G74.6-
ExCEL (Ours)S90 mins3.2 G78.078.4
", + "image_path": "df209429a16cb59305b2099d05e746d69b0062b25292e73bee4e669831ae8efc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "text", + "content": "Fully-supervised Counterparts. Tab. 6 presents a fairness comparison between WSSS methods and their fully-supervised counterparts using the same segmentation backbone. With CLIP's visual encoder as the backbone, ExCEL achieves " + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "inline_equation", + "content": "78.4\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "text", + "content": " mIoU, reaching " + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "inline_equation", + "content": "96.1\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "text", + "content": " of the fully-supervised performance. It significantly outperforms CLIP-based WeCLIP by " + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "text", + "content": " and demonstrates ExCEL's advantage over other multi-stage CLIP-based methods as well." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "text", + "content": "Training Efficiency Analysis. Our method only trains the adapter and decoder in a single-stage paradigm. Tab. 7 compares the training efficiency between ExCEL and recent methods. Without training, ExCEL requires just 2.9 GB of GPU memory and generates comparable CAMs to recent SOTAs. When training is included, the entire pipeline only takes 90 minutes and 3.2 GB of memory for SOTA performance. ExCEL just requires " + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "text", + "content": " training time of multi-stage MCTformer+ and " + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "inline_equation", + "content": "33.3\\%" + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "text", + "content": " of single-stage WeCLIP, highlighting ExCEL's remarkable training efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": "Attribute Response Analysis. We treat text prompting as an implicit attribute-hunting process to comprehensively enrich text representation semantics. To evaluate if the clustered attributes capture distinct object characteristics, we visualize 5 implicit attributes based on similarity scores. As shown in Fig. 5, given instances of " + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\{\\text{aeroplane}\\}" + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\{\\text{train}\\}" + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": ", our attributes highlight different object parts, which clearly validates that our TSE module enhances inte" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 551, + 134 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 551, + 134 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 551, + 134 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 551, + 134 + ], + "type": "image", + "image_path": "178f8d7f60bfa66fe2b0fd0d2b2000aaf26998844024d8e16377bf35b7bb58d7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 138, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 313, + 138, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 553, + 159 + ], + "type": "text", + "content": "Figure 5. Implicit attribute responses. Based on the TOPK similarity scores, 5 attributes are sampled for visualizations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 163, + 556, + 245 + ], + "blocks": [ + { + "bbox": [ + 316, + 163, + 556, + 245 + ], + "lines": [ + { + "bbox": [ + 316, + 163, + 556, + 245 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 556, + 245 + ], + "type": "image", + "image_path": "5d9e85d1175b51ab2d9ff7dd9fe40e8f79ab87d68d4b4f0c327585ba9676213e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 250, + 553, + 283 + ], + "lines": [ + { + "bbox": [ + 313, + 250, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 553, + 283 + ], + "type": "text", + "content": "Figure 6. Comparisons of attention maps from the last visual encoder layer. (a) Attention features from the query patches (marked by red stars). (b) Token relations measured by cosine similarity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 531, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 531, + 300 + ], + "type": "text", + "content": "gral visual responses by gathering relevant semantics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 455 + ], + "type": "text", + "content": "Feature Representation Analysis. CLIP lacks fine-grained details, leading to inaccurate patch-text responses. To explore further, we visualize the self-attention features in Fig. 6 (a). Given the query patch (red star), CLIP's q-k attention falls short in generating diverse features, supporting our claim in Sec. 3.3. MaskCLIP observes that " + }, + { + "bbox": [ + 313, + 300, + 553, + 455 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 300, + 553, + 455 + ], + "type": "text", + "content": " keeps diversity and takes it from the last layer for visual response. We visualize it by calculating " + }, + { + "bbox": [ + 313, + 300, + 553, + 455 + ], + "type": "inline_equation", + "content": "v - v" + }, + { + "bbox": [ + 313, + 300, + 553, + 455 + ], + "type": "text", + "content": " attention. Although effective, MaskCLIP still misses fine granularity. Instead, SVC calculates attention within each space and implements it from intermediate layers. LVC further diversifies the features with a dynamic adapter, both of which effectively generate features with clear boundaries and spatial details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 540 + ], + "type": "text", + "content": "Additionally, we explore the pairwise token relations in Fig. 6 (b). Unlike the smoother attention maps of CLIP or MaskCLIP, our approach distinctly groups tokens with similar semantics, aligning pairwise similarities with corresponding semantics. This validates that ExCEL successfully enhances the frozen features of CLIP by calibrating it towards distributions with more diverse spatial information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 550, + 387, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 387, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 387, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose ExCEL, a novel patch-text alignment method to explore CLIP's dense knowledge for WSSS, which provides a different insight to generate better pseudo labels based on CLIP. To this end, Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules are designed to improve the dense alignment across text and vision modalities. In addition, ExCEL generates CAMs in both training-free and efficient training modes, calibrating CLIP without altering its pre-trained weights. It retains CLIP's transferability while significantly reducing training cost. We believe ExCEL can inspire more future research to unlock CLIP's dense capabilities in the WSSS field." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20230" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "type": "text", + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In CVPR, pages 4981-4990, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 250, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 250, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 250, + 296, + 315 + ], + "type": "text", + "content": "[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 549-565. Springer, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 316, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 296, + 358 + ], + "type": "text", + "content": "[3] Liyi Chen, Chenyang Lei, Ruihuang Li, Shuai Li, Zhaoxiang Zhang, and Lei Zhang. Fpr: False positive rectification for weakly supervised semantic segmentation. In ICCV, pages 1108-1118, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 447 + ], + "type": "text", + "content": "[5] Zhaozheng Chen and Qianru Sun. Extracting class activation maps from non-discriminative features as well. In CVPR, pages 3135-3144, 2023. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 491 + ], + "type": "text", + "content": "[6] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, pages 969–978, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 493, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 296, + 557 + ], + "type": "text", + "content": "[7] Zesen Cheng, Pengchong Qiao, Kehan Li, Siheng Li, Pengxu Wei, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Out-of-candidate rectification for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23673-23684, 2023. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 559, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 591 + ], + "type": "text", + "content": "[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, pages 1635-1643, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 593, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 296, + 656 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "type": "text", + "content": "[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In CVPR, pages 4320–4329, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[11] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant No.82372097, Shanghai Sailing Program under Grant 22YF1409300, International Science and Technology Cooperation Program under the 2023 Shanghai Action Plan for Science under Grant 23410710400, Taishan Scholars Program under Grant NO.tsqn202408245." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[13] Soojin Jang, Jungmin Yun, Junehyoung Kwon, Eunju Lee, and Youngbin Kim. Dial: Dense image-text alignment for weakly supervised semantic segmentation. arXiv preprint arXiv:2409.15801, 2024. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "text", + "content": "[14] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[15] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 24, 2011. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "text", + "content": "[16] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "text", + "content": "[17] Xueyi Li, Tianfei Zhou, Jianwu Li, Yi Zhou, and Zhaoxiang Zhang. Group-wise semantic mining for weakly supervised semantic segmentation. In AAAI, pages 1984-1992, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 378, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 553, + 420 + ], + "type": "text", + "content": "[18] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In CVPR, pages 3159-3167, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 422, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 488 + ], + "type": "text", + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 490, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 543 + ], + "type": "text", + "content": "[20] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. arXiv preprint arXiv:2212.09506, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "text", + "content": "[21] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 569, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 555, + 633 + ], + "type": "text", + "content": "[22] Balamurali Murugesan, Rukhshanda Hussain, Rajarshi Bhattacharya, Ismail Ben Ayed, and Jose Dolz. Prompting classes: exploring the power of prompt class learning in weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 291-302, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[23] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In CVPR, pages 1713-1721, 2015. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20231" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[26] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "text", + "content": "[27] Simone Rossetti, Damiano Zappia, Marta Sanzari, Marco Schaerf, and Fiora Pirri. Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In ECCV, pages 446-463. Springer, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "type": "text", + "content": "[28] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: end-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "text", + "content": "[29] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. arXiv preprint arXiv:2303.01267, 2023. 2, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 419 + ], + "type": "text", + "content": "[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "text", + "content": "[31] Feilong Tang, Zhongxing Xu, Zhaojun Qu, Wei Feng, Xingjian Jiang, and Zongyuan Ge. Hunting attributes: Context prototype-aware learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3324-3334, 2024. 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "text", + "content": "[32] Paul Vernaza and Manmohan Chandraker. Learning random-walk label propagation for weakly-supervised semantic segmentation. In CVPR, pages 7158-7166, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[33] Xiang Wang, Shaodi You, Xi Li, and Huimin Ma. Weakly-supervised semantic segmentation by iteratively mining common object features. In CVPR, pages 1354-1362, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "text", + "content": "[34] Yuanchen Wu, Xiaogiang Li, Jide Li, Pinpin Zhu, Shaohua Zhang, et al. Dino is also a semantic guider: Exploiting class-aware affinity for weakly supervised semantic segmentation. In ACM Multimedia 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "text", + "content": "[35] Yuanchen Wu, Xichen Ye, Kequan Yang, Jide Li, and Xiaoqiang Li. Dupl: Dual student with trustworthy progressive learning for robust weakly supervised semantic segmentation. In CVPR, pages 3534-3543, 2024. 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Jinheng Xie, Xianxu Hou, Kai Ye, and Linlin Shen. Clims: cross language image matching for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4483-4492, 2022. 1, 3, 6, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 708 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[37] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[38] Lian Xu, Mohammed Bennamoun, Farid Boussaid, Hamid Laga, Wanli Ouyang, and Dan Xu. Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 2024. 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[39] Zhiwei Yang, Kexue Fu, Minghong Duan, Linhao Qu, Shuo Wang, and Zhijian Song. Separate and conquer: Decoupling co-occurrence via decomposition and representation for weakly supervised semantic segmentation. In CVPR, pages 3606–3615, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[40] Zhiwei Yang, Yuong Meng, Kexue Fu, Shuo Wang, and Zhijian Song. Tackling ambiguity from perspective of uncertainty inference and affinity diversification for weakly supervised semantic segmentation. arXiv preprint arXiv:2404.08195, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[41] Sung-Hoon Yoon, Hoyong Kwon, Hyeonseong Kim, and Kuk-Jin Yoon. Class tokens infusion for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3595-3605, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "text", + "content": "[42] Bingfeng Zhang, Xuru Gao, Siyue Yu, and Weifeng Liu. Enhanced online cam: Single-stage weakly supervised semantic segmentation via collaborative guidance. Pattern Recognition, 156:110787, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 397, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 453 + ], + "type": "text", + "content": "[43] Bingfeng Zhang, Siyue Yu, Yunchao Wei, Yao Zhao, and Jimin Xiao. Frozen clip: A strong backbone for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3796-3806, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "text", + "content": "[44] Xinqiao Zhao, Feilong Tang, Xiaoyang Wang, and Jimin Xiao. Sfc: Shared feature calibration in weakly supervised semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7525-7533, 2024. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[45] Xinqiao Zhao, Ziqian Yang, Tianhong Dai, Bingfeng Zhang, and Jimin Xiao. Psdpm: Prototype-based secondary discriminative pixels mining for weakly supervised semantic segmentation. In CVPR, pages 3437-3446, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "text", + "content": "[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 587, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 620 + ], + "type": "text", + "content": "[47] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 622, + 553, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 664 + ], + "type": "text", + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 666, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 666, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 666, + 553, + 708 + ], + "type": "text", + "content": "[49] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In CVPR, pages 4299-4309, 2022. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20232" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_content_list.json b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ad062d85a9adf9118df67522c27b58d0919d66b --- /dev/null +++ b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_content_list.json @@ -0,0 +1,1313 @@ +[ + { + "type": "text", + "text": "Exploring Contextual Attribute Density in Referring Expression Counting", + "text_level": 1, + "bbox": [ + 124, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhicheng Wang Zhiyu Pan* Zhan Peng Jian Cheng Liwen Xiao Wei Jiang Zhiguo Cao School of AIA, Huazhong University of Science and Technology", + "bbox": [ + 117, + 180, + 879, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhicheng_wang, zhiyupan, zgcao}@hust.edu.cn", + "bbox": [ + 300, + 219, + 692, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring expression counting (REC) algorithms are for more flexible and interactive counting ability across varied fine-grained text expressions. However, the requirement for fine-grained attribute understanding poses challenges for prior arts, as they struggle to accurately align attribute information with correct visual patterns. Given the proven importance of \"visual density\", it is presumed that the limitations of current REC approaches stem from an under-exploration of \"contextual attribute density\" (CAD). In the scope of REC, we define CAD as the measure of the information intensity of one certain fine-grained attribute in visual regions. To model the CAD, we propose a U-shape CAD estimator in which referring expression and multi-scale visual features from GroundingDINO can interact with each other. With additional density supervision, we can effectively encode CAD, which is subsequently decoded via a novel attention procedure with CAD-refined queries. Integrating all these contributions, our framework significantly outperforms state-of-the-art REC methods, achieves $30\\%$ error reduction in counting metrics and a $10\\%$ improvement in localization accuracy. The surprising results shed light on the significance of contextual attribute density for REC. Code will be at github.com/Xu3XiWang/CAD-GD.", + "bbox": [ + 89, + 301, + 483, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 681, + 220, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object counting algorithms estimate the number of specific objects in query images. Early research has explored various forms of object descriptions, including class-specific [26, 37], class-agnostic [27, 29, 32, 33], exemplar-guided [29], and text-guided [16, 36] approaches. However, these methods lack the flexibility to distinguish between fine-grained, open-world descriptions, e.g., differentiating a walking person from a person riding motorcycle. To address this, recent work [8] proposes the referring expression counting (REC) task which aims to count fine-grained and contextually nuanced objects within the same class.", + "bbox": [ + 89, + 707, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1dba6845c6edfd7cb3e9aff6d3eba562bb6085dd966702fe9e789ce9ad1f1993.jpg", + "image_caption": [ + "Figure 1. The efficacy of contextual attribute density (CAD). According to the qualitative comparison above, the REC baseline [8] often over-counts when it focuses primarily on class information rather than fine-grained attributes, and it under-counts in cases of occlusion or scale variations. By modeling CAD, our method more accurately aligns fine-grained attributes with corresponding visual regions, effectively reducing counting errors." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 906, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GroundingREC [8] is firstly developed as the REC baseline with the power of the open-world detection model GroundingDINO [22]. However, distinguishing between objects of the same class based on nuanced attribute differences is challenging, which makes the REC baseline struggle to align attributes with corresponding visual regions. As shown in Fig. 1, two primary types of failure cases can be observed in the REC baseline: (1) Over-counting: the baseline may place excessive emphasis on class information while neglecting fine-grained attributes, e.g., walking, leading to the inclusion of objects with wrong attributes; (2) Under-counting: the baseline might miss objects with the", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19587", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "given attribute when occlusion and scale variations arise. These two failure cases may be due to the baseline's reliance on one-to-one matching between texts and image regions, a typical detection-to-count pipeline. Many previous counting approaches [4, 6, 12, 15] have proved that visual density can provide information about scale-robust spatial distribution. However, in modern open-world perception models [18, 22, 30], the capacity to perceive density for a specific attribute is often overlooked. Inspired by the success of visual density modeling, we propose investigating whether modeling contextual attribute density (CAD) for an open-world perception model could help address the above issues in a cross-modal context.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we attempt to make the REC model aware of the CAD. Analogous to the concept of visual density, we define CAD as the information intensity measurement of one certain fine-grained attribute about one object conveyed in query image regions of different scales. Based on our definition, modeling CAD requires both the image and text information, with the image information encoded in a multi-scale format. Hence, we propose a CAD generate module with U-shape CAD estimator (CADE) that encodes the multi-layer CAD features. The CADE is provided with the visual features and similarity features between visual and text to estimates the CAD map for a referring expression, where the intermediate features serve as representations of CAD. After encoding the CAD, the next problem is how to facilitate the counting procedure with the CAD features. We fuse the CAD features with original visual features from GroundingDINO via both spatial and channel attention. In this way, spatial location and channel importance prior are provided for visual features for referring expression separately. Additionally, the counting queries in the localization decoder for counting are also dynamically initialized by both the text features and their corresponding CAD information. With this dynamic initialization module, the feature differentiation of the initial queries under different inputs is enhanced for better counting accuracy.", + "bbox": [ + 91, + 291, + 483, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Integrating CAD into our REC framework leads to substantial improvements. Experimental results on the public benchmark REC-8K [8] show that our framework outperforms previous state-of-the-art REC baselines by large margins, achieving $37.8\\%$ and $33.2\\%$ relative error reductions on the validation and test sets, respectively. Besides, We also conduct experiments on the previous class-agnostic counting benchmark FSC-147 [29] with a zero-shot setting, with results further underscoring the effectiveness of our framework. These findings highlight the significance of CAD. Our contributions are three-fold:", + "bbox": [ + 89, + 671, + 482, + 835 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We pioneer the concept of contextual attribute density in REC, enabling more precise differentiation between fine-grained descriptions of objects within the same class.", + "- We propose the contextual attribute density estimator and" + ], + "bbox": [ + 89, + 840, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "an attention procedure with a query enhancement, including the contextual attribute density into the REC pipeline.", + "bbox": [ + 526, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We achieve significant performance improvements on public benchmarks, demonstrating the value of contextual attribute density modeling for REC.", + "bbox": [ + 511, + 122, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 180, + 653, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The progression toward more generalized, fine-grained counting tasks has shifted from class-specific to class-agnostic methods, and also from visual to textual guidance. This section reviews prior works contributing to the development of open-world counting models, emphasizing advancements in visual density for counting tasks given our introduction of contextual attribute density.", + "bbox": [ + 511, + 205, + 903, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advancing to open-world counting. Counting tasks originally focus on specific categories, such as crowd [1, 4, 15], vehicle [11], wheat head [24, 25], and so on. These early methods are limited to predefined target types and struggle to generalize across diverse object categories. The class-agnostic counting (CAC) [23] marks a significant step toward more flexible counting approaches [27, 32, 33, 35]. Defined as a template-matching problem [23], CAC achieves generalized counting by using visual exemplars to count objects beyond specific classes. Notably, visual density-based methods [32, 33, 35] consistently outperform traditional detection-based counting methods such as MaskRCNN [10] and RetinaNet [31] on counting benchmarks. However, the reliance on visual templates is still a notable limitation in practical applications. To address this issue, textual-based counting [16, 36] is proposed by leveraging the multi-modal ability of CLIP [28]. Although incorporating visual guidance information has been proved beneficial [3], researchers pursue a more generalized setting: with only text expressions, fine-grained, general-purpose objects are expected to be counted, which is referring expression counting (REC) [8]. The REC baseline solely relies on the multi-modal localization ability of GroundingDINO [22], and interpreting some fine-grained expressions accurately remains challenging, as shown in Fig. 1.", + "bbox": [ + 511, + 311, + 903, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Density modeling in computer vision. Beyond the task of counting, there are some works in other fields using visual density features to provide the spatial and target number prior, e.g., instance segmentation [7] and tiny object detection [9, 13, 20, 38]. Cholakka et al. [7] use density maps to provide prior information on object counts and spatial locations, thereby assisting instance detection tasks under image-level supervision. Huang et al. [13] propose DQ-DETR for tiny object detection, which utilizes the density feature to provide the general target number and spatial prior for DETR [5] structure detection pipeline. In contrast to these methods, our approach focuses more on the role of contextual attribute density for fine-grained expression understanding in a multi-modal setting.", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19588", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d1a3e0b8b14e70ebc548d9f6b9d86b554208a0897259ace4a33371bb8542176.jpg", + "image_caption": [ + "Figure 2. The framework of Contextual Attribute Density Aware GroundingDINO (CAD-GD). A query image and referring expression are sent into backbones separately together with the feature enhancer to get the visual features and the text feature. Then we obtain the Contextual Attribute Density (CAD) features by the CAD Generate Module supervised by the GT contextual attribute density map. To leverage the CAD information, we design the CAD Attention Module and CAD Dynamic Query Module to enhance visual features and query contents separately. Then we sent the dynamic queries, CAD enhanced visual features, and the text feature into the localization decoder to get the final localization prediction." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 420, + 181, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the overview of our framework for Referring Expression Counting (REC), named Contextual Attribute Density Aware GroundingDINO (CAD-GD). The overview diagram of our approach is shown in Fig. 2. Given an input query image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ and a referring expression text $T = \\{w_i\\}_{i=1}^N$ with $N$ words, our model predicts a localization counting result. For each $(I, T)$ pair, we first send multi-scale image features and text features extracted from the image and text backbones into feature enhancer [22] to obtain cross-modality visual features $\\{F_{vi} \\in \\mathbb{R}^{h_i \\times w_i \\times C}\\}_{i=1}^4$ and text feature $F_t \\in \\mathbb{R}^{N \\times C}$ where the $C$ indicates the number of channels.", + "bbox": [ + 89, + 446, + 483, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Afterwards, we feed visual features $\\{F_{vi}\\}_{i = 1}^{4}$ and text feature $F_{t}$ into the U-shape contextual attribute density estimator (CADE) to obtain the contextual attribute density (CAD) features $\\{D_i\\in \\mathbb{R}^{h_i\\times w_i\\times C}\\}_{i = 1}^4$ . We then inject the CAD features $\\{D_i\\}_{i = 1}^4$ into visual features $\\{F_{vi}\\}_{i = 1}^{4}$ to enhance the visual feature with spatial and channel attention in CAD Attention Module. Besides, for the initialization of queries in the cross-modality localization decoder [22], we selected the $K$ visual feature positions with the highest scores among them as the initial query locations by calculating the similarity between the visual and text features, while for the initial content features of the query, we dynamically initialize them with text features $F_{t}$ and CAD features $\\{D_i\\}_{i = 1}^4$ corresponding to the query positions. Then we refine the queries by the cross-modality decoder at every layer with CAD enhanced visual and text features. The output queries of the last decoder layer will be used to predict the localization of referring expression objects.", + "bbox": [ + 89, + 628, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. CAD Generation Module", + "text_level": 1, + "bbox": [ + 513, + 421, + 743, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first step to model CAD for referring expression counting is to obtain the CAD features, which should be responsive to input text with fine-grained attributes. To obtain the satisfying CAD features, we identify several requirements for obtaining satisfying CAD features: (a) the image-text similarity features are required to guide the estimator to focus on areas corresponding to the input referring expression; (b) due to the density diversity of targets in the counting task, a multi-scale structure is of benefit for the robustness of counting [14].", + "bbox": [ + 511, + 444, + 906, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the above requirements, the CAD Generate Module is proposed, as in Fig. 2. Specifically, we first utilize the well-pretrained linear projector from GroundingDINO to project the visual features $\\{F_{vi}\\}_{i = 1}^{4}$ into the text feature space. Then we calculate the similarity feature $\\{S_i\\in \\mathbb{R}^{h_i\\times w_i\\times C}\\}_{i = 1}^4$ between the well-aligned visual feature $F_{vi}$ and corresponding referring expression feature $F_{t}$ as follow:", + "bbox": [ + 511, + 597, + 906, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i} = \\operatorname {P r o j} \\left(F _ {v i}\\right) \\cdot F _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 720, + 903, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the Proj consists of a single linear layer followed by a layer norm, and the $\\cdot$ represents the element-wise multiplication. Then we send the $\\{S_i\\}_{i=1}^4$ together with the visual feature $\\{F_{vi}\\}_{i=1}^4$ into a U-shape CAD estimator (CADE) to estimate the CAD map for a referring expression, where the intermediate features serve as the CAD features $\\{D_i\\}_{i=1}^4$ .", + "bbox": [ + 511, + 748, + 906, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervision Signal. The final output of CADE is the density map for referring expression. In order to guide the CADE to learn the contextual attribute density map, we adopt a conventional $\\ell_2$ loss as the counting loss $\\mathcal{L}_{\\mathrm{density}}$ :", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19589", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e n s i t y}} = \\left\\| D _ {\\text {p r e d}} (I, T) - D _ {\\mathrm {g t}} (I, T) \\right\\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 104, + 482, + 123 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{\\mathrm{pred}}(I,T)$ and $D_{\\mathrm{gt}}(I,T)$ denote the prediction and ground truth contextual attribute density map respectively.", + "bbox": [ + 89, + 128, + 482, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CAD Attention Module", + "text_level": 1, + "bbox": [ + 89, + 167, + 307, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the CAD features $\\{D_i\\}_{i=1}^4$ , the next core step is to consider how to inject the CAD information into the localization process, which follows a typically DETR decoder structure [22]. In the DETR-like structure, the visual features $\\{F_{vi}\\}_{i=1}^4$ are crucial for the localization process due to its cross-attention with queries. Motivated by the high relevance between the density feature and the attention map [35], we inject the CAD features into visual features by attention mechanism, as shown in Fig. 2.", + "bbox": [ + 89, + 188, + 482, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Attention. Since the visual features need to be matched with queries to find the target location, foreground enhancement for visual features on areas corresponding to the input referring expression can help to get accurate localization results. Hence we design the spatial attention mechanism to enhance the CAD corresponding foreground areas, utilizing distribution prior from CAD. In specific, for the $i_{th}$ scale of $\\{D_i\\}_{i=1}^4$ , we first apply the average pooling and max pooling for $D_i$ on the channel dimension. Then we concatenate these two channel pooled features together to get the spatial attention features $F_i^s \\in \\mathbb{R}^{h_i \\times w_i \\times 2}$ . Afterward, the spatial attention features $F_i^s$ are sent into CNN layers followed by a sigmoid function to obtain the spatial attention map for $F_{vi}$ . The process is as follows:", + "bbox": [ + 89, + 325, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} ^ {s} = \\operatorname {S i g m o i d} \\left(\\operatorname {C o n v} \\left(\\left[ \\operatorname {M a x} \\left(D _ {i}\\right), \\operatorname {A v g} \\left(D _ {i}\\right) \\right]\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 547, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{i}^{s}\\in \\mathbb{R}^{h_{i}\\times w_{i}\\times 1}$ is the spatial attention map for $F_{vi}$ . Then we multiple the visual feature $F_{vi}$ and corresponding spatial attention map as follow:", + "bbox": [ + 89, + 574, + 482, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {F} _ {v i} = F _ {v i} \\cdot A _ {i} ^ {s}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 628, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This results $\\{F_{vi}\\}_{i = 1}^{4}$ are the spatially attended visual features for each scale $i$ .", + "bbox": [ + 89, + 657, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel Attention. The previous work [32] shows that channel attention can help to distinguish different class targets. For the referring expression counting task, it is even harder to distinguish targets due to the contextual attribute understanding requirement. So it is beneficial to make use of channel attention to enhance the distinguishing ability of the REC model. However, the previous channel attention in the CAC task neglects the scale variance due to its relation to fixed-size exemplars. Thanks to the multi-layer CAD features, our channel attention can vary on different scales. Specifically, for the $i_{th}$ scale, we first apply average pooling and max pooling followed by shared MLP on the spatial attention enhanced visual feature $\\dot{F}_{vi}$ along the spatial dimension to get two channel attention features $F_{i}^{c} \\in \\mathbb{R}^{1 \\times C}$ .", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We add two features up, then feed the added feature into the sigmoid function to get the channel attention map $A_{i}^{c}$ :", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} ^ {c} = \\operatorname {S i g m o i d} \\left(M L P \\left(\\operatorname {A v g} \\left(\\dot {F} _ {v i}\\right) + \\operatorname {M a x} \\left(\\dot {F} _ {v i}\\right)\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 132, + 906, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterwards, we multiple the spatial attention enhanced visual feature $\\dot{F}_{vi}$ with channel attention map $A_{i}^{c}$ in every scale $i$ to obtain the final CAD enhanced visual features $\\{\\hat{F}_{vi}\\}_{i=1}^{4}$ as follows:", + "bbox": [ + 511, + 162, + 906, + 223 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {v i} = \\dot {F} _ {v i} \\cdot A _ {i} ^ {c}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 234, + 906, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. CAD Dynamic Query Module", + "text_level": 1, + "bbox": [ + 511, + 263, + 779, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The queries consisting of their locations and corresponding content features are another essential component for the localization decoder. For the initial location of queries, we follow the GroundingDINO strategy of selecting visual feature positions with the top $K$ similarity scores to the text feature. For the corresponding content features $Q \\in \\mathbb{R}^{K \\times C}$ of queries which are learnt embedding in GroundingDINO, we argue that the fixed embedding is not the ideal initialization for query contents in REC. In the REC task, the target with a specific attribute is given, hence the initialized $K$ content is expected to contain as many kinds belonging to this given referring expression as possible.", + "bbox": [ + 511, + 285, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text Init. Therefore, as mentioned above, for the referring expression counting task, since it has only one given target text, we propose to dynamically initialize the query contents with text features first as shown in Fig. 2. In specific, we first multiply the input text feature $F_{t}$ with the learning matrix $M \\in \\mathbb{R}^{C \\times C}$ , then send it into the GELU function to obtain the dynamic text feature $T \\in \\mathbb{R}^{N \\times C}$ , where $N$ is the words number. Afterward, we multiply the dynamic text feature $T$ with init queries $Q \\in \\mathbb{R}^{K \\times C}$ , where $K$ is the query number. We can get the dynamic text weight map $W \\in \\mathbb{R}^{K \\times N}$ . Finally, we multiply the dynamic weight with the text feature $F$ to get the text-dynamic init queries $Q_{t}$ . The process is as follows:", + "bbox": [ + 511, + 467, + 908, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {Q} = \\left(Q \\times \\left(F _ {t} \\times M\\right) ^ {\\top}\\right) \\times F _ {t}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 674, + 906, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to the original queries, the meaning of the text-dynamic init queries obtained in this method is more explicit and can be understood as a combination of different words in the input referring expression.", + "bbox": [ + 511, + 704, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Density Init. While using text combination-based initialization makes it possible for queries to contain different attribute weights, it completely ignores the variance of the given query image. Therefore, we propose to use CAD features which are highly related to the visual features of the given referring expression to further initialize query contents to improve the distinguishing ability. Specifically, we first find $K$ CAD features $D_{K}$ corresponding to the query positions, which has the top $K$ similarity scores to the text", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19590", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6902333488a229a89bb4a751138196b38fb906b309cc66f77223ab56df11ddef.jpg", + "image_caption": [ + "Figure 3. Visualization of queries. We visualize the init queries $Q$ , text dynamic initialized $\\dot{Q}$ and the CAD dynamic initialized queries $\\hat{Q}$ in the polar coordinate system by t-SNE [34]. (b) represent the feature distribution of init queries in the polar coordinate system, which are the same for different referring expressions. (c) represent the feature distribution of text init queries. As shown in (c), the distribution of \"bluish pen\" and \"greenish pen\" is quite similar even with much overlap. (d) represent the feature distribution of text & density init queries. As shown in (d), the distribution of queries is quite different, which means they can be distinguished easily." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 903, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "feature. Then, we input the corresponding CAD feature $D_K$ as key and value and the text-dynamic queries $\\dot{Q}$ as the query for a cross-attention layer to fuse. Finally, we can obtain the CAD dynamic initialized queries $\\hat{Q}$ .", + "bbox": [ + 88, + 484, + 480, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to verify that the CAD Dynamic Query Module is conducive to distinguishing between different referring expressions. We visualize the init queries $Q$ , text dynamic initialized $\\dot{Q}$ and the CAD dynamic initialized queries $\\hat{Q}$ in the polar coordinate system by t-SNE [34]. As shown in Fig. 3, for queries with only text initialization, it is hard to distinguish the bluish pen and greenish pen which are relatively close and similar in the feature space. However, with the help of CAD initialization, the query contents for different input referring expressions are distinguishable.", + "bbox": [ + 88, + 549, + 482, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Objective", + "text_level": 1, + "bbox": [ + 89, + 720, + 264, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our loss function consists of two components, the first one is the localization loss $\\mathcal{L}_{\\mathrm{loc}}$ as previous REC work based on GroundingDINO [8], and the second of which is the density loss $\\mathcal{L}_{\\mathrm{density}}$ . The more detail for $\\mathcal{L}_{\\mathrm{loc}}$ is on the supplementary. Hence the final loss is a weighted sum of two losses:", + "bbox": [ + 89, + 748, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {l o c}} + \\alpha \\cdot \\mathcal {L} _ {\\text {d e n s i t y}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 851, + 482, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a parameter to control the influence of $\\mathcal{L}_{\\mathrm{density}}$ .", + "bbox": [ + 89, + 885, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 483, + 645, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. REC Datasets & Metric", + "text_level": 1, + "bbox": [ + 511, + 515, + 733, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "REC-8k [8]. REC-8K is the first REC task dataset, which contains 8011 different images and 17122 Image-RE pairs. The dataset is split into train, val, test sets which contain 10555, 3336, 3231 Image-RE pairs, 4923, 1566, 1522 images, and 723, 341, 299 REs respectively. 80 REs are shared between train and val sets and 79 REs are shared between train and test sets. The original REC-8K dataset does not contain ground truth contextual attribute density maps, so we generate these maps using a fixed-size Gaussian kernel of 15, following the previous method [19].", + "bbox": [ + 511, + 537, + 906, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Following prior work on object counting, the Mean Absolute Error (MAE) and the Root Mean Squared Error (RMSE) are used to measure counting performance. For the REC-8K dataset, we include Precision, Recall, and F1 scores as metrics for localization as GroundingREC [8].", + "bbox": [ + 511, + 691, + 906, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 786, + 730, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is built on the GroundingDINO [22] where the text and visual backbone, feature enhancer, and localization decoder are from. We provide two versions of our model, trained with Swin-T and Swin-B as visual backbone separately. Please refer to the supplementary for more details about network architectures.", + "bbox": [ + 511, + 809, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19591", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ab730acf73802141c618c45b68df9e648a1de955aa760bff1b49b2529847eb67.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art approaches on the REC-8K dataset. The best performance is boldfaced, and the second is underlined. GroundingREC* is reproduced following the training setting in paper [8] using swin-b as the visual backbone because they do not report the result. Results with † indicate that we use the number of density map estimates to determine the number of queries selected for localization." + ], + "table_footnote": [], + "table_body": "
MethodBackboneVal setTest set
MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑
Mean-14.2827.75---13.7525.91---
ZSC [36]Res-5014.8431.30---14.9329.72---
ZSC [36]Swin-T12.9626.74---13.0029.07---
CounTX [2]ViT-B11.8827.04---11.8425.62---
GroundingDINO [22]Swin-T9.0321.980.560.760.658.8821.950.590.760.66
GroundingREC [8]Swin-T6.8018.130.650.710.686.5019.790.670.720.69
CAD-GD (ours)Swin-T5.4315.010.680.720.705.2917.080.710.730.72
CAD-GD† (ours)Swin-T4.5813.240.680.710.704.5914.680.720.700.71
GroundingREC* [8]Swin-B5.6615.240.660.770.715.4218.470.710.690.70
CAD-GD (ours)Swin-B4.8313.520.740.760.754.9414.650.750.770.76
CAD-GD† (ours)Swin-B4.2313.140.760.700.734.3412.930.770.710.74
", + "bbox": [ + 98, + 155, + 901, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. We freeze the visual and text backbone to keep the feature space aligned. For the REC-8K dataset, we use the AdamW as the optimizer to train the model for 20 epochs. The learning rate is set to $1e - 5$ and decays by $10\\times$ on the $10^{th}$ epoch. For each batch, we send one image with multiple corresponding referring expressions. Besides, data augmentations are not used.", + "bbox": [ + 88, + 400, + 480, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. For the inference stage, we provide two strategies. In the first strategy, we use a fixed threshold following GroundingREC [8]. For each output query, we compare its score with the CLS token by a threshold of 0.25 and scores with the rest of the text tokens by a threshold of 0.35. If all scores are higher than the threshold, we include the query in the final count. Eventually, the total number of queries that satisfy the threshold condition is considered to be the number of targets in the query image. The second strategy is to discard the threshold and use the target number estimated from the density map as the predicted number of outcomes $N$ . Then we find the $N$ queries among all queries that have the highest scores with the CLS token and add them to the predicted outcomes. We use $\\dagger$ to represent the model with the second strategy on Table 1.", + "bbox": [ + 89, + 507, + 482, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State of the Art", + "text_level": 1, + "bbox": [ + 89, + 742, + 383, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. In this section, we compare our CAD-GD with language-based counting methods, open-set detection model GroundingDINO, and state-of-the-art REC model GroundingREC on REC-8K datasets. Most of the results for comparison are obtained from previous paper [8] directly. For the GroundingREC, we train a Swin-B version of GroundingREC, which is marked with * on the Table 1, following the training setting in GroundingREC [8] because it only has the swin-T version model.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, for the Swin-T backbone model, CAD-GD† achieve a relative improvement of $32.6\\%$ w.r.t validation MAE and $29.4\\%$ w.r.t test MAE compared with GroundingDINO. Besides, our method reduces the validation RMSE by $27.0\\%$ and test RMSE by $25.8\\%$ . The reduction in the RMSE reflects the improved robustness of our method for query images of different densities. For the localization metrics, our model improves precision consistently for localization, while the recall between GroundingREC and our model is similar. This phenomenon may be due to that the goal of our approach is to enhance the capture of contextual attributes for referring expression, thus improving the accuracy of localization instead of recall.", + "bbox": [ + 511, + 400, + 906, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also give in the table the performance of the CADGD with Swin-B as the visual backbone. As can be seen from Table 2, our method maintains a steady increase in counting accuracy compared to GroundingREC with Swin-B visual backbone. Meanwhile, as the scale of the visual backbone increases, our method can better improve the localization accuracy compared to GroundingREC. The reason may lie in that the quality of contextual attribute density is highly related to the quality of visual features.", + "bbox": [ + 511, + 598, + 905, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Analysis. For qualitative analysis, we compare our Density-GD with the state-of-the-art REC method GroundingREC in various scenes as shown in Fig. 4. According to the difficulty of the input referring expression, we visualized examples with direct attributes (columns 1-3) and context-related attributes (columns 4-6), respectively. Compared to GroundingREC, even in different density environments, our method can better capture the attribute information of the right target, thus obtaining more accurate counting and localization results. Meanwhile, as shown in columns 4-6, our method can still achieve accurate counting", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19592", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3db9db66617754618344f988c98ea646f6c4d4ee40d301abcbe6861780ed416.jpg", + "image_caption": [ + "Figure 4. Qualitative results on the REC-8k dataset. Columns 1-3 contain examples with direct attributes for REs, and columns 4-6 contain those with context-related attributes. Our method consistently outperforms the GroundingREC with more precious locations." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 903, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results for referring expressions with contextual attributes. From the image corresponding to \"person not in a bus\", we can see that although the previous method can capture the two key information of \"person\" and \"bus\", it can not process the whole semantic features well, and thus wrongly count \"person in a bus\". However, our approach can capture the context-dependent attribute correctly.", + "bbox": [ + 88, + 338, + 480, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of CAD. To demonstrate the relation between the contextual attribute density and the final prediction results, we visualize the contextual attribute density map. As shown in Fig. 5, for the same input query image with different referring expressions, our contextual attribute density map can capture where the corresponding semantic features are located. As shown in the images in (a), the contextual attribute density map can distinguish different attributes on both appearance and spatial aspects. One can observe in (b) that our method can perform well even without any attributes for the target class.", + "bbox": [ + 88, + 444, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 619, + 243, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct the ablation study on REC-8K and provide quantitative results in Table 2. The plain version of CAD-GD is employed as the baseline which only consists of the standard backbone, feature enhancer, and localization decoder as GroundingREC [8].", + "bbox": [ + 89, + 642, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CAD Generate Module Effect. By comparing R1 and R2 in Table 2, one can observe that SDCM can introduce a relative improvement of $5.4\\%$ w.r.t. MAE on the validation set. This performance boost indicates that the counting and localization ability of the model can be improved even by implicitly introducing contextual attribute density features.", + "bbox": [ + 89, + 719, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CAD Attention Module Effect. The comparison of R2 vs. R4 demonstrates that the SDAM improves the validation MAE by 0.51 and the precision by 0.29. This result suggests that the counting accuracy of the model can be improved by introducing CAD to visual features directly through SDAM. CAD Dynamic Query Module Effect. By comparing R4", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on REC-8k dataset. A represents the CAD Generate Module. B represents the CAD Attention Module, where the B1 and B2 represent the spatial attention and channel attention separately. C represents the CAD Dynamic Query Module, where C1 and C2 represent the text and density feature init. D represents the second strategy for inference, which is to use the target number estimated from the density map as the predicted number.", + "bbox": [ + 511, + 337, + 906, + 434 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0ef466b1dd428dac1c8a4bf32a7651742049a96fe1243f46fbe7ac00bd9c2d1f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No.ABCDVal
B1B2C1C2MAERMSEPrecRecall
R16.5217.720.6320.703
R2✓6.1716.380.6410.709
R3✓✓5.8816.430.6620.722
R4✓✓✓5.6116.280.6700.710
R5✓✓✓✓5.6714.430.6780.703
R6✓✓✓✓✓5.4315.010.6830.718
R7✓✓✓✓✓✓4.8313.520.6790.711
", + "bbox": [ + 514, + 445, + 906, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and R6 in Table 2, one can observe that both the MAE and RMSE are improved. However, the comparison of R4 and R5 demonstrates that the validation MAE declines slightly by only adding text init. The reason may lie in that only text init will reduce the perception of the query image.", + "bbox": [ + 511, + 628, + 906, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Inference Strategy. By comparing R6 and R7 in Table 2, it can be observed that the TA can bring a relative improvement of $11.05\\%$ w.r.t MAE and $9.93\\%$ w.r.t RMSE on validation set. However, the performance on localization metrics even declines slightly. The phenomenon may lie in mismatch localization queries and density maps.", + "bbox": [ + 511, + 703, + 906, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Results on Zero-shot Counting", + "text_level": 1, + "bbox": [ + 511, + 803, + 785, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FSC-147. Besides the REC task, we also evaluate our method on class agnostic counting to show our method is generalizable to prior tasks. The quantitative results are presented in Table 3. We evaluate the benchmark dataset FSC-147 [29] in a zero-shot manner, which means the model is", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bc4ca07bd478b362a9d74f496a2811c43df2e7869400b6d252c7c33a829cda4.jpg", + "image_caption": [ + "Figure 5. Visualization of contextual attribute density map. The predicted density map can capture the locations of referring expressions." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/33ba2036e71fd674b26fd041ddb4fe8042dc146c4cca781e92917104074652aa.jpg", + "table_caption": [ + "Table 3. Comparison with the state-of-the-art approaches on the FSC-147 dataset. The upper and lower part of the table presents the results using visual exemplars and class text as guidance separately." + ], + "table_footnote": [], + "table_body": "
MethodValTest
MAERMSEMAERMSE
CounTR [21]13.1349.8311.9591.23
LOCA [33]10.2432.5610.7956.97
CACViT [35]10.6337.959.1348.96
DAVE [27]8.9128.088.6632.36
CountGD [35]7.1026.085.7424.09
Patch-selection [36]26.9388.6322.09115.17
CLIP-count [16]18.7961.1817.78106.62
VLCounter [17]18.0665.1317.05106.16
CounTX [2]17.1065.6115.88106.29
DAVE [27]15.4852.5714.90103.42
GroundingREC [8]10.0658.6210.12107.19
CountGD [3]12.1447.5112.9898.35
CAD-GD (ours)9.3040.9610.3586.88
", + "bbox": [ + 96, + 425, + 475, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trained on the FSC-147 train set and evaluated on the val and test set of novel classes. For training and inference, we use the same training and testing strategy as CountGD [3]. Please refer to the supplementary material for more training and testing details.", + "bbox": [ + 89, + 696, + 482, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3, our method achieves significantly lower counting errors than all prior zero-shot approaches. By comparing with the CAC methods, it can be seen that although our method only uses text as prompts, the counting error is still comparable with the method using extra visual exemplars as guidance.", + "bbox": [ + 89, + 776, + 482, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CARPK. To test cross-dataset generalization, we test our method on the CARPK [11] car counting dataset which is", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/482cd3f81a67f92764f5f848994ebf4d969cdd3a165950acf5d1cb81ce4ec770.jpg", + "table_caption": [ + "Table 4. Comparison with the state-of-the-art Sapproaches on the CARPK dataset. Tuned means the method is finetuned on the CARPK training dataset." + ], + "table_footnote": [], + "table_body": "
MethodPromptTunedTest
MAERMSE
CLIP-count [16]Text×11.9616.61
CounTX [2]Text✓8.1310.87
VLCounter [17]Text×6.468.68
CACViT [35]Visual✓4.916.49
CountGD [3]Text×3.835.41
CountGD [3]Text & Visual×3.685.17
CAD-GD (ours)Text×3.294.56
", + "bbox": [ + 519, + 412, + 898, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only trained on FSC-147 dataset. As shown in Table 4, CAD-GD achieves lower counting errors than other methods even with visual exemplars as guidance.", + "bbox": [ + 511, + 590, + 903, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 659, + 633, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we study why the previous referring expression counting (REC) methods cannot well distinguish between fine-grained attributes for one class. The core issue is that previous methods overlook the contextual attribute density (CAD) conveyed in visual regions. Therefore, we propose the CAD-GD, in which the CAD information is modeled by a CAD estimator and injected into visual features and decoder queries via a well-designed CAD attention module and a CAD dynamic query module. By leveraging the CAD information, consistent performance improvements can be observed on varied benchmarks, which demonstrates the significance of CAD. We hope this work can inspire researchers to explore the CAD in other multimodal tasks.", + "bbox": [ + 509, + 689, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19594", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2", + "[2] Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 6, 8", + "[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. arXiv preprint arXiv:2407.04619, 2024. 2, 8", + "[4] Xinkun Cao, Zhipeng Wang, Yanyun Zhao, and Fei Su. Scale aggregation network for accurate and efficient crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[6] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19638-19648, 2022. 2", + "[7] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 2", + "[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16985-16995, 2024. 1, 2, 5, 6, 7, 8", + "[9] Chengzhen Duan, Zhiwei Wei, Chi Zhang, Siying Qu, and Hongpeng Wang. Coarse-grained density map guided object detection in aerial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2789-2798, 2021. 2", + "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2", + "[11] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In Proceedings of the IEEE international conference on computer vision, pages 4145-4153, 2017. 2, 8", + "[12] Yutao Hu, Xiaolong Jiang, Xuhui Liu, Baochang Zhang, Jungong Han, Xianbin Cao, and David Doermann. Nascount: Counting-by-density with neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 747-766. Springer, 2020. 2", + "[13] Yi-Xin Huang, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Dq-detr: Detr with dynamic query for tiny object detection. arXiv preprint arXiv:2404.03507, 2024. 2", + "[14] Haroon Idrees, Imran Saleemi, Cody Seibert, and Mubarak Shah. Multi-source multi-scale counting in extremely dense" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "crowd images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2547-2554, 2013. 3", + "[15] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 2", + "[16] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 1, 2, 8", + "[17] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2714-2722, 2024. 8", + "[18] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2", + "[19] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 5", + "[20] Changlin Li, Taojiannan Yang, Sijie Zhu, Chen Chen, and Shanyue Guan. Density map guided object detection in aerial images. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 190-191, 2020. 2", + "[21] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. arXiv preprint arXiv:2208.13721, 2022. 8", + "[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1, 2, 3, 4, 5, 6", + "[23] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 2", + "[24] Hao Lu, Zhiguo Cao, Yang Xiao, Bohan Zhuang, and Chunhua Shen. Tasselnet: counting maize tassels in the wild via local counts regression network. Plant methods, 13:1-17, 2017. 2", + "[25] Simon Madec, Xiuliang Jin, Hao Lu, Benoit De Solan, Shouyang Liu, Florent Duyme, Emmanuelle Heritier, and Frederic Baret. Ear density estimation from high resolution rgb imagery using deep learning technique. Agricultural and forest meteorology, 264:225-234, 2019. 2", + "[26] Daniel Onoro-Rubio and Roberto J López-Sastre. Towards perspective-free object counting with deep learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 615-629. Springer, 2016. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Dave-a detect-and-verify paradigm for low-shot counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23293–23302, 2024. 1, 2, 8", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[29] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 1, 2, 7", + "[30] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the\" edge\" of open-set object detection. arXiv preprint arXiv:2405.10300. 2024. 2", + "[31] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 2", + "[32] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9529–9538, 2022. 1, 2, 4", + "[33] Nikola Dukić, Alan Lukežić, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18872-18881, 2023. 1, 2, 8", + "[34] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5", + "[35] Zhicheng Wang, Liwen Xiao, Zhiguo Cao, and Hao Lu. Vision transformer off-the-shelf: A surprising baseline for few-shot class-agnostic counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5832–5840, 2024. 2, 4, 8", + "[36] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 1, 2, 6, 8", + "[37] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 833-841, 2015. 1", + "[38] XQ Zhou, YX Zou, and Y Wang. Accurate small object detection via density map aided saliency estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 425-429. IEEE, 2017. 2" + ], + "bbox": [ + 91, + 92, + 482, + 852 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19596", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_model.json b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_model.json new file mode 100644 index 0000000000000000000000000000000000000000..266582a5b703f222fa5519141563c17c05232d7d --- /dev/null +++ b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_model.json @@ -0,0 +1,1760 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "Exploring Contextual Attribute Density in Referring Expression Counting" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.181, + 0.88, + 0.218 + ], + "angle": 0, + "content": "Zhicheng Wang Zhiyu Pan* Zhan Peng Jian Cheng Liwen Xiao Wei Jiang Zhiguo Cao School of AIA, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.22, + 0.693, + 0.234 + ], + "angle": 0, + "content": "{zhicheng_wang, zhiyupan, zgcao}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.485, + 0.651 + ], + "angle": 0, + "content": "Referring expression counting (REC) algorithms are for more flexible and interactive counting ability across varied fine-grained text expressions. However, the requirement for fine-grained attribute understanding poses challenges for prior arts, as they struggle to accurately align attribute information with correct visual patterns. Given the proven importance of \"visual density\", it is presumed that the limitations of current REC approaches stem from an under-exploration of \"contextual attribute density\" (CAD). In the scope of REC, we define CAD as the measure of the information intensity of one certain fine-grained attribute in visual regions. To model the CAD, we propose a U-shape CAD estimator in which referring expression and multi-scale visual features from GroundingDINO can interact with each other. With additional density supervision, we can effectively encode CAD, which is subsequently decoded via a novel attention procedure with CAD-refined queries. Integrating all these contributions, our framework significantly outperforms state-of-the-art REC methods, achieves \\(30\\%\\) error reduction in counting metrics and a \\(10\\%\\) improvement in localization accuracy. The surprising results shed light on the significance of contextual attribute density for REC. Code will be at github.com/Xu3XiWang/CAD-GD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.222, + 0.698 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Object counting algorithms estimate the number of specific objects in query images. Early research has explored various forms of object descriptions, including class-specific [26, 37], class-agnostic [27, 29, 32, 33], exemplar-guided [29], and text-guided [16, 36] approaches. However, these methods lack the flexibility to distinguish between fine-grained, open-world descriptions, e.g., differentiating a walking person from a person riding motorcycle. To address this, recent work [8] proposes the referring expression counting (REC) task which aims to count fine-grained and contextually nuanced objects within the same class." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.27, + 0.907, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.585, + 0.907, + 0.684 + ], + "angle": 0, + "content": "Figure 1. The efficacy of contextual attribute density (CAD). According to the qualitative comparison above, the REC baseline [8] often over-counts when it focuses primarily on class information rather than fine-grained attributes, and it under-counts in cases of occlusion or scale variations. By modeling CAD, our method more accurately aligns fine-grained attributes with corresponding visual regions, effectively reducing counting errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "GroundingREC [8] is firstly developed as the REC baseline with the power of the open-world detection model GroundingDINO [22]. However, distinguishing between objects of the same class based on nuanced attribute differences is challenging, which makes the REC baseline struggle to align attributes with corresponding visual regions. As shown in Fig. 1, two primary types of failure cases can be observed in the REC baseline: (1) Over-counting: the baseline may place excessive emphasis on class information while neglecting fine-grained attributes, e.g., walking, leading to the inclusion of objects with wrong attributes; (2) Under-counting: the baseline might miss objects with the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19587" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "given attribute when occlusion and scale variations arise. These two failure cases may be due to the baseline's reliance on one-to-one matching between texts and image regions, a typical detection-to-count pipeline. Many previous counting approaches [4, 6, 12, 15] have proved that visual density can provide information about scale-robust spatial distribution. However, in modern open-world perception models [18, 22, 30], the capacity to perceive density for a specific attribute is often overlooked. Inspired by the success of visual density modeling, we propose investigating whether modeling contextual attribute density (CAD) for an open-world perception model could help address the above issues in a cross-modal context." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.668 + ], + "angle": 0, + "content": "In this work, we attempt to make the REC model aware of the CAD. Analogous to the concept of visual density, we define CAD as the information intensity measurement of one certain fine-grained attribute about one object conveyed in query image regions of different scales. Based on our definition, modeling CAD requires both the image and text information, with the image information encoded in a multi-scale format. Hence, we propose a CAD generate module with U-shape CAD estimator (CADE) that encodes the multi-layer CAD features. The CADE is provided with the visual features and similarity features between visual and text to estimates the CAD map for a referring expression, where the intermediate features serve as representations of CAD. After encoding the CAD, the next problem is how to facilitate the counting procedure with the CAD features. We fuse the CAD features with original visual features from GroundingDINO via both spatial and channel attention. In this way, spatial location and channel importance prior are provided for visual features for referring expression separately. Additionally, the counting queries in the localization decoder for counting are also dynamically initialized by both the text features and their corresponding CAD information. With this dynamic initialization module, the feature differentiation of the initial queries under different inputs is enhanced for better counting accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Integrating CAD into our REC framework leads to substantial improvements. Experimental results on the public benchmark REC-8K [8] show that our framework outperforms previous state-of-the-art REC baselines by large margins, achieving \\(37.8\\%\\) and \\(33.2\\%\\) relative error reductions on the validation and test sets, respectively. Besides, We also conduct experiments on the previous class-agnostic counting benchmark FSC-147 [29] with a zero-shot setting, with results further underscoring the effectiveness of our framework. These findings highlight the significance of CAD. Our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We pioneer the concept of contextual attribute density in REC, enabling more precise differentiation between fine-grained descriptions of objects within the same class." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.887, + 0.482, + 0.9 + ], + "angle": 0, + "content": "- We propose the contextual attribute density estimator and" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "an attention procedure with a query enhancement, including the contextual attribute density into the REC pipeline." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.166 + ], + "angle": 0, + "content": "- We achieve significant performance improvements on public benchmarks, demonstrating the value of contextual attribute density modeling for REC." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.181, + 0.655, + 0.196 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.905, + 0.312 + ], + "angle": 0, + "content": "The progression toward more generalized, fine-grained counting tasks has shifted from class-specific to class-agnostic methods, and also from visual to textual guidance. This section reviews prior works contributing to the development of open-world counting models, emphasizing advancements in visual density for counting tasks given our introduction of contextual attribute density." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.313, + 0.905, + 0.689 + ], + "angle": 0, + "content": "Advancing to open-world counting. Counting tasks originally focus on specific categories, such as crowd [1, 4, 15], vehicle [11], wheat head [24, 25], and so on. These early methods are limited to predefined target types and struggle to generalize across diverse object categories. The class-agnostic counting (CAC) [23] marks a significant step toward more flexible counting approaches [27, 32, 33, 35]. Defined as a template-matching problem [23], CAC achieves generalized counting by using visual exemplars to count objects beyond specific classes. Notably, visual density-based methods [32, 33, 35] consistently outperform traditional detection-based counting methods such as MaskRCNN [10] and RetinaNet [31] on counting benchmarks. However, the reliance on visual templates is still a notable limitation in practical applications. To address this issue, textual-based counting [16, 36] is proposed by leveraging the multi-modal ability of CLIP [28]. Although incorporating visual guidance information has been proved beneficial [3], researchers pursue a more generalized setting: with only text expressions, fine-grained, general-purpose objects are expected to be counted, which is referring expression counting (REC) [8]. The REC baseline solely relies on the multi-modal localization ability of GroundingDINO [22], and interpreting some fine-grained expressions accurately remains challenging, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Density modeling in computer vision. Beyond the task of counting, there are some works in other fields using visual density features to provide the spatial and target number prior, e.g., instance segmentation [7] and tiny object detection [9, 13, 20, 38]. Cholakka et al. [7] use density maps to provide prior information on object counts and spatial locations, thereby assisting instance detection tasks under image-level supervision. Huang et al. [13] propose DQ-DETR for tiny object detection, which utilizes the density feature to provide the general target number and spatial prior for DETR [5] structure detection pipeline. In contrast to these methods, our approach focuses more on the role of contextual attribute density for fine-grained expression understanding in a multi-modal setting." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.312, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 2. The framework of Contextual Attribute Density Aware GroundingDINO (CAD-GD). A query image and referring expression are sent into backbones separately together with the feature enhancer to get the visual features and the text feature. Then we obtain the Contextual Attribute Density (CAD) features by the CAD Generate Module supervised by the GT contextual attribute density map. To leverage the CAD information, we design the CAD Attention Module and CAD Dynamic Query Module to enhance visual features and query contents separately. Then we sent the dynamic queries, CAD enhanced visual features, and the text feature into the localization decoder to get the final localization prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.182, + 0.436 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.627 + ], + "angle": 0, + "content": "In this section, we first introduce the overview of our framework for Referring Expression Counting (REC), named Contextual Attribute Density Aware GroundingDINO (CAD-GD). The overview diagram of our approach is shown in Fig. 2. Given an input query image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\) and a referring expression text \\( T = \\{w_i\\}_{i=1}^N \\) with \\( N \\) words, our model predicts a localization counting result. For each \\( (I, T) \\) pair, we first send multi-scale image features and text features extracted from the image and text backbones into feature enhancer [22] to obtain cross-modality visual features \\( \\{F_{vi} \\in \\mathbb{R}^{h_i \\times w_i \\times C}\\}_{i=1}^4 \\) and text feature \\( F_t \\in \\mathbb{R}^{N \\times C} \\) where the \\( C \\) indicates the number of channels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Afterwards, we feed visual features \\(\\{F_{vi}\\}_{i = 1}^{4}\\) and text feature \\(F_{t}\\) into the U-shape contextual attribute density estimator (CADE) to obtain the contextual attribute density (CAD) features \\(\\{D_i\\in \\mathbb{R}^{h_i\\times w_i\\times C}\\}_{i = 1}^4\\) . We then inject the CAD features \\(\\{D_i\\}_{i = 1}^4\\) into visual features \\(\\{F_{vi}\\}_{i = 1}^{4}\\) to enhance the visual feature with spatial and channel attention in CAD Attention Module. Besides, for the initialization of queries in the cross-modality localization decoder [22], we selected the \\(K\\) visual feature positions with the highest scores among them as the initial query locations by calculating the similarity between the visual and text features, while for the initial content features of the query, we dynamically initialize them with text features \\(F_{t}\\) and CAD features \\(\\{D_i\\}_{i = 1}^4\\) corresponding to the query positions. Then we refine the queries by the cross-modality decoder at every layer with CAD enhanced visual and text features. The output queries of the last decoder layer will be used to predict the localization of referring expression objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.422, + 0.744, + 0.436 + ], + "angle": 0, + "content": "3.1. CAD Generation Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.597 + ], + "angle": 0, + "content": "The first step to model CAD for referring expression counting is to obtain the CAD features, which should be responsive to input text with fine-grained attributes. To obtain the satisfying CAD features, we identify several requirements for obtaining satisfying CAD features: (a) the image-text similarity features are required to guide the estimator to focus on areas corresponding to the input referring expression; (b) due to the density diversity of targets in the counting task, a multi-scale structure is of benefit for the robustness of counting [14]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.598, + 0.907, + 0.717 + ], + "angle": 0, + "content": "Motivated by the above requirements, the CAD Generate Module is proposed, as in Fig. 2. Specifically, we first utilize the well-pretrained linear projector from GroundingDINO to project the visual features \\(\\{F_{vi}\\}_{i = 1}^{4}\\) into the text feature space. Then we calculate the similarity feature \\(\\{S_i\\in \\mathbb{R}^{h_i\\times w_i\\times C}\\}_{i = 1}^4\\) between the well-aligned visual feature \\(F_{vi}\\) and corresponding referring expression feature \\(F_{t}\\) as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.722, + 0.905, + 0.739 + ], + "angle": 0, + "content": "\\[\nS _ {i} = \\operatorname {P r o j} \\left(F _ {v i}\\right) \\cdot F _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.841 + ], + "angle": 0, + "content": "where the Proj consists of a single linear layer followed by a layer norm, and the \\(\\cdot\\) represents the element-wise multiplication. Then we send the \\(\\{S_i\\}_{i=1}^4\\) together with the visual feature \\(\\{F_{vi}\\}_{i=1}^4\\) into a U-shape CAD estimator (CADE) to estimate the CAD map for a referring expression, where the intermediate features serve as the CAD features \\(\\{D_i\\}_{i=1}^4\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Supervision Signal. The final output of CADE is the density map for referring expression. In order to guide the CADE to learn the contextual attribute density map, we adopt a conventional \\(\\ell_2\\) loss as the counting loss \\(\\mathcal{L}_{\\mathrm{density}}\\):" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19589" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.157, + 0.105, + 0.483, + 0.124 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e n s i t y}} = \\left\\| D _ {\\text {p r e d}} (I, T) - D _ {\\mathrm {g t}} (I, T) \\right\\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.483, + 0.16 + ], + "angle": 0, + "content": "where \\(D_{\\mathrm{pred}}(I,T)\\) and \\(D_{\\mathrm{gt}}(I,T)\\) denote the prediction and ground truth contextual attribute density map respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.168, + 0.308, + 0.183 + ], + "angle": 0, + "content": "3.2. CAD Attention Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.483, + 0.326 + ], + "angle": 0, + "content": "After obtaining the CAD features \\(\\{D_i\\}_{i=1}^4\\), the next core step is to consider how to inject the CAD information into the localization process, which follows a typically DETR decoder structure [22]. In the DETR-like structure, the visual features \\(\\{F_{vi}\\}_{i=1}^4\\) are crucial for the localization process due to its cross-attention with queries. Motivated by the high relevance between the density feature and the attention map [35], we inject the CAD features into visual features by attention mechanism, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.538 + ], + "angle": 0, + "content": "Spatial Attention. Since the visual features need to be matched with queries to find the target location, foreground enhancement for visual features on areas corresponding to the input referring expression can help to get accurate localization results. Hence we design the spatial attention mechanism to enhance the CAD corresponding foreground areas, utilizing distribution prior from CAD. In specific, for the \\( i_{th} \\) scale of \\( \\{D_i\\}_{i=1}^4 \\), we first apply the average pooling and max pooling for \\( D_i \\) on the channel dimension. Then we concatenate these two channel pooled features together to get the spatial attention features \\( F_i^s \\in \\mathbb{R}^{h_i \\times w_i \\times 2} \\). Afterward, the spatial attention features \\( F_i^s \\) are sent into CNN layers followed by a sigmoid function to obtain the spatial attention map for \\( F_{vi} \\). The process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.548, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nA _ {i} ^ {s} = \\operatorname {S i g m o i d} \\left(\\operatorname {C o n v} \\left(\\left[ \\operatorname {M a x} \\left(D _ {i}\\right), \\operatorname {A v g} \\left(D _ {i}\\right) \\right]\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.622 + ], + "angle": 0, + "content": "where \\(A_{i}^{s}\\in \\mathbb{R}^{h_{i}\\times w_{i}\\times 1}\\) is the spatial attention map for \\(F_{vi}\\). Then we multiple the visual feature \\(F_{vi}\\) and corresponding spatial attention map as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.63, + 0.483, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\dot {F} _ {v i} = F _ {v i} \\cdot A _ {i} ^ {s}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.688 + ], + "angle": 0, + "content": "This results \\(\\{F_{vi}\\}_{i = 1}^{4}\\) are the spatially attended visual features for each scale \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Channel Attention. The previous work [32] shows that channel attention can help to distinguish different class targets. For the referring expression counting task, it is even harder to distinguish targets due to the contextual attribute understanding requirement. So it is beneficial to make use of channel attention to enhance the distinguishing ability of the REC model. However, the previous channel attention in the CAC task neglects the scale variance due to its relation to fixed-size exemplars. Thanks to the multi-layer CAD features, our channel attention can vary on different scales. Specifically, for the \\( i_{th} \\) scale, we first apply average pooling and max pooling followed by shared MLP on the spatial attention enhanced visual feature \\( \\dot{F}_{vi} \\) along the spatial dimension to get two channel attention features \\( F_{i}^{c} \\in \\mathbb{R}^{1 \\times C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "We add two features up, then feed the added feature into the sigmoid function to get the channel attention map \\( A_{i}^{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.133, + 0.907, + 0.151 + ], + "angle": 0, + "content": "\\[\nA _ {i} ^ {c} = \\operatorname {S i g m o i d} \\left(M L P \\left(\\operatorname {A v g} \\left(\\dot {F} _ {v i}\\right) + \\operatorname {M a x} \\left(\\dot {F} _ {v i}\\right)\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.163, + 0.907, + 0.224 + ], + "angle": 0, + "content": "Afterwards, we multiple the spatial attention enhanced visual feature \\(\\dot{F}_{vi}\\) with channel attention map \\(A_{i}^{c}\\) in every scale \\(i\\) to obtain the final CAD enhanced visual features \\(\\{\\hat{F}_{vi}\\}_{i=1}^{4}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.235, + 0.907, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {v i} = \\dot {F} _ {v i} \\cdot A _ {i} ^ {c}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.264, + 0.78, + 0.28 + ], + "angle": 0, + "content": "3.3. CAD Dynamic Query Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.468 + ], + "angle": 0, + "content": "The queries consisting of their locations and corresponding content features are another essential component for the localization decoder. For the initial location of queries, we follow the GroundingDINO strategy of selecting visual feature positions with the top \\( K \\) similarity scores to the text feature. For the corresponding content features \\( Q \\in \\mathbb{R}^{K \\times C} \\) of queries which are learnt embedding in GroundingDINO, we argue that the fixed embedding is not the ideal initialization for query contents in REC. In the REC task, the target with a specific attribute is given, hence the initialized \\( K \\) content is expected to contain as many kinds belonging to this given referring expression as possible." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.468, + 0.909, + 0.664 + ], + "angle": 0, + "content": "Text Init. Therefore, as mentioned above, for the referring expression counting task, since it has only one given target text, we propose to dynamically initialize the query contents with text features first as shown in Fig. 2. In specific, we first multiply the input text feature \\( F_{t} \\) with the learning matrix \\( M \\in \\mathbb{R}^{C \\times C} \\), then send it into the GELU function to obtain the dynamic text feature \\( T \\in \\mathbb{R}^{N \\times C} \\), where \\( N \\) is the words number. Afterward, we multiply the dynamic text feature \\( T \\) with init queries \\( Q \\in \\mathbb{R}^{K \\times C} \\), where \\( K \\) is the query number. We can get the dynamic text weight map \\( W \\in \\mathbb{R}^{K \\times N} \\). Finally, we multiply the dynamic weight with the text feature \\( F \\) to get the text-dynamic init queries \\( Q_{t} \\). The process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.675, + 0.907, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\dot {Q} = \\left(Q \\times \\left(F _ {t} \\times M\\right) ^ {\\top}\\right) \\times F _ {t}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Compared to the original queries, the meaning of the text-dynamic init queries obtained in this method is more explicit and can be understood as a combination of different words in the input referring expression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Density Init. While using text combination-based initialization makes it possible for queries to contain different attribute weights, it completely ignores the variance of the given query image. Therefore, we propose to use CAD features which are highly related to the visual features of the given referring expression to further initialize query contents to improve the distinguishing ability. Specifically, we first find \\( K \\) CAD features \\( D_{K} \\) corresponding to the query positions, which has the top \\( K \\) similarity scores to the text" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19590" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.905, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.389, + 0.908, + 0.463 + ], + "angle": 0, + "content": "Figure 3. Visualization of queries. We visualize the init queries \\( Q \\), text dynamic initialized \\( \\dot{Q} \\) and the CAD dynamic initialized queries \\( \\hat{Q} \\) in the polar coordinate system by t-SNE [34]. (b) represent the feature distribution of init queries in the polar coordinate system, which are the same for different referring expressions. (c) represent the feature distribution of text init queries. As shown in (c), the distribution of \"bluish pen\" and \"greenish pen\" is quite similar even with much overlap. (d) represent the feature distribution of text & density init queries. As shown in (d), the distribution of queries is quite different, which means they can be distinguished easily." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.486, + 0.482, + 0.547 + ], + "angle": 0, + "content": "feature. Then, we input the corresponding CAD feature \\(D_K\\) as key and value and the text-dynamic queries \\(\\dot{Q}\\) as the query for a cross-attention layer to fuse. Finally, we can obtain the CAD dynamic initialized queries \\(\\hat{Q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.483, + 0.703 + ], + "angle": 0, + "content": "In order to verify that the CAD Dynamic Query Module is conducive to distinguishing between different referring expressions. We visualize the init queries \\( Q \\), text dynamic initialized \\( \\dot{Q} \\) and the CAD dynamic initialized queries \\( \\hat{Q} \\) in the polar coordinate system by t-SNE [34]. As shown in Fig. 3, for queries with only text initialization, it is hard to distinguish the bluish pen and greenish pen which are relatively close and similar in the feature space. However, with the help of CAD initialization, the query contents for different input referring expressions are distinguishable." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.722, + 0.265, + 0.74 + ], + "angle": 0, + "content": "3.4. Overall Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Our loss function consists of two components, the first one is the localization loss \\(\\mathcal{L}_{\\mathrm{loc}}\\) as previous REC work based on GroundingDINO [8], and the second of which is the density loss \\(\\mathcal{L}_{\\mathrm{density}}\\). The more detail for \\(\\mathcal{L}_{\\mathrm{loc}}\\) is on the supplementary. Hence the final loss is a weighted sum of two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.852, + 0.483, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {l o c}} + \\alpha \\cdot \\mathcal {L} _ {\\text {d e n s i t y}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.472, + 0.903 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a parameter to control the influence of \\(\\mathcal{L}_{\\mathrm{density}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.484, + 0.646, + 0.502 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.734, + 0.53 + ], + "angle": 0, + "content": "4.1. REC Datasets & Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.691 + ], + "angle": 0, + "content": "REC-8k [8]. REC-8K is the first REC task dataset, which contains 8011 different images and 17122 Image-RE pairs. The dataset is split into train, val, test sets which contain 10555, 3336, 3231 Image-RE pairs, 4923, 1566, 1522 images, and 723, 341, 299 REs respectively. 80 REs are shared between train and val sets and 79 REs are shared between train and test sets. The original REC-8K dataset does not contain ground truth contextual attribute density maps, so we generate these maps using a fixed-size Gaussian kernel of 15, following the previous method [19]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.692, + 0.907, + 0.768 + ], + "angle": 0, + "content": "Metrics. Following prior work on object counting, the Mean Absolute Error (MAE) and the Root Mean Squared Error (RMSE) are used to measure counting performance. For the REC-8K dataset, we include Precision, Recall, and F1 scores as metrics for localization as GroundingREC [8]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.731, + 0.803 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Our model is built on the GroundingDINO [22] where the text and visual backbone, feature enhancer, and localization decoder are from. We provide two versions of our model, trained with Swin-T and Swin-B as visual backbone separately. Please refer to the supplementary for more details about network architectures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19591" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.145 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art approaches on the REC-8K dataset. The best performance is boldfaced, and the second is underlined. GroundingREC* is reproduced following the training setting in paper [8] using swin-b as the visual backbone because they do not report the result. Results with † indicate that we use the number of density map estimates to determine the number of queries selected for localization." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.156, + 0.903, + 0.377 + ], + "angle": 0, + "content": "
MethodBackboneVal setTest set
MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑
Mean-14.2827.75---13.7525.91---
ZSC [36]Res-5014.8431.30---14.9329.72---
ZSC [36]Swin-T12.9626.74---13.0029.07---
CounTX [2]ViT-B11.8827.04---11.8425.62---
GroundingDINO [22]Swin-T9.0321.980.560.760.658.8821.950.590.760.66
GroundingREC [8]Swin-T6.8018.130.650.710.686.5019.790.670.720.69
CAD-GD (ours)Swin-T5.4315.010.680.720.705.2917.080.710.730.72
CAD-GD† (ours)Swin-T4.5813.240.680.710.704.5914.680.720.700.71
GroundingREC* [8]Swin-B5.6615.240.660.770.715.4218.470.710.690.70
CAD-GD (ours)Swin-B4.8313.520.740.760.754.9414.650.750.770.76
CAD-GD† (ours)Swin-B4.2313.140.760.700.734.3412.930.770.710.74
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.401, + 0.482, + 0.506 + ], + "angle": 0, + "content": "Training. We freeze the visual and text backbone to keep the feature space aligned. For the REC-8K dataset, we use the AdamW as the optimizer to train the model for 20 epochs. The learning rate is set to \\(1e - 5\\) and decays by \\(10\\times\\) on the \\(10^{th}\\) epoch. For each batch, we send one image with multiple corresponding referring expressions. Besides, data augmentations are not used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Inference. For the inference stage, we provide two strategies. In the first strategy, we use a fixed threshold following GroundingREC [8]. For each output query, we compare its score with the CLS token by a threshold of 0.25 and scores with the rest of the text tokens by a threshold of 0.35. If all scores are higher than the threshold, we include the query in the final count. Eventually, the total number of queries that satisfy the threshold condition is considered to be the number of targets in the query image. The second strategy is to discard the threshold and use the target number estimated from the density map as the predicted number of outcomes \\( N \\). Then we find the \\( N \\) queries among all queries that have the highest scores with the CLS token and add them to the predicted outcomes. We use \\( \\dagger \\) to represent the model with the second strategy on Table 1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.384, + 0.759 + ], + "angle": 0, + "content": "4.3. Comparison with State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. In this section, we compare our CAD-GD with language-based counting methods, open-set detection model GroundingDINO, and state-of-the-art REC model GroundingREC on REC-8K datasets. Most of the results for comparison are obtained from previous paper [8] directly. For the GroundingREC, we train a Swin-B version of GroundingREC, which is marked with * on the Table 1, following the training setting in GroundingREC [8] because it only has the swin-T version model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.598 + ], + "angle": 0, + "content": "As shown in Table 1, for the Swin-T backbone model, CAD-GD† achieve a relative improvement of \\(32.6\\%\\) w.r.t validation MAE and \\(29.4\\%\\) w.r.t test MAE compared with GroundingDINO. Besides, our method reduces the validation RMSE by \\(27.0\\%\\) and test RMSE by \\(25.8\\%\\). The reduction in the RMSE reflects the improved robustness of our method for query images of different densities. For the localization metrics, our model improves precision consistently for localization, while the recall between GroundingREC and our model is similar. This phenomenon may be due to that the goal of our approach is to enhance the capture of contextual attributes for referring expression, thus improving the accuracy of localization instead of recall." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.735 + ], + "angle": 0, + "content": "We also give in the table the performance of the CADGD with Swin-B as the visual backbone. As can be seen from Table 2, our method maintains a steady increase in counting accuracy compared to GroundingREC with Swin-B visual backbone. Meanwhile, as the scale of the visual backbone increases, our method can better improve the localization accuracy compared to GroundingREC. The reason may lie in that the quality of contextual attribute density is highly related to the quality of visual features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Qualitative Analysis. For qualitative analysis, we compare our Density-GD with the state-of-the-art REC method GroundingREC in various scenes as shown in Fig. 4. According to the difficulty of the input referring expression, we visualized examples with direct attributes (columns 1-3) and context-related attributes (columns 4-6), respectively. Compared to GroundingREC, even in different density environments, our method can better capture the attribute information of the right target, thus obtaining more accurate counting and localization results. Meanwhile, as shown in columns 4-6, our method can still achieve accurate counting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.904, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.284, + 0.908, + 0.315 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on the REC-8k dataset. Columns 1-3 contain examples with direct attributes for REs, and columns 4-6 contain those with context-related attributes. Our method consistently outperforms the GroundingREC with more precious locations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.482, + 0.445 + ], + "angle": 0, + "content": "results for referring expressions with contextual attributes. From the image corresponding to \"person not in a bus\", we can see that although the previous method can capture the two key information of \"person\" and \"bus\", it can not process the whole semantic features well, and thus wrongly count \"person in a bus\". However, our approach can capture the context-dependent attribute correctly." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Effect of CAD. To demonstrate the relation between the contextual attribute density and the final prediction results, we visualize the contextual attribute density map. As shown in Fig. 5, for the same input query image with different referring expressions, our contextual attribute density map can capture where the corresponding semantic features are located. As shown in the images in (a), the contextual attribute density map can distinguish different attributes on both appearance and spatial aspects. One can observe in (b) that our method can perform well even without any attributes for the target class." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.245, + 0.637 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.719 + ], + "angle": 0, + "content": "We conduct the ablation study on REC-8K and provide quantitative results in Table 2. The plain version of CAD-GD is employed as the baseline which only consists of the standard backbone, feature enhancer, and localization decoder as GroundingREC [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.81 + ], + "angle": 0, + "content": "CAD Generate Module Effect. By comparing R1 and R2 in Table 2, one can observe that SDCM can introduce a relative improvement of \\(5.4\\%\\) w.r.t. MAE on the validation set. This performance boost indicates that the counting and localization ability of the model can be improved even by implicitly introducing contextual attribute density features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "CAD Attention Module Effect. The comparison of R2 vs. R4 demonstrates that the SDAM improves the validation MAE by 0.51 and the precision by 0.29. This result suggests that the counting accuracy of the model can be improved by introducing CAD to visual features directly through SDAM. CAD Dynamic Query Module Effect. By comparing R4" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Table 2. Ablation study on REC-8k dataset. A represents the CAD Generate Module. B represents the CAD Attention Module, where the B1 and B2 represent the spatial attention and channel attention separately. C represents the CAD Dynamic Query Module, where C1 and C2 represent the text and density feature init. D represents the second strategy for inference, which is to use the target number estimated from the density map as the predicted number." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.446, + 0.907, + 0.606 + ], + "angle": 0, + "content": "
No.ABCDVal
B1B2C1C2MAERMSEPrecRecall
R16.5217.720.6320.703
R2✓6.1716.380.6410.709
R3✓✓5.8816.430.6620.722
R4✓✓✓5.6116.280.6700.710
R5✓✓✓✓5.6714.430.6780.703
R6✓✓✓✓✓5.4315.010.6830.718
R7✓✓✓✓✓✓4.8313.520.6790.711
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.704 + ], + "angle": 0, + "content": "and R6 in Table 2, one can observe that both the MAE and RMSE are improved. However, the comparison of R4 and R5 demonstrates that the validation MAE declines slightly by only adding text init. The reason may lie in that only text init will reduce the perception of the query image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Impact of Inference Strategy. By comparing R6 and R7 in Table 2, it can be observed that the TA can bring a relative improvement of \\(11.05\\%\\) w.r.t MAE and \\(9.93\\%\\) w.r.t RMSE on validation set. However, the performance on localization metrics even declines slightly. The phenomenon may lie in mismatch localization queries and density maps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.787, + 0.82 + ], + "angle": 0, + "content": "4.5. Results on Zero-shot Counting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "FSC-147. Besides the REC task, we also evaluate our method on class agnostic counting to show our method is generalizable to prior tasks. The quantitative results are presented in Table 3. We evaluate the benchmark dataset FSC-147 [29] in a zero-shot manner, which means the model is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19593" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.905, + 0.335 + ], + "angle": 0, + "content": "Figure 5. Visualization of contextual attribute density map. The predicted density map can capture the locations of referring expressions." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.359, + 0.483, + 0.415 + ], + "angle": 0, + "content": "Table 3. Comparison with the state-of-the-art approaches on the FSC-147 dataset. The upper and lower part of the table presents the results using visual exemplars and class text as guidance separately." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.426, + 0.477, + 0.665 + ], + "angle": 0, + "content": "
MethodValTest
MAERMSEMAERMSE
CounTR [21]13.1349.8311.9591.23
LOCA [33]10.2432.5610.7956.97
CACViT [35]10.6337.959.1348.96
DAVE [27]8.9128.088.6632.36
CountGD [35]7.1026.085.7424.09
Patch-selection [36]26.9388.6322.09115.17
CLIP-count [16]18.7961.1817.78106.62
VLCounter [17]18.0665.1317.05106.16
CounTX [2]17.1065.6115.88106.29
DAVE [27]15.4852.5714.90103.42
GroundingREC [8]10.0658.6210.12107.19
CountGD [3]12.1447.5112.9898.35
CAD-GD (ours)9.3040.9610.3586.88
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.773 + ], + "angle": 0, + "content": "trained on the FSC-147 train set and evaluated on the val and test set of novel classes. For training and inference, we use the same training and testing strategy as CountGD [3]. Please refer to the supplementary material for more training and testing details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.483, + 0.867 + ], + "angle": 0, + "content": "As shown in Table 3, our method achieves significantly lower counting errors than all prior zero-shot approaches. By comparing with the CAC methods, it can be seen that although our method only uses text as prompts, the counting error is still comparable with the method using extra visual exemplars as guidance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "CARPK. To test cross-dataset generalization, we test our method on the CARPK [11] car counting dataset which is" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.359, + 0.905, + 0.401 + ], + "angle": 0, + "content": "Table 4. Comparison with the state-of-the-art Sapproaches on the CARPK dataset. Tuned means the method is finetuned on the CARPK training dataset." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.413, + 0.9, + 0.558 + ], + "angle": 0, + "content": "
MethodPromptTunedTest
MAERMSE
CLIP-count [16]Text×11.9616.61
CounTX [2]Text✓8.1310.87
VLCounter [17]Text×6.468.68
CACViT [35]Visual✓4.916.49
CountGD [3]Text×3.835.41
CountGD [3]Text & Visual×3.685.17
CAD-GD (ours)Text×3.294.56
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.905, + 0.637 + ], + "angle": 0, + "content": "only trained on FSC-147 dataset. As shown in Table 4, CAD-GD achieves lower counting errors than other methods even with visual exemplars as guidance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.66, + 0.634, + 0.675 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.69, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this work, we study why the previous referring expression counting (REC) methods cannot well distinguish between fine-grained attributes for one class. The core issue is that previous methods overlook the contextual attribute density (CAD) conveyed in visual regions. Therefore, we propose the CAD-GD, in which the CAD information is modeled by a CAD estimator and injected into visual features and decoder queries via a well-designed CAD attention module and a CAD dynamic query module. By leveraging the CAD information, consistent performance improvements can be observed on varied benchmarks, which demonstrates the significance of CAD. We hope this work can inspire researchers to explore the CAD in other multimodal tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19594" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. arXiv preprint arXiv:2407.04619, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.312 + ], + "angle": 0, + "content": "[4] Xinkun Cao, Zhipeng Wang, Yanyun Zhao, and Fei Su. Scale aggregation network for accurate and efficient crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[6] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19638-19648, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[7] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16985-16995, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[9] Chengzhen Duan, Zhiwei Wei, Chi Zhang, Siying Qu, and Hongpeng Wang. Coarse-grained density map guided object detection in aerial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2789-2798, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In Proceedings of the IEEE international conference on computer vision, pages 4145-4153, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[12] Yutao Hu, Xiaolong Jiang, Xuhui Liu, Baochang Zhang, Jungong Han, Xianbin Cao, and David Doermann. Nascount: Counting-by-density with neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 747-766. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[13] Yi-Xin Huang, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Dq-detr: Detr with dynamic query for tiny object detection. arXiv preprint arXiv:2404.03507, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Haroon Idrees, Imran Saleemi, Cody Seibert, and Mubarak Shah. Multi-source multi-scale counting in extremely dense" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "crowd images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2547-2554, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[15] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[16] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[17] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2714-2722, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[18] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[19] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[20] Changlin Li, Taojiannan Yang, Sijie Zhu, Chen Chen, and Shanyue Guan. Density map guided object detection in aerial images. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 190-191, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.561 + ], + "angle": 0, + "content": "[21] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. arXiv preprint arXiv:2208.13721, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[23] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[24] Hao Lu, Zhiguo Cao, Yang Xiao, Bohan Zhuang, and Chunhua Shen. Tasselnet: counting maize tassels in the wild via local counts regression network. Plant methods, 13:1-17, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[25] Simon Madec, Xiuliang Jin, Hao Lu, Benoit De Solan, Shouyang Liu, Florent Duyme, Emmanuelle Heritier, and Frederic Baret. Ear density estimation from high resolution rgb imagery using deep learning technique. Agricultural and forest meteorology, 264:225-234, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Daniel Onoro-Rubio and Roberto J López-Sastre. Towards perspective-free object counting with deep learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 615-629. Springer, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19595" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Dave-a detect-and-verify paradigm for low-shot counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23293–23302, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[29] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[30] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the\" edge\" of open-set object detection. arXiv preprint arXiv:2405.10300. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[31] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[32] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9529–9538, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[33] Nikola Dukić, Alan Lukežić, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18872-18881, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[34] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[35] Zhicheng Wang, Liwen Xiao, Zhiguo Cao, and Hao Lu. Vision transformer off-the-shelf: A surprising baseline for few-shot class-agnostic counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5832–5840, 2024. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[36] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.795 + ], + "angle": 0, + "content": "[37] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 833-841, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.799, + 0.482, + 0.853 + ], + "angle": 0, + "content": "[38] XQ Zhou, YX Zou, and Y Wang. Accurate small object detection via density map aided saliency estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 425-429. IEEE, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19596" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_origin.pdf b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7f69b02b804bb27d87eb3805ccf835ba6e66fa72 --- /dev/null +++ b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/f78a8ba2-f99c-43c8-aadb-f099a2092f51_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9a57b802052e87c4165845c8746cd070ab7cd29a1c16492e7377829c8aa1598 +size 4821093 diff --git a/2025/Exploring Contextual Attribute Density in Referring Expression Counting/full.md b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a7d2ab7372b28d8950cb4f8bc1c99e8d61569023 --- /dev/null +++ b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/full.md @@ -0,0 +1,255 @@ +# Exploring Contextual Attribute Density in Referring Expression Counting + +Zhicheng Wang Zhiyu Pan* Zhan Peng Jian Cheng Liwen Xiao Wei Jiang Zhiguo Cao School of AIA, Huazhong University of Science and Technology + +{zhicheng_wang, zhiyupan, zgcao}@hust.edu.cn + +# Abstract + +Referring expression counting (REC) algorithms are for more flexible and interactive counting ability across varied fine-grained text expressions. However, the requirement for fine-grained attribute understanding poses challenges for prior arts, as they struggle to accurately align attribute information with correct visual patterns. Given the proven importance of "visual density", it is presumed that the limitations of current REC approaches stem from an under-exploration of "contextual attribute density" (CAD). In the scope of REC, we define CAD as the measure of the information intensity of one certain fine-grained attribute in visual regions. To model the CAD, we propose a U-shape CAD estimator in which referring expression and multi-scale visual features from GroundingDINO can interact with each other. With additional density supervision, we can effectively encode CAD, which is subsequently decoded via a novel attention procedure with CAD-refined queries. Integrating all these contributions, our framework significantly outperforms state-of-the-art REC methods, achieves $30\%$ error reduction in counting metrics and a $10\%$ improvement in localization accuracy. The surprising results shed light on the significance of contextual attribute density for REC. Code will be at github.com/Xu3XiWang/CAD-GD. + +# 1. Introduction + +Object counting algorithms estimate the number of specific objects in query images. Early research has explored various forms of object descriptions, including class-specific [26, 37], class-agnostic [27, 29, 32, 33], exemplar-guided [29], and text-guided [16, 36] approaches. However, these methods lack the flexibility to distinguish between fine-grained, open-world descriptions, e.g., differentiating a walking person from a person riding motorcycle. To address this, recent work [8] proposes the referring expression counting (REC) task which aims to count fine-grained and contextually nuanced objects within the same class. + +![](images/1dba6845c6edfd7cb3e9aff6d3eba562bb6085dd966702fe9e789ce9ad1f1993.jpg) +Figure 1. The efficacy of contextual attribute density (CAD). According to the qualitative comparison above, the REC baseline [8] often over-counts when it focuses primarily on class information rather than fine-grained attributes, and it under-counts in cases of occlusion or scale variations. By modeling CAD, our method more accurately aligns fine-grained attributes with corresponding visual regions, effectively reducing counting errors. + +GroundingREC [8] is firstly developed as the REC baseline with the power of the open-world detection model GroundingDINO [22]. However, distinguishing between objects of the same class based on nuanced attribute differences is challenging, which makes the REC baseline struggle to align attributes with corresponding visual regions. As shown in Fig. 1, two primary types of failure cases can be observed in the REC baseline: (1) Over-counting: the baseline may place excessive emphasis on class information while neglecting fine-grained attributes, e.g., walking, leading to the inclusion of objects with wrong attributes; (2) Under-counting: the baseline might miss objects with the + +given attribute when occlusion and scale variations arise. These two failure cases may be due to the baseline's reliance on one-to-one matching between texts and image regions, a typical detection-to-count pipeline. Many previous counting approaches [4, 6, 12, 15] have proved that visual density can provide information about scale-robust spatial distribution. However, in modern open-world perception models [18, 22, 30], the capacity to perceive density for a specific attribute is often overlooked. Inspired by the success of visual density modeling, we propose investigating whether modeling contextual attribute density (CAD) for an open-world perception model could help address the above issues in a cross-modal context. + +In this work, we attempt to make the REC model aware of the CAD. Analogous to the concept of visual density, we define CAD as the information intensity measurement of one certain fine-grained attribute about one object conveyed in query image regions of different scales. Based on our definition, modeling CAD requires both the image and text information, with the image information encoded in a multi-scale format. Hence, we propose a CAD generate module with U-shape CAD estimator (CADE) that encodes the multi-layer CAD features. The CADE is provided with the visual features and similarity features between visual and text to estimates the CAD map for a referring expression, where the intermediate features serve as representations of CAD. After encoding the CAD, the next problem is how to facilitate the counting procedure with the CAD features. We fuse the CAD features with original visual features from GroundingDINO via both spatial and channel attention. In this way, spatial location and channel importance prior are provided for visual features for referring expression separately. Additionally, the counting queries in the localization decoder for counting are also dynamically initialized by both the text features and their corresponding CAD information. With this dynamic initialization module, the feature differentiation of the initial queries under different inputs is enhanced for better counting accuracy. + +Integrating CAD into our REC framework leads to substantial improvements. Experimental results on the public benchmark REC-8K [8] show that our framework outperforms previous state-of-the-art REC baselines by large margins, achieving $37.8\%$ and $33.2\%$ relative error reductions on the validation and test sets, respectively. Besides, We also conduct experiments on the previous class-agnostic counting benchmark FSC-147 [29] with a zero-shot setting, with results further underscoring the effectiveness of our framework. These findings highlight the significance of CAD. Our contributions are three-fold: + +- We pioneer the concept of contextual attribute density in REC, enabling more precise differentiation between fine-grained descriptions of objects within the same class. +- We propose the contextual attribute density estimator and + +an attention procedure with a query enhancement, including the contextual attribute density into the REC pipeline. + +- We achieve significant performance improvements on public benchmarks, demonstrating the value of contextual attribute density modeling for REC. + +# 2. Related Work + +The progression toward more generalized, fine-grained counting tasks has shifted from class-specific to class-agnostic methods, and also from visual to textual guidance. This section reviews prior works contributing to the development of open-world counting models, emphasizing advancements in visual density for counting tasks given our introduction of contextual attribute density. + +Advancing to open-world counting. Counting tasks originally focus on specific categories, such as crowd [1, 4, 15], vehicle [11], wheat head [24, 25], and so on. These early methods are limited to predefined target types and struggle to generalize across diverse object categories. The class-agnostic counting (CAC) [23] marks a significant step toward more flexible counting approaches [27, 32, 33, 35]. Defined as a template-matching problem [23], CAC achieves generalized counting by using visual exemplars to count objects beyond specific classes. Notably, visual density-based methods [32, 33, 35] consistently outperform traditional detection-based counting methods such as MaskRCNN [10] and RetinaNet [31] on counting benchmarks. However, the reliance on visual templates is still a notable limitation in practical applications. To address this issue, textual-based counting [16, 36] is proposed by leveraging the multi-modal ability of CLIP [28]. Although incorporating visual guidance information has been proved beneficial [3], researchers pursue a more generalized setting: with only text expressions, fine-grained, general-purpose objects are expected to be counted, which is referring expression counting (REC) [8]. The REC baseline solely relies on the multi-modal localization ability of GroundingDINO [22], and interpreting some fine-grained expressions accurately remains challenging, as shown in Fig. 1. + +Density modeling in computer vision. Beyond the task of counting, there are some works in other fields using visual density features to provide the spatial and target number prior, e.g., instance segmentation [7] and tiny object detection [9, 13, 20, 38]. Cholakka et al. [7] use density maps to provide prior information on object counts and spatial locations, thereby assisting instance detection tasks under image-level supervision. Huang et al. [13] propose DQ-DETR for tiny object detection, which utilizes the density feature to provide the general target number and spatial prior for DETR [5] structure detection pipeline. In contrast to these methods, our approach focuses more on the role of contextual attribute density for fine-grained expression understanding in a multi-modal setting. + +![](images/9d1a3e0b8b14e70ebc548d9f6b9d86b554208a0897259ace4a33371bb8542176.jpg) +Figure 2. The framework of Contextual Attribute Density Aware GroundingDINO (CAD-GD). A query image and referring expression are sent into backbones separately together with the feature enhancer to get the visual features and the text feature. Then we obtain the Contextual Attribute Density (CAD) features by the CAD Generate Module supervised by the GT contextual attribute density map. To leverage the CAD information, we design the CAD Attention Module and CAD Dynamic Query Module to enhance visual features and query contents separately. Then we sent the dynamic queries, CAD enhanced visual features, and the text feature into the localization decoder to get the final localization prediction. + +# 3. Method + +In this section, we first introduce the overview of our framework for Referring Expression Counting (REC), named Contextual Attribute Density Aware GroundingDINO (CAD-GD). The overview diagram of our approach is shown in Fig. 2. Given an input query image $I \in \mathbb{R}^{H \times W \times 3}$ and a referring expression text $T = \{w_i\}_{i=1}^N$ with $N$ words, our model predicts a localization counting result. For each $(I, T)$ pair, we first send multi-scale image features and text features extracted from the image and text backbones into feature enhancer [22] to obtain cross-modality visual features $\{F_{vi} \in \mathbb{R}^{h_i \times w_i \times C}\}_{i=1}^4$ and text feature $F_t \in \mathbb{R}^{N \times C}$ where the $C$ indicates the number of channels. + +Afterwards, we feed visual features $\{F_{vi}\}_{i = 1}^{4}$ and text feature $F_{t}$ into the U-shape contextual attribute density estimator (CADE) to obtain the contextual attribute density (CAD) features $\{D_i\in \mathbb{R}^{h_i\times w_i\times C}\}_{i = 1}^4$ . We then inject the CAD features $\{D_i\}_{i = 1}^4$ into visual features $\{F_{vi}\}_{i = 1}^{4}$ to enhance the visual feature with spatial and channel attention in CAD Attention Module. Besides, for the initialization of queries in the cross-modality localization decoder [22], we selected the $K$ visual feature positions with the highest scores among them as the initial query locations by calculating the similarity between the visual and text features, while for the initial content features of the query, we dynamically initialize them with text features $F_{t}$ and CAD features $\{D_i\}_{i = 1}^4$ corresponding to the query positions. Then we refine the queries by the cross-modality decoder at every layer with CAD enhanced visual and text features. The output queries of the last decoder layer will be used to predict the localization of referring expression objects. + +# 3.1. CAD Generation Module + +The first step to model CAD for referring expression counting is to obtain the CAD features, which should be responsive to input text with fine-grained attributes. To obtain the satisfying CAD features, we identify several requirements for obtaining satisfying CAD features: (a) the image-text similarity features are required to guide the estimator to focus on areas corresponding to the input referring expression; (b) due to the density diversity of targets in the counting task, a multi-scale structure is of benefit for the robustness of counting [14]. + +Motivated by the above requirements, the CAD Generate Module is proposed, as in Fig. 2. Specifically, we first utilize the well-pretrained linear projector from GroundingDINO to project the visual features $\{F_{vi}\}_{i = 1}^{4}$ into the text feature space. Then we calculate the similarity feature $\{S_i\in \mathbb{R}^{h_i\times w_i\times C}\}_{i = 1}^4$ between the well-aligned visual feature $F_{vi}$ and corresponding referring expression feature $F_{t}$ as follow: + +$$ +S _ {i} = \operatorname {P r o j} \left(F _ {v i}\right) \cdot F _ {t}, \tag {1} +$$ + +where the Proj consists of a single linear layer followed by a layer norm, and the $\cdot$ represents the element-wise multiplication. Then we send the $\{S_i\}_{i=1}^4$ together with the visual feature $\{F_{vi}\}_{i=1}^4$ into a U-shape CAD estimator (CADE) to estimate the CAD map for a referring expression, where the intermediate features serve as the CAD features $\{D_i\}_{i=1}^4$ . + +Supervision Signal. The final output of CADE is the density map for referring expression. In order to guide the CADE to learn the contextual attribute density map, we adopt a conventional $\ell_2$ loss as the counting loss $\mathcal{L}_{\mathrm{density}}$ : + +$$ +\mathcal {L} _ {\text {d e n s i t y}} = \left\| D _ {\text {p r e d}} (I, T) - D _ {\mathrm {g t}} (I, T) \right\| _ {2} ^ {2}, \tag {2} +$$ + +where $D_{\mathrm{pred}}(I,T)$ and $D_{\mathrm{gt}}(I,T)$ denote the prediction and ground truth contextual attribute density map respectively. + +# 3.2. CAD Attention Module + +After obtaining the CAD features $\{D_i\}_{i=1}^4$ , the next core step is to consider how to inject the CAD information into the localization process, which follows a typically DETR decoder structure [22]. In the DETR-like structure, the visual features $\{F_{vi}\}_{i=1}^4$ are crucial for the localization process due to its cross-attention with queries. Motivated by the high relevance between the density feature and the attention map [35], we inject the CAD features into visual features by attention mechanism, as shown in Fig. 2. + +Spatial Attention. Since the visual features need to be matched with queries to find the target location, foreground enhancement for visual features on areas corresponding to the input referring expression can help to get accurate localization results. Hence we design the spatial attention mechanism to enhance the CAD corresponding foreground areas, utilizing distribution prior from CAD. In specific, for the $i_{th}$ scale of $\{D_i\}_{i=1}^4$ , we first apply the average pooling and max pooling for $D_i$ on the channel dimension. Then we concatenate these two channel pooled features together to get the spatial attention features $F_i^s \in \mathbb{R}^{h_i \times w_i \times 2}$ . Afterward, the spatial attention features $F_i^s$ are sent into CNN layers followed by a sigmoid function to obtain the spatial attention map for $F_{vi}$ . The process is as follows: + +$$ +A _ {i} ^ {s} = \operatorname {S i g m o i d} \left(\operatorname {C o n v} \left(\left[ \operatorname {M a x} \left(D _ {i}\right), \operatorname {A v g} \left(D _ {i}\right) \right]\right)\right), \tag {3} +$$ + +where $A_{i}^{s}\in \mathbb{R}^{h_{i}\times w_{i}\times 1}$ is the spatial attention map for $F_{vi}$ . Then we multiple the visual feature $F_{vi}$ and corresponding spatial attention map as follow: + +$$ +\dot {F} _ {v i} = F _ {v i} \cdot A _ {i} ^ {s}. \tag {4} +$$ + +This results $\{F_{vi}\}_{i = 1}^{4}$ are the spatially attended visual features for each scale $i$ . + +Channel Attention. The previous work [32] shows that channel attention can help to distinguish different class targets. For the referring expression counting task, it is even harder to distinguish targets due to the contextual attribute understanding requirement. So it is beneficial to make use of channel attention to enhance the distinguishing ability of the REC model. However, the previous channel attention in the CAC task neglects the scale variance due to its relation to fixed-size exemplars. Thanks to the multi-layer CAD features, our channel attention can vary on different scales. Specifically, for the $i_{th}$ scale, we first apply average pooling and max pooling followed by shared MLP on the spatial attention enhanced visual feature $\dot{F}_{vi}$ along the spatial dimension to get two channel attention features $F_{i}^{c} \in \mathbb{R}^{1 \times C}$ . + +We add two features up, then feed the added feature into the sigmoid function to get the channel attention map $A_{i}^{c}$ : + +$$ +A _ {i} ^ {c} = \operatorname {S i g m o i d} \left(M L P \left(\operatorname {A v g} \left(\dot {F} _ {v i}\right) + \operatorname {M a x} \left(\dot {F} _ {v i}\right)\right)\right). \tag {5} +$$ + +Afterwards, we multiple the spatial attention enhanced visual feature $\dot{F}_{vi}$ with channel attention map $A_{i}^{c}$ in every scale $i$ to obtain the final CAD enhanced visual features $\{\hat{F}_{vi}\}_{i=1}^{4}$ as follows: + +$$ +\hat {F} _ {v i} = \dot {F} _ {v i} \cdot A _ {i} ^ {c}. \tag {6} +$$ + +# 3.3. CAD Dynamic Query Module + +The queries consisting of their locations and corresponding content features are another essential component for the localization decoder. For the initial location of queries, we follow the GroundingDINO strategy of selecting visual feature positions with the top $K$ similarity scores to the text feature. For the corresponding content features $Q \in \mathbb{R}^{K \times C}$ of queries which are learnt embedding in GroundingDINO, we argue that the fixed embedding is not the ideal initialization for query contents in REC. In the REC task, the target with a specific attribute is given, hence the initialized $K$ content is expected to contain as many kinds belonging to this given referring expression as possible. + +Text Init. Therefore, as mentioned above, for the referring expression counting task, since it has only one given target text, we propose to dynamically initialize the query contents with text features first as shown in Fig. 2. In specific, we first multiply the input text feature $F_{t}$ with the learning matrix $M \in \mathbb{R}^{C \times C}$ , then send it into the GELU function to obtain the dynamic text feature $T \in \mathbb{R}^{N \times C}$ , where $N$ is the words number. Afterward, we multiply the dynamic text feature $T$ with init queries $Q \in \mathbb{R}^{K \times C}$ , where $K$ is the query number. We can get the dynamic text weight map $W \in \mathbb{R}^{K \times N}$ . Finally, we multiply the dynamic weight with the text feature $F$ to get the text-dynamic init queries $Q_{t}$ . The process is as follows: + +$$ +\dot {Q} = \left(Q \times \left(F _ {t} \times M\right) ^ {\top}\right) \times F _ {t}. \tag {7} +$$ + +Compared to the original queries, the meaning of the text-dynamic init queries obtained in this method is more explicit and can be understood as a combination of different words in the input referring expression. + +Density Init. While using text combination-based initialization makes it possible for queries to contain different attribute weights, it completely ignores the variance of the given query image. Therefore, we propose to use CAD features which are highly related to the visual features of the given referring expression to further initialize query contents to improve the distinguishing ability. Specifically, we first find $K$ CAD features $D_{K}$ corresponding to the query positions, which has the top $K$ similarity scores to the text + +![](images/6902333488a229a89bb4a751138196b38fb906b309cc66f77223ab56df11ddef.jpg) +Figure 3. Visualization of queries. We visualize the init queries $Q$ , text dynamic initialized $\dot{Q}$ and the CAD dynamic initialized queries $\hat{Q}$ in the polar coordinate system by t-SNE [34]. (b) represent the feature distribution of init queries in the polar coordinate system, which are the same for different referring expressions. (c) represent the feature distribution of text init queries. As shown in (c), the distribution of "bluish pen" and "greenish pen" is quite similar even with much overlap. (d) represent the feature distribution of text & density init queries. As shown in (d), the distribution of queries is quite different, which means they can be distinguished easily. + +feature. Then, we input the corresponding CAD feature $D_K$ as key and value and the text-dynamic queries $\dot{Q}$ as the query for a cross-attention layer to fuse. Finally, we can obtain the CAD dynamic initialized queries $\hat{Q}$ . + +In order to verify that the CAD Dynamic Query Module is conducive to distinguishing between different referring expressions. We visualize the init queries $Q$ , text dynamic initialized $\dot{Q}$ and the CAD dynamic initialized queries $\hat{Q}$ in the polar coordinate system by t-SNE [34]. As shown in Fig. 3, for queries with only text initialization, it is hard to distinguish the bluish pen and greenish pen which are relatively close and similar in the feature space. However, with the help of CAD initialization, the query contents for different input referring expressions are distinguishable. + +# 3.4. Overall Objective + +Our loss function consists of two components, the first one is the localization loss $\mathcal{L}_{\mathrm{loc}}$ as previous REC work based on GroundingDINO [8], and the second of which is the density loss $\mathcal{L}_{\mathrm{density}}$ . The more detail for $\mathcal{L}_{\mathrm{loc}}$ is on the supplementary. Hence the final loss is a weighted sum of two losses: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {l o c}} + \alpha \cdot \mathcal {L} _ {\text {d e n s i t y}}, \tag {8} +$$ + +where $\alpha$ is a parameter to control the influence of $\mathcal{L}_{\mathrm{density}}$ . + +# 4. Experiments + +# 4.1. REC Datasets & Metric + +REC-8k [8]. REC-8K is the first REC task dataset, which contains 8011 different images and 17122 Image-RE pairs. The dataset is split into train, val, test sets which contain 10555, 3336, 3231 Image-RE pairs, 4923, 1566, 1522 images, and 723, 341, 299 REs respectively. 80 REs are shared between train and val sets and 79 REs are shared between train and test sets. The original REC-8K dataset does not contain ground truth contextual attribute density maps, so we generate these maps using a fixed-size Gaussian kernel of 15, following the previous method [19]. + +Metrics. Following prior work on object counting, the Mean Absolute Error (MAE) and the Root Mean Squared Error (RMSE) are used to measure counting performance. For the REC-8K dataset, we include Precision, Recall, and F1 scores as metrics for localization as GroundingREC [8]. + +# 4.2. Implementation Details + +Our model is built on the GroundingDINO [22] where the text and visual backbone, feature enhancer, and localization decoder are from. We provide two versions of our model, trained with Swin-T and Swin-B as visual backbone separately. Please refer to the supplementary for more details about network architectures. + +Table 1. Comparison with the state-of-the-art approaches on the REC-8K dataset. The best performance is boldfaced, and the second is underlined. GroundingREC* is reproduced following the training setting in paper [8] using swin-b as the visual backbone because they do not report the result. Results with † indicate that we use the number of density map estimates to determine the number of queries selected for localization. + +
MethodBackboneVal setTest set
MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑
Mean-14.2827.75---13.7525.91---
ZSC [36]Res-5014.8431.30---14.9329.72---
ZSC [36]Swin-T12.9626.74---13.0029.07---
CounTX [2]ViT-B11.8827.04---11.8425.62---
GroundingDINO [22]Swin-T9.0321.980.560.760.658.8821.950.590.760.66
GroundingREC [8]Swin-T6.8018.130.650.710.686.5019.790.670.720.69
CAD-GD (ours)Swin-T5.4315.010.680.720.705.2917.080.710.730.72
CAD-GD† (ours)Swin-T4.5813.240.680.710.704.5914.680.720.700.71
GroundingREC* [8]Swin-B5.6615.240.660.770.715.4218.470.710.690.70
CAD-GD (ours)Swin-B4.8313.520.740.760.754.9414.650.750.770.76
CAD-GD† (ours)Swin-B4.2313.140.760.700.734.3412.930.770.710.74
+ +Training. We freeze the visual and text backbone to keep the feature space aligned. For the REC-8K dataset, we use the AdamW as the optimizer to train the model for 20 epochs. The learning rate is set to $1e - 5$ and decays by $10\times$ on the $10^{th}$ epoch. For each batch, we send one image with multiple corresponding referring expressions. Besides, data augmentations are not used. + +Inference. For the inference stage, we provide two strategies. In the first strategy, we use a fixed threshold following GroundingREC [8]. For each output query, we compare its score with the CLS token by a threshold of 0.25 and scores with the rest of the text tokens by a threshold of 0.35. If all scores are higher than the threshold, we include the query in the final count. Eventually, the total number of queries that satisfy the threshold condition is considered to be the number of targets in the query image. The second strategy is to discard the threshold and use the target number estimated from the density map as the predicted number of outcomes $N$ . Then we find the $N$ queries among all queries that have the highest scores with the CLS token and add them to the predicted outcomes. We use $\dagger$ to represent the model with the second strategy on Table 1. + +# 4.3. Comparison with State of the Art + +Quantitative Results. In this section, we compare our CAD-GD with language-based counting methods, open-set detection model GroundingDINO, and state-of-the-art REC model GroundingREC on REC-8K datasets. Most of the results for comparison are obtained from previous paper [8] directly. For the GroundingREC, we train a Swin-B version of GroundingREC, which is marked with * on the Table 1, following the training setting in GroundingREC [8] because it only has the swin-T version model. + +As shown in Table 1, for the Swin-T backbone model, CAD-GD† achieve a relative improvement of $32.6\%$ w.r.t validation MAE and $29.4\%$ w.r.t test MAE compared with GroundingDINO. Besides, our method reduces the validation RMSE by $27.0\%$ and test RMSE by $25.8\%$ . The reduction in the RMSE reflects the improved robustness of our method for query images of different densities. For the localization metrics, our model improves precision consistently for localization, while the recall between GroundingREC and our model is similar. This phenomenon may be due to that the goal of our approach is to enhance the capture of contextual attributes for referring expression, thus improving the accuracy of localization instead of recall. + +We also give in the table the performance of the CADGD with Swin-B as the visual backbone. As can be seen from Table 2, our method maintains a steady increase in counting accuracy compared to GroundingREC with Swin-B visual backbone. Meanwhile, as the scale of the visual backbone increases, our method can better improve the localization accuracy compared to GroundingREC. The reason may lie in that the quality of contextual attribute density is highly related to the quality of visual features. + +Qualitative Analysis. For qualitative analysis, we compare our Density-GD with the state-of-the-art REC method GroundingREC in various scenes as shown in Fig. 4. According to the difficulty of the input referring expression, we visualized examples with direct attributes (columns 1-3) and context-related attributes (columns 4-6), respectively. Compared to GroundingREC, even in different density environments, our method can better capture the attribute information of the right target, thus obtaining more accurate counting and localization results. Meanwhile, as shown in columns 4-6, our method can still achieve accurate counting + +![](images/c3db9db66617754618344f988c98ea646f6c4d4ee40d301abcbe6861780ed416.jpg) +Figure 4. Qualitative results on the REC-8k dataset. Columns 1-3 contain examples with direct attributes for REs, and columns 4-6 contain those with context-related attributes. Our method consistently outperforms the GroundingREC with more precious locations. + +results for referring expressions with contextual attributes. From the image corresponding to "person not in a bus", we can see that although the previous method can capture the two key information of "person" and "bus", it can not process the whole semantic features well, and thus wrongly count "person in a bus". However, our approach can capture the context-dependent attribute correctly. + +Effect of CAD. To demonstrate the relation between the contextual attribute density and the final prediction results, we visualize the contextual attribute density map. As shown in Fig. 5, for the same input query image with different referring expressions, our contextual attribute density map can capture where the corresponding semantic features are located. As shown in the images in (a), the contextual attribute density map can distinguish different attributes on both appearance and spatial aspects. One can observe in (b) that our method can perform well even without any attributes for the target class. + +# 4.4. Ablation Study + +We conduct the ablation study on REC-8K and provide quantitative results in Table 2. The plain version of CAD-GD is employed as the baseline which only consists of the standard backbone, feature enhancer, and localization decoder as GroundingREC [8]. + +CAD Generate Module Effect. By comparing R1 and R2 in Table 2, one can observe that SDCM can introduce a relative improvement of $5.4\%$ w.r.t. MAE on the validation set. This performance boost indicates that the counting and localization ability of the model can be improved even by implicitly introducing contextual attribute density features. + +CAD Attention Module Effect. The comparison of R2 vs. R4 demonstrates that the SDAM improves the validation MAE by 0.51 and the precision by 0.29. This result suggests that the counting accuracy of the model can be improved by introducing CAD to visual features directly through SDAM. CAD Dynamic Query Module Effect. By comparing R4 + +Table 2. Ablation study on REC-8k dataset. A represents the CAD Generate Module. B represents the CAD Attention Module, where the B1 and B2 represent the spatial attention and channel attention separately. C represents the CAD Dynamic Query Module, where C1 and C2 represent the text and density feature init. D represents the second strategy for inference, which is to use the target number estimated from the density map as the predicted number. + +
No.ABCDVal
B1B2C1C2MAERMSEPrecRecall
R16.5217.720.6320.703
R2✓6.1716.380.6410.709
R3✓✓5.8816.430.6620.722
R4✓✓✓5.6116.280.6700.710
R5✓✓✓✓5.6714.430.6780.703
R6✓✓✓✓✓5.4315.010.6830.718
R7✓✓✓✓✓✓4.8313.520.6790.711
+ +and R6 in Table 2, one can observe that both the MAE and RMSE are improved. However, the comparison of R4 and R5 demonstrates that the validation MAE declines slightly by only adding text init. The reason may lie in that only text init will reduce the perception of the query image. + +Impact of Inference Strategy. By comparing R6 and R7 in Table 2, it can be observed that the TA can bring a relative improvement of $11.05\%$ w.r.t MAE and $9.93\%$ w.r.t RMSE on validation set. However, the performance on localization metrics even declines slightly. The phenomenon may lie in mismatch localization queries and density maps. + +# 4.5. Results on Zero-shot Counting + +FSC-147. Besides the REC task, we also evaluate our method on class agnostic counting to show our method is generalizable to prior tasks. The quantitative results are presented in Table 3. We evaluate the benchmark dataset FSC-147 [29] in a zero-shot manner, which means the model is + +![](images/7bc4ca07bd478b362a9d74f496a2811c43df2e7869400b6d252c7c33a829cda4.jpg) +Figure 5. Visualization of contextual attribute density map. The predicted density map can capture the locations of referring expressions. + +Table 3. Comparison with the state-of-the-art approaches on the FSC-147 dataset. The upper and lower part of the table presents the results using visual exemplars and class text as guidance separately. + +
MethodValTest
MAERMSEMAERMSE
CounTR [21]13.1349.8311.9591.23
LOCA [33]10.2432.5610.7956.97
CACViT [35]10.6337.959.1348.96
DAVE [27]8.9128.088.6632.36
CountGD [35]7.1026.085.7424.09
Patch-selection [36]26.9388.6322.09115.17
CLIP-count [16]18.7961.1817.78106.62
VLCounter [17]18.0665.1317.05106.16
CounTX [2]17.1065.6115.88106.29
DAVE [27]15.4852.5714.90103.42
GroundingREC [8]10.0658.6210.12107.19
CountGD [3]12.1447.5112.9898.35
CAD-GD (ours)9.3040.9610.3586.88
+ +trained on the FSC-147 train set and evaluated on the val and test set of novel classes. For training and inference, we use the same training and testing strategy as CountGD [3]. Please refer to the supplementary material for more training and testing details. + +As shown in Table 3, our method achieves significantly lower counting errors than all prior zero-shot approaches. By comparing with the CAC methods, it can be seen that although our method only uses text as prompts, the counting error is still comparable with the method using extra visual exemplars as guidance. + +CARPK. To test cross-dataset generalization, we test our method on the CARPK [11] car counting dataset which is + +Table 4. Comparison with the state-of-the-art Sapproaches on the CARPK dataset. Tuned means the method is finetuned on the CARPK training dataset. + +
MethodPromptTunedTest
MAERMSE
CLIP-count [16]Text×11.9616.61
CounTX [2]Text✓8.1310.87
VLCounter [17]Text×6.468.68
CACViT [35]Visual✓4.916.49
CountGD [3]Text×3.835.41
CountGD [3]Text & Visual×3.685.17
CAD-GD (ours)Text×3.294.56
+ +only trained on FSC-147 dataset. As shown in Table 4, CAD-GD achieves lower counting errors than other methods even with visual exemplars as guidance. + +# 5. Conclusion + +In this work, we study why the previous referring expression counting (REC) methods cannot well distinguish between fine-grained attributes for one class. The core issue is that previous methods overlook the contextual attribute density (CAD) conveyed in visual regions. Therefore, we propose the CAD-GD, in which the CAD information is modeled by a CAD estimator and injected into visual features and decoder queries via a well-designed CAD attention module and a CAD dynamic query module. By leveraging the CAD information, consistent performance improvements can be observed on varied benchmarks, which demonstrates the significance of CAD. We hope this work can inspire researchers to explore the CAD in other multimodal tasks. + +# References + +[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2 +[2] Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 6, 8 +[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. arXiv preprint arXiv:2407.04619, 2024. 2, 8 +[4] Xinkun Cao, Zhipeng Wang, Yanyun Zhao, and Fei Su. Scale aggregation network for accurate and efficient crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[6] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19638-19648, 2022. 2 +[7] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 2 +[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16985-16995, 2024. 1, 2, 5, 6, 7, 8 +[9] Chengzhen Duan, Zhiwei Wei, Chi Zhang, Siying Qu, and Hongpeng Wang. Coarse-grained density map guided object detection in aerial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2789-2798, 2021. 2 +[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2 +[11] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In Proceedings of the IEEE international conference on computer vision, pages 4145-4153, 2017. 2, 8 +[12] Yutao Hu, Xiaolong Jiang, Xuhui Liu, Baochang Zhang, Jungong Han, Xianbin Cao, and David Doermann. Nascount: Counting-by-density with neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 747-766. Springer, 2020. 2 +[13] Yi-Xin Huang, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Dq-detr: Detr with dynamic query for tiny object detection. arXiv preprint arXiv:2404.03507, 2024. 2 +[14] Haroon Idrees, Imran Saleemi, Cody Seibert, and Mubarak Shah. Multi-source multi-scale counting in extremely dense + +crowd images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2547-2554, 2013. 3 +[15] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 2 +[16] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 1, 2, 8 +[17] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2714-2722, 2024. 8 +[18] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2 +[19] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 5 +[20] Changlin Li, Taojiannan Yang, Sijie Zhu, Chen Chen, and Shanyue Guan. Density map guided object detection in aerial images. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 190-191, 2020. 2 +[21] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. arXiv preprint arXiv:2208.13721, 2022. 8 +[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1, 2, 3, 4, 5, 6 +[23] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 2 +[24] Hao Lu, Zhiguo Cao, Yang Xiao, Bohan Zhuang, and Chunhua Shen. Tasselnet: counting maize tassels in the wild via local counts regression network. Plant methods, 13:1-17, 2017. 2 +[25] Simon Madec, Xiuliang Jin, Hao Lu, Benoit De Solan, Shouyang Liu, Florent Duyme, Emmanuelle Heritier, and Frederic Baret. Ear density estimation from high resolution rgb imagery using deep learning technique. Agricultural and forest meteorology, 264:225-234, 2019. 2 +[26] Daniel Onoro-Rubio and Roberto J López-Sastre. Towards perspective-free object counting with deep learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 615-629. Springer, 2016. 1 + +[27] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Dave-a detect-and-verify paradigm for low-shot counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23293–23302, 2024. 1, 2, 8 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[29] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 1, 2, 7 +[30] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the" edge" of open-set object detection. arXiv preprint arXiv:2405.10300. 2024. 2 +[31] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 2 +[32] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9529–9538, 2022. 1, 2, 4 +[33] Nikola Dukić, Alan Lukežić, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18872-18881, 2023. 1, 2, 8 +[34] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5 +[35] Zhicheng Wang, Liwen Xiao, Zhiguo Cao, and Hao Lu. Vision transformer off-the-shelf: A surprising baseline for few-shot class-agnostic counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5832–5840, 2024. 2, 4, 8 +[36] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 1, 2, 6, 8 +[37] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 833-841, 2015. 1 +[38] XQ Zhou, YX Zou, and Y Wang. Accurate small object detection via density map aided saliency estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 425-429. IEEE, 2017. 2 \ No newline at end of file diff --git a/2025/Exploring Contextual Attribute Density in Referring Expression Counting/images.zip b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cd5d7317d1edc90260ce8efe83de22da106fe831 --- /dev/null +++ b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10c165161e2ba6686693a71310d696b0eed0c4ebc0bb43d6fb185945a26d2c5b +size 889559 diff --git a/2025/Exploring Contextual Attribute Density in Referring Expression Counting/layout.json b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6aa559ca8f689820cbf520e504d82d0c2771bb4c --- /dev/null +++ b/2025/Exploring Contextual Attribute Density in Referring Expression Counting/layout.json @@ -0,0 +1,7274 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "type": "text", + "content": "Exploring Contextual Attribute Density in Referring Expression Counting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 143, + 538, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 143, + 538, + 172 + ], + "spans": [ + { + "bbox": [ + 72, + 143, + 538, + 172 + ], + "type": "text", + "content": "Zhicheng Wang Zhiyu Pan* Zhan Peng Jian Cheng Liwen Xiao Wei Jiang Zhiguo Cao School of AIA, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 174, + 424, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 174, + 424, + 185 + ], + "spans": [ + { + "bbox": [ + 184, + 174, + 424, + 185 + ], + "type": "text", + "content": "{zhicheng_wang, zhiyupan, zgcao}@hust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 515 + ], + "type": "text", + "content": "Referring expression counting (REC) algorithms are for more flexible and interactive counting ability across varied fine-grained text expressions. However, the requirement for fine-grained attribute understanding poses challenges for prior arts, as they struggle to accurately align attribute information with correct visual patterns. Given the proven importance of \"visual density\", it is presumed that the limitations of current REC approaches stem from an under-exploration of \"contextual attribute density\" (CAD). In the scope of REC, we define CAD as the measure of the information intensity of one certain fine-grained attribute in visual regions. To model the CAD, we propose a U-shape CAD estimator in which referring expression and multi-scale visual features from GroundingDINO can interact with each other. With additional density supervision, we can effectively encode CAD, which is subsequently decoded via a novel attention procedure with CAD-refined queries. Integrating all these contributions, our framework significantly outperforms state-of-the-art REC methods, achieves " + }, + { + "bbox": [ + 55, + 239, + 296, + 515 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 515 + ], + "type": "text", + "content": " error reduction in counting metrics and a " + }, + { + "bbox": [ + 55, + 239, + 296, + 515 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 515 + ], + "type": "text", + "content": " improvement in localization accuracy. The surprising results shed light on the significance of contextual attribute density for REC. Code will be at github.com/Xu3XiWang/CAD-GD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 135, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 135, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 135, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": "Object counting algorithms estimate the number of specific objects in query images. Early research has explored various forms of object descriptions, including class-specific [26, 37], class-agnostic [27, 29, 32, 33], exemplar-guided [29], and text-guided [16, 36] approaches. However, these methods lack the flexibility to distinguish between fine-grained, open-world descriptions, e.g., differentiating a walking person from a person riding motorcycle. To address this, recent work [8] proposes the referring expression counting (REC) task which aims to count fine-grained and contextually nuanced objects within the same class." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 555, + 454 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 555, + 454 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 555, + 454 + ], + "type": "image", + "image_path": "1dba6845c6edfd7cb3e9aff6d3eba562bb6085dd966702fe9e789ce9ad1f1993.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 463, + 555, + 541 + ], + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 541 + ], + "type": "text", + "content": "Figure 1. The efficacy of contextual attribute density (CAD). According to the qualitative comparison above, the REC baseline [8] often over-counts when it focuses primarily on class information rather than fine-grained attributes, and it under-counts in cases of occlusion or scale variations. By modeling CAD, our method more accurately aligns fine-grained attributes with corresponding visual regions, effectively reducing counting errors." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "GroundingREC [8] is firstly developed as the REC baseline with the power of the open-world detection model GroundingDINO [22]. However, distinguishing between objects of the same class based on nuanced attribute differences is challenging, which makes the REC baseline struggle to align attributes with corresponding visual regions. As shown in Fig. 1, two primary types of failure cases can be observed in the REC baseline: (1) Over-counting: the baseline may place excessive emphasis on class information while neglecting fine-grained attributes, e.g., walking, leading to the inclusion of objects with wrong attributes; (2) Under-counting: the baseline might miss objects with the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19587" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "given attribute when occlusion and scale variations arise. These two failure cases may be due to the baseline's reliance on one-to-one matching between texts and image regions, a typical detection-to-count pipeline. Many previous counting approaches [4, 6, 12, 15] have proved that visual density can provide information about scale-robust spatial distribution. However, in modern open-world perception models [18, 22, 30], the capacity to perceive density for a specific attribute is often overlooked. Inspired by the success of visual density modeling, we propose investigating whether modeling contextual attribute density (CAD) for an open-world perception model could help address the above issues in a cross-modal context." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 529 + ], + "type": "text", + "content": "In this work, we attempt to make the REC model aware of the CAD. Analogous to the concept of visual density, we define CAD as the information intensity measurement of one certain fine-grained attribute about one object conveyed in query image regions of different scales. Based on our definition, modeling CAD requires both the image and text information, with the image information encoded in a multi-scale format. Hence, we propose a CAD generate module with U-shape CAD estimator (CADE) that encodes the multi-layer CAD features. The CADE is provided with the visual features and similarity features between visual and text to estimates the CAD map for a referring expression, where the intermediate features serve as representations of CAD. After encoding the CAD, the next problem is how to facilitate the counting procedure with the CAD features. We fuse the CAD features with original visual features from GroundingDINO via both spatial and channel attention. In this way, spatial location and channel importance prior are provided for visual features for referring expression separately. Additionally, the counting queries in the localization decoder for counting are also dynamically initialized by both the text features and their corresponding CAD information. With this dynamic initialization module, the feature differentiation of the initial queries under different inputs is enhanced for better counting accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 532, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 662 + ], + "type": "text", + "content": "Integrating CAD into our REC framework leads to substantial improvements. Experimental results on the public benchmark REC-8K [8] show that our framework outperforms previous state-of-the-art REC baselines by large margins, achieving " + }, + { + "bbox": [ + 55, + 532, + 295, + 662 + ], + "type": "inline_equation", + "content": "37.8\\%" + }, + { + "bbox": [ + 55, + 532, + 295, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 532, + 295, + 662 + ], + "type": "inline_equation", + "content": "33.2\\%" + }, + { + "bbox": [ + 55, + 532, + 295, + 662 + ], + "type": "text", + "content": " relative error reductions on the validation and test sets, respectively. Besides, We also conduct experiments on the previous class-agnostic counting benchmark FSC-147 [29] with a zero-shot setting, with results further underscoring the effectiveness of our framework. These findings highlight the significance of CAD. Our contributions are three-fold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 666, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 666, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 700 + ], + "type": "text", + "content": "- We pioneer the concept of contextual attribute density in REC, enabling more precise differentiation between fine-grained descriptions of objects within the same class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 702, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 294, + 712 + ], + "type": "text", + "content": "- We propose the contextual attribute density estimator and" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "content": "an attention procedure with a query enhancement, including the contextual attribute density into the REC pipeline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 97, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 131 + ], + "type": "text", + "content": "- We achieve significant performance improvements on public benchmarks, demonstrating the value of contextual attribute density modeling for REC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 143, + 400, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 400, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 400, + 155 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 163, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 247 + ], + "type": "text", + "content": "The progression toward more generalized, fine-grained counting tasks has shifted from class-specific to class-agnostic methods, and also from visual to textual guidance. This section reviews prior works contributing to the development of open-world counting models, emphasizing advancements in visual density for counting tasks given our introduction of contextual attribute density." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 545 + ], + "type": "text", + "content": "Advancing to open-world counting. Counting tasks originally focus on specific categories, such as crowd [1, 4, 15], vehicle [11], wheat head [24, 25], and so on. These early methods are limited to predefined target types and struggle to generalize across diverse object categories. The class-agnostic counting (CAC) [23] marks a significant step toward more flexible counting approaches [27, 32, 33, 35]. Defined as a template-matching problem [23], CAC achieves generalized counting by using visual exemplars to count objects beyond specific classes. Notably, visual density-based methods [32, 33, 35] consistently outperform traditional detection-based counting methods such as MaskRCNN [10] and RetinaNet [31] on counting benchmarks. However, the reliance on visual templates is still a notable limitation in practical applications. To address this issue, textual-based counting [16, 36] is proposed by leveraging the multi-modal ability of CLIP [28]. Although incorporating visual guidance information has been proved beneficial [3], researchers pursue a more generalized setting: with only text expressions, fine-grained, general-purpose objects are expected to be counted, which is referring expression counting (REC) [8]. The REC baseline solely relies on the multi-modal localization ability of GroundingDINO [22], and interpreting some fine-grained expressions accurately remains challenging, as shown in Fig. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "Density modeling in computer vision. Beyond the task of counting, there are some works in other fields using visual density features to provide the spatial and target number prior, e.g., instance segmentation [7] and tiny object detection [9, 13, 20, 38]. Cholakka et al. [7] use density maps to provide prior information on object counts and spatial locations, thereby assisting instance detection tasks under image-level supervision. Huang et al. [13] propose DQ-DETR for tiny object detection, which utilizes the density feature to provide the general target number and spatial prior for DETR [5] structure detection pipeline. In contrast to these methods, our approach focuses more on the role of contextual attribute density for fine-grained expression understanding in a multi-modal setting." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19588" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 243 + ], + "type": "image", + "image_path": "9d1a3e0b8b14e70ebc548d9f6b9d86b554208a0897259ace4a33371bb8542176.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 314 + ], + "type": "text", + "content": "Figure 2. The framework of Contextual Attribute Density Aware GroundingDINO (CAD-GD). A query image and referring expression are sent into backbones separately together with the feature enhancer to get the visual features and the text feature. Then we obtain the Contextual Attribute Density (CAD) features by the CAD Generate Module supervised by the GT contextual attribute density map. To leverage the CAD information, we design the CAD Attention Module and CAD Dynamic Query Module to enhance visual features and query contents separately. Then we sent the dynamic queries, CAD enhanced visual features, and the text feature into the localization decoder to get the final localization prediction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 333, + 111, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 111, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 111, + 345 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": "In this section, we first introduce the overview of our framework for Referring Expression Counting (REC), named Contextual Attribute Density Aware GroundingDINO (CAD-GD). The overview diagram of our approach is shown in Fig. 2. Given an input query image " + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": " and a referring expression text " + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "inline_equation", + "content": "T = \\{w_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": " words, our model predicts a localization counting result. For each " + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "inline_equation", + "content": "(I, T)" + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": " pair, we first send multi-scale image features and text features extracted from the image and text backbones into feature enhancer [22] to obtain cross-modality visual features " + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\{F_{vi} \\in \\mathbb{R}^{h_i \\times w_i \\times C}\\}_{i=1}^4" + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "inline_equation", + "content": "F_t \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": " where the " + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 354, + 296, + 496 + ], + "type": "text", + "content": " indicates the number of channels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": "Afterwards, we feed visual features " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{F_{vi}\\}_{i = 1}^{4}" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " into the U-shape contextual attribute density estimator (CADE) to obtain the contextual attribute density (CAD) features " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{D_i\\in \\mathbb{R}^{h_i\\times w_i\\times C}\\}_{i = 1}^4" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " . We then inject the CAD features " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i = 1}^4" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " into visual features " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{F_{vi}\\}_{i = 1}^{4}" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " to enhance the visual feature with spatial and channel attention in CAD Attention Module. Besides, for the initialization of queries in the cross-modality localization decoder [22], we selected the " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " visual feature positions with the highest scores among them as the initial query locations by calculating the similarity between the visual and text features, while for the initial content features of the query, we dynamically initialize them with text features " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " and CAD features " + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i = 1}^4" + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": " corresponding to the query positions. Then we refine the queries by the cross-modality decoder at every layer with CAD enhanced visual and text features. The output queries of the last decoder layer will be used to predict the localization of referring expression objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 334, + 455, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 334, + 455, + 345 + ], + "spans": [ + { + "bbox": [ + 314, + 334, + 455, + 345 + ], + "type": "text", + "content": "3.1. CAD Generation Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 352, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 472 + ], + "type": "text", + "content": "The first step to model CAD for referring expression counting is to obtain the CAD features, which should be responsive to input text with fine-grained attributes. To obtain the satisfying CAD features, we identify several requirements for obtaining satisfying CAD features: (a) the image-text similarity features are required to guide the estimator to focus on areas corresponding to the input referring expression; (b) due to the density diversity of targets in the counting task, a multi-scale structure is of benefit for the robustness of counting [14]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "text", + "content": "Motivated by the above requirements, the CAD Generate Module is proposed, as in Fig. 2. Specifically, we first utilize the well-pretrained linear projector from GroundingDINO to project the visual features " + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\{F_{vi}\\}_{i = 1}^{4}" + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "text", + "content": " into the text feature space. Then we calculate the similarity feature " + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\{S_i\\in \\mathbb{R}^{h_i\\times w_i\\times C}\\}_{i = 1}^4" + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "text", + "content": " between the well-aligned visual feature " + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "inline_equation", + "content": "F_{vi}" + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "text", + "content": " and corresponding referring expression feature " + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 473, + 555, + 567 + ], + "type": "text", + "content": " as follow:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 390, + 571, + 553, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 571, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 390, + 571, + 553, + 585 + ], + "type": "interline_equation", + "content": "S _ {i} = \\operatorname {P r o j} \\left(F _ {v i}\\right) \\cdot F _ {t}, \\tag {1}", + "image_path": "b9ad8e198ceb0869d1013c5213ec47dd1ed73b58649becd5ea3b495dd3068478.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "text", + "content": "where the Proj consists of a single linear layer followed by a layer norm, and the " + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "text", + "content": " represents the element-wise multiplication. Then we send the " + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{S_i\\}_{i=1}^4" + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "text", + "content": " together with the visual feature " + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{F_{vi}\\}_{i=1}^4" + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "text", + "content": " into a U-shape CAD estimator (CADE) to estimate the CAD map for a referring expression, where the intermediate features serve as the CAD features " + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i=1}^4" + }, + { + "bbox": [ + 313, + 593, + 555, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Supervision Signal. The final output of CADE is the density map for referring expression. In order to guide the CADE to learn the contextual attribute density map, we adopt a conventional " + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": " loss as the counting loss " + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{density}}" + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19589" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 96, + 83, + 295, + 98 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 83, + 295, + 98 + ], + "spans": [ + { + "bbox": [ + 96, + 83, + 295, + 98 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e n s i t y}} = \\left\\| D _ {\\text {p r e d}} (I, T) - D _ {\\mathrm {g t}} (I, T) \\right\\| _ {2} ^ {2}, \\tag {2}", + "image_path": "9074e07a41f7682360bbb6568231f88c186fc0aa264837d1383e858ecf689715.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 295, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 295, + 126 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 102, + 295, + 126 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{pred}}(I,T)" + }, + { + "bbox": [ + 55, + 102, + 295, + 126 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 102, + 295, + 126 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{gt}}(I,T)" + }, + { + "bbox": [ + 55, + 102, + 295, + 126 + ], + "type": "text", + "content": " denote the prediction and ground truth contextual attribute density map respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 188, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 188, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 188, + 144 + ], + "type": "text", + "content": "3.2. CAD Attention Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 149, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 295, + 258 + ], + "type": "text", + "content": "After obtaining the CAD features " + }, + { + "bbox": [ + 55, + 149, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i=1}^4" + }, + { + "bbox": [ + 55, + 149, + 295, + 258 + ], + "type": "text", + "content": ", the next core step is to consider how to inject the CAD information into the localization process, which follows a typically DETR decoder structure [22]. In the DETR-like structure, the visual features " + }, + { + "bbox": [ + 55, + 149, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\{F_{vi}\\}_{i=1}^4" + }, + { + "bbox": [ + 55, + 149, + 295, + 258 + ], + "type": "text", + "content": " are crucial for the localization process due to its cross-attention with queries. Motivated by the high relevance between the density feature and the attention map [35], we inject the CAD features into visual features by attention mechanism, as shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "content": "Spatial Attention. Since the visual features need to be matched with queries to find the target location, foreground enhancement for visual features on areas corresponding to the input referring expression can help to get accurate localization results. Hence we design the spatial attention mechanism to enhance the CAD corresponding foreground areas, utilizing distribution prior from CAD. In specific, for the " + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "content": " scale of " + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i=1}^4" + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "content": ", we first apply the average pooling and max pooling for " + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "content": " on the channel dimension. Then we concatenate these two channel pooled features together to get the spatial attention features " + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "F_i^s \\in \\mathbb{R}^{h_i \\times w_i \\times 2}" + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "content": ". Afterward, the spatial attention features " + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "F_i^s" + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "content": " are sent into CNN layers followed by a sigmoid function to obtain the spatial attention map for " + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "F_{vi}" + }, + { + "bbox": [ + 55, + 258, + 296, + 426 + ], + "type": "text", + "content": ". The process is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 434, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 434, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 73, + 434, + 295, + 449 + ], + "type": "interline_equation", + "content": "A _ {i} ^ {s} = \\operatorname {S i g m o i d} \\left(\\operatorname {C o n v} \\left(\\left[ \\operatorname {M a x} \\left(D _ {i}\\right), \\operatorname {A v g} \\left(D _ {i}\\right) \\right]\\right)\\right), \\tag {3}", + "image_path": "d951f4f4445e33dd53ce98a330857523851d7cab10694bd316bb824574b10bf4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "inline_equation", + "content": "A_{i}^{s}\\in \\mathbb{R}^{h_{i}\\times w_{i}\\times 1}" + }, + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "text", + "content": " is the spatial attention map for " + }, + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "inline_equation", + "content": "F_{vi}" + }, + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "text", + "content": ". Then we multiple the visual feature " + }, + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "inline_equation", + "content": "F_{vi}" + }, + { + "bbox": [ + 55, + 455, + 295, + 492 + ], + "type": "text", + "content": " and corresponding spatial attention map as follow:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 498, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 498, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 141, + 498, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\dot {F} _ {v i} = F _ {v i} \\cdot A _ {i} ^ {s}. \\tag {4}", + "image_path": "20341e627ea17f07d3732f41243e5142a32a93dc0bfec07305e5540ae6f41763.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "text", + "content": "This results " + }, + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\{F_{vi}\\}_{i = 1}^{4}" + }, + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "text", + "content": " are the spatially attended visual features for each scale " + }, + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Channel Attention. The previous work [32] shows that channel attention can help to distinguish different class targets. For the referring expression counting task, it is even harder to distinguish targets due to the contextual attribute understanding requirement. So it is beneficial to make use of channel attention to enhance the distinguishing ability of the REC model. However, the previous channel attention in the CAC task neglects the scale variance due to its relation to fixed-size exemplars. Thanks to the multi-layer CAD features, our channel attention can vary on different scales. Specifically, for the " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " scale, we first apply average pooling and max pooling followed by shared MLP on the spatial attention enhanced visual feature " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\dot{F}_{vi}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " along the spatial dimension to get two channel attention features " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{i}^{c} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "We add two features up, then feed the added feature into the sigmoid function to get the channel attention map " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "A_{i}^{c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 327, + 105, + 555, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 105, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 327, + 105, + 555, + 119 + ], + "type": "interline_equation", + "content": "A _ {i} ^ {c} = \\operatorname {S i g m o i d} \\left(M L P \\left(\\operatorname {A v g} \\left(\\dot {F} _ {v i}\\right) + \\operatorname {M a x} \\left(\\dot {F} _ {v i}\\right)\\right)\\right). \\tag {5}", + "image_path": "a9241dfb6641fc5d62c5bb2ab415d4c8021f4b87d3fa9193041067627f32402f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "text", + "content": "Afterwards, we multiple the spatial attention enhanced visual feature " + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "inline_equation", + "content": "\\dot{F}_{vi}" + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "text", + "content": " with channel attention map " + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "inline_equation", + "content": "A_{i}^{c}" + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "text", + "content": " in every scale " + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "text", + "content": " to obtain the final CAD enhanced visual features " + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_{vi}\\}_{i=1}^{4}" + }, + { + "bbox": [ + 313, + 129, + 555, + 177 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 400, + 186, + 555, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 186, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 400, + 186, + 555, + 200 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {v i} = \\dot {F} _ {v i} \\cdot A _ {i} ^ {c}. \\tag {6}", + "image_path": "5359e49039049b252e154ede6d1a03483e99c3b5c6d7384719ad2ac0d0273a4f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 209, + 477, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 477, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 477, + 221 + ], + "type": "text", + "content": "3.3. CAD Dynamic Query Module" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "text", + "content": "The queries consisting of their locations and corresponding content features are another essential component for the localization decoder. For the initial location of queries, we follow the GroundingDINO strategy of selecting visual feature positions with the top " + }, + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "text", + "content": " similarity scores to the text feature. For the corresponding content features " + }, + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "text", + "content": " of queries which are learnt embedding in GroundingDINO, we argue that the fixed embedding is not the ideal initialization for query contents in REC. In the REC task, the target with a specific attribute is given, hence the initialized " + }, + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 226, + 555, + 370 + ], + "type": "text", + "content": " content is expected to contain as many kinds belonging to this given referring expression as possible." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": "Text Init. Therefore, as mentioned above, for the referring expression counting task, since it has only one given target text, we propose to dynamically initialize the query contents with text features first as shown in Fig. 2. In specific, we first multiply the input text feature " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": " with the learning matrix " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": ", then send it into the GELU function to obtain the dynamic text feature " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": " is the words number. Afterward, we multiply the dynamic text feature " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": " with init queries " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": " is the query number. We can get the dynamic text weight map " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{K \\times N}" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": ". Finally, we multiply the dynamic weight with the text feature " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": " to get the text-dynamic init queries " + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "inline_equation", + "content": "Q_{t}" + }, + { + "bbox": [ + 313, + 370, + 556, + 525 + ], + "type": "text", + "content": ". The process is as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 534, + 555, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 534, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 372, + 534, + 555, + 548 + ], + "type": "interline_equation", + "content": "\\dot {Q} = \\left(Q \\times \\left(F _ {t} \\times M\\right) ^ {\\top}\\right) \\times F _ {t}. \\tag {7}", + "image_path": "68f3e3b17d87da1ceac9c0d05d3b857b276f7c2a9c86435795d36d4362428612.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": "Compared to the original queries, the meaning of the text-dynamic init queries obtained in this method is more explicit and can be understood as a combination of different words in the input referring expression." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Density Init. While using text combination-based initialization makes it possible for queries to contain different attribute weights, it completely ignores the variance of the given query image. Therefore, we propose to use CAD features which are highly related to the visual features of the given referring expression to further initialize query contents to improve the distinguishing ability. Specifically, we first find " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " CAD features " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "D_{K}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " corresponding to the query positions, which has the top " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " similarity scores to the text" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19590" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 553, + 299 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 553, + 299 + ], + "type": "image", + "image_path": "6902333488a229a89bb4a751138196b38fb906b309cc66f77223ab56df11ddef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "type": "text", + "content": "Figure 3. Visualization of queries. We visualize the init queries " + }, + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "type": "text", + "content": ", text dynamic initialized " + }, + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\dot{Q}" + }, + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "type": "text", + "content": " and the CAD dynamic initialized queries " + }, + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\hat{Q}" + }, + { + "bbox": [ + 54, + 308, + 555, + 366 + ], + "type": "text", + "content": " in the polar coordinate system by t-SNE [34]. (b) represent the feature distribution of init queries in the polar coordinate system, which are the same for different referring expressions. (c) represent the feature distribution of text init queries. As shown in (c), the distribution of \"bluish pen\" and \"greenish pen\" is quite similar even with much overlap. (d) represent the feature distribution of text & density init queries. As shown in (d), the distribution of queries is quite different, which means they can be distinguished easily." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "text", + "content": "feature. Then, we input the corresponding CAD feature " + }, + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "inline_equation", + "content": "D_K" + }, + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "text", + "content": " as key and value and the text-dynamic queries " + }, + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "inline_equation", + "content": "\\dot{Q}" + }, + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "text", + "content": " as the query for a cross-attention layer to fuse. Finally, we can obtain the CAD dynamic initialized queries " + }, + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "inline_equation", + "content": "\\hat{Q}" + }, + { + "bbox": [ + 54, + 384, + 294, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "content": "In order to verify that the CAD Dynamic Query Module is conducive to distinguishing between different referring expressions. We visualize the init queries " + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "content": ", text dynamic initialized " + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\dot{Q}" + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "content": " and the CAD dynamic initialized queries " + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\hat{Q}" + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "content": " in the polar coordinate system by t-SNE [34]. As shown in Fig. 3, for queries with only text initialization, it is hard to distinguish the bluish pen and greenish pen which are relatively close and similar in the feature space. However, with the help of CAD initialization, the query contents for different input referring expressions are distinguishable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 571, + 162, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 162, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 162, + 586 + ], + "type": "text", + "content": "3.4. Overall Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "text", + "content": "Our loss function consists of two components, the first one is the localization loss " + }, + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "text", + "content": " as previous REC work based on GroundingDINO [8], and the second of which is the density loss " + }, + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{density}}" + }, + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "text", + "content": ". The more detail for " + }, + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 55, + 593, + 296, + 653 + ], + "type": "text", + "content": " is on the supplementary. Hence the final loss is a weighted sum of two losses:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 674, + 295, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 674, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 127, + 674, + 295, + 688 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {l o c}} + \\alpha \\cdot \\mathcal {L} _ {\\text {d e n s i t y}}, \\tag {8}", + "image_path": "f9893a8a7f1949d03986186bb719081fda5dc14724df62343b4cb928ae33cb45.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 701, + 288, + 715 + ], + "type": "text", + "content": " is a parameter to control the influence of " + }, + { + "bbox": [ + 55, + 701, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{density}}" + }, + { + "bbox": [ + 55, + 701, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 383, + 395, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 383, + 395, + 397 + ], + "spans": [ + { + "bbox": [ + 314, + 383, + 395, + 397 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 408, + 449, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 449, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 449, + 419 + ], + "type": "text", + "content": "4.1. REC Datasets & Metric" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 547 + ], + "type": "text", + "content": "REC-8k [8]. REC-8K is the first REC task dataset, which contains 8011 different images and 17122 Image-RE pairs. The dataset is split into train, val, test sets which contain 10555, 3336, 3231 Image-RE pairs, 4923, 1566, 1522 images, and 723, 341, 299 REs respectively. 80 REs are shared between train and val sets and 79 REs are shared between train and test sets. The original REC-8K dataset does not contain ground truth contextual attribute density maps, so we generate these maps using a fixed-size Gaussian kernel of 15, following the previous method [19]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 548, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 608 + ], + "type": "text", + "content": "Metrics. Following prior work on object counting, the Mean Absolute Error (MAE) and the Root Mean Squared Error (RMSE) are used to measure counting performance. For the REC-8K dataset, we include Precision, Recall, and F1 scores as metrics for localization as GroundingREC [8]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 623, + 447, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 447, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 447, + 635 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "type": "text", + "content": "Our model is built on the GroundingDINO [22] where the text and visual backbone, feature enhancer, and localization decoder are from. We provide two versions of our model, trained with Swin-T and Swin-B as visual backbone separately. Please refer to the supplementary for more details about network architectures." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19591" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 123, + 552, + 298 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 114 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art approaches on the REC-8K dataset. The best performance is boldfaced, and the second is underlined. GroundingREC* is reproduced following the training setting in paper [8] using swin-b as the visual backbone because they do not report the result. Results with † indicate that we use the number of density map estimates to determine the number of queries selected for localization." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 123, + 552, + 298 + ], + "lines": [ + { + "bbox": [ + 60, + 123, + 552, + 298 + ], + "spans": [ + { + "bbox": [ + 60, + 123, + 552, + 298 + ], + "type": "table", + "html": "
MethodBackboneVal setTest set
MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑MAE ↓RMSE ↓Prec ↑Rec ↑F1 ↑
Mean-14.2827.75---13.7525.91---
ZSC [36]Res-5014.8431.30---14.9329.72---
ZSC [36]Swin-T12.9626.74---13.0029.07---
CounTX [2]ViT-B11.8827.04---11.8425.62---
GroundingDINO [22]Swin-T9.0321.980.560.760.658.8821.950.590.760.66
GroundingREC [8]Swin-T6.8018.130.650.710.686.5019.790.670.720.69
CAD-GD (ours)Swin-T5.4315.010.680.720.705.2917.080.710.730.72
CAD-GD† (ours)Swin-T4.5813.240.680.710.704.5914.680.720.700.71
GroundingREC* [8]Swin-B5.6615.240.660.770.715.4218.470.710.690.70
CAD-GD (ours)Swin-B4.8313.520.740.760.754.9414.650.750.770.76
CAD-GD† (ours)Swin-B4.2313.140.760.700.734.3412.930.770.710.74
", + "image_path": "ab730acf73802141c618c45b68df9e648a1de955aa760bff1b49b2529847eb67.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "text", + "content": "Training. We freeze the visual and text backbone to keep the feature space aligned. For the REC-8K dataset, we use the AdamW as the optimizer to train the model for 20 epochs. The learning rate is set to " + }, + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "text", + "content": " and decays by " + }, + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 54, + 317, + 294, + 400 + ], + "type": "text", + "content": " epoch. For each batch, we send one image with multiple corresponding referring expressions. Besides, data augmentations are not used." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "text", + "content": "Inference. For the inference stage, we provide two strategies. In the first strategy, we use a fixed threshold following GroundingREC [8]. For each output query, we compare its score with the CLS token by a threshold of 0.25 and scores with the rest of the text tokens by a threshold of 0.35. If all scores are higher than the threshold, we include the query in the final count. Eventually, the total number of queries that satisfy the threshold condition is considered to be the number of targets in the query image. The second strategy is to discard the threshold and use the target number estimated from the density map as the predicted number of outcomes " + }, + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "text", + "content": ". Then we find the " + }, + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "text", + "content": " queries among all queries that have the highest scores with the CLS token and add them to the predicted outcomes. We use " + }, + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 402, + 295, + 582 + ], + "type": "text", + "content": " to represent the model with the second strategy on Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 588, + 235, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 235, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 235, + 601 + ], + "type": "text", + "content": "4.3. Comparison with State of the Art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Quantitative Results. In this section, we compare our CAD-GD with language-based counting methods, open-set detection model GroundingDINO, and state-of-the-art REC model GroundingREC on REC-8K datasets. Most of the results for comparison are obtained from previous paper [8] directly. For the GroundingREC, we train a Swin-B version of GroundingREC, which is marked with * on the Table 1, following the training setting in GroundingREC [8] because it only has the swin-T version model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "text", + "content": "As shown in Table 1, for the Swin-T backbone model, CAD-GD† achieve a relative improvement of " + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "text", + "content": " w.r.t validation MAE and " + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "inline_equation", + "content": "29.4\\%" + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "text", + "content": " w.r.t test MAE compared with GroundingDINO. Besides, our method reduces the validation RMSE by " + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "inline_equation", + "content": "27.0\\%" + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "text", + "content": " and test RMSE by " + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "inline_equation", + "content": "25.8\\%" + }, + { + "bbox": [ + 313, + 317, + 555, + 473 + ], + "type": "text", + "content": ". The reduction in the RMSE reflects the improved robustness of our method for query images of different densities. For the localization metrics, our model improves precision consistently for localization, while the recall between GroundingREC and our model is similar. This phenomenon may be due to that the goal of our approach is to enhance the capture of contextual attributes for referring expression, thus improving the accuracy of localization instead of recall." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 474, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 582 + ], + "type": "text", + "content": "We also give in the table the performance of the CADGD with Swin-B as the visual backbone. As can be seen from Table 2, our method maintains a steady increase in counting accuracy compared to GroundingREC with Swin-B visual backbone. Meanwhile, as the scale of the visual backbone increases, our method can better improve the localization accuracy compared to GroundingREC. The reason may lie in that the quality of contextual attribute density is highly related to the quality of visual features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Qualitative Analysis. For qualitative analysis, we compare our Density-GD with the state-of-the-art REC method GroundingREC in various scenes as shown in Fig. 4. According to the difficulty of the input referring expression, we visualized examples with direct attributes (columns 1-3) and context-related attributes (columns 4-6), respectively. Compared to GroundingREC, even in different density environments, our method can better capture the attribute information of the right target, thus obtaining more accurate counting and localization results. Meanwhile, as shown in columns 4-6, our method can still achieve accurate counting" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19592" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 553, + 221 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 221 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 221 + ], + "type": "image", + "image_path": "c3db9db66617754618344f988c98ea646f6c4d4ee40d301abcbe6861780ed416.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 555, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 249 + ], + "type": "text", + "content": "Figure 4. Qualitative results on the REC-8k dataset. Columns 1-3 contain examples with direct attributes for REs, and columns 4-6 contain those with context-related attributes. Our method consistently outperforms the GroundingREC with more precious locations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 268, + 294, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 294, + 352 + ], + "type": "text", + "content": "results for referring expressions with contextual attributes. From the image corresponding to \"person not in a bus\", we can see that although the previous method can capture the two key information of \"person\" and \"bus\", it can not process the whole semantic features well, and thus wrongly count \"person in a bus\". However, our approach can capture the context-dependent attribute correctly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 352, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 295, + 483 + ], + "type": "text", + "content": "Effect of CAD. To demonstrate the relation between the contextual attribute density and the final prediction results, we visualize the contextual attribute density map. As shown in Fig. 5, for the same input query image with different referring expressions, our contextual attribute density map can capture where the corresponding semantic features are located. As shown in the images in (a), the contextual attribute density map can distinguish different attributes on both appearance and spatial aspects. One can observe in (b) that our method can perform well even without any attributes for the target class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 569 + ], + "type": "text", + "content": "We conduct the ablation study on REC-8K and provide quantitative results in Table 2. The plain version of CAD-GD is employed as the baseline which only consists of the standard backbone, feature enhancer, and localization decoder as GroundingREC [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": "CAD Generate Module Effect. By comparing R1 and R2 in Table 2, one can observe that SDCM can introduce a relative improvement of " + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": " w.r.t. MAE on the validation set. This performance boost indicates that the counting and localization ability of the model can be improved even by implicitly introducing contextual attribute density features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "CAD Attention Module Effect. The comparison of R2 vs. R4 demonstrates that the SDAM improves the validation MAE by 0.51 and the precision by 0.29. This result suggests that the counting accuracy of the model can be improved by introducing CAD to visual features directly through SDAM. CAD Dynamic Query Module Effect. By comparing R4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 267, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 344 + ], + "type": "text", + "content": "Table 2. Ablation study on REC-8k dataset. A represents the CAD Generate Module. B represents the CAD Attention Module, where the B1 and B2 represent the spatial attention and channel attention separately. C represents the CAD Dynamic Query Module, where C1 and C2 represent the text and density feature init. D represents the second strategy for inference, which is to use the target number estimated from the density map as the predicted number." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 353, + 555, + 479 + ], + "blocks": [ + { + "bbox": [ + 315, + 353, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 315, + 353, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 555, + 479 + ], + "type": "table", + "html": "
No.ABCDVal
B1B2C1C2MAERMSEPrecRecall
R16.5217.720.6320.703
R2✓6.1716.380.6410.709
R3✓✓5.8816.430.6620.722
R4✓✓✓5.6116.280.6700.710
R5✓✓✓✓5.6714.430.6780.703
R6✓✓✓✓✓5.4315.010.6830.718
R7✓✓✓✓✓✓4.8313.520.6790.711
", + "image_path": "0ef466b1dd428dac1c8a4bf32a7651742049a96fe1243f46fbe7ac00bd9c2d1f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 557 + ], + "type": "text", + "content": "and R6 in Table 2, one can observe that both the MAE and RMSE are improved. However, the comparison of R4 and R5 demonstrates that the validation MAE declines slightly by only adding text init. The reason may lie in that only text init will reduce the perception of the query image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "text", + "content": "Impact of Inference Strategy. By comparing R6 and R7 in Table 2, it can be observed that the TA can bring a relative improvement of " + }, + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "inline_equation", + "content": "11.05\\%" + }, + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "text", + "content": " w.r.t MAE and " + }, + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "inline_equation", + "content": "9.93\\%" + }, + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "text", + "content": " w.r.t RMSE on validation set. However, the performance on localization metrics even declines slightly. The phenomenon may lie in mismatch localization queries and density maps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 636, + 481, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 481, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 481, + 649 + ], + "type": "text", + "content": "4.5. Results on Zero-shot Counting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "FSC-147. Besides the REC task, we also evaluate our method on class agnostic counting to show our method is generalizable to prior tasks. The quantitative results are presented in Table 3. We evaluate the benchmark dataset FSC-147 [29] in a zero-shot manner, which means the model is" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19593" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 251 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 251 + ], + "type": "image", + "image_path": "7bc4ca07bd478b362a9d74f496a2811c43df2e7869400b6d252c7c33a829cda4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 553, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 553, + 265 + ], + "type": "text", + "content": "Figure 5. Visualization of contextual attribute density map. The predicted density map can capture the locations of referring expressions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 337, + 291, + 526 + ], + "blocks": [ + { + "bbox": [ + 55, + 284, + 295, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 328 + ], + "type": "text", + "content": "Table 3. Comparison with the state-of-the-art approaches on the FSC-147 dataset. The upper and lower part of the table presents the results using visual exemplars and class text as guidance separately." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 337, + 291, + 526 + ], + "lines": [ + { + "bbox": [ + 59, + 337, + 291, + 526 + ], + "spans": [ + { + "bbox": [ + 59, + 337, + 291, + 526 + ], + "type": "table", + "html": "
MethodValTest
MAERMSEMAERMSE
CounTR [21]13.1349.8311.9591.23
LOCA [33]10.2432.5610.7956.97
CACViT [35]10.6337.959.1348.96
DAVE [27]8.9128.088.6632.36
CountGD [35]7.1026.085.7424.09
Patch-selection [36]26.9388.6322.09115.17
CLIP-count [16]18.7961.1817.78106.62
VLCounter [17]18.0665.1317.05106.16
CounTX [2]17.1065.6115.88106.29
DAVE [27]15.4852.5714.90103.42
GroundingREC [8]10.0658.6210.12107.19
CountGD [3]12.1447.5112.9898.35
CAD-GD (ours)9.3040.9610.3586.88
", + "image_path": "33ba2036e71fd674b26fd041ddb4fe8042dc146c4cca781e92917104074652aa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 612 + ], + "type": "text", + "content": "trained on the FSC-147 train set and evaluated on the val and test set of novel classes. For training and inference, we use the same training and testing strategy as CountGD [3]. Please refer to the supplementary material for more training and testing details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 615, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 686 + ], + "type": "text", + "content": "As shown in Table 3, our method achieves significantly lower counting errors than all prior zero-shot approaches. By comparing with the CAC methods, it can be seen that although our method only uses text as prompts, the counting error is still comparable with the method using extra visual exemplars as guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "CARPK. To test cross-dataset generalization, we test our method on the CARPK [11] car counting dataset which is" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 327, + 550, + 441 + ], + "blocks": [ + { + "bbox": [ + 313, + 284, + 553, + 317 + ], + "lines": [ + { + "bbox": [ + 313, + 284, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 553, + 317 + ], + "type": "text", + "content": "Table 4. Comparison with the state-of-the-art Sapproaches on the CARPK dataset. Tuned means the method is finetuned on the CARPK training dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 327, + 550, + 441 + ], + "lines": [ + { + "bbox": [ + 318, + 327, + 550, + 441 + ], + "spans": [ + { + "bbox": [ + 318, + 327, + 550, + 441 + ], + "type": "table", + "html": "
MethodPromptTunedTest
MAERMSE
CLIP-count [16]Text×11.9616.61
CounTX [2]Text✓8.1310.87
VLCounter [17]Text×6.468.68
CACViT [35]Visual✓4.916.49
CountGD [3]Text×3.835.41
CountGD [3]Text & Visual×3.685.17
CAD-GD (ours)Text×3.294.56
", + "image_path": "482cd3f81a67f92764f5f848994ebf4d969cdd3a165950acf5d1cb81ce4ec770.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 504 + ], + "type": "text", + "content": "only trained on FSC-147 dataset. As shown in Table 4, CAD-GD achieves lower counting errors than other methods even with visual exemplars as guidance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 388, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 388, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 388, + 534 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 546, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 555, + 712 + ], + "type": "text", + "content": "In this work, we study why the previous referring expression counting (REC) methods cannot well distinguish between fine-grained attributes for one class. The core issue is that previous methods overlook the contextual attribute density (CAD) conveyed in visual regions. Therefore, we propose the CAD-GD, in which the CAD information is modeled by a CAD estimator and injected into visual features and decoder queries via a well-designed CAD attention module and a CAD dynamic query module. By leveraging the CAD information, consistent performance improvements can be observed on varied benchmarks, which demonstrates the significance of CAD. We hope this work can inspire researchers to explore the CAD in other multimodal tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19594" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "text", + "content": "[2] Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 201 + ], + "type": "text", + "content": "[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. arXiv preprint arXiv:2407.04619, 2024. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 247 + ], + "type": "text", + "content": "[4] Xinkun Cao, Zhipeng Wang, Yanyun Zhao, and Fei Su. Scale aggregation network for accurate and efficient crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 248, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 295, + 291 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 346 + ], + "type": "text", + "content": "[6] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19638-19648, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "type": "text", + "content": "[7] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 445 + ], + "type": "text", + "content": "[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16985-16995, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 448, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 295, + 501 + ], + "type": "text", + "content": "[9] Chengzhen Duan, Zhiwei Wei, Chi Zhang, Siying Qu, and Hongpeng Wang. Coarse-grained density map guided object detection in aerial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2789-2798, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 535 + ], + "type": "text", + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[11] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In Proceedings of the IEEE international conference on computer vision, pages 4145-4153, 2017. 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "type": "text", + "content": "[12] Yutao Hu, Xiaolong Jiang, Xuhui Liu, Baochang Zhang, Jungong Han, Xianbin Cao, and David Doermann. Nascount: Counting-by-density with neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 747-766. Springer, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "text", + "content": "[13] Yi-Xin Huang, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Dq-detr: Detr with dynamic query for tiny object detection. arXiv preprint arXiv:2404.03507, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[14] Haroon Idrees, Imran Saleemi, Cody Seibert, and Mubarak Shah. Multi-source multi-scale counting in extremely dense" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 105 + ], + "type": "text", + "content": "crowd images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2547-2554, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 172 + ], + "type": "text", + "content": "[15] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "type": "text", + "content": "[16] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 1, 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "text", + "content": "[17] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2714-2722, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "text", + "content": "[18] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 321, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 554, + 354 + ], + "type": "text", + "content": "[19] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 554, + 408 + ], + "type": "text", + "content": "[20] Changlin Li, Taojiannan Yang, Sijie Zhu, Chen Chen, and Shanyue Guan. Density map guided object detection in aerial images. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 190-191, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "text", + "content": "[21] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. arXiv preprint arXiv:2208.13721, 2022. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "text", + "content": "[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "text", + "content": "[23] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 557, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 600 + ], + "type": "text", + "content": "[24] Hao Lu, Zhiguo Cao, Yang Xiao, Bohan Zhuang, and Chunhua Shen. Tasselnet: counting maize tassels in the wild via local counts regression network. Plant methods, 13:1-17, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 657 + ], + "type": "text", + "content": "[25] Simon Madec, Xiuliang Jin, Hao Lu, Benoit De Solan, Shouyang Liu, Florent Duyme, Emmanuelle Heritier, and Frederic Baret. Ear density estimation from high resolution rgb imagery using deep learning technique. Agricultural and forest meteorology, 264:225-234, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[26] Daniel Onoro-Rubio and Roberto J López-Sastre. Towards perspective-free object counting with deep learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 615-629. Springer, 2016. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19595" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 675 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Dave-a detect-and-verify paradigm for low-shot counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23293–23302, 2024. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "text", + "content": "[29] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 282 + ], + "type": "text", + "content": "[30] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the\" edge\" of open-set object detection. arXiv preprint arXiv:2405.10300. 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 327 + ], + "type": "text", + "content": "[31] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "text", + "content": "[32] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9529–9538, 2022. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 439 + ], + "type": "text", + "content": "[33] Nikola Dukić, Alan Lukežić, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18872-18881, 2023. 1, 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 473 + ], + "type": "text", + "content": "[34] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 476, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 529 + ], + "type": "text", + "content": "[35] Zhicheng Wang, Liwen Xiao, Zhiguo Cao, and Hao Lu. Vision transformer off-the-shelf: A surprising baseline for few-shot class-agnostic counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5832–5840, 2024. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "text", + "content": "[36] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 294, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 629 + ], + "type": "text", + "content": "[37] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 833-841, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 632, + 294, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 294, + 675 + ], + "type": "text", + "content": "[38] XQ Zhou, YX Zou, and Y Wang. Accurate small object detection via density map aided saliency estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 425-429. IEEE, 2017. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19596" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_content_list.json b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..87cb2883ee27a707b1f22fcab8a4e640183c1fe2 --- /dev/null +++ b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_content_list.json @@ -0,0 +1,1707 @@ +[ + { + "type": "text", + "text": "Exploring Historical Information for RGBE Visual Tracking with Mamba", + "text_level": 1, + "bbox": [ + 125, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuanyu Sun $^{1*}$ , Jiqing Zhang $^{2*}$ , Yang Wang $^{1}$ , Huilin Ge $^{3\\dagger}$ , Qianchen Xia $^{4\\dagger}$ , Baocai Yin $^{1,5\\dagger}$ , Xin Yang $^{1\\dagger}$ $^{1}$ Key Laboratory of Social Computing and Cognitive Intelligence, Dalian University of Technology \n $^{2}$ Dalian Maritime University ${}^{3}$ Jiangsu University of Science and Technology \n $^{4}$ The Future Laboratory, Tsinghua University, Beijing, China ${}^{5}$ Beijing University of Technology", + "bbox": [ + 107, + 179, + 892, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Combining the advantages of conventional and event cameras for robust visual tracing has drawn extensive interest. However, existing tracking approaches heavily engage in complex cross-modal fusion modules, leading to higher computational complexity and training challenges. Besides, these methods generally ignore the effective integration of historical information, which is crucial to grasping the change in the target's appearance and motion trends. Given the recent advancements in Mamba's long-range modeling and linear complexity, we explore its potential in addressing the above issues in RGBE tracking tasks. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to achieve effective modality-selective fusion. This module can be seamlessly integrated into the encoding layer of prevalent Transformer-based backbones. Moreover, we further present a novel historical decoder that leverages Mamba's advanced long sequence modeling to effectively capture the target appearance changes with autoregressive queries. Extensive experiments show that our proposed approach achieves state-of-the-art performance on multiple challenging short-term and long-term RGBE benchmarks. Besides, the effectiveness of each key Mamba-based component of our approach is evidenced by our thorough ablation study. Code will be released at: https://github.com/scy0712/MamTrack", + "bbox": [ + 88, + 319, + 485, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 744, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual object tracking (VOT) is a fundamental task in computer vision[21-23, 30, 46] that focuses on estimating the position and shape of a target within video sequences based on its initial state [1, 3, 5, 17, 35, 36, 38, 47].Owing to the advantages of high temporal resolution and dynamic range offered by event cameras [33, 34, 40], there has been a", + "bbox": [ + 89, + 770, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e669452b3e32356a7b28c2128f777926545126898152cda417372448738fc1c1.jpg", + "image_caption": [ + "Figure 1. Overview of MamTrack and its two main components: (i) Multimodal Encoder is responsible for extracting and fusing features from RGB and event modalities; (ii) Historical Decoder is utilized to mine the temporal information of target appearance changes." + ], + "image_footnote": [], + "bbox": [ + 522, + 286, + 901, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "growing trend of integrating them with conventional RGB cameras for robust visual tracking tasks [10, 28, 32, 41-43, 48]. These RGBE multimodal approaches can significantly enhance tracking performance in challenging conditions, such as fast motion and limited illumination. Although being effective, the above multimodal methods expend a substantial amount of labor on developing overly intricate fusion algorithms. For example, FENet [41] employs a total of 8 self-attention and 4 complex cross-domain attention operations to integrate event and frame features at various levels. Implementing these sophisticated strategies will significantly increase the computational complexity and challenge of the training.", + "bbox": [ + 509, + 565, + 906, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moreover, the complex fusion designs make existing RGBE methods usually overlook the mining and application of historical temporal information, which is crucial for tracking as it captures changes in the target's state. Several RGB-only trackers [1, 4, 38] have been developed to update appearance information by replacing or fusing the templates. Occasionally, these straightforward methods ineffectively filter out unreliable updated templates and thus compromise the performance. To address this issue, some", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6500", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7863685608b5213b98bff92b07b281025fe99d931457345ba6c098abf899009a.jpg", + "image_caption": [ + "Figure 2. A comparison of our method with other recent SOTA trackers in long-term sequence. Benefiting from the proposed Historical Decoder, our method can effectively capture target appearance changes to guide robust long-term tracking even in partial occlusion scenes." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods [35, 36] incorporate an autoregressive scheme into the Transformer to transfer temporal information across multiple search frames. However, the Transformer structure lacks intrinsic designs for temporal sequential modeling, resulting in an inherent constraint, particularly in handling long sequences.", + "bbox": [ + 89, + 335, + 483, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Mamba [11], a novel State Space Model (SSM) [9, 12, 26, 37], has emerged as a highly promising approach for efficient long sequence modeling due to convolutional computation and linear complexity. Drawing inspiration from this advancement, as shown in Figure 1, we present mamba-based modules to facilitate robust RGB-E tracking by efficient multimodal fusion and effective historical information transmission. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to enhance the relationship between search and template tokens, as well as the complementarity between RGB and event modalities. This module is then easily inserted into the encoding layer of prevalent transformer-based backbones for combining the advantages of self-attention. Furthermore, we present a novel historical information decoder, which investigates the great potential of mamba in long sequence modeling to capture the appearance of targets by introducing a set of additional autoregressive queries.", + "bbox": [ + 89, + 429, + 483, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on multiple challenging RGBE benchmarks show the effectiveness of our approach, which outperforms existing state-of-the-art (SOTA) methods in both short-term and long-term tracking (see Figure 2). Ablation studies verify the importance of each key component. More importantly, we reveal the considerable potential of Mamba architecture in object tracking tasks. In summary, we make the following contributions:", + "bbox": [ + 89, + 717, + 483, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a novel and effective framework, dubbed MamTrack, a pioneering approach to incorporate the Mamba mechanism into RGBE multimodal tracking tasks.", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our proposed Mamba-based fusion module and historical decoder are capable of adaptively fusing multimodal information and effectively capturing the historical temporal cues of the target appearance changes, respectively.", + "- The proposed trackers achieve SOTA performance on multiple challenging short-term and long-term RGBE tracking benchmarks." + ], + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 212, + 663, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Mamba", + "text_level": 1, + "bbox": [ + 513, + 237, + 609, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The concept of the State Space Model (SSM) was first introduced in the S4 model [12], presenting a distinctive architecture capable of effectively modeling global information compared to traditional convolutional neural networks and Transformers. In their general form, SSM can be written as follows:", + "bbox": [ + 511, + 260, + 905, + 349 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {x} (t) = A (t) x (t) + B (t) u (t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 364, + 906, + 381 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ny (t) = C (t) x (t) + D (t) u (t). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 383, + 905, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This process maps a 1-D input sequence $u(t) \\in \\mathbb{R}$ to a 1-D output sequence $y(t) \\in \\mathbb{R}$ through an implicit $N$ -D latent state sequence $x(t) \\in \\mathbb{R}^N$ . Building on the S4 model, the S5 model [26] reduces complexity to a linear level, while H3 [9] introduces it into language model tasks. Mamba [11] introduces a data-dependent SSM layer and a selection mechanism using parallel scan (S6), achieving higher inference speed and overall metrics compared to equivalent-scale Transformers. Recently, Mamba has been employed in discriminative computer vision models and drawn great attention [7, 14, 18, 20, 49]. For example, ViM [49] applies Mamba to the Vision Transformer [8] architecture. VMamba [20] designs a cross-scan mechanism to bridge the gap between 1-D array scanning and 2-D plain traversing. Unlike previous efforts, we explore a novel architecture for RGB-E tracking based on Mamba, further unleashing its potential in visual tasks.", + "bbox": [ + 511, + 412, + 906, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. RGB-Event Single Object Tracking", + "text_level": 1, + "bbox": [ + 511, + 681, + 821, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, due to the unreliable nature of RGB-only data in challenging scenarios, more and more studies introduced event cameras to enhance tracking performance. These multimodal trackers can be primarily categorised into two categories based on the vision backbone employed. The first one [29, 41, 42] utilizes conventional CNNs to extract features from both RGB and event modalities, followed by a meticulously designed cross-modal attention mechanism for fusion. The second approach [28, 32] adopts the Transformer as the backbone, leveraging the multi-head attention mechanism to effectively model the indispensable correlation relationship between the RGB and the event modalities. To tackle the challenge of limited multimodal tracking data,", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6501", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26e30b1b0babfe40b62458605b793384273b405907bf015cc8227c1ec88641af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 94, + 898, + 265 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92017b34b2ff1b95432d3efbbcc301a324fff692c55030a9d144dbf5345f80f2.jpg", + "image_caption": [ + "Figure 3. (a) Details of our MamTrack, which employs the Multimodal Encoder to extract and fuse the features of RGB and event modals and introduces the Historical Decoder to capture the temporal cues related to target state changes; and two key components in the Fusion Mamba (FM) module: (b) Target-Aware Scan (TAS) and (c) Cross-Modality Scan (CMS) module." + ], + "image_footnote": [], + "bbox": [ + 99, + 266, + 491, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/816a3f267be7141c2e4ea43ffef8c285fce45617f0c9d1d2d61c93f97f03aa7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 267, + 903, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "recent studies [2, 13, 48] incorporate prompt tuning strategies into the Transformer architecture to transfer pretrained RGB-based trackers to multimodal data.", + "bbox": [ + 89, + 450, + 482, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The methods mentioned above primarily emphasise multimodal fusion and overlook the significance of mining historical information, which is crucial for multimodal tracking. In contrast, we explore the great potential of Mamba for multimodal tracking tasks. By combining the gating signals of two modalities in Mamba and leveraging the temporal modeling mechanism of Mamba, our approach is capable of concurrently achieving adaptive RGB-E multimodal fusion and efficient utilization of historical cues.", + "bbox": [ + 89, + 498, + 483, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Tracking with Historical Temporal Information", + "text_level": 1, + "bbox": [ + 89, + 648, + 482, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The historical information can provide valuable insights into the target's movements, appearance, and behaviour, thereby enhancing the tracker's confidence. Based on the number of search images input into the tracker, existing tracking approaches that leverage historical temporal cues can be categorised into two groups. The first one involves only inputting one search image and implementing a dynamic kernel filter [1, 6, 44] or a specialized update branch [4, 27, 38] to capture the target appearance changes with time and dynamically update template images during tracking. However, these methods generally involve handcrafted strategies, requiring tedious tuning for various application conditions. Another line of work [25, 35, 36, 47] takes multiple search sequences as input and then models the se", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quential evolution of target trajectories across frames by explicitly decoding the entire trajectory sequence step by step. Despite these advances, Transformer architectures employed in these methods are not intrinsically designed for temporal sequential modeling, which presents an inherent limitation. Competing with Transformer, Mamba [11] has shown great potential for long sequence modeling. Therefore, our research endeavors to design a novel Mamba-based historical decoder to effectively transmit temporal information in a parallel training manner.", + "bbox": [ + 511, + 450, + 906, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 617, + 648, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of our proposed MamTrack is illustrated in Figure 3, which has two branches with shared weights: RGB branch and Event branch. Each branch receives a template and multiple search images as inputs and corresponding features can be extracted by the Multimodal Encoder. The core of the Multimodal Encoder is our proposed Fusion Mamba (FM) module, which is inserted between Transformer encoder layers to leverage the advantages of both RGB and event modalities. In particular, the RGB modal offers abundant rich texture information, while the event modal is adept at handling challenging scenarios. The fused features are then input into our other primary component, the Historical Encoder, which specifically aims to capture the target appearance changes and motion trends by utilizing temporal cues. Finally, the fused features are multiplied with the decoder's output and then fed into the prediction head to locate the targets.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6502", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multimodal Encoder", + "text_level": 1, + "bbox": [ + 89, + 90, + 287, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [36], we employ a gradual downsampling network as multimodal encoder, which uses a $4 \\times 4$ patch size for patch embedding. Subsequently, we adopt a total of eight multi-layer perceptron (MLP) layers and two merging layers for gradual downsampling. After this, we obtain the RGB template tokens $F_{z}^{R} \\in \\mathbb{R}^{N_{z} \\times D}$ , RGB search tokens $F_{x}^{R} \\in \\mathbb{R}^{N_{x} \\times D}$ , event template tokens $F_{z}^{E} \\in \\mathbb{R}^{N_{z} \\times D}$ , and event search tokens $F_{x}^{E} \\in \\mathbb{R}^{N_{x} \\times D}$ . Here, $N_{z} = \\frac{H_{z}W_{z}}{16^{2}}$ ; $N_{x} = \\frac{H_{x}W_{x}}{16^{2}}$ ; $D = 512$ ; $H_{z}W_{z}$ and $H_{x}W_{x}$ represent the size of the template and search images, respectively. Then, the template tokens and search tokens are concatenated and fed into the $N$ -layer Transformer-based encoder for joint modeling. The operation of our encoder is defined as:", + "bbox": [ + 89, + 112, + 483, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {0} ^ {m} = \\mathcal {C} \\left(F _ {z} ^ {m}, F _ {x} ^ {m}\\right), \\quad m \\in \\{R G B, E v e n t \\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 318, + 480, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {n} ^ {R}, F _ {n} ^ {E} = \\mathcal {E} _ {n} \\left(F _ {n - 1} ^ {R}, F _ {n - 1} ^ {E}\\right), \\quad n = 1, \\dots , N \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 338, + 480, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C}$ is the element-wise concatenation; $\\mathcal{E}_n$ denotes $n_{th}$ encoder layer; $F_n^R$ and $F_n^E$ are the RGB and event outputs of the $n_{th}$ layer, respectively. The above operations facilitate the interaction between the template and search tokens during the RGB and event feature extraction. Next, we describe in detail our Mamba-based fusion module inserted between encoding layers.", + "bbox": [ + 89, + 363, + 483, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Fusion Mamba Module", + "text_level": 1, + "bbox": [ + 89, + 476, + 307, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed Fusion Mamba (FM) module aims to fuse multimodal semantic information. The key to FM is designed based on the following observations: (i) The interaction between template and search features is crucial as it can help the tracker to learn and prioritize the target-relevant search features; (ii) Directly combining RGB and event modals disregards cases where one of modalities cannot provide valuable cues. In the FM, we present Target-Aware Scan module (TAS) and Cross-Modality Scan module (CMS) to address the above issues, respectively.", + "bbox": [ + 89, + 498, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Target-Aware Scan Module (TAS). The TAS module enhances the interaction between template and search regions by leveraging Mamba's global modeling capability. Compared with the self-attention in Transformer, our TAS employs the bidirectional SSM mechanism [49] to estimate the relevance between all template tokens and each search token. The completeness of template cues can encourage our tracker to focus on informative search tokens. Given $F_{n}^{m} \\in \\{F_{n}^{R}, F_{n}^{E}\\}$ , as shown in Figure 3 (b), the TAS module is formulated as:", + "bbox": [ + 89, + 648, + 483, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {n} ^ {m} = \\mathcal {L} \\left(\\mathcal {N} \\left(F _ {n} ^ {m}\\right)\\right), Z _ {n} ^ {m} = \\sigma \\left(\\mathcal {L} \\left(\\mathcal {N} \\left(F _ {n} ^ {m}\\right)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 808, + 480, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {f} = \\mathcal {S S M} _ {f} \\left(\\psi \\left(X _ {n} ^ {m}\\right)\\right), Y _ {b} = \\mathcal {S S M} _ {b} \\left(\\psi \\left(X _ {n} ^ {m}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 827, + 480, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {n} ^ {m} = \\mathcal {L} \\left(\\mathcal {N} \\left(Z _ {n} ^ {m} \\odot Y _ {f} + Z _ {n} ^ {m} \\odot Y _ {b}\\right)\\right) + F _ {n} ^ {m}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 845, + 480, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is Linear projection; $\\mathcal{N}$ denotes LayerNorm operation; $\\sigma$ is SiLU activation function; $\\psi$ is 1D convolu", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion operation with the kernel size of 4; $\\mathcal{N}$ means elementwise multiplication; $X_{n}^{m}$ is the projection of $F_{n}^{m}$ ; $Z_{n}^{m}$ is gate signal; $Y_{f}$ and $Y_{b}$ are the outputs through forward and backward directions of SSM (i.e., $\\mathcal{SSM}_f$ and $\\mathcal{SSM}_b$ ). $H_{n}^{m} \\in \\{H_{n}^{R}, H_{n}^{E}\\}$ is the enhanced output of our TAS module in the $n_{th}$ encoder layer.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Modality Scan Module (CMS). CMS is responsible for improving the relevance and complementarity of multimodal features through a simple yet effective cross-fusion scheme. Different from TAS, we only input search tokens into CMS to prevent interference from template features. For simplicity, we will use the RGB domain as an example to describe the specific operation since RGB and event domains function similarly. Given RGB search tokens $H_{n}^{R,x}$ in the $n_{th}$ layer, as shown in Figure 3 (c), our CMS is defined as:", + "bbox": [ + 511, + 181, + 906, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {n} ^ {R, x} = \\mathcal {L} \\left(\\mathcal {N} \\left(H _ {n} ^ {R, x}\\right)\\right), Z _ {n} ^ {R, x} = \\sigma \\left(\\mathcal {L} \\left(\\mathcal {N} \\left(H _ {n} ^ {R, x}\\right)\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 339, + 903, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY ^ {\\prime} = \\mathcal {S S M} _ {f} \\left(\\psi \\left(X _ {n} ^ {R, x}\\right)\\right) + \\mathcal {S S M} _ {b} \\left(\\psi \\left(X _ {n} ^ {R, x}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 361, + 903, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {n} ^ {R, x} = \\mathcal {L} \\left(\\mathcal {N} \\left(Y ^ {\\prime} \\odot Z _ {n} ^ {R, x} + Y ^ {\\prime} \\odot Z _ {n} ^ {E, x}\\right)\\right) + H _ {n} ^ {R, x}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 381, + 903, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The modality-specific gating signals $Z_{n}^{R,x}$ and $Z_{n}^{E,x}$ are utilized to direct the interaction between modalities, thereby strengthening the primary modality while suppressing the degraded modality. With the assistance of our proposed TAS and CMS, the multimodal encoder can efficiently facilitate the interaction between template and search tokens, as well as between RGB and events modalities. The RGB and event features of the last encoder layer are added together to get the final fused feature $\\mathcal{F}_s$ , which will be used to enhance the functionality of the prediction head and historical decoder.", + "bbox": [ + 511, + 409, + 905, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Historical Decoder", + "text_level": 1, + "bbox": [ + 513, + 584, + 692, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Historical temporal information is crucial in visual object tracking as it provides the state changes of targets across frames. To fully exploit the temporal cues, we propose the Historical Decoder with $M$ layers, each layer consists of two key components: Historical State Aware (HSA) module and Feature Sequence Attention (FSA) module.", + "bbox": [ + 511, + 604, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Historical State Aware Module (HSA). Mamba offers a natural advantage over conventional CNNs and Transformers when it comes to long sequence modeling. We thus present Mamba-based HSA module to capture the target historical states of $N$ frames within a sequence into $N$ one-dimensional learnable vectors $Q(q_{1},\\ldots ,q_{N})\\in \\mathbb{R}^{N\\times D}$ . Mamba models the sequence and outputs $Q^{\\prime} = (q^{\\prime},\\dots,q_N^{\\prime})\\in \\mathbb{R}^{N\\times D}$ , this is", + "bbox": [ + 511, + 696, + 905, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {Q} = \\mathcal {S S M} (\\psi (\\mathcal {L} (Q))), Z _ {Q} = \\sigma (\\mathcal {L} (Q)), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 825, + 903, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {\\prime} = \\mathcal {L} \\left(\\mathcal {N} \\left(X _ {Q} \\odot Z _ {Q}\\right) + Q. \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 845, + 903, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output $Q'$ will be subsequently utilized as the input of FSA. It is important to highlight that, with the help of the", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6503", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c5dbc520d5001a40e06eeddb685d967ed7e92e3b414773af54ead4a3c3d60602.jpg", + "image_caption": [ + "Figure 4. Overview of our Historical Decoder with two key components: Historical State Aware (HSA) module, and Feature-Sequence Attention (FSA) module." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 477, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SSM mechanism, the hidden state generated by each query at the current moment will subsequently be transmitted and contribute to the estimation in the following moment. In this way, we guarantee the long-term viability of modeling historical cues. Compared to Transformer-based autoregressive models, this significantly increases both inference and training speeds.", + "bbox": [ + 89, + 362, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature-Sequence Attention Module (FSA). Our proposed FSA module aims to combine multimodal spatial features $\\mathcal{F}_s$ and historical temporal features $Q^{\\prime}$ to improve the confidence of the prediction head. Our FSA module employs the multi-head attention mechanism, where historical information sequences are used as queries and multimodal features are used as keys and values:", + "bbox": [ + 89, + 468, + 483, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {\\prime \\prime} = \\mathcal {N} \\left(\\mathcal {C} \\left(\\left(H _ {1}, \\dots , H _ {n _ {h}}\\right) W ^ {O}\\right)\\right) + Q ^ {\\prime},\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 580, + 416, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {i} = \\mathcal {A T T N} \\left(Q ^ {\\prime} W _ {i} ^ {Q}, \\mathcal {F} _ {s} W _ {i} ^ {K}, \\mathcal {F} _ {s} W _ {i} ^ {V}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 602, + 480, + 626 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A T T N} (q, k, v) = \\varphi \\left(\\frac {q k ^ {\\top}}{\\sqrt {d _ {k}}}\\right) v,\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 625, + 369, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi$ is the Softmax function; $W_{i}^{Q}\\in \\mathbb{R}^{d_{m}\\times d_{k}}$ , $W_{i}^{K}\\in \\mathbb{R}^{d_{m}\\times d_{k}}$ , $W_{i}^{V}\\in \\mathbb{R}^{d_{m}\\times d_{v}}$ , and $W^{O}\\in \\mathbb{R}^{n_{h}\\times d_{m}}$ are all learnable parameters. In our method, we employ multihead attention with 8 heads, i.e., $n_h = 8$ , $d_m = 512$ , and $d_k = d_v = \\frac{d_m}{8} = 64$ . Then $Q''$ is fed into the following decoder layer to generate the final temporal queries $Q$ after being processed by the Feed Forward Network (FFN). Finally, we further combine the spatial and temporal information without introducing any parameters by:", + "bbox": [ + 89, + 667, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o u t} = \\mathcal {F} _ {s} \\odot \\mathcal {Q} ^ {\\top}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 813, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{out}$ represents the multimodal features enriched with historical temporal information, and highlights the position where the target may be located. Then it will be fed into the tracking head for the target bbox prediction.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Head and Loss", + "text_level": 1, + "bbox": [ + 513, + 90, + 665, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a center-based head as regular [39] to predict the center position and scale of the target. The prediction head outputs include the classification score map, the bounding box size, and the offset size. We use focal loss [19] and GIoU loss [24] for training the network. The total loss $L$ is:", + "bbox": [ + 511, + 112, + 906, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {c l s} + \\lambda_ {i o u} L _ {i o u} + \\lambda_ {L _ {1}} L _ {L _ {1}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 200, + 906, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{cls}$ denotes the weighted focal loss for classification, the generalized GIoU loss $L_{iou}$ and L1 loss $L_{1}$ are adopted for bounding box regression, $\\lambda_{iou} = 2$ and $\\lambda_{L_1} = 5$ are the hyperparameters.", + "bbox": [ + 511, + 229, + 905, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 306, + 645, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 330, + 720, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We implemented MamTrack using PyTorch and trained it on 8 NVIDIA RTX 4090 GPUs with a batch size of 16. We employed the AdamW optimizer with a weight decay of $10^{-4}$ and a learning rate of $4 \\times 10^{-4}$ . We utilized pretrained HiViT-Base [45] as the Transformer encoder with 20 layers. Fine-tuning MamTrack on the RGB-Event training set took 50 epochs. Our proposed Fusion Mamba module is inserted into the 12th, 15th, and 18th layers. The decoder consists of 3 layers (i.e., $M = 3$ ). Besides, we utilize the Hamming window to incorporate positional priors.", + "bbox": [ + 511, + 352, + 906, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt three widely used metrics for quantitatively evaluating the tracking performance: Success Rate (SR), Precision Rate (PR), and Overlap Precision $(\\mathrm{OP}_T)$ . The SR focuses on the frame where the overlap between ground truth and predicted bounding box is larger than a threshold; $\\mathrm{OP}_T$ matches SR but with a specific overlap threshold $T$ . The PR metric counts the frame on which the center distance between the ground truth and predicted bounding box is within a given threshold.", + "bbox": [ + 511, + 518, + 905, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 511, + 666, + 821, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed approach, we compare our tracker with representative state-of-the-art trackers on three widely utilized benchmarks: VisEvent [29], FELT [31], and FE108 [41].", + "bbox": [ + 511, + 688, + 905, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VisEvent. VisEvent is a large-scale visible-event tracking benchmark, which contains 500 training and 320 testing sequences with 17 attributes. As shown in Figure 5(a), our MamTrack achieves the top SR and PR of $61.6\\%$ and $79.2\\%$ , respectively, which beats all single- and multimodal SOTA trackers. For example, our method outperforms the pure Transformer-based temporal tracker AQA-Track [36] by $2.9\\%$ on SR and multimodal tracker SDSTrack [13] by $1.9\\%$ . Figure 6 further provides the tracking performance under 17 challenging attributes. We can see", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6504", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1feae14f806188fdf8dacf77337d8e2b57149d4a27936c4785feaafe897ed536.jpg", + "image_caption": [ + "(a) Success (left) and Precision (right) plot on VisEvent" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 285, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57337a777f67e11710bc187f7d303aae6a5f49de325ba53e43414584b266986e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 88, + 482, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/418ef6ec29d75a1b519eb585be683d153a9c365f80e1d86bb257bce9a788e516.jpg", + "image_caption": [ + "(b) Success (left) and Precision (right) plot on FELT" + ], + "image_footnote": [], + "bbox": [ + 94, + 223, + 287, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1a8f87336485c44e1cc0a2b1e1dea274e2f514837c37cb96d35a89c65f533e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 223, + 482, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a6d6f788155dc793cf33733125e6018267ba3e9e75c07a2c8f6c083a67a75c2.jpg", + "image_caption": [ + "Figure 5. Comparison of Success (left) and Precision (right) plots on VisEvent (a) and FELT (b)." + ], + "image_footnote": [], + "bbox": [ + 91, + 416, + 287, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f75e7012294b9642cb18319095b63aae79476f57174c17e9ee279cf8621a70e.jpg", + "image_caption": [ + "Figure 6. Comparison of SR (left) and PR (right) across various attributes using our approach against four top trackers on VisEvent, highlights the performance of Ours/Lowest." + ], + "image_footnote": [], + "bbox": [ + 287, + 417, + 482, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that our approach still fares best in all degraded conditions, demonstrating its ability to effectively fuse the strengths of RGB and Events.", + "bbox": [ + 89, + 626, + 483, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FELT. FELT is a large-scale long-term frame-event single object tracking dataset, which contains 742 videos, 1,594,474 RGB frames and corresponding events. The results in Figure 5(b) show that our MamTrack achieves a $48.9\\%$ overall SR and $60.8\\%$ PR. The results illustrate that our approach achieves effective long-term tracking by exploiting the historical temporal modeling capability of Mamba. We also offer an attribute-based evaluation in the supplementary material for further analysis, our MamTrack achieves the best performance on most attribute splits.", + "bbox": [ + 89, + 672, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FE108. FE108 is a visible-event multimodal tracking dataset of indoor scenes, which contains 76 videos for training and 32 for testing. As shown in Table 1, our MamTrack outperforms the previous best tracker ViPT by $0.6\\%$ in SR and $0.4\\%$ in PR. It demonstrates that our approach is also", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5b6f87f489cdee60f854a81c191711cd15692b961594f00e50b113a9f7a35f13.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ATOMDiMPPrDiMPFENetMixFormerTransT
52.8/80.053.4/88.255.2/86.863.4/92.460.4/87.856.7/89.0
SiamFc++STARKOSTrackViPTHDETrackOurs
54.5/85.355.4/83.757.1/89.365.8/93.859.8/92.266.4/94.2
", + "bbox": [ + 514, + 89, + 908, + 164 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f81d199d5e7670be4dd0eb9be379aa25f6e2cd3241359a066f767f8fbb4bca60.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of SR(\\%) / PR(\\%) against SOTA trackers on the FE108 dataset." + ], + "table_footnote": [], + "table_body": "
#RGBEventTASCMSCFFSAHSASR(%)PR(%)
1X40.556.1
2X59.075.8
3X60.778.1
4X60.277.4
5X60.077.1
6X60.678.0
7X60.477.8
861.679.2
", + "bbox": [ + 517, + 218, + 903, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation study for important components on VisEvent. Blank denotes the component is used by default, while $\\mathbf{x}$ represents the component is removed.", + "bbox": [ + 511, + 369, + 905, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "highly competitive for short-term tracking scenarios.", + "bbox": [ + 511, + 440, + 859, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 511, + 468, + 769, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of Multimodal Input. We assess the effectiveness of multimodal fusion by separately utilizing event-only and RGB-only inputs. As shown in Table 2 #1 and #2, the tracking results with unimodal input are significantly worse than those with multimodal input, demonstrating that our method effectively utilizes both modalities' advantages to improve tracking under degraded conditions.", + "bbox": [ + 511, + 489, + 906, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of Fusion Mamba Module. There are two key mamba-based components in our Fusion Mamba (FM) module: (i) Target-Aware Scan (TAS) facilitates the interaction between the template and search tokens; (ii) Cross-Modality Scan (CMS) focuses on the fusion of event and RGB modalities in an adaptive manner. To verify their effectiveness, we modify the original model by dropping each of the components and retraining the modified models. The tracking performance in Table 2 #3 and #4 has shown a decrease to varying degrees, indicating the effectiveness and rationality of our designed FM module. In CMS, the gate-based operation Cross-Fusion (CF) is the primary operation that supports the interaction between the two modalities. We thus eliminate the CF to showcase its indispensability. #5 shows that when the CF is removed, the SR drops by $1.6\\%$ . This illustrates that the deep interaction among features from both modalities plays a key role.", + "bbox": [ + 511, + 598, + 908, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One possible factor that impacts the contribution of FM to our tracker is the location of FM insertion in the encoder. To verify this, several additional sets of experiments were", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6505", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/be27743dc075a58a3af74a256523a95d52bd8dd71cb95fed2f56557cd1cca189.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LocationSR (%)PR(%)LocationSR (%)PR(%)
1260.077.014-17-2061.178.4
12-1561.078.312-13-1460.177.3
12-15-18 (Ours)61.679.218-19-2060.677.9
", + "bbox": [ + 91, + 88, + 491, + 165 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1e3ccfc731bfeca43d3f1d708cd007c0a5c65499f90a52ac55afde9b5433426.jpg", + "table_caption": [ + "Table 3. Comparison of different FM module locations on VisEvent. $x - y - z$ represents our FM module is inserted into the $x$ th, $y$ th, and $z$ th encoder layers." + ], + "table_footnote": [], + "table_body": "
Type of HSANumber of LayersSR(%)PR(%)FPS
None059.976.820.7
Mamba160.578.119.9
361.679.218.1
660.978.116.6
Transformer361.078.416.9
", + "bbox": [ + 101, + 233, + 473, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "conducted. As shown in Table 3, compared to other 3-stage insertion styles, the 12-15-18 style adopted in this paper provides the best performance, which can be attributed to our insertion strategy that focuses on both low-level texture features and high-level semantic features. Moreover, the tracking performance exhibits a progressive improvement from 12 to 12-15 and then to 12-15-18 style, revealing the growing synergy between RGB and event modalities facilitated by our FM module.", + "bbox": [ + 88, + 414, + 482, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Historical Decoder. The Historical Decoder is designed to capture and transmit historical information during tracking, which consists of two main components: Historical State Aware (HSA) and Feature Sequence Attention (FSA) modules. As shown in Table 2 #6, when the FSA is removed, the SR score drops by $1.0\\%$ , showing the importance of multimodal features for exploiting historical information. The performance drops by $1.2\\%$ when the HSA is removed as shown in #7. This illustrates the effectiveness of Mamba for modeling long sequences.", + "bbox": [ + 88, + 550, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also examine the influence of the number of historical decoder layers on tracking performance. As shown in Table 4, the utilization of the historical decoder greatly enhances the results, with the most performance observed when employing 3 decoder layers. To further verify the effectiveness of the Mamba in the historical decoder, we replace the Mamba architecture with Transformer. Table 4 shows our method with Mamba excels in both performance and speed due to Mamba's linear complexity and its capability to model long sequences.", + "bbox": [ + 88, + 702, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the Length of Historical Information. As shown in Table 5, the model's performance improves as the range of historical information increases from 1 to 4. This", + "bbox": [ + 88, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/495246e818fad6fbb4288ba81e3cc9aa4424fd063192fd84f8756a70bfcf7ec4.jpg", + "table_caption": [ + "Table 4. Ablation studies on the type and number of decoder layers." + ], + "table_footnote": [], + "table_body": "
LengthSR(%)PR(%)LengthSR(%)PR(%)
160.177.5859.876.9
260.477.91659.776.6
461.679.2
", + "bbox": [ + 542, + 88, + 880, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Impact of the length of historical information.", + "bbox": [ + 545, + 175, + 872, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1050d100ba3e2fda10a91f18d4754b16d31e34d9cf9572058ba0aba5d847899.jpg", + "image_caption": [ + "Figure 7. Comparison of SR scores for long-term tracking on VisEvent, where the horizontal axis is test sequence length." + ], + "image_footnote": [], + "bbox": [ + 514, + 204, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicates that the model can effectively capture the target's movement trends and appearance changes, resulting in improved performance. However, when the historical information range is extended to 8 and 16, the model's performance slightly declines. This decrease is due to the limited number of video sequences used in each iteration, which impairs the model's generalization ability.", + "bbox": [ + 511, + 443, + 906, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness for Long-term Tracking. To further demonstrate the effectiveness of our Mamba-based historical decoder in capturing temporal information, we analyze its performance in long-term tracking. Specifically, we evaluate the tracking performance of sequences containing more than $i$ frames on the VisEvent individually, where $i \\in \\{300, 500, 700, 1000, 1500, 2000\\}$ . Figure 7 shows that our MamTrack outperforms other SOTA Transformer-based temporal trackers and multimodal trackers by a large margin in SR across all settings. This suggests our Mamba-based approach proves superior in capturing long-term cues compared to the Transformer-based autoregressive scheme.", + "bbox": [ + 511, + 550, + 908, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Performance on RGB-T Tracking Dataset", + "text_level": 1, + "bbox": [ + 511, + 741, + 870, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate the generalization ability of our MamTrack, we conducted additional experiments on the RGB-T tracking task. Following the common settings used in previous work [48], we trained our tracker on the LasHeR [16] training set and tested it on both the LasHeR test set and the RGBT234 [15] dataset.", + "bbox": [ + 511, + 763, + 906, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on LasHeR Dataset. LasHeR is the largest RGB-T tracking dataset at present, containing 1224 aligned video sequences with diverse attribute annotations. Of", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6506", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b339d6bb256b02c7a923b689f16a35ce7876feb9c6ec7b5e3c4d11f296cb7585.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LasHeRmfDimp 34.3/ 44.7SGT++ 25.1/36.5OSTrack 41.2/51.5ProTrack 42.0/53.8HMFT 31.3/ 43.6APFNet 36.2/50.0CEUTrack 49.6/63.5ViPT 52.5/65.1Ours 54.2/67.4
RGBT234mfDimp 42.8/ 64.6CMPP 57.7/82.3OSTrack 54.9/72.9ProTrack 57.9/82.7HMFT 56.8/78.8APFNet 57.9/82.7MFGNet 53.5/78.3ViPT 61.7/ 83.5Ours 62.4/84.4
", + "bbox": [ + 91, + 88, + 934, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Performance comparison of SR(%) / PR(%) on LasHeR and RGBT234 dataset.", + "bbox": [ + 236, + 175, + 758, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/699df7a95cd61d3cdc1e7f38b9eb4821bafe859cde968c155142f372a76a7921.jpg", + "image_caption": [ + "Figure 8. Visual comparison of our MamTrack against other SOTA trackers under 4 challenging conditions on VisEvent." + ], + "image_footnote": [], + "bbox": [ + 93, + 218, + 480, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9703a62830c13349d42c2772401979646b81cbbef28cbbdd060325feeef42e78.jpg", + "image_caption": [ + "Figure 9. The comparison attention map in the MamTrack. The first row is the ground truth; the second and third rows display the attention maps in the multimodal encoder and historical decoder, respectively." + ], + "image_footnote": [], + "bbox": [ + 93, + 474, + 482, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "these, 245 sequences are designated as the testing set, while the remaining are used for training. As shown in Table 6, MamTrack performs exceptionally well on RGB-T tracking tasks, effectively leveraging the advantages of both thermal and visible information. This indicates its potential for other multimodal tracking tasks.", + "bbox": [ + 89, + 722, + 482, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on RGBT234 Dataset. RGBT234 is a large-scale RGB-T tracking dataset comprising 234 pairs of videos. As shown in Table 6, our MamTrack surpasses the most advantaged tracker ViPT [48] by $0.7\\%$ in Success Rate (SR) and $0.9\\%$ in Precision Rate (PR).", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 511, + 215, + 650, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first provide a visual comparison across four challenging conditions on the VisEvent dataset. As shown in Figure 8, our approach can accurately track the target under all conditions. In particular, our method effectively fuses the advantages of RGB and event to enhance tracker confidence when the target is in scenes with motion blur or fast motion. Besides, in situations where the target is partially occluded or the background is complex, the historical cues encourage our method to recall the target's appearance, thereby minimizing environmental interference.", + "bbox": [ + 511, + 239, + 906, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we also visualize the attention maps of the encoder and decoder to demonstrate the effectiveness of historical information. As shown in Figure 9, guided by historical information, our MamTrack pays more attention to the target's location under challenging conditions.", + "bbox": [ + 511, + 393, + 906, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 492, + 633, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose MamTrack, a novel approach jointly leverages multimodal spatial and historical temporal information for RGBE object tracking. Our historical decoder, based on the innovative Mamba architecture, effectively captures target appearance changes and motion trends, enabling the model to locate the target within complex backgrounds. We also introduce a simple yet effective Mamba-based fusion method for integrating cross-modality cues. Extensive evaluations demonstrate that our method achieves competitive performance compared to previous SOTA trackers. However, we primarily focuses on the effectiveness of Mamba in modality fusion, but not leverage the high temporal resolution characteristics of event data. Future research will explore this aspect to enhance the performance of the tracker.", + "bbox": [ + 511, + 520, + 906, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by the National Key Research and Development Program of China (No. 2022ZD0210500), the National Natural Science Foundation of China under Grant 62441216/62332019, the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01), the Ningbo Major Research and Development Plan Project of China (No. 2023Z225), Beijing Natural Science Foundation of China(Grant No.3254040), National Key Laboratory of Human Factors Engineering and China Postdoctoral Science Foundation under Grant (No. 2024M760315).", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6507", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6182-6191, 2019. 1, 3", + "[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 3", + "[3] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 1", + "[4] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13608-13618, 2022. 1, 3", + "[5] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4660-4669, 2019. 1", + "[6] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7183-7192, 2020. 3", + "[7] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-Mamba for Cross-modality Object Detection, 2024. 2", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2", + "[9] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[10] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. IEEE Transactions on Image Processing, 32:6129-6141, 2023. 1", + "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3", + "[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2022. 2", + "[13] Xiaojun Hou, Jiazheng Xing, Yijie Qian, Yaowei Guo, Shuo Xin, Junhao Chen, Kai Tang, Mengmeng Wang, Zhengkai Jiang, Liang Liu, et al. Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking. In" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26551-26561, 2024. 3, 5", + "[14] Ju Huang, Shiao Wang, Shuai Wang, Zhe Wu, Xiao Wang, and Bo Jiang. Mamba-fetrack: Frame-event tracking via state space model. arXiv preprint arXiv:2404.18174, 2024.2", + "[15] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. Rgb-t object tracking: Benchmark and baseline. Pattern Recognition, 96:106977, 2019. 7", + "[16] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for rgbt tracking. IEEE Transactions on Image Processing, 31:392-404, 2021. 7", + "[17] Kang Li, Fazhi He, Haiping Yu, and Xiao Chen. A parallel and robust object tracking approach synthesizing adaptive bayesian learning and improved incremental subspace learning. Frontiers of Computer Science, 13(5):1116, 2019. 1", + "[18] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2", + "[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[20] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual State Space Model, 2024. 2", + "[21] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3044-3053, 2021. 1", + "[22] Haiyang Mei, Bo Dong, Wen Dong, Jiaxi Yang, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Glass segmentation using intensity and spectral polarization cues. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12622-12631, 2022.", + "[23] Yu Qiao, Bo Dong, Ao Jin, Yu Fu, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Multiview spectral polarization propagation for video glass segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23218-23228, 2023. 1", + "[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5", + "[25] Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Shengping Zhang, and Xianxian Li. Explicit visual prompts for visual object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4838-4846, 2024. 3", + "[26] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6508", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In The Eleventh International Conference on Learning Representations, 2023. 2", + "[27] Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2321-2329, 2023. 3", + "[28] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1, 2", + "[29] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 2023. 2, 5", + "[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 7(1):30-40, 2023. 1", + "[31] Xiao Wang, Ju Huang, Shiao Wang, Chuanming Tang, Bo Jiang, Yonghong Tian, Jin Tang, and Bin Luo. Long-term frame-event visual tracking: Benchmark dataset and baseline. arXiv preprint arXiv:2403.05839, 2024. 5", + "[32] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19248-19257, 2024. 1, 2", + "[33] Yang Wang, Bo Dong, Yuji Zhang, Yunduo Zhou, Haiyang Mei, Ziqi Wei, and Xin Yang. Event-enhanced multi-modal spiking neural network for dynamic obstacle avoidance. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3138-3148, 2023. 1", + "[34] Yang Wang, Haiyang Mei, Qirui Bao, Ziqi Wei, Mike Zheng Shou, Haizhou Li, Bo Dong, and Xin Yang. Apprenticeship-inspired elegance: synergistic knowledge distillation empowers spiking neural networks for efficient single-eye emotion recognition. arXiv preprint arXiv:2407.09521, 2024. 1", + "[35] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9697-9706, 2023. 1, 2, 3", + "[36] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19300-19309, 2024. 1, 2, 3, 4, 5", + "[37] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2", + "[38] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10448-10457, 2021. 1, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European conference on computer vision, pages 341-357. Springer, 2022. 5", + "[40] Haiwei Zhang, Jiqing Zhang, Bo Dong, Pieter Peers, Wenwei Wu, Xiaopeng Wei, Felix Heide, and Xin Yang. In the blink of an eye: Event-based emotion recognition. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1", + "[41] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13043-13052, 2021. 1, 2, 5", + "[42] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9781-9790, 2023. 2", + "[43] Jiqing Zhang, Bo Dong, Yingkai Fu, Yuanchen Wang, Xiaopeng Wei, Baocai Yin, and Xin Yang. A universal event-based plug-in module for visual object tracking in degraded conditions. International Journal of Computer Vision, 132 (5):1857-1879, 2024. 1", + "[44] Lichao Zhang, Abel Gonzalez-Garcia, Joost Van De Weijer, Martin Danelljan, and Fahad Shahbaz Khan. Learning the model update for siamese trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4010-4019, 2019. 3", + "[45] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 5", + "[46] Aihua Zheng, Juncong Liu, Zi Wang, Lili Huang, Chenglong Li, and Bing Yin. Visible-infrared person re-identification via specific and shared representations learning. Visual Intelligence, 1(1):29, 2023. 1", + "[47] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7588-7596, 2024. 1, 3", + "[48] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9516-9526, 2023. 1, 3, 7, 8", + "[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model, 2024. 2, 4" + ], + "bbox": [ + 516, + 92, + 906, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6509", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_model.json b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..358d71fbe3953b15c288728e8cc16aafa5ecb5ff --- /dev/null +++ b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "Exploring Historical Information for RGBE Visual Tracking with Mamba" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.18, + 0.893, + 0.253 + ], + "angle": 0, + "content": "Chuanyu Sun\\(^{1*}\\), Jiqing Zhang\\(^{2*}\\), Yang Wang\\(^{1}\\), Huilin Ge\\(^{3\\dagger}\\), Qianchen Xia\\(^{4\\dagger}\\), Baocai Yin\\(^{1,5\\dagger}\\), Xin Yang\\(^{1\\dagger}\\) \n\\(^{1}\\)Key Laboratory of Social Computing and Cognitive Intelligence, Dalian University of Technology \n\\(^{2}\\)Dalian Maritime University \\({}^{3}\\)Jiangsu University of Science and Technology \n\\(^{4}\\)The Future Laboratory, Tsinghua University, Beijing, China \\({}^{5}\\)Beijing University of Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.486, + 0.714 + ], + "angle": 0, + "content": "Combining the advantages of conventional and event cameras for robust visual tracing has drawn extensive interest. However, existing tracking approaches heavily engage in complex cross-modal fusion modules, leading to higher computational complexity and training challenges. Besides, these methods generally ignore the effective integration of historical information, which is crucial to grasping the change in the target's appearance and motion trends. Given the recent advancements in Mamba's long-range modeling and linear complexity, we explore its potential in addressing the above issues in RGBE tracking tasks. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to achieve effective modality-selective fusion. This module can be seamlessly integrated into the encoding layer of prevalent Transformer-based backbones. Moreover, we further present a novel historical decoder that leverages Mamba's advanced long sequence modeling to effectively capture the target appearance changes with autoregressive queries. Extensive experiments show that our proposed approach achieves state-of-the-art performance on multiple challenging short-term and long-term RGBE benchmarks. Besides, the effectiveness of each key Mamba-based component of our approach is evidenced by our thorough ablation study. Code will be released at: https://github.com/scy0712/MamTrack" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.745, + 0.222, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Visual object tracking (VOT) is a fundamental task in computer vision[21-23, 30, 46] that focuses on estimating the position and shape of a target within video sequences based on its initial state [1, 3, 5, 17, 35, 36, 38, 47].Owing to the advantages of high temporal resolution and dynamic range offered by event cameras [33, 34, 40], there has been a" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.287, + 0.903, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.533 + ], + "angle": 0, + "content": "Figure 1. Overview of MamTrack and its two main components: (i) Multimodal Encoder is responsible for extracting and fusing features from RGB and event modalities; (ii) Historical Decoder is utilized to mine the temporal information of target appearance changes." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.566, + 0.907, + 0.763 + ], + "angle": 0, + "content": "growing trend of integrating them with conventional RGB cameras for robust visual tracking tasks [10, 28, 32, 41-43, 48]. These RGBE multimodal approaches can significantly enhance tracking performance in challenging conditions, such as fast motion and limited illumination. Although being effective, the above multimodal methods expend a substantial amount of labor on developing overly intricate fusion algorithms. For example, FENet [41] employs a total of 8 self-attention and 4 complex cross-domain attention operations to integrate event and frame features at various levels. Implementing these sophisticated strategies will significantly increase the computational complexity and challenge of the training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Moreover, the complex fusion designs make existing RGBE methods usually overlook the mining and application of historical temporal information, which is crucial for tracking as it captures changes in the target's state. Several RGB-only trackers [1, 4, 38] have been developed to update appearance information by replacing or fusing the templates. Occasionally, these straightforward methods ineffectively filter out unreliable updated templates and thus compromise the performance. To address this issue, some" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6500" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.237, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Figure 2. A comparison of our method with other recent SOTA trackers in long-term sequence. Benefiting from the proposed Historical Decoder, our method can effectively capture target appearance changes to guide robust long-term tracking even in partial occlusion scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.428 + ], + "angle": 0, + "content": "methods [35, 36] incorporate an autoregressive scheme into the Transformer to transfer temporal information across multiple search frames. However, the Transformer structure lacks intrinsic designs for temporal sequential modeling, resulting in an inherent constraint, particularly in handling long sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.717 + ], + "angle": 0, + "content": "Recently, Mamba [11], a novel State Space Model (SSM) [9, 12, 26, 37], has emerged as a highly promising approach for efficient long sequence modeling due to convolutional computation and linear complexity. Drawing inspiration from this advancement, as shown in Figure 1, we present mamba-based modules to facilitate robust RGB-E tracking by efficient multimodal fusion and effective historical information transmission. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to enhance the relationship between search and template tokens, as well as the complementarity between RGB and event modalities. This module is then easily inserted into the encoding layer of prevalent transformer-based backbones for combining the advantages of self-attention. Furthermore, we present a novel historical information decoder, which investigates the great potential of mamba in long sequence modeling to capture the appearance of targets by introducing a set of additional autoregressive queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Extensive experiments on multiple challenging RGBE benchmarks show the effectiveness of our approach, which outperforms existing state-of-the-art (SOTA) methods in both short-term and long-term tracking (see Figure 2). Ablation studies verify the importance of each key component. More importantly, we reveal the considerable potential of Mamba architecture in object tracking tasks. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "- We present a novel and effective framework, dubbed MamTrack, a pioneering approach to incorporate the Mamba mechanism into RGBE multimodal tracking tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "- Our proposed Mamba-based fusion module and historical decoder are capable of adaptively fusing multimodal information and effectively capturing the historical temporal cues of the target appearance changes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.907, + 0.198 + ], + "angle": 0, + "content": "- The proposed trackers achieve SOTA performance on multiple challenging short-term and long-term RGBE tracking benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.213, + 0.664, + 0.228 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.238, + 0.611, + 0.253 + ], + "angle": 0, + "content": "2.1. Mamba" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.906, + 0.351 + ], + "angle": 0, + "content": "The concept of the State Space Model (SSM) was first introduced in the S4 model [12], presenting a distinctive architecture capable of effectively modeling global information compared to traditional convolutional neural networks and Transformers. In their general form, SSM can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.365, + 0.907, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\dot {x} (t) = A (t) x (t) + B (t) u (t), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.384, + 0.906, + 0.401 + ], + "angle": 0, + "content": "\\[\ny (t) = C (t) x (t) + D (t) u (t). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.672 + ], + "angle": 0, + "content": "This process maps a 1-D input sequence \\( u(t) \\in \\mathbb{R} \\) to a 1-D output sequence \\( y(t) \\in \\mathbb{R} \\) through an implicit \\( N \\)-D latent state sequence \\( x(t) \\in \\mathbb{R}^N \\). Building on the S4 model, the S5 model [26] reduces complexity to a linear level, while H3 [9] introduces it into language model tasks. Mamba [11] introduces a data-dependent SSM layer and a selection mechanism using parallel scan (S6), achieving higher inference speed and overall metrics compared to equivalent-scale Transformers. Recently, Mamba has been employed in discriminative computer vision models and drawn great attention [7, 14, 18, 20, 49]. For example, ViM [49] applies Mamba to the Vision Transformer [8] architecture. VMamba [20] designs a cross-scan mechanism to bridge the gap between 1-D array scanning and 2-D plain traversing. Unlike previous efforts, we explore a novel architecture for RGB-E tracking based on Mamba, further unleashing its potential in visual tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.822, + 0.699 + ], + "angle": 0, + "content": "2.2. RGB-Event Single Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Recently, due to the unreliable nature of RGB-only data in challenging scenarios, more and more studies introduced event cameras to enhance tracking performance. These multimodal trackers can be primarily categorised into two categories based on the vision backbone employed. The first one [29, 41, 42] utilizes conventional CNNs to extract features from both RGB and event modalities, followed by a meticulously designed cross-modal attention mechanism for fusion. The second approach [28, 32] adopts the Transformer as the backbone, leveraging the multi-head attention mechanism to effectively model the indispensable correlation relationship between the RGB and the event modalities. To tackle the challenge of limited multimodal tracking data," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.095, + 0.899, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.267, + 0.492, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.268, + 0.904, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 3. (a) Details of our MamTrack, which employs the Multimodal Encoder to extract and fuse the features of RGB and event modals and introduces the Historical Decoder to capture the temporal cues related to target state changes; and two key components in the Fusion Mamba (FM) module: (b) Target-Aware Scan (TAS) and (c) Cross-Modality Scan (CMS) module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.496 + ], + "angle": 0, + "content": "recent studies [2, 13, 48] incorporate prompt tuning strategies into the Transformer architecture to transfer pretrained RGB-based trackers to multimodal data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.635 + ], + "angle": 0, + "content": "The methods mentioned above primarily emphasise multimodal fusion and overlook the significance of mining historical information, which is crucial for multimodal tracking. In contrast, we explore the great potential of Mamba for multimodal tracking tasks. By combining the gating signals of two modalities in Mamba and leveraging the temporal modeling mechanism of Mamba, our approach is capable of concurrently achieving adaptive RGB-E multimodal fusion and efficient utilization of historical cues." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.681 + ], + "angle": 0, + "content": "2.3. Tracking with Historical Temporal Information" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The historical information can provide valuable insights into the target's movements, appearance, and behaviour, thereby enhancing the tracker's confidence. Based on the number of search images input into the tracker, existing tracking approaches that leverage historical temporal cues can be categorised into two groups. The first one involves only inputting one search image and implementing a dynamic kernel filter [1, 6, 44] or a specialized update branch [4, 27, 38] to capture the target appearance changes with time and dynamically update template images during tracking. However, these methods generally involve handcrafted strategies, requiring tedious tuning for various application conditions. Another line of work [25, 35, 36, 47] takes multiple search sequences as input and then models the se" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.603 + ], + "angle": 0, + "content": "quential evolution of target trajectories across frames by explicitly decoding the entire trajectory sequence step by step. Despite these advances, Transformer architectures employed in these methods are not intrinsically designed for temporal sequential modeling, which presents an inherent limitation. Competing with Transformer, Mamba [11] has shown great potential for long sequence modeling. Therefore, our research endeavors to design a novel Mamba-based historical decoder to effectively transmit temporal information in a parallel training manner." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.649, + 0.635 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The overall architecture of our proposed MamTrack is illustrated in Figure 3, which has two branches with shared weights: RGB branch and Event branch. Each branch receives a template and multiple search images as inputs and corresponding features can be extracted by the Multimodal Encoder. The core of the Multimodal Encoder is our proposed Fusion Mamba (FM) module, which is inserted between Transformer encoder layers to leverage the advantages of both RGB and event modalities. In particular, the RGB modal offers abundant rich texture information, while the event modal is adept at handling challenging scenarios. The fused features are then input into our other primary component, the Historical Encoder, which specifically aims to capture the target appearance changes and motion trends by utilizing temporal cues. Finally, the fused features are multiplied with the decoder's output and then fed into the prediction head to locate the targets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6502" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.289, + 0.105 + ], + "angle": 0, + "content": "3.1. Multimodal Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.312 + ], + "angle": 0, + "content": "Following [36], we employ a gradual downsampling network as multimodal encoder, which uses a \\(4 \\times 4\\) patch size for patch embedding. Subsequently, we adopt a total of eight multi-layer perceptron (MLP) layers and two merging layers for gradual downsampling. After this, we obtain the RGB template tokens \\(F_{z}^{R} \\in \\mathbb{R}^{N_{z} \\times D}\\), RGB search tokens \\(F_{x}^{R} \\in \\mathbb{R}^{N_{x} \\times D}\\), event template tokens \\(F_{z}^{E} \\in \\mathbb{R}^{N_{z} \\times D}\\), and event search tokens \\(F_{x}^{E} \\in \\mathbb{R}^{N_{x} \\times D}\\). Here, \\(N_{z} = \\frac{H_{z}W_{z}}{16^{2}}\\); \\(N_{x} = \\frac{H_{x}W_{x}}{16^{2}}\\); \\(D = 512\\); \\(H_{z}W_{z}\\) and \\(H_{x}W_{x}\\) represent the size of the template and search images, respectively. Then, the template tokens and search tokens are concatenated and fed into the \\(N\\)-layer Transformer-based encoder for joint modeling. The operation of our encoder is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.319, + 0.482, + 0.337 + ], + "angle": 0, + "content": "\\[\nF _ {0} ^ {m} = \\mathcal {C} \\left(F _ {z} ^ {m}, F _ {x} ^ {m}\\right), \\quad m \\in \\{R G B, E v e n t \\} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.339, + 0.482, + 0.357 + ], + "angle": 0, + "content": "\\[\nF _ {n} ^ {R}, F _ {n} ^ {E} = \\mathcal {E} _ {n} \\left(F _ {n - 1} ^ {R}, F _ {n - 1} ^ {E}\\right), \\quad n = 1, \\dots , N \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.471 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}\\) is the element-wise concatenation; \\(\\mathcal{E}_n\\) denotes \\(n_{th}\\) encoder layer; \\(F_n^R\\) and \\(F_n^E\\) are the RGB and event outputs of the \\(n_{th}\\) layer, respectively. The above operations facilitate the interaction between the template and search tokens during the RGB and event feature extraction. Next, we describe in detail our Mamba-based fusion module inserted between encoding layers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.477, + 0.308, + 0.492 + ], + "angle": 0, + "content": "3.2. Fusion Mamba Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Our proposed Fusion Mamba (FM) module aims to fuse multimodal semantic information. The key to FM is designed based on the following observations: (i) The interaction between template and search features is crucial as it can help the tracker to learn and prioritize the target-relevant search features; (ii) Directly combining RGB and event modals disregards cases where one of modalities cannot provide valuable cues. In the FM, we present Target-Aware Scan module (TAS) and Cross-Modality Scan module (CMS) to address the above issues, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.801 + ], + "angle": 0, + "content": "Target-Aware Scan Module (TAS). The TAS module enhances the interaction between template and search regions by leveraging Mamba's global modeling capability. Compared with the self-attention in Transformer, our TAS employs the bidirectional SSM mechanism [49] to estimate the relevance between all template tokens and each search token. The completeness of template cues can encourage our tracker to focus on informative search tokens. Given \\( F_{n}^{m} \\in \\{F_{n}^{R}, F_{n}^{E}\\} \\), as shown in Figure 3 (b), the TAS module is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.809, + 0.482, + 0.826 + ], + "angle": 0, + "content": "\\[\nX _ {n} ^ {m} = \\mathcal {L} \\left(\\mathcal {N} \\left(F _ {n} ^ {m}\\right)\\right), Z _ {n} ^ {m} = \\sigma \\left(\\mathcal {L} \\left(\\mathcal {N} \\left(F _ {n} ^ {m}\\right)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.828, + 0.482, + 0.845 + ], + "angle": 0, + "content": "\\[\nY _ {f} = \\mathcal {S S M} _ {f} \\left(\\psi \\left(X _ {n} ^ {m}\\right)\\right), Y _ {b} = \\mathcal {S S M} _ {b} \\left(\\psi \\left(X _ {n} ^ {m}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.847, + 0.482, + 0.865 + ], + "angle": 0, + "content": "\\[\nH _ {n} ^ {m} = \\mathcal {L} \\left(\\mathcal {N} \\left(Z _ {n} ^ {m} \\odot Y _ {f} + Z _ {n} ^ {m} \\odot Y _ {b}\\right)\\right) + F _ {n} ^ {m}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is Linear projection; \\(\\mathcal{N}\\) denotes LayerNorm operation; \\(\\sigma\\) is SiLU activation function; \\(\\psi\\) is 1D convolu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "tion operation with the kernel size of 4; \\(\\mathcal{N}\\) means elementwise multiplication; \\(X_{n}^{m}\\) is the projection of \\(F_{n}^{m}\\); \\(Z_{n}^{m}\\) is gate signal; \\(Y_{f}\\) and \\(Y_{b}\\) are the outputs through forward and backward directions of SSM (i.e., \\(\\mathcal{SSM}_f\\) and \\(\\mathcal{SSM}_b\\)). \\(H_{n}^{m} \\in \\{H_{n}^{R}, H_{n}^{E}\\}\\) is the enhanced output of our TAS module in the \\(n_{th}\\) encoder layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.333 + ], + "angle": 0, + "content": "Cross-Modality Scan Module (CMS). CMS is responsible for improving the relevance and complementarity of multimodal features through a simple yet effective cross-fusion scheme. Different from TAS, we only input search tokens into CMS to prevent interference from template features. For simplicity, we will use the RGB domain as an example to describe the specific operation since RGB and event domains function similarly. Given RGB search tokens \\( H_{n}^{R,x} \\) in the \\( n_{th} \\) layer, as shown in Figure 3 (c), our CMS is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.34, + 0.905, + 0.36 + ], + "angle": 0, + "content": "\\[\nX _ {n} ^ {R, x} = \\mathcal {L} \\left(\\mathcal {N} \\left(H _ {n} ^ {R, x}\\right)\\right), Z _ {n} ^ {R, x} = \\sigma \\left(\\mathcal {L} \\left(\\mathcal {N} \\left(H _ {n} ^ {R, x}\\right)\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.362, + 0.905, + 0.38 + ], + "angle": 0, + "content": "\\[\nY ^ {\\prime} = \\mathcal {S S M} _ {f} \\left(\\psi \\left(X _ {n} ^ {R, x}\\right)\\right) + \\mathcal {S S M} _ {b} \\left(\\psi \\left(X _ {n} ^ {R, x}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.382, + 0.905, + 0.401 + ], + "angle": 0, + "content": "\\[\nY _ {n} ^ {R, x} = \\mathcal {L} \\left(\\mathcal {N} \\left(Y ^ {\\prime} \\odot Z _ {n} ^ {R, x} + Y ^ {\\prime} \\odot Z _ {n} ^ {E, x}\\right)\\right) + H _ {n} ^ {R, x}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.576 + ], + "angle": 0, + "content": "The modality-specific gating signals \\( Z_{n}^{R,x} \\) and \\( Z_{n}^{E,x} \\) are utilized to direct the interaction between modalities, thereby strengthening the primary modality while suppressing the degraded modality. With the assistance of our proposed TAS and CMS, the multimodal encoder can efficiently facilitate the interaction between template and search tokens, as well as between RGB and events modalities. The RGB and event features of the last encoder layer are added together to get the final fused feature \\( \\mathcal{F}_s \\), which will be used to enhance the functionality of the prediction head and historical decoder." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.585, + 0.694, + 0.599 + ], + "angle": 0, + "content": "3.3. Historical Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Historical temporal information is crucial in visual object tracking as it provides the state changes of targets across frames. To fully exploit the temporal cues, we propose the Historical Decoder with \\(M\\) layers, each layer consists of two key components: Historical State Aware (HSA) module and Feature Sequence Attention (FSA) module." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.697, + 0.906, + 0.82 + ], + "angle": 0, + "content": "Historical State Aware Module (HSA). Mamba offers a natural advantage over conventional CNNs and Transformers when it comes to long sequence modeling. We thus present Mamba-based HSA module to capture the target historical states of \\(N\\) frames within a sequence into \\(N\\) one-dimensional learnable vectors \\(Q(q_{1},\\ldots ,q_{N})\\in \\mathbb{R}^{N\\times D}\\). Mamba models the sequence and outputs \\(Q^{\\prime} = (q^{\\prime},\\dots,q_N^{\\prime})\\in \\mathbb{R}^{N\\times D}\\), this is" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.827, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\nX _ {Q} = \\mathcal {S S M} (\\psi (\\mathcal {L} (Q))), Z _ {Q} = \\sigma (\\mathcal {L} (Q)), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.846, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\nQ ^ {\\prime} = \\mathcal {L} \\left(\\mathcal {N} \\left(X _ {Q} \\odot Z _ {Q}\\right) + Q. \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "The output \\( Q' \\) will be subsequently utilized as the input of FSA. It is important to highlight that, with the help of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6503" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.091, + 0.478, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.486, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Overview of our Historical Decoder with two key components: Historical State Aware (HSA) module, and Feature-Sequence Attention (FSA) module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.468 + ], + "angle": 0, + "content": "SSM mechanism, the hidden state generated by each query at the current moment will subsequently be transmitted and contribute to the estimation in the following moment. In this way, we guarantee the long-term viability of modeling historical cues. Compared to Transformer-based autoregressive models, this significantly increases both inference and training speeds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Feature-Sequence Attention Module (FSA). Our proposed FSA module aims to combine multimodal spatial features \\(\\mathcal{F}_s\\) and historical temporal features \\(Q^{\\prime}\\) to improve the confidence of the prediction head. Our FSA module employs the multi-head attention mechanism, where historical information sequences are used as queries and multimodal features are used as keys and values:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.582, + 0.418, + 0.601 + ], + "angle": 0, + "content": "\\[\nQ ^ {\\prime \\prime} = \\mathcal {N} \\left(\\mathcal {C} \\left(\\left(H _ {1}, \\dots , H _ {n _ {h}}\\right) W ^ {O}\\right)\\right) + Q ^ {\\prime},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.603, + 0.482, + 0.627 + ], + "angle": 0, + "content": "\\[\nH _ {i} = \\mathcal {A T T N} \\left(Q ^ {\\prime} W _ {i} ^ {Q}, \\mathcal {F} _ {s} W _ {i} ^ {K}, \\mathcal {F} _ {s} W _ {i} ^ {V}\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.626, + 0.37, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A T T N} (q, k, v) = \\varphi \\left(\\frac {q k ^ {\\top}}{\\sqrt {d _ {k}}}\\right) v,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.807 + ], + "angle": 0, + "content": "where \\(\\varphi\\) is the Softmax function; \\(W_{i}^{Q}\\in \\mathbb{R}^{d_{m}\\times d_{k}}\\), \\(W_{i}^{K}\\in \\mathbb{R}^{d_{m}\\times d_{k}}\\), \\(W_{i}^{V}\\in \\mathbb{R}^{d_{m}\\times d_{v}}\\), and \\(W^{O}\\in \\mathbb{R}^{n_{h}\\times d_{m}}\\) are all learnable parameters. In our method, we employ multihead attention with 8 heads, i.e., \\(n_h = 8\\), \\(d_m = 512\\), and \\(d_k = d_v = \\frac{d_m}{8} = 64\\). Then \\(Q''\\) is fed into the following decoder layer to generate the final temporal queries \\(Q\\) after being processed by the Feed Forward Network (FFN). Finally, we further combine the spatial and temporal information without introducing any parameters by:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.814, + 0.483, + 0.832 + ], + "angle": 0, + "content": "\\[\nF _ {o u t} = \\mathcal {F} _ {s} \\odot \\mathcal {Q} ^ {\\top}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( F_{out} \\) represents the multimodal features enriched with historical temporal information, and highlights the position where the target may be located. Then it will be fed into the tracking head for the target bbox prediction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.666, + 0.105 + ], + "angle": 0, + "content": "3.4. Head and Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.189 + ], + "angle": 0, + "content": "We employ a center-based head as regular [39] to predict the center position and scale of the target. The prediction head outputs include the classification score map, the bounding box size, and the offset size. We use focal loss [19] and GIoU loss [24] for training the network. The total loss \\( L \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.202, + 0.907, + 0.218 + ], + "angle": 0, + "content": "\\[\nL = L _ {c l s} + \\lambda_ {i o u} L _ {i o u} + \\lambda_ {L _ {1}} L _ {L _ {1}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.231, + 0.906, + 0.293 + ], + "angle": 0, + "content": "where \\( L_{cls} \\) denotes the weighted focal loss for classification, the generalized GIoU loss \\( L_{iou} \\) and L1 loss \\( L_{1} \\) are adopted for bounding box regression, \\( \\lambda_{iou} = 2 \\) and \\( \\lambda_{L_1} = 5 \\) are the hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.307, + 0.646, + 0.323 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.331, + 0.722, + 0.348 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.519 + ], + "angle": 0, + "content": "Implementation Details. We implemented MamTrack using PyTorch and trained it on 8 NVIDIA RTX 4090 GPUs with a batch size of 16. We employed the AdamW optimizer with a weight decay of \\(10^{-4}\\) and a learning rate of \\(4 \\times 10^{-4}\\). We utilized pretrained HiViT-Base [45] as the Transformer encoder with 20 layers. Fine-tuning MamTrack on the RGB-Event training set took 50 epochs. Our proposed Fusion Mamba module is inserted into the 12th, 15th, and 18th layers. The decoder consists of 3 layers (i.e., \\(M = 3\\)). Besides, we utilize the Hamming window to incorporate positional priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt three widely used metrics for quantitatively evaluating the tracking performance: Success Rate (SR), Precision Rate (PR), and Overlap Precision \\((\\mathrm{OP}_T)\\). The SR focuses on the frame where the overlap between ground truth and predicted bounding box is larger than a threshold; \\(\\mathrm{OP}_T\\) matches SR but with a specific overlap threshold \\(T\\). The PR metric counts the frame on which the center distance between the ground truth and predicted bounding box is within a given threshold." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.822, + 0.683 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.749 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed approach, we compare our tracker with representative state-of-the-art trackers on three widely utilized benchmarks: VisEvent [29], FELT [31], and FE108 [41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "VisEvent. VisEvent is a large-scale visible-event tracking benchmark, which contains 500 training and 320 testing sequences with 17 attributes. As shown in Figure 5(a), our MamTrack achieves the top SR and PR of \\(61.6\\%\\) and \\(79.2\\%\\), respectively, which beats all single- and multimodal SOTA trackers. For example, our method outperforms the pure Transformer-based temporal tracker AQA-Track [36] by \\(2.9\\%\\) on SR and multimodal tracker SDSTrack [13] by \\(1.9\\%\\). Figure 6 further provides the tracking performance under 17 challenging attributes. We can see" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6504" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.287, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.483, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.213, + 0.375, + 0.222 + ], + "angle": 0, + "content": "(a) Success (left) and Precision (right) plot on VisEvent" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.224, + 0.288, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.224, + 0.483, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.348, + 0.362, + 0.357 + ], + "angle": 0, + "content": "(b) Success (left) and Precision (right) plot on FELT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.401 + ], + "angle": 0, + "content": "Figure 5. Comparison of Success (left) and Precision (right) plots on VisEvent (a) and FELT (b)." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.417, + 0.288, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.418, + 0.483, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Figure 6. Comparison of SR (left) and PR (right) across various attributes using our approach against four top trackers on VisEvent, highlights the performance of Ours/Lowest." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.671 + ], + "angle": 0, + "content": "that our approach still fares best in all degraded conditions, demonstrating its ability to effectively fuse the strengths of RGB and Events." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.824 + ], + "angle": 0, + "content": "FELT. FELT is a large-scale long-term frame-event single object tracking dataset, which contains 742 videos, 1,594,474 RGB frames and corresponding events. The results in Figure 5(b) show that our MamTrack achieves a \\(48.9\\%\\) overall SR and \\(60.8\\%\\) PR. The results illustrate that our approach achieves effective long-term tracking by exploiting the historical temporal modeling capability of Mamba. We also offer an attribute-based evaluation in the supplementary material for further analysis, our MamTrack achieves the best performance on most attribute splits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "FE108. FE108 is a visible-event multimodal tracking dataset of indoor scenes, which contains 76 videos for training and 32 for testing. As shown in Table 1, our MamTrack outperforms the previous best tracker ViPT by \\(0.6\\%\\) in SR and \\(0.4\\%\\) in PR. It demonstrates that our approach is also" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.909, + 0.165 + ], + "angle": 0, + "content": "
ATOMDiMPPrDiMPFENetMixFormerTransT
52.8/80.053.4/88.255.2/86.863.4/92.460.4/87.856.7/89.0
SiamFc++STARKOSTrackViPTHDETrackOurs
54.5/85.355.4/83.757.1/89.365.8/93.859.8/92.266.4/94.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.176, + 0.907, + 0.203 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of SR(\\%) / PR(\\%) against SOTA trackers on the FE108 dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.361 + ], + "angle": 0, + "content": "
#RGBEventTASCMSCFFSAHSASR(%)PR(%)
1X40.556.1
2X59.075.8
3X60.778.1
4X60.277.4
5X60.077.1
6X60.678.0
7X60.477.8
861.679.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.371, + 0.906, + 0.412 + ], + "angle": 0, + "content": "Table 2. Ablation study for important components on VisEvent. Blank denotes the component is used by default, while \\( \\mathbf{x} \\) represents the component is removed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.441, + 0.861, + 0.457 + ], + "angle": 0, + "content": "highly competitive for short-term tracking scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.771, + 0.485 + ], + "angle": 0, + "content": "4.3. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Impact of Multimodal Input. We assess the effectiveness of multimodal fusion by separately utilizing event-only and RGB-only inputs. As shown in Table 2 #1 and #2, the tracking results with unimodal input are significantly worse than those with multimodal input, demonstrating that our method effectively utilizes both modalities' advantages to improve tracking under degraded conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Effectiveness of Fusion Mamba Module. There are two key mamba-based components in our Fusion Mamba (FM) module: (i) Target-Aware Scan (TAS) facilitates the interaction between the template and search tokens; (ii) Cross-Modality Scan (CMS) focuses on the fusion of event and RGB modalities in an adaptive manner. To verify their effectiveness, we modify the original model by dropping each of the components and retraining the modified models. The tracking performance in Table 2 #3 and #4 has shown a decrease to varying degrees, indicating the effectiveness and rationality of our designed FM module. In CMS, the gate-based operation Cross-Fusion (CF) is the primary operation that supports the interaction between the two modalities. We thus eliminate the CF to showcase its indispensability. #5 shows that when the CF is removed, the SR drops by \\(1.6\\%\\). This illustrates that the deep interaction among features from both modalities plays a key role." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "One possible factor that impacts the contribution of FM to our tracker is the location of FM insertion in the encoder. To verify this, several additional sets of experiments were" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6505" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.493, + 0.166 + ], + "angle": 0, + "content": "
LocationSR (%)PR(%)LocationSR (%)PR(%)
1260.077.014-17-2061.178.4
12-1561.078.312-13-1460.177.3
12-15-18 (Ours)61.679.218-19-2060.677.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.176, + 0.483, + 0.219 + ], + "angle": 0, + "content": "Table 3. Comparison of different FM module locations on VisEvent. \\( x - y - z \\) represents our FM module is inserted into the \\( x \\)th, \\( y \\)th, and \\( z \\)th encoder layers." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.234, + 0.475, + 0.348 + ], + "angle": 0, + "content": "
Type of HSANumber of LayersSR(%)PR(%)FPS
None059.976.820.7
Mamba160.578.119.9
361.679.218.1
660.978.116.6
Transformer361.078.416.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.359, + 0.483, + 0.387 + ], + "angle": 0, + "content": "Table 4. Ablation studies on the type and number of decoder layers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.483, + 0.55 + ], + "angle": 0, + "content": "conducted. As shown in Table 3, compared to other 3-stage insertion styles, the 12-15-18 style adopted in this paper provides the best performance, which can be attributed to our insertion strategy that focuses on both low-level texture features and high-level semantic features. Moreover, the tracking performance exhibits a progressive improvement from 12 to 12-15 and then to 12-15-18 style, revealing the growing synergy between RGB and event modalities facilitated by our FM module." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Effectiveness of Historical Decoder. The Historical Decoder is designed to capture and transmit historical information during tracking, which consists of two main components: Historical State Aware (HSA) and Feature Sequence Attention (FSA) modules. As shown in Table 2 #6, when the FSA is removed, the SR score drops by \\(1.0\\%\\), showing the importance of multimodal features for exploiting historical information. The performance drops by \\(1.2\\%\\) when the HSA is removed as shown in #7. This illustrates the effectiveness of Mamba for modeling long sequences." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.703, + 0.483, + 0.856 + ], + "angle": 0, + "content": "We also examine the influence of the number of historical decoder layers on tracking performance. As shown in Table 4, the utilization of the historical decoder greatly enhances the results, with the most performance observed when employing 3 decoder layers. To further verify the effectiveness of the Mamba in the historical decoder, we replace the Mamba architecture with Transformer. Table 4 shows our method with Mamba excels in both performance and speed due to Mamba's linear complexity and its capability to model long sequences." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Impact of the Length of Historical Information. As shown in Table 5, the model's performance improves as the range of historical information increases from 1 to 4. This" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.089, + 0.881, + 0.166 + ], + "angle": 0, + "content": "
LengthSR(%)PR(%)LengthSR(%)PR(%)
160.177.5859.876.9
260.477.91659.776.6
461.679.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.176, + 0.874, + 0.191 + ], + "angle": 0, + "content": "Table 5. Impact of the length of historical information." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.205, + 0.905, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.389, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Figure 7. Comparison of SR scores for long-term tracking on VisEvent, where the horizontal axis is test sequence length." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.55 + ], + "angle": 0, + "content": "indicates that the model can effectively capture the target's movement trends and appearance changes, resulting in improved performance. However, when the historical information range is extended to 8 and 16, the model's performance slightly declines. This decrease is due to the limited number of video sequences used in each iteration, which impairs the model's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Effectiveness for Long-term Tracking. To further demonstrate the effectiveness of our Mamba-based historical decoder in capturing temporal information, we analyze its performance in long-term tracking. Specifically, we evaluate the tracking performance of sequences containing more than \\( i \\) frames on the VisEvent individually, where \\( i \\in \\{300, 500, 700, 1000, 1500, 2000\\} \\). Figure 7 shows that our MamTrack outperforms other SOTA Transformer-based temporal trackers and multimodal trackers by a large margin in SR across all settings. This suggests our Mamba-based approach proves superior in capturing long-term cues compared to the Transformer-based autoregressive scheme." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.871, + 0.759 + ], + "angle": 0, + "content": "4.4. Performance on RGB-T Tracking Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.854 + ], + "angle": 0, + "content": "To further validate the generalization ability of our MamTrack, we conducted additional experiments on the RGB-T tracking task. Following the common settings used in previous work [48], we trained our tracker on the LasHeR [16] training set and tested it on both the LasHeR test set and the RGBT234 [15] dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Results on LasHeR Dataset. LasHeR is the largest RGB-T tracking dataset at present, containing 1224 aligned video sequences with diverse attribute annotations. Of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6506" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.935, + 0.168 + ], + "angle": 0, + "content": "
LasHeRmfDimp 34.3/ 44.7SGT++ 25.1/36.5OSTrack 41.2/51.5ProTrack 42.0/53.8HMFT 31.3/ 43.6APFNet 36.2/50.0CEUTrack 49.6/63.5ViPT 52.5/65.1Ours 54.2/67.4
RGBT234mfDimp 42.8/ 64.6CMPP 57.7/82.3OSTrack 54.9/72.9ProTrack 57.9/82.7HMFT 56.8/78.8APFNet 57.9/82.7MFGNet 53.5/78.3ViPT 61.7/ 83.5Ours 62.4/84.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.237, + 0.176, + 0.759, + 0.19 + ], + "angle": 0, + "content": "Table 6. Performance comparison of SR(%) / PR(%) on LasHeR and RGBT234 dataset." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.219, + 0.482, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.438 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of our MamTrack against other SOTA trackers under 4 challenging conditions on VisEvent." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.475, + 0.483, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Figure 9. The comparison attention map in the MamTrack. The first row is the ground truth; the second and third rows display the attention maps in the multimodal encoder and historical decoder, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.483, + 0.814 + ], + "angle": 0, + "content": "these, 245 sequences are designated as the testing set, while the remaining are used for training. As shown in Table 6, MamTrack performs exceptionally well on RGB-T tracking tasks, effectively leveraging the advantages of both thermal and visible information. This indicates its potential for other multimodal tracking tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Results on RGBT234 Dataset. RGBT234 is a large-scale RGB-T tracking dataset comprising 234 pairs of videos. As shown in Table 6, our MamTrack surpasses the most advantaged tracker ViPT [48] by \\(0.7\\%\\) in Success Rate (SR) and \\(0.9\\%\\) in Precision Rate (PR)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.217, + 0.651, + 0.231 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.241, + 0.907, + 0.392 + ], + "angle": 0, + "content": "We first provide a visual comparison across four challenging conditions on the VisEvent dataset. As shown in Figure 8, our approach can accurately track the target under all conditions. In particular, our method effectively fuses the advantages of RGB and event to enhance tracker confidence when the target is in scenes with motion blur or fast motion. Besides, in situations where the target is partially occluded or the background is complex, the historical cues encourage our method to recall the target's appearance, thereby minimizing environmental interference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Furthermore, we also visualize the attention maps of the encoder and decoder to demonstrate the effectiveness of historical information. As shown in Figure 9, guided by historical information, our MamTrack pays more attention to the target's location under challenging conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.493, + 0.634, + 0.508 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.734 + ], + "angle": 0, + "content": "We propose MamTrack, a novel approach jointly leverages multimodal spatial and historical temporal information for RGBE object tracking. Our historical decoder, based on the innovative Mamba architecture, effectively captures target appearance changes and motion trends, enabling the model to locate the target within complex backgrounds. We also introduce a simple yet effective Mamba-based fusion method for integrating cross-modality cues. Extensive evaluations demonstrate that our method achieves competitive performance compared to previous SOTA trackers. However, we primarily focuses on the effectiveness of Mamba in modality fusion, but not leverage the high temporal resolution characteristics of event data. Future research will explore this aspect to enhance the performance of the tracker." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by the National Key Research and Development Program of China (No. 2022ZD0210500), the National Natural Science Foundation of China under Grant 62441216/62332019, the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01), the Ningbo Major Research and Development Plan Project of China (No. 2023Z225), Beijing Natural Science Foundation of China(Grant No.3254040), National Key Laboratory of Human Factors Engineering and China Postdoctoral Science Foundation under Grant (No. 2024M760315)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6507" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6182-6191, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.366 + ], + "angle": 0, + "content": "[4] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13608-13618, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[5] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4660-4669, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[6] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7183-7192, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[7] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-Mamba for Cross-modality Object Detection, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[9] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[10] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. IEEE Transactions on Image Processing, 32:6129-6141, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Xiaojun Hou, Jiazheng Xing, Yijie Qian, Yaowei Guo, Shuo Xin, Junhao Chen, Kai Tang, Mengmeng Wang, Zhengkai Jiang, Liang Liu, et al. Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26551-26561, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Ju Huang, Shiao Wang, Shuai Wang, Zhe Wu, Xiao Wang, and Bo Jiang. Mamba-fetrack: Frame-event tracking via state space model. arXiv preprint arXiv:2404.18174, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. Rgb-t object tracking: Benchmark and baseline. Pattern Recognition, 96:106977, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for rgbt tracking. IEEE Transactions on Image Processing, 31:392-404, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[17] Kang Li, Fazhi He, Haiping Yu, and Xiao Chen. A parallel and robust object tracking approach synthesizing adaptive bayesian learning and improved incremental subspace learning. Frontiers of Computer Science, 13(5):1116, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.391 + ], + "angle": 0, + "content": "[18] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.393, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.904, + 0.49 + ], + "angle": 0, + "content": "[20] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual State Space Model, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[21] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3044-3053, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[22] Haiyang Mei, Bo Dong, Wen Dong, Jiaxi Yang, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Glass segmentation using intensity and spectral polarization cues. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12622-12631, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[23] Yu Qiao, Bo Dong, Ao Jin, Yu Fu, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Multiview spectral polarization propagation for video glass segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23218-23228, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[25] Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Shengping Zhang, and Xianxian Li. Explicit visual prompts for visual object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4838-4846, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6508" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[27] Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2321-2329, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[28] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[29] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 7(1):30-40, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[31] Xiao Wang, Ju Huang, Shiao Wang, Chuanming Tang, Bo Jiang, Yonghong Tian, Jin Tang, and Bin Luo. Long-term frame-event visual tracking: Benchmark dataset and baseline. arXiv preprint arXiv:2403.05839, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[32] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19248-19257, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[33] Yang Wang, Bo Dong, Yuji Zhang, Yunduo Zhou, Haiyang Mei, Ziqi Wei, and Xin Yang. Event-enhanced multi-modal spiking neural network for dynamic obstacle avoidance. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3138-3148, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[34] Yang Wang, Haiyang Mei, Qirui Bao, Ziqi Wei, Mike Zheng Shou, Haizhou Li, Bo Dong, and Xin Yang. Apprenticeship-inspired elegance: synergistic knowledge distillation empowers spiking neural networks for efficient single-eye emotion recognition. arXiv preprint arXiv:2407.09521, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[35] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9697-9706, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[36] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19300-19309, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[37] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10448-10457, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[39] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European conference on computer vision, pages 341-357. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[40] Haiwei Zhang, Jiqing Zhang, Bo Dong, Pieter Peers, Wenwei Wu, Xiaopeng Wei, Felix Heide, and Xin Yang. In the blink of an eye: Event-based emotion recognition. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[41] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13043-13052, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[42] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9781-9790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.429 + ], + "angle": 0, + "content": "[43] Jiqing Zhang, Bo Dong, Yingkai Fu, Yuanchen Wang, Xiaopeng Wei, Baocai Yin, and Xin Yang. A universal event-based plug-in module for visual object tracking in degraded conditions. International Journal of Computer Vision, 132 (5):1857-1879, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.907, + 0.5 + ], + "angle": 0, + "content": "[44] Lichao Zhang, Abel Gonzalez-Garcia, Joost Van De Weijer, Martin Danelljan, and Fahad Shahbaz Khan. Learning the model update for siamese trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4010-4019, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[45] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[46] Aihua Zheng, Juncong Liu, Zi Wang, Lili Huang, Chenglong Li, and Bing Yin. Visible-infrared person re-identification via specific and shared representations learning. Visual Intelligence, 1(1):29, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[47] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7588-7596, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.699, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[48] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9516-9526, 2023. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.811 + ], + "angle": 0, + "content": "[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model, 2024. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6509" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_origin.pdf b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..053f4de061f1a248b9b85c13243142870101491a --- /dev/null +++ b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/36e76161-ac87-43e3-905e-9c2eccb37ec2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09b5d110d46c7e34c37cae01876f3509ab60b37c93ebbf848752f891c6d5aafd +size 9562897 diff --git a/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/full.md b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad98bac03be380d2a476c0315f5c88120d337b8f --- /dev/null +++ b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/full.md @@ -0,0 +1,350 @@ +# Exploring Historical Information for RGBE Visual Tracking with Mamba + +Chuanyu Sun $^{1*}$ , Jiqing Zhang $^{2*}$ , Yang Wang $^{1}$ , Huilin Ge $^{3\dagger}$ , Qianchen Xia $^{4\dagger}$ , Baocai Yin $^{1,5\dagger}$ , Xin Yang $^{1\dagger}$ $^{1}$ Key Laboratory of Social Computing and Cognitive Intelligence, Dalian University of Technology + $^{2}$ Dalian Maritime University ${}^{3}$ Jiangsu University of Science and Technology + $^{4}$ The Future Laboratory, Tsinghua University, Beijing, China ${}^{5}$ Beijing University of Technology + +# Abstract + +Combining the advantages of conventional and event cameras for robust visual tracing has drawn extensive interest. However, existing tracking approaches heavily engage in complex cross-modal fusion modules, leading to higher computational complexity and training challenges. Besides, these methods generally ignore the effective integration of historical information, which is crucial to grasping the change in the target's appearance and motion trends. Given the recent advancements in Mamba's long-range modeling and linear complexity, we explore its potential in addressing the above issues in RGBE tracking tasks. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to achieve effective modality-selective fusion. This module can be seamlessly integrated into the encoding layer of prevalent Transformer-based backbones. Moreover, we further present a novel historical decoder that leverages Mamba's advanced long sequence modeling to effectively capture the target appearance changes with autoregressive queries. Extensive experiments show that our proposed approach achieves state-of-the-art performance on multiple challenging short-term and long-term RGBE benchmarks. Besides, the effectiveness of each key Mamba-based component of our approach is evidenced by our thorough ablation study. Code will be released at: https://github.com/scy0712/MamTrack + +# 1. Introduction + +Visual object tracking (VOT) is a fundamental task in computer vision[21-23, 30, 46] that focuses on estimating the position and shape of a target within video sequences based on its initial state [1, 3, 5, 17, 35, 36, 38, 47].Owing to the advantages of high temporal resolution and dynamic range offered by event cameras [33, 34, 40], there has been a + +![](images/e669452b3e32356a7b28c2128f777926545126898152cda417372448738fc1c1.jpg) +Figure 1. Overview of MamTrack and its two main components: (i) Multimodal Encoder is responsible for extracting and fusing features from RGB and event modalities; (ii) Historical Decoder is utilized to mine the temporal information of target appearance changes. + +growing trend of integrating them with conventional RGB cameras for robust visual tracking tasks [10, 28, 32, 41-43, 48]. These RGBE multimodal approaches can significantly enhance tracking performance in challenging conditions, such as fast motion and limited illumination. Although being effective, the above multimodal methods expend a substantial amount of labor on developing overly intricate fusion algorithms. For example, FENet [41] employs a total of 8 self-attention and 4 complex cross-domain attention operations to integrate event and frame features at various levels. Implementing these sophisticated strategies will significantly increase the computational complexity and challenge of the training. + +Moreover, the complex fusion designs make existing RGBE methods usually overlook the mining and application of historical temporal information, which is crucial for tracking as it captures changes in the target's state. Several RGB-only trackers [1, 4, 38] have been developed to update appearance information by replacing or fusing the templates. Occasionally, these straightforward methods ineffectively filter out unreliable updated templates and thus compromise the performance. To address this issue, some + +![](images/7863685608b5213b98bff92b07b281025fe99d931457345ba6c098abf899009a.jpg) +Figure 2. A comparison of our method with other recent SOTA trackers in long-term sequence. Benefiting from the proposed Historical Decoder, our method can effectively capture target appearance changes to guide robust long-term tracking even in partial occlusion scenes. + +methods [35, 36] incorporate an autoregressive scheme into the Transformer to transfer temporal information across multiple search frames. However, the Transformer structure lacks intrinsic designs for temporal sequential modeling, resulting in an inherent constraint, particularly in handling long sequences. + +Recently, Mamba [11], a novel State Space Model (SSM) [9, 12, 26, 37], has emerged as a highly promising approach for efficient long sequence modeling due to convolutional computation and linear complexity. Drawing inspiration from this advancement, as shown in Figure 1, we present mamba-based modules to facilitate robust RGB-E tracking by efficient multimodal fusion and effective historical information transmission. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to enhance the relationship between search and template tokens, as well as the complementarity between RGB and event modalities. This module is then easily inserted into the encoding layer of prevalent transformer-based backbones for combining the advantages of self-attention. Furthermore, we present a novel historical information decoder, which investigates the great potential of mamba in long sequence modeling to capture the appearance of targets by introducing a set of additional autoregressive queries. + +Extensive experiments on multiple challenging RGBE benchmarks show the effectiveness of our approach, which outperforms existing state-of-the-art (SOTA) methods in both short-term and long-term tracking (see Figure 2). Ablation studies verify the importance of each key component. More importantly, we reveal the considerable potential of Mamba architecture in object tracking tasks. In summary, we make the following contributions: + +- We present a novel and effective framework, dubbed MamTrack, a pioneering approach to incorporate the Mamba mechanism into RGBE multimodal tracking tasks. + +- Our proposed Mamba-based fusion module and historical decoder are capable of adaptively fusing multimodal information and effectively capturing the historical temporal cues of the target appearance changes, respectively. +- The proposed trackers achieve SOTA performance on multiple challenging short-term and long-term RGBE tracking benchmarks. + +# 2. Related Works + +# 2.1. Mamba + +The concept of the State Space Model (SSM) was first introduced in the S4 model [12], presenting a distinctive architecture capable of effectively modeling global information compared to traditional convolutional neural networks and Transformers. In their general form, SSM can be written as follows: + +$$ +\dot {x} (t) = A (t) x (t) + B (t) u (t), \tag {1} +$$ + +$$ +y (t) = C (t) x (t) + D (t) u (t). \tag {2} +$$ + +This process maps a 1-D input sequence $u(t) \in \mathbb{R}$ to a 1-D output sequence $y(t) \in \mathbb{R}$ through an implicit $N$ -D latent state sequence $x(t) \in \mathbb{R}^N$ . Building on the S4 model, the S5 model [26] reduces complexity to a linear level, while H3 [9] introduces it into language model tasks. Mamba [11] introduces a data-dependent SSM layer and a selection mechanism using parallel scan (S6), achieving higher inference speed and overall metrics compared to equivalent-scale Transformers. Recently, Mamba has been employed in discriminative computer vision models and drawn great attention [7, 14, 18, 20, 49]. For example, ViM [49] applies Mamba to the Vision Transformer [8] architecture. VMamba [20] designs a cross-scan mechanism to bridge the gap between 1-D array scanning and 2-D plain traversing. Unlike previous efforts, we explore a novel architecture for RGB-E tracking based on Mamba, further unleashing its potential in visual tasks. + +# 2.2. RGB-Event Single Object Tracking + +Recently, due to the unreliable nature of RGB-only data in challenging scenarios, more and more studies introduced event cameras to enhance tracking performance. These multimodal trackers can be primarily categorised into two categories based on the vision backbone employed. The first one [29, 41, 42] utilizes conventional CNNs to extract features from both RGB and event modalities, followed by a meticulously designed cross-modal attention mechanism for fusion. The second approach [28, 32] adopts the Transformer as the backbone, leveraging the multi-head attention mechanism to effectively model the indispensable correlation relationship between the RGB and the event modalities. To tackle the challenge of limited multimodal tracking data, + +![](images/26e30b1b0babfe40b62458605b793384273b405907bf015cc8227c1ec88641af.jpg) + +![](images/92017b34b2ff1b95432d3efbbcc301a324fff692c55030a9d144dbf5345f80f2.jpg) +Figure 3. (a) Details of our MamTrack, which employs the Multimodal Encoder to extract and fuse the features of RGB and event modals and introduces the Historical Decoder to capture the temporal cues related to target state changes; and two key components in the Fusion Mamba (FM) module: (b) Target-Aware Scan (TAS) and (c) Cross-Modality Scan (CMS) module. + +![](images/816a3f267be7141c2e4ea43ffef8c285fce45617f0c9d1d2d61c93f97f03aa7f.jpg) + +recent studies [2, 13, 48] incorporate prompt tuning strategies into the Transformer architecture to transfer pretrained RGB-based trackers to multimodal data. + +The methods mentioned above primarily emphasise multimodal fusion and overlook the significance of mining historical information, which is crucial for multimodal tracking. In contrast, we explore the great potential of Mamba for multimodal tracking tasks. By combining the gating signals of two modalities in Mamba and leveraging the temporal modeling mechanism of Mamba, our approach is capable of concurrently achieving adaptive RGB-E multimodal fusion and efficient utilization of historical cues. + +# 2.3. Tracking with Historical Temporal Information + +The historical information can provide valuable insights into the target's movements, appearance, and behaviour, thereby enhancing the tracker's confidence. Based on the number of search images input into the tracker, existing tracking approaches that leverage historical temporal cues can be categorised into two groups. The first one involves only inputting one search image and implementing a dynamic kernel filter [1, 6, 44] or a specialized update branch [4, 27, 38] to capture the target appearance changes with time and dynamically update template images during tracking. However, these methods generally involve handcrafted strategies, requiring tedious tuning for various application conditions. Another line of work [25, 35, 36, 47] takes multiple search sequences as input and then models the se + +quential evolution of target trajectories across frames by explicitly decoding the entire trajectory sequence step by step. Despite these advances, Transformer architectures employed in these methods are not intrinsically designed for temporal sequential modeling, which presents an inherent limitation. Competing with Transformer, Mamba [11] has shown great potential for long sequence modeling. Therefore, our research endeavors to design a novel Mamba-based historical decoder to effectively transmit temporal information in a parallel training manner. + +# 3. Methodology + +The overall architecture of our proposed MamTrack is illustrated in Figure 3, which has two branches with shared weights: RGB branch and Event branch. Each branch receives a template and multiple search images as inputs and corresponding features can be extracted by the Multimodal Encoder. The core of the Multimodal Encoder is our proposed Fusion Mamba (FM) module, which is inserted between Transformer encoder layers to leverage the advantages of both RGB and event modalities. In particular, the RGB modal offers abundant rich texture information, while the event modal is adept at handling challenging scenarios. The fused features are then input into our other primary component, the Historical Encoder, which specifically aims to capture the target appearance changes and motion trends by utilizing temporal cues. Finally, the fused features are multiplied with the decoder's output and then fed into the prediction head to locate the targets. + +# 3.1. Multimodal Encoder + +Following [36], we employ a gradual downsampling network as multimodal encoder, which uses a $4 \times 4$ patch size for patch embedding. Subsequently, we adopt a total of eight multi-layer perceptron (MLP) layers and two merging layers for gradual downsampling. After this, we obtain the RGB template tokens $F_{z}^{R} \in \mathbb{R}^{N_{z} \times D}$ , RGB search tokens $F_{x}^{R} \in \mathbb{R}^{N_{x} \times D}$ , event template tokens $F_{z}^{E} \in \mathbb{R}^{N_{z} \times D}$ , and event search tokens $F_{x}^{E} \in \mathbb{R}^{N_{x} \times D}$ . Here, $N_{z} = \frac{H_{z}W_{z}}{16^{2}}$ ; $N_{x} = \frac{H_{x}W_{x}}{16^{2}}$ ; $D = 512$ ; $H_{z}W_{z}$ and $H_{x}W_{x}$ represent the size of the template and search images, respectively. Then, the template tokens and search tokens are concatenated and fed into the $N$ -layer Transformer-based encoder for joint modeling. The operation of our encoder is defined as: + +$$ +F _ {0} ^ {m} = \mathcal {C} \left(F _ {z} ^ {m}, F _ {x} ^ {m}\right), \quad m \in \{R G B, E v e n t \} \tag {3} +$$ + +$$ +F _ {n} ^ {R}, F _ {n} ^ {E} = \mathcal {E} _ {n} \left(F _ {n - 1} ^ {R}, F _ {n - 1} ^ {E}\right), \quad n = 1, \dots , N \tag {4} +$$ + +where $\mathcal{C}$ is the element-wise concatenation; $\mathcal{E}_n$ denotes $n_{th}$ encoder layer; $F_n^R$ and $F_n^E$ are the RGB and event outputs of the $n_{th}$ layer, respectively. The above operations facilitate the interaction between the template and search tokens during the RGB and event feature extraction. Next, we describe in detail our Mamba-based fusion module inserted between encoding layers. + +# 3.2. Fusion Mamba Module + +Our proposed Fusion Mamba (FM) module aims to fuse multimodal semantic information. The key to FM is designed based on the following observations: (i) The interaction between template and search features is crucial as it can help the tracker to learn and prioritize the target-relevant search features; (ii) Directly combining RGB and event modals disregards cases where one of modalities cannot provide valuable cues. In the FM, we present Target-Aware Scan module (TAS) and Cross-Modality Scan module (CMS) to address the above issues, respectively. + +Target-Aware Scan Module (TAS). The TAS module enhances the interaction between template and search regions by leveraging Mamba's global modeling capability. Compared with the self-attention in Transformer, our TAS employs the bidirectional SSM mechanism [49] to estimate the relevance between all template tokens and each search token. The completeness of template cues can encourage our tracker to focus on informative search tokens. Given $F_{n}^{m} \in \{F_{n}^{R}, F_{n}^{E}\}$ , as shown in Figure 3 (b), the TAS module is formulated as: + +$$ +X _ {n} ^ {m} = \mathcal {L} \left(\mathcal {N} \left(F _ {n} ^ {m}\right)\right), Z _ {n} ^ {m} = \sigma \left(\mathcal {L} \left(\mathcal {N} \left(F _ {n} ^ {m}\right)\right)\right), \tag {5} +$$ + +$$ +Y _ {f} = \mathcal {S S M} _ {f} \left(\psi \left(X _ {n} ^ {m}\right)\right), Y _ {b} = \mathcal {S S M} _ {b} \left(\psi \left(X _ {n} ^ {m}\right)\right), \tag {6} +$$ + +$$ +H _ {n} ^ {m} = \mathcal {L} \left(\mathcal {N} \left(Z _ {n} ^ {m} \odot Y _ {f} + Z _ {n} ^ {m} \odot Y _ {b}\right)\right) + F _ {n} ^ {m}, \tag {7} +$$ + +where $\mathcal{L}$ is Linear projection; $\mathcal{N}$ denotes LayerNorm operation; $\sigma$ is SiLU activation function; $\psi$ is 1D convolu + +tion operation with the kernel size of 4; $\mathcal{N}$ means elementwise multiplication; $X_{n}^{m}$ is the projection of $F_{n}^{m}$ ; $Z_{n}^{m}$ is gate signal; $Y_{f}$ and $Y_{b}$ are the outputs through forward and backward directions of SSM (i.e., $\mathcal{SSM}_f$ and $\mathcal{SSM}_b$ ). $H_{n}^{m} \in \{H_{n}^{R}, H_{n}^{E}\}$ is the enhanced output of our TAS module in the $n_{th}$ encoder layer. + +Cross-Modality Scan Module (CMS). CMS is responsible for improving the relevance and complementarity of multimodal features through a simple yet effective cross-fusion scheme. Different from TAS, we only input search tokens into CMS to prevent interference from template features. For simplicity, we will use the RGB domain as an example to describe the specific operation since RGB and event domains function similarly. Given RGB search tokens $H_{n}^{R,x}$ in the $n_{th}$ layer, as shown in Figure 3 (c), our CMS is defined as: + +$$ +X _ {n} ^ {R, x} = \mathcal {L} \left(\mathcal {N} \left(H _ {n} ^ {R, x}\right)\right), Z _ {n} ^ {R, x} = \sigma \left(\mathcal {L} \left(\mathcal {N} \left(H _ {n} ^ {R, x}\right)\right)\right), \tag {8} +$$ + +$$ +Y ^ {\prime} = \mathcal {S S M} _ {f} \left(\psi \left(X _ {n} ^ {R, x}\right)\right) + \mathcal {S S M} _ {b} \left(\psi \left(X _ {n} ^ {R, x}\right)\right), \tag {9} +$$ + +$$ +Y _ {n} ^ {R, x} = \mathcal {L} \left(\mathcal {N} \left(Y ^ {\prime} \odot Z _ {n} ^ {R, x} + Y ^ {\prime} \odot Z _ {n} ^ {E, x}\right)\right) + H _ {n} ^ {R, x}. \tag {10} +$$ + +The modality-specific gating signals $Z_{n}^{R,x}$ and $Z_{n}^{E,x}$ are utilized to direct the interaction between modalities, thereby strengthening the primary modality while suppressing the degraded modality. With the assistance of our proposed TAS and CMS, the multimodal encoder can efficiently facilitate the interaction between template and search tokens, as well as between RGB and events modalities. The RGB and event features of the last encoder layer are added together to get the final fused feature $\mathcal{F}_s$ , which will be used to enhance the functionality of the prediction head and historical decoder. + +# 3.3. Historical Decoder + +Historical temporal information is crucial in visual object tracking as it provides the state changes of targets across frames. To fully exploit the temporal cues, we propose the Historical Decoder with $M$ layers, each layer consists of two key components: Historical State Aware (HSA) module and Feature Sequence Attention (FSA) module. + +Historical State Aware Module (HSA). Mamba offers a natural advantage over conventional CNNs and Transformers when it comes to long sequence modeling. We thus present Mamba-based HSA module to capture the target historical states of $N$ frames within a sequence into $N$ one-dimensional learnable vectors $Q(q_{1},\ldots ,q_{N})\in \mathbb{R}^{N\times D}$ . Mamba models the sequence and outputs $Q^{\prime} = (q^{\prime},\dots,q_N^{\prime})\in \mathbb{R}^{N\times D}$ , this is + +$$ +X _ {Q} = \mathcal {S S M} (\psi (\mathcal {L} (Q))), Z _ {Q} = \sigma (\mathcal {L} (Q)), \tag {11} +$$ + +$$ +Q ^ {\prime} = \mathcal {L} \left(\mathcal {N} \left(X _ {Q} \odot Z _ {Q}\right) + Q. \right. \tag {12} +$$ + +The output $Q'$ will be subsequently utilized as the input of FSA. It is important to highlight that, with the help of the + +![](images/c5dbc520d5001a40e06eeddb685d967ed7e92e3b414773af54ead4a3c3d60602.jpg) +Figure 4. Overview of our Historical Decoder with two key components: Historical State Aware (HSA) module, and Feature-Sequence Attention (FSA) module. + +SSM mechanism, the hidden state generated by each query at the current moment will subsequently be transmitted and contribute to the estimation in the following moment. In this way, we guarantee the long-term viability of modeling historical cues. Compared to Transformer-based autoregressive models, this significantly increases both inference and training speeds. + +Feature-Sequence Attention Module (FSA). Our proposed FSA module aims to combine multimodal spatial features $\mathcal{F}_s$ and historical temporal features $Q^{\prime}$ to improve the confidence of the prediction head. Our FSA module employs the multi-head attention mechanism, where historical information sequences are used as queries and multimodal features are used as keys and values: + +$$ +Q ^ {\prime \prime} = \mathcal {N} \left(\mathcal {C} \left(\left(H _ {1}, \dots , H _ {n _ {h}}\right) W ^ {O}\right)\right) + Q ^ {\prime}, +$$ + +$$ +H _ {i} = \mathcal {A T T N} \left(Q ^ {\prime} W _ {i} ^ {Q}, \mathcal {F} _ {s} W _ {i} ^ {K}, \mathcal {F} _ {s} W _ {i} ^ {V}\right), \tag {13} +$$ + +$$ +\mathcal {A T T N} (q, k, v) = \varphi \left(\frac {q k ^ {\top}}{\sqrt {d _ {k}}}\right) v, +$$ + +where $\varphi$ is the Softmax function; $W_{i}^{Q}\in \mathbb{R}^{d_{m}\times d_{k}}$ , $W_{i}^{K}\in \mathbb{R}^{d_{m}\times d_{k}}$ , $W_{i}^{V}\in \mathbb{R}^{d_{m}\times d_{v}}$ , and $W^{O}\in \mathbb{R}^{n_{h}\times d_{m}}$ are all learnable parameters. In our method, we employ multihead attention with 8 heads, i.e., $n_h = 8$ , $d_m = 512$ , and $d_k = d_v = \frac{d_m}{8} = 64$ . Then $Q''$ is fed into the following decoder layer to generate the final temporal queries $Q$ after being processed by the Feed Forward Network (FFN). Finally, we further combine the spatial and temporal information without introducing any parameters by: + +$$ +F _ {o u t} = \mathcal {F} _ {s} \odot \mathcal {Q} ^ {\top}, \tag {14} +$$ + +where $F_{out}$ represents the multimodal features enriched with historical temporal information, and highlights the position where the target may be located. Then it will be fed into the tracking head for the target bbox prediction. + +# 3.4. Head and Loss + +We employ a center-based head as regular [39] to predict the center position and scale of the target. The prediction head outputs include the classification score map, the bounding box size, and the offset size. We use focal loss [19] and GIoU loss [24] for training the network. The total loss $L$ is: + +$$ +L = L _ {c l s} + \lambda_ {i o u} L _ {i o u} + \lambda_ {L _ {1}} L _ {L _ {1}}, \tag {15} +$$ + +where $L_{cls}$ denotes the weighted focal loss for classification, the generalized GIoU loss $L_{iou}$ and L1 loss $L_{1}$ are adopted for bounding box regression, $\lambda_{iou} = 2$ and $\lambda_{L_1} = 5$ are the hyperparameters. + +# 4. Experiments + +# 4.1. Experimental Settings + +Implementation Details. We implemented MamTrack using PyTorch and trained it on 8 NVIDIA RTX 4090 GPUs with a batch size of 16. We employed the AdamW optimizer with a weight decay of $10^{-4}$ and a learning rate of $4 \times 10^{-4}$ . We utilized pretrained HiViT-Base [45] as the Transformer encoder with 20 layers. Fine-tuning MamTrack on the RGB-Event training set took 50 epochs. Our proposed Fusion Mamba module is inserted into the 12th, 15th, and 18th layers. The decoder consists of 3 layers (i.e., $M = 3$ ). Besides, we utilize the Hamming window to incorporate positional priors. + +Evaluation Metrics. We adopt three widely used metrics for quantitatively evaluating the tracking performance: Success Rate (SR), Precision Rate (PR), and Overlap Precision $(\mathrm{OP}_T)$ . The SR focuses on the frame where the overlap between ground truth and predicted bounding box is larger than a threshold; $\mathrm{OP}_T$ matches SR but with a specific overlap threshold $T$ . The PR metric counts the frame on which the center distance between the ground truth and predicted bounding box is within a given threshold. + +# 4.2. Comparisons with State-of-the-arts + +To demonstrate the effectiveness of our proposed approach, we compare our tracker with representative state-of-the-art trackers on three widely utilized benchmarks: VisEvent [29], FELT [31], and FE108 [41]. + +VisEvent. VisEvent is a large-scale visible-event tracking benchmark, which contains 500 training and 320 testing sequences with 17 attributes. As shown in Figure 5(a), our MamTrack achieves the top SR and PR of $61.6\%$ and $79.2\%$ , respectively, which beats all single- and multimodal SOTA trackers. For example, our method outperforms the pure Transformer-based temporal tracker AQA-Track [36] by $2.9\%$ on SR and multimodal tracker SDSTrack [13] by $1.9\%$ . Figure 6 further provides the tracking performance under 17 challenging attributes. We can see + +![](images/1feae14f806188fdf8dacf77337d8e2b57149d4a27936c4785feaafe897ed536.jpg) +(a) Success (left) and Precision (right) plot on VisEvent + +![](images/57337a777f67e11710bc187f7d303aae6a5f49de325ba53e43414584b266986e.jpg) + +![](images/418ef6ec29d75a1b519eb585be683d153a9c365f80e1d86bb257bce9a788e516.jpg) +(b) Success (left) and Precision (right) plot on FELT + +![](images/f1a8f87336485c44e1cc0a2b1e1dea274e2f514837c37cb96d35a89c65f533e4.jpg) + +![](images/6a6d6f788155dc793cf33733125e6018267ba3e9e75c07a2c8f6c083a67a75c2.jpg) +Figure 5. Comparison of Success (left) and Precision (right) plots on VisEvent (a) and FELT (b). + +![](images/7f75e7012294b9642cb18319095b63aae79476f57174c17e9ee279cf8621a70e.jpg) +Figure 6. Comparison of SR (left) and PR (right) across various attributes using our approach against four top trackers on VisEvent, highlights the performance of Ours/Lowest. + +that our approach still fares best in all degraded conditions, demonstrating its ability to effectively fuse the strengths of RGB and Events. + +FELT. FELT is a large-scale long-term frame-event single object tracking dataset, which contains 742 videos, 1,594,474 RGB frames and corresponding events. The results in Figure 5(b) show that our MamTrack achieves a $48.9\%$ overall SR and $60.8\%$ PR. The results illustrate that our approach achieves effective long-term tracking by exploiting the historical temporal modeling capability of Mamba. We also offer an attribute-based evaluation in the supplementary material for further analysis, our MamTrack achieves the best performance on most attribute splits. + +FE108. FE108 is a visible-event multimodal tracking dataset of indoor scenes, which contains 76 videos for training and 32 for testing. As shown in Table 1, our MamTrack outperforms the previous best tracker ViPT by $0.6\%$ in SR and $0.4\%$ in PR. It demonstrates that our approach is also + +
ATOMDiMPPrDiMPFENetMixFormerTransT
52.8/80.053.4/88.255.2/86.863.4/92.460.4/87.856.7/89.0
SiamFc++STARKOSTrackViPTHDETrackOurs
54.5/85.355.4/83.757.1/89.365.8/93.859.8/92.266.4/94.2
+ +Table 1. Quantitative comparison of SR(\%) / PR(\%) against SOTA trackers on the FE108 dataset. + +
#RGBEventTASCMSCFFSAHSASR(%)PR(%)
1X40.556.1
2X59.075.8
3X60.778.1
4X60.277.4
5X60.077.1
6X60.678.0
7X60.477.8
861.679.2
+ +Table 2. Ablation study for important components on VisEvent. Blank denotes the component is used by default, while $\mathbf{x}$ represents the component is removed. + +highly competitive for short-term tracking scenarios. + +# 4.3. Ablation Study and Analysis + +Impact of Multimodal Input. We assess the effectiveness of multimodal fusion by separately utilizing event-only and RGB-only inputs. As shown in Table 2 #1 and #2, the tracking results with unimodal input are significantly worse than those with multimodal input, demonstrating that our method effectively utilizes both modalities' advantages to improve tracking under degraded conditions. + +Effectiveness of Fusion Mamba Module. There are two key mamba-based components in our Fusion Mamba (FM) module: (i) Target-Aware Scan (TAS) facilitates the interaction between the template and search tokens; (ii) Cross-Modality Scan (CMS) focuses on the fusion of event and RGB modalities in an adaptive manner. To verify their effectiveness, we modify the original model by dropping each of the components and retraining the modified models. The tracking performance in Table 2 #3 and #4 has shown a decrease to varying degrees, indicating the effectiveness and rationality of our designed FM module. In CMS, the gate-based operation Cross-Fusion (CF) is the primary operation that supports the interaction between the two modalities. We thus eliminate the CF to showcase its indispensability. #5 shows that when the CF is removed, the SR drops by $1.6\%$ . This illustrates that the deep interaction among features from both modalities plays a key role. + +One possible factor that impacts the contribution of FM to our tracker is the location of FM insertion in the encoder. To verify this, several additional sets of experiments were + +
LocationSR (%)PR(%)LocationSR (%)PR(%)
1260.077.014-17-2061.178.4
12-1561.078.312-13-1460.177.3
12-15-18 (Ours)61.679.218-19-2060.677.9
+ +Table 3. Comparison of different FM module locations on VisEvent. $x - y - z$ represents our FM module is inserted into the $x$ th, $y$ th, and $z$ th encoder layers. + +
Type of HSANumber of LayersSR(%)PR(%)FPS
None059.976.820.7
Mamba160.578.119.9
361.679.218.1
660.978.116.6
Transformer361.078.416.9
+ +conducted. As shown in Table 3, compared to other 3-stage insertion styles, the 12-15-18 style adopted in this paper provides the best performance, which can be attributed to our insertion strategy that focuses on both low-level texture features and high-level semantic features. Moreover, the tracking performance exhibits a progressive improvement from 12 to 12-15 and then to 12-15-18 style, revealing the growing synergy between RGB and event modalities facilitated by our FM module. + +Effectiveness of Historical Decoder. The Historical Decoder is designed to capture and transmit historical information during tracking, which consists of two main components: Historical State Aware (HSA) and Feature Sequence Attention (FSA) modules. As shown in Table 2 #6, when the FSA is removed, the SR score drops by $1.0\%$ , showing the importance of multimodal features for exploiting historical information. The performance drops by $1.2\%$ when the HSA is removed as shown in #7. This illustrates the effectiveness of Mamba for modeling long sequences. + +We also examine the influence of the number of historical decoder layers on tracking performance. As shown in Table 4, the utilization of the historical decoder greatly enhances the results, with the most performance observed when employing 3 decoder layers. To further verify the effectiveness of the Mamba in the historical decoder, we replace the Mamba architecture with Transformer. Table 4 shows our method with Mamba excels in both performance and speed due to Mamba's linear complexity and its capability to model long sequences. + +Impact of the Length of Historical Information. As shown in Table 5, the model's performance improves as the range of historical information increases from 1 to 4. This + +Table 4. Ablation studies on the type and number of decoder layers. + +
LengthSR(%)PR(%)LengthSR(%)PR(%)
160.177.5859.876.9
260.477.91659.776.6
461.679.2
+ +Table 5. Impact of the length of historical information. + +![](images/e1050d100ba3e2fda10a91f18d4754b16d31e34d9cf9572058ba0aba5d847899.jpg) +Figure 7. Comparison of SR scores for long-term tracking on VisEvent, where the horizontal axis is test sequence length. + +indicates that the model can effectively capture the target's movement trends and appearance changes, resulting in improved performance. However, when the historical information range is extended to 8 and 16, the model's performance slightly declines. This decrease is due to the limited number of video sequences used in each iteration, which impairs the model's generalization ability. + +Effectiveness for Long-term Tracking. To further demonstrate the effectiveness of our Mamba-based historical decoder in capturing temporal information, we analyze its performance in long-term tracking. Specifically, we evaluate the tracking performance of sequences containing more than $i$ frames on the VisEvent individually, where $i \in \{300, 500, 700, 1000, 1500, 2000\}$ . Figure 7 shows that our MamTrack outperforms other SOTA Transformer-based temporal trackers and multimodal trackers by a large margin in SR across all settings. This suggests our Mamba-based approach proves superior in capturing long-term cues compared to the Transformer-based autoregressive scheme. + +# 4.4. Performance on RGB-T Tracking Dataset + +To further validate the generalization ability of our MamTrack, we conducted additional experiments on the RGB-T tracking task. Following the common settings used in previous work [48], we trained our tracker on the LasHeR [16] training set and tested it on both the LasHeR test set and the RGBT234 [15] dataset. + +Results on LasHeR Dataset. LasHeR is the largest RGB-T tracking dataset at present, containing 1224 aligned video sequences with diverse attribute annotations. Of + +
LasHeRmfDimp 34.3/ 44.7SGT++ 25.1/36.5OSTrack 41.2/51.5ProTrack 42.0/53.8HMFT 31.3/ 43.6APFNet 36.2/50.0CEUTrack 49.6/63.5ViPT 52.5/65.1Ours 54.2/67.4
RGBT234mfDimp 42.8/ 64.6CMPP 57.7/82.3OSTrack 54.9/72.9ProTrack 57.9/82.7HMFT 56.8/78.8APFNet 57.9/82.7MFGNet 53.5/78.3ViPT 61.7/ 83.5Ours 62.4/84.4
+ +Table 6. Performance comparison of SR(%) / PR(%) on LasHeR and RGBT234 dataset. + +![](images/699df7a95cd61d3cdc1e7f38b9eb4821bafe859cde968c155142f372a76a7921.jpg) +Figure 8. Visual comparison of our MamTrack against other SOTA trackers under 4 challenging conditions on VisEvent. + +![](images/9703a62830c13349d42c2772401979646b81cbbef28cbbdd060325feeef42e78.jpg) +Figure 9. The comparison attention map in the MamTrack. The first row is the ground truth; the second and third rows display the attention maps in the multimodal encoder and historical decoder, respectively. + +these, 245 sequences are designated as the testing set, while the remaining are used for training. As shown in Table 6, MamTrack performs exceptionally well on RGB-T tracking tasks, effectively leveraging the advantages of both thermal and visible information. This indicates its potential for other multimodal tracking tasks. + +Results on RGBT234 Dataset. RGBT234 is a large-scale RGB-T tracking dataset comprising 234 pairs of videos. As shown in Table 6, our MamTrack surpasses the most advantaged tracker ViPT [48] by $0.7\%$ in Success Rate (SR) and $0.9\%$ in Precision Rate (PR). + +# 4.5. Visualization + +We first provide a visual comparison across four challenging conditions on the VisEvent dataset. As shown in Figure 8, our approach can accurately track the target under all conditions. In particular, our method effectively fuses the advantages of RGB and event to enhance tracker confidence when the target is in scenes with motion blur or fast motion. Besides, in situations where the target is partially occluded or the background is complex, the historical cues encourage our method to recall the target's appearance, thereby minimizing environmental interference. + +Furthermore, we also visualize the attention maps of the encoder and decoder to demonstrate the effectiveness of historical information. As shown in Figure 9, guided by historical information, our MamTrack pays more attention to the target's location under challenging conditions. + +# 5. Conclusion + +We propose MamTrack, a novel approach jointly leverages multimodal spatial and historical temporal information for RGBE object tracking. Our historical decoder, based on the innovative Mamba architecture, effectively captures target appearance changes and motion trends, enabling the model to locate the target within complex backgrounds. We also introduce a simple yet effective Mamba-based fusion method for integrating cross-modality cues. Extensive evaluations demonstrate that our method achieves competitive performance compared to previous SOTA trackers. However, we primarily focuses on the effectiveness of Mamba in modality fusion, but not leverage the high temporal resolution characteristics of event data. Future research will explore this aspect to enhance the performance of the tracker. + +Acknowledgements. This work is supported in part by the National Key Research and Development Program of China (No. 2022ZD0210500), the National Natural Science Foundation of China under Grant 62441216/62332019, the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01), the Ningbo Major Research and Development Plan Project of China (No. 2023Z225), Beijing Natural Science Foundation of China(Grant No.3254040), National Key Laboratory of Human Factors Engineering and China Postdoctoral Science Foundation under Grant (No. 2024M760315). + +# References + +[1] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6182-6191, 2019. 1, 3 +[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 3 +[3] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 1 +[4] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13608-13618, 2022. 1, 3 +[5] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4660-4669, 2019. 1 +[6] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7183-7192, 2020. 3 +[7] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-Mamba for Cross-modality Object Detection, 2024. 2 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2 +[9] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2 +[10] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. IEEE Transactions on Image Processing, 32:6129-6141, 2023. 1 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3 +[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2022. 2 +[13] Xiaojun Hou, Jiazheng Xing, Yijie Qian, Yaowei Guo, Shuo Xin, Junhao Chen, Kai Tang, Mengmeng Wang, Zhengkai Jiang, Liang Liu, et al. Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26551-26561, 2024. 3, 5 +[14] Ju Huang, Shiao Wang, Shuai Wang, Zhe Wu, Xiao Wang, and Bo Jiang. Mamba-fetrack: Frame-event tracking via state space model. arXiv preprint arXiv:2404.18174, 2024.2 +[15] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. Rgb-t object tracking: Benchmark and baseline. Pattern Recognition, 96:106977, 2019. 7 +[16] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for rgbt tracking. IEEE Transactions on Image Processing, 31:392-404, 2021. 7 +[17] Kang Li, Fazhi He, Haiping Yu, and Xiao Chen. A parallel and robust object tracking approach synthesizing adaptive bayesian learning and improved incremental subspace learning. Frontiers of Computer Science, 13(5):1116, 2019. 1 +[18] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2 +[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[20] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual State Space Model, 2024. 2 +[21] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3044-3053, 2021. 1 +[22] Haiyang Mei, Bo Dong, Wen Dong, Jiaxi Yang, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Glass segmentation using intensity and spectral polarization cues. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12622-12631, 2022. +[23] Yu Qiao, Bo Dong, Ao Jin, Yu Fu, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Multiview spectral polarization propagation for video glass segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23218-23228, 2023. 1 +[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5 +[25] Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Shengping Zhang, and Xianxian Li. Explicit visual prompts for visual object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4838-4846, 2024. 3 +[26] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. + +In The Eleventh International Conference on Learning Representations, 2023. 2 +[27] Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2321-2329, 2023. 3 +[28] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1, 2 +[29] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 2023. 2, 5 +[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 7(1):30-40, 2023. 1 +[31] Xiao Wang, Ju Huang, Shiao Wang, Chuanming Tang, Bo Jiang, Yonghong Tian, Jin Tang, and Bin Luo. Long-term frame-event visual tracking: Benchmark dataset and baseline. arXiv preprint arXiv:2403.05839, 2024. 5 +[32] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19248-19257, 2024. 1, 2 +[33] Yang Wang, Bo Dong, Yuji Zhang, Yunduo Zhou, Haiyang Mei, Ziqi Wei, and Xin Yang. Event-enhanced multi-modal spiking neural network for dynamic obstacle avoidance. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3138-3148, 2023. 1 +[34] Yang Wang, Haiyang Mei, Qirui Bao, Ziqi Wei, Mike Zheng Shou, Haizhou Li, Bo Dong, and Xin Yang. Apprenticeship-inspired elegance: synergistic knowledge distillation empowers spiking neural networks for efficient single-eye emotion recognition. arXiv preprint arXiv:2407.09521, 2024. 1 +[35] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9697-9706, 2023. 1, 2, 3 +[36] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19300-19309, 2024. 1, 2, 3, 4, 5 +[37] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2 +[38] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10448-10457, 2021. 1, 3 + +[39] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European conference on computer vision, pages 341-357. Springer, 2022. 5 +[40] Haiwei Zhang, Jiqing Zhang, Bo Dong, Pieter Peers, Wenwei Wu, Xiaopeng Wei, Felix Heide, and Xin Yang. In the blink of an eye: Event-based emotion recognition. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1 +[41] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13043-13052, 2021. 1, 2, 5 +[42] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9781-9790, 2023. 2 +[43] Jiqing Zhang, Bo Dong, Yingkai Fu, Yuanchen Wang, Xiaopeng Wei, Baocai Yin, and Xin Yang. A universal event-based plug-in module for visual object tracking in degraded conditions. International Journal of Computer Vision, 132 (5):1857-1879, 2024. 1 +[44] Lichao Zhang, Abel Gonzalez-Garcia, Joost Van De Weijer, Martin Danelljan, and Fahad Shahbaz Khan. Learning the model update for siamese trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4010-4019, 2019. 3 +[45] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 5 +[46] Aihua Zheng, Juncong Liu, Zi Wang, Lili Huang, Chenglong Li, and Bing Yin. Visible-infrared person re-identification via specific and shared representations learning. Visual Intelligence, 1(1):29, 2023. 1 +[47] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7588-7596, 2024. 1, 3 +[48] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9516-9526, 2023. 1, 3, 7, 8 +[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model, 2024. 2, 4 \ No newline at end of file diff --git a/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/images.zip b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..725c3bf2dc6c1a90119ef06399f8fa6c3561154f --- /dev/null +++ b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e4b543253e2538607992c4d2ccfe4fae0bc673b48874edb1e4bdc3bf60871d1 +size 726375 diff --git a/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/layout.json b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21c0460d02fb54aec46c4a4928eb22367149ae8e --- /dev/null +++ b/2025/Exploring Historical Information for RGBE Visual Tracking with Mamba/layout.json @@ -0,0 +1,9164 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "text", + "content": "Exploring Historical Information for RGBE Visual Tracking with Mamba" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "spans": [ + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": "Chuanyu Sun" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": ", Jiqing Zhang" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": ", Yang Wang" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": ", Huilin Ge" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": ", Qianchen Xia" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": ", Baocai Yin" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{1,5\\dagger}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": ", Xin Yang" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": "Key Laboratory of Social Computing and Cognitive Intelligence, Dalian University of Technology \n" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": "Dalian Maritime University " + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": "Jiangsu University of Science and Technology \n" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": "The Future Laboratory, Tsinghua University, Beijing, China " + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 66, + 142, + 546, + 200 + ], + "type": "text", + "content": "Beijing University of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 253, + 297, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 297, + 565 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 297, + 565 + ], + "type": "text", + "content": "Combining the advantages of conventional and event cameras for robust visual tracing has drawn extensive interest. However, existing tracking approaches heavily engage in complex cross-modal fusion modules, leading to higher computational complexity and training challenges. Besides, these methods generally ignore the effective integration of historical information, which is crucial to grasping the change in the target's appearance and motion trends. Given the recent advancements in Mamba's long-range modeling and linear complexity, we explore its potential in addressing the above issues in RGBE tracking tasks. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to achieve effective modality-selective fusion. This module can be seamlessly integrated into the encoding layer of prevalent Transformer-based backbones. Moreover, we further present a novel historical decoder that leverages Mamba's advanced long sequence modeling to effectively capture the target appearance changes with autoregressive queries. Extensive experiments show that our proposed approach achieves state-of-the-art performance on multiple challenging short-term and long-term RGBE benchmarks. Besides, the effectiveness of each key Mamba-based component of our approach is evidenced by our thorough ablation study. Code will be released at: https://github.com/scy0712/MamTrack" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 590, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 610, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 682 + ], + "type": "text", + "content": "Visual object tracking (VOT) is a fundamental task in computer vision[21-23, 30, 46] that focuses on estimating the position and shape of a target within video sequences based on its initial state [1, 3, 5, 17, 35, 36, 38, 47].Owing to the advantages of high temporal resolution and dynamic range offered by event cameras [33, 34, 40], there has been a" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 227, + 552, + 356 + ], + "blocks": [ + { + "bbox": [ + 320, + 227, + 552, + 356 + ], + "lines": [ + { + "bbox": [ + 320, + 227, + 552, + 356 + ], + "spans": [ + { + "bbox": [ + 320, + 227, + 552, + 356 + ], + "type": "image", + "image_path": "e669452b3e32356a7b28c2128f777926545126898152cda417372448738fc1c1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 366, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 422 + ], + "type": "text", + "content": "Figure 1. Overview of MamTrack and its two main components: (i) Multimodal Encoder is responsible for extracting and fusing features from RGB and event modalities; (ii) Historical Decoder is utilized to mine the temporal information of target appearance changes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 448, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 448, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 312, + 448, + 555, + 604 + ], + "type": "text", + "content": "growing trend of integrating them with conventional RGB cameras for robust visual tracking tasks [10, 28, 32, 41-43, 48]. These RGBE multimodal approaches can significantly enhance tracking performance in challenging conditions, such as fast motion and limited illumination. Although being effective, the above multimodal methods expend a substantial amount of labor on developing overly intricate fusion algorithms. For example, FENet [41] employs a total of 8 self-attention and 4 complex cross-domain attention operations to integrate event and frame features at various levels. Implementing these sophisticated strategies will significantly increase the computational complexity and challenge of the training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Moreover, the complex fusion designs make existing RGBE methods usually overlook the mining and application of historical temporal information, which is crucial for tracking as it captures changes in the target's state. Several RGB-only trackers [1, 4, 38] have been developed to update appearance information by replacing or fusing the templates. Occasionally, these straightforward methods ineffectively filter out unreliable updated templates and thus compromise the performance. To address this issue, some" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6500" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 180 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "type": "image", + "image_path": "7863685608b5213b98bff92b07b281025fe99d931457345ba6c098abf899009a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 187, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 243 + ], + "type": "text", + "content": "Figure 2. A comparison of our method with other recent SOTA trackers in long-term sequence. Benefiting from the proposed Historical Decoder, our method can effectively capture target appearance changes to guide robust long-term tracking even in partial occlusion scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "content": "methods [35, 36] incorporate an autoregressive scheme into the Transformer to transfer temporal information across multiple search frames. However, the Transformer structure lacks intrinsic designs for temporal sequential modeling, resulting in an inherent constraint, particularly in handling long sequences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 567 + ], + "type": "text", + "content": "Recently, Mamba [11], a novel State Space Model (SSM) [9, 12, 26, 37], has emerged as a highly promising approach for efficient long sequence modeling due to convolutional computation and linear complexity. Drawing inspiration from this advancement, as shown in Figure 1, we present mamba-based modules to facilitate robust RGB-E tracking by efficient multimodal fusion and effective historical information transmission. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to enhance the relationship between search and template tokens, as well as the complementarity between RGB and event modalities. This module is then easily inserted into the encoding layer of prevalent transformer-based backbones for combining the advantages of self-attention. Furthermore, we present a novel historical information decoder, which investigates the great potential of mamba in long sequence modeling to capture the appearance of targets by introducing a set of additional autoregressive queries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 664 + ], + "type": "text", + "content": "Extensive experiments on multiple challenging RGBE benchmarks show the effectiveness of our approach, which outperforms existing state-of-the-art (SOTA) methods in both short-term and long-term tracking (see Figure 2). Ablation studies verify the importance of each key component. More importantly, we reveal the considerable potential of Mamba architecture in object tracking tasks. In summary, we make the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "- We present a novel and effective framework, dubbed MamTrack, a pioneering approach to incorporate the Mamba mechanism into RGBE multimodal tracking tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 555, + 120 + ], + "type": "text", + "content": "- Our proposed Mamba-based fusion module and historical decoder are capable of adaptively fusing multimodal information and effectively capturing the historical temporal cues of the target appearance changes, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 156 + ], + "type": "text", + "content": "- The proposed trackers achieve SOTA performance on multiple challenging short-term and long-term RGBE tracking benchmarks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 168, + 406, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 406, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 406, + 180 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 188, + 373, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 188, + 373, + 200 + ], + "spans": [ + { + "bbox": [ + 314, + 188, + 373, + 200 + ], + "type": "text", + "content": "2.1. Mamba" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 206, + 554, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 554, + 277 + ], + "type": "text", + "content": "The concept of the State Space Model (SSM) was first introduced in the S4 model [12], presenting a distinctive architecture capable of effectively modeling global information compared to traditional convolutional neural networks and Transformers. In their general form, SSM can be written as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 289, + 555, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 289, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 373, + 289, + 555, + 302 + ], + "type": "interline_equation", + "content": "\\dot {x} (t) = A (t) x (t) + B (t) u (t), \\tag {1}", + "image_path": "e8e3998a532d31635ad85bbe25789b5fc626f480c0e686aea2442f070d043adf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 374, + 304, + 554, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 304, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 374, + 304, + 554, + 317 + ], + "type": "interline_equation", + "content": "y (t) = C (t) x (t) + D (t) u (t). \\tag {2}", + "image_path": "daa8698e413ac0b2cb9a07a17ce7604df9def5762a209262e427c98e09085914.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "text", + "content": "This process maps a 1-D input sequence " + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "inline_equation", + "content": "u(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "text", + "content": " to a 1-D output sequence " + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "inline_equation", + "content": "y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "text", + "content": " through an implicit " + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "text", + "content": "-D latent state sequence " + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 327, + 555, + 532 + ], + "type": "text", + "content": ". Building on the S4 model, the S5 model [26] reduces complexity to a linear level, while H3 [9] introduces it into language model tasks. Mamba [11] introduces a data-dependent SSM layer and a selection mechanism using parallel scan (S6), achieving higher inference speed and overall metrics compared to equivalent-scale Transformers. Recently, Mamba has been employed in discriminative computer vision models and drawn great attention [7, 14, 18, 20, 49]. For example, ViM [49] applies Mamba to the Vision Transformer [8] architecture. VMamba [20] designs a cross-scan mechanism to bridge the gap between 1-D array scanning and 2-D plain traversing. Unlike previous efforts, we explore a novel architecture for RGB-E tracking based on Mamba, further unleashing its potential in visual tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 540, + 503, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 503, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 503, + 553 + ], + "type": "text", + "content": "2.2. RGB-Event Single Object Tracking" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "Recently, due to the unreliable nature of RGB-only data in challenging scenarios, more and more studies introduced event cameras to enhance tracking performance. These multimodal trackers can be primarily categorised into two categories based on the vision backbone employed. The first one [29, 41, 42] utilizes conventional CNNs to extract features from both RGB and event modalities, followed by a meticulously designed cross-modal attention mechanism for fusion. The second approach [28, 32] adopts the Transformer as the backbone, leveraging the multi-head attention mechanism to effectively model the indispensable correlation relationship between the RGB and the event modalities. To tackle the challenge of limited multimodal tracking data," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6501" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 75, + 550, + 210 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 550, + 210 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 550, + 210 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 550, + 210 + ], + "type": "image", + "image_path": "26e30b1b0babfe40b62458605b793384273b405907bf015cc8227c1ec88641af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 211, + 301, + 293 + ], + "blocks": [ + { + "bbox": [ + 61, + 211, + 301, + 293 + ], + "lines": [ + { + "bbox": [ + 61, + 211, + 301, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 211, + 301, + 293 + ], + "type": "image", + "image_path": "92017b34b2ff1b95432d3efbbcc301a324fff692c55030a9d144dbf5345f80f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "type": "text", + "content": "Figure 3. (a) Details of our MamTrack, which employs the Multimodal Encoder to extract and fuse the features of RGB and event modals and introduces the Historical Decoder to capture the temporal cues related to target state changes; and two key components in the Fusion Mamba (FM) module: (b) Target-Aware Scan (TAS) and (c) Cross-Modality Scan (CMS) module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 212, + 553, + 293 + ], + "blocks": [ + { + "bbox": [ + 306, + 212, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 212, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 553, + 293 + ], + "type": "image", + "image_path": "816a3f267be7141c2e4ea43ffef8c285fce45617f0c9d1d2d61c93f97f03aa7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 392 + ], + "type": "text", + "content": "recent studies [2, 13, 48] incorporate prompt tuning strategies into the Transformer architecture to transfer pretrained RGB-based trackers to multimodal data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 395, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 502 + ], + "type": "text", + "content": "The methods mentioned above primarily emphasise multimodal fusion and overlook the significance of mining historical information, which is crucial for multimodal tracking. In contrast, we explore the great potential of Mamba for multimodal tracking tasks. By combining the gating signals of two modalities in Mamba and leveraging the temporal modeling mechanism of Mamba, our approach is capable of concurrently achieving adaptive RGB-E multimodal fusion and efficient utilization of historical cues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 295, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 539 + ], + "type": "text", + "content": "2.3. Tracking with Historical Temporal Information" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "The historical information can provide valuable insights into the target's movements, appearance, and behaviour, thereby enhancing the tracker's confidence. Based on the number of search images input into the tracker, existing tracking approaches that leverage historical temporal cues can be categorised into two groups. The first one involves only inputting one search image and implementing a dynamic kernel filter [1, 6, 44] or a specialized update branch [4, 27, 38] to capture the target appearance changes with time and dynamically update template images during tracking. However, these methods generally involve handcrafted strategies, requiring tedious tuning for various application conditions. Another line of work [25, 35, 36, 47] takes multiple search sequences as input and then models the se" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 357, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 477 + ], + "type": "text", + "content": "quential evolution of target trajectories across frames by explicitly decoding the entire trajectory sequence step by step. Despite these advances, Transformer architectures employed in these methods are not intrinsically designed for temporal sequential modeling, which presents an inherent limitation. Competing with Transformer, Mamba [11] has shown great potential for long sequence modeling. Therefore, our research endeavors to design a novel Mamba-based historical decoder to effectively transmit temporal information in a parallel training manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 489, + 397, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 397, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 397, + 502 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "The overall architecture of our proposed MamTrack is illustrated in Figure 3, which has two branches with shared weights: RGB branch and Event branch. Each branch receives a template and multiple search images as inputs and corresponding features can be extracted by the Multimodal Encoder. The core of the Multimodal Encoder is our proposed Fusion Mamba (FM) module, which is inserted between Transformer encoder layers to leverage the advantages of both RGB and event modalities. In particular, the RGB modal offers abundant rich texture information, while the event modal is adept at handling challenging scenarios. The fused features are then input into our other primary component, the Historical Encoder, which specifically aims to capture the target appearance changes and motion trends by utilizing temporal cues. Finally, the fused features are multiplied with the decoder's output and then fed into the prediction head to locate the targets." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6502" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "type": "text", + "content": "3.1. Multimodal Encoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": "Following [36], we employ a gradual downsampling network as multimodal encoder, which uses a " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": " patch size for patch embedding. Subsequently, we adopt a total of eight multi-layer perceptron (MLP) layers and two merging layers for gradual downsampling. After this, we obtain the RGB template tokens " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "F_{z}^{R} \\in \\mathbb{R}^{N_{z} \\times D}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": ", RGB search tokens " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "F_{x}^{R} \\in \\mathbb{R}^{N_{x} \\times D}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": ", event template tokens " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "F_{z}^{E} \\in \\mathbb{R}^{N_{z} \\times D}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": ", and event search tokens " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "F_{x}^{E} \\in \\mathbb{R}^{N_{x} \\times D}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "N_{z} = \\frac{H_{z}W_{z}}{16^{2}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "N_{x} = \\frac{H_{x}W_{x}}{16^{2}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "D = 512" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "H_{z}W_{z}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "H_{x}W_{x}" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": " represent the size of the template and search images, respectively. Then, the template tokens and search tokens are concatenated and fed into the " + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 89, + 296, + 247 + ], + "type": "text", + "content": "-layer Transformer-based encoder for joint modeling. The operation of our encoder is defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 252, + 294, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 252, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 89, + 252, + 294, + 266 + ], + "type": "interline_equation", + "content": "F _ {0} ^ {m} = \\mathcal {C} \\left(F _ {z} ^ {m}, F _ {x} ^ {m}\\right), \\quad m \\in \\{R G B, E v e n t \\} \\tag {3}", + "image_path": "3eb358ae1b45eceaa74f031d8428ab4366d5c59b3f6c50bec77f0a1b9c528194.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 268, + 294, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 268, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 71, + 268, + 294, + 282 + ], + "type": "interline_equation", + "content": "F _ {n} ^ {R}, F _ {n} ^ {E} = \\mathcal {E} _ {n} \\left(F _ {n - 1} ^ {R}, F _ {n - 1} ^ {E}\\right), \\quad n = 1, \\dots , N \\tag {4}", + "image_path": "57d260223323e988f3e3b84baa6d521ef4f4b1f3bb73b65c06dc6c5e0ec861d1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": " is the element-wise concatenation; " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_n" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "n_{th}" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": " encoder layer; " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "F_n^R" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "F_n^E" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": " are the RGB and event outputs of the " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "n_{th}" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": " layer, respectively. The above operations facilitate the interaction between the template and search tokens during the RGB and event feature extraction. Next, we describe in detail our Mamba-based fusion module inserted between encoding layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 377, + 188, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 188, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 188, + 389 + ], + "type": "text", + "content": "3.2. Fusion Mamba Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 395, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 514 + ], + "type": "text", + "content": "Our proposed Fusion Mamba (FM) module aims to fuse multimodal semantic information. The key to FM is designed based on the following observations: (i) The interaction between template and search features is crucial as it can help the tracker to learn and prioritize the target-relevant search features; (ii) Directly combining RGB and event modals disregards cases where one of modalities cannot provide valuable cues. In the FM, we present Target-Aware Scan module (TAS) and Cross-Modality Scan module (CMS) to address the above issues, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 634 + ], + "type": "text", + "content": "Target-Aware Scan Module (TAS). The TAS module enhances the interaction between template and search regions by leveraging Mamba's global modeling capability. Compared with the self-attention in Transformer, our TAS employs the bidirectional SSM mechanism [49] to estimate the relevance between all template tokens and each search token. The completeness of template cues can encourage our tracker to focus on informative search tokens. Given " + }, + { + "bbox": [ + 55, + 514, + 296, + 634 + ], + "type": "inline_equation", + "content": "F_{n}^{m} \\in \\{F_{n}^{R}, F_{n}^{E}\\}" + }, + { + "bbox": [ + 55, + 514, + 296, + 634 + ], + "type": "text", + "content": ", as shown in Figure 3 (b), the TAS module is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 640, + 294, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 640, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 73, + 640, + 294, + 654 + ], + "type": "interline_equation", + "content": "X _ {n} ^ {m} = \\mathcal {L} \\left(\\mathcal {N} \\left(F _ {n} ^ {m}\\right)\\right), Z _ {n} ^ {m} = \\sigma \\left(\\mathcal {L} \\left(\\mathcal {N} \\left(F _ {n} ^ {m}\\right)\\right)\\right), \\tag {5}", + "image_path": "23bfa04647a2c561438ae0bf12160f25cf351fd9ee859d6bff353f91b8d0fd6e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 73, + 655, + 294, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 655, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 73, + 655, + 294, + 669 + ], + "type": "interline_equation", + "content": "Y _ {f} = \\mathcal {S S M} _ {f} \\left(\\psi \\left(X _ {n} ^ {m}\\right)\\right), Y _ {b} = \\mathcal {S S M} _ {b} \\left(\\psi \\left(X _ {n} ^ {m}\\right)\\right), \\tag {6}", + "image_path": "90ed717020fa88e776412d9417d77b7f93027948b096da1da259a6dff9c1ddb9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 73, + 670, + 294, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 670, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 73, + 670, + 294, + 685 + ], + "type": "interline_equation", + "content": "H _ {n} ^ {m} = \\mathcal {L} \\left(\\mathcal {N} \\left(Z _ {n} ^ {m} \\odot Y _ {f} + Z _ {n} ^ {m} \\odot Y _ {b}\\right)\\right) + F _ {n} ^ {m}, \\tag {7}", + "image_path": "e6eb0a723ec4ed18f359a1b03fd4911bcadc8948cb4f3f04cd37ac8f412b2bad.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is Linear projection; " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " denotes LayerNorm operation; " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is SiLU activation function; " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is 1D convolu" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "tion operation with the kernel size of 4; " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " means elementwise multiplication; " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "X_{n}^{m}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " is the projection of " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "F_{n}^{m}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "Z_{n}^{m}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " is gate signal; " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "Y_{f}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "Y_{b}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " are the outputs through forward and backward directions of SSM (i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{SSM}_f" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{SSM}_b" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "H_{n}^{m} \\in \\{H_{n}^{R}, H_{n}^{E}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " is the enhanced output of our TAS module in the " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "n_{th}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " encoder layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": "Cross-Modality Scan Module (CMS). CMS is responsible for improving the relevance and complementarity of multimodal features through a simple yet effective cross-fusion scheme. Different from TAS, we only input search tokens into CMS to prevent interference from template features. For simplicity, we will use the RGB domain as an example to describe the specific operation since RGB and event domains function similarly. Given RGB search tokens " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "H_{n}^{R,x}" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "n_{th}" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " layer, as shown in Figure 3 (c), our CMS is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 269, + 553, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 269, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 322, + 269, + 553, + 285 + ], + "type": "interline_equation", + "content": "X _ {n} ^ {R, x} = \\mathcal {L} \\left(\\mathcal {N} \\left(H _ {n} ^ {R, x}\\right)\\right), Z _ {n} ^ {R, x} = \\sigma \\left(\\mathcal {L} \\left(\\mathcal {N} \\left(H _ {n} ^ {R, x}\\right)\\right)\\right), \\tag {8}", + "image_path": "78f35621dfda497f6723c6d833f4508cca8b3debe5cb85b6126532ee050aa5d0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 286, + 553, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 286, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 322, + 286, + 553, + 300 + ], + "type": "interline_equation", + "content": "Y ^ {\\prime} = \\mathcal {S S M} _ {f} \\left(\\psi \\left(X _ {n} ^ {R, x}\\right)\\right) + \\mathcal {S S M} _ {b} \\left(\\psi \\left(X _ {n} ^ {R, x}\\right)\\right), \\tag {9}", + "image_path": "30bc26c033a93aa7c4bbb9eefb4d85392fdbac891e69394c398bb1ace1817a88.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 302, + 553, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 302, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 322, + 302, + 553, + 317 + ], + "type": "interline_equation", + "content": "Y _ {n} ^ {R, x} = \\mathcal {L} \\left(\\mathcal {N} \\left(Y ^ {\\prime} \\odot Z _ {n} ^ {R, x} + Y ^ {\\prime} \\odot Z _ {n} ^ {E, x}\\right)\\right) + H _ {n} ^ {R, x}. \\tag {10}", + "image_path": "092a7d85d0b61d2b7bff996d4d3b952fdd2c738016e4eb33fd328f2cc9a0ebde.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "text", + "content": "The modality-specific gating signals " + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "inline_equation", + "content": "Z_{n}^{R,x}" + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "inline_equation", + "content": "Z_{n}^{E,x}" + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "text", + "content": " are utilized to direct the interaction between modalities, thereby strengthening the primary modality while suppressing the degraded modality. With the assistance of our proposed TAS and CMS, the multimodal encoder can efficiently facilitate the interaction between template and search tokens, as well as between RGB and events modalities. The RGB and event features of the last encoder layer are added together to get the final fused feature " + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "text", + "content": ", which will be used to enhance the functionality of the prediction head and historical decoder." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 463, + 424, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 463, + 424, + 474 + ], + "spans": [ + { + "bbox": [ + 314, + 463, + 424, + 474 + ], + "type": "text", + "content": "3.3. Historical Decoder" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "text", + "content": "Historical temporal information is crucial in visual object tracking as it provides the state changes of targets across frames. To fully exploit the temporal cues, we propose the Historical Decoder with " + }, + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "text", + "content": " layers, each layer consists of two key components: Historical State Aware (HSA) module and Feature Sequence Attention (FSA) module." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "content": "Historical State Aware Module (HSA). Mamba offers a natural advantage over conventional CNNs and Transformers when it comes to long sequence modeling. We thus present Mamba-based HSA module to capture the target historical states of " + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "content": " frames within a sequence into " + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "content": " one-dimensional learnable vectors " + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "inline_equation", + "content": "Q(q_{1},\\ldots ,q_{N})\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "content": ". Mamba models the sequence and outputs " + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "inline_equation", + "content": "Q^{\\prime} = (q^{\\prime},\\dots,q_N^{\\prime})\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "content": ", this is" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 348, + 654, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 654, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 348, + 654, + 553, + 668 + ], + "type": "interline_equation", + "content": "X _ {Q} = \\mathcal {S S M} (\\psi (\\mathcal {L} (Q))), Z _ {Q} = \\sigma (\\mathcal {L} (Q)), \\tag {11}", + "image_path": "101e7cd167b46672666918c7d0cb848a754f867ca238a767268e60591beabe87.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 353, + 670, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 670, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 353, + 670, + 553, + 683 + ], + "type": "interline_equation", + "content": "Q ^ {\\prime} = \\mathcal {L} \\left(\\mathcal {N} \\left(X _ {Q} \\odot Z _ {Q}\\right) + Q. \\right. \\tag {12}", + "image_path": "0b8e52d5ed6fd2020307fd356605fd73ef45b21fdfb6e994ef7330698301728d.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "The output " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " will be subsequently utilized as the input of FSA. It is important to highlight that, with the help of the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6503" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 292, + 223 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 292, + 223 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 292, + 223 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 292, + 223 + ], + "type": "image", + "image_path": "c5dbc520d5001a40e06eeddb685d967ed7e92e3b414773af54ead4a3c3d60602.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 297, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 297, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 297, + 269 + ], + "type": "text", + "content": "Figure 4. Overview of our Historical Decoder with two key components: Historical State Aware (HSA) module, and Feature-Sequence Attention (FSA) module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 287, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 370 + ], + "type": "text", + "content": "SSM mechanism, the hidden state generated by each query at the current moment will subsequently be transmitted and contribute to the estimation in the following moment. In this way, we guarantee the long-term viability of modeling historical cues. Compared to Transformer-based autoregressive models, this significantly increases both inference and training speeds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 455 + ], + "type": "text", + "content": "Feature-Sequence Attention Module (FSA). Our proposed FSA module aims to combine multimodal spatial features " + }, + { + "bbox": [ + 55, + 371, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 55, + 371, + 296, + 455 + ], + "type": "text", + "content": " and historical temporal features " + }, + { + "bbox": [ + 55, + 371, + 296, + 455 + ], + "type": "inline_equation", + "content": "Q^{\\prime}" + }, + { + "bbox": [ + 55, + 371, + 296, + 455 + ], + "type": "text", + "content": " to improve the confidence of the prediction head. Our FSA module employs the multi-head attention mechanism, where historical information sequences are used as queries and multimodal features are used as keys and values:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 460, + 255, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 460, + 255, + 475 + ], + "spans": [ + { + "bbox": [ + 89, + 460, + 255, + 475 + ], + "type": "interline_equation", + "content": "Q ^ {\\prime \\prime} = \\mathcal {N} \\left(\\mathcal {C} \\left(\\left(H _ {1}, \\dots , H _ {n _ {h}}\\right) W ^ {O}\\right)\\right) + Q ^ {\\prime},", + "image_path": "601a92dd172a7570879f551c2b76c8c58179c79e698e32adb44b94b2ec842faa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 477, + 294, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 477, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 91, + 477, + 294, + 496 + ], + "type": "interline_equation", + "content": "H _ {i} = \\mathcal {A T T N} \\left(Q ^ {\\prime} W _ {i} ^ {Q}, \\mathcal {F} _ {s} W _ {i} ^ {K}, \\mathcal {F} _ {s} W _ {i} ^ {V}\\right), \\tag {13}", + "image_path": "ee29c8e69429fed17c3c2bcb73a46450ff79080a98b535043baa45097e97e97d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 495, + 226, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 495, + 226, + 522 + ], + "spans": [ + { + "bbox": [ + 91, + 495, + 226, + 522 + ], + "type": "interline_equation", + "content": "\\mathcal {A T T N} (q, k, v) = \\varphi \\left(\\frac {q k ^ {\\top}}{\\sqrt {d _ {k}}}\\right) v,", + "image_path": "d7d0971774c5b636dcc14d2baa2aa4b6ee9d0606bc6b73f446fb38fc85729448.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": " is the Softmax function; " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "W_{i}^{Q}\\in \\mathbb{R}^{d_{m}\\times d_{k}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "W_{i}^{K}\\in \\mathbb{R}^{d_{m}\\times d_{k}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "W_{i}^{V}\\in \\mathbb{R}^{d_{m}\\times d_{v}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "W^{O}\\in \\mathbb{R}^{n_{h}\\times d_{m}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": " are all learnable parameters. In our method, we employ multihead attention with 8 heads, i.e., " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "n_h = 8" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "d_m = 512" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "d_k = d_v = \\frac{d_m}{8} = 64" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "Q''" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": " is fed into the following decoder layer to generate the final temporal queries " + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 529, + 296, + 639 + ], + "type": "text", + "content": " after being processed by the Feed Forward Network (FFN). Finally, we further combine the spatial and temporal information without introducing any parameters by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 644, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 644, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 137, + 644, + 295, + 658 + ], + "type": "interline_equation", + "content": "F _ {o u t} = \\mathcal {F} _ {s} \\odot \\mathcal {Q} ^ {\\top}, \\tag {14}", + "image_path": "86904535030178580f7626b65069ecd0da30be373139f8fc6ea9e096318d8d34.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represents the multimodal features enriched with historical temporal information, and highlights the position where the target may be located. Then it will be fed into the tracking head for the target bbox prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 407, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 407, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 407, + 83 + ], + "type": "text", + "content": "3.4. Head and Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "text", + "content": "We employ a center-based head as regular [39] to predict the center position and scale of the target. The prediction head outputs include the classification score map, the bounding box size, and the offset size. We use focal loss [19] and GIoU loss [24] for training the network. The total loss " + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 159, + 555, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 159, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 365, + 159, + 555, + 172 + ], + "type": "interline_equation", + "content": "L = L _ {c l s} + \\lambda_ {i o u} L _ {i o u} + \\lambda_ {L _ {1}} L _ {L _ {1}}, \\tag {15}", + "image_path": "722b42b27bd4c354e1c81bc9f0a427da3121ff0a277ad3ad403687b60e9962ec.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "text", + "content": " denotes the weighted focal loss for classification, the generalized GIoU loss " + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "inline_equation", + "content": "L_{iou}" + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "text", + "content": " and L1 loss " + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "text", + "content": " are adopted for bounding box regression, " + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "inline_equation", + "content": "\\lambda_{iou} = 2" + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "inline_equation", + "content": "\\lambda_{L_1} = 5" + }, + { + "bbox": [ + 313, + 182, + 554, + 232 + ], + "type": "text", + "content": " are the hyperparameters." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 243, + 395, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 395, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 395, + 255 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 262, + 441, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 441, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 441, + 275 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "text", + "content": "Implementation Details. We implemented MamTrack using PyTorch and trained it on 8 NVIDIA RTX 4090 GPUs with a batch size of 16. We employed the AdamW optimizer with a weight decay of " + }, + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "text", + "content": " and a learning rate of " + }, + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "text", + "content": ". We utilized pretrained HiViT-Base [45] as the Transformer encoder with 20 layers. Fine-tuning MamTrack on the RGB-Event training set took 50 epochs. Our proposed Fusion Mamba module is inserted into the 12th, 15th, and 18th layers. The decoder consists of 3 layers (i.e., " + }, + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 313, + 279, + 555, + 411 + ], + "type": "text", + "content": "). Besides, we utilize the Hamming window to incorporate positional priors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt three widely used metrics for quantitatively evaluating the tracking performance: Success Rate (SR), Precision Rate (PR), and Overlap Precision " + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "inline_equation", + "content": "(\\mathrm{OP}_T)" + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "text", + "content": ". The SR focuses on the frame where the overlap between ground truth and predicted bounding box is larger than a threshold; " + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{OP}_T" + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "text", + "content": " matches SR but with a specific overlap threshold " + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "text", + "content": ". The PR metric counts the frame on which the center distance between the ground truth and predicted bounding box is within a given threshold." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 528, + 503, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 503, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 503, + 540 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed approach, we compare our tracker with representative state-of-the-art trackers on three widely utilized benchmarks: VisEvent [29], FELT [31], and FE108 [41]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "VisEvent. VisEvent is a large-scale visible-event tracking benchmark, which contains 500 training and 320 testing sequences with 17 attributes. As shown in Figure 5(a), our MamTrack achieves the top SR and PR of " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "61.6\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "79.2\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", respectively, which beats all single- and multimodal SOTA trackers. For example, our method outperforms the pure Transformer-based temporal tracker AQA-Track [36] by " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " on SR and multimodal tracker SDSTrack [13] by " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". Figure 6 further provides the tracking performance under 17 challenging attributes. We can see" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6504" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 175, + 167 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 175, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 175, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 175, + 167 + ], + "type": "image", + "image_path": "1feae14f806188fdf8dacf77337d8e2b57149d4a27936c4785feaafe897ed536.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 168, + 229, + 175 + ], + "lines": [ + { + "bbox": [ + 107, + 168, + 229, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 168, + 229, + 175 + ], + "type": "text", + "content": "(a) Success (left) and Precision (right) plot on VisEvent" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 295, + 167 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 295, + 167 + ], + "type": "image", + "image_path": "57337a777f67e11710bc187f7d303aae6a5f49de325ba53e43414584b266986e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 177, + 176, + 273 + ], + "blocks": [ + { + "bbox": [ + 58, + 177, + 176, + 273 + ], + "lines": [ + { + "bbox": [ + 58, + 177, + 176, + 273 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 176, + 273 + ], + "type": "image", + "image_path": "418ef6ec29d75a1b519eb585be683d153a9c365f80e1d86bb257bce9a788e516.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 275, + 221, + 282 + ], + "lines": [ + { + "bbox": [ + 107, + 275, + 221, + 282 + ], + "spans": [ + { + "bbox": [ + 107, + 275, + 221, + 282 + ], + "type": "text", + "content": "(b) Success (left) and Precision (right) plot on FELT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 176, + 177, + 295, + 273 + ], + "blocks": [ + { + "bbox": [ + 176, + 177, + 295, + 273 + ], + "lines": [ + { + "bbox": [ + 176, + 177, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 176, + 177, + 295, + 273 + ], + "type": "image", + "image_path": "f1a8f87336485c44e1cc0a2b1e1dea274e2f514837c37cb96d35a89c65f533e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 330, + 176, + 426 + ], + "blocks": [ + { + "bbox": [ + 55, + 295, + 295, + 317 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 317 + ], + "type": "text", + "content": "Figure 5. Comparison of Success (left) and Precision (right) plots on VisEvent (a) and FELT (b)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 330, + 176, + 426 + ], + "lines": [ + { + "bbox": [ + 56, + 330, + 176, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 176, + 426 + ], + "type": "image", + "image_path": "6a6d6f788155dc793cf33733125e6018267ba3e9e75c07a2c8f6c083a67a75c2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 176, + 331, + 295, + 428 + ], + "blocks": [ + { + "bbox": [ + 176, + 331, + 295, + 428 + ], + "lines": [ + { + "bbox": [ + 176, + 331, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 176, + 331, + 295, + 428 + ], + "type": "image", + "image_path": "7f75e7012294b9642cb18319095b63aae79476f57174c17e9ee279cf8621a70e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 439, + 295, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 472 + ], + "type": "text", + "content": "Figure 6. Comparison of SR (left) and PR (right) across various attributes using our approach against four top trackers on VisEvent, highlights the performance of Ours/Lowest." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 496, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 531 + ], + "type": "text", + "content": "that our approach still fares best in all degraded conditions, demonstrating its ability to effectively fuse the strengths of RGB and Events." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "content": "FELT. FELT is a large-scale long-term frame-event single object tracking dataset, which contains 742 videos, 1,594,474 RGB frames and corresponding events. The results in Figure 5(b) show that our MamTrack achieves a " + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "inline_equation", + "content": "48.9\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "content": " overall SR and " + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "inline_equation", + "content": "60.8\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "content": " PR. The results illustrate that our approach achieves effective long-term tracking by exploiting the historical temporal modeling capability of Mamba. We also offer an attribute-based evaluation in the supplementary material for further analysis, our MamTrack achieves the best performance on most attribute splits." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "FE108. FE108 is a visible-event multimodal tracking dataset of indoor scenes, which contains 76 videos for training and 32 for testing. As shown in Table 1, our MamTrack outperforms the previous best tracker ViPT by " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " in SR and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " in PR. It demonstrates that our approach is also" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 556, + 130 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 556, + 130 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 556, + 130 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 556, + 130 + ], + "type": "table", + "html": "
ATOMDiMPPrDiMPFENetMixFormerTransT
52.8/80.053.4/88.255.2/86.863.4/92.460.4/87.856.7/89.0
SiamFc++STARKOSTrackViPTHDETrackOurs
54.5/85.355.4/83.757.1/89.365.8/93.859.8/92.266.4/94.2
", + "image_path": "5b6f87f489cdee60f854a81c191711cd15692b961594f00e50b113a9f7a35f13.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 317, + 173, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 313, + 139, + 555, + 160 + ], + "lines": [ + { + "bbox": [ + 313, + 139, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 555, + 160 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of SR(\\%) / PR(\\%) against SOTA trackers on the FE108 dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 173, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 285 + ], + "type": "table", + "html": "
#RGBEventTASCMSCFFSAHSASR(%)PR(%)
1X40.556.1
2X59.075.8
3X60.778.1
4X60.277.4
5X60.077.1
6X60.678.0
7X60.477.8
861.679.2
", + "image_path": "f81d199d5e7670be4dd0eb9be379aa25f6e2cd3241359a066f767f8fbb4bca60.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 293, + 554, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 326 + ], + "type": "text", + "content": "Table 2. Ablation study for important components on VisEvent. Blank denotes the component is used by default, while " + }, + { + "bbox": [ + 313, + 293, + 554, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 293, + 554, + 326 + ], + "type": "text", + "content": " represents the component is removed." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 349, + 526, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 526, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 526, + 361 + ], + "type": "text", + "content": "highly competitive for short-term tracking scenarios." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 371, + 471, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 471, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 471, + 384 + ], + "type": "text", + "content": "4.3. Ablation Study and Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 388, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 472 + ], + "type": "text", + "content": "Impact of Multimodal Input. We assess the effectiveness of multimodal fusion by separately utilizing event-only and RGB-only inputs. As shown in Table 2 #1 and #2, the tracking results with unimodal input are significantly worse than those with multimodal input, demonstrating that our method effectively utilizes both modalities' advantages to improve tracking under degraded conditions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 474, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 676 + ], + "type": "text", + "content": "Effectiveness of Fusion Mamba Module. There are two key mamba-based components in our Fusion Mamba (FM) module: (i) Target-Aware Scan (TAS) facilitates the interaction between the template and search tokens; (ii) Cross-Modality Scan (CMS) focuses on the fusion of event and RGB modalities in an adaptive manner. To verify their effectiveness, we modify the original model by dropping each of the components and retraining the modified models. The tracking performance in Table 2 #3 and #4 has shown a decrease to varying degrees, indicating the effectiveness and rationality of our designed FM module. In CMS, the gate-based operation Cross-Fusion (CF) is the primary operation that supports the interaction between the two modalities. We thus eliminate the CF to showcase its indispensability. #5 shows that when the CF is removed, the SR drops by " + }, + { + "bbox": [ + 313, + 474, + 556, + 676 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 474, + 556, + 676 + ], + "type": "text", + "content": ". This illustrates that the deep interaction among features from both modalities plays a key role." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "One possible factor that impacts the contribution of FM to our tracker is the location of FM insertion in the encoder. To verify this, several additional sets of experiments were" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6505" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 301, + 131 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 301, + 131 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 301, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 301, + 131 + ], + "type": "table", + "html": "
LocationSR (%)PR(%)LocationSR (%)PR(%)
1260.077.014-17-2061.178.4
12-1561.078.312-13-1460.177.3
12-15-18 (Ours)61.679.218-19-2060.677.9
", + "image_path": "be27743dc075a58a3af74a256523a95d52bd8dd71cb95fed2f56557cd1cca189.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 185, + 290, + 275 + ], + "blocks": [ + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "lines": [ + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "text", + "content": "Table 3. Comparison of different FM module locations on VisEvent. " + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "inline_equation", + "content": "x - y - z" + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "text", + "content": " represents our FM module is inserted into the " + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "text", + "content": "th, " + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "text", + "content": "th, and " + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 139, + 295, + 173 + ], + "type": "text", + "content": "th encoder layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 185, + 290, + 275 + ], + "lines": [ + { + "bbox": [ + 62, + 185, + 290, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 185, + 290, + 275 + ], + "type": "table", + "html": "
Type of HSANumber of LayersSR(%)PR(%)FPS
None059.976.820.7
Mamba160.578.119.9
361.679.218.1
660.978.116.6
Transformer361.078.416.9
", + "image_path": "b1e3ccfc731bfeca43d3f1d708cd007c0a5c65499f90a52ac55afde9b5433426.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 328, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 295, + 435 + ], + "type": "text", + "content": "conducted. As shown in Table 3, compared to other 3-stage insertion styles, the 12-15-18 style adopted in this paper provides the best performance, which can be attributed to our insertion strategy that focuses on both low-level texture features and high-level semantic features. Moreover, the tracking performance exhibits a progressive improvement from 12 to 12-15 and then to 12-15-18 style, revealing the growing synergy between RGB and event modalities facilitated by our FM module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 436, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 295, + 556 + ], + "type": "text", + "content": "Effectiveness of Historical Decoder. The Historical Decoder is designed to capture and transmit historical information during tracking, which consists of two main components: Historical State Aware (HSA) and Feature Sequence Attention (FSA) modules. As shown in Table 2 #6, when the FSA is removed, the SR score drops by " + }, + { + "bbox": [ + 54, + 436, + 295, + 556 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 54, + 436, + 295, + 556 + ], + "type": "text", + "content": ", showing the importance of multimodal features for exploiting historical information. The performance drops by " + }, + { + "bbox": [ + 54, + 436, + 295, + 556 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 54, + 436, + 295, + 556 + ], + "type": "text", + "content": " when the HSA is removed as shown in #7. This illustrates the effectiveness of Mamba for modeling long sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 556, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 556, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 556, + 295, + 677 + ], + "type": "text", + "content": "We also examine the influence of the number of historical decoder layers on tracking performance. As shown in Table 4, the utilization of the historical decoder greatly enhances the results, with the most performance observed when employing 3 decoder layers. To further verify the effectiveness of the Mamba in the historical decoder, we replace the Mamba architecture with Transformer. Table 4 shows our method with Mamba excels in both performance and speed due to Mamba's linear complexity and its capability to model long sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 296, + 713 + ], + "type": "text", + "content": "Impact of the Length of Historical Information. As shown in Table 5, the model's performance improves as the range of historical information increases from 1 to 4. This" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 332, + 70, + 539, + 131 + ], + "blocks": [ + { + "bbox": [ + 55, + 284, + 295, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 306 + ], + "type": "text", + "content": "Table 4. Ablation studies on the type and number of decoder layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 70, + 539, + 131 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 539, + 131 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 539, + 131 + ], + "type": "table", + "html": "
LengthSR(%)PR(%)LengthSR(%)PR(%)
160.177.5859.876.9
260.477.91659.776.6
461.679.2
", + "image_path": "495246e818fad6fbb4288ba81e3cc9aa4424fd063192fd84f8756a70bfcf7ec4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 139, + 534, + 151 + ], + "lines": [ + { + "bbox": [ + 334, + 139, + 534, + 151 + ], + "spans": [ + { + "bbox": [ + 334, + 139, + 534, + 151 + ], + "type": "text", + "content": "Table 5. Impact of the length of historical information." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 162, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 315, + 162, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 315, + 162, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 315, + 162, + 553, + 300 + ], + "type": "image", + "image_path": "e1050d100ba3e2fda10a91f18d4754b16d31e34d9cf9572058ba0aba5d847899.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 308, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 331 + ], + "type": "text", + "content": "Figure 7. Comparison of SR scores for long-term tracking on VisEvent, where the horizontal axis is test sequence length." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 351, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 435 + ], + "type": "text", + "content": "indicates that the model can effectively capture the target's movement trends and appearance changes, resulting in improved performance. However, when the historical information range is extended to 8 and 16, the model's performance slightly declines. This decrease is due to the limited number of video sequences used in each iteration, which impairs the model's generalization ability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 436, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 579 + ], + "type": "text", + "content": "Effectiveness for Long-term Tracking. To further demonstrate the effectiveness of our Mamba-based historical decoder in capturing temporal information, we analyze its performance in long-term tracking. Specifically, we evaluate the tracking performance of sequences containing more than " + }, + { + "bbox": [ + 313, + 436, + 556, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 436, + 556, + 579 + ], + "type": "text", + "content": " frames on the VisEvent individually, where " + }, + { + "bbox": [ + 313, + 436, + 556, + 579 + ], + "type": "inline_equation", + "content": "i \\in \\{300, 500, 700, 1000, 1500, 2000\\}" + }, + { + "bbox": [ + 313, + 436, + 556, + 579 + ], + "type": "text", + "content": ". Figure 7 shows that our MamTrack outperforms other SOTA Transformer-based temporal trackers and multimodal trackers by a large margin in SR across all settings. This suggests our Mamba-based approach proves superior in capturing long-term cues compared to the Transformer-based autoregressive scheme." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 587, + 533, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 533, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 533, + 601 + ], + "type": "text", + "content": "4.4. Performance on RGB-T Tracking Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "content": "To further validate the generalization ability of our MamTrack, we conducted additional experiments on the RGB-T tracking task. Following the common settings used in previous work [48], we trained our tracker on the LasHeR [16] training set and tested it on both the LasHeR test set and the RGBT234 [15] dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Results on LasHeR Dataset. LasHeR is the largest RGB-T tracking dataset at present, containing 1224 aligned video sequences with diverse attribute annotations. Of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6506" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 572, + 133 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 572, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 572, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 572, + 133 + ], + "type": "table", + "html": "
LasHeRmfDimp 34.3/ 44.7SGT++ 25.1/36.5OSTrack 41.2/51.5ProTrack 42.0/53.8HMFT 31.3/ 43.6APFNet 36.2/50.0CEUTrack 49.6/63.5ViPT 52.5/65.1Ours 54.2/67.4
RGBT234mfDimp 42.8/ 64.6CMPP 57.7/82.3OSTrack 54.9/72.9ProTrack 57.9/82.7HMFT 56.8/78.8APFNet 57.9/82.7MFGNet 53.5/78.3ViPT 61.7/ 83.5Ours 62.4/84.4
", + "image_path": "b339d6bb256b02c7a923b689f16a35ce7876feb9c6ec7b5e3c4d11f296cb7585.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 139, + 464, + 150 + ], + "lines": [ + { + "bbox": [ + 145, + 139, + 464, + 150 + ], + "spans": [ + { + "bbox": [ + 145, + 139, + 464, + 150 + ], + "type": "text", + "content": "Table 6. Performance comparison of SR(%) / PR(%) on LasHeR and RGBT234 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 173, + 294, + 315 + ], + "blocks": [ + { + "bbox": [ + 57, + 173, + 294, + 315 + ], + "lines": [ + { + "bbox": [ + 57, + 173, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 173, + 294, + 315 + ], + "type": "image", + "image_path": "699df7a95cd61d3cdc1e7f38b9eb4821bafe859cde968c155142f372a76a7921.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 324, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 346 + ], + "type": "text", + "content": "Figure 8. Visual comparison of our MamTrack against other SOTA trackers under 4 challenging conditions on VisEvent." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 376, + 295, + 480 + ], + "blocks": [ + { + "bbox": [ + 57, + 376, + 295, + 480 + ], + "lines": [ + { + "bbox": [ + 57, + 376, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 295, + 480 + ], + "type": "image", + "image_path": "9703a62830c13349d42c2772401979646b81cbbef28cbbdd060325feeef42e78.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 489, + 296, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 534 + ], + "type": "text", + "content": "Figure 9. The comparison attention map in the MamTrack. The first row is the ground truth; the second and third rows display the attention maps in the multimodal encoder and historical decoder, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 572, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 295, + 644 + ], + "type": "text", + "content": "these, 245 sequences are designated as the testing set, while the remaining are used for training. As shown in Table 6, MamTrack performs exceptionally well on RGB-T tracking tasks, effectively leveraging the advantages of both thermal and visible information. This indicates its potential for other multimodal tracking tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Results on RGBT234 Dataset. RGBT234 is a large-scale RGB-T tracking dataset comprising 234 pairs of videos. As shown in Table 6, our MamTrack surpasses the most advantaged tracker ViPT [48] by " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " in Success Rate (SR) and " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " in Precision Rate (PR)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 171, + 398, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 398, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 398, + 182 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 190, + 555, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 310 + ], + "type": "text", + "content": "We first provide a visual comparison across four challenging conditions on the VisEvent dataset. As shown in Figure 8, our approach can accurately track the target under all conditions. In particular, our method effectively fuses the advantages of RGB and event to enhance tracker confidence when the target is in scenes with motion blur or fast motion. Besides, in situations where the target is partially occluded or the background is complex, the historical cues encourage our method to recall the target's appearance, thereby minimizing environmental interference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 373 + ], + "type": "text", + "content": "Furthermore, we also visualize the attention maps of the encoder and decoder to demonstrate the effectiveness of historical information. As shown in Figure 9, guided by historical information, our MamTrack pays more attention to the target's location under challenging conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 390, + 388, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 388, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 388, + 402 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 412, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 581 + ], + "type": "text", + "content": "We propose MamTrack, a novel approach jointly leverages multimodal spatial and historical temporal information for RGBE object tracking. Our historical decoder, based on the innovative Mamba architecture, effectively captures target appearance changes and motion trends, enabling the model to locate the target within complex backgrounds. We also introduce a simple yet effective Mamba-based fusion method for integrating cross-modality cues. Extensive evaluations demonstrate that our method achieves competitive performance compared to previous SOTA trackers. However, we primarily focuses on the effectiveness of Mamba in modality fusion, but not leverage the high temporal resolution characteristics of event data. Future research will explore this aspect to enhance the performance of the tracker." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by the National Key Research and Development Program of China (No. 2022ZD0210500), the National Natural Science Foundation of China under Grant 62441216/62332019, the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01), the Ningbo Major Research and Development Plan Project of China (No. 2023Z225), Beijing Natural Science Foundation of China(Grant No.3254040), National Key Laboratory of Human Factors Engineering and China Postdoctoral Science Foundation under Grant (No. 2024M760315)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6507" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6182-6191, 2019. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "type": "text", + "content": "[4] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13608-13618, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "text", + "content": "[5] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4660-4669, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 390 + ], + "type": "text", + "content": "[6] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7183-7192, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "text", + "content": "[7] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-Mamba for Cross-modality Object Detection, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 555 + ], + "type": "text", + "content": "[9] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 601 + ], + "type": "text", + "content": "[10] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. IEEE Transactions on Image Processing, 32:6129-6141, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 635 + ], + "type": "text", + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 669 + ], + "type": "text", + "content": "[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[13] Xiaojun Hou, Jiazheng Xing, Yijie Qian, Yaowei Guo, Shuo Xin, Junhao Chen, Kai Tang, Mengmeng Wang, Zhengkai Jiang, Liang Liu, et al. Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26551-26561, 2024. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[14] Ju Huang, Shiao Wang, Shuai Wang, Zhe Wu, Xiao Wang, and Bo Jiang. Mamba-fetrack: Frame-event tracking via state space model. arXiv preprint arXiv:2404.18174, 2024.2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[15] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. Rgb-t object tracking: Benchmark and baseline. Pattern Recognition, 96:106977, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "text", + "content": "[16] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for rgbt tracking. IEEE Transactions on Image Processing, 31:392-404, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "type": "text", + "content": "[17] Kang Li, Fazhi He, Haiping Yu, and Xiao Chen. A parallel and robust object tracking approach synthesizing adaptive bayesian learning and improved incremental subspace learning. Frontiers of Computer Science, 13(5):1116, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 277, + 553, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 309 + ], + "type": "text", + "content": "[18] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 311, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 311, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 311, + 553, + 354 + ], + "type": "text", + "content": "[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[20] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual State Space Model, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "text", + "content": "[21] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3044-3053, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "text", + "content": "[22] Haiyang Mei, Bo Dong, Wen Dong, Jiaxi Yang, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Glass segmentation using intensity and spectral polarization cues. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12622-12631, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "type": "text", + "content": "[23] Yu Qiao, Bo Dong, Ao Jin, Yu Fu, Seung-Hwan Baek, Felix Heide, Pieter Peers, Xiaopeng Wei, and Xin Yang. Multiview spectral polarization propagation for video glass segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23218-23228, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "type": "text", + "content": "[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[25] Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Shengping Zhang, and Xianxian Li. Explicit visual prompts for visual object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4838-4846, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6508" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[27] Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2321-2329, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[28] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "type": "text", + "content": "[29] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 7(1):30-40, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 320 + ], + "type": "text", + "content": "[31] Xiao Wang, Ju Huang, Shiao Wang, Chuanming Tang, Bo Jiang, Yonghong Tian, Jin Tang, and Bin Luo. Long-term frame-event visual tracking: Benchmark dataset and baseline. arXiv preprint arXiv:2403.05839, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 322, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 387 + ], + "type": "text", + "content": "[32] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19248-19257, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 444 + ], + "type": "text", + "content": "[33] Yang Wang, Bo Dong, Yuji Zhang, Yunduo Zhou, Haiyang Mei, Ziqi Wei, and Xin Yang. Event-enhanced multi-modal spiking neural network for dynamic obstacle avoidance. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3138-3148, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[34] Yang Wang, Haiyang Mei, Qirui Bao, Ziqi Wei, Mike Zheng Shou, Haizhou Li, Bo Dong, and Xin Yang. Apprenticeship-inspired elegance: synergistic knowledge distillation empowers spiking neural networks for efficient single-eye emotion recognition. arXiv preprint arXiv:2407.09521, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "type": "text", + "content": "[35] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9697-9706, 2023. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 612 + ], + "type": "text", + "content": "[36] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19300-19309, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[37] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[38] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10448-10457, 2021. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 642 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[39] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European conference on computer vision, pages 341-357. Springer, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[40] Haiwei Zhang, Jiqing Zhang, Bo Dong, Pieter Peers, Wenwei Wu, Xiaopeng Wei, Felix Heide, and Xin Yang. In the blink of an eye: Event-based emotion recognition. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[41] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13043-13052, 2021. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "text", + "content": "[42] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9781-9790, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "text", + "content": "[43] Jiqing Zhang, Bo Dong, Yingkai Fu, Yuanchen Wang, Xiaopeng Wei, Baocai Yin, and Xin Yang. A universal event-based plug-in module for visual object tracking in degraded conditions. International Journal of Computer Vision, 132 (5):1857-1879, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "type": "text", + "content": "[44] Lichao Zhang, Abel Gonzalez-Garcia, Joost Van De Weijer, Martin Danelljan, and Fahad Shahbaz Khan. Learning the model update for siamese trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4010-4019, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "text", + "content": "[45] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "text", + "content": "[46] Aihua Zheng, Juncong Liu, Zi Wang, Lili Huang, Chenglong Li, and Bing Yin. Visible-infrared person re-identification via specific and shared representations learning. Visual Intelligence, 1(1):29, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "text", + "content": "[47] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7588-7596, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 553, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 553, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 553, + 553, + 597 + ], + "type": "text", + "content": "[48] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9516-9526, 2023. 1, 3, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 599, + 553, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 642 + ], + "type": "text", + "content": "[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model, 2024. 2, 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6509" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_content_list.json b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f5671344707c0a3fcf3fa1e0d56b61013fa34e7 --- /dev/null +++ b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_content_list.json @@ -0,0 +1,1523 @@ +[ + { + "type": "text", + "text": "Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection", + "text_level": 1, + "bbox": [ + 174, + 130, + 823, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Luo $^{1*}$ Yunkang Cao $^{2*}$ Haiming Yao $^{1*}$ Xiaotian Zhang $^{1}$ Jianan Lou $^{1}$ Yuqi Cheng $^{2}$ Weiming Shen $^{2}$ Wenyong Yu $^{2\\dagger}$ Tsinghua University ${}^{2}$ Huazhong University of Science and Technology {luow23, yhm22, zxt19, ljn22}@mails.tsinghua.edu.cn {cyk_hust, yuqicheng, shenwm, ywy}@hust.edu.cn", + "bbox": [ + 176, + 202, + 818, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly detection (AD) is essential for industrial inspection, yet existing methods typically rely on \"comparing\" test images to normal references from a training set. However, variations in appearance and positioning often complicate the alignment of these references with the test image, limiting detection accuracy. We observe that most anomalies manifest as local variations, meaning that even within anomalous images, valuable normal information remains. We argue that this information is useful and may be more aligned with the anomalies since both the anomalies and the normal information originate from the same image. Therefore, rather than relying on external normality from the training set, we propose INP-Former, a novel method that extracts Intrinsic Normal Prototypes (INPs) directly from the test image. Specifically, we introduce the INP Extractor, which linearly combines normal tokens to represent INPs. We further propose an INP Coherence Loss to ensure INPs can faithfully represent normality for the testing image. These INPs then guide the INP-Guided Decoder to reconstruct only normal tokens, with reconstruction errors serving as anomaly scores. Additionally, we propose a Soft Mining Loss to prioritize hard-to-optimize samples during training. INP-Former achieves state-of-the-art performance in single-class, multi-class, and few-shot AD tasks across MVtec-AD, VisA, and Real-IAD, positioning it as a versatile and universal solution for AD. Remarkably, INP-Former also demonstrates some zero-shot AD capability. Code is available at: https://github.com/luow23/INP-Former.", + "bbox": [ + 89, + 357, + 485, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 811, + 220, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised image anomaly detection (AD) [3, 39] seeks to identify abnormal patterns in images and localize anom", + "bbox": [ + 89, + 837, + 482, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lous regions by learning solely from normal samples. This technique has seen widespread application in industrial defect detection [2, 47] and medical disease screening [19]. Recently, various specialized tasks have emerged in response to real-world demands, from conventional single-class AD [28, 34] to more advanced few-shot AD [17, 22] and multi-class AD [14, 40, 42].", + "bbox": [ + 511, + 327, + 905, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although the composition of normal samples varies across these tasks, the fundamental principle remains unchanged: modeling normality in the training data and assessing whether a test image aligns with this learned normality. However, this approach can be limited due to misaligned normality between the training data and the test image. For instance, prototype-based methods [34] extract representative normal prototypes to capture the normality of training samples. In few-shot AD, intra-class variance may lead to poorly aligned prototypes [17], e.g., hazelnuts in different appearances and positions, as shown in Fig. 1(a). Increasing the sample size can mitigate this problem but at the cost of additional prototypes and reduced inference efficiency. When there are multiple classes, i.e., multi-class AD, prototypes from one class may resemble anomalies from another, like the normal background of hazelnut is similar to the anomalies in cable in Fig. 1(b), leading to misclassification.", + "bbox": [ + 511, + 436, + 906, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several works have focused on extracting normality that is more aligned with the test image. For instance, some studies [13, 17, 45] propose spatially aligning normality within a single class through geometrical transformations. However, spatial alignment is ineffective for certain objects, such as hazelnuts, which exhibit variations beyond spatial positions. Other approaches [27, 40, 41] attempt to divide the normality in the training set into smaller, specific portions and then compare the test image to the corresponding portion of normality, but may still fail to find perfect alignment because of intra-class variances.", + "bbox": [ + 511, + 714, + 908, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rather than attempting to extract more aligned normal-", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9974", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2de706df4c771685ae090d74c4c72bdc33bbc345986ce083e7ddb560289dc046.jpg", + "image_caption": [ + "Figure 1. Motivation for Intrinsic Normal Prototypes (INPs). (a) Pre-stored prototypes from few-shot normal samples may fail to represent all normal patterns. (b) Pre-stored prototypes from one class can be similar to anomalies in another class. (c) The extracted INPs are concise yet well-aligned to the test image, alleviating the issues in (a) and (b)." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 898, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity from the training set, we propose addressing the issue of misaligned normality by leveraging the normality within the test image itself as prototypes, which we term Intrinsic Normal Prototypes (INPs). As illustrated in Fig. 1(c), normal patches within an anomalous test image can function as INPs, and anomalies can be easily detected by comparing them with these INPs. These INPs provide more concise and well-aligned prototypes to the anomalies than those learned from training data, as they typically share the same geometrical context and similar appearances with the abnormal regions within the testing image itself. Accordingly, we explore the prevalence of INPs in various AD scenarios and evaluate their potential to improve AD performance.", + "bbox": [ + 88, + 275, + 480, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although previous work [1] has attempted to utilize INP for anomaly detection, it employs handcrafted aggregated features as prototypes, thus limiting the method to zero-shot texture anomaly detection. In contrast, we introduce a learnable INP Extractor to extract normal features with adaptable shapes as INPs. We also propose an INP Coherence Loss to ensure that the extracted INPs coherently represent the normality within the test image, avoiding the capture of anomalous regions. However, some weakly representative normal regions are challenging to model with a limited set of discrete INPs, resulting in background noise (Fig. 4(c)). To address this issue, we introduce an INP-Guided Decoder, which integrates INPs into a reconstruction-based framework. This decoder leverages combinations of discrete INPs to accurately reconstruct all normal regions while effectively suppressing the reconstruction of anomalous regions, with reconstruction errors serving as anomaly scores. Furthermore, we introduce a Soft Mining Loss that focuses on normal regions that are challenging to reconstruct, i.e., hard samples, thereby improving overall reconstruction quality and enhancing AD performance.", + "bbox": [ + 91, + 474, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach, termed INP-Former, primarily leverages vision transformers (ViTs) for both INP extraction and INP-guided reconstruction. Extensive experiments on MVtec-AD [2], VisA [47], and Real-IAD [37] demonstrate that INP-Former achieves superior performance across multiclass, single-class, and few-shot AD tasks, positioning", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "INP-Former as a universal AD solution. INP-Former also optimizes computational complexity by extracting concise INPs, e.g., images can be represented effectively using only six INPs, as shown in Sec. 4.3.2. Additionally, as demonstrated in Sec. 4.4.2, INP-Former exhibits strong generalization and can even extract INPs for unseen classes, enabling zero-shot AD capabilities. In summary, our main contributions are:", + "bbox": [ + 511, + 275, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate that a single image can contain Intrinsic Normal Prototypes (INPs), offering concise and aligned normality for anomaly detection.", + "- We propose the INP Extractor and incorporate INPs into a reconstruction-based anomaly detection framework using the INP-Guided Decoder.", + "- We introduce the INP Coherence Loss to extract representative INPs and the Soft Mining Loss to enhance reconstruction quality." + ], + "bbox": [ + 511, + 396, + 903, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 547, + 661, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Universal Anomaly Detection", + "text_level": 1, + "bbox": [ + 511, + 573, + 774, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are numerous unsupervised AD tasks, ranging from conventional single-class AD to recent few-shot and multiclass AD setups. We refer to these collectively as universal anomaly detection.", + "bbox": [ + 511, + 597, + 905, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-Class Anomaly Detection: This setup was originally introduced by MVTec-AD [2] and involves developing distinct AD models for each class. Typically, images are embedded into a feature space using a pre-trained encoder, after which various schemes, such as reconstruction-based [5, 29, 44], knowledge-distillation-based [9, 36], prototype-based [34, 45], and embedding-based [26, 43] methods, are employed to learn the normality of the given class. While these approaches achieve strong performance, their reliance on class-specific models limits scalability when dealing with a wide range of classes.", + "bbox": [ + 511, + 657, + 905, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot Anomaly Detection: In practical scenarios, the number of available normal samples may also be limited, motivating the development of few-shot AD methods. In this case, normal samples may not fully capture the variability of normality. To address this challenge, approaches", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9975", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a80e01ce54e0b4a20d117d19b6d5e6f0551585e619cc443ffd876fcfa29e1ab9.jpg", + "image_caption": [ + "Figure 2. Overview of our INP-Former framework for universal anomaly detection. (a) Our model consists of a pre-trained Encoder, an INP Extractor, a Bottleneck, and an INP-Guided decoder. The INP Extractor dynamically extracts intrinsic normal prototypes from a single image, which the INP-Guided Decoder leverages to effectively suppress anomalous features. (b) Detailed architecture of the INP Extractor. (c) Detailed architecture of each layer in the INP-Guided Decoder. (d) Comparison of computational complexity between INP-Guided Attention and Self Attention. It is important to note that the patch token (Encoding) and patch token (Decoding) refer to the patch tokens utilized during the encoding and decoding stages, respectively." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 880, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "such as spatial alignment [17] or contrastive learning [22] are used to create more compact and representative normal embeddings. Recently, Vision-Language Models (VLMs) like CLIP [33] have proven effective for few-shot AD due to their broad, pre-trained knowledge. These VLMs not only provide descriptive visual embeddings but also compute the similarity between text prompts and test images, as seen in works like WinCLIP [21], AnomalyGPT [12], and InCTRL [46]. Some approaches, such as AdaCLIP [4], even enable zero-shot AD through VLMs.", + "bbox": [ + 88, + 460, + 480, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Class Anomaly Detection: Developing separate models for each class can be resource-intensive, prompting interest in multi-class AD, also known as unified AD [42], which aims to build a single model for multiple classes. UniAD [42] pioneered a unified reconstruction framework for anomaly detection, followed by HVQ-Trans [27] addressed the identical shortcut problem using a vector quantization framework. More recent approaches, such as MambaAD [15] and Dinomaly [14], further enhance multi-class AD performance by leveraging advanced models, i.e., the State Space Model Mamba [11] and DINO [7], respectively. However, these methods lack the functionality to derive aligned normality with the test image. On the contrary, we extract INPs from the testing image, bringing aligned and precise normality for anomaly detection.", + "bbox": [ + 89, + 611, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prototype Learning", + "text_level": 1, + "bbox": [ + 89, + 847, + 277, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prototype learning [35] aims to extract representative prototypes from a given training set, which are then used for", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "classification by measuring their distances to a test sample in a metric space. This technique is widely used in few-shot learning [23]. Several AD methods also employ prototype learning. For example, PatchCore [34] extracts multiple normal prototypes to represent the normality of the training data, directly computing the minimal distances to the test sample for anomaly detection. Other approaches [10, 18, 30, 32] incorporate prototypes into the reconstruction process to avoid the identical shortcut issue. Specifically, they replace the original inputs with combinations of learned normal prototypes, ensuring that the inputs to the reconstruction model contain only normal elements. However, these methods rely on pre-stored normal prototypes extracted from the training set, which can suffer from the misaligned normality problem. In contrast, our INPs are dynamically extracted from the test image, providing more aligned alternatives for normality representation.", + "bbox": [ + 511, + 460, + 906, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method: INP-Former", + "text_level": 1, + "bbox": [ + 511, + 731, + 718, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 756, + 624, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To fully exploit the advantages of INPs in anomaly detection, we propose INP-Former, as depicted in Fig. 2(a). The model dynamically extracts INPs from a single image and utilizes them to guide the feature reconstruction process, with the reconstruction errors serving as anomaly scores. Specifically, it comprises four key modules: a fixed pre-trained Encoder $\\mathcal{Q}$ , an INP Extractor $\\mathcal{E}$ , a Bottleneck $\\mathcal{B}$ , and an INP-Guided Decoder $\\mathcal{D}$ . The input im", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9976", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "age $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ is first processed by the pre-trained Encoder $\\mathcal{Q}$ to extract multi-scale latent features $f_{\\mathcal{Q}} = \\{f_{\\mathcal{Q}}^{1}, \\ldots, f_{\\mathcal{Q}}^{L} | f_{\\mathcal{Q}}^{l} \\in \\mathbb{R}^{N \\times C}, N = \\frac{HW}{k^{2}}\\}$ , where $k$ represents the downsampling factor. Next, the INP Extractor $\\mathcal{E}$ extracts $M$ INPs $\\mathbf{P} = \\{p_{1}, \\ldots, p_{M} | p_{m} \\in \\mathbb{R}^{C}\\}$ from the pre-trained features, with an INP coherence loss ensuring that the extracted INPs consistently represent normal features during testing. The Bottleneck $\\mathcal{B}$ subsequently fuses the multi-scale latent features, producing the fused output $F_{\\mathcal{B}} = \\mathcal{B}(f_{\\mathcal{Q}})$ . Following the bottleneck, the extracted INPs are utilized to guide the Decoder $\\mathcal{D}$ to yield reconstruction outputs $f_{\\mathcal{D}} = \\{f_{\\mathcal{D}}^{1}, \\ldots, f_{\\mathcal{D}}^{L} | f_{\\mathcal{D}}^{l} \\in \\mathbb{R}^{N \\times C}\\}$ with only normal patterns, thus the reconstruction error between $f_{\\mathcal{Q}}$ and $f_{\\mathcal{D}}$ can serve as the anomaly score.", + "bbox": [ + 89, + 89, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. INP Extractor", + "text_level": 1, + "bbox": [ + 89, + 311, + 238, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing prototype-based methods [10, 31, 34] store local normal features from the training data and compare them with test images. However, the misaligned normality between these pre-stored prototypes and the test images and the lack of global information lead to suboptimal detection performance. To address these limitations, we propose the INP Extractor to dynamically extract INPs with global information from the test image itself.", + "bbox": [ + 89, + 334, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as illustrated in Fig. 2(b), instead of extracting representative local features as done in PatchCore [34], we employ cross attention to aggregate the global semantic information of the pre-trained features $\\mathbf{F}_{\\mathcal{Q}} \\in \\mathbb{R}^{N \\times C}$ with $M$ learnable tokens $\\mathbf{T} = \\{t_1, \\dots, t_M | t_m \\in \\mathbb{R}^C\\}$ . Here $\\mathbf{F}_{\\mathcal{Q}}$ is used as the key-value pairs, while $\\mathbf{T}$ serve as the query, allowing $\\mathbf{T}$ to linearly aggregate $\\mathbf{F}_{\\mathcal{Q}}$ into INPs $\\mathbf{P} = \\{p_1, \\dots, p_M | p_m \\in \\mathbb{R}^C\\}$ .", + "bbox": [ + 89, + 455, + 483, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathcal {Q}} = \\operatorname {s u m} \\left(\\left\\{f _ {\\mathcal {Q}} ^ {1}, \\dots , f _ {\\mathcal {Q}} ^ {L} \\right\\}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 587, + 331, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\mathbf {T} W ^ {Q}, K = \\mathbf {F} _ {\\mathcal {Q}} W ^ {K}, V = \\mathbf {F} _ {\\mathcal {Q}} W ^ {V} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 607, + 480, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} ^ {\\prime} = \\operatorname {A t t e n t i o n} (Q, K, V) + \\mathbf {T}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 627, + 361, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\mathrm {F F N} (\\mathbf {T} ^ {\\prime}) + \\mathbf {T} ^ {\\prime}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 647, + 290, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sum}(\\cdot)$ denotes the element-wise summation. $Q\\in$ $\\mathbb{R}^{M\\times C}$ and $K,V\\in \\mathbb{R}^{N\\times C}$ represent the query, key and value, respectively. $W^{Q},W^{k},W^{v}\\in \\mathbb{R}^{C\\times C}$ are the learnable projection parameters for $Q,K,V;\\mathrm{FFN}(\\cdot)$ represents the feed-forward network.", + "bbox": [ + 89, + 672, + 483, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure that INPs coherently represent normal features while minimizing the capture of anomalous features during the testing process, we propose an INP coherence loss $\\mathcal{L}_c$ to minimize the distances between individual normal features and the corresponding nearest INP.", + "bbox": [ + 89, + 750, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d _ {i} = \\min _ {m \\in \\{1, \\dots , M \\}} \\mathcal {S} (\\mathbf {F} _ {\\mathcal {Q}} (i), p _ {m}) \\\\ \\mathcal {L} _ {c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} d _ {i} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 835, + 480, + 902 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3c15be7c66ea7c3406bfd2ee66728257802b3a8b70a11ac6f0a6e83a1c52a401.jpg", + "table_caption": [ + "Table 1. Comparison of computational cost and memory usage." + ], + "table_footnote": [], + "table_body": "
Number of multiplicaiton and addition
CalculationVanilla Self AttentionINP-Guided Attention
Al=Ql(Kl)T943 496 9607 220 640
fdl-1'=AlVl943 509 5046 623 232
Memory usage (MB)
Ql/Kl/Vl/Al2.30/2.30/2.30/2.342.30/0.018/0.018/0.018
", + "bbox": [ + 524, + 108, + 893, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S(\\cdot, \\cdot)$ denotes the cosine distance. $d_{i}$ represents the distance between the query feature $\\mathbf{F}_{\\mathcal{Q}}(i)$ and the corresponding nearest INP item. Fig. 4 visually illustrates the effectiveness of $\\mathcal{L}_c$ .", + "bbox": [ + 511, + 231, + 903, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. INP-Guided Decoder", + "text_level": 1, + "bbox": [ + 511, + 299, + 712, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While we can use the distance between testing features and their nearest INPs for anomaly detection, as illustrated in Fig. 4(c), certain low-representative normal regions are difficult to model with a limited number of discrete INPs, leading to noisy distance maps between these INPs and testing features. To address this issue, we propose the INP-Guided Decoder, aiming to reconstruct these low-representative normal regions through a combination of multiple discrete INPs and suppress the reconstruction of anomalous regions. Additionally, this decoder provides a token-wise discrepancy that can be directly leveraged for anomaly detection. As shown in Fig. 2(c), INPs are incorporated into this decoder to guide the reconstruction process. Since INPs exclusively represent normal patterns in test images, we employ the extracted INPs as key-value pairs, ensuring that the output is a linear combination of normal INPs, thereby effectively suppressing the reconstruction of anomalous queries, i.e., the identical mapping issue [42]. Furthermore, we find that the first residual connection can directly introduce anomalous features into the subsequent reconstruction, so we remove this connection in our INP-Guided Decoder. Following the previous work [20], we also employ the ReLU activation function to mitigate the influence of weak correlations and noise on the attention maps.", + "bbox": [ + 511, + 321, + 906, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mathematically, let $f_{\\mathcal{D}}^{l - 1} \\in \\mathbb{R}^{N \\times C}$ denotes the output latent features from previous decoding layer. The output $f_{\\mathcal{D}}^{l} \\in \\mathbb{R}^{N \\times C}$ of the $l_{th}$ decoding layer is formulated as,", + "bbox": [ + 511, + 683, + 905, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {l} = f _ {\\mathcal {D}} ^ {l - 1} W _ {l} ^ {Q}, K _ {l} = \\mathbf {P} W _ {l} ^ {K}, V _ {l} = \\mathbf {P} W _ {l} ^ {V}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 738, + 846, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}} = A _ {l} V _ {l}, A _ {l} = \\operatorname {R e L U} \\left(Q _ {l} \\left(K _ {l}\\right) ^ {T}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 760, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {D}} ^ {l} = \\operatorname {F F N} \\left(f _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}}\\right) + f _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 782, + 751, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{l}\\in \\mathbb{R}^{N\\times C}$ and $K_{l},V_{l}\\in \\mathbb{R}^{M\\times C}$ denote the query, key and value of the $l_{th}$ decoding layer. $W_{l}^{Q},W_{l}^{k},W_{l}^{v}\\in$ $\\mathbb{R}^{C\\times C}$ denote the learnable projection parameters for $Q_{l},K_{l},V_{l}$ . $A_{l}\\in \\mathbb{R}^{N\\times M}$ represent the attention map.", + "bbox": [ + 511, + 808, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention Complexity Analysis: As depicted in Fig. 2(d), the computational complexity of vanilla self-attention is", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9977", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{O}(N^2 C)$ , while its memory usage is $\\mathcal{O}(N^2)$ . In contrast, our INP-Guided Attention reduces both the computational complexity and memory usage to $\\mathcal{O}(NMC)$ and $\\mathcal{O}(NM)$ , respectively, which can be approximated as $\\mathcal{O}(NC)$ and $\\mathcal{O}(N)$ due to $M \\ll N$ . Tab. 1 offers a detailed comparison of the complexity of vanilla self-attention and INP-Guided Attention. Appendix Sec. C compares the overall complexity between INP-Former and other methods. The light version of INP-Former can even be more efficient than MambaAD [15] yet demonstrate better performance.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Soft Mining Loss", + "text_level": 1, + "bbox": [ + 89, + 251, + 261, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by Focal Loss [25], different regions should be assigned varying weights based on their optimization difficulty. Accordingly, we propose Soft Mining Loss to encourage the model to focus more on difficult regions.", + "bbox": [ + 89, + 273, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the ratio of the reconstruction error of an individual normal region to the average reconstruction error of all normal regions can serve as an indicator of optimization difficulty. Given the encoder $f_{Q}^{l}$ and decoder $f_{D}^{l}$ features at layer $l$ , let $M^{l}$ denote the regional cosine distance. Our soft mining loss $\\mathcal{L}_{sm}$ is defined as follows:", + "bbox": [ + 89, + 335, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {l} (h, w) = \\left[ \\frac {M ^ {l} (h , w)}{u (M ^ {l})} \\right] ^ {\\gamma}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 436, + 325, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s m} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} 1 - \\frac {\\operatorname {v e c} \\left(f _ {\\mathcal {Q}} ^ {l}\\right) ^ {T} \\cdot \\operatorname {v e c} \\left(\\hat {f} _ {\\mathcal {D}} ^ {l}\\right)}{\\left| \\left| \\operatorname {v e c} \\left(f _ {\\mathcal {Q}} ^ {l}\\right) \\right| \\right| \\left| \\left| \\operatorname {v e c} \\left(\\hat {f} _ {\\mathcal {D}} ^ {l}\\right) \\right| \\right|} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 474, + 482, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {\\mathcal {D}} ^ {l} (h, w) = c g \\left(f _ {\\mathcal {D}} ^ {l} (h, w)\\right) _ {w ^ {l} (h, w)}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 518, + 369, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u(M^l)$ represents the average regional cosine distance within a batch, $\\gamma \\geq 0$ denotes the temperature hyperparameter, $cg(\\cdot)_{w^{l}(h,w)}$ denotes a gradient adjustment based on dynamic weight $w^{l}(h,w)$ , and $vec(\\cdot)$ denotes the flattening operation. The overall training loss of our INP-Former can be expressed as follows: $\\mathcal{L}_{total} = \\mathcal{L}_{sm} + \\lambda \\mathcal{L}_c$ .", + "bbox": [ + 89, + 549, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 656, + 223, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 680, + 297, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We conduct a comprehensive analysis of the proposed INP-Former on three widely used AD datasets: MVTec-AD [2], VisA [47], and Real-IAD [37]. MVTec-AD consists of 15 categories, with 3,629 normal images for training, and 1,982 anomalous images along with 498 normal images for testing. VisA contains 12 object categories, with 8,659 normal images for training and 962 normal images along with 1,200 anomalous images for testing.", + "bbox": [ + 89, + 704, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-IAD includes 30 different objects, with 36,645 normal images for training and 63,256 normal images along with 51,329 anomalous images for testing.", + "bbox": [ + 89, + 824, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: Following existing works [14, 15], we use the Area Under the Receiver Operating Characteristic Curve", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(AUROC), Average Precision (AP), and F1-score-max (F1_max) to evaluate anomaly detection and localization. For anomaly localization specifically, we use Area Under the Per-Region-Overlap (AUPRO) as an additional metric.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: INP-Former adopts ViT-Base/14 with DINO2-R [7] weights as the default pre-trained encoder. The number $M$ of INPs is set to six by default. The INP Extractor includes a standard Vision Transformer block. The layer number of the INP-Guided decoder is eight. All input images are resized to $448^2$ and then center-cropped to $392^2$ . The hyperparameters $\\gamma$ and $\\lambda$ are set to 3.0 and 0.2, respectively. We use the StableAdamW [38] optimizer with a learning rate $1e^{-3}$ and a weight decay of $1e^{-4}$ for 200 epochs. Notably, the above hyperparameters do not require any adjustment across the three datasets. Appendix Sec. A presents more implementation details. Appendix Sec. D, E, and F analyze the influence of input resolution, ViT architecture, and $\\lambda$ , respectively.", + "bbox": [ + 511, + 152, + 906, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 369, + 653, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Multi-Class Anomaly Detection", + "text_level": 1, + "bbox": [ + 511, + 393, + 787, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare the proposed INP-Former with several state-of-the-art (SOTA) methods for multi-class anomaly detection, including reconstruction-based methods RD4AD [9], UniAD [42], DiAD [16], MambaAD [15], and Dinomaly [14], and embedding-based methods SimpleNet [26] and DeSTSeg [43]. A detailed introduction to the comparison methods can be found in Appendix Sec. B.", + "bbox": [ + 511, + 417, + 905, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental results on the three AD datasets are presented in Tab. 2. On the widely used MVTec-AD dataset, our method achieves SOTA performance, with image-level metrics of 99.7/99.9/99.2 and pixel-level metrics of 98.5/71.0/69.7/94.9. On VisA, our method achieves competitive results, attaining the best image-level metrics of 98.9/99.0/96.6, and achieving the best or second-best pixel-level performance of 98.8/51.2/54.7/94.4. On the more complex and challenging Real-IAD dataset, our method reaches new SOTA performance, with image-level metrics of 90.5/88.1/81.5 and pixel-level metrics of 99.0/47.5/50.3/95.0. Compared to the second-best results, our method improves by $1.2\\uparrow /1.3\\uparrow /1.3\\uparrow$ at the image level and by $0.2\\uparrow /4.7\\uparrow /3.2\\uparrow /1.1\\uparrow$ at the pixel level. The SOTA performance achieved across the three datasets showcases the effectiveness and robustness of our method. The per-class performance metrics are presented in Appendix Sec. H. Appendix Sec. G presents the performance of INP-Former in a more challenging scenario, which we call supermulti-class anomaly detection, thus training INP-Former on several datasets - MVTec-AD, VisA, and Real-IAD - simultaneously. Results show that our method can even detect anomalies in more classes without significant performance degradation. Fig. 3 demonstrates the precise anomaly localization capability of our method. More qualitative results", + "bbox": [ + 511, + 523, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9978", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/93a44a51a33ac0d271d7ba3ebc9232b1b1734bb5df4d808f66efcd993ec20104.jpg", + "table_caption": [ + "Table 2. Multi-class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined." + ], + "table_footnote": [], + "table_body": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Metric →Image-level(I-AUROC/I-AP/I-F1_max)Pixel-level(P-AUROC/P-AP/P-F1_max/AUPRO)
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
RD4AD [9]94.6/96.5/95.296.1/48.6/53.8/91.192.4/92.4/89.698.1/38.0/42.6/91.882.4/79.0/73.997.3/25.0/32.7/89.6
UniAD [42]96.5/98.8/96.296.8/43.4/49.5/90.788.8/90.8/85.898.3/33.7/39.0/85.583.0/80.9/74.397.3/21.1/29.2/86.7
SimpleNet [26]95.3/98.4/95.896.9/45.9/49.7/86.587.2/87.0/81.896.8/34.7/37.8/81.457.2/53.4/61.575.7/2.8/6.5/39.0
DeSTSeg [43]89.2/95.5/91.693.1/54.3/50.9/64.888.9/89.0/85.296.1/39.6/43.4/67.482.3/79.2/73.294.6/37.9/41.7/40.6
DiAD [16]97.2/99.0/96.596.8/52.6/55.5/90.786.8/88.3/85.196.0/26.1/33.0/75.275.6/66.4/69.988.0/2.9/7.1/58.1
MambaAD [15]98.6/99.6/97.897.7/56.3/59.2/93.194.3/94.5/89.498.5/39.4/44.0/91.086.3/84.6/77.098.5/33.0/38.7/90.5
Dinomaly [14]99.6/99.8/99.098.4/69.3/69.2/94.898.7/98.9/96.298.7/53.2/55.7/94.589.3/86.8/80.298.8/42.8/47.1/93.9
INP-Former99.7/99.9/99.298.5/71.0/69.7/94.998.9/99.0/96.698.9/51.2/54.7/94.490.5/88.1/81.599.0/47.5/50.3/95.0
", + "bbox": [ + 109, + 109, + 883, + 303 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c98dd9ba501bbd8735457b61a545e3e876be3647c013a5ec863380f319b8c309.jpg", + "table_caption": [ + "Table 3. Few-shot (4-shot) anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined. $\\dagger$ indicates the results we reproduced using publicly available code." + ], + "table_footnote": [], + "table_body": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
SPADE [6]84.8/92.5/91.592.7/-/46.2/87.081.7/83.4/82.196.6/-/43.6/87.350.8†/45.8†/61.2†59.5†/0.2†/0.5†/19.2†
PaDiM [8]80.4/90.5/90.292.6/-/46.1/81.372.8/75.6/78.093.2/-/24.6/72.660.3†/53.5†/64.0†90.9†/2.1†/5.1†/67.6†
PatchCore [34]88.8/94.5/92.694.3/-/55.0/84.385.3/87.5/84.396.8/-/43.9/84.966.0†/62.2†/65.2†92.9†/9.8†/16.1†/68.6†
WinCLIP [21]95.2/97.3/94.796.2/-/59.5/89.087.3/88.8/84.297.2/-/47.0/87.673.0†/61.8†/61.0†93.8†/13.3†/21.0†/76.4†
PromptAD [24]96.6/-/-96.5/-/-/90.589.1/-/-97.4/-/-/86.259.7†/43.5†/52.9†86.9†/8.7†/16.1†/61.9†
INP-Former97.6/98.6/97.097.0/65.9/65.6/92.996.4/96.0/93.097.7/49.3/54.3/93.176.7/72.3/71.797.3/32.2/36.7/89.0
", + "bbox": [ + 112, + 339, + 883, + 477 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3b0024b7decfb692714c19723fd49bd86b99e1b36937b7a9b9d434c3df8fd2f.jpg", + "image_caption": [ + "Figure 3. Qualitative results of anomaly localization on the MVTec-AD [2], VisA [47], and Real-IAD [37] datasets for multi-class anomaly detection. The first row presents the input images with their ground truth, while the second row displays the corresponding anomaly maps." + ], + "image_footnote": [], + "bbox": [ + 114, + 479, + 883, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are presented in Appendix Sec. L.", + "bbox": [ + 89, + 667, + 316, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Few-Shot Anomaly Detection", + "text_level": 1, + "bbox": [ + 89, + 707, + 349, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with several SOTA approaches for few-shot anomaly detection, including prototype-based methods SPADE [6], PaDiM [8], and PatchCore [34] and recent advances that utilize VLMs, i.e., WinCLIP [21] and PromptAD [24].", + "bbox": [ + 89, + 733, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 3, our method significantly outperforms previous SOTAs on three different AD datasets. Compared to the second-best results, our method achieves improvements of $1.0\\uparrow /1.3\\uparrow /2.3\\uparrow$ in image-level scores and $0.5\\uparrow / - / 6.1\\uparrow /2.4\\uparrow$ in pixel-level scores on the MVTec-AD dataset. It similarly outperforms the second-best results on", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the VisA dataset, with enhancements of $7.3 \\uparrow / 7.2 \\uparrow / 8.7 \\uparrow$ for image-level and $0.3 \\uparrow / -7.3 \\uparrow / 5.5 \\uparrow$ for pixel-level scores. Additionally, on the Real-IAD dataset, our method surpasses the second-best results by $3.7 \\uparrow / 10.1 \\uparrow / 6.5 \\uparrow$ in image-level and $3.5 \\uparrow / 18.9 \\uparrow / 15.7 \\uparrow / 12.6 \\uparrow$ in pixel-level scores. The superior performance of our method in few-shot anomaly detection stems from its ability to extract INPs from a single image, eliminating the need for extensive normal data to pre-store prototypes. More comparison results on 1-shot and 2-shot are presented in Appendix Sec. I.", + "bbox": [ + 511, + 667, + 906, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 Single-Class Anomaly Detection", + "text_level": 1, + "bbox": [ + 511, + 830, + 790, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further compared our proposed INP-Former with current SOTA methods for single-class anomaly detection, as shown in Tab. 4. The results indicate that INP-Former", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9979", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dc3a92e4779bb3e4e57ed6bee5ca232c18eafc6bebfe6a70a59c4f4dea330733.jpg", + "table_caption": [ + "Table 4. Single class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined." + ], + "table_footnote": [], + "table_body": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓I-AUROCP-APAUPROI-AUROCP-APAUPROI-AUROCP-APAUPRO
PatchCore [34]99.156.193.595.140.191.289.4-91.5
RD4AD [9]98.558.093.996.027.770.987.1-93.8
SimpleNet [26]99.654.890.096.836.388.788.5-84.6
Dinomaly [14]99.768.995.098.950.795.192.045.295.1
INP-Former99.770.295.498.549.293.892.148.195.6
", + "bbox": [ + 171, + 108, + 821, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27119f24e8625ddf68999d64bda5e72e82ccffe0e92298a2df78716a9b4f5ef4.jpg", + "table_caption": [ + "Table 5. Overall ablation on MVTec-AD [2] and VisA [47] datasets. \"INP\" refers to the use of INP Extractor and INP-Guided Decoder." + ], + "table_footnote": [], + "table_body": "
Dataset →MVTec-AD [2]VisA [47]
“INP”LcLsmImage-levelPixel-levelImage-levelPixel-level
Module×××98.59/99.18/97.6397.19/61.73/62.94/92.7396.58/97.18/92.8997.50/47.24/51.90/82.85
✓××99.53/99.80/98.8198.32/69.82/69.38/94.6998.11/98.23/95.2298.41/50.34/54.23/93.63
✓✓×99.61/99.83/99.0298.39/70.01/69.53/95.1098.16/98.30/95.4798.46/51.09/54.46/93.71
✓✓✓99.67/99.88/99.2098.48/71.02/69.65/94.8798.90/99.02/96.5798.90/51.22/54.74/94.36
", + "bbox": [ + 133, + 248, + 864, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b419ae5c178a17756174ffa632fc883e6d2df30b7f6d1bd7c9d76d4464b8239b.jpg", + "image_caption": [ + "Figure 4. Visualization of the impact of INP coherence loss $\\mathcal{L}_c$ . (a) Input anomalous image and ground truth. (b) Distance map without $\\mathcal{L}_c$ . (c) Distance map with $\\mathcal{L}_c$ . The distance map is obtained by calculating the distance between the input features and their nearest INP terms, as described in Eq. 2." + ], + "image_footnote": [], + "bbox": [ + 96, + 375, + 480, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves new SOTA performance on the MVTec-AD and Real-IAD datasets and demonstrates competitive performance on the VisA dataset. Per-category performance of INP-Former is presented in Appendix Sec. J.", + "bbox": [ + 89, + 599, + 483, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 675, + 243, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Overall Ablation", + "text_level": 1, + "bbox": [ + 89, + 700, + 261, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 5, we conduct comprehensive experiments on MVTec-AD [2] and VisA [47] to validate the effectiveness of the proposed components, i.e., INP Extractor and INP-Guided Decoder (\"INP\"), INP Coherence Loss $(\\mathcal{L}_c)$ , and Soft Mining Loss $(\\mathcal{L}_{sm})$ . In the first row, we train a baseline model without incorporating any proposed module, similar to the RD4AD [9] framework. The results in the second row demonstrate that \"INP\" significantly enhances overall performance. This improvement arises from the fact that \"INP\" introduces an information bottleneck, which effectively helps the model preserve normal features while filtering out anomalous ones. The results in the third row", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22e7000012dce8eef883408ac061912895ee9b51e42ed37c700c83452e39baff.jpg", + "image_caption": [ + "Figure 5. Visualization of the impact of soft mining loss $\\mathcal{L}_{sm}$ . We plot the Kernel Density Estimation (KDE) for the chewing-gum and cashew categories in the VisA [47] dataset to estimate the probability density of the anomaly scores." + ], + "image_footnote": [], + "bbox": [ + 514, + 375, + 906, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicate that $\\mathcal{L}_c$ enhances the model's performance. This improvement stems from $\\mathcal{L}_c$ ensuring that the extracted INP coherently represents normal patterns, thereby avoiding the capture of anomalous ones and establishing a solid foundation for the subsequent suppression of anomalous feature reconstruction. Fig. 4 provides a more intuitive demonstration of the effectiveness of $\\mathcal{L}_c$ . The last row indicates that $\\mathcal{L}_{sm}$ boosts overall performance, as $\\mathcal{L}_{sm}$ directs the model's attention toward more challenging regions, thereby unlocking its optimal performance. Fig. 5 visually illustrates the impact of $\\mathcal{L}_{sm}$ , from which we can see that $\\mathcal{L}_{sm}$ contributes to a smaller overlap between anomaly score distributions of normal and abnormal pixels.", + "bbox": [ + 511, + 551, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Influence of the Number of INPs", + "text_level": 1, + "bbox": [ + 511, + 770, + 794, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 6, we conduct an ablation analysis on the number $M$ of INPs. The experimental results indicate that when $M$ exceeds four, the model's performance stabilizes. However, if $M$ becomes excessively large, the extracted INPs may also comprise information from abnormal tokens, leading to a slight decline in overall performance. In our study, we set $M$ to six.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9980", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13aceb1385fb3b9f411dd582bfdd2235a7cd319df49657c79fa68ca8f86cd493.jpg", + "image_caption": [ + "Figure 6. Influence of the number of INPs $M$ on model performance across the MVTec-AD [2] and VisA [47] datasets. Pixel-level AP and F1_max use the right vertical axis, while the other metrics share the left vertical axis." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 478, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32f97f3471c008bc4aefd2e8daa715f9bf00fd8aad342eb70d71a8d11f3da55a.jpg", + "image_caption": [ + "Figure 7. Visualizations of INPs. (a) Input anomalous image and ground truth. (b)-(g) Attention maps of six different INPs." + ], + "image_footnote": [], + "bbox": [ + 93, + 282, + 482, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Exploration on INPs", + "text_level": 1, + "bbox": [ + 89, + 518, + 285, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.1 Visualizations of INPs", + "text_level": 1, + "bbox": [ + 89, + 542, + 299, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 7, INPs effectively capture different semantic information. Specifically, the learned INPs focus on various regions of the image, including object areas (Fig. 7(b), (e) and (f)), object edges (Fig. 7(c) and (d)), and background areas (Fig. 7(g)). This diversity is attributed to our design of guiding the reconstruction process with INPs. Additionally, INP coherence loss ensures consistency in representing normal features, allowing INPs to concentrate solely on normal regions while ignoring anomalies. This mechanism ensures the decoder reconstructs features containing only normal patterns, thereby improving anomaly detection performance.", + "bbox": [ + 89, + 568, + 483, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2 Generalization capabilities of INP Extractor", + "text_level": 1, + "bbox": [ + 89, + 770, + 452, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 8, the INP Extractor trained on the RealIAD [37] dataset is capable of detecting INPs on the unseen MVTec-AD [2] dataset, and the distance maps to these INPs can serve for zero-shot anomaly detection. This effectively demonstrates the INP Extractor's ability to dynamically extract INPs from a single image, with the INP coherence loss $\\mathcal{L}_c$ ensuring that the extracted INPs coherently cap", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4aba017b2ce12e1416aaca3d32cc9a0a13ca6473f0728486e28f1483ab5f166.jpg", + "image_caption": [ + "Figure 8. Zero-shot anomaly detection results. Here INP-Former is trained on Real-IAD [37] and tested on MVtec-AD [2]. Distance maps are visualized." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 903, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ture normal patterns. As shown in Appendix Sec. K, without any specific training for zero-shot anomaly detection, our method can even outperform a specified method WinCLIP [21], achieving 88.0 and 88.7 pixel-level AUROCs on MVTec-AD and VisA, respectively.", + "bbox": [ + 511, + 236, + 903, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 335, + 633, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose INP-Former, a novel method for anomaly detection that explores the role of INPs. By learning to linearly combine normal tokens into INPs and using these INPs to guide the reconstruction of normal tokens, INP-Former significantly enhances anomaly detection performance. The introduction of the INP Coherence Loss and Soft Mining Loss further refines INP quality and optimizes the training process. Extensive experiments on MVTec-AD, VisA, and Real-IAD datasets demonstrate that INP-Former achieves SOTA or comparable performance across single-class, multi-class, and few-shot anomaly detection tasks. These results validate the existence and effectiveness of INPs, which can even be extracted from images in unseen categories, enabling zero-shot anomaly detection.", + "bbox": [ + 511, + 364, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & Future Works. Our method encounters certain limitations when detecting logical anomalies that closely resemble the background distribution, such as the misplaced anomalies in the Transistor class of the MVTec-AD dataset. This issue primarily arises because the misplaced anomaly in Transistor is highly similar to the background, causing INP Extractor to incorrectly extract this anomaly as INPs. A more detailed discussion can be found in Appendix Sec. M. In future work, we plan to combine the proposed INPs with pre-stored prototypes to address this limitation. While the pre-stored prototypes encapsulate comprehensive semantic information, the INPs exhibit strong alignment. This integration is expected to significantly improve the model's ability to detect logical anomalies that are similar to the background.", + "bbox": [ + 511, + 579, + 908, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledge", + "text_level": 1, + "bbox": [ + 513, + 830, + 633, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study was supported by the National Natural Science Foundation of China under Grant No. 52375494.", + "bbox": [ + 511, + 857, + 906, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9981", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5553-5561, 2023. 2", + "[2] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. The MVTec anomaly detection dataset: A comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision, 129(4):1038-1059, 2021. 1, 2, 5, 6, 7, 8", + "[3] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1", + "[4] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision. Springer, 2024. 3", + "[5] Yunkang Cao, Haiming Yao, Wei Luo, and Weiming Shen. Varad: Lightweight high-resolution image anomaly detection via visual autoregressive modeling. IEEE Transactions on Industrial Informatics, pages 1-10, 2025. 2", + "[6] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 6", + "[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 5", + "[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 6", + "[9] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 2, 5, 6, 7", + "[10] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3, 4", + "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3", + "[12] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 20 24. 3", + "[13] Hewei Guo, Liping Ren, Jingjing Fu, Yuwang Wang, Zhizheng Zhang, Cuiling Lan, Haoqian Wang, and Xinwen Hou. Template-guided hierarchical feature restoration for" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6447-6458, 2023. 1", + "[14] Jia Guo, Shuai Lu, Weihang Zhang, Fang Chen, Hongen Liao, and Huiqi Li. Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. arXiv preprint arXiv:2405.14325, 2024. 1, 3, 5, 6, 7", + "[15] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 6", + "[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 5, 6", + "[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3", + "[18] Chao Huang, Chengliang Liu, Zheng Zhang, Zhihao Wu, Jie Wen, Qiuping Jiang, and Yong Xu. Pixel-level anomaly detection via uncertainty-aware prototypical transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 521-530, 2022. 3", + "[19] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1", + "[20] Wenli Huang, Ye Deng, Siqi Hui, Yang Wu, Sanping Zhou, and Jinjun Wang. Sparse self-attention transformer for image inpainting. Pattern Recognition, 145:109897, 2024. 4", + "[21] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19606-19616, 2023. 3, 6, 8", + "[22] Yuxin Jiang, Yunkang Cao, and Weiming Shen. Prototypical learning guided context-aware segmentation network for few-shot anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2024. 1, 3", + "[23] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 3", + "[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16838-16848, 2024. 6", + "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In 2017" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9982", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5", + "[26] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 2, 5, 6, 7", + "[27] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36:8487-8500, 2023. 1, 3", + "[28] Wei Luo, Haiming Yao, and Wenyong Yu. Template-based feature aggregation network for industrial anomaly detection. Engineering Applications of Artificial Intelligence, 131: 107810, 2024. 1", + "[29] Wei Luo, Haiming Yao, Wenyong Yu, and Zhengyong Li. AMI-Net: Adaptive mask inpainting network for industrial anomaly detection and localization. IEEE Transactions on Automation Science and Engineering, 2024. 2", + "[30] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 3", + "[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4", + "[32] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 3", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3", + "[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 2, 3, 4, 6, 7", + "[35] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3", + "[36] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T.M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 2", + "[37] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detec" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22883-22892, 2024. 2, 5, 6, 7, 8", + "[38] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 5", + "[39] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Jiayi Lyu, Yong Liu, Chengjie Wang, Feng Zheng, and Yaochu Jin. IM-IAD: Industrial image anomaly detection benchmark in manufacturing. IEEE Transactions on Cybernetics, pages 1-14, 2024. 1", + "[40] Haiming Yao, Yunkang Cao, Wei Luo, Weihang Zhang, Wenyong Yu, and Weiming Shen. Prior normality prompt transformer for multiclass industrial image anomaly detection. IEEE Transactions on Industrial Informatics, 20(10): 11866-11876, 2024. 1", + "[41] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1", + "[42] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, pages 4571-4584, 2022. 1, 3, 4, 5, 6", + "[43] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 2, 5, 6", + "[44] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 2", + "[45] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 1, 2", + "[46] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 3", + "[47] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9983", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_model.json b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9431bd023555b5e449ada8cbc05848f6c2060924 --- /dev/null +++ b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.825, + 0.175 + ], + "angle": 0, + "content": "Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.203, + 0.819, + 0.294 + ], + "angle": 0, + "content": "Wei Luo\\(^{1*}\\) Yunkang Cao\\(^{2*}\\) Haiming Yao\\(^{1*}\\) Xiaotian Zhang\\(^{1}\\) Jianan Lou\\(^{1}\\) Yuqi Cheng\\(^{2}\\) Weiming Shen\\(^{2}\\) Wenyong Yu\\(^{2\\dagger}\\) Tsinghua University \\({}^{2}\\)Huazhong University of Science and Technology {luow23, yhm22, zxt19, ljn22}@mails.tsinghua.edu.cn {cyk_hust, yuqicheng, shenwm, ywy}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.795 + ], + "angle": 0, + "content": "Anomaly detection (AD) is essential for industrial inspection, yet existing methods typically rely on \"comparing\" test images to normal references from a training set. However, variations in appearance and positioning often complicate the alignment of these references with the test image, limiting detection accuracy. We observe that most anomalies manifest as local variations, meaning that even within anomalous images, valuable normal information remains. We argue that this information is useful and may be more aligned with the anomalies since both the anomalies and the normal information originate from the same image. Therefore, rather than relying on external normality from the training set, we propose INP-Former, a novel method that extracts Intrinsic Normal Prototypes (INPs) directly from the test image. Specifically, we introduce the INP Extractor, which linearly combines normal tokens to represent INPs. We further propose an INP Coherence Loss to ensure INPs can faithfully represent normality for the testing image. These INPs then guide the INP-Guided Decoder to reconstruct only normal tokens, with reconstruction errors serving as anomaly scores. Additionally, we propose a Soft Mining Loss to prioritize hard-to-optimize samples during training. INP-Former achieves state-of-the-art performance in single-class, multi-class, and few-shot AD tasks across MVtec-AD, VisA, and Real-IAD, positioning it as a versatile and universal solution for AD. Remarkably, INP-Former also demonstrates some zero-shot AD capability. Code is available at: https://github.com/luow23/INP-Former." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.812, + 0.222, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.838, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Unsupervised image anomaly detection (AD) [3, 39] seeks to identify abnormal patterns in images and localize anom" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.906, + 0.435 + ], + "angle": 0, + "content": "lous regions by learning solely from normal samples. This technique has seen widespread application in industrial defect detection [2, 47] and medical disease screening [19]. Recently, various specialized tasks have emerged in response to real-world demands, from conventional single-class AD [28, 34] to more advanced few-shot AD [17, 22] and multi-class AD [14, 40, 42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.908, + 0.71 + ], + "angle": 0, + "content": "Although the composition of normal samples varies across these tasks, the fundamental principle remains unchanged: modeling normality in the training data and assessing whether a test image aligns with this learned normality. However, this approach can be limited due to misaligned normality between the training data and the test image. For instance, prototype-based methods [34] extract representative normal prototypes to capture the normality of training samples. In few-shot AD, intra-class variance may lead to poorly aligned prototypes [17], e.g., hazelnuts in different appearances and positions, as shown in Fig. 1(a). Increasing the sample size can mitigate this problem but at the cost of additional prototypes and reduced inference efficiency. When there are multiple classes, i.e., multi-class AD, prototypes from one class may resemble anomalies from another, like the normal background of hazelnut is similar to the anomalies in cable in Fig. 1(b), leading to misclassification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.909, + 0.881 + ], + "angle": 0, + "content": "Several works have focused on extracting normality that is more aligned with the test image. For instance, some studies [13, 17, 45] propose spatially aligning normality within a single class through geometrical transformations. However, spatial alignment is ineffective for certain objects, such as hazelnuts, which exhibit variations beyond spatial positions. Other approaches [27, 40, 41] attempt to divide the normality in the training set into smaller, specific portions and then compare the test image to the corresponding portion of normality, but may still fail to find perfect alignment because of intra-class variances." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Rather than attempting to extract more aligned normal-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.9, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.218, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Figure 1. Motivation for Intrinsic Normal Prototypes (INPs). (a) Pre-stored prototypes from few-shot normal samples may fail to represent all normal patterns. (b) Pre-stored prototypes from one class can be similar to anomalies in another class. (c) The extracted INPs are concise yet well-aligned to the test image, alleviating the issues in (a) and (b)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.276, + 0.482, + 0.472 + ], + "angle": 0, + "content": "ity from the training set, we propose addressing the issue of misaligned normality by leveraging the normality within the test image itself as prototypes, which we term Intrinsic Normal Prototypes (INPs). As illustrated in Fig. 1(c), normal patches within an anomalous test image can function as INPs, and anomalies can be easily detected by comparing them with these INPs. These INPs provide more concise and well-aligned prototypes to the anomalies than those learned from training data, as they typically share the same geometrical context and similar appearances with the abnormal regions within the testing image itself. Accordingly, we explore the prevalence of INPs in various AD scenarios and evaluate their potential to improve AD performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Although previous work [1] has attempted to utilize INP for anomaly detection, it employs handcrafted aggregated features as prototypes, thus limiting the method to zero-shot texture anomaly detection. In contrast, we introduce a learnable INP Extractor to extract normal features with adaptable shapes as INPs. We also propose an INP Coherence Loss to ensure that the extracted INPs coherently represent the normality within the test image, avoiding the capture of anomalous regions. However, some weakly representative normal regions are challenging to model with a limited set of discrete INPs, resulting in background noise (Fig. 4(c)). To address this issue, we introduce an INP-Guided Decoder, which integrates INPs into a reconstruction-based framework. This decoder leverages combinations of discrete INPs to accurately reconstruct all normal regions while effectively suppressing the reconstruction of anomalous regions, with reconstruction errors serving as anomaly scores. Furthermore, we introduce a Soft Mining Loss that focuses on normal regions that are challenging to reconstruct, i.e., hard samples, thereby improving overall reconstruction quality and enhancing AD performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our approach, termed INP-Former, primarily leverages vision transformers (ViTs) for both INP extraction and INP-guided reconstruction. Extensive experiments on MVtec-AD [2], VisA [47], and Real-IAD [37] demonstrate that INP-Former achieves superior performance across multiclass, single-class, and few-shot AD tasks, positioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.395 + ], + "angle": 0, + "content": "INP-Former as a universal AD solution. INP-Former also optimizes computational complexity by extracting concise INPs, e.g., images can be represented effectively using only six INPs, as shown in Sec. 4.3.2. Additionally, as demonstrated in Sec. 4.4.2, INP-Former exhibits strong generalization and can even extract INPs for unseen classes, enabling zero-shot AD capabilities. In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.397, + 0.905, + 0.442 + ], + "angle": 0, + "content": "- We demonstrate that a single image can contain Intrinsic Normal Prototypes (INPs), offering concise and aligned normality for anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.443, + 0.905, + 0.486 + ], + "angle": 0, + "content": "- We propose the INP Extractor and incorporate INPs into a reconstruction-based anomaly detection framework using the INP-Guided Decoder." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.488, + 0.905, + 0.533 + ], + "angle": 0, + "content": "- We introduce the INP Coherence Loss to extract representative INPs and the Soft Mining Loss to enhance reconstruction quality." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.397, + 0.905, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.549, + 0.663, + 0.563 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.776, + 0.59 + ], + "angle": 0, + "content": "2.1. Universal Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.658 + ], + "angle": 0, + "content": "There are numerous unsupervised AD tasks, ranging from conventional single-class AD to recent few-shot and multiclass AD setups. We refer to these collectively as universal anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Single-Class Anomaly Detection: This setup was originally introduced by MVTec-AD [2] and involves developing distinct AD models for each class. Typically, images are embedded into a feature space using a pre-trained encoder, after which various schemes, such as reconstruction-based [5, 29, 44], knowledge-distillation-based [9, 36], prototype-based [34, 45], and embedding-based [26, 43] methods, are employed to learn the normality of the given class. While these approaches achieve strong performance, their reliance on class-specific models limits scalability when dealing with a wide range of classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Few-shot Anomaly Detection: In practical scenarios, the number of available normal samples may also be limited, motivating the development of few-shot AD methods. In this case, normal samples may not fully capture the variability of normality. To address this challenge, approaches" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9975" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.881, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.361, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Figure 2. Overview of our INP-Former framework for universal anomaly detection. (a) Our model consists of a pre-trained Encoder, an INP Extractor, a Bottleneck, and an INP-Guided decoder. The INP Extractor dynamically extracts intrinsic normal prototypes from a single image, which the INP-Guided Decoder leverages to effectively suppress anomalous features. (b) Detailed architecture of the INP Extractor. (c) Detailed architecture of each layer in the INP-Guided Decoder. (d) Comparison of computational complexity between INP-Guided Attention and Self Attention. It is important to note that the patch token (Encoding) and patch token (Decoding) refer to the patch tokens utilized during the encoding and decoding stages, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.482, + 0.612 + ], + "angle": 0, + "content": "such as spatial alignment [17] or contrastive learning [22] are used to create more compact and representative normal embeddings. Recently, Vision-Language Models (VLMs) like CLIP [33] have proven effective for few-shot AD due to their broad, pre-trained knowledge. These VLMs not only provide descriptive visual embeddings but also compute the similarity between text prompts and test images, as seen in works like WinCLIP [21], AnomalyGPT [12], and InCTRL [46]. Some approaches, such as AdaCLIP [4], even enable zero-shot AD through VLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Multi-Class Anomaly Detection: Developing separate models for each class can be resource-intensive, prompting interest in multi-class AD, also known as unified AD [42], which aims to build a single model for multiple classes. UniAD [42] pioneered a unified reconstruction framework for anomaly detection, followed by HVQ-Trans [27] addressed the identical shortcut problem using a vector quantization framework. More recent approaches, such as MambaAD [15] and Dinomaly [14], further enhance multi-class AD performance by leveraging advanced models, i.e., the State Space Model Mamba [11] and DINO [7], respectively. However, these methods lack the functionality to derive aligned normality with the test image. On the contrary, we extract INPs from the testing image, bringing aligned and precise normality for anomaly detection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.279, + 0.865 + ], + "angle": 0, + "content": "2.2. Prototype Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Prototype learning [35] aims to extract representative prototypes from a given training set, which are then used for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.718 + ], + "angle": 0, + "content": "classification by measuring their distances to a test sample in a metric space. This technique is widely used in few-shot learning [23]. Several AD methods also employ prototype learning. For example, PatchCore [34] extracts multiple normal prototypes to represent the normality of the training data, directly computing the minimal distances to the test sample for anomaly detection. Other approaches [10, 18, 30, 32] incorporate prototypes into the reconstruction process to avoid the identical shortcut issue. Specifically, they replace the original inputs with combinations of learned normal prototypes, ensuring that the inputs to the reconstruction model contain only normal elements. However, these methods rely on pre-stored normal prototypes extracted from the training set, which can suffer from the misaligned normality problem. In contrast, our INPs are dynamically extracted from the test image, providing more aligned alternatives for normality representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.732, + 0.72, + 0.747 + ], + "angle": 0, + "content": "3. Method: INP-Former" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.625, + 0.771 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To fully exploit the advantages of INPs in anomaly detection, we propose INP-Former, as depicted in Fig. 2(a). The model dynamically extracts INPs from a single image and utilizes them to guide the feature reconstruction process, with the reconstruction errors serving as anomaly scores. Specifically, it comprises four key modules: a fixed pre-trained Encoder \\(\\mathcal{Q}\\), an INP Extractor \\(\\mathcal{E}\\), a Bottleneck \\(\\mathcal{B}\\), and an INP-Guided Decoder \\(\\mathcal{D}\\). The input im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.304 + ], + "angle": 0, + "content": "age \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is first processed by the pre-trained Encoder \\(\\mathcal{Q}\\) to extract multi-scale latent features \\(f_{\\mathcal{Q}} = \\{f_{\\mathcal{Q}}^{1}, \\ldots, f_{\\mathcal{Q}}^{L} | f_{\\mathcal{Q}}^{l} \\in \\mathbb{R}^{N \\times C}, N = \\frac{HW}{k^{2}}\\}\\), where \\(k\\) represents the downsampling factor. Next, the INP Extractor \\(\\mathcal{E}\\) extracts \\(M\\) INPs \\(\\mathbf{P} = \\{p_{1}, \\ldots, p_{M} | p_{m} \\in \\mathbb{R}^{C}\\}\\) from the pre-trained features, with an INP coherence loss ensuring that the extracted INPs consistently represent normal features during testing. The Bottleneck \\(\\mathcal{B}\\) subsequently fuses the multi-scale latent features, producing the fused output \\(F_{\\mathcal{B}} = \\mathcal{B}(f_{\\mathcal{Q}})\\). Following the bottleneck, the extracted INPs are utilized to guide the Decoder \\(\\mathcal{D}\\) to yield reconstruction outputs \\(f_{\\mathcal{D}} = \\{f_{\\mathcal{D}}^{1}, \\ldots, f_{\\mathcal{D}}^{L} | f_{\\mathcal{D}}^{l} \\in \\mathbb{R}^{N \\times C}\\}\\) with only normal patterns, thus the reconstruction error between \\(f_{\\mathcal{Q}}\\) and \\(f_{\\mathcal{D}}\\) can serve as the anomaly score." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.312, + 0.24, + 0.327 + ], + "angle": 0, + "content": "3.2. INP Extractor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Existing prototype-based methods [10, 31, 34] store local normal features from the training data and compare them with test images. However, the misaligned normality between these pre-stored prototypes and the test images and the lack of global information lead to suboptimal detection performance. To address these limitations, we propose the INP Extractor to dynamically extract INPs with global information from the test image itself." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.58 + ], + "angle": 0, + "content": "Specifically, as illustrated in Fig. 2(b), instead of extracting representative local features as done in PatchCore [34], we employ cross attention to aggregate the global semantic information of the pre-trained features \\(\\mathbf{F}_{\\mathcal{Q}} \\in \\mathbb{R}^{N \\times C}\\) with \\(M\\) learnable tokens \\(\\mathbf{T} = \\{t_1, \\dots, t_M | t_m \\in \\mathbb{R}^C\\}\\). Here \\(\\mathbf{F}_{\\mathcal{Q}}\\) is used as the key-value pairs, while \\(\\mathbf{T}\\) serve as the query, allowing \\(\\mathbf{T}\\) to linearly aggregate \\(\\mathbf{F}_{\\mathcal{Q}}\\) into INPs \\(\\mathbf{P} = \\{p_1, \\dots, p_M | p_m \\in \\mathbb{R}^C\\}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.588, + 0.332, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathcal {Q}} = \\operatorname {s u m} \\left(\\left\\{f _ {\\mathcal {Q}} ^ {1}, \\dots , f _ {\\mathcal {Q}} ^ {L} \\right\\}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.608, + 0.482, + 0.631 + ], + "angle": 0, + "content": "\\[\nQ = \\mathbf {T} W ^ {Q}, K = \\mathbf {F} _ {\\mathcal {Q}} W ^ {K}, V = \\mathbf {F} _ {\\mathcal {Q}} W ^ {V} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.628, + 0.362, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} ^ {\\prime} = \\operatorname {A t t e n t i o n} (Q, K, V) + \\mathbf {T}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.648, + 0.291, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\mathrm {F F N} (\\mathbf {T} ^ {\\prime}) + \\mathbf {T} ^ {\\prime}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.748 + ], + "angle": 0, + "content": "where \\(\\mathrm{sum}(\\cdot)\\) denotes the element-wise summation. \\(Q\\in\\) \\(\\mathbb{R}^{M\\times C}\\) and \\(K,V\\in \\mathbb{R}^{N\\times C}\\) represent the query, key and value, respectively. \\(W^{Q},W^{k},W^{v}\\in \\mathbb{R}^{C\\times C}\\) are the learnable projection parameters for \\(Q,K,V;\\mathrm{FFN}(\\cdot)\\) represents the feed-forward network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.826 + ], + "angle": 0, + "content": "To ensure that INPs coherently represent normal features while minimizing the capture of anomalous features during the testing process, we propose an INP coherence loss \\(\\mathcal{L}_c\\) to minimize the distances between individual normal features and the corresponding nearest INP." + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.837, + 0.482, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d _ {i} = \\min _ {m \\in \\{1, \\dots , M \\}} \\mathcal {S} (\\mathbf {F} _ {\\mathcal {Q}} (i), p _ {m}) \\\\ \\mathcal {L} _ {c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} d _ {i} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.09, + 0.904, + 0.105 + ], + "angle": 0, + "content": "Table 1. Comparison of computational cost and memory usage." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.109, + 0.895, + 0.222 + ], + "angle": 0, + "content": "
Number of multiplicaiton and addition
CalculationVanilla Self AttentionINP-Guided Attention
Al=Ql(Kl)T943 496 9607 220 640
fdl-1'=AlVl943 509 5046 623 232
Memory usage (MB)
Ql/Kl/Vl/Al2.30/2.30/2.30/2.342.30/0.018/0.018/0.018
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.905, + 0.293 + ], + "angle": 0, + "content": "where \\(S(\\cdot, \\cdot)\\) denotes the cosine distance. \\(d_{i}\\) represents the distance between the query feature \\(\\mathbf{F}_{\\mathcal{Q}}(i)\\) and the corresponding nearest INP item. Fig. 4 visually illustrates the effectiveness of \\(\\mathcal{L}_c\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.3, + 0.714, + 0.315 + ], + "angle": 0, + "content": "3.3. INP-Guided Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.684 + ], + "angle": 0, + "content": "While we can use the distance between testing features and their nearest INPs for anomaly detection, as illustrated in Fig. 4(c), certain low-representative normal regions are difficult to model with a limited number of discrete INPs, leading to noisy distance maps between these INPs and testing features. To address this issue, we propose the INP-Guided Decoder, aiming to reconstruct these low-representative normal regions through a combination of multiple discrete INPs and suppress the reconstruction of anomalous regions. Additionally, this decoder provides a token-wise discrepancy that can be directly leveraged for anomaly detection. As shown in Fig. 2(c), INPs are incorporated into this decoder to guide the reconstruction process. Since INPs exclusively represent normal patterns in test images, we employ the extracted INPs as key-value pairs, ensuring that the output is a linear combination of normal INPs, thereby effectively suppressing the reconstruction of anomalous queries, i.e., the identical mapping issue [42]. Furthermore, we find that the first residual connection can directly introduce anomalous features into the subsequent reconstruction, so we remove this connection in our INP-Guided Decoder. Following the previous work [20], we also employ the ReLU activation function to mitigate the influence of weak correlations and noise on the attention maps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.684, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Mathematically, let \\( f_{\\mathcal{D}}^{l - 1} \\in \\mathbb{R}^{N \\times C} \\) denotes the output latent features from previous decoding layer. The output \\( f_{\\mathcal{D}}^{l} \\in \\mathbb{R}^{N \\times C} \\) of the \\( l_{th} \\) decoding layer is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.739, + 0.848, + 0.756 + ], + "angle": 0, + "content": "\\[\nQ _ {l} = f _ {\\mathcal {D}} ^ {l - 1} W _ {l} ^ {Q}, K _ {l} = \\mathbf {P} W _ {l} ^ {K}, V _ {l} = \\mathbf {P} W _ {l} ^ {V}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.761, + 0.904, + 0.78 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}} = A _ {l} V _ {l}, A _ {l} = \\operatorname {R e L U} \\left(Q _ {l} \\left(K _ {l}\\right) ^ {T}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.783, + 0.752, + 0.802 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {D}} ^ {l} = \\operatorname {F F N} \\left(f _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}}\\right) + f _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.809, + 0.905, + 0.871 + ], + "angle": 0, + "content": "where \\(Q_{l}\\in \\mathbb{R}^{N\\times C}\\) and \\(K_{l},V_{l}\\in \\mathbb{R}^{M\\times C}\\) denote the query, key and value of the \\(l_{th}\\) decoding layer. \\(W_{l}^{Q},W_{l}^{k},W_{l}^{v}\\in\\) \\(\\mathbb{R}^{C\\times C}\\) denote the learnable projection parameters for \\(Q_{l},K_{l},V_{l}\\). \\(A_{l}\\in \\mathbb{R}^{N\\times M}\\) represent the attention map." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Attention Complexity Analysis: As depicted in Fig. 2(d), the computational complexity of vanilla self-attention is" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9977" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.244 + ], + "angle": 0, + "content": "\\(\\mathcal{O}(N^2 C)\\), while its memory usage is \\(\\mathcal{O}(N^2)\\). In contrast, our INP-Guided Attention reduces both the computational complexity and memory usage to \\(\\mathcal{O}(NMC)\\) and \\(\\mathcal{O}(NM)\\), respectively, which can be approximated as \\(\\mathcal{O}(NC)\\) and \\(\\mathcal{O}(N)\\) due to \\(M \\ll N\\). Tab. 1 offers a detailed comparison of the complexity of vanilla self-attention and INP-Guided Attention. Appendix Sec. C compares the overall complexity between INP-Former and other methods. The light version of INP-Former can even be more efficient than MambaAD [15] yet demonstrate better performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.252, + 0.262, + 0.269 + ], + "angle": 0, + "content": "3.4. Soft Mining Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.484, + 0.336 + ], + "angle": 0, + "content": "Inspired by Focal Loss [25], different regions should be assigned varying weights based on their optimization difficulty. Accordingly, we propose Soft Mining Loss to encourage the model to focus more on difficult regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.427 + ], + "angle": 0, + "content": "Intuitively, the ratio of the reconstruction error of an individual normal region to the average reconstruction error of all normal regions can serve as an indicator of optimization difficulty. Given the encoder \\( f_{Q}^{l} \\) and decoder \\( f_{D}^{l} \\) features at layer \\( l \\), let \\( M^{l} \\) denote the regional cosine distance. Our soft mining loss \\( \\mathcal{L}_{sm} \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.437, + 0.326, + 0.473 + ], + "angle": 0, + "content": "\\[\nw ^ {l} (h, w) = \\left[ \\frac {M ^ {l} (h , w)}{u (M ^ {l})} \\right] ^ {\\gamma}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.476, + 0.483, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s m} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} 1 - \\frac {\\operatorname {v e c} \\left(f _ {\\mathcal {Q}} ^ {l}\\right) ^ {T} \\cdot \\operatorname {v e c} \\left(\\hat {f} _ {\\mathcal {D}} ^ {l}\\right)}{\\left| \\left| \\operatorname {v e c} \\left(f _ {\\mathcal {Q}} ^ {l}\\right) \\right| \\right| \\left| \\left| \\operatorname {v e c} \\left(\\hat {f} _ {\\mathcal {D}} ^ {l}\\right) \\right| \\right|} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.519, + 0.37, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {\\mathcal {D}} ^ {l} (h, w) = c g \\left(f _ {\\mathcal {D}} ^ {l} (h, w)\\right) _ {w ^ {l} (h, w)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.645 + ], + "angle": 0, + "content": "where \\( u(M^l) \\) represents the average regional cosine distance within a batch, \\( \\gamma \\geq 0 \\) denotes the temperature hyperparameter, \\( cg(\\cdot)_{w^{l}(h,w)} \\) denotes a gradient adjustment based on dynamic weight \\( w^{l}(h,w) \\), and \\( vec(\\cdot) \\) denotes the flattening operation. The overall training loss of our INP-Former can be expressed as follows: \\( \\mathcal{L}_{total} = \\mathcal{L}_{sm} + \\lambda \\mathcal{L}_c \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.657, + 0.224, + 0.674 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.299, + 0.698 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Datasets: We conduct a comprehensive analysis of the proposed INP-Former on three widely used AD datasets: MVTec-AD [2], VisA [47], and Real-IAD [37]. MVTec-AD consists of 15 categories, with 3,629 normal images for training, and 1,982 anomalous images along with 498 normal images for testing. VisA contains 12 object categories, with 8,659 normal images for training and 962 normal images along with 1,200 anomalous images for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Real-IAD includes 30 different objects, with 36,645 normal images for training and 63,256 normal images along with 51,329 anomalous images for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Metrics: Following existing works [14, 15], we use the Area Under the Receiver Operating Characteristic Curve" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "(AUROC), Average Precision (AP), and F1-score-max (F1_max) to evaluate anomaly detection and localization. For anomaly localization specifically, we use Area Under the Per-Region-Overlap (AUPRO) as an additional metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Implementation Details: INP-Former adopts ViT-Base/14 with DINO2-R [7] weights as the default pre-trained encoder. The number \\(M\\) of INPs is set to six by default. The INP Extractor includes a standard Vision Transformer block. The layer number of the INP-Guided decoder is eight. All input images are resized to \\(448^2\\) and then center-cropped to \\(392^2\\). The hyperparameters \\(\\gamma\\) and \\(\\lambda\\) are set to 3.0 and 0.2, respectively. We use the StableAdamW [38] optimizer with a learning rate \\(1e^{-3}\\) and a weight decay of \\(1e^{-4}\\) for 200 epochs. Notably, the above hyperparameters do not require any adjustment across the three datasets. Appendix Sec. A presents more implementation details. Appendix Sec. D, E, and F analyze the influence of input resolution, ViT architecture, and \\(\\lambda\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.654, + 0.387 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.788, + 0.41 + ], + "angle": 0, + "content": "4.2.1 Multi-Class Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.524 + ], + "angle": 0, + "content": "We compare the proposed INP-Former with several state-of-the-art (SOTA) methods for multi-class anomaly detection, including reconstruction-based methods RD4AD [9], UniAD [42], DiAD [16], MambaAD [15], and Dinomaly [14], and embedding-based methods SimpleNet [26] and DeSTSeg [43]. A detailed introduction to the comparison methods can be found in Appendix Sec. B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The experimental results on the three AD datasets are presented in Tab. 2. On the widely used MVTec-AD dataset, our method achieves SOTA performance, with image-level metrics of 99.7/99.9/99.2 and pixel-level metrics of 98.5/71.0/69.7/94.9. On VisA, our method achieves competitive results, attaining the best image-level metrics of 98.9/99.0/96.6, and achieving the best or second-best pixel-level performance of 98.8/51.2/54.7/94.4. On the more complex and challenging Real-IAD dataset, our method reaches new SOTA performance, with image-level metrics of 90.5/88.1/81.5 and pixel-level metrics of 99.0/47.5/50.3/95.0. Compared to the second-best results, our method improves by \\(1.2\\uparrow /1.3\\uparrow /1.3\\uparrow\\) at the image level and by \\(0.2\\uparrow /4.7\\uparrow /3.2\\uparrow /1.1\\uparrow\\) at the pixel level. The SOTA performance achieved across the three datasets showcases the effectiveness and robustness of our method. The per-class performance metrics are presented in Appendix Sec. H. Appendix Sec. G presents the performance of INP-Former in a more challenging scenario, which we call supermulti-class anomaly detection, thus training INP-Former on several datasets - MVTec-AD, VisA, and Real-IAD - simultaneously. Results show that our method can even detect anomalies in more classes without significant performance degradation. Fig. 3 demonstrates the precise anomaly localization capability of our method. More qualitative results" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9978" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.09, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Table 2. Multi-class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.11, + 0.885, + 0.304 + ], + "angle": 0, + "content": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Metric →Image-level(I-AUROC/I-AP/I-F1_max)Pixel-level(P-AUROC/P-AP/P-F1_max/AUPRO)
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
RD4AD [9]94.6/96.5/95.296.1/48.6/53.8/91.192.4/92.4/89.698.1/38.0/42.6/91.882.4/79.0/73.997.3/25.0/32.7/89.6
UniAD [42]96.5/98.8/96.296.8/43.4/49.5/90.788.8/90.8/85.898.3/33.7/39.0/85.583.0/80.9/74.397.3/21.1/29.2/86.7
SimpleNet [26]95.3/98.4/95.896.9/45.9/49.7/86.587.2/87.0/81.896.8/34.7/37.8/81.457.2/53.4/61.575.7/2.8/6.5/39.0
DeSTSeg [43]89.2/95.5/91.693.1/54.3/50.9/64.888.9/89.0/85.296.1/39.6/43.4/67.482.3/79.2/73.294.6/37.9/41.7/40.6
DiAD [16]97.2/99.0/96.596.8/52.6/55.5/90.786.8/88.3/85.196.0/26.1/33.0/75.275.6/66.4/69.988.0/2.9/7.1/58.1
MambaAD [15]98.6/99.6/97.897.7/56.3/59.2/93.194.3/94.5/89.498.5/39.4/44.0/91.086.3/84.6/77.098.5/33.0/38.7/90.5
Dinomaly [14]99.6/99.8/99.098.4/69.3/69.2/94.898.7/98.9/96.298.7/53.2/55.7/94.589.3/86.8/80.298.8/42.8/47.1/93.9
INP-Former99.7/99.9/99.298.5/71.0/69.7/94.998.9/99.0/96.698.9/51.2/54.7/94.490.5/88.1/81.599.0/47.5/50.3/95.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.905, + 0.336 + ], + "angle": 0, + "content": "Table 3. Few-shot (4-shot) anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined. \\(\\dagger\\) indicates the results we reproduced using publicly available code." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.34, + 0.885, + 0.478 + ], + "angle": 0, + "content": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
SPADE [6]84.8/92.5/91.592.7/-/46.2/87.081.7/83.4/82.196.6/-/43.6/87.350.8†/45.8†/61.2†59.5†/0.2†/0.5†/19.2†
PaDiM [8]80.4/90.5/90.292.6/-/46.1/81.372.8/75.6/78.093.2/-/24.6/72.660.3†/53.5†/64.0†90.9†/2.1†/5.1†/67.6†
PatchCore [34]88.8/94.5/92.694.3/-/55.0/84.385.3/87.5/84.396.8/-/43.9/84.966.0†/62.2†/65.2†92.9†/9.8†/16.1†/68.6†
WinCLIP [21]95.2/97.3/94.796.2/-/59.5/89.087.3/88.8/84.297.2/-/47.0/87.673.0†/61.8†/61.0†93.8†/13.3†/21.0†/76.4†
PromptAD [24]96.6/-/-96.5/-/-/90.589.1/-/-97.4/-/-/86.259.7†/43.5†/52.9†86.9†/8.7†/16.1†/61.9†
INP-Former97.6/98.6/97.097.0/65.9/65.6/92.996.4/96.0/93.097.7/49.3/54.3/93.176.7/72.3/71.797.3/32.2/36.7/89.0
" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.48, + 0.884, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.611, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of anomaly localization on the MVTec-AD [2], VisA [47], and Real-IAD [37] datasets for multi-class anomaly detection. The first row presents the input images with their ground truth, while the second row displays the corresponding anomaly maps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.669, + 0.318, + 0.685 + ], + "angle": 0, + "content": "are presented in Appendix Sec. L." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.708, + 0.35, + 0.723 + ], + "angle": 0, + "content": "4.2.2 Few-Shot Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We compare our method with several SOTA approaches for few-shot anomaly detection, including prototype-based methods SPADE [6], PaDiM [8], and PatchCore [34] and recent advances that utilize VLMs, i.e., WinCLIP [21] and PromptAD [24]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 3, our method significantly outperforms previous SOTAs on three different AD datasets. Compared to the second-best results, our method achieves improvements of \\(1.0\\uparrow /1.3\\uparrow /2.3\\uparrow\\) in image-level scores and \\(0.5\\uparrow / - / 6.1\\uparrow /2.4\\uparrow\\) in pixel-level scores on the MVTec-AD dataset. It similarly outperforms the second-best results on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.907, + 0.821 + ], + "angle": 0, + "content": "the VisA dataset, with enhancements of \\(7.3 \\uparrow / 7.2 \\uparrow / 8.7 \\uparrow\\) for image-level and \\(0.3 \\uparrow / -7.3 \\uparrow / 5.5 \\uparrow\\) for pixel-level scores. Additionally, on the Real-IAD dataset, our method surpasses the second-best results by \\(3.7 \\uparrow / 10.1 \\uparrow / 6.5 \\uparrow\\) in image-level and \\(3.5 \\uparrow / 18.9 \\uparrow / 15.7 \\uparrow / 12.6 \\uparrow\\) in pixel-level scores. The superior performance of our method in few-shot anomaly detection stems from its ability to extract INPs from a single image, eliminating the need for extensive normal data to pre-store prototypes. More comparison results on 1-shot and 2-shot are presented in Appendix Sec. I." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.791, + 0.846 + ], + "angle": 0, + "content": "4.2.3 Single-Class Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We further compared our proposed INP-Former with current SOTA methods for single-class anomaly detection, as shown in Tab. 4. The results indicate that INP-Former" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9979" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.089, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Table 4. Single class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined." + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.109, + 0.822, + 0.226 + ], + "angle": 0, + "content": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓I-AUROCP-APAUPROI-AUROCP-APAUPROI-AUROCP-APAUPRO
PatchCore [34]99.156.193.595.140.191.289.4-91.5
RD4AD [9]98.558.093.996.027.770.987.1-93.8
SimpleNet [26]99.654.890.096.836.388.788.5-84.6
Dinomaly [14]99.768.995.098.950.795.192.045.295.1
INP-Former99.770.295.498.549.293.892.148.195.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.229, + 0.904, + 0.243 + ], + "angle": 0, + "content": "Table 5. Overall ablation on MVTec-AD [2] and VisA [47] datasets. \"INP\" refers to the use of INP Extractor and INP-Guided Decoder." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.249, + 0.866, + 0.364 + ], + "angle": 0, + "content": "
Dataset →MVTec-AD [2]VisA [47]
“INP”LcLsmImage-levelPixel-levelImage-levelPixel-level
Module×××98.59/99.18/97.6397.19/61.73/62.94/92.7396.58/97.18/92.8997.50/47.24/51.90/82.85
✓××99.53/99.80/98.8198.32/69.82/69.38/94.6998.11/98.23/95.2298.41/50.34/54.23/93.63
✓✓×99.61/99.83/99.0298.39/70.01/69.53/95.1098.16/98.30/95.4798.46/51.09/54.46/93.71
✓✓✓99.67/99.88/99.2098.48/71.02/69.65/94.8798.90/99.02/96.5798.90/51.22/54.74/94.36
" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.375, + 0.482, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.586 + ], + "angle": 0, + "content": "Figure 4. Visualization of the impact of INP coherence loss \\(\\mathcal{L}_c\\). (a) Input anomalous image and ground truth. (b) Distance map without \\(\\mathcal{L}_c\\). (c) Distance map with \\(\\mathcal{L}_c\\). The distance map is obtained by calculating the distance between the input features and their nearest INP terms, as described in Eq. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.661 + ], + "angle": 0, + "content": "achieves new SOTA performance on the MVTec-AD and Real-IAD datasets and demonstrates competitive performance on the VisA dataset. Per-category performance of INP-Former is presented in Appendix Sec. J." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.676, + 0.245, + 0.691 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.263, + 0.715 + ], + "angle": 0, + "content": "4.3.1 Overall Ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 5, we conduct comprehensive experiments on MVTec-AD [2] and VisA [47] to validate the effectiveness of the proposed components, i.e., INP Extractor and INP-Guided Decoder (\"INP\"), INP Coherence Loss \\((\\mathcal{L}_c)\\), and Soft Mining Loss \\((\\mathcal{L}_{sm})\\). In the first row, we train a baseline model without incorporating any proposed module, similar to the RD4AD [9] framework. The results in the second row demonstrate that \"INP\" significantly enhances overall performance. This improvement arises from the fact that \"INP\" introduces an information bottleneck, which effectively helps the model preserve normal features while filtering out anomalous ones. The results in the third row" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.375, + 0.907, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.48, + 0.907, + 0.536 + ], + "angle": 0, + "content": "Figure 5. Visualization of the impact of soft mining loss \\(\\mathcal{L}_{sm}\\). We plot the Kernel Density Estimation (KDE) for the chewing-gum and cashew categories in the VisA [47] dataset to estimate the probability density of the anomaly scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.75 + ], + "angle": 0, + "content": "indicate that \\(\\mathcal{L}_c\\) enhances the model's performance. This improvement stems from \\(\\mathcal{L}_c\\) ensuring that the extracted INP coherently represents normal patterns, thereby avoiding the capture of anomalous ones and establishing a solid foundation for the subsequent suppression of anomalous feature reconstruction. Fig. 4 provides a more intuitive demonstration of the effectiveness of \\(\\mathcal{L}_c\\). The last row indicates that \\(\\mathcal{L}_{sm}\\) boosts overall performance, as \\(\\mathcal{L}_{sm}\\) directs the model's attention toward more challenging regions, thereby unlocking its optimal performance. Fig. 5 visually illustrates the impact of \\(\\mathcal{L}_{sm}\\), from which we can see that \\(\\mathcal{L}_{sm}\\) contributes to a smaller overlap between anomaly score distributions of normal and abnormal pixels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.771, + 0.795, + 0.785 + ], + "angle": 0, + "content": "4.3.2 Influence of the Number of INPs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 6, we conduct an ablation analysis on the number \\(M\\) of INPs. The experimental results indicate that when \\(M\\) exceeds four, the model's performance stabilizes. However, if \\(M\\) becomes excessively large, the extracted INPs may also comprise information from abnormal tokens, leading to a slight decline in overall performance. In our study, we set \\(M\\) to six." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9980" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.48, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.223, + 0.483, + 0.277 + ], + "angle": 0, + "content": "Figure 6. Influence of the number of INPs \\(M\\) on model performance across the MVTec-AD [2] and VisA [47] datasets. Pixel-level AP and F1_max use the right vertical axis, while the other metrics share the left vertical axis." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.283, + 0.483, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Figure 7. Visualizations of INPs. (a) Input anomalous image and ground truth. (b)-(g) Attention maps of six different INPs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.286, + 0.536 + ], + "angle": 0, + "content": "4.4. Exploration on INPs" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.3, + 0.558 + ], + "angle": 0, + "content": "4.4.1 Visualizations of INPs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.75 + ], + "angle": 0, + "content": "As shown in Fig. 7, INPs effectively capture different semantic information. Specifically, the learned INPs focus on various regions of the image, including object areas (Fig. 7(b), (e) and (f)), object edges (Fig. 7(c) and (d)), and background areas (Fig. 7(g)). This diversity is attributed to our design of guiding the reconstruction process with INPs. Additionally, INP coherence loss ensures consistency in representing normal features, allowing INPs to concentrate solely on normal regions while ignoring anomalies. This mechanism ensures the decoder reconstructs features containing only normal patterns, thereby improving anomaly detection performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.453, + 0.787 + ], + "angle": 0, + "content": "4.4.2 Generalization capabilities of INP Extractor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 8, the INP Extractor trained on the RealIAD [37] dataset is capable of detecting INPs on the unseen MVTec-AD [2] dataset, and the distance maps to these INPs can serve for zero-shot anomaly detection. This effectively demonstrates the INP Extractor's ability to dynamically extract INPs from a single image, with the INP coherence loss \\(\\mathcal{L}_c\\) ensuring that the extracted INPs coherently cap" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.09, + 0.905, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.179, + 0.905, + 0.221 + ], + "angle": 0, + "content": "Figure 8. Zero-shot anomaly detection results. Here INP-Former is trained on Real-IAD [37] and tested on MVtec-AD [2]. Distance maps are visualized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.905, + 0.313 + ], + "angle": 0, + "content": "ture normal patterns. As shown in Appendix Sec. K, without any specific training for zero-shot anomaly detection, our method can even outperform a specified method WinCLIP [21], achieving 88.0 and 88.7 pixel-level AUROCs on MVTec-AD and VisA, respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.337, + 0.634, + 0.352 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.577 + ], + "angle": 0, + "content": "We propose INP-Former, a novel method for anomaly detection that explores the role of INPs. By learning to linearly combine normal tokens into INPs and using these INPs to guide the reconstruction of normal tokens, INP-Former significantly enhances anomaly detection performance. The introduction of the INP Coherence Loss and Soft Mining Loss further refines INP quality and optimizes the training process. Extensive experiments on MVTec-AD, VisA, and Real-IAD datasets demonstrate that INP-Former achieves SOTA or comparable performance across single-class, multi-class, and few-shot anomaly detection tasks. These results validate the existence and effectiveness of INPs, which can even be extracted from images in unseen categories, enabling zero-shot anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Limitations & Future Works. Our method encounters certain limitations when detecting logical anomalies that closely resemble the background distribution, such as the misplaced anomalies in the Transistor class of the MVTec-AD dataset. This issue primarily arises because the misplaced anomaly in Transistor is highly similar to the background, causing INP Extractor to incorrectly extract this anomaly as INPs. A more detailed discussion can be found in Appendix Sec. M. In future work, we plan to combine the proposed INPs with pre-stored prototypes to address this limitation. While the pre-stored prototypes encapsulate comprehensive semantic information, the INPs exhibit strong alignment. This integration is expected to significantly improve the model's ability to detect logical anomalies that are similar to the background." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.831, + 0.634, + 0.848 + ], + "angle": 0, + "content": "Acknowledge" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.858, + 0.908, + 0.887 + ], + "angle": 0, + "content": "This study was supported by the National Natural Science Foundation of China under Grant No. 52375494." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9981" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5553-5561, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. The MVTec anomaly detection dataset: A comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision, 129(4):1038-1059, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[4] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Yunkang Cao, Haiming Yao, Wei Luo, and Weiming Shen. Varad: Lightweight high-resolution image anomaly detection via visual autoregressive modeling. IEEE Transactions on Industrial Informatics, pages 1-10, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[9] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 20 24. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Hewei Guo, Liping Ren, Jingjing Fu, Yuwang Wang, Zhizheng Zhang, Cuiling Lan, Haoqian Wang, and Xinwen Hou. Template-guided hierarchical feature restoration for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6447-6458, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Jia Guo, Shuai Lu, Weihang Zhang, Fang Chen, Hongen Liao, and Huiqi Li. Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. arXiv preprint arXiv:2405.14325, 2024. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[15] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[18] Chao Huang, Chengliang Liu, Zheng Zhang, Zhihao Wu, Jie Wen, Qiuping Jiang, and Yong Xu. Pixel-level anomaly detection via uncertainty-aware prototypical transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 521-530, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[19] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.552, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[20] Wenli Huang, Ye Deng, Siqi Hui, Yang Wu, Sanping Zhou, and Jinjun Wang. Sparse self-attention transformer for image inpainting. Pattern Recognition, 145:109897, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[21] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19606-19616, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[22] Yuxin Jiang, Yunkang Cao, and Weiming Shen. Prototypical learning guided context-aware segmentation network for few-shot anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[23] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16838-16848, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In 2017" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9982" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[27] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36:8487-8500, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[28] Wei Luo, Haiming Yao, and Wenyong Yu. Template-based feature aggregation network for industrial anomaly detection. Engineering Applications of Artificial Intelligence, 131: 107810, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[29] Wei Luo, Haiming Yao, Wenyong Yu, and Zhengyong Li. AMI-Net: Adaptive mask inpainting network for industrial anomaly detection and localization. IEEE Transactions on Automation Science and Engineering, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[32] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[35] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[36] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T.M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detec" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22883-22892, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[39] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Jiayi Lyu, Yong Liu, Chengjie Wang, Feng Zheng, and Yaochu Jin. IM-IAD: Industrial image anomaly detection benchmark in manufacturing. IEEE Transactions on Cybernetics, pages 1-14, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[40] Haiming Yao, Yunkang Cao, Wei Luo, Weihang Zhang, Wenyong Yu, and Weiming Shen. Prior normality prompt transformer for multiclass industrial image anomaly detection. IEEE Transactions on Industrial Informatics, 20(10): 11866-11876, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[41] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, pages 4571-4584, 2022. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.907, + 0.53 + ], + "angle": 0, + "content": "[43] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[44] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[45] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[46] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.905, + 0.811 + ], + "angle": 0, + "content": "[47] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9983" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_origin.pdf b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3a582369bc0451a5a8f44bfb72b7282e579e963 --- /dev/null +++ b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/8e301c2a-3c32-4f9e-8f87-33596952eadb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:373682a754f833394916f266d2e11db4bc523b0f6f4973e9202af724b0630112 +size 9584369 diff --git a/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/full.md b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..23d7bfd4778cf89acd925f502e6dec9316de1739 --- /dev/null +++ b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/full.md @@ -0,0 +1,307 @@ +# Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection + +Wei Luo $^{1*}$ Yunkang Cao $^{2*}$ Haiming Yao $^{1*}$ Xiaotian Zhang $^{1}$ Jianan Lou $^{1}$ Yuqi Cheng $^{2}$ Weiming Shen $^{2}$ Wenyong Yu $^{2\dagger}$ Tsinghua University ${}^{2}$ Huazhong University of Science and Technology {luow23, yhm22, zxt19, ljn22}@mails.tsinghua.edu.cn {cyk_hust, yuqicheng, shenwm, ywy}@hust.edu.cn + +# Abstract + +Anomaly detection (AD) is essential for industrial inspection, yet existing methods typically rely on "comparing" test images to normal references from a training set. However, variations in appearance and positioning often complicate the alignment of these references with the test image, limiting detection accuracy. We observe that most anomalies manifest as local variations, meaning that even within anomalous images, valuable normal information remains. We argue that this information is useful and may be more aligned with the anomalies since both the anomalies and the normal information originate from the same image. Therefore, rather than relying on external normality from the training set, we propose INP-Former, a novel method that extracts Intrinsic Normal Prototypes (INPs) directly from the test image. Specifically, we introduce the INP Extractor, which linearly combines normal tokens to represent INPs. We further propose an INP Coherence Loss to ensure INPs can faithfully represent normality for the testing image. These INPs then guide the INP-Guided Decoder to reconstruct only normal tokens, with reconstruction errors serving as anomaly scores. Additionally, we propose a Soft Mining Loss to prioritize hard-to-optimize samples during training. INP-Former achieves state-of-the-art performance in single-class, multi-class, and few-shot AD tasks across MVtec-AD, VisA, and Real-IAD, positioning it as a versatile and universal solution for AD. Remarkably, INP-Former also demonstrates some zero-shot AD capability. Code is available at: https://github.com/luow23/INP-Former. + +# 1. Introduction + +Unsupervised image anomaly detection (AD) [3, 39] seeks to identify abnormal patterns in images and localize anom + +lous regions by learning solely from normal samples. This technique has seen widespread application in industrial defect detection [2, 47] and medical disease screening [19]. Recently, various specialized tasks have emerged in response to real-world demands, from conventional single-class AD [28, 34] to more advanced few-shot AD [17, 22] and multi-class AD [14, 40, 42]. + +Although the composition of normal samples varies across these tasks, the fundamental principle remains unchanged: modeling normality in the training data and assessing whether a test image aligns with this learned normality. However, this approach can be limited due to misaligned normality between the training data and the test image. For instance, prototype-based methods [34] extract representative normal prototypes to capture the normality of training samples. In few-shot AD, intra-class variance may lead to poorly aligned prototypes [17], e.g., hazelnuts in different appearances and positions, as shown in Fig. 1(a). Increasing the sample size can mitigate this problem but at the cost of additional prototypes and reduced inference efficiency. When there are multiple classes, i.e., multi-class AD, prototypes from one class may resemble anomalies from another, like the normal background of hazelnut is similar to the anomalies in cable in Fig. 1(b), leading to misclassification. + +Several works have focused on extracting normality that is more aligned with the test image. For instance, some studies [13, 17, 45] propose spatially aligning normality within a single class through geometrical transformations. However, spatial alignment is ineffective for certain objects, such as hazelnuts, which exhibit variations beyond spatial positions. Other approaches [27, 40, 41] attempt to divide the normality in the training set into smaller, specific portions and then compare the test image to the corresponding portion of normality, but may still fail to find perfect alignment because of intra-class variances. + +Rather than attempting to extract more aligned normal- + +![](images/2de706df4c771685ae090d74c4c72bdc33bbc345986ce083e7ddb560289dc046.jpg) +Figure 1. Motivation for Intrinsic Normal Prototypes (INPs). (a) Pre-stored prototypes from few-shot normal samples may fail to represent all normal patterns. (b) Pre-stored prototypes from one class can be similar to anomalies in another class. (c) The extracted INPs are concise yet well-aligned to the test image, alleviating the issues in (a) and (b). + +ity from the training set, we propose addressing the issue of misaligned normality by leveraging the normality within the test image itself as prototypes, which we term Intrinsic Normal Prototypes (INPs). As illustrated in Fig. 1(c), normal patches within an anomalous test image can function as INPs, and anomalies can be easily detected by comparing them with these INPs. These INPs provide more concise and well-aligned prototypes to the anomalies than those learned from training data, as they typically share the same geometrical context and similar appearances with the abnormal regions within the testing image itself. Accordingly, we explore the prevalence of INPs in various AD scenarios and evaluate their potential to improve AD performance. + +Although previous work [1] has attempted to utilize INP for anomaly detection, it employs handcrafted aggregated features as prototypes, thus limiting the method to zero-shot texture anomaly detection. In contrast, we introduce a learnable INP Extractor to extract normal features with adaptable shapes as INPs. We also propose an INP Coherence Loss to ensure that the extracted INPs coherently represent the normality within the test image, avoiding the capture of anomalous regions. However, some weakly representative normal regions are challenging to model with a limited set of discrete INPs, resulting in background noise (Fig. 4(c)). To address this issue, we introduce an INP-Guided Decoder, which integrates INPs into a reconstruction-based framework. This decoder leverages combinations of discrete INPs to accurately reconstruct all normal regions while effectively suppressing the reconstruction of anomalous regions, with reconstruction errors serving as anomaly scores. Furthermore, we introduce a Soft Mining Loss that focuses on normal regions that are challenging to reconstruct, i.e., hard samples, thereby improving overall reconstruction quality and enhancing AD performance. + +Our approach, termed INP-Former, primarily leverages vision transformers (ViTs) for both INP extraction and INP-guided reconstruction. Extensive experiments on MVtec-AD [2], VisA [47], and Real-IAD [37] demonstrate that INP-Former achieves superior performance across multiclass, single-class, and few-shot AD tasks, positioning + +INP-Former as a universal AD solution. INP-Former also optimizes computational complexity by extracting concise INPs, e.g., images can be represented effectively using only six INPs, as shown in Sec. 4.3.2. Additionally, as demonstrated in Sec. 4.4.2, INP-Former exhibits strong generalization and can even extract INPs for unseen classes, enabling zero-shot AD capabilities. In summary, our main contributions are: + +- We demonstrate that a single image can contain Intrinsic Normal Prototypes (INPs), offering concise and aligned normality for anomaly detection. +- We propose the INP Extractor and incorporate INPs into a reconstruction-based anomaly detection framework using the INP-Guided Decoder. +- We introduce the INP Coherence Loss to extract representative INPs and the Soft Mining Loss to enhance reconstruction quality. + +# 2. Related Works + +# 2.1. Universal Anomaly Detection + +There are numerous unsupervised AD tasks, ranging from conventional single-class AD to recent few-shot and multiclass AD setups. We refer to these collectively as universal anomaly detection. + +Single-Class Anomaly Detection: This setup was originally introduced by MVTec-AD [2] and involves developing distinct AD models for each class. Typically, images are embedded into a feature space using a pre-trained encoder, after which various schemes, such as reconstruction-based [5, 29, 44], knowledge-distillation-based [9, 36], prototype-based [34, 45], and embedding-based [26, 43] methods, are employed to learn the normality of the given class. While these approaches achieve strong performance, their reliance on class-specific models limits scalability when dealing with a wide range of classes. + +Few-shot Anomaly Detection: In practical scenarios, the number of available normal samples may also be limited, motivating the development of few-shot AD methods. In this case, normal samples may not fully capture the variability of normality. To address this challenge, approaches + +![](images/a80e01ce54e0b4a20d117d19b6d5e6f0551585e619cc443ffd876fcfa29e1ab9.jpg) +Figure 2. Overview of our INP-Former framework for universal anomaly detection. (a) Our model consists of a pre-trained Encoder, an INP Extractor, a Bottleneck, and an INP-Guided decoder. The INP Extractor dynamically extracts intrinsic normal prototypes from a single image, which the INP-Guided Decoder leverages to effectively suppress anomalous features. (b) Detailed architecture of the INP Extractor. (c) Detailed architecture of each layer in the INP-Guided Decoder. (d) Comparison of computational complexity between INP-Guided Attention and Self Attention. It is important to note that the patch token (Encoding) and patch token (Decoding) refer to the patch tokens utilized during the encoding and decoding stages, respectively. + +such as spatial alignment [17] or contrastive learning [22] are used to create more compact and representative normal embeddings. Recently, Vision-Language Models (VLMs) like CLIP [33] have proven effective for few-shot AD due to their broad, pre-trained knowledge. These VLMs not only provide descriptive visual embeddings but also compute the similarity between text prompts and test images, as seen in works like WinCLIP [21], AnomalyGPT [12], and InCTRL [46]. Some approaches, such as AdaCLIP [4], even enable zero-shot AD through VLMs. + +Multi-Class Anomaly Detection: Developing separate models for each class can be resource-intensive, prompting interest in multi-class AD, also known as unified AD [42], which aims to build a single model for multiple classes. UniAD [42] pioneered a unified reconstruction framework for anomaly detection, followed by HVQ-Trans [27] addressed the identical shortcut problem using a vector quantization framework. More recent approaches, such as MambaAD [15] and Dinomaly [14], further enhance multi-class AD performance by leveraging advanced models, i.e., the State Space Model Mamba [11] and DINO [7], respectively. However, these methods lack the functionality to derive aligned normality with the test image. On the contrary, we extract INPs from the testing image, bringing aligned and precise normality for anomaly detection. + +# 2.2. Prototype Learning + +Prototype learning [35] aims to extract representative prototypes from a given training set, which are then used for + +classification by measuring their distances to a test sample in a metric space. This technique is widely used in few-shot learning [23]. Several AD methods also employ prototype learning. For example, PatchCore [34] extracts multiple normal prototypes to represent the normality of the training data, directly computing the minimal distances to the test sample for anomaly detection. Other approaches [10, 18, 30, 32] incorporate prototypes into the reconstruction process to avoid the identical shortcut issue. Specifically, they replace the original inputs with combinations of learned normal prototypes, ensuring that the inputs to the reconstruction model contain only normal elements. However, these methods rely on pre-stored normal prototypes extracted from the training set, which can suffer from the misaligned normality problem. In contrast, our INPs are dynamically extracted from the test image, providing more aligned alternatives for normality representation. + +# 3. Method: INP-Former + +# 3.1. Overview + +To fully exploit the advantages of INPs in anomaly detection, we propose INP-Former, as depicted in Fig. 2(a). The model dynamically extracts INPs from a single image and utilizes them to guide the feature reconstruction process, with the reconstruction errors serving as anomaly scores. Specifically, it comprises four key modules: a fixed pre-trained Encoder $\mathcal{Q}$ , an INP Extractor $\mathcal{E}$ , a Bottleneck $\mathcal{B}$ , and an INP-Guided Decoder $\mathcal{D}$ . The input im + +age $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ is first processed by the pre-trained Encoder $\mathcal{Q}$ to extract multi-scale latent features $f_{\mathcal{Q}} = \{f_{\mathcal{Q}}^{1}, \ldots, f_{\mathcal{Q}}^{L} | f_{\mathcal{Q}}^{l} \in \mathbb{R}^{N \times C}, N = \frac{HW}{k^{2}}\}$ , where $k$ represents the downsampling factor. Next, the INP Extractor $\mathcal{E}$ extracts $M$ INPs $\mathbf{P} = \{p_{1}, \ldots, p_{M} | p_{m} \in \mathbb{R}^{C}\}$ from the pre-trained features, with an INP coherence loss ensuring that the extracted INPs consistently represent normal features during testing. The Bottleneck $\mathcal{B}$ subsequently fuses the multi-scale latent features, producing the fused output $F_{\mathcal{B}} = \mathcal{B}(f_{\mathcal{Q}})$ . Following the bottleneck, the extracted INPs are utilized to guide the Decoder $\mathcal{D}$ to yield reconstruction outputs $f_{\mathcal{D}} = \{f_{\mathcal{D}}^{1}, \ldots, f_{\mathcal{D}}^{L} | f_{\mathcal{D}}^{l} \in \mathbb{R}^{N \times C}\}$ with only normal patterns, thus the reconstruction error between $f_{\mathcal{Q}}$ and $f_{\mathcal{D}}$ can serve as the anomaly score. + +# 3.2. INP Extractor + +Existing prototype-based methods [10, 31, 34] store local normal features from the training data and compare them with test images. However, the misaligned normality between these pre-stored prototypes and the test images and the lack of global information lead to suboptimal detection performance. To address these limitations, we propose the INP Extractor to dynamically extract INPs with global information from the test image itself. + +Specifically, as illustrated in Fig. 2(b), instead of extracting representative local features as done in PatchCore [34], we employ cross attention to aggregate the global semantic information of the pre-trained features $\mathbf{F}_{\mathcal{Q}} \in \mathbb{R}^{N \times C}$ with $M$ learnable tokens $\mathbf{T} = \{t_1, \dots, t_M | t_m \in \mathbb{R}^C\}$ . Here $\mathbf{F}_{\mathcal{Q}}$ is used as the key-value pairs, while $\mathbf{T}$ serve as the query, allowing $\mathbf{T}$ to linearly aggregate $\mathbf{F}_{\mathcal{Q}}$ into INPs $\mathbf{P} = \{p_1, \dots, p_M | p_m \in \mathbb{R}^C\}$ . + +$$ +\mathbf {F} _ {\mathcal {Q}} = \operatorname {s u m} \left(\left\{f _ {\mathcal {Q}} ^ {1}, \dots , f _ {\mathcal {Q}} ^ {L} \right\}\right) +$$ + +$$ +Q = \mathbf {T} W ^ {Q}, K = \mathbf {F} _ {\mathcal {Q}} W ^ {K}, V = \mathbf {F} _ {\mathcal {Q}} W ^ {V} \tag {1} +$$ + +$$ +\mathbf {T} ^ {\prime} = \operatorname {A t t e n t i o n} (Q, K, V) + \mathbf {T} +$$ + +$$ +\mathbf {P} = \mathrm {F F N} (\mathbf {T} ^ {\prime}) + \mathbf {T} ^ {\prime} +$$ + +where $\mathrm{sum}(\cdot)$ denotes the element-wise summation. $Q\in$ $\mathbb{R}^{M\times C}$ and $K,V\in \mathbb{R}^{N\times C}$ represent the query, key and value, respectively. $W^{Q},W^{k},W^{v}\in \mathbb{R}^{C\times C}$ are the learnable projection parameters for $Q,K,V;\mathrm{FFN}(\cdot)$ represents the feed-forward network. + +To ensure that INPs coherently represent normal features while minimizing the capture of anomalous features during the testing process, we propose an INP coherence loss $\mathcal{L}_c$ to minimize the distances between individual normal features and the corresponding nearest INP. + +$$ +\begin{array}{l} d _ {i} = \min _ {m \in \{1, \dots , M \}} \mathcal {S} (\mathbf {F} _ {\mathcal {Q}} (i), p _ {m}) \\ \mathcal {L} _ {c} = \frac {1}{N} \sum_ {i = 1} ^ {N} d _ {i} \tag {2} \\ \end{array} +$$ + +Table 1. Comparison of computational cost and memory usage. + +
Number of multiplicaiton and addition
CalculationVanilla Self AttentionINP-Guided Attention
Al=Ql(Kl)T943 496 9607 220 640
fdl-1'=AlVl943 509 5046 623 232
Memory usage (MB)
Ql/Kl/Vl/Al2.30/2.30/2.30/2.342.30/0.018/0.018/0.018
+ +where $S(\cdot, \cdot)$ denotes the cosine distance. $d_{i}$ represents the distance between the query feature $\mathbf{F}_{\mathcal{Q}}(i)$ and the corresponding nearest INP item. Fig. 4 visually illustrates the effectiveness of $\mathcal{L}_c$ . + +# 3.3. INP-Guided Decoder + +While we can use the distance between testing features and their nearest INPs for anomaly detection, as illustrated in Fig. 4(c), certain low-representative normal regions are difficult to model with a limited number of discrete INPs, leading to noisy distance maps between these INPs and testing features. To address this issue, we propose the INP-Guided Decoder, aiming to reconstruct these low-representative normal regions through a combination of multiple discrete INPs and suppress the reconstruction of anomalous regions. Additionally, this decoder provides a token-wise discrepancy that can be directly leveraged for anomaly detection. As shown in Fig. 2(c), INPs are incorporated into this decoder to guide the reconstruction process. Since INPs exclusively represent normal patterns in test images, we employ the extracted INPs as key-value pairs, ensuring that the output is a linear combination of normal INPs, thereby effectively suppressing the reconstruction of anomalous queries, i.e., the identical mapping issue [42]. Furthermore, we find that the first residual connection can directly introduce anomalous features into the subsequent reconstruction, so we remove this connection in our INP-Guided Decoder. Following the previous work [20], we also employ the ReLU activation function to mitigate the influence of weak correlations and noise on the attention maps. + +Mathematically, let $f_{\mathcal{D}}^{l - 1} \in \mathbb{R}^{N \times C}$ denotes the output latent features from previous decoding layer. The output $f_{\mathcal{D}}^{l} \in \mathbb{R}^{N \times C}$ of the $l_{th}$ decoding layer is formulated as, + +$$ +Q _ {l} = f _ {\mathcal {D}} ^ {l - 1} W _ {l} ^ {Q}, K _ {l} = \mathbf {P} W _ {l} ^ {K}, V _ {l} = \mathbf {P} W _ {l} ^ {V} +$$ + +$$ +f _ {\mathcal {D}} ^ {l - 1 ^ {\prime}} = A _ {l} V _ {l}, A _ {l} = \operatorname {R e L U} \left(Q _ {l} \left(K _ {l}\right) ^ {T}\right) \tag {3} +$$ + +$$ +f _ {\mathcal {D}} ^ {l} = \operatorname {F F N} \left(f _ {\mathcal {D}} ^ {l - 1 ^ {\prime}}\right) + f _ {\mathcal {D}} ^ {l - 1 ^ {\prime}} +$$ + +where $Q_{l}\in \mathbb{R}^{N\times C}$ and $K_{l},V_{l}\in \mathbb{R}^{M\times C}$ denote the query, key and value of the $l_{th}$ decoding layer. $W_{l}^{Q},W_{l}^{k},W_{l}^{v}\in$ $\mathbb{R}^{C\times C}$ denote the learnable projection parameters for $Q_{l},K_{l},V_{l}$ . $A_{l}\in \mathbb{R}^{N\times M}$ represent the attention map. + +Attention Complexity Analysis: As depicted in Fig. 2(d), the computational complexity of vanilla self-attention is + +$\mathcal{O}(N^2 C)$ , while its memory usage is $\mathcal{O}(N^2)$ . In contrast, our INP-Guided Attention reduces both the computational complexity and memory usage to $\mathcal{O}(NMC)$ and $\mathcal{O}(NM)$ , respectively, which can be approximated as $\mathcal{O}(NC)$ and $\mathcal{O}(N)$ due to $M \ll N$ . Tab. 1 offers a detailed comparison of the complexity of vanilla self-attention and INP-Guided Attention. Appendix Sec. C compares the overall complexity between INP-Former and other methods. The light version of INP-Former can even be more efficient than MambaAD [15] yet demonstrate better performance. + +# 3.4. Soft Mining Loss + +Inspired by Focal Loss [25], different regions should be assigned varying weights based on their optimization difficulty. Accordingly, we propose Soft Mining Loss to encourage the model to focus more on difficult regions. + +Intuitively, the ratio of the reconstruction error of an individual normal region to the average reconstruction error of all normal regions can serve as an indicator of optimization difficulty. Given the encoder $f_{Q}^{l}$ and decoder $f_{D}^{l}$ features at layer $l$ , let $M^{l}$ denote the regional cosine distance. Our soft mining loss $\mathcal{L}_{sm}$ is defined as follows: + +$$ +w ^ {l} (h, w) = \left[ \frac {M ^ {l} (h , w)}{u (M ^ {l})} \right] ^ {\gamma} +$$ + +$$ +\mathcal {L} _ {s m} = \frac {1}{L} \sum_ {l = 1} ^ {L} 1 - \frac {\operatorname {v e c} \left(f _ {\mathcal {Q}} ^ {l}\right) ^ {T} \cdot \operatorname {v e c} \left(\hat {f} _ {\mathcal {D}} ^ {l}\right)}{\left| \left| \operatorname {v e c} \left(f _ {\mathcal {Q}} ^ {l}\right) \right| \right| \left| \left| \operatorname {v e c} \left(\hat {f} _ {\mathcal {D}} ^ {l}\right) \right| \right|} \tag {4} +$$ + +$$ +\hat {f} _ {\mathcal {D}} ^ {l} (h, w) = c g \left(f _ {\mathcal {D}} ^ {l} (h, w)\right) _ {w ^ {l} (h, w)} +$$ + +where $u(M^l)$ represents the average regional cosine distance within a batch, $\gamma \geq 0$ denotes the temperature hyperparameter, $cg(\cdot)_{w^{l}(h,w)}$ denotes a gradient adjustment based on dynamic weight $w^{l}(h,w)$ , and $vec(\cdot)$ denotes the flattening operation. The overall training loss of our INP-Former can be expressed as follows: $\mathcal{L}_{total} = \mathcal{L}_{sm} + \lambda \mathcal{L}_c$ . + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets: We conduct a comprehensive analysis of the proposed INP-Former on three widely used AD datasets: MVTec-AD [2], VisA [47], and Real-IAD [37]. MVTec-AD consists of 15 categories, with 3,629 normal images for training, and 1,982 anomalous images along with 498 normal images for testing. VisA contains 12 object categories, with 8,659 normal images for training and 962 normal images along with 1,200 anomalous images for testing. + +Real-IAD includes 30 different objects, with 36,645 normal images for training and 63,256 normal images along with 51,329 anomalous images for testing. + +Metrics: Following existing works [14, 15], we use the Area Under the Receiver Operating Characteristic Curve + +(AUROC), Average Precision (AP), and F1-score-max (F1_max) to evaluate anomaly detection and localization. For anomaly localization specifically, we use Area Under the Per-Region-Overlap (AUPRO) as an additional metric. + +Implementation Details: INP-Former adopts ViT-Base/14 with DINO2-R [7] weights as the default pre-trained encoder. The number $M$ of INPs is set to six by default. The INP Extractor includes a standard Vision Transformer block. The layer number of the INP-Guided decoder is eight. All input images are resized to $448^2$ and then center-cropped to $392^2$ . The hyperparameters $\gamma$ and $\lambda$ are set to 3.0 and 0.2, respectively. We use the StableAdamW [38] optimizer with a learning rate $1e^{-3}$ and a weight decay of $1e^{-4}$ for 200 epochs. Notably, the above hyperparameters do not require any adjustment across the three datasets. Appendix Sec. A presents more implementation details. Appendix Sec. D, E, and F analyze the influence of input resolution, ViT architecture, and $\lambda$ , respectively. + +# 4.2. Main Results + +# 4.2.1 Multi-Class Anomaly Detection + +We compare the proposed INP-Former with several state-of-the-art (SOTA) methods for multi-class anomaly detection, including reconstruction-based methods RD4AD [9], UniAD [42], DiAD [16], MambaAD [15], and Dinomaly [14], and embedding-based methods SimpleNet [26] and DeSTSeg [43]. A detailed introduction to the comparison methods can be found in Appendix Sec. B. + +The experimental results on the three AD datasets are presented in Tab. 2. On the widely used MVTec-AD dataset, our method achieves SOTA performance, with image-level metrics of 99.7/99.9/99.2 and pixel-level metrics of 98.5/71.0/69.7/94.9. On VisA, our method achieves competitive results, attaining the best image-level metrics of 98.9/99.0/96.6, and achieving the best or second-best pixel-level performance of 98.8/51.2/54.7/94.4. On the more complex and challenging Real-IAD dataset, our method reaches new SOTA performance, with image-level metrics of 90.5/88.1/81.5 and pixel-level metrics of 99.0/47.5/50.3/95.0. Compared to the second-best results, our method improves by $1.2\uparrow /1.3\uparrow /1.3\uparrow$ at the image level and by $0.2\uparrow /4.7\uparrow /3.2\uparrow /1.1\uparrow$ at the pixel level. The SOTA performance achieved across the three datasets showcases the effectiveness and robustness of our method. The per-class performance metrics are presented in Appendix Sec. H. Appendix Sec. G presents the performance of INP-Former in a more challenging scenario, which we call supermulti-class anomaly detection, thus training INP-Former on several datasets - MVTec-AD, VisA, and Real-IAD - simultaneously. Results show that our method can even detect anomalies in more classes without significant performance degradation. Fig. 3 demonstrates the precise anomaly localization capability of our method. More qualitative results + +Table 2. Multi-class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined. + +
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Metric →Image-level(I-AUROC/I-AP/I-F1_max)Pixel-level(P-AUROC/P-AP/P-F1_max/AUPRO)
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
RD4AD [9]94.6/96.5/95.296.1/48.6/53.8/91.192.4/92.4/89.698.1/38.0/42.6/91.882.4/79.0/73.997.3/25.0/32.7/89.6
UniAD [42]96.5/98.8/96.296.8/43.4/49.5/90.788.8/90.8/85.898.3/33.7/39.0/85.583.0/80.9/74.397.3/21.1/29.2/86.7
SimpleNet [26]95.3/98.4/95.896.9/45.9/49.7/86.587.2/87.0/81.896.8/34.7/37.8/81.457.2/53.4/61.575.7/2.8/6.5/39.0
DeSTSeg [43]89.2/95.5/91.693.1/54.3/50.9/64.888.9/89.0/85.296.1/39.6/43.4/67.482.3/79.2/73.294.6/37.9/41.7/40.6
DiAD [16]97.2/99.0/96.596.8/52.6/55.5/90.786.8/88.3/85.196.0/26.1/33.0/75.275.6/66.4/69.988.0/2.9/7.1/58.1
MambaAD [15]98.6/99.6/97.897.7/56.3/59.2/93.194.3/94.5/89.498.5/39.4/44.0/91.086.3/84.6/77.098.5/33.0/38.7/90.5
Dinomaly [14]99.6/99.8/99.098.4/69.3/69.2/94.898.7/98.9/96.298.7/53.2/55.7/94.589.3/86.8/80.298.8/42.8/47.1/93.9
INP-Former99.7/99.9/99.298.5/71.0/69.7/94.998.9/99.0/96.698.9/51.2/54.7/94.490.5/88.1/81.599.0/47.5/50.3/95.0
+ +Table 3. Few-shot (4-shot) anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined. $\dagger$ indicates the results we reproduced using publicly available code. + +
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
SPADE [6]84.8/92.5/91.592.7/-/46.2/87.081.7/83.4/82.196.6/-/43.6/87.350.8†/45.8†/61.2†59.5†/0.2†/0.5†/19.2†
PaDiM [8]80.4/90.5/90.292.6/-/46.1/81.372.8/75.6/78.093.2/-/24.6/72.660.3†/53.5†/64.0†90.9†/2.1†/5.1†/67.6†
PatchCore [34]88.8/94.5/92.694.3/-/55.0/84.385.3/87.5/84.396.8/-/43.9/84.966.0†/62.2†/65.2†92.9†/9.8†/16.1†/68.6†
WinCLIP [21]95.2/97.3/94.796.2/-/59.5/89.087.3/88.8/84.297.2/-/47.0/87.673.0†/61.8†/61.0†93.8†/13.3†/21.0†/76.4†
PromptAD [24]96.6/-/-96.5/-/-/90.589.1/-/-97.4/-/-/86.259.7†/43.5†/52.9†86.9†/8.7†/16.1†/61.9†
INP-Former97.6/98.6/97.097.0/65.9/65.6/92.996.4/96.0/93.097.7/49.3/54.3/93.176.7/72.3/71.797.3/32.2/36.7/89.0
+ +![](images/c3b0024b7decfb692714c19723fd49bd86b99e1b36937b7a9b9d434c3df8fd2f.jpg) +Figure 3. Qualitative results of anomaly localization on the MVTec-AD [2], VisA [47], and Real-IAD [37] datasets for multi-class anomaly detection. The first row presents the input images with their ground truth, while the second row displays the corresponding anomaly maps. + +are presented in Appendix Sec. L. + +# 4.2.2 Few-Shot Anomaly Detection + +We compare our method with several SOTA approaches for few-shot anomaly detection, including prototype-based methods SPADE [6], PaDiM [8], and PatchCore [34] and recent advances that utilize VLMs, i.e., WinCLIP [21] and PromptAD [24]. + +As shown in Tab. 3, our method significantly outperforms previous SOTAs on three different AD datasets. Compared to the second-best results, our method achieves improvements of $1.0\uparrow /1.3\uparrow /2.3\uparrow$ in image-level scores and $0.5\uparrow / - / 6.1\uparrow /2.4\uparrow$ in pixel-level scores on the MVTec-AD dataset. It similarly outperforms the second-best results on + +the VisA dataset, with enhancements of $7.3 \uparrow / 7.2 \uparrow / 8.7 \uparrow$ for image-level and $0.3 \uparrow / -7.3 \uparrow / 5.5 \uparrow$ for pixel-level scores. Additionally, on the Real-IAD dataset, our method surpasses the second-best results by $3.7 \uparrow / 10.1 \uparrow / 6.5 \uparrow$ in image-level and $3.5 \uparrow / 18.9 \uparrow / 15.7 \uparrow / 12.6 \uparrow$ in pixel-level scores. The superior performance of our method in few-shot anomaly detection stems from its ability to extract INPs from a single image, eliminating the need for extensive normal data to pre-store prototypes. More comparison results on 1-shot and 2-shot are presented in Appendix Sec. I. + +# 4.2.3 Single-Class Anomaly Detection + +We further compared our proposed INP-Former with current SOTA methods for single-class anomaly detection, as shown in Tab. 4. The results indicate that INP-Former + +Table 4. Single class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined. + +
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓I-AUROCP-APAUPROI-AUROCP-APAUPROI-AUROCP-APAUPRO
PatchCore [34]99.156.193.595.140.191.289.4-91.5
RD4AD [9]98.558.093.996.027.770.987.1-93.8
SimpleNet [26]99.654.890.096.836.388.788.5-84.6
Dinomaly [14]99.768.995.098.950.795.192.045.295.1
INP-Former99.770.295.498.549.293.892.148.195.6
+ +Table 5. Overall ablation on MVTec-AD [2] and VisA [47] datasets. "INP" refers to the use of INP Extractor and INP-Guided Decoder. + +
Dataset →MVTec-AD [2]VisA [47]
“INP”LcLsmImage-levelPixel-levelImage-levelPixel-level
Module×××98.59/99.18/97.6397.19/61.73/62.94/92.7396.58/97.18/92.8997.50/47.24/51.90/82.85
✓××99.53/99.80/98.8198.32/69.82/69.38/94.6998.11/98.23/95.2298.41/50.34/54.23/93.63
✓✓×99.61/99.83/99.0298.39/70.01/69.53/95.1098.16/98.30/95.4798.46/51.09/54.46/93.71
✓✓✓99.67/99.88/99.2098.48/71.02/69.65/94.8798.90/99.02/96.5798.90/51.22/54.74/94.36
+ +![](images/b419ae5c178a17756174ffa632fc883e6d2df30b7f6d1bd7c9d76d4464b8239b.jpg) +Figure 4. Visualization of the impact of INP coherence loss $\mathcal{L}_c$ . (a) Input anomalous image and ground truth. (b) Distance map without $\mathcal{L}_c$ . (c) Distance map with $\mathcal{L}_c$ . The distance map is obtained by calculating the distance between the input features and their nearest INP terms, as described in Eq. 2. + +achieves new SOTA performance on the MVTec-AD and Real-IAD datasets and demonstrates competitive performance on the VisA dataset. Per-category performance of INP-Former is presented in Appendix Sec. J. + +# 4.3. Ablation Study + +# 4.3.1 Overall Ablation + +As shown in Tab. 5, we conduct comprehensive experiments on MVTec-AD [2] and VisA [47] to validate the effectiveness of the proposed components, i.e., INP Extractor and INP-Guided Decoder ("INP"), INP Coherence Loss $(\mathcal{L}_c)$ , and Soft Mining Loss $(\mathcal{L}_{sm})$ . In the first row, we train a baseline model without incorporating any proposed module, similar to the RD4AD [9] framework. The results in the second row demonstrate that "INP" significantly enhances overall performance. This improvement arises from the fact that "INP" introduces an information bottleneck, which effectively helps the model preserve normal features while filtering out anomalous ones. The results in the third row + +![](images/22e7000012dce8eef883408ac061912895ee9b51e42ed37c700c83452e39baff.jpg) +Figure 5. Visualization of the impact of soft mining loss $\mathcal{L}_{sm}$ . We plot the Kernel Density Estimation (KDE) for the chewing-gum and cashew categories in the VisA [47] dataset to estimate the probability density of the anomaly scores. + +indicate that $\mathcal{L}_c$ enhances the model's performance. This improvement stems from $\mathcal{L}_c$ ensuring that the extracted INP coherently represents normal patterns, thereby avoiding the capture of anomalous ones and establishing a solid foundation for the subsequent suppression of anomalous feature reconstruction. Fig. 4 provides a more intuitive demonstration of the effectiveness of $\mathcal{L}_c$ . The last row indicates that $\mathcal{L}_{sm}$ boosts overall performance, as $\mathcal{L}_{sm}$ directs the model's attention toward more challenging regions, thereby unlocking its optimal performance. Fig. 5 visually illustrates the impact of $\mathcal{L}_{sm}$ , from which we can see that $\mathcal{L}_{sm}$ contributes to a smaller overlap between anomaly score distributions of normal and abnormal pixels. + +# 4.3.2 Influence of the Number of INPs + +As shown in Fig. 6, we conduct an ablation analysis on the number $M$ of INPs. The experimental results indicate that when $M$ exceeds four, the model's performance stabilizes. However, if $M$ becomes excessively large, the extracted INPs may also comprise information from abnormal tokens, leading to a slight decline in overall performance. In our study, we set $M$ to six. + +![](images/13aceb1385fb3b9f411dd582bfdd2235a7cd319df49657c79fa68ca8f86cd493.jpg) +Figure 6. Influence of the number of INPs $M$ on model performance across the MVTec-AD [2] and VisA [47] datasets. Pixel-level AP and F1_max use the right vertical axis, while the other metrics share the left vertical axis. + +![](images/32f97f3471c008bc4aefd2e8daa715f9bf00fd8aad342eb70d71a8d11f3da55a.jpg) +Figure 7. Visualizations of INPs. (a) Input anomalous image and ground truth. (b)-(g) Attention maps of six different INPs. + +# 4.4. Exploration on INPs + +# 4.4.1 Visualizations of INPs + +As shown in Fig. 7, INPs effectively capture different semantic information. Specifically, the learned INPs focus on various regions of the image, including object areas (Fig. 7(b), (e) and (f)), object edges (Fig. 7(c) and (d)), and background areas (Fig. 7(g)). This diversity is attributed to our design of guiding the reconstruction process with INPs. Additionally, INP coherence loss ensures consistency in representing normal features, allowing INPs to concentrate solely on normal regions while ignoring anomalies. This mechanism ensures the decoder reconstructs features containing only normal patterns, thereby improving anomaly detection performance. + +# 4.4.2 Generalization capabilities of INP Extractor + +As shown in Fig. 8, the INP Extractor trained on the RealIAD [37] dataset is capable of detecting INPs on the unseen MVTec-AD [2] dataset, and the distance maps to these INPs can serve for zero-shot anomaly detection. This effectively demonstrates the INP Extractor's ability to dynamically extract INPs from a single image, with the INP coherence loss $\mathcal{L}_c$ ensuring that the extracted INPs coherently cap + +![](images/b4aba017b2ce12e1416aaca3d32cc9a0a13ca6473f0728486e28f1483ab5f166.jpg) +Figure 8. Zero-shot anomaly detection results. Here INP-Former is trained on Real-IAD [37] and tested on MVtec-AD [2]. Distance maps are visualized. + +ture normal patterns. As shown in Appendix Sec. K, without any specific training for zero-shot anomaly detection, our method can even outperform a specified method WinCLIP [21], achieving 88.0 and 88.7 pixel-level AUROCs on MVTec-AD and VisA, respectively. + +# 5. Conclusion + +We propose INP-Former, a novel method for anomaly detection that explores the role of INPs. By learning to linearly combine normal tokens into INPs and using these INPs to guide the reconstruction of normal tokens, INP-Former significantly enhances anomaly detection performance. The introduction of the INP Coherence Loss and Soft Mining Loss further refines INP quality and optimizes the training process. Extensive experiments on MVTec-AD, VisA, and Real-IAD datasets demonstrate that INP-Former achieves SOTA or comparable performance across single-class, multi-class, and few-shot anomaly detection tasks. These results validate the existence and effectiveness of INPs, which can even be extracted from images in unseen categories, enabling zero-shot anomaly detection. + +Limitations & Future Works. Our method encounters certain limitations when detecting logical anomalies that closely resemble the background distribution, such as the misplaced anomalies in the Transistor class of the MVTec-AD dataset. This issue primarily arises because the misplaced anomaly in Transistor is highly similar to the background, causing INP Extractor to incorrectly extract this anomaly as INPs. A more detailed discussion can be found in Appendix Sec. M. In future work, we plan to combine the proposed INPs with pre-stored prototypes to address this limitation. While the pre-stored prototypes encapsulate comprehensive semantic information, the INPs exhibit strong alignment. This integration is expected to significantly improve the model's ability to detect logical anomalies that are similar to the background. + +# Acknowledge + +This study was supported by the National Natural Science Foundation of China under Grant No. 52375494. + +# References + +[1] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5553-5561, 2023. 2 +[2] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. The MVTec anomaly detection dataset: A comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision, 129(4):1038-1059, 2021. 1, 2, 5, 6, 7, 8 +[3] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1 +[4] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision. Springer, 2024. 3 +[5] Yunkang Cao, Haiming Yao, Wei Luo, and Weiming Shen. Varad: Lightweight high-resolution image anomaly detection via visual autoregressive modeling. IEEE Transactions on Industrial Informatics, pages 1-10, 2025. 2 +[6] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 6 +[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 5 +[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 6 +[9] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 2, 5, 6, 7 +[10] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3, 4 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3 +[12] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 20 24. 3 +[13] Hewei Guo, Liping Ren, Jingjing Fu, Yuwang Wang, Zhizheng Zhang, Cuiling Lan, Haoqian Wang, and Xinwen Hou. Template-guided hierarchical feature restoration for + +anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6447-6458, 2023. 1 +[14] Jia Guo, Shuai Lu, Weihang Zhang, Fang Chen, Hongen Liao, and Huiqi Li. Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. arXiv preprint arXiv:2405.14325, 2024. 1, 3, 5, 6, 7 +[15] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 6 +[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 5, 6 +[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3 +[18] Chao Huang, Chengliang Liu, Zheng Zhang, Zhihao Wu, Jie Wen, Qiuping Jiang, and Yong Xu. Pixel-level anomaly detection via uncertainty-aware prototypical transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 521-530, 2022. 3 +[19] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1 +[20] Wenli Huang, Ye Deng, Siqi Hui, Yang Wu, Sanping Zhou, and Jinjun Wang. Sparse self-attention transformer for image inpainting. Pattern Recognition, 145:109897, 2024. 4 +[21] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19606-19616, 2023. 3, 6, 8 +[22] Yuxin Jiang, Yunkang Cao, and Weiming Shen. Prototypical learning guided context-aware segmentation network for few-shot anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2024. 1, 3 +[23] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 3 +[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16838-16848, 2024. 6 +[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In 2017 + +IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5 +[26] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 2, 5, 6, 7 +[27] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36:8487-8500, 2023. 1, 3 +[28] Wei Luo, Haiming Yao, and Wenyong Yu. Template-based feature aggregation network for industrial anomaly detection. Engineering Applications of Artificial Intelligence, 131: 107810, 2024. 1 +[29] Wei Luo, Haiming Yao, Wenyong Yu, and Zhengyong Li. AMI-Net: Adaptive mask inpainting network for industrial anomaly detection and localization. IEEE Transactions on Automation Science and Engineering, 2024. 2 +[30] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 3 +[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4 +[32] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 3 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3 +[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 2, 3, 4, 6, 7 +[35] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3 +[36] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T.M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 2 +[37] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detec + +tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22883-22892, 2024. 2, 5, 6, 7, 8 +[38] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 5 +[39] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Jiayi Lyu, Yong Liu, Chengjie Wang, Feng Zheng, and Yaochu Jin. IM-IAD: Industrial image anomaly detection benchmark in manufacturing. IEEE Transactions on Cybernetics, pages 1-14, 2024. 1 +[40] Haiming Yao, Yunkang Cao, Wei Luo, Weihang Zhang, Wenyong Yu, and Weiming Shen. Prior normality prompt transformer for multiclass industrial image anomaly detection. IEEE Transactions on Industrial Informatics, 20(10): 11866-11876, 2024. 1 +[41] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1 +[42] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, pages 4571-4584, 2022. 1, 3, 4, 5, 6 +[43] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 2, 5, 6 +[44] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 2 +[45] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 1, 2 +[46] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 3 +[47] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2, 5, 6, 7, 8 \ No newline at end of file diff --git a/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/images.zip b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c829883a4828610b1b3f642230993f522d68469 --- /dev/null +++ b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f39a58e71ac261d9b8c585f01f82239e854107a284d4bea0cd8bcb7918afe4e4 +size 883497 diff --git a/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/layout.json b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56c02df4365ed50e4a12c4f0d4218847f8e936c1 --- /dev/null +++ b/2025/Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection/layout.json @@ -0,0 +1,8613 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 504, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 504, + 138 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 504, + 138 + ], + "type": "text", + "content": "Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": "Wei Luo" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Yunkang Cao" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Haiming Yao" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Xiaotian Zhang" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Jianan Lou" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Yuqi Cheng" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Weiming Shen" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Wenyong Yu" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 108, + 160, + 501, + 232 + ], + "type": "text", + "content": "Huazhong University of Science and Technology {luow23, yhm22, zxt19, ljn22}@mails.tsinghua.edu.cn {cyk_hust, yuqicheng, shenwm, ywy}@hust.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 297, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 629 + ], + "type": "text", + "content": "Anomaly detection (AD) is essential for industrial inspection, yet existing methods typically rely on \"comparing\" test images to normal references from a training set. However, variations in appearance and positioning often complicate the alignment of these references with the test image, limiting detection accuracy. We observe that most anomalies manifest as local variations, meaning that even within anomalous images, valuable normal information remains. We argue that this information is useful and may be more aligned with the anomalies since both the anomalies and the normal information originate from the same image. Therefore, rather than relying on external normality from the training set, we propose INP-Former, a novel method that extracts Intrinsic Normal Prototypes (INPs) directly from the test image. Specifically, we introduce the INP Extractor, which linearly combines normal tokens to represent INPs. We further propose an INP Coherence Loss to ensure INPs can faithfully represent normality for the testing image. These INPs then guide the INP-Guided Decoder to reconstruct only normal tokens, with reconstruction errors serving as anomaly scores. Additionally, we propose a Soft Mining Loss to prioritize hard-to-optimize samples during training. INP-Former achieves state-of-the-art performance in single-class, multi-class, and few-shot AD tasks across MVtec-AD, VisA, and Real-IAD, positioning it as a versatile and universal solution for AD. Remarkably, INP-Former also demonstrates some zero-shot AD capability. Code is available at: https://github.com/luow23/INP-Former." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 643, + 135, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 643, + 135, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 643, + 135, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 663, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 295, + 688 + ], + "type": "text", + "content": "Unsupervised image anomaly detection (AD) [3, 39] seeks to identify abnormal patterns in images and localize anom" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 259, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 554, + 344 + ], + "type": "text", + "content": "lous regions by learning solely from normal samples. This technique has seen widespread application in industrial defect detection [2, 47] and medical disease screening [19]. Recently, various specialized tasks have emerged in response to real-world demands, from conventional single-class AD [28, 34] to more advanced few-shot AD [17, 22] and multi-class AD [14, 40, 42]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 346, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 562 + ], + "type": "text", + "content": "Although the composition of normal samples varies across these tasks, the fundamental principle remains unchanged: modeling normality in the training data and assessing whether a test image aligns with this learned normality. However, this approach can be limited due to misaligned normality between the training data and the test image. For instance, prototype-based methods [34] extract representative normal prototypes to capture the normality of training samples. In few-shot AD, intra-class variance may lead to poorly aligned prototypes [17], e.g., hazelnuts in different appearances and positions, as shown in Fig. 1(a). Increasing the sample size can mitigate this problem but at the cost of additional prototypes and reduced inference efficiency. When there are multiple classes, i.e., multi-class AD, prototypes from one class may resemble anomalies from another, like the normal background of hazelnut is similar to the anomalies in cable in Fig. 1(b), leading to misclassification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 566, + 556, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 556, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 556, + 697 + ], + "type": "text", + "content": "Several works have focused on extracting normality that is more aligned with the test image. For instance, some studies [13, 17, 45] propose spatially aligning normality within a single class through geometrical transformations. However, spatial alignment is ineffective for certain objects, such as hazelnuts, which exhibit variations beyond spatial positions. Other approaches [27, 40, 41] attempt to divide the normality in the training set into smaller, specific portions and then compare the test image to the corresponding portion of normality, but may still fail to find perfect alignment because of intra-class variances." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "Rather than attempting to extract more aligned normal-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9974" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 550, + 167 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 550, + 167 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 550, + 167 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 550, + 167 + ], + "type": "image", + "image_path": "2de706df4c771685ae090d74c4c72bdc33bbc345986ce083e7ddb560289dc046.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 172, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 555, + 206 + ], + "type": "text", + "content": "Figure 1. Motivation for Intrinsic Normal Prototypes (INPs). (a) Pre-stored prototypes from few-shot normal samples may fail to represent all normal patterns. (b) Pre-stored prototypes from one class can be similar to anomalies in another class. (c) The extracted INPs are concise yet well-aligned to the test image, alleviating the issues in (a) and (b)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 218, + 294, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 218, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 218, + 294, + 373 + ], + "type": "text", + "content": "ity from the training set, we propose addressing the issue of misaligned normality by leveraging the normality within the test image itself as prototypes, which we term Intrinsic Normal Prototypes (INPs). As illustrated in Fig. 1(c), normal patches within an anomalous test image can function as INPs, and anomalies can be easily detected by comparing them with these INPs. These INPs provide more concise and well-aligned prototypes to the anomalies than those learned from training data, as they typically share the same geometrical context and similar appearances with the abnormal regions within the testing image itself. Accordingly, we explore the prevalence of INPs in various AD scenarios and evaluate their potential to improve AD performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 376, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 639 + ], + "type": "text", + "content": "Although previous work [1] has attempted to utilize INP for anomaly detection, it employs handcrafted aggregated features as prototypes, thus limiting the method to zero-shot texture anomaly detection. In contrast, we introduce a learnable INP Extractor to extract normal features with adaptable shapes as INPs. We also propose an INP Coherence Loss to ensure that the extracted INPs coherently represent the normality within the test image, avoiding the capture of anomalous regions. However, some weakly representative normal regions are challenging to model with a limited set of discrete INPs, resulting in background noise (Fig. 4(c)). To address this issue, we introduce an INP-Guided Decoder, which integrates INPs into a reconstruction-based framework. This decoder leverages combinations of discrete INPs to accurately reconstruct all normal regions while effectively suppressing the reconstruction of anomalous regions, with reconstruction errors serving as anomaly scores. Furthermore, we introduce a Soft Mining Loss that focuses on normal regions that are challenging to reconstruct, i.e., hard samples, thereby improving overall reconstruction quality and enhancing AD performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Our approach, termed INP-Former, primarily leverages vision transformers (ViTs) for both INP extraction and INP-guided reconstruction. Extensive experiments on MVtec-AD [2], VisA [47], and Real-IAD [37] demonstrate that INP-Former achieves superior performance across multiclass, single-class, and few-shot AD tasks, positioning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 218, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 312 + ], + "type": "text", + "content": "INP-Former as a universal AD solution. INP-Former also optimizes computational complexity by extracting concise INPs, e.g., images can be represented effectively using only six INPs, as shown in Sec. 4.3.2. Additionally, as demonstrated in Sec. 4.4.2, INP-Former exhibits strong generalization and can even extract INPs for unseen classes, enabling zero-shot AD capabilities. In summary, our main contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 314, + 553, + 422 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 314, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 314, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 314, + 553, + 350 + ], + "type": "text", + "content": "- We demonstrate that a single image can contain Intrinsic Normal Prototypes (INPs), offering concise and aligned normality for anomaly detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 350, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 384 + ], + "type": "text", + "content": "- We propose the INP Extractor and incorporate INPs into a reconstruction-based anomaly detection framework using the INP-Guided Decoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 386, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 553, + 422 + ], + "type": "text", + "content": "- We introduce the INP Coherence Loss to extract representative INPs and the Soft Mining Loss to enhance reconstruction quality." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 434, + 405, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 405, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 405, + 445 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 454, + 474, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 474, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 474, + 467 + ], + "type": "text", + "content": "2.1. Universal Anomaly Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "type": "text", + "content": "There are numerous unsupervised AD tasks, ranging from conventional single-class AD to recent few-shot and multiclass AD setups. We refer to these collectively as universal anomaly detection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 653 + ], + "type": "text", + "content": "Single-Class Anomaly Detection: This setup was originally introduced by MVTec-AD [2] and involves developing distinct AD models for each class. Typically, images are embedded into a feature space using a pre-trained encoder, after which various schemes, such as reconstruction-based [5, 29, 44], knowledge-distillation-based [9, 36], prototype-based [34, 45], and embedding-based [26, 43] methods, are employed to learn the normality of the given class. While these approaches achieve strong performance, their reliance on class-specific models limits scalability when dealing with a wide range of classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Few-shot Anomaly Detection: In practical scenarios, the number of available normal samples may also be limited, motivating the development of few-shot AD methods. In this case, normal samples may not fully capture the variability of normality. To address this challenge, approaches" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9975" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 539, + 281 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 539, + 281 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 539, + 281 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 539, + 281 + ], + "type": "image", + "image_path": "a80e01ce54e0b4a20d117d19b6d5e6f0551585e619cc443ffd876fcfa29e1ab9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 353 + ], + "type": "text", + "content": "Figure 2. Overview of our INP-Former framework for universal anomaly detection. (a) Our model consists of a pre-trained Encoder, an INP Extractor, a Bottleneck, and an INP-Guided decoder. The INP Extractor dynamically extracts intrinsic normal prototypes from a single image, which the INP-Guided Decoder leverages to effectively suppress anomalous features. (b) Detailed architecture of the INP Extractor. (c) Detailed architecture of each layer in the INP-Guided Decoder. (d) Comparison of computational complexity between INP-Guided Attention and Self Attention. It is important to note that the patch token (Encoding) and patch token (Decoding) refer to the patch tokens utilized during the encoding and decoding stages, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 294, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 294, + 484 + ], + "type": "text", + "content": "such as spatial alignment [17] or contrastive learning [22] are used to create more compact and representative normal embeddings. Recently, Vision-Language Models (VLMs) like CLIP [33] have proven effective for few-shot AD due to their broad, pre-trained knowledge. These VLMs not only provide descriptive visual embeddings but also compute the similarity between text prompts and test images, as seen in works like WinCLIP [21], AnomalyGPT [12], and InCTRL [46]. Some approaches, such as AdaCLIP [4], even enable zero-shot AD through VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 664 + ], + "type": "text", + "content": "Multi-Class Anomaly Detection: Developing separate models for each class can be resource-intensive, prompting interest in multi-class AD, also known as unified AD [42], which aims to build a single model for multiple classes. UniAD [42] pioneered a unified reconstruction framework for anomaly detection, followed by HVQ-Trans [27] addressed the identical shortcut problem using a vector quantization framework. More recent approaches, such as MambaAD [15] and Dinomaly [14], further enhance multi-class AD performance by leveraging advanced models, i.e., the State Space Model Mamba [11] and DINO [7], respectively. However, these methods lack the functionality to derive aligned normality with the test image. On the contrary, we extract INPs from the testing image, bringing aligned and precise normality for anomaly detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 671, + 170, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 170, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 170, + 685 + ], + "type": "text", + "content": "2.2. Prototype Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Prototype learning [35] aims to extract representative prototypes from a given training set, which are then used for" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": "classification by measuring their distances to a test sample in a metric space. This technique is widely used in few-shot learning [23]. Several AD methods also employ prototype learning. For example, PatchCore [34] extracts multiple normal prototypes to represent the normality of the training data, directly computing the minimal distances to the test sample for anomaly detection. Other approaches [10, 18, 30, 32] incorporate prototypes into the reconstruction process to avoid the identical shortcut issue. Specifically, they replace the original inputs with combinations of learned normal prototypes, ensuring that the inputs to the reconstruction model contain only normal elements. However, these methods rely on pre-stored normal prototypes extracted from the training set, which can suffer from the misaligned normality problem. In contrast, our INPs are dynamically extracted from the test image, providing more aligned alternatives for normality representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 579, + 440, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 440, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 440, + 591 + ], + "type": "text", + "content": "3. Method: INP-Former" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 599, + 382, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 382, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 382, + 610 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "To fully exploit the advantages of INPs in anomaly detection, we propose INP-Former, as depicted in Fig. 2(a). The model dynamically extracts INPs from a single image and utilizes them to guide the feature reconstruction process, with the reconstruction errors serving as anomaly scores. Specifically, it comprises four key modules: a fixed pre-trained Encoder " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", an INP Extractor " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", a Bottleneck " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", and an INP-Guided Decoder " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ". The input im" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9976" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": "age " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " is first processed by the pre-trained Encoder " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " to extract multi-scale latent features " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{Q}} = \\{f_{\\mathcal{Q}}^{1}, \\ldots, f_{\\mathcal{Q}}^{L} | f_{\\mathcal{Q}}^{l} \\in \\mathbb{R}^{N \\times C}, N = \\frac{HW}{k^{2}}\\}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " represents the downsampling factor. Next, the INP Extractor " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " extracts " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " INPs " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{p_{1}, \\ldots, p_{M} | p_{m} \\in \\mathbb{R}^{C}\\}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " from the pre-trained features, with an INP coherence loss ensuring that the extracted INPs consistently represent normal features during testing. The Bottleneck " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " subsequently fuses the multi-scale latent features, producing the fused output " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "F_{\\mathcal{B}} = \\mathcal{B}(f_{\\mathcal{Q}})" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": ". Following the bottleneck, the extracted INPs are utilized to guide the Decoder " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " to yield reconstruction outputs " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}} = \\{f_{\\mathcal{D}}^{1}, \\ldots, f_{\\mathcal{D}}^{L} | f_{\\mathcal{D}}^{l} \\in \\mathbb{R}^{N \\times C}\\}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " with only normal patterns, thus the reconstruction error between " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{Q}}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 71, + 296, + 240 + ], + "type": "text", + "content": " can serve as the anomaly score." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 146, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 146, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 146, + 258 + ], + "type": "text", + "content": "3.2. INP Extractor" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "type": "text", + "content": "Existing prototype-based methods [10, 31, 34] store local normal features from the training data and compare them with test images. However, the misaligned normality between these pre-stored prototypes and the test images and the lack of global information lead to suboptimal detection performance. To address these limitations, we propose the INP Extractor to dynamically extract INPs with global information from the test image itself." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": "Specifically, as illustrated in Fig. 2(b), instead of extracting representative local features as done in PatchCore [34], we employ cross attention to aggregate the global semantic information of the pre-trained features " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathcal{Q}} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": " learnable tokens " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = \\{t_1, \\dots, t_M | t_m \\in \\mathbb{R}^C\\}" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathcal{Q}}" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": " is used as the key-value pairs, while " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": " serve as the query, allowing " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": " to linearly aggregate " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathcal{Q}}" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": " into INPs " + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{p_1, \\dots, p_M | p_m \\in \\mathbb{R}^C\\}" + }, + { + "bbox": [ + 55, + 361, + 296, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 465, + 203, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 465, + 203, + 479 + ], + "spans": [ + { + "bbox": [ + 93, + 465, + 203, + 479 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathcal {Q}} = \\operatorname {s u m} \\left(\\left\\{f _ {\\mathcal {Q}} ^ {1}, \\dots , f _ {\\mathcal {Q}} ^ {L} \\right\\}\\right)", + "image_path": "6563a4e1b20661e920204bd0de5cd7eff818281b4b4a49791cb5cd24bc7a9a7c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 481, + 294, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 481, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 93, + 481, + 294, + 499 + ], + "type": "interline_equation", + "content": "Q = \\mathbf {T} W ^ {Q}, K = \\mathbf {F} _ {\\mathcal {Q}} W ^ {K}, V = \\mathbf {F} _ {\\mathcal {Q}} W ^ {V} \\tag {1}", + "image_path": "f7b24a33f6e86957d9f6131f1cb0437d3ef67b2eb9b9bf51bd9e587877ea12cf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 497, + 221, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 497, + 221, + 510 + ], + "spans": [ + { + "bbox": [ + 93, + 497, + 221, + 510 + ], + "type": "interline_equation", + "content": "\\mathbf {T} ^ {\\prime} = \\operatorname {A t t e n t i o n} (Q, K, V) + \\mathbf {T}", + "image_path": "d2bdcc1d00f8f75ca8c1d835b78bcda5c2c48efc05191c289adb79c50bdf2d58.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 513, + 178, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 513, + 178, + 524 + ], + "spans": [ + { + "bbox": [ + 93, + 513, + 178, + 524 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\mathrm {F F N} (\\mathbf {T} ^ {\\prime}) + \\mathbf {T} ^ {\\prime}", + "image_path": "163016955236fb08df625d3937d1f87cf7ed02eb8054b98b086b3e15270855ac.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{sum}(\\cdot)" + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "text", + "content": " denotes the element-wise summation. " + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "inline_equation", + "content": "Q\\in" + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M\\times C}" + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "inline_equation", + "content": "K,V\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "text", + "content": " represent the query, key and value, respectively. " + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "inline_equation", + "content": "W^{Q},W^{k},W^{v}\\in \\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "text", + "content": " are the learnable projection parameters for " + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "inline_equation", + "content": "Q,K,V;\\mathrm{FFN}(\\cdot)" + }, + { + "bbox": [ + 55, + 533, + 296, + 592 + ], + "type": "text", + "content": " represents the feed-forward network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "text", + "content": "To ensure that INPs coherently represent normal features while minimizing the capture of anomalous features during the testing process, we propose an INP coherence loss " + }, + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "text", + "content": " to minimize the distances between individual normal features and the corresponding nearest INP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 662, + 294, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 662, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 113, + 662, + 294, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d _ {i} = \\min _ {m \\in \\{1, \\dots , M \\}} \\mathcal {S} (\\mathbf {F} _ {\\mathcal {Q}} (i), p _ {m}) \\\\ \\mathcal {L} _ {c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} d _ {i} \\tag {2} \\\\ \\end{array}", + "image_path": "7f636752f3665922d864807010e5197fde44bad2a924b301afacc27b518bbb60.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 86, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 83 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 83 + ], + "type": "text", + "content": "Table 1. Comparison of computational cost and memory usage." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 86, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 321, + 86, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 321, + 86, + 547, + 175 + ], + "type": "table", + "html": "
Number of multiplicaiton and addition
CalculationVanilla Self AttentionINP-Guided Attention
Al=Ql(Kl)T943 496 9607 220 640
fdl-1'=AlVl943 509 5046 623 232
Memory usage (MB)
Ql/Kl/Vl/Al2.30/2.30/2.30/2.342.30/0.018/0.018/0.018
", + "image_path": "3c15be7c66ea7c3406bfd2ee66728257802b3a8b70a11ac6f0a6e83a1c52a401.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "inline_equation", + "content": "S(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "text", + "content": " denotes the cosine distance. " + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "text", + "content": " represents the distance between the query feature " + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathcal{Q}}(i)" + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "text", + "content": " and the corresponding nearest INP item. Fig. 4 visually illustrates the effectiveness of " + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 183, + 553, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 237, + 436, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 436, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 436, + 249 + ], + "type": "text", + "content": "3.3. INP-Guided Decoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 255, + 555, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 541 + ], + "type": "text", + "content": "While we can use the distance between testing features and their nearest INPs for anomaly detection, as illustrated in Fig. 4(c), certain low-representative normal regions are difficult to model with a limited number of discrete INPs, leading to noisy distance maps between these INPs and testing features. To address this issue, we propose the INP-Guided Decoder, aiming to reconstruct these low-representative normal regions through a combination of multiple discrete INPs and suppress the reconstruction of anomalous regions. Additionally, this decoder provides a token-wise discrepancy that can be directly leveraged for anomaly detection. As shown in Fig. 2(c), INPs are incorporated into this decoder to guide the reconstruction process. Since INPs exclusively represent normal patterns in test images, we employ the extracted INPs as key-value pairs, ensuring that the output is a linear combination of normal INPs, thereby effectively suppressing the reconstruction of anomalous queries, i.e., the identical mapping issue [42]. Furthermore, we find that the first residual connection can directly introduce anomalous features into the subsequent reconstruction, so we remove this connection in our INP-Guided Decoder. Following the previous work [20], we also employ the ReLU activation function to mitigate the influence of weak correlations and noise on the attention maps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "text", + "content": "Mathematically, let " + }, + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}^{l - 1} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "text", + "content": " denotes the output latent features from previous decoding layer. The output " + }, + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}^{l} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "inline_equation", + "content": "l_{th}" + }, + { + "bbox": [ + 313, + 541, + 554, + 579 + ], + "type": "text", + "content": " decoding layer is formulated as," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 585, + 518, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 585, + 518, + 598 + ], + "spans": [ + { + "bbox": [ + 350, + 585, + 518, + 598 + ], + "type": "interline_equation", + "content": "Q _ {l} = f _ {\\mathcal {D}} ^ {l - 1} W _ {l} ^ {Q}, K _ {l} = \\mathbf {P} W _ {l} ^ {K}, V _ {l} = \\mathbf {P} W _ {l} ^ {V}", + "image_path": "2199f386013dcf4fa45f392969e4bd74657d768a0e0f46918eadeb9c457642c2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 350, + 602, + 553, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 602, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 350, + 602, + 553, + 617 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}} = A _ {l} V _ {l}, A _ {l} = \\operatorname {R e L U} \\left(Q _ {l} \\left(K _ {l}\\right) ^ {T}\\right) \\tag {3}", + "image_path": "f050447db882f9d0a634c1d2269d9f31a515e462522c529e1f549d34ff48091d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 620, + 460, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 620, + 460, + 635 + ], + "spans": [ + { + "bbox": [ + 350, + 620, + 460, + 635 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {D}} ^ {l} = \\operatorname {F F N} \\left(f _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}}\\right) + f _ {\\mathcal {D}} ^ {l - 1 ^ {\\prime}}", + "image_path": "044628f79965adc22e754a8d7436c6df83e782c1494649669eb48c32e62beb03.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "inline_equation", + "content": "Q_{l}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "inline_equation", + "content": "K_{l},V_{l}\\in \\mathbb{R}^{M\\times C}" + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "content": " denote the query, key and value of the " + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "inline_equation", + "content": "l_{th}" + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "content": " decoding layer. " + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "inline_equation", + "content": "W_{l}^{Q},W_{l}^{k},W_{l}^{v}\\in" + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "content": " denote the learnable projection parameters for " + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "inline_equation", + "content": "Q_{l},K_{l},V_{l}" + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "inline_equation", + "content": "A_{l}\\in \\mathbb{R}^{N\\times M}" + }, + { + "bbox": [ + 313, + 640, + 553, + 689 + ], + "type": "text", + "content": " represent the attention map." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Attention Complexity Analysis: As depicted in Fig. 2(d), the computational complexity of vanilla self-attention is" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9977" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N^2 C)" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", while its memory usage is " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N^2)" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ". In contrast, our INP-Guided Attention reduces both the computational complexity and memory usage to " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(NMC)" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(NM)" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", respectively, which can be approximated as " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(NC)" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " due to " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "M \\ll N" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ". Tab. 1 offers a detailed comparison of the complexity of vanilla self-attention and INP-Guided Attention. Appendix Sec. C compares the overall complexity between INP-Former and other methods. The light version of INP-Former can even be more efficient than MambaAD [15] yet demonstrate better performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 199, + 160, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 160, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 160, + 213 + ], + "type": "text", + "content": "3.4. Soft Mining Loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 266 + ], + "type": "text", + "content": "Inspired by Focal Loss [25], different regions should be assigned varying weights based on their optimization difficulty. Accordingly, we propose Soft Mining Loss to encourage the model to focus more on difficult regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "content": "Intuitively, the ratio of the reconstruction error of an individual normal region to the average reconstruction error of all normal regions can serve as an indicator of optimization difficulty. Given the encoder " + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "inline_equation", + "content": "f_{Q}^{l}" + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "inline_equation", + "content": "f_{D}^{l}" + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "content": " features at layer " + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "inline_equation", + "content": "M^{l}" + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "content": " denote the regional cosine distance. Our soft mining loss " + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sm}" + }, + { + "bbox": [ + 55, + 266, + 296, + 338 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 346, + 199, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 346, + 199, + 374 + ], + "spans": [ + { + "bbox": [ + 88, + 346, + 199, + 374 + ], + "type": "interline_equation", + "content": "w ^ {l} (h, w) = \\left[ \\frac {M ^ {l} (h , w)}{u (M ^ {l})} \\right] ^ {\\gamma}", + "image_path": "8895d6f6a84dec702744fd457feee198f7a60bbd95611cf14307f1ec6f7bb7f5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 376, + 295, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 376, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 88, + 376, + 295, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s m} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} 1 - \\frac {\\operatorname {v e c} \\left(f _ {\\mathcal {Q}} ^ {l}\\right) ^ {T} \\cdot \\operatorname {v e c} \\left(\\hat {f} _ {\\mathcal {D}} ^ {l}\\right)}{\\left| \\left| \\operatorname {v e c} \\left(f _ {\\mathcal {Q}} ^ {l}\\right) \\right| \\right| \\left| \\left| \\operatorname {v e c} \\left(\\hat {f} _ {\\mathcal {D}} ^ {l}\\right) \\right| \\right|} \\tag {4}", + "image_path": "5a823acbb9fa61f8fd7830ff19576de61ece076554a3f549d0d4b83245d1e6cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 411, + 226, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 411, + 226, + 426 + ], + "spans": [ + { + "bbox": [ + 88, + 411, + 226, + 426 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {\\mathcal {D}} ^ {l} (h, w) = c g \\left(f _ {\\mathcal {D}} ^ {l} (h, w)\\right) _ {w ^ {l} (h, w)}", + "image_path": "a0dbb700a3f75d74609be97afb0d1978ed8f13e775a96c56a065b3ad2f532150.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "inline_equation", + "content": "u(M^l)" + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "content": " represents the average regional cosine distance within a batch, " + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\gamma \\geq 0" + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "content": " denotes the temperature hyperparameter, " + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "inline_equation", + "content": "cg(\\cdot)_{w^{l}(h,w)}" + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "content": " denotes a gradient adjustment based on dynamic weight " + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "inline_equation", + "content": "w^{l}(h,w)" + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "inline_equation", + "content": "vec(\\cdot)" + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "content": " denotes the flattening operation. The overall training loss of our INP-Former can be expressed as follows: " + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total} = \\mathcal{L}_{sm} + \\lambda \\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 435, + 296, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 520, + 137, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 137, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 137, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 539, + 182, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 182, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 182, + 552 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": "Datasets: We conduct a comprehensive analysis of the proposed INP-Former on three widely used AD datasets: MVTec-AD [2], VisA [47], and Real-IAD [37]. MVTec-AD consists of 15 categories, with 3,629 normal images for training, and 1,982 anomalous images along with 498 normal images for testing. VisA contains 12 object categories, with 8,659 normal images for training and 962 normal images along with 1,200 anomalous images for testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "content": "Real-IAD includes 30 different objects, with 36,645 normal images for training and 63,256 normal images along with 51,329 anomalous images for testing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Metrics: Following existing works [14, 15], we use the Area Under the Receiver Operating Characteristic Curve" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "(AUROC), Average Precision (AP), and F1-score-max (F1_max) to evaluate anomaly detection and localization. For anomaly localization specifically, we use Area Under the Per-Region-Overlap (AUPRO) as an additional metric." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": "Implementation Details: INP-Former adopts ViT-Base/14 with DINO2-R [7] weights as the default pre-trained encoder. The number " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": " of INPs is set to six by default. The INP Extractor includes a standard Vision Transformer block. The layer number of the INP-Guided decoder is eight. All input images are resized to " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "448^2" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": " and then center-cropped to " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "392^2" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": ". The hyperparameters " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": " are set to 3.0 and 0.2, respectively. We use the StableAdamW [38] optimizer with a learning rate " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": " for 200 epochs. Notably, the above hyperparameters do not require any adjustment across the three datasets. Appendix Sec. A presents more implementation details. Appendix Sec. D, E, and F analyze the influence of input resolution, ViT architecture, and " + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 293, + 400, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 400, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 400, + 306 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 312, + 482, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 482, + 324 + ], + "type": "text", + "content": "4.2.1 Multi-Class Anomaly Detection" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 331, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 415 + ], + "type": "text", + "content": "We compare the proposed INP-Former with several state-of-the-art (SOTA) methods for multi-class anomaly detection, including reconstruction-based methods RD4AD [9], UniAD [42], DiAD [16], MambaAD [15], and Dinomaly [14], and embedding-based methods SimpleNet [26] and DeSTSeg [43]. A detailed introduction to the comparison methods can be found in Appendix Sec. B." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": "The experimental results on the three AD datasets are presented in Tab. 2. On the widely used MVTec-AD dataset, our method achieves SOTA performance, with image-level metrics of 99.7/99.9/99.2 and pixel-level metrics of 98.5/71.0/69.7/94.9. On VisA, our method achieves competitive results, attaining the best image-level metrics of 98.9/99.0/96.6, and achieving the best or second-best pixel-level performance of 98.8/51.2/54.7/94.4. On the more complex and challenging Real-IAD dataset, our method reaches new SOTA performance, with image-level metrics of 90.5/88.1/81.5 and pixel-level metrics of 99.0/47.5/50.3/95.0. Compared to the second-best results, our method improves by " + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.2\\uparrow /1.3\\uparrow /1.3\\uparrow" + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": " at the image level and by " + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "0.2\\uparrow /4.7\\uparrow /3.2\\uparrow /1.1\\uparrow" + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": " at the pixel level. The SOTA performance achieved across the three datasets showcases the effectiveness and robustness of our method. The per-class performance metrics are presented in Appendix Sec. H. Appendix Sec. G presents the performance of INP-Former in a more challenging scenario, which we call supermulti-class anomaly detection, thus training INP-Former on several datasets - MVTec-AD, VisA, and Real-IAD - simultaneously. Results show that our method can even detect anomalies in more classes without significant performance degradation. Fig. 3 demonstrates the precise anomaly localization capability of our method. More qualitative results" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9978" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 87, + 541, + 240 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 539, + 82 + ], + "type": "text", + "content": "Table 2. Multi-class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 87, + 541, + 240 + ], + "lines": [ + { + "bbox": [ + 67, + 87, + 541, + 240 + ], + "spans": [ + { + "bbox": [ + 67, + 87, + 541, + 240 + ], + "type": "table", + "html": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Metric →Image-level(I-AUROC/I-AP/I-F1_max)Pixel-level(P-AUROC/P-AP/P-F1_max/AUPRO)
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
RD4AD [9]94.6/96.5/95.296.1/48.6/53.8/91.192.4/92.4/89.698.1/38.0/42.6/91.882.4/79.0/73.997.3/25.0/32.7/89.6
UniAD [42]96.5/98.8/96.296.8/43.4/49.5/90.788.8/90.8/85.898.3/33.7/39.0/85.583.0/80.9/74.397.3/21.1/29.2/86.7
SimpleNet [26]95.3/98.4/95.896.9/45.9/49.7/86.587.2/87.0/81.896.8/34.7/37.8/81.457.2/53.4/61.575.7/2.8/6.5/39.0
DeSTSeg [43]89.2/95.5/91.693.1/54.3/50.9/64.888.9/89.0/85.296.1/39.6/43.4/67.482.3/79.2/73.294.6/37.9/41.7/40.6
DiAD [16]97.2/99.0/96.596.8/52.6/55.5/90.786.8/88.3/85.196.0/26.1/33.0/75.275.6/66.4/69.988.0/2.9/7.1/58.1
MambaAD [15]98.6/99.6/97.897.7/56.3/59.2/93.194.3/94.5/89.498.5/39.4/44.0/91.086.3/84.6/77.098.5/33.0/38.7/90.5
Dinomaly [14]99.6/99.8/99.098.4/69.3/69.2/94.898.7/98.9/96.298.7/53.2/55.7/94.589.3/86.8/80.298.8/42.8/47.1/93.9
INP-Former99.7/99.9/99.298.5/71.0/69.7/94.998.9/99.0/96.698.9/51.2/54.7/94.490.5/88.1/81.599.0/47.5/50.3/95.0
", + "image_path": "93a44a51a33ac0d271d7ba3ebc9232b1b1734bb5df4d808f66efcd993ec20104.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 269, + 541, + 378 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 553, + 266 + ], + "type": "text", + "content": "Table 3. Few-shot (4-shot) anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined. " + }, + { + "bbox": [ + 55, + 243, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 243, + 553, + 266 + ], + "type": "text", + "content": " indicates the results we reproduced using publicly available code." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 269, + 541, + 378 + ], + "lines": [ + { + "bbox": [ + 69, + 269, + 541, + 378 + ], + "spans": [ + { + "bbox": [ + 69, + 269, + 541, + 378 + ], + "type": "table", + "html": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓Image-levelPixel-levelImage-levelPixel-levelImage-levelPixel-level
SPADE [6]84.8/92.5/91.592.7/-/46.2/87.081.7/83.4/82.196.6/-/43.6/87.350.8†/45.8†/61.2†59.5†/0.2†/0.5†/19.2†
PaDiM [8]80.4/90.5/90.292.6/-/46.1/81.372.8/75.6/78.093.2/-/24.6/72.660.3†/53.5†/64.0†90.9†/2.1†/5.1†/67.6†
PatchCore [34]88.8/94.5/92.694.3/-/55.0/84.385.3/87.5/84.396.8/-/43.9/84.966.0†/62.2†/65.2†92.9†/9.8†/16.1†/68.6†
WinCLIP [21]95.2/97.3/94.796.2/-/59.5/89.087.3/88.8/84.297.2/-/47.0/87.673.0†/61.8†/61.0†93.8†/13.3†/21.0†/76.4†
PromptAD [24]96.6/-/-96.5/-/-/90.589.1/-/-97.4/-/-/86.259.7†/43.5†/52.9†86.9†/8.7†/16.1†/61.9†
INP-Former97.6/98.6/97.097.0/65.9/65.6/92.996.4/96.0/93.097.7/49.3/54.3/93.176.7/72.3/71.797.3/32.2/36.7/89.0
", + "image_path": "c98dd9ba501bbd8735457b61a545e3e876be3647c013a5ec863380f319b8c309.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 380, + 541, + 479 + ], + "blocks": [ + { + "bbox": [ + 70, + 380, + 541, + 479 + ], + "lines": [ + { + "bbox": [ + 70, + 380, + 541, + 479 + ], + "spans": [ + { + "bbox": [ + 70, + 380, + 541, + 479 + ], + "type": "image", + "image_path": "c3b0024b7decfb692714c19723fd49bd86b99e1b36937b7a9b9d434c3df8fd2f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 483, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 54, + 483, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 555, + 518 + ], + "type": "text", + "content": "Figure 3. Qualitative results of anomaly localization on the MVTec-AD [2], VisA [47], and Real-IAD [37] datasets for multi-class anomaly detection. The first row presents the input images with their ground truth, while the second row displays the corresponding anomaly maps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 194, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 194, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 194, + 542 + ], + "type": "text", + "content": "are presented in Appendix Sec. L." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 560, + 214, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 214, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 214, + 572 + ], + "type": "text", + "content": "4.2.2 Few-Shot Anomaly Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "type": "text", + "content": "We compare our method with several SOTA approaches for few-shot anomaly detection, including prototype-based methods SPADE [6], PaDiM [8], and PatchCore [34] and recent advances that utilize VLMs, i.e., WinCLIP [21] and PromptAD [24]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "As shown in Tab. 3, our method significantly outperforms previous SOTAs on three different AD datasets. Compared to the second-best results, our method achieves improvements of " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.0\\uparrow /1.3\\uparrow /2.3\\uparrow" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " in image-level scores and " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.5\\uparrow / - / 6.1\\uparrow /2.4\\uparrow" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " in pixel-level scores on the MVTec-AD dataset. It similarly outperforms the second-best results on" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "text", + "content": "the VisA dataset, with enhancements of " + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "inline_equation", + "content": "7.3 \\uparrow / 7.2 \\uparrow / 8.7 \\uparrow" + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "text", + "content": " for image-level and " + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "inline_equation", + "content": "0.3 \\uparrow / -7.3 \\uparrow / 5.5 \\uparrow" + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "text", + "content": " for pixel-level scores. Additionally, on the Real-IAD dataset, our method surpasses the second-best results by " + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "inline_equation", + "content": "3.7 \\uparrow / 10.1 \\uparrow / 6.5 \\uparrow" + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "text", + "content": " in image-level and " + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "inline_equation", + "content": "3.5 \\uparrow / 18.9 \\uparrow / 15.7 \\uparrow / 12.6 \\uparrow" + }, + { + "bbox": [ + 313, + 529, + 555, + 650 + ], + "type": "text", + "content": " in pixel-level scores. The superior performance of our method in few-shot anomaly detection stems from its ability to extract INPs from a single image, eliminating the need for extensive normal data to pre-store prototypes. More comparison results on 1-shot and 2-shot are presented in Appendix Sec. I." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 484, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 484, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 484, + 670 + ], + "type": "text", + "content": "4.2.3 Single-Class Anomaly Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We further compared our proposed INP-Former with current SOTA methods for single-class anomaly detection, as shown in Tab. 4. The results indicate that INP-Former" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9979" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 86, + 503, + 178 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 539, + 82 + ], + "type": "text", + "content": "Table 4. Single class anomaly detection performance on different AD datasets. The best in bold, the second-highest is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 86, + 503, + 178 + ], + "lines": [ + { + "bbox": [ + 105, + 86, + 503, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 86, + 503, + 178 + ], + "type": "table", + "html": "
Dataset →MVTec-AD [2]VisA [47]Real-IAD [37]
Method ↓I-AUROCP-APAUPROI-AUROCP-APAUPROI-AUROCP-APAUPRO
PatchCore [34]99.156.193.595.140.191.289.4-91.5
RD4AD [9]98.558.093.996.027.770.987.1-93.8
SimpleNet [26]99.654.890.096.836.388.788.5-84.6
Dinomaly [14]99.768.995.098.950.795.192.045.295.1
INP-Former99.770.295.498.549.293.892.148.195.6
", + "image_path": "dc3a92e4779bb3e4e57ed6bee5ca232c18eafc6bebfe6a70a59c4f4dea330733.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 82, + 197, + 529, + 288 + ], + "blocks": [ + { + "bbox": [ + 56, + 181, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 56, + 181, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 181, + 553, + 192 + ], + "type": "text", + "content": "Table 5. Overall ablation on MVTec-AD [2] and VisA [47] datasets. \"INP\" refers to the use of INP Extractor and INP-Guided Decoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 197, + 529, + 288 + ], + "lines": [ + { + "bbox": [ + 82, + 197, + 529, + 288 + ], + "spans": [ + { + "bbox": [ + 82, + 197, + 529, + 288 + ], + "type": "table", + "html": "
Dataset →MVTec-AD [2]VisA [47]
“INP”LcLsmImage-levelPixel-levelImage-levelPixel-level
Module×××98.59/99.18/97.6397.19/61.73/62.94/92.7396.58/97.18/92.8997.50/47.24/51.90/82.85
✓××99.53/99.80/98.8198.32/69.82/69.38/94.6998.11/98.23/95.2298.41/50.34/54.23/93.63
✓✓×99.61/99.83/99.0298.39/70.01/69.53/95.1098.16/98.30/95.4798.46/51.09/54.46/93.71
✓✓✓99.67/99.88/99.2098.48/71.02/69.65/94.8798.90/99.02/96.5798.90/51.22/54.74/94.36
", + "image_path": "27119f24e8625ddf68999d64bda5e72e82ccffe0e92298a2df78716a9b4f5ef4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 297, + 294, + 405 + ], + "blocks": [ + { + "bbox": [ + 59, + 297, + 294, + 405 + ], + "lines": [ + { + "bbox": [ + 59, + 297, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 59, + 297, + 294, + 405 + ], + "type": "image", + "image_path": "b419ae5c178a17756174ffa632fc883e6d2df30b7f6d1bd7c9d76d4464b8239b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "type": "text", + "content": "Figure 4. Visualization of the impact of INP coherence loss " + }, + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "type": "text", + "content": ". (a) Input anomalous image and ground truth. (b) Distance map without " + }, + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "type": "text", + "content": ". (c) Distance map with " + }, + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 408, + 296, + 464 + ], + "type": "text", + "content": ". The distance map is obtained by calculating the distance between the input features and their nearest INP terms, as described in Eq. 2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": "achieves new SOTA performance on the MVTec-AD and Real-IAD datasets and demonstrates competitive performance on the VisA dataset. Per-category performance of INP-Former is presented in Appendix Sec. J." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 535, + 149, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 149, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 149, + 547 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 555, + 160, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 160, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 160, + 566 + ], + "type": "text", + "content": "4.3.1 Overall Ablation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "As shown in Tab. 5, we conduct comprehensive experiments on MVTec-AD [2] and VisA [47] to validate the effectiveness of the proposed components, i.e., INP Extractor and INP-Guided Decoder (\"INP\"), INP Coherence Loss " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_c)" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", and Soft Mining Loss " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{sm})" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". In the first row, we train a baseline model without incorporating any proposed module, similar to the RD4AD [9] framework. The results in the second row demonstrate that \"INP\" significantly enhances overall performance. This improvement arises from the fact that \"INP\" introduces an information bottleneck, which effectively helps the model preserve normal features while filtering out anomalous ones. The results in the third row" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 297, + 555, + 376 + ], + "blocks": [ + { + "bbox": [ + 315, + 297, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 315, + 297, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 297, + 555, + 376 + ], + "type": "image", + "image_path": "22e7000012dce8eef883408ac061912895ee9b51e42ed37c700c83452e39baff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 424 + ], + "type": "text", + "content": "Figure 5. Visualization of the impact of soft mining loss " + }, + { + "bbox": [ + 313, + 380, + 555, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sm}" + }, + { + "bbox": [ + 313, + 380, + 555, + 424 + ], + "type": "text", + "content": ". We plot the Kernel Density Estimation (KDE) for the chewing-gum and cashew categories in the VisA [47] dataset to estimate the probability density of the anomaly scores." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": "indicate that " + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": " enhances the model's performance. This improvement stems from " + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": " ensuring that the extracted INP coherently represents normal patterns, thereby avoiding the capture of anomalous ones and establishing a solid foundation for the subsequent suppression of anomalous feature reconstruction. Fig. 4 provides a more intuitive demonstration of the effectiveness of " + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": ". The last row indicates that " + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sm}" + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": " boosts overall performance, as " + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sm}" + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": " directs the model's attention toward more challenging regions, thereby unlocking its optimal performance. Fig. 5 visually illustrates the impact of " + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sm}" + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": ", from which we can see that " + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sm}" + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": " contributes to a smaller overlap between anomaly score distributions of normal and abnormal pixels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 610, + 486, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 486, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 486, + 621 + ], + "type": "text", + "content": "4.3.2 Influence of the Number of INPs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "As shown in Fig. 6, we conduct an ablation analysis on the number " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " of INPs. The experimental results indicate that when " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " exceeds four, the model's performance stabilizes. However, if " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " becomes excessively large, the extracted INPs may also comprise information from abnormal tokens, leading to a slight decline in overall performance. In our study, we set " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " to six." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9980" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 293, + 171 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 293, + 171 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 293, + 171 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 293, + 171 + ], + "type": "image", + "image_path": "13aceb1385fb3b9f411dd582bfdd2235a7cd319df49657c79fa68ca8f86cd493.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 176, + 295, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 295, + 219 + ], + "type": "text", + "content": "Figure 6. Influence of the number of INPs " + }, + { + "bbox": [ + 55, + 176, + 295, + 219 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 176, + 295, + 219 + ], + "type": "text", + "content": " on model performance across the MVTec-AD [2] and VisA [47] datasets. Pixel-level AP and F1_max use the right vertical axis, while the other metrics share the left vertical axis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 224, + 295, + 371 + ], + "blocks": [ + { + "bbox": [ + 57, + 224, + 295, + 371 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 295, + 371 + ], + "type": "image", + "image_path": "32f97f3471c008bc4aefd2e8daa715f9bf00fd8aad342eb70d71a8d11f3da55a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "type": "text", + "content": "Figure 7. Visualizations of INPs. (a) Input anomalous image and ground truth. (b)-(g) Attention maps of six different INPs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 175, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 175, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 175, + 424 + ], + "type": "text", + "content": "4.4. Exploration on INPs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 430, + 183, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 183, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 183, + 441 + ], + "type": "text", + "content": "4.4.1 Visualizations of INPs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "text", + "content": "As shown in Fig. 7, INPs effectively capture different semantic information. Specifically, the learned INPs focus on various regions of the image, including object areas (Fig. 7(b), (e) and (f)), object edges (Fig. 7(c) and (d)), and background areas (Fig. 7(g)). This diversity is attributed to our design of guiding the reconstruction process with INPs. Additionally, INP coherence loss ensures consistency in representing normal features, allowing INPs to concentrate solely on normal regions while ignoring anomalies. This mechanism ensures the decoder reconstructs features containing only normal patterns, thereby improving anomaly detection performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 610, + 277, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 277, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 277, + 623 + ], + "type": "text", + "content": "4.4.2 Generalization capabilities of INP Extractor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "As shown in Fig. 8, the INP Extractor trained on the RealIAD [37] dataset is capable of detecting INPs on the unseen MVTec-AD [2] dataset, and the distance maps to these INPs can serve for zero-shot anomaly detection. This effectively demonstrates the INP Extractor's ability to dynamically extract INPs from a single image, with the INP coherence loss " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " ensuring that the extracted INPs coherently cap" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 553, + 138 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 138 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 138 + ], + "type": "image", + "image_path": "b4aba017b2ce12e1416aaca3d32cc9a0a13ca6473f0728486e28f1483ab5f166.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 141, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 313, + 141, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 553, + 175 + ], + "type": "text", + "content": "Figure 8. Zero-shot anomaly detection results. Here INP-Former is trained on Real-IAD [37] and tested on MVtec-AD [2]. Distance maps are visualized." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 187, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 247 + ], + "type": "text", + "content": "ture normal patterns. As shown in Appendix Sec. K, without any specific training for zero-shot anomaly detection, our method can even outperform a specified method WinCLIP [21], achieving 88.0 and 88.7 pixel-level AUROCs on MVTec-AD and VisA, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 266, + 388, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 388, + 278 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 388, + 278 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "type": "text", + "content": "We propose INP-Former, a novel method for anomaly detection that explores the role of INPs. By learning to linearly combine normal tokens into INPs and using these INPs to guide the reconstruction of normal tokens, INP-Former significantly enhances anomaly detection performance. The introduction of the INP Coherence Loss and Soft Mining Loss further refines INP quality and optimizes the training process. Extensive experiments on MVTec-AD, VisA, and Real-IAD datasets demonstrate that INP-Former achieves SOTA or comparable performance across single-class, multi-class, and few-shot anomaly detection tasks. These results validate the existence and effectiveness of INPs, which can even be extracted from images in unseen categories, enabling zero-shot anomaly detection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 459, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 639 + ], + "type": "text", + "content": "Limitations & Future Works. Our method encounters certain limitations when detecting logical anomalies that closely resemble the background distribution, such as the misplaced anomalies in the Transistor class of the MVTec-AD dataset. This issue primarily arises because the misplaced anomaly in Transistor is highly similar to the background, causing INP Extractor to incorrectly extract this anomaly as INPs. A more detailed discussion can be found in Appendix Sec. M. In future work, we plan to combine the proposed INPs with pre-stored prototypes to address this limitation. While the pre-stored prototypes encapsulate comprehensive semantic information, the INPs exhibit strong alignment. This integration is expected to significantly improve the model's ability to detect logical anomalies that are similar to the background." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 658, + 388, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 658, + 388, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 658, + 388, + 671 + ], + "type": "text", + "content": "Acknowledge" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 679, + 555, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 555, + 702 + ], + "type": "text", + "content": "This study was supported by the National Natural Science Foundation of China under Grant No. 52375494." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9981" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5553-5561, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. The MVTec anomaly detection dataset: A comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision, 129(4):1038-1059, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "text", + "content": "[3] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 301 + ], + "type": "text", + "content": "[4] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision. Springer, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "type": "text", + "content": "[5] Yunkang Cao, Haiming Yao, Wei Luo, and Weiming Shen. Varad: Lightweight high-resolution image anomaly detection via visual autoregressive modeling. IEEE Transactions on Industrial Informatics, pages 1-10, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": "[6] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "text", + "content": "[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "text", + "content": "[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 295, + 512 + ], + "type": "text", + "content": "[9] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 590 + ], + "type": "text", + "content": "[10] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "text", + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[12] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 20 24. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Hewei Guo, Liping Ren, Jingjing Fu, Yuwang Wang, Zhizheng Zhang, Cuiling Lan, Haoqian Wang, and Xinwen Hou. Template-guided hierarchical feature restoration for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6447-6458, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 150 + ], + "type": "text", + "content": "[14] Jia Guo, Shuai Lu, Weihang Zhang, Fang Chen, Hongen Liao, and Huiqi Li. Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. arXiv preprint arXiv:2405.14325, 2024. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 216 + ], + "type": "text", + "content": "[15] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "text", + "content": "[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "text", + "content": "[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "type": "text", + "content": "[18] Chao Huang, Chengliang Liu, Zheng Zhang, Zhihao Wu, Jie Wen, Qiuping Jiang, and Yong Xu. Pixel-level anomaly detection via uncertainty-aware prototypical transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 521-530, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 372, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 435 + ], + "type": "text", + "content": "[19] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 470 + ], + "type": "text", + "content": "[20] Wenli Huang, Ye Deng, Siqi Hui, Yang Wu, Sanping Zhou, and Jinjun Wang. Sparse self-attention transformer for image inpainting. Pattern Recognition, 145:109897, 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "type": "text", + "content": "[21] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19606-19616, 2023. 3, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "type": "text", + "content": "[22] Yuxin Jiang, Yunkang Cao, and Weiming Shen. Prototypical learning guided context-aware segmentation network for few-shot anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2024. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[23] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "text", + "content": "[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16838-16848, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In 2017" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9982" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[27] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36:8487-8500, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[28] Wei Luo, Haiming Yao, and Wenyong Yu. Template-based feature aggregation network for industrial anomaly detection. Engineering Applications of Artificial Intelligence, 131: 107810, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[29] Wei Luo, Haiming Yao, Wenyong Yu, and Zhengyong Li. AMI-Net: Adaptive mask inpainting network for industrial anomaly detection and localization. IEEE Transactions on Automation Science and Engineering, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "text", + "content": "[30] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "type": "text", + "content": "[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 400, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 444 + ], + "type": "text", + "content": "[32] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 511 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "type": "text", + "content": "[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 602 + ], + "type": "text", + "content": "[35] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "type": "text", + "content": "[36] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T.M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[37] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detec" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 642 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22883-22892, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[38] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[39] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Jiayi Lyu, Yong Liu, Chengjie Wang, Feng Zheng, and Yaochu Jin. IM-IAD: Industrial image anomaly detection benchmark in manufacturing. IEEE Transactions on Cybernetics, pages 1-14, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "text", + "content": "[40] Haiming Yao, Yunkang Cao, Wei Luo, Weihang Zhang, Wenyong Yu, and Weiming Shen. Prior normality prompt transformer for multiclass industrial image anomaly detection. IEEE Transactions on Industrial Informatics, 20(10): 11866-11876, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "text", + "content": "[41] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[42] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, pages 4571-4584, 2022. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "type": "text", + "content": "[43] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "type": "text", + "content": "[44] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 531 + ], + "type": "text", + "content": "[45] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 532, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 586 + ], + "type": "text", + "content": "[46] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 587, + 553, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 642 + ], + "type": "text", + "content": "[47] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9983" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_content_list.json b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db2dce1ddf7f999250420e517ee43f117891a0b6 --- /dev/null +++ b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_content_list.json @@ -0,0 +1,1663 @@ +[ + { + "type": "text", + "text": "Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation", + "text_level": 1, + "bbox": [ + 107, + 128, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuandong Liu $^{1,*}$ , Xingxing Weng $^{1,*}$ , Shuguo Jiang $^{1}$ , Pengcheng Li $^{3}$ , Lei Yu $^{2,4}$ , Gui-Song Xia $^{1,2,5,6}$", + "bbox": [ + 83, + 188, + 947, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Wuhan University", + "bbox": [ + 331, + 208, + 718, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, Wuhan University", + "bbox": [ + 318, + 224, + 728, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer Science and Technology, Hainan University", + "bbox": [ + 266, + 243, + 781, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Electronic Information, Wuhan University", + "bbox": [ + 313, + 260, + 733, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "${}^{5}$ State Key Lab. of LIESMARS, Wuhan University", + "bbox": [ + 321, + 277, + 723, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Institute for Math & AI, Wuhan University", + "bbox": [ + 346, + 296, + 694, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 348, + 326, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper explores scene affinity (AIScene), namely intrascene consistency and inter-scene correlation, for semisupervised LiDAR semantic segmentation in driving scenes. Adopting teacher-student training, AIScene employs a teacher network to generate pseudo-labeled scenes from unlabeled data, which then supervise the student network's learning. Unlike most methods that include all points in pseudo-labeled scenes for forward propagation but only pseudo-labeled points for backpropagation, AIScene removes points without pseudo-labels, ensuring consistency in both forward and backward propagation within the scene. This simple point erasure strategy effectively prevents unsupervised, semantically ambiguous points (excluded in backpropagation) from affecting the learning of pseudo-labeled points. Moreover, AIScene incorporates patch-based data augmentation, mixing multiple scenes at both scene and instance levels. Compared to existing augmentation techniques that typically perform scene-level mixing between two scenes, our method enhances the semantic diversity of labeled (or pseudo-labeled) scenes, thereby improving the semi-supervised performance of segmentation models. Experiments show that AIScene outperforms previous methods on two popular benchmarks across four settings, achieving notable improvements of $1.9\\%$ and $2.1\\%$ in the most challenging $1\\%$ labeled data. The code will be released at https://github.com/azhuantou/AIScene.", + "bbox": [ + 88, + 380, + 483, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation of LiDAR point clouds is essential for autonomous driving, allowing vehicles to perceive their 3D surroundings. Most state-of-the-art models for", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/def801321dd4c8177946d0f512f60a86825a85b1f08ced5b0721c55c9b4220f6.jpg", + "image_caption": [ + "pseudo-labeled scene" + ], + "image_footnote": [], + "bbox": [ + 519, + 349, + 704, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8668e7fd18e1f46cda9edb1914fe9c633dd144b490b322fed769740e01b97a53.jpg", + "image_caption": [ + "intra-scene inconsistency" + ], + "image_footnote": [], + "bbox": [ + 720, + 351, + 898, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e97d8aa7fb154af8d157afdc0c41046e42790ebe0dd503e8afdc2ab54d04c36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 474, + 705, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/616f6ba59fe1ade73aacc311e385c4e308b41102d64f54e5f49976764218bd9c.jpg", + "image_caption": [ + "scene-level mixing on two scenes", + "Figure 1. Top: Pseudo-labeled scenes generated from unlabeled scenes typically contain points with pseudo labels (colored) and without them (gray). Previous methods include all points in forward propagation but only pseudo-labeled points in backward propagation, leading to intra-scene inconsistency. This inconsistency may allow unsupervised, semantically ambiguous points to affect the learning of pseudo-labeled points (see Fig. 7). Bottom: Existing data augmentation techniques suitable for semi-supervised segmentation usually mix two scenes through scene-level operations like concatenation or swapping. Their resulting scenes may be constrained in terms of semantic diversity, potentially lacking objects like bicyclists and buses, which are important in driving scenes." + ], + "image_footnote": [], + "bbox": [ + 521, + 527, + 705, + 595 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2220de7a7303f910a842a4fb07e4cf2efb164c978fd7339f75293a20955be67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 474, + 900, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8963cb7d9482a7ce4dfa3d883a98e5d972939e035d312282a8017cde7fcee2b.jpg", + "image_caption": [ + "limited semantic diversity" + ], + "image_footnote": [], + "bbox": [ + 720, + 517, + 900, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "driving-scene LiDAR segmentation are dependent on high-quality, point-annotated datasets [12, 36, 57]. With numerous objects in driving scenes, densely annotating points is extremely labor-intensive. Consequently, semi-supervised LiDAR semantic segmentation has attracted growing atten", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ${}^{ \\ominus }$ Corresponding authors.", + "bbox": [ + 114, + 885, + 367, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27380", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion, where models are trained on a limited amount of labeled data supplemented by a large set of unlabeled data.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The common solution for learning from unlabeled data is to apply teacher-student training, where a teacher network generates pseudo-labeled scenes from unlabeled data, which are then used to supervise the learning of a student network. The student network's objective is to minimize segmentation loss while the teacher network's weights are updated via exponential moving average [39]. To further improve model learning with unlabeled data, existing studies implement designs such as utilizing language-vision models [22], adopting density-guided contrastive learning [21], or trying feature fusion between LiDAR and image data [5]. Despite the progress, two main issues remain to be addressed:", + "bbox": [ + 88, + 123, + 480, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) Given the teacher network's confident prediction, pseudo-labeled scenes typically contain points with and without pseudo-labels. During training, most studies include all points in pseudo-labeled scenes in the student network's forward propagation, but only pseudo-labeled points are involved in backward propagation. Points excluded from backpropagation are unsupervised and, therefore, semantically ambiguous. However, these points still contribute to the forward feature computation for pseudolabeled points, which may interfere with the model's ability to learn from them effectively. In short, the inconsistency between forward and backward propagation within the scene causes ambiguous points to obstruct the model's learning of pseudo-labeled points. Both point-based [12] and voxel-based [7, 57] methods suffer from this issue in feature computation, as illustrated at the top of Fig. 1.", + "bbox": [ + 91, + 321, + 482, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(2) Data augmentation is an effective practice in semi-supervised LiDAR semantic segmentation [13, 15]. The basic idea is to mix raw labeled scenes and pseudo-labeled scenes to enrich scene semantics. Although impressive advancements have been made [33, 52], they usually select two scenes (one labeled scene and one pseudo-label scene) and then perform scene-level mix through operations like concatenation [27] and swapping [15], which limits the semantic diversity of mixed scenes. As shown at the bottom of Fig. 1, multiple scenes, whether labeled or pseudo-labeled, often share similar spatial layouts. Moreover, in driving scenes, objects such as cars and pedestrians are the primary focus of an autonomous vehicle's attention. This naturally raises the question: can we improve the semantic diversity of scenes by mixing multiple scenes at the instance level?", + "bbox": [ + 89, + 566, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address inconsistency and implement the multi-scene mixing idea, we explore Scene Affinity for semi-supervised LiDAR semantic segmentation, named AIScene. The term scene affinity refers to intra-scene consistency and interscene correlation. Specifically, AIScene, built on the teacher-student training pipeline, employs a simple yet helpful point erasure strategy, removing points without", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pseudo-labels in forward propagation. This ensures intra-scene consistency in both forward and backward propagation, preventing semantically ambiguous points from disrupting learning from pseudo-labeled points. Building on this, AIScene incorporates patch-based data augmentation that mixes multiple scenes at both scene and instance levels. By maintaining two pools that store scene and instance patches from various scenes, it selects patches for mixing based on inter-scene correlation. This novel augmentation method not only enhances the semantic diversity of raw labeled scenes but also complements pseudo-labeled scenes where some points have been erased.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AIScene has been extensively evaluated through experiments and demonstrated superiority on two popular benchmarks for semi-supervised LiDAR semantic segmentation. It exceeds state-of-the-art methods by $1.3\\%$ on the SemanticKITTI [2] and $1.2\\%$ on the nuScenes [10] in terms of average performance across four settings (1%, 10%, 20%, and 50% labeled data). Notably, in the most challenging setting, with only $1\\%$ labeled data, AIScene achieves remarkable improvements of $1.9\\%$ and $2.1\\%$ on the two datasets. Additionally, our ablation studies confirm AIScene's potential to reduce the requirements for labeled data across different LiDAR semantic segmentation networks (see Tab. 1 and Tab. 2), with each component of AIScene able to function independently to facilitate semi-supervised LiDAR semantic segmentation (see Fig. 6).", + "bbox": [ + 511, + 272, + 903, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 532, + 500, + 846, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a point erasure strategy to ensure intra-scene consistency and facilitate network learning from pseudolabeled scenes. This strategy can serve as a plug-and-play tool for semi-supervised LiDAR semantic segmentation using the pseudo-label mechanism.", + "- We design a patch-based data augmentation to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method improves the semantics diversity of labeled scenes and complements pseudolabeled scenes.", + "- Extensive experiments demonstrate that AIScene achieves state-of-the-art results on popular benchmarks for semi-supervised LiDAR semantic segmentation. Ablation studies show that the novel erasure strategy and augmentation method can function independently." + ], + "bbox": [ + 511, + 513, + 903, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 753, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully-supervised LiDAR Semantic Segmentation: Numerous approaches for LiDAR semantic segmentation have been explored, which can be broadly grouped into three categories: voxel-based, 2D projection-based, and point-based methods. Voxel-based approaches [6, 7, 19, 37, 57] involve voxelizing sparse point clouds to generate regular voxel grids, allowing the application of standard 3D convolution neural networks for embedding learning. Conversely,", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27381", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c2265fe8a86ad5d13f4b41e40929e4b18f7188fd5d5ec75a9d88c507f5d1f2f2.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed AIScene training pipeline. The teacher network generates the pseudo-labels $y_{i}^{u}$ from the unlabeled scene $x_{i}^{u}$ , followed by an erasure operation to filter out low-confidence points, ensuring training consistency. For the labeled scene $x_{i}^{l}$ and the generated erased scene $\\hat{x}_i^u$ , we employ patch-based data augmentation to obtain augmented scenes $x_{i}^{lpf}$ and $\\hat{x}_i^{upf}$ , respectively. Subsequently, both scenes are fed into the student network to compute the loss of $\\mathcal{L}_l$ and $\\mathcal{L}_u$ ." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 893, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D projection-based methods [14, 26, 45, 48, 53, 55] project input point clouds into a 2D dense space, use 2D convolutional filters [31] for processing, and then map predictions back to the original 3D space. Finally, point-based methods [12, 18, 29, 46] operate directly on raw point clouds, enabling the extraction of point-wise semantic features.", + "bbox": [ + 88, + 470, + 480, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised LiDAR Semantic Segmentation: To reduce the need for 3D dense annotations, recent research has explored methods that require less supervision [3, 17, 24, 38, 40, 54]. Semi-supervised learning [4, 16, 21, 23, 30, 58] aims to utilize a large number of unlabeled data to alleviate the annotation burden. LaserMix [15], a pioneer work in semi-supervised LiDAR segmentation, leverages a simple laser beam mix strategy to achieve solid results. Subsequent research fuses LiDAR and camera data to extract deeper semantic information [5, 41]. DDsemi [21], focusing on point cloud density, builds on contrastive learning to enhance feature discriminability. The recently proposed LASS3D [22] leverages large language-vision models to support semi-supervised segmentation. Existing semi-supervised methods often build upon the mean-teacher paradigm [39] and utilize pseudo labels [1]. However, many points without pseudo-labels contribute equally to pseudo-labeled points in forward propagation but are ignored during backward propagation, leading to intra-scene inconsistency. To address this, our AIScene leverages a point erasure strategy to mitigate this inconsistency, complementing the pseudolabel mechanism.", + "bbox": [ + 89, + 568, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Point Cloud Data Augmentation: Point cloud data augmentation has been explored across various 3D vision tasks [9, 25, 34, 43, 44, 56]. Inspired by the success of Mixup [51] in image classification, several methods have investigated techniques for mixing selected point cloud samples. To preserve structural information, RSMix [20] introduces a neighboring function, and PointCutMix [52] establishes correspondences for sample fusion. Notably, these methods focus on local object-level augmentation. In response, several works have explored mixing strategies for point cloud scene augmentation. PloarMix [47] swaps points within the fan-shaped areas between two scenes, while CoSMix [32] aims to mitigate domain shift by mixing semantic point groups. Unlike these methods, our method is capable of modeling inter-scene information across multiple scenes to create more diversified data, while also extracting instance-level semantic information for individual classes to further enrich scenes.", + "bbox": [ + 511, + 470, + 908, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 756, + 607, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 781, + 643, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Before we describe our model and training process, we first provide the definition of semi-supervised point cloud segmentation. In detail, we have a small labeled training set $D_{l} = \\{x_{i}^{l},y_{i}^{l}\\}_{i = 1}^{N_{l}}$ and a large unlabeled training set $D_{u} = \\{x_{i}^{u}\\}_{i = 1}^{N_{u}}$ , where $x_{i}\\in \\mathbb{R}^{m\\times (3 + r)}$ is a point cloud scene of $m$ points with three-dimensional", + "bbox": [ + 511, + 803, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27382", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cartesian coordinates relative to the sensor and additional $r$ -dimensional feature (e.g., intensity, color, or reflectivity), $y_{i}$ represents its corresponding point-level segmentation labels, $N_{l}$ and $N_{u}$ are the numbers of labeled and unlabeled point cloud scenes, respectively, and $N_{u} \\gg N_{l}$ in ideal cases. The goal of SSL-based point cloud segmentation is to learn a segmentation model using limited labeled data and large-scale unlabeled data.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Teacher-Student Paradigm. Following current mainstream research [15, 21], our framework also builds upon the teacher-student paradigm, which includes two segmentation networks with the same architecture. Our AIScene can be instantiated with any 3D segmentation network [7, 19, 57], denoted as $\\Phi(\\cdot)$ . In detail, the teacher model generates pseudo-labels for unlabeled scenes and undergoes a gradual update through the exponential moving average (EMA) [39] of the student weights:", + "bbox": [ + 89, + 212, + 483, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathbb {t}} ^ {k + 1} = \\alpha \\cdot \\theta_ {\\mathbb {t}} ^ {k} + (1 - \\alpha) \\cdot \\theta_ {\\mathbb {S}} ^ {k}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 358, + 482, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is the EMA decay rate, which is set to 0.99 by default, $\\theta_{\\mathbb{t}}$ and $\\theta_{\\mathrm{s}}$ represent the weight parameters of the teacher and student models, respectively, and $k$ denotes the training step.", + "bbox": [ + 89, + 388, + 483, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Point Erasure Strategy", + "text_level": 1, + "bbox": [ + 89, + 459, + 303, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the teacher-student mutual learning framework [35], the overall architecture of AIScene is depicted in Fig. 2. Specifically, AIScene consists of two primary branches: one dedicated to labeled scenes and the other to unlabeled scenes. How to effectively leverage a large number of unlabeled scenes is crucial for semi-supervised learning. Hence, for the branch of unlabeled scenes, we use a teacher network to generate pseudo-labeled scenes $(x_{i}^{u},y_{i}^{u})$ based on a predefined confidence threshold $\\tau_{s}$ [15]. As mentioned previously, when using the pseudo-labeled scene for training, points without pseudo-labels are involved in the forward propagation but are ignored in the backward propagation due to the lack of corresponding supervision. This leads to intra-scene inconsistency, which causes semantic ambiguity of points. To address this issue, we propose erasing points without pseudo labels, resulting in an erased scene $(\\hat{x}_i^u,\\hat{y}_i^u)$ that contains fully annotated information, similar to fully supervised scenes. Formally, intrascene inconsistency addressed by the point erasure strategy can be defined as:", + "bbox": [ + 89, + 481, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {i} ^ {u} = \\left\\{x _ {i} ^ {u} \\mid \\Phi_ {s} \\left(x _ {i} ^ {u}\\right) \\geq \\tau_ {s} \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 795, + 482, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{x}_i^u$ is the erased scene with corresponding class labels $\\hat{y}_i^u$ , $\\Phi_s(\\cdot)$ represents the segmentation network that predicts the confidence score for each point in the scene $x_i^u$ , and $\\tau_s$ is a predefined confidence threshold, equivalent to the pseudolabel mining threshold, requiring no additional adjustment.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88e87573d075d5a475b65e720e6bfd27db4be3c86aa6cd6dd8071821c3aa6b1c.jpg", + "image_caption": [ + "Figure 3. Illustration of the pool initialization process. For each labeled point cloud scene, splitting it into multiple patches from the BEV perspective and assigning different instance categories to each patch enables the extraction of instances belonging to different classes. Subsequently, these patches and instances are stored individually in the patch pool and instance pool." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 898, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the proposed point erasure strategy is independent of the segmentation network architecture, it can be utilized as a plug-and-play tool for semi-supervised 3D semantic segmentation methods based on the pseudo-label mechanism, as depicted in Fig. 6 (a).", + "bbox": [ + 511, + 328, + 906, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Patch-based Data Augmentation", + "text_level": 1, + "bbox": [ + 511, + 412, + 799, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the labeled and erased scenes, we propose a patch-based data augmentation to enhance the semantic diversity of annotation information, guided by inter-scene correlation. The key idea is to construct patch pools to store scene-level and instance-level information, and then combine patches from multiple scenes to generate diverse training samples.", + "bbox": [ + 511, + 433, + 905, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pool Initialization. We first use the labeled training set $D_{l}$ to create two patch pools, which store scene-level and instance-level information, respectively, as shown in Fig. 3. Given the $i$ -th point cloud scene $x_{i}^{l}$ , we patchify $x_{i}^{l}$ into a sequence of $T$ regular, non-overlapping point patches $\\{p_{ij}^{l}\\}_{j=1}^{T}$ , each of size $(n \\times n)$ , in the bird's eye view. This process can be expressed as:", + "bbox": [ + 511, + 539, + 906, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{p _ {i j} ^ {l} \\right\\} _ {j = 1} ^ {T} = \\text {p a t c h i f y} \\left(x _ {i} ^ {l}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 654, + 906, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_{ij}^{l}$ denotes the $j$ -th point patch from the sequence. For all labeled scenes, by performing the operation in Eq. (3), the patch pool $\\mathcal{B}_{\\mathrm{Pat}}^{l}$ is constructed, which can be defined as:", + "bbox": [ + 511, + 689, + 905, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {\\mathrm {P a t}} ^ {l} = \\left\\{\\left\\{p _ {i j} ^ {l} \\right\\} _ {j = 1} ^ {T} \\right\\} _ {i = 1} ^ {N _ {l}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 756, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the patch pool, we can obtain the instance pool $\\mathcal{B}_{\\mathrm{Ins}}^l$ by assigning different class labels to each patch, as depicted in Fig. 3, which contains instance information for different categories (e.g., car and person).", + "bbox": [ + 511, + 794, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the same steps used for labeled scenes, we further create a pseudo patch pool $\\mathcal{B}_{\\mathrm{Pat}}^u$ and a pseudo instance pool $\\mathcal{B}_{\\mathrm{Ins}}^u$ from the erased scene $\\hat{x}_i^u$ . Unlike $\\mathcal{B}_{\\mathrm{Pat}}^l$ and $\\mathcal{B}_{\\mathrm{Ins}}^l$", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/14ae380cf8b360e30cec537917399150684b8664dff76e0c32c3e3c4f041f302.jpg", + "image_caption": [ + "Figure 4. (a) Illustration of the MixPatch process. (b) Illustration of selecting instances from the patch pool for filling. An instance will not be filled if there are no other points around the filling location or if the filled instance overlaps with existing instances." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 264, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b15d4aa449a510f82e7a03474a437a6dbc82c59f1749a8ed5fbd0a2bef996c45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 93, + 480, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "covering the whole training set $D_{l}$ , the pseudo pools $\\mathcal{B}_{\\mathrm{Pat}}^{u}$ and $\\mathcal{B}_{\\mathrm{Ins}}^{u}$ only contain scene and instance information from the current batch and are reset in each iteration. This prevents low-quality pseudo labels from persisting in pools and misleading model training. Additionally, if the number of points within a patch or instance falls below a certain threshold $\\tau_{\\mathrm{min}}$ , it will not be saved into the pool. In practice, we set $\\tau_{\\mathrm{min}}$ to 5 [49]. If a patch or instance has fewer than 5 points, its spatial structure exhibits high uncertainty, making it semantically ambiguous. Thus, it is excluded from the pools to prevent noise during data augmentation.", + "bbox": [ + 88, + 344, + 482, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MixPatch. To model the inter-scene correlation and then guide the transfer of annotation information between labeled scenes and unlabeled scenes, we combine multiple scenes using patch pools. Fig. 4 (a) illustrates the patch mixing (MixPatch) strategy, while Fig. 4 (b) presents the instance filling (InsFill) strategy, which is introduced in the next paragraph. During a training iteration, the mini-batch is comprised of labeled and unlabeled scenes. After the unlabeled scene $x_{i}^{u}$ is transformed into the erased scene $\\hat{x}_{i}^{u}$ using the point erasure strategy, we integrate information from the patch pool into $\\hat{x}_{i}^{u}$ by MixPatch. The MixPatch process can be described as follows:", + "bbox": [ + 89, + 511, + 482, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {i} ^ {u p} = \\left(\\hat {x} _ {i} ^ {u} \\odot \\mathcal {M}\\right) \\oplus \\left(\\mathcal {B} _ {\\text {P a t}} ^ {l} \\odot (1 - \\mathcal {M})\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 700, + 480, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is a binary mask indicating which patches from $\\hat{x}_i^u$ will be replaced with selected patches of the same patch index from the patch pool, $\\odot$ denotes element-wise multiplication, and $\\hat{x}_i^{up}$ represents the augmented scene. For labeled scenes, MixPatch is applied based on the current pseudo patch pool $\\mathcal{B}_{\\mathrm{Pat}}^u$ , which can be expressed as the following:", + "bbox": [ + 89, + 728, + 482, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} ^ {l p} = \\left(x _ {i} ^ {l} \\odot (1 - \\mathcal {M})\\right) \\oplus \\left(\\mathcal {B} _ {\\text {P a t}} ^ {u} \\odot \\mathcal {M}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 842, + 482, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is the same as defined in Eq. (5), thereby ensuring the full utilization of scene information from the unlabeled", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scenes. The binary mask $\\mathcal{M}$ and its complement $1 - \\mathcal{M}$ are used to merge inter-scene point cloud patches with the erased scene, facilitating the generation of diverse semantic information for labeled and unlabeled scenes. To alleviate bias in the patch mixing process, we apply a uniform sampling strategy to the patches in each scene, ensuring that each patch has an equal opportunity of being selected in the mixed scene.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "InsFill. In addition to patch mixing, we utilize the instances stored in the instance pool $\\mathcal{B}_{\\mathrm{Ins}}^l$ to further enhance the scene $\\hat{x}_i^{up}$ . Given a patch with location index $j$ in the scene, we randomly select an instance from the instance pool $\\mathcal{B}_{\\mathrm{Ins}}^l$ that was extracted from the $j$ -th patch. Patches with identical indices typically correspond to similar Li-DAR spatial layouts, ensuring consistency between the selected instance and its fill positions in the scene. As depicted in Fig. 4 (b), if the selected instance overlaps with existing instances in the scene or lacks sufficient contextual information, the filling operation is bypassed. This judgment process can be easily implemented using the maximum 3D bounding boxes. This data augmentation provides instance-level semantic information for the point cloud scene, enabling the augmented scene $\\hat{x}_i^{upf}$ and corresponding labels $\\hat{y}_i^{upf}$ to effectively support semi-supervised training. For the labeled branch, we also apply the same filling strategy as described, selecting instances from the pseudo instance pool $\\mathcal{B}_{\\mathrm{Ins}}^u$ to fill the current scene. This InsFill process yields the scene $x_i^{lpf}$ along with its corresponding label $y_i^{lpf}$ .", + "bbox": [ + 511, + 210, + 906, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Training", + "text_level": 1, + "bbox": [ + 511, + 523, + 669, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the standard teacher-student mutual learning paradigm [15, 35, 39], we construct a mini-batch of data with an equal composition of labeled and unlabeled scenes at each iteration. The unlabeled scenes are assigned with pseudo-labels by the teacher model and then train the student model jointly with labeled scenes. We use three types of loss functions to supervise the model training here, which is shown as follows:", + "bbox": [ + 511, + 547, + 905, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\mathcal {L} _ {\\text {s e g}} \\left(\\Phi_ {\\theta_ {\\mathfrak {s}}} \\left(x _ {i} ^ {l}\\right), y _ {i} ^ {l}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 675, + 903, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l} = \\mathcal {L} _ {\\text {s e g}} \\left(\\Phi_ {\\theta_ {\\mathrm {s}}} \\left(x _ {i} ^ {l p f}\\right), y _ {i} ^ {l p f}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 696, + 903, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\mathcal {L} _ {s e g} \\left(\\Phi_ {\\theta_ {\\mathrm {s}}} \\left(\\hat {x} _ {i} ^ {u p f}\\right), \\hat {y} _ {i} ^ {u p f}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 718, + 903, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_s, \\mathcal{L}_l$ , and $\\mathcal{L}_u$ denote the segmentation loss for the labeled scene $x_i^l$ , augmented labeled scene $x_i^{lpf}$ , and augmented unlabeled scene $\\hat{x}_i^{upf}$ , respectively, while $\\mathcal{L}_{seg}$ is the original loss of the segmentation network $\\Phi(\\cdot)$ . We sum up all the loss term to define the final training objective as:", + "bbox": [ + 511, + 744, + 905, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda_ {u} \\mathcal {L} _ {u} + \\lambda_ {l} \\mathcal {L} _ {l}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 845, + 903, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{u}$ and $\\lambda_{l}$ are loss weights that balance the two loss terms. After that, we update the teacher model weights us", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27384", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d24a81c8896292f6f5e3524af962285ea6d100d54a3a4e1ce4b58a8480b964ef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueSemanticKITTI [2]Avg.nuScenes [10]Avg.
1%10%20%50%1%10%20%50%
Meanteacher [39]NeurIPS'201745.457.159.260.055.451.666.067.171.764.1
CBST [58]ECCV'201848.858.359.459.756.653.066.569.671.665.2
CPS [4]CVPR'202146.758.759.660.556.452.966.370.072.565.4
LaserMix (Range View) [15]CVPR'202343.458.859.461.455.849.568.270.673.065.3
LaserMix (Voxel) [15]CVPR'202350.660.061.962.358.755.369.971.873.267.6
GPC [13]ICCV'202154.162.062.562.860.4-----
IGNet [41]WACV'202449.061.363.164.859.6-----
LiM3D [23]CVPR'202358.462.263.163.661.8-----
LASS3D [22]ECCV'202458.563.064.164.562.5-----
DDSemi [21]CVPR'202459.365.166.367.064.458.170.274.076.569.7
Ours (AIScene)-61.266.367.467.965.760.272.375.076.470.9
Improvement ↑-+1.9+1.2+1.1+0.9+1.3+2.1+2.1+1.0-0.1+1.2
", + "bbox": [ + 102, + 88, + 895, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Experimental results on the SemanticKITTI and nuScenes dataset compared with different SSL methods. The results are reported with mIoU scores (\\%). The bold and underline scores respectively denote the best and second results.", + "bbox": [ + 89, + 301, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing the EMA strategy described in Eq. (1).", + "bbox": [ + 89, + 342, + 372, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 369, + 223, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 393, + 375, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SemanticKITTI. Following state-of-the-art methods [15, 21], we evaluate our AISCene on the SemanticKITTI dataset [2], which is a large-scale segmentation dataset consisting of LiDAR data derived from the popular KITTI dataset [11]. It is composed of 22 sequences, with sequences 00-10 serving as the train set (of which 08 as the val set) and sequences 11-21 as the test set. Moreover, all the point cloud scenes are collected from real-world rural and urban road environments in Germany, and the semantic class labels are mapped to 19 classes in our experiments.", + "bbox": [ + 89, + 416, + 482, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nuScenes. The nuScenes dataset [10] is another well-known large autonomous driving dataset. This dataset is composed of 1000 sequences of 20s duration (850 for training and validation, 150 for testing), which are captured by the LiDAR sensor from Boston and Singapore. Following the official splitting protocol, the train and val sets contain 28,130 and 6,019 point cloud scenes, respectively, including per-points annotation for 16 semantic classes.", + "bbox": [ + 89, + 566, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For both the SemanticKITTI and nuScenes datasets, we sample different ratios of annotated train scenes to obtain semi-supervised data. Meanwhile, we report the performance on the validation set using the mean Intersection over Union (mIoU) metric [21, 23].", + "bbox": [ + 89, + 686, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 772, + 307, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments with $4 \\times$ NVIDIA V100 GPUs with 32GB RAM, based on MMDetection3D [8]. For the threshold $\\tau_{s}$ for pseudo-label generation and point erasure, we experimentally find that 0.9 yields relatively satisfactory results. We use the MinkowskiNet [7] and Cylinder3D [57] as the backbone network due to its widely adopted in previous methods [15, 28, 42]. To facilitate patch-based data", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "augmentation, we restrict our focus to point clouds within the X and Y dimensions in the range of $[-50, 50]m$ for both the SemanticKITTI and nuScenes datasets. Similar to mainstream work, we also adopt the consistency loss proposed in MeanTeacher [39] to align the predictions of the teacher and student networks. Additionally, the weights of the consistency loss are assigned as $1 \\times 10^{3}$ and $2 \\times 10^{3}$ for the SemanticKITTI and nuScenes datasets, respectively. The weights for $\\lambda_{u}$ and $\\lambda_{l}$ are set as 1 for both datasets. Our method excludes all forms of model ensemble and test-time augmentation during the evaluation.", + "bbox": [ + 511, + 342, + 906, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 516, + 651, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we compare our method with current SOTA semi-supervised point cloud segmentation methods (e.g., LaserMix [15], LiM3D [23] and DDsemi [21]), using sub-sampling of the datasets under various partition protocols [21]. Specifically, we split the entire training set into labeled subsets with ratios of $1\\%$ , $10\\%$ , $20\\%$ , and $50\\%$ in the SemanticKITTI [2] and nuScenes dataset [10]. For further comparison, we also include several classical semi-supervised learning methods [4, 39, 58] from the 2D image domain adapted to 3D segmentation.", + "bbox": [ + 511, + 537, + 906, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 summarizes the quantitative results on SemanticKITTI and nuScenes datasets. Our method consistently outperforms other comparative methods on both datasets, achieving average mIoU improvements of $1.3\\%$ and $1.2\\%$ , respectively. Notably, the proposed method demonstrates even greater gains on smaller sampling splits, such as $1\\%$ and $10\\%$ , which are particularly beneficial for the semi-supervised setting. Specifically, with a $1\\%$ ratio, our method shows mIoU improvements of up to $1.9\\%$ and $2.1\\%$ on the two datasets, respectively, compared to the recently proposed SOTA method, DDsemi [21]. These results also validate that our method is less sensitive to reduced labeled data sampling than other methods. Furthermore, we present qualitative results in Fig. 5. As indicated by the", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/effddf6ca379569f775d87cef2e34128b867ab0cf25ad621a9b973ae7717b936.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 220, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5f8f271224715696ef40a16a8b1fe3119772ca7c6b7e19b5cc27828c29e940c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 89, + 349, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24b274b44b46e6df8d1f622a281a78da544e9b0ec4a3cfd6fd3bbee7df68a8ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 89, + 477, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f30d5b9e6d29c1e0dad9788e69bbb7aa3490ff9deefa293c2664c459170ddedf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 214, + 220, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c6c6bba31f1e1a3e6f91667349620a7dff637e5e9ad7a6479bf06785397dae9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 214, + 349, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4ad28930e5c24a1c3a7be72f42e918babf091b018fe4ab6e23e22175728af48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 214, + 477, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6b546960433769d186988177e8b8bbd90eab545899925b2830433ae76184f8a.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison results from LiDAR bird's eye view on SemanticKITTI val set. We set the point with correct and incorrect predicted semantic class in blue and orange for better visualization. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 96, + 321, + 220, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2dc2d0e117ed974f7ad40beb905ac71fb8eca19aa2e2e50237c8d326c9574d5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 321, + 349, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6235198668efd37c506bd210494a39694a506d80792564ea57f657f35248e3e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 321, + 477, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "purple and red dashed circles, our method achieves better segmentation of boundary regions and small objects.", + "bbox": [ + 89, + 512, + 482, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 556, + 243, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, We first validate the adaptability of our method to different segmentation models. We then examine the impact of intra-scene consistency through the point erasure strategy and explore inter-scene correlation by incorporating patch-based mixing and patch-based instance filling data augmentation.", + "bbox": [ + 89, + 580, + 482, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptability with Different Segmentation Methods. Our AISCene is segmentor-agnostic and can be easily used to further improve off-the-shelf 3D fully-supervised segmentors under the semi-supervised setting. Tab. 2 shows that the proposed method can substantially boost the performance with Cylinder3D [57] as the backbone, indicating the generalization ability of our framework. For consistency in comparison, the subsequent experiments are based on Cylinder3D unless otherwise specified.", + "bbox": [ + 89, + 672, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Point Erasure Strategy. The $1^{\\text{st}}$ row in Tab. 3 reports the performance of the supervision-only Cylinder3D model [57], which serves as the baseline. When PtErase is applied to the baseline, it effectively boosts performance by ensuring consistency during forward and backward propagation, thus addressing the lim", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e2ce4b73cc4d3c6cf2e6da6ef372c139f6655510e61949881c8047c18c5fbe2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMethods1%10%20%50%Avg.
SemanticKITTICylinder3D [57]45.456.157.858.754.5
Ours (AIScene)54.563.363.764.361.5
Improvement+9.1+7.2+5.9+5.6+7.0
nuScenesCylinder3D [57]50.965.966.671.263.7
Ours (AIScene)56.670.272.873.968.4
Improvement+5.7+4.3+6.2+2.7+4.7
", + "bbox": [ + 517, + 88, + 905, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58576494990a01372e6985937fdfbdf161c180622ea3fd9ec75fec151e0c5961.jpg", + "table_caption": [ + "Table 2. Ablation study on the adaptability of our AIScene using Cylinder3D as the segmentation model." + ], + "table_footnote": [], + "table_body": "
Ours AISCeneptEraseMixPatchInsFillSplits on SemanticKITTIAvg.
1%10%20%50%
45.456.157.858.754.5
✓51.759.861.860.458.4
✓51.361.460.760.558.5
✓52.361.760.161.158.8
✓✓54.661.062.962.860.3
✓✓52.061.760.862.659.3
✓✓50.562.563.863.660.1
✓✓✓54.563.363.764.361.5
", + "bbox": [ + 517, + 229, + 905, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study of different components of our AIScene framework. PtErase refers to the point erasure strategy, while MixPatch and InsFill refer to the patch mix and instance filling data augmentation, respectively.", + "bbox": [ + 511, + 372, + 905, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ditions of the pseudo-label mechanism. This consistency is achieved through a plug-and-play point erasure strategy, which can be easily integrated into methods relying on the pseudo-label mechanism. To further validate the importance of maintaining consistency during forward and backward propagation, we embedded the PtErase strategy into LaserMix [15]. As illustrated in Fig. 6 (a), using the PtErase strategy improves performance across all ratios, underscoring the significance of intra-scene consistency and demonstrating the generalizability of our method. Moreover, Fig. 6 (b) shows that the proportion of points removed by the erasure strategy is initially high at the start of training, but gradually decreases as the network converges. Furthermore, We present a comparison of class probability distributions in Fig. 7 to validate the effectiveness of the point erasure strategy. The results demonstrate that this strategy reduces classification ambiguity, enabling the model to better capture the features of the correct class.", + "bbox": [ + 511, + 445, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Patch-based Data Augmentation. The patch-based data augmentation consists of MixPatch (patch-based mixing) and InsFill (instance filling). In Tab. 3, the results in the $3^{\\mathrm{rd}}$ and $4^{\\mathrm{th}}$ rows indicate that using MixPatch or InsFill individually yields an average mIoU improvement of $4.0\\%$ or $4.3\\%$ , respectively, over the baseline. This suggests that incorporating diverse semantic information across scenes benefits semi-supervised LiDAR segmentation. Furthermore, as shown in the $5^{\\mathrm{th}}$ and $6^{\\mathrm{th}}$ rows, combining MixPatch or InsFill with PtErase significantly enhances performance, highlighting the effectiveness of patch-based data augmentation. The combination of all", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27386", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/443f65983496f89a861c533cedc6247d0287a0a339650b8586a44767b443dbec.jpg", + "image_caption": [ + "Figure 6. (a) Illustration of performance differences whether using point erasure strategy on the LaserMix [15]. (b) Variation in the proportion of points removed by the point erasure strategy throughout the training process. (c) Comparison of different mixing-based techniques on SemanticKITTI dataset. (d) Results of mIoU against the number of patch split $n$ under the $10\\%$ SemanticKITTI split." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 903, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57f2f0d75921826a5c67b8ca4b5dd8175e66ccbdb5a34df49a31fa4ab3d93e93.jpg", + "image_caption": [ + "Figure 7. The left chart shows the initial output with distributed probabilities across multiple classes, indicating uncertainty in class prediction. Through our point erasure strategy, the right chart exhibits a distribution with a dominant class (\"person\") having a significantly higher probability," + ], + "image_footnote": [], + "bbox": [ + 91, + 297, + 478, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "three components yields substantial performance gains, as evidenced by the last row. Additionally, the patch pool and instance pool within the patch-based data augmentation facilitate the fusion of diverse semantic information across multiple scenes. To verify the effectiveness of interactions at the multi-scene level, we modify the pool implementation to allow interactions between only two scenes. The results in Tab. 4 indicate that multi-scene interactions yield superior performance compared to dual-scene interactions.", + "bbox": [ + 88, + 532, + 482, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Mixing Strategies. Fig. 6 (c) compares our patch-based data augmentation with existing mixing methods [27, 47, 50]. MixUp [51] shows a slight improvement over the baseline, likely due to its limited capacity for information transformation. Note that we reproduce Cylinder-Mix in the LiDAR branch [5] for comparison. Other methods demonstrate considerable improvement over the baseline by enriching point cloud distributions. In contrast, our AIScene achieves the most substantial gains by effectively leveraging inter-scene correlation, highlighting its effectiveness in facilitating diverse semantic information generation for labeled and unlabeled scenes. This is achieved through an innovative design that incorporates a patch pool and an instance pool spanning multiple scenes.", + "bbox": [ + 89, + 671, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Patch Split Number. As the patch split number $n$ plays", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/186132ff7da66aac59ca28d908eadb0e344be1e50ee484ea2e2ddc3a241d145e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Augmentation TypeMethodsSplits on SemanticAvg.
1%10%20%50%
dual-scene levelOurs53.461.361.862.259.7
multi-scene levelOurs54.563.363.764.361.5
Improvement-+1.1+2.0+1.9+2.1+1.8
", + "bbox": [ + 517, + 297, + 906, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison between dual-scene and multi-scene levels in the patch-based data augmentation of our AIScene.", + "bbox": [ + 511, + 375, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a critical role in our framework, we further analyze the impact of varying $n$ in patch-based data augmentation, with the results presented in Fig. 6 (d). When $n$ is small, it can yield a certain level of enhancement, but it may fail to generate data sufficiently divergent from the original, thereby impeding the effective generation of diverse semantic information between labeled and unlabeled scenes. Conversely, when $n$ is too large, the generated scenes may deviate excessively from the original data, resulting in a loss of essential prior scene information. Empirically, we find that setting $n$ to 18 yields optimal results. Throughout this paper, we consistently use $n$ set to 18 unless explicitly stated otherwise.", + "bbox": [ + 511, + 417, + 906, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 616, + 633, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper explores scene affinity for semi-supervised Li-DAR semantic segmentation in driving scenes, named AIScene. Specifically, AIScene addresses intra-scene inconsistency through a point erasure strategy, which enhances the learning of pseudo-labeled points. Moreover, it employs a patch-based data augmentation method to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method enriches the semantic diversity of labeled scenes and complements pseudo-labeled scenes. Extensive experiments show that AIScene achieves state-of-the-art results on popular benchmarks. Additionally, the novel erasure strategy and augmentation method can function independently to facilitate semi-supervised LiDAR semantic segmentation.", + "bbox": [ + 511, + 642, + 906, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (No.62325111, U22B2011, 62271354).", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, pages 1-8. IEEE, 2020. 3", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6", + "[3] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In CVPR, pages 7020-7030, 2023. 3", + "[4] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In CVPR, pages 2613-2622, 2021. 3, 6", + "[5] Yujun Chen, Xin Tan, Zhizhong Zhang, Yanyun Qu, and Yuan Xie. Beyond the label itself: Latent labels enhance semi-supervised point cloud panoptic segmentation. In AAAI, pages 1245-1253, 2024. 2, 3, 8", + "[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2", + "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 4, 6", + "[8] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6", + "[9] Jin Fang, Xinxin Zuo, Dingfu Zhou, Shengze Jin, Sen Wang, and Liangjun Zhang. Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In CVPR, pages 4710-4720, 2021. 3", + "[10] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE RA-L, 7(2):3795-3802, 2022. 2, 6", + "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6", + "[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 1, 2, 3", + "[13] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In ICCV, pages 6423-6432, 2021. 2, 6", + "[14] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In ICCV, pages 228-240, 2023. 3", + "[15] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu." + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lasermix for semi-supervised lidar semantic segmentation. In CVPR, pages 21705-21715, 2023. 2, 3, 4, 5, 6, 7, 8", + "[16] Lingdong Kong, Xiang Xu, Jiawei Ren, Wenwei Zhang, Liang Pan, Kai Chen, Wei Tsang Ooi, and Ziwei Liu. Multimodal data-efficient 3d scene understanding for autonomous driving. arXiv:2405.05258, 2024. 3", + "[17] Hyeokjun Kweon, Jihun Kim, and Kuk-Jin Yoon. Weakly supervised point cloud semantic segmentation via artificial oracle. In CVPR, pages 3721-3731, 2024. 3", + "[18] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 3", + "[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, pages 17545-17555, 2023. 2, 4", + "[20] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3", + "[21] Jianan Li and Qiulei Dong. Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In CVPR, pages 3260-3269, 2024. 2, 3, 4, 6", + "[22] Jianan Li and Qiulei Dong. Lass3d: Language-assisted semi-supervised 3d semantic segmentation with progressive unreliable data exploitation. In ECCV, pages 252-269. Springer, 2025. 2, 3, 6", + "[23] Li Li, Hubert PH Shum, and Toby P Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In CVPR, pages 9361-9371, 2023. 3, 6", + "[24] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In CVPR, pages 8428-8437, 2022. 3", + "[25] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In CVPR, pages 23819-23828, 2023. 3", + "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, pages 4213-4220. IEEE, 2019. 3", + "[27] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In 3DV, pages 116-125. IEEE, 2021. 2, 8", + "[28] Gilles Puy, Spyros Gidaris, Alexandre Boulch, Oriane Simeoni, Corentin Sautier, Patrick Pérez, Andrei Bursuc, and Renaud Marlet. Three pillars improving vision foundation model distillation for lidar. In CVPR, pages 21519-21529, 2024. 6", + "[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 3", + "[30] Laurenz Reichardt, Nikolas Ebert, and Oliver Wasenmüller. 360deg from a single camera: A few-shot approach for lidar segmentation. In ICCV, pages 1075-1083, 2023. 3" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27388", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3", + "[32] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Fabio Poiesi, and Elisa Ricci. Compositional semantic mix for domain adaptation in point cloud segmentation. IEEE TPAMI, 2023. 3", + "[33] Shivanand Venkanna Sheshappanavar, Vinit Veerendraveer Singh, and Chandra Kambhametu. Patchaugment: Local neighborhood augmentation in point cloud classification. In ICCV, pages 2118-2127, 2021. 2", + "[34] Jungwook Shin, Jaeill Kim, Kyungeun Lee, Hyunghun Cho, and Wonjong Rhee. Diversified and realistic 3d augmentation via iterative construction, random placement, and hpr occlusion. In AAAI, pages 2282-2291, 2023. 3", + "[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 4, 5", + "[36] Shihao Su, Jianyun Xu, Huanyu Wang, Zhenwei Miao, Xin Zhan, Dayang Hao, and Xi Li. Pups: Point cloud unified panoptic segmentation. In AAAI, pages 2339-2347, 2023. 1", + "[37] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702. Springer, 2020. 2", + "[38] Liyao Tang, Zhe Chen, Shanshan Zhao, Chaoyue Wang, and Dacheng Tao. All points matter: Entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS, 36, 2024. 3", + "[39] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, pages 1195–1204, 2017. 2, 3, 4, 5, 6", + "[40] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In CVPR, pages 2697-2707, 2022. 3", + "[41] Ozan Unal, Dengxin Dai, Lukas Hoyer, Yigit Baran Can, and Luc Van Gool. 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In WACV, pages 7336-7345, 2024. 3, 6", + "[42] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, pages 4917-4928, 2024. 6", + "[43] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 3", + "[44] Yi Wang, Jiaze Wang, Jinpeng Li, Zixu Zhao, Guangyong Chen, Anfeng Liu, and Pheng Ann Heng. Pointpatchmix: Point cloud mixing with patch scoring. In AAAI, pages 5686-5694, 2024. 3", + "[45] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, pages 4376-4382. IEEE, 2019. 3", + "[46] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 3", + "[47] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In NeurIPS, 2022. 3, 8", + "[48] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In ECCV, pages 1-19. Springer, 2020. 3", + "[49] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 5", + "[50] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 8", + "[51] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv:1710.09412, 2017. 3, 8", + "[52] Jinlai Zhang, Lyujie Chen, Bo Ouyang, Binbin Liu, Jihong Zhu, Yujin Chen, Yanmei Meng, and Danfeng Wu. Pointcutmix: Regularization strategy for point cloud classification. Neurocomputing, 505:58-67, 2022. 2, 3", + "[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, pages 9601-9610, 2020. 3", + "[54] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In AAAI, pages 3421-3429, 2021. 3", + "[55] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IROS, pages 4453-4458. IEEE, 2021. 3", + "[56] Qinfeng Zhu, Lei Fan, and Ningxin Weng. Advancements in point cloud data augmentation for deep learning: A survey. arXiv:2308.12113, 2023. 3", + "[57] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1, 2, 4, 6, 7, 8", + "[58] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, pages 289-305, 2018. 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 813 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27389", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_model.json b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d2f05f26fa46d7af34d15f70d79d423ce7cb8f9 --- /dev/null +++ b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_model.json @@ -0,0 +1,2332 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.888, + 0.154 + ], + "angle": 0, + "content": "Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.189, + 0.948, + 0.21 + ], + "angle": 0, + "content": "Chuandong Liu\\(^{1,*}\\), Xingxing Weng\\(^{1,*}\\), Shuguo Jiang\\(^{1}\\), Pengcheng Li\\(^{3}\\), Lei Yu\\(^{2,4}\\), Gui-Song Xia\\(^{1,2,5,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.209, + 0.719, + 0.226 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Wuhan University" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.226, + 0.73, + 0.244 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Artificial Intelligence, Wuhan University" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.244, + 0.782, + 0.261 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Computer Science and Technology, Hainan University" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.261, + 0.735, + 0.279 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Electronic Information, Wuhan University" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.279, + 0.725, + 0.297 + ], + "angle": 0, + "content": "\\( {}^{5} \\) State Key Lab. of LIESMARS, Wuhan University" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.297, + 0.695, + 0.314 + ], + "angle": 0, + "content": "\\(^{6}\\)Institute for Math & AI, Wuhan University" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.349, + 0.327, + 0.365 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.381, + 0.485, + 0.775 + ], + "angle": 0, + "content": "This paper explores scene affinity (AIScene), namely intrascene consistency and inter-scene correlation, for semisupervised LiDAR semantic segmentation in driving scenes. Adopting teacher-student training, AIScene employs a teacher network to generate pseudo-labeled scenes from unlabeled data, which then supervise the student network's learning. Unlike most methods that include all points in pseudo-labeled scenes for forward propagation but only pseudo-labeled points for backpropagation, AIScene removes points without pseudo-labels, ensuring consistency in both forward and backward propagation within the scene. This simple point erasure strategy effectively prevents unsupervised, semantically ambiguous points (excluded in backpropagation) from affecting the learning of pseudo-labeled points. Moreover, AIScene incorporates patch-based data augmentation, mixing multiple scenes at both scene and instance levels. Compared to existing augmentation techniques that typically perform scene-level mixing between two scenes, our method enhances the semantic diversity of labeled (or pseudo-labeled) scenes, thereby improving the semi-supervised performance of segmentation models. Experiments show that AIScene outperforms previous methods on two popular benchmarks across four settings, achieving notable improvements of \\(1.9\\%\\) and \\(2.1\\%\\) in the most challenging \\(1\\%\\) labeled data. The code will be released at https://github.com/azhuantou/AIScene." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Semantic segmentation of LiDAR point clouds is essential for autonomous driving, allowing vehicles to perceive their 3D surroundings. Most state-of-the-art models for" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.351, + 0.705, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.454, + 0.666, + 0.464 + ], + "angle": 0, + "content": "pseudo-labeled scene" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.352, + 0.9, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.454, + 0.871, + 0.464 + ], + "angle": 0, + "content": "intra-scene inconsistency" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.475, + 0.706, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.528, + 0.706, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.597, + 0.695, + 0.608 + ], + "angle": 0, + "content": "scene-level mixing on two scenes" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.476, + 0.901, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.518, + 0.901, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.597, + 0.872, + 0.608 + ], + "angle": 0, + "content": "limited semantic diversity" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.802 + ], + "angle": 0, + "content": "Figure 1. Top: Pseudo-labeled scenes generated from unlabeled scenes typically contain points with pseudo labels (colored) and without them (gray). Previous methods include all points in forward propagation but only pseudo-labeled points in backward propagation, leading to intra-scene inconsistency. This inconsistency may allow unsupervised, semantically ambiguous points to affect the learning of pseudo-labeled points (see Fig. 7). Bottom: Existing data augmentation techniques suitable for semi-supervised segmentation usually mix two scenes through scene-level operations like concatenation or swapping. Their resulting scenes may be constrained in terms of semantic diversity, potentially lacking objects like bicyclists and buses, which are important in driving scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "driving-scene LiDAR segmentation are dependent on high-quality, point-annotated datasets [12, 36, 57]. With numerous objects in driving scenes, densely annotating points is extremely labor-intensive. Consequently, semi-supervised LiDAR semantic segmentation has attracted growing atten" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.886, + 0.368, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. \\( {}^{ \\ominus } \\) Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27380" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "tion, where models are trained on a limited amount of labeled data supplemented by a large set of unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.482, + 0.319 + ], + "angle": 0, + "content": "The common solution for learning from unlabeled data is to apply teacher-student training, where a teacher network generates pseudo-labeled scenes from unlabeled data, which are then used to supervise the learning of a student network. The student network's objective is to minimize segmentation loss while the teacher network's weights are updated via exponential moving average [39]. To further improve model learning with unlabeled data, existing studies implement designs such as utilizing language-vision models [22], adopting density-guided contrastive learning [21], or trying feature fusion between LiDAR and image data [5]. Despite the progress, two main issues remain to be addressed:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.483, + 0.565 + ], + "angle": 0, + "content": "(1) Given the teacher network's confident prediction, pseudo-labeled scenes typically contain points with and without pseudo-labels. During training, most studies include all points in pseudo-labeled scenes in the student network's forward propagation, but only pseudo-labeled points are involved in backward propagation. Points excluded from backpropagation are unsupervised and, therefore, semantically ambiguous. However, these points still contribute to the forward feature computation for pseudolabeled points, which may interfere with the model's ability to learn from them effectively. In short, the inconsistency between forward and backward propagation within the scene causes ambiguous points to obstruct the model's learning of pseudo-labeled points. Both point-based [12] and voxel-based [7, 57] methods suffer from this issue in feature computation, as illustrated at the top of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.793 + ], + "angle": 0, + "content": "(2) Data augmentation is an effective practice in semi-supervised LiDAR semantic segmentation [13, 15]. The basic idea is to mix raw labeled scenes and pseudo-labeled scenes to enrich scene semantics. Although impressive advancements have been made [33, 52], they usually select two scenes (one labeled scene and one pseudo-label scene) and then perform scene-level mix through operations like concatenation [27] and swapping [15], which limits the semantic diversity of mixed scenes. As shown at the bottom of Fig. 1, multiple scenes, whether labeled or pseudo-labeled, often share similar spatial layouts. Moreover, in driving scenes, objects such as cars and pedestrians are the primary focus of an autonomous vehicle's attention. This naturally raises the question: can we improve the semantic diversity of scenes by mixing multiple scenes at the instance level?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To address inconsistency and implement the multi-scene mixing idea, we explore Scene Affinity for semi-supervised LiDAR semantic segmentation, named AIScene. The term scene affinity refers to intra-scene consistency and interscene correlation. Specifically, AIScene, built on the teacher-student training pipeline, employs a simple yet helpful point erasure strategy, removing points without" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "pseudo-labels in forward propagation. This ensures intra-scene consistency in both forward and backward propagation, preventing semantically ambiguous points from disrupting learning from pseudo-labeled points. Building on this, AIScene incorporates patch-based data augmentation that mixes multiple scenes at both scene and instance levels. By maintaining two pools that store scene and instance patches from various scenes, it selects patches for mixing based on inter-scene correlation. This novel augmentation method not only enhances the semantic diversity of raw labeled scenes but also complements pseudo-labeled scenes where some points have been erased." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.499 + ], + "angle": 0, + "content": "AIScene has been extensively evaluated through experiments and demonstrated superiority on two popular benchmarks for semi-supervised LiDAR semantic segmentation. It exceeds state-of-the-art methods by \\(1.3\\%\\) on the SemanticKITTI [2] and \\(1.2\\%\\) on the nuScenes [10] in terms of average performance across four settings (1%, 10%, 20%, and 50% labeled data). Notably, in the most challenging setting, with only \\(1\\%\\) labeled data, AIScene achieves remarkable improvements of \\(1.9\\%\\) and \\(2.1\\%\\) on the two datasets. Additionally, our ablation studies confirm AIScene's potential to reduce the requirements for labeled data across different LiDAR semantic segmentation networks (see Tab. 1 and Tab. 2), with each component of AIScene able to function independently to facilitate semi-supervised LiDAR semantic segmentation (see Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.5, + 0.848, + 0.513 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.589 + ], + "angle": 0, + "content": "- We propose a point erasure strategy to ensure intra-scene consistency and facilitate network learning from pseudolabeled scenes. This strategy can serve as a plug-and-play tool for semi-supervised LiDAR semantic segmentation using the pseudo-label mechanism." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.59, + 0.905, + 0.664 + ], + "angle": 0, + "content": "- We design a patch-based data augmentation to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method improves the semantics diversity of labeled scenes and complements pseudolabeled scenes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.665, + 0.905, + 0.741 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that AIScene achieves state-of-the-art results on popular benchmarks for semi-supervised LiDAR semantic segmentation. Ablation studies show that the novel erasure strategy and augmentation method can function independently." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.515, + 0.905, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.655, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Fully-supervised LiDAR Semantic Segmentation: Numerous approaches for LiDAR semantic segmentation have been explored, which can be broadly grouped into three categories: voxel-based, 2D projection-based, and point-based methods. Voxel-based approaches [6, 7, 19, 37, 57] involve voxelizing sparse point clouds to generate regular voxel grids, allowing the application of standard 3D convolution neural networks for embedding learning. Conversely," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27381" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.895, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.403, + 0.908, + 0.463 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed AIScene training pipeline. The teacher network generates the pseudo-labels \\( y_{i}^{u} \\) from the unlabeled scene \\( x_{i}^{u} \\), followed by an erasure operation to filter out low-confidence points, ensuring training consistency. For the labeled scene \\( x_{i}^{l} \\) and the generated erased scene \\( \\hat{x}_i^u \\), we employ patch-based data augmentation to obtain augmented scenes \\( x_{i}^{lpf} \\) and \\( \\hat{x}_i^{upf} \\), respectively. Subsequently, both scenes are fed into the student network to compute the loss of \\( \\mathcal{L}_l \\) and \\( \\mathcal{L}_u \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.482, + 0.563 + ], + "angle": 0, + "content": "2D projection-based methods [14, 26, 45, 48, 53, 55] project input point clouds into a 2D dense space, use 2D convolutional filters [31] for processing, and then map predictions back to the original 3D space. Finally, point-based methods [12, 18, 29, 46] operate directly on raw point clouds, enabling the extraction of point-wise semantic features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Semi-supervised LiDAR Semantic Segmentation: To reduce the need for 3D dense annotations, recent research has explored methods that require less supervision [3, 17, 24, 38, 40, 54]. Semi-supervised learning [4, 16, 21, 23, 30, 58] aims to utilize a large number of unlabeled data to alleviate the annotation burden. LaserMix [15], a pioneer work in semi-supervised LiDAR segmentation, leverages a simple laser beam mix strategy to achieve solid results. Subsequent research fuses LiDAR and camera data to extract deeper semantic information [5, 41]. DDsemi [21], focusing on point cloud density, builds on contrastive learning to enhance feature discriminability. The recently proposed LASS3D [22] leverages large language-vision models to support semi-supervised segmentation. Existing semi-supervised methods often build upon the mean-teacher paradigm [39] and utilize pseudo labels [1]. However, many points without pseudo-labels contribute equally to pseudo-labeled points in forward propagation but are ignored during backward propagation, leading to intra-scene inconsistency. To address this, our AIScene leverages a point erasure strategy to mitigate this inconsistency, complementing the pseudolabel mechanism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.744 + ], + "angle": 0, + "content": "3D Point Cloud Data Augmentation: Point cloud data augmentation has been explored across various 3D vision tasks [9, 25, 34, 43, 44, 56]. Inspired by the success of Mixup [51] in image classification, several methods have investigated techniques for mixing selected point cloud samples. To preserve structural information, RSMix [20] introduces a neighboring function, and PointCutMix [52] establishes correspondences for sample fusion. Notably, these methods focus on local object-level augmentation. In response, several works have explored mixing strategies for point cloud scene augmentation. PloarMix [47] swaps points within the fan-shaped areas between two scenes, while CoSMix [32] aims to mitigate domain shift by mixing semantic point groups. Unlike these methods, our method is capable of modeling inter-scene information across multiple scenes to create more diversified data, while also extracting instance-level semantic information for individual classes to further enrich scenes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.757, + 0.608, + 0.773 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.782, + 0.644, + 0.799 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.804, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Problem Formulation. Before we describe our model and training process, we first provide the definition of semi-supervised point cloud segmentation. In detail, we have a small labeled training set \\( D_{l} = \\{x_{i}^{l},y_{i}^{l}\\}_{i = 1}^{N_{l}} \\) and a large unlabeled training set \\( D_{u} = \\{x_{i}^{u}\\}_{i = 1}^{N_{u}} \\), where \\( x_{i}\\in \\mathbb{R}^{m\\times (3 + r)} \\) is a point cloud scene of \\( m \\) points with three-dimensional" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27382" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "Cartesian coordinates relative to the sensor and additional \\(r\\)-dimensional feature (e.g., intensity, color, or reflectivity), \\(y_{i}\\) represents its corresponding point-level segmentation labels, \\(N_{l}\\) and \\(N_{u}\\) are the numbers of labeled and unlabeled point cloud scenes, respectively, and \\(N_{u} \\gg N_{l}\\) in ideal cases. The goal of SSL-based point cloud segmentation is to learn a segmentation model using limited labeled data and large-scale unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.349 + ], + "angle": 0, + "content": "Teacher-Student Paradigm. Following current mainstream research [15, 21], our framework also builds upon the teacher-student paradigm, which includes two segmentation networks with the same architecture. Our AIScene can be instantiated with any 3D segmentation network [7, 19, 57], denoted as \\(\\Phi(\\cdot)\\). In detail, the teacher model generates pseudo-labels for unlabeled scenes and undergoes a gradual update through the exponential moving average (EMA) [39] of the student weights:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.359, + 0.483, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathbb {t}} ^ {k + 1} = \\alpha \\cdot \\theta_ {\\mathbb {t}} ^ {k} + (1 - \\alpha) \\cdot \\theta_ {\\mathbb {S}} ^ {k}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.451 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the EMA decay rate, which is set to 0.99 by default, \\(\\theta_{\\mathbb{t}}\\) and \\(\\theta_{\\mathrm{s}}\\) represent the weight parameters of the teacher and student models, respectively, and \\(k\\) denotes the training step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.46, + 0.304, + 0.476 + ], + "angle": 0, + "content": "3.2. Point Erasure Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.784 + ], + "angle": 0, + "content": "Based on the teacher-student mutual learning framework [35], the overall architecture of AIScene is depicted in Fig. 2. Specifically, AIScene consists of two primary branches: one dedicated to labeled scenes and the other to unlabeled scenes. How to effectively leverage a large number of unlabeled scenes is crucial for semi-supervised learning. Hence, for the branch of unlabeled scenes, we use a teacher network to generate pseudo-labeled scenes \\((x_{i}^{u},y_{i}^{u})\\) based on a predefined confidence threshold \\(\\tau_{s}\\) [15]. As mentioned previously, when using the pseudo-labeled scene for training, points without pseudo-labels are involved in the forward propagation but are ignored in the backward propagation due to the lack of corresponding supervision. This leads to intra-scene inconsistency, which causes semantic ambiguity of points. To address this issue, we propose erasing points without pseudo labels, resulting in an erased scene \\((\\hat{x}_i^u,\\hat{y}_i^u)\\) that contains fully annotated information, similar to fully supervised scenes. Formally, intrascene inconsistency addressed by the point erasure strategy can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.796, + 0.483, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {i} ^ {u} = \\left\\{x _ {i} ^ {u} \\mid \\Phi_ {s} \\left(x _ {i} ^ {u}\\right) \\geq \\tau_ {s} \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{x}_i^u\\) is the erased scene with corresponding class labels \\(\\hat{y}_i^u\\), \\(\\Phi_s(\\cdot)\\) represents the segmentation network that predicts the confidence score for each point in the scene \\(x_i^u\\), and \\(\\tau_s\\) is a predefined confidence threshold, equivalent to the pseudolabel mining threshold, requiring no additional adjustment." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.092, + 0.899, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.219, + 0.907, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Illustration of the pool initialization process. For each labeled point cloud scene, splitting it into multiple patches from the BEV perspective and assigning different instance categories to each patch enables the extraction of instances belonging to different classes. Subsequently, these patches and instances are stored individually in the patch pool and instance pool." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.907, + 0.405 + ], + "angle": 0, + "content": "Since the proposed point erasure strategy is independent of the segmentation network architecture, it can be utilized as a plug-and-play tool for semi-supervised 3D semantic segmentation methods based on the pseudo-label mechanism, as depicted in Fig. 6 (a)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.413, + 0.8, + 0.429 + ], + "angle": 0, + "content": "3.3. Patch-based Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.906, + 0.54 + ], + "angle": 0, + "content": "Based on the labeled and erased scenes, we propose a patch-based data augmentation to enhance the semantic diversity of annotation information, guided by inter-scene correlation. The key idea is to construct patch pools to store scene-level and instance-level information, and then combine patches from multiple scenes to generate diverse training samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.54, + 0.907, + 0.647 + ], + "angle": 0, + "content": "Pool Initialization. We first use the labeled training set \\(D_{l}\\) to create two patch pools, which store scene-level and instance-level information, respectively, as shown in Fig. 3. Given the \\(i\\)-th point cloud scene \\(x_{i}^{l}\\), we patchify \\(x_{i}^{l}\\) into a sequence of \\(T\\) regular, non-overlapping point patches \\(\\{p_{ij}^{l}\\}_{j=1}^{T}\\), each of size \\((n \\times n)\\), in the bird's eye view. This process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.655, + 0.907, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\left\\{p _ {i j} ^ {l} \\right\\} _ {j = 1} ^ {T} = \\text {p a t c h i f y} \\left(x _ {i} ^ {l}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.906, + 0.738 + ], + "angle": 0, + "content": "where \\(p_{ij}^{l}\\) denotes the \\(j\\)-th point patch from the sequence. For all labeled scenes, by performing the operation in Eq. (3), the patch pool \\(\\mathcal{B}_{\\mathrm{Pat}}^{l}\\) is constructed, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.757, + 0.907, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {\\mathrm {P a t}} ^ {l} = \\left\\{\\left\\{p _ {i j} ^ {l} \\right\\} _ {j = 1} ^ {T} \\right\\} _ {i = 1} ^ {N _ {l}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.795, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Based on the patch pool, we can obtain the instance pool \\(\\mathcal{B}_{\\mathrm{Ins}}^l\\) by assigning different class labels to each patch, as depicted in Fig. 3, which contains instance information for different categories (e.g., car and person)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Following the same steps used for labeled scenes, we further create a pseudo patch pool \\(\\mathcal{B}_{\\mathrm{Pat}}^u\\) and a pseudo instance pool \\(\\mathcal{B}_{\\mathrm{Ins}}^u\\) from the erased scene \\(\\hat{x}_i^u\\). Unlike \\(\\mathcal{B}_{\\mathrm{Pat}}^l\\) and \\(\\mathcal{B}_{\\mathrm{Ins}}^l\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.094, + 0.266, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.094, + 0.482, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.483, + 0.322 + ], + "angle": 0, + "content": "Figure 4. (a) Illustration of the MixPatch process. (b) Illustration of selecting instances from the patch pool for filling. An instance will not be filled if there are no other points around the filling location or if the filled instance overlaps with existing instances." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.483, + 0.512 + ], + "angle": 0, + "content": "covering the whole training set \\(D_{l}\\), the pseudo pools \\(\\mathcal{B}_{\\mathrm{Pat}}^{u}\\) and \\(\\mathcal{B}_{\\mathrm{Ins}}^{u}\\) only contain scene and instance information from the current batch and are reset in each iteration. This prevents low-quality pseudo labels from persisting in pools and misleading model training. Additionally, if the number of points within a patch or instance falls below a certain threshold \\(\\tau_{\\mathrm{min}}\\), it will not be saved into the pool. In practice, we set \\(\\tau_{\\mathrm{min}}\\) to 5 [49]. If a patch or instance has fewer than 5 points, its spatial structure exhibits high uncertainty, making it semantically ambiguous. Thus, it is excluded from the pools to prevent noise during data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.693 + ], + "angle": 0, + "content": "MixPatch. To model the inter-scene correlation and then guide the transfer of annotation information between labeled scenes and unlabeled scenes, we combine multiple scenes using patch pools. Fig. 4 (a) illustrates the patch mixing (MixPatch) strategy, while Fig. 4 (b) presents the instance filling (InsFill) strategy, which is introduced in the next paragraph. During a training iteration, the mini-batch is comprised of labeled and unlabeled scenes. After the unlabeled scene \\( x_{i}^{u} \\) is transformed into the erased scene \\( \\hat{x}_{i}^{u} \\) using the point erasure strategy, we integrate information from the patch pool into \\( \\hat{x}_{i}^{u} \\) by MixPatch. The MixPatch process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.702, + 0.482, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {i} ^ {u p} = \\left(\\hat {x} _ {i} ^ {u} \\odot \\mathcal {M}\\right) \\oplus \\left(\\mathcal {B} _ {\\text {P a t}} ^ {l} \\odot (1 - \\mathcal {M})\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.835 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is a binary mask indicating which patches from \\(\\hat{x}_i^u\\) will be replaced with selected patches of the same patch index from the patch pool, \\(\\odot\\) denotes element-wise multiplication, and \\(\\hat{x}_i^{up}\\) represents the augmented scene. For labeled scenes, MixPatch is applied based on the current pseudo patch pool \\(\\mathcal{B}_{\\mathrm{Pat}}^u\\), which can be expressed as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.843, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nx _ {i} ^ {l p} = \\left(x _ {i} ^ {l} \\odot (1 - \\mathcal {M})\\right) \\oplus \\left(\\mathcal {B} _ {\\text {P a t}} ^ {u} \\odot \\mathcal {M}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is the same as defined in Eq. (5), thereby ensuring the full utilization of scene information from the unlabeled" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.211 + ], + "angle": 0, + "content": "scenes. The binary mask \\(\\mathcal{M}\\) and its complement \\(1 - \\mathcal{M}\\) are used to merge inter-scene point cloud patches with the erased scene, facilitating the generation of diverse semantic information for labeled and unlabeled scenes. To alleviate bias in the patch mixing process, we apply a uniform sampling strategy to the patches in each scene, ensuring that each patch has an equal opportunity of being selected in the mixed scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.907, + 0.52 + ], + "angle": 0, + "content": "InsFill. In addition to patch mixing, we utilize the instances stored in the instance pool \\(\\mathcal{B}_{\\mathrm{Ins}}^l\\) to further enhance the scene \\(\\hat{x}_i^{up}\\). Given a patch with location index \\(j\\) in the scene, we randomly select an instance from the instance pool \\(\\mathcal{B}_{\\mathrm{Ins}}^l\\) that was extracted from the \\(j\\)-th patch. Patches with identical indices typically correspond to similar Li-DAR spatial layouts, ensuring consistency between the selected instance and its fill positions in the scene. As depicted in Fig. 4 (b), if the selected instance overlaps with existing instances in the scene or lacks sufficient contextual information, the filling operation is bypassed. This judgment process can be easily implemented using the maximum 3D bounding boxes. This data augmentation provides instance-level semantic information for the point cloud scene, enabling the augmented scene \\(\\hat{x}_i^{upf}\\) and corresponding labels \\(\\hat{y}_i^{upf}\\) to effectively support semi-supervised training. For the labeled branch, we also apply the same filling strategy as described, selecting instances from the pseudo instance pool \\(\\mathcal{B}_{\\mathrm{Ins}}^u\\) to fill the current scene. This InsFill process yields the scene \\(x_i^{lpf}\\) along with its corresponding label \\(y_i^{lpf}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.525, + 0.671, + 0.541 + ], + "angle": 0, + "content": "3.4. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.668 + ], + "angle": 0, + "content": "Following the standard teacher-student mutual learning paradigm [15, 35, 39], we construct a mini-batch of data with an equal composition of labeled and unlabeled scenes at each iteration. The unlabeled scenes are assigned with pseudo-labels by the teacher model and then train the student model jointly with labeled scenes. We use three types of loss functions to supervise the model training here, which is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.676, + 0.905, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\mathcal {L} _ {\\text {s e g}} \\left(\\Phi_ {\\theta_ {\\mathfrak {s}}} \\left(x _ {i} ^ {l}\\right), y _ {i} ^ {l}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.697, + 0.905, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l} = \\mathcal {L} _ {\\text {s e g}} \\left(\\Phi_ {\\theta_ {\\mathrm {s}}} \\left(x _ {i} ^ {l p f}\\right), y _ {i} ^ {l p f}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.719, + 0.905, + 0.738 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\mathcal {L} _ {s e g} \\left(\\Phi_ {\\theta_ {\\mathrm {s}}} \\left(\\hat {x} _ {i} ^ {u p f}\\right), \\hat {y} _ {i} ^ {u p f}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.823 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_s, \\mathcal{L}_l\\), and \\(\\mathcal{L}_u\\) denote the segmentation loss for the labeled scene \\(x_i^l\\), augmented labeled scene \\(x_i^{lpf}\\), and augmented unlabeled scene \\(\\hat{x}_i^{upf}\\), respectively, while \\(\\mathcal{L}_{seg}\\) is the original loss of the segmentation network \\(\\Phi(\\cdot)\\). We sum up all the loss term to define the final training objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.847, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda_ {u} \\mathcal {L} _ {u} + \\lambda_ {l} \\mathcal {L} _ {l}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda_{u}\\) and \\(\\lambda_{l}\\) are loss weights that balance the two loss terms. After that, we update the teacher model weights us" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27384" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.896, + 0.299 + ], + "angle": 0, + "content": "
MethodVenueSemanticKITTI [2]Avg.nuScenes [10]Avg.
1%10%20%50%1%10%20%50%
Meanteacher [39]NeurIPS'201745.457.159.260.055.451.666.067.171.764.1
CBST [58]ECCV'201848.858.359.459.756.653.066.569.671.665.2
CPS [4]CVPR'202146.758.759.660.556.452.966.370.072.565.4
LaserMix (Range View) [15]CVPR'202343.458.859.461.455.849.568.270.673.065.3
LaserMix (Voxel) [15]CVPR'202350.660.061.962.358.755.369.971.873.267.6
GPC [13]ICCV'202154.162.062.562.860.4-----
IGNet [41]WACV'202449.061.363.164.859.6-----
LiM3D [23]CVPR'202358.462.263.163.661.8-----
LASS3D [22]ECCV'202458.563.064.164.562.5-----
DDSemi [21]CVPR'202459.365.166.367.064.458.170.274.076.569.7
Ours (AIScene)-61.266.367.467.965.760.272.375.076.470.9
Improvement ↑-+1.9+1.2+1.1+0.9+1.3+2.1+2.1+1.0-0.1+1.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.302, + 0.908, + 0.332 + ], + "angle": 0, + "content": "Table 1. Experimental results on the SemanticKITTI and nuScenes dataset compared with different SSL methods. The results are reported with mIoU scores (\\%). The bold and underline scores respectively denote the best and second results." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.343, + 0.373, + 0.358 + ], + "angle": 0, + "content": "ing the EMA strategy described in Eq. (1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.371, + 0.224, + 0.387 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.395, + 0.377, + 0.41 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.483, + 0.568 + ], + "angle": 0, + "content": "SemanticKITTI. Following state-of-the-art methods [15, 21], we evaluate our AISCene on the SemanticKITTI dataset [2], which is a large-scale segmentation dataset consisting of LiDAR data derived from the popular KITTI dataset [11]. It is composed of 22 sequences, with sequences 00-10 serving as the train set (of which 08 as the val set) and sequences 11-21 as the test set. Moreover, all the point cloud scenes are collected from real-world rural and urban road environments in Germany, and the semantic class labels are mapped to 19 classes in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.688 + ], + "angle": 0, + "content": "nuScenes. The nuScenes dataset [10] is another well-known large autonomous driving dataset. This dataset is composed of 1000 sequences of 20s duration (850 for training and validation, 150 for testing), which are captured by the LiDAR sensor from Boston and Singapore. Following the official splitting protocol, the train and val sets contain 28,130 and 6,019 point cloud scenes, respectively, including per-points annotation for 16 semantic classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Evaluation Metrics. For both the SemanticKITTI and nuScenes datasets, we sample different ratios of annotated train scenes to obtain semi-supervised data. Meanwhile, we report the performance on the validation set using the mean Intersection over Union (mIoU) metric [21, 23]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.308, + 0.789 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments with \\(4 \\times\\) NVIDIA V100 GPUs with 32GB RAM, based on MMDetection3D [8]. For the threshold \\(\\tau_{s}\\) for pseudo-label generation and point erasure, we experimentally find that 0.9 yields relatively satisfactory results. We use the MinkowskiNet [7] and Cylinder3D [57] as the backbone network due to its widely adopted in previous methods [15, 28, 42]. To facilitate patch-based data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.509 + ], + "angle": 0, + "content": "augmentation, we restrict our focus to point clouds within the X and Y dimensions in the range of \\([-50, 50]m\\) for both the SemanticKITTI and nuScenes datasets. Similar to mainstream work, we also adopt the consistency loss proposed in MeanTeacher [39] to align the predictions of the teacher and student networks. Additionally, the weights of the consistency loss are assigned as \\(1 \\times 10^{3}\\) and \\(2 \\times 10^{3}\\) for the SemanticKITTI and nuScenes datasets, respectively. The weights for \\(\\lambda_{u}\\) and \\(\\lambda_{l}\\) are set as 1 for both datasets. Our method excludes all forms of model ensemble and test-time augmentation during the evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.652, + 0.531 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.689 + ], + "angle": 0, + "content": "In this subsection, we compare our method with current SOTA semi-supervised point cloud segmentation methods (e.g., LaserMix [15], LiM3D [23] and DDsemi [21]), using sub-sampling of the datasets under various partition protocols [21]. Specifically, we split the entire training set into labeled subsets with ratios of \\(1\\%\\), \\(10\\%\\), \\(20\\%\\), and \\(50\\%\\) in the SemanticKITTI [2] and nuScenes dataset [10]. For further comparison, we also include several classical semi-supervised learning methods [4, 39, 58] from the 2D image domain adapted to 3D segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 summarizes the quantitative results on SemanticKITTI and nuScenes datasets. Our method consistently outperforms other comparative methods on both datasets, achieving average mIoU improvements of \\(1.3\\%\\) and \\(1.2\\%\\), respectively. Notably, the proposed method demonstrates even greater gains on smaller sampling splits, such as \\(1\\%\\) and \\(10\\%\\), which are particularly beneficial for the semi-supervised setting. Specifically, with a \\(1\\%\\) ratio, our method shows mIoU improvements of up to \\(1.9\\%\\) and \\(2.1\\%\\) on the two datasets, respectively, compared to the recently proposed SOTA method, DDsemi [21]. These results also validate that our method is less sensitive to reduced labeled data sampling than other methods. Furthermore, we present qualitative results in Fig. 5. As indicated by the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.222, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.09, + 0.351, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.09, + 0.478, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.215, + 0.222, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.215, + 0.35, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.215, + 0.478, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.323, + 0.222, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.323, + 0.35, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.323, + 0.478, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.442, + 0.483, + 0.496 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison results from LiDAR bird's eye view on SemanticKITTI val set. We set the point with correct and incorrect predicted semantic class in blue and orange for better visualization. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.544 + ], + "angle": 0, + "content": "purple and red dashed circles, our method achieves better segmentation of boundary regions and small objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.245, + 0.573 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.671 + ], + "angle": 0, + "content": "In this subsection, We first validate the adaptability of our method to different segmentation models. We then examine the impact of intra-scene consistency through the point erasure strategy and explore inter-scene correlation by incorporating patch-based mixing and patch-based instance filling data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Adaptability with Different Segmentation Methods. Our AISCene is segmentor-agnostic and can be easily used to further improve off-the-shelf 3D fully-supervised segmentors under the semi-supervised setting. Tab. 2 shows that the proposed method can substantially boost the performance with Cylinder3D [57] as the backbone, indicating the generalization ability of our framework. For consistency in comparison, the subsequent experiments are based on Cylinder3D unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Point Erasure Strategy. The \\(1^{\\text{st}}\\) row in Tab. 3 reports the performance of the supervision-only Cylinder3D model [57], which serves as the baseline. When PtErase is applied to the baseline, it effectively boosts performance by ensuring consistency during forward and backward propagation, thus addressing the lim" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.188 + ], + "angle": 0, + "content": "
DatasetsMethods1%10%20%50%Avg.
SemanticKITTICylinder3D [57]45.456.157.858.754.5
Ours (AIScene)54.563.363.764.361.5
Improvement+9.1+7.2+5.9+5.6+7.0
nuScenesCylinder3D [57]50.965.966.671.263.7
Ours (AIScene)56.670.272.873.968.4
Improvement+5.7+4.3+6.2+2.7+4.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.199, + 0.906, + 0.226 + ], + "angle": 0, + "content": "Table 2. Ablation study on the adaptability of our AIScene using Cylinder3D as the segmentation model." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.23, + 0.906, + 0.363 + ], + "angle": 0, + "content": "
Ours AISCeneptEraseMixPatchInsFillSplits on SemanticKITTIAvg.
1%10%20%50%
45.456.157.858.754.5
✓51.759.861.860.458.4
✓51.361.460.760.558.5
✓52.361.760.161.158.8
✓✓54.661.062.962.860.3
✓✓52.061.760.862.659.3
✓✓50.562.563.863.660.1
✓✓✓54.563.363.764.361.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.373, + 0.906, + 0.43 + ], + "angle": 0, + "content": "Table 3. Ablation study of different components of our AIScene framework. PtErase refers to the point erasure strategy, while MixPatch and InsFill refer to the patch mix and instance filling data augmentation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.717 + ], + "angle": 0, + "content": "ditions of the pseudo-label mechanism. This consistency is achieved through a plug-and-play point erasure strategy, which can be easily integrated into methods relying on the pseudo-label mechanism. To further validate the importance of maintaining consistency during forward and backward propagation, we embedded the PtErase strategy into LaserMix [15]. As illustrated in Fig. 6 (a), using the PtErase strategy improves performance across all ratios, underscoring the significance of intra-scene consistency and demonstrating the generalizability of our method. Moreover, Fig. 6 (b) shows that the proportion of points removed by the erasure strategy is initially high at the start of training, but gradually decreases as the network converges. Furthermore, We present a comparison of class probability distributions in Fig. 7 to validate the effectiveness of the point erasure strategy. The results demonstrate that this strategy reduces classification ambiguity, enabling the model to better capture the features of the correct class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Patch-based Data Augmentation. The patch-based data augmentation consists of MixPatch (patch-based mixing) and InsFill (instance filling). In Tab. 3, the results in the \\(3^{\\mathrm{rd}}\\) and \\(4^{\\mathrm{th}}\\) rows indicate that using MixPatch or InsFill individually yields an average mIoU improvement of \\(4.0\\%\\) or \\(4.3\\%\\), respectively, over the baseline. This suggests that incorporating diverse semantic information across scenes benefits semi-supervised LiDAR segmentation. Furthermore, as shown in the \\(5^{\\mathrm{th}}\\) and \\(6^{\\mathrm{th}}\\) rows, combining MixPatch or InsFill with PtErase significantly enhances performance, highlighting the effectiveness of patch-based data augmentation. The combination of all" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.905, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.248, + 0.907, + 0.29 + ], + "angle": 0, + "content": "Figure 6. (a) Illustration of performance differences whether using point erasure strategy on the LaserMix [15]. (b) Variation in the proportion of points removed by the point erasure strategy throughout the training process. (c) Comparison of different mixing-based techniques on SemanticKITTI dataset. (d) Results of mIoU against the number of patch split \\( n \\) under the \\( 10\\% \\) SemanticKITTI split." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.298, + 0.48, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Figure 7. The left chart shows the initial output with distributed probabilities across multiple classes, indicating uncertainty in class prediction. Through our point erasure strategy, the right chart exhibits a distribution with a dominant class (\"person\") having a significantly higher probability," + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.534, + 0.483, + 0.67 + ], + "angle": 0, + "content": "three components yields substantial performance gains, as evidenced by the last row. Additionally, the patch pool and instance pool within the patch-based data augmentation facilitate the fusion of diverse semantic information across multiple scenes. To verify the effectiveness of interactions at the multi-scene level, we modify the pool implementation to allow interactions between only two scenes. The results in Tab. 4 indicate that multi-scene interactions yield superior performance compared to dual-scene interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Different Mixing Strategies. Fig. 6 (c) compares our patch-based data augmentation with existing mixing methods [27, 47, 50]. MixUp [51] shows a slight improvement over the baseline, likely due to its limited capacity for information transformation. Note that we reproduce Cylinder-Mix in the LiDAR branch [5] for comparison. Other methods demonstrate considerable improvement over the baseline by enriching point cloud distributions. In contrast, our AIScene achieves the most substantial gains by effectively leveraging inter-scene correlation, highlighting its effectiveness in facilitating diverse semantic information generation for labeled and unlabeled scenes. This is achieved through an innovative design that incorporates a patch pool and an instance pool spanning multiple scenes." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Patch Split Number. As the patch split number \\( n \\) plays" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.298, + 0.907, + 0.366 + ], + "angle": 0, + "content": "
Augmentation TypeMethodsSplits on SemanticAvg.
1%10%20%50%
dual-scene levelOurs53.461.361.862.259.7
multi-scene levelOurs54.563.363.764.361.5
Improvement-+1.1+2.0+1.9+2.1+1.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.376, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Table 4. Comparison between dual-scene and multi-scene levels in the patch-based data augmentation of our AIScene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.603 + ], + "angle": 0, + "content": "a critical role in our framework, we further analyze the impact of varying \\( n \\) in patch-based data augmentation, with the results presented in Fig. 6 (d). When \\( n \\) is small, it can yield a certain level of enhancement, but it may fail to generate data sufficiently divergent from the original, thereby impeding the effective generation of diverse semantic information between labeled and unlabeled scenes. Conversely, when \\( n \\) is too large, the generated scenes may deviate excessively from the original data, resulting in a loss of essential prior scene information. Empirically, we find that setting \\( n \\) to 18 yields optimal results. Throughout this paper, we consistently use \\( n \\) set to 18 unless explicitly stated otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.634, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.854 + ], + "angle": 0, + "content": "This paper explores scene affinity for semi-supervised Li-DAR semantic segmentation in driving scenes, named AIScene. Specifically, AIScene addresses intra-scene inconsistency through a point erasure strategy, which enhances the learning of pseudo-labeled points. Moreover, it employs a patch-based data augmentation method to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method enriches the semantic diversity of labeled scenes and complements pseudo-labeled scenes. Extensive experiments show that AIScene achieves state-of-the-art results on popular benchmarks. Additionally, the novel erasure strategy and augmentation method can function independently to facilitate semi-supervised LiDAR semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (No.62325111, U22B2011, 62271354)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27387" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.169 + ], + "angle": 0, + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, pages 1-8. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[3] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In CVPR, pages 7020-7030, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[4] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In CVPR, pages 2613-2622, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[5] Yujun Chen, Xin Tan, Zhizhong Zhang, Yanyun Qu, and Yuan Xie. Beyond the label itself: Latent labels enhance semi-supervised point cloud panoptic segmentation. In AAAI, pages 1245-1253, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[8] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[9] Jin Fang, Xinxin Zuo, Dingfu Zhou, Shengze Jin, Sen Wang, and Liangjun Zhang. Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In CVPR, pages 4710-4720, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[10] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE RA-L, 7(2):3795-3802, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[13] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In ICCV, pages 6423-6432, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.885 + ], + "angle": 0, + "content": "[14] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In ICCV, pages 228-240, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.888, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Lasermix for semi-supervised lidar semantic segmentation. In CVPR, pages 21705-21715, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[16] Lingdong Kong, Xiang Xu, Jiawei Ren, Wenwei Zhang, Liang Pan, Kai Chen, Wei Tsang Ooi, and Ziwei Liu. Multimodal data-efficient 3d scene understanding for autonomous driving. arXiv:2405.05258, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[17] Hyeokjun Kweon, Jihun Kim, and Kuk-Jin Yoon. Weakly supervised point cloud semantic segmentation via artificial oracle. In CVPR, pages 3721-3731, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[18] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.279, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, pages 17545-17555, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[20] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[21] Jianan Li and Qiulei Dong. Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In CVPR, pages 3260-3269, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[22] Jianan Li and Qiulei Dong. Lass3d: Language-assisted semi-supervised 3d semantic segmentation with progressive unreliable data exploitation. In ECCV, pages 252-269. Springer, 2025. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[23] Li Li, Hubert PH Shum, and Toby P Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In CVPR, pages 9361-9371, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[24] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In CVPR, pages 8428-8437, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[25] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In CVPR, pages 23819-23828, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, pages 4213-4220. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[27] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In 3DV, pages 116-125. IEEE, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[28] Gilles Puy, Spyros Gidaris, Alexandre Boulch, Oriane Simeoni, Corentin Sautier, Patrick Pérez, Andrei Bursuc, and Renaud Marlet. Three pillars improving vision foundation model distillation for lidar. In CVPR, pages 21519-21529, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Laurenz Reichardt, Nikolas Ebert, and Oliver Wasenmüller. 360deg from a single camera: A few-shot approach for lidar segmentation. In ICCV, pages 1075-1083, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27388" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[32] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Fabio Poiesi, and Elisa Ricci. Compositional semantic mix for domain adaptation in point cloud segmentation. IEEE TPAMI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[33] Shivanand Venkanna Sheshappanavar, Vinit Veerendraveer Singh, and Chandra Kambhametu. Patchaugment: Local neighborhood augmentation in point cloud classification. In ICCV, pages 2118-2127, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[34] Jungwook Shin, Jaeill Kim, Kyungeun Lee, Hyunghun Cho, and Wonjong Rhee. Diversified and realistic 3d augmentation via iterative construction, random placement, and hpr occlusion. In AAAI, pages 2282-2291, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[36] Shihao Su, Jianyun Xu, Huanyu Wang, Zhenwei Miao, Xin Zhan, Dayang Hao, and Xi Li. Pups: Point cloud unified panoptic segmentation. In AAAI, pages 2339-2347, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[37] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[38] Liyao Tang, Zhe Chen, Shanshan Zhao, Chaoyue Wang, and Dacheng Tao. All points matter: Entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[39] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, pages 1195–1204, 2017. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[40] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In CVPR, pages 2697-2707, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[41] Ozan Unal, Dengxin Dai, Lukas Hoyer, Yigit Baran Can, and Luc Van Gool. 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In WACV, pages 7336-7345, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[42] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, pages 4917-4928, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[43] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[44] Yi Wang, Jiaze Wang, Jinpeng Li, Zixu Zhao, Guangyong Chen, Anfeng Liu, and Pheng Ann Heng. Pointpatchmix: Point cloud mixing with patch scoring. In AAAI, pages 5686-5694, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, pages 4376-4382. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[46] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[47] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In NeurIPS, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[48] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In ECCV, pages 1-19. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[49] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[50] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[51] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv:1710.09412, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[52] Jinlai Zhang, Lyujie Chen, Bo Ouyang, Binbin Liu, Jihong Zhu, Yujin Chen, Yanmei Meng, and Danfeng Wu. Pointcutmix: Regularization strategy for point cloud classification. Neurocomputing, 505:58-67, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, pages 9601-9610, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[54] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In AAAI, pages 3421-3429, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[55] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IROS, pages 4453-4458. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[56] Qinfeng Zhu, Lei Fan, and Ningxin Weng. Advancements in point cloud data augmentation for deep learning: A survey. arXiv:2308.12113, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[57] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[58] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, pages 289-305, 2018. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27389" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_origin.pdf b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..af64c221b5452b137e806948ab700794ee9962fe --- /dev/null +++ b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/8e486940-e51f-442d-a6a0-a73d2630afa0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c472604735fda349d50e9a3ffcd44ed0373564d38503619387135560053750a1 +size 1742887 diff --git a/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/full.md b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ade021d948d9f59fd577ced0d1afd84b21fbfd16 --- /dev/null +++ b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/full.md @@ -0,0 +1,334 @@ +# Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation + +Chuandong Liu $^{1,*}$ , Xingxing Weng $^{1,*}$ , Shuguo Jiang $^{1}$ , Pengcheng Li $^{3}$ , Lei Yu $^{2,4}$ , Gui-Song Xia $^{1,2,5,6}$ + +$^{1}$ School of Computer Science, Wuhan University + +$^{2}$ School of Artificial Intelligence, Wuhan University + +$^{3}$ School of Computer Science and Technology, Hainan University + +$^{4}$ School of Electronic Information, Wuhan University + +${}^{5}$ State Key Lab. of LIESMARS, Wuhan University + +$^{6}$ Institute for Math & AI, Wuhan University + +# Abstract + +This paper explores scene affinity (AIScene), namely intrascene consistency and inter-scene correlation, for semisupervised LiDAR semantic segmentation in driving scenes. Adopting teacher-student training, AIScene employs a teacher network to generate pseudo-labeled scenes from unlabeled data, which then supervise the student network's learning. Unlike most methods that include all points in pseudo-labeled scenes for forward propagation but only pseudo-labeled points for backpropagation, AIScene removes points without pseudo-labels, ensuring consistency in both forward and backward propagation within the scene. This simple point erasure strategy effectively prevents unsupervised, semantically ambiguous points (excluded in backpropagation) from affecting the learning of pseudo-labeled points. Moreover, AIScene incorporates patch-based data augmentation, mixing multiple scenes at both scene and instance levels. Compared to existing augmentation techniques that typically perform scene-level mixing between two scenes, our method enhances the semantic diversity of labeled (or pseudo-labeled) scenes, thereby improving the semi-supervised performance of segmentation models. Experiments show that AIScene outperforms previous methods on two popular benchmarks across four settings, achieving notable improvements of $1.9\%$ and $2.1\%$ in the most challenging $1\%$ labeled data. The code will be released at https://github.com/azhuantou/AIScene. + +# 1. Introduction + +Semantic segmentation of LiDAR point clouds is essential for autonomous driving, allowing vehicles to perceive their 3D surroundings. Most state-of-the-art models for + +![](images/def801321dd4c8177946d0f512f60a86825a85b1f08ced5b0721c55c9b4220f6.jpg) +pseudo-labeled scene + +![](images/8668e7fd18e1f46cda9edb1914fe9c633dd144b490b322fed769740e01b97a53.jpg) +intra-scene inconsistency + +![](images/0e97d8aa7fb154af8d157afdc0c41046e42790ebe0dd503e8afdc2ab54d04c36.jpg) + +![](images/616f6ba59fe1ade73aacc311e385c4e308b41102d64f54e5f49976764218bd9c.jpg) +scene-level mixing on two scenes +Figure 1. Top: Pseudo-labeled scenes generated from unlabeled scenes typically contain points with pseudo labels (colored) and without them (gray). Previous methods include all points in forward propagation but only pseudo-labeled points in backward propagation, leading to intra-scene inconsistency. This inconsistency may allow unsupervised, semantically ambiguous points to affect the learning of pseudo-labeled points (see Fig. 7). Bottom: Existing data augmentation techniques suitable for semi-supervised segmentation usually mix two scenes through scene-level operations like concatenation or swapping. Their resulting scenes may be constrained in terms of semantic diversity, potentially lacking objects like bicyclists and buses, which are important in driving scenes. + +![](images/d2220de7a7303f910a842a4fb07e4cf2efb164c978fd7339f75293a20955be67.jpg) + +![](images/e8963cb7d9482a7ce4dfa3d883a98e5d972939e035d312282a8017cde7fcee2b.jpg) +limited semantic diversity + +driving-scene LiDAR segmentation are dependent on high-quality, point-annotated datasets [12, 36, 57]. With numerous objects in driving scenes, densely annotating points is extremely labor-intensive. Consequently, semi-supervised LiDAR semantic segmentation has attracted growing atten + +tion, where models are trained on a limited amount of labeled data supplemented by a large set of unlabeled data. + +The common solution for learning from unlabeled data is to apply teacher-student training, where a teacher network generates pseudo-labeled scenes from unlabeled data, which are then used to supervise the learning of a student network. The student network's objective is to minimize segmentation loss while the teacher network's weights are updated via exponential moving average [39]. To further improve model learning with unlabeled data, existing studies implement designs such as utilizing language-vision models [22], adopting density-guided contrastive learning [21], or trying feature fusion between LiDAR and image data [5]. Despite the progress, two main issues remain to be addressed: + +(1) Given the teacher network's confident prediction, pseudo-labeled scenes typically contain points with and without pseudo-labels. During training, most studies include all points in pseudo-labeled scenes in the student network's forward propagation, but only pseudo-labeled points are involved in backward propagation. Points excluded from backpropagation are unsupervised and, therefore, semantically ambiguous. However, these points still contribute to the forward feature computation for pseudolabeled points, which may interfere with the model's ability to learn from them effectively. In short, the inconsistency between forward and backward propagation within the scene causes ambiguous points to obstruct the model's learning of pseudo-labeled points. Both point-based [12] and voxel-based [7, 57] methods suffer from this issue in feature computation, as illustrated at the top of Fig. 1. + +(2) Data augmentation is an effective practice in semi-supervised LiDAR semantic segmentation [13, 15]. The basic idea is to mix raw labeled scenes and pseudo-labeled scenes to enrich scene semantics. Although impressive advancements have been made [33, 52], they usually select two scenes (one labeled scene and one pseudo-label scene) and then perform scene-level mix through operations like concatenation [27] and swapping [15], which limits the semantic diversity of mixed scenes. As shown at the bottom of Fig. 1, multiple scenes, whether labeled or pseudo-labeled, often share similar spatial layouts. Moreover, in driving scenes, objects such as cars and pedestrians are the primary focus of an autonomous vehicle's attention. This naturally raises the question: can we improve the semantic diversity of scenes by mixing multiple scenes at the instance level? + +To address inconsistency and implement the multi-scene mixing idea, we explore Scene Affinity for semi-supervised LiDAR semantic segmentation, named AIScene. The term scene affinity refers to intra-scene consistency and interscene correlation. Specifically, AIScene, built on the teacher-student training pipeline, employs a simple yet helpful point erasure strategy, removing points without + +pseudo-labels in forward propagation. This ensures intra-scene consistency in both forward and backward propagation, preventing semantically ambiguous points from disrupting learning from pseudo-labeled points. Building on this, AIScene incorporates patch-based data augmentation that mixes multiple scenes at both scene and instance levels. By maintaining two pools that store scene and instance patches from various scenes, it selects patches for mixing based on inter-scene correlation. This novel augmentation method not only enhances the semantic diversity of raw labeled scenes but also complements pseudo-labeled scenes where some points have been erased. + +AIScene has been extensively evaluated through experiments and demonstrated superiority on two popular benchmarks for semi-supervised LiDAR semantic segmentation. It exceeds state-of-the-art methods by $1.3\%$ on the SemanticKITTI [2] and $1.2\%$ on the nuScenes [10] in terms of average performance across four settings (1%, 10%, 20%, and 50% labeled data). Notably, in the most challenging setting, with only $1\%$ labeled data, AIScene achieves remarkable improvements of $1.9\%$ and $2.1\%$ on the two datasets. Additionally, our ablation studies confirm AIScene's potential to reduce the requirements for labeled data across different LiDAR semantic segmentation networks (see Tab. 1 and Tab. 2), with each component of AIScene able to function independently to facilitate semi-supervised LiDAR semantic segmentation (see Fig. 6). + +To summarize, our contributions are as follows: + +- We propose a point erasure strategy to ensure intra-scene consistency and facilitate network learning from pseudolabeled scenes. This strategy can serve as a plug-and-play tool for semi-supervised LiDAR semantic segmentation using the pseudo-label mechanism. +- We design a patch-based data augmentation to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method improves the semantics diversity of labeled scenes and complements pseudolabeled scenes. +- Extensive experiments demonstrate that AIScene achieves state-of-the-art results on popular benchmarks for semi-supervised LiDAR semantic segmentation. Ablation studies show that the novel erasure strategy and augmentation method can function independently. + +# 2. Related Work + +Fully-supervised LiDAR Semantic Segmentation: Numerous approaches for LiDAR semantic segmentation have been explored, which can be broadly grouped into three categories: voxel-based, 2D projection-based, and point-based methods. Voxel-based approaches [6, 7, 19, 37, 57] involve voxelizing sparse point clouds to generate regular voxel grids, allowing the application of standard 3D convolution neural networks for embedding learning. Conversely, + +![](images/c2265fe8a86ad5d13f4b41e40929e4b18f7188fd5d5ec75a9d88c507f5d1f2f2.jpg) +Figure 2. Overview of the proposed AIScene training pipeline. The teacher network generates the pseudo-labels $y_{i}^{u}$ from the unlabeled scene $x_{i}^{u}$ , followed by an erasure operation to filter out low-confidence points, ensuring training consistency. For the labeled scene $x_{i}^{l}$ and the generated erased scene $\hat{x}_i^u$ , we employ patch-based data augmentation to obtain augmented scenes $x_{i}^{lpf}$ and $\hat{x}_i^{upf}$ , respectively. Subsequently, both scenes are fed into the student network to compute the loss of $\mathcal{L}_l$ and $\mathcal{L}_u$ . + +2D projection-based methods [14, 26, 45, 48, 53, 55] project input point clouds into a 2D dense space, use 2D convolutional filters [31] for processing, and then map predictions back to the original 3D space. Finally, point-based methods [12, 18, 29, 46] operate directly on raw point clouds, enabling the extraction of point-wise semantic features. + +Semi-supervised LiDAR Semantic Segmentation: To reduce the need for 3D dense annotations, recent research has explored methods that require less supervision [3, 17, 24, 38, 40, 54]. Semi-supervised learning [4, 16, 21, 23, 30, 58] aims to utilize a large number of unlabeled data to alleviate the annotation burden. LaserMix [15], a pioneer work in semi-supervised LiDAR segmentation, leverages a simple laser beam mix strategy to achieve solid results. Subsequent research fuses LiDAR and camera data to extract deeper semantic information [5, 41]. DDsemi [21], focusing on point cloud density, builds on contrastive learning to enhance feature discriminability. The recently proposed LASS3D [22] leverages large language-vision models to support semi-supervised segmentation. Existing semi-supervised methods often build upon the mean-teacher paradigm [39] and utilize pseudo labels [1]. However, many points without pseudo-labels contribute equally to pseudo-labeled points in forward propagation but are ignored during backward propagation, leading to intra-scene inconsistency. To address this, our AIScene leverages a point erasure strategy to mitigate this inconsistency, complementing the pseudolabel mechanism. + +3D Point Cloud Data Augmentation: Point cloud data augmentation has been explored across various 3D vision tasks [9, 25, 34, 43, 44, 56]. Inspired by the success of Mixup [51] in image classification, several methods have investigated techniques for mixing selected point cloud samples. To preserve structural information, RSMix [20] introduces a neighboring function, and PointCutMix [52] establishes correspondences for sample fusion. Notably, these methods focus on local object-level augmentation. In response, several works have explored mixing strategies for point cloud scene augmentation. PloarMix [47] swaps points within the fan-shaped areas between two scenes, while CoSMix [32] aims to mitigate domain shift by mixing semantic point groups. Unlike these methods, our method is capable of modeling inter-scene information across multiple scenes to create more diversified data, while also extracting instance-level semantic information for individual classes to further enrich scenes. + +# 3. Method + +# 3.1. Preliminary + +Problem Formulation. Before we describe our model and training process, we first provide the definition of semi-supervised point cloud segmentation. In detail, we have a small labeled training set $D_{l} = \{x_{i}^{l},y_{i}^{l}\}_{i = 1}^{N_{l}}$ and a large unlabeled training set $D_{u} = \{x_{i}^{u}\}_{i = 1}^{N_{u}}$ , where $x_{i}\in \mathbb{R}^{m\times (3 + r)}$ is a point cloud scene of $m$ points with three-dimensional + +Cartesian coordinates relative to the sensor and additional $r$ -dimensional feature (e.g., intensity, color, or reflectivity), $y_{i}$ represents its corresponding point-level segmentation labels, $N_{l}$ and $N_{u}$ are the numbers of labeled and unlabeled point cloud scenes, respectively, and $N_{u} \gg N_{l}$ in ideal cases. The goal of SSL-based point cloud segmentation is to learn a segmentation model using limited labeled data and large-scale unlabeled data. + +Teacher-Student Paradigm. Following current mainstream research [15, 21], our framework also builds upon the teacher-student paradigm, which includes two segmentation networks with the same architecture. Our AIScene can be instantiated with any 3D segmentation network [7, 19, 57], denoted as $\Phi(\cdot)$ . In detail, the teacher model generates pseudo-labels for unlabeled scenes and undergoes a gradual update through the exponential moving average (EMA) [39] of the student weights: + +$$ +\theta_ {\mathbb {t}} ^ {k + 1} = \alpha \cdot \theta_ {\mathbb {t}} ^ {k} + (1 - \alpha) \cdot \theta_ {\mathbb {S}} ^ {k}, \tag {1} +$$ + +where $\alpha$ is the EMA decay rate, which is set to 0.99 by default, $\theta_{\mathbb{t}}$ and $\theta_{\mathrm{s}}$ represent the weight parameters of the teacher and student models, respectively, and $k$ denotes the training step. + +# 3.2. Point Erasure Strategy + +Based on the teacher-student mutual learning framework [35], the overall architecture of AIScene is depicted in Fig. 2. Specifically, AIScene consists of two primary branches: one dedicated to labeled scenes and the other to unlabeled scenes. How to effectively leverage a large number of unlabeled scenes is crucial for semi-supervised learning. Hence, for the branch of unlabeled scenes, we use a teacher network to generate pseudo-labeled scenes $(x_{i}^{u},y_{i}^{u})$ based on a predefined confidence threshold $\tau_{s}$ [15]. As mentioned previously, when using the pseudo-labeled scene for training, points without pseudo-labels are involved in the forward propagation but are ignored in the backward propagation due to the lack of corresponding supervision. This leads to intra-scene inconsistency, which causes semantic ambiguity of points. To address this issue, we propose erasing points without pseudo labels, resulting in an erased scene $(\hat{x}_i^u,\hat{y}_i^u)$ that contains fully annotated information, similar to fully supervised scenes. Formally, intrascene inconsistency addressed by the point erasure strategy can be defined as: + +$$ +\hat {x} _ {i} ^ {u} = \left\{x _ {i} ^ {u} \mid \Phi_ {s} \left(x _ {i} ^ {u}\right) \geq \tau_ {s} \right\}, \tag {2} +$$ + +where $\hat{x}_i^u$ is the erased scene with corresponding class labels $\hat{y}_i^u$ , $\Phi_s(\cdot)$ represents the segmentation network that predicts the confidence score for each point in the scene $x_i^u$ , and $\tau_s$ is a predefined confidence threshold, equivalent to the pseudolabel mining threshold, requiring no additional adjustment. + +![](images/88e87573d075d5a475b65e720e6bfd27db4be3c86aa6cd6dd8071821c3aa6b1c.jpg) +Figure 3. Illustration of the pool initialization process. For each labeled point cloud scene, splitting it into multiple patches from the BEV perspective and assigning different instance categories to each patch enables the extraction of instances belonging to different classes. Subsequently, these patches and instances are stored individually in the patch pool and instance pool. + +Since the proposed point erasure strategy is independent of the segmentation network architecture, it can be utilized as a plug-and-play tool for semi-supervised 3D semantic segmentation methods based on the pseudo-label mechanism, as depicted in Fig. 6 (a). + +# 3.3. Patch-based Data Augmentation + +Based on the labeled and erased scenes, we propose a patch-based data augmentation to enhance the semantic diversity of annotation information, guided by inter-scene correlation. The key idea is to construct patch pools to store scene-level and instance-level information, and then combine patches from multiple scenes to generate diverse training samples. + +Pool Initialization. We first use the labeled training set $D_{l}$ to create two patch pools, which store scene-level and instance-level information, respectively, as shown in Fig. 3. Given the $i$ -th point cloud scene $x_{i}^{l}$ , we patchify $x_{i}^{l}$ into a sequence of $T$ regular, non-overlapping point patches $\{p_{ij}^{l}\}_{j=1}^{T}$ , each of size $(n \times n)$ , in the bird's eye view. This process can be expressed as: + +$$ +\left\{p _ {i j} ^ {l} \right\} _ {j = 1} ^ {T} = \text {p a t c h i f y} \left(x _ {i} ^ {l}\right), \tag {3} +$$ + +where $p_{ij}^{l}$ denotes the $j$ -th point patch from the sequence. For all labeled scenes, by performing the operation in Eq. (3), the patch pool $\mathcal{B}_{\mathrm{Pat}}^{l}$ is constructed, which can be defined as: + +$$ +\mathcal {B} _ {\mathrm {P a t}} ^ {l} = \left\{\left\{p _ {i j} ^ {l} \right\} _ {j = 1} ^ {T} \right\} _ {i = 1} ^ {N _ {l}}. \tag {4} +$$ + +Based on the patch pool, we can obtain the instance pool $\mathcal{B}_{\mathrm{Ins}}^l$ by assigning different class labels to each patch, as depicted in Fig. 3, which contains instance information for different categories (e.g., car and person). + +Following the same steps used for labeled scenes, we further create a pseudo patch pool $\mathcal{B}_{\mathrm{Pat}}^u$ and a pseudo instance pool $\mathcal{B}_{\mathrm{Ins}}^u$ from the erased scene $\hat{x}_i^u$ . Unlike $\mathcal{B}_{\mathrm{Pat}}^l$ and $\mathcal{B}_{\mathrm{Ins}}^l$ + +![](images/14ae380cf8b360e30cec537917399150684b8664dff76e0c32c3e3c4f041f302.jpg) +Figure 4. (a) Illustration of the MixPatch process. (b) Illustration of selecting instances from the patch pool for filling. An instance will not be filled if there are no other points around the filling location or if the filled instance overlaps with existing instances. + +![](images/b15d4aa449a510f82e7a03474a437a6dbc82c59f1749a8ed5fbd0a2bef996c45.jpg) + +covering the whole training set $D_{l}$ , the pseudo pools $\mathcal{B}_{\mathrm{Pat}}^{u}$ and $\mathcal{B}_{\mathrm{Ins}}^{u}$ only contain scene and instance information from the current batch and are reset in each iteration. This prevents low-quality pseudo labels from persisting in pools and misleading model training. Additionally, if the number of points within a patch or instance falls below a certain threshold $\tau_{\mathrm{min}}$ , it will not be saved into the pool. In practice, we set $\tau_{\mathrm{min}}$ to 5 [49]. If a patch or instance has fewer than 5 points, its spatial structure exhibits high uncertainty, making it semantically ambiguous. Thus, it is excluded from the pools to prevent noise during data augmentation. + +MixPatch. To model the inter-scene correlation and then guide the transfer of annotation information between labeled scenes and unlabeled scenes, we combine multiple scenes using patch pools. Fig. 4 (a) illustrates the patch mixing (MixPatch) strategy, while Fig. 4 (b) presents the instance filling (InsFill) strategy, which is introduced in the next paragraph. During a training iteration, the mini-batch is comprised of labeled and unlabeled scenes. After the unlabeled scene $x_{i}^{u}$ is transformed into the erased scene $\hat{x}_{i}^{u}$ using the point erasure strategy, we integrate information from the patch pool into $\hat{x}_{i}^{u}$ by MixPatch. The MixPatch process can be described as follows: + +$$ +\hat {x} _ {i} ^ {u p} = \left(\hat {x} _ {i} ^ {u} \odot \mathcal {M}\right) \oplus \left(\mathcal {B} _ {\text {P a t}} ^ {l} \odot (1 - \mathcal {M})\right), \tag {5} +$$ + +where $\mathcal{M}$ is a binary mask indicating which patches from $\hat{x}_i^u$ will be replaced with selected patches of the same patch index from the patch pool, $\odot$ denotes element-wise multiplication, and $\hat{x}_i^{up}$ represents the augmented scene. For labeled scenes, MixPatch is applied based on the current pseudo patch pool $\mathcal{B}_{\mathrm{Pat}}^u$ , which can be expressed as the following: + +$$ +x _ {i} ^ {l p} = \left(x _ {i} ^ {l} \odot (1 - \mathcal {M})\right) \oplus \left(\mathcal {B} _ {\text {P a t}} ^ {u} \odot \mathcal {M}\right), \tag {6} +$$ + +where $\mathcal{M}$ is the same as defined in Eq. (5), thereby ensuring the full utilization of scene information from the unlabeled + +scenes. The binary mask $\mathcal{M}$ and its complement $1 - \mathcal{M}$ are used to merge inter-scene point cloud patches with the erased scene, facilitating the generation of diverse semantic information for labeled and unlabeled scenes. To alleviate bias in the patch mixing process, we apply a uniform sampling strategy to the patches in each scene, ensuring that each patch has an equal opportunity of being selected in the mixed scene. + +InsFill. In addition to patch mixing, we utilize the instances stored in the instance pool $\mathcal{B}_{\mathrm{Ins}}^l$ to further enhance the scene $\hat{x}_i^{up}$ . Given a patch with location index $j$ in the scene, we randomly select an instance from the instance pool $\mathcal{B}_{\mathrm{Ins}}^l$ that was extracted from the $j$ -th patch. Patches with identical indices typically correspond to similar Li-DAR spatial layouts, ensuring consistency between the selected instance and its fill positions in the scene. As depicted in Fig. 4 (b), if the selected instance overlaps with existing instances in the scene or lacks sufficient contextual information, the filling operation is bypassed. This judgment process can be easily implemented using the maximum 3D bounding boxes. This data augmentation provides instance-level semantic information for the point cloud scene, enabling the augmented scene $\hat{x}_i^{upf}$ and corresponding labels $\hat{y}_i^{upf}$ to effectively support semi-supervised training. For the labeled branch, we also apply the same filling strategy as described, selecting instances from the pseudo instance pool $\mathcal{B}_{\mathrm{Ins}}^u$ to fill the current scene. This InsFill process yields the scene $x_i^{lpf}$ along with its corresponding label $y_i^{lpf}$ . + +# 3.4. Model Training + +Following the standard teacher-student mutual learning paradigm [15, 35, 39], we construct a mini-batch of data with an equal composition of labeled and unlabeled scenes at each iteration. The unlabeled scenes are assigned with pseudo-labels by the teacher model and then train the student model jointly with labeled scenes. We use three types of loss functions to supervise the model training here, which is shown as follows: + +$$ +\mathcal {L} _ {s} = \mathcal {L} _ {\text {s e g}} \left(\Phi_ {\theta_ {\mathfrak {s}}} \left(x _ {i} ^ {l}\right), y _ {i} ^ {l}\right), \tag {7} +$$ + +$$ +\mathcal {L} _ {l} = \mathcal {L} _ {\text {s e g}} \left(\Phi_ {\theta_ {\mathrm {s}}} \left(x _ {i} ^ {l p f}\right), y _ {i} ^ {l p f}\right), \tag {8} +$$ + +$$ +\mathcal {L} _ {u} = \mathcal {L} _ {s e g} \left(\Phi_ {\theta_ {\mathrm {s}}} \left(\hat {x} _ {i} ^ {u p f}\right), \hat {y} _ {i} ^ {u p f}\right), \tag {9} +$$ + +where $\mathcal{L}_s, \mathcal{L}_l$ , and $\mathcal{L}_u$ denote the segmentation loss for the labeled scene $x_i^l$ , augmented labeled scene $x_i^{lpf}$ , and augmented unlabeled scene $\hat{x}_i^{upf}$ , respectively, while $\mathcal{L}_{seg}$ is the original loss of the segmentation network $\Phi(\cdot)$ . We sum up all the loss term to define the final training objective as: + +$$ +\mathcal {L} = \mathcal {L} _ {s} + \lambda_ {u} \mathcal {L} _ {u} + \lambda_ {l} \mathcal {L} _ {l}, \tag {10} +$$ + +where $\lambda_{u}$ and $\lambda_{l}$ are loss weights that balance the two loss terms. After that, we update the teacher model weights us + +
MethodVenueSemanticKITTI [2]Avg.nuScenes [10]Avg.
1%10%20%50%1%10%20%50%
Meanteacher [39]NeurIPS'201745.457.159.260.055.451.666.067.171.764.1
CBST [58]ECCV'201848.858.359.459.756.653.066.569.671.665.2
CPS [4]CVPR'202146.758.759.660.556.452.966.370.072.565.4
LaserMix (Range View) [15]CVPR'202343.458.859.461.455.849.568.270.673.065.3
LaserMix (Voxel) [15]CVPR'202350.660.061.962.358.755.369.971.873.267.6
GPC [13]ICCV'202154.162.062.562.860.4-----
IGNet [41]WACV'202449.061.363.164.859.6-----
LiM3D [23]CVPR'202358.462.263.163.661.8-----
LASS3D [22]ECCV'202458.563.064.164.562.5-----
DDSemi [21]CVPR'202459.365.166.367.064.458.170.274.076.569.7
Ours (AIScene)-61.266.367.467.965.760.272.375.076.470.9
Improvement ↑-+1.9+1.2+1.1+0.9+1.3+2.1+2.1+1.0-0.1+1.2
+ +Table 1. Experimental results on the SemanticKITTI and nuScenes dataset compared with different SSL methods. The results are reported with mIoU scores (\%). The bold and underline scores respectively denote the best and second results. + +ing the EMA strategy described in Eq. (1). + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +SemanticKITTI. Following state-of-the-art methods [15, 21], we evaluate our AISCene on the SemanticKITTI dataset [2], which is a large-scale segmentation dataset consisting of LiDAR data derived from the popular KITTI dataset [11]. It is composed of 22 sequences, with sequences 00-10 serving as the train set (of which 08 as the val set) and sequences 11-21 as the test set. Moreover, all the point cloud scenes are collected from real-world rural and urban road environments in Germany, and the semantic class labels are mapped to 19 classes in our experiments. + +nuScenes. The nuScenes dataset [10] is another well-known large autonomous driving dataset. This dataset is composed of 1000 sequences of 20s duration (850 for training and validation, 150 for testing), which are captured by the LiDAR sensor from Boston and Singapore. Following the official splitting protocol, the train and val sets contain 28,130 and 6,019 point cloud scenes, respectively, including per-points annotation for 16 semantic classes. + +Evaluation Metrics. For both the SemanticKITTI and nuScenes datasets, we sample different ratios of annotated train scenes to obtain semi-supervised data. Meanwhile, we report the performance on the validation set using the mean Intersection over Union (mIoU) metric [21, 23]. + +# 4.2. Implementation Details + +We conduct experiments with $4 \times$ NVIDIA V100 GPUs with 32GB RAM, based on MMDetection3D [8]. For the threshold $\tau_{s}$ for pseudo-label generation and point erasure, we experimentally find that 0.9 yields relatively satisfactory results. We use the MinkowskiNet [7] and Cylinder3D [57] as the backbone network due to its widely adopted in previous methods [15, 28, 42]. To facilitate patch-based data + +augmentation, we restrict our focus to point clouds within the X and Y dimensions in the range of $[-50, 50]m$ for both the SemanticKITTI and nuScenes datasets. Similar to mainstream work, we also adopt the consistency loss proposed in MeanTeacher [39] to align the predictions of the teacher and student networks. Additionally, the weights of the consistency loss are assigned as $1 \times 10^{3}$ and $2 \times 10^{3}$ for the SemanticKITTI and nuScenes datasets, respectively. The weights for $\lambda_{u}$ and $\lambda_{l}$ are set as 1 for both datasets. Our method excludes all forms of model ensemble and test-time augmentation during the evaluation. + +# 4.3. Main Results + +In this subsection, we compare our method with current SOTA semi-supervised point cloud segmentation methods (e.g., LaserMix [15], LiM3D [23] and DDsemi [21]), using sub-sampling of the datasets under various partition protocols [21]. Specifically, we split the entire training set into labeled subsets with ratios of $1\%$ , $10\%$ , $20\%$ , and $50\%$ in the SemanticKITTI [2] and nuScenes dataset [10]. For further comparison, we also include several classical semi-supervised learning methods [4, 39, 58] from the 2D image domain adapted to 3D segmentation. + +Tab. 1 summarizes the quantitative results on SemanticKITTI and nuScenes datasets. Our method consistently outperforms other comparative methods on both datasets, achieving average mIoU improvements of $1.3\%$ and $1.2\%$ , respectively. Notably, the proposed method demonstrates even greater gains on smaller sampling splits, such as $1\%$ and $10\%$ , which are particularly beneficial for the semi-supervised setting. Specifically, with a $1\%$ ratio, our method shows mIoU improvements of up to $1.9\%$ and $2.1\%$ on the two datasets, respectively, compared to the recently proposed SOTA method, DDsemi [21]. These results also validate that our method is less sensitive to reduced labeled data sampling than other methods. Furthermore, we present qualitative results in Fig. 5. As indicated by the + +![](images/effddf6ca379569f775d87cef2e34128b867ab0cf25ad621a9b973ae7717b936.jpg) + +![](images/f5f8f271224715696ef40a16a8b1fe3119772ca7c6b7e19b5cc27828c29e940c.jpg) + +![](images/24b274b44b46e6df8d1f622a281a78da544e9b0ec4a3cfd6fd3bbee7df68a8ed.jpg) + +![](images/f30d5b9e6d29c1e0dad9788e69bbb7aa3490ff9deefa293c2664c459170ddedf.jpg) + +![](images/7c6c6bba31f1e1a3e6f91667349620a7dff637e5e9ad7a6479bf06785397dae9.jpg) + +![](images/b4ad28930e5c24a1c3a7be72f42e918babf091b018fe4ab6e23e22175728af48.jpg) + +![](images/a6b546960433769d186988177e8b8bbd90eab545899925b2830433ae76184f8a.jpg) +Figure 5. Qualitative comparison results from LiDAR bird's eye view on SemanticKITTI val set. We set the point with correct and incorrect predicted semantic class in blue and orange for better visualization. Best viewed in color. + +![](images/2dc2d0e117ed974f7ad40beb905ac71fb8eca19aa2e2e50237c8d326c9574d5c.jpg) + +![](images/6235198668efd37c506bd210494a39694a506d80792564ea57f657f35248e3e9.jpg) + +purple and red dashed circles, our method achieves better segmentation of boundary regions and small objects. + +# 4.4. Ablation Study + +In this subsection, We first validate the adaptability of our method to different segmentation models. We then examine the impact of intra-scene consistency through the point erasure strategy and explore inter-scene correlation by incorporating patch-based mixing and patch-based instance filling data augmentation. + +Adaptability with Different Segmentation Methods. Our AISCene is segmentor-agnostic and can be easily used to further improve off-the-shelf 3D fully-supervised segmentors under the semi-supervised setting. Tab. 2 shows that the proposed method can substantially boost the performance with Cylinder3D [57] as the backbone, indicating the generalization ability of our framework. For consistency in comparison, the subsequent experiments are based on Cylinder3D unless otherwise specified. + +Effectiveness of Point Erasure Strategy. The $1^{\text{st}}$ row in Tab. 3 reports the performance of the supervision-only Cylinder3D model [57], which serves as the baseline. When PtErase is applied to the baseline, it effectively boosts performance by ensuring consistency during forward and backward propagation, thus addressing the lim + +
DatasetsMethods1%10%20%50%Avg.
SemanticKITTICylinder3D [57]45.456.157.858.754.5
Ours (AIScene)54.563.363.764.361.5
Improvement+9.1+7.2+5.9+5.6+7.0
nuScenesCylinder3D [57]50.965.966.671.263.7
Ours (AIScene)56.670.272.873.968.4
Improvement+5.7+4.3+6.2+2.7+4.7
+ +Table 2. Ablation study on the adaptability of our AIScene using Cylinder3D as the segmentation model. + +
Ours AISCeneptEraseMixPatchInsFillSplits on SemanticKITTIAvg.
1%10%20%50%
45.456.157.858.754.5
✓51.759.861.860.458.4
✓51.361.460.760.558.5
✓52.361.760.161.158.8
✓✓54.661.062.962.860.3
✓✓52.061.760.862.659.3
✓✓50.562.563.863.660.1
✓✓✓54.563.363.764.361.5
+ +Table 3. Ablation study of different components of our AIScene framework. PtErase refers to the point erasure strategy, while MixPatch and InsFill refer to the patch mix and instance filling data augmentation, respectively. + +ditions of the pseudo-label mechanism. This consistency is achieved through a plug-and-play point erasure strategy, which can be easily integrated into methods relying on the pseudo-label mechanism. To further validate the importance of maintaining consistency during forward and backward propagation, we embedded the PtErase strategy into LaserMix [15]. As illustrated in Fig. 6 (a), using the PtErase strategy improves performance across all ratios, underscoring the significance of intra-scene consistency and demonstrating the generalizability of our method. Moreover, Fig. 6 (b) shows that the proportion of points removed by the erasure strategy is initially high at the start of training, but gradually decreases as the network converges. Furthermore, We present a comparison of class probability distributions in Fig. 7 to validate the effectiveness of the point erasure strategy. The results demonstrate that this strategy reduces classification ambiguity, enabling the model to better capture the features of the correct class. + +Effectiveness of Patch-based Data Augmentation. The patch-based data augmentation consists of MixPatch (patch-based mixing) and InsFill (instance filling). In Tab. 3, the results in the $3^{\mathrm{rd}}$ and $4^{\mathrm{th}}$ rows indicate that using MixPatch or InsFill individually yields an average mIoU improvement of $4.0\%$ or $4.3\%$ , respectively, over the baseline. This suggests that incorporating diverse semantic information across scenes benefits semi-supervised LiDAR segmentation. Furthermore, as shown in the $5^{\mathrm{th}}$ and $6^{\mathrm{th}}$ rows, combining MixPatch or InsFill with PtErase significantly enhances performance, highlighting the effectiveness of patch-based data augmentation. The combination of all + +![](images/443f65983496f89a861c533cedc6247d0287a0a339650b8586a44767b443dbec.jpg) +Figure 6. (a) Illustration of performance differences whether using point erasure strategy on the LaserMix [15]. (b) Variation in the proportion of points removed by the point erasure strategy throughout the training process. (c) Comparison of different mixing-based techniques on SemanticKITTI dataset. (d) Results of mIoU against the number of patch split $n$ under the $10\%$ SemanticKITTI split. + +![](images/57f2f0d75921826a5c67b8ca4b5dd8175e66ccbdb5a34df49a31fa4ab3d93e93.jpg) +Figure 7. The left chart shows the initial output with distributed probabilities across multiple classes, indicating uncertainty in class prediction. Through our point erasure strategy, the right chart exhibits a distribution with a dominant class ("person") having a significantly higher probability, + +three components yields substantial performance gains, as evidenced by the last row. Additionally, the patch pool and instance pool within the patch-based data augmentation facilitate the fusion of diverse semantic information across multiple scenes. To verify the effectiveness of interactions at the multi-scene level, we modify the pool implementation to allow interactions between only two scenes. The results in Tab. 4 indicate that multi-scene interactions yield superior performance compared to dual-scene interactions. + +Different Mixing Strategies. Fig. 6 (c) compares our patch-based data augmentation with existing mixing methods [27, 47, 50]. MixUp [51] shows a slight improvement over the baseline, likely due to its limited capacity for information transformation. Note that we reproduce Cylinder-Mix in the LiDAR branch [5] for comparison. Other methods demonstrate considerable improvement over the baseline by enriching point cloud distributions. In contrast, our AIScene achieves the most substantial gains by effectively leveraging inter-scene correlation, highlighting its effectiveness in facilitating diverse semantic information generation for labeled and unlabeled scenes. This is achieved through an innovative design that incorporates a patch pool and an instance pool spanning multiple scenes. + +Patch Split Number. As the patch split number $n$ plays + +
Augmentation TypeMethodsSplits on SemanticAvg.
1%10%20%50%
dual-scene levelOurs53.461.361.862.259.7
multi-scene levelOurs54.563.363.764.361.5
Improvement-+1.1+2.0+1.9+2.1+1.8
+ +Table 4. Comparison between dual-scene and multi-scene levels in the patch-based data augmentation of our AIScene. + +a critical role in our framework, we further analyze the impact of varying $n$ in patch-based data augmentation, with the results presented in Fig. 6 (d). When $n$ is small, it can yield a certain level of enhancement, but it may fail to generate data sufficiently divergent from the original, thereby impeding the effective generation of diverse semantic information between labeled and unlabeled scenes. Conversely, when $n$ is too large, the generated scenes may deviate excessively from the original data, resulting in a loss of essential prior scene information. Empirically, we find that setting $n$ to 18 yields optimal results. Throughout this paper, we consistently use $n$ set to 18 unless explicitly stated otherwise. + +# 5. Conclusion + +This paper explores scene affinity for semi-supervised Li-DAR semantic segmentation in driving scenes, named AIScene. Specifically, AIScene addresses intra-scene inconsistency through a point erasure strategy, which enhances the learning of pseudo-labeled points. Moreover, it employs a patch-based data augmentation method to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method enriches the semantic diversity of labeled scenes and complements pseudo-labeled scenes. Extensive experiments show that AIScene achieves state-of-the-art results on popular benchmarks. Additionally, the novel erasure strategy and augmentation method can function independently to facilitate semi-supervised LiDAR semantic segmentation. + +Acknowledgments. This work is supported by the National Natural Science Foundation of China (No.62325111, U22B2011, 62271354). + +# References + +[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, pages 1-8. IEEE, 2020. 3 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6 +[3] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In CVPR, pages 7020-7030, 2023. 3 +[4] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In CVPR, pages 2613-2622, 2021. 3, 6 +[5] Yujun Chen, Xin Tan, Zhizhong Zhang, Yanyun Qu, and Yuan Xie. Beyond the label itself: Latent labels enhance semi-supervised point cloud panoptic segmentation. In AAAI, pages 1245-1253, 2024. 2, 3, 8 +[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2 +[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 4, 6 +[8] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6 +[9] Jin Fang, Xinxin Zuo, Dingfu Zhou, Shengze Jin, Sen Wang, and Liangjun Zhang. Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In CVPR, pages 4710-4720, 2021. 3 +[10] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE RA-L, 7(2):3795-3802, 2022. 2, 6 +[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6 +[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 1, 2, 3 +[13] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In ICCV, pages 6423-6432, 2021. 2, 6 +[14] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In ICCV, pages 228-240, 2023. 3 +[15] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. + +Lasermix for semi-supervised lidar semantic segmentation. In CVPR, pages 21705-21715, 2023. 2, 3, 4, 5, 6, 7, 8 +[16] Lingdong Kong, Xiang Xu, Jiawei Ren, Wenwei Zhang, Liang Pan, Kai Chen, Wei Tsang Ooi, and Ziwei Liu. Multimodal data-efficient 3d scene understanding for autonomous driving. arXiv:2405.05258, 2024. 3 +[17] Hyeokjun Kweon, Jihun Kim, and Kuk-Jin Yoon. Weakly supervised point cloud semantic segmentation via artificial oracle. In CVPR, pages 3721-3731, 2024. 3 +[18] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 3 +[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, pages 17545-17555, 2023. 2, 4 +[20] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3 +[21] Jianan Li and Qiulei Dong. Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In CVPR, pages 3260-3269, 2024. 2, 3, 4, 6 +[22] Jianan Li and Qiulei Dong. Lass3d: Language-assisted semi-supervised 3d semantic segmentation with progressive unreliable data exploitation. In ECCV, pages 252-269. Springer, 2025. 2, 3, 6 +[23] Li Li, Hubert PH Shum, and Toby P Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In CVPR, pages 9361-9371, 2023. 3, 6 +[24] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In CVPR, pages 8428-8437, 2022. 3 +[25] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In CVPR, pages 23819-23828, 2023. 3 +[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, pages 4213-4220. IEEE, 2019. 3 +[27] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In 3DV, pages 116-125. IEEE, 2021. 2, 8 +[28] Gilles Puy, Spyros Gidaris, Alexandre Boulch, Oriane Simeoni, Corentin Sautier, Patrick Pérez, Andrei Bursuc, and Renaud Marlet. Three pillars improving vision foundation model distillation for lidar. In CVPR, pages 21519-21529, 2024. 6 +[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 3 +[30] Laurenz Reichardt, Nikolas Ebert, and Oliver Wasenmüller. 360deg from a single camera: A few-shot approach for lidar segmentation. In ICCV, pages 1075-1083, 2023. 3 + +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3 +[32] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Fabio Poiesi, and Elisa Ricci. Compositional semantic mix for domain adaptation in point cloud segmentation. IEEE TPAMI, 2023. 3 +[33] Shivanand Venkanna Sheshappanavar, Vinit Veerendraveer Singh, and Chandra Kambhametu. Patchaugment: Local neighborhood augmentation in point cloud classification. In ICCV, pages 2118-2127, 2021. 2 +[34] Jungwook Shin, Jaeill Kim, Kyungeun Lee, Hyunghun Cho, and Wonjong Rhee. Diversified and realistic 3d augmentation via iterative construction, random placement, and hpr occlusion. In AAAI, pages 2282-2291, 2023. 3 +[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 4, 5 +[36] Shihao Su, Jianyun Xu, Huanyu Wang, Zhenwei Miao, Xin Zhan, Dayang Hao, and Xi Li. Pups: Point cloud unified panoptic segmentation. In AAAI, pages 2339-2347, 2023. 1 +[37] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702. Springer, 2020. 2 +[38] Liyao Tang, Zhe Chen, Shanshan Zhao, Chaoyue Wang, and Dacheng Tao. All points matter: Entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS, 36, 2024. 3 +[39] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, pages 1195–1204, 2017. 2, 3, 4, 5, 6 +[40] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In CVPR, pages 2697-2707, 2022. 3 +[41] Ozan Unal, Dengxin Dai, Lukas Hoyer, Yigit Baran Can, and Luc Van Gool. 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In WACV, pages 7336-7345, 2024. 3, 6 +[42] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, pages 4917-4928, 2024. 6 +[43] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 3 +[44] Yi Wang, Jiaze Wang, Jinpeng Li, Zixu Zhao, Guangyong Chen, Anfeng Liu, and Pheng Ann Heng. Pointpatchmix: Point cloud mixing with patch scoring. In AAAI, pages 5686-5694, 2024. 3 +[45] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and + +unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, pages 4376-4382. IEEE, 2019. 3 +[46] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 3 +[47] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In NeurIPS, 2022. 3, 8 +[48] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In ECCV, pages 1-19. Springer, 2020. 3 +[49] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 5 +[50] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 8 +[51] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv:1710.09412, 2017. 3, 8 +[52] Jinlai Zhang, Lyujie Chen, Bo Ouyang, Binbin Liu, Jihong Zhu, Yujin Chen, Yanmei Meng, and Danfeng Wu. Pointcutmix: Regularization strategy for point cloud classification. Neurocomputing, 505:58-67, 2022. 2, 3 +[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, pages 9601-9610, 2020. 3 +[54] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In AAAI, pages 3421-3429, 2021. 3 +[55] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IROS, pages 4453-4458. IEEE, 2021. 3 +[56] Qinfeng Zhu, Lei Fan, and Ningxin Weng. Advancements in point cloud data augmentation for deep learning: A survey. arXiv:2308.12113, 2023. 3 +[57] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1, 2, 4, 6, 7, 8 +[58] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, pages 289-305, 2018. 3, 6 \ No newline at end of file diff --git a/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/images.zip b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0313e8ba4f60ec03e3049449239e2c11a37dd512 --- /dev/null +++ b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d91291939db355d29542aed23b27e80530b4d16e66eddab5db4cf14eaf1004e4 +size 674150 diff --git a/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/layout.json b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f3a4e8b337106dc591aa01aa8c21ceea8ac38366 --- /dev/null +++ b/2025/Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation/layout.json @@ -0,0 +1,9885 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 543, + 121 + ], + "type": "text", + "content": "Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "spans": [ + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "text", + "content": "Chuandong Liu" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "text", + "content": ", Xingxing Weng" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "text", + "content": ", Shuguo Jiang" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "text", + "content": ", Pengcheng Li" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "text", + "content": ", Lei Yu" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "text", + "content": ", Gui-Song Xia" + }, + { + "bbox": [ + 51, + 149, + 580, + 166 + ], + "type": "inline_equation", + "content": "^{1,2,5,6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 165, + 440, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 165, + 440, + 178 + ], + "spans": [ + { + "bbox": [ + 203, + 165, + 440, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 203, + 165, + 440, + 178 + ], + "type": "text", + "content": "School of Computer Science, Wuhan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 195, + 178, + 446, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 178, + 446, + 193 + ], + "spans": [ + { + "bbox": [ + 195, + 178, + 446, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 195, + 178, + 446, + 193 + ], + "type": "text", + "content": "School of Artificial Intelligence, Wuhan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 163, + 193, + 478, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 193, + 478, + 206 + ], + "spans": [ + { + "bbox": [ + 163, + 193, + 478, + 206 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 163, + 193, + 478, + 206 + ], + "type": "text", + "content": "School of Computer Science and Technology, Hainan University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 206, + 449, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 206, + 449, + 220 + ], + "spans": [ + { + "bbox": [ + 192, + 206, + 449, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 192, + 206, + 449, + 220 + ], + "type": "text", + "content": "School of Electronic Information, Wuhan University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 197, + 220, + 443, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 220, + 443, + 235 + ], + "spans": [ + { + "bbox": [ + 197, + 220, + 443, + 235 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 197, + 220, + 443, + 235 + ], + "type": "text", + "content": " State Key Lab. of LIESMARS, Wuhan University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 212, + 235, + 425, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 235, + 425, + 248 + ], + "spans": [ + { + "bbox": [ + 212, + 235, + 425, + 248 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 212, + 235, + 425, + 248 + ], + "type": "text", + "content": "Institute for Math & AI, Wuhan University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 276, + 200, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 276, + 200, + 289 + ], + "spans": [ + { + "bbox": [ + 152, + 276, + 200, + 289 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "text", + "content": "This paper explores scene affinity (AIScene), namely intrascene consistency and inter-scene correlation, for semisupervised LiDAR semantic segmentation in driving scenes. Adopting teacher-student training, AIScene employs a teacher network to generate pseudo-labeled scenes from unlabeled data, which then supervise the student network's learning. Unlike most methods that include all points in pseudo-labeled scenes for forward propagation but only pseudo-labeled points for backpropagation, AIScene removes points without pseudo-labels, ensuring consistency in both forward and backward propagation within the scene. This simple point erasure strategy effectively prevents unsupervised, semantically ambiguous points (excluded in backpropagation) from affecting the learning of pseudo-labeled points. Moreover, AIScene incorporates patch-based data augmentation, mixing multiple scenes at both scene and instance levels. Compared to existing augmentation techniques that typically perform scene-level mixing between two scenes, our method enhances the semantic diversity of labeled (or pseudo-labeled) scenes, thereby improving the semi-supervised performance of segmentation models. Experiments show that AIScene outperforms previous methods on two popular benchmarks across four settings, achieving notable improvements of " + }, + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "text", + "content": " in the most challenging " + }, + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 301, + 296, + 613 + ], + "type": "text", + "content": " labeled data. The code will be released at https://github.com/azhuantou/AIScene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "Semantic segmentation of LiDAR point clouds is essential for autonomous driving, allowing vehicles to perceive their 3D surroundings. Most state-of-the-art models for" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 277, + 431, + 358 + ], + "blocks": [ + { + "bbox": [ + 318, + 277, + 431, + 358 + ], + "lines": [ + { + "bbox": [ + 318, + 277, + 431, + 358 + ], + "spans": [ + { + "bbox": [ + 318, + 277, + 431, + 358 + ], + "type": "image", + "image_path": "def801321dd4c8177946d0f512f60a86825a85b1f08ced5b0721c55c9b4220f6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 359, + 407, + 367 + ], + "lines": [ + { + "bbox": [ + 343, + 359, + 407, + 367 + ], + "spans": [ + { + "bbox": [ + 343, + 359, + 407, + 367 + ], + "type": "text", + "content": "pseudo-labeled scene" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 441, + 278, + 550, + 357 + ], + "blocks": [ + { + "bbox": [ + 441, + 278, + 550, + 357 + ], + "lines": [ + { + "bbox": [ + 441, + 278, + 550, + 357 + ], + "spans": [ + { + "bbox": [ + 441, + 278, + 550, + 357 + ], + "type": "image", + "image_path": "8668e7fd18e1f46cda9edb1914fe9c633dd144b490b322fed769740e01b97a53.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 359, + 533, + 367 + ], + "lines": [ + { + "bbox": [ + 458, + 359, + 533, + 367 + ], + "spans": [ + { + "bbox": [ + 458, + 359, + 533, + 367 + ], + "type": "text", + "content": "intra-scene inconsistency" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 318, + 376, + 432, + 417 + ], + "blocks": [ + { + "bbox": [ + 318, + 376, + 432, + 417 + ], + "lines": [ + { + "bbox": [ + 318, + 376, + 432, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 376, + 432, + 417 + ], + "type": "image", + "image_path": "0e97d8aa7fb154af8d157afdc0c41046e42790ebe0dd503e8afdc2ab54d04c36.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 319, + 418, + 432, + 472 + ], + "blocks": [ + { + "bbox": [ + 319, + 418, + 432, + 472 + ], + "lines": [ + { + "bbox": [ + 319, + 418, + 432, + 472 + ], + "spans": [ + { + "bbox": [ + 319, + 418, + 432, + 472 + ], + "type": "image", + "image_path": "616f6ba59fe1ade73aacc311e385c4e308b41102d64f54e5f49976764218bd9c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 472, + 425, + 481 + ], + "lines": [ + { + "bbox": [ + 326, + 472, + 425, + 481 + ], + "spans": [ + { + "bbox": [ + 326, + 472, + 425, + 481 + ], + "type": "text", + "content": "scene-level mixing on two scenes" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 491, + 555, + 635 + ], + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 635 + ], + "type": "text", + "content": "Figure 1. Top: Pseudo-labeled scenes generated from unlabeled scenes typically contain points with pseudo labels (colored) and without them (gray). Previous methods include all points in forward propagation but only pseudo-labeled points in backward propagation, leading to intra-scene inconsistency. This inconsistency may allow unsupervised, semantically ambiguous points to affect the learning of pseudo-labeled points (see Fig. 7). Bottom: Existing data augmentation techniques suitable for semi-supervised segmentation usually mix two scenes through scene-level operations like concatenation or swapping. Their resulting scenes may be constrained in terms of semantic diversity, potentially lacking objects like bicyclists and buses, which are important in driving scenes." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 441, + 376, + 551, + 408 + ], + "blocks": [ + { + "bbox": [ + 441, + 376, + 551, + 408 + ], + "lines": [ + { + "bbox": [ + 441, + 376, + 551, + 408 + ], + "spans": [ + { + "bbox": [ + 441, + 376, + 551, + 408 + ], + "type": "image", + "image_path": "d2220de7a7303f910a842a4fb07e4cf2efb164c978fd7339f75293a20955be67.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 441, + 410, + 551, + 472 + ], + "blocks": [ + { + "bbox": [ + 441, + 410, + 551, + 472 + ], + "lines": [ + { + "bbox": [ + 441, + 410, + 551, + 472 + ], + "spans": [ + { + "bbox": [ + 441, + 410, + 551, + 472 + ], + "type": "image", + "image_path": "e8963cb7d9482a7ce4dfa3d883a98e5d972939e035d312282a8017cde7fcee2b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 472, + 533, + 481 + ], + "lines": [ + { + "bbox": [ + 456, + 472, + 533, + 481 + ], + "spans": [ + { + "bbox": [ + 456, + 472, + 533, + 481 + ], + "type": "text", + "content": "limited semantic diversity" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "driving-scene LiDAR segmentation are dependent on high-quality, point-annotated datasets [12, 36, 57]. With numerous objects in driving scenes, densely annotating points is extremely labor-intensive. Consequently, semi-supervised LiDAR semantic segmentation has attracted growing atten" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 701, + 225, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 701, + 225, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 701, + 225, + 713 + ], + "type": "text", + "content": "* Equal contribution. " + }, + { + "bbox": [ + 70, + 701, + 225, + 713 + ], + "type": "inline_equation", + "content": "{}^{ \\ominus }" + }, + { + "bbox": [ + 70, + 701, + 225, + 713 + ], + "type": "text", + "content": " Corresponding authors." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27380" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "tion, where models are trained on a limited amount of labeled data supplemented by a large set of unlabeled data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 294, + 252 + ], + "type": "text", + "content": "The common solution for learning from unlabeled data is to apply teacher-student training, where a teacher network generates pseudo-labeled scenes from unlabeled data, which are then used to supervise the learning of a student network. The student network's objective is to minimize segmentation loss while the teacher network's weights are updated via exponential moving average [39]. To further improve model learning with unlabeled data, existing studies implement designs such as utilizing language-vision models [22], adopting density-guided contrastive learning [21], or trying feature fusion between LiDAR and image data [5]. Despite the progress, two main issues remain to be addressed:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 255, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 447 + ], + "type": "text", + "content": "(1) Given the teacher network's confident prediction, pseudo-labeled scenes typically contain points with and without pseudo-labels. During training, most studies include all points in pseudo-labeled scenes in the student network's forward propagation, but only pseudo-labeled points are involved in backward propagation. Points excluded from backpropagation are unsupervised and, therefore, semantically ambiguous. However, these points still contribute to the forward feature computation for pseudolabeled points, which may interfere with the model's ability to learn from them effectively. In short, the inconsistency between forward and backward propagation within the scene causes ambiguous points to obstruct the model's learning of pseudo-labeled points. Both point-based [12] and voxel-based [7, 57] methods suffer from this issue in feature computation, as illustrated at the top of Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "content": "(2) Data augmentation is an effective practice in semi-supervised LiDAR semantic segmentation [13, 15]. The basic idea is to mix raw labeled scenes and pseudo-labeled scenes to enrich scene semantics. Although impressive advancements have been made [33, 52], they usually select two scenes (one labeled scene and one pseudo-label scene) and then perform scene-level mix through operations like concatenation [27] and swapping [15], which limits the semantic diversity of mixed scenes. As shown at the bottom of Fig. 1, multiple scenes, whether labeled or pseudo-labeled, often share similar spatial layouts. Moreover, in driving scenes, objects such as cars and pedestrians are the primary focus of an autonomous vehicle's attention. This naturally raises the question: can we improve the semantic diversity of scenes by mixing multiple scenes at the instance level?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "To address inconsistency and implement the multi-scene mixing idea, we explore Scene Affinity for semi-supervised LiDAR semantic segmentation, named AIScene. The term scene affinity refers to intra-scene consistency and interscene correlation. Specifically, AIScene, built on the teacher-student training pipeline, employs a simple yet helpful point erasure strategy, removing points without" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "pseudo-labels in forward propagation. This ensures intra-scene consistency in both forward and backward propagation, preventing semantically ambiguous points from disrupting learning from pseudo-labeled points. Building on this, AIScene incorporates patch-based data augmentation that mixes multiple scenes at both scene and instance levels. By maintaining two pools that store scene and instance patches from various scenes, it selects patches for mixing based on inter-scene correlation. This novel augmentation method not only enhances the semantic diversity of raw labeled scenes but also complements pseudo-labeled scenes where some points have been erased." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "text", + "content": "AIScene has been extensively evaluated through experiments and demonstrated superiority on two popular benchmarks for semi-supervised LiDAR semantic segmentation. It exceeds state-of-the-art methods by " + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "text", + "content": " on the SemanticKITTI [2] and " + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "text", + "content": " on the nuScenes [10] in terms of average performance across four settings (1%, 10%, 20%, and 50% labeled data). Notably, in the most challenging setting, with only " + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "text", + "content": " labeled data, AIScene achieves remarkable improvements of " + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 395 + ], + "type": "text", + "content": " on the two datasets. Additionally, our ablation studies confirm AIScene's potential to reduce the requirements for labeled data across different LiDAR semantic segmentation networks (see Tab. 1 and Tab. 2), with each component of AIScene able to function independently to facilitate semi-supervised LiDAR semantic segmentation (see Fig. 6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 396, + 518, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 396, + 518, + 406 + ], + "spans": [ + { + "bbox": [ + 326, + 396, + 518, + 406 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 553, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 407, + 553, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 466 + ], + "type": "text", + "content": "- We propose a point erasure strategy to ensure intra-scene consistency and facilitate network learning from pseudolabeled scenes. This strategy can serve as a plug-and-play tool for semi-supervised LiDAR semantic segmentation using the pseudo-label mechanism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 467, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 553, + 525 + ], + "type": "text", + "content": "- We design a patch-based data augmentation to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method improves the semantics diversity of labeled scenes and complements pseudolabeled scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 526, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 553, + 586 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that AIScene achieves state-of-the-art results on popular benchmarks for semi-supervised LiDAR semantic segmentation. Ablation studies show that the novel erasure strategy and augmentation method can function independently." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Fully-supervised LiDAR Semantic Segmentation: Numerous approaches for LiDAR semantic segmentation have been explored, which can be broadly grouped into three categories: voxel-based, 2D projection-based, and point-based methods. Voxel-based approaches [6, 7, 19, 37, 57] involve voxelizing sparse point clouds to generate regular voxel grids, allowing the application of standard 3D convolution neural networks for embedding learning. Conversely," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27381" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 547, + 311 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 547, + 311 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 547, + 311 + ], + "type": "image", + "image_path": "c2265fe8a86ad5d13f4b41e40929e4b18f7188fd5d5ec75a9d88c507f5d1f2f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed AIScene training pipeline. The teacher network generates the pseudo-labels " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "y_{i}^{u}" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": " from the unlabeled scene " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": ", followed by an erasure operation to filter out low-confidence points, ensuring training consistency. For the labeled scene " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "x_{i}^{l}" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": " and the generated erased scene " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^u" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": ", we employ patch-based data augmentation to obtain augmented scenes " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "x_{i}^{lpf}" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{upf}" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": ", respectively. Subsequently, both scenes are fed into the student network to compute the loss of " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 294, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 294, + 445 + ], + "type": "text", + "content": "2D projection-based methods [14, 26, 45, 48, 53, 55] project input point clouds into a 2D dense space, use 2D convolutional filters [31] for processing, and then map predictions back to the original 3D space. Finally, point-based methods [12, 18, 29, 46] operate directly on raw point clouds, enabling the extraction of point-wise semantic features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 713 + ], + "type": "text", + "content": "Semi-supervised LiDAR Semantic Segmentation: To reduce the need for 3D dense annotations, recent research has explored methods that require less supervision [3, 17, 24, 38, 40, 54]. Semi-supervised learning [4, 16, 21, 23, 30, 58] aims to utilize a large number of unlabeled data to alleviate the annotation burden. LaserMix [15], a pioneer work in semi-supervised LiDAR segmentation, leverages a simple laser beam mix strategy to achieve solid results. Subsequent research fuses LiDAR and camera data to extract deeper semantic information [5, 41]. DDsemi [21], focusing on point cloud density, builds on contrastive learning to enhance feature discriminability. The recently proposed LASS3D [22] leverages large language-vision models to support semi-supervised segmentation. Existing semi-supervised methods often build upon the mean-teacher paradigm [39] and utilize pseudo labels [1]. However, many points without pseudo-labels contribute equally to pseudo-labeled points in forward propagation but are ignored during backward propagation, leading to intra-scene inconsistency. To address this, our AIScene leverages a point erasure strategy to mitigate this inconsistency, complementing the pseudolabel mechanism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 373, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 589 + ], + "type": "text", + "content": "3D Point Cloud Data Augmentation: Point cloud data augmentation has been explored across various 3D vision tasks [9, 25, 34, 43, 44, 56]. Inspired by the success of Mixup [51] in image classification, several methods have investigated techniques for mixing selected point cloud samples. To preserve structural information, RSMix [20] introduces a neighboring function, and PointCutMix [52] establishes correspondences for sample fusion. Notably, these methods focus on local object-level augmentation. In response, several works have explored mixing strategies for point cloud scene augmentation. PloarMix [47] swaps points within the fan-shaped areas between two scenes, while CoSMix [32] aims to mitigate domain shift by mixing semantic point groups. Unlike these methods, our method is capable of modeling inter-scene information across multiple scenes to create more diversified data, while also extracting instance-level semantic information for individual classes to further enrich scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 599, + 372, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 599, + 372, + 612 + ], + "spans": [ + { + "bbox": [ + 314, + 599, + 372, + 612 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 619, + 394, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 394, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 394, + 632 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "text", + "content": "Problem Formulation. Before we describe our model and training process, we first provide the definition of semi-supervised point cloud segmentation. In detail, we have a small labeled training set " + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_{l} = \\{x_{i}^{l},y_{i}^{l}\\}_{i = 1}^{N_{l}}" + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "text", + "content": " and a large unlabeled training set " + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_{u} = \\{x_{i}^{u}\\}_{i = 1}^{N_{u}}" + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}\\in \\mathbb{R}^{m\\times (3 + r)}" + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "text", + "content": " is a point cloud scene of " + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 636, + 556, + 715 + ], + "type": "text", + "content": " points with three-dimensional" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27382" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "Cartesian coordinates relative to the sensor and additional " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "-dimensional feature (e.g., intensity, color, or reflectivity), " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " represents its corresponding point-level segmentation labels, " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "N_{u}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " are the numbers of labeled and unlabeled point cloud scenes, respectively, and " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "N_{u} \\gg N_{l}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " in ideal cases. The goal of SSL-based point cloud segmentation is to learn a segmentation model using limited labeled data and large-scale unlabeled data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "content": "Teacher-Student Paradigm. Following current mainstream research [15, 21], our framework also builds upon the teacher-student paradigm, which includes two segmentation networks with the same architecture. Our AIScene can be instantiated with any 3D segmentation network [7, 19, 57], denoted as " + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "content": ". In detail, the teacher model generates pseudo-labels for unlabeled scenes and undergoes a gradual update through the exponential moving average (EMA) [39] of the student weights:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 284, + 295, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 284, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 113, + 284, + 295, + 300 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathbb {t}} ^ {k + 1} = \\alpha \\cdot \\theta_ {\\mathbb {t}} ^ {k} + (1 - \\alpha) \\cdot \\theta_ {\\mathbb {S}} ^ {k}, \\tag {1}", + "image_path": "615fd9b6822dfac4a49ed970246c3f9ed451ac7a1fca90388c5e1a2b88db6419.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "text", + "content": " is the EMA decay rate, which is set to 0.99 by default, " + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathbb{t}}" + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "text", + "content": " represent the weight parameters of the teacher and student models, respectively, and " + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 308, + 296, + 357 + ], + "type": "text", + "content": " denotes the training step." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 364, + 186, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 186, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 186, + 376 + ], + "type": "text", + "content": "3.2. Point Erasure Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "text", + "content": "Based on the teacher-student mutual learning framework [35], the overall architecture of AIScene is depicted in Fig. 2. Specifically, AIScene consists of two primary branches: one dedicated to labeled scenes and the other to unlabeled scenes. How to effectively leverage a large number of unlabeled scenes is crucial for semi-supervised learning. Hence, for the branch of unlabeled scenes, we use a teacher network to generate pseudo-labeled scenes " + }, + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "inline_equation", + "content": "(x_{i}^{u},y_{i}^{u})" + }, + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "text", + "content": " based on a predefined confidence threshold " + }, + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\tau_{s}" + }, + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "text", + "content": " [15]. As mentioned previously, when using the pseudo-labeled scene for training, points without pseudo-labels are involved in the forward propagation but are ignored in the backward propagation due to the lack of corresponding supervision. This leads to intra-scene inconsistency, which causes semantic ambiguity of points. To address this issue, we propose erasing points without pseudo labels, resulting in an erased scene " + }, + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "inline_equation", + "content": "(\\hat{x}_i^u,\\hat{y}_i^u)" + }, + { + "bbox": [ + 55, + 381, + 296, + 620 + ], + "type": "text", + "content": " that contains fully annotated information, similar to fully supervised scenes. Formally, intrascene inconsistency addressed by the point erasure strategy can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 630, + 295, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 630, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 122, + 630, + 295, + 644 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {i} ^ {u} = \\left\\{x _ {i} ^ {u} \\mid \\Phi_ {s} \\left(x _ {i} ^ {u}\\right) \\geq \\tau_ {s} \\right\\}, \\tag {2}", + "image_path": "f0bac12af22eefbfd215741e5d17cd4ebcde408c23c1eacf5a50c7c4a335f17e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^u" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is the erased scene with corresponding class labels " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^u" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_s(\\cdot)" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " represents the segmentation network that predicts the confidence score for each point in the scene " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_i^u" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau_s" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is a predefined confidence threshold, equivalent to the pseudolabel mining threshold, requiring no additional adjustment." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 550, + 162 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 550, + 162 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 550, + 162 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 550, + 162 + ], + "type": "image", + "image_path": "88e87573d075d5a475b65e720e6bfd27db4be3c86aa6cd6dd8071821c3aa6b1c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 173, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 240 + ], + "type": "text", + "content": "Figure 3. Illustration of the pool initialization process. For each labeled point cloud scene, splitting it into multiple patches from the BEV perspective and assigning different instance categories to each patch enables the extraction of instances belonging to different classes. Subsequently, these patches and instances are stored individually in the patch pool and instance pool." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 260, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 320 + ], + "type": "text", + "content": "Since the proposed point erasure strategy is independent of the segmentation network architecture, it can be utilized as a plug-and-play tool for semi-supervised 3D semantic segmentation methods based on the pseudo-label mechanism, as depicted in Fig. 6 (a)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 327, + 489, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 489, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 489, + 339 + ], + "type": "text", + "content": "3.3. Patch-based Data Augmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 343, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 554, + 427 + ], + "type": "text", + "content": "Based on the labeled and erased scenes, we propose a patch-based data augmentation to enhance the semantic diversity of annotation information, guided by inter-scene correlation. The key idea is to construct patch pools to store scene-level and instance-level information, and then combine patches from multiple scenes to generate diverse training samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": "Pool Initialization. We first use the labeled training set " + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": " to create two patch pools, which store scene-level and instance-level information, respectively, as shown in Fig. 3. Given the " + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": "-th point cloud scene " + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "inline_equation", + "content": "x_{i}^{l}" + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": ", we patchify " + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "inline_equation", + "content": "x_{i}^{l}" + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": " into a sequence of " + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": " regular, non-overlapping point patches " + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "inline_equation", + "content": "\\{p_{ij}^{l}\\}_{j=1}^{T}" + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": ", each of size " + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "inline_equation", + "content": "(n \\times n)" + }, + { + "bbox": [ + 313, + 427, + 555, + 512 + ], + "type": "text", + "content": ", in the bird's eye view. This process can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 381, + 518, + 555, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 518, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 381, + 518, + 555, + 539 + ], + "type": "interline_equation", + "content": "\\left\\{p _ {i j} ^ {l} \\right\\} _ {j = 1} ^ {T} = \\text {p a t c h i f y} \\left(x _ {i} ^ {l}\\right), \\tag {3}", + "image_path": "8fc4c140729a6418666f0173aa10b20dac6fd74c83c8b6186d4ec70d1c4d7a61.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "inline_equation", + "content": "p_{ij}^{l}" + }, + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "text", + "content": "-th point patch from the sequence. For all labeled scenes, by performing the operation in Eq. (3), the patch pool " + }, + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Pat}}^{l}" + }, + { + "bbox": [ + 313, + 546, + 554, + 584 + ], + "type": "text", + "content": " is constructed, which can be defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 599, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 599, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 383, + 599, + 555, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {\\mathrm {P a t}} ^ {l} = \\left\\{\\left\\{p _ {i j} ^ {l} \\right\\} _ {j = 1} ^ {T} \\right\\} _ {i = 1} ^ {N _ {l}}. \\tag {4}", + "image_path": "3bb1babeafc8a9ac96f5514a5dbeb94262fa5ce150d41b50afb77fa5e7c43601.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": "Based on the patch pool, we can obtain the instance pool " + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Ins}}^l" + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": " by assigning different class labels to each patch, as depicted in Fig. 3, which contains instance information for different categories (e.g., car and person)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Following the same steps used for labeled scenes, we further create a pseudo patch pool " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Pat}}^u" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " and a pseudo instance pool " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Ins}}^u" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " from the erased scene " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^u" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": ". Unlike " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Pat}}^l" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Ins}}^l" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27383" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 162, + 196 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 162, + 196 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 162, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 162, + 196 + ], + "type": "image", + "image_path": "14ae380cf8b360e30cec537917399150684b8664dff76e0c32c3e3c4f041f302.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 295, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 255 + ], + "type": "text", + "content": "Figure 4. (a) Illustration of the MixPatch process. (b) Illustration of selecting instances from the patch pool for filling. An instance will not be filled if there are no other points around the filling location or if the filled instance overlaps with existing instances." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 166, + 74, + 294, + 196 + ], + "blocks": [ + { + "bbox": [ + 166, + 74, + 294, + 196 + ], + "lines": [ + { + "bbox": [ + 166, + 74, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 166, + 74, + 294, + 196 + ], + "type": "image", + "image_path": "b15d4aa449a510f82e7a03474a437a6dbc82c59f1749a8ed5fbd0a2bef996c45.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "text", + "content": "covering the whole training set " + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "text", + "content": ", the pseudo pools " + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Pat}}^{u}" + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Ins}}^{u}" + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "text", + "content": " only contain scene and instance information from the current batch and are reset in each iteration. This prevents low-quality pseudo labels from persisting in pools and misleading model training. Additionally, if the number of points within a patch or instance falls below a certain threshold " + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{min}}" + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "text", + "content": ", it will not be saved into the pool. In practice, we set " + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{min}}" + }, + { + "bbox": [ + 54, + 273, + 295, + 405 + ], + "type": "text", + "content": " to 5 [49]. If a patch or instance has fewer than 5 points, its spatial structure exhibits high uncertainty, making it semantically ambiguous. Thus, it is excluded from the pools to prevent noise during data augmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "text", + "content": "MixPatch. To model the inter-scene correlation and then guide the transfer of annotation information between labeled scenes and unlabeled scenes, we combine multiple scenes using patch pools. Fig. 4 (a) illustrates the patch mixing (MixPatch) strategy, while Fig. 4 (b) presents the instance filling (InsFill) strategy, which is introduced in the next paragraph. During a training iteration, the mini-batch is comprised of labeled and unlabeled scenes. After the unlabeled scene " + }, + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "text", + "content": " is transformed into the erased scene " + }, + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\hat{x}_{i}^{u}" + }, + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "text", + "content": " using the point erasure strategy, we integrate information from the patch pool into " + }, + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\hat{x}_{i}^{u}" + }, + { + "bbox": [ + 55, + 405, + 295, + 548 + ], + "type": "text", + "content": " by MixPatch. The MixPatch process can be described as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 555, + 294, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 555, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 94, + 555, + 294, + 570 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {i} ^ {u p} = \\left(\\hat {x} _ {i} ^ {u} \\odot \\mathcal {M}\\right) \\oplus \\left(\\mathcal {B} _ {\\text {P a t}} ^ {l} \\odot (1 - \\mathcal {M})\\right), \\tag {5}", + "image_path": "b93328b60aadb794d6f3ed5ab2b89e19c0c431739f4a65cd9ec55f3af9bd71ee.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": " is a binary mask indicating which patches from " + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^u" + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": " will be replaced with selected patches of the same patch index from the patch pool, " + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": " denotes element-wise multiplication, and " + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{up}" + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": " represents the augmented scene. For labeled scenes, MixPatch is applied based on the current pseudo patch pool " + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Pat}}^u" + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": ", which can be expressed as the following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 667, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 667, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 96, + 667, + 295, + 682 + ], + "type": "interline_equation", + "content": "x _ {i} ^ {l p} = \\left(x _ {i} ^ {l} \\odot (1 - \\mathcal {M})\\right) \\oplus \\left(\\mathcal {B} _ {\\text {P a t}} ^ {u} \\odot \\mathcal {M}\\right), \\tag {6}", + "image_path": "566502cdc7573d19cc21fd2b297aaa21dc1bafd935072673a6cbb744e82f1182.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": " is the same as defined in Eq. (5), thereby ensuring the full utilization of scene information from the unlabeled" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "scenes. The binary mask " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " and its complement " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "1 - \\mathcal{M}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " are used to merge inter-scene point cloud patches with the erased scene, facilitating the generation of diverse semantic information for labeled and unlabeled scenes. To alleviate bias in the patch mixing process, we apply a uniform sampling strategy to the patches in each scene, ensuring that each patch has an equal opportunity of being selected in the mixed scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": "InsFill. In addition to patch mixing, we utilize the instances stored in the instance pool " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Ins}}^l" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": " to further enhance the scene " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{up}" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": ". Given a patch with location index " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": " in the scene, we randomly select an instance from the instance pool " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Ins}}^l" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": " that was extracted from the " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": "-th patch. Patches with identical indices typically correspond to similar Li-DAR spatial layouts, ensuring consistency between the selected instance and its fill positions in the scene. As depicted in Fig. 4 (b), if the selected instance overlaps with existing instances in the scene or lacks sufficient contextual information, the filling operation is bypassed. This judgment process can be easily implemented using the maximum 3D bounding boxes. This data augmentation provides instance-level semantic information for the point cloud scene, enabling the augmented scene " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{upf}" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": " and corresponding labels " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^{upf}" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": " to effectively support semi-supervised training. For the labeled branch, we also apply the same filling strategy as described, selecting instances from the pseudo instance pool " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{Ins}}^u" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": " to fill the current scene. This InsFill process yields the scene " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "x_i^{lpf}" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": " along with its corresponding label " + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "inline_equation", + "content": "y_i^{lpf}" + }, + { + "bbox": [ + 313, + 167, + 555, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 410, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 410, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 410, + 428 + ], + "type": "text", + "content": "3.4. Model Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 529 + ], + "type": "text", + "content": "Following the standard teacher-student mutual learning paradigm [15, 35, 39], we construct a mini-batch of data with an equal composition of labeled and unlabeled scenes at each iteration. The unlabeled scenes are assigned with pseudo-labels by the teacher model and then train the student model jointly with labeled scenes. We use three types of loss functions to supervise the model training here, which is shown as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 535, + 553, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 535, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 373, + 535, + 553, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\mathcal {L} _ {\\text {s e g}} \\left(\\Phi_ {\\theta_ {\\mathfrak {s}}} \\left(x _ {i} ^ {l}\\right), y _ {i} ^ {l}\\right), \\tag {7}", + "image_path": "e5e40873ae4a37822543c6ac8a787dcdbf94560fee12fb9bad0bcd524894b6fb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 552, + 553, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 552, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 373, + 552, + 553, + 567 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l} = \\mathcal {L} _ {\\text {s e g}} \\left(\\Phi_ {\\theta_ {\\mathrm {s}}} \\left(x _ {i} ^ {l p f}\\right), y _ {i} ^ {l p f}\\right), \\tag {8}", + "image_path": "fefc702d32e278045ddf96fd169fbfded02177cc6e1197de1e6cb9cd4a37e71e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 569, + 553, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 569, + 553, + 584 + ], + "spans": [ + { + "bbox": [ + 373, + 569, + 553, + 584 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\mathcal {L} _ {s e g} \\left(\\Phi_ {\\theta_ {\\mathrm {s}}} \\left(\\hat {x} _ {i} ^ {u p f}\\right), \\hat {y} _ {i} ^ {u p f}\\right), \\tag {9}", + "image_path": "da0774ff55ac60210a2f2524a05076c7191f07dc38686b2faa8136fb0644b31e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s, \\mathcal{L}_l" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": " denote the segmentation loss for the labeled scene " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "x_i^l" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": ", augmented labeled scene " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "x_i^{lpf}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": ", and augmented unlabeled scene " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{upf}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": ", respectively, while " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": " is the original loss of the segmentation network " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": ". We sum up all the loss term to define the final training objective as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 382, + 670, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 670, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 382, + 670, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda_ {u} \\mathcal {L} _ {u} + \\lambda_ {l} \\mathcal {L} _ {l}, \\tag {10}", + "image_path": "d4649adb14bac52565e55d71d34a88d112e2aebbb9927f2cc9fef512af3534f0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{u}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{l}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " are loss weights that balance the two loss terms. After that, we update the teacher model weights us" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27384" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 548, + 236 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 548, + 236 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 548, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 548, + 236 + ], + "type": "table", + "html": "
MethodVenueSemanticKITTI [2]Avg.nuScenes [10]Avg.
1%10%20%50%1%10%20%50%
Meanteacher [39]NeurIPS'201745.457.159.260.055.451.666.067.171.764.1
CBST [58]ECCV'201848.858.359.459.756.653.066.569.671.665.2
CPS [4]CVPR'202146.758.759.660.556.452.966.370.072.565.4
LaserMix (Range View) [15]CVPR'202343.458.859.461.455.849.568.270.673.065.3
LaserMix (Voxel) [15]CVPR'202350.660.061.962.358.755.369.971.873.267.6
GPC [13]ICCV'202154.162.062.562.860.4-----
IGNet [41]WACV'202449.061.363.164.859.6-----
LiM3D [23]CVPR'202358.462.263.163.661.8-----
LASS3D [22]ECCV'202458.563.064.164.562.5-----
DDSemi [21]CVPR'202459.365.166.367.064.458.170.274.076.569.7
Ours (AIScene)-61.266.367.467.965.760.272.375.076.470.9
Improvement ↑-+1.9+1.2+1.1+0.9+1.3+2.1+2.1+1.0-0.1+1.2
", + "image_path": "d24a81c8896292f6f5e3524af962285ea6d100d54a3a4e1ce4b58a8480b964ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "type": "text", + "content": "Table 1. Experimental results on the SemanticKITTI and nuScenes dataset compared with different SSL methods. The results are reported with mIoU scores (\\%). The bold and underline scores respectively denote the best and second results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 271, + 228, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 228, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 228, + 283 + ], + "type": "text", + "content": "ing the EMA strategy described in Eq. (1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 137, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 137, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 137, + 306 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 230, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 230, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 230, + 324 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 295, + 449 + ], + "type": "text", + "content": "SemanticKITTI. Following state-of-the-art methods [15, 21], we evaluate our AISCene on the SemanticKITTI dataset [2], which is a large-scale segmentation dataset consisting of LiDAR data derived from the popular KITTI dataset [11]. It is composed of 22 sequences, with sequences 00-10 serving as the train set (of which 08 as the val set) and sequences 11-21 as the test set. Moreover, all the point cloud scenes are collected from real-world rural and urban road environments in Germany, and the semantic class labels are mapped to 19 classes in our experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 544 + ], + "type": "text", + "content": "nuScenes. The nuScenes dataset [10] is another well-known large autonomous driving dataset. This dataset is composed of 1000 sequences of 20s duration (850 for training and validation, 150 for testing), which are captured by the LiDAR sensor from Boston and Singapore. Following the official splitting protocol, the train and val sets contain 28,130 and 6,019 point cloud scenes, respectively, including per-points annotation for 16 semantic classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": "Evaluation Metrics. For both the SemanticKITTI and nuScenes datasets, we sample different ratios of annotated train scenes to obtain semi-supervised data. Meanwhile, we report the performance on the validation set using the mean Intersection over Union (mIoU) metric [21, 23]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 188, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 188, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 188, + 624 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "We conduct experiments with " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " NVIDIA V100 GPUs with 32GB RAM, based on MMDetection3D [8]. For the threshold " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{s}" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " for pseudo-label generation and point erasure, we experimentally find that 0.9 yields relatively satisfactory results. We use the MinkowskiNet [7] and Cylinder3D [57] as the backbone network due to its widely adopted in previous methods [15, 28, 42]. To facilitate patch-based data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": "augmentation, we restrict our focus to point clouds within the X and Y dimensions in the range of " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "[-50, 50]m" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " for both the SemanticKITTI and nuScenes datasets. Similar to mainstream work, we also adopt the consistency loss proposed in MeanTeacher [39] to align the predictions of the teacher and student networks. Additionally, the weights of the consistency loss are assigned as " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "1 \\times 10^{3}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "2 \\times 10^{3}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " for the SemanticKITTI and nuScenes datasets, respectively. The weights for " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\lambda_{u}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\lambda_{l}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " are set as 1 for both datasets. Our method excludes all forms of model ensemble and test-time augmentation during the evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 409, + 399, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 399, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 399, + 420 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "content": "In this subsection, we compare our method with current SOTA semi-supervised point cloud segmentation methods (e.g., LaserMix [15], LiM3D [23] and DDsemi [21]), using sub-sampling of the datasets under various partition protocols [21]. Specifically, we split the entire training set into labeled subsets with ratios of " + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "content": " in the SemanticKITTI [2] and nuScenes dataset [10]. For further comparison, we also include several classical semi-supervised learning methods [4, 39, 58] from the 2D image domain adapted to 3D segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Tab. 1 summarizes the quantitative results on SemanticKITTI and nuScenes datasets. Our method consistently outperforms other comparative methods on both datasets, achieving average mIoU improvements of " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": ", respectively. Notably, the proposed method demonstrates even greater gains on smaller sampling splits, such as " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": ", which are particularly beneficial for the semi-supervised setting. Specifically, with a " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " ratio, our method shows mIoU improvements of up to " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " on the two datasets, respectively, compared to the recently proposed SOTA method, DDsemi [21]. These results also validate that our method is less sensitive to reduced labeled data sampling than other methods. Furthermore, we present qualitative results in Fig. 5. As indicated by the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27385" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 135, + 168 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 135, + 168 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 135, + 168 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 135, + 168 + ], + "type": "image", + "image_path": "effddf6ca379569f775d87cef2e34128b867ab0cf25ad621a9b973ae7717b936.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 71, + 214, + 168 + ], + "blocks": [ + { + "bbox": [ + 137, + 71, + 214, + 168 + ], + "lines": [ + { + "bbox": [ + 137, + 71, + 214, + 168 + ], + "spans": [ + { + "bbox": [ + 137, + 71, + 214, + 168 + ], + "type": "image", + "image_path": "f5f8f271224715696ef40a16a8b1fe3119772ca7c6b7e19b5cc27828c29e940c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 215, + 71, + 292, + 168 + ], + "blocks": [ + { + "bbox": [ + 215, + 71, + 292, + 168 + ], + "lines": [ + { + "bbox": [ + 215, + 71, + 292, + 168 + ], + "spans": [ + { + "bbox": [ + 215, + 71, + 292, + 168 + ], + "type": "image", + "image_path": "24b274b44b46e6df8d1f622a281a78da544e9b0ec4a3cfd6fd3bbee7df68a8ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 170, + 135, + 254 + ], + "blocks": [ + { + "bbox": [ + 59, + 170, + 135, + 254 + ], + "lines": [ + { + "bbox": [ + 59, + 170, + 135, + 254 + ], + "spans": [ + { + "bbox": [ + 59, + 170, + 135, + 254 + ], + "type": "image", + "image_path": "f30d5b9e6d29c1e0dad9788e69bbb7aa3490ff9deefa293c2664c459170ddedf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 170, + 214, + 254 + ], + "blocks": [ + { + "bbox": [ + 137, + 170, + 214, + 254 + ], + "lines": [ + { + "bbox": [ + 137, + 170, + 214, + 254 + ], + "spans": [ + { + "bbox": [ + 137, + 170, + 214, + 254 + ], + "type": "image", + "image_path": "7c6c6bba31f1e1a3e6f91667349620a7dff637e5e9ad7a6479bf06785397dae9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 170, + 292, + 254 + ], + "blocks": [ + { + "bbox": [ + 215, + 170, + 292, + 254 + ], + "lines": [ + { + "bbox": [ + 215, + 170, + 292, + 254 + ], + "spans": [ + { + "bbox": [ + 215, + 170, + 292, + 254 + ], + "type": "image", + "image_path": "b4ad28930e5c24a1c3a7be72f42e918babf091b018fe4ab6e23e22175728af48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 255, + 135, + 339 + ], + "blocks": [ + { + "bbox": [ + 59, + 255, + 135, + 339 + ], + "lines": [ + { + "bbox": [ + 59, + 255, + 135, + 339 + ], + "spans": [ + { + "bbox": [ + 59, + 255, + 135, + 339 + ], + "type": "image", + "image_path": "a6b546960433769d186988177e8b8bbd90eab545899925b2830433ae76184f8a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 295, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 392 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison results from LiDAR bird's eye view on SemanticKITTI val set. We set the point with correct and incorrect predicted semantic class in blue and orange for better visualization. Best viewed in color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 255, + 214, + 339 + ], + "blocks": [ + { + "bbox": [ + 137, + 255, + 214, + 339 + ], + "lines": [ + { + "bbox": [ + 137, + 255, + 214, + 339 + ], + "spans": [ + { + "bbox": [ + 137, + 255, + 214, + 339 + ], + "type": "image", + "image_path": "2dc2d0e117ed974f7ad40beb905ac71fb8eca19aa2e2e50237c8d326c9574d5c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 215, + 255, + 292, + 339 + ], + "blocks": [ + { + "bbox": [ + 215, + 255, + 292, + 339 + ], + "lines": [ + { + "bbox": [ + 215, + 255, + 292, + 339 + ], + "spans": [ + { + "bbox": [ + 215, + 255, + 292, + 339 + ], + "type": "image", + "image_path": "6235198668efd37c506bd210494a39694a506d80792564ea57f657f35248e3e9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 406, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 430 + ], + "type": "text", + "content": "purple and red dashed circles, our method achieves better segmentation of boundary regions and small objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 441, + 149, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 149, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 149, + 453 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 460, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 531 + ], + "type": "text", + "content": "In this subsection, We first validate the adaptability of our method to different segmentation models. We then examine the impact of intra-scene consistency through the point erasure strategy and explore inter-scene correlation by incorporating patch-based mixing and patch-based instance filling data augmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "content": "Adaptability with Different Segmentation Methods. Our AISCene is segmentor-agnostic and can be easily used to further improve off-the-shelf 3D fully-supervised segmentors under the semi-supervised setting. Tab. 2 shows that the proposed method can substantially boost the performance with Cylinder3D [57] as the backbone, indicating the generalization ability of our framework. For consistency in comparison, the subsequent experiments are based on Cylinder3D unless otherwise specified." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Effectiveness of Point Erasure Strategy. The " + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": " row in Tab. 3 reports the performance of the supervision-only Cylinder3D model [57], which serves as the baseline. When PtErase is applied to the baseline, it effectively boosts performance by ensuring consistency during forward and backward propagation, thus addressing the lim" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 554, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 148 + ], + "type": "table", + "html": "
DatasetsMethods1%10%20%50%Avg.
SemanticKITTICylinder3D [57]45.456.157.858.754.5
Ours (AIScene)54.563.363.764.361.5
Improvement+9.1+7.2+5.9+5.6+7.0
nuScenesCylinder3D [57]50.965.966.671.263.7
Ours (AIScene)56.670.272.873.968.4
Improvement+5.7+4.3+6.2+2.7+4.7
", + "image_path": "e2ce4b73cc4d3c6cf2e6da6ef372c139f6655510e61949881c8047c18c5fbe2d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 182, + 554, + 287 + ], + "blocks": [ + { + "bbox": [ + 313, + 157, + 554, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 554, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 554, + 178 + ], + "type": "text", + "content": "Table 2. Ablation study on the adaptability of our AIScene using Cylinder3D as the segmentation model." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 182, + 554, + 287 + ], + "lines": [ + { + "bbox": [ + 317, + 182, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 554, + 287 + ], + "type": "table", + "html": "
Ours AISCeneptEraseMixPatchInsFillSplits on SemanticKITTIAvg.
1%10%20%50%
45.456.157.858.754.5
✓51.759.861.860.458.4
✓51.361.460.760.558.5
✓52.361.760.161.158.8
✓✓54.661.062.962.860.3
✓✓52.061.760.862.659.3
✓✓50.562.563.863.660.1
✓✓✓54.563.363.764.361.5
", + "image_path": "58576494990a01372e6985937fdfbdf161c180622ea3fd9ec75fec151e0c5961.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 554, + 340 + ], + "type": "text", + "content": "Table 3. Ablation study of different components of our AIScene framework. PtErase refers to the point erasure strategy, while MixPatch and InsFill refer to the patch mix and instance filling data augmentation, respectively." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 353, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 567 + ], + "type": "text", + "content": "ditions of the pseudo-label mechanism. This consistency is achieved through a plug-and-play point erasure strategy, which can be easily integrated into methods relying on the pseudo-label mechanism. To further validate the importance of maintaining consistency during forward and backward propagation, we embedded the PtErase strategy into LaserMix [15]. As illustrated in Fig. 6 (a), using the PtErase strategy improves performance across all ratios, underscoring the significance of intra-scene consistency and demonstrating the generalizability of our method. Moreover, Fig. 6 (b) shows that the proportion of points removed by the erasure strategy is initially high at the start of training, but gradually decreases as the network converges. Furthermore, We present a comparison of class probability distributions in Fig. 7 to validate the effectiveness of the point erasure strategy. The results demonstrate that this strategy reduces classification ambiguity, enabling the model to better capture the features of the correct class." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Effectiveness of Patch-based Data Augmentation. The patch-based data augmentation consists of MixPatch (patch-based mixing) and InsFill (instance filling). In Tab. 3, the results in the " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " rows indicate that using MixPatch or InsFill individually yields an average mIoU improvement of " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", respectively, over the baseline. This suggests that incorporating diverse semantic information across scenes benefits semi-supervised LiDAR segmentation. Furthermore, as shown in the " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "5^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "6^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " rows, combining MixPatch or InsFill with PtErase significantly enhances performance, highlighting the effectiveness of patch-based data augmentation. The combination of all" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27386" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 553, + 195 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 553, + 195 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 553, + 195 + ], + "type": "image", + "image_path": "443f65983496f89a861c533cedc6247d0287a0a339650b8586a44767b443dbec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 196, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 555, + 229 + ], + "type": "text", + "content": "Figure 6. (a) Illustration of performance differences whether using point erasure strategy on the LaserMix [15]. (b) Variation in the proportion of points removed by the point erasure strategy throughout the training process. (c) Comparison of different mixing-based techniques on SemanticKITTI dataset. (d) Results of mIoU against the number of patch split " + }, + { + "bbox": [ + 55, + 196, + 555, + 229 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 196, + 555, + 229 + ], + "type": "text", + "content": " under the " + }, + { + "bbox": [ + 55, + 196, + 555, + 229 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 196, + 555, + 229 + ], + "type": "text", + "content": " SemanticKITTI split." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 236, + 293, + 353 + ], + "blocks": [ + { + "bbox": [ + 56, + 236, + 293, + 353 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 293, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 293, + 353 + ], + "type": "image", + "image_path": "57f2f0d75921826a5c67b8ca4b5dd8175e66ccbdb5a34df49a31fa4ab3d93e93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 354, + 295, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 409 + ], + "type": "text", + "content": "Figure 7. The left chart shows the initial output with distributed probabilities across multiple classes, indicating uncertainty in class prediction. Through our point erasure strategy, the right chart exhibits a distribution with a dominant class (\"person\") having a significantly higher probability," + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 422, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 295, + 530 + ], + "type": "text", + "content": "three components yields substantial performance gains, as evidenced by the last row. Additionally, the patch pool and instance pool within the patch-based data augmentation facilitate the fusion of diverse semantic information across multiple scenes. To verify the effectiveness of interactions at the multi-scene level, we modify the pool implementation to allow interactions between only two scenes. The results in Tab. 4 indicate that multi-scene interactions yield superior performance compared to dual-scene interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 532, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 700 + ], + "type": "text", + "content": "Different Mixing Strategies. Fig. 6 (c) compares our patch-based data augmentation with existing mixing methods [27, 47, 50]. MixUp [51] shows a slight improvement over the baseline, likely due to its limited capacity for information transformation. Note that we reproduce Cylinder-Mix in the LiDAR branch [5] for comparison. Other methods demonstrate considerable improvement over the baseline by enriching point cloud distributions. In contrast, our AIScene achieves the most substantial gains by effectively leveraging inter-scene correlation, highlighting its effectiveness in facilitating diverse semantic information generation for labeled and unlabeled scenes. This is achieved through an innovative design that incorporates a patch pool and an instance pool spanning multiple scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Patch Split Number. As the patch split number " + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": " plays" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 236, + 555, + 289 + ], + "blocks": [ + { + "bbox": [ + 317, + 236, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 317, + 236, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 555, + 289 + ], + "type": "table", + "html": "
Augmentation TypeMethodsSplits on SemanticAvg.
1%10%20%50%
dual-scene levelOurs53.461.361.862.259.7
multi-scene levelOurs54.563.363.764.361.5
Improvement-+1.1+2.0+1.9+2.1+1.8
", + "image_path": "186132ff7da66aac59ca28d908eadb0e344be1e50ee484ea2e2ddc3a241d145e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 319 + ], + "type": "text", + "content": "Table 4. Comparison between dual-scene and multi-scene levels in the patch-based data augmentation of our AIScene." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": "a critical role in our framework, we further analyze the impact of varying " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": " in patch-based data augmentation, with the results presented in Fig. 6 (d). When " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": " is small, it can yield a certain level of enhancement, but it may fail to generate data sufficiently divergent from the original, thereby impeding the effective generation of diverse semantic information between labeled and unlabeled scenes. Conversely, when " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": " is too large, the generated scenes may deviate excessively from the original data, resulting in a loss of essential prior scene information. Empirically, we find that setting " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": " to 18 yields optimal results. Throughout this paper, we consistently use " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": " set to 18 unless explicitly stated otherwise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 509, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 676 + ], + "type": "text", + "content": "This paper explores scene affinity for semi-supervised Li-DAR semantic segmentation in driving scenes, named AIScene. Specifically, AIScene addresses intra-scene inconsistency through a point erasure strategy, which enhances the learning of pseudo-labeled points. Moreover, it employs a patch-based data augmentation method to mix multiple scenes at both scene and instance levels, guided by inter-scene correlation. This method enriches the semantic diversity of labeled scenes and complements pseudo-labeled scenes. Extensive experiments show that AIScene achieves state-of-the-art results on popular benchmarks. Additionally, the novel erasure strategy and augmentation method can function independently to facilitate semi-supervised LiDAR semantic segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (No.62325111, U22B2011, 62271354)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27387" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 133 + ], + "type": "text", + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, pages 1-8. IEEE, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 295, + 179 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 225 + ], + "type": "text", + "content": "[3] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In CVPR, pages 7020-7030, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "type": "text", + "content": "[4] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In CVPR, pages 2613-2622, 2021. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 304 + ], + "type": "text", + "content": "[5] Yujun Chen, Xin Tan, Zhizhong Zhang, Yanyun Qu, and Yuan Xie. Beyond the label itself: Latent labels enhance semi-supervised point cloud panoptic segmentation. In AAAI, pages 1245-1253, 2024. 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "text", + "content": "[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "text", + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "text", + "content": "[8] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "type": "text", + "content": "[9] Jin Fang, Xinxin Zuo, Dingfu Zhou, Shengze Jin, Sen Wang, and Liangjun Zhang. Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In CVPR, pages 4710-4720, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 476, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 294, + 529 + ], + "type": "text", + "content": "[10] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE RA-L, 7(2):3795-3802, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 564 + ], + "type": "text", + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 609 + ], + "type": "text", + "content": "[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 612, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 294, + 655 + ], + "type": "text", + "content": "[13] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In ICCV, pages 6423-6432, 2021. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 700 + ], + "type": "text", + "content": "[14] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In ICCV, pages 228-240, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 703, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 294, + 713 + ], + "type": "text", + "content": "[15] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "Lasermix for semi-supervised lidar semantic segmentation. In CVPR, pages 21705-21715, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[16] Lingdong Kong, Xiang Xu, Jiawei Ren, Wenwei Zhang, Liang Pan, Kai Chen, Wei Tsang Ooi, and Ziwei Liu. Multimodal data-efficient 3d scene understanding for autonomous driving. arXiv:2405.05258, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "text", + "content": "[17] Hyeokjun Kweon, Jihun Kim, and Kuk-Jin Yoon. Weakly supervised point cloud semantic segmentation via artificial oracle. In CVPR, pages 3721-3731, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[18] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 220, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 253 + ], + "type": "text", + "content": "[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, pages 17545-17555, 2023. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "type": "text", + "content": "[20] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 331 + ], + "type": "text", + "content": "[21] Jianan Li and Qiulei Dong. Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In CVPR, pages 3260-3269, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 333, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 375 + ], + "type": "text", + "content": "[22] Jianan Li and Qiulei Dong. Lass3d: Language-assisted semi-supervised 3d semantic segmentation with progressive unreliable data exploitation. In ECCV, pages 252-269. Springer, 2025. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 420 + ], + "type": "text", + "content": "[23] Li Li, Hubert PH Shum, and Toby P Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In CVPR, pages 9361-9371, 2023. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 422, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 464 + ], + "type": "text", + "content": "[24] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In CVPR, pages 8428-8437, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[25] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In CVPR, pages 23819-23828, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 513, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 544 + ], + "type": "text", + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, pages 4213-4220. IEEE, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "type": "text", + "content": "[27] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In 3DV, pages 116-125. IEEE, 2021. 2, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "type": "text", + "content": "[28] Gilles Puy, Spyros Gidaris, Alexandre Boulch, Oriane Simeoni, Corentin Sautier, Patrick Pérez, Andrei Bursuc, and Renaud Marlet. Three pillars improving vision foundation model distillation for lidar. In CVPR, pages 21519-21529, 2024. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "text", + "content": "[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[30] Laurenz Reichardt, Nikolas Ebert, and Oliver Wasenmüller. 360deg from a single camera: A few-shot approach for lidar segmentation. In ICCV, pages 1075-1083, 2023. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27388" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[32] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Fabio Poiesi, and Elisa Ricci. Compositional semantic mix for domain adaptation in point cloud segmentation. IEEE TPAMI, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[33] Shivanand Venkanna Sheshappanavar, Vinit Veerendraveer Singh, and Chandra Kambhametu. Patchaugment: Local neighborhood augmentation in point cloud classification. In ICCV, pages 2118-2127, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 240 + ], + "type": "text", + "content": "[34] Jungwook Shin, Jaeill Kim, Kyungeun Lee, Hyunghun Cho, and Wonjong Rhee. Diversified and realistic 3d augmentation via iterative construction, random placement, and hpr occlusion. In AAAI, pages 2282-2291, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "text", + "content": "[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "type": "text", + "content": "[36] Shihao Su, Jianyun Xu, Huanyu Wang, Zhenwei Miao, Xin Zhan, Dayang Hao, and Xi Li. Pups: Point cloud unified panoptic segmentation. In AAAI, pages 2339-2347, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[37] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "text", + "content": "[38] Liyao Tang, Zhe Chen, Shanshan Zhao, Chaoyue Wang, and Dacheng Tao. All points matter: Entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS, 36, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[39] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, pages 1195–1204, 2017. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 498 + ], + "type": "text", + "content": "[40] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In CVPR, pages 2697-2707, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[41] Ozan Unal, Dengxin Dai, Lukas Hoyer, Yigit Baran Can, and Luc Van Gool. 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In WACV, pages 7336-7345, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[42] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, pages 4917-4928, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 645 + ], + "type": "text", + "content": "[43] Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2707-2715, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[44] Yi Wang, Jiaze Wang, Jinpeng Li, Zixu Zhao, Guangyong Chen, Anfeng Liu, and Pheng Ann Heng. Pointpatchmix: Point cloud mixing with patch scoring. In AAAI, pages 5686-5694, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[45] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 644 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, pages 4376-4382. IEEE, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[46] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[47] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In NeurIPS, 2022. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "text", + "content": "[48] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In ECCV, pages 1-19. Springer, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "type": "text", + "content": "[49] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "text", + "content": "[50] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 352 + ], + "type": "text", + "content": "[51] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv:1710.09412, 2017. 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[52] Jinlai Zhang, Lyujie Chen, Bo Ouyang, Binbin Liu, Jihong Zhu, Yujin Chen, Yanmei Meng, and Danfeng Wu. Pointcutmix: Regularization strategy for point cloud classification. Neurocomputing, 505:58-67, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, pages 9601-9610, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 444, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 486 + ], + "type": "text", + "content": "[54] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In AAAI, pages 3421-3429, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 488, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 521 + ], + "type": "text", + "content": "[55] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IROS, pages 4453-4458. IEEE, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[56] Qinfeng Zhu, Lei Fan, and Ningxin Weng. Advancements in point cloud data augmentation for deep learning: A survey. arXiv:2308.12113, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "text", + "content": "[57] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[58] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, pages 289-305, 2018. 3, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27389" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_content_list.json b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c29f3187c5e51ec72709e2b58b099aa441537a06 --- /dev/null +++ b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guanglu Dong Xiangyu Liao Mingyang Li Guihuan Guo Chao Ren\\*", + "bbox": [ + 151, + 203, + 851, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "College of Electronics and Information Engineering, Sichuan University, China", + "bbox": [ + 184, + 222, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dongguanglu, liaoxiangyul, 2024222055110, guoguihuan}@stu.scu.edu.cn, chaoren@scu.edu.cn", + "bbox": [ + 101, + 241, + 890, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Adversarial Networks (GANs) have been widely applied to image super-resolution (SR) to enhance the perceptual quality. However, most existing GAN-based SR methods typically perform coarse-grained discrimination directly on images and ignore the semantic information of images, making it challenging for the super resolution networks (SRN) to learn fine-grained and semantic-related texture details. To alleviate this issue, we propose a semantic feature discrimination method, SFD, for perceptual SR. Specifically, we first design a feature discriminator (FeatD), to discriminate the pixel-wise middle semantic features from CLIP, aligning the feature distributions of SR images with that of high-quality images. Additionally, we propose a text-guided discrimination method (TG-D) by introducing learnable prompt pairs (LPP) in an adversarial manner to perform discrimination on the more abstract output feature of CLIP, further enhancing the discriminative ability of our method. With both Feat-D and TG-D, our SFD can effectively distinguish between the semantic feature distributions of low-quality and high-quality images, encouraging SRN to generate more realistic and semantic-relevant textures. Furthermore, based on the trained Feat-D and LPP, we propose a novel opinion-unaware no-reference image quality assessment (OU NR-IQA) method, SFD-IQA, greatly improving OU NR-IQA performance without any additional targeted training. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our proposed methods. Code is available at https://github.com/GuangluDong0728/SFD.", + "bbox": [ + 88, + 321, + 485, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 790, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As one of the most important tasks in low-level computer vision [14, 19, 33, 36, 50, 52, 60, 67], single image super-resolution (SISR) [11, 18, 28, 34, 56, 57, 62, 80] aims to accurately reconstruct high-resolution (HR) images from the", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/396d96cff39202a8db442fad8ae6f06fc0e3a04d68748b4918e30ea48a121052.jpg", + "image_caption": [ + "Figure 1. Perception-distortion trade-off comparison between our proposed SFD and other SOTA GAN-based SISR methods." + ], + "image_footnote": [], + "bbox": [ + 513, + 289, + 906, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f08aef6845d08f676d8eb3de9d65efebb56b8000129b5db6bd79e1078d064fd.jpg", + "image_caption": [ + "Figure 2. Relative PLCC comparison between the proposed SFD-IQA and other OU NR-IQA methods." + ], + "image_footnote": [], + "bbox": [ + 521, + 542, + 906, + 761 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "given low-resolution (LR) counterparts. In recent years, with advancements in deep learning, SISR has achieved rapid and substantial progress. Since Dong et al. [11] first introduce deep learning to SISR task, numerous studies have emerged to explore more advanced network architectures, including", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28176", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb444b8f05e7e2e4847acef8eeb69dd4ea8581952263557cc70c5879bcb428b9.jpg", + "image_caption": [ + "(a) Vanilla GAN-based SISR methods" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 401, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7816e3f3230f2168f7d501b4f8569de0efe99dbd8fa7e05306790716d9ac8f40.jpg", + "image_caption": [ + "(b) Our proposed SFD which performs discrimination on the semantic features", + "Figure 3. (a) Vanilla GAN-based SISR methods: performing discrimination on images; (b) Our proposed SFD: performing discrimination on the semantic features. Vanilla GAN-based SISR typically uses an image-wise, patch-wise, or pixel-wise discriminator to determine if the distribution of SR images align with that of high-quality real-world images, ignoring the semantic information of images. Such coarse-grained discrimination makes it challenging for the SR network to reconstruct fine-grained, semantic-relevant textures. In contrast, we propose to perform discrimination on the semantic features, encouraging the SR network to generate more realistic semantic textures." + ], + "image_footnote": [], + "bbox": [ + 408, + 88, + 903, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CNN-based models [28, 34, 56, 57, 81, 82] and Transformer-based models [7, 13, 31, 38, 72, 80], achieving continuous improvements in fidelity-oriented performance. However, these works primarily employ pixel-wise L1 or MSE losses for the training process, which often produce overly smooth results and fail to generate realistic texture details [4, 29].", + "bbox": [ + 88, + 359, + 482, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the above limitations, a series of generative adversarial networks (GANs) [17] based SISR models [6, 15, 29, 30, 32, 40, 47, 48, 63, 65] have been proposed to generate more realistic images that align more closely with human visual perception by effectively balancing perceptual quality and fidelity. As shown in Figure 3 (a), most existing GAN-based methods perform discrimination on the images using an image-wise [54], patch-wise [23], or pixel-wise [53] discriminator to distinguish between the distributions of SR images and HR images. However, directly discriminating on the images is coarse-grained and lacks semantic awareness. This leads to that the textures of generated images do not match its semantic and often come with artificial artifacts, making it challenging to produce realistic semantic textures.", + "bbox": [ + 88, + 457, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, different from existing works, we propose to perform discrimination on the semantic features, aiming to align the semantic feature distributions of SR images more closely with that of HR images. Considering that CLIP [51] can extract more interpretable, semantic-aware and quality-relevant features [9, 20, 30, 61], we select CLIP's image encoder as the semantic feature extractor and propose two methods to perform joint discrimination on the semantic features. Specifically, as shown in Figure 3 (b), as for the pixel-wise middle semantic features, we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR images and HR images, encouraging the semantic feature distributions of SR images to match that of HR images. As for the final output feature which is more abstract and global, we propose a text-guided discrimination", + "bbox": [ + 88, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method, TG-D, by introducing learnable prompt pairs (LPP) into an adversarial training process to make them sensitive to global image quality and thus able to discriminate the abstract semantic features, further enhancing our method's overall discriminative capability. With both Feat-D and TG-D, the SR network is able to learn more fine-grained semantic feature distributions, ultimately producing more realistic and semantic-related texture details.", + "bbox": [ + 511, + 359, + 906, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, existing researches [39, 40, 83] have shown that during the training process where the discriminators in GAN-based SISR learn to distinguish the realness of image distributions, they simultaneously and implicitly learn to assess image quality, making them highly applicable to opinion-unaware no-reference image quality assessment (OU NR-IQA). Simultaneously, recent studies [20, 44] have also shown the advantage of CLIP for OU NR-IQA. In this paper, to further improve OU NR-IQA performance, we propose a novel OU NR-IQA method, SFD-IQA, by integrating the well trained Feat-D and TG-D which conveniently combines both of the above advantages. Thanks to the superior semantic feature discrimination capability of Feat-D and TG-D and the inherent image quality correlation of CLIP, we greatly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets.", + "bbox": [ + 511, + 484, + 908, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, our SFD demonstrates a better perception-distortion (PD) trade-off on the PSNR-LPIPS-DISTS plane. Compared to other state-of-the-art (SOTA) GAN-based SISR methods, our LPIPS and PSNR scores are the best among all methods with competitive DISTS scores. Additionally, Fig. 2 illustrates a comparison between our SFD-IQA and other SOTA OU NR-IQA methods, where SFD-IQA outperforms other methods across both SR IQA datasets and authentically distorted IQA datasets. The main contributions of this work are summarized below:", + "bbox": [ + 511, + 729, + 908, + 880 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a semantic feature discrimination (SFD)", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method for perceptual SISR, which consists of a feature discriminator (Feat-D) based discrimination and a text-guided discrimination (TG-D). Our SFD can effectively distinguish between the CLIP semantic feature distributions of SR and HR images, encouraging SR networks to learn more fine-grained and semantic-related textures, and achieving better perception-distortion (PD) trade-off.", + "bbox": [ + 102, + 90, + 482, + 196 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel OU NR-IQA method, SFD-IQA, by effectively integrating the two well trained feature discrimination methods. Without any additional task-specific training on IQA datasets, SFD-IQA can significantly enhance the OU NR-IQA performance for both SR images and other authentically distorted images.", + "- Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our method. SFD can help SR networks achieve better PD trade-off and generate more realistic semantic textures, and SFD-IQA is able to greatly improve OU NR-IQA performance. Moreover, our approach can be easily integrated into other GAN-based image restoration methods." + ], + "bbox": [ + 89, + 196, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 409, + 240, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Fidelity-Oriented SISR", + "text_level": 1, + "bbox": [ + 89, + 434, + 305, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, deep learning based single image super-resolution (SISR) [11, 18, 28, 29, 62, 70, 81, 82] has made tremendous advancements. Since SRCNN [11] first introduces CNN into SISR, numerous SISR backbones [18, 28, 34, 56, 57, 81, 82] have continued to emerge to enhance the performance of fidelity-oriented SISR. Specifically, RCAN [81] introduces channel attention into SISR network, and RDN [82] leverages dense connection to fully utilize hierarchical features for SISR. Furthermore, based on the success of the vision transformer [12] framework, [13, 31, 72, 80] further improve the SISR performance. However, in order to maximize fidelity, most of the aforementioned methods only employ pixel-wise loss functions to optimize the training process and ignore the perceptual quality, which often leads to overly smooth results and fails to achieve visually appealing results.", + "bbox": [ + 89, + 458, + 482, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Perception-Oriented GAN-based SISR", + "text_level": 1, + "bbox": [ + 89, + 696, + 423, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve perceptual quality, GAN-based SISR [15, 27, 30, 32, 39, 47, 64, 65, 75, 79] has received constant attention in recent years. Specifically, SRGAN [29] introduces GANs into SR tasks to obtain SR images more in line with human perception. DualFormer [39] applies an improved spectral discriminator to improve perceptual quality. Sed [30] incorporates the semantic features into the vanilla discriminator to finely distinguish the image distributions. EA-ADAM [55] develops a new optimizer from a multi-objective optimization perspective to achieve a better perception-distortion tradeoff. However, due to the lack of semantic-relevant perception, most GAN-based SR networks are difficult to", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "produce fine-grained semantic textures, result in exhibiting artificial artifacts. In this paper, we propose to discriminate the semantic features of CLIP to guide the SR network to learn more fine-grained semantic feature distributions and generate more realistic semantic textures.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Opinion-Unaware NR-IQA", + "text_level": 1, + "bbox": [ + 511, + 174, + 759, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OU NR-IQA [2, 43, 46, 61, 69, 71, 77, 83] aims to evaluate the quality of images without human-labeled data. Traditional OU NR-IQA methods [69, 76, 77] are typically based on natural scene statistics (NSS), which estimate distribution parameters to assess quality. Researches on deep learning-based OU NR-IQA are relatively sparse, with most methods centered around learning-to-rank [35, 43] and contrastive learning [44]. Recently, with the emergence of the pretrained vision-language model CLIP [51], [20, 61] leverage the rich visual-linguistic priors encapsulated in CLIP to evaluate image perceptual quality. Additionally, RecycleD [83] proposes to recycle the trained discriminator for GAN-based SISR to predict perceptual quality. Following [83], [39] and [40] further enhance the IQA ability of the discriminators in GAN-based SISR. However, these discriminator-based methods only rely on the trained semantic-unaware discriminator and ignore the assessment for semantic, which is difficult to achieve satisfactory performance.", + "bbox": [ + 511, + 196, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method and Analyses", + "text_level": 1, + "bbox": [ + 511, + 481, + 718, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide a detailed explanation and analysis of our method. Firstly, we describe the overall framework of our method in Sec. 3.1. Next, we provide a detailed introduction and analysis of the two proposed semantic feature discrimination methods in Sec. 3.2. Lastly, we discuss how to extend our method to OU NR-IQA in Sec. 3.3.", + "bbox": [ + 511, + 507, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 511, + 606, + 702, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel semantic feature discrimination (SFD) method for perceptual SISR, aiming to achieve better perception-distortion (PD) trade-off. Differing from vanilla GAN-based SR methods that only consider coarse-grained image distributions with poor semantic relevance, our SFD focuses on the distributions of semantic features. As illustrated in Fig. 4, given the input low-resolution image $I_{lr}$ , we can obtain its corresponding SR image $I_{sr}$ through the SRN. Then, through the pre-trained CLIP [51] image encoder $E_{i}$ we can obtain rich semantic features of $I_{sr}$ with different scales and characteristics, including the multi-scale pixelwise middle features $F_{smi}$ , $F_{smi}$ , $F_{smi} \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}$ ( $i = 1,2,3$ ), and the more abstract final output feature $F_{sout} \\in \\mathbb{R}^{1 \\times L}$ . To better leverage these semantic features to improve the robustness of our approach, we propose two methods to separately perform discrimination on them.", + "bbox": [ + 511, + 628, + 908, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On one hand, as for the pixel-wise middle features, we design a feature discriminator (Feat-D) to distinguish be", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28178", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92675a7137dc0bfde740b3776c091ecc7bb348491a2b0a1ae82a8a776f033263.jpg", + "image_caption": [ + "Figure 4. Framework of the proposed SFD. SFD consists of a feature discriminator (Feat-D) and a text-guided discrimination (TG-D) method, Feat-D and TG-D are used to perform discrimination on the middle features and final output features from CLIP, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 908, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tween the feature distributions of SR and HR images. On the other hand, we propose a text-guided discrimination (TG-D) method to perform discrimination on $F_{\\text{out}}$ , further enhancing the overall discriminative ability. With both Feat-D and TG-D, SFD is able to facilitate the SR network (SRN) to learn more fine-grained semantic feature distributions and generate more realistic and semantic-related textures.", + "bbox": [ + 88, + 412, + 483, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Discrimination on Semantic Features", + "text_level": 1, + "bbox": [ + 89, + 527, + 410, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Feature Discriminator. As for the middle features of CLIP which contain multi-scale pixel-wise semantic information [30], we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR and HR images, encouraging the semantic feature distribution of SR images to match that of HR images. Specifically, as shown in Figure 4, we can obtain the middle features $F_{sm1}$ , $F_{sm2}$ , $F_{sm3}$ of SR image $I_{sr}$ , and $F_{hm1}$ , $F_{hm2}$ , $F_{hm3}$ of HR image $I_{hr}$ through the freezing CLIP image encoder $E_i$ . Then, the Feat-D is used to perform discrimination on those pixel-wise semantic features. During the adversarial training process, the Feat-D and the SRN will be optimized respectively using the following adversarial loss functions:", + "bbox": [ + 89, + 550, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {D _ {f e a}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(1 - D \\left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\\right)\\right) \\tag {1} \\\\ + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log (D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 758, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {G _ {f e a}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(D \\left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\\right)\\right) \\tag {2} \\\\ + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log (1 - D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 810, + 483, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{F_{hr}}$ and $P_{F_{sr}}$ represent the feature distributions of HR and SR images, respectively. Benefiting from the excellent discriminative ability of Feat-D, the SRN can learn more", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fine-grained semantic feature distributions aligned with HR images, generating more semantically relevant textures.", + "bbox": [ + 511, + 412, + 903, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learnable Prompt Pairs Driven Text-Guided Discrimination. Discriminating the final output feature $F_{\\text{sout}}$ which is more global and abstract is expected to further enhance the overall performance of our method. However, differing from the middle features, $F_{\\text{sout}}$ lacks the pixel-wise semantic information due to its more abstract nature, making it less suitable for Feat-D's pixel-wise discrimination process. Therefore, leveraging the inherent visual-language connection of CLIP, we propose a text-guided discrimination method, TG-D, to perform discrimination on $F_{\\text{sout}}$ . Specifically, inspired by [20, 61], we introduce learnable prompt pairs (LPP) into our method. As shown in Fig. 4, given the positive prompt and negative prompt, we can obtain the text features $F_{t+}$ and $F_{t-}$ through the CLIP text encoder $E_t$ . Additionally, the final output features of SR image $F_{\\text{sout}}$ and HR image $F_{\\text{hout}}$ can be got from $E_i$ . Then, we adopt the same computation method as in [61] to calculate the relative cosine similarity score between $F_{\\text{sout}}$ and the text features. This takes two steps: firstly, the cosine similarities $s_+$ between $F_{\\text{sout}}$ and $F_{t+}$ , $s_-$ between $F_{\\text{sout}}$ and $f_{t-}$ can be respectively calculated as following:", + "bbox": [ + 511, + 441, + 906, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {+} = \\frac {F _ {s o u t} \\odot F _ {t +}}{| | F _ {s o u t} | | \\cdot | | F _ {t +} | |}, s _ {-} = \\frac {F _ {s o u t} \\odot F _ {t -}}{| | F _ {s o u t} | | \\cdot | | F _ {t -} | |}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 763, + 906, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes vector dot-product and $||\\cdot ||$ represents $L_{2}$ norm. Then, Softmax is used to calculate the final score:", + "bbox": [ + 511, + 801, + 903, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {s r} = \\frac {e ^ {s +}}{e ^ {s +} + e ^ {s -}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 835, + 906, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for $F_{hout}$ of HR image, we can also adopt Eq. (3) and Eq. (4) to calculate the cosine similarity score $s_{hr}$ . Then, we", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28179", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/183166f0e50134e223cd3019886b2a8c423a68e909a462e087c135b20af8fd42.jpg", + "image_caption": [ + "Figure 5. Framework of the proposed SFD-IQA. SFD-IQA is directly based on the well trained Feat-D and LPP, and it doesn't require any additional task-specific training on IQA datasets." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 483, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "utilize $s_{sr}$ and $s_{hr}$ to train the LPP and SRN in an adversarial manner:", + "bbox": [ + 89, + 295, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D _ {t g}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(1 - s _ {h r}\\right) + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log \\left(s _ {s r}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 335, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {G _ {t g}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(s _ {h r}\\right) + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log \\left(1 - s _ {s r}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 367, + 483, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{D_{tg}}$ and $L_{G_{tg}}$ represent the adversarial loss functions for LPP and SRN, respectively. During the adversarial training process, LPP will be more and more sensitive to image quality and TG-D will be able to discriminate $F_{sout}$ in the form of assigning a global score. Finally, the SRN is optimized by the combination of the following losses as:", + "bbox": [ + 89, + 397, + 483, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {t o t a l} = \\lambda_ {1} \\left\\| I _ {h r}, I _ {s r} \\right\\| _ {1} + \\lambda_ {2} \\left\\| \\phi_ {i} \\left(I _ {h r}\\right), \\phi_ {i} \\left(I _ {s r}\\right) \\right\\| _ {1} \\tag {7} \\\\ + \\lambda_ {3} L _ {G _ {f e a}} + \\lambda_ {4} L _ {G _ {t g}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 498, + 483, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $||\\cdot ||_1$ represents the $L_{1}$ norm, $\\phi_i(\\cdot)$ is the layers of pre-trained VGG-19 network[54], $\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4$ represent the hyperparameters for the above losses.", + "bbox": [ + 89, + 544, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Extension to Opinion-Unaware NR-IQA", + "text_level": 1, + "bbox": [ + 89, + 598, + 434, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the process where the Feat-D and TG-D learns to distinguish the distributions of semantic features in an adversarial manner, they simultaneously and implicitly learn to assess image quality. By integrating the well trained Feat-D and LPP, we design a novel OU NR-IQA method, SFD-IQA. As shown in Fig. 5, given an image $x$ , we can obtain its middle semantic features $F_{m1}$ , $F_{m2}$ , $F_{m3}$ , and the final output feature $F_{out}$ through the CLIP image encoder $E_{i}$ . Then, Feat-D will convert $F_{m1}$ , $F_{m2}$ , $F_{m3}$ into a score matrix:", + "bbox": [ + 89, + 619, + 485, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ s _ {i j} \\right] _ {h \\times w} = D \\left(F _ {m 1}, F _ {m 2}, F _ {m 3}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 767, + 483, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[s_{ij}]_{h\\times w}$ denotes the score matrix. Additionally, according to Eq. 2 and Eq. 3, we adopt $F_{out}$ and LPP to calculate the similarity score $s_{lp}$ . Due to CLIP's outstanding contrastive learning training scheme, $F_{out}$ is already sensitive to quality-related text prompts. To further enhance the robustness of SFD-IQA, we also adopt the original antonymic text prompts used in CLIP-IQA (e.g., \"Good photo\" and \"Bad", + "bbox": [ + 89, + 794, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "photo\") to calculate the similarity score $s_o$ with $F_{out}$ . Then we perform a weighted average on $s_o$ and $s_{lp}$ :", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {a v e r} = \\alpha_ {1} s _ {o} + \\alpha_ {2} s _ {l p}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 135, + 906, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ and $\\alpha_{2}$ denotes the weight coefficients. Finally, following [83], we adopt the product of $s_{aver}$ and the average of $[s_{ij}]_{h\\times w}$ as the final IQA score $s_d$ :", + "bbox": [ + 511, + 162, + 906, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {d} = \\left(s _ {a v e r}\\right) \\times \\frac {\\sum_ {i = 0} ^ {h} \\sum_ {j = 0} ^ {w} \\left(\\operatorname {s i g m o i d} \\left(\\left[ s _ {i j} \\right] _ {h \\times w}\\right)\\right)}{h \\times w}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 219, + 906, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where sigmoid denotes the Sigmoid function to adjust the score range to [0, 1]. Notably, we directly apply the Feat-D and TG-D to the above IQA process without any additional task-specific training. Benefiting from the semantic-related discriminative abilities of Feat-D and TG-D, SFD-IQA can significantly improve the OU NR-IQA performance.", + "bbox": [ + 511, + 265, + 906, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 369, + 645, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 395, + 702, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets. For classical SISR, following previous works [15, 30, 39], we train our network using 3450 images from DF2K [1, 58], and the bicubic downsampling is used to synthesize LR-HR training pairs. For real-world SISR and OU NR-IQA, following [39, 65, 83], we adopt DF2K and OutdoorSceneTraining [63] as the training dataset, and the LR-HR training pairs are generated by the degradation pipeline of Real-ESRGAN [65].", + "bbox": [ + 511, + 416, + 905, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test Datasets and Evaluation Metrics. For classical SISR, we test on 4 commonly used benchmarks: Set5 [3], Set14 [74], Urban100 [22], and DIV2K validation set [1], using 4 metrics to comprehensively evaluate the performance of different methods, including fidelity metrics: PSNR and SSIM [66] (calculated on the Y channel in YCbCr space), and perceptual quality metrics: LPIPS [78], DISTS [10]. For real-world SISR, following [62, 70], we employ 2 centeredropped real-world datasets, RealSR [5] and DRealSR [68]. We also create another dataset, RealD2K, by degrading DIV2K validation set under the same pipeline as that in training. We add another metric NIQE [77] for real-world SISR. For OU NR-IQA, following [39, 83], we use two categories of IQA datasets, including SR IQA datasets: PIPAL [24], Ma's dataset [41], and authentically distorted IQA datasets: KonIQ-10k [21], LIVE-itW [16]. To evaluate the IQA performance, we adopt three widely used metrics: PLCC, SRCC and KRCC [25].", + "bbox": [ + 511, + 537, + 908, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For the generator, following [30, 65], we use the competitive RRDB [64] as the backbone. We implement the experiments on NVIDIA 4090 GPUs with PyTorch [49]. We initialize the parameters with the pre-trained fidelity-oriented model. We use Adam [26] optimizer to train the network with a learning rate of $10^{-4}$ .", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28180", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b08e4a6d45ec14a68fa017488ee1262efd48b382c8e817b3660e023ef2b79a2d.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of our method vs. other SOTA methods for $\\times 4\\mathrm{{SR}}$ task. The best and the second-best are marked in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
BenchmarkMetricRankSR GAN [79]SRGAN [29]SFTGAN [63]ESRGAN [64]SPSR [42]LDL [32]DualFormer [39]WGSR [27]SeD [30]RRDB +OursSwinIR +\\( {L}_{GAN} \\)SwinIR +Ours
Set5PSNR ↑29.6530.5230.0630.4430.4031.0331.3031.3331.2231.6331.1031.78
SSIM ↑0.8380.8630.8480.8520.8440.8610.8690.8660.8680.8770.8690.877
LPIPS ↓0.0700.0690.0800.0740.0690.0660.0660.0650.0640.0630.0680.059
DISTS ↓0.1040.1080.1090.0970.0930.0930.0930.1100.0940.0980.0980.089
Set14PSNR ↑26.4527.0126.7426.2826.6427.1027.3927.4027.3727.7927.0527.94
SSIM ↑0.7030.7280.7180.6990.7140.7350.7390.7390.7360.7520.7310.754
LPIPS ↓0.1350.1230.1310.1310.1350.1310.1200.1380.1170.1170.1200.118
DISTS ↓0.1100.1090.1130.0990.0990.0980.0920.1120.0930.0960.0980.091
Urban100PSNR ↑24.4725.0224.3424.3524.8025.4925.6925.7825.9326.1825.8326.57
SSIM ↑0.7290.7780.7240.7340.7470.7670.7730.7810.7800.7870.7850.798
LPIPS ↓0.1380.1300.1340.1230.1210.1100.1150.1350.1070.1090.1000.099
DISTS ↓0.1040.1100.1060.0880.0860.0820.0850.1080.0790.0820.0810.079
DIV2K100PSNR ↑28.0328.1828.0928.2028.1828.9629.2529.1929.2729.7528.9929.90
SSIM ↑0.7670.7880.7710.7770.7720.7950.8020.8040.8030.8150.8030.818
LPIPS ↓0.1210.1220.1330.1150.1130.1010.0970.1040.0940.0970.0940.099
DISTS ↓0.0640.0630.0740.0590.0550.0530.0560.0540.0540.0530.0500.050
", + "bbox": [ + 120, + 128, + 877, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bfaf4422242b935e6f2854901770c6de0764d3d56474ad9661236473bfd15105.jpg", + "image_caption": [ + "Figure 6. Visual comparison of different GAN-based SR methods on $\\times 4$ super-resolution." + ], + "image_footnote": [], + "bbox": [ + 91, + 362, + 906, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Classical SISR", + "text_level": 1, + "bbox": [ + 89, + 710, + 323, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method, we compare it with several SOTA GAN-based SR methods, including [27, 29, 30, 32, 39, 42, 64, 79]. As shown in Tab. 1, compared to other approaches, our method achieves the best PSNR and SSIM scores on all benchmarks, while also holding highly competitive perceptual metrics, meaning that our method can sacrifice less fidelity in exchange for improved perceptual quality. To validate generalizability, we also employ another Transformer-based SR backbone, SwinIR [31]. It's clearly that our method also demonstrates superior performance with SwinIR. Fig. 6 presents the visual comparison, indicating", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that our method generates more realistic semantic textures with fewer artifacts than other methods.", + "bbox": [ + 511, + 712, + 906, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on Real-World SISR", + "text_level": 1, + "bbox": [ + 511, + 756, + 769, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For real-world SISR, we provide quantitative and visual comparison with representative SOTA real-world SISR models including [31, 32, 65, 75]. The quantitative comparison is demonstrated in Table 2. The results show that our SFD achieves the best SSIM score on all datasets, while also maintaining competitive perceptual metrics, especially LPIPS and DISTS, indicating that our approach can achieve better PD trade-off. As shown in Fig. 7, our method produces SR", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28181", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc10b71cceb15f3f3e6ad5da0237c259fc148e087f7b59cbc39c5579dd901707.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 205, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e7d99f196ba262691ed54647598466c018e8f995db6e80303b05c9bdda30a4b.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 207, + 88, + 321, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3688d96a7a6252d3e825b160d96cb4eeb6bd719c03da937210c3b5f686e7b8df.jpg", + "image_caption": [ + "BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 325, + 88, + 439, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39a5e3ce40fdf4ab19cb7a6caf6c2245f192737b2a9069ee043c618025edf165.jpg", + "image_caption": [ + "ReaESRGAN", + "Figure 7. Visual comparison of different GAN-based real-world SR methods on $\\times 4$ super-resolution." + ], + "image_footnote": [], + "bbox": [ + 441, + 88, + 555, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d510b580f555012f5ddfa30e091f618775918d9959ab2d1142e503f28c1f759.jpg", + "image_caption": [ + "LDL" + ], + "image_footnote": [], + "bbox": [ + 558, + 88, + 671, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6096729d2fd689112840df82219cbc58482ebb9e5c70ea378e264e5fafe0fa2a.jpg", + "image_caption": [ + "RealSwinIR" + ], + "image_footnote": [], + "bbox": [ + 674, + 88, + 787, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a932a7fc76a2ff3336e219fac6d526bb82b12ead11cdafef833b0625668a594b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 792, + 88, + 905, + 175 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a73b44f922574fc2f513ba3d3ae004e680a2e5777841898b0edb441c76d48d2f.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of our method vs. other state-of-the-art real-world SISR methods for $\\times 4\\mathrm{{SR}}$ task. The best and the second-best are marked in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
BenchmarkMetricsBSRGAN [75]RealESR GAN [65]LDL [32]RealSwinIR [31]SFD (Ours)
RealD2KPSNR ↑23.5123.5022.9422.6723.40
SSIM ↑0.5850.6040.5970.5810.606
LPIPS ↓0.3790.3680.3730.3680.358
DISTS ↓0.2800.2780.2840.2790.278
NIQE ↓7.5378.3698.3847.8668.600
RealSRPSNR ↑26.3925.6925.2825.9026.12
SSIM ↑0.7650.7620.7570.7680.772
LPIPS ↓0.2670.2730.2750.2620.256
DISTS ↓0.2120.2060.2120.2000.200
NIQE ↓5.6575.8305.9925.9265.797
DRealSRPSNR ↑28.7528.6428.2128.2129.20
SSIM ↑0.8030.8050.8130.7980.815
LPIPS ↓0.2880.2850.2790.2850.284
DISTS ↓0.2140.2090.2130.2100.221
NIQE ↓6.5196.6937.1436.7666.878
", + "bbox": [ + 94, + 279, + 480, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images with sharp edges and fine textures, accompanied by fewer undesirable artifacts. The above results suggest that our SFD can distinguish complex degraded real world images and facilitate the SR network to learn more fine-grained and semantically accurate textures.", + "bbox": [ + 89, + 502, + 483, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results on OU NR-IQA", + "text_level": 1, + "bbox": [ + 89, + 588, + 308, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of our SFD-IQA for OU NR-IQA tasks, we compared SFD-IQA with several representative OU NR-IQA methods, including traditional methods [37, 69, 71, 76, 77], deep learning-based methods [2, 35, 39, 43, 46, 61, 83]. The results are presented in Table 3. Thanks to the excellent ability of Feat-D and TG-D to distinguish between low-quality and high-quality images, learned during the real-world SISR training process, our SFD-IQA achieves the best performance across all four datasets, validating its enhanced ability to predict the perceptual quality of images.", + "bbox": [ + 88, + 609, + 485, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 772, + 254, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effects of Feat-D and TG-D on perceptual SISR. To investigate the effectiveness of our Feat-D and TG-D, we conduct ablation experiments on them using the DIV2K100 test set. As shown in Fig. 8, compared to the baseline (ESR-GAN), Feat-D and TG-D can both independently improve the perceptual quality with less decrease in fidelity. Furthermore, when both Feat-D and TG-D are introduced into the", + "bbox": [ + 89, + 794, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dc3df15b417a06c1b1d63769d21eb1885c8d4bd09404c4640e0c07145f46d51.jpg", + "image_caption": [ + "Figure 8. The effects of Feat-D and TG-D on perceptual SISR." + ], + "image_footnote": [], + "bbox": [ + 544, + 233, + 875, + 381 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c0230ae1c742bae1c7fbceaad6d6c0494be433b6472580bb7e40449a053dd88.jpg", + "image_caption": [ + "t-SNE of vanilla discriminator features", + "Figure 9. The t-SNE visualization of vanilla discriminator features and our Feat-D features. We select 8 categories of images from OST dataset, including \"bear\", \"building\", \"cat\", \"dog\", \"grass\", \"mountain\", \"sky\" and \"water\" and 50 images for each category." + ], + "image_footnote": [], + "bbox": [ + 519, + 412, + 707, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c500fb3b443b488235480078a749d3767fbe6f30231394676d9cc662ac5cddd.jpg", + "image_caption": [ + "t-SNE of our Feat-D features" + ], + "image_footnote": [], + "bbox": [ + 710, + 412, + 900, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "baseline, we can trade the minimum PSNR decrease for the best perceptual quality. The above ablation results strongly verify the superiority of Feat-D and TG-D.", + "bbox": [ + 511, + 628, + 905, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effects of Feat-D and TG-D in SFD-IQA. To investigate the contribution of our Feat-D and TG-D to the proposed SFD-IQA, we conduct ablation studies by respectively remove Feat-D and TG-D from our SFD-IQA. Tab. 4 presents the comparative results on the 3 SR related subtypes of PIPAL [24] training set and 3 sub-types of KonIQ-10k [21] dataset. Compared to the baseline without Feat-D and TG-D, it's clearly that each of Feat-D and TG-D can independently enhance the OU NR-IQA performance to a certain extent. With both Feat-D and TG-D, SFD-IQA significantly improves the performance in most cases. Especially on the GAN-based SR subset of PIPAL, it boosts the PLCC by $86.0\\%$ and SRCC by $28.8\\%$ . The above results demonstrate the effectiveness of Feat-D and TG-D.", + "bbox": [ + 511, + 674, + 908, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "T-SNE analysis for Feat-D. To further demonstrate the", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28182", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/afc02d2e84a292c8b8c2232d0174c63b63c2c403803aca2f86722cf7b5703be8.jpg", + "table_caption": [ + "Table 3. OU NR-IQA comparison between different methods. The best and the second-best are marked in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
IQA MethodsPIPAL [24]KonIQ-10k [21]Ma's dataset [41]LIVE-itW [16]
PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑
QAC [71]0.25660.22200.15030.37190.33970.23020.50710.43870.30550.27200.04570.0370
LPSI [69]0.28970.20310.13630.20660.22390.15040.52930.48960.35430.28770.08340.0524
NIQE [77]0.04390.06080.04010.46370.52910.36660.64270.62610.45200.47900.44930.3062
IL-NIQE [76]0.27480.24120.16300.53160.50570.35040.71550.65480.47510.50390.43940.2985
SNP-NIQE [37]0.18200.18010.12140.63400.62850.44350.62500.56210.39840.52010.46570.3162
dipIQ [43]0.28000.23040.15350.44290.23750.15940.51670.48660.36070.31800.17740.1207
RankIAQ [35]0.36980.37150.25220.50280.49830.34480.51780.49130.36480.45280.43070.2945
RecycleD [83]0.33960.33310.22650.64070.63640.44680.59830.55550.38740.48060.46610.3194
DualFormer [39]0.28960.27870.19730.65430.63210.44350.61720.60540.40440.50680.48970.3312
ContentSep [2]---0.62740.64010.4529---0.51300.50600.3450
CLIPIAQ [61]0.28310.27620.18620.43300.41450.28210.71750.70550.50630.58530.56470.4010
MDFS [46]0.26510.22730.15320.71230.73330.53440.76490.71420.52050.53640.48210.3274
SFD-IQA (Ours)0.42520.40920.28160.75740.74400.54430.80120.78700.58640.71110.69840.5090
", + "bbox": [ + 127, + 108, + 867, + 294 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70ad08bef6f7db25b60d64797937aab2c2b150821d64df80257faaee1f06827c.jpg", + "table_caption": [ + "Table 4. OU NR-IQA performance comparison on sub-types of PIPAL training set and KonIQ-10k dataset. \"Baseline\" refers to the model without neither Feat-D nor TG-D. \"W/ Feat-D\" and \"w/. TG-D\" refer to the models that only include Feat-D or TG-D." + ], + "table_footnote": [], + "table_body": "
MetricsDataset. Sub-TypesBaselinew/.Feat-Dw/.TG-DSFD-IQA (Ours)
PLCCPIPAL.Trad.SR0.37870.36840.39100.3797 (+0.26%)
PIPAL.PSNR.SR0.42900.45040.45470.4657 (+8.55%)
PIPAL.GAN.SR0.21540.39580.24300.4006 (+86.0%)
KonIQ-10k.Train0.70920.74840.74080.7581 (+6.90%)
KonIQ-10k.Val0.70550.74520.73500.7528 (+6.70%)
KonIQ-10k.Test0.68650.74930.74650.7576 (+10.4%)
SRCCPIPAL.Trad.SR0.35450.34200.36600.3522 (-0.64%)
PIPAL.PSNR.SR0.44320.47970.46720.4932 (+11.3%)
PIPAL.GAN.SR0.23760.30480.25980.3061 (+28.8%)
KonIQ-10k.Train0.68490.73250.71300.7441 (+4.92%)
KonIQ-10k.Val0.68030.72640.70720.7373 (+4.51%)
KonIQ-10k.Test0.68620.73590.71310.7475 (+8.89%)
", + "bbox": [ + 99, + 377, + 470, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "effectiveness of our Feat-D, we use t-SNE [59] to visualize the middle features of the VGG-style vanilla discriminator and our Feat-D. We select 8 categories of images from the OST [63] dataset, and 50 images are randomly selected for each category. We visualize features after the 3rd upsampling layer of our Feat-D and after the 2nd BN layer of the vanilla discriminator, respectively. As shown in Fig. 9, the feature clustering of our Feat-D is much better than that of the vanilla discriminator. This strongly proves that the proposed Feat-D is semantic-aware and can perform more fine-grained discrimination, encouraging the SR network to generate more more realistic semantic textures.", + "bbox": [ + 88, + 537, + 482, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of our method on other image restoration tasks. We preliminarily explore the effectiveness of our approach in addressing other image restoration tasks, such as image deraining and image deblurring. We select a commonly used image restoration baselines NAFNet [8], and conduct training and testing on the Rain100L [73] and GoPro [45] datasets for the two tasks. As shown in Tab. 5, compared to fidelity-oriented baseline methods and vanilla discriminator based methods, our method significantly improves perceptual metrics while maintaining a high level of fidelity on both two tasks, demonstrating the potential of our method to be integrated into other image", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/54e048b6ee35751bf4315ebca67a15d70cd64e120cbcde425eacb444d8c58530.jpg", + "table_caption": [ + "Table 5. The effectiveness of our method on image deraining and image deblurring. \"VD\" means vanilla discriminator." + ], + "table_footnote": [], + "table_body": "
MethodRain100L [73]GoPro [45]
PSNR ↑LPIPS ↓DISTS ↓PSNR ↑LPIPS ↓DISTS ↓
NAFNet37.000.0220.03132.870.0930.072
+ VD36.390.0150.02032.010.0680.036
+ Ours36.490.0140.01932.190.0640.034
", + "bbox": [ + 521, + 349, + 890, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "restoration tasks for better perceptual quality.", + "bbox": [ + 511, + 439, + 813, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 467, + 633, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a semantic feature discrimination method, SFD, for perceptual SISR and OU NR-IQA. We first design a feature discriminator (Feat-D) to discriminate the semantic features from CLIP, effectively encouraging the feature distributions of SR images to match that of HR images. Then we propose a text-guided discrimination (TG-D) method by training the learnable prompt pairs (LPP) in an adversarial manner to further enhance the discriminative ability. Our SFD facilitates the SR network to generate more fine-grained and more realistic semantic textures, achieving better perception-distortion trade-off. Furthermore, by effectively integrating the well trained Feat-D and LPP, we significantly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks validate the effectiveness of our approach.", + "bbox": [ + 509, + 492, + 906, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09, the TCL Science and Technology Innovation Fund under grant 25JZH008, and the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025.", + "bbox": [ + 509, + 750, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28183", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5", + "[2] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 3, 7, 8", + "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference (BMVC), 2012. 5", + "[4] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2", + "[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5", + "[6] Du Chen, Zhengqiang Zhang, Jie Liang, and Lei Zhang. Ssl: A self-similarity loss for improving generative image superresolution. In ACM Multimedia 2024, 2024. 2", + "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2", + "[8] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 8", + "[9] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25974-25984, 2024. 2", + "[10] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 5", + "[11] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1, 3", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 3", + "[13] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision and pattern recognition, pages 1103-1112, 2022. 2, 3", + "[14] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1", + "[15] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 3, 5", + "[16] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5, 8", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[18] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 3", + "[19] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1", + "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 3, 4", + "[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5, 7, 8", + "[22] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5", + "[23] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5, 7, 8", + "[25] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938. 5", + "[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[27] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 3, 6" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28184", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 1, 2, 3", + "[29] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 3, 6", + "[30] Bingchen Li, Xin Li, Hanxin Zhu, Yeying Jin, Ruoyu Feng, Zhizheng Zhang, and Zhibo Chen. Sed: Semantic-aware discriminator for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25784-25795, 2024. 2, 3, 4, 5, 6", + "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 6, 7", + "[32] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 2, 3, 6, 7", + "[33] Xiangyu Liao, Tianheng Zheng, Jiayu Zhong, Pingping Zhang, and Chao Ren. Asymmetric mask scheme for self-supervised real image denoising. arXiv preprint arXiv:2407.06514, 2024. 1", + "[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 3", + "[35] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 3, 7, 8", + "[36] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1", + "[37] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2019. 7, 8", + "[38] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2", + "[39] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13243-13253, 2023. 2, 3, 5, 6, 7, 8", + "[40] Chenxi Ma. Uncertainty-aware gan for single image super resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4071-4079, 2024. 2, 3", + "[41] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 158:1-16, 2017. 5, 8", + "[42] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020. 6", + "[43] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 3, 7, 8", + "[44] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 2, 3", + "[45] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 8", + "[46] Zhangkai Ni, Yue Liu, Keyan Ding, Wenhan Yang, Hanli Wang, and Shiqi Wang. Opinion-unaware blind image quality assessment using multi-scale deep feature statistics. IEEE Transactions on Multimedia, 2024. 3, 7, 8", + "[47] JoonKyu Park, Sanghyun Son, and Kyoung Mu Lee. Content-aware local gan for photo-realistic super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10585-10594, 2023. 2, 3", + "[48] Seung Ho Park, Young Su Moon, and Nam Ik Cho. Perception-oriented single image super-resolution using optimal objective estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1725-1735, 2023. 2", + "[49] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[50] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28185", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1", + "[53] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8207-8216, 2020. 2", + "[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2, 5", + "[55] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 3", + "[56] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 1, 2, 3", + "[57] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 1, 2, 3", + "[58] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5", + "[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8", + "[60] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 1", + "[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 4, 7, 8", + "[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 5", + "[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 2, 5, 6, 8", + "[64] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 3, 5, 6", + "[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF inter" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national conference on computer vision, pages 1905-1914, 2021. 2, 3, 5, 6, 7", + "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 1", + "[68] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision - ECCV 2020, pages 101-117. Springer International Publishing, 2020. 5", + "[69] Qingbo Wu, Zhou Wang, and Hongliang Li. A highly efficient method for blind image quality assessment. In 2015 IEEE International Conference on Image Processing (ICIP), pages 339-343. IEEE, 2015. 3, 7, 8", + "[70] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25456-25467, 2024. 3, 5", + "[71] Wufeng Xue, Lei Zhang, and Xuanqin Mou. Learning without human scores for blind image quality assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 995-1002, 2013. 3, 7, 8", + "[72] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5791-5800, 2020. 2, 3", + "[73] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 8", + "[74] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5", + "[75] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 3, 6, 7", + "[76] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 7, 8", + "[77] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 5, 7, 8", + "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28186", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[79] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Super resolution generative adversarial networks with learning to rank. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7149-7166, 2021. 3, 6", + "[80] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image super-resolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1, 2, 3", + "[81] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2, 3", + "[82] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3", + "[83] Yunan Zhu, Haichuan Ma, Jialun Peng, Dong Liu, and Zhiwei Xiong. Recycling discriminator: Towards opinion-unaware image quality assessment using Wasserstein gan. In Proceedings of the 29th ACM International Conference on Multimedia, pages 116-125, 2021. 2, 3, 5, 7, 8" + ], + "bbox": [ + 91, + 90, + 483, + 444 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "28187", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_model.json b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9c9effd5ee5d70b6542d38204755e6434edad43 --- /dev/null +++ b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_model.json @@ -0,0 +1,2677 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.177 + ], + "angle": 0, + "content": "Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.204, + 0.852, + 0.223 + ], + "angle": 0, + "content": "Guanglu Dong Xiangyu Liao Mingyang Li Guihuan Guo Chao Ren\\*" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.223, + 0.812, + 0.24 + ], + "angle": 0, + "content": "College of Electronics and Information Engineering, Sichuan University, China" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.242, + 0.892, + 0.257 + ], + "angle": 0, + "content": "{dongguanglu, liaoxiangyul, 2024222055110, guoguihuan}@stu.scu.edu.cn, chaoren@scu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.486, + 0.763 + ], + "angle": 0, + "content": "Generative Adversarial Networks (GANs) have been widely applied to image super-resolution (SR) to enhance the perceptual quality. However, most existing GAN-based SR methods typically perform coarse-grained discrimination directly on images and ignore the semantic information of images, making it challenging for the super resolution networks (SRN) to learn fine-grained and semantic-related texture details. To alleviate this issue, we propose a semantic feature discrimination method, SFD, for perceptual SR. Specifically, we first design a feature discriminator (FeatD), to discriminate the pixel-wise middle semantic features from CLIP, aligning the feature distributions of SR images with that of high-quality images. Additionally, we propose a text-guided discrimination method (TG-D) by introducing learnable prompt pairs (LPP) in an adversarial manner to perform discrimination on the more abstract output feature of CLIP, further enhancing the discriminative ability of our method. With both Feat-D and TG-D, our SFD can effectively distinguish between the semantic feature distributions of low-quality and high-quality images, encouraging SRN to generate more realistic and semantic-relevant textures. Furthermore, based on the trained Feat-D and LPP, we propose a novel opinion-unaware no-reference image quality assessment (OU NR-IQA) method, SFD-IQA, greatly improving OU NR-IQA performance without any additional targeted training. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our proposed methods. Code is available at https://github.com/GuangluDong0728/SFD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.222, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.485, + 0.879 + ], + "angle": 0, + "content": "As one of the most important tasks in low-level computer vision [14, 19, 33, 36, 50, 52, 60, 67], single image super-resolution (SISR) [11, 18, 28, 34, 56, 57, 62, 80] aims to accurately reconstruct high-resolution (HR) images from the" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.29, + 0.907, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.495, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Figure 1. Perception-distortion trade-off comparison between our proposed SFD and other SOTA GAN-based SISR methods." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.543, + 0.907, + 0.762 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.764, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Figure 2. Relative PLCC comparison between the proposed SFD-IQA and other OU NR-IQA methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "given low-resolution (LR) counterparts. In recent years, with advancements in deep learning, SISR has achieved rapid and substantial progress. Since Dong et al. [11] first introduce deep learning to SISR task, numerous studies have emerged to explore more advanced network architectures, including" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28176" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.403, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.251, + 0.35, + 0.262 + ], + "angle": 0, + "content": "(a) Vanilla GAN-based SISR methods" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.089, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.251, + 0.862, + 0.263 + ], + "angle": 0, + "content": "(b) Our proposed SFD which performs discrimination on the semantic features" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.265, + 0.907, + 0.335 + ], + "angle": 0, + "content": "Figure 3. (a) Vanilla GAN-based SISR methods: performing discrimination on images; (b) Our proposed SFD: performing discrimination on the semantic features. Vanilla GAN-based SISR typically uses an image-wise, patch-wise, or pixel-wise discriminator to determine if the distribution of SR images align with that of high-quality real-world images, ignoring the semantic information of images. Such coarse-grained discrimination makes it challenging for the SR network to reconstruct fine-grained, semantic-relevant textures. In contrast, we propose to perform discrimination on the semantic features, encouraging the SR network to generate more realistic semantic textures." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.483, + 0.452 + ], + "angle": 0, + "content": "CNN-based models [28, 34, 56, 57, 81, 82] and Transformer-based models [7, 13, 31, 38, 72, 80], achieving continuous improvements in fidelity-oriented performance. However, these works primarily employ pixel-wise L1 or MSE losses for the training process, which often produce overly smooth results and fail to generate realistic texture details [4, 29]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.458, + 0.483, + 0.669 + ], + "angle": 0, + "content": "To overcome the above limitations, a series of generative adversarial networks (GANs) [17] based SISR models [6, 15, 29, 30, 32, 40, 47, 48, 63, 65] have been proposed to generate more realistic images that align more closely with human visual perception by effectively balancing perceptual quality and fidelity. As shown in Figure 3 (a), most existing GAN-based methods perform discrimination on the images using an image-wise [54], patch-wise [23], or pixel-wise [53] discriminator to distinguish between the distributions of SR images and HR images. However, directly discriminating on the images is coarse-grained and lacks semantic awareness. This leads to that the textures of generated images do not match its semantic and often come with artificial artifacts, making it challenging to produce realistic semantic textures." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, different from existing works, we propose to perform discrimination on the semantic features, aiming to align the semantic feature distributions of SR images more closely with that of HR images. Considering that CLIP [51] can extract more interpretable, semantic-aware and quality-relevant features [9, 20, 30, 61], we select CLIP's image encoder as the semantic feature extractor and propose two methods to perform joint discrimination on the semantic features. Specifically, as shown in Figure 3 (b), as for the pixel-wise middle semantic features, we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR images and HR images, encouraging the semantic feature distributions of SR images to match that of HR images. As for the final output feature which is more abstract and global, we propose a text-guided discrimination" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.481 + ], + "angle": 0, + "content": "method, TG-D, by introducing learnable prompt pairs (LPP) into an adversarial training process to make them sensitive to global image quality and thus able to discriminate the abstract semantic features, further enhancing our method's overall discriminative capability. With both Feat-D and TG-D, the SR network is able to learn more fine-grained semantic feature distributions, ultimately producing more realistic and semantic-related texture details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.728 + ], + "angle": 0, + "content": "Additionally, existing researches [39, 40, 83] have shown that during the training process where the discriminators in GAN-based SISR learn to distinguish the realness of image distributions, they simultaneously and implicitly learn to assess image quality, making them highly applicable to opinion-unaware no-reference image quality assessment (OU NR-IQA). Simultaneously, recent studies [20, 44] have also shown the advantage of CLIP for OU NR-IQA. In this paper, to further improve OU NR-IQA performance, we propose a novel OU NR-IQA method, SFD-IQA, by integrating the well trained Feat-D and TG-D which conveniently combines both of the above advantages. Thanks to the superior semantic feature discrimination capability of Feat-D and TG-D and the inherent image quality correlation of CLIP, we greatly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.909, + 0.881 + ], + "angle": 0, + "content": "As shown in Fig. 1, our SFD demonstrates a better perception-distortion (PD) trade-off on the PSNR-LPIPS-DISTS plane. Compared to other state-of-the-art (SOTA) GAN-based SISR methods, our LPIPS and PSNR scores are the best among all methods with competitive DISTS scores. Additionally, Fig. 2 illustrates a comparison between our SFD-IQA and other SOTA OU NR-IQA methods, where SFD-IQA outperforms other methods across both SR IQA datasets and authentically distorted IQA datasets. The main contributions of this work are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "- We propose a semantic feature discrimination (SFD)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28177" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.483, + 0.197 + ], + "angle": 0, + "content": "method for perceptual SISR, which consists of a feature discriminator (Feat-D) based discrimination and a text-guided discrimination (TG-D). Our SFD can effectively distinguish between the CLIP semantic feature distributions of SR and HR images, encouraging SR networks to learn more fine-grained and semantic-related textures, and achieving better perception-distortion (PD) trade-off." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.198, + 0.483, + 0.287 + ], + "angle": 0, + "content": "- We propose a novel OU NR-IQA method, SFD-IQA, by effectively integrating the two well trained feature discrimination methods. Without any additional task-specific training on IQA datasets, SFD-IQA can significantly enhance the OU NR-IQA performance for both SR images and other authentically distorted images." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.288, + 0.483, + 0.395 + ], + "angle": 0, + "content": "- Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our method. SFD can help SR networks achieve better PD trade-off and generate more realistic semantic textures, and SFD-IQA is able to greatly improve OU NR-IQA performance. Moreover, our approach can be easily integrated into other GAN-based image restoration methods." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.198, + 0.483, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.241, + 0.426 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.307, + 0.452 + ], + "angle": 0, + "content": "2.1. Fidelity-Oriented SISR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Recently, deep learning based single image super-resolution (SISR) [11, 18, 28, 29, 62, 70, 81, 82] has made tremendous advancements. Since SRCNN [11] first introduces CNN into SISR, numerous SISR backbones [18, 28, 34, 56, 57, 81, 82] have continued to emerge to enhance the performance of fidelity-oriented SISR. Specifically, RCAN [81] introduces channel attention into SISR network, and RDN [82] leverages dense connection to fully utilize hierarchical features for SISR. Furthermore, based on the success of the vision transformer [12] framework, [13, 31, 72, 80] further improve the SISR performance. However, in order to maximize fidelity, most of the aforementioned methods only employ pixel-wise loss functions to optimize the training process and ignore the perceptual quality, which often leads to overly smooth results and fails to achieve visually appealing results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.424, + 0.714 + ], + "angle": 0, + "content": "2.2. Perception-Oriented GAN-based SISR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To improve perceptual quality, GAN-based SISR [15, 27, 30, 32, 39, 47, 64, 65, 75, 79] has received constant attention in recent years. Specifically, SRGAN [29] introduces GANs into SR tasks to obtain SR images more in line with human perception. DualFormer [39] applies an improved spectral discriminator to improve perceptual quality. Sed [30] incorporates the semantic features into the vanilla discriminator to finely distinguish the image distributions. EA-ADAM [55] develops a new optimizer from a multi-objective optimization perspective to achieve a better perception-distortion tradeoff. However, due to the lack of semantic-relevant perception, most GAN-based SR networks are difficult to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "produce fine-grained semantic textures, result in exhibiting artificial artifacts. In this paper, we propose to discriminate the semantic features of CLIP to guide the SR network to learn more fine-grained semantic feature distributions and generate more realistic semantic textures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.175, + 0.76, + 0.192 + ], + "angle": 0, + "content": "2.3. Opinion-Unaware NR-IQA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.908, + 0.471 + ], + "angle": 0, + "content": "OU NR-IQA [2, 43, 46, 61, 69, 71, 77, 83] aims to evaluate the quality of images without human-labeled data. Traditional OU NR-IQA methods [69, 76, 77] are typically based on natural scene statistics (NSS), which estimate distribution parameters to assess quality. Researches on deep learning-based OU NR-IQA are relatively sparse, with most methods centered around learning-to-rank [35, 43] and contrastive learning [44]. Recently, with the emergence of the pretrained vision-language model CLIP [51], [20, 61] leverage the rich visual-linguistic priors encapsulated in CLIP to evaluate image perceptual quality. Additionally, RecycleD [83] proposes to recycle the trained discriminator for GAN-based SISR to predict perceptual quality. Following [83], [39] and [40] further enhance the IQA ability of the discriminators in GAN-based SISR. However, these discriminator-based methods only rely on the trained semantic-unaware discriminator and ignore the assessment for semantic, which is difficult to achieve satisfactory performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.72, + 0.5 + ], + "angle": 0, + "content": "3. Method and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.908, + 0.6 + ], + "angle": 0, + "content": "In this section, we provide a detailed explanation and analysis of our method. Firstly, we describe the overall framework of our method in Sec. 3.1. Next, we provide a detailed introduction and analysis of the two proposed semantic feature discrimination methods in Sec. 3.2. Lastly, we discuss how to extend our method to OU NR-IQA in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.704, + 0.622 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.87 + ], + "angle": 0, + "content": "We propose a novel semantic feature discrimination (SFD) method for perceptual SISR, aiming to achieve better perception-distortion (PD) trade-off. Differing from vanilla GAN-based SR methods that only consider coarse-grained image distributions with poor semantic relevance, our SFD focuses on the distributions of semantic features. As illustrated in Fig. 4, given the input low-resolution image \\( I_{lr} \\), we can obtain its corresponding SR image \\( I_{sr} \\) through the SRN. Then, through the pre-trained CLIP [51] image encoder \\( E_{i} \\) we can obtain rich semantic features of \\( I_{sr} \\) with different scales and characteristics, including the multi-scale pixelwise middle features \\( F_{smi} \\), \\( F_{smi} \\), \\( F_{smi} \\in \\mathbb{R}^{C_i \\times H_i \\times W_i} \\) (\\( i = 1,2,3 \\)), and the more abstract final output feature \\( F_{sout} \\in \\mathbb{R}^{1 \\times L} \\). To better leverage these semantic features to improve the robustness of our approach, we propose two methods to separately perform discrimination on them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "On one hand, as for the pixel-wise middle features, we design a feature discriminator (Feat-D) to distinguish be" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28178" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.91, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.358, + 0.908, + 0.387 + ], + "angle": 0, + "content": "Figure 4. Framework of the proposed SFD. SFD consists of a feature discriminator (Feat-D) and a text-guided discrimination (TG-D) method, Feat-D and TG-D are used to perform discrimination on the middle features and final output features from CLIP, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.413, + 0.484, + 0.518 + ], + "angle": 0, + "content": "tween the feature distributions of SR and HR images. On the other hand, we propose a text-guided discrimination (TG-D) method to perform discrimination on \\( F_{\\text{out}} \\), further enhancing the overall discriminative ability. With both Feat-D and TG-D, SFD is able to facilitate the SR network (SRN) to learn more fine-grained semantic feature distributions and generate more realistic and semantic-related textures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.411, + 0.543 + ], + "angle": 0, + "content": "3.2. Discrimination on Semantic Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Semantic Feature Discriminator. As for the middle features of CLIP which contain multi-scale pixel-wise semantic information [30], we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR and HR images, encouraging the semantic feature distribution of SR images to match that of HR images. Specifically, as shown in Figure 4, we can obtain the middle features \\( F_{sm1} \\), \\( F_{sm2} \\), \\( F_{sm3} \\) of SR image \\( I_{sr} \\), and \\( F_{hm1} \\), \\( F_{hm2} \\), \\( F_{hm3} \\) of HR image \\( I_{hr} \\) through the freezing CLIP image encoder \\( E_i \\). Then, the Feat-D is used to perform discrimination on those pixel-wise semantic features. During the adversarial training process, the Feat-D and the SRN will be optimized respectively using the following adversarial loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.759, + 0.483, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {D _ {f e a}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(1 - D \\left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\\right)\\right) \\tag {1} \\\\ + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log (D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.811, + 0.484, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {G _ {f e a}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(D \\left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\\right)\\right) \\tag {2} \\\\ + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log (1 - D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( P_{F_{hr}} \\) and \\( P_{F_{sr}} \\) represent the feature distributions of HR and SR images, respectively. Benefiting from the excellent discriminative ability of Feat-D, the SRN can learn more" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.905, + 0.442 + ], + "angle": 0, + "content": "fine-grained semantic feature distributions aligned with HR images, generating more semantically relevant textures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.908, + 0.76 + ], + "angle": 0, + "content": "Learnable Prompt Pairs Driven Text-Guided Discrimination. Discriminating the final output feature \\( F_{\\text{sout}} \\) which is more global and abstract is expected to further enhance the overall performance of our method. However, differing from the middle features, \\( F_{\\text{sout}} \\) lacks the pixel-wise semantic information due to its more abstract nature, making it less suitable for Feat-D's pixel-wise discrimination process. Therefore, leveraging the inherent visual-language connection of CLIP, we propose a text-guided discrimination method, TG-D, to perform discrimination on \\( F_{\\text{sout}} \\). Specifically, inspired by [20, 61], we introduce learnable prompt pairs (LPP) into our method. As shown in Fig. 4, given the positive prompt and negative prompt, we can obtain the text features \\( F_{t+} \\) and \\( F_{t-} \\) through the CLIP text encoder \\( E_t \\). Additionally, the final output features of SR image \\( F_{\\text{sout}} \\) and HR image \\( F_{\\text{hout}} \\) can be got from \\( E_i \\). Then, we adopt the same computation method as in [61] to calculate the relative cosine similarity score between \\( F_{\\text{sout}} \\) and the text features. This takes two steps: firstly, the cosine similarities \\( s_+ \\) between \\( F_{\\text{sout}} \\) and \\( F_{t+} \\), \\( s_- \\) between \\( F_{\\text{sout}} \\) and \\( f_{t-} \\) can be respectively calculated as following:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.764, + 0.907, + 0.798 + ], + "angle": 0, + "content": "\\[\ns _ {+} = \\frac {F _ {s o u t} \\odot F _ {t +}}{| | F _ {s o u t} | | \\cdot | | F _ {t +} | |}, s _ {-} = \\frac {F _ {s o u t} \\odot F _ {t -}}{| | F _ {s o u t} | | \\cdot | | F _ {t -} | |}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.905, + 0.832 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes vector dot-product and \\(||\\cdot ||\\) represents \\(L_{2}\\) norm. Then, Softmax is used to calculate the final score:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.836, + 0.907, + 0.866 + ], + "angle": 0, + "content": "\\[\ns _ {s r} = \\frac {e ^ {s +}}{e ^ {s +} + e ^ {s -}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "As for \\( F_{hout} \\) of HR image, we can also adopt Eq. (3) and Eq. (4) to calculate the cosine similarity score \\( s_{hr} \\). Then, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.271 + ], + "angle": 0, + "content": "Figure 5. Framework of the proposed SFD-IQA. SFD-IQA is directly based on the well trained Feat-D and LPP, and it doesn't require any additional task-specific training on IQA datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.327 + ], + "angle": 0, + "content": "utilize \\( s_{sr} \\) and \\( s_{hr} \\) to train the LPP and SRN in an adversarial manner:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.336, + 0.484, + 0.368 + ], + "angle": 0, + "content": "\\[\nL _ {D _ {t g}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(1 - s _ {h r}\\right) + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log \\left(s _ {s r}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.398 + ], + "angle": 0, + "content": "\\[\nL _ {G _ {t g}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(s _ {h r}\\right) + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log \\left(1 - s _ {s r}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.485, + 0.49 + ], + "angle": 0, + "content": "where \\(L_{D_{tg}}\\) and \\(L_{G_{tg}}\\) represent the adversarial loss functions for LPP and SRN, respectively. During the adversarial training process, LPP will be more and more sensitive to image quality and TG-D will be able to discriminate \\(F_{sout}\\) in the form of assigning a global score. Finally, the SRN is optimized by the combination of the following losses as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.499, + 0.484, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {t o t a l} = \\lambda_ {1} \\left\\| I _ {h r}, I _ {s r} \\right\\| _ {1} + \\lambda_ {2} \\left\\| \\phi_ {i} \\left(I _ {h r}\\right), \\phi_ {i} \\left(I _ {s r}\\right) \\right\\| _ {1} \\tag {7} \\\\ + \\lambda_ {3} L _ {G _ {f e a}} + \\lambda_ {4} L _ {G _ {t g}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.484, + 0.59 + ], + "angle": 0, + "content": "where \\(||\\cdot ||_1\\) represents the \\(L_{1}\\) norm, \\(\\phi_i(\\cdot)\\) is the layers of pre-trained VGG-19 network[54], \\(\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4\\) represent the hyperparameters for the above losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.599, + 0.436, + 0.616 + ], + "angle": 0, + "content": "3.3. Extension to Opinion-Unaware NR-IQA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.486, + 0.759 + ], + "angle": 0, + "content": "During the process where the Feat-D and TG-D learns to distinguish the distributions of semantic features in an adversarial manner, they simultaneously and implicitly learn to assess image quality. By integrating the well trained Feat-D and LPP, we design a novel OU NR-IQA method, SFD-IQA. As shown in Fig. 5, given an image \\( x \\), we can obtain its middle semantic features \\( F_{m1} \\), \\( F_{m2} \\), \\( F_{m3} \\), and the final output feature \\( F_{out} \\) through the CLIP image encoder \\( E_{i} \\). Then, Feat-D will convert \\( F_{m1} \\), \\( F_{m2} \\), \\( F_{m3} \\) into a score matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.768, + 0.484, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\left[ s _ {i j} \\right] _ {h \\times w} = D \\left(F _ {m 1}, F _ {m 2}, F _ {m 3}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.902 + ], + "angle": 0, + "content": "where \\([s_{ij}]_{h\\times w}\\) denotes the score matrix. Additionally, according to Eq. 2 and Eq. 3, we adopt \\(F_{out}\\) and LPP to calculate the similarity score \\(s_{lp}\\). Due to CLIP's outstanding contrastive learning training scheme, \\(F_{out}\\) is already sensitive to quality-related text prompts. To further enhance the robustness of SFD-IQA, we also adopt the original antonymic text prompts used in CLIP-IQA (e.g., \"Good photo\" and \"Bad" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "photo\") to calculate the similarity score \\( s_o \\) with \\( F_{out} \\). Then we perform a weighted average on \\( s_o \\) and \\( s_{lp} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.136, + 0.907, + 0.151 + ], + "angle": 0, + "content": "\\[\ns _ {a v e r} = \\alpha_ {1} s _ {o} + \\alpha_ {2} s _ {l p}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.163, + 0.908, + 0.21 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) denotes the weight coefficients. Finally, following [83], we adopt the product of \\(s_{aver}\\) and the average of \\([s_{ij}]_{h\\times w}\\) as the final IQA score \\(s_d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.22, + 0.907, + 0.256 + ], + "angle": 0, + "content": "\\[\ns _ {d} = \\left(s _ {a v e r}\\right) \\times \\frac {\\sum_ {i = 0} ^ {h} \\sum_ {j = 0} ^ {w} \\left(\\operatorname {s i g m o i d} \\left(\\left[ s _ {i j} \\right] _ {h \\times w}\\right)\\right)}{h \\times w}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.357 + ], + "angle": 0, + "content": "where sigmoid denotes the Sigmoid function to adjust the score range to [0, 1]. Notably, we directly apply the Feat-D and TG-D to the above IQA process without any additional task-specific training. Benefiting from the semantic-related discriminative abilities of Feat-D and TG-D, SFD-IQA can significantly improve the OU NR-IQA performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.646, + 0.387 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.396, + 0.704, + 0.412 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.538 + ], + "angle": 0, + "content": "Training Datasets. For classical SISR, following previous works [15, 30, 39], we train our network using 3450 images from DF2K [1, 58], and the bicubic downsampling is used to synthesize LR-HR training pairs. For real-world SISR and OU NR-IQA, following [39, 65, 83], we adopt DF2K and OutdoorSceneTraining [63] as the training dataset, and the LR-HR training pairs are generated by the degradation pipeline of Real-ESRGAN [65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Test Datasets and Evaluation Metrics. For classical SISR, we test on 4 commonly used benchmarks: Set5 [3], Set14 [74], Urban100 [22], and DIV2K validation set [1], using 4 metrics to comprehensively evaluate the performance of different methods, including fidelity metrics: PSNR and SSIM [66] (calculated on the Y channel in YCbCr space), and perceptual quality metrics: LPIPS [78], DISTS [10]. For real-world SISR, following [62, 70], we employ 2 centeredropped real-world datasets, RealSR [5] and DRealSR [68]. We also create another dataset, RealD2K, by degrading DIV2K validation set under the same pipeline as that in training. We add another metric NIQE [77] for real-world SISR. For OU NR-IQA, following [39, 83], we use two categories of IQA datasets, including SR IQA datasets: PIPAL [24], Ma's dataset [41], and authentically distorted IQA datasets: KonIQ-10k [21], LIVE-itW [16]. To evaluate the IQA performance, we adopt three widely used metrics: PLCC, SRCC and KRCC [25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. For the generator, following [30, 65], we use the competitive RRDB [64] as the backbone. We implement the experiments on NVIDIA 4090 GPUs with PyTorch [49]. We initialize the parameters with the pre-trained fidelity-oriented model. We use Adam [26] optimizer to train the network with a learning rate of \\(10^{-4}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28180" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of our method vs. other SOTA methods for \\( \\times 4\\mathrm{{SR}} \\) task. The best and the second-best are marked in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.129, + 0.879, + 0.351 + ], + "angle": 0, + "content": "
BenchmarkMetricRankSR GAN [79]SRGAN [29]SFTGAN [63]ESRGAN [64]SPSR [42]LDL [32]DualFormer [39]WGSR [27]SeD [30]RRDB +OursSwinIR +\\( {L}_{GAN} \\)SwinIR +Ours
Set5PSNR ↑29.6530.5230.0630.4430.4031.0331.3031.3331.2231.6331.1031.78
SSIM ↑0.8380.8630.8480.8520.8440.8610.8690.8660.8680.8770.8690.877
LPIPS ↓0.0700.0690.0800.0740.0690.0660.0660.0650.0640.0630.0680.059
DISTS ↓0.1040.1080.1090.0970.0930.0930.0930.1100.0940.0980.0980.089
Set14PSNR ↑26.4527.0126.7426.2826.6427.1027.3927.4027.3727.7927.0527.94
SSIM ↑0.7030.7280.7180.6990.7140.7350.7390.7390.7360.7520.7310.754
LPIPS ↓0.1350.1230.1310.1310.1350.1310.1200.1380.1170.1170.1200.118
DISTS ↓0.1100.1090.1130.0990.0990.0980.0920.1120.0930.0960.0980.091
Urban100PSNR ↑24.4725.0224.3424.3524.8025.4925.6925.7825.9326.1825.8326.57
SSIM ↑0.7290.7780.7240.7340.7470.7670.7730.7810.7800.7870.7850.798
LPIPS ↓0.1380.1300.1340.1230.1210.1100.1150.1350.1070.1090.1000.099
DISTS ↓0.1040.1100.1060.0880.0860.0820.0850.1080.0790.0820.0810.079
DIV2K100PSNR ↑28.0328.1828.0928.2028.1828.9629.2529.1929.2729.7528.9929.90
SSIM ↑0.7670.7880.7710.7770.7720.7950.8020.8040.8030.8150.8030.818
LPIPS ↓0.1210.1220.1330.1150.1130.1010.0970.1040.0940.0970.0940.099
DISTS ↓0.0640.0630.0740.0590.0550.0530.0560.0540.0540.0530.0500.050
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.363, + 0.907, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.671, + 0.766, + 0.687 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of different GAN-based SR methods on \\(\\times 4\\) super-resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.325, + 0.726 + ], + "angle": 0, + "content": "4.2. Results on Classical SISR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To validate the effectiveness of our method, we compare it with several SOTA GAN-based SR methods, including [27, 29, 30, 32, 39, 42, 64, 79]. As shown in Tab. 1, compared to other approaches, our method achieves the best PSNR and SSIM scores on all benchmarks, while also holding highly competitive perceptual metrics, meaning that our method can sacrifice less fidelity in exchange for improved perceptual quality. To validate generalizability, we also employ another Transformer-based SR backbone, SwinIR [31]. It's clearly that our method also demonstrates superior performance with SwinIR. Fig. 6 presents the visual comparison, indicating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.907, + 0.743 + ], + "angle": 0, + "content": "that our method generates more realistic semantic textures with fewer artifacts than other methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.771, + 0.772 + ], + "angle": 0, + "content": "4.3. Results on Real-World SISR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For real-world SISR, we provide quantitative and visual comparison with representative SOTA real-world SISR models including [31, 32, 65, 75]. The quantitative comparison is demonstrated in Table 2. The results show that our SFD achieves the best SSIM score on all datasets, while also maintaining competitive perceptual metrics, especially LPIPS and DISTS, indicating that our approach can achieve better PD trade-off. As shown in Fig. 7, our method produces SR" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.207, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.179, + 0.158, + 0.188 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.089, + 0.323, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.179, + 0.286, + 0.189 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.089, + 0.44, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.179, + 0.407, + 0.189 + ], + "angle": 0, + "content": "BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.089, + 0.556, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.179, + 0.534, + 0.189 + ], + "angle": 0, + "content": "ReaESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.089, + 0.672, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.179, + 0.628, + 0.188 + ], + "angle": 0, + "content": "LDL" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.089, + 0.789, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.179, + 0.762, + 0.189 + ], + "angle": 0, + "content": "RealSwinIR" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.089, + 0.906, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.179, + 0.861, + 0.188 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.195, + 0.795, + 0.209 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of different GAN-based real-world SR methods on \\(\\times 4\\) super-resolution." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.234, + 0.485, + 0.276 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of our method vs. other state-of-the-art real-world SISR methods for \\( \\times 4\\mathrm{{SR}} \\) task. The best and the second-best are marked in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.28, + 0.481, + 0.476 + ], + "angle": 0, + "content": "
BenchmarkMetricsBSRGAN [75]RealESR GAN [65]LDL [32]RealSwinIR [31]SFD (Ours)
RealD2KPSNR ↑23.5123.5022.9422.6723.40
SSIM ↑0.5850.6040.5970.5810.606
LPIPS ↓0.3790.3680.3730.3680.358
DISTS ↓0.2800.2780.2840.2790.278
NIQE ↓7.5378.3698.3847.8668.600
RealSRPSNR ↑26.3925.6925.2825.9026.12
SSIM ↑0.7650.7620.7570.7680.772
LPIPS ↓0.2670.2730.2750.2620.256
DISTS ↓0.2120.2060.2120.2000.200
NIQE ↓5.6575.8305.9925.9265.797
DRealSRPSNR ↑28.7528.6428.2128.2129.20
SSIM ↑0.8030.8050.8130.7980.815
LPIPS ↓0.2880.2850.2790.2850.284
DISTS ↓0.2140.2090.2130.2100.221
NIQE ↓6.5196.6937.1436.7666.878
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.578 + ], + "angle": 0, + "content": "images with sharp edges and fine textures, accompanied by fewer undesirable artifacts. The above results suggest that our SFD can distinguish complex degraded real world images and facilitate the SR network to learn more fine-grained and semantically accurate textures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.31, + 0.604 + ], + "angle": 0, + "content": "4.4. Results on OU NR-IQA" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.611, + 0.486, + 0.763 + ], + "angle": 0, + "content": "To verify the effectiveness of our SFD-IQA for OU NR-IQA tasks, we compared SFD-IQA with several representative OU NR-IQA methods, including traditional methods [37, 69, 71, 76, 77], deep learning-based methods [2, 35, 39, 43, 46, 61, 83]. The results are presented in Table 3. Thanks to the excellent ability of Feat-D and TG-D to distinguish between low-quality and high-quality images, learned during the real-world SISR training process, our SFD-IQA achieves the best performance across all four datasets, validating its enhanced ability to predict the perceptual quality of images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.255, + 0.787 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.9 + ], + "angle": 0, + "content": "The effects of Feat-D and TG-D on perceptual SISR. To investigate the effectiveness of our Feat-D and TG-D, we conduct ablation experiments on them using the DIV2K100 test set. As shown in Fig. 8, compared to the baseline (ESR-GAN), Feat-D and TG-D can both independently improve the perceptual quality with less decrease in fidelity. Furthermore, when both Feat-D and TG-D are introduced into the" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.234, + 0.877, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.385, + 0.896, + 0.399 + ], + "angle": 0, + "content": "Figure 8. The effects of Feat-D and TG-D on perceptual SISR." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.413, + 0.709, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.532, + 0.701, + 0.542 + ], + "angle": 0, + "content": "t-SNE of vanilla discriminator features" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.413, + 0.901, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.532, + 0.871, + 0.542 + ], + "angle": 0, + "content": "t-SNE of our Feat-D features" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Figure 9. The t-SNE visualization of vanilla discriminator features and our Feat-D features. We select 8 categories of images from OST dataset, including \"bear\", \"building\", \"cat\", \"dog\", \"grass\", \"mountain\", \"sky\" and \"water\" and 50 images for each category." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.674 + ], + "angle": 0, + "content": "baseline, we can trade the minimum PSNR decrease for the best perceptual quality. The above ablation results strongly verify the superiority of Feat-D and TG-D." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.884 + ], + "angle": 0, + "content": "The effects of Feat-D and TG-D in SFD-IQA. To investigate the contribution of our Feat-D and TG-D to the proposed SFD-IQA, we conduct ablation studies by respectively remove Feat-D and TG-D from our SFD-IQA. Tab. 4 presents the comparative results on the 3 SR related subtypes of PIPAL [24] training set and 3 sub-types of KonIQ-10k [21] dataset. Compared to the baseline without Feat-D and TG-D, it's clearly that each of Feat-D and TG-D can independently enhance the OU NR-IQA performance to a certain extent. With both Feat-D and TG-D, SFD-IQA significantly improves the performance in most cases. Especially on the GAN-based SR subset of PIPAL, it boosts the PLCC by \\(86.0\\%\\) and SRCC by \\(28.8\\%\\). The above results demonstrate the effectiveness of Feat-D and TG-D." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "T-SNE analysis for Feat-D. To further demonstrate the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28182" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.09, + 0.889, + 0.104 + ], + "angle": 0, + "content": "Table 3. OU NR-IQA comparison between different methods. The best and the second-best are marked in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.109, + 0.868, + 0.295 + ], + "angle": 0, + "content": "
IQA MethodsPIPAL [24]KonIQ-10k [21]Ma's dataset [41]LIVE-itW [16]
PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑
QAC [71]0.25660.22200.15030.37190.33970.23020.50710.43870.30550.27200.04570.0370
LPSI [69]0.28970.20310.13630.20660.22390.15040.52930.48960.35430.28770.08340.0524
NIQE [77]0.04390.06080.04010.46370.52910.36660.64270.62610.45200.47900.44930.3062
IL-NIQE [76]0.27480.24120.16300.53160.50570.35040.71550.65480.47510.50390.43940.2985
SNP-NIQE [37]0.18200.18010.12140.63400.62850.44350.62500.56210.39840.52010.46570.3162
dipIQ [43]0.28000.23040.15350.44290.23750.15940.51670.48660.36070.31800.17740.1207
RankIAQ [35]0.36980.37150.25220.50280.49830.34480.51780.49130.36480.45280.43070.2945
RecycleD [83]0.33960.33310.22650.64070.63640.44680.59830.55550.38740.48060.46610.3194
DualFormer [39]0.28960.27870.19730.65430.63210.44350.61720.60540.40440.50680.48970.3312
ContentSep [2]---0.62740.64010.4529---0.51300.50600.3450
CLIPIAQ [61]0.28310.27620.18620.43300.41450.28210.71750.70550.50630.58530.56470.4010
MDFS [46]0.26510.22730.15320.71230.73330.53440.76490.71420.52050.53640.48210.3274
SFD-IQA (Ours)0.42520.40920.28160.75740.74400.54430.80120.78700.58640.71110.69840.5090
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.317, + 0.486, + 0.373 + ], + "angle": 0, + "content": "Table 4. OU NR-IQA performance comparison on sub-types of PIPAL training set and KonIQ-10k dataset. \"Baseline\" refers to the model without neither Feat-D nor TG-D. \"W/ Feat-D\" and \"w/. TG-D\" refer to the models that only include Feat-D or TG-D." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.378, + 0.472, + 0.523 + ], + "angle": 0, + "content": "
MetricsDataset. Sub-TypesBaselinew/.Feat-Dw/.TG-DSFD-IQA (Ours)
PLCCPIPAL.Trad.SR0.37870.36840.39100.3797 (+0.26%)
PIPAL.PSNR.SR0.42900.45040.45470.4657 (+8.55%)
PIPAL.GAN.SR0.21540.39580.24300.4006 (+86.0%)
KonIQ-10k.Train0.70920.74840.74080.7581 (+6.90%)
KonIQ-10k.Val0.70550.74520.73500.7528 (+6.70%)
KonIQ-10k.Test0.68650.74930.74650.7576 (+10.4%)
SRCCPIPAL.Trad.SR0.35450.34200.36600.3522 (-0.64%)
PIPAL.PSNR.SR0.44320.47970.46720.4932 (+11.3%)
PIPAL.GAN.SR0.23760.30480.25980.3061 (+28.8%)
KonIQ-10k.Train0.68490.73250.71300.7441 (+4.92%)
KonIQ-10k.Val0.68030.72640.70720.7373 (+4.51%)
KonIQ-10k.Test0.68620.73590.71310.7475 (+8.89%)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.718 + ], + "angle": 0, + "content": "effectiveness of our Feat-D, we use t-SNE [59] to visualize the middle features of the VGG-style vanilla discriminator and our Feat-D. We select 8 categories of images from the OST [63] dataset, and 50 images are randomly selected for each category. We visualize features after the 3rd upsampling layer of our Feat-D and after the 2nd BN layer of the vanilla discriminator, respectively. As shown in Fig. 9, the feature clustering of our Feat-D is much better than that of the vanilla discriminator. This strongly proves that the proposed Feat-D is semantic-aware and can perform more fine-grained discrimination, encouraging the SR network to generate more more realistic semantic textures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The effectiveness of our method on other image restoration tasks. We preliminarily explore the effectiveness of our approach in addressing other image restoration tasks, such as image deraining and image deblurring. We select a commonly used image restoration baselines NAFNet [8], and conduct training and testing on the Rain100L [73] and GoPro [45] datasets for the two tasks. As shown in Tab. 5, compared to fidelity-oriented baseline methods and vanilla discriminator based methods, our method significantly improves perceptual metrics while maintaining a high level of fidelity on both two tasks, demonstrating the potential of our method to be integrated into other image" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Table 5. The effectiveness of our method on image deraining and image deblurring. \"VD\" means vanilla discriminator." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.35, + 0.892, + 0.421 + ], + "angle": 0, + "content": "
MethodRain100L [73]GoPro [45]
PSNR ↑LPIPS ↓DISTS ↓PSNR ↑LPIPS ↓DISTS ↓
NAFNet37.000.0220.03132.870.0930.072
+ VD36.390.0150.02032.010.0680.036
+ Ours36.490.0140.01932.190.0640.034
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.815, + 0.456 + ], + "angle": 0, + "content": "restoration tasks for better perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.634, + 0.484 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.493, + 0.908, + 0.736 + ], + "angle": 0, + "content": "In this paper, we propose a semantic feature discrimination method, SFD, for perceptual SISR and OU NR-IQA. We first design a feature discriminator (Feat-D) to discriminate the semantic features from CLIP, effectively encouraging the feature distributions of SR images to match that of HR images. Then we propose a text-guided discrimination (TG-D) method by training the learnable prompt pairs (LPP) in an adversarial manner to further enhance the discriminative ability. Our SFD facilitates the SR network to generate more fine-grained and more realistic semantic textures, achieving better perception-distortion trade-off. Furthermore, by effectively integrating the well trained Feat-D and LPP, we significantly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks validate the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09, the TCL Science and Technology Innovation Fund under grant 25JZH008, and the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28183" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference (BMVC), 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[4] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.485, + 0.413 + ], + "angle": 0, + "content": "[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[6] Du Chen, Zhengqiang Zhang, Jie Liang, and Lei Zhang. Ssl: A self-similarity loss for improving generative image superresolution. In ACM Multimedia 2024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[8] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[9] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25974-25984, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[10] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[11] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "conference on computer vision and pattern recognition, pages 1103-1112, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[14] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[16] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.907, + 0.404 + ], + "angle": 0, + "content": "[18] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[19] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.907, + 0.574 + ], + "angle": 0, + "content": "[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[22] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[23] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.907, + 0.772 + ], + "angle": 0, + "content": "[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[25] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[27] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28184" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.486, + 0.26 + ], + "angle": 0, + "content": "[29] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.485, + 0.331 + ], + "angle": 0, + "content": "[30] Bingchen Li, Xin Li, Hanxin Zhu, Yeying Jin, Ruoyu Feng, Zhizheng Zhang, and Zhibo Chen. Sed: Semantic-aware discriminator for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25784-25795, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.485, + 0.402 + ], + "angle": 0, + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.485, + 0.473 + ], + "angle": 0, + "content": "[32] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.485, + 0.53 + ], + "angle": 0, + "content": "[33] Xiangyu Liao, Tianheng Zheng, Jiayu Zhong, Pingping Zhang, and Chao Ren. Asymmetric mask scheme for self-supervised real image denoising. arXiv preprint arXiv:2407.06514, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.485, + 0.601 + ], + "angle": 0, + "content": "[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.485, + 0.671 + ], + "angle": 0, + "content": "[35] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.485, + 0.729 + ], + "angle": 0, + "content": "[36] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.485, + 0.8 + ], + "angle": 0, + "content": "[37] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.485, + 0.87 + ], + "angle": 0, + "content": "[38] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[39] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13243-13253, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[40] Chenxi Ma. Uncertainty-aware gan for single image super resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4071-4079, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[41] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 158:1-16, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.908, + 0.3 + ], + "angle": 0, + "content": "[42] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.908, + 0.37 + ], + "angle": 0, + "content": "[43] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.907, + 0.426 + ], + "angle": 0, + "content": "[44] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.908, + 0.495 + ], + "angle": 0, + "content": "[45] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.907, + 0.552 + ], + "angle": 0, + "content": "[46] Zhangkai Ni, Yue Liu, Keyan Ding, Wenhan Yang, Hanli Wang, and Shiqi Wang. Opinion-unaware blind image quality assessment using multi-scale deep feature statistics. IEEE Transactions on Multimedia, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.908, + 0.609 + ], + "angle": 0, + "content": "[47] JoonKyu Park, Sanghyun Son, and Kyoung Mu Lee. Content-aware local gan for photo-realistic super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10585-10594, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.908, + 0.678 + ], + "angle": 0, + "content": "[48] Seung Ho Park, Young Su Moon, and Nam Ik Cho. Perception-oriented single image super-resolution using optimal objective estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1725-1735, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.908, + 0.761 + ], + "angle": 0, + "content": "[49] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.908, + 0.817 + ], + "angle": 0, + "content": "[50] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28185" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[52] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[53] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8207-8216, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[55] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[56] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[57] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[58] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[60] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[64] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF inter" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "national conference on computer vision, pages 1905-1914, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[68] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision - ECCV 2020, pages 101-117. Springer International Publishing, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[69] Qingbo Wu, Zhou Wang, and Hongliang Li. A highly efficient method for blind image quality assessment. In 2015 IEEE International Conference on Image Processing (ICIP), pages 339-343. IEEE, 2015. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[70] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25456-25467, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.907, + 0.504 + ], + "angle": 0, + "content": "[71] Wufeng Xue, Lei Zhang, and Xuanqin Mou. Learning without human scores for blind image quality assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 995-1002, 2013. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.573 + ], + "angle": 0, + "content": "[72] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5791-5800, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[73] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.907, + 0.715 + ], + "angle": 0, + "content": "[74] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.907, + 0.785 + ], + "angle": 0, + "content": "[75] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[76] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[77] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28186" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[79] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Super resolution generative adversarial networks with learning to rank. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7149-7166, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[80] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image super-resolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[81] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[82] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[83] Yunan Zhu, Haichuan Ma, Jialun Peng, Dong Liu, and Zhiwei Xiong. Recycling discriminator: Towards opinion-unaware image quality assessment using Wasserstein gan. In Proceedings of the 29th ACM International Conference on Multimedia, pages 116-125, 2021. 2, 3, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28187" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_origin.pdf b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..69dd5eb2da78488081a02f715c941c88db40d874 --- /dev/null +++ b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/cde8052d-f8eb-472e-b000-fd4d976b6a20_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4aed4be51fd48f799dda97401fa85225dd01611e1cda0461494f20f97eaa799c +size 8293418 diff --git a/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/full.md b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1aa767daed23df2e04a5466a6bbcd17b0d5b8183 --- /dev/null +++ b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/full.md @@ -0,0 +1,354 @@ +# Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment + +Guanglu Dong Xiangyu Liao Mingyang Li Guihuan Guo Chao Ren\* + +College of Electronics and Information Engineering, Sichuan University, China + +{dongguanglu, liaoxiangyul, 2024222055110, guoguihuan}@stu.scu.edu.cn, chaoren@scu.edu.cn + +# Abstract + +Generative Adversarial Networks (GANs) have been widely applied to image super-resolution (SR) to enhance the perceptual quality. However, most existing GAN-based SR methods typically perform coarse-grained discrimination directly on images and ignore the semantic information of images, making it challenging for the super resolution networks (SRN) to learn fine-grained and semantic-related texture details. To alleviate this issue, we propose a semantic feature discrimination method, SFD, for perceptual SR. Specifically, we first design a feature discriminator (FeatD), to discriminate the pixel-wise middle semantic features from CLIP, aligning the feature distributions of SR images with that of high-quality images. Additionally, we propose a text-guided discrimination method (TG-D) by introducing learnable prompt pairs (LPP) in an adversarial manner to perform discrimination on the more abstract output feature of CLIP, further enhancing the discriminative ability of our method. With both Feat-D and TG-D, our SFD can effectively distinguish between the semantic feature distributions of low-quality and high-quality images, encouraging SRN to generate more realistic and semantic-relevant textures. Furthermore, based on the trained Feat-D and LPP, we propose a novel opinion-unaware no-reference image quality assessment (OU NR-IQA) method, SFD-IQA, greatly improving OU NR-IQA performance without any additional targeted training. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our proposed methods. Code is available at https://github.com/GuangluDong0728/SFD. + +# 1. Introduction + +As one of the most important tasks in low-level computer vision [14, 19, 33, 36, 50, 52, 60, 67], single image super-resolution (SISR) [11, 18, 28, 34, 56, 57, 62, 80] aims to accurately reconstruct high-resolution (HR) images from the + +![](images/396d96cff39202a8db442fad8ae6f06fc0e3a04d68748b4918e30ea48a121052.jpg) +Figure 1. Perception-distortion trade-off comparison between our proposed SFD and other SOTA GAN-based SISR methods. + +![](images/9f08aef6845d08f676d8eb3de9d65efebb56b8000129b5db6bd79e1078d064fd.jpg) +Figure 2. Relative PLCC comparison between the proposed SFD-IQA and other OU NR-IQA methods. + +given low-resolution (LR) counterparts. In recent years, with advancements in deep learning, SISR has achieved rapid and substantial progress. Since Dong et al. [11] first introduce deep learning to SISR task, numerous studies have emerged to explore more advanced network architectures, including + +![](images/fb444b8f05e7e2e4847acef8eeb69dd4ea8581952263557cc70c5879bcb428b9.jpg) +(a) Vanilla GAN-based SISR methods + +![](images/7816e3f3230f2168f7d501b4f8569de0efe99dbd8fa7e05306790716d9ac8f40.jpg) +(b) Our proposed SFD which performs discrimination on the semantic features +Figure 3. (a) Vanilla GAN-based SISR methods: performing discrimination on images; (b) Our proposed SFD: performing discrimination on the semantic features. Vanilla GAN-based SISR typically uses an image-wise, patch-wise, or pixel-wise discriminator to determine if the distribution of SR images align with that of high-quality real-world images, ignoring the semantic information of images. Such coarse-grained discrimination makes it challenging for the SR network to reconstruct fine-grained, semantic-relevant textures. In contrast, we propose to perform discrimination on the semantic features, encouraging the SR network to generate more realistic semantic textures. + +CNN-based models [28, 34, 56, 57, 81, 82] and Transformer-based models [7, 13, 31, 38, 72, 80], achieving continuous improvements in fidelity-oriented performance. However, these works primarily employ pixel-wise L1 or MSE losses for the training process, which often produce overly smooth results and fail to generate realistic texture details [4, 29]. + +To overcome the above limitations, a series of generative adversarial networks (GANs) [17] based SISR models [6, 15, 29, 30, 32, 40, 47, 48, 63, 65] have been proposed to generate more realistic images that align more closely with human visual perception by effectively balancing perceptual quality and fidelity. As shown in Figure 3 (a), most existing GAN-based methods perform discrimination on the images using an image-wise [54], patch-wise [23], or pixel-wise [53] discriminator to distinguish between the distributions of SR images and HR images. However, directly discriminating on the images is coarse-grained and lacks semantic awareness. This leads to that the textures of generated images do not match its semantic and often come with artificial artifacts, making it challenging to produce realistic semantic textures. + +In this paper, different from existing works, we propose to perform discrimination on the semantic features, aiming to align the semantic feature distributions of SR images more closely with that of HR images. Considering that CLIP [51] can extract more interpretable, semantic-aware and quality-relevant features [9, 20, 30, 61], we select CLIP's image encoder as the semantic feature extractor and propose two methods to perform joint discrimination on the semantic features. Specifically, as shown in Figure 3 (b), as for the pixel-wise middle semantic features, we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR images and HR images, encouraging the semantic feature distributions of SR images to match that of HR images. As for the final output feature which is more abstract and global, we propose a text-guided discrimination + +method, TG-D, by introducing learnable prompt pairs (LPP) into an adversarial training process to make them sensitive to global image quality and thus able to discriminate the abstract semantic features, further enhancing our method's overall discriminative capability. With both Feat-D and TG-D, the SR network is able to learn more fine-grained semantic feature distributions, ultimately producing more realistic and semantic-related texture details. + +Additionally, existing researches [39, 40, 83] have shown that during the training process where the discriminators in GAN-based SISR learn to distinguish the realness of image distributions, they simultaneously and implicitly learn to assess image quality, making them highly applicable to opinion-unaware no-reference image quality assessment (OU NR-IQA). Simultaneously, recent studies [20, 44] have also shown the advantage of CLIP for OU NR-IQA. In this paper, to further improve OU NR-IQA performance, we propose a novel OU NR-IQA method, SFD-IQA, by integrating the well trained Feat-D and TG-D which conveniently combines both of the above advantages. Thanks to the superior semantic feature discrimination capability of Feat-D and TG-D and the inherent image quality correlation of CLIP, we greatly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets. + +As shown in Fig. 1, our SFD demonstrates a better perception-distortion (PD) trade-off on the PSNR-LPIPS-DISTS plane. Compared to other state-of-the-art (SOTA) GAN-based SISR methods, our LPIPS and PSNR scores are the best among all methods with competitive DISTS scores. Additionally, Fig. 2 illustrates a comparison between our SFD-IQA and other SOTA OU NR-IQA methods, where SFD-IQA outperforms other methods across both SR IQA datasets and authentically distorted IQA datasets. The main contributions of this work are summarized below: + +- We propose a semantic feature discrimination (SFD) + +method for perceptual SISR, which consists of a feature discriminator (Feat-D) based discrimination and a text-guided discrimination (TG-D). Our SFD can effectively distinguish between the CLIP semantic feature distributions of SR and HR images, encouraging SR networks to learn more fine-grained and semantic-related textures, and achieving better perception-distortion (PD) trade-off. + +- We propose a novel OU NR-IQA method, SFD-IQA, by effectively integrating the two well trained feature discrimination methods. Without any additional task-specific training on IQA datasets, SFD-IQA can significantly enhance the OU NR-IQA performance for both SR images and other authentically distorted images. +- Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our method. SFD can help SR networks achieve better PD trade-off and generate more realistic semantic textures, and SFD-IQA is able to greatly improve OU NR-IQA performance. Moreover, our approach can be easily integrated into other GAN-based image restoration methods. + +# 2. Related Works + +# 2.1. Fidelity-Oriented SISR + +Recently, deep learning based single image super-resolution (SISR) [11, 18, 28, 29, 62, 70, 81, 82] has made tremendous advancements. Since SRCNN [11] first introduces CNN into SISR, numerous SISR backbones [18, 28, 34, 56, 57, 81, 82] have continued to emerge to enhance the performance of fidelity-oriented SISR. Specifically, RCAN [81] introduces channel attention into SISR network, and RDN [82] leverages dense connection to fully utilize hierarchical features for SISR. Furthermore, based on the success of the vision transformer [12] framework, [13, 31, 72, 80] further improve the SISR performance. However, in order to maximize fidelity, most of the aforementioned methods only employ pixel-wise loss functions to optimize the training process and ignore the perceptual quality, which often leads to overly smooth results and fails to achieve visually appealing results. + +# 2.2. Perception-Oriented GAN-based SISR + +To improve perceptual quality, GAN-based SISR [15, 27, 30, 32, 39, 47, 64, 65, 75, 79] has received constant attention in recent years. Specifically, SRGAN [29] introduces GANs into SR tasks to obtain SR images more in line with human perception. DualFormer [39] applies an improved spectral discriminator to improve perceptual quality. Sed [30] incorporates the semantic features into the vanilla discriminator to finely distinguish the image distributions. EA-ADAM [55] develops a new optimizer from a multi-objective optimization perspective to achieve a better perception-distortion tradeoff. However, due to the lack of semantic-relevant perception, most GAN-based SR networks are difficult to + +produce fine-grained semantic textures, result in exhibiting artificial artifacts. In this paper, we propose to discriminate the semantic features of CLIP to guide the SR network to learn more fine-grained semantic feature distributions and generate more realistic semantic textures. + +# 2.3. Opinion-Unaware NR-IQA + +OU NR-IQA [2, 43, 46, 61, 69, 71, 77, 83] aims to evaluate the quality of images without human-labeled data. Traditional OU NR-IQA methods [69, 76, 77] are typically based on natural scene statistics (NSS), which estimate distribution parameters to assess quality. Researches on deep learning-based OU NR-IQA are relatively sparse, with most methods centered around learning-to-rank [35, 43] and contrastive learning [44]. Recently, with the emergence of the pretrained vision-language model CLIP [51], [20, 61] leverage the rich visual-linguistic priors encapsulated in CLIP to evaluate image perceptual quality. Additionally, RecycleD [83] proposes to recycle the trained discriminator for GAN-based SISR to predict perceptual quality. Following [83], [39] and [40] further enhance the IQA ability of the discriminators in GAN-based SISR. However, these discriminator-based methods only rely on the trained semantic-unaware discriminator and ignore the assessment for semantic, which is difficult to achieve satisfactory performance. + +# 3. Method and Analyses + +In this section, we provide a detailed explanation and analysis of our method. Firstly, we describe the overall framework of our method in Sec. 3.1. Next, we provide a detailed introduction and analysis of the two proposed semantic feature discrimination methods in Sec. 3.2. Lastly, we discuss how to extend our method to OU NR-IQA in Sec. 3.3. + +# 3.1. Overall Framework + +We propose a novel semantic feature discrimination (SFD) method for perceptual SISR, aiming to achieve better perception-distortion (PD) trade-off. Differing from vanilla GAN-based SR methods that only consider coarse-grained image distributions with poor semantic relevance, our SFD focuses on the distributions of semantic features. As illustrated in Fig. 4, given the input low-resolution image $I_{lr}$ , we can obtain its corresponding SR image $I_{sr}$ through the SRN. Then, through the pre-trained CLIP [51] image encoder $E_{i}$ we can obtain rich semantic features of $I_{sr}$ with different scales and characteristics, including the multi-scale pixelwise middle features $F_{smi}$ , $F_{smi}$ , $F_{smi} \in \mathbb{R}^{C_i \times H_i \times W_i}$ ( $i = 1,2,3$ ), and the more abstract final output feature $F_{sout} \in \mathbb{R}^{1 \times L}$ . To better leverage these semantic features to improve the robustness of our approach, we propose two methods to separately perform discrimination on them. + +On one hand, as for the pixel-wise middle features, we design a feature discriminator (Feat-D) to distinguish be + +![](images/92675a7137dc0bfde740b3776c091ecc7bb348491a2b0a1ae82a8a776f033263.jpg) +Figure 4. Framework of the proposed SFD. SFD consists of a feature discriminator (Feat-D) and a text-guided discrimination (TG-D) method, Feat-D and TG-D are used to perform discrimination on the middle features and final output features from CLIP, respectively. + +tween the feature distributions of SR and HR images. On the other hand, we propose a text-guided discrimination (TG-D) method to perform discrimination on $F_{\text{out}}$ , further enhancing the overall discriminative ability. With both Feat-D and TG-D, SFD is able to facilitate the SR network (SRN) to learn more fine-grained semantic feature distributions and generate more realistic and semantic-related textures. + +# 3.2. Discrimination on Semantic Features + +Semantic Feature Discriminator. As for the middle features of CLIP which contain multi-scale pixel-wise semantic information [30], we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR and HR images, encouraging the semantic feature distribution of SR images to match that of HR images. Specifically, as shown in Figure 4, we can obtain the middle features $F_{sm1}$ , $F_{sm2}$ , $F_{sm3}$ of SR image $I_{sr}$ , and $F_{hm1}$ , $F_{hm2}$ , $F_{hm3}$ of HR image $I_{hr}$ through the freezing CLIP image encoder $E_i$ . Then, the Feat-D is used to perform discrimination on those pixel-wise semantic features. During the adversarial training process, the Feat-D and the SRN will be optimized respectively using the following adversarial loss functions: + +$$ +\begin{array}{l} L _ {D _ {f e a}} = \mathbb {E} _ {F _ {h r} \sim P _ {F _ {h r}}} \log \left(1 - D \left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\right)\right) \tag {1} \\ + \mathbb {E} _ {F _ {s r} \sim P _ {F _ {s r}}} \log (D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\ \end{array} +$$ + +$$ +\begin{array}{l} L _ {G _ {f e a}} = \mathbb {E} _ {F _ {h r} \sim P _ {F _ {h r}}} \log \left(D \left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\right)\right) \tag {2} \\ + \mathbb {E} _ {F _ {s r} \sim P _ {F _ {s r}}} \log (1 - D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\ \end{array} +$$ + +where $P_{F_{hr}}$ and $P_{F_{sr}}$ represent the feature distributions of HR and SR images, respectively. Benefiting from the excellent discriminative ability of Feat-D, the SRN can learn more + +fine-grained semantic feature distributions aligned with HR images, generating more semantically relevant textures. + +Learnable Prompt Pairs Driven Text-Guided Discrimination. Discriminating the final output feature $F_{\text{sout}}$ which is more global and abstract is expected to further enhance the overall performance of our method. However, differing from the middle features, $F_{\text{sout}}$ lacks the pixel-wise semantic information due to its more abstract nature, making it less suitable for Feat-D's pixel-wise discrimination process. Therefore, leveraging the inherent visual-language connection of CLIP, we propose a text-guided discrimination method, TG-D, to perform discrimination on $F_{\text{sout}}$ . Specifically, inspired by [20, 61], we introduce learnable prompt pairs (LPP) into our method. As shown in Fig. 4, given the positive prompt and negative prompt, we can obtain the text features $F_{t+}$ and $F_{t-}$ through the CLIP text encoder $E_t$ . Additionally, the final output features of SR image $F_{\text{sout}}$ and HR image $F_{\text{hout}}$ can be got from $E_i$ . Then, we adopt the same computation method as in [61] to calculate the relative cosine similarity score between $F_{\text{sout}}$ and the text features. This takes two steps: firstly, the cosine similarities $s_+$ between $F_{\text{sout}}$ and $F_{t+}$ , $s_-$ between $F_{\text{sout}}$ and $f_{t-}$ can be respectively calculated as following: + +$$ +s _ {+} = \frac {F _ {s o u t} \odot F _ {t +}}{| | F _ {s o u t} | | \cdot | | F _ {t +} | |}, s _ {-} = \frac {F _ {s o u t} \odot F _ {t -}}{| | F _ {s o u t} | | \cdot | | F _ {t -} | |}, \tag {3} +$$ + +where $\odot$ denotes vector dot-product and $||\cdot ||$ represents $L_{2}$ norm. Then, Softmax is used to calculate the final score: + +$$ +s _ {s r} = \frac {e ^ {s +}}{e ^ {s +} + e ^ {s -}}. \tag {4} +$$ + +As for $F_{hout}$ of HR image, we can also adopt Eq. (3) and Eq. (4) to calculate the cosine similarity score $s_{hr}$ . Then, we + +![](images/183166f0e50134e223cd3019886b2a8c423a68e909a462e087c135b20af8fd42.jpg) +Figure 5. Framework of the proposed SFD-IQA. SFD-IQA is directly based on the well trained Feat-D and LPP, and it doesn't require any additional task-specific training on IQA datasets. + +utilize $s_{sr}$ and $s_{hr}$ to train the LPP and SRN in an adversarial manner: + +$$ +L _ {D _ {t g}} = \mathbb {E} _ {F _ {h r} \sim P _ {F _ {h r}}} \log \left(1 - s _ {h r}\right) + \mathbb {E} _ {F _ {s r} \sim P _ {F _ {s r}}} \log \left(s _ {s r}\right), \tag {5} +$$ + +$$ +L _ {G _ {t g}} = \mathbb {E} _ {F _ {h r} \sim P _ {F _ {h r}}} \log \left(s _ {h r}\right) + \mathbb {E} _ {F _ {s r} \sim P _ {F _ {s r}}} \log \left(1 - s _ {s r}\right), \tag {6} +$$ + +where $L_{D_{tg}}$ and $L_{G_{tg}}$ represent the adversarial loss functions for LPP and SRN, respectively. During the adversarial training process, LPP will be more and more sensitive to image quality and TG-D will be able to discriminate $F_{sout}$ in the form of assigning a global score. Finally, the SRN is optimized by the combination of the following losses as: + +$$ +\begin{array}{l} L _ {t o t a l} = \lambda_ {1} \left\| I _ {h r}, I _ {s r} \right\| _ {1} + \lambda_ {2} \left\| \phi_ {i} \left(I _ {h r}\right), \phi_ {i} \left(I _ {s r}\right) \right\| _ {1} \tag {7} \\ + \lambda_ {3} L _ {G _ {f e a}} + \lambda_ {4} L _ {G _ {t g}}, \\ \end{array} +$$ + +where $||\cdot ||_1$ represents the $L_{1}$ norm, $\phi_i(\cdot)$ is the layers of pre-trained VGG-19 network[54], $\lambda_1,\lambda_2,\lambda_3,\lambda_4$ represent the hyperparameters for the above losses. + +# 3.3. Extension to Opinion-Unaware NR-IQA + +During the process where the Feat-D and TG-D learns to distinguish the distributions of semantic features in an adversarial manner, they simultaneously and implicitly learn to assess image quality. By integrating the well trained Feat-D and LPP, we design a novel OU NR-IQA method, SFD-IQA. As shown in Fig. 5, given an image $x$ , we can obtain its middle semantic features $F_{m1}$ , $F_{m2}$ , $F_{m3}$ , and the final output feature $F_{out}$ through the CLIP image encoder $E_{i}$ . Then, Feat-D will convert $F_{m1}$ , $F_{m2}$ , $F_{m3}$ into a score matrix: + +$$ +\left[ s _ {i j} \right] _ {h \times w} = D \left(F _ {m 1}, F _ {m 2}, F _ {m 3}\right), \tag {8} +$$ + +where $[s_{ij}]_{h\times w}$ denotes the score matrix. Additionally, according to Eq. 2 and Eq. 3, we adopt $F_{out}$ and LPP to calculate the similarity score $s_{lp}$ . Due to CLIP's outstanding contrastive learning training scheme, $F_{out}$ is already sensitive to quality-related text prompts. To further enhance the robustness of SFD-IQA, we also adopt the original antonymic text prompts used in CLIP-IQA (e.g., "Good photo" and "Bad + +photo") to calculate the similarity score $s_o$ with $F_{out}$ . Then we perform a weighted average on $s_o$ and $s_{lp}$ : + +$$ +s _ {a v e r} = \alpha_ {1} s _ {o} + \alpha_ {2} s _ {l p}, \tag {9} +$$ + +where $\alpha_{1}$ and $\alpha_{2}$ denotes the weight coefficients. Finally, following [83], we adopt the product of $s_{aver}$ and the average of $[s_{ij}]_{h\times w}$ as the final IQA score $s_d$ : + +$$ +s _ {d} = \left(s _ {a v e r}\right) \times \frac {\sum_ {i = 0} ^ {h} \sum_ {j = 0} ^ {w} \left(\operatorname {s i g m o i d} \left(\left[ s _ {i j} \right] _ {h \times w}\right)\right)}{h \times w}, \tag {10} +$$ + +where sigmoid denotes the Sigmoid function to adjust the score range to [0, 1]. Notably, we directly apply the Feat-D and TG-D to the above IQA process without any additional task-specific training. Benefiting from the semantic-related discriminative abilities of Feat-D and TG-D, SFD-IQA can significantly improve the OU NR-IQA performance. + +# 4. Experiments + +# 4.1. Experimental Setup + +Training Datasets. For classical SISR, following previous works [15, 30, 39], we train our network using 3450 images from DF2K [1, 58], and the bicubic downsampling is used to synthesize LR-HR training pairs. For real-world SISR and OU NR-IQA, following [39, 65, 83], we adopt DF2K and OutdoorSceneTraining [63] as the training dataset, and the LR-HR training pairs are generated by the degradation pipeline of Real-ESRGAN [65]. + +Test Datasets and Evaluation Metrics. For classical SISR, we test on 4 commonly used benchmarks: Set5 [3], Set14 [74], Urban100 [22], and DIV2K validation set [1], using 4 metrics to comprehensively evaluate the performance of different methods, including fidelity metrics: PSNR and SSIM [66] (calculated on the Y channel in YCbCr space), and perceptual quality metrics: LPIPS [78], DISTS [10]. For real-world SISR, following [62, 70], we employ 2 centeredropped real-world datasets, RealSR [5] and DRealSR [68]. We also create another dataset, RealD2K, by degrading DIV2K validation set under the same pipeline as that in training. We add another metric NIQE [77] for real-world SISR. For OU NR-IQA, following [39, 83], we use two categories of IQA datasets, including SR IQA datasets: PIPAL [24], Ma's dataset [41], and authentically distorted IQA datasets: KonIQ-10k [21], LIVE-itW [16]. To evaluate the IQA performance, we adopt three widely used metrics: PLCC, SRCC and KRCC [25]. + +Implementation Details. For the generator, following [30, 65], we use the competitive RRDB [64] as the backbone. We implement the experiments on NVIDIA 4090 GPUs with PyTorch [49]. We initialize the parameters with the pre-trained fidelity-oriented model. We use Adam [26] optimizer to train the network with a learning rate of $10^{-4}$ . + +Table 1. Quantitative comparison of our method vs. other SOTA methods for $\times 4\mathrm{{SR}}$ task. The best and the second-best are marked in red and blue, respectively. + +
BenchmarkMetricRankSR GAN [79]SRGAN [29]SFTGAN [63]ESRGAN [64]SPSR [42]LDL [32]DualFormer [39]WGSR [27]SeD [30]RRDB +OursSwinIR +\( {L}_{GAN} \)SwinIR +Ours
Set5PSNR ↑29.6530.5230.0630.4430.4031.0331.3031.3331.2231.6331.1031.78
SSIM ↑0.8380.8630.8480.8520.8440.8610.8690.8660.8680.8770.8690.877
LPIPS ↓0.0700.0690.0800.0740.0690.0660.0660.0650.0640.0630.0680.059
DISTS ↓0.1040.1080.1090.0970.0930.0930.0930.1100.0940.0980.0980.089
Set14PSNR ↑26.4527.0126.7426.2826.6427.1027.3927.4027.3727.7927.0527.94
SSIM ↑0.7030.7280.7180.6990.7140.7350.7390.7390.7360.7520.7310.754
LPIPS ↓0.1350.1230.1310.1310.1350.1310.1200.1380.1170.1170.1200.118
DISTS ↓0.1100.1090.1130.0990.0990.0980.0920.1120.0930.0960.0980.091
Urban100PSNR ↑24.4725.0224.3424.3524.8025.4925.6925.7825.9326.1825.8326.57
SSIM ↑0.7290.7780.7240.7340.7470.7670.7730.7810.7800.7870.7850.798
LPIPS ↓0.1380.1300.1340.1230.1210.1100.1150.1350.1070.1090.1000.099
DISTS ↓0.1040.1100.1060.0880.0860.0820.0850.1080.0790.0820.0810.079
DIV2K100PSNR ↑28.0328.1828.0928.2028.1828.9629.2529.1929.2729.7528.9929.90
SSIM ↑0.7670.7880.7710.7770.7720.7950.8020.8040.8030.8150.8030.818
LPIPS ↓0.1210.1220.1330.1150.1130.1010.0970.1040.0940.0970.0940.099
DISTS ↓0.0640.0630.0740.0590.0550.0530.0560.0540.0540.0530.0500.050
+ +![](images/bfaf4422242b935e6f2854901770c6de0764d3d56474ad9661236473bfd15105.jpg) +Figure 6. Visual comparison of different GAN-based SR methods on $\times 4$ super-resolution. + +# 4.2. Results on Classical SISR + +To validate the effectiveness of our method, we compare it with several SOTA GAN-based SR methods, including [27, 29, 30, 32, 39, 42, 64, 79]. As shown in Tab. 1, compared to other approaches, our method achieves the best PSNR and SSIM scores on all benchmarks, while also holding highly competitive perceptual metrics, meaning that our method can sacrifice less fidelity in exchange for improved perceptual quality. To validate generalizability, we also employ another Transformer-based SR backbone, SwinIR [31]. It's clearly that our method also demonstrates superior performance with SwinIR. Fig. 6 presents the visual comparison, indicating + +that our method generates more realistic semantic textures with fewer artifacts than other methods. + +# 4.3. Results on Real-World SISR + +For real-world SISR, we provide quantitative and visual comparison with representative SOTA real-world SISR models including [31, 32, 65, 75]. The quantitative comparison is demonstrated in Table 2. The results show that our SFD achieves the best SSIM score on all datasets, while also maintaining competitive perceptual metrics, especially LPIPS and DISTS, indicating that our approach can achieve better PD trade-off. As shown in Fig. 7, our method produces SR + +![](images/bc10b71cceb15f3f3e6ad5da0237c259fc148e087f7b59cbc39c5579dd901707.jpg) +LR + +![](images/0e7d99f196ba262691ed54647598466c018e8f995db6e80303b05c9bdda30a4b.jpg) +Bicubic + +![](images/3688d96a7a6252d3e825b160d96cb4eeb6bd719c03da937210c3b5f686e7b8df.jpg) +BSRGAN + +![](images/39a5e3ce40fdf4ab19cb7a6caf6c2245f192737b2a9069ee043c618025edf165.jpg) +ReaESRGAN +Figure 7. Visual comparison of different GAN-based real-world SR methods on $\times 4$ super-resolution. + +![](images/2d510b580f555012f5ddfa30e091f618775918d9959ab2d1142e503f28c1f759.jpg) +LDL + +![](images/6096729d2fd689112840df82219cbc58482ebb9e5c70ea378e264e5fafe0fa2a.jpg) +RealSwinIR + +![](images/a932a7fc76a2ff3336e219fac6d526bb82b12ead11cdafef833b0625668a594b.jpg) +Ours + +Table 2. Quantitative comparison of our method vs. other state-of-the-art real-world SISR methods for $\times 4\mathrm{{SR}}$ task. The best and the second-best are marked in red and blue, respectively. + +
BenchmarkMetricsBSRGAN [75]RealESR GAN [65]LDL [32]RealSwinIR [31]SFD (Ours)
RealD2KPSNR ↑23.5123.5022.9422.6723.40
SSIM ↑0.5850.6040.5970.5810.606
LPIPS ↓0.3790.3680.3730.3680.358
DISTS ↓0.2800.2780.2840.2790.278
NIQE ↓7.5378.3698.3847.8668.600
RealSRPSNR ↑26.3925.6925.2825.9026.12
SSIM ↑0.7650.7620.7570.7680.772
LPIPS ↓0.2670.2730.2750.2620.256
DISTS ↓0.2120.2060.2120.2000.200
NIQE ↓5.6575.8305.9925.9265.797
DRealSRPSNR ↑28.7528.6428.2128.2129.20
SSIM ↑0.8030.8050.8130.7980.815
LPIPS ↓0.2880.2850.2790.2850.284
DISTS ↓0.2140.2090.2130.2100.221
NIQE ↓6.5196.6937.1436.7666.878
+ +images with sharp edges and fine textures, accompanied by fewer undesirable artifacts. The above results suggest that our SFD can distinguish complex degraded real world images and facilitate the SR network to learn more fine-grained and semantically accurate textures. + +# 4.4. Results on OU NR-IQA + +To verify the effectiveness of our SFD-IQA for OU NR-IQA tasks, we compared SFD-IQA with several representative OU NR-IQA methods, including traditional methods [37, 69, 71, 76, 77], deep learning-based methods [2, 35, 39, 43, 46, 61, 83]. The results are presented in Table 3. Thanks to the excellent ability of Feat-D and TG-D to distinguish between low-quality and high-quality images, learned during the real-world SISR training process, our SFD-IQA achieves the best performance across all four datasets, validating its enhanced ability to predict the perceptual quality of images. + +# 4.5. Ablation Studies + +The effects of Feat-D and TG-D on perceptual SISR. To investigate the effectiveness of our Feat-D and TG-D, we conduct ablation experiments on them using the DIV2K100 test set. As shown in Fig. 8, compared to the baseline (ESR-GAN), Feat-D and TG-D can both independently improve the perceptual quality with less decrease in fidelity. Furthermore, when both Feat-D and TG-D are introduced into the + +![](images/6dc3df15b417a06c1b1d63769d21eb1885c8d4bd09404c4640e0c07145f46d51.jpg) +Figure 8. The effects of Feat-D and TG-D on perceptual SISR. + +![](images/9c0230ae1c742bae1c7fbceaad6d6c0494be433b6472580bb7e40449a053dd88.jpg) +t-SNE of vanilla discriminator features +Figure 9. The t-SNE visualization of vanilla discriminator features and our Feat-D features. We select 8 categories of images from OST dataset, including "bear", "building", "cat", "dog", "grass", "mountain", "sky" and "water" and 50 images for each category. + +![](images/9c500fb3b443b488235480078a749d3767fbe6f30231394676d9cc662ac5cddd.jpg) +t-SNE of our Feat-D features + +baseline, we can trade the minimum PSNR decrease for the best perceptual quality. The above ablation results strongly verify the superiority of Feat-D and TG-D. + +The effects of Feat-D and TG-D in SFD-IQA. To investigate the contribution of our Feat-D and TG-D to the proposed SFD-IQA, we conduct ablation studies by respectively remove Feat-D and TG-D from our SFD-IQA. Tab. 4 presents the comparative results on the 3 SR related subtypes of PIPAL [24] training set and 3 sub-types of KonIQ-10k [21] dataset. Compared to the baseline without Feat-D and TG-D, it's clearly that each of Feat-D and TG-D can independently enhance the OU NR-IQA performance to a certain extent. With both Feat-D and TG-D, SFD-IQA significantly improves the performance in most cases. Especially on the GAN-based SR subset of PIPAL, it boosts the PLCC by $86.0\%$ and SRCC by $28.8\%$ . The above results demonstrate the effectiveness of Feat-D and TG-D. + +T-SNE analysis for Feat-D. To further demonstrate the + +Table 3. OU NR-IQA comparison between different methods. The best and the second-best are marked in red and blue, respectively. + +
IQA MethodsPIPAL [24]KonIQ-10k [21]Ma's dataset [41]LIVE-itW [16]
PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑
QAC [71]0.25660.22200.15030.37190.33970.23020.50710.43870.30550.27200.04570.0370
LPSI [69]0.28970.20310.13630.20660.22390.15040.52930.48960.35430.28770.08340.0524
NIQE [77]0.04390.06080.04010.46370.52910.36660.64270.62610.45200.47900.44930.3062
IL-NIQE [76]0.27480.24120.16300.53160.50570.35040.71550.65480.47510.50390.43940.2985
SNP-NIQE [37]0.18200.18010.12140.63400.62850.44350.62500.56210.39840.52010.46570.3162
dipIQ [43]0.28000.23040.15350.44290.23750.15940.51670.48660.36070.31800.17740.1207
RankIAQ [35]0.36980.37150.25220.50280.49830.34480.51780.49130.36480.45280.43070.2945
RecycleD [83]0.33960.33310.22650.64070.63640.44680.59830.55550.38740.48060.46610.3194
DualFormer [39]0.28960.27870.19730.65430.63210.44350.61720.60540.40440.50680.48970.3312
ContentSep [2]---0.62740.64010.4529---0.51300.50600.3450
CLIPIAQ [61]0.28310.27620.18620.43300.41450.28210.71750.70550.50630.58530.56470.4010
MDFS [46]0.26510.22730.15320.71230.73330.53440.76490.71420.52050.53640.48210.3274
SFD-IQA (Ours)0.42520.40920.28160.75740.74400.54430.80120.78700.58640.71110.69840.5090
+ +Table 4. OU NR-IQA performance comparison on sub-types of PIPAL training set and KonIQ-10k dataset. "Baseline" refers to the model without neither Feat-D nor TG-D. "W/ Feat-D" and "w/. TG-D" refer to the models that only include Feat-D or TG-D. + +
MetricsDataset. Sub-TypesBaselinew/.Feat-Dw/.TG-DSFD-IQA (Ours)
PLCCPIPAL.Trad.SR0.37870.36840.39100.3797 (+0.26%)
PIPAL.PSNR.SR0.42900.45040.45470.4657 (+8.55%)
PIPAL.GAN.SR0.21540.39580.24300.4006 (+86.0%)
KonIQ-10k.Train0.70920.74840.74080.7581 (+6.90%)
KonIQ-10k.Val0.70550.74520.73500.7528 (+6.70%)
KonIQ-10k.Test0.68650.74930.74650.7576 (+10.4%)
SRCCPIPAL.Trad.SR0.35450.34200.36600.3522 (-0.64%)
PIPAL.PSNR.SR0.44320.47970.46720.4932 (+11.3%)
PIPAL.GAN.SR0.23760.30480.25980.3061 (+28.8%)
KonIQ-10k.Train0.68490.73250.71300.7441 (+4.92%)
KonIQ-10k.Val0.68030.72640.70720.7373 (+4.51%)
KonIQ-10k.Test0.68620.73590.71310.7475 (+8.89%)
+ +effectiveness of our Feat-D, we use t-SNE [59] to visualize the middle features of the VGG-style vanilla discriminator and our Feat-D. We select 8 categories of images from the OST [63] dataset, and 50 images are randomly selected for each category. We visualize features after the 3rd upsampling layer of our Feat-D and after the 2nd BN layer of the vanilla discriminator, respectively. As shown in Fig. 9, the feature clustering of our Feat-D is much better than that of the vanilla discriminator. This strongly proves that the proposed Feat-D is semantic-aware and can perform more fine-grained discrimination, encouraging the SR network to generate more more realistic semantic textures. + +The effectiveness of our method on other image restoration tasks. We preliminarily explore the effectiveness of our approach in addressing other image restoration tasks, such as image deraining and image deblurring. We select a commonly used image restoration baselines NAFNet [8], and conduct training and testing on the Rain100L [73] and GoPro [45] datasets for the two tasks. As shown in Tab. 5, compared to fidelity-oriented baseline methods and vanilla discriminator based methods, our method significantly improves perceptual metrics while maintaining a high level of fidelity on both two tasks, demonstrating the potential of our method to be integrated into other image + +Table 5. The effectiveness of our method on image deraining and image deblurring. "VD" means vanilla discriminator. + +
MethodRain100L [73]GoPro [45]
PSNR ↑LPIPS ↓DISTS ↓PSNR ↑LPIPS ↓DISTS ↓
NAFNet37.000.0220.03132.870.0930.072
+ VD36.390.0150.02032.010.0680.036
+ Ours36.490.0140.01932.190.0640.034
+ +restoration tasks for better perceptual quality. + +# 5. Conclusion + +In this paper, we propose a semantic feature discrimination method, SFD, for perceptual SISR and OU NR-IQA. We first design a feature discriminator (Feat-D) to discriminate the semantic features from CLIP, effectively encouraging the feature distributions of SR images to match that of HR images. Then we propose a text-guided discrimination (TG-D) method by training the learnable prompt pairs (LPP) in an adversarial manner to further enhance the discriminative ability. Our SFD facilitates the SR network to generate more fine-grained and more realistic semantic textures, achieving better perception-distortion trade-off. Furthermore, by effectively integrating the well trained Feat-D and LPP, we significantly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks validate the effectiveness of our approach. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09, the TCL Science and Technology Innovation Fund under grant 25JZH008, and the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5 +[2] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 3, 7, 8 +[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference (BMVC), 2012. 5 +[4] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2 +[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5 +[6] Du Chen, Zhengqiang Zhang, Jie Liang, and Lei Zhang. Ssl: A self-similarity loss for improving generative image superresolution. In ACM Multimedia 2024, 2024. 2 +[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2 +[8] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 8 +[9] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25974-25984, 2024. 2 +[10] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 5 +[11] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1, 3 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 3 +[13] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In Proceedings of the IEEE/CVF + +conference on computer vision and pattern recognition, pages 1103-1112, 2022. 2, 3 +[14] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1 +[15] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 3, 5 +[16] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5, 8 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[18] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 3 +[19] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1 +[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 3, 4 +[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5, 7, 8 +[22] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5 +[23] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5, 7, 8 +[25] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938. 5 +[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[27] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 3, 6 + +[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 1, 2, 3 +[29] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 3, 6 +[30] Bingchen Li, Xin Li, Hanxin Zhu, Yeying Jin, Ruoyu Feng, Zhizheng Zhang, and Zhibo Chen. Sed: Semantic-aware discriminator for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25784-25795, 2024. 2, 3, 4, 5, 6 +[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 6, 7 +[32] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 2, 3, 6, 7 +[33] Xiangyu Liao, Tianheng Zheng, Jiayu Zhong, Pingping Zhang, and Chao Ren. Asymmetric mask scheme for self-supervised real image denoising. arXiv preprint arXiv:2407.06514, 2024. 1 +[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 3 +[35] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 3, 7, 8 +[36] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1 +[37] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2019. 7, 8 +[38] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2 +[39] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality + +improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13243-13253, 2023. 2, 3, 5, 6, 7, 8 +[40] Chenxi Ma. Uncertainty-aware gan for single image super resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4071-4079, 2024. 2, 3 +[41] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 158:1-16, 2017. 5, 8 +[42] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020. 6 +[43] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 3, 7, 8 +[44] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 2, 3 +[45] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 8 +[46] Zhangkai Ni, Yue Liu, Keyan Ding, Wenhan Yang, Hanli Wang, and Shiqi Wang. Opinion-unaware blind image quality assessment using multi-scale deep feature statistics. IEEE Transactions on Multimedia, 2024. 3, 7, 8 +[47] JoonKyu Park, Sanghyun Son, and Kyoung Mu Lee. Content-aware local gan for photo-realistic super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10585-10594, 2023. 2, 3 +[48] Seung Ho Park, Young Su Moon, and Nam Ik Cho. Perception-oriented single image super-resolution using optimal objective estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1725-1735, 2023. 2 +[49] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[50] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3 + +[52] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1 +[53] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8207-8216, 2020. 2 +[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2, 5 +[55] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 3 +[56] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 1, 2, 3 +[57] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 1, 2, 3 +[58] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5 +[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8 +[60] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 1 +[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 4, 7, 8 +[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 5 +[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 2, 5, 6, 8 +[64] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 3, 5, 6 +[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF inter + +national conference on computer vision, pages 1905-1914, 2021. 2, 3, 5, 6, 7 +[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 1 +[68] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision - ECCV 2020, pages 101-117. Springer International Publishing, 2020. 5 +[69] Qingbo Wu, Zhou Wang, and Hongliang Li. A highly efficient method for blind image quality assessment. In 2015 IEEE International Conference on Image Processing (ICIP), pages 339-343. IEEE, 2015. 3, 7, 8 +[70] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25456-25467, 2024. 3, 5 +[71] Wufeng Xue, Lei Zhang, and Xuanqin Mou. Learning without human scores for blind image quality assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 995-1002, 2013. 3, 7, 8 +[72] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5791-5800, 2020. 2, 3 +[73] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 8 +[74] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5 +[75] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 3, 6, 7 +[76] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 7, 8 +[77] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 5, 7, 8 +[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep + +features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[79] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Super resolution generative adversarial networks with learning to rank. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7149-7166, 2021. 3, 6 +[80] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image super-resolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1, 2, 3 +[81] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2, 3 +[82] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3 +[83] Yunan Zhu, Haichuan Ma, Jialun Peng, Dong Liu, and Zhiwei Xiong. Recycling discriminator: Towards opinion-unaware image quality assessment using Wasserstein gan. In Proceedings of the 29th ACM International Conference on Multimedia, pages 116-125, 2021. 2, 3, 5, 7, 8 \ No newline at end of file diff --git a/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/images.zip b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0f2e65258baf0b3a40cec6047253659b000eac1c --- /dev/null +++ b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc38f5a8907338c855a56f73a98c67008927b52011a1b6413d6a40992dbe0250 +size 938800 diff --git a/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/layout.json b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ffc56d4bb11d91a3092239ffb796b8254395ea7 --- /dev/null +++ b/2025/Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment/layout.json @@ -0,0 +1,9818 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "text", + "content": "Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 161, + 521, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 161, + 521, + 176 + ], + "spans": [ + { + "bbox": [ + 93, + 161, + 521, + 176 + ], + "type": "text", + "content": "Guanglu Dong Xiangyu Liao Mingyang Li Guihuan Guo Chao Ren\\*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 176, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 176, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 176, + 496, + 190 + ], + "type": "text", + "content": "College of Electronics and Information Engineering, Sichuan University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 191, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 545, + 203 + ], + "type": "text", + "content": "{dongguanglu, liaoxiangyul, 2024222055110, guoguihuan}@stu.scu.edu.cn, chaoren@scu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 255, + 297, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 604 + ], + "type": "text", + "content": "Generative Adversarial Networks (GANs) have been widely applied to image super-resolution (SR) to enhance the perceptual quality. However, most existing GAN-based SR methods typically perform coarse-grained discrimination directly on images and ignore the semantic information of images, making it challenging for the super resolution networks (SRN) to learn fine-grained and semantic-related texture details. To alleviate this issue, we propose a semantic feature discrimination method, SFD, for perceptual SR. Specifically, we first design a feature discriminator (FeatD), to discriminate the pixel-wise middle semantic features from CLIP, aligning the feature distributions of SR images with that of high-quality images. Additionally, we propose a text-guided discrimination method (TG-D) by introducing learnable prompt pairs (LPP) in an adversarial manner to perform discrimination on the more abstract output feature of CLIP, further enhancing the discriminative ability of our method. With both Feat-D and TG-D, our SFD can effectively distinguish between the semantic feature distributions of low-quality and high-quality images, encouraging SRN to generate more realistic and semantic-relevant textures. Furthermore, based on the trained Feat-D and LPP, we propose a novel opinion-unaware no-reference image quality assessment (OU NR-IQA) method, SFD-IQA, greatly improving OU NR-IQA performance without any additional targeted training. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our proposed methods. Code is available at https://github.com/GuangluDong0728/SFD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "As one of the most important tasks in low-level computer vision [14, 19, 33, 36, 50, 52, 60, 67], single image super-resolution (SISR) [11, 18, 28, 34, 56, 57, 62, 80] aims to accurately reconstruct high-resolution (HR) images from the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 314, + 229, + 555, + 390 + ], + "blocks": [ + { + "bbox": [ + 314, + 229, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 314, + 229, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 314, + 229, + 555, + 390 + ], + "type": "image", + "image_path": "396d96cff39202a8db442fad8ae6f06fc0e3a04d68748b4918e30ea48a121052.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "type": "text", + "content": "Figure 1. Perception-distortion trade-off comparison between our proposed SFD and other SOTA GAN-based SISR methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 430, + 555, + 603 + ], + "blocks": [ + { + "bbox": [ + 319, + 430, + 555, + 603 + ], + "lines": [ + { + "bbox": [ + 319, + 430, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 319, + 430, + 555, + 603 + ], + "type": "image", + "image_path": "9f08aef6845d08f676d8eb3de9d65efebb56b8000129b5db6bd79e1078d064fd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 605, + 556, + 628 + ], + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 628 + ], + "type": "text", + "content": "Figure 2. Relative PLCC comparison between the proposed SFD-IQA and other OU NR-IQA methods." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "given low-resolution (LR) counterparts. In recent years, with advancements in deep learning, SISR has achieved rapid and substantial progress. Since Dong et al. [11] first introduce deep learning to SISR task, numerous studies have emerged to explore more advanced network architectures, including" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28176" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 246, + 196 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 246, + 196 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 246, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 246, + 196 + ], + "type": "image", + "image_path": "fb444b8f05e7e2e4847acef8eeb69dd4ea8581952263557cc70c5879bcb428b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 198, + 214, + 207 + ], + "lines": [ + { + "bbox": [ + 93, + 198, + 214, + 207 + ], + "spans": [ + { + "bbox": [ + 93, + 198, + 214, + 207 + ], + "type": "text", + "content": "(a) Vanilla GAN-based SISR methods" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 250, + 70, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 250, + 70, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 250, + 70, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 250, + 70, + 553, + 196 + ], + "type": "image", + "image_path": "7816e3f3230f2168f7d501b4f8569de0efe99dbd8fa7e05306790716d9ac8f40.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 198, + 527, + 208 + ], + "lines": [ + { + "bbox": [ + 276, + 198, + 527, + 208 + ], + "spans": [ + { + "bbox": [ + 276, + 198, + 527, + 208 + ], + "type": "text", + "content": "(b) Our proposed SFD which performs discrimination on the semantic features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 209, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 54, + 209, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 555, + 265 + ], + "type": "text", + "content": "Figure 3. (a) Vanilla GAN-based SISR methods: performing discrimination on images; (b) Our proposed SFD: performing discrimination on the semantic features. Vanilla GAN-based SISR typically uses an image-wise, patch-wise, or pixel-wise discriminator to determine if the distribution of SR images align with that of high-quality real-world images, ignoring the semantic information of images. Such coarse-grained discrimination makes it challenging for the SR network to reconstruct fine-grained, semantic-relevant textures. In contrast, we propose to perform discrimination on the semantic features, encouraging the SR network to generate more realistic semantic textures." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 285, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 295, + 357 + ], + "type": "text", + "content": "CNN-based models [28, 34, 56, 57, 81, 82] and Transformer-based models [7, 13, 31, 38, 72, 80], achieving continuous improvements in fidelity-oriented performance. However, these works primarily employ pixel-wise L1 or MSE losses for the training process, which often produce overly smooth results and fail to generate realistic texture details [4, 29]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 362, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 362, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 362, + 295, + 529 + ], + "type": "text", + "content": "To overcome the above limitations, a series of generative adversarial networks (GANs) [17] based SISR models [6, 15, 29, 30, 32, 40, 47, 48, 63, 65] have been proposed to generate more realistic images that align more closely with human visual perception by effectively balancing perceptual quality and fidelity. As shown in Figure 3 (a), most existing GAN-based methods perform discrimination on the images using an image-wise [54], patch-wise [23], or pixel-wise [53] discriminator to distinguish between the distributions of SR images and HR images. However, directly discriminating on the images is coarse-grained and lacks semantic awareness. This leads to that the textures of generated images do not match its semantic and often come with artificial artifacts, making it challenging to produce realistic semantic textures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "type": "text", + "content": "In this paper, different from existing works, we propose to perform discrimination on the semantic features, aiming to align the semantic feature distributions of SR images more closely with that of HR images. Considering that CLIP [51] can extract more interpretable, semantic-aware and quality-relevant features [9, 20, 30, 61], we select CLIP's image encoder as the semantic feature extractor and propose two methods to perform joint discrimination on the semantic features. Specifically, as shown in Figure 3 (b), as for the pixel-wise middle semantic features, we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR images and HR images, encouraging the semantic feature distributions of SR images to match that of HR images. As for the final output feature which is more abstract and global, we propose a text-guided discrimination" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "type": "text", + "content": "method, TG-D, by introducing learnable prompt pairs (LPP) into an adversarial training process to make them sensitive to global image quality and thus able to discriminate the abstract semantic features, further enhancing our method's overall discriminative capability. With both Feat-D and TG-D, the SR network is able to learn more fine-grained semantic feature distributions, ultimately producing more realistic and semantic-related texture details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 556, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 576 + ], + "type": "text", + "content": "Additionally, existing researches [39, 40, 83] have shown that during the training process where the discriminators in GAN-based SISR learn to distinguish the realness of image distributions, they simultaneously and implicitly learn to assess image quality, making them highly applicable to opinion-unaware no-reference image quality assessment (OU NR-IQA). Simultaneously, recent studies [20, 44] have also shown the advantage of CLIP for OU NR-IQA. In this paper, to further improve OU NR-IQA performance, we propose a novel OU NR-IQA method, SFD-IQA, by integrating the well trained Feat-D and TG-D which conveniently combines both of the above advantages. Thanks to the superior semantic feature discrimination capability of Feat-D and TG-D and the inherent image quality correlation of CLIP, we greatly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 578, + 556, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 556, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 556, + 697 + ], + "type": "text", + "content": "As shown in Fig. 1, our SFD demonstrates a better perception-distortion (PD) trade-off on the PSNR-LPIPS-DISTS plane. Compared to other state-of-the-art (SOTA) GAN-based SISR methods, our LPIPS and PSNR scores are the best among all methods with competitive DISTS scores. Additionally, Fig. 2 illustrates a comparison between our SFD-IQA and other SOTA OU NR-IQA methods, where SFD-IQA outperforms other methods across both SR IQA datasets and authentically distorted IQA datasets. The main contributions of this work are summarized below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "- We propose a semantic feature discrimination (SFD)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28177" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 156 + ], + "type": "text", + "content": "method for perceptual SISR, which consists of a feature discriminator (Feat-D) based discrimination and a text-guided discrimination (TG-D). Our SFD can effectively distinguish between the CLIP semantic feature distributions of SR and HR images, encouraging SR networks to learn more fine-grained and semantic-related textures, and achieving better perception-distortion (PD) trade-off." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 295, + 312 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 156, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 227 + ], + "type": "text", + "content": "- We propose a novel OU NR-IQA method, SFD-IQA, by effectively integrating the two well trained feature discrimination methods. Without any additional task-specific training on IQA datasets, SFD-IQA can significantly enhance the OU NR-IQA performance for both SR images and other authentically distorted images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 228, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 312 + ], + "type": "text", + "content": "- Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks demonstrate the effectiveness of our method. SFD can help SR networks achieve better PD trade-off and generate more realistic semantic textures, and SFD-IQA is able to greatly improve OU NR-IQA performance. Moreover, our approach can be easily integrated into other GAN-based image restoration methods." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 324, + 147, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 147, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 147, + 337 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 344, + 187, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 187, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 187, + 357 + ], + "type": "text", + "content": "2.1. Fidelity-Oriented SISR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 363, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 542 + ], + "type": "text", + "content": "Recently, deep learning based single image super-resolution (SISR) [11, 18, 28, 29, 62, 70, 81, 82] has made tremendous advancements. Since SRCNN [11] first introduces CNN into SISR, numerous SISR backbones [18, 28, 34, 56, 57, 81, 82] have continued to emerge to enhance the performance of fidelity-oriented SISR. Specifically, RCAN [81] introduces channel attention into SISR network, and RDN [82] leverages dense connection to fully utilize hierarchical features for SISR. Furthermore, based on the success of the vision transformer [12] framework, [13, 31, 72, 80] further improve the SISR performance. However, in order to maximize fidelity, most of the aforementioned methods only employ pixel-wise loss functions to optimize the training process and ignore the perceptual quality, which often leads to overly smooth results and fails to achieve visually appealing results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 259, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 259, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 259, + 565 + ], + "type": "text", + "content": "2.2. Perception-Oriented GAN-based SISR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "To improve perceptual quality, GAN-based SISR [15, 27, 30, 32, 39, 47, 64, 65, 75, 79] has received constant attention in recent years. Specifically, SRGAN [29] introduces GANs into SR tasks to obtain SR images more in line with human perception. DualFormer [39] applies an improved spectral discriminator to improve perceptual quality. Sed [30] incorporates the semantic features into the vanilla discriminator to finely distinguish the image distributions. EA-ADAM [55] develops a new optimizer from a multi-objective optimization perspective to achieve a better perception-distortion tradeoff. However, due to the lack of semantic-relevant perception, most GAN-based SR networks are difficult to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "produce fine-grained semantic textures, result in exhibiting artificial artifacts. In this paper, we propose to discriminate the semantic features of CLIP to guide the SR network to learn more fine-grained semantic feature distributions and generate more realistic semantic textures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 138, + 465, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 465, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 465, + 152 + ], + "type": "text", + "content": "2.3. Opinion-Unaware NR-IQA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 373 + ], + "type": "text", + "content": "OU NR-IQA [2, 43, 46, 61, 69, 71, 77, 83] aims to evaluate the quality of images without human-labeled data. Traditional OU NR-IQA methods [69, 76, 77] are typically based on natural scene statistics (NSS), which estimate distribution parameters to assess quality. Researches on deep learning-based OU NR-IQA are relatively sparse, with most methods centered around learning-to-rank [35, 43] and contrastive learning [44]. Recently, with the emergence of the pretrained vision-language model CLIP [51], [20, 61] leverage the rich visual-linguistic priors encapsulated in CLIP to evaluate image perceptual quality. Additionally, RecycleD [83] proposes to recycle the trained discriminator for GAN-based SISR to predict perceptual quality. Following [83], [39] and [40] further enhance the IQA ability of the discriminators in GAN-based SISR. However, these discriminator-based methods only rely on the trained semantic-unaware discriminator and ignore the assessment for semantic, which is difficult to achieve satisfactory performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 381, + 440, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 440, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 440, + 396 + ], + "type": "text", + "content": "3. Method and Analyses" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 402, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 475 + ], + "type": "text", + "content": "In this section, we provide a detailed explanation and analysis of our method. Firstly, we describe the overall framework of our method in Sec. 3.1. Next, we provide a detailed introduction and analysis of the two proposed semantic feature discrimination methods in Sec. 3.2. Lastly, we discuss how to extend our method to OU NR-IQA in Sec. 3.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 480, + 430, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 430, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 430, + 492 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": "We propose a novel semantic feature discrimination (SFD) method for perceptual SISR, aiming to achieve better perception-distortion (PD) trade-off. Differing from vanilla GAN-based SR methods that only consider coarse-grained image distributions with poor semantic relevance, our SFD focuses on the distributions of semantic features. As illustrated in Fig. 4, given the input low-resolution image " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "I_{lr}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": ", we can obtain its corresponding SR image " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "I_{sr}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": " through the SRN. Then, through the pre-trained CLIP [51] image encoder " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": " we can obtain rich semantic features of " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "I_{sr}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": " with different scales and characteristics, including the multi-scale pixelwise middle features " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_{smi}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_{smi}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_{smi} \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "i = 1,2,3" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": "), and the more abstract final output feature " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_{sout} \\in \\mathbb{R}^{1 \\times L}" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": ". To better leverage these semantic features to improve the robustness of our approach, we propose two methods to separately perform discrimination on them." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "On one hand, as for the pixel-wise middle features, we design a feature discriminator (Feat-D) to distinguish be" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28178" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 556, + 281 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 556, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 556, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 556, + 281 + ], + "type": "image", + "image_path": "92675a7137dc0bfde740b3776c091ecc7bb348491a2b0a1ae82a8a776f033263.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 283, + 555, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 555, + 306 + ], + "type": "text", + "content": "Figure 4. Framework of the proposed SFD. SFD consists of a feature discriminator (Feat-D) and a text-guided discrimination (TG-D) method, Feat-D and TG-D are used to perform discrimination on the middle features and final output features from CLIP, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 327, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 296, + 410 + ], + "type": "text", + "content": "tween the feature distributions of SR and HR images. On the other hand, we propose a text-guided discrimination (TG-D) method to perform discrimination on " + }, + { + "bbox": [ + 54, + 327, + 296, + 410 + ], + "type": "inline_equation", + "content": "F_{\\text{out}}" + }, + { + "bbox": [ + 54, + 327, + 296, + 410 + ], + "type": "text", + "content": ", further enhancing the overall discriminative ability. With both Feat-D and TG-D, SFD is able to facilitate the SR network (SRN) to learn more fine-grained semantic feature distributions and generate more realistic and semantic-related textures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 251, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 251, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 251, + 430 + ], + "type": "text", + "content": "3.2. Discrimination on Semantic Features" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": "Semantic Feature Discriminator. As for the middle features of CLIP which contain multi-scale pixel-wise semantic information [30], we first design a feature discriminator, Feat-D, to distinguish between the feature distributions of SR and HR images, encouraging the semantic feature distribution of SR images to match that of HR images. Specifically, as shown in Figure 4, we can obtain the middle features " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "F_{sm1}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "F_{sm2}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "F_{sm3}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": " of SR image " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "I_{sr}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "F_{hm1}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "F_{hm2}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "F_{hm3}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": " of HR image " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "I_{hr}" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": " through the freezing CLIP image encoder " + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 55, + 436, + 296, + 593 + ], + "type": "text", + "content": ". Then, the Feat-D is used to perform discrimination on those pixel-wise semantic features. During the adversarial training process, the Feat-D and the SRN will be optimized respectively using the following adversarial loss functions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 601, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 601, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 61, + 601, + 295, + 632 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {D _ {f e a}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(1 - D \\left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\\right)\\right) \\tag {1} \\\\ + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log (D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\\\ \\end{array}", + "image_path": "6619194052b3afcf77170fe68bc99b11a392d5649eee4e6b0fdc988611576db8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 642, + 296, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 642, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 62, + 642, + 296, + 673 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {G _ {f e a}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(D \\left(F _ {h m 1}, F _ {h m 2}, F _ {h m 3}\\right)\\right) \\tag {2} \\\\ + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log (1 - D (F _ {s m 1}, F _ {s m 2}, F _ {s m 3})), \\\\ \\end{array}", + "image_path": "e85ec158ebf2eb74d129e698ae6e1cf258f704c7e740eaaa40ed0ed066276e2e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "P_{F_{hr}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "P_{F_{sr}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " represent the feature distributions of HR and SR images, respectively. Benefiting from the excellent discriminative ability of Feat-D, the SRN can learn more" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "text", + "content": "fine-grained semantic feature distributions aligned with HR images, generating more semantically relevant textures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": "Learnable Prompt Pairs Driven Text-Guided Discrimination. Discriminating the final output feature " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{sout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " which is more global and abstract is expected to further enhance the overall performance of our method. However, differing from the middle features, " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{sout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " lacks the pixel-wise semantic information due to its more abstract nature, making it less suitable for Feat-D's pixel-wise discrimination process. Therefore, leveraging the inherent visual-language connection of CLIP, we propose a text-guided discrimination method, TG-D, to perform discrimination on " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{sout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": ". Specifically, inspired by [20, 61], we introduce learnable prompt pairs (LPP) into our method. As shown in Fig. 4, given the positive prompt and negative prompt, we can obtain the text features " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{t+}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{t-}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " through the CLIP text encoder " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "E_t" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": ". Additionally, the final output features of SR image " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{sout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " and HR image " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{hout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " can be got from " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": ". Then, we adopt the same computation method as in [61] to calculate the relative cosine similarity score between " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{sout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " and the text features. This takes two steps: firstly, the cosine similarities " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "s_+" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{sout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{t+}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "s_-" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "F_{\\text{sout}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "inline_equation", + "content": "f_{t-}" + }, + { + "bbox": [ + 313, + 350, + 555, + 601 + ], + "type": "text", + "content": " can be respectively calculated as following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 327, + 605, + 555, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 605, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 327, + 605, + 555, + 632 + ], + "type": "interline_equation", + "content": "s _ {+} = \\frac {F _ {s o u t} \\odot F _ {t +}}{| | F _ {s o u t} | | \\cdot | | F _ {t +} | |}, s _ {-} = \\frac {F _ {s o u t} \\odot F _ {t -}}{| | F _ {s o u t} | | \\cdot | | F _ {t -} | |}, \\tag {3}", + "image_path": "d2e4b6e9f1ad8ef4f5253b6bb56c3b3b126850f99f8bcbd4e142d00c975a4cbe.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "text", + "content": " denotes vector dot-product and " + }, + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "inline_equation", + "content": "||\\cdot ||" + }, + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 635, + 553, + 658 + ], + "type": "text", + "content": " norm. Then, Softmax is used to calculate the final score:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 395, + 662, + 555, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 662, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 395, + 662, + 555, + 685 + ], + "type": "interline_equation", + "content": "s _ {s r} = \\frac {e ^ {s +}}{e ^ {s +} + e ^ {s -}}. \\tag {4}", + "image_path": "d472d064f4a36457f9dc4e28c2a4a3faf520dc09264d8b68485a1a0f40e6953a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "As for " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{hout}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " of HR image, we can also adopt Eq. (3) and Eq. (4) to calculate the cosine similarity score " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "s_{hr}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". Then, we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28179" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 296, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 178 + ], + "type": "image", + "image_path": "183166f0e50134e223cd3019886b2a8c423a68e909a462e087c135b20af8fd42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 296, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 214 + ], + "type": "text", + "content": "Figure 5. Framework of the proposed SFD-IQA. SFD-IQA is directly based on the well trained Feat-D and LPP, and it doesn't require any additional task-specific training on IQA datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 258 + ], + "type": "text", + "content": "utilize " + }, + { + "bbox": [ + 55, + 234, + 296, + 258 + ], + "type": "inline_equation", + "content": "s_{sr}" + }, + { + "bbox": [ + 55, + 234, + 296, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 234, + 296, + 258 + ], + "type": "inline_equation", + "content": "s_{hr}" + }, + { + "bbox": [ + 55, + 234, + 296, + 258 + ], + "type": "text", + "content": " to train the LPP and SRN in an adversarial manner:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 266, + 296, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 296, + 291 + ], + "type": "interline_equation", + "content": "L _ {D _ {t g}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(1 - s _ {h r}\\right) + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log \\left(s _ {s r}\\right), \\tag {5}", + "image_path": "5ef3259e9aab7614de5e1e42e3bc99d6fb67571f5c2ede920d559f628873e1f1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 291, + 296, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 315 + ], + "type": "interline_equation", + "content": "L _ {G _ {t g}} = \\mathbb {E} _ {F _ {h r} \\sim P _ {F _ {h r}}} \\log \\left(s _ {h r}\\right) + \\mathbb {E} _ {F _ {s r} \\sim P _ {F _ {s r}}} \\log \\left(1 - s _ {s r}\\right), \\tag {6}", + "image_path": "ec429ceed523523c441abb2118166eac44575c7a2935492755e774a2018633d7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "inline_equation", + "content": "L_{D_{tg}}" + }, + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "inline_equation", + "content": "L_{G_{tg}}" + }, + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "text", + "content": " represent the adversarial loss functions for LPP and SRN, respectively. During the adversarial training process, LPP will be more and more sensitive to image quality and TG-D will be able to discriminate " + }, + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "inline_equation", + "content": "F_{sout}" + }, + { + "bbox": [ + 55, + 315, + 296, + 388 + ], + "type": "text", + "content": " in the form of assigning a global score. Finally, the SRN is optimized by the combination of the following losses as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 395, + 296, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 395, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 67, + 395, + 296, + 425 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {t o t a l} = \\lambda_ {1} \\left\\| I _ {h r}, I _ {s r} \\right\\| _ {1} + \\lambda_ {2} \\left\\| \\phi_ {i} \\left(I _ {h r}\\right), \\phi_ {i} \\left(I _ {s r}\\right) \\right\\| _ {1} \\tag {7} \\\\ + \\lambda_ {3} L _ {G _ {f e a}} + \\lambda_ {4} L _ {G _ {t g}}, \\\\ \\end{array}", + "image_path": "8d9ba4f3bdbf431cea906ec252cca634db9777505053a368b2dc07c48a7965dd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "inline_equation", + "content": "||\\cdot ||_1" + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "text", + "content": " norm, " + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\phi_i(\\cdot)" + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "text", + "content": " is the layers of pre-trained VGG-19 network[54], " + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4" + }, + { + "bbox": [ + 55, + 431, + 296, + 467 + ], + "type": "text", + "content": " represent the hyperparameters for the above losses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 266, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 266, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 266, + 487 + ], + "type": "text", + "content": "3.3. Extension to Opinion-Unaware NR-IQA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": "During the process where the Feat-D and TG-D learns to distinguish the distributions of semantic features in an adversarial manner, they simultaneously and implicitly learn to assess image quality. By integrating the well trained Feat-D and LPP, we design a novel OU NR-IQA method, SFD-IQA. As shown in Fig. 5, given an image " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": ", we can obtain its middle semantic features " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "F_{m1}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "F_{m2}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "F_{m3}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": ", and the final output feature " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": " through the CLIP image encoder " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": ". Then, Feat-D will convert " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "F_{m1}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "F_{m2}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "inline_equation", + "content": "F_{m3}" + }, + { + "bbox": [ + 55, + 491, + 297, + 601 + ], + "type": "text", + "content": " into a score matrix:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 608, + 296, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 608, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 108, + 608, + 296, + 622 + ], + "type": "interline_equation", + "content": "\\left[ s _ {i j} \\right] _ {h \\times w} = D \\left(F _ {m 1}, F _ {m 2}, F _ {m 3}\\right), \\tag {8}", + "image_path": "04f560506dca2a606cc3d40a326402ee6e96bf9974fe446919fc2a568f352e9c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "[s_{ij}]_{h\\times w}" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " denotes the score matrix. Additionally, according to Eq. 2 and Eq. 3, we adopt " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " and LPP to calculate the similarity score " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "s_{lp}" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": ". Due to CLIP's outstanding contrastive learning training scheme, " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " is already sensitive to quality-related text prompts. To further enhance the robustness of SFD-IQA, we also adopt the original antonymic text prompts used in CLIP-IQA (e.g., \"Good photo\" and \"Bad" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "photo\") to calculate the similarity score " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": ". Then we perform a weighted average on " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "s_{lp}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 107, + 555, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 107, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 386, + 107, + 555, + 119 + ], + "type": "interline_equation", + "content": "s _ {a v e r} = \\alpha_ {1} s _ {o} + \\alpha_ {2} s _ {l p}, \\tag {9}", + "image_path": "db19826528d734e8165d089238cf8ce5e91b910562497e4533e7b4d7b4adac67.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "text", + "content": " denotes the weight coefficients. Finally, following [83], we adopt the product of " + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "inline_equation", + "content": "s_{aver}" + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "text", + "content": " and the average of " + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "inline_equation", + "content": "[s_{ij}]_{h\\times w}" + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "text", + "content": " as the final IQA score " + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "inline_equation", + "content": "s_d" + }, + { + "bbox": [ + 313, + 129, + 555, + 166 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 174, + 555, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 174, + 555, + 202 + ], + "spans": [ + { + "bbox": [ + 322, + 174, + 555, + 202 + ], + "type": "interline_equation", + "content": "s _ {d} = \\left(s _ {a v e r}\\right) \\times \\frac {\\sum_ {i = 0} ^ {h} \\sum_ {j = 0} ^ {w} \\left(\\operatorname {s i g m o i d} \\left(\\left[ s _ {i j} \\right] _ {h \\times w}\\right)\\right)}{h \\times w}, \\tag {10}", + "image_path": "e65b3fb010c7010610d3acd4e8c9f2beee4c37b275410d6bc0725bda32b2061f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "content": "where sigmoid denotes the Sigmoid function to adjust the score range to [0, 1]. Notably, we directly apply the Feat-D and TG-D to the above IQA process without any additional task-specific training. Benefiting from the semantic-related discriminative abilities of Feat-D and TG-D, SFD-IQA can significantly improve the OU NR-IQA performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 293, + 395, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 395, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 395, + 306 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 313, + 430, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 430, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 430, + 326 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 330, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 426 + ], + "type": "text", + "content": "Training Datasets. For classical SISR, following previous works [15, 30, 39], we train our network using 3450 images from DF2K [1, 58], and the bicubic downsampling is used to synthesize LR-HR training pairs. For real-world SISR and OU NR-IQA, following [39, 65, 83], we adopt DF2K and OutdoorSceneTraining [63] as the training dataset, and the LR-HR training pairs are generated by the degradation pipeline of Real-ESRGAN [65]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "content": "Test Datasets and Evaluation Metrics. For classical SISR, we test on 4 commonly used benchmarks: Set5 [3], Set14 [74], Urban100 [22], and DIV2K validation set [1], using 4 metrics to comprehensively evaluate the performance of different methods, including fidelity metrics: PSNR and SSIM [66] (calculated on the Y channel in YCbCr space), and perceptual quality metrics: LPIPS [78], DISTS [10]. For real-world SISR, following [62, 70], we employ 2 centeredropped real-world datasets, RealSR [5] and DRealSR [68]. We also create another dataset, RealD2K, by degrading DIV2K validation set under the same pipeline as that in training. We add another metric NIQE [77] for real-world SISR. For OU NR-IQA, following [39, 83], we use two categories of IQA datasets, including SR IQA datasets: PIPAL [24], Ma's dataset [41], and authentically distorted IQA datasets: KonIQ-10k [21], LIVE-itW [16]. To evaluate the IQA performance, we adopt three widely used metrics: PLCC, SRCC and KRCC [25]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Implementation Details. For the generator, following [30, 65], we use the competitive RRDB [64] as the backbone. We implement the experiments on NVIDIA 4090 GPUs with PyTorch [49]. We initialize the parameters with the pre-trained fidelity-oriented model. We use Adam [26] optimizer to train the network with a learning rate of " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28180" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 102, + 537, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of our method vs. other SOTA methods for " + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\times 4\\mathrm{{SR}}" + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": " task. The best and the second-best are marked in red and blue, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 102, + 537, + 277 + ], + "lines": [ + { + "bbox": [ + 74, + 102, + 537, + 277 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 537, + 277 + ], + "type": "table", + "html": "
BenchmarkMetricRankSR GAN [79]SRGAN [29]SFTGAN [63]ESRGAN [64]SPSR [42]LDL [32]DualFormer [39]WGSR [27]SeD [30]RRDB +OursSwinIR +\\( {L}_{GAN} \\)SwinIR +Ours
Set5PSNR ↑29.6530.5230.0630.4430.4031.0331.3031.3331.2231.6331.1031.78
SSIM ↑0.8380.8630.8480.8520.8440.8610.8690.8660.8680.8770.8690.877
LPIPS ↓0.0700.0690.0800.0740.0690.0660.0660.0650.0640.0630.0680.059
DISTS ↓0.1040.1080.1090.0970.0930.0930.0930.1100.0940.0980.0980.089
Set14PSNR ↑26.4527.0126.7426.2826.6427.1027.3927.4027.3727.7927.0527.94
SSIM ↑0.7030.7280.7180.6990.7140.7350.7390.7390.7360.7520.7310.754
LPIPS ↓0.1350.1230.1310.1310.1350.1310.1200.1380.1170.1170.1200.118
DISTS ↓0.1100.1090.1130.0990.0990.0980.0920.1120.0930.0960.0980.091
Urban100PSNR ↑24.4725.0224.3424.3524.8025.4925.6925.7825.9326.1825.8326.57
SSIM ↑0.7290.7780.7240.7340.7470.7670.7730.7810.7800.7870.7850.798
LPIPS ↓0.1380.1300.1340.1230.1210.1100.1150.1350.1070.1090.1000.099
DISTS ↓0.1040.1100.1060.0880.0860.0820.0850.1080.0790.0820.0810.079
DIV2K100PSNR ↑28.0328.1828.0928.2028.1828.9629.2529.1929.2729.7528.9929.90
SSIM ↑0.7670.7880.7710.7770.7720.7950.8020.8040.8030.8150.8030.818
LPIPS ↓0.1210.1220.1330.1150.1130.1010.0970.1040.0940.0970.0940.099
DISTS ↓0.0640.0630.0740.0590.0550.0530.0560.0540.0540.0530.0500.050
", + "image_path": "b08e4a6d45ec14a68fa017488ee1262efd48b382c8e817b3660e023ef2b79a2d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 287, + 555, + 529 + ], + "blocks": [ + { + "bbox": [ + 56, + 287, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 56, + 287, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 555, + 529 + ], + "type": "image", + "image_path": "bfaf4422242b935e6f2854901770c6de0764d3d56474ad9661236473bfd15105.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 531, + 468, + 544 + ], + "lines": [ + { + "bbox": [ + 141, + 531, + 468, + 544 + ], + "spans": [ + { + "bbox": [ + 141, + 531, + 468, + 544 + ], + "type": "text", + "content": "Figure 6. Visual comparison of different GAN-based SR methods on " + }, + { + "bbox": [ + 141, + 531, + 468, + 544 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 141, + 531, + 468, + 544 + ], + "type": "text", + "content": " super-resolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 563, + 198, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 198, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 198, + 574 + ], + "type": "text", + "content": "4.2. Results on Classical SISR" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "To validate the effectiveness of our method, we compare it with several SOTA GAN-based SR methods, including [27, 29, 30, 32, 39, 42, 64, 79]. As shown in Tab. 1, compared to other approaches, our method achieves the best PSNR and SSIM scores on all benchmarks, while also holding highly competitive perceptual metrics, meaning that our method can sacrifice less fidelity in exchange for improved perceptual quality. To validate generalizability, we also employ another Transformer-based SR backbone, SwinIR [31]. It's clearly that our method also demonstrates superior performance with SwinIR. Fig. 6 presents the visual comparison, indicating" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 564, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 555, + 588 + ], + "type": "text", + "content": "that our method generates more realistic semantic textures with fewer artifacts than other methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 599, + 471, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 471, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 471, + 611 + ], + "type": "text", + "content": "4.3. Results on Real-World SISR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "For real-world SISR, we provide quantitative and visual comparison with representative SOTA real-world SISR models including [31, 32, 65, 75]. The quantitative comparison is demonstrated in Table 2. The results show that our SFD achieves the best SSIM score on all datasets, while also maintaining competitive perceptual metrics, especially LPIPS and DISTS, indicating that our approach can achieve better PD trade-off. As shown in Fig. 7, our method produces SR" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28181" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 126, + 139 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 126, + 139 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 126, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 126, + 139 + ], + "type": "image", + "image_path": "bc10b71cceb15f3f3e6ad5da0237c259fc148e087f7b59cbc39c5579dd901707.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 141, + 96, + 148 + ], + "lines": [ + { + "bbox": [ + 85, + 141, + 96, + 148 + ], + "spans": [ + { + "bbox": [ + 85, + 141, + 96, + 148 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 127, + 70, + 197, + 139 + ], + "blocks": [ + { + "bbox": [ + 127, + 70, + 197, + 139 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 197, + 139 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 197, + 139 + ], + "type": "image", + "image_path": "0e7d99f196ba262691ed54647598466c018e8f995db6e80303b05c9bdda30a4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 141, + 175, + 149 + ], + "lines": [ + { + "bbox": [ + 150, + 141, + 175, + 149 + ], + "spans": [ + { + "bbox": [ + 150, + 141, + 175, + 149 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 199, + 70, + 269, + 139 + ], + "blocks": [ + { + "bbox": [ + 199, + 70, + 269, + 139 + ], + "lines": [ + { + "bbox": [ + 199, + 70, + 269, + 139 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 269, + 139 + ], + "type": "image", + "image_path": "3688d96a7a6252d3e825b160d96cb4eeb6bd719c03da937210c3b5f686e7b8df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 141, + 249, + 149 + ], + "lines": [ + { + "bbox": [ + 219, + 141, + 249, + 149 + ], + "spans": [ + { + "bbox": [ + 219, + 141, + 249, + 149 + ], + "type": "text", + "content": "BSRGAN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 270, + 70, + 340, + 139 + ], + "blocks": [ + { + "bbox": [ + 270, + 70, + 340, + 139 + ], + "lines": [ + { + "bbox": [ + 270, + 70, + 340, + 139 + ], + "spans": [ + { + "bbox": [ + 270, + 70, + 340, + 139 + ], + "type": "image", + "image_path": "39a5e3ce40fdf4ab19cb7a6caf6c2245f192737b2a9069ee043c618025edf165.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 141, + 326, + 149 + ], + "lines": [ + { + "bbox": [ + 284, + 141, + 326, + 149 + ], + "spans": [ + { + "bbox": [ + 284, + 141, + 326, + 149 + ], + "type": "text", + "content": "ReaESRGAN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 122, + 154, + 486, + 165 + ], + "lines": [ + { + "bbox": [ + 122, + 154, + 486, + 165 + ], + "spans": [ + { + "bbox": [ + 122, + 154, + 486, + 165 + ], + "type": "text", + "content": "Figure 7. Visual comparison of different GAN-based real-world SR methods on " + }, + { + "bbox": [ + 122, + 154, + 486, + 165 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 122, + 154, + 486, + 165 + ], + "type": "text", + "content": " super-resolution." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 70, + 411, + 139 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 411, + 139 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 411, + 139 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 411, + 139 + ], + "type": "image", + "image_path": "2d510b580f555012f5ddfa30e091f618775918d9959ab2d1142e503f28c1f759.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 141, + 384, + 148 + ], + "lines": [ + { + "bbox": [ + 369, + 141, + 384, + 148 + ], + "spans": [ + { + "bbox": [ + 369, + 141, + 384, + 148 + ], + "type": "text", + "content": "LDL" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 70, + 482, + 139 + ], + "blocks": [ + { + "bbox": [ + 413, + 70, + 482, + 139 + ], + "lines": [ + { + "bbox": [ + 413, + 70, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 413, + 70, + 482, + 139 + ], + "type": "image", + "image_path": "6096729d2fd689112840df82219cbc58482ebb9e5c70ea378e264e5fafe0fa2a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 141, + 466, + 149 + ], + "lines": [ + { + "bbox": [ + 430, + 141, + 466, + 149 + ], + "spans": [ + { + "bbox": [ + 430, + 141, + 466, + 149 + ], + "type": "text", + "content": "RealSwinIR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 485, + 70, + 554, + 139 + ], + "blocks": [ + { + "bbox": [ + 485, + 70, + 554, + 139 + ], + "lines": [ + { + "bbox": [ + 485, + 70, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 485, + 70, + 554, + 139 + ], + "type": "image", + "image_path": "a932a7fc76a2ff3336e219fac6d526bb82b12ead11cdafef833b0625668a594b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 141, + 526, + 148 + ], + "lines": [ + { + "bbox": [ + 511, + 141, + 526, + 148 + ], + "spans": [ + { + "bbox": [ + 511, + 141, + 526, + 148 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 58, + 221, + 294, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 185, + 296, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 218 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of our method vs. other state-of-the-art real-world SISR methods for " + }, + { + "bbox": [ + 55, + 185, + 296, + 218 + ], + "type": "inline_equation", + "content": "\\times 4\\mathrm{{SR}}" + }, + { + "bbox": [ + 55, + 185, + 296, + 218 + ], + "type": "text", + "content": " task. The best and the second-best are marked in red and blue, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 221, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 58, + 221, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 221, + 294, + 376 + ], + "type": "table", + "html": "
BenchmarkMetricsBSRGAN [75]RealESR GAN [65]LDL [32]RealSwinIR [31]SFD (Ours)
RealD2KPSNR ↑23.5123.5022.9422.6723.40
SSIM ↑0.5850.6040.5970.5810.606
LPIPS ↓0.3790.3680.3730.3680.358
DISTS ↓0.2800.2780.2840.2790.278
NIQE ↓7.5378.3698.3847.8668.600
RealSRPSNR ↑26.3925.6925.2825.9026.12
SSIM ↑0.7650.7620.7570.7680.772
LPIPS ↓0.2670.2730.2750.2620.256
DISTS ↓0.2120.2060.2120.2000.200
NIQE ↓5.6575.8305.9925.9265.797
DRealSRPSNR ↑28.7528.6428.2128.2129.20
SSIM ↑0.8030.8050.8130.7980.815
LPIPS ↓0.2880.2850.2790.2850.284
DISTS ↓0.2140.2090.2130.2100.221
NIQE ↓6.5196.6937.1436.7666.878
", + "image_path": "a73b44f922574fc2f513ba3d3ae004e680a2e5777841898b0edb441c76d48d2f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 398, + 296, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 457 + ], + "type": "text", + "content": "images with sharp edges and fine textures, accompanied by fewer undesirable artifacts. The above results suggest that our SFD can distinguish complex degraded real world images and facilitate the SR network to learn more fine-grained and semantically accurate textures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 466, + 189, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 189, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 189, + 478 + ], + "type": "text", + "content": "4.4. Results on OU NR-IQA" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 483, + 297, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 297, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 297, + 604 + ], + "type": "text", + "content": "To verify the effectiveness of our SFD-IQA for OU NR-IQA tasks, we compared SFD-IQA with several representative OU NR-IQA methods, including traditional methods [37, 69, 71, 76, 77], deep learning-based methods [2, 35, 39, 43, 46, 61, 83]. The results are presented in Table 3. Thanks to the excellent ability of Feat-D and TG-D to distinguish between low-quality and high-quality images, learned during the real-world SISR training process, our SFD-IQA achieves the best performance across all four datasets, validating its enhanced ability to predict the perceptual quality of images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 612, + 156, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 156, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 156, + 623 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "content": "The effects of Feat-D and TG-D on perceptual SISR. To investigate the effectiveness of our Feat-D and TG-D, we conduct ablation experiments on them using the DIV2K100 test set. As shown in Fig. 8, compared to the baseline (ESR-GAN), Feat-D and TG-D can both independently improve the perceptual quality with less decrease in fidelity. Furthermore, when both Feat-D and TG-D are introduced into the" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 333, + 185, + 536, + 302 + ], + "blocks": [ + { + "bbox": [ + 333, + 185, + 536, + 302 + ], + "lines": [ + { + "bbox": [ + 333, + 185, + 536, + 302 + ], + "spans": [ + { + "bbox": [ + 333, + 185, + 536, + 302 + ], + "type": "image", + "image_path": "6dc3df15b417a06c1b1d63769d21eb1885c8d4bd09404c4640e0c07145f46d51.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 304, + 548, + 316 + ], + "lines": [ + { + "bbox": [ + 320, + 304, + 548, + 316 + ], + "spans": [ + { + "bbox": [ + 320, + 304, + 548, + 316 + ], + "type": "text", + "content": "Figure 8. The effects of Feat-D and TG-D on perceptual SISR." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 318, + 327, + 433, + 419 + ], + "blocks": [ + { + "bbox": [ + 318, + 327, + 433, + 419 + ], + "lines": [ + { + "bbox": [ + 318, + 327, + 433, + 419 + ], + "spans": [ + { + "bbox": [ + 318, + 327, + 433, + 419 + ], + "type": "image", + "image_path": "9c0230ae1c742bae1c7fbceaad6d6c0494be433b6472580bb7e40449a053dd88.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 421, + 429, + 429 + ], + "lines": [ + { + "bbox": [ + 323, + 421, + 429, + 429 + ], + "spans": [ + { + "bbox": [ + 323, + 421, + 429, + 429 + ], + "type": "text", + "content": "t-SNE of vanilla discriminator features" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 433, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 477 + ], + "type": "text", + "content": "Figure 9. The t-SNE visualization of vanilla discriminator features and our Feat-D features. We select 8 categories of images from OST dataset, including \"bear\", \"building\", \"cat\", \"dog\", \"grass\", \"mountain\", \"sky\" and \"water\" and 50 images for each category." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 435, + 327, + 551, + 419 + ], + "blocks": [ + { + "bbox": [ + 435, + 327, + 551, + 419 + ], + "lines": [ + { + "bbox": [ + 435, + 327, + 551, + 419 + ], + "spans": [ + { + "bbox": [ + 435, + 327, + 551, + 419 + ], + "type": "image", + "image_path": "9c500fb3b443b488235480078a749d3767fbe6f30231394676d9cc662ac5cddd.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 421, + 533, + 429 + ], + "lines": [ + { + "bbox": [ + 454, + 421, + 533, + 429 + ], + "spans": [ + { + "bbox": [ + 454, + 421, + 533, + 429 + ], + "type": "text", + "content": "t-SNE of our Feat-D features" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 498, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 533 + ], + "type": "text", + "content": "baseline, we can trade the minimum PSNR decrease for the best perceptual quality. The above ablation results strongly verify the superiority of Feat-D and TG-D." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "text", + "content": "The effects of Feat-D and TG-D in SFD-IQA. To investigate the contribution of our Feat-D and TG-D to the proposed SFD-IQA, we conduct ablation studies by respectively remove Feat-D and TG-D from our SFD-IQA. Tab. 4 presents the comparative results on the 3 SR related subtypes of PIPAL [24] training set and 3 sub-types of KonIQ-10k [21] dataset. Compared to the baseline without Feat-D and TG-D, it's clearly that each of Feat-D and TG-D can independently enhance the OU NR-IQA performance to a certain extent. With both Feat-D and TG-D, SFD-IQA significantly improves the performance in most cases. Especially on the GAN-based SR subset of PIPAL, it boosts the PLCC by " + }, + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "inline_equation", + "content": "86.0\\%" + }, + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "text", + "content": " and SRCC by " + }, + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "inline_equation", + "content": "28.8\\%" + }, + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "text", + "content": ". The above results demonstrate the effectiveness of Feat-D and TG-D." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "T-SNE analysis for Feat-D. To further demonstrate the" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28182" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 86, + 531, + 233 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 544, + 82 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 544, + 82 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 544, + 82 + ], + "type": "text", + "content": "Table 3. OU NR-IQA comparison between different methods. The best and the second-best are marked in red and blue, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 86, + 531, + 233 + ], + "lines": [ + { + "bbox": [ + 78, + 86, + 531, + 233 + ], + "spans": [ + { + "bbox": [ + 78, + 86, + 531, + 233 + ], + "type": "table", + "html": "
IQA MethodsPIPAL [24]KonIQ-10k [21]Ma's dataset [41]LIVE-itW [16]
PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑PLCC ↑SRCC ↑KRCC ↑
QAC [71]0.25660.22200.15030.37190.33970.23020.50710.43870.30550.27200.04570.0370
LPSI [69]0.28970.20310.13630.20660.22390.15040.52930.48960.35430.28770.08340.0524
NIQE [77]0.04390.06080.04010.46370.52910.36660.64270.62610.45200.47900.44930.3062
IL-NIQE [76]0.27480.24120.16300.53160.50570.35040.71550.65480.47510.50390.43940.2985
SNP-NIQE [37]0.18200.18010.12140.63400.62850.44350.62500.56210.39840.52010.46570.3162
dipIQ [43]0.28000.23040.15350.44290.23750.15940.51670.48660.36070.31800.17740.1207
RankIAQ [35]0.36980.37150.25220.50280.49830.34480.51780.49130.36480.45280.43070.2945
RecycleD [83]0.33960.33310.22650.64070.63640.44680.59830.55550.38740.48060.46610.3194
DualFormer [39]0.28960.27870.19730.65430.63210.44350.61720.60540.40440.50680.48970.3312
ContentSep [2]---0.62740.64010.4529---0.51300.50600.3450
CLIPIAQ [61]0.28310.27620.18620.43300.41450.28210.71750.70550.50630.58530.56470.4010
MDFS [46]0.26510.22730.15320.71230.73330.53440.76490.71420.52050.53640.48210.3274
SFD-IQA (Ours)0.42520.40920.28160.75740.74400.54430.80120.78700.58640.71110.69840.5090
", + "image_path": "afc02d2e84a292c8b8c2232d0174c63b63c2c403803aca2f86722cf7b5703be8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 299, + 288, + 414 + ], + "blocks": [ + { + "bbox": [ + 55, + 251, + 297, + 295 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 295 + ], + "type": "text", + "content": "Table 4. OU NR-IQA performance comparison on sub-types of PIPAL training set and KonIQ-10k dataset. \"Baseline\" refers to the model without neither Feat-D nor TG-D. \"W/ Feat-D\" and \"w/. TG-D\" refer to the models that only include Feat-D or TG-D." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 299, + 288, + 414 + ], + "lines": [ + { + "bbox": [ + 61, + 299, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 61, + 299, + 288, + 414 + ], + "type": "table", + "html": "
MetricsDataset. Sub-TypesBaselinew/.Feat-Dw/.TG-DSFD-IQA (Ours)
PLCCPIPAL.Trad.SR0.37870.36840.39100.3797 (+0.26%)
PIPAL.PSNR.SR0.42900.45040.45470.4657 (+8.55%)
PIPAL.GAN.SR0.21540.39580.24300.4006 (+86.0%)
KonIQ-10k.Train0.70920.74840.74080.7581 (+6.90%)
KonIQ-10k.Val0.70550.74520.73500.7528 (+6.70%)
KonIQ-10k.Test0.68650.74930.74650.7576 (+10.4%)
SRCCPIPAL.Trad.SR0.35450.34200.36600.3522 (-0.64%)
PIPAL.PSNR.SR0.44320.47970.46720.4932 (+11.3%)
PIPAL.GAN.SR0.23760.30480.25980.3061 (+28.8%)
KonIQ-10k.Train0.68490.73250.71300.7441 (+4.92%)
KonIQ-10k.Val0.68030.72640.70720.7373 (+4.51%)
KonIQ-10k.Test0.68620.73590.71310.7475 (+8.89%)
", + "image_path": "70ad08bef6f7db25b60d64797937aab2c2b150821d64df80257faaee1f06827c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 426, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 568 + ], + "type": "text", + "content": "effectiveness of our Feat-D, we use t-SNE [59] to visualize the middle features of the VGG-style vanilla discriminator and our Feat-D. We select 8 categories of images from the OST [63] dataset, and 50 images are randomly selected for each category. We visualize features after the 3rd upsampling layer of our Feat-D and after the 2nd BN layer of the vanilla discriminator, respectively. As shown in Fig. 9, the feature clustering of our Feat-D is much better than that of the vanilla discriminator. This strongly proves that the proposed Feat-D is semantic-aware and can perform more fine-grained discrimination, encouraging the SR network to generate more more realistic semantic textures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "The effectiveness of our method on other image restoration tasks. We preliminarily explore the effectiveness of our approach in addressing other image restoration tasks, such as image deraining and image deblurring. We select a commonly used image restoration baselines NAFNet [8], and conduct training and testing on the Rain100L [73] and GoPro [45] datasets for the two tasks. As shown in Tab. 5, compared to fidelity-oriented baseline methods and vanilla discriminator based methods, our method significantly improves perceptual metrics while maintaining a high level of fidelity on both two tasks, demonstrating the potential of our method to be integrated into other image" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 319, + 277, + 545, + 333 + ], + "blocks": [ + { + "bbox": [ + 313, + 251, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 273 + ], + "type": "text", + "content": "Table 5. The effectiveness of our method on image deraining and image deblurring. \"VD\" means vanilla discriminator." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 277, + 545, + 333 + ], + "lines": [ + { + "bbox": [ + 319, + 277, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 319, + 277, + 545, + 333 + ], + "type": "table", + "html": "
MethodRain100L [73]GoPro [45]
PSNR ↑LPIPS ↓DISTS ↓PSNR ↑LPIPS ↓DISTS ↓
NAFNet37.000.0220.03132.870.0930.072
+ VD36.390.0150.02032.010.0680.036
+ Ours36.490.0140.01932.190.0640.034
", + "image_path": "54e048b6ee35751bf4315ebca67a15d70cd64e120cbcde425eacb444d8c58530.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 348, + 498, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 498, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 498, + 361 + ], + "type": "text", + "content": "restoration tasks for better perceptual quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 370, + 388, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 388, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 388, + 383 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 312, + 390, + 555, + 582 + ], + "type": "text", + "content": "In this paper, we propose a semantic feature discrimination method, SFD, for perceptual SISR and OU NR-IQA. We first design a feature discriminator (Feat-D) to discriminate the semantic features from CLIP, effectively encouraging the feature distributions of SR images to match that of HR images. Then we propose a text-guided discrimination (TG-D) method by training the learnable prompt pairs (LPP) in an adversarial manner to further enhance the discriminative ability. Our SFD facilitates the SR network to generate more fine-grained and more realistic semantic textures, achieving better perception-distortion trade-off. Furthermore, by effectively integrating the well trained Feat-D and LPP, we significantly improve the OU NR-IQA performance without any additional task-specific training on IQA datasets. Extensive experiments on classical SISR, real-world SISR, and OU NR-IQA tasks validate the effectiveness of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09, the TCL Science and Technology Innovation Fund under grant 25JZH008, and the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28183" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "text", + "content": "[2] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 237 + ], + "type": "text", + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference (BMVC), 2012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 296, + 272 + ], + "type": "text", + "content": "[4] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "type": "text", + "content": "[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "type": "text", + "content": "[6] Du Chen, Zhengqiang Zhang, Jie Liang, and Lei Zhang. Ssl: A self-similarity loss for improving generative image superresolution. In ACM Multimedia 2024, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "text", + "content": "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 421, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 296, + 463 + ], + "type": "text", + "content": "[8] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "text", + "content": "[9] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25974-25984, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 511, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 296, + 555 + ], + "type": "text", + "content": "[10] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 556, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 296, + 612 + ], + "type": "text", + "content": "[11] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 678 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[13] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "conference on computer vision and pattern recognition, pages 1103-1112, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[14] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[15] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 229 + ], + "type": "text", + "content": "[16] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 276, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 555, + 319 + ], + "type": "text", + "content": "[18] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 322, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 555, + 364 + ], + "type": "text", + "content": "[19] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 366, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 555, + 409 + ], + "type": "text", + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 555, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 454 + ], + "type": "text", + "content": "[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[22] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 501, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 555, + 544 + ], + "type": "text", + "content": "[23] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 546, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 555, + 611 + ], + "type": "text", + "content": "[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 634 + ], + "type": "text", + "content": "[25] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 657 + ], + "type": "text", + "content": "[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[27] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 3, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28184" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 297, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 297, + 205 + ], + "type": "text", + "content": "[29] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "text", + "content": "[30] Bingchen Li, Xin Li, Hanxin Zhu, Yeying Jin, Ruoyu Feng, Zhizheng Zhang, and Zhibo Chen. Sed: Semantic-aware discriminator for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25784-25795, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 296, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 318 + ], + "type": "text", + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "text", + "content": "[32] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 296, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 419 + ], + "type": "text", + "content": "[33] Xiangyu Liao, Tianheng Zheng, Jiayu Zhong, Pingping Zhang, and Chao Ren. Asymmetric mask scheme for self-supervised real image denoising. arXiv preprint arXiv:2407.06514, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 475 + ], + "type": "text", + "content": "[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 531 + ], + "type": "text", + "content": "[35] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 296, + 577 + ], + "type": "text", + "content": "[36] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 633 + ], + "type": "text", + "content": "[37] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2019. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "text", + "content": "[38] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "type": "text", + "content": "[39] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13243-13253, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "text", + "content": "[40] Chenxi Ma. Uncertainty-aware gan for single image super resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4071-4079, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[41] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 158:1-16, 2017. 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 555, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 237 + ], + "type": "text", + "content": "[42] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "text", + "content": "[43] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 295, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 555, + 337 + ], + "type": "text", + "content": "[44] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "text", + "content": "[45] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 555, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 437 + ], + "type": "text", + "content": "[46] Zhangkai Ni, Yue Liu, Keyan Ding, Wenhan Yang, Hanli Wang, and Shiqi Wang. Opinion-unaware blind image quality assessment using multi-scale deep feature statistics. IEEE Transactions on Multimedia, 2024. 3, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 438, + 555, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 555, + 482 + ], + "type": "text", + "content": "[47] JoonKyu Park, Sanghyun Son, and Kyoung Mu Lee. Content-aware local gan for photo-realistic super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10585-10594, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 483, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 555, + 536 + ], + "type": "text", + "content": "[48] Seung Ho Park, Young Su Moon, and Nam Ik Cho. Perception-oriented single image super-resolution using optimal objective estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1725-1735, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 602 + ], + "type": "text", + "content": "[49] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 604, + 555, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 647 + ], + "type": "text", + "content": "[50] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28185" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[52] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[53] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8207-8216, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "text", + "content": "[55] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 242, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 286 + ], + "type": "text", + "content": "[56] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "text", + "content": "[57] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "text", + "content": "[58] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "text", + "content": "[60] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 522 + ], + "type": "text", + "content": "[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "type": "text", + "content": "[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 2, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "text", + "content": "[64] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF inter" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "content": "national conference on computer vision, pages 1905-1914, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 196 + ], + "type": "text", + "content": "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "type": "text", + "content": "[68] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision - ECCV 2020, pages 101-117. Springer International Publishing, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "text", + "content": "[69] Qingbo Wu, Zhou Wang, and Hongliang Li. A highly efficient method for blind image quality assessment. In 2015 IEEE International Conference on Image Processing (ICIP), pages 339-343. IEEE, 2015. 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 354 + ], + "type": "text", + "content": "[70] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25456-25467, 2024. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 399 + ], + "type": "text", + "content": "[71] Wufeng Xue, Lei Zhang, and Xuanqin Mou. Learning without human scores for blind image quality assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 995-1002, 2013. 3, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "text", + "content": "[72] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5791-5800, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "type": "text", + "content": "[73] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 555, + 566 + ], + "type": "text", + "content": "[74] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 555, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 621 + ], + "type": "text", + "content": "[75] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 623, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 554, + 656 + ], + "type": "text", + "content": "[76] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 657, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 554, + 690 + ], + "type": "text", + "content": "[77] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 3, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28186" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 352 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[79] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Super resolution generative adversarial networks with learning to rank. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7149-7166, 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[80] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image super-resolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "text", + "content": "[81] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 296, + 297 + ], + "type": "text", + "content": "[82] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "type": "text", + "content": "[83] Yunan Zhu, Haichuan Ma, Jialun Peng, Dong Liu, and Zhiwei Xiong. Recycling discriminator: Towards opinion-unaware image quality assessment using Wasserstein gan. In Proceedings of the 29th ACM International Conference on Multimedia, pages 116-125, 2021. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28187" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_content_list.json b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e3d2b6a338bbda5d6e1b3431e817cde3f95146a --- /dev/null +++ b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_content_list.json @@ -0,0 +1,1688 @@ +[ + { + "type": "text", + "text": "Exploring Simple Open-Vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 196, + 130, + 800, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zihang Lai Visual Geometry Group (VGG), University of Oxford zlai@robots.ox.ac.uk", + "bbox": [ + 285, + 181, + 710, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary semantic segmentation models aim to accurately assign a semantic label to each pixel in an image from a set of arbitrary open-vocabulary texts. In order to learn such pixel-level alignment, current approaches typically rely on a combination of (i) image-level VL model (e.g. CLIP), (ii) ground truth masks, (iii) custom grouping encoders, and (iv) the Segment Anything Model (SAM). In this paper, we introduce S-Seg, a simple model that can achieve surprisingly strong performance without depending on any of the above elements. S-Seg leverages pseudo-masks and language features to train a MaskFormer, and can be easily trained from publicly available image-text datasets. Contrary to prior works, our model directly trains for pixel-level features and language alignment. Once trained, S-Seg generalizes well to multiple testing datasets without requiring fine-tuning. In addition, S-Seg has the extra benefits of scalability with data and consistently improving when augmented with self-training. We believe that our simple yet effective approach will serve as a solid baseline for future research. Project page: zlai0.github.io/S-Seg.", + "bbox": [ + 89, + 300, + 483, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 632, + 220, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary semantic segmentation presents a unique challenge as it requires assigning accurate semantic labels to each pixel in an image using arbitrary open-vocabulary texts, rather than a fixed set of classes. This means that the model must be able to segment and classify any arbitrary categories expressed in language. Achieving this requires a robust, pixel-level alignment between images and textual descriptions, which enables accurate association of each pixel with the most relevant class from a dynamically provided set of textual categories.", + "bbox": [ + 89, + 657, + 482, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A primary obstacle in this domain is that it is impossible to construct datasets that provide pixel-level annotations for all possible labels. This limitation often results in the adoption of weakly-supervised or semi-supervised learning approaches. Current methods typically rely on a combination of strategies to learn the required pixel-level alignment.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2caf8be573234d8bca9040c0c5fd01f221a52463bbf41bc5be8e273eb99324db.jpg", + "image_caption": [ + "Figure 1. S-Seg result on a web image. Our goal is to segment everything, including fictional characters like minions.", + "Figure 2. Our S-Seg framework leverages pseudo-mask and language to train a MaskFormer. We show that our method of directly training for pixel-level feature and language alignment yields superior results." + ], + "image_footnote": [], + "bbox": [ + 519, + 311, + 906, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One common tactic is adapting existing Vision-Language (VL) models, which are initially trained for image-level alignment (e.g., CLIP [37]), to perform at the pixel level. Another strategy involves training models on ground truth masks that are annotated for a select number of seen classes, thereby encouraging the model to extrapolate its knowledge to novel unseen classes during testing. Furthermore, specialized models such as GroupViT [47] and OVSegmentor [48], which are explicitly designed for open-", + "bbox": [ + 511, + 580, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vocabulary segmentation, are being explored. Lastly, recent literature adapts the Segment Anything Model (SAM), originally not designed to classify segments, for effective use in open-vocabulary segmentation.", + "bbox": [ + 511, + 717, + 905, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we report a model that can work surprisingly well with none of the above strategies. Our approach, named S-Seg, is built on top of a standard MaskFormer model. Our model directly trains for pixel-level feature and language alignment, using neither existing large image-level alignment models like CLIP [37] nor manually annotated segmentation or classification labels.", + "bbox": [ + 511, + 777, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the biggest challenges we face is finding the right", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30221", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fa11898c8567044470900c48196ba772a25b7a2f0f1e958c22cf8ba3fd43a72.jpg", + "image_caption": [ + "Figure 3. Qualitative results of S-Seg, evaluated using all dataset classes as queries. Our model copes with challenging situation, such as overlapping objects (col. 2) and small objects (col. 5). Our model is also capable of handling \"stuff\" categories such as water and floor (col. 3, 4). Moreover, our S-Seg+ model is able to correct small errors observed in the S-Seg method (col. 4). Finally, in the COCO dataset, which featured a significantly higher number of objects, our model is still able to achieve high accuracy in its predictions." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervision since annotated masks and labels are not available. To address this issue, we propose to leverage pseudomasks and language to supervise MaskFormer. Our strategy involves using a pseudo-mask generator to provide class-agnostic mask supervision by generating pseudo ground truth masks. We adopt a simple design that clusters image representations obtained through self-supervised representation learning methods like DINO [3]. Our experiments demonstrate that this approach delivers exceptional performance, which is essential for high-quality supervision, as well as rapid processing speed, which is necessary for efficient training. In addition, we use noisy web texts to provide semantic supervision. The image-text dataset contains a wide range of concepts and has demonstrated impressive zero-shot classification results [37]. We utilize a straightforward image-text contrastive loss, which has proven to be highly effective. Once trained, our model generalizes well to new categories without requiring fine-tuning.", + "bbox": [ + 88, + 378, + 480, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "S-Seg is a simple and effective model that can be trained using publicly available image-text datasets, such as Conceptual Captions [5, 40]. This makes it easy to reproduce and extend for further research. The S-Seg framework is also designed to be flexible with easily replaceable submodules. We prioritize simplicity in our subcomponent selection to focus on the general design of our framework, while remaining open to more advanced techniques that could result in further improvements.", + "bbox": [ + 89, + 654, + 482, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted a thorough evaluation of S-Seg using multiple benchmark datasets, and we show that our method achieve competitive results on three widely tested benchmarks (Pascal VOC, Pascal Context, and COCO). In addition, pseudo-mask and language provide scalable supervision and our model consistently improves in performance as more data became available. Finally, we find adding an ad", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ditional self-training step leads to an even greater improvement to our model, with an average increase of $5.5\\%$ mIoU over three datasets, highlighting the potential for further improvement of our approach.", + "bbox": [ + 511, + 378, + 903, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our simple solution suggests that the reliance on complex models and extensive ground truth data in open-vocabulary semantic segmentation may be reduced, leading to more streamlined and accessible framework for future developments in the field, and we hope our exploration can serve as a solid baseline for future research.", + "bbox": [ + 511, + 439, + 905, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 542, + 651, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary segmentation. The earliest efforts to employ language for image segmentation can be traced back to Duygulu et al.'s seminal work [13], where the authors tackled image segmentation by framing it as a machine translation problem. Current approaches leverage a combination of strategies to learn pixel-level image-text alignment.", + "bbox": [ + 511, + 568, + 905, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adapting image-level vision-language models. The first strategy involves the adapting pretrained vision-language models, originally designed for image-level alignment, to the more granular task of pixel-level alignment. This strategy is widely adopted in open-vocabulary methods [4, 15, 18, 19, 22, 24, 27, 32, 33, 38, 44, 48, 49, 52]. These works vary in their methods of refining image-level models for finer alignment tasks.", + "bbox": [ + 511, + 674, + 905, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MaskCLIP [52] demonstrates modifying the CLIP image encoder can significantly enhance its pixel-level alignment capabilities without requiring retraining. TCL [4] employs CLIP for initial text-to-image region grounding, followed by contrastive learning to refine the alignment between the text embedding and the grounded region. OpenSeg [15] fine-tunes ALIGN [20] using a grounding", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30222", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f344ff7f03ca2d473871f309a0bc51836d738753576f4feec0f0925555b3c2dc.jpg", + "image_caption": [ + "Figure 4. Overview of S-Seg. A MaskFormer model computes masks and mask features from an image input. A pseudo-mask generator produces segmentation maps to supervise mask predictions, while a text that describes the image, encoded by a language model trained together with the MaskFormer, provides supervision for mask features using image-text contrastive loss.", + "Figure 5. Pseudocode for training S-Seg with image-text pairs." + ], + "image_footnote": [], + "bbox": [ + 150, + 73, + 851, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "loss [17] to better align words in captions to segmentation masks. OpenSeg and DiffuMask [44] also explored the use of pseudo-masks. The primary distinction lies in their dependency on different sources for learning; OpenSeg uses annotated segments while DiffuMask employs masks generated through diffusion. In contrast, our method is entirely learned from pseudo masks. Also, our mask generator is entirely self-supervised, whereas their mask generator is fully-supervised.", + "bbox": [ + 88, + 271, + 480, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ground truth masks. Another effective strategy [11, 15, 18, 19, 24, 27, 33] involves training models using ground truth masks annotated for a limited set of seen classes. By training on seen annotations, models are encouraged to learn detailed features and patterns that are potentially applicable beyond the scope of the trained classes.", + "bbox": [ + 89, + 410, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Of most relevance, ZegFormer [11] trains a MaskFormer by decoupling zero-shot semantic segmentation into two sub-tasks, a class-agnostic grouping task and a zero-shot segment classification task. Our method has similar training paradigm but with notable distinctions. Similar to GroupViT, we train exclusively with image-text pairs and do not utilize a pretrained CLIP model. Notably, even without access to ground truth masks, labels, or CLIP, our method outperforms ZegFormer in unseen categories, indicating potentially stronger generalization.", + "bbox": [ + 88, + 503, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Custom grouping-based encoders. The third strategy employs custom-designed models specifically for open-vocabulary segmentation. GroupViT [47] groups pixels in an image hierarchically based on their attention scores with learnable group tokens. OVSegmentor [48] applies Slot Attention [31] for a similar pixel grouping process based on feature proximity.", + "bbox": [ + 89, + 656, + 483, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segment Anything Model (SAM). The final strategy adapts the Segment Anything Model (SAM) [23], a powerful segmentation model trained on over 11 million images and 1.1 billion masks, for open-vocabulary segmentation. While the SAM model segments object with high accuracy, it does not effectively classify these segments. To address this limitation, follow-up works has adapted the SAM model for grounding, closed-set, and open-vocabulary segmentation [25, 39, 51] by leveraging additional training", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "I [n, h, w, c] - minibatch of aligned images \nT [n, l] - minibatch of aligned texts \nN - number of MaskFormer queries \nC - number of pseudo masks \n# predict mask, mask feature, and text feature \nM, M_f = maskformer(I) # [n, N, H, W], [n, N, d_f] \nT_f = textEncoder(T) # [n, d_f] \n# aggregate all mask features [n, d_f] \nM_f = M_f.mean(axis=1) \n# generate pseudo masks [n, C, H, W] \nS = pseudo_mask_generation(I) \n# compute loss \nloss_c = contrastive_loss(M_f, T_f) \nloss_m = mask_loss(M, S) \nloss = (loss_c + loss_m)/2", + "guess_lang": "python", + "bbox": [ + 526, + 273, + 890, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data, adaptors, and distillation.", + "bbox": [ + 511, + 529, + 718, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model, S-Seg, can be conceptualized as a synergy of these approaches. It can be viewed as a CLIP model integrated with a MaskFormer image encoder, directly optimizing for pixel-level feature and language alignment. Alternatively, it resembles \"ZegFormer with pseudomask and language training\" or \"GroupVit with a MaskFormer as the grouping mechanism.\" Interestingly, our model relates to each method by omitting certain core architectural components or supervision method. Even so, our method is able to achieve competitive performance.", + "bbox": [ + 511, + 542, + 906, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 511, + 708, + 622, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed method, called S-Seg, is conceptually simple: we learn a MaskFormer model from pseudo-mask and language. Our method leverages image-text pairs solely, without relying on ground truth masks or large-scale pretrained models. In figure 5, we provide pseudocode for the core implementation of training S-Seg. Figure 4 provides a schematic layout of our approach.", + "bbox": [ + 511, + 733, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem definition", + "text_level": 1, + "bbox": [ + 511, + 848, + 692, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider the problem of open-vocabulary semantic segmentation, where we aim to learn a function $f$ that maps an", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30223", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image $I$ and a set of category names $\\mathcal{C} = \\{c_i\\}_{i = i}^{N_C}$ to a semantic segmentation map $S$ , where $c_{i}$ can be any category name expressed as open vocabulary texts.", + "bbox": [ + 89, + 89, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach adopts the unsupervised setting [4, 38, 47, 52], which aims to learn open-vocabulary segmentation from image-text pairs only, without learning from any dense annotation or class labels. Specifically, we use a web dataset of image-text pairs $\\mathcal{D} = \\{(I_i,T_i)\\}_{i = i}^{N_{\\mathcal{D}}}$ for training, where $T_{i}$ is a caption that describes the corresponding image $I_{i}$ . However, since the textual labels are gathered from the web, they may be noisy and contain errors. We do not use any additional manually annotated segmentation or classification labels during training.", + "bbox": [ + 89, + 137, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During testing, a set of category names $\\mathcal{C}$ is provided, and the model is tasked with assigning a semantic label $c_{i} \\in \\mathcal{C}$ to each pixel in an unlabeled image. The performance of the model is evaluated based on its mean Intersection over Union (mIoU) with the ground truth labels.", + "bbox": [ + 89, + 287, + 483, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adapting MaskFormer", + "text_level": 1, + "bbox": [ + 89, + 375, + 305, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach builds on top of MaskFormer [8]. Here, we begin by briefly review MaskFormer and explain the adjustments we made.", + "bbox": [ + 89, + 397, + 482, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Maskformer model takes an image as input and generates $N$ masks and mask features. First, the input image passes through a backbone model to produce feature maps at different output resolutions. These image features are then fed into a per-pixel encoder, which upsamples and aggregates them into a set of feature maps with higher resolution. Meanwhile, a transformer decoder uses $N$ learnable queries to cross-attend to the set of features with the lowest resolution and gather global information about each segment.", + "bbox": [ + 89, + 443, + 483, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the original Maskformer, a linear classifier and softmax activation were applied to the output of the decoder to predict class probabilities for a fixed list of categories. However, as we do not have a fixed list of categories, we remove this classifier branch and output the $N$ raw mask features instead.", + "bbox": [ + 89, + 594, + 483, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to predicting mask features, the Maskformer also predicts $N$ binary masks. To predict each mask, a dot product is taken between the mask embedding, generated from mask features, and the high resolution per-pixel feature. Finally, $N$ mask-feature pairs are combined to generate the output.", + "bbox": [ + 89, + 685, + 482, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. S-Seg", + "text_level": 1, + "bbox": [ + 89, + 786, + 171, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "S-Seg employs MaskFormer as its segmentation model, but in our weakly-supervised learning setting (where only texts are available), we face the challenge of not having annotated masks and labels. To overcome this, we utilize pseudo labels and language to as supervision.", + "bbox": [ + 89, + 809, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our training framework is illustrated in Figure 4. We first", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa5ed71c24a51ed37f63628f75ec0c2037ce28d0a2262483dae9acf82aa15c12.jpg", + "image_caption": [ + "Figure 6. Testing on S-Seg. During inference, S-Seg generalize to new categories by leveraging language features generated from a list of candidate classes in text." + ], + "image_footnote": [], + "bbox": [ + 526, + 89, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09e8812e7567f49aff372fd0a9f706aefb6a26a41a368434a184f8501d3b73f9.jpg", + "image_caption": [ + "Figure 7. Pseudo-mask generator generates pseudo-masks to supervise predicted mask during training. This module takes an image as its input, extracts its features using a DINO pre-trained ViT, and then employs K-means clustering to group the pixels into segments." + ], + "image_footnote": [], + "bbox": [ + 529, + 272, + 890, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generate a set of segmentation maps using our pseudo-mask generator (Sec. 3.4) and use them as supervision for mask prediction. Meanwhile, we use a language model to process input text and generate language embeddings. These embeddings provide supervision for mask features by leveraging image-text contrastive loss (Sec. 3.5).", + "bbox": [ + 511, + 429, + 903, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, unlike the supervised learning setting, where mask and label annotations are coupled, we decouple mask and semantic supervision. This enables us to utilize pseudomask and language as two distinct forms of supervision.", + "bbox": [ + 511, + 518, + 903, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the testing phase (as shown in figure 6), the trained MaskFormer model predicts $N$ masks and mask features from the input image. The language model takes as input a list of candidate category names (represented as texts) and extracts a set of language features. These features are then used to classify the mask features. This process is similar to the one used in CLIP [37], where the image and possible text inputs are encoded by their respective encoders to compute feature embeddings. The cosine similarity between these embeddings is calculated and adjusted by a learnable temperature parameter. The resulting values are normalized into a class probability distribution using a softmax function, and a combination module is used to takes $N$ mask-class pairs to produce the final segmentation map, similar to [8].", + "bbox": [ + 511, + 580, + 905, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we provide a details of the subcomponents in our framework.", + "bbox": [ + 511, + 806, + 903, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Pseudo-Mask Generator", + "text_level": 1, + "bbox": [ + 511, + 848, + 736, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our approach, we use a pseudo-mask generator (fig. 7) to produce a class-agnostic segmentation map from the input", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30224", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/99bed843566e1491365aec73ad4b7854aed94aa1d3e4161eb1414f3dc4196609.jpg", + "image_caption": [ + "Pseudo mask", + "Figure 8. Example pseudo-masks. Our pseudo-mask generator is capable of generating high-quality artificial masks. When provided with an oracle label, these masks demonstrate a high degree of overlap with the ground truth annotations." + ], + "image_footnote": [], + "bbox": [ + 107, + 103, + 468, + 233 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8d3ab9651217262ffd6d521b49c9a6c7f3408e27fd8db626f1da0c6c74759ab6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSup.P. VOC↑P. Context↑Time(s)↓
Spectral Clus. [41]*none49.243.20.543
K-Means [21]*none49.543.30.188
ImageNet [12]+[21]label68.858.10.079
GroupViT [47]text73.754.60.002
Pseudo-mask (Ours)self78.866.30.002
", + "bbox": [ + 94, + 301, + 478, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Our pseudo-mask generator achieves excellent oracle performance with rapid speed, making it an ideal mask supervision. We report amortised running time on a batch of 128 samples, simulating training time scenario. * We process downsampled image at $\\frac{H}{8} \\times \\frac{W}{8}$ resolution to obtain reasonable running time.", + "bbox": [ + 89, + 405, + 483, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image, which supervises the mask prediction of our model.", + "bbox": [ + 89, + 492, + 480, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To implement the pseudo-mask generator, we adopt a simple strategy that involves clustering tokens extracted from a self-supervised pre-trained ViT. Specifically, we use a DINO-pretrained ViT to compute a set of featurized tokens from the input image. We then apply a clustering algorithm (K-Means in our case) to these tokens, assigning each token a label that corresponds to the index of the cluster it belongs to. We reshape the resulting label map into an image and resize it to the original resolution to supervise the mask prediction of our segmentation model.", + "bbox": [ + 88, + 507, + 482, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite its simplicity, our pseudo-mask generator achieves both impressive performance, which is crucial for high-quality supervision, and fast processing speed, which is essential for efficient training. We evaluate its performance and compare against baseline methods, and the quantitative results are presented in Table 1, with example predictions visualized in 8. Our method significantly outperforms simple baselines such as K-Means and Spectral Clustering, which naively cluster image pixels, while running two orders of magnitude faster. We also observed that clustering DINO representation outperforms clustering ImageNet pre-trained ViT representation by a significant margin. Notably, our pseudo-mask generator even outperforms GroupViT, which has already employed vision-language training.", + "bbox": [ + 89, + 659, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the predicted masks are unordered, we need to", + "bbox": [ + 109, + 886, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "match the $N$ predicted masks with $K$ pseudo ground truth masks. To accomplish this, we utilize bipartite matching, as described in [2, 8], which assigns a pseudo-mask to each predicted mask such that the overall assignment cost is minimal in all possible assignments. Since each pseudo-mask is assigned to at most one predicted mask, $N - K$ pseudomasks are unassigned to no-object. Unlike standard training recipe [8], we do not penalize these no-object masks, nor do we use classification loss as an assignment cost. Finally, we compute the mask loss between predicted masks and their corresponding pseudo-mask, utilizing a combination of dice loss [34] and focal loss [29].", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\text {d i c e}} \\mathcal {L} _ {\\text {d i c e}} + \\lambda_ {\\text {f o c a l}} \\mathcal {L} _ {\\text {f o c a l}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 287, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Language Supervision", + "text_level": 1, + "bbox": [ + 511, + 311, + 720, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model learns to classify open-vocabulary concepts from language supervision. To train the model, we use an image-text contrastive loss [15, 37]. Specifically, we view $N$ mask features as representation of the input image, each capturing information about a different part of the image. We then compute a single feature that represents the entire image by taking the average of these mask features. To encode the text, we use a text transformer [42] and select the embedding corresponding to the [EOS] token, resulting in a textual feature. Since the visual and textual features may have different dimensions, we project each representation into a common embedding space using 2-layer MLPs. To compute the image-text contrastive loss, we calculate the cosine similarity between the image embeddings and the text embeddings within the same batch. Following common practice [26, 36, 37], we decouple the image-text contrastive loss into two parts:", + "bbox": [ + 511, + 333, + 906, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I \\rightarrow T} = - \\frac {1}{N} \\sum_ {i} ^ {N} \\log \\frac {\\exp \\left(x _ {i} ^ {\\intercal} y _ {i} / \\sigma\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(x _ {i} ^ {\\intercal} y _ {j} / \\sigma\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 611, + 905, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T \\rightarrow I} = - \\frac {1}{N} \\sum_ {i} ^ {N} \\log \\frac {\\exp \\left(y _ {i} ^ {\\intercal} x _ {i} / \\sigma\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(y _ {i} ^ {\\intercal} x _ {j} / \\sigma\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 662, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{i}$ and $y_{i}$ are L2-normalized embedding of image and text of the i-th pair. $N$ denotes batch size and $\\sigma$ is a learnable temperature parameter optimized together with the rest of the model. The total loss is the sum of these two losses, $\\mathcal{L}_{\\mathrm{contrastive}} = \\mathcal{L}_{I\\to T} + \\mathcal{L}_{T\\to I}$ .", + "bbox": [ + 511, + 709, + 906, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Training Loss", + "text_level": 1, + "bbox": [ + 511, + 792, + 656, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, mask loss (Sec. 3.4) and image-text contrastive loss (Sec. 3.5) complete the necessary mask and semantic supervision that is needed to train our model. The final loss is a weighted combination of the two losses:", + "bbox": [ + 511, + 814, + 905, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c o n t r a s t i v e}} \\mathcal {L} _ {\\text {c o n t r a s t i v e}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 885, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30225", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b463b9f7f82dd3bdca4137c9793207b02e5099766353ec239f90374432021be9.jpg", + "image_caption": [ + "Figure 9. Qualitative comparison with existing methods. CLIP [37] is primarily designed for classification and does not perform well in segmentation. MaskCLIP [52] adapts CLIP for segmentation, although it produces noisy predictions and cannot handle background classes. GroupViT [47] is a strong competitor, but it could struggle in challenging scenarios." + ], + "image_footnote": [], + "bbox": [ + 106, + 92, + 893, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiment, we use $\\lambda_{\\mathrm{mask}} = 1.0$ , $\\lambda_{\\mathrm{contrastive}} = 1.0$ , $\\lambda_{\\mathrm{dice}} = 1.0$ , $\\lambda_{\\mathrm{focal}} = 20.0$ .", + "bbox": [ + 89, + 449, + 483, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.7. Self-training", + "text_level": 1, + "bbox": [ + 89, + 489, + 225, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further improve our results, we introduce an optional step wherein we train a new model using the predictions generated by our current model. This process of self-training results in an augmented model, which we refer to as S-Seg+. More specifically, when we evaluate on a given dataset, we generate pseudo labels for the unlabeled images in the training set. Subsequently, we employ these pseudo labels to train a new segmentation model.", + "bbox": [ + 89, + 511, + 483, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-training improves the accuracy by leveraging additional data [46], augmentation [53], and bootstrapping [16]. In our situation, self-training offers even greater benefits since we can take advantage of additional information that is obtainable during testing: unlabeled images and testing categories. We show that this additional step improves our results significantly at no extra manual labelling cost.", + "bbox": [ + 89, + 632, + 483, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 752, + 223, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we empirically evaluate our method and compare to existing approaches. We show that, although our method is quite simple, it performs surprisingly well against more complex existing methods.", + "bbox": [ + 89, + 777, + 483, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 848, + 305, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture. Our experiments use MaskFormer [8] with a 6-layer transformer decoder and $N = 64$ queries. The", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hidden and output feature dimension is 256. The language model is a Transformer [42] with 12 layers, each with a hidden dimension of 256. We use a 2-layer MLP to project the visual and text feature into a common embedding space. We use DINO ViT-S/8 as the pretrained ViT in pseudo-mask generator.", + "bbox": [ + 511, + 449, + 906, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. During training, we used three publically available datasets: CC3M [40], CC12M [5], and RedCaps [10], containing 3M, 12M and 12M image-text pairs, respectively. Due to storage constraint, we use only first 11M data samples at a smaller resolution of $448 \\times 448$ when using RedCaps dataset. In total, we use at most 26M imagetext pairs for training - this is an order of magnitude fewer data than CLIP [37] and 1-4M fewer than GroupViT [47]. For more detailed training hyperparameters, please refer to the appendix.", + "bbox": [ + 511, + 554, + 908, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. We evaluate S-Seg on the validation set of three datasets: Pascal VOC 2012 [14], Pascal Context [35] and COCO [28]. The Pascal VOC dataset contains 1449 images for testing. Each image is labeled with 20 foreground classes and a background class. The Pascal Context dataset contains 5104 testing images with 59 foreground classes and a background class. The COCO dataset contains 5000 images for testing with 80 foreground classes and an additional background class. Following GroupViT [47], we threshold the maximum probability to obtain background prediction. During inference, we set the input resolution to $448 \\times 448$ , in consistent with [47].", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30226", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4fcad3b064431adf218ff7441212e3501181f791d10e03a52c13df06ac421296.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Linearly-probed classification models:
MoCo v3 [7]✗----34.321.3-
DINO [3]✗----39.120.4-
Open-vocabulary models:
CLIP [37]†✓✗✓✗✗13.58.15.9
MaskCLIP [52]†✓✗✓✗✗26.822.812.8
ViL-Seg [30]✓✗✓✗✗34.416.316.4
CLIPpy [38]✓✗✓✗✗52.2--
GroupViT [47]✓✗✗✓✗50.823.727.5
SegCLIP [32]✓✗✓✗✗52.624.726.5
OVSegmentor [48]✓✗✗✓✗53.820.425.1
TCL [4]✓✗✓✓✗55.030.431.6
SAM-CLIP [43]✓✓✓✗✓60.629.231.5
S-Seg (Ours)✓✗✗✗✗53.227.930.3
S-Seg+ (Ours)✓✗✗✗✗62.030.235.7
Fully-supervised segmentation models:
DeepLabV3+† [6]✗----78.746.455.7
MaskFormer† [8]✗----81.250.062.1
", + "bbox": [ + 91, + 88, + 486, + 344 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c254016d5bd2887f7cc2c08ae8c49cf8e5c23ea7748a1abeb2bdec0c89694f77.jpg", + "table_caption": [ + "Table 2. Open-vocabulary (OV) semantic segmentation results (background pixels included in evaluation). Benchmarked following standard evaluation protocols for unsupervised open-vocabulary models trained without annotated masks [4, 32, 47, 48]. Models labels: Mask (use GT mask), VLM (employ pretrained large VL model like CLIP), Enc. (require specialized grouping encoder), and SAM (use the SAM model). Our approach obtain second highest performance on average, without using any extra components such as CLIP and SAM. $\\dagger$ denotes our recomputed results. Higher values are better." + ], + "table_footnote": [], + "table_body": "
MethodP. VOCP. ContextCOCO3-Avg.
B1: Pseudo Mask + CLIP12.93.92.96.6
B2: Pseudo-mask ViT23.211.010.414.9
S-Seg (Ours)44.922.922.530.1
", + "bbox": [ + 91, + 503, + 478, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Simple baselines", + "text_level": 1, + "bbox": [ + 89, + 635, + 253, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The high quality of pseudo-masks (as shown in Figure 7) may lead one to assume that the primary challenge is simply classifying these masks, and that this can be accomplished by utilizing pre-existing methods such as CLIP. To test this assumption, we first develop two simple baselines.", + "bbox": [ + 89, + 657, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline 1: Pseudo-mask + CLIP. Firstly, our pseudo label generator is utilized to obtain pseudo segments. Then, we iterate through all the masks and apply the current mask to the original image. Next, the masked image is fed to CLIP for classification and the resulting class label is assigned to the corresponding segment.", + "bbox": [ + 89, + 734, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline 2: Pseudo-mask ViT. We introduce a new visual backbone that differs from the regular ViT. Instead of pooling all image tokens into a single feature, we first individually pool tokens in each segment of the pseudo-mask into segment features, and then pool these features into a vi", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9af991822afaa93e5da4c50f63c0062b651f5b6a47bba07212935a4e1f03f7d3.jpg", + "table_caption": [ + "Table 3. Simple baselines for open-vocabulary semantic segmentation. All models are trained on CC12M. Higher values are better. Two simple baselines fail to obtain satisfactory results, even using after using our pseudo masks and no less training data." + ], + "table_footnote": [], + "table_body": "
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Open-vocabulary models (annotated masks required for training):
ZS3Net [1]✓✗✗✗38.319.421.1
LSeg [24]✓✓✗✗52.3-27.2
OpenSeg [15]✓✓✗✗77.245.938.1
ZegFormer [11]✓✓✗✗80.7--
GKC [19]✓✓✗✗83.245.2-
ODISE [49]✓✓✓✗84.657.365.2*
DeOp [18]✓✓✗✗91.748.8-
OVSeg [27]✓✓✗✗94.555.7-
SAN [33]✓✓✓✗94.657.7-
FC-CLIP [50]✓✓✓✗95.458.4-
SED [45]✓✓✓✗96.160.6-
CAT-Seg [9]✓✓✓✗96.662.0-
Open-vocabulary models (annotated masks not required for training):
CLIP [37]†✓✗✓✗✗39.69.013.8
MaskCLIP [52]†✓✗✓✗✗49.525.523.6
GroupViT [47]†✓✗✗✓✗77.223.037.5
TCL [4]✓✗✓✓✗83.233.9-
S-Seg (Ours)✓✗✗✗✗81.827.242.4
S-Seg+ (Ours)✓✗✗✗✗84.731.653.0
Fully-supervised segmentation models:
DeepLabV3+† [6]✗GT---89.948.566.9
", + "bbox": [ + 514, + 88, + 903, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Open-vocabulary (OV) semantic segmentation results (background pixels excluded in evaluation). Benchmarked following standard protocol for evaluating supervised open-vocabulary models trained with annotated masks [11, 18, 19, 27, 33]. Similar to the previous setting, S-Seg achieves competitive performance compared to earlier methods trained without using any extra components such as CLIP and SAM. *COCO is used for training. Higher values are better.", + "bbox": [ + 511, + 402, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sual embedding. We train a CLIP-like model from scratch using this visual backbone. During testing, we classify each segment feature and assign the label to that segment.", + "bbox": [ + 511, + 531, + 905, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are presented in Table 3. As we can see, open-vocabulary segmentation is more complex than simply grouping image into segments and then categorizing them into classes, even when the segments are of high quality. Baseline 1 employs a significantly larger pretrained CLIP ViT/L-14 model that was also trained on a much larger dataset, while Baseline 2 is trained using the same data as ours. Nevertheless, both baselines fail to achieve satisfactory results, suggesting that open-vocabulary segmentation cannot be naively deconstructed in such ways. We hypothesize that a multi-task learning approach that jointly trains for segmentation and classification could yield significant advantages.", + "bbox": [ + 511, + 578, + 906, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation with background", + "text_level": 1, + "bbox": [ + 511, + 787, + 769, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In table 2, we evaluate our model and compare with existing method on open-vocabulary semantic segmentation task. Following standard evaluation protocols on open-vocabulary model trained without annotated masks [4, 32, 47, 48], we include background pixels in evaluation and obtain background prediction by setting a threshold for back", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30227", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b721ac53106a6ec030767e0a2492f8f379b321f43bc4e9cdf9d3891a60643ae4.jpg", + "image_caption": [ + "(a) Pascal VOC $(+18.3\\%)$" + ], + "image_footnote": [], + "bbox": [ + 120, + 114, + 370, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e734d2df69c37061f0c9e39ee04bd0a1c97ab62a32ab726f1bb095af0401aaf.jpg", + "image_caption": [ + "(b) Pascal Context $(+14.1\\%)$", + "Figure 10. Self-training improvement. We show average relative improvement in bracket on top of the plot. we observe that self-training consistently leads to significant improvement for S-Seg across all of our training and testing data settings." + ], + "image_footnote": [], + "bbox": [ + 372, + 114, + 620, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee1f9647d5223e15acd714d2948429a03e3fbbadf49143eb7e0c2e6f79cd2dc4.jpg", + "image_caption": [ + "(c) COCO $(+17.6\\%)$" + ], + "image_footnote": [], + "bbox": [ + 622, + 114, + 870, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c486708a1bcb6efd0e4d3f3136b9afedbfc396dd8e0dfc004d89c6342fad0e9.jpg", + "image_caption": [ + "S-Seg (details)", + "Figure 11. Visualizing effect of self-training. Our self-trained S-Seg+ model demonstrates the ability to accurately predict in regions overlooked by S-Seg, as shown in the colorful rectangles." + ], + "image_footnote": [], + "bbox": [ + 96, + 290, + 194, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ce6363644a1f6822ca7e8136f1ed7266e650839167e1a901f26cd6476c9a0f0.jpg", + "image_caption": [ + "S-Seg+ (details)" + ], + "image_footnote": [], + "bbox": [ + 196, + 291, + 292, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f03368b0991e209d40baec368fce736dfc4eec87514ba2da4a9505d7cb01590.jpg", + "image_caption": [ + "S-Seg (details)" + ], + "image_footnote": [], + "bbox": [ + 294, + 291, + 383, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f7be99205e5a2faaaf34a6852c2a9e528701758f4ac44753b33bbf3cb7d4bab.jpg", + "image_caption": [ + "S-Seg+ (details)" + ], + "image_footnote": [], + "bbox": [ + 388, + 291, + 480, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ground classes [47]. Despite the simplicity of S-Seg, our approach achieve competitive performance over previous open-vocabulary segmentation methods that does not require mask annotations. Our model has second highest performance on average and has better results than GroupViT on all datasets. Moreover, our self-trained model, S-Seg+, provides an impressive $5.5\\%$ mIoU improvement over our base model S-Seg (42.6% vs 37.1% 3-avg. mIoU), suggesting the efficacy of self-training.", + "bbox": [ + 89, + 426, + 483, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Evaluation without background", + "text_level": 1, + "bbox": [ + 89, + 573, + 370, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also evaluate our model on the evaluation protocol commonly used for evaluating open-vocabulary models with annotated masks [11, 18, 19], where the background pixels are excluded in evaluation. We note that this setting is easier because background class is more diverse in appearance and often requires additional processing such as thresholding. Table 4 shows the results. Similar to the previous setting, our S-Seg and S-Seg+ models achieve competitive performance compared to earlier methods.", + "bbox": [ + 89, + 595, + 483, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 742, + 250, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-training. We investigated the effectiveness of self-training for improving segmentation performance. To this end, we compared S-Seg and S-Seg+ on three datasets and evaluated the results using Figure 10. We found that self-training consistently improved the segmentation performance by a significant margin $(+5.5\\%)$ mIoU on average), regardless of the data size and test dataset. These results indicate that self-training is a reliable approach for enhancing the performance of S-Seg and can provide a de", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af176d28d05d8e0752f64c9d368d072671bcb27600d65eac8fc42210b288303f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 300, + 625, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/741b4de68b2938b0483b1ce2b153437fe5e017fd3acc6073b0eda4251538cb54.jpg", + "image_caption": [ + "(a) S-Seg" + ], + "image_footnote": [], + "bbox": [ + 625, + 300, + 715, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f2f1e9acf4fd12ae534eaecef9a0d4c10d261590773ce851440e933c7350a2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 300, + 802, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd2ece15d355ac639e64cc308f6fb2a0252fdad89f007022fc12d2c702857cee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 300, + 893, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e67a20d3295c34f1c94c2da4ad986ae53d7cba63160f48edbb119e9a7f017e17.jpg", + "image_caption": [ + "Figure 12. Scaling training data provide consistent gain in performance, with or without self-training. We train our model using different sizes of data: CC12M (12M), CC12M+CC3M (15M), and CC12M+CC3M+RedCaps (26M). We note a steady improvement in the performance as the data size increases." + ], + "image_footnote": [], + "bbox": [ + 521, + 416, + 625, + 510 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0254ff69ca81fa9847903f84bb286c9bf9754b2ac08670b2dcbbfa39d9c5af77.jpg", + "image_caption": [ + "(b) S-Seg+", + "Size of training data $(\\times 10^{6})$" + ], + "image_footnote": [], + "bbox": [ + 625, + 417, + 712, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62725b948e8f5fe8c6384d84f2f2233260535b92e14d179bc2b1414011bc6f6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 417, + 800, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54865075f1660a72c082a35ca530a6dee88b40b424f72a1877bb53b9c974b31a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 417, + 888, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sirable complement for further improvement.", + "bbox": [ + 511, + 614, + 812, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data scalability. To evaluate the scalability of our method, we trained S-Seg and S-Seg+ using three datasets of increasing sizes: 12M, 15M, and 26M. The results of the experiments are presented in Figure 12. We observed that both models achieve significant improvements in performance across all three testing datasets as the amount of data increased, suggesting that our method scales well with larger datasets.", + "bbox": [ + 511, + 630, + 905, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 763, + 632, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To summarize, we propose S-Seg, a simple and intuitive framework that enables accurate and generalizable open-vocabulary segmentation. Our algorithm directly trains for pixel-level feature and language alignment, and does not require manual segmentation annotations or extensive pretraining. We hope that our simple yet effective approach will serve as a solid baseline for future research.", + "bbox": [ + 511, + 790, + 905, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30228", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 7", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 7", + "[4] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 2, 4, 7", + "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 6", + "[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7", + "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 7", + "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. 2021. 4, 5, 6, 7", + "[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 7", + "[10] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431, 2021.6", + "[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 3, 7, 8", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 5", + "[13] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, 2002. 2", + "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2009. 6", + "[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3, 5, 7", + "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 6", + "[17] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In ECCV, 2020. 3", + "[18] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2, 3, 7, 8", + "[19] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, and Yunchao Wei. Global knowledge calibration for fast open-vocabulary segmentation. In ICCV, 2023. 2, 3, 7, 8", + "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. 2021. 2", + "[21] Tapas Kanungo, David M Mount, Nathan S Netanyahu, Christine D Piatko, Ruth Silverman, and Angela Y Wu. An efficient k-means clustering algorithm: Analysis and implementation. IEEE TPAMI, 2002. 5", + "[22] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316, 2023. 2", + "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3", + "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2, 3, 7", + "[25] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 3", + "[26] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. 2023. 5", + "[27] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 2, 3, 7", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6", + "[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5", + "[30] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In ECCV, 2022. 7", + "[31] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30229", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 3", + "[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. 2023. 2, 7", + "[33] Fangyun Wei Han Hu Xiang Bai Mengde Xu, Zheng Zhang. Side adapter network for open-vocabulary semantic segmentation. 2023. 2, 3, 7", + "[34] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 5", + "[35] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 6", + "[36] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 5", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 1, 2, 4, 5, 6, 7", + "[38] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. In ICCV, 2023. 2, 4, 7", + "[39] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3", + "[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2018, 2, 6", + "[41] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 5, 6", + "[43] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 7", + "[44] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In ICCV, 2023. 2, 3", + "[45] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 6", + "[47] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 1, 3, 4, 5, 6, 7, 8", + "[48] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 1, 2, 3, 7", + "[49] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2, 7", + "[50] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[51] Haobo Yuan, Xiangtai Li, Chong Zhou, Yining Li, Kai Chen, and Chen Change Loy. Open-vocabulary sam: Segment and recognize twenty-thousand classes interactively. arXiv preprint arXiv:2401.02955, 2024. 3", + "[52] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 2, 4, 6, 7", + "[53] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. *NeurIPS*, 2020. 6" + ], + "bbox": [ + 516, + 92, + 903, + 488 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30230", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_model.json b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..31160fc22ba0fb92644db9d8781951145b8d6fbb --- /dev/null +++ b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.802, + 0.154 + ], + "angle": 0, + "content": "Exploring Simple Open-Vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.182, + 0.712, + 0.233 + ], + "angle": 0, + "content": "Zihang Lai Visual Geometry Group (VGG), University of Oxford zlai@robots.ox.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.604 + ], + "angle": 0, + "content": "Open-vocabulary semantic segmentation models aim to accurately assign a semantic label to each pixel in an image from a set of arbitrary open-vocabulary texts. In order to learn such pixel-level alignment, current approaches typically rely on a combination of (i) image-level VL model (e.g. CLIP), (ii) ground truth masks, (iii) custom grouping encoders, and (iv) the Segment Anything Model (SAM). In this paper, we introduce S-Seg, a simple model that can achieve surprisingly strong performance without depending on any of the above elements. S-Seg leverages pseudo-masks and language features to train a MaskFormer, and can be easily trained from publicly available image-text datasets. Contrary to prior works, our model directly trains for pixel-level features and language alignment. Once trained, S-Seg generalizes well to multiple testing datasets without requiring fine-tuning. In addition, S-Seg has the extra benefits of scalability with data and consistently improving when augmented with self-training. We believe that our simple yet effective approach will serve as a solid baseline for future research. Project page: zlai0.github.io/S-Seg." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.633, + 0.222, + 0.648 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Open-vocabulary semantic segmentation presents a unique challenge as it requires assigning accurate semantic labels to each pixel in an image using arbitrary open-vocabulary texts, rather than a fixed set of classes. This means that the model must be able to segment and classify any arbitrary categories expressed in language. Achieving this requires a robust, pixel-level alignment between images and textual descriptions, which enables accurate association of each pixel with the most relevant class from a dynamically provided set of textual categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "A primary obstacle in this domain is that it is impossible to construct datasets that provide pixel-level annotations for all possible labels. This limitation often results in the adoption of weakly-supervised or semi-supervised learning approaches. Current methods typically rely on a combination of strategies to learn the required pixel-level alignment." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.269, + 0.907, + 0.297 + ], + "angle": 0, + "content": "Figure 1. S-Seg result on a web image. Our goal is to segment everything, including fictional characters like minions." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.313, + 0.907, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.506, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Figure 2. Our S-Seg framework leverages pseudo-mask and language to train a MaskFormer. We show that our method of directly training for pixel-level feature and language alignment yields superior results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.717 + ], + "angle": 0, + "content": "One common tactic is adapting existing Vision-Language (VL) models, which are initially trained for image-level alignment (e.g., CLIP [37]), to perform at the pixel level. Another strategy involves training models on ground truth masks that are annotated for a select number of seen classes, thereby encouraging the model to extrapolate its knowledge to novel unseen classes during testing. Furthermore, specialized models such as GroupViT [47] and OVSegmentor [48], which are explicitly designed for open-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.779 + ], + "angle": 0, + "content": "vocabulary segmentation, are being explored. Lastly, recent literature adapts the Segment Anything Model (SAM), originally not designed to classify segments, for effective use in open-vocabulary segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.884 + ], + "angle": 0, + "content": "In this paper, we report a model that can work surprisingly well with none of the above strategies. Our approach, named S-Seg, is built on top of a standard MaskFormer model. Our model directly trains for pixel-level feature and language alignment, using neither existing large image-level alignment models like CLIP [37] nor manually annotated segmentation or classification labels." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "One of the biggest challenges we face is finding the right" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30221" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.309, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of S-Seg, evaluated using all dataset classes as queries. Our model copes with challenging situation, such as overlapping objects (col. 2) and small objects (col. 5). Our model is also capable of handling \"stuff\" categories such as water and floor (col. 3, 4). Moreover, our S-Seg+ model is able to correct small errors observed in the S-Seg method (col. 4). Finally, in the COCO dataset, which featured a significantly higher number of objects, our model is still able to achieve high accuracy in its predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.482, + 0.652 + ], + "angle": 0, + "content": "supervision since annotated masks and labels are not available. To address this issue, we propose to leverage pseudomasks and language to supervise MaskFormer. Our strategy involves using a pseudo-mask generator to provide class-agnostic mask supervision by generating pseudo ground truth masks. We adopt a simple design that clusters image representations obtained through self-supervised representation learning methods like DINO [3]. Our experiments demonstrate that this approach delivers exceptional performance, which is essential for high-quality supervision, as well as rapid processing speed, which is necessary for efficient training. In addition, we use noisy web texts to provide semantic supervision. The image-text dataset contains a wide range of concepts and has demonstrated impressive zero-shot classification results [37]. We utilize a straightforward image-text contrastive loss, which has proven to be highly effective. Once trained, our model generalizes well to new categories without requiring fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.791 + ], + "angle": 0, + "content": "S-Seg is a simple and effective model that can be trained using publicly available image-text datasets, such as Conceptual Captions [5, 40]. This makes it easy to reproduce and extend for further research. The S-Seg framework is also designed to be flexible with easily replaceable submodules. We prioritize simplicity in our subcomponent selection to focus on the general design of our framework, while remaining open to more advanced techniques that could result in further improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conducted a thorough evaluation of S-Seg using multiple benchmark datasets, and we show that our method achieve competitive results on three widely tested benchmarks (Pascal VOC, Pascal Context, and COCO). In addition, pseudo-mask and language provide scalable supervision and our model consistently improves in performance as more data became available. Finally, we find adding an ad" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.439 + ], + "angle": 0, + "content": "ditional self-training step leads to an even greater improvement to our model, with an average increase of \\(5.5\\%\\) mIoU over three datasets, highlighting the potential for further improvement of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.906, + 0.529 + ], + "angle": 0, + "content": "Our simple solution suggests that the reliance on complex models and extensive ground truth data in open-vocabulary semantic segmentation may be reduced, leading to more streamlined and accessible framework for future developments in the field, and we hope our exploration can serve as a solid baseline for future research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.652, + 0.558 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Open-vocabulary segmentation. The earliest efforts to employ language for image segmentation can be traced back to Duygulu et al.'s seminal work [13], where the authors tackled image segmentation by framing it as a machine translation problem. Current approaches leverage a combination of strategies to learn pixel-level image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Adapting image-level vision-language models. The first strategy involves the adapting pretrained vision-language models, originally designed for image-level alignment, to the more granular task of pixel-level alignment. This strategy is widely adopted in open-vocabulary methods [4, 15, 18, 19, 22, 24, 27, 32, 33, 38, 44, 48, 49, 52]. These works vary in their methods of refining image-level models for finer alignment tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "MaskCLIP [52] demonstrates modifying the CLIP image encoder can significantly enhance its pixel-level alignment capabilities without requiring retraining. TCL [4] employs CLIP for initial text-to-image region grounding, followed by contrastive learning to refine the alignment between the text embedding and the grounded region. OpenSeg [15] fine-tunes ALIGN [20] using a grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30222" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.074, + 0.852, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.216, + 0.908, + 0.259 + ], + "angle": 0, + "content": "Figure 4. Overview of S-Seg. A MaskFormer model computes masks and mask features from an image input. A pseudo-mask generator produces segmentation maps to supervise mask predictions, while a text that describes the image, encoded by a language model trained together with the MaskFormer, provides supervision for mask features using image-text contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.272, + 0.482, + 0.408 + ], + "angle": 0, + "content": "loss [17] to better align words in captions to segmentation masks. OpenSeg and DiffuMask [44] also explored the use of pseudo-masks. The primary distinction lies in their dependency on different sources for learning; OpenSeg uses annotated segments while DiffuMask employs masks generated through diffusion. In contrast, our method is entirely learned from pseudo masks. Also, our mask generator is entirely self-supervised, whereas their mask generator is fully-supervised." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.501 + ], + "angle": 0, + "content": "Ground truth masks. Another effective strategy [11, 15, 18, 19, 24, 27, 33] involves training models using ground truth masks annotated for a limited set of seen classes. By training on seen annotations, models are encouraged to learn detailed features and patterns that are potentially applicable beyond the scope of the trained classes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Of most relevance, ZegFormer [11] trains a MaskFormer by decoupling zero-shot semantic segmentation into two sub-tasks, a class-agnostic grouping task and a zero-shot segment classification task. Our method has similar training paradigm but with notable distinctions. Similar to GroupViT, we train exclusively with image-text pairs and do not utilize a pretrained CLIP model. Notably, even without access to ground truth masks, labels, or CLIP, our method outperforms ZegFormer in unseen categories, indicating potentially stronger generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Custom grouping-based encoders. The third strategy employs custom-designed models specifically for open-vocabulary segmentation. GroupViT [47] groups pixels in an image hierarchically based on their attention scores with learnable group tokens. OVSegmentor [48] applies Slot Attention [31] for a similar pixel grouping process based on feature proximity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Segment Anything Model (SAM). The final strategy adapts the Segment Anything Model (SAM) [23], a powerful segmentation model trained on over 11 million images and 1.1 billion masks, for open-vocabulary segmentation. While the SAM model segments object with high accuracy, it does not effectively classify these segments. To address this limitation, follow-up works has adapted the SAM model for grounding, closed-set, and open-vocabulary segmentation [25, 39, 51] by leveraging additional training" + }, + { + "type": "code", + "bbox": [ + 0.527, + 0.274, + 0.892, + 0.489 + ], + "angle": 0, + "content": "I [n, h, w, c] - minibatch of aligned images \nT [n, l] - minibatch of aligned texts \nN - number of MaskFormer queries \nC - number of pseudo masks \n# predict mask, mask feature, and text feature \nM, M_f = maskformer(I) # [n, N, H, W], [n, N, d_f] \nT_f = textEncoder(T) # [n, d_f] \n# aggregate all mask features [n, d_f] \nM_f = M_f.mean(axis=1) \n# generate pseudo masks [n, C, H, W] \nS = pseudo_mask_generation(I) \n# compute loss \nloss_c = contrastive_loss(M_f, T_f) \nloss_m = mask_loss(M, S) \nloss = (loss_c + loss_m)/2" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.501, + 0.905, + 0.516 + ], + "angle": 0, + "content": "Figure 5. Pseudocode for training S-Seg with image-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.719, + 0.543 + ], + "angle": 0, + "content": "data, adaptors, and distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.696 + ], + "angle": 0, + "content": "Our model, S-Seg, can be conceptualized as a synergy of these approaches. It can be viewed as a CLIP model integrated with a MaskFormer image encoder, directly optimizing for pixel-level feature and language alignment. Alternatively, it resembles \"ZegFormer with pseudomask and language training\" or \"GroupVit with a MaskFormer as the grouping mechanism.\" Interestingly, our model relates to each method by omitting certain core architectural components or supervision method. Even so, our method is able to achieve competitive performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.623, + 0.726 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Our proposed method, called S-Seg, is conceptually simple: we learn a MaskFormer model from pseudo-mask and language. Our method leverages image-text pairs solely, without relying on ground truth masks or large-scale pretrained models. In figure 5, we provide pseudocode for the core implementation of training S-Seg. Figure 4 provides a schematic layout of our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.694, + 0.863 + ], + "angle": 0, + "content": "3.1. Problem definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We consider the problem of open-vocabulary semantic segmentation, where we aim to learn a function \\( f \\) that maps an" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30223" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.137 + ], + "angle": 0, + "content": "image \\(I\\) and a set of category names \\(\\mathcal{C} = \\{c_i\\}_{i = i}^{N_C}\\) to a semantic segmentation map \\(S\\), where \\(c_{i}\\) can be any category name expressed as open vocabulary texts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.288 + ], + "angle": 0, + "content": "Our approach adopts the unsupervised setting [4, 38, 47, 52], which aims to learn open-vocabulary segmentation from image-text pairs only, without learning from any dense annotation or class labels. Specifically, we use a web dataset of image-text pairs \\(\\mathcal{D} = \\{(I_i,T_i)\\}_{i = i}^{N_{\\mathcal{D}}}\\) for training, where \\(T_{i}\\) is a caption that describes the corresponding image \\(I_{i}\\). However, since the textual labels are gathered from the web, they may be noisy and contain errors. We do not use any additional manually annotated segmentation or classification labels during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.364 + ], + "angle": 0, + "content": "During testing, a set of category names \\(\\mathcal{C}\\) is provided, and the model is tasked with assigning a semantic label \\(c_{i} \\in \\mathcal{C}\\) to each pixel in an unlabeled image. The performance of the model is evaluated based on its mean Intersection over Union (mIoU) with the ground truth labels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.307, + 0.391 + ], + "angle": 0, + "content": "3.2. Adapting MaskFormer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Our approach builds on top of MaskFormer [8]. Here, we begin by briefly review MaskFormer and explain the adjustments we made." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.594 + ], + "angle": 0, + "content": "The Maskformer model takes an image as input and generates \\(N\\) masks and mask features. First, the input image passes through a backbone model to produce feature maps at different output resolutions. These image features are then fed into a per-pixel encoder, which upsamples and aggregates them into a set of feature maps with higher resolution. Meanwhile, a transformer decoder uses \\(N\\) learnable queries to cross-attend to the set of features with the lowest resolution and gather global information about each segment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.684 + ], + "angle": 0, + "content": "In the original Maskformer, a linear classifier and softmax activation were applied to the output of the decoder to predict class probabilities for a fixed list of categories. However, as we do not have a fixed list of categories, we remove this classifier branch and output the \\(N\\) raw mask features instead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.777 + ], + "angle": 0, + "content": "In addition to predicting mask features, the Maskformer also predicts \\(N\\) binary masks. To predict each mask, a dot product is taken between the mask embedding, generated from mask features, and the high resolution per-pixel feature. Finally, \\(N\\) mask-feature pairs are combined to generate the output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.172, + 0.804 + ], + "angle": 0, + "content": "3.3. S-Seg" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.885 + ], + "angle": 0, + "content": "S-Seg employs MaskFormer as its segmentation model, but in our weakly-supervised learning setting (where only texts are available), we face the challenge of not having annotated masks and labels. To overcome this, we utilize pseudo labels and language to as supervision." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our training framework is illustrated in Figure 4. We first" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.09, + 0.891, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.226, + 0.905, + 0.268 + ], + "angle": 0, + "content": "Figure 6. Testing on S-Seg. During inference, S-Seg generalize to new categories by leveraging language features generated from a list of candidate classes in text." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.273, + 0.891, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.345, + 0.905, + 0.413 + ], + "angle": 0, + "content": "Figure 7. Pseudo-mask generator generates pseudo-masks to supervise predicted mask during training. This module takes an image as its input, extracts its features using a DINO pre-trained ViT, and then employs K-means clustering to group the pixels into segments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.519 + ], + "angle": 0, + "content": "generate a set of segmentation maps using our pseudo-mask generator (Sec. 3.4) and use them as supervision for mask prediction. Meanwhile, we use a language model to process input text and generate language embeddings. These embeddings provide supervision for mask features by leveraging image-text contrastive loss (Sec. 3.5)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.905, + 0.58 + ], + "angle": 0, + "content": "Notably, unlike the supervised learning setting, where mask and label annotations are coupled, we decouple mask and semantic supervision. This enables us to utilize pseudomask and language as two distinct forms of supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.806 + ], + "angle": 0, + "content": "In the testing phase (as shown in figure 6), the trained MaskFormer model predicts \\( N \\) masks and mask features from the input image. The language model takes as input a list of candidate category names (represented as texts) and extracts a set of language features. These features are then used to classify the mask features. This process is similar to the one used in CLIP [37], where the image and possible text inputs are encoded by their respective encoders to compute feature embeddings. The cosine similarity between these embeddings is calculated and adjusted by a learnable temperature parameter. The resulting values are normalized into a class probability distribution using a softmax function, and a combination module is used to takes \\( N \\) mask-class pairs to produce the final segmentation map, similar to [8]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.808, + 0.905, + 0.837 + ], + "angle": 0, + "content": "Next, we provide a details of the subcomponents in our framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.738, + 0.863 + ], + "angle": 0, + "content": "3.4. Pseudo-Mask Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In our approach, we use a pseudo-mask generator (fig. 7) to produce a class-agnostic segmentation map from the input" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30224" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.092, + 0.255, + 0.103 + ], + "angle": 0, + "content": "Pseudo mask" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.104, + 0.47, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.483, + 0.301 + ], + "angle": 0, + "content": "Figure 8. Example pseudo-masks. Our pseudo-mask generator is capable of generating high-quality artificial masks. When provided with an oracle label, these masks demonstrate a high degree of overlap with the ground truth annotations." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.303, + 0.479, + 0.392 + ], + "angle": 0, + "content": "
MethodSup.P. VOC↑P. Context↑Time(s)↓
Spectral Clus. [41]*none49.243.20.543
K-Means [21]*none49.543.30.188
ImageNet [12]+[21]label68.858.10.079
GroupViT [47]text73.754.60.002
Pseudo-mask (Ours)self78.866.30.002
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.477 + ], + "angle": 0, + "content": "Table 1. Our pseudo-mask generator achieves excellent oracle performance with rapid speed, making it an ideal mask supervision. We report amortised running time on a batch of 128 samples, simulating training time scenario. * We process downsampled image at \\( \\frac{H}{8} \\times \\frac{W}{8} \\) resolution to obtain reasonable running time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.481, + 0.508 + ], + "angle": 0, + "content": "image, which supervises the mask prediction of our model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.658 + ], + "angle": 0, + "content": "To implement the pseudo-mask generator, we adopt a simple strategy that involves clustering tokens extracted from a self-supervised pre-trained ViT. Specifically, we use a DINO-pretrained ViT to compute a set of featurized tokens from the input image. We then apply a clustering algorithm (K-Means in our case) to these tokens, assigning each token a label that corresponds to the index of the cluster it belongs to. We reshape the resulting label map into an image and resize it to the original resolution to supervise the mask prediction of our segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Despite its simplicity, our pseudo-mask generator achieves both impressive performance, which is crucial for high-quality supervision, and fast processing speed, which is essential for efficient training. We evaluate its performance and compare against baseline methods, and the quantitative results are presented in Table 1, with example predictions visualized in 8. Our method significantly outperforms simple baselines such as K-Means and Spectral Clustering, which naively cluster image pixels, while running two orders of magnitude faster. We also observed that clustering DINO representation outperforms clustering ImageNet pre-trained ViT representation by a significant margin. Notably, our pseudo-mask generator even outperforms GroupViT, which has already employed vision-language training." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Since the predicted masks are unordered, we need to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "match the \\(N\\) predicted masks with \\(K\\) pseudo ground truth masks. To accomplish this, we utilize bipartite matching, as described in [2, 8], which assigns a pseudo-mask to each predicted mask such that the overall assignment cost is minimal in all possible assignments. Since each pseudo-mask is assigned to at most one predicted mask, \\(N - K\\) pseudomasks are unassigned to no-object. Unlike standard training recipe [8], we do not penalize these no-object masks, nor do we use classification loss as an assignment cost. Finally, we compute the mask loss between predicted masks and their corresponding pseudo-mask, utilizing a combination of dice loss [34] and focal loss [29]." + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.288, + 0.907, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\text {d i c e}} \\mathcal {L} _ {\\text {d i c e}} + \\lambda_ {\\text {f o c a l}} \\mathcal {L} _ {\\text {f o c a l}} \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.312, + 0.722, + 0.329 + ], + "angle": 0, + "content": "3.5. Language Supervision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Our model learns to classify open-vocabulary concepts from language supervision. To train the model, we use an image-text contrastive loss [15, 37]. Specifically, we view \\(N\\) mask features as representation of the input image, each capturing information about a different part of the image. We then compute a single feature that represents the entire image by taking the average of these mask features. To encode the text, we use a text transformer [42] and select the embedding corresponding to the [EOS] token, resulting in a textual feature. Since the visual and textual features may have different dimensions, we project each representation into a common embedding space using 2-layer MLPs. To compute the image-text contrastive loss, we calculate the cosine similarity between the image embeddings and the text embeddings within the same batch. Following common practice [26, 36, 37], we decouple the image-text contrastive loss into two parts:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.612, + 0.906, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I \\rightarrow T} = - \\frac {1}{N} \\sum_ {i} ^ {N} \\log \\frac {\\exp \\left(x _ {i} ^ {\\intercal} y _ {i} / \\sigma\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(x _ {i} ^ {\\intercal} y _ {j} / \\sigma\\right)} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.664, + 0.905, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T \\rightarrow I} = - \\frac {1}{N} \\sum_ {i} ^ {N} \\log \\frac {\\exp \\left(y _ {i} ^ {\\intercal} x _ {i} / \\sigma\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(y _ {i} ^ {\\intercal} x _ {j} / \\sigma\\right)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.907, + 0.786 + ], + "angle": 0, + "content": "where \\( x_{i} \\) and \\( y_{i} \\) are L2-normalized embedding of image and text of the i-th pair. \\( N \\) denotes batch size and \\( \\sigma \\) is a learnable temperature parameter optimized together with the rest of the model. The total loss is the sum of these two losses, \\( \\mathcal{L}_{\\mathrm{contrastive}} = \\mathcal{L}_{I\\to T} + \\mathcal{L}_{T\\to I} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.658, + 0.809 + ], + "angle": 0, + "content": "3.6. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.906, + 0.875 + ], + "angle": 0, + "content": "Overall, mask loss (Sec. 3.4) and image-text contrastive loss (Sec. 3.5) complete the necessary mask and semantic supervision that is needed to train our model. The final loss is a weighted combination of the two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c o n t r a s t i v e}} \\mathcal {L} _ {\\text {c o n t r a s t i v e}} \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30225" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.093, + 0.895, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.394, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison with existing methods. CLIP [37] is primarily designed for classification and does not perform well in segmentation. MaskCLIP [52] adapts CLIP for segmentation, although it produces noisy predictions and cannot handle background classes. GroupViT [47] is a strong competitor, but it could struggle in challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.481 + ], + "angle": 0, + "content": "In our experiment, we use \\(\\lambda_{\\mathrm{mask}} = 1.0\\), \\(\\lambda_{\\mathrm{contrastive}} = 1.0\\), \\(\\lambda_{\\mathrm{dice}} = 1.0\\), \\(\\lambda_{\\mathrm{focal}} = 20.0\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.49, + 0.226, + 0.507 + ], + "angle": 0, + "content": "3.7. Self-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.633 + ], + "angle": 0, + "content": "To further improve our results, we introduce an optional step wherein we train a new model using the predictions generated by our current model. This process of self-training results in an augmented model, which we refer to as S-Seg+. More specifically, when we evaluate on a given dataset, we generate pseudo labels for the unlabeled images in the training set. Subsequently, we employ these pseudo labels to train a new segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.484, + 0.74 + ], + "angle": 0, + "content": "Self-training improves the accuracy by leveraging additional data [46], augmentation [53], and bootstrapping [16]. In our situation, self-training offers even greater benefits since we can take advantage of additional information that is obtainable during testing: unlabeled images and testing categories. We show that this additional step improves our results significantly at no extra manual labelling cost." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.224, + 0.77 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.84 + ], + "angle": 0, + "content": "In this section, we empirically evaluate our method and compare to existing approaches. We show that, although our method is quite simple, it performs surprisingly well against more complex existing methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.306, + 0.866 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Architecture. Our experiments use MaskFormer [8] with a 6-layer transformer decoder and \\( N = 64 \\) queries. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.542 + ], + "angle": 0, + "content": "hidden and output feature dimension is 256. The language model is a Transformer [42] with 12 layers, each with a hidden dimension of 256. We use a 2-layer MLP to project the visual and text feature into a common embedding space. We use DINO ViT-S/8 as the pretrained ViT in pseudo-mask generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.909, + 0.707 + ], + "angle": 0, + "content": "Training. During training, we used three publically available datasets: CC3M [40], CC12M [5], and RedCaps [10], containing 3M, 12M and 12M image-text pairs, respectively. Due to storage constraint, we use only first 11M data samples at a smaller resolution of \\(448 \\times 448\\) when using RedCaps dataset. In total, we use at most 26M imagetext pairs for training - this is an order of magnitude fewer data than CLIP [37] and 1-4M fewer than GroupViT [47]. For more detailed training hyperparameters, please refer to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Inference. We evaluate S-Seg on the validation set of three datasets: Pascal VOC 2012 [14], Pascal Context [35] and COCO [28]. The Pascal VOC dataset contains 1449 images for testing. Each image is labeled with 20 foreground classes and a background class. The Pascal Context dataset contains 5104 testing images with 59 foreground classes and a background class. The COCO dataset contains 5000 images for testing with 80 foreground classes and an additional background class. Following GroupViT [47], we threshold the maximum probability to obtain background prediction. During inference, we set the input resolution to \\(448 \\times 448\\), in consistent with [47]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "30226" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.487, + 0.345 + ], + "angle": 0, + "content": "
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Linearly-probed classification models:
MoCo v3 [7]✗----34.321.3-
DINO [3]✗----39.120.4-
Open-vocabulary models:
CLIP [37]†✓✗✓✗✗13.58.15.9
MaskCLIP [52]†✓✗✓✗✗26.822.812.8
ViL-Seg [30]✓✗✓✗✗34.416.316.4
CLIPpy [38]✓✗✓✗✗52.2--
GroupViT [47]✓✗✗✓✗50.823.727.5
SegCLIP [32]✓✗✓✗✗52.624.726.5
OVSegmentor [48]✓✗✗✓✗53.820.425.1
TCL [4]✓✗✓✓✗55.030.431.6
SAM-CLIP [43]✓✓✓✗✓60.629.231.5
S-Seg (Ours)✓✗✗✗✗53.227.930.3
S-Seg+ (Ours)✓✗✗✗✗62.030.235.7
Fully-supervised segmentation models:
DeepLabV3+† [6]✗----78.746.455.7
MaskFormer† [8]✗----81.250.062.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.496 + ], + "angle": 0, + "content": "Table 2. Open-vocabulary (OV) semantic segmentation results (background pixels included in evaluation). Benchmarked following standard evaluation protocols for unsupervised open-vocabulary models trained without annotated masks [4, 32, 47, 48]. Models labels: Mask (use GT mask), VLM (employ pretrained large VL model like CLIP), Enc. (require specialized grouping encoder), and SAM (use the SAM model). Our approach obtain second highest performance on average, without using any extra components such as CLIP and SAM. \\(\\dagger\\) denotes our recomputed results. Higher values are better." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.505, + 0.48, + 0.571 + ], + "angle": 0, + "content": "
MethodP. VOCP. ContextCOCO3-Avg.
B1: Pseudo Mask + CLIP12.93.92.96.6
B2: Pseudo-mask ViT23.211.010.414.9
S-Seg (Ours)44.922.922.530.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.632 + ], + "angle": 0, + "content": "Table 3. Simple baselines for open-vocabulary semantic segmentation. All models are trained on CC12M. Higher values are better. Two simple baselines fail to obtain satisfactory results, even using after using our pseudo masks and no less training data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.254, + 0.652 + ], + "angle": 0, + "content": "4.2. Simple baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.734 + ], + "angle": 0, + "content": "The high quality of pseudo-masks (as shown in Figure 7) may lead one to assume that the primary challenge is simply classifying these masks, and that this can be accomplished by utilizing pre-existing methods such as CLIP. To test this assumption, we first develop two simple baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Baseline 1: Pseudo-mask + CLIP. Firstly, our pseudo label generator is utilized to obtain pseudo segments. Then, we iterate through all the masks and apply the current mask to the original image. Next, the masked image is fed to CLIP for classification and the resulting class label is assigned to the corresponding segment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Baseline 2: Pseudo-mask ViT. We introduce a new visual backbone that differs from the regular ViT. Instead of pooling all image tokens into a single feature, we first individually pool tokens in each segment of the pseudo-mask into segment features, and then pool these features into a vi" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.905, + 0.393 + ], + "angle": 0, + "content": "
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Open-vocabulary models (annotated masks required for training):
ZS3Net [1]✓✗✗✗38.319.421.1
LSeg [24]✓✓✗✗52.3-27.2
OpenSeg [15]✓✓✗✗77.245.938.1
ZegFormer [11]✓✓✗✗80.7--
GKC [19]✓✓✗✗83.245.2-
ODISE [49]✓✓✓✗84.657.365.2*
DeOp [18]✓✓✗✗91.748.8-
OVSeg [27]✓✓✗✗94.555.7-
SAN [33]✓✓✓✗94.657.7-
FC-CLIP [50]✓✓✓✗95.458.4-
SED [45]✓✓✓✗96.160.6-
CAT-Seg [9]✓✓✓✗96.662.0-
Open-vocabulary models (annotated masks not required for training):
CLIP [37]†✓✗✓✗✗39.69.013.8
MaskCLIP [52]†✓✗✓✗✗49.525.523.6
GroupViT [47]†✓✗✗✓✗77.223.037.5
TCL [4]✓✗✓✓✗83.233.9-
S-Seg (Ours)✓✗✗✗✗81.827.242.4
S-Seg+ (Ours)✓✗✗✗✗84.731.653.0
Fully-supervised segmentation models:
DeepLabV3+† [6]✗GT---89.948.566.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Table 4. Open-vocabulary (OV) semantic segmentation results (background pixels excluded in evaluation). Benchmarked following standard protocol for evaluating supervised open-vocabulary models trained with annotated masks [11, 18, 19, 27, 33]. Similar to the previous setting, S-Seg achieves competitive performance compared to earlier methods trained without using any extra components such as CLIP and SAM. *COCO is used for training. Higher values are better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.578 + ], + "angle": 0, + "content": "sual embedding. We train a CLIP-like model from scratch using this visual backbone. During testing, we classify each segment feature and assign the label to that segment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.777 + ], + "angle": 0, + "content": "The results are presented in Table 3. As we can see, open-vocabulary segmentation is more complex than simply grouping image into segments and then categorizing them into classes, even when the segments are of high quality. Baseline 1 employs a significantly larger pretrained CLIP ViT/L-14 model that was also trained on a much larger dataset, while Baseline 2 is trained using the same data as ours. Nevertheless, both baselines fail to achieve satisfactory results, suggesting that open-vocabulary segmentation cannot be naively deconstructed in such ways. We hypothesize that a multi-task learning approach that jointly trains for segmentation and classification could yield significant advantages." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.771, + 0.804 + ], + "angle": 0, + "content": "4.3. Evaluation with background" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In table 2, we evaluate our model and compare with existing method on open-vocabulary semantic segmentation task. Following standard evaluation protocols on open-vocabulary model trained without annotated masks [4, 32, 47, 48], we include background pixels in evaluation and obtain background prediction by setting a threshold for back" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30227" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.1, + 0.321, + 0.112 + ], + "angle": 0, + "content": "(a) Pascal VOC \\((+18.3\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.115, + 0.371, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.1, + 0.578, + 0.112 + ], + "angle": 0, + "content": "(b) Pascal Context \\((+14.1\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.115, + 0.622, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.1, + 0.808, + 0.112 + ], + "angle": 0, + "content": "(c) COCO \\((+17.6\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.115, + 0.872, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Figure 10. Self-training improvement. We show average relative improvement in bracket on top of the plot. we observe that self-training consistently leads to significant improvement for S-Seg across all of our training and testing data settings." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.291, + 0.195, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.351, + 0.183, + 0.361 + ], + "angle": 0, + "content": "S-Seg (details)" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.292, + 0.293, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.35, + 0.282, + 0.361 + ], + "angle": 0, + "content": "S-Seg+ (details)" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.292, + 0.385, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.35, + 0.376, + 0.361 + ], + "angle": 0, + "content": "S-Seg (details)" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.292, + 0.481, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.35, + 0.475, + 0.361 + ], + "angle": 0, + "content": "S-Seg+ (details)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.413 + ], + "angle": 0, + "content": "Figure 11. Visualizing effect of self-training. Our self-trained S-Seg+ model demonstrates the ability to accurately predict in regions overlooked by S-Seg, as shown in the colorful rectangles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.563 + ], + "angle": 0, + "content": "ground classes [47]. Despite the simplicity of S-Seg, our approach achieve competitive performance over previous open-vocabulary segmentation methods that does not require mask annotations. Our model has second highest performance on average and has better results than GroupViT on all datasets. Moreover, our self-trained model, S-Seg+, provides an impressive \\(5.5\\%\\) mIoU improvement over our base model S-Seg (42.6% vs 37.1% 3-avg. mIoU), suggesting the efficacy of self-training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.371, + 0.59 + ], + "angle": 0, + "content": "4.4. Evaluation without background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.732 + ], + "angle": 0, + "content": "We also evaluate our model on the evaluation protocol commonly used for evaluating open-vocabulary models with annotated masks [11, 18, 19], where the background pixels are excluded in evaluation. We note that this setting is easier because background class is more diverse in appearance and often requires additional processing such as thresholding. Table 4 shows the results. Similar to the previous setting, our S-Seg and S-Seg+ models achieve competitive performance compared to earlier methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.251, + 0.757 + ], + "angle": 0, + "content": "4.5. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Self-training. We investigated the effectiveness of self-training for improving segmentation performance. To this end, we compared S-Seg and S-Seg+ on three datasets and evaluated the results using Figure 10. We found that self-training consistently improved the segmentation performance by a significant margin \\((+5.5\\%)\\) mIoU on average), regardless of the data size and test dataset. These results indicate that self-training is a reliable approach for enhancing the performance of S-Seg and can provide a de" + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.286, + 0.736, + 0.298 + ], + "angle": 0, + "content": "(a) S-Seg" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.301, + 0.627, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.301, + 0.716, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.301, + 0.803, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.301, + 0.895, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.402, + 0.741, + 0.414 + ], + "angle": 0, + "content": "(b) S-Seg+" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.417, + 0.626, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.418, + 0.714, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.418, + 0.801, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.418, + 0.889, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.513, + 0.776, + 0.523 + ], + "angle": 0, + "content": "Size of training data \\((\\times 10^{6})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.602 + ], + "angle": 0, + "content": "Figure 12. Scaling training data provide consistent gain in performance, with or without self-training. We train our model using different sizes of data: CC12M (12M), CC12M+CC3M (15M), and CC12M+CC3M+RedCaps (26M). We note a steady improvement in the performance as the data size increases." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.616, + 0.813, + 0.631 + ], + "angle": 0, + "content": "sirable complement for further improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.906, + 0.751 + ], + "angle": 0, + "content": "Data scalability. To evaluate the scalability of our method, we trained S-Seg and S-Seg+ using three datasets of increasing sizes: 12M, 15M, and 26M. The results of the experiments are presented in Figure 12. We observed that both models achieve significant improvements in performance across all three testing datasets as the amount of data increased, suggesting that our method scales well with larger datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.633, + 0.781 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.906, + 0.896 + ], + "angle": 0, + "content": "To summarize, we propose S-Seg, a simple and intuitive framework that enables accurate and generalizable open-vocabulary segmentation. Our algorithm directly trains for pixel-level feature and language alignment, and does not require manual segmentation annotations or extensive pretraining. We hope that our simple yet effective approach will serve as a solid baseline for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30228" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.186 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[4] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.482, + 0.456 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. 2021. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[10] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.616, + 0.482, + 0.656 + ], + "angle": 0, + "content": "[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[13] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.787, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[17] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[18] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[19] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, and Yunchao Wei. Global knowledge calibration for fast open-vocabulary segmentation. In ICCV, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[21] Tapas Kanungo, David M Mount, Nathan S Netanyahu, Christine D Piatko, Ruth Silverman, and Angela Y Wu. An efficient k-means clustering algorithm: Analysis and implementation. IEEE TPAMI, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[22] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.518, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[25] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[26] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[27] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[30] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30229" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[33] Fangyun Wei Han Hu Xiang Bai Mengde Xu, Zheng Zhang. Side adapter network for open-vocabulary semantic segmentation. 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[34] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[35] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[36] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[38] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. In ICCV, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[39] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2018, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[41] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[43] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[44] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[46] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[47] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[48] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[49] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[50] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[51] Haobo Yuan, Xiangtai Li, Chong Zhou, Yining Li, Kai Chen, and Chen Change Loy. Open-vocabulary sam: Segment and recognize twenty-thousand classes interactively. arXiv preprint arXiv:2401.02955, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[52] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[53] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. *NeurIPS*, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30230" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_origin.pdf b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..31aa8425bc4932f55dd27c374f80f37086aec3b3 --- /dev/null +++ b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/495afadb-de9a-431f-ba2d-9db6ebb6d17d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d59b9378f0eb74af27d6c8f093b7e0e1687bc69e95e6fdc18067953c8d5c0490 +size 2897932 diff --git a/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/full.md b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..15090a33ea1631df50581a91fa3b30108b920f66 --- /dev/null +++ b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/full.md @@ -0,0 +1,349 @@ +# Exploring Simple Open-Vocabulary Semantic Segmentation + +Zihang Lai Visual Geometry Group (VGG), University of Oxford zlai@robots.ox.ac.uk + +# Abstract + +Open-vocabulary semantic segmentation models aim to accurately assign a semantic label to each pixel in an image from a set of arbitrary open-vocabulary texts. In order to learn such pixel-level alignment, current approaches typically rely on a combination of (i) image-level VL model (e.g. CLIP), (ii) ground truth masks, (iii) custom grouping encoders, and (iv) the Segment Anything Model (SAM). In this paper, we introduce S-Seg, a simple model that can achieve surprisingly strong performance without depending on any of the above elements. S-Seg leverages pseudo-masks and language features to train a MaskFormer, and can be easily trained from publicly available image-text datasets. Contrary to prior works, our model directly trains for pixel-level features and language alignment. Once trained, S-Seg generalizes well to multiple testing datasets without requiring fine-tuning. In addition, S-Seg has the extra benefits of scalability with data and consistently improving when augmented with self-training. We believe that our simple yet effective approach will serve as a solid baseline for future research. Project page: zlai0.github.io/S-Seg. + +# 1. Introduction + +Open-vocabulary semantic segmentation presents a unique challenge as it requires assigning accurate semantic labels to each pixel in an image using arbitrary open-vocabulary texts, rather than a fixed set of classes. This means that the model must be able to segment and classify any arbitrary categories expressed in language. Achieving this requires a robust, pixel-level alignment between images and textual descriptions, which enables accurate association of each pixel with the most relevant class from a dynamically provided set of textual categories. + +A primary obstacle in this domain is that it is impossible to construct datasets that provide pixel-level annotations for all possible labels. This limitation often results in the adoption of weakly-supervised or semi-supervised learning approaches. Current methods typically rely on a combination of strategies to learn the required pixel-level alignment. + +![](images/2caf8be573234d8bca9040c0c5fd01f221a52463bbf41bc5be8e273eb99324db.jpg) +Figure 1. S-Seg result on a web image. Our goal is to segment everything, including fictional characters like minions. +Figure 2. Our S-Seg framework leverages pseudo-mask and language to train a MaskFormer. We show that our method of directly training for pixel-level feature and language alignment yields superior results. + +One common tactic is adapting existing Vision-Language (VL) models, which are initially trained for image-level alignment (e.g., CLIP [37]), to perform at the pixel level. Another strategy involves training models on ground truth masks that are annotated for a select number of seen classes, thereby encouraging the model to extrapolate its knowledge to novel unseen classes during testing. Furthermore, specialized models such as GroupViT [47] and OVSegmentor [48], which are explicitly designed for open- + +vocabulary segmentation, are being explored. Lastly, recent literature adapts the Segment Anything Model (SAM), originally not designed to classify segments, for effective use in open-vocabulary segmentation. + +In this paper, we report a model that can work surprisingly well with none of the above strategies. Our approach, named S-Seg, is built on top of a standard MaskFormer model. Our model directly trains for pixel-level feature and language alignment, using neither existing large image-level alignment models like CLIP [37] nor manually annotated segmentation or classification labels. + +One of the biggest challenges we face is finding the right + +![](images/9fa11898c8567044470900c48196ba772a25b7a2f0f1e958c22cf8ba3fd43a72.jpg) +Figure 3. Qualitative results of S-Seg, evaluated using all dataset classes as queries. Our model copes with challenging situation, such as overlapping objects (col. 2) and small objects (col. 5). Our model is also capable of handling "stuff" categories such as water and floor (col. 3, 4). Moreover, our S-Seg+ model is able to correct small errors observed in the S-Seg method (col. 4). Finally, in the COCO dataset, which featured a significantly higher number of objects, our model is still able to achieve high accuracy in its predictions. + +supervision since annotated masks and labels are not available. To address this issue, we propose to leverage pseudomasks and language to supervise MaskFormer. Our strategy involves using a pseudo-mask generator to provide class-agnostic mask supervision by generating pseudo ground truth masks. We adopt a simple design that clusters image representations obtained through self-supervised representation learning methods like DINO [3]. Our experiments demonstrate that this approach delivers exceptional performance, which is essential for high-quality supervision, as well as rapid processing speed, which is necessary for efficient training. In addition, we use noisy web texts to provide semantic supervision. The image-text dataset contains a wide range of concepts and has demonstrated impressive zero-shot classification results [37]. We utilize a straightforward image-text contrastive loss, which has proven to be highly effective. Once trained, our model generalizes well to new categories without requiring fine-tuning. + +S-Seg is a simple and effective model that can be trained using publicly available image-text datasets, such as Conceptual Captions [5, 40]. This makes it easy to reproduce and extend for further research. The S-Seg framework is also designed to be flexible with easily replaceable submodules. We prioritize simplicity in our subcomponent selection to focus on the general design of our framework, while remaining open to more advanced techniques that could result in further improvements. + +We conducted a thorough evaluation of S-Seg using multiple benchmark datasets, and we show that our method achieve competitive results on three widely tested benchmarks (Pascal VOC, Pascal Context, and COCO). In addition, pseudo-mask and language provide scalable supervision and our model consistently improves in performance as more data became available. Finally, we find adding an ad + +ditional self-training step leads to an even greater improvement to our model, with an average increase of $5.5\%$ mIoU over three datasets, highlighting the potential for further improvement of our approach. + +Our simple solution suggests that the reliance on complex models and extensive ground truth data in open-vocabulary semantic segmentation may be reduced, leading to more streamlined and accessible framework for future developments in the field, and we hope our exploration can serve as a solid baseline for future research. + +# 2. Related work + +Open-vocabulary segmentation. The earliest efforts to employ language for image segmentation can be traced back to Duygulu et al.'s seminal work [13], where the authors tackled image segmentation by framing it as a machine translation problem. Current approaches leverage a combination of strategies to learn pixel-level image-text alignment. + +Adapting image-level vision-language models. The first strategy involves the adapting pretrained vision-language models, originally designed for image-level alignment, to the more granular task of pixel-level alignment. This strategy is widely adopted in open-vocabulary methods [4, 15, 18, 19, 22, 24, 27, 32, 33, 38, 44, 48, 49, 52]. These works vary in their methods of refining image-level models for finer alignment tasks. + +MaskCLIP [52] demonstrates modifying the CLIP image encoder can significantly enhance its pixel-level alignment capabilities without requiring retraining. TCL [4] employs CLIP for initial text-to-image region grounding, followed by contrastive learning to refine the alignment between the text embedding and the grounded region. OpenSeg [15] fine-tunes ALIGN [20] using a grounding + +![](images/f344ff7f03ca2d473871f309a0bc51836d738753576f4feec0f0925555b3c2dc.jpg) +Figure 4. Overview of S-Seg. A MaskFormer model computes masks and mask features from an image input. A pseudo-mask generator produces segmentation maps to supervise mask predictions, while a text that describes the image, encoded by a language model trained together with the MaskFormer, provides supervision for mask features using image-text contrastive loss. +Figure 5. Pseudocode for training S-Seg with image-text pairs. + +loss [17] to better align words in captions to segmentation masks. OpenSeg and DiffuMask [44] also explored the use of pseudo-masks. The primary distinction lies in their dependency on different sources for learning; OpenSeg uses annotated segments while DiffuMask employs masks generated through diffusion. In contrast, our method is entirely learned from pseudo masks. Also, our mask generator is entirely self-supervised, whereas their mask generator is fully-supervised. + +Ground truth masks. Another effective strategy [11, 15, 18, 19, 24, 27, 33] involves training models using ground truth masks annotated for a limited set of seen classes. By training on seen annotations, models are encouraged to learn detailed features and patterns that are potentially applicable beyond the scope of the trained classes. + +Of most relevance, ZegFormer [11] trains a MaskFormer by decoupling zero-shot semantic segmentation into two sub-tasks, a class-agnostic grouping task and a zero-shot segment classification task. Our method has similar training paradigm but with notable distinctions. Similar to GroupViT, we train exclusively with image-text pairs and do not utilize a pretrained CLIP model. Notably, even without access to ground truth masks, labels, or CLIP, our method outperforms ZegFormer in unseen categories, indicating potentially stronger generalization. + +Custom grouping-based encoders. The third strategy employs custom-designed models specifically for open-vocabulary segmentation. GroupViT [47] groups pixels in an image hierarchically based on their attention scores with learnable group tokens. OVSegmentor [48] applies Slot Attention [31] for a similar pixel grouping process based on feature proximity. + +Segment Anything Model (SAM). The final strategy adapts the Segment Anything Model (SAM) [23], a powerful segmentation model trained on over 11 million images and 1.1 billion masks, for open-vocabulary segmentation. While the SAM model segments object with high accuracy, it does not effectively classify these segments. To address this limitation, follow-up works has adapted the SAM model for grounding, closed-set, and open-vocabulary segmentation [25, 39, 51] by leveraging additional training + +```python +I [n, h, w, c] - minibatch of aligned images +T [n, l] - minibatch of aligned texts +N - number of MaskFormer queries +C - number of pseudo masks +# predict mask, mask feature, and text feature +M, M_f = maskformer(I) # [n, N, H, W], [n, N, d_f] +T_f = textEncoder(T) # [n, d_f] +# aggregate all mask features [n, d_f] +M_f = M_f.mean(axis=1) +# generate pseudo masks [n, C, H, W] +S = pseudo_mask_generation(I) +# compute loss +loss_c = contrastive_loss(M_f, T_f) +loss_m = mask_loss(M, S) +loss = (loss_c + loss_m)/2 +``` + +data, adaptors, and distillation. + +Our model, S-Seg, can be conceptualized as a synergy of these approaches. It can be viewed as a CLIP model integrated with a MaskFormer image encoder, directly optimizing for pixel-level feature and language alignment. Alternatively, it resembles "ZegFormer with pseudomask and language training" or "GroupVit with a MaskFormer as the grouping mechanism." Interestingly, our model relates to each method by omitting certain core architectural components or supervision method. Even so, our method is able to achieve competitive performance. + +# 3. Approach + +Our proposed method, called S-Seg, is conceptually simple: we learn a MaskFormer model from pseudo-mask and language. Our method leverages image-text pairs solely, without relying on ground truth masks or large-scale pretrained models. In figure 5, we provide pseudocode for the core implementation of training S-Seg. Figure 4 provides a schematic layout of our approach. + +# 3.1. Problem definition + +We consider the problem of open-vocabulary semantic segmentation, where we aim to learn a function $f$ that maps an + +image $I$ and a set of category names $\mathcal{C} = \{c_i\}_{i = i}^{N_C}$ to a semantic segmentation map $S$ , where $c_{i}$ can be any category name expressed as open vocabulary texts. + +Our approach adopts the unsupervised setting [4, 38, 47, 52], which aims to learn open-vocabulary segmentation from image-text pairs only, without learning from any dense annotation or class labels. Specifically, we use a web dataset of image-text pairs $\mathcal{D} = \{(I_i,T_i)\}_{i = i}^{N_{\mathcal{D}}}$ for training, where $T_{i}$ is a caption that describes the corresponding image $I_{i}$ . However, since the textual labels are gathered from the web, they may be noisy and contain errors. We do not use any additional manually annotated segmentation or classification labels during training. + +During testing, a set of category names $\mathcal{C}$ is provided, and the model is tasked with assigning a semantic label $c_{i} \in \mathcal{C}$ to each pixel in an unlabeled image. The performance of the model is evaluated based on its mean Intersection over Union (mIoU) with the ground truth labels. + +# 3.2. Adapting MaskFormer + +Our approach builds on top of MaskFormer [8]. Here, we begin by briefly review MaskFormer and explain the adjustments we made. + +The Maskformer model takes an image as input and generates $N$ masks and mask features. First, the input image passes through a backbone model to produce feature maps at different output resolutions. These image features are then fed into a per-pixel encoder, which upsamples and aggregates them into a set of feature maps with higher resolution. Meanwhile, a transformer decoder uses $N$ learnable queries to cross-attend to the set of features with the lowest resolution and gather global information about each segment. + +In the original Maskformer, a linear classifier and softmax activation were applied to the output of the decoder to predict class probabilities for a fixed list of categories. However, as we do not have a fixed list of categories, we remove this classifier branch and output the $N$ raw mask features instead. + +In addition to predicting mask features, the Maskformer also predicts $N$ binary masks. To predict each mask, a dot product is taken between the mask embedding, generated from mask features, and the high resolution per-pixel feature. Finally, $N$ mask-feature pairs are combined to generate the output. + +# 3.3. S-Seg + +S-Seg employs MaskFormer as its segmentation model, but in our weakly-supervised learning setting (where only texts are available), we face the challenge of not having annotated masks and labels. To overcome this, we utilize pseudo labels and language to as supervision. + +Our training framework is illustrated in Figure 4. We first + +![](images/fa5ed71c24a51ed37f63628f75ec0c2037ce28d0a2262483dae9acf82aa15c12.jpg) +Figure 6. Testing on S-Seg. During inference, S-Seg generalize to new categories by leveraging language features generated from a list of candidate classes in text. + +![](images/09e8812e7567f49aff372fd0a9f706aefb6a26a41a368434a184f8501d3b73f9.jpg) +Figure 7. Pseudo-mask generator generates pseudo-masks to supervise predicted mask during training. This module takes an image as its input, extracts its features using a DINO pre-trained ViT, and then employs K-means clustering to group the pixels into segments. + +generate a set of segmentation maps using our pseudo-mask generator (Sec. 3.4) and use them as supervision for mask prediction. Meanwhile, we use a language model to process input text and generate language embeddings. These embeddings provide supervision for mask features by leveraging image-text contrastive loss (Sec. 3.5). + +Notably, unlike the supervised learning setting, where mask and label annotations are coupled, we decouple mask and semantic supervision. This enables us to utilize pseudomask and language as two distinct forms of supervision. + +In the testing phase (as shown in figure 6), the trained MaskFormer model predicts $N$ masks and mask features from the input image. The language model takes as input a list of candidate category names (represented as texts) and extracts a set of language features. These features are then used to classify the mask features. This process is similar to the one used in CLIP [37], where the image and possible text inputs are encoded by their respective encoders to compute feature embeddings. The cosine similarity between these embeddings is calculated and adjusted by a learnable temperature parameter. The resulting values are normalized into a class probability distribution using a softmax function, and a combination module is used to takes $N$ mask-class pairs to produce the final segmentation map, similar to [8]. + +Next, we provide a details of the subcomponents in our framework. + +# 3.4. Pseudo-Mask Generator + +In our approach, we use a pseudo-mask generator (fig. 7) to produce a class-agnostic segmentation map from the input + +![](images/99bed843566e1491365aec73ad4b7854aed94aa1d3e4161eb1414f3dc4196609.jpg) +Pseudo mask +Figure 8. Example pseudo-masks. Our pseudo-mask generator is capable of generating high-quality artificial masks. When provided with an oracle label, these masks demonstrate a high degree of overlap with the ground truth annotations. + +
MethodSup.P. VOC↑P. Context↑Time(s)↓
Spectral Clus. [41]*none49.243.20.543
K-Means [21]*none49.543.30.188
ImageNet [12]+[21]label68.858.10.079
GroupViT [47]text73.754.60.002
Pseudo-mask (Ours)self78.866.30.002
+ +Table 1. Our pseudo-mask generator achieves excellent oracle performance with rapid speed, making it an ideal mask supervision. We report amortised running time on a batch of 128 samples, simulating training time scenario. * We process downsampled image at $\frac{H}{8} \times \frac{W}{8}$ resolution to obtain reasonable running time. + +image, which supervises the mask prediction of our model. + +To implement the pseudo-mask generator, we adopt a simple strategy that involves clustering tokens extracted from a self-supervised pre-trained ViT. Specifically, we use a DINO-pretrained ViT to compute a set of featurized tokens from the input image. We then apply a clustering algorithm (K-Means in our case) to these tokens, assigning each token a label that corresponds to the index of the cluster it belongs to. We reshape the resulting label map into an image and resize it to the original resolution to supervise the mask prediction of our segmentation model. + +Despite its simplicity, our pseudo-mask generator achieves both impressive performance, which is crucial for high-quality supervision, and fast processing speed, which is essential for efficient training. We evaluate its performance and compare against baseline methods, and the quantitative results are presented in Table 1, with example predictions visualized in 8. Our method significantly outperforms simple baselines such as K-Means and Spectral Clustering, which naively cluster image pixels, while running two orders of magnitude faster. We also observed that clustering DINO representation outperforms clustering ImageNet pre-trained ViT representation by a significant margin. Notably, our pseudo-mask generator even outperforms GroupViT, which has already employed vision-language training. + +Since the predicted masks are unordered, we need to + +match the $N$ predicted masks with $K$ pseudo ground truth masks. To accomplish this, we utilize bipartite matching, as described in [2, 8], which assigns a pseudo-mask to each predicted mask such that the overall assignment cost is minimal in all possible assignments. Since each pseudo-mask is assigned to at most one predicted mask, $N - K$ pseudomasks are unassigned to no-object. Unlike standard training recipe [8], we do not penalize these no-object masks, nor do we use classification loss as an assignment cost. Finally, we compute the mask loss between predicted masks and their corresponding pseudo-mask, utilizing a combination of dice loss [34] and focal loss [29]. + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {\text {d i c e}} \mathcal {L} _ {\text {d i c e}} + \lambda_ {\text {f o c a l}} \mathcal {L} _ {\text {f o c a l}} \tag {1} +$$ + +# 3.5. Language Supervision + +Our model learns to classify open-vocabulary concepts from language supervision. To train the model, we use an image-text contrastive loss [15, 37]. Specifically, we view $N$ mask features as representation of the input image, each capturing information about a different part of the image. We then compute a single feature that represents the entire image by taking the average of these mask features. To encode the text, we use a text transformer [42] and select the embedding corresponding to the [EOS] token, resulting in a textual feature. Since the visual and textual features may have different dimensions, we project each representation into a common embedding space using 2-layer MLPs. To compute the image-text contrastive loss, we calculate the cosine similarity between the image embeddings and the text embeddings within the same batch. Following common practice [26, 36, 37], we decouple the image-text contrastive loss into two parts: + +$$ +\mathcal {L} _ {I \rightarrow T} = - \frac {1}{N} \sum_ {i} ^ {N} \log \frac {\exp \left(x _ {i} ^ {\intercal} y _ {i} / \sigma\right)}{\sum_ {j = 1} ^ {N} \exp \left(x _ {i} ^ {\intercal} y _ {j} / \sigma\right)} \tag {2} +$$ + +$$ +\mathcal {L} _ {T \rightarrow I} = - \frac {1}{N} \sum_ {i} ^ {N} \log \frac {\exp \left(y _ {i} ^ {\intercal} x _ {i} / \sigma\right)}{\sum_ {j = 1} ^ {N} \exp \left(y _ {i} ^ {\intercal} x _ {j} / \sigma\right)} \tag {3} +$$ + +where $x_{i}$ and $y_{i}$ are L2-normalized embedding of image and text of the i-th pair. $N$ denotes batch size and $\sigma$ is a learnable temperature parameter optimized together with the rest of the model. The total loss is the sum of these two losses, $\mathcal{L}_{\mathrm{contrastive}} = \mathcal{L}_{I\to T} + \mathcal{L}_{T\to I}$ . + +# 3.6. Training Loss + +Overall, mask loss (Sec. 3.4) and image-text contrastive loss (Sec. 3.5) complete the necessary mask and semantic supervision that is needed to train our model. The final loss is a weighted combination of the two losses: + +$$ +L = \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + \lambda_ {\text {c o n t r a s t i v e}} \mathcal {L} _ {\text {c o n t r a s t i v e}} \tag {4} +$$ + +![](images/b463b9f7f82dd3bdca4137c9793207b02e5099766353ec239f90374432021be9.jpg) +Figure 9. Qualitative comparison with existing methods. CLIP [37] is primarily designed for classification and does not perform well in segmentation. MaskCLIP [52] adapts CLIP for segmentation, although it produces noisy predictions and cannot handle background classes. GroupViT [47] is a strong competitor, but it could struggle in challenging scenarios. + +In our experiment, we use $\lambda_{\mathrm{mask}} = 1.0$ , $\lambda_{\mathrm{contrastive}} = 1.0$ , $\lambda_{\mathrm{dice}} = 1.0$ , $\lambda_{\mathrm{focal}} = 20.0$ . + +# 3.7. Self-training + +To further improve our results, we introduce an optional step wherein we train a new model using the predictions generated by our current model. This process of self-training results in an augmented model, which we refer to as S-Seg+. More specifically, when we evaluate on a given dataset, we generate pseudo labels for the unlabeled images in the training set. Subsequently, we employ these pseudo labels to train a new segmentation model. + +Self-training improves the accuracy by leveraging additional data [46], augmentation [53], and bootstrapping [16]. In our situation, self-training offers even greater benefits since we can take advantage of additional information that is obtainable during testing: unlabeled images and testing categories. We show that this additional step improves our results significantly at no extra manual labelling cost. + +# 4. Experiments + +In this section, we empirically evaluate our method and compare to existing approaches. We show that, although our method is quite simple, it performs surprisingly well against more complex existing methods. + +# 4.1. Implementation details + +Architecture. Our experiments use MaskFormer [8] with a 6-layer transformer decoder and $N = 64$ queries. The + +hidden and output feature dimension is 256. The language model is a Transformer [42] with 12 layers, each with a hidden dimension of 256. We use a 2-layer MLP to project the visual and text feature into a common embedding space. We use DINO ViT-S/8 as the pretrained ViT in pseudo-mask generator. + +Training. During training, we used three publically available datasets: CC3M [40], CC12M [5], and RedCaps [10], containing 3M, 12M and 12M image-text pairs, respectively. Due to storage constraint, we use only first 11M data samples at a smaller resolution of $448 \times 448$ when using RedCaps dataset. In total, we use at most 26M imagetext pairs for training - this is an order of magnitude fewer data than CLIP [37] and 1-4M fewer than GroupViT [47]. For more detailed training hyperparameters, please refer to the appendix. + +Inference. We evaluate S-Seg on the validation set of three datasets: Pascal VOC 2012 [14], Pascal Context [35] and COCO [28]. The Pascal VOC dataset contains 1449 images for testing. Each image is labeled with 20 foreground classes and a background class. The Pascal Context dataset contains 5104 testing images with 59 foreground classes and a background class. The COCO dataset contains 5000 images for testing with 80 foreground classes and an additional background class. Following GroupViT [47], we threshold the maximum probability to obtain background prediction. During inference, we set the input resolution to $448 \times 448$ , in consistent with [47]. + +
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Linearly-probed classification models:
MoCo v3 [7]✗----34.321.3-
DINO [3]✗----39.120.4-
Open-vocabulary models:
CLIP [37]†✓✗✓✗✗13.58.15.9
MaskCLIP [52]†✓✗✓✗✗26.822.812.8
ViL-Seg [30]✓✗✓✗✗34.416.316.4
CLIPpy [38]✓✗✓✗✗52.2--
GroupViT [47]✓✗✗✓✗50.823.727.5
SegCLIP [32]✓✗✓✗✗52.624.726.5
OVSegmentor [48]✓✗✗✓✗53.820.425.1
TCL [4]✓✗✓✓✗55.030.431.6
SAM-CLIP [43]✓✓✓✗✓60.629.231.5
S-Seg (Ours)✓✗✗✗✗53.227.930.3
S-Seg+ (Ours)✓✗✗✗✗62.030.235.7
Fully-supervised segmentation models:
DeepLabV3+† [6]✗----78.746.455.7
MaskFormer† [8]✗----81.250.062.1
+ +Table 2. Open-vocabulary (OV) semantic segmentation results (background pixels included in evaluation). Benchmarked following standard evaluation protocols for unsupervised open-vocabulary models trained without annotated masks [4, 32, 47, 48]. Models labels: Mask (use GT mask), VLM (employ pretrained large VL model like CLIP), Enc. (require specialized grouping encoder), and SAM (use the SAM model). Our approach obtain second highest performance on average, without using any extra components such as CLIP and SAM. $\dagger$ denotes our recomputed results. Higher values are better. + +
MethodP. VOCP. ContextCOCO3-Avg.
B1: Pseudo Mask + CLIP12.93.92.96.6
B2: Pseudo-mask ViT23.211.010.414.9
S-Seg (Ours)44.922.922.530.1
+ +# 4.2. Simple baselines + +The high quality of pseudo-masks (as shown in Figure 7) may lead one to assume that the primary challenge is simply classifying these masks, and that this can be accomplished by utilizing pre-existing methods such as CLIP. To test this assumption, we first develop two simple baselines. + +Baseline 1: Pseudo-mask + CLIP. Firstly, our pseudo label generator is utilized to obtain pseudo segments. Then, we iterate through all the masks and apply the current mask to the original image. Next, the masked image is fed to CLIP for classification and the resulting class label is assigned to the corresponding segment. + +Baseline 2: Pseudo-mask ViT. We introduce a new visual backbone that differs from the regular ViT. Instead of pooling all image tokens into a single feature, we first individually pool tokens in each segment of the pseudo-mask into segment features, and then pool these features into a vi + +Table 3. Simple baselines for open-vocabulary semantic segmentation. All models are trained on CC12M. Higher values are better. Two simple baselines fail to obtain satisfactory results, even using after using our pseudo masks and no less training data. + +
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Open-vocabulary models (annotated masks required for training):
ZS3Net [1]✓✗✗✗38.319.421.1
LSeg [24]✓✓✗✗52.3-27.2
OpenSeg [15]✓✓✗✗77.245.938.1
ZegFormer [11]✓✓✗✗80.7--
GKC [19]✓✓✗✗83.245.2-
ODISE [49]✓✓✓✗84.657.365.2*
DeOp [18]✓✓✗✗91.748.8-
OVSeg [27]✓✓✗✗94.555.7-
SAN [33]✓✓✓✗94.657.7-
FC-CLIP [50]✓✓✓✗95.458.4-
SED [45]✓✓✓✗96.160.6-
CAT-Seg [9]✓✓✓✗96.662.0-
Open-vocabulary models (annotated masks not required for training):
CLIP [37]†✓✗✓✗✗39.69.013.8
MaskCLIP [52]†✓✗✓✗✗49.525.523.6
GroupViT [47]†✓✗✗✓✗77.223.037.5
TCL [4]✓✗✓✓✗83.233.9-
S-Seg (Ours)✓✗✗✗✗81.827.242.4
S-Seg+ (Ours)✓✗✗✗✗84.731.653.0
Fully-supervised segmentation models:
DeepLabV3+† [6]✗GT---89.948.566.9
+ +Table 4. Open-vocabulary (OV) semantic segmentation results (background pixels excluded in evaluation). Benchmarked following standard protocol for evaluating supervised open-vocabulary models trained with annotated masks [11, 18, 19, 27, 33]. Similar to the previous setting, S-Seg achieves competitive performance compared to earlier methods trained without using any extra components such as CLIP and SAM. *COCO is used for training. Higher values are better. + +sual embedding. We train a CLIP-like model from scratch using this visual backbone. During testing, we classify each segment feature and assign the label to that segment. + +The results are presented in Table 3. As we can see, open-vocabulary segmentation is more complex than simply grouping image into segments and then categorizing them into classes, even when the segments are of high quality. Baseline 1 employs a significantly larger pretrained CLIP ViT/L-14 model that was also trained on a much larger dataset, while Baseline 2 is trained using the same data as ours. Nevertheless, both baselines fail to achieve satisfactory results, suggesting that open-vocabulary segmentation cannot be naively deconstructed in such ways. We hypothesize that a multi-task learning approach that jointly trains for segmentation and classification could yield significant advantages. + +# 4.3. Evaluation with background + +In table 2, we evaluate our model and compare with existing method on open-vocabulary semantic segmentation task. Following standard evaluation protocols on open-vocabulary model trained without annotated masks [4, 32, 47, 48], we include background pixels in evaluation and obtain background prediction by setting a threshold for back + +![](images/b721ac53106a6ec030767e0a2492f8f379b321f43bc4e9cdf9d3891a60643ae4.jpg) +(a) Pascal VOC $(+18.3\%)$ + +![](images/7e734d2df69c37061f0c9e39ee04bd0a1c97ab62a32ab726f1bb095af0401aaf.jpg) +(b) Pascal Context $(+14.1\%)$ +Figure 10. Self-training improvement. We show average relative improvement in bracket on top of the plot. we observe that self-training consistently leads to significant improvement for S-Seg across all of our training and testing data settings. + +![](images/ee1f9647d5223e15acd714d2948429a03e3fbbadf49143eb7e0c2e6f79cd2dc4.jpg) +(c) COCO $(+17.6\%)$ + +![](images/0c486708a1bcb6efd0e4d3f3136b9afedbfc396dd8e0dfc004d89c6342fad0e9.jpg) +S-Seg (details) +Figure 11. Visualizing effect of self-training. Our self-trained S-Seg+ model demonstrates the ability to accurately predict in regions overlooked by S-Seg, as shown in the colorful rectangles. + +![](images/4ce6363644a1f6822ca7e8136f1ed7266e650839167e1a901f26cd6476c9a0f0.jpg) +S-Seg+ (details) + +![](images/8f03368b0991e209d40baec368fce736dfc4eec87514ba2da4a9505d7cb01590.jpg) +S-Seg (details) + +![](images/3f7be99205e5a2faaaf34a6852c2a9e528701758f4ac44753b33bbf3cb7d4bab.jpg) +S-Seg+ (details) + +ground classes [47]. Despite the simplicity of S-Seg, our approach achieve competitive performance over previous open-vocabulary segmentation methods that does not require mask annotations. Our model has second highest performance on average and has better results than GroupViT on all datasets. Moreover, our self-trained model, S-Seg+, provides an impressive $5.5\%$ mIoU improvement over our base model S-Seg (42.6% vs 37.1% 3-avg. mIoU), suggesting the efficacy of self-training. + +# 4.4. Evaluation without background + +We also evaluate our model on the evaluation protocol commonly used for evaluating open-vocabulary models with annotated masks [11, 18, 19], where the background pixels are excluded in evaluation. We note that this setting is easier because background class is more diverse in appearance and often requires additional processing such as thresholding. Table 4 shows the results. Similar to the previous setting, our S-Seg and S-Seg+ models achieve competitive performance compared to earlier methods. + +# 4.5. Ablation studies + +Self-training. We investigated the effectiveness of self-training for improving segmentation performance. To this end, we compared S-Seg and S-Seg+ on three datasets and evaluated the results using Figure 10. We found that self-training consistently improved the segmentation performance by a significant margin $(+5.5\%)$ mIoU on average), regardless of the data size and test dataset. These results indicate that self-training is a reliable approach for enhancing the performance of S-Seg and can provide a de + +![](images/af176d28d05d8e0752f64c9d368d072671bcb27600d65eac8fc42210b288303f.jpg) + +![](images/741b4de68b2938b0483b1ce2b153437fe5e017fd3acc6073b0eda4251538cb54.jpg) +(a) S-Seg + +![](images/4f2f1e9acf4fd12ae534eaecef9a0d4c10d261590773ce851440e933c7350a2b.jpg) + +![](images/bd2ece15d355ac639e64cc308f6fb2a0252fdad89f007022fc12d2c702857cee.jpg) + +![](images/e67a20d3295c34f1c94c2da4ad986ae53d7cba63160f48edbb119e9a7f017e17.jpg) +Figure 12. Scaling training data provide consistent gain in performance, with or without self-training. We train our model using different sizes of data: CC12M (12M), CC12M+CC3M (15M), and CC12M+CC3M+RedCaps (26M). We note a steady improvement in the performance as the data size increases. + +![](images/0254ff69ca81fa9847903f84bb286c9bf9754b2ac08670b2dcbbfa39d9c5af77.jpg) +(b) S-Seg+ +Size of training data $(\times 10^{6})$ + +![](images/62725b948e8f5fe8c6384d84f2f2233260535b92e14d179bc2b1414011bc6f6d.jpg) + +![](images/54865075f1660a72c082a35ca530a6dee88b40b424f72a1877bb53b9c974b31a.jpg) + +sirable complement for further improvement. + +Data scalability. To evaluate the scalability of our method, we trained S-Seg and S-Seg+ using three datasets of increasing sizes: 12M, 15M, and 26M. The results of the experiments are presented in Figure 12. We observed that both models achieve significant improvements in performance across all three testing datasets as the amount of data increased, suggesting that our method scales well with larger datasets. + +# 5. Conclusion + +To summarize, we propose S-Seg, a simple and intuitive framework that enables accurate and generalizable open-vocabulary segmentation. Our algorithm directly trains for pixel-level feature and language alignment, and does not require manual segmentation annotations or extensive pretraining. We hope that our simple yet effective approach will serve as a solid baseline for future research. + +# References + +[1] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 7 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 7 +[4] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 2, 4, 7 +[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 6 +[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7 +[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 7 +[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. 2021. 4, 5, 6, 7 +[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 7 +[10] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431, 2021.6 +[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 3, 7, 8 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 5 +[13] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, 2002. 2 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2009. 6 +[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3, 5, 7 +[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, + +Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 6 +[17] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In ECCV, 2020. 3 +[18] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2, 3, 7, 8 +[19] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, and Yunchao Wei. Global knowledge calibration for fast open-vocabulary segmentation. In ICCV, 2023. 2, 3, 7, 8 +[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. 2021. 2 +[21] Tapas Kanungo, David M Mount, Nathan S Netanyahu, Christine D Piatko, Ruth Silverman, and Angela Y Wu. An efficient k-means clustering algorithm: Analysis and implementation. IEEE TPAMI, 2002. 5 +[22] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316, 2023. 2 +[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3 +[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2, 3, 7 +[25] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 3 +[26] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. 2023. 5 +[27] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 2, 3, 7 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6 +[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5 +[30] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In ECCV, 2022. 7 +[31] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob + +Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 3 +[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. 2023. 2, 7 +[33] Fangyun Wei Han Hu Xiang Bai Mengde Xu, Zheng Zhang. Side adapter network for open-vocabulary semantic segmentation. 2023. 2, 3, 7 +[34] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 5 +[35] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 6 +[36] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 5 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 1, 2, 4, 5, 6, 7 +[38] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. In ICCV, 2023. 2, 4, 7 +[39] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3 +[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2018, 2, 6 +[41] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5 +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 5, 6 +[43] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 7 +[44] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In ICCV, 2023. 2, 3 +[45] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 7 + +[46] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 6 +[47] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 1, 3, 4, 5, 6, 7, 8 +[48] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 1, 2, 3, 7 +[49] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2, 7 +[50] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. Advances in Neural Information Processing Systems, 36, 2024. 7 +[51] Haobo Yuan, Xiangtai Li, Chong Zhou, Yining Li, Kai Chen, and Chen Change Loy. Open-vocabulary sam: Segment and recognize twenty-thousand classes interactively. arXiv preprint arXiv:2401.02955, 2024. 3 +[52] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 2, 4, 6, 7 +[53] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. *NeurIPS*, 2020. 6 \ No newline at end of file diff --git a/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/images.zip b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5bfc570afd757a9c7764426c83a2deb3e9f1e4a2 --- /dev/null +++ b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a1826906f52c53726558d85ab5f30f9fc4d0bff15061a509df0aa8bea7eca56 +size 628861 diff --git a/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/layout.json b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..212c7566cea95c2a033c036ef3a8b289510cdad5 --- /dev/null +++ b/2025/Exploring Simple Open-Vocabulary Semantic Segmentation/layout.json @@ -0,0 +1,8302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 490, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 490, + 121 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 490, + 121 + ], + "type": "text", + "content": "Exploring Simple Open-Vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 144, + 435, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 144, + 435, + 184 + ], + "spans": [ + { + "bbox": [ + 175, + 144, + 435, + 184 + ], + "type": "text", + "content": "Zihang Lai Visual Geometry Group (VGG), University of Oxford zlai@robots.ox.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 478 + ], + "type": "text", + "content": "Open-vocabulary semantic segmentation models aim to accurately assign a semantic label to each pixel in an image from a set of arbitrary open-vocabulary texts. In order to learn such pixel-level alignment, current approaches typically rely on a combination of (i) image-level VL model (e.g. CLIP), (ii) ground truth masks, (iii) custom grouping encoders, and (iv) the Segment Anything Model (SAM). In this paper, we introduce S-Seg, a simple model that can achieve surprisingly strong performance without depending on any of the above elements. S-Seg leverages pseudo-masks and language features to train a MaskFormer, and can be easily trained from publicly available image-text datasets. Contrary to prior works, our model directly trains for pixel-level features and language alignment. Once trained, S-Seg generalizes well to multiple testing datasets without requiring fine-tuning. In addition, S-Seg has the extra benefits of scalability with data and consistently improving when augmented with self-training. We believe that our simple yet effective approach will serve as a solid baseline for future research. Project page: zlai0.github.io/S-Seg." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 501, + 135, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 135, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 135, + 513 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "text", + "content": "Open-vocabulary semantic segmentation presents a unique challenge as it requires assigning accurate semantic labels to each pixel in an image using arbitrary open-vocabulary texts, rather than a fixed set of classes. This means that the model must be able to segment and classify any arbitrary categories expressed in language. Achieving this requires a robust, pixel-level alignment between images and textual descriptions, which enables accurate association of each pixel with the most relevant class from a dynamically provided set of textual categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "A primary obstacle in this domain is that it is impossible to construct datasets that provide pixel-level annotations for all possible labels. This limitation often results in the adoption of weakly-supervised or semi-supervised learning approaches. Current methods typically rely on a combination of strategies to learn the required pixel-level alignment." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 247, + 555, + 392 + ], + "blocks": [ + { + "bbox": [ + 313, + 213, + 555, + 235 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 235 + ], + "type": "text", + "content": "Figure 1. S-Seg result on a web image. Our goal is to segment everything, including fictional characters like minions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 247, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 318, + 247, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 555, + 392 + ], + "type": "image", + "image_path": "2caf8be573234d8bca9040c0c5fd01f221a52463bbf41bc5be8e273eb99324db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 400, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 444 + ], + "type": "text", + "content": "Figure 2. Our S-Seg framework leverages pseudo-mask and language to train a MaskFormer. We show that our method of directly training for pixel-level feature and language alignment yields superior results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 567 + ], + "type": "text", + "content": "One common tactic is adapting existing Vision-Language (VL) models, which are initially trained for image-level alignment (e.g., CLIP [37]), to perform at the pixel level. Another strategy involves training models on ground truth masks that are annotated for a select number of seen classes, thereby encouraging the model to extrapolate its knowledge to novel unseen classes during testing. Furthermore, specialized models such as GroupViT [47] and OVSegmentor [48], which are explicitly designed for open-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": "vocabulary segmentation, are being explored. Lastly, recent literature adapts the Segment Anything Model (SAM), originally not designed to classify segments, for effective use in open-vocabulary segmentation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "type": "text", + "content": "In this paper, we report a model that can work surprisingly well with none of the above strategies. Our approach, named S-Seg, is built on top of a standard MaskFormer model. Our model directly trains for pixel-level feature and language alignment, using neither existing large image-level alignment models like CLIP [37] nor manually annotated segmentation or classification labels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "One of the biggest challenges we face is finding the right" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30221" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 235 + ], + "type": "image", + "image_path": "9fa11898c8567044470900c48196ba772a25b7a2f0f1e958c22cf8ba3fd43a72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 244, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 244, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 555, + 289 + ], + "type": "text", + "content": "Figure 3. Qualitative results of S-Seg, evaluated using all dataset classes as queries. Our model copes with challenging situation, such as overlapping objects (col. 2) and small objects (col. 5). Our model is also capable of handling \"stuff\" categories such as water and floor (col. 3, 4). Moreover, our S-Seg+ model is able to correct small errors observed in the S-Seg method (col. 4). Finally, in the COCO dataset, which featured a significantly higher number of objects, our model is still able to achieve high accuracy in its predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 300, + 294, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 294, + 516 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 294, + 516 + ], + "type": "text", + "content": "supervision since annotated masks and labels are not available. To address this issue, we propose to leverage pseudomasks and language to supervise MaskFormer. Our strategy involves using a pseudo-mask generator to provide class-agnostic mask supervision by generating pseudo ground truth masks. We adopt a simple design that clusters image representations obtained through self-supervised representation learning methods like DINO [3]. Our experiments demonstrate that this approach delivers exceptional performance, which is essential for high-quality supervision, as well as rapid processing speed, which is necessary for efficient training. In addition, we use noisy web texts to provide semantic supervision. The image-text dataset contains a wide range of concepts and has demonstrated impressive zero-shot classification results [37]. We utilize a straightforward image-text contrastive loss, which has proven to be highly effective. Once trained, our model generalizes well to new categories without requiring fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 518, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 626 + ], + "type": "text", + "content": "S-Seg is a simple and effective model that can be trained using publicly available image-text datasets, such as Conceptual Captions [5, 40]. This makes it easy to reproduce and extend for further research. The S-Seg framework is also designed to be flexible with easily replaceable submodules. We prioritize simplicity in our subcomponent selection to focus on the general design of our framework, while remaining open to more advanced techniques that could result in further improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "We conducted a thorough evaluation of S-Seg using multiple benchmark datasets, and we show that our method achieve competitive results on three widely tested benchmarks (Pascal VOC, Pascal Context, and COCO). In addition, pseudo-mask and language provide scalable supervision and our model consistently improves in performance as more data became available. Finally, we find adding an ad" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 300, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 347 + ], + "type": "text", + "content": "ditional self-training step leads to an even greater improvement to our model, with an average increase of " + }, + { + "bbox": [ + 313, + 300, + 553, + 347 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 313, + 300, + 553, + 347 + ], + "type": "text", + "content": " mIoU over three datasets, highlighting the potential for further improvement of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 348, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 554, + 418 + ], + "type": "text", + "content": "Our simple solution suggests that the reliance on complex models and extensive ground truth data in open-vocabulary semantic segmentation may be reduced, leading to more streamlined and accessible framework for future developments in the field, and we hope our exploration can serve as a solid baseline for future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 430, + 399, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 399, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 399, + 441 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": "Open-vocabulary segmentation. The earliest efforts to employ language for image segmentation can be traced back to Duygulu et al.'s seminal work [13], where the authors tackled image segmentation by framing it as a machine translation problem. Current approaches leverage a combination of strategies to learn pixel-level image-text alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 629 + ], + "type": "text", + "content": "Adapting image-level vision-language models. The first strategy involves the adapting pretrained vision-language models, originally designed for image-level alignment, to the more granular task of pixel-level alignment. This strategy is widely adopted in open-vocabulary methods [4, 15, 18, 19, 22, 24, 27, 32, 33, 38, 44, 48, 49, 52]. These works vary in their methods of refining image-level models for finer alignment tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "MaskCLIP [52] demonstrates modifying the CLIP image encoder can significantly enhance its pixel-level alignment capabilities without requiring retraining. TCL [4] employs CLIP for initial text-to-image region grounding, followed by contrastive learning to refine the alignment between the text embedding and the grounded region. OpenSeg [15] fine-tunes ALIGN [20] using a grounding" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30222" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 58, + 521, + 168 + ], + "blocks": [ + { + "bbox": [ + 92, + 58, + 521, + 168 + ], + "lines": [ + { + "bbox": [ + 92, + 58, + 521, + 168 + ], + "spans": [ + { + "bbox": [ + 92, + 58, + 521, + 168 + ], + "type": "image", + "image_path": "f344ff7f03ca2d473871f309a0bc51836d738753576f4feec0f0925555b3c2dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 171, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 54, + 171, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 555, + 205 + ], + "type": "text", + "content": "Figure 4. Overview of S-Seg. A MaskFormer model computes masks and mask features from an image input. A pseudo-mask generator produces segmentation maps to supervise mask predictions, while a text that describes the image, encoded by a language model trained together with the MaskFormer, provides supervision for mask features using image-text contrastive loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 396, + 553, + 408 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 408 + ], + "type": "text", + "content": "Figure 5. Pseudocode for training S-Seg with image-text pairs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 215, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 215, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 215, + 294, + 323 + ], + "type": "text", + "content": "loss [17] to better align words in captions to segmentation masks. OpenSeg and DiffuMask [44] also explored the use of pseudo-masks. The primary distinction lies in their dependency on different sources for learning; OpenSeg uses annotated segments while DiffuMask employs masks generated through diffusion. In contrast, our method is entirely learned from pseudo masks. Also, our mask generator is entirely self-supervised, whereas their mask generator is fully-supervised." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 325, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 396 + ], + "type": "text", + "content": "Ground truth masks. Another effective strategy [11, 15, 18, 19, 24, 27, 33] involves training models using ground truth masks annotated for a limited set of seen classes. By training on seen annotations, models are encouraged to learn detailed features and patterns that are potentially applicable beyond the scope of the trained classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 399, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 296, + 518 + ], + "type": "text", + "content": "Of most relevance, ZegFormer [11] trains a MaskFormer by decoupling zero-shot semantic segmentation into two sub-tasks, a class-agnostic grouping task and a zero-shot segment classification task. Our method has similar training paradigm but with notable distinctions. Similar to GroupViT, we train exclusively with image-text pairs and do not utilize a pretrained CLIP model. Notably, even without access to ground truth masks, labels, or CLIP, our method outperforms ZegFormer in unseen categories, indicating potentially stronger generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 604 + ], + "type": "text", + "content": "Custom grouping-based encoders. The third strategy employs custom-designed models specifically for open-vocabulary segmentation. GroupViT [47] groups pixels in an image hierarchically based on their attention scores with learnable group tokens. OVSegmentor [48] applies Slot Attention [31] for a similar pixel grouping process based on feature proximity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Segment Anything Model (SAM). The final strategy adapts the Segment Anything Model (SAM) [23], a powerful segmentation model trained on over 11 million images and 1.1 billion masks, for open-vocabulary segmentation. While the SAM model segments object with high accuracy, it does not effectively classify these segments. To address this limitation, follow-up works has adapted the SAM model for grounding, closed-set, and open-vocabulary segmentation [25, 39, 51] by leveraging additional training" + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 322, + 217, + 545, + 387 + ], + "blocks": [ + { + "bbox": [ + 322, + 217, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 322, + 217, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 322, + 217, + 545, + 387 + ], + "type": "text", + "content": "I [n, h, w, c] - minibatch of aligned images \nT [n, l] - minibatch of aligned texts \nN - number of MaskFormer queries \nC - number of pseudo masks \n# predict mask, mask feature, and text feature \nM, M_f = maskformer(I) # [n, N, H, W], [n, N, d_f] \nT_f = textEncoder(T) # [n, d_f] \n# aggregate all mask features [n, d_f] \nM_f = M_f.mean(axis=1) \n# generate pseudo masks [n, C, H, W] \nS = pseudo_mask_generation(I) \n# compute loss \nloss_c = contrastive_loss(M_f, T_f) \nloss_m = mask_loss(M, S) \nloss = (loss_c + loss_m)/2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 313, + 419, + 440, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 440, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 440, + 430 + ], + "type": "text", + "content": "data, adaptors, and distillation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 551 + ], + "type": "text", + "content": "Our model, S-Seg, can be conceptualized as a synergy of these approaches. It can be viewed as a CLIP model integrated with a MaskFormer image encoder, directly optimizing for pixel-level feature and language alignment. Alternatively, it resembles \"ZegFormer with pseudomask and language training\" or \"GroupVit with a MaskFormer as the grouping mechanism.\" Interestingly, our model relates to each method by omitting certain core architectural components or supervision method. Even so, our method is able to achieve competitive performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 561, + 381, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 381, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 381, + 574 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 581, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 665 + ], + "type": "text", + "content": "Our proposed method, called S-Seg, is conceptually simple: we learn a MaskFormer model from pseudo-mask and language. Our method leverages image-text pairs solely, without relying on ground truth masks or large-scale pretrained models. In figure 5, we provide pseudocode for the core implementation of training S-Seg. Figure 4 provides a schematic layout of our approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 424, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 424, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 424, + 683 + ], + "type": "text", + "content": "3.1. Problem definition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We consider the problem of open-vocabulary semantic segmentation, where we aim to learn a function " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " that maps an" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30223" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "text", + "content": "image " + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "text", + "content": " and a set of category names " + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_i\\}_{i = i}^{N_C}" + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "text", + "content": " to a semantic segmentation map " + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 71, + 295, + 108 + ], + "type": "text", + "content": " can be any category name expressed as open vocabulary texts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "text", + "content": "Our approach adopts the unsupervised setting [4, 38, 47, 52], which aims to learn open-vocabulary segmentation from image-text pairs only, without learning from any dense annotation or class labels. Specifically, we use a web dataset of image-text pairs " + }, + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(I_i,T_i)\\}_{i = i}^{N_{\\mathcal{D}}}" + }, + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "text", + "content": " for training, where " + }, + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "text", + "content": " is a caption that describes the corresponding image " + }, + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 109, + 296, + 228 + ], + "type": "text", + "content": ". However, since the textual labels are gathered from the web, they may be noisy and contain errors. We do not use any additional manually annotated segmentation or classification labels during training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 288 + ], + "type": "text", + "content": "During testing, a set of category names " + }, + { + "bbox": [ + 55, + 228, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 228, + 296, + 288 + ], + "type": "text", + "content": " is provided, and the model is tasked with assigning a semantic label " + }, + { + "bbox": [ + 55, + 228, + 296, + 288 + ], + "type": "inline_equation", + "content": "c_{i} \\in \\mathcal{C}" + }, + { + "bbox": [ + 55, + 228, + 296, + 288 + ], + "type": "text", + "content": " to each pixel in an unlabeled image. The performance of the model is evaluated based on its mean Intersection over Union (mIoU) with the ground truth labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 187, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 187, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 187, + 309 + ], + "type": "text", + "content": "3.2. Adapting MaskFormer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 350 + ], + "type": "text", + "content": "Our approach builds on top of MaskFormer [8]. Here, we begin by briefly review MaskFormer and explain the adjustments we made." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 351, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 470 + ], + "type": "text", + "content": "The Maskformer model takes an image as input and generates " + }, + { + "bbox": [ + 55, + 351, + 296, + 470 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 351, + 296, + 470 + ], + "type": "text", + "content": " masks and mask features. First, the input image passes through a backbone model to produce feature maps at different output resolutions. These image features are then fed into a per-pixel encoder, which upsamples and aggregates them into a set of feature maps with higher resolution. Meanwhile, a transformer decoder uses " + }, + { + "bbox": [ + 55, + 351, + 296, + 470 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 351, + 296, + 470 + ], + "type": "text", + "content": " learnable queries to cross-attend to the set of features with the lowest resolution and gather global information about each segment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 541 + ], + "type": "text", + "content": "In the original Maskformer, a linear classifier and softmax activation were applied to the output of the decoder to predict class probabilities for a fixed list of categories. However, as we do not have a fixed list of categories, we remove this classifier branch and output the " + }, + { + "bbox": [ + 55, + 471, + 296, + 541 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 471, + 296, + 541 + ], + "type": "text", + "content": " raw mask features instead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": "In addition to predicting mask features, the Maskformer also predicts " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": " binary masks. To predict each mask, a dot product is taken between the mask embedding, generated from mask features, and the high resolution per-pixel feature. Finally, " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": " mask-feature pairs are combined to generate the output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 623, + 105, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 105, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 105, + 636 + ], + "type": "text", + "content": "3.3. S-Seg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": "S-Seg employs MaskFormer as its segmentation model, but in our weakly-supervised learning setting (where only texts are available), we face the challenge of not having annotated masks and labels. To overcome this, we utilize pseudo labels and language to as supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Our training framework is illustrated in Figure 4. We first" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 71, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 545, + 168 + ], + "type": "image", + "image_path": "fa5ed71c24a51ed37f63628f75ec0c2037ce28d0a2262483dae9acf82aa15c12.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "text", + "content": "Figure 6. Testing on S-Seg. During inference, S-Seg generalize to new categories by leveraging language features generated from a list of candidate classes in text." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 324, + 216, + 545, + 262 + ], + "blocks": [ + { + "bbox": [ + 324, + 216, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 324, + 216, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 324, + 216, + 545, + 262 + ], + "type": "image", + "image_path": "09e8812e7567f49aff372fd0a9f706aefb6a26a41a368434a184f8501d3b73f9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 273, + 553, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 273, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 553, + 327 + ], + "type": "text", + "content": "Figure 7. Pseudo-mask generator generates pseudo-masks to supervise predicted mask during training. This module takes an image as its input, extracts its features using a DINO pre-trained ViT, and then employs K-means clustering to group the pixels into segments." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 340, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 411 + ], + "type": "text", + "content": "generate a set of segmentation maps using our pseudo-mask generator (Sec. 3.4) and use them as supervision for mask prediction. Meanwhile, we use a language model to process input text and generate language embeddings. These embeddings provide supervision for mask features by leveraging image-text contrastive loss (Sec. 3.5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 411, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 459 + ], + "type": "text", + "content": "Notably, unlike the supervised learning setting, where mask and label annotations are coupled, we decouple mask and semantic supervision. This enables us to utilize pseudomask and language as two distinct forms of supervision." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 460, + 554, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 638 + ], + "type": "text", + "content": "In the testing phase (as shown in figure 6), the trained MaskFormer model predicts " + }, + { + "bbox": [ + 313, + 460, + 554, + 638 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 460, + 554, + 638 + ], + "type": "text", + "content": " masks and mask features from the input image. The language model takes as input a list of candidate category names (represented as texts) and extracts a set of language features. These features are then used to classify the mask features. This process is similar to the one used in CLIP [37], where the image and possible text inputs are encoded by their respective encoders to compute feature embeddings. The cosine similarity between these embeddings is calculated and adjusted by a learnable temperature parameter. The resulting values are normalized into a class probability distribution using a softmax function, and a combination module is used to takes " + }, + { + "bbox": [ + 313, + 460, + 554, + 638 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 460, + 554, + 638 + ], + "type": "text", + "content": " mask-class pairs to produce the final segmentation map, similar to [8]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 639, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 553, + 662 + ], + "type": "text", + "content": "Next, we provide a details of the subcomponents in our framework." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 672, + 451, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 451, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 451, + 683 + ], + "type": "text", + "content": "3.4. Pseudo-Mask Generator" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "In our approach, we use a pseudo-mask generator (fig. 7) to produce a class-agnostic segmentation map from the input" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30224" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 82, + 287, + 185 + ], + "blocks": [ + { + "bbox": [ + 105, + 72, + 156, + 81 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 156, + 81 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 156, + 81 + ], + "type": "text", + "content": "Pseudo mask" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 82, + 287, + 185 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 287, + 185 + ], + "type": "image", + "image_path": "99bed843566e1491365aec73ad4b7854aed94aa1d3e4161eb1414f3dc4196609.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 295, + 238 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 238 + ], + "type": "text", + "content": "Figure 8. Example pseudo-masks. Our pseudo-mask generator is capable of generating high-quality artificial masks. When provided with an oracle label, these masks demonstrate a high degree of overlap with the ground truth annotations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 239, + 293, + 310 + ], + "blocks": [ + { + "bbox": [ + 58, + 239, + 293, + 310 + ], + "lines": [ + { + "bbox": [ + 58, + 239, + 293, + 310 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 293, + 310 + ], + "type": "table", + "html": "
MethodSup.P. VOC↑P. Context↑Time(s)↓
Spectral Clus. [41]*none49.243.20.543
K-Means [21]*none49.543.30.188
ImageNet [12]+[21]label68.858.10.079
GroupViT [47]text73.754.60.002
Pseudo-mask (Ours)self78.866.30.002
", + "image_path": "8d3ab9651217262ffd6d521b49c9a6c7f3408e27fd8db626f1da0c6c74759ab6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 321, + 296, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 377 + ], + "type": "text", + "content": "Table 1. Our pseudo-mask generator achieves excellent oracle performance with rapid speed, making it an ideal mask supervision. We report amortised running time on a batch of 128 samples, simulating training time scenario. * We process downsampled image at " + }, + { + "bbox": [ + 55, + 321, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\frac{H}{8} \\times \\frac{W}{8}" + }, + { + "bbox": [ + 55, + 321, + 296, + 377 + ], + "type": "text", + "content": " resolution to obtain reasonable running time." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 390, + 294, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 294, + 402 + ], + "type": "text", + "content": "image, which supervises the mask prediction of our model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 402, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 521 + ], + "type": "text", + "content": "To implement the pseudo-mask generator, we adopt a simple strategy that involves clustering tokens extracted from a self-supervised pre-trained ViT. Specifically, we use a DINO-pretrained ViT to compute a set of featurized tokens from the input image. We then apply a clustering algorithm (K-Means in our case) to these tokens, assigning each token a label that corresponds to the index of the cluster it belongs to. We reshape the resulting label map into an image and resize it to the original resolution to supervise the mask prediction of our segmentation model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 700 + ], + "type": "text", + "content": "Despite its simplicity, our pseudo-mask generator achieves both impressive performance, which is crucial for high-quality supervision, and fast processing speed, which is essential for efficient training. We evaluate its performance and compare against baseline methods, and the quantitative results are presented in Table 1, with example predictions visualized in 8. Our method significantly outperforms simple baselines such as K-Means and Spectral Clustering, which naively cluster image pixels, while running two orders of magnitude faster. We also observed that clustering DINO representation outperforms clustering ImageNet pre-trained ViT representation by a significant margin. Notably, our pseudo-mask generator even outperforms GroupViT, which has already employed vision-language training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "content": "Since the predicted masks are unordered, we need to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "match the " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " predicted masks with " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " pseudo ground truth masks. To accomplish this, we utilize bipartite matching, as described in [2, 8], which assigns a pseudo-mask to each predicted mask such that the overall assignment cost is minimal in all possible assignments. Since each pseudo-mask is assigned to at most one predicted mask, " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "N - K" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " pseudomasks are unassigned to no-object. Unlike standard training recipe [8], we do not penalize these no-object masks, nor do we use classification loss as an assignment cost. Finally, we compute the mask loss between predicted masks and their corresponding pseudo-mask, utilizing a combination of dice loss [34] and focal loss [29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 369, + 228, + 555, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 228, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 369, + 228, + 555, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\text {d i c e}} \\mathcal {L} _ {\\text {d i c e}} + \\lambda_ {\\text {f o c a l}} \\mathcal {L} _ {\\text {f o c a l}} \\tag {1}", + "image_path": "1c329832c2d648857b68fe2cae6b95e4c5d18390e37c3bf6c0ce4a4d217be761.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 441, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 441, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 441, + 260 + ], + "type": "text", + "content": "3.5. Language Supervision" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 264, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 468 + ], + "type": "text", + "content": "Our model learns to classify open-vocabulary concepts from language supervision. To train the model, we use an image-text contrastive loss [15, 37]. Specifically, we view " + }, + { + "bbox": [ + 313, + 264, + 555, + 468 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 264, + 555, + 468 + ], + "type": "text", + "content": " mask features as representation of the input image, each capturing information about a different part of the image. We then compute a single feature that represents the entire image by taking the average of these mask features. To encode the text, we use a text transformer [42] and select the embedding corresponding to the [EOS] token, resulting in a textual feature. Since the visual and textual features may have different dimensions, we project each representation into a common embedding space using 2-layer MLPs. To compute the image-text contrastive loss, we calculate the cosine similarity between the image embeddings and the text embeddings within the same batch. Following common practice [26, 36, 37], we decouple the image-text contrastive loss into two parts:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 346, + 484, + 554, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 484, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 346, + 484, + 554, + 518 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I \\rightarrow T} = - \\frac {1}{N} \\sum_ {i} ^ {N} \\log \\frac {\\exp \\left(x _ {i} ^ {\\intercal} y _ {i} / \\sigma\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(x _ {i} ^ {\\intercal} y _ {j} / \\sigma\\right)} \\tag {2}", + "image_path": "ae0a03e2c82b9a0866a35afacf190bfc8b9095699e74f14643c7bd88eb7feb4c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 346, + 525, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 525, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 346, + 525, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T \\rightarrow I} = - \\frac {1}{N} \\sum_ {i} ^ {N} \\log \\frac {\\exp \\left(y _ {i} ^ {\\intercal} x _ {i} / \\sigma\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(y _ {i} ^ {\\intercal} x _ {j} / \\sigma\\right)} \\tag {3}", + "image_path": "4cbd151367b44d3ec243ed8429833b494eed069938afdb271ec28da6f255c11a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "text", + "content": " are L2-normalized embedding of image and text of the i-th pair. " + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "text", + "content": " denotes batch size and " + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "text", + "content": " is a learnable temperature parameter optimized together with the rest of the model. The total loss is the sum of these two losses, " + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrastive}} = \\mathcal{L}_{I\\to T} + \\mathcal{L}_{T\\to I}" + }, + { + "bbox": [ + 313, + 562, + 555, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 628, + 402, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 402, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 402, + 640 + ], + "type": "text", + "content": "3.6. Training Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": "Overall, mask loss (Sec. 3.4) and image-text contrastive loss (Sec. 3.5) complete the necessary mask and semantic supervision that is needed to train our model. The final loss is a weighted combination of the two losses:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 701, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 358, + 701, + 555, + 714 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c o n t r a s t i v e}} \\mathcal {L} _ {\\text {c o n t r a s t i v e}} \\tag {4}", + "image_path": "3380e29613e05a3baad2f1e98b42876b51ca495e858975446a0304e42c3c5fdd.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30225" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 73, + 547, + 308 + ], + "blocks": [ + { + "bbox": [ + 65, + 73, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 65, + 73, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 65, + 73, + 547, + 308 + ], + "type": "image", + "image_path": "b463b9f7f82dd3bdca4137c9793207b02e5099766353ec239f90374432021be9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 312, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 312, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 555, + 346 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison with existing methods. CLIP [37] is primarily designed for classification and does not perform well in segmentation. MaskCLIP [52] adapts CLIP for segmentation, although it produces noisy predictions and cannot handle background classes. GroupViT [47] is a strong competitor, but it could struggle in challenging scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "text", + "content": "In our experiment, we use " + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mask}} = 1.0" + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{contrastive}} = 1.0" + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dice}} = 1.0" + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{focal}} = 20.0" + }, + { + "bbox": [ + 55, + 356, + 296, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 138, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 138, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 138, + 401 + ], + "type": "text", + "content": "3.7. Self-training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 501 + ], + "type": "text", + "content": "To further improve our results, we introduce an optional step wherein we train a new model using the predictions generated by our current model. This process of self-training results in an augmented model, which we refer to as S-Seg+. More specifically, when we evaluate on a given dataset, we generate pseudo labels for the unlabeled images in the training set. Subsequently, we employ these pseudo labels to train a new segmentation model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 501, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 296, + 586 + ], + "type": "text", + "content": "Self-training improves the accuracy by leveraging additional data [46], augmentation [53], and bootstrapping [16]. In our situation, self-training offers even greater benefits since we can take advantage of additional information that is obtainable during testing: unlabeled images and testing categories. We show that this additional step improves our results significantly at no extra manual labelling cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": "In this section, we empirically evaluate our method and compare to existing approaches. We show that, although our method is quite simple, it performs surprisingly well against more complex existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 187, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 187, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 187, + 685 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Architecture. Our experiments use MaskFormer [8] with a 6-layer transformer decoder and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "N = 64" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " queries. The" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 356, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 429 + ], + "type": "text", + "content": "hidden and output feature dimension is 256. The language model is a Transformer [42] with 12 layers, each with a hidden dimension of 256. We use a 2-layer MLP to project the visual and text feature into a common embedding space. We use DINO ViT-S/8 as the pretrained ViT in pseudo-mask generator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 439, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 556, + 559 + ], + "type": "text", + "content": "Training. During training, we used three publically available datasets: CC3M [40], CC12M [5], and RedCaps [10], containing 3M, 12M and 12M image-text pairs, respectively. Due to storage constraint, we use only first 11M data samples at a smaller resolution of " + }, + { + "bbox": [ + 313, + 439, + 556, + 559 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 313, + 439, + 556, + 559 + ], + "type": "text", + "content": " when using RedCaps dataset. In total, we use at most 26M imagetext pairs for training - this is an order of magnitude fewer data than CLIP [37] and 1-4M fewer than GroupViT [47]. For more detailed training hyperparameters, please refer to the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Inference. We evaluate S-Seg on the validation set of three datasets: Pascal VOC 2012 [14], Pascal Context [35] and COCO [28]. The Pascal VOC dataset contains 1449 images for testing. Each image is labeled with 20 foreground classes and a background class. The Pascal Context dataset contains 5104 testing images with 59 foreground classes and a background class. The COCO dataset contains 5000 images for testing with 80 foreground classes and an additional background class. Following GroupViT [47], we threshold the maximum probability to obtain background prediction. During inference, we set the input resolution to " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ", in consistent with [47]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30226" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 273 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 273 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 273 + ], + "type": "table", + "html": "
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Linearly-probed classification models:
MoCo v3 [7]✗----34.321.3-
DINO [3]✗----39.120.4-
Open-vocabulary models:
CLIP [37]†✓✗✓✗✗13.58.15.9
MaskCLIP [52]†✓✗✓✗✗26.822.812.8
ViL-Seg [30]✓✗✓✗✗34.416.316.4
CLIPpy [38]✓✗✓✗✗52.2--
GroupViT [47]✓✗✗✓✗50.823.727.5
SegCLIP [32]✓✗✓✗✗52.624.726.5
OVSegmentor [48]✓✗✗✓✗53.820.425.1
TCL [4]✓✗✓✓✗55.030.431.6
SAM-CLIP [43]✓✓✓✗✓60.629.231.5
S-Seg (Ours)✓✗✗✗✗53.227.930.3
S-Seg+ (Ours)✓✗✗✗✗62.030.235.7
Fully-supervised segmentation models:
DeepLabV3+† [6]✗----78.746.455.7
MaskFormer† [8]✗----81.250.062.1
", + "image_path": "4fcad3b064431adf218ff7441212e3501181f791d10e03a52c13df06ac421296.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 399, + 293, + 452 + ], + "blocks": [ + { + "bbox": [ + 55, + 283, + 296, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 392 + ], + "type": "text", + "content": "Table 2. Open-vocabulary (OV) semantic segmentation results (background pixels included in evaluation). Benchmarked following standard evaluation protocols for unsupervised open-vocabulary models trained without annotated masks [4, 32, 47, 48]. Models labels: Mask (use GT mask), VLM (employ pretrained large VL model like CLIP), Enc. (require specialized grouping encoder), and SAM (use the SAM model). Our approach obtain second highest performance on average, without using any extra components such as CLIP and SAM. " + }, + { + "bbox": [ + 55, + 283, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 283, + 296, + 392 + ], + "type": "text", + "content": " denotes our recomputed results. Higher values are better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 399, + 293, + 452 + ], + "lines": [ + { + "bbox": [ + 56, + 399, + 293, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 293, + 452 + ], + "type": "table", + "html": "
MethodP. VOCP. ContextCOCO3-Avg.
B1: Pseudo Mask + CLIP12.93.92.96.6
B2: Pseudo-mask ViT23.211.010.414.9
S-Seg (Ours)44.922.922.530.1
", + "image_path": "c254016d5bd2887f7cc2c08ae8c49cf8e5c23ea7748a1abeb2bdec0c89694f77.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 503, + 155, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 155, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 155, + 516 + ], + "type": "text", + "content": "4.2. Simple baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 581 + ], + "type": "text", + "content": "The high quality of pseudo-masks (as shown in Figure 7) may lead one to assume that the primary challenge is simply classifying these masks, and that this can be accomplished by utilizing pre-existing methods such as CLIP. To test this assumption, we first develop two simple baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": "Baseline 1: Pseudo-mask + CLIP. Firstly, our pseudo label generator is utilized to obtain pseudo segments. Then, we iterate through all the masks and apply the current mask to the original image. Next, the masked image is fed to CLIP for classification and the resulting class label is assigned to the corresponding segment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Baseline 2: Pseudo-mask ViT. We introduce a new visual backbone that differs from the regular ViT. Instead of pooling all image tokens into a single feature, we first individually pool tokens in each segment of the pseudo-mask into segment features, and then pool these features into a vi" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 553, + 311 + ], + "blocks": [ + { + "bbox": [ + 55, + 456, + 295, + 500 + ], + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 500 + ], + "type": "text", + "content": "Table 3. Simple baselines for open-vocabulary semantic segmentation. All models are trained on CC12M. Higher values are better. Two simple baselines fail to obtain satisfactory results, even using after using our pseudo masks and no less training data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 311 + ], + "type": "table", + "html": "
MethodOV|Mask VLM Enc. SAM|P. VOC P. Context COCO
Open-vocabulary models (annotated masks required for training):
ZS3Net [1]✓✗✗✗38.319.421.1
LSeg [24]✓✓✗✗52.3-27.2
OpenSeg [15]✓✓✗✗77.245.938.1
ZegFormer [11]✓✓✗✗80.7--
GKC [19]✓✓✗✗83.245.2-
ODISE [49]✓✓✓✗84.657.365.2*
DeOp [18]✓✓✗✗91.748.8-
OVSeg [27]✓✓✗✗94.555.7-
SAN [33]✓✓✓✗94.657.7-
FC-CLIP [50]✓✓✓✗95.458.4-
SED [45]✓✓✓✗96.160.6-
CAT-Seg [9]✓✓✓✗96.662.0-
Open-vocabulary models (annotated masks not required for training):
CLIP [37]†✓✗✓✗✗39.69.013.8
MaskCLIP [52]†✓✗✓✗✗49.525.523.6
GroupViT [47]†✓✗✗✓✗77.223.037.5
TCL [4]✓✗✓✓✗83.233.9-
S-Seg (Ours)✓✗✗✗✗81.827.242.4
S-Seg+ (Ours)✓✗✗✗✗84.731.653.0
Fully-supervised segmentation models:
DeepLabV3+† [6]✗GT---89.948.566.9
", + "image_path": "9af991822afaa93e5da4c50f63c0062b651f5b6a47bba07212935a4e1f03f7d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 319, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 407 + ], + "type": "text", + "content": "Table 4. Open-vocabulary (OV) semantic segmentation results (background pixels excluded in evaluation). Benchmarked following standard protocol for evaluating supervised open-vocabulary models trained with annotated masks [11, 18, 19, 27, 33]. Similar to the previous setting, S-Seg achieves competitive performance compared to earlier methods trained without using any extra components such as CLIP and SAM. *COCO is used for training. Higher values are better." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 421, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 457 + ], + "type": "text", + "content": "sual embedding. We train a CLIP-like model from scratch using this visual backbone. During testing, we classify each segment feature and assign the label to that segment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 458, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 615 + ], + "type": "text", + "content": "The results are presented in Table 3. As we can see, open-vocabulary segmentation is more complex than simply grouping image into segments and then categorizing them into classes, even when the segments are of high quality. Baseline 1 employs a significantly larger pretrained CLIP ViT/L-14 model that was also trained on a much larger dataset, while Baseline 2 is trained using the same data as ours. Nevertheless, both baselines fail to achieve satisfactory results, suggesting that open-vocabulary segmentation cannot be naively deconstructed in such ways. We hypothesize that a multi-task learning approach that jointly trains for segmentation and classification could yield significant advantages." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 471, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 471, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 471, + 636 + ], + "type": "text", + "content": "4.3. Evaluation with background" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "In table 2, we evaluate our model and compare with existing method on open-vocabulary semantic segmentation task. Following standard evaluation protocols on open-vocabulary model trained without annotated masks [4, 32, 47, 48], we include background pixels in evaluation and obtain background prediction by setting a threshold for back" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30227" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 91, + 227, + 200 + ], + "blocks": [ + { + "bbox": [ + 107, + 79, + 196, + 88 + ], + "lines": [ + { + "bbox": [ + 107, + 79, + 196, + 88 + ], + "spans": [ + { + "bbox": [ + 107, + 79, + 196, + 88 + ], + "type": "text", + "content": "(a) Pascal VOC " + }, + { + "bbox": [ + 107, + 79, + 196, + 88 + ], + "type": "inline_equation", + "content": "(+18.3\\%)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 91, + 227, + 200 + ], + "lines": [ + { + "bbox": [ + 74, + 91, + 227, + 200 + ], + "spans": [ + { + "bbox": [ + 74, + 91, + 227, + 200 + ], + "type": "image", + "image_path": "b721ac53106a6ec030767e0a2492f8f379b321f43bc4e9cdf9d3891a60643ae4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 228, + 91, + 380, + 200 + ], + "blocks": [ + { + "bbox": [ + 256, + 79, + 353, + 88 + ], + "lines": [ + { + "bbox": [ + 256, + 79, + 353, + 88 + ], + "spans": [ + { + "bbox": [ + 256, + 79, + 353, + 88 + ], + "type": "text", + "content": "(b) Pascal Context " + }, + { + "bbox": [ + 256, + 79, + 353, + 88 + ], + "type": "inline_equation", + "content": "(+14.1\\%)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 91, + 380, + 200 + ], + "lines": [ + { + "bbox": [ + 228, + 91, + 380, + 200 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 380, + 200 + ], + "type": "image", + "image_path": "7e734d2df69c37061f0c9e39ee04bd0a1c97ab62a32ab726f1bb095af0401aaf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 203, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 553, + 225 + ], + "type": "text", + "content": "Figure 10. Self-training improvement. We show average relative improvement in bracket on top of the plot. we observe that self-training consistently leads to significant improvement for S-Seg across all of our training and testing data settings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 381, + 91, + 533, + 199 + ], + "blocks": [ + { + "bbox": [ + 423, + 79, + 494, + 88 + ], + "lines": [ + { + "bbox": [ + 423, + 79, + 494, + 88 + ], + "spans": [ + { + "bbox": [ + 423, + 79, + 494, + 88 + ], + "type": "text", + "content": "(c) COCO " + }, + { + "bbox": [ + 423, + 79, + 494, + 88 + ], + "type": "inline_equation", + "content": "(+17.6\\%)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 91, + 533, + 199 + ], + "lines": [ + { + "bbox": [ + 381, + 91, + 533, + 199 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 533, + 199 + ], + "type": "image", + "image_path": "ee1f9647d5223e15acd714d2948429a03e3fbbadf49143eb7e0c2e6f79cd2dc4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 230, + 119, + 275 + ], + "blocks": [ + { + "bbox": [ + 59, + 230, + 119, + 275 + ], + "lines": [ + { + "bbox": [ + 59, + 230, + 119, + 275 + ], + "spans": [ + { + "bbox": [ + 59, + 230, + 119, + 275 + ], + "type": "image", + "image_path": "0c486708a1bcb6efd0e4d3f3136b9afedbfc396dd8e0dfc004d89c6342fad0e9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 277, + 111, + 285 + ], + "lines": [ + { + "bbox": [ + 65, + 277, + 111, + 285 + ], + "spans": [ + { + "bbox": [ + 65, + 277, + 111, + 285 + ], + "type": "text", + "content": "S-Seg (details)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "text", + "content": "Figure 11. Visualizing effect of self-training. Our self-trained S-Seg+ model demonstrates the ability to accurately predict in regions overlooked by S-Seg, as shown in the colorful rectangles." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 120, + 231, + 179, + 275 + ], + "blocks": [ + { + "bbox": [ + 120, + 231, + 179, + 275 + ], + "lines": [ + { + "bbox": [ + 120, + 231, + 179, + 275 + ], + "spans": [ + { + "bbox": [ + 120, + 231, + 179, + 275 + ], + "type": "image", + "image_path": "4ce6363644a1f6822ca7e8136f1ed7266e650839167e1a901f26cd6476c9a0f0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 277, + 172, + 285 + ], + "lines": [ + { + "bbox": [ + 121, + 277, + 172, + 285 + ], + "spans": [ + { + "bbox": [ + 121, + 277, + 172, + 285 + ], + "type": "text", + "content": "S-Seg+ (details)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 180, + 231, + 235, + 275 + ], + "blocks": [ + { + "bbox": [ + 180, + 231, + 235, + 275 + ], + "lines": [ + { + "bbox": [ + 180, + 231, + 235, + 275 + ], + "spans": [ + { + "bbox": [ + 180, + 231, + 235, + 275 + ], + "type": "image", + "image_path": "8f03368b0991e209d40baec368fce736dfc4eec87514ba2da4a9505d7cb01590.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 277, + 230, + 285 + ], + "lines": [ + { + "bbox": [ + 183, + 277, + 230, + 285 + ], + "spans": [ + { + "bbox": [ + 183, + 277, + 230, + 285 + ], + "type": "text", + "content": "S-Seg (details)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 238, + 231, + 294, + 275 + ], + "blocks": [ + { + "bbox": [ + 238, + 231, + 294, + 275 + ], + "lines": [ + { + "bbox": [ + 238, + 231, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 238, + 231, + 294, + 275 + ], + "type": "image", + "image_path": "3f7be99205e5a2faaaf34a6852c2a9e528701758f4ac44753b33bbf3cb7d4bab.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 277, + 290, + 285 + ], + "lines": [ + { + "bbox": [ + 239, + 277, + 290, + 285 + ], + "spans": [ + { + "bbox": [ + 239, + 277, + 290, + 285 + ], + "type": "text", + "content": "S-Seg+ (details)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 338, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 445 + ], + "type": "text", + "content": "ground classes [47]. Despite the simplicity of S-Seg, our approach achieve competitive performance over previous open-vocabulary segmentation methods that does not require mask annotations. Our model has second highest performance on average and has better results than GroupViT on all datasets. Moreover, our self-trained model, S-Seg+, provides an impressive " + }, + { + "bbox": [ + 55, + 338, + 296, + 445 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 55, + 338, + 296, + 445 + ], + "type": "text", + "content": " mIoU improvement over our base model S-Seg (42.6% vs 37.1% 3-avg. mIoU), suggesting the efficacy of self-training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 454, + 227, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 227, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 227, + 467 + ], + "type": "text", + "content": "4.4. Evaluation without background" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 472, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 579 + ], + "type": "text", + "content": "We also evaluate our model on the evaluation protocol commonly used for evaluating open-vocabulary models with annotated masks [11, 18, 19], where the background pixels are excluded in evaluation. We note that this setting is easier because background class is more diverse in appearance and often requires additional processing such as thresholding. Table 4 shows the results. Similar to the previous setting, our S-Seg and S-Seg+ models achieve competitive performance compared to earlier methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 588, + 153, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 153, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 153, + 599 + ], + "type": "text", + "content": "4.5. Ablation studies" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Self-training. We investigated the effectiveness of self-training for improving segmentation performance. To this end, we compared S-Seg and S-Seg+ on three datasets and evaluated the results using Figure 10. We found that self-training consistently improved the segmentation performance by a significant margin " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "(+5.5\\%)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " mIoU on average), regardless of the data size and test dataset. These results indicate that self-training is a reliable approach for enhancing the performance of S-Seg and can provide a de" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 319, + 238, + 383, + 312 + ], + "blocks": [ + { + "bbox": [ + 319, + 238, + 383, + 312 + ], + "lines": [ + { + "bbox": [ + 319, + 238, + 383, + 312 + ], + "spans": [ + { + "bbox": [ + 319, + 238, + 383, + 312 + ], + "type": "image", + "image_path": "af176d28d05d8e0752f64c9d368d072671bcb27600d65eac8fc42210b288303f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 383, + 238, + 438, + 312 + ], + "blocks": [ + { + "bbox": [ + 419, + 226, + 450, + 236 + ], + "lines": [ + { + "bbox": [ + 419, + 226, + 450, + 236 + ], + "spans": [ + { + "bbox": [ + 419, + 226, + 450, + 236 + ], + "type": "text", + "content": "(a) S-Seg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 238, + 438, + 312 + ], + "lines": [ + { + "bbox": [ + 383, + 238, + 438, + 312 + ], + "spans": [ + { + "bbox": [ + 383, + 238, + 438, + 312 + ], + "type": "image", + "image_path": "741b4de68b2938b0483b1ce2b153437fe5e017fd3acc6073b0eda4251538cb54.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 438, + 238, + 491, + 312 + ], + "blocks": [ + { + "bbox": [ + 438, + 238, + 491, + 312 + ], + "lines": [ + { + "bbox": [ + 438, + 238, + 491, + 312 + ], + "spans": [ + { + "bbox": [ + 438, + 238, + 491, + 312 + ], + "type": "image", + "image_path": "4f2f1e9acf4fd12ae534eaecef9a0d4c10d261590773ce851440e933c7350a2b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 491, + 238, + 547, + 312 + ], + "blocks": [ + { + "bbox": [ + 491, + 238, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 491, + 238, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 491, + 238, + 547, + 312 + ], + "type": "image", + "image_path": "bd2ece15d355ac639e64cc308f6fb2a0252fdad89f007022fc12d2c702857cee.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 319, + 330, + 383, + 404 + ], + "blocks": [ + { + "bbox": [ + 319, + 330, + 383, + 404 + ], + "lines": [ + { + "bbox": [ + 319, + 330, + 383, + 404 + ], + "spans": [ + { + "bbox": [ + 319, + 330, + 383, + 404 + ], + "type": "image", + "image_path": "e67a20d3295c34f1c94c2da4ad986ae53d7cba63160f48edbb119e9a7f017e17.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 422, + 554, + 476 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 476 + ], + "type": "text", + "content": "Figure 12. Scaling training data provide consistent gain in performance, with or without self-training. We train our model using different sizes of data: CC12M (12M), CC12M+CC3M (15M), and CC12M+CC3M+RedCaps (26M). We note a steady improvement in the performance as the data size increases." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 383, + 331, + 436, + 403 + ], + "blocks": [ + { + "bbox": [ + 416, + 318, + 453, + 327 + ], + "lines": [ + { + "bbox": [ + 416, + 318, + 453, + 327 + ], + "spans": [ + { + "bbox": [ + 416, + 318, + 453, + 327 + ], + "type": "text", + "content": "(b) S-Seg+" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 331, + 436, + 403 + ], + "lines": [ + { + "bbox": [ + 383, + 331, + 436, + 403 + ], + "spans": [ + { + "bbox": [ + 383, + 331, + 436, + 403 + ], + "type": "image", + "image_path": "0254ff69ca81fa9847903f84bb286c9bf9754b2ac08670b2dcbbfa39d9c5af77.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 406, + 474, + 414 + ], + "lines": [ + { + "bbox": [ + 406, + 406, + 474, + 414 + ], + "spans": [ + { + "bbox": [ + 406, + 406, + 474, + 414 + ], + "type": "text", + "content": "Size of training data " + }, + { + "bbox": [ + 406, + 406, + 474, + 414 + ], + "type": "inline_equation", + "content": "(\\times 10^{6})" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 436, + 331, + 490, + 403 + ], + "blocks": [ + { + "bbox": [ + 436, + 331, + 490, + 403 + ], + "lines": [ + { + "bbox": [ + 436, + 331, + 490, + 403 + ], + "spans": [ + { + "bbox": [ + 436, + 331, + 490, + 403 + ], + "type": "image", + "image_path": "62725b948e8f5fe8c6384d84f2f2233260535b92e14d179bc2b1414011bc6f6d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 490, + 331, + 544, + 404 + ], + "blocks": [ + { + "bbox": [ + 490, + 331, + 544, + 404 + ], + "lines": [ + { + "bbox": [ + 490, + 331, + 544, + 404 + ], + "spans": [ + { + "bbox": [ + 490, + 331, + 544, + 404 + ], + "type": "image", + "image_path": "54865075f1660a72c082a35ca530a6dee88b40b424f72a1877bb53b9c974b31a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 487, + 497, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 497, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 497, + 499 + ], + "type": "text", + "content": "sirable complement for further improvement." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 499, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 554, + 594 + ], + "type": "text", + "content": "Data scalability. To evaluate the scalability of our method, we trained S-Seg and S-Seg+ using three datasets of increasing sizes: 12M, 15M, and 26M. The results of the experiments are presented in Figure 12. We observed that both models achieve significant improvements in performance across all three testing datasets as the amount of data increased, suggesting that our method scales well with larger datasets." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 605, + 387, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 387, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 387, + 618 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 626, + 554, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 554, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 554, + 709 + ], + "type": "text", + "content": "To summarize, we propose S-Seg, a simple and intuitive framework that enables accurate and generalizable open-vocabulary segmentation. Our algorithm directly trains for pixel-level feature and language alignment, and does not require manual segmentation annotations or extensive pretraining. We hope that our simple yet effective approach will serve as a solid baseline for future research." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30228" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 147 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 294, + 191 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 236 + ], + "type": "text", + "content": "[4] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 2, 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 281 + ], + "type": "text", + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "type": "text", + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 363, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 294, + 396 + ], + "type": "text", + "content": "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. 2021. 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 397, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 294, + 441 + ], + "type": "text", + "content": "[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 57, + 442, + 294, + 485 + ], + "type": "text", + "content": "[10] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431, 2021.6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 487, + 294, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 294, + 519 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 294, + 519 + ], + "type": "text", + "content": "[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 522, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 576 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 578, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 620 + ], + "type": "text", + "content": "[13] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, 2002. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 623, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 623, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 623, + 294, + 655 + ], + "type": "text", + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2009. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "text", + "content": "[17] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In ECCV, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "text", + "content": "[18] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 229 + ], + "type": "text", + "content": "[19] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, and Yunchao Wei. Global knowledge calibration for fast open-vocabulary segmentation. In ICCV, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[21] Tapas Kanungo, David M Mount, Nathan S Netanyahu, Christine D Piatko, Ruth Silverman, and Angela Y Wu. An efficient k-means clustering algorithm: Analysis and implementation. IEEE TPAMI, 2002. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 320, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 553, + 353 + ], + "type": "text", + "content": "[22] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "text", + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 410, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 410, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 410, + 553, + 442 + ], + "type": "text", + "content": "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[25] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 489, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 521 + ], + "type": "text", + "content": "[26] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "text", + "content": "[27] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 2, 3, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[30] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In ECCV, 2022. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 691, + 553, + 713 + ], + "type": "text", + "content": "[31] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30229" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. 2023. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[33] Fangyun Wei Han Hu Xiang Bai Mengde Xu, Zheng Zhang. Side adapter network for open-vocabulary semantic segmentation. 2023. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[34] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[35] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 255, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 286 + ], + "type": "text", + "content": "[36] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 288, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 342 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "text", + "content": "[38] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. In ICCV, 2023. 2, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 389, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 294, + 442 + ], + "type": "text", + "content": "[39] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 509 + ], + "type": "text", + "content": "[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2018, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 533 + ], + "type": "text", + "content": "[41] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 569, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 632 + ], + "type": "text", + "content": "[43] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[44] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[45] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 387 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[46] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[47] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "text", + "content": "[48] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[49] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "text", + "content": "[50] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "type": "text", + "content": "[51] Haobo Yuan, Xiangtai Li, Chong Zhou, Yining Li, Kai Chen, and Chen Change Loy. Open-vocabulary sam: Segment and recognize twenty-thousand classes interactively. arXiv preprint arXiv:2401.02955, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 353 + ], + "type": "text", + "content": "[52] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[53] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. *NeurIPS*, 2020. 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30230" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_content_list.json b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c290880723612a8990d4d85cb0d198ade64f8a6f --- /dev/null +++ b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_content_list.json @@ -0,0 +1,1662 @@ +[ + { + "type": "text", + "text": "Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis", + "text_level": 1, + "bbox": [ + 145, + 130, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiapeng Zhu1,2", + "bbox": [ + 166, + 179, + 285, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ceyuan Yang3 \nYifei Zhang1,4", + "bbox": [ + 305, + 180, + 423, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kecheng Zheng1 \nQifeng Chen2", + "bbox": [ + 442, + 180, + 575, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yinghao Xu $^{1,3}$ \nYujun Shen $^{1}$", + "bbox": [ + 586, + 180, + 712, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zifan Shi $^{1,2}$", + "bbox": [ + 733, + 180, + 828, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Ant Group", + "bbox": [ + 191, + 220, + 287, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ HKUST", + "bbox": [ + 326, + 220, + 403, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ CUHK", + "bbox": [ + 442, + 220, + 509, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Shanghai Jiao Tong Univeristy", + "bbox": [ + 549, + 220, + 800, + 241 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dfb66adaf22deacd14135b67a068e138504a8804a86a1e608be3db7eb3d0d558.jpg", + "image_caption": [ + "A black and white photo of a dog with big eyes." + ], + "image_footnote": [], + "bbox": [ + 94, + 275, + 433, + 536 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee1e7d9be98346aee67f0576d33f31be2ed149d288837b2ab934e57f0214eba1.jpg", + "image_caption": [ + "A photo of a colorful sports car, with mountain background." + ], + "image_footnote": [], + "bbox": [ + 434, + 275, + 588, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a51ccde226dac9754610c73802c93c5e3aee485d9173acc14ce01cd2a422f7f0.jpg", + "image_caption": [ + "Sunrise over the clouds in the mountains.", + "Figure 1. Example results at $512 \\times 512$ resolution synthesized by our proposed Aurora, a large-scale GAN-based text-to-image generator." + ], + "image_footnote": [], + "bbox": [ + 436, + 412, + 588, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/755b522bbd2b2962a16f0eaba4182c4192565977b33c08cea7be6f054b48019f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 275, + 741, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/586a04845d97e4ef594442235b7360a7b200931700e30b02ab5db169b928643d.jpg", + "image_caption": [ + "A steaming plate of classic fried rice with vegetables, eggs.", + "A coral reef with starfish, sea anemones and fish." + ], + "image_footnote": [], + "bbox": [ + 589, + 411, + 741, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/11b332b209e7fb2ea9a32c89e5f98f4dc57f087c55492b92ab4075321a45f89e.jpg", + "image_caption": [ + "Oil painting of a flower garden, distant cottage, soft sunlight." + ], + "image_footnote": [], + "bbox": [ + 743, + 275, + 895, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/889c1bf433a3700bace27f2c773a2768c62b5167e7d91dba223dff1f550c8084.jpg", + "image_caption": [ + "A pencil sketch of a woman with deep, almond-shaped eyes in." + ], + "image_footnote": [], + "bbox": [ + 745, + 412, + 895, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 597, + 326, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the difficulty in scaling up, generative adversarial networks (GANs) seem to be falling out of grace with the task of text-conditioned image synthesis. Sparsely activated mixture-of-experts (MoE) has recently been demonstrated as a valid solution to training large-scale models with limited resources. Inspired by this, we present Aurora, a GAN-based text-to-image generator that employs a collection of experts to learn feature processing, together with a sparse router to adaptively select the most suitable expert for each feature point. We adopt a two-stage training strategy, which first learns a base model at $64 \\times 64$ resolution followed by an upsampler to produce $512 \\times 512$ images. Trained with only public data, our approach encouragingly closes the performance gap between GANs and industry-level diffusion models, maintaining a fast inference speed. We release the code and checkpoints here to facilitate the community for further development.", + "bbox": [ + 86, + 632, + 485, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 597, + 643, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As one of the most popular tasks in the field of content generation, text-conditioned image synthesis [3, 12, 19, 34, 60, 61, 65, 67, 100, 108] allows users to customize a picture with natural language instructions. Among diverse types of generative models, such as variational autoencoders (VAEs) [42], autoregressive models (ARs) [97, 98], and generative adversarial networks (GANs) [23], diffusion models [15, 28, 87] have blossomed into a dominant solution with a thriving community (e.g., stable diffusion [65]), leaving few efforts on improving other frameworks.", + "bbox": [ + 511, + 625, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we put our focus on GANs, which are usually criticized for training instability [1, 53] and limited mode coverage [70]. However, we argue that the task of text-to-image (T2I) generation is more like a translation task (i.e., shifting an informative data distribution to another) instead of a pure generation task (i.e., modeling the observed distribution with random noise as the input), and accordingly that the introduction of the text condition", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "can already help alleviate the above two issues to some extent, which is also empirically supported by some recent studies [34, 72]. Even so, the difficulty in scaling up the model size still manifests itself as a major drawback of GANs. Concretely, unlike diffusion models that produce an image through an iterative process [15, 28, 65], the generator of a GAN typically adopts a feedforward network to render an image. Considering the limited computational resources (e.g., GPU memory) in practice, directly learning a sufficiently large feedforward model remains challenging.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The advent of sparse mixture-of-experts (MoE) offers a promising way to enlarge the model scale with affordable computing cost [20, 78]. The rationale behind is to employ multiple expert modules yet only activate one or a few of them for each feedforward execution. That way, the growth of the number of modules could expand the model capacity but barely increase the number of computations.", + "bbox": [ + 89, + 244, + 480, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We decently introduce the philosophy of sparse MoE into the design of GAN generators and propose Aurora for fast T2I generation. Specifically, given an intermediate feature map at a particular level of degree, we engage a set of experts, each of which is originally designed to process a single-point feature, as well as a highly sparse router to help select the most suitable expert for each feature point. It is noteworthy that, different from most existing spare MoE methods [20, 32, 78] that determine the routing path simply based on the information from the input feature map, our router also takes into account the text-integrated global latent code such that it can adaptively make the best decision via fully considering the sampling stochasticity and the text condition. Additionally, Aurora employs a two-stage pipeline [34, 67] involving a base generator and an upsampler. The base generator creates images at a resolution of $64 \\times 64$ , are then enhanced by an upsampler to produce final images at $512 \\times 512$ . Our base generator learned on LAION2B-en [74] and COYO-700M [8] achieves 6.2 zero-shot FID [27] on MS COCO [46] at the resolution of $64 \\times 64$ , surpassing other existing methods. The upsampler further refines the image to a final resolution of $512 \\times 512$ pixels, achieving a zero-shot FID score of 8.74 on MS COCO, which sets a new benchmark for GANs.", + "bbox": [ + 91, + 354, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides the fast inference speed, GANs also enjoy a well-studied latent space [33, 81, 105] and the high flexibility to incorporate other priors (e.g., involving the volumetric rendering pipeline for 3D-aware image synthesis) [9, 10, 22, 24, 54, 57, 75, 76, 83-86, 103]. However, the lack of an open-sourced large-scale GAN model hinders such a powerful tool from being adequately developed and analyzed. We thus release the code and checkpoints to facilitate the research community and plan for continuous improvement from the perspectives of both performance and functionality. We believe that our work could encourage more studies in the field of visual content generation.", + "bbox": [ + 89, + 719, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative adversarial networks (GANs). Formulated as a two-player game between a generator and a discriminator, GANs [23] are designed to model the observed data distribution through adversarial training. Thanks to the flexible formulation (i.e., there are no particular assumptions on the data format or on the generation process), GANs have shown great potential in a variety of visual generative tasks, such as image generation [7, 36-39], image editing [4, 25, 33, 80, 81, 90, 101, 102, 105, 113-115], video generation [68, 96, 104], and 3D-aware image synthesis [9, 10, 24, 54, 75, 83, 103]. Some studies also explore conditional GANs [52], which incorporate signals like class labels [7, 71, 110], texts [34, 72, 93, 112], and reference images [31, 43, 107] into the generator for interactive control. Despite the aforementioned success, GANs usually struggle in learning from large-scale, diverse datasets (e.g., LAION [74]) due to the difficulty in scaling up [34].", + "bbox": [ + 511, + 114, + 903, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-conditioned image synthesis. With the aid of visual-language models like CLIP [59], text-to-image (T2I) generation [2, 3, 11, 12, 19, 26, 34, 58, 60, 61, 65, 67, 108] manages to decode an open-vocabulary description to a vivid image, and hence allows users to visualize their creativity and imagination. Towards such a challenging task, existing efforts usually opt to trade computing efficiency for visual quality. For example, autoregressive models [16, 21, 26, 60, 100, 108] and diffusion models [3, 12, 19, 55, 58, 61, 65, 67] have demonstrated impressive results through an iterative generation process, which can be time-consuming and costly to make a synthesis. Differently, GANs employ a feedforward network as the generator and thus naturally support fast generation without having to use an acceleration [40, 48, 88] or distillation [69, 73, 89] algorithm. Some recent attempts [34, 72] have confirmed the promise of GANs in text-to-image synthesis yet left the issue of enlarging the model capacity unsolved. In addition, there currently lacks a publicly available GAN-based T2I model in the community, making it difficult to fully study the properties of GANs, let alone integrate the advantages of GANs into other types of generative models.", + "bbox": [ + 511, + 387, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture-of-experts (MoE). Sharing a similar philosophy with ensemble [56, 64], MoE [32] learns a set of expert modules to process the input signal, together with a gating module (also known as a router) to help combine the outputs of different experts. Under such a design, each expert is expected to focus on the learning of a sub-region of the input space, while the router is responsible for the weight assignment across experts [18]. Sparsely activated MoE [17, 20, 78] distinguishes itself from other variants of MoE because of the discrete routing mechanism. By selecting only one or a few experts from a large candidate pool for each incoming example, sparse MoE offers a", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18412", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "feasible solution to scaling up the model size with barely increased computational cost [20, 63, 79]. In this work, we introduce the idea of sparse MoE into the design of a GAN generator and adequately adapt the routing mechanism for the task of text-to-image generation by taking into account the text-integrated stochasticity.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 196, + 181, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recall that we aim to develop an efficient text-to-image (T2I) generator based on the framework of GANs [23], which consists of a generator and a discriminator competing with each other. On the one hand, the generator takes in a global latent code, $\\mathbf{z} \\in \\mathbb{R}^{128}$ , along with a text description (e.g., a sentence), $\\mathbf{c}$ , and outputs an image, $\\mathbf{x} = G(\\mathbf{z},\\mathbf{c})$ . On the other hand, the discriminator takes in an image together with an associated text and differentiates synthesized samples from real samples using a realiznescore, $D(\\mathbf{x},\\mathbf{c})$ . As discussed above, the major contribution of this work is to enlarge the capacity of the base generator; therefore, we directly borrow the discriminator architecture from the prior art [34]. In the following, we first introduce the common unit blocks utilized in both the base generator and the upsampler (Sec. 3.1). We then provide a detailed overview of the architecture of our base generator (Sec. 3.2) and upsampler (Sec. 3.3), followed by a discussion of the supervision signals and training strategy (Sec. 3.4).", + "bbox": [ + 89, + 220, + 483, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unit Generative Block", + "text_level": 1, + "bbox": [ + 89, + 503, + 299, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Like most modern deep models [37, 99], our base generator and upsampler follow the principle of block-wise design. As shown in Fig. 2, a generative unit stacks a convolution block and an attention block to process the input feature map, $\\mathbf{f}_{in}$ , where the attention block adopts the sparse routing mechanism for scaling up. Additionally, in this section, $\\mathbf{w}$ refers to the intermediate latent code and will be introduced in detail in Sec. 3.2 and Sec. 3.3.", + "bbox": [ + 89, + 526, + 483, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text encoder. We employ CLIP [59] as the text encoder, which transforms the input text, $\\mathbf{c}$ , into both a global token, $\\mathbf{t}_g$ , and a token sequence, $\\mathbf{t}_{seq}$ . Specifically, $\\mathbf{t}_g$ is the pooled text embedding that encapsulates the meaning of the entire sentence, while $\\mathbf{t}_{seq}$ is derived from the penultimate layer of the text encoder, capturing information for each individual word in the text. We formulate these as follows", + "bbox": [ + 89, + 647, + 483, + 753 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {t} _ {g}, \\mathbf {t} _ {\\text {s e q}} \\right\\} = \\operatorname {C L I P} (\\mathbf {c}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 766, + 482, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Convolution block. Latent modulation [37], which is developed from adaptive instance normalization (AdaIN) [29], has demonstrated its effectiveness in the generation field and hence serves as a basic operation in GAN-based models [34, 71, 72]. Such a technique is further improved by integrating it into the convolution operation with demodulation (i.e., modulated convolution,", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ModConv) [38] and by equipping it with learnable feature deformation (i.e., modulated transformation module, MTM) [106]. We employ two MTMs in the convolution block and construct the block using skip connection. Consequently, the output of the convolutional block is given by", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\text {c o n v}} = \\operatorname {M T M} \\left(\\operatorname {M T M} \\left(\\mathbf {f} _ {\\text {i n}}, \\mathbf {w}\\right), \\mathbf {w}\\right) + \\mathbf {f} _ {\\text {i n}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 189, + 906, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, the learnable offsets in MTM [106] are only introduced when the feature resolution is equal to or less than $16 \\times 16$ .", + "bbox": [ + 511, + 220, + 905, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention block. We can tell from Eq. (2) that, in the convolution block, the text condition, $\\mathbf{c}$ , only interacts with the generative features through the global token, $\\mathbf{t}_g$ , with $\\mathbf{w}$ in Eq. (7) as an intermediary. To also make sufficient use of the text token sequence, $\\mathbf{t}_{seq}$ , we introduce an attention block for text-image interaction, which is widely used in multi-modality learning [45, 92]. Specifically, the attention block consists of a self-attention layer, a cross-attention layer, and a feedforward network [99]. We thereby have", + "bbox": [ + 511, + 266, + 906, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {s a} = \\operatorname {S e l f A t t n} \\left(\\mathbf {f} _ {\\text {c o n v}}, \\mathbf {w}\\right) + \\mathbf {f} _ {\\text {c o n v}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 411, + 903, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {c a} = \\operatorname {C r o s s A t t n} \\left(\\mathbf {f} _ {s a}, \\mathbf {t} _ {\\text {s e q}}\\right) + \\mathbf {f} _ {s a}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 430, + 903, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {f f n} = \\mathrm {F F N} \\left(\\mathbf {f} _ {c a}\\right) + \\mathbf {f} _ {c a}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 449, + 903, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathsf{FFN}$ in Eq. (5) processes each feature point individually using a two-layer MLP. Here, we follow the prior work [34, 44] that employs $\\ell 2$ -distance instead of the dot product for attention logits to stabilize training [41].", + "bbox": [ + 511, + 473, + 905, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiple experts and sparse routing. Recall that the main motivation of this work is to enlarge the capacity of the base generator in GANs. One straightforward solution is to make the network deeper or wider. However, it may cause an out-of-memory issue. Inspired by the practice in natural language processing (NLP) [20], we incorporate the philosophy of sparse MoE [78] into the design of our unit generative block to enlarge the base generator capacity. Concretely, we treat the FFN in Eq. (5) as an expert, which specializes in processing a single-point feature, and engage $N$ experts in total, $\\{ \\mathsf{FFN}_j \\}_{j=1}^N$ , forming a candidate pool. We then employ a sparse router to help select the most suitable expert for each feature point. Unlike existing sparse MoE methods [17, 20, 78], our router adaptively makes its decision by taking into account not only the input feature but also the text-integrated global latent code, $\\mathbf{w}$ , which better matches the target of T2I generation (i.e., decoding the text condition and the sampling stochasticity to an image). Under such a routing strategy, we can reformulate Eq. (5) as", + "bbox": [ + 511, + 532, + 906, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {f f n} ^ {(k)} = \\operatorname {F F N} _ {j} \\left(\\mathbf {f} _ {c a} ^ {(k)}\\right) + \\mathbf {f} _ {c a} ^ {(k)} \\text {s . t .} j = \\operatorname {R o u t e r} \\left(\\mathbf {f} _ {c a} ^ {(k)}, \\mathbf {w}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 842, + 903, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, the superscript $(k)$ denotes the $k$ -th feature point. In addition to increasing the capacity of the generator, MoE", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18413", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4903de1974a77a4a90d05484159018fa3d202c7baa518e077f5ffb5e3b8d7af.jpg", + "image_caption": [ + "Figure 2. Illustration of base generator in Aurora. First, we utilize a pretrained CLIP text encoder to obtain the text embeddings $\\mathbf{t}_g$ and $\\mathbf{t}_{seq}$ . Then, starting from the Fourier feature, the generator processes it through several unit generative blocks. Each unit block is composed of a convolution block and an attention block. There are two MTM [106] layers within the convolution block. In the attention block, the FFN acts as the expert, specializing in processing a single-point feature. A total of $N$ experts are assembled to form a candidate pool, from which selections are made by the router based on the input features and the latent code $\\mathbf{w}$ ." + ], + "image_footnote": [], + "bbox": [ + 93, + 95, + 911, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "also frees the generator from modeling complex data distributions by assigning different objects or parts of an image to specialized experts, reducing the learning difficulty.", + "bbox": [ + 88, + 428, + 480, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Base Generator", + "text_level": 1, + "bbox": [ + 89, + 481, + 246, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The base generator comprises two components: a mapping network and a synthesis network. Both networks accept text and latent code as inputs, which we will introduce next.", + "bbox": [ + 89, + 503, + 482, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mapping network. As pointed out by StyleGAN [37], utilizing a learnable MLP-based mapping network on the latent code, $\\mathbf{z}$ , could create a more disentangled latent space, $\\mathcal{W}$ , than the native latent space, $\\mathcal{Z}$ . In the class-conditional generation, a class label is first concatenated with the latent code $\\mathbf{z}$ before being passed to the mapping network [37, 71]. Under the T2I setting, it is necessary to shift from class conditioning to text conditioning to make the sampling process better aware of the text information. Hence, we concatenate $\\mathbf{t}_g$ with the sampled latent code $\\mathbf{z}$ for further processing, as follows:", + "bbox": [ + 89, + 549, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} = \\operatorname {M L P} \\left(\\operatorname {c o n c a t} \\left(\\mathbf {z}, \\mathbf {t} _ {g}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 724, + 482, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}$ stands for the more disentangled latent code.", + "bbox": [ + 89, + 750, + 447, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthesis network. The synthesis network starts with Fourier features [39] and is subsequently followed by a series of convolutional blocks, attention blocks, including the MoE, and finally outputs the image at a resolution of $64 \\times 64$ , as shown in Fig. 2. The $\\mathbf{w}$ generated by the mapping network is used to modulate the convolutional and self-attention layers, and it also functions in the MoE to select the appropriate expert for each input feature point. Additionally, the local token $\\mathbf{t}_{seq}$ , is incorporated into the", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cross-attention layers as the textual condition to further enhance text-image alignment. The generator produces RGB images at each resolution and employs output skip connections [38]. Unlike StyleGAN2 [38], which only utilizes the highest resolution image as the final output, our approach takes advantage of all generated images, similar to MSG-GAN [35] and GigaGAN [34]. Each image is then used to calculate the training loss, mirroring the progressive training approach described in [36], which can improve the final performance.", + "bbox": [ + 511, + 428, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Upsampler", + "text_level": 1, + "bbox": [ + 511, + 590, + 635, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The base generator produces images at a resolution of $64 \\times 64$ . To create high-resolution images, we train a text-conditioned super-resolution model that upsamples the images generated by the base generator to a resolution of $512 \\times 512$ . The architectures of the upsampler are derived from the base model with specific modifications. First, additional convolution blocks are integrated to process the input image, resulting in a final architecture resembling a U-Net. Second, we set the number of experts to 1 based on empirical evidence indicating that this model does not require an extensive parameter count. Third, instead of acquiring the global code (i.e., $\\mathbf{w}$ in base generator) from the mapping network, we obtain it from the encoder part of the U-Net, which is then concatenated with $\\mathbf{t}_g$ to modulate the decoder part of the U-Net. Additionally, to enhance memory efficiency and increase the speed of the upsampler, following Imagen [67], we reverse the order of the downsampling and upsampling operations in our U-Net architecture compared to the approach used in prior art [66].", + "bbox": [ + 511, + 613, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18414", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5706d4f00532f59629316fadc9e3819270fe63690f104f498b29c7e586e49a97.jpg", + "image_caption": [ + "A lighthouse under a starry night sky." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 356, + 267 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bb6f7377e22e5a8ba20bdf93d04c2a20bac8fd9c4e41bed32b3a707318e03591.jpg", + "image_caption": [ + "A painting of flowers in a blue vase on a table." + ], + "image_footnote": [], + "bbox": [ + 362, + 90, + 624, + 267 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d34b245c552d56f490fc8859aa59a338a901dafdf94e8e72154eee7e74e65196.jpg", + "image_caption": [ + "An oil painting depicting a large crowd of people in front of a church." + ], + "image_footnote": [], + "bbox": [ + 630, + 90, + 892, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/09e08fc0b6c22dab3c7342bb1f664964fab04bd310927ca217c2a1f90d861cab.jpg", + "image_caption": [ + "A road that is surrounded by trees with yellow leaves." + ], + "image_footnote": [], + "bbox": [ + 96, + 299, + 357, + 477 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/04cf4a994a2d64b21f15b47f024099ddc80521927791bf4ae322e1eb92baa0ea.jpg", + "image_caption": [ + "Sunset casts a golden glow on \nmajestic mountain peaks." + ], + "image_footnote": [], + "bbox": [ + 364, + 299, + 625, + 477 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad99aa22ecbe136b116b845fc712ba40d7effc81e016eb176c1aef92e5fbd166.jpg", + "image_caption": [ + "Landscape, lake, buildings, bridge, Chinese ink style.", + "Figure 3. Diverse results generated by Aurora, In each group, the images on the left are randomly generated by our base generator at the resolution of $64 \\times 64$ . We also highlight an example at $512 \\times 512$ on the right to demonstrate high-quality synthesis capability." + ], + "image_footnote": [], + "bbox": [ + 630, + 300, + 892, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is noteworthy that the upsampler is a super-resolution model primarily trained on real images. Therefore, it is crucial to minimize the gap between real images and those generated by the base generator. We address this gap through three aspects. First, we augment the input data by introducing additional augmentations, such as Gaussian noise and Gaussian blurring. Second, the skip connection is removed since it provides linearity between the input and output [49], which helps reduce discrepancies when transitioning from real to synthesized images. Third, we increase the downsampling rate beyond the conventional rate of 8, lowering the feature map resolution from the standard $8 \\times 8$ [34, 66], which produces more abstracted features by omitting details in the input image, further bridging the gap between real and synthesized images.", + "bbox": [ + 89, + 570, + 482, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Learning Objectives and Training", + "text_level": 1, + "bbox": [ + 89, + 815, + 387, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The primary focus of this work is to scale up the model size of a GAN generator instead of exploring a better supervision signal for GAN training. Thus, we directly inherit some well-studied loss functions from prior arts.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Still, to make the presentation self-contained, we list below all the objectives used in the training phase.", + "bbox": [ + 511, + 570, + 903, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial loss. Adversarial training [23] is the core concept shared by all GANs and owns many variants in the literature [1, 5, 50]. We adopt the approach from GigaGAN [34], implementing a multi-scale input and multi-scale output logistic non-saturating loss [37] with $R_{1}$ penalty [51] as the regularizer.", + "bbox": [ + 511, + 604, + 905, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Matching-aware loss. Unlike the discriminator in unconditional GANs that rejects a sample by merely considering its plausibility, the discriminator in T2I GANs is also expected to take into account the alignment between the condition and the synthesized result. We follow prior arts [34, 62, 109] to employ a matching-aware discriminator and introduce the matching-aware loss that urges the discriminator to reject mismatched text-image pair.", + "bbox": [ + 511, + 699, + 905, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-level CLIP loss. CLIP loss [59] follows the idea of contrastive learning [13], and forces the feature of a generated image close to the feature of its associated text yet away from the feature of irrelevant text. We introduce a multi-level CLIP loss that measures CLIP loss at different", + "bbox": [ + 511, + 824, + 908, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image resolutions, promoting increased text-image similarity. Such a supervision signal not only stabilizes training but also complements the matching-aware loss, aiding the generator in more effectively decoding the text condition.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MoE loss. Recall that we introduce a collection of experts to enlarge the model capacity and expect them to process the input feature in various ways. For each network execution, only one expert will be activated for a feature point. To balance the workload across experts (i.e., to avoid that some experts are never activated), we introduce the MoE loss following Switch Transformer [20].", + "bbox": [ + 88, + 151, + 480, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. We first train our base generator using the aforementioned losses. When training the upsampler, we incorporate an additional LPIPS loss [111], MoE loss is omitted because we set the number of experts to one for the upsampler. Furthermore, similar to the training procedure of the autoencoder in StableDiffusion [66], the adversarial loss is not employed at the beginning of the upsampler's training process. Once the training of the upsampler on real images is complete, we proceed to finetune it using synthesized images generated by the base generator with adversarial loss only. It is noteworthy that our approach adopts the mechanisms of both adversarial learning and sparse MoE, yet we do not encounter any instability throughout the entire training process, whether in the training of the base generator or the upsampler.", + "bbox": [ + 91, + 257, + 482, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 496, + 222, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of our proposed Aurora both quantitatively (Sec. 4.2) and qualitatively (Sec. 4.3). Additionally, we provide detailed analyses to deepen the reader's understanding of text-to-image (T2I) GANs, discuss the significance of our project, and highlight how our open-sourced code and checkpoints can benefit the research community (Sec. 4.4).", + "bbox": [ + 89, + 521, + 482, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 637, + 287, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The training data is sourced from a diverse set of public datasets, including LAION2B-en [74], COYO-700M [8], CC3M [77], and YFCC15M [95]. All datasets are filtered based on the following criteria: clip score $>0.3$ , resolution $>512$ , aesthetic score $>5.0$ , and an estimated watermark probability $<0.5$ , which results in a total of 120M text-image pairs for training. Additionally, we use InstructBLIP [14] to obtain better captions for training our model. We do not use any self-collected data to simply pursue the improvement of synthesis quality. We believe that our open-sourced codebase can facilitate learning a more capable model using cleaner and higher-quality data.", + "bbox": [ + 89, + 657, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. We develop our algorithm using Hammer [82]. The base generator was trained for one week on 320 NVIDIA A100 GPUs, while the upsampler underwent training for five days using 128 NVIDIA A100 GPUs. We", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/914e0e45e38658032d1bcc9d90b18e10078a917a6321eb4ab2910613a3dc6fc6.jpg", + "table_caption": [ + "Table 1. Ablation study of our base model, which targets generating images at $64 \\times 64$ resolution. FID [27] is employed as the evaluation metric, where a smaller number indicates better performance. “ $\\mathrm{FID}_{10K}$ ” reports the results on the training set, while “zero-shot $\\mathrm{FID}_{30K}$ ” reports the results on MS COCO [46]." + ], + "table_footnote": [], + "table_body": "
ModelZero-shot FID30KFID10K# Params.
Base (64 GPUs)22.4420.770.367B
+ Better caption16.7214.310.367B
+ MoE13.8112.800.981B
+ ConvNeXt12.0311.100.986B
Scaling (128 GPUs)7.638.540.986B
Scaling (320 GPUs)6.207.960.986B
", + "bbox": [ + 516, + 162, + 903, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "utilized the AdamW optimizer [47] with hyper-parameters set to $\\beta_{1} = 0$ , $\\beta_{2} = 0.99$ , and a weight decay of $1e^{-5}$ . The base generator and discriminator were jointly optimized with a learning rate of 0.0025, whereas the upsampler was trained with a learning rate of 0.001. We utilize OpenCLIP ConvNeXt-XXLarge [30] as our text encoder and for computing the CLIP loss during training. Additional training details are provided in the supplementary materials.", + "bbox": [ + 511, + 304, + 906, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 438, + 705, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this part, we quantitatively compare our Aurora with existing T2I generators to give readers an overall picture of its synthesis performance. Following prior arts [7, 34, 37, 38, 71, 72], we adopt FID [27] as the metric, which evaluates a generator at the distribution level. Here, we compute FID on two datasets to evaluate our model, one is \"FID $_{10K}$ \", which is obtained by comparing the inception features [91] of $10K$ synthesized images and $10K$ real images randomly sampled from the training set, while the other \"zero-shot $\\mathrm{FID}_{30K}$ \" is obtained by comparing the inception features of $30K$ synthesized images and the full validation set of MS COCO.", + "bbox": [ + 511, + 460, + 906, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Base generator study. In Tab. 1, we report the evaluation results on two different data domains under different configurations of the base generator. We begin by implementing our base generator following the StyleGAN architecture and incorporate attention blocks as described in Sec. 3.1. Besides, we set the number of experts to 1 for this baseline model, resulting in 367M parameters. First, the baseline model is trained using the original captions, resulting in a zero-shot $\\mathrm{FID}_{30K}$ score of 22.44. Second, we retrain the model with re-captioned data, which significantly improves its performance. Following this, we increase the number of experts in the attention blocks, expanding the model size to 981M parameters. Additionally, we enlarge the CLIP text encoder by replacing the CLIP ViT-L/14 with OpenCLIP ConvNeXt-XXLarge, further increasing the model parameters. These two modifications result in our final base generator, which has 986M parameters and", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18416", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3d124f80939e902da3e5bf57d1e52cd9be54268cc9fcd7727b9325f4f06e413.jpg", + "image_caption": [ + "Figure 4. Synthesized results through text prompt interpolation, where we fix the global latent code, $\\mathbf{z}$ , and interpolate the text tokens, $\\{\\mathbf{t}_{seq},\\mathbf{t}_g\\}$ , extracted from two different text conditions, $\\mathbf{c}$ ." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 890, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5e46dd65bc7a32b65e988a1357534b40f1a6ea0b00bca1238c30f445fb30823.jpg", + "table_caption": [ + "Table 2. Quantitative comparison between Aurora and existing methods at $64 \\times 64$ resolution. “*” means that we downsample the samples using Lanczos [72] for a fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodTypeFID10KFID30K# Params.
eDiff-I* [2]DM-7.609.10B
SD* [65]DM-8.400.94B
LAFITE [112]GAN-14.8075M
StyleGAN-T [72]GAN-7.30-
GigaGAN [34]GAN9.18-0.65B
Aurora (ours)GAN7.966.200.98B
", + "bbox": [ + 93, + 415, + 478, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exhibits additional performance improvements. Finally, we conduct experiments by training our model using more GPUs, while keeping all other configurations the same. We observe consistent performance improvements with the use of additional GPUs and have not yet encountered any performance saturation related to the number of GPUs utilized. Our final base generator, trained using 320 GPUs, achieves a zero-shot $\\mathrm{FID}_{30K}$ score of 6.2.", + "bbox": [ + 88, + 564, + 482, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with existing methods. We first compare our base generator with existing methods at the resolution of $64 \\times 64$ in Tab. 2. The evaluation results demonstrate that Aurora consistently outperforms all other methods across both metrics at this resolution, irrespective of the model type. Tab. 3 presents the comparison results with the existing methods on higher resolution. According to Tab. 3, our method not only achieves the best performance among GANs in terms of the FID but also has the largest model size. However, it lags behind diffusion models and autoregressive models. We conclude that more powerful superresolution techniques or larger models are necessary, and exploring how to train a high-resolution GAN on a larger model remains worthwhile.", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e07d03090364885c1e0b5379b49857caeb9c0ac104dec08aec09c3ae740b0e7.jpg", + "table_caption": [ + "Table 3. Quantitative comparison between our Aurora and existing text-to-image generators at $256 \\times 256$ resolution. We report the model size, total images seen during training, \"zero-shot $\\mathrm{FID}_{30K}$ \", and inference speed. \"\\*\" means that we downsample the generated samples using Lanczos [72] for a fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodType# Params.# ImagesFID30KSpeed
DALL-E [60]AR12.0B1.54B27.50-
Parti-20B [108]AR20.0B3.69B7.23-
GLIDE [55]DM5.0B5.94B12.2415.0s
SD* [65]DM0.9B3.16B8.592.9s
Imagen [67]DM3.0B15.36B7.279.1s
eDiff-I[2]DM9.1B11.47B6.9532.0s
LAFITE [112]GAN75M0.03B26.940.02s
StyleGAN-T [72]GAN1.0B0.25B13.900.1s
GALIP [94]GAN0.32B0.012B12.540.04s
GigaGAN* [34]GAN1.0B0.98B9.030.13s
Aurora* (ours)GAN1.32B0.12B8.740.16s
", + "bbox": [ + 516, + 443, + 903, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 666, + 694, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-to-image synthesis. We showcase some visual results in Fig. 3 under diverse text conditions. We can tell that, besides performing well on image synthesis within a narrow distribution (e.g., human faces, horses, or bedrooms), GANs are also capable of open-vocabulary T2I generation, allowing users to play with their creativity and imagination. It is noteworthy that our generator excels at generating scenic images (see Figs. 1 and 3), producing pictures that are more realistic than those generated by diffusion models (see supplementary materials for the comparison).", + "bbox": [ + 509, + 688, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text interpolation. Our generator produces an image with two independently sampled inputs, namely, a global latent code and a text prompt. Here, we investigate the properties of these two different input spaces through interpolation.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18417", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/974d4041d56d6907ace876cb9bc8337cd82ec8c9684e70fb539788abb3980522.jpg", + "image_caption": [ + "Figure 5. Synthesized results through latent code interpolation, where we fix the text condition, $\\mathbf{c}$ , and interpolate the latent codes within the disentangled latent space, $\\mathcal{W}$ ." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 906, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we do the interpolation in prompt space, we present some results in Fig. 4, where we can observe a smooth semantic transition. For instance, when we interpolate between \"a vase filled with lots of pink flowers\" and \"a moongate in the snowy ice\", the shapes of the flowers smoothly merge with the moongate. Such a result suggests that our model could faithfully preserve semantic continuity when learning to decode text conditions.", + "bbox": [ + 88, + 316, + 480, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Latent interpolation. Latent space interpolation is widely studied in the literature [7, 36, 37] to verify that the generator delivers the generative ability instead of memorizing the dataset. We also perform interpolation between global latent codes, with results presented in Fig. 5. Here, we conduct experiments in $\\mathcal{W}$ space and observe smooth interpolation as well. For instance, the pose of the man's head transitions smoothly from left to right, indicating that the continuity of the latent space in GANs is maintained even with highly complex datasets.", + "bbox": [ + 88, + 439, + 482, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Analyses and Discussion", + "text_level": 1, + "bbox": [ + 89, + 604, + 313, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse router. Recall that our motivation is to enlarge the generator capacity in GANs following the philosophy of sparse MoE [20, 32, 78]. Different from current methods, our sparse router makes its decision considering not only the input feature map but also the conditional text and the sampling stochasticity (Eq. (6)). Here, we visualize some routing paths to verify the effectiveness of our routing mechanism. We can tell from Fig. 6 that, within a synthesized image, pixels with similar visual concepts tend to be processed by the same expert (i.e., having the same color in Fig. 6). Such an observation holds at all resolutions, verifying the rationality of our design.", + "bbox": [ + 88, + 627, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. The lack of an open-sourced large-scale GAN model makes it hard for the community to fully study the past favorite generative weapon, leaving some stereotypes, including both negative and positive ones, on GANs after scaling up for T2I generation. For example, instead of the commonly believed training instability, we find that", + "bbox": [ + 88, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4692986f08633da6107ea238fafada82c557cf6d76813c7695f0542db7ef402e.jpg", + "image_caption": [ + "Figure 6. Analysis of our learned sparse router, where different colors represent the indices of the experts chosen by the router for each pixel. We can tell that the router tends to cluster the pixels by various visual concepts." + ], + "image_footnote": [], + "bbox": [ + 532, + 316, + 903, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GANs can be stably optimized after introducing text as the condition. Besides, we would like to point out that our model is decently trained on publicly available datasets (see Sec. 4.1) without fine-tuning on any self-collected data for performance improvement. While our release is intended for research purposes only, we have observed that using higher-quality datasets consistently improves performance, as discussed in the diffusion model [6]. We leave such an exploration as the future work.", + "bbox": [ + 511, + 523, + 906, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 679, + 769, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we investigate how to expand the capacity of GANs to match the demand of open-vocabulary text-to-image generation. Drawing lessons from sparse MoE, our proposed Aurora matches the synthesis performance of existing diffusion-based and autoregressive counterparts, both quantitatively (e.g., FID scores) and qualitatively (particularly in scene image generation). Moreover, it offers significantly faster inference and maintains favorable continuous latent space properties. However, our method still struggles when generating complex images, particularly those involving multiple objects, intricate shapes, or fine-grained details. We believe that issues with the discriminator might be contributing to these challenges.", + "bbox": [ + 511, + 705, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18418", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Int. Conf. Mach. Learn., 2017. 1, 5", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. eDiffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 7", + "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 2", + "[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN dissection: Visualizing and understanding generative adversarial networks. In Int. Conf. Learn. Represent., 2019. 2", + "[5] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 5", + "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2023.8", + "[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In Int. Conf. Learn. Represent., 2019. 2, 6, 8", + "[8] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2.6", + "[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2", + "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[11] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Int. Conf. Mach. Learn., 2023. 2", + "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2", + "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Int. Conf. Mach. Learn., 2020. 5" + ], + "bbox": [ + 101, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Adv. Neural Inform. Process. Syst., 2024. 6", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In Adv. Neural Inform. Process. Syst., 2021. 1, 2", + "[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In Adv. Neural Inform. Process. Syst., 2021. 2", + "[17] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. GLaM: Efficient scaling of language models with mixture-of-experts. In Int. Conf. Mach. Learn., 2022. 2, 3", + "[18] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In Int. Conf. Learn. Represent. Worksh., 2013. 2", + "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. Learn., 2024. 1, 2", + "[20] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 2022, 2, 3, 6, 8", + "[21] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-A-Scene: Scene-based text-to-image generation with human priors. In Eur. Conf. Comput. Vis., 2022. 2", + "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3D: A generative model of high quality 3D textured shapes learned from images. In Adv. Neural Inform. Process. Syst., 2022. 2", + "[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Adv. Neural Inform. Process. Syst., 2014. 1, 2, 3, 5", + "[24] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2", + "[25] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2", + "[26] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 2", + "[27] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained" + ], + "bbox": [ + 524, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18419", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 2017, 2, 6", + "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Adv. Neural Inform. Process. Syst., 2020. 1, 2", + "[29] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Int. Conf. Comput. Vis., 2017. 3", + "[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.6", + "[31] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2", + "[32] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2, 8", + "[33] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In Int. Conf. Learn. Represent., 2020. 2", + "[34] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1, 2, 3, 4, 5, 6, 7", + "[35] Animesh Karnewar and Oliver Wang. *Msg-gan: Multiscale gradients for generative adversarial networks.* arXiv preprint arXiv:1903.06048, 2019. 4", + "[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Int. Conf. Learn. Represent., 2018. 2, 4, 8", + "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4, 5, 6, 8", + "[38] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3, 4, 6", + "[39] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Adv. Neural Inform. Process. Syst., 2021. 2, 4", + "[40] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2", + "[41] Hyunjik Kim, George Papamakarios, and Andriy Mnih. The lipschitz constant of self-attention. In Int. Conf. Mach. Learn., 2021. 3", + "[42] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In Int. Conf. Learn. Represent., 2014. 1" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2", + "[44] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. In Int. Conf. Learn. Represent., 2022. 3", + "[45] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before Fuse: Vision and language representation learning with momentum distillation. In Adv. Neural Inform. Process. Syst., 2021. 3", + "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis., 2014. 2, 6", + "[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 6", + "[48] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Adv. Neural Inform. Process. Syst., 2022. 2", + "[49] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 5", + "[50] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5", + "[51] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for GANs do actually converge? In Int. Conf. Mach. Learn., 2018. 5", + "[52] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2", + "[53] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Int. Conf. Learn. Represent., 2018. 1", + "[54] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In Int. Conf. Comput. Vis., 2019. 2", + "[55] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., 2022. 2, 7", + "[56] David W. Opitz and Richard Maclin. Popular ensemble methods: An empirical study. J. Artif. Intell. Res., 11:169-198, 1999. 2", + "[57] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman." + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18420", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "StyleSDF: High-resolution 3D-consistent image and geometry generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., 2021. 2, 3, 5", + "[60] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Int. Conf. Mach. Learn., 2021. 1, 2, 7", + "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", + "[62] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Int. Conf. Mach. Learn., 2016. 5", + "[63] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Adv. Neural Inform. Process. Syst., 2021. 3", + "[64] Lior Rokach. Ensemble-based classifiers. Artif. Intell. Rev., 33(1-2):1-39, 2010. 2", + "[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 7", + "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4, 5, 6", + "[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2, 4, 7", + "[68] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Int. Conf. Comput. Vis., 2017. 2", + "[69] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 2", + "[70] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. Adv. Neural Inform. Process. Syst., 2016. 1" + ], + "bbox": [ + 99, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGAN-XL: Scaling StyleGAN to large diverse datasets. In ACM SIGGRAPH, 2022. 2, 3, 4, 6", + "[72] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In Int. Conf. Mach. Learn., 2023. 2, 3, 6, 7", + "[73] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In Eur. Conf. Comput. Vis., 2025. 2", + "[74] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 2, 6", + "[75] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Adv. Neural Inform. Process. Syst., 2020. 2", + "[76] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3D-aware image synthesis with sparse voxel grids. In Adv. Neural Inform. Process. Syst., 2022. 2", + "[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Association for Computational Linguistics, 2018. 6", + "[78] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Int. Conf. Learn. Represent., 2017. 2, 3, 8", + "[79] Sheng Shen, Zhewei Yao, Chunyuan Li, Trevor Darrell, Kurt Keutzer, and Yuxiong He. Scaling vision-language models with sparse mixture of experts. arXiv preprint arXiv:2303.07226, 2023. 3", + "[80] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2", + "[81] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2", + "[82] Yujun Shen, Zhiyi Zhang, Dingdong Yang, Yinghao Xu, Ceyuan Yang, and Jiapeng Zhu. Hammer: An efficient toolkit for training deep models. https://github.com/bytedance/Hammer, 2022.6", + "[83] Zifan Shi, Sida Peng, Yinghao Xu, Geiger Andreas, Yiyi Liao, and Yujun Shen. Deep generative models on 3D representations: A survey. arXiv preprint arXiv:2210.15663, 2022. 2", + "[84] Zifan Shi, Yinghao Xu, Yujun Shen, Deli Zhao, Qifeng Chen, and Dit-Yan Yeung. Improving 3D-aware image synthesis with a geometry-aware discriminator. In Adv. Neural Inform. Process. Syst., 2022.", + "[85] Zifan Shi, Yujun Shen, Yinghao Xu, Sida Peng, Yiyi Liao, Sheng Guo, Qifeng Chen, and Dit-Yan Yeung. Learning" + ], + "bbox": [ + 522, + 92, + 903, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d-aware image synthesis with unknown pose distribution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023.", + "[86] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3D GANs. In Adv. Neural Inform. Process. Syst., 2022. 2", + "[87] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Int. Conf. Mach. Learn., 2015. 1", + "[88] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2020. 2", + "[89] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Int. Conf. Mach. Learn., 2023. 2", + "[90] Ryohei Suzuki, Masanori Koyama, Takeru Miyato, Taizan Yonetsuji, and Huachun Zhu. Spatially controllable image synthesis with internal representation collaging. arXiv preprint arXiv:1811.10153, 2018. 2", + "[91] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2818-2826, 2016. 6", + "[92] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2019. 3", + "[93] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. GALIP: Generative adversarial CLIPs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2", + "[94] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7", + "[95] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 6", + "[96] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2", + "[97] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In Adv. Neural Inform. Process. Syst., 2016. 1", + "[98] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Adv. Neural Inform. Process. Syst., 2017. 1", + "[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Adv. Neural Inform. Process. Syst., 2017. 3", + "[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[101] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2", + "[102] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2", + "[103] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3D-aware image synthesis via learning structural and textural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[104] Yiran Xu, Taesung Park, Richard Zhang, Yang Zhou, Eli Shechtman, Feng Liu, Jia-Bin Huang, and Difan Liu. Videogigagan: Towards detail-rich video super-resolution. arXiv preprint arXiv:2404.12388, 2024. 2", + "[105] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. Int. J. Comput. Vis., 2020. 2", + "[106] Ceyuan Yang, Qihang Zhang, Yinghao Xu, Jiapeng Zhu, Yujun Shen, and Bo Dai. Learning modulated transformation in GANs. arXiv preprint arXiv:2308.15472, 2023. 3, 4", + "[107] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2", + "[108] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022. 1, 2, 7", + "[109] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5", + "[110] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In Int. Conf. Mach. Learn., pages 7354–7363, 2019. 2", + "[111] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[112] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiaxiang Gu, Jinhui Xu, and Tong Sun. LAFITE: Towards language-free training for text-to-image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 7", + "[113] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zhengjun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in GANs. In Adv. Neural Inform. Process. Syst., 2021. 2", + "[114] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in GANs. In Int. Conf. Mach. Learn., 2022.", + "[115] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18422", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "natural image manifold. In Eur. Conf. Comput. Vis., 2016. 2", + "bbox": [ + 130, + 90, + 482, + 119 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "18423", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_model.json b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..96e889e38c92b3f4e2fb93786d1a2c313ad2ec76 --- /dev/null +++ b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_model.json @@ -0,0 +1,2987 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.852, + 0.154 + ], + "angle": 0, + "content": "Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.18, + 0.287, + 0.2 + ], + "angle": 0, + "content": "Jiapeng Zhu1,2" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.181, + 0.424, + 0.217 + ], + "angle": 0, + "content": "Ceyuan Yang3 \nYifei Zhang1,4" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.181, + 0.576, + 0.217 + ], + "angle": 0, + "content": "Kecheng Zheng1 \nQifeng Chen2" + }, + { + "type": "text", + "bbox": [ + 0.587, + 0.181, + 0.714, + 0.218 + ], + "angle": 0, + "content": "Yinghao Xu\\(^{1,3}\\) \nYujun Shen\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.735, + 0.181, + 0.829, + 0.198 + ], + "angle": 0, + "content": "Zifan Shi\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.222, + 0.288, + 0.242 + ], + "angle": 0, + "content": "1Ant Group" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.222, + 0.404, + 0.239 + ], + "angle": 0, + "content": "\\(^{2}\\)HKUST" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.222, + 0.51, + 0.239 + ], + "angle": 0, + "content": "\\(^{3}\\)CUHK" + }, + { + "type": "text", + "bbox": [ + 0.55, + 0.222, + 0.802, + 0.242 + ], + "angle": 0, + "content": "4Shanghai Jiao Tong Univeristy" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.276, + 0.434, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.538, + 0.315, + 0.549 + ], + "angle": 0, + "content": "A black and white photo of a dog with big eyes." + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.276, + 0.589, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.393, + 0.587, + 0.412 + ], + "angle": 0, + "content": "A photo of a colorful sports car, with mountain background." + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.413, + 0.589, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.53, + 0.575, + 0.548 + ], + "angle": 0, + "content": "Sunrise over the clouds in the mountains." + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.276, + 0.743, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.393, + 0.734, + 0.412 + ], + "angle": 0, + "content": "A steaming plate of classic fried rice with vegetables, eggs." + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.412, + 0.742, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.529, + 0.728, + 0.548 + ], + "angle": 0, + "content": "A coral reef with starfish, sea anemones and fish." + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.276, + 0.896, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.393, + 0.893, + 0.412 + ], + "angle": 0, + "content": "Oil painting of a flower garden, distant cottage, soft sunlight." + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.413, + 0.896, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.529, + 0.896, + 0.549 + ], + "angle": 0, + "content": "A pencil sketch of a woman with deep, almond-shaped eyes in." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.556, + 0.905, + 0.571 + ], + "angle": 0, + "content": "Figure 1. Example results at \\(512 \\times 512\\) resolution synthesized by our proposed Aurora, a large-scale GAN-based text-to-image generator." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.598, + 0.327, + 0.614 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.633, + 0.486, + 0.89 + ], + "angle": 0, + "content": "Due to the difficulty in scaling up, generative adversarial networks (GANs) seem to be falling out of grace with the task of text-conditioned image synthesis. Sparsely activated mixture-of-experts (MoE) has recently been demonstrated as a valid solution to training large-scale models with limited resources. Inspired by this, we present Aurora, a GAN-based text-to-image generator that employs a collection of experts to learn feature processing, together with a sparse router to adaptively select the most suitable expert for each feature point. We adopt a two-stage training strategy, which first learns a base model at \\(64 \\times 64\\) resolution followed by an upsampler to produce \\(512 \\times 512\\) images. Trained with only public data, our approach encouragingly closes the performance gap between GANs and industry-level diffusion models, maintaining a fast inference speed. We release the code and checkpoints here to facilitate the community for further development." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.598, + 0.645, + 0.614 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.779 + ], + "angle": 0, + "content": "As one of the most popular tasks in the field of content generation, text-conditioned image synthesis [3, 12, 19, 34, 60, 61, 65, 67, 100, 108] allows users to customize a picture with natural language instructions. Among diverse types of generative models, such as variational autoencoders (VAEs) [42], autoregressive models (ARs) [97, 98], and generative adversarial networks (GANs) [23], diffusion models [15, 28, 87] have blossomed into a dominant solution with a thriving community (e.g., stable diffusion [65]), leaving few efforts on improving other frameworks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we put our focus on GANs, which are usually criticized for training instability [1, 53] and limited mode coverage [70]. However, we argue that the task of text-to-image (T2I) generation is more like a translation task (i.e., shifting an informative data distribution to another) instead of a pure generation task (i.e., modeling the observed distribution with random noise as the input), and accordingly that the introduction of the text condition" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18411" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "can already help alleviate the above two issues to some extent, which is also empirically supported by some recent studies [34, 72]. Even so, the difficulty in scaling up the model size still manifests itself as a major drawback of GANs. Concretely, unlike diffusion models that produce an image through an iterative process [15, 28, 65], the generator of a GAN typically adopts a feedforward network to render an image. Considering the limited computational resources (e.g., GPU memory) in practice, directly learning a sufficiently large feedforward model remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.482, + 0.352 + ], + "angle": 0, + "content": "The advent of sparse mixture-of-experts (MoE) offers a promising way to enlarge the model scale with affordable computing cost [20, 78]. The rationale behind is to employ multiple expert modules yet only activate one or a few of them for each feedforward execution. That way, the growth of the number of modules could expand the model capacity but barely increase the number of computations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.355, + 0.483, + 0.716 + ], + "angle": 0, + "content": "We decently introduce the philosophy of sparse MoE into the design of GAN generators and propose Aurora for fast T2I generation. Specifically, given an intermediate feature map at a particular level of degree, we engage a set of experts, each of which is originally designed to process a single-point feature, as well as a highly sparse router to help select the most suitable expert for each feature point. It is noteworthy that, different from most existing spare MoE methods [20, 32, 78] that determine the routing path simply based on the information from the input feature map, our router also takes into account the text-integrated global latent code such that it can adaptively make the best decision via fully considering the sampling stochasticity and the text condition. Additionally, Aurora employs a two-stage pipeline [34, 67] involving a base generator and an upsampler. The base generator creates images at a resolution of \\(64 \\times 64\\), are then enhanced by an upsampler to produce final images at \\(512 \\times 512\\). Our base generator learned on LAION2B-en [74] and COYO-700M [8] achieves 6.2 zero-shot FID [27] on MS COCO [46] at the resolution of \\(64 \\times 64\\), surpassing other existing methods. The upsampler further refines the image to a final resolution of \\(512 \\times 512\\) pixels, achieving a zero-shot FID score of 8.74 on MS COCO, which sets a new benchmark for GANs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Besides the fast inference speed, GANs also enjoy a well-studied latent space [33, 81, 105] and the high flexibility to incorporate other priors (e.g., involving the volumetric rendering pipeline for 3D-aware image synthesis) [9, 10, 22, 24, 54, 57, 75, 76, 83-86, 103]. However, the lack of an open-sourced large-scale GAN model hinders such a powerful tool from being adequately developed and analyzed. We thus release the code and checkpoints to facilitate the research community and plan for continuous improvement from the perspectives of both performance and functionality. We believe that our work could encourage more studies in the field of visual content generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.386 + ], + "angle": 0, + "content": "Generative adversarial networks (GANs). Formulated as a two-player game between a generator and a discriminator, GANs [23] are designed to model the observed data distribution through adversarial training. Thanks to the flexible formulation (i.e., there are no particular assumptions on the data format or on the generation process), GANs have shown great potential in a variety of visual generative tasks, such as image generation [7, 36-39], image editing [4, 25, 33, 80, 81, 90, 101, 102, 105, 113-115], video generation [68, 96, 104], and 3D-aware image synthesis [9, 10, 24, 54, 75, 83, 103]. Some studies also explore conditional GANs [52], which incorporate signals like class labels [7, 71, 110], texts [34, 72, 93, 112], and reference images [31, 43, 107] into the generator for interactive control. Despite the aforementioned success, GANs usually struggle in learning from large-scale, diverse datasets (e.g., LAION [74]) due to the difficulty in scaling up [34]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.905, + 0.719 + ], + "angle": 0, + "content": "Text-conditioned image synthesis. With the aid of visual-language models like CLIP [59], text-to-image (T2I) generation [2, 3, 11, 12, 19, 26, 34, 58, 60, 61, 65, 67, 108] manages to decode an open-vocabulary description to a vivid image, and hence allows users to visualize their creativity and imagination. Towards such a challenging task, existing efforts usually opt to trade computing efficiency for visual quality. For example, autoregressive models [16, 21, 26, 60, 100, 108] and diffusion models [3, 12, 19, 55, 58, 61, 65, 67] have demonstrated impressive results through an iterative generation process, which can be time-consuming and costly to make a synthesis. Differently, GANs employ a feedforward network as the generator and thus naturally support fast generation without having to use an acceleration [40, 48, 88] or distillation [69, 73, 89] algorithm. Some recent attempts [34, 72] have confirmed the promise of GANs in text-to-image synthesis yet left the issue of enlarging the model capacity unsolved. In addition, there currently lacks a publicly available GAN-based T2I model in the community, making it difficult to fully study the properties of GANs, let alone integrate the advantages of GANs into other types of generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Mixture-of-experts (MoE). Sharing a similar philosophy with ensemble [56, 64], MoE [32] learns a set of expert modules to process the input signal, together with a gating module (also known as a router) to help combine the outputs of different experts. Under such a design, each expert is expected to focus on the learning of a sub-region of the input space, while the router is responsible for the weight assignment across experts [18]. Sparsely activated MoE [17, 20, 78] distinguishes itself from other variants of MoE because of the discrete routing mechanism. By selecting only one or a few experts from a large candidate pool for each incoming example, sparse MoE offers a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18412" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "feasible solution to scaling up the model size with barely increased computational cost [20, 63, 79]. In this work, we introduce the idea of sparse MoE into the design of a GAN generator and adequately adapt the routing mechanism for the task of text-to-image generation by taking into account the text-integrated stochasticity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.197, + 0.182, + 0.212 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.495 + ], + "angle": 0, + "content": "Recall that we aim to develop an efficient text-to-image (T2I) generator based on the framework of GANs [23], which consists of a generator and a discriminator competing with each other. On the one hand, the generator takes in a global latent code, \\(\\mathbf{z} \\in \\mathbb{R}^{128}\\), along with a text description (e.g., a sentence), \\(\\mathbf{c}\\), and outputs an image, \\(\\mathbf{x} = G(\\mathbf{z},\\mathbf{c})\\). On the other hand, the discriminator takes in an image together with an associated text and differentiates synthesized samples from real samples using a realiznescore, \\(D(\\mathbf{x},\\mathbf{c})\\). As discussed above, the major contribution of this work is to enlarge the capacity of the base generator; therefore, we directly borrow the discriminator architecture from the prior art [34]. In the following, we first introduce the common unit blocks utilized in both the base generator and the upsampler (Sec. 3.1). We then provide a detailed overview of the architecture of our base generator (Sec. 3.2) and upsampler (Sec. 3.3), followed by a discussion of the supervision signals and training strategy (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.505, + 0.3, + 0.519 + ], + "angle": 0, + "content": "3.1. Unit Generative Block" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.648 + ], + "angle": 0, + "content": "Like most modern deep models [37, 99], our base generator and upsampler follow the principle of block-wise design. As shown in Fig. 2, a generative unit stacks a convolution block and an attention block to process the input feature map, \\(\\mathbf{f}_{in}\\), where the attention block adopts the sparse routing mechanism for scaling up. Additionally, in this section, \\(\\mathbf{w}\\) refers to the intermediate latent code and will be introduced in detail in Sec. 3.2 and Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.484, + 0.754 + ], + "angle": 0, + "content": "Text encoder. We employ CLIP [59] as the text encoder, which transforms the input text, \\(\\mathbf{c}\\), into both a global token, \\(\\mathbf{t}_g\\), and a token sequence, \\(\\mathbf{t}_{seq}\\). Specifically, \\(\\mathbf{t}_g\\) is the pooled text embedding that encapsulates the meaning of the entire sentence, while \\(\\mathbf{t}_{seq}\\) is derived from the penultimate layer of the text encoder, capturing information for each individual word in the text. We formulate these as follows" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.767, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {t} _ {g}, \\mathbf {t} _ {\\text {s e q}} \\right\\} = \\operatorname {C L I P} (\\mathbf {c}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Convolution block. Latent modulation [37], which is developed from adaptive instance normalization (AdaIN) [29], has demonstrated its effectiveness in the generation field and hence serves as a basic operation in GAN-based models [34, 71, 72]. Such a technique is further improved by integrating it into the convolution operation with demodulation (i.e., modulated convolution," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "ModConv) [38] and by equipping it with learnable feature deformation (i.e., modulated transformation module, MTM) [106]. We employ two MTMs in the convolution block and construct the block using skip connection. Consequently, the output of the convolutional block is given by" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.19, + 0.907, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\text {c o n v}} = \\operatorname {M T M} \\left(\\operatorname {M T M} \\left(\\mathbf {f} _ {\\text {i n}}, \\mathbf {w}\\right), \\mathbf {w}\\right) + \\mathbf {f} _ {\\text {i n}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.906, + 0.265 + ], + "angle": 0, + "content": "Notably, the learnable offsets in MTM [106] are only introduced when the feature resolution is equal to or less than \\( 16 \\times 16 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Attention block. We can tell from Eq. (2) that, in the convolution block, the text condition, \\(\\mathbf{c}\\), only interacts with the generative features through the global token, \\(\\mathbf{t}_g\\), with \\(\\mathbf{w}\\) in Eq. (7) as an intermediary. To also make sufficient use of the text token sequence, \\(\\mathbf{t}_{seq}\\), we introduce an attention block for text-image interaction, which is widely used in multi-modality learning [45, 92]. Specifically, the attention block consists of a self-attention layer, a cross-attention layer, and a feedforward network [99]. We thereby have" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.412, + 0.905, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {s a} = \\operatorname {S e l f A t t n} \\left(\\mathbf {f} _ {\\text {c o n v}}, \\mathbf {w}\\right) + \\mathbf {f} _ {\\text {c o n v}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.431, + 0.905, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {c a} = \\operatorname {C r o s s A t t n} \\left(\\mathbf {f} _ {s a}, \\mathbf {t} _ {\\text {s e q}}\\right) + \\mathbf {f} _ {s a}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.45, + 0.905, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {f f n} = \\mathrm {F F N} \\left(\\mathbf {f} _ {c a}\\right) + \\mathbf {f} _ {c a}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.534 + ], + "angle": 0, + "content": "where \\( \\mathsf{FFN} \\) in Eq. (5) processes each feature point individually using a two-layer MLP. Here, we follow the prior work [34, 44] that employs \\( \\ell 2 \\)-distance instead of the dot product for attention logits to stabilize training [41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Multiple experts and sparse routing. Recall that the main motivation of this work is to enlarge the capacity of the base generator in GANs. One straightforward solution is to make the network deeper or wider. However, it may cause an out-of-memory issue. Inspired by the practice in natural language processing (NLP) [20], we incorporate the philosophy of sparse MoE [78] into the design of our unit generative block to enlarge the base generator capacity. Concretely, we treat the FFN in Eq. (5) as an expert, which specializes in processing a single-point feature, and engage \\( N \\) experts in total, \\( \\{ \\mathsf{FFN}_j \\}_{j=1}^N \\), forming a candidate pool. We then employ a sparse router to help select the most suitable expert for each feature point. Unlike existing sparse MoE methods [17, 20, 78], our router adaptively makes its decision by taking into account not only the input feature but also the text-integrated global latent code, \\( \\mathbf{w} \\), which better matches the target of T2I generation (i.e., decoding the text condition and the sampling stochasticity to an image). Under such a routing strategy, we can reformulate Eq. (5) as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.843, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {f f n} ^ {(k)} = \\operatorname {F F N} _ {j} \\left(\\mathbf {f} _ {c a} ^ {(k)}\\right) + \\mathbf {f} _ {c a} ^ {(k)} \\text {s . t .} j = \\operatorname {R o u t e r} \\left(\\mathbf {f} _ {c a} ^ {(k)}, \\mathbf {w}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Here, the superscript \\((k)\\) denotes the \\(k\\)-th feature point. In addition to increasing the capacity of the generator, MoE" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18413" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.096, + 0.912, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 2. Illustration of base generator in Aurora. First, we utilize a pretrained CLIP text encoder to obtain the text embeddings \\(\\mathbf{t}_g\\) and \\(\\mathbf{t}_{seq}\\). Then, starting from the Fourier feature, the generator processes it through several unit generative blocks. Each unit block is composed of a convolution block and an attention block. There are two MTM [106] layers within the convolution block. In the attention block, the FFN acts as the expert, specializing in processing a single-point feature. A total of \\(N\\) experts are assembled to form a candidate pool, from which selections are made by the router based on the input features and the latent code \\(\\mathbf{w}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.482, + 0.476 + ], + "angle": 0, + "content": "also frees the generator from modeling complex data distributions by assigning different objects or parts of an image to specialized experts, reducing the learning difficulty." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.248, + 0.497 + ], + "angle": 0, + "content": "3.2. Base Generator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.549 + ], + "angle": 0, + "content": "The base generator comprises two components: a mapping network and a synthesis network. Both networks accept text and latent code as inputs, which we will introduce next." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Mapping network. As pointed out by StyleGAN [37], utilizing a learnable MLP-based mapping network on the latent code, \\(\\mathbf{z}\\), could create a more disentangled latent space, \\(\\mathcal{W}\\), than the native latent space, \\(\\mathcal{Z}\\). In the class-conditional generation, a class label is first concatenated with the latent code \\(\\mathbf{z}\\) before being passed to the mapping network [37, 71]. Under the T2I setting, it is necessary to shift from class conditioning to text conditioning to make the sampling process better aware of the text information. Hence, we concatenate \\(\\mathbf{t}_g\\) with the sampled latent code \\(\\mathbf{z}\\) for further processing, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.726, + 0.483, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} = \\operatorname {M L P} \\left(\\operatorname {c o n c a t} \\left(\\mathbf {z}, \\mathbf {t} _ {g}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.449, + 0.765 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}\\) stands for the more disentangled latent code." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Synthesis network. The synthesis network starts with Fourier features [39] and is subsequently followed by a series of convolutional blocks, attention blocks, including the MoE, and finally outputs the image at a resolution of \\(64 \\times 64\\), as shown in Fig. 2. The \\(\\mathbf{w}\\) generated by the mapping network is used to modulate the convolutional and self-attention layers, and it also functions in the MoE to select the appropriate expert for each input feature point. Additionally, the local token \\(\\mathbf{t}_{seq}\\), is incorporated into the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.581 + ], + "angle": 0, + "content": "cross-attention layers as the textual condition to further enhance text-image alignment. The generator produces RGB images at each resolution and employs output skip connections [38]. Unlike StyleGAN2 [38], which only utilizes the highest resolution image as the final output, our approach takes advantage of all generated images, similar to MSG-GAN [35] and GigaGAN [34]. Each image is then used to calculate the training loss, mirroring the progressive training approach described in [36], which can improve the final performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.637, + 0.608 + ], + "angle": 0, + "content": "3.3. Upsampler" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The base generator produces images at a resolution of \\(64 \\times 64\\). To create high-resolution images, we train a text-conditioned super-resolution model that upsamples the images generated by the base generator to a resolution of \\(512 \\times 512\\). The architectures of the upsampler are derived from the base model with specific modifications. First, additional convolution blocks are integrated to process the input image, resulting in a final architecture resembling a U-Net. Second, we set the number of experts to 1 based on empirical evidence indicating that this model does not require an extensive parameter count. Third, instead of acquiring the global code (i.e., \\(\\mathbf{w}\\) in base generator) from the mapping network, we obtain it from the encoder part of the U-Net, which is then concatenated with \\(\\mathbf{t}_g\\) to modulate the decoder part of the U-Net. Additionally, to enhance memory efficiency and increase the speed of the upsampler, following Imagen [67], we reverse the order of the downsampling and upsampling operations in our U-Net architecture compared to the approach used in prior art [66]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18414" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.357, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.27, + 0.335, + 0.299 + ], + "angle": 0, + "content": "A lighthouse under a starry night sky." + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.091, + 0.625, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.269, + 0.613, + 0.297 + ], + "angle": 0, + "content": "A painting of flowers in a blue vase on a table." + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.091, + 0.893, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.27, + 0.894, + 0.3 + ], + "angle": 0, + "content": "An oil painting depicting a large crowd of people in front of a church." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.3, + 0.358, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.48, + 0.341, + 0.51 + ], + "angle": 0, + "content": "A road that is surrounded by trees with yellow leaves." + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.3, + 0.626, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.48, + 0.576, + 0.511 + ], + "angle": 0, + "content": "Sunset casts a golden glow on \nmajestic mountain peaks." + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.301, + 0.893, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.48, + 0.87, + 0.511 + ], + "angle": 0, + "content": "Landscape, lake, buildings, bridge, Chinese ink style." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.522, + 0.908, + 0.552 + ], + "angle": 0, + "content": "Figure 3. Diverse results generated by Aurora, In each group, the images on the left are randomly generated by our base generator at the resolution of \\(64 \\times 64\\). We also highlight an example at \\(512 \\times 512\\) on the right to demonstrate high-quality synthesis capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.798 + ], + "angle": 0, + "content": "It is noteworthy that the upsampler is a super-resolution model primarily trained on real images. Therefore, it is crucial to minimize the gap between real images and those generated by the base generator. We address this gap through three aspects. First, we augment the input data by introducing additional augmentations, such as Gaussian noise and Gaussian blurring. Second, the skip connection is removed since it provides linearity between the input and output [49], which helps reduce discrepancies when transitioning from real to synthesized images. Third, we increase the downsampling rate beyond the conventional rate of 8, lowering the feature map resolution from the standard \\(8 \\times 8\\) [34, 66], which produces more abstracted features by omitting details in the input image, further bridging the gap between real and synthesized images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.388, + 0.833 + ], + "angle": 0, + "content": "3.4. Learning Objectives and Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The primary focus of this work is to scale up the model size of a GAN generator instead of exploring a better supervision signal for GAN training. Thus, we directly inherit some well-studied loss functions from prior arts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.905, + 0.601 + ], + "angle": 0, + "content": "Still, to make the presentation self-contained, we list below all the objectives used in the training phase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Adversarial loss. Adversarial training [23] is the core concept shared by all GANs and owns many variants in the literature [1, 5, 50]. We adopt the approach from GigaGAN [34], implementing a multi-scale input and multi-scale output logistic non-saturating loss [37] with \\( R_{1} \\) penalty [51] as the regularizer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Matching-aware loss. Unlike the discriminator in unconditional GANs that rejects a sample by merely considering its plausibility, the discriminator in T2I GANs is also expected to take into account the alignment between the condition and the synthesized result. We follow prior arts [34, 62, 109] to employ a matching-aware discriminator and introduce the matching-aware loss that urges the discriminator to reject mismatched text-image pair." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Multi-level CLIP loss. CLIP loss [59] follows the idea of contrastive learning [13], and forces the feature of a generated image close to the feature of its associated text yet away from the feature of irrelevant text. We introduce a multi-level CLIP loss that measures CLIP loss at different" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18415" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "image resolutions, promoting increased text-image similarity. Such a supervision signal not only stabilizes training but also complements the matching-aware loss, aiding the generator in more effectively decoding the text condition." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.152, + 0.482, + 0.257 + ], + "angle": 0, + "content": "MoE loss. Recall that we introduce a collection of experts to enlarge the model capacity and expect them to process the input feature in various ways. For each network execution, only one expert will be activated for a feature point. To balance the workload across experts (i.e., to avoid that some experts are never activated), we introduce the MoE loss following Switch Transformer [20]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.258, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Training. We first train our base generator using the aforementioned losses. When training the upsampler, we incorporate an additional LPIPS loss [111], MoE loss is omitted because we set the number of experts to one for the upsampler. Furthermore, similar to the training procedure of the autoencoder in StableDiffusion [66], the adversarial loss is not employed at the beginning of the upsampler's training process. Once the training of the upsampler on real images is complete, we proceed to finetune it using synthesized images generated by the base generator with adversarial loss only. It is noteworthy that our approach adopts the mechanisms of both adversarial learning and sparse MoE, yet we do not encounter any instability throughout the entire training process, whether in the training of the base generator or the upsampler." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.497, + 0.223, + 0.515 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.629 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of our proposed Aurora both quantitatively (Sec. 4.2) and qualitatively (Sec. 4.3). Additionally, we provide detailed analyses to deepen the reader's understanding of text-to-image (T2I) GANs, discuss the significance of our project, and highlight how our open-sourced code and checkpoints can benefit the research community (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.288, + 0.654 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Datasets. The training data is sourced from a diverse set of public datasets, including LAION2B-en [74], COYO-700M [8], CC3M [77], and YFCC15M [95]. All datasets are filtered based on the following criteria: clip score \\(>0.3\\), resolution \\(>512\\), aesthetic score \\(>5.0\\), and an estimated watermark probability \\(<0.5\\), which results in a total of 120M text-image pairs for training. Additionally, we use InstructBLIP [14] to obtain better captions for training our model. We do not use any self-collected data to simply pursue the improvement of synthesis quality. We believe that our open-sourced codebase can facilitate learning a more capable model using cleaner and higher-quality data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training details. We develop our algorithm using Hammer [82]. The base generator was trained for one week on 320 NVIDIA A100 GPUs, while the upsampler underwent training for five days using 128 NVIDIA A100 GPUs. We" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.159 + ], + "angle": 0, + "content": "Table 1. Ablation study of our base model, which targets generating images at \\(64 \\times 64\\) resolution. FID [27] is employed as the evaluation metric, where a smaller number indicates better performance. “\\(\\mathrm{FID}_{10K}\\)” reports the results on the training set, while “zero-shot \\(\\mathrm{FID}_{30K}\\)” reports the results on MS COCO [46]." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.283 + ], + "angle": 0, + "content": "
ModelZero-shot FID30KFID10K# Params.
Base (64 GPUs)22.4420.770.367B
+ Better caption16.7214.310.367B
+ MoE13.8112.800.981B
+ ConvNeXt12.0311.100.986B
Scaling (128 GPUs)7.638.540.986B
Scaling (320 GPUs)6.207.960.986B
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.427 + ], + "angle": 0, + "content": "utilized the AdamW optimizer [47] with hyper-parameters set to \\(\\beta_{1} = 0\\), \\(\\beta_{2} = 0.99\\), and a weight decay of \\(1e^{-5}\\). The base generator and discriminator were jointly optimized with a learning rate of 0.0025, whereas the upsampler was trained with a learning rate of 0.001. We utilize OpenCLIP ConvNeXt-XXLarge [30] as our text encoder and for computing the CLIP loss during training. Additional training details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.707, + 0.455 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.642 + ], + "angle": 0, + "content": "In this part, we quantitatively compare our Aurora with existing T2I generators to give readers an overall picture of its synthesis performance. Following prior arts [7, 34, 37, 38, 71, 72], we adopt FID [27] as the metric, which evaluates a generator at the distribution level. Here, we compute FID on two datasets to evaluate our model, one is \"FID\\(_{10K}\\)\", which is obtained by comparing the inception features [91] of \\(10K\\) synthesized images and \\(10K\\) real images randomly sampled from the training set, while the other \"zero-shot \\(\\mathrm{FID}_{30K}\\)\" is obtained by comparing the inception features of \\(30K\\) synthesized images and the full validation set of MS COCO." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Base generator study. In Tab. 1, we report the evaluation results on two different data domains under different configurations of the base generator. We begin by implementing our base generator following the StyleGAN architecture and incorporate attention blocks as described in Sec. 3.1. Besides, we set the number of experts to 1 for this baseline model, resulting in 367M parameters. First, the baseline model is trained using the original captions, resulting in a zero-shot \\(\\mathrm{FID}_{30K}\\) score of 22.44. Second, we retrain the model with re-captioned data, which significantly improves its performance. Following this, we increase the number of experts in the attention blocks, expanding the model size to 981M parameters. Additionally, we enlarge the CLIP text encoder by replacing the CLIP ViT-L/14 with OpenCLIP ConvNeXt-XXLarge, further increasing the model parameters. These two modifications result in our final base generator, which has 986M parameters and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18416" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.892, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.325, + 0.905, + 0.355 + ], + "angle": 0, + "content": "Figure 4. Synthesized results through text prompt interpolation, where we fix the global latent code, \\(\\mathbf{z}\\), and interpolate the text tokens, \\(\\{\\mathbf{t}_{seq},\\mathbf{t}_g\\}\\), extracted from two different text conditions, \\(\\mathbf{c}\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison between Aurora and existing methods at \\(64 \\times 64\\) resolution. “*” means that we downsample the samples using Lanczos [72] for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.416, + 0.479, + 0.538 + ], + "angle": 0, + "content": "
MethodTypeFID10KFID30K# Params.
eDiff-I* [2]DM-7.609.10B
SD* [65]DM-8.400.94B
LAFITE [112]GAN-14.8075M
StyleGAN-T [72]GAN-7.30-
GigaGAN [34]GAN9.18-0.65B
Aurora (ours)GAN7.966.200.98B
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.565, + 0.483, + 0.686 + ], + "angle": 0, + "content": "exhibits additional performance improvements. Finally, we conduct experiments by training our model using more GPUs, while keeping all other configurations the same. We observe consistent performance improvements with the use of additional GPUs and have not yet encountered any performance saturation related to the number of GPUs utilized. Our final base generator, trained using 320 GPUs, achieves a zero-shot \\(\\mathrm{FID}_{30K}\\) score of 6.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison with existing methods. We first compare our base generator with existing methods at the resolution of \\(64 \\times 64\\) in Tab. 2. The evaluation results demonstrate that Aurora consistently outperforms all other methods across both metrics at this resolution, irrespective of the model type. Tab. 3 presents the comparison results with the existing methods on higher resolution. According to Tab. 3, our method not only achieves the best performance among GANs in terms of the FID but also has the largest model size. However, it lags behind diffusion models and autoregressive models. We conclude that more powerful superresolution techniques or larger models are necessary, and exploring how to train a high-resolution GAN on a larger model remains worthwhile." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.372, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison between our Aurora and existing text-to-image generators at \\(256 \\times 256\\) resolution. We report the model size, total images seen during training, \"zero-shot \\(\\mathrm{FID}_{30K}\\)\", and inference speed. \"\\*\" means that we downsample the generated samples using Lanczos [72] for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.444, + 0.905, + 0.647 + ], + "angle": 0, + "content": "
MethodType# Params.# ImagesFID30KSpeed
DALL-E [60]AR12.0B1.54B27.50-
Parti-20B [108]AR20.0B3.69B7.23-
GLIDE [55]DM5.0B5.94B12.2415.0s
SD* [65]DM0.9B3.16B8.592.9s
Imagen [67]DM3.0B15.36B7.279.1s
eDiff-I[2]DM9.1B11.47B6.9532.0s
LAFITE [112]GAN75M0.03B26.940.02s
StyleGAN-T [72]GAN1.0B0.25B13.900.1s
GALIP [94]GAN0.32B0.012B12.540.04s
GigaGAN* [34]GAN1.0B0.98B9.030.13s
Aurora* (ours)GAN1.32B0.12B8.740.16s
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.696, + 0.683 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Text-to-image synthesis. We showcase some visual results in Fig. 3 under diverse text conditions. We can tell that, besides performing well on image synthesis within a narrow distribution (e.g., human faces, horses, or bedrooms), GANs are also capable of open-vocabulary T2I generation, allowing users to play with their creativity and imagination. It is noteworthy that our generator excels at generating scenic images (see Figs. 1 and 3), producing pictures that are more realistic than those generated by diffusion models (see supplementary materials for the comparison)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Text interpolation. Our generator produces an image with two independently sampled inputs, namely, a global latent code and a text prompt. Here, we investigate the properties of these two different input spaces through interpolation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18417" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.907, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Figure 5. Synthesized results through latent code interpolation, where we fix the text condition, \\(\\mathbf{c}\\), and interpolate the latent codes within the disentangled latent space, \\(\\mathcal{W}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.482, + 0.437 + ], + "angle": 0, + "content": "First, we do the interpolation in prompt space, we present some results in Fig. 4, where we can observe a smooth semantic transition. For instance, when we interpolate between \"a vase filled with lots of pink flowers\" and \"a moongate in the snowy ice\", the shapes of the flowers smoothly merge with the moongate. Such a result suggests that our model could faithfully preserve semantic continuity when learning to decode text conditions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.44, + 0.483, + 0.592 + ], + "angle": 0, + "content": "Latent interpolation. Latent space interpolation is widely studied in the literature [7, 36, 37] to verify that the generator delivers the generative ability instead of memorizing the dataset. We also perform interpolation between global latent codes, with results presented in Fig. 5. Here, we conduct experiments in \\(\\mathcal{W}\\) space and observe smooth interpolation as well. For instance, the pose of the man's head transitions smoothly from left to right, indicating that the continuity of the latent space in GANs is maintained even with highly complex datasets." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.605, + 0.315, + 0.62 + ], + "angle": 0, + "content": "4.4. Analyses and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Sparse router. Recall that our motivation is to enlarge the generator capacity in GANs following the philosophy of sparse MoE [20, 32, 78]. Different from current methods, our sparse router makes its decision considering not only the input feature map but also the conditional text and the sampling stochasticity (Eq. (6)). Here, we visualize some routing paths to verify the effectiveness of our routing mechanism. We can tell from Fig. 6 that, within a synthesized image, pixels with similar visual concepts tend to be processed by the same expert (i.e., having the same color in Fig. 6). Such an observation holds at all resolutions, verifying the rationality of our design." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Discussion. The lack of an open-sourced large-scale GAN model makes it hard for the community to fully study the past favorite generative weapon, leaving some stereotypes, including both negative and positive ones, on GANs after scaling up for T2I generation. For example, instead of the commonly believed training instability, we find that" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.317, + 0.905, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.452, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Figure 6. Analysis of our learned sparse router, where different colors represent the indices of the experts chosen by the router for each pixel. We can tell that the router tends to cluster the pixels by various visual concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.907, + 0.662 + ], + "angle": 0, + "content": "GANs can be stably optimized after introducing text as the condition. Besides, we would like to point out that our model is decently trained on publicly available datasets (see Sec. 4.1) without fine-tuning on any self-collected data for performance improvement. While our release is intended for research purposes only, we have observed that using higher-quality datasets consistently improves performance, as discussed in the diffusion model [6]. We leave such an exploration as the future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.771, + 0.696 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.888 + ], + "angle": 0, + "content": "In this work, we investigate how to expand the capacity of GANs to match the demand of open-vocabulary text-to-image generation. Drawing lessons from sparse MoE, our proposed Aurora matches the synthesis performance of existing diffusion-based and autoregressive counterparts, both quantitatively (e.g., FID scores) and qualitatively (particularly in scene image generation). Moreover, it offers significantly faster inference and maintains favorable continuous latent space properties. However, our method still struggles when generating complex images, particularly those involving multiple objects, intricate shapes, or fine-grained details. We believe that issues with the discriminator might be contributing to these challenges." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18418" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Int. Conf. Mach. Learn., 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.159, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. eDiffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.228, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.285, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN dissection: Visualizing and understanding generative adversarial networks. In Int. Conf. Learn. Represent., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.355, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[5] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.398, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.468, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In Int. Conf. Learn. Represent., 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.51, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[8] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.566, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[11] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Int. Conf. Mach. Learn., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Int. Conf. Mach. Learn., 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.115, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Adv. Neural Inform. Process. Syst., 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In Adv. Neural Inform. Process. Syst., 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In Adv. Neural Inform. Process. Syst., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.277, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[17] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. GLaM: Efficient scaling of language models with mixture-of-experts. In Int. Conf. Mach. Learn., 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[18] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In Int. Conf. Learn. Represent. Worksh., 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. Learn., 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.461, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[20] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 2022, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.518, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[21] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-A-Scene: Scene-based text-to-image generation with human priors. In Eur. Conf. Comput. Vis., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.575, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3D: A generative model of high quality 3D textured shapes learned from images. In Adv. Neural Inform. Process. Syst., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Adv. Neural Inform. Process. Syst., 2014. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[24] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.759, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[25] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18419" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 2017, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Adv. Neural Inform. Process. Syst., 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.165, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[29] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Int. Conf. Comput. Vis., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.208, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[31] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[32] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[33] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In Int. Conf. Learn. Represent., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.435, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[34] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[35] Animesh Karnewar and Oliver Wang. *Msg-gan: Multiscale gradients for generative adversarial networks.* arXiv preprint arXiv:1903.06048, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Int. Conf. Learn. Represent., 2018. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[38] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[39] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Adv. Neural Inform. Process. Syst., 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[40] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[41] Hyunjik Kim, George Papamakarios, and Andriy Mnih. The lipschitz constant of self-attention. In Int. Conf. Mach. Learn., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[42] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In Int. Conf. Learn. Represent., 2014. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[43] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[44] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. In Int. Conf. Learn. Represent., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.223, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[45] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before Fuse: Vision and language representation learning with momentum distillation. In Adv. Neural Inform. Process. Syst., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.295, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis., 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.352, + 0.904, + 0.38 + ], + "angle": 0, + "content": "[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.382, + 0.905, + 0.436 + ], + "angle": 0, + "content": "[48] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Adv. Neural Inform. Process. Syst., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.44, + 0.905, + 0.507 + ], + "angle": 0, + "content": "[49] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[50] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.569, + 0.905, + 0.61 + ], + "angle": 0, + "content": "[51] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for GANs do actually converge? In Int. Conf. Mach. Learn., 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.613, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[52] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.643, + 0.905, + 0.696 + ], + "angle": 0, + "content": "[53] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Int. Conf. Learn. Represent., 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.7, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[54] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In Int. Conf. Comput. Vis., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.758, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[55] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.829, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[56] David W. Opitz and Richard Maclin. Popular ensemble methods: An empirical study. J. Artif. Intell. Res., 11:169-198, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[57] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18420" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "StyleSDF: High-resolution 3D-consistent image and geometry generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.137, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.208, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.281, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[60] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Int. Conf. Mach. Learn., 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.338, + 0.483, + 0.393 + ], + "angle": 0, + "content": "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[62] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Int. Conf. Mach. Learn., 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[63] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Adv. Neural Inform. Process. Syst., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[64] Lior Rokach. Ensemble-based classifiers. Artif. Intell. Rev., 33(1-2):1-39, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.614, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.758, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[68] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Int. Conf. Comput. Vis., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[69] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[70] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. Adv. Neural Inform. Process. Syst., 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[71] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGAN-XL: Scaling StyleGAN to large diverse datasets. In ACM SIGGRAPH, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[72] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In Int. Conf. Mach. Learn., 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[73] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In Eur. Conf. Comput. Vis., 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.233, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[74] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[75] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Adv. Neural Inform. Process. Syst., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[76] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3D-aware image synthesis with sparse voxel grids. In Adv. Neural Inform. Process. Syst., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Association for Computational Linguistics, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[78] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Int. Conf. Learn. Represent., 2017. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[79] Sheng Shen, Zhewei Yao, Chunyuan Li, Trevor Darrell, Kurt Keutzer, and Yuxiong He. Scaling vision-language models with sparse mixture of experts. arXiv preprint arXiv:2303.07226, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.609, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[80] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[81] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[82] Yujun Shen, Zhiyi Zhang, Dingdong Yang, Yinghao Xu, Ceyuan Yang, and Jiapeng Zhu. Hammer: An efficient toolkit for training deep models. https://github.com/bytedance/Hammer, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.762, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[83] Zifan Shi, Sida Peng, Yinghao Xu, Geiger Andreas, Yiyi Liao, and Yujun Shen. Deep generative models on 3D representations: A survey. arXiv preprint arXiv:2210.15663, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[84] Zifan Shi, Yinghao Xu, Yujun Shen, Deli Zhao, Qifeng Chen, and Dit-Yan Yeung. Improving 3D-aware image synthesis with a geometry-aware discriminator. In Adv. Neural Inform. Process. Syst., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[85] Zifan Shi, Yujun Shen, Yinghao Xu, Sida Peng, Yiyi Liao, Sheng Guo, Qifeng Chen, and Dit-Yan Yeung. Learning" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18421" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "3d-aware image synthesis with unknown pose distribution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[86] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3D GANs. In Adv. Neural Inform. Process. Syst., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.163, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[87] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Int. Conf. Mach. Learn., 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[88] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[89] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Int. Conf. Mach. Learn., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[90] Ryohei Suzuki, Masanori Koyama, Takeru Miyato, Taizan Yonetsuji, and Huachun Zhu. Spatially controllable image synthesis with internal representation collaging. arXiv preprint arXiv:1811.10153, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[91] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2818-2826, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[92] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[93] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. GALIP: Generative adversarial CLIPs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.526, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[94] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[95] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[96] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[97] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In Adv. Neural Inform. Process. Syst., 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[98] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Adv. Neural Inform. Process. Syst., 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Adv. Neural Inform. Process. Syst., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[101] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[102] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[103] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3D-aware image synthesis via learning structural and textural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[104] Yiran Xu, Taesung Park, Richard Zhang, Yang Zhou, Eli Shechtman, Feng Liu, Jia-Bin Huang, and Difan Liu. Videogigagan: Towards detail-rich video super-resolution. arXiv preprint arXiv:2404.12388, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[105] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. Int. J. Comput. Vis., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[106] Ceyuan Yang, Qihang Zhang, Yinghao Xu, Jiapeng Zhu, Yujun Shen, and Bo Dai. Learning modulated transformation in GANs. arXiv preprint arXiv:2308.15472, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[107] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[108] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[109] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[110] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In Int. Conf. Mach. Learn., pages 7354–7363, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[111] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[112] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiaxiang Gu, Jinhui Xu, and Tong Sun. LAFITE: Towards language-free training for text-to-image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[113] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zhengjun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in GANs. In Adv. Neural Inform. Process. Syst., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[114] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in GANs. In Int. Conf. Mach. Learn., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[115] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.131, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "natural image manifold. In Eur. Conf. Comput. Vis., 2016. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18423" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_origin.pdf b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e91626196ba7cf37188102de4065c2b31c56e2cb --- /dev/null +++ b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/7078b157-7879-47cb-991a-09f81047db7a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16a9f48dd8643fdd8f9328974852c67f27eb32fc953460646b76a82a292b8ebc +size 5662701 diff --git a/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/full.md b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f73989b1105eb47d1a40c61ebfe478cb49365312 --- /dev/null +++ b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/full.md @@ -0,0 +1,380 @@ +# Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis + +Jiapeng Zhu1,2 + +Ceyuan Yang3 +Yifei Zhang1,4 + +Kecheng Zheng1 +Qifeng Chen2 + +Yinghao Xu $^{1,3}$ +Yujun Shen $^{1}$ + +Zifan Shi $^{1,2}$ + +1Ant Group + +$^{2}$ HKUST + +$^{3}$ CUHK + +4Shanghai Jiao Tong Univeristy + +![](images/dfb66adaf22deacd14135b67a068e138504a8804a86a1e608be3db7eb3d0d558.jpg) +A black and white photo of a dog with big eyes. + +![](images/ee1e7d9be98346aee67f0576d33f31be2ed149d288837b2ab934e57f0214eba1.jpg) +A photo of a colorful sports car, with mountain background. + +![](images/a51ccde226dac9754610c73802c93c5e3aee485d9173acc14ce01cd2a422f7f0.jpg) +Sunrise over the clouds in the mountains. +Figure 1. Example results at $512 \times 512$ resolution synthesized by our proposed Aurora, a large-scale GAN-based text-to-image generator. + +![](images/755b522bbd2b2962a16f0eaba4182c4192565977b33c08cea7be6f054b48019f.jpg) + +![](images/586a04845d97e4ef594442235b7360a7b200931700e30b02ab5db169b928643d.jpg) +A steaming plate of classic fried rice with vegetables, eggs. +A coral reef with starfish, sea anemones and fish. + +![](images/11b332b209e7fb2ea9a32c89e5f98f4dc57f087c55492b92ab4075321a45f89e.jpg) +Oil painting of a flower garden, distant cottage, soft sunlight. + +![](images/889c1bf433a3700bace27f2c773a2768c62b5167e7d91dba223dff1f550c8084.jpg) +A pencil sketch of a woman with deep, almond-shaped eyes in. + +# Abstract + +Due to the difficulty in scaling up, generative adversarial networks (GANs) seem to be falling out of grace with the task of text-conditioned image synthesis. Sparsely activated mixture-of-experts (MoE) has recently been demonstrated as a valid solution to training large-scale models with limited resources. Inspired by this, we present Aurora, a GAN-based text-to-image generator that employs a collection of experts to learn feature processing, together with a sparse router to adaptively select the most suitable expert for each feature point. We adopt a two-stage training strategy, which first learns a base model at $64 \times 64$ resolution followed by an upsampler to produce $512 \times 512$ images. Trained with only public data, our approach encouragingly closes the performance gap between GANs and industry-level diffusion models, maintaining a fast inference speed. We release the code and checkpoints here to facilitate the community for further development. + +# 1. Introduction + +As one of the most popular tasks in the field of content generation, text-conditioned image synthesis [3, 12, 19, 34, 60, 61, 65, 67, 100, 108] allows users to customize a picture with natural language instructions. Among diverse types of generative models, such as variational autoencoders (VAEs) [42], autoregressive models (ARs) [97, 98], and generative adversarial networks (GANs) [23], diffusion models [15, 28, 87] have blossomed into a dominant solution with a thriving community (e.g., stable diffusion [65]), leaving few efforts on improving other frameworks. + +In this work, we put our focus on GANs, which are usually criticized for training instability [1, 53] and limited mode coverage [70]. However, we argue that the task of text-to-image (T2I) generation is more like a translation task (i.e., shifting an informative data distribution to another) instead of a pure generation task (i.e., modeling the observed distribution with random noise as the input), and accordingly that the introduction of the text condition + +can already help alleviate the above two issues to some extent, which is also empirically supported by some recent studies [34, 72]. Even so, the difficulty in scaling up the model size still manifests itself as a major drawback of GANs. Concretely, unlike diffusion models that produce an image through an iterative process [15, 28, 65], the generator of a GAN typically adopts a feedforward network to render an image. Considering the limited computational resources (e.g., GPU memory) in practice, directly learning a sufficiently large feedforward model remains challenging. + +The advent of sparse mixture-of-experts (MoE) offers a promising way to enlarge the model scale with affordable computing cost [20, 78]. The rationale behind is to employ multiple expert modules yet only activate one or a few of them for each feedforward execution. That way, the growth of the number of modules could expand the model capacity but barely increase the number of computations. + +We decently introduce the philosophy of sparse MoE into the design of GAN generators and propose Aurora for fast T2I generation. Specifically, given an intermediate feature map at a particular level of degree, we engage a set of experts, each of which is originally designed to process a single-point feature, as well as a highly sparse router to help select the most suitable expert for each feature point. It is noteworthy that, different from most existing spare MoE methods [20, 32, 78] that determine the routing path simply based on the information from the input feature map, our router also takes into account the text-integrated global latent code such that it can adaptively make the best decision via fully considering the sampling stochasticity and the text condition. Additionally, Aurora employs a two-stage pipeline [34, 67] involving a base generator and an upsampler. The base generator creates images at a resolution of $64 \times 64$ , are then enhanced by an upsampler to produce final images at $512 \times 512$ . Our base generator learned on LAION2B-en [74] and COYO-700M [8] achieves 6.2 zero-shot FID [27] on MS COCO [46] at the resolution of $64 \times 64$ , surpassing other existing methods. The upsampler further refines the image to a final resolution of $512 \times 512$ pixels, achieving a zero-shot FID score of 8.74 on MS COCO, which sets a new benchmark for GANs. + +Besides the fast inference speed, GANs also enjoy a well-studied latent space [33, 81, 105] and the high flexibility to incorporate other priors (e.g., involving the volumetric rendering pipeline for 3D-aware image synthesis) [9, 10, 22, 24, 54, 57, 75, 76, 83-86, 103]. However, the lack of an open-sourced large-scale GAN model hinders such a powerful tool from being adequately developed and analyzed. We thus release the code and checkpoints to facilitate the research community and plan for continuous improvement from the perspectives of both performance and functionality. We believe that our work could encourage more studies in the field of visual content generation. + +# 2. Related Work + +Generative adversarial networks (GANs). Formulated as a two-player game between a generator and a discriminator, GANs [23] are designed to model the observed data distribution through adversarial training. Thanks to the flexible formulation (i.e., there are no particular assumptions on the data format or on the generation process), GANs have shown great potential in a variety of visual generative tasks, such as image generation [7, 36-39], image editing [4, 25, 33, 80, 81, 90, 101, 102, 105, 113-115], video generation [68, 96, 104], and 3D-aware image synthesis [9, 10, 24, 54, 75, 83, 103]. Some studies also explore conditional GANs [52], which incorporate signals like class labels [7, 71, 110], texts [34, 72, 93, 112], and reference images [31, 43, 107] into the generator for interactive control. Despite the aforementioned success, GANs usually struggle in learning from large-scale, diverse datasets (e.g., LAION [74]) due to the difficulty in scaling up [34]. + +Text-conditioned image synthesis. With the aid of visual-language models like CLIP [59], text-to-image (T2I) generation [2, 3, 11, 12, 19, 26, 34, 58, 60, 61, 65, 67, 108] manages to decode an open-vocabulary description to a vivid image, and hence allows users to visualize their creativity and imagination. Towards such a challenging task, existing efforts usually opt to trade computing efficiency for visual quality. For example, autoregressive models [16, 21, 26, 60, 100, 108] and diffusion models [3, 12, 19, 55, 58, 61, 65, 67] have demonstrated impressive results through an iterative generation process, which can be time-consuming and costly to make a synthesis. Differently, GANs employ a feedforward network as the generator and thus naturally support fast generation without having to use an acceleration [40, 48, 88] or distillation [69, 73, 89] algorithm. Some recent attempts [34, 72] have confirmed the promise of GANs in text-to-image synthesis yet left the issue of enlarging the model capacity unsolved. In addition, there currently lacks a publicly available GAN-based T2I model in the community, making it difficult to fully study the properties of GANs, let alone integrate the advantages of GANs into other types of generative models. + +Mixture-of-experts (MoE). Sharing a similar philosophy with ensemble [56, 64], MoE [32] learns a set of expert modules to process the input signal, together with a gating module (also known as a router) to help combine the outputs of different experts. Under such a design, each expert is expected to focus on the learning of a sub-region of the input space, while the router is responsible for the weight assignment across experts [18]. Sparsely activated MoE [17, 20, 78] distinguishes itself from other variants of MoE because of the discrete routing mechanism. By selecting only one or a few experts from a large candidate pool for each incoming example, sparse MoE offers a + +feasible solution to scaling up the model size with barely increased computational cost [20, 63, 79]. In this work, we introduce the idea of sparse MoE into the design of a GAN generator and adequately adapt the routing mechanism for the task of text-to-image generation by taking into account the text-integrated stochasticity. + +# 3. Method + +Recall that we aim to develop an efficient text-to-image (T2I) generator based on the framework of GANs [23], which consists of a generator and a discriminator competing with each other. On the one hand, the generator takes in a global latent code, $\mathbf{z} \in \mathbb{R}^{128}$ , along with a text description (e.g., a sentence), $\mathbf{c}$ , and outputs an image, $\mathbf{x} = G(\mathbf{z},\mathbf{c})$ . On the other hand, the discriminator takes in an image together with an associated text and differentiates synthesized samples from real samples using a realiznescore, $D(\mathbf{x},\mathbf{c})$ . As discussed above, the major contribution of this work is to enlarge the capacity of the base generator; therefore, we directly borrow the discriminator architecture from the prior art [34]. In the following, we first introduce the common unit blocks utilized in both the base generator and the upsampler (Sec. 3.1). We then provide a detailed overview of the architecture of our base generator (Sec. 3.2) and upsampler (Sec. 3.3), followed by a discussion of the supervision signals and training strategy (Sec. 3.4). + +# 3.1. Unit Generative Block + +Like most modern deep models [37, 99], our base generator and upsampler follow the principle of block-wise design. As shown in Fig. 2, a generative unit stacks a convolution block and an attention block to process the input feature map, $\mathbf{f}_{in}$ , where the attention block adopts the sparse routing mechanism for scaling up. Additionally, in this section, $\mathbf{w}$ refers to the intermediate latent code and will be introduced in detail in Sec. 3.2 and Sec. 3.3. + +Text encoder. We employ CLIP [59] as the text encoder, which transforms the input text, $\mathbf{c}$ , into both a global token, $\mathbf{t}_g$ , and a token sequence, $\mathbf{t}_{seq}$ . Specifically, $\mathbf{t}_g$ is the pooled text embedding that encapsulates the meaning of the entire sentence, while $\mathbf{t}_{seq}$ is derived from the penultimate layer of the text encoder, capturing information for each individual word in the text. We formulate these as follows + +$$ +\left\{\mathbf {t} _ {g}, \mathbf {t} _ {\text {s e q}} \right\} = \operatorname {C L I P} (\mathbf {c}). \tag {1} +$$ + +Convolution block. Latent modulation [37], which is developed from adaptive instance normalization (AdaIN) [29], has demonstrated its effectiveness in the generation field and hence serves as a basic operation in GAN-based models [34, 71, 72]. Such a technique is further improved by integrating it into the convolution operation with demodulation (i.e., modulated convolution, + +ModConv) [38] and by equipping it with learnable feature deformation (i.e., modulated transformation module, MTM) [106]. We employ two MTMs in the convolution block and construct the block using skip connection. Consequently, the output of the convolutional block is given by + +$$ +\mathbf {f} _ {\text {c o n v}} = \operatorname {M T M} \left(\operatorname {M T M} \left(\mathbf {f} _ {\text {i n}}, \mathbf {w}\right), \mathbf {w}\right) + \mathbf {f} _ {\text {i n}}. \tag {2} +$$ + +Notably, the learnable offsets in MTM [106] are only introduced when the feature resolution is equal to or less than $16 \times 16$ . + +Attention block. We can tell from Eq. (2) that, in the convolution block, the text condition, $\mathbf{c}$ , only interacts with the generative features through the global token, $\mathbf{t}_g$ , with $\mathbf{w}$ in Eq. (7) as an intermediary. To also make sufficient use of the text token sequence, $\mathbf{t}_{seq}$ , we introduce an attention block for text-image interaction, which is widely used in multi-modality learning [45, 92]. Specifically, the attention block consists of a self-attention layer, a cross-attention layer, and a feedforward network [99]. We thereby have + +$$ +\mathbf {f} _ {s a} = \operatorname {S e l f A t t n} \left(\mathbf {f} _ {\text {c o n v}}, \mathbf {w}\right) + \mathbf {f} _ {\text {c o n v}}, \tag {3} +$$ + +$$ +\mathbf {f} _ {c a} = \operatorname {C r o s s A t t n} \left(\mathbf {f} _ {s a}, \mathbf {t} _ {\text {s e q}}\right) + \mathbf {f} _ {s a}, \tag {4} +$$ + +$$ +\mathbf {f} _ {f f n} = \mathrm {F F N} \left(\mathbf {f} _ {c a}\right) + \mathbf {f} _ {c a}, \tag {5} +$$ + +where $\mathsf{FFN}$ in Eq. (5) processes each feature point individually using a two-layer MLP. Here, we follow the prior work [34, 44] that employs $\ell 2$ -distance instead of the dot product for attention logits to stabilize training [41]. + +Multiple experts and sparse routing. Recall that the main motivation of this work is to enlarge the capacity of the base generator in GANs. One straightforward solution is to make the network deeper or wider. However, it may cause an out-of-memory issue. Inspired by the practice in natural language processing (NLP) [20], we incorporate the philosophy of sparse MoE [78] into the design of our unit generative block to enlarge the base generator capacity. Concretely, we treat the FFN in Eq. (5) as an expert, which specializes in processing a single-point feature, and engage $N$ experts in total, $\{ \mathsf{FFN}_j \}_{j=1}^N$ , forming a candidate pool. We then employ a sparse router to help select the most suitable expert for each feature point. Unlike existing sparse MoE methods [17, 20, 78], our router adaptively makes its decision by taking into account not only the input feature but also the text-integrated global latent code, $\mathbf{w}$ , which better matches the target of T2I generation (i.e., decoding the text condition and the sampling stochasticity to an image). Under such a routing strategy, we can reformulate Eq. (5) as + +$$ +\mathbf {f} _ {f f n} ^ {(k)} = \operatorname {F F N} _ {j} \left(\mathbf {f} _ {c a} ^ {(k)}\right) + \mathbf {f} _ {c a} ^ {(k)} \text {s . t .} j = \operatorname {R o u t e r} \left(\mathbf {f} _ {c a} ^ {(k)}, \mathbf {w}\right). \tag {6} +$$ + +Here, the superscript $(k)$ denotes the $k$ -th feature point. In addition to increasing the capacity of the generator, MoE + +![](images/f4903de1974a77a4a90d05484159018fa3d202c7baa518e077f5ffb5e3b8d7af.jpg) +Figure 2. Illustration of base generator in Aurora. First, we utilize a pretrained CLIP text encoder to obtain the text embeddings $\mathbf{t}_g$ and $\mathbf{t}_{seq}$ . Then, starting from the Fourier feature, the generator processes it through several unit generative blocks. Each unit block is composed of a convolution block and an attention block. There are two MTM [106] layers within the convolution block. In the attention block, the FFN acts as the expert, specializing in processing a single-point feature. A total of $N$ experts are assembled to form a candidate pool, from which selections are made by the router based on the input features and the latent code $\mathbf{w}$ . + +also frees the generator from modeling complex data distributions by assigning different objects or parts of an image to specialized experts, reducing the learning difficulty. + +# 3.2. Base Generator + +The base generator comprises two components: a mapping network and a synthesis network. Both networks accept text and latent code as inputs, which we will introduce next. + +Mapping network. As pointed out by StyleGAN [37], utilizing a learnable MLP-based mapping network on the latent code, $\mathbf{z}$ , could create a more disentangled latent space, $\mathcal{W}$ , than the native latent space, $\mathcal{Z}$ . In the class-conditional generation, a class label is first concatenated with the latent code $\mathbf{z}$ before being passed to the mapping network [37, 71]. Under the T2I setting, it is necessary to shift from class conditioning to text conditioning to make the sampling process better aware of the text information. Hence, we concatenate $\mathbf{t}_g$ with the sampled latent code $\mathbf{z}$ for further processing, as follows: + +$$ +\mathbf {w} = \operatorname {M L P} \left(\operatorname {c o n c a t} \left(\mathbf {z}, \mathbf {t} _ {g}\right)\right), \tag {7} +$$ + +where $\mathbf{w}$ stands for the more disentangled latent code. + +Synthesis network. The synthesis network starts with Fourier features [39] and is subsequently followed by a series of convolutional blocks, attention blocks, including the MoE, and finally outputs the image at a resolution of $64 \times 64$ , as shown in Fig. 2. The $\mathbf{w}$ generated by the mapping network is used to modulate the convolutional and self-attention layers, and it also functions in the MoE to select the appropriate expert for each input feature point. Additionally, the local token $\mathbf{t}_{seq}$ , is incorporated into the + +cross-attention layers as the textual condition to further enhance text-image alignment. The generator produces RGB images at each resolution and employs output skip connections [38]. Unlike StyleGAN2 [38], which only utilizes the highest resolution image as the final output, our approach takes advantage of all generated images, similar to MSG-GAN [35] and GigaGAN [34]. Each image is then used to calculate the training loss, mirroring the progressive training approach described in [36], which can improve the final performance. + +# 3.3. Upsampler + +The base generator produces images at a resolution of $64 \times 64$ . To create high-resolution images, we train a text-conditioned super-resolution model that upsamples the images generated by the base generator to a resolution of $512 \times 512$ . The architectures of the upsampler are derived from the base model with specific modifications. First, additional convolution blocks are integrated to process the input image, resulting in a final architecture resembling a U-Net. Second, we set the number of experts to 1 based on empirical evidence indicating that this model does not require an extensive parameter count. Third, instead of acquiring the global code (i.e., $\mathbf{w}$ in base generator) from the mapping network, we obtain it from the encoder part of the U-Net, which is then concatenated with $\mathbf{t}_g$ to modulate the decoder part of the U-Net. Additionally, to enhance memory efficiency and increase the speed of the upsampler, following Imagen [67], we reverse the order of the downsampling and upsampling operations in our U-Net architecture compared to the approach used in prior art [66]. + +![](images/5706d4f00532f59629316fadc9e3819270fe63690f104f498b29c7e586e49a97.jpg) +A lighthouse under a starry night sky. + +![](images/bb6f7377e22e5a8ba20bdf93d04c2a20bac8fd9c4e41bed32b3a707318e03591.jpg) +A painting of flowers in a blue vase on a table. + +![](images/d34b245c552d56f490fc8859aa59a338a901dafdf94e8e72154eee7e74e65196.jpg) +An oil painting depicting a large crowd of people in front of a church. + +![](images/09e08fc0b6c22dab3c7342bb1f664964fab04bd310927ca217c2a1f90d861cab.jpg) +A road that is surrounded by trees with yellow leaves. + +![](images/04cf4a994a2d64b21f15b47f024099ddc80521927791bf4ae322e1eb92baa0ea.jpg) +Sunset casts a golden glow on +majestic mountain peaks. + +![](images/ad99aa22ecbe136b116b845fc712ba40d7effc81e016eb176c1aef92e5fbd166.jpg) +Landscape, lake, buildings, bridge, Chinese ink style. +Figure 3. Diverse results generated by Aurora, In each group, the images on the left are randomly generated by our base generator at the resolution of $64 \times 64$ . We also highlight an example at $512 \times 512$ on the right to demonstrate high-quality synthesis capability. + +It is noteworthy that the upsampler is a super-resolution model primarily trained on real images. Therefore, it is crucial to minimize the gap between real images and those generated by the base generator. We address this gap through three aspects. First, we augment the input data by introducing additional augmentations, such as Gaussian noise and Gaussian blurring. Second, the skip connection is removed since it provides linearity between the input and output [49], which helps reduce discrepancies when transitioning from real to synthesized images. Third, we increase the downsampling rate beyond the conventional rate of 8, lowering the feature map resolution from the standard $8 \times 8$ [34, 66], which produces more abstracted features by omitting details in the input image, further bridging the gap between real and synthesized images. + +# 3.4. Learning Objectives and Training + +The primary focus of this work is to scale up the model size of a GAN generator instead of exploring a better supervision signal for GAN training. Thus, we directly inherit some well-studied loss functions from prior arts. + +Still, to make the presentation self-contained, we list below all the objectives used in the training phase. + +Adversarial loss. Adversarial training [23] is the core concept shared by all GANs and owns many variants in the literature [1, 5, 50]. We adopt the approach from GigaGAN [34], implementing a multi-scale input and multi-scale output logistic non-saturating loss [37] with $R_{1}$ penalty [51] as the regularizer. + +Matching-aware loss. Unlike the discriminator in unconditional GANs that rejects a sample by merely considering its plausibility, the discriminator in T2I GANs is also expected to take into account the alignment between the condition and the synthesized result. We follow prior arts [34, 62, 109] to employ a matching-aware discriminator and introduce the matching-aware loss that urges the discriminator to reject mismatched text-image pair. + +Multi-level CLIP loss. CLIP loss [59] follows the idea of contrastive learning [13], and forces the feature of a generated image close to the feature of its associated text yet away from the feature of irrelevant text. We introduce a multi-level CLIP loss that measures CLIP loss at different + +image resolutions, promoting increased text-image similarity. Such a supervision signal not only stabilizes training but also complements the matching-aware loss, aiding the generator in more effectively decoding the text condition. + +MoE loss. Recall that we introduce a collection of experts to enlarge the model capacity and expect them to process the input feature in various ways. For each network execution, only one expert will be activated for a feature point. To balance the workload across experts (i.e., to avoid that some experts are never activated), we introduce the MoE loss following Switch Transformer [20]. + +Training. We first train our base generator using the aforementioned losses. When training the upsampler, we incorporate an additional LPIPS loss [111], MoE loss is omitted because we set the number of experts to one for the upsampler. Furthermore, similar to the training procedure of the autoencoder in StableDiffusion [66], the adversarial loss is not employed at the beginning of the upsampler's training process. Once the training of the upsampler on real images is complete, we proceed to finetune it using synthesized images generated by the base generator with adversarial loss only. It is noteworthy that our approach adopts the mechanisms of both adversarial learning and sparse MoE, yet we do not encounter any instability throughout the entire training process, whether in the training of the base generator or the upsampler. + +# 4. Experiments + +In this section, we evaluate the performance of our proposed Aurora both quantitatively (Sec. 4.2) and qualitatively (Sec. 4.3). Additionally, we provide detailed analyses to deepen the reader's understanding of text-to-image (T2I) GANs, discuss the significance of our project, and highlight how our open-sourced code and checkpoints can benefit the research community (Sec. 4.4). + +# 4.1. Experimental Setups + +Datasets. The training data is sourced from a diverse set of public datasets, including LAION2B-en [74], COYO-700M [8], CC3M [77], and YFCC15M [95]. All datasets are filtered based on the following criteria: clip score $>0.3$ , resolution $>512$ , aesthetic score $>5.0$ , and an estimated watermark probability $<0.5$ , which results in a total of 120M text-image pairs for training. Additionally, we use InstructBLIP [14] to obtain better captions for training our model. We do not use any self-collected data to simply pursue the improvement of synthesis quality. We believe that our open-sourced codebase can facilitate learning a more capable model using cleaner and higher-quality data. + +Training details. We develop our algorithm using Hammer [82]. The base generator was trained for one week on 320 NVIDIA A100 GPUs, while the upsampler underwent training for five days using 128 NVIDIA A100 GPUs. We + +Table 1. Ablation study of our base model, which targets generating images at $64 \times 64$ resolution. FID [27] is employed as the evaluation metric, where a smaller number indicates better performance. “ $\mathrm{FID}_{10K}$ ” reports the results on the training set, while “zero-shot $\mathrm{FID}_{30K}$ ” reports the results on MS COCO [46]. + +
ModelZero-shot FID30KFID10K# Params.
Base (64 GPUs)22.4420.770.367B
+ Better caption16.7214.310.367B
+ MoE13.8112.800.981B
+ ConvNeXt12.0311.100.986B
Scaling (128 GPUs)7.638.540.986B
Scaling (320 GPUs)6.207.960.986B
+ +utilized the AdamW optimizer [47] with hyper-parameters set to $\beta_{1} = 0$ , $\beta_{2} = 0.99$ , and a weight decay of $1e^{-5}$ . The base generator and discriminator were jointly optimized with a learning rate of 0.0025, whereas the upsampler was trained with a learning rate of 0.001. We utilize OpenCLIP ConvNeXt-XXLarge [30] as our text encoder and for computing the CLIP loss during training. Additional training details are provided in the supplementary materials. + +# 4.2. Quantitative Results + +In this part, we quantitatively compare our Aurora with existing T2I generators to give readers an overall picture of its synthesis performance. Following prior arts [7, 34, 37, 38, 71, 72], we adopt FID [27] as the metric, which evaluates a generator at the distribution level. Here, we compute FID on two datasets to evaluate our model, one is "FID $_{10K}$ ", which is obtained by comparing the inception features [91] of $10K$ synthesized images and $10K$ real images randomly sampled from the training set, while the other "zero-shot $\mathrm{FID}_{30K}$ " is obtained by comparing the inception features of $30K$ synthesized images and the full validation set of MS COCO. + +Base generator study. In Tab. 1, we report the evaluation results on two different data domains under different configurations of the base generator. We begin by implementing our base generator following the StyleGAN architecture and incorporate attention blocks as described in Sec. 3.1. Besides, we set the number of experts to 1 for this baseline model, resulting in 367M parameters. First, the baseline model is trained using the original captions, resulting in a zero-shot $\mathrm{FID}_{30K}$ score of 22.44. Second, we retrain the model with re-captioned data, which significantly improves its performance. Following this, we increase the number of experts in the attention blocks, expanding the model size to 981M parameters. Additionally, we enlarge the CLIP text encoder by replacing the CLIP ViT-L/14 with OpenCLIP ConvNeXt-XXLarge, further increasing the model parameters. These two modifications result in our final base generator, which has 986M parameters and + +![](images/c3d124f80939e902da3e5bf57d1e52cd9be54268cc9fcd7727b9325f4f06e413.jpg) +Figure 4. Synthesized results through text prompt interpolation, where we fix the global latent code, $\mathbf{z}$ , and interpolate the text tokens, $\{\mathbf{t}_{seq},\mathbf{t}_g\}$ , extracted from two different text conditions, $\mathbf{c}$ . + +Table 2. Quantitative comparison between Aurora and existing methods at $64 \times 64$ resolution. “*” means that we downsample the samples using Lanczos [72] for a fair comparison. + +
MethodTypeFID10KFID30K# Params.
eDiff-I* [2]DM-7.609.10B
SD* [65]DM-8.400.94B
LAFITE [112]GAN-14.8075M
StyleGAN-T [72]GAN-7.30-
GigaGAN [34]GAN9.18-0.65B
Aurora (ours)GAN7.966.200.98B
+ +exhibits additional performance improvements. Finally, we conduct experiments by training our model using more GPUs, while keeping all other configurations the same. We observe consistent performance improvements with the use of additional GPUs and have not yet encountered any performance saturation related to the number of GPUs utilized. Our final base generator, trained using 320 GPUs, achieves a zero-shot $\mathrm{FID}_{30K}$ score of 6.2. + +Comparison with existing methods. We first compare our base generator with existing methods at the resolution of $64 \times 64$ in Tab. 2. The evaluation results demonstrate that Aurora consistently outperforms all other methods across both metrics at this resolution, irrespective of the model type. Tab. 3 presents the comparison results with the existing methods on higher resolution. According to Tab. 3, our method not only achieves the best performance among GANs in terms of the FID but also has the largest model size. However, it lags behind diffusion models and autoregressive models. We conclude that more powerful superresolution techniques or larger models are necessary, and exploring how to train a high-resolution GAN on a larger model remains worthwhile. + +Table 3. Quantitative comparison between our Aurora and existing text-to-image generators at $256 \times 256$ resolution. We report the model size, total images seen during training, "zero-shot $\mathrm{FID}_{30K}$ ", and inference speed. "\*" means that we downsample the generated samples using Lanczos [72] for a fair comparison. + +
MethodType# Params.# ImagesFID30KSpeed
DALL-E [60]AR12.0B1.54B27.50-
Parti-20B [108]AR20.0B3.69B7.23-
GLIDE [55]DM5.0B5.94B12.2415.0s
SD* [65]DM0.9B3.16B8.592.9s
Imagen [67]DM3.0B15.36B7.279.1s
eDiff-I[2]DM9.1B11.47B6.9532.0s
LAFITE [112]GAN75M0.03B26.940.02s
StyleGAN-T [72]GAN1.0B0.25B13.900.1s
GALIP [94]GAN0.32B0.012B12.540.04s
GigaGAN* [34]GAN1.0B0.98B9.030.13s
Aurora* (ours)GAN1.32B0.12B8.740.16s
+ +# 4.3. Qualitative Results + +Text-to-image synthesis. We showcase some visual results in Fig. 3 under diverse text conditions. We can tell that, besides performing well on image synthesis within a narrow distribution (e.g., human faces, horses, or bedrooms), GANs are also capable of open-vocabulary T2I generation, allowing users to play with their creativity and imagination. It is noteworthy that our generator excels at generating scenic images (see Figs. 1 and 3), producing pictures that are more realistic than those generated by diffusion models (see supplementary materials for the comparison). + +Text interpolation. Our generator produces an image with two independently sampled inputs, namely, a global latent code and a text prompt. Here, we investigate the properties of these two different input spaces through interpolation. + +![](images/974d4041d56d6907ace876cb9bc8337cd82ec8c9684e70fb539788abb3980522.jpg) +Figure 5. Synthesized results through latent code interpolation, where we fix the text condition, $\mathbf{c}$ , and interpolate the latent codes within the disentangled latent space, $\mathcal{W}$ . + +First, we do the interpolation in prompt space, we present some results in Fig. 4, where we can observe a smooth semantic transition. For instance, when we interpolate between "a vase filled with lots of pink flowers" and "a moongate in the snowy ice", the shapes of the flowers smoothly merge with the moongate. Such a result suggests that our model could faithfully preserve semantic continuity when learning to decode text conditions. + +Latent interpolation. Latent space interpolation is widely studied in the literature [7, 36, 37] to verify that the generator delivers the generative ability instead of memorizing the dataset. We also perform interpolation between global latent codes, with results presented in Fig. 5. Here, we conduct experiments in $\mathcal{W}$ space and observe smooth interpolation as well. For instance, the pose of the man's head transitions smoothly from left to right, indicating that the continuity of the latent space in GANs is maintained even with highly complex datasets. + +# 4.4. Analyses and Discussion + +Sparse router. Recall that our motivation is to enlarge the generator capacity in GANs following the philosophy of sparse MoE [20, 32, 78]. Different from current methods, our sparse router makes its decision considering not only the input feature map but also the conditional text and the sampling stochasticity (Eq. (6)). Here, we visualize some routing paths to verify the effectiveness of our routing mechanism. We can tell from Fig. 6 that, within a synthesized image, pixels with similar visual concepts tend to be processed by the same expert (i.e., having the same color in Fig. 6). Such an observation holds at all resolutions, verifying the rationality of our design. + +Discussion. The lack of an open-sourced large-scale GAN model makes it hard for the community to fully study the past favorite generative weapon, leaving some stereotypes, including both negative and positive ones, on GANs after scaling up for T2I generation. For example, instead of the commonly believed training instability, we find that + +![](images/4692986f08633da6107ea238fafada82c557cf6d76813c7695f0542db7ef402e.jpg) +Figure 6. Analysis of our learned sparse router, where different colors represent the indices of the experts chosen by the router for each pixel. We can tell that the router tends to cluster the pixels by various visual concepts. + +GANs can be stably optimized after introducing text as the condition. Besides, we would like to point out that our model is decently trained on publicly available datasets (see Sec. 4.1) without fine-tuning on any self-collected data for performance improvement. While our release is intended for research purposes only, we have observed that using higher-quality datasets consistently improves performance, as discussed in the diffusion model [6]. We leave such an exploration as the future work. + +# 5. Conclusion and Limitations + +In this work, we investigate how to expand the capacity of GANs to match the demand of open-vocabulary text-to-image generation. Drawing lessons from sparse MoE, our proposed Aurora matches the synthesis performance of existing diffusion-based and autoregressive counterparts, both quantitatively (e.g., FID scores) and qualitatively (particularly in scene image generation). Moreover, it offers significantly faster inference and maintains favorable continuous latent space properties. However, our method still struggles when generating complex images, particularly those involving multiple objects, intricate shapes, or fine-grained details. We believe that issues with the discriminator might be contributing to these challenges. + +# References + +[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Int. Conf. Mach. Learn., 2017. 1, 5 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. eDiffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 7 +[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 2 +[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN dissection: Visualizing and understanding generative adversarial networks. In Int. Conf. Learn. Represent., 2019. 2 +[5] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 5 +[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2023.8 +[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In Int. Conf. Learn. Represent., 2019. 2, 6, 8 +[8] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2.6 +[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2 +[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[11] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Int. Conf. Mach. Learn., 2023. 2 +[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2 +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Int. Conf. Mach. Learn., 2020. 5 + +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Adv. Neural Inform. Process. Syst., 2024. 6 +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In Adv. Neural Inform. Process. Syst., 2021. 1, 2 +[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In Adv. Neural Inform. Process. Syst., 2021. 2 +[17] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. GLaM: Efficient scaling of language models with mixture-of-experts. In Int. Conf. Mach. Learn., 2022. 2, 3 +[18] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In Int. Conf. Learn. Represent. Worksh., 2013. 2 +[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. Learn., 2024. 1, 2 +[20] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 2022, 2, 3, 6, 8 +[21] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-A-Scene: Scene-based text-to-image generation with human priors. In Eur. Conf. Comput. Vis., 2022. 2 +[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3D: A generative model of high quality 3D textured shapes learned from images. In Adv. Neural Inform. Process. Syst., 2022. 2 +[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Adv. Neural Inform. Process. Syst., 2014. 1, 2, 3, 5 +[24] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2 +[25] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2 +[26] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 2 +[27] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained + +by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 2017, 2, 6 +[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Adv. Neural Inform. Process. Syst., 2020. 1, 2 +[29] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Int. Conf. Comput. Vis., 2017. 3 +[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.6 +[31] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2 +[32] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2, 8 +[33] Ali Jahanian, Lucy Chai, and Phillip Isola. On the "steerability" of generative adversarial networks. In Int. Conf. Learn. Represent., 2020. 2 +[34] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1, 2, 3, 4, 5, 6, 7 +[35] Animesh Karnewar and Oliver Wang. *Msg-gan: Multiscale gradients for generative adversarial networks.* arXiv preprint arXiv:1903.06048, 2019. 4 +[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Int. Conf. Learn. Represent., 2018. 2, 4, 8 +[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4, 5, 6, 8 +[38] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3, 4, 6 +[39] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Adv. Neural Inform. Process. Syst., 2021. 2, 4 +[40] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2 +[41] Hyunjik Kim, George Papamakarios, and Andriy Mnih. The lipschitz constant of self-attention. In Int. Conf. Mach. Learn., 2021. 3 +[42] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In Int. Conf. Learn. Represent., 2014. 1 + +[43] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2 +[44] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. In Int. Conf. Learn. Represent., 2022. 3 +[45] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before Fuse: Vision and language representation learning with momentum distillation. In Adv. Neural Inform. Process. Syst., 2021. 3 +[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis., 2014. 2, 6 +[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 6 +[48] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Adv. Neural Inform. Process. Syst., 2022. 2 +[49] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 5 +[50] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5 +[51] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for GANs do actually converge? In Int. Conf. Mach. Learn., 2018. 5 +[52] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2 +[53] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Int. Conf. Learn. Represent., 2018. 1 +[54] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In Int. Conf. Comput. Vis., 2019. 2 +[55] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., 2022. 2, 7 +[56] David W. Opitz and Richard Maclin. Popular ensemble methods: An empirical study. J. Artif. Intell. Res., 11:169-198, 1999. 2 +[57] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. + +StyleSDF: High-resolution 3D-consistent image and geometry generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., 2021. 2, 3, 5 +[60] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Int. Conf. Mach. Learn., 2021. 1, 2, 7 +[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2 +[62] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Int. Conf. Mach. Learn., 2016. 5 +[63] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Adv. Neural Inform. Process. Syst., 2021. 3 +[64] Lior Rokach. Ensemble-based classifiers. Artif. Intell. Rev., 33(1-2):1-39, 2010. 2 +[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 7 +[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4, 5, 6 +[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2, 4, 7 +[68] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Int. Conf. Comput. Vis., 2017. 2 +[69] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 2 +[70] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. Adv. Neural Inform. Process. Syst., 2016. 1 + +[71] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGAN-XL: Scaling StyleGAN to large diverse datasets. In ACM SIGGRAPH, 2022. 2, 3, 4, 6 +[72] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In Int. Conf. Mach. Learn., 2023. 2, 3, 6, 7 +[73] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In Eur. Conf. Comput. Vis., 2025. 2 +[74] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 2, 6 +[75] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Adv. Neural Inform. Process. Syst., 2020. 2 +[76] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3D-aware image synthesis with sparse voxel grids. In Adv. Neural Inform. Process. Syst., 2022. 2 +[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Association for Computational Linguistics, 2018. 6 +[78] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Int. Conf. Learn. Represent., 2017. 2, 3, 8 +[79] Sheng Shen, Zhewei Yao, Chunyuan Li, Trevor Darrell, Kurt Keutzer, and Yuxiong He. Scaling vision-language models with sparse mixture of experts. arXiv preprint arXiv:2303.07226, 2023. 3 +[80] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2 +[81] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2 +[82] Yujun Shen, Zhiyi Zhang, Dingdong Yang, Yinghao Xu, Ceyuan Yang, and Jiapeng Zhu. Hammer: An efficient toolkit for training deep models. https://github.com/bytedance/Hammer, 2022.6 +[83] Zifan Shi, Sida Peng, Yinghao Xu, Geiger Andreas, Yiyi Liao, and Yujun Shen. Deep generative models on 3D representations: A survey. arXiv preprint arXiv:2210.15663, 2022. 2 +[84] Zifan Shi, Yinghao Xu, Yujun Shen, Deli Zhao, Qifeng Chen, and Dit-Yan Yeung. Improving 3D-aware image synthesis with a geometry-aware discriminator. In Adv. Neural Inform. Process. Syst., 2022. +[85] Zifan Shi, Yujun Shen, Yinghao Xu, Sida Peng, Yiyi Liao, Sheng Guo, Qifeng Chen, and Dit-Yan Yeung. Learning + +3d-aware image synthesis with unknown pose distribution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. +[86] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3D GANs. In Adv. Neural Inform. Process. Syst., 2022. 2 +[87] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Int. Conf. Mach. Learn., 2015. 1 +[88] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2020. 2 +[89] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Int. Conf. Mach. Learn., 2023. 2 +[90] Ryohei Suzuki, Masanori Koyama, Takeru Miyato, Taizan Yonetsuji, and Huachun Zhu. Spatially controllable image synthesis with internal representation collaging. arXiv preprint arXiv:1811.10153, 2018. 2 +[91] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2818-2826, 2016. 6 +[92] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2019. 3 +[93] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. GALIP: Generative adversarial CLIPs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2 +[94] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7 +[95] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 6 +[96] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2 +[97] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In Adv. Neural Inform. Process. Syst., 2016. 1 +[98] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Adv. Neural Inform. Process. Syst., 2017. 1 +[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Adv. Neural Inform. Process. Syst., 2017. 3 +[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2 + +[101] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2 +[102] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2 +[103] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3D-aware image synthesis via learning structural and textural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[104] Yiran Xu, Taesung Park, Richard Zhang, Yang Zhou, Eli Shechtman, Feng Liu, Jia-Bin Huang, and Difan Liu. Videogigagan: Towards detail-rich video super-resolution. arXiv preprint arXiv:2404.12388, 2024. 2 +[105] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. Int. J. Comput. Vis., 2020. 2 +[106] Ceyuan Yang, Qihang Zhang, Yinghao Xu, Jiapeng Zhu, Yujun Shen, and Bo Dai. Learning modulated transformation in GANs. arXiv preprint arXiv:2308.15472, 2023. 3, 4 +[107] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2 +[108] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022. 1, 2, 7 +[109] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5 +[110] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In Int. Conf. Mach. Learn., pages 7354–7363, 2019. 2 +[111] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[112] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiaxiang Gu, Jinhui Xu, and Tong Sun. LAFITE: Towards language-free training for text-to-image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 7 +[113] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zhengjun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in GANs. In Adv. Neural Inform. Process. Syst., 2021. 2 +[114] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in GANs. In Int. Conf. Mach. Learn., 2022. +[115] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the + +natural image manifold. In Eur. Conf. Comput. Vis., 2016. 2 \ No newline at end of file diff --git a/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/images.zip b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3f8ce3f588c2ec94733badc47db156c211490e57 --- /dev/null +++ b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f11770543ab64470fdaa31aa0a0f0672d1ff9ae7d6ece565450ecc97b21a7a00 +size 822192 diff --git a/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/layout.json b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dcf13f4130edd278a3a15af2a7de4cf61a437b03 --- /dev/null +++ b/2025/Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis/layout.json @@ -0,0 +1,10764 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "text", + "content": "Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 175, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 175, + 158 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 175, + 158 + ], + "type": "text", + "content": "Jiapeng Zhu1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 143, + 259, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 143, + 259, + 171 + ], + "spans": [ + { + "bbox": [ + 187, + 143, + 259, + 171 + ], + "type": "text", + "content": "Ceyuan Yang3 \nYifei Zhang1,4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 271, + 143, + 352, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 143, + 352, + 171 + ], + "spans": [ + { + "bbox": [ + 271, + 143, + 352, + 171 + ], + "type": "text", + "content": "Kecheng Zheng1 \nQifeng Chen2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 359, + 143, + 436, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 143, + 436, + 172 + ], + "spans": [ + { + "bbox": [ + 359, + 143, + 436, + 172 + ], + "type": "text", + "content": "Yinghao Xu" + }, + { + "bbox": [ + 359, + 143, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 359, + 143, + 436, + 172 + ], + "type": "text", + "content": " \nYujun Shen" + }, + { + "bbox": [ + 359, + 143, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 449, + 143, + 507, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 449, + 143, + 507, + 156 + ], + "spans": [ + { + "bbox": [ + 449, + 143, + 507, + 156 + ], + "type": "text", + "content": "Zifan Shi" + }, + { + "bbox": [ + 449, + 143, + 507, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 175, + 176, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 175, + 176, + 191 + ], + "spans": [ + { + "bbox": [ + 117, + 175, + 176, + 191 + ], + "type": "text", + "content": "1Ant Group" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 175, + 247, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 175, + 247, + 189 + ], + "spans": [ + { + "bbox": [ + 200, + 175, + 247, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 200, + 175, + 247, + 189 + ], + "type": "text", + "content": "HKUST" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 271, + 175, + 312, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 175, + 312, + 189 + ], + "spans": [ + { + "bbox": [ + 271, + 175, + 312, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 271, + 175, + 312, + 189 + ], + "type": "text", + "content": "CUHK" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 336, + 175, + 490, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 175, + 490, + 191 + ], + "spans": [ + { + "bbox": [ + 336, + 175, + 490, + 191 + ], + "type": "text", + "content": "4Shanghai Jiao Tong Univeristy" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 218, + 265, + 425 + ], + "blocks": [ + { + "bbox": [ + 58, + 218, + 265, + 425 + ], + "lines": [ + { + "bbox": [ + 58, + 218, + 265, + 425 + ], + "spans": [ + { + "bbox": [ + 58, + 218, + 265, + 425 + ], + "type": "image", + "image_path": "dfb66adaf22deacd14135b67a068e138504a8804a86a1e608be3db7eb3d0d558.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 426, + 192, + 434 + ], + "lines": [ + { + "bbox": [ + 60, + 426, + 192, + 434 + ], + "spans": [ + { + "bbox": [ + 60, + 426, + 192, + 434 + ], + "type": "text", + "content": "A black and white photo of a dog with big eyes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 266, + 218, + 360, + 310 + ], + "blocks": [ + { + "bbox": [ + 266, + 218, + 360, + 310 + ], + "lines": [ + { + "bbox": [ + 266, + 218, + 360, + 310 + ], + "spans": [ + { + "bbox": [ + 266, + 218, + 360, + 310 + ], + "type": "image", + "image_path": "ee1e7d9be98346aee67f0576d33f31be2ed149d288837b2ab934e57f0214eba1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 311, + 359, + 326 + ], + "lines": [ + { + "bbox": [ + 269, + 311, + 359, + 326 + ], + "spans": [ + { + "bbox": [ + 269, + 311, + 359, + 326 + ], + "type": "text", + "content": "A photo of a colorful sports car, with mountain background." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 267, + 327, + 360, + 418 + ], + "blocks": [ + { + "bbox": [ + 267, + 327, + 360, + 418 + ], + "lines": [ + { + "bbox": [ + 267, + 327, + 360, + 418 + ], + "spans": [ + { + "bbox": [ + 267, + 327, + 360, + 418 + ], + "type": "image", + "image_path": "a51ccde226dac9754610c73802c93c5e3aee485d9173acc14ce01cd2a422f7f0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 419, + 351, + 434 + ], + "lines": [ + { + "bbox": [ + 269, + 419, + 351, + 434 + ], + "spans": [ + { + "bbox": [ + 269, + 419, + 351, + 434 + ], + "type": "text", + "content": "Sunrise over the clouds in the mountains." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 440, + 553, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 440, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 553, + 452 + ], + "type": "text", + "content": "Figure 1. Example results at " + }, + { + "bbox": [ + 55, + 440, + 553, + 452 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 440, + 553, + 452 + ], + "type": "text", + "content": " resolution synthesized by our proposed Aurora, a large-scale GAN-based text-to-image generator." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 361, + 218, + 454, + 310 + ], + "blocks": [ + { + "bbox": [ + 361, + 218, + 454, + 310 + ], + "lines": [ + { + "bbox": [ + 361, + 218, + 454, + 310 + ], + "spans": [ + { + "bbox": [ + 361, + 218, + 454, + 310 + ], + "type": "image", + "image_path": "755b522bbd2b2962a16f0eaba4182c4192565977b33c08cea7be6f054b48019f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 361, + 326, + 454, + 418 + ], + "blocks": [ + { + "bbox": [ + 363, + 311, + 449, + 326 + ], + "lines": [ + { + "bbox": [ + 363, + 311, + 449, + 326 + ], + "spans": [ + { + "bbox": [ + 363, + 311, + 449, + 326 + ], + "type": "text", + "content": "A steaming plate of classic fried rice with vegetables, eggs." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 326, + 454, + 418 + ], + "lines": [ + { + "bbox": [ + 361, + 326, + 454, + 418 + ], + "spans": [ + { + "bbox": [ + 361, + 326, + 454, + 418 + ], + "type": "image", + "image_path": "586a04845d97e4ef594442235b7360a7b200931700e30b02ab5db169b928643d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 418, + 445, + 434 + ], + "lines": [ + { + "bbox": [ + 362, + 418, + 445, + 434 + ], + "spans": [ + { + "bbox": [ + 362, + 418, + 445, + 434 + ], + "type": "text", + "content": "A coral reef with starfish, sea anemones and fish." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 455, + 218, + 548, + 311 + ], + "blocks": [ + { + "bbox": [ + 455, + 218, + 548, + 311 + ], + "lines": [ + { + "bbox": [ + 455, + 218, + 548, + 311 + ], + "spans": [ + { + "bbox": [ + 455, + 218, + 548, + 311 + ], + "type": "image", + "image_path": "11b332b209e7fb2ea9a32c89e5f98f4dc57f087c55492b92ab4075321a45f89e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 311, + 546, + 326 + ], + "lines": [ + { + "bbox": [ + 457, + 311, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 457, + 311, + 546, + 326 + ], + "type": "text", + "content": "Oil painting of a flower garden, distant cottage, soft sunlight." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 456, + 327, + 548, + 418 + ], + "blocks": [ + { + "bbox": [ + 456, + 327, + 548, + 418 + ], + "lines": [ + { + "bbox": [ + 456, + 327, + 548, + 418 + ], + "spans": [ + { + "bbox": [ + 456, + 327, + 548, + 418 + ], + "type": "image", + "image_path": "889c1bf433a3700bace27f2c773a2768c62b5167e7d91dba223dff1f550c8084.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 418, + 548, + 434 + ], + "lines": [ + { + "bbox": [ + 457, + 418, + 548, + 434 + ], + "spans": [ + { + "bbox": [ + 457, + 418, + 548, + 434 + ], + "type": "text", + "content": "A pencil sketch of a woman with deep, almond-shaped eyes in." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 152, + 473, + 200, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 473, + 200, + 486 + ], + "spans": [ + { + "bbox": [ + 152, + 473, + 200, + 486 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 53, + 501, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 297, + 704 + ], + "type": "text", + "content": "Due to the difficulty in scaling up, generative adversarial networks (GANs) seem to be falling out of grace with the task of text-conditioned image synthesis. Sparsely activated mixture-of-experts (MoE) has recently been demonstrated as a valid solution to training large-scale models with limited resources. Inspired by this, we present Aurora, a GAN-based text-to-image generator that employs a collection of experts to learn feature processing, together with a sparse router to adaptively select the most suitable expert for each feature point. We adopt a two-stage training strategy, which first learns a base model at " + }, + { + "bbox": [ + 53, + 501, + 297, + 704 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 53, + 501, + 297, + 704 + ], + "type": "text", + "content": " resolution followed by an upsampler to produce " + }, + { + "bbox": [ + 53, + 501, + 297, + 704 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 53, + 501, + 297, + 704 + ], + "type": "text", + "content": " images. Trained with only public data, our approach encouragingly closes the performance gap between GANs and industry-level diffusion models, maintaining a fast inference speed. We release the code and checkpoints here to facilitate the community for further development." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 473, + 394, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 473, + 394, + 486 + ], + "spans": [ + { + "bbox": [ + 315, + 473, + 394, + 486 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 495, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 616 + ], + "type": "text", + "content": "As one of the most popular tasks in the field of content generation, text-conditioned image synthesis [3, 12, 19, 34, 60, 61, 65, 67, 100, 108] allows users to customize a picture with natural language instructions. Among diverse types of generative models, such as variational autoencoders (VAEs) [42], autoregressive models (ARs) [97, 98], and generative adversarial networks (GANs) [23], diffusion models [15, 28, 87] have blossomed into a dominant solution with a thriving community (e.g., stable diffusion [65]), leaving few efforts on improving other frameworks." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "In this work, we put our focus on GANs, which are usually criticized for training instability [1, 53] and limited mode coverage [70]. However, we argue that the task of text-to-image (T2I) generation is more like a translation task (i.e., shifting an informative data distribution to another) instead of a pure generation task (i.e., modeling the observed distribution with random noise as the input), and accordingly that the introduction of the text condition" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18411" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "can already help alleviate the above two issues to some extent, which is also empirically supported by some recent studies [34, 72]. Even so, the difficulty in scaling up the model size still manifests itself as a major drawback of GANs. Concretely, unlike diffusion models that produce an image through an iterative process [15, 28, 65], the generator of a GAN typically adopts a feedforward network to render an image. Considering the limited computational resources (e.g., GPU memory) in practice, directly learning a sufficiently large feedforward model remains challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 294, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 278 + ], + "type": "text", + "content": "The advent of sparse mixture-of-experts (MoE) offers a promising way to enlarge the model scale with affordable computing cost [20, 78]. The rationale behind is to employ multiple expert modules yet only activate one or a few of them for each feedforward execution. That way, the growth of the number of modules could expand the model capacity but barely increase the number of computations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "text", + "content": "We decently introduce the philosophy of sparse MoE into the design of GAN generators and propose Aurora for fast T2I generation. Specifically, given an intermediate feature map at a particular level of degree, we engage a set of experts, each of which is originally designed to process a single-point feature, as well as a highly sparse router to help select the most suitable expert for each feature point. It is noteworthy that, different from most existing spare MoE methods [20, 32, 78] that determine the routing path simply based on the information from the input feature map, our router also takes into account the text-integrated global latent code such that it can adaptively make the best decision via fully considering the sampling stochasticity and the text condition. Additionally, Aurora employs a two-stage pipeline [34, 67] involving a base generator and an upsampler. The base generator creates images at a resolution of " + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "text", + "content": ", are then enhanced by an upsampler to produce final images at " + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "text", + "content": ". Our base generator learned on LAION2B-en [74] and COYO-700M [8] achieves 6.2 zero-shot FID [27] on MS COCO [46] at the resolution of " + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "text", + "content": ", surpassing other existing methods. The upsampler further refines the image to a final resolution of " + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 56, + 281, + 295, + 567 + ], + "type": "text", + "content": " pixels, achieving a zero-shot FID score of 8.74 on MS COCO, which sets a new benchmark for GANs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "content": "Besides the fast inference speed, GANs also enjoy a well-studied latent space [33, 81, 105] and the high flexibility to incorporate other priors (e.g., involving the volumetric rendering pipeline for 3D-aware image synthesis) [9, 10, 22, 24, 54, 57, 75, 76, 83-86, 103]. However, the lack of an open-sourced large-scale GAN model hinders such a powerful tool from being adequately developed and analyzed. We thus release the code and checkpoints to facilitate the research community and plan for continuous improvement from the perspectives of both performance and functionality. We believe that our work could encourage more studies in the field of visual content generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 553, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 305 + ], + "type": "text", + "content": "Generative adversarial networks (GANs). Formulated as a two-player game between a generator and a discriminator, GANs [23] are designed to model the observed data distribution through adversarial training. Thanks to the flexible formulation (i.e., there are no particular assumptions on the data format or on the generation process), GANs have shown great potential in a variety of visual generative tasks, such as image generation [7, 36-39], image editing [4, 25, 33, 80, 81, 90, 101, 102, 105, 113-115], video generation [68, 96, 104], and 3D-aware image synthesis [9, 10, 24, 54, 75, 83, 103]. Some studies also explore conditional GANs [52], which incorporate signals like class labels [7, 71, 110], texts [34, 72, 93, 112], and reference images [31, 43, 107] into the generator for interactive control. Despite the aforementioned success, GANs usually struggle in learning from large-scale, diverse datasets (e.g., LAION [74]) due to the difficulty in scaling up [34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 307, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 569 + ], + "type": "text", + "content": "Text-conditioned image synthesis. With the aid of visual-language models like CLIP [59], text-to-image (T2I) generation [2, 3, 11, 12, 19, 26, 34, 58, 60, 61, 65, 67, 108] manages to decode an open-vocabulary description to a vivid image, and hence allows users to visualize their creativity and imagination. Towards such a challenging task, existing efforts usually opt to trade computing efficiency for visual quality. For example, autoregressive models [16, 21, 26, 60, 100, 108] and diffusion models [3, 12, 19, 55, 58, 61, 65, 67] have demonstrated impressive results through an iterative generation process, which can be time-consuming and costly to make a synthesis. Differently, GANs employ a feedforward network as the generator and thus naturally support fast generation without having to use an acceleration [40, 48, 88] or distillation [69, 73, 89] algorithm. Some recent attempts [34, 72] have confirmed the promise of GANs in text-to-image synthesis yet left the issue of enlarging the model capacity unsolved. In addition, there currently lacks a publicly available GAN-based T2I model in the community, making it difficult to fully study the properties of GANs, let alone integrate the advantages of GANs into other types of generative models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Mixture-of-experts (MoE). Sharing a similar philosophy with ensemble [56, 64], MoE [32] learns a set of expert modules to process the input signal, together with a gating module (also known as a router) to help combine the outputs of different experts. Under such a design, each expert is expected to focus on the learning of a sub-region of the input space, while the router is responsible for the weight assignment across experts [18]. Sparsely activated MoE [17, 20, 78] distinguishes itself from other variants of MoE because of the discrete routing mechanism. By selecting only one or a few experts from a large candidate pool for each incoming example, sparse MoE offers a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18412" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "feasible solution to scaling up the model size with barely increased computational cost [20, 63, 79]. In this work, we introduce the idea of sparse MoE into the design of a GAN generator and adequately adapt the routing mechanism for the task of text-to-image generation by taking into account the text-integrated stochasticity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 111, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 111, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 111, + 167 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "text", + "content": "Recall that we aim to develop an efficient text-to-image (T2I) generator based on the framework of GANs [23], which consists of a generator and a discriminator competing with each other. On the one hand, the generator takes in a global latent code, " + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "text", + "content": ", along with a text description (e.g., a sentence), " + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "text", + "content": ", and outputs an image, " + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = G(\\mathbf{z},\\mathbf{c})" + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "text", + "content": ". On the other hand, the discriminator takes in an image together with an associated text and differentiates synthesized samples from real samples using a realiznescore, " + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "inline_equation", + "content": "D(\\mathbf{x},\\mathbf{c})" + }, + { + "bbox": [ + 55, + 175, + 296, + 392 + ], + "type": "text", + "content": ". As discussed above, the major contribution of this work is to enlarge the capacity of the base generator; therefore, we directly borrow the discriminator architecture from the prior art [34]. In the following, we first introduce the common unit blocks utilized in both the base generator and the upsampler (Sec. 3.1). We then provide a detailed overview of the architecture of our base generator (Sec. 3.2) and upsampler (Sec. 3.3), followed by a discussion of the supervision signals and training strategy (Sec. 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 183, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 183, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 183, + 411 + ], + "type": "text", + "content": "3.1. Unit Generative Block" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 417, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 513 + ], + "type": "text", + "content": "Like most modern deep models [37, 99], our base generator and upsampler follow the principle of block-wise design. As shown in Fig. 2, a generative unit stacks a convolution block and an attention block to process the input feature map, " + }, + { + "bbox": [ + 55, + 417, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{in}" + }, + { + "bbox": [ + 55, + 417, + 296, + 513 + ], + "type": "text", + "content": ", where the attention block adopts the sparse routing mechanism for scaling up. Additionally, in this section, " + }, + { + "bbox": [ + 55, + 417, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 417, + 296, + 513 + ], + "type": "text", + "content": " refers to the intermediate latent code and will be introduced in detail in Sec. 3.2 and Sec. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "text", + "content": "Text encoder. We employ CLIP [59] as the text encoder, which transforms the input text, " + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "text", + "content": ", into both a global token, " + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "text", + "content": ", and a token sequence, " + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{seq}" + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "text", + "content": " is the pooled text embedding that encapsulates the meaning of the entire sentence, while " + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{seq}" + }, + { + "bbox": [ + 55, + 513, + 296, + 597 + ], + "type": "text", + "content": " is derived from the penultimate layer of the text encoder, capturing information for each individual word in the text. We formulate these as follows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 607, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 607, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 127, + 607, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {t} _ {g}, \\mathbf {t} _ {\\text {s e q}} \\right\\} = \\operatorname {C L I P} (\\mathbf {c}). \\tag {1}", + "image_path": "211793bd63feb9efc0e7bc390f9e50f73d4da77ac9d3b2af30d4951d33d9a0b4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Convolution block. Latent modulation [37], which is developed from adaptive instance normalization (AdaIN) [29], has demonstrated its effectiveness in the generation field and hence serves as a basic operation in GAN-based models [34, 71, 72]. Such a technique is further improved by integrating it into the convolution operation with demodulation (i.e., modulated convolution," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "ModConv) [38] and by equipping it with learnable feature deformation (i.e., modulated transformation module, MTM) [106]. We employ two MTMs in the convolution block and construct the block using skip connection. Consequently, the output of the convolutional block is given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 150, + 555, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 150, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 355, + 150, + 555, + 171 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\text {c o n v}} = \\operatorname {M T M} \\left(\\operatorname {M T M} \\left(\\mathbf {f} _ {\\text {i n}}, \\mathbf {w}\\right), \\mathbf {w}\\right) + \\mathbf {f} _ {\\text {i n}}. \\tag {2}", + "image_path": "734da0591ea60afabbd86225a266443ea187772ce9a6dcc41abad847127e6420.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "type": "text", + "content": "Notably, the learnable offsets in MTM [106] are only introduced when the feature resolution is equal to or less than " + }, + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "content": "Attention block. We can tell from Eq. (2) that, in the convolution block, the text condition, " + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "content": ", only interacts with the generative features through the global token, " + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "content": " in Eq. (7) as an intermediary. To also make sufficient use of the text token sequence, " + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{seq}" + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "content": ", we introduce an attention block for text-image interaction, which is widely used in multi-modality learning [45, 92]. Specifically, the attention block consists of a self-attention layer, a cross-attention layer, and a feedforward network [99]. We thereby have" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 326, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 326, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 326, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {s a} = \\operatorname {S e l f A t t n} \\left(\\mathbf {f} _ {\\text {c o n v}}, \\mathbf {w}\\right) + \\mathbf {f} _ {\\text {c o n v}}, \\tag {3}", + "image_path": "68293aae2cc751524e5b70ce618966f3139e7cd68a4b4202056775381a823a84.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 341, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 341, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 361, + 341, + 553, + 354 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {c a} = \\operatorname {C r o s s A t t n} \\left(\\mathbf {f} _ {s a}, \\mathbf {t} _ {\\text {s e q}}\\right) + \\mathbf {f} _ {s a}, \\tag {4}", + "image_path": "478b5f15e269476073995404f131418f29934ad39a3566edceba5d158c42aee5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 356, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 356, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 354, + 356, + 553, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {f f n} = \\mathrm {F F N} \\left(\\mathbf {f} _ {c a}\\right) + \\mathbf {f} _ {c a}, \\tag {5}", + "image_path": "a1c98199f6c0d472c9a271f9eb27fe9e69c69bdaa58a7e2683628f34c4a50c82.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathsf{FFN}" + }, + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "text", + "content": " in Eq. (5) processes each feature point individually using a two-layer MLP. Here, we follow the prior work [34, 44] that employs " + }, + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\ell 2" + }, + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "text", + "content": "-distance instead of the dot product for attention logits to stabilize training [41]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "text", + "content": "Multiple experts and sparse routing. Recall that the main motivation of this work is to enlarge the capacity of the base generator in GANs. One straightforward solution is to make the network deeper or wider. However, it may cause an out-of-memory issue. Inspired by the practice in natural language processing (NLP) [20], we incorporate the philosophy of sparse MoE [78] into the design of our unit generative block to enlarge the base generator capacity. Concretely, we treat the FFN in Eq. (5) as an expert, which specializes in processing a single-point feature, and engage " + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "text", + "content": " experts in total, " + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\{ \\mathsf{FFN}_j \\}_{j=1}^N" + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "text", + "content": ", forming a candidate pool. We then employ a sparse router to help select the most suitable expert for each feature point. Unlike existing sparse MoE methods [17, 20, 78], our router adaptively makes its decision by taking into account not only the input feature but also the text-integrated global latent code, " + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "text", + "content": ", which better matches the target of T2I generation (i.e., decoding the text condition and the sampling stochasticity to an image). Under such a routing strategy, we can reformulate Eq. (5) as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 667, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 667, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 318, + 667, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {f f n} ^ {(k)} = \\operatorname {F F N} _ {j} \\left(\\mathbf {f} _ {c a} ^ {(k)}\\right) + \\mathbf {f} _ {c a} ^ {(k)} \\text {s . t .} j = \\operatorname {R o u t e r} \\left(\\mathbf {f} _ {c a} ^ {(k)}, \\mathbf {w}\\right). \\tag {6}", + "image_path": "fa2f319f7d1da2e7d3540f7cdb3c9a05ad505174181b818a42dd6810d0d97d72.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Here, the superscript " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "-th feature point. In addition to increasing the capacity of the generator, MoE" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18413" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 76, + 558, + 262 + ], + "blocks": [ + { + "bbox": [ + 57, + 76, + 558, + 262 + ], + "lines": [ + { + "bbox": [ + 57, + 76, + 558, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 76, + 558, + 262 + ], + "type": "image", + "image_path": "f4903de1974a77a4a90d05484159018fa3d202c7baa518e077f5ffb5e3b8d7af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "text", + "content": "Figure 2. Illustration of base generator in Aurora. First, we utilize a pretrained CLIP text encoder to obtain the text embeddings " + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{seq}" + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "text", + "content": ". Then, starting from the Fourier feature, the generator processes it through several unit generative blocks. Each unit block is composed of a convolution block and an attention block. There are two MTM [106] layers within the convolution block. In the attention block, the FFN acts as the expert, specializing in processing a single-point feature. A total of " + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "text", + "content": " experts are assembled to form a candidate pool, from which selections are made by the router based on the input features and the latent code " + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 54, + 268, + 555, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 339, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 294, + 376 + ], + "type": "text", + "content": "also frees the generator from modeling complex data distributions by assigning different objects or parts of an image to specialized experts, reducing the learning difficulty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 151, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 151, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 151, + 393 + ], + "type": "text", + "content": "3.2. Base Generator" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 434 + ], + "type": "text", + "content": "The base generator comprises two components: a mapping network and a synthesis network. Both networks accept text and latent code as inputs, which we will introduce next." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": "Mapping network. As pointed out by StyleGAN [37], utilizing a learnable MLP-based mapping network on the latent code, " + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": ", could create a more disentangled latent space, " + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": ", than the native latent space, " + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": ". In the class-conditional generation, a class label is first concatenated with the latent code " + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": " before being passed to the mapping network [37, 71]. Under the T2I setting, it is necessary to shift from class conditioning to text conditioning to make the sampling process better aware of the text information. Hence, we concatenate " + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": " with the sampled latent code " + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": " for further processing, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 574, + 295, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 574, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 117, + 574, + 295, + 587 + ], + "type": "interline_equation", + "content": "\\mathbf {w} = \\operatorname {M L P} \\left(\\operatorname {c o n c a t} \\left(\\mathbf {z}, \\mathbf {t} _ {g}\\right)\\right), \\tag {7}", + "image_path": "21de8a894046a17594459725791a89771bc3dfc78ef56d83ddc9d73dc5828f29.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 274, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 274, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 274, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 274, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 594, + 274, + 605 + ], + "type": "text", + "content": " stands for the more disentangled latent code." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Synthesis network. The synthesis network starts with Fourier features [39] and is subsequently followed by a series of convolutional blocks, attention blocks, including the MoE, and finally outputs the image at a resolution of " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", as shown in Fig. 2. The " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " generated by the mapping network is used to modulate the convolutional and self-attention layers, and it also functions in the MoE to select the appropriate expert for each input feature point. Additionally, the local token " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{seq}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", is incorporated into the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 339, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 460 + ], + "type": "text", + "content": "cross-attention layers as the textual condition to further enhance text-image alignment. The generator produces RGB images at each resolution and employs output skip connections [38]. Unlike StyleGAN2 [38], which only utilizes the highest resolution image as the final output, our approach takes advantage of all generated images, similar to MSG-GAN [35] and GigaGAN [34]. Each image is then used to calculate the training loss, mirroring the progressive training approach described in [36], which can improve the final performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 389, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 389, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 389, + 481 + ], + "type": "text", + "content": "3.3. Upsampler" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": "The base generator produces images at a resolution of " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": ". To create high-resolution images, we train a text-conditioned super-resolution model that upsamples the images generated by the base generator to a resolution of " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": ". The architectures of the upsampler are derived from the base model with specific modifications. First, additional convolution blocks are integrated to process the input image, resulting in a final architecture resembling a U-Net. Second, we set the number of experts to 1 based on empirical evidence indicating that this model does not require an extensive parameter count. Third, instead of acquiring the global code (i.e., " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " in base generator) from the mapping network, we obtain it from the encoder part of the U-Net, which is then concatenated with " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " to modulate the decoder part of the U-Net. Additionally, to enhance memory efficiency and increase the speed of the upsampler, following Imagen [67], we reverse the order of the downsampling and upsampling operations in our U-Net architecture compared to the approach used in prior art [66]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18414" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 218, + 212 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 218, + 212 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 218, + 212 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 218, + 212 + ], + "type": "image", + "image_path": "5706d4f00532f59629316fadc9e3819270fe63690f104f498b29c7e586e49a97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 213, + 205, + 236 + ], + "lines": [ + { + "bbox": [ + 62, + 213, + 205, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 205, + 236 + ], + "type": "text", + "content": "A lighthouse under a starry night sky." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 72, + 382, + 212 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 382, + 212 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 382, + 212 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 382, + 212 + ], + "type": "image", + "image_path": "bb6f7377e22e5a8ba20bdf93d04c2a20bac8fd9c4e41bed32b3a707318e03591.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 213, + 375, + 235 + ], + "lines": [ + { + "bbox": [ + 226, + 213, + 375, + 235 + ], + "spans": [ + { + "bbox": [ + 226, + 213, + 375, + 235 + ], + "type": "text", + "content": "A painting of flowers in a blue vase on a table." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 546, + 213 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 546, + 213 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 546, + 213 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 546, + 213 + ], + "type": "image", + "image_path": "d34b245c552d56f490fc8859aa59a338a901dafdf94e8e72154eee7e74e65196.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 213, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 390, + 213, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 390, + 213, + 547, + 237 + ], + "type": "text", + "content": "An oil painting depicting a large crowd of people in front of a church." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 237, + 219, + 378 + ], + "blocks": [ + { + "bbox": [ + 59, + 237, + 219, + 378 + ], + "lines": [ + { + "bbox": [ + 59, + 237, + 219, + 378 + ], + "spans": [ + { + "bbox": [ + 59, + 237, + 219, + 378 + ], + "type": "image", + "image_path": "09e08fc0b6c22dab3c7342bb1f664964fab04bd310927ca217c2a1f90d861cab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 380, + 208, + 403 + ], + "lines": [ + { + "bbox": [ + 62, + 380, + 208, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 208, + 403 + ], + "type": "text", + "content": "A road that is surrounded by trees with yellow leaves." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 237, + 383, + 378 + ], + "blocks": [ + { + "bbox": [ + 223, + 237, + 383, + 378 + ], + "lines": [ + { + "bbox": [ + 223, + 237, + 383, + 378 + ], + "spans": [ + { + "bbox": [ + 223, + 237, + 383, + 378 + ], + "type": "image", + "image_path": "04cf4a994a2d64b21f15b47f024099ddc80521927791bf4ae322e1eb92baa0ea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 380, + 352, + 404 + ], + "lines": [ + { + "bbox": [ + 226, + 380, + 352, + 404 + ], + "spans": [ + { + "bbox": [ + 226, + 380, + 352, + 404 + ], + "type": "text", + "content": "Sunset casts a golden glow on \nmajestic mountain peaks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 386, + 238, + 546, + 378 + ], + "blocks": [ + { + "bbox": [ + 386, + 238, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 386, + 238, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 386, + 238, + 546, + 378 + ], + "type": "image", + "image_path": "ad99aa22ecbe136b116b845fc712ba40d7effc81e016eb176c1aef92e5fbd166.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 380, + 532, + 404 + ], + "lines": [ + { + "bbox": [ + 390, + 380, + 532, + 404 + ], + "spans": [ + { + "bbox": [ + 390, + 380, + 532, + 404 + ], + "type": "text", + "content": "Landscape, lake, buildings, bridge, Chinese ink style." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "type": "text", + "content": "Figure 3. Diverse results generated by Aurora, In each group, the images on the left are randomly generated by our base generator at the resolution of " + }, + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "type": "text", + "content": ". We also highlight an example at " + }, + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "type": "text", + "content": " on the right to demonstrate high-quality synthesis capability." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 452, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 295, + 632 + ], + "type": "text", + "content": "It is noteworthy that the upsampler is a super-resolution model primarily trained on real images. Therefore, it is crucial to minimize the gap between real images and those generated by the base generator. We address this gap through three aspects. First, we augment the input data by introducing additional augmentations, such as Gaussian noise and Gaussian blurring. Second, the skip connection is removed since it provides linearity between the input and output [49], which helps reduce discrepancies when transitioning from real to synthesized images. Third, we increase the downsampling rate beyond the conventional rate of 8, lowering the feature map resolution from the standard " + }, + { + "bbox": [ + 55, + 452, + 295, + 632 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 452, + 295, + 632 + ], + "type": "text", + "content": " [34, 66], which produces more abstracted features by omitting details in the input image, further bridging the gap between real and synthesized images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 646, + 237, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 237, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 237, + 659 + ], + "type": "text", + "content": "3.4. Learning Objectives and Training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The primary focus of this work is to scale up the model size of a GAN generator instead of exploring a better supervision signal for GAN training. Thus, we directly inherit some well-studied loss functions from prior arts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 452, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 553, + 475 + ], + "type": "text", + "content": "Still, to make the presentation self-contained, we list below all the objectives used in the training phase." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "text", + "content": "Adversarial loss. Adversarial training [23] is the core concept shared by all GANs and owns many variants in the literature [1, 5, 50]. We adopt the approach from GigaGAN [34], implementing a multi-scale input and multi-scale output logistic non-saturating loss [37] with " + }, + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 313, + 479, + 554, + 552 + ], + "type": "text", + "content": " penalty [51] as the regularizer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 554, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 554, + 651 + ], + "type": "text", + "content": "Matching-aware loss. Unlike the discriminator in unconditional GANs that rejects a sample by merely considering its plausibility, the discriminator in T2I GANs is also expected to take into account the alignment between the condition and the synthesized result. We follow prior arts [34, 62, 109] to employ a matching-aware discriminator and introduce the matching-aware loss that urges the discriminator to reject mismatched text-image pair." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": "Multi-level CLIP loss. CLIP loss [59] follows the idea of contrastive learning [13], and forces the feature of a generated image close to the feature of its associated text yet away from the feature of irrelevant text. We introduce a multi-level CLIP loss that measures CLIP loss at different" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18415" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "image resolutions, promoting increased text-image similarity. Such a supervision signal not only stabilizes training but also complements the matching-aware loss, aiding the generator in more effectively decoding the text condition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 120, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 120, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 54, + 120, + 294, + 203 + ], + "type": "text", + "content": "MoE loss. Recall that we introduce a collection of experts to enlarge the model capacity and expect them to process the input feature in various ways. For each network execution, only one expert will be activated for a feature point. To balance the workload across experts (i.e., to avoid that some experts are never activated), we introduce the MoE loss following Switch Transformer [20]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 204, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 295, + 384 + ], + "type": "text", + "content": "Training. We first train our base generator using the aforementioned losses. When training the upsampler, we incorporate an additional LPIPS loss [111], MoE loss is omitted because we set the number of experts to one for the upsampler. Furthermore, similar to the training procedure of the autoencoder in StableDiffusion [66], the adversarial loss is not employed at the beginning of the upsampler's training process. Once the training of the upsampler on real images is complete, we proceed to finetune it using synthesized images generated by the base generator with adversarial loss only. It is noteworthy that our approach adopts the mechanisms of both adversarial learning and sparse MoE, yet we do not encounter any instability throughout the entire training process, whether in the training of the base generator or the upsampler." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 393, + 136, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 136, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 136, + 407 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 498 + ], + "type": "text", + "content": "In this section, we evaluate the performance of our proposed Aurora both quantitatively (Sec. 4.2) and qualitatively (Sec. 4.3). Additionally, we provide detailed analyses to deepen the reader's understanding of text-to-image (T2I) GANs, discuss the significance of our project, and highlight how our open-sourced code and checkpoints can benefit the research community (Sec. 4.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 505, + 176, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 176, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 176, + 517 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": "Datasets. The training data is sourced from a diverse set of public datasets, including LAION2B-en [74], COYO-700M [8], CC3M [77], and YFCC15M [95]. All datasets are filtered based on the following criteria: clip score " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": ">0.3" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": ", resolution " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": ">512" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": ", aesthetic score " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": ">5.0" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": ", and an estimated watermark probability " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": "<0.5" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": ", which results in a total of 120M text-image pairs for training. Additionally, we use InstructBLIP [14] to obtain better captions for training our model. We do not use any self-collected data to simply pursue the improvement of synthesis quality. We believe that our open-sourced codebase can facilitate learning a more capable model using cleaner and higher-quality data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Training details. We develop our algorithm using Hammer [82]. The base generator was trained for one week on 320 NVIDIA A100 GPUs, while the upsampler underwent training for five days using 128 NVIDIA A100 GPUs. We" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 129, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "text", + "content": "Table 1. Ablation study of our base model, which targets generating images at " + }, + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "text", + "content": " resolution. FID [27] is employed as the evaluation metric, where a smaller number indicates better performance. “" + }, + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{10K}" + }, + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "text", + "content": "” reports the results on the training set, while “zero-shot " + }, + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{30K}" + }, + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "text", + "content": "” reports the results on MS COCO [46]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 129, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 224 + ], + "type": "table", + "html": "
ModelZero-shot FID30KFID10K# Params.
Base (64 GPUs)22.4420.770.367B
+ Better caption16.7214.310.367B
+ MoE13.8112.800.981B
+ ConvNeXt12.0311.100.986B
Scaling (128 GPUs)7.638.540.986B
Scaling (320 GPUs)6.207.960.986B
", + "image_path": "914e0e45e38658032d1bcc9d90b18e10078a917a6321eb4ab2910613a3dc6fc6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "text", + "content": "utilized the AdamW optimizer [47] with hyper-parameters set to " + }, + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0" + }, + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "text", + "content": ", and a weight decay of " + }, + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 313, + 241, + 555, + 338 + ], + "type": "text", + "content": ". The base generator and discriminator were jointly optimized with a learning rate of 0.0025, whereas the upsampler was trained with a learning rate of 0.001. We utilize OpenCLIP ConvNeXt-XXLarge [30] as our text encoder and for computing the CLIP loss during training. Additional training details are provided in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 347, + 432, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 432, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 432, + 360 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "text", + "content": "In this part, we quantitatively compare our Aurora with existing T2I generators to give readers an overall picture of its synthesis performance. Following prior arts [7, 34, 37, 38, 71, 72], we adopt FID [27] as the metric, which evaluates a generator at the distribution level. Here, we compute FID on two datasets to evaluate our model, one is \"FID" + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "inline_equation", + "content": "_{10K}" + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "text", + "content": "\", which is obtained by comparing the inception features [91] of " + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "text", + "content": " synthesized images and " + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "text", + "content": " real images randomly sampled from the training set, while the other \"zero-shot " + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{30K}" + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "text", + "content": "\" is obtained by comparing the inception features of " + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "inline_equation", + "content": "30K" + }, + { + "bbox": [ + 313, + 365, + 555, + 508 + ], + "type": "text", + "content": " synthesized images and the full validation set of MS COCO." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Base generator study. In Tab. 1, we report the evaluation results on two different data domains under different configurations of the base generator. We begin by implementing our base generator following the StyleGAN architecture and incorporate attention blocks as described in Sec. 3.1. Besides, we set the number of experts to 1 for this baseline model, resulting in 367M parameters. First, the baseline model is trained using the original captions, resulting in a zero-shot " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{30K}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " score of 22.44. Second, we retrain the model with re-captioned data, which significantly improves its performance. Following this, we increase the number of experts in the attention blocks, expanding the model size to 981M parameters. Additionally, we enlarge the CLIP text encoder by replacing the CLIP ViT-L/14 with OpenCLIP ConvNeXt-XXLarge, further increasing the model parameters. These two modifications result in our final base generator, which has 986M parameters and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18416" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 545, + 253 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 545, + 253 + ], + "type": "image", + "image_path": "c3d124f80939e902da3e5bf57d1e52cd9be54268cc9fcd7727b9325f4f06e413.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "type": "text", + "content": "Figure 4. Synthesized results through text prompt interpolation, where we fix the global latent code, " + }, + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "type": "text", + "content": ", and interpolate the text tokens, " + }, + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_{seq},\\mathbf{t}_g\\}" + }, + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "type": "text", + "content": ", extracted from two different text conditions, " + }, + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 257, + 553, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 329, + 293, + 426 + ], + "blocks": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "type": "text", + "content": "Table 2. Quantitative comparison between Aurora and existing methods at " + }, + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "type": "text", + "content": " resolution. “*” means that we downsample the samples using Lanczos [72] for a fair comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 329, + 293, + 426 + ], + "lines": [ + { + "bbox": [ + 57, + 329, + 293, + 426 + ], + "spans": [ + { + "bbox": [ + 57, + 329, + 293, + 426 + ], + "type": "table", + "html": "
MethodTypeFID10KFID30K# Params.
eDiff-I* [2]DM-7.609.10B
SD* [65]DM-8.400.94B
LAFITE [112]GAN-14.8075M
StyleGAN-T [72]GAN-7.30-
GigaGAN [34]GAN9.18-0.65B
Aurora (ours)GAN7.966.200.98B
", + "image_path": "d5e46dd65bc7a32b65e988a1357534b40f1a6ea0b00bca1238c30f445fb30823.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 447, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 447, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 295, + 543 + ], + "type": "text", + "content": "exhibits additional performance improvements. Finally, we conduct experiments by training our model using more GPUs, while keeping all other configurations the same. We observe consistent performance improvements with the use of additional GPUs and have not yet encountered any performance saturation related to the number of GPUs utilized. Our final base generator, trained using 320 GPUs, achieves a zero-shot " + }, + { + "bbox": [ + 54, + 447, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{30K}" + }, + { + "bbox": [ + 54, + 447, + 295, + 543 + ], + "type": "text", + "content": " score of 6.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Comparison with existing methods. We first compare our base generator with existing methods at the resolution of " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " in Tab. 2. The evaluation results demonstrate that Aurora consistently outperforms all other methods across both metrics at this resolution, irrespective of the model type. Tab. 3 presents the comparison results with the existing methods on higher resolution. According to Tab. 3, our method not only achieves the best performance among GANs in terms of the FID but also has the largest model size. However, it lags behind diffusion models and autoregressive models. We conclude that more powerful superresolution techniques or larger models are necessary, and exploring how to train a high-resolution GAN on a larger model remains worthwhile." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 351, + 553, + 512 + ], + "blocks": [ + { + "bbox": [ + 313, + 294, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 350 + ], + "type": "text", + "content": "Table 3. Quantitative comparison between our Aurora and existing text-to-image generators at " + }, + { + "bbox": [ + 313, + 294, + 555, + 350 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 294, + 555, + 350 + ], + "type": "text", + "content": " resolution. We report the model size, total images seen during training, \"zero-shot " + }, + { + "bbox": [ + 313, + 294, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{30K}" + }, + { + "bbox": [ + 313, + 294, + 555, + 350 + ], + "type": "text", + "content": "\", and inference speed. \"\\*\" means that we downsample the generated samples using Lanczos [72] for a fair comparison." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 351, + 553, + 512 + ], + "lines": [ + { + "bbox": [ + 316, + 351, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 351, + 553, + 512 + ], + "type": "table", + "html": "
MethodType# Params.# ImagesFID30KSpeed
DALL-E [60]AR12.0B1.54B27.50-
Parti-20B [108]AR20.0B3.69B7.23-
GLIDE [55]DM5.0B5.94B12.2415.0s
SD* [65]DM0.9B3.16B8.592.9s
Imagen [67]DM3.0B15.36B7.279.1s
eDiff-I[2]DM9.1B11.47B6.9532.0s
LAFITE [112]GAN75M0.03B26.940.02s
StyleGAN-T [72]GAN1.0B0.25B13.900.1s
GALIP [94]GAN0.32B0.012B12.540.04s
GigaGAN* [34]GAN1.0B0.98B9.030.13s
Aurora* (ours)GAN1.32B0.12B8.740.16s
", + "image_path": "8e07d03090364885c1e0b5379b49857caeb9c0ac104dec08aec09c3ae740b0e7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 528, + 425, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 425, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 425, + 540 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 545, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 665 + ], + "type": "text", + "content": "Text-to-image synthesis. We showcase some visual results in Fig. 3 under diverse text conditions. We can tell that, besides performing well on image synthesis within a narrow distribution (e.g., human faces, horses, or bedrooms), GANs are also capable of open-vocabulary T2I generation, allowing users to play with their creativity and imagination. It is noteworthy that our generator excels at generating scenic images (see Figs. 1 and 3), producing pictures that are more realistic than those generated by diffusion models (see supplementary materials for the comparison)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Text interpolation. Our generator produces an image with two independently sampled inputs, namely, a global latent code and a text prompt. Here, we investigate the properties of these two different input spaces through interpolation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18417" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 555, + 208 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 555, + 208 + ], + "type": "image", + "image_path": "974d4041d56d6907ace876cb9bc8337cd82ec8c9684e70fb539788abb3980522.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": "Figure 5. Synthesized results through latent code interpolation, where we fix the text condition, " + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": ", and interpolate the latent codes within the disentangled latent space, " + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 251, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 294, + 346 + ], + "type": "text", + "content": "First, we do the interpolation in prompt space, we present some results in Fig. 4, where we can observe a smooth semantic transition. For instance, when we interpolate between \"a vase filled with lots of pink flowers\" and \"a moongate in the snowy ice\", the shapes of the flowers smoothly merge with the moongate. Such a result suggests that our model could faithfully preserve semantic continuity when learning to decode text conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 348, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 348, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 295, + 468 + ], + "type": "text", + "content": "Latent interpolation. Latent space interpolation is widely studied in the literature [7, 36, 37] to verify that the generator delivers the generative ability instead of memorizing the dataset. We also perform interpolation between global latent codes, with results presented in Fig. 5. Here, we conduct experiments in " + }, + { + "bbox": [ + 54, + 348, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 54, + 348, + 295, + 468 + ], + "type": "text", + "content": " space and observe smooth interpolation as well. For instance, the pose of the man's head transitions smoothly from left to right, indicating that the continuity of the latent space in GANs is maintained even with highly complex datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 192, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 192, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 192, + 491 + ], + "type": "text", + "content": "4.4. Analyses and Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 497, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 640 + ], + "type": "text", + "content": "Sparse router. Recall that our motivation is to enlarge the generator capacity in GANs following the philosophy of sparse MoE [20, 32, 78]. Different from current methods, our sparse router makes its decision considering not only the input feature map but also the conditional text and the sampling stochasticity (Eq. (6)). Here, we visualize some routing paths to verify the effectiveness of our routing mechanism. We can tell from Fig. 6 that, within a synthesized image, pixels with similar visual concepts tend to be processed by the same expert (i.e., having the same color in Fig. 6). Such an observation holds at all resolutions, verifying the rationality of our design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "content": "Discussion. The lack of an open-sourced large-scale GAN model makes it hard for the community to fully study the past favorite generative weapon, leaving some stereotypes, including both negative and positive ones, on GANs after scaling up for T2I generation. For example, instead of the commonly believed training instability, we find that" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 251, + 553, + 355 + ], + "blocks": [ + { + "bbox": [ + 326, + 251, + 553, + 355 + ], + "lines": [ + { + "bbox": [ + 326, + 251, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 326, + 251, + 553, + 355 + ], + "type": "image", + "image_path": "4692986f08633da6107ea238fafada82c557cf6d76813c7695f0542db7ef402e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 357, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 401 + ], + "type": "text", + "content": "Figure 6. Analysis of our learned sparse router, where different colors represent the indices of the experts chosen by the router for each pixel. We can tell that the router tends to cluster the pixels by various visual concepts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 415, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 524 + ], + "type": "text", + "content": "GANs can be stably optimized after introducing text as the condition. Besides, we would like to point out that our model is decently trained on publicly available datasets (see Sec. 4.1) without fine-tuning on any self-collected data for performance improvement. While our release is intended for research purposes only, we have observed that using higher-quality datasets consistently improves performance, as discussed in the diffusion model [6]. We leave such an exploration as the future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 538, + 471, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 471, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 471, + 551 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 559, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 703 + ], + "type": "text", + "content": "In this work, we investigate how to expand the capacity of GANs to match the demand of open-vocabulary text-to-image generation. Drawing lessons from sparse MoE, our proposed Aurora matches the synthesis performance of existing diffusion-based and autoregressive counterparts, both quantitatively (e.g., FID scores) and qualitatively (particularly in scene image generation). Moreover, it offers significantly faster inference and maintains favorable continuous latent space properties. However, our method still struggles when generating complex images, particularly those involving multiple objects, intricate shapes, or fine-grained details. We believe that issues with the discriminator might be contributing to these challenges." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18418" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Int. Conf. Mach. Learn., 2017. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 125, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 125, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 66, + 125, + 295, + 179 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. eDiffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 180, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 180, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 66, + 180, + 295, + 224 + ], + "type": "text", + "content": "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 225, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 225, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 66, + 225, + 295, + 280 + ], + "type": "text", + "content": "[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN dissection: Visualizing and understanding generative adversarial networks. In Int. Conf. Learn. Represent., 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 281, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 281, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 295, + 313 + ], + "type": "text", + "content": "[5] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 315, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 315, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 66, + 315, + 295, + 369 + ], + "type": "text", + "content": "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2023.8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 370, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 370, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 66, + 370, + 295, + 403 + ], + "type": "text", + "content": "[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In Int. Conf. Learn. Represent., 2019. 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 403, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 403, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 66, + 403, + 295, + 446 + ], + "type": "text", + "content": "[8] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 448, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 448, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 66, + 448, + 295, + 491 + ], + "type": "text", + "content": "[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 492, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 557 + ], + "type": "text", + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 295, + 613 + ], + "type": "text", + "content": "[11] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Int. Conf. Mach. Learn., 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 614, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 295, + 669 + ], + "type": "text", + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Int. Conf. Mach. Learn., 2020. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 128 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Adv. Neural Inform. Process. Syst., 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 321, + 129, + 553, + 162 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In Adv. Neural Inform. Process. Syst., 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 163, + 553, + 217 + ], + "type": "text", + "content": "[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In Adv. Neural Inform. Process. Syst., 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 219, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 219, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 219, + 553, + 274 + ], + "type": "text", + "content": "[17] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. GLaM: Efficient scaling of language models with mixture-of-experts. In Int. Conf. Mach. Learn., 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 308 + ], + "type": "text", + "content": "[18] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In Int. Conf. Learn. Represent. Worksh., 2013. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 321, + 309, + 553, + 364 + ], + "type": "text", + "content": "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. Learn., 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 553, + 408 + ], + "type": "text", + "content": "[20] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 2022, 2, 3, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 410, + 553, + 453 + ], + "type": "text", + "content": "[21] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-A-Scene: Scene-based text-to-image generation with human priors. In Eur. Conf. Comput. Vis., 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 455, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 455, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 455, + 553, + 510 + ], + "type": "text", + "content": "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3D: A generative model of high quality 3D textured shapes learned from images. In Adv. Neural Inform. Process. Syst., 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 555 + ], + "type": "text", + "content": "[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Adv. Neural Inform. Process. Syst., 2014. 1, 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 321, + 556, + 553, + 599 + ], + "type": "text", + "content": "[24] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 321, + 601, + 553, + 634 + ], + "type": "text", + "content": "[25] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "type": "text", + "content": "[26] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 321, + 691, + 553, + 714 + ], + "type": "text", + "content": "[27] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18419" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 80, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 294, + 95 + ], + "type": "text", + "content": "by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 2017, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 129 + ], + "type": "text", + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Adv. Neural Inform. Process. Syst., 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 130, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 130, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 295, + 163 + ], + "type": "text", + "content": "[29] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Int. Conf. Comput. Vis., 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 164, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 164, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 164, + 295, + 230 + ], + "type": "text", + "content": "[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "text", + "content": "[31] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 276, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 295, + 308 + ], + "type": "text", + "content": "[32] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 310, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 295, + 342 + ], + "type": "text", + "content": "[33] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In Int. Conf. Learn. Represent., 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 344, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 344, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 295, + 387 + ], + "type": "text", + "content": "[34] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 422 + ], + "type": "text", + "content": "[35] Animesh Karnewar and Oliver Wang. *Msg-gan: Multiscale gradients for generative adversarial networks.* arXiv preprint arXiv:1903.06048, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 423, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 295, + 464 + ], + "type": "text", + "content": "[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Int. Conf. Learn. Represent., 2018. 2, 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 510 + ], + "type": "text", + "content": "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 295, + 555 + ], + "type": "text", + "content": "[38] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 557, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 295, + 600 + ], + "type": "text", + "content": "[39] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Adv. Neural Inform. Process. Syst., 2021. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 602, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 295, + 645 + ], + "type": "text", + "content": "[40] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 647, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 678 + ], + "type": "text", + "content": "[41] Hyunjik Kim, George Papamakarios, and Andriy Mnih. The lipschitz constant of self-attention. In Int. Conf. Mach. Learn., 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 295, + 712 + ], + "type": "text", + "content": "[42] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In Int. Conf. Learn. Represent., 2014. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 139 + ], + "type": "text", + "content": "[43] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 174 + ], + "type": "text", + "content": "[44] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. In Int. Conf. Learn. Represent., 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 176, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 176, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 320, + 176, + 553, + 230 + ], + "type": "text", + "content": "[45] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before Fuse: Vision and language representation learning with momentum distillation. In Adv. Neural Inform. Process. Syst., 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 233, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 233, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 320, + 233, + 553, + 276 + ], + "type": "text", + "content": "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis., 2014. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 278, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 278, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 320, + 278, + 553, + 300 + ], + "type": "text", + "content": "[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 302, + 553, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 302, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 320, + 302, + 553, + 345 + ], + "type": "text", + "content": "[48] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Adv. Neural Inform. Process. Syst., 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 348, + 553, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 348, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 320, + 348, + 553, + 401 + ], + "type": "text", + "content": "[49] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 553, + 447 + ], + "type": "text", + "content": "[50] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 450, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 450, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 320, + 450, + 553, + 483 + ], + "type": "text", + "content": "[51] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for GANs do actually converge? In Int. Conf. Mach. Learn., 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 485, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 485, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 320, + 485, + 553, + 506 + ], + "type": "text", + "content": "[52] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 509, + 553, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 509, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 320, + 509, + 553, + 551 + ], + "type": "text", + "content": "[53] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Int. Conf. Learn. Represent., 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 320, + 554, + 553, + 597 + ], + "type": "text", + "content": "[54] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In Int. Conf. Comput. Vis., 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 600, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 600, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 320, + 600, + 553, + 654 + ], + "type": "text", + "content": "[55] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., 2022. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 656, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 656, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 656, + 553, + 689 + ], + "type": "text", + "content": "[56] David W. Opitz and Richard Maclin. Popular ensemble methods: An empirical study. J. Artif. Intell. Res., 11:169-198, 1999. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "text", + "content": "[57] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18420" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "text", + "content": "StyleSDF: High-resolution 3D-consistent image and geometry generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 108, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 108, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 108, + 295, + 162 + ], + "type": "text", + "content": "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 164, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 164, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 61, + 164, + 295, + 219 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., 2021. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 222, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 222, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 62, + 222, + 294, + 265 + ], + "type": "text", + "content": "[60] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Int. Conf. Mach. Learn., 2021. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 267, + 295, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 267, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 62, + 267, + 295, + 311 + ], + "type": "text", + "content": "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 294, + 356 + ], + "type": "text", + "content": "[62] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Int. Conf. Mach. Learn., 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 359, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 294, + 413 + ], + "type": "text", + "content": "[63] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Adv. Neural Inform. Process. Syst., 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 437 + ], + "type": "text", + "content": "[64] Lior Rokach. Ensemble-based classifiers. Artif. Intell. Rev., 33(1-2):1-39, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 440, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 294, + 483 + ], + "type": "text", + "content": "[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 486, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 486, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 486, + 294, + 529 + ], + "type": "text", + "content": "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 532, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 294, + 597 + ], + "type": "text", + "content": "[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 294, + 632 + ], + "type": "text", + "content": "[68] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Int. Conf. Comput. Vis., 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 667 + ], + "type": "text", + "content": "[69] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[70] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. Adv. Neural Inform. Process. Syst., 2016. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[71] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGAN-XL: Scaling StyleGAN to large diverse datasets. In ACM SIGGRAPH, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 553, + 149 + ], + "type": "text", + "content": "[72] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In Int. Conf. Mach. Learn., 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 151, + 553, + 183 + ], + "type": "text", + "content": "[73] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In Eur. Conf. Comput. Vis., 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 184, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 553, + 249 + ], + "type": "text", + "content": "[74] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 250, + 553, + 293 + ], + "type": "text", + "content": "[75] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Adv. Neural Inform. Process. Syst., 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 320, + 294, + 553, + 337 + ], + "type": "text", + "content": "[76] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3D-aware image synthesis with sparse voxel grids. In Adv. Neural Inform. Process. Syst., 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 553, + 381 + ], + "type": "text", + "content": "[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Association for Computational Linguistics, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "type": "text", + "content": "[78] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Int. Conf. Learn. Represent., 2017. 2, 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 553, + 480 + ], + "type": "text", + "content": "[79] Sheng Shen, Zhewei Yao, Chunyuan Li, Trevor Darrell, Kurt Keutzer, and Yuxiong He. Scaling vision-language models with sparse mixture of experts. arXiv preprint arXiv:2303.07226, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 553, + 514 + ], + "type": "text", + "content": "[80] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 320, + 515, + 553, + 557 + ], + "type": "text", + "content": "[81] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 553, + 602 + ], + "type": "text", + "content": "[82] Yujun Shen, Zhiyi Zhang, Dingdong Yang, Yinghao Xu, Ceyuan Yang, and Jiapeng Zhu. Hammer: An efficient toolkit for training deep models. https://github.com/bytedance/Hammer, 2022.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 603, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 603, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 320, + 603, + 553, + 645 + ], + "type": "text", + "content": "[83] Zifan Shi, Sida Peng, Yinghao Xu, Geiger Andreas, Yiyi Liao, and Yujun Shen. Deep generative models on 3D representations: A survey. arXiv preprint arXiv:2210.15663, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "text", + "content": "[84] Zifan Shi, Yinghao Xu, Yujun Shen, Deli Zhao, Qifeng Chen, and Dit-Yan Yeung. Improving 3D-aware image synthesis with a geometry-aware discriminator. In Adv. Neural Inform. Process. Syst., 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 714 + ], + "type": "text", + "content": "[85] Zifan Shi, Yujun Shen, Yinghao Xu, Sida Peng, Yiyi Liao, Sheng Guo, Qifeng Chen, and Dit-Yan Yeung. Learning" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18421" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "3d-aware image synthesis with unknown pose distribution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 61, + 95, + 295, + 128 + ], + "type": "text", + "content": "[86] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3D GANs. In Adv. Neural Inform. Process. Syst., 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 129, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 294, + 171 + ], + "type": "text", + "content": "[87] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Int. Conf. Mach. Learn., 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 294, + 205 + ], + "type": "text", + "content": "[88] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 206, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 294, + 237 + ], + "type": "text", + "content": "[89] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Int. Conf. Mach. Learn., 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "text", + "content": "[90] Ryohei Suzuki, Masanori Koyama, Takeru Miyato, Taizan Yonetsuji, and Huachun Zhu. Spatially controllable image synthesis with internal representation collaging. arXiv preprint arXiv:1811.10153, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "text", + "content": "[91] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2818-2826, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 328, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 294, + 371 + ], + "type": "text", + "content": "[92] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 372, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 414 + ], + "type": "text", + "content": "[93] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. GALIP: Generative adversarial CLIPs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 416, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 416, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 416, + 294, + 458 + ], + "type": "text", + "content": "[94] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 460, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 503 + ], + "type": "text", + "content": "[95] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 504, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 294, + 547 + ], + "type": "text", + "content": "[96] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "type": "text", + "content": "[97] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In Adv. Neural Inform. Process. Syst., 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 593, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 294, + 624 + ], + "type": "text", + "content": "[98] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Adv. Neural Inform. Process. Syst., 2017. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 669 + ], + "type": "text", + "content": "[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Adv. Neural Inform. Process. Syst., 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 670, + 294, + 713 + ], + "type": "text", + "content": "[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[101] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[102] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[103] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3D-aware image synthesis via learning structural and textural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[104] Yiran Xu, Taesung Park, Richard Zhang, Yang Zhou, Eli Shechtman, Feng Liu, Jia-Bin Huang, and Difan Liu. Videogigagan: Towards detail-rich video super-resolution. arXiv preprint arXiv:2404.12388, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 282 + ], + "type": "text", + "content": "[105] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. Int. J. Comput. Vis., 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 284, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 325 + ], + "type": "text", + "content": "[106] Ceyuan Yang, Qihang Zhang, Yinghao Xu, Jiapeng Zhu, Yujun Shen, and Bo Dai. Learning modulated transformation in GANs. arXiv preprint arXiv:2308.15472, 2023. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[107] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "text", + "content": "[108] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "type": "text", + "content": "[109] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Int. Conf. Comput. Vis., 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "text", + "content": "[110] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In Int. Conf. Mach. Learn., pages 7354–7363, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "text", + "content": "[111] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[112] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiaxiang Gu, Jinhui Xu, and Tong Sun. LAFITE: Towards language-free training for text-to-image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "text", + "content": "[113] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zhengjun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in GANs. In Adv. Neural Inform. Process. Syst., 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "text", + "content": "[114] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in GANs. In Int. Conf. Mach. Learn., 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[115] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18422" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 80, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 295, + 95 + ], + "type": "text", + "content": "natural image manifold. In Eur. Conf. Comput. Vis., 2016. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18423" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_content_list.json b/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..14fd43f7dcabfbdde3451c773ba49d364c7f2404 --- /dev/null +++ b/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_content_list.json @@ -0,0 +1,1613 @@ +[ + { + "type": "text", + "text": "Exploring Temporally-Aware Features for Point Tracking", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inès Hyeonsu Kim1* Seokju Cho1* Jiahui Huang2 Jung Yi1 \nJoon-Young Lee2 Seungryong Kim1 \n1KAIST AI 2Adobe Research", + "bbox": [ + 212, + 178, + 781, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 281, + 326, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point tracking in videos is a fundamental task with applications in robotics, video editing, and more. While many vision tasks benefit from pre-trained feature backbones to improve generalizability, point tracking has primarily relied on simpler backbones trained from scratch on synthetic data, which may limit robustness in real-world scenarios. Additionally, point tracking requires temporal awareness to ensure coherence across frames, but using temporally-aware features is still underexplored. Most current methods often employ a two-stage process: an initial coarse prediction followed by a refinement stage to inject temporal information and correct errors from the coarse stage. These approach, however, is computationally expensive and potentially redundant if the feature backbone itself captures sufficient temporal information.", + "bbox": [ + 88, + 313, + 483, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce Chrono, a feature backbone specifically designed for point tracking with built-in temporal awareness. Leveraging pre-trained representations from self-supervised learner DINOv2 and enhanced with a temporal adapter, Chrono effectively captures long-term temporal context, enabling precise prediction even without the refinement stage. Experimental results demonstrate that Chrono achieves state-of-the-art performance in a refiner-free setting on the TAP-Vid-DAVIS and TAP-Vid-Kinetics datasets, among common feature backbones used in point tracking as well as DINOv2, with exceptional efficiency. Project page: https://cvlab-kaist.github.io/Chrono/", + "bbox": [ + 88, + 539, + 483, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point tracking aims to track any point within a casual video, which has wide-ranging applications such as robotics [48], video editing [23], and view synthesis [50]. It involves establishing correspondences of specific points across frames to determine their positions over time and whether they are visible or occluded. Accurate point tracking is challenging", + "bbox": [ + 89, + 787, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd814143d84d78e9371f27836ae18b44e17cfd22e719dbe0901b359d17ef3451.jpg", + "image_caption": [ + "Figure 1. Chrono is a highly precise, temporally-aware feature backbone specifically designed for point tracking. Unlike prior methods that rely on heavy iterative refinement for each query point [11-13, 19, 25] or test-time optimization [46, 49], Chrono achieves competitive tracking performance through simple feature matching using a non-parametric operation. Moreover, its performance can be further improved by an optional refinement step." + ], + "image_footnote": [], + "bbox": [ + 516, + 279, + 901, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "due to complex motions, occlusions, and deformations [11- 14]. Achieving robust tracking requires effective matching of points across frames and a comprehensive understanding of temporal dynamics.", + "bbox": [ + 511, + 582, + 905, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In many computer vision tasks such as segmentation, object detection, and depth estimation, it is often beneficial to use pre-trained feature backbones along with task-specific heads [4, 6-8, 21]. These backbones, trained on large-scale datasets, offer robust representations that enhance generalizability to real-world data. In contrast, point tracking models often rely on simpler backbones such as ResNet [20] and are frequently trained from scratch [11-13, 19, 25] on synthetic dataset [16, 54]. This difference raises an important question: Could point tracking similarly benefit from adopting pre-trained feature backbones?", + "bbox": [ + 509, + 643, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While these backbones are often pre-trained on single-image tasks [15, 33, 37, 46, 53], point tracking inherently demands temporal awareness to maintain coherence across video frames. Unlike single-image tasks, point tracking requires consistent point matching over time, necessitating a model capable of capturing temporal informa", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 114, + 887, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1962", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/091872ec61229992bb6f64163161730e9f3fbbf1b8d04eda6e1cd377e6d3fca3.jpg", + "image_caption": [ + "Figure 2. Comparison of initial trajectories from Chrono and existing feature backbone for point tracking. Chrono demonstrates a significant improvement over existing feature backbones such as TSM-ResNet [32]." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion [11, 14, 25, 26, 31, 48]. Therefore, features used in point tracking also need to be temporally aware.", + "bbox": [ + 89, + 402, + 482, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its critical role, temporally-aware backbone has been relatively under-explored in this area. Early efforts, such as simply incorporating TSM-ResNet [32] in point tracking models [12, 13], yielded limited improvements. This outcome may be attributed to two factors: First, training on small synthetic datasets from scratch [16] may limit generalization to real-world data. Second, TSM-ResNet's temporal awareness, constrained to adjacent frames, may not provide long-term temporal context for point tracking.", + "bbox": [ + 89, + 431, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, point tracking methods often adopt a two-stage process [11, 13, 14]: an initial stage that predicts coarse tracks directly from the simple feature backbone such as ResNet, followed by a refinement stage that introduces temporal information by analyzing trajectories across frames and iteratively refining the prediction. This pipeline has become standard, with refinement methods aiming to mitigate noise and enhance track smoothness [11, 14, 25, 26, 31, 48]. However, these refinement steps can be computationally demanding and query-dependent [10, 11], often requiring temporal refinement for each query point individually, which impacts efficiency. Moreover, a strong reliance on post-hoc refinement adds a substantial burden in correcting initial prediction errors from the first stage. This error could potentially be mitigated by leveraging features with built-in temporal awareness.", + "bbox": [ + 89, + 568, + 482, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we introduce a temporally-aware feature backbone for point tracking, called Chrono, as illustrated in Figure 1. Our key insights are twofold: First, we leverage robust feature representations learned from large-scale, real-world datasets. We employ DINOv2 [37], a pre-trained model known for", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "strong feature representations across various tasks such as segmentation, classification, and localization [30, 37]. However, these representations are not directly compatible with point tracking due to a lack of temporal awareness. Second, we incorporate temporal awareness by designing a temporal adapter that enables the pre-trained backbone to process data with temporal context, without losing its learned knowledge. This enables the use of these strong feature backbones for point tracking tasks. Unlike the adjacent-frame context used in TSM-ResNet [32], our approach incorporates a temporal context that is six times longer, enabling the capture of complex dynamics and enhancing tracking performance, as shown in Figure 2.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results show that tracks estimated from Chrono features, computed solely using soft argmax without any learnable layers or additional temporal information after feature extraction, significantly outperform traditional feature backbones commonly used in point tracking as well as DINOv2 [37]. Specifically, Chrono achieves a $+20.6\\%$ p increase in the position accuracy on the TAP-ViddAVIS [12] dataset compared to TSM-ResNet-18 [32]. Additionally, our method demonstrates superior efficiency and comparable performance compared to point tracking models equipped with refiners that inject temporal information post-feature extraction. Despite the absence of learnable layers after feature extraction, Chrono achieves high precision, underscoring that directly embedding temporal information within the features is both efficient and powerful. Optionally, we can add iterative refinement on top of our backbone, which can further enhance the performance of the refiner compared to when an existing backbone is used. In summary, our contributions are as follows:", + "bbox": [ + 511, + 287, + 906, + 574 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We highlight the lack of temporally-aware feature backbones and the reliance on computationally intensive refinement processes in current point tracking methods.", + "- We propose a feature backbone designed for point tracking that incorporates a long-range temporal adapter, enhancing temporal awareness over extended sequences.", + "- We demonstrate that our backbone produces accurate initial tracks in simple and effective manner, reducing the need for extensive refinement and achieving both improved performance and efficiency." + ], + "bbox": [ + 511, + 574, + 903, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 739, + 653, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point tracking. PIPs [19] independently tracks a point by fetching a local correlation around the point estimate and gradually refining the tracking result through iteration. TAP-Net [12] utilizes a shallow TSM-ResNet [32] for point tracking, with lightweight layers added to the backbone. TAPIR [13] integrates TAP-Net with the iterative refinement from PIPs, modifying the architecture to a convolution-based model for temporal processing. CoTracker tracks multiple points simultaneously, modeling", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1963", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "their dependencies with a Transformer architecture. LocoTrack [11] achieves improved correspondence using local 4D correlation, inspired by dense matching literature [9]. TAPTR [31] introduces the DETR-like [6] architecture for point tracking. Another line of work employs test-time optimization with regularization, such as tracking smoothness, geometric constraints, and cycle consistency [46, 49]. While these methods focus on designing better architectures for track refinement, the exploration of more effective feature backbones has been relatively underexplored. Our work focuses on developing an improved feature backbone.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature backbone in point tracking. Early works on point tracking [12] explored the performance of self-supervised feature backbones, such as VFS [51], as well as temporally-aware features like TSM-ResNet [32]. However, the exploration of feature backbones has been relatively underdeveloped compared to iterative track refinement [10, 11, 13, 25]. DINO-Tracker [46] explored the use of DINO in point tracking but it requires an hour of optimization for each video. Recently, [2] investigated the tracking capabilities of foundational models, such as Stable Diffusion [41] and DINOv2 [37], though the potential to extend these models for temporal applications remains unexplored.", + "bbox": [ + 89, + 263, + 482, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adapting large feature backbone. With recent advances in large-scale self-supervised training [22, 37], effectively transferring its vast knowledge has become a key challenge. Specifically, the strong features of DINOv2 have been adapted to various tasks. [43] fine-tunes DINO for video tasks in a self-supervised manner. Additionally, DINO has been applied to various correspondence tasks [1, 17, 30, 35, 44] and has demonstrated robustness in establishing correspondence and semantic segmentation [1, 18, 36], showing its rich semantics. We focus on adapting this representation to the point tracking task, enhancing its temporal consistency while preserving its pre-trained knowledge.", + "bbox": [ + 89, + 465, + 482, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 660, + 181, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our temporally aware feature backbone. We begin by formally defining the point tracking problem and its associated challenges. We then describe the design of our temporal adapter and its integration with a pre-trained backbone to incorporate temporal awareness. Finally, we present our approach for adapting the pretrained backbone and training the temporal adapter for point tracking.", + "bbox": [ + 89, + 686, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries and Motivation", + "text_level": 1, + "bbox": [ + 89, + 816, + 352, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task definition. Point tracking [10-14, 25, 31] in videos involves establishing correspondences of specific points across frames, determining their positions over time, and identifying whether they are visible or occluded. Formally,", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "given a video sequence $\\{\\mathcal{I}_t\\}_{t = 0}^{T - 1}$ , where $\\mathcal{I}_t \\in \\mathbb{R}^{H \\times W \\times 3}$ represents the $t$ -th frame of height $H$ and width $W$ , and a query point $q = (x_q, y_q, t_q) \\in \\mathbb{R}^3$ at frame $t_q$ , the goal is to produce a trajectory $\\mathcal{T} = \\{\\hat{p}_t\\}_{t = 0}^{T - 1}$ , where $\\hat{p}_t \\in \\mathbb{R}^2$ denotes the estimated position of the point at time $t$ , and an associated visibility probability $\\mathcal{V} = \\{v_t\\}_{t = 0}^{T - 1}$ , where $v_t \\in [0,1]$ indicates the likelihood of the point being visible at frame $t$ . In this work, we focus on predicting accurate positions rather than occlusions, as our primary goal is to achieve better point correspondences.", + "bbox": [ + 511, + 89, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. Accurate point tracking in videos requires robust point matching across frames, a challenging task due to complex motions, occlusions, deformations, and scale variations [11, 25, 49]. Because videos consist of sequences over time, considering both spatial and temporal aspects is essential to understand point dynamics. More specifically, a good feature for point tracking must satisfy two key criteria. Spatially, it should effectively model complex real-world data with powerful feature representations vital for robust matching, the cornerstone of point tracking [11]. Temporally, it must capture complex motions and understand the dynamics of points over time to handle their movement across frames.", + "bbox": [ + 511, + 251, + 903, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To fulfill the spatial criterion, we utilize DINOv2 [37], a self-supervised model trained on large-scale real-world data, renowned for its robust feature representations across various tasks [5, 30, 37]. However, DINOv2 lacks inherent temporal awareness, which is essential for point tracking in videos. To address this limitation, we introduce a temporal adapter that supplements DINOv2's features with temporal information. This adapter integrates temporal understanding directly into the feature extraction process, creating a feature backbone tailored for point tracking that combines strong spatial representations with temporal dynamics.", + "bbox": [ + 511, + 448, + 905, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Temporal Adapter", + "text_level": 1, + "bbox": [ + 511, + 625, + 692, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal awareness is essential for accurately tracking points across frames, as it enables the model to capture motion patterns and temporal dependencies. To embed this capability within DINOv2, we design a temporal adapter that allows the backbone to incorporate information from adjacent frames, enhancing its temporal sensitivity. Figure 3 illustrates the overall architecture of Chrono.", + "bbox": [ + 511, + 648, + 903, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Design of the temporal adapter. Our temporal adapter is placed between each transformer block of DINOv2 [37] to enhance temporal modeling across multiple feature levels. By computing adjacent features from other frames at once, the adapter connects features from different time steps, enabling it to recognize motion across frames.", + "bbox": [ + 511, + 763, + 903, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each adapter follows a bottleneck structure inspired by the ResNet [20] architecture. It begins with a 2D convolutional layer Conv2D $^{\\text{down}}$ with stride $s$ that reduces the", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1964", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d6a3b9ea0b55e61a19776ae4b3754c0051a6447a21a3533b136c82fa30583377.jpg", + "image_caption": [ + "Figure 3. Overall architecture of Chrono. Temporal adapters between transformer blocks use 2D convolution and 1D local attention along the temporal axis to output temporally-aware features." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spatial dimensions from $H_{p} \\times W_{p} \\times C_{\\mathrm{in}}$ to a compact $H_{p}^{\\prime} \\times W_{p}^{\\prime} \\times C_{\\mathrm{out}}$ , where $H_{p}^{\\prime} = H_{p} / s$ and $W_{p}^{\\prime} = W_{p} / s$ . This spatial downsampling step optimizes the representation for efficient processing while expanding the spatial receptive field for subsequent operations. This is formally defined as:", + "bbox": [ + 89, + 536, + 483, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t} ^ {\\text {d o w n}} = \\operatorname {C o n v 2 D} ^ {\\text {d o w n}} \\left(\\mathbf {f} _ {t} ^ {\\text {i n}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 623, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_t^{\\mathrm{in}}\\in \\mathbb{R}^{H_p\\times W_p\\times C_{\\mathrm{in}}}$ and $\\mathbf{f}_t^{\\mathrm{out}}\\in \\mathbb{R}^{H_p'\\times W_p'\\times C_{\\mathrm{out}}}$ are the input and output feature maps of Conv2D $^{\\mathrm{down}}$ at time $t$ , respectively.", + "bbox": [ + 89, + 655, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After downsampling, we apply a temporal attention layer that captures dependencies within a local temporal window of size $N$ , focusing on temporal correlations across time [3, 42]. The local window attention is carefully calibrated to balance computational efficiency with effective motion capture, ensuring a broad enough temporal window without excessive resource demands. This layer operates over the neighboring frames within the window $[t - k, t + k]$ where $k = \\frac{N - 1}{2}$ . For each spatial location $(x,y)$ , we compute attention weights based on the similarity between the query vector at time $t$ and key vectors from neighboring times. The aggregated feature at time $t$ and location $(x,y)$ , referred to as $\\mathbf{f}_t^{\\mathrm{att}}(x,y)$ , is computed as a weighted sum of", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e87dfa2071da79a9c3801c5383b65643a682d562e60656cacc48098952c367ec.jpg", + "image_caption": [ + "Figure 4. Point track prediction. To predict point positions, we simply match the query points with features from other frames, without using any learnable layers." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the value vectors from these neighboring frames:", + "bbox": [ + 511, + 277, + 838, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t} ^ {\\mathrm {a t t n}} (x, y) = \\sum_ {n = - k} ^ {k} \\alpha^ {(t, n)} (x, y) \\cdot \\mathbf {V} _ {t + n} (x, y), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 301, + 905, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha^{(t,n)}(x,y)$ represents the attention weights for time offset $n$ , computed as:", + "bbox": [ + 511, + 354, + 905, + 386 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha^ {(t, n)} (x, y) = \\frac {\\exp \\left(\\mathbf {Q} _ {t} (x , y) \\cdot \\mathbf {K} _ {t + n} (x , y) / \\sqrt {d _ {k}}\\right)}{\\sum_ {n ^ {\\prime} = - k} ^ {k} \\exp \\left(\\mathbf {Q} _ {t} (x , y) \\cdot \\mathbf {K} _ {t + n ^ {\\prime}} (x , y) / \\sqrt {d _ {k}}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 393, + 913, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{k}$ is a scaling factor [47], $\\mathbf{Q},\\mathbf{K}$ , and $\\mathbf{V}$ are projections of $\\mathbf{f}^{\\mathrm{down}}$ , obtained through the linear projection layers $\\mathbf{W}_Q,\\mathbf{W}_K$ , and $\\mathbf{W}_V$ .", + "bbox": [ + 511, + 446, + 905, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After temporal attention, another 2D convolutional layer restores the spatial dimensions back to $H_{p} \\times W_{p} \\times C_{\\mathrm{in}}$ :", + "bbox": [ + 511, + 491, + 905, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t} ^ {\\mathrm {u p}} = \\operatorname {C o n v 2 D} ^ {\\mathrm {u p}} \\left(\\mathbf {f} _ {t} ^ {\\mathrm {a t t n}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 530, + 903, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_t^{\\mathrm{up}}\\in \\mathbb{R}^{H_p\\times W_p\\times C_{\\mathrm{in}}}$ . Finally, a residual connection adds the original input feature map to the output of the adapter to preserve the feature representation of DINOv2:", + "bbox": [ + 511, + 555, + 905, + 603 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t} ^ {\\text {o u t}} = \\mathbf {f} _ {t} ^ {\\text {u p}} + \\mathbf {f} _ {t} ^ {\\text {i n}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 611, + 903, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Point Prediction with the Feature Backbone", + "text_level": 1, + "bbox": [ + 511, + 638, + 883, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 4, we track a query point by simply performing feature matching using features from Chrono. This approach avoids the need for learnable modules such as iterative refinement, as used in [10, 11, 13, 25].", + "bbox": [ + 511, + 661, + 905, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Correlation construction. Given the query point $q = (x_q, y_q, t_q)$ , we extract the query feature $\\mathbf{f}_q$ from the feature map of frame $t_q$ at position $(x_q, y_q)$ using bilinear interpolation. For each frame $t$ , we compute the correlation map $\\mathcal{C}_t$ by calculating the cosine similarity between $\\mathbf{f}_q$ and the feature map $\\mathbf{f}_t$ at every spatial location:", + "bbox": [ + 511, + 726, + 905, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {t} (x, y) = \\frac {\\mathbf {f} _ {q} ^ {\\top} \\mathbf {f} _ {t} (x , y)}{\\| \\mathbf {f} _ {q} \\| \\| \\mathbf {f} _ {t} (x , y) \\|}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 825, + 903, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_t(x,y)$ is the feature vector at position $(x,y)$ in frame $t$ , $\\mathbf{f}_q^\\top \\mathbf{f}_t(x,y)$ denotes the dot product between $\\mathbf{f}_q$ and", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1965", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{f}_t(x,y)$ , and $\\| \\mathbf{f}_q\\|$ and $\\| \\mathbf{f}_t(x,y)\\|$ indicate the Euclidean norms of $\\mathbf{f}_q$ and $\\mathbf{f}_t(x,y)$ , respectively. This correlation map $\\mathcal{C}_t$ quantifies the similarity between the query feature and features across all spatial locations in each frame $t$ .", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point prediction. To estimate the position of points in frame $t$ , we apply the soft-argmax operation [29] to the correlation map $C_t$ . The soft-argmax computes a weighted average of all spatial positions, with weights given by the softmax of the correlation values:", + "bbox": [ + 89, + 156, + 482, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {p}} _ {t} = \\sum_ {(x, y)} \\sigma \\left(C _ {t} (x, y)\\right) \\cdot (x, y), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 239, + 482, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma(C_t(x, y))$ is the softmax over all spatial positions within frame $t$ :", + "bbox": [ + 89, + 281, + 480, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma \\left(C _ {t} (x, y)\\right) = \\frac {\\exp \\left(C _ {t} (x , y) \\cdot \\tau\\right)}{\\sum_ {\\left(x ^ {\\prime} , y ^ {\\prime}\\right)} \\exp \\left(C _ {t} \\left(x ^ {\\prime} , y ^ {\\prime}\\right) \\cdot \\tau\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 316, + 482, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is a softmax temperature. The soft argmax is nonlearnable and provides a differentiable method to estimate the point's position based on the correlation map. To enhance precision, we mask out positions more than $M$ pixels away from the maximum correlation value, focusing the soft argmax computation on a local neighborhood and reducing the influence of irrelevant regions [29].", + "bbox": [ + 89, + 359, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. To train Chrono, we use the Huber loss [24] to supervise the estimated positions. The Huber loss is chosen for its robustness to outliers compared to the squared error loss. At each time step $t$ , the loss is defined as:", + "bbox": [ + 89, + 470, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {H u b e r}} \\left(\\hat {p} _ {t}, p _ {t}\\right) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\left\\| \\hat {p} _ {t} - p _ {t} \\right\\| ^ {2}, & \\text {i f} \\| \\hat {p} _ {t} - p _ {t} \\| \\leq \\delta , \\\\ \\delta \\cdot \\left(\\left\\| \\hat {p} _ {t} - p _ {t} \\right\\| - \\frac {1}{2} \\delta\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 540, + 493, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_t$ is the ground truth position, and $\\delta$ is a threshold parameter. For occluded points indicated by the ground truth visibility status $v_t$ , we exclude the loss computation, effectively masking out those time steps:", + "bbox": [ + 89, + 594, + 483, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\left(1 - v _ {t}\\right) \\cdot \\mathcal {L} _ {\\text {H u b e r}} \\left(\\hat {p} _ {t}, p _ {t}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 664, + 482, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By minimizing this loss over all time steps and query points, we train the backbone to produce features for predicting accurate tracked positions.", + "bbox": [ + 89, + 689, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 746, + 223, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 771, + 307, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our method using PyTorch [38]. During training, we use AdamW optimizer [34] with a learning rate of $10^{-4}$ , a weight decay of $10^{-4}$ , and a batch size of 1 per GPU. All models are trained for 100,000 iterations on 4 A100 GPUs, employing a cosine learning rate scheduler with warmup. For training, we use the Kubric Panning-MOVi-E dataset [11, 13] and sample 256 query points per", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "batch. The hyperparameters are set as follows: the softmax temperature is $\\tau = 20.0$ , the soft argmax pixel threshold is $M = 5$ and the local temporal window size is $N = 13$ , and the 2D convolution stride in the temporal adapter is $s = 4$ .", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 513, + 159, + 702, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation datasets. To assess our approach, we employ the TAP-Vid benchmark [12], which comprises both real and synthetic video sets. The real videos come with accurate annotation tracks, while the synthetic videos are paired with perfect ground-truth trajectories. The benchmark consists of three datasets: TAP-Vid-RGB-Stacking, TAP-Vid-Kinetics, and TAP-Vid-DAVIS. TAP-Vid-RGB-Stacking [28] contains synthetic videos of a robot gripper stacking objects and includes 50 videos. The TAP-Vid-Kinetics [27] consists of 1,189 YouTube videos with diverse difficulties, including intense motion blur and abrupt scene transitions, while the TAP-Vid-DAVIS dataset [39] provides 30 videos featuring challenges such as substantial scale shifts of dynamic object.", + "bbox": [ + 511, + 183, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We assess the accuracy of the predicted tracks using two metrics: position accuracy at various thresholds and average position accuracy $(< \\delta_{\\mathrm{avg}}^x)$ . Position accuracy is evaluated at five threshold levels: $< \\delta^0$ , $< \\delta^1$ , $< \\delta^2$ , $< \\delta^3$ , and $< \\delta^4$ , corresponding to accuracies within pixel distances of 1, 2, 4, 8, and 16, respectively. Each $< \\delta^x$ score represents the percentage of visible ground-truth points whose predicted positions fall within the specified threshold. The $< \\delta_{\\mathrm{avg}}^x$ score is the average across all thresholds.", + "bbox": [ + 511, + 398, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [12], we evaluate the datasets using two modes: strided query mode and first query mode. In strided query mode, query points are sampled along the groundtruth trajectory at intervals of 5 frames. In first query mode, query points are sampled from the first visible frame.", + "bbox": [ + 511, + 551, + 905, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 513, + 635, + 653, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative comparison. We evaluate our method against the backbones [20, 32] commonly used in point tracking [11-13, 25, 26] and DINOv2 [37] in both the strided query (Table 1) and the first query modes (Table 2). Specifically, we compare with ResNet-18 [20] from TAPIR [13], TSM-ResNet-18 [32] from TAP-Net [12], and the backbone from the pre-trained CoTracker model [25], all of which are trained for point tracking. We report position accuracy to demonstrate the effectiveness of temporal information in our model.", + "bbox": [ + 511, + 659, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our small model, Chrono (ViT-S/14) with DINOv2 (ViTs/14), achieves state-of-the-art performance in position accuracy across most thresholds, outperforming other backbones on $< \\delta_{\\mathrm{avg}}^x$ . Chrono (ViT-B/14) with DINOv2 (ViTB/14) delivers even better results. Across RGB-Stacking, Kinetics and DAVIS datasets, in the strided and first query", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1966", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/758f437a94c7e47de91ed06044205f5c6456706dad8f2b4c6a7dadd597473707.jpg", + "image_caption": [ + "Reference frame $(t = 0)$" + ], + "image_footnote": [], + "bbox": [ + 212, + 89, + 392, + 169 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dbc3950ad79c04f1f1bcb987b020c236d8a0e1114f48e20a0eebb41a7a149f0a.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of complex real-world video tracking. We qualitatively compare the results generated by Chrono with those from other commonly used backbones in point tracking and as well as DINOv2. Our model demonstrates better smoothness and precision than other competitors." + ], + "image_footnote": [], + "bbox": [ + 93, + 181, + 498, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e5e5573f9b7c0868138d27a6a33b152b823c908cb453c9dbb885779a11389c2.jpg", + "image_caption": [ + "Reference frame $(t = 0)$" + ], + "image_footnote": [], + "bbox": [ + 614, + 89, + 794, + 169 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2f90cf1178d19ef93d654e3f9df62fc1c18d26e7fba46695a1284af62f7bd97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 181, + 901, + 435 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/23fd4d8548e9f1e23f8da2676d2a74846b74ed1866419c7708277ee9e1739f12.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneRGB-Stacking-StridedKinetics-StridedDAVIS-Strided
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]36.664.682.790.994.773.910.535.365.781.388.656.39.731.160.678.387.053.3
TSM-ResNet-18 [12, 32]18.155.583.789.892.567.99.1533.264.679.286.554.58.226.853.673.583.949.2
CoTraker [25] Backbone36.960.174.982.587.668.431.031.050.570.674.958.233.752.563.468.372.058.0
DINOv2 (ViT-S/14) [37]3.213.644.977.487.645.34.915.841.973.886.044.57.322.752.980.089.550.4
DINOv2 (ViT-B/14) [37]4.215.246.678.988.146.65.918.645.975.786.946.610.028.659.582.987.454.4
Chrono (ViT-S/14)65.581.788.892.393.784.332.255.273.984.088.766.829.756.476.786.890.968.0
Chrono (ViT-B/14)68.384.090.392.994.586.033.557.275.885.890.268.531.859.278.888.492.270.1
", + "bbox": [ + 93, + 484, + 903, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison on the TAP-Vid datasets [12] with the strided query mode. Best scores are in bold and second best are underlined for each column.", + "bbox": [ + 89, + 599, + 906, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "modes, Chrono (ViT-S/14) and Chrono (ViT-B/14) consistently outperform competitors in PCK accuracy at most thresholds. The high accuracy at $< \\delta^0$ underscores the precision of Chrono features in point tracking, with further gains at higher thresholds. At $< \\delta^4$ , only Chrono backbones achieve over $90\\%$ accuracy on the DAVIS dataset in strided mode. Both Chrono (ViT-S/14) and Chrono (ViT-B/14) reach the highest position accuracy among backbones on average, demonstrating Chrono's ability to effectively model spatial and temporal information in videos.", + "bbox": [ + 89, + 633, + 482, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison. We visualize the estimated tracks from the DAVIS dataset, with a qualitative comparison shown in Figure 5. Unlike other methods, which produce highly jittery and inconsistent tracks over time, Chrono generates temporally smooth and accurate tracks. This jitter in other methods is expected as they lack awareness of neighboring frames. In contrast, Chrono utilizes", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "temporal adapters to maintain frame-to-frame consistency, allowing it to produce smooth tracks even when individual frame estimates are less stable.", + "bbox": [ + 511, + 635, + 906, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Analysis and Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 694, + 769, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to two-stage point tracking models with iterative refiner. As shown in Table 3, Chrono, although solely a feature backbone without a refiner, achieves point estimation precision comparable to full pipelines with refiners, while demonstrating significantly higher efficiency in terms of throughput. Unlike other models, which require refiners to inject temporal information for each query point, Chrono embeds temporal awareness directly in the feature backbone. This design substantially boosts throughput: for example, Chrono (ViT-B/14) achieves $12.5 \\times$ the speed of TAPIR with only a $3.5\\%$ drop in DAVIS accuracy, and even surpasses TAPIR in RGB-Stacking by $11.4\\%$ p. Simi", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1967", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5215449ac9ec1eb0393585d1c47c54b821932882105054b505f6230c835935a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneRGB-Stacking-FirstKinetics-FirstDAVIS-First
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]31.458.377.887.792.469.68.628.856.574.283.350.39.027.354.973.784.149.8
TSM-ResNet-18 [12, 32]19.251.778.786.890.565.47.828.155.271.480.248.67.323.146.766.679.244.6
CoTraker [25] Backbone32.054.870.278.284.063.924.141.955.162.267.050.127.846.458.664.168.553.1
DINOv2 (ViT-S/14) [37]3.011.838.972.284.142.04.213.636.365.979.239.86.019.947.573.784.546.3
DINOv2 (ViT-B/14) [37]3.613.241.174.184.843.45.116.040.067.980.441.98.924.753.877.087.150.3
Chrono (ViT-S/14)58.277.285.489.191.380.224.846.265.877.583.159.524.049.271.282.887.963.0
Chrono (ViT-B/14)61.879.686.990.292.282.126.048.468.279.885.361.626.152.674.584.990.065.6
", + "bbox": [ + 94, + 88, + 903, + 198 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/721b5d0b5a513878f39660154b05f262d0bfc763180b998ce7c96a321ad5c6bf.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on the TAP-Vid datasets [12] with the query-first mode. Best scores are in bold and second best are underlined for each column." + ], + "table_footnote": [], + "table_body": "
MethodRGB-Stacking-SDAVIS-SThroughput (points/sec)# of Refiner Params.Additional Params.
RAFT [45]58.646.323,405.714.2M4.2M
TAP-Net [12]72.853.129,535.985.5K5.5K
TAPIR [13]74.673.62,097.3225.9M25.9M
PIPs [19]51.059.446.4326.0M26.0M
Chrono (ViT-S/14)84.368.034,396.300M16.2M
Chrono (ViT-B/14)86.070.126,139.860M26.0M
", + "bbox": [ + 93, + 241, + 480, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dd1bc3a6e9c89324f3cdc939f99332dbb5419bb3ebc00ccf8d42732dd79751d1.jpg", + "table_caption": [ + "Table 3. Comparison on point tracking pipelines using refiners. Chrono achieves competitive or superior accuracy while offering significantly higher throughput, despite not using a heavy iterative refiner. TAP-Vid-RGB-Stacking and TAP-Vid-DAVIS metrics report $< \\delta_{avg}^{x}$ . Throughput is measured on a 24-frame video using a single NVIDIA RTX 3090 GPU." + ], + "table_footnote": [], + "table_body": "
MethodRGB-StackingKineticsDAVIS
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
Kubric-VFS-Like [16]57.972.691.940.559.080.033.148.579.4
TAP-Net [12]59.972.890.446.660.985.038.453.182.3
PIPs [19]37.351.091.635.354.877.442.059.482.1
RAFT [45]44.058.690.434.552.579.730.046.379.6
TAPIR [13]62.774.691.657.270.187.861.373.688.8
FlowTrack [10]------66.079.887.2
CoTracker [25]------65.979.489.9
LocoTrack [11]77.186.993.259.573.088.567.879.689.9
Chrono + LocoTrack83.291.094.260.773.789.568.280.291.2
", + "bbox": [ + 91, + 421, + 480, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "larly, Chrono (ViT-S/14) delivers $16.4 \\times$ higher throughput. These results demonstrate that Chrono provides a simple yet highly effective solution, offering performance comparable to refiner-based models without requiring iterative processing.", + "bbox": [ + 89, + 619, + 482, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on Chrono as a backbone in existing point tracking pipeline. In Table 4, we present an analysis of the impact of adding a refiner after our model to investigate its ability as a feature backbone within a point tracking pipeline that uses iterative refinement [11, 13, 19, 25]. To assess the ability to predict precise occlusion status, we use occlusion accuracy (OA) and the Average Jaccard (AJ) metric [12], which measures both position and occlusion.", + "bbox": [ + 89, + 703, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For our experiment, we incorporate the iterative refiner from LocoTrack [11]. Since LocoTrack utilizes three levels of hierarchical features from ResNet, we apply a simple convolutional upsampler to our single-resolution feature map. Specifically, using transposed convolution [52], we", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/708a576f115c05b86e1080c2d1789befdc00665858a98fe8d78e1bfd7778692d.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of Chrono, adapted to LocoTrack [11], on the TAP-Vid dataset with the strided query mode. Our model shows a performance boost over LocoTrack on all datasets, with a particularly significant improvement on RGB-Stacking." + ], + "table_footnote": [], + "table_body": "
MethodDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
1D Conv.26.553.174.585.490.165.9
3D Conv.27.554.374.985.590.066.4
1D Attn. (Ours)29.756.476.786.890.968.0
", + "bbox": [ + 517, + 242, + 903, + 329 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f3c36f51df653c47060201359747574ad41813d483d144b9f3a62ca467491749.jpg", + "table_caption": [ + "Table 5. Ablation on temporal aggregation design choice. 1D attention demonstrates the best performance compared to both 1D and 3D convolution layers." + ], + "table_footnote": [], + "table_body": "
Configuration# of \nAdaptersDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
Early Blocks623.147.869.881.486.661.7
Later Blocks626.652.974.385.390.165.8
Alternating Blocks626.953.474.385.390.065.9
All Blocks1129.756.476.786.890.968.0
", + "bbox": [ + 517, + 376, + 903, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation on the number and position of temporal adapter. Rather than placing the temporal adapter into early or later blocks, or alternating blocks in DINOv2, applying it to all layers significantly boosts performance.", + "bbox": [ + 511, + 462, + 906, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generate a $2 \\times$ feature map with a $4 \\times$ channel reduction and then a $4 \\times$ feature map with a further $2 \\times$ channel reduction. We freeze our model and train the convolutional upsampler and the LocoTrack refiner on the Kubric [16] panning MOVi-E dataset [13].", + "bbox": [ + 511, + 532, + 905, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model combined with LocoTrack surpasses the original LocoTrack and also surpasses the performance of recent state-of-the-art point trackers [10, 25]. Our model shows a boost on all datasets, and notably, it shows a huge boost in the RGB-Stacking dataset, achieving a $+6.1$ improvement in the AJ score. This result demonstrates the effectiveness of our approach when adopted into an existing point tracking pipeline.", + "bbox": [ + 511, + 608, + 905, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on temporal aggregation design choice. To model temporal information in our temporal adapter, we tested two approaches for temporal aggregation: convolution and attention. In Table 5, we present an ablation study where only the temporal aggregation layer within Chrono is modified, comparing 1D convolution, 3D convolution, and 1D attention, with all other components kept constant. While 1D convolution and 1D attention aggregate along the temporal axis only, 3D convolution considers both spatial and temporal dimensions. Due to computational constraints, this study is conducted on Chrono (ViT-S/14).", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1968", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af6dc6c3a6c1029debf1616cdddc83394e3d1acd82f9aa64f3cd9afa7bbe18ac.jpg", + "image_caption": [ + "Figure 6. Visualization of features. We visualize the features generated by our model and DINOv2 using PCA. The results show that our model demonstrates improved temporal smoothness and finer-grained feature representation relative to DINOv2." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 903, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results suggest that 1D attention is effective for temporal aggregation, as it allows the model to capture motion dynamics by adapting to correlations between frames. This adaptability may contribute to improved tracking accuracy by enabling the model to weigh information from different frames and handle varied motion patterns and occlusions. In contrast, temporal convolutions, with their fixed weights, may be less effective in capturing these complex temporal relationships.", + "bbox": [ + 89, + 354, + 483, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on the number and position of temporal adapters. In Table 6, we conduct an ablation study on the placement of temporal adapters in Chrono, which consists of 12 transformer blocks. We evaluate three configurations: placing six temporal adapters within the early six blocks, within the later six blocks, and in alternating blocks.", + "bbox": [ + 89, + 500, + 483, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Early transformer blocks often capture local details, while later blocks tend to focus on broader, global patterns and complex relationships [15, 40]. Therefore, it's plausible that adapters in the initial blocks may learn local motion cues, while those in later blocks might capture overall motion patterns. Placing adapters only in the early or later blocks might emphasize either local or global motion, respectively. While the alternating configuration shows some improvement, it may not be optimal. Integrating temporal adapters between each block could potentially allow for a more balanced capture of both local and global motion patterns, making features sensitive to multilevel motions and leading to better performance.", + "bbox": [ + 89, + 590, + 483, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computation and latency comparison. Table 7 compares our inference time and computation with that of DINOv2 [37], measuring the time and GPU memory taken to extract features for 24 frames. While our method exhibits comparable computation and slower inference due to temporal injection into the feature, Chrono achieves a significantly higher score than vanilla DINOv2. This higher accu", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a6cdad2c9931e202a8077c24c4958dc55616c7526ee90de3b3e80d5a59b10ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method<δxavgInference Time ↓ (sec)GPU Memory ↓ (MiB)
DINOv2 (ViT-S/14) [37]39.80.1941,460
DINOv2 (ViT-B/14) [37]41.90.4432,816
Chrono (ViT-S/14)68.00.5751,550
Chrono (ViT-B/14)70.11.1972,998
", + "bbox": [ + 517, + 352, + 905, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Computation and latency comparison with DINOv2. We measure inference time and GPU memory usage for processing 24 frames without point prediction.", + "bbox": [ + 511, + 446, + 906, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "racy compensates for the increased time required for feature extraction.", + "bbox": [ + 511, + 498, + 905, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of feature with PCA. We visually compare the feature of Chrono and DINOv2 [37] with PCA (Principal Component Analysis) in Figure 6. We reduce the feature dimensions to three, represented as RGB in the figure.", + "bbox": [ + 511, + 539, + 905, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While DINOv2's visualizations demonstrate a noisy background and tend to show a uniform representation within a single object, our model produces consistent and smooth representations over time and also shows fine granularity within the same semantic object. We believe the temporally smooth representation and fine-grained feature detail lead to better point tracking.", + "bbox": [ + 511, + 599, + 905, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented Chrono, a temporally-aware feature backbone for point tracking that integrates pretrained DINOv2 representations with a temporal adapter, enabling long-term temporal context capture in feature space. Through extensive experiments, we demonstrate that embedding temporal information directly into the feature backbone, Chrono reduces reliance on costly refiners while achieving both accuracy and efficiency. We anticipate Chrono will inspire advancements in efficient, temporally-aware backbones for point tracking.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1969", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 260, + 187, + 276 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 3", + "[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 3", + "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 4", + "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1", + "[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 3", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3", + "[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017.", + "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 1", + "[9] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. Advances in Neural Information Processing Systems, 34:9011-9023, 2021. 3", + "[10] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268-19277, 2024. 2, 3, 4, 7" + ], + "bbox": [ + 93, + 285, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Acknowledgements This research was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, RS-2024-00509279, RS-2025-II212068, RS-2023-00227592) and the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (RS-2024-00345025, RS-2024-00333068), and National Research Foundation of Korea (RS-2024-00346597).", + "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 1, 2, 3, 4, 5, 7", + "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 3, 5, 6, 7", + "[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 1, 2, 3, 4, 5, 6, 7", + "[14] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 1, 2, 3", + "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 8", + "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 1, 2, 7", + "[17] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 3", + "[18] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3", + "[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 7", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 3, 5, 6, 7", + "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3", + "[23] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 1" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1970", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5", + "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2, 3, 4, 5, 6, 7", + "[26] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 5", + "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5", + "[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In 5th Annual Conference on Robot Learning, 2021. 5", + "[29] Junghyup Lee, Dohyung Kim, Jean Ponce, and Bumsub Ham. Sfnet: Learning object-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2278-2287, 2019. 5", + "[30] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3", + "[31] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2, 3", + "[32] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 2, 3, 5, 6, 7", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1", + "[34] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 3", + "[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5, 6, 7, 8", + "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5", + "[40] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 8", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[42] Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers. Transactions of the Association for Computational Linguistics, 9:53-68, 2021. 4", + "[43] Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Time does tell: Self-supervised timetuning of dense image representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16536-16547, 2023. 3", + "[44] Aleksandar Shtedritski, Andrea Vedaldi, and Christian Rupprecht. Learning universal semantic correspondences with no supervision and automatic data curation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 933-943, 2023. 3", + "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7", + "[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision, pages 367-385. Springer, 2025. 1, 3", + "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4", + "[48] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. arXiv preprint arXiv:2308.15975, 2023. 1, 2", + "[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely." + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1971", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 1, 3", + "[50] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 1", + "[51] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021. 3", + "[52] Matthew D Zeiler, Dilip Krishnan, Graham W Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on computer vision and pattern recognition, pages 2528-2535. IEEE, 2010. 7", + "[53] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 1", + "[54] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 1" + ], + "bbox": [ + 91, + 90, + 483, + 445 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1972", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_model.json b/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05282f4c12bcac7c077ac454bce356c5a7cd483b --- /dev/null +++ b/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Exploring Temporally-Aware Features for Point Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.782, + 0.248 + ], + "angle": 0, + "content": "Inès Hyeonsu Kim1* Seokju Cho1* Jiahui Huang2 Jung Yi1 \nJoon-Young Lee2 Seungryong Kim1 \n1KAIST AI 2Adobe Research" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.282, + 0.327, + 0.298 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.314, + 0.484, + 0.54 + ], + "angle": 0, + "content": "Point tracking in videos is a fundamental task with applications in robotics, video editing, and more. While many vision tasks benefit from pre-trained feature backbones to improve generalizability, point tracking has primarily relied on simpler backbones trained from scratch on synthetic data, which may limit robustness in real-world scenarios. Additionally, point tracking requires temporal awareness to ensure coherence across frames, but using temporally-aware features is still underexplored. Most current methods often employ a two-stage process: an initial coarse prediction followed by a refinement stage to inject temporal information and correct errors from the coarse stage. These approach, however, is computationally expensive and potentially redundant if the feature backbone itself captures sufficient temporal information." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.484, + 0.735 + ], + "angle": 0, + "content": "In this work, we introduce Chrono, a feature backbone specifically designed for point tracking with built-in temporal awareness. Leveraging pre-trained representations from self-supervised learner DINOv2 and enhanced with a temporal adapter, Chrono effectively captures long-term temporal context, enabling precise prediction even without the refinement stage. Experimental results demonstrate that Chrono achieves state-of-the-art performance in a refiner-free setting on the TAP-Vid-DAVIS and TAP-Vid-Kinetics datasets, among common feature backbones used in point tracking as well as DINOv2, with exceptional efficiency. Project page: https://cvlab-kaist.github.io/Chrono/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Point tracking aims to track any point within a casual video, which has wide-ranging applications such as robotics [48], video editing [23], and view synthesis [50]. It involves establishing correspondences of specific points across frames to determine their positions over time and whether they are visible or occluded. Accurate point tracking is challenging" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.28, + 0.903, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Chrono is a highly precise, temporally-aware feature backbone specifically designed for point tracking. Unlike prior methods that rely on heavy iterative refinement for each query point [11-13, 19, 25] or test-time optimization [46, 49], Chrono achieves competitive tracking performance through simple feature matching using a non-parametric operation. Moreover, its performance can be further improved by an optional refinement step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.643 + ], + "angle": 0, + "content": "due to complex motions, occlusions, and deformations [11- 14]. Achieving robust tracking requires effective matching of points across frames and a comprehensive understanding of temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.809 + ], + "angle": 0, + "content": "In many computer vision tasks such as segmentation, object detection, and depth estimation, it is often beneficial to use pre-trained feature backbones along with task-specific heads [4, 6-8, 21]. These backbones, trained on large-scale datasets, offer robust representations that enhance generalizability to real-world data. In contrast, point tracking models often rely on simpler backbones such as ResNet [20] and are frequently trained from scratch [11-13, 19, 25] on synthetic dataset [16, 54]. This difference raises an important question: Could point tracking similarly benefit from adopting pre-trained feature backbones?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While these backbones are often pre-trained on single-image tasks [15, 33, 37, 46, 53], point tracking inherently demands temporal awareness to maintain coherence across video frames. Unlike single-image tasks, point tracking requires consistent point matching over time, necessitating a model capable of capturing temporal informa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1962" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.485, + 0.389 + ], + "angle": 0, + "content": "Figure 2. Comparison of initial trajectories from Chrono and existing feature backbone for point tracking. Chrono demonstrates a significant improvement over existing feature backbones such as TSM-ResNet [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.432 + ], + "angle": 0, + "content": "tion [11, 14, 25, 26, 31, 48]. Therefore, features used in point tracking also need to be temporally aware." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Despite its critical role, temporally-aware backbone has been relatively under-explored in this area. Early efforts, such as simply incorporating TSM-ResNet [32] in point tracking models [12, 13], yielded limited improvements. This outcome may be attributed to two factors: First, training on small synthetic datasets from scratch [16] may limit generalization to real-world data. Second, TSM-ResNet's temporal awareness, constrained to adjacent frames, may not provide long-term temporal context for point tracking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.81 + ], + "angle": 0, + "content": "On the other hand, point tracking methods often adopt a two-stage process [11, 13, 14]: an initial stage that predicts coarse tracks directly from the simple feature backbone such as ResNet, followed by a refinement stage that introduces temporal information by analyzing trajectories across frames and iteratively refining the prediction. This pipeline has become standard, with refinement methods aiming to mitigate noise and enhance track smoothness [11, 14, 25, 26, 31, 48]. However, these refinement steps can be computationally demanding and query-dependent [10, 11], often requiring temporal refinement for each query point individually, which impacts efficiency. Moreover, a strong reliance on post-hoc refinement adds a substantial burden in correcting initial prediction errors from the first stage. This error could potentially be mitigated by leveraging features with built-in temporal awareness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To overcome these limitations, we introduce a temporally-aware feature backbone for point tracking, called Chrono, as illustrated in Figure 1. Our key insights are twofold: First, we leverage robust feature representations learned from large-scale, real-world datasets. We employ DINOv2 [37], a pre-trained model known for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "strong feature representations across various tasks such as segmentation, classification, and localization [30, 37]. However, these representations are not directly compatible with point tracking due to a lack of temporal awareness. Second, we incorporate temporal awareness by designing a temporal adapter that enables the pre-trained backbone to process data with temporal context, without losing its learned knowledge. This enables the use of these strong feature backbones for point tracking tasks. Unlike the adjacent-frame context used in TSM-ResNet [32], our approach incorporates a temporal context that is six times longer, enabling the capture of complex dynamics and enhancing tracking performance, as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Experimental results show that tracks estimated from Chrono features, computed solely using soft argmax without any learnable layers or additional temporal information after feature extraction, significantly outperform traditional feature backbones commonly used in point tracking as well as DINOv2 [37]. Specifically, Chrono achieves a \\(+20.6\\%\\) p increase in the position accuracy on the TAP-ViddAVIS [12] dataset compared to TSM-ResNet-18 [32]. Additionally, our method demonstrates superior efficiency and comparable performance compared to point tracking models equipped with refiners that inject temporal information post-feature extraction. Despite the absence of learnable layers after feature extraction, Chrono achieves high precision, underscoring that directly embedding temporal information within the features is both efficient and powerful. Optionally, we can add iterative refinement on top of our backbone, which can further enhance the performance of the refiner compared to when an existing backbone is used. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.575, + 0.905, + 0.619 + ], + "angle": 0, + "content": "- We highlight the lack of temporally-aware feature backbones and the reliance on computationally intensive refinement processes in current point tracking methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.62, + 0.905, + 0.664 + ], + "angle": 0, + "content": "- We propose a feature backbone designed for point tracking that incorporates a long-range temporal adapter, enhancing temporal awareness over extended sequences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.665, + 0.905, + 0.726 + ], + "angle": 0, + "content": "- We demonstrate that our backbone produces accurate initial tracks in simple and effective manner, reducing the need for extensive refinement and achieving both improved performance and efficiency." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.575, + 0.905, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.655, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Point tracking. PIPs [19] independently tracks a point by fetching a local correlation around the point estimate and gradually refining the tracking result through iteration. TAP-Net [12] utilizes a shallow TSM-ResNet [32] for point tracking, with lightweight layers added to the backbone. TAPIR [13] integrates TAP-Net with the iterative refinement from PIPs, modifying the architecture to a convolution-based model for temporal processing. CoTracker tracks multiple points simultaneously, modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1963" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "their dependencies with a Transformer architecture. LocoTrack [11] achieves improved correspondence using local 4D correlation, inspired by dense matching literature [9]. TAPTR [31] introduces the DETR-like [6] architecture for point tracking. Another line of work employs test-time optimization with regularization, such as tracking smoothness, geometric constraints, and cycle consistency [46, 49]. While these methods focus on designing better architectures for track refinement, the exploration of more effective feature backbones has been relatively underexplored. Our work focuses on developing an improved feature backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.483, + 0.461 + ], + "angle": 0, + "content": "Feature backbone in point tracking. Early works on point tracking [12] explored the performance of self-supervised feature backbones, such as VFS [51], as well as temporally-aware features like TSM-ResNet [32]. However, the exploration of feature backbones has been relatively underdeveloped compared to iterative track refinement [10, 11, 13, 25]. DINO-Tracker [46] explored the use of DINO in point tracking but it requires an hour of optimization for each video. Recently, [2] investigated the tracking capabilities of foundational models, such as Stable Diffusion [41] and DINOv2 [37], though the potential to extend these models for temporal applications remains unexplored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.648 + ], + "angle": 0, + "content": "Adapting large feature backbone. With recent advances in large-scale self-supervised training [22, 37], effectively transferring its vast knowledge has become a key challenge. Specifically, the strong features of DINOv2 have been adapted to various tasks. [43] fine-tunes DINO for video tasks in a self-supervised manner. Additionally, DINO has been applied to various correspondence tasks [1, 17, 30, 35, 44] and has demonstrated robustness in establishing correspondence and semantic segmentation [1, 18, 36], showing its rich semantics. We focus on adapting this representation to the point tracking task, enhancing its temporal consistency while preserving its pre-trained knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.182, + 0.677 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.809 + ], + "angle": 0, + "content": "In this section, we introduce our temporally aware feature backbone. We begin by formally defining the point tracking problem and its associated challenges. We then describe the design of our temporal adapter and its integration with a pre-trained backbone to incorporate temporal awareness. Finally, we present our approach for adapting the pretrained backbone and training the temporal adapter for point tracking." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.353, + 0.832 + ], + "angle": 0, + "content": "3.1. Preliminaries and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Task definition. Point tracking [10-14, 25, 31] in videos involves establishing correspondences of specific points across frames, determining their positions over time, and identifying whether they are visible or occluded. Formally," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.905, + 0.243 + ], + "angle": 0, + "content": "given a video sequence \\(\\{\\mathcal{I}_t\\}_{t = 0}^{T - 1}\\), where \\(\\mathcal{I}_t \\in \\mathbb{R}^{H \\times W \\times 3}\\) represents the \\(t\\)-th frame of height \\(H\\) and width \\(W\\), and a query point \\(q = (x_q, y_q, t_q) \\in \\mathbb{R}^3\\) at frame \\(t_q\\), the goal is to produce a trajectory \\(\\mathcal{T} = \\{\\hat{p}_t\\}_{t = 0}^{T - 1}\\), where \\(\\hat{p}_t \\in \\mathbb{R}^2\\) denotes the estimated position of the point at time \\(t\\), and an associated visibility probability \\(\\mathcal{V} = \\{v_t\\}_{t = 0}^{T - 1}\\), where \\(v_t \\in [0,1]\\) indicates the likelihood of the point being visible at frame \\(t\\). In this work, we focus on predicting accurate positions rather than occlusions, as our primary goal is to achieve better point correspondences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Motivation. Accurate point tracking in videos requires robust point matching across frames, a challenging task due to complex motions, occlusions, deformations, and scale variations [11, 25, 49]. Because videos consist of sequences over time, considering both spatial and temporal aspects is essential to understand point dynamics. More specifically, a good feature for point tracking must satisfy two key criteria. Spatially, it should effectively model complex real-world data with powerful feature representations vital for robust matching, the cornerstone of point tracking [11]. Temporally, it must capture complex motions and understand the dynamics of points over time to handle their movement across frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.616 + ], + "angle": 0, + "content": "To fulfill the spatial criterion, we utilize DINOv2 [37], a self-supervised model trained on large-scale real-world data, renowned for its robust feature representations across various tasks [5, 30, 37]. However, DINOv2 lacks inherent temporal awareness, which is essential for point tracking in videos. To address this limitation, we introduce a temporal adapter that supplements DINOv2's features with temporal information. This adapter integrates temporal understanding directly into the feature extraction process, creating a feature backbone tailored for point tracking that combines strong spatial representations with temporal dynamics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.693, + 0.642 + ], + "angle": 0, + "content": "3.2. Temporal Adapter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.905, + 0.755 + ], + "angle": 0, + "content": "Temporal awareness is essential for accurately tracking points across frames, as it enables the model to capture motion patterns and temporal dependencies. To embed this capability within DINOv2, we design a temporal adapter that allows the backbone to incorporate information from adjacent frames, enhancing its temporal sensitivity. Figure 3 illustrates the overall architecture of Chrono." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.854 + ], + "angle": 0, + "content": "Design of the temporal adapter. Our temporal adapter is placed between each transformer block of DINOv2 [37] to enhance temporal modeling across multiple feature levels. By computing adjacent features from other frames at once, the adapter connects features from different time steps, enabling it to recognize motion across frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Each adapter follows a bottleneck structure inspired by the ResNet [20] architecture. It begins with a 2D convolutional layer Conv2D\\(^{\\text{down}}\\) with stride \\(s\\) that reduces the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1964" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.484, + 0.483, + 0.527 + ], + "angle": 0, + "content": "Figure 3. Overall architecture of Chrono. Temporal adapters between transformer blocks use 2D convolution and 1D local attention along the temporal axis to output temporally-aware features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.613 + ], + "angle": 0, + "content": "spatial dimensions from \\( H_{p} \\times W_{p} \\times C_{\\mathrm{in}} \\) to a compact \\( H_{p}^{\\prime} \\times W_{p}^{\\prime} \\times C_{\\mathrm{out}} \\), where \\( H_{p}^{\\prime} = H_{p} / s \\) and \\( W_{p}^{\\prime} = W_{p} / s \\). This spatial downsampling step optimizes the representation for efficient processing while expanding the spatial receptive field for subsequent operations. This is formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.624, + 0.483, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t} ^ {\\text {d o w n}} = \\operatorname {C o n v 2 D} ^ {\\text {d o w n}} \\left(\\mathbf {f} _ {t} ^ {\\text {i n}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.703 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_t^{\\mathrm{in}}\\in \\mathbb{R}^{H_p\\times W_p\\times C_{\\mathrm{in}}}\\) and \\(\\mathbf{f}_t^{\\mathrm{out}}\\in \\mathbb{R}^{H_p'\\times W_p'\\times C_{\\mathrm{out}}}\\) are the input and output feature maps of Conv2D\\(^{\\mathrm{down}}\\) at time \\(t\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After downsampling, we apply a temporal attention layer that captures dependencies within a local temporal window of size \\(N\\), focusing on temporal correlations across time [3, 42]. The local window attention is carefully calibrated to balance computational efficiency with effective motion capture, ensuring a broad enough temporal window without excessive resource demands. This layer operates over the neighboring frames within the window \\([t - k, t + k]\\) where \\(k = \\frac{N - 1}{2}\\). For each spatial location \\((x,y)\\), we compute attention weights based on the similarity between the query vector at time \\(t\\) and key vectors from neighboring times. The aggregated feature at time \\(t\\) and location \\((x,y)\\), referred to as \\(\\mathbf{f}_t^{\\mathrm{att}}(x,y)\\), is computed as a weighted sum of" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.225, + 0.907, + 0.267 + ], + "angle": 0, + "content": "Figure 4. Point track prediction. To predict point positions, we simply match the query points with features from other frames, without using any learnable layers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.279, + 0.839, + 0.294 + ], + "angle": 0, + "content": "the value vectors from these neighboring frames:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.303, + 0.906, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t} ^ {\\mathrm {a t t n}} (x, y) = \\sum_ {n = - k} ^ {k} \\alpha^ {(t, n)} (x, y) \\cdot \\mathbf {V} _ {t + n} (x, y), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.906, + 0.387 + ], + "angle": 0, + "content": "where \\(\\alpha^{(t,n)}(x,y)\\) represents the attention weights for time offset \\(n\\), computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.395, + 0.915, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\alpha^ {(t, n)} (x, y) = \\frac {\\exp \\left(\\mathbf {Q} _ {t} (x , y) \\cdot \\mathbf {K} _ {t + n} (x , y) / \\sqrt {d _ {k}}\\right)}{\\sum_ {n ^ {\\prime} = - k} ^ {k} \\exp \\left(\\mathbf {Q} _ {t} (x , y) \\cdot \\mathbf {K} _ {t + n ^ {\\prime}} (x , y) / \\sqrt {d _ {k}}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.906, + 0.492 + ], + "angle": 0, + "content": "where \\(d_{k}\\) is a scaling factor [47], \\(\\mathbf{Q},\\mathbf{K}\\), and \\(\\mathbf{V}\\) are projections of \\(\\mathbf{f}^{\\mathrm{down}}\\), obtained through the linear projection layers \\(\\mathbf{W}_Q,\\mathbf{W}_K\\), and \\(\\mathbf{W}_V\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.492, + 0.906, + 0.525 + ], + "angle": 0, + "content": "After temporal attention, another 2D convolutional layer restores the spatial dimensions back to \\( H_{p} \\times W_{p} \\times C_{\\mathrm{in}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.531, + 0.905, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t} ^ {\\mathrm {u p}} = \\operatorname {C o n v 2 D} ^ {\\mathrm {u p}} \\left(\\mathbf {f} _ {t} ^ {\\mathrm {a t t n}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.604 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_t^{\\mathrm{up}}\\in \\mathbb{R}^{H_p\\times W_p\\times C_{\\mathrm{in}}}\\). Finally, a residual connection adds the original input feature map to the output of the adapter to preserve the feature representation of DINOv2:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.612, + 0.905, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t} ^ {\\text {o u t}} = \\mathbf {f} _ {t} ^ {\\text {u p}} + \\mathbf {f} _ {t} ^ {\\text {i n}}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.885, + 0.654 + ], + "angle": 0, + "content": "3.3. Point Prediction with the Feature Backbone" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.662, + 0.906, + 0.723 + ], + "angle": 0, + "content": "As shown in Figure 4, we track a query point by simply performing feature matching using features from Chrono. This approach avoids the need for learnable modules such as iterative refinement, as used in [10, 11, 13, 25]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.727, + 0.906, + 0.819 + ], + "angle": 0, + "content": "Correlation construction. Given the query point \\( q = (x_q, y_q, t_q) \\), we extract the query feature \\( \\mathbf{f}_q \\) from the feature map of frame \\( t_q \\) at position \\( (x_q, y_q) \\) using bilinear interpolation. For each frame \\( t \\), we compute the correlation map \\( \\mathcal{C}_t \\) by calculating the cosine similarity between \\( \\mathbf{f}_q \\) and the feature map \\( \\mathbf{f}_t \\) at every spatial location:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.826, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {t} (x, y) = \\frac {\\mathbf {f} _ {q} ^ {\\top} \\mathbf {f} _ {t} (x , y)}{\\| \\mathbf {f} _ {q} \\| \\| \\mathbf {f} _ {t} (x , y) \\|}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.904 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_t(x,y)\\) is the feature vector at position \\((x,y)\\) in frame \\(t\\), \\(\\mathbf{f}_q^\\top \\mathbf{f}_t(x,y)\\) denotes the dot product between \\(\\mathbf{f}_q\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1965" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.152 + ], + "angle": 0, + "content": "\\(\\mathbf{f}_t(x,y)\\), and \\(\\| \\mathbf{f}_q\\|\\) and \\(\\| \\mathbf{f}_t(x,y)\\|\\) indicate the Euclidean norms of \\(\\mathbf{f}_q\\) and \\(\\mathbf{f}_t(x,y)\\), respectively. This correlation map \\(\\mathcal{C}_t\\) quantifies the similarity between the query feature and features across all spatial locations in each frame \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.157, + 0.483, + 0.232 + ], + "angle": 0, + "content": "Point prediction. To estimate the position of points in frame \\( t \\), we apply the soft-argmax operation [29] to the correlation map \\( C_t \\). The soft-argmax computes a weighted average of all spatial positions, with weights given by the softmax of the correlation values:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.24, + 0.483, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {p}} _ {t} = \\sum_ {(x, y)} \\sigma \\left(C _ {t} (x, y)\\right) \\cdot (x, y), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.482, + 0.312 + ], + "angle": 0, + "content": "where \\(\\sigma(C_t(x, y))\\) is the softmax over all spatial positions within frame \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.318, + 0.483, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\sigma \\left(C _ {t} (x, y)\\right) = \\frac {\\exp \\left(C _ {t} (x , y) \\cdot \\tau\\right)}{\\sum_ {\\left(x ^ {\\prime} , y ^ {\\prime}\\right)} \\exp \\left(C _ {t} \\left(x ^ {\\prime} , y ^ {\\prime}\\right) \\cdot \\tau\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.468 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a softmax temperature. The soft argmax is nonlearnable and provides a differentiable method to estimate the point's position based on the correlation map. To enhance precision, we mask out positions more than \\(M\\) pixels away from the maximum correlation value, focusing the soft argmax computation on a local neighborhood and reducing the influence of irrelevant regions [29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.533 + ], + "angle": 0, + "content": "Training. To train Chrono, we use the Huber loss [24] to supervise the estimated positions. The Huber loss is chosen for its robustness to outliers compared to the squared error loss. At each time step \\( t \\), the loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.541, + 0.495, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {H u b e r}} \\left(\\hat {p} _ {t}, p _ {t}\\right) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\left\\| \\hat {p} _ {t} - p _ {t} \\right\\| ^ {2}, & \\text {i f} \\| \\hat {p} _ {t} - p _ {t} \\| \\leq \\delta , \\\\ \\delta \\cdot \\left(\\left\\| \\hat {p} _ {t} - p _ {t} \\right\\| - \\frac {1}{2} \\delta\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.655 + ], + "angle": 0, + "content": "where \\( p_t \\) is the ground truth position, and \\( \\delta \\) is a threshold parameter. For occluded points indicated by the ground truth visibility status \\( v_t \\), we exclude the loss computation, effectively masking out those time steps:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.665, + 0.483, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\left(1 - v _ {t}\\right) \\cdot \\mathcal {L} _ {\\text {H u b e r}} \\left(\\hat {p} _ {t}, p _ {t}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.735 + ], + "angle": 0, + "content": "By minimizing this loss over all time steps and query points, we train the backbone to produce features for predicting accurate tracked positions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.224, + 0.764 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.308, + 0.788 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We implement our method using PyTorch [38]. During training, we use AdamW optimizer [34] with a learning rate of \\(10^{-4}\\), a weight decay of \\(10^{-4}\\), and a batch size of 1 per GPU. All models are trained for 100,000 iterations on 4 A100 GPUs, employing a cosine learning rate scheduler with warmup. For training, we use the Kubric Panning-MOVi-E dataset [11, 13] and sample 256 query points per" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "batch. The hyperparameters are set as follows: the softmax temperature is \\(\\tau = 20.0\\), the soft argmax pixel threshold is \\(M = 5\\) and the local temporal window size is \\(N = 13\\), and the 2D convolution stride in the temporal adapter is \\(s = 4\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.16, + 0.704, + 0.175 + ], + "angle": 0, + "content": "4.2. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Evaluation datasets. To assess our approach, we employ the TAP-Vid benchmark [12], which comprises both real and synthetic video sets. The real videos come with accurate annotation tracks, while the synthetic videos are paired with perfect ground-truth trajectories. The benchmark consists of three datasets: TAP-Vid-RGB-Stacking, TAP-Vid-Kinetics, and TAP-Vid-DAVIS. TAP-Vid-RGB-Stacking [28] contains synthetic videos of a robot gripper stacking objects and includes 50 videos. The TAP-Vid-Kinetics [27] consists of 1,189 YouTube videos with diverse difficulties, including intense motion blur and abrupt scene transitions, while the TAP-Vid-DAVIS dataset [39] provides 30 videos featuring challenges such as substantial scale shifts of dynamic object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Evaluation metrics. We assess the accuracy of the predicted tracks using two metrics: position accuracy at various thresholds and average position accuracy \\((< \\delta_{\\mathrm{avg}}^x)\\). Position accuracy is evaluated at five threshold levels: \\(< \\delta^0\\), \\(< \\delta^1\\), \\(< \\delta^2\\), \\(< \\delta^3\\), and \\(< \\delta^4\\), corresponding to accuracies within pixel distances of 1, 2, 4, 8, and 16, respectively. Each \\(< \\delta^x\\) score represents the percentage of visible ground-truth points whose predicted positions fall within the specified threshold. The \\(< \\delta_{\\mathrm{avg}}^x\\) score is the average across all thresholds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Following [12], we evaluate the datasets using two modes: strided query mode and first query mode. In strided query mode, query points are sampled along the groundtruth trajectory at intervals of 5 frames. In first query mode, query points are sampled from the first visible frame." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.636, + 0.654, + 0.651 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Quantitative comparison. We evaluate our method against the backbones [20, 32] commonly used in point tracking [11-13, 25, 26] and DINOv2 [37] in both the strided query (Table 1) and the first query modes (Table 2). Specifically, we compare with ResNet-18 [20] from TAPIR [13], TSM-ResNet-18 [32] from TAP-Net [12], and the backbone from the pre-trained CoTracker model [25], all of which are trained for point tracking. We report position accuracy to demonstrate the effectiveness of temporal information in our model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Our small model, Chrono (ViT-S/14) with DINOv2 (ViTs/14), achieves state-of-the-art performance in position accuracy across most thresholds, outperforming other backbones on \\(< \\delta_{\\mathrm{avg}}^x\\). Chrono (ViT-B/14) with DINOv2 (ViTB/14) delivers even better results. Across RGB-Stacking, Kinetics and DAVIS datasets, in the strided and first query" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1966" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.214, + 0.09, + 0.393, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.17, + 0.353, + 0.179 + ], + "angle": 0, + "content": "Reference frame \\((t = 0)\\)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.183, + 0.499, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.09, + 0.795, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.169, + 0.756, + 0.178 + ], + "angle": 0, + "content": "Reference frame \\((t = 0)\\)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.183, + 0.902, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.442, + 0.907, + 0.484 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of complex real-world video tracking. We qualitatively compare the results generated by Chrono with those from other commonly used backbones in point tracking and as well as DINOv2. Our model demonstrates better smoothness and precision than other competitors." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.486, + 0.905, + 0.595 + ], + "angle": 0, + "content": "
BackboneRGB-Stacking-StridedKinetics-StridedDAVIS-Strided
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]36.664.682.790.994.773.910.535.365.781.388.656.39.731.160.678.387.053.3
TSM-ResNet-18 [12, 32]18.155.583.789.892.567.99.1533.264.679.286.554.58.226.853.673.583.949.2
CoTraker [25] Backbone36.960.174.982.587.668.431.031.050.570.674.958.233.752.563.468.372.058.0
DINOv2 (ViT-S/14) [37]3.213.644.977.487.645.34.915.841.973.886.044.57.322.752.980.089.550.4
DINOv2 (ViT-B/14) [37]4.215.246.678.988.146.65.918.645.975.786.946.610.028.659.582.987.454.4
Chrono (ViT-S/14)65.581.788.892.393.784.332.255.273.984.088.766.829.756.476.786.890.968.0
Chrono (ViT-B/14)68.384.090.392.994.586.033.557.275.885.890.268.531.859.278.888.492.270.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.6, + 0.907, + 0.628 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on the TAP-Vid datasets [12] with the strided query mode. Best scores are in bold and second best are underlined for each column." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.483, + 0.787 + ], + "angle": 0, + "content": "modes, Chrono (ViT-S/14) and Chrono (ViT-B/14) consistently outperform competitors in PCK accuracy at most thresholds. The high accuracy at \\(< \\delta^0\\) underscores the precision of Chrono features in point tracking, with further gains at higher thresholds. At \\(< \\delta^4\\), only Chrono backbones achieve over \\(90\\%\\) accuracy on the DAVIS dataset in strided mode. Both Chrono (ViT-S/14) and Chrono (ViT-B/14) reach the highest position accuracy among backbones on average, demonstrating Chrono's ability to effectively model spatial and temporal information in videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Qualitative comparison. We visualize the estimated tracks from the DAVIS dataset, with a qualitative comparison shown in Figure 5. Unlike other methods, which produce highly jittery and inconsistent tracks over time, Chrono generates temporally smooth and accurate tracks. This jitter in other methods is expected as they lack awareness of neighboring frames. In contrast, Chrono utilizes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.681 + ], + "angle": 0, + "content": "temporal adapters to maintain frame-to-frame consistency, allowing it to produce smooth tracks even when individual frame estimates are less stable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.771, + 0.712 + ], + "angle": 0, + "content": "4.4. Analysis and Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison to two-stage point tracking models with iterative refiner. As shown in Table 3, Chrono, although solely a feature backbone without a refiner, achieves point estimation precision comparable to full pipelines with refiners, while demonstrating significantly higher efficiency in terms of throughput. Unlike other models, which require refiners to inject temporal information for each query point, Chrono embeds temporal awareness directly in the feature backbone. This design substantially boosts throughput: for example, Chrono (ViT-B/14) achieves \\(12.5 \\times\\) the speed of TAPIR with only a \\(3.5\\%\\) drop in DAVIS accuracy, and even surpasses TAPIR in RGB-Stacking by \\(11.4\\%\\) p. Simi" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1967" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.199 + ], + "angle": 0, + "content": "
BackboneRGB-Stacking-FirstKinetics-FirstDAVIS-First
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]31.458.377.887.792.469.68.628.856.574.283.350.39.027.354.973.784.149.8
TSM-ResNet-18 [12, 32]19.251.778.786.890.565.47.828.155.271.480.248.67.323.146.766.679.244.6
CoTraker [25] Backbone32.054.870.278.284.063.924.141.955.162.267.050.127.846.458.664.168.553.1
DINOv2 (ViT-S/14) [37]3.011.838.972.284.142.04.213.636.365.979.239.86.019.947.573.784.546.3
DINOv2 (ViT-B/14) [37]3.613.241.174.184.843.45.116.040.067.980.441.98.924.753.877.087.150.3
Chrono (ViT-S/14)58.277.285.489.191.380.224.846.265.877.583.159.524.049.271.282.887.963.0
Chrono (ViT-B/14)61.879.686.990.292.282.126.048.468.279.885.361.626.152.674.584.990.065.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.203, + 0.907, + 0.231 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on the TAP-Vid datasets [12] with the query-first mode. Best scores are in bold and second best are underlined for each column." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.242, + 0.482, + 0.332 + ], + "angle": 0, + "content": "
MethodRGB-Stacking-SDAVIS-SThroughput (points/sec)# of Refiner Params.Additional Params.
RAFT [45]58.646.323,405.714.2M4.2M
TAP-Net [12]72.853.129,535.985.5K5.5K
TAPIR [13]74.673.62,097.3225.9M25.9M
PIPs [19]51.059.446.4326.0M26.0M
Chrono (ViT-S/14)84.368.034,396.300M16.2M
Chrono (ViT-B/14)86.070.126,139.860M26.0M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.419 + ], + "angle": 0, + "content": "Table 3. Comparison on point tracking pipelines using refiners. Chrono achieves competitive or superior accuracy while offering significantly higher throughput, despite not using a heavy iterative refiner. TAP-Vid-RGB-Stacking and TAP-Vid-DAVIS metrics report \\(< \\delta_{avg}^{x}\\). Throughput is measured on a 24-frame video using a single NVIDIA RTX 3090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.422, + 0.482, + 0.538 + ], + "angle": 0, + "content": "
MethodRGB-StackingKineticsDAVIS
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
Kubric-VFS-Like [16]57.972.691.940.559.080.033.148.579.4
TAP-Net [12]59.972.890.446.660.985.038.453.182.3
PIPs [19]37.351.091.635.354.877.442.059.482.1
RAFT [45]44.058.690.434.552.579.730.046.379.6
TAPIR [13]62.774.691.657.270.187.861.373.688.8
FlowTrack [10]------66.079.887.2
CoTracker [25]------65.979.489.9
LocoTrack [11]77.186.993.259.573.088.567.879.689.9
Chrono + LocoTrack83.291.094.260.773.789.568.280.291.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of Chrono, adapted to LocoTrack [11], on the TAP-Vid dataset with the strided query mode. Our model shows a performance boost over LocoTrack on all datasets, with a particularly significant improvement on RGB-Stacking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.696 + ], + "angle": 0, + "content": "larly, Chrono (ViT-S/14) delivers \\(16.4 \\times\\) higher throughput. These results demonstrate that Chrono provides a simple yet highly effective solution, offering performance comparable to refiner-based models without requiring iterative processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Analysis on Chrono as a backbone in existing point tracking pipeline. In Table 4, we present an analysis of the impact of adding a refiner after our model to investigate its ability as a feature backbone within a point tracking pipeline that uses iterative refinement [11, 13, 19, 25]. To assess the ability to predict precise occlusion status, we use occlusion accuracy (OA) and the Average Jaccard (AJ) metric [12], which measures both position and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For our experiment, we incorporate the iterative refiner from LocoTrack [11]. Since LocoTrack utilizes three levels of hierarchical features from ResNet, we apply a simple convolutional upsampler to our single-resolution feature map. Specifically, using transposed convolution [52], we" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.243, + 0.905, + 0.33 + ], + "angle": 0, + "content": "
MethodDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
1D Conv.26.553.174.585.490.165.9
3D Conv.27.554.374.985.590.066.4
1D Attn. (Ours)29.756.476.786.890.968.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.333, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Table 5. Ablation on temporal aggregation design choice. 1D attention demonstrates the best performance compared to both 1D and 3D convolution layers." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.46 + ], + "angle": 0, + "content": "
Configuration# of \nAdaptersDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
Early Blocks623.147.869.881.486.661.7
Later Blocks626.652.974.385.390.165.8
Alternating Blocks626.953.474.385.390.065.9
All Blocks1129.756.476.786.890.968.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.463, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Table 6. Ablation on the number and position of temporal adapter. Rather than placing the temporal adapter into early or later blocks, or alternating blocks in DINOv2, applying it to all layers significantly boosts performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.608 + ], + "angle": 0, + "content": "generate a \\(2 \\times\\) feature map with a \\(4 \\times\\) channel reduction and then a \\(4 \\times\\) feature map with a further \\(2 \\times\\) channel reduction. We freeze our model and train the convolutional upsampler and the LocoTrack refiner on the Kubric [16] panning MOVi-E dataset [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.731 + ], + "angle": 0, + "content": "Our model combined with LocoTrack surpasses the original LocoTrack and also surpasses the performance of recent state-of-the-art point trackers [10, 25]. Our model shows a boost on all datasets, and notably, it shows a huge boost in the RGB-Stacking dataset, achieving a \\(+6.1\\) improvement in the AJ score. This result demonstrates the effectiveness of our approach when adopted into an existing point tracking pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ablation on temporal aggregation design choice. To model temporal information in our temporal adapter, we tested two approaches for temporal aggregation: convolution and attention. In Table 5, we present an ablation study where only the temporal aggregation layer within Chrono is modified, comparing 1D convolution, 3D convolution, and 1D attention, with all other components kept constant. While 1D convolution and 1D attention aggregate along the temporal axis only, 3D convolution considers both spatial and temporal dimensions. Due to computational constraints, this study is conducted on Chrono (ViT-S/14)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1968" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.904, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Figure 6. Visualization of features. We visualize the features generated by our model and DINOv2 using PCA. The results show that our model demonstrates improved temporal smoothness and finer-grained feature representation relative to DINOv2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.492 + ], + "angle": 0, + "content": "The results suggest that 1D attention is effective for temporal aggregation, as it allows the model to capture motion dynamics by adapting to correlations between frames. This adaptability may contribute to improved tracking accuracy by enabling the model to weigh information from different frames and handle varied motion patterns and occlusions. In contrast, temporal convolutions, with their fixed weights, may be less effective in capturing these complex temporal relationships." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Ablation on the number and position of temporal adapters. In Table 6, we conduct an ablation study on the placement of temporal adapters in Chrono, which consists of 12 transformer blocks. We evaluate three configurations: placing six temporal adapters within the early six blocks, within the later six blocks, and in alternating blocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.788 + ], + "angle": 0, + "content": "Early transformer blocks often capture local details, while later blocks tend to focus on broader, global patterns and complex relationships [15, 40]. Therefore, it's plausible that adapters in the initial blocks may learn local motion cues, while those in later blocks might capture overall motion patterns. Placing adapters only in the early or later blocks might emphasize either local or global motion, respectively. While the alternating configuration shows some improvement, it may not be optimal. Integrating temporal adapters between each block could potentially allow for a more balanced capture of both local and global motion patterns, making features sensitive to multilevel motions and leading to better performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Computation and latency comparison. Table 7 compares our inference time and computation with that of DINOv2 [37], measuring the time and GPU memory taken to extract features for 24 frames. While our method exhibits comparable computation and slower inference due to temporal injection into the feature, Chrono achieves a significantly higher score than vanilla DINOv2. This higher accu" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.353, + 0.906, + 0.442 + ], + "angle": 0, + "content": "
Method<δxavgInference Time ↓ (sec)GPU Memory ↓ (MiB)
DINOv2 (ViT-S/14) [37]39.80.1941,460
DINOv2 (ViT-B/14) [37]41.90.4432,816
Chrono (ViT-S/14)68.00.5751,550
Chrono (ViT-B/14)70.11.1972,998
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.447, + 0.907, + 0.49 + ], + "angle": 0, + "content": "Table 7. Computation and latency comparison with DINOv2. We measure inference time and GPU memory usage for processing 24 frames without point prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.906, + 0.528 + ], + "angle": 0, + "content": "racy compensates for the increased time required for feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.906, + 0.601 + ], + "angle": 0, + "content": "Visualization of feature with PCA. We visually compare the feature of Chrono and DINOv2 [37] with PCA (Principal Component Analysis) in Figure 6. We reduce the feature dimensions to three, represented as RGB in the figure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.708 + ], + "angle": 0, + "content": "While DINOv2's visualizations demonstrate a noisy background and tend to show a uniform representation within a single object, our model produces consistent and smooth representations over time and also shows fine granularity within the same semantic object. We believe the temporally smooth representation and fine-grained feature detail lead to better point tracking." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we presented Chrono, a temporally-aware feature backbone for point tracking that integrates pretrained DINOv2 representations with a temporal adapter, enabling long-term temporal context capture in feature space. Through extensive experiments, we demonstrate that embedding temporal information directly into the feature backbone, Chrono reduces reliance on costly refiners while achieving both accuracy and efficiency. We anticipate Chrono will inspire advancements in efficient, temporally-aware backbones for point tracking." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1969" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.232 + ], + "angle": 0, + "content": "Acknowledgements This research was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, RS-2024-00509279, RS-2025-II212068, RS-2023-00227592) and the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (RS-2024-00345025, RS-2024-00333068), and National Research Foundation of Korea (RS-2024-00346597)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.261, + 0.188, + 0.277 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.286, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.33, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. Advances in Neural Information Processing Systems, 34:9011-9023, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268-19277, 2024. 2, 3, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.286, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.152, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.906, + 0.305 + ], + "angle": 0, + "content": "[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.906, + 0.377 + ], + "angle": 0, + "content": "[14] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.381, + 0.906, + 0.422 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.425, + 0.906, + 0.508 + ], + "angle": 0, + "content": "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[17] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.583, + 0.906, + 0.638 + ], + "angle": 0, + "content": "[18] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.641, + 0.906, + 0.682 + ], + "angle": 0, + "content": "[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.686, + 0.906, + 0.741 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.744, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1970" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[24] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[26] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In 5th Annual Conference on Robot Learning, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[29] Junghyup Lee, Dohyung Kim, Jean Ponce, and Bumsub Ham. Sfnet: Learning object-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2278-2287, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[30] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[31] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[32] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[34] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[40] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[42] Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers. Transactions of the Association for Computational Linguistics, 9:53-68, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[43] Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Time does tell: Self-supervised timetuning of dense image representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16536-16547, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[44] Aleksandar Shtedritski, Andrea Vedaldi, and Christian Rupprecht. Learning universal semantic correspondences with no supervision and automatic data curation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 933-943, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision, pages 367-385. Springer, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[48] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. arXiv preprint arXiv:2308.15975, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1971" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[50] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[51] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[52] Matthew D Zeiler, Dilip Krishnan, Graham W Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on computer vision and pattern recognition, pages 2528-2535. IEEE, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[53] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[54] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1972" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_origin.pdf b/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a3e6c7bf632ca38d1273a233a9f17296a6580f9d --- /dev/null +++ b/2025/Exploring Temporally-Aware Features for Point Tracking/ac98a5e9-90fa-4012-89c2-06839140a6cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61937b1e48df766480aea47ca68da40f17a394a98c8bbbc22bfc926a42fcd16d +size 7004225 diff --git a/2025/Exploring Temporally-Aware Features for Point Tracking/full.md b/2025/Exploring Temporally-Aware Features for Point Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a3371bdf49576bfaa0b029fee4cca206f28cb232 --- /dev/null +++ b/2025/Exploring Temporally-Aware Features for Point Tracking/full.md @@ -0,0 +1,327 @@ +# Exploring Temporally-Aware Features for Point Tracking + +Inès Hyeonsu Kim1* Seokju Cho1* Jiahui Huang2 Jung Yi1 +Joon-Young Lee2 Seungryong Kim1 +1KAIST AI 2Adobe Research + +# Abstract + +Point tracking in videos is a fundamental task with applications in robotics, video editing, and more. While many vision tasks benefit from pre-trained feature backbones to improve generalizability, point tracking has primarily relied on simpler backbones trained from scratch on synthetic data, which may limit robustness in real-world scenarios. Additionally, point tracking requires temporal awareness to ensure coherence across frames, but using temporally-aware features is still underexplored. Most current methods often employ a two-stage process: an initial coarse prediction followed by a refinement stage to inject temporal information and correct errors from the coarse stage. These approach, however, is computationally expensive and potentially redundant if the feature backbone itself captures sufficient temporal information. + +In this work, we introduce Chrono, a feature backbone specifically designed for point tracking with built-in temporal awareness. Leveraging pre-trained representations from self-supervised learner DINOv2 and enhanced with a temporal adapter, Chrono effectively captures long-term temporal context, enabling precise prediction even without the refinement stage. Experimental results demonstrate that Chrono achieves state-of-the-art performance in a refiner-free setting on the TAP-Vid-DAVIS and TAP-Vid-Kinetics datasets, among common feature backbones used in point tracking as well as DINOv2, with exceptional efficiency. Project page: https://cvlab-kaist.github.io/Chrono/ + +# 1. Introduction + +Point tracking aims to track any point within a casual video, which has wide-ranging applications such as robotics [48], video editing [23], and view synthesis [50]. It involves establishing correspondences of specific points across frames to determine their positions over time and whether they are visible or occluded. Accurate point tracking is challenging + +![](images/fd814143d84d78e9371f27836ae18b44e17cfd22e719dbe0901b359d17ef3451.jpg) +Figure 1. Chrono is a highly precise, temporally-aware feature backbone specifically designed for point tracking. Unlike prior methods that rely on heavy iterative refinement for each query point [11-13, 19, 25] or test-time optimization [46, 49], Chrono achieves competitive tracking performance through simple feature matching using a non-parametric operation. Moreover, its performance can be further improved by an optional refinement step. + +due to complex motions, occlusions, and deformations [11- 14]. Achieving robust tracking requires effective matching of points across frames and a comprehensive understanding of temporal dynamics. + +In many computer vision tasks such as segmentation, object detection, and depth estimation, it is often beneficial to use pre-trained feature backbones along with task-specific heads [4, 6-8, 21]. These backbones, trained on large-scale datasets, offer robust representations that enhance generalizability to real-world data. In contrast, point tracking models often rely on simpler backbones such as ResNet [20] and are frequently trained from scratch [11-13, 19, 25] on synthetic dataset [16, 54]. This difference raises an important question: Could point tracking similarly benefit from adopting pre-trained feature backbones? + +While these backbones are often pre-trained on single-image tasks [15, 33, 37, 46, 53], point tracking inherently demands temporal awareness to maintain coherence across video frames. Unlike single-image tasks, point tracking requires consistent point matching over time, necessitating a model capable of capturing temporal informa + +![](images/091872ec61229992bb6f64163161730e9f3fbbf1b8d04eda6e1cd377e6d3fca3.jpg) +Figure 2. Comparison of initial trajectories from Chrono and existing feature backbone for point tracking. Chrono demonstrates a significant improvement over existing feature backbones such as TSM-ResNet [32]. + +tion [11, 14, 25, 26, 31, 48]. Therefore, features used in point tracking also need to be temporally aware. + +Despite its critical role, temporally-aware backbone has been relatively under-explored in this area. Early efforts, such as simply incorporating TSM-ResNet [32] in point tracking models [12, 13], yielded limited improvements. This outcome may be attributed to two factors: First, training on small synthetic datasets from scratch [16] may limit generalization to real-world data. Second, TSM-ResNet's temporal awareness, constrained to adjacent frames, may not provide long-term temporal context for point tracking. + +On the other hand, point tracking methods often adopt a two-stage process [11, 13, 14]: an initial stage that predicts coarse tracks directly from the simple feature backbone such as ResNet, followed by a refinement stage that introduces temporal information by analyzing trajectories across frames and iteratively refining the prediction. This pipeline has become standard, with refinement methods aiming to mitigate noise and enhance track smoothness [11, 14, 25, 26, 31, 48]. However, these refinement steps can be computationally demanding and query-dependent [10, 11], often requiring temporal refinement for each query point individually, which impacts efficiency. Moreover, a strong reliance on post-hoc refinement adds a substantial burden in correcting initial prediction errors from the first stage. This error could potentially be mitigated by leveraging features with built-in temporal awareness. + +To overcome these limitations, we introduce a temporally-aware feature backbone for point tracking, called Chrono, as illustrated in Figure 1. Our key insights are twofold: First, we leverage robust feature representations learned from large-scale, real-world datasets. We employ DINOv2 [37], a pre-trained model known for + +strong feature representations across various tasks such as segmentation, classification, and localization [30, 37]. However, these representations are not directly compatible with point tracking due to a lack of temporal awareness. Second, we incorporate temporal awareness by designing a temporal adapter that enables the pre-trained backbone to process data with temporal context, without losing its learned knowledge. This enables the use of these strong feature backbones for point tracking tasks. Unlike the adjacent-frame context used in TSM-ResNet [32], our approach incorporates a temporal context that is six times longer, enabling the capture of complex dynamics and enhancing tracking performance, as shown in Figure 2. + +Experimental results show that tracks estimated from Chrono features, computed solely using soft argmax without any learnable layers or additional temporal information after feature extraction, significantly outperform traditional feature backbones commonly used in point tracking as well as DINOv2 [37]. Specifically, Chrono achieves a $+20.6\%$ p increase in the position accuracy on the TAP-ViddAVIS [12] dataset compared to TSM-ResNet-18 [32]. Additionally, our method demonstrates superior efficiency and comparable performance compared to point tracking models equipped with refiners that inject temporal information post-feature extraction. Despite the absence of learnable layers after feature extraction, Chrono achieves high precision, underscoring that directly embedding temporal information within the features is both efficient and powerful. Optionally, we can add iterative refinement on top of our backbone, which can further enhance the performance of the refiner compared to when an existing backbone is used. In summary, our contributions are as follows: + +- We highlight the lack of temporally-aware feature backbones and the reliance on computationally intensive refinement processes in current point tracking methods. +- We propose a feature backbone designed for point tracking that incorporates a long-range temporal adapter, enhancing temporal awareness over extended sequences. +- We demonstrate that our backbone produces accurate initial tracks in simple and effective manner, reducing the need for extensive refinement and achieving both improved performance and efficiency. + +# 2. Related Work + +Point tracking. PIPs [19] independently tracks a point by fetching a local correlation around the point estimate and gradually refining the tracking result through iteration. TAP-Net [12] utilizes a shallow TSM-ResNet [32] for point tracking, with lightweight layers added to the backbone. TAPIR [13] integrates TAP-Net with the iterative refinement from PIPs, modifying the architecture to a convolution-based model for temporal processing. CoTracker tracks multiple points simultaneously, modeling + +their dependencies with a Transformer architecture. LocoTrack [11] achieves improved correspondence using local 4D correlation, inspired by dense matching literature [9]. TAPTR [31] introduces the DETR-like [6] architecture for point tracking. Another line of work employs test-time optimization with regularization, such as tracking smoothness, geometric constraints, and cycle consistency [46, 49]. While these methods focus on designing better architectures for track refinement, the exploration of more effective feature backbones has been relatively underexplored. Our work focuses on developing an improved feature backbone. + +Feature backbone in point tracking. Early works on point tracking [12] explored the performance of self-supervised feature backbones, such as VFS [51], as well as temporally-aware features like TSM-ResNet [32]. However, the exploration of feature backbones has been relatively underdeveloped compared to iterative track refinement [10, 11, 13, 25]. DINO-Tracker [46] explored the use of DINO in point tracking but it requires an hour of optimization for each video. Recently, [2] investigated the tracking capabilities of foundational models, such as Stable Diffusion [41] and DINOv2 [37], though the potential to extend these models for temporal applications remains unexplored. + +Adapting large feature backbone. With recent advances in large-scale self-supervised training [22, 37], effectively transferring its vast knowledge has become a key challenge. Specifically, the strong features of DINOv2 have been adapted to various tasks. [43] fine-tunes DINO for video tasks in a self-supervised manner. Additionally, DINO has been applied to various correspondence tasks [1, 17, 30, 35, 44] and has demonstrated robustness in establishing correspondence and semantic segmentation [1, 18, 36], showing its rich semantics. We focus on adapting this representation to the point tracking task, enhancing its temporal consistency while preserving its pre-trained knowledge. + +# 3. Method + +In this section, we introduce our temporally aware feature backbone. We begin by formally defining the point tracking problem and its associated challenges. We then describe the design of our temporal adapter and its integration with a pre-trained backbone to incorporate temporal awareness. Finally, we present our approach for adapting the pretrained backbone and training the temporal adapter for point tracking. + +# 3.1. Preliminaries and Motivation + +Task definition. Point tracking [10-14, 25, 31] in videos involves establishing correspondences of specific points across frames, determining their positions over time, and identifying whether they are visible or occluded. Formally, + +given a video sequence $\{\mathcal{I}_t\}_{t = 0}^{T - 1}$ , where $\mathcal{I}_t \in \mathbb{R}^{H \times W \times 3}$ represents the $t$ -th frame of height $H$ and width $W$ , and a query point $q = (x_q, y_q, t_q) \in \mathbb{R}^3$ at frame $t_q$ , the goal is to produce a trajectory $\mathcal{T} = \{\hat{p}_t\}_{t = 0}^{T - 1}$ , where $\hat{p}_t \in \mathbb{R}^2$ denotes the estimated position of the point at time $t$ , and an associated visibility probability $\mathcal{V} = \{v_t\}_{t = 0}^{T - 1}$ , where $v_t \in [0,1]$ indicates the likelihood of the point being visible at frame $t$ . In this work, we focus on predicting accurate positions rather than occlusions, as our primary goal is to achieve better point correspondences. + +Motivation. Accurate point tracking in videos requires robust point matching across frames, a challenging task due to complex motions, occlusions, deformations, and scale variations [11, 25, 49]. Because videos consist of sequences over time, considering both spatial and temporal aspects is essential to understand point dynamics. More specifically, a good feature for point tracking must satisfy two key criteria. Spatially, it should effectively model complex real-world data with powerful feature representations vital for robust matching, the cornerstone of point tracking [11]. Temporally, it must capture complex motions and understand the dynamics of points over time to handle their movement across frames. + +To fulfill the spatial criterion, we utilize DINOv2 [37], a self-supervised model trained on large-scale real-world data, renowned for its robust feature representations across various tasks [5, 30, 37]. However, DINOv2 lacks inherent temporal awareness, which is essential for point tracking in videos. To address this limitation, we introduce a temporal adapter that supplements DINOv2's features with temporal information. This adapter integrates temporal understanding directly into the feature extraction process, creating a feature backbone tailored for point tracking that combines strong spatial representations with temporal dynamics. + +# 3.2. Temporal Adapter + +Temporal awareness is essential for accurately tracking points across frames, as it enables the model to capture motion patterns and temporal dependencies. To embed this capability within DINOv2, we design a temporal adapter that allows the backbone to incorporate information from adjacent frames, enhancing its temporal sensitivity. Figure 3 illustrates the overall architecture of Chrono. + +Design of the temporal adapter. Our temporal adapter is placed between each transformer block of DINOv2 [37] to enhance temporal modeling across multiple feature levels. By computing adjacent features from other frames at once, the adapter connects features from different time steps, enabling it to recognize motion across frames. + +Each adapter follows a bottleneck structure inspired by the ResNet [20] architecture. It begins with a 2D convolutional layer Conv2D $^{\text{down}}$ with stride $s$ that reduces the + +![](images/d6a3b9ea0b55e61a19776ae4b3754c0051a6447a21a3533b136c82fa30583377.jpg) +Figure 3. Overall architecture of Chrono. Temporal adapters between transformer blocks use 2D convolution and 1D local attention along the temporal axis to output temporally-aware features. + +spatial dimensions from $H_{p} \times W_{p} \times C_{\mathrm{in}}$ to a compact $H_{p}^{\prime} \times W_{p}^{\prime} \times C_{\mathrm{out}}$ , where $H_{p}^{\prime} = H_{p} / s$ and $W_{p}^{\prime} = W_{p} / s$ . This spatial downsampling step optimizes the representation for efficient processing while expanding the spatial receptive field for subsequent operations. This is formally defined as: + +$$ +\mathbf {f} _ {t} ^ {\text {d o w n}} = \operatorname {C o n v 2 D} ^ {\text {d o w n}} \left(\mathbf {f} _ {t} ^ {\text {i n}}\right), \tag {1} +$$ + +where $\mathbf{f}_t^{\mathrm{in}}\in \mathbb{R}^{H_p\times W_p\times C_{\mathrm{in}}}$ and $\mathbf{f}_t^{\mathrm{out}}\in \mathbb{R}^{H_p'\times W_p'\times C_{\mathrm{out}}}$ are the input and output feature maps of Conv2D $^{\mathrm{down}}$ at time $t$ , respectively. + +After downsampling, we apply a temporal attention layer that captures dependencies within a local temporal window of size $N$ , focusing on temporal correlations across time [3, 42]. The local window attention is carefully calibrated to balance computational efficiency with effective motion capture, ensuring a broad enough temporal window without excessive resource demands. This layer operates over the neighboring frames within the window $[t - k, t + k]$ where $k = \frac{N - 1}{2}$ . For each spatial location $(x,y)$ , we compute attention weights based on the similarity between the query vector at time $t$ and key vectors from neighboring times. The aggregated feature at time $t$ and location $(x,y)$ , referred to as $\mathbf{f}_t^{\mathrm{att}}(x,y)$ , is computed as a weighted sum of + +![](images/e87dfa2071da79a9c3801c5383b65643a682d562e60656cacc48098952c367ec.jpg) +Figure 4. Point track prediction. To predict point positions, we simply match the query points with features from other frames, without using any learnable layers. + +the value vectors from these neighboring frames: + +$$ +\mathbf {f} _ {t} ^ {\mathrm {a t t n}} (x, y) = \sum_ {n = - k} ^ {k} \alpha^ {(t, n)} (x, y) \cdot \mathbf {V} _ {t + n} (x, y), \tag {2} +$$ + +where $\alpha^{(t,n)}(x,y)$ represents the attention weights for time offset $n$ , computed as: + +$$ +\alpha^ {(t, n)} (x, y) = \frac {\exp \left(\mathbf {Q} _ {t} (x , y) \cdot \mathbf {K} _ {t + n} (x , y) / \sqrt {d _ {k}}\right)}{\sum_ {n ^ {\prime} = - k} ^ {k} \exp \left(\mathbf {Q} _ {t} (x , y) \cdot \mathbf {K} _ {t + n ^ {\prime}} (x , y) / \sqrt {d _ {k}}\right)}, \tag {3} +$$ + +where $d_{k}$ is a scaling factor [47], $\mathbf{Q},\mathbf{K}$ , and $\mathbf{V}$ are projections of $\mathbf{f}^{\mathrm{down}}$ , obtained through the linear projection layers $\mathbf{W}_Q,\mathbf{W}_K$ , and $\mathbf{W}_V$ . + +After temporal attention, another 2D convolutional layer restores the spatial dimensions back to $H_{p} \times W_{p} \times C_{\mathrm{in}}$ : + +$$ +\mathbf {f} _ {t} ^ {\mathrm {u p}} = \operatorname {C o n v 2 D} ^ {\mathrm {u p}} \left(\mathbf {f} _ {t} ^ {\mathrm {a t t n}}\right), \tag {4} +$$ + +where $\mathbf{f}_t^{\mathrm{up}}\in \mathbb{R}^{H_p\times W_p\times C_{\mathrm{in}}}$ . Finally, a residual connection adds the original input feature map to the output of the adapter to preserve the feature representation of DINOv2: + +$$ +\mathbf {f} _ {t} ^ {\text {o u t}} = \mathbf {f} _ {t} ^ {\text {u p}} + \mathbf {f} _ {t} ^ {\text {i n}}. \tag {5} +$$ + +# 3.3. Point Prediction with the Feature Backbone + +As shown in Figure 4, we track a query point by simply performing feature matching using features from Chrono. This approach avoids the need for learnable modules such as iterative refinement, as used in [10, 11, 13, 25]. + +Correlation construction. Given the query point $q = (x_q, y_q, t_q)$ , we extract the query feature $\mathbf{f}_q$ from the feature map of frame $t_q$ at position $(x_q, y_q)$ using bilinear interpolation. For each frame $t$ , we compute the correlation map $\mathcal{C}_t$ by calculating the cosine similarity between $\mathbf{f}_q$ and the feature map $\mathbf{f}_t$ at every spatial location: + +$$ +\mathcal {C} _ {t} (x, y) = \frac {\mathbf {f} _ {q} ^ {\top} \mathbf {f} _ {t} (x , y)}{\| \mathbf {f} _ {q} \| \| \mathbf {f} _ {t} (x , y) \|}, \tag {6} +$$ + +where $\mathbf{f}_t(x,y)$ is the feature vector at position $(x,y)$ in frame $t$ , $\mathbf{f}_q^\top \mathbf{f}_t(x,y)$ denotes the dot product between $\mathbf{f}_q$ and + +$\mathbf{f}_t(x,y)$ , and $\| \mathbf{f}_q\|$ and $\| \mathbf{f}_t(x,y)\|$ indicate the Euclidean norms of $\mathbf{f}_q$ and $\mathbf{f}_t(x,y)$ , respectively. This correlation map $\mathcal{C}_t$ quantifies the similarity between the query feature and features across all spatial locations in each frame $t$ . + +Point prediction. To estimate the position of points in frame $t$ , we apply the soft-argmax operation [29] to the correlation map $C_t$ . The soft-argmax computes a weighted average of all spatial positions, with weights given by the softmax of the correlation values: + +$$ +\hat {\mathbf {p}} _ {t} = \sum_ {(x, y)} \sigma \left(C _ {t} (x, y)\right) \cdot (x, y), \tag {7} +$$ + +where $\sigma(C_t(x, y))$ is the softmax over all spatial positions within frame $t$ : + +$$ +\sigma \left(C _ {t} (x, y)\right) = \frac {\exp \left(C _ {t} (x , y) \cdot \tau\right)}{\sum_ {\left(x ^ {\prime} , y ^ {\prime}\right)} \exp \left(C _ {t} \left(x ^ {\prime} , y ^ {\prime}\right) \cdot \tau\right)}, \tag {8} +$$ + +where $\tau$ is a softmax temperature. The soft argmax is nonlearnable and provides a differentiable method to estimate the point's position based on the correlation map. To enhance precision, we mask out positions more than $M$ pixels away from the maximum correlation value, focusing the soft argmax computation on a local neighborhood and reducing the influence of irrelevant regions [29]. + +Training. To train Chrono, we use the Huber loss [24] to supervise the estimated positions. The Huber loss is chosen for its robustness to outliers compared to the squared error loss. At each time step $t$ , the loss is defined as: + +$$ +\mathcal {L} _ {\text {H u b e r}} \left(\hat {p} _ {t}, p _ {t}\right) = \left\{ \begin{array}{l l} \frac {1}{2} \left\| \hat {p} _ {t} - p _ {t} \right\| ^ {2}, & \text {i f} \| \hat {p} _ {t} - p _ {t} \| \leq \delta , \\ \delta \cdot \left(\left\| \hat {p} _ {t} - p _ {t} \right\| - \frac {1}{2} \delta\right), & \text {o t h e r w i s e}, \end{array} \right. \tag {9} +$$ + +where $p_t$ is the ground truth position, and $\delta$ is a threshold parameter. For occluded points indicated by the ground truth visibility status $v_t$ , we exclude the loss computation, effectively masking out those time steps: + +$$ +\mathcal {L} _ {t} = \left(1 - v _ {t}\right) \cdot \mathcal {L} _ {\text {H u b e r}} \left(\hat {p} _ {t}, p _ {t}\right). \tag {10} +$$ + +By minimizing this loss over all time steps and query points, we train the backbone to produce features for predicting accurate tracked positions. + +# 4. Experiments + +# 4.1. Implementation Details + +We implement our method using PyTorch [38]. During training, we use AdamW optimizer [34] with a learning rate of $10^{-4}$ , a weight decay of $10^{-4}$ , and a batch size of 1 per GPU. All models are trained for 100,000 iterations on 4 A100 GPUs, employing a cosine learning rate scheduler with warmup. For training, we use the Kubric Panning-MOVi-E dataset [11, 13] and sample 256 query points per + +batch. The hyperparameters are set as follows: the softmax temperature is $\tau = 20.0$ , the soft argmax pixel threshold is $M = 5$ and the local temporal window size is $N = 13$ , and the 2D convolution stride in the temporal adapter is $s = 4$ . + +# 4.2. Evaluation Protocol + +Evaluation datasets. To assess our approach, we employ the TAP-Vid benchmark [12], which comprises both real and synthetic video sets. The real videos come with accurate annotation tracks, while the synthetic videos are paired with perfect ground-truth trajectories. The benchmark consists of three datasets: TAP-Vid-RGB-Stacking, TAP-Vid-Kinetics, and TAP-Vid-DAVIS. TAP-Vid-RGB-Stacking [28] contains synthetic videos of a robot gripper stacking objects and includes 50 videos. The TAP-Vid-Kinetics [27] consists of 1,189 YouTube videos with diverse difficulties, including intense motion blur and abrupt scene transitions, while the TAP-Vid-DAVIS dataset [39] provides 30 videos featuring challenges such as substantial scale shifts of dynamic object. + +Evaluation metrics. We assess the accuracy of the predicted tracks using two metrics: position accuracy at various thresholds and average position accuracy $(< \delta_{\mathrm{avg}}^x)$ . Position accuracy is evaluated at five threshold levels: $< \delta^0$ , $< \delta^1$ , $< \delta^2$ , $< \delta^3$ , and $< \delta^4$ , corresponding to accuracies within pixel distances of 1, 2, 4, 8, and 16, respectively. Each $< \delta^x$ score represents the percentage of visible ground-truth points whose predicted positions fall within the specified threshold. The $< \delta_{\mathrm{avg}}^x$ score is the average across all thresholds. + +Following [12], we evaluate the datasets using two modes: strided query mode and first query mode. In strided query mode, query points are sampled along the groundtruth trajectory at intervals of 5 frames. In first query mode, query points are sampled from the first visible frame. + +# 4.3. Main Results + +Quantitative comparison. We evaluate our method against the backbones [20, 32] commonly used in point tracking [11-13, 25, 26] and DINOv2 [37] in both the strided query (Table 1) and the first query modes (Table 2). Specifically, we compare with ResNet-18 [20] from TAPIR [13], TSM-ResNet-18 [32] from TAP-Net [12], and the backbone from the pre-trained CoTracker model [25], all of which are trained for point tracking. We report position accuracy to demonstrate the effectiveness of temporal information in our model. + +Our small model, Chrono (ViT-S/14) with DINOv2 (ViTs/14), achieves state-of-the-art performance in position accuracy across most thresholds, outperforming other backbones on $< \delta_{\mathrm{avg}}^x$ . Chrono (ViT-B/14) with DINOv2 (ViTB/14) delivers even better results. Across RGB-Stacking, Kinetics and DAVIS datasets, in the strided and first query + +![](images/758f437a94c7e47de91ed06044205f5c6456706dad8f2b4c6a7dadd597473707.jpg) +Reference frame $(t = 0)$ + +![](images/dbc3950ad79c04f1f1bcb987b020c236d8a0e1114f48e20a0eebb41a7a149f0a.jpg) +Figure 5. Qualitative comparison of complex real-world video tracking. We qualitatively compare the results generated by Chrono with those from other commonly used backbones in point tracking and as well as DINOv2. Our model demonstrates better smoothness and precision than other competitors. + +![](images/2e5e5573f9b7c0868138d27a6a33b152b823c908cb453c9dbb885779a11389c2.jpg) +Reference frame $(t = 0)$ + +![](images/e2f90cf1178d19ef93d654e3f9df62fc1c18d26e7fba46695a1284af62f7bd97.jpg) + +
BackboneRGB-Stacking-StridedKinetics-StridedDAVIS-Strided
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]36.664.682.790.994.773.910.535.365.781.388.656.39.731.160.678.387.053.3
TSM-ResNet-18 [12, 32]18.155.583.789.892.567.99.1533.264.679.286.554.58.226.853.673.583.949.2
CoTraker [25] Backbone36.960.174.982.587.668.431.031.050.570.674.958.233.752.563.468.372.058.0
DINOv2 (ViT-S/14) [37]3.213.644.977.487.645.34.915.841.973.886.044.57.322.752.980.089.550.4
DINOv2 (ViT-B/14) [37]4.215.246.678.988.146.65.918.645.975.786.946.610.028.659.582.987.454.4
Chrono (ViT-S/14)65.581.788.892.393.784.332.255.273.984.088.766.829.756.476.786.890.968.0
Chrono (ViT-B/14)68.384.090.392.994.586.033.557.275.885.890.268.531.859.278.888.492.270.1
+ +Table 1. Quantitative comparison on the TAP-Vid datasets [12] with the strided query mode. Best scores are in bold and second best are underlined for each column. + +modes, Chrono (ViT-S/14) and Chrono (ViT-B/14) consistently outperform competitors in PCK accuracy at most thresholds. The high accuracy at $< \delta^0$ underscores the precision of Chrono features in point tracking, with further gains at higher thresholds. At $< \delta^4$ , only Chrono backbones achieve over $90\%$ accuracy on the DAVIS dataset in strided mode. Both Chrono (ViT-S/14) and Chrono (ViT-B/14) reach the highest position accuracy among backbones on average, demonstrating Chrono's ability to effectively model spatial and temporal information in videos. + +Qualitative comparison. We visualize the estimated tracks from the DAVIS dataset, with a qualitative comparison shown in Figure 5. Unlike other methods, which produce highly jittery and inconsistent tracks over time, Chrono generates temporally smooth and accurate tracks. This jitter in other methods is expected as they lack awareness of neighboring frames. In contrast, Chrono utilizes + +temporal adapters to maintain frame-to-frame consistency, allowing it to produce smooth tracks even when individual frame estimates are less stable. + +# 4.4. Analysis and Ablation Study + +Comparison to two-stage point tracking models with iterative refiner. As shown in Table 3, Chrono, although solely a feature backbone without a refiner, achieves point estimation precision comparable to full pipelines with refiners, while demonstrating significantly higher efficiency in terms of throughput. Unlike other models, which require refiners to inject temporal information for each query point, Chrono embeds temporal awareness directly in the feature backbone. This design substantially boosts throughput: for example, Chrono (ViT-B/14) achieves $12.5 \times$ the speed of TAPIR with only a $3.5\%$ drop in DAVIS accuracy, and even surpasses TAPIR in RGB-Stacking by $11.4\%$ p. Simi + +
BackboneRGB-Stacking-FirstKinetics-FirstDAVIS-First
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]31.458.377.887.792.469.68.628.856.574.283.350.39.027.354.973.784.149.8
TSM-ResNet-18 [12, 32]19.251.778.786.890.565.47.828.155.271.480.248.67.323.146.766.679.244.6
CoTraker [25] Backbone32.054.870.278.284.063.924.141.955.162.267.050.127.846.458.664.168.553.1
DINOv2 (ViT-S/14) [37]3.011.838.972.284.142.04.213.636.365.979.239.86.019.947.573.784.546.3
DINOv2 (ViT-B/14) [37]3.613.241.174.184.843.45.116.040.067.980.441.98.924.753.877.087.150.3
Chrono (ViT-S/14)58.277.285.489.191.380.224.846.265.877.583.159.524.049.271.282.887.963.0
Chrono (ViT-B/14)61.879.686.990.292.282.126.048.468.279.885.361.626.152.674.584.990.065.6
+ +Table 2. Quantitative comparison on the TAP-Vid datasets [12] with the query-first mode. Best scores are in bold and second best are underlined for each column. + +
MethodRGB-Stacking-SDAVIS-SThroughput (points/sec)# of Refiner Params.Additional Params.
RAFT [45]58.646.323,405.714.2M4.2M
TAP-Net [12]72.853.129,535.985.5K5.5K
TAPIR [13]74.673.62,097.3225.9M25.9M
PIPs [19]51.059.446.4326.0M26.0M
Chrono (ViT-S/14)84.368.034,396.300M16.2M
Chrono (ViT-B/14)86.070.126,139.860M26.0M
+ +Table 3. Comparison on point tracking pipelines using refiners. Chrono achieves competitive or superior accuracy while offering significantly higher throughput, despite not using a heavy iterative refiner. TAP-Vid-RGB-Stacking and TAP-Vid-DAVIS metrics report $< \delta_{avg}^{x}$ . Throughput is measured on a 24-frame video using a single NVIDIA RTX 3090 GPU. + +
MethodRGB-StackingKineticsDAVIS
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
Kubric-VFS-Like [16]57.972.691.940.559.080.033.148.579.4
TAP-Net [12]59.972.890.446.660.985.038.453.182.3
PIPs [19]37.351.091.635.354.877.442.059.482.1
RAFT [45]44.058.690.434.552.579.730.046.379.6
TAPIR [13]62.774.691.657.270.187.861.373.688.8
FlowTrack [10]------66.079.887.2
CoTracker [25]------65.979.489.9
LocoTrack [11]77.186.993.259.573.088.567.879.689.9
Chrono + LocoTrack83.291.094.260.773.789.568.280.291.2
+ +larly, Chrono (ViT-S/14) delivers $16.4 \times$ higher throughput. These results demonstrate that Chrono provides a simple yet highly effective solution, offering performance comparable to refiner-based models without requiring iterative processing. + +Analysis on Chrono as a backbone in existing point tracking pipeline. In Table 4, we present an analysis of the impact of adding a refiner after our model to investigate its ability as a feature backbone within a point tracking pipeline that uses iterative refinement [11, 13, 19, 25]. To assess the ability to predict precise occlusion status, we use occlusion accuracy (OA) and the Average Jaccard (AJ) metric [12], which measures both position and occlusion. + +For our experiment, we incorporate the iterative refiner from LocoTrack [11]. Since LocoTrack utilizes three levels of hierarchical features from ResNet, we apply a simple convolutional upsampler to our single-resolution feature map. Specifically, using transposed convolution [52], we + +Table 4. Quantitative comparison of Chrono, adapted to LocoTrack [11], on the TAP-Vid dataset with the strided query mode. Our model shows a performance boost over LocoTrack on all datasets, with a particularly significant improvement on RGB-Stacking. + +
MethodDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
1D Conv.26.553.174.585.490.165.9
3D Conv.27.554.374.985.590.066.4
1D Attn. (Ours)29.756.476.786.890.968.0
+ +Table 5. Ablation on temporal aggregation design choice. 1D attention demonstrates the best performance compared to both 1D and 3D convolution layers. + +
Configuration# of +AdaptersDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
Early Blocks623.147.869.881.486.661.7
Later Blocks626.652.974.385.390.165.8
Alternating Blocks626.953.474.385.390.065.9
All Blocks1129.756.476.786.890.968.0
+ +Table 6. Ablation on the number and position of temporal adapter. Rather than placing the temporal adapter into early or later blocks, or alternating blocks in DINOv2, applying it to all layers significantly boosts performance. + +generate a $2 \times$ feature map with a $4 \times$ channel reduction and then a $4 \times$ feature map with a further $2 \times$ channel reduction. We freeze our model and train the convolutional upsampler and the LocoTrack refiner on the Kubric [16] panning MOVi-E dataset [13]. + +Our model combined with LocoTrack surpasses the original LocoTrack and also surpasses the performance of recent state-of-the-art point trackers [10, 25]. Our model shows a boost on all datasets, and notably, it shows a huge boost in the RGB-Stacking dataset, achieving a $+6.1$ improvement in the AJ score. This result demonstrates the effectiveness of our approach when adopted into an existing point tracking pipeline. + +Ablation on temporal aggregation design choice. To model temporal information in our temporal adapter, we tested two approaches for temporal aggregation: convolution and attention. In Table 5, we present an ablation study where only the temporal aggregation layer within Chrono is modified, comparing 1D convolution, 3D convolution, and 1D attention, with all other components kept constant. While 1D convolution and 1D attention aggregate along the temporal axis only, 3D convolution considers both spatial and temporal dimensions. Due to computational constraints, this study is conducted on Chrono (ViT-S/14). + +![](images/af6dc6c3a6c1029debf1616cdddc83394e3d1acd82f9aa64f3cd9afa7bbe18ac.jpg) +Figure 6. Visualization of features. We visualize the features generated by our model and DINOv2 using PCA. The results show that our model demonstrates improved temporal smoothness and finer-grained feature representation relative to DINOv2. + +The results suggest that 1D attention is effective for temporal aggregation, as it allows the model to capture motion dynamics by adapting to correlations between frames. This adaptability may contribute to improved tracking accuracy by enabling the model to weigh information from different frames and handle varied motion patterns and occlusions. In contrast, temporal convolutions, with their fixed weights, may be less effective in capturing these complex temporal relationships. + +Ablation on the number and position of temporal adapters. In Table 6, we conduct an ablation study on the placement of temporal adapters in Chrono, which consists of 12 transformer blocks. We evaluate three configurations: placing six temporal adapters within the early six blocks, within the later six blocks, and in alternating blocks. + +Early transformer blocks often capture local details, while later blocks tend to focus on broader, global patterns and complex relationships [15, 40]. Therefore, it's plausible that adapters in the initial blocks may learn local motion cues, while those in later blocks might capture overall motion patterns. Placing adapters only in the early or later blocks might emphasize either local or global motion, respectively. While the alternating configuration shows some improvement, it may not be optimal. Integrating temporal adapters between each block could potentially allow for a more balanced capture of both local and global motion patterns, making features sensitive to multilevel motions and leading to better performance. + +Computation and latency comparison. Table 7 compares our inference time and computation with that of DINOv2 [37], measuring the time and GPU memory taken to extract features for 24 frames. While our method exhibits comparable computation and slower inference due to temporal injection into the feature, Chrono achieves a significantly higher score than vanilla DINOv2. This higher accu + +
Method<δxavgInference Time ↓ (sec)GPU Memory ↓ (MiB)
DINOv2 (ViT-S/14) [37]39.80.1941,460
DINOv2 (ViT-B/14) [37]41.90.4432,816
Chrono (ViT-S/14)68.00.5751,550
Chrono (ViT-B/14)70.11.1972,998
+ +Table 7. Computation and latency comparison with DINOv2. We measure inference time and GPU memory usage for processing 24 frames without point prediction. + +racy compensates for the increased time required for feature extraction. + +Visualization of feature with PCA. We visually compare the feature of Chrono and DINOv2 [37] with PCA (Principal Component Analysis) in Figure 6. We reduce the feature dimensions to three, represented as RGB in the figure. + +While DINOv2's visualizations demonstrate a noisy background and tend to show a uniform representation within a single object, our model produces consistent and smooth representations over time and also shows fine granularity within the same semantic object. We believe the temporally smooth representation and fine-grained feature detail lead to better point tracking. + +# 5. Conclusion + +In this paper, we presented Chrono, a temporally-aware feature backbone for point tracking that integrates pretrained DINOv2 representations with a temporal adapter, enabling long-term temporal context capture in feature space. Through extensive experiments, we demonstrate that embedding temporal information directly into the feature backbone, Chrono reduces reliance on costly refiners while achieving both accuracy and efficiency. We anticipate Chrono will inspire advancements in efficient, temporally-aware backbones for point tracking. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 3 +[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 3 +[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 4 +[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1 +[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 3 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3 +[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. +[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 1 +[9] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. Advances in Neural Information Processing Systems, 34:9011-9023, 2021. 3 +[10] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268-19277, 2024. 2, 3, 4, 7 + +Acknowledgements This research was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, RS-2024-00509279, RS-2025-II212068, RS-2023-00227592) and the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (RS-2024-00345025, RS-2024-00333068), and National Research Foundation of Korea (RS-2024-00346597). +[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 1, 2, 3, 4, 5, 7 +[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 3, 5, 6, 7 +[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 1, 2, 3, 4, 5, 6, 7 +[14] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 1, 2, 3 +[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 8 +[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 1, 2, 7 +[17] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 3 +[18] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3 +[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 7 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 3, 5, 6, 7 +[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3 +[23] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 1 + +[24] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5 +[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2, 3, 4, 5, 6, 7 +[26] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 5 +[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5 +[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In 5th Annual Conference on Robot Learning, 2021. 5 +[29] Junghyup Lee, Dohyung Kim, Jean Ponce, and Bumsub Ham. Sfnet: Learning object-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2278-2287, 2019. 5 +[30] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3 +[31] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2, 3 +[32] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 2, 3, 5, 6, 7 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1 +[34] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 3 +[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3 + +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5, 6, 7, 8 +[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5 +[40] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 8 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[42] Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers. Transactions of the Association for Computational Linguistics, 9:53-68, 2021. 4 +[43] Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Time does tell: Self-supervised timetuning of dense image representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16536-16547, 2023. 3 +[44] Aleksandar Shtedritski, Andrea Vedaldi, and Christian Rupprecht. Learning universal semantic correspondences with no supervision and automatic data curation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 933-943, 2023. 3 +[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7 +[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision, pages 367-385. Springer, 2025. 1, 3 +[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4 +[48] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. arXiv preprint arXiv:2308.15975, 2023. 1, 2 +[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. + +Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 1, 3 +[50] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 1 +[51] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021. 3 +[52] Matthew D Zeiler, Dilip Krishnan, Graham W Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on computer vision and pattern recognition, pages 2528-2535. IEEE, 2010. 7 +[53] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 1 +[54] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 1 \ No newline at end of file diff --git a/2025/Exploring Temporally-Aware Features for Point Tracking/images.zip b/2025/Exploring Temporally-Aware Features for Point Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..529f7fca59dd39d64e24121ba6a31fb52f8c6645 --- /dev/null +++ b/2025/Exploring Temporally-Aware Features for Point Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a91c2c973ab7e84de38ced7afb6ab9d12ca3cd7e290ac3d2920750c99910c7b2 +size 906431 diff --git a/2025/Exploring Temporally-Aware Features for Point Tracking/layout.json b/2025/Exploring Temporally-Aware Features for Point Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0df000e600f6213158aa3007565bd4c14608ba61 --- /dev/null +++ b/2025/Exploring Temporally-Aware Features for Point Tracking/layout.json @@ -0,0 +1,8931 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "Exploring Temporally-Aware Features for Point Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 478, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 478, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 478, + 196 + ], + "type": "text", + "content": "Inès Hyeonsu Kim1* Seokju Cho1* Jiahui Huang2 Jung Yi1 \nJoon-Young Lee2 Seungryong Kim1 \n1KAIST AI 2Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 223, + 200, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 223, + 200, + 236 + ], + "spans": [ + { + "bbox": [ + 151, + 223, + 200, + 236 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 248, + 296, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 248, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 248, + 296, + 427 + ], + "type": "text", + "content": "Point tracking in videos is a fundamental task with applications in robotics, video editing, and more. While many vision tasks benefit from pre-trained feature backbones to improve generalizability, point tracking has primarily relied on simpler backbones trained from scratch on synthetic data, which may limit robustness in real-world scenarios. Additionally, point tracking requires temporal awareness to ensure coherence across frames, but using temporally-aware features is still underexplored. Most current methods often employ a two-stage process: an initial coarse prediction followed by a refinement stage to inject temporal information and correct errors from the coarse stage. These approach, however, is computationally expensive and potentially redundant if the feature backbone itself captures sufficient temporal information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 427, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 296, + 582 + ], + "type": "text", + "content": "In this work, we introduce Chrono, a feature backbone specifically designed for point tracking with built-in temporal awareness. Leveraging pre-trained representations from self-supervised learner DINOv2 and enhanced with a temporal adapter, Chrono effectively captures long-term temporal context, enabling precise prediction even without the refinement stage. Experimental results demonstrate that Chrono achieves state-of-the-art performance in a refiner-free setting on the TAP-Vid-DAVIS and TAP-Vid-Kinetics datasets, among common feature backbones used in point tracking as well as DINOv2, with exceptional efficiency. Project page: https://cvlab-kaist.github.io/Chrono/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": "Point tracking aims to track any point within a casual video, which has wide-ranging applications such as robotics [48], video editing [23], and view synthesis [50]. It involves establishing correspondences of specific points across frames to determine their positions over time and whether they are visible or occluded. Accurate point tracking is challenging" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 221, + 552, + 365 + ], + "blocks": [ + { + "bbox": [ + 316, + 221, + 552, + 365 + ], + "lines": [ + { + "bbox": [ + 316, + 221, + 552, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 221, + 552, + 365 + ], + "type": "image", + "image_path": "fd814143d84d78e9371f27836ae18b44e17cfd22e719dbe0901b359d17ef3451.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 372, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. Chrono is a highly precise, temporally-aware feature backbone specifically designed for point tracking. Unlike prior methods that rely on heavy iterative refinement for each query point [11-13, 19, 25] or test-time optimization [46, 49], Chrono achieves competitive tracking performance through simple feature matching using a non-parametric operation. Moreover, its performance can be further improved by an optional refinement step." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 509 + ], + "type": "text", + "content": "due to complex motions, occlusions, and deformations [11- 14]. Achieving robust tracking requires effective matching of points across frames and a comprehensive understanding of temporal dynamics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 510, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 640 + ], + "type": "text", + "content": "In many computer vision tasks such as segmentation, object detection, and depth estimation, it is often beneficial to use pre-trained feature backbones along with task-specific heads [4, 6-8, 21]. These backbones, trained on large-scale datasets, offer robust representations that enhance generalizability to real-world data. In contrast, point tracking models often rely on simpler backbones such as ResNet [20] and are frequently trained from scratch [11-13, 19, 25] on synthetic dataset [16, 54]. This difference raises an important question: Could point tracking similarly benefit from adopting pre-trained feature backbones?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "While these backbones are often pre-trained on single-image tasks [15, 33, 37, 46, 53], point tracking inherently demands temporal awareness to maintain coherence across video frames. Unlike single-image tasks, point tracking requires consistent point matching over time, necessitating a model capable of capturing temporal informa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 137, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1962" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 255 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 255 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 255 + ], + "type": "image", + "image_path": "091872ec61229992bb6f64163161730e9f3fbbf1b8d04eda6e1cd377e6d3fca3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "type": "text", + "content": "Figure 2. Comparison of initial trajectories from Chrono and existing feature backbone for point tracking. Chrono demonstrates a significant improvement over existing feature backbones such as TSM-ResNet [32]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 342 + ], + "type": "text", + "content": "tion [11, 14, 25, 26, 31, 48]. Therefore, features used in point tracking also need to be temporally aware." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 450 + ], + "type": "text", + "content": "Despite its critical role, temporally-aware backbone has been relatively under-explored in this area. Early efforts, such as simply incorporating TSM-ResNet [32] in point tracking models [12, 13], yielded limited improvements. This outcome may be attributed to two factors: First, training on small synthetic datasets from scratch [16] may limit generalization to real-world data. Second, TSM-ResNet's temporal awareness, constrained to adjacent frames, may not provide long-term temporal context for point tracking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 641 + ], + "type": "text", + "content": "On the other hand, point tracking methods often adopt a two-stage process [11, 13, 14]: an initial stage that predicts coarse tracks directly from the simple feature backbone such as ResNet, followed by a refinement stage that introduces temporal information by analyzing trajectories across frames and iteratively refining the prediction. This pipeline has become standard, with refinement methods aiming to mitigate noise and enhance track smoothness [11, 14, 25, 26, 31, 48]. However, these refinement steps can be computationally demanding and query-dependent [10, 11], often requiring temporal refinement for each query point individually, which impacts efficiency. Moreover, a strong reliance on post-hoc refinement adds a substantial burden in correcting initial prediction errors from the first stage. This error could potentially be mitigated by leveraging features with built-in temporal awareness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "To overcome these limitations, we introduce a temporally-aware feature backbone for point tracking, called Chrono, as illustrated in Figure 1. Our key insights are twofold: First, we leverage robust feature representations learned from large-scale, real-world datasets. We employ DINOv2 [37], a pre-trained model known for" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "strong feature representations across various tasks such as segmentation, classification, and localization [30, 37]. However, these representations are not directly compatible with point tracking due to a lack of temporal awareness. Second, we incorporate temporal awareness by designing a temporal adapter that enables the pre-trained backbone to process data with temporal context, without losing its learned knowledge. This enables the use of these strong feature backbones for point tracking tasks. Unlike the adjacent-frame context used in TSM-ResNet [32], our approach incorporates a temporal context that is six times longer, enabling the capture of complex dynamics and enhancing tracking performance, as shown in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": "Experimental results show that tracks estimated from Chrono features, computed solely using soft argmax without any learnable layers or additional temporal information after feature extraction, significantly outperform traditional feature backbones commonly used in point tracking as well as DINOv2 [37]. Specifically, Chrono achieves a " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "+20.6\\%" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " p increase in the position accuracy on the TAP-ViddAVIS [12] dataset compared to TSM-ResNet-18 [32]. Additionally, our method demonstrates superior efficiency and comparable performance compared to point tracking models equipped with refiners that inject temporal information post-feature extraction. Despite the absence of learnable layers after feature extraction, Chrono achieves high precision, underscoring that directly embedding temporal information within the features is both efficient and powerful. Optionally, we can add iterative refinement on top of our backbone, which can further enhance the performance of the refiner compared to when an existing backbone is used. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 455, + 553, + 574 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 455, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 455, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 314, + 455, + 553, + 490 + ], + "type": "text", + "content": "- We highlight the lack of temporally-aware feature backbones and the reliance on computationally intensive refinement processes in current point tracking methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "content": "- We propose a feature backbone designed for point tracking that incorporates a long-range temporal adapter, enhancing temporal awareness over extended sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 526, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 553, + 574 + ], + "type": "text", + "content": "- We demonstrate that our backbone produces accurate initial tracks in simple and effective manner, reducing the need for extensive refinement and achieving both improved performance and efficiency." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Point tracking. PIPs [19] independently tracks a point by fetching a local correlation around the point estimate and gradually refining the tracking result through iteration. TAP-Net [12] utilizes a shallow TSM-ResNet [32] for point tracking, with lightweight layers added to the backbone. TAPIR [13] integrates TAP-Net with the iterative refinement from PIPs, modifying the architecture to a convolution-based model for temporal processing. CoTracker tracks multiple points simultaneously, modeling" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1963" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "their dependencies with a Transformer architecture. LocoTrack [11] achieves improved correspondence using local 4D correlation, inspired by dense matching literature [9]. TAPTR [31] introduces the DETR-like [6] architecture for point tracking. Another line of work employs test-time optimization with regularization, such as tracking smoothness, geometric constraints, and cycle consistency [46, 49]. While these methods focus on designing better architectures for track refinement, the exploration of more effective feature backbones has been relatively underexplored. Our work focuses on developing an improved feature backbone." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 365 + ], + "type": "text", + "content": "Feature backbone in point tracking. Early works on point tracking [12] explored the performance of self-supervised feature backbones, such as VFS [51], as well as temporally-aware features like TSM-ResNet [32]. However, the exploration of feature backbones has been relatively underdeveloped compared to iterative track refinement [10, 11, 13, 25]. DINO-Tracker [46] explored the use of DINO in point tracking but it requires an hour of optimization for each video. Recently, [2] investigated the tracking capabilities of foundational models, such as Stable Diffusion [41] and DINOv2 [37], though the potential to extend these models for temporal applications remains unexplored." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 369, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 513 + ], + "type": "text", + "content": "Adapting large feature backbone. With recent advances in large-scale self-supervised training [22, 37], effectively transferring its vast knowledge has become a key challenge. Specifically, the strong features of DINOv2 have been adapted to various tasks. [43] fine-tunes DINO for video tasks in a self-supervised manner. Additionally, DINO has been applied to various correspondence tasks [1, 17, 30, 35, 44] and has demonstrated robustness in establishing correspondence and semantic segmentation [1, 18, 36], showing its rich semantics. We focus on adapting this representation to the point tracking task, enhancing its temporal consistency while preserving its pre-trained knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 523, + 111, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 111, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 111, + 536 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": "In this section, we introduce our temporally aware feature backbone. We begin by formally defining the point tracking problem and its associated challenges. We then describe the design of our temporal adapter and its integration with a pre-trained backbone to incorporate temporal awareness. Finally, we present our approach for adapting the pretrained backbone and training the temporal adapter for point tracking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 216, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 216, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 216, + 658 + ], + "type": "text", + "content": "3.1. Preliminaries and Motivation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Task definition. Point tracking [10-14, 25, 31] in videos involves establishing correspondences of specific points across frames, determining their positions over time, and identifying whether they are visible or occluded. Formally," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": "given a video sequence " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_t\\}_{t = 0}^{T - 1}" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": "-th frame of height " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": ", and a query point " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "q = (x_q, y_q, t_q) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "t_q" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": ", the goal is to produce a trajectory " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\hat{p}_t\\}_{t = 0}^{T - 1}" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\hat{p}_t \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": " denotes the estimated position of the point at time " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": ", and an associated visibility probability " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_t\\}_{t = 0}^{T - 1}" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "v_t \\in [0,1]" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": " indicates the likelihood of the point being visible at frame " + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 71, + 553, + 192 + ], + "type": "text", + "content": ". In this work, we focus on predicting accurate positions rather than occlusions, as our primary goal is to achieve better point correspondences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 199, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 353 + ], + "type": "text", + "content": "Motivation. Accurate point tracking in videos requires robust point matching across frames, a challenging task due to complex motions, occlusions, deformations, and scale variations [11, 25, 49]. Because videos consist of sequences over time, considering both spatial and temporal aspects is essential to understand point dynamics. More specifically, a good feature for point tracking must satisfy two key criteria. Spatially, it should effectively model complex real-world data with powerful feature representations vital for robust matching, the cornerstone of point tracking [11]. Temporally, it must capture complex motions and understand the dynamics of points over time to handle their movement across frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 355, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 487 + ], + "type": "text", + "content": "To fulfill the spatial criterion, we utilize DINOv2 [37], a self-supervised model trained on large-scale real-world data, renowned for its robust feature representations across various tasks [5, 30, 37]. However, DINOv2 lacks inherent temporal awareness, which is essential for point tracking in videos. To address this limitation, we introduce a temporal adapter that supplements DINOv2's features with temporal information. This adapter integrates temporal understanding directly into the feature extraction process, creating a feature backbone tailored for point tracking that combines strong spatial representations with temporal dynamics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 424, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 424, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 424, + 508 + ], + "type": "text", + "content": "3.2. Temporal Adapter" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 514, + 553, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 553, + 597 + ], + "type": "text", + "content": "Temporal awareness is essential for accurately tracking points across frames, as it enables the model to capture motion patterns and temporal dependencies. To embed this capability within DINOv2, we design a temporal adapter that allows the backbone to incorporate information from adjacent frames, enhancing its temporal sensitivity. Figure 3 illustrates the overall architecture of Chrono." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 676 + ], + "type": "text", + "content": "Design of the temporal adapter. Our temporal adapter is placed between each transformer block of DINOv2 [37] to enhance temporal modeling across multiple feature levels. By computing adjacent features from other frames at once, the adapter connects features from different time steps, enabling it to recognize motion across frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Each adapter follows a bottleneck structure inspired by the ResNet [20] architecture. It begins with a 2D convolutional layer Conv2D" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "^{\\text{down}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " with stride " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " that reduces the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1964" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 376 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 376 + ], + "type": "image", + "image_path": "d6a3b9ea0b55e61a19776ae4b3754c0051a6447a21a3533b136c82fa30583377.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 383, + 295, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 383, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 295, + 417 + ], + "type": "text", + "content": "Figure 3. Overall architecture of Chrono. Temporal adapters between transformer blocks use 2D convolution and 1D local attention along the temporal axis to output temporally-aware features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "text", + "content": "spatial dimensions from " + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "inline_equation", + "content": "H_{p} \\times W_{p} \\times C_{\\mathrm{in}}" + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "text", + "content": " to a compact " + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "inline_equation", + "content": "H_{p}^{\\prime} \\times W_{p}^{\\prime} \\times C_{\\mathrm{out}}" + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "inline_equation", + "content": "H_{p}^{\\prime} = H_{p} / s" + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "inline_equation", + "content": "W_{p}^{\\prime} = W_{p} / s" + }, + { + "bbox": [ + 55, + 425, + 296, + 485 + ], + "type": "text", + "content": ". This spatial downsampling step optimizes the representation for efficient processing while expanding the spatial receptive field for subsequent operations. This is formally defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 494, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 494, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 121, + 494, + 295, + 510 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t} ^ {\\text {d o w n}} = \\operatorname {C o n v 2 D} ^ {\\text {d o w n}} \\left(\\mathbf {f} _ {t} ^ {\\text {i n}}\\right), \\tag {1}", + "image_path": "b010e579bdcd326a2c2465353d435c84bb83a189b8c4e6b6ff95ea10407922d4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{\\mathrm{in}}\\in \\mathbb{R}^{H_p\\times W_p\\times C_{\\mathrm{in}}}" + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{\\mathrm{out}}\\in \\mathbb{R}^{H_p'\\times W_p'\\times C_{\\mathrm{out}}}" + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "text", + "content": " are the input and output feature maps of Conv2D" + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "inline_equation", + "content": "^{\\mathrm{down}}" + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 519, + 295, + 556 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "After downsampling, we apply a temporal attention layer that captures dependencies within a local temporal window of size " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", focusing on temporal correlations across time [3, 42]. The local window attention is carefully calibrated to balance computational efficiency with effective motion capture, ensuring a broad enough temporal window without excessive resource demands. This layer operates over the neighboring frames within the window " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "[t - k, t + k]" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "k = \\frac{N - 1}{2}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ". For each spatial location " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", we compute attention weights based on the similarity between the query vector at time " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and key vectors from neighboring times. The aggregated feature at time " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and location " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", referred to as " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{\\mathrm{att}}(x,y)" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", is computed as a weighted sum of" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 170 + ], + "type": "image", + "image_path": "e87dfa2071da79a9c3801c5383b65643a682d562e60656cacc48098952c367ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 178, + 555, + 211 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 211 + ], + "type": "text", + "content": "Figure 4. Point track prediction. To predict point positions, we simply match the query points with features from other frames, without using any learnable layers." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 220, + 513, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 513, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 513, + 232 + ], + "type": "text", + "content": "the value vectors from these neighboring frames:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 239, + 554, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 239, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 343, + 239, + 554, + 273 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t} ^ {\\mathrm {a t t n}} (x, y) = \\sum_ {n = - k} ^ {k} \\alpha^ {(t, n)} (x, y) \\cdot \\mathbf {V} _ {t + n} (x, y), \\tag {2}", + "image_path": "385ecb342b764312c4c48457f61167eb6e47077548cf80e8214401ef3a789d36.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "inline_equation", + "content": "\\alpha^{(t,n)}(x,y)" + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "content": " represents the attention weights for time offset " + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "content": ", computed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 559, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 559, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 559, + 353 + ], + "type": "interline_equation", + "content": "\\alpha^ {(t, n)} (x, y) = \\frac {\\exp \\left(\\mathbf {Q} _ {t} (x , y) \\cdot \\mathbf {K} _ {t + n} (x , y) / \\sqrt {d _ {k}}\\right)}{\\sum_ {n ^ {\\prime} = - k} ^ {k} \\exp \\left(\\mathbf {Q} _ {t} (x , y) \\cdot \\mathbf {K} _ {t + n ^ {\\prime}} (x , y) / \\sqrt {d _ {k}}\\right)}, \\tag {3}", + "image_path": "4ca3283adb13a833d1d928ce9e5e5c336b4f70b527893b524bd7d77a38d09edf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": " is a scaling factor [47], " + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q},\\mathbf{K}" + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": " are projections of " + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathrm{down}}" + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": ", obtained through the linear projection layers " + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q,\\mathbf{W}_K" + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V" + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 415 + ], + "type": "text", + "content": "After temporal attention, another 2D convolutional layer restores the spatial dimensions back to " + }, + { + "bbox": [ + 313, + 389, + 554, + 415 + ], + "type": "inline_equation", + "content": "H_{p} \\times W_{p} \\times C_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 389, + 554, + 415 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 420, + 553, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 420, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 386, + 420, + 553, + 434 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t} ^ {\\mathrm {u p}} = \\operatorname {C o n v 2 D} ^ {\\mathrm {u p}} \\left(\\mathbf {f} _ {t} ^ {\\mathrm {a t t n}}\\right), \\tag {4}", + "image_path": "9e7ddeaaa8aa0af7b8b726d5c46896ef4a23a0de0749b5f65763eca172346518.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 440, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 440, + 554, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{\\mathrm{up}}\\in \\mathbb{R}^{H_p\\times W_p\\times C_{\\mathrm{in}}}" + }, + { + "bbox": [ + 313, + 440, + 554, + 478 + ], + "type": "text", + "content": ". Finally, a residual connection adds the original input feature map to the output of the adapter to preserve the feature representation of DINOv2:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 399, + 484, + 553, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 484, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 399, + 484, + 553, + 499 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t} ^ {\\text {o u t}} = \\mathbf {f} _ {t} ^ {\\text {u p}} + \\mathbf {f} _ {t} ^ {\\text {i n}}. \\tag {5}", + "image_path": "bec55a66577c2b760c8d413377799f87bc6b4807cd83473a4e86da75ec5749a8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 506, + 541, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 541, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 541, + 517 + ], + "type": "text", + "content": "3.3. Point Prediction with the Feature Backbone" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 524, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 554, + 572 + ], + "type": "text", + "content": "As shown in Figure 4, we track a query point by simply performing feature matching using features from Chrono. This approach avoids the need for learnable modules such as iterative refinement, as used in [10, 11, 13, 25]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": "Correlation construction. Given the query point " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "q = (x_q, y_q, t_q)" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": ", we extract the query feature " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_q" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": " from the feature map of frame " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "t_q" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": " at position " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "(x_q, y_q)" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": " using bilinear interpolation. For each frame " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": ", we compute the correlation map " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": " by calculating the cosine similarity between " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_q" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": " and the feature map " + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 313, + 575, + 554, + 648 + ], + "type": "text", + "content": " at every spatial location:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 377, + 654, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 654, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 377, + 654, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {t} (x, y) = \\frac {\\mathbf {f} _ {q} ^ {\\top} \\mathbf {f} _ {t} (x , y)}{\\| \\mathbf {f} _ {q} \\| \\| \\mathbf {f} _ {t} (x , y) \\|}, \\tag {6}", + "image_path": "0230d95a99826b755699132b2f12b41f483d3ac461d07b034972cab32ce6c8a6.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t(x,y)" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " is the feature vector at position " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_q^\\top \\mathbf{f}_t(x,y)" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " denotes the dot product between " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_q" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1965" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t(x,y)" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{f}_q\\|" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{f}_t(x,y)\\|" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " indicate the Euclidean norms of " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_q" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t(x,y)" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ", respectively. This correlation map " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " quantifies the similarity between the query feature and features across all spatial locations in each frame " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 124, + 295, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 295, + 183 + ], + "type": "text", + "content": "Point prediction. To estimate the position of points in frame " + }, + { + "bbox": [ + 55, + 124, + 295, + 183 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 124, + 295, + 183 + ], + "type": "text", + "content": ", we apply the soft-argmax operation [29] to the correlation map " + }, + { + "bbox": [ + 55, + 124, + 295, + 183 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 55, + 124, + 295, + 183 + ], + "type": "text", + "content": ". The soft-argmax computes a weighted average of all spatial positions, with weights given by the softmax of the correlation values:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 190, + 295, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 190, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 111, + 190, + 295, + 217 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {p}} _ {t} = \\sum_ {(x, y)} \\sigma \\left(C _ {t} (x, y)\\right) \\cdot (x, y), \\tag {7}", + "image_path": "d4bf5890b1d69d7d5e34fe6564d0ac1c4bec53c93876042b1490aec45172e5e4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 294, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 294, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 223, + 294, + 247 + ], + "type": "inline_equation", + "content": "\\sigma(C_t(x, y))" + }, + { + "bbox": [ + 55, + 223, + 294, + 247 + ], + "type": "text", + "content": " is the softmax over all spatial positions within frame " + }, + { + "bbox": [ + 55, + 223, + 294, + 247 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 223, + 294, + 247 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 251, + 295, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 251, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 83, + 251, + 295, + 280 + ], + "type": "interline_equation", + "content": "\\sigma \\left(C _ {t} (x, y)\\right) = \\frac {\\exp \\left(C _ {t} (x , y) \\cdot \\tau\\right)}{\\sum_ {\\left(x ^ {\\prime} , y ^ {\\prime}\\right)} \\exp \\left(C _ {t} \\left(x ^ {\\prime} , y ^ {\\prime}\\right) \\cdot \\tau\\right)}, \\tag {8}", + "image_path": "eea9be7b2e711d4d44a199808f94c503b2972d3bb2d7d5054fd7ab1aff80150d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 285, + 296, + 370 + ], + "type": "text", + "content": " is a softmax temperature. The soft argmax is nonlearnable and provides a differentiable method to estimate the point's position based on the correlation map. To enhance precision, we mask out positions more than " + }, + { + "bbox": [ + 55, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 285, + 296, + 370 + ], + "type": "text", + "content": " pixels away from the maximum correlation value, focusing the soft argmax computation on a local neighborhood and reducing the influence of irrelevant regions [29]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 373, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 422 + ], + "type": "text", + "content": "Training. To train Chrono, we use the Huber loss [24] to supervise the estimated positions. The Huber loss is chosen for its robustness to outliers compared to the squared error loss. At each time step " + }, + { + "bbox": [ + 55, + 373, + 296, + 422 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 373, + 296, + 422 + ], + "type": "text", + "content": ", the loss is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 428, + 302, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 302, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 302, + 470 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {H u b e r}} \\left(\\hat {p} _ {t}, p _ {t}\\right) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\left\\| \\hat {p} _ {t} - p _ {t} \\right\\| ^ {2}, & \\text {i f} \\| \\hat {p} _ {t} - p _ {t} \\| \\leq \\delta , \\\\ \\delta \\cdot \\left(\\left\\| \\hat {p} _ {t} - p _ {t} \\right\\| - \\frac {1}{2} \\delta\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}", + "image_path": "e7739d7a2985a9af573ec5c4dfc62915aab2277dc81b1343c4d513c468c4aafe.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "text", + "content": " is the ground truth position, and " + }, + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "text", + "content": " is a threshold parameter. For occluded points indicated by the ground truth visibility status " + }, + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 55, + 471, + 296, + 518 + ], + "type": "text", + "content": ", we exclude the loss computation, effectively masking out those time steps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 526, + 295, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 526, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 113, + 526, + 295, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\left(1 - v _ {t}\\right) \\cdot \\mathcal {L} _ {\\text {H u b e r}} \\left(\\hat {p} _ {t}, p _ {t}\\right). \\tag {10}", + "image_path": "e1f57e732cbb45837ed3e0b9ede1cd16c5089231a5a3407f02bb88bf368e3ea7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 546, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 582 + ], + "type": "text", + "content": "By minimizing this loss over all time steps and query points, we train the backbone to produce features for predicting accurate tracked positions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 591, + 137, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 137, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 137, + 605 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 611, + 188, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 188, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 188, + 624 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "We implement our method using PyTorch [38]. During training, we use AdamW optimizer [34] with a learning rate of " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ", a weight decay of " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ", and a batch size of 1 per GPU. All models are trained for 100,000 iterations on 4 A100 GPUs, employing a cosine learning rate scheduler with warmup. For training, we use the Kubric Panning-MOVi-E dataset [11, 13] and sample 256 query points per" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "batch. The hyperparameters are set as follows: the softmax temperature is " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\tau = 20.0" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ", the soft argmax pixel threshold is " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "M = 5" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " and the local temporal window size is " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "N = 13" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ", and the 2D convolution stride in the temporal adapter is " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "s = 4" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 126, + 430, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 126, + 430, + 138 + ], + "spans": [ + { + "bbox": [ + 314, + 126, + 430, + 138 + ], + "type": "text", + "content": "4.2. Evaluation Protocol" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 145, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 312 + ], + "type": "text", + "content": "Evaluation datasets. To assess our approach, we employ the TAP-Vid benchmark [12], which comprises both real and synthetic video sets. The real videos come with accurate annotation tracks, while the synthetic videos are paired with perfect ground-truth trajectories. The benchmark consists of three datasets: TAP-Vid-RGB-Stacking, TAP-Vid-Kinetics, and TAP-Vid-DAVIS. TAP-Vid-RGB-Stacking [28] contains synthetic videos of a robot gripper stacking objects and includes 50 videos. The TAP-Vid-Kinetics [27] consists of 1,189 YouTube videos with diverse difficulties, including intense motion blur and abrupt scene transitions, while the TAP-Vid-DAVIS dataset [39] provides 30 videos featuring challenges such as substantial scale shifts of dynamic object." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": "Evaluation metrics. We assess the accuracy of the predicted tracks using two metrics: position accuracy at various thresholds and average position accuracy " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "(< \\delta_{\\mathrm{avg}}^x)" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": ". Position accuracy is evaluated at five threshold levels: " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "< \\delta^0" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "< \\delta^1" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "< \\delta^2" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "< \\delta^3" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "< \\delta^4" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": ", corresponding to accuracies within pixel distances of 1, 2, 4, 8, and 16, respectively. Each " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "< \\delta^x" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": " score represents the percentage of visible ground-truth points whose predicted positions fall within the specified threshold. The " + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "inline_equation", + "content": "< \\delta_{\\mathrm{avg}}^x" + }, + { + "bbox": [ + 313, + 316, + 554, + 437 + ], + "type": "text", + "content": " score is the average across all thresholds." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "content": "Following [12], we evaluate the datasets using two modes: strided query mode and first query mode. In strided query mode, query points are sampled along the groundtruth trajectory at intervals of 5 frames. In first query mode, query points are sampled from the first visible frame." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 503, + 400, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 400, + 515 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 400, + 515 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": "Quantitative comparison. We evaluate our method against the backbones [20, 32] commonly used in point tracking [11-13, 25, 26] and DINOv2 [37] in both the strided query (Table 1) and the first query modes (Table 2). Specifically, we compare with ResNet-18 [20] from TAPIR [13], TSM-ResNet-18 [32] from TAP-Net [12], and the backbone from the pre-trained CoTracker model [25], all of which are trained for point tracking. We report position accuracy to demonstrate the effectiveness of temporal information in our model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "Our small model, Chrono (ViT-S/14) with DINOv2 (ViTs/14), achieves state-of-the-art performance in position accuracy across most thresholds, outperforming other backbones on " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "< \\delta_{\\mathrm{avg}}^x" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": ". Chrono (ViT-B/14) with DINOv2 (ViTB/14) delivers even better results. Across RGB-Stacking, Kinetics and DAVIS datasets, in the strided and first query" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1966" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 71, + 240, + 134 + ], + "blocks": [ + { + "bbox": [ + 130, + 71, + 240, + 134 + ], + "lines": [ + { + "bbox": [ + 130, + 71, + 240, + 134 + ], + "spans": [ + { + "bbox": [ + 130, + 71, + 240, + 134 + ], + "type": "image", + "image_path": "758f437a94c7e47de91ed06044205f5c6456706dad8f2b4c6a7dadd597473707.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 134, + 216, + 141 + ], + "lines": [ + { + "bbox": [ + 154, + 134, + 216, + 141 + ], + "spans": [ + { + "bbox": [ + 154, + 134, + 216, + 141 + ], + "type": "text", + "content": "Reference frame " + }, + { + "bbox": [ + 154, + 134, + 216, + 141 + ], + "type": "inline_equation", + "content": "(t = 0)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 144, + 305, + 345 + ], + "blocks": [ + { + "bbox": [ + 57, + 144, + 305, + 345 + ], + "lines": [ + { + "bbox": [ + 57, + 144, + 305, + 345 + ], + "spans": [ + { + "bbox": [ + 57, + 144, + 305, + 345 + ], + "type": "image", + "image_path": "dbc3950ad79c04f1f1bcb987b020c236d8a0e1114f48e20a0eebb41a7a149f0a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 383 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of complex real-world video tracking. We qualitatively compare the results generated by Chrono with those from other commonly used backbones in point tracking and as well as DINOv2. Our model demonstrates better smoothness and precision than other competitors." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 486, + 134 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 486, + 134 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 486, + 134 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 486, + 134 + ], + "type": "image", + "image_path": "2e5e5573f9b7c0868138d27a6a33b152b823c908cb453c9dbb885779a11389c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 133, + 462, + 140 + ], + "lines": [ + { + "bbox": [ + 401, + 133, + 462, + 140 + ], + "spans": [ + { + "bbox": [ + 401, + 133, + 462, + 140 + ], + "type": "text", + "content": "Reference frame " + }, + { + "bbox": [ + 401, + 133, + 462, + 140 + ], + "type": "inline_equation", + "content": "(t = 0)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 144, + 552, + 345 + ], + "blocks": [ + { + "bbox": [ + 310, + 144, + 552, + 345 + ], + "lines": [ + { + "bbox": [ + 310, + 144, + 552, + 345 + ], + "spans": [ + { + "bbox": [ + 310, + 144, + 552, + 345 + ], + "type": "image", + "image_path": "e2f90cf1178d19ef93d654e3f9df62fc1c18d26e7fba46695a1284af62f7bd97.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 57, + 384, + 553, + 471 + ], + "blocks": [ + { + "bbox": [ + 57, + 384, + 553, + 471 + ], + "lines": [ + { + "bbox": [ + 57, + 384, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 57, + 384, + 553, + 471 + ], + "type": "table", + "html": "
BackboneRGB-Stacking-StridedKinetics-StridedDAVIS-Strided
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]36.664.682.790.994.773.910.535.365.781.388.656.39.731.160.678.387.053.3
TSM-ResNet-18 [12, 32]18.155.583.789.892.567.99.1533.264.679.286.554.58.226.853.673.583.949.2
CoTraker [25] Backbone36.960.174.982.587.668.431.031.050.570.674.958.233.752.563.468.372.058.0
DINOv2 (ViT-S/14) [37]3.213.644.977.487.645.34.915.841.973.886.044.57.322.752.980.089.550.4
DINOv2 (ViT-B/14) [37]4.215.246.678.988.146.65.918.645.975.786.946.610.028.659.582.987.454.4
Chrono (ViT-S/14)65.581.788.892.393.784.332.255.273.984.088.766.829.756.476.786.890.968.0
Chrono (ViT-B/14)68.384.090.392.994.586.033.557.275.885.890.268.531.859.278.888.492.270.1
", + "image_path": "23fd4d8548e9f1e23f8da2676d2a74846b74ed1866419c7708277ee9e1739f12.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 475, + 555, + 497 + ], + "lines": [ + { + "bbox": [ + 55, + 475, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 555, + 497 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on the TAP-Vid datasets [12] with the strided query mode. Best scores are in bold and second best are underlined for each column." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "text", + "content": "modes, Chrono (ViT-S/14) and Chrono (ViT-B/14) consistently outperform competitors in PCK accuracy at most thresholds. The high accuracy at " + }, + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "inline_equation", + "content": "< \\delta^0" + }, + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "text", + "content": " underscores the precision of Chrono features in point tracking, with further gains at higher thresholds. At " + }, + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "inline_equation", + "content": "< \\delta^4" + }, + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "text", + "content": ", only Chrono backbones achieve over " + }, + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 502, + 295, + 623 + ], + "type": "text", + "content": " accuracy on the DAVIS dataset in strided mode. Both Chrono (ViT-S/14) and Chrono (ViT-B/14) reach the highest position accuracy among backbones on average, demonstrating Chrono's ability to effectively model spatial and temporal information in videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Qualitative comparison. We visualize the estimated tracks from the DAVIS dataset, with a qualitative comparison shown in Figure 5. Unlike other methods, which produce highly jittery and inconsistent tracks over time, Chrono generates temporally smooth and accurate tracks. This jitter in other methods is expected as they lack awareness of neighboring frames. In contrast, Chrono utilizes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 503, + 555, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 539 + ], + "type": "text", + "content": "temporal adapters to maintain frame-to-frame consistency, allowing it to produce smooth tracks even when individual frame estimates are less stable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 550, + 471, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 471, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 471, + 563 + ], + "type": "text", + "content": "4.4. Analysis and Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Comparison to two-stage point tracking models with iterative refiner. As shown in Table 3, Chrono, although solely a feature backbone without a refiner, achieves point estimation precision comparable to full pipelines with refiners, while demonstrating significantly higher efficiency in terms of throughput. Unlike other models, which require refiners to inject temporal information for each query point, Chrono embeds temporal awareness directly in the feature backbone. This design substantially boosts throughput: for example, Chrono (ViT-B/14) achieves " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "12.5 \\times" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " the speed of TAPIR with only a " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " drop in DAVIS accuracy, and even surpasses TAPIR in RGB-Stacking by " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " p. Simi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1967" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 157 + ], + "type": "table", + "html": "
BackboneRGB-Stacking-FirstKinetics-FirstDAVIS-First
<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg<δ0<δ1<δ2<δ3<δ4<δavg
ResNet-18 [13, 20]31.458.377.887.792.469.68.628.856.574.283.350.39.027.354.973.784.149.8
TSM-ResNet-18 [12, 32]19.251.778.786.890.565.47.828.155.271.480.248.67.323.146.766.679.244.6
CoTraker [25] Backbone32.054.870.278.284.063.924.141.955.162.267.050.127.846.458.664.168.553.1
DINOv2 (ViT-S/14) [37]3.011.838.972.284.142.04.213.636.365.979.239.86.019.947.573.784.546.3
DINOv2 (ViT-B/14) [37]3.613.241.174.184.843.45.116.040.067.980.441.98.924.753.877.087.150.3
Chrono (ViT-S/14)58.277.285.489.191.380.224.846.265.877.583.159.524.049.271.282.887.963.0
Chrono (ViT-B/14)61.879.686.990.292.282.126.048.468.279.885.361.626.152.674.584.990.065.6
", + "image_path": "5215449ac9ec1eb0393585d1c47c54b821932882105054b505f6230c835935a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 191, + 294, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 160, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 555, + 182 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on the TAP-Vid datasets [12] with the query-first mode. Best scores are in bold and second best are underlined for each column." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 191, + 294, + 262 + ], + "lines": [ + { + "bbox": [ + 57, + 191, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 191, + 294, + 262 + ], + "type": "table", + "html": "
MethodRGB-Stacking-SDAVIS-SThroughput (points/sec)# of Refiner Params.Additional Params.
RAFT [45]58.646.323,405.714.2M4.2M
TAP-Net [12]72.853.129,535.985.5K5.5K
TAPIR [13]74.673.62,097.3225.9M25.9M
PIPs [19]51.059.446.4326.0M26.0M
Chrono (ViT-S/14)84.368.034,396.300M16.2M
Chrono (ViT-B/14)86.070.126,139.860M26.0M
", + "image_path": "721b5d0b5a513878f39660154b05f262d0bfc763180b998ce7c96a321ad5c6bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 334, + 294, + 426 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 331 + ], + "type": "text", + "content": "Table 3. Comparison on point tracking pipelines using refiners. Chrono achieves competitive or superior accuracy while offering significantly higher throughput, despite not using a heavy iterative refiner. TAP-Vid-RGB-Stacking and TAP-Vid-DAVIS metrics report " + }, + { + "bbox": [ + 55, + 266, + 295, + 331 + ], + "type": "inline_equation", + "content": "< \\delta_{avg}^{x}" + }, + { + "bbox": [ + 55, + 266, + 295, + 331 + ], + "type": "text", + "content": ". Throughput is measured on a 24-frame video using a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 334, + 294, + 426 + ], + "lines": [ + { + "bbox": [ + 56, + 334, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 334, + 294, + 426 + ], + "type": "table", + "html": "
MethodRGB-StackingKineticsDAVIS
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
Kubric-VFS-Like [16]57.972.691.940.559.080.033.148.579.4
TAP-Net [12]59.972.890.446.660.985.038.453.182.3
PIPs [19]37.351.091.635.354.877.442.059.482.1
RAFT [45]44.058.690.434.552.579.730.046.379.6
TAPIR [13]62.774.691.657.270.187.861.373.688.8
FlowTrack [10]------66.079.887.2
CoTracker [25]------65.979.489.9
LocoTrack [11]77.186.993.259.573.088.567.879.689.9
Chrono + LocoTrack83.291.094.260.773.789.568.280.291.2
", + "image_path": "dd1bc3a6e9c89324f3cdc939f99332dbb5419bb3ebc00ccf8d42732dd79751d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 551 + ], + "type": "text", + "content": "larly, Chrono (ViT-S/14) delivers " + }, + { + "bbox": [ + 55, + 491, + 295, + 551 + ], + "type": "inline_equation", + "content": "16.4 \\times" + }, + { + "bbox": [ + 55, + 491, + 295, + 551 + ], + "type": "text", + "content": " higher throughput. These results demonstrate that Chrono provides a simple yet highly effective solution, offering performance comparable to refiner-based models without requiring iterative processing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": "Analysis on Chrono as a backbone in existing point tracking pipeline. In Table 4, we present an analysis of the impact of adding a refiner after our model to investigate its ability as a feature backbone within a point tracking pipeline that uses iterative refinement [11, 13, 19, 25]. To assess the ability to predict precise occlusion status, we use occlusion accuracy (OA) and the Average Jaccard (AJ) metric [12], which measures both position and occlusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "For our experiment, we incorporate the iterative refiner from LocoTrack [11]. Since LocoTrack utilizes three levels of hierarchical features from ResNet, we apply a simple convolutional upsampler to our single-resolution feature map. Specifically, using transposed convolution [52], we" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 192, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 55, + 429, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 483 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of Chrono, adapted to LocoTrack [11], on the TAP-Vid dataset with the strided query mode. Our model shows a performance boost over LocoTrack on all datasets, with a particularly significant improvement on RGB-Stacking." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 192, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 317, + 192, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 192, + 553, + 261 + ], + "type": "table", + "html": "
MethodDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
1D Conv.26.553.174.585.490.165.9
3D Conv.27.554.374.985.590.066.4
1D Attn. (Ours)29.756.476.786.890.968.0
", + "image_path": "708a576f115c05b86e1080c2d1789befdc00665858a98fe8d78e1bfd7778692d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 298, + 553, + 364 + ], + "blocks": [ + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "type": "text", + "content": "Table 5. Ablation on temporal aggregation design choice. 1D attention demonstrates the best performance compared to both 1D and 3D convolution layers." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 298, + 553, + 364 + ], + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 364 + ], + "type": "table", + "html": "
Configuration# of \nAdaptersDAVIS
<δ0<δ1<δ2<δ3<δ4<δxavg
Early Blocks623.147.869.881.486.661.7
Later Blocks626.652.974.385.390.165.8
Alternating Blocks626.953.474.385.390.065.9
All Blocks1129.756.476.786.890.968.0
", + "image_path": "f3c36f51df653c47060201359747574ad41813d483d144b9f3a62ca467491749.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 366, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 411 + ], + "type": "text", + "content": "Table 6. Ablation on the number and position of temporal adapter. Rather than placing the temporal adapter into early or later blocks, or alternating blocks in DINOv2, applying it to all layers significantly boosts performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "text", + "content": "generate a " + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "text", + "content": " feature map with a " + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "text", + "content": " channel reduction and then a " + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "text", + "content": " feature map with a further " + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 422, + 554, + 481 + ], + "type": "text", + "content": " channel reduction. We freeze our model and train the convolutional upsampler and the LocoTrack refiner on the Kubric [16] panning MOVi-E dataset [13]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 482, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 578 + ], + "type": "text", + "content": "Our model combined with LocoTrack surpasses the original LocoTrack and also surpasses the performance of recent state-of-the-art point trackers [10, 25]. Our model shows a boost on all datasets, and notably, it shows a huge boost in the RGB-Stacking dataset, achieving a " + }, + { + "bbox": [ + 313, + 482, + 554, + 578 + ], + "type": "inline_equation", + "content": "+6.1" + }, + { + "bbox": [ + 313, + 482, + 554, + 578 + ], + "type": "text", + "content": " improvement in the AJ score. This result demonstrates the effectiveness of our approach when adopted into an existing point tracking pipeline." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Ablation on temporal aggregation design choice. To model temporal information in our temporal adapter, we tested two approaches for temporal aggregation: convolution and attention. In Table 5, we present an ablation study where only the temporal aggregation layer within Chrono is modified, comparing 1D convolution, 3D convolution, and 1D attention, with all other components kept constant. While 1D convolution and 1D attention aggregate along the temporal axis only, 3D convolution considers both spatial and temporal dimensions. Due to computational constraints, this study is conducted on Chrono (ViT-S/14)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1968" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 553, + 251 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 553, + 251 + ], + "type": "image", + "image_path": "af6dc6c3a6c1029debf1616cdddc83394e3d1acd82f9aa64f3cd9afa7bbe18ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 553, + 275 + ], + "type": "text", + "content": "Figure 6. Visualization of features. We visualize the features generated by our model and DINOv2 using PCA. The results show that our model demonstrates improved temporal smoothness and finer-grained feature representation relative to DINOv2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 389 + ], + "type": "text", + "content": "The results suggest that 1D attention is effective for temporal aggregation, as it allows the model to capture motion dynamics by adapting to correlations between frames. This adaptability may contribute to improved tracking accuracy by enabling the model to weigh information from different frames and handle varied motion patterns and occlusions. In contrast, temporal convolutions, with their fixed weights, may be less effective in capturing these complex temporal relationships." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 467 + ], + "type": "text", + "content": "Ablation on the number and position of temporal adapters. In Table 6, we conduct an ablation study on the placement of temporal adapters in Chrono, which consists of 12 transformer blocks. We evaluate three configurations: placing six temporal adapters within the early six blocks, within the later six blocks, and in alternating blocks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 624 + ], + "type": "text", + "content": "Early transformer blocks often capture local details, while later blocks tend to focus on broader, global patterns and complex relationships [15, 40]. Therefore, it's plausible that adapters in the initial blocks may learn local motion cues, while those in later blocks might capture overall motion patterns. Placing adapters only in the early or later blocks might emphasize either local or global motion, respectively. While the alternating configuration shows some improvement, it may not be optimal. Integrating temporal adapters between each block could potentially allow for a more balanced capture of both local and global motion patterns, making features sensitive to multilevel motions and leading to better performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Computation and latency comparison. Table 7 compares our inference time and computation with that of DINOv2 [37], measuring the time and GPU memory taken to extract features for 24 frames. While our method exhibits comparable computation and slower inference due to temporal injection into the feature, Chrono achieves a significantly higher score than vanilla DINOv2. This higher accu" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 279, + 554, + 350 + ], + "blocks": [ + { + "bbox": [ + 317, + 279, + 554, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 279, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 279, + 554, + 350 + ], + "type": "table", + "html": "
Method<δxavgInference Time ↓ (sec)GPU Memory ↓ (MiB)
DINOv2 (ViT-S/14) [37]39.80.1941,460
DINOv2 (ViT-B/14) [37]41.90.4432,816
Chrono (ViT-S/14)68.00.5751,550
Chrono (ViT-B/14)70.11.1972,998
", + "image_path": "5a6cdad2c9931e202a8077c24c4958dc55616c7526ee90de3b3e80d5a59b10ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 354, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 388 + ], + "type": "text", + "content": "Table 7. Computation and latency comparison with DINOv2. We measure inference time and GPU memory usage for processing 24 frames without point prediction." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "text", + "content": "racy compensates for the increased time required for feature extraction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 427, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 554, + 475 + ], + "type": "text", + "content": "Visualization of feature with PCA. We visually compare the feature of Chrono and DINOv2 [37] with PCA (Principal Component Analysis) in Figure 6. We reduce the feature dimensions to three, represented as RGB in the figure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "type": "text", + "content": "While DINOv2's visualizations demonstrate a noisy background and tend to show a uniform representation within a single object, our model produces consistent and smooth representations over time and also shows fine granularity within the same semantic object. We believe the temporally smooth representation and fine-grained feature detail lead to better point tracking." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we presented Chrono, a temporally-aware feature backbone for point tracking that integrates pretrained DINOv2 representations with a temporal adapter, enabling long-term temporal context capture in feature space. Through extensive experiments, we demonstrate that embedding temporal information directly into the feature backbone, Chrono reduces reliance on costly refiners while achieving both accuracy and efficiency. We anticipate Chrono will inspire advancements in efficient, temporally-aware backbones for point tracking." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1969" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 206, + 115, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 115, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 115, + 219 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 226, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 226, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 226, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 296, + 258 + ], + "type": "text", + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 261, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 296, + 294 + ], + "type": "text", + "content": "[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 296, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 328 + ], + "type": "text", + "content": "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 331, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 374 + ], + "type": "text", + "content": "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 430 + ], + "type": "text", + "content": "[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 433, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 296, + 485 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "text", + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "text", + "content": "[9] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. Advances in Neural Information Processing Systems, 34:9011-9023, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268-19277, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 183 + ], + "type": "text", + "content": "Acknowledgements This research was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, RS-2024-00509279, RS-2025-II212068, RS-2023-00227592) and the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (RS-2024-00345025, RS-2024-00333068), and National Research Foundation of Korea (RS-2024-00346597)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 120, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 554, + 174 + ], + "type": "text", + "content": "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 177, + 554, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 554, + 241 + ], + "type": "text", + "content": "[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 244, + 554, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 554, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 554, + 298 + ], + "type": "text", + "content": "[14] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 301, + 554, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 301, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 301, + 554, + 334 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 336, + 554, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 336, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 336, + 554, + 402 + ], + "type": "text", + "content": "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 404, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 554, + 459 + ], + "type": "text", + "content": "[17] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 461, + 554, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 461, + 554, + 505 + ], + "spans": [ + { + "bbox": [ + 317, + 461, + 554, + 505 + ], + "type": "text", + "content": "[18] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 507, + 554, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 507, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 507, + 554, + 540 + ], + "type": "text", + "content": "[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 543, + 554, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 543, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 543, + 554, + 586 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 589, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 554, + 621 + ], + "type": "text", + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 624, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 677 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[23] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1970" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[24] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "text", + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[26] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 330 + ], + "type": "text", + "content": "[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In 5th Annual Conference on Robot Learning, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "text", + "content": "[29] Junghyup Lee, Dohyung Kim, Jean Ponce, and Bumsub Ham. Sfnet: Learning object-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2278-2287, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[30] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "text", + "content": "[31] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "type": "text", + "content": "[32] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 590 + ], + "type": "text", + "content": "[34] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 591, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 645 + ], + "type": "text", + "content": "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "text", + "content": "[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[40] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[42] Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers. Transactions of the Association for Computational Linguistics, 9:53-68, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[43] Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Time does tell: Self-supervised timetuning of dense image representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16536-16547, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "type": "text", + "content": "[44] Aleksandar Shtedritski, Andrea Vedaldi, and Christian Rupprecht. Learning universal semantic correspondences with no supervision and automatic data curation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 933-943, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "type": "text", + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "text", + "content": "[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision, pages 367-385. Springer, 2025. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 612, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 634 + ], + "type": "text", + "content": "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[48] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. arXiv preprint arXiv:2308.15975, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "text", + "content": "[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1971" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 353 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[50] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "type": "text", + "content": "[51] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 252 + ], + "type": "text", + "content": "[52] Matthew D Zeiler, Dilip Krishnan, Graham W Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on computer vision and pattern recognition, pages 2528-2535. IEEE, 2010. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "text", + "content": "[53] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 353 + ], + "type": "text", + "content": "[54] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1972" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_content_list.json b/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1670b2a6982e59434c56e0f406fd23a0afdca519 --- /dev/null +++ b/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_content_list.json @@ -0,0 +1,1363 @@ +[ + { + "type": "text", + "text": "Exploring Timeline Control for Facial Motion Generation", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifeng Ma $^{1}$ , Jinwei Qi $^{2}$ , Chaonan Ji $^{2}$ , Peng Zhang $^{2}$ , Bang Zhang $^{2}$ , Zhidong Deng $^{1}$ , Liefeng Bo $^{2}$ $^{1}$ Department of Computer Science and Technology, Tsinghua University \n $^{2}$ Tongyi Lab, Alibaba Group", + "bbox": [ + 106, + 178, + 888, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e50501a6f43d035fff80db1ff38bb2eed2754aee5a0ed9f16b35a1ad89799ed6.jpg", + "image_caption": [ + "Figure 1. We introduce a new control signal for facial motion generation: timeline control. We first utilize a labor-efficient approach to annotate the time intervals of facial motion at a frame-level granularity. Using the annotations, we propose a model that can generate natural facial motions aligned with an input timeline. Compared to previous controls like audio and text, timeline control enables precise temporal control of facial motions. In this paper, facial motions are rendered into photorealistic videos for better visualization." + ], + "image_footnote": [], + "bbox": [ + 93, + 282, + 486, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04eafa1a458ae0753af68bc385382fdf61400b2b1fc50d896007467810c822e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 282, + 901, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 553, + 326, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces a new control signal for facial motion generation: timeline control. Compared to audio and text signals, timelines provide more fine-grained control, such as generating specific facial motions with precise timing. Users can specify a multi-track timeline of facial actions arranged in temporal intervals, allowing precise control over the timing of each action. To model the timeline control capability, We first annotate the time intervals of facial actions in natural facial motion sequences at a frame-level granularity. This process is facilitated by Toeplitz Inverse Covariance-based Clustering to minimize human labor. Based on the annotations, we propose a diffusion-based generation model capable of generating facial motions that are natural and accurately aligned with input timelines. Our method supports text-guided motion generation by using ChatGPT to convert text into timelines. Experimental results show that our method can annotate facial action intervals with satisfactory accuracy, and produces natural facial motions accurately aligned with timelines.", + "bbox": [ + 88, + 595, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 553, + 643, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating vivid facial motions has drawn growing attention due to its broad applications, including digital human generation and filmmaking. To produce desired facial motions, current methods use audio or text to provide guidance. However, these methods lack a critical ability that users often require: generating specific facial motions with precise timing. For example, users may want to generate a brow raise between frames 10 and 30 while simultaneously generating a smile between frames 14 and 43. Audiodriven methods [43, 54] can only generate motions synchronized with the audio. Text signals offer only coarse-grained descriptions of facial motions and lack frame-level detail. Text-driven approaches [4, 44, 49, 51] rely on temporal adverbial cues (e.g., then) for rough timing guidance.", + "bbox": [ + 509, + 580, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve more fine-grained control, we introduce a novel control signal: timeline control for facial motion generation. In this setup, users can generate facial motions by inputting an intuitive timeline that contains several temporal intervals, each corresponding to a desired facial action. This setup enables users to manage the timing of each action.", + "bbox": [ + 511, + 792, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Achieving timeline control is very challenging. Frame", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1940", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "level control of facial motions requires the model to achieve exceptional precision in generating accurate actions. To model timeline control capability, it is crucial to annotate the precise start and end frames of facial actions, a challenge that existing methods have yet to overcome. Existing methods [44, 51] represent facial motion sequences as a time series of motion descriptors, such as Facial Action Units (AUs) [10] or blendshapes [46], and then leverage ChatGPT to summarize these time series for annotation. However, ChatGPT's limited sensitivity to the temporal dynamics of facial motions prevents it from determining the exact start and end frames of facial actions. Another annotation approach is using thresholds, like labeling the eye as \"closed\" if the blendshape eyeBlink exceeds 0.4. However, setting thresholds for complex actions, such as brow motions, is challenging. Additionally, some actions are determined by the relative values of multiple motion descriptors rather than a single one.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we adopt Toeplitz Inverse Covariance-based Clustering (TICC) [18] to annotate the start and end frames of facial actions. TICC can segment the time series into a sequence of intervals, with each interval containing a single motion pattern. Each interval has clearly defined start and end frames. Once the start and end frames of the intervals are identified, the remaining task is to determine the facial action represented by each interval. This can be addressed by another feature of TICC: its ability to group the segmented intervals into several clusters based on the similarity of their motion patterns, with each cluster containing intervals that exhibit similar motion patterns. The automatic clustering process eliminates the need for manual threshold setting and considers the relationships between multiple motion descriptors. Once these clusters are obtained, we can determine the overall action (e.g. significantly raised brows) represented by each cluster by analyzing a few representative patterns within it. This process requires minimal labor. Using this procedure, we analyze motions in different facial regions individually and achieve frame-level annotation for the brows, eyes, mouth, gaze, and head motions in a labor-efficient manner.", + "bbox": [ + 91, + 367, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using frame-level facial motion annotations, we propose a diffusion-based generation model that produces natural facial motions accurately aligned with the input timeline. The model generates motions for each facial region (e.g. upper face, lower face) separately. This decoupling can help address the adverse impact of motion coupling in the data, thereby improving accuracy. However, some couplings are necessary for motion naturalness, such as a slight head lift during a brow raise. To properly manage coupling, the generation model is divided into a base network and multiple branch networks, with each branch network dedicated to a facial region. The base network encodes global facial motion couplings into base features. The branch network takes", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "base features and the timeline of the corresponding region as input to accurately generate facial motions for that region while maintaining natural couplings. We use FaceVerse[46] 3DMM coefficients to represent facial motions for generation. Since photorealistic facial motion enables more accurate assessments of motion realism than mesh-based motion [31], we render motions into photorealistic videos by a diffusion-based renderer.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method enables text-guided facial motion generation by using ChatGPT to convert text into timelines. ChatGPT effectively translates simple natural language descriptions into timelines, which our model transforms into facial motions. Users can modify these timelines, allowing precise timing control for text-driven applications.", + "bbox": [ + 511, + 213, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 531, + 304, + 831, + 319 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to develop a labor-efficient approach to annotate temporal intervals of facial actions. Such fine-grained annotations enable precise modeling of the temporal dynamics of facial motions.", + "- We are the first to achieve timeline control for facial motion generation. This enables users to generate specific facial motions with precise timing. Our method also supports generating actions from natural language text.", + "- Detailed evaluations show that our method generates accurate motion annotations and produces natural, precise facial motions from timeline inputs." + ], + "bbox": [ + 513, + 324, + 903, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 510, + 653, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Facial Motion Annotation. Compared to human motion annotation [15, 16, 29, 36], facial motion annotation is in its early stages. Existing methods [28, 44, 49, 51, 60] generate annotations that either overlook temporal changes or describe them only roughly with temporal adverbs. Some datasets [38, 56] label intervals of macro/micro expressions rather than facial actions used for facial motion generation. These datasets also require intensive manual labeling. Facial expression spotting methods [9, 17, 48, 57, 59] are also restricted to label macro/micro-expressions. In this paper, we aim to labor-efficiently generate frame-level facial action annotations.", + "bbox": [ + 511, + 537, + 906, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Facial Motion Generation. Existing methods use audio or text as the control signal and cannot generate specific motions at precise timing. Audio-driven methods (2D [6, 14, 19, 23, 26, 32, 37, 42, 43, 47, 50, 53, 54, 58, 63, 64] and 3D methods [1, 7, 8, 11, 12, 24, 25, 34, 41, 52, 55, 65]) can only generate facial motions synchronized with the audio. Text-driven methods [4, 27, 28, 44, 49, 51] can only use temporal adverbial phrases to control timing roughly. Rule-based methods [5, 33] offer temporal control but often produce unnatural motions due to deviations from real movement distributions. This paper introduces timeline control to generate natural facial motions with precise timing.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1941", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05bf3c89485f9c1bd19fb9bdc8078a6f9b836256d59c72e733ac79156d9a2c24.jpg", + "image_caption": [ + "Figure 2. The pipeline of frame-level facial motion annotation (using brow motions as an example). We first extract facial motion descriptors (blendshapes) from natural facial motion videos and concatenate the results to create a facial motion time series for time series analysis. This analysis can simultaneously segment the sequence into a series of motion patterns and cluster similar patterns, resulting in multiple clusters, each containing consistent facial motion patterns. Then, by inspecting a few patterns, we identify the facial motions each cluster represents, thereby obtaining frame-level facial motion annotations for all videos." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/b704b1959310eac8003d5eb21974bc95cb431a975d8cbf5c80d9cfa907be5b5f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Facial RegionMotion Categories
BrowBrowUp, BrowDown, Neutral
EyeEyeSquint, EyeWiden, EyeClose, Neutral
MouthSoftSmile, Smile, MouthFrown, Neutral
GazeLeft, Right, Up, Down, Neutral
HeadLeft, Right, Up, Down, Neutral
", + "bbox": [ + 93, + 369, + 478, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Facial actions annotated by our method.", + "bbox": [ + 140, + 470, + 434, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Timeline control has been utilized in human motion generation [2, 3, 35, 39] by first generating motion segments and then piecing them together. However, due to the rapid changes in facial movements, this approach is unsuitable, necessitating a new model structure for effective generation.", + "bbox": [ + 89, + 497, + 483, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 585, + 183, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we first annotate the temporal intervals of facial action for natural facial motion sequences. Based on these annotations, we develop a generation model that generates natural facial motions that are accurately aligned with an input timeline. Our method also supports text-guided motion generation by leveraging ChatGPT to convert text into timelines.", + "bbox": [ + 89, + 612, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Annotating temporal intervals of facial actions", + "text_level": 1, + "bbox": [ + 89, + 727, + 483, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. For each video $V$ in dataset, our method aims to generate frame-level facial motion annotations $A = [a_i]_{i=1}^L$ for each video frame $[v_i]_{i=1}^L$ . Tab. 1 shows the annotated facial actions. The annotation $a_i$ for each frame is a vector of binary values, with each dimension representing a non-neutral facial action. A dimension is set to 1 if the action occurs, otherwise, it is 0.", + "bbox": [ + 89, + 750, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Annotation Process Overview. We achieve the annotation through time series analysis TICC. Fig. 2 shows the pipeline: (1) We extract proper motion descriptors for each", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "facial region to obtain facial motion time series. (2) We apply TICC to simultaneously segment the time series into a sequence of motion patterns and cluster similar patterns. This results in several clusters of intervals, each containing similar facial motions. (3) Finally, for each cluster, we can infer the actions of all intervals by examining only a few intervals. Once the action for each cluster is identified, the actions for all intervals are determined, resulting in facial action annotations at a frame-level granularity.", + "bbox": [ + 511, + 371, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Facial Motion Descriptor. We observe that previously used AUs lack sufficient precision, while ARKit blendshapes offer a more accurate representation. The publicly available MediaPipe blendshape detector struggles to detect certain blendshapes (e.g. eyeWide, mouthFrown), so we use an in-house blendshape detector. ARKit blendshapes are a set of coefficients that describe facial expressions, with each coefficient representing the movement of a specific facial region. Each coefficient ranges from 0 to 1, indicating the motion intensity.", + "bbox": [ + 511, + 507, + 908, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the blendshapes corresponding to each facial region to construct motion time series for each region, resulting in time series data for the eye, brow, and mouth areas. For the eye region, we selected eyeBlink, eyeSquint, eyeWide, and for the brow region, we selected browDown, browInnerUp, browOuterUp. For the mouth region, we use only expression-related blendshapes (mouthSmile, mouthStretch, mouthFrown) to avoid interference from speech-related movements. ARKit blendshapes provide coefficients for the left and right sides. As facial motions in our dataset are largely symmetrical, we use only the left-side coefficients to simplify analysis.", + "bbox": [ + 511, + 657, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For gaze and head pose motion, we use a 3DMM model FaceVerse [46] to detect coefficients. For gaze, we select eye coefficients, for head pose, we select angle coefficients.", + "bbox": [ + 511, + 839, + 908, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analyzing Facial Motion Time Series. We utilize", + "bbox": [ + 514, + 885, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1942", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b0f9f501d79aaaa308dd00981cfc6d9d4b2e064ecc9ccc9bf9b7d1de77400e9e.jpg", + "image_caption": [ + "(a) Generation Model's Base-Branch Design" + ], + "image_footnote": [], + "bbox": [ + 99, + 95, + 578, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/625412adf6fba8cbd446dae4e043a4f7619ddd2fae6281a06c61e3d7760582c9.jpg", + "image_caption": [ + "(b) Base/Branch Network's Architecture", + "Figure 3. Illustration of generation model. (a) Base-Branch Design. The base network takes the timelines of all facial regions as input and outputs base features that model the global facial motion couplings. Through timeline selection, each region's timeline is directed to its respective branch network. Since head pose is interconnected with all facial movements, the pose branch receives timelines of all regions. Each branch network takes the timeline of its corresponding region to generate the facial motions for that region. These motions are then combined to produce the overall motion of the entire face. Lin. Proj. denotes Linear Projection. (b) Base/Branch Network's Architecture. Timeline control guides motion generation through cross-attention. The initial timeline tokens remain unchanged and are added at each layer. For clarity, the diffusion step (omitted in sub-figure (a) for clarity) is applied to each base and branch network." + ], + "image_footnote": [], + "bbox": [ + 583, + 93, + 893, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TICC [18] for time series analysis. TICC is designed to handle a single long sequence (CubeMarker [22] can analyze multiple time series analysis but performs poorly.). However, our data consists of multiple short sequences. To address this issue, we concatenate all the short sequences into a single long sequence for analysis. To avoid the impact of abrupt transitions between different videos, we introduce a \"null sequence\" between videos to separate them. The null sequence is implemented as a length-100 sequence with values set to -1. These null sequences will eventually be grouped into a single cluster, which will not affect others.", + "bbox": [ + 89, + 378, + 482, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the time series analysis, the video is annotated with a sequence of motion pattern intervals and similar patterns are clustered into a few clusters. Therefore, for each cluster, we inspect a few intervals to determine the facial action they represent. This facial action is then considered to be the action for all intervals within that cluster. Different clusters may represent variations of the same facial action category, such as a high eyebrow raise versus a moderate one. We categorized the clusters into facial motion categories, such as eyebrow raise, frown, and neutral, to obtain frame-level annotations.", + "bbox": [ + 89, + 547, + 482, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that eye closure frequently produces sharp spikes in the time series, which disrupts the analysis of squinting and widening. Therefore, we analyze eye closure and squinting/widening separately. We also find that threshold-based annotation, which classifies actions using predefined thresholds (e.g., labeling the eye as \"closed\" if eyeBlink exceeds 0.4), achieves high accuracy for eye closure, gaze, and head pose. Consequently, we use this method for these regions, while applying time series analysis for others.", + "bbox": [ + 89, + 717, + 482, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detect Facial Motion for Unseen Videos. After learning various facial motion patterns from the data, TICC can also", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be applied to previously unseen facial motion sequences and detect facial action intervals. We use this function to assess the accuracy of facial motion generation in our evaluation.", + "bbox": [ + 511, + 378, + 906, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Facial Motion Generation from Timelines", + "text_level": 1, + "bbox": [ + 511, + 436, + 867, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Formulation. Given a timeline control $C = [c_i]_{i=1}^L$ for each video frame, the generation model $\\mathcal{G}$ aims to generate natural facial motions $M = [m_l]_{l=1}^L$ that align with the timeline control. The control for each frame $c_i$ follows the same format as the annotation. Generated facial motions are rendered by a diffusion-based renderer into photorealistic videos for better visualization.", + "bbox": [ + 511, + 460, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion: The coupling of facial motions and its impact on motion accuracy and naturalness. Facial motions across different regions are coupled. For example, when people smile, they might also squint and lower their eyebrows. These couplings can reduce the precision of generated motions. For instance, when generating a smile, the model may also learn to produce eyebrow-lowering actions, which could conflict with user-specified eyebrow motions (e.g. brow raise). To enhance accuracy, it is essential to decouple the generation of different facial regions. However, certain facial motion couplings are crucial for conveying naturalness—for instance, a subtle head movement that accompanies a shift in gaze. Therefore, the model must achieve a delicate balance, selectively decoupling motions to improve accuracy while preserving necessary couplings to enhance naturalness.", + "bbox": [ + 511, + 566, + 908, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation Model $\\mathcal{G}$ . $\\mathcal{G}$ is a diffusion-based model. Diffusion models consist of two Markov chains [21, 40]: the forward chain incrementally injects Gaussian noise into the original signal, while the reverse chain sequentially reconstructs the original signal from the noise. $\\mathcal{G}$ predicts the original signal instead of noise, and the loss function is in", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1943", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "produced as:", + "bbox": [ + 89, + 90, + 174, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e n o i s e}} = \\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ]}, M _ {(0)}, C (\\| M _ {(0)} - \\mathcal {G} (M _ {(t)}, t, C) \\| ^ {2}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 113, + 482, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ denotes the diffusion step, $M_{(0)}$ is the original facial motion sequence, and $M_{(t)}$ is the noisy sequence produced by the diffusion forward function $q(M_{(t)}|M_{(t - 1)}) = \\mathcal{N}(\\sqrt{\\alpha_n} M_{(t - 1)},(1 - \\alpha_n)\\pmb {I})$ . $\\pmb{C}$ is the timeline condition.", + "bbox": [ + 89, + 146, + 482, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To balance decoupling and coupling for both accuracy and naturalness, the generation model employs a base-branch design (Fig. 3). It consists of a base network and individual branch networks for each facial region (e.g. upper face, lower face). The base network takes timelines of all facial regions and noisy motions as input and encodes global motion couplings into base features. The branch network takes the base features the timelines relevant only to its designated facial region as inputs, and generates decoupled motions for each region.", + "bbox": [ + 88, + 207, + 482, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the expression, eye, and pose coefficients of FaceVerse [46] 3DMM model as target motion representation for generation. Its expression coefficients align with ARKit blendshapes, allowing separate representations of different facial regions. When dividing facial motions into different regions for separate generations, we observe that splitting them into the upper face, lower face, and pose & other regions strikes a better balance between decoupling and necessary coupling, rather than assigning a branch network for each individual region (e.g. further split eye and brow). Specifically, the upper face includes eye, brow, and gaze; the lower face includes mouth and jaw; and the pose & other regions cover the head pose and remaining regions like cheek and nose. It is important to note that, since pose is coupled with facial motions in all regions, the timeline control for the pose & other branch includes all facial regions (rather than just the control for pose).", + "bbox": [ + 89, + 357, + 482, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the specific implementation, the timeline control for each frame $c_{i}$ is implemented as a 16-dimensional vector composed of values 0 and 1, with each dimension representing an action. A value of 1 indicates that the action is performed in that frame, and a value of 0 indicates that the action is not present. The timeline control is transformed into timeline tokens through a linear projection. The base network and branch networks share the same structure. Their inputs consist of timeline tokens and motion tokens. The motion tokens are derived either from the noisy motion through linear projection or from base features. The motion tokens learn the temporal control from the timeline tokens through cross-attention in multiple transformer encoder layers. When incorporated into each encoder layer, the timeline tokens always use the initial timeline tokens rather than the output tokens from the previous layer. This prevents the temporal information in timeline tokens from being altered, thereby enhancing motion accuracy. The motion tokens for each layer come from the output of the motion tokens from", + "bbox": [ + 89, + 614, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the previous layer. The network's final output consists of the motion tokens from the last layer, which are then linearly projected to produce the facial motion. Each branch network generates the facial motion for its corresponding region. These motions are then combined to produce the motion for the entire face.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve flexibility and generalization, we use classifier-free guidance [20] to train our model. During training, we randomly drop the condition of each facial region timeline. We use a dropout probability of 0.5 for each condition independently, with a 0.1 probability of dropping all conditions and a 0.1 probability of maintaining all conditions. When the timeline condition for a certain region is dropped, its value is set to -1.", + "bbox": [ + 511, + 181, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering. We develop a diffusion-based portrait animation method to render motions as images. Since the renderer is not the main contribution of this paper, its details are reported in the Supplementary Material.", + "bbox": [ + 511, + 301, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Translating Natural Language to Timelines", + "text_level": 1, + "bbox": [ + 511, + 371, + 880, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that ChatGPT can generate a reasonable timeline from natural language descriptions, enabling our model to produce facial motions based on natural language input. The generated timeline can be further edited by users to fit their needs. To enhance the realism of the generated movements, we manually annotated several timeline descriptions as examples, allowing ChatGPT to perform few-shot learning. The details of the prompts and generated results are reported in Supplementary Material.", + "bbox": [ + 511, + 393, + 903, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 542, + 643, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. To effectively model natural facial motions, a dataset capturing authentic facial expressions is essential. For this purpose, we utilize the RealTalk dataset [13]. The RealTalk dataset contains 692 videos captured from genuine, unscripted conversations, resulting in highly natural facial movements with rich temporal dynamics. Previously used lab-recorded, scripted datasets, such as MEAD [45], lack these qualities. We use 1412 video clips extracted from Realtalk as the dataset. The size of the dataset is approximately 600,000 frames.", + "bbox": [ + 511, + 566, + 906, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Facial Motion Annotation", + "text_level": 1, + "bbox": [ + 511, + 727, + 746, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. No previous methods have attempted annotating intervals of facial actions in a labor-efficient manner. We experimented with time series analysis methods such as AutoPlait [30], CubeMarker [22], and TICC. However, AutoPlait and CubeMarker failed to produce satisfactory results, as the motions of the segments within each cluster were inconsistent, making annotation impossible. Therefore, we primarily relied on TICC. Empirical results [18] indicate TICC's annotation performance can be enhanced by tuning key hyperparameters, including the number of clusters and", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1944", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00442194334990496a745553a085f0fab248c496c8122189aee9b66f78234c69.jpg", + "image_caption": [ + "Figure 4. An example of brow motion annotation." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 477, + 231 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5397c53be53ed135df9f44b2aed5567accfc67d023ad8e6ec9be4459d5ec14b6.jpg", + "image_caption": [ + "Figure 5. An example of eye motion annotation." + ], + "image_footnote": [], + "bbox": [ + 96, + 260, + 468, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "beta, a hyperparameter that facilitates temporal consistency. We examine the effects of these parameters.", + "bbox": [ + 89, + 441, + 480, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. Annotating intervals of facial actions is essentially a multi-class classification problem, where each data point represents a frame and each class represents an action. To evaluate annotation accuracy, we manually annotate the facial actions on 50 videos to serve as ground truth, and use Macro-F1 as the metric for evaluation. For each class, the F1 score is the harmonic mean of the precision and recall of our estimate. Then, the macro-F1 score is the average of the F1 scores for all the classes. A higher Macro-F1 score indicates greater annotation accuracy. We calculated the macro-F1 score for the eye region only on segments when the eyes are not closed.", + "bbox": [ + 88, + 474, + 482, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 4 shows brow motion annotation. Our method can precisely segment different facial motion patterns and cluster similar ones. For a specific facial motion brow raise, our approach can distinguish multiple variants, such as a significant or moderate raise. In this work, these finer categories are grouped under one motion category for simplicity, but future methods could generate each subcategory separately. Annotation for the mouth is similar to that for the brow. Fig. 5 shows eye motion annotation. As stated in Sec. 3.1, only eye squint/widen are annotated using TICC. Eye closure is determined by a threshold on the blendshape eyeBlink and may occur within intervals of other eye motions. We also conduct a user study to evaluate annotation accuracy and the results are reported in Sec. 4.2.", + "bbox": [ + 89, + 657, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. We select the optimal hyperparameters based on the highest macro-F1 score. Experiments", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0096b72dc3303fa56a17645c5b5ab0e98b86411f6dedf63d24c5393b1c02493.jpg", + "image_caption": [ + "Figure 6. The annotation accuracy for different number of clusters." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 895, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a561be1bc336f3cb2533f87371732a9997c3e33bceb978b18043b3535ba1bc92.jpg", + "image_caption": [ + "Figure 7. The annotation accuracy in brows for different $\\beta$ , concatenating videos without separating, and using AU as descriptors." + ], + "image_footnote": [], + "bbox": [ + 517, + 247, + 895, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "show that the optimal beta value for different regions is 5, with the best number of clusters being 8 for the eye region, 9 for the brow region, and 9 for the mouth region. (The number of clusters here refers to the count of valid clusters, excluding the one that contain only null sequences). The optimal scores for brow, eyes, and mouth are 0.90, 0.91, and 0.87, respectively. When using threshold-based annotation, the scores for eye closure, pose, and gaze are 0.95, 0.87, and 0.89, respectively.", + "bbox": [ + 511, + 428, + 905, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The impact of different cluster numbers is shown in Fig. 6 (with beta set to 5). Too few clusters in TICC lead to under-segmentation, grouping distinct patterns into single clusters. Conversely, too many clusters decreases intercluster variability, making clusters less distinct and blurring boundaries.", + "bbox": [ + 511, + 565, + 905, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 7 illustrates the impact of $\\beta$ , using brow as an example (with the number of clusters set to 8). $\\beta$ controls the smoothness of segment transitions. Too large $\\beta$ leads to over-smoothing, Conversely, a small $\\beta$ makes the model overly sensitive, leading to noisy segments that misrepresent the data's structure. Fig. 7 shows that when using AUs, the score decreases to 0.73. This highlights the importance of accurate facial motion descriptor quality for annotation. Fig. 7 also shows that separating different video clips when analyzing improves performance.", + "bbox": [ + 511, + 657, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To conduct a preliminary test of TICC's effectiveness in detecting unseen facial motion sequences, we remove 50 manually annotated videos from the dataset, refit TICC, and then have it annotate these videos to calculate the macro-F1 score. The scores are similar to those obtained when the 50 videos are included (brow: 0.89; eyes: 0.88; mouth: 0.85).", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1945", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8a887ffba982e6cc951eab7cc89be000a6e3c7dca1e308855d2d9ba3de6165e.jpg", + "image_caption": [ + "Figure 8. Qualitative results of facial motion generation from the timeline. Better viewed in Supp. Video. The timelines for the unplotted regions are set to 0. The same applies to the subsequent figures." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 478, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95d3507b18adee8901178e274db79a145c8753efb88c030a669024210595b756.jpg", + "image_caption": [ + "Figure 9. Qualitative results of ablation study. Green arrows indicate subtle natural head motions, red arrows indicate missing subtle natural head motions, and yellow arrows highlight inaccurate motions. Better viewed in Supp. Video." + ], + "image_footnote": [], + "bbox": [ + 89, + 358, + 478, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Facial Motion Generation", + "text_level": 1, + "bbox": [ + 89, + 755, + 326, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metric. We utilize 100 timelines unseen during training to generate samples for evaluating the motion accuracy and naturalness. To evaluate accuracy, we use TICC or threshold-based approach to annotate facial motion intervals within the generated video. These intervals are then compared to the input timeline to calculate the macro-F1 score for each facial region. The average of these scores across all regions is used to assess the alignment between", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/297e3a4f5555df9646a5167435b1b126e9bd233d817928c33f9a165673cc7410.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsVar→FIDfm↓FIDΔfm↓SND ↓TAS ↑
w/o branch0.687.390.147.530.66
w/o base0.6412.40.1812.580.81
all decoup.0.4128.40.2328.630.69
w/o time con.0.715.380.105.480.79
branchL10.706.250.116.360.76
branchL30.685.760.105.860.84
branchL40.696.640.126.760.83
drop 00.626.880.137.010.78
drop 0.30.675.930.116.040.83
drop 0.70.784.120.094.210.68
Ours0.704.540.094.630.84
", + "bbox": [ + 522, + 88, + 897, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results of ablation study. The unit of Var in the table is $10^{-2}$ , meaning that 0.70 in the table represents $0.70 \\times 10^{-2}$ . The unit of $\\mathrm{FID}_{fm}$ and $\\mathrm{FID}_{\\Delta fm}$ is $10^{-1}$ . Our results differ in magnitude from those of [62] because we use a different 3DMM (FaceVerse) for the calculations. $\\mathbf{\\nabla}^{\\prime \\prime} \\rightarrow$ means results are better if they are close to the variance of real data, which is 0.73.", + "bbox": [ + 511, + 285, + 906, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the generated motion and the input timeline. We refer to this score as Timeline Alignment Score (TAS). To evaluate naturalness, we follow previous methods [44, 61] and use Var (variance of generated facial motions, with values closer to GT indicating better performance), $\\mathbf{FID}_{fm}$ (FID score of 3DMM coefficients), $\\mathbf{FID}_{\\Delta fm}$ (FID score of 3DMM coefficient difference between consecutive frames), SND (sum of $\\mathbf{FID}_{fm}$ and $\\mathbf{FID}_{\\Delta fm}$ ).", + "bbox": [ + 511, + 398, + 906, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 8 shows the results generated by our model. Our model is capable of producing natural facial movements that are aligned with the input timeline. Note that because the training data includes videos of people speaking, our generated videos sometimes include speaking. Since FaceVerse 3DMM decouples the motion of the left and right face, we can generate motion separately for each side, allowing for generating asymmetric expressions. Our method can generate forehead wrinkles when the speaker raises brows. By introducing descriptors that capture more complex texture-related motions, we can extend our method to control them.", + "bbox": [ + 511, + 521, + 908, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results & Ablation Study. To evaluate the impact of our design choices, we conduct an ablation study with several variants: (1) w/o branch remove all branches and only use the base network to predict the facial motions; (2) w/o base remove the base network and only use branch networks to generate each region independently; (3) all decoup. decouple all facial and head motions into different branch networks rather than limiting the decoupling to the upper and lower halves of the face. (4) w/o time con. starting from the second layer of the base/branch network, the initial timeline token input is replaced with the timeline tokens from the previous layer. We investigate the optimal number of layers for the branch network. The entire net", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1946", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21bb55d107fdb2850495767d6e8000864a8bbad61ef995b31e7948cbebe502b7.jpg", + "image_caption": [ + "Figure 10. User study results." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 495, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work consists of 8 layers, with branchLx indicating that the branch network uses $x$ layers while the base network uses $8 - x$ layers. In the optimal version Ours, the branch network has 2 layers. We investigate the optimal drop probability. $\\text{drop} x$ indicates the drop probability is $x$ , the optimal probability used in Ours is 0.5", + "bbox": [ + 89, + 273, + 482, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 2 shows quantitative results and Fig. 9 shows qualitative results from a timeline. w/o branch fails to generate accurate motions, highlighting the necessity of using the branch network to enhance accuracy. When the head is oriented to the left, subtle movements may occur due to changes in gaze and brow. While these subtle motions do not affect the overall head direction, they are essential for motion naturalness. w/o base fails to generate these subtle motions (Fig. 9). Even though the pose branch receives the complete motion timeline, it still cannot produce such coupling. This highlights the necessity of using the base network to learn the coupling of facial movements. all decoup produces stiff and inaccurate motion changes (brow raised too early in Fig. 9) due to overfitting to the movements of individual regions, resulting in weak generalization to timelines outside the training set. The performance of w/o time con. indicates that adding the initial timeline token at each layer can improve accuracy. branchLx's results show that increasing the number of branch layers excessively reduces naturalness. This is because having too few base layers is insufficient to model the natural coupling between motions. drop x's results show that too low or too high drop probability both cannot generate good performance. Too low probability makes the model overly rely on the condition and hampers its generalization capability. Too high probability makes the model receive too few conditioning signals, reducing its accuracy. drop 0.7 get better FID scores, but its accuracy is low.", + "bbox": [ + 91, + 369, + 483, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparisons with previous methods. No prior methods have achieved fine-grained timeline control of facial motion generation. Similar approaches, such as AgentAvatar [44] and InstructAvatar [49], can only describe temporal changes coarsely using temporal adverbs. Neither of these methods provides open-source code. We can only conduct a qualitative comparison using their demos. Since the temporal", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12fc34c9961d90662c415fcc27e46bfcd56f3d99b197d9802ba5480edcb64d3e.jpg", + "image_caption": [ + "Figure 11. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dynamics of facial motion are better demonstrated through video, these comparisons are included in Supp. Video.", + "bbox": [ + 511, + 250, + 903, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. We conduct a user study of 21 participants. The participants are required to rate three types of pairs: (1) GT: Real videos paired with manually annotated facial motion timelines. (2) Annotation: Real videos paired with timelines generated using our method. (3) Generation: Videos generated by our method paired with the input facial motion timelines. Participants need to evaluate their level of agreement (on a 5-point scale: strongly agree/disagree, agree/disagree, don't know) on two aspects: (1) Whether the motions in the video accurately match the corresponding timelines. (2) Whether the motions in the video appear natural. Each participant evaluates 15 pairs sampled from the test data for each type of pair, resulting in a total of 45 pairs per participant. Fig. 9 shows the results. $92\\%$ of the evaluations agree that our annotation is accurate. $89\\%$ and $86\\%$ of the evaluations consider the motions in our generated videos to be accurate and natural, respectively.", + "bbox": [ + 511, + 282, + 905, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generating Motions Using Natural Language. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines. As shown in Fig. 11, ChatGPT can infer that when a person is reminiscing, their head may turn to the side, and their gaze shifts accordingly. When a happy memory comes to mind, the person begins to smile.", + "bbox": [ + 511, + 540, + 906, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explore a new control signal: timeline control for facial motion generation. Timeline allows for more fine-grained control than audio or text, enabling users to generate specific motions with precise timing. To model this capability, we first develop a labor-efficient approach to annotate the temporal intervals of facial actions. The annotation process relies on time series analysis of facial motion descriptors. Based on the annotations, we propose a generation model that can generate natural facial motions aligned with timelines. The generation model utilize a base-branch design to effectively manage motion couplings across different facial regions. Our method supports text-guided generation by using ChatGPT to translate text into timelines. Experiments validate the effectiveness of our method.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1947", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21263-21273, 2024. 2", + "[2] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 3", + "[3] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 3", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2", + "[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2", + "[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2", + "[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, pages 10101-10111, 2019. 2", + "[8] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael J Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. arXiv preprint arXiv:2306.08990, 2023. 2", + "[9] Yicheng Deng, Hideaki Hayashi, and Hajime Nagahara. Spotformer: Multi-scale spatio-temporal transformer for facial expression spotting. arXiv preprint arXiv:2407.20799, 2024. 2", + "[10] Paul Ekman. Facial action coding system (facs). A Human Face, Salt Lake City, 2002. 2", + "[11] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through a unified model. arXiv preprint arXiv:2408.00762, 2024. 2", + "[12] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2", + "[13] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. arXiv preprint arXiv:2301.10939, 2023. 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jiazhi Guan, Zhiliang Xu, Hang Zhou, Kaisiyuan Wang, Shengyi He, Zhanwang Zhang, Borong Liang, Haocheng Feng, Errui Ding, Jingtuo Liu, et al. Resyncer: Rewiring style-based generator for unified audio-visually syncsed facial performer. arXiv preprint arXiv:2408.03284, 2024. 2", + "[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021–2029, 2020. 2", + "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2", + "[17] Xupeng Guo, Xiaobiao Zhang, Lei Li, and Zhaoqiang Xia. Micro-expression spotting with multi-scale local transformer in long videos. Pattern Recognition Letters, 168:146-152, 2023. 2", + "[18] David Hallac, Sagar Vare, Stephen Boyd, and Jure Leskovec. Toeplitz inverse covariance-based clustering of multivariate time series data. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 215-223, 2017. 2, 4, 5", + "[19] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2", + "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 4", + "[22] Takato Honda, Yasuko Matsubara, Ryo Neyama, Mutsumi Abe, and Yasushi Sakurai. Multi-aspect mining of complex sensor sequences. In 2019 IEEE International Conference on Data Mining (ICDM), pages 299-308. IEEE, 2019. 4, 5", + "[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2", + "[24] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2", + "[25] Haojie Li and Shuangping Huang. Kmtalk: Speech-driven 3d facial animation with key motion embedding. 2024. 2", + "[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 2", + "[27] Jun Ling, Yiwen Wang, Han Xue, Rong Xie, and Li Song. Posetalk: Text-and-audio-based pose control and motion refinement for one-shot talking head generation. arXiv preprint arXiv:2409.02657, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1948", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yifeng Ma, Suzhen Wang, Yu Ding, Bowen Ma, Tangjie Lv, Changjie Fan, Zhipeng Hu, Zhidong Deng, and Xin Yu. Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334, 2023. 2", + "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2", + "[30] Yasuko Matsubara, Yasushi Sakurai, and Christos Faloutisos. Autoplait: Automatic mining of co-evolving time sequences. In Proceedings of the 2014 ACM SIGMOD international conference on Management of data, pages 193-204, 2014. 5", + "[31] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[32] Federico Nocentini, Thomas Besnier, Claudio Ferrari, Sylvain Arguillere, Stefano Berretti, and Mohamed Daoudi. Scantalk: 3d talking heads from unregistered scans. arXiv preprint arXiv:2403.10942, 2024. 2", + "[33] Catherine Pelachaud, Norman I Badler, and Mark Steedman. Generating facial expressions for speech. Cognitive science, 20(1):1-46, 1996. 2", + "[34] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2", + "[35] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1921, 2024. 3", + "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2", + "[37] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2", + "[38] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. $\\text{Cas}(\\text{me})^2$ : A database for spontaneous macro-expression and micro-expression spotting and recognition. IEEE Transactions on Affective Computing, 9 (4):424-436, 2018. 2", + "[39] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 3", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2", + "[42] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 2", + "[43] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2", + "[44] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Agentavatar: Disentangling planning, driving and rendering for photorealistic avatar agents. arXiv preprint arXiv:2311.17465, 2023. 1, 2, 7, 8", + "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, pages 700–717. Springer, 2020. 5", + "[46] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20333-20342, 2022. 2, 3, 5", + "[47] Suzhen Wang, Yifeng Ma, Yu Ding, Zhipeng Hu, Changjie Fan, Tangjie Lv, Zhidong Deng, and Xin Yu. Stylistalk++: A unified framework for controlling the speaking styles of talking heads. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[48] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2", + "[49] Yuchi Wang, Junliang Guo, Jianhong Bai, Runyi Yu, Tianyu He, Xu Tan, Xu Sun, and Jiang Bian. Instructavator: Text-guided emotion and motion control for avatar generation. arXiv preprint arXiv:2405.15758, 2024. 1, 2, 8", + "[50] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2", + "[51] Sijing Wu, Yunhao Li, Yichao Yan, Huiyu Duan, Zwei Liu, and Guangtao Zhai. Mmhead: Towards fine-grained multi-modal 3d facial animation. arXiv preprint arXiv:2410.07757, 2024. 1, 2", + "[52] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 2", + "[53] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1949", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 2", + "[55] Karren D Yang, Anurag Ranjan, Jen-Hao Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks methods and applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27294-27303, 2024. 2", + "[56] Chuin Hong Yap, Connah Kendrick, and Moi Hoon Yap. Samm long videos: A spontaneous facial micro-and macro-expressions dataset. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 771-776. IEEE, 2020. 2", + "[57] Chuin Hong Yap, Moi Hoon Yap, Adrian Davison, Connah Kendrick, Jingting Li, Su-Jing Wang, and Ryan Cunningham. 3d-cnn for facial micro-and macro-expression spotting on long video sequences using temporal oriented reference frame. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7016-7020, 2022. 2", + "[58] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2", + "[59] Jun Yu, Zhongpeng Cai, Zepeng Liu, Guochen Xie, and Peng He. Facial expression spotting based on optical flow features. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7205-7209, 2022. 2", + "[60] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. CelebV-Text: A large-scale facial text-video dataset. In CVPR, 2023. 2", + "[61] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 7", + "[62] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7645–7655, 2023. 7", + "[63] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 2", + "[64] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audiodriven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2", + "[65] Qingcheng Zhao, Pengyu Long, Qixuan Zhang, Dafei Qin, Han Liang, Longwen Zhang, Yingliang Zhang, Jingyi Yu, and Lan Xu. Media2face: Co-speech facial animation generation with multi-modality guidance. In ACM SIGGRAPH 2024 Conference Papers, pages 1-13, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 880 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1950", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_model.json b/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_model.json new file mode 100644 index 0000000000000000000000000000000000000000..da277c8c4416c2b9c3d373cd13b8f02c273201ce --- /dev/null +++ b/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Exploring Timeline Control for Facial Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.179, + 0.89, + 0.236 + ], + "angle": 0, + "content": "Yifeng Ma\\(^{1}\\), Jinwei Qi\\(^{2}\\), Chaonan Ji\\(^{2}\\), Peng Zhang\\(^{2}\\), Bang Zhang\\(^{2}\\), Zhidong Deng\\(^{1}\\), Liefeng Bo\\(^{2}\\) \n\\(^{1}\\) Department of Computer Science and Technology, Tsinghua University \n\\(^{2}\\) Tongyi Lab, Alibaba Group" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.283, + 0.487, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.283, + 0.902, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.483, + 0.908, + 0.541 + ], + "angle": 0, + "content": "Figure 1. We introduce a new control signal for facial motion generation: timeline control. We first utilize a labor-efficient approach to annotate the time intervals of facial motion at a frame-level granularity. Using the annotations, we propose a model that can generate natural facial motions aligned with an input timeline. Compared to previous controls like audio and text, timeline control enables precise temporal control of facial motions. In this paper, facial motions are rendered into photorealistic videos for better visualization." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.554, + 0.327, + 0.57 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.596, + 0.486, + 0.884 + ], + "angle": 0, + "content": "This paper introduces a new control signal for facial motion generation: timeline control. Compared to audio and text signals, timelines provide more fine-grained control, such as generating specific facial motions with precise timing. Users can specify a multi-track timeline of facial actions arranged in temporal intervals, allowing precise control over the timing of each action. To model the timeline control capability, We first annotate the time intervals of facial actions in natural facial motion sequences at a frame-level granularity. This process is facilitated by Toeplitz Inverse Covariance-based Clustering to minimize human labor. Based on the annotations, we propose a diffusion-based generation model capable of generating facial motions that are natural and accurately aligned with input timelines. Our method supports text-guided motion generation by using ChatGPT to convert text into timelines. Experimental results show that our method can annotate facial action intervals with satisfactory accuracy, and produces natural facial motions accurately aligned with timelines." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.554, + 0.645, + 0.57 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.581, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Generating vivid facial motions has drawn growing attention due to its broad applications, including digital human generation and filmmaking. To produce desired facial motions, current methods use audio or text to provide guidance. However, these methods lack a critical ability that users often require: generating specific facial motions with precise timing. For example, users may want to generate a brow raise between frames 10 and 30 while simultaneously generating a smile between frames 14 and 43. Audiodriven methods [43, 54] can only generate motions synchronized with the audio. Text signals offer only coarse-grained descriptions of facial motions and lack frame-level detail. Text-driven approaches [4, 44, 49, 51] rely on temporal adverbial cues (e.g., then) for rough timing guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.907, + 0.884 + ], + "angle": 0, + "content": "To achieve more fine-grained control, we introduce a novel control signal: timeline control for facial motion generation. In this setup, users can generate facial motions by inputting an intuitive timeline that contains several temporal intervals, each corresponding to a desired facial action. This setup enables users to manage the timing of each action." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Achieving timeline control is very challenging. Frame" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1940" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "level control of facial motions requires the model to achieve exceptional precision in generating accurate actions. To model timeline control capability, it is crucial to annotate the precise start and end frames of facial actions, a challenge that existing methods have yet to overcome. Existing methods [44, 51] represent facial motion sequences as a time series of motion descriptors, such as Facial Action Units (AUs) [10] or blendshapes [46], and then leverage ChatGPT to summarize these time series for annotation. However, ChatGPT's limited sensitivity to the temporal dynamics of facial motions prevents it from determining the exact start and end frames of facial actions. Another annotation approach is using thresholds, like labeling the eye as \"closed\" if the blendshape eyeBlink exceeds 0.4. However, setting thresholds for complex actions, such as brow motions, is challenging. Additionally, some actions are determined by the relative values of multiple motion descriptors rather than a single one." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.368, + 0.483, + 0.7 + ], + "angle": 0, + "content": "To address this issue, we adopt Toeplitz Inverse Covariance-based Clustering (TICC) [18] to annotate the start and end frames of facial actions. TICC can segment the time series into a sequence of intervals, with each interval containing a single motion pattern. Each interval has clearly defined start and end frames. Once the start and end frames of the intervals are identified, the remaining task is to determine the facial action represented by each interval. This can be addressed by another feature of TICC: its ability to group the segmented intervals into several clusters based on the similarity of their motion patterns, with each cluster containing intervals that exhibit similar motion patterns. The automatic clustering process eliminates the need for manual threshold setting and considers the relationships between multiple motion descriptors. Once these clusters are obtained, we can determine the overall action (e.g. significantly raised brows) represented by each cluster by analyzing a few representative patterns within it. This process requires minimal labor. Using this procedure, we analyze motions in different facial regions individually and achieve frame-level annotation for the brows, eyes, mouth, gaze, and head motions in a labor-efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Using frame-level facial motion annotations, we propose a diffusion-based generation model that produces natural facial motions accurately aligned with the input timeline. The model generates motions for each facial region (e.g. upper face, lower face) separately. This decoupling can help address the adverse impact of motion coupling in the data, thereby improving accuracy. However, some couplings are necessary for motion naturalness, such as a slight head lift during a brow raise. To properly manage coupling, the generation model is divided into a base network and multiple branch networks, with each branch network dedicated to a facial region. The base network encodes global facial motion couplings into base features. The branch network takes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "base features and the timeline of the corresponding region as input to accurately generate facial motions for that region while maintaining natural couplings. We use FaceVerse[46] 3DMM coefficients to represent facial motions for generation. Since photorealistic facial motion enables more accurate assessments of motion realism than mesh-based motion [31], we render motions into photorealistic videos by a diffusion-based renderer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Our method enables text-guided facial motion generation by using ChatGPT to convert text into timelines. ChatGPT effectively translates simple natural language descriptions into timelines, which our model transforms into facial motions. Users can modify these timelines, allowing precise timing control for text-driven applications." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.305, + 0.833, + 0.32 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.325, + 0.905, + 0.385 + ], + "angle": 0, + "content": "- We are the first to develop a labor-efficient approach to annotate temporal intervals of facial actions. Such fine-grained annotations enable precise modeling of the temporal dynamics of facial motions." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.385, + 0.905, + 0.446 + ], + "angle": 0, + "content": "- We are the first to achieve timeline control for facial motion generation. This enables users to generate specific facial motions with precise timing. Our method also supports generating actions from natural language text." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.446, + 0.905, + 0.491 + ], + "angle": 0, + "content": "- Detailed evaluations show that our method generates accurate motion annotations and produces natural, precise facial motions from timeline inputs." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.325, + 0.905, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.511, + 0.655, + 0.527 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Facial Motion Annotation. Compared to human motion annotation [15, 16, 29, 36], facial motion annotation is in its early stages. Existing methods [28, 44, 49, 51, 60] generate annotations that either overlook temporal changes or describe them only roughly with temporal adverbs. Some datasets [38, 56] label intervals of macro/micro expressions rather than facial actions used for facial motion generation. These datasets also require intensive manual labeling. Facial expression spotting methods [9, 17, 48, 57, 59] are also restricted to label macro/micro-expressions. In this paper, we aim to labor-efficiently generate frame-level facial action annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Facial Motion Generation. Existing methods use audio or text as the control signal and cannot generate specific motions at precise timing. Audio-driven methods (2D [6, 14, 19, 23, 26, 32, 37, 42, 43, 47, 50, 53, 54, 58, 63, 64] and 3D methods [1, 7, 8, 11, 12, 24, 25, 34, 41, 52, 55, 65]) can only generate facial motions synchronized with the audio. Text-driven methods [4, 27, 28, 44, 49, 51] can only use temporal adverbial phrases to control timing roughly. Rule-based methods [5, 33] offer temporal control but often produce unnatural motions due to deviations from real movement distributions. This paper introduces timeline control to generate natural facial motions with precise timing." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1941" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.895, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 2. The pipeline of frame-level facial motion annotation (using brow motions as an example). We first extract facial motion descriptors (blendshapes) from natural facial motion videos and concatenate the results to create a facial motion time series for time series analysis. This analysis can simultaneously segment the sequence into a series of motion patterns and cluster similar patterns, resulting in multiple clusters, each containing consistent facial motion patterns. Then, by inspecting a few patterns, we identify the facial motions each cluster represents, thereby obtaining frame-level facial motion annotations for all videos." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.37, + 0.479, + 0.472 + ], + "angle": 0, + "content": "
Facial RegionMotion Categories
BrowBrowUp, BrowDown, Neutral
EyeEyeSquint, EyeWiden, EyeClose, Neutral
MouthSoftSmile, Smile, MouthFrown, Neutral
GazeLeft, Right, Up, Down, Neutral
HeadLeft, Right, Up, Down, Neutral
" + }, + { + "type": "table_caption", + "bbox": [ + 0.142, + 0.472, + 0.435, + 0.486 + ], + "angle": 0, + "content": "Table 1. Facial actions annotated by our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Timeline control has been utilized in human motion generation [2, 3, 35, 39] by first generating motion segments and then piecing them together. However, due to the rapid changes in facial movements, this approach is unsuitable, necessitating a new model structure for effective generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.184, + 0.603 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.718 + ], + "angle": 0, + "content": "In this work, we first annotate the temporal intervals of facial action for natural facial motion sequences. Based on these annotations, we develop a generation model that generates natural facial motions that are accurately aligned with an input timeline. Our method also supports text-guided motion generation by leveraging ChatGPT to convert text into timelines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.484, + 0.745 + ], + "angle": 0, + "content": "3.1. Annotating temporal intervals of facial actions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Problem Formulation. For each video \\( V \\) in dataset, our method aims to generate frame-level facial motion annotations \\( A = [a_i]_{i=1}^L \\) for each video frame \\( [v_i]_{i=1}^L \\). Tab. 1 shows the annotated facial actions. The annotation \\( a_i \\) for each frame is a vector of binary values, with each dimension representing a non-neutral facial action. A dimension is set to 1 if the action occurs, otherwise, it is 0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Annotation Process Overview. We achieve the annotation through time series analysis TICC. Fig. 2 shows the pipeline: (1) We extract proper motion descriptors for each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.508 + ], + "angle": 0, + "content": "facial region to obtain facial motion time series. (2) We apply TICC to simultaneously segment the time series into a sequence of motion patterns and cluster similar patterns. This results in several clusters of intervals, each containing similar facial motions. (3) Finally, for each cluster, we can infer the actions of all intervals by examining only a few intervals. Once the action for each cluster is identified, the actions for all intervals are determined, resulting in facial action annotations at a frame-level granularity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.659 + ], + "angle": 0, + "content": "Facial Motion Descriptor. We observe that previously used AUs lack sufficient precision, while ARKit blendshapes offer a more accurate representation. The publicly available MediaPipe blendshape detector struggles to detect certain blendshapes (e.g. eyeWide, mouthFrown), so we use an in-house blendshape detector. ARKit blendshapes are a set of coefficients that describe facial expressions, with each coefficient representing the movement of a specific facial region. Each coefficient ranges from 0 to 1, indicating the motion intensity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.84 + ], + "angle": 0, + "content": "We use the blendshapes corresponding to each facial region to construct motion time series for each region, resulting in time series data for the eye, brow, and mouth areas. For the eye region, we selected eyeBlink, eyeSquint, eyeWide, and for the brow region, we selected browDown, browInnerUp, browOuterUp. For the mouth region, we use only expression-related blendshapes (mouthSmile, mouthStretch, mouthFrown) to avoid interference from speech-related movements. ARKit blendshapes provide coefficients for the left and right sides. As facial motions in our dataset are largely symmetrical, we use only the left-side coefficients to simplify analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.884 + ], + "angle": 0, + "content": "For gaze and head pose motion, we use a 3DMM model FaceVerse [46] to detect coefficients. For gaze, we select eye coefficients, for head pose, we select angle coefficients." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Analyzing Facial Motion Time Series. We utilize" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1942" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.096, + 0.58, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.254, + 0.478, + 0.266 + ], + "angle": 0, + "content": "(a) Generation Model's Base-Branch Design" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.094, + 0.895, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.254, + 0.877, + 0.265 + ], + "angle": 0, + "content": "(b) Base/Branch Network's Architecture" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Illustration of generation model. (a) Base-Branch Design. The base network takes the timelines of all facial regions as input and outputs base features that model the global facial motion couplings. Through timeline selection, each region's timeline is directed to its respective branch network. Since head pose is interconnected with all facial movements, the pose branch receives timelines of all regions. Each branch network takes the timeline of its corresponding region to generate the facial motions for that region. These motions are then combined to produce the overall motion of the entire face. Lin. Proj. denotes Linear Projection. (b) Base/Branch Network's Architecture. Timeline control guides motion generation through cross-attention. The initial timeline tokens remain unchanged and are added at each layer. For clarity, the diffusion step (omitted in sub-figure (a) for clarity) is applied to each base and branch network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.546 + ], + "angle": 0, + "content": "TICC [18] for time series analysis. TICC is designed to handle a single long sequence (CubeMarker [22] can analyze multiple time series analysis but performs poorly.). However, our data consists of multiple short sequences. To address this issue, we concatenate all the short sequences into a single long sequence for analysis. To avoid the impact of abrupt transitions between different videos, we introduce a \"null sequence\" between videos to separate them. The null sequence is implemented as a length-100 sequence with values set to -1. These null sequences will eventually be grouped into a single cluster, which will not affect others." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.715 + ], + "angle": 0, + "content": "After the time series analysis, the video is annotated with a sequence of motion pattern intervals and similar patterns are clustered into a few clusters. Therefore, for each cluster, we inspect a few intervals to determine the facial action they represent. This facial action is then considered to be the action for all intervals within that cluster. Different clusters may represent variations of the same facial action category, such as a high eyebrow raise versus a moderate one. We categorized the clusters into facial motion categories, such as eyebrow raise, frown, and neutral, to obtain frame-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.867 + ], + "angle": 0, + "content": "We observe that eye closure frequently produces sharp spikes in the time series, which disrupts the analysis of squinting and widening. Therefore, we analyze eye closure and squinting/widening separately. We also find that threshold-based annotation, which classifies actions using predefined thresholds (e.g., labeling the eye as \"closed\" if eyeBlink exceeds 0.4), achieves high accuracy for eye closure, gaze, and head pose. Consequently, we use this method for these regions, while applying time series analysis for others." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Detect Facial Motion for Unseen Videos. After learning various facial motion patterns from the data, TICC can also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.426 + ], + "angle": 0, + "content": "be applied to previously unseen facial motion sequences and detect facial action intervals. We use this function to assess the accuracy of facial motion generation in our evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.869, + 0.453 + ], + "angle": 0, + "content": "3.2. Facial Motion Generation from Timelines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.567 + ], + "angle": 0, + "content": "Problem Formulation. Given a timeline control \\( C = [c_i]_{i=1}^L \\) for each video frame, the generation model \\( \\mathcal{G} \\) aims to generate natural facial motions \\( M = [m_l]_{l=1}^L \\) that align with the timeline control. The control for each frame \\( c_i \\) follows the same format as the annotation. Generated facial motions are rendered by a diffusion-based renderer into photorealistic videos for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Discussion: The coupling of facial motions and its impact on motion accuracy and naturalness. Facial motions across different regions are coupled. For example, when people smile, they might also squint and lower their eyebrows. These couplings can reduce the precision of generated motions. For instance, when generating a smile, the model may also learn to produce eyebrow-lowering actions, which could conflict with user-specified eyebrow motions (e.g. brow raise). To enhance accuracy, it is essential to decouple the generation of different facial regions. However, certain facial motion couplings are crucial for conveying naturalness—for instance, a subtle head movement that accompanies a shift in gaze. Therefore, the model must achieve a delicate balance, selectively decoupling motions to improve accuracy while preserving necessary couplings to enhance naturalness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Generation Model \\(\\mathcal{G}\\). \\(\\mathcal{G}\\) is a diffusion-based model. Diffusion models consist of two Markov chains [21, 40]: the forward chain incrementally injects Gaussian noise into the original signal, while the reverse chain sequentially reconstructs the original signal from the noise. \\(\\mathcal{G}\\) predicts the original signal instead of noise, and the loss function is in" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.176, + 0.106 + ], + "angle": 0, + "content": "produced as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.114, + 0.483, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e n o i s e}} = \\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ]}, M _ {(0)}, C (\\| M _ {(0)} - \\mathcal {G} (M _ {(t)}, t, C) \\| ^ {2}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.147, + 0.483, + 0.207 + ], + "angle": 0, + "content": "where \\(t\\) denotes the diffusion step, \\(M_{(0)}\\) is the original facial motion sequence, and \\(M_{(t)}\\) is the noisy sequence produced by the diffusion forward function \\(q(M_{(t)}|M_{(t - 1)}) = \\mathcal{N}(\\sqrt{\\alpha_n} M_{(t - 1)},(1 - \\alpha_n)\\pmb {I})\\). \\(\\pmb{C}\\) is the timeline condition." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.208, + 0.483, + 0.357 + ], + "angle": 0, + "content": "To balance decoupling and coupling for both accuracy and naturalness, the generation model employs a base-branch design (Fig. 3). It consists of a base network and individual branch networks for each facial region (e.g. upper face, lower face). The base network takes timelines of all facial regions and noisy motions as input and encodes global motion couplings into base features. The branch network takes the base features the timelines relevant only to its designated facial region as inputs, and generates decoupled motions for each region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.614 + ], + "angle": 0, + "content": "We use the expression, eye, and pose coefficients of FaceVerse [46] 3DMM model as target motion representation for generation. Its expression coefficients align with ARKit blendshapes, allowing separate representations of different facial regions. When dividing facial motions into different regions for separate generations, we observe that splitting them into the upper face, lower face, and pose & other regions strikes a better balance between decoupling and necessary coupling, rather than assigning a branch network for each individual region (e.g. further split eye and brow). Specifically, the upper face includes eye, brow, and gaze; the lower face includes mouth and jaw; and the pose & other regions cover the head pose and remaining regions like cheek and nose. It is important to note that, since pose is coupled with facial motions in all regions, the timeline control for the pose & other branch includes all facial regions (rather than just the control for pose)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As for the specific implementation, the timeline control for each frame \\( c_{i} \\) is implemented as a 16-dimensional vector composed of values 0 and 1, with each dimension representing an action. A value of 1 indicates that the action is performed in that frame, and a value of 0 indicates that the action is not present. The timeline control is transformed into timeline tokens through a linear projection. The base network and branch networks share the same structure. Their inputs consist of timeline tokens and motion tokens. The motion tokens are derived either from the noisy motion through linear projection or from base features. The motion tokens learn the temporal control from the timeline tokens through cross-attention in multiple transformer encoder layers. When incorporated into each encoder layer, the timeline tokens always use the initial timeline tokens rather than the output tokens from the previous layer. This prevents the temporal information in timeline tokens from being altered, thereby enhancing motion accuracy. The motion tokens for each layer come from the output of the motion tokens from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "the previous layer. The network's final output consists of the motion tokens from the last layer, which are then linearly projected to produce the facial motion. Each branch network generates the facial motion for its corresponding region. These motions are then combined to produce the motion for the entire face." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.302 + ], + "angle": 0, + "content": "To improve flexibility and generalization, we use classifier-free guidance [20] to train our model. During training, we randomly drop the condition of each facial region timeline. We use a dropout probability of 0.5 for each condition independently, with a 0.1 probability of dropping all conditions and a 0.1 probability of maintaining all conditions. When the timeline condition for a certain region is dropped, its value is set to -1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.363 + ], + "angle": 0, + "content": "Rendering. We develop a diffusion-based portrait animation method to render motions as images. Since the renderer is not the main contribution of this paper, its details are reported in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.372, + 0.882, + 0.387 + ], + "angle": 0, + "content": "3.3. Translating Natural Language to Timelines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.53 + ], + "angle": 0, + "content": "We observe that ChatGPT can generate a reasonable timeline from natural language descriptions, enabling our model to produce facial motions based on natural language input. The generated timeline can be further edited by users to fit their needs. To enhance the realism of the generated movements, we manually annotated several timeline descriptions as examples, allowing ChatGPT to perform few-shot learning. The details of the prompts and generated results are reported in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.645, + 0.56 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Dataset. To effectively model natural facial motions, a dataset capturing authentic facial expressions is essential. For this purpose, we utilize the RealTalk dataset [13]. The RealTalk dataset contains 692 videos captured from genuine, unscripted conversations, resulting in highly natural facial movements with rich temporal dynamics. Previously used lab-recorded, scripted datasets, such as MEAD [45], lack these qualities. We use 1412 video clips extracted from Realtalk as the dataset. The size of the dataset is approximately 600,000 frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.748, + 0.742 + ], + "angle": 0, + "content": "4.1. Facial Motion Annotation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Baselines. No previous methods have attempted annotating intervals of facial actions in a labor-efficient manner. We experimented with time series analysis methods such as AutoPlait [30], CubeMarker [22], and TICC. However, AutoPlait and CubeMarker failed to produce satisfactory results, as the motions of the segments within each cluster were inconsistent, making annotation impossible. Therefore, we primarily relied on TICC. Empirical results [18] indicate TICC's annotation performance can be enhanced by tuning key hyperparameters, including the number of clusters and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.478, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.235, + 0.435, + 0.249 + ], + "angle": 0, + "content": "Figure 4. An example of brow motion annotation." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.261, + 0.47, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.407, + 0.431, + 0.42 + ], + "angle": 0, + "content": "Figure 5. An example of eye motion annotation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.481, + 0.473 + ], + "angle": 0, + "content": "beta, a hyperparameter that facilitates temporal consistency. We examine the effects of these parameters." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Evaluation Metric. Annotating intervals of facial actions is essentially a multi-class classification problem, where each data point represents a frame and each class represents an action. To evaluate annotation accuracy, we manually annotate the facial actions on 50 videos to serve as ground truth, and use Macro-F1 as the metric for evaluation. For each class, the F1 score is the harmonic mean of the precision and recall of our estimate. Then, the macro-F1 score is the average of the F1 scores for all the classes. A higher Macro-F1 score indicates greater annotation accuracy. We calculated the macro-F1 score for the eye region only on segments when the eyes are not closed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Qualitative results. Fig. 4 shows brow motion annotation. Our method can precisely segment different facial motion patterns and cluster similar ones. For a specific facial motion brow raise, our approach can distinguish multiple variants, such as a significant or moderate raise. In this work, these finer categories are grouped under one motion category for simplicity, but future methods could generate each subcategory separately. Annotation for the mouth is similar to that for the brow. Fig. 5 shows eye motion annotation. As stated in Sec. 3.1, only eye squint/widen are annotated using TICC. Eye closure is determined by a threshold on the blendshape eyeBlink and may occur within intervals of other eye motions. We also conduct a user study to evaluate annotation accuracy and the results are reported in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. We select the optimal hyperparameters based on the highest macro-F1 score. Experiments" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.092, + 0.897, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.222, + 0.905, + 0.236 + ], + "angle": 0, + "content": "Figure 6. The annotation accuracy for different number of clusters." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.248, + 0.897, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.378, + 0.905, + 0.407 + ], + "angle": 0, + "content": "Figure 7. The annotation accuracy in brows for different \\(\\beta\\), concatenating videos without separating, and using AU as descriptors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.565 + ], + "angle": 0, + "content": "show that the optimal beta value for different regions is 5, with the best number of clusters being 8 for the eye region, 9 for the brow region, and 9 for the mouth region. (The number of clusters here refers to the count of valid clusters, excluding the one that contain only null sequences). The optimal scores for brow, eyes, and mouth are 0.90, 0.91, and 0.87, respectively. When using threshold-based annotation, the scores for eye closure, pose, and gaze are 0.95, 0.87, and 0.89, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.655 + ], + "angle": 0, + "content": "The impact of different cluster numbers is shown in Fig. 6 (with beta set to 5). Too few clusters in TICC lead to under-segmentation, grouping distinct patterns into single clusters. Conversely, too many clusters decreases intercluster variability, making clusters less distinct and blurring boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Fig. 7 illustrates the impact of \\(\\beta\\), using brow as an example (with the number of clusters set to 8). \\(\\beta\\) controls the smoothness of segment transitions. Too large \\(\\beta\\) leads to over-smoothing, Conversely, a small \\(\\beta\\) makes the model overly sensitive, leading to noisy segments that misrepresent the data's structure. Fig. 7 shows that when using AUs, the score decreases to 0.73. This highlights the importance of accurate facial motion descriptor quality for annotation. Fig. 7 also shows that separating different video clips when analyzing improves performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "To conduct a preliminary test of TICC's effectiveness in detecting unseen facial motion sequences, we remove 50 manually annotated videos from the dataset, refit TICC, and then have it annotate these videos to calculate the macro-F1 score. The scores are similar to those obtained when the 50 videos are included (brow: 0.89; eyes: 0.88; mouth: 0.85)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.088, + 0.48, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Figure 8. Qualitative results of facial motion generation from the timeline. Better viewed in Supp. Video. The timelines for the unplotted regions are set to 0. The same applies to the subsequent figures." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.359, + 0.48, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.677, + 0.483, + 0.733 + ], + "angle": 0, + "content": "Figure 9. Qualitative results of ablation study. Green arrows indicate subtle natural head motions, red arrows indicate missing subtle natural head motions, and yellow arrows highlight inaccurate motions. Better viewed in Supp. Video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.327, + 0.77 + ], + "angle": 0, + "content": "4.2. Facial Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metric. We utilize 100 timelines unseen during training to generate samples for evaluating the motion accuracy and naturalness. To evaluate accuracy, we use TICC or threshold-based approach to annotate facial motion intervals within the generated video. These intervals are then compared to the input timeline to calculate the macro-F1 score for each facial region. The average of these scores across all regions is used to assess the alignment between" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.089, + 0.898, + 0.273 + ], + "angle": 0, + "content": "
MethodsVar→FIDfm↓FIDΔfm↓SND ↓TAS ↑
w/o branch0.687.390.147.530.66
w/o base0.6412.40.1812.580.81
all decoup.0.4128.40.2328.630.69
w/o time con.0.715.380.105.480.79
branchL10.706.250.116.360.76
branchL30.685.760.105.860.84
branchL40.696.640.126.760.83
drop 00.626.880.137.010.78
drop 0.30.675.930.116.040.83
drop 0.70.784.120.094.210.68
Ours0.704.540.094.630.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.37 + ], + "angle": 0, + "content": "Table 2. Results of ablation study. The unit of Var in the table is \\(10^{-2}\\), meaning that 0.70 in the table represents \\(0.70 \\times 10^{-2}\\). The unit of \\(\\mathrm{FID}_{fm}\\) and \\(\\mathrm{FID}_{\\Delta fm}\\) is \\(10^{-1}\\). Our results differ in magnitude from those of [62] because we use a different 3DMM (FaceVerse) for the calculations. \\(\\mathbf{\\nabla}^{\\prime \\prime} \\rightarrow\\) means results are better if they are close to the variance of real data, which is 0.73." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.522 + ], + "angle": 0, + "content": "the generated motion and the input timeline. We refer to this score as Timeline Alignment Score (TAS). To evaluate naturalness, we follow previous methods [44, 61] and use Var (variance of generated facial motions, with values closer to GT indicating better performance), \\(\\mathbf{FID}_{fm}\\) (FID score of 3DMM coefficients), \\(\\mathbf{FID}_{\\Delta fm}\\) (FID score of 3DMM coefficient difference between consecutive frames), SND (sum of \\(\\mathbf{FID}_{fm}\\) and \\(\\mathbf{FID}_{\\Delta fm}\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.702 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 8 shows the results generated by our model. Our model is capable of producing natural facial movements that are aligned with the input timeline. Note that because the training data includes videos of people speaking, our generated videos sometimes include speaking. Since FaceVerse 3DMM decouples the motion of the left and right face, we can generate motion separately for each side, allowing for generating asymmetric expressions. Our method can generate forehead wrinkles when the speaker raises brows. By introducing descriptors that capture more complex texture-related motions, we can extend our method to control them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results & Ablation Study. To evaluate the impact of our design choices, we conduct an ablation study with several variants: (1) w/o branch remove all branches and only use the base network to predict the facial motions; (2) w/o base remove the base network and only use branch networks to generate each region independently; (3) all decoup. decouple all facial and head motions into different branch networks rather than limiting the decoupling to the upper and lower halves of the face. (4) w/o time con. starting from the second layer of the base/branch network, the initial timeline token input is replaced with the timeline tokens from the previous layer. We investigate the optimal number of layers for the branch network. The entire net" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.496, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.256, + 0.376, + 0.27 + ], + "angle": 0, + "content": "Figure 10. User study results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.483, + 0.366 + ], + "angle": 0, + "content": "work consists of 8 layers, with branchLx indicating that the branch network uses \\( x \\) layers while the base network uses \\( 8 - x \\) layers. In the optimal version Ours, the branch network has 2 layers. We investigate the optimal drop probability. \\( \\text{drop} x \\) indicates the drop probability is \\( x \\), the optimal probability used in Ours is 0.5" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.37, + 0.485, + 0.791 + ], + "angle": 0, + "content": "Tab. 2 shows quantitative results and Fig. 9 shows qualitative results from a timeline. w/o branch fails to generate accurate motions, highlighting the necessity of using the branch network to enhance accuracy. When the head is oriented to the left, subtle movements may occur due to changes in gaze and brow. While these subtle motions do not affect the overall head direction, they are essential for motion naturalness. w/o base fails to generate these subtle motions (Fig. 9). Even though the pose branch receives the complete motion timeline, it still cannot produce such coupling. This highlights the necessity of using the base network to learn the coupling of facial movements. all decoup produces stiff and inaccurate motion changes (brow raised too early in Fig. 9) due to overfitting to the movements of individual regions, resulting in weak generalization to timelines outside the training set. The performance of w/o time con. indicates that adding the initial timeline token at each layer can improve accuracy. branchLx's results show that increasing the number of branch layers excessively reduces naturalness. This is because having too few base layers is insufficient to model the natural coupling between motions. drop x's results show that too low or too high drop probability both cannot generate good performance. Too low probability makes the model overly rely on the condition and hampers its generalization capability. Too high probability makes the model receive too few conditioning signals, reducing its accuracy. drop 0.7 get better FID scores, but its accuracy is low." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparisons with previous methods. No prior methods have achieved fine-grained timeline control of facial motion generation. Similar approaches, such as AgentAvatar [44] and InstructAvatar [49], can only describe temporal changes coarsely using temporal adverbs. Neither of these methods provides open-source code. We can only conduct a qualitative comparison using their demos. Since the temporal" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.202, + 0.905, + 0.229 + ], + "angle": 0, + "content": "Figure 11. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.905, + 0.281 + ], + "angle": 0, + "content": "dynamics of facial motion are better demonstrated through video, these comparisons are included in Supp. Video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.906, + 0.539 + ], + "angle": 0, + "content": "User Study. We conduct a user study of 21 participants. The participants are required to rate three types of pairs: (1) GT: Real videos paired with manually annotated facial motion timelines. (2) Annotation: Real videos paired with timelines generated using our method. (3) Generation: Videos generated by our method paired with the input facial motion timelines. Participants need to evaluate their level of agreement (on a 5-point scale: strongly agree/disagree, agree/disagree, don't know) on two aspects: (1) Whether the motions in the video accurately match the corresponding timelines. (2) Whether the motions in the video appear natural. Each participant evaluates 15 pairs sampled from the test data for each type of pair, resulting in a total of 45 pairs per participant. Fig. 9 shows the results. \\(92\\%\\) of the evaluations agree that our annotation is accurate. \\(89\\%\\) and \\(86\\%\\) of the evaluations consider the motions in our generated videos to be accurate and natural, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Generating Motions Using Natural Language. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines. As shown in Fig. 11, ChatGPT can infer that when a person is reminiscing, their head may turn to the side, and their gaze shifts accordingly. When a happy memory comes to mind, the person begins to smile." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.634, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we explore a new control signal: timeline control for facial motion generation. Timeline allows for more fine-grained control than audio or text, enabling users to generate specific motions with precise timing. To model this capability, we first develop a labor-efficient approach to annotate the temporal intervals of facial actions. The annotation process relies on time series analysis of facial motion descriptors. Based on the annotations, we propose a generation model that can generate natural facial motions aligned with timelines. The generation model utilize a base-branch design to effectively manage motion couplings across different facial regions. Our method supports text-guided generation by using ChatGPT to translate text into timelines. Experiments validate the effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1947" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21263-21273, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, pages 10101-10111, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[8] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael J Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. arXiv preprint arXiv:2306.08990, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[9] Yicheng Deng, Hideaki Hayashi, and Hajime Nagahara. Spotformer: Multi-scale spatio-temporal transformer for facial expression spotting. arXiv preprint arXiv:2407.20799, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[10] Paul Ekman. Facial action coding system (facs). A Human Face, Salt Lake City, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[11] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through a unified model. arXiv preprint arXiv:2408.00762, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. arXiv preprint arXiv:2301.10939, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Jiazhi Guan, Zhiliang Xu, Hang Zhou, Kaisiyuan Wang, Shengyi He, Zhanwang Zhang, Borong Liang, Haocheng Feng, Errui Ding, Jingtuo Liu, et al. Resyncer: Rewiring style-based generator for unified audio-visually syncsed facial performer. arXiv preprint arXiv:2408.03284, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.162, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021–2029, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[17] Xupeng Guo, Xiaobiao Zhang, Lei Li, and Zhaoqiang Xia. Micro-expression spotting with multi-scale local transformer in long videos. Pattern Recognition Letters, 168:146-152, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[18] David Hallac, Sagar Vare, Stephen Boyd, and Jure Leskovec. Toeplitz inverse covariance-based clustering of multivariate time series data. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 215-223, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[19] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[22] Takato Honda, Yasuko Matsubara, Ryo Neyama, Mutsumi Abe, and Yasushi Sakurai. Multi-aspect mining of complex sensor sequences. In 2019 IEEE International Conference on Data Mining (ICDM), pages 299-308. IEEE, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[24] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[25] Haojie Li and Shuangping Huang. Kmtalk: Speech-driven 3d facial animation with key motion embedding. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Jun Ling, Yiwen Wang, Han Xue, Rong Xie, and Li Song. Posetalk: Text-and-audio-based pose control and motion refinement for one-shot talking head generation. arXiv preprint arXiv:2409.02657, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1948" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[28] Yifeng Ma, Suzhen Wang, Yu Ding, Bowen Ma, Tangjie Lv, Changjie Fan, Zhipeng Hu, Zhidong Deng, and Xin Yu. Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[30] Yasuko Matsubara, Yasushi Sakurai, and Christos Faloutisos. Autoplait: Automatic mining of co-evolving time sequences. In Proceedings of the 2014 ACM SIGMOD international conference on Management of data, pages 193-204, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[31] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[32] Federico Nocentini, Thomas Besnier, Claudio Ferrari, Sylvain Arguillere, Stefano Berretti, and Mohamed Daoudi. Scantalk: 3d talking heads from unregistered scans. arXiv preprint arXiv:2403.10942, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[33] Catherine Pelachaud, Norman I Badler, and Mark Steedman. Generating facial expressions for speech. Cognitive science, 20(1):1-46, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[34] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[35] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1921, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[37] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[38] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. \\( \\text{Cas}(\\text{me})^2 \\): A database for spontaneous macro-expression and micro-expression spotting and recognition. IEEE Transactions on Affective Computing, 9 (4):424-436, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[39] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[42] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[43] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[44] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Agentavatar: Disentangling planning, driving and rendering for photorealistic avatar agents. arXiv preprint arXiv:2311.17465, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, pages 700–717. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[46] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20333-20342, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[47] Suzhen Wang, Yifeng Ma, Yu Ding, Zhipeng Hu, Changjie Fan, Tangjie Lv, Zhidong Deng, and Xin Yu. Stylistalk++: A unified framework for controlling the speaking styles of talking heads. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[48] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[49] Yuchi Wang, Junliang Guo, Jianhong Bai, Runyi Yu, Tianyu He, Xu Tan, Xu Sun, and Jiang Bian. Instructavator: Text-guided emotion and motion control for avatar generation. arXiv preprint arXiv:2405.15758, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[50] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[51] Sijing Wu, Yunhao Li, Yichao Yan, Huiyu Duan, Zwei Liu, and Guangtao Zhai. Mmhead: Towards fine-grained multi-modal 3d facial animation. arXiv preprint arXiv:2410.07757, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[52] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[53] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1949" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[54] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[55] Karren D Yang, Anurag Ranjan, Jen-Hao Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks methods and applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27294-27303, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[56] Chuin Hong Yap, Connah Kendrick, and Moi Hoon Yap. Samm long videos: A spontaneous facial micro-and macro-expressions dataset. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 771-776. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[57] Chuin Hong Yap, Moi Hoon Yap, Adrian Davison, Connah Kendrick, Jingting Li, Su-Jing Wang, and Ryan Cunningham. 3d-cnn for facial micro-and macro-expression spotting on long video sequences using temporal oriented reference frame. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7016-7020, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[58] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.459, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[59] Jun Yu, Zhongpeng Cai, Zepeng Liu, Guochen Xie, and Peng He. Facial expression spotting based on optical flow features. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7205-7209, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[60] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. CelebV-Text: A large-scale facial text-video dataset. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[61] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[62] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7645–7655, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[63] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[64] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audiodriven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.813, + 0.482, + 0.881 + ], + "angle": 0, + "content": "[65] Qingcheng Zhao, Pengyu Long, Qixuan Zhang, Dafei Qin, Han Liang, Longwen Zhang, Yingliang Zhang, Jingyi Yu, and Lan Xu. Media2face: Co-speech facial animation generation with multi-modality guidance. In ACM SIGGRAPH 2024 Conference Papers, pages 1-13, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.881 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1950" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_origin.pdf b/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c5e06f35b9a515ad75225132979e503713325fa8 --- /dev/null +++ b/2025/Exploring Timeline Control for Facial Motion Generation/65b84757-8067-4189-a894-056e4be1d124_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da76d4c9818e2d33fdba3e339d34d8c5411bb29519382e5d25be71d74b8b4469 +size 4229439 diff --git a/2025/Exploring Timeline Control for Facial Motion Generation/full.md b/2025/Exploring Timeline Control for Facial Motion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1e767c0dc0896f313f45b01d1bdf6ed0829bb662 --- /dev/null +++ b/2025/Exploring Timeline Control for Facial Motion Generation/full.md @@ -0,0 +1,271 @@ +# Exploring Timeline Control for Facial Motion Generation + +Yifeng Ma $^{1}$ , Jinwei Qi $^{2}$ , Chaonan Ji $^{2}$ , Peng Zhang $^{2}$ , Bang Zhang $^{2}$ , Zhidong Deng $^{1}$ , Liefeng Bo $^{2}$ $^{1}$ Department of Computer Science and Technology, Tsinghua University + $^{2}$ Tongyi Lab, Alibaba Group + +![](images/e50501a6f43d035fff80db1ff38bb2eed2754aee5a0ed9f16b35a1ad89799ed6.jpg) +Figure 1. We introduce a new control signal for facial motion generation: timeline control. We first utilize a labor-efficient approach to annotate the time intervals of facial motion at a frame-level granularity. Using the annotations, we propose a model that can generate natural facial motions aligned with an input timeline. Compared to previous controls like audio and text, timeline control enables precise temporal control of facial motions. In this paper, facial motions are rendered into photorealistic videos for better visualization. + +![](images/04eafa1a458ae0753af68bc385382fdf61400b2b1fc50d896007467810c822e7.jpg) + +# Abstract + +This paper introduces a new control signal for facial motion generation: timeline control. Compared to audio and text signals, timelines provide more fine-grained control, such as generating specific facial motions with precise timing. Users can specify a multi-track timeline of facial actions arranged in temporal intervals, allowing precise control over the timing of each action. To model the timeline control capability, We first annotate the time intervals of facial actions in natural facial motion sequences at a frame-level granularity. This process is facilitated by Toeplitz Inverse Covariance-based Clustering to minimize human labor. Based on the annotations, we propose a diffusion-based generation model capable of generating facial motions that are natural and accurately aligned with input timelines. Our method supports text-guided motion generation by using ChatGPT to convert text into timelines. Experimental results show that our method can annotate facial action intervals with satisfactory accuracy, and produces natural facial motions accurately aligned with timelines. + +# 1. Introduction + +Generating vivid facial motions has drawn growing attention due to its broad applications, including digital human generation and filmmaking. To produce desired facial motions, current methods use audio or text to provide guidance. However, these methods lack a critical ability that users often require: generating specific facial motions with precise timing. For example, users may want to generate a brow raise between frames 10 and 30 while simultaneously generating a smile between frames 14 and 43. Audiodriven methods [43, 54] can only generate motions synchronized with the audio. Text signals offer only coarse-grained descriptions of facial motions and lack frame-level detail. Text-driven approaches [4, 44, 49, 51] rely on temporal adverbial cues (e.g., then) for rough timing guidance. + +To achieve more fine-grained control, we introduce a novel control signal: timeline control for facial motion generation. In this setup, users can generate facial motions by inputting an intuitive timeline that contains several temporal intervals, each corresponding to a desired facial action. This setup enables users to manage the timing of each action. + +Achieving timeline control is very challenging. Frame + +level control of facial motions requires the model to achieve exceptional precision in generating accurate actions. To model timeline control capability, it is crucial to annotate the precise start and end frames of facial actions, a challenge that existing methods have yet to overcome. Existing methods [44, 51] represent facial motion sequences as a time series of motion descriptors, such as Facial Action Units (AUs) [10] or blendshapes [46], and then leverage ChatGPT to summarize these time series for annotation. However, ChatGPT's limited sensitivity to the temporal dynamics of facial motions prevents it from determining the exact start and end frames of facial actions. Another annotation approach is using thresholds, like labeling the eye as "closed" if the blendshape eyeBlink exceeds 0.4. However, setting thresholds for complex actions, such as brow motions, is challenging. Additionally, some actions are determined by the relative values of multiple motion descriptors rather than a single one. + +To address this issue, we adopt Toeplitz Inverse Covariance-based Clustering (TICC) [18] to annotate the start and end frames of facial actions. TICC can segment the time series into a sequence of intervals, with each interval containing a single motion pattern. Each interval has clearly defined start and end frames. Once the start and end frames of the intervals are identified, the remaining task is to determine the facial action represented by each interval. This can be addressed by another feature of TICC: its ability to group the segmented intervals into several clusters based on the similarity of their motion patterns, with each cluster containing intervals that exhibit similar motion patterns. The automatic clustering process eliminates the need for manual threshold setting and considers the relationships between multiple motion descriptors. Once these clusters are obtained, we can determine the overall action (e.g. significantly raised brows) represented by each cluster by analyzing a few representative patterns within it. This process requires minimal labor. Using this procedure, we analyze motions in different facial regions individually and achieve frame-level annotation for the brows, eyes, mouth, gaze, and head motions in a labor-efficient manner. + +Using frame-level facial motion annotations, we propose a diffusion-based generation model that produces natural facial motions accurately aligned with the input timeline. The model generates motions for each facial region (e.g. upper face, lower face) separately. This decoupling can help address the adverse impact of motion coupling in the data, thereby improving accuracy. However, some couplings are necessary for motion naturalness, such as a slight head lift during a brow raise. To properly manage coupling, the generation model is divided into a base network and multiple branch networks, with each branch network dedicated to a facial region. The base network encodes global facial motion couplings into base features. The branch network takes + +base features and the timeline of the corresponding region as input to accurately generate facial motions for that region while maintaining natural couplings. We use FaceVerse[46] 3DMM coefficients to represent facial motions for generation. Since photorealistic facial motion enables more accurate assessments of motion realism than mesh-based motion [31], we render motions into photorealistic videos by a diffusion-based renderer. + +Our method enables text-guided facial motion generation by using ChatGPT to convert text into timelines. ChatGPT effectively translates simple natural language descriptions into timelines, which our model transforms into facial motions. Users can modify these timelines, allowing precise timing control for text-driven applications. + +In summary, our contributions are as follows: + +- We are the first to develop a labor-efficient approach to annotate temporal intervals of facial actions. Such fine-grained annotations enable precise modeling of the temporal dynamics of facial motions. +- We are the first to achieve timeline control for facial motion generation. This enables users to generate specific facial motions with precise timing. Our method also supports generating actions from natural language text. +- Detailed evaluations show that our method generates accurate motion annotations and produces natural, precise facial motions from timeline inputs. + +# 2. Related Work + +Facial Motion Annotation. Compared to human motion annotation [15, 16, 29, 36], facial motion annotation is in its early stages. Existing methods [28, 44, 49, 51, 60] generate annotations that either overlook temporal changes or describe them only roughly with temporal adverbs. Some datasets [38, 56] label intervals of macro/micro expressions rather than facial actions used for facial motion generation. These datasets also require intensive manual labeling. Facial expression spotting methods [9, 17, 48, 57, 59] are also restricted to label macro/micro-expressions. In this paper, we aim to labor-efficiently generate frame-level facial action annotations. + +Facial Motion Generation. Existing methods use audio or text as the control signal and cannot generate specific motions at precise timing. Audio-driven methods (2D [6, 14, 19, 23, 26, 32, 37, 42, 43, 47, 50, 53, 54, 58, 63, 64] and 3D methods [1, 7, 8, 11, 12, 24, 25, 34, 41, 52, 55, 65]) can only generate facial motions synchronized with the audio. Text-driven methods [4, 27, 28, 44, 49, 51] can only use temporal adverbial phrases to control timing roughly. Rule-based methods [5, 33] offer temporal control but often produce unnatural motions due to deviations from real movement distributions. This paper introduces timeline control to generate natural facial motions with precise timing. + +![](images/05bf3c89485f9c1bd19fb9bdc8078a6f9b836256d59c72e733ac79156d9a2c24.jpg) +Figure 2. The pipeline of frame-level facial motion annotation (using brow motions as an example). We first extract facial motion descriptors (blendshapes) from natural facial motion videos and concatenate the results to create a facial motion time series for time series analysis. This analysis can simultaneously segment the sequence into a series of motion patterns and cluster similar patterns, resulting in multiple clusters, each containing consistent facial motion patterns. Then, by inspecting a few patterns, we identify the facial motions each cluster represents, thereby obtaining frame-level facial motion annotations for all videos. + +
Facial RegionMotion Categories
BrowBrowUp, BrowDown, Neutral
EyeEyeSquint, EyeWiden, EyeClose, Neutral
MouthSoftSmile, Smile, MouthFrown, Neutral
GazeLeft, Right, Up, Down, Neutral
HeadLeft, Right, Up, Down, Neutral
+ +Table 1. Facial actions annotated by our method. + +Timeline control has been utilized in human motion generation [2, 3, 35, 39] by first generating motion segments and then piecing them together. However, due to the rapid changes in facial movements, this approach is unsuitable, necessitating a new model structure for effective generation. + +# 3. Method + +In this work, we first annotate the temporal intervals of facial action for natural facial motion sequences. Based on these annotations, we develop a generation model that generates natural facial motions that are accurately aligned with an input timeline. Our method also supports text-guided motion generation by leveraging ChatGPT to convert text into timelines. + +# 3.1. Annotating temporal intervals of facial actions + +Problem Formulation. For each video $V$ in dataset, our method aims to generate frame-level facial motion annotations $A = [a_i]_{i=1}^L$ for each video frame $[v_i]_{i=1}^L$ . Tab. 1 shows the annotated facial actions. The annotation $a_i$ for each frame is a vector of binary values, with each dimension representing a non-neutral facial action. A dimension is set to 1 if the action occurs, otherwise, it is 0. + +Annotation Process Overview. We achieve the annotation through time series analysis TICC. Fig. 2 shows the pipeline: (1) We extract proper motion descriptors for each + +facial region to obtain facial motion time series. (2) We apply TICC to simultaneously segment the time series into a sequence of motion patterns and cluster similar patterns. This results in several clusters of intervals, each containing similar facial motions. (3) Finally, for each cluster, we can infer the actions of all intervals by examining only a few intervals. Once the action for each cluster is identified, the actions for all intervals are determined, resulting in facial action annotations at a frame-level granularity. + +Facial Motion Descriptor. We observe that previously used AUs lack sufficient precision, while ARKit blendshapes offer a more accurate representation. The publicly available MediaPipe blendshape detector struggles to detect certain blendshapes (e.g. eyeWide, mouthFrown), so we use an in-house blendshape detector. ARKit blendshapes are a set of coefficients that describe facial expressions, with each coefficient representing the movement of a specific facial region. Each coefficient ranges from 0 to 1, indicating the motion intensity. + +We use the blendshapes corresponding to each facial region to construct motion time series for each region, resulting in time series data for the eye, brow, and mouth areas. For the eye region, we selected eyeBlink, eyeSquint, eyeWide, and for the brow region, we selected browDown, browInnerUp, browOuterUp. For the mouth region, we use only expression-related blendshapes (mouthSmile, mouthStretch, mouthFrown) to avoid interference from speech-related movements. ARKit blendshapes provide coefficients for the left and right sides. As facial motions in our dataset are largely symmetrical, we use only the left-side coefficients to simplify analysis. + +For gaze and head pose motion, we use a 3DMM model FaceVerse [46] to detect coefficients. For gaze, we select eye coefficients, for head pose, we select angle coefficients. + +Analyzing Facial Motion Time Series. We utilize + +![](images/b0f9f501d79aaaa308dd00981cfc6d9d4b2e064ecc9ccc9bf9b7d1de77400e9e.jpg) +(a) Generation Model's Base-Branch Design + +![](images/625412adf6fba8cbd446dae4e043a4f7619ddd2fae6281a06c61e3d7760582c9.jpg) +(b) Base/Branch Network's Architecture +Figure 3. Illustration of generation model. (a) Base-Branch Design. The base network takes the timelines of all facial regions as input and outputs base features that model the global facial motion couplings. Through timeline selection, each region's timeline is directed to its respective branch network. Since head pose is interconnected with all facial movements, the pose branch receives timelines of all regions. Each branch network takes the timeline of its corresponding region to generate the facial motions for that region. These motions are then combined to produce the overall motion of the entire face. Lin. Proj. denotes Linear Projection. (b) Base/Branch Network's Architecture. Timeline control guides motion generation through cross-attention. The initial timeline tokens remain unchanged and are added at each layer. For clarity, the diffusion step (omitted in sub-figure (a) for clarity) is applied to each base and branch network. + +TICC [18] for time series analysis. TICC is designed to handle a single long sequence (CubeMarker [22] can analyze multiple time series analysis but performs poorly.). However, our data consists of multiple short sequences. To address this issue, we concatenate all the short sequences into a single long sequence for analysis. To avoid the impact of abrupt transitions between different videos, we introduce a "null sequence" between videos to separate them. The null sequence is implemented as a length-100 sequence with values set to -1. These null sequences will eventually be grouped into a single cluster, which will not affect others. + +After the time series analysis, the video is annotated with a sequence of motion pattern intervals and similar patterns are clustered into a few clusters. Therefore, for each cluster, we inspect a few intervals to determine the facial action they represent. This facial action is then considered to be the action for all intervals within that cluster. Different clusters may represent variations of the same facial action category, such as a high eyebrow raise versus a moderate one. We categorized the clusters into facial motion categories, such as eyebrow raise, frown, and neutral, to obtain frame-level annotations. + +We observe that eye closure frequently produces sharp spikes in the time series, which disrupts the analysis of squinting and widening. Therefore, we analyze eye closure and squinting/widening separately. We also find that threshold-based annotation, which classifies actions using predefined thresholds (e.g., labeling the eye as "closed" if eyeBlink exceeds 0.4), achieves high accuracy for eye closure, gaze, and head pose. Consequently, we use this method for these regions, while applying time series analysis for others. + +Detect Facial Motion for Unseen Videos. After learning various facial motion patterns from the data, TICC can also + +be applied to previously unseen facial motion sequences and detect facial action intervals. We use this function to assess the accuracy of facial motion generation in our evaluation. + +# 3.2. Facial Motion Generation from Timelines + +Problem Formulation. Given a timeline control $C = [c_i]_{i=1}^L$ for each video frame, the generation model $\mathcal{G}$ aims to generate natural facial motions $M = [m_l]_{l=1}^L$ that align with the timeline control. The control for each frame $c_i$ follows the same format as the annotation. Generated facial motions are rendered by a diffusion-based renderer into photorealistic videos for better visualization. + +Discussion: The coupling of facial motions and its impact on motion accuracy and naturalness. Facial motions across different regions are coupled. For example, when people smile, they might also squint and lower their eyebrows. These couplings can reduce the precision of generated motions. For instance, when generating a smile, the model may also learn to produce eyebrow-lowering actions, which could conflict with user-specified eyebrow motions (e.g. brow raise). To enhance accuracy, it is essential to decouple the generation of different facial regions. However, certain facial motion couplings are crucial for conveying naturalness—for instance, a subtle head movement that accompanies a shift in gaze. Therefore, the model must achieve a delicate balance, selectively decoupling motions to improve accuracy while preserving necessary couplings to enhance naturalness. + +Generation Model $\mathcal{G}$ . $\mathcal{G}$ is a diffusion-based model. Diffusion models consist of two Markov chains [21, 40]: the forward chain incrementally injects Gaussian noise into the original signal, while the reverse chain sequentially reconstructs the original signal from the noise. $\mathcal{G}$ predicts the original signal instead of noise, and the loss function is in + +produced as: + +$$ +\mathcal {L} _ {\text {d e n o i s e}} = \mathbb {E} _ {t \sim \mathcal {U} [ 1, T ]}, M _ {(0)}, C (\| M _ {(0)} - \mathcal {G} (M _ {(t)}, t, C) \| ^ {2}), \tag {1} +$$ + +where $t$ denotes the diffusion step, $M_{(0)}$ is the original facial motion sequence, and $M_{(t)}$ is the noisy sequence produced by the diffusion forward function $q(M_{(t)}|M_{(t - 1)}) = \mathcal{N}(\sqrt{\alpha_n} M_{(t - 1)},(1 - \alpha_n)\pmb {I})$ . $\pmb{C}$ is the timeline condition. + +To balance decoupling and coupling for both accuracy and naturalness, the generation model employs a base-branch design (Fig. 3). It consists of a base network and individual branch networks for each facial region (e.g. upper face, lower face). The base network takes timelines of all facial regions and noisy motions as input and encodes global motion couplings into base features. The branch network takes the base features the timelines relevant only to its designated facial region as inputs, and generates decoupled motions for each region. + +We use the expression, eye, and pose coefficients of FaceVerse [46] 3DMM model as target motion representation for generation. Its expression coefficients align with ARKit blendshapes, allowing separate representations of different facial regions. When dividing facial motions into different regions for separate generations, we observe that splitting them into the upper face, lower face, and pose & other regions strikes a better balance between decoupling and necessary coupling, rather than assigning a branch network for each individual region (e.g. further split eye and brow). Specifically, the upper face includes eye, brow, and gaze; the lower face includes mouth and jaw; and the pose & other regions cover the head pose and remaining regions like cheek and nose. It is important to note that, since pose is coupled with facial motions in all regions, the timeline control for the pose & other branch includes all facial regions (rather than just the control for pose). + +As for the specific implementation, the timeline control for each frame $c_{i}$ is implemented as a 16-dimensional vector composed of values 0 and 1, with each dimension representing an action. A value of 1 indicates that the action is performed in that frame, and a value of 0 indicates that the action is not present. The timeline control is transformed into timeline tokens through a linear projection. The base network and branch networks share the same structure. Their inputs consist of timeline tokens and motion tokens. The motion tokens are derived either from the noisy motion through linear projection or from base features. The motion tokens learn the temporal control from the timeline tokens through cross-attention in multiple transformer encoder layers. When incorporated into each encoder layer, the timeline tokens always use the initial timeline tokens rather than the output tokens from the previous layer. This prevents the temporal information in timeline tokens from being altered, thereby enhancing motion accuracy. The motion tokens for each layer come from the output of the motion tokens from + +the previous layer. The network's final output consists of the motion tokens from the last layer, which are then linearly projected to produce the facial motion. Each branch network generates the facial motion for its corresponding region. These motions are then combined to produce the motion for the entire face. + +To improve flexibility and generalization, we use classifier-free guidance [20] to train our model. During training, we randomly drop the condition of each facial region timeline. We use a dropout probability of 0.5 for each condition independently, with a 0.1 probability of dropping all conditions and a 0.1 probability of maintaining all conditions. When the timeline condition for a certain region is dropped, its value is set to -1. + +Rendering. We develop a diffusion-based portrait animation method to render motions as images. Since the renderer is not the main contribution of this paper, its details are reported in the Supplementary Material. + +# 3.3. Translating Natural Language to Timelines + +We observe that ChatGPT can generate a reasonable timeline from natural language descriptions, enabling our model to produce facial motions based on natural language input. The generated timeline can be further edited by users to fit their needs. To enhance the realism of the generated movements, we manually annotated several timeline descriptions as examples, allowing ChatGPT to perform few-shot learning. The details of the prompts and generated results are reported in Supplementary Material. + +# 4. Experiments + +Dataset. To effectively model natural facial motions, a dataset capturing authentic facial expressions is essential. For this purpose, we utilize the RealTalk dataset [13]. The RealTalk dataset contains 692 videos captured from genuine, unscripted conversations, resulting in highly natural facial movements with rich temporal dynamics. Previously used lab-recorded, scripted datasets, such as MEAD [45], lack these qualities. We use 1412 video clips extracted from Realtalk as the dataset. The size of the dataset is approximately 600,000 frames. + +# 4.1. Facial Motion Annotation + +Baselines. No previous methods have attempted annotating intervals of facial actions in a labor-efficient manner. We experimented with time series analysis methods such as AutoPlait [30], CubeMarker [22], and TICC. However, AutoPlait and CubeMarker failed to produce satisfactory results, as the motions of the segments within each cluster were inconsistent, making annotation impossible. Therefore, we primarily relied on TICC. Empirical results [18] indicate TICC's annotation performance can be enhanced by tuning key hyperparameters, including the number of clusters and + +![](images/00442194334990496a745553a085f0fab248c496c8122189aee9b66f78234c69.jpg) +Figure 4. An example of brow motion annotation. + +![](images/5397c53be53ed135df9f44b2aed5567accfc67d023ad8e6ec9be4459d5ec14b6.jpg) +Figure 5. An example of eye motion annotation. + +beta, a hyperparameter that facilitates temporal consistency. We examine the effects of these parameters. + +Evaluation Metric. Annotating intervals of facial actions is essentially a multi-class classification problem, where each data point represents a frame and each class represents an action. To evaluate annotation accuracy, we manually annotate the facial actions on 50 videos to serve as ground truth, and use Macro-F1 as the metric for evaluation. For each class, the F1 score is the harmonic mean of the precision and recall of our estimate. Then, the macro-F1 score is the average of the F1 scores for all the classes. A higher Macro-F1 score indicates greater annotation accuracy. We calculated the macro-F1 score for the eye region only on segments when the eyes are not closed. + +Qualitative results. Fig. 4 shows brow motion annotation. Our method can precisely segment different facial motion patterns and cluster similar ones. For a specific facial motion brow raise, our approach can distinguish multiple variants, such as a significant or moderate raise. In this work, these finer categories are grouped under one motion category for simplicity, but future methods could generate each subcategory separately. Annotation for the mouth is similar to that for the brow. Fig. 5 shows eye motion annotation. As stated in Sec. 3.1, only eye squint/widen are annotated using TICC. Eye closure is determined by a threshold on the blendshape eyeBlink and may occur within intervals of other eye motions. We also conduct a user study to evaluate annotation accuracy and the results are reported in Sec. 4.2. + +Quantitative results. We select the optimal hyperparameters based on the highest macro-F1 score. Experiments + +![](images/e0096b72dc3303fa56a17645c5b5ab0e98b86411f6dedf63d24c5393b1c02493.jpg) +Figure 6. The annotation accuracy for different number of clusters. + +![](images/a561be1bc336f3cb2533f87371732a9997c3e33bceb978b18043b3535ba1bc92.jpg) +Figure 7. The annotation accuracy in brows for different $\beta$ , concatenating videos without separating, and using AU as descriptors. + +show that the optimal beta value for different regions is 5, with the best number of clusters being 8 for the eye region, 9 for the brow region, and 9 for the mouth region. (The number of clusters here refers to the count of valid clusters, excluding the one that contain only null sequences). The optimal scores for brow, eyes, and mouth are 0.90, 0.91, and 0.87, respectively. When using threshold-based annotation, the scores for eye closure, pose, and gaze are 0.95, 0.87, and 0.89, respectively. + +The impact of different cluster numbers is shown in Fig. 6 (with beta set to 5). Too few clusters in TICC lead to under-segmentation, grouping distinct patterns into single clusters. Conversely, too many clusters decreases intercluster variability, making clusters less distinct and blurring boundaries. + +Fig. 7 illustrates the impact of $\beta$ , using brow as an example (with the number of clusters set to 8). $\beta$ controls the smoothness of segment transitions. Too large $\beta$ leads to over-smoothing, Conversely, a small $\beta$ makes the model overly sensitive, leading to noisy segments that misrepresent the data's structure. Fig. 7 shows that when using AUs, the score decreases to 0.73. This highlights the importance of accurate facial motion descriptor quality for annotation. Fig. 7 also shows that separating different video clips when analyzing improves performance. + +To conduct a preliminary test of TICC's effectiveness in detecting unseen facial motion sequences, we remove 50 manually annotated videos from the dataset, refit TICC, and then have it annotate these videos to calculate the macro-F1 score. The scores are similar to those obtained when the 50 videos are included (brow: 0.89; eyes: 0.88; mouth: 0.85). + +![](images/e8a887ffba982e6cc951eab7cc89be000a6e3c7dca1e308855d2d9ba3de6165e.jpg) +Figure 8. Qualitative results of facial motion generation from the timeline. Better viewed in Supp. Video. The timelines for the unplotted regions are set to 0. The same applies to the subsequent figures. + +![](images/95d3507b18adee8901178e274db79a145c8753efb88c030a669024210595b756.jpg) +Figure 9. Qualitative results of ablation study. Green arrows indicate subtle natural head motions, red arrows indicate missing subtle natural head motions, and yellow arrows highlight inaccurate motions. Better viewed in Supp. Video. + +# 4.2. Facial Motion Generation + +Evaluation Metric. We utilize 100 timelines unseen during training to generate samples for evaluating the motion accuracy and naturalness. To evaluate accuracy, we use TICC or threshold-based approach to annotate facial motion intervals within the generated video. These intervals are then compared to the input timeline to calculate the macro-F1 score for each facial region. The average of these scores across all regions is used to assess the alignment between + +
MethodsVar→FIDfm↓FIDΔfm↓SND ↓TAS ↑
w/o branch0.687.390.147.530.66
w/o base0.6412.40.1812.580.81
all decoup.0.4128.40.2328.630.69
w/o time con.0.715.380.105.480.79
branchL10.706.250.116.360.76
branchL30.685.760.105.860.84
branchL40.696.640.126.760.83
drop 00.626.880.137.010.78
drop 0.30.675.930.116.040.83
drop 0.70.784.120.094.210.68
Ours0.704.540.094.630.84
+ +Table 2. Results of ablation study. The unit of Var in the table is $10^{-2}$ , meaning that 0.70 in the table represents $0.70 \times 10^{-2}$ . The unit of $\mathrm{FID}_{fm}$ and $\mathrm{FID}_{\Delta fm}$ is $10^{-1}$ . Our results differ in magnitude from those of [62] because we use a different 3DMM (FaceVerse) for the calculations. $\mathbf{\nabla}^{\prime \prime} \rightarrow$ means results are better if they are close to the variance of real data, which is 0.73. + +the generated motion and the input timeline. We refer to this score as Timeline Alignment Score (TAS). To evaluate naturalness, we follow previous methods [44, 61] and use Var (variance of generated facial motions, with values closer to GT indicating better performance), $\mathbf{FID}_{fm}$ (FID score of 3DMM coefficients), $\mathbf{FID}_{\Delta fm}$ (FID score of 3DMM coefficient difference between consecutive frames), SND (sum of $\mathbf{FID}_{fm}$ and $\mathbf{FID}_{\Delta fm}$ ). + +Qualitative Results. Fig. 8 shows the results generated by our model. Our model is capable of producing natural facial movements that are aligned with the input timeline. Note that because the training data includes videos of people speaking, our generated videos sometimes include speaking. Since FaceVerse 3DMM decouples the motion of the left and right face, we can generate motion separately for each side, allowing for generating asymmetric expressions. Our method can generate forehead wrinkles when the speaker raises brows. By introducing descriptors that capture more complex texture-related motions, we can extend our method to control them. + +Quantitative Results & Ablation Study. To evaluate the impact of our design choices, we conduct an ablation study with several variants: (1) w/o branch remove all branches and only use the base network to predict the facial motions; (2) w/o base remove the base network and only use branch networks to generate each region independently; (3) all decoup. decouple all facial and head motions into different branch networks rather than limiting the decoupling to the upper and lower halves of the face. (4) w/o time con. starting from the second layer of the base/branch network, the initial timeline token input is replaced with the timeline tokens from the previous layer. We investigate the optimal number of layers for the branch network. The entire net + +![](images/21bb55d107fdb2850495767d6e8000864a8bbad61ef995b31e7948cbebe502b7.jpg) +Figure 10. User study results. + +work consists of 8 layers, with branchLx indicating that the branch network uses $x$ layers while the base network uses $8 - x$ layers. In the optimal version Ours, the branch network has 2 layers. We investigate the optimal drop probability. $\text{drop} x$ indicates the drop probability is $x$ , the optimal probability used in Ours is 0.5 + +Tab. 2 shows quantitative results and Fig. 9 shows qualitative results from a timeline. w/o branch fails to generate accurate motions, highlighting the necessity of using the branch network to enhance accuracy. When the head is oriented to the left, subtle movements may occur due to changes in gaze and brow. While these subtle motions do not affect the overall head direction, they are essential for motion naturalness. w/o base fails to generate these subtle motions (Fig. 9). Even though the pose branch receives the complete motion timeline, it still cannot produce such coupling. This highlights the necessity of using the base network to learn the coupling of facial movements. all decoup produces stiff and inaccurate motion changes (brow raised too early in Fig. 9) due to overfitting to the movements of individual regions, resulting in weak generalization to timelines outside the training set. The performance of w/o time con. indicates that adding the initial timeline token at each layer can improve accuracy. branchLx's results show that increasing the number of branch layers excessively reduces naturalness. This is because having too few base layers is insufficient to model the natural coupling between motions. drop x's results show that too low or too high drop probability both cannot generate good performance. Too low probability makes the model overly rely on the condition and hampers its generalization capability. Too high probability makes the model receive too few conditioning signals, reducing its accuracy. drop 0.7 get better FID scores, but its accuracy is low. + +Comparisons with previous methods. No prior methods have achieved fine-grained timeline control of facial motion generation. Similar approaches, such as AgentAvatar [44] and InstructAvatar [49], can only describe temporal changes coarsely using temporal adverbs. Neither of these methods provides open-source code. We can only conduct a qualitative comparison using their demos. Since the temporal + +![](images/12fc34c9961d90662c415fcc27e46bfcd56f3d99b197d9802ba5480edcb64d3e.jpg) +Figure 11. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines. + +dynamics of facial motion are better demonstrated through video, these comparisons are included in Supp. Video. + +User Study. We conduct a user study of 21 participants. The participants are required to rate three types of pairs: (1) GT: Real videos paired with manually annotated facial motion timelines. (2) Annotation: Real videos paired with timelines generated using our method. (3) Generation: Videos generated by our method paired with the input facial motion timelines. Participants need to evaluate their level of agreement (on a 5-point scale: strongly agree/disagree, agree/disagree, don't know) on two aspects: (1) Whether the motions in the video accurately match the corresponding timelines. (2) Whether the motions in the video appear natural. Each participant evaluates 15 pairs sampled from the test data for each type of pair, resulting in a total of 45 pairs per participant. Fig. 9 shows the results. $92\%$ of the evaluations agree that our annotation is accurate. $89\%$ and $86\%$ of the evaluations consider the motions in our generated videos to be accurate and natural, respectively. + +Generating Motions Using Natural Language. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines. As shown in Fig. 11, ChatGPT can infer that when a person is reminiscing, their head may turn to the side, and their gaze shifts accordingly. When a happy memory comes to mind, the person begins to smile. + +# 5. Conclusion + +In this paper, we explore a new control signal: timeline control for facial motion generation. Timeline allows for more fine-grained control than audio or text, enabling users to generate specific motions with precise timing. To model this capability, we first develop a labor-efficient approach to annotate the temporal intervals of facial actions. The annotation process relies on time series analysis of facial motion descriptors. Based on the annotations, we propose a generation model that can generate natural facial motions aligned with timelines. The generation model utilize a base-branch design to effectively manage motion couplings across different facial regions. Our method supports text-guided generation by using ChatGPT to translate text into timelines. Experiments validate the effectiveness of our method. + +# References + +[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21263-21273, 2024. 2 +[2] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 3 +[3] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 3 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2 +[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2 +[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2 +[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, pages 10101-10111, 2019. 2 +[8] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael J Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. arXiv preprint arXiv:2306.08990, 2023. 2 +[9] Yicheng Deng, Hideaki Hayashi, and Hajime Nagahara. Spotformer: Multi-scale spatio-temporal transformer for facial expression spotting. arXiv preprint arXiv:2407.20799, 2024. 2 +[10] Paul Ekman. Facial action coding system (facs). A Human Face, Salt Lake City, 2002. 2 +[11] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through a unified model. arXiv preprint arXiv:2408.00762, 2024. 2 +[12] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2 +[13] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. arXiv preprint arXiv:2301.10939, 2023. 5 + +[14] Jiazhi Guan, Zhiliang Xu, Hang Zhou, Kaisiyuan Wang, Shengyi He, Zhanwang Zhang, Borong Liang, Haocheng Feng, Errui Ding, Jingtuo Liu, et al. Resyncer: Rewiring style-based generator for unified audio-visually syncsed facial performer. arXiv preprint arXiv:2408.03284, 2024. 2 +[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021–2029, 2020. 2 +[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2 +[17] Xupeng Guo, Xiaobiao Zhang, Lei Li, and Zhaoqiang Xia. Micro-expression spotting with multi-scale local transformer in long videos. Pattern Recognition Letters, 168:146-152, 2023. 2 +[18] David Hallac, Sagar Vare, Stephen Boyd, and Jure Leskovec. Toeplitz inverse covariance-based clustering of multivariate time series data. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 215-223, 2017. 2, 4, 5 +[19] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2 +[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 4 +[22] Takato Honda, Yasuko Matsubara, Ryo Neyama, Mutsumi Abe, and Yasushi Sakurai. Multi-aspect mining of complex sensor sequences. In 2019 IEEE International Conference on Data Mining (ICDM), pages 299-308. IEEE, 2019. 4, 5 +[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2 +[24] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2 +[25] Haojie Li and Shuangping Huang. Kmtalk: Speech-driven 3d facial animation with key motion embedding. 2024. 2 +[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 2 +[27] Jun Ling, Yiwen Wang, Han Xue, Rong Xie, and Li Song. Posetalk: Text-and-audio-based pose control and motion refinement for one-shot talking head generation. arXiv preprint arXiv:2409.02657, 2024. 2 + +[28] Yifeng Ma, Suzhen Wang, Yu Ding, Bowen Ma, Tangjie Lv, Changjie Fan, Zhipeng Hu, Zhidong Deng, and Xin Yu. Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334, 2023. 2 +[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2 +[30] Yasuko Matsubara, Yasushi Sakurai, and Christos Faloutisos. Autoplait: Automatic mining of co-evolving time sequences. In Proceedings of the 2014 ACM SIGMOD international conference on Management of data, pages 193-204, 2014. 5 +[31] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[32] Federico Nocentini, Thomas Besnier, Claudio Ferrari, Sylvain Arguillere, Stefano Berretti, and Mohamed Daoudi. Scantalk: 3d talking heads from unregistered scans. arXiv preprint arXiv:2403.10942, 2024. 2 +[33] Catherine Pelachaud, Norman I Badler, and Mark Steedman. Generating facial expressions for speech. Cognitive science, 20(1):1-46, 1996. 2 +[34] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2 +[35] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1921, 2024. 3 +[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2 +[37] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2 +[38] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. $\text{Cas}(\text{me})^2$ : A database for spontaneous macro-expression and micro-expression spotting and recognition. IEEE Transactions on Affective Computing, 9 (4):424-436, 2018. 2 +[39] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 3 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4 + +[41] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2 +[42] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 2 +[43] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2 +[44] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Agentavatar: Disentangling planning, driving and rendering for photorealistic avatar agents. arXiv preprint arXiv:2311.17465, 2023. 1, 2, 7, 8 +[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, pages 700–717. Springer, 2020. 5 +[46] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20333-20342, 2022. 2, 3, 5 +[47] Suzhen Wang, Yifeng Ma, Yu Ding, Zhipeng Hu, Changjie Fan, Tangjie Lv, Zhidong Deng, and Xin Yu. Stylistalk++: A unified framework for controlling the speaking styles of talking heads. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[48] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2 +[49] Yuchi Wang, Junliang Guo, Jianhong Bai, Runyi Yu, Tianyu He, Xu Tan, Xu Sun, and Jiang Bian. Instructavator: Text-guided emotion and motion control for avatar generation. arXiv preprint arXiv:2405.15758, 2024. 1, 2, 8 +[50] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2 +[51] Sijing Wu, Yunhao Li, Yichao Yan, Huiyu Duan, Zwei Liu, and Guangtao Zhai. Mmhead: Towards fine-grained multi-modal 3d facial animation. arXiv preprint arXiv:2410.07757, 2024. 1, 2 +[52] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 2 +[53] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2 + +[54] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 2 +[55] Karren D Yang, Anurag Ranjan, Jen-Hao Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks methods and applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27294-27303, 2024. 2 +[56] Chuin Hong Yap, Connah Kendrick, and Moi Hoon Yap. Samm long videos: A spontaneous facial micro-and macro-expressions dataset. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 771-776. IEEE, 2020. 2 +[57] Chuin Hong Yap, Moi Hoon Yap, Adrian Davison, Connah Kendrick, Jingting Li, Su-Jing Wang, and Ryan Cunningham. 3d-cnn for facial micro-and macro-expression spotting on long video sequences using temporal oriented reference frame. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7016-7020, 2022. 2 +[58] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2 +[59] Jun Yu, Zhongpeng Cai, Zepeng Liu, Guochen Xie, and Peng He. Facial expression spotting based on optical flow features. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7205-7209, 2022. 2 +[60] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. CelebV-Text: A large-scale facial text-video dataset. In CVPR, 2023. 2 +[61] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 7 +[62] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7645–7655, 2023. 7 +[63] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 2 +[64] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audiodriven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2 +[65] Qingcheng Zhao, Pengyu Long, Qixuan Zhang, Dafei Qin, Han Liang, Longwen Zhang, Yingliang Zhang, Jingyi Yu, and Lan Xu. Media2face: Co-speech facial animation generation with multi-modality guidance. In ACM SIGGRAPH 2024 Conference Papers, pages 1-13, 2024. 2 \ No newline at end of file diff --git a/2025/Exploring Timeline Control for Facial Motion Generation/images.zip b/2025/Exploring Timeline Control for Facial Motion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb170f9853b75ad6ea983cdcb6f7bf78c7c1a281 --- /dev/null +++ b/2025/Exploring Timeline Control for Facial Motion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:881857138465c2ce2938b83c1fe3f20c187d6ccecb50bf112d0cb8efd729aeef +size 551341 diff --git a/2025/Exploring Timeline Control for Facial Motion Generation/layout.json b/2025/Exploring Timeline Control for Facial Motion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..000c967d9eb4b84f35659a6250800089e9ae242b --- /dev/null +++ b/2025/Exploring Timeline Control for Facial Motion Generation/layout.json @@ -0,0 +1,7469 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "Exploring Timeline Control for Facial Motion Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": "Yifeng Ma" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": ", Jinwei Qi" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": ", Chaonan Ji" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": ", Peng Zhang" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": ", Bang Zhang" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": ", Zhidong Deng" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": ", Liefeng Bo" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": " Department of Computer Science and Technology, Tsinghua University \n" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 141, + 544, + 186 + ], + "type": "text", + "content": " Tongyi Lab, Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 224, + 298, + 375 + ], + "blocks": [ + { + "bbox": [ + 57, + 224, + 298, + 375 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 298, + 375 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 298, + 375 + ], + "type": "image", + "image_path": "e50501a6f43d035fff80db1ff38bb2eed2754aee5a0ed9f16b35a1ad89799ed6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 382, + 555, + 428 + ], + "lines": [ + { + "bbox": [ + 54, + 382, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 555, + 428 + ], + "type": "text", + "content": "Figure 1. We introduce a new control signal for facial motion generation: timeline control. We first utilize a labor-efficient approach to annotate the time intervals of facial motion at a frame-level granularity. Using the annotations, we propose a model that can generate natural facial motions aligned with an input timeline. Compared to previous controls like audio and text, timeline control enables precise temporal control of facial motions. In this paper, facial motions are rendered into photorealistic videos for better visualization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 224, + 552, + 375 + ], + "blocks": [ + { + "bbox": [ + 310, + 224, + 552, + 375 + ], + "lines": [ + { + "bbox": [ + 310, + 224, + 552, + 375 + ], + "spans": [ + { + "bbox": [ + 310, + 224, + 552, + 375 + ], + "type": "image", + "image_path": "04eafa1a458ae0753af68bc385382fdf61400b2b1fc50d896007467810c822e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 438, + 200, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 438, + 200, + 451 + ], + "spans": [ + { + "bbox": [ + 151, + 438, + 200, + 451 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 472, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 297, + 700 + ], + "type": "text", + "content": "This paper introduces a new control signal for facial motion generation: timeline control. Compared to audio and text signals, timelines provide more fine-grained control, such as generating specific facial motions with precise timing. Users can specify a multi-track timeline of facial actions arranged in temporal intervals, allowing precise control over the timing of each action. To model the timeline control capability, We first annotate the time intervals of facial actions in natural facial motion sequences at a frame-level granularity. This process is facilitated by Toeplitz Inverse Covariance-based Clustering to minimize human labor. Based on the annotations, we propose a diffusion-based generation model capable of generating facial motions that are natural and accurately aligned with input timelines. Our method supports text-guided motion generation by using ChatGPT to convert text into timelines. Experimental results show that our method can annotate facial action intervals with satisfactory accuracy, and produces natural facial motions accurately aligned with timelines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 438, + 394, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 438, + 394, + 451 + ], + "spans": [ + { + "bbox": [ + 314, + 438, + 394, + 451 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "text", + "content": "Generating vivid facial motions has drawn growing attention due to its broad applications, including digital human generation and filmmaking. To produce desired facial motions, current methods use audio or text to provide guidance. However, these methods lack a critical ability that users often require: generating specific facial motions with precise timing. For example, users may want to generate a brow raise between frames 10 and 30 while simultaneously generating a smile between frames 14 and 43. Audiodriven methods [43, 54] can only generate motions synchronized with the audio. Text signals offer only coarse-grained descriptions of facial motions and lack frame-level detail. Text-driven approaches [4, 44, 49, 51] rely on temporal adverbial cues (e.g., then) for rough timing guidance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 628, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 700 + ], + "type": "text", + "content": "To achieve more fine-grained control, we introduce a novel control signal: timeline control for facial motion generation. In this setup, users can generate facial motions by inputting an intuitive timeline that contains several temporal intervals, each corresponding to a desired facial action. This setup enables users to manage the timing of each action." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "Achieving timeline control is very challenging. Frame" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1940" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "level control of facial motions requires the model to achieve exceptional precision in generating accurate actions. To model timeline control capability, it is crucial to annotate the precise start and end frames of facial actions, a challenge that existing methods have yet to overcome. Existing methods [44, 51] represent facial motion sequences as a time series of motion descriptors, such as Facial Action Units (AUs) [10] or blendshapes [46], and then leverage ChatGPT to summarize these time series for annotation. However, ChatGPT's limited sensitivity to the temporal dynamics of facial motions prevents it from determining the exact start and end frames of facial actions. Another annotation approach is using thresholds, like labeling the eye as \"closed\" if the blendshape eyeBlink exceeds 0.4. However, setting thresholds for complex actions, such as brow motions, is challenging. Additionally, some actions are determined by the relative values of multiple motion descriptors rather than a single one." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 291, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 295, + 554 + ], + "type": "text", + "content": "To address this issue, we adopt Toeplitz Inverse Covariance-based Clustering (TICC) [18] to annotate the start and end frames of facial actions. TICC can segment the time series into a sequence of intervals, with each interval containing a single motion pattern. Each interval has clearly defined start and end frames. Once the start and end frames of the intervals are identified, the remaining task is to determine the facial action represented by each interval. This can be addressed by another feature of TICC: its ability to group the segmented intervals into several clusters based on the similarity of their motion patterns, with each cluster containing intervals that exhibit similar motion patterns. The automatic clustering process eliminates the need for manual threshold setting and considers the relationships between multiple motion descriptors. Once these clusters are obtained, we can determine the overall action (e.g. significantly raised brows) represented by each cluster by analyzing a few representative patterns within it. This process requires minimal labor. Using this procedure, we analyze motions in different facial regions individually and achieve frame-level annotation for the brows, eyes, mouth, gaze, and head motions in a labor-efficient manner." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Using frame-level facial motion annotations, we propose a diffusion-based generation model that produces natural facial motions accurately aligned with the input timeline. The model generates motions for each facial region (e.g. upper face, lower face) separately. This decoupling can help address the adverse impact of motion coupling in the data, thereby improving accuracy. However, some couplings are necessary for motion naturalness, such as a slight head lift during a brow raise. To properly manage coupling, the generation model is divided into a base network and multiple branch networks, with each branch network dedicated to a facial region. The base network encodes global facial motion couplings into base features. The branch network takes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "base features and the timeline of the corresponding region as input to accurately generate facial motions for that region while maintaining natural couplings. We use FaceVerse[46] 3DMM coefficients to represent facial motions for generation. Since photorealistic facial motion enables more accurate assessments of motion realism than mesh-based motion [31], we render motions into photorealistic videos by a diffusion-based renderer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 169, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 240 + ], + "type": "text", + "content": "Our method enables text-guided facial motion generation by using ChatGPT to convert text into timelines. ChatGPT effectively translates simple natural language descriptions into timelines, which our model transforms into facial motions. Users can modify these timelines, allowing precise timing control for text-driven applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 325, + 241, + 509, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 241, + 509, + 253 + ], + "spans": [ + { + "bbox": [ + 325, + 241, + 509, + 253 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 257, + 553, + 388 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 257, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 257, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 257, + 553, + 304 + ], + "type": "text", + "content": "- We are the first to develop a labor-efficient approach to annotate temporal intervals of facial actions. Such fine-grained annotations enable precise modeling of the temporal dynamics of facial motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 304, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 304, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 315, + 304, + 553, + 353 + ], + "type": "text", + "content": "- We are the first to achieve timeline control for facial motion generation. This enables users to generate specific facial motions with precise timing. Our method also supports generating actions from natural language text." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 353, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 353, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 553, + 388 + ], + "type": "text", + "content": "- Detailed evaluations show that our method generates accurate motion annotations and produces natural, precise facial motions from timeline inputs." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 404, + 400, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 404, + 400, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 404, + 400, + 417 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 568 + ], + "type": "text", + "content": "Facial Motion Annotation. Compared to human motion annotation [15, 16, 29, 36], facial motion annotation is in its early stages. Existing methods [28, 44, 49, 51, 60] generate annotations that either overlook temporal changes or describe them only roughly with temporal adverbs. Some datasets [38, 56] label intervals of macro/micro expressions rather than facial actions used for facial motion generation. These datasets also require intensive manual labeling. Facial expression spotting methods [9, 17, 48, 57, 59] are also restricted to label macro/micro-expressions. In this paper, we aim to labor-efficiently generate frame-level facial action annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Facial Motion Generation. Existing methods use audio or text as the control signal and cannot generate specific motions at precise timing. Audio-driven methods (2D [6, 14, 19, 23, 26, 32, 37, 42, 43, 47, 50, 53, 54, 58, 63, 64] and 3D methods [1, 7, 8, 11, 12, 24, 25, 34, 41, 52, 55, 65]) can only generate facial motions synchronized with the audio. Text-driven methods [4, 27, 28, 44, 49, 51] can only use temporal adverbial phrases to control timing roughly. Rule-based methods [5, 33] offer temporal control but often produce unnatural motions due to deviations from real movement distributions. This paper introduces timeline control to generate natural facial motions with precise timing." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1941" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 547, + 228 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 547, + 228 + ], + "type": "image", + "image_path": "05bf3c89485f9c1bd19fb9bdc8078a6f9b836256d59c72e733ac79156d9a2c24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 229, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. The pipeline of frame-level facial motion annotation (using brow motions as an example). We first extract facial motion descriptors (blendshapes) from natural facial motion videos and concatenate the results to create a facial motion time series for time series analysis. This analysis can simultaneously segment the sequence into a series of motion patterns and cluster similar patterns, resulting in multiple clusters, each containing consistent facial motion patterns. Then, by inspecting a few patterns, we identify the facial motions each cluster represents, thereby obtaining frame-level facial motion annotations for all videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 293, + 293, + 373 + ], + "blocks": [ + { + "bbox": [ + 57, + 293, + 293, + 373 + ], + "lines": [ + { + "bbox": [ + 57, + 293, + 293, + 373 + ], + "spans": [ + { + "bbox": [ + 57, + 293, + 293, + 373 + ], + "type": "table", + "html": "
Facial RegionMotion Categories
BrowBrowUp, BrowDown, Neutral
EyeEyeSquint, EyeWiden, EyeClose, Neutral
MouthSoftSmile, Smile, MouthFrown, Neutral
GazeLeft, Right, Up, Down, Neutral
HeadLeft, Right, Up, Down, Neutral
", + "image_path": "b704b1959310eac8003d5eb21974bc95cb431a975d8cbf5c80d9cfa907be5b5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 373, + 266, + 384 + ], + "lines": [ + { + "bbox": [ + 86, + 373, + 266, + 384 + ], + "spans": [ + { + "bbox": [ + 86, + 373, + 266, + 384 + ], + "type": "text", + "content": "Table 1. Facial actions annotated by our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "text", + "content": "Timeline control has been utilized in human motion generation [2, 3, 35, 39] by first generating motion segments and then piecing them together. However, due to the rapid changes in facial movements, this approach is unsuitable, necessitating a new model structure for effective generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 112, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 112, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 112, + 477 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "type": "text", + "content": "In this work, we first annotate the temporal intervals of facial action for natural facial motion sequences. Based on these annotations, we develop a generation model that generates natural facial motions that are accurately aligned with an input timeline. Our method also supports text-guided motion generation by leveraging ChatGPT to convert text into timelines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 296, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 590 + ], + "type": "text", + "content": "3.1. Annotating temporal intervals of facial actions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": "Problem Formulation. For each video " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " in dataset, our method aims to generate frame-level facial motion annotations " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "A = [a_i]_{i=1}^L" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " for each video frame " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "[v_i]_{i=1}^L" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": ". Tab. 1 shows the annotated facial actions. The annotation " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " for each frame is a vector of binary values, with each dimension representing a non-neutral facial action. A dimension is set to 1 if the action occurs, otherwise, it is 0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Annotation Process Overview. We achieve the annotation through time series analysis TICC. Fig. 2 shows the pipeline: (1) We extract proper motion descriptors for each" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "type": "text", + "content": "facial region to obtain facial motion time series. (2) We apply TICC to simultaneously segment the time series into a sequence of motion patterns and cluster similar patterns. This results in several clusters of intervals, each containing similar facial motions. (3) Finally, for each cluster, we can infer the actions of all intervals by examining only a few intervals. Once the action for each cluster is identified, the actions for all intervals are determined, resulting in facial action annotations at a frame-level granularity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 521 + ], + "type": "text", + "content": "Facial Motion Descriptor. We observe that previously used AUs lack sufficient precision, while ARKit blendshapes offer a more accurate representation. The publicly available MediaPipe blendshape detector struggles to detect certain blendshapes (e.g. eyeWide, mouthFrown), so we use an in-house blendshape detector. ARKit blendshapes are a set of coefficients that describe facial expressions, with each coefficient representing the movement of a specific facial region. Each coefficient ranges from 0 to 1, indicating the motion intensity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": "We use the blendshapes corresponding to each facial region to construct motion time series for each region, resulting in time series data for the eye, brow, and mouth areas. For the eye region, we selected eyeBlink, eyeSquint, eyeWide, and for the brow region, we selected browDown, browInnerUp, browOuterUp. For the mouth region, we use only expression-related blendshapes (mouthSmile, mouthStretch, mouthFrown) to avoid interference from speech-related movements. ARKit blendshapes provide coefficients for the left and right sides. As facial motions in our dataset are largely symmetrical, we use only the left-side coefficients to simplify analysis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 700 + ], + "type": "text", + "content": "For gaze and head pose motion, we use a 3DMM model FaceVerse [46] to detect coefficients. For gaze, we select eye coefficients, for head pose, we select angle coefficients." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 701, + 555, + 714 + ], + "type": "text", + "content": "Analyzing Facial Motion Time Series. We utilize" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1942" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 76, + 354, + 194 + ], + "blocks": [ + { + "bbox": [ + 61, + 76, + 354, + 194 + ], + "lines": [ + { + "bbox": [ + 61, + 76, + 354, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 354, + 194 + ], + "type": "image", + "image_path": "b0f9f501d79aaaa308dd00981cfc6d9d4b2e064ecc9ccc9bf9b7d1de77400e9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 201, + 292, + 210 + ], + "lines": [ + { + "bbox": [ + 157, + 201, + 292, + 210 + ], + "spans": [ + { + "bbox": [ + 157, + 201, + 292, + 210 + ], + "type": "text", + "content": "(a) Generation Model's Base-Branch Design" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 357, + 74, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 357, + 74, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 357, + 74, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 357, + 74, + 547, + 196 + ], + "type": "image", + "image_path": "625412adf6fba8cbd446dae4e043a4f7619ddd2fae6281a06c61e3d7760582c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 201, + 536, + 209 + ], + "lines": [ + { + "bbox": [ + 414, + 201, + 536, + 209 + ], + "spans": [ + { + "bbox": [ + 414, + 201, + 536, + 209 + ], + "type": "text", + "content": "(b) Base/Branch Network's Architecture" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 214, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 555, + 291 + ], + "type": "text", + "content": "Figure 3. Illustration of generation model. (a) Base-Branch Design. The base network takes the timelines of all facial regions as input and outputs base features that model the global facial motion couplings. Through timeline selection, each region's timeline is directed to its respective branch network. Since head pose is interconnected with all facial movements, the pose branch receives timelines of all regions. Each branch network takes the timeline of its corresponding region to generate the facial motions for that region. These motions are then combined to produce the overall motion of the entire face. Lin. Proj. denotes Linear Projection. (b) Base/Branch Network's Architecture. Timeline control guides motion generation through cross-attention. The initial timeline tokens remain unchanged and are added at each layer. For clarity, the diffusion step (omitted in sub-figure (a) for clarity) is applied to each base and branch network." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 432 + ], + "type": "text", + "content": "TICC [18] for time series analysis. TICC is designed to handle a single long sequence (CubeMarker [22] can analyze multiple time series analysis but performs poorly.). However, our data consists of multiple short sequences. To address this issue, we concatenate all the short sequences into a single long sequence for analysis. To avoid the impact of abrupt transitions between different videos, we introduce a \"null sequence\" between videos to separate them. The null sequence is implemented as a length-100 sequence with values set to -1. These null sequences will eventually be grouped into a single cluster, which will not affect others." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 566 + ], + "type": "text", + "content": "After the time series analysis, the video is annotated with a sequence of motion pattern intervals and similar patterns are clustered into a few clusters. Therefore, for each cluster, we inspect a few intervals to determine the facial action they represent. This facial action is then considered to be the action for all intervals within that cluster. Different clusters may represent variations of the same facial action category, such as a high eyebrow raise versus a moderate one. We categorized the clusters into facial motion categories, such as eyebrow raise, frown, and neutral, to obtain frame-level annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 568, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 686 + ], + "type": "text", + "content": "We observe that eye closure frequently produces sharp spikes in the time series, which disrupts the analysis of squinting and widening. Therefore, we analyze eye closure and squinting/widening separately. We also find that threshold-based annotation, which classifies actions using predefined thresholds (e.g., labeling the eye as \"closed\" if eyeBlink exceeds 0.4), achieves high accuracy for eye closure, gaze, and head pose. Consequently, we use this method for these regions, while applying time series analysis for others." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Detect Facial Motion for Unseen Videos. After learning various facial motion patterns from the data, TICC can also" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 337 + ], + "type": "text", + "content": "be applied to previously unseen facial motion sequences and detect facial action intervals. We use this function to assess the accuracy of facial motion generation in our evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 531, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 531, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 531, + 358 + ], + "type": "text", + "content": "3.2. Facial Motion Generation from Timelines" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "text", + "content": "Problem Formulation. Given a timeline control " + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "C = [c_i]_{i=1}^L" + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "text", + "content": " for each video frame, the generation model " + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "text", + "content": " aims to generate natural facial motions " + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "M = [m_l]_{l=1}^L" + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "text", + "content": " that align with the timeline control. The control for each frame " + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 365, + 555, + 449 + ], + "type": "text", + "content": " follows the same format as the annotation. Generated facial motions are rendered by a diffusion-based renderer into photorealistic videos for better visualization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 640 + ], + "type": "text", + "content": "Discussion: The coupling of facial motions and its impact on motion accuracy and naturalness. Facial motions across different regions are coupled. For example, when people smile, they might also squint and lower their eyebrows. These couplings can reduce the precision of generated motions. For instance, when generating a smile, the model may also learn to produce eyebrow-lowering actions, which could conflict with user-specified eyebrow motions (e.g. brow raise). To enhance accuracy, it is essential to decouple the generation of different facial regions. However, certain facial motion couplings are crucial for conveying naturalness—for instance, a subtle head movement that accompanies a shift in gaze. Therefore, the model must achieve a delicate balance, selectively decoupling motions to improve accuracy while preserving necessary couplings to enhance naturalness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Generation Model " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " is a diffusion-based model. Diffusion models consist of two Markov chains [21, 40]: the forward chain incrementally injects Gaussian noise into the original signal, while the reverse chain sequentially reconstructs the original signal from the noise. " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " predicts the original signal instead of noise, and the loss function is in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1943" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 107, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 107, + 83 + ], + "type": "text", + "content": "produced as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 90, + 295, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 90, + 295, + 114 + ], + "spans": [ + { + "bbox": [ + 59, + 90, + 295, + 114 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e n o i s e}} = \\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ]}, M _ {(0)}, C (\\| M _ {(0)} - \\mathcal {G} (M _ {(t)}, t, C) \\| ^ {2}), \\tag {1}", + "image_path": "6cc145d0bf2c255318d10cc46415631497beb71a4f5ce710b0a677999b38b3fa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "text", + "content": " denotes the diffusion step, " + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "inline_equation", + "content": "M_{(0)}" + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "text", + "content": " is the original facial motion sequence, and " + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "inline_equation", + "content": "M_{(t)}" + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "text", + "content": " is the noisy sequence produced by the diffusion forward function " + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "inline_equation", + "content": "q(M_{(t)}|M_{(t - 1)}) = \\mathcal{N}(\\sqrt{\\alpha_n} M_{(t - 1)},(1 - \\alpha_n)\\pmb {I})" + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{C}" + }, + { + "bbox": [ + 55, + 116, + 295, + 163 + ], + "type": "text", + "content": " is the timeline condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 164, + 295, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 164, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 164, + 295, + 282 + ], + "type": "text", + "content": "To balance decoupling and coupling for both accuracy and naturalness, the generation model employs a base-branch design (Fig. 3). It consists of a base network and individual branch networks for each facial region (e.g. upper face, lower face). The base network takes timelines of all facial regions and noisy motions as input and encodes global motion couplings into base features. The branch network takes the base features the timelines relevant only to its designated facial region as inputs, and generates decoupled motions for each region." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 283, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 486 + ], + "type": "text", + "content": "We use the expression, eye, and pose coefficients of FaceVerse [46] 3DMM model as target motion representation for generation. Its expression coefficients align with ARKit blendshapes, allowing separate representations of different facial regions. When dividing facial motions into different regions for separate generations, we observe that splitting them into the upper face, lower face, and pose & other regions strikes a better balance between decoupling and necessary coupling, rather than assigning a branch network for each individual region (e.g. further split eye and brow). Specifically, the upper face includes eye, brow, and gaze; the lower face includes mouth and jaw; and the pose & other regions cover the head pose and remaining regions like cheek and nose. It is important to note that, since pose is coupled with facial motions in all regions, the timeline control for the pose & other branch includes all facial regions (rather than just the control for pose)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "text", + "content": "As for the specific implementation, the timeline control for each frame " + }, + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "text", + "content": " is implemented as a 16-dimensional vector composed of values 0 and 1, with each dimension representing an action. A value of 1 indicates that the action is performed in that frame, and a value of 0 indicates that the action is not present. The timeline control is transformed into timeline tokens through a linear projection. The base network and branch networks share the same structure. Their inputs consist of timeline tokens and motion tokens. The motion tokens are derived either from the noisy motion through linear projection or from base features. The motion tokens learn the temporal control from the timeline tokens through cross-attention in multiple transformer encoder layers. When incorporated into each encoder layer, the timeline tokens always use the initial timeline tokens rather than the output tokens from the previous layer. This prevents the temporal information in timeline tokens from being altered, thereby enhancing motion accuracy. The motion tokens for each layer come from the output of the motion tokens from" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "the previous layer. The network's final output consists of the motion tokens from the last layer, which are then linearly projected to produce the facial motion. Each branch network generates the facial motion for its corresponding region. These motions are then combined to produce the motion for the entire face." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 239 + ], + "type": "text", + "content": "To improve flexibility and generalization, we use classifier-free guidance [20] to train our model. During training, we randomly drop the condition of each facial region timeline. We use a dropout probability of 0.5 for each condition independently, with a 0.1 probability of dropping all conditions and a 0.1 probability of maintaining all conditions. When the timeline condition for a certain region is dropped, its value is set to -1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 239, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 287 + ], + "type": "text", + "content": "Rendering. We develop a diffusion-based portrait animation method to render motions as images. Since the renderer is not the main contribution of this paper, its details are reported in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 294, + 539, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 539, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 539, + 306 + ], + "type": "text", + "content": "3.3. Translating Natural Language to Timelines" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 419 + ], + "type": "text", + "content": "We observe that ChatGPT can generate a reasonable timeline from natural language descriptions, enabling our model to produce facial motions based on natural language input. The generated timeline can be further edited by users to fit their needs. To enhance the realism of the generated movements, we manually annotated several timeline descriptions as examples, allowing ChatGPT to perform few-shot learning. The details of the prompts and generated results are reported in Supplementary Material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 394, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 394, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 394, + 443 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 568 + ], + "type": "text", + "content": "Dataset. To effectively model natural facial motions, a dataset capturing authentic facial expressions is essential. For this purpose, we utilize the RealTalk dataset [13]. The RealTalk dataset contains 692 videos captured from genuine, unscripted conversations, resulting in highly natural facial movements with rich temporal dynamics. Previously used lab-recorded, scripted datasets, such as MEAD [45], lack these qualities. We use 1412 video clips extracted from Realtalk as the dataset. The size of the dataset is approximately 600,000 frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 576, + 457, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 457, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 457, + 587 + ], + "type": "text", + "content": "4.1. Facial Motion Annotation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Baselines. No previous methods have attempted annotating intervals of facial actions in a labor-efficient manner. We experimented with time series analysis methods such as AutoPlait [30], CubeMarker [22], and TICC. However, AutoPlait and CubeMarker failed to produce satisfactory results, as the motions of the segments within each cluster were inconsistent, making annotation impossible. Therefore, we primarily relied on TICC. Empirical results [18] indicate TICC's annotation performance can be enhanced by tuning key hyperparameters, including the number of clusters and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1944" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 292, + 183 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 292, + 183 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 292, + 183 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 292, + 183 + ], + "type": "image", + "image_path": "00442194334990496a745553a085f0fab248c496c8122189aee9b66f78234c69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 186, + 266, + 197 + ], + "lines": [ + { + "bbox": [ + 84, + 186, + 266, + 197 + ], + "spans": [ + { + "bbox": [ + 84, + 186, + 266, + 197 + ], + "type": "text", + "content": "Figure 4. An example of brow motion annotation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 206, + 287, + 319 + ], + "blocks": [ + { + "bbox": [ + 59, + 206, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 59, + 206, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 59, + 206, + 287, + 319 + ], + "type": "image", + "image_path": "5397c53be53ed135df9f44b2aed5567accfc67d023ad8e6ec9be4459d5ec14b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 322, + 263, + 332 + ], + "lines": [ + { + "bbox": [ + 87, + 322, + 263, + 332 + ], + "spans": [ + { + "bbox": [ + 87, + 322, + 263, + 332 + ], + "type": "text", + "content": "Figure 5. An example of eye motion annotation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 294, + 374 + ], + "type": "text", + "content": "beta, a hyperparameter that facilitates temporal consistency. We examine the effects of these parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 376, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 518 + ], + "type": "text", + "content": "Evaluation Metric. Annotating intervals of facial actions is essentially a multi-class classification problem, where each data point represents a frame and each class represents an action. To evaluate annotation accuracy, we manually annotate the facial actions on 50 videos to serve as ground truth, and use Macro-F1 as the metric for evaluation. For each class, the F1 score is the harmonic mean of the precision and recall of our estimate. Then, the macro-F1 score is the average of the F1 scores for all the classes. A higher Macro-F1 score indicates greater annotation accuracy. We calculated the macro-F1 score for the eye region only on segments when the eyes are not closed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "content": "Qualitative results. Fig. 4 shows brow motion annotation. Our method can precisely segment different facial motion patterns and cluster similar ones. For a specific facial motion brow raise, our approach can distinguish multiple variants, such as a significant or moderate raise. In this work, these finer categories are grouped under one motion category for simplicity, but future methods could generate each subcategory separately. Annotation for the mouth is similar to that for the brow. Fig. 5 shows eye motion annotation. As stated in Sec. 3.1, only eye squint/widen are annotated using TICC. Eye closure is determined by a threshold on the blendshape eyeBlink and may occur within intervals of other eye motions. We also conduct a user study to evaluate annotation accuracy and the results are reported in Sec. 4.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Quantitative results. We select the optimal hyperparameters based on the highest macro-F1 score. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 548, + 171 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 548, + 171 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 548, + 171 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 548, + 171 + ], + "type": "image", + "image_path": "e0096b72dc3303fa56a17645c5b5ab0e98b86411f6dedf63d24c5393b1c02493.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 186 + ], + "type": "text", + "content": "Figure 6. The annotation accuracy for different number of clusters." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 196, + 548, + 294 + ], + "blocks": [ + { + "bbox": [ + 317, + 196, + 548, + 294 + ], + "lines": [ + { + "bbox": [ + 317, + 196, + 548, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 196, + 548, + 294 + ], + "type": "image", + "image_path": "a561be1bc336f3cb2533f87371732a9997c3e33bceb978b18043b3535ba1bc92.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 299, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 313, + 299, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 553, + 322 + ], + "type": "text", + "content": "Figure 7. The annotation accuracy in brows for different " + }, + { + "bbox": [ + 313, + 299, + 553, + 322 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 299, + 553, + 322 + ], + "type": "text", + "content": ", concatenating videos without separating, and using AU as descriptors." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 339, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 447 + ], + "type": "text", + "content": "show that the optimal beta value for different regions is 5, with the best number of clusters being 8 for the eye region, 9 for the brow region, and 9 for the mouth region. (The number of clusters here refers to the count of valid clusters, excluding the one that contain only null sequences). The optimal scores for brow, eyes, and mouth are 0.90, 0.91, and 0.87, respectively. When using threshold-based annotation, the scores for eye closure, pose, and gaze are 0.95, 0.87, and 0.89, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 448, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 518 + ], + "type": "text", + "content": "The impact of different cluster numbers is shown in Fig. 6 (with beta set to 5). Too few clusters in TICC lead to under-segmentation, grouping distinct patterns into single clusters. Conversely, too many clusters decreases intercluster variability, making clusters less distinct and blurring boundaries." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": "Fig. 7 illustrates the impact of " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": ", using brow as an example (with the number of clusters set to 8). " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": " controls the smoothness of segment transitions. Too large " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": " leads to over-smoothing, Conversely, a small " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": " makes the model overly sensitive, leading to noisy segments that misrepresent the data's structure. Fig. 7 shows that when using AUs, the score decreases to 0.73. This highlights the importance of accurate facial motion descriptor quality for annotation. Fig. 7 also shows that separating different video clips when analyzing improves performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "To conduct a preliminary test of TICC's effectiveness in detecting unseen facial motion sequences, we remove 50 manually annotated videos from the dataset, refit TICC, and then have it annotate these videos to calculate the macro-F1 score. The scores are similar to those obtained when the 50 videos are included (brow: 0.89; eyes: 0.88; mouth: 0.85)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1945" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 293, + 231 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 293, + 231 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 293, + 231 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 293, + 231 + ], + "type": "image", + "image_path": "e8a887ffba982e6cc951eab7cc89be000a6e3c7dca1e308855d2d9ba3de6165e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 275 + ], + "type": "text", + "content": "Figure 8. Qualitative results of facial motion generation from the timeline. Better viewed in Supp. Video. The timelines for the unplotted regions are set to 0. The same applies to the subsequent figures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 284, + 293, + 535 + ], + "blocks": [ + { + "bbox": [ + 55, + 284, + 293, + 535 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 293, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 293, + 535 + ], + "type": "image", + "image_path": "95d3507b18adee8901178e274db79a145c8753efb88c030a669024210595b756.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 536, + 295, + 580 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 580 + ], + "type": "text", + "content": "Figure 9. Qualitative results of ablation study. Green arrows indicate subtle natural head motions, red arrows indicate missing subtle natural head motions, and yellow arrows highlight inaccurate motions. Better viewed in Supp. Video." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 598, + 200, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 200, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 200, + 609 + ], + "type": "text", + "content": "4.2. Facial Motion Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Evaluation Metric. We utilize 100 timelines unseen during training to generate samples for evaluating the motion accuracy and naturalness. To evaluate accuracy, we use TICC or threshold-based approach to annotate facial motion intervals within the generated video. These intervals are then compared to the input timeline to calculate the macro-F1 score for each facial region. The average of these scores across all regions is used to assess the alignment between" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 70, + 549, + 216 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 549, + 216 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 549, + 216 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 549, + 216 + ], + "type": "table", + "html": "
MethodsVar→FIDfm↓FIDΔfm↓SND ↓TAS ↑
w/o branch0.687.390.147.530.66
w/o base0.6412.40.1812.580.81
all decoup.0.4128.40.2328.630.69
w/o time con.0.715.380.105.480.79
branchL10.706.250.116.360.76
branchL30.685.760.105.860.84
branchL40.696.640.126.760.83
drop 00.626.880.137.010.78
drop 0.30.675.930.116.040.83
drop 0.70.784.120.094.210.68
Ours0.704.540.094.630.84
", + "image_path": "297e3a4f5555df9646a5167435b1b126e9bd233d817928c33f9a165673cc7410.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "text", + "content": "Table 2. Results of ablation study. The unit of Var in the table is " + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "text", + "content": ", meaning that 0.70 in the table represents " + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "inline_equation", + "content": "0.70 \\times 10^{-2}" + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "text", + "content": ". The unit of " + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{fm}" + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "inline_equation", + "content": "10^{-1}" + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "text", + "content": ". Our results differ in magnitude from those of [62] because we use a different 3DMM (FaceVerse) for the calculations. " + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{\\nabla}^{\\prime \\prime} \\rightarrow" + }, + { + "bbox": [ + 313, + 226, + 555, + 293 + ], + "type": "text", + "content": " means results are better if they are close to the variance of real data, which is 0.73." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "content": "the generated motion and the input timeline. We refer to this score as Timeline Alignment Score (TAS). To evaluate naturalness, we follow previous methods [44, 61] and use Var (variance of generated facial motions, with values closer to GT indicating better performance), " + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{fm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "content": " (FID score of 3DMM coefficients), " + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "content": " (FID score of 3DMM coefficient difference between consecutive frames), SND (sum of " + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{fm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 413, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 555 + ], + "type": "text", + "content": "Qualitative Results. Fig. 8 shows the results generated by our model. Our model is capable of producing natural facial movements that are aligned with the input timeline. Note that because the training data includes videos of people speaking, our generated videos sometimes include speaking. Since FaceVerse 3DMM decouples the motion of the left and right face, we can generate motion separately for each side, allowing for generating asymmetric expressions. Our method can generate forehead wrinkles when the speaker raises brows. By introducing descriptors that capture more complex texture-related motions, we can extend our method to control them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Quantitative Results & Ablation Study. To evaluate the impact of our design choices, we conduct an ablation study with several variants: (1) w/o branch remove all branches and only use the base network to predict the facial motions; (2) w/o base remove the base network and only use branch networks to generate each region independently; (3) all decoup. decouple all facial and head motions into different branch networks rather than limiting the decoupling to the upper and lower halves of the face. (4) w/o time con. starting from the second layer of the base/branch network, the initial timeline token input is replaced with the timeline tokens from the previous layer. We investigate the optimal number of layers for the branch network. The entire net" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1946" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 303, + 194 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 303, + 194 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 303, + 194 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 303, + 194 + ], + "type": "image", + "image_path": "21bb55d107fdb2850495767d6e8000864a8bbad61ef995b31e7948cbebe502b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 202, + 230, + 213 + ], + "lines": [ + { + "bbox": [ + 120, + 202, + 230, + 213 + ], + "spans": [ + { + "bbox": [ + 120, + 202, + 230, + 213 + ], + "type": "text", + "content": "Figure 10. User study results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "text", + "content": "work consists of 8 layers, with branchLx indicating that the branch network uses " + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "text", + "content": " layers while the base network uses " + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "inline_equation", + "content": "8 - x" + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "text", + "content": " layers. In the optimal version Ours, the branch network has 2 layers. We investigate the optimal drop probability. " + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\text{drop} x" + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "text", + "content": " indicates the drop probability is " + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 217, + 295, + 289 + ], + "type": "text", + "content": ", the optimal probability used in Ours is 0.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 293, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 626 + ], + "type": "text", + "content": "Tab. 2 shows quantitative results and Fig. 9 shows qualitative results from a timeline. w/o branch fails to generate accurate motions, highlighting the necessity of using the branch network to enhance accuracy. When the head is oriented to the left, subtle movements may occur due to changes in gaze and brow. While these subtle motions do not affect the overall head direction, they are essential for motion naturalness. w/o base fails to generate these subtle motions (Fig. 9). Even though the pose branch receives the complete motion timeline, it still cannot produce such coupling. This highlights the necessity of using the base network to learn the coupling of facial movements. all decoup produces stiff and inaccurate motion changes (brow raised too early in Fig. 9) due to overfitting to the movements of individual regions, resulting in weak generalization to timelines outside the training set. The performance of w/o time con. indicates that adding the initial timeline token at each layer can improve accuracy. branchLx's results show that increasing the number of branch layers excessively reduces naturalness. This is because having too few base layers is insufficient to model the natural coupling between motions. drop x's results show that too low or too high drop probability both cannot generate good performance. Too low probability makes the model overly rely on the condition and hampers its generalization capability. Too high probability makes the model receive too few conditioning signals, reducing its accuracy. drop 0.7 get better FID scores, but its accuracy is low." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Comparisons with previous methods. No prior methods have achieved fine-grained timeline control of facial motion generation. Similar approaches, such as AgentAvatar [44] and InstructAvatar [49], can only describe temporal changes coarsely using temporal adverbs. Neither of these methods provides open-source code. We can only conduct a qualitative comparison using their demos. Since the temporal" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 157 + ], + "type": "image", + "image_path": "12fc34c9961d90662c415fcc27e46bfcd56f3d99b197d9802ba5480edcb64d3e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 159, + 553, + 181 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 181 + ], + "type": "text", + "content": "Figure 11. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 198, + 553, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 222 + ], + "type": "text", + "content": "dynamics of facial motion are better demonstrated through video, these comparisons are included in Supp. Video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "text", + "content": "User Study. We conduct a user study of 21 participants. The participants are required to rate three types of pairs: (1) GT: Real videos paired with manually annotated facial motion timelines. (2) Annotation: Real videos paired with timelines generated using our method. (3) Generation: Videos generated by our method paired with the input facial motion timelines. Participants need to evaluate their level of agreement (on a 5-point scale: strongly agree/disagree, agree/disagree, don't know) on two aspects: (1) Whether the motions in the video accurately match the corresponding timelines. (2) Whether the motions in the video appear natural. Each participant evaluates 15 pairs sampled from the test data for each type of pair, resulting in a total of 45 pairs per participant. Fig. 9 shows the results. " + }, + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "text", + "content": " of the evaluations agree that our annotation is accurate. " + }, + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "inline_equation", + "content": "86\\%" + }, + { + "bbox": [ + 313, + 224, + 554, + 426 + ], + "type": "text", + "content": " of the evaluations consider the motions in our generated videos to be accurate and natural, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 428, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 511 + ], + "type": "text", + "content": "Generating Motions Using Natural Language. Our method supports text-guided facial motion generation by leveraging ChatGPT to convert text into timelines. As shown in Fig. 11, ChatGPT can infer that when a person is reminiscing, their head may turn to the side, and their gaze shifts accordingly. When a happy memory comes to mind, the person begins to smile." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we explore a new control signal: timeline control for facial motion generation. Timeline allows for more fine-grained control than audio or text, enabling users to generate specific motions with precise timing. To model this capability, we first develop a labor-efficient approach to annotate the temporal intervals of facial actions. The annotation process relies on time series analysis of facial motion descriptors. Based on the annotations, we propose a generation model that can generate natural facial motions aligned with timelines. The generation model utilize a base-branch design to effectively manage motion couplings across different facial regions. Our method supports text-guided generation by using ChatGPT to translate text into timelines. Experiments validate the effectiveness of our method." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1947" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21263-21273, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "type": "text", + "content": "[3] German Barquero, Sergio Escalera, and Cristina Palmero. Seamless human motion composition with blended positional encodings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 457-469, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 301, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 295, + 378 + ], + "type": "text", + "content": "[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "type": "text", + "content": "[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, pages 10101-10111, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "text", + "content": "[8] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael J Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. arXiv preprint arXiv:2306.08990, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "type": "text", + "content": "[9] Yicheng Deng, Hideaki Hayashi, and Hajime Nagahara. Spotformer: Multi-scale spatio-temporal transformer for facial expression spotting. arXiv preprint arXiv:2407.20799, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 557, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 579 + ], + "type": "text", + "content": "[10] Paul Ekman. Facial action coding system (facs). A Human Face, Salt Lake City, 2002. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 623 + ], + "type": "text", + "content": "[11] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through a unified model. arXiv preprint arXiv:2408.00762, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 679 + ], + "type": "text", + "content": "[12] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[13] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. arXiv preprint arXiv:2301.10939, 2023. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Jiazhi Guan, Zhiliang Xu, Hang Zhou, Kaisiyuan Wang, Shengyi He, Zhanwang Zhang, Borong Liang, Haocheng Feng, Errui Ding, Jingtuo Liu, et al. Resyncer: Rewiring style-based generator for unified audio-visually syncsed facial performer. arXiv preprint arXiv:2408.03284, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 128, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 128, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 553, + 183 + ], + "type": "text", + "content": "[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021–2029, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "text", + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 238, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 553, + 281 + ], + "type": "text", + "content": "[17] Xupeng Guo, Xiaobiao Zhang, Lei Li, and Zhaoqiang Xia. Micro-expression spotting with multi-scale local transformer in long videos. Pattern Recognition Letters, 168:146-152, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 282, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 337 + ], + "type": "text", + "content": "[18] David Hallac, Sagar Vare, Stephen Boyd, and Jure Leskovec. Toeplitz inverse covariance-based clustering of multivariate time series data. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 215-223, 2017. 2, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[19] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "text", + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "text", + "content": "[22] Takato Honda, Yasuko Matsubara, Ryo Neyama, Mutsumi Abe, and Yasushi Sakurai. Multi-aspect mining of complex sensor sequences. In 2019 IEEE International Conference on Data Mining (ICDM), pages 299-308. IEEE, 2019. 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 503, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 548 + ], + "type": "text", + "content": "[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[24] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "text", + "content": "[25] Haojie Li and Shuangping Huang. Kmtalk: Speech-driven 3d facial animation with key motion embedding. 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 614, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 667 + ], + "type": "text", + "content": "[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[27] Jun Ling, Yiwen Wang, Han Xue, Rong Xie, and Li Song. Posetalk: Text-and-audio-based pose control and motion refinement for one-shot talking head generation. arXiv preprint arXiv:2409.02657, 2024. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1948" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[28] Yifeng Ma, Suzhen Wang, Yu Ding, Bowen Ma, Tangjie Lv, Changjie Fan, Zhipeng Hu, Zhidong Deng, and Xin Yu. Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "text", + "content": "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[30] Yasuko Matsubara, Yasushi Sakurai, and Christos Faloutisos. Autoplait: Automatic mining of co-evolving time sequences. In Proceedings of the 2014 ACM SIGMOD international conference on Management of data, pages 193-204, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "text", + "content": "[31] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[32] Federico Nocentini, Thomas Besnier, Claudio Ferrari, Sylvain Arguillere, Stefano Berretti, and Mohamed Daoudi. Scantalk: 3d talking heads from unregistered scans. arXiv preprint arXiv:2403.10942, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "text", + "content": "[33] Catherine Pelachaud, Norman I Badler, and Mark Steedman. Generating facial expressions for speech. Cognitive science, 20(1):1-46, 1996. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 432 + ], + "type": "text", + "content": "[34] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[35] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1921, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "text", + "content": "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[37] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[38] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. " + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "inline_equation", + "content": "\\text{Cas}(\\text{me})^2" + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": ": A database for spontaneous macro-expression and micro-expression spotting and recognition. IEEE Transactions on Affective Computing, 9 (4):424-436, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[39] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[41] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[42] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "type": "text", + "content": "[43] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[44] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Agentavatar: Disentangling planning, driving and rendering for photorealistic avatar agents. arXiv preprint arXiv:2311.17465, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "text", + "content": "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, pages 700–717. Springer, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 380 + ], + "type": "text", + "content": "[46] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20333-20342, 2022. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "type": "text", + "content": "[47] Suzhen Wang, Yifeng Ma, Yu Ding, Zhipeng Hu, Changjie Fan, Tangjie Lv, Zhidong Deng, and Xin Yu. Stylistalk++: A unified framework for controlling the speaking styles of talking heads. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "type": "text", + "content": "[48] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[49] Yuchi Wang, Junliang Guo, Jianhong Bai, Runyi Yu, Tianyu He, Xu Tan, Xu Sun, and Jiang Bian. Instructavator: Text-guided emotion and motion control for avatar generation. arXiv preprint arXiv:2405.15758, 2024. 1, 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[50] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[51] Sijing Wu, Yunhao Li, Yichao Yan, Huiyu Duan, Zwei Liu, and Guangtao Zhai. Mmhead: Towards fine-grained multi-modal 3d facial animation. arXiv preprint arXiv:2410.07757, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "type": "text", + "content": "[52] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[53] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1949" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 697 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[54] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[55] Karren D Yang, Anurag Ranjan, Jen-Hao Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks methods and applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27294-27303, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 250 + ], + "type": "text", + "content": "[56] Chuin Hong Yap, Connah Kendrick, and Moi Hoon Yap. Samm long videos: A spontaneous facial micro-and macro-expressions dataset. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 771-776. IEEE, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 317 + ], + "type": "text", + "content": "[57] Chuin Hong Yap, Moi Hoon Yap, Adrian Davison, Connah Kendrick, Jingting Li, Su-Jing Wang, and Ryan Cunningham. 3d-cnn for facial micro-and macro-expression spotting on long video sequences using temporal oriented reference frame. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7016-7020, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "text", + "content": "[58] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "type": "text", + "content": "[59] Jun Yu, Zhongpeng Cai, Zepeng Liu, Guochen Xie, and Peng He. Facial expression spotting based on optical flow features. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7205-7209, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 409, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 440 + ], + "type": "text", + "content": "[60] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. CelebV-Text: A large-scale facial text-video dataset. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "text", + "content": "[61] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "type": "text", + "content": "[62] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7645–7655, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "type": "text", + "content": "[63] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 642 + ], + "type": "text", + "content": "[64] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audiodriven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 643, + 294, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 643, + 294, + 697 + ], + "spans": [ + { + "bbox": [ + 56, + 643, + 294, + 697 + ], + "type": "text", + "content": "[65] Qingcheng Zhao, Pengyu Long, Qixuan Zhang, Dafei Qin, Han Liang, Longwen Zhang, Yingliang Zhang, Jingyi Yu, and Lan Xu. Media2face: Co-speech facial animation generation with multi-modality guidance. In ACM SIGGRAPH 2024 Conference Papers, pages 1-13, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1950" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_content_list.json b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..287a240acb68fb78ba99114fdc6527c9d7dabb2b --- /dev/null +++ b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_content_list.json @@ -0,0 +1,1407 @@ +[ + { + "type": "text", + "text": "Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models", + "text_level": 1, + "bbox": [ + 142, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuyang Hao $^{1}$ Bryan Hooi $^{2}$ Jun Liu $^{3}$ Kai-Wei Chang $^{4}$ Zi Huang $^{5}$ Yujun Cai $^{5*}$ Southeast University, $^{2}$ National University of Singapore, $^{3}$ Lancaster University $^{4}$ University of California, Los Angeles, $^{5}$ University of Queensland haosy@seu.edu.cn, dcsbhk@nus.edu.sg, j.liu81@lancaster.ac.uk, kw@kwchang.net, huang@itee.uq.edu.au, yujun.cai@uq.edu.au", + "bbox": [ + 132, + 200, + 869, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 305, + 326, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite inheriting security measures from underlying language models, Vision-Language Models (VLMs) may still be vulnerable to safety alignment issues. Through empirical analysis, we uncover two critical findings: scenario-matched images can significantly amplify harmful outputs, and contrary to common assumptions in gradient-based attacks, minimal loss values do not guarantee optimal attack effectiveness. Building on these insights, we introduce MLAI (Multi-Loss Adversarial Images), a novel jailbreak framework that leverages scenario-aware image generation for semantic alignment, exploits flat minima theory for robust adversarial image selection, and employs multi-image collaborative attacks for enhanced effectiveness. Extensive experiments demonstrate MLAI's significant impact, achieving attack success rates of $77.75\\%$ on MiniGPT-4 and $82.80\\%$ on LLaVA-2, substantially outperforming existing methods by margins of $34.37\\%$ and $12.77\\%$ respectively. Furthermore, MLAI shows considerable transferability to commercial black-box VLMs, achieving up to $60.11\\%$ success rate. Our work reveals fundamental visual vulnerabilities in current VLMs safety mechanisms and underscores the need for stronger defenses. Warning: This paper contains potentially harmful example text.", + "bbox": [ + 88, + 338, + 483, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 714, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) have demonstrated remarkable capabilities in understanding and generating multimodal content, enabling various applications from content creation to visual reasoning. To prevent potential misuse, these models typically undergo comprehensive safety protocols [23, 30] including safety-focused pre-training and reinforcement learning with human feedback (RLHF) [15, 28, 33]. Despite these safeguards, ensuring the safe deployment of these increasingly powerful models remains a criti-", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0125f4c2e3c819b2eb7f0430bee864c98b8447df6902dfc100cf6fac7f4334be.jpg", + "image_caption": [ + "Figure 1. An example to show the Visual Vulnerabilities of the scenario-matched images on alignment of VLMs and the unreliability of using image with minimal loss (the cross-entropy loss between the model's output and the target in gradient-based optimization). We can find that: (1) matching images are better than irrelevant or no images, and (2) since only the last jailbreak is successful, lower loss is not always better." + ], + "image_footnote": [], + "bbox": [ + 519, + 305, + 903, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal challenge, as recent studies [10, 18, 29, 32, 37] have revealed that visual inputs can significantly compromise these", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19890", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "safety mechanisms.", + "bbox": [ + 89, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The vulnerability of VLMs to visual inputs is particularly concerning as these models are increasingly deployed in both open-source and commercial platforms. In this paper, we investigate jailbreak attacks that aim to circumvent VLMs' safety mechanisms through carefully crafted adversarial images. This vulnerability is particularly critical as VLMs are increasingly deployed in both open-source and commercial platforms. Our study primarily focuses on white-box [19, 40] settings where model parameters are accessible, with additional evaluation on commercial blackbox VLMs [5, 7, 11] to assess real-world implications.", + "bbox": [ + 89, + 108, + 483, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For white-box attacks, current approaches [18, 29, 37] typically rely on gradient-based optimization of adversarial images. These methods iteratively refine random noise patterns by minimizing the cross-entropy loss between the model's generated response and the desired harmful output. Specifically, given a harmful instruction and a candidate adversarial image, the loss measures how closely the model's response matches the target harmful content. Through systematic analysis, we identify two critical limitations in this paradigm. As illustrated in Fig. 1, first, adversarial images with minimal loss values do not necessarily lead to successful attacks, challenging a fundamental assumption in gradient-based optimization. When repeating attacks with different initial images, we observe significant variations in success rates, suggesting that loss minimization alone is not a reliable strategy. Second, using a single image for attacking all scenarios is not an effective approach, as the harmfulness of the VLM's response is strongly correlated with the semantic relevance of the image content to the target scenario, which is also demonstrated in previous work [18].", + "bbox": [ + 89, + 276, + 483, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these insights, we propose a novel jailbreak method called MLAI, which emphasizes targeted image generation and the use of Multi-Loss Adversarial Images to amplify harmful intent and disrupt multi-modal alignment. Specifically, our method introduces a three-stage attack strategy. First, MLAI generates initial images that match the target scenario. In this step, we generate the image with the highest potential for harmful output in the specific scenario while ensuring that the adversarial image remains visually natural, making the model more susceptible to generating unsafe responses. Second, MLAI optimizes the adversarial image through gradient updates and saves results that fall within a predefined loss range during the iterative process. In the process of iterative refinement, the harmfulness of the image is progressively amplified. Third, MLAI carefully calculate the loss range and select adversarial images within the loss range for collaborative attack. The determination of the loss range takes into account the balance between attack success rate and computational cost.", + "bbox": [ + 89, + 580, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interestingly, MLAI demonstrates considerable transferability across different models. Our experiments show", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that adversarial images trained on white-box models maintain significant attack success rates when applied to unseen black-box commercial VLMs. This transferability suggests fundamental vulnerabilities in current VLM safety mechanisms that transcend specific model architectures and training approaches. To mitigate this problem, we propose a similarity-based dedduplication defense that effectively mitigates MLAI's impact by limiting the number of similar adversarial input images.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are as follows:", + "bbox": [ + 531, + 227, + 861, + 241 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We undertake a comprehensive analysis of the gradient-based attack process. Through targeted experiments, we explore the selection of initial images and examine the outcomes of adversarial images with varying loss values. The findings highlight that strategies relying exclusively on the adversarial image with the lowest loss or using a single initial image are suboptimal, offering new insights into attack effectiveness.", + "- We introduce a novel jailbreak method, MLAI, which leverages images that match the target scenario along with a collaborative approach involving multi-loss adversarial images. This combined strategy amplifies the harmful intent of the attack, successfully disrupting multi-modal alignment in VLMs.", + "- Experimental results show that MLAI achieves a remarkable attack success rate (ASR) of $77.75\\%$ on MiniGPT-4 and $82.80\\%$ on LLaVA-2, making existing white-box VLMs highly vulnerable. Additionally, we demonstrate the potential of MLAI against black-box commercial VLMs and propose a countermeasure to defend against these advanced attacks." + ], + "bbox": [ + 513, + 242, + 903, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 571, + 653, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Language Models (VLMs). Leveraging the remarkable capabilities of LLMs, the field has seen the emergence of several prominent VLMs. Notable among these are LLaVA [19] and MiniGPT-4 [40], which integrate an LLM with a vision encoder and a cross-modal connector. These models adopt a two-stage training framework, comprising cross-modal representation alignment and visual instruction tuning. Building upon this foundational approach, a variety of open-source VLMs [20] have been developed, which enhance multimodal performance through several key advancements: the use of higher-resolution inputs, more powerful LLM architectures, expanded pretraining corpora [2], and the introduction of specialized pretraining tasks [2, 6]. These efforts have significantly improved the models' ability to process and generate content across modalities, further advancing the state of multimodal understanding and generation.", + "bbox": [ + 511, + 598, + 906, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alignment Vulnerabilities in VLMs. Similar to LLMs, VLMs are also vulnerable to malicious inputs, which can", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19891", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lead to the generation of harmful content [29, 38]. However, existing benchmarks predominantly focus on evaluating the effectiveness of VLMs, with insufficient attention given to their alignment vulnerabilities, particularly regarding harmlessness [9, 16, 24]. To address this gap, recent research has developed red-teaming attack benchmarks that systematically assess the alignment vulnerabilities of VLMs across various scenarios [17, 24, 35]. In line with LLM research, studies on jailbreaking VLMs can be divided into white-box, and black-box attack strategies. For white-box attacks, existing methods [3, 8, 29, 31, 32, 34] can be further categorized based on their target: input images or visual embeddings. Attacks on input images often generate adversarial images constrained to provoke harmful responses [8, 29, 31] or use teacher-forcing optimization [3]. For visual embeddings, adversarial images are crafted to appear harmless but mimic the embeddings of harmful images [32], thus bypassing content filters. On the other hand, black-box attacks typically undermine the alignment of VLMs by techniques such as system prompt attacks [4, 36], converting harmful information into text-oriented images [10], leveraging surrogate models to generate adversarial images [39], or utilizing maximum likelihood-based jailbreaking methods [25]. Our work extends this line of research by systematically analyzing the visual-semantic vulnerabilities in VLM safety alignment and introducing novel attack strategies that leverage scenario-matched images and flat minima characteristics.", + "bbox": [ + 89, + 90, + 483, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 527, + 264, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 551, + 202, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies [10, 18, 29, 32, 37] have revealed that VLMs, despite inheriting safety measures from their foundational LLMs, remain vulnerable to jailbreak attacks. Given a vision-language model built based on a LLM M, an image encoder $E$ and a projection layer $W$ , its generation process can be formulated as:", + "bbox": [ + 89, + 574, + 483, + 665 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = \\mathbf {M} ([ W \\cdot E (i), t ]), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 676, + 482, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $i$ and $t$ are the input image and text, and $y$ is the model's output. Through empirical analysis of gradient-based attack methods, we observe two key limitations: (1) arbitrarily selected initial images fail to fully exploit the correlation between image content and model output harmfulness, and (2) our experiments show that adversarial images with minimal loss do not necessarily lead to the highest attack success rate.", + "bbox": [ + 89, + 704, + 483, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these challenges, we propose MLAI, a novel jailbreak framework that enhances attack effectiveness through scenario-aware image generation and multi-image collaborative attack. As illustrated in Fig. 3, MLAI first generates initial images that match the target scenario,", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d2e31a66865086b8353a7c44f8c5e00e2dfd90fd12439c43774aa07a0ae69a74.jpg", + "image_caption": [ + "Figure 2. The ASR of transferability across scenarios. The heat map shows that images generated under the IA scenario achieve high transferability when applied to MG contexts, while FR scenario images transfer effectively to LO settings. These transferability patterns are bidirectional between paired scenarios. In contrast, PO scenario images show minimal transfer effectiveness across other contexts." + ], + "image_footnote": [], + "bbox": [ + 573, + 87, + 851, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "then obtains a set of adversarial images through gradient optimization, and finally utilizes multiple images within an optimal loss range for the collaborative attack. The complete attack process can be formulated as:", + "bbox": [ + 511, + 390, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny ^ {*} = \\sum_ {k} ^ {\\Theta} \\mathbf {M} (\\left[ W \\cdot E (I _ {t} ^ {a d v, k}), t \\right]), \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 463, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y^{*}$ is the attack generation, $I_{t}^{adv,k}$ represents the k-th adversarial image and $\\Theta$ denotes the selected loss range.", + "bbox": [ + 511, + 517, + 906, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Scenario-aware Image Generation", + "text_level": 1, + "bbox": [ + 511, + 559, + 815, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through training on human preference data, existing LLMs [15, 28, 33] have learned to align with human values and refuse to respond to harmful text inputs. However, recent studies [18] indicate that incorporating images into inputs can significantly increase the likelihood of generating harmful content, and the harmfulness of VLM outputs is often positively correlated with the harmfulness of the visual content. This observation suggests that visual components introduce additional vulnerabilities, which can be exploited to circumvent the alignment of the underlying LLMs.", + "bbox": [ + 511, + 583, + 906, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By comparing the transferability of generated images across different scenarios, we further validate the importance of semantic matching between images and attack targets. As shown in Fig. 2, images generated under certain scenarios demonstrate high transferability to semantically related contexts. For instance, instructions within Illegal Activity (IA) and Malware Generation (MG) scenarios share common harmful action patterns like \"creating\" or \"utilizing\", leading to effective cross-scenario transfers. Similarly, Fraud (FR) and Legal Opinion (LO) scenarios often involve abstract harmful concepts like \"concealing\"", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19892", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d8a4e932007eee462732a713a262cdbe14f427acb5dc3505926d210b66ffb3ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 89, + 500, + 188 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3bbd64101760aa6daeb6a68f554d5e4abc795adb9a87499d1b2c794ef435e23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 195, + 498, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a4b2e2df1c879b35f0e7e7afeb40d14e020be11b59ca0f9120a5c7302d8cac7.jpg", + "image_caption": [ + "Figure 3. Our MLAI framework involves a three-step procedure: (1) generate an image that matches the text scenario as initial image, (2) obtain the adversarial image set based on the initial image by gradient update, and (3) calculate the loss range and select adversarial images within the loss range for collaborative attack." + ], + "image_footnote": [], + "bbox": [ + 109, + 380, + 496, + 563 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6b2434129c3d7a81b01c5bc66874756e731174728dd7cd9bbf34606b790c6b34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 890, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "or \"deceiving\". In contrast, scenarios with distinct semantic frameworks, such as Pornography (PO), show minimal transfer effectiveness, highlighting the crucial role of semantic alignment in attack success.", + "bbox": [ + 89, + 643, + 485, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by these findings, we propose to generate initial images that match the semantic content of jailbreak targets. Given that jailbreak-related scenarios often involve abstract concepts that are challenging to accurately depict, we first classify jailbreak targets $t$ into 13 distinct categories $C = \\{c_{1}, c_{2}, \\dots, c_{13}\\}$ based on common harmful behaviors (e.g., Illegal Activity, Hate Speech and Fraud). For each category, we carefully craft prompts that encode contextual information while avoiding direct depiction of harmful elements. For example, when targeting fraud-related scenarios, we generate images of relevant but benign contexts like business meetings or document processing.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These carefully crafted prompts are then used with DALL-E 3 [26] to generate the initial images:", + "bbox": [ + 511, + 643, + 906, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} = G (P (t)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 681, + 906, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(\\cdot)$ generates the scenario-specific prompt and $G(\\cdot)$ represents the image generation model. Through this design, the generated images establish strong semantic connections with the jailbreak targets while maintaining a proper balance between attack effectiveness and content safety. Our subsequent experiments in Sec. 4.3 demonstrate that such scenario-aware initial images significantly improve the success rate of gradient-based attacks compared to randomly selected images.", + "bbox": [ + 511, + 703, + 906, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adversarial Image Optimization", + "text_level": 1, + "bbox": [ + 511, + 847, + 802, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the scenario-aware initial image $I_{t}$ , we further optimize it through gradient updates to enhance the attack effec", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19893", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a81ad4b175b629cd621173ed4ec51d9b84bc2d8139997379a1ec865338a6e15.jpg", + "image_caption": [ + "(a) ASR of different Loss." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 285, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e706e8e2656b10bcc9583fb276927fee48da97a4cf9255a65a9fe6cb347ae25f.jpg", + "image_caption": [ + "(b) ASR of Min/Multi-Loss." + ], + "image_footnote": [], + "bbox": [ + 290, + 95, + 482, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77e0a0f2c07284bc2ce7f5a20e2f57c316f89318e34ab876457a29086670a652.jpg", + "image_caption": [ + "Figure 4. Results of the effect of various Loss on ASR. We can clearly observe the limitations of the minimum loss strategy.", + "Figure 5. Comparison of the effects of flat and sharp minima at test set shift. Here the test set shift is simulated by curve translation." + ], + "image_footnote": [], + "bbox": [ + 102, + 280, + 475, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tiveness. First, we construct a small corpus $Y := \\{y_k\\}_{k=1}^m$ consisting of harmful statements and affirmative responses, which serves as target outputs for optimization. For each batch of harmful instructions $t'$ and corresponding adversarial images $I_t^{adv}$ , we compute the cross-entropy loss between model outputs and the target responses:", + "bbox": [ + 89, + 470, + 483, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} ^ {a d v, k} = \\arg \\min _ {\\widehat {I} _ {t} ^ {a d v} \\in \\mathbb {R}} \\sum_ {k = 1} ^ {m} - l o g \\left(p _ {\\varphi} \\left(y _ {k} \\mid t _ {k} ^ {\\prime}, \\widehat {I} _ {t} ^ {a d v, k}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 571, + 482, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{R}$ is some constraint applied to the input space in which we search for adversarial examples, and $p_{\\varphi}$ represents the conditional probability generated by the target VLM. To ensure the adversarial images remain visually natural, we constrain the perturbation magnitude by $\\left\\| I_t^{adv,k} - I_t\\right\\|_\\infty \\leq \\varepsilon$ where $\\varepsilon$ is set to 32/255.", + "bbox": [ + 89, + 619, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During extensive experiments with this optimization process, we observe that adversarial images with minimal loss values do not necessarily lead to successful attacks. Specifically, we find cases where images with relatively low loss fail to jailbreak the target VLM, while those with moderately higher loss values achieve better success rates. This observation persists across different models and attack scenarios.", + "bbox": [ + 89, + 718, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-image Collaborative Attack", + "text_level": 1, + "bbox": [ + 89, + 847, + 387, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better understand the observation that adversarial images with minimal loss do not necessarily lead to the highest at", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tack success rate, we conduct systematic experiments analyzing attack performance across different loss ranges. As shown in Fig. 4a, our results reveal that the optimal attack performance often comes from adversarial images within a certain loss range, rather than those with minimal loss. This finding aligns with the flat minima phenomenon [12] in optimization, where solutions in flatter regions of the loss landscape tend to be more robust than those in sharp minima. As illustrated in Fig. 5, when the loss landscape exhibits a sharp minimum, even small perturbations can lead to significant performance degradation. In contrast, solutions in flat regions are more resilient to input variations, which is particularly relevant for adversarial attacks where the test-time scenario may differ from the optimization objective.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this insight, we propose to select adversarial images within a carefully determined loss range for collaborative attack. The selection process starts by analyzing the loss landscape around the minimal loss point. Specifically, we select 20 nearest points on each side of the minimum and fit straight lines using least squares method. The loss range $\\Theta$ is then determined by:", + "bbox": [ + 511, + 316, + 903, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta = \\left(\\left\\lceil \\frac {k _ {l}}{K} \\right\\rceil \\times \\rho , \\left\\lceil \\frac {k _ {r}}{K} \\right\\rceil \\times \\rho\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 431, + 903, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k_{l}$ and $k_{r}$ are the fitted slopes representing the local geometry of the loss landscape, $K$ is the standard slope derived from normal distribution, and $\\rho$ is a range coefficient set to 6. As demonstrated in Fig. 4b, this approach adaptively selects a narrow range in flat regions for computational efficiency, while expanding the range in sharp regions for better robustness.", + "bbox": [ + 511, + 476, + 905, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final jailbreak output is produced through the collaborative effort of multiple adversarial images within $\\Theta$ , as formulated in Eq. (2). This strategy enables us to maintain both attack effectiveness and computational efficiency.", + "bbox": [ + 511, + 582, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 656, + 638, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 681, + 702, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and Datasets. To comprehensively evaluate MLAI's effectiveness, we conduct experiments on both open-source and commercial VLMs. For open-source models, we focus on MiniGPT-4 [40] (Vicuna 13B) and LLaVA-2 [19] (13B) due to their widespread adoption and strong performance. We use the official weights provided in their respective repositories. For commercial models, we evaluate on Gemini [11], ChatGLM [5], and Qwen [7] to validate real-world applicability. We evaluate our approaches using two common datasets: SafetyBench [22] and AdvBench [41]. SafetyBench serves as a benchmark for assessing the safety of VLMs, encompassing 13 typically prohibited scenarios or behaviors as outlined in", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19894", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3615682832fc2c4ea4048e0a8b7da9face3378db3fe564a078a85ac63af7746d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScenarioMiniGPT-4 [40]LLaVA-2 [19]
Plain TextLiu et al. [21]Qi et al. [29]OursPlain TextLiu et al. [21]Qi et al. [29]Ours
Illegal Activity (IA)1.9214.5411.5570.243.8274.7748.0383.73
Hate Speech (HS)1.6811.923.9766.573.0840.9740.2180.36
Malware Generation (MG)3.3219.8815.5250.2830.7564.6353.8579.10
Physical Harm (PH)2.9824.5123.4372.0213.8861.0450.3179.62
Economic Harm (EH)5.684.916.8070.154.1518.427.9850.18
Fraud (FR)3.1718.5614.7150.674.9757.2047.9775.85
Pornography (PO)4.1420.9419.1169.0120.6461.2341.5568.98
Political Lobbying (PL)67.6779.1176.3396.7474.5991.1585.6293.45
Privacy Violence (PV)8.9710.5012.9778.9220.9249.3637.3374.82
Legal Opinion (LO)74.5685.2882.4895.1680.1695.4682.9493.59
Financial Advice (FA)84.3388.1290.9310087.4897.2187.6996.77
Health Consultation (HC)76.5093.9491.2293.0985.6110090.04100
Gov Decision (GD)90.2991.7591.2597.8587.7498.9791.22100
Average32.7143.3841.5677.7539.8370.0358.8382.80
", + "bbox": [ + 94, + 88, + 903, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. ASR results of jailbreaking MiniGPT4 and LLaVA-2 in 13 scenarios. Our attack achieves the best attacking performance in both modules and almost every scenario. We achieve improvements of $34.37\\%$ and $12.77\\%$ over existing state-of-the-art approaches on MiniGPT-4 and LLaVA-2, respectively. Additionally, substantial ASR differences can be observed across scenarios, with notable variations in performance improvement across different contexts.", + "bbox": [ + 89, + 333, + 906, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the usage policies [13, 27] of OpenAI [1] and Meta [42]. AdvBench has been extensively utilized in prior research concerning LLM jailbreak attacks and includes 521 harmful behaviors. We following the setting of BAP [37], removed duplicate entries from AdvBench and integrated each unique item into SafetyBench according to the corresponding scenario for our experiments.", + "bbox": [ + 89, + 415, + 483, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric. We following the setting of HADES [18], employ Attack Success Rate (ASR) as our primary metric to assess the effectiveness of our approach, which is calculated by:", + "bbox": [ + 89, + 537, + 483, + 597 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA S R = \\frac {\\sum_ {k = 1} ^ {N} \\mathbf {B} (J (y _ {k}) = T r u e)}{N}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 604, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y_{k}$ is the model's response, $\\mathbf{B}$ is an indicator function that equals to 1 if $J(y_{k}) = True$ and 0 otherwise, $N$ is the total number of instructions and $J(\\cdot)$ is the harmfulness judging model, outputting True or False to indicate whether $y_{k}$ is harmful. We adopt Beaver-dam-7B [14] as $J(\\cdot)$ , which has been trained on high-quality human feedback data about the above harmful categories.", + "bbox": [ + 89, + 643, + 483, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared attacks. We compare MLAI with two state-of-the-art jailbreak attacks: Liu et al. [21] and Qi et al. [29]. Liu et al. [21] integrated images associated with aggressive intent alongside typographic text to serve as visual adversarial prompts. Qi et al. [29], on the other hand, refined these visual adversarial prompts by leveraging a corpus specific to certain scenarios. Additionally, we include a \"Plain Text\" baseline where harmful queries are directly input to evaluate models' base vulnerability.", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiment Results", + "text_level": 1, + "bbox": [ + 513, + 415, + 702, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "White-box Attacks. Our experimental evaluation focuses on three key aspects: attack effectiveness across models, vulnerability patterns in different scenarios, and performance variations among scenarios.", + "bbox": [ + 511, + 440, + 906, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results in Tab. 1 show that MLAI achieves strong attack performance on both tested models. For MiniGPT-4 [40], MLAI reaches $77.75\\%$ ASR, surpassing the previous best method $(43.38\\%)$ by $34.37\\%$ . Similarly on LLaVA-2 [19], MLAI achieves $82.80\\%$ ASR, outperforming existing approaches by $12.77\\%$ .", + "bbox": [ + 511, + 503, + 905, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis across scenarios reveals distinct vulnerability patterns in both models. Security-critical scenarios like Malware Generation (MG) and Fraud (FR) show stronger resistance with ASR of $50.28\\%$ and $50.67\\%$ , while advisory scenarios such as Financial Advice (FA) and Gov Decision (GD) exhibit higher vulnerability $(\\mathrm{ASR} > 90\\%)$ . This pattern reflects the models' internal defense mechanisms: security-critical scenarios are equipped with more rigorous detection systems due to their well-defined harmful patterns, and these high-risk scenarios are typically protected by enhanced safety measures.", + "bbox": [ + 509, + 595, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effectiveness of MLAI varies significantly across scenarios. Most notably, while achieving a $55.70\\%$ improvement in Illegal Activity (IA) scenario, the gain in Gov Decision (GD) scenario is merely $0.85\\%$ . This disparity can be primarily explained by two aspects. First, scenarios like Gov Decision (GD) already demonstrate high base vulnerability $(90.29\\%)$ ASR without attacks), leaving limited room for improvement. Second, the nature of different scenarios affects our attack strategy: Illegal Activity (IA) scenario", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19895", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/804a39b90fa6c2d1d4bece21c81c2a68765a61a32e08032f8908ec83a182ea48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScenarioGemini [11]ChatGLM [5]Qwen [7]
Illegal Activity49.5960.1154.97
Malware Generation39.8442.5645.28
Physical Harm43.3545.6248.46
", + "bbox": [ + 96, + 88, + 478, + 154 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. ASR results for black-box models. The ASR differences across various models within the same scenario are minor, and the ranking of ASR aligns closely with those observed during testing on white-box models.", + "bbox": [ + 89, + 164, + 482, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "involve concrete elements (e.g., drug manufacturing) that can be effectively represented in images, whereas Gov Decision (GD) scenario contain abstract concepts that are inherently challenging to visualize, thereby limiting the potential of image-based attacks.", + "bbox": [ + 89, + 234, + 483, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Black-box Attacks. Besides the open-source models, we also evaluate our attack on three black-box commercial VLMs (i.e., Gemini [11], ChatGLM [5] and Qwen [7]). We selected 3 most harmful scenarios (Illegal Activity, Malware Generation, and Physical Harm) and 60 related queries to attack each model. Since direct access to the internal data of the model is unavailable, we employ images trained on MiniGPT-4 [40] to perform a black-box model jailbreak. As shown in Tab. 2, MLAI maintains considerable attack effectiveness on commercial models, achieving ASRs of $49.59\\%$ , $60.11\\%$ , and $54.97\\%$ on Gemini, ChatGLM, and Qwen respectively. We also notice that compared to white-box attacks, there exists a consistent performance drop ( $16.43\\%$ lower on average). This gap likely stems from commercial models' enhanced defense mechanisms and architectural differences. Nevertheless, the consistent attack performance across different commercial models suggests that our method can effectively generalize beyond the training model, despite the presence of unknown defense strategies.", + "bbox": [ + 91, + 324, + 483, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 635, + 254, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this part, we further discuss the effectiveness of our proposed approach, including image harmfulness optimization and loss range determination. We note that except for the explicitly stated modifications, all other parameters are identical to the baseline configuration.", + "bbox": [ + 89, + 659, + 482, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image harmfulness optimization. To validate the effectiveness of the optimization process for image generation discussed in Sec. 3.2, we conduct a specific experiment with selecting different types of images to examine the attack performance. We selected two scenarios, Hate Speech and Fraud, and conducted experiments on the MiniGPT-4 [40] and LLaVA-2 [19] models, respectively. Moreover, the Irrelevant Image used in the experiments are fixed and consist of 12 images that is not associated with any of the jailbreak scenarios. As shown in Fig. 6, when blank images", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1c07cdd7ffc11e1e5ea53691b5a8cc2e25fb9217723e3932ea2b2d902f30f7c.jpg", + "image_caption": [ + "(a) ASR of Fraud." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 709, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c05aa229a841b7e3777d92510e539a571923261bc60e87416ce9e754ca45af9.jpg", + "image_caption": [ + "(b) ASR of Hate Speech." + ], + "image_footnote": [], + "bbox": [ + 714, + 88, + 905, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed976887c3dcd4706752a5ab8f67807ab290d7393a26b7f99f9156b01e1dce97.jpg", + "image_caption": [ + "Figure 6. The ASR of using different types of initial images. We can find that simply adding a blank image can lead to a substantial increase in the success rate. Compared to blank image, further selecting images that contain information but are unrelated to the jailbreak scenario results in a smaller improvement. However, when images are introduced that closely match the specific scenario, the performance reaches state-of-the-art levels.", + "Figure 7. The ASR of using different numbers of images. According to Eq. (5), since $K$ is a constant value, we set $v = \\left\\lceil \\frac{k_{L}}{K} \\right\\rceil$ and assume that the slopes on both sides are equal without loss of generality, where $v$ represents the flatness of the minimum value, in order to better focus on the changing trends of ASR." + ], + "image_footnote": [], + "bbox": [ + 571, + 342, + 851, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are introduced, the alignment of the underlying LLM fails to cover the unforeseen domains introduced by the visual modality. This misalignment effectively circumvents the inherent alignment mechanisms of the LLM within the VLM framework, leading to a marked increase in the ASR. However, when choosing unrelated images, for the target scene, it merely increases the information content and complexity of the image, which does not significantly aid in bypassing restrictions. Only by selecting images that match the scenario can meaningful semantic information be introduced, thereby amplifying the potential risk of the image and ultimately further increasing the ASR.", + "bbox": [ + 511, + 626, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss range determination. To further validate the impact of using images in different loss range on the improvement of jailbreak success rate, we conduct experiments on four scenarios on the MiniGPT-4 [40] model and the results are shown in Fig. 7. We observe that, overall, the ASR increases", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19896", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b27b1cce159879bbbae84bd425c0a879a97c55e623f840580405628c4b1ad60c.jpg", + "image_caption": [ + "Figure 8. The average ASR of using different $\\rho$ . The average ASR increases gradually with the value of $\\rho$ , with the rate of increase remaining relatively stable. However, once $\\rho$ reaches a critical threshold (which is 5 in this experiment), the growth rate drops significantly, and the ASR quickly reaches its maximum value." + ], + "image_footnote": [], + "bbox": [ + 147, + 88, + 423, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "significantly when the number of Images initially increases. However, the growth patterns of the curves in different scenarios show considerable variation, with differences in both the rate of increase and the point at which the growth stops. Specifically, the values of $v$ for Illegal Activity (IA) and Hate Speech (HS) are relatively large, indicating that they have relatively sharp minima, whereas Legal Opinion (LO) and Financial Advice (FA) exhibit flatter minima. Correspondingly, the curves for Illegal Activity (IA) and Hate Speech (HS) stop increasing early (almost cease to grow when the number of images reaches 5), while the curves for Legal Opinion (LO) and Financial Advice (FA) maintain a prolonged growth phase (only stopping near the point where the number of images reaches 9). The results demonstrate that to achieve an optimal ASR, image sets with flat minima require fewer images, while image sets with sharp minima incur higher costs to achieve the desired outcome. These findings prove that flatter minima indeed provide greater robustness, allowing models to perform better when the test set deviates from the training set.", + "bbox": [ + 91, + 369, + 483, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we also examined the values of $\\rho$ . We tested the average ASR across all scenarios for different values of $\\rho$ on both MiniGPT-4 [40] and LLaVA-2 [19]. As shown in Fig. 8, the average ASR reaches its near-maximum value when $\\rho = 6$ , while the computational complexity increases at a linear rate. Further increases in $\\rho$ yield diminishing returns. Therefore, based on empirical observations, we set the baseline value of $\\rho = 6$ . Through the above experiments, we can prove the effectiveness of the flat minimum, the optimal choice of $\\rho$ , and we finally find the balance point between the attack success rate and the computational cost.", + "bbox": [ + 89, + 672, + 483, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Defense Analysis", + "text_level": 1, + "bbox": [ + 89, + 847, + 256, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the perspective of developmental trends, defense mechanisms always occupy the weaker position. With the", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af9cfd232eaf701e8b6d9df2d60aea6839da0e0fa17f18bcbbe6858c000b5683.jpg", + "image_caption": [ + "(a) Results of 10 specific questions in fraud scenario.", + "Figure 9. The demonstration of our defense motivation and results." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 707, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb1fa30aeba1e9f044177f294863bd17d3e3373707d1b4d5b345d131164efce9.jpg", + "image_caption": [ + "(b) ASR of original/defensive circumstance." + ], + "image_footnote": [], + "bbox": [ + 712, + 88, + 905, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "emergence of an increasing number of VLMs and the continuous improvement of jailbreak methods, comprehensive defense becomes progressively more challenging. In this section, we analyze an effective defense strategy against Multi-Loss Adversarial Images attack.", + "bbox": [ + 511, + 287, + 905, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 9a, we observe that under the same scenario, images with different loss not only vary in overall attack potency but also demonstrate significant differences in their effectiveness against individual queries. Based on this finding, we propose introducing a similarity detection module before image sets are input into VLMs. If a current image is detected to be highly similar to a previously encountered image, it is removed from the input set, thus preventing it from being fed into the VLMs. This approach transforms a Multi-Loss Adversarial Attack back into a Single-Loss Adversarial Attack, and due to the diversity of queries in the scenario, the average ASR achieves robust defense. We conducted experiments on MiniGPT-4 [40], and the results, as shown in Fig. 9b, indicate that our defense method effectively reduces the success rate of the Multi-Loss Adversarial Attack, achieving a $22.99\\%$ reduction compared to the model without the defense module.", + "bbox": [ + 511, + 363, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we conducted a systematic analysis of visual vulnerabilities in VLMs safety alignment and introduced MLAI, a novel jailbreak framework that effectively exploits these weaknesses. Our investigation revealed two critical findings: the unreliability of minimal-loss optimization in gradient-based attacks, and the crucial role of semantic alignment between visual content and target scenarios. Based on these insights, MLAI achieves superior performance through scenario-aware image generation, flat minima-based optimization, and multi-image collaborative attacks. Extensive experiments demonstrate MLAI's effectiveness, achieving attack success rates of $77.75\\%$ on MiniGPT-4 and $82.80\\%$ on LLaVA-2. The strong transferability to commercial black-box VLMs further highlights fundamental challenges in achieving robust safety alignment for multi-modal systems.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19897", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research has been partially supported by Australian Research Council Discovery Project DP230101196.", + "bbox": [ + 89, + 114, + 483, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 186 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sam Altman and et al. OpenAI, 2015. 6", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2", + "[3] Nicholas Carlini, Milad Nasr, Christopher A. Choquette-Choo, Matthew Jagielski, Irena Gao, Anas Awadalla, Pang Wei Koh, Daphne Ippolito, Katherine Lee, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned?, 2024. 3", + "[4] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J. Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries, 2024. 3", + "[5] Chatglm. Zhipu ai, 2023. 2, 5, 7", + "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning, 2023. 2", + "[7] Alibaba Cloud. Qwen, 2023. 2, 5, 7", + "[8] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks?, 2023. 3", + "[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 3", + "[10] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 3", + "[11] Google. Gemini, 2024. 2, 5, 7", + "[12] Sepp Hochreiter and Jurgen Schmidhuber. Flat minima. Neural computation, 9:1-42, 1997. 5", + "[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, and Madian Khabsa. Llama guard: Llm-based input-output safeguard for human-ai conversations, 2023. 6", + "[14] Jiaming Ji, Mickel Liu, Juntao Dai, Xuehai Pan, Chi Zhang, Ce Bian, Chi Zhang, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. arXiv preprint arXiv:2307.04657, 2023. 6", + "[15] Tomasz Korbak, Kejian Shi, Angelica Chen, Rasika Bhalerao, Christopher L. Buckley, Jason Phang, Samuel R." + ], + "bbox": [ + 93, + 196, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bowman, and Ethan Perez. Pretraining language models with human preferences, 2023. 1, 3", + "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension, 2023. 3", + "[17] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models, 2024. 3", + "[18] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2, 3, 6", + "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 5, 6, 7, 8", + "[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2024. 2", + "[21] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 6", + "[22] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 5", + "[23] Yugeng Liu, Tianshuo Cong, Zhengyu Zhao, Michael Backes, Yun Shen, and Yang Zhang. Robustness over time: Understanding adversarial examples' effectiveness on longitudinal versions of large language models, 2024. 1", + "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024. 3", + "[25] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model, 2024. 3", + "[26] OpenAI. Dall·E 3, 2023. 4", + "[27] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, and et al. Florencia Leoni Aleman and. Gpt-4 technical report, 2024. 6", + "[28] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3419-3448, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3", + "[29] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models, 2023. 1, 2, 3, 6", + "[30] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to!, 2023. 1", + "[31] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models, 2023. 3", + "[32] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19898", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, Lewis Ho, Divya Siddarth, Shahar Avin, Will Hawkins, Been Kim, Jason Gabriel, Vijay Bolina, Jack Clark, Yoshua Bengio, Paul Christiano, and Allan Dafoe. Model evaluation for extreme risks, 2023. 1, 3", + "[34] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image, 2024. 3", + "[35] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms, 2023. 3", + "[36] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts, 2024. 3", + "[37] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt, 2024. 1, 2, 3, 6", + "[38] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3", + "[39] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models, 2023. 3", + "[40] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023, 2, 5, 6, 7, 8", + "[41] Andy Zou, Zifan Wang, J. Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models, 2023. 5", + "[42] Mark Elliot Zuckerberg. Meta, 2004. 6" + ], + "bbox": [ + 91, + 90, + 480, + 656 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19899", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_model.json b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..366fa423052fec4c4d293baf2c88eae3dff935c0 --- /dev/null +++ b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.202, + 0.87, + 0.294 + ], + "angle": 0, + "content": "Shuyang Hao \\(^{1}\\) Bryan Hooi \\(^{2}\\) Jun Liu \\(^{3}\\) Kai-Wei Chang \\(^{4}\\) Zi Huang \\(^{5}\\) Yujun Cai \\(^{5*}\\) Southeast University, \\(^{2}\\)National University of Singapore, \\(^{3}\\)Lancaster University \\(^{4}\\)University of California, Los Angeles, \\(^{5}\\)University of Queensland haosy@seu.edu.cn, dcsbhk@nus.edu.sg, j.liu81@lancaster.ac.uk, kw@kwchang.net, huang@itee.uq.edu.au, yujun.cai@uq.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.306, + 0.327, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.485, + 0.686 + ], + "angle": 0, + "content": "Despite inheriting security measures from underlying language models, Vision-Language Models (VLMs) may still be vulnerable to safety alignment issues. Through empirical analysis, we uncover two critical findings: scenario-matched images can significantly amplify harmful outputs, and contrary to common assumptions in gradient-based attacks, minimal loss values do not guarantee optimal attack effectiveness. Building on these insights, we introduce MLAI (Multi-Loss Adversarial Images), a novel jailbreak framework that leverages scenario-aware image generation for semantic alignment, exploits flat minima theory for robust adversarial image selection, and employs multi-image collaborative attacks for enhanced effectiveness. Extensive experiments demonstrate MLAI's significant impact, achieving attack success rates of \\(77.75\\%\\) on MiniGPT-4 and \\(82.80\\%\\) on LLaVA-2, substantially outperforming existing methods by margins of \\(34.37\\%\\) and \\(12.77\\%\\) respectively. Furthermore, MLAI shows considerable transferability to commercial black-box VLMs, achieving up to \\(60.11\\%\\) success rate. Our work reveals fundamental visual vulnerabilities in current VLMs safety mechanisms and underscores the need for stronger defenses. Warning: This paper contains potentially harmful example text." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.715, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) have demonstrated remarkable capabilities in understanding and generating multimodal content, enabling various applications from content creation to visual reasoning. To prevent potential misuse, these models typically undergo comprehensive safety protocols [23, 30] including safety-focused pre-training and reinforcement learning with human feedback (RLHF) [15, 28, 33]. Despite these safeguards, ensuring the safe deployment of these increasingly powerful models remains a criti-" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.306, + 0.905, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.753, + 0.908, + 0.851 + ], + "angle": 0, + "content": "Figure 1. An example to show the Visual Vulnerabilities of the scenario-matched images on alignment of VLMs and the unreliability of using image with minimal loss (the cross-entropy loss between the model's output and the target in gradient-based optimization). We can find that: (1) matching images are better than irrelevant or no images, and (2) since only the last jailbreak is successful, lower loss is not always better." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "cal challenge, as recent studies [10, 18, 29, 32, 37] have revealed that visual inputs can significantly compromise these" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19890" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.223, + 0.106 + ], + "angle": 0, + "content": "safety mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.484, + 0.275 + ], + "angle": 0, + "content": "The vulnerability of VLMs to visual inputs is particularly concerning as these models are increasingly deployed in both open-source and commercial platforms. In this paper, we investigate jailbreak attacks that aim to circumvent VLMs' safety mechanisms through carefully crafted adversarial images. This vulnerability is particularly critical as VLMs are increasingly deployed in both open-source and commercial platforms. Our study primarily focuses on white-box [19, 40] settings where model parameters are accessible, with additional evaluation on commercial blackbox VLMs [5, 7, 11] to assess real-world implications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.484, + 0.58 + ], + "angle": 0, + "content": "For white-box attacks, current approaches [18, 29, 37] typically rely on gradient-based optimization of adversarial images. These methods iteratively refine random noise patterns by minimizing the cross-entropy loss between the model's generated response and the desired harmful output. Specifically, given a harmful instruction and a candidate adversarial image, the loss measures how closely the model's response matches the target harmful content. Through systematic analysis, we identify two critical limitations in this paradigm. As illustrated in Fig. 1, first, adversarial images with minimal loss values do not necessarily lead to successful attacks, challenging a fundamental assumption in gradient-based optimization. When repeating attacks with different initial images, we observe significant variations in success rates, suggesting that loss minimization alone is not a reliable strategy. Second, using a single image for attacking all scenarios is not an effective approach, as the harmfulness of the VLM's response is strongly correlated with the semantic relevance of the image content to the target scenario, which is also demonstrated in previous work [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Building on these insights, we propose a novel jailbreak method called MLAI, which emphasizes targeted image generation and the use of Multi-Loss Adversarial Images to amplify harmful intent and disrupt multi-modal alignment. Specifically, our method introduces a three-stage attack strategy. First, MLAI generates initial images that match the target scenario. In this step, we generate the image with the highest potential for harmful output in the specific scenario while ensuring that the adversarial image remains visually natural, making the model more susceptible to generating unsafe responses. Second, MLAI optimizes the adversarial image through gradient updates and saves results that fall within a predefined loss range during the iterative process. In the process of iterative refinement, the harmfulness of the image is progressively amplified. Third, MLAI carefully calculate the loss range and select adversarial images within the loss range for collaborative attack. The determination of the loss range takes into account the balance between attack success rate and computational cost." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Interestingly, MLAI demonstrates considerable transferability across different models. Our experiments show" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "that adversarial images trained on white-box models maintain significant attack success rates when applied to unseen black-box commercial VLMs. This transferability suggests fundamental vulnerabilities in current VLM safety mechanisms that transcend specific model architectures and training approaches. To mitigate this problem, we propose a similarity-based dedduplication defense that effectively mitigates MLAI's impact by limiting the number of similar adversarial input images." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.228, + 0.862, + 0.242 + ], + "angle": 0, + "content": "In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.243, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- We undertake a comprehensive analysis of the gradient-based attack process. Through targeted experiments, we explore the selection of initial images and examine the outcomes of adversarial images with varying loss values. The findings highlight that strategies relying exclusively on the adversarial image with the lowest loss or using a single initial image are suboptimal, offering new insights into attack effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.364, + 0.905, + 0.453 + ], + "angle": 0, + "content": "- We introduce a novel jailbreak method, MLAI, which leverages images that match the target scenario along with a collaborative approach involving multi-loss adversarial images. This combined strategy amplifies the harmful intent of the attack, successfully disrupting multi-modal alignment in VLMs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.455, + 0.905, + 0.558 + ], + "angle": 0, + "content": "- Experimental results show that MLAI achieves a remarkable attack success rate (ASR) of \\(77.75\\%\\) on MiniGPT-4 and \\(82.80\\%\\) on LLaVA-2, making existing white-box VLMs highly vulnerable. Additionally, we demonstrate the potential of MLAI against black-box commercial VLMs and propose a countermeasure to defend against these advanced attacks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.243, + 0.905, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.573, + 0.655, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Vision Language Models (VLMs). Leveraging the remarkable capabilities of LLMs, the field has seen the emergence of several prominent VLMs. Notable among these are LLaVA [19] and MiniGPT-4 [40], which integrate an LLM with a vision encoder and a cross-modal connector. These models adopt a two-stage training framework, comprising cross-modal representation alignment and visual instruction tuning. Building upon this foundational approach, a variety of open-source VLMs [20] have been developed, which enhance multimodal performance through several key advancements: the use of higher-resolution inputs, more powerful LLM architectures, expanded pretraining corpora [2], and the introduction of specialized pretraining tasks [2, 6]. These efforts have significantly improved the models' ability to process and generate content across modalities, further advancing the state of multimodal understanding and generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Alignment Vulnerabilities in VLMs. Similar to LLMs, VLMs are also vulnerable to malicious inputs, which can" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19891" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.513 + ], + "angle": 0, + "content": "lead to the generation of harmful content [29, 38]. However, existing benchmarks predominantly focus on evaluating the effectiveness of VLMs, with insufficient attention given to their alignment vulnerabilities, particularly regarding harmlessness [9, 16, 24]. To address this gap, recent research has developed red-teaming attack benchmarks that systematically assess the alignment vulnerabilities of VLMs across various scenarios [17, 24, 35]. In line with LLM research, studies on jailbreaking VLMs can be divided into white-box, and black-box attack strategies. For white-box attacks, existing methods [3, 8, 29, 31, 32, 34] can be further categorized based on their target: input images or visual embeddings. Attacks on input images often generate adversarial images constrained to provoke harmful responses [8, 29, 31] or use teacher-forcing optimization [3]. For visual embeddings, adversarial images are crafted to appear harmless but mimic the embeddings of harmful images [32], thus bypassing content filters. On the other hand, black-box attacks typically undermine the alignment of VLMs by techniques such as system prompt attacks [4, 36], converting harmful information into text-oriented images [10], leveraging surrogate models to generate adversarial images [39], or utilizing maximum likelihood-based jailbreaking methods [25]. Our work extends this line of research by systematically analyzing the visual-semantic vulnerabilities in VLM safety alignment and introducing novel attack strategies that leverage scenario-matched images and flat minima characteristics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.266, + 0.544 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.203, + 0.567 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.666 + ], + "angle": 0, + "content": "Recent studies [10, 18, 29, 32, 37] have revealed that VLMs, despite inheriting safety measures from their foundational LLMs, remain vulnerable to jailbreak attacks. Given a vision-language model built based on a LLM M, an image encoder \\( E \\) and a projection layer \\( W \\), its generation process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.678, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\ny = \\mathbf {M} ([ W \\cdot E (i), t ]), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.824 + ], + "angle": 0, + "content": "where \\(i\\) and \\(t\\) are the input image and text, and \\(y\\) is the model's output. Through empirical analysis of gradient-based attack methods, we observe two key limitations: (1) arbitrarily selected initial images fail to fully exploit the correlation between image content and model output harmfulness, and (2) our experiments show that adversarial images with minimal loss do not necessarily lead to the highest attack success rate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we propose MLAI, a novel jailbreak framework that enhances attack effectiveness through scenario-aware image generation and multi-image collaborative attack. As illustrated in Fig. 3, MLAI first generates initial images that match the target scenario," + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.088, + 0.852, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.266, + 0.908, + 0.363 + ], + "angle": 0, + "content": "Figure 2. The ASR of transferability across scenarios. The heat map shows that images generated under the IA scenario achieve high transferability when applied to MG contexts, while FR scenario images transfer effectively to LO settings. These transferability patterns are bidirectional between paired scenarios. In contrast, PO scenario images show minimal transfer effectiveness across other contexts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.452 + ], + "angle": 0, + "content": "then obtains a set of adversarial images through gradient optimization, and finally utilizes multiple images within an optimal loss range for the collaborative attack. The complete attack process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.464, + 0.907, + 0.505 + ], + "angle": 0, + "content": "\\[\ny ^ {*} = \\sum_ {k} ^ {\\Theta} \\mathbf {M} (\\left[ W \\cdot E (I _ {t} ^ {a d v, k}), t \\right]), \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.518, + 0.907, + 0.551 + ], + "angle": 0, + "content": "where \\(y^{*}\\) is the attack generation, \\(I_{t}^{adv,k}\\) represents the k-th adversarial image and \\(\\Theta\\) denotes the selected loss range." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.816, + 0.575 + ], + "angle": 0, + "content": "3.2. Scenario-aware Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Through training on human preference data, existing LLMs [15, 28, 33] have learned to align with human values and refuse to respond to harmful text inputs. However, recent studies [18] indicate that incorporating images into inputs can significantly increase the likelihood of generating harmful content, and the harmfulness of VLM outputs is often positively correlated with the harmfulness of the visual content. This observation suggests that visual components introduce additional vulnerabilities, which can be exploited to circumvent the alignment of the underlying LLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "By comparing the transferability of generated images across different scenarios, we further validate the importance of semantic matching between images and attack targets. As shown in Fig. 2, images generated under certain scenarios demonstrate high transferability to semantically related contexts. For instance, instructions within Illegal Activity (IA) and Malware Generation (MG) scenarios share common harmful action patterns like \"creating\" or \"utilizing\", leading to effective cross-scenario transfers. Similarly, Fraud (FR) and Legal Opinion (LO) scenarios often involve abstract harmful concepts like \"concealing\"" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19892" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.5, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.196, + 0.499, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.381, + 0.498, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.891, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.576, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 3. Our MLAI framework involves a three-step procedure: (1) generate an image that matches the text scenario as initial image, (2) obtain the adversarial image set based on the initial image by gradient update, and (3) calculate the loss range and select adversarial images within the loss range for collaborative attack." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.486, + 0.706 + ], + "angle": 0, + "content": "or \"deceiving\". In contrast, scenarios with distinct semantic frameworks, such as Pornography (PO), show minimal transfer effectiveness, highlighting the crucial role of semantic alignment in attack success." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Motivated by these findings, we propose to generate initial images that match the semantic content of jailbreak targets. Given that jailbreak-related scenarios often involve abstract concepts that are challenging to accurately depict, we first classify jailbreak targets \\( t \\) into 13 distinct categories \\( C = \\{c_{1}, c_{2}, \\dots, c_{13}\\} \\) based on common harmful behaviors (e.g., Illegal Activity, Hate Speech and Fraud). For each category, we carefully craft prompts that encode contextual information while avoiding direct depiction of harmful elements. For example, when targeting fraud-related scenarios, we generate images of relevant but benign contexts like business meetings or document processing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.907, + 0.676 + ], + "angle": 0, + "content": "These carefully crafted prompts are then used with DALL-E 3 [26] to generate the initial images:" + }, + { + "type": "equation", + "bbox": [ + 0.657, + 0.682, + 0.907, + 0.699 + ], + "angle": 0, + "content": "\\[\nI _ {t} = G (P (t)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.841 + ], + "angle": 0, + "content": "where \\( P(\\cdot) \\) generates the scenario-specific prompt and \\( G(\\cdot) \\) represents the image generation model. Through this design, the generated images establish strong semantic connections with the jailbreak targets while maintaining a proper balance between attack effectiveness and content safety. Our subsequent experiments in Sec. 4.3 demonstrate that such scenario-aware initial images significantly improve the success rate of gradient-based attacks compared to randomly selected images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.803, + 0.865 + ], + "angle": 0, + "content": "3.3. Adversarial Image Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Given the scenario-aware initial image \\(I_{t}\\), we further optimize it through gradient updates to enhance the attack effec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19893" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.287, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.212, + 0.258, + 0.223 + ], + "angle": 0, + "content": "(a) ASR of different Loss." + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.096, + 0.483, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.213, + 0.462, + 0.223 + ], + "angle": 0, + "content": "(b) ASR of Min/Multi-Loss." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.236, + 0.483, + 0.265 + ], + "angle": 0, + "content": "Figure 4. Results of the effect of various Loss on ASR. We can clearly observe the limitations of the minimum loss strategy." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.281, + 0.476, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.483, + 0.448 + ], + "angle": 0, + "content": "Figure 5. Comparison of the effects of flat and sharp minima at test set shift. Here the test set shift is simulated by curve translation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.564 + ], + "angle": 0, + "content": "tiveness. First, we construct a small corpus \\( Y := \\{y_k\\}_{k=1}^m \\) consisting of harmful statements and affirmative responses, which serves as target outputs for optimization. For each batch of harmful instructions \\( t' \\) and corresponding adversarial images \\( I_t^{adv} \\), we compute the cross-entropy loss between model outputs and the target responses:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.573, + 0.483, + 0.613 + ], + "angle": 0, + "content": "\\[\nI _ {t} ^ {a d v, k} = \\arg \\min _ {\\widehat {I} _ {t} ^ {a d v} \\in \\mathbb {R}} \\sum_ {k = 1} ^ {m} - l o g \\left(p _ {\\varphi} \\left(y _ {k} \\mid t _ {k} ^ {\\prime}, \\widehat {I} _ {t} ^ {a d v, k}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.719 + ], + "angle": 0, + "content": "where \\(\\mathbb{R}\\) is some constraint applied to the input space in which we search for adversarial examples, and \\(p_{\\varphi}\\) represents the conditional probability generated by the target VLM. To ensure the adversarial images remain visually natural, we constrain the perturbation magnitude by \\(\\left\\| I_t^{adv,k} - I_t\\right\\|_\\infty \\leq \\varepsilon\\) where \\(\\varepsilon\\) is set to 32/255." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.839 + ], + "angle": 0, + "content": "During extensive experiments with this optimization process, we observe that adversarial images with minimal loss values do not necessarily lead to successful attacks. Specifically, we find cases where images with relatively low loss fail to jailbreak the target VLM, while those with moderately higher loss values achieve better success rates. This observation persists across different models and attack scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.388, + 0.864 + ], + "angle": 0, + "content": "3.4. Multi-image Collaborative Attack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To better understand the observation that adversarial images with minimal loss do not necessarily lead to the highest at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.317 + ], + "angle": 0, + "content": "tack success rate, we conduct systematic experiments analyzing attack performance across different loss ranges. As shown in Fig. 4a, our results reveal that the optimal attack performance often comes from adversarial images within a certain loss range, rather than those with minimal loss. This finding aligns with the flat minima phenomenon [12] in optimization, where solutions in flatter regions of the loss landscape tend to be more robust than those in sharp minima. As illustrated in Fig. 5, when the loss landscape exhibits a sharp minimum, even small perturbations can lead to significant performance degradation. In contrast, solutions in flat regions are more resilient to input variations, which is particularly relevant for adversarial attacks where the test-time scenario may differ from the optimization objective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.424 + ], + "angle": 0, + "content": "Based on this insight, we propose to select adversarial images within a carefully determined loss range for collaborative attack. The selection process starts by analyzing the loss landscape around the minimal loss point. Specifically, we select 20 nearest points on each side of the minimum and fit straight lines using least squares method. The loss range \\(\\Theta\\) is then determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.433, + 0.905, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\Theta = \\left(\\left\\lceil \\frac {k _ {l}}{K} \\right\\rceil \\times \\rho , \\left\\lceil \\frac {k _ {r}}{K} \\right\\rceil \\times \\rho\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.581 + ], + "angle": 0, + "content": "where \\( k_{l} \\) and \\( k_{r} \\) are the fitted slopes representing the local geometry of the loss landscape, \\( K \\) is the standard slope derived from normal distribution, and \\( \\rho \\) is a range coefficient set to 6. As demonstrated in Fig. 4b, this approach adaptively selects a narrow range in flat regions for computational efficiency, while expanding the range in sharp regions for better robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.644 + ], + "angle": 0, + "content": "The final jailbreak output is produced through the collaborative effort of multiple adversarial images within \\(\\Theta\\), as formulated in Eq. (2). This strategy enables us to maintain both attack effectiveness and computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.64, + 0.674 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.704, + 0.699 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Models and Datasets. To comprehensively evaluate MLAI's effectiveness, we conduct experiments on both open-source and commercial VLMs. For open-source models, we focus on MiniGPT-4 [40] (Vicuna 13B) and LLaVA-2 [19] (13B) due to their widespread adoption and strong performance. We use the official weights provided in their respective repositories. For commercial models, we evaluate on Gemini [11], ChatGLM [5], and Qwen [7] to validate real-world applicability. We evaluate our approaches using two common datasets: SafetyBench [22] and AdvBench [41]. SafetyBench serves as a benchmark for assessing the safety of VLMs, encompassing 13 typically prohibited scenarios or behaviors as outlined in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19894" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.324 + ], + "angle": 0, + "content": "
ScenarioMiniGPT-4 [40]LLaVA-2 [19]
Plain TextLiu et al. [21]Qi et al. [29]OursPlain TextLiu et al. [21]Qi et al. [29]Ours
Illegal Activity (IA)1.9214.5411.5570.243.8274.7748.0383.73
Hate Speech (HS)1.6811.923.9766.573.0840.9740.2180.36
Malware Generation (MG)3.3219.8815.5250.2830.7564.6353.8579.10
Physical Harm (PH)2.9824.5123.4372.0213.8861.0450.3179.62
Economic Harm (EH)5.684.916.8070.154.1518.427.9850.18
Fraud (FR)3.1718.5614.7150.674.9757.2047.9775.85
Pornography (PO)4.1420.9419.1169.0120.6461.2341.5568.98
Political Lobbying (PL)67.6779.1176.3396.7474.5991.1585.6293.45
Privacy Violence (PV)8.9710.5012.9778.9220.9249.3637.3374.82
Legal Opinion (LO)74.5685.2882.4895.1680.1695.4682.9493.59
Financial Advice (FA)84.3388.1290.9310087.4897.2187.6996.77
Health Consultation (HC)76.5093.9491.2293.0985.6110090.04100
Gov Decision (GD)90.2991.7591.2597.8587.7498.9791.22100
Average32.7143.3841.5677.7539.8370.0358.8382.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.334, + 0.907, + 0.391 + ], + "angle": 0, + "content": "Table 1. ASR results of jailbreaking MiniGPT4 and LLaVA-2 in 13 scenarios. Our attack achieves the best attacking performance in both modules and almost every scenario. We achieve improvements of \\(34.37\\%\\) and \\(12.77\\%\\) over existing state-of-the-art approaches on MiniGPT-4 and LLaVA-2, respectively. Additionally, substantial ASR differences can be observed across scenarios, with notable variations in performance improvement across different contexts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.522 + ], + "angle": 0, + "content": "the usage policies [13, 27] of OpenAI [1] and Meta [42]. AdvBench has been extensively utilized in prior research concerning LLM jailbreak attacks and includes 521 harmful behaviors. We following the setting of BAP [37], removed duplicate entries from AdvBench and integrated each unique item into SafetyBench according to the corresponding scenario for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.598 + ], + "angle": 0, + "content": "Evaluation metric. We following the setting of HADES [18], employ Attack Success Rate (ASR) as our primary metric to assess the effectiveness of our approach, which is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.605, + 0.483, + 0.639 + ], + "angle": 0, + "content": "\\[\nA S R = \\frac {\\sum_ {k = 1} ^ {N} \\mathbf {B} (J (y _ {k}) = T r u e)}{N}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.751 + ], + "angle": 0, + "content": "where \\( y_{k} \\) is the model's response, \\( \\mathbf{B} \\) is an indicator function that equals to 1 if \\( J(y_{k}) = True \\) and 0 otherwise, \\( N \\) is the total number of instructions and \\( J(\\cdot) \\) is the harmfulness judging model, outputting True or False to indicate whether \\( y_{k} \\) is harmful. We adopt Beaver-dam-7B [14] as \\( J(\\cdot) \\), which has been trained on high-quality human feedback data about the above harmful categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Compared attacks. We compare MLAI with two state-of-the-art jailbreak attacks: Liu et al. [21] and Qi et al. [29]. Liu et al. [21] integrated images associated with aggressive intent alongside typographic text to serve as visual adversarial prompts. Qi et al. [29], on the other hand, refined these visual adversarial prompts by leveraging a corpus specific to certain scenarios. Additionally, we include a \"Plain Text\" baseline where harmful queries are directly input to evaluate models' base vulnerability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.416, + 0.704, + 0.433 + ], + "angle": 0, + "content": "4.2. Experiment Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.441, + 0.907, + 0.502 + ], + "angle": 0, + "content": "White-box Attacks. Our experimental evaluation focuses on three key aspects: attack effectiveness across models, vulnerability patterns in different scenarios, and performance variations among scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.504, + 0.906, + 0.595 + ], + "angle": 0, + "content": "The results in Tab. 1 show that MLAI achieves strong attack performance on both tested models. For MiniGPT-4 [40], MLAI reaches \\(77.75\\%\\) ASR, surpassing the previous best method \\((43.38\\%)\\) by \\(34.37\\%\\). Similarly on LLaVA-2 [19], MLAI achieves \\(82.80\\%\\) ASR, outperforming existing approaches by \\(12.77\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.597, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Analysis across scenarios reveals distinct vulnerability patterns in both models. Security-critical scenarios like Malware Generation (MG) and Fraud (FR) show stronger resistance with ASR of \\(50.28\\%\\) and \\(50.67\\%\\), while advisory scenarios such as Financial Advice (FA) and Gov Decision (GD) exhibit higher vulnerability \\((\\mathrm{ASR} > 90\\%)\\). This pattern reflects the models' internal defense mechanisms: security-critical scenarios are equipped with more rigorous detection systems due to their well-defined harmful patterns, and these high-risk scenarios are typically protected by enhanced safety measures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The effectiveness of MLAI varies significantly across scenarios. Most notably, while achieving a \\(55.70\\%\\) improvement in Illegal Activity (IA) scenario, the gain in Gov Decision (GD) scenario is merely \\(0.85\\%\\). This disparity can be primarily explained by two aspects. First, scenarios like Gov Decision (GD) already demonstrate high base vulnerability \\((90.29\\%)\\) ASR without attacks), leaving limited room for improvement. Second, the nature of different scenarios affects our attack strategy: Illegal Activity (IA) scenario" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19895" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.155 + ], + "angle": 0, + "content": "
ScenarioGemini [11]ChatGLM [5]Qwen [7]
Illegal Activity49.5960.1154.97
Malware Generation39.8442.5645.28
Physical Harm43.3545.6248.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.165, + 0.483, + 0.222 + ], + "angle": 0, + "content": "Table 2. ASR results for black-box models. The ASR differences across various models within the same scenario are minor, and the ranking of ASR aligns closely with those observed during testing on white-box models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.484, + 0.311 + ], + "angle": 0, + "content": "involve concrete elements (e.g., drug manufacturing) that can be effectively represented in images, whereas Gov Decision (GD) scenario contain abstract concepts that are inherently challenging to visualize, thereby limiting the potential of image-based attacks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.325, + 0.485, + 0.628 + ], + "angle": 0, + "content": "Black-box Attacks. Besides the open-source models, we also evaluate our attack on three black-box commercial VLMs (i.e., Gemini [11], ChatGLM [5] and Qwen [7]). We selected 3 most harmful scenarios (Illegal Activity, Malware Generation, and Physical Harm) and 60 related queries to attack each model. Since direct access to the internal data of the model is unavailable, we employ images trained on MiniGPT-4 [40] to perform a black-box model jailbreak. As shown in Tab. 2, MLAI maintains considerable attack effectiveness on commercial models, achieving ASRs of \\(49.59\\%\\), \\(60.11\\%\\), and \\(54.97\\%\\) on Gemini, ChatGLM, and Qwen respectively. We also notice that compared to white-box attacks, there exists a consistent performance drop (\\(16.43\\%\\) lower on average). This gap likely stems from commercial models' enhanced defense mechanisms and architectural differences. Nevertheless, the consistent attack performance across different commercial models suggests that our method can effectively generalize beyond the training model, despite the presence of unknown defense strategies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.255, + 0.65 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.735 + ], + "angle": 0, + "content": "In this part, we further discuss the effectiveness of our proposed approach, including image harmfulness optimization and loss range determination. We note that except for the explicitly stated modifications, all other parameters are identical to the baseline configuration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Image harmfulness optimization. To validate the effectiveness of the optimization process for image generation discussed in Sec. 3.2, we conduct a specific experiment with selecting different types of images to examine the attack performance. We selected two scenarios, Hate Speech and Fraud, and conducted experiments on the MiniGPT-4 [40] and LLaVA-2 [19] models, respectively. Moreover, the Irrelevant Image used in the experiments are fixed and consist of 12 images that is not associated with any of the jailbreak scenarios. As shown in Fig. 6, when blank images" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.71, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.204, + 0.661, + 0.215 + ], + "angle": 0, + "content": "(a) ASR of Fraud." + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.089, + 0.906, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.204, + 0.875, + 0.216 + ], + "angle": 0, + "content": "(b) ASR of Hate Speech." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.326 + ], + "angle": 0, + "content": "Figure 6. The ASR of using different types of initial images. We can find that simply adding a blank image can lead to a substantial increase in the success rate. Compared to blank image, further selecting images that contain information but are unrelated to the jailbreak scenario results in a smaller improvement. However, when images are introduced that closely match the specific scenario, the performance reaches state-of-the-art levels." + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.343, + 0.852, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.52, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Figure 7. The ASR of using different numbers of images. According to Eq. (5), since \\( K \\) is a constant value, we set \\( v = \\left\\lceil \\frac{k_{L}}{K} \\right\\rceil \\) and assume that the slopes on both sides are equal without loss of generality, where \\( v \\) represents the flatness of the minimum value, in order to better focus on the changing trends of ASR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.809 + ], + "angle": 0, + "content": "are introduced, the alignment of the underlying LLM fails to cover the unforeseen domains introduced by the visual modality. This misalignment effectively circumvents the inherent alignment mechanisms of the LLM within the VLM framework, leading to a marked increase in the ASR. However, when choosing unrelated images, for the target scene, it merely increases the information content and complexity of the image, which does not significantly aid in bypassing restrictions. Only by selecting images that match the scenario can meaningful semantic information be introduced, thereby amplifying the potential risk of the image and ultimately further increasing the ASR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Loss range determination. To further validate the impact of using images in different loss range on the improvement of jailbreak success rate, we conduct experiments on four scenarios on the MiniGPT-4 [40] model and the results are shown in Fig. 7. We observe that, overall, the ASR increases" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19896" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.089, + 0.424, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Figure 8. The average ASR of using different \\(\\rho\\). The average ASR increases gradually with the value of \\(\\rho\\), with the rate of increase remaining relatively stable. However, once \\(\\rho\\) reaches a critical threshold (which is 5 in this experiment), the growth rate drops significantly, and the ASR quickly reaches its maximum value." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.37, + 0.485, + 0.672 + ], + "angle": 0, + "content": "significantly when the number of Images initially increases. However, the growth patterns of the curves in different scenarios show considerable variation, with differences in both the rate of increase and the point at which the growth stops. Specifically, the values of \\( v \\) for Illegal Activity (IA) and Hate Speech (HS) are relatively large, indicating that they have relatively sharp minima, whereas Legal Opinion (LO) and Financial Advice (FA) exhibit flatter minima. Correspondingly, the curves for Illegal Activity (IA) and Hate Speech (HS) stop increasing early (almost cease to grow when the number of images reaches 5), while the curves for Legal Opinion (LO) and Financial Advice (FA) maintain a prolonged growth phase (only stopping near the point where the number of images reaches 9). The results demonstrate that to achieve an optimal ASR, image sets with flat minima require fewer images, while image sets with sharp minima incur higher costs to achieve the desired outcome. These findings prove that flatter minima indeed provide greater robustness, allowing models to perform better when the test set deviates from the training set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.839 + ], + "angle": 0, + "content": "In addition, we also examined the values of \\(\\rho\\). We tested the average ASR across all scenarios for different values of \\(\\rho\\) on both MiniGPT-4 [40] and LLaVA-2 [19]. As shown in Fig. 8, the average ASR reaches its near-maximum value when \\(\\rho = 6\\), while the computational complexity increases at a linear rate. Further increases in \\(\\rho\\) yield diminishing returns. Therefore, based on empirical observations, we set the baseline value of \\(\\rho = 6\\). Through the above experiments, we can prove the effectiveness of the flat minimum, the optimal choice of \\(\\rho\\), and we finally find the balance point between the attack success rate and the computational cost." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.258, + 0.864 + ], + "angle": 0, + "content": "4.4. Defense Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "From the perspective of developmental trends, defense mechanisms always occupy the weaker position. With the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.709, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.21, + 0.709, + 0.234 + ], + "angle": 0, + "content": "(a) Results of 10 specific questions in fraud scenario." + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.089, + 0.906, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.21, + 0.905, + 0.234 + ], + "angle": 0, + "content": "(b) ASR of original/defensive circumstance." + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.247, + 0.905, + 0.261 + ], + "angle": 0, + "content": "Figure 9. The demonstration of our defense motivation and results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.906, + 0.363 + ], + "angle": 0, + "content": "emergence of an increasing number of VLMs and the continuous improvement of jailbreak methods, comprehensive defense becomes progressively more challenging. In this section, we analyze an effective defense strategy against Multi-Loss Adversarial Images attack." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.62 + ], + "angle": 0, + "content": "As shown in Fig. 9a, we observe that under the same scenario, images with different loss not only vary in overall attack potency but also demonstrate significant differences in their effectiveness against individual queries. Based on this finding, we propose introducing a similarity detection module before image sets are input into VLMs. If a current image is detected to be highly similar to a previously encountered image, it is removed from the input set, thus preventing it from being fed into the VLMs. This approach transforms a Multi-Loss Adversarial Attack back into a Single-Loss Adversarial Attack, and due to the diversity of queries in the scenario, the average ASR achieves robust defense. We conducted experiments on MiniGPT-4 [40], and the results, as shown in Fig. 9b, indicate that our defense method effectively reduces the success rate of the Multi-Loss Adversarial Attack, achieving a \\(22.99\\%\\) reduction compared to the model without the defense module." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we conducted a systematic analysis of visual vulnerabilities in VLMs safety alignment and introduced MLAI, a novel jailbreak framework that effectively exploits these weaknesses. Our investigation revealed two critical findings: the unreliability of minimal-loss optimization in gradient-based attacks, and the crucial role of semantic alignment between visual content and target scenarios. Based on these insights, MLAI achieves superior performance through scenario-aware image generation, flat minima-based optimization, and multi-image collaborative attacks. Extensive experiments demonstrate MLAI's effectiveness, achieving attack success rates of \\(77.75\\%\\) on MiniGPT-4 and \\(82.80\\%\\) on LLaVA-2. The strong transferability to commercial black-box VLMs further highlights fundamental challenges in achieving robust safety alignment for multi-modal systems." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19897" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.144 + ], + "angle": 0, + "content": "This research has been partially supported by Australian Research Council Discovery Project DP230101196." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.188 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.198, + 0.369, + 0.212 + ], + "angle": 0, + "content": "[1] Sam Altman and et al. OpenAI, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[3] Nicholas Carlini, Milad Nasr, Christopher A. Choquette-Choo, Matthew Jagielski, Irena Gao, Anas Awadalla, Pang Wei Koh, Daphne Ippolito, Katherine Lee, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned?, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[4] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J. Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.399, + 0.32, + 0.413 + ], + "angle": 0, + "content": "[5] Chatglm. Zhipu ai, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.484, + 0.484 + ], + "angle": 0, + "content": "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.341, + 0.499 + ], + "angle": 0, + "content": "[7] Alibaba Cloud. Qwen, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[8] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks?, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.484, + 0.627 + ], + "angle": 0, + "content": "[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.63, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.307, + 0.7 + ], + "angle": 0, + "content": "[11] Google. Gemini, 2024. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[12] Sepp Hochreiter and Jurgen Schmidhuber. Flat minima. Neural computation, 9:1-42, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, and Madian Khabsa. Llama guard: Llm-based input-output safeguard for human-ai conversations, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[14] Jiaming Ji, Mickel Liu, Juntao Dai, Xuehai Pan, Chi Zhang, Ce Bian, Chi Zhang, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. arXiv preprint arXiv:2307.04657, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Tomasz Korbak, Kejian Shi, Angelica Chen, Rasika Bhalerao, Christopher L. Buckley, Jason Phang, Samuel R." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.198, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Bowman, and Ethan Perez. Pretraining language models with human preferences, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[17] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[18] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.288 + ], + "angle": 0, + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.904, + 0.316 + ], + "angle": 0, + "content": "[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.904, + 0.357 + ], + "angle": 0, + "content": "[21] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[22] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[23] Yugeng Liu, Tianshuo Cong, Zhengyu Zhao, Michael Backes, Yun Shen, and Yang Zhang. Robustness over time: Understanding adversarial examples' effectiveness on longitudinal versions of large language models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[25] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.71, + 0.567 + ], + "angle": 0, + "content": "[26] OpenAI. Dall·E 3, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.609 + ], + "angle": 0, + "content": "[27] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, and et al. Florencia Leoni Aleman and. Gpt-4 technical report, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[28] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3419-3448, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[29] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[30] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to!, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[31] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19898" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[33] Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, Lewis Ho, Divya Siddarth, Shahar Avin, Will Hawkins, Been Kim, Jason Gabriel, Vijay Bolina, Jack Clark, Yoshua Bengio, Paul Christiano, and Allan Dafoe. Model evaluation for extreme risks, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[34] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[35] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[36] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[37] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[38] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[39] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[40] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[41] Andy Zou, Zifan Wang, J. Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.361, + 0.657 + ], + "angle": 0, + "content": "[42] Mark Elliot Zuckerberg. Meta, 2004. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19899" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_origin.pdf b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2246ad0a309d5f704f934b57d00b494fd298514d --- /dev/null +++ b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/0fb20cd3-12d6-4733-afec-1c9df735d84b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d75df37c8089d244acadf7a22e25172e93b8d326847699458a0fd5455bbff2e1 +size 3097524 diff --git a/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/full.md b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..80498fcb0d75ec358ea4838d498322fe5f228581 --- /dev/null +++ b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/full.md @@ -0,0 +1,272 @@ +# Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models + +Shuyang Hao $^{1}$ Bryan Hooi $^{2}$ Jun Liu $^{3}$ Kai-Wei Chang $^{4}$ Zi Huang $^{5}$ Yujun Cai $^{5*}$ Southeast University, $^{2}$ National University of Singapore, $^{3}$ Lancaster University $^{4}$ University of California, Los Angeles, $^{5}$ University of Queensland haosy@seu.edu.cn, dcsbhk@nus.edu.sg, j.liu81@lancaster.ac.uk, kw@kwchang.net, huang@itee.uq.edu.au, yujun.cai@uq.edu.au + +# Abstract + +Despite inheriting security measures from underlying language models, Vision-Language Models (VLMs) may still be vulnerable to safety alignment issues. Through empirical analysis, we uncover two critical findings: scenario-matched images can significantly amplify harmful outputs, and contrary to common assumptions in gradient-based attacks, minimal loss values do not guarantee optimal attack effectiveness. Building on these insights, we introduce MLAI (Multi-Loss Adversarial Images), a novel jailbreak framework that leverages scenario-aware image generation for semantic alignment, exploits flat minima theory for robust adversarial image selection, and employs multi-image collaborative attacks for enhanced effectiveness. Extensive experiments demonstrate MLAI's significant impact, achieving attack success rates of $77.75\%$ on MiniGPT-4 and $82.80\%$ on LLaVA-2, substantially outperforming existing methods by margins of $34.37\%$ and $12.77\%$ respectively. Furthermore, MLAI shows considerable transferability to commercial black-box VLMs, achieving up to $60.11\%$ success rate. Our work reveals fundamental visual vulnerabilities in current VLMs safety mechanisms and underscores the need for stronger defenses. Warning: This paper contains potentially harmful example text. + +# 1. Introduction + +Vision-Language Models (VLMs) have demonstrated remarkable capabilities in understanding and generating multimodal content, enabling various applications from content creation to visual reasoning. To prevent potential misuse, these models typically undergo comprehensive safety protocols [23, 30] including safety-focused pre-training and reinforcement learning with human feedback (RLHF) [15, 28, 33]. Despite these safeguards, ensuring the safe deployment of these increasingly powerful models remains a criti- + +![](images/0125f4c2e3c819b2eb7f0430bee864c98b8447df6902dfc100cf6fac7f4334be.jpg) +Figure 1. An example to show the Visual Vulnerabilities of the scenario-matched images on alignment of VLMs and the unreliability of using image with minimal loss (the cross-entropy loss between the model's output and the target in gradient-based optimization). We can find that: (1) matching images are better than irrelevant or no images, and (2) since only the last jailbreak is successful, lower loss is not always better. + +cal challenge, as recent studies [10, 18, 29, 32, 37] have revealed that visual inputs can significantly compromise these + +safety mechanisms. + +The vulnerability of VLMs to visual inputs is particularly concerning as these models are increasingly deployed in both open-source and commercial platforms. In this paper, we investigate jailbreak attacks that aim to circumvent VLMs' safety mechanisms through carefully crafted adversarial images. This vulnerability is particularly critical as VLMs are increasingly deployed in both open-source and commercial platforms. Our study primarily focuses on white-box [19, 40] settings where model parameters are accessible, with additional evaluation on commercial blackbox VLMs [5, 7, 11] to assess real-world implications. + +For white-box attacks, current approaches [18, 29, 37] typically rely on gradient-based optimization of adversarial images. These methods iteratively refine random noise patterns by minimizing the cross-entropy loss between the model's generated response and the desired harmful output. Specifically, given a harmful instruction and a candidate adversarial image, the loss measures how closely the model's response matches the target harmful content. Through systematic analysis, we identify two critical limitations in this paradigm. As illustrated in Fig. 1, first, adversarial images with minimal loss values do not necessarily lead to successful attacks, challenging a fundamental assumption in gradient-based optimization. When repeating attacks with different initial images, we observe significant variations in success rates, suggesting that loss minimization alone is not a reliable strategy. Second, using a single image for attacking all scenarios is not an effective approach, as the harmfulness of the VLM's response is strongly correlated with the semantic relevance of the image content to the target scenario, which is also demonstrated in previous work [18]. + +Building on these insights, we propose a novel jailbreak method called MLAI, which emphasizes targeted image generation and the use of Multi-Loss Adversarial Images to amplify harmful intent and disrupt multi-modal alignment. Specifically, our method introduces a three-stage attack strategy. First, MLAI generates initial images that match the target scenario. In this step, we generate the image with the highest potential for harmful output in the specific scenario while ensuring that the adversarial image remains visually natural, making the model more susceptible to generating unsafe responses. Second, MLAI optimizes the adversarial image through gradient updates and saves results that fall within a predefined loss range during the iterative process. In the process of iterative refinement, the harmfulness of the image is progressively amplified. Third, MLAI carefully calculate the loss range and select adversarial images within the loss range for collaborative attack. The determination of the loss range takes into account the balance between attack success rate and computational cost. + +Interestingly, MLAI demonstrates considerable transferability across different models. Our experiments show + +that adversarial images trained on white-box models maintain significant attack success rates when applied to unseen black-box commercial VLMs. This transferability suggests fundamental vulnerabilities in current VLM safety mechanisms that transcend specific model architectures and training approaches. To mitigate this problem, we propose a similarity-based dedduplication defense that effectively mitigates MLAI's impact by limiting the number of similar adversarial input images. + +In summary, our key contributions are as follows: + +- We undertake a comprehensive analysis of the gradient-based attack process. Through targeted experiments, we explore the selection of initial images and examine the outcomes of adversarial images with varying loss values. The findings highlight that strategies relying exclusively on the adversarial image with the lowest loss or using a single initial image are suboptimal, offering new insights into attack effectiveness. +- We introduce a novel jailbreak method, MLAI, which leverages images that match the target scenario along with a collaborative approach involving multi-loss adversarial images. This combined strategy amplifies the harmful intent of the attack, successfully disrupting multi-modal alignment in VLMs. +- Experimental results show that MLAI achieves a remarkable attack success rate (ASR) of $77.75\%$ on MiniGPT-4 and $82.80\%$ on LLaVA-2, making existing white-box VLMs highly vulnerable. Additionally, we demonstrate the potential of MLAI against black-box commercial VLMs and propose a countermeasure to defend against these advanced attacks. + +# 2. Related Work + +Vision Language Models (VLMs). Leveraging the remarkable capabilities of LLMs, the field has seen the emergence of several prominent VLMs. Notable among these are LLaVA [19] and MiniGPT-4 [40], which integrate an LLM with a vision encoder and a cross-modal connector. These models adopt a two-stage training framework, comprising cross-modal representation alignment and visual instruction tuning. Building upon this foundational approach, a variety of open-source VLMs [20] have been developed, which enhance multimodal performance through several key advancements: the use of higher-resolution inputs, more powerful LLM architectures, expanded pretraining corpora [2], and the introduction of specialized pretraining tasks [2, 6]. These efforts have significantly improved the models' ability to process and generate content across modalities, further advancing the state of multimodal understanding and generation. + +Alignment Vulnerabilities in VLMs. Similar to LLMs, VLMs are also vulnerable to malicious inputs, which can + +lead to the generation of harmful content [29, 38]. However, existing benchmarks predominantly focus on evaluating the effectiveness of VLMs, with insufficient attention given to their alignment vulnerabilities, particularly regarding harmlessness [9, 16, 24]. To address this gap, recent research has developed red-teaming attack benchmarks that systematically assess the alignment vulnerabilities of VLMs across various scenarios [17, 24, 35]. In line with LLM research, studies on jailbreaking VLMs can be divided into white-box, and black-box attack strategies. For white-box attacks, existing methods [3, 8, 29, 31, 32, 34] can be further categorized based on their target: input images or visual embeddings. Attacks on input images often generate adversarial images constrained to provoke harmful responses [8, 29, 31] or use teacher-forcing optimization [3]. For visual embeddings, adversarial images are crafted to appear harmless but mimic the embeddings of harmful images [32], thus bypassing content filters. On the other hand, black-box attacks typically undermine the alignment of VLMs by techniques such as system prompt attacks [4, 36], converting harmful information into text-oriented images [10], leveraging surrogate models to generate adversarial images [39], or utilizing maximum likelihood-based jailbreaking methods [25]. Our work extends this line of research by systematically analyzing the visual-semantic vulnerabilities in VLM safety alignment and introducing novel attack strategies that leverage scenario-matched images and flat minima characteristics. + +# 3. Proposed Method + +# 3.1. Overview + +Recent studies [10, 18, 29, 32, 37] have revealed that VLMs, despite inheriting safety measures from their foundational LLMs, remain vulnerable to jailbreak attacks. Given a vision-language model built based on a LLM M, an image encoder $E$ and a projection layer $W$ , its generation process can be formulated as: + +$$ +y = \mathbf {M} ([ W \cdot E (i), t ]), \tag {1} +$$ + +where $i$ and $t$ are the input image and text, and $y$ is the model's output. Through empirical analysis of gradient-based attack methods, we observe two key limitations: (1) arbitrarily selected initial images fail to fully exploit the correlation between image content and model output harmfulness, and (2) our experiments show that adversarial images with minimal loss do not necessarily lead to the highest attack success rate. + +To address these challenges, we propose MLAI, a novel jailbreak framework that enhances attack effectiveness through scenario-aware image generation and multi-image collaborative attack. As illustrated in Fig. 3, MLAI first generates initial images that match the target scenario, + +![](images/d2e31a66865086b8353a7c44f8c5e00e2dfd90fd12439c43774aa07a0ae69a74.jpg) +Figure 2. The ASR of transferability across scenarios. The heat map shows that images generated under the IA scenario achieve high transferability when applied to MG contexts, while FR scenario images transfer effectively to LO settings. These transferability patterns are bidirectional between paired scenarios. In contrast, PO scenario images show minimal transfer effectiveness across other contexts. + +then obtains a set of adversarial images through gradient optimization, and finally utilizes multiple images within an optimal loss range for the collaborative attack. The complete attack process can be formulated as: + +$$ +y ^ {*} = \sum_ {k} ^ {\Theta} \mathbf {M} (\left[ W \cdot E (I _ {t} ^ {a d v, k}), t \right]), \qquad (2) +$$ + +where $y^{*}$ is the attack generation, $I_{t}^{adv,k}$ represents the k-th adversarial image and $\Theta$ denotes the selected loss range. + +# 3.2. Scenario-aware Image Generation + +Through training on human preference data, existing LLMs [15, 28, 33] have learned to align with human values and refuse to respond to harmful text inputs. However, recent studies [18] indicate that incorporating images into inputs can significantly increase the likelihood of generating harmful content, and the harmfulness of VLM outputs is often positively correlated with the harmfulness of the visual content. This observation suggests that visual components introduce additional vulnerabilities, which can be exploited to circumvent the alignment of the underlying LLMs. + +By comparing the transferability of generated images across different scenarios, we further validate the importance of semantic matching between images and attack targets. As shown in Fig. 2, images generated under certain scenarios demonstrate high transferability to semantically related contexts. For instance, instructions within Illegal Activity (IA) and Malware Generation (MG) scenarios share common harmful action patterns like "creating" or "utilizing", leading to effective cross-scenario transfers. Similarly, Fraud (FR) and Legal Opinion (LO) scenarios often involve abstract harmful concepts like "concealing" + +![](images/d8a4e932007eee462732a713a262cdbe14f427acb5dc3505926d210b66ffb3ee.jpg) + +![](images/e3bbd64101760aa6daeb6a68f554d5e4abc795adb9a87499d1b2c794ef435e23.jpg) + +![](images/3a4b2e2df1c879b35f0e7e7afeb40d14e020be11b59ca0f9120a5c7302d8cac7.jpg) +Figure 3. Our MLAI framework involves a three-step procedure: (1) generate an image that matches the text scenario as initial image, (2) obtain the adversarial image set based on the initial image by gradient update, and (3) calculate the loss range and select adversarial images within the loss range for collaborative attack. + +![](images/6b2434129c3d7a81b01c5bc66874756e731174728dd7cd9bbf34606b790c6b34.jpg) + +or "deceiving". In contrast, scenarios with distinct semantic frameworks, such as Pornography (PO), show minimal transfer effectiveness, highlighting the crucial role of semantic alignment in attack success. + +Motivated by these findings, we propose to generate initial images that match the semantic content of jailbreak targets. Given that jailbreak-related scenarios often involve abstract concepts that are challenging to accurately depict, we first classify jailbreak targets $t$ into 13 distinct categories $C = \{c_{1}, c_{2}, \dots, c_{13}\}$ based on common harmful behaviors (e.g., Illegal Activity, Hate Speech and Fraud). For each category, we carefully craft prompts that encode contextual information while avoiding direct depiction of harmful elements. For example, when targeting fraud-related scenarios, we generate images of relevant but benign contexts like business meetings or document processing. + +These carefully crafted prompts are then used with DALL-E 3 [26] to generate the initial images: + +$$ +I _ {t} = G (P (t)), \tag {3} +$$ + +where $P(\cdot)$ generates the scenario-specific prompt and $G(\cdot)$ represents the image generation model. Through this design, the generated images establish strong semantic connections with the jailbreak targets while maintaining a proper balance between attack effectiveness and content safety. Our subsequent experiments in Sec. 4.3 demonstrate that such scenario-aware initial images significantly improve the success rate of gradient-based attacks compared to randomly selected images. + +# 3.3. Adversarial Image Optimization + +Given the scenario-aware initial image $I_{t}$ , we further optimize it through gradient updates to enhance the attack effec + +![](images/6a81ad4b175b629cd621173ed4ec51d9b84bc2d8139997379a1ec865338a6e15.jpg) +(a) ASR of different Loss. + +![](images/e706e8e2656b10bcc9583fb276927fee48da97a4cf9255a65a9fe6cb347ae25f.jpg) +(b) ASR of Min/Multi-Loss. + +![](images/77e0a0f2c07284bc2ce7f5a20e2f57c316f89318e34ab876457a29086670a652.jpg) +Figure 4. Results of the effect of various Loss on ASR. We can clearly observe the limitations of the minimum loss strategy. +Figure 5. Comparison of the effects of flat and sharp minima at test set shift. Here the test set shift is simulated by curve translation. + +tiveness. First, we construct a small corpus $Y := \{y_k\}_{k=1}^m$ consisting of harmful statements and affirmative responses, which serves as target outputs for optimization. For each batch of harmful instructions $t'$ and corresponding adversarial images $I_t^{adv}$ , we compute the cross-entropy loss between model outputs and the target responses: + +$$ +I _ {t} ^ {a d v, k} = \arg \min _ {\widehat {I} _ {t} ^ {a d v} \in \mathbb {R}} \sum_ {k = 1} ^ {m} - l o g \left(p _ {\varphi} \left(y _ {k} \mid t _ {k} ^ {\prime}, \widehat {I} _ {t} ^ {a d v, k}\right)\right), \tag {4} +$$ + +where $\mathbb{R}$ is some constraint applied to the input space in which we search for adversarial examples, and $p_{\varphi}$ represents the conditional probability generated by the target VLM. To ensure the adversarial images remain visually natural, we constrain the perturbation magnitude by $\left\| I_t^{adv,k} - I_t\right\|_\infty \leq \varepsilon$ where $\varepsilon$ is set to 32/255. + +During extensive experiments with this optimization process, we observe that adversarial images with minimal loss values do not necessarily lead to successful attacks. Specifically, we find cases where images with relatively low loss fail to jailbreak the target VLM, while those with moderately higher loss values achieve better success rates. This observation persists across different models and attack scenarios. + +# 3.4. Multi-image Collaborative Attack + +To better understand the observation that adversarial images with minimal loss do not necessarily lead to the highest at + +tack success rate, we conduct systematic experiments analyzing attack performance across different loss ranges. As shown in Fig. 4a, our results reveal that the optimal attack performance often comes from adversarial images within a certain loss range, rather than those with minimal loss. This finding aligns with the flat minima phenomenon [12] in optimization, where solutions in flatter regions of the loss landscape tend to be more robust than those in sharp minima. As illustrated in Fig. 5, when the loss landscape exhibits a sharp minimum, even small perturbations can lead to significant performance degradation. In contrast, solutions in flat regions are more resilient to input variations, which is particularly relevant for adversarial attacks where the test-time scenario may differ from the optimization objective. + +Based on this insight, we propose to select adversarial images within a carefully determined loss range for collaborative attack. The selection process starts by analyzing the loss landscape around the minimal loss point. Specifically, we select 20 nearest points on each side of the minimum and fit straight lines using least squares method. The loss range $\Theta$ is then determined by: + +$$ +\Theta = \left(\left\lceil \frac {k _ {l}}{K} \right\rceil \times \rho , \left\lceil \frac {k _ {r}}{K} \right\rceil \times \rho\right), \tag {5} +$$ + +where $k_{l}$ and $k_{r}$ are the fitted slopes representing the local geometry of the loss landscape, $K$ is the standard slope derived from normal distribution, and $\rho$ is a range coefficient set to 6. As demonstrated in Fig. 4b, this approach adaptively selects a narrow range in flat regions for computational efficiency, while expanding the range in sharp regions for better robustness. + +The final jailbreak output is produced through the collaborative effort of multiple adversarial images within $\Theta$ , as formulated in Eq. (2). This strategy enables us to maintain both attack effectiveness and computational efficiency. + +# 4. Experiment + +# 4.1. Experimental Setup + +Models and Datasets. To comprehensively evaluate MLAI's effectiveness, we conduct experiments on both open-source and commercial VLMs. For open-source models, we focus on MiniGPT-4 [40] (Vicuna 13B) and LLaVA-2 [19] (13B) due to their widespread adoption and strong performance. We use the official weights provided in their respective repositories. For commercial models, we evaluate on Gemini [11], ChatGLM [5], and Qwen [7] to validate real-world applicability. We evaluate our approaches using two common datasets: SafetyBench [22] and AdvBench [41]. SafetyBench serves as a benchmark for assessing the safety of VLMs, encompassing 13 typically prohibited scenarios or behaviors as outlined in + +
ScenarioMiniGPT-4 [40]LLaVA-2 [19]
Plain TextLiu et al. [21]Qi et al. [29]OursPlain TextLiu et al. [21]Qi et al. [29]Ours
Illegal Activity (IA)1.9214.5411.5570.243.8274.7748.0383.73
Hate Speech (HS)1.6811.923.9766.573.0840.9740.2180.36
Malware Generation (MG)3.3219.8815.5250.2830.7564.6353.8579.10
Physical Harm (PH)2.9824.5123.4372.0213.8861.0450.3179.62
Economic Harm (EH)5.684.916.8070.154.1518.427.9850.18
Fraud (FR)3.1718.5614.7150.674.9757.2047.9775.85
Pornography (PO)4.1420.9419.1169.0120.6461.2341.5568.98
Political Lobbying (PL)67.6779.1176.3396.7474.5991.1585.6293.45
Privacy Violence (PV)8.9710.5012.9778.9220.9249.3637.3374.82
Legal Opinion (LO)74.5685.2882.4895.1680.1695.4682.9493.59
Financial Advice (FA)84.3388.1290.9310087.4897.2187.6996.77
Health Consultation (HC)76.5093.9491.2293.0985.6110090.04100
Gov Decision (GD)90.2991.7591.2597.8587.7498.9791.22100
Average32.7143.3841.5677.7539.8370.0358.8382.80
+ +Table 1. ASR results of jailbreaking MiniGPT4 and LLaVA-2 in 13 scenarios. Our attack achieves the best attacking performance in both modules and almost every scenario. We achieve improvements of $34.37\%$ and $12.77\%$ over existing state-of-the-art approaches on MiniGPT-4 and LLaVA-2, respectively. Additionally, substantial ASR differences can be observed across scenarios, with notable variations in performance improvement across different contexts. + +the usage policies [13, 27] of OpenAI [1] and Meta [42]. AdvBench has been extensively utilized in prior research concerning LLM jailbreak attacks and includes 521 harmful behaviors. We following the setting of BAP [37], removed duplicate entries from AdvBench and integrated each unique item into SafetyBench according to the corresponding scenario for our experiments. + +Evaluation metric. We following the setting of HADES [18], employ Attack Success Rate (ASR) as our primary metric to assess the effectiveness of our approach, which is calculated by: + +$$ +A S R = \frac {\sum_ {k = 1} ^ {N} \mathbf {B} (J (y _ {k}) = T r u e)}{N}, \tag {6} +$$ + +where $y_{k}$ is the model's response, $\mathbf{B}$ is an indicator function that equals to 1 if $J(y_{k}) = True$ and 0 otherwise, $N$ is the total number of instructions and $J(\cdot)$ is the harmfulness judging model, outputting True or False to indicate whether $y_{k}$ is harmful. We adopt Beaver-dam-7B [14] as $J(\cdot)$ , which has been trained on high-quality human feedback data about the above harmful categories. + +Compared attacks. We compare MLAI with two state-of-the-art jailbreak attacks: Liu et al. [21] and Qi et al. [29]. Liu et al. [21] integrated images associated with aggressive intent alongside typographic text to serve as visual adversarial prompts. Qi et al. [29], on the other hand, refined these visual adversarial prompts by leveraging a corpus specific to certain scenarios. Additionally, we include a "Plain Text" baseline where harmful queries are directly input to evaluate models' base vulnerability. + +# 4.2. Experiment Results + +White-box Attacks. Our experimental evaluation focuses on three key aspects: attack effectiveness across models, vulnerability patterns in different scenarios, and performance variations among scenarios. + +The results in Tab. 1 show that MLAI achieves strong attack performance on both tested models. For MiniGPT-4 [40], MLAI reaches $77.75\%$ ASR, surpassing the previous best method $(43.38\%)$ by $34.37\%$ . Similarly on LLaVA-2 [19], MLAI achieves $82.80\%$ ASR, outperforming existing approaches by $12.77\%$ . + +Analysis across scenarios reveals distinct vulnerability patterns in both models. Security-critical scenarios like Malware Generation (MG) and Fraud (FR) show stronger resistance with ASR of $50.28\%$ and $50.67\%$ , while advisory scenarios such as Financial Advice (FA) and Gov Decision (GD) exhibit higher vulnerability $(\mathrm{ASR} > 90\%)$ . This pattern reflects the models' internal defense mechanisms: security-critical scenarios are equipped with more rigorous detection systems due to their well-defined harmful patterns, and these high-risk scenarios are typically protected by enhanced safety measures. + +The effectiveness of MLAI varies significantly across scenarios. Most notably, while achieving a $55.70\%$ improvement in Illegal Activity (IA) scenario, the gain in Gov Decision (GD) scenario is merely $0.85\%$ . This disparity can be primarily explained by two aspects. First, scenarios like Gov Decision (GD) already demonstrate high base vulnerability $(90.29\%)$ ASR without attacks), leaving limited room for improvement. Second, the nature of different scenarios affects our attack strategy: Illegal Activity (IA) scenario + +
ScenarioGemini [11]ChatGLM [5]Qwen [7]
Illegal Activity49.5960.1154.97
Malware Generation39.8442.5645.28
Physical Harm43.3545.6248.46
+ +Table 2. ASR results for black-box models. The ASR differences across various models within the same scenario are minor, and the ranking of ASR aligns closely with those observed during testing on white-box models. + +involve concrete elements (e.g., drug manufacturing) that can be effectively represented in images, whereas Gov Decision (GD) scenario contain abstract concepts that are inherently challenging to visualize, thereby limiting the potential of image-based attacks. + +Black-box Attacks. Besides the open-source models, we also evaluate our attack on three black-box commercial VLMs (i.e., Gemini [11], ChatGLM [5] and Qwen [7]). We selected 3 most harmful scenarios (Illegal Activity, Malware Generation, and Physical Harm) and 60 related queries to attack each model. Since direct access to the internal data of the model is unavailable, we employ images trained on MiniGPT-4 [40] to perform a black-box model jailbreak. As shown in Tab. 2, MLAI maintains considerable attack effectiveness on commercial models, achieving ASRs of $49.59\%$ , $60.11\%$ , and $54.97\%$ on Gemini, ChatGLM, and Qwen respectively. We also notice that compared to white-box attacks, there exists a consistent performance drop ( $16.43\%$ lower on average). This gap likely stems from commercial models' enhanced defense mechanisms and architectural differences. Nevertheless, the consistent attack performance across different commercial models suggests that our method can effectively generalize beyond the training model, despite the presence of unknown defense strategies. + +# 4.3. Ablation Studies + +In this part, we further discuss the effectiveness of our proposed approach, including image harmfulness optimization and loss range determination. We note that except for the explicitly stated modifications, all other parameters are identical to the baseline configuration. + +Image harmfulness optimization. To validate the effectiveness of the optimization process for image generation discussed in Sec. 3.2, we conduct a specific experiment with selecting different types of images to examine the attack performance. We selected two scenarios, Hate Speech and Fraud, and conducted experiments on the MiniGPT-4 [40] and LLaVA-2 [19] models, respectively. Moreover, the Irrelevant Image used in the experiments are fixed and consist of 12 images that is not associated with any of the jailbreak scenarios. As shown in Fig. 6, when blank images + +![](images/e1c07cdd7ffc11e1e5ea53691b5a8cc2e25fb9217723e3932ea2b2d902f30f7c.jpg) +(a) ASR of Fraud. + +![](images/6c05aa229a841b7e3777d92510e539a571923261bc60e87416ce9e754ca45af9.jpg) +(b) ASR of Hate Speech. + +![](images/ed976887c3dcd4706752a5ab8f67807ab290d7393a26b7f99f9156b01e1dce97.jpg) +Figure 6. The ASR of using different types of initial images. We can find that simply adding a blank image can lead to a substantial increase in the success rate. Compared to blank image, further selecting images that contain information but are unrelated to the jailbreak scenario results in a smaller improvement. However, when images are introduced that closely match the specific scenario, the performance reaches state-of-the-art levels. +Figure 7. The ASR of using different numbers of images. According to Eq. (5), since $K$ is a constant value, we set $v = \left\lceil \frac{k_{L}}{K} \right\rceil$ and assume that the slopes on both sides are equal without loss of generality, where $v$ represents the flatness of the minimum value, in order to better focus on the changing trends of ASR. + +are introduced, the alignment of the underlying LLM fails to cover the unforeseen domains introduced by the visual modality. This misalignment effectively circumvents the inherent alignment mechanisms of the LLM within the VLM framework, leading to a marked increase in the ASR. However, when choosing unrelated images, for the target scene, it merely increases the information content and complexity of the image, which does not significantly aid in bypassing restrictions. Only by selecting images that match the scenario can meaningful semantic information be introduced, thereby amplifying the potential risk of the image and ultimately further increasing the ASR. + +Loss range determination. To further validate the impact of using images in different loss range on the improvement of jailbreak success rate, we conduct experiments on four scenarios on the MiniGPT-4 [40] model and the results are shown in Fig. 7. We observe that, overall, the ASR increases + +![](images/b27b1cce159879bbbae84bd425c0a879a97c55e623f840580405628c4b1ad60c.jpg) +Figure 8. The average ASR of using different $\rho$ . The average ASR increases gradually with the value of $\rho$ , with the rate of increase remaining relatively stable. However, once $\rho$ reaches a critical threshold (which is 5 in this experiment), the growth rate drops significantly, and the ASR quickly reaches its maximum value. + +significantly when the number of Images initially increases. However, the growth patterns of the curves in different scenarios show considerable variation, with differences in both the rate of increase and the point at which the growth stops. Specifically, the values of $v$ for Illegal Activity (IA) and Hate Speech (HS) are relatively large, indicating that they have relatively sharp minima, whereas Legal Opinion (LO) and Financial Advice (FA) exhibit flatter minima. Correspondingly, the curves for Illegal Activity (IA) and Hate Speech (HS) stop increasing early (almost cease to grow when the number of images reaches 5), while the curves for Legal Opinion (LO) and Financial Advice (FA) maintain a prolonged growth phase (only stopping near the point where the number of images reaches 9). The results demonstrate that to achieve an optimal ASR, image sets with flat minima require fewer images, while image sets with sharp minima incur higher costs to achieve the desired outcome. These findings prove that flatter minima indeed provide greater robustness, allowing models to perform better when the test set deviates from the training set. + +In addition, we also examined the values of $\rho$ . We tested the average ASR across all scenarios for different values of $\rho$ on both MiniGPT-4 [40] and LLaVA-2 [19]. As shown in Fig. 8, the average ASR reaches its near-maximum value when $\rho = 6$ , while the computational complexity increases at a linear rate. Further increases in $\rho$ yield diminishing returns. Therefore, based on empirical observations, we set the baseline value of $\rho = 6$ . Through the above experiments, we can prove the effectiveness of the flat minimum, the optimal choice of $\rho$ , and we finally find the balance point between the attack success rate and the computational cost. + +# 4.4. Defense Analysis + +From the perspective of developmental trends, defense mechanisms always occupy the weaker position. With the + +![](images/af9cfd232eaf701e8b6d9df2d60aea6839da0e0fa17f18bcbbe6858c000b5683.jpg) +(a) Results of 10 specific questions in fraud scenario. +Figure 9. The demonstration of our defense motivation and results. + +![](images/eb1fa30aeba1e9f044177f294863bd17d3e3373707d1b4d5b345d131164efce9.jpg) +(b) ASR of original/defensive circumstance. + +emergence of an increasing number of VLMs and the continuous improvement of jailbreak methods, comprehensive defense becomes progressively more challenging. In this section, we analyze an effective defense strategy against Multi-Loss Adversarial Images attack. + +As shown in Fig. 9a, we observe that under the same scenario, images with different loss not only vary in overall attack potency but also demonstrate significant differences in their effectiveness against individual queries. Based on this finding, we propose introducing a similarity detection module before image sets are input into VLMs. If a current image is detected to be highly similar to a previously encountered image, it is removed from the input set, thus preventing it from being fed into the VLMs. This approach transforms a Multi-Loss Adversarial Attack back into a Single-Loss Adversarial Attack, and due to the diversity of queries in the scenario, the average ASR achieves robust defense. We conducted experiments on MiniGPT-4 [40], and the results, as shown in Fig. 9b, indicate that our defense method effectively reduces the success rate of the Multi-Loss Adversarial Attack, achieving a $22.99\%$ reduction compared to the model without the defense module. + +# 5. Conclusion + +In this paper, we conducted a systematic analysis of visual vulnerabilities in VLMs safety alignment and introduced MLAI, a novel jailbreak framework that effectively exploits these weaknesses. Our investigation revealed two critical findings: the unreliability of minimal-loss optimization in gradient-based attacks, and the crucial role of semantic alignment between visual content and target scenarios. Based on these insights, MLAI achieves superior performance through scenario-aware image generation, flat minima-based optimization, and multi-image collaborative attacks. Extensive experiments demonstrate MLAI's effectiveness, achieving attack success rates of $77.75\%$ on MiniGPT-4 and $82.80\%$ on LLaVA-2. The strong transferability to commercial black-box VLMs further highlights fundamental challenges in achieving robust safety alignment for multi-modal systems. + +# 6. Acknowledgement + +This research has been partially supported by Australian Research Council Discovery Project DP230101196. + +# References + +[1] Sam Altman and et al. OpenAI, 2015. 6 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2 +[3] Nicholas Carlini, Milad Nasr, Christopher A. Choquette-Choo, Matthew Jagielski, Irena Gao, Anas Awadalla, Pang Wei Koh, Daphne Ippolito, Katherine Lee, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned?, 2024. 3 +[4] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J. Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries, 2024. 3 +[5] Chatglm. Zhipu ai, 2023. 2, 5, 7 +[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning, 2023. 2 +[7] Alibaba Cloud. Qwen, 2023. 2, 5, 7 +[8] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks?, 2023. 3 +[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 3 +[10] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 3 +[11] Google. Gemini, 2024. 2, 5, 7 +[12] Sepp Hochreiter and Jurgen Schmidhuber. Flat minima. Neural computation, 9:1-42, 1997. 5 +[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, and Madian Khabsa. Llama guard: Llm-based input-output safeguard for human-ai conversations, 2023. 6 +[14] Jiaming Ji, Mickel Liu, Juntao Dai, Xuehai Pan, Chi Zhang, Ce Bian, Chi Zhang, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. arXiv preprint arXiv:2307.04657, 2023. 6 +[15] Tomasz Korbak, Kejian Shi, Angelica Chen, Rasika Bhalerao, Christopher L. Buckley, Jason Phang, Samuel R. + +Bowman, and Ethan Perez. Pretraining language models with human preferences, 2023. 1, 3 +[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension, 2023. 3 +[17] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models, 2024. 3 +[18] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2, 3, 6 +[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 5, 6, 7, 8 +[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2024. 2 +[21] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 6 +[22] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 5 +[23] Yugeng Liu, Tianshuo Cong, Zhengyu Zhao, Michael Backes, Yun Shen, and Yang Zhang. Robustness over time: Understanding adversarial examples' effectiveness on longitudinal versions of large language models, 2024. 1 +[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024. 3 +[25] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model, 2024. 3 +[26] OpenAI. Dall·E 3, 2023. 4 +[27] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, and et al. Florencia Leoni Aleman and. Gpt-4 technical report, 2024. 6 +[28] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3419-3448, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3 +[29] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models, 2023. 1, 2, 3, 6 +[30] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to!, 2023. 1 +[31] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models, 2023. 3 +[32] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3 + +[33] Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, Lewis Ho, Divya Siddarth, Shahar Avin, Will Hawkins, Been Kim, Jason Gabriel, Vijay Bolina, Jack Clark, Yoshua Bengio, Paul Christiano, and Allan Dafoe. Model evaluation for extreme risks, 2023. 1, 3 +[34] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image, 2024. 3 +[35] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms, 2023. 3 +[36] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts, 2024. 3 +[37] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt, 2024. 1, 2, 3, 6 +[38] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3 +[39] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models, 2023. 3 +[40] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023, 2, 5, 6, 7, 8 +[41] Andy Zou, Zifan Wang, J. Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models, 2023. 5 +[42] Mark Elliot Zuckerberg. Meta, 2004. 6 \ No newline at end of file diff --git a/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/images.zip b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..97b12aab25a0b51cbbb4da382b70e17f5ef1dcb8 --- /dev/null +++ b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32f3bc0d26c7f212058363f4f07d527027ac788a5d70b20adf24daa57c5732dd +size 615919 diff --git a/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/layout.json b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e0b7fc5dd3c71726a53459c1ce2e83fff70ca04 --- /dev/null +++ b/2025/Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models/layout.json @@ -0,0 +1,7720 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "text", + "content": "Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "spans": [ + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": "Shuyang Hao " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": " Bryan Hooi " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": " Jun Liu " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": " Kai-Wei Chang " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": " Zi Huang " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": " Yujun Cai " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{5*}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": " Southeast University, " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": "National University of Singapore, " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": "Lancaster University " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": "University of California, Los Angeles, " + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 81, + 159, + 532, + 232 + ], + "type": "text", + "content": "University of Queensland haosy@seu.edu.cn, dcsbhk@nus.edu.sg, j.liu81@lancaster.ac.uk, kw@kwchang.net, huang@itee.uq.edu.au, yujun.cai@uq.edu.au" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 242, + 200, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 242, + 200, + 254 + ], + "spans": [ + { + "bbox": [ + 151, + 242, + 200, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "text", + "content": "Despite inheriting security measures from underlying language models, Vision-Language Models (VLMs) may still be vulnerable to safety alignment issues. Through empirical analysis, we uncover two critical findings: scenario-matched images can significantly amplify harmful outputs, and contrary to common assumptions in gradient-based attacks, minimal loss values do not guarantee optimal attack effectiveness. Building on these insights, we introduce MLAI (Multi-Loss Adversarial Images), a novel jailbreak framework that leverages scenario-aware image generation for semantic alignment, exploits flat minima theory for robust adversarial image selection, and employs multi-image collaborative attacks for enhanced effectiveness. Extensive experiments demonstrate MLAI's significant impact, achieving attack success rates of " + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "inline_equation", + "content": "77.75\\%" + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "text", + "content": " on MiniGPT-4 and " + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "inline_equation", + "content": "82.80\\%" + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "text", + "content": " on LLaVA-2, substantially outperforming existing methods by margins of " + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "inline_equation", + "content": "34.37\\%" + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "inline_equation", + "content": "12.77\\%" + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "text", + "content": " respectively. Furthermore, MLAI shows considerable transferability to commercial black-box VLMs, achieving up to " + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "inline_equation", + "content": "60.11\\%" + }, + { + "bbox": [ + 54, + 268, + 296, + 543 + ], + "type": "text", + "content": " success rate. Our work reveals fundamental visual vulnerabilities in current VLMs safety mechanisms and underscores the need for stronger defenses. Warning: This paper contains potentially harmful example text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 566, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) have demonstrated remarkable capabilities in understanding and generating multimodal content, enabling various applications from content creation to visual reasoning. To prevent potential misuse, these models typically undergo comprehensive safety protocols [23, 30] including safety-focused pre-training and reinforcement learning with human feedback (RLHF) [15, 28, 33]. Despite these safeguards, ensuring the safe deployment of these increasingly powerful models remains a criti-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 242, + 553, + 586 + ], + "blocks": [ + { + "bbox": [ + 318, + 242, + 553, + 586 + ], + "lines": [ + { + "bbox": [ + 318, + 242, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 318, + 242, + 553, + 586 + ], + "type": "image", + "image_path": "0125f4c2e3c819b2eb7f0430bee864c98b8447df6902dfc100cf6fac7f4334be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 596, + 555, + 673 + ], + "lines": [ + { + "bbox": [ + 313, + 596, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 555, + 673 + ], + "type": "text", + "content": "Figure 1. An example to show the Visual Vulnerabilities of the scenario-matched images on alignment of VLMs and the unreliability of using image with minimal loss (the cross-entropy loss between the model's output and the target in gradient-based optimization). We can find that: (1) matching images are better than irrelevant or no images, and (2) since only the last jailbreak is successful, lower loss is not always better." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "cal challenge, as recent studies [10, 18, 29, 32, 37] have revealed that visual inputs can significantly compromise these" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19890" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 136, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 136, + 83 + ], + "type": "text", + "content": "safety mechanisms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 296, + 217 + ], + "type": "text", + "content": "The vulnerability of VLMs to visual inputs is particularly concerning as these models are increasingly deployed in both open-source and commercial platforms. In this paper, we investigate jailbreak attacks that aim to circumvent VLMs' safety mechanisms through carefully crafted adversarial images. This vulnerability is particularly critical as VLMs are increasingly deployed in both open-source and commercial platforms. Our study primarily focuses on white-box [19, 40] settings where model parameters are accessible, with additional evaluation on commercial blackbox VLMs [5, 7, 11] to assess real-world implications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 219, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 459 + ], + "type": "text", + "content": "For white-box attacks, current approaches [18, 29, 37] typically rely on gradient-based optimization of adversarial images. These methods iteratively refine random noise patterns by minimizing the cross-entropy loss between the model's generated response and the desired harmful output. Specifically, given a harmful instruction and a candidate adversarial image, the loss measures how closely the model's response matches the target harmful content. Through systematic analysis, we identify two critical limitations in this paradigm. As illustrated in Fig. 1, first, adversarial images with minimal loss values do not necessarily lead to successful attacks, challenging a fundamental assumption in gradient-based optimization. When repeating attacks with different initial images, we observe significant variations in success rates, suggesting that loss minimization alone is not a reliable strategy. Second, using a single image for attacking all scenarios is not an effective approach, as the harmfulness of the VLM's response is strongly correlated with the semantic relevance of the image content to the target scenario, which is also demonstrated in previous work [18]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 687 + ], + "type": "text", + "content": "Building on these insights, we propose a novel jailbreak method called MLAI, which emphasizes targeted image generation and the use of Multi-Loss Adversarial Images to amplify harmful intent and disrupt multi-modal alignment. Specifically, our method introduces a three-stage attack strategy. First, MLAI generates initial images that match the target scenario. In this step, we generate the image with the highest potential for harmful output in the specific scenario while ensuring that the adversarial image remains visually natural, making the model more susceptible to generating unsafe responses. Second, MLAI optimizes the adversarial image through gradient updates and saves results that fall within a predefined loss range during the iterative process. In the process of iterative refinement, the harmfulness of the image is progressively amplified. Third, MLAI carefully calculate the loss range and select adversarial images within the loss range for collaborative attack. The determination of the loss range takes into account the balance between attack success rate and computational cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Interestingly, MLAI demonstrates considerable transferability across different models. Our experiments show" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "that adversarial images trained on white-box models maintain significant attack success rates when applied to unseen black-box commercial VLMs. This transferability suggests fundamental vulnerabilities in current VLM safety mechanisms that transcend specific model architectures and training approaches. To mitigate this problem, we propose a similarity-based dedduplication defense that effectively mitigates MLAI's impact by limiting the number of similar adversarial input images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 180, + 527, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 180, + 527, + 191 + ], + "spans": [ + { + "bbox": [ + 325, + 180, + 527, + 191 + ], + "type": "text", + "content": "In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 192, + 553, + 441 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 192, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 553, + 287 + ], + "type": "text", + "content": "- We undertake a comprehensive analysis of the gradient-based attack process. Through targeted experiments, we explore the selection of initial images and examine the outcomes of adversarial images with varying loss values. The findings highlight that strategies relying exclusively on the adversarial image with the lowest loss or using a single initial image are suboptimal, offering new insights into attack effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 288, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 553, + 358 + ], + "type": "text", + "content": "- We introduce a novel jailbreak method, MLAI, which leverages images that match the target scenario along with a collaborative approach involving multi-loss adversarial images. This combined strategy amplifies the harmful intent of the attack, successfully disrupting multi-modal alignment in VLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 360, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 553, + 441 + ], + "type": "text", + "content": "- Experimental results show that MLAI achieves a remarkable attack success rate (ASR) of " + }, + { + "bbox": [ + 314, + 360, + 553, + 441 + ], + "type": "inline_equation", + "content": "77.75\\%" + }, + { + "bbox": [ + 314, + 360, + 553, + 441 + ], + "type": "text", + "content": " on MiniGPT-4 and " + }, + { + "bbox": [ + 314, + 360, + 553, + 441 + ], + "type": "inline_equation", + "content": "82.80\\%" + }, + { + "bbox": [ + 314, + 360, + 553, + 441 + ], + "type": "text", + "content": " on LLaVA-2, making existing white-box VLMs highly vulnerable. Additionally, we demonstrate the potential of MLAI against black-box commercial VLMs and propose a countermeasure to defend against these advanced attacks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 453, + 400, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 400, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 400, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": "Vision Language Models (VLMs). Leveraging the remarkable capabilities of LLMs, the field has seen the emergence of several prominent VLMs. Notable among these are LLaVA [19] and MiniGPT-4 [40], which integrate an LLM with a vision encoder and a cross-modal connector. These models adopt a two-stage training framework, comprising cross-modal representation alignment and visual instruction tuning. Building upon this foundational approach, a variety of open-source VLMs [20] have been developed, which enhance multimodal performance through several key advancements: the use of higher-resolution inputs, more powerful LLM architectures, expanded pretraining corpora [2], and the introduction of specialized pretraining tasks [2, 6]. These efforts have significantly improved the models' ability to process and generate content across modalities, further advancing the state of multimodal understanding and generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Alignment Vulnerabilities in VLMs. Similar to LLMs, VLMs are also vulnerable to malicious inputs, which can" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19891" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 406 + ], + "type": "text", + "content": "lead to the generation of harmful content [29, 38]. However, existing benchmarks predominantly focus on evaluating the effectiveness of VLMs, with insufficient attention given to their alignment vulnerabilities, particularly regarding harmlessness [9, 16, 24]. To address this gap, recent research has developed red-teaming attack benchmarks that systematically assess the alignment vulnerabilities of VLMs across various scenarios [17, 24, 35]. In line with LLM research, studies on jailbreaking VLMs can be divided into white-box, and black-box attack strategies. For white-box attacks, existing methods [3, 8, 29, 31, 32, 34] can be further categorized based on their target: input images or visual embeddings. Attacks on input images often generate adversarial images constrained to provoke harmful responses [8, 29, 31] or use teacher-forcing optimization [3]. For visual embeddings, adversarial images are crafted to appear harmless but mimic the embeddings of harmful images [32], thus bypassing content filters. On the other hand, black-box attacks typically undermine the alignment of VLMs by techniques such as system prompt attacks [4, 36], converting harmful information into text-oriented images [10], leveraging surrogate models to generate adversarial images [39], or utilizing maximum likelihood-based jailbreaking methods [25]. Our work extends this line of research by systematically analyzing the visual-semantic vulnerabilities in VLM safety alignment and introducing novel attack strategies that leverage scenario-matched images and flat minima characteristics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 418, + 162, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 162, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 162, + 430 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 437, + 124, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 124, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 124, + 449 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": "Recent studies [10, 18, 29, 32, 37] have revealed that VLMs, despite inheriting safety measures from their foundational LLMs, remain vulnerable to jailbreak attacks. Given a vision-language model built based on a LLM M, an image encoder " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": " and a projection layer " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": ", its generation process can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 536, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 536, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 128, + 536, + 295, + 550 + ], + "type": "interline_equation", + "content": "y = \\mathbf {M} ([ W \\cdot E (i), t ]), \\tag {1}", + "image_path": "e1b810392aee6ace35cf3f0485f75cdd5bf5f2e9ec93b52cf9d62ef68c7697c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": " are the input image and text, and " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": " is the model's output. Through empirical analysis of gradient-based attack methods, we observe two key limitations: (1) arbitrarily selected initial images fail to fully exploit the correlation between image content and model output harmfulness, and (2) our experiments show that adversarial images with minimal loss do not necessarily lead to the highest attack success rate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "To address these challenges, we propose MLAI, a novel jailbreak framework that enhances attack effectiveness through scenario-aware image generation and multi-image collaborative attack. As illustrated in Fig. 3, MLAI first generates initial images that match the target scenario," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 351, + 69, + 521, + 203 + ], + "blocks": [ + { + "bbox": [ + 351, + 69, + 521, + 203 + ], + "lines": [ + { + "bbox": [ + 351, + 69, + 521, + 203 + ], + "spans": [ + { + "bbox": [ + 351, + 69, + 521, + 203 + ], + "type": "image", + "image_path": "d2e31a66865086b8353a7c44f8c5e00e2dfd90fd12439c43774aa07a0ae69a74.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 210, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 287 + ], + "type": "text", + "content": "Figure 2. The ASR of transferability across scenarios. The heat map shows that images generated under the IA scenario achieve high transferability when applied to MG contexts, while FR scenario images transfer effectively to LO settings. These transferability patterns are bidirectional between paired scenarios. In contrast, PO scenario images show minimal transfer effectiveness across other contexts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "content": "then obtains a set of adversarial images through gradient optimization, and finally utilizes multiple images within an optimal loss range for the collaborative attack. The complete attack process can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 367, + 555, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 367, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 364, + 367, + 555, + 399 + ], + "type": "interline_equation", + "content": "y ^ {*} = \\sum_ {k} ^ {\\Theta} \\mathbf {M} (\\left[ W \\cdot E (I _ {t} ^ {a d v, k}), t \\right]), \\qquad (2)", + "image_path": "b1c946dddfe1273daf37c3313ce2b7a8c612c93e04a64c798d661b7f548f693d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "inline_equation", + "content": "y^{*}" + }, + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "text", + "content": " is the attack generation, " + }, + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "inline_equation", + "content": "I_{t}^{adv,k}" + }, + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "text", + "content": " represents the k-th adversarial image and " + }, + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 410, + 555, + 436 + ], + "type": "text", + "content": " denotes the selected loss range." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 443, + 499, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 499, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 499, + 455 + ], + "type": "text", + "content": "3.2. Scenario-aware Image Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 581 + ], + "type": "text", + "content": "Through training on human preference data, existing LLMs [15, 28, 33] have learned to align with human values and refuse to respond to harmful text inputs. However, recent studies [18] indicate that incorporating images into inputs can significantly increase the likelihood of generating harmful content, and the harmfulness of VLM outputs is often positively correlated with the harmfulness of the visual content. This observation suggests that visual components introduce additional vulnerabilities, which can be exploited to circumvent the alignment of the underlying LLMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "By comparing the transferability of generated images across different scenarios, we further validate the importance of semantic matching between images and attack targets. As shown in Fig. 2, images generated under certain scenarios demonstrate high transferability to semantically related contexts. For instance, instructions within Illegal Activity (IA) and Malware Generation (MG) scenarios share common harmful action patterns like \"creating\" or \"utilizing\", leading to effective cross-scenario transfers. Similarly, Fraud (FR) and Legal Opinion (LO) scenarios often involve abstract harmful concepts like \"concealing\"" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19892" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 306, + 149 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 306, + 149 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 306, + 149 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 306, + 149 + ], + "type": "image", + "image_path": "d8a4e932007eee462732a713a262cdbe14f427acb5dc3505926d210b66ffb3ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 155, + 305, + 297 + ], + "blocks": [ + { + "bbox": [ + 67, + 155, + 305, + 297 + ], + "lines": [ + { + "bbox": [ + 67, + 155, + 305, + 297 + ], + "spans": [ + { + "bbox": [ + 67, + 155, + 305, + 297 + ], + "type": "image", + "image_path": "e3bbd64101760aa6daeb6a68f554d5e4abc795adb9a87499d1b2c794ef435e23.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 67, + 301, + 304, + 446 + ], + "blocks": [ + { + "bbox": [ + 67, + 301, + 304, + 446 + ], + "lines": [ + { + "bbox": [ + 67, + 301, + 304, + 446 + ], + "spans": [ + { + "bbox": [ + 67, + 301, + 304, + 446 + ], + "type": "image", + "image_path": "3a4b2e2df1c879b35f0e7e7afeb40d14e020be11b59ca0f9120a5c7302d8cac7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 456, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 456, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 555, + 489 + ], + "type": "text", + "content": "Figure 3. Our MLAI framework involves a three-step procedure: (1) generate an image that matches the text scenario as initial image, (2) obtain the adversarial image set based on the initial image by gradient update, and (3) calculate the loss range and select adversarial images within the loss range for collaborative attack." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 545, + 446 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 446 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 446 + ], + "type": "image", + "image_path": "6b2434129c3d7a81b01c5bc66874756e731174728dd7cd9bbf34606b790c6b34.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 297, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 559 + ], + "type": "text", + "content": "or \"deceiving\". In contrast, scenarios with distinct semantic frameworks, such as Pornography (PO), show minimal transfer effectiveness, highlighting the crucial role of semantic alignment in attack success." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Motivated by these findings, we propose to generate initial images that match the semantic content of jailbreak targets. Given that jailbreak-related scenarios often involve abstract concepts that are challenging to accurately depict, we first classify jailbreak targets " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " into 13 distinct categories " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "C = \\{c_{1}, c_{2}, \\dots, c_{13}\\}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " based on common harmful behaviors (e.g., Illegal Activity, Hate Speech and Fraud). For each category, we carefully craft prompts that encode contextual information while avoiding direct depiction of harmful elements. For example, when targeting fraud-related scenarios, we generate images of relevant but benign contexts like business meetings or document processing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "text", + "content": "These carefully crafted prompts are then used with DALL-E 3 [26] to generate the initial images:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 402, + 540, + 555, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 540, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 402, + 540, + 555, + 553 + ], + "type": "interline_equation", + "content": "I _ {t} = G (P (t)), \\tag {3}", + "image_path": "93a87b58ce90c6f56478076abbb08e1005fd73d97cd28608fafe2864940dc098.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 557, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 557, + 555, + 666 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 313, + 557, + 555, + 666 + ], + "type": "text", + "content": " generates the scenario-specific prompt and " + }, + { + "bbox": [ + 313, + 557, + 555, + 666 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 313, + 557, + 555, + 666 + ], + "type": "text", + "content": " represents the image generation model. Through this design, the generated images establish strong semantic connections with the jailbreak targets while maintaining a proper balance between attack effectiveness and content safety. Our subsequent experiments in Sec. 4.3 demonstrate that such scenario-aware initial images significantly improve the success rate of gradient-based attacks compared to randomly selected images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 671, + 491, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 491, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 491, + 685 + ], + "type": "text", + "content": "3.3. Adversarial Image Optimization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Given the scenario-aware initial image " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ", we further optimize it through gradient updates to enhance the attack effec" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19893" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 175, + 165 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 175, + 165 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 175, + 165 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 175, + 165 + ], + "type": "image", + "image_path": "6a81ad4b175b629cd621173ed4ec51d9b84bc2d8139997379a1ec865338a6e15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 167, + 157, + 176 + ], + "lines": [ + { + "bbox": [ + 72, + 167, + 157, + 176 + ], + "spans": [ + { + "bbox": [ + 72, + 167, + 157, + 176 + ], + "type": "text", + "content": "(a) ASR of different Loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 76, + 295, + 167 + ], + "blocks": [ + { + "bbox": [ + 178, + 76, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 178, + 76, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 178, + 76, + 295, + 167 + ], + "type": "image", + "image_path": "e706e8e2656b10bcc9583fb276927fee48da97a4cf9255a65a9fe6cb347ae25f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 168, + 282, + 176 + ], + "lines": [ + { + "bbox": [ + 190, + 168, + 282, + 176 + ], + "spans": [ + { + "bbox": [ + 190, + 168, + 282, + 176 + ], + "type": "text", + "content": "(b) ASR of Min/Multi-Loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 222, + 291, + 323 + ], + "blocks": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "type": "text", + "content": "Figure 4. Results of the effect of various Loss on ASR. We can clearly observe the limitations of the minimum loss strategy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 222, + 291, + 323 + ], + "lines": [ + { + "bbox": [ + 63, + 222, + 291, + 323 + ], + "spans": [ + { + "bbox": [ + 63, + 222, + 291, + 323 + ], + "type": "image", + "image_path": "77e0a0f2c07284bc2ce7f5a20e2f57c316f89318e34ab876457a29086670a652.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 295, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 354 + ], + "type": "text", + "content": "Figure 5. Comparison of the effects of flat and sharp minima at test set shift. Here the test set shift is simulated by curve translation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": "tiveness. First, we construct a small corpus " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "Y := \\{y_k\\}_{k=1}^m" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": " consisting of harmful statements and affirmative responses, which serves as target outputs for optimization. For each batch of harmful instructions " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": " and corresponding adversarial images " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "I_t^{adv}" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": ", we compute the cross-entropy loss between model outputs and the target responses:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 453, + 295, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 453, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 67, + 453, + 295, + 485 + ], + "type": "interline_equation", + "content": "I _ {t} ^ {a d v, k} = \\arg \\min _ {\\widehat {I} _ {t} ^ {a d v} \\in \\mathbb {R}} \\sum_ {k = 1} ^ {m} - l o g \\left(p _ {\\varphi} \\left(y _ {k} \\mid t _ {k} ^ {\\prime}, \\widehat {I} _ {t} ^ {a d v, k}\\right)\\right), \\tag {4}", + "image_path": "c397ca5e92bb94a349854a92c77e3f36519086643e3f3487827d3ebab1c26dfb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "text", + "content": " is some constraint applied to the input space in which we search for adversarial examples, and " + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "inline_equation", + "content": "p_{\\varphi}" + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "text", + "content": " represents the conditional probability generated by the target VLM. To ensure the adversarial images remain visually natural, we constrain the perturbation magnitude by " + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\left\\| I_t^{adv,k} - I_t\\right\\|_\\infty \\leq \\varepsilon" + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 491, + 295, + 569 + ], + "type": "text", + "content": " is set to 32/255." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 569, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 664 + ], + "type": "text", + "content": "During extensive experiments with this optimization process, we observe that adversarial images with minimal loss values do not necessarily lead to successful attacks. Specifically, we find cases where images with relatively low loss fail to jailbreak the target VLM, while those with moderately higher loss values achieve better success rates. This observation persists across different models and attack scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 237, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 237, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 237, + 684 + ], + "type": "text", + "content": "3.4. Multi-image Collaborative Attack" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "To better understand the observation that adversarial images with minimal loss do not necessarily lead to the highest at" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "tack success rate, we conduct systematic experiments analyzing attack performance across different loss ranges. As shown in Fig. 4a, our results reveal that the optimal attack performance often comes from adversarial images within a certain loss range, rather than those with minimal loss. This finding aligns with the flat minima phenomenon [12] in optimization, where solutions in flatter regions of the loss landscape tend to be more robust than those in sharp minima. As illustrated in Fig. 5, when the loss landscape exhibits a sharp minimum, even small perturbations can lead to significant performance degradation. In contrast, solutions in flat regions are more resilient to input variations, which is particularly relevant for adversarial attacks where the test-time scenario may differ from the optimization objective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 251, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 335 + ], + "type": "text", + "content": "Based on this insight, we propose to select adversarial images within a carefully determined loss range for collaborative attack. The selection process starts by analyzing the loss landscape around the minimal loss point. Specifically, we select 20 nearest points on each side of the minimum and fit straight lines using least squares method. The loss range " + }, + { + "bbox": [ + 313, + 251, + 553, + 335 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 251, + 553, + 335 + ], + "type": "text", + "content": " is then determined by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 342, + 553, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 342, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 369, + 342, + 553, + 370 + ], + "type": "interline_equation", + "content": "\\Theta = \\left(\\left\\lceil \\frac {k _ {l}}{K} \\right\\rceil \\times \\rho , \\left\\lceil \\frac {k _ {r}}{K} \\right\\rceil \\times \\rho\\right), \\tag {5}", + "image_path": "3219824d661817e88cfae832d3b7a7382433664cd83b3a419efc8d343019911d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "inline_equation", + "content": "k_{r}" + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "text", + "content": " are the fitted slopes representing the local geometry of the loss landscape, " + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "text", + "content": " is the standard slope derived from normal distribution, and " + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 377, + 554, + 460 + ], + "type": "text", + "content": " is a range coefficient set to 6. As demonstrated in Fig. 4b, this approach adaptively selects a narrow range in flat regions for computational efficiency, while expanding the range in sharp regions for better robustness." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 461, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 510 + ], + "type": "text", + "content": "The final jailbreak output is produced through the collaborative effort of multiple adversarial images within " + }, + { + "bbox": [ + 313, + 461, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 461, + 554, + 510 + ], + "type": "text", + "content": ", as formulated in Eq. (2). This strategy enables us to maintain both attack effectiveness and computational efficiency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 520, + 391, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 391, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 391, + 533 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "Models and Datasets. To comprehensively evaluate MLAI's effectiveness, we conduct experiments on both open-source and commercial VLMs. For open-source models, we focus on MiniGPT-4 [40] (Vicuna 13B) and LLaVA-2 [19] (13B) due to their widespread adoption and strong performance. We use the official weights provided in their respective repositories. For commercial models, we evaluate on Gemini [11], ChatGLM [5], and Qwen [7] to validate real-world applicability. We evaluate our approaches using two common datasets: SafetyBench [22] and AdvBench [41]. SafetyBench serves as a benchmark for assessing the safety of VLMs, encompassing 13 typically prohibited scenarios or behaviors as outlined in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19894" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 256 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 256 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 256 + ], + "type": "table", + "html": "
ScenarioMiniGPT-4 [40]LLaVA-2 [19]
Plain TextLiu et al. [21]Qi et al. [29]OursPlain TextLiu et al. [21]Qi et al. [29]Ours
Illegal Activity (IA)1.9214.5411.5570.243.8274.7748.0383.73
Hate Speech (HS)1.6811.923.9766.573.0840.9740.2180.36
Malware Generation (MG)3.3219.8815.5250.2830.7564.6353.8579.10
Physical Harm (PH)2.9824.5123.4372.0213.8861.0450.3179.62
Economic Harm (EH)5.684.916.8070.154.1518.427.9850.18
Fraud (FR)3.1718.5614.7150.674.9757.2047.9775.85
Pornography (PO)4.1420.9419.1169.0120.6461.2341.5568.98
Political Lobbying (PL)67.6779.1176.3396.7474.5991.1585.6293.45
Privacy Violence (PV)8.9710.5012.9778.9220.9249.3637.3374.82
Legal Opinion (LO)74.5685.2882.4895.1680.1695.4682.9493.59
Financial Advice (FA)84.3388.1290.9310087.4897.2187.6996.77
Health Consultation (HC)76.5093.9491.2293.0985.6110090.04100
Gov Decision (GD)90.2991.7591.2597.8587.7498.9791.22100
Average32.7143.3841.5677.7539.8370.0358.8382.80
", + "image_path": "3615682832fc2c4ea4048e0a8b7da9face3378db3fe564a078a85ac63af7746d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 555, + 309 + ], + "type": "text", + "content": "Table 1. ASR results of jailbreaking MiniGPT4 and LLaVA-2 in 13 scenarios. Our attack achieves the best attacking performance in both modules and almost every scenario. We achieve improvements of " + }, + { + "bbox": [ + 55, + 264, + 555, + 309 + ], + "type": "inline_equation", + "content": "34.37\\%" + }, + { + "bbox": [ + 55, + 264, + 555, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 264, + 555, + 309 + ], + "type": "inline_equation", + "content": "12.77\\%" + }, + { + "bbox": [ + 55, + 264, + 555, + 309 + ], + "type": "text", + "content": " over existing state-of-the-art approaches on MiniGPT-4 and LLaVA-2, respectively. Additionally, substantial ASR differences can be observed across scenarios, with notable variations in performance improvement across different contexts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 329, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 413 + ], + "type": "text", + "content": "the usage policies [13, 27] of OpenAI [1] and Meta [42]. AdvBench has been extensively utilized in prior research concerning LLM jailbreak attacks and includes 521 harmful behaviors. We following the setting of BAP [37], removed duplicate entries from AdvBench and integrated each unique item into SafetyBench according to the corresponding scenario for our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 473 + ], + "type": "text", + "content": "Evaluation metric. We following the setting of HADES [18], employ Attack Success Rate (ASR) as our primary metric to assess the effectiveness of our approach, which is calculated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 479, + 295, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 479, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 102, + 479, + 295, + 506 + ], + "type": "interline_equation", + "content": "A S R = \\frac {\\sum_ {k = 1} ^ {N} \\mathbf {B} (J (y _ {k}) = T r u e)}{N}, \\tag {6}", + "image_path": "8a25e8b4140c59d5658adc24ec2f64b1c523db1d0d0df1fcb54d0f3fae611795.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " is the model's response, " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " is an indicator function that equals to 1 if " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "J(y_{k}) = True" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " and 0 otherwise, " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " is the total number of instructions and " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "J(\\cdot)" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " is the harmfulness judging model, outputting True or False to indicate whether " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " is harmful. We adopt Beaver-dam-7B [14] as " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "J(\\cdot)" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": ", which has been trained on high-quality human feedback data about the above harmful categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Compared attacks. We compare MLAI with two state-of-the-art jailbreak attacks: Liu et al. [21] and Qi et al. [29]. Liu et al. [21] integrated images associated with aggressive intent alongside typographic text to serve as visual adversarial prompts. Qi et al. [29], on the other hand, refined these visual adversarial prompts by leveraging a corpus specific to certain scenarios. Additionally, we include a \"Plain Text\" baseline where harmful queries are directly input to evaluate models' base vulnerability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 329, + 430, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 329, + 430, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 329, + 430, + 342 + ], + "type": "text", + "content": "4.2. Experiment Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "text", + "content": "White-box Attacks. Our experimental evaluation focuses on three key aspects: attack effectiveness across models, vulnerability patterns in different scenarios, and performance variations among scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": "The results in Tab. 1 show that MLAI achieves strong attack performance on both tested models. For MiniGPT-4 [40], MLAI reaches " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "77.75\\%" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": " ASR, surpassing the previous best method " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "(43.38\\%)" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "34.37\\%" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": ". Similarly on LLaVA-2 [19], MLAI achieves " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "82.80\\%" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": " ASR, outperforming existing approaches by " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "12.77\\%" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "text", + "content": "Analysis across scenarios reveals distinct vulnerability patterns in both models. Security-critical scenarios like Malware Generation (MG) and Fraud (FR) show stronger resistance with ASR of " + }, + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "inline_equation", + "content": "50.28\\%" + }, + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "inline_equation", + "content": "50.67\\%" + }, + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "text", + "content": ", while advisory scenarios such as Financial Advice (FA) and Gov Decision (GD) exhibit higher vulnerability " + }, + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "inline_equation", + "content": "(\\mathrm{ASR} > 90\\%)" + }, + { + "bbox": [ + 312, + 472, + 555, + 604 + ], + "type": "text", + "content": ". This pattern reflects the models' internal defense mechanisms: security-critical scenarios are equipped with more rigorous detection systems due to their well-defined harmful patterns, and these high-risk scenarios are typically protected by enhanced safety measures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "The effectiveness of MLAI varies significantly across scenarios. Most notably, while achieving a " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "55.70\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " improvement in Illegal Activity (IA) scenario, the gain in Gov Decision (GD) scenario is merely " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.85\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ". This disparity can be primarily explained by two aspects. First, scenarios like Gov Decision (GD) already demonstrate high base vulnerability " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "(90.29\\%)" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " ASR without attacks), leaving limited room for improvement. Second, the nature of different scenarios affects our attack strategy: Illegal Activity (IA) scenario" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19895" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 122 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 122 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 122 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 122 + ], + "type": "table", + "html": "
ScenarioGemini [11]ChatGLM [5]Qwen [7]
Illegal Activity49.5960.1154.97
Malware Generation39.8442.5645.28
Physical Harm43.3545.6248.46
", + "image_path": "804a39b90fa6c2d1d4bece21c81c2a68765a61a32e08032f8908ec83a182ea48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 130, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 130, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 295, + 175 + ], + "type": "text", + "content": "Table 2. ASR results for black-box models. The ASR differences across various models within the same scenario are minor, and the ranking of ASR aligns closely with those observed during testing on white-box models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 186, + 296, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 246 + ], + "type": "text", + "content": "involve concrete elements (e.g., drug manufacturing) that can be effectively represented in images, whereas Gov Decision (GD) scenario contain abstract concepts that are inherently challenging to visualize, thereby limiting the potential of image-based attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "text", + "content": "Black-box Attacks. Besides the open-source models, we also evaluate our attack on three black-box commercial VLMs (i.e., Gemini [11], ChatGLM [5] and Qwen [7]). We selected 3 most harmful scenarios (Illegal Activity, Malware Generation, and Physical Harm) and 60 related queries to attack each model. Since direct access to the internal data of the model is unavailable, we employ images trained on MiniGPT-4 [40] to perform a black-box model jailbreak. As shown in Tab. 2, MLAI maintains considerable attack effectiveness on commercial models, achieving ASRs of " + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "inline_equation", + "content": "49.59\\%" + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "inline_equation", + "content": "60.11\\%" + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "inline_equation", + "content": "54.97\\%" + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "text", + "content": " on Gemini, ChatGLM, and Qwen respectively. We also notice that compared to white-box attacks, there exists a consistent performance drop (" + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "inline_equation", + "content": "16.43\\%" + }, + { + "bbox": [ + 56, + 257, + 296, + 497 + ], + "type": "text", + "content": " lower on average). This gap likely stems from commercial models' enhanced defense mechanisms and architectural differences. Nevertheless, the consistent attack performance across different commercial models suggests that our method can effectively generalize beyond the training model, despite the presence of unknown defense strategies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 503, + 156, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 156, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 156, + 514 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "text", + "content": "In this part, we further discuss the effectiveness of our proposed approach, including image harmfulness optimization and loss range determination. We note that except for the explicitly stated modifications, all other parameters are identical to the baseline configuration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Image harmfulness optimization. To validate the effectiveness of the optimization process for image generation discussed in Sec. 3.2, we conduct a specific experiment with selecting different types of images to examine the attack performance. We selected two scenarios, Hate Speech and Fraud, and conducted experiments on the MiniGPT-4 [40] and LLaVA-2 [19] models, respectively. Moreover, the Irrelevant Image used in the experiments are fixed and consist of 12 images that is not associated with any of the jailbreak scenarios. As shown in Fig. 6, when blank images" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 434, + 159 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 434, + 159 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 434, + 159 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 434, + 159 + ], + "type": "image", + "image_path": "e1c07cdd7ffc11e1e5ea53691b5a8cc2e25fb9217723e3932ea2b2d902f30f7c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 161, + 404, + 170 + ], + "lines": [ + { + "bbox": [ + 345, + 161, + 404, + 170 + ], + "spans": [ + { + "bbox": [ + 345, + 161, + 404, + 170 + ], + "type": "text", + "content": "(a) ASR of Fraud." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 70, + 554, + 159 + ], + "blocks": [ + { + "bbox": [ + 437, + 70, + 554, + 159 + ], + "lines": [ + { + "bbox": [ + 437, + 70, + 554, + 159 + ], + "spans": [ + { + "bbox": [ + 437, + 70, + 554, + 159 + ], + "type": "image", + "image_path": "6c05aa229a841b7e3777d92510e539a571923261bc60e87416ce9e754ca45af9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 161, + 535, + 171 + ], + "lines": [ + { + "bbox": [ + 454, + 161, + 535, + 171 + ], + "spans": [ + { + "bbox": [ + 454, + 161, + 535, + 171 + ], + "type": "text", + "content": "(b) ASR of Hate Speech." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 350, + 271, + 521, + 403 + ], + "blocks": [ + { + "bbox": [ + 313, + 180, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 258 + ], + "type": "text", + "content": "Figure 6. The ASR of using different types of initial images. We can find that simply adding a blank image can lead to a substantial increase in the success rate. Compared to blank image, further selecting images that contain information but are unrelated to the jailbreak scenario results in a smaller improvement. However, when images are introduced that closely match the specific scenario, the performance reaches state-of-the-art levels." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 271, + 521, + 403 + ], + "lines": [ + { + "bbox": [ + 350, + 271, + 521, + 403 + ], + "spans": [ + { + "bbox": [ + 350, + 271, + 521, + 403 + ], + "type": "image", + "image_path": "ed976887c3dcd4706752a5ab8f67807ab290d7393a26b7f99f9156b01e1dce97.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "content": "Figure 7. The ASR of using different numbers of images. According to Eq. (5), since " + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "content": " is a constant value, we set " + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "inline_equation", + "content": "v = \\left\\lceil \\frac{k_{L}}{K} \\right\\rceil" + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "content": " and assume that the slopes on both sides are equal without loss of generality, where " + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "content": " represents the flatness of the minimum value, in order to better focus on the changing trends of ASR." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 496, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 640 + ], + "type": "text", + "content": "are introduced, the alignment of the underlying LLM fails to cover the unforeseen domains introduced by the visual modality. This misalignment effectively circumvents the inherent alignment mechanisms of the LLM within the VLM framework, leading to a marked increase in the ASR. However, when choosing unrelated images, for the target scene, it merely increases the information content and complexity of the image, which does not significantly aid in bypassing restrictions. Only by selecting images that match the scenario can meaningful semantic information be introduced, thereby amplifying the potential risk of the image and ultimately further increasing the ASR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Loss range determination. To further validate the impact of using images in different loss range on the improvement of jailbreak success rate, we conduct experiments on four scenarios on the MiniGPT-4 [40] model and the results are shown in Fig. 7. We observe that, overall, the ASR increases" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19896" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 70, + 259, + 208 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 259, + 208 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 259, + 208 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 259, + 208 + ], + "type": "image", + "image_path": "b27b1cce159879bbbae84bd425c0a879a97c55e623f840580405628c4b1ad60c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "text", + "content": "Figure 8. The average ASR of using different " + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "text", + "content": ". The average ASR increases gradually with the value of " + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "text", + "content": ", with the rate of increase remaining relatively stable. However, once " + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "text", + "content": " reaches a critical threshold (which is 5 in this experiment), the growth rate drops significantly, and the ASR quickly reaches its maximum value." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 293, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 532 + ], + "type": "text", + "content": "significantly when the number of Images initially increases. However, the growth patterns of the curves in different scenarios show considerable variation, with differences in both the rate of increase and the point at which the growth stops. Specifically, the values of " + }, + { + "bbox": [ + 56, + 293, + 296, + 532 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 56, + 293, + 296, + 532 + ], + "type": "text", + "content": " for Illegal Activity (IA) and Hate Speech (HS) are relatively large, indicating that they have relatively sharp minima, whereas Legal Opinion (LO) and Financial Advice (FA) exhibit flatter minima. Correspondingly, the curves for Illegal Activity (IA) and Hate Speech (HS) stop increasing early (almost cease to grow when the number of images reaches 5), while the curves for Legal Opinion (LO) and Financial Advice (FA) maintain a prolonged growth phase (only stopping near the point where the number of images reaches 9). The results demonstrate that to achieve an optimal ASR, image sets with flat minima require fewer images, while image sets with sharp minima incur higher costs to achieve the desired outcome. These findings prove that flatter minima indeed provide greater robustness, allowing models to perform better when the test set deviates from the training set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": "In addition, we also examined the values of " + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": ". We tested the average ASR across all scenarios for different values of " + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": " on both MiniGPT-4 [40] and LLaVA-2 [19]. As shown in Fig. 8, the average ASR reaches its near-maximum value when " + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\rho = 6" + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": ", while the computational complexity increases at a linear rate. Further increases in " + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": " yield diminishing returns. Therefore, based on empirical observations, we set the baseline value of " + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\rho = 6" + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": ". Through the above experiments, we can prove the effectiveness of the flat minimum, the optimal choice of " + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": ", and we finally find the balance point between the attack success rate and the computational cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 671, + 157, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 157, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 157, + 684 + ], + "type": "text", + "content": "4.4. Defense Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "From the perspective of developmental trends, defense mechanisms always occupy the weaker position. With the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 433, + 164 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 433, + 164 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 433, + 164 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 433, + 164 + ], + "type": "image", + "image_path": "af9cfd232eaf701e8b6d9df2d60aea6839da0e0fa17f18bcbbe6858c000b5683.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 166, + 433, + 185 + ], + "lines": [ + { + "bbox": [ + 315, + 166, + 433, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 166, + 433, + 185 + ], + "type": "text", + "content": "(a) Results of 10 specific questions in fraud scenario." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 195, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 314, + 195, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 195, + 553, + 206 + ], + "type": "text", + "content": "Figure 9. The demonstration of our defense motivation and results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 436, + 70, + 554, + 164 + ], + "blocks": [ + { + "bbox": [ + 436, + 70, + 554, + 164 + ], + "lines": [ + { + "bbox": [ + 436, + 70, + 554, + 164 + ], + "spans": [ + { + "bbox": [ + 436, + 70, + 554, + 164 + ], + "type": "image", + "image_path": "eb1fa30aeba1e9f044177f294863bd17d3e3373707d1b4d5b345d131164efce9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 166, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 435, + 166, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 435, + 166, + 553, + 185 + ], + "type": "text", + "content": "(b) ASR of original/defensive circumstance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 287 + ], + "type": "text", + "content": "emergence of an increasing number of VLMs and the continuous improvement of jailbreak methods, comprehensive defense becomes progressively more challenging. In this section, we analyze an effective defense strategy against Multi-Loss Adversarial Images attack." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 288, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 491 + ], + "type": "text", + "content": "As shown in Fig. 9a, we observe that under the same scenario, images with different loss not only vary in overall attack potency but also demonstrate significant differences in their effectiveness against individual queries. Based on this finding, we propose introducing a similarity detection module before image sets are input into VLMs. If a current image is detected to be highly similar to a previously encountered image, it is removed from the input set, thus preventing it from being fed into the VLMs. This approach transforms a Multi-Loss Adversarial Attack back into a Single-Loss Adversarial Attack, and due to the diversity of queries in the scenario, the average ASR achieves robust defense. We conducted experiments on MiniGPT-4 [40], and the results, as shown in Fig. 9b, indicate that our defense method effectively reduces the success rate of the Multi-Loss Adversarial Attack, achieving a " + }, + { + "bbox": [ + 313, + 288, + 555, + 491 + ], + "type": "inline_equation", + "content": "22.99\\%" + }, + { + "bbox": [ + 313, + 288, + 555, + 491 + ], + "type": "text", + "content": " reduction compared to the model without the defense module." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we conducted a systematic analysis of visual vulnerabilities in VLMs safety alignment and introduced MLAI, a novel jailbreak framework that effectively exploits these weaknesses. Our investigation revealed two critical findings: the unreliability of minimal-loss optimization in gradient-based attacks, and the crucial role of semantic alignment between visual content and target scenarios. Based on these insights, MLAI achieves superior performance through scenario-aware image generation, flat minima-based optimization, and multi-image collaborative attacks. Extensive experiments demonstrate MLAI's effectiveness, achieving attack success rates of " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "77.75\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " on MiniGPT-4 and " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "82.80\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " on LLaVA-2. The strong transferability to commercial black-box VLMs further highlights fundamental challenges in achieving robust safety alignment for multi-modal systems." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19897" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 114 + ], + "type": "text", + "content": "This research has been partially supported by Australian Research Council Discovery Project DP230101196." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 148 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 156, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 156, + 225, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 156, + 225, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 156, + 225, + 167 + ], + "type": "text", + "content": "[1] Sam Altman and et al. OpenAI, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 214, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 296, + 269 + ], + "type": "text", + "content": "[3] Nicholas Carlini, Milad Nasr, Christopher A. Choquette-Choo, Matthew Jagielski, Irena Gao, Anas Awadalla, Pang Wei Koh, Daphne Ippolito, Katherine Lee, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned?, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "text", + "content": "[4] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J. Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 316, + 195, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 195, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 195, + 327 + ], + "type": "text", + "content": "[5] Chatglm. Zhipu ai, 2023. 2, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 328, + 296, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 296, + 383 + ], + "type": "text", + "content": "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 384, + 208, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 208, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 208, + 395 + ], + "type": "text", + "content": "[7] Alibaba Cloud. Qwen, 2023. 2, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 396, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 296, + 440 + ], + "type": "text", + "content": "[8] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks?, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 441, + 296, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 496 + ], + "type": "text", + "content": "[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 498, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 295, + 542 + ], + "type": "text", + "content": "[10] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 544, + 187, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 187, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 187, + 554 + ], + "type": "text", + "content": "[11] Google. Gemini, 2024. 2, 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 555, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 577 + ], + "type": "text", + "content": "[12] Sepp Hochreiter and Jurgen Schmidhuber. Flat minima. Neural computation, 9:1-42, 1997. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "type": "text", + "content": "[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, and Madian Khabsa. Llama guard: Llm-based input-output safeguard for human-ai conversations, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[14] Jiaming Ji, Mickel Liu, Juntao Dai, Xuehai Pan, Chi Zhang, Ce Bian, Chi Zhang, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. arXiv preprint arXiv:2307.04657, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[15] Tomasz Korbak, Kejian Shi, Angelica Chen, Rasika Bhalerao, Christopher L. Buckley, Jason Phang, Samuel R." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Bowman, and Ethan Perez. Pretraining language models with human preferences, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "type": "text", + "content": "[17] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[18] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 206, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 228 + ], + "type": "text", + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 228, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 250 + ], + "type": "text", + "content": "[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "text", + "content": "[21] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "text", + "content": "[22] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 317, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 360 + ], + "type": "text", + "content": "[23] Yugeng Liu, Tianshuo Cong, Zhengyu Zhao, Michael Backes, Yun Shen, and Yang Zhang. Robustness over time: Understanding adversarial examples' effectiveness on longitudinal versions of large language models, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "text", + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "text", + "content": "[25] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 438, + 434, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 434, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 434, + 449 + ], + "type": "text", + "content": "[26] OpenAI. Dall·E 3, 2023. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 449, + 553, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 482 + ], + "type": "text", + "content": "[27] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, and et al. Florencia Leoni Aleman and. Gpt-4 technical report, 2024. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 483, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 559 + ], + "type": "text", + "content": "[28] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3419-3448, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[29] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "text", + "content": "[30] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to!, 2023. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "text", + "content": "[31] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models, 2023. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[32] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19898" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 520 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "type": "text", + "content": "[33] Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, Lewis Ho, Divya Siddarth, Shahar Avin, Will Hawkins, Been Kim, Jason Gabriel, Vijay Bolina, Jack Clark, Yoshua Bengio, Paul Christiano, and Allan Dafoe. Model evaluation for extreme risks, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 150, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 184 + ], + "type": "text", + "content": "[34] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[35] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "text", + "content": "[36] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "text", + "content": "[37] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 385 + ], + "type": "text", + "content": "[38] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "text", + "content": "[39] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "text", + "content": "[40] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 508 + ], + "type": "text", + "content": "[41] Andy Zou, Zifan Wang, J. Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 220, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 220, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 220, + 520 + ], + "type": "text", + "content": "[42] Mark Elliot Zuckerberg. Meta, 2004. 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19899" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_content_list.json b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..652f45ade2734129b6f2c4afa58ad98242ce6e55 --- /dev/null +++ b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_content_list.json @@ -0,0 +1,2133 @@ +[ + { + "type": "text", + "text": "Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis", + "text_level": 1, + "bbox": [ + 153, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bingda Tang $^{1}$ Boyang Zheng $^{1}$ Sayak Paul $^{2}$ Saining Xie $^{1}$", + "bbox": [ + 235, + 202, + 759, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 New York University 2 Hugging Face", + "bbox": [ + 333, + 227, + 663, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 280, + 326, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper does not describe a new method; instead, it provides a thorough exploration of an important yet understudied design space related to recent advances in text-to-image synthesis—specifically, the deep fusion of large language models (LLMs) with diffusion transformers (DiTs) for multimodal generation. Previous studies mainly focused on overall system performance rather than detailed comparisons with alternative methods, and key design details and training recipes were often left undisclosed. These gaps create uncertainty about the real potential of this approach. To fill these gaps, we conduct an empirical study on text-to-image generation, performing controlled comparisons with established baselines, analyzing important design choices, and providing a clear, reproducible recipe for training at scale. We hope this work offers meaningful data points and practical guidelines for future research in multimodal generation.", + "bbox": [ + 89, + 313, + 483, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 585, + 220, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models have made remarkable progress in generating high-quality images from descriptive texts. Current state-of-the-art systems [2, 5, 10, 17, 30] typically derive text representations from specialized encoders, such as CLIP [31] and T5 [32]. With the rise of decoder-only large language models (LLMs), there has been a growing amount of interest in their potential as replacements for these traditional text encoders [7, 15, 21, 22, 25, 28, 48, 51]. However, simply substituting LLMs has not yielded expected performance gains unless coupled with sophisticated architectural adaptations [15, 25, 47]. Prior work [25] attributes this to the misalignment between the next token prediction training objective of LLMs and the need for discriminative text representations in diffusion models.", + "bbox": [ + 89, + 612, + 483, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements [20, 26, 46, 50] have successfully unified auto-regressive decoding and denoising diffusion within a single transformer [43], enabling seamless multi-modal generation. This unified approach supports a wide range of tasks, including instructed image-to-text,", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2990b2bdacd0d68abdaa088ddd2b9dcb6414fa2edebd3e18de8f202c2d3f0c1d.jpg", + "image_caption": [ + "(a) Shallow-Fusion" + ], + "image_footnote": [], + "bbox": [ + 517, + 279, + 712, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/afc01f31583e83df2d8990efbdc392754d30cf73a6ce7972ba37023b9146e5f9.jpg", + "image_caption": [ + "(b) Deep-Fusion", + "Figure 1. Illustration of the deep fusion approach and baselines. We conduct controlled comparisons with baseline methods that incorporate text representations from a single text encoder layer into each DiT layer using late fusion within the attention mechanism, a strategy we term the shallow fusion approach." + ], + "image_footnote": [], + "bbox": [ + 718, + 279, + 895, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text-to-image synthesis, and interleaved image-text generation. While earlier methods often relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pre-training of LLMs by deeply fusing them with diffusion transformers (DiTs) [29] through layer-wise shared self-attention. This design facilitates rich cross-modal interactions while maintaining modality-specific computation by using distinct sets of weights. When optimized for text-to-image generation, it claims state-of-the-art performance [21].", + "bbox": [ + 511, + 489, + 906, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep fusion presents a compelling alternative to existing architectures for text-to-image synthesis, which typically conditions directly on representations from a single text encoder layer. By aligning diffusion models with the auto-regressive decoding nature of LLMs, deep fusion enables a more natural and tight-knit use of these models. However, despite the existing positive signals, its true potential remains uncertain. Current research [21, 38] prioritizes system-level benchmarks over controlled comparisons with established baselines, obscuring its position within the broader research landscape. More critically, the design space remains severely underexplored, and essential implementation details, such as training recipes, are often undisclosed. These limitations impede reproducibility and hinder broader adoption within the research community.", + "bbox": [ + 511, + 642, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we bridge these gaps through an empirical study on the deep fusion of a frozen LLM and a train", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28586", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "able DiT for text-to-image synthesis. We conduct controlled comparisons between the deep fusion approach and baselines, examine key design choices, and introduce a scalable, reproducible training recipe that delivers competitive performance on the established benchmarks for text-to-image generation. We believe that the evidence and unresolved questions highlighted in this study are of significant importance. We hope this work serves as a valuable resource, offering meaningful data points and practical guidelines to drive future advancements in multi-modal generation.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 253, + 232, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditioning mechanisms in diffusion models. Numerous studies have examined effective methods for integrating linguistic conditional information into diffusion models to facilitate text-to-image synthesis. Latent diffusion models (LDM) [35] pioneered the use of cross-attention between image and text features, a technique that has since become standard in U-Net [36] based architectures [18, 30].", + "bbox": [ + 89, + 279, + 482, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rise of DiTs [29], vision transformer [8] has emerged as the dominant architecture for diffusion models, prompting a reassessment of conditioning mechanisms. The vanilla DiT employs adaLN-Zero modulation to inject conditional information. However, this approach is limited to using pooled text representations, which capture only coarse-grained information. Subsequent architectures have explored cross-attention [5, 6, 47] and self-attention [10, 11, 17, 51] for text conditioning, which typically extract representations from a single text encoder layer, usually the last or penultimate one. While this strategy aligns well with CLIP [31] and T5 [32] text encoders, it is inherently mismatched with the next-token prediction training objective of LLMs, where the last layer focuses on next-token prediction instead of learning discriminative representation. In contrast, the deep fusion approach shows promise in harnessing the internal information flow of LLMs, aligning with their in-context self-attention mechanism for processing information.", + "bbox": [ + 89, + 385, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taming LLMs for diffusion models. The prevalence of decoder-only LLMs has driven extensive efforts to tame them for text-to-image diffusion models. The most effective and widely adopted practice involves leveraging LLMs to enrich input prompts [2], as well as employing multimodal LLMs (MLLMs) to generate synthetic captions for image data [2, 5, 10, 21, 47]. Another popular direction integrates (M)LLMs as system components, such as planners and discriminators [19, 44, 45, 49]. While these approaches have demonstrated effectiveness, they do not tap into the architecture of diffusion models.", + "bbox": [ + 89, + 689, + 482, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To integrate LLMs into diffusion models, previous works attempt to replace text encoders with LLMs, either by training from scratch [25, 47, 51] or by aligning feature", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/58bec034de58bec47fbb3d5554dd44df984efe3fe8d58554c69688dd07471973.jpg", + "image_caption": [ + "Figure 2. Illustration of the attention mask. Each dotted square indicates whether the row can attend to the column." + ], + "image_footnote": [], + "bbox": [ + 594, + 88, + 823, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "spaces [7, 15, 22, 28]. However, this substitution alone has not yielded the expected performance gains unless paired with sophisticated architectural adaptations [15, 25, 47].", + "bbox": [ + 511, + 310, + 903, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research [20, 26, 46, 50] has proven that autoregressive decoding and denoising diffusion can be effectively unified within a single transformer for multi-modal generation. While earlier methods relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pretraining of LLMs by deeply fusing them with DiTs. From a text-to-image perspective, this architecture introduces a novel way to tame LLMs for text-to-image diffusion models, with the potential for achieving state-of-the-art performance [21].", + "bbox": [ + 511, + 356, + 906, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Deep Fusion of LLMs and DiTs", + "text_level": 1, + "bbox": [ + 511, + 521, + 802, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 511, + 546, + 700, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the deep fusion approach, we integrate a frozen decoder-only LLM with a trainable DiT using layer-wise shared self-attention (Fig. 1). The DiT mirrors the LLM's transformer architecture, differing only in its input/output layers and timestep conditioning modules. This implements a two-stream transformer architecture that facilitates rich cross-modal interactions while maintaining modality-specific computation by utilizing distinct weight sets for processing tokens from different modalities.", + "bbox": [ + 511, + 568, + 905, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The fused model processes text embeddings through the LLM stream and noisy image latents through the DiT stream. At each layer's self-attention operation, we concatenate token sequences from both streams, enabling the DiT to extract conditional information from the linguistic context. To preserve the pretrained LLM's functionality, we apply a causal attention mask to the text sequence and a bidirectional mask to the image sequence, permitting the image tokens to attend to text tokens but not vice versa (Fig. 2). After the final layer, we discard text tokens and use only image tokens to predict velocity, as typically done when training rectified flow models [23].", + "bbox": [ + 511, + 704, + 905, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notably, only the key and value states of the text hidden", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "states are needed for self-attention. These remain constant throughout the diffusion process, allowing them to be efficiently cached and reused during inference.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Training Objective", + "text_level": 1, + "bbox": [ + 89, + 146, + 272, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt the rectified flow formulation [23] to learn transport maps between the standard Gaussian noise distribution $\\pi_0$ and the data distribution $\\pi_1$ , by connecting straight paths $x_{t} = tx_{1} + (1 - t)x_{0}$ between samples $x_0\\sim \\pi_0,x_1\\sim \\pi_1$ and learning an ODE model $dz_{t} = v_{\\theta}(z_{t},t)dt$ on time $t\\in [0,1]$ which converts $z_{0}$ from $\\pi_0$ to a $z_{1}$ following $\\pi_1$ .", + "bbox": [ + 89, + 169, + 482, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We fit the velocity $v$ with $x_{1} - x_{0}$ with a prescribed time distribution $\\pi_{t}$ by solving the following regression problem:", + "bbox": [ + 89, + 261, + 482, + 291 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {v} \\mathbb {E} _ {t \\sim \\pi_ {t}, x _ {0} \\sim \\pi_ {0}, x _ {1} \\sim \\pi_ {1}} [ \\| v (x _ {t}, t) - (x _ {1} - x _ {0}) \\| ^ {2} ] d t \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 311, + 482, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We parameterize $v$ with network $\\theta$ and solve Eq. (1) by stochastic optimization with empirical draws.", + "bbox": [ + 89, + 345, + 482, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 390, + 267, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure fully open and reproducible comparison between the deep and shallow fusion approaches, we provide comprehensive details on the experimental setup, including the model, dataset, training, inference, and evaluation. For the same purpose, we exclusively use open-sourced pre-trained LLMs and publicly available datasets.", + "bbox": [ + 89, + 416, + 483, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model. We employ a frozen Gemma 2B [27] as the base LLM for all experiments (excluding Sec. 7). We pair it with a randomly initialized 2.5B-parameter DiT. The transformer configurations of the DiT strictly follows the base LLM, including the hidden size, number of layers, number of attention heads, FFN design, and other architectural details, ensuring both models have an identical 2B-parameter backbone. Following the vanilla DiT setup [29], we use 2D frequency absolute positional encoding, AdaLN-Zero timesteps conditioning, ViT [8]-style weight initialization, and a patch size of 2. To further stabilize training, we apply QK normalization to all layers. For all experiments we adopt the same 16-channel VAE from Stable Diffusion 3 (SD 3) [10].", + "bbox": [ + 89, + 527, + 483, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset. We use the CC12M [4] dataset with community-sourced synthetic captions [9] as our training set for all experiments excluding Sec. 7. Our downloaded version of the dataset includes 10.9M image-caption pairs. The images are resized and center-cropped to $512 \\times 512$ and the texts are padded or truncated to 256 tokens.", + "bbox": [ + 89, + 744, + 482, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training. We train all models with a batch size of 512 using AdamW [24] optimizer $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ in BF16 mixed precision. We use a constant learning rate of", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$1 \\times 10^{-4}$ , a weight decay of $1 \\times 10^{-4}$ , and gradient clipping with a threshold of 1.0. Exponential moving average of the weights are gathered by a decay factor of 0.99 every 100 steps. We employ the same logit-normal distribution as used in SD 3 for timestep sampling. During training, $10\\%$ of the texts are randomly dropped to learn unconditional generation. Training is carried out using Google TPU v4-256 pods and FSDP implemented by PyTorch / XLA SPMD.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inference. We conduct inference using Euler discretization with 25 sampling steps, and a classifier-free guidance scale of 6 which we find to be near optimal for text-image alignment. We employ identical sampling steps and guidance scale across all experiments.", + "bbox": [ + 511, + 234, + 906, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation. We evaluate image-text alignment using GenEval [12] and DPG-Bench [15] metrics, prioritizing GenEval for its robustness. While both benchmarks provide valuable insights, DPG-Bench exhibits certain limitations, such as rapid performance saturation and potential measurement errors [21]. To ensure a comprehensive evaluation, we also provide visual quality measurements using FID [14] on MJHQ-30K [18]. Notably, image-text alignment does not always correlate positively with visual quality, often presenting trade-offs. Our sampling and evaluation are primarily carried out using NVIDIA L40S GPUs.", + "bbox": [ + 511, + 334, + 906, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5. Comparing Deep and Shallow Fusion", + "text_level": 1, + "bbox": [ + 511, + 516, + 849, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The deep fusion approach fuses an LLM and a DiT through layer-wise shared self-attention, creating interconnections throughout the network. However, established architectures typically condition on representations from a single text encoder layer. To investigate the true potential of deep fusion, we conduct controlled comparisons with baseline methods.", + "bbox": [ + 511, + 542, + 905, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a fair and meaningful comparison, we examine a common architectural paradigm which we refer to as shallow fusion. In this approach, representations from a single text encoder layer are integrated into each DiT layer through late fusion within the attention operation. Unlike deep fusion, which involves multi-layered interactions, shallow fusion maintains a fixed connection between each DiT layer and a prescribed text encoder layer (Fig. 1).", + "bbox": [ + 511, + 635, + 905, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1. Shallow Fusion Baselines", + "text_level": 1, + "bbox": [ + 511, + 767, + 743, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider two shallow fusion architectures that condition on last-layer hidden states of LLMs as our baselines. As illustrated in Fig. 3, the two architecture differ in how they aggregate information from the condition.", + "bbox": [ + 511, + 790, + 905, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Self-attention DiT. In this design, text representations are projected to key and value states and then concatenated with those of image hidden states in self-attention, which", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28588", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/26fb8174e61731ae5a33672dea3ec54381e8ac694b6cb96d9be6721b3e5532e4.jpg", + "image_caption": [ + "(a) Self-Attention" + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 313, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7d7637fd3db9bb8167e746807b8a70b2de942f3ba91772f80d5f7573c1b08c51.jpg", + "image_caption": [ + "(b) Cross-Attention", + "Figure 3. Illustration of cross-modal attention in the shallow fusion baselines. The key and query states of the condition are directly projected from text representations." + ], + "image_footnote": [], + "bbox": [ + 344, + 87, + 483, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can also be decoupled by running self-attention and cross-attention in parallel and merging their outputs. This approach resembles architectures proposed by [11, 51].", + "bbox": [ + 102, + 252, + 482, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Cross-attention DiT. This design also projects text representations to key and values states. However, unlike the previous approach, they are used for additional cross-attention with image hidden states, applied after the self-attention in each layer. This architecture follows the methodology employed in [5, 6, 47].", + "bbox": [ + 89, + 297, + 483, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to the deep fusion approach, both baseline models include RMS normalization and a linear layer for text representations before passing them through additional key and value projection layers. Additionally, the cross-attention DiT model uses extra query projections in its cross-attention mechanism for image hidden states. Other model configurations follow the DiT design in Sec. 4.", + "bbox": [ + 89, + 393, + 483, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, the deep fusion approach can be reinterpreted as a variant of the self-attention DiT architecture, as they both aggregate conditional information through in-context self-attention. The key difference lies in how the key and value states of the condition are derived: self-attention DiT employs a trainable projection to generate these states from a single text encoder layer, whereas deep fusion extracts them from corresponding LLM layers. Despite their similarities, the two approaches have fundamentally different conceptual implications: deep fusion treats the LLM and DiT as equal components of a unified model.", + "bbox": [ + 89, + 500, + 483, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from the aforementioned designs, SD3 [10] introduced an alternative fusion strategy that uses a two-stream transformer (MM-DiT) to jointly processes noised image latents and linguistic representations. While this is another popular approach worthy of investigation, an apples-to-apples fair comparison between the deep fusion approach and MM-DiT is not feasible, as both streams in MM-DiT are trainable. Consequently, our analysis focuses solely on the shallow fusion baselines detailed in this section.", + "bbox": [ + 89, + 667, + 483, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. Controlled Comparison", + "text_level": 1, + "bbox": [ + 89, + 816, + 308, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a fair comparison, we design self-attention DiT, cross-attention DiT, and the deep fusion model with similar architectures, and train them for 300K steps following the setup detailed in Sec. 4.", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4ad6daea267291d9d27033456f39ebb502b3aa1d89cba3a628c6a0fa2e9605c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 89, + 620, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69b9d2ebe405c405dc29059100fd208a75a1ddcdb93ead233685eee93f343571.jpg", + "image_caption": [ + "Figure 4. Illustration of timestep conditioning strategies. Removing timesetp conditioning leads to the fewest parameters and the best overall performance." + ], + "image_footnote": [], + "bbox": [ + 622, + 89, + 723, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17652829af896ca34a07b42ef4554cbbaf992c28c08a29ac8e203faf660a732e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 89, + 805, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/af6cc08e1cbca3253e4642d2b0f6589d81c71a372713008802f73974d59841ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 89, + 888, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Tab. 1, the deep fusion model achieve significantly better performance in image-text alignment than the self-attention DiT model and also surpass the cross-attention DiT model, while shallow fusion models demonstrate better visual quality. In terms of inference efficiency, deep fusion also demonstrates competitive performance, as shown in Tab. 2. This positive evidence underscores the compelling positioning of the deep fusion approach within the current landscape.", + "bbox": [ + 511, + 250, + 906, + 385 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/adc207133b2678079d562e2a718edb98fbd5a08d62396098972e2e3b91b64b03.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParams.GenEval ↑DPG ↑FID ↓
Self-Attention2.47B0.4273.926.16
Cross-Attention2.62B0.4976.324.00
Deep Fusion2.45B0.5176.627.33
", + "bbox": [ + 514, + 388, + 911, + 443 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/aa717091f0cf85075b07a7bf51e53deb9015074c37a34a64294643a3ac546fe0.jpg", + "table_caption": [ + "Table 1. Comparison of performance between deep and shallow fusion models. Deep fusion beats shallow fusion in text-image alignment while underperforms in visual quality." + ], + "table_footnote": [], + "table_body": "
MethodParams.Inference Latency (s)
Self-Attention2.47B1.75
Cross-Attention2.62B1.86
Deep Fusion2.45B1.66
", + "bbox": [ + 540, + 494, + 875, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Comparison of inference latency between deep and shallow fusion models. The numbers are measured with a batch size of 1 in automatic mixed precision on an NVIDIA A100 GPU.", + "bbox": [ + 511, + 550, + 906, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6. Examining Key Design Choices", + "text_level": 1, + "bbox": [ + 511, + 601, + 799, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we examine key design choices of the deep fusion approach through a text-to-image-centric lens. We begin by assessing the necessity and potential redundancy of parameters for timestep conditioning, determining whether to optimize their use or eliminate them altogether (Sec. 6.1). Next, we compare various positional encoding strategies (Sec. 6.2). Additionally, we investigate how the choice of base LLM and the use of instruction prompts impacts text-to-image performance (Sec. 6.3).", + "bbox": [ + 511, + 627, + 905, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For all experiments, we use the default design as the baseline and train the models for 300K steps, following the setup detailed in Sec. 4.", + "bbox": [ + 511, + 763, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6.1. Timestep Conditioning", + "text_level": 1, + "bbox": [ + 511, + 818, + 725, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DiT [29] has introduced AdaLN-Zero as the standard mechanism for injecting timestep and class label information. The AdaLN modules typically accounts for a large proportion of the model parameters, 0.5B out of a total 2.5B in our", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28589", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "case. However, since our text-to-image model does not use class labels, AdaLN serves only for timestep conditioning. This substantial parameter allocation raises a critical question about its necessity and potential redundancy: Could these parameters be utilized more effectively, or are they redundant altogether?", + "bbox": [ + 88, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question 1.1. Can adaLN parameters be utilized more effectively by integrating additional text conditioning?", + "bbox": [ + 86, + 181, + 480, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior research [10, 11, 17] has integrated text information into AdaLN-Zero by augmenting timestep embeddings with pooled text representations through summation. To optimize parameter efficiency, we follow this approach by leveraging embeddings1 from the CLIP L/14 text encoder [31], which provides high-quality linguistic representations trained on large-scale multi-modal data.", + "bbox": [ + 88, + 212, + 480, + 316 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7829ffd7d522fc4647aceeb2c92e2a416c56d7bb3a319366b6a91e3545c8df13.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGenEval ↑DPG ↑FID ↓
adaLN-Zero0.5176.627.33
+ CLIP L/140.5076.224.00
", + "bbox": [ + 122, + 321, + 452, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results indicate that integrating text modulation results in slight improvements in FID but weakens image-text alignment. Additionally, it further increases compute.", + "bbox": [ + 89, + 372, + 482, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question 1.2. Can we shrink the parameters for timestep conditioning?", + "bbox": [ + 89, + 417, + 482, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We explore the possibility of eliminating timestep conditioning parameters to develop a more streamlined architecture akin to that of the LLM. Specifically, we compare four timestep conditioning strategies from previous work, each with a progressively reduced number of parameters, as illustrated in Fig. 4.", + "bbox": [ + 89, + 448, + 482, + 537 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- adaLN-Zero. Following the vanilla DiT [29], we regress zero-initialized modulation parameters for each layer from the timestep embedding.", + "- adaLN-Single. Following PixArt- $\\alpha$ [6], we compute a global set of modulation parameters and refine them per layer by adding learnable embeddings.", + "- Addition. Following Transfusion [50], we directly add the timestep embedding to all image tokens.", + "- w/o Timestep. Inspired by a recent study [42], we completely remove the timestep conditioning from the model." + ], + "bbox": [ + 89, + 540, + 482, + 691 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/833c4512f9f0b66e26e14c83e2116e4685ebb968b0c628221f05232639ce28bc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParams.GenEval ↑DPG ↑FID ↓
adaLN-Zero2.47B0.5176.627.33
adaLN-Single2.01B0.4775.227.09
addition1.99B0.4775.626.40
w/o timestep1.98B0.4976.721.27
", + "bbox": [ + 91, + 700, + 486, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Surprisingly, the results in Tab. 3 indicate that reducing the number of parameters in timestep conditioning consistently enhances visual quality, whereas the performance in image-text alignment exhibits fluctuations.", + "bbox": [ + 89, + 806, + 480, + 867 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/28523ec2faf641ffb71f38442a8a4572a5538838f27bc6d135936db2f28921f5.jpg", + "table_caption": [ + "Table 3. Evaluation of timestep conditioning strategies. Removing timestep conditioning yields surprisingly strong results." + ], + "table_footnote": [], + "table_body": "
Input SequenceTXT 1TXT 2[PAD]IMG 1IMG 2IMG 3IMG 4
1D-RoPE + APE[0][1]—[2D Frequency APE]
1D-RoPE[0][1]—[2][3][4][5]
1D + 2D-RoPE[0][1]—[2, 2][2, 3][3, 2][3, 3]
M-RoPE[0, 0][1, 1]—[2, 2][2, 3][3, 2][3, 3]
", + "bbox": [ + 521, + 89, + 898, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, the strategy that completely removes timestep conditioning not only achieves significantly better FID but also maintains comparable GenEval and DPG-Bench performance. This finding is consistent with [42], where timestep conditioning removal improved FID in rectified flow models trained on smaller datasets. Furthermore, the complete removal of timestep conditioning eliminates the need for associated parameters, resulting in a $20\\%$ reduction in the total number of model parameters. Although this approach slightly lags behind AdaLN-Zero in terms of text-image alignment metrics, we prefer it due to its parameter efficiency and architectural simplicity.", + "bbox": [ + 509, + 194, + 903, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.2. Positional Encoding", + "text_level": 1, + "bbox": [ + 511, + 383, + 702, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Absolute positional encoding (APE) is widely used in text-to-image diffusion models, whereas in the context of LLMs, rotary positional embedding (RoPE) [40] is the predominant choice. As deep fusion models are inherently multimodal and differ from traditional text-to-image diffusion models, it is unclear which positional encoding (or their combinations) is best suited for mixed-modal sequences.", + "bbox": [ + 511, + 405, + 903, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question 2.1. Is RoPE more advantageous than APE for enhancing the performance of deep fusion models?", + "bbox": [ + 511, + 512, + 903, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To explore this, we experiment with three positional encoding methods, as illustrated in Fig. 5:", + "bbox": [ + 511, + 542, + 903, + 571 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 1D RoPE + APE: We apply 1D RoPE to the text sequence and APE to the image sequence respectively.", + "- 1D RoPE: We extend 1D RoPE to encompass both text and image sequences.", + "- 1D + 2D RoPE: We apply 1D RoPE to the text sequence and 2D RoPE to the image sequence respectively." + ], + "bbox": [ + 511, + 573, + 903, + 664 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d3a9e8e522d2797c0f736845a817322975c3705e0e78d1789cdb2210df72a716.jpg", + "table_caption": [ + "Figure 5. Illustration of RoPE. The indices denote position IDs." + ], + "table_footnote": [], + "table_body": "
MethodGenEval ↑DPG ↑FID ↓
1D-RoPE + APE0.5176.627.33
1D-RoPE0.4677.027.94
1D + 2D-RoPE0.5176.425.42
", + "bbox": [ + 542, + 672, + 870, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4. Comparing different positional encoding strategies. $1\\mathrm{D} + 2\\mathrm{D}$ RoPE achieves the best overall performance.", + "bbox": [ + 511, + 727, + 903, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Tab. 4, the 1D + 2D-RoPE configuration achieves the best overall performance, with only a marginal decrease in DPGBench compared to the 1D-RoPE + APE variant. The superiority of 2D-RoPE over APE suggests RoPE is more effective for modeling image sequences in deep fusion models. Using only 1D-RoPE slightly reduces performance, indicating that while deep fusion models treat text and image sequences as a unified input, their distinct positional characteristics are best modeled separately.", + "bbox": [ + 509, + 763, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1We apply RMS-normalization and an MLP to the text embeddings before adding them to the timestep embeddings.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28590", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question 2.2. Does deep fusion models benefit from RoPE specifically designed for mixed-modal sequences?", + "text_level": 1, + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous work on MLLMs has explored RoPE strategies for handling mixed-modal sequences. Naturally, we are curious whether deep fusion models can benefit from these positional encodings. Follow Qwen2-VL, we implement M-RoPE, a variant of RoPE that applies 2D positional IDs to chunked 1D RoPE frequencies, allowing it to function as 1D RoPE for text sequences while approximating 2D RoPE for image sequences.", + "bbox": [ + 89, + 121, + 483, + 242 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0f65c5adf3b3fae989aa7f6ce5436297adb20801b06a9469b0c0be4bfca3a9ce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGenEval ↑DPG ↑FID ↓
1D + 2D-RoPE0.5176.425.42
M-RoPE0.4974.927.60
", + "bbox": [ + 119, + 251, + 452, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although M-RoPE elegantly unifies 1D and 2D-RoPEs, it still falls short compared to their direct combination. This underscores the challenge of designing position encodings for mixed-modal sequences.", + "bbox": [ + 89, + 301, + 483, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Base LLM", + "text_level": 1, + "bbox": [ + 89, + 372, + 209, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LLMs trained with different paradigms and data demonstrate diverse capabilities and behaviors. In this section, we examine how the choice of base LLM impacts the performance of deep fusion models.", + "bbox": [ + 89, + 395, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Question 3.1 Can instruction tuning, combined with instruction prompts, improve text-to-image performance?", + "bbox": [ + 89, + 455, + 482, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction tuning enables LLMs to effectively follow complex instructions. We explore whether this process can also enhance their internal information flow, leading to more contextualized and discriminative representations for text-to-image synthesis. We compare Gemma 2B with Gemma 2B IT, its instruction-tuned variant. We also experiment with using it with a simple instruction prompt, \"Imagine: \", which we find sufficient for guiding the LLM to generate detailed and relevant expansions of the input.", + "bbox": [ + 89, + 486, + 483, + 622 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d6c999776831b8f9e3600963c7d14cdf9d4aeedbfe4c7638490aaa0b55223f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ instruction tuning0.4975.427.04
+ instruction prompt0.5075.825.28
", + "bbox": [ + 104, + 628, + 467, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 5, instruction tuning appears to have a slightly negative impact on performance. While the use of an instruction prompt mitigates this effect to some extent, consistent with findings in [25, 47], it still falls short of the baseline. This result highlights a challenge in effectively leveraging the instruction-following capabilities of LLMs.", + "bbox": [ + 89, + 719, + 482, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Question 3.2 Can multi-modal tuning improve text-to-image performance?", + "bbox": [ + 89, + 810, + 482, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we're also interested in the effect of multi-modal tuning. While multi-modal tuning differs significantly from our setup, its shift in data distribution may still potentially enhance adaptability to multi-modal tasks. We", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compare Gemma 2B with the base LLM of PaliGemma 3B PT [3], a multi-modal extension of Gemma 2B that undergoes additional pretraining on image-text data.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/840ea00435f9a93b36946d8deec1f074d77584864d0f7fa02323ee0be338b705.jpg", + "table_caption": [ + "Table 5. Evaluating the effect of instruction tuning. Using instruction-tuned LLMs does not improve performance." + ], + "table_footnote": [], + "table_body": "
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ multi-modal tuning0.5276.226.30
", + "bbox": [ + 527, + 138, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which yields small improvements in performance. This observation indicates that multi-modal finetuning provides some benefit to the deep-fusion model.", + "bbox": [ + 511, + 186, + 903, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Question 3.3 Does improved LLM capabilities translate to stronger text-to-image performance?", + "bbox": [ + 511, + 231, + 903, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, as the base LLM become more proficient in understanding and generating text, they could potentially foster synergistic improvements in DiT performance. We compare Gemma 2B with Gemma 2 2B [34], the next generation of Gemma 2B which demonstrates a $6\\%$ absolute performance improvement $(0.44 \\rightarrow 0.50)$ on an average of 8 language-only benchmarks. Notably, Gemma 2 2B features a different transformer architecture than Gemma 2B, prompting us to adjust the DiT architecture accordingly. This modification increases the number of adaLN parameters by 0.3B. However, as demonstrated in Sec. 6.1, these parameters do not contribute to model performance.", + "bbox": [ + 511, + 261, + 905, + 441 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2e7a5a0a021a348e9fde7d85ccc5db2174f94e3a0e49d14788f2b0447a7040f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
Gemma 2 2B0.5479.123.94
", + "bbox": [ + 544, + 446, + 875, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Upgrading from Gemma 2B to Gemma 2 2B yields a drastic performance boost. This finding suggests that the DiT's performance in deep fusion models is strongly dependent on the capabilities of the underlying base LLM.", + "bbox": [ + 511, + 493, + 905, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7. Training at Scale", + "text_level": 1, + "bbox": [ + 511, + 566, + 681, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present a final recipe for the deep fusion model, building on the original framework while incorporating key insights from previous exploration. We conduct large-scale training to benchmark our model against established systems, showcasing its scalability and competitive performance on the leaderboard.", + "bbox": [ + 511, + 592, + 905, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7.1. Final Recipe", + "text_level": 1, + "bbox": [ + 511, + 691, + 645, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on the insights from Sec. 6, we introduce the following design modifications to our model:", + "bbox": [ + 511, + 714, + 903, + 744 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Remove AdaLN-Zero modules.", + "- Replace 1D-RoPE + APE with 1D + 2D-RoPE.", + "- Replace Gemma 2B with Gemma 2 2B, adjusting the transformer configurations of the DiT accordingly." + ], + "bbox": [ + 511, + 747, + 903, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model for 800K steps on a mixed dataset comprising CC12M [4], SA-1B [16], and the training subset of JourneyDB [41], amounting to approximately 26M image-caption pairs. Notably, state-of-the-art text-to-image models typically rely on high-quality datasets of much larger scale to achieve superior performance. For CC12M", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/151612be0ce56445f20a9dcae542bed4e3577b9ed0541072b02516f8e5f18734.jpg", + "image_caption": [ + "(a) Fully Parallel", + "Figure 6. Illustration of architecture alignment. Dashed boxes indicate parameters that have been reduced by decreasing either the hidden size or the number of layers." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 220, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfae5e7ebb34949aa90ffbd1cdfa03240bc612a3d957b161f39361aadc715f26.jpg", + "image_caption": [ + "(b) Skipped Layers" + ], + "image_footnote": [], + "bbox": [ + 225, + 90, + 349, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea62c125e010eff6f16c791f75ab09bd4575bff41bfa09b4f2fcfe0c71831853.jpg", + "image_caption": [ + "(c) Decoupled Dimension" + ], + "image_footnote": [], + "bbox": [ + 354, + 90, + 478, + 178 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and SA-1B, we utilize synthetic captions [6, 9]. Other experimental setup follows Sec. 4.", + "bbox": [ + 89, + 253, + 482, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 89, + 295, + 326, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ff2663f46bf36b2ea20940472ba63a5ebf29e1283d28c043b5394b203dda6b37.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParamsDataGen.↑DPG↑FID↓
SD 1.5 [35]0.9B4.8B0.4363.2—
DALL-E 2 [33]4.2B2.6B0.52——
SDXL [30]2.6B1.6B0.5574.76.63
PG 2.5 [18]2.6B—0.5675.56.09
SD 3 M [10]2B1B0.6284.111.92
DALL-E 3 [2]——0.6783.5—
FLUX.1 [dev] [17]12B—0.6784.010.15
PG 3 [21]24B—0.7687.0—
MicroDiT [37]1.2B37M0.46——
PixArt-α [6]0.6B25M0.4871.1—
Lumina Next [11]2B—0.4674.67.58
PixArt-Σ [5]0.6B46M0.5480.56.15
Transfusion [50]7.3B3.5B0.63——
Sana 1.0 1.6B [47]1.6B—0.6684.85.76
Ours2B26M0.6081.67.54
", + "bbox": [ + 91, + 321, + 483, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our model with the most advanced text-to-image diffusion models in Tab. 6. Despite being trained with limited compute and data in a simplified setting, our model surpasses many industry-standard systems and delivers competitive results.", + "bbox": [ + 89, + 614, + 482, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present qualitative examples from our model in Figure 7. Our model demonstrates the ability to generate high-quality images with superior prompt alignment.", + "bbox": [ + 89, + 691, + 483, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8. Further Exploration", + "text_level": 1, + "bbox": [ + 89, + 751, + 287, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present preliminary studies exploring more aggressive modifications to the deep fusion approach. For all experiments, we train the models for 300K steps, following the setup detailed in Sec. 4.", + "bbox": [ + 89, + 777, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.1. Architecture Alignment", + "text_level": 1, + "bbox": [ + 89, + 848, + 310, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Up to this point, our approach has followed prior work by aligning the LLM and DiT backbones in a layer-by-layer", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "approach, strictly enforcing identical transformer configurations for both models. However, this rigid constraint limits the flexibility of deep fusion. In practice, we need the ability to scale the LLM and DiT independently, as different modalities follow distinct scaling laws and network design principles, and training and deployment scenarios vary.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this, we explore modifying the DiT model's hidden size and number of transformer layers, as illustrated in Fig. 6. The adapted model is fused into the middle layers of the LLM, which contain richer semantic information [39]. Additionally, in self-attention, hidden states are still projected to query, key, and value states that match the LLM's dimensionality, ensuring compatibility. Since the most successful DiTs are generally much smaller than state-of-the-art LLMs, we focus on shrinking the size of our DiT.", + "bbox": [ + 511, + 184, + 906, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/980420d20b71dc9966e04cc750c07050f71d644d18e5ecf1ab5aa1deea275892.jpg", + "table_caption": [ + "Table 6. Comparison of performance with state-of-the-art systems. Table adapted from [48, 50]. Industrial baselines are presented at the top, while academic baselines and our model are listed at the bottom." + ], + "table_footnote": [], + "table_body": "
Hidden sizeParams.GenEval ↑DPG ↑FID ↓
20482.5B0.5176.627.33
17922.1B0.5077.124.27
15361.8B0.4976.225.46
12801.4B0.4874.824.64
", + "bbox": [ + 524, + 330, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/39991e9a795bd1cb0b4a8b28a1567af9a327dbf7f53cdce514980d8a322521b1.jpg", + "table_caption": [ + "Table 7. Evaluating models of different hidden sizes. The default hidden size is 2048." + ], + "table_footnote": [], + "table_body": "
LayersParams.GenEval ↑DPG ↑FID ↓
182.5B0.5176.627.33
141.9B0.4774.623.46
101.4B0.3368.028.34
", + "bbox": [ + 524, + 443, + 890, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Evaluating different numbers of layers. The default number of layer is 18.", + "bbox": [ + 511, + 498, + 903, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table Tab. 7, the model's performance degrades gracefully as we reduce the hidden size, with visual quality actually improving in some cases. While decreasing the number of transformer layers (Tab. 8) also yields acceptable results, performance deteriorates more quickly. We hypothesize this occurs because Gemma 2B already employs fewer layers than typical model architectures. These findings suggest that LLM and DiT model designs can be effectively decoupled, enabling the application of separate scaling laws and design principles for each component.", + "bbox": [ + 511, + 541, + 906, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.2. Attention Mechanism", + "text_level": 1, + "bbox": [ + 511, + 709, + 718, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Sec. 3, we build on prior work by defining the deep fusion architecture with shared self-attention to bridge the LLM and DiT. Inspired by extensive research on cross-attention in MLLMs [1, 13] and our findings in Sec. 5, which highlight the superior performance of cross-attention DiT over self-attention DiT, we explore an alternative deep fusion variant. This new approach replaces shared self-attention with cross-attention mechanisms, similar to cross-attention DiT but with a key distinction: we substitute the projected linguistic key and value states in traditional cross-attention DiT with corresponding states from LLM layers.", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28592", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f046d53be3e87e89c215ac852dffa20cc94ed407976b407fe95dd2b052df5582.jpg", + "image_caption": [ + "Dew on blue rose petals, HD, close up, detail" + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 297, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/273f3bb4e7aee303f6939030bf3ed037185f825f6eb8f3ef41a7678d6f8dc403.jpg", + "image_caption": [ + "A corgi wearing sunglasses walks on the beach of a tropical island" + ], + "image_footnote": [], + "bbox": [ + 300, + 90, + 496, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfe7a970366132155da3fd780030eca52ffc88ae64bfd7a3263d04105a80115f.jpg", + "image_caption": [ + "A rally car taking a fast turn on a track" + ], + "image_footnote": [], + "bbox": [ + 496, + 90, + 697, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd47c8b9bc51ebe7a21cefd83c4e74962037de8b6b2fa7546420c68462475560.jpg", + "image_caption": [ + "An armchair in the shape of an avocado" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 898, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28e5c9516d6dd5f4a1afd11cbf5efd2a7820a07cfde5c41cdd31dc20b8c1f789.jpg", + "image_caption": [ + "A man riding a horse through the Gobi Desert with a beautiful sunset behind him, movie quality." + ], + "image_footnote": [], + "bbox": [ + 98, + 273, + 297, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e8e8b5e639fad0c7071a397ea7b1941c2ce13106225d94f9562a34e10203f4e7.jpg", + "image_caption": [ + "Pirate ship trapped in a cosmic maelstrom nebula" + ], + "image_footnote": [], + "bbox": [ + 297, + 273, + 496, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18705d9ef123edb53eb534838fe38b5ac85508776d8fb05d7cdf78373ab50ddd.jpg", + "image_caption": [ + "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" + ], + "image_footnote": [], + "bbox": [ + 496, + 273, + 697, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bac32400729d933a1bf4611893eda1268d340793d61fcf4b90e94ac8cc7e1bb.jpg", + "image_caption": [ + "A person surfing on a wave of stars in outer space." + ], + "image_footnote": [], + "bbox": [ + 697, + 273, + 898, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de8a836baaf5eedb97e37dbc2a581bd03d26316c737bc1af27ab0268806ff3f7.jpg", + "image_caption": [ + "a picturesque autumn scene where a quaint cottage with a thatched roof sits beside a tranquil lake, surrounded by trees with leaves in vibrant shades of orange, red, and yellow. The cottage's wooden exterior is complemented by white-framed windows, and a stone chimney rises above the roofline. The lake reflects the warm fall colors, creating a mirror image of the foliage and the small structure on its calm surface." + ], + "image_footnote": [], + "bbox": [ + 98, + 465, + 299, + 621 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e1fb7678a63b70f38190cabfd67b47f9d74ede94322dab8314a643b32a52fe7.jpg", + "image_caption": [ + "A portrait of a human growing colorful flowers from her hair. Hyperrealistic oil painting. Intricate details." + ], + "image_footnote": [], + "bbox": [ + 300, + 465, + 496, + 621 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/084ba814fbc087a2d54f5940196c570a463401e793865aa3382705d6b3d6fd4b.jpg", + "image_caption": [ + "a black and white picture of a woman looking through the window, in the style of Duffy Sheridan, Anna Razumovskaya, smooth and shiny, wavy, Patrick Demarchelier, album covers, lush and detailed", + "13 year old farm boy, red tousled hair, green eyes, soft pale skin, freckles, stub nose, very beautiful, very pretty, dark fantastic medieval" + ], + "image_footnote": [], + "bbox": [ + 496, + 465, + 696, + 622 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db084e5ab4541cd00906c718f8d7893db1d011a0033645b3e5a12b5b0d59d5ef.jpg", + "image_caption": [ + "Figure 7. Samples generated by our final recipe." + ], + "image_footnote": [], + "bbox": [ + 694, + 465, + 898, + 621 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a3328bff66a73947f8242944fced02bed6e2ab1501e75cd647e80b4c97fbb44.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGenEval ↑DPG ↑FID ↓
self-attention0.5176.627.33
cross-attention0.5276.526.57
", + "bbox": [ + 122, + 719, + 450, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This modification yields minor gains, though at a cost to the LLM-DiT parity. Additionally, we find that although cross-attention introduces a negligible increase in FLOPs and parameter count, it leads to approximately a $12\\%$ increase in latency $^{2}$ (1.66s vs. 1.86s). Therefore, we retained the self-attention design in our final configuration.", + "bbox": [ + 89, + 771, + 480, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 720, + 632, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have studied the recently popular deep fusion of a frozen LLM with a trainable DiT for text-to-image synthesis. Our findings provide empirical evidence supporting its advantages over baselines. We highlight key design choices, identify unresolved problems, and offer meaningful data points alongside practical guidelines. We hope our empirical work help advance multi-modal generation and bridge the gap between auto-regressive decoding and denoising diffusion.", + "bbox": [ + 511, + 746, + 906, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2Latency is measured with a batch size of 1 under automatic mixed precision on an NVIDIA A100 GPU", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank the reviewers for their constructive feedback. We also thank Shusheng Yang, Shengbang Tong, Wenhao Chai, Nanye Ma, Sihan Xu, and Chenyu Li for insightful discussions. This work was mainly supported by the Google TPU Research Cloud (TRC) program, and the Google Cloud Research Credits program (GCP19980904) and Open Path AI Foundation. SX also acknowledges support from Intel AI SRS, IITP grant funded by the Korean Government (MSIT) (RS-2024-00457882, National AI Research Lab Project), Amazon Research Award, and NSF Award IIS-2443404.", + "bbox": [ + 89, + 113, + 483, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 279, + 187, + 294 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 7", + "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.opencai.com/papers/dall-e-3.pdf, 2023.1,2,7", + "[3] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv:2407.07726, 2024. 6", + "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 3, 6", + "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024.1,2,4,7", + "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\alpha$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 7", + "[7] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[9] Caption Emporium. conceptual-captions-cc12m-llavanext. https://huggingface.co/datasets/CaptionEmporium / conceptual - captions - cc12m-llavanext, 2024.3,7" + ], + "bbox": [ + 99, + 303, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 2, 3, 4, 5, 7", + "[11] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv:2405.05945, 2024. 2, 4, 5, 7", + "[12] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. arXiv:2310.11513, 2023. 3", + "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 7", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 3", + "[15] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv:2403.05135, 2024. 1, 2, 3", + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6", + "[17] Black Forest Labs. Announcing black forest labs. https://blackforestlabs.ai/announcing-black-forest-labs/, 2024. 1, 2, 5, 7", + "[18] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv:2402.17245, 2024. 2, 3, 7", + "[19] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv:2305.13655, 2023. 2", + "[20] Weixin Liang, Lili Yu, Liang Luo, Srinivasan Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, et al. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. arXiv:2411.04996, 2024. 1, 2", + "[21] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv:2409.10695, 2024. 1, 2, 3, 7", + "[22] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv:2407.00737, 2024. 1, 2", + "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv:2209.03003, 2022. 2, 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28594", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 3", + "[25] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv:2406.11831, 2024. 1, 2, 6", + "[26] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv:2411.07975, 2024. 1, 2", + "[27] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv:2403.08295, 2024. 3", + "[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 1, 2", + "[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1, 2, 3, 4, 5", + "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1, 2, 7", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5", + "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In JMLR, 2020. 1, 2", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 7", + "[34] Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv:2408.00118, 2024. 6", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 7", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2", + "[37] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv:2407.15811, 2024. 7", + "[38] Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, and Lili Yu. Lmfusion:" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Adapting pretrained language models for multimodal generation. arXiv:2412.15188, 2024. 1, 2", + "[39] Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models. arXiv:2502.02013, 2025. 7", + "[40] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 5", + "[41] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. In NeurIPS, 2024. 6", + "[42] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv:2502.13129, 2025. 5", + "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1", + "[44] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv:2401.15688, 2024. 2", + "[45] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2", + "[46] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv:2409.11340, 2024. 1, 2", + "[47] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv:2410.10629, 2024. 1, 2, 4, 6, 7", + "[48] Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv:2501.18427, 2025. 1, 7", + "[49] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2", + "[50] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv:2408.11039, 2024. 1, 2, 5, 7", + "[51] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv:2406.18583, 2024. 1, 2, 4" + ], + "bbox": [ + 516, + 92, + 903, + 827 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28595", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_model.json b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22773a3418b57c0163ddd2ac294fdfcbc73e88a8 --- /dev/null +++ b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_model.json @@ -0,0 +1,2926 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.203, + 0.761, + 0.223 + ], + "angle": 0, + "content": "Bingda Tang \\(^{1}\\) Boyang Zheng \\(^{1}\\) Sayak Paul \\(^{2}\\) Saining Xie \\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.228, + 0.664, + 0.248 + ], + "angle": 0, + "content": "1 New York University 2 Hugging Face" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.281, + 0.327, + 0.297 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.485, + 0.556 + ], + "angle": 0, + "content": "This paper does not describe a new method; instead, it provides a thorough exploration of an important yet understudied design space related to recent advances in text-to-image synthesis—specifically, the deep fusion of large language models (LLMs) with diffusion transformers (DiTs) for multimodal generation. Previous studies mainly focused on overall system performance rather than detailed comparisons with alternative methods, and key design details and training recipes were often left undisclosed. These gaps create uncertainty about the real potential of this approach. To fill these gaps, we conduct an empirical study on text-to-image generation, performing controlled comparisons with established baselines, analyzing important design choices, and providing a clear, reproducible recipe for training at scale. We hope this work offers meaningful data points and practical guidelines for future research in multimodal generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.222, + 0.603 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Text-to-image diffusion models have made remarkable progress in generating high-quality images from descriptive texts. Current state-of-the-art systems [2, 5, 10, 17, 30] typically derive text representations from specialized encoders, such as CLIP [31] and T5 [32]. With the rise of decoder-only large language models (LLMs), there has been a growing amount of interest in their potential as replacements for these traditional text encoders [7, 15, 21, 22, 25, 28, 48, 51]. However, simply substituting LLMs has not yielded expected performance gains unless coupled with sophisticated architectural adaptations [15, 25, 47]. Prior work [25] attributes this to the misalignment between the next token prediction training objective of LLMs and the need for discriminative text representations in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Recent advancements [20, 26, 46, 50] have successfully unified auto-regressive decoding and denoising diffusion within a single transformer [43], enabling seamless multi-modal generation. This unified approach supports a wide range of tasks, including instructed image-to-text," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.28, + 0.713, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.388, + 0.659, + 0.397 + ], + "angle": 0, + "content": "(a) Shallow-Fusion" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.28, + 0.897, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.388, + 0.846, + 0.398 + ], + "angle": 0, + "content": "(b) Deep-Fusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 1. Illustration of the deep fusion approach and baselines. We conduct controlled comparisons with baseline methods that incorporate text representations from a single text encoder layer into each DiT layer using late fusion within the attention mechanism, a strategy we term the shallow fusion approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.642 + ], + "angle": 0, + "content": "text-to-image synthesis, and interleaved image-text generation. While earlier methods often relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pre-training of LLMs by deeply fusing them with diffusion transformers (DiTs) [29] through layer-wise shared self-attention. This design facilitates rich cross-modal interactions while maintaining modality-specific computation by using distinct sets of weights. When optimized for text-to-image generation, it claims state-of-the-art performance [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Deep fusion presents a compelling alternative to existing architectures for text-to-image synthesis, which typically conditions directly on representations from a single text encoder layer. By aligning diffusion models with the auto-regressive decoding nature of LLMs, deep fusion enables a more natural and tight-knit use of these models. However, despite the existing positive signals, its true potential remains uncertain. Current research [21, 38] prioritizes system-level benchmarks over controlled comparisons with established baselines, obscuring its position within the broader research landscape. More critically, the design space remains severely underexplored, and essential implementation details, such as training recipes, are often undisclosed. These limitations impede reproducibility and hinder broader adoption within the research community." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In this paper, we bridge these gaps through an empirical study on the deep fusion of a frozen LLM and a train" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28586" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "able DiT for text-to-image synthesis. We conduct controlled comparisons between the deep fusion approach and baselines, examine key design choices, and introduce a scalable, reproducible training recipe that delivers competitive performance on the established benchmarks for text-to-image generation. We believe that the evidence and unresolved questions highlighted in this study are of significant importance. We hope this work serves as a valuable resource, offering meaningful data points and practical guidelines to drive future advancements in multi-modal generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.255, + 0.233, + 0.27 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.483, + 0.385 + ], + "angle": 0, + "content": "Conditioning mechanisms in diffusion models. Numerous studies have examined effective methods for integrating linguistic conditional information into diffusion models to facilitate text-to-image synthesis. Latent diffusion models (LDM) [35] pioneered the use of cross-attention between image and text features, a technique that has since become standard in U-Net [36] based architectures [18, 30]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.673 + ], + "angle": 0, + "content": "With the rise of DiTs [29], vision transformer [8] has emerged as the dominant architecture for diffusion models, prompting a reassessment of conditioning mechanisms. The vanilla DiT employs adaLN-Zero modulation to inject conditional information. However, this approach is limited to using pooled text representations, which capture only coarse-grained information. Subsequent architectures have explored cross-attention [5, 6, 47] and self-attention [10, 11, 17, 51] for text conditioning, which typically extract representations from a single text encoder layer, usually the last or penultimate one. While this strategy aligns well with CLIP [31] and T5 [32] text encoders, it is inherently mismatched with the next-token prediction training objective of LLMs, where the last layer focuses on next-token prediction instead of learning discriminative representation. In contrast, the deep fusion approach shows promise in harnessing the internal information flow of LLMs, aligning with their in-context self-attention mechanism for processing information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Taming LLMs for diffusion models. The prevalence of decoder-only LLMs has driven extensive efforts to tame them for text-to-image diffusion models. The most effective and widely adopted practice involves leveraging LLMs to enrich input prompts [2], as well as employing multimodal LLMs (MLLMs) to generate synthetic captions for image data [2, 5, 10, 21, 47]. Another popular direction integrates (M)LLMs as system components, such as planners and discriminators [19, 44, 45, 49]. While these approaches have demonstrated effectiveness, they do not tap into the architecture of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To integrate LLMs into diffusion models, previous works attempt to replace text encoders with LLMs, either by training from scratch [25, 47, 51] or by aligning feature" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.089, + 0.825, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.905, + 0.303 + ], + "angle": 0, + "content": "Figure 2. Illustration of the attention mask. Each dotted square indicates whether the row can attend to the column." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.356 + ], + "angle": 0, + "content": "spaces [7, 15, 22, 28]. However, this substitution alone has not yielded the expected performance gains unless paired with sophisticated architectural adaptations [15, 25, 47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Recent research [20, 26, 46, 50] has proven that autoregressive decoding and denoising diffusion can be effectively unified within a single transformer for multi-modal generation. While earlier methods relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pretraining of LLMs by deeply fusing them with DiTs. From a text-to-image perspective, this architecture introduces a novel way to tame LLMs for text-to-image diffusion models, with the potential for achieving state-of-the-art performance [21]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.803, + 0.538 + ], + "angle": 0, + "content": "3. Deep Fusion of LLMs and DiTs" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.702, + 0.561 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.704 + ], + "angle": 0, + "content": "In the deep fusion approach, we integrate a frozen decoder-only LLM with a trainable DiT using layer-wise shared self-attention (Fig. 1). The DiT mirrors the LLM's transformer architecture, differing only in its input/output layers and timestep conditioning modules. This implements a two-stream transformer architecture that facilitates rich cross-modal interactions while maintaining modality-specific computation by utilizing distinct weight sets for processing tokens from different modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.885 + ], + "angle": 0, + "content": "The fused model processes text embeddings through the LLM stream and noisy image latents through the DiT stream. At each layer's self-attention operation, we concatenate token sequences from both streams, enabling the DiT to extract conditional information from the linguistic context. To preserve the pretrained LLM's functionality, we apply a causal attention mask to the text sequence and a bidirectional mask to the image sequence, permitting the image tokens to attend to text tokens but not vice versa (Fig. 2). After the final layer, we discard text tokens and use only image tokens to predict velocity, as typically done when training rectified flow models [23]." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Notably, only the key and value states of the text hidden" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28587" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "states are needed for self-attention. These remain constant throughout the diffusion process, allowing them to be efficiently cached and reused during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.147, + 0.273, + 0.164 + ], + "angle": 0, + "content": "3.2. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.261 + ], + "angle": 0, + "content": "We adopt the rectified flow formulation [23] to learn transport maps between the standard Gaussian noise distribution \\(\\pi_0\\) and the data distribution \\(\\pi_1\\), by connecting straight paths \\(x_{t} = tx_{1} + (1 - t)x_{0}\\) between samples \\(x_0\\sim \\pi_0,x_1\\sim \\pi_1\\) and learning an ODE model \\(dz_{t} = v_{\\theta}(z_{t},t)dt\\) on time \\(t\\in [0,1]\\) which converts \\(z_{0}\\) from \\(\\pi_0\\) to a \\(z_{1}\\) following \\(\\pi_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.292 + ], + "angle": 0, + "content": "We fit the velocity \\( v \\) with \\( x_{1} - x_{0} \\) with a prescribed time distribution \\( \\pi_{t} \\) by solving the following regression problem:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.312, + 0.483, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\min _ {v} \\mathbb {E} _ {t \\sim \\pi_ {t}, x _ {0} \\sim \\pi_ {0}, x _ {1} \\sim \\pi_ {1}} [ \\| v (x _ {t}, t) - (x _ {1} - x _ {0}) \\| ^ {2} ] d t \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.483, + 0.378 + ], + "angle": 0, + "content": "We parameterize \\( v \\) with network \\( \\theta \\) and solve Eq. (1) by stochastic optimization with empirical draws." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.268, + 0.409 + ], + "angle": 0, + "content": "4. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.507 + ], + "angle": 0, + "content": "To ensure fully open and reproducible comparison between the deep and shallow fusion approaches, we provide comprehensive details on the experimental setup, including the model, dataset, training, inference, and evaluation. For the same purpose, we exclusively use open-sourced pre-trained LLMs and publicly available datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.725 + ], + "angle": 0, + "content": "Model. We employ a frozen Gemma 2B [27] as the base LLM for all experiments (excluding Sec. 7). We pair it with a randomly initialized 2.5B-parameter DiT. The transformer configurations of the DiT strictly follows the base LLM, including the hidden size, number of layers, number of attention heads, FFN design, and other architectural details, ensuring both models have an identical 2B-parameter backbone. Following the vanilla DiT setup [29], we use 2D frequency absolute positional encoding, AdaLN-Zero timesteps conditioning, ViT [8]-style weight initialization, and a patch size of 2. To further stabilize training, we apply QK normalization to all layers. For all experiments we adopt the same 16-channel VAE from Stable Diffusion 3 (SD 3) [10]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Dataset. We use the CC12M [4] dataset with community-sourced synthetic captions [9] as our training set for all experiments excluding Sec. 7. Our downloaded version of the dataset includes 10.9M image-caption pairs. The images are resized and center-cropped to \\(512 \\times 512\\) and the texts are padded or truncated to 256 tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training. We train all models with a batch size of 512 using AdamW [24] optimizer \\((\\beta_{1} = 0.9, \\beta_{2} = 0.999)\\) in BF16 mixed precision. We use a constant learning rate of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "\\(1 \\times 10^{-4}\\), a weight decay of \\(1 \\times 10^{-4}\\), and gradient clipping with a threshold of 1.0. Exponential moving average of the weights are gathered by a decay factor of 0.99 every 100 steps. We employ the same logit-normal distribution as used in SD 3 for timestep sampling. During training, \\(10\\%\\) of the texts are randomly dropped to learn unconditional generation. Training is carried out using Google TPU v4-256 pods and FSDP implemented by PyTorch / XLA SPMD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.907, + 0.313 + ], + "angle": 0, + "content": "Inference. We conduct inference using Euler discretization with 25 sampling steps, and a classifier-free guidance scale of 6 which we find to be near optimal for text-image alignment. We employ identical sampling steps and guidance scale across all experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Evaluation. We evaluate image-text alignment using GenEval [12] and DPG-Bench [15] metrics, prioritizing GenEval for its robustness. While both benchmarks provide valuable insights, DPG-Bench exhibits certain limitations, such as rapid performance saturation and potential measurement errors [21]. To ensure a comprehensive evaluation, we also provide visual quality measurements using FID [14] on MJHQ-30K [18]. Notably, image-text alignment does not always correlate positively with visual quality, often presenting trade-offs. Our sampling and evaluation are primarily carried out using NVIDIA L40S GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.851, + 0.535 + ], + "angle": 0, + "content": "5. Comparing Deep and Shallow Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.906, + 0.634 + ], + "angle": 0, + "content": "The deep fusion approach fuses an LLM and a DiT through layer-wise shared self-attention, creating interconnections throughout the network. However, established architectures typically condition on representations from a single text encoder layer. To investigate the true potential of deep fusion, we conduct controlled comparisons with baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.906, + 0.757 + ], + "angle": 0, + "content": "For a fair and meaningful comparison, we examine a common architectural paradigm which we refer to as shallow fusion. In this approach, representations from a single text encoder layer are integrated into each DiT layer through late fusion within the attention operation. Unlike deep fusion, which involves multi-layered interactions, shallow fusion maintains a fixed connection between each DiT layer and a prescribed text encoder layer (Fig. 1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.744, + 0.783 + ], + "angle": 0, + "content": "5.1. Shallow Fusion Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.906, + 0.852 + ], + "angle": 0, + "content": "We consider two shallow fusion architectures that condition on last-layer hidden states of LLMs as our baselines. As illustrated in Fig. 3, the two architecture differ in how they aggregate information from the condition." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- Self-attention DiT. In this design, text representations are projected to key and value states and then concatenated with those of image hidden states in self-attention, which" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.315, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.175, + 0.25, + 0.187 + ], + "angle": 0, + "content": "(a) Self-Attention" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.088, + 0.484, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.175, + 0.466, + 0.187 + ], + "angle": 0, + "content": "(b) Cross-Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Figure 3. Illustration of cross-modal attention in the shallow fusion baselines. The key and query states of the condition are directly projected from text representations." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.253, + 0.483, + 0.299 + ], + "angle": 0, + "content": "can also be decoupled by running self-attention and cross-attention in parallel and merging their outputs. This approach resembles architectures proposed by [11, 51]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.298, + 0.484, + 0.388 + ], + "angle": 0, + "content": "- Cross-attention DiT. This design also projects text representations to key and values states. However, unlike the previous approach, they are used for additional cross-attention with image hidden states, applied after the self-attention in each layer. This architecture follows the methodology employed in [5, 6, 47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Compared to the deep fusion approach, both baseline models include RMS normalization and a linear layer for text representations before passing them through additional key and value projection layers. Additionally, the cross-attention DiT model uses extra query projections in its cross-attention mechanism for image hidden states. Other model configurations follow the DiT design in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.667 + ], + "angle": 0, + "content": "Notably, the deep fusion approach can be reinterpreted as a variant of the self-attention DiT architecture, as they both aggregate conditional information through in-context self-attention. The key difference lies in how the key and value states of the condition are derived: self-attention DiT employs a trainable projection to generate these states from a single text encoder layer, whereas deep fusion extracts them from corresponding LLM layers. Despite their similarities, the two approaches have fundamentally different conceptual implications: deep fusion treats the LLM and DiT as equal components of a unified model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.484, + 0.803 + ], + "angle": 0, + "content": "Apart from the aforementioned designs, SD3 [10] introduced an alternative fusion strategy that uses a two-stream transformer (MM-DiT) to jointly processes noised image latents and linguistic representations. While this is another popular approach worthy of investigation, an apples-to-apples fair comparison between the deep fusion approach and MM-DiT is not feasible, as both streams in MM-DiT are trainable. Consequently, our analysis focuses solely on the shallow fusion baselines detailed in this section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.31, + 0.833 + ], + "angle": 0, + "content": "5.2. Controlled Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "For a fair comparison, we design self-attention DiT, cross-attention DiT, and the deep fusion model with similar architectures, and train them for 300K steps following the setup detailed in Sec. 4." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.621, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.09, + 0.724, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.09, + 0.807, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.09, + 0.89, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.201, + 0.907, + 0.243 + ], + "angle": 0, + "content": "Figure 4. Illustration of timestep conditioning strategies. Removing timesetp conditioning leads to the fewest parameters and the best overall performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.907, + 0.386 + ], + "angle": 0, + "content": "As shown in Tab. 1, the deep fusion model achieve significantly better performance in image-text alignment than the self-attention DiT model and also surpass the cross-attention DiT model, while shallow fusion models demonstrate better visual quality. In terms of inference efficiency, deep fusion also demonstrates competitive performance, as shown in Tab. 2. This positive evidence underscores the compelling positioning of the deep fusion approach within the current landscape." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.39, + 0.912, + 0.444 + ], + "angle": 0, + "content": "
MethodParams.GenEval ↑DPG ↑FID ↓
Self-Attention2.47B0.4273.926.16
Cross-Attention2.62B0.4976.324.00
Deep Fusion2.45B0.5176.627.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.446, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Table 1. Comparison of performance between deep and shallow fusion models. Deep fusion beats shallow fusion in text-image alignment while underperforms in visual quality." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.496, + 0.876, + 0.549 + ], + "angle": 0, + "content": "
MethodParams.Inference Latency (s)
Self-Attention2.47B1.75
Cross-Attention2.62B1.86
Deep Fusion2.45B1.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.551, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Table 2. Comparison of inference latency between deep and shallow fusion models. The numbers are measured with a batch size of 1 in automatic mixed precision on an NVIDIA A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.8, + 0.62 + ], + "angle": 0, + "content": "6. Examining Key Design Choices" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.763 + ], + "angle": 0, + "content": "In this section, we examine key design choices of the deep fusion approach through a text-to-image-centric lens. We begin by assessing the necessity and potential redundancy of parameters for timestep conditioning, determining whether to optimize their use or eliminate them altogether (Sec. 6.1). Next, we compare various positional encoding strategies (Sec. 6.2). Additionally, we investigate how the choice of base LLM and the use of instruction prompts impacts text-to-image performance (Sec. 6.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.809 + ], + "angle": 0, + "content": "For all experiments, we use the default design as the baseline and train the models for 300K steps, following the setup detailed in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.727, + 0.835 + ], + "angle": 0, + "content": "6.1. Timestep Conditioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "DiT [29] has introduced AdaLN-Zero as the standard mechanism for injecting timestep and class label information. The AdaLN modules typically accounts for a large proportion of the model parameters, 0.5B out of a total 2.5B in our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28589" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "case. However, since our text-to-image model does not use class labels, AdaLN serves only for timestep conditioning. This substantial parameter allocation raises a critical question about its necessity and potential redundancy: Could these parameters be utilized more effectively, or are they redundant altogether?" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.183, + 0.482, + 0.213 + ], + "angle": 0, + "content": "Question 1.1. Can adaLN parameters be utilized more effectively by integrating additional text conditioning?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.213, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Prior research [10, 11, 17] has integrated text information into AdaLN-Zero by augmenting timestep embeddings with pooled text representations through summation. To optimize parameter efficiency, we follow this approach by leveraging embeddings1 from the CLIP L/14 text encoder [31], which provides high-quality linguistic representations trained on large-scale multi-modal data." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.322, + 0.453, + 0.365 + ], + "angle": 0, + "content": "
MethodGenEval ↑DPG ↑FID ↓
adaLN-Zero0.5176.627.33
+ CLIP L/140.5076.224.00
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.418 + ], + "angle": 0, + "content": "The results indicate that integrating text modulation results in slight improvements in FID but weakens image-text alignment. Additionally, it further increases compute." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.448 + ], + "angle": 0, + "content": "Question 1.2. Can we shrink the parameters for timestep conditioning?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.539 + ], + "angle": 0, + "content": "We explore the possibility of eliminating timestep conditioning parameters to develop a more streamlined architecture akin to that of the LLM. Specifically, we compare four timestep conditioning strategies from previous work, each with a progressively reduced number of parameters, as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.541, + 0.483, + 0.586 + ], + "angle": 0, + "content": "- adaLN-Zero. Following the vanilla DiT [29], we regress zero-initialized modulation parameters for each layer from the timestep embedding." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.587, + 0.483, + 0.631 + ], + "angle": 0, + "content": "- adaLN-Single. Following PixArt-\\(\\alpha\\) [6], we compute a global set of modulation parameters and refine them per layer by adding learnable embeddings." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.632, + 0.483, + 0.662 + ], + "angle": 0, + "content": "- Addition. Following Transfusion [50], we directly add the timestep embedding to all image tokens." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.662, + 0.483, + 0.692 + ], + "angle": 0, + "content": "- w/o Timestep. Inspired by a recent study [42], we completely remove the timestep conditioning from the model." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.541, + 0.483, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.701, + 0.488, + 0.768 + ], + "angle": 0, + "content": "
MethodParams.GenEval ↑DPG ↑FID ↓
adaLN-Zero2.47B0.5176.627.33
adaLN-Single2.01B0.4775.227.09
addition1.99B0.4775.626.40
w/o timestep1.98B0.4976.721.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.769, + 0.482, + 0.798 + ], + "angle": 0, + "content": "Table 3. Evaluation of timestep conditioning strategies. Removing timestep conditioning yields surprisingly strong results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.482, + 0.868 + ], + "angle": 0, + "content": "Surprisingly, the results in Tab. 3 indicate that reducing the number of parameters in timestep conditioning consistently enhances visual quality, whereas the performance in image-text alignment exhibits fluctuations." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.09, + 0.9, + 0.169 + ], + "angle": 0, + "content": "
Input SequenceTXT 1TXT 2[PAD]IMG 1IMG 2IMG 3IMG 4
1D-RoPE + APE[0][1]—[2D Frequency APE]
1D-RoPE[0][1]—[2][3][4][5]
1D + 2D-RoPE[0][1]—[2, 2][2, 3][3, 2][3, 3]
M-RoPE[0, 0][1, 1]—[2, 2][2, 3][3, 2][3, 3]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.174, + 0.902, + 0.188 + ], + "angle": 0, + "content": "Figure 5. Illustration of RoPE. The indices denote position IDs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.195, + 0.905, + 0.376 + ], + "angle": 0, + "content": "Notably, the strategy that completely removes timestep conditioning not only achieves significantly better FID but also maintains comparable GenEval and DPG-Bench performance. This finding is consistent with [42], where timestep conditioning removal improved FID in rectified flow models trained on smaller datasets. Furthermore, the complete removal of timestep conditioning eliminates the need for associated parameters, resulting in a \\(20\\%\\) reduction in the total number of model parameters. Although this approach slightly lags behind AdaLN-Zero in terms of text-image alignment metrics, we prefer it due to its parameter efficiency and architectural simplicity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.385, + 0.704, + 0.401 + ], + "angle": 0, + "content": "6.2. Positional Encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.905, + 0.512 + ], + "angle": 0, + "content": "Absolute positional encoding (APE) is widely used in text-to-image diffusion models, whereas in the context of LLMs, rotary positional embedding (RoPE) [40] is the predominant choice. As deep fusion models are inherently multimodal and differ from traditional text-to-image diffusion models, it is unclear which positional encoding (or their combinations) is best suited for mixed-modal sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.905, + 0.542 + ], + "angle": 0, + "content": "Question 2.1. Is RoPE more advantageous than APE for enhancing the performance of deep fusion models?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.905, + 0.572 + ], + "angle": 0, + "content": "To explore this, we experiment with three positional encoding methods, as illustrated in Fig. 5:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.574, + 0.905, + 0.604 + ], + "angle": 0, + "content": "- 1D RoPE + APE: We apply 1D RoPE to the text sequence and APE to the image sequence respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.605, + 0.905, + 0.635 + ], + "angle": 0, + "content": "- 1D RoPE: We extend 1D RoPE to encompass both text and image sequences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.635, + 0.905, + 0.665 + ], + "angle": 0, + "content": "- 1D + 2D RoPE: We apply 1D RoPE to the text sequence and 2D RoPE to the image sequence respectively." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.574, + 0.905, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.673, + 0.872, + 0.726 + ], + "angle": 0, + "content": "
MethodGenEval ↑DPG ↑FID ↓
1D-RoPE + APE0.5176.627.33
1D-RoPE0.4677.027.94
1D + 2D-RoPE0.5176.425.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.728, + 0.905, + 0.756 + ], + "angle": 0, + "content": "Table 4. Comparing different positional encoding strategies. \\(1\\mathrm{D} + 2\\mathrm{D}\\) RoPE achieves the best overall performance." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.905, + 0.902 + ], + "angle": 0, + "content": "As shown in Tab. 4, the 1D + 2D-RoPE configuration achieves the best overall performance, with only a marginal decrease in DPGBench compared to the 1D-RoPE + APE variant. The superiority of 2D-RoPE over APE suggests RoPE is more effective for modeling image sequences in deep fusion models. Using only 1D-RoPE slightly reduces performance, indicating that while deep fusion models treat text and image sequences as a unified input, their distinct positional characteristics are best modeled separately." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1We apply RMS-normalization and an MLP to the text embeddings before adding them to the timestep embeddings." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28590" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "Question 2.2. Does deep fusion models benefit from RoPE specifically designed for mixed-modal sequences?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.243 + ], + "angle": 0, + "content": "Previous work on MLLMs has explored RoPE strategies for handling mixed-modal sequences. Naturally, we are curious whether deep fusion models can benefit from these positional encodings. Follow Qwen2-VL, we implement M-RoPE, a variant of RoPE that applies 2D positional IDs to chunked 1D RoPE frequencies, allowing it to function as 1D RoPE for text sequences while approximating 2D RoPE for image sequences." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.252, + 0.454, + 0.294 + ], + "angle": 0, + "content": "
MethodGenEval ↑DPG ↑FID ↓
1D + 2D-RoPE0.5176.425.42
M-RoPE0.4974.927.60
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.364 + ], + "angle": 0, + "content": "Although M-RoPE elegantly unifies 1D and 2D-RoPEs, it still falls short compared to their direct combination. This underscores the challenge of designing position encodings for mixed-modal sequences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.21, + 0.388 + ], + "angle": 0, + "content": "6.3. Base LLM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.456 + ], + "angle": 0, + "content": "LLMs trained with different paradigms and data demonstrate diverse capabilities and behaviors. In this section, we examine how the choice of base LLM impacts the performance of deep fusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.487 + ], + "angle": 0, + "content": "Question 3.1 Can instruction tuning, combined with instruction prompts, improve text-to-image performance?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Instruction tuning enables LLMs to effectively follow complex instructions. We explore whether this process can also enhance their internal information flow, leading to more contextualized and discriminative representations for text-to-image synthesis. We compare Gemma 2B with Gemma 2B IT, its instruction-tuned variant. We also experiment with using it with a simple instruction prompt, \"Imagine: \", which we find sufficient for guiding the LLM to generate detailed and relevant expansions of the input." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.63, + 0.468, + 0.684 + ], + "angle": 0, + "content": "
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ instruction tuning0.4975.427.04
+ instruction prompt0.5075.825.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Table 5. Evaluating the effect of instruction tuning. Using instruction-tuned LLMs does not improve performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.81 + ], + "angle": 0, + "content": "As shown in Tab. 5, instruction tuning appears to have a slightly negative impact on performance. While the use of an instruction prompt mitigates this effect to some extent, consistent with findings in [25, 47], it still falls short of the baseline. This result highlights a challenge in effectively leveraging the instruction-following capabilities of LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Question 3.2 Can multi-modal tuning improve text-to-image performance?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Additionally, we're also interested in the effect of multi-modal tuning. While multi-modal tuning differs significantly from our setup, its shift in data distribution may still potentially enhance adaptability to multi-modal tasks. We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "compare Gemma 2B with the base LLM of PaliGemma 3B PT [3], a multi-modal extension of Gemma 2B that undergoes additional pretraining on image-text data." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.14, + 0.892, + 0.182 + ], + "angle": 0, + "content": "
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ multi-modal tuning0.5276.226.30
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.187, + 0.905, + 0.231 + ], + "angle": 0, + "content": "which yields small improvements in performance. This observation indicates that multi-modal finetuning provides some benefit to the deep-fusion model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.261 + ], + "angle": 0, + "content": "Question 3.3 Does improved LLM capabilities translate to stronger text-to-image performance?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.906, + 0.443 + ], + "angle": 0, + "content": "Finally, as the base LLM become more proficient in understanding and generating text, they could potentially foster synergistic improvements in DiT performance. We compare Gemma 2B with Gemma 2 2B [34], the next generation of Gemma 2B which demonstrates a \\(6\\%\\) absolute performance improvement \\((0.44 \\rightarrow 0.50)\\) on an average of 8 language-only benchmarks. Notably, Gemma 2 2B features a different transformer architecture than Gemma 2B, prompting us to adjust the DiT architecture accordingly. This modification increases the number of adaLN parameters by 0.3B. However, as demonstrated in Sec. 6.1, these parameters do not contribute to model performance." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.447, + 0.876, + 0.489 + ], + "angle": 0, + "content": "
ModelGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
Gemma 2 2B0.5479.123.94
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Upgrading from Gemma 2B to Gemma 2 2B yields a drastic performance boost. This finding suggests that the DiT's performance in deep fusion models is strongly dependent on the capabilities of the underlying base LLM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.568, + 0.682, + 0.585 + ], + "angle": 0, + "content": "7. Training at Scale" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.684 + ], + "angle": 0, + "content": "In this section, we present a final recipe for the deep fusion model, building on the original framework while incorporating key insights from previous exploration. We conduct large-scale training to benchmark our model against established systems, showcasing its scalability and competitive performance on the leaderboard." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.647, + 0.71 + ], + "angle": 0, + "content": "7.1. Final Recipe" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.905, + 0.745 + ], + "angle": 0, + "content": "Building on the insights from Sec. 6, we introduce the following design modifications to our model:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.748, + 0.74, + 0.761 + ], + "angle": 0, + "content": "- Remove AdaLN-Zero modules." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.763, + 0.842, + 0.777 + ], + "angle": 0, + "content": "- Replace 1D-RoPE + APE with 1D + 2D-RoPE." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.778, + 0.905, + 0.808 + ], + "angle": 0, + "content": "- Replace Gemma 2B with Gemma 2 2B, adjusting the transformer configurations of the DiT accordingly." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.748, + 0.905, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We train our model for 800K steps on a mixed dataset comprising CC12M [4], SA-1B [16], and the training subset of JourneyDB [41], amounting to approximately 26M image-caption pairs. Notably, state-of-the-art text-to-image models typically rely on high-quality datasets of much larger scale to achieve superior performance. For CC12M" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.221, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.18, + 0.19, + 0.189 + ], + "angle": 0, + "content": "(a) Fully Parallel" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.091, + 0.35, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.181, + 0.324, + 0.19 + ], + "angle": 0, + "content": "(b) Skipped Layers" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.091, + 0.48, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.181, + 0.466, + 0.19 + ], + "angle": 0, + "content": "(c) Decoupled Dimension" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.205, + 0.483, + 0.247 + ], + "angle": 0, + "content": "Figure 6. Illustration of architecture alignment. Dashed boxes indicate parameters that have been reduced by decreasing either the hidden size or the number of layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.287 + ], + "angle": 0, + "content": "and SA-1B, we utilize synthetic captions [6, 9]. Other experimental setup follows Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.296, + 0.327, + 0.312 + ], + "angle": 0, + "content": "7.2. Performance Comparison" + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.322, + 0.485, + 0.535 + ], + "angle": 0, + "content": "
ModelParamsDataGen.↑DPG↑FID↓
SD 1.5 [35]0.9B4.8B0.4363.2—
DALL-E 2 [33]4.2B2.6B0.52——
SDXL [30]2.6B1.6B0.5574.76.63
PG 2.5 [18]2.6B—0.5675.56.09
SD 3 M [10]2B1B0.6284.111.92
DALL-E 3 [2]——0.6783.5—
FLUX.1 [dev] [17]12B—0.6784.010.15
PG 3 [21]24B—0.7687.0—
MicroDiT [37]1.2B37M0.46——
PixArt-α [6]0.6B25M0.4871.1—
Lumina Next [11]2B—0.4674.67.58
PixArt-Σ [5]0.6B46M0.5480.56.15
Transfusion [50]7.3B3.5B0.63——
Sana 1.0 1.6B [47]1.6B—0.6684.85.76
Ours2B26M0.6081.67.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Table 6. Comparison of performance with state-of-the-art systems. Table adapted from [48, 50]. Industrial baselines are presented at the top, while academic baselines and our model are listed at the bottom." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.483, + 0.691 + ], + "angle": 0, + "content": "We compare our model with the most advanced text-to-image diffusion models in Tab. 6. Despite being trained with limited compute and data in a simplified setting, our model surpasses many industry-standard systems and delivers competitive results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.739 + ], + "angle": 0, + "content": "We present qualitative examples from our model in Figure 7. Our model demonstrates the ability to generate high-quality images with superior prompt alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.752, + 0.288, + 0.77 + ], + "angle": 0, + "content": "8. Further Exploration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.839 + ], + "angle": 0, + "content": "In this section, we present preliminary studies exploring more aggressive modifications to the deep fusion approach. For all experiments, we train the models for 300K steps, following the setup detailed in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.311, + 0.866 + ], + "angle": 0, + "content": "8.1. Architecture Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Up to this point, our approach has followed prior work by aligning the LLM and DiT backbones in a layer-by-layer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "approach, strictly enforcing identical transformer configurations for both models. However, this rigid constraint limits the flexibility of deep fusion. In practice, we need the ability to scale the LLM and DiT independently, as different modalities follow distinct scaling laws and network design principles, and training and deployment scenarios vary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.322 + ], + "angle": 0, + "content": "To address this, we explore modifying the DiT model's hidden size and number of transformer layers, as illustrated in Fig. 6. The adapted model is fused into the middle layers of the LLM, which contain richer semantic information [39]. Additionally, in self-attention, hidden states are still projected to query, key, and value states that match the LLM's dimensionality, ensuring compatibility. Since the most successful DiTs are generally much smaller than state-of-the-art LLMs, we focus on shrinking the size of our DiT." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.332, + 0.892, + 0.399 + ], + "angle": 0, + "content": "
Hidden sizeParams.GenEval ↑DPG ↑FID ↓
20482.5B0.5176.627.33
17922.1B0.5077.124.27
15361.8B0.4976.225.46
12801.4B0.4874.824.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.4, + 0.905, + 0.427 + ], + "angle": 0, + "content": "Table 7. Evaluating models of different hidden sizes. The default hidden size is 2048." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.444, + 0.892, + 0.497 + ], + "angle": 0, + "content": "
LayersParams.GenEval ↑DPG ↑FID ↓
182.5B0.5176.627.33
141.9B0.4774.623.46
101.4B0.3368.028.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.499, + 0.905, + 0.527 + ], + "angle": 0, + "content": "Table 8. Evaluating different numbers of layers. The default number of layer is 18." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.693 + ], + "angle": 0, + "content": "As shown in Table Tab. 7, the model's performance degrades gracefully as we reduce the hidden size, with visual quality actually improving in some cases. While decreasing the number of transformer layers (Tab. 8) also yields acceptable results, performance deteriorates more quickly. We hypothesize this occurs because Gemma 2B already employs fewer layers than typical model architectures. These findings suggest that LLM and DiT model designs can be effectively decoupled, enabling the application of separate scaling laws and design principles for each component." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.719, + 0.725 + ], + "angle": 0, + "content": "8.2. Attention Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In Sec. 3, we build on prior work by defining the deep fusion architecture with shared self-attention to bridge the LLM and DiT. Inspired by extensive research on cross-attention in MLLMs [1, 13] and our findings in Sec. 5, which highlight the superior performance of cross-attention DiT over self-attention DiT, we explore an alternative deep fusion variant. This new approach replaces shared self-attention with cross-attention mechanisms, similar to cross-attention DiT but with a key distinction: we substitute the projected linguistic key and value states in traditional cross-attention DiT with corresponding states from LLM layers." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.091, + 0.299, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.249, + 0.286, + 0.271 + ], + "angle": 0, + "content": "Dew on blue rose petals, HD, close up, detail" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.091, + 0.498, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.249, + 0.49, + 0.271 + ], + "angle": 0, + "content": "A corgi wearing sunglasses walks on the beach of a tropical island" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.699, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.249, + 0.695, + 0.26 + ], + "angle": 0, + "content": "A rally car taking a fast turn on a track" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.091, + 0.899, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.249, + 0.858, + 0.27 + ], + "angle": 0, + "content": "An armchair in the shape of an avocado" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.274, + 0.299, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.432, + 0.286, + 0.464 + ], + "angle": 0, + "content": "A man riding a horse through the Gobi Desert with a beautiful sunset behind him, movie quality." + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.274, + 0.498, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.432, + 0.462, + 0.453 + ], + "angle": 0, + "content": "Pirate ship trapped in a cosmic maelstrom nebula" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.274, + 0.698, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.432, + 0.68, + 0.453 + ], + "angle": 0, + "content": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.274, + 0.899, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.432, + 0.894, + 0.453 + ], + "angle": 0, + "content": "A person surfing on a wave of stars in outer space." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.467, + 0.3, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.625, + 0.298, + 0.685 + ], + "angle": 0, + "content": "a picturesque autumn scene where a quaint cottage with a thatched roof sits beside a tranquil lake, surrounded by trees with leaves in vibrant shades of orange, red, and yellow. The cottage's wooden exterior is complemented by white-framed windows, and a stone chimney rises above the roofline. The lake reflects the warm fall colors, creating a mirror image of the foliage and the small structure on its calm surface." + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.467, + 0.498, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.625, + 0.485, + 0.668 + ], + "angle": 0, + "content": "A portrait of a human growing colorful flowers from her hair. Hyperrealistic oil painting. Intricate details." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.467, + 0.697, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.625, + 0.69, + 0.683 + ], + "angle": 0, + "content": "a black and white picture of a woman looking through the window, in the style of Duffy Sheridan, Anna Razumovskaya, smooth and shiny, wavy, Patrick Demarchelier, album covers, lush and detailed" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.467, + 0.899, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.625, + 0.879, + 0.669 + ], + "angle": 0, + "content": "13 year old farm boy, red tousled hair, green eyes, soft pale skin, freckles, stub nose, very beautiful, very pretty, dark fantastic medieval" + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.701, + 0.641, + 0.715 + ], + "angle": 0, + "content": "Figure 7. Samples generated by our final recipe." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.72, + 0.452, + 0.761 + ], + "angle": 0, + "content": "
MethodGenEval ↑DPG ↑FID ↓
self-attention0.5176.627.33
cross-attention0.5276.526.57
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.482, + 0.863 + ], + "angle": 0, + "content": "This modification yields minor gains, though at a cost to the LLM-DiT parity. Additionally, we find that although cross-attention introduces a negligible increase in FLOPs and parameter count, it leads to approximately a \\(12\\%\\) increase in latency\\(^{2}\\) (1.66s vs. 1.86s). Therefore, we retained the self-attention design in our final configuration." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.722, + 0.633, + 0.737 + ], + "angle": 0, + "content": "9. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.907, + 0.868 + ], + "angle": 0, + "content": "We have studied the recently popular deep fusion of a frozen LLM with a trainable DiT for text-to-image synthesis. Our findings provide empirical evidence supporting its advantages over baselines. We highlight key design choices, identify unresolved problems, and offer meaningful data points alongside practical guidelines. We hope our empirical work help advance multi-modal generation and bridge the gap between auto-regressive decoding and denoising diffusion." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2Latency is measured with a batch size of 1 under automatic mixed precision on an NVIDIA A100 GPU" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28593" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.254 + ], + "angle": 0, + "content": "We thank the reviewers for their constructive feedback. We also thank Shusheng Yang, Shengbang Tong, Wenhao Chai, Nanye Ma, Sihan Xu, and Chenyu Li for insightful discussions. This work was mainly supported by the Google TPU Research Cloud (TRC) program, and the Google Cloud Research Credits program (GCP19980904) and Open Path AI Foundation. SX also acknowledges support from Intel AI SRS, IITP grant funded by the Korean Government (MSIT) (RS-2024-00457882, National AI Research Lab Project), Amazon Research Award, and NSF Award IIS-2443404." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.28, + 0.188, + 0.295 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.374, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.opencai.com/papers/dall-e-3.pdf, 2023.1,2,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.444, + 0.483, + 0.512 + ], + "angle": 0, + "content": "[3] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv:2407.07726, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.567 + ], + "angle": 0, + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.484, + 0.637 + ], + "angle": 0, + "content": "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\sigma\\) : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024.1,2,4,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.706 + ], + "angle": 0, + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\alpha\\): Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[7] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Caption Emporium. conceptual-captions-cc12m-llavanext. https://huggingface.co/datasets/CaptionEmporium / conceptual - captions - cc12m-llavanext, 2024.3,7" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[11] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv:2405.05945, 2024. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[12] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. arXiv:2310.11513, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[15] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv:2403.05135, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[17] Black Forest Labs. Announcing black forest labs. https://blackforestlabs.ai/announcing-black-forest-labs/, 2024. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[18] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv:2402.17245, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[19] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv:2305.13655, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[20] Weixin Liang, Lili Yu, Liang Luo, Srinivasan Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, et al. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. arXiv:2411.04996, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[21] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv:2409.10695, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[22] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv:2407.00737, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv:2209.03003, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28594" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[25] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv:2406.11831, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[26] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv:2411.07975, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.25, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[27] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv:2403.08295, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.377, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.477, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.562, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In JMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[34] Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv:2408.00118, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[37] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv:2407.15811, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, and Lili Yu. Lmfusion:" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Adapting pretrained language models for multimodal generation. arXiv:2412.15188, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[39] Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models. arXiv:2502.02013, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[40] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[41] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. In NeurIPS, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[42] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv:2502.13129, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[44] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv:2401.15688, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[45] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[46] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv:2409.11340, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[47] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv:2410.10629, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[48] Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv:2501.18427, 2025. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[49] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[50] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv:2408.11039, 2024. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[51] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv:2406.18583, 2024. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.828 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28595" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_origin.pdf b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f3ad439efffcb18336512c317b41ff9a16c8b392 --- /dev/null +++ b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/3434d4fb-2873-49b0-8ac0-297681602f31_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:231f2d8ce652a03086e17881cb580233eeca979dfe5fe9e9a54f26a6bd3a1e58 +size 1327809 diff --git a/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/full.md b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5309563c2bb4f297a3bf4c9d03c7181b70ab823 --- /dev/null +++ b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/full.md @@ -0,0 +1,407 @@ +# Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis + +Bingda Tang $^{1}$ Boyang Zheng $^{1}$ Sayak Paul $^{2}$ Saining Xie $^{1}$ + +1 New York University 2 Hugging Face + +# Abstract + +This paper does not describe a new method; instead, it provides a thorough exploration of an important yet understudied design space related to recent advances in text-to-image synthesis—specifically, the deep fusion of large language models (LLMs) with diffusion transformers (DiTs) for multimodal generation. Previous studies mainly focused on overall system performance rather than detailed comparisons with alternative methods, and key design details and training recipes were often left undisclosed. These gaps create uncertainty about the real potential of this approach. To fill these gaps, we conduct an empirical study on text-to-image generation, performing controlled comparisons with established baselines, analyzing important design choices, and providing a clear, reproducible recipe for training at scale. We hope this work offers meaningful data points and practical guidelines for future research in multimodal generation. + +# 1. Introduction + +Text-to-image diffusion models have made remarkable progress in generating high-quality images from descriptive texts. Current state-of-the-art systems [2, 5, 10, 17, 30] typically derive text representations from specialized encoders, such as CLIP [31] and T5 [32]. With the rise of decoder-only large language models (LLMs), there has been a growing amount of interest in their potential as replacements for these traditional text encoders [7, 15, 21, 22, 25, 28, 48, 51]. However, simply substituting LLMs has not yielded expected performance gains unless coupled with sophisticated architectural adaptations [15, 25, 47]. Prior work [25] attributes this to the misalignment between the next token prediction training objective of LLMs and the need for discriminative text representations in diffusion models. + +Recent advancements [20, 26, 46, 50] have successfully unified auto-regressive decoding and denoising diffusion within a single transformer [43], enabling seamless multi-modal generation. This unified approach supports a wide range of tasks, including instructed image-to-text, + +![](images/2990b2bdacd0d68abdaa088ddd2b9dcb6414fa2edebd3e18de8f202c2d3f0c1d.jpg) +(a) Shallow-Fusion + +![](images/afc01f31583e83df2d8990efbdc392754d30cf73a6ce7972ba37023b9146e5f9.jpg) +(b) Deep-Fusion +Figure 1. Illustration of the deep fusion approach and baselines. We conduct controlled comparisons with baseline methods that incorporate text representations from a single text encoder layer into each DiT layer using late fusion within the attention mechanism, a strategy we term the shallow fusion approach. + +text-to-image synthesis, and interleaved image-text generation. While earlier methods often relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pre-training of LLMs by deeply fusing them with diffusion transformers (DiTs) [29] through layer-wise shared self-attention. This design facilitates rich cross-modal interactions while maintaining modality-specific computation by using distinct sets of weights. When optimized for text-to-image generation, it claims state-of-the-art performance [21]. + +Deep fusion presents a compelling alternative to existing architectures for text-to-image synthesis, which typically conditions directly on representations from a single text encoder layer. By aligning diffusion models with the auto-regressive decoding nature of LLMs, deep fusion enables a more natural and tight-knit use of these models. However, despite the existing positive signals, its true potential remains uncertain. Current research [21, 38] prioritizes system-level benchmarks over controlled comparisons with established baselines, obscuring its position within the broader research landscape. More critically, the design space remains severely underexplored, and essential implementation details, such as training recipes, are often undisclosed. These limitations impede reproducibility and hinder broader adoption within the research community. + +In this paper, we bridge these gaps through an empirical study on the deep fusion of a frozen LLM and a train + +able DiT for text-to-image synthesis. We conduct controlled comparisons between the deep fusion approach and baselines, examine key design choices, and introduce a scalable, reproducible training recipe that delivers competitive performance on the established benchmarks for text-to-image generation. We believe that the evidence and unresolved questions highlighted in this study are of significant importance. We hope this work serves as a valuable resource, offering meaningful data points and practical guidelines to drive future advancements in multi-modal generation. + +# 2. Related Work + +Conditioning mechanisms in diffusion models. Numerous studies have examined effective methods for integrating linguistic conditional information into diffusion models to facilitate text-to-image synthesis. Latent diffusion models (LDM) [35] pioneered the use of cross-attention between image and text features, a technique that has since become standard in U-Net [36] based architectures [18, 30]. + +With the rise of DiTs [29], vision transformer [8] has emerged as the dominant architecture for diffusion models, prompting a reassessment of conditioning mechanisms. The vanilla DiT employs adaLN-Zero modulation to inject conditional information. However, this approach is limited to using pooled text representations, which capture only coarse-grained information. Subsequent architectures have explored cross-attention [5, 6, 47] and self-attention [10, 11, 17, 51] for text conditioning, which typically extract representations from a single text encoder layer, usually the last or penultimate one. While this strategy aligns well with CLIP [31] and T5 [32] text encoders, it is inherently mismatched with the next-token prediction training objective of LLMs, where the last layer focuses on next-token prediction instead of learning discriminative representation. In contrast, the deep fusion approach shows promise in harnessing the internal information flow of LLMs, aligning with their in-context self-attention mechanism for processing information. + +Taming LLMs for diffusion models. The prevalence of decoder-only LLMs has driven extensive efforts to tame them for text-to-image diffusion models. The most effective and widely adopted practice involves leveraging LLMs to enrich input prompts [2], as well as employing multimodal LLMs (MLLMs) to generate synthetic captions for image data [2, 5, 10, 21, 47]. Another popular direction integrates (M)LLMs as system components, such as planners and discriminators [19, 44, 45, 49]. While these approaches have demonstrated effectiveness, they do not tap into the architecture of diffusion models. + +To integrate LLMs into diffusion models, previous works attempt to replace text encoders with LLMs, either by training from scratch [25, 47, 51] or by aligning feature + +![](images/58bec034de58bec47fbb3d5554dd44df984efe3fe8d58554c69688dd07471973.jpg) +Figure 2. Illustration of the attention mask. Each dotted square indicates whether the row can attend to the column. + +spaces [7, 15, 22, 28]. However, this substitution alone has not yielded the expected performance gains unless paired with sophisticated architectural adaptations [15, 25, 47]. + +Recent research [20, 26, 46, 50] has proven that autoregressive decoding and denoising diffusion can be effectively unified within a single transformer for multi-modal generation. While earlier methods relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pretraining of LLMs by deeply fusing them with DiTs. From a text-to-image perspective, this architecture introduces a novel way to tame LLMs for text-to-image diffusion models, with the potential for achieving state-of-the-art performance [21]. + +# 3. Deep Fusion of LLMs and DiTs + +# 3.1. Model Architecture + +In the deep fusion approach, we integrate a frozen decoder-only LLM with a trainable DiT using layer-wise shared self-attention (Fig. 1). The DiT mirrors the LLM's transformer architecture, differing only in its input/output layers and timestep conditioning modules. This implements a two-stream transformer architecture that facilitates rich cross-modal interactions while maintaining modality-specific computation by utilizing distinct weight sets for processing tokens from different modalities. + +The fused model processes text embeddings through the LLM stream and noisy image latents through the DiT stream. At each layer's self-attention operation, we concatenate token sequences from both streams, enabling the DiT to extract conditional information from the linguistic context. To preserve the pretrained LLM's functionality, we apply a causal attention mask to the text sequence and a bidirectional mask to the image sequence, permitting the image tokens to attend to text tokens but not vice versa (Fig. 2). After the final layer, we discard text tokens and use only image tokens to predict velocity, as typically done when training rectified flow models [23]. + +Notably, only the key and value states of the text hidden + +states are needed for self-attention. These remain constant throughout the diffusion process, allowing them to be efficiently cached and reused during inference. + +# 3.2. Training Objective + +We adopt the rectified flow formulation [23] to learn transport maps between the standard Gaussian noise distribution $\pi_0$ and the data distribution $\pi_1$ , by connecting straight paths $x_{t} = tx_{1} + (1 - t)x_{0}$ between samples $x_0\sim \pi_0,x_1\sim \pi_1$ and learning an ODE model $dz_{t} = v_{\theta}(z_{t},t)dt$ on time $t\in [0,1]$ which converts $z_{0}$ from $\pi_0$ to a $z_{1}$ following $\pi_1$ . + +We fit the velocity $v$ with $x_{1} - x_{0}$ with a prescribed time distribution $\pi_{t}$ by solving the following regression problem: + +$$ +\min _ {v} \mathbb {E} _ {t \sim \pi_ {t}, x _ {0} \sim \pi_ {0}, x _ {1} \sim \pi_ {1}} [ \| v (x _ {t}, t) - (x _ {1} - x _ {0}) \| ^ {2} ] d t \tag {1} +$$ + +We parameterize $v$ with network $\theta$ and solve Eq. (1) by stochastic optimization with empirical draws. + +# 4. Experiment Setup + +To ensure fully open and reproducible comparison between the deep and shallow fusion approaches, we provide comprehensive details on the experimental setup, including the model, dataset, training, inference, and evaluation. For the same purpose, we exclusively use open-sourced pre-trained LLMs and publicly available datasets. + +Model. We employ a frozen Gemma 2B [27] as the base LLM for all experiments (excluding Sec. 7). We pair it with a randomly initialized 2.5B-parameter DiT. The transformer configurations of the DiT strictly follows the base LLM, including the hidden size, number of layers, number of attention heads, FFN design, and other architectural details, ensuring both models have an identical 2B-parameter backbone. Following the vanilla DiT setup [29], we use 2D frequency absolute positional encoding, AdaLN-Zero timesteps conditioning, ViT [8]-style weight initialization, and a patch size of 2. To further stabilize training, we apply QK normalization to all layers. For all experiments we adopt the same 16-channel VAE from Stable Diffusion 3 (SD 3) [10]. + +Dataset. We use the CC12M [4] dataset with community-sourced synthetic captions [9] as our training set for all experiments excluding Sec. 7. Our downloaded version of the dataset includes 10.9M image-caption pairs. The images are resized and center-cropped to $512 \times 512$ and the texts are padded or truncated to 256 tokens. + +Training. We train all models with a batch size of 512 using AdamW [24] optimizer $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ in BF16 mixed precision. We use a constant learning rate of + +$1 \times 10^{-4}$ , a weight decay of $1 \times 10^{-4}$ , and gradient clipping with a threshold of 1.0. Exponential moving average of the weights are gathered by a decay factor of 0.99 every 100 steps. We employ the same logit-normal distribution as used in SD 3 for timestep sampling. During training, $10\%$ of the texts are randomly dropped to learn unconditional generation. Training is carried out using Google TPU v4-256 pods and FSDP implemented by PyTorch / XLA SPMD. + +Inference. We conduct inference using Euler discretization with 25 sampling steps, and a classifier-free guidance scale of 6 which we find to be near optimal for text-image alignment. We employ identical sampling steps and guidance scale across all experiments. + +Evaluation. We evaluate image-text alignment using GenEval [12] and DPG-Bench [15] metrics, prioritizing GenEval for its robustness. While both benchmarks provide valuable insights, DPG-Bench exhibits certain limitations, such as rapid performance saturation and potential measurement errors [21]. To ensure a comprehensive evaluation, we also provide visual quality measurements using FID [14] on MJHQ-30K [18]. Notably, image-text alignment does not always correlate positively with visual quality, often presenting trade-offs. Our sampling and evaluation are primarily carried out using NVIDIA L40S GPUs. + +# 5. Comparing Deep and Shallow Fusion + +The deep fusion approach fuses an LLM and a DiT through layer-wise shared self-attention, creating interconnections throughout the network. However, established architectures typically condition on representations from a single text encoder layer. To investigate the true potential of deep fusion, we conduct controlled comparisons with baseline methods. + +For a fair and meaningful comparison, we examine a common architectural paradigm which we refer to as shallow fusion. In this approach, representations from a single text encoder layer are integrated into each DiT layer through late fusion within the attention operation. Unlike deep fusion, which involves multi-layered interactions, shallow fusion maintains a fixed connection between each DiT layer and a prescribed text encoder layer (Fig. 1). + +# 5.1. Shallow Fusion Baselines + +We consider two shallow fusion architectures that condition on last-layer hidden states of LLMs as our baselines. As illustrated in Fig. 3, the two architecture differ in how they aggregate information from the condition. + +- Self-attention DiT. In this design, text representations are projected to key and value states and then concatenated with those of image hidden states in self-attention, which + +![](images/26fb8174e61731ae5a33672dea3ec54381e8ac694b6cb96d9be6721b3e5532e4.jpg) +(a) Self-Attention + +![](images/7d7637fd3db9bb8167e746807b8a70b2de942f3ba91772f80d5f7573c1b08c51.jpg) +(b) Cross-Attention +Figure 3. Illustration of cross-modal attention in the shallow fusion baselines. The key and query states of the condition are directly projected from text representations. + +can also be decoupled by running self-attention and cross-attention in parallel and merging their outputs. This approach resembles architectures proposed by [11, 51]. + +- Cross-attention DiT. This design also projects text representations to key and values states. However, unlike the previous approach, they are used for additional cross-attention with image hidden states, applied after the self-attention in each layer. This architecture follows the methodology employed in [5, 6, 47]. + +Compared to the deep fusion approach, both baseline models include RMS normalization and a linear layer for text representations before passing them through additional key and value projection layers. Additionally, the cross-attention DiT model uses extra query projections in its cross-attention mechanism for image hidden states. Other model configurations follow the DiT design in Sec. 4. + +Notably, the deep fusion approach can be reinterpreted as a variant of the self-attention DiT architecture, as they both aggregate conditional information through in-context self-attention. The key difference lies in how the key and value states of the condition are derived: self-attention DiT employs a trainable projection to generate these states from a single text encoder layer, whereas deep fusion extracts them from corresponding LLM layers. Despite their similarities, the two approaches have fundamentally different conceptual implications: deep fusion treats the LLM and DiT as equal components of a unified model. + +Apart from the aforementioned designs, SD3 [10] introduced an alternative fusion strategy that uses a two-stream transformer (MM-DiT) to jointly processes noised image latents and linguistic representations. While this is another popular approach worthy of investigation, an apples-to-apples fair comparison between the deep fusion approach and MM-DiT is not feasible, as both streams in MM-DiT are trainable. Consequently, our analysis focuses solely on the shallow fusion baselines detailed in this section. + +# 5.2. Controlled Comparison + +For a fair comparison, we design self-attention DiT, cross-attention DiT, and the deep fusion model with similar architectures, and train them for 300K steps following the setup detailed in Sec. 4. + +![](images/4ad6daea267291d9d27033456f39ebb502b3aa1d89cba3a628c6a0fa2e9605c3.jpg) + +![](images/69b9d2ebe405c405dc29059100fd208a75a1ddcdb93ead233685eee93f343571.jpg) +Figure 4. Illustration of timestep conditioning strategies. Removing timesetp conditioning leads to the fewest parameters and the best overall performance. + +![](images/17652829af896ca34a07b42ef4554cbbaf992c28c08a29ac8e203faf660a732e.jpg) + +![](images/af6cc08e1cbca3253e4642d2b0f6589d81c71a372713008802f73974d59841ae.jpg) + +As shown in Tab. 1, the deep fusion model achieve significantly better performance in image-text alignment than the self-attention DiT model and also surpass the cross-attention DiT model, while shallow fusion models demonstrate better visual quality. In terms of inference efficiency, deep fusion also demonstrates competitive performance, as shown in Tab. 2. This positive evidence underscores the compelling positioning of the deep fusion approach within the current landscape. + +
MethodParams.GenEval ↑DPG ↑FID ↓
Self-Attention2.47B0.4273.926.16
Cross-Attention2.62B0.4976.324.00
Deep Fusion2.45B0.5176.627.33
+ +Table 1. Comparison of performance between deep and shallow fusion models. Deep fusion beats shallow fusion in text-image alignment while underperforms in visual quality. + +
MethodParams.Inference Latency (s)
Self-Attention2.47B1.75
Cross-Attention2.62B1.86
Deep Fusion2.45B1.66
+ +Table 2. Comparison of inference latency between deep and shallow fusion models. The numbers are measured with a batch size of 1 in automatic mixed precision on an NVIDIA A100 GPU. + +# 6. Examining Key Design Choices + +In this section, we examine key design choices of the deep fusion approach through a text-to-image-centric lens. We begin by assessing the necessity and potential redundancy of parameters for timestep conditioning, determining whether to optimize their use or eliminate them altogether (Sec. 6.1). Next, we compare various positional encoding strategies (Sec. 6.2). Additionally, we investigate how the choice of base LLM and the use of instruction prompts impacts text-to-image performance (Sec. 6.3). + +For all experiments, we use the default design as the baseline and train the models for 300K steps, following the setup detailed in Sec. 4. + +# 6.1. Timestep Conditioning + +DiT [29] has introduced AdaLN-Zero as the standard mechanism for injecting timestep and class label information. The AdaLN modules typically accounts for a large proportion of the model parameters, 0.5B out of a total 2.5B in our + +case. However, since our text-to-image model does not use class labels, AdaLN serves only for timestep conditioning. This substantial parameter allocation raises a critical question about its necessity and potential redundancy: Could these parameters be utilized more effectively, or are they redundant altogether? + +Question 1.1. Can adaLN parameters be utilized more effectively by integrating additional text conditioning? + +Prior research [10, 11, 17] has integrated text information into AdaLN-Zero by augmenting timestep embeddings with pooled text representations through summation. To optimize parameter efficiency, we follow this approach by leveraging embeddings1 from the CLIP L/14 text encoder [31], which provides high-quality linguistic representations trained on large-scale multi-modal data. + +
MethodGenEval ↑DPG ↑FID ↓
adaLN-Zero0.5176.627.33
+ CLIP L/140.5076.224.00
+ +The results indicate that integrating text modulation results in slight improvements in FID but weakens image-text alignment. Additionally, it further increases compute. + +Question 1.2. Can we shrink the parameters for timestep conditioning? + +We explore the possibility of eliminating timestep conditioning parameters to develop a more streamlined architecture akin to that of the LLM. Specifically, we compare four timestep conditioning strategies from previous work, each with a progressively reduced number of parameters, as illustrated in Fig. 4. + +- adaLN-Zero. Following the vanilla DiT [29], we regress zero-initialized modulation parameters for each layer from the timestep embedding. +- adaLN-Single. Following PixArt- $\alpha$ [6], we compute a global set of modulation parameters and refine them per layer by adding learnable embeddings. +- Addition. Following Transfusion [50], we directly add the timestep embedding to all image tokens. +- w/o Timestep. Inspired by a recent study [42], we completely remove the timestep conditioning from the model. + +
MethodParams.GenEval ↑DPG ↑FID ↓
adaLN-Zero2.47B0.5176.627.33
adaLN-Single2.01B0.4775.227.09
addition1.99B0.4775.626.40
w/o timestep1.98B0.4976.721.27
+ +Surprisingly, the results in Tab. 3 indicate that reducing the number of parameters in timestep conditioning consistently enhances visual quality, whereas the performance in image-text alignment exhibits fluctuations. + +Table 3. Evaluation of timestep conditioning strategies. Removing timestep conditioning yields surprisingly strong results. + +
Input SequenceTXT 1TXT 2[PAD]IMG 1IMG 2IMG 3IMG 4
1D-RoPE + APE[0][1]—[2D Frequency APE]
1D-RoPE[0][1]—[2][3][4][5]
1D + 2D-RoPE[0][1]—[2, 2][2, 3][3, 2][3, 3]
M-RoPE[0, 0][1, 1]—[2, 2][2, 3][3, 2][3, 3]
+ +Notably, the strategy that completely removes timestep conditioning not only achieves significantly better FID but also maintains comparable GenEval and DPG-Bench performance. This finding is consistent with [42], where timestep conditioning removal improved FID in rectified flow models trained on smaller datasets. Furthermore, the complete removal of timestep conditioning eliminates the need for associated parameters, resulting in a $20\%$ reduction in the total number of model parameters. Although this approach slightly lags behind AdaLN-Zero in terms of text-image alignment metrics, we prefer it due to its parameter efficiency and architectural simplicity. + +# 6.2. Positional Encoding + +Absolute positional encoding (APE) is widely used in text-to-image diffusion models, whereas in the context of LLMs, rotary positional embedding (RoPE) [40] is the predominant choice. As deep fusion models are inherently multimodal and differ from traditional text-to-image diffusion models, it is unclear which positional encoding (or their combinations) is best suited for mixed-modal sequences. + +Question 2.1. Is RoPE more advantageous than APE for enhancing the performance of deep fusion models? + +To explore this, we experiment with three positional encoding methods, as illustrated in Fig. 5: + +- 1D RoPE + APE: We apply 1D RoPE to the text sequence and APE to the image sequence respectively. +- 1D RoPE: We extend 1D RoPE to encompass both text and image sequences. +- 1D + 2D RoPE: We apply 1D RoPE to the text sequence and 2D RoPE to the image sequence respectively. + +Figure 5. Illustration of RoPE. The indices denote position IDs. + +
MethodGenEval ↑DPG ↑FID ↓
1D-RoPE + APE0.5176.627.33
1D-RoPE0.4677.027.94
1D + 2D-RoPE0.5176.425.42
+ +Table 4. Comparing different positional encoding strategies. $1\mathrm{D} + 2\mathrm{D}$ RoPE achieves the best overall performance. + +As shown in Tab. 4, the 1D + 2D-RoPE configuration achieves the best overall performance, with only a marginal decrease in DPGBench compared to the 1D-RoPE + APE variant. The superiority of 2D-RoPE over APE suggests RoPE is more effective for modeling image sequences in deep fusion models. Using only 1D-RoPE slightly reduces performance, indicating that while deep fusion models treat text and image sequences as a unified input, their distinct positional characteristics are best modeled separately. + +# Question 2.2. Does deep fusion models benefit from RoPE specifically designed for mixed-modal sequences? + +Previous work on MLLMs has explored RoPE strategies for handling mixed-modal sequences. Naturally, we are curious whether deep fusion models can benefit from these positional encodings. Follow Qwen2-VL, we implement M-RoPE, a variant of RoPE that applies 2D positional IDs to chunked 1D RoPE frequencies, allowing it to function as 1D RoPE for text sequences while approximating 2D RoPE for image sequences. + +
MethodGenEval ↑DPG ↑FID ↓
1D + 2D-RoPE0.5176.425.42
M-RoPE0.4974.927.60
+ +Although M-RoPE elegantly unifies 1D and 2D-RoPEs, it still falls short compared to their direct combination. This underscores the challenge of designing position encodings for mixed-modal sequences. + +# 6.3. Base LLM + +LLMs trained with different paradigms and data demonstrate diverse capabilities and behaviors. In this section, we examine how the choice of base LLM impacts the performance of deep fusion models. + +Question 3.1 Can instruction tuning, combined with instruction prompts, improve text-to-image performance? + +Instruction tuning enables LLMs to effectively follow complex instructions. We explore whether this process can also enhance their internal information flow, leading to more contextualized and discriminative representations for text-to-image synthesis. We compare Gemma 2B with Gemma 2B IT, its instruction-tuned variant. We also experiment with using it with a simple instruction prompt, "Imagine: ", which we find sufficient for guiding the LLM to generate detailed and relevant expansions of the input. + +
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ instruction tuning0.4975.427.04
+ instruction prompt0.5075.825.28
+ +As shown in Tab. 5, instruction tuning appears to have a slightly negative impact on performance. While the use of an instruction prompt mitigates this effect to some extent, consistent with findings in [25, 47], it still falls short of the baseline. This result highlights a challenge in effectively leveraging the instruction-following capabilities of LLMs. + +Question 3.2 Can multi-modal tuning improve text-to-image performance? + +Additionally, we're also interested in the effect of multi-modal tuning. While multi-modal tuning differs significantly from our setup, its shift in data distribution may still potentially enhance adaptability to multi-modal tasks. We + +compare Gemma 2B with the base LLM of PaliGemma 3B PT [3], a multi-modal extension of Gemma 2B that undergoes additional pretraining on image-text data. + +Table 5. Evaluating the effect of instruction tuning. Using instruction-tuned LLMs does not improve performance. + +
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ multi-modal tuning0.5276.226.30
+ +which yields small improvements in performance. This observation indicates that multi-modal finetuning provides some benefit to the deep-fusion model. + +Question 3.3 Does improved LLM capabilities translate to stronger text-to-image performance? + +Finally, as the base LLM become more proficient in understanding and generating text, they could potentially foster synergistic improvements in DiT performance. We compare Gemma 2B with Gemma 2 2B [34], the next generation of Gemma 2B which demonstrates a $6\%$ absolute performance improvement $(0.44 \rightarrow 0.50)$ on an average of 8 language-only benchmarks. Notably, Gemma 2 2B features a different transformer architecture than Gemma 2B, prompting us to adjust the DiT architecture accordingly. This modification increases the number of adaLN parameters by 0.3B. However, as demonstrated in Sec. 6.1, these parameters do not contribute to model performance. + +
ModelGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
Gemma 2 2B0.5479.123.94
+ +Upgrading from Gemma 2B to Gemma 2 2B yields a drastic performance boost. This finding suggests that the DiT's performance in deep fusion models is strongly dependent on the capabilities of the underlying base LLM. + +# 7. Training at Scale + +In this section, we present a final recipe for the deep fusion model, building on the original framework while incorporating key insights from previous exploration. We conduct large-scale training to benchmark our model against established systems, showcasing its scalability and competitive performance on the leaderboard. + +# 7.1. Final Recipe + +Building on the insights from Sec. 6, we introduce the following design modifications to our model: + +- Remove AdaLN-Zero modules. +- Replace 1D-RoPE + APE with 1D + 2D-RoPE. +- Replace Gemma 2B with Gemma 2 2B, adjusting the transformer configurations of the DiT accordingly. + +We train our model for 800K steps on a mixed dataset comprising CC12M [4], SA-1B [16], and the training subset of JourneyDB [41], amounting to approximately 26M image-caption pairs. Notably, state-of-the-art text-to-image models typically rely on high-quality datasets of much larger scale to achieve superior performance. For CC12M + +![](images/151612be0ce56445f20a9dcae542bed4e3577b9ed0541072b02516f8e5f18734.jpg) +(a) Fully Parallel +Figure 6. Illustration of architecture alignment. Dashed boxes indicate parameters that have been reduced by decreasing either the hidden size or the number of layers. + +![](images/cfae5e7ebb34949aa90ffbd1cdfa03240bc612a3d957b161f39361aadc715f26.jpg) +(b) Skipped Layers + +![](images/ea62c125e010eff6f16c791f75ab09bd4575bff41bfa09b4f2fcfe0c71831853.jpg) +(c) Decoupled Dimension + +and SA-1B, we utilize synthetic captions [6, 9]. Other experimental setup follows Sec. 4. + +# 7.2. Performance Comparison + +
ModelParamsDataGen.↑DPG↑FID↓
SD 1.5 [35]0.9B4.8B0.4363.2—
DALL-E 2 [33]4.2B2.6B0.52——
SDXL [30]2.6B1.6B0.5574.76.63
PG 2.5 [18]2.6B—0.5675.56.09
SD 3 M [10]2B1B0.6284.111.92
DALL-E 3 [2]——0.6783.5—
FLUX.1 [dev] [17]12B—0.6784.010.15
PG 3 [21]24B—0.7687.0—
MicroDiT [37]1.2B37M0.46——
PixArt-α [6]0.6B25M0.4871.1—
Lumina Next [11]2B—0.4674.67.58
PixArt-Σ [5]0.6B46M0.5480.56.15
Transfusion [50]7.3B3.5B0.63——
Sana 1.0 1.6B [47]1.6B—0.6684.85.76
Ours2B26M0.6081.67.54
+ +We compare our model with the most advanced text-to-image diffusion models in Tab. 6. Despite being trained with limited compute and data in a simplified setting, our model surpasses many industry-standard systems and delivers competitive results. + +We present qualitative examples from our model in Figure 7. Our model demonstrates the ability to generate high-quality images with superior prompt alignment. + +# 8. Further Exploration + +In this section, we present preliminary studies exploring more aggressive modifications to the deep fusion approach. For all experiments, we train the models for 300K steps, following the setup detailed in Sec. 4. + +# 8.1. Architecture Alignment + +Up to this point, our approach has followed prior work by aligning the LLM and DiT backbones in a layer-by-layer + +approach, strictly enforcing identical transformer configurations for both models. However, this rigid constraint limits the flexibility of deep fusion. In practice, we need the ability to scale the LLM and DiT independently, as different modalities follow distinct scaling laws and network design principles, and training and deployment scenarios vary. + +To address this, we explore modifying the DiT model's hidden size and number of transformer layers, as illustrated in Fig. 6. The adapted model is fused into the middle layers of the LLM, which contain richer semantic information [39]. Additionally, in self-attention, hidden states are still projected to query, key, and value states that match the LLM's dimensionality, ensuring compatibility. Since the most successful DiTs are generally much smaller than state-of-the-art LLMs, we focus on shrinking the size of our DiT. + +Table 6. Comparison of performance with state-of-the-art systems. Table adapted from [48, 50]. Industrial baselines are presented at the top, while academic baselines and our model are listed at the bottom. + +
Hidden sizeParams.GenEval ↑DPG ↑FID ↓
20482.5B0.5176.627.33
17922.1B0.5077.124.27
15361.8B0.4976.225.46
12801.4B0.4874.824.64
+ +Table 7. Evaluating models of different hidden sizes. The default hidden size is 2048. + +
LayersParams.GenEval ↑DPG ↑FID ↓
182.5B0.5176.627.33
141.9B0.4774.623.46
101.4B0.3368.028.34
+ +Table 8. Evaluating different numbers of layers. The default number of layer is 18. + +As shown in Table Tab. 7, the model's performance degrades gracefully as we reduce the hidden size, with visual quality actually improving in some cases. While decreasing the number of transformer layers (Tab. 8) also yields acceptable results, performance deteriorates more quickly. We hypothesize this occurs because Gemma 2B already employs fewer layers than typical model architectures. These findings suggest that LLM and DiT model designs can be effectively decoupled, enabling the application of separate scaling laws and design principles for each component. + +# 8.2. Attention Mechanism + +In Sec. 3, we build on prior work by defining the deep fusion architecture with shared self-attention to bridge the LLM and DiT. Inspired by extensive research on cross-attention in MLLMs [1, 13] and our findings in Sec. 5, which highlight the superior performance of cross-attention DiT over self-attention DiT, we explore an alternative deep fusion variant. This new approach replaces shared self-attention with cross-attention mechanisms, similar to cross-attention DiT but with a key distinction: we substitute the projected linguistic key and value states in traditional cross-attention DiT with corresponding states from LLM layers. + +![](images/f046d53be3e87e89c215ac852dffa20cc94ed407976b407fe95dd2b052df5582.jpg) +Dew on blue rose petals, HD, close up, detail + +![](images/273f3bb4e7aee303f6939030bf3ed037185f825f6eb8f3ef41a7678d6f8dc403.jpg) +A corgi wearing sunglasses walks on the beach of a tropical island + +![](images/dfe7a970366132155da3fd780030eca52ffc88ae64bfd7a3263d04105a80115f.jpg) +A rally car taking a fast turn on a track + +![](images/cd47c8b9bc51ebe7a21cefd83c4e74962037de8b6b2fa7546420c68462475560.jpg) +An armchair in the shape of an avocado + +![](images/28e5c9516d6dd5f4a1afd11cbf5efd2a7820a07cfde5c41cdd31dc20b8c1f789.jpg) +A man riding a horse through the Gobi Desert with a beautiful sunset behind him, movie quality. + +![](images/e8e8b5e639fad0c7071a397ea7b1941c2ce13106225d94f9562a34e10203f4e7.jpg) +Pirate ship trapped in a cosmic maelstrom nebula + +![](images/18705d9ef123edb53eb534838fe38b5ac85508776d8fb05d7cdf78373ab50ddd.jpg) +Astronaut in a jungle, cold color palette, muted colors, detailed, 8k + +![](images/5bac32400729d933a1bf4611893eda1268d340793d61fcf4b90e94ac8cc7e1bb.jpg) +A person surfing on a wave of stars in outer space. + +![](images/de8a836baaf5eedb97e37dbc2a581bd03d26316c737bc1af27ab0268806ff3f7.jpg) +a picturesque autumn scene where a quaint cottage with a thatched roof sits beside a tranquil lake, surrounded by trees with leaves in vibrant shades of orange, red, and yellow. The cottage's wooden exterior is complemented by white-framed windows, and a stone chimney rises above the roofline. The lake reflects the warm fall colors, creating a mirror image of the foliage and the small structure on its calm surface. + +![](images/3e1fb7678a63b70f38190cabfd67b47f9d74ede94322dab8314a643b32a52fe7.jpg) +A portrait of a human growing colorful flowers from her hair. Hyperrealistic oil painting. Intricate details. + +![](images/084ba814fbc087a2d54f5940196c570a463401e793865aa3382705d6b3d6fd4b.jpg) +a black and white picture of a woman looking through the window, in the style of Duffy Sheridan, Anna Razumovskaya, smooth and shiny, wavy, Patrick Demarchelier, album covers, lush and detailed +13 year old farm boy, red tousled hair, green eyes, soft pale skin, freckles, stub nose, very beautiful, very pretty, dark fantastic medieval + +![](images/db084e5ab4541cd00906c718f8d7893db1d011a0033645b3e5a12b5b0d59d5ef.jpg) +Figure 7. Samples generated by our final recipe. + +
MethodGenEval ↑DPG ↑FID ↓
self-attention0.5176.627.33
cross-attention0.5276.526.57
+ +This modification yields minor gains, though at a cost to the LLM-DiT parity. Additionally, we find that although cross-attention introduces a negligible increase in FLOPs and parameter count, it leads to approximately a $12\%$ increase in latency $^{2}$ (1.66s vs. 1.86s). Therefore, we retained the self-attention design in our final configuration. + +# 9. Conclusion + +We have studied the recently popular deep fusion of a frozen LLM with a trainable DiT for text-to-image synthesis. Our findings provide empirical evidence supporting its advantages over baselines. We highlight key design choices, identify unresolved problems, and offer meaningful data points alongside practical guidelines. We hope our empirical work help advance multi-modal generation and bridge the gap between auto-regressive decoding and denoising diffusion. + +# Acknowledgments + +We thank the reviewers for their constructive feedback. We also thank Shusheng Yang, Shengbang Tong, Wenhao Chai, Nanye Ma, Sihan Xu, and Chenyu Li for insightful discussions. This work was mainly supported by the Google TPU Research Cloud (TRC) program, and the Google Cloud Research Credits program (GCP19980904) and Open Path AI Foundation. SX also acknowledges support from Intel AI SRS, IITP grant funded by the Korean Government (MSIT) (RS-2024-00457882, National AI Research Lab Project), Amazon Research Award, and NSF Award IIS-2443404. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 7 +[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.opencai.com/papers/dall-e-3.pdf, 2023.1,2,7 +[3] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv:2407.07726, 2024. 6 +[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 3, 6 +[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024.1,2,4,7 +[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\alpha$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 7 +[7] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[9] Caption Emporium. conceptual-captions-cc12m-llavanext. https://huggingface.co/datasets/CaptionEmporium / conceptual - captions - cc12m-llavanext, 2024.3,7 + +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 2, 3, 4, 5, 7 +[11] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv:2405.05945, 2024. 2, 4, 5, 7 +[12] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. arXiv:2310.11513, 2023. 3 +[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 7 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 3 +[15] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv:2403.05135, 2024. 1, 2, 3 +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6 +[17] Black Forest Labs. Announcing black forest labs. https://blackforestlabs.ai/announcing-black-forest-labs/, 2024. 1, 2, 5, 7 +[18] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv:2402.17245, 2024. 2, 3, 7 +[19] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv:2305.13655, 2023. 2 +[20] Weixin Liang, Lili Yu, Liang Luo, Srinivasan Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, et al. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. arXiv:2411.04996, 2024. 1, 2 +[21] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv:2409.10695, 2024. 1, 2, 3, 7 +[22] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv:2407.00737, 2024. 1, 2 +[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv:2209.03003, 2022. 2, 3 + +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 3 +[25] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv:2406.11831, 2024. 1, 2, 6 +[26] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv:2411.07975, 2024. 1, 2 +[27] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv:2403.08295, 2024. 3 +[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 1, 2 +[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1, 2, 3, 4, 5 +[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1, 2, 7 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5 +[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In JMLR, 2020. 1, 2 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 7 +[34] Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv:2408.00118, 2024. 6 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 7 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2 +[37] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv:2407.15811, 2024. 7 +[38] Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, and Lili Yu. Lmfusion: + +Adapting pretrained language models for multimodal generation. arXiv:2412.15188, 2024. 1, 2 +[39] Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models. arXiv:2502.02013, 2025. 7 +[40] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 5 +[41] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. In NeurIPS, 2024. 6 +[42] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv:2502.13129, 2025. 5 +[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1 +[44] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv:2401.15688, 2024. 2 +[45] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2 +[46] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv:2409.11340, 2024. 1, 2 +[47] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv:2410.10629, 2024. 1, 2, 4, 6, 7 +[48] Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv:2501.18427, 2025. 1, 7 +[49] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2 +[50] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv:2408.11039, 2024. 1, 2, 5, 7 +[51] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv:2406.18583, 2024. 1, 2, 4 \ No newline at end of file diff --git a/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/images.zip b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dcb1ae70f340d9f60c4d235b501da4f33180b460 --- /dev/null +++ b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f978cdc223ddfad0d63369f2cb34c15dee3b44cd114f33f349cad88601fc187a +size 522240 diff --git a/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/layout.json b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c28a631c1fc8d8e8c8cd4aa99e5949839cabf0aa --- /dev/null +++ b/2025/Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis/layout.json @@ -0,0 +1,9806 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "text", + "content": "Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "spans": [ + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "text", + "content": "Bingda Tang " + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "text", + "content": " Boyang Zheng " + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "text", + "content": " Sayak Paul " + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "text", + "content": " Saining Xie " + }, + { + "bbox": [ + 144, + 160, + 465, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 180, + 406, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 180, + 406, + 196 + ], + "spans": [ + { + "bbox": [ + 204, + 180, + 406, + 196 + ], + "type": "text", + "content": "1 New York University 2 Hugging Face" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 222, + 200, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 222, + 200, + 235 + ], + "spans": [ + { + "bbox": [ + 151, + 222, + 200, + 235 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 248, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 440 + ], + "type": "text", + "content": "This paper does not describe a new method; instead, it provides a thorough exploration of an important yet understudied design space related to recent advances in text-to-image synthesis—specifically, the deep fusion of large language models (LLMs) with diffusion transformers (DiTs) for multimodal generation. Previous studies mainly focused on overall system performance rather than detailed comparisons with alternative methods, and key design details and training recipes were often left undisclosed. These gaps create uncertainty about the real potential of this approach. To fill these gaps, we conduct an empirical study on text-to-image generation, performing controlled comparisons with established baselines, analyzing important design choices, and providing a clear, reproducible recipe for training at scale. We hope this work offers meaningful data points and practical guidelines for future research in multimodal generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 135, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 135, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 135, + 477 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 485, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 652 + ], + "type": "text", + "content": "Text-to-image diffusion models have made remarkable progress in generating high-quality images from descriptive texts. Current state-of-the-art systems [2, 5, 10, 17, 30] typically derive text representations from specialized encoders, such as CLIP [31] and T5 [32]. With the rise of decoder-only large language models (LLMs), there has been a growing amount of interest in their potential as replacements for these traditional text encoders [7, 15, 21, 22, 25, 28, 48, 51]. However, simply substituting LLMs has not yielded expected performance gains unless coupled with sophisticated architectural adaptations [15, 25, 47]. Prior work [25] attributes this to the misalignment between the next token prediction training objective of LLMs and the need for discriminative text representations in diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Recent advancements [20, 26, 46, 50] have successfully unified auto-regressive decoding and denoising diffusion within a single transformer [43], enabling seamless multi-modal generation. This unified approach supports a wide range of tasks, including instructed image-to-text," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 221, + 436, + 304 + ], + "blocks": [ + { + "bbox": [ + 317, + 221, + 436, + 304 + ], + "lines": [ + { + "bbox": [ + 317, + 221, + 436, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 221, + 436, + 304 + ], + "type": "image", + "image_path": "2990b2bdacd0d68abdaa088ddd2b9dcb6414fa2edebd3e18de8f202c2d3f0c1d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 307, + 403, + 314 + ], + "lines": [ + { + "bbox": [ + 349, + 307, + 403, + 314 + ], + "spans": [ + { + "bbox": [ + 349, + 307, + 403, + 314 + ], + "type": "text", + "content": "(a) Shallow-Fusion" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 221, + 548, + 304 + ], + "blocks": [ + { + "bbox": [ + 440, + 221, + 548, + 304 + ], + "lines": [ + { + "bbox": [ + 440, + 221, + 548, + 304 + ], + "spans": [ + { + "bbox": [ + 440, + 221, + 548, + 304 + ], + "type": "image", + "image_path": "afc01f31583e83df2d8990efbdc392754d30cf73a6ce7972ba37023b9146e5f9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 307, + 517, + 315 + ], + "lines": [ + { + "bbox": [ + 471, + 307, + 517, + 315 + ], + "spans": [ + { + "bbox": [ + 471, + 307, + 517, + 315 + ], + "type": "text", + "content": "(b) Deep-Fusion" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "text", + "content": "Figure 1. Illustration of the deep fusion approach and baselines. We conduct controlled comparisons with baseline methods that incorporate text representations from a single text encoder layer into each DiT layer using late fusion within the attention mechanism, a strategy we term the shallow fusion approach." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 388, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 508 + ], + "type": "text", + "content": "text-to-image synthesis, and interleaved image-text generation. While earlier methods often relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pre-training of LLMs by deeply fusing them with diffusion transformers (DiTs) [29] through layer-wise shared self-attention. This design facilitates rich cross-modal interactions while maintaining modality-specific computation by using distinct sets of weights. When optimized for text-to-image generation, it claims state-of-the-art performance [21]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 509, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 689 + ], + "type": "text", + "content": "Deep fusion presents a compelling alternative to existing architectures for text-to-image synthesis, which typically conditions directly on representations from a single text encoder layer. By aligning diffusion models with the auto-regressive decoding nature of LLMs, deep fusion enables a more natural and tight-knit use of these models. However, despite the existing positive signals, its true potential remains uncertain. Current research [21, 38] prioritizes system-level benchmarks over controlled comparisons with established baselines, obscuring its position within the broader research landscape. More critically, the design space remains severely underexplored, and essential implementation details, such as training recipes, are often undisclosed. These limitations impede reproducibility and hinder broader adoption within the research community." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we bridge these gaps through an empirical study on the deep fusion of a frozen LLM and a train" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28586" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "able DiT for text-to-image synthesis. We conduct controlled comparisons between the deep fusion approach and baselines, examine key design choices, and introduce a scalable, reproducible training recipe that delivers competitive performance on the established benchmarks for text-to-image generation. We believe that the evidence and unresolved questions highlighted in this study are of significant importance. We hope this work serves as a valuable resource, offering meaningful data points and practical guidelines to drive future advancements in multi-modal generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 142, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 142, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 142, + 213 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 295, + 304 + ], + "type": "text", + "content": "Conditioning mechanisms in diffusion models. Numerous studies have examined effective methods for integrating linguistic conditional information into diffusion models to facilitate text-to-image synthesis. Latent diffusion models (LDM) [35] pioneered the use of cross-attention between image and text features, a technique that has since become standard in U-Net [36] based architectures [18, 30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 305, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 533 + ], + "type": "text", + "content": "With the rise of DiTs [29], vision transformer [8] has emerged as the dominant architecture for diffusion models, prompting a reassessment of conditioning mechanisms. The vanilla DiT employs adaLN-Zero modulation to inject conditional information. However, this approach is limited to using pooled text representations, which capture only coarse-grained information. Subsequent architectures have explored cross-attention [5, 6, 47] and self-attention [10, 11, 17, 51] for text conditioning, which typically extract representations from a single text encoder layer, usually the last or penultimate one. While this strategy aligns well with CLIP [31] and T5 [32] text encoders, it is inherently mismatched with the next-token prediction training objective of LLMs, where the last layer focuses on next-token prediction instead of learning discriminative representation. In contrast, the deep fusion approach shows promise in harnessing the internal information flow of LLMs, aligning with their in-context self-attention mechanism for processing information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 676 + ], + "type": "text", + "content": "Taming LLMs for diffusion models. The prevalence of decoder-only LLMs has driven extensive efforts to tame them for text-to-image diffusion models. The most effective and widely adopted practice involves leveraging LLMs to enrich input prompts [2], as well as employing multimodal LLMs (MLLMs) to generate synthetic captions for image data [2, 5, 10, 21, 47]. Another popular direction integrates (M)LLMs as system components, such as planners and discriminators [19, 44, 45, 49]. While these approaches have demonstrated effectiveness, they do not tap into the architecture of diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To integrate LLMs into diffusion models, previous works attempt to replace text encoders with LLMs, either by training from scratch [25, 47, 51] or by aligning feature" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 364, + 70, + 504, + 210 + ], + "blocks": [ + { + "bbox": [ + 364, + 70, + 504, + 210 + ], + "lines": [ + { + "bbox": [ + 364, + 70, + 504, + 210 + ], + "spans": [ + { + "bbox": [ + 364, + 70, + 504, + 210 + ], + "type": "image", + "image_path": "58bec034de58bec47fbb3d5554dd44df984efe3fe8d58554c69688dd07471973.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "type": "text", + "content": "Figure 2. Illustration of the attention mask. Each dotted square indicates whether the row can attend to the column." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "type": "text", + "content": "spaces [7, 15, 22, 28]. However, this substitution alone has not yielded the expected performance gains unless paired with sophisticated architectural adaptations [15, 25, 47]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 403 + ], + "type": "text", + "content": "Recent research [20, 26, 46, 50] has proven that autoregressive decoding and denoising diffusion can be effectively unified within a single transformer for multi-modal generation. While earlier methods relied on large-scale pretraining of the entire model, latest research [21, 38] flexibly leverages the computationally intensive pretraining of LLMs by deeply fusing them with DiTs. From a text-to-image perspective, this architecture introduces a novel way to tame LLMs for text-to-image diffusion models, with the potential for achieving state-of-the-art performance [21]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 491, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 491, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 491, + 426 + ], + "type": "text", + "content": "3. Deep Fusion of LLMs and DiTs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 433, + 429, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 429, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 429, + 444 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 557 + ], + "type": "text", + "content": "In the deep fusion approach, we integrate a frozen decoder-only LLM with a trainable DiT using layer-wise shared self-attention (Fig. 1). The DiT mirrors the LLM's transformer architecture, differing only in its input/output layers and timestep conditioning modules. This implements a two-stream transformer architecture that facilitates rich cross-modal interactions while maintaining modality-specific computation by utilizing distinct weight sets for processing tokens from different modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 700 + ], + "type": "text", + "content": "The fused model processes text embeddings through the LLM stream and noisy image latents through the DiT stream. At each layer's self-attention operation, we concatenate token sequences from both streams, enabling the DiT to extract conditional information from the linguistic context. To preserve the pretrained LLM's functionality, we apply a causal attention mask to the text sequence and a bidirectional mask to the image sequence, permitting the image tokens to attend to text tokens but not vice versa (Fig. 2). After the final layer, we discard text tokens and use only image tokens to predict velocity, as typically done when training rectified flow models [23]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Notably, only the key and value states of the text hidden" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28587" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "states are needed for self-attention. These remain constant throughout the diffusion process, allowing them to be efficiently cached and reused during inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 116, + 167, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 167, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 167, + 129 + ], + "type": "text", + "content": "3.2. Training Objective" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": "We adopt the rectified flow formulation [23] to learn transport maps between the standard Gaussian noise distribution " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " and the data distribution " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "\\pi_1" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": ", by connecting straight paths " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "x_{t} = tx_{1} + (1 - t)x_{0}" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " between samples " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "x_0\\sim \\pi_0,x_1\\sim \\pi_1" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " and learning an ODE model " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "dz_{t} = v_{\\theta}(z_{t},t)dt" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " on time " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " which converts " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "z_{1}" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": " following " + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "inline_equation", + "content": "\\pi_1" + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "content": "We fit the velocity " + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "inline_equation", + "content": "x_{1} - x_{0}" + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "content": " with a prescribed time distribution " + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "inline_equation", + "content": "\\pi_{t}" + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "content": " by solving the following regression problem:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 247, + 295, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 247, + 295, + 265 + ], + "spans": [ + { + "bbox": [ + 65, + 247, + 295, + 265 + ], + "type": "interline_equation", + "content": "\\min _ {v} \\mathbb {E} _ {t \\sim \\pi_ {t}, x _ {0} \\sim \\pi_ {0}, x _ {1} \\sim \\pi_ {1}} [ \\| v (x _ {t}, t) - (x _ {1} - x _ {0}) \\| ^ {2} ] d t \\tag {1}", + "image_path": "a16a90aa44929ad17ec1b42101bc3212210c07f0a791a13af858bf2e7a429a67.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "text", + "content": "We parameterize " + }, + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "text", + "content": " with network " + }, + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "text", + "content": " and solve Eq. (1) by stochastic optimization with empirical draws." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 309, + 164, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 164, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 164, + 323 + ], + "type": "text", + "content": "4. Experiment Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": "To ensure fully open and reproducible comparison between the deep and shallow fusion approaches, we provide comprehensive details on the experimental setup, including the model, dataset, training, inference, and evaluation. For the same purpose, we exclusively use open-sourced pre-trained LLMs and publicly available datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 418, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 574 + ], + "type": "text", + "content": "Model. We employ a frozen Gemma 2B [27] as the base LLM for all experiments (excluding Sec. 7). We pair it with a randomly initialized 2.5B-parameter DiT. The transformer configurations of the DiT strictly follows the base LLM, including the hidden size, number of layers, number of attention heads, FFN design, and other architectural details, ensuring both models have an identical 2B-parameter backbone. Following the vanilla DiT setup [29], we use 2D frequency absolute positional encoding, AdaLN-Zero timesteps conditioning, ViT [8]-style weight initialization, and a patch size of 2. To further stabilize training, we apply QK normalization to all layers. For all experiments we adopt the same 16-channel VAE from Stable Diffusion 3 (SD 3) [10]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "text", + "content": "Dataset. We use the CC12M [4] dataset with community-sourced synthetic captions [9] as our training set for all experiments excluding Sec. 7. Our downloaded version of the dataset includes 10.9M image-caption pairs. The images are resized and center-cropped to " + }, + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "text", + "content": " and the texts are padded or truncated to 256 tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Training. We train all models with a batch size of 512 using AdamW [24] optimizer " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " in BF16 mixed precision. We use a constant learning rate of" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", a weight decay of " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", and gradient clipping with a threshold of 1.0. Exponential moving average of the weights are gathered by a decay factor of 0.99 every 100 steps. We employ the same logit-normal distribution as used in SD 3 for timestep sampling. During training, " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " of the texts are randomly dropped to learn unconditional generation. Training is carried out using Google TPU v4-256 pods and FSDP implemented by PyTorch / XLA SPMD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 186, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 247 + ], + "type": "text", + "content": "Inference. We conduct inference using Euler discretization with 25 sampling steps, and a classifier-free guidance scale of 6 which we find to be near optimal for text-image alignment. We employ identical sampling steps and guidance scale across all experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 265, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 397 + ], + "type": "text", + "content": "Evaluation. We evaluate image-text alignment using GenEval [12] and DPG-Bench [15] metrics, prioritizing GenEval for its robustness. While both benchmarks provide valuable insights, DPG-Bench exhibits certain limitations, such as rapid performance saturation and potential measurement errors [21]. To ensure a comprehensive evaluation, we also provide visual quality measurements using FID [14] on MJHQ-30K [18]. Notably, image-text alignment does not always correlate positively with visual quality, often presenting trade-offs. Our sampling and evaluation are primarily carried out using NVIDIA L40S GPUs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 409, + 520, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 520, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 520, + 423 + ], + "type": "text", + "content": "5. Comparing Deep and Shallow Fusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "content": "The deep fusion approach fuses an LLM and a DiT through layer-wise shared self-attention, creating interconnections throughout the network. However, established architectures typically condition on representations from a single text encoder layer. To investigate the true potential of deep fusion, we conduct controlled comparisons with baseline methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 503, + 554, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 554, + 599 + ], + "type": "text", + "content": "For a fair and meaningful comparison, we examine a common architectural paradigm which we refer to as shallow fusion. In this approach, representations from a single text encoder layer are integrated into each DiT layer through late fusion within the attention operation. Unlike deep fusion, which involves multi-layered interactions, shallow fusion maintains a fixed connection between each DiT layer and a prescribed text encoder layer (Fig. 1)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "type": "text", + "content": "5.1. Shallow Fusion Baselines" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 626, + 554, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 554, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 554, + 674 + ], + "type": "text", + "content": "We consider two shallow fusion architectures that condition on last-layer hidden states of LLMs as our baselines. As illustrated in Fig. 3, the two architecture differ in how they aggregate information from the condition." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "- Self-attention DiT. In this design, text representations are projected to key and value states and then concatenated with those of image hidden states in self-attention, which" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28588" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 192, + 137 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 192, + 137 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 192, + 137 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 192, + 137 + ], + "type": "image", + "image_path": "26fb8174e61731ae5a33672dea3ec54381e8ac694b6cb96d9be6721b3e5532e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 138, + 153, + 148 + ], + "lines": [ + { + "bbox": [ + 94, + 138, + 153, + 148 + ], + "spans": [ + { + "bbox": [ + 94, + 138, + 153, + 148 + ], + "type": "text", + "content": "(a) Self-Attention" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 69, + 296, + 137 + ], + "blocks": [ + { + "bbox": [ + 211, + 69, + 296, + 137 + ], + "lines": [ + { + "bbox": [ + 211, + 69, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 211, + 69, + 296, + 137 + ], + "type": "image", + "image_path": "7d7637fd3db9bb8167e746807b8a70b2de942f3ba91772f80d5f7573c1b08c51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 138, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 220, + 138, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 220, + 138, + 285, + 148 + ], + "type": "text", + "content": "(b) Cross-Attention" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 158, + 295, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 192 + ], + "type": "text", + "content": "Figure 3. Illustration of cross-modal attention in the shallow fusion baselines. The key and query states of the condition are directly projected from text representations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 200, + 295, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 200, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 200, + 295, + 236 + ], + "type": "text", + "content": "can also be decoupled by running self-attention and cross-attention in parallel and merging their outputs. This approach resembles architectures proposed by [11, 51]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "type": "text", + "content": "- Cross-attention DiT. This design also projects text representations to key and values states. However, unlike the previous approach, they are used for additional cross-attention with image hidden states, applied after the self-attention in each layer. This architecture follows the methodology employed in [5, 6, 47]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 312, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 395 + ], + "type": "text", + "content": "Compared to the deep fusion approach, both baseline models include RMS normalization and a linear layer for text representations before passing them through additional key and value projection layers. Additionally, the cross-attention DiT model uses extra query projections in its cross-attention mechanism for image hidden states. Other model configurations follow the DiT design in Sec. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 396, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 528 + ], + "type": "text", + "content": "Notably, the deep fusion approach can be reinterpreted as a variant of the self-attention DiT architecture, as they both aggregate conditional information through in-context self-attention. The key difference lies in how the key and value states of the condition are derived: self-attention DiT employs a trainable projection to generate these states from a single text encoder layer, whereas deep fusion extracts them from corresponding LLM layers. Despite their similarities, the two approaches have fundamentally different conceptual implications: deep fusion treats the LLM and DiT as equal components of a unified model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 635 + ], + "type": "text", + "content": "Apart from the aforementioned designs, SD3 [10] introduced an alternative fusion strategy that uses a two-stream transformer (MM-DiT) to jointly processes noised image latents and linguistic representations. While this is another popular approach worthy of investigation, an apples-to-apples fair comparison between the deep fusion approach and MM-DiT is not feasible, as both streams in MM-DiT are trainable. Consequently, our analysis focuses solely on the shallow fusion baselines detailed in this section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 189, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 189, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 189, + 659 + ], + "type": "text", + "content": "5.2. Controlled Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "For a fair comparison, we design self-attention DiT, cross-attention DiT, and the deep fusion model with similar architectures, and train them for 300K steps following the setup detailed in Sec. 4." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 380, + 148 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 380, + 148 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 380, + 148 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 380, + 148 + ], + "type": "image", + "image_path": "4ad6daea267291d9d27033456f39ebb502b3aa1d89cba3a628c6a0fa2e9605c3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 381, + 71, + 443, + 150 + ], + "blocks": [ + { + "bbox": [ + 381, + 71, + 443, + 150 + ], + "lines": [ + { + "bbox": [ + 381, + 71, + 443, + 150 + ], + "spans": [ + { + "bbox": [ + 381, + 71, + 443, + 150 + ], + "type": "image", + "image_path": "69b9d2ebe405c405dc29059100fd208a75a1ddcdb93ead233685eee93f343571.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 159, + 555, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 192 + ], + "type": "text", + "content": "Figure 4. Illustration of timestep conditioning strategies. Removing timesetp conditioning leads to the fewest parameters and the best overall performance." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 449, + 71, + 493, + 148 + ], + "blocks": [ + { + "bbox": [ + 449, + 71, + 493, + 148 + ], + "lines": [ + { + "bbox": [ + 449, + 71, + 493, + 148 + ], + "spans": [ + { + "bbox": [ + 449, + 71, + 493, + 148 + ], + "type": "image", + "image_path": "17652829af896ca34a07b42ef4554cbbaf992c28c08a29ac8e203faf660a732e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 499, + 71, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 499, + 71, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 499, + 71, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 499, + 71, + 544, + 148 + ], + "type": "image", + "image_path": "af6cc08e1cbca3253e4642d2b0f6589d81c71a372713008802f73974d59841ae.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "text", + "content": "As shown in Tab. 1, the deep fusion model achieve significantly better performance in image-text alignment than the self-attention DiT model and also surpass the cross-attention DiT model, while shallow fusion models demonstrate better visual quality. In terms of inference efficiency, deep fusion also demonstrates competitive performance, as shown in Tab. 2. This positive evidence underscores the compelling positioning of the deep fusion approach within the current landscape." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 315, + 308, + 558, + 351 + ], + "blocks": [ + { + "bbox": [ + 315, + 308, + 558, + 351 + ], + "lines": [ + { + "bbox": [ + 315, + 308, + 558, + 351 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 558, + 351 + ], + "type": "table", + "html": "
MethodParams.GenEval ↑DPG ↑FID ↓
Self-Attention2.47B0.4273.926.16
Cross-Attention2.62B0.4976.324.00
Deep Fusion2.45B0.5176.627.33
", + "image_path": "adc207133b2678079d562e2a718edb98fbd5a08d62396098972e2e3b91b64b03.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 331, + 392, + 536, + 434 + ], + "blocks": [ + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "type": "text", + "content": "Table 1. Comparison of performance between deep and shallow fusion models. Deep fusion beats shallow fusion in text-image alignment while underperforms in visual quality." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 392, + 536, + 434 + ], + "lines": [ + { + "bbox": [ + 331, + 392, + 536, + 434 + ], + "spans": [ + { + "bbox": [ + 331, + 392, + 536, + 434 + ], + "type": "table", + "html": "
MethodParams.Inference Latency (s)
Self-Attention2.47B1.75
Cross-Attention2.62B1.86
Deep Fusion2.45B1.66
", + "image_path": "aa717091f0cf85075b07a7bf51e53deb9015074c37a34a64294643a3ac546fe0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 436, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 468 + ], + "type": "text", + "content": "Table 2. Comparison of inference latency between deep and shallow fusion models. The numbers are measured with a batch size of 1 in automatic mixed precision on an NVIDIA A100 GPU." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 476, + 489, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 489, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 489, + 491 + ], + "type": "text", + "content": "6. Examining Key Design Choices" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 497, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 604 + ], + "type": "text", + "content": "In this section, we examine key design choices of the deep fusion approach through a text-to-image-centric lens. We begin by assessing the necessity and potential redundancy of parameters for timestep conditioning, determining whether to optimize their use or eliminate them altogether (Sec. 6.1). Next, we compare various positional encoding strategies (Sec. 6.2). Additionally, we investigate how the choice of base LLM and the use of instruction prompts impacts text-to-image performance (Sec. 6.3)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "content": "For all experiments, we use the default design as the baseline and train the models for 300K steps, following the setup detailed in Sec. 4." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 648, + 444, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 444, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 444, + 661 + ], + "type": "text", + "content": "6.1. Timestep Conditioning" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "DiT [29] has introduced AdaLN-Zero as the standard mechanism for injecting timestep and class label information. The AdaLN modules typically accounts for a large proportion of the model parameters, 0.5B out of a total 2.5B in our" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28589" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "content": "case. However, since our text-to-image model does not use class labels, AdaLN serves only for timestep conditioning. This substantial parameter allocation raises a critical question about its necessity and potential redundancy: Could these parameters be utilized more effectively, or are they redundant altogether?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 144, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 144, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 144, + 294, + 168 + ], + "type": "text", + "content": "Question 1.1. Can adaLN parameters be utilized more effectively by integrating additional text conditioning?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 168, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 168, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 168, + 294, + 251 + ], + "type": "text", + "content": "Prior research [10, 11, 17] has integrated text information into AdaLN-Zero by augmenting timestep embeddings with pooled text representations through summation. To optimize parameter efficiency, we follow this approach by leveraging embeddings1 from the CLIP L/14 text encoder [31], which provides high-quality linguistic representations trained on large-scale multi-modal data." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 75, + 255, + 277, + 289 + ], + "blocks": [ + { + "bbox": [ + 75, + 255, + 277, + 289 + ], + "lines": [ + { + "bbox": [ + 75, + 255, + 277, + 289 + ], + "spans": [ + { + "bbox": [ + 75, + 255, + 277, + 289 + ], + "type": "table", + "html": "
MethodGenEval ↑DPG ↑FID ↓
adaLN-Zero0.5176.627.33
+ CLIP L/140.5076.224.00
", + "image_path": "7829ffd7d522fc4647aceeb2c92e2a416c56d7bb3a319366b6a91e3545c8df13.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": "The results indicate that integrating text modulation results in slight improvements in FID but weakens image-text alignment. Additionally, it further increases compute." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 354 + ], + "type": "text", + "content": "Question 1.2. Can we shrink the parameters for timestep conditioning?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 355, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 426 + ], + "type": "text", + "content": "We explore the possibility of eliminating timestep conditioning parameters to develop a more streamlined architecture akin to that of the LLM. Specifically, we compare four timestep conditioning strategies from previous work, each with a progressively reduced number of parameters, as illustrated in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 428, + 295, + 548 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 428, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 464 + ], + "type": "text", + "content": "- adaLN-Zero. Following the vanilla DiT [29], we regress zero-initialized modulation parameters for each layer from the timestep embedding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "type": "text", + "content": "- adaLN-Single. Following PixArt-" + }, + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "type": "text", + "content": " [6], we compute a global set of modulation parameters and refine them per layer by adding learnable embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 500, + 295, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 524 + ], + "type": "text", + "content": "- Addition. Following Transfusion [50], we directly add the timestep embedding to all image tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 524, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 295, + 548 + ], + "type": "text", + "content": "- w/o Timestep. Inspired by a recent study [42], we completely remove the timestep conditioning from the model." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 56, + 555, + 298, + 608 + ], + "blocks": [ + { + "bbox": [ + 56, + 555, + 298, + 608 + ], + "lines": [ + { + "bbox": [ + 56, + 555, + 298, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 298, + 608 + ], + "type": "table", + "html": "
MethodParams.GenEval ↑DPG ↑FID ↓
adaLN-Zero2.47B0.5176.627.33
adaLN-Single2.01B0.4775.227.09
addition1.99B0.4775.626.40
w/o timestep1.98B0.4976.721.27
", + "image_path": "833c4512f9f0b66e26e14c83e2116e4685ebb968b0c628221f05232639ce28bc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 639, + 294, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 294, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 294, + 687 + ], + "type": "text", + "content": "Surprisingly, the results in Tab. 3 indicate that reducing the number of parameters in timestep conditioning consistently enhances visual quality, whereas the performance in image-text alignment exhibits fluctuations." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 319, + 71, + 550, + 133 + ], + "blocks": [ + { + "bbox": [ + 55, + 609, + 294, + 632 + ], + "lines": [ + { + "bbox": [ + 55, + 609, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 294, + 632 + ], + "type": "text", + "content": "Table 3. Evaluation of timestep conditioning strategies. Removing timestep conditioning yields surprisingly strong results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 71, + 550, + 133 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 550, + 133 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 550, + 133 + ], + "type": "table", + "html": "
Input SequenceTXT 1TXT 2[PAD]IMG 1IMG 2IMG 3IMG 4
1D-RoPE + APE[0][1]—[2D Frequency APE]
1D-RoPE[0][1]—[2][3][4][5]
1D + 2D-RoPE[0][1]—[2, 2][2, 3][3, 2][3, 3]
M-RoPE[0, 0][1, 1]—[2, 2][2, 3][3, 2][3, 3]
", + "image_path": "28523ec2faf641ffb71f38442a8a4572a5538838f27bc6d135936db2f28921f5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 154, + 553, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 154, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 154, + 553, + 297 + ], + "type": "text", + "content": "Notably, the strategy that completely removes timestep conditioning not only achieves significantly better FID but also maintains comparable GenEval and DPG-Bench performance. This finding is consistent with [42], where timestep conditioning removal improved FID in rectified flow models trained on smaller datasets. Furthermore, the complete removal of timestep conditioning eliminates the need for associated parameters, resulting in a " + }, + { + "bbox": [ + 312, + 154, + 553, + 297 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 312, + 154, + 553, + 297 + ], + "type": "text", + "content": " reduction in the total number of model parameters. Although this approach slightly lags behind AdaLN-Zero in terms of text-image alignment metrics, we prefer it due to its parameter efficiency and architectural simplicity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 304, + 430, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 430, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 430, + 317 + ], + "type": "text", + "content": "6.2. Positional Encoding" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 321, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 553, + 405 + ], + "type": "text", + "content": "Absolute positional encoding (APE) is widely used in text-to-image diffusion models, whereas in the context of LLMs, rotary positional embedding (RoPE) [40] is the predominant choice. As deep fusion models are inherently multimodal and differ from traditional text-to-image diffusion models, it is unclear which positional encoding (or their combinations) is best suited for mixed-modal sequences." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 406, + 553, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 429 + ], + "type": "text", + "content": "Question 2.1. Is RoPE more advantageous than APE for enhancing the performance of deep fusion models?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 430, + 553, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 553, + 453 + ], + "type": "text", + "content": "To explore this, we experiment with three positional encoding methods, as illustrated in Fig. 5:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 454, + 553, + 526 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 313, + 454, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 553, + 478 + ], + "type": "text", + "content": "- 1D RoPE + APE: We apply 1D RoPE to the text sequence and APE to the image sequence respectively." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 479, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 502 + ], + "type": "text", + "content": "- 1D RoPE: We extend 1D RoPE to encompass both text and image sequences." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 502, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 553, + 526 + ], + "type": "text", + "content": "- 1D + 2D RoPE: We apply 1D RoPE to the text sequence and 2D RoPE to the image sequence respectively." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 332, + 533, + 533, + 574 + ], + "blocks": [ + { + "bbox": [ + 315, + 137, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 137, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 552, + 148 + ], + "type": "text", + "content": "Figure 5. Illustration of RoPE. The indices denote position IDs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 533, + 533, + 574 + ], + "lines": [ + { + "bbox": [ + 332, + 533, + 533, + 574 + ], + "spans": [ + { + "bbox": [ + 332, + 533, + 533, + 574 + ], + "type": "table", + "html": "
MethodGenEval ↑DPG ↑FID ↓
1D-RoPE + APE0.5176.627.33
1D-RoPE0.4677.027.94
1D + 2D-RoPE0.5176.425.42
", + "image_path": "d3a9e8e522d2797c0f736845a817322975c3705e0e78d1789cdb2210df72a716.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 576, + 553, + 598 + ], + "lines": [ + { + "bbox": [ + 313, + 576, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 553, + 598 + ], + "type": "text", + "content": "Table 4. Comparing different positional encoding strategies. " + }, + { + "bbox": [ + 313, + 576, + 553, + 598 + ], + "type": "inline_equation", + "content": "1\\mathrm{D} + 2\\mathrm{D}" + }, + { + "bbox": [ + 313, + 576, + 553, + 598 + ], + "type": "text", + "content": " RoPE achieves the best overall performance." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 605, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 553, + 714 + ], + "type": "text", + "content": "As shown in Tab. 4, the 1D + 2D-RoPE configuration achieves the best overall performance, with only a marginal decrease in DPGBench compared to the 1D-RoPE + APE variant. The superiority of 2D-RoPE over APE suggests RoPE is more effective for modeling image sequences in deep fusion models. Using only 1D-RoPE slightly reduces performance, indicating that while deep fusion models treat text and image sequences as a unified input, their distinct positional characteristics are best modeled separately." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1We apply RMS-normalization and an MLP to the text embeddings before adding them to the timestep embeddings." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28590" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "Question 2.2. Does deep fusion models benefit from RoPE specifically designed for mixed-modal sequences?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 192 + ], + "type": "text", + "content": "Previous work on MLLMs has explored RoPE strategies for handling mixed-modal sequences. Naturally, we are curious whether deep fusion models can benefit from these positional encodings. Follow Qwen2-VL, we implement M-RoPE, a variant of RoPE that applies 2D positional IDs to chunked 1D RoPE frequencies, allowing it to function as 1D RoPE for text sequences while approximating 2D RoPE for image sequences." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 73, + 199, + 277, + 232 + ], + "blocks": [ + { + "bbox": [ + 73, + 199, + 277, + 232 + ], + "lines": [ + { + "bbox": [ + 73, + 199, + 277, + 232 + ], + "spans": [ + { + "bbox": [ + 73, + 199, + 277, + 232 + ], + "type": "table", + "html": "
MethodGenEval ↑DPG ↑FID ↓
1D + 2D-RoPE0.5176.425.42
M-RoPE0.4974.927.60
", + "image_path": "0f65c5adf3b3fae989aa7f6ce5436297adb20801b06a9469b0c0be4bfca3a9ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "type": "text", + "content": "Although M-RoPE elegantly unifies 1D and 2D-RoPEs, it still falls short compared to their direct combination. This underscores the challenge of designing position encodings for mixed-modal sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 128, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 128, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 128, + 307 + ], + "type": "text", + "content": "6.3. Base LLM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 313, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 361 + ], + "type": "text", + "content": "LLMs trained with different paradigms and data demonstrate diverse capabilities and behaviors. In this section, we examine how the choice of base LLM impacts the performance of deep fusion models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 361, + 295, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 385 + ], + "type": "text", + "content": "Question 3.1 Can instruction tuning, combined with instruction prompts, improve text-to-image performance?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 385, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 493 + ], + "type": "text", + "content": "Instruction tuning enables LLMs to effectively follow complex instructions. We explore whether this process can also enhance their internal information flow, leading to more contextualized and discriminative representations for text-to-image synthesis. We compare Gemma 2B with Gemma 2B IT, its instruction-tuned variant. We also experiment with using it with a simple instruction prompt, \"Imagine: \", which we find sufficient for guiding the LLM to generate detailed and relevant expansions of the input." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 64, + 498, + 286, + 541 + ], + "blocks": [ + { + "bbox": [ + 64, + 498, + 286, + 541 + ], + "lines": [ + { + "bbox": [ + 64, + 498, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 64, + 498, + 286, + 541 + ], + "type": "table", + "html": "
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ instruction tuning0.4975.427.04
+ instruction prompt0.5075.825.28
", + "image_path": "8d6c999776831b8f9e3600963c7d14cdf9d4aeedbfe4c7638490aaa0b55223f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": "As shown in Tab. 5, instruction tuning appears to have a slightly negative impact on performance. While the use of an instruction prompt mitigates this effect to some extent, consistent with findings in [25, 47], it still falls short of the baseline. This result highlights a challenge in effectively leveraging the instruction-following capabilities of LLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 666 + ], + "type": "text", + "content": "Question 3.2 Can multi-modal tuning improve text-to-image performance?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "Additionally, we're also interested in the effect of multi-modal tuning. While multi-modal tuning differs significantly from our setup, its shift in data distribution may still potentially enhance adaptability to multi-modal tasks. We" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "compare Gemma 2B with the base LLM of PaliGemma 3B PT [3], a multi-modal extension of Gemma 2B that undergoes additional pretraining on image-text data." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 323, + 110, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "type": "text", + "content": "Table 5. Evaluating the effect of instruction tuning. Using instruction-tuned LLMs does not improve performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 110, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 323, + 110, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 323, + 110, + 545, + 144 + ], + "type": "table", + "html": "
MethodGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
+ multi-modal tuning0.5276.226.30
", + "image_path": "840ea00435f9a93b36946d8deec1f074d77584864d0f7fa02323ee0be338b705.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 148, + 553, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 553, + 182 + ], + "type": "text", + "content": "which yields small improvements in performance. This observation indicates that multi-modal finetuning provides some benefit to the deep-fusion model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 183, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 206 + ], + "type": "text", + "content": "Question 3.3 Does improved LLM capabilities translate to stronger text-to-image performance?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 207, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 554, + 350 + ], + "type": "text", + "content": "Finally, as the base LLM become more proficient in understanding and generating text, they could potentially foster synergistic improvements in DiT performance. We compare Gemma 2B with Gemma 2 2B [34], the next generation of Gemma 2B which demonstrates a " + }, + { + "bbox": [ + 313, + 207, + 554, + 350 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 207, + 554, + 350 + ], + "type": "text", + "content": " absolute performance improvement " + }, + { + "bbox": [ + 313, + 207, + 554, + 350 + ], + "type": "inline_equation", + "content": "(0.44 \\rightarrow 0.50)" + }, + { + "bbox": [ + 313, + 207, + 554, + 350 + ], + "type": "text", + "content": " on an average of 8 language-only benchmarks. Notably, Gemma 2 2B features a different transformer architecture than Gemma 2B, prompting us to adjust the DiT architecture accordingly. This modification increases the number of adaLN parameters by 0.3B. However, as demonstrated in Sec. 6.1, these parameters do not contribute to model performance." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 333, + 354, + 536, + 387 + ], + "blocks": [ + { + "bbox": [ + 333, + 354, + 536, + 387 + ], + "lines": [ + { + "bbox": [ + 333, + 354, + 536, + 387 + ], + "spans": [ + { + "bbox": [ + 333, + 354, + 536, + 387 + ], + "type": "table", + "html": "
ModelGenEval ↑DPG ↑FID ↓
Gemma 2B0.5176.627.33
Gemma 2 2B0.5479.123.94
", + "image_path": "a2e7a5a0a021a348e9fde7d85ccc5db2174f94e3a0e49d14788f2b0447a7040f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 391, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 440 + ], + "type": "text", + "content": "Upgrading from Gemma 2B to Gemma 2 2B yields a drastic performance boost. This finding suggests that the DiT's performance in deep fusion models is strongly dependent on the capabilities of the underlying base LLM." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 449, + 417, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 417, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 417, + 463 + ], + "type": "text", + "content": "7. Training at Scale" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 469, + 554, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 541 + ], + "type": "text", + "content": "In this section, we present a final recipe for the deep fusion model, building on the original framework while incorporating key insights from previous exploration. We conduct large-scale training to benchmark our model against established systems, showcasing its scalability and competitive performance on the leaderboard." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 548, + 395, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 395, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 395, + 562 + ], + "type": "text", + "content": "7.1. Final Recipe" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 566, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 553, + 590 + ], + "type": "text", + "content": "Building on the insights from Sec. 6, we introduce the following design modifications to our model:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 592, + 553, + 639 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 313, + 592, + 452, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 452, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 452, + 602 + ], + "type": "text", + "content": "- Remove AdaLN-Zero modules." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 604, + 515, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 515, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 515, + 615 + ], + "type": "text", + "content": "- Replace 1D-RoPE + APE with 1D + 2D-RoPE." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 616, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 639 + ], + "type": "text", + "content": "- Replace Gemma 2B with Gemma 2 2B, adjusting the transformer configurations of the DiT accordingly." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "We train our model for 800K steps on a mixed dataset comprising CC12M [4], SA-1B [16], and the training subset of JourneyDB [41], amounting to approximately 26M image-caption pairs. Notably, state-of-the-art text-to-image models typically rely on high-quality datasets of much larger scale to achieve superior performance. For CC12M" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28591" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 135, + 142 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 135, + 142 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 135, + 142 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 135, + 142 + ], + "type": "image", + "image_path": "151612be0ce56445f20a9dcae542bed4e3577b9ed0541072b02516f8e5f18734.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 142, + 116, + 149 + ], + "lines": [ + { + "bbox": [ + 77, + 142, + 116, + 149 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 116, + 149 + ], + "type": "text", + "content": "(a) Fully Parallel" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 162, + 295, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 295, + 195 + ], + "type": "text", + "content": "Figure 6. Illustration of architecture alignment. Dashed boxes indicate parameters that have been reduced by decreasing either the hidden size or the number of layers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 72, + 214, + 141 + ], + "blocks": [ + { + "bbox": [ + 138, + 72, + 214, + 141 + ], + "lines": [ + { + "bbox": [ + 138, + 72, + 214, + 141 + ], + "spans": [ + { + "bbox": [ + 138, + 72, + 214, + 141 + ], + "type": "image", + "image_path": "cfae5e7ebb34949aa90ffbd1cdfa03240bc612a3d957b161f39361aadc715f26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 143, + 198, + 150 + ], + "lines": [ + { + "bbox": [ + 154, + 143, + 198, + 150 + ], + "spans": [ + { + "bbox": [ + 154, + 143, + 198, + 150 + ], + "type": "text", + "content": "(b) Skipped Layers" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 72, + 293, + 141 + ], + "blocks": [ + { + "bbox": [ + 217, + 72, + 293, + 141 + ], + "lines": [ + { + "bbox": [ + 217, + 72, + 293, + 141 + ], + "spans": [ + { + "bbox": [ + 217, + 72, + 293, + 141 + ], + "type": "image", + "image_path": "ea62c125e010eff6f16c791f75ab09bd4575bff41bfa09b4f2fcfe0c71831853.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 143, + 285, + 150 + ], + "lines": [ + { + "bbox": [ + 226, + 143, + 285, + 150 + ], + "spans": [ + { + "bbox": [ + 226, + 143, + 285, + 150 + ], + "type": "text", + "content": "(c) Decoupled Dimension" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 201, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 227 + ], + "type": "text", + "content": "and SA-1B, we utilize synthetic captions [6, 9]. Other experimental setup follows Sec. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 234, + 200, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 200, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 200, + 247 + ], + "type": "text", + "content": "7.2. Performance Comparison" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 56, + 255, + 296, + 423 + ], + "blocks": [ + { + "bbox": [ + 56, + 255, + 296, + 423 + ], + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 423 + ], + "type": "table", + "html": "
ModelParamsDataGen.↑DPG↑FID↓
SD 1.5 [35]0.9B4.8B0.4363.2—
DALL-E 2 [33]4.2B2.6B0.52——
SDXL [30]2.6B1.6B0.5574.76.63
PG 2.5 [18]2.6B—0.5675.56.09
SD 3 M [10]2B1B0.6284.111.92
DALL-E 3 [2]——0.6783.5—
FLUX.1 [dev] [17]12B—0.6784.010.15
PG 3 [21]24B—0.7687.0—
MicroDiT [37]1.2B37M0.46——
PixArt-α [6]0.6B25M0.4871.1—
Lumina Next [11]2B—0.4674.67.58
PixArt-Σ [5]0.6B46M0.5480.56.15
Transfusion [50]7.3B3.5B0.63——
Sana 1.0 1.6B [47]1.6B—0.6684.85.76
Ours2B26M0.6081.67.54
", + "image_path": "ff2663f46bf36b2ea20940472ba63a5ebf29e1283d28c043b5394b203dda6b37.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 547 + ], + "type": "text", + "content": "We compare our model with the most advanced text-to-image diffusion models in Tab. 6. Despite being trained with limited compute and data in a simplified setting, our model surpasses many industry-standard systems and delivers competitive results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 548, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 585 + ], + "type": "text", + "content": "We present qualitative examples from our model in Figure 7. Our model demonstrates the ability to generate high-quality images with superior prompt alignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 595, + 176, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 176, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 176, + 609 + ], + "type": "text", + "content": "8. Further Exploration" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": "In this section, we present preliminary studies exploring more aggressive modifications to the deep fusion approach. For all experiments, we train the models for 300K steps, following the setup detailed in Sec. 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 672, + 190, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 190, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 190, + 685 + ], + "type": "text", + "content": "8.1. Architecture Alignment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Up to this point, our approach has followed prior work by aligning the LLM and DiT backbones in a layer-by-layer" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "approach, strictly enforcing identical transformer configurations for both models. However, this rigid constraint limits the flexibility of deep fusion. In practice, we need the ability to scale the LLM and DiT independently, as different modalities follow distinct scaling laws and network design principles, and training and deployment scenarios vary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 146, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 255 + ], + "type": "text", + "content": "To address this, we explore modifying the DiT model's hidden size and number of transformer layers, as illustrated in Fig. 6. The adapted model is fused into the middle layers of the LLM, which contain richer semantic information [39]. Additionally, in self-attention, hidden states are still projected to query, key, and value states that match the LLM's dimensionality, ensuring compatibility. Since the most successful DiTs are generally much smaller than state-of-the-art LLMs, we focus on shrinking the size of our DiT." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 321, + 262, + 545, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 430, + 296, + 474 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 474 + ], + "type": "text", + "content": "Table 6. Comparison of performance with state-of-the-art systems. Table adapted from [48, 50]. Industrial baselines are presented at the top, while academic baselines and our model are listed at the bottom." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 262, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 321, + 262, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 262, + 545, + 316 + ], + "type": "table", + "html": "
Hidden sizeParams.GenEval ↑DPG ↑FID ↓
20482.5B0.5176.627.33
17922.1B0.5077.124.27
15361.8B0.4976.225.46
12801.4B0.4874.824.64
", + "image_path": "980420d20b71dc9966e04cc750c07050f71d644d18e5ecf1ab5aa1deea275892.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 321, + 351, + 545, + 393 + ], + "blocks": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "type": "text", + "content": "Table 7. Evaluating models of different hidden sizes. The default hidden size is 2048." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 351, + 545, + 393 + ], + "lines": [ + { + "bbox": [ + 321, + 351, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 321, + 351, + 545, + 393 + ], + "type": "table", + "html": "
LayersParams.GenEval ↑DPG ↑FID ↓
182.5B0.5176.627.33
141.9B0.4774.623.46
101.4B0.3368.028.34
", + "image_path": "39991e9a795bd1cb0b4a8b28a1567af9a327dbf7f53cdce514980d8a322521b1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 395, + 553, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 395, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 553, + 417 + ], + "type": "text", + "content": "Table 8. Evaluating different numbers of layers. The default number of layer is 18." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 429, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 548 + ], + "type": "text", + "content": "As shown in Table Tab. 7, the model's performance degrades gracefully as we reduce the hidden size, with visual quality actually improving in some cases. While decreasing the number of transformer layers (Tab. 8) also yields acceptable results, performance deteriorates more quickly. We hypothesize this occurs because Gemma 2B already employs fewer layers than typical model architectures. These findings suggest that LLM and DiT model designs can be effectively decoupled, enabling the application of separate scaling laws and design principles for each component." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 562, + 440, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 440, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 440, + 574 + ], + "type": "text", + "content": "8.2. Attention Mechanism" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "In Sec. 3, we build on prior work by defining the deep fusion architecture with shared self-attention to bridge the LLM and DiT. Inspired by extensive research on cross-attention in MLLMs [1, 13] and our findings in Sec. 5, which highlight the superior performance of cross-attention DiT over self-attention DiT, we explore an alternative deep fusion variant. This new approach replaces shared self-attention with cross-attention mechanisms, similar to cross-attention DiT but with a key distinction: we substitute the projected linguistic key and value states in traditional cross-attention DiT with corresponding states from LLM layers." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28592" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 182, + 194 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 182, + 194 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 182, + 194 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 182, + 194 + ], + "type": "image", + "image_path": "f046d53be3e87e89c215ac852dffa20cc94ed407976b407fe95dd2b052df5582.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 197, + 175, + 214 + ], + "lines": [ + { + "bbox": [ + 62, + 197, + 175, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 175, + 214 + ], + "type": "text", + "content": "Dew on blue rose petals, HD, close up, detail" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 72, + 304, + 194 + ], + "blocks": [ + { + "bbox": [ + 184, + 72, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 184, + 72, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 184, + 72, + 304, + 194 + ], + "type": "image", + "image_path": "273f3bb4e7aee303f6939030bf3ed037185f825f6eb8f3ef41a7678d6f8dc403.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 197, + 299, + 214 + ], + "lines": [ + { + "bbox": [ + 184, + 197, + 299, + 214 + ], + "spans": [ + { + "bbox": [ + 184, + 197, + 299, + 214 + ], + "type": "text", + "content": "A corgi wearing sunglasses walks on the beach of a tropical island" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 427, + 194 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 427, + 194 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 427, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 427, + 194 + ], + "type": "image", + "image_path": "dfe7a970366132155da3fd780030eca52ffc88ae64bfd7a3263d04105a80115f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 197, + 425, + 205 + ], + "lines": [ + { + "bbox": [ + 304, + 197, + 425, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 425, + 205 + ], + "type": "text", + "content": "A rally car taking a fast turn on a track" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 550, + 194 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 550, + 194 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 550, + 194 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 550, + 194 + ], + "type": "image", + "image_path": "cd47c8b9bc51ebe7a21cefd83c4e74962037de8b6b2fa7546420c68462475560.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 197, + 525, + 213 + ], + "lines": [ + { + "bbox": [ + 427, + 197, + 525, + 213 + ], + "spans": [ + { + "bbox": [ + 427, + 197, + 525, + 213 + ], + "type": "text", + "content": "An armchair in the shape of an avocado" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 217, + 182, + 339 + ], + "blocks": [ + { + "bbox": [ + 60, + 217, + 182, + 339 + ], + "lines": [ + { + "bbox": [ + 60, + 217, + 182, + 339 + ], + "spans": [ + { + "bbox": [ + 60, + 217, + 182, + 339 + ], + "type": "image", + "image_path": "28e5c9516d6dd5f4a1afd11cbf5efd2a7820a07cfde5c41cdd31dc20b8c1f789.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 342, + 175, + 367 + ], + "lines": [ + { + "bbox": [ + 62, + 342, + 175, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 175, + 367 + ], + "type": "text", + "content": "A man riding a horse through the Gobi Desert with a beautiful sunset behind him, movie quality." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 182, + 217, + 304, + 338 + ], + "blocks": [ + { + "bbox": [ + 182, + 217, + 304, + 338 + ], + "lines": [ + { + "bbox": [ + 182, + 217, + 304, + 338 + ], + "spans": [ + { + "bbox": [ + 182, + 217, + 304, + 338 + ], + "type": "image", + "image_path": "e8e8b5e639fad0c7071a397ea7b1941c2ce13106225d94f9562a34e10203f4e7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 342, + 282, + 358 + ], + "lines": [ + { + "bbox": [ + 184, + 342, + 282, + 358 + ], + "spans": [ + { + "bbox": [ + 184, + 342, + 282, + 358 + ], + "type": "text", + "content": "Pirate ship trapped in a cosmic maelstrom nebula" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 304, + 217, + 427, + 339 + ], + "blocks": [ + { + "bbox": [ + 304, + 217, + 427, + 339 + ], + "lines": [ + { + "bbox": [ + 304, + 217, + 427, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 427, + 339 + ], + "type": "image", + "image_path": "18705d9ef123edb53eb534838fe38b5ac85508776d8fb05d7cdf78373ab50ddd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 342, + 416, + 358 + ], + "lines": [ + { + "bbox": [ + 306, + 342, + 416, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 416, + 358 + ], + "type": "text", + "content": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 217, + 550, + 339 + ], + "blocks": [ + { + "bbox": [ + 427, + 217, + 550, + 339 + ], + "lines": [ + { + "bbox": [ + 427, + 217, + 550, + 339 + ], + "spans": [ + { + "bbox": [ + 427, + 217, + 550, + 339 + ], + "type": "image", + "image_path": "5bac32400729d933a1bf4611893eda1268d340793d61fcf4b90e94ac8cc7e1bb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 342, + 547, + 358 + ], + "lines": [ + { + "bbox": [ + 427, + 342, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 427, + 342, + 547, + 358 + ], + "type": "text", + "content": "A person surfing on a wave of stars in outer space." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 60, + 369, + 183, + 492 + ], + "blocks": [ + { + "bbox": [ + 60, + 369, + 183, + 492 + ], + "lines": [ + { + "bbox": [ + 60, + 369, + 183, + 492 + ], + "spans": [ + { + "bbox": [ + 60, + 369, + 183, + 492 + ], + "type": "image", + "image_path": "de8a836baaf5eedb97e37dbc2a581bd03d26316c737bc1af27ab0268806ff3f7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 495, + 182, + 542 + ], + "lines": [ + { + "bbox": [ + 59, + 495, + 182, + 542 + ], + "spans": [ + { + "bbox": [ + 59, + 495, + 182, + 542 + ], + "type": "text", + "content": "a picturesque autumn scene where a quaint cottage with a thatched roof sits beside a tranquil lake, surrounded by trees with leaves in vibrant shades of orange, red, and yellow. The cottage's wooden exterior is complemented by white-framed windows, and a stone chimney rises above the roofline. The lake reflects the warm fall colors, creating a mirror image of the foliage and the small structure on its calm surface." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 184, + 369, + 304, + 492 + ], + "blocks": [ + { + "bbox": [ + 184, + 369, + 304, + 492 + ], + "lines": [ + { + "bbox": [ + 184, + 369, + 304, + 492 + ], + "spans": [ + { + "bbox": [ + 184, + 369, + 304, + 492 + ], + "type": "image", + "image_path": "3e1fb7678a63b70f38190cabfd67b47f9d74ede94322dab8314a643b32a52fe7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 495, + 296, + 529 + ], + "lines": [ + { + "bbox": [ + 184, + 495, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 184, + 495, + 296, + 529 + ], + "type": "text", + "content": "A portrait of a human growing colorful flowers from her hair. Hyperrealistic oil painting. Intricate details." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 304, + 369, + 426, + 493 + ], + "blocks": [ + { + "bbox": [ + 304, + 369, + 426, + 493 + ], + "lines": [ + { + "bbox": [ + 304, + 369, + 426, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 426, + 493 + ], + "type": "image", + "image_path": "084ba814fbc087a2d54f5940196c570a463401e793865aa3382705d6b3d6fd4b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 495, + 422, + 540 + ], + "lines": [ + { + "bbox": [ + 304, + 495, + 422, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 422, + 540 + ], + "type": "text", + "content": "a black and white picture of a woman looking through the window, in the style of Duffy Sheridan, Anna Razumovskaya, smooth and shiny, wavy, Patrick Demarchelier, album covers, lush and detailed" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 495, + 537, + 529 + ], + "lines": [ + { + "bbox": [ + 427, + 495, + 537, + 529 + ], + "spans": [ + { + "bbox": [ + 427, + 495, + 537, + 529 + ], + "type": "text", + "content": "13 year old farm boy, red tousled hair, green eyes, soft pale skin, freckles, stub nose, very beautiful, very pretty, dark fantastic medieval" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 425, + 369, + 550, + 492 + ], + "blocks": [ + { + "bbox": [ + 425, + 369, + 550, + 492 + ], + "lines": [ + { + "bbox": [ + 425, + 369, + 550, + 492 + ], + "spans": [ + { + "bbox": [ + 425, + 369, + 550, + 492 + ], + "type": "image", + "image_path": "db084e5ab4541cd00906c718f8d7893db1d011a0033645b3e5a12b5b0d59d5ef.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 555, + 392, + 566 + ], + "lines": [ + { + "bbox": [ + 216, + 555, + 392, + 566 + ], + "spans": [ + { + "bbox": [ + 216, + 555, + 392, + 566 + ], + "type": "text", + "content": "Figure 7. Samples generated by our final recipe." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 75, + 570, + 276, + 602 + ], + "blocks": [ + { + "bbox": [ + 75, + 570, + 276, + 602 + ], + "lines": [ + { + "bbox": [ + 75, + 570, + 276, + 602 + ], + "spans": [ + { + "bbox": [ + 75, + 570, + 276, + 602 + ], + "type": "table", + "html": "
MethodGenEval ↑DPG ↑FID ↓
self-attention0.5176.627.33
cross-attention0.5276.526.57
", + "image_path": "1a3328bff66a73947f8242944fced02bed6e2ab1501e75cd647e80b4c97fbb44.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 611, + 294, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 294, + 683 + ], + "type": "text", + "content": "This modification yields minor gains, though at a cost to the LLM-DiT parity. Additionally, we find that although cross-attention introduces a negligible increase in FLOPs and parameter count, it leads to approximately a " + }, + { + "bbox": [ + 55, + 611, + 294, + 683 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 55, + 611, + 294, + 683 + ], + "type": "text", + "content": " increase in latency" + }, + { + "bbox": [ + 55, + 611, + 294, + 683 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 611, + 294, + 683 + ], + "type": "text", + "content": " (1.66s vs. 1.86s). Therefore, we retained the self-attention design in our final configuration." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 571, + 387, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 571, + 387, + 583 + ], + "spans": [ + { + "bbox": [ + 314, + 571, + 387, + 583 + ], + "type": "text", + "content": "9. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 591, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 555, + 687 + ], + "type": "text", + "content": "We have studied the recently popular deep fusion of a frozen LLM with a trainable DiT for text-to-image synthesis. Our findings provide empirical evidence supporting its advantages over baselines. We highlight key design choices, identify unresolved problems, and offer meaningful data points alongside practical guidelines. We hope our empirical work help advance multi-modal generation and bridge the gap between auto-regressive decoding and denoising diffusion." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2Latency is measured with a batch size of 1 under automatic mixed precision on an NVIDIA A100 GPU" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28593" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 201 + ], + "type": "text", + "content": "We thank the reviewers for their constructive feedback. We also thank Shusheng Yang, Shengbang Tong, Wenhao Chai, Nanye Ma, Sihan Xu, and Chenyu Li for insightful discussions. This work was mainly supported by the Google TPU Research Cloud (TRC) program, and the Google Cloud Research Credits program (GCP19980904) and Open Path AI Foundation. SX also acknowledges support from Intel AI SRS, IITP grant funded by the Korean Government (MSIT) (RS-2024-00457882, National AI Research Lab Project), Amazon Research Award, and NSF Award IIS-2443404." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 221, + 115, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 221, + 115, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 221, + 115, + 233 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 240, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "type": "text", + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.opencai.com/papers/dall-e-3.pdf, 2023.1,2,7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 351, + 295, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 351, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 351, + 295, + 405 + ], + "type": "text", + "content": "[3] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv:2407.07726, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "text", + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "type": "text", + "content": "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "type": "text", + "content": " : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024.1,2,4,7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "text", + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "text", + "content": ": Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 603 + ], + "type": "text", + "content": "[7] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 604, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 296, + 669 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[9] Caption Emporium. conceptual-captions-cc12m-llavanext. https://huggingface.co/datasets/CaptionEmporium / conceptual - captions - cc12m-llavanext, 2024.3,7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[11] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv:2405.05945, 2024. 2, 4, 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 218 + ], + "type": "text", + "content": "[12] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. arXiv:2310.11513, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "text", + "content": "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 309, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 352 + ], + "type": "text", + "content": "[15] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv:2403.05135, 2024. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "text", + "content": "[17] Black Forest Labs. Announcing black forest labs. https://blackforestlabs.ai/announcing-black-forest-labs/, 2024. 1, 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "type": "text", + "content": "[18] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv:2402.17245, 2024. 2, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "text", + "content": "[19] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv:2305.13655, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "text", + "content": "[20] Weixin Liang, Lili Yu, Liang Luo, Srinivasan Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, et al. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. arXiv:2411.04996, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 634 + ], + "type": "text", + "content": "[21] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv:2409.10695, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "type": "text", + "content": "[22] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv:2407.00737, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv:2209.03003, 2022. 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28594" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "text", + "content": "[25] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv:2406.11831, 2024. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 195 + ], + "type": "text", + "content": "[26] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv:2411.07975, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 294, + 251 + ], + "type": "text", + "content": "[27] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv:2403.08295, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 57, + 253, + 294, + 296 + ], + "type": "text", + "content": "[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 298, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 298, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 298, + 294, + 319 + ], + "type": "text", + "content": "[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 57, + 321, + 294, + 374 + ], + "type": "text", + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 377, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 377, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 57, + 377, + 294, + 441 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 445, + 294, + 487 + ], + "type": "text", + "content": "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In JMLR, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 489, + 294, + 520 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 523, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 294, + 576 + ], + "type": "text", + "content": "[34] Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv:2408.00118, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 579, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 610 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[37] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv:2407.15811, 2024. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "text", + "content": "[38] Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, and Lili Yu. Lmfusion:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 655 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Adapting pretrained language models for multimodal generation. arXiv:2412.15188, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[39] Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models. arXiv:2502.02013, 2025. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[40] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "text", + "content": "[41] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. In NeurIPS, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "text", + "content": "[42] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv:2502.13129, 2025. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "text", + "content": "[44] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv:2401.15688, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "type": "text", + "content": "[45] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "type": "text", + "content": "[46] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv:2409.11340, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[47] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv:2410.10629, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[48] Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv:2501.18427, 2025. 1, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[49] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[50] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv:2408.11039, 2024. 1, 2, 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "text", + "content": "[51] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv:2406.18583, 2024. 1, 2, 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28595" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_content_list.json b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a59fe5d756452f438d2fec5681194f2d260190f1 --- /dev/null +++ b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_content_list.json @@ -0,0 +1,1807 @@ +[ + { + "type": "text", + "text": "Exposure-slot: Exposure-centric representations learning with Slot-in-Slot Attention for Region-aware Exposure Correction", + "text_level": 1, + "bbox": [ + 184, + 130, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Donggoo Jung\\*, Daehyun Kim\\*, Guanghui Wang, Tae Hyun Kim† Hanyang University, Toronto Metropolitan University", + "bbox": [ + 215, + 202, + 779, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dgjung, daehyun, taehyunkim}@hanyang.ac.kr, {wangcs}@torontomu.ca", + "bbox": [ + 207, + 241, + 787, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image exposure correction enhances images captured under diverse real-world conditions by addressing issues of under- and over-exposure, which can result in the loss of critical details and hinder content recognition. While significant advancements have been made, current methods often fail to achieve optimal feature learning for effective correction. To overcome these challenges, we propose Exposure-slot, a novel framework that integrates a prompt-based slot-in-slot attention mechanism to cluster exposed feature regions and learn exposure-centric features for each cluster. By extending the Slot Attention algorithm with a hierarchical structure, our approach progressively clusters features, enabling precise and region-aware correction. In particular, learnable prompts tailored to exposure characteristics of slots further enhance feature quality, adapting dynamically to varying conditions. Our method delivers superior performance on benchmark datasets, surpassing the current state-of-the-art with a PSNR improvement of over 1.85 dB on the SICE dataset and 0.4 dB on the LCDP dataset, thereby establishing a new benchmark for multi-exposure correction. The source code can be found at: https://github.com/kdhRick2222/Exposure-slot.", + "bbox": [ + 86, + 323, + 485, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 684, + 220, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image exposure correction enhances images captured under diverse real-world conditions. Under- or over-exposed images can appear excessively dark or bright, losing essential details and compromising accurate recognition. Despite advancements in imaging technology, robust automatic exposure correction remains a significant challenge.", + "bbox": [ + 89, + 710, + 482, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to its critical importance, image exposure correction has been the focus of considerable research. Early efforts primarily treated under-exposure and over-exposure as separate issues [10, 37, 40]. More recent advancements, however, have introduced multi-exposure correction mod", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1e3638b18597a88089e950bfec1c3792e50485f2d9030b3ff6aab4306168e0d.jpg", + "image_caption": [ + "Figure 1. (Top) The Slot-in-Slot Attention mechanism hierarchically partitions exposure-aware regions using the attention maps of main and sub-slot attention, performing corrections with learnable prompts based on exposure levels. (Bottom) Comparison of the Exposure-slot with existing approaches. From left to right: ECLNet [14], FECNet [13], DRBN-ENC [12], CSEC [22] and our proposed approach, Exposure-slot." + ], + "image_footnote": [], + "bbox": [ + 517, + 290, + 906, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els capable of handling a wide range of exposure levels within a unified framework. For instance, MSEC [1] introduced an approach that considers both color and detail enhancement while performing multi-exposure correction in an end-to-end manner, and provided a dataset of images with different exposure errors for both training and evaluation. In addition, LCDPNet [35] provided another dataset of diverse scenes, covering both over-exposure and under-exposure, and incorporates retinax theory to account for local color distributions within images. Recent approaches increasingly focus on the separate processing of features based on their physical characteristics, such as exposure, frequency, contrast, color, and details. FECNet [13] uses Fourier transform to facilitate interactions between local spatial feature information and global frequency information, while ECLNet [14] uses a bilateral activation mechanism to process over- and under-exposed regions independently. In addition, DA [39] introduces a decoupling and", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 107, + 886, + 362, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17892", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aggregation scheme that separately enhances image details and contrast.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, relying on separating image features based on these physical properties may not ensure the optimal features for effective exposure correction. For example, at the bottom of Fig. 1, we visualize this limitation, showing that previous works [13-15, 22] often generate inappropriate artifacts in sky regions and struggle to accurately adjust sunset exposure levels.", + "bbox": [ + 89, + 122, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these challenges, we propose a novel prompt-based slot-in-slot attention mechanism that clusters regions with similar exposure levels, enabling exposure-centric feature learning within each cluster. We utilize Slot Attention [24], designed to cluster features based on object representations, to identify distinct exposure regions within an image. In our approach, slots represent features grouped by exposure levels, and the corresponding attention maps act as exposure-aware region maps, guiding the grouping of features according to exposure characteristics. Therefore, these maps highlight regions based on exposure levels, enabling targeted exposure correction for each area.", + "bbox": [ + 89, + 229, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To support effective correction, we introduce learnable prompt vectors that operate based on exposure-aware region maps, which are attention maps of each slot. These prompts adapt to the characteristics of each region defined by the attention map, helping the model learn targeted correction to improve under- or over-exposed areas. Specifically, we introduce Slot-Prompt Interaction Module (SPIM), which combines slot and prompt information using cross-attention. The cross-attention component of SPIM uses the prompts learned through Slot-in-Slot Attention as conditioning factors, enhancing the interaction between features and prompts.", + "bbox": [ + 89, + 412, + 482, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach builds upon standard slot attention with a hierarchical design, called Slot-in-Slot Attention, which processes slots in levels. This mechanism operates across multiple structure levels: the first level coarsely partitions regions based on distinct exposure characteristics, while subsequent levels progressively refine these regions with more partitions. Each slot captures exposure-specific regional information, distinguishing areas with varying exposure levels to enhance feature separation and improve the precision of local details.", + "bbox": [ + 89, + 597, + 482, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overall framework of the proposed approach is illustrated at the top of Fig. 1. The Slot-in-Slot Attention is applied to intermediate features between the encoder and decoder, generating attention maps for slots that are hierarchically partitioned according to exposure characteristics, as depicted in Fig. 1. The generated attention maps are then used as weight maps for learnable prompts, resulting in refined intermediate features that are subsequently fed into the decoder. We refer to our model as Exposure-Slot, as it is the first model to use Slot Attention for unsupervised feature", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "partitioning based on exposure levels in exposure correction, and the first to apply region-aware prompts for feature enhancement. Our main contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Exposure-slot is the first approach to leverage Slot Attention mechanism for optimized exposure-specific feature partitioning.", + "- We introduce the slot-in-slot attention that enables sophisticated feature partitioning and learning.", + "- We apply exposure-aware prompts that enhance the exposure-centric characteristics of each image feature.", + "- Exposure-slot achieves state-of-the-art results on multi-exposure benchmark datasets [1, 5, 35], setting a new standard in this field." + ], + "bbox": [ + 511, + 138, + 903, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 305, + 653, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Exposure Correction", + "text_level": 1, + "bbox": [ + 513, + 330, + 710, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rise of deep neural networks (DNN), exposure correction has seen the emergence of a range of DNN-based methods leveraging diverse concepts. For under-exposed image enhancement, methods [36, 40, 41, 43, 44] inspired by retinax theory have been proposed. Retinex-based models like CMEC [27] have also incorporated attention mechanisms to address multi-exposure correction, while LCDP-Net [35] introduced a novel focus on local color distribution. To support multi-exposure correction tasks, dedicated datasets such as MSEC [1] and SICE [5] were developed for training and evaluation. MSEC further introduced a Laplacian pyramid architecture to handle varying exposure levels. ENC [12] contributed an exposure normalization module that refined feature maps by transforming diverse exposure features into an exposure-invariant feature space. CSNorm [42] also enhanced model generalization by selectively normalizing lightness-sensitive channels and ERL [15] proposed exposure relationship learning through regularization techniques for multi-exposure correction.", + "bbox": [ + 511, + 353, + 906, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are several recent studies that emphasize the feature separation based on their physical characteristics. ECLNet [14] used a bilateral activation mechanism to adjust processing across different exposure conditions, while FECNet [13] proposed a lightweight model that utilizes spatial-frequency interactions with a Fourier-based approach. CLODE [17] also utilizes image curve adjustment and formulates brightness changes through continuous curve map extraction using Neural ODEs [6]. The latest advancement, CSEC [22], addresses color distribution shifts by defining darkened and brightened feature maps.", + "bbox": [ + 511, + 641, + 905, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Slot Attention", + "text_level": 1, + "bbox": [ + 513, + 816, + 656, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Slot Attention [24] is a mechanism for clustering features corresponding to slots which is the main object of object-centric learning (OCL). Specifically, this mechanism iteratively employs dot-product attention [25] with attention co", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17893", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c370d44eacae54c2cff85279b6d38a663e2fbe90d27a4f7d16e5dff89a2c3918.jpg", + "image_caption": [ + "Figure 2. Overview of Exposure-Slot: Exposure-Slot operates within a $U$ -shaped encoder-decoder network, with the Slot-Prompt Interaction Module (SPIM) bridging the encoder and decoder stages. In SPIM, Slot-in-Slot Attention Block (SSAB) adaptively partitions regions and generates prompt features to enhance the feature representation. In addition, a Slot Decoder $(\\mathbf{Dec}_{slot})$ is employed during training to reconstruct slot features to the target. This ensures that slot attention maps capture exposure-centric information, helping prompts learn the relevant information for enhancement within each corresponding slot." + ], + "image_footnote": [], + "bbox": [ + 130, + 85, + 468, + 334 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7493085b6a97d7e99cfe6816dda50509285af7115b32dc2ead4cd8362e9012a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 85, + 861, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "efficients to update slots, which function as queries within the attention process. Through multiple rounds of recurrent attention, Slot Attention achieves effective clustering and feature separation in OCL without the need for annotations.", + "bbox": [ + 88, + 435, + 482, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Slot Attention has been effectively applied to image classification, as seen in SCOUTER [21, 33], which uses a Slot-Attention-based classifier for explainable recognition. In segmentation, Slot Attention has proven particularly valuable for video segmentation [3, 8, 19, 29]. For example, GSANet [20] utilizes Slot Attention to separate center objects from background elements across video frames.", + "bbox": [ + 88, + 497, + 482, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In reconstruction and restoration tasks, Slot-VAE [38] integrates Slot Attention with a hierarchical Variational Autoencoder (VAE) framework, enhancing object-centric scene generation. Moreover, AID [18] employs Slot Attention to capture implicit representations of illuminant chromaticities, with each slot vector capturing the characteristics of a specific illuminant. This enables AID to generate accurate chromaticity and weight maps for each light source. In this paper, we leverage Slot Attention for partitioning in unsupervised way for the exposure correction. This allows our method to train each slot to capture feature-relevant exposure representations and effectively to cluster features implicitly for exposure correction without requiring annotations.", + "bbox": [ + 89, + 604, + 482, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Prompt-Based Learning", + "text_level": 1, + "bbox": [ + 89, + 830, + 313, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The prompt-based learning method was first introduced by [4], proposing the concept of optimizing input prompts. This approach gained popularity in the field of natural lan", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guage processing, inspiring numerous follow-up studies. CoOp [45] introduced a novel approach by treating prompt contexts as learnable parameters, moving beyond methods limited to fixed-format prompts, and demonstrating that learnable prompts can outperform handcrafted ones. On the basis of CoOp, several methods have been developed for dynamically generating suitable prompts [7, 11, 32]. For example, CODA [32] generates prompts tailored to specific inputs, while HyperPrompt [11] addresses multi-task learning through the use of a prompt generator. In the computer vision field, visual prompts involve adding trainable parameters to modify inputs for efficient model adaptation. VPT [16] demonstrated substantial performance improvements over conventional fine-tuning methods by applying visual prompts to a fixed transformer backbone. Furthermore, [2] presented visual prompts compatible with input images, specifically optimized for CLIP. For low-level vision tasks, methods like PromptIR [28] and PromptReformer [34] leverage prompts to encode degradation-specific information, guiding restoration networks to adapt to various types and intensities of degradation. In our method, we introduce exposure-specific prompts that guide feature separation and generate region-aware features specifically tailored to distinct exposure characteristics.", + "bbox": [ + 511, + 435, + 906, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 513, + 808, + 687, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Flow", + "text_level": 1, + "bbox": [ + 513, + 832, + 650, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, the proposed Exposure-slot architecture is a $U$ -shaped residual network composed of an encoder-decoder structure connected via skip connections", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17894", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Slot-in-Slot Attention Block (SSAB). For clarity, we represent the algorithm using a 2-level hierarchical structure, consisting of main- and sub-slots." + ], + "code_body": "Input: K, V, slots $^{main}$ , slots $^{sub}$ , prompts $^{main}$ , prompts $^{sub}$ \n1: function SLOT_UPDATE(slots, k, v) \n2: slots $_{prev}$ = slots \n3: slots = LayerNorm(slots) \n4: q = to_q(slots) \n5: attn = Softmax $\\left(\\frac{1}{\\sqrt{D}} k \\cdot q^T\\right)$ \n6: updates = WeightedMean(weights=attn + ε, values = v) \n7: slots = GRU(state=slot $_{prev}$ , inputs=updates) \n8: return slots, attn \n9: end function \n10: \n11: k = K, v = V \n12: for $t_{main} = 0 \\ldots T$ do \n13: slots $^{main}$ , attn $^{main}$ = SLOT_UPDATE(slots $^{main}$ , k, v) \n14: P $^{main}$ = prompts $^{main}$ × attn $^{main}$ \n15: k, v = k $\\cdot$ P $^{main}$ , v $\\cdot$ P $^{main}$ \n16: for $t_{sub} = 0 \\ldots T$ do \n17: slots $^{sub}$ , attn $^{sub}$ = SLOT_UPDATE(slots $^{sub}$ , k, v) \n18: end for \n19: P $^{sub}$ = prompts $^{sub}$ × attn $^{sub}$ \n20: k, v = k $\\cdot$ P $^{sub}$ , v $\\cdot$ P $^{sub}$ \n21: end for \n22: \n23: P $^{final}$ = P $^{main}$ $\\cdot$ P $^{sub}$ \n24: S $^{final}$ = (slots $^{main}$ × attn $^{main}$ ) $\\cdot$ (slots $^{sub}$ × attn $^{sub}$ ) \n25: return P $^{final}$ , S $^{final}$", + "bbox": [ + 89, + 122, + 906, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and Slot-Prompt Interaction Module (SPIM).", + "bbox": [ + 89, + 523, + 390, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the encoder processes the poorly exposed input image $I_{in}$ to produce a latent feature representation $F \\in \\mathbb{R}^{H \\times W \\times C}$ as $F = \\mathbf{Enc}(I_{in})$ . Then, SPIM, positioned between the encoder and decoder, refines the encoder output $F$ using input-specific, region-aware prompts guided by the Slot-in-Slot Attention mechanism. The refined latent feature $F'$ is subsequently passed to the image decoder (Dec_enhance), which processes it to produce the output image $I_{out}$ with well-enhanced exposure.", + "bbox": [ + 89, + 539, + 483, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Slot-Prompt Interaction Module (SPIM)", + "text_level": 1, + "bbox": [ + 89, + 681, + 434, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SPIM is designed as a combination of the Slot-in-Slot Attention Block (SSAB) and a subsequent cross-attention process. SPIM initially projects $F$ into query, key, and value representations, while SSAB receives the key and value to estimate the corresponding prompt features of $F$ . Specifically, SSAB hierarchically applies slot attention to achieve coarse-to-fine region separation, which supports effective exposure correction. Then the derived prompt feature serves as conditioning elements in the cross-attention step involving the query, key, and value representations.", + "bbox": [ + 89, + 703, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as illustrated in Fig. 2, the encoded feature $F \\in \\mathbb{R}^{H \\times W \\times C}$ is first flattened into $X \\in \\mathbb{R}^{HW \\times C}$ . Subsequently, $X$ is projected into query $\\mathbf{Q} \\in \\mathbb{R}^{HW \\times C}$ , key", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{K} \\in \\mathbb{R}^{HW \\times C}$ , and value $\\mathbf{V} \\in \\mathbb{R}^{HW \\times C}$ through the linear projections $W_{\\mathbf{Q}}, W_{\\mathbf{K}}$ , and $W_{\\mathbf{V}}$ as follows:", + "bbox": [ + 511, + 522, + 906, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = X W _ {\\mathbf {Q}} ^ {T}, \\quad \\mathbf {K} = X W _ {\\mathbf {K}} ^ {T}, \\quad \\mathbf {V} = X W _ {\\mathbf {V}} ^ {T}. \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 564, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within our SSAB, $\\mathbf{K}$ and $\\mathbf{V}$ are used to generate the prompt feature $\\mathbf{P}^{final}$ , which serves as a condition for the subsequent cross-attention process to produce the refined feature map $F' \\in \\mathbb{R}^{H \\times W \\times C}$ as follows:", + "bbox": [ + 511, + 593, + 906, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\prime} = r e s h a p e (s o f t m a x (\\frac {\\mathbf {Q K} ^ {T}}{\\sqrt {d}}) \\cdot (\\mathbf {V} \\cdot \\mathbf {P} ^ {f i n a l})), \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 664, + 906, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ denotes a learnable parameter that adaptively scales matrix multiplication.", + "bbox": [ + 511, + 708, + 905, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Slot-in-Slot Attention Block (SSAB)", + "text_level": 1, + "bbox": [ + 511, + 744, + 808, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the proposed Slot-in-Slot Attention Block within SPIM, we generate a refined feature map by integrating a hierarchically structured slot mechanism, which partitions regions with different exposure levels in a hierarchical and iterative manner, with learned prompts specifically tailored to distinct exposure characteristics.", + "bbox": [ + 511, + 763, + 905, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For instance, in a 2-level slot-in-slot structure, the first level partitions regions with similar exposure characteristics using a relatively smaller number of slots. Building on", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17895", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "these results, the second level employs more slots to cluster the regions more finely. By iteratively repeating this process, the sub-slots gradually learn finer details of differently exposed regions. Notably, the hierarchical structure can extend beyond two levels, enabling the identification of exposure-centric feature maps with increasing precision. Next, we introduce exposure-specific prompts that enhance the image by targeting distinct regional exposure characteristics, multiplied by attention maps from the slot attention mechanism. Since the attention maps identify regions with different exposure levels, multiplying them with exposure-specific prompts generates region-aware guidance, enabling balanced exposure correction.", + "bbox": [ + 89, + 90, + 480, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Alg. 1, the pseudo-code for this 2-level slot-in-slot structure is provided, refines key $(k)$ and value $(v)$ through two nested loops: an outer loop and an inner loop. The outer loop updates the main-slots, which are then combined with prompts to refine $k$ and $v$ . In the inner loop, the refined $k$ and $v$ are used to update the sub-slots, which are further combined with prompts to refine $k$ and $v$ with fine details. This process is repeated until the outer loop is completed. Notably, to update the slots, we employ the SLOT_UPDATE function introduced process in [24]. The key difference is that each loop iteration uses the updated $k$ and $v$ as inputs.", + "bbox": [ + 89, + 287, + 482, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Main-slot Attention In the outer loop of Alg. 1, we perform Main-slot Attention to update the main-slots (slots $^{main}$ ), which represent the first level of the hierarchical slot structure. Notably, slots $^{main} \\in \\mathbb{R}^{K^{main} \\times D_{slot}^{main}}$ , where $K^{main}$ , and $D_{slot}^{main}$ indicate the number of slots and the dimensionality of each slot, respectively, can be initialized with learnable parameters.", + "bbox": [ + 89, + 465, + 482, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, given $\\mathbf{K}$ and $\\mathbf{V}$ from the Encoder, assigned to $k$ and $v$ , respectively, we update the slots using the SLOT_UPDATE function as follows:", + "bbox": [ + 89, + 571, + 482, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s l o t s} ^ {\\text {m a i n}}, \\mathbf {a t t n} ^ {\\text {m a i n}} = \\mathbf {S L O T} _ {-} \\mathbf {U P D A T E} (\\mathbf {s l o t s} ^ {\\text {m a i n}}, k, v), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 625, + 482, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{attn}^{main}$ represents the attended region of each slots $^{main}$ . Then, the attention map $\\mathbf{attn}^{main}$ is used to produce prompt feature $\\mathbf{P}^{main}$ as:", + "bbox": [ + 89, + 656, + 482, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {\\text {m a i n}} = \\text {p r o m p t s} ^ {\\text {m a i n}} \\times \\text {a t t n} ^ {\\text {m a i n}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 708, + 480, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}^{\\text{main}} \\in \\mathbb{R}^{K^{\\text{main}} \\times D_{\\text{slot}}^{\\text{main}}}$ denotes learnable prompts tailored to each exposure characteristics. Thus, by multiplying the exposure-specific prompts with the attention map containing distinct exposure region information, our prompt feature $\\mathbf{P}^{\\text{main}}$ effectively encapsulates exposure-centric and region-aware information. Finally, prompt feature $\\mathbf{P}^{\\text{main}}$ is used to update the key and value as:", + "bbox": [ + 89, + 729, + 482, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nk, v = k \\cdot \\mathbf {P} ^ {\\text {m a i n}}, v \\cdot \\mathbf {P} ^ {\\text {m a i n}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 852, + 480, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the resulting $k$ and $v$ are used as inputs for the subsequent inner loop process, referred to as Sub-slot Attention.", + "bbox": [ + 89, + 875, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sub-slot Attention Within the inner loop of Alg. 1, we perform Sub-slot Attention to update the sub-slots (slot $s u b$ ), which represent the second level of hierarchical structure. Particularly, slots $s u b \\in \\mathbb{R}^{K^{s u b} \\times D_{s l o t}^{s u b}}$ , where slots $s u b$ are also initialized using learnable parameters. Here, $K^{s u b}$ denotes the number of sub-slots, $D_{s l o t}^{s u b}$ denotes their dimensionality, and $K^{s u b}$ is configured to ensure finer feature separation.", + "bbox": [ + 511, + 90, + 903, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, the updated $k$ and $v$ from the Main-slot Attention in Eq. 5 are used to update slots $^{sub}$ via the SLOT_UPDATE function. Similarly to Eq. 3, Sub-slot Attention is updated as follows:", + "bbox": [ + 511, + 214, + 903, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s l o t s} ^ {\\text {s u b}}, \\text {a t t n} ^ {\\text {s u b}} = \\text {S L O T} _ {-} \\text {U P D A T E} (\\text {s l o t s} ^ {\\text {s u b}}, k, v), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 285, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{attn}^{sub}$ represents the attended region of each slots $^{sub}$ , and this update process is repeated for $T$ iterations. Subsequently, $\\mathbf{attn}^{sub}$ is used to obtain sub-prompt feature $\\mathbf{P}^{sub}$ from the Sub-slot Attention as:", + "bbox": [ + 511, + 313, + 903, + 373 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {s u b} = \\text {p r o m p t s} ^ {s u b} \\times \\text {a t t n} ^ {s u b}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 383, + 903, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}^{\\mathrm{sub}} \\in \\mathbb{R}^{K^{\\mathrm{sub}} \\times D_{\\mathrm{slot}}^{\\mathrm{sub}}}$ is also learnable parameters tailored to each exposure characteristic in Sub-slot Attention. Finally, by multiplying $\\mathbf{P}^{\\mathrm{sub}}$ with $k$ and $v$ , we transfer information from the Sub-slot Attention for the next iteration of the Main-slot Attention as:", + "bbox": [ + 511, + 412, + 903, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nk, v = k \\cdot \\mathbf {P} ^ {s u b}, v \\cdot \\mathbf {P} ^ {s u b}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 501, + 903, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize, the result of the Main-slot Attention is integrated into the Sub-slot Attention, and the output from the Sub-slot Attention is iteratively fed back into the Main-slot Attention loop. This complementary updating process is repeated $T$ times. At the end of SSAB, the final outputs $\\mathbf{P}^{final}$ and $\\mathbf{S}^{final}$ are obtained as follows:", + "bbox": [ + 511, + 531, + 903, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {P} ^ {f i n a l} = \\mathbf {P} ^ {m a i n} \\cdot \\mathbf {P} ^ {s u b}, \\\\ \\mathbf {S} ^ {\\text {f i n a l}} = \\left(\\mathbf {s l o t s} ^ {\\text {m a i n}} \\times \\mathbf {a t t n} ^ {\\text {m a i n}}\\right) \\cdot \\left(\\mathbf {s l o t s} ^ {\\text {s u b}} \\times \\mathbf {a t t n} ^ {\\text {s u b}}\\right). \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 628, + 903, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\mathbf{P}^{final}$ is integrated with $\\mathbf{Q}$ , $\\mathbf{K}$ , and $\\mathbf{V}$ as described in Eq. 2, and is used to decode the final output. Meanwhile $\\mathbf{S}^{final}$ serves as an input to a separate decoder to facilitate slot training.", + "bbox": [ + 511, + 678, + 903, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.Decoder Process", + "text_level": 1, + "bbox": [ + 511, + 748, + 674, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The refined feature $F'$ and slots $S^{final}$ from SPIM are used as inputs for the final decoding procedures. Specifically, our Exposure-slot uses the enhanced feature $F'$ to predict an enhanced image through the decoder, as follows:", + "bbox": [ + 511, + 768, + 903, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {o u t} = \\operatorname {D e c} _ {\\text {e n h a n c e}} \\left(F ^ {\\prime}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 842, + 903, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Dec}_{\\text{enhance}}$ is image decoder and $I_{out}$ is the final exposure-enhanced output.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17896", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cd420e70bd6f84a7a28ad8ab20cd915288fc6ce16619c5f8dbbad03cfa04e786.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative results on SICE [5], MSEC [1], and LCDP [35] in terms of PSNR↑/SSIM↑. The best score is displayed in red, the second in blue. Additionally, the number of parameters (#Params) required for inference is also specified. Exposure-slot is lightweight compared to the previous SOTA approach [22] while consistently outperforming it across all datasets on average." + ], + "table_body": "
Model#Params (M)SICE [5]MSEC [1]LCDP [35] Avg.
UnderOverAvg.UnderOverAvg.
CLAHE [46]-12.69 / 0.503710.21 / 0.487811.45 / 0.494216.77 / 0.621114.45 / 0.584215.38 / 0.599016.33 / 0.6420
RetinexNet [40]0.84012.94 / 0.517112.87 / 0.525212.90 / 0.521212.13 / 0.620910.47 / 0.595311.14 / 0.604819.25 / 0.7041
ZeroDCE [9]0.07916.92 / 0.63307.11 / 0.429212.02 / 0.531114.55 / 0.588710.40 / 0.514212.06 / 0.544112.59 / 0.6530
RUAS [23]0.00216.63 / 0.55894.54 / 0.319610.59 / 0.439313.43 / 0.68076.39 / 0.46559.20 / 0.551513.76 / 0.6060
SCI [26]0.00117.86 / 0.64014.45 / 0.362912.49 / 0.50519.97 / 0.66815.84 / 0.51907.49 / 0.578611.87 / 0.5234
MSEC [1]7.04019.62 / 0.651217.59 / 0.656018.58 / 0.653620.52 / 0.812919.79 / 0.815620.08 / 0.821020.38 / 0.7800
LCDPNet [35]0.96017.45 / 0.562217.04 / 0.646317.25 / 0.604322.35 / 0.865022.17 / 0.847622.30 / 0.855223.24 / 0.8420
ECLNet [14]0.01822.05 / 0.689319.25 / 0.687220.65 / 0.686122.37 / 0.856622.70 / 0.867322.57 / 0.863122.44 / 0.8061
FECNet [13]0.15022.01 / 0.673719.91 / 0.696120.96 / 0.684922.96 / 0.859823.22 / 0.874823.12 / 0.868822.41 / 0.8402
DRBN-ENC [12]0.58021.89 / 0.707119.09 / 0.722920.49 / 0.715022.72 / 0.859022.11 / 0.852122.35 / 0.853022.09 / 0.8271
MSEC+DA [39]7.04020.94 / 0.754617.49 / 0.664019.22 / 0.709321.53 / 0.859021.55 / 0.875021.54 / 0.867021.05 / 0.8119
ECLNet+ERL [15]0.01822.14 / 0.690819.47 / 0.698220.81 / 0.694522.90 / 0.862422.58 / 0.867622.70 / 0.865522.63 / 0.8096
PromptIR [28]34.16422.51 / 0.695519.29 / 0.684920.90 / 0.690215.80 / 0.739116.73 / 0.785216.36 / 0.766823.49 / 0.8513
CSEC [22]1.36420.79 / 0.703120.02 / 0.709320.41 / 0.706222.18 / 0.850222.69 / 0.866222.73 / 0.863823.63 / 0.8550
Exposure-slot1.22923.85 / 0.709221.77 / 0.737522.81 / 0.723923.09 / 0.860123.24 / 0.876223.18 / 0.869724.03 / 0.8592
", + "bbox": [ + 91, + 88, + 906, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meanwhile, $\\mathbf{S}^{final}$ is decoded into a pseudo-corrected image using the slot reconstruction decoder:", + "bbox": [ + 89, + 367, + 482, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {p s e u d o} = \\mathbf {D e c} _ {s l o t} \\left(\\mathbf {S} ^ {f i n a l}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 402, + 480, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{Dec}_{slot}$ represents the slot reconstruction decoder and $I_{pseudo}$ is the resulting pseudo-corrected RGB image. This training strategy encourages each slot to focus on exposure-centric learning, similar to the set prediction approach used in [24]. It is important to note that the slot reconstruction decoder is used only during training and is not required during inference.", + "bbox": [ + 89, + 426, + 483, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Loss functions", + "text_level": 1, + "bbox": [ + 89, + 539, + 238, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Exposure-slot is trained using two loss functions: image enhancement loss $\\mathcal{L}_{\\text {enhance }}$ and slot reconstruction loss $\\mathcal{L}_{\\text {slot }}$ . The overall training objective is to minimize the combined loss, promoting both accurate image enhancement and effective slot-based reconstruction.", + "bbox": [ + 89, + 561, + 483, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Enhancement loss. The image enhancement loss $\\mathcal{L}_{\\text {enhance }}$ is defined using the $L_{1}$ distance between the exposure enhanced image $I_{\\text {out }}$ and the ground truth image $I_{\\text {gt }}$ as follows:", + "bbox": [ + 89, + 654, + 483, + 714 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n h a n c e}} = \\left\\| I _ {\\text {o u t}} - I _ {\\text {g t}} \\right\\| _ {1}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 720, + 482, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Slot Reconstruction loss. The slot reconstruction loss $\\mathcal{L}_{slot}$ is similarly defined using the $L_{1}$ distance, but between the pseudo-corrected image $I_{pseudo}$ , generated by the slot reconstruction decoder, and the ground truth image $I_{gt}$ , as follows:", + "bbox": [ + 89, + 753, + 482, + 827 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s l o t}} = \\left\\| I _ {\\text {p s e u d o}} - I _ {\\text {g t}} \\right\\| _ {1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 829, + 480, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final objective function $\\mathcal{L}_{final}$ is defined as the sum of the two losses:", + "bbox": [ + 89, + 848, + 482, + 877 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {e n h a n c e}} + \\mathcal {L} _ {\\text {s l o t}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 885, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 366, + 645, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 391, + 702, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We trained our model using the Adam optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ using a patch size of $256 \\times 256$ and a batch size of 16. The learning rate was set to $2 \\times 10^{-4}$ , and training was conducted for 500 epochs. Additionally, the parameters $K^{main}$ and $K^{sub}$ were set to 3 and 7, respectively.", + "bbox": [ + 511, + 412, + 906, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Comparative methods. Our training and benchmarking settings align with established standards for existing exposure correction tasks [13, 14, 22, 35]. We train our network on three multi-exposure datasets: Single Image Contrast Enhancement (SICE) [5], Multi-Scale Exposure Correction (MSEC) [1], and LCDP [35].", + "bbox": [ + 511, + 522, + 905, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our Exposure-slot to existing state-of-the-art exposure correction methods, including LCPDNet [35], ERL [15], ENC [12], DA [39], ECLNet [14], FECNet [13], and CSEC [22]. In PromptIR [28], the number of prompts is set to match the number of exposure values in each dataset: 2 for SICE, 5 for MSEC, and 2 for LCDP. Evaluations are conducted using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics.", + "bbox": [ + 511, + 613, + 905, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Evaluation", + "text_level": 1, + "bbox": [ + 511, + 742, + 736, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the performance of our method on three representative multi-exposure datasets: SICE [5], MSEC [1], and LCDP [35]. On the SICE dataset, our approach achieves the highest performance overall, except for one SSIM value in the under-exposed condition, where it ranks second. Similarly, for the MSEC dataset, Exposure-slot consistently outperforms previous methods in PSNR and SSIM, achieving top scores in all cases except for under-exposure in SSIM. On the LCDP dataset,", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17897", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba670b43b7537856e6885fb394c266099a0fc24e94cb56f7d28f24868741e703.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 225, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dad1efa62d23312b168e065c5935e798aa6239a9521753cf4875b3963cbf380.jpg", + "image_caption": [ + "FECNet" + ], + "image_footnote": [], + "bbox": [ + 227, + 88, + 362, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d4ec2cb75a0e1e3f0b703d99ebb245de48981e003cec059ebb17fa22c4f6fab.jpg", + "image_caption": [ + "DRBN-ENC" + ], + "image_footnote": [], + "bbox": [ + 364, + 88, + 500, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/245d1b303be3e1b67351a352f979c2c7a333dca2702abaa695e28d70906ef4bd.jpg", + "image_caption": [ + "CSEC" + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 632, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad4773be4e0498c0114c5c21680f50d3d7743114f415a62c1cc879781abe72f0.jpg", + "image_caption": [ + "Exposure-slot" + ], + "image_footnote": [], + "bbox": [ + 635, + 88, + 769, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e7b96fef407037412207c9d6a19689e03a7e4f4c4a17af58d7d676290527755.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 771, + 88, + 903, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3766b1f22c0bcf28b2ed52149cf91edf6595048fe24c100e9dea9fd8cabb6105.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison (FECNet [13], ENC [12], CSEC [22]) on the SICE [5] and MSEC [1] dataset. Examples of images enhanced from under-exposed condition (Top) and images enhanced from over-exposed condition (Bottom).", + "Figure 4. Visualization of slot attention maps (attn) of main- and sub-slot. It demonstrates how exposure-centric characteristics are effectively partitioned and refined through the slot attention mechanisms, transforming the input image into the corrected output." + ], + "image_footnote": [], + "bbox": [ + 93, + 282, + 903, + 366 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5c2d2f822966dadae820dfad26c2d2568237ec372fc9d76b95a9faae2c773c1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelΔEB2000↓ [31]ΔEab↓ [30]
ZeroDCE [9]23.7829.31
RUAS [23]29.5937.05
SCI [26]25.6431.22
ECLNet [14]9.1511.58
FECNet [13]8.8511.19
DRBN-ENC [12]8.6811.02
PromptIR [28]9.2711.59
CSEC [22]8.7211.39
Exposure-slot6.948.89
", + "bbox": [ + 99, + 406, + 473, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparisons with color difference metrics, $\\Delta {E}_{2000} \\downarrow$ and $\\Delta {E}_{ab} \\downarrow$ ,on the SICE [5] dataset. The highest score is high- lighted in red, while the second-highest is marked in blue.", + "bbox": [ + 89, + 555, + 482, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method also demonstrates superior performance, surpassing CSEC [22]. Compared to previous state-of-the-art methods [12, 13], our approach shows a notable gain of 1.85 dB in PSNR on the SICE dataset and 0.4 dB on the LCDP dataset compared to CSEC [22], highlighting its strong performance advantage.", + "bbox": [ + 89, + 611, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig.3 shows qualitative comparisons on the SICE [5] and MSEC [1] datasets, demonstrating the performance of our model against other exposure correction methods. Under under-exposed conditions, Exposure-slot achieves superior color fidelity and detail recovery, closely matching the ground truth without excessive brightening or color distortion, unlike other models [13, 22]. In over-exposed areas, the zoomed-in views reveal that other models, particularly CSEC [22], produce artifacts and color inconsistencies due to over-correction. In contrast, Exposure-slot maintains stable color restoration and texture enhancement, preserving natural details and color harmony. These results highlight Exposure-slot's capability to handle diverse exposure con", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/835e3d8d76172bf4dbc130c189f03c7ad86b41e7ea8bd1efa402f09ed6d671c7.jpg", + "image_caption": [ + "Figure 5. t-SNE results of features before and after prompts. Features from the same sub-slot are represented by the same color. With prompts, the features exhibit clearer separation." + ], + "image_footnote": [], + "bbox": [ + 555, + 406, + 864, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ditions, achieving state-of-the-art performance in complex exposure correction tasks.", + "bbox": [ + 511, + 580, + 903, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, to evaluate color correction performance, we conduct a comparison using $\\Delta E_{2000}$ [31] and $\\Delta E_{ab}$ [30] metrics in the LAB color space. Table 2 presents the results, demonstrating that our approach also excels in color correction. Exposure-slot demonstrates a performance improvement of 1.78 in $\\Delta E_{2000}$ and 2.13 in $\\Delta E_{ab}$ compared to previous state-of-the-art methods.", + "bbox": [ + 511, + 612, + 905, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 visualizes the attention maps for main and sub-slots predicted by SSAB. These results demonstrate that each map effectively generates prompts region maps, and Exposure-slot leverages this partitioning information to achieve robust enhancement across different exposure conditions of the same scene. In Fig. 5, the t-SNE visualization of $V \\cdot P^{final}$ in Eq. 2, demonstrates that our prompts improve clustering and effectively facilitate feature separation. Notably, the improved feature identification within each slot suggests that the model effectively captures exposure-specific or region-based information. For further visualizations of attention maps and additional details on", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17898", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27f5f3de9fb3ab137342016cc8714df99f761f5a49f8390e570b5f29c933d8d5.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 218, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33bbd8ada2de18220583b532fcfec2f05c70ebe8fa5d946b21c6c14be10d7784.jpg", + "image_caption": [ + "w/o Prompts" + ], + "image_footnote": [], + "bbox": [ + 223, + 88, + 349, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1f7df9028430fda4f3c9e2158169fffef59860a2618f6a5c830c87921f58dd5.jpg", + "image_caption": [ + "w/o SSAB" + ], + "image_footnote": [], + "bbox": [ + 354, + 88, + 480, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13bd8490eba7320186ea4b188ed2a37c62f3cb5ba7e3550c19dbc8659b8fb905.jpg", + "image_caption": [ + "w/o Decslot" + ], + "image_footnote": [], + "bbox": [ + 91, + 157, + 218, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74c9e209e38e658f31f26d51ba5eaff63c6b094a88b305e43623f51b6fa4642d.jpg", + "image_caption": [ + "Exposure-slot" + ], + "image_footnote": [], + "bbox": [ + 223, + 157, + 349, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84153e20cf1e9f04031ee208cf16656b19f72a8e627c6b441fe702cc1a94887c.jpg", + "image_caption": [ + "CSEC", + "Figure 6. The visualization of the ablation study. From the top left, the images correspond to (e), (f), (g), and (h) in the Table 3, with the previous SOTA method, CSEC [22], included for comparison." + ], + "image_footnote": [], + "bbox": [ + 354, + 157, + 480, + 213 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8946ae756b543ab7d9f24141b5612f70bc65a8dd2386b5aadc3a21bd1e00550a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDecslotSSAB (2-level)PromptsPSNR↑/SSIM↑
(a)21.92/0.7091
(b)✓22.02/0.7154
(c)✓21.76/0.6998
(d)✓22.09/0.7062
(e)✓✓22.10/0.7065
(f)✓✓22.09/0.7210
(g)✓✓21.95/0.7149
(h)✓✓✓22.81/0.7239
", + "bbox": [ + 91, + 277, + 480, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "t-SNE, please refer to the supplementary material.", + "bbox": [ + 89, + 420, + 421, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 444, + 243, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to assess the effectiveness of the Exposure-slot, as well as the model configuration and structural levels.", + "bbox": [ + 89, + 465, + 482, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Proposed Modules. This section evaluates the effectiveness of SSAB, $\\mathbf{Dec}_{slot}$ , and prompts through ablation studies. Table 3 shows results for each ablation case on the SICE [5] dataset. Specifically, configurations without SSAB (2-level) refer to an SSAB with a single level consisting of 7 slots (i.e. main-slots only). Configurations without Prompts directly use the slot feature $\\mathbf{S}^{final}$ for cross-attention in Eq. 2.", + "bbox": [ + 89, + 516, + 482, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, a comparison between cases with and without $\\mathbf{Dec}_{slot}$ shows consistent performance improvements when $\\mathbf{Dec}_{slot}$ is included. This result suggests that training $\\mathbf{Dec}_{slot}$ helps the slots more accurately identify and represent regions with distinct exposure characteristics, leading to enhanced performance. When employing SSAB (Model (e) and (h)) also shows significant performance improvement when used together with $\\mathbf{Dec}_{slot}$ . Regarding prompts, although Model (d) shows a slight decrease in SSIM compared to Model (a), all other cases utilizing prompts demonstrate improved performance. Notably, our full model (h) achieves the highest performance, underscoring the importance of each component in enhancing model performance. In Fig. 6, we provide a part of ablations as visual results within state-of-the-art method CSEC [22].", + "bbox": [ + 89, + 638, + 482, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Investigation of Model Configuration. Table 4 presents ablation studies on the number of main-slots ( $K^{main}$ ), sub-", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ce61d466fefa68a4326083173f24c14e64620595a917b67ae9fcda2c373f0ef2.jpg", + "table_caption": [ + "Table 3. Ablation studies on proposed modules." + ], + "table_footnote": [], + "table_body": "
ModelKmainKsubTPSNR↑SSIM↑Time (S)
(a)37322.810.72360.06759
(b)27322.610.72450.06759
(c)47322.360.72010.06759
(d)36321.920.70380.06759
(e)38322.780.72410.06759
(f)37221.150.70580.06747
(g)37421.870.72220.06996
", + "bbox": [ + 514, + 88, + 905, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3cac51099489f85d26b2135968740ab0fb25762ad384b8e8dc22e7c66dd9b994.jpg", + "table_caption": [ + "Table 4. Investigation of the number of slots ( $K^{main}$ , $K^{sub}$ ) and number of iterations ( $T$ ) on the SICE [5] dataset." + ], + "table_footnote": [], + "table_body": "
ModelKmainK*sub-1K*sub-2PSNR↑SSIM↑ΔE2000↓[31]ΔEab↓[30]
1-level3--22.020.71317.129.41
2-level37-22.810.72366.948.89
3-level371023.060.73066.848.69
", + "bbox": [ + 514, + 244, + 903, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on $n$ -level SSAB $\\left( {n = 1,2,3}\\right)$", + "bbox": [ + 540, + 287, + 877, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "slots $(K^{sub})$ , and iterations $(T)$ within SSAB to evaluate their impact on performance. The results include PSNR and SSIM metrics on the SICE [5] dataset, with runtime measurements taken on a single $844 \\times 1500$ RGB image using an NVIDIA RTX 4090 GPU. In this analysis, we vary $K^{main}$ , $K^{sub}$ , and $T$ by incrementing or decrementing each parameter by 1. Based on the PSNR values, the configuration of $K^{main} = 3$ , $K^{sub} = 7$ , and $T = 3$ yields the highest performance and is therefore selected for our method.", + "bbox": [ + 511, + 314, + 906, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Structural Levels. In Sec.3, we use a 2-level SSAB structure as the default configuration. To further investigate the potential of SSAB, we perform experiments using both 1-level and 3-level SSAB structures on the SICE [5] dataset, as summarized in Table 5. In the 3-level configuration, we set $K^{sub-2} = 10$ for the number of second sub-slots, resulting in PSNR and SSIM improvements of 0.25 and 0.07 compared to 2-level SSAB, respectively. These results indicate that adding more SSAB levels can enhance performance. While adding additional levels improves performance, it also leads to an exponential increase in time complexity. Therefore, considering this trade-off, we adopt the 2-level SSAB as the default structure.", + "bbox": [ + 511, + 465, + 906, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Exposure-slot, a novel framework for exposure correction that integrates slot-in-slot attention with learnable prompts to achieve precise, exposure-centric feature learning. By hierarchically clustering and refining exposure regions, Exposure-slot facilitates accurate, region-aware adjustments and achieves state-of-the-art performance on multiple multi-exposure benchmarks. Our approach demonstrates significant improvements in both quantitative and qualitative metrics, highlighting the effectiveness of structured attention mechanisms for challenging exposure correction scenarios. Furthermore, this work advances exposure correction and paves the way for exploring slot-based architectures in other low-level vision tasks.", + "bbox": [ + 511, + 704, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17899", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 236, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2023-00222776), and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06, Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Ministry of Science and ICT (No.RS-2020-II201373, Artificial Intelligence Graduate School Program [Hanyang University]).", + "bbox": [ + 89, + 112, + 485, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 337, + 187, + 353 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Afifi, Konstantinos G Derpanis, Bjorn Ommer, and Michael S Brown. Learning multi-scale photo exposure correction. In CVPR, 2021. 1, 2, 6, 7", + "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3", + "[3] Ondrej Biza, Sjoerd Van Steenkiste, Mehdi S. M. Sajjadi, Gamaleldin F. Elsayed, Aravindh Mahendran, and Thomas Kipf. Invariant slot attention: object discovery with slot-centric reference frames. In ICML, 2023. 3", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3", + "[5] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27, 2018. 2, 6, 7, 8", + "[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2", + "[7] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, 2023. 3", + "[8] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In CVPR, 2024. 3", + "[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 6, 7", + "[10] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 2016. 1", + "[11] Yun He, Steven Zheng, Yi Tay, Jai Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, YaGuang Li, Zhao Chen, Donald Met" + ], + "bbox": [ + 93, + 362, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "zler, et al. Hyperprompt: Prompt-based task-conditioning of transformers. In ICML. PMLR, 2022. 3", + "[12] Jie Huang, Yajing Liu, Xueyang Fu, Man Zhou, Yang Wang, Feng Zhao, and Zhiwei Xiong. Exposure normalization and compensation for multiple-exposure correction. In CVPR, 2022. 1, 2, 6, 7", + "[13] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In ECCV, 2022. 1, 2, 6, 7", + "[14] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In ACMMM, 2022. 1, 2, 6, 7", + "[15] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In CVPR, 2023. 2, 6", + "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3", + "[17] Donggoo Jung, Daehyun Kim, and Tae Hyun Kim. Continuous exposure learning for low-light image enhancement using Neural ODEs. In ICLR, 2025. 2", + "[18] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In CVPR, 2024. 3", + "[19] Markus Krimmel, Jan Achterhold, and Joerg Stueckler. Attention normalization impacts cardinality generalization in slot attention. Transactions on Machine Learning Research, 2024. 3", + "[20] Minhyeok Lee, Suhwan Cho, Dogyoon Lee, Chaewon Park, Jungho Lee, and Sangyoun Lee. Guided slot attention for unsupervised video object segmentation. In CVPR, 2024. 3", + "[21] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In CVPR, 2021. 3", + "[22] Yiyu Li, Ke Xu, Gerhard Petrus Hancke, and Rynson WH Lau. Color shift estimation-and-correction for image enhancement. In CVPR, 2024. 1, 2, 6, 7, 8", + "[23] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 6, 7", + "[24] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 2, 5, 6", + "[25] Minh-Thang Luong. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 2", + "[26] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 6, 7", + "[27] Ntumba Elie Nsampi, Zhongyun Hu, and Qing Wang. Learning exposure correction via consistency modeling. In BMVC, 2021. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17900", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. PromptIR: Prompting for all-in-one image restoration. In NeurIPS, 2023. 3, 6, 7", + "[29] Rishav Pramanik, José-Fabian Villa-Vásquez, and Marco Pedersoli. Masked Multi-Query slot attention for unsupervised object discovery. In IJCNN, 2024. 3", + "[30] Gaurav Sharma and Raja Bala. Digital color imaging handbook. CRC press, 2017. 7, 8", + "[31] Gaurav Sharma, Wencheng Wu, and Edul N Dalal. The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application, 2005. 7, 8", + "[32] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. CODA-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, 2023. 3", + "[33] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, 2023. 3", + "[34] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 3", + "[35] Haoyuan Wang, Ke Xu, and Rynson WH Lau. Local color distributions prior for image enhancement. In ECCV. Springer, 2022. 1, 2, 6", + "[36] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 2", + "[37] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 1", + "[38] Yanbo Wang, Letao Liu, and Justin Dauwels. Slot-vae: Object-centric scene generation with slot attention. In ICML, 2023. 3", + "[39] Yang Wang, Long Peng, Liang Li, Yang Cao, and Zheng-Jun Zha. Decoupling-and-aggregating for image exposure correction. In CVPR, 2023. 1, 6", + "[40] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep Retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 6", + "[41] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. URetinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 2", + "[42] Mingde Yao, Jie Huang, Xin Jin, Ruikang Xu, Shenglong Zhou, Man Zhou, and Zhiwei Xiong. Generalized lightness adaptation with channel selective normalization. In CVPR, 2023. 2", + "[43] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In ACMM, 2019. 2", + "[44] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129, 2021. 2" + ], + "bbox": [ + 89, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3", + "[46] Karel Zuiderveld. Contrast limited adaptive histogram equalization. Graphics gems, 1994. 6" + ], + "bbox": [ + 514, + 92, + 903, + 162 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17901", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_model.json b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_model.json new file mode 100644 index 0000000000000000000000000000000000000000..61e578bbced35e19d0f1d5b161a4e545e2ce0f56 --- /dev/null +++ b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.186, + 0.131, + 0.817, + 0.177 + ], + "angle": 0, + "content": "Exposure-slot: Exposure-centric representations learning with Slot-in-Slot Attention for Region-aware Exposure Correction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.203, + 0.78, + 0.24 + ], + "angle": 0, + "content": "Donggoo Jung\\*, Daehyun Kim\\*, Guanghui Wang, Tae Hyun Kim† Hanyang University, Toronto Metropolitan University" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.242, + 0.788, + 0.258 + ], + "angle": 0, + "content": "{dgjung, daehyun, taehyunkim}@hanyang.ac.kr, {wangcs}@torontomu.ca" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.486, + 0.657 + ], + "angle": 0, + "content": "Image exposure correction enhances images captured under diverse real-world conditions by addressing issues of under- and over-exposure, which can result in the loss of critical details and hinder content recognition. While significant advancements have been made, current methods often fail to achieve optimal feature learning for effective correction. To overcome these challenges, we propose Exposure-slot, a novel framework that integrates a prompt-based slot-in-slot attention mechanism to cluster exposed feature regions and learn exposure-centric features for each cluster. By extending the Slot Attention algorithm with a hierarchical structure, our approach progressively clusters features, enabling precise and region-aware correction. In particular, learnable prompts tailored to exposure characteristics of slots further enhance feature quality, adapting dynamically to varying conditions. Our method delivers superior performance on benchmark datasets, surpassing the current state-of-the-art with a PSNR improvement of over 1.85 dB on the SICE dataset and 0.4 dB on the LCDP dataset, thereby establishing a new benchmark for multi-exposure correction. The source code can be found at: https://github.com/kdhRick2222/Exposure-slot." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.685, + 0.222, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Image exposure correction enhances images captured under diverse real-world conditions. Under- or over-exposed images can appear excessively dark or bright, losing essential details and compromising accurate recognition. Despite advancements in imaging technology, robust automatic exposure correction remains a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Due to its critical importance, image exposure correction has been the focus of considerable research. Early efforts primarily treated under-exposure and over-exposure as separate issues [10, 37, 40]. More recent advancements, however, have introduced multi-exposure correction mod" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.907, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Figure 1. (Top) The Slot-in-Slot Attention mechanism hierarchically partitions exposure-aware regions using the attention maps of main and sub-slot attention, performing corrections with learnable prompts based on exposure levels. (Bottom) Comparison of the Exposure-slot with existing approaches. From left to right: ECLNet [14], FECNet [13], DRBN-ENC [12], CSEC [22] and our proposed approach, Exposure-slot." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "els capable of handling a wide range of exposure levels within a unified framework. For instance, MSEC [1] introduced an approach that considers both color and detail enhancement while performing multi-exposure correction in an end-to-end manner, and provided a dataset of images with different exposure errors for both training and evaluation. In addition, LCDPNet [35] provided another dataset of diverse scenes, covering both over-exposure and under-exposure, and incorporates retinax theory to account for local color distributions within images. Recent approaches increasingly focus on the separate processing of features based on their physical characteristics, such as exposure, frequency, contrast, color, and details. FECNet [13] uses Fourier transform to facilitate interactions between local spatial feature information and global frequency information, while ECLNet [14] uses a bilateral activation mechanism to process over- and under-exposed regions independently. In addition, DA [39] introduces a decoupling and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.363, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17892" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "aggregation scheme that separately enhances image details and contrast." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.482, + 0.228 + ], + "angle": 0, + "content": "However, relying on separating image features based on these physical properties may not ensure the optimal features for effective exposure correction. For example, at the bottom of Fig. 1, we visualize this limitation, showing that previous works [13-15, 22] often generate inappropriate artifacts in sky regions and struggle to accurately adjust sunset exposure levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.412 + ], + "angle": 0, + "content": "To overcome these challenges, we propose a novel prompt-based slot-in-slot attention mechanism that clusters regions with similar exposure levels, enabling exposure-centric feature learning within each cluster. We utilize Slot Attention [24], designed to cluster features based on object representations, to identify distinct exposure regions within an image. In our approach, slots represent features grouped by exposure levels, and the corresponding attention maps act as exposure-aware region maps, guiding the grouping of features according to exposure characteristics. Therefore, these maps highlight regions based on exposure levels, enabling targeted exposure correction for each area." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.595 + ], + "angle": 0, + "content": "To support effective correction, we introduce learnable prompt vectors that operate based on exposure-aware region maps, which are attention maps of each slot. These prompts adapt to the characteristics of each region defined by the attention map, helping the model learn targeted correction to improve under- or over-exposed areas. Specifically, we introduce Slot-Prompt Interaction Module (SPIM), which combines slot and prompt information using cross-attention. The cross-attention component of SPIM uses the prompts learned through Slot-in-Slot Attention as conditioning factors, enhancing the interaction between features and prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Our approach builds upon standard slot attention with a hierarchical design, called Slot-in-Slot Attention, which processes slots in levels. This mechanism operates across multiple structure levels: the first level coarsely partitions regions based on distinct exposure characteristics, while subsequent levels progressively refine these regions with more partitions. Each slot captures exposure-specific regional information, distinguishing areas with varying exposure levels to enhance feature separation and improve the precision of local details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The overall framework of the proposed approach is illustrated at the top of Fig. 1. The Slot-in-Slot Attention is applied to intermediate features between the encoder and decoder, generating attention maps for slots that are hierarchically partitioned according to exposure characteristics, as depicted in Fig. 1. The generated attention maps are then used as weight maps for learnable prompts, resulting in refined intermediate features that are subsequently fed into the decoder. We refer to our model as Exposure-Slot, as it is the first model to use Slot Attention for unsupervised feature" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "partitioning based on exposure levels in exposure correction, and the first to apply region-aware prompts for feature enhancement. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.14, + 0.905, + 0.185 + ], + "angle": 0, + "content": "- Exposure-slot is the first approach to leverage Slot Attention mechanism for optimized exposure-specific feature partitioning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.185, + 0.905, + 0.216 + ], + "angle": 0, + "content": "- We introduce the slot-in-slot attention that enables sophisticated feature partitioning and learning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.217, + 0.905, + 0.247 + ], + "angle": 0, + "content": "- We apply exposure-aware prompts that enhance the exposure-centric characteristics of each image feature." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.247, + 0.905, + 0.291 + ], + "angle": 0, + "content": "- Exposure-slot achieves state-of-the-art results on multi-exposure benchmark datasets [1, 5, 35], setting a new standard in this field." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.14, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.306, + 0.655, + 0.322 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.332, + 0.712, + 0.348 + ], + "angle": 0, + "content": "2.1. Exposure Correction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.354, + 0.907, + 0.641 + ], + "angle": 0, + "content": "With the rise of deep neural networks (DNN), exposure correction has seen the emergence of a range of DNN-based methods leveraging diverse concepts. For under-exposed image enhancement, methods [36, 40, 41, 43, 44] inspired by retinax theory have been proposed. Retinex-based models like CMEC [27] have also incorporated attention mechanisms to address multi-exposure correction, while LCDP-Net [35] introduced a novel focus on local color distribution. To support multi-exposure correction tasks, dedicated datasets such as MSEC [1] and SICE [5] were developed for training and evaluation. MSEC further introduced a Laplacian pyramid architecture to handle varying exposure levels. ENC [12] contributed an exposure normalization module that refined feature maps by transforming diverse exposure features into an exposure-invariant feature space. CSNorm [42] also enhanced model generalization by selectively normalizing lightness-sensitive channels and ERL [15] proposed exposure relationship learning through regularization techniques for multi-exposure correction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.642, + 0.906, + 0.808 + ], + "angle": 0, + "content": "There are several recent studies that emphasize the feature separation based on their physical characteristics. ECLNet [14] used a bilateral activation mechanism to adjust processing across different exposure conditions, while FECNet [13] proposed a lightweight model that utilizes spatial-frequency interactions with a Fourier-based approach. CLODE [17] also utilizes image curve adjustment and formulates brightness changes through continuous curve map extraction using Neural ODEs [6]. The latest advancement, CSEC [22], addresses color distribution shifts by defining darkened and brightened feature maps." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.818, + 0.658, + 0.833 + ], + "angle": 0, + "content": "2.2. Slot Attention" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Slot Attention [24] is a mechanism for clustering features corresponding to slots which is the main object of object-centric learning (OCL). Specifically, this mechanism iteratively employs dot-product attention [25] with attention co" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17893" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.086, + 0.47, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.087, + 0.862, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.346, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Overview of Exposure-Slot: Exposure-Slot operates within a \\(U\\)-shaped encoder-decoder network, with the Slot-Prompt Interaction Module (SPIM) bridging the encoder and decoder stages. In SPIM, Slot-in-Slot Attention Block (SSAB) adaptively partitions regions and generates prompt features to enhance the feature representation. In addition, a Slot Decoder \\((\\mathbf{Dec}_{slot})\\) is employed during training to reconstruct slot features to the target. This ensures that slot attention maps capture exposure-centric information, helping prompts learn the relevant information for enhancement within each corresponding slot." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.436, + 0.483, + 0.496 + ], + "angle": 0, + "content": "efficients to update slots, which function as queries within the attention process. Through multiple rounds of recurrent attention, Slot Attention achieves effective clustering and feature separation in OCL without the need for annotations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.498, + 0.483, + 0.604 + ], + "angle": 0, + "content": "Slot Attention has been effectively applied to image classification, as seen in SCOUTER [21, 33], which uses a Slot-Attention-based classifier for explainable recognition. In segmentation, Slot Attention has proven particularly valuable for video segmentation [3, 8, 19, 29]. For example, GSANet [20] utilizes Slot Attention to separate center objects from background elements across video frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.483, + 0.817 + ], + "angle": 0, + "content": "In reconstruction and restoration tasks, Slot-VAE [38] integrates Slot Attention with a hierarchical Variational Autoencoder (VAE) framework, enhancing object-centric scene generation. Moreover, AID [18] employs Slot Attention to capture implicit representations of illuminant chromaticities, with each slot vector capturing the characteristics of a specific illuminant. This enables AID to generate accurate chromaticity and weight maps for each light source. In this paper, we leverage Slot Attention for partitioning in unsupervised way for the exposure correction. This allows our method to train each slot to capture feature-relevant exposure representations and effectively to cluster features implicitly for exposure correction without requiring annotations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.315, + 0.849 + ], + "angle": 0, + "content": "2.3. Prompt-Based Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The prompt-based learning method was first introduced by [4], proposing the concept of optimizing input prompts. This approach gained popularity in the field of natural lan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.799 + ], + "angle": 0, + "content": "guage processing, inspiring numerous follow-up studies. CoOp [45] introduced a novel approach by treating prompt contexts as learnable parameters, moving beyond methods limited to fixed-format prompts, and demonstrating that learnable prompts can outperform handcrafted ones. On the basis of CoOp, several methods have been developed for dynamically generating suitable prompts [7, 11, 32]. For example, CODA [32] generates prompts tailored to specific inputs, while HyperPrompt [11] addresses multi-task learning through the use of a prompt generator. In the computer vision field, visual prompts involve adding trainable parameters to modify inputs for efficient model adaptation. VPT [16] demonstrated substantial performance improvements over conventional fine-tuning methods by applying visual prompts to a fixed transformer backbone. Furthermore, [2] presented visual prompts compatible with input images, specifically optimized for CLIP. For low-level vision tasks, methods like PromptIR [28] and PromptReformer [34] leverage prompts to encode degradation-specific information, guiding restoration networks to adapt to various types and intensities of degradation. In our method, we introduce exposure-specific prompts that guide feature separation and generate region-aware features specifically tailored to distinct exposure characteristics." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.809, + 0.688, + 0.826 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.833, + 0.651, + 0.849 + ], + "angle": 0, + "content": "3.1. Overall Flow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, the proposed Exposure-slot architecture is a \\(U\\)-shaped residual network composed of an encoder-decoder structure connected via skip connections" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17894" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.092, + 0.091, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Algorithm 1 Slot-in-Slot Attention Block (SSAB). For clarity, we represent the algorithm using a 2-level hierarchical structure, consisting of main- and sub-slots." + }, + { + "type": "algorithm", + "bbox": [ + 0.09, + 0.123, + 0.907, + 0.499 + ], + "angle": 0, + "content": "Input: K, V, slots\\(^{main}\\), slots\\(^{sub}\\), prompts\\(^{main}\\), prompts\\(^{sub}\\) \n1: function SLOT_UPDATE(slots, k, v) \n2: slots\\(_{prev}\\) = slots \n3: slots = LayerNorm(slots) \n4: q = to_q(slots) \n5: attn = Softmax \\(\\left(\\frac{1}{\\sqrt{D}} k \\cdot q^T\\right)\\) \n6: updates = WeightedMean(weights=attn + ε, values = v) \n7: slots = GRU(state=slot\\(_{prev}\\), inputs=updates) \n8: return slots, attn \n9: end function \n10: \n11: k = K, v = V \n12: for \\(t_{main} = 0 \\ldots T\\) do \n13: slots\\(^{main}\\), attn\\(^{main}\\) = SLOT_UPDATE(slots\\(^{main}\\), k, v) \n14: P\\(^{main}\\) = prompts\\(^{main}\\) × attn\\(^{main}\\) \n15: k, v = k \\(\\cdot\\) P\\(^{main}\\), v \\(\\cdot\\) P\\(^{main}\\) \n16: for \\(t_{sub} = 0 \\ldots T\\) do \n17: slots\\(^{sub}\\), attn\\(^{sub}\\) = SLOT_UPDATE(slots\\(^{sub}\\), k, v) \n18: end for \n19: P\\(^{sub}\\) = prompts\\(^{sub}\\) × attn\\(^{sub}\\) \n20: k, v = k \\(\\cdot\\) P\\(^{sub}\\), v \\(\\cdot\\) P\\(^{sub}\\) \n21: end for \n22: \n23: P\\(^{final}\\) = P\\(^{main}\\) \\(\\cdot\\) P\\(^{sub}\\) \n24: S\\(^{final}\\) = (slots\\(^{main}\\) × attn\\(^{main}\\)) \\(\\cdot\\) (slots\\(^{sub}\\) × attn\\(^{sub}\\)) \n25: return P\\(^{final}\\), S\\(^{final}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.391, + 0.539 + ], + "angle": 0, + "content": "and Slot-Prompt Interaction Module (SPIM)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.676 + ], + "angle": 0, + "content": "First, the encoder processes the poorly exposed input image \\( I_{in} \\) to produce a latent feature representation \\( F \\in \\mathbb{R}^{H \\times W \\times C} \\) as \\( F = \\mathbf{Enc}(I_{in}) \\). Then, SPIM, positioned between the encoder and decoder, refines the encoder output \\( F \\) using input-specific, region-aware prompts guided by the Slot-in-Slot Attention mechanism. The refined latent feature \\( F' \\) is subsequently passed to the image decoder (Dec_enhance), which processes it to produce the output image \\( I_{out} \\) with well-enhanced exposure." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.436, + 0.699 + ], + "angle": 0, + "content": "3.2. Slot-Prompt Interaction Module (SPIM)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.856 + ], + "angle": 0, + "content": "SPIM is designed as a combination of the Slot-in-Slot Attention Block (SSAB) and a subsequent cross-attention process. SPIM initially projects \\( F \\) into query, key, and value representations, while SSAB receives the key and value to estimate the corresponding prompt features of \\( F \\). Specifically, SSAB hierarchically applies slot attention to achieve coarse-to-fine region separation, which supports effective exposure correction. Then the derived prompt feature serves as conditioning elements in the cross-attention step involving the query, key, and value representations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specifically, as illustrated in Fig. 2, the encoded feature \\(F \\in \\mathbb{R}^{H \\times W \\times C}\\) is first flattened into \\(X \\in \\mathbb{R}^{HW \\times C}\\). Subsequently, \\(X\\) is projected into query \\(\\mathbf{Q} \\in \\mathbb{R}^{HW \\times C}\\), key" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.556 + ], + "angle": 0, + "content": "\\(\\mathbf{K} \\in \\mathbb{R}^{HW \\times C}\\), and value \\(\\mathbf{V} \\in \\mathbb{R}^{HW \\times C}\\) through the linear projections \\(W_{\\mathbf{Q}}, W_{\\mathbf{K}}\\), and \\(W_{\\mathbf{V}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.565, + 0.907, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = X W _ {\\mathbf {Q}} ^ {T}, \\quad \\mathbf {K} = X W _ {\\mathbf {K}} ^ {T}, \\quad \\mathbf {V} = X W _ {\\mathbf {V}} ^ {T}. \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Within our SSAB, \\(\\mathbf{K}\\) and \\(\\mathbf{V}\\) are used to generate the prompt feature \\(\\mathbf{P}^{final}\\), which serves as a condition for the subsequent cross-attention process to produce the refined feature map \\(F' \\in \\mathbb{R}^{H \\times W \\times C}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.665, + 0.907, + 0.7 + ], + "angle": 0, + "content": "\\[\nF ^ {\\prime} = r e s h a p e (s o f t m a x (\\frac {\\mathbf {Q K} ^ {T}}{\\sqrt {d}}) \\cdot (\\mathbf {V} \\cdot \\mathbf {P} ^ {f i n a l})), \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.906, + 0.739 + ], + "angle": 0, + "content": "where \\(d\\) denotes a learnable parameter that adaptively scales matrix multiplication." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.809, + 0.761 + ], + "angle": 0, + "content": "3.2.1. Slot-in-Slot Attention Block (SSAB)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.854 + ], + "angle": 0, + "content": "In the proposed Slot-in-Slot Attention Block within SPIM, we generate a refined feature map by integrating a hierarchically structured slot mechanism, which partitions regions with different exposure levels in a hierarchical and iterative manner, with learned prompts specifically tailored to distinct exposure characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "For instance, in a 2-level slot-in-slot structure, the first level partitions regions with similar exposure characteristics using a relatively smaller number of slots. Building on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17895" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.287 + ], + "angle": 0, + "content": "these results, the second level employs more slots to cluster the regions more finely. By iteratively repeating this process, the sub-slots gradually learn finer details of differently exposed regions. Notably, the hierarchical structure can extend beyond two levels, enabling the identification of exposure-centric feature maps with increasing precision. Next, we introduce exposure-specific prompts that enhance the image by targeting distinct regional exposure characteristics, multiplied by attention maps from the slot attention mechanism. Since the attention maps identify regions with different exposure levels, multiplying them with exposure-specific prompts generates region-aware guidance, enabling balanced exposure correction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.483, + 0.455 + ], + "angle": 0, + "content": "In Alg. 1, the pseudo-code for this 2-level slot-in-slot structure is provided, refines key \\((k)\\) and value \\((v)\\) through two nested loops: an outer loop and an inner loop. The outer loop updates the main-slots, which are then combined with prompts to refine \\(k\\) and \\(v\\). In the inner loop, the refined \\(k\\) and \\(v\\) are used to update the sub-slots, which are further combined with prompts to refine \\(k\\) and \\(v\\) with fine details. This process is repeated until the outer loop is completed. Notably, to update the slots, we employ the SLOT_UPDATE function introduced process in [24]. The key difference is that each loop iteration uses the updated \\(k\\) and \\(v\\) as inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.483, + 0.572 + ], + "angle": 0, + "content": "Main-slot Attention In the outer loop of Alg. 1, we perform Main-slot Attention to update the main-slots (slots\\(^{main}\\)), which represent the first level of the hierarchical slot structure. Notably, slots\\(^{main} \\in \\mathbb{R}^{K^{main} \\times D_{slot}^{main}}\\), where \\(K^{main}\\), and \\(D_{slot}^{main}\\) indicate the number of slots and the dimensionality of each slot, respectively, can be initialized with learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.618 + ], + "angle": 0, + "content": "First, given \\(\\mathbf{K}\\) and \\(\\mathbf{V}\\) from the Encoder, assigned to \\(k\\) and \\(v\\), respectively, we update the slots using the SLOT_UPDATE function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.625, + 0.483, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s l o t s} ^ {\\text {m a i n}}, \\mathbf {a t t n} ^ {\\text {m a i n}} = \\mathbf {S L O T} _ {-} \\mathbf {U P D A T E} (\\mathbf {s l o t s} ^ {\\text {m a i n}}, k, v), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.702 + ], + "angle": 0, + "content": "where \\(\\mathbf{attn}^{main}\\) represents the attended region of each slots\\(^{main}\\). Then, the attention map \\(\\mathbf{attn}^{main}\\) is used to produce prompt feature \\(\\mathbf{P}^{main}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.709, + 0.482, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {\\text {m a i n}} = \\text {p r o m p t s} ^ {\\text {m a i n}} \\times \\text {a t t n} ^ {\\text {m a i n}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.852 + ], + "angle": 0, + "content": "where \\( \\mathbf{p}^{\\text{main}} \\in \\mathbb{R}^{K^{\\text{main}} \\times D_{\\text{slot}}^{\\text{main}}} \\) denotes learnable prompts tailored to each exposure characteristics. Thus, by multiplying the exposure-specific prompts with the attention map containing distinct exposure region information, our prompt feature \\( \\mathbf{P}^{\\text{main}} \\) effectively encapsulates exposure-centric and region-aware information. Finally, prompt feature \\( \\mathbf{P}^{\\text{main}} \\) is used to update the key and value as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.853, + 0.482, + 0.87 + ], + "angle": 0, + "content": "\\[\nk, v = k \\cdot \\mathbf {P} ^ {\\text {m a i n}}, v \\cdot \\mathbf {P} ^ {\\text {m a i n}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.906 + ], + "angle": 0, + "content": "and the resulting \\( k \\) and \\( v \\) are used as inputs for the subsequent inner loop process, referred to as Sub-slot Attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.214 + ], + "angle": 0, + "content": "Sub-slot Attention Within the inner loop of Alg. 1, we perform Sub-slot Attention to update the sub-slots (slot \\(s u b\\)), which represent the second level of hierarchical structure. Particularly, slots \\(s u b \\in \\mathbb{R}^{K^{s u b} \\times D_{s l o t}^{s u b}}\\), where slots \\(s u b\\) are also initialized using learnable parameters. Here, \\(K^{s u b}\\) denotes the number of sub-slots, \\(D_{s l o t}^{s u b}\\) denotes their dimensionality, and \\(K^{s u b}\\) is configured to ensure finer feature separation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.905, + 0.274 + ], + "angle": 0, + "content": "First, the updated \\( k \\) and \\( v \\) from the Main-slot Attention in Eq. 5 are used to update slots\\(^{sub}\\) via the SLOT_UPDATE function. Similarly to Eq. 3, Sub-slot Attention is updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.286, + 0.905, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\text {s l o t s} ^ {\\text {s u b}}, \\text {a t t n} ^ {\\text {s u b}} = \\text {S L O T} _ {-} \\text {U P D A T E} (\\text {s l o t s} ^ {\\text {s u b}}, k, v), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.314, + 0.905, + 0.374 + ], + "angle": 0, + "content": "where \\(\\mathbf{attn}^{sub}\\) represents the attended region of each slots\\(^{sub}\\), and this update process is repeated for \\(T\\) iterations. Subsequently, \\(\\mathbf{attn}^{sub}\\) is used to obtain sub-prompt feature \\(\\mathbf{P}^{sub}\\) from the Sub-slot Attention as:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.385, + 0.905, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {s u b} = \\text {p r o m p t s} ^ {s u b} \\times \\text {a t t n} ^ {s u b}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.414, + 0.905, + 0.491 + ], + "angle": 0, + "content": "where \\( \\mathbf{p}^{\\mathrm{sub}} \\in \\mathbb{R}^{K^{\\mathrm{sub}} \\times D_{\\mathrm{slot}}^{\\mathrm{sub}}} \\) is also learnable parameters tailored to each exposure characteristic in Sub-slot Attention. Finally, by multiplying \\( \\mathbf{P}^{\\mathrm{sub}} \\) with \\( k \\) and \\( v \\), we transfer information from the Sub-slot Attention for the next iteration of the Main-slot Attention as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.502, + 0.905, + 0.519 + ], + "angle": 0, + "content": "\\[\nk, v = k \\cdot \\mathbf {P} ^ {s u b}, v \\cdot \\mathbf {P} ^ {s u b}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.905, + 0.621 + ], + "angle": 0, + "content": "To summarize, the result of the Main-slot Attention is integrated into the Sub-slot Attention, and the output from the Sub-slot Attention is iteratively fed back into the Main-slot Attention loop. This complementary updating process is repeated \\( T \\) times. At the end of SSAB, the final outputs \\( \\mathbf{P}^{final} \\) and \\( \\mathbf{S}^{final} \\) are obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.629, + 0.905, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {P} ^ {f i n a l} = \\mathbf {P} ^ {m a i n} \\cdot \\mathbf {P} ^ {s u b}, \\\\ \\mathbf {S} ^ {\\text {f i n a l}} = \\left(\\mathbf {s l o t s} ^ {\\text {m a i n}} \\times \\mathbf {a t t n} ^ {\\text {m a i n}}\\right) \\cdot \\left(\\mathbf {s l o t s} ^ {\\text {s u b}} \\times \\mathbf {a t t n} ^ {\\text {s u b}}\\right). \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.905, + 0.74 + ], + "angle": 0, + "content": "Note that \\(\\mathbf{P}^{final}\\) is integrated with \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), and \\(\\mathbf{V}\\) as described in Eq. 2, and is used to decode the final output. Meanwhile \\(\\mathbf{S}^{final}\\) serves as an input to a separate decoder to facilitate slot training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.676, + 0.763 + ], + "angle": 0, + "content": "3.3.Decoder Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.905, + 0.831 + ], + "angle": 0, + "content": "The refined feature \\( F' \\) and slots \\( S^{final} \\) from SPIM are used as inputs for the final decoding procedures. Specifically, our Exposure-slot uses the enhanced feature \\( F' \\) to predict an enhanced image through the decoder, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.843, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\nI _ {o u t} = \\operatorname {D e c} _ {\\text {e n h a n c e}} \\left(F ^ {\\prime}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{Dec}_{\\text{enhance}}\\) is image decoder and \\(I_{out}\\) is the final exposure-enhanced output." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17896" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.299 + ], + "angle": 0, + "content": "
Model#Params (M)SICE [5]MSEC [1]LCDP [35] Avg.
UnderOverAvg.UnderOverAvg.
CLAHE [46]-12.69 / 0.503710.21 / 0.487811.45 / 0.494216.77 / 0.621114.45 / 0.584215.38 / 0.599016.33 / 0.6420
RetinexNet [40]0.84012.94 / 0.517112.87 / 0.525212.90 / 0.521212.13 / 0.620910.47 / 0.595311.14 / 0.604819.25 / 0.7041
ZeroDCE [9]0.07916.92 / 0.63307.11 / 0.429212.02 / 0.531114.55 / 0.588710.40 / 0.514212.06 / 0.544112.59 / 0.6530
RUAS [23]0.00216.63 / 0.55894.54 / 0.319610.59 / 0.439313.43 / 0.68076.39 / 0.46559.20 / 0.551513.76 / 0.6060
SCI [26]0.00117.86 / 0.64014.45 / 0.362912.49 / 0.50519.97 / 0.66815.84 / 0.51907.49 / 0.578611.87 / 0.5234
MSEC [1]7.04019.62 / 0.651217.59 / 0.656018.58 / 0.653620.52 / 0.812919.79 / 0.815620.08 / 0.821020.38 / 0.7800
LCDPNet [35]0.96017.45 / 0.562217.04 / 0.646317.25 / 0.604322.35 / 0.865022.17 / 0.847622.30 / 0.855223.24 / 0.8420
ECLNet [14]0.01822.05 / 0.689319.25 / 0.687220.65 / 0.686122.37 / 0.856622.70 / 0.867322.57 / 0.863122.44 / 0.8061
FECNet [13]0.15022.01 / 0.673719.91 / 0.696120.96 / 0.684922.96 / 0.859823.22 / 0.874823.12 / 0.868822.41 / 0.8402
DRBN-ENC [12]0.58021.89 / 0.707119.09 / 0.722920.49 / 0.715022.72 / 0.859022.11 / 0.852122.35 / 0.853022.09 / 0.8271
MSEC+DA [39]7.04020.94 / 0.754617.49 / 0.664019.22 / 0.709321.53 / 0.859021.55 / 0.875021.54 / 0.867021.05 / 0.8119
ECLNet+ERL [15]0.01822.14 / 0.690819.47 / 0.698220.81 / 0.694522.90 / 0.862422.58 / 0.867622.70 / 0.865522.63 / 0.8096
PromptIR [28]34.16422.51 / 0.695519.29 / 0.684920.90 / 0.690215.80 / 0.739116.73 / 0.785216.36 / 0.766823.49 / 0.8513
CSEC [22]1.36420.79 / 0.703120.02 / 0.709320.41 / 0.706222.18 / 0.850222.69 / 0.866222.73 / 0.863823.63 / 0.8550
Exposure-slot1.22923.85 / 0.709221.77 / 0.737522.81 / 0.723923.09 / 0.860123.24 / 0.876223.18 / 0.869724.03 / 0.8592
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.301, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 1. Quantitative results on SICE [5], MSEC [1], and LCDP [35] in terms of PSNR↑/SSIM↑. The best score is displayed in red, the second in blue. Additionally, the number of parameters (#Params) required for inference is also specified. Exposure-slot is lightweight compared to the previous SOTA approach [22] while consistently outperforming it across all datasets on average." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.399 + ], + "angle": 0, + "content": "Meanwhile, \\(\\mathbf{S}^{final}\\) is decoded into a pseudo-corrected image using the slot reconstruction decoder:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.403, + 0.482, + 0.423 + ], + "angle": 0, + "content": "\\[\nI _ {p s e u d o} = \\mathbf {D e c} _ {s l o t} \\left(\\mathbf {S} ^ {f i n a l}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.534 + ], + "angle": 0, + "content": "where \\(\\mathbf{Dec}_{slot}\\) represents the slot reconstruction decoder and \\(I_{pseudo}\\) is the resulting pseudo-corrected RGB image. This training strategy encourages each slot to focus on exposure-centric learning, similar to the set prediction approach used in [24]. It is important to note that the slot reconstruction decoder is used only during training and is not required during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.54, + 0.24, + 0.555 + ], + "angle": 0, + "content": "3.4. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.637 + ], + "angle": 0, + "content": "Exposure-slot is trained using two loss functions: image enhancement loss \\(\\mathcal{L}_{\\text {enhance }}\\) and slot reconstruction loss \\(\\mathcal{L}_{\\text {slot }}\\). The overall training objective is to minimize the combined loss, promoting both accurate image enhancement and effective slot-based reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.715 + ], + "angle": 0, + "content": "Image Enhancement loss. The image enhancement loss \\(\\mathcal{L}_{\\text {enhance }}\\) is defined using the \\(L_{1}\\) distance between the exposure enhanced image \\(I_{\\text {out }}\\) and the ground truth image \\(I_{\\text {gt }}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.721, + 0.483, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n h a n c e}} = \\left\\| I _ {\\text {o u t}} - I _ {\\text {g t}} \\right\\| _ {1}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.483, + 0.828 + ], + "angle": 0, + "content": "Slot Reconstruction loss. The slot reconstruction loss \\(\\mathcal{L}_{slot}\\) is similarly defined using the \\(L_{1}\\) distance, but between the pseudo-corrected image \\(I_{pseudo}\\), generated by the slot reconstruction decoder, and the ground truth image \\(I_{gt}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.83, + 0.482, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s l o t}} = \\left\\| I _ {\\text {p s e u d o}} - I _ {\\text {g t}} \\right\\| _ {1}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.878 + ], + "angle": 0, + "content": "The final objective function \\(\\mathcal{L}_{final}\\) is defined as the sum of the two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.886, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {e n h a n c e}} + \\mathcal {L} _ {\\text {s l o t}}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.367, + 0.646, + 0.384 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.704, + 0.408 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.505 + ], + "angle": 0, + "content": "Implementation Details. We trained our model using the Adam optimizer with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\) using a patch size of \\(256 \\times 256\\) and a batch size of 16. The learning rate was set to \\(2 \\times 10^{-4}\\), and training was conducted for 500 epochs. Additionally, the parameters \\(K^{main}\\) and \\(K^{sub}\\) were set to 3 and 7, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.613 + ], + "angle": 0, + "content": "Datasets and Comparative methods. Our training and benchmarking settings align with established standards for existing exposure correction tasks [13, 14, 22, 35]. We train our network on three multi-exposure datasets: Single Image Contrast Enhancement (SICE) [5], Multi-Scale Exposure Correction (MSEC) [1], and LCDP [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.734 + ], + "angle": 0, + "content": "We compare our Exposure-slot to existing state-of-the-art exposure correction methods, including LCPDNet [35], ERL [15], ENC [12], DA [39], ECLNet [14], FECNet [13], and CSEC [22]. In PromptIR [28], the number of prompts is set to match the number of exposure values in each dataset: 2 for SICE, 5 for MSEC, and 2 for LCDP. Evaluations are conducted using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.738, + 0.758 + ], + "angle": 0, + "content": "4.2. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the performance of our method on three representative multi-exposure datasets: SICE [5], MSEC [1], and LCDP [35]. On the SICE dataset, our approach achieves the highest performance overall, except for one SSIM value in the under-exposed condition, where it ranks second. Similarly, for the MSEC dataset, Exposure-slot consistently outperforms previous methods in PSNR and SSIM, achieving top scores in all cases except for under-exposure in SSIM. On the LCDP dataset," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.227, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.235, + 0.18, + 0.247 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.363, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.235, + 0.323, + 0.246 + ], + "angle": 0, + "content": "FECNet" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.089, + 0.5, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.235, + 0.471, + 0.246 + ], + "angle": 0, + "content": "DRBN-ENC" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.633, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.235, + 0.588, + 0.246 + ], + "angle": 0, + "content": "CSEC" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.77, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.235, + 0.746, + 0.247 + ], + "angle": 0, + "content": "Exposure-slot" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.089, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.235, + 0.882, + 0.246 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.905, + 0.278 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison (FECNet [13], ENC [12], CSEC [22]) on the SICE [5] and MSEC [1] dataset. Examples of images enhanced from under-exposed condition (Top) and images enhanced from over-exposed condition (Bottom)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.283, + 0.905, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.368, + 0.905, + 0.397 + ], + "angle": 0, + "content": "Figure 4. Visualization of slot attention maps (attn) of main- and sub-slot. It demonstrates how exposure-centric characteristics are effectively partitioned and refined through the slot attention mechanisms, transforming the input image into the corrected output." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.407, + 0.475, + 0.552 + ], + "angle": 0, + "content": "
ModelΔEB2000↓ [31]ΔEab↓ [30]
ZeroDCE [9]23.7829.31
RUAS [23]29.5937.05
SCI [26]25.6431.22
ECLNet [14]9.1511.58
FECNet [13]8.8511.19
DRBN-ENC [12]8.6811.02
PromptIR [28]9.2711.59
CSEC [22]8.7211.39
Exposure-slot6.948.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Table 2. Comparisons with color difference metrics, \\( \\Delta {E}_{2000} \\downarrow \\) and \\( \\Delta {E}_{ab} \\downarrow \\) ,on the SICE [5] dataset. The highest score is high- lighted in red, while the second-highest is marked in blue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.702 + ], + "angle": 0, + "content": "our method also demonstrates superior performance, surpassing CSEC [22]. Compared to previous state-of-the-art methods [12, 13], our approach shows a notable gain of 1.85 dB in PSNR on the SICE dataset and 0.4 dB on the LCDP dataset compared to CSEC [22], highlighting its strong performance advantage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Fig.3 shows qualitative comparisons on the SICE [5] and MSEC [1] datasets, demonstrating the performance of our model against other exposure correction methods. Under under-exposed conditions, Exposure-slot achieves superior color fidelity and detail recovery, closely matching the ground truth without excessive brightening or color distortion, unlike other models [13, 22]. In over-exposed areas, the zoomed-in views reveal that other models, particularly CSEC [22], produce artifacts and color inconsistencies due to over-correction. In contrast, Exposure-slot maintains stable color restoration and texture enhancement, preserving natural details and color harmony. These results highlight Exposure-slot's capability to handle diverse exposure con" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.407, + 0.865, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.523, + 0.905, + 0.566 + ], + "angle": 0, + "content": "Figure 5. t-SNE results of features before and after prompts. Features from the same sub-slot are represented by the same color. With prompts, the features exhibit clearer separation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.905, + 0.611 + ], + "angle": 0, + "content": "ditions, achieving state-of-the-art performance in complex exposure correction tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Additionally, to evaluate color correction performance, we conduct a comparison using \\(\\Delta E_{2000}\\) [31] and \\(\\Delta E_{ab}\\) [30] metrics in the LAB color space. Table 2 presents the results, demonstrating that our approach also excels in color correction. Exposure-slot demonstrates a performance improvement of 1.78 in \\(\\Delta E_{2000}\\) and 2.13 in \\(\\Delta E_{ab}\\) compared to previous state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 visualizes the attention maps for main and sub-slots predicted by SSAB. These results demonstrate that each map effectively generates prompts region maps, and Exposure-slot leverages this partitioning information to achieve robust enhancement across different exposure conditions of the same scene. In Fig. 5, the t-SNE visualization of \\( V \\cdot P^{final} \\) in Eq. 2, demonstrates that our prompts improve clustering and effectively facilitate feature separation. Notably, the improved feature identification within each slot suggests that the model effectively captures exposure-specific or region-based information. For further visualizations of attention maps and additional details on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.22, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.146, + 0.174, + 0.156 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.089, + 0.351, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.146, + 0.323, + 0.156 + ], + "angle": 0, + "content": "w/o Prompts" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.089, + 0.481, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.146, + 0.448, + 0.155 + ], + "angle": 0, + "content": "w/o SSAB" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.158, + 0.22, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.216, + 0.19, + 0.227 + ], + "angle": 0, + "content": "w/o Decslot" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.158, + 0.351, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.216, + 0.328, + 0.227 + ], + "angle": 0, + "content": "Exposure-slot" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.158, + 0.482, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.215, + 0.438, + 0.226 + ], + "angle": 0, + "content": "CSEC" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.234, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Figure 6. The visualization of the ablation study. From the top left, the images correspond to (e), (f), (g), and (h) in the Table 3, with the previous SOTA method, CSEC [22], included for comparison." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.399 + ], + "angle": 0, + "content": "
ModelDecslotSSAB (2-level)PromptsPSNR↑/SSIM↑
(a)21.92/0.7091
(b)✓22.02/0.7154
(c)✓21.76/0.6998
(d)✓22.09/0.7062
(e)✓✓22.10/0.7065
(f)✓✓22.09/0.7210
(g)✓✓21.95/0.7149
(h)✓✓✓22.81/0.7239
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.4, + 0.428, + 0.412 + ], + "angle": 0, + "content": "Table 3. Ablation studies on proposed modules." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.421, + 0.422, + 0.435 + ], + "angle": 0, + "content": "t-SNE, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.445, + 0.244, + 0.46 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.512 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to assess the effectiveness of the Exposure-slot, as well as the model configuration and structural levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.638 + ], + "angle": 0, + "content": "Effectiveness of the Proposed Modules. This section evaluates the effectiveness of SSAB, \\(\\mathbf{Dec}_{slot}\\), and prompts through ablation studies. Table 3 shows results for each ablation case on the SICE [5] dataset. Specifically, configurations without SSAB (2-level) refer to an SSAB with a single level consisting of 7 slots (i.e. main-slots only). Configurations without Prompts directly use the slot feature \\(\\mathbf{S}^{final}\\) for cross-attention in Eq. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.865 + ], + "angle": 0, + "content": "First, a comparison between cases with and without \\(\\mathbf{Dec}_{slot}\\) shows consistent performance improvements when \\(\\mathbf{Dec}_{slot}\\) is included. This result suggests that training \\(\\mathbf{Dec}_{slot}\\) helps the slots more accurately identify and represent regions with distinct exposure characteristics, leading to enhanced performance. When employing SSAB (Model (e) and (h)) also shows significant performance improvement when used together with \\(\\mathbf{Dec}_{slot}\\). Regarding prompts, although Model (d) shows a slight decrease in SSIM compared to Model (a), all other cases utilizing prompts demonstrate improved performance. Notably, our full model (h) achieves the highest performance, underscoring the importance of each component in enhancing model performance. In Fig. 6, we provide a part of ablations as visual results within state-of-the-art method CSEC [22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Investigation of Model Configuration. Table 4 presents ablation studies on the number of main-slots (\\( K^{main} \\)), sub-" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.906, + 0.206 + ], + "angle": 0, + "content": "
ModelKmainKsubTPSNR↑SSIM↑Time (S)
(a)37322.810.72360.06759
(b)27322.610.72450.06759
(c)47322.360.72010.06759
(d)36321.920.70380.06759
(e)38322.780.72410.06759
(f)37221.150.70580.06747
(g)37421.870.72220.06996
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.211, + 0.906, + 0.24 + ], + "angle": 0, + "content": "Table 4. Investigation of the number of slots (\\( K^{main} \\), \\( K^{sub} \\)) and number of iterations (\\( T \\)) on the SICE [5] dataset." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.246, + 0.905, + 0.287 + ], + "angle": 0, + "content": "
ModelKmainK*sub-1K*sub-2PSNR↑SSIM↑ΔE2000↓[31]ΔEab↓[30]
1-level3--22.020.71317.129.41
2-level37-22.810.72366.948.89
3-level371023.060.73066.848.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.288, + 0.878, + 0.301 + ], + "angle": 0, + "content": "Table 5. Ablation studies on \\( n \\) -level SSAB \\( \\left( {n = 1,2,3}\\right) \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.907, + 0.452 + ], + "angle": 0, + "content": "slots \\((K^{sub})\\), and iterations \\((T)\\) within SSAB to evaluate their impact on performance. The results include PSNR and SSIM metrics on the SICE [5] dataset, with runtime measurements taken on a single \\(844 \\times 1500\\) RGB image using an NVIDIA RTX 4090 GPU. In this analysis, we vary \\(K^{main}\\), \\(K^{sub}\\), and \\(T\\) by incrementing or decrementing each parameter by 1. Based on the PSNR values, the configuration of \\(K^{main} = 3\\), \\(K^{sub} = 7\\), and \\(T = 3\\) yields the highest performance and is therefore selected for our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.662 + ], + "angle": 0, + "content": "Effectiveness of Structural Levels. In Sec.3, we use a 2-level SSAB structure as the default configuration. To further investigate the potential of SSAB, we perform experiments using both 1-level and 3-level SSAB structures on the SICE [5] dataset, as summarized in Table 5. In the 3-level configuration, we set \\( K^{sub-2} = 10 \\) for the number of second sub-slots, resulting in PSNR and SSIM improvements of 0.25 and 0.07 compared to 2-level SSAB, respectively. These results indicate that adding more SSAB levels can enhance performance. While adding additional levels improves performance, it also leads to an exponential increase in time complexity. Therefore, considering this trade-off, we adopt the 2-level SSAB as the default structure." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we present Exposure-slot, a novel framework for exposure correction that integrates slot-in-slot attention with learnable prompts to achieve precise, exposure-centric feature learning. By hierarchically clustering and refining exposure regions, Exposure-slot facilitates accurate, region-aware adjustments and achieves state-of-the-art performance on multiple multi-exposure benchmarks. Our approach demonstrates significant improvements in both quantitative and qualitative metrics, highlighting the effectiveness of structured attention mechanisms for challenging exposure correction scenarios. Furthermore, this work advances exposure correction and paves the way for exploring slot-based architectures in other low-level vision tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17899" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.238, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.326 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2023-00222776), and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06, Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Ministry of Science and ICT (No.RS-2020-II201373, Artificial Intelligence Graduate School Program [Hanyang University])." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.338, + 0.188, + 0.354 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.363, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[1] Mahmoud Afifi, Konstantinos G Derpanis, Bjorn Ommer, and Michael S Brown. Learning multi-scale photo exposure correction. In CVPR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.406, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.449, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[3] Ondrej Biza, Sjoerd Van Steenkiste, Mehdi S. M. Sajjadi, Gamaleldin F. Elsayed, Aravindh Mahendran, and Thomas Kipf. Invariant slot attention: object discovery with slot-centric reference frames. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[5] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27, 2018. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.648, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[7] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.772 + ], + "angle": 0, + "content": "[8] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[10] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Yun He, Steven Zheng, Yi Tay, Jai Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, YaGuang Li, Zhao Chen, Donald Met" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.363, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "zler, et al. Hyperprompt: Prompt-based task-conditioning of transformers. In ICML. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[12] Jie Huang, Yajing Liu, Xueyang Fu, Man Zhou, Yang Wang, Feng Zhao, and Zhiwei Xiong. Exposure normalization and compensation for multiple-exposure correction. In CVPR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[13] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In ECCV, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[14] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In ACMMM, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.329 + ], + "angle": 0, + "content": "[15] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In CVPR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.906, + 0.372 + ], + "angle": 0, + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[17] Donggoo Jung, Daehyun Kim, and Tae Hyun Kim. Continuous exposure learning for low-light image enhancement using Neural ODEs. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[18] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.906, + 0.51 + ], + "angle": 0, + "content": "[19] Markus Krimmel, Jan Achterhold, and Joerg Stueckler. Attention normalization impacts cardinality generalization in slot attention. Transactions on Machine Learning Research, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.906, + 0.553 + ], + "angle": 0, + "content": "[20] Minhyeok Lee, Suhwan Cho, Dogyoon Lee, Chaewon Park, Jungho Lee, and Sangyoun Lee. Guided slot attention for unsupervised video object segmentation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.555, + 0.906, + 0.608 + ], + "angle": 0, + "content": "[21] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[22] Yiyu Li, Ke Xu, Gerhard Petrus Hancke, and Rynson WH Lau. Color shift estimation-and-correction for image enhancement. In CVPR, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.652, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[23] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[24] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[25] Minh-Thang Luong. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[26] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.861, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[27] Ntumba Elie Nsampi, Zhongyun Hu, and Qing Wang. Learning exposure correction via consistency modeling. In BMVC, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17900" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[28] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. PromptIR: Prompting for all-in-one image restoration. In NeurIPS, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[29] Rishav Pramanik, José-Fabian Villa-Vásquez, and Marco Pedersoli. Masked Multi-Query slot attention for unsupervised object discovery. In IJCNN, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.18, + 0.482, + 0.208 + ], + "angle": 0, + "content": "[30] Gaurav Sharma and Raja Bala. Digital color imaging handbook. CRC press, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.21, + 0.483, + 0.265 + ], + "angle": 0, + "content": "[31] Gaurav Sharma, Wencheng Wu, and Edul N Dalal. The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application, 2005. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.267, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[32] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. CODA-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.338, + 0.482, + 0.379 + ], + "angle": 0, + "content": "[33] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.382, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[34] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.439, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[35] Haoyuan Wang, Ke Xu, and Rynson WH Lau. Local color distributions prior for image enhancement. In ECCV. Springer, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.483, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[36] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.527, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[37] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.571, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[38] Yanbo Wang, Letao Liu, and Justin Dauwels. Slot-vae: Object-centric scene generation with slot attention. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.614, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[39] Yang Wang, Long Peng, Liang Li, Yang Cao, and Zheng-Jun Zha. Decoupling-and-aggregating for image exposure correction. In CVPR, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.657, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[40] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep Retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.701, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[41] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. URetinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[42] Mingde Yao, Jie Huang, Xin Jin, Ruikang Xu, Shenglong Zhou, Man Zhou, and Zhiwei Xiong. Generalized lightness adaptation with channel selective normalization. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[43] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In ACMM, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[46] Karel Zuiderveld. Contrast limited adaptive histogram equalization. Graphics gems, 1994. 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17901" + } + ] +] \ No newline at end of file diff --git a/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_origin.pdf b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1948f23244ba6d4fbdb828e493486cbcea85cdda --- /dev/null +++ b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/83f68ecd-43a7-49b5-b8e2-1ddf3578f781_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:448db1064e9015b174f3322c0efbe7dffca731a7a24dee3e8fc927c50654b278 +size 7582371 diff --git a/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/full.md b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..85dcf7467c10d87b3d9a113a46158c78d7c03b4a --- /dev/null +++ b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/full.md @@ -0,0 +1,391 @@ +# Exposure-slot: Exposure-centric representations learning with Slot-in-Slot Attention for Region-aware Exposure Correction + +Donggoo Jung\*, Daehyun Kim\*, Guanghui Wang, Tae Hyun Kim† Hanyang University, Toronto Metropolitan University + +{dgjung, daehyun, taehyunkim}@hanyang.ac.kr, {wangcs}@torontomu.ca + +# Abstract + +Image exposure correction enhances images captured under diverse real-world conditions by addressing issues of under- and over-exposure, which can result in the loss of critical details and hinder content recognition. While significant advancements have been made, current methods often fail to achieve optimal feature learning for effective correction. To overcome these challenges, we propose Exposure-slot, a novel framework that integrates a prompt-based slot-in-slot attention mechanism to cluster exposed feature regions and learn exposure-centric features for each cluster. By extending the Slot Attention algorithm with a hierarchical structure, our approach progressively clusters features, enabling precise and region-aware correction. In particular, learnable prompts tailored to exposure characteristics of slots further enhance feature quality, adapting dynamically to varying conditions. Our method delivers superior performance on benchmark datasets, surpassing the current state-of-the-art with a PSNR improvement of over 1.85 dB on the SICE dataset and 0.4 dB on the LCDP dataset, thereby establishing a new benchmark for multi-exposure correction. The source code can be found at: https://github.com/kdhRick2222/Exposure-slot. + +# 1. Introduction + +Image exposure correction enhances images captured under diverse real-world conditions. Under- or over-exposed images can appear excessively dark or bright, losing essential details and compromising accurate recognition. Despite advancements in imaging technology, robust automatic exposure correction remains a significant challenge. + +Due to its critical importance, image exposure correction has been the focus of considerable research. Early efforts primarily treated under-exposure and over-exposure as separate issues [10, 37, 40]. More recent advancements, however, have introduced multi-exposure correction mod + +![](images/b1e3638b18597a88089e950bfec1c3792e50485f2d9030b3ff6aab4306168e0d.jpg) +Figure 1. (Top) The Slot-in-Slot Attention mechanism hierarchically partitions exposure-aware regions using the attention maps of main and sub-slot attention, performing corrections with learnable prompts based on exposure levels. (Bottom) Comparison of the Exposure-slot with existing approaches. From left to right: ECLNet [14], FECNet [13], DRBN-ENC [12], CSEC [22] and our proposed approach, Exposure-slot. + +els capable of handling a wide range of exposure levels within a unified framework. For instance, MSEC [1] introduced an approach that considers both color and detail enhancement while performing multi-exposure correction in an end-to-end manner, and provided a dataset of images with different exposure errors for both training and evaluation. In addition, LCDPNet [35] provided another dataset of diverse scenes, covering both over-exposure and under-exposure, and incorporates retinax theory to account for local color distributions within images. Recent approaches increasingly focus on the separate processing of features based on their physical characteristics, such as exposure, frequency, contrast, color, and details. FECNet [13] uses Fourier transform to facilitate interactions between local spatial feature information and global frequency information, while ECLNet [14] uses a bilateral activation mechanism to process over- and under-exposed regions independently. In addition, DA [39] introduces a decoupling and + +aggregation scheme that separately enhances image details and contrast. + +However, relying on separating image features based on these physical properties may not ensure the optimal features for effective exposure correction. For example, at the bottom of Fig. 1, we visualize this limitation, showing that previous works [13-15, 22] often generate inappropriate artifacts in sky regions and struggle to accurately adjust sunset exposure levels. + +To overcome these challenges, we propose a novel prompt-based slot-in-slot attention mechanism that clusters regions with similar exposure levels, enabling exposure-centric feature learning within each cluster. We utilize Slot Attention [24], designed to cluster features based on object representations, to identify distinct exposure regions within an image. In our approach, slots represent features grouped by exposure levels, and the corresponding attention maps act as exposure-aware region maps, guiding the grouping of features according to exposure characteristics. Therefore, these maps highlight regions based on exposure levels, enabling targeted exposure correction for each area. + +To support effective correction, we introduce learnable prompt vectors that operate based on exposure-aware region maps, which are attention maps of each slot. These prompts adapt to the characteristics of each region defined by the attention map, helping the model learn targeted correction to improve under- or over-exposed areas. Specifically, we introduce Slot-Prompt Interaction Module (SPIM), which combines slot and prompt information using cross-attention. The cross-attention component of SPIM uses the prompts learned through Slot-in-Slot Attention as conditioning factors, enhancing the interaction between features and prompts. + +Our approach builds upon standard slot attention with a hierarchical design, called Slot-in-Slot Attention, which processes slots in levels. This mechanism operates across multiple structure levels: the first level coarsely partitions regions based on distinct exposure characteristics, while subsequent levels progressively refine these regions with more partitions. Each slot captures exposure-specific regional information, distinguishing areas with varying exposure levels to enhance feature separation and improve the precision of local details. + +The overall framework of the proposed approach is illustrated at the top of Fig. 1. The Slot-in-Slot Attention is applied to intermediate features between the encoder and decoder, generating attention maps for slots that are hierarchically partitioned according to exposure characteristics, as depicted in Fig. 1. The generated attention maps are then used as weight maps for learnable prompts, resulting in refined intermediate features that are subsequently fed into the decoder. We refer to our model as Exposure-Slot, as it is the first model to use Slot Attention for unsupervised feature + +partitioning based on exposure levels in exposure correction, and the first to apply region-aware prompts for feature enhancement. Our main contributions are as follows: + +- Exposure-slot is the first approach to leverage Slot Attention mechanism for optimized exposure-specific feature partitioning. +- We introduce the slot-in-slot attention that enables sophisticated feature partitioning and learning. +- We apply exposure-aware prompts that enhance the exposure-centric characteristics of each image feature. +- Exposure-slot achieves state-of-the-art results on multi-exposure benchmark datasets [1, 5, 35], setting a new standard in this field. + +# 2. Related Work + +# 2.1. Exposure Correction + +With the rise of deep neural networks (DNN), exposure correction has seen the emergence of a range of DNN-based methods leveraging diverse concepts. For under-exposed image enhancement, methods [36, 40, 41, 43, 44] inspired by retinax theory have been proposed. Retinex-based models like CMEC [27] have also incorporated attention mechanisms to address multi-exposure correction, while LCDP-Net [35] introduced a novel focus on local color distribution. To support multi-exposure correction tasks, dedicated datasets such as MSEC [1] and SICE [5] were developed for training and evaluation. MSEC further introduced a Laplacian pyramid architecture to handle varying exposure levels. ENC [12] contributed an exposure normalization module that refined feature maps by transforming diverse exposure features into an exposure-invariant feature space. CSNorm [42] also enhanced model generalization by selectively normalizing lightness-sensitive channels and ERL [15] proposed exposure relationship learning through regularization techniques for multi-exposure correction. + +There are several recent studies that emphasize the feature separation based on their physical characteristics. ECLNet [14] used a bilateral activation mechanism to adjust processing across different exposure conditions, while FECNet [13] proposed a lightweight model that utilizes spatial-frequency interactions with a Fourier-based approach. CLODE [17] also utilizes image curve adjustment and formulates brightness changes through continuous curve map extraction using Neural ODEs [6]. The latest advancement, CSEC [22], addresses color distribution shifts by defining darkened and brightened feature maps. + +# 2.2. Slot Attention + +Slot Attention [24] is a mechanism for clustering features corresponding to slots which is the main object of object-centric learning (OCL). Specifically, this mechanism iteratively employs dot-product attention [25] with attention co + +![](images/c370d44eacae54c2cff85279b6d38a663e2fbe90d27a4f7d16e5dff89a2c3918.jpg) +Figure 2. Overview of Exposure-Slot: Exposure-Slot operates within a $U$ -shaped encoder-decoder network, with the Slot-Prompt Interaction Module (SPIM) bridging the encoder and decoder stages. In SPIM, Slot-in-Slot Attention Block (SSAB) adaptively partitions regions and generates prompt features to enhance the feature representation. In addition, a Slot Decoder $(\mathbf{Dec}_{slot})$ is employed during training to reconstruct slot features to the target. This ensures that slot attention maps capture exposure-centric information, helping prompts learn the relevant information for enhancement within each corresponding slot. + +![](images/7493085b6a97d7e99cfe6816dda50509285af7115b32dc2ead4cd8362e9012a7.jpg) + +efficients to update slots, which function as queries within the attention process. Through multiple rounds of recurrent attention, Slot Attention achieves effective clustering and feature separation in OCL without the need for annotations. + +Slot Attention has been effectively applied to image classification, as seen in SCOUTER [21, 33], which uses a Slot-Attention-based classifier for explainable recognition. In segmentation, Slot Attention has proven particularly valuable for video segmentation [3, 8, 19, 29]. For example, GSANet [20] utilizes Slot Attention to separate center objects from background elements across video frames. + +In reconstruction and restoration tasks, Slot-VAE [38] integrates Slot Attention with a hierarchical Variational Autoencoder (VAE) framework, enhancing object-centric scene generation. Moreover, AID [18] employs Slot Attention to capture implicit representations of illuminant chromaticities, with each slot vector capturing the characteristics of a specific illuminant. This enables AID to generate accurate chromaticity and weight maps for each light source. In this paper, we leverage Slot Attention for partitioning in unsupervised way for the exposure correction. This allows our method to train each slot to capture feature-relevant exposure representations and effectively to cluster features implicitly for exposure correction without requiring annotations. + +# 2.3. Prompt-Based Learning + +The prompt-based learning method was first introduced by [4], proposing the concept of optimizing input prompts. This approach gained popularity in the field of natural lan + +guage processing, inspiring numerous follow-up studies. CoOp [45] introduced a novel approach by treating prompt contexts as learnable parameters, moving beyond methods limited to fixed-format prompts, and demonstrating that learnable prompts can outperform handcrafted ones. On the basis of CoOp, several methods have been developed for dynamically generating suitable prompts [7, 11, 32]. For example, CODA [32] generates prompts tailored to specific inputs, while HyperPrompt [11] addresses multi-task learning through the use of a prompt generator. In the computer vision field, visual prompts involve adding trainable parameters to modify inputs for efficient model adaptation. VPT [16] demonstrated substantial performance improvements over conventional fine-tuning methods by applying visual prompts to a fixed transformer backbone. Furthermore, [2] presented visual prompts compatible with input images, specifically optimized for CLIP. For low-level vision tasks, methods like PromptIR [28] and PromptReformer [34] leverage prompts to encode degradation-specific information, guiding restoration networks to adapt to various types and intensities of degradation. In our method, we introduce exposure-specific prompts that guide feature separation and generate region-aware features specifically tailored to distinct exposure characteristics. + +# 3. Proposed Method + +# 3.1. Overall Flow + +As illustrated in Fig. 2, the proposed Exposure-slot architecture is a $U$ -shaped residual network composed of an encoder-decoder structure connected via skip connections + +Algorithm 1 Slot-in-Slot Attention Block (SSAB). For clarity, we represent the algorithm using a 2-level hierarchical structure, consisting of main- and sub-slots. +Input: K, V, slots $^{main}$ , slots $^{sub}$ , prompts $^{main}$ , prompts $^{sub}$ +1: function SLOT_UPDATE(slots, k, v) +2: slots $_{prev}$ = slots +3: slots = LayerNorm(slots) +4: q = to_q(slots) +5: attn = Softmax $\left(\frac{1}{\sqrt{D}} k \cdot q^T\right)$ +6: updates = WeightedMean(weights=attn + ε, values = v) +7: slots = GRU(state=slot $_{prev}$ , inputs=updates) +8: return slots, attn +9: end function +10: +11: k = K, v = V +12: for $t_{main} = 0 \ldots T$ do +13: slots $^{main}$ , attn $^{main}$ = SLOT_UPDATE(slots $^{main}$ , k, v) +14: P $^{main}$ = prompts $^{main}$ × attn $^{main}$ +15: k, v = k $\cdot$ P $^{main}$ , v $\cdot$ P $^{main}$ +16: for $t_{sub} = 0 \ldots T$ do +17: slots $^{sub}$ , attn $^{sub}$ = SLOT_UPDATE(slots $^{sub}$ , k, v) +18: end for +19: P $^{sub}$ = prompts $^{sub}$ × attn $^{sub}$ +20: k, v = k $\cdot$ P $^{sub}$ , v $\cdot$ P $^{sub}$ +21: end for +22: +23: P $^{final}$ = P $^{main}$ $\cdot$ P $^{sub}$ +24: S $^{final}$ = (slots $^{main}$ × attn $^{main}$ ) $\cdot$ (slots $^{sub}$ × attn $^{sub}$ ) +25: return P $^{final}$ , S $^{final}$ + +and Slot-Prompt Interaction Module (SPIM). + +First, the encoder processes the poorly exposed input image $I_{in}$ to produce a latent feature representation $F \in \mathbb{R}^{H \times W \times C}$ as $F = \mathbf{Enc}(I_{in})$ . Then, SPIM, positioned between the encoder and decoder, refines the encoder output $F$ using input-specific, region-aware prompts guided by the Slot-in-Slot Attention mechanism. The refined latent feature $F'$ is subsequently passed to the image decoder (Dec_enhance), which processes it to produce the output image $I_{out}$ with well-enhanced exposure. + +# 3.2. Slot-Prompt Interaction Module (SPIM) + +SPIM is designed as a combination of the Slot-in-Slot Attention Block (SSAB) and a subsequent cross-attention process. SPIM initially projects $F$ into query, key, and value representations, while SSAB receives the key and value to estimate the corresponding prompt features of $F$ . Specifically, SSAB hierarchically applies slot attention to achieve coarse-to-fine region separation, which supports effective exposure correction. Then the derived prompt feature serves as conditioning elements in the cross-attention step involving the query, key, and value representations. + +Specifically, as illustrated in Fig. 2, the encoded feature $F \in \mathbb{R}^{H \times W \times C}$ is first flattened into $X \in \mathbb{R}^{HW \times C}$ . Subsequently, $X$ is projected into query $\mathbf{Q} \in \mathbb{R}^{HW \times C}$ , key + +$\mathbf{K} \in \mathbb{R}^{HW \times C}$ , and value $\mathbf{V} \in \mathbb{R}^{HW \times C}$ through the linear projections $W_{\mathbf{Q}}, W_{\mathbf{K}}$ , and $W_{\mathbf{V}}$ as follows: + +$$ +\mathbf {Q} = X W _ {\mathbf {Q}} ^ {T}, \quad \mathbf {K} = X W _ {\mathbf {K}} ^ {T}, \quad \mathbf {V} = X W _ {\mathbf {V}} ^ {T}. \qquad (1) +$$ + +Within our SSAB, $\mathbf{K}$ and $\mathbf{V}$ are used to generate the prompt feature $\mathbf{P}^{final}$ , which serves as a condition for the subsequent cross-attention process to produce the refined feature map $F' \in \mathbb{R}^{H \times W \times C}$ as follows: + +$$ +F ^ {\prime} = r e s h a p e (s o f t m a x (\frac {\mathbf {Q K} ^ {T}}{\sqrt {d}}) \cdot (\mathbf {V} \cdot \mathbf {P} ^ {f i n a l})), \quad (2) +$$ + +where $d$ denotes a learnable parameter that adaptively scales matrix multiplication. + +# 3.2.1. Slot-in-Slot Attention Block (SSAB) + +In the proposed Slot-in-Slot Attention Block within SPIM, we generate a refined feature map by integrating a hierarchically structured slot mechanism, which partitions regions with different exposure levels in a hierarchical and iterative manner, with learned prompts specifically tailored to distinct exposure characteristics. + +For instance, in a 2-level slot-in-slot structure, the first level partitions regions with similar exposure characteristics using a relatively smaller number of slots. Building on + +these results, the second level employs more slots to cluster the regions more finely. By iteratively repeating this process, the sub-slots gradually learn finer details of differently exposed regions. Notably, the hierarchical structure can extend beyond two levels, enabling the identification of exposure-centric feature maps with increasing precision. Next, we introduce exposure-specific prompts that enhance the image by targeting distinct regional exposure characteristics, multiplied by attention maps from the slot attention mechanism. Since the attention maps identify regions with different exposure levels, multiplying them with exposure-specific prompts generates region-aware guidance, enabling balanced exposure correction. + +In Alg. 1, the pseudo-code for this 2-level slot-in-slot structure is provided, refines key $(k)$ and value $(v)$ through two nested loops: an outer loop and an inner loop. The outer loop updates the main-slots, which are then combined with prompts to refine $k$ and $v$ . In the inner loop, the refined $k$ and $v$ are used to update the sub-slots, which are further combined with prompts to refine $k$ and $v$ with fine details. This process is repeated until the outer loop is completed. Notably, to update the slots, we employ the SLOT_UPDATE function introduced process in [24]. The key difference is that each loop iteration uses the updated $k$ and $v$ as inputs. + +Main-slot Attention In the outer loop of Alg. 1, we perform Main-slot Attention to update the main-slots (slots $^{main}$ ), which represent the first level of the hierarchical slot structure. Notably, slots $^{main} \in \mathbb{R}^{K^{main} \times D_{slot}^{main}}$ , where $K^{main}$ , and $D_{slot}^{main}$ indicate the number of slots and the dimensionality of each slot, respectively, can be initialized with learnable parameters. + +First, given $\mathbf{K}$ and $\mathbf{V}$ from the Encoder, assigned to $k$ and $v$ , respectively, we update the slots using the SLOT_UPDATE function as follows: + +$$ +\mathbf {s l o t s} ^ {\text {m a i n}}, \mathbf {a t t n} ^ {\text {m a i n}} = \mathbf {S L O T} _ {-} \mathbf {U P D A T E} (\mathbf {s l o t s} ^ {\text {m a i n}}, k, v), \tag {3} +$$ + +where $\mathbf{attn}^{main}$ represents the attended region of each slots $^{main}$ . Then, the attention map $\mathbf{attn}^{main}$ is used to produce prompt feature $\mathbf{P}^{main}$ as: + +$$ +\mathbf {P} ^ {\text {m a i n}} = \text {p r o m p t s} ^ {\text {m a i n}} \times \text {a t t n} ^ {\text {m a i n}}, \tag {4} +$$ + +where $\mathbf{p}^{\text{main}} \in \mathbb{R}^{K^{\text{main}} \times D_{\text{slot}}^{\text{main}}}$ denotes learnable prompts tailored to each exposure characteristics. Thus, by multiplying the exposure-specific prompts with the attention map containing distinct exposure region information, our prompt feature $\mathbf{P}^{\text{main}}$ effectively encapsulates exposure-centric and region-aware information. Finally, prompt feature $\mathbf{P}^{\text{main}}$ is used to update the key and value as: + +$$ +k, v = k \cdot \mathbf {P} ^ {\text {m a i n}}, v \cdot \mathbf {P} ^ {\text {m a i n}}, \tag {5} +$$ + +and the resulting $k$ and $v$ are used as inputs for the subsequent inner loop process, referred to as Sub-slot Attention. + +Sub-slot Attention Within the inner loop of Alg. 1, we perform Sub-slot Attention to update the sub-slots (slot $s u b$ ), which represent the second level of hierarchical structure. Particularly, slots $s u b \in \mathbb{R}^{K^{s u b} \times D_{s l o t}^{s u b}}$ , where slots $s u b$ are also initialized using learnable parameters. Here, $K^{s u b}$ denotes the number of sub-slots, $D_{s l o t}^{s u b}$ denotes their dimensionality, and $K^{s u b}$ is configured to ensure finer feature separation. + +First, the updated $k$ and $v$ from the Main-slot Attention in Eq. 5 are used to update slots $^{sub}$ via the SLOT_UPDATE function. Similarly to Eq. 3, Sub-slot Attention is updated as follows: + +$$ +\text {s l o t s} ^ {\text {s u b}}, \text {a t t n} ^ {\text {s u b}} = \text {S L O T} _ {-} \text {U P D A T E} (\text {s l o t s} ^ {\text {s u b}}, k, v), \tag {6} +$$ + +where $\mathbf{attn}^{sub}$ represents the attended region of each slots $^{sub}$ , and this update process is repeated for $T$ iterations. Subsequently, $\mathbf{attn}^{sub}$ is used to obtain sub-prompt feature $\mathbf{P}^{sub}$ from the Sub-slot Attention as: + +$$ +\mathbf {P} ^ {s u b} = \text {p r o m p t s} ^ {s u b} \times \text {a t t n} ^ {s u b}, \tag {7} +$$ + +where $\mathbf{p}^{\mathrm{sub}} \in \mathbb{R}^{K^{\mathrm{sub}} \times D_{\mathrm{slot}}^{\mathrm{sub}}}$ is also learnable parameters tailored to each exposure characteristic in Sub-slot Attention. Finally, by multiplying $\mathbf{P}^{\mathrm{sub}}$ with $k$ and $v$ , we transfer information from the Sub-slot Attention for the next iteration of the Main-slot Attention as: + +$$ +k, v = k \cdot \mathbf {P} ^ {s u b}, v \cdot \mathbf {P} ^ {s u b}. \tag {8} +$$ + +To summarize, the result of the Main-slot Attention is integrated into the Sub-slot Attention, and the output from the Sub-slot Attention is iteratively fed back into the Main-slot Attention loop. This complementary updating process is repeated $T$ times. At the end of SSAB, the final outputs $\mathbf{P}^{final}$ and $\mathbf{S}^{final}$ are obtained as follows: + +$$ +\begin{array}{l} \mathbf {P} ^ {f i n a l} = \mathbf {P} ^ {m a i n} \cdot \mathbf {P} ^ {s u b}, \\ \mathbf {S} ^ {\text {f i n a l}} = \left(\mathbf {s l o t s} ^ {\text {m a i n}} \times \mathbf {a t t n} ^ {\text {m a i n}}\right) \cdot \left(\mathbf {s l o t s} ^ {\text {s u b}} \times \mathbf {a t t n} ^ {\text {s u b}}\right). \tag {9} \\ \end{array} +$$ + +Note that $\mathbf{P}^{final}$ is integrated with $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ as described in Eq. 2, and is used to decode the final output. Meanwhile $\mathbf{S}^{final}$ serves as an input to a separate decoder to facilitate slot training. + +# 3.3.Decoder Process + +The refined feature $F'$ and slots $S^{final}$ from SPIM are used as inputs for the final decoding procedures. Specifically, our Exposure-slot uses the enhanced feature $F'$ to predict an enhanced image through the decoder, as follows: + +$$ +I _ {o u t} = \operatorname {D e c} _ {\text {e n h a n c e}} \left(F ^ {\prime}\right), \tag {10} +$$ + +where $\mathbf{Dec}_{\text{enhance}}$ is image decoder and $I_{out}$ is the final exposure-enhanced output. + +
Model#Params (M)SICE [5]MSEC [1]LCDP [35] Avg.
UnderOverAvg.UnderOverAvg.
CLAHE [46]-12.69 / 0.503710.21 / 0.487811.45 / 0.494216.77 / 0.621114.45 / 0.584215.38 / 0.599016.33 / 0.6420
RetinexNet [40]0.84012.94 / 0.517112.87 / 0.525212.90 / 0.521212.13 / 0.620910.47 / 0.595311.14 / 0.604819.25 / 0.7041
ZeroDCE [9]0.07916.92 / 0.63307.11 / 0.429212.02 / 0.531114.55 / 0.588710.40 / 0.514212.06 / 0.544112.59 / 0.6530
RUAS [23]0.00216.63 / 0.55894.54 / 0.319610.59 / 0.439313.43 / 0.68076.39 / 0.46559.20 / 0.551513.76 / 0.6060
SCI [26]0.00117.86 / 0.64014.45 / 0.362912.49 / 0.50519.97 / 0.66815.84 / 0.51907.49 / 0.578611.87 / 0.5234
MSEC [1]7.04019.62 / 0.651217.59 / 0.656018.58 / 0.653620.52 / 0.812919.79 / 0.815620.08 / 0.821020.38 / 0.7800
LCDPNet [35]0.96017.45 / 0.562217.04 / 0.646317.25 / 0.604322.35 / 0.865022.17 / 0.847622.30 / 0.855223.24 / 0.8420
ECLNet [14]0.01822.05 / 0.689319.25 / 0.687220.65 / 0.686122.37 / 0.856622.70 / 0.867322.57 / 0.863122.44 / 0.8061
FECNet [13]0.15022.01 / 0.673719.91 / 0.696120.96 / 0.684922.96 / 0.859823.22 / 0.874823.12 / 0.868822.41 / 0.8402
DRBN-ENC [12]0.58021.89 / 0.707119.09 / 0.722920.49 / 0.715022.72 / 0.859022.11 / 0.852122.35 / 0.853022.09 / 0.8271
MSEC+DA [39]7.04020.94 / 0.754617.49 / 0.664019.22 / 0.709321.53 / 0.859021.55 / 0.875021.54 / 0.867021.05 / 0.8119
ECLNet+ERL [15]0.01822.14 / 0.690819.47 / 0.698220.81 / 0.694522.90 / 0.862422.58 / 0.867622.70 / 0.865522.63 / 0.8096
PromptIR [28]34.16422.51 / 0.695519.29 / 0.684920.90 / 0.690215.80 / 0.739116.73 / 0.785216.36 / 0.766823.49 / 0.8513
CSEC [22]1.36420.79 / 0.703120.02 / 0.709320.41 / 0.706222.18 / 0.850222.69 / 0.866222.73 / 0.863823.63 / 0.8550
Exposure-slot1.22923.85 / 0.709221.77 / 0.737522.81 / 0.723923.09 / 0.860123.24 / 0.876223.18 / 0.869724.03 / 0.8592
+ +Table 1. Quantitative results on SICE [5], MSEC [1], and LCDP [35] in terms of PSNR↑/SSIM↑. The best score is displayed in red, the second in blue. Additionally, the number of parameters (#Params) required for inference is also specified. Exposure-slot is lightweight compared to the previous SOTA approach [22] while consistently outperforming it across all datasets on average. + +Meanwhile, $\mathbf{S}^{final}$ is decoded into a pseudo-corrected image using the slot reconstruction decoder: + +$$ +I _ {p s e u d o} = \mathbf {D e c} _ {s l o t} \left(\mathbf {S} ^ {f i n a l}\right), \tag {11} +$$ + +where $\mathbf{Dec}_{slot}$ represents the slot reconstruction decoder and $I_{pseudo}$ is the resulting pseudo-corrected RGB image. This training strategy encourages each slot to focus on exposure-centric learning, similar to the set prediction approach used in [24]. It is important to note that the slot reconstruction decoder is used only during training and is not required during inference. + +# 3.4. Loss functions + +Exposure-slot is trained using two loss functions: image enhancement loss $\mathcal{L}_{\text {enhance }}$ and slot reconstruction loss $\mathcal{L}_{\text {slot }}$ . The overall training objective is to minimize the combined loss, promoting both accurate image enhancement and effective slot-based reconstruction. + +Image Enhancement loss. The image enhancement loss $\mathcal{L}_{\text {enhance }}$ is defined using the $L_{1}$ distance between the exposure enhanced image $I_{\text {out }}$ and the ground truth image $I_{\text {gt }}$ as follows: + +$$ +\mathcal {L} _ {\text {e n h a n c e}} = \left\| I _ {\text {o u t}} - I _ {\text {g t}} \right\| _ {1}. \tag {12} +$$ + +Slot Reconstruction loss. The slot reconstruction loss $\mathcal{L}_{slot}$ is similarly defined using the $L_{1}$ distance, but between the pseudo-corrected image $I_{pseudo}$ , generated by the slot reconstruction decoder, and the ground truth image $I_{gt}$ , as follows: + +$$ +\mathcal {L} _ {\text {s l o t}} = \left\| I _ {\text {p s e u d o}} - I _ {\text {g t}} \right\| _ {1}. \tag {13} +$$ + +The final objective function $\mathcal{L}_{final}$ is defined as the sum of the two losses: + +$$ +\mathcal {L} _ {\text {f i n a l}} = \mathcal {L} _ {\text {e n h a n c e}} + \mathcal {L} _ {\text {s l o t}}. \tag {14} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation Details. We trained our model using the Adam optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ using a patch size of $256 \times 256$ and a batch size of 16. The learning rate was set to $2 \times 10^{-4}$ , and training was conducted for 500 epochs. Additionally, the parameters $K^{main}$ and $K^{sub}$ were set to 3 and 7, respectively. + +Datasets and Comparative methods. Our training and benchmarking settings align with established standards for existing exposure correction tasks [13, 14, 22, 35]. We train our network on three multi-exposure datasets: Single Image Contrast Enhancement (SICE) [5], Multi-Scale Exposure Correction (MSEC) [1], and LCDP [35]. + +We compare our Exposure-slot to existing state-of-the-art exposure correction methods, including LCPDNet [35], ERL [15], ENC [12], DA [39], ECLNet [14], FECNet [13], and CSEC [22]. In PromptIR [28], the number of prompts is set to match the number of exposure values in each dataset: 2 for SICE, 5 for MSEC, and 2 for LCDP. Evaluations are conducted using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics. + +# 4.2. Performance Evaluation + +Table 1 presents the performance of our method on three representative multi-exposure datasets: SICE [5], MSEC [1], and LCDP [35]. On the SICE dataset, our approach achieves the highest performance overall, except for one SSIM value in the under-exposed condition, where it ranks second. Similarly, for the MSEC dataset, Exposure-slot consistently outperforms previous methods in PSNR and SSIM, achieving top scores in all cases except for under-exposure in SSIM. On the LCDP dataset, + +![](images/ba670b43b7537856e6885fb394c266099a0fc24e94cb56f7d28f24868741e703.jpg) +Input + +![](images/3dad1efa62d23312b168e065c5935e798aa6239a9521753cf4875b3963cbf380.jpg) +FECNet + +![](images/6d4ec2cb75a0e1e3f0b703d99ebb245de48981e003cec059ebb17fa22c4f6fab.jpg) +DRBN-ENC + +![](images/245d1b303be3e1b67351a352f979c2c7a333dca2702abaa695e28d70906ef4bd.jpg) +CSEC + +![](images/ad4773be4e0498c0114c5c21680f50d3d7743114f415a62c1cc879781abe72f0.jpg) +Exposure-slot + +![](images/8e7b96fef407037412207c9d6a19689e03a7e4f4c4a17af58d7d676290527755.jpg) +Ground Truth + +![](images/3766b1f22c0bcf28b2ed52149cf91edf6595048fe24c100e9dea9fd8cabb6105.jpg) +Figure 3. Qualitative comparison (FECNet [13], ENC [12], CSEC [22]) on the SICE [5] and MSEC [1] dataset. Examples of images enhanced from under-exposed condition (Top) and images enhanced from over-exposed condition (Bottom). +Figure 4. Visualization of slot attention maps (attn) of main- and sub-slot. It demonstrates how exposure-centric characteristics are effectively partitioned and refined through the slot attention mechanisms, transforming the input image into the corrected output. + +
ModelΔEB2000↓ [31]ΔEab↓ [30]
ZeroDCE [9]23.7829.31
RUAS [23]29.5937.05
SCI [26]25.6431.22
ECLNet [14]9.1511.58
FECNet [13]8.8511.19
DRBN-ENC [12]8.6811.02
PromptIR [28]9.2711.59
CSEC [22]8.7211.39
Exposure-slot6.948.89
+ +Table 2. Comparisons with color difference metrics, $\Delta {E}_{2000} \downarrow$ and $\Delta {E}_{ab} \downarrow$ ,on the SICE [5] dataset. The highest score is high- lighted in red, while the second-highest is marked in blue. + +our method also demonstrates superior performance, surpassing CSEC [22]. Compared to previous state-of-the-art methods [12, 13], our approach shows a notable gain of 1.85 dB in PSNR on the SICE dataset and 0.4 dB on the LCDP dataset compared to CSEC [22], highlighting its strong performance advantage. + +Fig.3 shows qualitative comparisons on the SICE [5] and MSEC [1] datasets, demonstrating the performance of our model against other exposure correction methods. Under under-exposed conditions, Exposure-slot achieves superior color fidelity and detail recovery, closely matching the ground truth without excessive brightening or color distortion, unlike other models [13, 22]. In over-exposed areas, the zoomed-in views reveal that other models, particularly CSEC [22], produce artifacts and color inconsistencies due to over-correction. In contrast, Exposure-slot maintains stable color restoration and texture enhancement, preserving natural details and color harmony. These results highlight Exposure-slot's capability to handle diverse exposure con + +![](images/835e3d8d76172bf4dbc130c189f03c7ad86b41e7ea8bd1efa402f09ed6d671c7.jpg) +Figure 5. t-SNE results of features before and after prompts. Features from the same sub-slot are represented by the same color. With prompts, the features exhibit clearer separation. + +ditions, achieving state-of-the-art performance in complex exposure correction tasks. + +Additionally, to evaluate color correction performance, we conduct a comparison using $\Delta E_{2000}$ [31] and $\Delta E_{ab}$ [30] metrics in the LAB color space. Table 2 presents the results, demonstrating that our approach also excels in color correction. Exposure-slot demonstrates a performance improvement of 1.78 in $\Delta E_{2000}$ and 2.13 in $\Delta E_{ab}$ compared to previous state-of-the-art methods. + +Fig. 4 visualizes the attention maps for main and sub-slots predicted by SSAB. These results demonstrate that each map effectively generates prompts region maps, and Exposure-slot leverages this partitioning information to achieve robust enhancement across different exposure conditions of the same scene. In Fig. 5, the t-SNE visualization of $V \cdot P^{final}$ in Eq. 2, demonstrates that our prompts improve clustering and effectively facilitate feature separation. Notably, the improved feature identification within each slot suggests that the model effectively captures exposure-specific or region-based information. For further visualizations of attention maps and additional details on + +![](images/27f5f3de9fb3ab137342016cc8714df99f761f5a49f8390e570b5f29c933d8d5.jpg) +Input + +![](images/33bbd8ada2de18220583b532fcfec2f05c70ebe8fa5d946b21c6c14be10d7784.jpg) +w/o Prompts + +![](images/b1f7df9028430fda4f3c9e2158169fffef59860a2618f6a5c830c87921f58dd5.jpg) +w/o SSAB + +![](images/13bd8490eba7320186ea4b188ed2a37c62f3cb5ba7e3550c19dbc8659b8fb905.jpg) +w/o Decslot + +![](images/74c9e209e38e658f31f26d51ba5eaff63c6b094a88b305e43623f51b6fa4642d.jpg) +Exposure-slot + +![](images/84153e20cf1e9f04031ee208cf16656b19f72a8e627c6b441fe702cc1a94887c.jpg) +CSEC +Figure 6. The visualization of the ablation study. From the top left, the images correspond to (e), (f), (g), and (h) in the Table 3, with the previous SOTA method, CSEC [22], included for comparison. + +
ModelDecslotSSAB (2-level)PromptsPSNR↑/SSIM↑
(a)21.92/0.7091
(b)✓22.02/0.7154
(c)✓21.76/0.6998
(d)✓22.09/0.7062
(e)✓✓22.10/0.7065
(f)✓✓22.09/0.7210
(g)✓✓21.95/0.7149
(h)✓✓✓22.81/0.7239
+ +t-SNE, please refer to the supplementary material. + +# 4.3. Ablation Study + +In this section, we conduct ablation studies to assess the effectiveness of the Exposure-slot, as well as the model configuration and structural levels. + +Effectiveness of the Proposed Modules. This section evaluates the effectiveness of SSAB, $\mathbf{Dec}_{slot}$ , and prompts through ablation studies. Table 3 shows results for each ablation case on the SICE [5] dataset. Specifically, configurations without SSAB (2-level) refer to an SSAB with a single level consisting of 7 slots (i.e. main-slots only). Configurations without Prompts directly use the slot feature $\mathbf{S}^{final}$ for cross-attention in Eq. 2. + +First, a comparison between cases with and without $\mathbf{Dec}_{slot}$ shows consistent performance improvements when $\mathbf{Dec}_{slot}$ is included. This result suggests that training $\mathbf{Dec}_{slot}$ helps the slots more accurately identify and represent regions with distinct exposure characteristics, leading to enhanced performance. When employing SSAB (Model (e) and (h)) also shows significant performance improvement when used together with $\mathbf{Dec}_{slot}$ . Regarding prompts, although Model (d) shows a slight decrease in SSIM compared to Model (a), all other cases utilizing prompts demonstrate improved performance. Notably, our full model (h) achieves the highest performance, underscoring the importance of each component in enhancing model performance. In Fig. 6, we provide a part of ablations as visual results within state-of-the-art method CSEC [22]. + +Investigation of Model Configuration. Table 4 presents ablation studies on the number of main-slots ( $K^{main}$ ), sub- + +Table 3. Ablation studies on proposed modules. + +
ModelKmainKsubTPSNR↑SSIM↑Time (S)
(a)37322.810.72360.06759
(b)27322.610.72450.06759
(c)47322.360.72010.06759
(d)36321.920.70380.06759
(e)38322.780.72410.06759
(f)37221.150.70580.06747
(g)37421.870.72220.06996
+ +Table 4. Investigation of the number of slots ( $K^{main}$ , $K^{sub}$ ) and number of iterations ( $T$ ) on the SICE [5] dataset. + +
ModelKmainK*sub-1K*sub-2PSNR↑SSIM↑ΔE2000↓[31]ΔEab↓[30]
1-level3--22.020.71317.129.41
2-level37-22.810.72366.948.89
3-level371023.060.73066.848.69
+ +Table 5. Ablation studies on $n$ -level SSAB $\left( {n = 1,2,3}\right)$ + +slots $(K^{sub})$ , and iterations $(T)$ within SSAB to evaluate their impact on performance. The results include PSNR and SSIM metrics on the SICE [5] dataset, with runtime measurements taken on a single $844 \times 1500$ RGB image using an NVIDIA RTX 4090 GPU. In this analysis, we vary $K^{main}$ , $K^{sub}$ , and $T$ by incrementing or decrementing each parameter by 1. Based on the PSNR values, the configuration of $K^{main} = 3$ , $K^{sub} = 7$ , and $T = 3$ yields the highest performance and is therefore selected for our method. + +Effectiveness of Structural Levels. In Sec.3, we use a 2-level SSAB structure as the default configuration. To further investigate the potential of SSAB, we perform experiments using both 1-level and 3-level SSAB structures on the SICE [5] dataset, as summarized in Table 5. In the 3-level configuration, we set $K^{sub-2} = 10$ for the number of second sub-slots, resulting in PSNR and SSIM improvements of 0.25 and 0.07 compared to 2-level SSAB, respectively. These results indicate that adding more SSAB levels can enhance performance. While adding additional levels improves performance, it also leads to an exponential increase in time complexity. Therefore, considering this trade-off, we adopt the 2-level SSAB as the default structure. + +# 5. Conclusion + +In this paper, we present Exposure-slot, a novel framework for exposure correction that integrates slot-in-slot attention with learnable prompts to achieve precise, exposure-centric feature learning. By hierarchically clustering and refining exposure regions, Exposure-slot facilitates accurate, region-aware adjustments and achieves state-of-the-art performance on multiple multi-exposure benchmarks. Our approach demonstrates significant improvements in both quantitative and qualitative metrics, highlighting the effectiveness of structured attention mechanisms for challenging exposure correction scenarios. Furthermore, this work advances exposure correction and paves the way for exploring slot-based architectures in other low-level vision tasks. + +# Acknowledgments + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2023-00222776), and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06, Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Ministry of Science and ICT (No.RS-2020-II201373, Artificial Intelligence Graduate School Program [Hanyang University]). + +# References + +[1] Mahmoud Afifi, Konstantinos G Derpanis, Bjorn Ommer, and Michael S Brown. Learning multi-scale photo exposure correction. In CVPR, 2021. 1, 2, 6, 7 +[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3 +[3] Ondrej Biza, Sjoerd Van Steenkiste, Mehdi S. M. Sajjadi, Gamaleldin F. Elsayed, Aravindh Mahendran, and Thomas Kipf. Invariant slot attention: object discovery with slot-centric reference frames. In ICML, 2023. 3 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3 +[5] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27, 2018. 2, 6, 7, 8 +[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2 +[7] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, 2023. 3 +[8] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In CVPR, 2024. 3 +[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 6, 7 +[10] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 2016. 1 +[11] Yun He, Steven Zheng, Yi Tay, Jai Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, YaGuang Li, Zhao Chen, Donald Met + +zler, et al. Hyperprompt: Prompt-based task-conditioning of transformers. In ICML. PMLR, 2022. 3 +[12] Jie Huang, Yajing Liu, Xueyang Fu, Man Zhou, Yang Wang, Feng Zhao, and Zhiwei Xiong. Exposure normalization and compensation for multiple-exposure correction. In CVPR, 2022. 1, 2, 6, 7 +[13] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In ECCV, 2022. 1, 2, 6, 7 +[14] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In ACMMM, 2022. 1, 2, 6, 7 +[15] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In CVPR, 2023. 2, 6 +[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3 +[17] Donggoo Jung, Daehyun Kim, and Tae Hyun Kim. Continuous exposure learning for low-light image enhancement using Neural ODEs. In ICLR, 2025. 2 +[18] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In CVPR, 2024. 3 +[19] Markus Krimmel, Jan Achterhold, and Joerg Stueckler. Attention normalization impacts cardinality generalization in slot attention. Transactions on Machine Learning Research, 2024. 3 +[20] Minhyeok Lee, Suhwan Cho, Dogyoon Lee, Chaewon Park, Jungho Lee, and Sangyoun Lee. Guided slot attention for unsupervised video object segmentation. In CVPR, 2024. 3 +[21] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In CVPR, 2021. 3 +[22] Yiyu Li, Ke Xu, Gerhard Petrus Hancke, and Rynson WH Lau. Color shift estimation-and-correction for image enhancement. In CVPR, 2024. 1, 2, 6, 7, 8 +[23] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 6, 7 +[24] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 2, 5, 6 +[25] Minh-Thang Luong. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 2 +[26] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 6, 7 +[27] Ntumba Elie Nsampi, Zhongyun Hu, and Qing Wang. Learning exposure correction via consistency modeling. In BMVC, 2021. 2 + +[28] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. PromptIR: Prompting for all-in-one image restoration. In NeurIPS, 2023. 3, 6, 7 +[29] Rishav Pramanik, José-Fabian Villa-Vásquez, and Marco Pedersoli. Masked Multi-Query slot attention for unsupervised object discovery. In IJCNN, 2024. 3 +[30] Gaurav Sharma and Raja Bala. Digital color imaging handbook. CRC press, 2017. 7, 8 +[31] Gaurav Sharma, Wencheng Wu, and Edul N Dalal. The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application, 2005. 7, 8 +[32] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. CODA-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, 2023. 3 +[33] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, 2023. 3 +[34] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 3 +[35] Haoyuan Wang, Ke Xu, and Rynson WH Lau. Local color distributions prior for image enhancement. In ECCV. Springer, 2022. 1, 2, 6 +[36] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 2 +[37] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 1 +[38] Yanbo Wang, Letao Liu, and Justin Dauwels. Slot-vae: Object-centric scene generation with slot attention. In ICML, 2023. 3 +[39] Yang Wang, Long Peng, Liang Li, Yang Cao, and Zheng-Jun Zha. Decoupling-and-aggregating for image exposure correction. In CVPR, 2023. 1, 6 +[40] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep Retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 6 +[41] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. URetinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 2 +[42] Mingde Yao, Jie Huang, Xin Jin, Ruikang Xu, Shenglong Zhou, Man Zhou, and Zhiwei Xiong. Generalized lightness adaptation with channel selective normalization. In CVPR, 2023. 2 +[43] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In ACMM, 2019. 2 +[44] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129, 2021. 2 + +[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3 +[46] Karel Zuiderveld. Contrast limited adaptive histogram equalization. Graphics gems, 1994. 6 \ No newline at end of file diff --git a/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/images.zip b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ebb6cb74d76086cbe6263936c46a1f184fead4b --- /dev/null +++ b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:980b9c72c4900221278c63c6cd416ccb46f9b612454f87d7444c193153a82e14 +size 685994 diff --git a/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/layout.json b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a3cbbffbbfde7755e02bcbdeb3cb5299e91b4802 --- /dev/null +++ b/2025/Exposure-slot_ Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction/layout.json @@ -0,0 +1,10977 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "type": "text", + "content": "Exposure-slot: Exposure-centric representations learning with Slot-in-Slot Attention for Region-aware Exposure Correction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 477, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 477, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 477, + 190 + ], + "type": "text", + "content": "Donggoo Jung\\*, Daehyun Kim\\*, Guanghui Wang, Tae Hyun Kim† Hanyang University, Toronto Metropolitan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 191, + 482, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 191, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 127, + 191, + 482, + 204 + ], + "type": "text", + "content": "{dgjung, daehyun, taehyunkim}@hanyang.ac.kr, {wangcs}@torontomu.ca" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 256, + 297, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 297, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 297, + 520 + ], + "type": "text", + "content": "Image exposure correction enhances images captured under diverse real-world conditions by addressing issues of under- and over-exposure, which can result in the loss of critical details and hinder content recognition. While significant advancements have been made, current methods often fail to achieve optimal feature learning for effective correction. To overcome these challenges, we propose Exposure-slot, a novel framework that integrates a prompt-based slot-in-slot attention mechanism to cluster exposed feature regions and learn exposure-centric features for each cluster. By extending the Slot Attention algorithm with a hierarchical structure, our approach progressively clusters features, enabling precise and region-aware correction. In particular, learnable prompts tailored to exposure characteristics of slots further enhance feature quality, adapting dynamically to varying conditions. Our method delivers superior performance on benchmark datasets, surpassing the current state-of-the-art with a PSNR improvement of over 1.85 dB on the SICE dataset and 0.4 dB on the LCDP dataset, thereby establishing a new benchmark for multi-exposure correction. The source code can be found at: https://github.com/kdhRick2222/Exposure-slot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 542, + 135, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 135, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 135, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "type": "text", + "content": "Image exposure correction enhances images captured under diverse real-world conditions. Under- or over-exposed images can appear excessively dark or bright, losing essential details and compromising accurate recognition. Despite advancements in imaging technology, robust automatic exposure correction remains a significant challenge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "Due to its critical importance, image exposure correction has been the focus of considerable research. Early efforts primarily treated under-exposure and over-exposure as separate issues [10, 37, 40]. More recent advancements, however, have introduced multi-exposure correction mod" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 555, + 401 + ], + "type": "image", + "image_path": "b1e3638b18597a88089e950bfec1c3792e50485f2d9030b3ff6aab4306168e0d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 405, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 483 + ], + "type": "text", + "content": "Figure 1. (Top) The Slot-in-Slot Attention mechanism hierarchically partitions exposure-aware regions using the attention maps of main and sub-slot attention, performing corrections with learnable prompts based on exposure levels. (Bottom) Comparison of the Exposure-slot with existing approaches. From left to right: ECLNet [14], FECNet [13], DRBN-ENC [12], CSEC [22] and our proposed approach, Exposure-slot." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "els capable of handling a wide range of exposure levels within a unified framework. For instance, MSEC [1] introduced an approach that considers both color and detail enhancement while performing multi-exposure correction in an end-to-end manner, and provided a dataset of images with different exposure errors for both training and evaluation. In addition, LCDPNet [35] provided another dataset of diverse scenes, covering both over-exposure and under-exposure, and incorporates retinax theory to account for local color distributions within images. Recent approaches increasingly focus on the separate processing of features based on their physical characteristics, such as exposure, frequency, contrast, color, and details. FECNet [13] uses Fourier transform to facilitate interactions between local spatial feature information and global frequency information, while ECLNet [14] uses a bilateral activation mechanism to process over- and under-exposed regions independently. In addition, DA [39] introduces a decoupling and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 222, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17892" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "aggregation scheme that separately enhances image details and contrast." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 294, + 180 + ], + "type": "text", + "content": "However, relying on separating image features based on these physical properties may not ensure the optimal features for effective exposure correction. For example, at the bottom of Fig. 1, we visualize this limitation, showing that previous works [13-15, 22] often generate inappropriate artifacts in sky regions and struggle to accurately adjust sunset exposure levels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 326 + ], + "type": "text", + "content": "To overcome these challenges, we propose a novel prompt-based slot-in-slot attention mechanism that clusters regions with similar exposure levels, enabling exposure-centric feature learning within each cluster. We utilize Slot Attention [24], designed to cluster features based on object representations, to identify distinct exposure regions within an image. In our approach, slots represent features grouped by exposure levels, and the corresponding attention maps act as exposure-aware region maps, guiding the grouping of features according to exposure characteristics. Therefore, these maps highlight regions based on exposure levels, enabling targeted exposure correction for each area." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 471 + ], + "type": "text", + "content": "To support effective correction, we introduce learnable prompt vectors that operate based on exposure-aware region maps, which are attention maps of each slot. These prompts adapt to the characteristics of each region defined by the attention map, helping the model learn targeted correction to improve under- or over-exposed areas. Specifically, we introduce Slot-Prompt Interaction Module (SPIM), which combines slot and prompt information using cross-attention. The cross-attention component of SPIM uses the prompts learned through Slot-in-Slot Attention as conditioning factors, enhancing the interaction between features and prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 592 + ], + "type": "text", + "content": "Our approach builds upon standard slot attention with a hierarchical design, called Slot-in-Slot Attention, which processes slots in levels. This mechanism operates across multiple structure levels: the first level coarsely partitions regions based on distinct exposure characteristics, while subsequent levels progressively refine these regions with more partitions. Each slot captures exposure-specific regional information, distinguishing areas with varying exposure levels to enhance feature separation and improve the precision of local details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "The overall framework of the proposed approach is illustrated at the top of Fig. 1. The Slot-in-Slot Attention is applied to intermediate features between the encoder and decoder, generating attention maps for slots that are hierarchically partitioned according to exposure characteristics, as depicted in Fig. 1. The generated attention maps are then used as weight maps for learnable prompts, resulting in refined intermediate features that are subsequently fed into the decoder. We refer to our model as Exposure-Slot, as it is the first model to use Slot Attention for unsupervised feature" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "partitioning based on exposure levels in exposure correction, and the first to apply region-aware prompts for feature enhancement. Our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 110, + 553, + 230 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 110, + 553, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 110, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 314, + 110, + 553, + 146 + ], + "type": "text", + "content": "- Exposure-slot is the first approach to leverage Slot Attention mechanism for optimized exposure-specific feature partitioning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 146, + 553, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 553, + 171 + ], + "type": "text", + "content": "- We introduce the slot-in-slot attention that enables sophisticated feature partitioning and learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "content": "- We apply exposure-aware prompts that enhance the exposure-centric characteristics of each image feature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 195, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 553, + 230 + ], + "type": "text", + "content": "- Exposure-slot achieves state-of-the-art results on multi-exposure benchmark datasets [1, 5, 35], setting a new standard in this field." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 242, + 400, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 242, + 400, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 400, + 255 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 262, + 435, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 435, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 435, + 275 + ], + "type": "text", + "content": "2.1. Exposure Correction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 280, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 507 + ], + "type": "text", + "content": "With the rise of deep neural networks (DNN), exposure correction has seen the emergence of a range of DNN-based methods leveraging diverse concepts. For under-exposed image enhancement, methods [36, 40, 41, 43, 44] inspired by retinax theory have been proposed. Retinex-based models like CMEC [27] have also incorporated attention mechanisms to address multi-exposure correction, while LCDP-Net [35] introduced a novel focus on local color distribution. To support multi-exposure correction tasks, dedicated datasets such as MSEC [1] and SICE [5] were developed for training and evaluation. MSEC further introduced a Laplacian pyramid architecture to handle varying exposure levels. ENC [12] contributed an exposure normalization module that refined feature maps by transforming diverse exposure features into an exposure-invariant feature space. CSNorm [42] also enhanced model generalization by selectively normalizing lightness-sensitive channels and ERL [15] proposed exposure relationship learning through regularization techniques for multi-exposure correction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 508, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 639 + ], + "type": "text", + "content": "There are several recent studies that emphasize the feature separation based on their physical characteristics. ECLNet [14] used a bilateral activation mechanism to adjust processing across different exposure conditions, while FECNet [13] proposed a lightweight model that utilizes spatial-frequency interactions with a Fourier-based approach. CLODE [17] also utilizes image curve adjustment and formulates brightness changes through continuous curve map extraction using Neural ODEs [6]. The latest advancement, CSEC [22], addresses color distribution shifts by defining darkened and brightened feature maps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 647, + 402, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 647, + 402, + 659 + ], + "spans": [ + { + "bbox": [ + 314, + 647, + 402, + 659 + ], + "type": "text", + "content": "2.2. Slot Attention" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Slot Attention [24] is a mechanism for clustering features corresponding to slots which is the main object of object-centric learning (OCL). Specifically, this mechanism iteratively employs dot-product attention [25] with attention co" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17893" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 68, + 287, + 265 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 287, + 265 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 287, + 265 + ], + "type": "image", + "image_path": "c370d44eacae54c2cff85279b6d38a663e2fbe90d27a4f7d16e5dff89a2c3918.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 274, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 329 + ], + "type": "text", + "content": "Figure 2. Overview of Exposure-Slot: Exposure-Slot operates within a " + }, + { + "bbox": [ + 54, + 274, + 555, + 329 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 54, + 274, + 555, + 329 + ], + "type": "text", + "content": "-shaped encoder-decoder network, with the Slot-Prompt Interaction Module (SPIM) bridging the encoder and decoder stages. In SPIM, Slot-in-Slot Attention Block (SSAB) adaptively partitions regions and generates prompt features to enhance the feature representation. In addition, a Slot Decoder " + }, + { + "bbox": [ + 54, + 274, + 555, + 329 + ], + "type": "inline_equation", + "content": "(\\mathbf{Dec}_{slot})" + }, + { + "bbox": [ + 54, + 274, + 555, + 329 + ], + "type": "text", + "content": " is employed during training to reconstruct slot features to the target. This ensures that slot attention maps capture exposure-centric information, helping prompts learn the relevant information for enhancement within each corresponding slot." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 288, + 68, + 527, + 264 + ], + "blocks": [ + { + "bbox": [ + 288, + 68, + 527, + 264 + ], + "lines": [ + { + "bbox": [ + 288, + 68, + 527, + 264 + ], + "spans": [ + { + "bbox": [ + 288, + 68, + 527, + 264 + ], + "type": "image", + "image_path": "7493085b6a97d7e99cfe6816dda50509285af7115b32dc2ead4cd8362e9012a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 345, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 345, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 295, + 392 + ], + "type": "text", + "content": "efficients to update slots, which function as queries within the attention process. Through multiple rounds of recurrent attention, Slot Attention achieves effective clustering and feature separation in OCL without the need for annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 394, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 394, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 295, + 478 + ], + "type": "text", + "content": "Slot Attention has been effectively applied to image classification, as seen in SCOUTER [21, 33], which uses a Slot-Attention-based classifier for explainable recognition. In segmentation, Slot Attention has proven particularly valuable for video segmentation [3, 8, 19, 29]. For example, GSANet [20] utilizes Slot Attention to separate center objects from background elements across video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 647 + ], + "type": "text", + "content": "In reconstruction and restoration tasks, Slot-VAE [38] integrates Slot Attention with a hierarchical Variational Autoencoder (VAE) framework, enhancing object-centric scene generation. Moreover, AID [18] employs Slot Attention to capture implicit representations of illuminant chromaticities, with each slot vector capturing the characteristics of a specific illuminant. This enables AID to generate accurate chromaticity and weight maps for each light source. In this paper, we leverage Slot Attention for partitioning in unsupervised way for the exposure correction. This allows our method to train each slot to capture feature-relevant exposure representations and effectively to cluster features implicitly for exposure correction without requiring annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 658, + 192, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 192, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 192, + 672 + ], + "type": "text", + "content": "2.3. Prompt-Based Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "The prompt-based learning method was first introduced by [4], proposing the concept of optimizing input prompts. This approach gained popularity in the field of natural lan" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 345, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 632 + ], + "type": "text", + "content": "guage processing, inspiring numerous follow-up studies. CoOp [45] introduced a novel approach by treating prompt contexts as learnable parameters, moving beyond methods limited to fixed-format prompts, and demonstrating that learnable prompts can outperform handcrafted ones. On the basis of CoOp, several methods have been developed for dynamically generating suitable prompts [7, 11, 32]. For example, CODA [32] generates prompts tailored to specific inputs, while HyperPrompt [11] addresses multi-task learning through the use of a prompt generator. In the computer vision field, visual prompts involve adding trainable parameters to modify inputs for efficient model adaptation. VPT [16] demonstrated substantial performance improvements over conventional fine-tuning methods by applying visual prompts to a fixed transformer backbone. Furthermore, [2] presented visual prompts compatible with input images, specifically optimized for CLIP. For low-level vision tasks, methods like PromptIR [28] and PromptReformer [34] leverage prompts to encode degradation-specific information, guiding restoration networks to adapt to various types and intensities of degradation. In our method, we introduce exposure-specific prompts that guide feature separation and generate region-aware features specifically tailored to distinct exposure characteristics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 640, + 421, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 640, + 421, + 654 + ], + "spans": [ + { + "bbox": [ + 314, + 640, + 421, + 654 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 659, + 398, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 398, + 672 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 398, + 672 + ], + "type": "text", + "content": "3.1. Overall Flow" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 2, the proposed Exposure-slot architecture is a " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "-shaped residual network composed of an encoder-decoder structure connected via skip connections" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17894" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 55, + 97, + 555, + 395 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 555, + 95 + ], + "type": "text", + "content": "Algorithm 1 Slot-in-Slot Attention Block (SSAB). For clarity, we represent the algorithm using a 2-level hierarchical structure, consisting of main- and sub-slots." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": "Input: K, V, slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", prompts" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", prompts" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " \n1: function SLOT_UPDATE(slots, k, v) \n2: slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "_{prev}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " = slots \n3: slots = LayerNorm(slots) \n4: q = to_q(slots) \n5: attn = Softmax " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\left(\\frac{1}{\\sqrt{D}} k \\cdot q^T\\right)" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " \n6: updates = WeightedMean(weights=attn + ε, values = v) \n7: slots = GRU(state=slot" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "_{prev}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", inputs=updates) \n8: return slots, attn \n9: end function \n10: \n11: k = K, v = V \n12: for " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "t_{main} = 0 \\ldots T" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " do \n13: slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", attn" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " = SLOT_UPDATE(slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", k, v) \n14: P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " = prompts" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " × attn" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " \n15: k, v = k " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", v " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " \n16: for " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "t_{sub} = 0 \\ldots T" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " do \n17: slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", attn" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " = SLOT_UPDATE(slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", k, v) \n18: end for \n19: P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " = prompts" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " × attn" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " \n20: k, v = k " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", v " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " \n21: end for \n22: \n23: P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{final}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " = P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " \n24: S" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{final}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " = (slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " × attn" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ") " + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " (slots" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": " × attn" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ") \n25: return P" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{final}" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "text", + "content": ", S" + }, + { + "bbox": [ + 55, + 97, + 555, + 395 + ], + "type": "inline_equation", + "content": "^{final}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 415, + 239, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 239, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 239, + 426 + ], + "type": "text", + "content": "and Slot-Prompt Interaction Module (SPIM)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": "First, the encoder processes the poorly exposed input image " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": " to produce a latent feature representation " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "F = \\mathbf{Enc}(I_{in})" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": ". Then, SPIM, positioned between the encoder and decoder, refines the encoder output " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": " using input-specific, region-aware prompts guided by the Slot-in-Slot Attention mechanism. The refined latent feature " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": " is subsequently passed to the image decoder (Dec_enhance), which processes it to produce the output image " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "I_{out}" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": " with well-enhanced exposure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 266, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 266, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 266, + 553 + ], + "type": "text", + "content": "3.2. Slot-Prompt Interaction Module (SPIM)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 677 + ], + "type": "text", + "content": "SPIM is designed as a combination of the Slot-in-Slot Attention Block (SSAB) and a subsequent cross-attention process. SPIM initially projects " + }, + { + "bbox": [ + 55, + 557, + 296, + 677 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 557, + 296, + 677 + ], + "type": "text", + "content": " into query, key, and value representations, while SSAB receives the key and value to estimate the corresponding prompt features of " + }, + { + "bbox": [ + 55, + 557, + 296, + 677 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 557, + 296, + 677 + ], + "type": "text", + "content": ". Specifically, SSAB hierarchically applies slot attention to achieve coarse-to-fine region separation, which supports effective exposure correction. Then the derived prompt feature serves as conditioning elements in the cross-attention step involving the query, key, and value representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Specifically, as illustrated in Fig. 2, the encoded feature " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is first flattened into " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is projected into query " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", key" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": " through the linear projections " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "W_{\\mathbf{Q}}, W_{\\mathbf{K}}" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "W_{\\mathbf{V}}" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 343, + 447, + 555, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 447, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 343, + 447, + 555, + 464 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = X W _ {\\mathbf {Q}} ^ {T}, \\quad \\mathbf {K} = X W _ {\\mathbf {K}} ^ {T}, \\quad \\mathbf {V} = X W _ {\\mathbf {V}} ^ {T}. \\qquad (1)", + "image_path": "3e23b56ad2e4d30e6e9cd84df747a5b02eacc433a20d7981b0251071d682e15d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "text", + "content": "Within our SSAB, " + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "text", + "content": " are used to generate the prompt feature " + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{final}" + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "text", + "content": ", which serves as a condition for the subsequent cross-attention process to produce the refined feature map " + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "inline_equation", + "content": "F' \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 470, + 555, + 518 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 526, + 555, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 526, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 326, + 526, + 555, + 554 + ], + "type": "interline_equation", + "content": "F ^ {\\prime} = r e s h a p e (s o f t m a x (\\frac {\\mathbf {Q K} ^ {T}}{\\sqrt {d}}) \\cdot (\\mathbf {V} \\cdot \\mathbf {P} ^ {f i n a l})), \\quad (2)", + "image_path": "27817105788fc23f590836b848912a3372b2ad2049719d3f40c58c3cd631de05.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 561, + 554, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 554, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 561, + 554, + 585 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 561, + 554, + 585 + ], + "type": "text", + "content": " denotes a learnable parameter that adaptively scales matrix multiplication." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 495, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 495, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 495, + 602 + ], + "type": "text", + "content": "3.2.1. Slot-in-Slot Attention Block (SSAB)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 676 + ], + "type": "text", + "content": "In the proposed Slot-in-Slot Attention Block within SPIM, we generate a refined feature map by integrating a hierarchically structured slot mechanism, which partitions regions with different exposure levels in a hierarchical and iterative manner, with learned prompts specifically tailored to distinct exposure characteristics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "For instance, in a 2-level slot-in-slot structure, the first level partitions regions with similar exposure characteristics using a relatively smaller number of slots. Building on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17895" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "content": "these results, the second level employs more slots to cluster the regions more finely. By iteratively repeating this process, the sub-slots gradually learn finer details of differently exposed regions. Notably, the hierarchical structure can extend beyond two levels, enabling the identification of exposure-centric feature maps with increasing precision. Next, we introduce exposure-specific prompts that enhance the image by targeting distinct regional exposure characteristics, multiplied by attention maps from the slot attention mechanism. Since the attention maps identify regions with different exposure levels, multiplying them with exposure-specific prompts generates region-aware guidance, enabling balanced exposure correction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": "In Alg. 1, the pseudo-code for this 2-level slot-in-slot structure is provided, refines key " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "(v)" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " through two nested loops: an outer loop and an inner loop. The outer loop updates the main-slots, which are then combined with prompts to refine " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": ". In the inner loop, the refined " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " are used to update the sub-slots, which are further combined with prompts to refine " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " with fine details. This process is repeated until the outer loop is completed. Notably, to update the slots, we employ the SLOT_UPDATE function introduced process in [24]. The key difference is that each loop iteration uses the updated " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": " as inputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "text", + "content": "Main-slot Attention In the outer loop of Alg. 1, we perform Main-slot Attention to update the main-slots (slots" + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "text", + "content": "), which represent the first level of the hierarchical slot structure. Notably, slots" + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "inline_equation", + "content": "^{main} \\in \\mathbb{R}^{K^{main} \\times D_{slot}^{main}}" + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "inline_equation", + "content": "K^{main}" + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "inline_equation", + "content": "D_{slot}^{main}" + }, + { + "bbox": [ + 55, + 369, + 295, + 453 + ], + "type": "text", + "content": " indicate the number of slots and the dimensionality of each slot, respectively, can be initialized with learnable parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "text", + "content": "First, given " + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "text", + "content": " from the Encoder, assigned to " + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 453, + 295, + 489 + ], + "type": "text", + "content": ", respectively, we update the slots using the SLOT_UPDATE function as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 495, + 295, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 495, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 59, + 495, + 295, + 518 + ], + "type": "interline_equation", + "content": "\\mathbf {s l o t s} ^ {\\text {m a i n}}, \\mathbf {a t t n} ^ {\\text {m a i n}} = \\mathbf {S L O T} _ {-} \\mathbf {U P D A T E} (\\mathbf {s l o t s} ^ {\\text {m a i n}}, k, v), \\tag {3}", + "image_path": "8f8cc333437cf715582f7e32cd08890c320830e2e9d8849303ced558a6af2d8f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{attn}^{main}" + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": " represents the attended region of each slots" + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "inline_equation", + "content": "^{main}" + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": ". Then, the attention map " + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{attn}^{main}" + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": " is used to produce prompt feature " + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{main}" + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 561, + 294, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 561, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 94, + 561, + 294, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {\\text {m a i n}} = \\text {p r o m p t s} ^ {\\text {m a i n}} \\times \\text {a t t n} ^ {\\text {m a i n}}, \\tag {4}", + "image_path": "604005824d1abaa6252ae85e02762785fb360d43a13b79c442875102c93817b5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\text{main}} \\in \\mathbb{R}^{K^{\\text{main}} \\times D_{\\text{slot}}^{\\text{main}}}" + }, + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "text", + "content": " denotes learnable prompts tailored to each exposure characteristics. Thus, by multiplying the exposure-specific prompts with the attention map containing distinct exposure region information, our prompt feature " + }, + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\text{main}}" + }, + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "text", + "content": " effectively encapsulates exposure-centric and region-aware information. Finally, prompt feature " + }, + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\text{main}}" + }, + { + "bbox": [ + 55, + 578, + 295, + 674 + ], + "type": "text", + "content": " is used to update the key and value as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 675, + 294, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 675, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 116, + 675, + 294, + 689 + ], + "type": "interline_equation", + "content": "k, v = k \\cdot \\mathbf {P} ^ {\\text {m a i n}}, v \\cdot \\mathbf {P} ^ {\\text {m a i n}}, \\tag {5}", + "image_path": "441e8a51be98bc68b7915f904e44b90580e874500b16202cfb6e82de9dc12513.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 693, + 295, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 717 + ], + "type": "text", + "content": "and the resulting " + }, + { + "bbox": [ + 55, + 693, + 295, + 717 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 693, + 295, + 717 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 693, + 295, + 717 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 693, + 295, + 717 + ], + "type": "text", + "content": " are used as inputs for the subsequent inner loop process, referred to as Sub-slot Attention." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "content": "Sub-slot Attention Within the inner loop of Alg. 1, we perform Sub-slot Attention to update the sub-slots (slot " + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "inline_equation", + "content": "s u b" + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "content": "), which represent the second level of hierarchical structure. Particularly, slots " + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "inline_equation", + "content": "s u b \\in \\mathbb{R}^{K^{s u b} \\times D_{s l o t}^{s u b}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "content": ", where slots " + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "inline_equation", + "content": "s u b" + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "content": " are also initialized using learnable parameters. Here, " + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "inline_equation", + "content": "K^{s u b}" + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "content": " denotes the number of sub-slots, " + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "inline_equation", + "content": "D_{s l o t}^{s u b}" + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "content": " denotes their dimensionality, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "inline_equation", + "content": "K^{s u b}" + }, + { + "bbox": [ + 313, + 72, + 553, + 169 + ], + "type": "text", + "content": " is configured to ensure finer feature separation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "text", + "content": "First, the updated " + }, + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "text", + "content": " from the Main-slot Attention in Eq. 5 are used to update slots" + }, + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 313, + 170, + 553, + 217 + ], + "type": "text", + "content": " via the SLOT_UPDATE function. Similarly to Eq. 3, Sub-slot Attention is updated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 226, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 226, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 226, + 553, + 239 + ], + "type": "interline_equation", + "content": "\\text {s l o t s} ^ {\\text {s u b}}, \\text {a t t n} ^ {\\text {s u b}} = \\text {S L O T} _ {-} \\text {U P D A T E} (\\text {s l o t s} ^ {\\text {s u b}}, k, v), \\tag {6}", + "image_path": "e1b1b8c8c9c8f098904204c013dfe16cae1d73b13e35091ae187e493bfa8fcde.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{attn}^{sub}" + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "text", + "content": " represents the attended region of each slots" + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "inline_equation", + "content": "^{sub}" + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "text", + "content": ", and this update process is repeated for " + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "text", + "content": " iterations. Subsequently, " + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{attn}^{sub}" + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "text", + "content": " is used to obtain sub-prompt feature " + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{sub}" + }, + { + "bbox": [ + 313, + 248, + 553, + 296 + ], + "type": "text", + "content": " from the Sub-slot Attention as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 304, + 553, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 304, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 364, + 304, + 553, + 319 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {s u b} = \\text {p r o m p t s} ^ {s u b} \\times \\text {a t t n} ^ {s u b}, \\tag {7}", + "image_path": "f9159cbd0fe59b015a7c7a77e1aa16629aeb971ecfa6fc2b32b0aa4743e18ecc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\mathrm{sub}} \\in \\mathbb{R}^{K^{\\mathrm{sub}} \\times D_{\\mathrm{slot}}^{\\mathrm{sub}}}" + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "content": " is also learnable parameters tailored to each exposure characteristic in Sub-slot Attention. Finally, by multiplying " + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{sub}}" + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "content": ", we transfer information from the Sub-slot Attention for the next iteration of the Main-slot Attention as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 397, + 553, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 397, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 381, + 397, + 553, + 411 + ], + "type": "interline_equation", + "content": "k, v = k \\cdot \\mathbf {P} ^ {s u b}, v \\cdot \\mathbf {P} ^ {s u b}. \\tag {8}", + "image_path": "6867113a4bed044cde797a45610798790d2817762b7b4beef0729f6549400912.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "text", + "content": "To summarize, the result of the Main-slot Attention is integrated into the Sub-slot Attention, and the output from the Sub-slot Attention is iteratively fed back into the Main-slot Attention loop. This complementary updating process is repeated " + }, + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "text", + "content": " times. At the end of SSAB, the final outputs " + }, + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{final}" + }, + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{final}" + }, + { + "bbox": [ + 313, + 421, + 553, + 491 + ], + "type": "text", + "content": " are obtained as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 323, + 498, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 498, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 323, + 498, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {P} ^ {f i n a l} = \\mathbf {P} ^ {m a i n} \\cdot \\mathbf {P} ^ {s u b}, \\\\ \\mathbf {S} ^ {\\text {f i n a l}} = \\left(\\mathbf {s l o t s} ^ {\\text {m a i n}} \\times \\mathbf {a t t n} ^ {\\text {m a i n}}\\right) \\cdot \\left(\\mathbf {s l o t s} ^ {\\text {s u b}} \\times \\mathbf {a t t n} ^ {\\text {s u b}}\\right). \\tag {9} \\\\ \\end{array}", + "image_path": "02466501cdc79bfba1bf69f91d7e37135102224b335337b559e27bb2fbd45fb8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{final}" + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "text", + "content": " is integrated with " + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "text", + "content": " as described in Eq. 2, and is used to decode the final output. Meanwhile " + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{final}" + }, + { + "bbox": [ + 313, + 537, + 553, + 586 + ], + "type": "text", + "content": " serves as an input to a separate decoder to facilitate slot training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 593, + 413, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 413, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 413, + 604 + ], + "type": "text", + "content": "3.3.Decoder Process" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "text", + "content": "The refined feature " + }, + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "text", + "content": " and slots " + }, + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "inline_equation", + "content": "S^{final}" + }, + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "text", + "content": " from SPIM are used as inputs for the final decoding procedures. Specifically, our Exposure-slot uses the enhanced feature " + }, + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 313, + 609, + 553, + 658 + ], + "type": "text", + "content": " to predict an enhanced image through the decoder, as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 382, + 667, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 667, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 382, + 667, + 553, + 681 + ], + "type": "interline_equation", + "content": "I _ {o u t} = \\operatorname {D e c} _ {\\text {e n h a n c e}} \\left(F ^ {\\prime}\\right), \\tag {10}", + "image_path": "7abdee68715baac82ffe939d39b11dd33dba46be3ff9a0cd7d453b42fbc0390c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Dec}_{\\text{enhance}}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " is image decoder and " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{out}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " is the final exposure-enhanced output." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17896" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 236 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 236 + ], + "type": "table", + "html": "
Model#Params (M)SICE [5]MSEC [1]LCDP [35] Avg.
UnderOverAvg.UnderOverAvg.
CLAHE [46]-12.69 / 0.503710.21 / 0.487811.45 / 0.494216.77 / 0.621114.45 / 0.584215.38 / 0.599016.33 / 0.6420
RetinexNet [40]0.84012.94 / 0.517112.87 / 0.525212.90 / 0.521212.13 / 0.620910.47 / 0.595311.14 / 0.604819.25 / 0.7041
ZeroDCE [9]0.07916.92 / 0.63307.11 / 0.429212.02 / 0.531114.55 / 0.588710.40 / 0.514212.06 / 0.544112.59 / 0.6530
RUAS [23]0.00216.63 / 0.55894.54 / 0.319610.59 / 0.439313.43 / 0.68076.39 / 0.46559.20 / 0.551513.76 / 0.6060
SCI [26]0.00117.86 / 0.64014.45 / 0.362912.49 / 0.50519.97 / 0.66815.84 / 0.51907.49 / 0.578611.87 / 0.5234
MSEC [1]7.04019.62 / 0.651217.59 / 0.656018.58 / 0.653620.52 / 0.812919.79 / 0.815620.08 / 0.821020.38 / 0.7800
LCDPNet [35]0.96017.45 / 0.562217.04 / 0.646317.25 / 0.604322.35 / 0.865022.17 / 0.847622.30 / 0.855223.24 / 0.8420
ECLNet [14]0.01822.05 / 0.689319.25 / 0.687220.65 / 0.686122.37 / 0.856622.70 / 0.867322.57 / 0.863122.44 / 0.8061
FECNet [13]0.15022.01 / 0.673719.91 / 0.696120.96 / 0.684922.96 / 0.859823.22 / 0.874823.12 / 0.868822.41 / 0.8402
DRBN-ENC [12]0.58021.89 / 0.707119.09 / 0.722920.49 / 0.715022.72 / 0.859022.11 / 0.852122.35 / 0.853022.09 / 0.8271
MSEC+DA [39]7.04020.94 / 0.754617.49 / 0.664019.22 / 0.709321.53 / 0.859021.55 / 0.875021.54 / 0.867021.05 / 0.8119
ECLNet+ERL [15]0.01822.14 / 0.690819.47 / 0.698220.81 / 0.694522.90 / 0.862422.58 / 0.867622.70 / 0.865522.63 / 0.8096
PromptIR [28]34.16422.51 / 0.695519.29 / 0.684920.90 / 0.690215.80 / 0.739116.73 / 0.785216.36 / 0.766823.49 / 0.8513
CSEC [22]1.36420.79 / 0.703120.02 / 0.709320.41 / 0.706222.18 / 0.850222.69 / 0.866222.73 / 0.863823.63 / 0.8550
Exposure-slot1.22923.85 / 0.709221.77 / 0.737522.81 / 0.723923.09 / 0.860123.24 / 0.876223.18 / 0.869724.03 / 0.8592
", + "image_path": "cd420e70bd6f84a7a28ad8ab20cd915288fc6ce16619c5f8dbbad03cfa04e786.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 270 + ], + "type": "text", + "content": "Table 1. Quantitative results on SICE [5], MSEC [1], and LCDP [35] in terms of PSNR↑/SSIM↑. The best score is displayed in red, the second in blue. Additionally, the number of parameters (#Params) required for inference is also specified. Exposure-slot is lightweight compared to the previous SOTA approach [22] while consistently outperforming it across all datasets on average." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 316 + ], + "type": "text", + "content": "Meanwhile, " + }, + { + "bbox": [ + 55, + 291, + 295, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{final}" + }, + { + "bbox": [ + 55, + 291, + 295, + 316 + ], + "type": "text", + "content": " is decoded into a pseudo-corrected image using the slot reconstruction decoder:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 319, + 294, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 319, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 117, + 319, + 294, + 335 + ], + "type": "interline_equation", + "content": "I _ {p s e u d o} = \\mathbf {D e c} _ {s l o t} \\left(\\mathbf {S} ^ {f i n a l}\\right), \\tag {11}", + "image_path": "5a927c12733e8f31363002117156ed08ca7a6c5772cff441a366cdc49d6ddffc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{Dec}_{slot}" + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "content": " represents the slot reconstruction decoder and " + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "I_{pseudo}" + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "content": " is the resulting pseudo-corrected RGB image. This training strategy encourages each slot to focus on exposure-centric learning, similar to the set prediction approach used in [24]. It is important to note that the slot reconstruction decoder is used only during training and is not required during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 427, + 146, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 146, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 146, + 439 + ], + "type": "text", + "content": "3.4. Loss functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 504 + ], + "type": "text", + "content": "Exposure-slot is trained using two loss functions: image enhancement loss " + }, + { + "bbox": [ + 55, + 445, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {enhance }}" + }, + { + "bbox": [ + 55, + 445, + 296, + 504 + ], + "type": "text", + "content": " and slot reconstruction loss " + }, + { + "bbox": [ + 55, + 445, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {slot }}" + }, + { + "bbox": [ + 55, + 445, + 296, + 504 + ], + "type": "text", + "content": ". The overall training objective is to minimize the combined loss, promoting both accurate image enhancement and effective slot-based reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "text", + "content": "Image Enhancement loss. The image enhancement loss " + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {enhance }}" + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "text", + "content": " is defined using the " + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "text", + "content": " distance between the exposure enhanced image " + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "inline_equation", + "content": "I_{\\text {out }}" + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "text", + "content": " and the ground truth image " + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "inline_equation", + "content": "I_{\\text {gt }}" + }, + { + "bbox": [ + 55, + 518, + 296, + 566 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 571, + 295, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 571, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 119, + 571, + 295, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n h a n c e}} = \\left\\| I _ {\\text {o u t}} - I _ {\\text {g t}} \\right\\| _ {1}. \\tag {12}", + "image_path": "29c395fe33e8c6c7de3d2333d9698a0db4603739effcc1deec2ac8b644d9b795.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "text", + "content": "Slot Reconstruction loss. The slot reconstruction loss " + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{slot}" + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "text", + "content": " is similarly defined using the " + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "text", + "content": " distance, but between the pseudo-corrected image " + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "inline_equation", + "content": "I_{pseudo}" + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "text", + "content": ", generated by the slot reconstruction decoder, and the ground truth image " + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 55, + 597, + 295, + 655 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 121, + 657, + 294, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 657, + 294, + 670 + ], + "spans": [ + { + "bbox": [ + 121, + 657, + 294, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s l o t}} = \\left\\| I _ {\\text {p s e u d o}} - I _ {\\text {g t}} \\right\\| _ {1}. \\tag {13}", + "image_path": "28369823d0a46a770bfd946f9eec26a43e679f358520e10b8848ce5e5a09accb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "text", + "content": "The final objective function " + }, + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{final}" + }, + { + "bbox": [ + 55, + 672, + 295, + 695 + ], + "type": "text", + "content": " is defined as the sum of the two losses:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 701, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 117, + 701, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {e n h a n c e}} + \\mathcal {L} _ {\\text {s l o t}}. \\tag {14}", + "image_path": "8a886e3ea9b8f59d150ed60d953063d4c82946a8c71bea6caab62fef4ca9bdc7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 290, + 395, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 395, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 395, + 304 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 310, + 430, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 430, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 430, + 323 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": "Implementation Details. We trained our model using the Adam optimizer with " + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": " using a patch size of " + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": " and a batch size of 16. The learning rate was set to " + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": ", and training was conducted for 500 epochs. Additionally, the parameters " + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "inline_equation", + "content": "K^{main}" + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "inline_equation", + "content": "K^{sub}" + }, + { + "bbox": [ + 313, + 327, + 555, + 399 + ], + "type": "text", + "content": " were set to 3 and 7, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 414, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 485 + ], + "type": "text", + "content": "Datasets and Comparative methods. Our training and benchmarking settings align with established standards for existing exposure correction tasks [13, 14, 22, 35]. We train our network on three multi-exposure datasets: Single Image Contrast Enhancement (SICE) [5], Multi-Scale Exposure Correction (MSEC) [1], and LCDP [35]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": "We compare our Exposure-slot to existing state-of-the-art exposure correction methods, including LCPDNet [35], ERL [15], ENC [12], DA [39], ECLNet [14], FECNet [13], and CSEC [22]. In PromptIR [28], the number of prompts is set to match the number of exposure values in each dataset: 2 for SICE, 5 for MSEC, and 2 for LCDP. Evaluations are conducted using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 588, + 451, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 451, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 451, + 600 + ], + "type": "text", + "content": "4.2. Performance Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Table 1 presents the performance of our method on three representative multi-exposure datasets: SICE [5], MSEC [1], and LCDP [35]. On the SICE dataset, our approach achieves the highest performance overall, except for one SSIM value in the under-exposed condition, where it ranks second. Similarly, for the MSEC dataset, Exposure-slot consistently outperforms previous methods in PSNR and SSIM, achieving top scores in all cases except for under-exposure in SSIM. On the LCDP dataset," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17897" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 138, + 185 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 138, + 185 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 138, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 138, + 185 + ], + "type": "image", + "image_path": "ba670b43b7537856e6885fb394c266099a0fc24e94cb56f7d28f24868741e703.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 186, + 110, + 195 + ], + "lines": [ + { + "bbox": [ + 86, + 186, + 110, + 195 + ], + "spans": [ + { + "bbox": [ + 86, + 186, + 110, + 195 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 222, + 184 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 222, + 184 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 222, + 184 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 222, + 184 + ], + "type": "image", + "image_path": "3dad1efa62d23312b168e065c5935e798aa6239a9521753cf4875b3963cbf380.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 186, + 197, + 194 + ], + "lines": [ + { + "bbox": [ + 165, + 186, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 165, + 186, + 197, + 194 + ], + "type": "text", + "content": "FECNet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 306, + 184 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 306, + 184 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 306, + 184 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 306, + 184 + ], + "type": "image", + "image_path": "6d4ec2cb75a0e1e3f0b703d99ebb245de48981e003cec059ebb17fa22c4f6fab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 186, + 288, + 194 + ], + "lines": [ + { + "bbox": [ + 241, + 186, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 241, + 186, + 288, + 194 + ], + "type": "text", + "content": "DRBN-ENC" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 387, + 185 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 387, + 185 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 387, + 185 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 387, + 185 + ], + "type": "image", + "image_path": "245d1b303be3e1b67351a352f979c2c7a333dca2702abaa695e28d70906ef4bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 186, + 359, + 194 + ], + "lines": [ + { + "bbox": [ + 334, + 186, + 359, + 194 + ], + "spans": [ + { + "bbox": [ + 334, + 186, + 359, + 194 + ], + "type": "text", + "content": "CSEC" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 471, + 185 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 471, + 185 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 471, + 185 + ], + "type": "image", + "image_path": "ad4773be4e0498c0114c5c21680f50d3d7743114f415a62c1cc879781abe72f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 186, + 456, + 195 + ], + "lines": [ + { + "bbox": [ + 403, + 186, + 456, + 195 + ], + "spans": [ + { + "bbox": [ + 403, + 186, + 456, + 195 + ], + "type": "text", + "content": "Exposure-slot" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 472, + 70, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 472, + 70, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 472, + 70, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 472, + 70, + 553, + 185 + ], + "type": "image", + "image_path": "8e7b96fef407037412207c9d6a19689e03a7e4f4c4a17af58d7d676290527755.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 186, + 539, + 194 + ], + "lines": [ + { + "bbox": [ + 485, + 186, + 539, + 194 + ], + "spans": [ + { + "bbox": [ + 485, + 186, + 539, + 194 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 57, + 224, + 553, + 290 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 553, + 220 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison (FECNet [13], ENC [12], CSEC [22]) on the SICE [5] and MSEC [1] dataset. Examples of images enhanced from under-exposed condition (Top) and images enhanced from over-exposed condition (Bottom)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 224, + 553, + 290 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 553, + 290 + ], + "type": "image", + "image_path": "3766b1f22c0bcf28b2ed52149cf91edf6595048fe24c100e9dea9fd8cabb6105.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 291, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 553, + 314 + ], + "type": "text", + "content": "Figure 4. Visualization of slot attention maps (attn) of main- and sub-slot. It demonstrates how exposure-centric characteristics are effectively partitioned and refined through the slot attention mechanisms, transforming the input image into the corrected output." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 61, + 322, + 290, + 437 + ], + "blocks": [ + { + "bbox": [ + 61, + 322, + 290, + 437 + ], + "lines": [ + { + "bbox": [ + 61, + 322, + 290, + 437 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 290, + 437 + ], + "type": "table", + "html": "
ModelΔEB2000↓ [31]ΔEab↓ [30]
ZeroDCE [9]23.7829.31
RUAS [23]29.5937.05
SCI [26]25.6431.22
ECLNet [14]9.1511.58
FECNet [13]8.8511.19
DRBN-ENC [12]8.6811.02
PromptIR [28]9.2711.59
CSEC [22]8.7211.39
Exposure-slot6.948.89
", + "image_path": "f5c2d2f822966dadae820dfad26c2d2568237ec372fc9d76b95a9faae2c773c1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "type": "text", + "content": "Table 2. Comparisons with color difference metrics, " + }, + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\Delta {E}_{2000} \\downarrow" + }, + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\Delta {E}_{ab} \\downarrow" + }, + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "type": "text", + "content": " ,on the SICE [5] dataset. The highest score is high- lighted in red, while the second-highest is marked in blue." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "text", + "content": "our method also demonstrates superior performance, surpassing CSEC [22]. Compared to previous state-of-the-art methods [12, 13], our approach shows a notable gain of 1.85 dB in PSNR on the SICE dataset and 0.4 dB on the LCDP dataset compared to CSEC [22], highlighting its strong performance advantage." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Fig.3 shows qualitative comparisons on the SICE [5] and MSEC [1] datasets, demonstrating the performance of our model against other exposure correction methods. Under under-exposed conditions, Exposure-slot achieves superior color fidelity and detail recovery, closely matching the ground truth without excessive brightening or color distortion, unlike other models [13, 22]. In over-exposed areas, the zoomed-in views reveal that other models, particularly CSEC [22], produce artifacts and color inconsistencies due to over-correction. In contrast, Exposure-slot maintains stable color restoration and texture enhancement, preserving natural details and color harmony. These results highlight Exposure-slot's capability to handle diverse exposure con" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 340, + 322, + 529, + 413 + ], + "blocks": [ + { + "bbox": [ + 340, + 322, + 529, + 413 + ], + "lines": [ + { + "bbox": [ + 340, + 322, + 529, + 413 + ], + "spans": [ + { + "bbox": [ + 340, + 322, + 529, + 413 + ], + "type": "image", + "image_path": "835e3d8d76172bf4dbc130c189f03c7ad86b41e7ea8bd1efa402f09ed6d671c7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 414, + 553, + 448 + ], + "lines": [ + { + "bbox": [ + 313, + 414, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 553, + 448 + ], + "type": "text", + "content": "Figure 5. t-SNE results of features before and after prompts. Features from the same sub-slot are represented by the same color. With prompts, the features exhibit clearer separation." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 460, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 553, + 483 + ], + "type": "text", + "content": "ditions, achieving state-of-the-art performance in complex exposure correction tasks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": "Additionally, to evaluate color correction performance, we conduct a comparison using " + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\Delta E_{2000}" + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": " [31] and " + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\Delta E_{ab}" + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": " [30] metrics in the LAB color space. Table 2 presents the results, demonstrating that our approach also excels in color correction. Exposure-slot demonstrates a performance improvement of 1.78 in " + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\Delta E_{2000}" + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": " and 2.13 in " + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\Delta E_{ab}" + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": " compared to previous state-of-the-art methods." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Fig. 4 visualizes the attention maps for main and sub-slots predicted by SSAB. These results demonstrate that each map effectively generates prompts region maps, and Exposure-slot leverages this partitioning information to achieve robust enhancement across different exposure conditions of the same scene. In Fig. 5, the t-SNE visualization of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "V \\cdot P^{final}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " in Eq. 2, demonstrates that our prompts improve clustering and effectively facilitate feature separation. Notably, the improved feature identification within each slot suggests that the model effectively captures exposure-specific or region-based information. For further visualizations of attention maps and additional details on" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17898" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 134, + 114 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 134, + 114 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 134, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 134, + 114 + ], + "type": "image", + "image_path": "27f5f3de9fb3ab137342016cc8714df99f761f5a49f8390e570b5f29c933d8d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 115, + 106, + 123 + ], + "lines": [ + { + "bbox": [ + 85, + 115, + 106, + 123 + ], + "spans": [ + { + "bbox": [ + 85, + 115, + 106, + 123 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 70, + 214, + 114 + ], + "blocks": [ + { + "bbox": [ + 137, + 70, + 214, + 114 + ], + "lines": [ + { + "bbox": [ + 137, + 70, + 214, + 114 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 214, + 114 + ], + "type": "image", + "image_path": "33bbd8ada2de18220583b532fcfec2f05c70ebe8fa5d946b21c6c14be10d7784.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 115, + 197, + 123 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 197, + 123 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 197, + 123 + ], + "type": "text", + "content": "w/o Prompts" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 70, + 294, + 114 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 294, + 114 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 294, + 114 + ], + "type": "image", + "image_path": "b1f7df9028430fda4f3c9e2158169fffef59860a2618f6a5c830c87921f58dd5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 115, + 274, + 122 + ], + "lines": [ + { + "bbox": [ + 238, + 115, + 274, + 122 + ], + "spans": [ + { + "bbox": [ + 238, + 115, + 274, + 122 + ], + "type": "text", + "content": "w/o SSAB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 125, + 134, + 169 + ], + "blocks": [ + { + "bbox": [ + 56, + 125, + 134, + 169 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 134, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 134, + 169 + ], + "type": "image", + "image_path": "13bd8490eba7320186ea4b188ed2a37c62f3cb5ba7e3550c19dbc8659b8fb905.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 171, + 116, + 179 + ], + "lines": [ + { + "bbox": [ + 75, + 171, + 116, + 179 + ], + "spans": [ + { + "bbox": [ + 75, + 171, + 116, + 179 + ], + "type": "text", + "content": "w/o Decslot" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 125, + 214, + 169 + ], + "blocks": [ + { + "bbox": [ + 137, + 125, + 214, + 169 + ], + "lines": [ + { + "bbox": [ + 137, + 125, + 214, + 169 + ], + "spans": [ + { + "bbox": [ + 137, + 125, + 214, + 169 + ], + "type": "image", + "image_path": "74c9e209e38e658f31f26d51ba5eaff63c6b094a88b305e43623f51b6fa4642d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 171, + 200, + 179 + ], + "lines": [ + { + "bbox": [ + 151, + 171, + 200, + 179 + ], + "spans": [ + { + "bbox": [ + 151, + 171, + 200, + 179 + ], + "type": "text", + "content": "Exposure-slot" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 217, + 125, + 294, + 169 + ], + "blocks": [ + { + "bbox": [ + 217, + 125, + 294, + 169 + ], + "lines": [ + { + "bbox": [ + 217, + 125, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 217, + 125, + 294, + 169 + ], + "type": "image", + "image_path": "84153e20cf1e9f04031ee208cf16656b19f72a8e627c6b441fe702cc1a94887c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 245, + 170, + 268, + 178 + ], + "lines": [ + { + "bbox": [ + 245, + 170, + 268, + 178 + ], + "spans": [ + { + "bbox": [ + 245, + 170, + 268, + 178 + ], + "type": "text", + "content": "CSEC" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "type": "text", + "content": "Figure 6. The visualization of the ablation study. From the top left, the images correspond to (e), (f), (g), and (h) in the Table 3, with the previous SOTA method, CSEC [22], included for comparison." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 56, + 220, + 294, + 316 + ], + "blocks": [ + { + "bbox": [ + 56, + 220, + 294, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 316 + ], + "type": "table", + "html": "
ModelDecslotSSAB (2-level)PromptsPSNR↑/SSIM↑
(a)21.92/0.7091
(b)✓22.02/0.7154
(c)✓21.76/0.6998
(d)✓22.09/0.7062
(e)✓✓22.10/0.7065
(f)✓✓22.09/0.7210
(g)✓✓21.95/0.7149
(h)✓✓✓22.81/0.7239
", + "image_path": "8946ae756b543ab7d9f24141b5612f70bc65a8dd2386b5aadc3a21bd1e00550a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 333, + 258, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 258, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 258, + 344 + ], + "type": "text", + "content": "t-SNE, please refer to the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 352, + 149, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 149, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 149, + 364 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 369, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 405 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to assess the effectiveness of the Exposure-slot, as well as the model configuration and structural levels." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 409, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 505 + ], + "type": "text", + "content": "Effectiveness of the Proposed Modules. This section evaluates the effectiveness of SSAB, " + }, + { + "bbox": [ + 55, + 409, + 295, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{Dec}_{slot}" + }, + { + "bbox": [ + 55, + 409, + 295, + 505 + ], + "type": "text", + "content": ", and prompts through ablation studies. Table 3 shows results for each ablation case on the SICE [5] dataset. Specifically, configurations without SSAB (2-level) refer to an SSAB with a single level consisting of 7 slots (i.e. main-slots only). Configurations without Prompts directly use the slot feature " + }, + { + "bbox": [ + 55, + 409, + 295, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{final}" + }, + { + "bbox": [ + 55, + 409, + 295, + 505 + ], + "type": "text", + "content": " for cross-attention in Eq. 2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "content": "First, a comparison between cases with and without " + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{Dec}_{slot}" + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "content": " shows consistent performance improvements when " + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{Dec}_{slot}" + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "content": " is included. This result suggests that training " + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{Dec}_{slot}" + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "content": " helps the slots more accurately identify and represent regions with distinct exposure characteristics, leading to enhanced performance. When employing SSAB (Model (e) and (h)) also shows significant performance improvement when used together with " + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{Dec}_{slot}" + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "content": ". Regarding prompts, although Model (d) shows a slight decrease in SSIM compared to Model (a), all other cases utilizing prompts demonstrate improved performance. Notably, our full model (h) achieves the highest performance, underscoring the importance of each component in enhancing model performance. In Fig. 6, we provide a part of ablations as visual results within state-of-the-art method CSEC [22]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Investigation of Model Configuration. Table 4 presents ablation studies on the number of main-slots (" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "K^{main}" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "), sub-" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 163 + ], + "blocks": [ + { + "bbox": [ + 88, + 316, + 261, + 326 + ], + "lines": [ + { + "bbox": [ + 88, + 316, + 261, + 326 + ], + "spans": [ + { + "bbox": [ + 88, + 316, + 261, + 326 + ], + "type": "text", + "content": "Table 3. Ablation studies on proposed modules." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 554, + 163 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 163 + ], + "type": "table", + "html": "
ModelKmainKsubTPSNR↑SSIM↑Time (S)
(a)37322.810.72360.06759
(b)27322.610.72450.06759
(c)47322.360.72010.06759
(d)36321.920.70380.06759
(e)38322.780.72410.06759
(f)37221.150.70580.06747
(g)37421.870.72220.06996
", + "image_path": "ce61d466fefa68a4326083173f24c14e64620595a917b67ae9fcda2c373f0ef2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 315, + 194, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "type": "text", + "content": "Table 4. Investigation of the number of slots (" + }, + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "type": "inline_equation", + "content": "K^{main}" + }, + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "type": "inline_equation", + "content": "K^{sub}" + }, + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "type": "text", + "content": ") and number of iterations (" + }, + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 167, + 554, + 190 + ], + "type": "text", + "content": ") on the SICE [5] dataset." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 194, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 315, + 194, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 315, + 194, + 553, + 227 + ], + "type": "table", + "html": "
ModelKmainK*sub-1K*sub-2PSNR↑SSIM↑ΔE2000↓[31]ΔEab↓[30]
1-level3--22.020.71317.129.41
2-level37-22.810.72366.948.89
3-level371023.060.73066.848.69
", + "image_path": "3cac51099489f85d26b2135968740ab0fb25762ad384b8e8dc22e7c66dd9b994.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 331, + 228, + 537, + 238 + ], + "lines": [ + { + "bbox": [ + 331, + 228, + 537, + 238 + ], + "spans": [ + { + "bbox": [ + 331, + 228, + 537, + 238 + ], + "type": "text", + "content": "Table 5. Ablation studies on " + }, + { + "bbox": [ + 331, + 228, + 537, + 238 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 331, + 228, + 537, + 238 + ], + "type": "text", + "content": " -level SSAB " + }, + { + "bbox": [ + 331, + 228, + 537, + 238 + ], + "type": "inline_equation", + "content": "\\left( {n = 1,2,3}\\right)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": "slots " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "(K^{sub})" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": ", and iterations " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": " within SSAB to evaluate their impact on performance. The results include PSNR and SSIM metrics on the SICE [5] dataset, with runtime measurements taken on a single " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "844 \\times 1500" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": " RGB image using an NVIDIA RTX 4090 GPU. In this analysis, we vary " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "K^{main}" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "K^{sub}" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": " by incrementing or decrementing each parameter by 1. Based on the PSNR values, the configuration of " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "K^{main} = 3" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "K^{sub} = 7" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": " yields the highest performance and is therefore selected for our method." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 369, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 524 + ], + "type": "text", + "content": "Effectiveness of Structural Levels. In Sec.3, we use a 2-level SSAB structure as the default configuration. To further investigate the potential of SSAB, we perform experiments using both 1-level and 3-level SSAB structures on the SICE [5] dataset, as summarized in Table 5. In the 3-level configuration, we set " + }, + { + "bbox": [ + 313, + 369, + 555, + 524 + ], + "type": "inline_equation", + "content": "K^{sub-2} = 10" + }, + { + "bbox": [ + 313, + 369, + 555, + 524 + ], + "type": "text", + "content": " for the number of second sub-slots, resulting in PSNR and SSIM improvements of 0.25 and 0.07 compared to 2-level SSAB, respectively. These results indicate that adding more SSAB levels can enhance performance. While adding additional levels improves performance, it also leads to an exponential increase in time complexity. Therefore, considering this trade-off, we adopt the 2-level SSAB as the default structure." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we present Exposure-slot, a novel framework for exposure correction that integrates slot-in-slot attention with learnable prompts to achieve precise, exposure-centric feature learning. By hierarchically clustering and refining exposure regions, Exposure-slot facilitates accurate, region-aware adjustments and achieves state-of-the-art performance on multiple multi-exposure benchmarks. Our approach demonstrates significant improvements in both quantitative and qualitative metrics, highlighting the effectiveness of structured attention mechanisms for challenging exposure correction scenarios. Furthermore, this work advances exposure correction and paves the way for exploring slot-based architectures in other low-level vision tasks." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17899" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 145, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 145, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 145, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 258 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2023-00222776), and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06, Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Ministry of Science and ICT (No.RS-2020-II201373, Artificial Intelligence Graduate School Program [Hanyang University])." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 267, + 115, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 115, + 280 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 115, + 280 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 287, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 287, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 287, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 287, + 295, + 319 + ], + "type": "text", + "content": "[1] Mahmoud Afifi, Konstantinos G Derpanis, Bjorn Ommer, and Michael S Brown. Learning multi-scale photo exposure correction. In CVPR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 321, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 296, + 354 + ], + "type": "text", + "content": "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 355, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 355, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 355, + 295, + 399 + ], + "type": "text", + "content": "[3] Ondrej Biza, Sjoerd Van Steenkiste, Mehdi S. M. Sajjadi, Gamaleldin F. Elsayed, Aravindh Mahendran, and Thomas Kipf. Invariant slot attention: object discovery with slot-centric reference frames. In ICML, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 400, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 296, + 443 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 477 + ], + "type": "text", + "content": "[5] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27, 2018. 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 510 + ], + "type": "text", + "content": "[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 513, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 513, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 513, + 296, + 566 + ], + "type": "text", + "content": "[7] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 296, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 611 + ], + "type": "text", + "content": "[8] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In CVPR, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 656 + ], + "type": "text", + "content": "[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 58, + 658, + 295, + 690 + ], + "type": "text", + "content": "[10] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[11] Yun He, Steven Zheng, Yi Tay, Jai Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, YaGuang Li, Zhao Chen, Donald Met" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "zler, et al. Hyperprompt: Prompt-based task-conditioning of transformers. In ICML. PMLR, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "text", + "content": "[12] Jie Huang, Yajing Liu, Xueyang Fu, Man Zhou, Yang Wang, Feng Zhao, and Zhiwei Xiong. Exposure normalization and compensation for multiple-exposure correction. In CVPR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "text", + "content": "[13] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In ECCV, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 554, + 227 + ], + "type": "text", + "content": "[14] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In ACMMM, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 554, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 260 + ], + "type": "text", + "content": "[15] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In CVPR, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 554, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 554, + 294 + ], + "type": "text", + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 295, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 554, + 327 + ], + "type": "text", + "content": "[17] Donggoo Jung, Daehyun Kim, and Tae Hyun Kim. Continuous exposure learning for low-light image enhancement using Neural ODEs. In ICLR, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 328, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 554, + 360 + ], + "type": "text", + "content": "[18] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In CVPR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 403 + ], + "type": "text", + "content": "[19] Markus Krimmel, Jan Achterhold, and Joerg Stueckler. Attention normalization impacts cardinality generalization in slot attention. Transactions on Machine Learning Research, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 405, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 554, + 437 + ], + "type": "text", + "content": "[20] Minhyeok Lee, Suhwan Cho, Dogyoon Lee, Chaewon Park, Jungho Lee, and Sangyoun Lee. Guided slot attention for unsupervised video object segmentation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 439, + 554, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 439, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 439, + 554, + 481 + ], + "type": "text", + "content": "[21] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In CVPR, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 483, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 554, + 514 + ], + "type": "text", + "content": "[22] Yiyu Li, Ke Xu, Gerhard Petrus Hancke, and Rynson WH Lau. Color shift estimation-and-correction for image enhancement. In CVPR, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 516, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 516, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 516, + 554, + 557 + ], + "type": "text", + "content": "[23] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 612 + ], + "type": "text", + "content": "[24] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 614, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 646 + ], + "type": "text", + "content": "[25] Minh-Thang Luong. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 680 + ], + "type": "text", + "content": "[26] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "type": "text", + "content": "[27] Ntumba Elie Nsampi, Zhongyun Hu, and Qing Wang. Learning exposure correction via consistency modeling. In BMVC, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17900" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[28] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. PromptIR: Prompting for all-in-one image restoration. In NeurIPS, 2023. 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 140 + ], + "type": "text", + "content": "[29] Rishav Pramanik, José-Fabian Villa-Vásquez, and Marco Pedersoli. Masked Multi-Query slot attention for unsupervised object discovery. In IJCNN, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 142, + 294, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 142, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 57, + 142, + 294, + 164 + ], + "type": "text", + "content": "[30] Gaurav Sharma and Raja Bala. Digital color imaging handbook. CRC press, 2017. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 166, + 295, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 166, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 57, + 166, + 295, + 209 + ], + "type": "text", + "content": "[31] Gaurav Sharma, Wencheng Wu, and Edul N Dalal. The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application, 2005. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 211, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 211, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 211, + 294, + 266 + ], + "type": "text", + "content": "[32] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. CODA-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 267, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 267, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 267, + 294, + 300 + ], + "type": "text", + "content": "[33] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 302, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 302, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 57, + 302, + 294, + 346 + ], + "type": "text", + "content": "[34] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 347, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 347, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 57, + 347, + 294, + 380 + ], + "type": "text", + "content": "[35] Haoyuan Wang, Ke Xu, and Rynson WH Lau. Local color distributions prior for image enhancement. In ECCV. Springer, 2022. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 382, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 382, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 382, + 294, + 415 + ], + "type": "text", + "content": "[36] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 417, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 417, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 57, + 417, + 294, + 450 + ], + "type": "text", + "content": "[37] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 452, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 452, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 57, + 452, + 294, + 483 + ], + "type": "text", + "content": "[38] Yanbo Wang, Letao Liu, and Justin Dauwels. Slot-vae: Object-centric scene generation with slot attention. In ICML, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 486, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 486, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 57, + 486, + 294, + 518 + ], + "type": "text", + "content": "[39] Yang Wang, Long Peng, Liang Li, Yang Cao, and Zheng-Jun Zha. Decoupling-and-aggregating for image exposure correction. In CVPR, 2023. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 520, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 520, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 520, + 294, + 553 + ], + "type": "text", + "content": "[40] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep Retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 555, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 597 + ], + "type": "text", + "content": "[41] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. URetinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "type": "text", + "content": "[42] Mingde Yao, Jie Huang, Xin Jin, Ruikang Xu, Shenglong Zhou, Man Zhou, and Zhiwei Xiong. Generalized lightness adaptation with channel selective normalization. In CVPR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "text", + "content": "[43] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In ACMM, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[44] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129, 2021. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 129 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[46] Karel Zuiderveld. Contrast limited adaptive histogram equalization. Graphics gems, 1994. 6" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17901" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_content_list.json b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a639118326a7bd0953db735de08112396e9688b --- /dev/null +++ b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_content_list.json @@ -0,0 +1,1549 @@ +[ + { + "type": "text", + "text": "Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think", + "text_level": 1, + "bbox": [ + 161, + 130, + 834, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jie Tian $^{1\\dagger*}$ Xiaoye Qu $^{1\\dagger}$ Zhenyi Lu $^{1\\dagger}$ Wei Wei $^{1\\boxtimes}$ Sichen Liu $^{1}$ Yu Cheng $^{2}$ $^{1}$ School of Computer Science & Technology, Huazhong University of Science and Technology, \n $^{2}$ The Chinese University of Hong Kong", + "bbox": [ + 125, + 200, + 872, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 306, + 326, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-to-Video (I2V) generation aims to synthesize a video clip according to a given image and condition (e.g., text). The key challenge of this task lies in simultaneously generating natural motions while preserving the original appearance of the images. However, current I2V diffusion models (I2V-DMs) often produce videos with limited motion degrees or exhibit uncontrollable motion that conflicts with the textual condition. To address these limitations, we propose a novel Extrapolating and Decoupling framework, which introduces model merging techniques to the I2V domain for the first time. Specifically, our framework consists of three separate stages: (1) Starting with a base I2V-DM, we explicitly inject the textual condition into the temporal module using a lightweight, learnable adapter and fine-tune the integrated model to improve motion controllability. (2) We introduce a training-free extrapolation strategy to amplify the dynamic range of the motion, effectively reversing the fine-tuning process to enhance the motion degree significantly. (3) With the above two-stage models excelling in motion controllability and degree, we decouple the relevant parameters associated with each type of motion ability and inject them into the base I2V-DM. Since the I2V-DM handles different levels of motion controllability and dynamics at various denoising time steps, we adjust the motion-aware parameters accordingly over time. Extensive qualitative and quantitative experiments have been conducted to demonstrate the superiority of our framework over existing methods. Code is available at https://github.com/Chuge0335/EDG", + "bbox": [ + 89, + 339, + 483, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Image-to-Video generation (I2V) [11, 23, 40, 45, 46] has attracted increasing attention. Given a single im", + "bbox": [ + 89, + 833, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0e69f353bba46926400601b34edbf1fed709f97bd8d08924a1a3b7bcdfa5426.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 299, + 885, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/45439e2b85fa9d023fe70d258dd5ed8df22883f215f6341db0c20d45c41c8b38.jpg", + "image_caption": [ + "Figure 1. In the first example, Consist2V [27] yields a video with limited motion. In the second, DynamiCrafter [40] fails to follow the text condition and lacks controllability. Our method, however, achieves improved motion degree and better controllability." + ], + "image_footnote": [], + "bbox": [ + 537, + 449, + 885, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age and a text condition, it aims to convert the still image into videos that conform to text instruction while preserving the original appearance of the images. This paper focuses on open-domain images and supports free text as guidance, rather than limited class labels. This conditioned I2V task is substantially challenging as it struggles with simultaneously maintaining appearance consistency and producing natural motions in the generated frames.", + "bbox": [ + 511, + 670, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike image synthesis, motion plays a significantly crucial role in image-to-video generation. However, most existing works lack effective motion modeling when generating videos. For instance, ConsistI2V [27] enhances the first frame conditioning through spatiotemporal attention, leading to limited expression ability for modeling motion degree as the model tends to over-rely on the conditional im", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution; * Work done during an internship at Shanghai AI Laboratory; $\\boxtimes$ corresponding authors.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12512", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age. DynamiCrafter [40] interacts the text embedding and conditional image embedding with the U-Net spatial layers through a dual cross-attention, neglecting the temporal property of the motion and suffering from motion controllability. In this way, the generated videos either contradict the motion described in the text or are still image sequences with slow change. As shown in Figure 1, the videos generated by the previous model show cars that barely move, or the videos of the castle contradict the textual descriptions.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we present a novel Extrapolating and Decoupling framework to effectively model motion for image-to-video generation, improving both motion degree and motion controllability. Specifically, it consists of three stages. Firstly, we improve motion controllability by explicitly injecting the textual condition into the I2V diffusion model (I2V-DM). A lightweight learnable adapter is introduced to enhance the temporal attention module. Fine-tuning the integrated model improves motion controllability without altering the original I2V-DM architecture. Secondly, we observe degraded motion dynamics in our fine-tuned model, a phenomenon has also been observed by Zhao et al. [46]. To boost the motion degree without extra training, we introduce a novel training-free extrapolation strategy to restore and enhance motion degree dynamics by reversing the fine-tuning process. Thirdly, we decouple the parameters contributing to motion controllability and dynamics from the previous two stages, selectively integrating them into the final model. Recognizing that different denoising steps focus on varying motion ability and temporal consistency, we dynamically adjust the motion-aware parameters over time. With these three stages, we can generate videos with both natural motion and temporal consistency.", + "bbox": [ + 91, + 229, + 482, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To verify the effectiveness of our framework, we perform experiments on the VBench I2V benchmark and UCF101 dataset. Extensive qualitative and quantitative experiments demonstrate the notable superiority of our framework over existing I2V methods. Furthermore, we offer discussion and analysis of some insightful designs in our framework. To sum up, our contributions are summarized as follows:", + "bbox": [ + 89, + 592, + 480, + 699 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Extrapolating and Decoupling framework for video generation, which models motion controllability and degree dynamics. To best of our knowledge, this is the first use of model merging techniques in the I2V domain.", + "- We introduce a lightweight adapter to enhance motion controllability while preserving the base I2V-DM structure and propose a novel training-free extrapolation strategy to amplify motion degree.", + "- We make the first attempt to decouple parameters related to motion controllability and degree, dynamically injecting them into the I2V-DM across varying denoising steps.", + "- We conduct extensive experiments on multiple bench" + ], + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "marks, where our proposed framework consistently outperforms previous state-of-the-art methods.", + "bbox": [ + 526, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 150, + 653, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Motion Modeling in I2V Generation", + "text_level": 1, + "bbox": [ + 513, + 180, + 826, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models (DMs) [15, 33] have recently demonstrated remarkable generative capabilities in image and video generation [12, 24, 26, 28, 29]. In this paper, we focus on the image-conditioned video (I2V) generation task, where motion modeling aims to predict object movements. Traditionally, optical flow estimates the displacement field between consecutive frames. To effectively model motion in I2V tasks, LFDM [23] proposes a latent flow diffusion model that synthesizes a latent optical flow sequence to warp the given image in latent space for generating new videos. Motion-I2V [31] factorizes I2V into two stages: a diffusion-based motion field predictor for pixel trajectories and motion-augmented temporal attention to enhance 1-D temporal attention in video latent diffusion models. Recent works [38, 40, 44] enable diffusion models to handle motion modeling and video generation simultaneously. For instance, Dynamiccafter [40] uses a dual-stream image injection mechanism to leverage the motion prior of text-to-video diffusion models. In this paper, we build on DynamamiCrafter to design an Extrapolating and Decoupling framework to improve both motion controllability and degree.", + "bbox": [ + 511, + 208, + 906, + 525 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Model Merging", + "text_level": 1, + "bbox": [ + 513, + 550, + 669, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model merging, widely explored in natural language processing and image classification, typically combines multiple fine-tuned task-specific models into a unified multitask model without additional training. Early methods [1, 8-10, 17, 22] leverage interpolation techniques to integrate multi-task knowledge effectively. Task-Arithmetic [16] extends these approaches by enabling complex arithmetic operations in weight space, offering finer control over the merging process. Recent studies [21, 25, 41-43] have further reduced interference during merging. Beyond interpolation-based methods, recent works have shown that sparsifying [43] or amplifying delta weights [47] can enhance model knowledge in merging. Additionally, merging techniques have been introduced to multimodal understanding such as question-answering [6, 32]. However, the video generation domain has remained unexplored until now.", + "bbox": [ + 511, + 578, + 906, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We pioneer the introduction of model merging techniques into the video generation domain. Specifically, we leverage model merging to extrapolate motion degree knowledge and extract fine-grained decoupled knowledge, which are selectively merged at different diffusion stages.", + "bbox": [ + 511, + 825, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b10612b13869a7de2d99ae36aa38144f6812edc4bbb5e7186a9d9cd542b3aac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 88, + 495, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa1dd35b126828d0a192f07f99ee42b3f22f1e9fbde6ed7c55e873502a93321d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 88, + 883, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5b6642b6e4e59b307915a7e2d7dcbfbcbd330e1a77e6a7de9d29f44ad5e0eb3.jpg", + "image_caption": [ + "Figure 2. The overview of our framework, consisting of three stages. (a) We improve the motion controllability by injecting the textual conditions into the temporal attention module in I2V-DM model and then fine-tune the integrated model. (b) The I2V-DM model suffers from a decrease in motion degree after fine-tuning. To mitigate it, we propose a training-free extrapolation to boost the motion degree by reversing the fine-tuning progress. (c) We decouple the relevant parameters contributing to motion controllability and motion degree from the last two-stage models. Moreover, we selectively inject these parameters into the I2V model along with the de-noising process." + ], + "image_footnote": [], + "bbox": [ + 109, + 229, + 764, + 381 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4cb4cb9d813853827865359c69548fb1b1782960468e3f12906af4db2c7b7d84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 234, + 883, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 477, + 181, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a reference image $I_0$ and a text prompt $c$ , image-to-video generation (I2V) targets at generating a sequence of subsequent video frames $V = \\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_N\\}$ . The key objective of I2V generation is to ensure that the generated video not only exhibits plausible motion but also faithfully preserves the visual appearance of the provided reference image. The overall framework of our method, shown in Figure 2, consists of three stages. First, we improve motion controllability guided by the text condition. Second, we propose a training-free extrapolation strategy to enhance motion dynamics. Finally, we decouple parameters related to motion controllability and dynamics, allowing flexible injection during the denoising process to generate videos with natural motion and temporal consistency.", + "bbox": [ + 89, + 503, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Latent Video Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 726, + 467, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before introducing our framework, we first review the Latent Diffusion Models (LDMs) [28]. Specifically, it extends diffusion models [15] by operating in a more compact latent space rather than directly on pixels, reducing computational complexity. In LDMs, an input $x_0$ is first encoded into a latent representation $z_0$ using a pretrained Variational Autoencoder (VAE). This latent variable $z_0$ is then corrupted gradually over $T$ steps by adding Gaussian noise, producing a sequence of noisy latents $\\{z_t\\}_{t=1}^T$ . The noise process follows $z_t \\mid z_0 \\sim \\mathcal{N}(z_t; \\sqrt{\\bar{\\alpha}_t} z_0, (1 - \\bar{\\alpha}_t)I)$ , where", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}$ controls the noise schedule. To recover the original latent, LDMs learn a reverse process that denoises $z_{t}$ back to $z_{0}$ through a sequence of steps, guided by the objective:", + "bbox": [ + 511, + 478, + 906, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {z _ {0}, t, \\epsilon} \\| \\epsilon_ {\\theta} (z _ {t}, t) - \\epsilon \\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 539, + 903, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0, I)$ represents Gaussian noise, $t$ is a randomly sampled timestep, and $\\epsilon_{\\theta}(z_t, t)$ is the learned noise estimator parameterized by $\\theta$ .", + "bbox": [ + 511, + 564, + 906, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we adopt an open-source video LDM, DynamiCrafter [40] as the base model. Given a video $x \\in \\mathbb{R}^{L \\times 3 \\times H \\times W}$ , we first encode it into a latent representation $z = \\mathcal{E}(x)$ , where $z \\in \\mathbb{R}^{L \\times C \\times h \\times w}$ is derived on a frame-by-frame basis. Both the forward diffusion process $z_{t} = p(z_{0}, t)$ and the backward denoising process $z_{t} = p_{\\theta}(z_{t-1}, c, t)$ are conducted in this latent space, with $c$ representing potential denoising conditions such as text prompts. Ultimately, the generated videos are obtained through the decoder $\\hat{x} = \\mathcal{D}(z)$ .", + "bbox": [ + 511, + 611, + 908, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Motion-Controllability Adaptation", + "text_level": 1, + "bbox": [ + 511, + 772, + 816, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In I2V generation, motion controllability is important for following textual conditions. To model motion, current I2VDMs [40] typically employ spatial attention to share information within a single frame and temporal attention to capture time sequence consistency [13]. However, methods like DynamiCrafter [40] only use text condition embedding in the spatial attention module, thus it does not share condi", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12514", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tional information across different time sequences, limiting its motion controllability.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate this problem, we explicitly inject textual conditions into DynamiCrafter's temporal attention with a lightweight adapter, as depicted in Figure 2(a). The spatial attention module remains unmodified as it already incorporates textual information. The adapter is non-intrusive and preserves the original model structure. Specifically, we extracts spatial features $z_{t} \\in \\mathbb{R}^{f \\times d \\times w \\times h}$ and the text embeddings $t_{\\mathrm{embd}} \\in \\mathbb{R}^{l \\times d}$ from the spatial attention layer and the CLIP text encoder respectively, where $f$ is the number of frames, $d$ is the feature dimension, and $w, h$ are spatial dimensions. Then, the condition information in text embeddings are further compressed by a trainable Q-Former [19] by retaining the motion-relevant tokens:", + "bbox": [ + 89, + 121, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {\\text {e m b d}} ^ {\\prime} = \\mathrm {Q} - \\operatorname {F o r m e r} \\left(t _ {\\text {e m b d}}\\right) \\in \\mathbb {R} ^ {f \\times d} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 328, + 482, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$t_{\\mathrm{embd}}^{\\prime}$ is then concatenated with spatial information $z_{t}$ along the feature dimension, projected through $W_{\\mathrm{proj}}$ to reduce the dimension to $d$ , and passed through temporal attention to inject the text control information into each frame. We finetune the model on WebVid-2M [2], updating only the adapter (the Q-Former and the linear projection $W_{\\mathrm{proj}}$ ) and temporal attention weights, ensuring minimal added parameters and efficient training.", + "bbox": [ + 89, + 358, + 483, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Motion-Degree Extrapolation", + "text_level": 1, + "bbox": [ + 89, + 489, + 354, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After fine-tuning I2V-DM DynamiCrafter with motion-controllability adaptation, we observe a reduction in motion degree, a common issue attributed to conditional image leakage [46]. During fine-tuning, Dynamicrafter concatenates the first frame with noise, which may cause the model to over-rely on the conditional image and neglect noisy inputs, especially at early timesteps. While adding time-dependent noise partially mitigates motion loss [46], it requires retraining and has limited effectiveness.", + "bbox": [ + 89, + 511, + 483, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the motion dynamics beyond the pre-trained model, we introduce a novel training-free and model-agnostic extrapolation approach, as shown in Figure 2 (b), which extremely pushes the \"unlearning\" process of motion controllability. In this way, we can achieve a model with amplified motion degree dynamics. Specifically, we estimate a degree-enhanced model, $\\theta_{dyn}$ , by extrapolating between a pre-trained model $\\theta_{pre}$ and its fine-tuned version $\\theta_{sft}$ with motion-controllability adaptation, amplifying their motion-dynamic differences:", + "bbox": [ + 89, + 648, + 483, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {d y n} = \\theta_ {p r e} + \\alpha \\left(\\theta_ {p r e} - \\theta_ {s f t}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 811, + 482, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where the hyperparameter $\\alpha$ controls the strength of this extrapolation. In this equation, we compute the difference of the model parameters except for the adapter introduced in Section 3.2. This process requires only two checkpoints and", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "obtaining the motion degree-enhanced model without training. To help further understand our extrapolating method, we provide a theoretical analysis below.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical Analysis. Assume that we have a motion degree metric score denoted as $\\mathcal{D}$ . Diffusion models are commonly regularized by a KL constraint in Gaussian distribution, implying that $\\theta_{pre}$ lies within a small vicinity of $\\theta_{sft}$ . Therefore, a first-order Taylor expansion can be applied:", + "bbox": [ + 511, + 154, + 906, + 231 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) = \\mathcal {D} \\left(\\theta_ {\\text {p r e}} + \\alpha \\Delta \\theta\\right) \\tag {4} \\\\ \\approx \\mathcal {D} (\\theta_ {p r e}) + \\alpha \\langle \\nabla \\mathcal {D} (\\theta_ {p r e}), \\Delta \\theta \\rangle \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 241, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta$ denotes the parameter update from $\\theta_{sft}$ to $\\theta_{pre}$ . The change in the motion degree is then given by:", + "bbox": [ + 511, + 282, + 905, + 313 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) - \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) = \\alpha \\langle \\nabla \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right), \\Delta \\theta \\rangle \\tag {5} \\\\ \\approx \\alpha \\langle \\nabla \\mathcal {D} (\\theta_ {p r e}), \\nabla \\mathcal {D} (\\theta_ {p r e}) \\rangle \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 323, + 903, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming that the parameter update $\\Delta \\theta$ during finetuning is approximately aligned with the gradient of $\\mathcal{D}$ [14, 16], we can express $\\Delta \\theta$ as:", + "bbox": [ + 511, + 369, + 906, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta \\approx \\gamma \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 429, + 903, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus by substitution, we have:", + "bbox": [ + 513, + 452, + 718, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) - \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) \\approx \\alpha \\left\\langle \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right), \\gamma \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) \\right\\rangle \\\\ = \\alpha \\gamma \\left\\langle \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right), \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right) \\right\\rangle \\tag {7} \\\\ = \\alpha \\gamma \\| \\nabla_ {\\theta} \\mathcal {D} (\\theta_ {p r e}) \\| ^ {2} \\geq 0. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 477, + 903, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the norm of the gradient $\\| \\nabla_{\\theta}\\mathcal{D}(\\theta_{pre})\\| ^2$ is always non-negative, the dynamic motion degree metric $\\mathcal{D}$ shall increases after the extrapolation, provided that the pretrained weight $\\theta_{pre}$ is not at a local maximum of $\\mathcal{D}$ .", + "bbox": [ + 511, + 547, + 905, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Decoupled Injection and Sampling", + "text_level": 1, + "bbox": [ + 511, + 618, + 815, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While motion-controllability adaptation and training-free degree extrapolation can enhance motion control and diversity respectively, directly combining them poses a significant challenge due to their opposing optimization directions: fine-tuning for control versus reversed extrapolation for degree. To balance motion controllability and degree without compromising image consistency, we propose a decoupling strategy that separates degree-related and control-related parameters, applying them at distinct denoising stages.", + "bbox": [ + 511, + 641, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Isolating Parameter Sets. Here, we categorize the obtained knowledge into three parameter sets that support core abilities: motion control ( $\\theta_{adt}$ ), motion degree ( $\\theta_{deg}$ ), and video consistency ( $\\theta_{con}$ ). We aim to integrate degree and control knowledge in the early denoising stage for long-term temporal planning, while emphasize consistency and", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12515", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "control later for refining subject appearance. However, they are embedded within $\\theta_{pre}$ and $\\theta_{sft}$ . Therefore, we need parameter isolation to decouple three types of knowledge.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We isolate these sets based on the following insight: (1) Based on the implicit gradient assumption (Section 3.3), $\\theta_{sft} - \\theta_{pre}$ captures the implicit control knowledge learned during the adaptation process, serves as $\\theta_{adt}$ paired with $(\\theta_{adapter})$ , incorporating elements like Q-Formers and projectors. (2) $\\theta_{sft}$ exhibits degraded degree knowledge (11.67 vs. 68.54 for $\\theta_{pre}$ in Table 1) and primarily contains control and consistency knowledge. Therefore, consistency knowledge $\\theta_{con}$ can be directly derived as $\\theta_{sft} - \\theta_{adt}$ . (3) The extrapolated $\\theta_{dyn}$ has similar control/consistency proportions as $\\theta_{pre}$ , therefore $\\theta_{deg} = \\theta_{dyn} - \\theta_{pre}$ can remove control/consistency components, isolating degree knowledge, even though it originates from $\\theta_{sft}$ . (4) Fine-tuned parameters often contain redundant elements, which can cause conflicts during merging [21, 43]. To minimize overlap between different knowledge while preserving performance, we apply DARE pruning [43] at a pruning rate $p$ (e.g., $p = 70\\%$ ). Distinct parameter groups are created using random masks $\\mathcal{M}$ amplified from a Bernoulli distribution:", + "bbox": [ + 89, + 136, + 483, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {1} \\sim \\operatorname {B e r n o u l l i} (p), \\quad \\mathcal {M} _ {2} \\sim \\operatorname {B e r n o u l l i} (p),\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 434, + 421, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {a d t} = \\left\\{\\theta_ {a d p a t e r} \\right\\} \\cup \\left\\{\\mathrm {D A R E} _ {\\mathcal {M} _ {2}} \\left(\\theta_ {s f t} - \\theta_ {p r e}\\right) \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 453, + 480, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {d e g} = \\mathrm {D A R E} _ {\\mathcal {M} _ {1}} \\left(\\theta_ {d y n} - \\theta_ {p r e}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 473, + 344, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {c o n} = \\theta_ {s f t} - \\theta_ {a d t},\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 492, + 259, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, Task-Arithmetic [16] is employed to inject control parameters $\\theta_{adt}$ into both $\\theta_{deg}$ and $\\theta_{con}$ , deriving two enhanced models: the degree-enhanced model $\\theta_{dyn}^{*}$ and the consistency-enhanced model $\\theta_{con}^{*}$ .", + "bbox": [ + 89, + 518, + 483, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {d y n} ^ {*} = \\text {T a s k - A r i t h m e t i c} (\\theta_ {d e g}, \\theta_ {a d t}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 590, + 480, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {c o n} ^ {*} = \\text {T a k - A r i t h m e t i c} \\left(\\theta_ {c o n}, \\theta_ {a d t}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 611, + 406, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoupled Video Sampling. During the denoising process for video generation, we typically observe that in the early stages, I2V models prioritize long-term temporal planning, including the motion of both subjects and textures, while later stages often focus more on refining subject appearance, which is mainly related to content consistency and overall image quality [37, 39]. Therefore, in the first $K$ steps of the denoising process, we adjust the video generator to make it more dynamic-enhanced and control-enhanced to increase motion controllability. In the later stages of the denoising process, we re-weight the motion-related modules, allowing the model to focus more on restoring the specific details of each frame of the subject, thereby generating high-consistency videos. Specifically, we define a generated model $\\theta_{gen}$ by blending the degree-enhanced model $\\theta_{dyn}^{*}$ and the consistency-merged model $\\theta_{con}^{*}$ based on the denoising step $t$ from $T$ to 0:", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {g e n} = \\alpha_ {t} \\theta_ {d y n} ^ {*} + (1 - \\alpha_ {t}) \\theta_ {c o n} ^ {*},\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 103, + 815, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {t} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} t > T - K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 123, + 903, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $T$ is the total number of denoising steps, and $K$ is a threshold that determines the switching point.", + "bbox": [ + 511, + 170, + 906, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 215, + 637, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 513, + 239, + 594, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In all experiments, videos and images are resized to a $320 \\times 512$ center-cropped format. During model training, we set the learning rate to $1 \\times 10^{-5}$ , a batch size of 64, and trained for 100K steps on 8 A100 GPUs. For inference, we use DDIM [34] with 50 steps and a classifier-free guidance scale of 7.5, generating 16 frames for a 2-second video.", + "bbox": [ + 511, + 262, + 905, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use the WebVid-2M dataset [2] for supervised fine-tuning, which contains two million video-text pairs from stock footage websites. For motion controllability annotation, we use CoTracker [18] to annotate motion information on the WebVid-2M dataset.", + "bbox": [ + 511, + 354, + 905, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. The evaluation benchmarks include VBench [27] and UCF101 [35]. VBench is a comprehensive benchmark for video generation, incorporating recent models and extensive evaluation metrics. UCF101 contains 101 human action categories with diverse background and lighting variations. For VBench, following previous work [20, 46], we utilize the official dataset to assess the model across several metrics: I2V-Subject Consistency (I2V Sub.), I2V-Background Consistency (I2V Back.), video quality, camera motion, and dynamic degree. The video quality evaluation considers dimensions including Subject Consistency, Background Consistency, Smoothness, Aesthetic Score, Imaging Quality, and Temporal Flickering. The Avg. score is calculated as $\\frac{1}{10} * (\\text{I2V Sub.} + \\text{I2V Back.} + 6 * \\text{Video Quality} + \\text{Motion Degree} + \\text{Motion Control})$ , which is consistent with VBench I2V original setting. For UCF101, similar to [39, 40], we use Fréchet Video Distance (FVD) [36] and Inception Score (IS) [30] to assess video quality.", + "bbox": [ + 511, + 429, + 906, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 513, + 712, + 735, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VBenchmark Results Analysis. We compare our method with two types of approaches on VBenchmark. The first kind of comparing methods include those requiring large-scale video pretraining, such as animate-Anything [7], SVD [3], and DynamiCrafter [40]. The other comparison methods are finetuning/inference strategies based on DynamiCrafter, including naive finetuning with the same training setup as our method (Naive FT) and Conditional Image Leakage (CIL) methods [46] (CIL Infer and CIL FT). Notably, our method performs best in Motion Degree and Motion Control across all baselines, even including large-scale pretrained methods,", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12516", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c29093831d82561806031391d99ba4a074f2bbefd7f6084885a55729afb3385e.jpg", + "table_caption": [ + "Table 1. Quantitative results on the VBench I2V benchmark for various I2V-DM pretrained baselines and finetuning/inference techniques, including our methods based on DynamiCrafter. \"PT.\" in the table denotes whether the method requires video pretraining, and \"FT.\" denotes whether the method requires supervised finetuning on WebVid-2M. The global best and DynamiCrafter-based best performing metrics are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
ModelAvg.PT.FT.I2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
ConsistI2V (Ren et al.)75.88✓✗95.8295.9585.7418.6233.92
SEINE (Chen et al.)76.56✓✗96.5796.8085.7134.3123.67
Animate-Anything (Dai et al.)74.61✓✗98.7698.5888.842.6813.08
VideoCrafter (Chen et al.)76.96✓✗91.1796.9487.5522.6033.60
SVD-XT-1.0 (Blattmann et al.)-✓✗97.5297.6386.5452.36-
SVD-XT-1.1 (Blattmann et al.)-✓✗97.5197.6286.6643.17-
I2VGen-XL (Zhang et al.)76.00✓✗96.4896.8387.0226.1018.48
DynamiCrafter (Xing et al.)79.93✓✗94.9195.3084.9568.5430.85
DynamiCrafter-Naive FT72.91✗✓97.9298.7083.5211.6717.82
DynamiCrafter-CIL Infer (Zhao et al.)80.47✗✗96.9693.3485.2869.4331.29
DynamiCrafter-CIL FT (Zhao et al.)77.55✗✓96.9097.7987.1025.6930.50
DynamiCrafter-Ours84.25✗✓97.0396.6286.2287.6443.87
DynamiCrafter-Ours w/ CIL Infer84.81✗✓97.1195.4587.1089.2143.61
", + "bbox": [ + 91, + 157, + 903, + 363 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aeee6cc04f222681c693565a64bd87e40510ec19c96359b39f65628c10f2fa8f.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of the DynamiCrafter-based method on the UCF-101 benchmark." + ], + "table_footnote": [], + "table_body": "
ModelFVD↓IS↑Motion Degree↑
DynamiCrafter288.0113.3168.17
DynamiCrafter-Naive FT373.7713.1947.66
DynamiCrafter-CIL FT409.6913.6853.78
DynamiCrafter-Ours324.7913.7177.93
", + "bbox": [ + 99, + 425, + 470, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as shown in Table 1. We observe that Naive FT causes a significant degradation in Motion Degree (11.67 vs 68.54), consistent with previous findings [46]. Naive FT achieves the highest I2V Subject and Background Consistency (I2V Sub. and I2V Back.) among the DynamiCrafter-based methods because lower motion degree results in higher consistency scores. Interestingly, motion control is also negatively affected in Naive FT (17.82 vs 11.54). We speculate that this is due to the insufficient motion degree, which hinders the model's ability to follow control instructions effectively. The CIL method can mitigate these effects, either by employing the inference strategy on the pre-trained model (CIL Infer) or by repairing the Naive FT (CIL FT). However, its enhancements in Motion Control and Motion Degree are limited. In contrast, our method, combining adaptation and extrapolation, significantly improves both motion degree and motion control (11.67 vs 87.64) while preserving similar video quality (87.10). Furthermore, our method is orthogonal to the CIL Infer method, and integrating it with our method further boosts performance (Avg. 84.25 vs 84.81).", + "bbox": [ + 88, + 534, + 482, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UCF101 Results Analysis. In the UCF101 dataset, our method still exhibits superior motion degree ability despite", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ed9ba45f26921b99ce254f23e4ff288dfa41ad75080ac5b285a73d28723d812.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 388, + 707, + 491 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3f40f575d7aa348e86614ede1c24599cbc04d2273e70f480d76c5ef2f277c940.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 388, + 900, + 491 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4d002a3e35f155dc61373a026e10214b25d26439c895465d0c823c2afe78ec0.jpg", + "image_caption": [ + "Figure 3. Human Evaluation of our method Compared to SVD, VideoCrafter (VC), DynamiCrafter (DC), and its variant of Native Fine-Tuned (DC-FT) and CIL-Based Methods." + ], + "image_footnote": [], + "bbox": [ + 519, + 491, + 709, + 585 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ea3af99211531cb8a6bce2d1bae0298866fcd76b618e0e58c85ae32e0145e95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 491, + 900, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the low resolution and poor image quality. Both Naive FT and CIL FT suffer from degradation in Fréchet Video Distance (FVD), as shown in Table 2. Unlike the performance on VBenchmark, CIL performs worse than Naive FT on UCF101, likely due to the added noise schedule disrupting the already noisy and low-resolution images. Our method, however, is more stable in FVD and outperforms in Inception Score (IS) and Motion Degree.", + "bbox": [ + 511, + 676, + 906, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study. As shown in Figure 3, our method excels in Motion Degree and Motion Control while remaining competitive in video quality and I2V consistency. The results are consistent with the above findings, supporting that our strategies enhance both video motion dynamics and control", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12517", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70134cf65ce9b4e42969d2f7e8755b2e5dbb3e5ab5e5f336e9dded2823ba7d88.jpg", + "image_caption": [ + "Figure 4. In the visual comparisons, we generated 16-frame videos using DynamiCrafter, CIL, Consistent-I2V, and our model. We selected 3 to 4 frames at regular intervals for display at the top. The upper section visualizes the overall aesthetic quality, while the lower section uses Optical Flow to illustrate the dynamic qualities of the generated videos for each method." + ], + "image_footnote": [], + "bbox": [ + 143, + 85, + 852, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while preserving quality and consistency, yielding superior overall performance.", + "bbox": [ + 89, + 684, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 726, + 302, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To empirically evaluate the visual quality of our method, we compare it with several state-of-the-art image-to-video generation approaches, including DynamiCrafter [40], CIL [46], and ConsistI2V [27]. In Figure 4, we observe that our method generates temporally coherent videos with a sufficient range of motion that closely adhere to the input text conditions. In contrast, DynamiCrafter exhibits inconsistencies in text control, likely due to challenges in capturing the semantic content of the input image. The ConsistI2V method produces videos that remain static over time, with", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the color points on the figures almost unchanged, indicating a lack of motion dynamics and control. While CIL generates videos that semantically resemble the input images and have enhanced motion compared to ConsistI2V, its improvements are limited. CIL does not sufficiently vary the content dynamically to match the text prompts, resulting in limited motion and control. Our method, however, preserves image consistency while achieving a large motion degree and exhibiting good controllability.", + "bbox": [ + 511, + 684, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 832, + 666, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the importance of our extrapolation and decoupling methods in Table 3: (1) On top of our adaptation training, using our extrapolation and decoupling techniques", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12518", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/500c845d1cccb95aa48819a6580b918f8a077a560669423dce1a486486811c35.jpg", + "table_caption": [ + "Table 3. Ablation study of the three stages of our framework on VBenchmark." + ], + "table_footnote": [], + "table_body": "
(a) Adaptation(b)Extrapolation(c) DecoupleI2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
✓XX97.5196.3986.6122.8549.31
X✓X94.1495.6283.5298.2117.63
XX✓94.6096.2183.9569.6132.37
✓✓X91.1192.8684.3271.6826.71
X✓✓96.0194.6084.3493.3321.85
✓✓✓97.0396.6286.2287.6443.87
", + "bbox": [ + 122, + 114, + 867, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "significantly improves the motion degree, from 22.85 to 87.64. (2) Using extrapolation alone can drastically improve the motion degree (69.61 vs 93.33), but it comes at the cost of degradation in I2V-Subject Consistency (I2V Sub.) and I2V-Background Consistency (I2V Back.), as well as motion control. This is because reversing the training process of the pre-trained model to the SFT model unlearns the camera motion ability, causing significant interference if directly merged [16]. (3) Using the decoupling technique alone can improve the motion degree while preserving other performance metrics. This indicates that the I2V model indeed exhibits functional differentiation during the sampling process. (4) Adding adaptation, though it increases training costs, greatly enhances the Motion Control aspect of the model. The degradation in motion degree indicates that direct supervision is suboptimal and can lead to minor motion dynamics, a finding consistent with previous research [46]. (5) When paired with the decoupling technique, we achieve optimal results, integrating high motion degree and control ability while preserving consistency and video quality.", + "bbox": [ + 89, + 241, + 483, + 544 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8487239c8707ead2a78c9aa4c4b52f7f9bc17c48f0322fb265445bf9289bceff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 558, + 285, + 647 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7faf79a98499b80709c802e758f7d607147326f4cef447eff43e28345feebab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 559, + 460, + 647 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/325f4dc6793b1cab5af21ad17c16945fe8ea437377c277c73802937aa093950c.jpg", + "image_caption": [ + "(a) Effect of extrapolate strength $\\alpha$" + ], + "image_footnote": [], + "bbox": [ + 107, + 651, + 282, + 739 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/492b8ccea4fb5c586769897af4ed4a12542d7dca40672b3c8b585e66d86a7a42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 651, + 460, + 739 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5c7037dfa3ced0fdb6a8d959d5372e7a35d4e0435b4499d7cd63ea6dc51f1ea.jpg", + "image_caption": [ + "(b) Effect of different switching time and model choices." + ], + "image_footnote": [], + "bbox": [ + 112, + 760, + 285, + 838 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a8bff6232c301f3098ca966ef3cb1baf2a2f925d0d41aa64555c1cb374d75af.jpg", + "image_caption": [ + "Figure 5. The impact of different extrapolate strengths and different sampling strategies." + ], + "image_footnote": [], + "bbox": [ + 290, + 760, + 460, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Effect of Extrapolate Strength", + "text_level": 1, + "bbox": [ + 511, + 241, + 781, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our extrapolation experiments, we analyzed the impact of different $\\alpha$ values on performance using Equation 3. As shown in Figure 5a, as $\\alpha$ increases, the motion degree first rapidly rises within the 0-0.7 range, peaks, and then gradually declines between 1.2-2.0. Simultaneously, quality and consistency metrics decline. This indicates that extrapolation enhances motion degree but sacrifices quality and consistency. This aligns with the findings in Section 4.4 and underscores the importance of appropriate decoupling.", + "bbox": [ + 511, + 267, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Analysis on Decoupled Sampling Method", + "text_level": 1, + "bbox": [ + 511, + 425, + 864, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To study the impact of switch model time and choice, we conducted experiments illustrated in Figure 5b. We outline two model switch strategies during the denoising process from time $T$ to 0: (1) Initially using $\\theta_{dyn}^{*}$ and then switching to $\\theta_{sft}^{*}$ . (2) Starting with $\\theta_{sft}^{*}$ and then transitioning to $\\theta_{dyn}^{*}$ . The performance trends are shown in Figure 5b. The first strategy, which initially uses $\\theta_{dyn}^{*}$ , results in higher motion degree with a later switch (smaller $T$ ), but with poorer quality. Conversely, an earlier switch (larger $T$ ) reduces the degree. The second strategy, introducing $\\theta_{dyn}^{*}$ later, has a less pronounced impact on the degree. This aligns with our hypothesis in Section 3.4, which suggests that the early sampling phase significantly influences the degree.", + "bbox": [ + 511, + 450, + 906, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 674, + 633, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel Extrapolating and Decoupling framework for image-to-video generation. Our framework consists of three stages and can effectively model the motion control and motion degree. In the first stage, we explicitly inject the textual condition into I2V-DM to improve motion controllability. Then, we devise a training-free extrapolation strategy to boost the motion degree. Finally, we decouple the motion-relevant parameters and dynamically inject them into the I2V-DM along different de-noising steps. With our framework, we can generate videos with both natural motion and temporal consistency. To the best of our knowledge, we are the first to apply merging techniques in video generation.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12519", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 4, 5", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 6", + "[4] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 6", + "[5] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 6", + "[6] Feng Cheng, Ziyang Wang, Yi-Lin Sung, Yan-Bo Lin, Mohit Bansal, and Gedas Bertasius. Dam: Dynamic adapter merging for continual video qa learning, 2024. 2", + "[7] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Finegrained open domain image animation with motion guidance, 2023. 5, 6", + "[8] Felix Draxler, Kambis Veschgini, Manfred Salmhofer, and Fred Hamprecht. Essentially no barriers in neural network energy landscape. In International conference on machine learning, pages 1309-1318. PMLR, 2018. 2", + "[9] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020.", + "[10] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2", + "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1", + "[12] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 3", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[16] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 2, 4, 5, 8", + "[17] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together, 2024. 5", + "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 4", + "[20] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 5", + "[21] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 2, 5", + "[22] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 2022. 2", + "[23] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 2", + "[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[25] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[27] Weiming Ren, Harry Yang, Ge Zhang, Cong Wei, Xinrun Du, Stephen Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation. arXiv preprint arXiv:2402.04324, 2024. 1, 5, 6, 7" + ], + "bbox": [ + 517, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12520", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[30] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping, 2017. 5", + "[31] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[32] Mustafa Shukor, Corentin Dancette, Alexandre Rame, and Matthieu Cord. Unival: Unified model for image, video, audio and language tasks, 2023. 2", + "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 5", + "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11):1-7, 2012. 5", + "[36] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation, 2019. 5", + "[37] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation, 2023. 5", + "[38] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[39] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 5", + "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 1, 2, 3, 5, 6, 7", + "[41] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "when merging models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[42] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024.", + "[43] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*. 2, 5", + "[44] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 6", + "[45] Yiming Zhang, Zhening Xing, Yanhong Zeng, Youqing Fang, and Kai Chen. Pia: Your personalized image animator via plug-and-play modules in text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7747-7756, 2024. 1", + "[46] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. arXiv preprint arXiv:2406.15735, 2024. 1, 2, 4, 5, 6, 7, 8", + "[47] Chujie Zheng, Ziqi Wang, Heng Ji, Minlie Huang, and Nanyun Peng. Weak-to-strong extrapolation expedites alignment, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 484 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12521", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_model.json b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c29d6a74cb6a13363a2d12067893b5006c576698 --- /dev/null +++ b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.835, + 0.177 + ], + "angle": 0, + "content": "Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.202, + 0.874, + 0.259 + ], + "angle": 0, + "content": "Jie Tian\\(^{1\\dagger*}\\) Xiaoye Qu\\(^{1\\dagger}\\) Zhenyi Lu\\(^{1\\dagger}\\) Wei Wei\\(^{1\\boxtimes}\\) Sichen Liu\\(^{1}\\) Yu Cheng\\(^{2}\\) \n\\(^{1}\\)School of Computer Science & Technology, Huazhong University of Science and Technology, \n\\(^{2}\\)The Chinese University of Hong Kong" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.307, + 0.327, + 0.323 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.485, + 0.779 + ], + "angle": 0, + "content": "Image-to-Video (I2V) generation aims to synthesize a video clip according to a given image and condition (e.g., text). The key challenge of this task lies in simultaneously generating natural motions while preserving the original appearance of the images. However, current I2V diffusion models (I2V-DMs) often produce videos with limited motion degrees or exhibit uncontrollable motion that conflicts with the textual condition. To address these limitations, we propose a novel Extrapolating and Decoupling framework, which introduces model merging techniques to the I2V domain for the first time. Specifically, our framework consists of three separate stages: (1) Starting with a base I2V-DM, we explicitly inject the textual condition into the temporal module using a lightweight, learnable adapter and fine-tune the integrated model to improve motion controllability. (2) We introduce a training-free extrapolation strategy to amplify the dynamic range of the motion, effectively reversing the fine-tuning process to enhance the motion degree significantly. (3) With the above two-stage models excelling in motion controllability and degree, we decouple the relevant parameters associated with each type of motion ability and inject them into the base I2V-DM. Since the I2V-DM handles different levels of motion controllability and dynamics at various denoising time steps, we adjust the motion-aware parameters accordingly over time. Extensive qualitative and quantitative experiments have been conducted to demonstrate the superiority of our framework over existing methods. Code is available at https://github.com/Chuge0335/EDG" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.834, + 0.483, + 0.865 + ], + "angle": 0, + "content": "Recently, Image-to-Video generation (I2V) [11, 23, 40, 45, 46] has attracted increasing attention. Given a single im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution; * Work done during an internship at Shanghai AI Laboratory; \\(\\boxtimes\\) corresponding authors." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.3, + 0.887, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.45, + 0.887, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Figure 1. In the first example, Consist2V [27] yields a video with limited motion. In the second, DynamiCrafter [40] fails to follow the text condition and lacks controllability. Our method, however, achieves improved motion degree and better controllability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.793 + ], + "angle": 0, + "content": "age and a text condition, it aims to convert the still image into videos that conform to text instruction while preserving the original appearance of the images. This paper focuses on open-domain images and supports free text as guidance, rather than limited class labels. This conditioned I2V task is substantially challenging as it struggles with simultaneously maintaining appearance consistency and producing natural motions in the generated frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Unlike image synthesis, motion plays a significantly crucial role in image-to-video generation. However, most existing works lack effective motion modeling when generating videos. For instance, ConsistI2V [27] enhances the first frame conditioning through spatiotemporal attention, leading to limited expression ability for modeling motion degree as the model tends to over-rely on the conditional im" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12512" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "age. DynamiCrafter [40] interacts the text embedding and conditional image embedding with the U-Net spatial layers through a dual cross-attention, neglecting the temporal property of the motion and suffering from motion controllability. In this way, the generated videos either contradict the motion described in the text or are still image sequences with slow change. As shown in Figure 1, the videos generated by the previous model show cars that barely move, or the videos of the castle contradict the textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.23, + 0.483, + 0.591 + ], + "angle": 0, + "content": "To address the aforementioned issues, we present a novel Extrapolating and Decoupling framework to effectively model motion for image-to-video generation, improving both motion degree and motion controllability. Specifically, it consists of three stages. Firstly, we improve motion controllability by explicitly injecting the textual condition into the I2V diffusion model (I2V-DM). A lightweight learnable adapter is introduced to enhance the temporal attention module. Fine-tuning the integrated model improves motion controllability without altering the original I2V-DM architecture. Secondly, we observe degraded motion dynamics in our fine-tuned model, a phenomenon has also been observed by Zhao et al. [46]. To boost the motion degree without extra training, we introduce a novel training-free extrapolation strategy to restore and enhance motion degree dynamics by reversing the fine-tuning process. Thirdly, we decouple the parameters contributing to motion controllability and dynamics from the previous two stages, selectively integrating them into the final model. Recognizing that different denoising steps focus on varying motion ability and temporal consistency, we dynamically adjust the motion-aware parameters over time. With these three stages, we can generate videos with both natural motion and temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.482, + 0.7 + ], + "angle": 0, + "content": "To verify the effectiveness of our framework, we perform experiments on the VBench I2V benchmark and UCF101 dataset. Extensive qualitative and quantitative experiments demonstrate the notable superiority of our framework over existing I2V methods. Furthermore, we offer discussion and analysis of some insightful designs in our framework. To sum up, our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.779 + ], + "angle": 0, + "content": "- We propose a novel Extrapolating and Decoupling framework for video generation, which models motion controllability and degree dynamics. To best of our knowledge, this is the first use of model merging techniques in the I2V domain." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.841 + ], + "angle": 0, + "content": "- We introduce a lightweight adapter to enhance motion controllability while preserving the base I2V-DM structure and propose a novel training-free extrapolation strategy to amplify motion degree." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.483, + 0.886 + ], + "angle": 0, + "content": "- We make the first attempt to decouple parameters related to motion controllability and degree, dynamically injecting them into the I2V-DM across varying denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We conduct extensive experiments on multiple bench" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "marks, where our proposed framework consistently outperforms previous state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.151, + 0.655, + 0.167 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.181, + 0.828, + 0.198 + ], + "angle": 0, + "content": "2.1. Motion Modeling in I2V Generation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.209, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Diffusion models (DMs) [15, 33] have recently demonstrated remarkable generative capabilities in image and video generation [12, 24, 26, 28, 29]. In this paper, we focus on the image-conditioned video (I2V) generation task, where motion modeling aims to predict object movements. Traditionally, optical flow estimates the displacement field between consecutive frames. To effectively model motion in I2V tasks, LFDM [23] proposes a latent flow diffusion model that synthesizes a latent optical flow sequence to warp the given image in latent space for generating new videos. Motion-I2V [31] factorizes I2V into two stages: a diffusion-based motion field predictor for pixel trajectories and motion-augmented temporal attention to enhance 1-D temporal attention in video latent diffusion models. Recent works [38, 40, 44] enable diffusion models to handle motion modeling and video generation simultaneously. For instance, Dynamiccafter [40] uses a dual-stream image injection mechanism to leverage the motion prior of text-to-video diffusion models. In this paper, we build on DynamamiCrafter to design an Extrapolating and Decoupling framework to improve both motion controllability and degree." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.551, + 0.671, + 0.568 + ], + "angle": 0, + "content": "2.2. Model Merging" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.579, + 0.907, + 0.821 + ], + "angle": 0, + "content": "Model merging, widely explored in natural language processing and image classification, typically combines multiple fine-tuned task-specific models into a unified multitask model without additional training. Early methods [1, 8-10, 17, 22] leverage interpolation techniques to integrate multi-task knowledge effectively. Task-Arithmetic [16] extends these approaches by enabling complex arithmetic operations in weight space, offering finer control over the merging process. Recent studies [21, 25, 41-43] have further reduced interference during merging. Beyond interpolation-based methods, recent works have shown that sparsifying [43] or amplifying delta weights [47] can enhance model knowledge in merging. Additionally, merging techniques have been introduced to multimodal understanding such as question-answering [6, 32]. However, the video generation domain has remained unexplored until now." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We pioneer the introduction of model merging techniques into the video generation domain. Specifically, we leverage model merging to extrapolate motion degree knowledge and extract fine-grained decoupled knowledge, which are selectively merged at different diffusion stages." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12513" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.496, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.885, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.23, + 0.766, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.236, + 0.885, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.394, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Figure 2. The overview of our framework, consisting of three stages. (a) We improve the motion controllability by injecting the textual conditions into the temporal attention module in I2V-DM model and then fine-tune the integrated model. (b) The I2V-DM model suffers from a decrease in motion degree after fine-tuning. To mitigate it, we propose a training-free extrapolation to boost the motion degree by reversing the fine-tuning progress. (c) We decouple the relevant parameters contributing to motion controllability and motion degree from the last two-stage models. Moreover, we selectively inject these parameters into the I2V model along with the de-noising process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.182, + 0.494 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Given a reference image \\( I_0 \\) and a text prompt \\( c \\), image-to-video generation (I2V) targets at generating a sequence of subsequent video frames \\( V = \\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_N\\} \\). The key objective of I2V generation is to ensure that the generated video not only exhibits plausible motion but also faithfully preserves the visual appearance of the provided reference image. The overall framework of our method, shown in Figure 2, consists of three stages. First, we improve motion controllability guided by the text condition. Second, we propose a training-free extrapolation strategy to enhance motion dynamics. Finally, we decouple parameters related to motion controllability and dynamics, allowing flexible injection during the denoising process to generate videos with natural motion and temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.727, + 0.468, + 0.743 + ], + "angle": 0, + "content": "3.1. Preliminary: Latent Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Before introducing our framework, we first review the Latent Diffusion Models (LDMs) [28]. Specifically, it extends diffusion models [15] by operating in a more compact latent space rather than directly on pixels, reducing computational complexity. In LDMs, an input \\( x_0 \\) is first encoded into a latent representation \\( z_0 \\) using a pretrained Variational Autoencoder (VAE). This latent variable \\( z_0 \\) is then corrupted gradually over \\( T \\) steps by adding Gaussian noise, producing a sequence of noisy latents \\( \\{z_t\\}_{t=1}^T \\). The noise process follows \\( z_t \\mid z_0 \\sim \\mathcal{N}(z_t; \\sqrt{\\bar{\\alpha}_t} z_0, (1 - \\bar{\\alpha}_t)I) \\), where" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.907, + 0.539 + ], + "angle": 0, + "content": "\\(\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}\\) controls the noise schedule. To recover the original latent, LDMs learn a reverse process that denoises \\(z_{t}\\) back to \\(z_{0}\\) through a sequence of steps, guided by the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.54, + 0.905, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {z _ {0}, t, \\epsilon} \\| \\epsilon_ {\\theta} (z _ {t}, t) - \\epsilon \\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.611 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0, I)\\) represents Gaussian noise, \\(t\\) is a randomly sampled timestep, and \\(\\epsilon_{\\theta}(z_t, t)\\) is the learned noise estimator parameterized by \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.763 + ], + "angle": 0, + "content": "In this paper, we adopt an open-source video LDM, DynamiCrafter [40] as the base model. Given a video \\( x \\in \\mathbb{R}^{L \\times 3 \\times H \\times W} \\), we first encode it into a latent representation \\( z = \\mathcal{E}(x) \\), where \\( z \\in \\mathbb{R}^{L \\times C \\times h \\times w} \\) is derived on a frame-by-frame basis. Both the forward diffusion process \\( z_{t} = p(z_{0}, t) \\) and the backward denoising process \\( z_{t} = p_{\\theta}(z_{t-1}, c, t) \\) are conducted in this latent space, with \\( c \\) representing potential denoising conditions such as text prompts. Ultimately, the generated videos are obtained through the decoder \\( \\hat{x} = \\mathcal{D}(z) \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.818, + 0.79 + ], + "angle": 0, + "content": "3.2. Motion-Controllability Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In I2V generation, motion controllability is important for following textual conditions. To model motion, current I2VDMs [40] typically employ spatial attention to share information within a single frame and temporal attention to capture time sequence consistency [13]. However, methods like DynamiCrafter [40] only use text condition embedding in the spatial attention module, thus it does not share condi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "tional information across different time sequences, limiting its motion controllability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.318 + ], + "angle": 0, + "content": "To alleviate this problem, we explicitly inject textual conditions into DynamiCrafter's temporal attention with a lightweight adapter, as depicted in Figure 2(a). The spatial attention module remains unmodified as it already incorporates textual information. The adapter is non-intrusive and preserves the original model structure. Specifically, we extracts spatial features \\( z_{t} \\in \\mathbb{R}^{f \\times d \\times w \\times h} \\) and the text embeddings \\( t_{\\mathrm{embd}} \\in \\mathbb{R}^{l \\times d} \\) from the spatial attention layer and the CLIP text encoder respectively, where \\( f \\) is the number of frames, \\( d \\) is the feature dimension, and \\( w, h \\) are spatial dimensions. Then, the condition information in text embeddings are further compressed by a trainable Q-Former [19] by retaining the motion-relevant tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.329, + 0.483, + 0.349 + ], + "angle": 0, + "content": "\\[\nt _ {\\text {e m b d}} ^ {\\prime} = \\mathrm {Q} - \\operatorname {F o r m e r} \\left(t _ {\\text {e m b d}}\\right) \\in \\mathbb {R} ^ {f \\times d} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.48 + ], + "angle": 0, + "content": "\\(t_{\\mathrm{embd}}^{\\prime}\\) is then concatenated with spatial information \\(z_{t}\\) along the feature dimension, projected through \\(W_{\\mathrm{proj}}\\) to reduce the dimension to \\(d\\), and passed through temporal attention to inject the text control information into each frame. We finetune the model on WebVid-2M [2], updating only the adapter (the Q-Former and the linear projection \\(W_{\\mathrm{proj}}\\)) and temporal attention weights, ensuring minimal added parameters and efficient training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.355, + 0.507 + ], + "angle": 0, + "content": "3.3. Motion-Degree Extrapolation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.648 + ], + "angle": 0, + "content": "After fine-tuning I2V-DM DynamiCrafter with motion-controllability adaptation, we observe a reduction in motion degree, a common issue attributed to conditional image leakage [46]. During fine-tuning, Dynamicrafter concatenates the first frame with noise, which may cause the model to over-rely on the conditional image and neglect noisy inputs, especially at early timesteps. While adding time-dependent noise partially mitigates motion loss [46], it requires retraining and has limited effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.8 + ], + "angle": 0, + "content": "To improve the motion dynamics beyond the pre-trained model, we introduce a novel training-free and model-agnostic extrapolation approach, as shown in Figure 2 (b), which extremely pushes the \"unlearning\" process of motion controllability. In this way, we can achieve a model with amplified motion degree dynamics. Specifically, we estimate a degree-enhanced model, \\(\\theta_{dyn}\\), by extrapolating between a pre-trained model \\(\\theta_{pre}\\) and its fine-tuned version \\(\\theta_{sft}\\) with motion-controllability adaptation, amplifying their motion-dynamic differences:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.813, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\theta_ {d y n} = \\theta_ {p r e} + \\alpha \\left(\\theta_ {p r e} - \\theta_ {s f t}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Where the hyperparameter \\(\\alpha\\) controls the strength of this extrapolation. In this equation, we compute the difference of the model parameters except for the adapter introduced in Section 3.2. This process requires only two checkpoints and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "obtaining the motion degree-enhanced model without training. To help further understand our extrapolating method, we provide a theoretical analysis below." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.155, + 0.907, + 0.232 + ], + "angle": 0, + "content": "Theoretical Analysis. Assume that we have a motion degree metric score denoted as \\(\\mathcal{D}\\). Diffusion models are commonly regularized by a KL constraint in Gaussian distribution, implying that \\(\\theta_{pre}\\) lies within a small vicinity of \\(\\theta_{sft}\\). Therefore, a first-order Taylor expansion can be applied:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.242, + 0.905, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) = \\mathcal {D} \\left(\\theta_ {\\text {p r e}} + \\alpha \\Delta \\theta\\right) \\tag {4} \\\\ \\approx \\mathcal {D} (\\theta_ {p r e}) + \\alpha \\langle \\nabla \\mathcal {D} (\\theta_ {p r e}), \\Delta \\theta \\rangle \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.283, + 0.906, + 0.314 + ], + "angle": 0, + "content": "where \\(\\Delta\\) denotes the parameter update from \\(\\theta_{sft}\\) to \\(\\theta_{pre}\\). The change in the motion degree is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.324, + 0.905, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) - \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) = \\alpha \\langle \\nabla \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right), \\Delta \\theta \\rangle \\tag {5} \\\\ \\approx \\alpha \\langle \\nabla \\mathcal {D} (\\theta_ {p r e}), \\nabla \\mathcal {D} (\\theta_ {p r e}) \\rangle \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Assuming that the parameter update \\(\\Delta \\theta\\) during finetuning is approximately aligned with the gradient of \\(\\mathcal{D}\\) [14, 16], we can express \\(\\Delta \\theta\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.43, + 0.905, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta \\approx \\gamma \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.453, + 0.719, + 0.468 + ], + "angle": 0, + "content": "Thus by substitution, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.478, + 0.905, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) - \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) \\approx \\alpha \\left\\langle \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right), \\gamma \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) \\right\\rangle \\\\ = \\alpha \\gamma \\left\\langle \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right), \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right) \\right\\rangle \\tag {7} \\\\ = \\alpha \\gamma \\| \\nabla_ {\\theta} \\mathcal {D} (\\theta_ {p r e}) \\| ^ {2} \\geq 0. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.611 + ], + "angle": 0, + "content": "Since the norm of the gradient \\(\\| \\nabla_{\\theta}\\mathcal{D}(\\theta_{pre})\\| ^2\\) is always non-negative, the dynamic motion degree metric \\(\\mathcal{D}\\) shall increases after the extrapolation, provided that the pretrained weight \\(\\theta_{pre}\\) is not at a local maximum of \\(\\mathcal{D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.816, + 0.636 + ], + "angle": 0, + "content": "3.4. Decoupled Injection and Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.793 + ], + "angle": 0, + "content": "While motion-controllability adaptation and training-free degree extrapolation can enhance motion control and diversity respectively, directly combining them poses a significant challenge due to their opposing optimization directions: fine-tuning for control versus reversed extrapolation for degree. To balance motion controllability and degree without compromising image consistency, we propose a decoupling strategy that separates degree-related and control-related parameters, applying them at distinct denoising stages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Isolating Parameter Sets. Here, we categorize the obtained knowledge into three parameter sets that support core abilities: motion control (\\(\\theta_{adt}\\)), motion degree (\\(\\theta_{deg}\\)), and video consistency (\\(\\theta_{con}\\)). We aim to integrate degree and control knowledge in the early denoising stage for long-term temporal planning, while emphasize consistency and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12515" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "control later for refining subject appearance. However, they are embedded within \\(\\theta_{pre}\\) and \\(\\theta_{sft}\\). Therefore, we need parameter isolation to decouple three types of knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.424 + ], + "angle": 0, + "content": "We isolate these sets based on the following insight: (1) Based on the implicit gradient assumption (Section 3.3), \\(\\theta_{sft} - \\theta_{pre}\\) captures the implicit control knowledge learned during the adaptation process, serves as \\(\\theta_{adt}\\) paired with \\((\\theta_{adapter})\\), incorporating elements like Q-Formers and projectors. (2) \\(\\theta_{sft}\\) exhibits degraded degree knowledge (11.67 vs. 68.54 for \\(\\theta_{pre}\\) in Table 1) and primarily contains control and consistency knowledge. Therefore, consistency knowledge \\(\\theta_{con}\\) can be directly derived as \\(\\theta_{sft} - \\theta_{adt}\\). (3) The extrapolated \\(\\theta_{dyn}\\) has similar control/consistency proportions as \\(\\theta_{pre}\\), therefore \\(\\theta_{deg} = \\theta_{dyn} - \\theta_{pre}\\) can remove control/consistency components, isolating degree knowledge, even though it originates from \\(\\theta_{sft}\\). (4) Fine-tuned parameters often contain redundant elements, which can cause conflicts during merging [21, 43]. To minimize overlap between different knowledge while preserving performance, we apply DARE pruning [43] at a pruning rate \\(p\\) (e.g., \\(p = 70\\%\\)). Distinct parameter groups are created using random masks \\(\\mathcal{M}\\) amplified from a Bernoulli distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.435, + 0.423, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {1} \\sim \\operatorname {B e r n o u l l i} (p), \\quad \\mathcal {M} _ {2} \\sim \\operatorname {B e r n o u l l i} (p),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.454, + 0.481, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\theta_ {a d t} = \\left\\{\\theta_ {a d p a t e r} \\right\\} \\cup \\left\\{\\mathrm {D A R E} _ {\\mathcal {M} _ {2}} \\left(\\theta_ {s f t} - \\theta_ {p r e}\\right) \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.474, + 0.346, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\theta_ {d e g} = \\mathrm {D A R E} _ {\\mathcal {M} _ {1}} \\left(\\theta_ {d y n} - \\theta_ {p r e}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.493, + 0.261, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\theta_ {c o n} = \\theta_ {s f t} - \\theta_ {a d t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.58 + ], + "angle": 0, + "content": "Finally, Task-Arithmetic [16] is employed to inject control parameters \\(\\theta_{adt}\\) into both \\(\\theta_{deg}\\) and \\(\\theta_{con}\\), deriving two enhanced models: the degree-enhanced model \\(\\theta_{dyn}^{*}\\) and the consistency-enhanced model \\(\\theta_{con}^{*}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.591, + 0.482, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\theta_ {d y n} ^ {*} = \\text {T a s k - A r i t h m e t i c} (\\theta_ {d e g}, \\theta_ {a d t}), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.612, + 0.408, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\theta_ {c o n} ^ {*} = \\text {T a k - A r i t h m e t i c} \\left(\\theta_ {c o n}, \\theta_ {a d t}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Decoupled Video Sampling. During the denoising process for video generation, we typically observe that in the early stages, I2V models prioritize long-term temporal planning, including the motion of both subjects and textures, while later stages often focus more on refining subject appearance, which is mainly related to content consistency and overall image quality [37, 39]. Therefore, in the first \\(K\\) steps of the denoising process, we adjust the video generator to make it more dynamic-enhanced and control-enhanced to increase motion controllability. In the later stages of the denoising process, we re-weight the motion-related modules, allowing the model to focus more on restoring the specific details of each frame of the subject, thereby generating high-consistency videos. Specifically, we define a generated model \\(\\theta_{gen}\\) by blending the degree-enhanced model \\(\\theta_{dyn}^{*}\\) and the consistency-merged model \\(\\theta_{con}^{*}\\) based on the denoising step \\(t\\) from \\(T\\) to 0:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.104, + 0.816, + 0.122 + ], + "angle": 0, + "content": "\\[\n\\theta_ {g e n} = \\alpha_ {t} \\theta_ {d y n} ^ {*} + (1 - \\alpha_ {t}) \\theta_ {c o n} ^ {*},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.124, + 0.905, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {t} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} t > T - K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.171, + 0.907, + 0.202 + ], + "angle": 0, + "content": "Here, \\( T \\) is the total number of denoising steps, and \\( K \\) is a threshold that determines the switching point." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.216, + 0.638, + 0.233 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.241, + 0.596, + 0.258 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.263, + 0.906, + 0.354 + ], + "angle": 0, + "content": "In all experiments, videos and images are resized to a \\(320 \\times 512\\) center-cropped format. During model training, we set the learning rate to \\(1 \\times 10^{-5}\\), a batch size of 64, and trained for 100K steps on 8 A100 GPUs. For inference, we use DDIM [34] with 50 steps and a classifier-free guidance scale of 7.5, generating 16 frames for a 2-second video." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.355, + 0.906, + 0.429 + ], + "angle": 0, + "content": "Datasets. We use the WebVid-2M dataset [2] for supervised fine-tuning, which contains two million video-text pairs from stock footage websites. For motion controllability annotation, we use CoTracker [18] to annotate motion information on the WebVid-2M dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Evaluation. The evaluation benchmarks include VBench [27] and UCF101 [35]. VBench is a comprehensive benchmark for video generation, incorporating recent models and extensive evaluation metrics. UCF101 contains 101 human action categories with diverse background and lighting variations. For VBench, following previous work [20, 46], we utilize the official dataset to assess the model across several metrics: I2V-Subject Consistency (I2V Sub.), I2V-Background Consistency (I2V Back.), video quality, camera motion, and dynamic degree. The video quality evaluation considers dimensions including Subject Consistency, Background Consistency, Smoothness, Aesthetic Score, Imaging Quality, and Temporal Flickering. The Avg. score is calculated as \\(\\frac{1}{10} * (\\text{I2V Sub.} + \\text{I2V Back.} + 6 * \\text{Video Quality} + \\text{Motion Degree} + \\text{Motion Control})\\), which is consistent with VBench I2V original setting. For UCF101, similar to [39, 40], we use Fréchet Video Distance (FVD) [36] and Inception Score (IS) [30] to assess video quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.713, + 0.736, + 0.729 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "VBenchmark Results Analysis. We compare our method with two types of approaches on VBenchmark. The first kind of comparing methods include those requiring large-scale video pretraining, such as animate-Anything [7], SVD [3], and DynamiCrafter [40]. The other comparison methods are finetuning/inference strategies based on DynamiCrafter, including naive finetuning with the same training setup as our method (Naive FT) and Conditional Image Leakage (CIL) methods [46] (CIL Infer and CIL FT). Notably, our method performs best in Motion Degree and Motion Control across all baselines, even including large-scale pretrained methods," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12516" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the VBench I2V benchmark for various I2V-DM pretrained baselines and finetuning/inference techniques, including our methods based on DynamiCrafter. \"PT.\" in the table denotes whether the method requires video pretraining, and \"FT.\" denotes whether the method requires supervised finetuning on WebVid-2M. The global best and DynamiCrafter-based best performing metrics are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.158, + 0.905, + 0.364 + ], + "angle": 0, + "content": "
ModelAvg.PT.FT.I2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
ConsistI2V (Ren et al.)75.88✓✗95.8295.9585.7418.6233.92
SEINE (Chen et al.)76.56✓✗96.5796.8085.7134.3123.67
Animate-Anything (Dai et al.)74.61✓✗98.7698.5888.842.6813.08
VideoCrafter (Chen et al.)76.96✓✗91.1796.9487.5522.6033.60
SVD-XT-1.0 (Blattmann et al.)-✓✗97.5297.6386.5452.36-
SVD-XT-1.1 (Blattmann et al.)-✓✗97.5197.6286.6643.17-
I2VGen-XL (Zhang et al.)76.00✓✗96.4896.8387.0226.1018.48
DynamiCrafter (Xing et al.)79.93✓✗94.9195.3084.9568.5430.85
DynamiCrafter-Naive FT72.91✗✓97.9298.7083.5211.6717.82
DynamiCrafter-CIL Infer (Zhao et al.)80.47✗✗96.9693.3485.2869.4331.29
DynamiCrafter-CIL FT (Zhao et al.)77.55✗✓96.9097.7987.1025.6930.50
DynamiCrafter-Ours84.25✗✓97.0396.6286.2287.6443.87
DynamiCrafter-Ours w/ CIL Infer84.81✗✓97.1195.4587.1089.2143.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.387, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of the DynamiCrafter-based method on the UCF-101 benchmark." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.426, + 0.472, + 0.512 + ], + "angle": 0, + "content": "
ModelFVD↓IS↑Motion Degree↑
DynamiCrafter288.0113.3168.17
DynamiCrafter-Naive FT373.7713.1947.66
DynamiCrafter-CIL FT409.6913.6853.78
DynamiCrafter-Ours324.7913.7177.93
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.852 + ], + "angle": 0, + "content": "as shown in Table 1. We observe that Naive FT causes a significant degradation in Motion Degree (11.67 vs 68.54), consistent with previous findings [46]. Naive FT achieves the highest I2V Subject and Background Consistency (I2V Sub. and I2V Back.) among the DynamiCrafter-based methods because lower motion degree results in higher consistency scores. Interestingly, motion control is also negatively affected in Naive FT (17.82 vs 11.54). We speculate that this is due to the insufficient motion degree, which hinders the model's ability to follow control instructions effectively. The CIL method can mitigate these effects, either by employing the inference strategy on the pre-trained model (CIL Infer) or by repairing the Naive FT (CIL FT). However, its enhancements in Motion Control and Motion Degree are limited. In contrast, our method, combining adaptation and extrapolation, significantly improves both motion degree and motion control (11.67 vs 87.64) while preserving similar video quality (87.10). Furthermore, our method is orthogonal to the CIL Infer method, and integrating it with our method further boosts performance (Avg. 84.25 vs 84.81)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "UCF101 Results Analysis. In the UCF101 dataset, our method still exhibits superior motion degree ability despite" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.389, + 0.709, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.389, + 0.901, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.492, + 0.71, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.492, + 0.901, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.648 + ], + "angle": 0, + "content": "Figure 3. Human Evaluation of our method Compared to SVD, VideoCrafter (VC), DynamiCrafter (DC), and its variant of Native Fine-Tuned (DC-FT) and CIL-Based Methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.907, + 0.799 + ], + "angle": 0, + "content": "the low resolution and poor image quality. Both Naive FT and CIL FT suffer from degradation in Fréchet Video Distance (FVD), as shown in Table 2. Unlike the performance on VBenchmark, CIL performs worse than Naive FT on UCF101, likely due to the added noise schedule disrupting the already noisy and low-resolution images. Our method, however, is more stable in FVD and outperforms in Inception Score (IS) and Motion Degree." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "User Study. As shown in Figure 3, our method excels in Motion Degree and Motion Control while remaining competitive in video quality and I2V consistency. The results are consistent with the above findings, supporting that our strategies enhance both video motion dynamics and control" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12517" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.087, + 0.854, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.623, + 0.908, + 0.667 + ], + "angle": 0, + "content": "Figure 4. In the visual comparisons, we generated 16-frame videos using DynamiCrafter, CIL, Consistent-I2V, and our model. We selected 3 to 4 frames at regular intervals for display at the top. The upper section visualizes the overall aesthetic quality, while the lower section uses Optical Flow to illustrate the dynamic qualities of the generated videos for each method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.716 + ], + "angle": 0, + "content": "while preserving quality and consistency, yielding superior overall performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.727, + 0.303, + 0.743 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To empirically evaluate the visual quality of our method, we compare it with several state-of-the-art image-to-video generation approaches, including DynamiCrafter [40], CIL [46], and ConsistI2V [27]. In Figure 4, we observe that our method generates temporally coherent videos with a sufficient range of motion that closely adhere to the input text conditions. In contrast, DynamiCrafter exhibits inconsistencies in text control, likely due to challenges in capturing the semantic content of the input image. The ConsistI2V method produces videos that remain static over time, with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.822 + ], + "angle": 0, + "content": "the color points on the figures almost unchanged, indicating a lack of motion dynamics and control. While CIL generates videos that semantically resemble the input images and have enhanced motion compared to ConsistI2V, its improvements are limited. CIL does not sufficiently vary the content dynamically to match the text prompts, resulting in limited motion and control. Our method, however, preserves image consistency while achieving a large motion degree and exhibiting good controllability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.667, + 0.85 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We demonstrate the importance of our extrapolation and decoupling methods in Table 3: (1) On top of our adaptation training, using our extrapolation and decoupling techniques" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12518" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.282, + 0.09, + 0.713, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation study of the three stages of our framework on VBenchmark." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.115, + 0.868, + 0.217 + ], + "angle": 0, + "content": "
(a) Adaptation(b)Extrapolation(c) DecoupleI2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
✓XX97.5196.3986.6122.8549.31
X✓X94.1495.6283.5298.2117.63
XX✓94.6096.2183.9569.6132.37
✓✓X91.1192.8684.3271.6826.71
X✓✓96.0194.6084.3493.3321.85
✓✓✓97.0396.6286.2287.6443.87
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.545 + ], + "angle": 0, + "content": "significantly improves the motion degree, from 22.85 to 87.64. (2) Using extrapolation alone can drastically improve the motion degree (69.61 vs 93.33), but it comes at the cost of degradation in I2V-Subject Consistency (I2V Sub.) and I2V-Background Consistency (I2V Back.), as well as motion control. This is because reversing the training process of the pre-trained model to the SFT model unlearns the camera motion ability, causing significant interference if directly merged [16]. (3) Using the decoupling technique alone can improve the motion degree while preserving other performance metrics. This indicates that the I2V model indeed exhibits functional differentiation during the sampling process. (4) Adding adaptation, though it increases training costs, greatly enhances the Motion Control aspect of the model. The degradation in motion degree indicates that direct supervision is suboptimal and can lead to minor motion dynamics, a finding consistent with previous research [46]. (5) When paired with the decoupling technique, we achieve optimal results, integrating high motion degree and control ability while preserving consistency and video quality." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.559, + 0.287, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.56, + 0.462, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.652, + 0.284, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.652, + 0.462, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.746, + 0.381, + 0.758 + ], + "angle": 0, + "content": "(a) Effect of extrapolate strength \\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.761, + 0.287, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.761, + 0.462, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.844, + 0.436, + 0.856 + ], + "angle": 0, + "content": "(b) Effect of different switching time and model choices." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.868, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Figure 5. The impact of different extrapolate strengths and different sampling strategies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.242, + 0.782, + 0.259 + ], + "angle": 0, + "content": "4.5. Effect of Extrapolate Strength" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.404 + ], + "angle": 0, + "content": "In our extrapolation experiments, we analyzed the impact of different \\(\\alpha\\) values on performance using Equation 3. As shown in Figure 5a, as \\(\\alpha\\) increases, the motion degree first rapidly rises within the 0-0.7 range, peaks, and then gradually declines between 1.2-2.0. Simultaneously, quality and consistency metrics decline. This indicates that extrapolation enhances motion degree but sacrifices quality and consistency. This aligns with the findings in Section 4.4 and underscores the importance of appropriate decoupling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.866, + 0.442 + ], + "angle": 0, + "content": "4.6. Analysis on Decoupled Sampling Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.65 + ], + "angle": 0, + "content": "To study the impact of switch model time and choice, we conducted experiments illustrated in Figure 5b. We outline two model switch strategies during the denoising process from time \\( T \\) to 0: (1) Initially using \\( \\theta_{dyn}^{*} \\) and then switching to \\( \\theta_{sft}^{*} \\). (2) Starting with \\( \\theta_{sft}^{*} \\) and then transitioning to \\( \\theta_{dyn}^{*} \\). The performance trends are shown in Figure 5b. The first strategy, which initially uses \\( \\theta_{dyn}^{*} \\), results in higher motion degree with a later switch (smaller \\( T \\)), but with poorer quality. Conversely, an earlier switch (larger \\( T \\)) reduces the degree. The second strategy, introducing \\( \\theta_{dyn}^{*} \\) later, has a less pronounced impact on the degree. This aligns with our hypothesis in Section 3.4, which suggests that the early sampling phase significantly influences the degree." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.675, + 0.634, + 0.69 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a novel Extrapolating and Decoupling framework for image-to-video generation. Our framework consists of three stages and can effectively model the motion control and motion degree. In the first stage, we explicitly inject the textual condition into I2V-DM to improve motion controllability. Then, we devise a training-free extrapolation strategy to boost the motion degree. Finally, we decouple the motion-relevant parameters and dynamically inject them into the I2V-DM along different de-noising steps. With our framework, we can generate videos with both natural motion and temporal consistency. To the best of our knowledge, we are the first to apply merging techniques in video generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12519" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[5] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[6] Feng Cheng, Ziyang Wang, Yi-Lin Sung, Yan-Bo Lin, Mohit Bansal, and Gedas Bertasius. Dam: Dynamic adapter merging for continual video qa learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[7] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Finegrained open domain image animation with motion guidance, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[8] Felix Draxler, Kambis Veschgini, Manfred Salmhofer, and Fred Hamprecht. Essentially no barriers in neural network energy landscape. In International conference on machine learning, pages 1309-1318. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[9] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[10] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.884 + ], + "angle": 0, + "content": "[12] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.15 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.152, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[17] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.308, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.351, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[20] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[21] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[22] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.907, + 0.66 + ], + "angle": 0, + "content": "[23] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[25] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Weiming Ren, Harry Yang, Ge Zhang, Cong Wei, Xinrun Du, Stephen Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation. arXiv preprint arXiv:2402.04324, 2024. 1, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12520" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[30] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[31] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[32] Mustafa Shukor, Corentin Dancette, Alexandre Rame, and Matthieu Cord. Unival: Unified model for image, video, audio and language tasks, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11):1-7, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[36] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[37] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[38] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[39] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "when merging models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[42] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[43] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[44] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[45] Yiming Zhang, Zhening Xing, Yanhong Zeng, Youqing Fang, and Kai Chen. Pia: Your personalized image animator via plug-and-play modules in text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7747-7756, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[46] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. arXiv preprint arXiv:2406.15735, 2024. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[47] Chujie Zheng, Ziqi Wang, Heng Ji, Minlie Huang, and Nanyun Peng. Weak-to-strong extrapolation expedites alignment, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12521" + } + ] +] \ No newline at end of file diff --git a/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_origin.pdf b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f8832004d4d86c51736379f927ae89db6f28eb1b --- /dev/null +++ b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/34056abe-9c83-4a9a-ac86-67d3126b13e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:838f95091e44fdd253a139551ab6b153374b0c81f7f4e7f1d661a50709bbab44 +size 6676628 diff --git a/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/full.md b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c8e72c6b391a4581e3a6ce448ad540c6f0307286 --- /dev/null +++ b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/full.md @@ -0,0 +1,317 @@ +# Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think + +Jie Tian $^{1\dagger*}$ Xiaoye Qu $^{1\dagger}$ Zhenyi Lu $^{1\dagger}$ Wei Wei $^{1\boxtimes}$ Sichen Liu $^{1}$ Yu Cheng $^{2}$ $^{1}$ School of Computer Science & Technology, Huazhong University of Science and Technology, + $^{2}$ The Chinese University of Hong Kong + +# Abstract + +Image-to-Video (I2V) generation aims to synthesize a video clip according to a given image and condition (e.g., text). The key challenge of this task lies in simultaneously generating natural motions while preserving the original appearance of the images. However, current I2V diffusion models (I2V-DMs) often produce videos with limited motion degrees or exhibit uncontrollable motion that conflicts with the textual condition. To address these limitations, we propose a novel Extrapolating and Decoupling framework, which introduces model merging techniques to the I2V domain for the first time. Specifically, our framework consists of three separate stages: (1) Starting with a base I2V-DM, we explicitly inject the textual condition into the temporal module using a lightweight, learnable adapter and fine-tune the integrated model to improve motion controllability. (2) We introduce a training-free extrapolation strategy to amplify the dynamic range of the motion, effectively reversing the fine-tuning process to enhance the motion degree significantly. (3) With the above two-stage models excelling in motion controllability and degree, we decouple the relevant parameters associated with each type of motion ability and inject them into the base I2V-DM. Since the I2V-DM handles different levels of motion controllability and dynamics at various denoising time steps, we adjust the motion-aware parameters accordingly over time. Extensive qualitative and quantitative experiments have been conducted to demonstrate the superiority of our framework over existing methods. Code is available at https://github.com/Chuge0335/EDG + +# 1. Introduction + +Recently, Image-to-Video generation (I2V) [11, 23, 40, 45, 46] has attracted increasing attention. Given a single im + +![](images/e0e69f353bba46926400601b34edbf1fed709f97bd8d08924a1a3b7bcdfa5426.jpg) + +![](images/45439e2b85fa9d023fe70d258dd5ed8df22883f215f6341db0c20d45c41c8b38.jpg) +Figure 1. In the first example, Consist2V [27] yields a video with limited motion. In the second, DynamiCrafter [40] fails to follow the text condition and lacks controllability. Our method, however, achieves improved motion degree and better controllability. + +age and a text condition, it aims to convert the still image into videos that conform to text instruction while preserving the original appearance of the images. This paper focuses on open-domain images and supports free text as guidance, rather than limited class labels. This conditioned I2V task is substantially challenging as it struggles with simultaneously maintaining appearance consistency and producing natural motions in the generated frames. + +Unlike image synthesis, motion plays a significantly crucial role in image-to-video generation. However, most existing works lack effective motion modeling when generating videos. For instance, ConsistI2V [27] enhances the first frame conditioning through spatiotemporal attention, leading to limited expression ability for modeling motion degree as the model tends to over-rely on the conditional im + +age. DynamiCrafter [40] interacts the text embedding and conditional image embedding with the U-Net spatial layers through a dual cross-attention, neglecting the temporal property of the motion and suffering from motion controllability. In this way, the generated videos either contradict the motion described in the text or are still image sequences with slow change. As shown in Figure 1, the videos generated by the previous model show cars that barely move, or the videos of the castle contradict the textual descriptions. + +To address the aforementioned issues, we present a novel Extrapolating and Decoupling framework to effectively model motion for image-to-video generation, improving both motion degree and motion controllability. Specifically, it consists of three stages. Firstly, we improve motion controllability by explicitly injecting the textual condition into the I2V diffusion model (I2V-DM). A lightweight learnable adapter is introduced to enhance the temporal attention module. Fine-tuning the integrated model improves motion controllability without altering the original I2V-DM architecture. Secondly, we observe degraded motion dynamics in our fine-tuned model, a phenomenon has also been observed by Zhao et al. [46]. To boost the motion degree without extra training, we introduce a novel training-free extrapolation strategy to restore and enhance motion degree dynamics by reversing the fine-tuning process. Thirdly, we decouple the parameters contributing to motion controllability and dynamics from the previous two stages, selectively integrating them into the final model. Recognizing that different denoising steps focus on varying motion ability and temporal consistency, we dynamically adjust the motion-aware parameters over time. With these three stages, we can generate videos with both natural motion and temporal consistency. + +To verify the effectiveness of our framework, we perform experiments on the VBench I2V benchmark and UCF101 dataset. Extensive qualitative and quantitative experiments demonstrate the notable superiority of our framework over existing I2V methods. Furthermore, we offer discussion and analysis of some insightful designs in our framework. To sum up, our contributions are summarized as follows: + +- We propose a novel Extrapolating and Decoupling framework for video generation, which models motion controllability and degree dynamics. To best of our knowledge, this is the first use of model merging techniques in the I2V domain. +- We introduce a lightweight adapter to enhance motion controllability while preserving the base I2V-DM structure and propose a novel training-free extrapolation strategy to amplify motion degree. +- We make the first attempt to decouple parameters related to motion controllability and degree, dynamically injecting them into the I2V-DM across varying denoising steps. +- We conduct extensive experiments on multiple bench + +marks, where our proposed framework consistently outperforms previous state-of-the-art methods. + +# 2. Related Work + +# 2.1. Motion Modeling in I2V Generation + +Diffusion models (DMs) [15, 33] have recently demonstrated remarkable generative capabilities in image and video generation [12, 24, 26, 28, 29]. In this paper, we focus on the image-conditioned video (I2V) generation task, where motion modeling aims to predict object movements. Traditionally, optical flow estimates the displacement field between consecutive frames. To effectively model motion in I2V tasks, LFDM [23] proposes a latent flow diffusion model that synthesizes a latent optical flow sequence to warp the given image in latent space for generating new videos. Motion-I2V [31] factorizes I2V into two stages: a diffusion-based motion field predictor for pixel trajectories and motion-augmented temporal attention to enhance 1-D temporal attention in video latent diffusion models. Recent works [38, 40, 44] enable diffusion models to handle motion modeling and video generation simultaneously. For instance, Dynamiccafter [40] uses a dual-stream image injection mechanism to leverage the motion prior of text-to-video diffusion models. In this paper, we build on DynamamiCrafter to design an Extrapolating and Decoupling framework to improve both motion controllability and degree. + +# 2.2. Model Merging + +Model merging, widely explored in natural language processing and image classification, typically combines multiple fine-tuned task-specific models into a unified multitask model without additional training. Early methods [1, 8-10, 17, 22] leverage interpolation techniques to integrate multi-task knowledge effectively. Task-Arithmetic [16] extends these approaches by enabling complex arithmetic operations in weight space, offering finer control over the merging process. Recent studies [21, 25, 41-43] have further reduced interference during merging. Beyond interpolation-based methods, recent works have shown that sparsifying [43] or amplifying delta weights [47] can enhance model knowledge in merging. Additionally, merging techniques have been introduced to multimodal understanding such as question-answering [6, 32]. However, the video generation domain has remained unexplored until now. + +We pioneer the introduction of model merging techniques into the video generation domain. Specifically, we leverage model merging to extrapolate motion degree knowledge and extract fine-grained decoupled knowledge, which are selectively merged at different diffusion stages. + +![](images/5b10612b13869a7de2d99ae36aa38144f6812edc4bbb5e7186a9d9cd542b3aac.jpg) + +![](images/fa1dd35b126828d0a192f07f99ee42b3f22f1e9fbde6ed7c55e873502a93321d.jpg) + +![](images/f5b6642b6e4e59b307915a7e2d7dcbfbcbd330e1a77e6a7de9d29f44ad5e0eb3.jpg) +Figure 2. The overview of our framework, consisting of three stages. (a) We improve the motion controllability by injecting the textual conditions into the temporal attention module in I2V-DM model and then fine-tune the integrated model. (b) The I2V-DM model suffers from a decrease in motion degree after fine-tuning. To mitigate it, we propose a training-free extrapolation to boost the motion degree by reversing the fine-tuning progress. (c) We decouple the relevant parameters contributing to motion controllability and motion degree from the last two-stage models. Moreover, we selectively inject these parameters into the I2V model along with the de-noising process. + +![](images/4cb4cb9d813853827865359c69548fb1b1782960468e3f12906af4db2c7b7d84.jpg) + +# 3. Method + +Given a reference image $I_0$ and a text prompt $c$ , image-to-video generation (I2V) targets at generating a sequence of subsequent video frames $V = \{\hat{I}_1, \hat{I}_2, \dots, \hat{I}_N\}$ . The key objective of I2V generation is to ensure that the generated video not only exhibits plausible motion but also faithfully preserves the visual appearance of the provided reference image. The overall framework of our method, shown in Figure 2, consists of three stages. First, we improve motion controllability guided by the text condition. Second, we propose a training-free extrapolation strategy to enhance motion dynamics. Finally, we decouple parameters related to motion controllability and dynamics, allowing flexible injection during the denoising process to generate videos with natural motion and temporal consistency. + +# 3.1. Preliminary: Latent Video Diffusion Models + +Before introducing our framework, we first review the Latent Diffusion Models (LDMs) [28]. Specifically, it extends diffusion models [15] by operating in a more compact latent space rather than directly on pixels, reducing computational complexity. In LDMs, an input $x_0$ is first encoded into a latent representation $z_0$ using a pretrained Variational Autoencoder (VAE). This latent variable $z_0$ is then corrupted gradually over $T$ steps by adding Gaussian noise, producing a sequence of noisy latents $\{z_t\}_{t=1}^T$ . The noise process follows $z_t \mid z_0 \sim \mathcal{N}(z_t; \sqrt{\bar{\alpha}_t} z_0, (1 - \bar{\alpha}_t)I)$ , where + +$\bar{\alpha}_{t} = \prod_{i=1}^{t} \alpha_{i}$ controls the noise schedule. To recover the original latent, LDMs learn a reverse process that denoises $z_{t}$ back to $z_{0}$ through a sequence of steps, guided by the objective: + +$$ +\mathcal {L} = \mathbb {E} _ {z _ {0}, t, \epsilon} \| \epsilon_ {\theta} (z _ {t}, t) - \epsilon \| ^ {2}, \tag {1} +$$ + +where $\epsilon \sim \mathcal{N}(0, I)$ represents Gaussian noise, $t$ is a randomly sampled timestep, and $\epsilon_{\theta}(z_t, t)$ is the learned noise estimator parameterized by $\theta$ . + +In this paper, we adopt an open-source video LDM, DynamiCrafter [40] as the base model. Given a video $x \in \mathbb{R}^{L \times 3 \times H \times W}$ , we first encode it into a latent representation $z = \mathcal{E}(x)$ , where $z \in \mathbb{R}^{L \times C \times h \times w}$ is derived on a frame-by-frame basis. Both the forward diffusion process $z_{t} = p(z_{0}, t)$ and the backward denoising process $z_{t} = p_{\theta}(z_{t-1}, c, t)$ are conducted in this latent space, with $c$ representing potential denoising conditions such as text prompts. Ultimately, the generated videos are obtained through the decoder $\hat{x} = \mathcal{D}(z)$ . + +# 3.2. Motion-Controllability Adaptation + +In I2V generation, motion controllability is important for following textual conditions. To model motion, current I2VDMs [40] typically employ spatial attention to share information within a single frame and temporal attention to capture time sequence consistency [13]. However, methods like DynamiCrafter [40] only use text condition embedding in the spatial attention module, thus it does not share condi + +tional information across different time sequences, limiting its motion controllability. + +To alleviate this problem, we explicitly inject textual conditions into DynamiCrafter's temporal attention with a lightweight adapter, as depicted in Figure 2(a). The spatial attention module remains unmodified as it already incorporates textual information. The adapter is non-intrusive and preserves the original model structure. Specifically, we extracts spatial features $z_{t} \in \mathbb{R}^{f \times d \times w \times h}$ and the text embeddings $t_{\mathrm{embd}} \in \mathbb{R}^{l \times d}$ from the spatial attention layer and the CLIP text encoder respectively, where $f$ is the number of frames, $d$ is the feature dimension, and $w, h$ are spatial dimensions. Then, the condition information in text embeddings are further compressed by a trainable Q-Former [19] by retaining the motion-relevant tokens: + +$$ +t _ {\text {e m b d}} ^ {\prime} = \mathrm {Q} - \operatorname {F o r m e r} \left(t _ {\text {e m b d}}\right) \in \mathbb {R} ^ {f \times d} \tag {2} +$$ + +$t_{\mathrm{embd}}^{\prime}$ is then concatenated with spatial information $z_{t}$ along the feature dimension, projected through $W_{\mathrm{proj}}$ to reduce the dimension to $d$ , and passed through temporal attention to inject the text control information into each frame. We finetune the model on WebVid-2M [2], updating only the adapter (the Q-Former and the linear projection $W_{\mathrm{proj}}$ ) and temporal attention weights, ensuring minimal added parameters and efficient training. + +# 3.3. Motion-Degree Extrapolation + +After fine-tuning I2V-DM DynamiCrafter with motion-controllability adaptation, we observe a reduction in motion degree, a common issue attributed to conditional image leakage [46]. During fine-tuning, Dynamicrafter concatenates the first frame with noise, which may cause the model to over-rely on the conditional image and neglect noisy inputs, especially at early timesteps. While adding time-dependent noise partially mitigates motion loss [46], it requires retraining and has limited effectiveness. + +To improve the motion dynamics beyond the pre-trained model, we introduce a novel training-free and model-agnostic extrapolation approach, as shown in Figure 2 (b), which extremely pushes the "unlearning" process of motion controllability. In this way, we can achieve a model with amplified motion degree dynamics. Specifically, we estimate a degree-enhanced model, $\theta_{dyn}$ , by extrapolating between a pre-trained model $\theta_{pre}$ and its fine-tuned version $\theta_{sft}$ with motion-controllability adaptation, amplifying their motion-dynamic differences: + +$$ +\theta_ {d y n} = \theta_ {p r e} + \alpha \left(\theta_ {p r e} - \theta_ {s f t}\right) \tag {3} +$$ + +Where the hyperparameter $\alpha$ controls the strength of this extrapolation. In this equation, we compute the difference of the model parameters except for the adapter introduced in Section 3.2. This process requires only two checkpoints and + +obtaining the motion degree-enhanced model without training. To help further understand our extrapolating method, we provide a theoretical analysis below. + +Theoretical Analysis. Assume that we have a motion degree metric score denoted as $\mathcal{D}$ . Diffusion models are commonly regularized by a KL constraint in Gaussian distribution, implying that $\theta_{pre}$ lies within a small vicinity of $\theta_{sft}$ . Therefore, a first-order Taylor expansion can be applied: + +$$ +\begin{array}{l} \mathcal {D} \left(\theta_ {\text {d y n}}\right) = \mathcal {D} \left(\theta_ {\text {p r e}} + \alpha \Delta \theta\right) \tag {4} \\ \approx \mathcal {D} (\theta_ {p r e}) + \alpha \langle \nabla \mathcal {D} (\theta_ {p r e}), \Delta \theta \rangle \\ \end{array} +$$ + +where $\Delta$ denotes the parameter update from $\theta_{sft}$ to $\theta_{pre}$ . The change in the motion degree is then given by: + +$$ +\begin{array}{l} \mathcal {D} \left(\theta_ {\text {d y n}}\right) - \mathcal {D} \left(\theta_ {\text {p r e}}\right) = \alpha \langle \nabla \mathcal {D} \left(\theta_ {\text {p r e}}\right), \Delta \theta \rangle \tag {5} \\ \approx \alpha \langle \nabla \mathcal {D} (\theta_ {p r e}), \nabla \mathcal {D} (\theta_ {p r e}) \rangle \\ \end{array} +$$ + +Assuming that the parameter update $\Delta \theta$ during finetuning is approximately aligned with the gradient of $\mathcal{D}$ [14, 16], we can express $\Delta \theta$ as: + +$$ +\Delta \theta \approx \gamma \nabla_ {\theta} \mathcal {D} \left(\theta_ {p r e}\right) \tag {6} +$$ + +Thus by substitution, we have: + +$$ +\begin{array}{l} \mathcal {D} \left(\theta_ {\text {d y n}}\right) - \mathcal {D} \left(\theta_ {\text {p r e}}\right) \approx \alpha \left\langle \nabla_ {\theta} \mathcal {D} \left(\theta_ {\text {p r e}}\right), \gamma \nabla_ {\theta} \mathcal {D} \left(\theta_ {\text {p r e}}\right) \right\rangle \\ = \alpha \gamma \left\langle \nabla_ {\theta} \mathcal {D} \left(\theta_ {p r e}\right), \nabla_ {\theta} \mathcal {D} \left(\theta_ {p r e}\right) \right\rangle \tag {7} \\ = \alpha \gamma \| \nabla_ {\theta} \mathcal {D} (\theta_ {p r e}) \| ^ {2} \geq 0. \\ \end{array} +$$ + +Since the norm of the gradient $\| \nabla_{\theta}\mathcal{D}(\theta_{pre})\| ^2$ is always non-negative, the dynamic motion degree metric $\mathcal{D}$ shall increases after the extrapolation, provided that the pretrained weight $\theta_{pre}$ is not at a local maximum of $\mathcal{D}$ . + +# 3.4. Decoupled Injection and Sampling + +While motion-controllability adaptation and training-free degree extrapolation can enhance motion control and diversity respectively, directly combining them poses a significant challenge due to their opposing optimization directions: fine-tuning for control versus reversed extrapolation for degree. To balance motion controllability and degree without compromising image consistency, we propose a decoupling strategy that separates degree-related and control-related parameters, applying them at distinct denoising stages. + +Isolating Parameter Sets. Here, we categorize the obtained knowledge into three parameter sets that support core abilities: motion control ( $\theta_{adt}$ ), motion degree ( $\theta_{deg}$ ), and video consistency ( $\theta_{con}$ ). We aim to integrate degree and control knowledge in the early denoising stage for long-term temporal planning, while emphasize consistency and + +control later for refining subject appearance. However, they are embedded within $\theta_{pre}$ and $\theta_{sft}$ . Therefore, we need parameter isolation to decouple three types of knowledge. + +We isolate these sets based on the following insight: (1) Based on the implicit gradient assumption (Section 3.3), $\theta_{sft} - \theta_{pre}$ captures the implicit control knowledge learned during the adaptation process, serves as $\theta_{adt}$ paired with $(\theta_{adapter})$ , incorporating elements like Q-Formers and projectors. (2) $\theta_{sft}$ exhibits degraded degree knowledge (11.67 vs. 68.54 for $\theta_{pre}$ in Table 1) and primarily contains control and consistency knowledge. Therefore, consistency knowledge $\theta_{con}$ can be directly derived as $\theta_{sft} - \theta_{adt}$ . (3) The extrapolated $\theta_{dyn}$ has similar control/consistency proportions as $\theta_{pre}$ , therefore $\theta_{deg} = \theta_{dyn} - \theta_{pre}$ can remove control/consistency components, isolating degree knowledge, even though it originates from $\theta_{sft}$ . (4) Fine-tuned parameters often contain redundant elements, which can cause conflicts during merging [21, 43]. To minimize overlap between different knowledge while preserving performance, we apply DARE pruning [43] at a pruning rate $p$ (e.g., $p = 70\%$ ). Distinct parameter groups are created using random masks $\mathcal{M}$ amplified from a Bernoulli distribution: + +$$ +\mathcal {M} _ {1} \sim \operatorname {B e r n o u l l i} (p), \quad \mathcal {M} _ {2} \sim \operatorname {B e r n o u l l i} (p), +$$ + +$$ +\theta_ {a d t} = \left\{\theta_ {a d p a t e r} \right\} \cup \left\{\mathrm {D A R E} _ {\mathcal {M} _ {2}} \left(\theta_ {s f t} - \theta_ {p r e}\right) \right\}, \tag {8} +$$ + +$$ +\theta_ {d e g} = \mathrm {D A R E} _ {\mathcal {M} _ {1}} \left(\theta_ {d y n} - \theta_ {p r e}\right), +$$ + +$$ +\theta_ {c o n} = \theta_ {s f t} - \theta_ {a d t}, +$$ + +Finally, Task-Arithmetic [16] is employed to inject control parameters $\theta_{adt}$ into both $\theta_{deg}$ and $\theta_{con}$ , deriving two enhanced models: the degree-enhanced model $\theta_{dyn}^{*}$ and the consistency-enhanced model $\theta_{con}^{*}$ . + +$$ +\theta_ {d y n} ^ {*} = \text {T a s k - A r i t h m e t i c} (\theta_ {d e g}, \theta_ {a d t}), \tag {9} +$$ + +$$ +\theta_ {c o n} ^ {*} = \text {T a k - A r i t h m e t i c} \left(\theta_ {c o n}, \theta_ {a d t}\right). +$$ + +Decoupled Video Sampling. During the denoising process for video generation, we typically observe that in the early stages, I2V models prioritize long-term temporal planning, including the motion of both subjects and textures, while later stages often focus more on refining subject appearance, which is mainly related to content consistency and overall image quality [37, 39]. Therefore, in the first $K$ steps of the denoising process, we adjust the video generator to make it more dynamic-enhanced and control-enhanced to increase motion controllability. In the later stages of the denoising process, we re-weight the motion-related modules, allowing the model to focus more on restoring the specific details of each frame of the subject, thereby generating high-consistency videos. Specifically, we define a generated model $\theta_{gen}$ by blending the degree-enhanced model $\theta_{dyn}^{*}$ and the consistency-merged model $\theta_{con}^{*}$ based on the denoising step $t$ from $T$ to 0: + +$$ +\theta_ {g e n} = \alpha_ {t} \theta_ {d y n} ^ {*} + (1 - \alpha_ {t}) \theta_ {c o n} ^ {*}, +$$ + +$$ +\alpha_ {t} = \left\{ \begin{array}{l l} 1 & \text {i f} t > T - K \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {10} +$$ + +Here, $T$ is the total number of denoising steps, and $K$ is a threshold that determines the switching point. + +# 4. Experiment + +# 4.1. Setup + +In all experiments, videos and images are resized to a $320 \times 512$ center-cropped format. During model training, we set the learning rate to $1 \times 10^{-5}$ , a batch size of 64, and trained for 100K steps on 8 A100 GPUs. For inference, we use DDIM [34] with 50 steps and a classifier-free guidance scale of 7.5, generating 16 frames for a 2-second video. + +Datasets. We use the WebVid-2M dataset [2] for supervised fine-tuning, which contains two million video-text pairs from stock footage websites. For motion controllability annotation, we use CoTracker [18] to annotate motion information on the WebVid-2M dataset. + +Evaluation. The evaluation benchmarks include VBench [27] and UCF101 [35]. VBench is a comprehensive benchmark for video generation, incorporating recent models and extensive evaluation metrics. UCF101 contains 101 human action categories with diverse background and lighting variations. For VBench, following previous work [20, 46], we utilize the official dataset to assess the model across several metrics: I2V-Subject Consistency (I2V Sub.), I2V-Background Consistency (I2V Back.), video quality, camera motion, and dynamic degree. The video quality evaluation considers dimensions including Subject Consistency, Background Consistency, Smoothness, Aesthetic Score, Imaging Quality, and Temporal Flickering. The Avg. score is calculated as $\frac{1}{10} * (\text{I2V Sub.} + \text{I2V Back.} + 6 * \text{Video Quality} + \text{Motion Degree} + \text{Motion Control})$ , which is consistent with VBench I2V original setting. For UCF101, similar to [39, 40], we use Fréchet Video Distance (FVD) [36] and Inception Score (IS) [30] to assess video quality. + +# 4.2. Quantitative Evaluation + +VBenchmark Results Analysis. We compare our method with two types of approaches on VBenchmark. The first kind of comparing methods include those requiring large-scale video pretraining, such as animate-Anything [7], SVD [3], and DynamiCrafter [40]. The other comparison methods are finetuning/inference strategies based on DynamiCrafter, including naive finetuning with the same training setup as our method (Naive FT) and Conditional Image Leakage (CIL) methods [46] (CIL Infer and CIL FT). Notably, our method performs best in Motion Degree and Motion Control across all baselines, even including large-scale pretrained methods, + +Table 1. Quantitative results on the VBench I2V benchmark for various I2V-DM pretrained baselines and finetuning/inference techniques, including our methods based on DynamiCrafter. "PT." in the table denotes whether the method requires video pretraining, and "FT." denotes whether the method requires supervised finetuning on WebVid-2M. The global best and DynamiCrafter-based best performing metrics are highlighted in bold and underline, respectively. + +
ModelAvg.PT.FT.I2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
ConsistI2V (Ren et al.)75.88✓✗95.8295.9585.7418.6233.92
SEINE (Chen et al.)76.56✓✗96.5796.8085.7134.3123.67
Animate-Anything (Dai et al.)74.61✓✗98.7698.5888.842.6813.08
VideoCrafter (Chen et al.)76.96✓✗91.1796.9487.5522.6033.60
SVD-XT-1.0 (Blattmann et al.)-✓✗97.5297.6386.5452.36-
SVD-XT-1.1 (Blattmann et al.)-✓✗97.5197.6286.6643.17-
I2VGen-XL (Zhang et al.)76.00✓✗96.4896.8387.0226.1018.48
DynamiCrafter (Xing et al.)79.93✓✗94.9195.3084.9568.5430.85
DynamiCrafter-Naive FT72.91✗✓97.9298.7083.5211.6717.82
DynamiCrafter-CIL Infer (Zhao et al.)80.47✗✗96.9693.3485.2869.4331.29
DynamiCrafter-CIL FT (Zhao et al.)77.55✗✓96.9097.7987.1025.6930.50
DynamiCrafter-Ours84.25✗✓97.0396.6286.2287.6443.87
DynamiCrafter-Ours w/ CIL Infer84.81✗✓97.1195.4587.1089.2143.61
+ +Table 2. Quantitative evaluation of the DynamiCrafter-based method on the UCF-101 benchmark. + +
ModelFVD↓IS↑Motion Degree↑
DynamiCrafter288.0113.3168.17
DynamiCrafter-Naive FT373.7713.1947.66
DynamiCrafter-CIL FT409.6913.6853.78
DynamiCrafter-Ours324.7913.7177.93
+ +as shown in Table 1. We observe that Naive FT causes a significant degradation in Motion Degree (11.67 vs 68.54), consistent with previous findings [46]. Naive FT achieves the highest I2V Subject and Background Consistency (I2V Sub. and I2V Back.) among the DynamiCrafter-based methods because lower motion degree results in higher consistency scores. Interestingly, motion control is also negatively affected in Naive FT (17.82 vs 11.54). We speculate that this is due to the insufficient motion degree, which hinders the model's ability to follow control instructions effectively. The CIL method can mitigate these effects, either by employing the inference strategy on the pre-trained model (CIL Infer) or by repairing the Naive FT (CIL FT). However, its enhancements in Motion Control and Motion Degree are limited. In contrast, our method, combining adaptation and extrapolation, significantly improves both motion degree and motion control (11.67 vs 87.64) while preserving similar video quality (87.10). Furthermore, our method is orthogonal to the CIL Infer method, and integrating it with our method further boosts performance (Avg. 84.25 vs 84.81). + +UCF101 Results Analysis. In the UCF101 dataset, our method still exhibits superior motion degree ability despite + +![](images/3ed9ba45f26921b99ce254f23e4ff288dfa41ad75080ac5b285a73d28723d812.jpg) + +![](images/3f40f575d7aa348e86614ede1c24599cbc04d2273e70f480d76c5ef2f277c940.jpg) + +![](images/b4d002a3e35f155dc61373a026e10214b25d26439c895465d0c823c2afe78ec0.jpg) +Figure 3. Human Evaluation of our method Compared to SVD, VideoCrafter (VC), DynamiCrafter (DC), and its variant of Native Fine-Tuned (DC-FT) and CIL-Based Methods. + +![](images/9ea3af99211531cb8a6bce2d1bae0298866fcd76b618e0e58c85ae32e0145e95.jpg) + +the low resolution and poor image quality. Both Naive FT and CIL FT suffer from degradation in Fréchet Video Distance (FVD), as shown in Table 2. Unlike the performance on VBenchmark, CIL performs worse than Naive FT on UCF101, likely due to the added noise schedule disrupting the already noisy and low-resolution images. Our method, however, is more stable in FVD and outperforms in Inception Score (IS) and Motion Degree. + +User Study. As shown in Figure 3, our method excels in Motion Degree and Motion Control while remaining competitive in video quality and I2V consistency. The results are consistent with the above findings, supporting that our strategies enhance both video motion dynamics and control + +![](images/70134cf65ce9b4e42969d2f7e8755b2e5dbb3e5ab5e5f336e9dded2823ba7d88.jpg) +Figure 4. In the visual comparisons, we generated 16-frame videos using DynamiCrafter, CIL, Consistent-I2V, and our model. We selected 3 to 4 frames at regular intervals for display at the top. The upper section visualizes the overall aesthetic quality, while the lower section uses Optical Flow to illustrate the dynamic qualities of the generated videos for each method. + +while preserving quality and consistency, yielding superior overall performance. + +# 4.3. Qualitative Evaluation + +To empirically evaluate the visual quality of our method, we compare it with several state-of-the-art image-to-video generation approaches, including DynamiCrafter [40], CIL [46], and ConsistI2V [27]. In Figure 4, we observe that our method generates temporally coherent videos with a sufficient range of motion that closely adhere to the input text conditions. In contrast, DynamiCrafter exhibits inconsistencies in text control, likely due to challenges in capturing the semantic content of the input image. The ConsistI2V method produces videos that remain static over time, with + +the color points on the figures almost unchanged, indicating a lack of motion dynamics and control. While CIL generates videos that semantically resemble the input images and have enhanced motion compared to ConsistI2V, its improvements are limited. CIL does not sufficiently vary the content dynamically to match the text prompts, resulting in limited motion and control. Our method, however, preserves image consistency while achieving a large motion degree and exhibiting good controllability. + +# 4.4. Ablation Study + +We demonstrate the importance of our extrapolation and decoupling methods in Table 3: (1) On top of our adaptation training, using our extrapolation and decoupling techniques + +Table 3. Ablation study of the three stages of our framework on VBenchmark. + +
(a) Adaptation(b)Extrapolation(c) DecoupleI2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
✓XX97.5196.3986.6122.8549.31
X✓X94.1495.6283.5298.2117.63
XX✓94.6096.2183.9569.6132.37
✓✓X91.1192.8684.3271.6826.71
X✓✓96.0194.6084.3493.3321.85
✓✓✓97.0396.6286.2287.6443.87
+ +significantly improves the motion degree, from 22.85 to 87.64. (2) Using extrapolation alone can drastically improve the motion degree (69.61 vs 93.33), but it comes at the cost of degradation in I2V-Subject Consistency (I2V Sub.) and I2V-Background Consistency (I2V Back.), as well as motion control. This is because reversing the training process of the pre-trained model to the SFT model unlearns the camera motion ability, causing significant interference if directly merged [16]. (3) Using the decoupling technique alone can improve the motion degree while preserving other performance metrics. This indicates that the I2V model indeed exhibits functional differentiation during the sampling process. (4) Adding adaptation, though it increases training costs, greatly enhances the Motion Control aspect of the model. The degradation in motion degree indicates that direct supervision is suboptimal and can lead to minor motion dynamics, a finding consistent with previous research [46]. (5) When paired with the decoupling technique, we achieve optimal results, integrating high motion degree and control ability while preserving consistency and video quality. + +![](images/8487239c8707ead2a78c9aa4c4b52f7f9bc17c48f0322fb265445bf9289bceff.jpg) + +![](images/c7faf79a98499b80709c802e758f7d607147326f4cef447eff43e28345feebab.jpg) + +![](images/325f4dc6793b1cab5af21ad17c16945fe8ea437377c277c73802937aa093950c.jpg) +(a) Effect of extrapolate strength $\alpha$ + +![](images/492b8ccea4fb5c586769897af4ed4a12542d7dca40672b3c8b585e66d86a7a42.jpg) + +![](images/f5c7037dfa3ced0fdb6a8d959d5372e7a35d4e0435b4499d7cd63ea6dc51f1ea.jpg) +(b) Effect of different switching time and model choices. + +![](images/9a8bff6232c301f3098ca966ef3cb1baf2a2f925d0d41aa64555c1cb374d75af.jpg) +Figure 5. The impact of different extrapolate strengths and different sampling strategies. + +# 4.5. Effect of Extrapolate Strength + +In our extrapolation experiments, we analyzed the impact of different $\alpha$ values on performance using Equation 3. As shown in Figure 5a, as $\alpha$ increases, the motion degree first rapidly rises within the 0-0.7 range, peaks, and then gradually declines between 1.2-2.0. Simultaneously, quality and consistency metrics decline. This indicates that extrapolation enhances motion degree but sacrifices quality and consistency. This aligns with the findings in Section 4.4 and underscores the importance of appropriate decoupling. + +# 4.6. Analysis on Decoupled Sampling Method + +To study the impact of switch model time and choice, we conducted experiments illustrated in Figure 5b. We outline two model switch strategies during the denoising process from time $T$ to 0: (1) Initially using $\theta_{dyn}^{*}$ and then switching to $\theta_{sft}^{*}$ . (2) Starting with $\theta_{sft}^{*}$ and then transitioning to $\theta_{dyn}^{*}$ . The performance trends are shown in Figure 5b. The first strategy, which initially uses $\theta_{dyn}^{*}$ , results in higher motion degree with a later switch (smaller $T$ ), but with poorer quality. Conversely, an earlier switch (larger $T$ ) reduces the degree. The second strategy, introducing $\theta_{dyn}^{*}$ later, has a less pronounced impact on the degree. This aligns with our hypothesis in Section 3.4, which suggests that the early sampling phase significantly influences the degree. + +# 5. Conclusion + +In this paper, we introduce a novel Extrapolating and Decoupling framework for image-to-video generation. Our framework consists of three stages and can effectively model the motion control and motion degree. In the first stage, we explicitly inject the textual condition into I2V-DM to improve motion controllability. Then, we devise a training-free extrapolation strategy to boost the motion degree. Finally, we decouple the motion-relevant parameters and dynamically inject them into the I2V-DM along different de-noising steps. With our framework, we can generate videos with both natural motion and temporal consistency. To the best of our knowledge, we are the first to apply merging techniques in video generation. + +# References + +[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2 +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 4, 5 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 6 +[4] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 6 +[5] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 6 +[6] Feng Cheng, Ziyang Wang, Yi-Lin Sung, Yan-Bo Lin, Mohit Bansal, and Gedas Bertasius. Dam: Dynamic adapter merging for continual video qa learning, 2024. 2 +[7] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Finegrained open domain image animation with motion guidance, 2023. 5, 6 +[8] Felix Draxler, Kambis Veschgini, Manfred Salmhofer, and Fred Hamprecht. Essentially no barriers in neural network energy landscape. In International conference on machine learning, pages 1309-1318. PMLR, 2018. 2 +[9] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. +[10] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2 +[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1 +[12] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2 +[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and + +Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 3 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[16] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 2, 4, 5, 8 +[17] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2022. 2 +[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together, 2024. 5 +[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 4 +[20] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 5 +[21] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 2, 5 +[22] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 2022. 2 +[23] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 2 +[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[25] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[27] Weiming Ren, Harry Yang, Ge Zhang, Cong Wei, Xinrun Du, Stephen Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation. arXiv preprint arXiv:2402.04324, 2024. 1, 5, 6, 7 + +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[30] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping, 2017. 5 +[31] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[32] Mustafa Shukor, Corentin Dancette, Alexandre Rame, and Matthieu Cord. Unival: Unified model for image, video, audio and language tasks, 2023. 2 +[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 5 +[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11):1-7, 2012. 5 +[36] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation, 2019. 5 +[37] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation, 2023. 5 +[38] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2 +[39] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 5 +[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 1, 2, 3, 5, 6, 7 +[41] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference + +when merging models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[42] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. +[43] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*. 2, 5 +[44] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 6 +[45] Yiming Zhang, Zhening Xing, Yanhong Zeng, Youqing Fang, and Kai Chen. Pia: Your personalized image animator via plug-and-play modules in text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7747-7756, 2024. 1 +[46] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. arXiv preprint arXiv:2406.15735, 2024. 1, 2, 4, 5, 6, 7, 8 +[47] Chujie Zheng, Ziqi Wang, Heng Ji, Minlie Huang, and Nanyun Peng. Weak-to-strong extrapolation expedites alignment, 2024. 2 \ No newline at end of file diff --git a/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/images.zip b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d949f31850aa6a6825e9a6e3380e1c3aaf12737 --- /dev/null +++ b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:221978aac1b8b1991bc164b29d94f4b6216cbc90b34befff7db48f6f031b6699 +size 828665 diff --git a/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/layout.json b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b84703a3fb94bb6d88abbc88ce9d7a51320d4010 --- /dev/null +++ b/2025/Extrapolating and Decoupling Image-to-Video Generation Models_ Motion Modeling is Easier Than You Think/layout.json @@ -0,0 +1,8606 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 511, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 511, + 140 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 511, + 140 + ], + "type": "text", + "content": "Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": "Jie Tian" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger*}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": " Xiaoye Qu" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": " Zhenyi Lu" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": " Wei Wei" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": " Sichen Liu" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": " Yu Cheng" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": "School of Computer Science & Technology, Huazhong University of Science and Technology, \n" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 159, + 534, + 205 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 243, + 200, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 243, + 200, + 255 + ], + "spans": [ + { + "bbox": [ + 152, + 243, + 200, + 255 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 269, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 616 + ], + "type": "text", + "content": "Image-to-Video (I2V) generation aims to synthesize a video clip according to a given image and condition (e.g., text). The key challenge of this task lies in simultaneously generating natural motions while preserving the original appearance of the images. However, current I2V diffusion models (I2V-DMs) often produce videos with limited motion degrees or exhibit uncontrollable motion that conflicts with the textual condition. To address these limitations, we propose a novel Extrapolating and Decoupling framework, which introduces model merging techniques to the I2V domain for the first time. Specifically, our framework consists of three separate stages: (1) Starting with a base I2V-DM, we explicitly inject the textual condition into the temporal module using a lightweight, learnable adapter and fine-tune the integrated model to improve motion controllability. (2) We introduce a training-free extrapolation strategy to amplify the dynamic range of the motion, effectively reversing the fine-tuning process to enhance the motion degree significantly. (3) With the above two-stage models excelling in motion controllability and degree, we decouple the relevant parameters associated with each type of motion ability and inject them into the base I2V-DM. Since the I2V-DM handles different levels of motion controllability and dynamics at various denoising time steps, we adjust the motion-aware parameters accordingly over time. Extensive qualitative and quantitative experiments have been conducted to demonstrate the superiority of our framework over existing methods. Code is available at https://github.com/Chuge0335/EDG" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "type": "text", + "content": "Recently, Image-to-Video generation (I2V) [11, 23, 40, 45, 46] has attracted increasing attention. Given a single im" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 237, + 542, + 354 + ], + "blocks": [ + { + "bbox": [ + 328, + 237, + 542, + 354 + ], + "lines": [ + { + "bbox": [ + 328, + 237, + 542, + 354 + ], + "spans": [ + { + "bbox": [ + 328, + 237, + 542, + 354 + ], + "type": "image", + "image_path": "e0e69f353bba46926400601b34edbf1fed709f97bd8d08924a1a3b7bcdfa5426.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 356, + 542, + 458 + ], + "blocks": [ + { + "bbox": [ + 329, + 356, + 542, + 458 + ], + "lines": [ + { + "bbox": [ + 329, + 356, + 542, + 458 + ], + "spans": [ + { + "bbox": [ + 329, + 356, + 542, + 458 + ], + "type": "image", + "image_path": "45439e2b85fa9d023fe70d258dd5ed8df22883f215f6341db0c20d45c41c8b38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 474, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 519 + ], + "type": "text", + "content": "Figure 1. In the first example, Consist2V [27] yields a video with limited motion. In the second, DynamiCrafter [40] fails to follow the text condition and lacks controllability. Our method, however, achieves improved motion degree and better controllability." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 531, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 628 + ], + "type": "text", + "content": "age and a text condition, it aims to convert the still image into videos that conform to text instruction while preserving the original appearance of the images. This paper focuses on open-domain images and supports free text as guidance, rather than limited class labels. This conditioned I2V task is substantially challenging as it struggles with simultaneously maintaining appearance consistency and producing natural motions in the generated frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Unlike image synthesis, motion plays a significantly crucial role in image-to-video generation. However, most existing works lack effective motion modeling when generating videos. For instance, ConsistI2V [27] enhances the first frame conditioning through spatiotemporal attention, leading to limited expression ability for modeling motion degree as the model tends to over-rely on the conditional im" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " Equal contribution; * Work done during an internship at Shanghai AI Laboratory; " + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " corresponding authors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12512" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "age. DynamiCrafter [40] interacts the text embedding and conditional image embedding with the U-Net spatial layers through a dual cross-attention, neglecting the temporal property of the motion and suffering from motion controllability. In this way, the generated videos either contradict the motion described in the text or are still image sequences with slow change. As shown in Figure 1, the videos generated by the previous model show cars that barely move, or the videos of the castle contradict the textual descriptions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 295, + 468 + ], + "type": "text", + "content": "To address the aforementioned issues, we present a novel Extrapolating and Decoupling framework to effectively model motion for image-to-video generation, improving both motion degree and motion controllability. Specifically, it consists of three stages. Firstly, we improve motion controllability by explicitly injecting the textual condition into the I2V diffusion model (I2V-DM). A lightweight learnable adapter is introduced to enhance the temporal attention module. Fine-tuning the integrated model improves motion controllability without altering the original I2V-DM architecture. Secondly, we observe degraded motion dynamics in our fine-tuned model, a phenomenon has also been observed by Zhao et al. [46]. To boost the motion degree without extra training, we introduce a novel training-free extrapolation strategy to restore and enhance motion degree dynamics by reversing the fine-tuning process. Thirdly, we decouple the parameters contributing to motion controllability and dynamics from the previous two stages, selectively integrating them into the final model. Recognizing that different denoising steps focus on varying motion ability and temporal consistency, we dynamically adjust the motion-aware parameters over time. With these three stages, we can generate videos with both natural motion and temporal consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 469, + 294, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 294, + 554 + ], + "type": "text", + "content": "To verify the effectiveness of our framework, we perform experiments on the VBench I2V benchmark and UCF101 dataset. Extensive qualitative and quantitative experiments demonstrate the notable superiority of our framework over existing I2V methods. Furthermore, we offer discussion and analysis of some insightful designs in our framework. To sum up, our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 558, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 616 + ], + "type": "text", + "content": "- We propose a novel Extrapolating and Decoupling framework for video generation, which models motion controllability and degree dynamics. To best of our knowledge, this is the first use of model merging techniques in the I2V domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "content": "- We introduce a lightweight adapter to enhance motion controllability while preserving the base I2V-DM structure and propose a novel training-free extrapolation strategy to amplify motion degree." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 701 + ], + "type": "text", + "content": "- We make the first attempt to decouple parameters related to motion controllability and degree, dynamically injecting them into the I2V-DM across varying denoising steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "content": "- We conduct extensive experiments on multiple bench" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "content": "marks, where our proposed framework consistently outperforms previous state-of-the-art methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 119, + 400, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 119, + 400, + 132 + ], + "spans": [ + { + "bbox": [ + 314, + 119, + 400, + 132 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 143, + 506, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 143, + 506, + 156 + ], + "type": "text", + "content": "2.1. Motion Modeling in I2V Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 165, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 555, + 416 + ], + "type": "text", + "content": "Diffusion models (DMs) [15, 33] have recently demonstrated remarkable generative capabilities in image and video generation [12, 24, 26, 28, 29]. In this paper, we focus on the image-conditioned video (I2V) generation task, where motion modeling aims to predict object movements. Traditionally, optical flow estimates the displacement field between consecutive frames. To effectively model motion in I2V tasks, LFDM [23] proposes a latent flow diffusion model that synthesizes a latent optical flow sequence to warp the given image in latent space for generating new videos. Motion-I2V [31] factorizes I2V into two stages: a diffusion-based motion field predictor for pixel trajectories and motion-augmented temporal attention to enhance 1-D temporal attention in video latent diffusion models. Recent works [38, 40, 44] enable diffusion models to handle motion modeling and video generation simultaneously. For instance, Dynamiccafter [40] uses a dual-stream image injection mechanism to leverage the motion prior of text-to-video diffusion models. In this paper, we build on DynamamiCrafter to design an Extrapolating and Decoupling framework to improve both motion controllability and degree." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 436, + 410, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 410, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 410, + 449 + ], + "type": "text", + "content": "2.2. Model Merging" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "content": "Model merging, widely explored in natural language processing and image classification, typically combines multiple fine-tuned task-specific models into a unified multitask model without additional training. Early methods [1, 8-10, 17, 22] leverage interpolation techniques to integrate multi-task knowledge effectively. Task-Arithmetic [16] extends these approaches by enabling complex arithmetic operations in weight space, offering finer control over the merging process. Recent studies [21, 25, 41-43] have further reduced interference during merging. Beyond interpolation-based methods, recent works have shown that sparsifying [43] or amplifying delta weights [47] can enhance model knowledge in merging. Additionally, merging techniques have been introduced to multimodal understanding such as question-answering [6, 32]. However, the video generation domain has remained unexplored until now." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "content": "We pioneer the introduction of model merging techniques into the video generation domain. Specifically, we leverage model merging to extrapolate motion degree knowledge and extract fine-grained decoupled knowledge, which are selectively merged at different diffusion stages." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12513" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 303, + 180 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 303, + 180 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 303, + 180 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 303, + 180 + ], + "type": "image", + "image_path": "5b10612b13869a7de2d99ae36aa38144f6812edc4bbb5e7186a9d9cd542b3aac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 541, + 180 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 541, + 180 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 541, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 541, + 180 + ], + "type": "image", + "image_path": "fa1dd35b126828d0a192f07f99ee42b3f22f1e9fbde6ed7c55e873502a93321d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 67, + 182, + 468, + 302 + ], + "blocks": [ + { + "bbox": [ + 67, + 182, + 468, + 302 + ], + "lines": [ + { + "bbox": [ + 67, + 182, + 468, + 302 + ], + "spans": [ + { + "bbox": [ + 67, + 182, + 468, + 302 + ], + "type": "image", + "image_path": "f5b6642b6e4e59b307915a7e2d7dcbfbcbd330e1a77e6a7de9d29f44ad5e0eb3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 312, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 54, + 312, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 555, + 369 + ], + "type": "text", + "content": "Figure 2. The overview of our framework, consisting of three stages. (a) We improve the motion controllability by injecting the textual conditions into the temporal attention module in I2V-DM model and then fine-tune the integrated model. (b) The I2V-DM model suffers from a decrease in motion degree after fine-tuning. To mitigate it, we propose a training-free extrapolation to boost the motion degree by reversing the fine-tuning progress. (c) We decouple the relevant parameters contributing to motion controllability and motion degree from the last two-stage models. Moreover, we selectively inject these parameters into the I2V model along with the de-noising process." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 471, + 186, + 541, + 302 + ], + "blocks": [ + { + "bbox": [ + 471, + 186, + 541, + 302 + ], + "lines": [ + { + "bbox": [ + 471, + 186, + 541, + 302 + ], + "spans": [ + { + "bbox": [ + 471, + 186, + 541, + 302 + ], + "type": "image", + "image_path": "4cb4cb9d813853827865359c69548fb1b1782960468e3f12906af4db2c7b7d84.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 111, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 111, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 111, + 391 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": "Given a reference image " + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": ", image-to-video generation (I2V) targets at generating a sequence of subsequent video frames " + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "inline_equation", + "content": "V = \\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_N\\}" + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": ". The key objective of I2V generation is to ensure that the generated video not only exhibits plausible motion but also faithfully preserves the visual appearance of the provided reference image. The overall framework of our method, shown in Figure 2, consists of three stages. First, we improve motion controllability guided by the text condition. Second, we propose a training-free extrapolation strategy to enhance motion dynamics. Finally, we decouple parameters related to motion controllability and dynamics, allowing flexible injection during the denoising process to generate videos with natural motion and temporal consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 286, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 286, + 588 + ], + "type": "text", + "content": "3.1. Preliminary: Latent Video Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Before introducing our framework, we first review the Latent Diffusion Models (LDMs) [28]. Specifically, it extends diffusion models [15] by operating in a more compact latent space rather than directly on pixels, reducing computational complexity. In LDMs, an input " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " is first encoded into a latent representation " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " using a pretrained Variational Autoencoder (VAE). This latent variable " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " is then corrupted gradually over " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " steps by adding Gaussian noise, producing a sequence of noisy latents " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{z_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": ". The noise process follows " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_t \\mid z_0 \\sim \\mathcal{N}(z_t; \\sqrt{\\bar{\\alpha}_t} z_0, (1 - \\bar{\\alpha}_t)I)" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}" + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "content": " controls the noise schedule. To recover the original latent, LDMs learn a reverse process that denoises " + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "content": " through a sequence of steps, guided by the objective:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 427, + 553, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 427, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 375, + 427, + 553, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {z _ {0}, t, \\epsilon} \\| \\epsilon_ {\\theta} (z _ {t}, t) - \\epsilon \\| ^ {2}, \\tag {1}", + "image_path": "b4753b7a5c1dcddd5f3f248698cd83beee237b35489505821ec7c963df622c2a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " represents Gaussian noise, " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " is a randomly sampled timestep, and " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t, t)" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " is the learned noise estimator parameterized by " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": "In this paper, we adopt an open-source video LDM, DynamiCrafter [40] as the base model. Given a video " + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{L \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": ", we first encode it into a latent representation " + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{L \\times C \\times h \\times w}" + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": " is derived on a frame-by-frame basis. Both the forward diffusion process " + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "inline_equation", + "content": "z_{t} = p(z_{0}, t)" + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": " and the backward denoising process " + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "inline_equation", + "content": "z_{t} = p_{\\theta}(z_{t-1}, c, t)" + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": " are conducted in this latent space, with " + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": " representing potential denoising conditions such as text prompts. Ultimately, the generated videos are obtained through the decoder " + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\mathcal{D}(z)" + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 612, + 500, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 500, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 500, + 625 + ], + "type": "text", + "content": "3.2. Motion-Controllability Adaptation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "In I2V generation, motion controllability is important for following textual conditions. To model motion, current I2VDMs [40] typically employ spatial attention to share information within a single frame and temporal attention to capture time sequence consistency [13]. However, methods like DynamiCrafter [40] only use text condition embedding in the spatial attention module, thus it does not share condi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12514" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "tional information across different time sequences, limiting its motion controllability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": "To alleviate this problem, we explicitly inject textual conditions into DynamiCrafter's temporal attention with a lightweight adapter, as depicted in Figure 2(a). The spatial attention module remains unmodified as it already incorporates textual information. The adapter is non-intrusive and preserves the original model structure. Specifically, we extracts spatial features " + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "inline_equation", + "content": "z_{t} \\in \\mathbb{R}^{f \\times d \\times w \\times h}" + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": " and the text embeddings " + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{embd}} \\in \\mathbb{R}^{l \\times d}" + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": " from the spatial attention layer and the CLIP text encoder respectively, where " + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": " is the number of frames, " + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": " is the feature dimension, and " + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "inline_equation", + "content": "w, h" + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": " are spatial dimensions. Then, the condition information in text embeddings are further compressed by a trainable Q-Former [19] by retaining the motion-relevant tokens:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 260, + 295, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 260, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 295, + 276 + ], + "type": "interline_equation", + "content": "t _ {\\text {e m b d}} ^ {\\prime} = \\mathrm {Q} - \\operatorname {F o r m e r} \\left(t _ {\\text {e m b d}}\\right) \\in \\mathbb {R} ^ {f \\times d} \\tag {2}", + "image_path": "0b3f2a0be143e822d3042a69e28fb25ba51947063464839d42813d812a222df1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{embd}}^{\\prime}" + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "content": " is then concatenated with spatial information " + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "content": " along the feature dimension, projected through " + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{proj}}" + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "content": " to reduce the dimension to " + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "content": ", and passed through temporal attention to inject the text control information into each frame. We finetune the model on WebVid-2M [2], updating only the adapter (the Q-Former and the linear projection " + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{proj}}" + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "content": ") and temporal attention weights, ensuring minimal added parameters and efficient training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 217, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 217, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 217, + 401 + ], + "type": "text", + "content": "3.3. Motion-Degree Extrapolation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": "After fine-tuning I2V-DM DynamiCrafter with motion-controllability adaptation, we observe a reduction in motion degree, a common issue attributed to conditional image leakage [46]. During fine-tuning, Dynamicrafter concatenates the first frame with noise, which may cause the model to over-rely on the conditional image and neglect noisy inputs, especially at early timesteps. While adding time-dependent noise partially mitigates motion loss [46], it requires retraining and has limited effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": "To improve the motion dynamics beyond the pre-trained model, we introduce a novel training-free and model-agnostic extrapolation approach, as shown in Figure 2 (b), which extremely pushes the \"unlearning\" process of motion controllability. In this way, we can achieve a model with amplified motion degree dynamics. Specifically, we estimate a degree-enhanced model, " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": ", by extrapolating between a pre-trained model " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": " and its fine-tuned version " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\theta_{sft}" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": " with motion-controllability adaptation, amplifying their motion-dynamic differences:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 643, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 643, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 114, + 643, + 295, + 657 + ], + "type": "interline_equation", + "content": "\\theta_ {d y n} = \\theta_ {p r e} + \\alpha \\left(\\theta_ {p r e} - \\theta_ {s f t}\\right) \\tag {3}", + "image_path": "73ec7c3ab067dfdf6c148c3093210ff56813062fdce997276dae65af29c35b8f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Where the hyperparameter " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " controls the strength of this extrapolation. In this equation, we compute the difference of the model parameters except for the adapter introduced in Section 3.2. This process requires only two checkpoints and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "obtaining the motion degree-enhanced model without training. To help further understand our extrapolating method, we provide a theoretical analysis below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "text", + "content": "Theoretical Analysis. Assume that we have a motion degree metric score denoted as " + }, + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "text", + "content": ". Diffusion models are commonly regularized by a KL constraint in Gaussian distribution, implying that " + }, + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "text", + "content": " lies within a small vicinity of " + }, + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\theta_{sft}" + }, + { + "bbox": [ + 313, + 122, + 555, + 183 + ], + "type": "text", + "content": ". Therefore, a first-order Taylor expansion can be applied:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 191, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 191, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 348, + 191, + 553, + 220 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) = \\mathcal {D} \\left(\\theta_ {\\text {p r e}} + \\alpha \\Delta \\theta\\right) \\tag {4} \\\\ \\approx \\mathcal {D} (\\theta_ {p r e}) + \\alpha \\langle \\nabla \\mathcal {D} (\\theta_ {p r e}), \\Delta \\theta \\rangle \\\\ \\end{array}", + "image_path": "146f4139a3e9a2b8543fb14fa91ac52bf91f6c5c798fbae577fb71f0f96d19ca.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "content": " denotes the parameter update from " + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "inline_equation", + "content": "\\theta_{sft}" + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "content": ". The change in the motion degree is then given by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 256, + 553, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 256, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 329, + 256, + 553, + 285 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) - \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) = \\alpha \\langle \\nabla \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right), \\Delta \\theta \\rangle \\tag {5} \\\\ \\approx \\alpha \\langle \\nabla \\mathcal {D} (\\theta_ {p r e}), \\nabla \\mathcal {D} (\\theta_ {p r e}) \\rangle \\\\ \\end{array}", + "image_path": "993241de3d7292cea5e5d36e2b9e990d4164c2fcbbdc4b48b5fb4c3b388eed5d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "content": "Assuming that the parameter update " + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "content": " during finetuning is approximately aligned with the gradient of " + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "content": " [14, 16], we can express " + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 394, + 340, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 340, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 394, + 340, + 553, + 354 + ], + "type": "interline_equation", + "content": "\\Delta \\theta \\approx \\gamma \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right) \\tag {6}", + "image_path": "95c1759d2d38cdcc9d652285eb84c136e5c3ef02fac490f973d8dd37dbff2e51.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 358, + 440, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 358, + 440, + 370 + ], + "spans": [ + { + "bbox": [ + 314, + 358, + 440, + 370 + ], + "type": "text", + "content": "Thus by substitution, we have:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 378, + 553, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 378, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 321, + 378, + 553, + 424 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} \\left(\\theta_ {\\text {d y n}}\\right) - \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) \\approx \\alpha \\left\\langle \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right), \\gamma \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {\\text {p r e}}\\right) \\right\\rangle \\\\ = \\alpha \\gamma \\left\\langle \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right), \\nabla_ {\\theta} \\mathcal {D} \\left(\\theta_ {p r e}\\right) \\right\\rangle \\tag {7} \\\\ = \\alpha \\gamma \\| \\nabla_ {\\theta} \\mathcal {D} (\\theta_ {p r e}) \\| ^ {2} \\geq 0. \\\\ \\end{array}", + "image_path": "190f40608b5b0c8d4da9c5574bd9313e587c3f4592fdef24b96e39f9e6111afb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "content": "Since the norm of the gradient " + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\| \\nabla_{\\theta}\\mathcal{D}(\\theta_{pre})\\| ^2" + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "content": " is always non-negative, the dynamic motion degree metric " + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "content": " shall increases after the extrapolation, provided that the pretrained weight " + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "content": " is not at a local maximum of " + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 490, + 499, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 499, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 499, + 503 + ], + "type": "text", + "content": "3.4. Decoupled Injection and Sampling" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 508, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 628 + ], + "type": "text", + "content": "While motion-controllability adaptation and training-free degree extrapolation can enhance motion control and diversity respectively, directly combining them poses a significant challenge due to their opposing optimization directions: fine-tuning for control versus reversed extrapolation for degree. To balance motion controllability and degree without compromising image consistency, we propose a decoupling strategy that separates degree-related and control-related parameters, applying them at distinct denoising stages." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Isolating Parameter Sets. Here, we categorize the obtained knowledge into three parameter sets that support core abilities: motion control (" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{adt}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "), motion degree (" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{deg}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "), and video consistency (" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{con}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "). We aim to integrate degree and control knowledge in the early denoising stage for long-term temporal planning, while emphasize consistency and" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12515" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "control later for refining subject appearance. However, they are embedded within " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\theta_{sft}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ". Therefore, we need parameter isolation to decouple three types of knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": "We isolate these sets based on the following insight: (1) Based on the implicit gradient assumption (Section 3.3), " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{sft} - \\theta_{pre}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " captures the implicit control knowledge learned during the adaptation process, serves as " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{adt}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " paired with " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "(\\theta_{adapter})" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": ", incorporating elements like Q-Formers and projectors. (2) " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{sft}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " exhibits degraded degree knowledge (11.67 vs. 68.54 for " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " in Table 1) and primarily contains control and consistency knowledge. Therefore, consistency knowledge " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{con}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " can be directly derived as " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{sft} - \\theta_{adt}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": ". (3) The extrapolated " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " has similar control/consistency proportions as " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": ", therefore " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{deg} = \\theta_{dyn} - \\theta_{pre}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " can remove control/consistency components, isolating degree knowledge, even though it originates from " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\theta_{sft}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": ". (4) Fine-tuned parameters often contain redundant elements, which can cause conflicts during merging [21, 43]. To minimize overlap between different knowledge while preserving performance, we apply DARE pruning [43] at a pruning rate " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "p = 70\\%" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": "). Distinct parameter groups are created using random masks " + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 108, + 296, + 335 + ], + "type": "text", + "content": " amplified from a Bernoulli distribution:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 344, + 258, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 344, + 258, + 357 + ], + "spans": [ + { + "bbox": [ + 80, + 344, + 258, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {1} \\sim \\operatorname {B e r n o u l l i} (p), \\quad \\mathcal {M} _ {2} \\sim \\operatorname {B e r n o u l l i} (p),", + "image_path": "8ef34e137acd1dcc03fbec4567e71f644c96bebb645e606cbe554bf4245d4b01.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 359, + 294, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 359, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 80, + 359, + 294, + 378 + ], + "type": "interline_equation", + "content": "\\theta_ {a d t} = \\left\\{\\theta_ {a d p a t e r} \\right\\} \\cup \\left\\{\\mathrm {D A R E} _ {\\mathcal {M} _ {2}} \\left(\\theta_ {s f t} - \\theta_ {p r e}\\right) \\right\\}, \\tag {8}", + "image_path": "2d25ef558f6d1638134c3662e197959e2d87b1f9822f7bd3ff194721569cd8d7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 375, + 211, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 375, + 211, + 388 + ], + "spans": [ + { + "bbox": [ + 80, + 375, + 211, + 388 + ], + "type": "interline_equation", + "content": "\\theta_ {d e g} = \\mathrm {D A R E} _ {\\mathcal {M} _ {1}} \\left(\\theta_ {d y n} - \\theta_ {p r e}\\right),", + "image_path": "27b4a180bf7dced04d018af07f597c075e6fda8d0135de8a0250cb18323c1d12.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 390, + 159, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 390, + 159, + 402 + ], + "spans": [ + { + "bbox": [ + 80, + 390, + 159, + 402 + ], + "type": "interline_equation", + "content": "\\theta_ {c o n} = \\theta_ {s f t} - \\theta_ {a d t},", + "image_path": "4c2e32a8b70c33b7939e653cb75983802718820545feddf273044e825c8b4eb6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "text", + "content": "Finally, Task-Arithmetic [16] is employed to inject control parameters " + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\theta_{adt}" + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "text", + "content": " into both " + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\theta_{deg}" + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\theta_{con}" + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "text", + "content": ", deriving two enhanced models: the degree-enhanced model " + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}^{*}" + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "text", + "content": " and the consistency-enhanced model " + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\theta_{con}^{*}" + }, + { + "bbox": [ + 55, + 411, + 296, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 468, + 294, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 468, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 100, + 468, + 294, + 487 + ], + "type": "interline_equation", + "content": "\\theta_ {d y n} ^ {*} = \\text {T a s k - A r i t h m e t i c} (\\theta_ {d e g}, \\theta_ {a d t}), \\tag {9}", + "image_path": "be2636e19964d36202e7ac2e7367c1196da118bcf31d7b80ce114390a3e13dc1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 484, + 249, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 484, + 249, + 497 + ], + "spans": [ + { + "bbox": [ + 102, + 484, + 249, + 497 + ], + "type": "interline_equation", + "content": "\\theta_ {c o n} ^ {*} = \\text {T a k - A r i t h m e t i c} \\left(\\theta_ {c o n}, \\theta_ {a d t}\\right).", + "image_path": "b97222ac6e07b3c2b90013cb0565b0d7f19bdc26cfdb23c4c238da5d86c13a15.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Decoupled Video Sampling. During the denoising process for video generation, we typically observe that in the early stages, I2V models prioritize long-term temporal planning, including the motion of both subjects and textures, while later stages often focus more on refining subject appearance, which is mainly related to content consistency and overall image quality [37, 39]. Therefore, in the first " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " steps of the denoising process, we adjust the video generator to make it more dynamic-enhanced and control-enhanced to increase motion controllability. In the later stages of the denoising process, we re-weight the motion-related modules, allowing the model to focus more on restoring the specific details of each frame of the subject, thereby generating high-consistency videos. Specifically, we define a generated model " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{gen}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " by blending the degree-enhanced model " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}^{*}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " and the consistency-merged model " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{con}^{*}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " based on the denoising step " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " to 0:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 368, + 82, + 499, + 96 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 82, + 499, + 96 + ], + "spans": [ + { + "bbox": [ + 368, + 82, + 499, + 96 + ], + "type": "interline_equation", + "content": "\\theta_ {g e n} = \\alpha_ {t} \\theta_ {d y n} ^ {*} + (1 - \\alpha_ {t}) \\theta_ {c o n} ^ {*},", + "image_path": "783e91019a0b1707172f27c5f7d586cfd6b80f3db3997f84913bc2f8b4ad0661.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 98, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 98, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 377, + 98, + 553, + 129 + ], + "type": "interline_equation", + "content": "\\alpha_ {t} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} t > T - K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {10}", + "image_path": "3afd567ce2b5d890f7d7538887c146c8cd77de0790d09de31e3755399ad55f64.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 135, + 555, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 555, + 159 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 135, + 555, + 159 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 135, + 555, + 159 + ], + "type": "text", + "content": " is the total number of denoising steps, and " + }, + { + "bbox": [ + 313, + 135, + 555, + 159 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 135, + 555, + 159 + ], + "type": "text", + "content": " is a threshold that determines the switching point." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 171, + 390, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 171, + 390, + 184 + ], + "spans": [ + { + "bbox": [ + 314, + 171, + 390, + 184 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 190, + 364, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 364, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 364, + 204 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "content": "In all experiments, videos and images are resized to a " + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "inline_equation", + "content": "320 \\times 512" + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "content": " center-cropped format. During model training, we set the learning rate to " + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "content": ", a batch size of 64, and trained for 100K steps on 8 A100 GPUs. For inference, we use DDIM [34] with 50 steps and a classifier-free guidance scale of 7.5, generating 16 frames for a 2-second video." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 281, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 339 + ], + "type": "text", + "content": "Datasets. We use the WebVid-2M dataset [2] for supervised fine-tuning, which contains two million video-text pairs from stock footage websites. For motion controllability annotation, we use CoTracker [18] to annotate motion information on the WebVid-2M dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 340, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 556 + ], + "type": "text", + "content": "Evaluation. The evaluation benchmarks include VBench [27] and UCF101 [35]. VBench is a comprehensive benchmark for video generation, incorporating recent models and extensive evaluation metrics. UCF101 contains 101 human action categories with diverse background and lighting variations. For VBench, following previous work [20, 46], we utilize the official dataset to assess the model across several metrics: I2V-Subject Consistency (I2V Sub.), I2V-Background Consistency (I2V Back.), video quality, camera motion, and dynamic degree. The video quality evaluation considers dimensions including Subject Consistency, Background Consistency, Smoothness, Aesthetic Score, Imaging Quality, and Temporal Flickering. The Avg. score is calculated as " + }, + { + "bbox": [ + 313, + 340, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\frac{1}{10} * (\\text{I2V Sub.} + \\text{I2V Back.} + 6 * \\text{Video Quality} + \\text{Motion Degree} + \\text{Motion Control})" + }, + { + "bbox": [ + 313, + 340, + 555, + 556 + ], + "type": "text", + "content": ", which is consistent with VBench I2V original setting. For UCF101, similar to [39, 40], we use Fréchet Video Distance (FVD) [36] and Inception Score (IS) [30] to assess video quality." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 564, + 450, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 564, + 450, + 577 + ], + "spans": [ + { + "bbox": [ + 314, + 564, + 450, + 577 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "VBenchmark Results Analysis. We compare our method with two types of approaches on VBenchmark. The first kind of comparing methods include those requiring large-scale video pretraining, such as animate-Anything [7], SVD [3], and DynamiCrafter [40]. The other comparison methods are finetuning/inference strategies based on DynamiCrafter, including naive finetuning with the same training setup as our method (Naive FT) and Conditional Image Leakage (CIL) methods [46] (CIL Infer and CIL FT). Notably, our method performs best in Motion Degree and Motion Control across all baselines, even including large-scale pretrained methods," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12516" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 125, + 553, + 288 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Quantitative results on the VBench I2V benchmark for various I2V-DM pretrained baselines and finetuning/inference techniques, including our methods based on DynamiCrafter. \"PT.\" in the table denotes whether the method requires video pretraining, and \"FT.\" denotes whether the method requires supervised finetuning on WebVid-2M. The global best and DynamiCrafter-based best performing metrics are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 125, + 553, + 288 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 553, + 288 + ], + "type": "table", + "html": "
ModelAvg.PT.FT.I2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
ConsistI2V (Ren et al.)75.88✓✗95.8295.9585.7418.6233.92
SEINE (Chen et al.)76.56✓✗96.5796.8085.7134.3123.67
Animate-Anything (Dai et al.)74.61✓✗98.7698.5888.842.6813.08
VideoCrafter (Chen et al.)76.96✓✗91.1796.9487.5522.6033.60
SVD-XT-1.0 (Blattmann et al.)-✓✗97.5297.6386.5452.36-
SVD-XT-1.1 (Blattmann et al.)-✓✗97.5197.6286.6643.17-
I2VGen-XL (Zhang et al.)76.00✓✗96.4896.8387.0226.1018.48
DynamiCrafter (Xing et al.)79.93✓✗94.9195.3084.9568.5430.85
DynamiCrafter-Naive FT72.91✗✓97.9298.7083.5211.6717.82
DynamiCrafter-CIL Infer (Zhao et al.)80.47✗✗96.9693.3485.2869.4331.29
DynamiCrafter-CIL FT (Zhao et al.)77.55✗✓96.9097.7987.1025.6930.50
DynamiCrafter-Ours84.25✗✓97.0396.6286.2287.6443.87
DynamiCrafter-Ours w/ CIL Infer84.81✗✓97.1195.4587.1089.2143.61
", + "image_path": "c29093831d82561806031391d99ba4a074f2bbefd7f6084885a55729afb3385e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 337, + 288, + 405 + ], + "blocks": [ + { + "bbox": [ + 55, + 306, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 327 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of the DynamiCrafter-based method on the UCF-101 benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 337, + 288, + 405 + ], + "lines": [ + { + "bbox": [ + 61, + 337, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 337, + 288, + 405 + ], + "type": "table", + "html": "
ModelFVD↓IS↑Motion Degree↑
DynamiCrafter288.0113.3168.17
DynamiCrafter-Naive FT373.7713.1947.66
DynamiCrafter-CIL FT409.6913.6853.78
DynamiCrafter-Ours324.7913.7177.93
", + "image_path": "aeee6cc04f222681c693565a64bd87e40510ec19c96359b39f65628c10f2fa8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 423, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 674 + ], + "type": "text", + "content": "as shown in Table 1. We observe that Naive FT causes a significant degradation in Motion Degree (11.67 vs 68.54), consistent with previous findings [46]. Naive FT achieves the highest I2V Subject and Background Consistency (I2V Sub. and I2V Back.) among the DynamiCrafter-based methods because lower motion degree results in higher consistency scores. Interestingly, motion control is also negatively affected in Naive FT (17.82 vs 11.54). We speculate that this is due to the insufficient motion degree, which hinders the model's ability to follow control instructions effectively. The CIL method can mitigate these effects, either by employing the inference strategy on the pre-trained model (CIL Infer) or by repairing the Naive FT (CIL FT). However, its enhancements in Motion Control and Motion Degree are limited. In contrast, our method, combining adaptation and extrapolation, significantly improves both motion degree and motion control (11.67 vs 87.64) while preserving similar video quality (87.10). Furthermore, our method is orthogonal to the CIL Infer method, and integrating it with our method further boosts performance (Avg. 84.25 vs 84.81)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "UCF101 Results Analysis. In the UCF101 dataset, our method still exhibits superior motion degree ability despite" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 308, + 433, + 389 + ], + "blocks": [ + { + "bbox": [ + 318, + 308, + 433, + 389 + ], + "lines": [ + { + "bbox": [ + 318, + 308, + 433, + 389 + ], + "spans": [ + { + "bbox": [ + 318, + 308, + 433, + 389 + ], + "type": "image", + "image_path": "3ed9ba45f26921b99ce254f23e4ff288dfa41ad75080ac5b285a73d28723d812.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 436, + 308, + 551, + 389 + ], + "blocks": [ + { + "bbox": [ + 436, + 308, + 551, + 389 + ], + "lines": [ + { + "bbox": [ + 436, + 308, + 551, + 389 + ], + "spans": [ + { + "bbox": [ + 436, + 308, + 551, + 389 + ], + "type": "image", + "image_path": "3f40f575d7aa348e86614ede1c24599cbc04d2273e70f480d76c5ef2f277c940.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 389, + 434, + 464 + ], + "blocks": [ + { + "bbox": [ + 318, + 389, + 434, + 464 + ], + "lines": [ + { + "bbox": [ + 318, + 389, + 434, + 464 + ], + "spans": [ + { + "bbox": [ + 318, + 389, + 434, + 464 + ], + "type": "image", + "image_path": "b4d002a3e35f155dc61373a026e10214b25d26439c895465d0c823c2afe78ec0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 479, + 555, + 513 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 513 + ], + "type": "text", + "content": "Figure 3. Human Evaluation of our method Compared to SVD, VideoCrafter (VC), DynamiCrafter (DC), and its variant of Native Fine-Tuned (DC-FT) and CIL-Based Methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 436, + 389, + 551, + 464 + ], + "blocks": [ + { + "bbox": [ + 436, + 389, + 551, + 464 + ], + "lines": [ + { + "bbox": [ + 436, + 389, + 551, + 464 + ], + "spans": [ + { + "bbox": [ + 436, + 389, + 551, + 464 + ], + "type": "image", + "image_path": "9ea3af99211531cb8a6bce2d1bae0298866fcd76b618e0e58c85ae32e0145e95.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 536, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 632 + ], + "type": "text", + "content": "the low resolution and poor image quality. Both Naive FT and CIL FT suffer from degradation in Fréchet Video Distance (FVD), as shown in Table 2. Unlike the performance on VBenchmark, CIL performs worse than Naive FT on UCF101, likely due to the added noise schedule disrupting the already noisy and low-resolution images. Our method, however, is more stable in FVD and outperforms in Inception Score (IS) and Motion Degree." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "User Study. As shown in Figure 3, our method excels in Motion Degree and Motion Control while remaining competitive in video quality and I2V consistency. The results are consistent with the above findings, supporting that our strategies enhance both video motion dynamics and control" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12517" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 68, + 522, + 475 + ], + "blocks": [ + { + "bbox": [ + 88, + 68, + 522, + 475 + ], + "lines": [ + { + "bbox": [ + 88, + 68, + 522, + 475 + ], + "spans": [ + { + "bbox": [ + 88, + 68, + 522, + 475 + ], + "type": "image", + "image_path": "70134cf65ce9b4e42969d2f7e8755b2e5dbb3e5ab5e5f336e9dded2823ba7d88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 493, + 555, + 528 + ], + "lines": [ + { + "bbox": [ + 55, + 493, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 555, + 528 + ], + "type": "text", + "content": "Figure 4. In the visual comparisons, we generated 16-frame videos using DynamiCrafter, CIL, Consistent-I2V, and our model. We selected 3 to 4 frames at regular intervals for display at the top. The upper section visualizes the overall aesthetic quality, while the lower section uses Optical Flow to illustrate the dynamic qualities of the generated videos for each method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 542, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 567 + ], + "type": "text", + "content": "while preserving quality and consistency, yielding superior overall performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 575, + 185, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 185, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 185, + 588 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "To empirically evaluate the visual quality of our method, we compare it with several state-of-the-art image-to-video generation approaches, including DynamiCrafter [40], CIL [46], and ConsistI2V [27]. In Figure 4, we observe that our method generates temporally coherent videos with a sufficient range of motion that closely adhere to the input text conditions. In contrast, DynamiCrafter exhibits inconsistencies in text control, likely due to challenges in capturing the semantic content of the input image. The ConsistI2V method produces videos that remain static over time, with" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 542, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 651 + ], + "type": "text", + "content": "the color points on the figures almost unchanged, indicating a lack of motion dynamics and control. While CIL generates videos that semantically resemble the input images and have enhanced motion compared to ConsistI2V, its improvements are limited. CIL does not sufficiently vary the content dynamically to match the text prompts, resulting in limited motion and control. Our method, however, preserves image consistency while achieving a large motion degree and exhibiting good controllability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 659, + 408, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 408, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 408, + 673 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We demonstrate the importance of our extrapolation and decoupling methods in Table 3: (1) On top of our adaptation training, using our extrapolation and decoupling techniques" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12518" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 91, + 531, + 171 + ], + "blocks": [ + { + "bbox": [ + 172, + 71, + 436, + 82 + ], + "lines": [ + { + "bbox": [ + 172, + 71, + 436, + 82 + ], + "spans": [ + { + "bbox": [ + 172, + 71, + 436, + 82 + ], + "type": "text", + "content": "Table 3. Ablation study of the three stages of our framework on VBenchmark." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 91, + 531, + 171 + ], + "lines": [ + { + "bbox": [ + 75, + 91, + 531, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 91, + 531, + 171 + ], + "type": "table", + "html": "
(a) Adaptation(b)Extrapolation(c) DecoupleI2V Sub.I2V Back.Video QualityMotion DegreeMotion Control
✓XX97.5196.3986.6122.8549.31
X✓X94.1495.6283.5298.2117.63
XX✓94.6096.2183.9569.6132.37
✓✓X91.1192.8684.3271.6826.71
X✓✓96.0194.6084.3493.3321.85
✓✓✓97.0396.6286.2287.6443.87
", + "image_path": "500c845d1cccb95aa48819a6580b918f8a077a560669423dce1a486486811c35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 431 + ], + "type": "text", + "content": "significantly improves the motion degree, from 22.85 to 87.64. (2) Using extrapolation alone can drastically improve the motion degree (69.61 vs 93.33), but it comes at the cost of degradation in I2V-Subject Consistency (I2V Sub.) and I2V-Background Consistency (I2V Back.), as well as motion control. This is because reversing the training process of the pre-trained model to the SFT model unlearns the camera motion ability, causing significant interference if directly merged [16]. (3) Using the decoupling technique alone can improve the motion degree while preserving other performance metrics. This indicates that the I2V model indeed exhibits functional differentiation during the sampling process. (4) Adding adaptation, though it increases training costs, greatly enhances the Motion Control aspect of the model. The degradation in motion degree indicates that direct supervision is suboptimal and can lead to minor motion dynamics, a finding consistent with previous research [46]. (5) When paired with the decoupling technique, we achieve optimal results, integrating high motion degree and control ability while preserving consistency and video quality." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 66, + 442, + 175, + 513 + ], + "blocks": [ + { + "bbox": [ + 66, + 442, + 175, + 513 + ], + "lines": [ + { + "bbox": [ + 66, + 442, + 175, + 513 + ], + "spans": [ + { + "bbox": [ + 66, + 442, + 175, + 513 + ], + "type": "image", + "image_path": "8487239c8707ead2a78c9aa4c4b52f7f9bc17c48f0322fb265445bf9289bceff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 178, + 443, + 282, + 513 + ], + "blocks": [ + { + "bbox": [ + 178, + 443, + 282, + 513 + ], + "lines": [ + { + "bbox": [ + 178, + 443, + 282, + 513 + ], + "spans": [ + { + "bbox": [ + 178, + 443, + 282, + 513 + ], + "type": "image", + "image_path": "c7faf79a98499b80709c802e758f7d607147326f4cef447eff43e28345feebab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 66, + 516, + 173, + 586 + ], + "blocks": [ + { + "bbox": [ + 66, + 516, + 173, + 586 + ], + "lines": [ + { + "bbox": [ + 66, + 516, + 173, + 586 + ], + "spans": [ + { + "bbox": [ + 66, + 516, + 173, + 586 + ], + "type": "image", + "image_path": "325f4dc6793b1cab5af21ad17c16945fe8ea437377c277c73802937aa093950c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 590, + 233, + 600 + ], + "lines": [ + { + "bbox": [ + 118, + 590, + 233, + 600 + ], + "spans": [ + { + "bbox": [ + 118, + 590, + 233, + 600 + ], + "type": "text", + "content": "(a) Effect of extrapolate strength " + }, + { + "bbox": [ + 118, + 590, + 233, + 600 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 178, + 516, + 282, + 586 + ], + "blocks": [ + { + "bbox": [ + 178, + 516, + 282, + 586 + ], + "lines": [ + { + "bbox": [ + 178, + 516, + 282, + 586 + ], + "spans": [ + { + "bbox": [ + 178, + 516, + 282, + 586 + ], + "type": "image", + "image_path": "492b8ccea4fb5c586769897af4ed4a12542d7dca40672b3c8b585e66d86a7a42.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 69, + 602, + 175, + 664 + ], + "blocks": [ + { + "bbox": [ + 69, + 602, + 175, + 664 + ], + "lines": [ + { + "bbox": [ + 69, + 602, + 175, + 664 + ], + "spans": [ + { + "bbox": [ + 69, + 602, + 175, + 664 + ], + "type": "image", + "image_path": "f5c7037dfa3ced0fdb6a8d959d5372e7a35d4e0435b4499d7cd63ea6dc51f1ea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 668, + 266, + 677 + ], + "lines": [ + { + "bbox": [ + 84, + 668, + 266, + 677 + ], + "spans": [ + { + "bbox": [ + 84, + 668, + 266, + 677 + ], + "type": "text", + "content": "(b) Effect of different switching time and model choices." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 178, + 602, + 282, + 664 + ], + "blocks": [ + { + "bbox": [ + 178, + 602, + 282, + 664 + ], + "lines": [ + { + "bbox": [ + 178, + 602, + 282, + 664 + ], + "spans": [ + { + "bbox": [ + 178, + 602, + 282, + 664 + ], + "type": "image", + "image_path": "9a8bff6232c301f3098ca966ef3cb1baf2a2f925d0d41aa64555c1cb374d75af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 687, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 687, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 295, + 710 + ], + "type": "text", + "content": "Figure 5. The impact of different extrapolate strengths and different sampling strategies." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 191, + 478, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 478, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 478, + 205 + ], + "type": "text", + "content": "4.5. Effect of Extrapolate Strength" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": "In our extrapolation experiments, we analyzed the impact of different " + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": " values on performance using Equation 3. As shown in Figure 5a, as " + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 212, + 555, + 319 + ], + "type": "text", + "content": " increases, the motion degree first rapidly rises within the 0-0.7 range, peaks, and then gradually declines between 1.2-2.0. Simultaneously, quality and consistency metrics decline. This indicates that extrapolation enhances motion degree but sacrifices quality and consistency. This aligns with the findings in Section 4.4 and underscores the importance of appropriate decoupling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 337, + 529, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 529, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 529, + 350 + ], + "type": "text", + "content": "4.6. Analysis on Decoupled Sampling Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": "To study the impact of switch model time and choice, we conducted experiments illustrated in Figure 5b. We outline two model switch strategies during the denoising process from time " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": " to 0: (1) Initially using " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}^{*}" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": " and then switching to " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\theta_{sft}^{*}" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": ". (2) Starting with " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\theta_{sft}^{*}" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": " and then transitioning to " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}^{*}" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": ". The performance trends are shown in Figure 5b. The first strategy, which initially uses " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}^{*}" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": ", results in higher motion degree with a later switch (smaller " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": "), but with poorer quality. Conversely, an earlier switch (larger " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": ") reduces the degree. The second strategy, introducing " + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\theta_{dyn}^{*}" + }, + { + "bbox": [ + 313, + 357, + 555, + 514 + ], + "type": "text", + "content": " later, has a less pronounced impact on the degree. This aligns with our hypothesis in Section 3.4, which suggests that the early sampling phase significantly influences the degree." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 534, + 388, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 388, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 388, + 546 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a novel Extrapolating and Decoupling framework for image-to-video generation. Our framework consists of three stages and can effectively model the motion control and motion degree. In the first stage, we explicitly inject the textual condition into I2V-DM to improve motion controllability. Then, we devise a training-free extrapolation strategy to boost the motion degree. Finally, we decouple the motion-relevant parameters and dynamically inject them into the I2V-DM along different de-noising steps. With our framework, we can generate videos with both natural motion and temporal consistency. To the best of our knowledge, we are the first to apply merging techniques in video generation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12519" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "text", + "content": "[4] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 295, + 360 + ], + "type": "text", + "content": "[5] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 295, + 396 + ], + "type": "text", + "content": "[6] Feng Cheng, Ziyang Wang, Yi-Lin Sung, Yan-Bo Lin, Mohit Bansal, and Gedas Bertasius. Dam: Dynamic adapter merging for continual video qa learning, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "type": "text", + "content": "[7] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Finegrained open domain image animation with motion guidance, 2023. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 486 + ], + "type": "text", + "content": "[8] Felix Draxler, Kambis Veschgini, Manfred Salmhofer, and Fred Hamprecht. Essentially no barriers in neural network energy landscape. In International conference on machine learning, pages 1309-1318. PMLR, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 531 + ], + "type": "text", + "content": "[9] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 577 + ], + "type": "text", + "content": "[10] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "text", + "content": "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 700 + ], + "type": "text", + "content": "[12] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "type": "text", + "content": "[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 120, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 553, + 152 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "type": "text", + "content": "[16] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 2, 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 198, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 242 + ], + "type": "text", + "content": "[17] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 243, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 276 + ], + "type": "text", + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "text", + "content": "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 333, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 388 + ], + "type": "text", + "content": "[20] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 432 + ], + "type": "text", + "content": "[21] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 466 + ], + "type": "text", + "content": "[22] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 468, + 555, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 522 + ], + "type": "text", + "content": "[23] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "text", + "content": "[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 579, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 624 + ], + "type": "text", + "content": "[25] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "text", + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 714 + ], + "type": "text", + "content": "[27] Weiming Ren, Harry Yang, Ge Zhang, Cong Wei, Xinrun Du, Stephen Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation. arXiv preprint arXiv:2402.04324, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12520" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 194 + ], + "type": "text", + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 226 + ], + "type": "text", + "content": "[30] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "type": "text", + "content": "[31] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "text", + "content": "[32] Mustafa Shukor, Corentin Dancette, Alexandre Rame, and Matthieu Cord. Unival: Unified model for image, video, audio and language tasks, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "text", + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 381, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 294, + 403 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "text", + "content": "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11):1-7, 2012. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "text", + "content": "[36] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "text", + "content": "[37] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[38] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[39] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "text", + "content": "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[41] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "when merging models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[42] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[43] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[44] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 307 + ], + "type": "text", + "content": "[45] Yiming Zhang, Zhening Xing, Yanhong Zeng, Youqing Fang, and Kai Chen. Pia: Your personalized image animator via plug-and-play modules in text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7747-7756, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[46] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. arXiv preprint arXiv:2406.15735, 2024. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 384 + ], + "type": "text", + "content": "[47] Chujie Zheng, Ziqi Wang, Heng Ji, Minlie Huang, and Nanyun Peng. Weak-to-strong extrapolation expedites alignment, 2024. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12521" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_content_list.json b/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c4ed502c8ccf5f55acdf261eae22284b7adcf19 --- /dev/null +++ b/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_content_list.json @@ -0,0 +1,1529 @@ +[ + { + "type": "text", + "text": "Extreme Rotation Estimation in the Wild", + "text_level": 1, + "bbox": [ + 290, + 130, + 707, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hana Bezalel $^{1}$ Dotan Ankri $^{1}$ Ruojin Cai $^{2}$ Hadar Averbach-Elor $^{1,2}$ $^{1}$ Tel Aviv University $^{2}$ Cornell University", + "bbox": [ + 218, + 179, + 795, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://tau-vailab.github.io/ExtremeRotationsInTheWild/", + "bbox": [ + 263, + 233, + 754, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 282, + 326, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a technique and benchmark dataset for estimating the relative 3D orientation between a pair of Internet images captured in an extreme setting, where the images have limited or non-overlapping field of views. Prior work targeting extreme rotation estimation assume constrained 3D environments and emulate perspective images by cropping regions from panoramic views. However, real images captured in the wild are highly diverse, exhibiting variation in both appearance and camera intrinsics. In this work, we propose a Transformer-based method for estimating relative rotations in extreme real-world settings, and contribute the ExtremeLandmarkPairs dataset, assembled from scene-level Internet photo collections. Our evaluation demonstrates that our approach succeeds in estimating the relative rotations in a wide variety of extreme-view Internet image pairs, outperforming various baselines, including dedicated rotation estimation techniques and contemporary 3D reconstruction methods.", + "bbox": [ + 89, + 315, + 483, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 617, + 220, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of estimating the relative 3D orientation between a pair of images is embodied in fundamental computer vision tasks, such as camera localization [6, 34, 35] and 3D reconstruction [29, 36, 42]. Establishing pixel correspondences (either explicitly or implicitly) is typically a prerequisite for computing the relative rotation between the images. Correspondences, however, cannot be extracted in extreme settings where the images have little to no overlap. As dense imagery may not necessarily be available for many practical applications, a natural question arises: How can we estimate the relative rotation between non-overlapping RGB images, without the use of additional data (such as depth or temporal information)?", + "bbox": [ + 89, + 642, + 482, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We have recently seen pioneering efforts addressing the task of relative rotation estimation in such extreme non-overlapping settings [8, 12]. Prior work has proposed end-to-end neural architectures, demonstrating that hidden cues,", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f2c5c753c36243e5c09bc76366efbfdb87a6d1c8637dd6c805dbb6f6d469c73.jpg", + "image_caption": [ + "Figure 1. Given a pair of (possibly) non-overlapping images captured in the wild—e.g., under arbitrary illumination and intrinsic camera parameters—such as the images of the Dam Square in Amsterdam depicted in red and blue boxes above*, our technique estimates the relative 3D rotation between the images. *The background panorama is illustrated for visualization purposes only." + ], + "image_footnote": [], + "bbox": [ + 542, + 282, + 879, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as vanishing points or the directions of cast shadows, can implicitly guide the model for inferring the relative orientation between the images. To facilitate learning and evaluation, datasets constructed from panoramic views were adopted. These datasets emulate perspective views by cropping sub-areas from these panoramas, enabling generation of image pairs with various degrees of overlap. However, while such emulated views perhaps capture some of the challenges associated with extreme-view imagery, are they sufficient for representing real images—particularly, images captured in the wild?", + "bbox": [ + 511, + 566, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a new approach that tackles the problem of extreme rotation estimation in the wild. Consider the boxed images in Figure 1. Internet (i.e., in the wild) images may vary due to a wide range of factors, including transient objects, weather conditions, time of day, and the cameras' intrinsic parameters. To explore this problem, we introduce a new dataset, ExtremeLandmarkPairs (ELP), assembled from publicly-available scene-level Internet image collections. We observe that the set of real extreme-view image pairs is limited, as Internet datasets are typically scene-centric, with nearby cameras commonly", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1061", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "capturing overlapping views. Therefore, to facilitate training, we propose a progressive learning scheme that leverages and augments images cropped from panoramic views, allowing for gradually generalizing the model onto real Internet data. In particular, we construct datasets with varying field of views, that better resemble the distribution of real data, and perform image-level appearance augmentations by leveraging recent advancements in text-to-image Diffusion models [7, 22, 31].", + "bbox": [ + 93, + 90, + 480, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To estimate extreme rotations in the wild, we propose a Transformer-based model that is provided with auxiliary channels, including the spatial distribution of local keypoints and matches and semantic segmentation maps, allowing for better reasoning over real image pairs with little or no overlap. Our results demonstrate that our model can accurately predict the relative rotations for a wide variety of extreme-view image pairs that vary in illumination, dynamic regions, and intrinsic parameters. We conduct extensive experiments, quantifying performance both over real Internet data and also over emulated perspective images cropped from panoramic views. Our evaluation shows that our model significantly improves over strong baselines when considering real images, while achieving comparable performance over emulated perspective image pairs.", + "bbox": [ + 93, + 227, + 480, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 94, + 467, + 228, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Relative pose estimation is a fundamental task in computer vision, typically studied for overlapping camera views. Traditionally, this task has been divided into two stages: correspondence estimation from local feature matching, followed by geometry-based pose estimation. In recent years, feature matching methods have advanced from using heuristic feature descriptors [5, 27, 32] and RANSAC-based matches [21] to learning-based feature extraction [4, 14-16, 18, 40, 45] and matching techniques [3, 26, 33], with several methods performing both feature extraction and matching using unified learning-based frameworks [4, 17, 38].", + "bbox": [ + 93, + 492, + 480, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These methods are generally invariant to changes in illumination and appearance, demonstrating robust performance across various scene scales and also over in-the-wild datasets. However, their reliability diminishes in extreme view scenarios due to their dependence on visual overlap.", + "bbox": [ + 93, + 659, + 480, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extreme-view scenarios lacking local pixelwise correspondences necessitate the use of end-to-end learning-based pose estimation methods which directly infer the 3D relationships and geometry from sparse and extreme-view images. Indeed, large-scale 3D object datasets have paved the way for learning-based methods which estimate camera pose directly from sparse views [20, 25, 37, 43, 44, 50-52]. However, these methods primarily concentrate on object-centric scenes, under experimental settings that typically assume similar lighting and camera intrinsics for the input views. Furthermore, these methods often utilize bounding", + "bbox": [ + 93, + 734, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "box inputs defining the object of interest, which is less suitable in the case of images depicting large-scale scenes.", + "bbox": [ + 517, + 90, + 901, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several prior works address a sparse view setting at scene-scale. In particular, Chen et al. [11] propose to learn a discrete distributions of pose space, Agarwala et al. [1] simplify scene reconstruction using a plane representation, and Rockwell et al. [30] introduce an inductive bias of the 8-point algorithm into a vision transformer architecture. Recently, models that directly predict pixel-aligned point maps from input pairs (or sparse image collections), such as DUSt3R [46] and Mast3R [23], have demonstrated promising results on pose estimation and scene-scale 3D reconstruction of Internet data with a wide baseline.", + "bbox": [ + 517, + 122, + 903, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera pose estimation for non-overlapping views presents a greater challenge. Earlier efforts [10] explored searching for consistent temporal behavior. Several work utilize pairwise RGB and depth scan data to estimate relative pose among such extreme pairs [48, 49]. Cai et al. [8] tackle pose estimation for non-overlapping views without the use of additional data, by introducing a learning-based network leveraging cross-correlation volume to exploit implicit cues. This correlation volume is later enhanced through the integration of transformer attention modules [12]. Nonetheless, these approaches assume constrained 3D environments, including the assumption of consistent lighting and camera intrinsics, and are designed for camera distribution of emulated perspective views cropped from panoramas. In this work, we aim to address pose estimation for realistic in-the-wild non-overlapping image pairs, enhancing the applicability of extreme pose estimation to Internet photos and real-world data.", + "bbox": [ + 517, + 287, + 903, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. The ExtremeLandmarkPairs Dataset", + "text_level": 1, + "bbox": [ + 517, + 571, + 836, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works on extreme pose estimation use panoramic views, cropping from it sub-areas to emulate perspective views [8, 12]. To evaluate and train models on real perspective image pairs, we propose a new benchmark and dataset, ExtremeLandmarkPairs (ELP), constructed from Internet image pairs from the MegaDepth [24], Cambridge Landmarks [2], and MegaScenes [39] datasets. In this section, we first describe the dataset construction procedure (Section 3.1), and then present details regarding dataset size and train and test splits (Table 1).", + "bbox": [ + 517, + 597, + 903, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Dataset Construction", + "text_level": 1, + "bbox": [ + 517, + 756, + 710, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To construct a dataset of real perspective image pairs with varying degrees of overlap, we leverage available scene-level training data. Existing Internet image collections typically contain camera poses (predicted up to scale), which are determined using Structure-from-Motion (SfM) algorithms, such as COLMAP [36]. In what follows, we describe how we extract real image pairs from this data, which can then be used for training and evaluating models.", + "bbox": [ + 517, + 780, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1062", + "bbox": [ + 485, + 946, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ab05a16a75e49499ee349e8ad6e82bcbb1e38fab076b2a7571a92230872b43b6.jpg", + "image_caption": [ + "(a) Reconstruction and camera distribution of all images" + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 290, + 207 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eb534a776b1ea06d3928ea43616cafac13240a4ffebb243d11b2ce45c0e8840b.jpg", + "image_caption": [ + "(b) Camera distribution of large-overlapping pairs" + ], + "image_footnote": [], + "bbox": [ + 295, + 93, + 477, + 202 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/03ee3c05ba1b6fd796e0cf91f897272bc55e5adab7324d19d6661b4f09c50f17.jpg", + "image_caption": [ + "(c) Camera distribution of non-overlapping pairs", + "Figure 2. Camera distribution of the Vatican, Rome scene from the ExtremeLandmarkPairs Dataset. We construct a dataset of real perspective image pairs with predominant rotational motion shown in (b) and (c) from the dense imagery reconstruction in (a)." + ], + "image_footnote": [], + "bbox": [ + 98, + 243, + 478, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identifying Pairs with Predominant Rotational Motion.", + "text_level": 1, + "bbox": [ + 89, + 628, + 482, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior works targeting relative rotation estimation, in particular for non-overlapping views, mostly utilize panoramic views, focusing on image pairs with purely rotational motion. Pairs belonging to real image collections almost always contain a non-negligible translation component. Furthermore, unlike in the StreetLearn [28] dataset used by prior work [8, 12] that provides exact translation values between consecutive panoramas which allows for filtering pairs with predominant rotational motions, reconstructed relative poses are only provided up to scale. The scale varies among different reconstructed scenes, and therefore there's no global threshold on the relative translation values which can be used for identifying pairs with predominant rotational motion.", + "bbox": [ + 88, + 643, + 482, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To automatically identify such pairs, we observe that available Internet collections require the existence of dense imagery, to compensate for the vast number of unknowns", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the SfM optimization. We therefore construct mutual nearest neighbors edge-weighted graphs, with one graph per landmark. In each graph $G$ , nodes $v \\in V$ correspond to images, and two images are connected by an edge $e \\in E$ if they are both among each other's $K$ nearest neighbors, considering L2 distances between their translations ( $K$ is empirically set to 5). Note that images captured from sparser (outlier) regions in space are unlikely to be within the mutual $K$ nearest neighbors of images captured within denser regions, and hence won't be included in $G$ . Finally, we select a subset of image pairs containing relatively small distances from each scene graph $G$ , yielding a set of image pairs with predominant rotational motion; see the supplementary material for additional details.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extracting Level of Overlap. Following prior work [8, 12], we are interested in training and evaluating models according to three different categories: Large, Small and None, indicating image pairs with a varying amount of overlap. However, unlike prior work that use cropped images with a fixed $90^{\\circ}$ FoV, Internet images contain varying FoV values. Thus, the relative rotation angle is not sufficient for extracting the pair's overlap level.", + "bbox": [ + 511, + 305, + 903, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote the FoV values of image $i\\in [1,2]$ as $[\\mathrm{fov}_x^i,\\mathrm{fov}_y^i ]$ We can parameterize a 3D rotation matrix $\\mathbf{R}$ using three Euler angles $[\\alpha ,\\beta ,\\gamma ]$ denoting the relative roll, pitch and yaw angles, respectively:", + "bbox": [ + 511, + 426, + 903, + 486 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} (\\alpha , \\beta , \\gamma) = \\mathbf {R} _ {z} (\\alpha) \\mathbf {R} _ {y} (\\beta) \\mathbf {R} _ {x} (\\gamma), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 501, + 903, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "following Dense Correlation Volumes' coordinate system. We use the following conditions to determine the overlap level $o$ :", + "bbox": [ + 511, + 521, + 903, + 565 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no = \\left\\{ \\begin{array}{l l} L a r g e & | \\gamma | < \\frac {\\operatorname {f o v} _ {x} ^ {1} + \\operatorname {f o v} _ {x} ^ {2}}{4} \\wedge | \\beta | < \\frac {\\operatorname {f o v} _ {y} ^ {1} + \\operatorname {f o v} _ {y} ^ {2}}{4} \\\\ N o n e & | \\gamma | > \\frac {\\operatorname {f o v} _ {x} ^ {1} + \\operatorname {f o v} _ {x} ^ {2}}{2} \\wedge | \\beta | > \\frac {\\operatorname {f o v} _ {y} ^ {1} + \\operatorname {f o v} _ {y} ^ {2}}{2} \\\\ S m a l l & e l s e \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 585, + 903, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In other words, pairs with relative yaw and pitch angles that are smaller than a quarter of the average corresponding FoV values are considered as pairs with a large overlap ratio. Likewise, pairs with relative yaw and pitch angles that are larger than half the average corresponding FoV values are considered non-overlapping pairs. Pairs in-between these conditions are considered pairs with a small overlap ratio.", + "bbox": [ + 511, + 656, + 906, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additional Filtering. In the scene-scale datasets we explore, large FoV disparities could result in one image focusing on specific architectural details like a statue, while the other captures a much broader scene perspective, further complicating the problem of estimating the relative rotations. To extract pairs consistent in scale, we limit the difference between the FoV values to be at most $5^{\\circ}$ . Furthermore, as most images contain small roll values, we rotate the scenes to match the gravity axis and horizontal axis", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1063", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/52c9c17204310d35fe7f126af2981f620270843ef4044f4a21efd7d5c7c9d38f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SubsetSource#Pairs
#ScenesLargeSmallNoneTotal
Train[39]588333430136842948176595
Validation[39]515339871041308238
Validation Balanced[39]1779255707854
sELP[2]6251282719615300
wELP[24]1727008296434172
", + "bbox": [ + 91, + 89, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. ExtremeLandmarkPairs Dataset Statistics. Above, we report the number of image pairs extracted for each overlap level, split into train and test (with sELP denoting a single camera setting and wELP denoting the \"in the wild\" setting).", + "bbox": [ + 89, + 184, + 482, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and filtered images with rolls exceeding $10^{\\circ}$ . Finally, we excluded images captured from an aerial perspective, by filtering images with translation along y-axis exceeding a global threshold, empirically set to 1.", + "bbox": [ + 89, + 267, + 482, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Size and Splits", + "text_level": 1, + "bbox": [ + 89, + 338, + 302, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply the procedure described above over landmarks from the MegaScenes [39] dataset to create a training set and a validation set. As we focus on outdoor scenes in our work, we filter reconstructions that capture indoor scenes. We obtain a total of nearly 34K non-overlapping pairs originating from over 2K unique landmarks. The validation set was also balanced the set over the overall angle.", + "bbox": [ + 89, + 361, + 482, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For evaluation, we create two test sets, to separately examine image pairs captured in a single camera setting with constant illumination (sELP) and image pairs captured in the wild (wELP). Image pairs in the sELP test set contain images from the Cambridge Landmarks [2] dataset, which contains videos capturing six different landmarks. Image pairs in the wELP test set contain images from the MegaDepth dataset [24], which contains Internet photos from Flickr for a set of large-scale landmarks.", + "bbox": [ + 89, + 467, + 482, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As there is some overlap between the landmarks in MegaScenes [39] and MegaDepth [24], we performed additional filtering, ensuring no overlap exists between the train and test sets. Furthermore, we filter test pairs if one of the images is highly transient—i.e., if transient objects dominate the image. We quantify this using a pretrained segmentation model [19], filtering images containing transient objects in over $40\\%$ of the pixels. We also manually validate non-overlapping image pairs, allowing to further filter unlabeled objects, such as a stage or a market stand. Finally, for non-overlapping image pairs, we observed that the overall relative rotations are highly imbalanced, and therefore we balance this set by the overall relative angle. Table 1 summarizes the number of image pairs and landmarks used for both training and test.", + "bbox": [ + 89, + 604, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 844, + 181, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a pair of Internet images with (possibly) extreme relative motion, we estimate the relative rotation $\\mathbf{R}$ between", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the images. Following prior works on extreme rotation estimation [8, 12], we assume a camera-centric setting, where the two cameras have limited translation. However, our approach departs from prior works by operating on outdoor images captured by a crowd of photographers, with varying intrinsic parameters as well as appearances— $e.g.$ , due to illumination changes and dynamic objects.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model (detailed in Section 4.1; see Figure 3 for an overview) outputs three discrete Euler angles, denoting the relative roll, pitch and yaw angles. As illustrated in prior work [8], this parameterization enables using a simple cross-entropy loss for training. In Section 4.2, we describe our progressive learning scheme, allowing for gradually adapting the model to extreme Internet imagery.", + "bbox": [ + 511, + 198, + 903, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Model", + "text_level": 1, + "bbox": [ + 511, + 315, + 599, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We extract image features using a pretrained LoFTR model [38]. In contrast to the features extracted using common convolutional neural networks pretrained on ImageNet [13], LoFTR is a Transformer-based model trained on Internet pairs, with the goal of extracting local feature matches – a setting and task which is highly related to the one we address in our work, thus enabling extraction of better (i.e., more relevant) features.", + "bbox": [ + 511, + 339, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we are interested in designing a network that can also reason over image pairs with little or no overlap, we combine the extracted features with additional auxiliary channels; see Figure 3 (bottom left). These include keypoint and pairwise matches masks, utilized previously for disambiguating images for similar structures [9]. Intuitively, knowledge over pairwise matches can assist the model in cases of small overlap and for generalizing across different camera intrinsic properties. We also incorporate a segmentation map as an additional auxiliary channel, which segments images into several categories (such as sky, building, road and sidewalk). This channel allows for identifying additional cues, such as the skyline or transient objects, which can aid in determining the rotation for nonoverlapping pairs. In Section 5, we demonstrate the benefit of incorporating these auxiliary channels in our model.", + "bbox": [ + 511, + 462, + 905, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then reshape extracted features (and auxiliary channels) to tokens, concatenating these image tokens with learnable Euler angle position embeddings. These tokens are processed by our Rotation Estimation Transformer module, which uses a transformer decoder architecture [41]. The output Euler angle tokens obtain information from image features and auxiliary channels within transformer attention modules. These tokens are then processed by three different prediction heads (one per angle, denoted as MLP in Figure 3). Each output prediction head uses as input averaged image tokens and one of the output Euler angle tokens, which provides the model with additional angle-specific information, allowing for achieving improved performance,", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1064", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e05a606a2a754adc8b223383658e7090df6e090689bc37d52b33935ae342b1ef.jpg", + "image_caption": [ + "Figure 3. Method architecture. Given a pair of input Internet images, we extract image features using pretrained LoFTR. These features are combined with auxiliary channels, including keypoint and pairwise matches masks, and segmentation maps (visualized on the bottom left). These image features are reshaped into tokens and concatenated with Euler angle position embeddings, which are then processed by our Rotation Estimation Transformer module. The output Euler angle tokens and averaged image tokens are concatenated and processed by MLPs to predict the probability distribution of Euler angles, representing the relative 3D rotation between the input images." + ], + "image_footnote": [], + "bbox": [ + 107, + 98, + 875, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as we show in the supplementary material. The prediction heads output a probability distribution over $N = 360$ bins, capturing an angle in the range $[-180^{\\circ}, 180^{\\circ}]$ .", + "bbox": [ + 88, + 369, + 482, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Learning", + "text_level": 1, + "bbox": [ + 89, + 426, + 199, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As detailed in Section 3, we assemble real image pairs from Internet imagery, which can be used for training and evaluating models. However, even with our proposed ExtremeLandmarkPairs dataset, the set of real image pairs is limited—e.g., only $\\sim 36\\mathrm{K}$ non-overlapping image pairs are extracted, as available Internet datasets are typically scene-centric, with nearby cameras usually capturing overlapping views. Therefore, in what follows, we propose a progressive learning scheme, which leverages panoramic images, and allows for gradually generalizing the model to images captured in-the-wild. All learning stages are optimized using a cross-entropy loss for each Euler angle prediction. Additional details are provided in the supplementary material.", + "bbox": [ + 88, + 450, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization. We begin by training our model on the perspective views cropped from panoramic images, using the data created by Cai et al [8]. Specifically, we use the image pairs cropped from panoramic images included in the StreetLearn [28] dataset, depicting various streets in Manhattan. As further detailed in Cai et al [8], this training set includes roughly 1M image pairs sampled from the same panorama, split according to the overlap level.", + "bbox": [ + 88, + 652, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training with Data Augmentations. We observe that the dataset constructed by prior work could be augmented to better capture the distribution of image pairs captured inthe-wild. In particular, we focus on two types of data augments, which we elaborate on next: (i) field of view (FoV) augmentations (denoted henceforth as $\\Delta \\mathrm{FoV}$ ) and (ii) image-level appearance augmentations (denoted henceforth as $\\Delta \\mathrm{Im}$ ). In Section 5, we demonstrate how both types", + "bbox": [ + 88, + 779, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f41d914e36d412fff29247b302c386059ab2a8610ed6d9bee580eec36fb3de22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 366, + 612, + 439 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c65f2ab42e9dd59fba11dcdd8ead8234b7fcb4a21ccd1c3e872215dbd203f15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 366, + 712, + 439 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/179fa16bb1559513abd0d439d256fc57c4c1938319cfa3ed368a5802381e7de8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 366, + 807, + 439 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34f1f671f5b12ba90fecc9decf60c69be8821fa88b6785834ff8cc8f11c91da0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 366, + 903, + 439 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17acefd6b1feb0660a88f8c83f58cff293545ec1c6968b0d355f850e1af137d2.jpg", + "image_caption": [ + "busy street" + ], + "image_footnote": [], + "bbox": [ + 517, + 443, + 612, + 527 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/26e848fbc28cb8551610bc23778892c0d8f002adf89b77ddaeb44add11918361.jpg", + "image_caption": [ + "Figure 4. Augmenting perspective images cropped from panoramic views with image-level appearance modifications. Given an input image (left) and a target text prompt \"Make it $\\langle w\\rangle$ (\" $\\langle w\\rangle$ is specified above), we use a conditional Diffusion model [7] to create semantic appearance augmentations which modify both the global image characteristics as well as local image regions." + ], + "image_footnote": [], + "bbox": [ + 620, + 443, + 712, + 527 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b992877c1d45e880bb2fccb8cbf65d94ee8db9d7f75927dcb981a4b9e4c64921.jpg", + "image_caption": [ + "night" + ], + "image_footnote": [], + "bbox": [ + 715, + 443, + 807, + 527 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e07e2abc8014def5d0a62032f02037bbd9b705d41c1a6536c7bd78cd209c4f2.jpg", + "image_caption": [ + "Christmas" + ], + "image_footnote": [], + "bbox": [ + 808, + 443, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of augmentations improve the model's ability in generalizing to real Internet scenes.", + "bbox": [ + 511, + 669, + 903, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rather than cropping perspective images with a fixed $90^{\\circ}$ FoV as was done in prior work, we analyze the FoV values of the images belonging to the ELP training set. We compute the mean and standard-deviation values, denoted as $\\mu$ and $\\sigma$ , respectively. We then construct new perspective images by sampling from a Gaussian distribution $G(\\mu, a \\cdot \\sigma)$ that resembles the distribution of real data, setting $a = 1.5$ for obtaining a more diverse set, which also bears higher similarity to the perspective images used during initialization. We also allow for FoV differences (of up to $5^{\\circ}$ ) between the two images paired together. Additionally, rather than providing the model with the full content within these regions, we construct crops with various aspect ratios to fur", + "bbox": [ + 509, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1065", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ther emulate real image pairs.", + "bbox": [ + 89, + 90, + 289, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-the-wild images differ not only in their intrinsic parameters, but also in their appearance - due to presence of varying illumination and dynamic objects. Therefore, in addition to augmenting the dataset with images of various field of views, we also perform image-level appearance augmentations, leveraging recent advancements in text-to-image Diffusion models [7, 22, 31]. In particular, we apply the conditional InstructPix2Pix [7] model on a subset of our data, using a set of text prompts of the format \"Make it $\\langle w\\rangle$ \", where $\\langle w\\rangle$ captures diverse appearance modifications. See Figure 4 for a few illustrative examples; note how some of these modify the global illumination (e.g., altering the time of day or the season) while others yield local edits (such as adding more people to make it \"a busy street\"). We performed additional filtering on the augmented set, to avoid significant edits that also altered the structure of the scene; see the supplementary material for additional details.", + "bbox": [ + 89, + 107, + 483, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training on Real Data. Finally, we finetune the model over real image pairs from our proposed ELP dataset. To prioritize learning over extreme image pairs, we first pass a batch of non-overlapping image pairs, followed by a batch of overlapping image pairs.", + "bbox": [ + 89, + 364, + 483, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results and Evaluation", + "text_level": 1, + "bbox": [ + 89, + 452, + 310, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate our method, we compare our model with prior relative rotation estimation methods on the sELP and wELP test sets. In particular, our experiments seek to answer the following questions:", + "bbox": [ + 89, + 477, + 483, + 537 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- How well do previous methods perform on our proposed task of extreme rotation estimation in the wild?", + "- How important is our progressive training scheme, and to what extent can the improved performance be attributed to training on the ExtremeLandmarkPairs dataset?", + "- What is the impact of our design choices?" + ], + "bbox": [ + 89, + 539, + 482, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also present qualitative results over different overlap levels in Figure 5. These results illustrate our model's performance, and its robustness to varying illumination conditions and transient objects (such as the screen and the beer cup in the rightmost examples). Please refer to the supplementary material for many more qualitative results, including visualizations for baseline methods.", + "bbox": [ + 89, + 628, + 483, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Baselines", + "text_level": 1, + "bbox": [ + 89, + 741, + 199, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method to correspondence-based techniques, dedicated (end-to-end) relative rotation estimation techniques as well as recent relative pose estimation techniques. Specifically, we consider two correspondence-based methods: SIFT [27] and LoFTR [38]. These methods extract keypoint matches to compute an essential matrix using the RANSAC algorithm and calculate the rotation that stems from the essential matrix. We also compare against two prior works specifically targeting extreme", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rotation estimation using end-to-end deep learning frameworks: DenseCorrVol [8] and CascadedAtt [12]. These methods are trained and evaluated on images sampled from panoramic views from StreetLearn [28]. Results for CascadedAtt are reported over a reproduced model, as a pretrained model is not provided; additional details are provided in the supplementary material.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also consider 8PointViT [30], which estimates the relative pose between two images with a Vision Transformer that is modified to be close to the eight point algorithm. Following prior work [12], we only report results over overlapping settings for this baseline. Furthermore, as it assumes a single camera setting, we evaluate it only on the sELP test set. Finally, we consider Dust3R [46], a recent technique for dense and unconstrained stereo 3D reconstruction. While this work does not specifically target the setting of extreme rotation estimation, we examine to what extent they can be adapted for this task in the wild. Note that Dust3R was trained on MegaDepth. Since our wELP test set is constructed from image pairs from this dataset, it is not entirely fair comparison; therefore, Dust3R's results on the wELP test set are highlighted in gray.", + "bbox": [ + 511, + 198, + 906, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 513, + 431, + 697, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each image pair, we compute the geodesic error, defined as follows:", + "bbox": [ + 511, + 455, + 905, + 484 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {e r r} = \\arccos \\left(\\frac {t r \\left(\\mathbf {R} ^ {T} \\mathbf {R} ^ {*}\\right) - 1}{2}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 492, + 903, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{R}$ is the predicted rotation matrix, and $\\mathbf{R}^*$ is the ground truth relative rotation matrix. We report the median geodesic error (MGE) and relative rotation accuracy (RRA) for each test set and overlap ratio. RRA is reported over a predefined threshold $\\tau$ , indicating the percentage of image pairs with relative rotation error below $\\tau$ . We report this metric for $\\tau = 15^\\circ$ ( $\\mathrm{RRA}_{15}$ ) and $\\tau = 30^\\circ$ ( $\\mathrm{RRA}_{30}$ ).", + "bbox": [ + 511, + 536, + 905, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, as we discretize the space of rotations and estimate three rotation angles, each predicted as a probability distribution over $N$ bins, we also report the performance of the Top 5 predictions in our ablation study. The Top 5 prediction reports the lowest error, considering the top-5 peaks in the relative yaw prediction only, as we observe that errors are mostly a function of the relative yaw angles.", + "bbox": [ + 511, + 642, + 905, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 756, + 746, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The main quantitative results comparing our method to other methods on rotation estimation in the wild are reported in Table 2. In the supplementary material, we conduct an additional evaluation in the more constrained setting examined by prior work targeting extreme rotation estimation (i.e., images cropped from panoramas).", + "bbox": [ + 511, + 780, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Table 2, SIFT and LoFTR, which are correspondence-based methods, exhibit some robustness", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1066", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e81900b6e1f44b20a73b6e30ac58d8caf95356429ab9755d839210925da916ae.jpg", + "image_caption": [ + "Figure 5. Qualitative results on the wELP test set. We visualize the results of our model over different overlap levels, where the images on the left serve as the reference points, and their coordinate system determines the relative rotation, which defines the images on the right. The ellipsoids representing the ground truth are color-coded to match their respective images, with the estimated relative rotation illustrated by a cyan dashed line. As illustrated by the examples above, our method can accurately predict relative rotations for diverse image pairs containing varying appearances and intrinsic parameters. Please refer to the supplementary material for additional qualitative results." + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 851, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dbbb878b0485d7a9fbd49223b3cc75f9f4189569eee2979fac297e41c9581914.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsELPwELP
MGE↓RRA15↑RRA30↑MGE↓RRA15↑RRA30↑
LargeSIFT* [27]1.9592.395.32.9474.680.8
LoFTR* [38]1.7697.299.12.1385.293.8
DenseCorrVol [8]98.5125.933.3120.537.013.0
CascadedAtt [12]29.7542.750.0170.627.39.2
8PointViT [30]22.3331.964.8---
Dust3R [46]0.7799.799.91.0198.499.2
Ours2.4596.796.82.4197.597.9
SmallSIFT* [27]5.0764.771.37.2761.468.4
LoFTR* [38]2.7081.593.66.8066.681.2
DenseCorrVol [8]143.472.89.4125.733.19.4
CascadedAtt [12]148.440.03.0139.142.74.4
8PointViT [30]51.301.712.8---
Dust3R [46]1.9695.994.62.8089.894.4
Ours4.3588.389.04.4787.291.6
NoneSIFT* [27]121.942.75.4122.840.02.0
LoFTR* [38]---56.540.033.0
DenseCorrVol [8]77.109.027.082.042.913.7
CascadedAtt [12]69.698.423.178.607.520.8
Dust3R [46]114.3319.823.981.2115.426.9
Ours13.6252.759.726.9736.150.7
", + "bbox": [ + 91, + 345, + 480, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Rotation Estimation in the Wild. We evaluate performance over the sELP and wELP test sets, separately considering Large (top), Small (middle) and Non-overlapping (bottom) pairs. * indicates median errors are computed only over successful image pairs, for which these algorithms output a pose estimate (failure over more than $50\\%$ of the test pairs is shown in gray). Note that Dust3R was trained on images from wELP.", + "bbox": [ + 89, + 628, + 482, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "when handling highly overlapping Internet images, compared to methods that trained on images cropped from panoramas, achieving a median error of less than $3^{\\circ}$ for Large overlap pairs in both test sets. However, these methods rely on image overlap and may not always provide an output of estimated camera pose, as the geometric verification requires a sufficient number of detected inliers. Therefore, they struggle to produce reliable matches in cases of limited overlapping regions, as also observed in prior work", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "focusing on extreme scenarios [8, 12].", + "bbox": [ + 511, + 348, + 767, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DenseCorrVol, CascadedAtt and 8PointViT exhibit relatively poor performance on the ELP test sets, illustrating that models trained on images sampled from panoramicas cannot easily generalize to Internet photos. In the supplementary material, we show that our model significantly outperforms these prior work even when trained on the same dataset (i.e., images with a constant FoV and illumination cropped from panoramicas). As further discussed there, we believe this gap can be partially attributed to the usage of a pretrained LoFTR feature extractor, which is capable of encoding knowledge between Internet image pairs (which vary in their intrinsic and appearance).", + "bbox": [ + 509, + 364, + 906, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dust3R operates on Internet datasets without the need for calibration and performs exceptionally well on overlapping pairs, achieving the lowest median error and highest success rates for Large and Small overlap categories for sELP test set. However, this method is designed for overlapping pairs as it initializes the model using pretrained CroCo [47], which is trained to address cross-view completion problem from two overlapping views. Furthermore, its output consists of a unified dense point cloud for each pair of images. Due to its design for overlapping pairs, its performance on non-overlapping views is extremely low, also on the wELP test set, which contains images from scenes included in its training set. Moreover, in terms of size, Dust3R (with a DPT head) contains 577 million parameters. In contrast, our model is significantly more compact, with only 80 million parameters. As shown in the table, despite its smaller size, our model yields significantly improved performance over non-overlapping pairs on both test sets.", + "bbox": [ + 511, + 547, + 908, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 830, + 665, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct various ablations analyzing the effect of our progressive training scheme and other design choices,also reporting performance over the Top 5 predictions (consider", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1067", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24c18a512effaa132fe8c1390581fb13c506d1e92d0f29311f750d7aeb8fc445.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Ablation study, evaluating the effect of our progressive training scheme over the wELP test set. All experiments start with the cropped panoramas used in Cai et al. [8]." + ], + "table_body": "
OverlapTrain dataTop 1Top 5
MGE ↓RRA15 ↑RRA30 ↑MGE ↓RRA15 ↑RRA30 ↑
Large[8]13.6535.473.512.2261.484.7
+ΔFoV4.6179.781.14.4190.398.9
+ΔIm4.4690.492.44.4394.399.1
+ELP2.4197.597.92.4198.499.4
Small[8]55.283.729.129.8315.050.3
+ΔFoV12.9156.268.210.9766.085.4
+ΔIm11.4662.580.610.7368.091.0
+ELP4.4787.291.64.2491.197.2
None[8]74.9412.825.325.1126.158.8
+ΔFoV61.6225.038.416.8244.275.0
+ΔIm68.3125.036.116.2145.778.2
+ELP26.9736.150.712.8557.185.8
", + "bbox": [ + 91, + 88, + 480, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ering the top-5 peaks in the relative yaw prediction, as detailed in Section 5.2). Additional ablations, including an analysis of architectural components, are provided in the supplementary material.", + "bbox": [ + 93, + 339, + 480, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of our progressive training scheme. We conduct multiple experiments ablating the effect of our progressive training scheme. Our training process consists of four stages: initialization (following prior work [8]), incorporating multiple FoVs $(+\\Delta \\mathrm{FoV})$ , training with image-level appearance augmentations $(+\\Delta \\mathrm{Im})$ , and training with ExtremeLandmarkPairs pairs $(+ELP)$ .", + "bbox": [ + 93, + 402, + 480, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 illustrates the impact of each training stage on the results. As can be observed from the table, each training stage further refines the model's performance, often in a significant manner. For instance, the median error in the wELP small overlap test set decreased from $55.3^{\\circ}$ to $12.9^{\\circ}$ when FoV augmentations were added. Additionally, the ExtremeLandmarkPairs training set plays a crucial role in finalizing our training process, yielding a significant reduction in the median error (more than half) for small and non-overlapping scenarios in the wELP test set. While this median error remains relatively high for non-overlapping pairs, we observe that the top-5 scores show significant improvements, e.g., reducing the median error from $27.0^{\\circ}$ to $12.8^{\\circ}$ . This demonstrates that the model has learned this knowledge, although it cannot necessarily be recovered by the largest peak alone. In the supplementary material, we demonstrate that the intermediate training stages are indeed important and that the improved performance cannot be obtained with the ExtremeLandmarkPairs training set alone.", + "bbox": [ + 93, + 508, + 480, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, evidenced in Table 3, our model demonstrates strong generalization capabilities even when trained exclusively on panorama-cropped images. The model generalizes to the wELP test set significantly better (e.g., for Large overlap cases, our model achieves an MGE of $13.65^{\\circ}$ compared to values higher than $120^{\\circ}$ achieved by baselines), demonstrating that the improvement is not just due to our", + "bbox": [ + 93, + 795, + 480, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa6689a546af5aea062cda41d378ce8bb9d6bd32f04764d6e93ac0e913aed203.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Ablation study, evaluating the effect of the auxiliary channels added as input to our network. We train models without the keypoints and matches (KP) and without the segmentation maps (SM), comparing to our model over the wELP test set , after using a validation split that is balanced over the overall angle." + ], + "table_body": "
OverlapKPSMTop 1Top 5
MGE ↓RRA15↑RRA30↑MGE ↓RRA15↑RRA30↑
Large××2.1897.498.12.1897.498.1
×✓2.3097.097.42.3098.599.4
✓×2.4497.698.32.3198.499.7
✓✓2.4197.597.92.4198.499.4
Small××4.5087.991.64.5087.991.7
×✓4.4988.192.04.4691.296.7
✓×4.4187.592.24.3291.997.6
✓✓4.4787.291.64.2491.197.2
None××48.8134.044.112.5657.584.6
×✓43.0731.244.213.9953.583.2
✓×41.3935.346.813.0456.986.2
✓✓26.9736.150.712.8557.185.8
", + "bbox": [ + 516, + 89, + 903, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "progressive training scheme. We further examined architectural differences by conducting an additional ablation study, applying our progressive training scheme to the baseline models (see Table 9 in the supplementary materials). This experiment revealed that prior models are not directly suitable for real-world applications, as these baseline models showed significantly poorer performance across all metrics.", + "bbox": [ + 516, + 372, + 903, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of adding auxiliary channels. We ablate the effect of adding auxiliary in Table 4, training models without keypoints and matches (KP) or the segmentation map (SM) provided as additional inputs. As illustrated in the table, for non-overlapping cases, these auxiliary channels boost performance almost across all metrics. In particular, both channels play a role in reducing the errors (reducing the median error from over $40^{\\circ}$ to $27.0^{\\circ}$ ).", + "bbox": [ + 516, + 481, + 903, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 516, + 614, + 630, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a method and benchmark dataset for estimating relative 3D rotations between pairs of (possibly) non-overlapping RGB images. Our approach extends prior work addressing extreme rotations to real-world data that exhibit variation in both appearance in intrinsic camera parameters. While our model shows promising results on real-world Internet image pairs, it also highlights the inherent difficulty of the underlying task, suggesting that considerable progress can be achieved with future techniques that leverage our dataset. Our paired data could also potentially serve for exploring the challenging task of estimating extreme translations in real-world settings. Future work could also consider incorporating more views for enhancing performance in such extreme non-overlapping scenarios.", + "bbox": [ + 516, + 641, + 903, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was partially supported by ISF (grant number 2510/23).", + "bbox": [ + 516, + 869, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1068", + "bbox": [ + 485, + 946, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In European Conference on Computer Vision, pages 192-209. Springer, 2022. 2", + "[2] Roberto Cipolla Alex Kendall, Matthew Grimes. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV -International Conference on Computer Vision, pages 2938-2946, 2015. 2, 4", + "[3] Lionel Baboud, Martin Čadík, Elmar Eisemann, and Hans-Peter Seidel. Automatic photo-to-terrain alignment for the annotation of mountain pictures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41–48. IEEE, 2011. 2", + "[4] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In European Conference on Computer Vision, pages 751-767. Springer, 2018. 2", + "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417. Springer, 2006. 2", + "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 1", + "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR IEEE / CVF Computer Vision and Pattern, pages 18392-18402, 2023. 2, 5, 6", + "[8] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[9] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 4", + "[10] Yaron Caspi and Michal Irani. Aligning non-overlapping sequences. IJCV, 48(1):39-51, 2002. 2", + "[11] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 2", + "[12] Shay Dekel, Yosi Keller, and Martin Cadik. Estimating extreme 3d image rotations using cascaded attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2588-2598, 2024. 1, 2, 3, 4, 6, 7", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2", + "[15] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. Camnet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2871-2880, 2019.", + "[16] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2", + "[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17775, 2023. 2", + "[18] Johan Edstedt, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. arXiv preprint arXiv:2308.08479, 2023. 2", + "[19] Xie Enze, Wang Wenhai, Yu Zhiding, Anandkumar Anima, M. Alvarez Jose, and Luo Ping. Segformer: Simple and efficient design for semantic segmentation with transformers. In Conference on Neural Information Processing Systems (NEURIPS), 2021. 4", + "[20] Zhiwen Fan, Panwang Pan, Peihao Wang, Yifan Jiang, Dejia Xu, Hanwen Jiang, and Zhangyang Wang. Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727, 2023. 2", + "[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM (CACM), 24(6):381-395, 1981. 2", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6", + "[23] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. 2", + "[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2, 4", + "[25] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 2", + "[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching atokens are frozen. arXiv preprint arXiv:2306.13643, 2023. 2", + "[27] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 60(2):91-110, 2004. 2, 6, 7", + "[28] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1069", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 3, 5, 6", + "[29] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1", + "[30] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2, 6, 7", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6", + "[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2", + "[33] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 2", + "[34] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1", + "[35] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8601-8610, 2018. 1", + "[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2", + "[37] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 2", + "[38] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 4, 6, 7", + "[39] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 2, 4", + "[40] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[42] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5722-5731, 2021. 1", + "[43] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 2", + "[44] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 2", + "[45] Qianqian Wang, Xiaowei Zhou, Bharath Hariharan, and Noah Snavely. Learning feature descriptors using camera pose supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 757-774. Springer, 2020. 2", + "[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. arXiv preprint arXiv:2312.14132, 2023. 2, 6, 7", + "[47] Philippe Weinzaepfel, Vincent Leroy, Thomas Lucas, Romain Brégier, Yohann Cabon, Vaibhav Arora, Leonid Antsfeld, Boris Chidlovskii, Gabriela Csurka, and Jérôme Revaud. Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. Advances in Neural Information Processing Systems, 35:3502-3516, 2022. 7", + "[48] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2", + "[49] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2", + "[50] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 2", + "[51] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Reply: Predicting probabilistic relative rotation for single objects in the wild. In European Conference on Computer Vision, pages 592-611. Springer, 2022.", + "[52] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 2" + ], + "bbox": [ + 516, + 90, + 905, + 867 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1070", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_model.json b/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7b7252dbc2fdd118fcdc8b16638ff438728ac71 --- /dev/null +++ b/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.291, + 0.131, + 0.709, + 0.151 + ], + "angle": 0, + "content": "Extreme Rotation Estimation in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.18, + 0.796, + 0.225 + ], + "angle": 0, + "content": "Hana Bezalel\\(^{1}\\) Dotan Ankri\\(^{1}\\) Ruojin Cai\\(^{2}\\) Hadar Averbach-Elor\\(^{1,2}\\) \n\\(^{1}\\)Tel Aviv University \\(^{2}\\)Cornell University" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.234, + 0.755, + 0.248 + ], + "angle": 0, + "content": "https://tau-vailab.github.io/ExtremeRotationsInTheWild/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.283, + 0.327, + 0.3 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.485, + 0.588 + ], + "angle": 0, + "content": "We present a technique and benchmark dataset for estimating the relative 3D orientation between a pair of Internet images captured in an extreme setting, where the images have limited or non-overlapping field of views. Prior work targeting extreme rotation estimation assume constrained 3D environments and emulate perspective images by cropping regions from panoramic views. However, real images captured in the wild are highly diverse, exhibiting variation in both appearance and camera intrinsics. In this work, we propose a Transformer-based method for estimating relative rotations in extreme real-world settings, and contribute the ExtremeLandmarkPairs dataset, assembled from scene-level Internet photo collections. Our evaluation demonstrates that our approach succeeds in estimating the relative rotations in a wide variety of extreme-view Internet image pairs, outperforming various baselines, including dedicated rotation estimation techniques and contemporary 3D reconstruction methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.222, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.84 + ], + "angle": 0, + "content": "The problem of estimating the relative 3D orientation between a pair of images is embodied in fundamental computer vision tasks, such as camera localization [6, 34, 35] and 3D reconstruction [29, 36, 42]. Establishing pixel correspondences (either explicitly or implicitly) is typically a prerequisite for computing the relative rotation between the images. Correspondences, however, cannot be extracted in extreme settings where the images have little to no overlap. As dense imagery may not necessarily be available for many practical applications, a natural question arises: How can we estimate the relative rotation between non-overlapping RGB images, without the use of additional data (such as depth or temporal information)?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We have recently seen pioneering efforts addressing the task of relative rotation estimation in such extreme non-overlapping settings [8, 12]. Prior work has proposed end-to-end neural architectures, demonstrating that hidden cues," + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.283, + 0.88, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.455, + 0.908, + 0.541 + ], + "angle": 0, + "content": "Figure 1. Given a pair of (possibly) non-overlapping images captured in the wild—e.g., under arbitrary illumination and intrinsic camera parameters—such as the images of the Dam Square in Amsterdam depicted in red and blue boxes above*, our technique estimates the relative 3D rotation between the images. *The background panorama is illustrated for visualization purposes only." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.734 + ], + "angle": 0, + "content": "such as vanishing points or the directions of cast shadows, can implicitly guide the model for inferring the relative orientation between the images. To facilitate learning and evaluation, datasets constructed from panoramic views were adopted. These datasets emulate perspective views by cropping sub-areas from these panoramas, enabling generation of image pairs with various degrees of overlap. However, while such emulated views perhaps capture some of the challenges associated with extreme-view imagery, are they sufficient for representing real images—particularly, images captured in the wild?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present a new approach that tackles the problem of extreme rotation estimation in the wild. Consider the boxed images in Figure 1. Internet (i.e., in the wild) images may vary due to a wide range of factors, including transient objects, weather conditions, time of day, and the cameras' intrinsic parameters. To explore this problem, we introduce a new dataset, ExtremeLandmarkPairs (ELP), assembled from publicly-available scene-level Internet image collections. We observe that the set of real extreme-view image pairs is limited, as Internet datasets are typically scene-centric, with nearby cameras commonly" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1061" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.481, + 0.226 + ], + "angle": 0, + "content": "capturing overlapping views. Therefore, to facilitate training, we propose a progressive learning scheme that leverages and augments images cropped from panoramic views, allowing for gradually generalizing the model onto real Internet data. In particular, we construct datasets with varying field of views, that better resemble the distribution of real data, and perform image-level appearance augmentations by leveraging recent advancements in text-to-image Diffusion models [7, 22, 31]." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.228, + 0.481, + 0.453 + ], + "angle": 0, + "content": "To estimate extreme rotations in the wild, we propose a Transformer-based model that is provided with auxiliary channels, including the spatial distribution of local keypoints and matches and semantic segmentation maps, allowing for better reasoning over real image pairs with little or no overlap. Our results demonstrate that our model can accurately predict the relative rotations for a wide variety of extreme-view image pairs that vary in illumination, dynamic regions, and intrinsic parameters. We conduct extensive experiments, quantifying performance both over real Internet data and also over emulated perspective images cropped from panoramic views. Our evaluation shows that our model significantly improves over strong baselines when considering real images, while achieving comparable performance over emulated perspective image pairs." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.468, + 0.23, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.493, + 0.481, + 0.658 + ], + "angle": 0, + "content": "Relative pose estimation is a fundamental task in computer vision, typically studied for overlapping camera views. Traditionally, this task has been divided into two stages: correspondence estimation from local feature matching, followed by geometry-based pose estimation. In recent years, feature matching methods have advanced from using heuristic feature descriptors [5, 27, 32] and RANSAC-based matches [21] to learning-based feature extraction [4, 14-16, 18, 40, 45] and matching techniques [3, 26, 33], with several methods performing both feature extraction and matching using unified learning-based frameworks [4, 17, 38]." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.66, + 0.481, + 0.734 + ], + "angle": 0, + "content": "These methods are generally invariant to changes in illumination and appearance, demonstrating robust performance across various scene scales and also over in-the-wild datasets. However, their reliability diminishes in extreme view scenarios due to their dependence on visual overlap." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.736, + 0.481, + 0.9 + ], + "angle": 0, + "content": "Extreme-view scenarios lacking local pixelwise correspondences necessitate the use of end-to-end learning-based pose estimation methods which directly infer the 3D relationships and geometry from sparse and extreme-view images. Indeed, large-scale 3D object datasets have paved the way for learning-based methods which estimate camera pose directly from sparse views [20, 25, 37, 43, 44, 50-52]. However, these methods primarily concentrate on object-centric scenes, under experimental settings that typically assume similar lighting and camera intrinsics for the input views. Furthermore, these methods often utilize bounding" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.903, + 0.121 + ], + "angle": 0, + "content": "box inputs defining the object of interest, which is less suitable in the case of images depicting large-scale scenes." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.123, + 0.904, + 0.286 + ], + "angle": 0, + "content": "Several prior works address a sparse view setting at scene-scale. In particular, Chen et al. [11] propose to learn a discrete distributions of pose space, Agarwala et al. [1] simplify scene reconstruction using a plane representation, and Rockwell et al. [30] introduce an inductive bias of the 8-point algorithm into a vision transformer architecture. Recently, models that directly predict pixel-aligned point maps from input pairs (or sparse image collections), such as DUSt3R [46] and Mast3R [23], have demonstrated promising results on pose estimation and scene-scale 3D reconstruction of Internet data with a wide baseline." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.288, + 0.904, + 0.558 + ], + "angle": 0, + "content": "Camera pose estimation for non-overlapping views presents a greater challenge. Earlier efforts [10] explored searching for consistent temporal behavior. Several work utilize pairwise RGB and depth scan data to estimate relative pose among such extreme pairs [48, 49]. Cai et al. [8] tackle pose estimation for non-overlapping views without the use of additional data, by introducing a learning-based network leveraging cross-correlation volume to exploit implicit cues. This correlation volume is later enhanced through the integration of transformer attention modules [12]. Nonetheless, these approaches assume constrained 3D environments, including the assumption of consistent lighting and camera intrinsics, and are designed for camera distribution of emulated perspective views cropped from panoramas. In this work, we aim to address pose estimation for realistic in-the-wild non-overlapping image pairs, enhancing the applicability of extreme pose estimation to Internet photos and real-world data." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.572, + 0.837, + 0.588 + ], + "angle": 0, + "content": "3. The ExtremeLandmarkPairs Dataset" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.598, + 0.904, + 0.748 + ], + "angle": 0, + "content": "Prior works on extreme pose estimation use panoramic views, cropping from it sub-areas to emulate perspective views [8, 12]. To evaluate and train models on real perspective image pairs, we propose a new benchmark and dataset, ExtremeLandmarkPairs (ELP), constructed from Internet image pairs from the MegaDepth [24], Cambridge Landmarks [2], and MegaScenes [39] datasets. In this section, we first describe the dataset construction procedure (Section 3.1), and then present details regarding dataset size and train and test splits (Table 1)." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.757, + 0.712, + 0.771 + ], + "angle": 0, + "content": "3.1. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.781, + 0.904, + 0.9 + ], + "angle": 0, + "content": "To construct a dataset of real perspective image pairs with varying degrees of overlap, we leverage available scene-level training data. Existing Internet image collections typically contain camera poses (predicted up to scale), which are determined using Structure-from-Motion (SfM) algorithms, such as COLMAP [36]. In what follows, we describe how we extract real image pairs from this data, which can then be used for training and evaluating models." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.947, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1062" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.094, + 0.291, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.21, + 0.285, + 0.236 + ], + "angle": 0, + "content": "(a) Reconstruction and camera distribution of all images" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.094, + 0.478, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.209, + 0.47, + 0.236 + ], + "angle": 0, + "content": "(b) Camera distribution of large-overlapping pairs" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.244, + 0.48, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.525, + 0.434, + 0.539 + ], + "angle": 0, + "content": "(c) Camera distribution of non-overlapping pairs" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Figure 2. Camera distribution of the Vatican, Rome scene from the ExtremeLandmarkPairs Dataset. We construct a dataset of real perspective image pairs with predominant rotational motion shown in (b) and (c) from the dense imagery reconstruction in (a)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.644 + ], + "angle": 0, + "content": "Identifying Pairs with Predominant Rotational Motion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Prior works targeting relative rotation estimation, in particular for non-overlapping views, mostly utilize panoramic views, focusing on image pairs with purely rotational motion. Pairs belonging to real image collections almost always contain a non-negligible translation component. Furthermore, unlike in the StreetLearn [28] dataset used by prior work [8, 12] that provides exact translation values between consecutive panoramas which allows for filtering pairs with predominant rotational motions, reconstructed relative poses are only provided up to scale. The scale varies among different reconstructed scenes, and therefore there's no global threshold on the relative translation values which can be used for identifying pairs with predominant rotational motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To automatically identify such pairs, we observe that available Internet collections require the existence of dense imagery, to compensate for the vast number of unknowns" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "in the SfM optimization. We therefore construct mutual nearest neighbors edge-weighted graphs, with one graph per landmark. In each graph \\( G \\), nodes \\( v \\in V \\) correspond to images, and two images are connected by an edge \\( e \\in E \\) if they are both among each other's \\( K \\) nearest neighbors, considering L2 distances between their translations (\\( K \\) is empirically set to 5). Note that images captured from sparser (outlier) regions in space are unlikely to be within the mutual \\( K \\) nearest neighbors of images captured within denser regions, and hence won't be included in \\( G \\). Finally, we select a subset of image pairs containing relatively small distances from each scene graph \\( G \\), yielding a set of image pairs with predominant rotational motion; see the supplementary material for additional details." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.306, + 0.905, + 0.427 + ], + "angle": 0, + "content": "Extracting Level of Overlap. Following prior work [8, 12], we are interested in training and evaluating models according to three different categories: Large, Small and None, indicating image pairs with a varying amount of overlap. However, unlike prior work that use cropped images with a fixed \\(90^{\\circ}\\) FoV, Internet images contain varying FoV values. Thus, the relative rotation angle is not sufficient for extracting the pair's overlap level." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.905, + 0.487 + ], + "angle": 0, + "content": "Denote the FoV values of image \\(i\\in [1,2]\\) as \\([\\mathrm{fov}_x^i,\\mathrm{fov}_y^i ]\\) We can parameterize a 3D rotation matrix \\(\\mathbf{R}\\) using three Euler angles \\([\\alpha ,\\beta ,\\gamma ]\\) denoting the relative roll, pitch and yaw angles, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.502, + 0.904, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} (\\alpha , \\beta , \\gamma) = \\mathbf {R} _ {z} (\\alpha) \\mathbf {R} _ {y} (\\beta) \\mathbf {R} _ {x} (\\gamma), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.905, + 0.566 + ], + "angle": 0, + "content": "following Dense Correlation Volumes' coordinate system. We use the following conditions to determine the overlap level \\(o\\):" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.587, + 0.905, + 0.651 + ], + "angle": 0, + "content": "\\[\no = \\left\\{ \\begin{array}{l l} L a r g e & | \\gamma | < \\frac {\\operatorname {f o v} _ {x} ^ {1} + \\operatorname {f o v} _ {x} ^ {2}}{4} \\wedge | \\beta | < \\frac {\\operatorname {f o v} _ {y} ^ {1} + \\operatorname {f o v} _ {y} ^ {2}}{4} \\\\ N o n e & | \\gamma | > \\frac {\\operatorname {f o v} _ {x} ^ {1} + \\operatorname {f o v} _ {x} ^ {2}}{2} \\wedge | \\beta | > \\frac {\\operatorname {f o v} _ {y} ^ {1} + \\operatorname {f o v} _ {y} ^ {2}}{2} \\\\ S m a l l & e l s e \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.763 + ], + "angle": 0, + "content": "In other words, pairs with relative yaw and pitch angles that are smaller than a quarter of the average corresponding FoV values are considered as pairs with a large overlap ratio. Likewise, pairs with relative yaw and pitch angles that are larger than half the average corresponding FoV values are considered non-overlapping pairs. Pairs in-between these conditions are considered pairs with a small overlap ratio." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Additional Filtering. In the scene-scale datasets we explore, large FoV disparities could result in one image focusing on specific architectural details like a statue, while the other captures a much broader scene perspective, further complicating the problem of estimating the relative rotations. To extract pairs consistent in scale, we limit the difference between the FoV values to be at most \\(5^{\\circ}\\). Furthermore, as most images contain small roll values, we rotate the scenes to match the gravity axis and horizontal axis" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1063" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.09, + 0.482, + 0.183 + ], + "angle": 0, + "content": "
SubsetSource#Pairs
#ScenesLargeSmallNoneTotal
Train[39]588333430136842948176595
Validation[39]515339871041308238
Validation Balanced[39]1779255707854
sELP[2]6251282719615300
wELP[24]1727008296434172
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Table 1. ExtremeLandmarkPairs Dataset Statistics. Above, we report the number of image pairs extracted for each overlap level, split into train and test (with sELP denoting a single camera setting and wELP denoting the \"in the wild\" setting)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.33 + ], + "angle": 0, + "content": "and filtered images with rolls exceeding \\(10^{\\circ}\\). Finally, we excluded images captured from an aerial perspective, by filtering images with translation along y-axis exceeding a global threshold, empirically set to 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.339, + 0.303, + 0.357 + ], + "angle": 0, + "content": "3.2. Dataset Size and Splits" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.468 + ], + "angle": 0, + "content": "We apply the procedure described above over landmarks from the MegaScenes [39] dataset to create a training set and a validation set. As we focus on outdoor scenes in our work, we filter reconstructions that capture indoor scenes. We obtain a total of nearly 34K non-overlapping pairs originating from over 2K unique landmarks. The validation set was also balanced the set over the overall angle." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.483, + 0.603 + ], + "angle": 0, + "content": "For evaluation, we create two test sets, to separately examine image pairs captured in a single camera setting with constant illumination (sELP) and image pairs captured in the wild (wELP). Image pairs in the sELP test set contain images from the Cambridge Landmarks [2] dataset, which contains videos capturing six different landmarks. Image pairs in the wELP test set contain images from the MegaDepth dataset [24], which contains Internet photos from Flickr for a set of large-scale landmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.831 + ], + "angle": 0, + "content": "As there is some overlap between the landmarks in MegaScenes [39] and MegaDepth [24], we performed additional filtering, ensuring no overlap exists between the train and test sets. Furthermore, we filter test pairs if one of the images is highly transient—i.e., if transient objects dominate the image. We quantify this using a pretrained segmentation model [19], filtering images containing transient objects in over \\(40\\%\\) of the pixels. We also manually validate non-overlapping image pairs, allowing to further filter unlabeled objects, such as a stage or a market stand. Finally, for non-overlapping image pairs, we observed that the overall relative rotations are highly imbalanced, and therefore we balance this set by the overall relative angle. Table 1 summarizes the number of image pairs and landmarks used for both training and test." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.182, + 0.86 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given a pair of Internet images with (possibly) extreme relative motion, we estimate the relative rotation \\(\\mathbf{R}\\) between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "the images. Following prior works on extreme rotation estimation [8, 12], we assume a camera-centric setting, where the two cameras have limited translation. However, our approach departs from prior works by operating on outdoor images captured by a crowd of photographers, with varying intrinsic parameters as well as appearances—\\(e.g.\\), due to illumination changes and dynamic objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.306 + ], + "angle": 0, + "content": "Our model (detailed in Section 4.1; see Figure 3 for an overview) outputs three discrete Euler angles, denoting the relative roll, pitch and yaw angles. As illustrated in prior work [8], this parameterization enables using a simple cross-entropy loss for training. In Section 4.2, we describe our progressive learning scheme, allowing for gradually adapting the model to extreme Internet imagery." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.316, + 0.6, + 0.331 + ], + "angle": 0, + "content": "4.1. Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.461 + ], + "angle": 0, + "content": "We extract image features using a pretrained LoFTR model [38]. In contrast to the features extracted using common convolutional neural networks pretrained on ImageNet [13], LoFTR is a Transformer-based model trained on Internet pairs, with the goal of extracting local feature matches – a setting and task which is highly related to the one we address in our work, thus enabling extraction of better (i.e., more relevant) features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.703 + ], + "angle": 0, + "content": "As we are interested in designing a network that can also reason over image pairs with little or no overlap, we combine the extracted features with additional auxiliary channels; see Figure 3 (bottom left). These include keypoint and pairwise matches masks, utilized previously for disambiguating images for similar structures [9]. Intuitively, knowledge over pairwise matches can assist the model in cases of small overlap and for generalizing across different camera intrinsic properties. We also incorporate a segmentation map as an additional auxiliary channel, which segments images into several categories (such as sky, building, road and sidewalk). This channel allows for identifying additional cues, such as the skyline or transient objects, which can aid in determining the rotation for nonoverlapping pairs. In Section 5, we demonstrate the benefit of incorporating these auxiliary channels in our model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We then reshape extracted features (and auxiliary channels) to tokens, concatenating these image tokens with learnable Euler angle position embeddings. These tokens are processed by our Rotation Estimation Transformer module, which uses a transformer decoder architecture [41]. The output Euler angle tokens obtain information from image features and auxiliary channels within transformer attention modules. These tokens are then processed by three different prediction heads (one per angle, denoted as MLP in Figure 3). Each output prediction head uses as input averaged image tokens and one of the output Euler angle tokens, which provides the model with additional angle-specific information, allowing for achieving improved performance," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1064" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.099, + 0.877, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.273, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 3. Method architecture. Given a pair of input Internet images, we extract image features using pretrained LoFTR. These features are combined with auxiliary channels, including keypoint and pairwise matches masks, and segmentation maps (visualized on the bottom left). These image features are reshaped into tokens and concatenated with Euler angle position embeddings, which are then processed by our Rotation Estimation Transformer module. The output Euler angle tokens and averaged image tokens are concatenated and processed by MLPs to predict the probability distribution of Euler angles, representing the relative 3D rotation between the input images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.483, + 0.418 + ], + "angle": 0, + "content": "as we show in the supplementary material. The prediction heads output a probability distribution over \\( N = 360 \\) bins, capturing an angle in the range \\([-180^{\\circ}, 180^{\\circ}]\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.428, + 0.2, + 0.444 + ], + "angle": 0, + "content": "4.2. Learning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.483, + 0.649 + ], + "angle": 0, + "content": "As detailed in Section 3, we assemble real image pairs from Internet imagery, which can be used for training and evaluating models. However, even with our proposed ExtremeLandmarkPairs dataset, the set of real image pairs is limited—e.g., only \\(\\sim 36\\mathrm{K}\\) non-overlapping image pairs are extracted, as available Internet datasets are typically scene-centric, with nearby cameras usually capturing overlapping views. Therefore, in what follows, we propose a progressive learning scheme, which leverages panoramic images, and allows for gradually generalizing the model to images captured in-the-wild. All learning stages are optimized using a cross-entropy loss for each Euler angle prediction. Additional details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.653, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Initialization. We begin by training our model on the perspective views cropped from panoramic images, using the data created by Cai et al [8]. Specifically, we use the image pairs cropped from panoramic images included in the StreetLearn [28] dataset, depicting various streets in Manhattan. As further detailed in Cai et al [8], this training set includes roughly 1M image pairs sampled from the same panorama, split according to the overlap level." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Training with Data Augmentations. We observe that the dataset constructed by prior work could be augmented to better capture the distribution of image pairs captured inthe-wild. In particular, we focus on two types of data augments, which we elaborate on next: (i) field of view (FoV) augmentations (denoted henceforth as \\(\\Delta \\mathrm{FoV}\\)) and (ii) image-level appearance augmentations (denoted henceforth as \\(\\Delta \\mathrm{Im}\\)). In Section 5, we demonstrate how both types" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.367, + 0.613, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.367, + 0.714, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.367, + 0.808, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.367, + 0.904, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.444, + 0.613, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.444, + 0.714, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.444, + 0.808, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.444, + 0.904, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.533, + 0.704, + 0.547 + ], + "angle": 0, + "content": "busy street" + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.533, + 0.78, + 0.546 + ], + "angle": 0, + "content": "night" + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.533, + 0.891, + 0.546 + ], + "angle": 0, + "content": "Christmas" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Figure 4. Augmenting perspective images cropped from panoramic views with image-level appearance modifications. Given an input image (left) and a target text prompt \"Make it \\(\\langle w\\rangle\\) (\" \\(\\langle w\\rangle\\) is specified above), we use a conditional Diffusion model [7] to create semantic appearance augmentations which modify both the global image characteristics as well as local image regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.905, + 0.7 + ], + "angle": 0, + "content": "of augmentations improve the model's ability in generalizing to real Internet scenes." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Rather than cropping perspective images with a fixed \\(90^{\\circ}\\) FoV as was done in prior work, we analyze the FoV values of the images belonging to the ELP training set. We compute the mean and standard-deviation values, denoted as \\(\\mu\\) and \\(\\sigma\\), respectively. We then construct new perspective images by sampling from a Gaussian distribution \\(G(\\mu, a \\cdot \\sigma)\\) that resembles the distribution of real data, setting \\(a = 1.5\\) for obtaining a more diverse set, which also bears higher similarity to the perspective images used during initialization. We also allow for FoV differences (of up to \\(5^{\\circ}\\)) between the two images paired together. Additionally, rather than providing the model with the full content within these regions, we construct crops with various aspect ratios to fur" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.29, + 0.107 + ], + "angle": 0, + "content": "ther emulate real image pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.364 + ], + "angle": 0, + "content": "In-the-wild images differ not only in their intrinsic parameters, but also in their appearance - due to presence of varying illumination and dynamic objects. Therefore, in addition to augmenting the dataset with images of various field of views, we also perform image-level appearance augmentations, leveraging recent advancements in text-to-image Diffusion models [7, 22, 31]. In particular, we apply the conditional InstructPix2Pix [7] model on a subset of our data, using a set of text prompts of the format \"Make it \\(\\langle w\\rangle\\) \", where \\(\\langle w\\rangle\\) captures diverse appearance modifications. See Figure 4 for a few illustrative examples; note how some of these modify the global illumination (e.g., altering the time of day or the season) while others yield local edits (such as adding more people to make it \"a busy street\"). We performed additional filtering on the augmented set, to avoid significant edits that also altered the structure of the scene; see the supplementary material for additional details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.442 + ], + "angle": 0, + "content": "Training on Real Data. Finally, we finetune the model over real image pairs from our proposed ELP dataset. To prioritize learning over extreme image pairs, we first pass a batch of non-overlapping image pairs, followed by a batch of overlapping image pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.312, + 0.469 + ], + "angle": 0, + "content": "5. Results and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.539 + ], + "angle": 0, + "content": "To validate our method, we compare our model with prior relative rotation estimation methods on the sELP and wELP test sets. In particular, our experiments seek to answer the following questions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.54, + 0.483, + 0.568 + ], + "angle": 0, + "content": "- How well do previous methods perform on our proposed task of extreme rotation estimation in the wild?" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.569, + 0.483, + 0.613 + ], + "angle": 0, + "content": "- How important is our progressive training scheme, and to what extent can the improved performance be attributed to training on the ExtremeLandmarkPairs dataset?" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.614, + 0.384, + 0.629 + ], + "angle": 0, + "content": "- What is the impact of our design choices?" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.54, + 0.483, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.734 + ], + "angle": 0, + "content": "We also present qualitative results over different overlap levels in Figure 5. These results illustrate our model's performance, and its robustness to varying illumination conditions and transient objects (such as the screen and the beer cup in the rightmost examples). Please refer to the supplementary material for many more qualitative results, including visualizations for baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.2, + 0.756 + ], + "angle": 0, + "content": "5.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare our method to correspondence-based techniques, dedicated (end-to-end) relative rotation estimation techniques as well as recent relative pose estimation techniques. Specifically, we consider two correspondence-based methods: SIFT [27] and LoFTR [38]. These methods extract keypoint matches to compute an essential matrix using the RANSAC algorithm and calculate the rotation that stems from the essential matrix. We also compare against two prior works specifically targeting extreme" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "rotation estimation using end-to-end deep learning frameworks: DenseCorrVol [8] and CascadedAtt [12]. These methods are trained and evaluated on images sampled from panoramic views from StreetLearn [28]. Results for CascadedAtt are reported over a reproduced model, as a pretrained model is not provided; additional details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.425 + ], + "angle": 0, + "content": "We also consider 8PointViT [30], which estimates the relative pose between two images with a Vision Transformer that is modified to be close to the eight point algorithm. Following prior work [12], we only report results over overlapping settings for this baseline. Furthermore, as it assumes a single camera setting, we evaluate it only on the sELP test set. Finally, we consider Dust3R [46], a recent technique for dense and unconstrained stereo 3D reconstruction. While this work does not specifically target the setting of extreme rotation estimation, we examine to what extent they can be adapted for this task in the wild. Note that Dust3R was trained on MegaDepth. Since our wELP test set is constructed from image pairs from this dataset, it is not entirely fair comparison; therefore, Dust3R's results on the wELP test set are highlighted in gray." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.433, + 0.698, + 0.447 + ], + "angle": 0, + "content": "5.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.485 + ], + "angle": 0, + "content": "For each image pair, we compute the geodesic error, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.493, + 0.905, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\operatorname {e r r} = \\arccos \\left(\\frac {t r \\left(\\mathbf {R} ^ {T} \\mathbf {R} ^ {*}\\right) - 1}{2}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.643 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}\\) is the predicted rotation matrix, and \\(\\mathbf{R}^*\\) is the ground truth relative rotation matrix. We report the median geodesic error (MGE) and relative rotation accuracy (RRA) for each test set and overlap ratio. RRA is reported over a predefined threshold \\(\\tau\\), indicating the percentage of image pairs with relative rotation error below \\(\\tau\\). We report this metric for \\(\\tau = 15^\\circ\\) (\\(\\mathrm{RRA}_{15}\\)) and \\(\\tau = 30^\\circ\\) (\\(\\mathrm{RRA}_{30}\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Additionally, as we discretize the space of rotations and estimate three rotation angles, each predicted as a probability distribution over \\(N\\) bins, we also report the performance of the Top 5 predictions in our ablation study. The Top 5 prediction reports the lowest error, considering the top-5 peaks in the relative yaw prediction only, as we observe that errors are mostly a function of the relative yaw angles." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.747, + 0.773 + ], + "angle": 0, + "content": "5.3.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.871 + ], + "angle": 0, + "content": "The main quantitative results comparing our method to other methods on rotation estimation in the wild are reported in Table 2. In the supplementary material, we conduct an additional evaluation in the more constrained setting examined by prior work targeting extreme rotation estimation (i.e., images cropped from panoramas)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Table 2, SIFT and LoFTR, which are correspondence-based methods, exhibit some robustness" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.089, + 0.852, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on the wELP test set. We visualize the results of our model over different overlap levels, where the images on the left serve as the reference points, and their coordinate system determines the relative rotation, which defines the images on the right. The ellipsoids representing the ground truth are color-coded to match their respective images, with the estimated relative rotation illustrated by a cyan dashed line. As illustrated by the examples above, our method can accurately predict relative rotations for diverse image pairs containing varying appearances and intrinsic parameters. Please refer to the supplementary material for additional qualitative results." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.627 + ], + "angle": 0, + "content": "
MethodsELPwELP
MGE↓RRA15↑RRA30↑MGE↓RRA15↑RRA30↑
LargeSIFT* [27]1.9592.395.32.9474.680.8
LoFTR* [38]1.7697.299.12.1385.293.8
DenseCorrVol [8]98.5125.933.3120.537.013.0
CascadedAtt [12]29.7542.750.0170.627.39.2
8PointViT [30]22.3331.964.8---
Dust3R [46]0.7799.799.91.0198.499.2
Ours2.4596.796.82.4197.597.9
SmallSIFT* [27]5.0764.771.37.2761.468.4
LoFTR* [38]2.7081.593.66.8066.681.2
DenseCorrVol [8]143.472.89.4125.733.19.4
CascadedAtt [12]148.440.03.0139.142.74.4
8PointViT [30]51.301.712.8---
Dust3R [46]1.9695.994.62.8089.894.4
Ours4.3588.389.04.4787.291.6
NoneSIFT* [27]121.942.75.4122.840.02.0
LoFTR* [38]---56.540.033.0
DenseCorrVol [8]77.109.027.082.042.913.7
CascadedAtt [12]69.698.423.178.607.520.8
Dust3R [46]114.3319.823.981.2115.426.9
Ours13.6252.759.726.9736.150.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Table 2. Rotation Estimation in the Wild. We evaluate performance over the sELP and wELP test sets, separately considering Large (top), Small (middle) and Non-overlapping (bottom) pairs. * indicates median errors are computed only over successful image pairs, for which these algorithms output a pose estimate (failure over more than \\(50\\%\\) of the test pairs is shown in gray). Note that Dust3R was trained on images from wELP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "when handling highly overlapping Internet images, compared to methods that trained on images cropped from panoramas, achieving a median error of less than \\(3^{\\circ}\\) for Large overlap pairs in both test sets. However, these methods rely on image overlap and may not always provide an output of estimated camera pose, as the geometric verification requires a sufficient number of detected inliers. Therefore, they struggle to produce reliable matches in cases of limited overlapping regions, as also observed in prior work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.768, + 0.364 + ], + "angle": 0, + "content": "focusing on extreme scenarios [8, 12]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.365, + 0.907, + 0.546 + ], + "angle": 0, + "content": "DenseCorrVol, CascadedAtt and 8PointViT exhibit relatively poor performance on the ELP test sets, illustrating that models trained on images sampled from panoramicas cannot easily generalize to Internet photos. In the supplementary material, we show that our model significantly outperforms these prior work even when trained on the same dataset (i.e., images with a constant FoV and illumination cropped from panoramicas). As further discussed there, we believe this gap can be partially attributed to the usage of a pretrained LoFTR feature extractor, which is capable of encoding knowledge between Internet image pairs (which vary in their intrinsic and appearance)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.82 + ], + "angle": 0, + "content": "Dust3R operates on Internet datasets without the need for calibration and performs exceptionally well on overlapping pairs, achieving the lowest median error and highest success rates for Large and Small overlap categories for sELP test set. However, this method is designed for overlapping pairs as it initializes the model using pretrained CroCo [47], which is trained to address cross-view completion problem from two overlapping views. Furthermore, its output consists of a unified dense point cloud for each pair of images. Due to its design for overlapping pairs, its performance on non-overlapping views is extremely low, also on the wELP test set, which contains images from scenes included in its training set. Moreover, in terms of size, Dust3R (with a DPT head) contains 577 million parameters. In contrast, our model is significantly more compact, with only 80 million parameters. As shown in the table, despite its smaller size, our model yields significantly improved performance over non-overlapping pairs on both test sets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.666, + 0.848 + ], + "angle": 0, + "content": "5.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We conduct various ablations analyzing the effect of our progressive training scheme and other design choices,also reporting performance over the Top 5 predictions (consider" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1067" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.267 + ], + "angle": 0, + "content": "
OverlapTrain dataTop 1Top 5
MGE ↓RRA15 ↑RRA30 ↑MGE ↓RRA15 ↑RRA30 ↑
Large[8]13.6535.473.512.2261.484.7
+ΔFoV4.6179.781.14.4190.398.9
+ΔIm4.4690.492.44.4394.399.1
+ELP2.4197.597.92.4198.499.4
Small[8]55.283.729.129.8315.050.3
+ΔFoV12.9156.268.210.9766.085.4
+ΔIm11.4662.580.610.7368.091.0
+ELP4.4787.291.64.2491.197.2
None[8]74.9412.825.325.1126.158.8
+ΔFoV61.6225.038.416.8244.275.0
+ΔIm68.3125.036.116.2145.778.2
+ELP26.9736.150.712.8557.185.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.271, + 0.482, + 0.311 + ], + "angle": 0, + "content": "Table 3. Ablation study, evaluating the effect of our progressive training scheme over the wELP test set. All experiments start with the cropped panoramas used in Cai et al. [8]." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.34, + 0.481, + 0.398 + ], + "angle": 0, + "content": "ering the top-5 peaks in the relative yaw prediction, as detailed in Section 5.2). Additional ablations, including an analysis of architectural components, are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.403, + 0.482, + 0.508 + ], + "angle": 0, + "content": "The effect of our progressive training scheme. We conduct multiple experiments ablating the effect of our progressive training scheme. Our training process consists of four stages: initialization (following prior work [8]), incorporating multiple FoVs \\((+\\Delta \\mathrm{FoV})\\), training with image-level appearance augmentations \\((+\\Delta \\mathrm{Im})\\), and training with ExtremeLandmarkPairs pairs \\((+ELP)\\)." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.509, + 0.482, + 0.794 + ], + "angle": 0, + "content": "Table 3 illustrates the impact of each training stage on the results. As can be observed from the table, each training stage further refines the model's performance, often in a significant manner. For instance, the median error in the wELP small overlap test set decreased from \\(55.3^{\\circ}\\) to \\(12.9^{\\circ}\\) when FoV augmentations were added. Additionally, the ExtremeLandmarkPairs training set plays a crucial role in finalizing our training process, yielding a significant reduction in the median error (more than half) for small and non-overlapping scenarios in the wELP test set. While this median error remains relatively high for non-overlapping pairs, we observe that the top-5 scores show significant improvements, e.g., reducing the median error from \\(27.0^{\\circ}\\) to \\(12.8^{\\circ}\\). This demonstrates that the model has learned this knowledge, although it cannot necessarily be recovered by the largest peak alone. In the supplementary material, we demonstrate that the intermediate training stages are indeed important and that the improved performance cannot be obtained with the ExtremeLandmarkPairs training set alone." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.796, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Moreover, evidenced in Table 3, our model demonstrates strong generalization capabilities even when trained exclusively on panorama-cropped images. The model generalizes to the wELP test set significantly better (e.g., for Large overlap cases, our model achieves an MGE of \\(13.65^{\\circ}\\) compared to values higher than \\(120^{\\circ}\\) achieved by baselines), demonstrating that the improvement is not just due to our" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.274 + ], + "angle": 0, + "content": "
OverlapKPSMTop 1Top 5
MGE ↓RRA15↑RRA30↑MGE ↓RRA15↑RRA30↑
Large××2.1897.498.12.1897.498.1
×✓2.3097.097.42.3098.599.4
✓×2.4497.698.32.3198.499.7
✓✓2.4197.597.92.4198.499.4
Small××4.5087.991.64.5087.991.7
×✓4.4988.192.04.4691.296.7
✓×4.4187.592.24.3291.997.6
✓✓4.4787.291.64.2491.197.2
None××48.8134.044.112.5657.584.6
×✓43.0731.244.213.9953.583.2
✓×41.3935.346.813.0456.986.2
✓✓26.9736.150.712.8557.185.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "Table 4. Ablation study, evaluating the effect of the auxiliary channels added as input to our network. We train models without the keypoints and matches (KP) and without the segmentation maps (SM), comparing to our model over the wELP test set , after using a validation split that is balanced over the overall angle." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.478 + ], + "angle": 0, + "content": "progressive training scheme. We further examined architectural differences by conducting an additional ablation study, applying our progressive training scheme to the baseline models (see Table 9 in the supplementary materials). This experiment revealed that prior models are not directly suitable for real-world applications, as these baseline models showed significantly poorer performance across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.482, + 0.905, + 0.602 + ], + "angle": 0, + "content": "The effect of adding auxiliary channels. We ablate the effect of adding auxiliary in Table 4, training models without keypoints and matches (KP) or the segmentation map (SM) provided as additional inputs. As illustrated in the table, for non-overlapping cases, these auxiliary channels boost performance almost across all metrics. In particular, both channels play a role in reducing the errors (reducing the median error from over \\(40^{\\circ}\\) to \\(27.0^{\\circ}\\))." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.616, + 0.631, + 0.631 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.642, + 0.905, + 0.852 + ], + "angle": 0, + "content": "We present a method and benchmark dataset for estimating relative 3D rotations between pairs of (possibly) non-overlapping RGB images. Our approach extends prior work addressing extreme rotations to real-world data that exhibit variation in both appearance in intrinsic camera parameters. While our model shows promising results on real-world Internet image pairs, it also highlights the inherent difficulty of the underlying task, suggesting that considerable progress can be achieved with future techniques that leverage our dataset. Our paired data could also potentially serve for exploring the challenging task of estimating extreme translations in real-world settings. Future work could also consider incorporating more views for enhancing performance in such extreme non-overlapping scenarios." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.871, + 0.904, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments This work was partially supported by ISF (grant number 2510/23)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.947, + 0.515, + 0.957 + ], + "angle": 0, + "content": "1068" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In European Conference on Computer Vision, pages 192-209. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Roberto Cipolla Alex Kendall, Matthew Grimes. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV -International Conference on Computer Vision, pages 2938-2946, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[3] Lionel Baboud, Martin Čadík, Elmar Eisemann, and Hans-Peter Seidel. Automatic photo-to-terrain alignment for the annotation of mountain pictures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41–48. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In European Conference on Computer Vision, pages 751-767. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417. Springer, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR IEEE / CVF Computer Vision and Pattern, pages 18392-18402, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[8] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[9] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[10] Yaron Caspi and Michal Irani. Aligning non-overlapping sequences. IJCV, 48(1):39-51, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[11] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[12] Shay Dekel, Yosi Keller, and Martin Cadik. Estimating extreme 3d image rotations using cascaded attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2588-2598, 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[15] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. Camnet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2871-2880, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[16] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17775, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[18] Johan Edstedt, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. arXiv preprint arXiv:2308.08479, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[19] Xie Enze, Wang Wenhai, Yu Zhiding, Anandkumar Anima, M. Alvarez Jose, and Luo Ping. Segformer: Simple and efficient design for semantic segmentation with transformers. In Conference on Neural Information Processing Systems (NEURIPS), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[20] Zhiwen Fan, Panwang Pan, Peihao Wang, Yifan Jiang, Dejia Xu, Hanwen Jiang, and Zhangyang Wang. Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM (CACM), 24(6):381-395, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[23] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[25] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching atokens are frozen. arXiv preprint arXiv:2306.13643, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 60(2):91-110, 2004. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1069" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[29] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[30] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[33] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[34] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[35] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8601-8610, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[37] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[38] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[39] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[40] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[42] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5722-5731, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[43] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.906, + 0.359 + ], + "angle": 0, + "content": "[44] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.906, + 0.43 + ], + "angle": 0, + "content": "[45] Qianqian Wang, Xiaowei Zhou, Bharath Hariharan, and Noah Snavely. Learning feature descriptors using camera pose supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 757-774. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.906, + 0.486 + ], + "angle": 0, + "content": "[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. arXiv preprint arXiv:2312.14132, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.906, + 0.571 + ], + "angle": 0, + "content": "[47] Philippe Weinzaepfel, Vincent Leroy, Thomas Lucas, Romain Brégier, Yohann Cabon, Vaibhav Arora, Leonid Antsfeld, Boris Chidlovskii, Gabriela Csurka, and Jérôme Revaud. Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. Advances in Neural Information Processing Systems, 35:3502-3516, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.906, + 0.627 + ], + "angle": 0, + "content": "[48] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.906, + 0.669 + ], + "angle": 0, + "content": "[49] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.906, + 0.753 + ], + "angle": 0, + "content": "[50] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.756, + 0.906, + 0.811 + ], + "angle": 0, + "content": "[51] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Reply: Predicting probabilistic relative rotation for single objects in the wild. In European Conference on Computer Vision, pages 592-611. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.813, + 0.906, + 0.868 + ], + "angle": 0, + "content": "[52] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1070" + } + ] +] \ No newline at end of file diff --git a/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_origin.pdf b/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..69cac9e933ac717b86ba97e8d4b1141184f3897e --- /dev/null +++ b/2025/Extreme Rotation Estimation in the Wild/1bd2da5f-45ff-4055-bc43-55a1a6080681_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb27db18f055495e54711bc2b8b98b8cd6890ed8179d82f152b11f5d41182c47 +size 2928363 diff --git a/2025/Extreme Rotation Estimation in the Wild/full.md b/2025/Extreme Rotation Estimation in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..40dd04581ca68d598376ed5363163d96f7dab9fb --- /dev/null +++ b/2025/Extreme Rotation Estimation in the Wild/full.md @@ -0,0 +1,297 @@ +# Extreme Rotation Estimation in the Wild + +Hana Bezalel $^{1}$ Dotan Ankri $^{1}$ Ruojin Cai $^{2}$ Hadar Averbach-Elor $^{1,2}$ $^{1}$ Tel Aviv University $^{2}$ Cornell University + +https://tau-vailab.github.io/ExtremeRotationsInTheWild/ + +# Abstract + +We present a technique and benchmark dataset for estimating the relative 3D orientation between a pair of Internet images captured in an extreme setting, where the images have limited or non-overlapping field of views. Prior work targeting extreme rotation estimation assume constrained 3D environments and emulate perspective images by cropping regions from panoramic views. However, real images captured in the wild are highly diverse, exhibiting variation in both appearance and camera intrinsics. In this work, we propose a Transformer-based method for estimating relative rotations in extreme real-world settings, and contribute the ExtremeLandmarkPairs dataset, assembled from scene-level Internet photo collections. Our evaluation demonstrates that our approach succeeds in estimating the relative rotations in a wide variety of extreme-view Internet image pairs, outperforming various baselines, including dedicated rotation estimation techniques and contemporary 3D reconstruction methods. + +# 1. Introduction + +The problem of estimating the relative 3D orientation between a pair of images is embodied in fundamental computer vision tasks, such as camera localization [6, 34, 35] and 3D reconstruction [29, 36, 42]. Establishing pixel correspondences (either explicitly or implicitly) is typically a prerequisite for computing the relative rotation between the images. Correspondences, however, cannot be extracted in extreme settings where the images have little to no overlap. As dense imagery may not necessarily be available for many practical applications, a natural question arises: How can we estimate the relative rotation between non-overlapping RGB images, without the use of additional data (such as depth or temporal information)? + +We have recently seen pioneering efforts addressing the task of relative rotation estimation in such extreme non-overlapping settings [8, 12]. Prior work has proposed end-to-end neural architectures, demonstrating that hidden cues, + +![](images/6f2c5c753c36243e5c09bc76366efbfdb87a6d1c8637dd6c805dbb6f6d469c73.jpg) +Figure 1. Given a pair of (possibly) non-overlapping images captured in the wild—e.g., under arbitrary illumination and intrinsic camera parameters—such as the images of the Dam Square in Amsterdam depicted in red and blue boxes above*, our technique estimates the relative 3D rotation between the images. *The background panorama is illustrated for visualization purposes only. + +such as vanishing points or the directions of cast shadows, can implicitly guide the model for inferring the relative orientation between the images. To facilitate learning and evaluation, datasets constructed from panoramic views were adopted. These datasets emulate perspective views by cropping sub-areas from these panoramas, enabling generation of image pairs with various degrees of overlap. However, while such emulated views perhaps capture some of the challenges associated with extreme-view imagery, are they sufficient for representing real images—particularly, images captured in the wild? + +In this paper, we present a new approach that tackles the problem of extreme rotation estimation in the wild. Consider the boxed images in Figure 1. Internet (i.e., in the wild) images may vary due to a wide range of factors, including transient objects, weather conditions, time of day, and the cameras' intrinsic parameters. To explore this problem, we introduce a new dataset, ExtremeLandmarkPairs (ELP), assembled from publicly-available scene-level Internet image collections. We observe that the set of real extreme-view image pairs is limited, as Internet datasets are typically scene-centric, with nearby cameras commonly + +capturing overlapping views. Therefore, to facilitate training, we propose a progressive learning scheme that leverages and augments images cropped from panoramic views, allowing for gradually generalizing the model onto real Internet data. In particular, we construct datasets with varying field of views, that better resemble the distribution of real data, and perform image-level appearance augmentations by leveraging recent advancements in text-to-image Diffusion models [7, 22, 31]. + +To estimate extreme rotations in the wild, we propose a Transformer-based model that is provided with auxiliary channels, including the spatial distribution of local keypoints and matches and semantic segmentation maps, allowing for better reasoning over real image pairs with little or no overlap. Our results demonstrate that our model can accurately predict the relative rotations for a wide variety of extreme-view image pairs that vary in illumination, dynamic regions, and intrinsic parameters. We conduct extensive experiments, quantifying performance both over real Internet data and also over emulated perspective images cropped from panoramic views. Our evaluation shows that our model significantly improves over strong baselines when considering real images, while achieving comparable performance over emulated perspective image pairs. + +# 2. Related Work + +Relative pose estimation is a fundamental task in computer vision, typically studied for overlapping camera views. Traditionally, this task has been divided into two stages: correspondence estimation from local feature matching, followed by geometry-based pose estimation. In recent years, feature matching methods have advanced from using heuristic feature descriptors [5, 27, 32] and RANSAC-based matches [21] to learning-based feature extraction [4, 14-16, 18, 40, 45] and matching techniques [3, 26, 33], with several methods performing both feature extraction and matching using unified learning-based frameworks [4, 17, 38]. + +These methods are generally invariant to changes in illumination and appearance, demonstrating robust performance across various scene scales and also over in-the-wild datasets. However, their reliability diminishes in extreme view scenarios due to their dependence on visual overlap. + +Extreme-view scenarios lacking local pixelwise correspondences necessitate the use of end-to-end learning-based pose estimation methods which directly infer the 3D relationships and geometry from sparse and extreme-view images. Indeed, large-scale 3D object datasets have paved the way for learning-based methods which estimate camera pose directly from sparse views [20, 25, 37, 43, 44, 50-52]. However, these methods primarily concentrate on object-centric scenes, under experimental settings that typically assume similar lighting and camera intrinsics for the input views. Furthermore, these methods often utilize bounding + +box inputs defining the object of interest, which is less suitable in the case of images depicting large-scale scenes. + +Several prior works address a sparse view setting at scene-scale. In particular, Chen et al. [11] propose to learn a discrete distributions of pose space, Agarwala et al. [1] simplify scene reconstruction using a plane representation, and Rockwell et al. [30] introduce an inductive bias of the 8-point algorithm into a vision transformer architecture. Recently, models that directly predict pixel-aligned point maps from input pairs (or sparse image collections), such as DUSt3R [46] and Mast3R [23], have demonstrated promising results on pose estimation and scene-scale 3D reconstruction of Internet data with a wide baseline. + +Camera pose estimation for non-overlapping views presents a greater challenge. Earlier efforts [10] explored searching for consistent temporal behavior. Several work utilize pairwise RGB and depth scan data to estimate relative pose among such extreme pairs [48, 49]. Cai et al. [8] tackle pose estimation for non-overlapping views without the use of additional data, by introducing a learning-based network leveraging cross-correlation volume to exploit implicit cues. This correlation volume is later enhanced through the integration of transformer attention modules [12]. Nonetheless, these approaches assume constrained 3D environments, including the assumption of consistent lighting and camera intrinsics, and are designed for camera distribution of emulated perspective views cropped from panoramas. In this work, we aim to address pose estimation for realistic in-the-wild non-overlapping image pairs, enhancing the applicability of extreme pose estimation to Internet photos and real-world data. + +# 3. The ExtremeLandmarkPairs Dataset + +Prior works on extreme pose estimation use panoramic views, cropping from it sub-areas to emulate perspective views [8, 12]. To evaluate and train models on real perspective image pairs, we propose a new benchmark and dataset, ExtremeLandmarkPairs (ELP), constructed from Internet image pairs from the MegaDepth [24], Cambridge Landmarks [2], and MegaScenes [39] datasets. In this section, we first describe the dataset construction procedure (Section 3.1), and then present details regarding dataset size and train and test splits (Table 1). + +# 3.1. Dataset Construction + +To construct a dataset of real perspective image pairs with varying degrees of overlap, we leverage available scene-level training data. Existing Internet image collections typically contain camera poses (predicted up to scale), which are determined using Structure-from-Motion (SfM) algorithms, such as COLMAP [36]. In what follows, we describe how we extract real image pairs from this data, which can then be used for training and evaluating models. + +![](images/ab05a16a75e49499ee349e8ad6e82bcbb1e38fab076b2a7571a92230872b43b6.jpg) +(a) Reconstruction and camera distribution of all images + +![](images/eb534a776b1ea06d3928ea43616cafac13240a4ffebb243d11b2ce45c0e8840b.jpg) +(b) Camera distribution of large-overlapping pairs + +![](images/03ee3c05ba1b6fd796e0cf91f897272bc55e5adab7324d19d6661b4f09c50f17.jpg) +(c) Camera distribution of non-overlapping pairs +Figure 2. Camera distribution of the Vatican, Rome scene from the ExtremeLandmarkPairs Dataset. We construct a dataset of real perspective image pairs with predominant rotational motion shown in (b) and (c) from the dense imagery reconstruction in (a). + +# Identifying Pairs with Predominant Rotational Motion. + +Prior works targeting relative rotation estimation, in particular for non-overlapping views, mostly utilize panoramic views, focusing on image pairs with purely rotational motion. Pairs belonging to real image collections almost always contain a non-negligible translation component. Furthermore, unlike in the StreetLearn [28] dataset used by prior work [8, 12] that provides exact translation values between consecutive panoramas which allows for filtering pairs with predominant rotational motions, reconstructed relative poses are only provided up to scale. The scale varies among different reconstructed scenes, and therefore there's no global threshold on the relative translation values which can be used for identifying pairs with predominant rotational motion. + +To automatically identify such pairs, we observe that available Internet collections require the existence of dense imagery, to compensate for the vast number of unknowns + +in the SfM optimization. We therefore construct mutual nearest neighbors edge-weighted graphs, with one graph per landmark. In each graph $G$ , nodes $v \in V$ correspond to images, and two images are connected by an edge $e \in E$ if they are both among each other's $K$ nearest neighbors, considering L2 distances between their translations ( $K$ is empirically set to 5). Note that images captured from sparser (outlier) regions in space are unlikely to be within the mutual $K$ nearest neighbors of images captured within denser regions, and hence won't be included in $G$ . Finally, we select a subset of image pairs containing relatively small distances from each scene graph $G$ , yielding a set of image pairs with predominant rotational motion; see the supplementary material for additional details. + +Extracting Level of Overlap. Following prior work [8, 12], we are interested in training and evaluating models according to three different categories: Large, Small and None, indicating image pairs with a varying amount of overlap. However, unlike prior work that use cropped images with a fixed $90^{\circ}$ FoV, Internet images contain varying FoV values. Thus, the relative rotation angle is not sufficient for extracting the pair's overlap level. + +Denote the FoV values of image $i\in [1,2]$ as $[\mathrm{fov}_x^i,\mathrm{fov}_y^i ]$ We can parameterize a 3D rotation matrix $\mathbf{R}$ using three Euler angles $[\alpha ,\beta ,\gamma ]$ denoting the relative roll, pitch and yaw angles, respectively: + +$$ +\mathbf {R} (\alpha , \beta , \gamma) = \mathbf {R} _ {z} (\alpha) \mathbf {R} _ {y} (\beta) \mathbf {R} _ {x} (\gamma), \tag {1} +$$ + +following Dense Correlation Volumes' coordinate system. We use the following conditions to determine the overlap level $o$ : + +$$ +o = \left\{ \begin{array}{l l} L a r g e & | \gamma | < \frac {\operatorname {f o v} _ {x} ^ {1} + \operatorname {f o v} _ {x} ^ {2}}{4} \wedge | \beta | < \frac {\operatorname {f o v} _ {y} ^ {1} + \operatorname {f o v} _ {y} ^ {2}}{4} \\ N o n e & | \gamma | > \frac {\operatorname {f o v} _ {x} ^ {1} + \operatorname {f o v} _ {x} ^ {2}}{2} \wedge | \beta | > \frac {\operatorname {f o v} _ {y} ^ {1} + \operatorname {f o v} _ {y} ^ {2}}{2} \\ S m a l l & e l s e \end{array} \right. \tag {2} +$$ + +In other words, pairs with relative yaw and pitch angles that are smaller than a quarter of the average corresponding FoV values are considered as pairs with a large overlap ratio. Likewise, pairs with relative yaw and pitch angles that are larger than half the average corresponding FoV values are considered non-overlapping pairs. Pairs in-between these conditions are considered pairs with a small overlap ratio. + +Additional Filtering. In the scene-scale datasets we explore, large FoV disparities could result in one image focusing on specific architectural details like a statue, while the other captures a much broader scene perspective, further complicating the problem of estimating the relative rotations. To extract pairs consistent in scale, we limit the difference between the FoV values to be at most $5^{\circ}$ . Furthermore, as most images contain small roll values, we rotate the scenes to match the gravity axis and horizontal axis + +
SubsetSource#Pairs
#ScenesLargeSmallNoneTotal
Train[39]588333430136842948176595
Validation[39]515339871041308238
Validation Balanced[39]1779255707854
sELP[2]6251282719615300
wELP[24]1727008296434172
+ +Table 1. ExtremeLandmarkPairs Dataset Statistics. Above, we report the number of image pairs extracted for each overlap level, split into train and test (with sELP denoting a single camera setting and wELP denoting the "in the wild" setting). + +and filtered images with rolls exceeding $10^{\circ}$ . Finally, we excluded images captured from an aerial perspective, by filtering images with translation along y-axis exceeding a global threshold, empirically set to 1. + +# 3.2. Dataset Size and Splits + +We apply the procedure described above over landmarks from the MegaScenes [39] dataset to create a training set and a validation set. As we focus on outdoor scenes in our work, we filter reconstructions that capture indoor scenes. We obtain a total of nearly 34K non-overlapping pairs originating from over 2K unique landmarks. The validation set was also balanced the set over the overall angle. + +For evaluation, we create two test sets, to separately examine image pairs captured in a single camera setting with constant illumination (sELP) and image pairs captured in the wild (wELP). Image pairs in the sELP test set contain images from the Cambridge Landmarks [2] dataset, which contains videos capturing six different landmarks. Image pairs in the wELP test set contain images from the MegaDepth dataset [24], which contains Internet photos from Flickr for a set of large-scale landmarks. + +As there is some overlap between the landmarks in MegaScenes [39] and MegaDepth [24], we performed additional filtering, ensuring no overlap exists between the train and test sets. Furthermore, we filter test pairs if one of the images is highly transient—i.e., if transient objects dominate the image. We quantify this using a pretrained segmentation model [19], filtering images containing transient objects in over $40\%$ of the pixels. We also manually validate non-overlapping image pairs, allowing to further filter unlabeled objects, such as a stage or a market stand. Finally, for non-overlapping image pairs, we observed that the overall relative rotations are highly imbalanced, and therefore we balance this set by the overall relative angle. Table 1 summarizes the number of image pairs and landmarks used for both training and test. + +# 4. Method + +Given a pair of Internet images with (possibly) extreme relative motion, we estimate the relative rotation $\mathbf{R}$ between + +the images. Following prior works on extreme rotation estimation [8, 12], we assume a camera-centric setting, where the two cameras have limited translation. However, our approach departs from prior works by operating on outdoor images captured by a crowd of photographers, with varying intrinsic parameters as well as appearances— $e.g.$ , due to illumination changes and dynamic objects. + +Our model (detailed in Section 4.1; see Figure 3 for an overview) outputs three discrete Euler angles, denoting the relative roll, pitch and yaw angles. As illustrated in prior work [8], this parameterization enables using a simple cross-entropy loss for training. In Section 4.2, we describe our progressive learning scheme, allowing for gradually adapting the model to extreme Internet imagery. + +# 4.1. Model + +We extract image features using a pretrained LoFTR model [38]. In contrast to the features extracted using common convolutional neural networks pretrained on ImageNet [13], LoFTR is a Transformer-based model trained on Internet pairs, with the goal of extracting local feature matches – a setting and task which is highly related to the one we address in our work, thus enabling extraction of better (i.e., more relevant) features. + +As we are interested in designing a network that can also reason over image pairs with little or no overlap, we combine the extracted features with additional auxiliary channels; see Figure 3 (bottom left). These include keypoint and pairwise matches masks, utilized previously for disambiguating images for similar structures [9]. Intuitively, knowledge over pairwise matches can assist the model in cases of small overlap and for generalizing across different camera intrinsic properties. We also incorporate a segmentation map as an additional auxiliary channel, which segments images into several categories (such as sky, building, road and sidewalk). This channel allows for identifying additional cues, such as the skyline or transient objects, which can aid in determining the rotation for nonoverlapping pairs. In Section 5, we demonstrate the benefit of incorporating these auxiliary channels in our model. + +We then reshape extracted features (and auxiliary channels) to tokens, concatenating these image tokens with learnable Euler angle position embeddings. These tokens are processed by our Rotation Estimation Transformer module, which uses a transformer decoder architecture [41]. The output Euler angle tokens obtain information from image features and auxiliary channels within transformer attention modules. These tokens are then processed by three different prediction heads (one per angle, denoted as MLP in Figure 3). Each output prediction head uses as input averaged image tokens and one of the output Euler angle tokens, which provides the model with additional angle-specific information, allowing for achieving improved performance, + +![](images/e05a606a2a754adc8b223383658e7090df6e090689bc37d52b33935ae342b1ef.jpg) +Figure 3. Method architecture. Given a pair of input Internet images, we extract image features using pretrained LoFTR. These features are combined with auxiliary channels, including keypoint and pairwise matches masks, and segmentation maps (visualized on the bottom left). These image features are reshaped into tokens and concatenated with Euler angle position embeddings, which are then processed by our Rotation Estimation Transformer module. The output Euler angle tokens and averaged image tokens are concatenated and processed by MLPs to predict the probability distribution of Euler angles, representing the relative 3D rotation between the input images. + +as we show in the supplementary material. The prediction heads output a probability distribution over $N = 360$ bins, capturing an angle in the range $[-180^{\circ}, 180^{\circ}]$ . + +# 4.2. Learning + +As detailed in Section 3, we assemble real image pairs from Internet imagery, which can be used for training and evaluating models. However, even with our proposed ExtremeLandmarkPairs dataset, the set of real image pairs is limited—e.g., only $\sim 36\mathrm{K}$ non-overlapping image pairs are extracted, as available Internet datasets are typically scene-centric, with nearby cameras usually capturing overlapping views. Therefore, in what follows, we propose a progressive learning scheme, which leverages panoramic images, and allows for gradually generalizing the model to images captured in-the-wild. All learning stages are optimized using a cross-entropy loss for each Euler angle prediction. Additional details are provided in the supplementary material. + +Initialization. We begin by training our model on the perspective views cropped from panoramic images, using the data created by Cai et al [8]. Specifically, we use the image pairs cropped from panoramic images included in the StreetLearn [28] dataset, depicting various streets in Manhattan. As further detailed in Cai et al [8], this training set includes roughly 1M image pairs sampled from the same panorama, split according to the overlap level. + +Training with Data Augmentations. We observe that the dataset constructed by prior work could be augmented to better capture the distribution of image pairs captured inthe-wild. In particular, we focus on two types of data augments, which we elaborate on next: (i) field of view (FoV) augmentations (denoted henceforth as $\Delta \mathrm{FoV}$ ) and (ii) image-level appearance augmentations (denoted henceforth as $\Delta \mathrm{Im}$ ). In Section 5, we demonstrate how both types + +![](images/f41d914e36d412fff29247b302c386059ab2a8610ed6d9bee580eec36fb3de22.jpg) + +![](images/8c65f2ab42e9dd59fba11dcdd8ead8234b7fcb4a21ccd1c3e872215dbd203f15.jpg) + +![](images/179fa16bb1559513abd0d439d256fc57c4c1938319cfa3ed368a5802381e7de8.jpg) + +![](images/34f1f671f5b12ba90fecc9decf60c69be8821fa88b6785834ff8cc8f11c91da0.jpg) + +![](images/17acefd6b1feb0660a88f8c83f58cff293545ec1c6968b0d355f850e1af137d2.jpg) +busy street + +![](images/26e848fbc28cb8551610bc23778892c0d8f002adf89b77ddaeb44add11918361.jpg) +Figure 4. Augmenting perspective images cropped from panoramic views with image-level appearance modifications. Given an input image (left) and a target text prompt "Make it $\langle w\rangle$ (" $\langle w\rangle$ is specified above), we use a conditional Diffusion model [7] to create semantic appearance augmentations which modify both the global image characteristics as well as local image regions. + +![](images/b992877c1d45e880bb2fccb8cbf65d94ee8db9d7f75927dcb981a4b9e4c64921.jpg) +night + +![](images/1e07e2abc8014def5d0a62032f02037bbd9b705d41c1a6536c7bd78cd209c4f2.jpg) +Christmas + +of augmentations improve the model's ability in generalizing to real Internet scenes. + +Rather than cropping perspective images with a fixed $90^{\circ}$ FoV as was done in prior work, we analyze the FoV values of the images belonging to the ELP training set. We compute the mean and standard-deviation values, denoted as $\mu$ and $\sigma$ , respectively. We then construct new perspective images by sampling from a Gaussian distribution $G(\mu, a \cdot \sigma)$ that resembles the distribution of real data, setting $a = 1.5$ for obtaining a more diverse set, which also bears higher similarity to the perspective images used during initialization. We also allow for FoV differences (of up to $5^{\circ}$ ) between the two images paired together. Additionally, rather than providing the model with the full content within these regions, we construct crops with various aspect ratios to fur + +ther emulate real image pairs. + +In-the-wild images differ not only in their intrinsic parameters, but also in their appearance - due to presence of varying illumination and dynamic objects. Therefore, in addition to augmenting the dataset with images of various field of views, we also perform image-level appearance augmentations, leveraging recent advancements in text-to-image Diffusion models [7, 22, 31]. In particular, we apply the conditional InstructPix2Pix [7] model on a subset of our data, using a set of text prompts of the format "Make it $\langle w\rangle$ ", where $\langle w\rangle$ captures diverse appearance modifications. See Figure 4 for a few illustrative examples; note how some of these modify the global illumination (e.g., altering the time of day or the season) while others yield local edits (such as adding more people to make it "a busy street"). We performed additional filtering on the augmented set, to avoid significant edits that also altered the structure of the scene; see the supplementary material for additional details. + +Training on Real Data. Finally, we finetune the model over real image pairs from our proposed ELP dataset. To prioritize learning over extreme image pairs, we first pass a batch of non-overlapping image pairs, followed by a batch of overlapping image pairs. + +# 5. Results and Evaluation + +To validate our method, we compare our model with prior relative rotation estimation methods on the sELP and wELP test sets. In particular, our experiments seek to answer the following questions: + +- How well do previous methods perform on our proposed task of extreme rotation estimation in the wild? +- How important is our progressive training scheme, and to what extent can the improved performance be attributed to training on the ExtremeLandmarkPairs dataset? +- What is the impact of our design choices? + +We also present qualitative results over different overlap levels in Figure 5. These results illustrate our model's performance, and its robustness to varying illumination conditions and transient objects (such as the screen and the beer cup in the rightmost examples). Please refer to the supplementary material for many more qualitative results, including visualizations for baseline methods. + +# 5.1. Baselines + +We compare our method to correspondence-based techniques, dedicated (end-to-end) relative rotation estimation techniques as well as recent relative pose estimation techniques. Specifically, we consider two correspondence-based methods: SIFT [27] and LoFTR [38]. These methods extract keypoint matches to compute an essential matrix using the RANSAC algorithm and calculate the rotation that stems from the essential matrix. We also compare against two prior works specifically targeting extreme + +rotation estimation using end-to-end deep learning frameworks: DenseCorrVol [8] and CascadedAtt [12]. These methods are trained and evaluated on images sampled from panoramic views from StreetLearn [28]. Results for CascadedAtt are reported over a reproduced model, as a pretrained model is not provided; additional details are provided in the supplementary material. + +We also consider 8PointViT [30], which estimates the relative pose between two images with a Vision Transformer that is modified to be close to the eight point algorithm. Following prior work [12], we only report results over overlapping settings for this baseline. Furthermore, as it assumes a single camera setting, we evaluate it only on the sELP test set. Finally, we consider Dust3R [46], a recent technique for dense and unconstrained stereo 3D reconstruction. While this work does not specifically target the setting of extreme rotation estimation, we examine to what extent they can be adapted for this task in the wild. Note that Dust3R was trained on MegaDepth. Since our wELP test set is constructed from image pairs from this dataset, it is not entirely fair comparison; therefore, Dust3R's results on the wELP test set are highlighted in gray. + +# 5.2. Evaluation Metrics + +For each image pair, we compute the geodesic error, defined as follows: + +$$ +\operatorname {e r r} = \arccos \left(\frac {t r \left(\mathbf {R} ^ {T} \mathbf {R} ^ {*}\right) - 1}{2}\right) \tag {3} +$$ + +where $\mathbf{R}$ is the predicted rotation matrix, and $\mathbf{R}^*$ is the ground truth relative rotation matrix. We report the median geodesic error (MGE) and relative rotation accuracy (RRA) for each test set and overlap ratio. RRA is reported over a predefined threshold $\tau$ , indicating the percentage of image pairs with relative rotation error below $\tau$ . We report this metric for $\tau = 15^\circ$ ( $\mathrm{RRA}_{15}$ ) and $\tau = 30^\circ$ ( $\mathrm{RRA}_{30}$ ). + +Additionally, as we discretize the space of rotations and estimate three rotation angles, each predicted as a probability distribution over $N$ bins, we also report the performance of the Top 5 predictions in our ablation study. The Top 5 prediction reports the lowest error, considering the top-5 peaks in the relative yaw prediction only, as we observe that errors are mostly a function of the relative yaw angles. + +# 5.3.Quantitative Comparison + +The main quantitative results comparing our method to other methods on rotation estimation in the wild are reported in Table 2. In the supplementary material, we conduct an additional evaluation in the more constrained setting examined by prior work targeting extreme rotation estimation (i.e., images cropped from panoramas). + +As illustrated in Table 2, SIFT and LoFTR, which are correspondence-based methods, exhibit some robustness + +![](images/e81900b6e1f44b20a73b6e30ac58d8caf95356429ab9755d839210925da916ae.jpg) +Figure 5. Qualitative results on the wELP test set. We visualize the results of our model over different overlap levels, where the images on the left serve as the reference points, and their coordinate system determines the relative rotation, which defines the images on the right. The ellipsoids representing the ground truth are color-coded to match their respective images, with the estimated relative rotation illustrated by a cyan dashed line. As illustrated by the examples above, our method can accurately predict relative rotations for diverse image pairs containing varying appearances and intrinsic parameters. Please refer to the supplementary material for additional qualitative results. + +
MethodsELPwELP
MGE↓RRA15↑RRA30↑MGE↓RRA15↑RRA30↑
LargeSIFT* [27]1.9592.395.32.9474.680.8
LoFTR* [38]1.7697.299.12.1385.293.8
DenseCorrVol [8]98.5125.933.3120.537.013.0
CascadedAtt [12]29.7542.750.0170.627.39.2
8PointViT [30]22.3331.964.8---
Dust3R [46]0.7799.799.91.0198.499.2
Ours2.4596.796.82.4197.597.9
SmallSIFT* [27]5.0764.771.37.2761.468.4
LoFTR* [38]2.7081.593.66.8066.681.2
DenseCorrVol [8]143.472.89.4125.733.19.4
CascadedAtt [12]148.440.03.0139.142.74.4
8PointViT [30]51.301.712.8---
Dust3R [46]1.9695.994.62.8089.894.4
Ours4.3588.389.04.4787.291.6
NoneSIFT* [27]121.942.75.4122.840.02.0
LoFTR* [38]---56.540.033.0
DenseCorrVol [8]77.109.027.082.042.913.7
CascadedAtt [12]69.698.423.178.607.520.8
Dust3R [46]114.3319.823.981.2115.426.9
Ours13.6252.759.726.9736.150.7
+ +Table 2. Rotation Estimation in the Wild. We evaluate performance over the sELP and wELP test sets, separately considering Large (top), Small (middle) and Non-overlapping (bottom) pairs. * indicates median errors are computed only over successful image pairs, for which these algorithms output a pose estimate (failure over more than $50\%$ of the test pairs is shown in gray). Note that Dust3R was trained on images from wELP. + +when handling highly overlapping Internet images, compared to methods that trained on images cropped from panoramas, achieving a median error of less than $3^{\circ}$ for Large overlap pairs in both test sets. However, these methods rely on image overlap and may not always provide an output of estimated camera pose, as the geometric verification requires a sufficient number of detected inliers. Therefore, they struggle to produce reliable matches in cases of limited overlapping regions, as also observed in prior work + +focusing on extreme scenarios [8, 12]. + +DenseCorrVol, CascadedAtt and 8PointViT exhibit relatively poor performance on the ELP test sets, illustrating that models trained on images sampled from panoramicas cannot easily generalize to Internet photos. In the supplementary material, we show that our model significantly outperforms these prior work even when trained on the same dataset (i.e., images with a constant FoV and illumination cropped from panoramicas). As further discussed there, we believe this gap can be partially attributed to the usage of a pretrained LoFTR feature extractor, which is capable of encoding knowledge between Internet image pairs (which vary in their intrinsic and appearance). + +Dust3R operates on Internet datasets without the need for calibration and performs exceptionally well on overlapping pairs, achieving the lowest median error and highest success rates for Large and Small overlap categories for sELP test set. However, this method is designed for overlapping pairs as it initializes the model using pretrained CroCo [47], which is trained to address cross-view completion problem from two overlapping views. Furthermore, its output consists of a unified dense point cloud for each pair of images. Due to its design for overlapping pairs, its performance on non-overlapping views is extremely low, also on the wELP test set, which contains images from scenes included in its training set. Moreover, in terms of size, Dust3R (with a DPT head) contains 577 million parameters. In contrast, our model is significantly more compact, with only 80 million parameters. As shown in the table, despite its smaller size, our model yields significantly improved performance over non-overlapping pairs on both test sets. + +# 5.4. Ablation study + +We conduct various ablations analyzing the effect of our progressive training scheme and other design choices,also reporting performance over the Top 5 predictions (consider + +
OverlapTrain dataTop 1Top 5
MGE ↓RRA15 ↑RRA30 ↑MGE ↓RRA15 ↑RRA30 ↑
Large[8]13.6535.473.512.2261.484.7
+ΔFoV4.6179.781.14.4190.398.9
+ΔIm4.4690.492.44.4394.399.1
+ELP2.4197.597.92.4198.499.4
Small[8]55.283.729.129.8315.050.3
+ΔFoV12.9156.268.210.9766.085.4
+ΔIm11.4662.580.610.7368.091.0
+ELP4.4787.291.64.2491.197.2
None[8]74.9412.825.325.1126.158.8
+ΔFoV61.6225.038.416.8244.275.0
+ΔIm68.3125.036.116.2145.778.2
+ELP26.9736.150.712.8557.185.8
+ +Table 3. Ablation study, evaluating the effect of our progressive training scheme over the wELP test set. All experiments start with the cropped panoramas used in Cai et al. [8]. + +ering the top-5 peaks in the relative yaw prediction, as detailed in Section 5.2). Additional ablations, including an analysis of architectural components, are provided in the supplementary material. + +The effect of our progressive training scheme. We conduct multiple experiments ablating the effect of our progressive training scheme. Our training process consists of four stages: initialization (following prior work [8]), incorporating multiple FoVs $(+\Delta \mathrm{FoV})$ , training with image-level appearance augmentations $(+\Delta \mathrm{Im})$ , and training with ExtremeLandmarkPairs pairs $(+ELP)$ . + +Table 3 illustrates the impact of each training stage on the results. As can be observed from the table, each training stage further refines the model's performance, often in a significant manner. For instance, the median error in the wELP small overlap test set decreased from $55.3^{\circ}$ to $12.9^{\circ}$ when FoV augmentations were added. Additionally, the ExtremeLandmarkPairs training set plays a crucial role in finalizing our training process, yielding a significant reduction in the median error (more than half) for small and non-overlapping scenarios in the wELP test set. While this median error remains relatively high for non-overlapping pairs, we observe that the top-5 scores show significant improvements, e.g., reducing the median error from $27.0^{\circ}$ to $12.8^{\circ}$ . This demonstrates that the model has learned this knowledge, although it cannot necessarily be recovered by the largest peak alone. In the supplementary material, we demonstrate that the intermediate training stages are indeed important and that the improved performance cannot be obtained with the ExtremeLandmarkPairs training set alone. + +Moreover, evidenced in Table 3, our model demonstrates strong generalization capabilities even when trained exclusively on panorama-cropped images. The model generalizes to the wELP test set significantly better (e.g., for Large overlap cases, our model achieves an MGE of $13.65^{\circ}$ compared to values higher than $120^{\circ}$ achieved by baselines), demonstrating that the improvement is not just due to our + +
OverlapKPSMTop 1Top 5
MGE ↓RRA15↑RRA30↑MGE ↓RRA15↑RRA30↑
Large××2.1897.498.12.1897.498.1
×✓2.3097.097.42.3098.599.4
✓×2.4497.698.32.3198.499.7
✓✓2.4197.597.92.4198.499.4
Small××4.5087.991.64.5087.991.7
×✓4.4988.192.04.4691.296.7
✓×4.4187.592.24.3291.997.6
✓✓4.4787.291.64.2491.197.2
None××48.8134.044.112.5657.584.6
×✓43.0731.244.213.9953.583.2
✓×41.3935.346.813.0456.986.2
✓✓26.9736.150.712.8557.185.8
+ +Table 4. Ablation study, evaluating the effect of the auxiliary channels added as input to our network. We train models without the keypoints and matches (KP) and without the segmentation maps (SM), comparing to our model over the wELP test set , after using a validation split that is balanced over the overall angle. + +progressive training scheme. We further examined architectural differences by conducting an additional ablation study, applying our progressive training scheme to the baseline models (see Table 9 in the supplementary materials). This experiment revealed that prior models are not directly suitable for real-world applications, as these baseline models showed significantly poorer performance across all metrics. + +The effect of adding auxiliary channels. We ablate the effect of adding auxiliary in Table 4, training models without keypoints and matches (KP) or the segmentation map (SM) provided as additional inputs. As illustrated in the table, for non-overlapping cases, these auxiliary channels boost performance almost across all metrics. In particular, both channels play a role in reducing the errors (reducing the median error from over $40^{\circ}$ to $27.0^{\circ}$ ). + +# 6. Conclusion + +We present a method and benchmark dataset for estimating relative 3D rotations between pairs of (possibly) non-overlapping RGB images. Our approach extends prior work addressing extreme rotations to real-world data that exhibit variation in both appearance in intrinsic camera parameters. While our model shows promising results on real-world Internet image pairs, it also highlights the inherent difficulty of the underlying task, suggesting that considerable progress can be achieved with future techniques that leverage our dataset. Our paired data could also potentially serve for exploring the challenging task of estimating extreme translations in real-world settings. Future work could also consider incorporating more views for enhancing performance in such extreme non-overlapping scenarios. + +Acknowledgments This work was partially supported by ISF (grant number 2510/23). + +# References + +[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In European Conference on Computer Vision, pages 192-209. Springer, 2022. 2 +[2] Roberto Cipolla Alex Kendall, Matthew Grimes. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV -International Conference on Computer Vision, pages 2938-2946, 2015. 2, 4 +[3] Lionel Baboud, Martin Čadík, Elmar Eisemann, and Hans-Peter Seidel. Automatic photo-to-terrain alignment for the annotation of mountain pictures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41–48. IEEE, 2011. 2 +[4] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In European Conference on Computer Vision, pages 751-767. Springer, 2018. 2 +[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417. Springer, 2006. 2 +[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 1 +[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR IEEE / CVF Computer Vision and Pattern, pages 18392-18402, 2023. 2, 5, 6 +[8] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[9] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 4 +[10] Yaron Caspi and Michal Irani. Aligning non-overlapping sequences. IJCV, 48(1):39-51, 2002. 2 +[11] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 2 +[12] Shay Dekel, Yosi Keller, and Martin Cadik. Estimating extreme 3d image rotations using cascaded attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2588-2598, 2024. 1, 2, 3, 4, 6, 7 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4 + +[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2 +[15] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. Camnet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2871-2880, 2019. +[16] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2 +[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17775, 2023. 2 +[18] Johan Edstedt, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. arXiv preprint arXiv:2308.08479, 2023. 2 +[19] Xie Enze, Wang Wenhai, Yu Zhiding, Anandkumar Anima, M. Alvarez Jose, and Luo Ping. Segformer: Simple and efficient design for semantic segmentation with transformers. In Conference on Neural Information Processing Systems (NEURIPS), 2021. 4 +[20] Zhiwen Fan, Panwang Pan, Peihao Wang, Yifan Jiang, Dejia Xu, Hanwen Jiang, and Zhangyang Wang. Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727, 2023. 2 +[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM (CACM), 24(6):381-395, 1981. 2 +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6 +[23] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. 2 +[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2, 4 +[25] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 2 +[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching atokens are frozen. arXiv preprint arXiv:2306.13643, 2023. 2 +[27] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 60(2):91-110, 2004. 2, 6, 7 +[28] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray + +Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 3, 5, 6 +[29] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1 +[30] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2, 6, 7 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6 +[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2 +[33] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 2 +[34] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1 +[35] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8601-8610, 2018. 1 +[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2 +[37] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 2 +[38] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 4, 6, 7 +[39] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 2, 4 +[40] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2 + +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[42] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5722-5731, 2021. 1 +[43] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 2 +[44] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 2 +[45] Qianqian Wang, Xiaowei Zhou, Bharath Hariharan, and Noah Snavely. Learning feature descriptors using camera pose supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 757-774. Springer, 2020. 2 +[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. arXiv preprint arXiv:2312.14132, 2023. 2, 6, 7 +[47] Philippe Weinzaepfel, Vincent Leroy, Thomas Lucas, Romain Brégier, Yohann Cabon, Vaibhav Arora, Leonid Antsfeld, Boris Chidlovskii, Gabriela Csurka, and Jérôme Revaud. Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. Advances in Neural Information Processing Systems, 35:3502-3516, 2022. 7 +[48] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2 +[49] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2 +[50] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 2 +[51] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Reply: Predicting probabilistic relative rotation for single objects in the wild. In European Conference on Computer Vision, pages 592-611. Springer, 2022. +[52] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 2 \ No newline at end of file diff --git a/2025/Extreme Rotation Estimation in the Wild/images.zip b/2025/Extreme Rotation Estimation in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..20599b419baa347e1a280529d4b19b304749ae65 --- /dev/null +++ b/2025/Extreme Rotation Estimation in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eede324e52126e4911ef4eeb20c87eb23f08ac9c5ad3b940fa165c5b58346e6a +size 561688 diff --git a/2025/Extreme Rotation Estimation in the Wild/layout.json b/2025/Extreme Rotation Estimation in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..72730ca223789594a2cca16bbb5d768276e6498f --- /dev/null +++ b/2025/Extreme Rotation Estimation in the Wild/layout.json @@ -0,0 +1,7876 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 178, + 103, + 433, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 103, + 433, + 119 + ], + "spans": [ + { + "bbox": [ + 178, + 103, + 433, + 119 + ], + "type": "text", + "content": "Extreme Rotation Estimation in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "spans": [ + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "text", + "content": "Hana Bezalel" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "text", + "content": " Dotan Ankri" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "text", + "content": " Ruojin Cai" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "text", + "content": " Hadar Averbach-Elor" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "text", + "content": "Tel Aviv University " + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 487, + 178 + ], + "type": "text", + "content": "Cornell University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 185, + 462, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 185, + 462, + 196 + ], + "spans": [ + { + "bbox": [ + 161, + 185, + 462, + 196 + ], + "type": "text", + "content": "https://tau-vailab.github.io/ExtremeRotationsInTheWild/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 224, + 200, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 224, + 200, + 237 + ], + "spans": [ + { + "bbox": [ + 151, + 224, + 200, + 237 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 250, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 465 + ], + "type": "text", + "content": "We present a technique and benchmark dataset for estimating the relative 3D orientation between a pair of Internet images captured in an extreme setting, where the images have limited or non-overlapping field of views. Prior work targeting extreme rotation estimation assume constrained 3D environments and emulate perspective images by cropping regions from panoramic views. However, real images captured in the wild are highly diverse, exhibiting variation in both appearance and camera intrinsics. In this work, we propose a Transformer-based method for estimating relative rotations in extreme real-world settings, and contribute the ExtremeLandmarkPairs dataset, assembled from scene-level Internet photo collections. Our evaluation demonstrates that our approach succeeds in estimating the relative rotations in a wide variety of extreme-view Internet image pairs, outperforming various baselines, including dedicated rotation estimation techniques and contemporary 3D reconstruction methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 489, + 135, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 135, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 135, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 509, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 665 + ], + "type": "text", + "content": "The problem of estimating the relative 3D orientation between a pair of images is embodied in fundamental computer vision tasks, such as camera localization [6, 34, 35] and 3D reconstruction [29, 36, 42]. Establishing pixel correspondences (either explicitly or implicitly) is typically a prerequisite for computing the relative rotation between the images. Correspondences, however, cannot be extracted in extreme settings where the images have little to no overlap. As dense imagery may not necessarily be available for many practical applications, a natural question arises: How can we estimate the relative rotation between non-overlapping RGB images, without the use of additional data (such as depth or temporal information)?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We have recently seen pioneering efforts addressing the task of relative rotation estimation in such extreme non-overlapping settings [8, 12]. Prior work has proposed end-to-end neural architectures, demonstrating that hidden cues," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 332, + 224, + 538, + 352 + ], + "blocks": [ + { + "bbox": [ + 332, + 224, + 538, + 352 + ], + "lines": [ + { + "bbox": [ + 332, + 224, + 538, + 352 + ], + "spans": [ + { + "bbox": [ + 332, + 224, + 538, + 352 + ], + "type": "image", + "image_path": "6f2c5c753c36243e5c09bc76366efbfdb87a6d1c8637dd6c805dbb6f6d469c73.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 360, + 555, + 428 + ], + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 428 + ], + "type": "text", + "content": "Figure 1. Given a pair of (possibly) non-overlapping images captured in the wild—e.g., under arbitrary illumination and intrinsic camera parameters—such as the images of the Dam Square in Amsterdam depicted in red and blue boxes above*, our technique estimates the relative 3D rotation between the images. *The background panorama is illustrated for visualization purposes only." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": "such as vanishing points or the directions of cast shadows, can implicitly guide the model for inferring the relative orientation between the images. To facilitate learning and evaluation, datasets constructed from panoramic views were adopted. These datasets emulate perspective views by cropping sub-areas from these panoramas, enabling generation of image pairs with various degrees of overlap. However, while such emulated views perhaps capture some of the challenges associated with extreme-view imagery, are they sufficient for representing real images—particularly, images captured in the wild?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we present a new approach that tackles the problem of extreme rotation estimation in the wild. Consider the boxed images in Figure 1. Internet (i.e., in the wild) images may vary due to a wide range of factors, including transient objects, weather conditions, time of day, and the cameras' intrinsic parameters. To explore this problem, we introduce a new dataset, ExtremeLandmarkPairs (ELP), assembled from publicly-available scene-level Internet image collections. We observe that the set of real extreme-view image pairs is limited, as Internet datasets are typically scene-centric, with nearby cameras commonly" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1061" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 178 + ], + "type": "text", + "content": "capturing overlapping views. Therefore, to facilitate training, we propose a progressive learning scheme that leverages and augments images cropped from panoramic views, allowing for gradually generalizing the model onto real Internet data. In particular, we construct datasets with varying field of views, that better resemble the distribution of real data, and perform image-level appearance augmentations by leveraging recent advancements in text-to-image Diffusion models [7, 22, 31]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 180, + 294, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 180, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 57, + 180, + 294, + 358 + ], + "type": "text", + "content": "To estimate extreme rotations in the wild, we propose a Transformer-based model that is provided with auxiliary channels, including the spatial distribution of local keypoints and matches and semantic segmentation maps, allowing for better reasoning over real image pairs with little or no overlap. Our results demonstrate that our model can accurately predict the relative rotations for a wide variety of extreme-view image pairs that vary in illumination, dynamic regions, and intrinsic parameters. We conduct extensive experiments, quantifying performance both over real Internet data and also over emulated perspective images cropped from panoramic views. Our evaluation shows that our model significantly improves over strong baselines when considering real images, while achieving comparable performance over emulated perspective image pairs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 370, + 140, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 370, + 140, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 140, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 390, + 294, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 390, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 390, + 294, + 521 + ], + "type": "text", + "content": "Relative pose estimation is a fundamental task in computer vision, typically studied for overlapping camera views. Traditionally, this task has been divided into two stages: correspondence estimation from local feature matching, followed by geometry-based pose estimation. In recent years, feature matching methods have advanced from using heuristic feature descriptors [5, 27, 32] and RANSAC-based matches [21] to learning-based feature extraction [4, 14-16, 18, 40, 45] and matching techniques [3, 26, 33], with several methods performing both feature extraction and matching using unified learning-based frameworks [4, 17, 38]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 522, + 294, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 581 + ], + "type": "text", + "content": "These methods are generally invariant to changes in illumination and appearance, demonstrating robust performance across various scene scales and also over in-the-wild datasets. However, their reliability diminishes in extreme view scenarios due to their dependence on visual overlap." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 582, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 712 + ], + "type": "text", + "content": "Extreme-view scenarios lacking local pixelwise correspondences necessitate the use of end-to-end learning-based pose estimation methods which directly infer the 3D relationships and geometry from sparse and extreme-view images. Indeed, large-scale 3D object datasets have paved the way for learning-based methods which estimate camera pose directly from sparse views [20, 25, 37, 43, 44, 50-52]. However, these methods primarily concentrate on object-centric scenes, under experimental settings that typically assume similar lighting and camera intrinsics for the input views. Furthermore, these methods often utilize bounding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 72, + 552, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 95 + ], + "type": "text", + "content": "box inputs defining the object of interest, which is less suitable in the case of images depicting large-scale scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 97, + 553, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 226 + ], + "type": "text", + "content": "Several prior works address a sparse view setting at scene-scale. In particular, Chen et al. [11] propose to learn a discrete distributions of pose space, Agarwala et al. [1] simplify scene reconstruction using a plane representation, and Rockwell et al. [30] introduce an inductive bias of the 8-point algorithm into a vision transformer architecture. Recently, models that directly predict pixel-aligned point maps from input pairs (or sparse image collections), such as DUSt3R [46] and Mast3R [23], have demonstrated promising results on pose estimation and scene-scale 3D reconstruction of Internet data with a wide baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 228, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 441 + ], + "type": "text", + "content": "Camera pose estimation for non-overlapping views presents a greater challenge. Earlier efforts [10] explored searching for consistent temporal behavior. Several work utilize pairwise RGB and depth scan data to estimate relative pose among such extreme pairs [48, 49]. Cai et al. [8] tackle pose estimation for non-overlapping views without the use of additional data, by introducing a learning-based network leveraging cross-correlation volume to exploit implicit cues. This correlation volume is later enhanced through the integration of transformer attention modules [12]. Nonetheless, these approaches assume constrained 3D environments, including the assumption of consistent lighting and camera intrinsics, and are designed for camera distribution of emulated perspective views cropped from panoramas. In this work, we aim to address pose estimation for realistic in-the-wild non-overlapping image pairs, enhancing the applicability of extreme pose estimation to Internet photos and real-world data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 453, + 512, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 453, + 512, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 512, + 465 + ], + "type": "text", + "content": "3. The ExtremeLandmarkPairs Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 473, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 473, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 473, + 553, + 592 + ], + "type": "text", + "content": "Prior works on extreme pose estimation use panoramic views, cropping from it sub-areas to emulate perspective views [8, 12]. To evaluate and train models on real perspective image pairs, we propose a new benchmark and dataset, ExtremeLandmarkPairs (ELP), constructed from Internet image pairs from the MegaDepth [24], Cambridge Landmarks [2], and MegaScenes [39] datasets. In this section, we first describe the dataset construction procedure (Section 3.1), and then present details regarding dataset size and train and test splits (Table 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 599, + 435, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 599, + 435, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 599, + 435, + 610 + ], + "type": "text", + "content": "3.1. Dataset Construction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 618, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 618, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 618, + 553, + 712 + ], + "type": "text", + "content": "To construct a dataset of real perspective image pairs with varying degrees of overlap, we leverage available scene-level training data. Existing Internet image collections typically contain camera poses (predicted up to scale), which are determined using Structure-from-Motion (SfM) algorithms, such as COLMAP [36]. In what follows, we describe how we extract real image pairs from this data, which can then be used for training and evaluating models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "text", + "content": "1062" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 178, + 164 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 178, + 164 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 178, + 164 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 178, + 164 + ], + "type": "image", + "image_path": "ab05a16a75e49499ee349e8ad6e82bcbb1e38fab076b2a7571a92230872b43b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 166, + 174, + 186 + ], + "lines": [ + { + "bbox": [ + 61, + 166, + 174, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 174, + 186 + ], + "type": "text", + "content": "(a) Reconstruction and camera distribution of all images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 74, + 292, + 160 + ], + "blocks": [ + { + "bbox": [ + 181, + 74, + 292, + 160 + ], + "lines": [ + { + "bbox": [ + 181, + 74, + 292, + 160 + ], + "spans": [ + { + "bbox": [ + 181, + 74, + 292, + 160 + ], + "type": "image", + "image_path": "eb534a776b1ea06d3928ea43616cafac13240a4ffebb243d11b2ce45c0e8840b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 165, + 287, + 186 + ], + "lines": [ + { + "bbox": [ + 189, + 165, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 189, + 165, + 287, + 186 + ], + "type": "text", + "content": "(b) Camera distribution of large-overlapping pairs" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 193, + 293, + 413 + ], + "blocks": [ + { + "bbox": [ + 60, + 193, + 293, + 413 + ], + "lines": [ + { + "bbox": [ + 60, + 193, + 293, + 413 + ], + "spans": [ + { + "bbox": [ + 60, + 193, + 293, + 413 + ], + "type": "image", + "image_path": "03ee3c05ba1b6fd796e0cf91f897272bc55e5adab7324d19d6661b4f09c50f17.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 415, + 265, + 426 + ], + "lines": [ + { + "bbox": [ + 86, + 415, + 265, + 426 + ], + "spans": [ + { + "bbox": [ + 86, + 415, + 265, + 426 + ], + "type": "text", + "content": "(c) Camera distribution of non-overlapping pairs" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 432, + 295, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 477 + ], + "type": "text", + "content": "Figure 2. Camera distribution of the Vatican, Rome scene from the ExtremeLandmarkPairs Dataset. We construct a dataset of real perspective image pairs with predominant rotational motion shown in (b) and (c) from the dense imagery reconstruction in (a)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 510 + ], + "type": "text", + "content": "Identifying Pairs with Predominant Rotational Motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 510, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 676 + ], + "type": "text", + "content": "Prior works targeting relative rotation estimation, in particular for non-overlapping views, mostly utilize panoramic views, focusing on image pairs with purely rotational motion. Pairs belonging to real image collections almost always contain a non-negligible translation component. Furthermore, unlike in the StreetLearn [28] dataset used by prior work [8, 12] that provides exact translation values between consecutive panoramas which allows for filtering pairs with predominant rotational motions, reconstructed relative poses are only provided up to scale. The scale varies among different reconstructed scenes, and therefore there's no global threshold on the relative translation values which can be used for identifying pairs with predominant rotational motion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To automatically identify such pairs, we observe that available Internet collections require the existence of dense imagery, to compensate for the vast number of unknowns" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "in the SfM optimization. We therefore construct mutual nearest neighbors edge-weighted graphs, with one graph per landmark. In each graph " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": ", nodes " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "v \\in V" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " correspond to images, and two images are connected by an edge " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "e \\in E" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " if they are both among each other's " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " nearest neighbors, considering L2 distances between their translations (" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " is empirically set to 5). Note that images captured from sparser (outlier) regions in space are unlikely to be within the mutual " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " nearest neighbors of images captured within denser regions, and hence won't be included in " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": ". Finally, we select a subset of image pairs containing relatively small distances from each scene graph " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": ", yielding a set of image pairs with predominant rotational motion; see the supplementary material for additional details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 242, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 338 + ], + "type": "text", + "content": "Extracting Level of Overlap. Following prior work [8, 12], we are interested in training and evaluating models according to three different categories: Large, Small and None, indicating image pairs with a varying amount of overlap. However, unlike prior work that use cropped images with a fixed " + }, + { + "bbox": [ + 313, + 242, + 553, + 338 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 242, + 553, + 338 + ], + "type": "text", + "content": " FoV, Internet images contain varying FoV values. Thus, the relative rotation angle is not sufficient for extracting the pair's overlap level." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "text", + "content": "Denote the FoV values of image " + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "inline_equation", + "content": "i\\in [1,2]" + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "inline_equation", + "content": "[\\mathrm{fov}_x^i,\\mathrm{fov}_y^i ]" + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "text", + "content": " We can parameterize a 3D rotation matrix " + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "text", + "content": " using three Euler angles " + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "inline_equation", + "content": "[\\alpha ,\\beta ,\\gamma ]" + }, + { + "bbox": [ + 313, + 338, + 553, + 385 + ], + "type": "text", + "content": " denoting the relative roll, pitch and yaw angles, respectively:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 397, + 553, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 397, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 361, + 397, + 553, + 410 + ], + "type": "interline_equation", + "content": "\\mathbf {R} (\\alpha , \\beta , \\gamma) = \\mathbf {R} _ {z} (\\alpha) \\mathbf {R} _ {y} (\\beta) \\mathbf {R} _ {x} (\\gamma), \\tag {1}", + "image_path": "31d398e146ceff198e4c1f906d8b985677bf07359e905cad740cfabeb730a221.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 413, + 553, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 553, + 448 + ], + "type": "text", + "content": "following Dense Correlation Volumes' coordinate system. We use the following conditions to determine the overlap level " + }, + { + "bbox": [ + 313, + 413, + 553, + 448 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 413, + 553, + 448 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 464, + 553, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 464, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 324, + 464, + 553, + 515 + ], + "type": "interline_equation", + "content": "o = \\left\\{ \\begin{array}{l l} L a r g e & | \\gamma | < \\frac {\\operatorname {f o v} _ {x} ^ {1} + \\operatorname {f o v} _ {x} ^ {2}}{4} \\wedge | \\beta | < \\frac {\\operatorname {f o v} _ {y} ^ {1} + \\operatorname {f o v} _ {y} ^ {2}}{4} \\\\ N o n e & | \\gamma | > \\frac {\\operatorname {f o v} _ {x} ^ {1} + \\operatorname {f o v} _ {x} ^ {2}}{2} \\wedge | \\beta | > \\frac {\\operatorname {f o v} _ {y} ^ {1} + \\operatorname {f o v} _ {y} ^ {2}}{2} \\\\ S m a l l & e l s e \\end{array} \\right. \\tag {2}", + "image_path": "ae6b4933b11e6ed79a56b710d3a68c6474f473725c87db1c5626d1932a76f062.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 520, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 604 + ], + "type": "text", + "content": "In other words, pairs with relative yaw and pitch angles that are smaller than a quarter of the average corresponding FoV values are considered as pairs with a large overlap ratio. Likewise, pairs with relative yaw and pitch angles that are larger than half the average corresponding FoV values are considered non-overlapping pairs. Pairs in-between these conditions are considered pairs with a small overlap ratio." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Additional Filtering. In the scene-scale datasets we explore, large FoV disparities could result in one image focusing on specific architectural details like a statue, while the other captures a much broader scene perspective, further complicating the problem of estimating the relative rotations. To extract pairs consistent in scale, we limit the difference between the FoV values to be at most " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". Furthermore, as most images contain small roll values, we rotate the scenes to match the gravity axis and horizontal axis" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1063" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 294, + 144 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 144 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 144 + ], + "type": "table", + "html": "
SubsetSource#Pairs
#ScenesLargeSmallNoneTotal
Train[39]588333430136842948176595
Validation[39]515339871041308238
Validation Balanced[39]1779255707854
sELP[2]6251282719615300
wELP[24]1727008296434172
", + "image_path": "52c9c17204310d35fe7f126af2981f620270843ef4044f4a21efd7d5c7c9d38f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 191 + ], + "type": "text", + "content": "Table 1. ExtremeLandmarkPairs Dataset Statistics. Above, we report the number of image pairs extracted for each overlap level, split into train and test (with sELP denoting a single camera setting and wELP denoting the \"in the wild\" setting)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 212, + 295, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 261 + ], + "type": "text", + "content": "and filtered images with rolls exceeding " + }, + { + "bbox": [ + 55, + 212, + 295, + 261 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 55, + 212, + 295, + 261 + ], + "type": "text", + "content": ". Finally, we excluded images captured from an aerial perspective, by filtering images with translation along y-axis exceeding a global threshold, empirically set to 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 268, + 185, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 185, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 185, + 282 + ], + "type": "text", + "content": "3.2. Dataset Size and Splits" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 286, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 370 + ], + "type": "text", + "content": "We apply the procedure described above over landmarks from the MegaScenes [39] dataset to create a training set and a validation set. As we focus on outdoor scenes in our work, we filter reconstructions that capture indoor scenes. We obtain a total of nearly 34K non-overlapping pairs originating from over 2K unique landmarks. The validation set was also balanced the set over the overall angle." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 370, + 295, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 295, + 477 + ], + "type": "text", + "content": "For evaluation, we create two test sets, to separately examine image pairs captured in a single camera setting with constant illumination (sELP) and image pairs captured in the wild (wELP). Image pairs in the sELP test set contain images from the Cambridge Landmarks [2] dataset, which contains videos capturing six different landmarks. Image pairs in the wELP test set contain images from the MegaDepth dataset [24], which contains Internet photos from Flickr for a set of large-scale landmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "text", + "content": "As there is some overlap between the landmarks in MegaScenes [39] and MegaDepth [24], we performed additional filtering, ensuring no overlap exists between the train and test sets. Furthermore, we filter test pairs if one of the images is highly transient—i.e., if transient objects dominate the image. We quantify this using a pretrained segmentation model [19], filtering images containing transient objects in over " + }, + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "text", + "content": " of the pixels. We also manually validate non-overlapping image pairs, allowing to further filter unlabeled objects, such as a stage or a market stand. Finally, for non-overlapping image pairs, we observed that the overall relative rotations are highly imbalanced, and therefore we balance this set by the overall relative angle. Table 1 summarizes the number of image pairs and landmarks used for both training and test." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Given a pair of Internet images with (possibly) extreme relative motion, we estimate the relative rotation " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " between" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "the images. Following prior works on extreme rotation estimation [8, 12], we assume a camera-centric setting, where the two cameras have limited translation. However, our approach departs from prior works by operating on outdoor images captured by a crowd of photographers, with varying intrinsic parameters as well as appearances—" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "e.g." + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", due to illumination changes and dynamic objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 553, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 242 + ], + "type": "text", + "content": "Our model (detailed in Section 4.1; see Figure 3 for an overview) outputs three discrete Euler angles, denoting the relative roll, pitch and yaw angles. As illustrated in prior work [8], this parameterization enables using a simple cross-entropy loss for training. In Section 4.2, we describe our progressive learning scheme, allowing for gradually adapting the model to extreme Internet imagery." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 250, + 367, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 367, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 367, + 262 + ], + "type": "text", + "content": "4.1. Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 269, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 365 + ], + "type": "text", + "content": "We extract image features using a pretrained LoFTR model [38]. In contrast to the features extracted using common convolutional neural networks pretrained on ImageNet [13], LoFTR is a Transformer-based model trained on Internet pairs, with the goal of extracting local feature matches – a setting and task which is highly related to the one we address in our work, thus enabling extraction of better (i.e., more relevant) features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 366, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 556 + ], + "type": "text", + "content": "As we are interested in designing a network that can also reason over image pairs with little or no overlap, we combine the extracted features with additional auxiliary channels; see Figure 3 (bottom left). These include keypoint and pairwise matches masks, utilized previously for disambiguating images for similar structures [9]. Intuitively, knowledge over pairwise matches can assist the model in cases of small overlap and for generalizing across different camera intrinsic properties. We also incorporate a segmentation map as an additional auxiliary channel, which segments images into several categories (such as sky, building, road and sidewalk). This channel allows for identifying additional cues, such as the skyline or transient objects, which can aid in determining the rotation for nonoverlapping pairs. In Section 5, we demonstrate the benefit of incorporating these auxiliary channels in our model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "We then reshape extracted features (and auxiliary channels) to tokens, concatenating these image tokens with learnable Euler angle position embeddings. These tokens are processed by our Rotation Estimation Transformer module, which uses a transformer decoder architecture [41]. The output Euler angle tokens obtain information from image features and auxiliary channels within transformer attention modules. These tokens are then processed by three different prediction heads (one per angle, denoted as MLP in Figure 3). Each output prediction head uses as input averaged image tokens and one of the output Euler angle tokens, which provides the model with additional angle-specific information, allowing for achieving improved performance," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1064" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 78, + 536, + 212 + ], + "blocks": [ + { + "bbox": [ + 66, + 78, + 536, + 212 + ], + "lines": [ + { + "bbox": [ + 66, + 78, + 536, + 212 + ], + "spans": [ + { + "bbox": [ + 66, + 78, + 536, + 212 + ], + "type": "image", + "image_path": "e05a606a2a754adc8b223383658e7090df6e090689bc37d52b33935ae342b1ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 216, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 54, + 216, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 555, + 273 + ], + "type": "text", + "content": "Figure 3. Method architecture. Given a pair of input Internet images, we extract image features using pretrained LoFTR. These features are combined with auxiliary channels, including keypoint and pairwise matches masks, and segmentation maps (visualized on the bottom left). These image features are reshaped into tokens and concatenated with Euler angle position embeddings, which are then processed by our Rotation Estimation Transformer module. The output Euler angle tokens and averaged image tokens are concatenated and processed by MLPs to predict the probability distribution of Euler angles, representing the relative 3D rotation between the input images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 293, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 331 + ], + "type": "text", + "content": "as we show in the supplementary material. The prediction heads output a probability distribution over " + }, + { + "bbox": [ + 54, + 293, + 295, + 331 + ], + "type": "inline_equation", + "content": "N = 360" + }, + { + "bbox": [ + 54, + 293, + 295, + 331 + ], + "type": "text", + "content": " bins, capturing an angle in the range " + }, + { + "bbox": [ + 54, + 293, + 295, + 331 + ], + "type": "inline_equation", + "content": "[-180^{\\circ}, 180^{\\circ}]" + }, + { + "bbox": [ + 54, + 293, + 295, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 122, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 122, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 122, + 351 + ], + "type": "text", + "content": "4.2. Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 357, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 514 + ], + "type": "text", + "content": "As detailed in Section 3, we assemble real image pairs from Internet imagery, which can be used for training and evaluating models. However, even with our proposed ExtremeLandmarkPairs dataset, the set of real image pairs is limited—e.g., only " + }, + { + "bbox": [ + 54, + 357, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\sim 36\\mathrm{K}" + }, + { + "bbox": [ + 54, + 357, + 295, + 514 + ], + "type": "text", + "content": " non-overlapping image pairs are extracted, as available Internet datasets are typically scene-centric, with nearby cameras usually capturing overlapping views. Therefore, in what follows, we propose a progressive learning scheme, which leverages panoramic images, and allows for gradually generalizing the model to images captured in-the-wild. All learning stages are optimized using a cross-entropy loss for each Euler angle prediction. Additional details are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 517, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 517, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 54, + 517, + 295, + 615 + ], + "type": "text", + "content": "Initialization. We begin by training our model on the perspective views cropped from panoramic images, using the data created by Cai et al [8]. Specifically, we use the image pairs cropped from panoramic images included in the StreetLearn [28] dataset, depicting various streets in Manhattan. As further detailed in Cai et al [8], this training set includes roughly 1M image pairs sampled from the same panorama, split according to the overlap level." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 715 + ], + "type": "text", + "content": "Training with Data Augmentations. We observe that the dataset constructed by prior work could be augmented to better capture the distribution of image pairs captured inthe-wild. In particular, we focus on two types of data augments, which we elaborate on next: (i) field of view (FoV) augmentations (denoted henceforth as " + }, + { + "bbox": [ + 54, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{FoV}" + }, + { + "bbox": [ + 54, + 617, + 295, + 715 + ], + "type": "text", + "content": ") and (ii) image-level appearance augmentations (denoted henceforth as " + }, + { + "bbox": [ + 54, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{Im}" + }, + { + "bbox": [ + 54, + 617, + 295, + 715 + ], + "type": "text", + "content": "). In Section 5, we demonstrate how both types" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 290, + 375, + 348 + ], + "blocks": [ + { + "bbox": [ + 317, + 290, + 375, + 348 + ], + "lines": [ + { + "bbox": [ + 317, + 290, + 375, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 290, + 375, + 348 + ], + "type": "image", + "image_path": "f41d914e36d412fff29247b302c386059ab2a8610ed6d9bee580eec36fb3de22.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 380, + 290, + 436, + 348 + ], + "blocks": [ + { + "bbox": [ + 380, + 290, + 436, + 348 + ], + "lines": [ + { + "bbox": [ + 380, + 290, + 436, + 348 + ], + "spans": [ + { + "bbox": [ + 380, + 290, + 436, + 348 + ], + "type": "image", + "image_path": "8c65f2ab42e9dd59fba11dcdd8ead8234b7fcb4a21ccd1c3e872215dbd203f15.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 290, + 494, + 348 + ], + "blocks": [ + { + "bbox": [ + 438, + 290, + 494, + 348 + ], + "lines": [ + { + "bbox": [ + 438, + 290, + 494, + 348 + ], + "spans": [ + { + "bbox": [ + 438, + 290, + 494, + 348 + ], + "type": "image", + "image_path": "179fa16bb1559513abd0d439d256fc57c4c1938319cfa3ed368a5802381e7de8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 495, + 290, + 553, + 348 + ], + "blocks": [ + { + "bbox": [ + 495, + 290, + 553, + 348 + ], + "lines": [ + { + "bbox": [ + 495, + 290, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 495, + 290, + 553, + 348 + ], + "type": "image", + "image_path": "34f1f671f5b12ba90fecc9decf60c69be8821fa88b6785834ff8cc8f11c91da0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 351, + 375, + 418 + ], + "blocks": [ + { + "bbox": [ + 317, + 351, + 375, + 418 + ], + "lines": [ + { + "bbox": [ + 317, + 351, + 375, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 351, + 375, + 418 + ], + "type": "image", + "image_path": "17acefd6b1feb0660a88f8c83f58cff293545ec1c6968b0d355f850e1af137d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 422, + 430, + 433 + ], + "lines": [ + { + "bbox": [ + 386, + 422, + 430, + 433 + ], + "spans": [ + { + "bbox": [ + 386, + 422, + 430, + 433 + ], + "type": "text", + "content": "busy street" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 380, + 351, + 436, + 418 + ], + "blocks": [ + { + "bbox": [ + 380, + 351, + 436, + 418 + ], + "lines": [ + { + "bbox": [ + 380, + 351, + 436, + 418 + ], + "spans": [ + { + "bbox": [ + 380, + 351, + 436, + 418 + ], + "type": "image", + "image_path": "26e848fbc28cb8551610bc23778892c0d8f002adf89b77ddaeb44add11918361.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 436, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 502 + ], + "type": "text", + "content": "Figure 4. Augmenting perspective images cropped from panoramic views with image-level appearance modifications. Given an input image (left) and a target text prompt \"Make it " + }, + { + "bbox": [ + 313, + 436, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\langle w\\rangle" + }, + { + "bbox": [ + 313, + 436, + 555, + 502 + ], + "type": "text", + "content": " (\" " + }, + { + "bbox": [ + 313, + 436, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\langle w\\rangle" + }, + { + "bbox": [ + 313, + 436, + 555, + 502 + ], + "type": "text", + "content": " is specified above), we use a conditional Diffusion model [7] to create semantic appearance augmentations which modify both the global image characteristics as well as local image regions." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 438, + 351, + 494, + 418 + ], + "blocks": [ + { + "bbox": [ + 438, + 351, + 494, + 418 + ], + "lines": [ + { + "bbox": [ + 438, + 351, + 494, + 418 + ], + "spans": [ + { + "bbox": [ + 438, + 351, + 494, + 418 + ], + "type": "image", + "image_path": "b992877c1d45e880bb2fccb8cbf65d94ee8db9d7f75927dcb981a4b9e4c64921.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 422, + 477, + 432 + ], + "lines": [ + { + "bbox": [ + 455, + 422, + 477, + 432 + ], + "spans": [ + { + "bbox": [ + 455, + 422, + 477, + 432 + ], + "type": "text", + "content": "night" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 495, + 351, + 553, + 418 + ], + "blocks": [ + { + "bbox": [ + 495, + 351, + 553, + 418 + ], + "lines": [ + { + "bbox": [ + 495, + 351, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 495, + 351, + 553, + 418 + ], + "type": "image", + "image_path": "1e07e2abc8014def5d0a62032f02037bbd9b705d41c1a6536c7bd78cd209c4f2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 422, + 545, + 432 + ], + "lines": [ + { + "bbox": [ + 503, + 422, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 503, + 422, + 545, + 432 + ], + "type": "text", + "content": "Christmas" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 530, + 553, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 553, + 554 + ], + "type": "text", + "content": "of augmentations improve the model's ability in generalizing to real Internet scenes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": "Rather than cropping perspective images with a fixed " + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": " FoV as was done in prior work, we analyze the FoV values of the images belonging to the ELP training set. We compute the mean and standard-deviation values, denoted as " + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": ", respectively. We then construct new perspective images by sampling from a Gaussian distribution " + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "G(\\mu, a \\cdot \\sigma)" + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": " that resembles the distribution of real data, setting " + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "a = 1.5" + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": " for obtaining a more diverse set, which also bears higher similarity to the perspective images used during initialization. We also allow for FoV differences (of up to " + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": ") between the two images paired together. Additionally, rather than providing the model with the full content within these regions, we construct crops with various aspect ratios to fur" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1065" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 177, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 177, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 177, + 84 + ], + "type": "text", + "content": "ther emulate real image pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 288 + ], + "type": "text", + "content": "In-the-wild images differ not only in their intrinsic parameters, but also in their appearance - due to presence of varying illumination and dynamic objects. Therefore, in addition to augmenting the dataset with images of various field of views, we also perform image-level appearance augmentations, leveraging recent advancements in text-to-image Diffusion models [7, 22, 31]. In particular, we apply the conditional InstructPix2Pix [7] model on a subset of our data, using a set of text prompts of the format \"Make it " + }, + { + "bbox": [ + 55, + 85, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\langle w\\rangle" + }, + { + "bbox": [ + 55, + 85, + 296, + 288 + ], + "type": "text", + "content": " \", where " + }, + { + "bbox": [ + 55, + 85, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\langle w\\rangle" + }, + { + "bbox": [ + 55, + 85, + 296, + 288 + ], + "type": "text", + "content": " captures diverse appearance modifications. See Figure 4 for a few illustrative examples; note how some of these modify the global illumination (e.g., altering the time of day or the season) while others yield local edits (such as adding more people to make it \"a busy street\"). We performed additional filtering on the augmented set, to avoid significant edits that also altered the structure of the scene; see the supplementary material for additional details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": "Training on Real Data. Finally, we finetune the model over real image pairs from our proposed ELP dataset. To prioritize learning over extreme image pairs, we first pass a batch of non-overlapping image pairs, followed by a batch of overlapping image pairs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 190, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 190, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 190, + 371 + ], + "type": "text", + "content": "5. Results and Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 426 + ], + "type": "text", + "content": "To validate our method, we compare our model with prior relative rotation estimation methods on the sELP and wELP test sets. In particular, our experiments seek to answer the following questions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 427, + 295, + 498 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 427, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 449 + ], + "type": "text", + "content": "- How well do previous methods perform on our proposed task of extreme rotation estimation in the wild?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 485 + ], + "type": "text", + "content": "- How important is our progressive training scheme, and to what extent can the improved performance be attributed to training on the ExtremeLandmarkPairs dataset?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 235, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 235, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 235, + 498 + ], + "type": "text", + "content": "- What is the impact of our design choices?" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": "We also present qualitative results over different overlap levels in Figure 5. These results illustrate our model's performance, and its robustness to varying illumination conditions and transient objects (such as the screen and the beer cup in the rightmost examples). Please refer to the supplementary material for many more qualitative results, including visualizations for baseline methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 587, + 122, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 122, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 122, + 598 + ], + "type": "text", + "content": "5.1. Baselines" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "We compare our method to correspondence-based techniques, dedicated (end-to-end) relative rotation estimation techniques as well as recent relative pose estimation techniques. Specifically, we consider two correspondence-based methods: SIFT [27] and LoFTR [38]. These methods extract keypoint matches to compute an essential matrix using the RANSAC algorithm and calculate the rotation that stems from the essential matrix. We also compare against two prior works specifically targeting extreme" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "rotation estimation using end-to-end deep learning frameworks: DenseCorrVol [8] and CascadedAtt [12]. These methods are trained and evaluated on images sampled from panoramic views from StreetLearn [28]. Results for CascadedAtt are reported over a reproduced model, as a pretrained model is not provided; additional details are provided in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 157, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 336 + ], + "type": "text", + "content": "We also consider 8PointViT [30], which estimates the relative pose between two images with a Vision Transformer that is modified to be close to the eight point algorithm. Following prior work [12], we only report results over overlapping settings for this baseline. Furthermore, as it assumes a single camera setting, we evaluate it only on the sELP test set. Finally, we consider Dust3R [46], a recent technique for dense and unconstrained stereo 3D reconstruction. While this work does not specifically target the setting of extreme rotation estimation, we examine to what extent they can be adapted for this task in the wild. Note that Dust3R was trained on MegaDepth. Since our wELP test set is constructed from image pairs from this dataset, it is not entirely fair comparison; therefore, Dust3R's results on the wELP test set are highlighted in gray." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 342, + 427, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 427, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 427, + 354 + ], + "type": "text", + "content": "5.2. Evaluation Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 384 + ], + "type": "text", + "content": "For each image pair, we compute the geodesic error, defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 390, + 553, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 390, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 367, + 390, + 553, + 418 + ], + "type": "interline_equation", + "content": "\\operatorname {e r r} = \\arccos \\left(\\frac {t r \\left(\\mathbf {R} ^ {T} \\mathbf {R} ^ {*}\\right) - 1}{2}\\right) \\tag {3}", + "image_path": "3d41c790881f9625c3245e9c6666657e5864da8d3454ca015adee0fe717e5e53.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": " is the predicted rotation matrix, and " + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^*" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": " is the ground truth relative rotation matrix. We report the median geodesic error (MGE) and relative rotation accuracy (RRA) for each test set and overlap ratio. RRA is reported over a predefined threshold " + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": ", indicating the percentage of image pairs with relative rotation error below " + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": ". We report this metric for " + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\tau = 15^\\circ" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathrm{RRA}_{15}" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\tau = 30^\\circ" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathrm{RRA}_{30}" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 509, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 593 + ], + "type": "text", + "content": "Additionally, as we discretize the space of rotations and estimate three rotation angles, each predicted as a probability distribution over " + }, + { + "bbox": [ + 313, + 509, + 554, + 593 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 509, + 554, + 593 + ], + "type": "text", + "content": " bins, we also report the performance of the Top 5 predictions in our ablation study. The Top 5 prediction reports the lowest error, considering the top-5 peaks in the relative yaw prediction only, as we observe that errors are mostly a function of the relative yaw angles." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 599, + 457, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 457, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 457, + 612 + ], + "type": "text", + "content": "5.3.Quantitative Comparison" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 689 + ], + "type": "text", + "content": "The main quantitative results comparing our method to other methods on rotation estimation in the wild are reported in Table 2. In the supplementary material, we conduct an additional evaluation in the more constrained setting examined by prior work targeting extreme rotation estimation (i.e., images cropped from panoramas)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "As illustrated in Table 2, SIFT and LoFTR, which are correspondence-based methods, exhibit some robustness" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1066" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 521, + 196 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 521, + 196 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 521, + 196 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 521, + 196 + ], + "type": "image", + "image_path": "e81900b6e1f44b20a73b6e30ac58d8caf95356429ab9755d839210925da916ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": "Figure 5. Qualitative results on the wELP test set. We visualize the results of our model over different overlap levels, where the images on the left serve as the reference points, and their coordinate system determines the relative rotation, which defines the images on the right. The ellipsoids representing the ground truth are color-coded to match their respective images, with the estimated relative rotation illustrated by a cyan dashed line. As illustrated by the examples above, our method can accurately predict relative rotations for diverse image pairs containing varying appearances and intrinsic parameters. Please refer to the supplementary material for additional qualitative results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 274, + 294, + 496 + ], + "blocks": [ + { + "bbox": [ + 56, + 274, + 294, + 496 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 496 + ], + "type": "table", + "html": "
MethodsELPwELP
MGE↓RRA15↑RRA30↑MGE↓RRA15↑RRA30↑
LargeSIFT* [27]1.9592.395.32.9474.680.8
LoFTR* [38]1.7697.299.12.1385.293.8
DenseCorrVol [8]98.5125.933.3120.537.013.0
CascadedAtt [12]29.7542.750.0170.627.39.2
8PointViT [30]22.3331.964.8---
Dust3R [46]0.7799.799.91.0198.499.2
Ours2.4596.796.82.4197.597.9
SmallSIFT* [27]5.0764.771.37.2761.468.4
LoFTR* [38]2.7081.593.66.8066.681.2
DenseCorrVol [8]143.472.89.4125.733.19.4
CascadedAtt [12]148.440.03.0139.142.74.4
8PointViT [30]51.301.712.8---
Dust3R [46]1.9695.994.62.8089.894.4
Ours4.3588.389.04.4787.291.6
NoneSIFT* [27]121.942.75.4122.840.02.0
LoFTR* [38]---56.540.033.0
DenseCorrVol [8]77.109.027.082.042.913.7
CascadedAtt [12]69.698.423.178.607.520.8
Dust3R [46]114.3319.823.981.2115.426.9
Ours13.6252.759.726.9736.150.7
", + "image_path": "dbbb878b0485d7a9fbd49223b3cc75f9f4189569eee2979fac297e41c9581914.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 578 + ], + "type": "text", + "content": "Table 2. Rotation Estimation in the Wild. We evaluate performance over the sELP and wELP test sets, separately considering Large (top), Small (middle) and Non-overlapping (bottom) pairs. * indicates median errors are computed only over successful image pairs, for which these algorithms output a pose estimate (failure over more than " + }, + { + "bbox": [ + 55, + 498, + 295, + 578 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 498, + 295, + 578 + ], + "type": "text", + "content": " of the test pairs is shown in gray). Note that Dust3R was trained on images from wELP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "when handling highly overlapping Internet images, compared to methods that trained on images cropped from panoramas, achieving a median error of less than " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "3^{\\circ}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " for Large overlap pairs in both test sets. However, these methods rely on image overlap and may not always provide an output of estimated camera pose, as the geometric verification requires a sufficient number of detected inliers. Therefore, they struggle to produce reliable matches in cases of limited overlapping regions, as also observed in prior work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 276, + 470, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 470, + 288 + ], + "type": "text", + "content": "focusing on extreme scenarios [8, 12]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 289, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 289, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 312, + 289, + 555, + 432 + ], + "type": "text", + "content": "DenseCorrVol, CascadedAtt and 8PointViT exhibit relatively poor performance on the ELP test sets, illustrating that models trained on images sampled from panoramicas cannot easily generalize to Internet photos. In the supplementary material, we show that our model significantly outperforms these prior work even when trained on the same dataset (i.e., images with a constant FoV and illumination cropped from panoramicas). As further discussed there, we believe this gap can be partially attributed to the usage of a pretrained LoFTR feature extractor, which is capable of encoding knowledge between Internet image pairs (which vary in their intrinsic and appearance)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 434, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 649 + ], + "type": "text", + "content": "Dust3R operates on Internet datasets without the need for calibration and performs exceptionally well on overlapping pairs, achieving the lowest median error and highest success rates for Large and Small overlap categories for sELP test set. However, this method is designed for overlapping pairs as it initializes the model using pretrained CroCo [47], which is trained to address cross-view completion problem from two overlapping views. Furthermore, its output consists of a unified dense point cloud for each pair of images. Due to its design for overlapping pairs, its performance on non-overlapping views is extremely low, also on the wELP test set, which contains images from scenes included in its training set. Moreover, in terms of size, Dust3R (with a DPT head) contains 577 million parameters. In contrast, our model is significantly more compact, with only 80 million parameters. As shown in the table, despite its smaller size, our model yields significantly improved performance over non-overlapping pairs on both test sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 658, + 407, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 407, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 407, + 671 + ], + "type": "text", + "content": "5.4. Ablation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We conduct various ablations analyzing the effect of our progressive training scheme and other design choices,also reporting performance over the Top 5 predictions (consider" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1067" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 211 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 211 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 211 + ], + "type": "table", + "html": "
OverlapTrain dataTop 1Top 5
MGE ↓RRA15 ↑RRA30 ↑MGE ↓RRA15 ↑RRA30 ↑
Large[8]13.6535.473.512.2261.484.7
+ΔFoV4.6179.781.14.4190.398.9
+ΔIm4.4690.492.44.4394.399.1
+ELP2.4197.597.92.4198.499.4
Small[8]55.283.729.129.8315.050.3
+ΔFoV12.9156.268.210.9766.085.4
+ΔIm11.4662.580.610.7368.091.0
+ELP4.4787.291.64.2491.197.2
None[8]74.9412.825.325.1126.158.8
+ΔFoV61.6225.038.416.8244.275.0
+ΔIm68.3125.036.116.2145.778.2
+ELP26.9736.150.712.8557.185.8
", + "image_path": "24c18a512effaa132fe8c1390581fb13c506d1e92d0f29311f750d7aeb8fc445.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 214, + 294, + 246 + ], + "lines": [ + { + "bbox": [ + 56, + 214, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 294, + 246 + ], + "type": "text", + "content": "Table 3. Ablation study, evaluating the effect of our progressive training scheme over the wELP test set. All experiments start with the cropped panoramas used in Cai et al. [8]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 269, + 294, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 269, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 269, + 294, + 315 + ], + "type": "text", + "content": "ering the top-5 peaks in the relative yaw prediction, as detailed in Section 5.2). Additional ablations, including an analysis of architectural components, are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "text", + "content": "The effect of our progressive training scheme. We conduct multiple experiments ablating the effect of our progressive training scheme. Our training process consists of four stages: initialization (following prior work [8]), incorporating multiple FoVs " + }, + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "inline_equation", + "content": "(+\\Delta \\mathrm{FoV})" + }, + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "text", + "content": ", training with image-level appearance augmentations " + }, + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "inline_equation", + "content": "(+\\Delta \\mathrm{Im})" + }, + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "text", + "content": ", and training with ExtremeLandmarkPairs pairs " + }, + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "inline_equation", + "content": "(+ELP)" + }, + { + "bbox": [ + 57, + 319, + 294, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "text", + "content": "Table 3 illustrates the impact of each training stage on the results. As can be observed from the table, each training stage further refines the model's performance, often in a significant manner. For instance, the median error in the wELP small overlap test set decreased from " + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "inline_equation", + "content": "55.3^{\\circ}" + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "inline_equation", + "content": "12.9^{\\circ}" + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "text", + "content": " when FoV augmentations were added. Additionally, the ExtremeLandmarkPairs training set plays a crucial role in finalizing our training process, yielding a significant reduction in the median error (more than half) for small and non-overlapping scenarios in the wELP test set. While this median error remains relatively high for non-overlapping pairs, we observe that the top-5 scores show significant improvements, e.g., reducing the median error from " + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "inline_equation", + "content": "27.0^{\\circ}" + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "inline_equation", + "content": "12.8^{\\circ}" + }, + { + "bbox": [ + 57, + 403, + 294, + 628 + ], + "type": "text", + "content": ". This demonstrates that the model has learned this knowledge, although it cannot necessarily be recovered by the largest peak alone. In the supplementary material, we demonstrate that the intermediate training stages are indeed important and that the improved performance cannot be obtained with the ExtremeLandmarkPairs training set alone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "content": "Moreover, evidenced in Table 3, our model demonstrates strong generalization capabilities even when trained exclusively on panorama-cropped images. The model generalizes to the wELP test set significantly better (e.g., for Large overlap cases, our model achieves an MGE of " + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "inline_equation", + "content": "13.65^{\\circ}" + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "content": " compared to values higher than " + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "inline_equation", + "content": "120^{\\circ}" + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "content": " achieved by baselines), demonstrating that the improvement is not just due to our" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 217 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 217 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 217 + ], + "type": "table", + "html": "
OverlapKPSMTop 1Top 5
MGE ↓RRA15↑RRA30↑MGE ↓RRA15↑RRA30↑
Large××2.1897.498.12.1897.498.1
×✓2.3097.097.42.3098.599.4
✓×2.4497.698.32.3198.499.7
✓✓2.4197.597.92.4198.499.4
Small××4.5087.991.64.5087.991.7
×✓4.4988.192.04.4691.296.7
✓×4.4187.592.24.3291.997.6
✓✓4.4787.291.64.2491.197.2
None××48.8134.044.112.5657.584.6
×✓43.0731.244.213.9953.583.2
✓×41.3935.346.813.0456.986.2
✓✓26.9736.150.712.8557.185.8
", + "image_path": "aa6689a546af5aea062cda41d378ce8bb9d6bd32f04764d6e93ac0e913aed203.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "Table 4. Ablation study, evaluating the effect of the auxiliary channels added as input to our network. We train models without the keypoints and matches (KP) and without the segmentation maps (SM), comparing to our model over the wELP test set , after using a validation split that is balanced over the overall angle." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 295, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 378 + ], + "type": "text", + "content": "progressive training scheme. We further examined architectural differences by conducting an additional ablation study, applying our progressive training scheme to the baseline models (see Table 9 in the supplementary materials). This experiment revealed that prior models are not directly suitable for real-world applications, as these baseline models showed significantly poorer performance across all metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 381, + 553, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 381, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 381, + 553, + 476 + ], + "type": "text", + "content": "The effect of adding auxiliary channels. We ablate the effect of adding auxiliary in Table 4, training models without keypoints and matches (KP) or the segmentation map (SM) provided as additional inputs. As illustrated in the table, for non-overlapping cases, these auxiliary channels boost performance almost across all metrics. In particular, both channels play a role in reducing the errors (reducing the median error from over " + }, + { + "bbox": [ + 316, + 381, + 553, + 476 + ], + "type": "inline_equation", + "content": "40^{\\circ}" + }, + { + "bbox": [ + 316, + 381, + 553, + 476 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 381, + 553, + 476 + ], + "type": "inline_equation", + "content": "27.0^{\\circ}" + }, + { + "bbox": [ + 316, + 381, + 553, + 476 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 487, + 386, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 386, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 386, + 499 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 508, + 553, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 553, + 674 + ], + "type": "text", + "content": "We present a method and benchmark dataset for estimating relative 3D rotations between pairs of (possibly) non-overlapping RGB images. Our approach extends prior work addressing extreme rotations to real-world data that exhibit variation in both appearance in intrinsic camera parameters. While our model shows promising results on real-world Internet image pairs, it also highlights the inherent difficulty of the underlying task, suggesting that considerable progress can be achieved with future techniques that leverage our dataset. Our paired data could also potentially serve for exploring the challenging task of estimating extreme translations in real-world settings. Future work could also consider incorporating more views for enhancing performance in such extreme non-overlapping scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 689, + 553, + 712 + ], + "type": "text", + "content": "Acknowledgments This work was partially supported by ISF (grant number 2510/23)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "text", + "content": "1068" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In European Conference on Computer Vision, pages 192-209. Springer, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Roberto Cipolla Alex Kendall, Matthew Grimes. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV -International Conference on Computer Vision, pages 2938-2946, 2015. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 236 + ], + "type": "text", + "content": "[3] Lionel Baboud, Martin Čadík, Elmar Eisemann, and Hans-Peter Seidel. Automatic photo-to-terrain alignment for the annotation of mountain pictures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41–48. IEEE, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[4] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In European Conference on Computer Vision, pages 751-767. Springer, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 317 + ], + "type": "text", + "content": "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417. Springer, 2006. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "type": "text", + "content": "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "type": "text", + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR IEEE / CVF Computer Vision and Pattern, pages 18392-18402, 2023. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "text", + "content": "[8] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "text", + "content": "[9] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 555 + ], + "type": "text", + "content": "[10] Yaron Caspi and Michal Irani. Aligning non-overlapping sequences. IJCV, 48(1):39-51, 2002. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 556, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 610 + ], + "type": "text", + "content": "[11] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3258-3268, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[12] Shay Dekel, Yosi Keller, and Martin Cadik. Estimating extreme 3d image rotations using cascaded attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2588-2598, 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[15] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. Camnet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2871-2880, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "text", + "content": "[16] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "type": "text", + "content": "[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17775, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[18] Johan Edstedt, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Dedode: Detect, don't describe—describe, don't detect for local feature matching. arXiv preprint arXiv:2308.08479, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[19] Xie Enze, Wang Wenhai, Yu Zhiding, Anandkumar Anima, M. Alvarez Jose, and Luo Ping. Segformer: Simple and efficient design for semantic segmentation with transformers. In Conference on Neural Information Processing Systems (NEURIPS), 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "text", + "content": "[20] Zhiwen Fan, Panwang Pan, Peihao Wang, Yifan Jiang, Dejia Xu, Hanwen Jiang, and Zhangyang Wang. Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "text", + "content": "[21] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM (CACM), 24(6):381-395, 1981. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 522, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 543 + ], + "type": "text", + "content": "[23] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "text", + "content": "[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "type": "text", + "content": "[25] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching atokens are frozen. arXiv preprint arXiv:2306.13643, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "text", + "content": "[27] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 60(2):91-110, 2004. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1069" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 139 + ], + "type": "text", + "content": "[29] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[30] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "text", + "content": "[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 295, + 342 + ], + "type": "text", + "content": "[33] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 420 + ], + "type": "text", + "content": "[34] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "text", + "content": "[35] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8601-8610, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 510 + ], + "type": "text", + "content": "[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[37] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "text", + "content": "[38] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "text", + "content": "[39] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[40] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 687 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "text", + "content": "[42] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5722-5731, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "type": "text", + "content": "[43] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 554, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 554, + 284 + ], + "type": "text", + "content": "[44] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 554, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 554, + 340 + ], + "type": "text", + "content": "[45] Qianqian Wang, Xiaowei Zhou, Bharath Hariharan, and Noah Snavely. Learning feature descriptors using camera pose supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 757-774. Springer, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 554, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 384 + ], + "type": "text", + "content": "[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. arXiv preprint arXiv:2312.14132, 2023. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 554, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 554, + 452 + ], + "type": "text", + "content": "[47] Philippe Weinzaepfel, Vincent Leroy, Thomas Lucas, Romain Brégier, Yohann Cabon, Vaibhav Arora, Leonid Antsfeld, Boris Chidlovskii, Gabriela Csurka, and Jérôme Revaud. Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. Advances in Neural Information Processing Systems, 35:3502-3516, 2022. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 554, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 554, + 496 + ], + "type": "text", + "content": "[48] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 554, + 529 + ], + "type": "text", + "content": "[49] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 531, + 554, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 554, + 596 + ], + "type": "text", + "content": "[50] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 598, + 554, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 598, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 316, + 598, + 554, + 642 + ], + "type": "text", + "content": "[51] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Reply: Predicting probabilistic relative rotation for single objects in the wild. In European Conference on Computer Vision, pages 592-611. Springer, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 643, + 554, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 643, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 316, + 643, + 554, + 687 + ], + "type": "text", + "content": "[52] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1070" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_content_list.json b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ea5db8d7c27bdb39337819bee8d55737f0c8ebe --- /dev/null +++ b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_content_list.json @@ -0,0 +1,1547 @@ +[ + { + "type": "text", + "text": "F-LMM: Grounding Frozen Large Multimodal Models", + "text_level": 1, + "bbox": [ + 218, + 130, + 777, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Size $\\mathbf{W}\\mathbf{u}^{1}$ Sheng Jin2 Wenwei Zhang3 Lumin $\\mathrm{Xu^4}$ Wentao Liu2,3 Wei Li1 Chen Change Loy' 1 S-Lab, Nanyang Technological University SenseTime Research and Tetras.AI 3 Shanghai AI Laboratory 4 The Chinese University of Hong Kong size001@e.ntu.edu.sg ccloy@ntu.edu.sg", + "bbox": [ + 104, + 179, + 915, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3e1e1ddd8c03e43d3c0dd350064ba99f869d5854bc1ac492e6c926ed0e144a9.jpg", + "image_caption": [ + "User: Is there a truck in the image? Answer the question using a single word or phrase. \nGLaMM: Sure, the segmentation is [SEG]." + ], + "image_footnote": [], + "bbox": [ + 94, + 261, + 330, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ca29c51bc8342b4e04f86ee755c6631376ef687c3e887d215817b7886cfc69a.jpg", + "image_caption": [ + "Figure 1. An example of user-AI conversation around an image. Left: The current state-of-the-art grounding model GLaMM [60] is effective for grounded conversation when prompted by \"answer with interleaved masks\", but fails to follow user instruction to answer a single word (yes or no) and misunderstands the question as a referring segmentation prompt. Right: Our F-LMM preserves instruction-following ability while being able to perform visual grounding." + ], + "image_footnote": [], + "bbox": [ + 339, + 258, + 687, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12eb4d80a541344329879bb511f1fd89872ab10b0b0c670f1b2d7b5c39d537d3.jpg", + "image_caption": [ + "User: Is there a truck in the image? Answer the question using a single word or phrase. F-LMM: No." + ], + "image_footnote": [], + "bbox": [ + 692, + 261, + 900, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 527, + 326, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Endowing Large Multimodal Models (LMMs) with visual grounding capability can significantly enhance AIs' understanding of the visual world and their interaction with humans. However, existing methods typically fine-tune the parameters of LMMs to learn additional segmentation tokens and overfit grounding and segmentation datasets. Such a design would inevitably cause a catastrophic diminution in the indispensable conversational capability of general AI assistants. In this paper, we comprehensively evaluate state-of-the-art grounding LMMs across a suite of multimodal question-answering benchmarks, observing drastic performance drops that indicate vanishing general knowledge comprehension and weakened instruction following ability. To address this issue, we present F-LMM—grounding frozen off-the-shelf LMMs in human-AI conversations—a straightforward yet effective design based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention mechanism of well-trained LMMs. Using only a few trainable CNN layers, we can translate word-pixel attention weights to mask logits, which a SAM-based mask refiner can further optimise. Our F-LMM neither learns special segmentation to", + "bbox": [ + 88, + 568, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kens nor utilises high-quality grounded instruction-tuning data, but achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks while completely preserving LMMs' original conversational ability. Additionally, with instruction-following ability preserved and grounding ability obtained, F-LMM can be directly applied to complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning. Our code can be found at https://github.com/wusize/F-LMM.", + "bbox": [ + 511, + 529, + 906, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 709, + 645, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs), which integrate Large Language Models (LLMs) with visual signals, have demonstrated remarkable success in multimodal understanding, reasoning and interaction [34, 38-40, 44, 65, 78]. To further advance LMMs with better perception capability, a recent line of research [28, 60, 61, 70, 80, 84] that visually grounds language contents in user-model conversations has drawn increasing attention. This explicit association between key phrases/words and visual objects greatly enhances LMMs' understanding of the visual world and allows for more intuitive and meaningful human-AI interactions.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24710", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ccb99f61929303a3bfe0feec5c8d55ed6f07a48f4eedf1504262f92de3cbe7b.jpg", + "image_caption": [ + "\"two girls\" \"a fence\"" + ], + "image_footnote": [], + "bbox": [ + 109, + 74, + 356, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3cf1e4093fa2e9ae3275d2a9fc469c196e35cda35230f4852ef30471d0dd714.jpg", + "image_caption": [ + "\"two elephants\" \"grass\" \"hills\" \"sky\"" + ], + "image_footnote": [], + "bbox": [ + 375, + 74, + 622, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ac40a0fd40619123168822b2cf63c72a0ad923dbcefa9bab400bd792aae79a0.jpg", + "image_caption": [ + "\"bread\" \"the plate\"" + ], + "image_footnote": [], + "bbox": [ + 640, + 74, + 887, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d618b2c3430986b89a4d9d51ae5a1caa9d714006f84a572b0f4cee7ad8f10634.jpg", + "image_caption": [ + "(a) Visualisations of word-image attention maps in Frozen LMMs via KMeans.", + "(b) Existing Grounding LMMs.", + "Figure 2. (a) Geometric and spatial cues conducive to visual grounding are observed in the visualisations of word-image attention maps in frozen LMMs. (b) Existing grounding LMMs are fine-tuned to generate a special mask token (e.g., [SEG]) for visual grounding purposes, which ruins the original conversational ability. (c) Our F-LMM translates word-image attention maps from frozen LMMs to grounding masks, while fully preserving the general-purpose chat capability." + ], + "image_footnote": [], + "bbox": [ + 107, + 180, + 470, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb98c5d08c498d3122b05491d35a0e3d19d347881dd2424529f021dc64b7dfc9.jpg", + "image_caption": [ + "(c) F-LMM (Ours)." + ], + "image_footnote": [], + "bbox": [ + 527, + 181, + 890, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By design, one commonly adopted build (Figure 2(b)) for visually grounding language contents is connecting LMMs with a mask head (e.g., Segment Anything Model (SAM) [26]), wherein both the LLM backbone and the mask head are fine-tuned with well-prepared visual grounding data that contains segmentation annotations. Also, some additional learnable tokens (e.g., [SEG]) are introduced to the LMMs' vocabulary, to directly associate key phrases or words with visual objects in conversations. However, this design will inevitably provoke a catastrophic diminution in general knowledge comprehension and instruction-following ability due to the following reasons. First, existing segmentation and visual grounding data only contain elementary patterns for answering simple grounding prompts. Second, during the fine-tuning stage, the LMMs are mainly optimised for effectively modelling the relationship between key phrases or words and special segmentation tokens, i.e., overfitting the segmentation and grounding data. Therefore, the conversational ability is sacrificed. For instance, the state-of-the-art grounding model GLaMM [60] fails to answer a simple yes-or-no question (Figure 1). Moreover, quantitative evaluations of existing grounding LMMs in conversational ability are presented in Table 1, with zero or near-zero scores on general multimodal question-answering benchmarks necessitating instruction-following ability.", + "bbox": [ + 91, + 417, + 483, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One possible option to deal with this dilemma is to collect high-quality training data encompassing both meaningful conversations and mask annotations. For example, LLaVA-G [80] annotates the 150k LLaVA-Instruct data samples [39] with segmentation masks so that the LMMs", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "simultaneously learn to chat and segment. Nonetheless, annotating high-quality grounded conversation data is costly and hard to scale. Despite being trained on costly annotated data, LLaVA-G still lags behind general-purpose LMMs on multimodal understanding tasks. Furthermore, training on large-scale annotated data normally consumes significant computational resources, which is, obviously not a resource-efficient solution.", + "bbox": [ + 511, + 417, + 906, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet effective design, i.e., grounding frozen LMMs (dubbed as F-LMM) in human-AI conversations. We argue that freezing the parameters of well-trained LMMs is the most practical design choice for fully preserving the original excellent conversational ability when building general-purpose grounding LMMs. In particular, we take inspiration from the built-in interpretability of the attention mechanism in transformers [6, 68] that represents interrelations between text and image contents in design. We observe that off-the-shelf LMMs already produce word-pixel correspondences necessary for visual grounding, despite they were not explicitly pre-trained with region or pixel annotations. As illustrated in Figure 2(a), we visualise word-image attention maps from frozen LMMs via K-Means clustering, revealing notable geometric and spatial cues of the objects1. For example, coarse visual grounding masks for key phrases (e.g., \"two girls\", \"two elephants\", and \"the plate\") in language sentences emerge from attention maps in LMMs. Therefore, our F-LMM takes these visual-language correspondences as useful segmentation priors for decoding ground", + "bbox": [ + 511, + 537, + 908, + 857 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1For better visibility, we perform K-Means clustering on the stack-up of all attention maps collected in a forward pass instead of selecting a single attention map.", + "bbox": [ + 511, + 862, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24711", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing masks, without further tuning the LMMs' weights or learning a special segmentation token to model object locations, as shown in Figure 2(c).", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The only trainable part of our F-LMM is a mask head plus a keyword selector. The mask head comprises a CNN-based mask decoder (a tiny U-Net [62]) that translates stacked attention maps to mask logits and a lightweight mask refiner (retrofitted from SAM [26]'s mask head) that uses additional image and language cues to refine the semantic-agnostic masks from the mask decoder. The keyword selector is a linear layer that discovers object nouns in text sequences, automating the process of grounded human-AI conversation. Moreover, we only use the RefCOCO $(+ / g)$ [24, 46] and PNG [16] datasets as our training data, enabling LMMs to segment user-described objects and ground key phrases or words in a text sequence. Unlike previous works [60, 61, 80], our F-LMM eliminates the necessity for high-quality conversation data that are annotated with masks to preserve conversational ability when learning grounding.", + "bbox": [ + 91, + 136, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our experiments demonstrate that F-LMM maintains the original excellence of off-the-shelf LMMs on general question-answering benchmarks, while achieving competitive results on referring segmentation and phrase grounding. In more complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning, F-LMM achieves better or comparable results when contrasted with models specially trained for such tasks. Compared with existing grounding LMMs, F-LMM offers the best balance between grounding and chat capabilities.", + "bbox": [ + 89, + 393, + 482, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 573, + 232, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Multimodal Models. Recent advancements in LMMs [2, 3, 13, 29-31, 34, 36, 38-40, 44, 47, 65, 75] have been fueled by the success of LLMs [1, 4, 12, 22, 48, 51, 66, 67, 81] since the debut of GPT series [1, 4, 56, 57] that feature an auto-regressive framework based on transformer decoders [68]. These LLMs possess general world knowledge and excellent conversational ability to follow human instructions, thanks to large-scale generative pre-training [4] and supervised finetuning on instruction-tuning data [71] or human feedback [50]. By integrating image representations from vision encoders [58, 79] to LLMs, LMMs enable visual understanding and reasoning in AI assistants. This integration is usually established by a multilayer perceptron (MLP) that directly maps image features to the LLMs' input embedding space [34, 38-40, 44, 65, 78] or a cross-attention module that abstracts the image contents with a set of query embeddings [2, 3, 31, 75]. In our research, we build F-LMM on LMMs of the former type (MLP-based), which preserves images' 2-D topological structure in the cross-modal integration.", + "bbox": [ + 91, + 599, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Segmentation. The task of predicting 2D masks for visual objects is known as image segmentation, which can be categorised into semantic segmentation [5, 8, 10, 87], instance segmentation [11, 19, 83] and panoptic segmentation [9, 25, 32, 73] depending on whether the goal is to differentiate pixel semantics or object instances. These standard segmentation approaches rely on a pre-defined set of object classes for recognition. In contrast, referring expression segmentation (RES) [24, 37, 45, 46, 49, 74, 88] involves segmenting objects based on free-form human language descriptions, allowing for enhanced human-model interaction. Additionally, panoptic narrative grounding (PNG) [14, 16, 17, 69] requires segmenting masks for key phrases or words in a sentence. In this study, we mainly leverage RES and PNG tasks to evaluate the grounding capability of LMMs. Besides, we also test LMMs' segmentation ability in complex scenarios that necessitate reasoning [28, 60, 63]. Moreover, the prompt-based SAM [26] pre-trained on billion-scale high-quality mask data has become a constituent component in many grounding LMMs to boost segmentation performance. We also adopt SAM's mask head to initialise our mask refiner.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Grounding Large Multimodal Models. Grounding Large Multimodal Models [3, 7, 28, 53, 54, 60, 61, 70, 72, 76, 78, 80, 84, 85] can localise language contents during user-model conversations. Some approaches [3, 7, 53, 76] represent coordinates of bounding boxes as texts and train LMMs to predict the coordinates in a generative manner. Several recent works [28, 54, 60, 61, 70, 80] train LMMs to predict a special segmentation token for encoding the grounded object and utilise a segmentation head (e.g., SAM [26]) to decode object masks. This study mainly focuses on grounding LMMs with segmentation ability for visual perception. To obtain competitive visual grounding performance, existing works extensively fine-tune the parameters of LMMs on a large amount of segmentation [5, 18, 25, 59, 87] and grounding [16, 23, 24, 27, 46, 55] datasets. And to balance the LMMs' grounding and conversational abilities, there are efforts [61, 80, 84] to collect high-quality instruction-tuning data annotated with segmentation masks. In contrast, we make the first attempt to build grounding LMMs on top of off-the-shelf LMMs without fine-tuning their parameters. Furthermore, we bypass the need for grounded instruction-tuning data to preserve decent chat ability.", + "bbox": [ + 511, + 422, + 903, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 768, + 604, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our F-LMM by first probing the causal attention mechanism in LMMs with visualisations of word-image attention maps in Sec 3.1. Then, we elaborate on F-LMM exploiting segmentation priors from frozen LMMs for visual grounding using the mask head in Sec 3.2. Finally, we show how to automate the process of grounded conversation with a linear keyword selector to in-", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24712", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bdd35d9e9487d868b1abd530b35df37f4c3c27f477c1e56b89ec8768f052fa50.jpg", + "image_caption": [ + "Figure 3. The overall pipeline of F-LMM. The word-image attention maps from the frozen LLM serve as segmentation priors for the mask head. The keyword selector discovers object nouns in the text sequence. The mask head encompasses a mask decoder that translates attention weights to mask logits and a mask refiner that optimises the mask decoder's predictions. $M$ and $N$ represent the numbers of transformer layers and attention heads." + ], + "image_footnote": [], + "bbox": [ + 109, + 92, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dace6dd8828545b4d0e50fc39618584f53839d77dc475e4854ea7a1a53622d83.jpg", + "image_caption": [ + "Figure 4. Visualisations of word-image attention maps. The letters $m$ and $n$ indicate that the attention map is derived from the $n$ -th attention head of the $m$ -th transformer layer." + ], + "image_footnote": [], + "bbox": [ + 101, + 358, + 473, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dicate words of grounding targets in Sec 3.3. The overall pipeline is illustrated in Figure 3.", + "bbox": [ + 89, + 592, + 482, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Segmentation Priors from Frozen LMM", + "text_level": 1, + "bbox": [ + 89, + 631, + 434, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Language Sequence. A typical build of a LMM $^2$ comprises an image encoder $f_{v}$ (e.g., CLIP [58] $^3$ ), a vision-language projector $f_{p}$ , and a LLM, denoted as $f_{\\mathrm{llm}}$ . The inputs to an LMM are usually an image $\\mathbf{X}_v \\in \\mathbb{R}^{3 \\times H \\times W}$ and the associated text $\\mathbf{X}_t$ . The input image is first encoded by the vision encoder $f_{v}$ and then mapped to the input space of the LLM by the projector $f_{p}$ :", + "bbox": [ + 89, + 652, + 482, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {v} = f _ {p} (\\text {F l a t t e n} (f _ {v} (\\mathbf {X} _ {v}))) \\in \\mathbb {R} ^ {h w \\times d},\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 768, + 424, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h$ and $w$ are the height and width of projected feature maps via $f_{v}$ . The Flatten operation unfolds the 2-D", + "bbox": [ + 89, + 799, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image feature map to a 1-D sequence. The constant $d$ is the hidden state dimension of the LLM. Likewise, the text input is first encoded as discrete tokens and then mapped to text embeddings:", + "bbox": [ + 511, + 357, + 906, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {t} = \\operatorname {E m b e d} (\\operatorname {T o k e n i z e} (\\mathbf {X} _ {t})) \\in \\mathbb {R} ^ {L \\times d},\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 430, + 849, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L$ denotes the length of text embeddings. The visual-language sequence input to the LLM is a concatenation of image and text embeddings: $\\mathbf{Z} = \\{\\mathbf{Z}_v,\\mathbf{Z}_t\\} \\in \\mathbb{R}^{(hw + L)\\times d}$", + "bbox": [ + 511, + 460, + 906, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Segmentation Priors in Self-Attention. The vision-language sequence is mainly processed by causal self-attentions [56, 68] in the LLM, including inner product and weighted-sum operations. Specifically, for a word token with position index $i$ in the vision-language sequence $\\mathbf{Z}$ , its embedding $z^i$ is updated by the weighted sum of the first $i$ embeddings: $\\hat{z}^i = \\mathrm{SoftMax}\\left(\\frac{z^i \\cdot \\mathbf{Z}[i]}{d}\\right) \\cdot \\mathbf{Z}[i]$ , where $\\mathrm{SoftMax}\\left(\\frac{z^i \\cdot \\mathbf{Z}[i]}{d}\\right)$ is the attention weights. Here, we omit the residual layers and feedforward layers for brevity.", + "bbox": [ + 511, + 507, + 905, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the word-image interaction in the multimodal scenario, we can select the word token's attention weights with the image embeddings from the overall vision-language attention weights:", + "bbox": [ + 511, + 650, + 905, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} ^ {i} = \\text {U n f l a t t e n} (\\operatorname {S o f t M a x} (\\frac {\\boldsymbol {z} ^ {i} \\cdot \\mathbf {Z} [: i ]}{d}) [: h w ]) \\in \\mathbb {R} ^ {h \\times w},\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 723, + 906, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Un flatten restores the 2-D spatial structure from the 1-D sequence to form an attention map. In Figure 4, we visualise such word-image attention maps from various transformers layers and attention heads in an LMM (i.e., DeepSeekVL-7B [44]). The objects' shape and location can be observed in word-image attention maps of certain layers or heads. The visibility is further enhanced when we stack the attention maps from all layers and heads and perform K-Means clustering. It can be observed that the attention", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2In this paper, the term 'multimodal' stands for vision and language modalities.", + "bbox": [ + 89, + 837, + 482, + 862 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3The image encoder might be any vision model that is pre-trained on image-text pairs. We use the classic term 'CLIP' in this paper to represent all such models for brevity.", + "bbox": [ + 89, + 862, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24713", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/5c1eb9ea76ab325d12f813e6d4e8be6107e39960f57cc897c056858a644d76b4.jpg", + "table_caption": [ + "Table 1. The main evaluation results on question-answering benchmarks, referring expression segmentation (RES) benchmark and panoptic narrative grounding (PNG) benchmark. LLaVAW: LLaVA-In-the-Wild. LLaVA-1.6 and MGM-HD take high-resolution image inputs. LLaVA-1.6-M-7B means the model is based on Mistral-7B [22]. GLaMM-FS-7B means we use the 'FullScope' version of GLaMM." + ], + "table_footnote": [], + "table_body": "
ModelMultimodal Question AnsweringRESPNG
MMEMMBenchMMVetLLaVAWRefCOCORefCOCO+RefCOCOgAllThingStuff
PixelLM-7B [61]309/13517.415.946.473.066.369.343.141.047.9
LISA-7B [28]1/10.419.147.574.965.167.9---
PerceptionGPT-7B [54]----75.168.570.3---
LLaVA-G-7B [80]---55.877.168.871.5---
GroundHog-7B [84]----78.570.574.166.865.069.4
GLaMM-FS-7B [60]14/936.810.332.078.670.574.855.852.962.3
GSVA-7B [72]446/1817.819.438.377.768.273.241.839.646.6
LaSagnA-7B [70]0/00.016.734.576.866.470.6---
F-LMM (DeepSeekVL-1.3B [44])1307/22564.634.851.175.062.868.264.963.468.3
F-LMM (MGM-2B [34])1341/31259.831.165.975.063.767.365.664.468.4
F-LMM (LLaVA-1.5-7B [38])1511/34864.330.569.075.263.767.164.863.468.2
F-LMM (HPT-Air-6B [65])1010/ 25869.831.359.274.364.067.565.564.068.8
F-LMM (HPT-Air-1.5-8B [65])1476/30875.236.362.176.364.568.565.464.168.5
F-LMM (MGM-7B [34])1523/31669.340.875.875.764.868.366.365.368.6
F-LMM (DeepSeekVL-7B [44])1468/29873.241.577.876.166.470.165.764.568.5
F-LMM (LLaVA-1.6-7B [40])1519/32268.144.172.375.865.870.166.365.169.0
F-LMM (LLaVA-1.6-M-7B [40])1501/32469.547.871.775.766.570.166.565.469.1
F-LMM (MGM-HD-7B [34])1546/31965.841.374.076.165.268.566.765.669.1
", + "bbox": [ + 91, + 142, + 910, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "maps offer meaningful segmentation priors with spatial and geometric cues for grounding objects visually.", + "bbox": [ + 89, + 420, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Language Cues. In addition to the spatial and geometric cues from word-image attention maps, F-LMM can also capitalise on the object's corresponding text embeddings from the LLM $f_{\\mathrm{llm}}$ , which provide extra language cues for the grounding of visual objects.", + "bbox": [ + 89, + 452, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Visual Grounding with Mask Head", + "text_level": 1, + "bbox": [ + 89, + 542, + 395, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the segmentation priors from the frozen LMM for pixel-level grounding, with the help of a mask head consisting of a mask decoder and a mask refiner.", + "bbox": [ + 89, + 566, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask Decoder. The mask decoder $f_{d}$ is a 2-D CNN model that transforms the word-image attention maps of grounded objects into mask logits, which is instantiated by a 3-stage U-Net [62]. Please refer to the supplemental material for details on the mask decoder. The extraction of word-image attention map $\\pmb{a}^{i}$ for a word token with position index $i$ is illustrated in Eq. 3.1 and Figure 3. For an object described by multiple words, we merge their corresponding word-image attention maps to a single attention map $\\pmb{a}$ via element-wise average or max operation. The attention map $\\pmb{a}$ is further normalised as $\\pmb{a} / \\mathrm{sum}(\\pmb{a})$ so that all elements sum to 1. Considering $M$ layers and $N$ attention heads, we stack the $MN$ attention maps as $\\pmb{A} \\in \\mathbb{R}^{MN \\times h \\times w}$ , which forms the input to a mask decoder. Given the importance of high input resolution for segmentation models, we upsample the stacked attention maps $\\pmb{A}$ to $h' \\times w'$ by bilinear interpolation before feeding it to a mask decoder, where $h' > h$ and $w' > w$ . In practice, we set $h' = w' = 64$ . Then, the mask decoder maps $\\pmb{A}$ into mask", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "logits: $\\mathbf{M}_{\\mathrm{logits}} = f_d(\\mathbf{A})$ . We derive the corresponding binary mask via $\\mathbf{M}_{\\mathrm{pred}} = \\mathbf{M}_{\\mathrm{logits}} > 0$ . During training, the mask decoder is optimised with BCE and DICE losses [64].", + "bbox": [ + 511, + 420, + 906, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask Refiner. The mask refiner $f_{r}$ is retrofitted from the mask head of SAM [26], which predicts masks based on prompts as well as image embeddings from SAM's ViT-based image encoder. To refine the output of the mask decoder $f_{d}$ , we re-use SAM's prompt encoder to transform $\\mathbf{M}_{\\mathrm{logits}}$ into dense prompt embeddings (i.e., a 2-D feature map) and the bounding box of $\\mathbf{M}_{\\mathrm{pred}}$ to box embeddings. In addition to the location cues from the mask and the box, the language cues, i.e., the object's corresponding text embeddings, are also utilised by $f_{r}$ . Considering text embeddings from the $M$ transformer layers, we train $M$ learnable scalars to calculate a weighted sum of these text embeddings. The weighted-summed text embeddings are processed by a linear layer and then concatenated with the box embeddings to form sparse prompt embeddings. The dense and sparse prompt embeddings, together with SAM's image embeddings, are fed to the mask refiner $f_{r}$ for finer-grained mask predictions $\\mathbf{M}_{\\mathrm{pred}}^{\\prime}$ . During training, we keep the ViT-based image encoder of SAM frozen and optimise the mask refiner $f_{r}$ using BCE loss and DICE loss [64]. For more details on the SAM's prompt-based mask head, please refer to the original SAM paper [26].", + "bbox": [ + 511, + 468, + 908, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Keyword Selector for Grounded Conversation", + "text_level": 1, + "bbox": [ + 511, + 816, + 903, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In grounded conversation [60] with interleaved segmentation masks and words, existing grounding LMMs [60, 61] expand LLMs' vocabularies with special tokens that indicate the start and end of grounding targets, which is infeas-", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24714", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ef6829c46d797781aec4481e13d2c870c390d736b2de8c01d1536c0def5404ee.jpg", + "table_caption": [ + "Table 2. Reasoning Segmentation." + ], + "table_footnote": [], + "table_body": "
ModelVal -ShortTest LongAll
X-Decoder [88]22.620.422.221.7
SEEM [89]25.520.125.624.3
GroundingSAM [41]26.017.822.421.3
OVSeg [35]28.518.028.726.1
LISA [28]44.437.636.636.8
F-LMM46.736.949.146.2
", + "bbox": [ + 122, + 107, + 416, + 212 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/365061a1da6fe95af870719a3f5099826e7aedec19d93abce4c4010f6474cc66.jpg", + "table_caption": [ + "Table 3. Grounded Conversation Generation (GCG). M. stands for METEOR." + ], + "table_footnote": [], + "table_body": "
ModelGCG TrainingM.Val mIoURecallM.Test mIoURecall
LISA [28]✓13.062.036.312.961.735.5
OMG-LLaVA [82]✓14.965.5-14.564.7-
GLaMM [60]✓16.266.341.815.865.640.8
BuboGPT [86]✗17.254.029.417.154.127.0
KOSMOS-2 [52]✗16.155.628.315.856.829.0
F-LMM✗17.663.542.017.463.638.6
", + "bbox": [ + 431, + 107, + 875, + 212 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2dfde106f92c4d7e31e5eb409930beeda159ade71bc3e8742b68768ba92c315.jpg", + "table_caption": [ + "Table 4. Unleashing visual chain-of-thought reasoning with both excellent grounding and instruction-following ability." + ], + "table_footnote": [], + "table_body": "
ModelVisual CoTDocVQATextCapsVisCoT BenchmarkPOPE
TextVQADUDESROIEInfographicsAccF1
VisCoT-7B [63]✓47.667.577.538.647.032.486.5-
F-LMM✗43.263.574.532.028.443.287.086.0
F-LMM✓53.867.978.442.344.149.188.087.7
", + "bbox": [ + 156, + 239, + 834, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ble in F-LMM given its ‘frozen’ nature. A common practice to discover visual objects in text sequences offline is using external tools such as SpaCy [20], which parse all nouns from a sentence including unwanted non-object words. Instead of adopting such offline tools that produce noisy results, we automate generating interleaved words and masks by training a linear layer to directly predict whether a word is to be grounded or not.", + "bbox": [ + 88, + 340, + 480, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, the linear layer (keyword selector) is placed on top of the LLM's transformer layers, projecting $d$ -dimension hidden state vectors into one-dimension scores, followed by a sigmoid function that normalizes the scores to $[0,1]$ . During training, the score prediction is supervised by a BCE loss. During inference, word tokens with scores exceeding a threshold $\\lambda$ are regarded as positive for visual grounding. In practice, we set $\\lambda = 0.3$ . Adjacent positive word tokens are grouped to indicate a single visual object. After the word tokens of visual objects are selected, the corresponding attention maps and text embeddings are fed to the mask head for visual grounding.", + "bbox": [ + 89, + 462, + 482, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 657, + 223, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 681, + 307, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Architectures. We build F-LMM on several open-sourced LMMs, including LLaVA-1.5 [38], LLaVA-Next [40], MiniGemini [34], DeepSeekVL [44] and HPT-Air [65]. The main experiment covers 10 LMMs with model sizes ranging from 1.3B to 8B. We employ a lightweight 3-stage U-Net [62] as the mask decoder to transform segmentation priors from frozen LLMs. The U-Net architecture features an encoder-decoder structure with skip connections, wherein feature maps are downsampled in the encoder and upsampled in the decoder. Please check the supplementary material for more details on the mask decoder. As for the SAM-based mask refiner, we choose SAM ViT-L [26] that balances cost and performance well.", + "bbox": [ + 89, + 703, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Training. We train F-LMM on RefCOCO(+/g) [24, 46] and PNG [16] datasets with about 190k data samples on a single machine with 8 NVIDIA A800-40G GPUs, which costs about 20 hours for each round of model training. We set the batch size to 8 and train models for 8 epochs, with gradient clipping at a max norm of 1.0. The AdamW [43] optimiser is used with a learning rate of 1e-4, a weight decay of 0.01, and betas as (0.9, 0.999). We choose a warm-up ratio of 0.03 to stabilise model optimisation.", + "bbox": [ + 511, + 340, + 906, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Standard QA and Grounding Tasks", + "text_level": 1, + "bbox": [ + 511, + 486, + 823, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a comprehensive study of LMMs' conversational and grounding capabilities, we first evaluate models under standard question-answering and grounding benchmarks separately. We summarise the evaluation results of grounding LMMs in Table 1. Please refer to the supplementary material for more detailed results.", + "bbox": [ + 511, + 508, + 905, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarks. For comprehensive conversational ability evaluation, we choose four widely used general question-answering benchmarks including MME [15], MMBench [42], LLaVA-In-the-Wild [39] and MMVet [77]. The MME and MMBench require an LMM to strictly follow the instruction to reply with single words (yes or no) or answer MCQs with alphabetical letters (i.e., answering A, B, C, or D). The LLaVA-In-the-Wild and MMVet benchmarks ask a model to respond with open-ended texts while demanding general world knowledge comprehension. In terms of grounding ability evaluation, we assess the LMMs' ability to segment user-described objects on referring expression segmentation (RES) [24, 46] benchmarks including RefCOCO, RefCOCO+, and RefCOCOg, using the cIoU metric. Due to limited space, we only report results on the Val splits of RefCOCO(+/g) in Table 1. We also test the LMMs' ability to ground key phrases or words in user-model conversations on the Panoptic Narrative Grounding (PNG) [16] benchmark, measuring individual mask recalls on thing/stuff objects and overall recall scores.", + "bbox": [ + 511, + 599, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24715", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0210f968d82afa7d617a0ba78a1b43c3eb27779d9453f770ab632a437e6458b0.jpg", + "image_caption": [ + "Figure 5. Ablation study of the mask decoder." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 480, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with Existing Methods. We compare F-LMM with existing grounding LMMs. As shown in Table 1, our F-LMM provides the best balance with conversational and grounding abilities among compared methods. On the question-answering benchmarks, existing grounding LMMs obtain zero or near-zero scores on MMBench and MME while lagging significantly behind general-purpose LMMs on MM Vet and LLaVA-In-the-Wild benchmarks, indicating compromised instruction-following ability and weakened general knowledge comprehension. On the RES and PNG benchmarks, our F-LMM achieves comparable results despite not having the parameters of LMMs fine-tuned for grounding purposes.", + "bbox": [ + 89, + 325, + 483, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Complex Scenarios", + "text_level": 1, + "bbox": [ + 89, + 531, + 274, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate grounding LMMs under more complex scenarios that typically require the LMMs to perform both reasoning and segmentation. The base LMM we use is DeepSeekVL-7B [44], considering its flexibility in supporting both single and multiple image inputs. And the size of models compared in this section is also 7B.", + "bbox": [ + 89, + 551, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reasoning Segmentation is proposed in LISA [28] that requires a model to infer what object to segment from common-sense knowledge or via logical reasoning. The evaluation results are provided in Table 2 and the metric is gIoU. F-LMM can effectively perform reasoning segmentation even though it is not trained on such type of data. It is remarkable that F-LMM significantly outperforms existing models on the subset of long sentences, reflecting the advantage of F-LMM in handling complex contexts.", + "bbox": [ + 89, + 643, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Grounded Conversation Generation (GCG) is proposed in GLaMM [60], which requires a model to generate interleaved segmentation masks and texts. For performance evaluation, METEOR (M.) and mIoU are used to measure the quality of generated texts and masks, respectively. In addition, we report the recall of object masks. As shown in Table 3, F-LMM exhibits the best zero-shot performance while being comparable with models fine-tuned on GCG", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/19fde3c27477195ccb4cd6ee35d4fcc89fb22f0a3127bc086e84bfa422adff41.jpg", + "table_caption": [ + "Table 5. Ablation study of the mask refiner on PNG benchmark." + ], + "table_footnote": [], + "table_body": "
(a) Effects of Mask Refiner(b) SAM Variants
Mask RefinerPNGSAMPNG
maskboxtextAllThingStuffAllThingStuff
XXX50.848.655.9ViT-B63.061.466.8
✓XX63.462.066.8ViT-L64.963.468.3
✓✓X63.762.267.1ViT-H65.063.568.3
✓✓✓64.963.468.3
", + "bbox": [ + 521, + 104, + 870, + 184 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/54f491662994f1b30a6d73b7eea9abb6011b4800fcf4e03dfc764a68f9e49e4b.jpg", + "table_caption": [ + "Table 6. Ablation study of keyword selection." + ], + "table_footnote": [], + "table_body": "
MethodF1RecallPrecision
SpaCy Parser [20]57.897.341.1
Linear Keyword Selector82.896.672.5
", + "bbox": [ + 563, + 209, + 848, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training dataset.", + "bbox": [ + 511, + 273, + 622, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Chain-of-Thought Reasoning. In human-AI conversations that involve Visual Chain-of-Thought Reasoning (VisCoT) [63], an LMM first localises the region/object relevant to the human's question and then generates the final answer by zooming in on the question-related region. Here we evaluate F-LMM and VisCoT-7B on the VisCoT benchmark [63]. As shown in Table 4, F-LMM achieves remarkable performance gains when prompted in the VisCoT manner. It is noticeable that F-LMM even outperforms VisCoT-7B [63] that has been well-tuned on the VisCoT training data [63]. Furthermore, we perform VisCoT on the object hallucination benchmark POPE [33] and observe significant performance gain in resisting object hallucinations.", + "bbox": [ + 511, + 289, + 906, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 498, + 666, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate the effects of design choices of F-LMM. All the ablation studies are conducted on the PNG dataset and we use DeepSeekVL-1.3B [44] unless otherwise stated.", + "bbox": [ + 511, + 521, + 905, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask Decoder. We summarise our analyses of the mask decoder in Figure 5. Note that the mask refiner is not involved in this part. We first consider attention maps from different transformer layers, i.e., early $(6^{\\mathrm{th}})$ , mid $(12^{\\mathrm{th}})$ and late $(24^{\\mathrm{th}})$ layers. As shown in Figure 5(a), attention maps from late layers perform the worst, conforming to prior studies [21] indicating that deeper transformer layers tend to focus on abstract concepts instead of visual details. And using attention maps from all layers achieves the best performance. Next, we study how to merge the word-image attention maps of multi-word objects, as shown in Figure 5(b). We find that the average operation outperforms the max operation by a margin of 0.7. In Figure 5(c), we show that normalizing the inputs to the mask decoder provides a 0.5 performance gain. Finally, we experiment with different input sizes for the mask decoder. As shown in Figure 5(d), using $64 \\times 64$ yields the best performance-cost trade-off.", + "bbox": [ + 511, + 566, + 905, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask Refiner. We study the effects of the mask refiner on segmentation in Table 5a. The performance of using only the mask decoder is shown in the first row of Table 5a. With the masks fed to the mask refiner, we observe a significant 12.6 performance gain on the PNG benchmark. Adding", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24716", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4be90301f9dd0c9dcba24b227771950867c268898c016fa13265a3b49fbf376c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelChat ↓Ground ↓
DeepseekVL-1.3B [44]7.758.33
MGM-2B [34]6.008.33
LLaVA-1.5-7B [38]6.757.83
HPT-Air-6B [65]9.007.16
HPT-Air-1.5-8B [65]6.507.00
MGM-7B [34]5.754.83
DeepseekVL-7B [44]3.754.00
LLaVA-1.6-7B [40]2.753.00
LLaVA-1.6-M-7B [40]3.251.66
MGM-HD-7B [34]3.502.83
", + "bbox": [ + 125, + 102, + 424, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53cbdeaf3b34fa1ac6590ea6b3a3019e18fad366495ab50bbf568e2261c97a23.jpg", + "image_caption": [ + "Figure 6. The relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. The left table shows the average ranks of each LMM on question-answering ('Chat') and grounding benchmarks ('Ground'). $\\downarrow$ means the lower the better. The dashed line in the right figure is the linear fit of the rank data points, indicating a positive correlation between abilities to chat and ground." + ], + "image_footnote": [], + "bbox": [ + 468, + 98, + 869, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "box and text prompts to the mask refiner further improves the performance by 0.3 and 1.2. Then we experiment with different SAM [26] model variants, i.e., ViT-B(ase), ViT-L(arge) and ViT-H(uge) as mask refiners. As shown in Table 5b, the performance grows with model sizes. For a good trade-off between cost and performance, we select the ViT-L variant of SAM as the default mask refiner.", + "bbox": [ + 88, + 342, + 480, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Keyword Selector. We analyse the keyword selector, implemented as a linear layer, on the PNG dataset, using the F1 score as the main metric. As shown in Table 6, our keyword selector achieves significantly higher F1 scores than the external SpaCy tool [20]. We also report recall and precision scores. Our keyword selector achieves comparable recall while being much more precise compared to SpaCy, which enumerates all nouns in a sentence.", + "bbox": [ + 88, + 449, + 482, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis & Visualisation", + "text_level": 1, + "bbox": [ + 89, + 575, + 316, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scalability: Does Better Chatting Lead to Better Grounding? We study the relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. Specifically, we examine the correlation between performance on the question-answering and grounding benchmarks. For the ten models reported in Table 1, we calculate their average ranks in each benchmark category. In Figure 6, we plot these ranks as 2D coordinates, i.e., (Chat Rank, Ground Rank), and apply a linear fit to the data points. As indicated by the blue dashed line, frozen LMMs with stronger conversational ability can serve as better backbones for grounding. We also observe that larger LMMs generally excel in both conversation and grounding tasks, and LMMs with larger input resolutions (e.g., LLaVA-1.6 and MGM-HD) can handle both tasks better.", + "bbox": [ + 91, + 598, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From Attention Maps to Segmentation Masks. We visualise the word-image attention maps by applying KMeans clustering to the stacked attention maps that are collected from all transformer layers and attention heads. The attention maps of multiple-word objects are merged by element", + "bbox": [ + 88, + 825, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd312f2ccd62ce4444726e5e51a247ba653854aa36e1ba77d139b90c5b90d27f.jpg", + "image_caption": [ + "Caption: There is a person in this picture on the surfing board who is surfing on a ocean tide. There is sky in the background." + ], + "image_footnote": [], + "bbox": [ + 526, + 340, + 645, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a60d565c64fcb6f3b191f7a3b4df50b0b63e4782a1aafab0d240b78bd8a56a90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 340, + 769, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98cce25cf2d0904342af9c95d8e4fe9dbe72cb8d26a79d1576ae6bf388fd900f.jpg", + "image_caption": [ + "Figure 7. Visualisations of KMeans clustering on attention maps and segmentation results from the CNN-based mask decoder and the SAM-based mask refiner." + ], + "image_footnote": [], + "bbox": [ + 650, + 407, + 769, + 470 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81a5618bbb37dfc7e59121dee86850d9f6d6919423138f242064951398aa79e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 340, + 890, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/637f4819a3f7a8728f55e5ea02b3da3d80d0c18e6329df10340492ece7d3dd1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 407, + 890, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "wise average. As shown in Figure 7, we observe that the pixels of objects are roughly clustered together (top-left). With the CNN-Based mask decoder, the attention weights are mapped to 2D binary masks (top-right), which are then further optimised by the SAM-based mask refiner (bottom-left). The model used is DeepSeekVL-1.3B [44].", + "bbox": [ + 511, + 527, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 631, + 633, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have studied the limitation of existing grounding LMMs, i.e., the loss of general world knowledge and instruction-following ability. To address this issue, we make the first attempt to ground fully frozen LMMs, which are already well-trained for user-model conversation, based on the insight that the geometric and spatial cues needed for visual grounding are inherently present within the self-attention mechanism of LMMs. By incorporating a CNN-based mask decoder and a SAM-based mask refiner, we achieve competitive visual grounding performance without sacrificing any conversational abilities of pre-trained LMMs. With this combination of strong conversational and visual grounding capabilities, these LMMs show promise for complex perception and reasoning tasks, such as segmentation with reasoning, grounded conversation generation, and visual chain-of-thought reasoning.", + "bbox": [ + 511, + 656, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24717", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-PhD-2023-08-048T), the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 226 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022.", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020.", + "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In $CVPR$ , pages 1209-1218, 2018.", + "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021.", + "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023.", + "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017.", + "[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020.", + "[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems, 34:17864-17875, 2021." + ], + "bbox": [ + 93, + 236, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022.", + "[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023.", + "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024.", + "[14] Zihan Ding, Zi-han Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Xiaolin Wei, and Si Liu. Ppmn: Pixel-phrase matching network for one-stage panoptic narrative grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5537–5546, 2022.", + "[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023.", + "[16] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021.", + "[17] Tianyu Guo, Haowei Wang, Yiwei Ma, Jiayi Ji, and Xiaoshuai Sun. Improving panoptic narrative grounding by harnessing semantic relationships and visual confirmation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1985-1993, 2024.", + "[18] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022.", + "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017.", + "[20] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017.", + "[21] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024.", + "[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lam-" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24718", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ple, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023.", + "[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014.", + "[24] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014.", + "[25] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. CVPR, 2019.", + "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023.", + "[27] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017.", + "[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023.", + "[29] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024.", + "[30] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023.", + "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023.", + "[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 214-223, 2021.", + "[33] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023.", + "[34] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023.", + "[35] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Con-" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023.", + "[36] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533, 2023.", + "[37] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023.", + "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023.", + "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023.", + "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.", + "[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.", + "[42] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024.", + "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019.", + "[44] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024.", + "[45] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020.", + "[46] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016.", + "[47] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611, 2024.", + "[48] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024.", + "[49] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24719", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016.", + "[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022.", + "[51] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023.", + "[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023.", + "[53] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023.", + "[54] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023.", + "[55] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015.", + "[56] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018.", + "[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.", + "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021.", + "[59] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023.", + "[60] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.", + "[61] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228, 2023.", + "[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015.", + "[63] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024.", + "[64] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240–248. Springer, 2017.", + "[65] HyperGAI Team. Hpt 1.5 air: Best open-sourced 8b multimodal llm with llama 3, 2024.", + "[66] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.", + "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.", + "[69] Haowei Wang, Jiayi Ji, Yiyi Zhou, Yongjian Wu, and Xiaoshuai Sun. Towards real-time panoptic narrative grounding by an end-to-end grounding network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2528-2536, 2023.", + "[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024.", + "[71] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021.", + "[72] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024.", + "[73] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019.", + "[74] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24720", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023.", + "[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023.", + "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024.", + "[78] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023.", + "[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023.", + "[80] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, and Jianwei Yang. Llava-grounding: Grounded visual chat with large multimodal models, 2023.", + "[81] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.", + "[82] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024.", + "[83] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021.", + "[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. arXiv preprint arXiv:2402.16846, 2024.", + "[85] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2024.", + "[86] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023.", + "[87] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023.", + "[89] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 232 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24721", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_model.json b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..60bbcb57f4212eee08803e9fb4b995752df3dd7a --- /dev/null +++ b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_model.json @@ -0,0 +1,2622 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.131, + 0.778, + 0.152 + ], + "angle": 0, + "content": "F-LMM: Grounding Frozen Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.18, + 0.916, + 0.253 + ], + "angle": 0, + "content": "Size \\(\\mathbf{W}\\mathbf{u}^{1}\\) Sheng Jin2 Wenwei Zhang3 Lumin \\(\\mathrm{Xu^4}\\) Wentao Liu2,3 Wei Li1 Chen Change Loy' 1 S-Lab, Nanyang Technological University SenseTime Research and Tetras.AI 3 Shanghai AI Laboratory 4 The Chinese University of Hong Kong size001@e.ntu.edu.sg ccloy@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.262, + 0.331, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.391, + 0.332, + 0.428 + ], + "angle": 0, + "content": "User: Is there a truck in the image? Answer the question using a single word or phrase. \nGLaMM: Sure, the segmentation is [SEG]." + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.259, + 0.688, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.262, + 0.901, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.38, + 0.893, + 0.427 + ], + "angle": 0, + "content": "User: Is there a truck in the image? Answer the question using a single word or phrase. F-LMM: No." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.447, + 0.908, + 0.504 + ], + "angle": 0, + "content": "Figure 1. An example of user-AI conversation around an image. Left: The current state-of-the-art grounding model GLaMM [60] is effective for grounded conversation when prompted by \"answer with interleaved masks\", but fails to follow user instruction to answer a single word (yes or no) and misunderstands the question as a referring segmentation prompt. Right: Our F-LMM preserves instruction-following ability while being able to perform visual grounding." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.528, + 0.327, + 0.543 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.569, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Endowing Large Multimodal Models (LMMs) with visual grounding capability can significantly enhance AIs' understanding of the visual world and their interaction with humans. However, existing methods typically fine-tune the parameters of LMMs to learn additional segmentation tokens and overfit grounding and segmentation datasets. Such a design would inevitably cause a catastrophic diminution in the indispensable conversational capability of general AI assistants. In this paper, we comprehensively evaluate state-of-the-art grounding LMMs across a suite of multimodal question-answering benchmarks, observing drastic performance drops that indicate vanishing general knowledge comprehension and weakened instruction following ability. To address this issue, we present F-LMM—grounding frozen off-the-shelf LMMs in human-AI conversations—a straightforward yet effective design based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention mechanism of well-trained LMMs. Using only a few trainable CNN layers, we can translate word-pixel attention weights to mask logits, which a SAM-based mask refiner can further optimise. Our F-LMM neither learns special segmentation to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.682 + ], + "angle": 0, + "content": "kens nor utilises high-quality grounded instruction-tuning data, but achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks while completely preserving LMMs' original conversational ability. Additionally, with instruction-following ability preserved and grounding ability obtained, F-LMM can be directly applied to complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning. Our code can be found at https://github.com/wusize/F-LMM." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.646, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs), which integrate Large Language Models (LLMs) with visual signals, have demonstrated remarkable success in multimodal understanding, reasoning and interaction [34, 38-40, 44, 65, 78]. To further advance LMMs with better perception capability, a recent line of research [28, 60, 61, 70, 80, 84] that visually grounds language contents in user-model conversations has drawn increasing attention. This explicit association between key phrases/words and visual objects greatly enhances LMMs' understanding of the visual world and allows for more intuitive and meaningful human-AI interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24710" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.075, + 0.357, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.155, + 0.292, + 0.167 + ], + "angle": 0, + "content": "\"two girls\" \"a fence\"" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.075, + 0.623, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.155, + 0.601, + 0.168 + ], + "angle": 0, + "content": "\"two elephants\" \"grass\" \"hills\" \"sky\"" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.075, + 0.888, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.155, + 0.818, + 0.169 + ], + "angle": 0, + "content": "\"bread\" \"the plate\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.168, + 0.696, + 0.18 + ], + "angle": 0, + "content": "(a) Visualisations of word-image attention maps in Frozen LMMs via KMeans." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.181, + 0.471, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.316, + 0.376, + 0.328 + ], + "angle": 0, + "content": "(b) Existing Grounding LMMs." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.182, + 0.891, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.315, + 0.766, + 0.328 + ], + "angle": 0, + "content": "(c) F-LMM (Ours)." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.343, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Figure 2. (a) Geometric and spatial cues conducive to visual grounding are observed in the visualisations of word-image attention maps in frozen LMMs. (b) Existing grounding LMMs are fine-tuned to generate a special mask token (e.g., [SEG]) for visual grounding purposes, which ruins the original conversational ability. (c) Our F-LMM translates word-image attention maps from frozen LMMs to grounding masks, while fully preserving the general-purpose chat capability." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.418, + 0.485, + 0.812 + ], + "angle": 0, + "content": "By design, one commonly adopted build (Figure 2(b)) for visually grounding language contents is connecting LMMs with a mask head (e.g., Segment Anything Model (SAM) [26]), wherein both the LLM backbone and the mask head are fine-tuned with well-prepared visual grounding data that contains segmentation annotations. Also, some additional learnable tokens (e.g., [SEG]) are introduced to the LMMs' vocabulary, to directly associate key phrases or words with visual objects in conversations. However, this design will inevitably provoke a catastrophic diminution in general knowledge comprehension and instruction-following ability due to the following reasons. First, existing segmentation and visual grounding data only contain elementary patterns for answering simple grounding prompts. Second, during the fine-tuning stage, the LMMs are mainly optimised for effectively modelling the relationship between key phrases or words and special segmentation tokens, i.e., overfitting the segmentation and grounding data. Therefore, the conversational ability is sacrificed. For instance, the state-of-the-art grounding model GLaMM [60] fails to answer a simple yes-or-no question (Figure 1). Moreover, quantitative evaluations of existing grounding LMMs in conversational ability are presented in Table 1, with zero or near-zero scores on general multimodal question-answering benchmarks necessitating instruction-following ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.903 + ], + "angle": 0, + "content": "One possible option to deal with this dilemma is to collect high-quality training data encompassing both meaningful conversations and mask annotations. For example, LLaVA-G [80] annotates the 150k LLaVA-Instruct data samples [39] with segmentation masks so that the LMMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.538 + ], + "angle": 0, + "content": "simultaneously learn to chat and segment. Nonetheless, annotating high-quality grounded conversation data is costly and hard to scale. Despite being trained on costly annotated data, LLaVA-G still lags behind general-purpose LMMs on multimodal understanding tasks. Furthermore, training on large-scale annotated data normally consumes significant computational resources, which is, obviously not a resource-efficient solution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.858 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet effective design, i.e., grounding frozen LMMs (dubbed as F-LMM) in human-AI conversations. We argue that freezing the parameters of well-trained LMMs is the most practical design choice for fully preserving the original excellent conversational ability when building general-purpose grounding LMMs. In particular, we take inspiration from the built-in interpretability of the attention mechanism in transformers [6, 68] that represents interrelations between text and image contents in design. We observe that off-the-shelf LMMs already produce word-pixel correspondences necessary for visual grounding, despite they were not explicitly pre-trained with region or pixel annotations. As illustrated in Figure 2(a), we visualise word-image attention maps from frozen LMMs via K-Means clustering, revealing notable geometric and spatial cues of the objects1. For example, coarse visual grounding masks for key phrases (e.g., \"two girls\", \"two elephants\", and \"the plate\") in language sentences emerge from attention maps in LMMs. Therefore, our F-LMM takes these visual-language correspondences as useful segmentation priors for decoding ground" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.909, + 0.903 + ], + "angle": 0, + "content": "1For better visibility, we perform K-Means clustering on the stack-up of all attention maps collected in a forward pass instead of selecting a single attention map." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24711" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "ing masks, without further tuning the LMMs' weights or learning a special segmentation token to model object locations, as shown in Figure 2(c)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.137, + 0.483, + 0.394 + ], + "angle": 0, + "content": "The only trainable part of our F-LMM is a mask head plus a keyword selector. The mask head comprises a CNN-based mask decoder (a tiny U-Net [62]) that translates stacked attention maps to mask logits and a lightweight mask refiner (retrofitted from SAM [26]'s mask head) that uses additional image and language cues to refine the semantic-agnostic masks from the mask decoder. The keyword selector is a linear layer that discovers object nouns in text sequences, automating the process of grounded human-AI conversation. Moreover, we only use the RefCOCO \\((+ / g)\\) [24, 46] and PNG [16] datasets as our training data, enabling LMMs to segment user-described objects and ground key phrases or words in a text sequence. Unlike previous works [60, 61, 80], our F-LMM eliminates the necessity for high-quality conversation data that are annotated with masks to preserve conversational ability when learning grounding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Our experiments demonstrate that F-LMM maintains the original excellence of off-the-shelf LMMs on general question-answering benchmarks, while achieving competitive results on referring segmentation and phrase grounding. In more complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning, F-LMM achieves better or comparable results when contrasted with models specially trained for such tasks. Compared with existing grounding LMMs, F-LMM offers the best balance between grounding and chat capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.233, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.6, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Large Multimodal Models. Recent advancements in LMMs [2, 3, 13, 29-31, 34, 36, 38-40, 44, 47, 65, 75] have been fueled by the success of LLMs [1, 4, 12, 22, 48, 51, 66, 67, 81] since the debut of GPT series [1, 4, 56, 57] that feature an auto-regressive framework based on transformer decoders [68]. These LLMs possess general world knowledge and excellent conversational ability to follow human instructions, thanks to large-scale generative pre-training [4] and supervised finetuning on instruction-tuning data [71] or human feedback [50]. By integrating image representations from vision encoders [58, 79] to LLMs, LMMs enable visual understanding and reasoning in AI assistants. This integration is usually established by a multilayer perceptron (MLP) that directly maps image features to the LLMs' input embedding space [34, 38-40, 44, 65, 78] or a cross-attention module that abstracts the image contents with a set of query embeddings [2, 3, 31, 75]. In our research, we build F-LMM on LMMs of the former type (MLP-based), which preserves images' 2-D topological structure in the cross-modal integration." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "Visual Segmentation. The task of predicting 2D masks for visual objects is known as image segmentation, which can be categorised into semantic segmentation [5, 8, 10, 87], instance segmentation [11, 19, 83] and panoptic segmentation [9, 25, 32, 73] depending on whether the goal is to differentiate pixel semantics or object instances. These standard segmentation approaches rely on a pre-defined set of object classes for recognition. In contrast, referring expression segmentation (RES) [24, 37, 45, 46, 49, 74, 88] involves segmenting objects based on free-form human language descriptions, allowing for enhanced human-model interaction. Additionally, panoptic narrative grounding (PNG) [14, 16, 17, 69] requires segmenting masks for key phrases or words in a sentence. In this study, we mainly leverage RES and PNG tasks to evaluate the grounding capability of LMMs. Besides, we also test LMMs' segmentation ability in complex scenarios that necessitate reasoning [28, 60, 63]. Moreover, the prompt-based SAM [26] pre-trained on billion-scale high-quality mask data has become a constituent component in many grounding LMMs to boost segmentation performance. We also adopt SAM's mask head to initialise our mask refiner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.905, + 0.756 + ], + "angle": 0, + "content": "Grounding Large Multimodal Models. Grounding Large Multimodal Models [3, 7, 28, 53, 54, 60, 61, 70, 72, 76, 78, 80, 84, 85] can localise language contents during user-model conversations. Some approaches [3, 7, 53, 76] represent coordinates of bounding boxes as texts and train LMMs to predict the coordinates in a generative manner. Several recent works [28, 54, 60, 61, 70, 80] train LMMs to predict a special segmentation token for encoding the grounded object and utilise a segmentation head (e.g., SAM [26]) to decode object masks. This study mainly focuses on grounding LMMs with segmentation ability for visual perception. To obtain competitive visual grounding performance, existing works extensively fine-tune the parameters of LMMs on a large amount of segmentation [5, 18, 25, 59, 87] and grounding [16, 23, 24, 27, 46, 55] datasets. And to balance the LMMs' grounding and conversational abilities, there are efforts [61, 80, 84] to collect high-quality instruction-tuning data annotated with segmentation masks. In contrast, we make the first attempt to build grounding LMMs on top of off-the-shelf LMMs without fine-tuning their parameters. Furthermore, we bypass the need for grounded instruction-tuning data to preserve decent chat ability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.605, + 0.785 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce our F-LMM by first probing the causal attention mechanism in LMMs with visualisations of word-image attention maps in Sec 3.1. Then, we elaborate on F-LMM exploiting segmentation priors from frozen LMMs for visual grounding using the mask head in Sec 3.2. Finally, we show how to automate the process of grounded conversation with a linear keyword selector to in-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24712" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.093, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 3. The overall pipeline of F-LMM. The word-image attention maps from the frozen LLM serve as segmentation priors for the mask head. The keyword selector discovers object nouns in the text sequence. The mask head encompasses a mask decoder that translates attention weights to mask logits and a mask refiner that optimises the mask decoder's predictions. \\(M\\) and \\(N\\) represent the numbers of transformer layers and attention heads." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.359, + 0.475, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Figure 4. Visualisations of word-image attention maps. The letters \\( m \\) and \\( n \\) indicate that the attention map is derived from the \\( n \\)-th attention head of the \\( m \\)-th transformer layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.623 + ], + "angle": 0, + "content": "dicate words of grounding targets in Sec 3.3. The overall pipeline is illustrated in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.436, + 0.648 + ], + "angle": 0, + "content": "3.1. Segmentation Priors from Frozen LMM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Vision-Language Sequence. A typical build of a LMM \\(^2\\) comprises an image encoder \\(f_{v}\\) (e.g., CLIP [58] \\(^3\\)), a vision-language projector \\(f_{p}\\), and a LLM, denoted as \\(f_{\\mathrm{llm}}\\). The inputs to an LMM are usually an image \\(\\mathbf{X}_v \\in \\mathbb{R}^{3 \\times H \\times W}\\) and the associated text \\(\\mathbf{X}_t\\). The input image is first encoded by the vision encoder \\(f_{v}\\) and then mapped to the input space of the LLM by the projector \\(f_{p}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.77, + 0.426, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {v} = f _ {p} (\\text {F l a t t e n} (f _ {v} (\\mathbf {X} _ {v}))) \\in \\mathbb {R} ^ {h w \\times d},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.483, + 0.831 + ], + "angle": 0, + "content": "where \\( h \\) and \\( w \\) are the height and width of projected feature maps via \\( f_{v} \\). The Flatten operation unfolds the 2-D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.838, + 0.483, + 0.863 + ], + "angle": 0, + "content": "2In this paper, the term 'multimodal' stands for vision and language modalities." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": "3The image encoder might be any vision model that is pre-trained on image-text pairs. We use the classic term 'CLIP' in this paper to represent all such models for brevity." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.838, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.419 + ], + "angle": 0, + "content": "image feature map to a 1-D sequence. The constant \\( d \\) is the hidden state dimension of the LLM. Likewise, the text input is first encoded as discrete tokens and then mapped to text embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.431, + 0.851, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {t} = \\operatorname {E m b e d} (\\operatorname {T o k e n i z e} (\\mathbf {X} _ {t})) \\in \\mathbb {R} ^ {L \\times d},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.507 + ], + "angle": 0, + "content": "where \\(L\\) denotes the length of text embeddings. The visual-language sequence input to the LLM is a concatenation of image and text embeddings: \\(\\mathbf{Z} = \\{\\mathbf{Z}_v,\\mathbf{Z}_t\\} \\in \\mathbb{R}^{(hw + L)\\times d}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.651 + ], + "angle": 0, + "content": "Segmentation Priors in Self-Attention. The vision-language sequence is mainly processed by causal self-attentions [56, 68] in the LLM, including inner product and weighted-sum operations. Specifically, for a word token with position index \\( i \\) in the vision-language sequence \\( \\mathbf{Z} \\), its embedding \\( z^i \\) is updated by the weighted sum of the first \\( i \\) embeddings: \\( \\hat{z}^i = \\mathrm{SoftMax}\\left(\\frac{z^i \\cdot \\mathbf{Z}[i]}{d}\\right) \\cdot \\mathbf{Z}[i] \\), where \\( \\mathrm{SoftMax}\\left(\\frac{z^i \\cdot \\mathbf{Z}[i]}{d}\\right) \\) is the attention weights. Here, we omit the residual layers and feedforward layers for brevity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.906, + 0.713 + ], + "angle": 0, + "content": "Considering the word-image interaction in the multimodal scenario, we can select the word token's attention weights with the image embeddings from the overall vision-language attention weights:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.724, + 0.908, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} ^ {i} = \\text {U n f l a t t e n} (\\operatorname {S o f t M a x} (\\frac {\\boldsymbol {z} ^ {i} \\cdot \\mathbf {Z} [: i ]}{d}) [: h w ]) \\in \\mathbb {R} ^ {h \\times w},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where Un flatten restores the 2-D spatial structure from the 1-D sequence to form an attention map. In Figure 4, we visualise such word-image attention maps from various transformers layers and attention heads in an LMM (i.e., DeepSeekVL-7B [44]). The objects' shape and location can be observed in word-image attention maps of certain layers or heads. The visibility is further enhanced when we stack the attention maps from all layers and heads and perform K-Means clustering. It can be observed that the attention" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24713" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. The main evaluation results on question-answering benchmarks, referring expression segmentation (RES) benchmark and panoptic narrative grounding (PNG) benchmark. LLaVAW: LLaVA-In-the-Wild. LLaVA-1.6 and MGM-HD take high-resolution image inputs. LLaVA-1.6-M-7B means the model is based on Mistral-7B [22]. GLaMM-FS-7B means we use the 'FullScope' version of GLaMM." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.143, + 0.911, + 0.396 + ], + "angle": 0, + "content": "
ModelMultimodal Question AnsweringRESPNG
MMEMMBenchMMVetLLaVAWRefCOCORefCOCO+RefCOCOgAllThingStuff
PixelLM-7B [61]309/13517.415.946.473.066.369.343.141.047.9
LISA-7B [28]1/10.419.147.574.965.167.9---
PerceptionGPT-7B [54]----75.168.570.3---
LLaVA-G-7B [80]---55.877.168.871.5---
GroundHog-7B [84]----78.570.574.166.865.069.4
GLaMM-FS-7B [60]14/936.810.332.078.670.574.855.852.962.3
GSVA-7B [72]446/1817.819.438.377.768.273.241.839.646.6
LaSagnA-7B [70]0/00.016.734.576.866.470.6---
F-LMM (DeepSeekVL-1.3B [44])1307/22564.634.851.175.062.868.264.963.468.3
F-LMM (MGM-2B [34])1341/31259.831.165.975.063.767.365.664.468.4
F-LMM (LLaVA-1.5-7B [38])1511/34864.330.569.075.263.767.164.863.468.2
F-LMM (HPT-Air-6B [65])1010/ 25869.831.359.274.364.067.565.564.068.8
F-LMM (HPT-Air-1.5-8B [65])1476/30875.236.362.176.364.568.565.464.168.5
F-LMM (MGM-7B [34])1523/31669.340.875.875.764.868.366.365.368.6
F-LMM (DeepSeekVL-7B [44])1468/29873.241.577.876.166.470.165.764.568.5
F-LMM (LLaVA-1.6-7B [40])1519/32268.144.172.375.865.870.166.365.169.0
F-LMM (LLaVA-1.6-M-7B [40])1501/32469.547.871.775.766.570.166.565.469.1
F-LMM (MGM-HD-7B [34])1546/31965.841.374.076.165.268.566.765.669.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.483, + 0.451 + ], + "angle": 0, + "content": "maps offer meaningful segmentation priors with spatial and geometric cues for grounding objects visually." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.529 + ], + "angle": 0, + "content": "Language Cues. In addition to the spatial and geometric cues from word-image attention maps, F-LMM can also capitalise on the object's corresponding text embeddings from the LLM \\( f_{\\mathrm{llm}} \\), which provide extra language cues for the grounding of visual objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.396, + 0.56 + ], + "angle": 0, + "content": "3.2. Visual Grounding with Mask Head" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.612 + ], + "angle": 0, + "content": "We use the segmentation priors from the frozen LMM for pixel-level grounding, with the help of a mask head consisting of a mask decoder and a mask refiner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Mask Decoder. The mask decoder \\( f_{d} \\) is a 2-D CNN model that transforms the word-image attention maps of grounded objects into mask logits, which is instantiated by a 3-stage U-Net [62]. Please refer to the supplemental material for details on the mask decoder. The extraction of word-image attention map \\( \\pmb{a}^{i} \\) for a word token with position index \\( i \\) is illustrated in Eq. 3.1 and Figure 3. For an object described by multiple words, we merge their corresponding word-image attention maps to a single attention map \\( \\pmb{a} \\) via element-wise average or max operation. The attention map \\( \\pmb{a} \\) is further normalised as \\( \\pmb{a} / \\mathrm{sum}(\\pmb{a}) \\) so that all elements sum to 1. Considering \\( M \\) layers and \\( N \\) attention heads, we stack the \\( MN \\) attention maps as \\( \\pmb{A} \\in \\mathbb{R}^{MN \\times h \\times w} \\), which forms the input to a mask decoder. Given the importance of high input resolution for segmentation models, we upsample the stacked attention maps \\( \\pmb{A} \\) to \\( h' \\times w' \\) by bilinear interpolation before feeding it to a mask decoder, where \\( h' > h \\) and \\( w' > w \\). In practice, we set \\( h' = w' = 64 \\). Then, the mask decoder maps \\( \\pmb{A} \\) into mask" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.467 + ], + "angle": 0, + "content": "logits: \\(\\mathbf{M}_{\\mathrm{logits}} = f_d(\\mathbf{A})\\). We derive the corresponding binary mask via \\(\\mathbf{M}_{\\mathrm{pred}} = \\mathbf{M}_{\\mathrm{logits}} > 0\\). During training, the mask decoder is optimised with BCE and DICE losses [64]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.802 + ], + "angle": 0, + "content": "Mask Refiner. The mask refiner \\( f_{r} \\) is retrofitted from the mask head of SAM [26], which predicts masks based on prompts as well as image embeddings from SAM's ViT-based image encoder. To refine the output of the mask decoder \\( f_{d} \\), we re-use SAM's prompt encoder to transform \\( \\mathbf{M}_{\\mathrm{logits}} \\) into dense prompt embeddings (i.e., a 2-D feature map) and the bounding box of \\( \\mathbf{M}_{\\mathrm{pred}} \\) to box embeddings. In addition to the location cues from the mask and the box, the language cues, i.e., the object's corresponding text embeddings, are also utilised by \\( f_{r} \\). Considering text embeddings from the \\( M \\) transformer layers, we train \\( M \\) learnable scalars to calculate a weighted sum of these text embeddings. The weighted-summed text embeddings are processed by a linear layer and then concatenated with the box embeddings to form sparse prompt embeddings. The dense and sparse prompt embeddings, together with SAM's image embeddings, are fed to the mask refiner \\( f_{r} \\) for finer-grained mask predictions \\( \\mathbf{M}_{\\mathrm{pred}}^{\\prime} \\). During training, we keep the ViT-based image encoder of SAM frozen and optimise the mask refiner \\( f_{r} \\) using BCE loss and DICE loss [64]. For more details on the SAM's prompt-based mask head, please refer to the original SAM paper [26]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.905, + 0.833 + ], + "angle": 0, + "content": "3.3. Keyword Selector for Grounded Conversation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In grounded conversation [60] with interleaved segmentation masks and words, existing grounding LMMs [60, 61] expand LLMs' vocabularies with special tokens that indicate the start and end of grounding targets, which is infeas-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24714" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.09, + 0.36, + 0.104 + ], + "angle": 0, + "content": "Table 2. Reasoning Segmentation." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.108, + 0.418, + 0.213 + ], + "angle": 0, + "content": "
ModelVal -ShortTest LongAll
X-Decoder [88]22.620.422.221.7
SEEM [89]25.520.125.624.3
GroundingSAM [41]26.017.822.421.3
OVSeg [35]28.518.028.726.1
LISA [28]44.437.636.636.8
F-LMM46.736.949.146.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.427, + 0.09, + 0.887, + 0.103 + ], + "angle": 0, + "content": "Table 3. Grounded Conversation Generation (GCG). M. stands for METEOR." + }, + { + "type": "table", + "bbox": [ + 0.432, + 0.108, + 0.877, + 0.213 + ], + "angle": 0, + "content": "
ModelGCG TrainingM.Val mIoURecallM.Test mIoURecall
LISA [28]✓13.062.036.312.961.735.5
OMG-LLaVA [82]✓14.965.5-14.564.7-
GLaMM [60]✓16.266.341.815.865.640.8
BuboGPT [86]✗17.254.029.417.154.127.0
KOSMOS-2 [52]✗16.155.628.315.856.829.0
F-LMM✗17.663.542.017.463.638.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.222, + 0.848, + 0.236 + ], + "angle": 0, + "content": "Table 4. Unleashing visual chain-of-thought reasoning with both excellent grounding and instruction-following ability." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.24, + 0.835, + 0.317 + ], + "angle": 0, + "content": "
ModelVisual CoTDocVQATextCapsVisCoT BenchmarkPOPE
TextVQADUDESROIEInfographicsAccF1
VisCoT-7B [63]✓47.667.577.538.647.032.486.5-
F-LMM✗43.263.574.532.028.443.287.086.0
F-LMM✓53.867.978.442.344.149.188.087.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.482, + 0.462 + ], + "angle": 0, + "content": "ble in F-LMM given its ‘frozen’ nature. A common practice to discover visual objects in text sequences offline is using external tools such as SpaCy [20], which parse all nouns from a sentence including unwanted non-object words. Instead of adopting such offline tools that produce noisy results, we automate generating interleaved words and masks by training a linear layer to directly predict whether a word is to be grounded or not." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Specifically, the linear layer (keyword selector) is placed on top of the LLM's transformer layers, projecting \\(d\\)-dimension hidden state vectors into one-dimension scores, followed by a sigmoid function that normalizes the scores to \\([0,1]\\). During training, the score prediction is supervised by a BCE loss. During inference, word tokens with scores exceeding a threshold \\(\\lambda\\) are regarded as positive for visual grounding. In practice, we set \\(\\lambda = 0.3\\). Adjacent positive word tokens are grouped to indicate a single visual object. After the word tokens of visual objects are selected, the corresponding attention maps and text embeddings are fed to the mask head for visual grounding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.224, + 0.674 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.308, + 0.699 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Model Architectures. We build F-LMM on several open-sourced LMMs, including LLaVA-1.5 [38], LLaVA-Next [40], MiniGemini [34], DeepSeekVL [44] and HPT-Air [65]. The main experiment covers 10 LMMs with model sizes ranging from 1.3B to 8B. We employ a lightweight 3-stage U-Net [62] as the mask decoder to transform segmentation priors from frozen LLMs. The U-Net architecture features an encoder-decoder structure with skip connections, wherein feature maps are downsampled in the encoder and upsampled in the decoder. Please check the supplementary material for more details on the mask decoder. As for the SAM-based mask refiner, we choose SAM ViT-L [26] that balances cost and performance well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Model Training. We train F-LMM on RefCOCO(+/g) [24, 46] and PNG [16] datasets with about 190k data samples on a single machine with 8 NVIDIA A800-40G GPUs, which costs about 20 hours for each round of model training. We set the batch size to 8 and train models for 8 epochs, with gradient clipping at a max norm of 1.0. The AdamW [43] optimiser is used with a learning rate of 1e-4, a weight decay of 0.01, and betas as (0.9, 0.999). We choose a warm-up ratio of 0.03 to stabilise model optimisation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.487, + 0.824, + 0.503 + ], + "angle": 0, + "content": "4.2. Standard QA and Grounding Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.598 + ], + "angle": 0, + "content": "For a comprehensive study of LMMs' conversational and grounding capabilities, we first evaluate models under standard question-answering and grounding benchmarks separately. We summarise the evaluation results of grounding LMMs in Table 1. Please refer to the supplementary material for more detailed results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Benchmarks. For comprehensive conversational ability evaluation, we choose four widely used general question-answering benchmarks including MME [15], MMBench [42], LLaVA-In-the-Wild [39] and MMVet [77]. The MME and MMBench require an LMM to strictly follow the instruction to reply with single words (yes or no) or answer MCQs with alphabetical letters (i.e., answering A, B, C, or D). The LLaVA-In-the-Wild and MMVet benchmarks ask a model to respond with open-ended texts while demanding general world knowledge comprehension. In terms of grounding ability evaluation, we assess the LMMs' ability to segment user-described objects on referring expression segmentation (RES) [24, 46] benchmarks including RefCOCO, RefCOCO+, and RefCOCOg, using the cIoU metric. Due to limited space, we only report results on the Val splits of RefCOCO(+/g) in Table 1. We also test the LMMs' ability to ground key phrases or words in user-model conversations on the Panoptic Narrative Grounding (PNG) [16] benchmark, measuring individual mask recalls on thing/stuff objects and overall recall scores." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24715" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.482, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.298, + 0.424, + 0.311 + ], + "angle": 0, + "content": "Figure 5. Ablation study of the mask decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.523 + ], + "angle": 0, + "content": "Comparisons with Existing Methods. We compare F-LMM with existing grounding LMMs. As shown in Table 1, our F-LMM provides the best balance with conversational and grounding abilities among compared methods. On the question-answering benchmarks, existing grounding LMMs obtain zero or near-zero scores on MMBench and MME while lagging significantly behind general-purpose LMMs on MM Vet and LLaVA-In-the-Wild benchmarks, indicating compromised instruction-following ability and weakened general knowledge comprehension. On the RES and PNG benchmarks, our F-LMM achieves comparable results despite not having the parameters of LMMs fine-tuned for grounding purposes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.275, + 0.547 + ], + "angle": 0, + "content": "4.3. Complex Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.644 + ], + "angle": 0, + "content": "In this section, we evaluate grounding LMMs under more complex scenarios that typically require the LMMs to perform both reasoning and segmentation. The base LMM we use is DeepSeekVL-7B [44], considering its flexibility in supporting both single and multiple image inputs. And the size of models compared in this section is also 7B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Reasoning Segmentation is proposed in LISA [28] that requires a model to infer what object to segment from common-sense knowledge or via logical reasoning. The evaluation results are provided in Table 2 and the metric is gIoU. F-LMM can effectively perform reasoning segmentation even though it is not trained on such type of data. It is remarkable that F-LMM significantly outperforms existing models on the subset of long sentences, reflecting the advantage of F-LMM in handling complex contexts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Grounded Conversation Generation (GCG) is proposed in GLaMM [60], which requires a model to generate interleaved segmentation masks and texts. For performance evaluation, METEOR (M.) and mIoU are used to measure the quality of generated texts and masks, respectively. In addition, we report the recall of object masks. As shown in Table 3, F-LMM exhibits the best zero-shot performance while being comparable with models fine-tuned on GCG" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.09, + 0.899, + 0.103 + ], + "angle": 0, + "content": "Table 5. Ablation study of the mask refiner on PNG benchmark." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.106, + 0.872, + 0.185 + ], + "angle": 0, + "content": "
(a) Effects of Mask Refiner(b) SAM Variants
Mask RefinerPNGSAMPNG
maskboxtextAllThingStuffAllThingStuff
XXX50.848.655.9ViT-B63.061.466.8
✓XX63.462.066.8ViT-L64.963.468.3
✓✓X63.762.267.1ViT-H65.063.568.3
✓✓✓64.963.468.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.574, + 0.195, + 0.845, + 0.208 + ], + "angle": 0, + "content": "Table 6. Ablation study of keyword selection." + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.21, + 0.849, + 0.253 + ], + "angle": 0, + "content": "
MethodF1RecallPrecision
SpaCy Parser [20]57.897.341.1
Linear Keyword Selector82.896.672.5
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.275, + 0.623, + 0.289 + ], + "angle": 0, + "content": "training dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Visual Chain-of-Thought Reasoning. In human-AI conversations that involve Visual Chain-of-Thought Reasoning (VisCoT) [63], an LMM first localises the region/object relevant to the human's question and then generates the final answer by zooming in on the question-related region. Here we evaluate F-LMM and VisCoT-7B on the VisCoT benchmark [63]. As shown in Table 4, F-LMM achieves remarkable performance gains when prompted in the VisCoT manner. It is noticeable that F-LMM even outperforms VisCoT-7B [63] that has been well-tuned on the VisCoT training data [63]. Furthermore, we perform VisCoT on the object hallucination benchmark POPE [33] and observe significant performance gain in resisting object hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.667, + 0.515 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.567 + ], + "angle": 0, + "content": "We investigate the effects of design choices of F-LMM. All the ablation studies are conducted on the PNG dataset and we use DeepSeekVL-1.3B [44] unless otherwise stated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Mask Decoder. We summarise our analyses of the mask decoder in Figure 5. Note that the mask refiner is not involved in this part. We first consider attention maps from different transformer layers, i.e., early \\((6^{\\mathrm{th}})\\), mid \\((12^{\\mathrm{th}})\\) and late \\((24^{\\mathrm{th}})\\) layers. As shown in Figure 5(a), attention maps from late layers perform the worst, conforming to prior studies [21] indicating that deeper transformer layers tend to focus on abstract concepts instead of visual details. And using attention maps from all layers achieves the best performance. Next, we study how to merge the word-image attention maps of multi-word objects, as shown in Figure 5(b). We find that the average operation outperforms the max operation by a margin of 0.7. In Figure 5(c), we show that normalizing the inputs to the mask decoder provides a 0.5 performance gain. Finally, we experiment with different input sizes for the mask decoder. As shown in Figure 5(d), using \\(64 \\times 64\\) yields the best performance-cost trade-off." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Mask Refiner. We study the effects of the mask refiner on segmentation in Table 5a. The performance of using only the mask decoder is shown in the first row of Table 5a. With the masks fed to the mask refiner, we observe a significant 12.6 performance gain on the PNG benchmark. Adding" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24716" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.103, + 0.426, + 0.269 + ], + "angle": 0, + "content": "
ModelChat ↓Ground ↓
DeepseekVL-1.3B [44]7.758.33
MGM-2B [34]6.008.33
LLaVA-1.5-7B [38]6.757.83
HPT-Air-6B [65]9.007.16
HPT-Air-1.5-8B [65]6.507.00
MGM-7B [34]5.754.83
DeepseekVL-7B [44]3.754.00
LLaVA-1.6-7B [40]2.753.00
LLaVA-1.6-M-7B [40]3.251.66
MGM-HD-7B [34]3.502.83
" + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.099, + 0.87, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 6. The relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. The left table shows the average ranks of each LMM on question-answering ('Chat') and grounding benchmarks ('Ground'). \\(\\downarrow\\) means the lower the better. The dashed line in the right figure is the linear fit of the rank data points, indicating a positive correlation between abilities to chat and ground." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.482, + 0.448 + ], + "angle": 0, + "content": "box and text prompts to the mask refiner further improves the performance by 0.3 and 1.2. Then we experiment with different SAM [26] model variants, i.e., ViT-B(ase), ViT-L(arge) and ViT-H(uge) as mask refiners. As shown in Table 5b, the performance grows with model sizes. For a good trade-off between cost and performance, we select the ViT-L variant of SAM as the default mask refiner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Keyword Selector. We analyse the keyword selector, implemented as a linear layer, on the PNG dataset, using the F1 score as the main metric. As shown in Table 6, our keyword selector achieves significantly higher F1 scores than the external SpaCy tool [20]. We also report recall and precision scores. Our keyword selector achieves comparable recall while being much more precise compared to SpaCy, which enumerates all nouns in a sentence." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.577, + 0.317, + 0.592 + ], + "angle": 0, + "content": "4.5. Analysis & Visualisation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Scalability: Does Better Chatting Lead to Better Grounding? We study the relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. Specifically, we examine the correlation between performance on the question-answering and grounding benchmarks. For the ten models reported in Table 1, we calculate their average ranks in each benchmark category. In Figure 6, we plot these ranks as 2D coordinates, i.e., (Chat Rank, Ground Rank), and apply a linear fit to the data points. As indicated by the blue dashed line, frozen LMMs with stronger conversational ability can serve as better backbones for grounding. We also observe that larger LMMs generally excel in both conversation and grounding tasks, and LMMs with larger input resolutions (e.g., LLaVA-1.6 and MGM-HD) can handle both tasks better." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "From Attention Maps to Segmentation Masks. We visualise the word-image attention maps by applying KMeans clustering to the stacked attention maps that are collected from all transformer layers and attention heads. The attention maps of multiple-word objects are merged by element" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.342, + 0.646, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.41, + 0.644, + 0.471 + ], + "angle": 0, + "content": "Caption: There is a person in this picture on the surfing board who is surfing on a ocean tide. There is sky in the background." + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.342, + 0.77, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.409, + 0.77, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.342, + 0.892, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.409, + 0.892, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.473, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Figure 7. Visualisations of KMeans clustering on attention maps and segmentation results from the CNN-based mask decoder and the SAM-based mask refiner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.62 + ], + "angle": 0, + "content": "wise average. As shown in Figure 7, we observe that the pixels of objects are roughly clustered together (top-left). With the CNN-Based mask decoder, the attention weights are mapped to 2D binary masks (top-right), which are then further optimised by the SAM-based mask refiner (bottom-left). The model used is DeepSeekVL-1.3B [44]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.634, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.9 + ], + "angle": 0, + "content": "In this work, we have studied the limitation of existing grounding LMMs, i.e., the loss of general world knowledge and instruction-following ability. To address this issue, we make the first attempt to ground fully frozen LMMs, which are already well-trained for user-model conversation, based on the insight that the geometric and spatial cues needed for visual grounding are inherently present within the self-attention mechanism of LMMs. By incorporating a CNN-based mask decoder and a SAM-based mask refiner, we achieve competitive visual grounding performance without sacrificing any conversational abilities of pre-trained LMMs. With this combination of strong conversational and visual grounding capabilities, these LMMs show promise for complex perception and reasoning tasks, such as segmentation with reasoning, grounded conversation generation, and visual chain-of-thought reasoning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24717" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Acknowledgement. This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-PhD-2023-08-048T), the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.211, + 0.188, + 0.227 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.308, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.449, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.52, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In \\(CVPR\\), pages 1209-1218, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems, 34:17864-17875, 2021." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[14] Zihan Ding, Zi-han Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Xiaolin Wei, and Si Liu. Ppmn: Pixel-phrase matching network for one-stage panoptic narrative grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5537–5546, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[16] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[17] Tianyu Guo, Haowei Wang, Yiwei Ma, Jiayi Ji, and Xiaoshuai Sun. Improving panoptic narrative grounding by harnessing semantic relationships and visual confirmation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1985-1993, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[18] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[20] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[21] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lam-" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24718" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "ple, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[24] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.362 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[27] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[29] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[30] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 214-223, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[33] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[34] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[36] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[37] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.904, + 0.263 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.904, + 0.292 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.463 + ], + "angle": 0, + "content": "[42] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.904, + 0.491 + ], + "angle": 0, + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.562 + ], + "angle": 0, + "content": "[44] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[45] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[46] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[47] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24719" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[51] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[53] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[54] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[55] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[56] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[59] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[60] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[61] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[63] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[64] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240–248. Springer, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.904, + 0.342 + ], + "angle": 0, + "content": "[65] HyperGAI Team. Hpt 1.5 air: Best open-sourced 8b multimodal llm with llama 3, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[66] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[69] Haowei Wang, Jiayi Ji, Yiyi Zhou, Yongjian Wu, and Xiaoshuai Sun. Towards real-time panoptic narrative grounding by an end-to-end grounding network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2528-2536, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[71] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.907, + 0.762 + ], + "angle": 0, + "content": "[72] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[73] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[74] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24720" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[75] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[78] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[80] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, and Jianwei Yang. Llava-grounding: Grounded visual chat with large multimodal models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[81] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[82] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[83] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. arXiv preprint arXiv:2402.16846, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[85] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[86] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[87] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[88] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[89] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24721" + } + ] +] \ No newline at end of file diff --git a/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_origin.pdf b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a411cc6cf2cdcecc87381b69f7c42b33823b84a --- /dev/null +++ b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/6e34bb31-55f8-4e82-8db1-d754fa78afe2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c84ebf3a31bdac40f65646459d9863aed5bcb7ce0a0662afee8a66944d2b1253 +size 3242624 diff --git a/2025/F-LMM_ Grounding Frozen Large Multimodal Models/full.md b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75ebc79d587f5b8f79228d6430a0d1cc61320e7a --- /dev/null +++ b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/full.md @@ -0,0 +1,325 @@ +# F-LMM: Grounding Frozen Large Multimodal Models + +Size $\mathbf{W}\mathbf{u}^{1}$ Sheng Jin2 Wenwei Zhang3 Lumin $\mathrm{Xu^4}$ Wentao Liu2,3 Wei Li1 Chen Change Loy' 1 S-Lab, Nanyang Technological University SenseTime Research and Tetras.AI 3 Shanghai AI Laboratory 4 The Chinese University of Hong Kong size001@e.ntu.edu.sg ccloy@ntu.edu.sg + +![](images/a3e1e1ddd8c03e43d3c0dd350064ba99f869d5854bc1ac492e6c926ed0e144a9.jpg) +User: Is there a truck in the image? Answer the question using a single word or phrase. +GLaMM: Sure, the segmentation is [SEG]. + +![](images/2ca29c51bc8342b4e04f86ee755c6631376ef687c3e887d215817b7886cfc69a.jpg) +Figure 1. An example of user-AI conversation around an image. Left: The current state-of-the-art grounding model GLaMM [60] is effective for grounded conversation when prompted by "answer with interleaved masks", but fails to follow user instruction to answer a single word (yes or no) and misunderstands the question as a referring segmentation prompt. Right: Our F-LMM preserves instruction-following ability while being able to perform visual grounding. + +![](images/12eb4d80a541344329879bb511f1fd89872ab10b0b0c670f1b2d7b5c39d537d3.jpg) +User: Is there a truck in the image? Answer the question using a single word or phrase. F-LMM: No. + +# Abstract + +Endowing Large Multimodal Models (LMMs) with visual grounding capability can significantly enhance AIs' understanding of the visual world and their interaction with humans. However, existing methods typically fine-tune the parameters of LMMs to learn additional segmentation tokens and overfit grounding and segmentation datasets. Such a design would inevitably cause a catastrophic diminution in the indispensable conversational capability of general AI assistants. In this paper, we comprehensively evaluate state-of-the-art grounding LMMs across a suite of multimodal question-answering benchmarks, observing drastic performance drops that indicate vanishing general knowledge comprehension and weakened instruction following ability. To address this issue, we present F-LMM—grounding frozen off-the-shelf LMMs in human-AI conversations—a straightforward yet effective design based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention mechanism of well-trained LMMs. Using only a few trainable CNN layers, we can translate word-pixel attention weights to mask logits, which a SAM-based mask refiner can further optimise. Our F-LMM neither learns special segmentation to + +kens nor utilises high-quality grounded instruction-tuning data, but achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks while completely preserving LMMs' original conversational ability. Additionally, with instruction-following ability preserved and grounding ability obtained, F-LMM can be directly applied to complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning. Our code can be found at https://github.com/wusize/F-LMM. + +# 1. Introduction + +Large Multimodal Models (LMMs), which integrate Large Language Models (LLMs) with visual signals, have demonstrated remarkable success in multimodal understanding, reasoning and interaction [34, 38-40, 44, 65, 78]. To further advance LMMs with better perception capability, a recent line of research [28, 60, 61, 70, 80, 84] that visually grounds language contents in user-model conversations has drawn increasing attention. This explicit association between key phrases/words and visual objects greatly enhances LMMs' understanding of the visual world and allows for more intuitive and meaningful human-AI interactions. + +![](images/5ccb99f61929303a3bfe0feec5c8d55ed6f07a48f4eedf1504262f92de3cbe7b.jpg) +"two girls" "a fence" + +![](images/c3cf1e4093fa2e9ae3275d2a9fc469c196e35cda35230f4852ef30471d0dd714.jpg) +"two elephants" "grass" "hills" "sky" + +![](images/4ac40a0fd40619123168822b2cf63c72a0ad923dbcefa9bab400bd792aae79a0.jpg) +"bread" "the plate" + +![](images/d618b2c3430986b89a4d9d51ae5a1caa9d714006f84a572b0f4cee7ad8f10634.jpg) +(a) Visualisations of word-image attention maps in Frozen LMMs via KMeans. +(b) Existing Grounding LMMs. +Figure 2. (a) Geometric and spatial cues conducive to visual grounding are observed in the visualisations of word-image attention maps in frozen LMMs. (b) Existing grounding LMMs are fine-tuned to generate a special mask token (e.g., [SEG]) for visual grounding purposes, which ruins the original conversational ability. (c) Our F-LMM translates word-image attention maps from frozen LMMs to grounding masks, while fully preserving the general-purpose chat capability. + +![](images/eb98c5d08c498d3122b05491d35a0e3d19d347881dd2424529f021dc64b7dfc9.jpg) +(c) F-LMM (Ours). + +By design, one commonly adopted build (Figure 2(b)) for visually grounding language contents is connecting LMMs with a mask head (e.g., Segment Anything Model (SAM) [26]), wherein both the LLM backbone and the mask head are fine-tuned with well-prepared visual grounding data that contains segmentation annotations. Also, some additional learnable tokens (e.g., [SEG]) are introduced to the LMMs' vocabulary, to directly associate key phrases or words with visual objects in conversations. However, this design will inevitably provoke a catastrophic diminution in general knowledge comprehension and instruction-following ability due to the following reasons. First, existing segmentation and visual grounding data only contain elementary patterns for answering simple grounding prompts. Second, during the fine-tuning stage, the LMMs are mainly optimised for effectively modelling the relationship between key phrases or words and special segmentation tokens, i.e., overfitting the segmentation and grounding data. Therefore, the conversational ability is sacrificed. For instance, the state-of-the-art grounding model GLaMM [60] fails to answer a simple yes-or-no question (Figure 1). Moreover, quantitative evaluations of existing grounding LMMs in conversational ability are presented in Table 1, with zero or near-zero scores on general multimodal question-answering benchmarks necessitating instruction-following ability. + +One possible option to deal with this dilemma is to collect high-quality training data encompassing both meaningful conversations and mask annotations. For example, LLaVA-G [80] annotates the 150k LLaVA-Instruct data samples [39] with segmentation masks so that the LMMs + +simultaneously learn to chat and segment. Nonetheless, annotating high-quality grounded conversation data is costly and hard to scale. Despite being trained on costly annotated data, LLaVA-G still lags behind general-purpose LMMs on multimodal understanding tasks. Furthermore, training on large-scale annotated data normally consumes significant computational resources, which is, obviously not a resource-efficient solution. + +In this paper, we propose a simple yet effective design, i.e., grounding frozen LMMs (dubbed as F-LMM) in human-AI conversations. We argue that freezing the parameters of well-trained LMMs is the most practical design choice for fully preserving the original excellent conversational ability when building general-purpose grounding LMMs. In particular, we take inspiration from the built-in interpretability of the attention mechanism in transformers [6, 68] that represents interrelations between text and image contents in design. We observe that off-the-shelf LMMs already produce word-pixel correspondences necessary for visual grounding, despite they were not explicitly pre-trained with region or pixel annotations. As illustrated in Figure 2(a), we visualise word-image attention maps from frozen LMMs via K-Means clustering, revealing notable geometric and spatial cues of the objects1. For example, coarse visual grounding masks for key phrases (e.g., "two girls", "two elephants", and "the plate") in language sentences emerge from attention maps in LMMs. Therefore, our F-LMM takes these visual-language correspondences as useful segmentation priors for decoding ground + +ing masks, without further tuning the LMMs' weights or learning a special segmentation token to model object locations, as shown in Figure 2(c). + +The only trainable part of our F-LMM is a mask head plus a keyword selector. The mask head comprises a CNN-based mask decoder (a tiny U-Net [62]) that translates stacked attention maps to mask logits and a lightweight mask refiner (retrofitted from SAM [26]'s mask head) that uses additional image and language cues to refine the semantic-agnostic masks from the mask decoder. The keyword selector is a linear layer that discovers object nouns in text sequences, automating the process of grounded human-AI conversation. Moreover, we only use the RefCOCO $(+ / g)$ [24, 46] and PNG [16] datasets as our training data, enabling LMMs to segment user-described objects and ground key phrases or words in a text sequence. Unlike previous works [60, 61, 80], our F-LMM eliminates the necessity for high-quality conversation data that are annotated with masks to preserve conversational ability when learning grounding. + +Our experiments demonstrate that F-LMM maintains the original excellence of off-the-shelf LMMs on general question-answering benchmarks, while achieving competitive results on referring segmentation and phrase grounding. In more complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning, F-LMM achieves better or comparable results when contrasted with models specially trained for such tasks. Compared with existing grounding LMMs, F-LMM offers the best balance between grounding and chat capabilities. + +# 2. Related Work + +Large Multimodal Models. Recent advancements in LMMs [2, 3, 13, 29-31, 34, 36, 38-40, 44, 47, 65, 75] have been fueled by the success of LLMs [1, 4, 12, 22, 48, 51, 66, 67, 81] since the debut of GPT series [1, 4, 56, 57] that feature an auto-regressive framework based on transformer decoders [68]. These LLMs possess general world knowledge and excellent conversational ability to follow human instructions, thanks to large-scale generative pre-training [4] and supervised finetuning on instruction-tuning data [71] or human feedback [50]. By integrating image representations from vision encoders [58, 79] to LLMs, LMMs enable visual understanding and reasoning in AI assistants. This integration is usually established by a multilayer perceptron (MLP) that directly maps image features to the LLMs' input embedding space [34, 38-40, 44, 65, 78] or a cross-attention module that abstracts the image contents with a set of query embeddings [2, 3, 31, 75]. In our research, we build F-LMM on LMMs of the former type (MLP-based), which preserves images' 2-D topological structure in the cross-modal integration. + +Visual Segmentation. The task of predicting 2D masks for visual objects is known as image segmentation, which can be categorised into semantic segmentation [5, 8, 10, 87], instance segmentation [11, 19, 83] and panoptic segmentation [9, 25, 32, 73] depending on whether the goal is to differentiate pixel semantics or object instances. These standard segmentation approaches rely on a pre-defined set of object classes for recognition. In contrast, referring expression segmentation (RES) [24, 37, 45, 46, 49, 74, 88] involves segmenting objects based on free-form human language descriptions, allowing for enhanced human-model interaction. Additionally, panoptic narrative grounding (PNG) [14, 16, 17, 69] requires segmenting masks for key phrases or words in a sentence. In this study, we mainly leverage RES and PNG tasks to evaluate the grounding capability of LMMs. Besides, we also test LMMs' segmentation ability in complex scenarios that necessitate reasoning [28, 60, 63]. Moreover, the prompt-based SAM [26] pre-trained on billion-scale high-quality mask data has become a constituent component in many grounding LMMs to boost segmentation performance. We also adopt SAM's mask head to initialise our mask refiner. + +Grounding Large Multimodal Models. Grounding Large Multimodal Models [3, 7, 28, 53, 54, 60, 61, 70, 72, 76, 78, 80, 84, 85] can localise language contents during user-model conversations. Some approaches [3, 7, 53, 76] represent coordinates of bounding boxes as texts and train LMMs to predict the coordinates in a generative manner. Several recent works [28, 54, 60, 61, 70, 80] train LMMs to predict a special segmentation token for encoding the grounded object and utilise a segmentation head (e.g., SAM [26]) to decode object masks. This study mainly focuses on grounding LMMs with segmentation ability for visual perception. To obtain competitive visual grounding performance, existing works extensively fine-tune the parameters of LMMs on a large amount of segmentation [5, 18, 25, 59, 87] and grounding [16, 23, 24, 27, 46, 55] datasets. And to balance the LMMs' grounding and conversational abilities, there are efforts [61, 80, 84] to collect high-quality instruction-tuning data annotated with segmentation masks. In contrast, we make the first attempt to build grounding LMMs on top of off-the-shelf LMMs without fine-tuning their parameters. Furthermore, we bypass the need for grounded instruction-tuning data to preserve decent chat ability. + +# 3. Method + +In this section, we introduce our F-LMM by first probing the causal attention mechanism in LMMs with visualisations of word-image attention maps in Sec 3.1. Then, we elaborate on F-LMM exploiting segmentation priors from frozen LMMs for visual grounding using the mask head in Sec 3.2. Finally, we show how to automate the process of grounded conversation with a linear keyword selector to in- + +![](images/bdd35d9e9487d868b1abd530b35df37f4c3c27f477c1e56b89ec8768f052fa50.jpg) +Figure 3. The overall pipeline of F-LMM. The word-image attention maps from the frozen LLM serve as segmentation priors for the mask head. The keyword selector discovers object nouns in the text sequence. The mask head encompasses a mask decoder that translates attention weights to mask logits and a mask refiner that optimises the mask decoder's predictions. $M$ and $N$ represent the numbers of transformer layers and attention heads. + +![](images/dace6dd8828545b4d0e50fc39618584f53839d77dc475e4854ea7a1a53622d83.jpg) +Figure 4. Visualisations of word-image attention maps. The letters $m$ and $n$ indicate that the attention map is derived from the $n$ -th attention head of the $m$ -th transformer layer. + +dicate words of grounding targets in Sec 3.3. The overall pipeline is illustrated in Figure 3. + +# 3.1. Segmentation Priors from Frozen LMM + +Vision-Language Sequence. A typical build of a LMM $^2$ comprises an image encoder $f_{v}$ (e.g., CLIP [58] $^3$ ), a vision-language projector $f_{p}$ , and a LLM, denoted as $f_{\mathrm{llm}}$ . The inputs to an LMM are usually an image $\mathbf{X}_v \in \mathbb{R}^{3 \times H \times W}$ and the associated text $\mathbf{X}_t$ . The input image is first encoded by the vision encoder $f_{v}$ and then mapped to the input space of the LLM by the projector $f_{p}$ : + +$$ +\mathbf {Z} _ {v} = f _ {p} (\text {F l a t t e n} (f _ {v} (\mathbf {X} _ {v}))) \in \mathbb {R} ^ {h w \times d}, +$$ + +where $h$ and $w$ are the height and width of projected feature maps via $f_{v}$ . The Flatten operation unfolds the 2-D + +image feature map to a 1-D sequence. The constant $d$ is the hidden state dimension of the LLM. Likewise, the text input is first encoded as discrete tokens and then mapped to text embeddings: + +$$ +\mathbf {Z} _ {t} = \operatorname {E m b e d} (\operatorname {T o k e n i z e} (\mathbf {X} _ {t})) \in \mathbb {R} ^ {L \times d}, +$$ + +where $L$ denotes the length of text embeddings. The visual-language sequence input to the LLM is a concatenation of image and text embeddings: $\mathbf{Z} = \{\mathbf{Z}_v,\mathbf{Z}_t\} \in \mathbb{R}^{(hw + L)\times d}$ + +Segmentation Priors in Self-Attention. The vision-language sequence is mainly processed by causal self-attentions [56, 68] in the LLM, including inner product and weighted-sum operations. Specifically, for a word token with position index $i$ in the vision-language sequence $\mathbf{Z}$ , its embedding $z^i$ is updated by the weighted sum of the first $i$ embeddings: $\hat{z}^i = \mathrm{SoftMax}\left(\frac{z^i \cdot \mathbf{Z}[i]}{d}\right) \cdot \mathbf{Z}[i]$ , where $\mathrm{SoftMax}\left(\frac{z^i \cdot \mathbf{Z}[i]}{d}\right)$ is the attention weights. Here, we omit the residual layers and feedforward layers for brevity. + +Considering the word-image interaction in the multimodal scenario, we can select the word token's attention weights with the image embeddings from the overall vision-language attention weights: + +$$ +\boldsymbol {a} ^ {i} = \text {U n f l a t t e n} (\operatorname {S o f t M a x} (\frac {\boldsymbol {z} ^ {i} \cdot \mathbf {Z} [: i ]}{d}) [: h w ]) \in \mathbb {R} ^ {h \times w}, +$$ + +where Un flatten restores the 2-D spatial structure from the 1-D sequence to form an attention map. In Figure 4, we visualise such word-image attention maps from various transformers layers and attention heads in an LMM (i.e., DeepSeekVL-7B [44]). The objects' shape and location can be observed in word-image attention maps of certain layers or heads. The visibility is further enhanced when we stack the attention maps from all layers and heads and perform K-Means clustering. It can be observed that the attention + +Table 1. The main evaluation results on question-answering benchmarks, referring expression segmentation (RES) benchmark and panoptic narrative grounding (PNG) benchmark. LLaVAW: LLaVA-In-the-Wild. LLaVA-1.6 and MGM-HD take high-resolution image inputs. LLaVA-1.6-M-7B means the model is based on Mistral-7B [22]. GLaMM-FS-7B means we use the 'FullScope' version of GLaMM. + +
ModelMultimodal Question AnsweringRESPNG
MMEMMBenchMMVetLLaVAWRefCOCORefCOCO+RefCOCOgAllThingStuff
PixelLM-7B [61]309/13517.415.946.473.066.369.343.141.047.9
LISA-7B [28]1/10.419.147.574.965.167.9---
PerceptionGPT-7B [54]----75.168.570.3---
LLaVA-G-7B [80]---55.877.168.871.5---
GroundHog-7B [84]----78.570.574.166.865.069.4
GLaMM-FS-7B [60]14/936.810.332.078.670.574.855.852.962.3
GSVA-7B [72]446/1817.819.438.377.768.273.241.839.646.6
LaSagnA-7B [70]0/00.016.734.576.866.470.6---
F-LMM (DeepSeekVL-1.3B [44])1307/22564.634.851.175.062.868.264.963.468.3
F-LMM (MGM-2B [34])1341/31259.831.165.975.063.767.365.664.468.4
F-LMM (LLaVA-1.5-7B [38])1511/34864.330.569.075.263.767.164.863.468.2
F-LMM (HPT-Air-6B [65])1010/ 25869.831.359.274.364.067.565.564.068.8
F-LMM (HPT-Air-1.5-8B [65])1476/30875.236.362.176.364.568.565.464.168.5
F-LMM (MGM-7B [34])1523/31669.340.875.875.764.868.366.365.368.6
F-LMM (DeepSeekVL-7B [44])1468/29873.241.577.876.166.470.165.764.568.5
F-LMM (LLaVA-1.6-7B [40])1519/32268.144.172.375.865.870.166.365.169.0
F-LMM (LLaVA-1.6-M-7B [40])1501/32469.547.871.775.766.570.166.565.469.1
F-LMM (MGM-HD-7B [34])1546/31965.841.374.076.165.268.566.765.669.1
+ +maps offer meaningful segmentation priors with spatial and geometric cues for grounding objects visually. + +Language Cues. In addition to the spatial and geometric cues from word-image attention maps, F-LMM can also capitalise on the object's corresponding text embeddings from the LLM $f_{\mathrm{llm}}$ , which provide extra language cues for the grounding of visual objects. + +# 3.2. Visual Grounding with Mask Head + +We use the segmentation priors from the frozen LMM for pixel-level grounding, with the help of a mask head consisting of a mask decoder and a mask refiner. + +Mask Decoder. The mask decoder $f_{d}$ is a 2-D CNN model that transforms the word-image attention maps of grounded objects into mask logits, which is instantiated by a 3-stage U-Net [62]. Please refer to the supplemental material for details on the mask decoder. The extraction of word-image attention map $\pmb{a}^{i}$ for a word token with position index $i$ is illustrated in Eq. 3.1 and Figure 3. For an object described by multiple words, we merge their corresponding word-image attention maps to a single attention map $\pmb{a}$ via element-wise average or max operation. The attention map $\pmb{a}$ is further normalised as $\pmb{a} / \mathrm{sum}(\pmb{a})$ so that all elements sum to 1. Considering $M$ layers and $N$ attention heads, we stack the $MN$ attention maps as $\pmb{A} \in \mathbb{R}^{MN \times h \times w}$ , which forms the input to a mask decoder. Given the importance of high input resolution for segmentation models, we upsample the stacked attention maps $\pmb{A}$ to $h' \times w'$ by bilinear interpolation before feeding it to a mask decoder, where $h' > h$ and $w' > w$ . In practice, we set $h' = w' = 64$ . Then, the mask decoder maps $\pmb{A}$ into mask + +logits: $\mathbf{M}_{\mathrm{logits}} = f_d(\mathbf{A})$ . We derive the corresponding binary mask via $\mathbf{M}_{\mathrm{pred}} = \mathbf{M}_{\mathrm{logits}} > 0$ . During training, the mask decoder is optimised with BCE and DICE losses [64]. + +Mask Refiner. The mask refiner $f_{r}$ is retrofitted from the mask head of SAM [26], which predicts masks based on prompts as well as image embeddings from SAM's ViT-based image encoder. To refine the output of the mask decoder $f_{d}$ , we re-use SAM's prompt encoder to transform $\mathbf{M}_{\mathrm{logits}}$ into dense prompt embeddings (i.e., a 2-D feature map) and the bounding box of $\mathbf{M}_{\mathrm{pred}}$ to box embeddings. In addition to the location cues from the mask and the box, the language cues, i.e., the object's corresponding text embeddings, are also utilised by $f_{r}$ . Considering text embeddings from the $M$ transformer layers, we train $M$ learnable scalars to calculate a weighted sum of these text embeddings. The weighted-summed text embeddings are processed by a linear layer and then concatenated with the box embeddings to form sparse prompt embeddings. The dense and sparse prompt embeddings, together with SAM's image embeddings, are fed to the mask refiner $f_{r}$ for finer-grained mask predictions $\mathbf{M}_{\mathrm{pred}}^{\prime}$ . During training, we keep the ViT-based image encoder of SAM frozen and optimise the mask refiner $f_{r}$ using BCE loss and DICE loss [64]. For more details on the SAM's prompt-based mask head, please refer to the original SAM paper [26]. + +# 3.3. Keyword Selector for Grounded Conversation + +In grounded conversation [60] with interleaved segmentation masks and words, existing grounding LMMs [60, 61] expand LLMs' vocabularies with special tokens that indicate the start and end of grounding targets, which is infeas- + +Table 2. Reasoning Segmentation. + +
ModelVal -ShortTest LongAll
X-Decoder [88]22.620.422.221.7
SEEM [89]25.520.125.624.3
GroundingSAM [41]26.017.822.421.3
OVSeg [35]28.518.028.726.1
LISA [28]44.437.636.636.8
F-LMM46.736.949.146.2
+ +Table 3. Grounded Conversation Generation (GCG). M. stands for METEOR. + +
ModelGCG TrainingM.Val mIoURecallM.Test mIoURecall
LISA [28]✓13.062.036.312.961.735.5
OMG-LLaVA [82]✓14.965.5-14.564.7-
GLaMM [60]✓16.266.341.815.865.640.8
BuboGPT [86]✗17.254.029.417.154.127.0
KOSMOS-2 [52]✗16.155.628.315.856.829.0
F-LMM✗17.663.542.017.463.638.6
+ +Table 4. Unleashing visual chain-of-thought reasoning with both excellent grounding and instruction-following ability. + +
ModelVisual CoTDocVQATextCapsVisCoT BenchmarkPOPE
TextVQADUDESROIEInfographicsAccF1
VisCoT-7B [63]✓47.667.577.538.647.032.486.5-
F-LMM✗43.263.574.532.028.443.287.086.0
F-LMM✓53.867.978.442.344.149.188.087.7
+ +ble in F-LMM given its ‘frozen’ nature. A common practice to discover visual objects in text sequences offline is using external tools such as SpaCy [20], which parse all nouns from a sentence including unwanted non-object words. Instead of adopting such offline tools that produce noisy results, we automate generating interleaved words and masks by training a linear layer to directly predict whether a word is to be grounded or not. + +Specifically, the linear layer (keyword selector) is placed on top of the LLM's transformer layers, projecting $d$ -dimension hidden state vectors into one-dimension scores, followed by a sigmoid function that normalizes the scores to $[0,1]$ . During training, the score prediction is supervised by a BCE loss. During inference, word tokens with scores exceeding a threshold $\lambda$ are regarded as positive for visual grounding. In practice, we set $\lambda = 0.3$ . Adjacent positive word tokens are grouped to indicate a single visual object. After the word tokens of visual objects are selected, the corresponding attention maps and text embeddings are fed to the mask head for visual grounding. + +# 4. Experiments + +# 4.1. Implementation Details + +Model Architectures. We build F-LMM on several open-sourced LMMs, including LLaVA-1.5 [38], LLaVA-Next [40], MiniGemini [34], DeepSeekVL [44] and HPT-Air [65]. The main experiment covers 10 LMMs with model sizes ranging from 1.3B to 8B. We employ a lightweight 3-stage U-Net [62] as the mask decoder to transform segmentation priors from frozen LLMs. The U-Net architecture features an encoder-decoder structure with skip connections, wherein feature maps are downsampled in the encoder and upsampled in the decoder. Please check the supplementary material for more details on the mask decoder. As for the SAM-based mask refiner, we choose SAM ViT-L [26] that balances cost and performance well. + +Model Training. We train F-LMM on RefCOCO(+/g) [24, 46] and PNG [16] datasets with about 190k data samples on a single machine with 8 NVIDIA A800-40G GPUs, which costs about 20 hours for each round of model training. We set the batch size to 8 and train models for 8 epochs, with gradient clipping at a max norm of 1.0. The AdamW [43] optimiser is used with a learning rate of 1e-4, a weight decay of 0.01, and betas as (0.9, 0.999). We choose a warm-up ratio of 0.03 to stabilise model optimisation. + +# 4.2. Standard QA and Grounding Tasks + +For a comprehensive study of LMMs' conversational and grounding capabilities, we first evaluate models under standard question-answering and grounding benchmarks separately. We summarise the evaluation results of grounding LMMs in Table 1. Please refer to the supplementary material for more detailed results. + +Benchmarks. For comprehensive conversational ability evaluation, we choose four widely used general question-answering benchmarks including MME [15], MMBench [42], LLaVA-In-the-Wild [39] and MMVet [77]. The MME and MMBench require an LMM to strictly follow the instruction to reply with single words (yes or no) or answer MCQs with alphabetical letters (i.e., answering A, B, C, or D). The LLaVA-In-the-Wild and MMVet benchmarks ask a model to respond with open-ended texts while demanding general world knowledge comprehension. In terms of grounding ability evaluation, we assess the LMMs' ability to segment user-described objects on referring expression segmentation (RES) [24, 46] benchmarks including RefCOCO, RefCOCO+, and RefCOCOg, using the cIoU metric. Due to limited space, we only report results on the Val splits of RefCOCO(+/g) in Table 1. We also test the LMMs' ability to ground key phrases or words in user-model conversations on the Panoptic Narrative Grounding (PNG) [16] benchmark, measuring individual mask recalls on thing/stuff objects and overall recall scores. + +![](images/0210f968d82afa7d617a0ba78a1b43c3eb27779d9453f770ab632a437e6458b0.jpg) +Figure 5. Ablation study of the mask decoder. + +Comparisons with Existing Methods. We compare F-LMM with existing grounding LMMs. As shown in Table 1, our F-LMM provides the best balance with conversational and grounding abilities among compared methods. On the question-answering benchmarks, existing grounding LMMs obtain zero or near-zero scores on MMBench and MME while lagging significantly behind general-purpose LMMs on MM Vet and LLaVA-In-the-Wild benchmarks, indicating compromised instruction-following ability and weakened general knowledge comprehension. On the RES and PNG benchmarks, our F-LMM achieves comparable results despite not having the parameters of LMMs fine-tuned for grounding purposes. + +# 4.3. Complex Scenarios + +In this section, we evaluate grounding LMMs under more complex scenarios that typically require the LMMs to perform both reasoning and segmentation. The base LMM we use is DeepSeekVL-7B [44], considering its flexibility in supporting both single and multiple image inputs. And the size of models compared in this section is also 7B. + +Reasoning Segmentation is proposed in LISA [28] that requires a model to infer what object to segment from common-sense knowledge or via logical reasoning. The evaluation results are provided in Table 2 and the metric is gIoU. F-LMM can effectively perform reasoning segmentation even though it is not trained on such type of data. It is remarkable that F-LMM significantly outperforms existing models on the subset of long sentences, reflecting the advantage of F-LMM in handling complex contexts. + +Grounded Conversation Generation (GCG) is proposed in GLaMM [60], which requires a model to generate interleaved segmentation masks and texts. For performance evaluation, METEOR (M.) and mIoU are used to measure the quality of generated texts and masks, respectively. In addition, we report the recall of object masks. As shown in Table 3, F-LMM exhibits the best zero-shot performance while being comparable with models fine-tuned on GCG + +Table 5. Ablation study of the mask refiner on PNG benchmark. + +
(a) Effects of Mask Refiner(b) SAM Variants
Mask RefinerPNGSAMPNG
maskboxtextAllThingStuffAllThingStuff
XXX50.848.655.9ViT-B63.061.466.8
✓XX63.462.066.8ViT-L64.963.468.3
✓✓X63.762.267.1ViT-H65.063.568.3
✓✓✓64.963.468.3
+ +Table 6. Ablation study of keyword selection. + +
MethodF1RecallPrecision
SpaCy Parser [20]57.897.341.1
Linear Keyword Selector82.896.672.5
+ +training dataset. + +Visual Chain-of-Thought Reasoning. In human-AI conversations that involve Visual Chain-of-Thought Reasoning (VisCoT) [63], an LMM first localises the region/object relevant to the human's question and then generates the final answer by zooming in on the question-related region. Here we evaluate F-LMM and VisCoT-7B on the VisCoT benchmark [63]. As shown in Table 4, F-LMM achieves remarkable performance gains when prompted in the VisCoT manner. It is noticeable that F-LMM even outperforms VisCoT-7B [63] that has been well-tuned on the VisCoT training data [63]. Furthermore, we perform VisCoT on the object hallucination benchmark POPE [33] and observe significant performance gain in resisting object hallucinations. + +# 4.4. Ablation Study + +We investigate the effects of design choices of F-LMM. All the ablation studies are conducted on the PNG dataset and we use DeepSeekVL-1.3B [44] unless otherwise stated. + +Mask Decoder. We summarise our analyses of the mask decoder in Figure 5. Note that the mask refiner is not involved in this part. We first consider attention maps from different transformer layers, i.e., early $(6^{\mathrm{th}})$ , mid $(12^{\mathrm{th}})$ and late $(24^{\mathrm{th}})$ layers. As shown in Figure 5(a), attention maps from late layers perform the worst, conforming to prior studies [21] indicating that deeper transformer layers tend to focus on abstract concepts instead of visual details. And using attention maps from all layers achieves the best performance. Next, we study how to merge the word-image attention maps of multi-word objects, as shown in Figure 5(b). We find that the average operation outperforms the max operation by a margin of 0.7. In Figure 5(c), we show that normalizing the inputs to the mask decoder provides a 0.5 performance gain. Finally, we experiment with different input sizes for the mask decoder. As shown in Figure 5(d), using $64 \times 64$ yields the best performance-cost trade-off. + +Mask Refiner. We study the effects of the mask refiner on segmentation in Table 5a. The performance of using only the mask decoder is shown in the first row of Table 5a. With the masks fed to the mask refiner, we observe a significant 12.6 performance gain on the PNG benchmark. Adding + +
ModelChat ↓Ground ↓
DeepseekVL-1.3B [44]7.758.33
MGM-2B [34]6.008.33
LLaVA-1.5-7B [38]6.757.83
HPT-Air-6B [65]9.007.16
HPT-Air-1.5-8B [65]6.507.00
MGM-7B [34]5.754.83
DeepseekVL-7B [44]3.754.00
LLaVA-1.6-7B [40]2.753.00
LLaVA-1.6-M-7B [40]3.251.66
MGM-HD-7B [34]3.502.83
+ +![](images/53cbdeaf3b34fa1ac6590ea6b3a3019e18fad366495ab50bbf568e2261c97a23.jpg) +Figure 6. The relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. The left table shows the average ranks of each LMM on question-answering ('Chat') and grounding benchmarks ('Ground'). $\downarrow$ means the lower the better. The dashed line in the right figure is the linear fit of the rank data points, indicating a positive correlation between abilities to chat and ground. + +box and text prompts to the mask refiner further improves the performance by 0.3 and 1.2. Then we experiment with different SAM [26] model variants, i.e., ViT-B(ase), ViT-L(arge) and ViT-H(uge) as mask refiners. As shown in Table 5b, the performance grows with model sizes. For a good trade-off between cost and performance, we select the ViT-L variant of SAM as the default mask refiner. + +Keyword Selector. We analyse the keyword selector, implemented as a linear layer, on the PNG dataset, using the F1 score as the main metric. As shown in Table 6, our keyword selector achieves significantly higher F1 scores than the external SpaCy tool [20]. We also report recall and precision scores. Our keyword selector achieves comparable recall while being much more precise compared to SpaCy, which enumerates all nouns in a sentence. + +# 4.5. Analysis & Visualisation + +Scalability: Does Better Chatting Lead to Better Grounding? We study the relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. Specifically, we examine the correlation between performance on the question-answering and grounding benchmarks. For the ten models reported in Table 1, we calculate their average ranks in each benchmark category. In Figure 6, we plot these ranks as 2D coordinates, i.e., (Chat Rank, Ground Rank), and apply a linear fit to the data points. As indicated by the blue dashed line, frozen LMMs with stronger conversational ability can serve as better backbones for grounding. We also observe that larger LMMs generally excel in both conversation and grounding tasks, and LMMs with larger input resolutions (e.g., LLaVA-1.6 and MGM-HD) can handle both tasks better. + +From Attention Maps to Segmentation Masks. We visualise the word-image attention maps by applying KMeans clustering to the stacked attention maps that are collected from all transformer layers and attention heads. The attention maps of multiple-word objects are merged by element + +![](images/cd312f2ccd62ce4444726e5e51a247ba653854aa36e1ba77d139b90c5b90d27f.jpg) +Caption: There is a person in this picture on the surfing board who is surfing on a ocean tide. There is sky in the background. + +![](images/a60d565c64fcb6f3b191f7a3b4df50b0b63e4782a1aafab0d240b78bd8a56a90.jpg) + +![](images/98cce25cf2d0904342af9c95d8e4fe9dbe72cb8d26a79d1576ae6bf388fd900f.jpg) +Figure 7. Visualisations of KMeans clustering on attention maps and segmentation results from the CNN-based mask decoder and the SAM-based mask refiner. + +![](images/81a5618bbb37dfc7e59121dee86850d9f6d6919423138f242064951398aa79e6.jpg) + +![](images/637f4819a3f7a8728f55e5ea02b3da3d80d0c18e6329df10340492ece7d3dd1a.jpg) + +wise average. As shown in Figure 7, we observe that the pixels of objects are roughly clustered together (top-left). With the CNN-Based mask decoder, the attention weights are mapped to 2D binary masks (top-right), which are then further optimised by the SAM-based mask refiner (bottom-left). The model used is DeepSeekVL-1.3B [44]. + +# 5. Conclusion + +In this work, we have studied the limitation of existing grounding LMMs, i.e., the loss of general world knowledge and instruction-following ability. To address this issue, we make the first attempt to ground fully frozen LMMs, which are already well-trained for user-model conversation, based on the insight that the geometric and spatial cues needed for visual grounding are inherently present within the self-attention mechanism of LMMs. By incorporating a CNN-based mask decoder and a SAM-based mask refiner, we achieve competitive visual grounding performance without sacrificing any conversational abilities of pre-trained LMMs. With this combination of strong conversational and visual grounding capabilities, these LMMs show promise for complex perception and reasoning tasks, such as segmentation with reasoning, grounded conversation generation, and visual chain-of-thought reasoning. + +Acknowledgement. This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-PhD-2023-08-048T), the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In $CVPR$ , pages 1209-1218, 2018. +[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. +[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. +[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. +[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020. +[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems, 34:17864-17875, 2021. + +[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. +[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. +[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. +[14] Zihan Ding, Zi-han Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Xiaolin Wei, and Si Liu. Ppmn: Pixel-phrase matching network for one-stage panoptic narrative grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5537–5546, 2022. +[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. +[16] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021. +[17] Tianyu Guo, Haowei Wang, Yiwei Ma, Jiayi Ji, and Xiaoshuai Sun. Improving panoptic narrative grounding by harnessing semantic relationships and visual confirmation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1985-1993, 2024. +[18] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. +[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. +[20] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. +[21] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024. +[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lam- + +ple, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. +[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. +[24] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. +[25] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. CVPR, 2019. +[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. +[27] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. +[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. +[29] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. +[30] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. +[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. +[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 214-223, 2021. +[33] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. +[34] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023. +[35] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. +[36] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533, 2023. +[37] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. +[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. +[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. +[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. +[42] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024. +[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. +[44] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024. +[45] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. +[46] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. +[47] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611, 2024. +[48] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024. +[49] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European + +Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. +[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. +[51] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. +[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. +[53] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. +[54] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023. +[55] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. +[56] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. +[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. +[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. +[59] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. +[60] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. +[61] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228, 2023. +[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted + +intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. +[63] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. +[64] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240–248. Springer, 2017. +[65] HyperGAI Team. Hpt 1.5 air: Best open-sourced 8b multimodal llm with llama 3, 2024. +[66] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. +[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. +[69] Haowei Wang, Jiayi Ji, Yiyi Zhou, Yongjian Wu, and Xiaoshuai Sun. Towards real-time panoptic narrative grounding by an end-to-end grounding network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2528-2536, 2023. +[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. +[71] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. +[72] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024. +[73] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019. +[74] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. + +[75] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. +[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. +[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024. +[78] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. +[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. +[80] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, and Jianwei Yang. Llava-grounding: Grounded visual chat with large multimodal models, 2023. +[81] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. +[82] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. +[83] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021. +[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. arXiv preprint arXiv:2402.16846, 2024. +[85] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2024. +[86] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. +[87] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. + +[88] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. +[89] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. \ No newline at end of file diff --git a/2025/F-LMM_ Grounding Frozen Large Multimodal Models/images.zip b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a337e8a55e93a7d1ad7e485e93c6926954125b4 --- /dev/null +++ b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dac47b5a5879cb08f2c2cb450132e7089b7e445c41d24bca2c4ec689ec726e5 +size 631836 diff --git a/2025/F-LMM_ Grounding Frozen Large Multimodal Models/layout.json b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..df328e8b1014183614c5a2e0a63ab15a99b2b475 --- /dev/null +++ b/2025/F-LMM_ Grounding Frozen Large Multimodal Models/layout.json @@ -0,0 +1,9409 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 476, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 476, + 120 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 476, + 120 + ], + "type": "text", + "content": "F-LMM: Grounding Frozen Large Multimodal Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 142, + 560, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 142, + 560, + 200 + ], + "spans": [ + { + "bbox": [ + 64, + 142, + 560, + 200 + ], + "type": "text", + "content": "Size " + }, + { + "bbox": [ + 64, + 142, + 560, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{1}" + }, + { + "bbox": [ + 64, + 142, + 560, + 200 + ], + "type": "text", + "content": " Sheng Jin2 Wenwei Zhang3 Lumin " + }, + { + "bbox": [ + 64, + 142, + 560, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu^4}" + }, + { + "bbox": [ + 64, + 142, + 560, + 200 + ], + "type": "text", + "content": " Wentao Liu2,3 Wei Li1 Chen Change Loy' 1 S-Lab, Nanyang Technological University SenseTime Research and Tetras.AI 3 Shanghai AI Laboratory 4 The Chinese University of Hong Kong size001@e.ntu.edu.sg ccloy@ntu.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 207, + 202, + 304 + ], + "blocks": [ + { + "bbox": [ + 58, + 207, + 202, + 304 + ], + "lines": [ + { + "bbox": [ + 58, + 207, + 202, + 304 + ], + "spans": [ + { + "bbox": [ + 58, + 207, + 202, + 304 + ], + "type": "image", + "image_path": "a3e1e1ddd8c03e43d3c0dd350064ba99f869d5854bc1ac492e6c926ed0e144a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 309, + 203, + 338 + ], + "lines": [ + { + "bbox": [ + 59, + 309, + 203, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 309, + 203, + 338 + ], + "type": "text", + "content": "User: Is there a truck in the image? Answer the question using a single word or phrase. \nGLaMM: Sure, the segmentation is [SEG]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 208, + 205, + 421, + 340 + ], + "blocks": [ + { + "bbox": [ + 208, + 205, + 421, + 340 + ], + "lines": [ + { + "bbox": [ + 208, + 205, + 421, + 340 + ], + "spans": [ + { + "bbox": [ + 208, + 205, + 421, + 340 + ], + "type": "image", + "image_path": "2ca29c51bc8342b4e04f86ee755c6631376ef687c3e887d215817b7886cfc69a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 354, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 54, + 354, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 555, + 399 + ], + "type": "text", + "content": "Figure 1. An example of user-AI conversation around an image. Left: The current state-of-the-art grounding model GLaMM [60] is effective for grounded conversation when prompted by \"answer with interleaved masks\", but fails to follow user instruction to answer a single word (yes or no) and misunderstands the question as a referring segmentation prompt. Right: Our F-LMM preserves instruction-following ability while being able to perform visual grounding." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 424, + 207, + 551, + 293 + ], + "blocks": [ + { + "bbox": [ + 424, + 207, + 551, + 293 + ], + "lines": [ + { + "bbox": [ + 424, + 207, + 551, + 293 + ], + "spans": [ + { + "bbox": [ + 424, + 207, + 551, + 293 + ], + "type": "image", + "image_path": "12eb4d80a541344329879bb511f1fd89872ab10b0b0c670f1b2d7b5c39d537d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 300, + 546, + 338 + ], + "lines": [ + { + "bbox": [ + 424, + 300, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 424, + 300, + 546, + 338 + ], + "type": "text", + "content": "User: Is there a truck in the image? Answer the question using a single word or phrase. F-LMM: No." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 418, + 200, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 418, + 200, + 430 + ], + "spans": [ + { + "bbox": [ + 152, + 418, + 200, + 430 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 450, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 450, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 297, + 714 + ], + "type": "text", + "content": "Endowing Large Multimodal Models (LMMs) with visual grounding capability can significantly enhance AIs' understanding of the visual world and their interaction with humans. However, existing methods typically fine-tune the parameters of LMMs to learn additional segmentation tokens and overfit grounding and segmentation datasets. Such a design would inevitably cause a catastrophic diminution in the indispensable conversational capability of general AI assistants. In this paper, we comprehensively evaluate state-of-the-art grounding LMMs across a suite of multimodal question-answering benchmarks, observing drastic performance drops that indicate vanishing general knowledge comprehension and weakened instruction following ability. To address this issue, we present F-LMM—grounding frozen off-the-shelf LMMs in human-AI conversations—a straightforward yet effective design based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention mechanism of well-trained LMMs. Using only a few trainable CNN layers, we can translate word-pixel attention weights to mask logits, which a SAM-based mask refiner can further optimise. Our F-LMM neither learns special segmentation to" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": "kens nor utilises high-quality grounded instruction-tuning data, but achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks while completely preserving LMMs' original conversational ability. Additionally, with instruction-following ability preserved and grounding ability obtained, F-LMM can be directly applied to complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning. Our code can be found at https://github.com/wusize/F-LMM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs), which integrate Large Language Models (LLMs) with visual signals, have demonstrated remarkable success in multimodal understanding, reasoning and interaction [34, 38-40, 44, 65, 78]. To further advance LMMs with better perception capability, a recent line of research [28, 60, 61, 70, 80, 84] that visually grounds language contents in user-model conversations has drawn increasing attention. This explicit association between key phrases/words and visual objects greatly enhances LMMs' understanding of the visual world and allows for more intuitive and meaningful human-AI interactions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24710" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 59, + 218, + 121 + ], + "blocks": [ + { + "bbox": [ + 67, + 59, + 218, + 121 + ], + "lines": [ + { + "bbox": [ + 67, + 59, + 218, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 59, + 218, + 121 + ], + "type": "image", + "image_path": "5ccb99f61929303a3bfe0feec5c8d55ed6f07a48f4eedf1504262f92de3cbe7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 122, + 178, + 132 + ], + "lines": [ + { + "bbox": [ + 108, + 122, + 178, + 132 + ], + "spans": [ + { + "bbox": [ + 108, + 122, + 178, + 132 + ], + "type": "text", + "content": "\"two girls\" \"a fence\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 59, + 381, + 121 + ], + "blocks": [ + { + "bbox": [ + 230, + 59, + 381, + 121 + ], + "lines": [ + { + "bbox": [ + 230, + 59, + 381, + 121 + ], + "spans": [ + { + "bbox": [ + 230, + 59, + 381, + 121 + ], + "type": "image", + "image_path": "c3cf1e4093fa2e9ae3275d2a9fc469c196e35cda35230f4852ef30471d0dd714.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 122, + 367, + 133 + ], + "lines": [ + { + "bbox": [ + 243, + 122, + 367, + 133 + ], + "spans": [ + { + "bbox": [ + 243, + 122, + 367, + 133 + ], + "type": "text", + "content": "\"two elephants\" \"grass\" \"hills\" \"sky\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 392, + 59, + 543, + 121 + ], + "blocks": [ + { + "bbox": [ + 392, + 59, + 543, + 121 + ], + "lines": [ + { + "bbox": [ + 392, + 59, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 392, + 59, + 543, + 121 + ], + "type": "image", + "image_path": "4ac40a0fd40619123168822b2cf63c72a0ad923dbcefa9bab400bd792aae79a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 122, + 500, + 133 + ], + "lines": [ + { + "bbox": [ + 436, + 122, + 500, + 133 + ], + "spans": [ + { + "bbox": [ + 436, + 122, + 500, + 133 + ], + "type": "text", + "content": "\"bread\" \"the plate\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 66, + 143, + 288, + 247 + ], + "blocks": [ + { + "bbox": [ + 183, + 133, + 425, + 142 + ], + "lines": [ + { + "bbox": [ + 183, + 133, + 425, + 142 + ], + "spans": [ + { + "bbox": [ + 183, + 133, + 425, + 142 + ], + "type": "text", + "content": "(a) Visualisations of word-image attention maps in Frozen LMMs via KMeans." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 143, + 288, + 247 + ], + "lines": [ + { + "bbox": [ + 66, + 143, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 66, + 143, + 288, + 247 + ], + "type": "image", + "image_path": "d618b2c3430986b89a4d9d51ae5a1caa9d714006f84a572b0f4cee7ad8f10634.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 250, + 230, + 259 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 230, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 230, + 259 + ], + "type": "text", + "content": "(b) Existing Grounding LMMs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 271, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 54, + 271, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 555, + 316 + ], + "type": "text", + "content": "Figure 2. (a) Geometric and spatial cues conducive to visual grounding are observed in the visualisations of word-image attention maps in frozen LMMs. (b) Existing grounding LMMs are fine-tuned to generate a special mask token (e.g., [SEG]) for visual grounding purposes, which ruins the original conversational ability. (c) Our F-LMM translates word-image attention maps from frozen LMMs to grounding masks, while fully preserving the general-purpose chat capability." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 144, + 545, + 244 + ], + "blocks": [ + { + "bbox": [ + 323, + 144, + 545, + 244 + ], + "lines": [ + { + "bbox": [ + 323, + 144, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 323, + 144, + 545, + 244 + ], + "type": "image", + "image_path": "eb98c5d08c498d3122b05491d35a0e3d19d347881dd2424529f021dc64b7dfc9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 249, + 468, + 259 + ], + "lines": [ + { + "bbox": [ + 408, + 249, + 468, + 259 + ], + "spans": [ + { + "bbox": [ + 408, + 249, + 468, + 259 + ], + "type": "text", + "content": "(c) F-LMM (Ours)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 331, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 643 + ], + "type": "text", + "content": "By design, one commonly adopted build (Figure 2(b)) for visually grounding language contents is connecting LMMs with a mask head (e.g., Segment Anything Model (SAM) [26]), wherein both the LLM backbone and the mask head are fine-tuned with well-prepared visual grounding data that contains segmentation annotations. Also, some additional learnable tokens (e.g., [SEG]) are introduced to the LMMs' vocabulary, to directly associate key phrases or words with visual objects in conversations. However, this design will inevitably provoke a catastrophic diminution in general knowledge comprehension and instruction-following ability due to the following reasons. First, existing segmentation and visual grounding data only contain elementary patterns for answering simple grounding prompts. Second, during the fine-tuning stage, the LMMs are mainly optimised for effectively modelling the relationship between key phrases or words and special segmentation tokens, i.e., overfitting the segmentation and grounding data. Therefore, the conversational ability is sacrificed. For instance, the state-of-the-art grounding model GLaMM [60] fails to answer a simple yes-or-no question (Figure 1). Moreover, quantitative evaluations of existing grounding LMMs in conversational ability are presented in Table 1, with zero or near-zero scores on general multimodal question-answering benchmarks necessitating instruction-following ability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "One possible option to deal with this dilemma is to collect high-quality training data encompassing both meaningful conversations and mask annotations. For example, LLaVA-G [80] annotates the 150k LLaVA-Instruct data samples [39] with segmentation masks so that the LMMs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 331, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 426 + ], + "type": "text", + "content": "simultaneously learn to chat and segment. Nonetheless, annotating high-quality grounded conversation data is costly and hard to scale. Despite being trained on costly annotated data, LLaVA-G still lags behind general-purpose LMMs on multimodal understanding tasks. Furthermore, training on large-scale annotated data normally consumes significant computational resources, which is, obviously not a resource-efficient solution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 426, + 556, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 679 + ], + "type": "text", + "content": "In this paper, we propose a simple yet effective design, i.e., grounding frozen LMMs (dubbed as F-LMM) in human-AI conversations. We argue that freezing the parameters of well-trained LMMs is the most practical design choice for fully preserving the original excellent conversational ability when building general-purpose grounding LMMs. In particular, we take inspiration from the built-in interpretability of the attention mechanism in transformers [6, 68] that represents interrelations between text and image contents in design. We observe that off-the-shelf LMMs already produce word-pixel correspondences necessary for visual grounding, despite they were not explicitly pre-trained with region or pixel annotations. As illustrated in Figure 2(a), we visualise word-image attention maps from frozen LMMs via K-Means clustering, revealing notable geometric and spatial cues of the objects1. For example, coarse visual grounding masks for key phrases (e.g., \"two girls\", \"two elephants\", and \"the plate\") in language sentences emerge from attention maps in LMMs. Therefore, our F-LMM takes these visual-language correspondences as useful segmentation priors for decoding ground" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 556, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 556, + 715 + ], + "type": "text", + "content": "1For better visibility, we perform K-Means clustering on the stack-up of all attention maps collected in a forward pass instead of selecting a single attention map." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24711" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "ing masks, without further tuning the LMMs' weights or learning a special segmentation token to model object locations, as shown in Figure 2(c)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 312 + ], + "type": "text", + "content": "The only trainable part of our F-LMM is a mask head plus a keyword selector. The mask head comprises a CNN-based mask decoder (a tiny U-Net [62]) that translates stacked attention maps to mask logits and a lightweight mask refiner (retrofitted from SAM [26]'s mask head) that uses additional image and language cues to refine the semantic-agnostic masks from the mask decoder. The keyword selector is a linear layer that discovers object nouns in text sequences, automating the process of grounded human-AI conversation. Moreover, we only use the RefCOCO " + }, + { + "bbox": [ + 56, + 108, + 295, + 312 + ], + "type": "inline_equation", + "content": "(+ / g)" + }, + { + "bbox": [ + 56, + 108, + 295, + 312 + ], + "type": "text", + "content": " [24, 46] and PNG [16] datasets as our training data, enabling LMMs to segment user-described objects and ground key phrases or words in a text sequence. Unlike previous works [60, 61, 80], our F-LMM eliminates the necessity for high-quality conversation data that are annotated with masks to preserve conversational ability when learning grounding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 443 + ], + "type": "text", + "content": "Our experiments demonstrate that F-LMM maintains the original excellence of off-the-shelf LMMs on general question-answering benchmarks, while achieving competitive results on referring segmentation and phrase grounding. In more complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning, F-LMM achieves better or comparable results when contrasted with models specially trained for such tasks. Compared with existing grounding LMMs, F-LMM offers the best balance between grounding and chat capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 454, + 142, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 142, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 142, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 475, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 713 + ], + "type": "text", + "content": "Large Multimodal Models. Recent advancements in LMMs [2, 3, 13, 29-31, 34, 36, 38-40, 44, 47, 65, 75] have been fueled by the success of LLMs [1, 4, 12, 22, 48, 51, 66, 67, 81] since the debut of GPT series [1, 4, 56, 57] that feature an auto-regressive framework based on transformer decoders [68]. These LLMs possess general world knowledge and excellent conversational ability to follow human instructions, thanks to large-scale generative pre-training [4] and supervised finetuning on instruction-tuning data [71] or human feedback [50]. By integrating image representations from vision encoders [58, 79] to LLMs, LMMs enable visual understanding and reasoning in AI assistants. This integration is usually established by a multilayer perceptron (MLP) that directly maps image features to the LLMs' input embedding space [34, 38-40, 44, 65, 78] or a cross-attention module that abstracts the image contents with a set of query embeddings [2, 3, 31, 75]. In our research, we build F-LMM on LMMs of the former type (MLP-based), which preserves images' 2-D topological structure in the cross-modal integration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "Visual Segmentation. The task of predicting 2D masks for visual objects is known as image segmentation, which can be categorised into semantic segmentation [5, 8, 10, 87], instance segmentation [11, 19, 83] and panoptic segmentation [9, 25, 32, 73] depending on whether the goal is to differentiate pixel semantics or object instances. These standard segmentation approaches rely on a pre-defined set of object classes for recognition. In contrast, referring expression segmentation (RES) [24, 37, 45, 46, 49, 74, 88] involves segmenting objects based on free-form human language descriptions, allowing for enhanced human-model interaction. Additionally, panoptic narrative grounding (PNG) [14, 16, 17, 69] requires segmenting masks for key phrases or words in a sentence. In this study, we mainly leverage RES and PNG tasks to evaluate the grounding capability of LMMs. Besides, we also test LMMs' segmentation ability in complex scenarios that necessitate reasoning [28, 60, 63]. Moreover, the prompt-based SAM [26] pre-trained on billion-scale high-quality mask data has become a constituent component in many grounding LMMs to boost segmentation performance. We also adopt SAM's mask head to initialise our mask refiner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 335, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 598 + ], + "type": "text", + "content": "Grounding Large Multimodal Models. Grounding Large Multimodal Models [3, 7, 28, 53, 54, 60, 61, 70, 72, 76, 78, 80, 84, 85] can localise language contents during user-model conversations. Some approaches [3, 7, 53, 76] represent coordinates of bounding boxes as texts and train LMMs to predict the coordinates in a generative manner. Several recent works [28, 54, 60, 61, 70, 80] train LMMs to predict a special segmentation token for encoding the grounded object and utilise a segmentation head (e.g., SAM [26]) to decode object masks. This study mainly focuses on grounding LMMs with segmentation ability for visual perception. To obtain competitive visual grounding performance, existing works extensively fine-tune the parameters of LMMs on a large amount of segmentation [5, 18, 25, 59, 87] and grounding [16, 23, 24, 27, 46, 55] datasets. And to balance the LMMs' grounding and conversational abilities, there are efforts [61, 80, 84] to collect high-quality instruction-tuning data annotated with segmentation masks. In contrast, we make the first attempt to build grounding LMMs on top of off-the-shelf LMMs without fine-tuning their parameters. Furthermore, we bypass the need for grounded instruction-tuning data to preserve decent chat ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 609, + 370, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 370, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 370, + 621 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "In this section, we introduce our F-LMM by first probing the causal attention mechanism in LMMs with visualisations of word-image attention maps in Sec 3.1. Then, we elaborate on F-LMM exploiting segmentation priors from frozen LMMs for visual grounding using the mask head in Sec 3.2. Finally, we show how to automate the process of grounded conversation with a linear keyword selector to in-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24712" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 73, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 67, + 73, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 67, + 73, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 545, + 216 + ], + "type": "image", + "image_path": "bdd35d9e9487d868b1abd530b35df37f4c3c27f477c1e56b89ec8768f052fa50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. The overall pipeline of F-LMM. The word-image attention maps from the frozen LLM serve as segmentation priors for the mask head. The keyword selector discovers object nouns in the text sequence. The mask head encompasses a mask decoder that translates attention weights to mask logits and a mask refiner that optimises the mask decoder's predictions. " + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": " represent the numbers of transformer layers and attention heads." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 284, + 290, + 407 + ], + "blocks": [ + { + "bbox": [ + 62, + 284, + 290, + 407 + ], + "lines": [ + { + "bbox": [ + 62, + 284, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 290, + 407 + ], + "type": "image", + "image_path": "dace6dd8828545b4d0e50fc39618584f53839d77dc475e4854ea7a1a53622d83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "text", + "content": "Figure 4. Visualisations of word-image attention maps. The letters " + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "text", + "content": " indicate that the attention map is derived from the " + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "text", + "content": "-th attention head of the " + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 415, + 295, + 448 + ], + "type": "text", + "content": "-th transformer layer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "text", + "content": "dicate words of grounding targets in Sec 3.3. The overall pipeline is illustrated in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 500, + 266, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 266, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 266, + 513 + ], + "type": "text", + "content": "3.1. Segmentation Priors from Frozen LMM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": "Vision-Language Sequence. A typical build of a LMM " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": " comprises an image encoder " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": " (e.g., CLIP [58] " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": "), a vision-language projector " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": ", and a LLM, denoted as " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{llm}}" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": ". The inputs to an LMM are usually an image " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": " and the associated text " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": ". The input image is first encoded by the vision encoder " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": " and then mapped to the input space of the LLM by the projector " + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 517, + 295, + 602 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 609, + 260, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 609, + 260, + 624 + ], + "spans": [ + { + "bbox": [ + 90, + 609, + 260, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {v} = f _ {p} (\\text {F l a t t e n} (f _ {v} (\\mathbf {X} _ {v}))) \\in \\mathbb {R} ^ {h w \\times d},", + "image_path": "0355f36da5b637cb8a062f5bccf307938e1caa814712809b1b7c2e11b7cd3d68.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "text", + "content": " are the height and width of projected feature maps via " + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "text", + "content": ". The Flatten operation unfolds the 2-D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 283, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 331 + ], + "type": "text", + "content": "image feature map to a 1-D sequence. The constant " + }, + { + "bbox": [ + 313, + 283, + 555, + 331 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 283, + 555, + 331 + ], + "type": "text", + "content": " is the hidden state dimension of the LLM. Likewise, the text input is first encoded as discrete tokens and then mapped to text embeddings:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 341, + 520, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 341, + 520, + 355 + ], + "spans": [ + { + "bbox": [ + 348, + 341, + 520, + 355 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {t} = \\operatorname {E m b e d} (\\operatorname {T o k e n i z e} (\\mathbf {X} _ {t})) \\in \\mathbb {R} ^ {L \\times d},", + "image_path": "19330783b4819f2599bca445c16e48c45aa3e8e217864b6c64e08407951a2de0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "text", + "content": " denotes the length of text embeddings. The visual-language sequence input to the LLM is a concatenation of image and text embeddings: " + }, + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\{\\mathbf{Z}_v,\\mathbf{Z}_t\\} \\in \\mathbb{R}^{(hw + L)\\times d}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "content": "Segmentation Priors in Self-Attention. The vision-language sequence is mainly processed by causal self-attentions [56, 68] in the LLM, including inner product and weighted-sum operations. Specifically, for a word token with position index " + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "content": " in the vision-language sequence " + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "content": ", its embedding " + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "inline_equation", + "content": "z^i" + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "content": " is updated by the weighted sum of the first " + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "content": " embeddings: " + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "inline_equation", + "content": "\\hat{z}^i = \\mathrm{SoftMax}\\left(\\frac{z^i \\cdot \\mathbf{Z}[i]}{d}\\right) \\cdot \\mathbf{Z}[i]" + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{SoftMax}\\left(\\frac{z^i \\cdot \\mathbf{Z}[i]}{d}\\right)" + }, + { + "bbox": [ + 313, + 402, + 554, + 515 + ], + "type": "text", + "content": " is the attention weights. Here, we omit the residual layers and feedforward layers for brevity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 515, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 554, + 564 + ], + "type": "text", + "content": "Considering the word-image interaction in the multimodal scenario, we can select the word token's attention weights with the image embeddings from the overall vision-language attention weights:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 573, + 555, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 555, + 597 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} ^ {i} = \\text {U n f l a t t e n} (\\operatorname {S o f t M a x} (\\frac {\\boldsymbol {z} ^ {i} \\cdot \\mathbf {Z} [: i ]}{d}) [: h w ]) \\in \\mathbb {R} ^ {h \\times w},", + "image_path": "916e9c632cf6ac5d4761795189922d3cd7f329b125c63b453453c63dde3a9acd.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "where Un flatten restores the 2-D spatial structure from the 1-D sequence to form an attention map. In Figure 4, we visualise such word-image attention maps from various transformers layers and attention heads in an LMM (i.e., DeepSeekVL-7B [44]). The objects' shape and location can be observed in word-image attention maps of certain layers or heads. The visibility is further enhanced when we stack the attention maps from all layers and heads and perform K-Means clustering. It can be observed that the attention" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 663, + 295, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 295, + 683 + ], + "type": "text", + "content": "2In this paper, the term 'multimodal' stands for vision and language modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 683, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 712 + ], + "type": "text", + "content": "3The image encoder might be any vision model that is pre-trained on image-text pairs. We use the classic term 'CLIP' in this paper to represent all such models for brevity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24713" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 557, + 313 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. The main evaluation results on question-answering benchmarks, referring expression segmentation (RES) benchmark and panoptic narrative grounding (PNG) benchmark. LLaVAW: LLaVA-In-the-Wild. LLaVA-1.6 and MGM-HD take high-resolution image inputs. LLaVA-1.6-M-7B means the model is based on Mistral-7B [22]. GLaMM-FS-7B means we use the 'FullScope' version of GLaMM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 557, + 313 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 557, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 557, + 313 + ], + "type": "table", + "html": "
ModelMultimodal Question AnsweringRESPNG
MMEMMBenchMMVetLLaVAWRefCOCORefCOCO+RefCOCOgAllThingStuff
PixelLM-7B [61]309/13517.415.946.473.066.369.343.141.047.9
LISA-7B [28]1/10.419.147.574.965.167.9---
PerceptionGPT-7B [54]----75.168.570.3---
LLaVA-G-7B [80]---55.877.168.871.5---
GroundHog-7B [84]----78.570.574.166.865.069.4
GLaMM-FS-7B [60]14/936.810.332.078.670.574.855.852.962.3
GSVA-7B [72]446/1817.819.438.377.768.273.241.839.646.6
LaSagnA-7B [70]0/00.016.734.576.866.470.6---
F-LMM (DeepSeekVL-1.3B [44])1307/22564.634.851.175.062.868.264.963.468.3
F-LMM (MGM-2B [34])1341/31259.831.165.975.063.767.365.664.468.4
F-LMM (LLaVA-1.5-7B [38])1511/34864.330.569.075.263.767.164.863.468.2
F-LMM (HPT-Air-6B [65])1010/ 25869.831.359.274.364.067.565.564.068.8
F-LMM (HPT-Air-1.5-8B [65])1476/30875.236.362.176.364.568.565.464.168.5
F-LMM (MGM-7B [34])1523/31669.340.875.875.764.868.366.365.368.6
F-LMM (DeepSeekVL-7B [44])1468/29873.241.577.876.166.470.165.764.568.5
F-LMM (LLaVA-1.6-7B [40])1519/32268.144.172.375.865.870.166.365.169.0
F-LMM (LLaVA-1.6-M-7B [40])1501/32469.547.871.775.766.570.166.565.469.1
F-LMM (MGM-HD-7B [34])1546/31965.841.374.076.165.268.566.765.669.1
", + "image_path": "5c1eb9ea76ab325d12f813e6d4e8be6107e39960f57cc897c056858a644d76b4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 333, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 295, + 357 + ], + "type": "text", + "content": "maps offer meaningful segmentation priors with spatial and geometric cues for grounding objects visually." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 418 + ], + "type": "text", + "content": "Language Cues. In addition to the spatial and geometric cues from word-image attention maps, F-LMM can also capitalise on the object's corresponding text embeddings from the LLM " + }, + { + "bbox": [ + 55, + 358, + 296, + 418 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{llm}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 418 + ], + "type": "text", + "content": ", which provide extra language cues for the grounding of visual objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 242, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 242, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 242, + 443 + ], + "type": "text", + "content": "3.2. Visual Grounding with Mask Head" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 484 + ], + "type": "text", + "content": "We use the segmentation priors from the frozen LMM for pixel-level grounding, with the help of a mask head consisting of a mask decoder and a mask refiner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "Mask Decoder. The mask decoder " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " is a 2-D CNN model that transforms the word-image attention maps of grounded objects into mask logits, which is instantiated by a 3-stage U-Net [62]. Please refer to the supplemental material for details on the mask decoder. The extraction of word-image attention map " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{a}^{i}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " for a word token with position index " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " is illustrated in Eq. 3.1 and Figure 3. For an object described by multiple words, we merge their corresponding word-image attention maps to a single attention map " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " via element-wise average or max operation. The attention map " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " is further normalised as " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{a} / \\mathrm{sum}(\\pmb{a})" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " so that all elements sum to 1. Considering " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " layers and " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " attention heads, we stack the " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "MN" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " attention maps as " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{A} \\in \\mathbb{R}^{MN \\times h \\times w}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": ", which forms the input to a mask decoder. Given the importance of high input resolution for segmentation models, we upsample the stacked attention maps " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{A}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "h' \\times w'" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " by bilinear interpolation before feeding it to a mask decoder, where " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "h' > h" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "w' > w" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": ". In practice, we set " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "h' = w' = 64" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": ". Then, the mask decoder maps " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{A}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " into mask" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "content": "logits: " + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{logits}} = f_d(\\mathbf{A})" + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "content": ". We derive the corresponding binary mask via " + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{pred}} = \\mathbf{M}_{\\mathrm{logits}} > 0" + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "content": ". During training, the mask decoder is optimised with BCE and DICE losses [64]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": "Mask Refiner. The mask refiner " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": " is retrofitted from the mask head of SAM [26], which predicts masks based on prompts as well as image embeddings from SAM's ViT-based image encoder. To refine the output of the mask decoder " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": ", we re-use SAM's prompt encoder to transform " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{logits}}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": " into dense prompt embeddings (i.e., a 2-D feature map) and the bounding box of " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": " to box embeddings. In addition to the location cues from the mask and the box, the language cues, i.e., the object's corresponding text embeddings, are also utilised by " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": ". Considering text embeddings from the " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": " transformer layers, we train " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": " learnable scalars to calculate a weighted sum of these text embeddings. The weighted-summed text embeddings are processed by a linear layer and then concatenated with the box embeddings to form sparse prompt embeddings. The dense and sparse prompt embeddings, together with SAM's image embeddings, are fed to the mask refiner " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": " for finer-grained mask predictions " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{pred}}^{\\prime}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": ". During training, we keep the ViT-based image encoder of SAM frozen and optimise the mask refiner " + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": " using BCE loss and DICE loss [64]. For more details on the SAM's prompt-based mask head, please refer to the original SAM paper [26]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 647, + 553, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 553, + 659 + ], + "type": "text", + "content": "3.3. Keyword Selector for Grounded Conversation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "In grounded conversation [60] with interleaved segmentation masks and words, existing grounding LMMs [60, 61] expand LLMs' vocabularies with special tokens that indicate the start and end of grounding targets, which is infeas-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24714" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 85, + 255, + 168 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 220, + 82 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 220, + 82 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 220, + 82 + ], + "type": "text", + "content": "Table 2. Reasoning Segmentation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 85, + 255, + 168 + ], + "lines": [ + { + "bbox": [ + 75, + 85, + 255, + 168 + ], + "spans": [ + { + "bbox": [ + 75, + 85, + 255, + 168 + ], + "type": "table", + "html": "
ModelVal -ShortTest LongAll
X-Decoder [88]22.620.422.221.7
SEEM [89]25.520.125.624.3
GroundingSAM [41]26.017.822.421.3
OVSeg [35]28.518.028.726.1
LISA [28]44.437.636.636.8
F-LMM46.736.949.146.2
", + "image_path": "ef6829c46d797781aec4481e13d2c870c390d736b2de8c01d1536c0def5404ee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 264, + 85, + 536, + 168 + ], + "blocks": [ + { + "bbox": [ + 261, + 71, + 542, + 81 + ], + "lines": [ + { + "bbox": [ + 261, + 71, + 542, + 81 + ], + "spans": [ + { + "bbox": [ + 261, + 71, + 542, + 81 + ], + "type": "text", + "content": "Table 3. Grounded Conversation Generation (GCG). M. stands for METEOR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 264, + 85, + 536, + 168 + ], + "lines": [ + { + "bbox": [ + 264, + 85, + 536, + 168 + ], + "spans": [ + { + "bbox": [ + 264, + 85, + 536, + 168 + ], + "type": "table", + "html": "
ModelGCG TrainingM.Val mIoURecallM.Test mIoURecall
LISA [28]✓13.062.036.312.961.735.5
OMG-LLaVA [82]✓14.965.5-14.564.7-
GLaMM [60]✓16.266.341.815.865.640.8
BuboGPT [86]✗17.254.029.417.154.127.0
KOSMOS-2 [52]✗16.155.628.315.856.829.0
F-LMM✗17.663.542.017.463.638.6
", + "image_path": "365061a1da6fe95af870719a3f5099826e7aedec19d93abce4c4010f6474cc66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 96, + 190, + 511, + 251 + ], + "blocks": [ + { + "bbox": [ + 89, + 175, + 518, + 186 + ], + "lines": [ + { + "bbox": [ + 89, + 175, + 518, + 186 + ], + "spans": [ + { + "bbox": [ + 89, + 175, + 518, + 186 + ], + "type": "text", + "content": "Table 4. Unleashing visual chain-of-thought reasoning with both excellent grounding and instruction-following ability." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 190, + 511, + 251 + ], + "lines": [ + { + "bbox": [ + 96, + 190, + 511, + 251 + ], + "spans": [ + { + "bbox": [ + 96, + 190, + 511, + 251 + ], + "type": "table", + "html": "
ModelVisual CoTDocVQATextCapsVisCoT BenchmarkPOPE
TextVQADUDESROIEInfographicsAccF1
VisCoT-7B [63]✓47.667.577.538.647.032.486.5-
F-LMM✗43.263.574.532.028.443.287.086.0
F-LMM✓53.867.978.442.344.149.188.087.7
", + "image_path": "a2dfde106f92c4d7e31e5eb409930beeda159ade71bc3e8742b68768ba92c315.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 270, + 294, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 294, + 365 + ], + "type": "text", + "content": "ble in F-LMM given its ‘frozen’ nature. A common practice to discover visual objects in text sequences offline is using external tools such as SpaCy [20], which parse all nouns from a sentence including unwanted non-object words. Instead of adopting such offline tools that produce noisy results, we automate generating interleaved words and masks by training a linear layer to directly predict whether a word is to be grounded or not." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "content": "Specifically, the linear layer (keyword selector) is placed on top of the LLM's transformer layers, projecting " + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "content": "-dimension hidden state vectors into one-dimension scores, followed by a sigmoid function that normalizes the scores to " + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "content": ". During training, the score prediction is supervised by a BCE loss. During inference, word tokens with scores exceeding a threshold " + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "content": " are regarded as positive for visual grounding. In practice, we set " + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "content": ". Adjacent positive word tokens are grouped to indicate a single visual object. After the word tokens of visual objects are selected, the corresponding attention maps and text embeddings are fed to the mask head for visual grounding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 137, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 137, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 137, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 540, + 188, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 188, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 188, + 553 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": "Model Architectures. We build F-LMM on several open-sourced LMMs, including LLaVA-1.5 [38], LLaVA-Next [40], MiniGemini [34], DeepSeekVL [44] and HPT-Air [65]. The main experiment covers 10 LMMs with model sizes ranging from 1.3B to 8B. We employ a lightweight 3-stage U-Net [62] as the mask decoder to transform segmentation priors from frozen LLMs. The U-Net architecture features an encoder-decoder structure with skip connections, wherein feature maps are downsampled in the encoder and upsampled in the decoder. Please check the supplementary material for more details on the mask decoder. As for the SAM-based mask refiner, we choose SAM ViT-L [26] that balances cost and performance well." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 378 + ], + "type": "text", + "content": "Model Training. We train F-LMM on RefCOCO(+/g) [24, 46] and PNG [16] datasets with about 190k data samples on a single machine with 8 NVIDIA A800-40G GPUs, which costs about 20 hours for each round of model training. We set the batch size to 8 and train models for 8 epochs, with gradient clipping at a max norm of 1.0. The AdamW [43] optimiser is used with a learning rate of 1e-4, a weight decay of 0.01, and betas as (0.9, 0.999). We choose a warm-up ratio of 0.03 to stabilise model optimisation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 385, + 504, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 504, + 398 + ], + "type": "text", + "content": "4.2. Standard QA and Grounding Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 473 + ], + "type": "text", + "content": "For a comprehensive study of LMMs' conversational and grounding capabilities, we first evaluate models under standard question-answering and grounding benchmarks separately. We summarise the evaluation results of grounding LMMs in Table 1. Please refer to the supplementary material for more detailed results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "type": "text", + "content": "Benchmarks. For comprehensive conversational ability evaluation, we choose four widely used general question-answering benchmarks including MME [15], MMBench [42], LLaVA-In-the-Wild [39] and MMVet [77]. The MME and MMBench require an LMM to strictly follow the instruction to reply with single words (yes or no) or answer MCQs with alphabetical letters (i.e., answering A, B, C, or D). The LLaVA-In-the-Wild and MMVet benchmarks ask a model to respond with open-ended texts while demanding general world knowledge comprehension. In terms of grounding ability evaluation, we assess the LMMs' ability to segment user-described objects on referring expression segmentation (RES) [24, 46] benchmarks including RefCOCO, RefCOCO+, and RefCOCOg, using the cIoU metric. Due to limited space, we only report results on the Val splits of RefCOCO(+/g) in Table 1. We also test the LMMs' ability to ground key phrases or words in user-model conversations on the Panoptic Narrative Grounding (PNG) [16] benchmark, measuring individual mask recalls on thing/stuff objects and overall recall scores." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24715" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 294, + 228 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 294, + 228 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 294, + 228 + ], + "type": "image", + "image_path": "0210f968d82afa7d617a0ba78a1b43c3eb27779d9453f770ab632a437e6458b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 236, + 259, + 246 + ], + "lines": [ + { + "bbox": [ + 91, + 236, + 259, + 246 + ], + "spans": [ + { + "bbox": [ + 91, + 236, + 259, + 246 + ], + "type": "text", + "content": "Figure 5. Ablation study of the mask decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 414 + ], + "type": "text", + "content": "Comparisons with Existing Methods. We compare F-LMM with existing grounding LMMs. As shown in Table 1, our F-LMM provides the best balance with conversational and grounding abilities among compared methods. On the question-answering benchmarks, existing grounding LMMs obtain zero or near-zero scores on MMBench and MME while lagging significantly behind general-purpose LMMs on MM Vet and LLaVA-In-the-Wild benchmarks, indicating compromised instruction-following ability and weakened general knowledge comprehension. On the RES and PNG benchmarks, our F-LMM achieves comparable results despite not having the parameters of LMMs fine-tuned for grounding purposes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 168, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 168, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 168, + 433 + ], + "type": "text", + "content": "4.3. Complex Scenarios" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": "In this section, we evaluate grounding LMMs under more complex scenarios that typically require the LMMs to perform both reasoning and segmentation. The base LMM we use is DeepSeekVL-7B [44], considering its flexibility in supporting both single and multiple image inputs. And the size of models compared in this section is also 7B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "content": "Reasoning Segmentation is proposed in LISA [28] that requires a model to infer what object to segment from common-sense knowledge or via logical reasoning. The evaluation results are provided in Table 2 and the metric is gIoU. F-LMM can effectively perform reasoning segmentation even though it is not trained on such type of data. It is remarkable that F-LMM significantly outperforms existing models on the subset of long sentences, reflecting the advantage of F-LMM in handling complex contexts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Grounded Conversation Generation (GCG) is proposed in GLaMM [60], which requires a model to generate interleaved segmentation masks and texts. For performance evaluation, METEOR (M.) and mIoU are used to measure the quality of generated texts and masks, respectively. In addition, we report the recall of object masks. As shown in Table 3, F-LMM exhibits the best zero-shot performance while being comparable with models fine-tuned on GCG" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 83, + 533, + 146 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 550, + 81 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 550, + 81 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 550, + 81 + ], + "type": "text", + "content": "Table 5. Ablation study of the mask refiner on PNG benchmark." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 83, + 533, + 146 + ], + "lines": [ + { + "bbox": [ + 319, + 83, + 533, + 146 + ], + "spans": [ + { + "bbox": [ + 319, + 83, + 533, + 146 + ], + "type": "table", + "html": "
(a) Effects of Mask Refiner(b) SAM Variants
Mask RefinerPNGSAMPNG
maskboxtextAllThingStuffAllThingStuff
XXX50.848.655.9ViT-B63.061.466.8
✓XX63.462.066.8ViT-L64.963.468.3
✓✓X63.762.267.1ViT-H65.063.568.3
✓✓✓64.963.468.3
", + "image_path": "19fde3c27477195ccb4cd6ee35d4fcc89fb22f0a3127bc086e84bfa422adff41.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 345, + 166, + 519, + 200 + ], + "blocks": [ + { + "bbox": [ + 351, + 154, + 517, + 164 + ], + "lines": [ + { + "bbox": [ + 351, + 154, + 517, + 164 + ], + "spans": [ + { + "bbox": [ + 351, + 154, + 517, + 164 + ], + "type": "text", + "content": "Table 6. Ablation study of keyword selection." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 166, + 519, + 200 + ], + "lines": [ + { + "bbox": [ + 345, + 166, + 519, + 200 + ], + "spans": [ + { + "bbox": [ + 345, + 166, + 519, + 200 + ], + "type": "table", + "html": "
MethodF1RecallPrecision
SpaCy Parser [20]57.897.341.1
Linear Keyword Selector82.896.672.5
", + "image_path": "54f491662994f1b30a6d73b7eea9abb6011b4800fcf4e03dfc764a68f9e49e4b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 381, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 381, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 381, + 228 + ], + "type": "text", + "content": "training dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 229, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 385 + ], + "type": "text", + "content": "Visual Chain-of-Thought Reasoning. In human-AI conversations that involve Visual Chain-of-Thought Reasoning (VisCoT) [63], an LMM first localises the region/object relevant to the human's question and then generates the final answer by zooming in on the question-related region. Here we evaluate F-LMM and VisCoT-7B on the VisCoT benchmark [63]. As shown in Table 4, F-LMM achieves remarkable performance gains when prompted in the VisCoT manner. It is noticeable that F-LMM even outperforms VisCoT-7B [63] that has been well-tuned on the VisCoT training data [63]. Furthermore, we perform VisCoT on the object hallucination benchmark POPE [33] and observe significant performance gain in resisting object hallucinations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 395, + 408, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 408, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 408, + 407 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "text", + "content": "We investigate the effects of design choices of F-LMM. All the ablation studies are conducted on the PNG dataset and we use DeepSeekVL-1.3B [44] unless otherwise stated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "text", + "content": "Mask Decoder. We summarise our analyses of the mask decoder in Figure 5. Note that the mask refiner is not involved in this part. We first consider attention maps from different transformer layers, i.e., early " + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "inline_equation", + "content": "(6^{\\mathrm{th}})" + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "text", + "content": ", mid " + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "inline_equation", + "content": "(12^{\\mathrm{th}})" + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "text", + "content": " and late " + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "inline_equation", + "content": "(24^{\\mathrm{th}})" + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "text", + "content": " layers. As shown in Figure 5(a), attention maps from late layers perform the worst, conforming to prior studies [21] indicating that deeper transformer layers tend to focus on abstract concepts instead of visual details. And using attention maps from all layers achieves the best performance. Next, we study how to merge the word-image attention maps of multi-word objects, as shown in Figure 5(b). We find that the average operation outperforms the max operation by a margin of 0.7. In Figure 5(c), we show that normalizing the inputs to the mask decoder provides a 0.5 performance gain. Finally, we experiment with different input sizes for the mask decoder. As shown in Figure 5(d), using " + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 449, + 554, + 653 + ], + "type": "text", + "content": " yields the best performance-cost trade-off." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Mask Refiner. We study the effects of the mask refiner on segmentation in Table 5a. The performance of using only the mask decoder is shown in the first row of Table 5a. With the masks fed to the mask refiner, we observe a significant 12.6 performance gain on the PNG benchmark. Adding" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24716" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 81, + 260, + 213 + ], + "blocks": [ + { + "bbox": [ + 77, + 81, + 260, + 213 + ], + "lines": [ + { + "bbox": [ + 77, + 81, + 260, + 213 + ], + "spans": [ + { + "bbox": [ + 77, + 81, + 260, + 213 + ], + "type": "table", + "html": "
ModelChat ↓Ground ↓
DeepseekVL-1.3B [44]7.758.33
MGM-2B [34]6.008.33
LLaVA-1.5-7B [38]6.757.83
HPT-Air-6B [65]9.007.16
HPT-Air-1.5-8B [65]6.507.00
MGM-7B [34]5.754.83
DeepseekVL-7B [44]3.754.00
LLaVA-1.6-7B [40]2.753.00
LLaVA-1.6-M-7B [40]3.251.66
MGM-HD-7B [34]3.502.83
", + "image_path": "4be90301f9dd0c9dcba24b227771950867c268898c016fa13265a3b49fbf376c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 287, + 78, + 532, + 218 + ], + "blocks": [ + { + "bbox": [ + 287, + 78, + 532, + 218 + ], + "lines": [ + { + "bbox": [ + 287, + 78, + 532, + 218 + ], + "spans": [ + { + "bbox": [ + 287, + 78, + 532, + 218 + ], + "type": "image", + "image_path": "53cbdeaf3b34fa1ac6590ea6b3a3019e18fad366495ab50bbf568e2261c97a23.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 221, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 257 + ], + "type": "text", + "content": "Figure 6. The relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. The left table shows the average ranks of each LMM on question-answering ('Chat') and grounding benchmarks ('Ground'). " + }, + { + "bbox": [ + 54, + 221, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 54, + 221, + 555, + 257 + ], + "type": "text", + "content": " means the lower the better. The dashed line in the right figure is the linear fit of the rank data points, indicating a positive correlation between abilities to chat and ground." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 271, + 294, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 294, + 354 + ], + "type": "text", + "content": "box and text prompts to the mask refiner further improves the performance by 0.3 and 1.2. Then we experiment with different SAM [26] model variants, i.e., ViT-B(ase), ViT-L(arge) and ViT-H(uge) as mask refiners. As shown in Table 5b, the performance grows with model sizes. For a good trade-off between cost and performance, we select the ViT-L variant of SAM as the default mask refiner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 356, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 295, + 452 + ], + "type": "text", + "content": "Keyword Selector. We analyse the keyword selector, implemented as a linear layer, on the PNG dataset, using the F1 score as the main metric. As shown in Table 6, our keyword selector achieves significantly higher F1 scores than the external SpaCy tool [20]. We also report recall and precision scores. Our keyword selector achieves comparable recall while being much more precise compared to SpaCy, which enumerates all nouns in a sentence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 194, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 194, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 194, + 468 + ], + "type": "text", + "content": "4.5. Analysis & Visualisation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 474, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 653 + ], + "type": "text", + "content": "Scalability: Does Better Chatting Lead to Better Grounding? We study the relevance between an LMM's chat ability and the grounding ability of the F-LMM built on it. Specifically, we examine the correlation between performance on the question-answering and grounding benchmarks. For the ten models reported in Table 1, we calculate their average ranks in each benchmark category. In Figure 6, we plot these ranks as 2D coordinates, i.e., (Chat Rank, Ground Rank), and apply a linear fit to the data points. As indicated by the blue dashed line, frozen LMMs with stronger conversational ability can serve as better backbones for grounding. We also observe that larger LMMs generally excel in both conversation and grounding tasks, and LMMs with larger input resolutions (e.g., LLaVA-1.6 and MGM-HD) can handle both tasks better." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 295, + 714 + ], + "type": "text", + "content": "From Attention Maps to Segmentation Masks. We visualise the word-image attention maps by applying KMeans clustering to the stacked attention maps that are collected from all transformer layers and attention heads. The attention maps of multiple-word objects are merged by element" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 270, + 395, + 323 + ], + "blocks": [ + { + "bbox": [ + 322, + 270, + 395, + 323 + ], + "lines": [ + { + "bbox": [ + 322, + 270, + 395, + 323 + ], + "spans": [ + { + "bbox": [ + 322, + 270, + 395, + 323 + ], + "type": "image", + "image_path": "cd312f2ccd62ce4444726e5e51a247ba653854aa36e1ba77d139b90c5b90d27f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 324, + 394, + 373 + ], + "lines": [ + { + "bbox": [ + 324, + 324, + 394, + 373 + ], + "spans": [ + { + "bbox": [ + 324, + 324, + 394, + 373 + ], + "type": "text", + "content": "Caption: There is a person in this picture on the surfing board who is surfing on a ocean tide. There is sky in the background." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 397, + 270, + 471, + 322 + ], + "blocks": [ + { + "bbox": [ + 397, + 270, + 471, + 322 + ], + "lines": [ + { + "bbox": [ + 397, + 270, + 471, + 322 + ], + "spans": [ + { + "bbox": [ + 397, + 270, + 471, + 322 + ], + "type": "image", + "image_path": "a60d565c64fcb6f3b191f7a3b4df50b0b63e4782a1aafab0d240b78bd8a56a90.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 398, + 323, + 471, + 373 + ], + "blocks": [ + { + "bbox": [ + 398, + 323, + 471, + 373 + ], + "lines": [ + { + "bbox": [ + 398, + 323, + 471, + 373 + ], + "spans": [ + { + "bbox": [ + 398, + 323, + 471, + 373 + ], + "type": "image", + "image_path": "98cce25cf2d0904342af9c95d8e4fe9dbe72cb8d26a79d1576ae6bf388fd900f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 374, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 407 + ], + "type": "text", + "content": "Figure 7. Visualisations of KMeans clustering on attention maps and segmentation results from the CNN-based mask decoder and the SAM-based mask refiner." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 472, + 270, + 545, + 322 + ], + "blocks": [ + { + "bbox": [ + 472, + 270, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 472, + 270, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 472, + 270, + 545, + 322 + ], + "type": "image", + "image_path": "81a5618bbb37dfc7e59121dee86850d9f6d6919423138f242064951398aa79e6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 472, + 323, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 472, + 323, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 472, + 323, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 472, + 323, + 545, + 373 + ], + "type": "image", + "image_path": "637f4819a3f7a8728f55e5ea02b3da3d80d0c18e6329df10340492ece7d3dd1a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 418, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 491 + ], + "type": "text", + "content": "wise average. As shown in Figure 7, we observe that the pixels of objects are roughly clustered together (top-left). With the CNN-Based mask decoder, the attention weights are mapped to 2D binary masks (top-right), which are then further optimised by the SAM-based mask refiner (bottom-left). The model used is DeepSeekVL-1.3B [44]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 500, + 388, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 388, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 388, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 520, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 712 + ], + "type": "text", + "content": "In this work, we have studied the limitation of existing grounding LMMs, i.e., the loss of general world knowledge and instruction-following ability. To address this issue, we make the first attempt to ground fully frozen LMMs, which are already well-trained for user-model conversation, based on the insight that the geometric and spatial cues needed for visual grounding are inherently present within the self-attention mechanism of LMMs. By incorporating a CNN-based mask decoder and a SAM-based mask refiner, we achieve competitive visual grounding performance without sacrificing any conversational abilities of pre-trained LMMs. With this combination of strong conversational and visual grounding capabilities, these LMMs show promise for complex perception and reasoning tasks, such as segmentation with reasoning, grounded conversation generation, and visual chain-of-thought reasoning." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24717" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Acknowledgement. This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-PhD-2023-08-048T), the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 187, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 187, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 294, + 241 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 243, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 294, + 308 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 310, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 294, + 354 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 355, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 355, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 355, + 294, + 409 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "text", + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In " + }, + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "text", + "content": ", pages 1209-1218, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 446, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 294, + 498 + ], + "type": "text", + "content": "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 544 + ], + "type": "text", + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "type": "text", + "content": "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 603, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 668 + ], + "type": "text", + "content": "[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems, 34:17864-17875, 2021." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 317, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 127 + ], + "type": "text", + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 130, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 553, + 194 + ], + "type": "text", + "content": "[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 553, + 251 + ], + "type": "text", + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 253, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 307 + ], + "type": "text", + "content": "[14] Zihan Ding, Zi-han Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Xiaolin Wei, and Si Liu. Ppmn: Pixel-phrase matching network for one-stage panoptic narrative grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5537–5546, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "text", + "content": "[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 418 + ], + "type": "text", + "content": "[16] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[17] Tianyu Guo, Haowei Wang, Yiwei Ma, Jiayi Ji, and Xiaoshuai Sun. Improving panoptic narrative grounding by harnessing semantic relationships and visual confirmation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1985-1993, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "text", + "content": "[18] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "type": "text", + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 569, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 601 + ], + "type": "text", + "content": "[20] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 603, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 678 + ], + "type": "text", + "content": "[21] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lam-" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24718" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "ple, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[24] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 239 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. CVPR, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 286 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 295, + 342 + ], + "type": "text", + "content": "[27] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 386 + ], + "type": "text", + "content": "[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 422 + ], + "type": "text", + "content": "[29] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "text", + "content": "[30] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 578 + ], + "type": "text", + "content": "[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 214-223, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "text", + "content": "[33] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "text", + "content": "[34] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[35] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[36] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[37] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 208 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 231 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 232, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 266 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 266, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 320 + ], + "type": "text", + "content": "[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 322, + 553, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 366 + ], + "type": "text", + "content": "[42] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 367, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 553, + 388 + ], + "type": "text", + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 389, + 553, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 445 + ], + "type": "text", + "content": "[44] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 445, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 511 + ], + "type": "text", + "content": "[45] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "text", + "content": "[46] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "text", + "content": "[47] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[48] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "text", + "content": "[49] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24719" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "text", + "content": "Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "type": "text", + "content": "[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "text", + "content": "[51] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[53] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "text", + "content": "[54] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[55] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[56] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 449 + ], + "type": "text", + "content": "[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "type": "text", + "content": "[59] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "type": "text", + "content": "[60] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "text", + "content": "[61] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "text", + "content": "[63] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 249 + ], + "type": "text", + "content": "[64] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240–248. Springer, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 270 + ], + "type": "text", + "content": "[65] HyperGAI Team. Hpt 1.5 air: Best open-sourced 8b multimodal llm with llama 3, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "type": "text", + "content": "[66] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "text", + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 469 + ], + "type": "text", + "content": "[69] Haowei Wang, Jiayi Ji, Yiyi Zhou, Yongjian Wu, and Xiaoshuai Sun. Towards real-time panoptic narrative grounding by an end-to-end grounding network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2528-2536, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "text", + "content": "[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "text", + "content": "[71] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "type": "text", + "content": "[72] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "text", + "content": "[73] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[74] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24720" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[75] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In International conference on machine learning. PMLR, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[78] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "text", + "content": "[80] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, and Jianwei Yang. Llava-grounding: Grounded visual chat with large multimodal models, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 419 + ], + "type": "text", + "content": "[81] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[82] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[83] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. arXiv preprint arXiv:2402.16846, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "text", + "content": "[85] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "text", + "content": "[86] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[87] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 184 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[88] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "text", + "content": "[89] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24721" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_content_list.json b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..76206037fbaade85a5bdc378097a15b86d48c59a --- /dev/null +++ b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_content_list.json @@ -0,0 +1,1949 @@ +[ + { + "type": "text", + "text": "FADA: Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyun Zhong $^{1*†}$ , Chao Liang $^{2*}$ , Jianwen Jiang $^{2*‡}$ , Gaojie Lin $^{2}$ , Jiaqi Yang $^{2}$ , Zhou Zhao $^{1}$ , Zhejiang University, $^{2}$ Bytedance Intelligent Creation", + "bbox": [ + 135, + 202, + 859, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhongtianyun@zju.edu.cn", + "bbox": [ + 406, + 243, + 591, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liangchao.0412, jianwen.alan, yjq850207131}@gmail.com", + "bbox": [ + 292, + 258, + 705, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://fadavatar.github.io/", + "bbox": [ + 370, + 277, + 622, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 328, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based audio-driven talking avatar methods have recently gained attention for their high-fidelity, vivid, and expressive results. However, their slow inference speed limits practical applications. Despite the development of various distillation techniques for diffusion models, we found that naive diffusion distillation methods do not yield satisfactory results. Distilled models exhibit reduced robustness with open-set input images and a decreased correlation between audio and video compared to teacher models, undermining the advantages of diffusion models. To address this, we propose FADA (Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation). We first designed a mixed-supervised loss to leverage data of varying quality and enhance the overall model capability as well as robustness. Additionally, we propose a multi-CFG distillation with learnable tokens to utilize the correlation between audio and reference image conditions, reducing the threefold inference runs caused by multi-CFG with acceptable quality degradation. Extensive experiments across multiple datasets show that FADA generates vivid videos comparable to recent diffusion model-based methods while achieving an NFE speedup of 4.17-12.5 times. Demos are available at our webpage https://fadavatar.github.io.", + "bbox": [ + 88, + 358, + 485, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 724, + 223, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Talking avatar synthesis aims to animate a given portrait image using driven video or audio. Due to the ease of obtaining audio inputs and their low usage barrier, there has been a surge of recent work in this area. Based on the technical route, talking avatar synthesis methods can be categorized into GAN-based[4, 8, 19, 34, 44, 47, 48], NeRF-based[5, 30, 38-41], and the more recent diffusion", + "bbox": [ + 89, + 750, + 483, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model-based approaches[1, 2, 10, 14, 31, 33, 36]. Diffusion model-based methods, in particular, have gained attention due to their strong generative capabilities. These models exhibit excellent robustness in input images and produce vivid and highly expressive videos that naturally align with the audio. However, the current diffusion model-based talking head generation methods suffer from slow inference speeds, which hinder their practical application. Methods[1, 10, 36] based on Stable Diffusion (SD)[23], all require multiple denoising steps. Moreover, to ensure the correlation between the audio input and the image input in the final video, multi-CFG (Classifier-Free Guidance) inference is often employed, further increasing the runtime. Although most methods achieve 30-40 denoising steps with the help of DDIM[27], they still face speed challenges due to CFG calculations.", + "bbox": [ + 511, + 327, + 906, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several works have explored diffusion distillation in text-to-image [3, 12, 15, 17, 21, 24, 28, 46] and text-to-video tasks [18] to accelerate inference. Nonetheless, there has been no direct exploration of diffusion distillation for talking avatar synthesis tasks. Our observations indicate that simply applying text-to-image and text-to-video diffusion distillation methods results in a noticeable decline in performance. Unlike text to image or video tasks, talking avatar synthesis involves two key conditions: audio and a reference image. Since audio has a relatively weaker influence on video generation[10, 31], the final output is jointly influenced by both conditions, necessitating careful parameter tuning in multi-CFG settings. The sensitivity to condition guidance often leads to severe artifacts when applying diffusion distillation methods with few-step inference.", + "bbox": [ + 511, + 561, + 908, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose FADA (FAst Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation) to address these issues from two perspectives, resulting in an audio-driven talking avatar diffusion model that balances both speed and quality. First, we address the robustness of the distilled model. In this task, the mapping between audio and portrait movement must be learned from data. Previous works have emphasized meticulous data se", + "bbox": [ + 511, + 782, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 107, + 862, + 220, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Done during an internship at ByteDance.", + "bbox": [ + 109, + 875, + 331, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Project Lead", + "bbox": [ + 109, + 887, + 184, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3101", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lection [1, 10, 14], such as removing instances with excessive movement or asynchronous audio, to ensure high audio-portrait movement correlation. This stringent selection process excludes a significant amount of data that could enhance robustness. Considering that a well-trained teacher model already provides high audio-portrait correlation guidance, this discarded data can be utilized in training the student model. We designed a mixed-supervised learning strategy that adaptively adjusts the weights between learning from the teacher model's guidance and the student model's own learning from previously discarded data. Secondly, we address the speed issues of multimodal conditions in the distillation process. In audio-driven methods, the final output is achieved through composite CFG inference [10, 14, 36]. To make this process transparent to the student model during training, we introduce a set of learnable tokens. These tokens mimic the weighted calculations of multi-CFG with its coefficients and are injected into the network as control signals, enabling the student model to better learn the relationships between multiple conditions, including those present during inference. Since the learnable token conditions replicate the multi-CFG inference calculation process, we can reduce the number of multi-CFG model inferences during actual inference, further decreasing inference time on top of the denoising step distillation.", + "bbox": [ + 93, + 90, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above considerations and designed methods, we implemented FADA, a diffusion distillation method for audio-driven talking head generation, achieving up to a $12.5 \\times$ speed-up while generating vivid portrait videos comparable to recent undistilled diffusion models. In summary, the contributions of this paper include:", + "bbox": [ + 93, + 457, + 480, + 545 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose FADA, the first diffusion-based distillation framework for audio-driven talking avatar tasks. It includes a mixed-supervised loss to learn from data of varying quality and learnable token conditions to mimic the multi-CFG inference process, reducing performance loss by distillation and multi-CFG inferences.", + "2. Quantitative and qualitative experiments demonstrate that FADA matches state-of-the-art generation quality while achieving significantly higher efficiency." + ], + "bbox": [ + 91, + 547, + 482, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 93, + 693, + 238, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Audio-Driven Talking Avatar", + "text_level": 1, + "bbox": [ + 93, + 717, + 349, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The task of audio-driven talking avatars has gained increasing attention recently, propelled by advancements in videogeneration technologies. GAN-based methods [4, 8, 19, 34, 44, 47, 48] typically employ a two-stage pipeline involving an audio-to-motion module and a motion-to-video module. Sadtalker [42] introduces a 3DMM-based motion representation along with a conditional VAE for generating high-fidelity talking heads. LivePortrait [4] expands the training dataset incorporating images and videos, using implicit keypoints for high-quality generation. NeRF-based methods [5, 30, 38-41] explore animating talking heads in 3D neural", + "bbox": [ + 93, + 739, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "spaces. Real3D-Portrait [41] combines a pre-trained image-to-triplane module with a motion adapter for zero-shot and one-reference 3D talking head generation. MimicTalk [40] captures the talking style of individuals through in-context-learning audio-to-motion modules and rapid fine-tuning of triplanes for mimic talking head generation.", + "bbox": [ + 516, + 90, + 903, + 179 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based talking avatar synthesis has swiftly become a crucial and efficient approach [1, 2, 10, 14, 31, 33, 36]. EMO [31] pioneers a double UNet framework in audio-driven avatar creation, yielding impressive and lifelike results. Concurrently, EchoMimic [1] trains on audio and facial landmarks for versatile control inputs. Hallo [36] incorporates a hierarchical audio-driven visual synthesis module to refine audio-visual alignments, while Hallo2 [2] explores image augmentation strategies for 4K avatar video synthesis. Loopy [10] introduces a motion-frame squeeze technique and audio-to-latent training for expressive results. Despite their progress in expressive audio-driven generation, diffusion-based methods still face challenges with slow inference speeds.", + "bbox": [ + 516, + 184, + 903, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion Model Acceleration", + "text_level": 1, + "bbox": [ + 516, + 415, + 772, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion acceleration involves lowering costs for denoising runs and reducing the number of runs. Various denoising methods aim to cut costs, such as utilizing Model quantization [6, 13, 25, 26, 29, 45] to decrease the model parameters directly. ViDiT-Q [47] introduced a new diffusion transformer quantization framework for video and image with mixed-precision quantization. DeepCache [20] takes cues from video compression to propagate reusable result regions through the model's internal structure, while Faster Diffusion [12] emphasizes re-using computed encoder features.", + "bbox": [ + 516, + 441, + 903, + 601 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reducing the number of denoising runs is a direct way to accelerate diffusion models. SD-Turbo[24] leverages the ideas from GAN and introduces adversarial diffusion distillation to enhance the generation fidelity. SnapFusion[12] proposes progressive distillation to achieve fewer steps by learning the previous student model progressively. SwiftBrush[21] and SwiftBrushV2[3] utilizing variational score distillation to make it possible for texts-only distillation in text-to-image tasks. SDXL-lightning[15] mixes progressive distillation and adversarial distillation to achieve one-step image generation with a delicate training process. Applying consistency models [17, 28] to text-to-image tasks also helps reduce diffusion inference steps. LCM[17] views the guided reverse diffusion process as solving an augmented probability flow ODE. TCD[46] proposes trajectory consistency distillation to accurately trace the entire trajectory of the probability flow ODE in semilinear form with an exponential integrator. However, there is no specially-designed distillation framework and techniques for talking avatar synthesis tasks to our knowledge.", + "bbox": [ + 516, + 607, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3102", + "bbox": [ + 483, + 946, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 89, + 225, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed method, FADA, a diffusion distillation framework for avatar synthesis, as illustrated in Figure 1. Section 3.1 provides an overview of the diffusion distillation framework, based on recent PeR-Flow [37]. Section 3.2 delves into the specifics of the adaptive mixed-supervised distillation process. In Section 3.3, we present the multiple-CFG distillation technique with a learnable token implementation.", + "bbox": [ + 89, + 114, + 483, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework and Preliminary", + "text_level": 1, + "bbox": [ + 89, + 241, + 410, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon cutting-edge diffusion-based methods for talking avatar synthesis [1, 2, 9, 10, 14, 31, 33, 36], we introduce a dual-Unet architecture for the teacher model, leveraging pre-trained weights from Stable Diffusion[23] 1.5. As depicted in Figure 1, the reference image and motion frames (i.e., images from the last clip) undergo processing by a fixed VAE [11] and reference net. The output from each reference layer is then input to the corresponding layer in the denoising net for spatial attention. To facilitate audio-driven controls and ensure temporal consistency in talking avatar synthesis, audio attention and temporal attention layers are integrated into the denoising net. Initially following the Loopy design [10], we simplify the model by removing the TSM and Audio2Latents modules, streamlining subsequent model distillation. The student model mirrors the teacher's structure, with the addition of a CFG control layer, detailed in Section 3.3.", + "bbox": [ + 89, + 263, + 483, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We choose the recent PeRFlow [37] as the optimization target for distillation. Initially, we train the teacher model on carefully curated high-quality datasets that prioritize motion amplitude and audio-visual synchronization. The teacher model is optimized using the $\\epsilon$ -prediction DDPM loss [7], where it predicts noise $\\epsilon_{\\mathrm{tea}}(\\cdot)$ from noisy latent $z_{t}$ at timestep $t$ . The training objective can be succinctly described as:", + "bbox": [ + 89, + 512, + 483, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {t e a}} = \\mathbb {E} _ {z _ {t}, c, t, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\mathrm {t e a}} (z _ {t}, t, c; \\theta_ {\\mathrm {t e a}}) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 638, + 482, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c$ includes the audio feature, reference image and motion frames.", + "bbox": [ + 89, + 667, + 482, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that it is not necessary for the teacher model to be trained with flow-matching loss, and both $\\epsilon$ -prediction and $v$ -prediction are valid for PeRFlow distillation. During the student distillation training, firstly same-interval $K$ time windows $\\{[t_k, t_{k-1})\\}_{k=K}^1$ are created where $t_k = k / K, k \\in \\{0, 1, \\dots, K\\}$ . A time window $[t_k, t_{k-1})$ is randomly sampled out, and the standpoint $z_{t_k}$ can be derived from the marginal distribution of the ground-truth targets, where $z_{t_k} = \\sqrt{1 - \\sigma^2(t_k)} z_0 + \\sigma(t_k)\\epsilon$ and $\\sigma(t)$ denotes the noise schedule. Then the teacher-predicted ending point $\\hat{z}_{t_{k-1}}$ of this time window can be predicted and solved by ODE solver $\\Phi(\\cdot)$ :", + "bbox": [ + 89, + 696, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t _ {k - 1}} = \\Phi \\left(z _ {t _ {k}}, t _ {k}, t _ {k - 1}; \\theta_ {\\text {t e a}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 885, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi (\\cdot)$ denotes the DDIM solver[27] which performs several inference steps iteratively to reach the endpoint of this time window. Note that the parameters of the teacher model are frozen and no gradient computations are needed here. After that, the input noisy latent of student $\\hat{z}_t$ will be derived from a linear interpolation between starting point $t_k$ and ending point $t_{k - 1}$ :", + "bbox": [ + 511, + 90, + 905, + 195 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t} = \\frac {z _ {t _ {k}} - \\hat {z} _ {t _ {k - 1}}}{t _ {k} - t _ {k - 1}} (t - t _ {k}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 200, + 905, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through the linear interpolation above, the ODE flow will be rectified to piecewise straight lines so only a few steps are enough to denoising in a time window, which reduces the number of timesteps at inference stage. Meanwhile, we can figure out the target noise $\\hat{\\epsilon} (t)$ via parameterization[37]:", + "bbox": [ + 511, + 239, + 905, + 328 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} (t) = \\frac {\\hat {z} _ {t _ {k - 1}} - \\lambda_ {k} z _ {t _ {k}}}{\\eta_ {k}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 335, + 905, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $\\lambda_{k} = \\sqrt{\\alpha_{t_{k - 1}}} / \\sqrt{\\alpha_{t_{k}}}$ and $\\eta_{k} = \\sqrt{1 - \\alpha_{t_{k - 1}}} -$ $\\sqrt{1 - \\alpha_{t_k}}\\sqrt{\\alpha_{t_{k - 1}}}\\big{/}\\sqrt{\\alpha_{t_k}}$ respectively. Finally, the training objective of the distillation process can be obtained as", + "bbox": [ + 511, + 378, + 905, + 421 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i s t i l l}} = \\mathbb {E} _ {k, t \\in [ t _ {k}, t _ {k - 1}), z _ {t}, c, \\epsilon} \\left[ \\| \\hat {\\epsilon} (t) - \\epsilon_ {\\mathrm {s t u}} (\\hat {z} _ {t}, t, c; \\theta_ {\\mathrm {s t u}}) \\| ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 431, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\mathrm{stu}}(\\cdot)$ refers to the predicted noise from the student model and $L_{\\mathrm{distill}}$ represents the distillation loss. Note that $L_{\\mathrm{distill}}$ is only a trivial part of our framework, the mixed-supervised training and multi-CFG distillation will be proposed in Section 3.2 and 3.3.", + "bbox": [ + 511, + 462, + 905, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adaptive Mixed-Supervised Distillation", + "text_level": 1, + "bbox": [ + 511, + 542, + 852, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments, we found that distillation significantly degrades the model's performance, especially overall video synthesis quality, as shown in Table 3. Increasing data can improve performance, but in audio-driven portrait generation, the weak control of audio over motion requires the model to learn high-quality motion patterns to fill in details not covered by audio. This necessitates stricter data filtering, reducing the available data quantity.", + "bbox": [ + 511, + 564, + 905, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fortunately, this problem can be mitigated in our proposed distillation process. In our design, we first maintain the teacher training unchanged with only a high-quality dataset $\\mathcal{A}$ to ensure its output stability. Then during the distillation stage, the data filtering strategies will be relaxed and some moderate-quality data samples (a larger dataset $\\mathcal{B}$ that includes $\\mathcal{A}$ where $|\\mathcal{B}| > |\\mathcal{A}|$ .) will be fed to student training. Since the student model mimics the teacher model's output, it can maintain the high-quality generation capability of the teacher model even with the inclusion of lower-quality data. To further leverage the new data during the distillation stage, we introduce a ground-truth supervised loss $L_{\\mathrm{gt}}$ . The ground-truth noisy latent $z_{t}^{*}$ and ground-truth target noise $\\epsilon^{*}(t)$ can be simply obtained by replacing $\\hat{z}_{t_{k-1}}$ with ground truth ending point $z_{t_{k-1}}$ in For", + "bbox": [ + 511, + 681, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3103", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cc474c8bd90a591ad224fc6e3f96fe418ae127a148d1f9a8109d9a57cb648be5.jpg", + "image_caption": [ + "Figure 1. Overall distillation framework of FADA. The teacher model is trained only with high-quality data, which is omitted in the figure. The student model is trained by a mixed loss of ground-truth and teacher-supervised loss to leverage data of varying quality. Learnable token-based CFG conditions enable the student model to mimic the multi-CFG process, further reducing inference times. For simplicity, we have omitted some components commonly used in previous methods." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 903, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mula 3 and 4. Similar to distillation loss computation, the ground-truth loss $L_{\\mathrm{GT}}$ and total mixed loss $L_{\\mathrm{total}}$ can be described as", + "bbox": [ + 89, + 546, + 485, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {G T}} = \\mathbb {E} _ {k, t \\in \\left[ t _ {k}, t _ {k - 1}\\right), z _ {t}, c, \\epsilon} \\left[ \\| \\epsilon^ {*} (t) - \\epsilon_ {\\mathrm {s t u}} \\left(z _ {t} ^ {*}, t, c; \\theta_ {\\mathrm {s t u}}\\right) \\| _ {2} ^ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 599, + 483, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o t a l}} = L _ {\\text {d i s t i l l}} + \\mathcal {W} \\cdot L _ {\\mathrm {G T}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 642, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ represents the loss weight of $L_{GT}$ . Through this approach, the student model can better learn high-quality motion pattern generation from the teacher model while also learning more common patterns from a broader range of data to enhance its generalization capability. A trivial implementation idea is to set a fixed hyper-parameter to $\\mathcal{W}$ but this is not very effective because it assumes every sample has the same learning value, which is inaccurate.", + "bbox": [ + 89, + 662, + 483, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, an adaptive method to balance teacher-supervised loss $L_{\\mathrm{teacher}}$ and ground-truth-supervised loss $L_{\\mathrm{gt}}$ is necessary, and the potential information behind the two losses should be taken into account. Considering the ratio $\\mathcal{R} = L_{\\mathrm{gt}} / L_{\\mathrm{teacher}}$ , it indicates the difference level between the $L_{\\mathrm{gt}}$ and $L_{\\mathrm{teacher}}$ . Based on qualitative observations of the $\\mathcal{R}$ value, there are two trends: as $\\mathcal{R}$ increases, the sample is more likely to be a valuable learning case,", + "bbox": [ + 89, + 782, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "such as a singing video, since it is often outside the teacher-training dataset $\\mathcal{A}$ . However, if $\\mathcal{R}$ is too large, the sample is more likely to be low-quality, such as one with poor audio-visual sync, because the teacher results are well synced. The statistical observations are discussed in Appendix A.2.", + "bbox": [ + 511, + 546, + 906, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we should gradually increase the loss weight when $\\mathcal{R}$ increases in a reasonable range, and also punish the loss weight if $\\mathcal{R}$ increases out of the peak threshold $\\mathcal{R}_p$ . For those cases whose $\\mathcal{R}$ is bigger than the dead threshold $\\mathcal{R}_d$ , the ground-truth supervised loss weight should be zero. The detailed formulation for loss weight $W$ can be obtained as:", + "bbox": [ + 511, + 621, + 906, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} = \\mathcal {W} _ {0} \\cdot \\left\\{ \\begin{array}{l l} \\mathcal {R} ^ {s}, & \\mathcal {R} \\in [ 0, \\mathcal {R} _ {p}) \\\\ \\mathcal {R} _ {p} ^ {s} (\\mathcal {R} _ {d} - \\mathcal {R}) / (\\mathcal {R} _ {d} - \\mathcal {R} _ {p}), & \\mathcal {R} \\in [ \\mathcal {R} _ {p}, \\mathcal {R} _ {d}) \\\\ 0, & \\mathcal {R} \\in [ \\mathcal {R} _ {d}, + \\infty) \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 719, + 903, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s, \\mathcal{W}_0, \\mathcal{R}_p$ and $\\mathcal{R}_d$ are hyper-parameters, they are set to 0.25, 0.2, 30, 100 respectively in this paper. Note that these hyper-parameters are relatively robust across different data distributions because they are used to select data within a relatively moderate range for training.", + "bbox": [ + 511, + 781, + 906, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through the proposed design, FADA learns general generation capabilities from more ordinary data during distillation, enhancing model robustness. Our experiments validate", + "bbox": [ + 511, + 856, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3104", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the proposed loss's effectiveness. Additionally, our adaptive mixed-supervised distillation can extend to tasks like text-to-image or text-to-video without specific restrictions.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-CFG Distillation with Learnable Token", + "text_level": 1, + "bbox": [ + 89, + 148, + 475, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In audio-driven talking avatar synthesis tasks, the MultiCFG technique is utilized [10, 14] to achieve robust control of multiple conditions (reference image and audio) effectively. Specifically, the single CFG calculates the direction from unconditional predicted values to a certain conditional predicted value and then moves a certain distance from the unconditional predicted values to obtain conditional predicted values. Furthermore, the MultiCFG recursively computes the next CFG based on the first CFG. In this paper, multiCFG inference result noise $\\epsilon_{cfg}$ can be derived as follows[10]:", + "bbox": [ + 89, + 171, + 482, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {c f g} = \\operatorname {c f g} _ {a} \\times \\left(\\hat {\\epsilon} _ {a} - \\hat {\\epsilon} _ {r}\\right) + \\operatorname {c f g} _ {r} \\times \\left(\\hat {\\epsilon} _ {r} - \\hat {\\epsilon} _ {b}\\right) + \\hat {\\epsilon} _ {b}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 344, + 482, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{cfg}_a$ and $\\operatorname{cfg}_r$ indicate the CFG guidance scale of audio and reference condition. $\\hat{\\epsilon}_a$ refers to the noise prediction with both audio and reference conditions, while $\\hat{\\epsilon}_r$ removes the audio condition and $\\hat{\\epsilon}_b$ lacks both audio and reference.", + "bbox": [ + 89, + 371, + 482, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, during inference, the model needs to run three times for each denoising step, which is time-consuming. It naturally comes to mind to inject the CFG guidance scale as a condition into the network, and then let the student model learn the CFG reasoning characteristics of the teacher model during the distillation process, which echoes a similar concept to the $\\omega$ -condition [20]. Specifically, the teacher model will conduct complete CFG reasoning during training to obtain predictions $\\hat{\\epsilon}_{\\mathrm{cfg}}$ controlled by CFG. Subsequently, these predictions are further integrated into Formulas 2, 3 and 4 to derive input noisy latent $\\hat{z}_t^{\\mathrm{cfg}}$ and target noise $\\hat{\\epsilon}^{\\mathrm{cfg}}(t)$ controlled by CFG, eventually undergoing a similar distillation procedure as Formula 5.", + "bbox": [ + 89, + 431, + 482, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, directly adding the CFG value as a condition to the network does not yield satisfactory results. A careful design is needed for the injection method of the CFG scale condition. By observing the calculation of Multi-CFG reasoning in Formula 9, it is evident that this is akin to performing linear transformations on three independent vectors $\\epsilon_{b}$ , $\\epsilon_{a}$ , and $\\epsilon_{r}$ in the noise space using the given CFG strengths $\\mathrm{cfg}_{a}$ and $\\mathrm{cfg}_{r}$ . Meanwhile, $\\epsilon_{r}$ serves as a crucial link between the $\\epsilon_{b}$ and $\\epsilon_{a}$ . Simply embedding the CFG values overlooks the aforementioned relationships. Therefore, we introduce learnable tokens $\\gamma_{b}, \\gamma_{r}$ , and $\\gamma_{r}$ into the model to learn the features required for CFG strength control and help the student model mimic the multi-CFG process. To achieve this, we design the CFG embedding $Emb_{\\mathrm{cfg}}$ of the CFG guidance scale $\\mathrm{cfg}_{a}$ and $\\mathrm{cfg}_{r}$ as follows:", + "bbox": [ + 89, + 625, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {\\text {c f g}} = \\operatorname {c f g} _ {a} \\times \\left(\\gamma_ {a} - \\gamma_ {r}\\right) + \\operatorname {c f g} _ {r} \\times \\left(\\gamma_ {r} - \\gamma_ {b}\\right) + \\gamma_ {b}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 857, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the CFG embedding $Emb_{\\mathrm{cfg}}$ , we sim", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "imilarly introduce a CFG layer after each audio layer in the denoising network, injecting the information from $Emb_{\\mathrm{cfg}}$ into the denoising process through cross attention. Importantly, the CFG layer does not introduce excessive performance overhead, especially when compared to the original cost of running three CFG runs. Therefore, our proposed multi-CFG distillation with learnable tokens can achieve nearly three times faster operation speed.", + "bbox": [ + 511, + 90, + 903, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 224, + 645, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we will introduce the experimental results. In Section 4.1, we will present the details of the implementation of FADA and the meta-information of our training and testing datasets. In Section 4.2, we will compare our proposed methods with other state-of-the-art diffusion-based talking avatar methods via quantitative and qualitative comparison. In Section 4.3, we will analyze our proposed techniques including the basic distillation method, adaptive mixed-supervised distillation and multi-CFG distillation, and prove their effectiveness.", + "bbox": [ + 511, + 250, + 903, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 409, + 720, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets Regarding the training dataset, we primarily used speaker videos obtained from the internet, sliced and cropped to durations ranging from two to ten seconds with facial frames, and resized to $512 \\times 512$ resolution. We designed multi-dimensional filters for this dataset, including filters for data source, image quality, audio-visual synchronization, background stability and degree of movement, each with different thresholds to control the filtering intensity. The details of data collection and data filter strategies will be presented in Appendix D. As elucidated in Section 3.2, by using strict filtering thresholds, we obtained approximately 160 hours of high-quality training data $\\mathcal{A}$ , which was utilized in teacher pre-training. Additionally, by using lenient filtering thresholds, we obtained around 1300 hours of moderate-quality training data $\\mathcal{B}$ for distillation training. Following the approach of Echomimic [1], we selected 100 random samples from the open-source test sets HDTF [43] and CelebV-HQ [49] to validate the model's performance on various real human scenarios. Furthermore, to evaluate the model's performance in more open and extreme scenarios, we adopted the openset test set from Loopy [10] for qualitative comparisons, which contains various kinds of portraits and audios.", + "bbox": [ + 511, + 430, + 906, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details The training process consists of teacher pre-training and student distillation. The training process of the teacher model closely aligns with the first two stages of recent methods[10, 31] with high-quality dataset $\\mathcal{A}$ . In student distillation with the moderate-quality dataset $\\mathcal{B}$ , the student model is initialized from the teacher model and integrated with CFG learnable tokens and CFG layers. During testing, the CFG guidance scale is set to 2.5 and 6.5 for reference and audio conditions in models without CFG", + "bbox": [ + 511, + 768, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3105", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "distillation, while the scales were set to 2.0 and 6.5 for models with CFG distillation. The number of time windows in PeRFlow[37] is 4. All results of FADA are conducted with 6-step inferences, where the two time-windows near $T = 1$ contain two inference steps while the others contain one inference step. More details about the teacher and student training will be discussed in Appendix E.", + "bbox": [ + 89, + 90, + 480, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 200, + 395, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics and Baselines For the purpose of quantitative comparisons, we evaluate the generated talking avatar videos with several metrics. We assess the image quality with IQA[35] metrics and evaluate the audio-visual synchronization distance with a pre-trained SyncNet[22] as Sync-D. FID and FVD[32] (with 3D ResNet) are assessed to evaluate the image/video-level fidelity difference between generated results and ground truths in the image inception space. E-FID (Expression-FID)[31] measures the inception distance between the generated and ground truths in expression parameter space. NFE-D indicates the number of function evaluations of the denoising net. For instance, our teacher model conducts 25-step DDIM inference with 3 CFG runs, leading to 75 NFE-D. Since our proposed FADA is a diffusion distillation framework, we mainly compare recent open-source diffusion-based state-of-the-art methods in talking portrait generation, including Hallo[36], EchoMimic[1], and V-Express[33]. Additionally, for reference, we also include the state-of-the-art GAN-based method SadTalker[42].", + "bbox": [ + 91, + 220, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison As shown in Table 1, we conduct quantitative comparison experiments with CelebV-HQ and HDTF test sets. Since the differences in IQA are subtle, we provide $95\\%$ confidence intervals to show their statistical significance. Ours-Balanced refers to our proposed method with mixed-supervised distillation with PeRFlow loss. Based on it, Ours-Fast utilizes multi-CFG distillation to achieve about 3-time speedup and gets a slight performance drop. Ours-Balanced achieves better IQA, FVD, FID, and E-FID on the challenging open-set CelebVHQ test set compared to the baselines. It also leads or is comparable in most metrics on the simpler HDTF test set. Ours-Fast maintains similar performance in IQA and Sync-D metrics with a slight drop in FVD, FID, and E-FID, which is still comparable to the baselines in most metrics. In terms of audio synchronization, the trend is similar. Notably, VExpress results show almost only lip movements, which, despite good metrics, do not produce satisfactory video quality. Finally, in terms of speedup, it is evident that both Ours-Balanced and Ours-Fast achieve significant speedup using the same SD base model as the baselines. The subjective evaluation will be conducted in Appendix F.", + "bbox": [ + 91, + 517, + 483, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison As illustrated in Figure 2 and in the supplementary video materials, Ours-Fast and Ours-Balanced maintain good video quality and vivid audiovisual expressiveness. We highly recommend readers watch", + "bbox": [ + 89, + 840, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the supplementary videos since it is difficult for the static image to show the audio-motion synchronization and temporal consistency. V-Express shows color differences and inferior identity-preserving abilities in all cases. Sadtalker generates good identity-preserving videos at the image level, but the head motions and lip movements are not well-synced with the emotion and pace of the given audio. EchoMimic and Hallo generate acceptable expressions, but there are temporal jitters that damage the visual quality. Moreover, by observing case-4 and case-6, we can find that FADA produces more full and plump pronunciations than baselines.", + "bbox": [ + 511, + 90, + 906, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 275, + 666, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of Distillation Methods. As illustrated in Table 2, we first compared various basic methods of diffusion distillation, including TCD [46], RFlow [16], and PeRFlow [37]. It was observed that compared to the TCD method, the PeRFlow method showed similar performance in IQA, FVD, Sync-D, and FID metrics, but with a significant advantage in E-FID. Additionally, the basic RFlow method could not handle the 6-step inference in our task setting. Therefore, we selected PeRFlow as the basic distillation method. We also tried simply adding moderate-quality data during teacher model training, which led to a decrease in model performance. These results indicate that existing distillation methods cannot directly produce a high-quality student model. Additionally, attempting to enhance the teacher model's capabilities by relaxing data filtering criteria and increasing data volume does not yield positive results. The comparison with the adversarial-based method is shown in Appendix G.", + "bbox": [ + 511, + 297, + 908, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of Adaptive Mixed-Supervised Distillation Here, based on the previously selected PeRFlow, we added dataset $\\mathcal{B}$ and a mixed-supervised loss to investigate the effectiveness of our Adaptive Mixed-Supervised Distillation. As shown in Table 4, below the hdashline, we provide the performance of Mixed-Supervised Distillation with the addition of dataset $\\mathcal{B}$ under different parameters. Comparing the effects of different Fixed-x values, it is evident that the ground-truth loss improves model performance, and a smaller loss weight value of 0.2 yields better results. We further investigated the effectiveness of the adaptive loss weight strategy. In the adaptive-unl. setting (unlimited, without using peak and dead thresholds), the adaptive loss weight increases infinitely as $\\mathcal{R}$ increases, resulting in slightly inferior performance compared to the Fixed-0.2 setting. However, when employing the complete adaptive strategy, the ground-truth loss of low-quality samples is automatically discarded, leading to a distillation effect that surpasses the Fixed setting and approaches the performance of the teacher model, especially in terms of the FVD metrics. We also compared different values of $s$ and ultimately chose 0.25 due to its slight advantage.", + "bbox": [ + 511, + 564, + 908, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, by comparing the performance of teacher", + "bbox": [ + 532, + 886, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3106", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3474dd35f04ccdfadd78c3afc920ffd9cdeb78bdcae854511697f7b7bcb9eb0a.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with state-of-the-art methods on CelebV-HQ & HDTF test sets. Bold and underlined numbers refer to the best and the second-best result in this test set." + ], + "table_footnote": [], + "table_body": "
MethodNFE-DCelebV-HQ TestHDTF Test
IQA↑Sync-D↓FVD-R↓FID↓E-FID↓IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Sadtalker[42]-2.953±.0868.765171.836.642.2483.435±.0697.87024.9325.351.559
Hallo[36]803.505±.0919.07953.9935.962.4263.922±.0797.91721.7120.151.337
EchoMimic[1]603.307±.10410.3754.7135.373.0183.994±.0839.39118.7119.011.328
V-Express[33]502.946±.1239.415117.865.092.4143.482±.0717.38248.0330.911.506
Ours-Balanced183.588±.0918.99850.3633.672.2023.951±.0827.92513.3218.391.362
Ours-Fast63.550±.0938.74654.6935.012.6043.927±.0817.99616.6718.511.635
", + "bbox": [ + 91, + 128, + 916, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91ac309c50f1000e1a1f1a7c9eda42cb03a94e974d3cf58f5e5d09697b2fd84e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 292, + 220, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e47b696e285041304bef896edc2b845bf848231933631ee11b3c8d66a8503b89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 367, + 220, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a5eae541ab7cb35e98e41a4b13cd2eafe748569191bf586c3da83aa56e371dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 438, + 220, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3a71f3f6bb9cd5e51e9315a0ba7d9a550ee6aa23b1e4421bf657f26c5b1322c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 511, + 220, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c8bb52902793ce0e54ec1e518fe6bdbe972994bfc42d0edfa30354aebace100.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 584, + 220, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ebe079194feb410e47ff8b998a1dba7ba7b5894d57d16d58ec9e7c4399ff87d.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 127, + 657, + 220, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59a8c40cc6ff11b07f0b28fdc001ccf36c2fe83107e7dcdf5992fe532eb26826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 294, + 323, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf39d59b677c3bdae860cefcb21f3882471aa53208ace999153bf869280f9830.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 367, + 323, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b30623fd9b7cf9dd57fa0efffa31693ee4b65f2877427f0326174655ee909120.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 439, + 323, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2760ce2115c605ef9b3ce2b01c69c56c6ce85707a20f351bf57a3f255f97db21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 511, + 323, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8aff1b9766452f80838de5fa68a38aee60d7b2e83eef941bdcbfb1650c6b71bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 583, + 323, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59172be50245465e690d2222ac9df98d8c7460137f7bd7b3047653052629289e.jpg", + "image_caption": [ + "SadTalker" + ], + "image_footnote": [], + "bbox": [ + 230, + 657, + 323, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8015fb8dbe326105aeef7266da0bc8d5032495dfb21fe450aa13ea4e4b60254f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 294, + 419, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f843bfe93a8854f9c9832d7b2074e8702fd7c2fa40ed67b3aa7be1e019513f97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 367, + 419, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9468345cef7c2d54ac66f1843359f0a378928dec857ed3cf5ee881b6a088568.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 439, + 419, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10c2f135edbbee68023574edbbfada5a9f68ef62327731391183cd1bcd0da3b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 511, + 419, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05a631022453d1660e114aaaba2cc53d4bb5096327f33afc328fc3ba0f7166ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 583, + 419, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab648c7601b18a21a2b02621c86d26979812cb9633f52c63312e8f69069c3f0a.jpg", + "image_caption": [ + "VExpress" + ], + "image_footnote": [], + "bbox": [ + 328, + 657, + 419, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8aaed49357062d3953f5c4816ca89b0eb6053b1604c5c01f724e218f38354fef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 294, + 517, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79b39434fbf19a06bc7772134181cbbeaf88825ccc5cda9d5a203de14e06464e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 367, + 517, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5e6955aa23c03e5239673851426de0ef2f48c75c88458c2863b7e58019a0129.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 439, + 517, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2ba2acbdab5a12d821be172edc8a24f987205b67dad25b3b6715a21cbf1e02e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 511, + 517, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d65e08e5b3d5505f7380ca15dc11f8fda68f532449289a90865b226d235c3710.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 583, + 517, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cad8d5239aa04828f15dcbbcb471133146b6d36a281e6ae3da917846c8ff8689.jpg", + "image_caption": [ + "Hallo" + ], + "image_footnote": [], + "bbox": [ + 424, + 657, + 517, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/170bea71c5883092017da89934c2b99ee8b7a04f49ccbf7797d32b4d27dacb5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 294, + 614, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d071fb204fd7dc3bdefe55a6ab3ba66de87d9575d4e0464bdd1a02e995d736cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 367, + 614, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecddce9a1170a25585f47cf0e07c8680858a799b6fd0a877cbac2b2d5dd90d9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 439, + 614, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4fff0c994435277323fba42650c8717f8b8efd995ff56ebee5da37a818be8046.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 511, + 614, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4da9d5aee4760b00c4809d7f2720463143c9036e1ea50580f84b242ad9061242.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 583, + 614, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efc6d9aa44cbff41048963277ad56a377c79705f67955bc0c55a330936b74c2b.jpg", + "image_caption": [ + "EchoMimic", + "Figure 2. Qualitative comparisons between FADA and baselines across different portraits and pronunciations in openset." + ], + "image_footnote": [], + "bbox": [ + 521, + 657, + 614, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ade7d7553527da62a23fc74917a0a83fb79bc35b6ae3dbccf97f087533e07a53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 294, + 710, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/179103c7bb34be9b57af474cf32e2e3b34996a5250d640e3a86f2a58a1b5f427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 367, + 710, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eeede619c8276360fadb778ee703fb6bb041e366a93daf1dfd2e0452807f379e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 439, + 710, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/461adbc1a22ec6f2ae2925680533c46176e854c91a9277fda14911d6b72b4e4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 511, + 710, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f706423712a94d3b05334856ca25061f3ddaccaab26a328d2ad3009f60bd2f8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 583, + 710, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/649a842008a9a793d8093ad06fafa5cc5a627e43e0d67eaea9684c24e3143702.jpg", + "image_caption": [ + "Ours-Fast" + ], + "image_footnote": [], + "bbox": [ + 617, + 657, + 710, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b128be656c081bf8542c7de2fe9e4d0282de8ccd3f6e7f784cdf2dabdf713da.jpg", + "image_caption": [ + "[ə]" + ], + "image_footnote": [], + "bbox": [ + 712, + 294, + 807, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb5ca9432ba958afd9bc166698949d523baa2b6c695b39135eae47c5f319c303.jpg", + "image_caption": [ + "[I]" + ], + "image_footnote": [], + "bbox": [ + 712, + 367, + 807, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e528c0a630a73b22b50fc3ffa9bbf19e25a5ae6cb836dbf90ab43c89c19ad00.jpg", + "image_caption": [ + "[u:]" + ], + "image_footnote": [], + "bbox": [ + 712, + 439, + 807, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bf0e77c96ef04f4ef787db97483c415c20c498bfaac4e025072dcd12c1cf1e6.jpg", + "image_caption": [ + "[æ]" + ], + "image_footnote": [], + "bbox": [ + 712, + 511, + 807, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a1d2467dd1c9f02e0cbfef74f46a3195208027f05e316d84eaadb916798c216.jpg", + "image_caption": [ + "[eI]" + ], + "image_footnote": [], + "bbox": [ + 712, + 583, + 807, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f34cf9edabf69e80faffe5dd1da9ac4fd34612b4d7bf5bf9906af72e208d53f.jpg", + "image_caption": [ + "Ours-Balanced", + "[1]" + ], + "image_footnote": [], + "bbox": [ + 712, + 657, + 807, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and student models under different data qualities and scales, it can be demonstrated that mixed supervised distillation actually enhances the model's utilization of moderate-quality data. As depicted in Tables 2 and 3, when the teacher model is trained with a large-scale moderate-quality dataset, a decline in model performance is observed, whereas the student model exhibits the opposite trend. Briefly, only mixed-supervised distillation can take advantage of the moderate-", + "bbox": [ + 89, + 782, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "quality data. More analysis of mixed-supervised distillation will be discussed and more hyper-parameters experiments will be conducted in Appendix A.1.", + "bbox": [ + 511, + 782, + 906, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Multi-CFG Distillation As demonstrated in Table 4, we further explore the effects of multi-CFG distillation on top of mixed-supervised distillation. Initially, we tested the results of conditional reasoning without CFG dis", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3107", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b89f26b7e5334109cfceabdfca7281cf4d08629414ff98da3b1c04943ee2472.jpg", + "table_caption": [ + "Table 2. Ablation study of basic distillation methods on HDTF test set. All methods in this table are free from mixed-supervised and CFG distillation." + ], + "table_footnote": [], + "table_body": "
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Teacher with B753.8858.12319.6421.401.689
TCD183.8277.93723.0123.151.893
RFlow182.8368.06681.1049.732.484
PeRFlow183.8238.00121.9121.261.477
", + "bbox": [ + 101, + 142, + 470, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cd92bf2faa5dbb59391b5164bb9963392b7811728b0745b78e534eec8c573fb7.jpg", + "table_caption": [ + "Table 3. Ablation study of adaptive mixed-supervised distillation on HDTF test set. All methods use PeRFlow with dataset $\\mathcal{B}$ without CFG distillation. Fixed-x indicates fixed ground-truth loss with weight x. Adaptive-unlimited-s and Adaptive-s indicate adaptive ground-truth loss with hyper-parameter $s$ , with the former being without peak and dead threshold." + ], + "table_footnote": [], + "table_body": "
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Stu. with A183.8238.00121.9121.261.477
Stu. with B (Fixed-0)183.7627.90925.0621.471.462
+ Fixed-0.2183.8637.93413.7119.941.434
+ Fixed-1.0183.8127.90324.5321.421.494
+ Fixed-0.2, only A183.8807.96215.3220.711.512
+ Adaptive-unl.-0.5183.8247.91415.1520.741.471
+ Adaptive-unl.-0.25183.8467.90614.2019.921.424
+ Adaptive-0.5183.9497.91913.8318.641.395
+ Adaptive-0.25183.9517.92513.3218.391.362
", + "bbox": [ + 98, + 334, + 478, + 483 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d549731a693bc54c5f60526cee23a89a2b6e711e8029982291efc3f3bfac2a46.jpg", + "table_caption": [ + "Table 4. Ablation study of multi-CFG distillation on HDTF test set. All methods in this table are PeRFlow + mixed-supervised distillation. CFG time embed refers to CFG distillation through timestep embedding injection." + ], + "table_footnote": [], + "table_body": "
MethodNFE-D IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
w.o. CFG63.6749.02027.1026.562.096
+ CFG time embed63.7018.57723.2422.471.791
+ CFG layer63.7467.90620.0520.941.620
+ CFG layer w. tokens63.9277.99616.6718.511.635
", + "bbox": [ + 94, + 556, + 478, + 647 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61c03eae7ba7a2ff50a84cec046c655e0339ca057af1a5cc2d4866890c852c42.jpg", + "image_caption": [ + "Figure 3. Line charts showing the variations of FVD and Sync-D metrics with different audio CFG on HDTF test set. Reference CFG is set to 2.0 in this figure." + ], + "image_footnote": [], + "bbox": [ + 102, + 650, + 465, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tillation, which significantly deteriorated the model performance, especially leading to a doubling of the FVD metric representing video quality. For techniques like time embedding injection similar to $\\omega$ -condition [20], they partially restored model performance but with limited efficacy. In con", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92f17b15c377942032f3430bc05d4cd9c1e12232f53f0a0bde41f39fb07bed04.jpg", + "image_caption": [ + "Figure 4. Line charts showing the variations of FVD and Sync-D metrics with different ref CFG on HDTF test set. Audio CFG is set to 6.5 in this figure." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 890, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trast, our proposed learnable token-based CFG control layers significantly enhance model performance, nearly matching the teacher model and even surpassing it in FVD. Notably, our model requires only $8\\%$ of the teacher model's NFE. We also tested without learnable tokens, and while the pure CFG control layer had some effect, it could not match the teacher model, especially in overall video quality metrics like FVD.", + "bbox": [ + 511, + 277, + 906, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify whether our proposed multi-CFG distillation effectively captures the influence of audio and reference conditions on the generated results, we plotted line charts of FVD and Sync-D metrics with variations in audio CFG and reference CFG, as shown in Figures 3 and 4. First, we observed improvements in both FVD and Sync-D as audio CFG increased, reaching a stable maximum value after 6.5. Hence, we ultimately chose 6.5 as the audio CFG for inference. Second, the influence of reference CFG on FVD and Sync-D metrics displayed an initially positive, then negative trend, with optimal performance achieved around the 2.0 position. These results indicate that our CFG control layer effectively mimics the multi-CFG process and achieves good results. More visualized examples and performance analysis under different scenes will be presented in Appendix C.", + "bbox": [ + 511, + 400, + 908, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 664, + 633, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, we proposed FADA, a fast diffusion avatar synthesis framework with mixed-supervised multi-CFG distillation. By designing a mixed-supervised loss, we leverage data of varying quality to enhance the robustness of generated results. Furthermore, a learnable token-based multi-CFG condition design is introduced to maintain the correlation between the audio and the generated video in the distilled models. The quantitative and qualitative experiments across multiple datasets show that our balanced setting achieves fast inference and high-fidelity generation ability. Ablation studies prove our proposed methods are effective in FADA. The limitations, future works and ethical concerns of this paper will be discussed in Appendix B, H and I respectively.", + "bbox": [ + 511, + 695, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3108", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is sponsored by National Natural Science Foundation of China under Grant No.62222211 and National Natural Science Foundation of China under Grant No.U24A20326.", + "bbox": [ + 89, + 114, + 485, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 5, 6, 7", + "[2] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2: Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 1, 2, 3", + "[3] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 1, 2", + "[4] Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2", + "[5] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2", + "[6] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[8] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 1, 2", + "[9] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3", + "[10] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3, 5", + "[11] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[12] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + ], + "bbox": [ + 93, + 212, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[14] Gaojie Lin, Jianwen Jiang, Chao Liang, Tianyun Zhong, Ji-aqi Yang, and Yanbo Zheng. Cyberhost: Taming audiodriven avatar diffusion model with region codebook attention. arXiv preprint arXiv:2409.01876, 2024. 1, 2, 3, 5", + "[15] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2", + "[16] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 6", + "[17] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2", + "[18] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 1", + "[19] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 1, 2", + "[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 5, 8", + "[21] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 1, 2", + "[22] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 6", + "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[24] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2", + "[25] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 2", + "[26] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3109", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3", + "[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1, 2", + "[29] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2", + "[30] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1, 2", + "[31] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2, 3, 5, 6", + "[32] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. OpenReview, 2019. 6", + "[33] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 3, 6, 7", + "[34] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2", + "[35] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 6", + "[36] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1, 2, 3, 6, 7", + "[37] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 3, 6", + "[38] Zhenhui Ye, Jinzheng He, Ziyue Jiang, Rongjie Huang, Jiawei Huang, Jinglin Liu, Yi Ren, Xiang Yin, Zejun Ma, and Zhou Zhao. Geneface++: Generalized and stable real-time audio-driven 3d talking face generation. arXiv preprint arXiv:2305.00787, 2023. 1, 2", + "[39] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023.", + "[40] Zhenhui Ye, Tianyun Zhong, Yi Ren, Ziyue Jiang, Jiawei Huang, Rongjie Huang, Jinglin Liu, Jinzheng He, Chen" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhang, Zehan Wang, et al. Mimictalk: Mimicking a personalized and expressive 3d talking face in minutes. arXiv preprint arXiv:2410.06734, 2024. 2", + "[41] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 1, 2", + "[42] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6, 7", + "[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5", + "[44] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 1, 2", + "[45] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Wan Rui, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, et al. Vedit-q: Efficient and accurate quantization of diffusion transformers for image and video generation. arXiv preprint arXiv:2406.02540, 2024. 2", + "[46] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation. arXiv preprint arXiv:2402.19159, 2024. 1, 2, 6", + "[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 1, 2", + "[48] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 1, 2", + "[49] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5" + ], + "bbox": [ + 516, + 92, + 906, + 744 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3110", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_model.json b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c7caf68759275abd505e355b02f9db6fd2eb4797 --- /dev/null +++ b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_model.json @@ -0,0 +1,2552 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "FADA: Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.203, + 0.861, + 0.24 + ], + "angle": 0, + "content": "Tianyun Zhong\\(^{1*†}\\), Chao Liang\\(^{2*}\\), Jianwen Jiang\\(^{2*‡}\\), Gaojie Lin\\(^{2}\\), Jiaqi Yang\\(^{2}\\), Zhou Zhao\\(^{1}\\), Zhejiang University, \\(^{2}\\)Bytedance Intelligent Creation" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.244, + 0.593, + 0.257 + ], + "angle": 0, + "content": "zhongtianyun@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.26, + 0.706, + 0.275 + ], + "angle": 0, + "content": "{liangchao.0412, jianwen.alan, yjq850207131}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.278, + 0.624, + 0.292 + ], + "angle": 0, + "content": "https://fadavatar.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.329, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.697 + ], + "angle": 0, + "content": "Diffusion-based audio-driven talking avatar methods have recently gained attention for their high-fidelity, vivid, and expressive results. However, their slow inference speed limits practical applications. Despite the development of various distillation techniques for diffusion models, we found that naive diffusion distillation methods do not yield satisfactory results. Distilled models exhibit reduced robustness with open-set input images and a decreased correlation between audio and video compared to teacher models, undermining the advantages of diffusion models. To address this, we propose FADA (Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation). We first designed a mixed-supervised loss to leverage data of varying quality and enhance the overall model capability as well as robustness. Additionally, we propose a multi-CFG distillation with learnable tokens to utilize the correlation between audio and reference image conditions, reducing the threefold inference runs caused by multi-CFG with acceptable quality degradation. Extensive experiments across multiple datasets show that FADA generates vivid videos comparable to recent diffusion model-based methods while achieving an NFE speedup of 4.17-12.5 times. Demos are available at our webpage https://fadavatar.github.io." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.225, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.854 + ], + "angle": 0, + "content": "Talking avatar synthesis aims to animate a given portrait image using driven video or audio. Due to the ease of obtaining audio inputs and their low usage barrier, there has been a surge of recent work in this area. Based on the technical route, talking avatar synthesis methods can be categorized into GAN-based[4, 8, 19, 34, 44, 47, 48], NeRF-based[5, 30, 38-41], and the more recent diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.562 + ], + "angle": 0, + "content": "model-based approaches[1, 2, 10, 14, 31, 33, 36]. Diffusion model-based methods, in particular, have gained attention due to their strong generative capabilities. These models exhibit excellent robustness in input images and produce vivid and highly expressive videos that naturally align with the audio. However, the current diffusion model-based talking head generation methods suffer from slow inference speeds, which hinder their practical application. Methods[1, 10, 36] based on Stable Diffusion (SD)[23], all require multiple denoising steps. Moreover, to ensure the correlation between the audio input and the image input in the final video, multi-CFG (Classifier-Free Guidance) inference is often employed, further increasing the runtime. Although most methods achieve 30-40 denoising steps with the help of DDIM[27], they still face speed challenges due to CFG calculations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.782 + ], + "angle": 0, + "content": "Recently, several works have explored diffusion distillation in text-to-image [3, 12, 15, 17, 21, 24, 28, 46] and text-to-video tasks [18] to accelerate inference. Nonetheless, there has been no direct exploration of diffusion distillation for talking avatar synthesis tasks. Our observations indicate that simply applying text-to-image and text-to-video diffusion distillation methods results in a noticeable decline in performance. Unlike text to image or video tasks, talking avatar synthesis involves two key conditions: audio and a reference image. Since audio has a relatively weaker influence on video generation[10, 31], the final output is jointly influenced by both conditions, necessitating careful parameter tuning in multi-CFG settings. The sensitivity to condition guidance often leads to severe artifacts when applying diffusion distillation methods with few-step inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.784, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose FADA (FAst Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation) to address these issues from two perspectives, resulting in an audio-driven talking avatar diffusion model that balances both speed and quality. First, we address the robustness of the distilled model. In this task, the mapping between audio and portrait movement must be learned from data. Previous works have emphasized meticulous data se" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.221, + 0.876 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.333, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Done during an internship at ByteDance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.186, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\)Project Lead" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.863, + 0.333, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.457 + ], + "angle": 0, + "content": "lection [1, 10, 14], such as removing instances with excessive movement or asynchronous audio, to ensure high audio-portrait movement correlation. This stringent selection process excludes a significant amount of data that could enhance robustness. Considering that a well-trained teacher model already provides high audio-portrait correlation guidance, this discarded data can be utilized in training the student model. We designed a mixed-supervised learning strategy that adaptively adjusts the weights between learning from the teacher model's guidance and the student model's own learning from previously discarded data. Secondly, we address the speed issues of multimodal conditions in the distillation process. In audio-driven methods, the final output is achieved through composite CFG inference [10, 14, 36]. To make this process transparent to the student model during training, we introduce a set of learnable tokens. These tokens mimic the weighted calculations of multi-CFG with its coefficients and are injected into the network as control signals, enabling the student model to better learn the relationships between multiple conditions, including those present during inference. Since the learnable token conditions replicate the multi-CFG inference calculation process, we can reduce the number of multi-CFG model inferences during actual inference, further decreasing inference time on top of the denoising step distillation." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.458, + 0.482, + 0.546 + ], + "angle": 0, + "content": "Based on the above considerations and designed methods, we implemented FADA, a diffusion distillation method for audio-driven talking head generation, achieving up to a \\(12.5 \\times\\) speed-up while generating vivid portrait videos comparable to recent undistilled diffusion models. In summary, the contributions of this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.548, + 0.483, + 0.635 + ], + "angle": 0, + "content": "1. We propose FADA, the first diffusion-based distillation framework for audio-driven talking avatar tasks. It includes a mixed-supervised loss to learn from data of varying quality and learnable token conditions to mimic the multi-CFG inference process, reducing performance loss by distillation and multi-CFG inferences." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.636, + 0.483, + 0.68 + ], + "angle": 0, + "content": "2. Quantitative and qualitative experiments demonstrate that FADA matches state-of-the-art generation quality while achieving significantly higher efficiency." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.548, + 0.483, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.694, + 0.24, + 0.709 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.718, + 0.35, + 0.734 + ], + "angle": 0, + "content": "2.1. Audio-Driven Talking Avatar" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.74, + 0.482, + 0.9 + ], + "angle": 0, + "content": "The task of audio-driven talking avatars has gained increasing attention recently, propelled by advancements in videogeneration technologies. GAN-based methods [4, 8, 19, 34, 44, 47, 48] typically employ a two-stage pipeline involving an audio-to-motion module and a motion-to-video module. Sadtalker [42] introduces a 3DMM-based motion representation along with a conditional VAE for generating high-fidelity talking heads. LivePortrait [4] expands the training dataset incorporating images and videos, using implicit keypoints for high-quality generation. NeRF-based methods [5, 30, 38-41] explore animating talking heads in 3D neural" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.18 + ], + "angle": 0, + "content": "spaces. Real3D-Portrait [41] combines a pre-trained image-to-triplane module with a motion adapter for zero-shot and one-reference 3D talking head generation. MimicTalk [40] captures the talking style of individuals through in-context-learning audio-to-motion modules and rapid fine-tuning of triplanes for mimic talking head generation." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.185, + 0.905, + 0.39 + ], + "angle": 0, + "content": "Diffusion-based talking avatar synthesis has swiftly become a crucial and efficient approach [1, 2, 10, 14, 31, 33, 36]. EMO [31] pioneers a double UNet framework in audio-driven avatar creation, yielding impressive and lifelike results. Concurrently, EchoMimic [1] trains on audio and facial landmarks for versatile control inputs. Hallo [36] incorporates a hierarchical audio-driven visual synthesis module to refine audio-visual alignments, while Hallo2 [2] explores image augmentation strategies for 4K avatar video synthesis. Loopy [10] introduces a motion-frame squeeze technique and audio-to-latent training for expressive results. Despite their progress in expressive audio-driven generation, diffusion-based methods still face challenges with slow inference speeds." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.416, + 0.774, + 0.43 + ], + "angle": 0, + "content": "2.2. Diffusion Model Acceleration" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.602 + ], + "angle": 0, + "content": "Diffusion acceleration involves lowering costs for denoising runs and reducing the number of runs. Various denoising methods aim to cut costs, such as utilizing Model quantization [6, 13, 25, 26, 29, 45] to decrease the model parameters directly. ViDiT-Q [47] introduced a new diffusion transformer quantization framework for video and image with mixed-precision quantization. DeepCache [20] takes cues from video compression to propagate reusable result regions through the model's internal structure, while Faster Diffusion [12] emphasizes re-using computed encoder features." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Reducing the number of denoising runs is a direct way to accelerate diffusion models. SD-Turbo[24] leverages the ideas from GAN and introduces adversarial diffusion distillation to enhance the generation fidelity. SnapFusion[12] proposes progressive distillation to achieve fewer steps by learning the previous student model progressively. SwiftBrush[21] and SwiftBrushV2[3] utilizing variational score distillation to make it possible for texts-only distillation in text-to-image tasks. SDXL-lightning[15] mixes progressive distillation and adversarial distillation to achieve one-step image generation with a delicate training process. Applying consistency models [17, 28] to text-to-image tasks also helps reduce diffusion inference steps. LCM[17] views the guided reverse diffusion process as solving an augmented probability flow ODE. TCD[46] proposes trajectory consistency distillation to accurately trace the entire trajectory of the probability flow ODE in semilinear form with an exponential integrator. However, there is no specially-designed distillation framework and techniques for talking avatar synthesis tasks to our knowledge." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3102" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.226, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.233 + ], + "angle": 0, + "content": "In this section, we introduce our proposed method, FADA, a diffusion distillation framework for avatar synthesis, as illustrated in Figure 1. Section 3.1 provides an overview of the diffusion distillation framework, based on recent PeR-Flow [37]. Section 3.2 delves into the specifics of the adaptive mixed-supervised distillation process. In Section 3.3, we present the multiple-CFG distillation technique with a learnable token implementation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.242, + 0.411, + 0.259 + ], + "angle": 0, + "content": "3.1. Overall Framework and Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.484, + 0.512 + ], + "angle": 0, + "content": "Building upon cutting-edge diffusion-based methods for talking avatar synthesis [1, 2, 9, 10, 14, 31, 33, 36], we introduce a dual-Unet architecture for the teacher model, leveraging pre-trained weights from Stable Diffusion[23] 1.5. As depicted in Figure 1, the reference image and motion frames (i.e., images from the last clip) undergo processing by a fixed VAE [11] and reference net. The output from each reference layer is then input to the corresponding layer in the denoising net for spatial attention. To facilitate audio-driven controls and ensure temporal consistency in talking avatar synthesis, audio attention and temporal attention layers are integrated into the denoising net. Initially following the Loopy design [10], we simplify the model by removing the TSM and Audio2Latents modules, streamlining subsequent model distillation. The student model mirrors the teacher's structure, with the addition of a CFG control layer, detailed in Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.629 + ], + "angle": 0, + "content": "We choose the recent PeRFlow [37] as the optimization target for distillation. Initially, we train the teacher model on carefully curated high-quality datasets that prioritize motion amplitude and audio-visual synchronization. The teacher model is optimized using the \\(\\epsilon\\)-prediction DDPM loss [7], where it predicts noise \\(\\epsilon_{\\mathrm{tea}}(\\cdot)\\) from noisy latent \\(z_{t}\\) at timestep \\(t\\). The training objective can be succinctly described as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.64, + 0.483, + 0.66 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {t e a}} = \\mathbb {E} _ {z _ {t}, c, t, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\mathrm {t e a}} (z _ {t}, t, c; \\theta_ {\\mathrm {t e a}}) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.697 + ], + "angle": 0, + "content": "where \\( c \\) includes the audio feature, reference image and motion frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Note that it is not necessary for the teacher model to be trained with flow-matching loss, and both \\(\\epsilon\\)-prediction and \\(v\\)-prediction are valid for PeRFlow distillation. During the student distillation training, firstly same-interval \\(K\\) time windows \\(\\{[t_k, t_{k-1})\\}_{k=K}^1\\) are created where \\(t_k = k / K, k \\in \\{0, 1, \\dots, K\\}\\). A time window \\([t_k, t_{k-1})\\) is randomly sampled out, and the standpoint \\(z_{t_k}\\) can be derived from the marginal distribution of the ground-truth targets, where \\(z_{t_k} = \\sqrt{1 - \\sigma^2(t_k)} z_0 + \\sigma(t_k)\\epsilon\\) and \\(\\sigma(t)\\) denotes the noise schedule. Then the teacher-predicted ending point \\(\\hat{z}_{t_{k-1}}\\) of this time window can be predicted and solved by ODE solver \\(\\Phi(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.886, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t _ {k - 1}} = \\Phi \\left(z _ {t _ {k}}, t _ {k}, t _ {k - 1}; \\theta_ {\\text {t e a}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.196 + ], + "angle": 0, + "content": "where \\(\\Phi (\\cdot)\\) denotes the DDIM solver[27] which performs several inference steps iteratively to reach the endpoint of this time window. Note that the parameters of the teacher model are frozen and no gradient computations are needed here. After that, the input noisy latent of student \\(\\hat{z}_t\\) will be derived from a linear interpolation between starting point \\(t_k\\) and ending point \\(t_{k - 1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.202, + 0.906, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t} = \\frac {z _ {t _ {k}} - \\hat {z} _ {t _ {k - 1}}}{t _ {k} - t _ {k - 1}} (t - t _ {k}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.241, + 0.906, + 0.329 + ], + "angle": 0, + "content": "Through the linear interpolation above, the ODE flow will be rectified to piecewise straight lines so only a few steps are enough to denoising in a time window, which reduces the number of timesteps at inference stage. Meanwhile, we can figure out the target noise \\(\\hat{\\epsilon} (t)\\) via parameterization[37]:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.336, + 0.906, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} (t) = \\frac {\\hat {z} _ {t _ {k - 1}} - \\lambda_ {k} z _ {t _ {k}}}{\\eta_ {k}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.906, + 0.422 + ], + "angle": 0, + "content": "Where \\(\\lambda_{k} = \\sqrt{\\alpha_{t_{k - 1}}} / \\sqrt{\\alpha_{t_{k}}}\\) and \\(\\eta_{k} = \\sqrt{1 - \\alpha_{t_{k - 1}}} -\\) \\(\\sqrt{1 - \\alpha_{t_k}}\\sqrt{\\alpha_{t_{k - 1}}}\\big{/}\\sqrt{\\alpha_{t_k}}\\) respectively. Finally, the training objective of the distillation process can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.432, + 0.908, + 0.462 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i s t i l l}} = \\mathbb {E} _ {k, t \\in [ t _ {k}, t _ {k - 1}), z _ {t}, c, \\epsilon} \\left[ \\| \\hat {\\epsilon} (t) - \\epsilon_ {\\mathrm {s t u}} (\\hat {z} _ {t}, t, c; \\theta_ {\\mathrm {s t u}}) \\| ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.535 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\mathrm{stu}}(\\cdot)\\) refers to the predicted noise from the student model and \\(L_{\\mathrm{distill}}\\) represents the distillation loss. Note that \\(L_{\\mathrm{distill}}\\) is only a trivial part of our framework, the mixed-supervised training and multi-CFG distillation will be proposed in Section 3.2 and 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.853, + 0.559 + ], + "angle": 0, + "content": "3.2. Adaptive Mixed-Supervised Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.682 + ], + "angle": 0, + "content": "In our experiments, we found that distillation significantly degrades the model's performance, especially overall video synthesis quality, as shown in Table 3. Increasing data can improve performance, but in audio-driven portrait generation, the weak control of audio over motion requires the model to learn high-quality motion patterns to fill in details not covered by audio. This necessitates stricter data filtering, reducing the available data quantity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Fortunately, this problem can be mitigated in our proposed distillation process. In our design, we first maintain the teacher training unchanged with only a high-quality dataset \\(\\mathcal{A}\\) to ensure its output stability. Then during the distillation stage, the data filtering strategies will be relaxed and some moderate-quality data samples (a larger dataset \\(\\mathcal{B}\\) that includes \\(\\mathcal{A}\\) where \\(|\\mathcal{B}| > |\\mathcal{A}|\\).) will be fed to student training. Since the student model mimics the teacher model's output, it can maintain the high-quality generation capability of the teacher model even with the inclusion of lower-quality data. To further leverage the new data during the distillation stage, we introduce a ground-truth supervised loss \\(L_{\\mathrm{gt}}\\). The ground-truth noisy latent \\(z_{t}^{*}\\) and ground-truth target noise \\(\\epsilon^{*}(t)\\) can be simply obtained by replacing \\(\\hat{z}_{t_{k-1}}\\) with ground truth ending point \\(z_{t_{k-1}}\\) in For" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.09, + 0.904, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.471, + 0.908, + 0.528 + ], + "angle": 0, + "content": "Figure 1. Overall distillation framework of FADA. The teacher model is trained only with high-quality data, which is omitted in the figure. The student model is trained by a mixed loss of ground-truth and teacher-supervised loss to leverage data of varying quality. Learnable token-based CFG conditions enable the student model to mimic the multi-CFG process, further reducing inference times. For simplicity, we have omitted some components commonly used in previous methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.486, + 0.59 + ], + "angle": 0, + "content": "mula 3 and 4. Similar to distillation loss computation, the ground-truth loss \\( L_{\\mathrm{GT}} \\) and total mixed loss \\( L_{\\mathrm{total}} \\) can be described as" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.601, + 0.484, + 0.633 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {G T}} = \\mathbb {E} _ {k, t \\in \\left[ t _ {k}, t _ {k - 1}\\right), z _ {t}, c, \\epsilon} \\left[ \\| \\epsilon^ {*} (t) - \\epsilon_ {\\mathrm {s t u}} \\left(z _ {t} ^ {*}, t, c; \\theta_ {\\mathrm {s t u}}\\right) \\| _ {2} ^ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.643, + 0.483, + 0.658 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o t a l}} = L _ {\\text {d i s t i l l}} + \\mathcal {W} \\cdot L _ {\\mathrm {G T}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.781 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) represents the loss weight of \\(L_{GT}\\). Through this approach, the student model can better learn high-quality motion pattern generation from the teacher model while also learning more common patterns from a broader range of data to enhance its generalization capability. A trivial implementation idea is to set a fixed hyper-parameter to \\(\\mathcal{W}\\) but this is not very effective because it assumes every sample has the same learning value, which is inaccurate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Hence, an adaptive method to balance teacher-supervised loss \\( L_{\\mathrm{teacher}} \\) and ground-truth-supervised loss \\( L_{\\mathrm{gt}} \\) is necessary, and the potential information behind the two losses should be taken into account. Considering the ratio \\( \\mathcal{R} = L_{\\mathrm{gt}} / L_{\\mathrm{teacher}} \\), it indicates the difference level between the \\( L_{\\mathrm{gt}} \\) and \\( L_{\\mathrm{teacher}} \\). Based on qualitative observations of the \\( \\mathcal{R} \\) value, there are two trends: as \\( \\mathcal{R} \\) increases, the sample is more likely to be a valuable learning case," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.621 + ], + "angle": 0, + "content": "such as a singing video, since it is often outside the teacher-training dataset \\(\\mathcal{A}\\). However, if \\(\\mathcal{R}\\) is too large, the sample is more likely to be low-quality, such as one with poor audio-visual sync, because the teacher results are well synced. The statistical observations are discussed in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.907, + 0.71 + ], + "angle": 0, + "content": "Therefore, we should gradually increase the loss weight when \\(\\mathcal{R}\\) increases in a reasonable range, and also punish the loss weight if \\(\\mathcal{R}\\) increases out of the peak threshold \\(\\mathcal{R}_p\\). For those cases whose \\(\\mathcal{R}\\) is bigger than the dead threshold \\(\\mathcal{R}_d\\), the ground-truth supervised loss weight should be zero. The detailed formulation for loss weight \\(W\\) can be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} = \\mathcal {W} _ {0} \\cdot \\left\\{ \\begin{array}{l l} \\mathcal {R} ^ {s}, & \\mathcal {R} \\in [ 0, \\mathcal {R} _ {p}) \\\\ \\mathcal {R} _ {p} ^ {s} (\\mathcal {R} _ {d} - \\mathcal {R}) / (\\mathcal {R} _ {d} - \\mathcal {R} _ {p}), & \\mathcal {R} \\in [ \\mathcal {R} _ {p}, \\mathcal {R} _ {d}) \\\\ 0, & \\mathcal {R} \\in [ \\mathcal {R} _ {d}, + \\infty) \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.907, + 0.855 + ], + "angle": 0, + "content": "where \\(s, \\mathcal{W}_0, \\mathcal{R}_p\\) and \\(\\mathcal{R}_d\\) are hyper-parameters, they are set to 0.25, 0.2, 30, 100 respectively in this paper. Note that these hyper-parameters are relatively robust across different data distributions because they are used to select data within a relatively moderate range for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Through the proposed design, FADA learns general generation capabilities from more ordinary data during distillation, enhancing model robustness. Our experiments validate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "the proposed loss's effectiveness. Additionally, our adaptive mixed-supervised distillation can extend to tasks like text-to-image or text-to-video without specific restrictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.149, + 0.476, + 0.166 + ], + "angle": 0, + "content": "3.3. Multi-CFG Distillation with Learnable Token" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.483, + 0.333 + ], + "angle": 0, + "content": "In audio-driven talking avatar synthesis tasks, the MultiCFG technique is utilized [10, 14] to achieve robust control of multiple conditions (reference image and audio) effectively. Specifically, the single CFG calculates the direction from unconditional predicted values to a certain conditional predicted value and then moves a certain distance from the unconditional predicted values to obtain conditional predicted values. Furthermore, the MultiCFG recursively computes the next CFG based on the first CFG. In this paper, multiCFG inference result noise \\(\\epsilon_{cfg}\\) can be derived as follows[10]:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.345, + 0.483, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {c f g} = \\operatorname {c f g} _ {a} \\times \\left(\\hat {\\epsilon} _ {a} - \\hat {\\epsilon} _ {r}\\right) + \\operatorname {c f g} _ {r} \\times \\left(\\hat {\\epsilon} _ {r} - \\hat {\\epsilon} _ {b}\\right) + \\hat {\\epsilon} _ {b}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.431 + ], + "angle": 0, + "content": "where \\(\\operatorname{cfg}_a\\) and \\(\\operatorname{cfg}_r\\) indicate the CFG guidance scale of audio and reference condition. \\(\\hat{\\epsilon}_a\\) refers to the noise prediction with both audio and reference conditions, while \\(\\hat{\\epsilon}_r\\) removes the audio condition and \\(\\hat{\\epsilon}_b\\) lacks both audio and reference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.623 + ], + "angle": 0, + "content": "Therefore, during inference, the model needs to run three times for each denoising step, which is time-consuming. It naturally comes to mind to inject the CFG guidance scale as a condition into the network, and then let the student model learn the CFG reasoning characteristics of the teacher model during the distillation process, which echoes a similar concept to the \\(\\omega\\)-condition [20]. Specifically, the teacher model will conduct complete CFG reasoning during training to obtain predictions \\(\\hat{\\epsilon}_{\\mathrm{cfg}}\\) controlled by CFG. Subsequently, these predictions are further integrated into Formulas 2, 3 and 4 to derive input noisy latent \\(\\hat{z}_t^{\\mathrm{cfg}}\\) and target noise \\(\\hat{\\epsilon}^{\\mathrm{cfg}}(t)\\) controlled by CFG, eventually undergoing a similar distillation procedure as Formula 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.845 + ], + "angle": 0, + "content": "However, directly adding the CFG value as a condition to the network does not yield satisfactory results. A careful design is needed for the injection method of the CFG scale condition. By observing the calculation of Multi-CFG reasoning in Formula 9, it is evident that this is akin to performing linear transformations on three independent vectors \\(\\epsilon_{b}\\), \\(\\epsilon_{a}\\), and \\(\\epsilon_{r}\\) in the noise space using the given CFG strengths \\(\\mathrm{cfg}_{a}\\) and \\(\\mathrm{cfg}_{r}\\). Meanwhile, \\(\\epsilon_{r}\\) serves as a crucial link between the \\(\\epsilon_{b}\\) and \\(\\epsilon_{a}\\). Simply embedding the CFG values overlooks the aforementioned relationships. Therefore, we introduce learnable tokens \\(\\gamma_{b}, \\gamma_{r}\\), and \\(\\gamma_{r}\\) into the model to learn the features required for CFG strength control and help the student model mimic the multi-CFG process. To achieve this, we design the CFG embedding \\(Emb_{\\mathrm{cfg}}\\) of the CFG guidance scale \\(\\mathrm{cfg}_{a}\\) and \\(\\mathrm{cfg}_{r}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.858, + 0.483, + 0.875 + ], + "angle": 0, + "content": "\\[\nE m b _ {\\text {c f g}} = \\operatorname {c f g} _ {a} \\times \\left(\\gamma_ {a} - \\gamma_ {r}\\right) + \\operatorname {c f g} _ {r} \\times \\left(\\gamma_ {r} - \\gamma_ {b}\\right) + \\gamma_ {b}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "After obtaining the CFG embedding \\(Emb_{\\mathrm{cfg}}\\), we sim" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.21 + ], + "angle": 0, + "content": "imilarly introduce a CFG layer after each audio layer in the denoising network, injecting the information from \\(Emb_{\\mathrm{cfg}}\\) into the denoising process through cross attention. Importantly, the CFG layer does not introduce excessive performance overhead, especially when compared to the original cost of running three CFG runs. Therefore, our proposed multi-CFG distillation with learnable tokens can achieve nearly three times faster operation speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.646, + 0.243 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.905, + 0.397 + ], + "angle": 0, + "content": "In this section, we will introduce the experimental results. In Section 4.1, we will present the details of the implementation of FADA and the meta-information of our training and testing datasets. In Section 4.2, we will compare our proposed methods with other state-of-the-art diffusion-based talking avatar methods via quantitative and qualitative comparison. In Section 4.3, we will analyze our proposed techniques including the basic distillation method, adaptive mixed-supervised distillation and multi-CFG distillation, and prove their effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.722, + 0.426 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.768 + ], + "angle": 0, + "content": "Datasets Regarding the training dataset, we primarily used speaker videos obtained from the internet, sliced and cropped to durations ranging from two to ten seconds with facial frames, and resized to \\(512 \\times 512\\) resolution. We designed multi-dimensional filters for this dataset, including filters for data source, image quality, audio-visual synchronization, background stability and degree of movement, each with different thresholds to control the filtering intensity. The details of data collection and data filter strategies will be presented in Appendix D. As elucidated in Section 3.2, by using strict filtering thresholds, we obtained approximately 160 hours of high-quality training data \\(\\mathcal{A}\\), which was utilized in teacher pre-training. Additionally, by using lenient filtering thresholds, we obtained around 1300 hours of moderate-quality training data \\(\\mathcal{B}\\) for distillation training. Following the approach of Echomimic [1], we selected 100 random samples from the open-source test sets HDTF [43] and CelebV-HQ [49] to validate the model's performance on various real human scenarios. Furthermore, to evaluate the model's performance in more open and extreme scenarios, we adopted the openset test set from Loopy [10] for qualitative comparisons, which contains various kinds of portraits and audios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation Details The training process consists of teacher pre-training and student distillation. The training process of the teacher model closely aligns with the first two stages of recent methods[10, 31] with high-quality dataset \\(\\mathcal{A}\\). In student distillation with the moderate-quality dataset \\(\\mathcal{B}\\), the student model is initialized from the teacher model and integrated with CFG learnable tokens and CFG layers. During testing, the CFG guidance scale is set to 2.5 and 6.5 for reference and audio conditions in models without CFG" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.196 + ], + "angle": 0, + "content": "distillation, while the scales were set to 2.0 and 6.5 for models with CFG distillation. The number of time windows in PeRFlow[37] is 4. All results of FADA are conducted with 6-step inferences, where the two time-windows near \\( T = 1 \\) contain two inference steps while the others contain one inference step. More details about the teacher and student training will be discussed in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.201, + 0.396, + 0.216 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.222, + 0.485, + 0.515 + ], + "angle": 0, + "content": "Metrics and Baselines For the purpose of quantitative comparisons, we evaluate the generated talking avatar videos with several metrics. We assess the image quality with IQA[35] metrics and evaluate the audio-visual synchronization distance with a pre-trained SyncNet[22] as Sync-D. FID and FVD[32] (with 3D ResNet) are assessed to evaluate the image/video-level fidelity difference between generated results and ground truths in the image inception space. E-FID (Expression-FID)[31] measures the inception distance between the generated and ground truths in expression parameter space. NFE-D indicates the number of function evaluations of the denoising net. For instance, our teacher model conducts 25-step DDIM inference with 3 CFG runs, leading to 75 NFE-D. Since our proposed FADA is a diffusion distillation framework, we mainly compare recent open-source diffusion-based state-of-the-art methods in talking portrait generation, including Hallo[36], EchoMimic[1], and V-Express[33]. Additionally, for reference, we also include the state-of-the-art GAN-based method SadTalker[42]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.518, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Quantitative Comparison As shown in Table 1, we conduct quantitative comparison experiments with CelebV-HQ and HDTF test sets. Since the differences in IQA are subtle, we provide \\(95\\%\\) confidence intervals to show their statistical significance. Ours-Balanced refers to our proposed method with mixed-supervised distillation with PeRFlow loss. Based on it, Ours-Fast utilizes multi-CFG distillation to achieve about 3-time speedup and gets a slight performance drop. Ours-Balanced achieves better IQA, FVD, FID, and E-FID on the challenging open-set CelebVHQ test set compared to the baselines. It also leads or is comparable in most metrics on the simpler HDTF test set. Ours-Fast maintains similar performance in IQA and Sync-D metrics with a slight drop in FVD, FID, and E-FID, which is still comparable to the baselines in most metrics. In terms of audio synchronization, the trend is similar. Notably, VExpress results show almost only lip movements, which, despite good metrics, do not produce satisfactory video quality. Finally, in terms of speedup, it is evident that both Ours-Balanced and Ours-Fast achieve significant speedup using the same SD base model as the baselines. The subjective evaluation will be conducted in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.842, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparison As illustrated in Figure 2 and in the supplementary video materials, Ours-Fast and Ours-Balanced maintain good video quality and vivid audiovisual expressiveness. We highly recommend readers watch" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.268 + ], + "angle": 0, + "content": "the supplementary videos since it is difficult for the static image to show the audio-motion synchronization and temporal consistency. V-Express shows color differences and inferior identity-preserving abilities in all cases. Sadtalker generates good identity-preserving videos at the image level, but the head motions and lip movements are not well-synced with the emotion and pace of the given audio. EchoMimic and Hallo generate acceptable expressions, but there are temporal jitters that damage the visual quality. Moreover, by observing case-4 and case-6, we can find that FADA produces more full and plump pronunciations than baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.276, + 0.667, + 0.292 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.909, + 0.56 + ], + "angle": 0, + "content": "Analysis of Distillation Methods. As illustrated in Table 2, we first compared various basic methods of diffusion distillation, including TCD [46], RFlow [16], and PeRFlow [37]. It was observed that compared to the TCD method, the PeRFlow method showed similar performance in IQA, FVD, Sync-D, and FID metrics, but with a significant advantage in E-FID. Additionally, the basic RFlow method could not handle the 6-step inference in our task setting. Therefore, we selected PeRFlow as the basic distillation method. We also tried simply adding moderate-quality data during teacher model training, which led to a decrease in model performance. These results indicate that existing distillation methods cannot directly produce a high-quality student model. Additionally, attempting to enhance the teacher model's capabilities by relaxing data filtering criteria and increasing data volume does not yield positive results. The comparison with the adversarial-based method is shown in Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.886 + ], + "angle": 0, + "content": "Analysis of Adaptive Mixed-Supervised Distillation Here, based on the previously selected PeRFlow, we added dataset \\(\\mathcal{B}\\) and a mixed-supervised loss to investigate the effectiveness of our Adaptive Mixed-Supervised Distillation. As shown in Table 4, below the hdashline, we provide the performance of Mixed-Supervised Distillation with the addition of dataset \\(\\mathcal{B}\\) under different parameters. Comparing the effects of different Fixed-x values, it is evident that the ground-truth loss improves model performance, and a smaller loss weight value of 0.2 yields better results. We further investigated the effectiveness of the adaptive loss weight strategy. In the adaptive-unl. setting (unlimited, without using peak and dead thresholds), the adaptive loss weight increases infinitely as \\(\\mathcal{R}\\) increases, resulting in slightly inferior performance compared to the Fixed-0.2 setting. However, when employing the complete adaptive strategy, the ground-truth loss of low-quality samples is automatically discarded, leading to a distillation effect that surpasses the Fixed setting and approaches the performance of the teacher model, especially in terms of the FVD metrics. We also compared different values of \\(s\\) and ultimately chose 0.25 due to its slight advantage." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Additionally, by comparing the performance of teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3106" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with state-of-the-art methods on CelebV-HQ & HDTF test sets. Bold and underlined numbers refer to the best and the second-best result in this test set." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.129, + 0.917, + 0.273 + ], + "angle": 0, + "content": "
MethodNFE-DCelebV-HQ TestHDTF Test
IQA↑Sync-D↓FVD-R↓FID↓E-FID↓IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Sadtalker[42]-2.953±.0868.765171.836.642.2483.435±.0697.87024.9325.351.559
Hallo[36]803.505±.0919.07953.9935.962.4263.922±.0797.91721.7120.151.337
EchoMimic[1]603.307±.10410.3754.7135.373.0183.994±.0839.39118.7119.011.328
V-Express[33]502.946±.1239.415117.865.092.4143.482±.0717.38248.0330.911.506
Ours-Balanced183.588±.0918.99850.3633.672.2023.951±.0827.92513.3218.391.362
Ours-Fast63.550±.0938.74654.6935.012.6043.927±.0817.99616.6718.511.635
" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.294, + 0.221, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.368, + 0.221, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.439, + 0.221, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.512, + 0.221, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.585, + 0.221, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.658, + 0.221, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.735, + 0.203, + 0.749 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.295, + 0.325, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.368, + 0.325, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.44, + 0.325, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.512, + 0.325, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.584, + 0.325, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.658, + 0.325, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.735, + 0.318, + 0.749 + ], + "angle": 0, + "content": "SadTalker" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.295, + 0.421, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.368, + 0.421, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.44, + 0.421, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.512, + 0.421, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.584, + 0.421, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.658, + 0.421, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.735, + 0.411, + 0.751 + ], + "angle": 0, + "content": "VExpress" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.295, + 0.518, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.368, + 0.518, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.44, + 0.518, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.512, + 0.518, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.584, + 0.518, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.658, + 0.518, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.735, + 0.496, + 0.749 + ], + "angle": 0, + "content": "Hallo" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.295, + 0.616, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.368, + 0.616, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.44, + 0.616, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.512, + 0.616, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.584, + 0.616, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.658, + 0.616, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.735, + 0.612, + 0.749 + ], + "angle": 0, + "content": "EchoMimic" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.295, + 0.712, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.368, + 0.712, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.44, + 0.712, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.512, + 0.712, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.584, + 0.712, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.658, + 0.712, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.735, + 0.702, + 0.749 + ], + "angle": 0, + "content": "Ours-Fast" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.295, + 0.808, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.368, + 0.808, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.44, + 0.808, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.512, + 0.808, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.584, + 0.808, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.658, + 0.808, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.735, + 0.818, + 0.749 + ], + "angle": 0, + "content": "Ours-Balanced" + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.319, + 0.863, + 0.345 + ], + "angle": 0, + "content": "[ə]" + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.387, + 0.859, + 0.413 + ], + "angle": 0, + "content": "[I]" + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.462, + 0.872, + 0.488 + ], + "angle": 0, + "content": "[u:]" + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.535, + 0.874, + 0.561 + ], + "angle": 0, + "content": "[æ]" + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.605, + 0.872, + 0.631 + ], + "angle": 0, + "content": "[eI]" + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.676, + 0.862, + 0.704 + ], + "angle": 0, + "content": "[1]" + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.757, + 0.876, + 0.772 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparisons between FADA and baselines across different portraits and pronunciations in openset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.901 + ], + "angle": 0, + "content": "and student models under different data qualities and scales, it can be demonstrated that mixed supervised distillation actually enhances the model's utilization of moderate-quality data. As depicted in Tables 2 and 3, when the teacher model is trained with a large-scale moderate-quality dataset, a decline in model performance is observed, whereas the student model exhibits the opposite trend. Briefly, only mixed-supervised distillation can take advantage of the moderate-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.784, + 0.907, + 0.829 + ], + "angle": 0, + "content": "quality data. More analysis of mixed-supervised distillation will be discussed and more hyper-parameters experiments will be conducted in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Analysis of Multi-CFG Distillation As demonstrated in Table 4, we further explore the effects of multi-CFG distillation on top of mixed-supervised distillation. Initially, we tested the results of conditional reasoning without CFG dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3107" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 2. Ablation study of basic distillation methods on HDTF test set. All methods in this table are free from mixed-supervised and CFG distillation." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.143, + 0.472, + 0.232 + ], + "angle": 0, + "content": "
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Teacher with B753.8858.12319.6421.401.689
TCD183.8277.93723.0123.151.893
RFlow182.8368.06681.1049.732.484
PeRFlow183.8238.00121.9121.261.477
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.247, + 0.485, + 0.331 + ], + "angle": 0, + "content": "Table 3. Ablation study of adaptive mixed-supervised distillation on HDTF test set. All methods use PeRFlow with dataset \\(\\mathcal{B}\\) without CFG distillation. Fixed-x indicates fixed ground-truth loss with weight x. Adaptive-unlimited-s and Adaptive-s indicate adaptive ground-truth loss with hyper-parameter \\(s\\), with the former being without peak and dead threshold." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.335, + 0.479, + 0.484 + ], + "angle": 0, + "content": "
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Stu. with A183.8238.00121.9121.261.477
Stu. with B (Fixed-0)183.7627.90925.0621.471.462
+ Fixed-0.2183.8637.93413.7119.941.434
+ Fixed-1.0183.8127.90324.5321.421.494
+ Fixed-0.2, only A183.8807.96215.3220.711.512
+ Adaptive-unl.-0.5183.8247.91415.1520.741.471
+ Adaptive-unl.-0.25183.8467.90614.2019.921.424
+ Adaptive-0.5183.9497.91913.8318.641.395
+ Adaptive-0.25183.9517.92513.3218.391.362
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.555 + ], + "angle": 0, + "content": "Table 4. Ablation study of multi-CFG distillation on HDTF test set. All methods in this table are PeRFlow + mixed-supervised distillation. CFG time embed refers to CFG distillation through timestep embedding injection." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.558, + 0.48, + 0.648 + ], + "angle": 0, + "content": "
MethodNFE-D IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
w.o. CFG63.6749.02027.1026.562.096
+ CFG time embed63.7018.57723.2422.471.791
+ CFG layer63.7467.90620.0520.941.620
+ CFG layer w. tokens63.9277.99616.6718.511.635
" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.651, + 0.466, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.771, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Figure 3. Line charts showing the variations of FVD and Sync-D metrics with different audio CFG on HDTF test set. Reference CFG is set to 2.0 in this figure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.484, + 0.903 + ], + "angle": 0, + "content": "tillation, which significantly deteriorated the model performance, especially leading to a doubling of the FVD metric representing video quality. For techniques like time embedding injection similar to \\(\\omega\\)-condition [20], they partially restored model performance but with limited efficacy. In con" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.091, + 0.891, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.213, + 0.908, + 0.257 + ], + "angle": 0, + "content": "Figure 4. Line charts showing the variations of FVD and Sync-D metrics with different ref CFG on HDTF test set. Audio CFG is set to 6.5 in this figure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.907, + 0.397 + ], + "angle": 0, + "content": "trast, our proposed learnable token-based CFG control layers significantly enhance model performance, nearly matching the teacher model and even surpassing it in FVD. Notably, our model requires only \\(8\\%\\) of the teacher model's NFE. We also tested without learnable tokens, and while the pure CFG control layer had some effect, it could not match the teacher model, especially in overall video quality metrics like FVD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.909, + 0.636 + ], + "angle": 0, + "content": "To verify whether our proposed multi-CFG distillation effectively captures the influence of audio and reference conditions on the generated results, we plotted line charts of FVD and Sync-D metrics with variations in audio CFG and reference CFG, as shown in Figures 3 and 4. First, we observed improvements in both FVD and Sync-D as audio CFG increased, reaching a stable maximum value after 6.5. Hence, we ultimately chose 6.5 as the audio CFG for inference. Second, the influence of reference CFG on FVD and Sync-D metrics displayed an initially positive, then negative trend, with optimal performance achieved around the 2.0 position. These results indicate that our CFG control layer effectively mimics the multi-CFG process and achieves good results. More visualized examples and performance analysis under different scenes will be presented in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.665, + 0.634, + 0.681 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In summary, we proposed FADA, a fast diffusion avatar synthesis framework with mixed-supervised multi-CFG distillation. By designing a mixed-supervised loss, we leverage data of varying quality to enhance the robustness of generated results. Furthermore, a learnable token-based multi-CFG condition design is introduced to maintain the correlation between the audio and the generated video in the distilled models. The quantitative and qualitative experiments across multiple datasets show that our balanced setting achieves fast inference and high-fidelity generation ability. Ablation studies prove our proposed methods are effective in FADA. The limitations, future works and ethical concerns of this paper will be discussed in Appendix B, H and I respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3108" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.175 + ], + "angle": 0, + "content": "This project is sponsored by National Natural Science Foundation of China under Grant No.62222211 and National Natural Science Foundation of China under Grant No.U24A20326." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[1] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.269, + 0.484, + 0.334 + ], + "angle": 0, + "content": "[2] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2: Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.337, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[3] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.406, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[4] Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.461, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[5] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.484, + 0.583 + ], + "angle": 0, + "content": "[6] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.484, + 0.625 + ], + "angle": 0, + "content": "[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.627, + 0.484, + 0.693 + ], + "angle": 0, + "content": "[8] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.695, + 0.484, + 0.749 + ], + "angle": 0, + "content": "[9] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.75, + 0.483, + 0.804 + ], + "angle": 0, + "content": "[10] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.806, + 0.483, + 0.833 + ], + "angle": 0, + "content": "[11] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.834, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.146 + ], + "angle": 0, + "content": "[13] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[14] Gaojie Lin, Jianwen Jiang, Chao Liang, Tianyun Zhong, Ji-aqi Yang, and Yanbo Zheng. Cyberhost: Taming audiodriven avatar diffusion model with region codebook attention. arXiv preprint arXiv:2409.01876, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[16] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[17] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[18] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.402, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[19] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[21] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.906, + 0.676 + ], + "angle": 0, + "content": "[22] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.747 + ], + "angle": 0, + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[24] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[25] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3109" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[30] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.36, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[32] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. OpenReview, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.415, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[33] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[34] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[35] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[36] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[37] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[38] Zhenhui Ye, Jinzheng He, Ziyue Jiang, Rongjie Huang, Jiawei Huang, Jinglin Liu, Yi Ren, Xiang Yin, Zejun Ma, and Zhou Zhao. Geneface++: Generalized and stable real-time audio-driven 3d talking face generation. arXiv preprint arXiv:2305.00787, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[39] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Zhenhui Ye, Tianyun Zhong, Yi Ren, Ziyue Jiang, Jiawei Huang, Rongjie Huang, Jinglin Liu, Jinzheng He, Chen" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Zhang, Zehan Wang, et al. Mimictalk: Mimicking a personalized and expressive 3d talking face in minutes. arXiv preprint arXiv:2410.06734, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.134, + 0.905, + 0.2 + ], + "angle": 0, + "content": "[41] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.203, + 0.905, + 0.282 + ], + "angle": 0, + "content": "[42] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.284, + 0.907, + 0.351 + ], + "angle": 0, + "content": "[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.353, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[44] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[45] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Wan Rui, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, et al. Vedit-q: Efficient and accurate quantization of diffusion transformers for image and video generation. arXiv preprint arXiv:2406.02540, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.542 + ], + "angle": 0, + "content": "[46] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation. arXiv preprint arXiv:2402.19159, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.623 + ], + "angle": 0, + "content": "[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.625, + 0.905, + 0.679 + ], + "angle": 0, + "content": "[48] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.681, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[49] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3110" + } + ] +] \ No newline at end of file diff --git a/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_origin.pdf b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64bb06c5bda8bf38b5560884f416fb9c9daefaa0 --- /dev/null +++ b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41cba168efa07718bc71772a5b39549edf06840bea0048957a04fcc1ee5c9afb +size 1638565 diff --git a/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/full.md b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fe4353d8390422daeca1d99cb6f97434ecbfc28e --- /dev/null +++ b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/full.md @@ -0,0 +1,373 @@ +# FADA: Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation + +Tianyun Zhong $^{1*†}$ , Chao Liang $^{2*}$ , Jianwen Jiang $^{2*‡}$ , Gaojie Lin $^{2}$ , Jiaqi Yang $^{2}$ , Zhou Zhao $^{1}$ , Zhejiang University, $^{2}$ Bytedance Intelligent Creation + +zhongtianyun@zju.edu.cn + +{liangchao.0412, jianwen.alan, yjq850207131}@gmail.com + +https://fadavatar.github.io/ + +# Abstract + +Diffusion-based audio-driven talking avatar methods have recently gained attention for their high-fidelity, vivid, and expressive results. However, their slow inference speed limits practical applications. Despite the development of various distillation techniques for diffusion models, we found that naive diffusion distillation methods do not yield satisfactory results. Distilled models exhibit reduced robustness with open-set input images and a decreased correlation between audio and video compared to teacher models, undermining the advantages of diffusion models. To address this, we propose FADA (Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation). We first designed a mixed-supervised loss to leverage data of varying quality and enhance the overall model capability as well as robustness. Additionally, we propose a multi-CFG distillation with learnable tokens to utilize the correlation between audio and reference image conditions, reducing the threefold inference runs caused by multi-CFG with acceptable quality degradation. Extensive experiments across multiple datasets show that FADA generates vivid videos comparable to recent diffusion model-based methods while achieving an NFE speedup of 4.17-12.5 times. Demos are available at our webpage https://fadavatar.github.io. + +# 1. Introduction + +Talking avatar synthesis aims to animate a given portrait image using driven video or audio. Due to the ease of obtaining audio inputs and their low usage barrier, there has been a surge of recent work in this area. Based on the technical route, talking avatar synthesis methods can be categorized into GAN-based[4, 8, 19, 34, 44, 47, 48], NeRF-based[5, 30, 38-41], and the more recent diffusion + +model-based approaches[1, 2, 10, 14, 31, 33, 36]. Diffusion model-based methods, in particular, have gained attention due to their strong generative capabilities. These models exhibit excellent robustness in input images and produce vivid and highly expressive videos that naturally align with the audio. However, the current diffusion model-based talking head generation methods suffer from slow inference speeds, which hinder their practical application. Methods[1, 10, 36] based on Stable Diffusion (SD)[23], all require multiple denoising steps. Moreover, to ensure the correlation between the audio input and the image input in the final video, multi-CFG (Classifier-Free Guidance) inference is often employed, further increasing the runtime. Although most methods achieve 30-40 denoising steps with the help of DDIM[27], they still face speed challenges due to CFG calculations. + +Recently, several works have explored diffusion distillation in text-to-image [3, 12, 15, 17, 21, 24, 28, 46] and text-to-video tasks [18] to accelerate inference. Nonetheless, there has been no direct exploration of diffusion distillation for talking avatar synthesis tasks. Our observations indicate that simply applying text-to-image and text-to-video diffusion distillation methods results in a noticeable decline in performance. Unlike text to image or video tasks, talking avatar synthesis involves two key conditions: audio and a reference image. Since audio has a relatively weaker influence on video generation[10, 31], the final output is jointly influenced by both conditions, necessitating careful parameter tuning in multi-CFG settings. The sensitivity to condition guidance often leads to severe artifacts when applying diffusion distillation methods with few-step inference. + +In this paper, we propose FADA (FAst Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation) to address these issues from two perspectives, resulting in an audio-driven talking avatar diffusion model that balances both speed and quality. First, we address the robustness of the distilled model. In this task, the mapping between audio and portrait movement must be learned from data. Previous works have emphasized meticulous data se + +lection [1, 10, 14], such as removing instances with excessive movement or asynchronous audio, to ensure high audio-portrait movement correlation. This stringent selection process excludes a significant amount of data that could enhance robustness. Considering that a well-trained teacher model already provides high audio-portrait correlation guidance, this discarded data can be utilized in training the student model. We designed a mixed-supervised learning strategy that adaptively adjusts the weights between learning from the teacher model's guidance and the student model's own learning from previously discarded data. Secondly, we address the speed issues of multimodal conditions in the distillation process. In audio-driven methods, the final output is achieved through composite CFG inference [10, 14, 36]. To make this process transparent to the student model during training, we introduce a set of learnable tokens. These tokens mimic the weighted calculations of multi-CFG with its coefficients and are injected into the network as control signals, enabling the student model to better learn the relationships between multiple conditions, including those present during inference. Since the learnable token conditions replicate the multi-CFG inference calculation process, we can reduce the number of multi-CFG model inferences during actual inference, further decreasing inference time on top of the denoising step distillation. + +Based on the above considerations and designed methods, we implemented FADA, a diffusion distillation method for audio-driven talking head generation, achieving up to a $12.5 \times$ speed-up while generating vivid portrait videos comparable to recent undistilled diffusion models. In summary, the contributions of this paper include: + +1. We propose FADA, the first diffusion-based distillation framework for audio-driven talking avatar tasks. It includes a mixed-supervised loss to learn from data of varying quality and learnable token conditions to mimic the multi-CFG inference process, reducing performance loss by distillation and multi-CFG inferences. +2. Quantitative and qualitative experiments demonstrate that FADA matches state-of-the-art generation quality while achieving significantly higher efficiency. + +# 2. Related Works + +# 2.1. Audio-Driven Talking Avatar + +The task of audio-driven talking avatars has gained increasing attention recently, propelled by advancements in videogeneration technologies. GAN-based methods [4, 8, 19, 34, 44, 47, 48] typically employ a two-stage pipeline involving an audio-to-motion module and a motion-to-video module. Sadtalker [42] introduces a 3DMM-based motion representation along with a conditional VAE for generating high-fidelity talking heads. LivePortrait [4] expands the training dataset incorporating images and videos, using implicit keypoints for high-quality generation. NeRF-based methods [5, 30, 38-41] explore animating talking heads in 3D neural + +spaces. Real3D-Portrait [41] combines a pre-trained image-to-triplane module with a motion adapter for zero-shot and one-reference 3D talking head generation. MimicTalk [40] captures the talking style of individuals through in-context-learning audio-to-motion modules and rapid fine-tuning of triplanes for mimic talking head generation. + +Diffusion-based talking avatar synthesis has swiftly become a crucial and efficient approach [1, 2, 10, 14, 31, 33, 36]. EMO [31] pioneers a double UNet framework in audio-driven avatar creation, yielding impressive and lifelike results. Concurrently, EchoMimic [1] trains on audio and facial landmarks for versatile control inputs. Hallo [36] incorporates a hierarchical audio-driven visual synthesis module to refine audio-visual alignments, while Hallo2 [2] explores image augmentation strategies for 4K avatar video synthesis. Loopy [10] introduces a motion-frame squeeze technique and audio-to-latent training for expressive results. Despite their progress in expressive audio-driven generation, diffusion-based methods still face challenges with slow inference speeds. + +# 2.2. Diffusion Model Acceleration + +Diffusion acceleration involves lowering costs for denoising runs and reducing the number of runs. Various denoising methods aim to cut costs, such as utilizing Model quantization [6, 13, 25, 26, 29, 45] to decrease the model parameters directly. ViDiT-Q [47] introduced a new diffusion transformer quantization framework for video and image with mixed-precision quantization. DeepCache [20] takes cues from video compression to propagate reusable result regions through the model's internal structure, while Faster Diffusion [12] emphasizes re-using computed encoder features. + +Reducing the number of denoising runs is a direct way to accelerate diffusion models. SD-Turbo[24] leverages the ideas from GAN and introduces adversarial diffusion distillation to enhance the generation fidelity. SnapFusion[12] proposes progressive distillation to achieve fewer steps by learning the previous student model progressively. SwiftBrush[21] and SwiftBrushV2[3] utilizing variational score distillation to make it possible for texts-only distillation in text-to-image tasks. SDXL-lightning[15] mixes progressive distillation and adversarial distillation to achieve one-step image generation with a delicate training process. Applying consistency models [17, 28] to text-to-image tasks also helps reduce diffusion inference steps. LCM[17] views the guided reverse diffusion process as solving an augmented probability flow ODE. TCD[46] proposes trajectory consistency distillation to accurately trace the entire trajectory of the probability flow ODE in semilinear form with an exponential integrator. However, there is no specially-designed distillation framework and techniques for talking avatar synthesis tasks to our knowledge. + +# 3. Methodology + +In this section, we introduce our proposed method, FADA, a diffusion distillation framework for avatar synthesis, as illustrated in Figure 1. Section 3.1 provides an overview of the diffusion distillation framework, based on recent PeR-Flow [37]. Section 3.2 delves into the specifics of the adaptive mixed-supervised distillation process. In Section 3.3, we present the multiple-CFG distillation technique with a learnable token implementation. + +# 3.1. Overall Framework and Preliminary + +Building upon cutting-edge diffusion-based methods for talking avatar synthesis [1, 2, 9, 10, 14, 31, 33, 36], we introduce a dual-Unet architecture for the teacher model, leveraging pre-trained weights from Stable Diffusion[23] 1.5. As depicted in Figure 1, the reference image and motion frames (i.e., images from the last clip) undergo processing by a fixed VAE [11] and reference net. The output from each reference layer is then input to the corresponding layer in the denoising net for spatial attention. To facilitate audio-driven controls and ensure temporal consistency in talking avatar synthesis, audio attention and temporal attention layers are integrated into the denoising net. Initially following the Loopy design [10], we simplify the model by removing the TSM and Audio2Latents modules, streamlining subsequent model distillation. The student model mirrors the teacher's structure, with the addition of a CFG control layer, detailed in Section 3.3. + +We choose the recent PeRFlow [37] as the optimization target for distillation. Initially, we train the teacher model on carefully curated high-quality datasets that prioritize motion amplitude and audio-visual synchronization. The teacher model is optimized using the $\epsilon$ -prediction DDPM loss [7], where it predicts noise $\epsilon_{\mathrm{tea}}(\cdot)$ from noisy latent $z_{t}$ at timestep $t$ . The training objective can be succinctly described as: + +$$ +L _ {\mathrm {t e a}} = \mathbb {E} _ {z _ {t}, c, t, \epsilon} \left[ \| \epsilon - \epsilon_ {\mathrm {t e a}} (z _ {t}, t, c; \theta_ {\mathrm {t e a}}) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $c$ includes the audio feature, reference image and motion frames. + +Note that it is not necessary for the teacher model to be trained with flow-matching loss, and both $\epsilon$ -prediction and $v$ -prediction are valid for PeRFlow distillation. During the student distillation training, firstly same-interval $K$ time windows $\{[t_k, t_{k-1})\}_{k=K}^1$ are created where $t_k = k / K, k \in \{0, 1, \dots, K\}$ . A time window $[t_k, t_{k-1})$ is randomly sampled out, and the standpoint $z_{t_k}$ can be derived from the marginal distribution of the ground-truth targets, where $z_{t_k} = \sqrt{1 - \sigma^2(t_k)} z_0 + \sigma(t_k)\epsilon$ and $\sigma(t)$ denotes the noise schedule. Then the teacher-predicted ending point $\hat{z}_{t_{k-1}}$ of this time window can be predicted and solved by ODE solver $\Phi(\cdot)$ : + +$$ +\hat {z} _ {t _ {k - 1}} = \Phi \left(z _ {t _ {k}}, t _ {k}, t _ {k - 1}; \theta_ {\text {t e a}}\right), \tag {2} +$$ + +where $\Phi (\cdot)$ denotes the DDIM solver[27] which performs several inference steps iteratively to reach the endpoint of this time window. Note that the parameters of the teacher model are frozen and no gradient computations are needed here. After that, the input noisy latent of student $\hat{z}_t$ will be derived from a linear interpolation between starting point $t_k$ and ending point $t_{k - 1}$ : + +$$ +\hat {z} _ {t} = \frac {z _ {t _ {k}} - \hat {z} _ {t _ {k - 1}}}{t _ {k} - t _ {k - 1}} (t - t _ {k}). \tag {3} +$$ + +Through the linear interpolation above, the ODE flow will be rectified to piecewise straight lines so only a few steps are enough to denoising in a time window, which reduces the number of timesteps at inference stage. Meanwhile, we can figure out the target noise $\hat{\epsilon} (t)$ via parameterization[37]: + +$$ +\hat {\epsilon} (t) = \frac {\hat {z} _ {t _ {k - 1}} - \lambda_ {k} z _ {t _ {k}}}{\eta_ {k}}, \tag {4} +$$ + +Where $\lambda_{k} = \sqrt{\alpha_{t_{k - 1}}} / \sqrt{\alpha_{t_{k}}}$ and $\eta_{k} = \sqrt{1 - \alpha_{t_{k - 1}}} -$ $\sqrt{1 - \alpha_{t_k}}\sqrt{\alpha_{t_{k - 1}}}\big{/}\sqrt{\alpha_{t_k}}$ respectively. Finally, the training objective of the distillation process can be obtained as + +$$ +L _ {\text {d i s t i l l}} = \mathbb {E} _ {k, t \in [ t _ {k}, t _ {k - 1}), z _ {t}, c, \epsilon} \left[ \| \hat {\epsilon} (t) - \epsilon_ {\mathrm {s t u}} (\hat {z} _ {t}, t, c; \theta_ {\mathrm {s t u}}) \| ^ {2} \right], \tag {5} +$$ + +where $\epsilon_{\mathrm{stu}}(\cdot)$ refers to the predicted noise from the student model and $L_{\mathrm{distill}}$ represents the distillation loss. Note that $L_{\mathrm{distill}}$ is only a trivial part of our framework, the mixed-supervised training and multi-CFG distillation will be proposed in Section 3.2 and 3.3. + +# 3.2. Adaptive Mixed-Supervised Distillation + +In our experiments, we found that distillation significantly degrades the model's performance, especially overall video synthesis quality, as shown in Table 3. Increasing data can improve performance, but in audio-driven portrait generation, the weak control of audio over motion requires the model to learn high-quality motion patterns to fill in details not covered by audio. This necessitates stricter data filtering, reducing the available data quantity. + +Fortunately, this problem can be mitigated in our proposed distillation process. In our design, we first maintain the teacher training unchanged with only a high-quality dataset $\mathcal{A}$ to ensure its output stability. Then during the distillation stage, the data filtering strategies will be relaxed and some moderate-quality data samples (a larger dataset $\mathcal{B}$ that includes $\mathcal{A}$ where $|\mathcal{B}| > |\mathcal{A}|$ .) will be fed to student training. Since the student model mimics the teacher model's output, it can maintain the high-quality generation capability of the teacher model even with the inclusion of lower-quality data. To further leverage the new data during the distillation stage, we introduce a ground-truth supervised loss $L_{\mathrm{gt}}$ . The ground-truth noisy latent $z_{t}^{*}$ and ground-truth target noise $\epsilon^{*}(t)$ can be simply obtained by replacing $\hat{z}_{t_{k-1}}$ with ground truth ending point $z_{t_{k-1}}$ in For + +![](images/cc474c8bd90a591ad224fc6e3f96fe418ae127a148d1f9a8109d9a57cb648be5.jpg) +Figure 1. Overall distillation framework of FADA. The teacher model is trained only with high-quality data, which is omitted in the figure. The student model is trained by a mixed loss of ground-truth and teacher-supervised loss to leverage data of varying quality. Learnable token-based CFG conditions enable the student model to mimic the multi-CFG process, further reducing inference times. For simplicity, we have omitted some components commonly used in previous methods. + +mula 3 and 4. Similar to distillation loss computation, the ground-truth loss $L_{\mathrm{GT}}$ and total mixed loss $L_{\mathrm{total}}$ can be described as + +$$ +L _ {\mathrm {G T}} = \mathbb {E} _ {k, t \in \left[ t _ {k}, t _ {k - 1}\right), z _ {t}, c, \epsilon} \left[ \| \epsilon^ {*} (t) - \epsilon_ {\mathrm {s t u}} \left(z _ {t} ^ {*}, t, c; \theta_ {\mathrm {s t u}}\right) \| _ {2} ^ {2} \right], \tag {6} +$$ + +$$ +L _ {\text {t o t a l}} = L _ {\text {d i s t i l l}} + \mathcal {W} \cdot L _ {\mathrm {G T}}, \tag {7} +$$ + +where $\mathcal{W}$ represents the loss weight of $L_{GT}$ . Through this approach, the student model can better learn high-quality motion pattern generation from the teacher model while also learning more common patterns from a broader range of data to enhance its generalization capability. A trivial implementation idea is to set a fixed hyper-parameter to $\mathcal{W}$ but this is not very effective because it assumes every sample has the same learning value, which is inaccurate. + +Hence, an adaptive method to balance teacher-supervised loss $L_{\mathrm{teacher}}$ and ground-truth-supervised loss $L_{\mathrm{gt}}$ is necessary, and the potential information behind the two losses should be taken into account. Considering the ratio $\mathcal{R} = L_{\mathrm{gt}} / L_{\mathrm{teacher}}$ , it indicates the difference level between the $L_{\mathrm{gt}}$ and $L_{\mathrm{teacher}}$ . Based on qualitative observations of the $\mathcal{R}$ value, there are two trends: as $\mathcal{R}$ increases, the sample is more likely to be a valuable learning case, + +such as a singing video, since it is often outside the teacher-training dataset $\mathcal{A}$ . However, if $\mathcal{R}$ is too large, the sample is more likely to be low-quality, such as one with poor audio-visual sync, because the teacher results are well synced. The statistical observations are discussed in Appendix A.2. + +Therefore, we should gradually increase the loss weight when $\mathcal{R}$ increases in a reasonable range, and also punish the loss weight if $\mathcal{R}$ increases out of the peak threshold $\mathcal{R}_p$ . For those cases whose $\mathcal{R}$ is bigger than the dead threshold $\mathcal{R}_d$ , the ground-truth supervised loss weight should be zero. The detailed formulation for loss weight $W$ can be obtained as: + +$$ +\mathcal {W} = \mathcal {W} _ {0} \cdot \left\{ \begin{array}{l l} \mathcal {R} ^ {s}, & \mathcal {R} \in [ 0, \mathcal {R} _ {p}) \\ \mathcal {R} _ {p} ^ {s} (\mathcal {R} _ {d} - \mathcal {R}) / (\mathcal {R} _ {d} - \mathcal {R} _ {p}), & \mathcal {R} \in [ \mathcal {R} _ {p}, \mathcal {R} _ {d}) \\ 0, & \mathcal {R} \in [ \mathcal {R} _ {d}, + \infty) \end{array} \right. \tag {8} +$$ + +where $s, \mathcal{W}_0, \mathcal{R}_p$ and $\mathcal{R}_d$ are hyper-parameters, they are set to 0.25, 0.2, 30, 100 respectively in this paper. Note that these hyper-parameters are relatively robust across different data distributions because they are used to select data within a relatively moderate range for training. + +Through the proposed design, FADA learns general generation capabilities from more ordinary data during distillation, enhancing model robustness. Our experiments validate + +the proposed loss's effectiveness. Additionally, our adaptive mixed-supervised distillation can extend to tasks like text-to-image or text-to-video without specific restrictions. + +# 3.3. Multi-CFG Distillation with Learnable Token + +In audio-driven talking avatar synthesis tasks, the MultiCFG technique is utilized [10, 14] to achieve robust control of multiple conditions (reference image and audio) effectively. Specifically, the single CFG calculates the direction from unconditional predicted values to a certain conditional predicted value and then moves a certain distance from the unconditional predicted values to obtain conditional predicted values. Furthermore, the MultiCFG recursively computes the next CFG based on the first CFG. In this paper, multiCFG inference result noise $\epsilon_{cfg}$ can be derived as follows[10]: + +$$ +\hat {\epsilon} _ {c f g} = \operatorname {c f g} _ {a} \times \left(\hat {\epsilon} _ {a} - \hat {\epsilon} _ {r}\right) + \operatorname {c f g} _ {r} \times \left(\hat {\epsilon} _ {r} - \hat {\epsilon} _ {b}\right) + \hat {\epsilon} _ {b}, \tag {9} +$$ + +where $\operatorname{cfg}_a$ and $\operatorname{cfg}_r$ indicate the CFG guidance scale of audio and reference condition. $\hat{\epsilon}_a$ refers to the noise prediction with both audio and reference conditions, while $\hat{\epsilon}_r$ removes the audio condition and $\hat{\epsilon}_b$ lacks both audio and reference. + +Therefore, during inference, the model needs to run three times for each denoising step, which is time-consuming. It naturally comes to mind to inject the CFG guidance scale as a condition into the network, and then let the student model learn the CFG reasoning characteristics of the teacher model during the distillation process, which echoes a similar concept to the $\omega$ -condition [20]. Specifically, the teacher model will conduct complete CFG reasoning during training to obtain predictions $\hat{\epsilon}_{\mathrm{cfg}}$ controlled by CFG. Subsequently, these predictions are further integrated into Formulas 2, 3 and 4 to derive input noisy latent $\hat{z}_t^{\mathrm{cfg}}$ and target noise $\hat{\epsilon}^{\mathrm{cfg}}(t)$ controlled by CFG, eventually undergoing a similar distillation procedure as Formula 5. + +However, directly adding the CFG value as a condition to the network does not yield satisfactory results. A careful design is needed for the injection method of the CFG scale condition. By observing the calculation of Multi-CFG reasoning in Formula 9, it is evident that this is akin to performing linear transformations on three independent vectors $\epsilon_{b}$ , $\epsilon_{a}$ , and $\epsilon_{r}$ in the noise space using the given CFG strengths $\mathrm{cfg}_{a}$ and $\mathrm{cfg}_{r}$ . Meanwhile, $\epsilon_{r}$ serves as a crucial link between the $\epsilon_{b}$ and $\epsilon_{a}$ . Simply embedding the CFG values overlooks the aforementioned relationships. Therefore, we introduce learnable tokens $\gamma_{b}, \gamma_{r}$ , and $\gamma_{r}$ into the model to learn the features required for CFG strength control and help the student model mimic the multi-CFG process. To achieve this, we design the CFG embedding $Emb_{\mathrm{cfg}}$ of the CFG guidance scale $\mathrm{cfg}_{a}$ and $\mathrm{cfg}_{r}$ as follows: + +$$ +E m b _ {\text {c f g}} = \operatorname {c f g} _ {a} \times \left(\gamma_ {a} - \gamma_ {r}\right) + \operatorname {c f g} _ {r} \times \left(\gamma_ {r} - \gamma_ {b}\right) + \gamma_ {b}. \tag {10} +$$ + +After obtaining the CFG embedding $Emb_{\mathrm{cfg}}$ , we sim + +imilarly introduce a CFG layer after each audio layer in the denoising network, injecting the information from $Emb_{\mathrm{cfg}}$ into the denoising process through cross attention. Importantly, the CFG layer does not introduce excessive performance overhead, especially when compared to the original cost of running three CFG runs. Therefore, our proposed multi-CFG distillation with learnable tokens can achieve nearly three times faster operation speed. + +# 4. Experiments + +In this section, we will introduce the experimental results. In Section 4.1, we will present the details of the implementation of FADA and the meta-information of our training and testing datasets. In Section 4.2, we will compare our proposed methods with other state-of-the-art diffusion-based talking avatar methods via quantitative and qualitative comparison. In Section 4.3, we will analyze our proposed techniques including the basic distillation method, adaptive mixed-supervised distillation and multi-CFG distillation, and prove their effectiveness. + +# 4.1. Experimental Settings + +Datasets Regarding the training dataset, we primarily used speaker videos obtained from the internet, sliced and cropped to durations ranging from two to ten seconds with facial frames, and resized to $512 \times 512$ resolution. We designed multi-dimensional filters for this dataset, including filters for data source, image quality, audio-visual synchronization, background stability and degree of movement, each with different thresholds to control the filtering intensity. The details of data collection and data filter strategies will be presented in Appendix D. As elucidated in Section 3.2, by using strict filtering thresholds, we obtained approximately 160 hours of high-quality training data $\mathcal{A}$ , which was utilized in teacher pre-training. Additionally, by using lenient filtering thresholds, we obtained around 1300 hours of moderate-quality training data $\mathcal{B}$ for distillation training. Following the approach of Echomimic [1], we selected 100 random samples from the open-source test sets HDTF [43] and CelebV-HQ [49] to validate the model's performance on various real human scenarios. Furthermore, to evaluate the model's performance in more open and extreme scenarios, we adopted the openset test set from Loopy [10] for qualitative comparisons, which contains various kinds of portraits and audios. + +Implementation Details The training process consists of teacher pre-training and student distillation. The training process of the teacher model closely aligns with the first two stages of recent methods[10, 31] with high-quality dataset $\mathcal{A}$ . In student distillation with the moderate-quality dataset $\mathcal{B}$ , the student model is initialized from the teacher model and integrated with CFG learnable tokens and CFG layers. During testing, the CFG guidance scale is set to 2.5 and 6.5 for reference and audio conditions in models without CFG + +distillation, while the scales were set to 2.0 and 6.5 for models with CFG distillation. The number of time windows in PeRFlow[37] is 4. All results of FADA are conducted with 6-step inferences, where the two time-windows near $T = 1$ contain two inference steps while the others contain one inference step. More details about the teacher and student training will be discussed in Appendix E. + +# 4.2. Comparison with State-of-the-Arts + +Metrics and Baselines For the purpose of quantitative comparisons, we evaluate the generated talking avatar videos with several metrics. We assess the image quality with IQA[35] metrics and evaluate the audio-visual synchronization distance with a pre-trained SyncNet[22] as Sync-D. FID and FVD[32] (with 3D ResNet) are assessed to evaluate the image/video-level fidelity difference between generated results and ground truths in the image inception space. E-FID (Expression-FID)[31] measures the inception distance between the generated and ground truths in expression parameter space. NFE-D indicates the number of function evaluations of the denoising net. For instance, our teacher model conducts 25-step DDIM inference with 3 CFG runs, leading to 75 NFE-D. Since our proposed FADA is a diffusion distillation framework, we mainly compare recent open-source diffusion-based state-of-the-art methods in talking portrait generation, including Hallo[36], EchoMimic[1], and V-Express[33]. Additionally, for reference, we also include the state-of-the-art GAN-based method SadTalker[42]. + +Quantitative Comparison As shown in Table 1, we conduct quantitative comparison experiments with CelebV-HQ and HDTF test sets. Since the differences in IQA are subtle, we provide $95\%$ confidence intervals to show their statistical significance. Ours-Balanced refers to our proposed method with mixed-supervised distillation with PeRFlow loss. Based on it, Ours-Fast utilizes multi-CFG distillation to achieve about 3-time speedup and gets a slight performance drop. Ours-Balanced achieves better IQA, FVD, FID, and E-FID on the challenging open-set CelebVHQ test set compared to the baselines. It also leads or is comparable in most metrics on the simpler HDTF test set. Ours-Fast maintains similar performance in IQA and Sync-D metrics with a slight drop in FVD, FID, and E-FID, which is still comparable to the baselines in most metrics. In terms of audio synchronization, the trend is similar. Notably, VExpress results show almost only lip movements, which, despite good metrics, do not produce satisfactory video quality. Finally, in terms of speedup, it is evident that both Ours-Balanced and Ours-Fast achieve significant speedup using the same SD base model as the baselines. The subjective evaluation will be conducted in Appendix F. + +Qualitative Comparison As illustrated in Figure 2 and in the supplementary video materials, Ours-Fast and Ours-Balanced maintain good video quality and vivid audiovisual expressiveness. We highly recommend readers watch + +the supplementary videos since it is difficult for the static image to show the audio-motion synchronization and temporal consistency. V-Express shows color differences and inferior identity-preserving abilities in all cases. Sadtalker generates good identity-preserving videos at the image level, but the head motions and lip movements are not well-synced with the emotion and pace of the given audio. EchoMimic and Hallo generate acceptable expressions, but there are temporal jitters that damage the visual quality. Moreover, by observing case-4 and case-6, we can find that FADA produces more full and plump pronunciations than baselines. + +# 4.3. Ablation Study + +Analysis of Distillation Methods. As illustrated in Table 2, we first compared various basic methods of diffusion distillation, including TCD [46], RFlow [16], and PeRFlow [37]. It was observed that compared to the TCD method, the PeRFlow method showed similar performance in IQA, FVD, Sync-D, and FID metrics, but with a significant advantage in E-FID. Additionally, the basic RFlow method could not handle the 6-step inference in our task setting. Therefore, we selected PeRFlow as the basic distillation method. We also tried simply adding moderate-quality data during teacher model training, which led to a decrease in model performance. These results indicate that existing distillation methods cannot directly produce a high-quality student model. Additionally, attempting to enhance the teacher model's capabilities by relaxing data filtering criteria and increasing data volume does not yield positive results. The comparison with the adversarial-based method is shown in Appendix G. + +Analysis of Adaptive Mixed-Supervised Distillation Here, based on the previously selected PeRFlow, we added dataset $\mathcal{B}$ and a mixed-supervised loss to investigate the effectiveness of our Adaptive Mixed-Supervised Distillation. As shown in Table 4, below the hdashline, we provide the performance of Mixed-Supervised Distillation with the addition of dataset $\mathcal{B}$ under different parameters. Comparing the effects of different Fixed-x values, it is evident that the ground-truth loss improves model performance, and a smaller loss weight value of 0.2 yields better results. We further investigated the effectiveness of the adaptive loss weight strategy. In the adaptive-unl. setting (unlimited, without using peak and dead thresholds), the adaptive loss weight increases infinitely as $\mathcal{R}$ increases, resulting in slightly inferior performance compared to the Fixed-0.2 setting. However, when employing the complete adaptive strategy, the ground-truth loss of low-quality samples is automatically discarded, leading to a distillation effect that surpasses the Fixed setting and approaches the performance of the teacher model, especially in terms of the FVD metrics. We also compared different values of $s$ and ultimately chose 0.25 due to its slight advantage. + +Additionally, by comparing the performance of teacher + +Table 1. Quantitative comparisons with state-of-the-art methods on CelebV-HQ & HDTF test sets. Bold and underlined numbers refer to the best and the second-best result in this test set. + +
MethodNFE-DCelebV-HQ TestHDTF Test
IQA↑Sync-D↓FVD-R↓FID↓E-FID↓IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Sadtalker[42]-2.953±.0868.765171.836.642.2483.435±.0697.87024.9325.351.559
Hallo[36]803.505±.0919.07953.9935.962.4263.922±.0797.91721.7120.151.337
EchoMimic[1]603.307±.10410.3754.7135.373.0183.994±.0839.39118.7119.011.328
V-Express[33]502.946±.1239.415117.865.092.4143.482±.0717.38248.0330.911.506
Ours-Balanced183.588±.0918.99850.3633.672.2023.951±.0827.92513.3218.391.362
Ours-Fast63.550±.0938.74654.6935.012.6043.927±.0817.99616.6718.511.635
+ +![](images/91ac309c50f1000e1a1f1a7c9eda42cb03a94e974d3cf58f5e5d09697b2fd84e.jpg) + +![](images/e47b696e285041304bef896edc2b845bf848231933631ee11b3c8d66a8503b89.jpg) + +![](images/0a5eae541ab7cb35e98e41a4b13cd2eafe748569191bf586c3da83aa56e371dc.jpg) + +![](images/c3a71f3f6bb9cd5e51e9315a0ba7d9a550ee6aa23b1e4421bf657f26c5b1322c.jpg) + +![](images/0c8bb52902793ce0e54ec1e518fe6bdbe972994bfc42d0edfa30354aebace100.jpg) + +![](images/6ebe079194feb410e47ff8b998a1dba7ba7b5894d57d16d58ec9e7c4399ff87d.jpg) +Source + +![](images/59a8c40cc6ff11b07f0b28fdc001ccf36c2fe83107e7dcdf5992fe532eb26826.jpg) + +![](images/bf39d59b677c3bdae860cefcb21f3882471aa53208ace999153bf869280f9830.jpg) + +![](images/b30623fd9b7cf9dd57fa0efffa31693ee4b65f2877427f0326174655ee909120.jpg) + +![](images/2760ce2115c605ef9b3ce2b01c69c56c6ce85707a20f351bf57a3f255f97db21.jpg) + +![](images/8aff1b9766452f80838de5fa68a38aee60d7b2e83eef941bdcbfb1650c6b71bf.jpg) + +![](images/59172be50245465e690d2222ac9df98d8c7460137f7bd7b3047653052629289e.jpg) +SadTalker + +![](images/8015fb8dbe326105aeef7266da0bc8d5032495dfb21fe450aa13ea4e4b60254f.jpg) + +![](images/f843bfe93a8854f9c9832d7b2074e8702fd7c2fa40ed67b3aa7be1e019513f97.jpg) + +![](images/e9468345cef7c2d54ac66f1843359f0a378928dec857ed3cf5ee881b6a088568.jpg) + +![](images/10c2f135edbbee68023574edbbfada5a9f68ef62327731391183cd1bcd0da3b8.jpg) + +![](images/05a631022453d1660e114aaaba2cc53d4bb5096327f33afc328fc3ba0f7166ff.jpg) + +![](images/ab648c7601b18a21a2b02621c86d26979812cb9633f52c63312e8f69069c3f0a.jpg) +VExpress + +![](images/8aaed49357062d3953f5c4816ca89b0eb6053b1604c5c01f724e218f38354fef.jpg) + +![](images/79b39434fbf19a06bc7772134181cbbeaf88825ccc5cda9d5a203de14e06464e.jpg) + +![](images/c5e6955aa23c03e5239673851426de0ef2f48c75c88458c2863b7e58019a0129.jpg) + +![](images/d2ba2acbdab5a12d821be172edc8a24f987205b67dad25b3b6715a21cbf1e02e.jpg) + +![](images/d65e08e5b3d5505f7380ca15dc11f8fda68f532449289a90865b226d235c3710.jpg) + +![](images/cad8d5239aa04828f15dcbbcb471133146b6d36a281e6ae3da917846c8ff8689.jpg) +Hallo + +![](images/170bea71c5883092017da89934c2b99ee8b7a04f49ccbf7797d32b4d27dacb5f.jpg) + +![](images/d071fb204fd7dc3bdefe55a6ab3ba66de87d9575d4e0464bdd1a02e995d736cc.jpg) + +![](images/ecddce9a1170a25585f47cf0e07c8680858a799b6fd0a877cbac2b2d5dd90d9c.jpg) + +![](images/4fff0c994435277323fba42650c8717f8b8efd995ff56ebee5da37a818be8046.jpg) + +![](images/4da9d5aee4760b00c4809d7f2720463143c9036e1ea50580f84b242ad9061242.jpg) + +![](images/efc6d9aa44cbff41048963277ad56a377c79705f67955bc0c55a330936b74c2b.jpg) +EchoMimic +Figure 2. Qualitative comparisons between FADA and baselines across different portraits and pronunciations in openset. + +![](images/ade7d7553527da62a23fc74917a0a83fb79bc35b6ae3dbccf97f087533e07a53.jpg) + +![](images/179103c7bb34be9b57af474cf32e2e3b34996a5250d640e3a86f2a58a1b5f427.jpg) + +![](images/eeede619c8276360fadb778ee703fb6bb041e366a93daf1dfd2e0452807f379e.jpg) + +![](images/461adbc1a22ec6f2ae2925680533c46176e854c91a9277fda14911d6b72b4e4b.jpg) + +![](images/f706423712a94d3b05334856ca25061f3ddaccaab26a328d2ad3009f60bd2f8a.jpg) + +![](images/649a842008a9a793d8093ad06fafa5cc5a627e43e0d67eaea9684c24e3143702.jpg) +Ours-Fast + +![](images/3b128be656c081bf8542c7de2fe9e4d0282de8ccd3f6e7f784cdf2dabdf713da.jpg) +[ə] + +![](images/cb5ca9432ba958afd9bc166698949d523baa2b6c695b39135eae47c5f319c303.jpg) +[I] + +![](images/9e528c0a630a73b22b50fc3ffa9bbf19e25a5ae6cb836dbf90ab43c89c19ad00.jpg) +[u:] + +![](images/1bf0e77c96ef04f4ef787db97483c415c20c498bfaac4e025072dcd12c1cf1e6.jpg) +[æ] + +![](images/3a1d2467dd1c9f02e0cbfef74f46a3195208027f05e316d84eaadb916798c216.jpg) +[eI] + +![](images/2f34cf9edabf69e80faffe5dd1da9ac4fd34612b4d7bf5bf9906af72e208d53f.jpg) +Ours-Balanced +[1] + +and student models under different data qualities and scales, it can be demonstrated that mixed supervised distillation actually enhances the model's utilization of moderate-quality data. As depicted in Tables 2 and 3, when the teacher model is trained with a large-scale moderate-quality dataset, a decline in model performance is observed, whereas the student model exhibits the opposite trend. Briefly, only mixed-supervised distillation can take advantage of the moderate- + +quality data. More analysis of mixed-supervised distillation will be discussed and more hyper-parameters experiments will be conducted in Appendix A.1. + +Analysis of Multi-CFG Distillation As demonstrated in Table 4, we further explore the effects of multi-CFG distillation on top of mixed-supervised distillation. Initially, we tested the results of conditional reasoning without CFG dis + +Table 2. Ablation study of basic distillation methods on HDTF test set. All methods in this table are free from mixed-supervised and CFG distillation. + +
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Teacher with B753.8858.12319.6421.401.689
TCD183.8277.93723.0123.151.893
RFlow182.8368.06681.1049.732.484
PeRFlow183.8238.00121.9121.261.477
+ +Table 3. Ablation study of adaptive mixed-supervised distillation on HDTF test set. All methods use PeRFlow with dataset $\mathcal{B}$ without CFG distillation. Fixed-x indicates fixed ground-truth loss with weight x. Adaptive-unlimited-s and Adaptive-s indicate adaptive ground-truth loss with hyper-parameter $s$ , with the former being without peak and dead threshold. + +
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Stu. with A183.8238.00121.9121.261.477
Stu. with B (Fixed-0)183.7627.90925.0621.471.462
+ Fixed-0.2183.8637.93413.7119.941.434
+ Fixed-1.0183.8127.90324.5321.421.494
+ Fixed-0.2, only A183.8807.96215.3220.711.512
+ Adaptive-unl.-0.5183.8247.91415.1520.741.471
+ Adaptive-unl.-0.25183.8467.90614.2019.921.424
+ Adaptive-0.5183.9497.91913.8318.641.395
+ Adaptive-0.25183.9517.92513.3218.391.362
+ +Table 4. Ablation study of multi-CFG distillation on HDTF test set. All methods in this table are PeRFlow + mixed-supervised distillation. CFG time embed refers to CFG distillation through timestep embedding injection. + +
MethodNFE-D IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
w.o. CFG63.6749.02027.1026.562.096
+ CFG time embed63.7018.57723.2422.471.791
+ CFG layer63.7467.90620.0520.941.620
+ CFG layer w. tokens63.9277.99616.6718.511.635
+ +![](images/61c03eae7ba7a2ff50a84cec046c655e0339ca057af1a5cc2d4866890c852c42.jpg) +Figure 3. Line charts showing the variations of FVD and Sync-D metrics with different audio CFG on HDTF test set. Reference CFG is set to 2.0 in this figure. + +tillation, which significantly deteriorated the model performance, especially leading to a doubling of the FVD metric representing video quality. For techniques like time embedding injection similar to $\omega$ -condition [20], they partially restored model performance but with limited efficacy. In con + +![](images/92f17b15c377942032f3430bc05d4cd9c1e12232f53f0a0bde41f39fb07bed04.jpg) +Figure 4. Line charts showing the variations of FVD and Sync-D metrics with different ref CFG on HDTF test set. Audio CFG is set to 6.5 in this figure. + +trast, our proposed learnable token-based CFG control layers significantly enhance model performance, nearly matching the teacher model and even surpassing it in FVD. Notably, our model requires only $8\%$ of the teacher model's NFE. We also tested without learnable tokens, and while the pure CFG control layer had some effect, it could not match the teacher model, especially in overall video quality metrics like FVD. + +To verify whether our proposed multi-CFG distillation effectively captures the influence of audio and reference conditions on the generated results, we plotted line charts of FVD and Sync-D metrics with variations in audio CFG and reference CFG, as shown in Figures 3 and 4. First, we observed improvements in both FVD and Sync-D as audio CFG increased, reaching a stable maximum value after 6.5. Hence, we ultimately chose 6.5 as the audio CFG for inference. Second, the influence of reference CFG on FVD and Sync-D metrics displayed an initially positive, then negative trend, with optimal performance achieved around the 2.0 position. These results indicate that our CFG control layer effectively mimics the multi-CFG process and achieves good results. More visualized examples and performance analysis under different scenes will be presented in Appendix C. + +# 5. Conclusion + +In summary, we proposed FADA, a fast diffusion avatar synthesis framework with mixed-supervised multi-CFG distillation. By designing a mixed-supervised loss, we leverage data of varying quality to enhance the robustness of generated results. Furthermore, a learnable token-based multi-CFG condition design is introduced to maintain the correlation between the audio and the generated video in the distilled models. The quantitative and qualitative experiments across multiple datasets show that our balanced setting achieves fast inference and high-fidelity generation ability. Ablation studies prove our proposed methods are effective in FADA. The limitations, future works and ethical concerns of this paper will be discussed in Appendix B, H and I respectively. + +# 6. Acknowledgements + +This project is sponsored by National Natural Science Foundation of China under Grant No.62222211 and National Natural Science Foundation of China under Grant No.U24A20326. + +# References + +[1] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 5, 6, 7 +[2] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2: Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 1, 2, 3 +[3] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 1, 2 +[4] Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2 +[5] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2 +[6] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[8] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 1, 2 +[9] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3 +[10] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3, 5 +[11] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[12] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 + +[13] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 2 +[14] Gaojie Lin, Jianwen Jiang, Chao Liang, Tianyun Zhong, Ji-aqi Yang, and Yanbo Zheng. Cyberhost: Taming audiodriven avatar diffusion model with region codebook attention. arXiv preprint arXiv:2409.01876, 2024. 1, 2, 3, 5 +[15] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2 +[16] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 6 +[17] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2 +[18] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 1 +[19] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 1, 2 +[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 5, 8 +[21] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 1, 2 +[22] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 6 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[24] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2 +[25] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 2 +[26] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 + +[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3 +[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1, 2 +[29] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2 +[30] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1, 2 +[31] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2, 3, 5, 6 +[32] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. OpenReview, 2019. 6 +[33] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 3, 6, 7 +[34] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2 +[35] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 6 +[36] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1, 2, 3, 6, 7 +[37] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 3, 6 +[38] Zhenhui Ye, Jinzheng He, Ziyue Jiang, Rongjie Huang, Jiawei Huang, Jinglin Liu, Yi Ren, Xiang Yin, Zejun Ma, and Zhou Zhao. Geneface++: Generalized and stable real-time audio-driven 3d talking face generation. arXiv preprint arXiv:2305.00787, 2023. 1, 2 +[39] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. +[40] Zhenhui Ye, Tianyun Zhong, Yi Ren, Ziyue Jiang, Jiawei Huang, Rongjie Huang, Jinglin Liu, Jinzheng He, Chen + +Zhang, Zehan Wang, et al. Mimictalk: Mimicking a personalized and expressive 3d talking face in minutes. arXiv preprint arXiv:2410.06734, 2024. 2 +[41] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 1, 2 +[42] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6, 7 +[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5 +[44] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 1, 2 +[45] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Wan Rui, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, et al. Vedit-q: Efficient and accurate quantization of diffusion transformers for image and video generation. arXiv preprint arXiv:2406.02540, 2024. 2 +[46] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation. arXiv preprint arXiv:2402.19159, 2024. 1, 2, 6 +[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 1, 2 +[48] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 1, 2 +[49] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5 \ No newline at end of file diff --git a/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/images.zip b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51f06c2a6bdd69e1c8bbcc7c091591f07100f334 --- /dev/null +++ b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcd307a9c1e2554709690dc681b4518309262b25e1b7e8c7c8b57dd5cfc0d1f2 +size 602149 diff --git a/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/layout.json b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c4c188df7ccb3d49b3eed352e36ee8b332e9882 --- /dev/null +++ b/2025/FADA_ Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation/layout.json @@ -0,0 +1,10310 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "text", + "content": "FADA: Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": "Tianyun Zhong" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": ", Chao Liang" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": ", Jianwen Jiang" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "inline_equation", + "content": "^{2*‡}" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": ", Gaojie Lin" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": ", Jiaqi Yang" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": ", Zhou Zhao" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": ", Zhejiang University, " + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 190 + ], + "type": "text", + "content": "Bytedance Intelligent Creation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 249, + 193, + 362, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 193, + 362, + 203 + ], + "spans": [ + { + "bbox": [ + 249, + 193, + 362, + 203 + ], + "type": "text", + "content": "zhongtianyun@zju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 205, + 432, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 205, + 432, + 217 + ], + "spans": [ + { + "bbox": [ + 179, + 205, + 432, + 217 + ], + "type": "text", + "content": "{liangchao.0412, jianwen.alan, yjq850207131}@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 220, + 381, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 220, + 381, + 231 + ], + "spans": [ + { + "bbox": [ + 227, + 220, + 381, + 231 + ], + "type": "text", + "content": "https://fadavatar.github.io/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 201, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 201, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 201, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 284, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 552 + ], + "type": "text", + "content": "Diffusion-based audio-driven talking avatar methods have recently gained attention for their high-fidelity, vivid, and expressive results. However, their slow inference speed limits practical applications. Despite the development of various distillation techniques for diffusion models, we found that naive diffusion distillation methods do not yield satisfactory results. Distilled models exhibit reduced robustness with open-set input images and a decreased correlation between audio and video compared to teacher models, undermining the advantages of diffusion models. To address this, we propose FADA (Fast Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation). We first designed a mixed-supervised loss to leverage data of varying quality and enhance the overall model capability as well as robustness. Additionally, we propose a multi-CFG distillation with learnable tokens to utilize the correlation between audio and reference image conditions, reducing the threefold inference runs caused by multi-CFG with acceptable quality degradation. Extensive experiments across multiple datasets show that FADA generates vivid videos comparable to recent diffusion model-based methods while achieving an NFE speedup of 4.17-12.5 times. Demos are available at our webpage https://fadavatar.github.io." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 574, + 137, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 137, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 137, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 676 + ], + "type": "text", + "content": "Talking avatar synthesis aims to animate a given portrait image using driven video or audio. Due to the ease of obtaining audio inputs and their low usage barrier, there has been a surge of recent work in this area. Based on the technical route, talking avatar synthesis methods can be categorized into GAN-based[4, 8, 19, 34, 44, 47, 48], NeRF-based[5, 30, 38-41], and the more recent diffusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 259, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 445 + ], + "type": "text", + "content": "model-based approaches[1, 2, 10, 14, 31, 33, 36]. Diffusion model-based methods, in particular, have gained attention due to their strong generative capabilities. These models exhibit excellent robustness in input images and produce vivid and highly expressive videos that naturally align with the audio. However, the current diffusion model-based talking head generation methods suffer from slow inference speeds, which hinder their practical application. Methods[1, 10, 36] based on Stable Diffusion (SD)[23], all require multiple denoising steps. Moreover, to ensure the correlation between the audio input and the image input in the final video, multi-CFG (Classifier-Free Guidance) inference is often employed, further increasing the runtime. Although most methods achieve 30-40 denoising steps with the help of DDIM[27], they still face speed challenges due to CFG calculations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 445, + 556, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 619 + ], + "type": "text", + "content": "Recently, several works have explored diffusion distillation in text-to-image [3, 12, 15, 17, 21, 24, 28, 46] and text-to-video tasks [18] to accelerate inference. Nonetheless, there has been no direct exploration of diffusion distillation for talking avatar synthesis tasks. Our observations indicate that simply applying text-to-image and text-to-video diffusion distillation methods results in a noticeable decline in performance. Unlike text to image or video tasks, talking avatar synthesis involves two key conditions: audio and a reference image. Since audio has a relatively weaker influence on video generation[10, 31], the final output is jointly influenced by both conditions, necessitating careful parameter tuning in multi-CFG settings. The sensitivity to condition guidance often leads to severe artifacts when applying diffusion distillation methods with few-step inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 620, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 557, + 715 + ], + "type": "text", + "content": "In this paper, we propose FADA (FAst Diffusion Avatar Synthesis with Mixed-Supervised Multi-CFG Distillation) to address these issues from two perspectives, resulting in an audio-driven talking avatar diffusion model that balances both speed and quality. First, we address the robustness of the distilled model. In this task, the mapping between audio and portrait movement must be learned from data. Previous works have emphasized meticulous data se" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 135, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 135, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 135, + 693 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 693, + 203, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 203, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 203, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 693, + 203, + 703 + ], + "type": "text", + "content": " Done during an internship at ByteDance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 113, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 113, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 113, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 113, + 712 + ], + "type": "text", + "content": "Project Lead" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3101" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 361 + ], + "type": "text", + "content": "lection [1, 10, 14], such as removing instances with excessive movement or asynchronous audio, to ensure high audio-portrait movement correlation. This stringent selection process excludes a significant amount of data that could enhance robustness. Considering that a well-trained teacher model already provides high audio-portrait correlation guidance, this discarded data can be utilized in training the student model. We designed a mixed-supervised learning strategy that adaptively adjusts the weights between learning from the teacher model's guidance and the student model's own learning from previously discarded data. Secondly, we address the speed issues of multimodal conditions in the distillation process. In audio-driven methods, the final output is achieved through composite CFG inference [10, 14, 36]. To make this process transparent to the student model during training, we introduce a set of learnable tokens. These tokens mimic the weighted calculations of multi-CFG with its coefficients and are injected into the network as control signals, enabling the student model to better learn the relationships between multiple conditions, including those present during inference. Since the learnable token conditions replicate the multi-CFG inference calculation process, we can reduce the number of multi-CFG model inferences during actual inference, further decreasing inference time on top of the denoising step distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 362, + 294, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 362, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 362, + 294, + 432 + ], + "type": "text", + "content": "Based on the above considerations and designed methods, we implemented FADA, a diffusion distillation method for audio-driven talking head generation, achieving up to a " + }, + { + "bbox": [ + 57, + 362, + 294, + 432 + ], + "type": "inline_equation", + "content": "12.5 \\times" + }, + { + "bbox": [ + 57, + 362, + 294, + 432 + ], + "type": "text", + "content": " speed-up while generating vivid portrait videos comparable to recent undistilled diffusion models. In summary, the contributions of this paper include:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 434, + 295, + 538 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 434, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 502 + ], + "type": "text", + "content": "1. We propose FADA, the first diffusion-based distillation framework for audio-driven talking avatar tasks. It includes a mixed-supervised loss to learn from data of varying quality and learnable token conditions to mimic the multi-CFG inference process, reducing performance loss by distillation and multi-CFG inferences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 503, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 538 + ], + "type": "text", + "content": "2. Quantitative and qualitative experiments demonstrate that FADA matches state-of-the-art generation quality while achieving significantly higher efficiency." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 57, + 549, + 146, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 549, + 146, + 561 + ], + "spans": [ + { + "bbox": [ + 57, + 549, + 146, + 561 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 568, + 214, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 214, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 214, + 581 + ], + "type": "text", + "content": "2.1. Audio-Driven Talking Avatar" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 586, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 586, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 586, + 294, + 712 + ], + "type": "text", + "content": "The task of audio-driven talking avatars has gained increasing attention recently, propelled by advancements in videogeneration technologies. GAN-based methods [4, 8, 19, 34, 44, 47, 48] typically employ a two-stage pipeline involving an audio-to-motion module and a motion-to-video module. Sadtalker [42] introduces a 3DMM-based motion representation along with a conditional VAE for generating high-fidelity talking heads. LivePortrait [4] expands the training dataset incorporating images and videos, using implicit keypoints for high-quality generation. NeRF-based methods [5, 30, 38-41] explore animating talking heads in 3D neural" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 72, + 553, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 142 + ], + "type": "text", + "content": "spaces. Real3D-Portrait [41] combines a pre-trained image-to-triplane module with a motion adapter for zero-shot and one-reference 3D talking head generation. MimicTalk [40] captures the talking style of individuals through in-context-learning audio-to-motion modules and rapid fine-tuning of triplanes for mimic talking head generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 146, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 146, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 146, + 553, + 308 + ], + "type": "text", + "content": "Diffusion-based talking avatar synthesis has swiftly become a crucial and efficient approach [1, 2, 10, 14, 31, 33, 36]. EMO [31] pioneers a double UNet framework in audio-driven avatar creation, yielding impressive and lifelike results. Concurrently, EchoMimic [1] trains on audio and facial landmarks for versatile control inputs. Hallo [36] incorporates a hierarchical audio-driven visual synthesis module to refine audio-visual alignments, while Hallo2 [2] explores image augmentation strategies for 4K avatar video synthesis. Loopy [10] introduces a motion-frame squeeze technique and audio-to-latent training for expressive results. Despite their progress in expressive audio-driven generation, diffusion-based methods still face challenges with slow inference speeds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 329, + 473, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 473, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 473, + 340 + ], + "type": "text", + "content": "2.2. Diffusion Model Acceleration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 350, + 553, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 476 + ], + "type": "text", + "content": "Diffusion acceleration involves lowering costs for denoising runs and reducing the number of runs. Various denoising methods aim to cut costs, such as utilizing Model quantization [6, 13, 25, 26, 29, 45] to decrease the model parameters directly. ViDiT-Q [47] introduced a new diffusion transformer quantization framework for video and image with mixed-precision quantization. DeepCache [20] takes cues from video compression to propagate reusable result regions through the model's internal structure, while Faster Diffusion [12] emphasizes re-using computed encoder features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 481, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 712 + ], + "type": "text", + "content": "Reducing the number of denoising runs is a direct way to accelerate diffusion models. SD-Turbo[24] leverages the ideas from GAN and introduces adversarial diffusion distillation to enhance the generation fidelity. SnapFusion[12] proposes progressive distillation to achieve fewer steps by learning the previous student model progressively. SwiftBrush[21] and SwiftBrushV2[3] utilizing variational score distillation to make it possible for texts-only distillation in text-to-image tasks. SDXL-lightning[15] mixes progressive distillation and adversarial distillation to achieve one-step image generation with a delicate training process. Applying consistency models [17, 28] to text-to-image tasks also helps reduce diffusion inference steps. LCM[17] views the guided reverse diffusion process as solving an augmented probability flow ODE. TCD[46] proposes trajectory consistency distillation to accurately trace the entire trajectory of the probability flow ODE in semilinear form with an exponential integrator. However, there is no specially-designed distillation framework and techniques for talking avatar synthesis tasks to our knowledge." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "text", + "content": "3102" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 184 + ], + "type": "text", + "content": "In this section, we introduce our proposed method, FADA, a diffusion distillation framework for avatar synthesis, as illustrated in Figure 1. Section 3.1 provides an overview of the diffusion distillation framework, based on recent PeR-Flow [37]. Section 3.2 delves into the specifics of the adaptive mixed-supervised distillation process. In Section 3.3, we present the multiple-CFG distillation technique with a learnable token implementation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 251, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 251, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 251, + 205 + ], + "type": "text", + "content": "3.1. Overall Framework and Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 405 + ], + "type": "text", + "content": "Building upon cutting-edge diffusion-based methods for talking avatar synthesis [1, 2, 9, 10, 14, 31, 33, 36], we introduce a dual-Unet architecture for the teacher model, leveraging pre-trained weights from Stable Diffusion[23] 1.5. As depicted in Figure 1, the reference image and motion frames (i.e., images from the last clip) undergo processing by a fixed VAE [11] and reference net. The output from each reference layer is then input to the corresponding layer in the denoising net for spatial attention. To facilitate audio-driven controls and ensure temporal consistency in talking avatar synthesis, audio attention and temporal attention layers are integrated into the denoising net. Initially following the Loopy design [10], we simplify the model by removing the TSM and Audio2Latents modules, streamlining subsequent model distillation. The student model mirrors the teacher's structure, with the addition of a CFG control layer, detailed in Section 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "text", + "content": "We choose the recent PeRFlow [37] as the optimization target for distillation. Initially, we train the teacher model on carefully curated high-quality datasets that prioritize motion amplitude and audio-visual synchronization. The teacher model is optimized using the " + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "text", + "content": "-prediction DDPM loss [7], where it predicts noise " + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{tea}}(\\cdot)" + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "text", + "content": " from noisy latent " + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 406, + 296, + 498 + ], + "type": "text", + "content": ". The training objective can be succinctly described as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 506, + 295, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 506, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 86, + 506, + 295, + 522 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {t e a}} = \\mathbb {E} _ {z _ {t}, c, t, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\mathrm {t e a}} (z _ {t}, t, c; \\theta_ {\\mathrm {t e a}}) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "3565e04e665d04d7d6159ff17b20be2d8926271de2622cf220a2de692ac0ed9d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "text", + "content": " includes the audio feature, reference image and motion frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": "Note that it is not necessary for the teacher model to be trained with flow-matching loss, and both " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": "-prediction and " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": "-prediction are valid for PeRFlow distillation. During the student distillation training, firstly same-interval " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": " time windows " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\{[t_k, t_{k-1})\\}_{k=K}^1" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": " are created where " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "t_k = k / K, k \\in \\{0, 1, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": ". A time window " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "[t_k, t_{k-1})" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": " is randomly sampled out, and the standpoint " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "z_{t_k}" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": " can be derived from the marginal distribution of the ground-truth targets, where " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "z_{t_k} = \\sqrt{1 - \\sigma^2(t_k)} z_0 + \\sigma(t_k)\\epsilon" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\sigma(t)" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": " denotes the noise schedule. Then the teacher-predicted ending point " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t_{k-1}}" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": " of this time window can be predicted and solved by ODE solver " + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 55, + 552, + 296, + 693 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 701, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 113, + 701, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t _ {k - 1}} = \\Phi \\left(z _ {t _ {k}}, t _ {k}, t _ {k - 1}; \\theta_ {\\text {t e a}}\\right), \\tag {2}", + "image_path": "3565b124406cec220d33bb4f4e6c3ea011d93d37d47da9213053e61f27fba4bf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "inline_equation", + "content": "\\Phi (\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "text", + "content": " denotes the DDIM solver[27] which performs several inference steps iteratively to reach the endpoint of this time window. Note that the parameters of the teacher model are frozen and no gradient computations are needed here. After that, the input noisy latent of student " + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "text", + "content": " will be derived from a linear interpolation between starting point " + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "text", + "content": " and ending point " + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "inline_equation", + "content": "t_{k - 1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 155 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 380, + 159, + 554, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 159, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 380, + 159, + 554, + 185 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t} = \\frac {z _ {t _ {k}} - \\hat {z} _ {t _ {k - 1}}}{t _ {k} - t _ {k - 1}} (t - t _ {k}). \\tag {3}", + "image_path": "bf2dd9d756da491df8b32bc2ff59585122dcda247905b8d2337d14ac25969b50.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 190, + 554, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 554, + 260 + ], + "type": "text", + "content": "Through the linear interpolation above, the ODE flow will be rectified to piecewise straight lines so only a few steps are enough to denoising in a time window, which reduces the number of timesteps at inference stage. Meanwhile, we can figure out the target noise " + }, + { + "bbox": [ + 313, + 190, + 554, + 260 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} (t)" + }, + { + "bbox": [ + 313, + 190, + 554, + 260 + ], + "type": "text", + "content": " via parameterization[37]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 387, + 266, + 554, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 266, + 554, + 291 + ], + "spans": [ + { + "bbox": [ + 387, + 266, + 554, + 291 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} (t) = \\frac {\\hat {z} _ {t _ {k - 1}} - \\lambda_ {k} z _ {t _ {k}}}{\\eta_ {k}}, \\tag {4}", + "image_path": "d93a8cac0a3dc0a787516391509ff17ded961ba17fab65a0f82186b962ef65f3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\lambda_{k} = \\sqrt{\\alpha_{t_{k - 1}}} / \\sqrt{\\alpha_{t_{k}}}" + }, + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\eta_{k} = \\sqrt{1 - \\alpha_{t_{k - 1}}} -" + }, + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\sqrt{1 - \\alpha_{t_k}}\\sqrt{\\alpha_{t_{k - 1}}}\\big{/}\\sqrt{\\alpha_{t_k}}" + }, + { + "bbox": [ + 313, + 300, + 554, + 334 + ], + "type": "text", + "content": " respectively. Finally, the training objective of the distillation process can be obtained as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 342, + 555, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 555, + 365 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i s t i l l}} = \\mathbb {E} _ {k, t \\in [ t _ {k}, t _ {k - 1}), z _ {t}, c, \\epsilon} \\left[ \\| \\hat {\\epsilon} (t) - \\epsilon_ {\\mathrm {s t u}} (\\hat {z} _ {t}, t, c; \\theta_ {\\mathrm {s t u}}) \\| ^ {2} \\right], \\tag {5}", + "image_path": "202f1027b5a57312548225e50ba8a466e84e358a06c9e1210d73c699f42c32cd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{stu}}(\\cdot)" + }, + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "text", + "content": " refers to the predicted noise from the student model and " + }, + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{distill}}" + }, + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "text", + "content": " represents the distillation loss. Note that " + }, + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{distill}}" + }, + { + "bbox": [ + 313, + 366, + 554, + 423 + ], + "type": "text", + "content": " is only a trivial part of our framework, the mixed-supervised training and multi-CFG distillation will be proposed in Section 3.2 and 3.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 522, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 522, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 522, + 442 + ], + "type": "text", + "content": "3.2. Adaptive Mixed-Supervised Distillation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 447, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 540 + ], + "type": "text", + "content": "In our experiments, we found that distillation significantly degrades the model's performance, especially overall video synthesis quality, as shown in Table 3. Increasing data can improve performance, but in audio-driven portrait generation, the weak control of audio over motion requires the model to learn high-quality motion patterns to fill in details not covered by audio. This necessitates stricter data filtering, reducing the available data quantity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": "Fortunately, this problem can be mitigated in our proposed distillation process. In our design, we first maintain the teacher training unchanged with only a high-quality dataset " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": " to ensure its output stability. Then during the distillation stage, the data filtering strategies will be relaxed and some moderate-quality data samples (a larger dataset " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": " that includes " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}| > |\\mathcal{A}|" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": ".) will be fed to student training. Since the student model mimics the teacher model's output, it can maintain the high-quality generation capability of the teacher model even with the inclusion of lower-quality data. To further leverage the new data during the distillation stage, we introduce a ground-truth supervised loss " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": ". The ground-truth noisy latent " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": " and ground-truth target noise " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}(t)" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": " can be simply obtained by replacing " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t_{k-1}}" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": " with ground truth ending point " + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "inline_equation", + "content": "z_{t_{k-1}}" + }, + { + "bbox": [ + 313, + 540, + 555, + 714 + ], + "type": "text", + "content": " in For" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3103" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 553, + 364 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 553, + 364 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 553, + 364 + ], + "type": "image", + "image_path": "cc474c8bd90a591ad224fc6e3f96fe418ae127a148d1f9a8109d9a57cb648be5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 373, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 418 + ], + "type": "text", + "content": "Figure 1. Overall distillation framework of FADA. The teacher model is trained only with high-quality data, which is omitted in the figure. The student model is trained by a mixed loss of ground-truth and teacher-supervised loss to leverage data of varying quality. Learnable token-based CFG conditions enable the student model to mimic the multi-CFG process, further reducing inference times. For simplicity, we have omitted some components commonly used in previous methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 433, + 297, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 297, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 297, + 467 + ], + "type": "text", + "content": "mula 3 and 4. Similar to distillation loss computation, the ground-truth loss " + }, + { + "bbox": [ + 55, + 433, + 297, + 467 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{GT}}" + }, + { + "bbox": [ + 55, + 433, + 297, + 467 + ], + "type": "text", + "content": " and total mixed loss " + }, + { + "bbox": [ + 55, + 433, + 297, + 467 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 433, + 297, + 467 + ], + "type": "text", + "content": " can be described as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 475, + 296, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 501 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {G T}} = \\mathbb {E} _ {k, t \\in \\left[ t _ {k}, t _ {k - 1}\\right), z _ {t}, c, \\epsilon} \\left[ \\| \\epsilon^ {*} (t) - \\epsilon_ {\\mathrm {s t u}} \\left(z _ {t} ^ {*}, t, c; \\theta_ {\\mathrm {s t u}}\\right) \\| _ {2} ^ {2} \\right], \\tag {6}", + "image_path": "a75b2027322b5274cdf5bc7018efa87062f743ffb32c3e1edfc90f95e886cdf3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 509, + 295, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 509, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 116, + 509, + 295, + 521 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o t a l}} = L _ {\\text {d i s t i l l}} + \\mathcal {W} \\cdot L _ {\\mathrm {G T}}, \\tag {7}", + "image_path": "76c88afc3eb73cb8c1a570d3916be8818b3c4517db52728fc1702335efd9a87e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "text", + "content": " represents the loss weight of " + }, + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "inline_equation", + "content": "L_{GT}" + }, + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "text", + "content": ". Through this approach, the student model can better learn high-quality motion pattern generation from the teacher model while also learning more common patterns from a broader range of data to enhance its generalization capability. A trivial implementation idea is to set a fixed hyper-parameter to " + }, + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 525, + 296, + 618 + ], + "type": "text", + "content": " but this is not very effective because it assumes every sample has the same learning value, which is inaccurate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": "Hence, an adaptive method to balance teacher-supervised loss " + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{teacher}}" + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": " and ground-truth-supervised loss " + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gt}}" + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": " is necessary, and the potential information behind the two losses should be taken into account. Considering the ratio " + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = L_{\\mathrm{gt}} / L_{\\mathrm{teacher}}" + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": ", it indicates the difference level between the " + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gt}}" + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{teacher}}" + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": ". Based on qualitative observations of the " + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": " value, there are two trends: as " + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 620, + 296, + 714 + ], + "type": "text", + "content": " increases, the sample is more likely to be a valuable learning case," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 433, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 491 + ], + "type": "text", + "content": "such as a singing video, since it is often outside the teacher-training dataset " + }, + { + "bbox": [ + 313, + 433, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 433, + 555, + 491 + ], + "type": "text", + "content": ". However, if " + }, + { + "bbox": [ + 313, + 433, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 433, + 555, + 491 + ], + "type": "text", + "content": " is too large, the sample is more likely to be low-quality, such as one with poor audio-visual sync, because the teacher results are well synced. The statistical observations are discussed in Appendix A.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "content": "Therefore, we should gradually increase the loss weight when " + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "content": " increases in a reasonable range, and also punish the loss weight if " + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "content": " increases out of the peak threshold " + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_p" + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "content": ". For those cases whose " + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "content": " is bigger than the dead threshold " + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_d" + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "content": ", the ground-truth supervised loss weight should be zero. The detailed formulation for loss weight " + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 492, + 555, + 562 + ], + "type": "text", + "content": " can be obtained as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 553, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 618 + ], + "type": "interline_equation", + "content": "\\mathcal {W} = \\mathcal {W} _ {0} \\cdot \\left\\{ \\begin{array}{l l} \\mathcal {R} ^ {s}, & \\mathcal {R} \\in [ 0, \\mathcal {R} _ {p}) \\\\ \\mathcal {R} _ {p} ^ {s} (\\mathcal {R} _ {d} - \\mathcal {R}) / (\\mathcal {R} _ {d} - \\mathcal {R} _ {p}), & \\mathcal {R} \\in [ \\mathcal {R} _ {p}, \\mathcal {R} _ {d}) \\\\ 0, & \\mathcal {R} \\in [ \\mathcal {R} _ {d}, + \\infty) \\end{array} \\right. \\tag {8}", + "image_path": "e86c34c99bea61cbd060bc57e7cda6ba07ab831d74d1f7930dbfd82ab1c1a0ae.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 619, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 619, + 555, + 677 + ], + "type": "inline_equation", + "content": "s, \\mathcal{W}_0, \\mathcal{R}_p" + }, + { + "bbox": [ + 313, + 619, + 555, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 619, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_d" + }, + { + "bbox": [ + 313, + 619, + 555, + 677 + ], + "type": "text", + "content": " are hyper-parameters, they are set to 0.25, 0.2, 30, 100 respectively in this paper. Note that these hyper-parameters are relatively robust across different data distributions because they are used to select data within a relatively moderate range for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 678, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 556, + 713 + ], + "type": "text", + "content": "Through the proposed design, FADA learns general generation capabilities from more ordinary data during distillation, enhancing model robustness. Our experiments validate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3104" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "the proposed loss's effectiveness. Additionally, our adaptive mixed-supervised distillation can extend to tasks like text-to-image or text-to-video without specific restrictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 291, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 291, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 291, + 131 + ], + "type": "text", + "content": "3.3. Multi-CFG Distillation with Learnable Token" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 295, + 263 + ], + "type": "text", + "content": "In audio-driven talking avatar synthesis tasks, the MultiCFG technique is utilized [10, 14] to achieve robust control of multiple conditions (reference image and audio) effectively. Specifically, the single CFG calculates the direction from unconditional predicted values to a certain conditional predicted value and then moves a certain distance from the unconditional predicted values to obtain conditional predicted values. Furthermore, the MultiCFG recursively computes the next CFG based on the first CFG. In this paper, multiCFG inference result noise " + }, + { + "bbox": [ + 55, + 136, + 295, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{cfg}" + }, + { + "bbox": [ + 55, + 136, + 295, + 263 + ], + "type": "text", + "content": " can be derived as follows[10]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 273, + 295, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 273, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 69, + 273, + 295, + 287 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {c f g} = \\operatorname {c f g} _ {a} \\times \\left(\\hat {\\epsilon} _ {a} - \\hat {\\epsilon} _ {r}\\right) + \\operatorname {c f g} _ {r} \\times \\left(\\hat {\\epsilon} _ {r} - \\hat {\\epsilon} _ {b}\\right) + \\hat {\\epsilon} _ {b}, \\tag {9}", + "image_path": "8244c18ab4ab227bd412f900e09d0e4241c9d46199bc177d20f2d28a7db19541.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\operatorname{cfg}_a" + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\operatorname{cfg}_r" + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "text", + "content": " indicate the CFG guidance scale of audio and reference condition. " + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_a" + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "text", + "content": " refers to the noise prediction with both audio and reference conditions, while " + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_r" + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "text", + "content": " removes the audio condition and " + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_b" + }, + { + "bbox": [ + 55, + 294, + 295, + 341 + ], + "type": "text", + "content": " lacks both audio and reference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "text", + "content": "Therefore, during inference, the model needs to run three times for each denoising step, which is time-consuming. It naturally comes to mind to inject the CFG guidance scale as a condition into the network, and then let the student model learn the CFG reasoning characteristics of the teacher model during the distillation process, which echoes a similar concept to the " + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "text", + "content": "-condition [20]. Specifically, the teacher model will conduct complete CFG reasoning during training to obtain predictions " + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\mathrm{cfg}}" + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "text", + "content": " controlled by CFG. Subsequently, these predictions are further integrated into Formulas 2, 3 and 4 to derive input noisy latent " + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^{\\mathrm{cfg}}" + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "text", + "content": " and target noise " + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}^{\\mathrm{cfg}}(t)" + }, + { + "bbox": [ + 55, + 342, + 295, + 493 + ], + "type": "text", + "content": " controlled by CFG, eventually undergoing a similar distillation procedure as Formula 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": "However, directly adding the CFG value as a condition to the network does not yield satisfactory results. A careful design is needed for the injection method of the CFG scale condition. By observing the calculation of Multi-CFG reasoning in Formula 9, it is evident that this is akin to performing linear transformations on three independent vectors " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\epsilon_{b}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\epsilon_{a}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\epsilon_{r}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " in the noise space using the given CFG strengths " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{cfg}_{a}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{cfg}_{r}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\epsilon_{r}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " serves as a crucial link between the " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\epsilon_{b}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\epsilon_{a}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": ". Simply embedding the CFG values overlooks the aforementioned relationships. Therefore, we introduce learnable tokens " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\gamma_{b}, \\gamma_{r}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\gamma_{r}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " into the model to learn the features required for CFG strength control and help the student model mimic the multi-CFG process. To achieve this, we design the CFG embedding " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "Emb_{\\mathrm{cfg}}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " of the CFG guidance scale " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{cfg}_{a}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{cfg}_{r}" + }, + { + "bbox": [ + 55, + 495, + 295, + 669 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 679, + 295, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 679, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 61, + 679, + 295, + 693 + ], + "type": "interline_equation", + "content": "E m b _ {\\text {c f g}} = \\operatorname {c f g} _ {a} \\times \\left(\\gamma_ {a} - \\gamma_ {r}\\right) + \\operatorname {c f g} _ {r} \\times \\left(\\gamma_ {r} - \\gamma_ {b}\\right) + \\gamma_ {b}. \\tag {10}", + "image_path": "52c067715fd648e9faf9c96d636f3c35f0ebe0608774b460dad2a9775d314123.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "After obtaining the CFG embedding " + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "Emb_{\\mathrm{cfg}}" + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": ", we sim" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 166 + ], + "type": "text", + "content": "imilarly introduce a CFG layer after each audio layer in the denoising network, injecting the information from " + }, + { + "bbox": [ + 313, + 72, + 553, + 166 + ], + "type": "inline_equation", + "content": "Emb_{\\mathrm{cfg}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 166 + ], + "type": "text", + "content": " into the denoising process through cross attention. Importantly, the CFG layer does not introduce excessive performance overhead, especially when compared to the original cost of running three CFG runs. Therefore, our proposed multi-CFG distillation with learnable tokens can achieve nearly three times faster operation speed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 198, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 314 + ], + "type": "text", + "content": "In this section, we will introduce the experimental results. In Section 4.1, we will present the details of the implementation of FADA and the meta-information of our training and testing datasets. In Section 4.2, we will compare our proposed methods with other state-of-the-art diffusion-based talking avatar methods via quantitative and qualitative comparison. In Section 4.3, we will analyze our proposed techniques including the basic distillation method, adaptive mixed-supervised distillation and multi-CFG distillation, and prove their effectiveness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 324, + 441, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 441, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 441, + 337 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "text", + "content": "Datasets Regarding the training dataset, we primarily used speaker videos obtained from the internet, sliced and cropped to durations ranging from two to ten seconds with facial frames, and resized to " + }, + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "text", + "content": " resolution. We designed multi-dimensional filters for this dataset, including filters for data source, image quality, audio-visual synchronization, background stability and degree of movement, each with different thresholds to control the filtering intensity. The details of data collection and data filter strategies will be presented in Appendix D. As elucidated in Section 3.2, by using strict filtering thresholds, we obtained approximately 160 hours of high-quality training data " + }, + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "text", + "content": ", which was utilized in teacher pre-training. Additionally, by using lenient filtering thresholds, we obtained around 1300 hours of moderate-quality training data " + }, + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 341, + 555, + 608 + ], + "type": "text", + "content": " for distillation training. Following the approach of Echomimic [1], we selected 100 random samples from the open-source test sets HDTF [43] and CelebV-HQ [49] to validate the model's performance on various real human scenarios. Furthermore, to evaluate the model's performance in more open and extreme scenarios, we adopted the openset test set from Loopy [10] for qualitative comparisons, which contains various kinds of portraits and audios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 609, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 556, + 713 + ], + "type": "text", + "content": "Implementation Details The training process consists of teacher pre-training and student distillation. The training process of the teacher model closely aligns with the first two stages of recent methods[10, 31] with high-quality dataset " + }, + { + "bbox": [ + 313, + 609, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 609, + 556, + 713 + ], + "type": "text", + "content": ". In student distillation with the moderate-quality dataset " + }, + { + "bbox": [ + 313, + 609, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 609, + 556, + 713 + ], + "type": "text", + "content": ", the student model is initialized from the teacher model and integrated with CFG learnable tokens and CFG layers. During testing, the CFG guidance scale is set to 2.5 and 6.5 for reference and audio conditions in models without CFG" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3105" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": "distillation, while the scales were set to 2.0 and 6.5 for models with CFG distillation. The number of time windows in PeRFlow[37] is 4. All results of FADA are conducted with 6-step inferences, where the two time-windows near " + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": " contain two inference steps while the others contain one inference step. More details about the teacher and student training will be discussed in Appendix E." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 242, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 242, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 242, + 171 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 407 + ], + "type": "text", + "content": "Metrics and Baselines For the purpose of quantitative comparisons, we evaluate the generated talking avatar videos with several metrics. We assess the image quality with IQA[35] metrics and evaluate the audio-visual synchronization distance with a pre-trained SyncNet[22] as Sync-D. FID and FVD[32] (with 3D ResNet) are assessed to evaluate the image/video-level fidelity difference between generated results and ground truths in the image inception space. E-FID (Expression-FID)[31] measures the inception distance between the generated and ground truths in expression parameter space. NFE-D indicates the number of function evaluations of the denoising net. For instance, our teacher model conducts 25-step DDIM inference with 3 CFG runs, leading to 75 NFE-D. Since our proposed FADA is a diffusion distillation framework, we mainly compare recent open-source diffusion-based state-of-the-art methods in talking portrait generation, including Hallo[36], EchoMimic[1], and V-Express[33]. Additionally, for reference, we also include the state-of-the-art GAN-based method SadTalker[42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 410, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 296, + 665 + ], + "type": "text", + "content": "Quantitative Comparison As shown in Table 1, we conduct quantitative comparison experiments with CelebV-HQ and HDTF test sets. Since the differences in IQA are subtle, we provide " + }, + { + "bbox": [ + 56, + 410, + 296, + 665 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 56, + 410, + 296, + 665 + ], + "type": "text", + "content": " confidence intervals to show their statistical significance. Ours-Balanced refers to our proposed method with mixed-supervised distillation with PeRFlow loss. Based on it, Ours-Fast utilizes multi-CFG distillation to achieve about 3-time speedup and gets a slight performance drop. Ours-Balanced achieves better IQA, FVD, FID, and E-FID on the challenging open-set CelebVHQ test set compared to the baselines. It also leads or is comparable in most metrics on the simpler HDTF test set. Ours-Fast maintains similar performance in IQA and Sync-D metrics with a slight drop in FVD, FID, and E-FID, which is still comparable to the baselines in most metrics. In terms of audio synchronization, the trend is similar. Notably, VExpress results show almost only lip movements, which, despite good metrics, do not produce satisfactory video quality. Finally, in terms of speedup, it is evident that both Ours-Balanced and Ours-Fast achieve significant speedup using the same SD base model as the baselines. The subjective evaluation will be conducted in Appendix F." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": "Qualitative Comparison As illustrated in Figure 2 and in the supplementary video materials, Ours-Fast and Ours-Balanced maintain good video quality and vivid audiovisual expressiveness. We highly recommend readers watch" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": "the supplementary videos since it is difficult for the static image to show the audio-motion synchronization and temporal consistency. V-Express shows color differences and inferior identity-preserving abilities in all cases. Sadtalker generates good identity-preserving videos at the image level, but the head motions and lip movements are not well-synced with the emotion and pace of the given audio. EchoMimic and Hallo generate acceptable expressions, but there are temporal jitters that damage the visual quality. Moreover, by observing case-4 and case-6, we can find that FADA produces more full and plump pronunciations than baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 218, + 408, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 408, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 408, + 231 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 556, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 556, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 556, + 443 + ], + "type": "text", + "content": "Analysis of Distillation Methods. As illustrated in Table 2, we first compared various basic methods of diffusion distillation, including TCD [46], RFlow [16], and PeRFlow [37]. It was observed that compared to the TCD method, the PeRFlow method showed similar performance in IQA, FVD, Sync-D, and FID metrics, but with a significant advantage in E-FID. Additionally, the basic RFlow method could not handle the 6-step inference in our task setting. Therefore, we selected PeRFlow as the basic distillation method. We also tried simply adding moderate-quality data during teacher model training, which led to a decrease in model performance. These results indicate that existing distillation methods cannot directly produce a high-quality student model. Additionally, attempting to enhance the teacher model's capabilities by relaxing data filtering criteria and increasing data volume does not yield positive results. The comparison with the adversarial-based method is shown in Appendix G." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "text", + "content": "Analysis of Adaptive Mixed-Supervised Distillation Here, based on the previously selected PeRFlow, we added dataset " + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "text", + "content": " and a mixed-supervised loss to investigate the effectiveness of our Adaptive Mixed-Supervised Distillation. As shown in Table 4, below the hdashline, we provide the performance of Mixed-Supervised Distillation with the addition of dataset " + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "text", + "content": " under different parameters. Comparing the effects of different Fixed-x values, it is evident that the ground-truth loss improves model performance, and a smaller loss weight value of 0.2 yields better results. We further investigated the effectiveness of the adaptive loss weight strategy. In the adaptive-unl. setting (unlimited, without using peak and dead thresholds), the adaptive loss weight increases infinitely as " + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "text", + "content": " increases, resulting in slightly inferior performance compared to the Fixed-0.2 setting. However, when employing the complete adaptive strategy, the ground-truth loss of low-quality samples is automatically discarded, leading to a distillation effect that surpasses the Fixed setting and approaches the performance of the teacher model, especially in terms of the FVD metrics. We also compared different values of " + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 447, + 556, + 701 + ], + "type": "text", + "content": " and ultimately chose 0.25 due to its slight advantage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 702, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 702, + 553, + 714 + ], + "type": "text", + "content": "Additionally, by comparing the performance of teacher" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3106" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 561, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with state-of-the-art methods on CelebV-HQ & HDTF test sets. Bold and underlined numbers refer to the best and the second-best result in this test set." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 561, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 561, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 561, + 216 + ], + "type": "table", + "html": "
MethodNFE-DCelebV-HQ TestHDTF Test
IQA↑Sync-D↓FVD-R↓FID↓E-FID↓IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Sadtalker[42]-2.953±.0868.765171.836.642.2483.435±.0697.87024.9325.351.559
Hallo[36]803.505±.0919.07953.9935.962.4263.922±.0797.91721.7120.151.337
EchoMimic[1]603.307±.10410.3754.7135.373.0183.994±.0839.39118.7119.011.328
V-Express[33]502.946±.1239.415117.865.092.4143.482±.0717.38248.0330.911.506
Ours-Balanced183.588±.0918.99850.3633.672.2023.951±.0827.92513.3218.391.362
Ours-Fast63.550±.0938.74654.6935.012.6043.927±.0817.99616.6718.511.635
", + "image_path": "3474dd35f04ccdfadd78c3afc920ffd9cdeb78bdcae854511697f7b7bcb9eb0a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 78, + 232, + 135, + 290 + ], + "blocks": [ + { + "bbox": [ + 78, + 232, + 135, + 290 + ], + "lines": [ + { + "bbox": [ + 78, + 232, + 135, + 290 + ], + "spans": [ + { + "bbox": [ + 78, + 232, + 135, + 290 + ], + "type": "image", + "image_path": "91ac309c50f1000e1a1f1a7c9eda42cb03a94e974d3cf58f5e5d09697b2fd84e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 78, + 291, + 135, + 347 + ], + "blocks": [ + { + "bbox": [ + 78, + 291, + 135, + 347 + ], + "lines": [ + { + "bbox": [ + 78, + 291, + 135, + 347 + ], + "spans": [ + { + "bbox": [ + 78, + 291, + 135, + 347 + ], + "type": "image", + "image_path": "e47b696e285041304bef896edc2b845bf848231933631ee11b3c8d66a8503b89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 78, + 347, + 135, + 404 + ], + "blocks": [ + { + "bbox": [ + 78, + 347, + 135, + 404 + ], + "lines": [ + { + "bbox": [ + 78, + 347, + 135, + 404 + ], + "spans": [ + { + "bbox": [ + 78, + 347, + 135, + 404 + ], + "type": "image", + "image_path": "0a5eae541ab7cb35e98e41a4b13cd2eafe748569191bf586c3da83aa56e371dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 78, + 405, + 135, + 461 + ], + "blocks": [ + { + "bbox": [ + 78, + 405, + 135, + 461 + ], + "lines": [ + { + "bbox": [ + 78, + 405, + 135, + 461 + ], + "spans": [ + { + "bbox": [ + 78, + 405, + 135, + 461 + ], + "type": "image", + "image_path": "c3a71f3f6bb9cd5e51e9315a0ba7d9a550ee6aa23b1e4421bf657f26c5b1322c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 78, + 463, + 135, + 519 + ], + "blocks": [ + { + "bbox": [ + 78, + 463, + 135, + 519 + ], + "lines": [ + { + "bbox": [ + 78, + 463, + 135, + 519 + ], + "spans": [ + { + "bbox": [ + 78, + 463, + 135, + 519 + ], + "type": "image", + "image_path": "0c8bb52902793ce0e54ec1e518fe6bdbe972994bfc42d0edfa30354aebace100.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 78, + 521, + 135, + 578 + ], + "blocks": [ + { + "bbox": [ + 78, + 521, + 135, + 578 + ], + "lines": [ + { + "bbox": [ + 78, + 521, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 78, + 521, + 135, + 578 + ], + "type": "image", + "image_path": "6ebe079194feb410e47ff8b998a1dba7ba7b5894d57d16d58ec9e7c4399ff87d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 582, + 124, + 593 + ], + "lines": [ + { + "bbox": [ + 89, + 582, + 124, + 593 + ], + "spans": [ + { + "bbox": [ + 89, + 582, + 124, + 593 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 141, + 233, + 198, + 290 + ], + "blocks": [ + { + "bbox": [ + 141, + 233, + 198, + 290 + ], + "lines": [ + { + "bbox": [ + 141, + 233, + 198, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 198, + 290 + ], + "type": "image", + "image_path": "59a8c40cc6ff11b07f0b28fdc001ccf36c2fe83107e7dcdf5992fe532eb26826.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 141, + 291, + 198, + 347 + ], + "blocks": [ + { + "bbox": [ + 141, + 291, + 198, + 347 + ], + "lines": [ + { + "bbox": [ + 141, + 291, + 198, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 291, + 198, + 347 + ], + "type": "image", + "image_path": "bf39d59b677c3bdae860cefcb21f3882471aa53208ace999153bf869280f9830.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 141, + 348, + 198, + 404 + ], + "blocks": [ + { + "bbox": [ + 141, + 348, + 198, + 404 + ], + "lines": [ + { + "bbox": [ + 141, + 348, + 198, + 404 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 198, + 404 + ], + "type": "image", + "image_path": "b30623fd9b7cf9dd57fa0efffa31693ee4b65f2877427f0326174655ee909120.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 141, + 405, + 198, + 461 + ], + "blocks": [ + { + "bbox": [ + 141, + 405, + 198, + 461 + ], + "lines": [ + { + "bbox": [ + 141, + 405, + 198, + 461 + ], + "spans": [ + { + "bbox": [ + 141, + 405, + 198, + 461 + ], + "type": "image", + "image_path": "2760ce2115c605ef9b3ce2b01c69c56c6ce85707a20f351bf57a3f255f97db21.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 462, + 198, + 519 + ], + "blocks": [ + { + "bbox": [ + 141, + 462, + 198, + 519 + ], + "lines": [ + { + "bbox": [ + 141, + 462, + 198, + 519 + ], + "spans": [ + { + "bbox": [ + 141, + 462, + 198, + 519 + ], + "type": "image", + "image_path": "8aff1b9766452f80838de5fa68a38aee60d7b2e83eef941bdcbfb1650c6b71bf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 141, + 521, + 198, + 578 + ], + "blocks": [ + { + "bbox": [ + 141, + 521, + 198, + 578 + ], + "lines": [ + { + "bbox": [ + 141, + 521, + 198, + 578 + ], + "spans": [ + { + "bbox": [ + 141, + 521, + 198, + 578 + ], + "type": "image", + "image_path": "59172be50245465e690d2222ac9df98d8c7460137f7bd7b3047653052629289e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 582, + 194, + 593 + ], + "lines": [ + { + "bbox": [ + 148, + 582, + 194, + 593 + ], + "spans": [ + { + "bbox": [ + 148, + 582, + 194, + 593 + ], + "type": "text", + "content": "SadTalker" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 201, + 233, + 257, + 290 + ], + "blocks": [ + { + "bbox": [ + 201, + 233, + 257, + 290 + ], + "lines": [ + { + "bbox": [ + 201, + 233, + 257, + 290 + ], + "spans": [ + { + "bbox": [ + 201, + 233, + 257, + 290 + ], + "type": "image", + "image_path": "8015fb8dbe326105aeef7266da0bc8d5032495dfb21fe450aa13ea4e4b60254f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 201, + 291, + 257, + 347 + ], + "blocks": [ + { + "bbox": [ + 201, + 291, + 257, + 347 + ], + "lines": [ + { + "bbox": [ + 201, + 291, + 257, + 347 + ], + "spans": [ + { + "bbox": [ + 201, + 291, + 257, + 347 + ], + "type": "image", + "image_path": "f843bfe93a8854f9c9832d7b2074e8702fd7c2fa40ed67b3aa7be1e019513f97.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 201, + 348, + 257, + 404 + ], + "blocks": [ + { + "bbox": [ + 201, + 348, + 257, + 404 + ], + "lines": [ + { + "bbox": [ + 201, + 348, + 257, + 404 + ], + "spans": [ + { + "bbox": [ + 201, + 348, + 257, + 404 + ], + "type": "image", + "image_path": "e9468345cef7c2d54ac66f1843359f0a378928dec857ed3cf5ee881b6a088568.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 201, + 405, + 257, + 461 + ], + "blocks": [ + { + "bbox": [ + 201, + 405, + 257, + 461 + ], + "lines": [ + { + "bbox": [ + 201, + 405, + 257, + 461 + ], + "spans": [ + { + "bbox": [ + 201, + 405, + 257, + 461 + ], + "type": "image", + "image_path": "10c2f135edbbee68023574edbbfada5a9f68ef62327731391183cd1bcd0da3b8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 201, + 462, + 257, + 519 + ], + "blocks": [ + { + "bbox": [ + 201, + 462, + 257, + 519 + ], + "lines": [ + { + "bbox": [ + 201, + 462, + 257, + 519 + ], + "spans": [ + { + "bbox": [ + 201, + 462, + 257, + 519 + ], + "type": "image", + "image_path": "05a631022453d1660e114aaaba2cc53d4bb5096327f33afc328fc3ba0f7166ff.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 201, + 521, + 257, + 578 + ], + "blocks": [ + { + "bbox": [ + 201, + 521, + 257, + 578 + ], + "lines": [ + { + "bbox": [ + 201, + 521, + 257, + 578 + ], + "spans": [ + { + "bbox": [ + 201, + 521, + 257, + 578 + ], + "type": "image", + "image_path": "ab648c7601b18a21a2b02621c86d26979812cb9633f52c63312e8f69069c3f0a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 582, + 251, + 594 + ], + "lines": [ + { + "bbox": [ + 208, + 582, + 251, + 594 + ], + "spans": [ + { + "bbox": [ + 208, + 582, + 251, + 594 + ], + "type": "text", + "content": "VExpress" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 260, + 233, + 317, + 290 + ], + "blocks": [ + { + "bbox": [ + 260, + 233, + 317, + 290 + ], + "lines": [ + { + "bbox": [ + 260, + 233, + 317, + 290 + ], + "spans": [ + { + "bbox": [ + 260, + 233, + 317, + 290 + ], + "type": "image", + "image_path": "8aaed49357062d3953f5c4816ca89b0eb6053b1604c5c01f724e218f38354fef.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 260, + 291, + 317, + 347 + ], + "blocks": [ + { + "bbox": [ + 260, + 291, + 317, + 347 + ], + "lines": [ + { + "bbox": [ + 260, + 291, + 317, + 347 + ], + "spans": [ + { + "bbox": [ + 260, + 291, + 317, + 347 + ], + "type": "image", + "image_path": "79b39434fbf19a06bc7772134181cbbeaf88825ccc5cda9d5a203de14e06464e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 260, + 348, + 317, + 404 + ], + "blocks": [ + { + "bbox": [ + 260, + 348, + 317, + 404 + ], + "lines": [ + { + "bbox": [ + 260, + 348, + 317, + 404 + ], + "spans": [ + { + "bbox": [ + 260, + 348, + 317, + 404 + ], + "type": "image", + "image_path": "c5e6955aa23c03e5239673851426de0ef2f48c75c88458c2863b7e58019a0129.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 260, + 405, + 317, + 461 + ], + "blocks": [ + { + "bbox": [ + 260, + 405, + 317, + 461 + ], + "lines": [ + { + "bbox": [ + 260, + 405, + 317, + 461 + ], + "spans": [ + { + "bbox": [ + 260, + 405, + 317, + 461 + ], + "type": "image", + "image_path": "d2ba2acbdab5a12d821be172edc8a24f987205b67dad25b3b6715a21cbf1e02e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 260, + 462, + 317, + 519 + ], + "blocks": [ + { + "bbox": [ + 260, + 462, + 317, + 519 + ], + "lines": [ + { + "bbox": [ + 260, + 462, + 317, + 519 + ], + "spans": [ + { + "bbox": [ + 260, + 462, + 317, + 519 + ], + "type": "image", + "image_path": "d65e08e5b3d5505f7380ca15dc11f8fda68f532449289a90865b226d235c3710.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 260, + 521, + 317, + 578 + ], + "blocks": [ + { + "bbox": [ + 260, + 521, + 317, + 578 + ], + "lines": [ + { + "bbox": [ + 260, + 521, + 317, + 578 + ], + "spans": [ + { + "bbox": [ + 260, + 521, + 317, + 578 + ], + "type": "image", + "image_path": "cad8d5239aa04828f15dcbbcb471133146b6d36a281e6ae3da917846c8ff8689.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 582, + 303, + 593 + ], + "lines": [ + { + "bbox": [ + 278, + 582, + 303, + 593 + ], + "spans": [ + { + "bbox": [ + 278, + 582, + 303, + 593 + ], + "type": "text", + "content": "Hallo" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 319, + 233, + 376, + 290 + ], + "blocks": [ + { + "bbox": [ + 319, + 233, + 376, + 290 + ], + "lines": [ + { + "bbox": [ + 319, + 233, + 376, + 290 + ], + "spans": [ + { + "bbox": [ + 319, + 233, + 376, + 290 + ], + "type": "image", + "image_path": "170bea71c5883092017da89934c2b99ee8b7a04f49ccbf7797d32b4d27dacb5f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 319, + 291, + 376, + 347 + ], + "blocks": [ + { + "bbox": [ + 319, + 291, + 376, + 347 + ], + "lines": [ + { + "bbox": [ + 319, + 291, + 376, + 347 + ], + "spans": [ + { + "bbox": [ + 319, + 291, + 376, + 347 + ], + "type": "image", + "image_path": "d071fb204fd7dc3bdefe55a6ab3ba66de87d9575d4e0464bdd1a02e995d736cc.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 319, + 348, + 376, + 404 + ], + "blocks": [ + { + "bbox": [ + 319, + 348, + 376, + 404 + ], + "lines": [ + { + "bbox": [ + 319, + 348, + 376, + 404 + ], + "spans": [ + { + "bbox": [ + 319, + 348, + 376, + 404 + ], + "type": "image", + "image_path": "ecddce9a1170a25585f47cf0e07c8680858a799b6fd0a877cbac2b2d5dd90d9c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 319, + 405, + 376, + 461 + ], + "blocks": [ + { + "bbox": [ + 319, + 405, + 376, + 461 + ], + "lines": [ + { + "bbox": [ + 319, + 405, + 376, + 461 + ], + "spans": [ + { + "bbox": [ + 319, + 405, + 376, + 461 + ], + "type": "image", + "image_path": "4fff0c994435277323fba42650c8717f8b8efd995ff56ebee5da37a818be8046.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 319, + 462, + 376, + 519 + ], + "blocks": [ + { + "bbox": [ + 319, + 462, + 376, + 519 + ], + "lines": [ + { + "bbox": [ + 319, + 462, + 376, + 519 + ], + "spans": [ + { + "bbox": [ + 319, + 462, + 376, + 519 + ], + "type": "image", + "image_path": "4da9d5aee4760b00c4809d7f2720463143c9036e1ea50580f84b242ad9061242.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 319, + 521, + 376, + 578 + ], + "blocks": [ + { + "bbox": [ + 319, + 521, + 376, + 578 + ], + "lines": [ + { + "bbox": [ + 319, + 521, + 376, + 578 + ], + "spans": [ + { + "bbox": [ + 319, + 521, + 376, + 578 + ], + "type": "image", + "image_path": "efc6d9aa44cbff41048963277ad56a377c79705f67955bc0c55a330936b74c2b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 582, + 374, + 593 + ], + "lines": [ + { + "bbox": [ + 323, + 582, + 374, + 593 + ], + "spans": [ + { + "bbox": [ + 323, + 582, + 374, + 593 + ], + "type": "text", + "content": "EchoMimic" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 73, + 599, + 536, + 611 + ], + "lines": [ + { + "bbox": [ + 73, + 599, + 536, + 611 + ], + "spans": [ + { + "bbox": [ + 73, + 599, + 536, + 611 + ], + "type": "text", + "content": "Figure 2. Qualitative comparisons between FADA and baselines across different portraits and pronunciations in openset." + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 378, + 233, + 435, + 290 + ], + "blocks": [ + { + "bbox": [ + 378, + 233, + 435, + 290 + ], + "lines": [ + { + "bbox": [ + 378, + 233, + 435, + 290 + ], + "spans": [ + { + "bbox": [ + 378, + 233, + 435, + 290 + ], + "type": "image", + "image_path": "ade7d7553527da62a23fc74917a0a83fb79bc35b6ae3dbccf97f087533e07a53.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 378, + 291, + 435, + 347 + ], + "blocks": [ + { + "bbox": [ + 378, + 291, + 435, + 347 + ], + "lines": [ + { + "bbox": [ + 378, + 291, + 435, + 347 + ], + "spans": [ + { + "bbox": [ + 378, + 291, + 435, + 347 + ], + "type": "image", + "image_path": "179103c7bb34be9b57af474cf32e2e3b34996a5250d640e3a86f2a58a1b5f427.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 378, + 348, + 435, + 404 + ], + "blocks": [ + { + "bbox": [ + 378, + 348, + 435, + 404 + ], + "lines": [ + { + "bbox": [ + 378, + 348, + 435, + 404 + ], + "spans": [ + { + "bbox": [ + 378, + 348, + 435, + 404 + ], + "type": "image", + "image_path": "eeede619c8276360fadb778ee703fb6bb041e366a93daf1dfd2e0452807f379e.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 378, + 405, + 435, + 461 + ], + "blocks": [ + { + "bbox": [ + 378, + 405, + 435, + 461 + ], + "lines": [ + { + "bbox": [ + 378, + 405, + 435, + 461 + ], + "spans": [ + { + "bbox": [ + 378, + 405, + 435, + 461 + ], + "type": "image", + "image_path": "461adbc1a22ec6f2ae2925680533c46176e854c91a9277fda14911d6b72b4e4b.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 378, + 462, + 435, + 519 + ], + "blocks": [ + { + "bbox": [ + 378, + 462, + 435, + 519 + ], + "lines": [ + { + "bbox": [ + 378, + 462, + 435, + 519 + ], + "spans": [ + { + "bbox": [ + 378, + 462, + 435, + 519 + ], + "type": "image", + "image_path": "f706423712a94d3b05334856ca25061f3ddaccaab26a328d2ad3009f60bd2f8a.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 378, + 521, + 435, + 578 + ], + "blocks": [ + { + "bbox": [ + 378, + 521, + 435, + 578 + ], + "lines": [ + { + "bbox": [ + 378, + 521, + 435, + 578 + ], + "spans": [ + { + "bbox": [ + 378, + 521, + 435, + 578 + ], + "type": "image", + "image_path": "649a842008a9a793d8093ad06fafa5cc5a627e43e0d67eaea9684c24e3143702.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 582, + 429, + 593 + ], + "lines": [ + { + "bbox": [ + 386, + 582, + 429, + 593 + ], + "spans": [ + { + "bbox": [ + 386, + 582, + 429, + 593 + ], + "type": "text", + "content": "Ours-Fast" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 436, + 233, + 494, + 290 + ], + "blocks": [ + { + "bbox": [ + 436, + 233, + 494, + 290 + ], + "lines": [ + { + "bbox": [ + 436, + 233, + 494, + 290 + ], + "spans": [ + { + "bbox": [ + 436, + 233, + 494, + 290 + ], + "type": "image", + "image_path": "3b128be656c081bf8542c7de2fe9e4d0282de8ccd3f6e7f784cdf2dabdf713da.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 252, + 528, + 273 + ], + "lines": [ + { + "bbox": [ + 507, + 252, + 528, + 273 + ], + "spans": [ + { + "bbox": [ + 507, + 252, + 528, + 273 + ], + "type": "text", + "content": "[ə]" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 436, + 291, + 494, + 347 + ], + "blocks": [ + { + "bbox": [ + 436, + 291, + 494, + 347 + ], + "lines": [ + { + "bbox": [ + 436, + 291, + 494, + 347 + ], + "spans": [ + { + "bbox": [ + 436, + 291, + 494, + 347 + ], + "type": "image", + "image_path": "cb5ca9432ba958afd9bc166698949d523baa2b6c695b39135eae47c5f319c303.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 306, + 525, + 327 + ], + "lines": [ + { + "bbox": [ + 509, + 306, + 525, + 327 + ], + "spans": [ + { + "bbox": [ + 509, + 306, + 525, + 327 + ], + "type": "text", + "content": "[I]" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 436, + 348, + 494, + 404 + ], + "blocks": [ + { + "bbox": [ + 436, + 348, + 494, + 404 + ], + "lines": [ + { + "bbox": [ + 436, + 348, + 494, + 404 + ], + "spans": [ + { + "bbox": [ + 436, + 348, + 494, + 404 + ], + "type": "image", + "image_path": "9e528c0a630a73b22b50fc3ffa9bbf19e25a5ae6cb836dbf90ab43c89c19ad00.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 365, + 533, + 386 + ], + "lines": [ + { + "bbox": [ + 507, + 365, + 533, + 386 + ], + "spans": [ + { + "bbox": [ + 507, + 365, + 533, + 386 + ], + "type": "text", + "content": "[u:]" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 436, + 405, + 494, + 461 + ], + "blocks": [ + { + "bbox": [ + 436, + 405, + 494, + 461 + ], + "lines": [ + { + "bbox": [ + 436, + 405, + 494, + 461 + ], + "spans": [ + { + "bbox": [ + 436, + 405, + 494, + 461 + ], + "type": "image", + "image_path": "1bf0e77c96ef04f4ef787db97483c415c20c498bfaac4e025072dcd12c1cf1e6.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 423, + 534, + 444 + ], + "lines": [ + { + "bbox": [ + 507, + 423, + 534, + 444 + ], + "spans": [ + { + "bbox": [ + 507, + 423, + 534, + 444 + ], + "type": "text", + "content": "[æ]" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 436, + 462, + 494, + 519 + ], + "blocks": [ + { + "bbox": [ + 436, + 462, + 494, + 519 + ], + "lines": [ + { + "bbox": [ + 436, + 462, + 494, + 519 + ], + "spans": [ + { + "bbox": [ + 436, + 462, + 494, + 519 + ], + "type": "image", + "image_path": "3a1d2467dd1c9f02e0cbfef74f46a3195208027f05e316d84eaadb916798c216.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 479, + 533, + 499 + ], + "lines": [ + { + "bbox": [ + 507, + 479, + 533, + 499 + ], + "spans": [ + { + "bbox": [ + 507, + 479, + 533, + 499 + ], + "type": "text", + "content": "[eI]" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 436, + 521, + 494, + 578 + ], + "blocks": [ + { + "bbox": [ + 436, + 521, + 494, + 578 + ], + "lines": [ + { + "bbox": [ + 436, + 521, + 494, + 578 + ], + "spans": [ + { + "bbox": [ + 436, + 521, + 494, + 578 + ], + "type": "image", + "image_path": "2f34cf9edabf69e80faffe5dd1da9ac4fd34612b4d7bf5bf9906af72e208d53f.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 582, + 500, + 593 + ], + "lines": [ + { + "bbox": [ + 436, + 582, + 500, + 593 + ], + "spans": [ + { + "bbox": [ + 436, + 582, + 500, + 593 + ], + "type": "text", + "content": "Ours-Balanced" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 509, + 535, + 527, + 557 + ], + "lines": [ + { + "bbox": [ + 509, + 535, + 527, + 557 + ], + "spans": [ + { + "bbox": [ + 509, + 535, + 527, + 557 + ], + "type": "text", + "content": "[1]" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "text", + "content": "and student models under different data qualities and scales, it can be demonstrated that mixed supervised distillation actually enhances the model's utilization of moderate-quality data. As depicted in Tables 2 and 3, when the teacher model is trained with a large-scale moderate-quality dataset, a decline in model performance is observed, whereas the student model exhibits the opposite trend. Briefly, only mixed-supervised distillation can take advantage of the moderate-" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 313, + 620, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 555, + 656 + ], + "type": "text", + "content": "quality data. More analysis of mixed-supervised distillation will be discussed and more hyper-parameters experiments will be conducted in Appendix A.1." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Analysis of Multi-CFG Distillation As demonstrated in Table 4, we further explore the effects of multi-CFG distillation on top of mixed-supervised distillation. Initially, we tested the results of conditional reasoning without CFG dis" + } + ] + } + ], + "index": 60 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3107" + } + ] + } + ], + "index": 61 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 113, + 288, + 183 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": "Table 2. Ablation study of basic distillation methods on HDTF test set. All methods in this table are free from mixed-supervised and CFG distillation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 113, + 288, + 183 + ], + "lines": [ + { + "bbox": [ + 62, + 113, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 62, + 113, + 288, + 183 + ], + "type": "table", + "html": "
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Teacher with B753.8858.12319.6421.401.689
TCD183.8277.93723.0123.151.893
RFlow182.8368.06681.1049.732.484
PeRFlow183.8238.00121.9121.261.477
", + "image_path": "4b89f26b7e5334109cfceabdfca7281cf4d08629414ff98da3b1c04943ee2472.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 60, + 265, + 293, + 383 + ], + "blocks": [ + { + "bbox": [ + 55, + 195, + 296, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 262 + ], + "type": "text", + "content": "Table 3. Ablation study of adaptive mixed-supervised distillation on HDTF test set. All methods use PeRFlow with dataset " + }, + { + "bbox": [ + 55, + 195, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 195, + 296, + 262 + ], + "type": "text", + "content": " without CFG distillation. Fixed-x indicates fixed ground-truth loss with weight x. Adaptive-unlimited-s and Adaptive-s indicate adaptive ground-truth loss with hyper-parameter " + }, + { + "bbox": [ + 55, + 195, + 296, + 262 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 195, + 296, + 262 + ], + "type": "text", + "content": ", with the former being without peak and dead threshold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 265, + 293, + 383 + ], + "lines": [ + { + "bbox": [ + 60, + 265, + 293, + 383 + ], + "spans": [ + { + "bbox": [ + 60, + 265, + 293, + 383 + ], + "type": "table", + "html": "
MethodNFE-DIQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
Stu. with A183.8238.00121.9121.261.477
Stu. with B (Fixed-0)183.7627.90925.0621.471.462
+ Fixed-0.2183.8637.93413.7119.941.434
+ Fixed-1.0183.8127.90324.5321.421.494
+ Fixed-0.2, only A183.8807.96215.3220.711.512
+ Adaptive-unl.-0.5183.8247.91415.1520.741.471
+ Adaptive-unl.-0.25183.8467.90614.2019.921.424
+ Adaptive-0.5183.9497.91913.8318.641.395
+ Adaptive-0.25183.9517.92513.3218.391.362
", + "image_path": "cd92bf2faa5dbb59391b5164bb9963392b7811728b0745b78e534eec8c573fb7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 441, + 293, + 513 + ], + "blocks": [ + { + "bbox": [ + 55, + 394, + 295, + 439 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 439 + ], + "type": "text", + "content": "Table 4. Ablation study of multi-CFG distillation on HDTF test set. All methods in this table are PeRFlow + mixed-supervised distillation. CFG time embed refers to CFG distillation through timestep embedding injection." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 441, + 293, + 513 + ], + "lines": [ + { + "bbox": [ + 58, + 441, + 293, + 513 + ], + "spans": [ + { + "bbox": [ + 58, + 441, + 293, + 513 + ], + "type": "table", + "html": "
MethodNFE-D IQA↑Sync-D↓FVD-R↓FID↓E-FID↓
Teacher with A753.9987.84918.2718.491.365
w.o. CFG63.6749.02027.1026.562.096
+ CFG time embed63.7018.57723.2422.471.791
+ CFG layer63.7467.90620.0520.941.620
+ CFG layer w. tokens63.9277.99616.6718.511.635
", + "image_path": "d549731a693bc54c5f60526cee23a89a2b6e711e8029982291efc3f3bfac2a46.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 63, + 515, + 285, + 605 + ], + "blocks": [ + { + "bbox": [ + 63, + 515, + 285, + 605 + ], + "lines": [ + { + "bbox": [ + 63, + 515, + 285, + 605 + ], + "spans": [ + { + "bbox": [ + 63, + 515, + 285, + 605 + ], + "type": "image", + "image_path": "61c03eae7ba7a2ff50a84cec046c655e0339ca057af1a5cc2d4866890c852c42.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 610, + 295, + 644 + ], + "lines": [ + { + "bbox": [ + 55, + 610, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 295, + 644 + ], + "type": "text", + "content": "Figure 3. Line charts showing the variations of FVD and Sync-D metrics with different audio CFG on HDTF test set. Reference CFG is set to 2.0 in this figure." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "tillation, which significantly deteriorated the model performance, especially leading to a doubling of the FVD metric representing video quality. For techniques like time embedding injection similar to " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "-condition [20], they partially restored model performance but with limited efficacy. In con" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 545, + 163 + ], + "type": "image", + "image_path": "92f17b15c377942032f3430bc05d4cd9c1e12232f53f0a0bde41f39fb07bed04.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 168, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 203 + ], + "type": "text", + "content": "Figure 4. Line charts showing the variations of FVD and Sync-D metrics with different ref CFG on HDTF test set. Audio CFG is set to 6.5 in this figure." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 220, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 314 + ], + "type": "text", + "content": "trast, our proposed learnable token-based CFG control layers significantly enhance model performance, nearly matching the teacher model and even surpassing it in FVD. Notably, our model requires only " + }, + { + "bbox": [ + 313, + 220, + 555, + 314 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 220, + 555, + 314 + ], + "type": "text", + "content": " of the teacher model's NFE. We also tested without learnable tokens, and while the pure CFG control layer had some effect, it could not match the teacher model, especially in overall video quality metrics like FVD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 317, + 556, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 556, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 556, + 503 + ], + "type": "text", + "content": "To verify whether our proposed multi-CFG distillation effectively captures the influence of audio and reference conditions on the generated results, we plotted line charts of FVD and Sync-D metrics with variations in audio CFG and reference CFG, as shown in Figures 3 and 4. First, we observed improvements in both FVD and Sync-D as audio CFG increased, reaching a stable maximum value after 6.5. Hence, we ultimately chose 6.5 as the audio CFG for inference. Second, the influence of reference CFG on FVD and Sync-D metrics displayed an initially positive, then negative trend, with optimal performance achieved around the 2.0 position. These results indicate that our CFG control layer effectively mimics the multi-CFG process and achieves good results. More visualized examples and performance analysis under different scenes will be presented in Appendix C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 526, + 388, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 526, + 388, + 539 + ], + "spans": [ + { + "bbox": [ + 314, + 526, + 388, + 539 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 556, + 715 + ], + "type": "text", + "content": "In summary, we proposed FADA, a fast diffusion avatar synthesis framework with mixed-supervised multi-CFG distillation. By designing a mixed-supervised loss, we leverage data of varying quality to enhance the robustness of generated results. Furthermore, a learnable token-based multi-CFG condition design is introduced to maintain the correlation between the audio and the generated video in the distilled models. The quantitative and qualitative experiments across multiple datasets show that our balanced setting achieves fast inference and high-fidelity generation ability. Ablation studies prove our proposed methods are effective in FADA. The limitations, future works and ethical concerns of this paper will be discussed in Appendix B, H and I respectively." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3108" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "content": "This project is sponsored by National Natural Science Foundation of China under Grant No.62222211 and National Natural Science Foundation of China under Grant No.U24A20326." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 168, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 168, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 296, + 212 + ], + "type": "text", + "content": "[1] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 213, + 296, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 296, + 264 + ], + "type": "text", + "content": "[2] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2: Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 266, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 296, + 319 + ], + "type": "text", + "content": "[3] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 321, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 296, + 364 + ], + "type": "text", + "content": "[4] Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 365, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 365, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 61, + 365, + 296, + 418 + ], + "type": "text", + "content": "[5] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 296, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 296, + 461 + ], + "type": "text", + "content": "[6] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 463, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 296, + 495 + ], + "type": "text", + "content": "[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 496, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 496, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 496, + 296, + 548 + ], + "type": "text", + "content": "[8] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 550, + 296, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 550, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 62, + 550, + 296, + 593 + ], + "type": "text", + "content": "[9] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 594, + 295, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 594, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 57, + 594, + 295, + 636 + ], + "type": "text", + "content": "[10] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 638, + 295, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 638, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 57, + 638, + 295, + 659 + ], + "type": "text", + "content": "[11] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 660, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 660, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 660, + 295, + 713 + ], + "type": "text", + "content": "[12] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 115 + ], + "type": "text", + "content": "[13] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "text", + "content": "[14] Gaojie Lin, Jianwen Jiang, Chao Liang, Tianyun Zhong, Ji-aqi Yang, and Yanbo Zheng. Cyberhost: Taming audiodriven avatar diffusion model with region codebook attention. arXiv preprint arXiv:2409.01876, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[15] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 226 + ], + "type": "text", + "content": "[16] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "text", + "content": "[17] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "text", + "content": "[18] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 318, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 318, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 318, + 554, + 371 + ], + "type": "text", + "content": "[19] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "text", + "content": "[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 428, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 480 + ], + "type": "text", + "content": "[21] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 483, + 554, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 554, + 535 + ], + "type": "text", + "content": "[22] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 537, + 554, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 591 + ], + "type": "text", + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[24] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 626, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 553, + 669 + ], + "type": "text", + "content": "[25] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[26] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3109" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "type": "text", + "content": "[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[29] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 185, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 237 + ], + "type": "text", + "content": "[30] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[31] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 285, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 325 + ], + "type": "text", + "content": "[32] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. OpenReview, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 328, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 328, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 294, + 381 + ], + "type": "text", + "content": "[33] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 435 + ], + "type": "text", + "content": "[34] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[35] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 545 + ], + "type": "text", + "content": "[36] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "text", + "content": "[37] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 644 + ], + "type": "text", + "content": "[38] Zhenhui Ye, Jinzheng He, Ziyue Jiang, Rongjie Huang, Jiawei Huang, Jinglin Liu, Yi Ren, Xiang Yin, Zejun Ma, and Zhou Zhao. Geneface++: Generalized and stable real-time audio-driven 3d talking face generation. arXiv preprint arXiv:2305.00787, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[39] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[40] Zhenhui Ye, Tianyun Zhong, Yi Ren, Ziyue Jiang, Jiawei Huang, Rongjie Huang, Jinglin Liu, Jinzheng He, Chen" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 590 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Zhang, Zehan Wang, et al. Mimictalk: Mimicking a personalized and expressive 3d talking face in minutes. arXiv preprint arXiv:2410.06734, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 158 + ], + "type": "text", + "content": "[41] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 160, + 553, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 160, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 316, + 160, + 553, + 223 + ], + "type": "text", + "content": "[42] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 224, + 555, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 555, + 277 + ], + "type": "text", + "content": "[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 279, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 279, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 279, + 553, + 320 + ], + "type": "text", + "content": "[44] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 322, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 384 + ], + "type": "text", + "content": "[45] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Wan Rui, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, et al. Vedit-q: Efficient and accurate quantization of diffusion transformers for image and video generation. arXiv preprint arXiv:2406.02540, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 553, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 429 + ], + "type": "text", + "content": "[46] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation. arXiv preprint arXiv:2402.19159, 2024. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 430, + 553, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 493 + ], + "type": "text", + "content": "[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 495, + 553, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 495, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 316, + 495, + 553, + 537 + ], + "type": "text", + "content": "[48] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 539, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 539, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 539, + 553, + 590 + ], + "type": "text", + "content": "[49] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3110" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_content_list.json b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4836db9e57a71f11fedac0f45736a5d36209cdfb --- /dev/null +++ b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_content_list.json @@ -0,0 +1,1679 @@ +[ + { + "type": "text", + "text": "FADE: Frequency-Aware Diffusion Model Factorization for Video Editing", + "text_level": 1, + "bbox": [ + 125, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixuan Zhu $^{1}$ Haolin Wang $^{1}$ Shilin Ma $^{2}$ Wenliang Zhao $^{1}$ Yansong Tang $^{2}$ Lei Chen $^{1,\\dagger}$ Jie Zhou $^{1}$", + "bbox": [ + 99, + 179, + 926, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Automation, Tsinghua University $^{2}$ Tsinghua Shenzhen International Graduate School, Tsinghua University", + "bbox": [ + 223, + 215, + 800, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7ec1f7755813d7983b26a29f2acb0334a1a86be45e65dab002f59ad5f3e74e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 268, + 194, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7c5c16893f07e6a046cc93ab0d18a974bfa57651d290e006f017b643157ef50.jpg", + "image_caption": [ + "Source Prompt: A black swan swimming in a river beside bushes" + ], + "image_footnote": [], + "bbox": [ + 196, + 268, + 295, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de04e1b6a5ea46e26317a6f9fedaa4e7f39e0ed250dcfeb677fa8e8650432a84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 268, + 395, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6be7ba932d243646a4dd7595f157220bf0b59f917c832fb7f8eaab46974916ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 268, + 496, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02067fb34f344144eeebb4f042d07bfb4b7487e1288cb4d16d7f0578c9242a2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 268, + 596, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae9adeb1a96b522f8b4d2da0bb3d2b6c33ef2a1b5c01fd5849109133726903be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 268, + 697, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2901c2e83ce290be7cf7275af39cf2613eb278b30c07ff381767bacc5955d98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 268, + 799, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ddfe18cb1a32226afe3e51c09950dccfd2fdd4b4cab147730c255f8f1d1bc78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 268, + 900, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6af6ec504268b3657596b005b39175d03f56635d471a31991416c054144a8f6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 364, + 194, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47c7f95d9301121bcac84846bb6d68da6fd5a8b65203569ebdbd597f52a94379.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 364, + 295, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ae7fc593bf28322649e45f1a4c47f570798fbb33748e01ef081fbddd4e8478a.jpg", + "image_caption": [ + "Appearance Edit: black swan $\\rightarrow$ white duck" + ], + "image_footnote": [], + "bbox": [ + 297, + 364, + 395, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4eefc3ed9ede70715c6e624a9def7df13a20cfb7592d6260367d44fb5f702a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 364, + 496, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9eff71770f5449bb33a63adfe012fdcc60dc32ba03e811186894603972ab3fae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 364, + 596, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59547d8b164f4e1c8f7664d5413d54cff4f73d9395e5fc30ee2449a6c2e318de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 364, + 697, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa07e97e6c01fece5f6d9569f5773fc3b1f020e3b2e5d9d86453c133bd96d6a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 364, + 799, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f03fc8cbf1ebb92c424e541918311addc84248eebf43b64b24328c5bc6d07e15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 364, + 900, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fb1a935cbfc356e0c7af102a8ceabdd7583e04e2ed7e1e724b1c2de5d9d2b7e.jpg", + "image_caption": [ + "Figure 1. Diverse video editing results of FADE. Our training-free approach, utilizing frequency-aware factorization and modulation, achieves high-fidelity, coherent edits across a variety of video types. FADE handles both appearance and motion adjustments with impressive robustness, ensuring precise alignment with input prompts and maintaining temporal consistency. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 94, + 458, + 194, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/971902c1377cd3ca873fa0f90cdd6442930b4fd1ad8e51d30949fcd2624e0834.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 458, + 295, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/391538d544dc45bd5d2a6c6dad3187bf0b988ab1bf5413402fb4daa60cce96f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 458, + 395, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a9398fc8ecb7acd2b2d9427b72d645f1de3a11380f38c9a25ad88aa0408a891.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 458, + 496, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3501d1ea27decfb3d8cbb1ca18336ab0fc1f562e6cb1ece2651ebfe0bee9d22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 458, + 596, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a63ce0aeb98d19eb4ab8482ac965b5153f3dab3474a6b724c6516fcf074de616.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 458, + 697, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b8226ae7b20ff1cb84c46fc0c22c46ef2889992f617420ad2e0b94040e28459.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 458, + 799, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7ff6632fc4a9321f61f72a784c56390b9dbbce202bb21b0493977df17a459ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 458, + 900, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motion Edit: swimming $\\rightarrow$ flapping the wings", + "bbox": [ + 336, + 536, + 653, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 611, + 326, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in diffusion frameworks have significantly enhanced video editing, achieving high fidelity and strong alignment with textual prompts. However, conventional approaches using image diffusion models fall short in handling video dynamics, particularly for challenging temporal edits like motion adjustments. While current video diffusion models produce high-quality results, adapting them for efficient editing remains difficult due to the heavy computational demands that prevent the direct application of previous image editing techniques. To overcome these limitations, we introduce FADE—a training-free yet highly effective video editing approach that fully leverages the inherent priors from pre-trained video diffusion models via frequency-aware factorization. Rather than simply using these models, we first analyze the attenuated performance of the proposed system.", + "bbox": [ + 88, + 643, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion patterns within the video model to reveal how video priors are distributed across different components. Building on these insights, we propose a factorization strategy to optimize each component's specialized role. Furthermore, we devise spectrum-guided modulation to refine the sampling trajectory with frequency domain cues, preventing information leakage and supporting efficient, versatile edits while preserving the basic spatial and temporal structure. Extensive experiments on real-world videos demonstrate that our method consistently delivers high-quality, realistic and temporally coherent editing results both qualitatively and quantitatively. Code is available at https://github.com/EternalEvan/FADE.", + "bbox": [ + 511, + 613, + 906, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 843, + 645, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The objective of video editing is to modify specific visual content consistently across all frames while preserving the", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 114, + 887, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28426", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ceada60a161a26322941d9a46db038f2bc1795e6a32a141b3bfe2a214153506.jpg", + "image_caption": [ + "Figure 2. (a) In typical T2V models, video tokens flattened along the temporal axis yield full-attention maps split into cross, spatial, and temporal attention. (b) Video priors in T2V models are factorized based on attention feature patterns. We visualize the attention map $M$ and result $F$ along with its spectrum for two blocks. The early block sketches the foundational structure and movement, while the later block refines details. This insight allows us to leverage each block's unique role for more efficient editing." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 480, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "integrity of unchanged regions. Recent achievements in diffusion models have greatly elevated video editing capabilities [10, 12, 15, 19, 20, 28, 32], bringing remarkable generative power to video editing tasks that require both spatial details and temporal coherence. While previous methods [1, 8, 13, 18, 30] have adapted text-to-image (T2I) diffusion models for prompt-driven image editing, their effectiveness diminishes when applied to video due to the complexities of managing spatial-temporal relationships across frames and a lack of generalized video prior knowledge. These limitations underscore the need for an efficient and robust editing strategy capable of handling the increased data and temporal coherence demands of videos. In response, we aim to design an accessible, high-quality editing solution that fully leverages the rich video priors embedded in pre-trained video diffusion models, enabling realistic, coherent, and versatile edits, including both appearance and motion adjustments, across complex video content.", + "bbox": [ + 88, + 460, + 482, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge in video editing is balancing changes to the target object while preserving unedited content. Previous methods largely rely on the generative power of T2I models for controlled edits. For instance, [15, 33] achieve this through null-text inversion [18], which minimizes inversion errors via test-time optimization. However, this process requires extensive iterations to compute a series of null-text embeddings for every sampling step, making it time-consuming and limiting its real-world applicability. Another line of work is attention feature injection [5, 15, 23, 27], where attention maps are modified by", + "bbox": [ + 89, + 734, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "replacing or interleaving to maintain content consistency. Despite this, the absence of video priors often results in unrealistic temporal dynamics and limits motion edit capacity. Some methods, like [19], directly utilize video diffusion models by adding noise to the source video. This straightforward approach fails to fully exploit video priors, resulting in suboptimal editing quality. With the recent emergence of advanced text-to-video (T2V) models, new opportunities arise to harness their rich video knowledge. However, adapting previous T2I techniques to these models is challenging, given the extensive resources required, including memory, computation and time, to handle the increased data and complexity inherent in large-scale video networks.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these challenges, we introduce FADE, a training-free yet potent approach for real-world video editing. Our goal is to harness the rich video priors about video appearance and dynamics within pre-trained T2V models to enable realistic and diverse editing. To achieve this, we start by examining the specific properties of these models. While previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps, we further observe that a similar frequency progression occurs block-wise within T2V models. We explore how video priors function in each block by analyzing their roles and patterns, revealing that early blocks sketch foundational spatial layouts and temporal dynamics while later blocks refine high-frequency details (Figure 2). Leveraging this insight, we factorize the editing problem by the distinct functions of each component, enhancing editing efficiency. Then, instead of performing optimization or attention injection, we design a spectrum-guided modulation strategy that utilizes the full-attention outputs in these sketching blocks to derive frequency-aware guidance through 3D Fourier transformation. This modulation adjusts the sampling trajectory to preserve low-frequency structure from the source video while allowing high-frequency details to support flexible and intricate edits. This cue from the frequency domain not only mitigates information leakage during guidance computation but also unleashes the potential of video diffusion priors, enabling sophisticated spatial and temporal edits with high fidelity and coherence.", + "bbox": [ + 511, + 292, + 906, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method using extensive real-world video clips across various editing tasks. As shown in Figure 1, FADE exhibits proficiency in producing realistic and consistent edits, covering both appearance and motion modifications. Quantitative comparisons on video editing benchmark DAVIS [22] demonstrate our approach achieves superior textual alignment and temporal consistency, with 0.3561 CLIP score and 43.57 OSV. Moreover, we maintain competitive input fidelity with 21.54 MaskPSNR and 0.3297 LPIPS, preserving unedited regions effectively. Compared to previous methods, our use of general video priors with frequency-aware factorization deliv", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28427", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/408f92fb59348413bcb1f228068ff6dc11626f7bf0d4d6e1cac88f616319973b.jpg", + "image_caption": [ + "Figure 3. The overall framework of FADE. FADE is a training-free framework for video editing. Given an input video, we first perform video inversion to obtain noise $\\mathbf{z}_T$ and trajectory $\\{\\mathbf{z}_t^*\\}_{t=0}^T$ . Starting from this noise, we use sketching blocks in the T2V model to produce full-attention results $\\mathbf{F}_t$ and $\\mathbf{F}_t^*$ from $\\mathbf{z}_t$ and $\\mathbf{z}_t^*$ . These results are transformed to the frequency domain, where a low-pass filter isolates spatial and temporal structures. The frequency error then modulates the sampling process, enabling high-quality, consistent edits." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 897, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ers improved efficiency and a broader editing range, highlighting FADE's outstanding versatility and practicality.", + "bbox": [ + 89, + 381, + 482, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 431, + 241, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Diffusion Models. In recent years, denoising diffusion models [9, 29] have emerged as a leading class of generative models, notable for their synthesis quality and controllability. After training on large-scale datasets, text-to-image (T2I) diffusion models [6, 25] exhibit exceptional image synthesis capability. For video generation, some works [2, 11] adapt image diffusion models by inflating the denoising UNet [26] with temporal layers to handle video inputs. However, replicating the success of T2I models for video generation remains challenging due to the lack of large-scale video datasets and the increased data complexity. Most recently, Sora [4] and following studies [16, 35, 37] pioneer high-quality, high-resolution text-to-video (T2V) diffusion models. These models commonly use the diffusion transformer (DiT) architecture [21] to capture complex spatial and temporal coherence, outperforming previous UNet-based methods. However, this improved performance comes at the cost of high computational demands, mostly due to the extensive attention mechanism, which requires processing a large number of tokens. Limited research has explored the pattern within these models to reduce the computational intensity. To address this, we delve into their internal structure to analyze the roles of individual blocks, leveraging these insights to facilitate their applicability and efficiency in video editing.", + "bbox": [ + 91, + 460, + 482, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Editing. Significant progress in image editing [3, 8, 17, 18, 20, 31] has driven advancements in video editing [7, 15, 19, 23, 34]. However, unique challenges in video editing arise from the need for consistent, coherent", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": " edits across frames, including both appearance and motion changes. Recent works [15, 23] adopt the attention feature injection techniques, manipulating various attention results to guide edits. They also employ model inflation to tune a T2I model with the input video following [34]. These approaches, however, are inefficient since they require a separate model for each sample and lack the general video priors to handle motion edits. In contrast, [19] utilizes a pre-trained video diffusion model for editing, beginning with corrupting the video by downsampling and noise. However, this corruption can result in global inconsistency with the source video and limits control over specific edits. Overall, there remains significant potential for advancing the use of video diffusion models in video editing. This application of the video diffusion model is nontrivial because (1) it requires maintaining global consistency in the generated video without relying on one-shot tuning or null-text optimization [18], and (2) applying previous attention feature injection methods is difficult with modern video models, as they involve dozens of transformer blocks with memory-intensive, full-attention computations. Our frequency-aware approach effectively addresses these challenges, setting our method apart from previous work.", + "bbox": [ + 511, + 381, + 906, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 750, + 604, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present FADE, a training-free framework that achieves real-world video editing with frequency-aware factorization of diffusion models. Our key idea is to investigate the inherent patterns in a pre-trained text-to-video diffusion model and activate each block's specific function for various video editing tasks. We will start by reviewing the background of video diffusion models, and then describe our designs of FADE, including how to factorize video dif", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28428", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fusion models according to frequency cues and adjust the sampling trajectory via spectrum-guided modulation. The overall framework of our FADE is illustrated in Figure 3.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminaries: Video Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 145, + 421, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video diffusion models are a class of generative models that reconstruct the distribution of video data by modeling the reverse trajectory of a diffusion process. Given a random noise $\\epsilon \\sim \\mathcal{N}(0, I)$ , the diffusion process is defined as:", + "bbox": [ + 89, + 167, + 482, + 229 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t} = \\alpha_ {t} \\boldsymbol {z} _ {0} + \\sigma_ {t} \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 239, + 482, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_0$ and $\\mathbf{z}_t$ are the clean and noisy video, and $\\left\\{ (\\alpha_t, \\sigma_t) \\right\\}_{t=1}^T$ denotes the noise schedule. With proper reparameterization, the training objective of diffusion models can be expressed as:", + "bbox": [ + 89, + 263, + 483, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D M}} = \\mathbb {E} _ {\\boldsymbol {z} _ {0}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 333, + 482, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{t}$ is computed as Equation (1). The denoising model $\\epsilon_{\\theta}$ is learned to predict the noise given the prompt $y$ . Previous T2I models often implement $\\epsilon_{\\theta}$ as UNet [26]. For video generation, recent models [4, 16, 35, 37] usually employ a series of cascaded transformer blocks to capture the spatial-temporal relationships more effectively.", + "bbox": [ + 89, + 359, + 483, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate a video in a small number of denoising steps, diffusion models usually use deterministic DDIM sampling:", + "bbox": [ + 89, + 450, + 483, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t - 1} = \\frac {\\alpha_ {t - 1}}{\\alpha_ {t}} \\boldsymbol {z} _ {t} + \\left(\\sigma_ {t - 1} - \\frac {\\alpha_ {t - 1}}{\\alpha_ {t}} \\sigma_ {t}\\right) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 501, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conversely, to convert a video to its corresponding noise, we can apply DDIM inversion, the reverse of Equation (3):", + "bbox": [ + 89, + 537, + 482, + 569 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t + 1} = \\frac {\\alpha_ {t + 1}}{\\alpha_ {t}} \\boldsymbol {z} _ {t} + \\left(\\sigma_ {t + 1} - \\frac {\\alpha_ {t + 1}}{\\alpha_ {t}} \\sigma_ {t}\\right) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 575, + 482, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, DDIM inversion introduces slight errors in every step, making it unreliable for fully reconstructing a real-world video or preserving fine details in the source video.", + "bbox": [ + 89, + 612, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Frequency-Aware Factorization of T2V Model", + "text_level": 1, + "bbox": [ + 89, + 665, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modern T2V models typically employ cascades of transformer blocks as the denoising model, with each block containing attention layers that process extensive video tokens. Given that dozens of identical blocks are involved in calculating the same form of attention, a key question arises: Do all these blocks serve the same purpose and contribute equally to video synthesis? To explore this, we analyze the behavior within each block of a pre-trained T2V model during video generation. Previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps. Inspired by this, we adopt a frequency-aware perspective to uncover if similar trends occur at the block level in video models. Specifically, we capture and visualize each block's attention maps", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{M_i\\}_{i=1}^N$ and outputs $\\{F_i\\}_{i=1}^N$ , along with their Fourier spectra, over all $N$ blocks during sampling. As illustrated in Figure 2, a typical attention map in T2V models contains three types of attention regions: cross-attention for multimodal alignments, spatial attention for capturing intra-frame spatial structures, and temporal attention for tracking dynamics across frames. Our observation is that spatial and temporal attention are densely aligned along diagonal lines in the early blocks, creating multiple parallel diagonals in the attention maps. The main diagonal line suggests that these early blocks focus on outlining the intra-frame spatial structures, defining key shapes and boundaries of the composition. The secondary diagonal lines indicate that these blocks are simultaneously establishing temporal correspondences across frames within the same regions. Collectively, these early blocks construct the low-frequency, foundational structure of the video, setting up basic object placement and movement. We refer to them as sketching blocks. In contrast, the following blocks—which constitute the majority of the denoising model—display relatively sparse attention distribution, with scores more evenly spread across elements in the map. This implies that these later blocks focus on refining the high-frequency refinements, such as color, texture, and subtle vibration, providing nuanced adjustments for visual details. We call them sharpening blocks, as their primary role is to enhance details, enriching the visual appeal of the final output.", + "bbox": [ + 511, + 90, + 906, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The variation in attention outputs of these blocks further supports our inference. The initial sketching blocks produce blurry outputs, with spectra focused on low frequencies that capture basic spatial and temporal layouts. In contrast, the sharpening blocks reveal clear foreground and background regions, involving much more high-frequency details. This distinction suggests that these two classes of blocks process features of differing levels and serve different purposes. We further verify that each block's function remains consistent throughout the sampling steps by analyzing these patterns across all steps. Thus, for video content editing—where maintaining underlying structures while adjusting finer details is essential—we can factorize each block's role, using sketching blocks for structural reconstruction and sharpening blocks for detailed generation. This division offers two key advantages: (1) it enables each block to perform optimally within its specialized function, and (2) it greatly reduces computational load by enabling frequency-aware guidance through selective use of blocks during sampling.", + "bbox": [ + 511, + 500, + 908, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Spectrum-Guided Modulation", + "text_level": 1, + "bbox": [ + 511, + 800, + 782, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After factorizing the video model based on frequency-informed attention patterns, our next step is to activate each block's specific function and develop a robust sampling strategy that optimally balances reconstruction with generation. We begin by employing the inversion process de", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28429", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scribed in Equation (4) with a pre-trained video diffusion model $\\epsilon_{\\theta}$ . This transforms the real-world video $z_0^*$ to an approximated noise $z_T^*$ and produces a reverse trajectory $\\{z_t^*\\}_{t=0}^T$ . However, as discussed in prior works, DDIM inversion can introduce non-negligible errors, making it difficult to accurately reconstruct the source video from $z_T^*$ . This issue is particularly pronounced in video inversion, where the predicted noise $z_T^*$ often deviates far from the standard Gaussian distribution. We, therefore, focus on refining the sampling process using the observed spectrum patterns to correct the inversion error and create an ideal trajectory $\\{z_t\\}_{t=0}^T$ to a plausible edited video $z_0$ . Following previous image editing approaches [18, 30], we adopt a dual branch strategy that leverages the inversion trajectory to enhance consistency with the source video. However, existing attention feature injection methods require extensive memory due to full-attention maps and limit editing flexibility through restrictive attention fusion techniques. While these methods can preserve the spatial layout, they often struggle with more complex edits involving pose, shape and motion.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose a more flexible, spectrum-guided approach that factorizes visual information in the frequency domain to separate content for preservation and editing. This approach offers greater freedom for a wider range of edits. As shown in Figure 3, our first design is to harness the sketching blocks, which care about the overall spatial layout and temporal movement. Rather than swapping or interleaving the attention maps directly, we gather the attention outputs of the source video $\\boldsymbol{z}_t^*$ and the target video $\\boldsymbol{z}_t$ within these blocks, denoted as $\\boldsymbol{F}_t^*$ and $\\boldsymbol{F}_t$ , and use their difference to derive a guidance term for the sampling process. Specifically, these attention outputs are computed using the same source prompt $\\boldsymbol{y}_{src}$ through the full-attention layers. This guidance term provides a more robust and adaptive way to adjust the sampling trajectory without directly altering the intermediate features within the diffusion model. Furthermore, since these features are derived from the sketching blocks, the guidance focuses on the coarse spatial and temporal content, such as basic layouts and object movements, leaving ample room for detailed edits to be made in the sharpening blocks.", + "bbox": [ + 89, + 397, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, native guidance can lead to information leakage from the source video to the target video, as the attention outputs contain the full range of visual content. To address this, it is crucial to determine which aspects of the content should be retained and which should be modified. For this purpose, we turn to the frequency domain for assistance. Real-world images and videos often exhibit redundancy across their spatial and temporal dimensions, meaning that most of their energy is concentrated in the low-frequency spectrum, while the high-frequency components contribute much less. This characteristic underpins popular compression algorithms such as JPEG for images and", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MPEG for videos, and we aim to apply this principle here. Since we cannot get a clean video during the denoising process, we use attention outputs composed of tokens representing the noisy video. Unlike the discrete cosine transform (DCT) used in the spatial domain for MPEG, we employ the discrete Fourier transform on both spatial and temporal dimensions (3D DFT) to capture the spectral characteristics of intra-frame content and inter-frame motion:", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {t} = \\mathrm {D F T} _ {\\mathrm {3 D}} \\left(\\boldsymbol {F} _ {t}\\right) \\in \\mathbb {R} ^ {h \\times w \\times \\tau \\times c}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 223, + 903, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h$ , $w$ and $\\tau$ are the spatial and temporal size of the video tokens, and $c$ is the channel dimension. As we know, most of the energy in $\\mathcal{F}_t$ is concentrated in the low-frequency range. Thus, with the observed attention patterns, we can assume that fundamental spatial structures and temporal movements are encapsulated within this range, while the high-frequency components contain Gaussian noise and specific details like texture and subtle shaking—elements we wish to suppress or remove during editing. To achieve this, we apply a low-pass filter to isolate the low-frequency content of $\\mathcal{F}_t$ . The same process is applied to $\\mathcal{F}_t^*$ to compute the spectrum guidance $\\mathcal{G}_t$ :", + "bbox": [ + 511, + 253, + 906, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {t} = \\left\\| \\mathrm {L P} \\left(\\mathcal {F} _ {t}\\right) - \\mathrm {L P} \\left(\\mathcal {F} _ {t} ^ {*}\\right) \\right\\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 446, + 903, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where LP denotes the low-pass filter. This guidance measures the deviation of the target video $\\mathbf{z}_t$ from the source video in terms of basic structure and movement. This distance is then used to modulate the trajectory during DDIM sampling by the gradient of $\\mathcal{G}_t$ with respect to $\\mathbf{z}_t$ :", + "bbox": [ + 511, + 476, + 905, + 553 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t - 1} = \\operatorname {D D I M} \\left(\\boldsymbol {\\epsilon} _ {\\theta}, \\boldsymbol {z} _ {t}, t, \\boldsymbol {y} _ {t g t}\\right) - \\lambda \\operatorname {N o r m} \\left(\\nabla_ {\\boldsymbol {z} _ {t}} \\mathcal {G} _ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 565, + 903, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where DDIM denotes the sampling process in Equation (3), $\\lambda$ is the guidance weight and Norm normalizes the gradient to avoid numerical instability. This modulation strategy iteratively denoises the edited video towards the ideal direction, ensuring high consistency and realistic edits.", + "bbox": [ + 511, + 594, + 905, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Versatile Video Editing Framework", + "text_level": 1, + "bbox": [ + 511, + 680, + 821, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize a pre-trained T2V diffusion model for video inversion and sampling. This model leverages a 3D VAE, comprising an encoder $\\mathcal{E}$ and a decoder $\\mathcal{D}$ , to enable spatial and temporal compression, facilitating a reversible transformation between pixel space and latent space. In addition to text embeddings, the T2V model incorporates FPS information to better handle video data. Unlike one-shot tuned video models from T2I models [15], using a well-trained video diffusion model ensures a more stable understanding of the complex spatial and temporal relationships within videos. Moreover, this eliminates the need for model tuning, enhancing efficiency and reducing the necessity of storing individual models for specific videos. For edit types,", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28430", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4dbc2aeb531ffdc4ab3147667131a9d851c7f0a8b51f003cc40a8f4df8c69548.jpg", + "image_caption": [ + "Input video", + "Target: bushes $\\rightarrow$ snowfield" + ], + "image_footnote": [], + "bbox": [ + 122, + 92, + 488, + 164 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da7e528571eca048c37c8a75cf3a829de56378f63f479acabb3a0590ba816ba3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 181, + 488, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0d33bd845d0a8f78efc3a9699ec8eec5803970f6dd3a8b97e61522c8d924e78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 255, + 488, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40f2debd73c2cf9fc0e95ca8c21954e428b39df3dd23d21524d191285ffe679d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 327, + 488, + 400 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc2b030c8f9d36c14d41f7958f41bf6d61e56a538f0b50a7c98999aef5ee73dc.jpg", + "image_caption": [ + "Target: bear $\\rightarrow$ panda" + ], + "image_footnote": [], + "bbox": [ + 96, + 410, + 488, + 482 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e4dc63f2b95716796af97e70af5865969eeaf017b793bc4ea717bf2e46f2e5f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 500, + 488, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35f4bad7762172f262605bbbfaabb3f2bc1d0747169f811ae1f78c4b5a67ff51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 574, + 488, + 645 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c662cb31fc0838798ed20c80a3dda2158dc82962562750868f8d593edc26a7b.jpg", + "image_caption": [ + "(a) Appearance Edits", + "Figure 4. Qualitative comparisons on real-world videos. We evaluate both (a) appearance and (b) motion edits. Our framework consistently achieves accurate textual alignment across frames and high fidelity in unedited regions. Notably, compared to other methods based on video models, we produce more coherent and realistic motion edits. More results are available in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 96, + 646, + 488, + 717 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/806885260d8a4837dfc356114ed573ae44499672db366f0a1124559f466222ea.jpg", + "image_caption": [ + "Input video", + "Target: turn the head $\\rightarrow$ lower the head" + ], + "image_footnote": [], + "bbox": [ + 535, + 92, + 900, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d126405e4c73934466888b3ea1d79e04d457f1d35b5166970973c4228292122.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 181, + 900, + 252 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7947c3bcf39e7289c959346cd2d07e878b01a2db0ba9b76ef9f5b494b3b5291a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 255, + 900, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/192b1ff75e1e116beaed25a639b512cf8309f2916b51bdac0e4694299d7b9502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 327, + 900, + 400 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a276093bcbd25dd356bf4c6429d7ac1dfc2170be2c5b9b8062e5a46bb37a72f1.jpg", + "image_caption": [ + "Target: close the eyes $\\rightarrow$ open the eyes" + ], + "image_footnote": [], + "bbox": [ + 508, + 411, + 900, + 482 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1516098ad6e4c17b6c078b092c14a992770223264aa780b0030ddaea145934c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 500, + 900, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20abeeaeeacdfb9483f944557878d4ea6ed9bdf4ddbbee81887e1b553c0cc28d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 571, + 900, + 645 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2252fc07fa7183872b460c2288fae4591475ed5bfe3fbd5f04995a9ff7c348ab.jpg", + "image_caption": [ + "(b) Motion Edits" + ], + "image_footnote": [], + "bbox": [ + 508, + 646, + 900, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we broadly categorize prompt-driven modifications into appearance edits and motion edits. Appearance edits involve altering nouns, such as changing the main subject or background, focusing on the texture, color, and style (e.g., transforming a \"cat\" to a \"tiger\" in Figure 3). Motion edits, on", + "bbox": [ + 89, + 816, + 480, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the other hand, involve changing verbs, which can alter spatial and temporal dynamics. By applying spectrum-guided modulation to harness inherent video priors within the T2V model, our method achieves realistic motion edits, distinguishing it from most previous approaches.", + "bbox": [ + 511, + 816, + 906, + 892 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28431", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24da7d9a88027b5a22f02ae61f6cbf9f976ed24fe4a885e12388a71177e9a8fe.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on real-world video data. We compare our FADE framework to previous methods on both appearance and motion edits, measuring editing accuracy, input fidelity, temporal consistency, and human preference. FADE achieves competitive performance across diverse editing tasks." + ], + "table_footnote": [], + "table_body": "
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓PF ↑
appearance edits
Tune-A-Video [34]0.352219.860.462535.010.12
Video-P2P [15]0.358920.570.319934.710.15
FateZero [23]0.356220.650.305733.230.29
CogVideoX-V2V [35]0.375418.960.481131.450.09
FADE (Ours)0.376220.690.308531.360.35
motion edits
Tune-A-Video [34]0.328118.680.463735.850.10
Video-P2P [15]0.331418.930.388535.440.15
FateZero [23]0.325919.020.371234.470.13
CogVideoX-V2V [35]0.367818.170.492835.520.19
FADE (Ours)0.368319.260.369232.280.43
", + "bbox": [ + 91, + 169, + 480, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 359, + 225, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of our method, we conduct comprehensive experiments across various video editing tasks. We will start by describing the experimental settings and then present our main results. We will also provide detailed ablation studies and analyses of our approach to highlight the contributions of each component in our framework and to validate the impact of our design choices.", + "bbox": [ + 89, + 383, + 483, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Experiment Setups", + "text_level": 1, + "bbox": [ + 89, + 500, + 274, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. To utilize video generation priors, we adopt a recent, widely used T2V model [35] for video inversion and generation. This model's denoising component employs the DiT architecture with 48 transformer blocks, each containing a full-attention layer. Based on observed attention map patterns, we designate the first 4 blocks as sketching blocks. We calculate the guidance over the interval $[0, 0.6T]$ of the DDIM sampling steps, with total steps set to $T = 50$ . To generate textual prompts from the source video, we use a multimodal language model [14] as the captioner. The guidance weight varies between 10 and 15, tailored to each editing task. We perform all evaluations on a single NVIDIA LS20 GPU with 48GB VRAM.", + "bbox": [ + 89, + 523, + 483, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Metrics. We evaluate our method using the DAVIS dataset [22] as well as various in-the-wild videos, following the approach of previous studies [15, 23]. Alongside qualitative results, we provide quantitative evaluations across multiple aspects using different metrics. Editing accuracy is assessed through CLIP score [24], which reflects the degree of textual alignment. To evaluate temporal consistency, a key factor for video coherence, we compute object semantic variance (OSV), following [15], to capture semantic consistency across frames. Additionally, we assess fidelity in unedited regions by calculating Mask PSNR and LPIPS [36], which quantify the preservation of content that", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "should remain unchanged. Furthermore, given the importance of visual quality in video editing, we conduct a user study to gauge human preference across all editing tasks.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 147, + 653, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our editing method with state-of-the-art approaches, including Video-P2P [15], FateZero [23], TuneA-Video (TAV) [34] and CogVideoX-V2V (CogV2V) [35]. As illustrated in Figure 4, our method consistently yields superior results across all aspects when compared to previous methods. The editing results of Video-P2P and FateZero exhibit significant limitations, often failing to modify content accurately, resulting in noticeable textual mismatches. Similarly, while TAV and CogV2V leverage T2V models, they struggle with motion editing, producing subpar results in capturing dynamic changes. Leveraging the inherent priors from the T2V model and spectrum-guided modulation, our method achieves the best performance in preserving local structure, maintaining temporal consistency, and ensuring text-video alignment. We evaluate our method on 20 videos and measure 5 metrics for quantitative analysis. As shown in Table 1, our method achieves the highest CLIP and OSV scores, demonstrating its superiority in textual alignment and temporal consistency. We also obtain competitive Mask PSNR and LPIPS, indicating that our method preserves unedited regions effectively. Moreover, we achieve the highest human preference score, reflecting the favorable visual quality and appeal of our edits across tasks. Notably, our method operates without any optimization or fine-tuning, completing various edits on real-world videos in approximately 3 minutes, a significant efficiency gain compared to previous methods, which require over 15 minutes.", + "bbox": [ + 511, + 171, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 511, + 589, + 616, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficient Editing via Sketching Blocks. We assess the function of different blocks in the T2V model by testing various block combinations to calculate modulation, holding other settings constant. We consider three configurations: (1) using sketching blocks only (ours), (2) symmetrically combining sketching and sharpening blocks (symm. blocks), and (3) using all blocks without factorization (w/o factor.). As shown in Table 2, configuration (1) provides comparable results to configurations (2) and (3), despite requiring fewer blocks for modulation. Although configurations (2) and (3) offer slightly improved fidelity, they significantly increase test time and memory demands. Figure 5 further demonstrates that relying on only the first four transformer blocks, or sketching blocks, achieves the best editing performance. Interestingly, contrary to common assumptions, including the sharpening blocks actually diminishes editing quality. We assume this is because the sharpening blocks primarily focus on elements like color, texture, and subtle vibrations, which often do not need", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28432", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b17bef75167506b7f24da7971274f502a76e0da9e419430faa7998ee33e0936.jpg", + "image_caption": [ + "Input video", + "Target: duck $\\rightarrow$ yellow chick + open eyes", + "Figure 5. Qualitative comparisons of the ablations. We find the variant frameworks fall short in terms of editing accuracy, temporal consistency and input fidelity, yielding unsatisfactory results." + ], + "image_footnote": [], + "bbox": [ + 98, + 106, + 187, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ab7d17fbdb49e07c3edf267bff92809baec101f02192a9396581245a3088dfd.jpg", + "image_caption": [ + "w/o factor." + ], + "image_footnote": [], + "bbox": [ + 192, + 106, + 284, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b6d16e720d60da30d7ad732e5f4b20a1050b3307a311373b6bf9c602920cb49.jpg", + "image_caption": [ + "w/o filter" + ], + "image_footnote": [], + "bbox": [ + 289, + 106, + 380, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/615cb00bb2207ffa78aedbfa41e1e47aa973edd39f2ac863f82db0d602280004.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 383, + 106, + 475, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to remain unchanged in most video editing tasks. Relying on these blocks for guidance may, therefore, misdirect the model, diverting its attention from building the foundational layout and movement of the input video, ultimately resulting in reduced performance. These results underscore the effectiveness of our frequency-aware factorization approach, demonstrating that our selective focus on sketching blocks provides an optimal balance between maintaining high-quality edits and achieving computational efficiency.", + "bbox": [ + 88, + 460, + 482, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Spectrum-Guided Modulation. In Figure 5 and Table 2, we examine our spectrum-guided modulation by removing the low-pass filter (w/o filter) and analyzing its effect on the editing performance. When we use the entire spectrum of attention outputs, high-frequency components are retained in the guidance term, causing the target object to retain too much of the source's original characteristics, which weakens the text-video alignment. Experiments show that using about two-thirds of the frequency components in the guidance term balances reconstruction and editing quality. The quantitative results further underscore the efficacy of spectrum-guided modulation, as shown by superior performance on key metrics like textual alignment and temporal consistency. By selectively modulating the frequency components, our method achieves precise edits that align well with the given textual prompt and maintain consistency across frames, both of which are crucial for high-quality video editing. These findings highlight that spectrum-guided modulation not only supports a stable structural foundation for the edits but also enhances flexibil", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6e29b377fd380ca94abf2795bfd230389479d3bcb557db53b11f011a6c22b3ea.jpg", + "table_caption": [ + "Table 2. Ablations studies. We perform ablations to verify the effectiveness of the components in FADE and the impact of the model factorization. We find that FADE achieves similar performance with only a subset of sketching blocks, compared to using symmetric or all blocks. Moreover, spectrum-guided modulation further enhances both preservation and precision in edits." + ], + "table_footnote": [], + "table_body": "
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Test-time
symm. blocks0.365920.730.336732.615 min
w/o factor.0.369120.940.332832.0512 min
w/o filter0.361220.890.336432.283 min
FADE (Ours)0.372820.870.335231.773 min
", + "bbox": [ + 514, + 183, + 901, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ity, allowing detailed adjustments for various editing tasks while ensuring input fidelity across the video.", + "bbox": [ + 511, + 273, + 903, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effective Exploitation of Video Diffusion Model. Our framework capitalizes on the rich video priors embedded in the pre-trained T2V model, enabling more effective utilization of the model's inherent knowledge of spatial layouts, temporal coherence, and object motions. The video diffusion model captures nuanced understanding across frames, making it an ideal foundation for edits that require both high fidelity and temporal consistency. To validate our superior exploitation of these priors, we compare FADE with CogVideoX-V2V [35], which uses the same underlying T2V model. As shown in Figure 4 and Table 1, FADE outperforms it both qualitatively and quantitatively across various editing tasks, demonstrating our framework's ability to harness the model's rich knowledge more effectively. Limitations. Despite FADE's effectiveness, its performance relies on the underlying video model and struggles with heavy occlusions requiring advanced temporal reasoning. We will address these in future work.", + "bbox": [ + 511, + 303, + 905, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present FADE, a novel, training-free framework tailored for a diverse range of video editing tasks, encompassing both appearance and motion adjustments. Through frequency-aware factorization of the T2V model, FADE efficiently harnesses the model's inherent structural and temporal patterns, allowing each block to excel in its designated role. We further design the spectrum-guided modulation, which adjusts the sampling trajectory using frequency cues to preserve foundational video elements and enhance text-video alignment. Extensive experiments demonstrate our framework's competitive performance and generalization across various video editing challenges. We envision our work will inspire future research on versatile video editing and efficient video diffusion model. Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2024YFB4708100, and in part by the National Natural Science Foundation of China under Grant Grant 62321005, Grant 62306031", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28433", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 2", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3, 4", + "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3", + "[7] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3", + "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3, 4", + "[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 3", + "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2", + "[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, pages 6007-6017, 2023. 2", + "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 7", + "[15] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In CVPR, pages 8599–8608, 2024. 2, 3, 5, 7" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 4", + "[17] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3", + "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2, 3, 5, 1", + "[19] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3", + "[20] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, pages 1-11, 2023. 2, 3", + "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3", + "[22] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7", + "[23] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3, 7", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 7", + "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3", + "[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3, 4", + "[27] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. In ACML, pages 1215-1230. PMLR, 2024. 2", + "[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2", + "[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3", + "[30] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-rescale: Self-guidance mechanism for effective tuning-free real image editing. arXiv preprint arXiv:2409.01322, 2024. 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28434", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 3", + "[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2", + "[33] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2", + "[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 7", + "[35] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3, 4, 7, 8, 1", + "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7", + "[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. URL https://github.com/hpcaitech/Open-Sora, 2024.3,4", + "[38] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1" + ], + "bbox": [ + 91, + 92, + 480, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28435", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_model.json b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d3671860dafaeb059f3441b3ea4ec43d3cd109b --- /dev/null +++ b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "FADE: Frequency-Aware Diffusion Model Factorization for Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.928, + 0.201 + ], + "angle": 0, + "content": "Yixuan Zhu\\(^{1}\\) Haolin Wang\\(^{1}\\) Shilin Ma\\(^{2}\\) Wenliang Zhao\\(^{1}\\) Yansong Tang\\(^{2}\\) Lei Chen\\(^{1,\\dagger}\\) Jie Zhou\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.216, + 0.802, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Automation, Tsinghua University \\(^{2}\\)Tsinghua Shenzhen International Graduate School, Tsinghua University" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.27, + 0.195, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.27, + 0.296, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.27, + 0.397, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.27, + 0.497, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.27, + 0.598, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.27, + 0.699, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.27, + 0.8, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.27, + 0.901, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.349, + 0.722, + 0.363 + ], + "angle": 0, + "content": "Source Prompt: A black swan swimming in a river beside bushes" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.365, + 0.195, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.365, + 0.296, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.365, + 0.397, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.365, + 0.497, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.365, + 0.598, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.365, + 0.699, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.365, + 0.8, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.365, + 0.901, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.444, + 0.65, + 0.457 + ], + "angle": 0, + "content": "Appearance Edit: black swan \\(\\rightarrow\\) white duck" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.459, + 0.195, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.459, + 0.296, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.459, + 0.397, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.459, + 0.497, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.459, + 0.598, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.459, + 0.699, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.459, + 0.8, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.459, + 0.901, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.537, + 0.655, + 0.553 + ], + "angle": 0, + "content": "Motion Edit: swimming \\(\\rightarrow\\) flapping the wings" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.564, + 0.908, + 0.607 + ], + "angle": 0, + "content": "Figure 1. Diverse video editing results of FADE. Our training-free approach, utilizing frequency-aware factorization and modulation, achieves high-fidelity, coherent edits across a variety of video types. FADE handles both appearance and motion adjustments with impressive robustness, ensuring precise alignment with input prompts and maintaining temporal consistency. Best viewed in color." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.612, + 0.327, + 0.629 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.645, + 0.486, + 0.873 + ], + "angle": 0, + "content": "Recent advancements in diffusion frameworks have significantly enhanced video editing, achieving high fidelity and strong alignment with textual prompts. However, conventional approaches using image diffusion models fall short in handling video dynamics, particularly for challenging temporal edits like motion adjustments. While current video diffusion models produce high-quality results, adapting them for efficient editing remains difficult due to the heavy computational demands that prevent the direct application of previous image editing techniques. To overcome these limitations, we introduce FADE—a training-free yet highly effective video editing approach that fully leverages the inherent priors from pre-trained video diffusion models via frequency-aware factorization. Rather than simply using these models, we first analyze the attenuated performance of the proposed system." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.81 + ], + "angle": 0, + "content": "tion patterns within the video model to reveal how video priors are distributed across different components. Building on these insights, we propose a factorization strategy to optimize each component's specialized role. Furthermore, we devise spectrum-guided modulation to refine the sampling trajectory with frequency domain cues, preventing information leakage and supporting efficient, versatile edits while preserving the basic spatial and temporal structure. Extensive experiments on real-world videos demonstrate that our method consistently delivers high-quality, realistic and temporally coherent editing results both qualitatively and quantitatively. Code is available at https://github.com/EternalEvan/FADE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.844, + 0.646, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The objective of video editing is to modify specific visual content consistently across all frames while preserving the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.245, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.482, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.432 + ], + "angle": 0, + "content": "Figure 2. (a) In typical T2V models, video tokens flattened along the temporal axis yield full-attention maps split into cross, spatial, and temporal attention. (b) Video priors in T2V models are factorized based on attention feature patterns. We visualize the attention map \\( M \\) and result \\( F \\) along with its spectrum for two blocks. The early block sketches the foundational structure and movement, while the later block refines details. This insight allows us to leverage each block's unique role for more efficient editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.734 + ], + "angle": 0, + "content": "integrity of unchanged regions. Recent achievements in diffusion models have greatly elevated video editing capabilities [10, 12, 15, 19, 20, 28, 32], bringing remarkable generative power to video editing tasks that require both spatial details and temporal coherence. While previous methods [1, 8, 13, 18, 30] have adapted text-to-image (T2I) diffusion models for prompt-driven image editing, their effectiveness diminishes when applied to video due to the complexities of managing spatial-temporal relationships across frames and a lack of generalized video prior knowledge. These limitations underscore the need for an efficient and robust editing strategy capable of handling the increased data and temporal coherence demands of videos. In response, we aim to design an accessible, high-quality editing solution that fully leverages the rich video priors embedded in pre-trained video diffusion models, enabling realistic, coherent, and versatile edits, including both appearance and motion adjustments, across complex video content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The main challenge in video editing is balancing changes to the target object while preserving unedited content. Previous methods largely rely on the generative power of T2I models for controlled edits. For instance, [15, 33] achieve this through null-text inversion [18], which minimizes inversion errors via test-time optimization. However, this process requires extensive iterations to compute a series of null-text embeddings for every sampling step, making it time-consuming and limiting its real-world applicability. Another line of work is attention feature injection [5, 15, 23, 27], where attention maps are modified by" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.289 + ], + "angle": 0, + "content": "replacing or interleaving to maintain content consistency. Despite this, the absence of video priors often results in unrealistic temporal dynamics and limits motion edit capacity. Some methods, like [19], directly utilize video diffusion models by adding noise to the source video. This straightforward approach fails to fully exploit video priors, resulting in suboptimal editing quality. With the recent emergence of advanced text-to-video (T2V) models, new opportunities arise to harness their rich video knowledge. However, adapting previous T2I techniques to these models is challenging, given the extensive resources required, including memory, computation and time, to handle the increased data and complexity inherent in large-scale video networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.716 + ], + "angle": 0, + "content": "To overcome these challenges, we introduce FADE, a training-free yet potent approach for real-world video editing. Our goal is to harness the rich video priors about video appearance and dynamics within pre-trained T2V models to enable realistic and diverse editing. To achieve this, we start by examining the specific properties of these models. While previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps, we further observe that a similar frequency progression occurs block-wise within T2V models. We explore how video priors function in each block by analyzing their roles and patterns, revealing that early blocks sketch foundational spatial layouts and temporal dynamics while later blocks refine high-frequency details (Figure 2). Leveraging this insight, we factorize the editing problem by the distinct functions of each component, enhancing editing efficiency. Then, instead of performing optimization or attention injection, we design a spectrum-guided modulation strategy that utilizes the full-attention outputs in these sketching blocks to derive frequency-aware guidance through 3D Fourier transformation. This modulation adjusts the sampling trajectory to preserve low-frequency structure from the source video while allowing high-frequency details to support flexible and intricate edits. This cue from the frequency domain not only mitigates information leakage during guidance computation but also unleashes the potential of video diffusion priors, enabling sophisticated spatial and temporal edits with high fidelity and coherence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We evaluate our method using extensive real-world video clips across various editing tasks. As shown in Figure 1, FADE exhibits proficiency in producing realistic and consistent edits, covering both appearance and motion modifications. Quantitative comparisons on video editing benchmark DAVIS [22] demonstrate our approach achieves superior textual alignment and temporal consistency, with 0.3561 CLIP score and 43.57 OSV. Moreover, we maintain competitive input fidelity with 21.54 MaskPSNR and 0.3297 LPIPS, preserving unedited regions effectively. Compared to previous methods, our use of general video priors with frequency-aware factorization deliv" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28427" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.898, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Figure 3. The overall framework of FADE. FADE is a training-free framework for video editing. Given an input video, we first perform video inversion to obtain noise \\( \\mathbf{z}_T \\) and trajectory \\( \\{\\mathbf{z}_t^*\\}_{t=0}^T \\). Starting from this noise, we use sketching blocks in the T2V model to produce full-attention results \\( \\mathbf{F}_t \\) and \\( \\mathbf{F}_t^* \\) from \\( \\mathbf{z}_t \\) and \\( \\mathbf{z}_t^* \\). These results are transformed to the frequency domain, where a low-pass filter isolates spatial and temporal structures. The frequency error then modulates the sampling process, enabling high-quality, consistent edits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.413 + ], + "angle": 0, + "content": "ers improved efficiency and a broader editing range, highlighting FADE's outstanding versatility and practicality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.242, + 0.449 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Video Diffusion Models. In recent years, denoising diffusion models [9, 29] have emerged as a leading class of generative models, notable for their synthesis quality and controllability. After training on large-scale datasets, text-to-image (T2I) diffusion models [6, 25] exhibit exceptional image synthesis capability. For video generation, some works [2, 11] adapt image diffusion models by inflating the denoising UNet [26] with temporal layers to handle video inputs. However, replicating the success of T2I models for video generation remains challenging due to the lack of large-scale video datasets and the increased data complexity. Most recently, Sora [4] and following studies [16, 35, 37] pioneer high-quality, high-resolution text-to-video (T2V) diffusion models. These models commonly use the diffusion transformer (DiT) architecture [21] to capture complex spatial and temporal coherence, outperforming previous UNet-based methods. However, this improved performance comes at the cost of high computational demands, mostly due to the extensive attention mechanism, which requires processing a large number of tokens. Limited research has explored the pattern within these models to reduce the computational intensity. To address this, we delve into their internal structure to analyze the roles of individual blocks, leveraging these insights to facilitate their applicability and efficiency in video editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Video Editing. Significant progress in image editing [3, 8, 17, 18, 20, 31] has driven advancements in video editing [7, 15, 19, 23, 34]. However, unique challenges in video editing arise from the need for consistent, coherent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.908, + 0.73 + ], + "angle": 0, + "content": " edits across frames, including both appearance and motion changes. Recent works [15, 23] adopt the attention feature injection techniques, manipulating various attention results to guide edits. They also employ model inflation to tune a T2I model with the input video following [34]. These approaches, however, are inefficient since they require a separate model for each sample and lack the general video priors to handle motion edits. In contrast, [19] utilizes a pre-trained video diffusion model for editing, beginning with corrupting the video by downsampling and noise. However, this corruption can result in global inconsistency with the source video and limits control over specific edits. Overall, there remains significant potential for advancing the use of video diffusion models in video editing. This application of the video diffusion model is nontrivial because (1) it requires maintaining global consistency in the generated video without relying on one-shot tuning or null-text optimization [18], and (2) applying previous attention feature injection methods is difficult with modern video models, as they involve dozens of transformer blocks with memory-intensive, full-attention computations. Our frequency-aware approach effectively addresses these challenges, setting our method apart from previous work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.751, + 0.605, + 0.767 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we present FADE, a training-free framework that achieves real-world video editing with frequency-aware factorization of diffusion models. Our key idea is to investigate the inherent patterns in a pre-trained text-to-video diffusion model and activate each block's specific function for various video editing tasks. We will start by reviewing the background of video diffusion models, and then describe our designs of FADE, including how to factorize video dif" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28428" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "fusion models according to frequency cues and adjust the sampling trajectory via spectrum-guided modulation. The overall framework of our FADE is illustrated in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.146, + 0.423, + 0.161 + ], + "angle": 0, + "content": "3.1. Preliminaries: Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.23 + ], + "angle": 0, + "content": "Video diffusion models are a class of generative models that reconstruct the distribution of video data by modeling the reverse trajectory of a diffusion process. Given a random noise \\(\\epsilon \\sim \\mathcal{N}(0, I)\\), the diffusion process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.24, + 0.483, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t} = \\alpha_ {t} \\boldsymbol {z} _ {0} + \\sigma_ {t} \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.484, + 0.326 + ], + "angle": 0, + "content": "where \\( \\mathbf{z}_0 \\) and \\( \\mathbf{z}_t \\) are the clean and noisy video, and \\( \\left\\{ (\\alpha_t, \\sigma_t) \\right\\}_{t=1}^T \\) denotes the noise schedule. With proper reparameterization, the training objective of diffusion models can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.334, + 0.483, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D M}} = \\mathbb {E} _ {\\boldsymbol {z} _ {0}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.451 + ], + "angle": 0, + "content": "where \\( z_{t} \\) is computed as Equation (1). The denoising model \\( \\epsilon_{\\theta} \\) is learned to predict the noise given the prompt \\( y \\). Previous T2I models often implement \\( \\epsilon_{\\theta} \\) as UNet [26]. For video generation, recent models [4, 16, 35, 37] usually employ a series of cascaded transformer blocks to capture the spatial-temporal relationships more effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.482 + ], + "angle": 0, + "content": "To generate a video in a small number of denoising steps, diffusion models usually use deterministic DDIM sampling:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.502, + 0.483, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t - 1} = \\frac {\\alpha_ {t - 1}}{\\alpha_ {t}} \\boldsymbol {z} _ {t} + \\left(\\sigma_ {t - 1} - \\frac {\\alpha_ {t - 1}}{\\alpha_ {t}} \\sigma_ {t}\\right) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.57 + ], + "angle": 0, + "content": "Conversely, to convert a video to its corresponding noise, we can apply DDIM inversion, the reverse of Equation (3):" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.577, + 0.483, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t + 1} = \\frac {\\alpha_ {t + 1}}{\\alpha_ {t}} \\boldsymbol {z} _ {t} + \\left(\\sigma_ {t + 1} - \\frac {\\alpha_ {t + 1}}{\\alpha_ {t}} \\sigma_ {t}\\right) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.658 + ], + "angle": 0, + "content": "In practice, DDIM inversion introduces slight errors in every step, making it unreliable for fully reconstructing a real-world video or preserving fine details in the source video." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.683 + ], + "angle": 0, + "content": "3.2. Frequency-Aware Factorization of T2V Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Modern T2V models typically employ cascades of transformer blocks as the denoising model, with each block containing attention layers that process extensive video tokens. Given that dozens of identical blocks are involved in calculating the same form of attention, a key question arises: Do all these blocks serve the same purpose and contribute equally to video synthesis? To explore this, we analyze the behavior within each block of a pre-trained T2V model during video generation. Previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps. Inspired by this, we adopt a frequency-aware perspective to uncover if similar trends occur at the block level in video models. Specifically, we capture and visualize each block's attention maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.5 + ], + "angle": 0, + "content": "\\(\\{M_i\\}_{i=1}^N\\) and outputs \\(\\{F_i\\}_{i=1}^N\\), along with their Fourier spectra, over all \\(N\\) blocks during sampling. As illustrated in Figure 2, a typical attention map in T2V models contains three types of attention regions: cross-attention for multimodal alignments, spatial attention for capturing intra-frame spatial structures, and temporal attention for tracking dynamics across frames. Our observation is that spatial and temporal attention are densely aligned along diagonal lines in the early blocks, creating multiple parallel diagonals in the attention maps. The main diagonal line suggests that these early blocks focus on outlining the intra-frame spatial structures, defining key shapes and boundaries of the composition. The secondary diagonal lines indicate that these blocks are simultaneously establishing temporal correspondences across frames within the same regions. Collectively, these early blocks construct the low-frequency, foundational structure of the video, setting up basic object placement and movement. We refer to them as sketching blocks. In contrast, the following blocks—which constitute the majority of the denoising model—display relatively sparse attention distribution, with scores more evenly spread across elements in the map. This implies that these later blocks focus on refining the high-frequency refinements, such as color, texture, and subtle vibration, providing nuanced adjustments for visual details. We call them sharpening blocks, as their primary role is to enhance details, enriching the visual appeal of the final output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.789 + ], + "angle": 0, + "content": "The variation in attention outputs of these blocks further supports our inference. The initial sketching blocks produce blurry outputs, with spectra focused on low frequencies that capture basic spatial and temporal layouts. In contrast, the sharpening blocks reveal clear foreground and background regions, involving much more high-frequency details. This distinction suggests that these two classes of blocks process features of differing levels and serve different purposes. We further verify that each block's function remains consistent throughout the sampling steps by analyzing these patterns across all steps. Thus, for video content editing—where maintaining underlying structures while adjusting finer details is essential—we can factorize each block's role, using sketching blocks for structural reconstruction and sharpening blocks for detailed generation. This division offers two key advantages: (1) it enables each block to perform optimally within its specialized function, and (2) it greatly reduces computational load by enabling frequency-aware guidance through selective use of blocks during sampling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.783, + 0.817 + ], + "angle": 0, + "content": "3.3. Spectrum-Guided Modulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "After factorizing the video model based on frequency-informed attention patterns, our next step is to activate each block's specific function and develop a robust sampling strategy that optimally balances reconstruction with generation. We begin by employing the inversion process de" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28429" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.395 + ], + "angle": 0, + "content": "scribed in Equation (4) with a pre-trained video diffusion model \\(\\epsilon_{\\theta}\\). This transforms the real-world video \\(z_0^*\\) to an approximated noise \\(z_T^*\\) and produces a reverse trajectory \\(\\{z_t^*\\}_{t=0}^T\\). However, as discussed in prior works, DDIM inversion can introduce non-negligible errors, making it difficult to accurately reconstruct the source video from \\(z_T^*\\). This issue is particularly pronounced in video inversion, where the predicted noise \\(z_T^*\\) often deviates far from the standard Gaussian distribution. We, therefore, focus on refining the sampling process using the observed spectrum patterns to correct the inversion error and create an ideal trajectory \\(\\{z_t\\}_{t=0}^T\\) to a plausible edited video \\(z_0\\). Following previous image editing approaches [18, 30], we adopt a dual branch strategy that leverages the inversion trajectory to enhance consistency with the source video. However, existing attention feature injection methods require extensive memory due to full-attention maps and limit editing flexibility through restrictive attention fusion techniques. While these methods can preserve the spatial layout, they often struggle with more complex edits involving pose, shape and motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.716 + ], + "angle": 0, + "content": "To overcome these limitations, we propose a more flexible, spectrum-guided approach that factorizes visual information in the frequency domain to separate content for preservation and editing. This approach offers greater freedom for a wider range of edits. As shown in Figure 3, our first design is to harness the sketching blocks, which care about the overall spatial layout and temporal movement. Rather than swapping or interleaving the attention maps directly, we gather the attention outputs of the source video \\( \\boldsymbol{z}_t^* \\) and the target video \\( \\boldsymbol{z}_t \\) within these blocks, denoted as \\( \\boldsymbol{F}_t^* \\) and \\( \\boldsymbol{F}_t \\), and use their difference to derive a guidance term for the sampling process. Specifically, these attention outputs are computed using the same source prompt \\( \\boldsymbol{y}_{src} \\) through the full-attention layers. This guidance term provides a more robust and adaptive way to adjust the sampling trajectory without directly altering the intermediate features within the diffusion model. Furthermore, since these features are derived from the sketching blocks, the guidance focuses on the coarse spatial and temporal content, such as basic layouts and object movements, leaving ample room for detailed edits to be made in the sharpening blocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "However, native guidance can lead to information leakage from the source video to the target video, as the attention outputs contain the full range of visual content. To address this, it is crucial to determine which aspects of the content should be retained and which should be modified. For this purpose, we turn to the frequency domain for assistance. Real-world images and videos often exhibit redundancy across their spatial and temporal dimensions, meaning that most of their energy is concentrated in the low-frequency spectrum, while the high-frequency components contribute much less. This characteristic underpins popular compression algorithms such as JPEG for images and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "MPEG for videos, and we aim to apply this principle here. Since we cannot get a clean video during the denoising process, we use attention outputs composed of tokens representing the noisy video. Unlike the discrete cosine transform (DCT) used in the spatial domain for MPEG, we employ the discrete Fourier transform on both spatial and temporal dimensions (3D DFT) to capture the spectral characteristics of intra-frame content and inter-frame motion:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.224, + 0.905, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {t} = \\mathrm {D F T} _ {\\mathrm {3 D}} \\left(\\boldsymbol {F} _ {t}\\right) \\in \\mathbb {R} ^ {h \\times w \\times \\tau \\times c}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.907, + 0.436 + ], + "angle": 0, + "content": "where \\( h \\), \\( w \\) and \\( \\tau \\) are the spatial and temporal size of the video tokens, and \\( c \\) is the channel dimension. As we know, most of the energy in \\( \\mathcal{F}_t \\) is concentrated in the low-frequency range. Thus, with the observed attention patterns, we can assume that fundamental spatial structures and temporal movements are encapsulated within this range, while the high-frequency components contain Gaussian noise and specific details like texture and subtle shaking—elements we wish to suppress or remove during editing. To achieve this, we apply a low-pass filter to isolate the low-frequency content of \\( \\mathcal{F}_t \\). The same process is applied to \\( \\mathcal{F}_t^* \\) to compute the spectrum guidance \\( \\mathcal{G}_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.447, + 0.905, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {t} = \\left\\| \\mathrm {L P} \\left(\\mathcal {F} _ {t}\\right) - \\mathrm {L P} \\left(\\mathcal {F} _ {t} ^ {*}\\right) \\right\\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.906, + 0.554 + ], + "angle": 0, + "content": "where LP denotes the low-pass filter. This guidance measures the deviation of the target video \\( \\mathbf{z}_t \\) from the source video in terms of basic structure and movement. This distance is then used to modulate the trajectory during DDIM sampling by the gradient of \\( \\mathcal{G}_t \\) with respect to \\( \\mathbf{z}_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.566, + 0.905, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t - 1} = \\operatorname {D D I M} \\left(\\boldsymbol {\\epsilon} _ {\\theta}, \\boldsymbol {z} _ {t}, t, \\boldsymbol {y} _ {t g t}\\right) - \\lambda \\operatorname {N o r m} \\left(\\nabla_ {\\boldsymbol {z} _ {t}} \\mathcal {G} _ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.671 + ], + "angle": 0, + "content": "where DDIM denotes the sampling process in Equation (3), \\(\\lambda\\) is the guidance weight and Norm normalizes the gradient to avoid numerical instability. This modulation strategy iteratively denoises the edited video towards the ideal direction, ensuring high consistency and realistic edits." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.822, + 0.697 + ], + "angle": 0, + "content": "3.4. Versatile Video Editing Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We utilize a pre-trained T2V diffusion model for video inversion and sampling. This model leverages a 3D VAE, comprising an encoder \\(\\mathcal{E}\\) and a decoder \\(\\mathcal{D}\\), to enable spatial and temporal compression, facilitating a reversible transformation between pixel space and latent space. In addition to text embeddings, the T2V model incorporates FPS information to better handle video data. Unlike one-shot tuned video models from T2I models [15], using a well-trained video diffusion model ensures a more stable understanding of the complex spatial and temporal relationships within videos. Moreover, this eliminates the need for model tuning, enhancing efficiency and reducing the necessity of storing individual models for specific videos. For edit types," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28430" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.097, + 0.116, + 0.156 + ], + "angle": 270, + "content": "Input video" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.093, + 0.489, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.166, + 0.408, + 0.179 + ], + "angle": 0, + "content": "Target: bushes \\(\\rightarrow\\) snowfield" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.183, + 0.489, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.256, + 0.489, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.328, + 0.489, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.411, + 0.489, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.484, + 0.387, + 0.497 + ], + "angle": 0, + "content": "Target: bear \\(\\rightarrow\\) panda" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.501, + 0.489, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.575, + 0.489, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.647, + 0.489, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.724, + 0.373, + 0.738 + ], + "angle": 0, + "content": "(a) Appearance Edits" + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.097, + 0.529, + 0.156 + ], + "angle": 270, + "content": "Input video" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.093, + 0.901, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.166, + 0.869, + 0.179 + ], + "angle": 0, + "content": "Target: turn the head \\(\\rightarrow\\) lower the head" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.183, + 0.901, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.256, + 0.901, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.328, + 0.901, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.412, + 0.901, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.484, + 0.863, + 0.497 + ], + "angle": 0, + "content": "Target: close the eyes \\(\\rightarrow\\) open the eyes" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.501, + 0.901, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.573, + 0.901, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.647, + 0.901, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.724, + 0.771, + 0.738 + ], + "angle": 0, + "content": "(b) Motion Edits" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.748, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on real-world videos. We evaluate both (a) appearance and (b) motion edits. Our framework consistently achieves accurate textual alignment across frames and high fidelity in unedited regions. Notably, compared to other methods based on video models, we produce more coherent and realistic motion edits. More results are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.482, + 0.892 + ], + "angle": 0, + "content": "we broadly categorize prompt-driven modifications into appearance edits and motion edits. Appearance edits involve altering nouns, such as changing the main subject or background, focusing on the texture, color, and style (e.g., transforming a \"cat\" to a \"tiger\" in Figure 3). Motion edits, on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.907, + 0.893 + ], + "angle": 0, + "content": "the other hand, involve changing verbs, which can alter spatial and temporal dynamics. By applying spectrum-guided modulation to harness inherent video priors within the T2V model, our method achieves realistic motion edits, distinguishing it from most previous approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28431" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.161 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on real-world video data. We compare our FADE framework to previous methods on both appearance and motion edits, measuring editing accuracy, input fidelity, temporal consistency, and human preference. FADE achieves competitive performance across diverse editing tasks." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.17, + 0.482, + 0.355 + ], + "angle": 0, + "content": "
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓PF ↑
appearance edits
Tune-A-Video [34]0.352219.860.462535.010.12
Video-P2P [15]0.358920.570.319934.710.15
FateZero [23]0.356220.650.305733.230.29
CogVideoX-V2V [35]0.375418.960.481131.450.09
FADE (Ours)0.376220.690.308531.360.35
motion edits
Tune-A-Video [34]0.328118.680.463735.850.10
Video-P2P [15]0.331418.930.388535.440.15
FateZero [23]0.325919.020.371234.470.13
CogVideoX-V2V [35]0.367818.170.492835.520.19
FADE (Ours)0.368319.260.369232.280.43
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.226, + 0.377 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.485, + 0.492 + ], + "angle": 0, + "content": "To verify the effectiveness of our method, we conduct comprehensive experiments across various video editing tasks. We will start by describing the experimental settings and then present our main results. We will also provide detailed ablation studies and analyses of our approach to highlight the contributions of each component in our framework and to validate the impact of our design choices." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.276, + 0.517 + ], + "angle": 0, + "content": "4.1. Experiment Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Implementation Details. To utilize video generation priors, we adopt a recent, widely used T2V model [35] for video inversion and generation. This model's denoising component employs the DiT architecture with 48 transformer blocks, each containing a full-attention layer. Based on observed attention map patterns, we designate the first 4 blocks as sketching blocks. We calculate the guidance over the interval \\([0, 0.6T]\\) of the DDIM sampling steps, with total steps set to \\(T = 50\\). To generate textual prompts from the source video, we use a multimodal language model [14] as the captioner. The guidance weight varies between 10 and 15, tailored to each editing task. We perform all evaluations on a single NVIDIA LS20 GPU with 48GB VRAM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets and Metrics. We evaluate our method using the DAVIS dataset [22] as well as various in-the-wild videos, following the approach of previous studies [15, 23]. Alongside qualitative results, we provide quantitative evaluations across multiple aspects using different metrics. Editing accuracy is assessed through CLIP score [24], which reflects the degree of textual alignment. To evaluate temporal consistency, a key factor for video coherence, we compute object semantic variance (OSV), following [15], to capture semantic consistency across frames. Additionally, we assess fidelity in unedited regions by calculating Mask PSNR and LPIPS [36], which quantify the preservation of content that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "should remain unchanged. Furthermore, given the importance of visual quality in video editing, we conduct a user study to gauge human preference across all editing tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.655, + 0.164 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.172, + 0.908, + 0.581 + ], + "angle": 0, + "content": "We compare our editing method with state-of-the-art approaches, including Video-P2P [15], FateZero [23], TuneA-Video (TAV) [34] and CogVideoX-V2V (CogV2V) [35]. As illustrated in Figure 4, our method consistently yields superior results across all aspects when compared to previous methods. The editing results of Video-P2P and FateZero exhibit significant limitations, often failing to modify content accurately, resulting in noticeable textual mismatches. Similarly, while TAV and CogV2V leverage T2V models, they struggle with motion editing, producing subpar results in capturing dynamic changes. Leveraging the inherent priors from the T2V model and spectrum-guided modulation, our method achieves the best performance in preserving local structure, maintaining temporal consistency, and ensuring text-video alignment. We evaluate our method on 20 videos and measure 5 metrics for quantitative analysis. As shown in Table 1, our method achieves the highest CLIP and OSV scores, demonstrating its superiority in textual alignment and temporal consistency. We also obtain competitive Mask PSNR and LPIPS, indicating that our method preserves unedited regions effectively. Moreover, we achieve the highest human preference score, reflecting the favorable visual quality and appeal of our edits across tasks. Notably, our method operates without any optimization or fine-tuning, completing various edits on real-world videos in approximately 3 minutes, a significant efficiency gain compared to previous methods, which require over 15 minutes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.617, + 0.607 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Efficient Editing via Sketching Blocks. We assess the function of different blocks in the T2V model by testing various block combinations to calculate modulation, holding other settings constant. We consider three configurations: (1) using sketching blocks only (ours), (2) symmetrically combining sketching and sharpening blocks (symm. blocks), and (3) using all blocks without factorization (w/o factor.). As shown in Table 2, configuration (1) provides comparable results to configurations (2) and (3), despite requiring fewer blocks for modulation. Although configurations (2) and (3) offer slightly improved fidelity, they significantly increase test time and memory demands. Figure 5 further demonstrates that relying on only the first four transformer blocks, or sketching blocks, achieves the best editing performance. Interestingly, contrary to common assumptions, including the sharpening blocks actually diminishes editing quality. We assume this is because the sharpening blocks primarily focus on elements like color, texture, and subtle vibrations, which often do not need" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28432" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.091, + 0.184, + 0.104 + ], + "angle": 0, + "content": "Input video" + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.091, + 0.273, + 0.103 + ], + "angle": 0, + "content": "w/o factor." + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.091, + 0.361, + 0.103 + ], + "angle": 0, + "content": "w/o filter" + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.091, + 0.446, + 0.103 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.107, + 0.189, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.107, + 0.285, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.107, + 0.381, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.107, + 0.476, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.383, + 0.432, + 0.397 + ], + "angle": 0, + "content": "Target: duck \\(\\rightarrow\\) yellow chick + open eyes" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.452 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of the ablations. We find the variant frameworks fall short in terms of editing accuracy, temporal consistency and input fidelity, yielding unsatisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.599 + ], + "angle": 0, + "content": "to remain unchanged in most video editing tasks. Relying on these blocks for guidance may, therefore, misdirect the model, diverting its attention from building the foundational layout and movement of the input video, ultimately resulting in reduced performance. These results underscore the effectiveness of our frequency-aware factorization approach, demonstrating that our selective focus on sketching blocks provides an optimal balance between maintaining high-quality edits and achieving computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Impact of Spectrum-Guided Modulation. In Figure 5 and Table 2, we examine our spectrum-guided modulation by removing the low-pass filter (w/o filter) and analyzing its effect on the editing performance. When we use the entire spectrum of attention outputs, high-frequency components are retained in the guidance term, causing the target object to retain too much of the source's original characteristics, which weakens the text-video alignment. Experiments show that using about two-thirds of the frequency components in the guidance term balances reconstruction and editing quality. The quantitative results further underscore the efficacy of spectrum-guided modulation, as shown by superior performance on key metrics like textual alignment and temporal consistency. By selectively modulating the frequency components, our method achieves precise edits that align well with the given textual prompt and maintain consistency across frames, both of which are crucial for high-quality video editing. These findings highlight that spectrum-guided modulation not only supports a stable structural foundation for the edits but also enhances flexibil" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.173 + ], + "angle": 0, + "content": "Table 2. Ablations studies. We perform ablations to verify the effectiveness of the components in FADE and the impact of the model factorization. We find that FADE achieves similar performance with only a subset of sketching blocks, compared to using symmetric or all blocks. Moreover, spectrum-guided modulation further enhances both preservation and precision in edits." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.184, + 0.903, + 0.263 + ], + "angle": 0, + "content": "
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Test-time
symm. blocks0.365920.730.336732.615 min
w/o factor.0.369120.940.332832.0512 min
w/o filter0.361220.890.336432.283 min
FADE (Ours)0.372820.870.335231.773 min
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.304 + ], + "angle": 0, + "content": "ity, allowing detailed adjustments for various editing tasks while ensuring input fidelity across the video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.906, + 0.576 + ], + "angle": 0, + "content": "Effective Exploitation of Video Diffusion Model. Our framework capitalizes on the rich video priors embedded in the pre-trained T2V model, enabling more effective utilization of the model's inherent knowledge of spatial layouts, temporal coherence, and object motions. The video diffusion model captures nuanced understanding across frames, making it an ideal foundation for edits that require both high fidelity and temporal consistency. To validate our superior exploitation of these priors, we compare FADE with CogVideoX-V2V [35], which uses the same underlying T2V model. As shown in Figure 4 and Table 1, FADE outperforms it both qualitatively and quantitatively across various editing tasks, demonstrating our framework's ability to harness the model's rich knowledge more effectively. Limitations. Despite FADE's effectiveness, its performance relies on the underlying video model and struggles with heavy occlusions requiring advanced temporal reasoning. We will address these in future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present FADE, a novel, training-free framework tailored for a diverse range of video editing tasks, encompassing both appearance and motion adjustments. Through frequency-aware factorization of the T2V model, FADE efficiently harnesses the model's inherent structural and temporal patterns, allowing each block to excel in its designated role. We further design the spectrum-guided modulation, which adjusts the sampling trajectory using frequency cues to preserve foundational video elements and enhance text-video alignment. Extensive experiments demonstrate our framework's competitive performance and generalization across various video editing challenges. We envision our work will inspire future research on versatile video editing and efficient video diffusion model. Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2024YFB4708100, and in part by the National Natural Science Foundation of China under Grant Grant 62321005, Grant 62306031" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28433" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.482, + 0.271 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[7] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In CVPR, pages 8599–8608, 2024. 2, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[16] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[17] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2, 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.267, + 0.905, + 0.321 + ], + "angle": 0, + "content": "[19] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.324, + 0.905, + 0.365 + ], + "angle": 0, + "content": "[20] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, pages 1-11, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.904, + 0.395 + ], + "angle": 0, + "content": "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.398, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[22] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[23] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.513, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.584, + 0.905, + 0.639 + ], + "angle": 0, + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.642, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.686, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[27] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. In ACML, pages 1215-1230. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.744, + 0.905, + 0.811 + ], + "angle": 0, + "content": "[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.904, + 0.843 + ], + "angle": 0, + "content": "[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-rescale: Self-guidance mechanism for effective tuning-free real image editing. arXiv preprint arXiv:2409.01322, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28434" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[33] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[35] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3, 4, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. URL https://github.com/hpcaitech/Open-Sora, 2024.3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[38] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28435" + } + ] +] \ No newline at end of file diff --git a/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_origin.pdf b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4bec2bdf072d285b55717b38c04aef7594941f4b --- /dev/null +++ b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8265345dea305946eec3732bf26f2efd3e090a7bd677a0ef660d6d230052b1c5 +size 5321802 diff --git a/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/full.md b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b44254fec3b9414a6e678357ce8d0fa0859944c2 --- /dev/null +++ b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/full.md @@ -0,0 +1,314 @@ +# FADE: Frequency-Aware Diffusion Model Factorization for Video Editing + +Yixuan Zhu $^{1}$ Haolin Wang $^{1}$ Shilin Ma $^{2}$ Wenliang Zhao $^{1}$ Yansong Tang $^{2}$ Lei Chen $^{1,\dagger}$ Jie Zhou $^{1}$ + +$^{1}$ Department of Automation, Tsinghua University $^{2}$ Tsinghua Shenzhen International Graduate School, Tsinghua University + +![](images/d7ec1f7755813d7983b26a29f2acb0334a1a86be45e65dab002f59ad5f3e74e8.jpg) + +![](images/d7c5c16893f07e6a046cc93ab0d18a974bfa57651d290e006f017b643157ef50.jpg) +Source Prompt: A black swan swimming in a river beside bushes + +![](images/de04e1b6a5ea46e26317a6f9fedaa4e7f39e0ed250dcfeb677fa8e8650432a84.jpg) + +![](images/6be7ba932d243646a4dd7595f157220bf0b59f917c832fb7f8eaab46974916ca.jpg) + +![](images/02067fb34f344144eeebb4f042d07bfb4b7487e1288cb4d16d7f0578c9242a2b.jpg) + +![](images/ae9adeb1a96b522f8b4d2da0bb3d2b6c33ef2a1b5c01fd5849109133726903be.jpg) + +![](images/b2901c2e83ce290be7cf7275af39cf2613eb278b30c07ff381767bacc5955d98.jpg) + +![](images/8ddfe18cb1a32226afe3e51c09950dccfd2fdd4b4cab147730c255f8f1d1bc78.jpg) + +![](images/6af6ec504268b3657596b005b39175d03f56635d471a31991416c054144a8f6b.jpg) + +![](images/47c7f95d9301121bcac84846bb6d68da6fd5a8b65203569ebdbd597f52a94379.jpg) + +![](images/5ae7fc593bf28322649e45f1a4c47f570798fbb33748e01ef081fbddd4e8478a.jpg) +Appearance Edit: black swan $\rightarrow$ white duck + +![](images/d4eefc3ed9ede70715c6e624a9def7df13a20cfb7592d6260367d44fb5f702a9.jpg) + +![](images/9eff71770f5449bb33a63adfe012fdcc60dc32ba03e811186894603972ab3fae.jpg) + +![](images/59547d8b164f4e1c8f7664d5413d54cff4f73d9395e5fc30ee2449a6c2e318de.jpg) + +![](images/fa07e97e6c01fece5f6d9569f5773fc3b1f020e3b2e5d9d86453c133bd96d6a5.jpg) + +![](images/f03fc8cbf1ebb92c424e541918311addc84248eebf43b64b24328c5bc6d07e15.jpg) + +![](images/7fb1a935cbfc356e0c7af102a8ceabdd7583e04e2ed7e1e724b1c2de5d9d2b7e.jpg) +Figure 1. Diverse video editing results of FADE. Our training-free approach, utilizing frequency-aware factorization and modulation, achieves high-fidelity, coherent edits across a variety of video types. FADE handles both appearance and motion adjustments with impressive robustness, ensuring precise alignment with input prompts and maintaining temporal consistency. Best viewed in color. + +![](images/971902c1377cd3ca873fa0f90cdd6442930b4fd1ad8e51d30949fcd2624e0834.jpg) + +![](images/391538d544dc45bd5d2a6c6dad3187bf0b988ab1bf5413402fb4daa60cce96f0.jpg) + +![](images/3a9398fc8ecb7acd2b2d9427b72d645f1de3a11380f38c9a25ad88aa0408a891.jpg) + +![](images/c3501d1ea27decfb3d8cbb1ca18336ab0fc1f562e6cb1ece2651ebfe0bee9d22.jpg) + +![](images/a63ce0aeb98d19eb4ab8482ac965b5153f3dab3474a6b724c6516fcf074de616.jpg) + +![](images/0b8226ae7b20ff1cb84c46fc0c22c46ef2889992f617420ad2e0b94040e28459.jpg) + +![](images/b7ff6632fc4a9321f61f72a784c56390b9dbbce202bb21b0493977df17a459ff.jpg) + +Motion Edit: swimming $\rightarrow$ flapping the wings + +# Abstract + +Recent advancements in diffusion frameworks have significantly enhanced video editing, achieving high fidelity and strong alignment with textual prompts. However, conventional approaches using image diffusion models fall short in handling video dynamics, particularly for challenging temporal edits like motion adjustments. While current video diffusion models produce high-quality results, adapting them for efficient editing remains difficult due to the heavy computational demands that prevent the direct application of previous image editing techniques. To overcome these limitations, we introduce FADE—a training-free yet highly effective video editing approach that fully leverages the inherent priors from pre-trained video diffusion models via frequency-aware factorization. Rather than simply using these models, we first analyze the attenuated performance of the proposed system. + +tion patterns within the video model to reveal how video priors are distributed across different components. Building on these insights, we propose a factorization strategy to optimize each component's specialized role. Furthermore, we devise spectrum-guided modulation to refine the sampling trajectory with frequency domain cues, preventing information leakage and supporting efficient, versatile edits while preserving the basic spatial and temporal structure. Extensive experiments on real-world videos demonstrate that our method consistently delivers high-quality, realistic and temporally coherent editing results both qualitatively and quantitatively. Code is available at https://github.com/EternalEvan/FADE. + +# 1. Introduction + +The objective of video editing is to modify specific visual content consistently across all frames while preserving the + +![](images/1ceada60a161a26322941d9a46db038f2bc1795e6a32a141b3bfe2a214153506.jpg) +Figure 2. (a) In typical T2V models, video tokens flattened along the temporal axis yield full-attention maps split into cross, spatial, and temporal attention. (b) Video priors in T2V models are factorized based on attention feature patterns. We visualize the attention map $M$ and result $F$ along with its spectrum for two blocks. The early block sketches the foundational structure and movement, while the later block refines details. This insight allows us to leverage each block's unique role for more efficient editing. + +integrity of unchanged regions. Recent achievements in diffusion models have greatly elevated video editing capabilities [10, 12, 15, 19, 20, 28, 32], bringing remarkable generative power to video editing tasks that require both spatial details and temporal coherence. While previous methods [1, 8, 13, 18, 30] have adapted text-to-image (T2I) diffusion models for prompt-driven image editing, their effectiveness diminishes when applied to video due to the complexities of managing spatial-temporal relationships across frames and a lack of generalized video prior knowledge. These limitations underscore the need for an efficient and robust editing strategy capable of handling the increased data and temporal coherence demands of videos. In response, we aim to design an accessible, high-quality editing solution that fully leverages the rich video priors embedded in pre-trained video diffusion models, enabling realistic, coherent, and versatile edits, including both appearance and motion adjustments, across complex video content. + +The main challenge in video editing is balancing changes to the target object while preserving unedited content. Previous methods largely rely on the generative power of T2I models for controlled edits. For instance, [15, 33] achieve this through null-text inversion [18], which minimizes inversion errors via test-time optimization. However, this process requires extensive iterations to compute a series of null-text embeddings for every sampling step, making it time-consuming and limiting its real-world applicability. Another line of work is attention feature injection [5, 15, 23, 27], where attention maps are modified by + +replacing or interleaving to maintain content consistency. Despite this, the absence of video priors often results in unrealistic temporal dynamics and limits motion edit capacity. Some methods, like [19], directly utilize video diffusion models by adding noise to the source video. This straightforward approach fails to fully exploit video priors, resulting in suboptimal editing quality. With the recent emergence of advanced text-to-video (T2V) models, new opportunities arise to harness their rich video knowledge. However, adapting previous T2I techniques to these models is challenging, given the extensive resources required, including memory, computation and time, to handle the increased data and complexity inherent in large-scale video networks. + +To overcome these challenges, we introduce FADE, a training-free yet potent approach for real-world video editing. Our goal is to harness the rich video priors about video appearance and dynamics within pre-trained T2V models to enable realistic and diverse editing. To achieve this, we start by examining the specific properties of these models. While previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps, we further observe that a similar frequency progression occurs block-wise within T2V models. We explore how video priors function in each block by analyzing their roles and patterns, revealing that early blocks sketch foundational spatial layouts and temporal dynamics while later blocks refine high-frequency details (Figure 2). Leveraging this insight, we factorize the editing problem by the distinct functions of each component, enhancing editing efficiency. Then, instead of performing optimization or attention injection, we design a spectrum-guided modulation strategy that utilizes the full-attention outputs in these sketching blocks to derive frequency-aware guidance through 3D Fourier transformation. This modulation adjusts the sampling trajectory to preserve low-frequency structure from the source video while allowing high-frequency details to support flexible and intricate edits. This cue from the frequency domain not only mitigates information leakage during guidance computation but also unleashes the potential of video diffusion priors, enabling sophisticated spatial and temporal edits with high fidelity and coherence. + +We evaluate our method using extensive real-world video clips across various editing tasks. As shown in Figure 1, FADE exhibits proficiency in producing realistic and consistent edits, covering both appearance and motion modifications. Quantitative comparisons on video editing benchmark DAVIS [22] demonstrate our approach achieves superior textual alignment and temporal consistency, with 0.3561 CLIP score and 43.57 OSV. Moreover, we maintain competitive input fidelity with 21.54 MaskPSNR and 0.3297 LPIPS, preserving unedited regions effectively. Compared to previous methods, our use of general video priors with frequency-aware factorization deliv + +![](images/408f92fb59348413bcb1f228068ff6dc11626f7bf0d4d6e1cac88f616319973b.jpg) +Figure 3. The overall framework of FADE. FADE is a training-free framework for video editing. Given an input video, we first perform video inversion to obtain noise $\mathbf{z}_T$ and trajectory $\{\mathbf{z}_t^*\}_{t=0}^T$ . Starting from this noise, we use sketching blocks in the T2V model to produce full-attention results $\mathbf{F}_t$ and $\mathbf{F}_t^*$ from $\mathbf{z}_t$ and $\mathbf{z}_t^*$ . These results are transformed to the frequency domain, where a low-pass filter isolates spatial and temporal structures. The frequency error then modulates the sampling process, enabling high-quality, consistent edits. + +ers improved efficiency and a broader editing range, highlighting FADE's outstanding versatility and practicality. + +# 2. Related Works + +Video Diffusion Models. In recent years, denoising diffusion models [9, 29] have emerged as a leading class of generative models, notable for their synthesis quality and controllability. After training on large-scale datasets, text-to-image (T2I) diffusion models [6, 25] exhibit exceptional image synthesis capability. For video generation, some works [2, 11] adapt image diffusion models by inflating the denoising UNet [26] with temporal layers to handle video inputs. However, replicating the success of T2I models for video generation remains challenging due to the lack of large-scale video datasets and the increased data complexity. Most recently, Sora [4] and following studies [16, 35, 37] pioneer high-quality, high-resolution text-to-video (T2V) diffusion models. These models commonly use the diffusion transformer (DiT) architecture [21] to capture complex spatial and temporal coherence, outperforming previous UNet-based methods. However, this improved performance comes at the cost of high computational demands, mostly due to the extensive attention mechanism, which requires processing a large number of tokens. Limited research has explored the pattern within these models to reduce the computational intensity. To address this, we delve into their internal structure to analyze the roles of individual blocks, leveraging these insights to facilitate their applicability and efficiency in video editing. + +Video Editing. Significant progress in image editing [3, 8, 17, 18, 20, 31] has driven advancements in video editing [7, 15, 19, 23, 34]. However, unique challenges in video editing arise from the need for consistent, coherent + +edits across frames, including both appearance and motion changes. Recent works [15, 23] adopt the attention feature injection techniques, manipulating various attention results to guide edits. They also employ model inflation to tune a T2I model with the input video following [34]. These approaches, however, are inefficient since they require a separate model for each sample and lack the general video priors to handle motion edits. In contrast, [19] utilizes a pre-trained video diffusion model for editing, beginning with corrupting the video by downsampling and noise. However, this corruption can result in global inconsistency with the source video and limits control over specific edits. Overall, there remains significant potential for advancing the use of video diffusion models in video editing. This application of the video diffusion model is nontrivial because (1) it requires maintaining global consistency in the generated video without relying on one-shot tuning or null-text optimization [18], and (2) applying previous attention feature injection methods is difficult with modern video models, as they involve dozens of transformer blocks with memory-intensive, full-attention computations. Our frequency-aware approach effectively addresses these challenges, setting our method apart from previous work. + +# 3. Method + +In this section, we present FADE, a training-free framework that achieves real-world video editing with frequency-aware factorization of diffusion models. Our key idea is to investigate the inherent patterns in a pre-trained text-to-video diffusion model and activate each block's specific function for various video editing tasks. We will start by reviewing the background of video diffusion models, and then describe our designs of FADE, including how to factorize video dif + +fusion models according to frequency cues and adjust the sampling trajectory via spectrum-guided modulation. The overall framework of our FADE is illustrated in Figure 3. + +# 3.1. Preliminaries: Video Diffusion Models + +Video diffusion models are a class of generative models that reconstruct the distribution of video data by modeling the reverse trajectory of a diffusion process. Given a random noise $\epsilon \sim \mathcal{N}(0, I)$ , the diffusion process is defined as: + +$$ +\boldsymbol {z} _ {t} = \alpha_ {t} \boldsymbol {z} _ {0} + \sigma_ {t} \epsilon , \tag {1} +$$ + +where $\mathbf{z}_0$ and $\mathbf{z}_t$ are the clean and noisy video, and $\left\{ (\alpha_t, \sigma_t) \right\}_{t=1}^T$ denotes the noise schedule. With proper reparameterization, the training objective of diffusion models can be expressed as: + +$$ +\mathcal {L} _ {\mathrm {D M}} = \mathbb {E} _ {\boldsymbol {z} _ {0}, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} (\boldsymbol {z} _ {t}, t, \boldsymbol {y}) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $z_{t}$ is computed as Equation (1). The denoising model $\epsilon_{\theta}$ is learned to predict the noise given the prompt $y$ . Previous T2I models often implement $\epsilon_{\theta}$ as UNet [26]. For video generation, recent models [4, 16, 35, 37] usually employ a series of cascaded transformer blocks to capture the spatial-temporal relationships more effectively. + +To generate a video in a small number of denoising steps, diffusion models usually use deterministic DDIM sampling: + +$$ +\boldsymbol {z} _ {t - 1} = \frac {\alpha_ {t - 1}}{\alpha_ {t}} \boldsymbol {z} _ {t} + \left(\sigma_ {t - 1} - \frac {\alpha_ {t - 1}}{\alpha_ {t}} \sigma_ {t}\right) \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {y}\right). \tag {3} +$$ + +Conversely, to convert a video to its corresponding noise, we can apply DDIM inversion, the reverse of Equation (3): + +$$ +\boldsymbol {z} _ {t + 1} = \frac {\alpha_ {t + 1}}{\alpha_ {t}} \boldsymbol {z} _ {t} + \left(\sigma_ {t + 1} - \frac {\alpha_ {t + 1}}{\alpha_ {t}} \sigma_ {t}\right) \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {y}\right). \tag {4} +$$ + +In practice, DDIM inversion introduces slight errors in every step, making it unreliable for fully reconstructing a real-world video or preserving fine details in the source video. + +# 3.2. Frequency-Aware Factorization of T2V Model + +Modern T2V models typically employ cascades of transformer blocks as the denoising model, with each block containing attention layers that process extensive video tokens. Given that dozens of identical blocks are involved in calculating the same form of attention, a key question arises: Do all these blocks serve the same purpose and contribute equally to video synthesis? To explore this, we analyze the behavior within each block of a pre-trained T2V model during video generation. Previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps. Inspired by this, we adopt a frequency-aware perspective to uncover if similar trends occur at the block level in video models. Specifically, we capture and visualize each block's attention maps + +$\{M_i\}_{i=1}^N$ and outputs $\{F_i\}_{i=1}^N$ , along with their Fourier spectra, over all $N$ blocks during sampling. As illustrated in Figure 2, a typical attention map in T2V models contains three types of attention regions: cross-attention for multimodal alignments, spatial attention for capturing intra-frame spatial structures, and temporal attention for tracking dynamics across frames. Our observation is that spatial and temporal attention are densely aligned along diagonal lines in the early blocks, creating multiple parallel diagonals in the attention maps. The main diagonal line suggests that these early blocks focus on outlining the intra-frame spatial structures, defining key shapes and boundaries of the composition. The secondary diagonal lines indicate that these blocks are simultaneously establishing temporal correspondences across frames within the same regions. Collectively, these early blocks construct the low-frequency, foundational structure of the video, setting up basic object placement and movement. We refer to them as sketching blocks. In contrast, the following blocks—which constitute the majority of the denoising model—display relatively sparse attention distribution, with scores more evenly spread across elements in the map. This implies that these later blocks focus on refining the high-frequency refinements, such as color, texture, and subtle vibration, providing nuanced adjustments for visual details. We call them sharpening blocks, as their primary role is to enhance details, enriching the visual appeal of the final output. + +The variation in attention outputs of these blocks further supports our inference. The initial sketching blocks produce blurry outputs, with spectra focused on low frequencies that capture basic spatial and temporal layouts. In contrast, the sharpening blocks reveal clear foreground and background regions, involving much more high-frequency details. This distinction suggests that these two classes of blocks process features of differing levels and serve different purposes. We further verify that each block's function remains consistent throughout the sampling steps by analyzing these patterns across all steps. Thus, for video content editing—where maintaining underlying structures while adjusting finer details is essential—we can factorize each block's role, using sketching blocks for structural reconstruction and sharpening blocks for detailed generation. This division offers two key advantages: (1) it enables each block to perform optimally within its specialized function, and (2) it greatly reduces computational load by enabling frequency-aware guidance through selective use of blocks during sampling. + +# 3.3. Spectrum-Guided Modulation + +After factorizing the video model based on frequency-informed attention patterns, our next step is to activate each block's specific function and develop a robust sampling strategy that optimally balances reconstruction with generation. We begin by employing the inversion process de + +scribed in Equation (4) with a pre-trained video diffusion model $\epsilon_{\theta}$ . This transforms the real-world video $z_0^*$ to an approximated noise $z_T^*$ and produces a reverse trajectory $\{z_t^*\}_{t=0}^T$ . However, as discussed in prior works, DDIM inversion can introduce non-negligible errors, making it difficult to accurately reconstruct the source video from $z_T^*$ . This issue is particularly pronounced in video inversion, where the predicted noise $z_T^*$ often deviates far from the standard Gaussian distribution. We, therefore, focus on refining the sampling process using the observed spectrum patterns to correct the inversion error and create an ideal trajectory $\{z_t\}_{t=0}^T$ to a plausible edited video $z_0$ . Following previous image editing approaches [18, 30], we adopt a dual branch strategy that leverages the inversion trajectory to enhance consistency with the source video. However, existing attention feature injection methods require extensive memory due to full-attention maps and limit editing flexibility through restrictive attention fusion techniques. While these methods can preserve the spatial layout, they often struggle with more complex edits involving pose, shape and motion. + +To overcome these limitations, we propose a more flexible, spectrum-guided approach that factorizes visual information in the frequency domain to separate content for preservation and editing. This approach offers greater freedom for a wider range of edits. As shown in Figure 3, our first design is to harness the sketching blocks, which care about the overall spatial layout and temporal movement. Rather than swapping or interleaving the attention maps directly, we gather the attention outputs of the source video $\boldsymbol{z}_t^*$ and the target video $\boldsymbol{z}_t$ within these blocks, denoted as $\boldsymbol{F}_t^*$ and $\boldsymbol{F}_t$ , and use their difference to derive a guidance term for the sampling process. Specifically, these attention outputs are computed using the same source prompt $\boldsymbol{y}_{src}$ through the full-attention layers. This guidance term provides a more robust and adaptive way to adjust the sampling trajectory without directly altering the intermediate features within the diffusion model. Furthermore, since these features are derived from the sketching blocks, the guidance focuses on the coarse spatial and temporal content, such as basic layouts and object movements, leaving ample room for detailed edits to be made in the sharpening blocks. + +However, native guidance can lead to information leakage from the source video to the target video, as the attention outputs contain the full range of visual content. To address this, it is crucial to determine which aspects of the content should be retained and which should be modified. For this purpose, we turn to the frequency domain for assistance. Real-world images and videos often exhibit redundancy across their spatial and temporal dimensions, meaning that most of their energy is concentrated in the low-frequency spectrum, while the high-frequency components contribute much less. This characteristic underpins popular compression algorithms such as JPEG for images and + +MPEG for videos, and we aim to apply this principle here. Since we cannot get a clean video during the denoising process, we use attention outputs composed of tokens representing the noisy video. Unlike the discrete cosine transform (DCT) used in the spatial domain for MPEG, we employ the discrete Fourier transform on both spatial and temporal dimensions (3D DFT) to capture the spectral characteristics of intra-frame content and inter-frame motion: + +$$ +\mathcal {F} _ {t} = \mathrm {D F T} _ {\mathrm {3 D}} \left(\boldsymbol {F} _ {t}\right) \in \mathbb {R} ^ {h \times w \times \tau \times c}, \tag {5} +$$ + +where $h$ , $w$ and $\tau$ are the spatial and temporal size of the video tokens, and $c$ is the channel dimension. As we know, most of the energy in $\mathcal{F}_t$ is concentrated in the low-frequency range. Thus, with the observed attention patterns, we can assume that fundamental spatial structures and temporal movements are encapsulated within this range, while the high-frequency components contain Gaussian noise and specific details like texture and subtle shaking—elements we wish to suppress or remove during editing. To achieve this, we apply a low-pass filter to isolate the low-frequency content of $\mathcal{F}_t$ . The same process is applied to $\mathcal{F}_t^*$ to compute the spectrum guidance $\mathcal{G}_t$ : + +$$ +\mathcal {G} _ {t} = \left\| \mathrm {L P} \left(\mathcal {F} _ {t}\right) - \mathrm {L P} \left(\mathcal {F} _ {t} ^ {*}\right) \right\| _ {2} ^ {2}, \tag {6} +$$ + +where LP denotes the low-pass filter. This guidance measures the deviation of the target video $\mathbf{z}_t$ from the source video in terms of basic structure and movement. This distance is then used to modulate the trajectory during DDIM sampling by the gradient of $\mathcal{G}_t$ with respect to $\mathbf{z}_t$ : + +$$ +\boldsymbol {z} _ {t - 1} = \operatorname {D D I M} \left(\boldsymbol {\epsilon} _ {\theta}, \boldsymbol {z} _ {t}, t, \boldsymbol {y} _ {t g t}\right) - \lambda \operatorname {N o r m} \left(\nabla_ {\boldsymbol {z} _ {t}} \mathcal {G} _ {t}\right), \tag {7} +$$ + +where DDIM denotes the sampling process in Equation (3), $\lambda$ is the guidance weight and Norm normalizes the gradient to avoid numerical instability. This modulation strategy iteratively denoises the edited video towards the ideal direction, ensuring high consistency and realistic edits. + +# 3.4. Versatile Video Editing Framework + +We utilize a pre-trained T2V diffusion model for video inversion and sampling. This model leverages a 3D VAE, comprising an encoder $\mathcal{E}$ and a decoder $\mathcal{D}$ , to enable spatial and temporal compression, facilitating a reversible transformation between pixel space and latent space. In addition to text embeddings, the T2V model incorporates FPS information to better handle video data. Unlike one-shot tuned video models from T2I models [15], using a well-trained video diffusion model ensures a more stable understanding of the complex spatial and temporal relationships within videos. Moreover, this eliminates the need for model tuning, enhancing efficiency and reducing the necessity of storing individual models for specific videos. For edit types, + +![](images/4dbc2aeb531ffdc4ab3147667131a9d851c7f0a8b51f003cc40a8f4df8c69548.jpg) +Input video +Target: bushes $\rightarrow$ snowfield + +![](images/da7e528571eca048c37c8a75cf3a829de56378f63f479acabb3a0590ba816ba3.jpg) + +![](images/c0d33bd845d0a8f78efc3a9699ec8eec5803970f6dd3a8b97e61522c8d924e78.jpg) + +![](images/40f2debd73c2cf9fc0e95ca8c21954e428b39df3dd23d21524d191285ffe679d.jpg) + +![](images/fc2b030c8f9d36c14d41f7958f41bf6d61e56a538f0b50a7c98999aef5ee73dc.jpg) +Target: bear $\rightarrow$ panda + +![](images/e4dc63f2b95716796af97e70af5865969eeaf017b793bc4ea717bf2e46f2e5f3.jpg) + +![](images/35f4bad7762172f262605bbbfaabb3f2bc1d0747169f811ae1f78c4b5a67ff51.jpg) + +![](images/7c662cb31fc0838798ed20c80a3dda2158dc82962562750868f8d593edc26a7b.jpg) +(a) Appearance Edits +Figure 4. Qualitative comparisons on real-world videos. We evaluate both (a) appearance and (b) motion edits. Our framework consistently achieves accurate textual alignment across frames and high fidelity in unedited regions. Notably, compared to other methods based on video models, we produce more coherent and realistic motion edits. More results are available in the supplementary materials. + +![](images/806885260d8a4837dfc356114ed573ae44499672db366f0a1124559f466222ea.jpg) +Input video +Target: turn the head $\rightarrow$ lower the head + +![](images/0d126405e4c73934466888b3ea1d79e04d457f1d35b5166970973c4228292122.jpg) + +![](images/7947c3bcf39e7289c959346cd2d07e878b01a2db0ba9b76ef9f5b494b3b5291a.jpg) + +![](images/192b1ff75e1e116beaed25a639b512cf8309f2916b51bdac0e4694299d7b9502.jpg) + +![](images/a276093bcbd25dd356bf4c6429d7ac1dfc2170be2c5b9b8062e5a46bb37a72f1.jpg) +Target: close the eyes $\rightarrow$ open the eyes + +![](images/1516098ad6e4c17b6c078b092c14a992770223264aa780b0030ddaea145934c7.jpg) + +![](images/20abeeaeeacdfb9483f944557878d4ea6ed9bdf4ddbbee81887e1b553c0cc28d.jpg) + +![](images/2252fc07fa7183872b460c2288fae4591475ed5bfe3fbd5f04995a9ff7c348ab.jpg) +(b) Motion Edits + +we broadly categorize prompt-driven modifications into appearance edits and motion edits. Appearance edits involve altering nouns, such as changing the main subject or background, focusing on the texture, color, and style (e.g., transforming a "cat" to a "tiger" in Figure 3). Motion edits, on + +the other hand, involve changing verbs, which can alter spatial and temporal dynamics. By applying spectrum-guided modulation to harness inherent video priors within the T2V model, our method achieves realistic motion edits, distinguishing it from most previous approaches. + +Table 1. Quantitative comparisons on real-world video data. We compare our FADE framework to previous methods on both appearance and motion edits, measuring editing accuracy, input fidelity, temporal consistency, and human preference. FADE achieves competitive performance across diverse editing tasks. + +
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓PF ↑
appearance edits
Tune-A-Video [34]0.352219.860.462535.010.12
Video-P2P [15]0.358920.570.319934.710.15
FateZero [23]0.356220.650.305733.230.29
CogVideoX-V2V [35]0.375418.960.481131.450.09
FADE (Ours)0.376220.690.308531.360.35
motion edits
Tune-A-Video [34]0.328118.680.463735.850.10
Video-P2P [15]0.331418.930.388535.440.15
FateZero [23]0.325919.020.371234.470.13
CogVideoX-V2V [35]0.367818.170.492835.520.19
FADE (Ours)0.368319.260.369232.280.43
+ +# 4. Experiments + +To verify the effectiveness of our method, we conduct comprehensive experiments across various video editing tasks. We will start by describing the experimental settings and then present our main results. We will also provide detailed ablation studies and analyses of our approach to highlight the contributions of each component in our framework and to validate the impact of our design choices. + +# 4.1. Experiment Setups + +Implementation Details. To utilize video generation priors, we adopt a recent, widely used T2V model [35] for video inversion and generation. This model's denoising component employs the DiT architecture with 48 transformer blocks, each containing a full-attention layer. Based on observed attention map patterns, we designate the first 4 blocks as sketching blocks. We calculate the guidance over the interval $[0, 0.6T]$ of the DDIM sampling steps, with total steps set to $T = 50$ . To generate textual prompts from the source video, we use a multimodal language model [14] as the captioner. The guidance weight varies between 10 and 15, tailored to each editing task. We perform all evaluations on a single NVIDIA LS20 GPU with 48GB VRAM. + +Datasets and Metrics. We evaluate our method using the DAVIS dataset [22] as well as various in-the-wild videos, following the approach of previous studies [15, 23]. Alongside qualitative results, we provide quantitative evaluations across multiple aspects using different metrics. Editing accuracy is assessed through CLIP score [24], which reflects the degree of textual alignment. To evaluate temporal consistency, a key factor for video coherence, we compute object semantic variance (OSV), following [15], to capture semantic consistency across frames. Additionally, we assess fidelity in unedited regions by calculating Mask PSNR and LPIPS [36], which quantify the preservation of content that + +should remain unchanged. Furthermore, given the importance of visual quality in video editing, we conduct a user study to gauge human preference across all editing tasks. + +# 4.2. Main Results + +We compare our editing method with state-of-the-art approaches, including Video-P2P [15], FateZero [23], TuneA-Video (TAV) [34] and CogVideoX-V2V (CogV2V) [35]. As illustrated in Figure 4, our method consistently yields superior results across all aspects when compared to previous methods. The editing results of Video-P2P and FateZero exhibit significant limitations, often failing to modify content accurately, resulting in noticeable textual mismatches. Similarly, while TAV and CogV2V leverage T2V models, they struggle with motion editing, producing subpar results in capturing dynamic changes. Leveraging the inherent priors from the T2V model and spectrum-guided modulation, our method achieves the best performance in preserving local structure, maintaining temporal consistency, and ensuring text-video alignment. We evaluate our method on 20 videos and measure 5 metrics for quantitative analysis. As shown in Table 1, our method achieves the highest CLIP and OSV scores, demonstrating its superiority in textual alignment and temporal consistency. We also obtain competitive Mask PSNR and LPIPS, indicating that our method preserves unedited regions effectively. Moreover, we achieve the highest human preference score, reflecting the favorable visual quality and appeal of our edits across tasks. Notably, our method operates without any optimization or fine-tuning, completing various edits on real-world videos in approximately 3 minutes, a significant efficiency gain compared to previous methods, which require over 15 minutes. + +# 4.3. Analysis + +Efficient Editing via Sketching Blocks. We assess the function of different blocks in the T2V model by testing various block combinations to calculate modulation, holding other settings constant. We consider three configurations: (1) using sketching blocks only (ours), (2) symmetrically combining sketching and sharpening blocks (symm. blocks), and (3) using all blocks without factorization (w/o factor.). As shown in Table 2, configuration (1) provides comparable results to configurations (2) and (3), despite requiring fewer blocks for modulation. Although configurations (2) and (3) offer slightly improved fidelity, they significantly increase test time and memory demands. Figure 5 further demonstrates that relying on only the first four transformer blocks, or sketching blocks, achieves the best editing performance. Interestingly, contrary to common assumptions, including the sharpening blocks actually diminishes editing quality. We assume this is because the sharpening blocks primarily focus on elements like color, texture, and subtle vibrations, which often do not need + +![](images/0b17bef75167506b7f24da7971274f502a76e0da9e419430faa7998ee33e0936.jpg) +Input video +Target: duck $\rightarrow$ yellow chick + open eyes +Figure 5. Qualitative comparisons of the ablations. We find the variant frameworks fall short in terms of editing accuracy, temporal consistency and input fidelity, yielding unsatisfactory results. + +![](images/3ab7d17fbdb49e07c3edf267bff92809baec101f02192a9396581245a3088dfd.jpg) +w/o factor. + +![](images/9b6d16e720d60da30d7ad732e5f4b20a1050b3307a311373b6bf9c602920cb49.jpg) +w/o filter + +![](images/615cb00bb2207ffa78aedbfa41e1e47aa973edd39f2ac863f82db0d602280004.jpg) +Ours + +to remain unchanged in most video editing tasks. Relying on these blocks for guidance may, therefore, misdirect the model, diverting its attention from building the foundational layout and movement of the input video, ultimately resulting in reduced performance. These results underscore the effectiveness of our frequency-aware factorization approach, demonstrating that our selective focus on sketching blocks provides an optimal balance between maintaining high-quality edits and achieving computational efficiency. + +Impact of Spectrum-Guided Modulation. In Figure 5 and Table 2, we examine our spectrum-guided modulation by removing the low-pass filter (w/o filter) and analyzing its effect on the editing performance. When we use the entire spectrum of attention outputs, high-frequency components are retained in the guidance term, causing the target object to retain too much of the source's original characteristics, which weakens the text-video alignment. Experiments show that using about two-thirds of the frequency components in the guidance term balances reconstruction and editing quality. The quantitative results further underscore the efficacy of spectrum-guided modulation, as shown by superior performance on key metrics like textual alignment and temporal consistency. By selectively modulating the frequency components, our method achieves precise edits that align well with the given textual prompt and maintain consistency across frames, both of which are crucial for high-quality video editing. These findings highlight that spectrum-guided modulation not only supports a stable structural foundation for the edits but also enhances flexibil + +Table 2. Ablations studies. We perform ablations to verify the effectiveness of the components in FADE and the impact of the model factorization. We find that FADE achieves similar performance with only a subset of sketching blocks, compared to using symmetric or all blocks. Moreover, spectrum-guided modulation further enhances both preservation and precision in edits. + +
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Test-time
symm. blocks0.365920.730.336732.615 min
w/o factor.0.369120.940.332832.0512 min
w/o filter0.361220.890.336432.283 min
FADE (Ours)0.372820.870.335231.773 min
+ +ity, allowing detailed adjustments for various editing tasks while ensuring input fidelity across the video. + +Effective Exploitation of Video Diffusion Model. Our framework capitalizes on the rich video priors embedded in the pre-trained T2V model, enabling more effective utilization of the model's inherent knowledge of spatial layouts, temporal coherence, and object motions. The video diffusion model captures nuanced understanding across frames, making it an ideal foundation for edits that require both high fidelity and temporal consistency. To validate our superior exploitation of these priors, we compare FADE with CogVideoX-V2V [35], which uses the same underlying T2V model. As shown in Figure 4 and Table 1, FADE outperforms it both qualitatively and quantitatively across various editing tasks, demonstrating our framework's ability to harness the model's rich knowledge more effectively. Limitations. Despite FADE's effectiveness, its performance relies on the underlying video model and struggles with heavy occlusions requiring advanced temporal reasoning. We will address these in future work. + +# 5. Conclusion + +In this paper, we present FADE, a novel, training-free framework tailored for a diverse range of video editing tasks, encompassing both appearance and motion adjustments. Through frequency-aware factorization of the T2V model, FADE efficiently harnesses the model's inherent structural and temporal patterns, allowing each block to excel in its designated role. We further design the spectrum-guided modulation, which adjusts the sampling trajectory using frequency cues to preserve foundational video elements and enhance text-video alignment. Extensive experiments demonstrate our framework's competitive performance and generalization across various video editing challenges. We envision our work will inspire future research on versatile video editing and efficient video diffusion model. Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2024YFB4708100, and in part by the National Natural Science Foundation of China under Grant Grant 62321005, Grant 62306031 + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 2 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3, 4 +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3 +[7] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3 +[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3, 4 +[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 3 +[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2 +[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, pages 6007-6017, 2023. 2 +[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 7 +[15] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In CVPR, pages 8599–8608, 2024. 2, 3, 5, 7 + +[16] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 4 +[17] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3 +[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2, 3, 5, 1 +[19] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3 +[20] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, pages 1-11, 2023. 2, 3 +[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3 +[22] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7 +[23] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3, 7 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 7 +[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3 +[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3, 4 +[27] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. In ACML, pages 1215-1230. PMLR, 2024. 2 +[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2 +[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3 +[30] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-rescale: Self-guidance mechanism for effective tuning-free real image editing. arXiv preprint arXiv:2409.01322, 2024. 2, 5 + +[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 3 +[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2 +[33] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2 +[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 7 +[35] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3, 4, 7, 8, 1 +[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7 +[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. URL https://github.com/hpcaitech/Open-Sora, 2024.3,4 +[38] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1 \ No newline at end of file diff --git a/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/images.zip b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..655a08e883f41f252d9d40479e040e574597e29b --- /dev/null +++ b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad2bc66f227cd6c6271ad11852bd36aaa005a04127ba347cb0469f91a90bd283 +size 791296 diff --git a/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/layout.json b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00b1f8e9eddf140c0f809f8ca8676df6e32e8b00 --- /dev/null +++ b/2025/FADE_ Frequency-Aware Diffusion Model Factorization for Video Editing/layout.json @@ -0,0 +1,8296 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "text", + "content": "FADE: Frequency-Aware Diffusion Model Factorization for Video Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "content": "Yixuan Zhu" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "content": " Haolin Wang" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "content": " Shilin Ma" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "content": " Wenliang Zhao" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "content": " Yansong Tang" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "content": " Lei Chen" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "text", + "content": " Jie Zhou" + }, + { + "bbox": [ + 61, + 142, + 567, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 171, + 490, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 171, + 490, + 200 + ], + "spans": [ + { + "bbox": [ + 137, + 171, + 490, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 171, + 490, + 200 + ], + "type": "text", + "content": "Department of Automation, Tsinghua University " + }, + { + "bbox": [ + 137, + 171, + 490, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 171, + 490, + 200 + ], + "type": "text", + "content": "Tsinghua Shenzhen International Graduate School, Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 213, + 119, + 274 + ], + "blocks": [ + { + "bbox": [ + 58, + 213, + 119, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 119, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 119, + 274 + ], + "type": "image", + "image_path": "d7ec1f7755813d7983b26a29f2acb0334a1a86be45e65dab002f59ad5f3e74e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 120, + 213, + 181, + 274 + ], + "blocks": [ + { + "bbox": [ + 120, + 213, + 181, + 274 + ], + "lines": [ + { + "bbox": [ + 120, + 213, + 181, + 274 + ], + "spans": [ + { + "bbox": [ + 120, + 213, + 181, + 274 + ], + "type": "image", + "image_path": "d7c5c16893f07e6a046cc93ab0d18a974bfa57651d290e006f017b643157ef50.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 276, + 441, + 287 + ], + "lines": [ + { + "bbox": [ + 166, + 276, + 441, + 287 + ], + "spans": [ + { + "bbox": [ + 166, + 276, + 441, + 287 + ], + "type": "text", + "content": "Source Prompt: A black swan swimming in a river beside bushes" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 182, + 213, + 242, + 274 + ], + "blocks": [ + { + "bbox": [ + 182, + 213, + 242, + 274 + ], + "lines": [ + { + "bbox": [ + 182, + 213, + 242, + 274 + ], + "spans": [ + { + "bbox": [ + 182, + 213, + 242, + 274 + ], + "type": "image", + "image_path": "de04e1b6a5ea46e26317a6f9fedaa4e7f39e0ed250dcfeb677fa8e8650432a84.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 244, + 213, + 304, + 274 + ], + "blocks": [ + { + "bbox": [ + 244, + 213, + 304, + 274 + ], + "lines": [ + { + "bbox": [ + 244, + 213, + 304, + 274 + ], + "spans": [ + { + "bbox": [ + 244, + 213, + 304, + 274 + ], + "type": "image", + "image_path": "6be7ba932d243646a4dd7595f157220bf0b59f917c832fb7f8eaab46974916ca.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 213, + 365, + 274 + ], + "blocks": [ + { + "bbox": [ + 306, + 213, + 365, + 274 + ], + "lines": [ + { + "bbox": [ + 306, + 213, + 365, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 365, + 274 + ], + "type": "image", + "image_path": "02067fb34f344144eeebb4f042d07bfb4b7487e1288cb4d16d7f0578c9242a2b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 367, + 213, + 427, + 274 + ], + "blocks": [ + { + "bbox": [ + 367, + 213, + 427, + 274 + ], + "lines": [ + { + "bbox": [ + 367, + 213, + 427, + 274 + ], + "spans": [ + { + "bbox": [ + 367, + 213, + 427, + 274 + ], + "type": "image", + "image_path": "ae9adeb1a96b522f8b4d2da0bb3d2b6c33ef2a1b5c01fd5849109133726903be.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 429, + 213, + 489, + 274 + ], + "blocks": [ + { + "bbox": [ + 429, + 213, + 489, + 274 + ], + "lines": [ + { + "bbox": [ + 429, + 213, + 489, + 274 + ], + "spans": [ + { + "bbox": [ + 429, + 213, + 489, + 274 + ], + "type": "image", + "image_path": "b2901c2e83ce290be7cf7275af39cf2613eb278b30c07ff381767bacc5955d98.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 490, + 213, + 551, + 274 + ], + "blocks": [ + { + "bbox": [ + 490, + 213, + 551, + 274 + ], + "lines": [ + { + "bbox": [ + 490, + 213, + 551, + 274 + ], + "spans": [ + { + "bbox": [ + 490, + 213, + 551, + 274 + ], + "type": "image", + "image_path": "8ddfe18cb1a32226afe3e51c09950dccfd2fdd4b4cab147730c255f8f1d1bc78.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 58, + 289, + 119, + 349 + ], + "blocks": [ + { + "bbox": [ + 58, + 289, + 119, + 349 + ], + "lines": [ + { + "bbox": [ + 58, + 289, + 119, + 349 + ], + "spans": [ + { + "bbox": [ + 58, + 289, + 119, + 349 + ], + "type": "image", + "image_path": "6af6ec504268b3657596b005b39175d03f56635d471a31991416c054144a8f6b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 120, + 289, + 181, + 349 + ], + "blocks": [ + { + "bbox": [ + 120, + 289, + 181, + 349 + ], + "lines": [ + { + "bbox": [ + 120, + 289, + 181, + 349 + ], + "spans": [ + { + "bbox": [ + 120, + 289, + 181, + 349 + ], + "type": "image", + "image_path": "47c7f95d9301121bcac84846bb6d68da6fd5a8b65203569ebdbd597f52a94379.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 182, + 289, + 242, + 349 + ], + "blocks": [ + { + "bbox": [ + 182, + 289, + 242, + 349 + ], + "lines": [ + { + "bbox": [ + 182, + 289, + 242, + 349 + ], + "spans": [ + { + "bbox": [ + 182, + 289, + 242, + 349 + ], + "type": "image", + "image_path": "5ae7fc593bf28322649e45f1a4c47f570798fbb33748e01ef081fbddd4e8478a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 351, + 397, + 361 + ], + "lines": [ + { + "bbox": [ + 212, + 351, + 397, + 361 + ], + "spans": [ + { + "bbox": [ + 212, + 351, + 397, + 361 + ], + "type": "text", + "content": "Appearance Edit: black swan " + }, + { + "bbox": [ + 212, + 351, + 397, + 361 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 212, + 351, + 397, + 361 + ], + "type": "text", + "content": " white duck" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 244, + 289, + 304, + 349 + ], + "blocks": [ + { + "bbox": [ + 244, + 289, + 304, + 349 + ], + "lines": [ + { + "bbox": [ + 244, + 289, + 304, + 349 + ], + "spans": [ + { + "bbox": [ + 244, + 289, + 304, + 349 + ], + "type": "image", + "image_path": "d4eefc3ed9ede70715c6e624a9def7df13a20cfb7592d6260367d44fb5f702a9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 306, + 289, + 365, + 349 + ], + "blocks": [ + { + "bbox": [ + 306, + 289, + 365, + 349 + ], + "lines": [ + { + "bbox": [ + 306, + 289, + 365, + 349 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 365, + 349 + ], + "type": "image", + "image_path": "9eff71770f5449bb33a63adfe012fdcc60dc32ba03e811186894603972ab3fae.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 367, + 289, + 427, + 349 + ], + "blocks": [ + { + "bbox": [ + 367, + 289, + 427, + 349 + ], + "lines": [ + { + "bbox": [ + 367, + 289, + 427, + 349 + ], + "spans": [ + { + "bbox": [ + 367, + 289, + 427, + 349 + ], + "type": "image", + "image_path": "59547d8b164f4e1c8f7664d5413d54cff4f73d9395e5fc30ee2449a6c2e318de.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 429, + 289, + 489, + 349 + ], + "blocks": [ + { + "bbox": [ + 429, + 289, + 489, + 349 + ], + "lines": [ + { + "bbox": [ + 429, + 289, + 489, + 349 + ], + "spans": [ + { + "bbox": [ + 429, + 289, + 489, + 349 + ], + "type": "image", + "image_path": "fa07e97e6c01fece5f6d9569f5773fc3b1f020e3b2e5d9d86453c133bd96d6a5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 490, + 289, + 551, + 349 + ], + "blocks": [ + { + "bbox": [ + 490, + 289, + 551, + 349 + ], + "lines": [ + { + "bbox": [ + 490, + 289, + 551, + 349 + ], + "spans": [ + { + "bbox": [ + 490, + 289, + 551, + 349 + ], + "type": "image", + "image_path": "f03fc8cbf1ebb92c424e541918311addc84248eebf43b64b24328c5bc6d07e15.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 58, + 363, + 119, + 423 + ], + "blocks": [ + { + "bbox": [ + 58, + 363, + 119, + 423 + ], + "lines": [ + { + "bbox": [ + 58, + 363, + 119, + 423 + ], + "spans": [ + { + "bbox": [ + 58, + 363, + 119, + 423 + ], + "type": "image", + "image_path": "7fb1a935cbfc356e0c7af102a8ceabdd7583e04e2ed7e1e724b1c2de5d9d2b7e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 446, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 54, + 446, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 54, + 446, + 555, + 480 + ], + "type": "text", + "content": "Figure 1. Diverse video editing results of FADE. Our training-free approach, utilizing frequency-aware factorization and modulation, achieves high-fidelity, coherent edits across a variety of video types. FADE handles both appearance and motion adjustments with impressive robustness, ensuring precise alignment with input prompts and maintaining temporal consistency. Best viewed in color." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 120, + 363, + 181, + 423 + ], + "blocks": [ + { + "bbox": [ + 120, + 363, + 181, + 423 + ], + "lines": [ + { + "bbox": [ + 120, + 363, + 181, + 423 + ], + "spans": [ + { + "bbox": [ + 120, + 363, + 181, + 423 + ], + "type": "image", + "image_path": "971902c1377cd3ca873fa0f90cdd6442930b4fd1ad8e51d30949fcd2624e0834.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 182, + 363, + 242, + 423 + ], + "blocks": [ + { + "bbox": [ + 182, + 363, + 242, + 423 + ], + "lines": [ + { + "bbox": [ + 182, + 363, + 242, + 423 + ], + "spans": [ + { + "bbox": [ + 182, + 363, + 242, + 423 + ], + "type": "image", + "image_path": "391538d544dc45bd5d2a6c6dad3187bf0b988ab1bf5413402fb4daa60cce96f0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 244, + 363, + 304, + 422 + ], + "blocks": [ + { + "bbox": [ + 244, + 363, + 304, + 422 + ], + "lines": [ + { + "bbox": [ + 244, + 363, + 304, + 422 + ], + "spans": [ + { + "bbox": [ + 244, + 363, + 304, + 422 + ], + "type": "image", + "image_path": "3a9398fc8ecb7acd2b2d9427b72d645f1de3a11380f38c9a25ad88aa0408a891.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 306, + 363, + 365, + 422 + ], + "blocks": [ + { + "bbox": [ + 306, + 363, + 365, + 422 + ], + "lines": [ + { + "bbox": [ + 306, + 363, + 365, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 363, + 365, + 422 + ], + "type": "image", + "image_path": "c3501d1ea27decfb3d8cbb1ca18336ab0fc1f562e6cb1ece2651ebfe0bee9d22.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 367, + 363, + 427, + 422 + ], + "blocks": [ + { + "bbox": [ + 367, + 363, + 427, + 422 + ], + "lines": [ + { + "bbox": [ + 367, + 363, + 427, + 422 + ], + "spans": [ + { + "bbox": [ + 367, + 363, + 427, + 422 + ], + "type": "image", + "image_path": "a63ce0aeb98d19eb4ab8482ac965b5153f3dab3474a6b724c6516fcf074de616.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 429, + 363, + 489, + 423 + ], + "blocks": [ + { + "bbox": [ + 429, + 363, + 489, + 423 + ], + "lines": [ + { + "bbox": [ + 429, + 363, + 489, + 423 + ], + "spans": [ + { + "bbox": [ + 429, + 363, + 489, + 423 + ], + "type": "image", + "image_path": "0b8226ae7b20ff1cb84c46fc0c22c46ef2889992f617420ad2e0b94040e28459.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 490, + 363, + 551, + 423 + ], + "blocks": [ + { + "bbox": [ + 490, + 363, + 551, + 423 + ], + "lines": [ + { + "bbox": [ + 490, + 363, + 551, + 423 + ], + "spans": [ + { + "bbox": [ + 490, + 363, + 551, + 423 + ], + "type": "image", + "image_path": "b7ff6632fc4a9321f61f72a784c56390b9dbbce202bb21b0493977df17a459ff.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 206, + 425, + 400, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 425, + 400, + 437 + ], + "spans": [ + { + "bbox": [ + 206, + 425, + 400, + 437 + ], + "type": "text", + "content": "Motion Edit: swimming " + }, + { + "bbox": [ + 206, + 425, + 400, + 437 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 206, + 425, + 400, + 437 + ], + "type": "text", + "content": " flapping the wings" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 151, + 484, + 200, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 484, + 200, + 498 + ], + "spans": [ + { + "bbox": [ + 151, + 484, + 200, + 498 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 54, + 510, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 691 + ], + "type": "text", + "content": "Recent advancements in diffusion frameworks have significantly enhanced video editing, achieving high fidelity and strong alignment with textual prompts. However, conventional approaches using image diffusion models fall short in handling video dynamics, particularly for challenging temporal edits like motion adjustments. While current video diffusion models produce high-quality results, adapting them for efficient editing remains difficult due to the heavy computational demands that prevent the direct application of previous image editing techniques. To overcome these limitations, we introduce FADE—a training-free yet highly effective video editing approach that fully leverages the inherent priors from pre-trained video diffusion models via frequency-aware factorization. Rather than simply using these models, we first analyze the attenuated performance of the proposed system." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 486, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 641 + ], + "type": "text", + "content": "tion patterns within the video model to reveal how video priors are distributed across different components. Building on these insights, we propose a factorization strategy to optimize each component's specialized role. Furthermore, we devise spectrum-guided modulation to refine the sampling trajectory with frequency domain cues, preventing information leakage and supporting efficient, versatile edits while preserving the basic spatial and temporal structure. Extensive experiments on real-world videos demonstrate that our method consistently delivers high-quality, realistic and temporally coherent editing results both qualitatively and quantitatively. Code is available at https://github.com/EternalEvan/FADE." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 314, + 668, + 395, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 668, + 395, + 680 + ], + "spans": [ + { + "bbox": [ + 314, + 668, + 395, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "The objective of video editing is to modify specific visual content consistently across all frames while preserving the" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28426" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 294, + 243 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 243 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 243 + ], + "type": "image", + "image_path": "1ceada60a161a26322941d9a46db038f2bc1795e6a32a141b3bfe2a214153506.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 296, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 342 + ], + "type": "text", + "content": "Figure 2. (a) In typical T2V models, video tokens flattened along the temporal axis yield full-attention maps split into cross, spatial, and temporal attention. (b) Video priors in T2V models are factorized based on attention feature patterns. We visualize the attention map " + }, + { + "bbox": [ + 55, + 252, + 296, + 342 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 252, + 296, + 342 + ], + "type": "text", + "content": " and result " + }, + { + "bbox": [ + 55, + 252, + 296, + 342 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 252, + 296, + 342 + ], + "type": "text", + "content": " along with its spectrum for two blocks. The early block sketches the foundational structure and movement, while the later block refines details. This insight allows us to leverage each block's unique role for more efficient editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "type": "text", + "content": "integrity of unchanged regions. Recent achievements in diffusion models have greatly elevated video editing capabilities [10, 12, 15, 19, 20, 28, 32], bringing remarkable generative power to video editing tasks that require both spatial details and temporal coherence. While previous methods [1, 8, 13, 18, 30] have adapted text-to-image (T2I) diffusion models for prompt-driven image editing, their effectiveness diminishes when applied to video due to the complexities of managing spatial-temporal relationships across frames and a lack of generalized video prior knowledge. These limitations underscore the need for an efficient and robust editing strategy capable of handling the increased data and temporal coherence demands of videos. In response, we aim to design an accessible, high-quality editing solution that fully leverages the rich video priors embedded in pre-trained video diffusion models, enabling realistic, coherent, and versatile edits, including both appearance and motion adjustments, across complex video content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": "The main challenge in video editing is balancing changes to the target object while preserving unedited content. Previous methods largely rely on the generative power of T2I models for controlled edits. For instance, [15, 33] achieve this through null-text inversion [18], which minimizes inversion errors via test-time optimization. However, this process requires extensive iterations to compute a series of null-text embeddings for every sampling step, making it time-consuming and limiting its real-world applicability. Another line of work is attention feature injection [5, 15, 23, 27], where attention maps are modified by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "replacing or interleaving to maintain content consistency. Despite this, the absence of video priors often results in unrealistic temporal dynamics and limits motion edit capacity. Some methods, like [19], directly utilize video diffusion models by adding noise to the source video. This straightforward approach fails to fully exploit video priors, resulting in suboptimal editing quality. With the recent emergence of advanced text-to-video (T2V) models, new opportunities arise to harness their rich video knowledge. However, adapting previous T2I techniques to these models is challenging, given the extensive resources required, including memory, computation and time, to handle the increased data and complexity inherent in large-scale video networks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 232, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 567 + ], + "type": "text", + "content": "To overcome these challenges, we introduce FADE, a training-free yet potent approach for real-world video editing. Our goal is to harness the rich video priors about video appearance and dynamics within pre-trained T2V models to enable realistic and diverse editing. To achieve this, we start by examining the specific properties of these models. While previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps, we further observe that a similar frequency progression occurs block-wise within T2V models. We explore how video priors function in each block by analyzing their roles and patterns, revealing that early blocks sketch foundational spatial layouts and temporal dynamics while later blocks refine high-frequency details (Figure 2). Leveraging this insight, we factorize the editing problem by the distinct functions of each component, enhancing editing efficiency. Then, instead of performing optimization or attention injection, we design a spectrum-guided modulation strategy that utilizes the full-attention outputs in these sketching blocks to derive frequency-aware guidance through 3D Fourier transformation. This modulation adjusts the sampling trajectory to preserve low-frequency structure from the source video while allowing high-frequency details to support flexible and intricate edits. This cue from the frequency domain not only mitigates information leakage during guidance computation but also unleashes the potential of video diffusion priors, enabling sophisticated spatial and temporal edits with high fidelity and coherence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "We evaluate our method using extensive real-world video clips across various editing tasks. As shown in Figure 1, FADE exhibits proficiency in producing realistic and consistent edits, covering both appearance and motion modifications. Quantitative comparisons on video editing benchmark DAVIS [22] demonstrate our approach achieves superior textual alignment and temporal consistency, with 0.3561 CLIP score and 43.57 OSV. Moreover, we maintain competitive input fidelity with 21.54 MaskPSNR and 0.3297 LPIPS, preserving unedited regions effectively. Compared to previous methods, our use of general video priors with frequency-aware factorization deliv" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28427" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 549, + 238 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 549, + 238 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 549, + 238 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 549, + 238 + ], + "type": "image", + "image_path": "408f92fb59348413bcb1f228068ff6dc11626f7bf0d4d6e1cac88f616319973b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": "Figure 3. The overall framework of FADE. FADE is a training-free framework for video editing. Given an input video, we first perform video inversion to obtain noise " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": " and trajectory " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_t^*\\}_{t=0}^T" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": ". Starting from this noise, we use sketching blocks in the T2V model to produce full-attention results " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^*" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^*" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": ". These results are transformed to the frequency domain, where a low-pass filter isolates spatial and temporal structures. The frequency error then modulates the sampling process, enabling high-quality, consistent edits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "type": "text", + "content": "ers improved efficiency and a broader editing range, highlighting FADE's outstanding versatility and practicality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 148, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 148, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 148, + 355 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 365, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 663 + ], + "type": "text", + "content": "Video Diffusion Models. In recent years, denoising diffusion models [9, 29] have emerged as a leading class of generative models, notable for their synthesis quality and controllability. After training on large-scale datasets, text-to-image (T2I) diffusion models [6, 25] exhibit exceptional image synthesis capability. For video generation, some works [2, 11] adapt image diffusion models by inflating the denoising UNet [26] with temporal layers to handle video inputs. However, replicating the success of T2I models for video generation remains challenging due to the lack of large-scale video datasets and the increased data complexity. Most recently, Sora [4] and following studies [16, 35, 37] pioneer high-quality, high-resolution text-to-video (T2V) diffusion models. These models commonly use the diffusion transformer (DiT) architecture [21] to capture complex spatial and temporal coherence, outperforming previous UNet-based methods. However, this improved performance comes at the cost of high computational demands, mostly due to the extensive attention mechanism, which requires processing a large number of tokens. Limited research has explored the pattern within these models to reduce the computational intensity. To address this, we delve into their internal structure to analyze the roles of individual blocks, leveraging these insights to facilitate their applicability and efficiency in video editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Video Editing. Significant progress in image editing [3, 8, 17, 18, 20, 31] has driven advancements in video editing [7, 15, 19, 23, 34]. However, unique challenges in video editing arise from the need for consistent, coherent" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 302, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 578 + ], + "type": "text", + "content": " edits across frames, including both appearance and motion changes. Recent works [15, 23] adopt the attention feature injection techniques, manipulating various attention results to guide edits. They also employ model inflation to tune a T2I model with the input video following [34]. These approaches, however, are inefficient since they require a separate model for each sample and lack the general video priors to handle motion edits. In contrast, [19] utilizes a pre-trained video diffusion model for editing, beginning with corrupting the video by downsampling and noise. However, this corruption can result in global inconsistency with the source video and limits control over specific edits. Overall, there remains significant potential for advancing the use of video diffusion models in video editing. This application of the video diffusion model is nontrivial because (1) it requires maintaining global consistency in the generated video without relying on one-shot tuning or null-text optimization [18], and (2) applying previous attention feature injection methods is difficult with modern video models, as they involve dozens of transformer blocks with memory-intensive, full-attention computations. Our frequency-aware approach effectively addresses these challenges, setting our method apart from previous work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 594, + 370, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 370, + 607 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 370, + 607 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "In this section, we present FADE, a training-free framework that achieves real-world video editing with frequency-aware factorization of diffusion models. Our key idea is to investigate the inherent patterns in a pre-trained text-to-video diffusion model and activate each block's specific function for various video editing tasks. We will start by reviewing the background of video diffusion models, and then describe our designs of FADE, including how to factorize video dif" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28428" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "fusion models according to frequency cues and adjust the sampling trajectory via spectrum-guided modulation. The overall framework of our FADE is illustrated in Figure 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 115, + 258, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 115, + 258, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 258, + 127 + ], + "type": "text", + "content": "3.1. Preliminaries: Video Diffusion Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 182 + ], + "type": "text", + "content": "Video diffusion models are a class of generative models that reconstruct the distribution of video data by modeling the reverse trajectory of a diffusion process. Given a random noise " + }, + { + "bbox": [ + 55, + 133, + 295, + 182 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 133, + 295, + 182 + ], + "type": "text", + "content": ", the diffusion process is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 190, + 295, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 295, + 201 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t} = \\alpha_ {t} \\boldsymbol {z} _ {0} + \\sigma_ {t} \\epsilon , \\tag {1}", + "image_path": "966b51b93da30a59eb5fc6252eb9df271cad8ad37edc7d1ad6b40318199c665b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "text", + "content": " are the clean and noisy video, and " + }, + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\left\\{ (\\alpha_t, \\sigma_t) \\right\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 209, + 296, + 258 + ], + "type": "text", + "content": " denotes the noise schedule. With proper reparameterization, the training objective of diffusion models can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 264, + 295, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 264, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 97, + 264, + 295, + 279 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D M}} = \\mathbb {E} _ {\\boldsymbol {z} _ {0}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "41c7239231c1e449fc75d63b04cf35c621425aa60326d191dc4ef8625bb0f350.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "text", + "content": " is computed as Equation (1). The denoising model " + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "text", + "content": " is learned to predict the noise given the prompt " + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "text", + "content": ". Previous T2I models often implement " + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 285, + 296, + 357 + ], + "type": "text", + "content": " as UNet [26]. For video generation, recent models [4, 16, 35, 37] usually employ a series of cascaded transformer blocks to capture the spatial-temporal relationships more effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 357, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 381 + ], + "type": "text", + "content": "To generate a video in a small number of denoising steps, diffusion models usually use deterministic DDIM sampling:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 397, + 295, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 397, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 70, + 397, + 295, + 420 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t - 1} = \\frac {\\alpha_ {t - 1}}{\\alpha_ {t}} \\boldsymbol {z} _ {t} + \\left(\\sigma_ {t - 1} - \\frac {\\alpha_ {t - 1}}{\\alpha_ {t}} \\sigma_ {t}\\right) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right). \\tag {3}", + "image_path": "10829e6a84b2775d28d91c0d75d740678a79754104cf7a8341de9cb8b2738b0e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 426, + 295, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 451 + ], + "type": "text", + "content": "Conversely, to convert a video to its corresponding noise, we can apply DDIM inversion, the reverse of Equation (3):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 456, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 456, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 70, + 456, + 295, + 479 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t + 1} = \\frac {\\alpha_ {t + 1}}{\\alpha_ {t}} \\boldsymbol {z} _ {t} + \\left(\\sigma_ {t + 1} - \\frac {\\alpha_ {t + 1}}{\\alpha_ {t}} \\sigma_ {t}\\right) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right). \\tag {4}", + "image_path": "85c640d1440d242fb1f11ae47cfa7bddb7d5af2f5b2dc841b3910406614d3818.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "type": "text", + "content": "In practice, DDIM inversion introduces slight errors in every step, making it unreliable for fully reconstructing a real-world video or preserving fine details in the source video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 527, + 295, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 540 + ], + "type": "text", + "content": "3.2. Frequency-Aware Factorization of T2V Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Modern T2V models typically employ cascades of transformer blocks as the denoising model, with each block containing attention layers that process extensive video tokens. Given that dozens of identical blocks are involved in calculating the same form of attention, a key question arises: Do all these blocks serve the same purpose and contribute equally to video synthesis? To explore this, we analyze the behavior within each block of a pre-trained T2V model during video generation. Previous work [9] shows that diffusion models naturally emphasize different frequency components across sampling timesteps. Inspired by this, we adopt a frequency-aware perspective to uncover if similar trends occur at the block level in video models. Specifically, we capture and visualize each block's attention maps" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\{M_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "text", + "content": " and outputs " + }, + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "text", + "content": ", along with their Fourier spectra, over all " + }, + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 555, + 396 + ], + "type": "text", + "content": " blocks during sampling. As illustrated in Figure 2, a typical attention map in T2V models contains three types of attention regions: cross-attention for multimodal alignments, spatial attention for capturing intra-frame spatial structures, and temporal attention for tracking dynamics across frames. Our observation is that spatial and temporal attention are densely aligned along diagonal lines in the early blocks, creating multiple parallel diagonals in the attention maps. The main diagonal line suggests that these early blocks focus on outlining the intra-frame spatial structures, defining key shapes and boundaries of the composition. The secondary diagonal lines indicate that these blocks are simultaneously establishing temporal correspondences across frames within the same regions. Collectively, these early blocks construct the low-frequency, foundational structure of the video, setting up basic object placement and movement. We refer to them as sketching blocks. In contrast, the following blocks—which constitute the majority of the denoising model—display relatively sparse attention distribution, with scores more evenly spread across elements in the map. This implies that these later blocks focus on refining the high-frequency refinements, such as color, texture, and subtle vibration, providing nuanced adjustments for visual details. We call them sharpening blocks, as their primary role is to enhance details, enriching the visual appeal of the final output." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 624 + ], + "type": "text", + "content": "The variation in attention outputs of these blocks further supports our inference. The initial sketching blocks produce blurry outputs, with spectra focused on low frequencies that capture basic spatial and temporal layouts. In contrast, the sharpening blocks reveal clear foreground and background regions, involving much more high-frequency details. This distinction suggests that these two classes of blocks process features of differing levels and serve different purposes. We further verify that each block's function remains consistent throughout the sampling steps by analyzing these patterns across all steps. Thus, for video content editing—where maintaining underlying structures while adjusting finer details is essential—we can factorize each block's role, using sketching blocks for structural reconstruction and sharpening blocks for detailed generation. This division offers two key advantages: (1) it enables each block to perform optimally within its specialized function, and (2) it greatly reduces computational load by enabling frequency-aware guidance through selective use of blocks during sampling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 634, + 479, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 479, + 647 + ], + "type": "text", + "content": "3.3. Spectrum-Guided Modulation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "After factorizing the video model based on frequency-informed attention patterns, our next step is to activate each block's specific function and develop a robust sampling strategy that optimally balances reconstruction with generation. We begin by employing the inversion process de" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28429" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "scribed in Equation (4) with a pre-trained video diffusion model " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": ". This transforms the real-world video " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "z_0^*" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": " to an approximated noise " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "z_T^*" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": " and produces a reverse trajectory " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "\\{z_t^*\\}_{t=0}^T" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": ". However, as discussed in prior works, DDIM inversion can introduce non-negligible errors, making it difficult to accurately reconstruct the source video from " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "z_T^*" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": ". This issue is particularly pronounced in video inversion, where the predicted noise " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "z_T^*" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": " often deviates far from the standard Gaussian distribution. We, therefore, focus on refining the sampling process using the observed spectrum patterns to correct the inversion error and create an ideal trajectory " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "\\{z_t\\}_{t=0}^T" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": " to a plausible edited video " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": ". Following previous image editing approaches [18, 30], we adopt a dual branch strategy that leverages the inversion trajectory to enhance consistency with the source video. However, existing attention feature injection methods require extensive memory due to full-attention maps and limit editing flexibility through restrictive attention fusion techniques. While these methods can preserve the spatial layout, they often struggle with more complex edits involving pose, shape and motion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "text", + "content": "To overcome these limitations, we propose a more flexible, spectrum-guided approach that factorizes visual information in the frequency domain to separate content for preservation and editing. This approach offers greater freedom for a wider range of edits. As shown in Figure 3, our first design is to harness the sketching blocks, which care about the overall spatial layout and temporal movement. Rather than swapping or interleaving the attention maps directly, we gather the attention outputs of the source video " + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t^*" + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "text", + "content": " and the target video " + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "text", + "content": " within these blocks, denoted as " + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{F}_t^*" + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{F}_t" + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "text", + "content": ", and use their difference to derive a guidance term for the sampling process. Specifically, these attention outputs are computed using the same source prompt " + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}_{src}" + }, + { + "bbox": [ + 55, + 315, + 295, + 567 + ], + "type": "text", + "content": " through the full-attention layers. This guidance term provides a more robust and adaptive way to adjust the sampling trajectory without directly altering the intermediate features within the diffusion model. Furthermore, since these features are derived from the sketching blocks, the guidance focuses on the coarse spatial and temporal content, such as basic layouts and object movements, leaving ample room for detailed edits to be made in the sharpening blocks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "However, native guidance can lead to information leakage from the source video to the target video, as the attention outputs contain the full range of visual content. To address this, it is crucial to determine which aspects of the content should be retained and which should be modified. For this purpose, we turn to the frequency domain for assistance. Real-world images and videos often exhibit redundancy across their spatial and temporal dimensions, meaning that most of their energy is concentrated in the low-frequency spectrum, while the high-frequency components contribute much less. This characteristic underpins popular compression algorithms such as JPEG for images and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "MPEG for videos, and we aim to apply this principle here. Since we cannot get a clean video during the denoising process, we use attention outputs composed of tokens representing the noisy video. Unlike the discrete cosine transform (DCT) used in the spatial domain for MPEG, we employ the discrete Fourier transform on both spatial and temporal dimensions (3D DFT) to capture the spectral characteristics of intra-frame content and inter-frame motion:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 365, + 177, + 553, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 177, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 365, + 177, + 553, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {t} = \\mathrm {D F T} _ {\\mathrm {3 D}} \\left(\\boldsymbol {F} _ {t}\\right) \\in \\mathbb {R} ^ {h \\times w \\times \\tau \\times c}, \\tag {5}", + "image_path": "c771a8c64e456de08793159461bcaa363798af7eb21404af13d747e40a4970ba.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": " are the spatial and temporal size of the video tokens, and " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": " is the channel dimension. As we know, most of the energy in " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": " is concentrated in the low-frequency range. Thus, with the observed attention patterns, we can assume that fundamental spatial structures and temporal movements are encapsulated within this range, while the high-frequency components contain Gaussian noise and specific details like texture and subtle shaking—elements we wish to suppress or remove during editing. To achieve this, we apply a low-pass filter to isolate the low-frequency content of " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ". The same process is applied to " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t^*" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": " to compute the spectrum guidance " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 372, + 354, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 354, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 372, + 354, + 553, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {t} = \\left\\| \\mathrm {L P} \\left(\\mathcal {F} _ {t}\\right) - \\mathrm {L P} \\left(\\mathcal {F} _ {t} ^ {*}\\right) \\right\\| _ {2} ^ {2}, \\tag {6}", + "image_path": "ba22de50c7812167dd25c0abe2662a5f6b2d1a08ea55697e3d2be6086c07fcf9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "text", + "content": "where LP denotes the low-pass filter. This guidance measures the deviation of the target video " + }, + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "text", + "content": " from the source video in terms of basic structure and movement. This distance is then used to modulate the trajectory during DDIM sampling by the gradient of " + }, + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 377, + 554, + 438 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 448, + 553, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 448, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 326, + 448, + 553, + 461 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t - 1} = \\operatorname {D D I M} \\left(\\boldsymbol {\\epsilon} _ {\\theta}, \\boldsymbol {z} _ {t}, t, \\boldsymbol {y} _ {t g t}\\right) - \\lambda \\operatorname {N o r m} \\left(\\nabla_ {\\boldsymbol {z} _ {t}} \\mathcal {G} _ {t}\\right), \\tag {7}", + "image_path": "9f216a77fab14ea0b1af310f07279f3f8747253eb705167c58a2463468687c11.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 471, + 554, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 531 + ], + "type": "text", + "content": "where DDIM denotes the sampling process in Equation (3), " + }, + { + "bbox": [ + 313, + 471, + 554, + 531 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 471, + 554, + 531 + ], + "type": "text", + "content": " is the guidance weight and Norm normalizes the gradient to avoid numerical instability. This modulation strategy iteratively denoises the edited video towards the ideal direction, ensuring high consistency and realistic edits." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 539, + 503, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 503, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 503, + 552 + ], + "type": "text", + "content": "3.4. Versatile Video Editing Framework" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "We utilize a pre-trained T2V diffusion model for video inversion and sampling. This model leverages a 3D VAE, comprising an encoder " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": ", to enable spatial and temporal compression, facilitating a reversible transformation between pixel space and latent space. In addition to text embeddings, the T2V model incorporates FPS information to better handle video data. Unlike one-shot tuned video models from T2I models [15], using a well-trained video diffusion model ensures a more stable understanding of the complex spatial and temporal relationships within videos. Moreover, this eliminates the need for model tuning, enhancing efficiency and reducing the necessity of storing individual models for specific videos. For edit types," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28430" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 73, + 299, + 130 + ], + "blocks": [ + { + "bbox": [ + 59, + 76, + 70, + 123 + ], + "lines": [ + { + "bbox": [ + 59, + 76, + 70, + 123 + ], + "spans": [ + { + "bbox": [ + 59, + 76, + 70, + 123 + ], + "type": "text", + "content": "Input video" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 73, + 299, + 130 + ], + "lines": [ + { + "bbox": [ + 75, + 73, + 299, + 130 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 299, + 130 + ], + "type": "image", + "image_path": "4dbc2aeb531ffdc4ab3147667131a9d851c7f0a8b51f003cc40a8f4df8c69548.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 131, + 249, + 141 + ], + "lines": [ + { + "bbox": [ + 127, + 131, + 249, + 141 + ], + "spans": [ + { + "bbox": [ + 127, + 131, + 249, + 141 + ], + "type": "text", + "content": "Target: bushes " + }, + { + "bbox": [ + 127, + 131, + 249, + 141 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 127, + 131, + 249, + 141 + ], + "type": "text", + "content": " snowfield" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 144, + 299, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 144, + 299, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 144, + 299, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 144, + 299, + 201 + ], + "type": "image", + "image_path": "da7e528571eca048c37c8a75cf3a829de56378f63f479acabb3a0590ba816ba3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 202, + 299, + 258 + ], + "blocks": [ + { + "bbox": [ + 59, + 202, + 299, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 202, + 299, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 202, + 299, + 258 + ], + "type": "image", + "image_path": "c0d33bd845d0a8f78efc3a9699ec8eec5803970f6dd3a8b97e61522c8d924e78.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 259, + 299, + 317 + ], + "blocks": [ + { + "bbox": [ + 59, + 259, + 299, + 317 + ], + "lines": [ + { + "bbox": [ + 59, + 259, + 299, + 317 + ], + "spans": [ + { + "bbox": [ + 59, + 259, + 299, + 317 + ], + "type": "image", + "image_path": "40f2debd73c2cf9fc0e95ca8c21954e428b39df3dd23d21524d191285ffe679d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 325, + 299, + 382 + ], + "blocks": [ + { + "bbox": [ + 59, + 325, + 299, + 382 + ], + "lines": [ + { + "bbox": [ + 59, + 325, + 299, + 382 + ], + "spans": [ + { + "bbox": [ + 59, + 325, + 299, + 382 + ], + "type": "image", + "image_path": "fc2b030c8f9d36c14d41f7958f41bf6d61e56a538f0b50a7c98999aef5ee73dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 383, + 236, + 393 + ], + "lines": [ + { + "bbox": [ + 140, + 383, + 236, + 393 + ], + "spans": [ + { + "bbox": [ + 140, + 383, + 236, + 393 + ], + "type": "text", + "content": "Target: bear " + }, + { + "bbox": [ + 140, + 383, + 236, + 393 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 140, + 383, + 236, + 393 + ], + "type": "text", + "content": " panda" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 396, + 299, + 453 + ], + "blocks": [ + { + "bbox": [ + 59, + 396, + 299, + 453 + ], + "lines": [ + { + "bbox": [ + 59, + 396, + 299, + 453 + ], + "spans": [ + { + "bbox": [ + 59, + 396, + 299, + 453 + ], + "type": "image", + "image_path": "e4dc63f2b95716796af97e70af5865969eeaf017b793bc4ea717bf2e46f2e5f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 455, + 299, + 511 + ], + "blocks": [ + { + "bbox": [ + 59, + 455, + 299, + 511 + ], + "lines": [ + { + "bbox": [ + 59, + 455, + 299, + 511 + ], + "spans": [ + { + "bbox": [ + 59, + 455, + 299, + 511 + ], + "type": "image", + "image_path": "35f4bad7762172f262605bbbfaabb3f2bc1d0747169f811ae1f78c4b5a67ff51.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 512, + 299, + 568 + ], + "blocks": [ + { + "bbox": [ + 59, + 512, + 299, + 568 + ], + "lines": [ + { + "bbox": [ + 59, + 512, + 299, + 568 + ], + "spans": [ + { + "bbox": [ + 59, + 512, + 299, + 568 + ], + "type": "image", + "image_path": "7c662cb31fc0838798ed20c80a3dda2158dc82962562750868f8d593edc26a7b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 573, + 228, + 584 + ], + "lines": [ + { + "bbox": [ + 144, + 573, + 228, + 584 + ], + "spans": [ + { + "bbox": [ + 144, + 573, + 228, + 584 + ], + "type": "text", + "content": "(a) Appearance Edits" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 592, + 555, + 626 + ], + "lines": [ + { + "bbox": [ + 55, + 592, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 555, + 626 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on real-world videos. We evaluate both (a) appearance and (b) motion edits. Our framework consistently achieves accurate textual alignment across frames and high fidelity in unedited regions. Notably, compared to other methods based on video models, we produce more coherent and realistic motion edits. More results are available in the supplementary materials." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 73, + 551, + 129 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 323, + 123 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 323, + 123 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 323, + 123 + ], + "type": "text", + "content": "Input video" + } + ] + } + ], + "index": 12, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 73, + 551, + 129 + ], + "lines": [ + { + "bbox": [ + 328, + 73, + 551, + 129 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 551, + 129 + ], + "type": "image", + "image_path": "806885260d8a4837dfc356114ed573ae44499672db366f0a1124559f466222ea.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 131, + 531, + 141 + ], + "lines": [ + { + "bbox": [ + 351, + 131, + 531, + 141 + ], + "spans": [ + { + "bbox": [ + 351, + 131, + 531, + 141 + ], + "type": "text", + "content": "Target: turn the head " + }, + { + "bbox": [ + 351, + 131, + 531, + 141 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 351, + 131, + 531, + 141 + ], + "type": "text", + "content": " lower the head" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 312, + 144, + 551, + 200 + ], + "blocks": [ + { + "bbox": [ + 312, + 144, + 551, + 200 + ], + "lines": [ + { + "bbox": [ + 312, + 144, + 551, + 200 + ], + "spans": [ + { + "bbox": [ + 312, + 144, + 551, + 200 + ], + "type": "image", + "image_path": "0d126405e4c73934466888b3ea1d79e04d457f1d35b5166970973c4228292122.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 202, + 551, + 258 + ], + "blocks": [ + { + "bbox": [ + 312, + 202, + 551, + 258 + ], + "lines": [ + { + "bbox": [ + 312, + 202, + 551, + 258 + ], + "spans": [ + { + "bbox": [ + 312, + 202, + 551, + 258 + ], + "type": "image", + "image_path": "7947c3bcf39e7289c959346cd2d07e878b01a2db0ba9b76ef9f5b494b3b5291a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 312, + 259, + 551, + 317 + ], + "blocks": [ + { + "bbox": [ + 312, + 259, + 551, + 317 + ], + "lines": [ + { + "bbox": [ + 312, + 259, + 551, + 317 + ], + "spans": [ + { + "bbox": [ + 312, + 259, + 551, + 317 + ], + "type": "image", + "image_path": "192b1ff75e1e116beaed25a639b512cf8309f2916b51bdac0e4694299d7b9502.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 311, + 326, + 551, + 382 + ], + "blocks": [ + { + "bbox": [ + 311, + 326, + 551, + 382 + ], + "lines": [ + { + "bbox": [ + 311, + 326, + 551, + 382 + ], + "spans": [ + { + "bbox": [ + 311, + 326, + 551, + 382 + ], + "type": "image", + "image_path": "a276093bcbd25dd356bf4c6429d7ac1dfc2170be2c5b9b8062e5a46bb37a72f1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 383, + 528, + 393 + ], + "lines": [ + { + "bbox": [ + 354, + 383, + 528, + 393 + ], + "spans": [ + { + "bbox": [ + 354, + 383, + 528, + 393 + ], + "type": "text", + "content": "Target: close the eyes " + }, + { + "bbox": [ + 354, + 383, + 528, + 393 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 354, + 383, + 528, + 393 + ], + "type": "text", + "content": " open the eyes" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 311, + 396, + 551, + 453 + ], + "blocks": [ + { + "bbox": [ + 311, + 396, + 551, + 453 + ], + "lines": [ + { + "bbox": [ + 311, + 396, + 551, + 453 + ], + "spans": [ + { + "bbox": [ + 311, + 396, + 551, + 453 + ], + "type": "image", + "image_path": "1516098ad6e4c17b6c078b092c14a992770223264aa780b0030ddaea145934c7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 311, + 453, + 551, + 511 + ], + "blocks": [ + { + "bbox": [ + 311, + 453, + 551, + 511 + ], + "lines": [ + { + "bbox": [ + 311, + 453, + 551, + 511 + ], + "spans": [ + { + "bbox": [ + 311, + 453, + 551, + 511 + ], + "type": "image", + "image_path": "20abeeaeeacdfb9483f944557878d4ea6ed9bdf4ddbbee81887e1b553c0cc28d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 311, + 512, + 551, + 568 + ], + "blocks": [ + { + "bbox": [ + 311, + 512, + 551, + 568 + ], + "lines": [ + { + "bbox": [ + 311, + 512, + 551, + 568 + ], + "spans": [ + { + "bbox": [ + 311, + 512, + 551, + 568 + ], + "type": "image", + "image_path": "2252fc07fa7183872b460c2288fae4591475ed5bfe3fbd5f04995a9ff7c348ab.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 573, + 471, + 584 + ], + "lines": [ + { + "bbox": [ + 405, + 573, + 471, + 584 + ], + "spans": [ + { + "bbox": [ + 405, + 573, + 471, + 584 + ], + "type": "text", + "content": "(b) Motion Edits" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 647, + 294, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 294, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 294, + 706 + ], + "type": "text", + "content": "we broadly categorize prompt-driven modifications into appearance edits and motion edits. Appearance edits involve altering nouns, such as changing the main subject or background, focusing on the texture, color, and style (e.g., transforming a \"cat\" to a \"tiger\" in Figure 3). Motion edits, on" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 647, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 707 + ], + "type": "text", + "content": "the other hand, involve changing verbs, which can alter spatial and temporal dynamics. By applying spectrum-guided modulation to harness inherent video priors within the T2V model, our method achieves realistic motion edits, distinguishing it from most previous approaches." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28431" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 134, + 294, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on real-world video data. We compare our FADE framework to previous methods on both appearance and motion edits, measuring editing accuracy, input fidelity, temporal consistency, and human preference. FADE achieves competitive performance across diverse editing tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 134, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 134, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 294, + 281 + ], + "type": "table", + "html": "
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓PF ↑
appearance edits
Tune-A-Video [34]0.352219.860.462535.010.12
Video-P2P [15]0.358920.570.319934.710.15
FateZero [23]0.356220.650.305733.230.29
CogVideoX-V2V [35]0.375418.960.481131.450.09
FADE (Ours)0.376220.690.308531.360.35
motion edits
Tune-A-Video [34]0.328118.680.463735.850.10
Video-P2P [15]0.331418.930.388535.440.15
FateZero [23]0.325919.020.371234.470.13
CogVideoX-V2V [35]0.367818.170.492835.520.19
FADE (Ours)0.368319.260.369232.280.43
", + "image_path": "24da7d9a88027b5a22f02ae61f6cbf9f976ed24fe4a885e12388a71177e9a8fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 138, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 138, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 138, + 298 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 389 + ], + "type": "text", + "content": "To verify the effectiveness of our method, we conduct comprehensive experiments across various video editing tasks. We will start by describing the experimental settings and then present our main results. We will also provide detailed ablation studies and analyses of our approach to highlight the contributions of each component in our framework and to validate the impact of our design choices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 168, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 168, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 168, + 409 + ], + "type": "text", + "content": "4.1. Experiment Setups" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 569 + ], + "type": "text", + "content": "Implementation Details. To utilize video generation priors, we adopt a recent, widely used T2V model [35] for video inversion and generation. This model's denoising component employs the DiT architecture with 48 transformer blocks, each containing a full-attention layer. Based on observed attention map patterns, we designate the first 4 blocks as sketching blocks. We calculate the guidance over the interval " + }, + { + "bbox": [ + 55, + 415, + 296, + 569 + ], + "type": "inline_equation", + "content": "[0, 0.6T]" + }, + { + "bbox": [ + 55, + 415, + 296, + 569 + ], + "type": "text", + "content": " of the DDIM sampling steps, with total steps set to " + }, + { + "bbox": [ + 55, + 415, + 296, + 569 + ], + "type": "inline_equation", + "content": "T = 50" + }, + { + "bbox": [ + 55, + 415, + 296, + 569 + ], + "type": "text", + "content": ". To generate textual prompts from the source video, we use a multimodal language model [14] as the captioner. The guidance weight varies between 10 and 15, tailored to each editing task. We perform all evaluations on a single NVIDIA LS20 GPU with 48GB VRAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Datasets and Metrics. We evaluate our method using the DAVIS dataset [22] as well as various in-the-wild videos, following the approach of previous studies [15, 23]. Alongside qualitative results, we provide quantitative evaluations across multiple aspects using different metrics. Editing accuracy is assessed through CLIP score [24], which reflects the degree of textual alignment. To evaluate temporal consistency, a key factor for video coherence, we compute object semantic variance (OSV), following [15], to capture semantic consistency across frames. Additionally, we assess fidelity in unedited regions by calculating Mask PSNR and LPIPS [36], which quantify the preservation of content that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "should remain unchanged. Furthermore, given the importance of visual quality in video editing, we conduct a user study to gauge human preference across all editing tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 117, + 400, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 400, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 400, + 129 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 136, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 555, + 460 + ], + "type": "text", + "content": "We compare our editing method with state-of-the-art approaches, including Video-P2P [15], FateZero [23], TuneA-Video (TAV) [34] and CogVideoX-V2V (CogV2V) [35]. As illustrated in Figure 4, our method consistently yields superior results across all aspects when compared to previous methods. The editing results of Video-P2P and FateZero exhibit significant limitations, often failing to modify content accurately, resulting in noticeable textual mismatches. Similarly, while TAV and CogV2V leverage T2V models, they struggle with motion editing, producing subpar results in capturing dynamic changes. Leveraging the inherent priors from the T2V model and spectrum-guided modulation, our method achieves the best performance in preserving local structure, maintaining temporal consistency, and ensuring text-video alignment. We evaluate our method on 20 videos and measure 5 metrics for quantitative analysis. As shown in Table 1, our method achieves the highest CLIP and OSV scores, demonstrating its superiority in textual alignment and temporal consistency. We also obtain competitive Mask PSNR and LPIPS, indicating that our method preserves unedited regions effectively. Moreover, we achieve the highest human preference score, reflecting the favorable visual quality and appeal of our edits across tasks. Notably, our method operates without any optimization or fine-tuning, completing various edits on real-world videos in approximately 3 minutes, a significant efficiency gain compared to previous methods, which require over 15 minutes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 467, + 377, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 377, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 377, + 480 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "Efficient Editing via Sketching Blocks. We assess the function of different blocks in the T2V model by testing various block combinations to calculate modulation, holding other settings constant. We consider three configurations: (1) using sketching blocks only (ours), (2) symmetrically combining sketching and sharpening blocks (symm. blocks), and (3) using all blocks without factorization (w/o factor.). As shown in Table 2, configuration (1) provides comparable results to configurations (2) and (3), despite requiring fewer blocks for modulation. Although configurations (2) and (3) offer slightly improved fidelity, they significantly increase test time and memory demands. Figure 5 further demonstrates that relying on only the first four transformer blocks, or sketching blocks, achieves the best editing performance. Interestingly, contrary to common assumptions, including the sharpening blocks actually diminishes editing quality. We assume this is because the sharpening blocks primarily focus on elements like color, texture, and subtle vibrations, which often do not need" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28432" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 84, + 115, + 300 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 112, + 82 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 112, + 82 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 112, + 82 + ], + "type": "text", + "content": "Input video" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 84, + 115, + 300 + ], + "lines": [ + { + "bbox": [ + 60, + 84, + 115, + 300 + ], + "spans": [ + { + "bbox": [ + 60, + 84, + 115, + 300 + ], + "type": "image", + "image_path": "0b17bef75167506b7f24da7971274f502a76e0da9e419430faa7998ee33e0936.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 303, + 264, + 314 + ], + "lines": [ + { + "bbox": [ + 87, + 303, + 264, + 314 + ], + "spans": [ + { + "bbox": [ + 87, + 303, + 264, + 314 + ], + "type": "text", + "content": "Target: duck " + }, + { + "bbox": [ + 87, + 303, + 264, + 314 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 87, + 303, + 264, + 314 + ], + "type": "text", + "content": " yellow chick + open eyes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of the ablations. We find the variant frameworks fall short in terms of editing accuracy, temporal consistency and input fidelity, yielding unsatisfactory results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 118, + 84, + 174, + 300 + ], + "blocks": [ + { + "bbox": [ + 121, + 72, + 167, + 81 + ], + "lines": [ + { + "bbox": [ + 121, + 72, + 167, + 81 + ], + "spans": [ + { + "bbox": [ + 121, + 72, + 167, + 81 + ], + "type": "text", + "content": "w/o factor." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 118, + 84, + 174, + 300 + ], + "lines": [ + { + "bbox": [ + 118, + 84, + 174, + 300 + ], + "spans": [ + { + "bbox": [ + 118, + 84, + 174, + 300 + ], + "type": "image", + "image_path": "3ab7d17fbdb49e07c3edf267bff92809baec101f02192a9396581245a3088dfd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 177, + 84, + 233, + 300 + ], + "blocks": [ + { + "bbox": [ + 176, + 72, + 220, + 81 + ], + "lines": [ + { + "bbox": [ + 176, + 72, + 220, + 81 + ], + "spans": [ + { + "bbox": [ + 176, + 72, + 220, + 81 + ], + "type": "text", + "content": "w/o filter" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 84, + 233, + 300 + ], + "lines": [ + { + "bbox": [ + 177, + 84, + 233, + 300 + ], + "spans": [ + { + "bbox": [ + 177, + 84, + 233, + 300 + ], + "type": "image", + "image_path": "9b6d16e720d60da30d7ad732e5f4b20a1050b3307a311373b6bf9c602920cb49.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 235, + 84, + 291, + 300 + ], + "blocks": [ + { + "bbox": [ + 235, + 72, + 272, + 81 + ], + "lines": [ + { + "bbox": [ + 235, + 72, + 272, + 81 + ], + "spans": [ + { + "bbox": [ + 235, + 72, + 272, + 81 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 84, + 291, + 300 + ], + "lines": [ + { + "bbox": [ + 235, + 84, + 291, + 300 + ], + "spans": [ + { + "bbox": [ + 235, + 84, + 291, + 300 + ], + "type": "image", + "image_path": "615cb00bb2207ffa78aedbfa41e1e47aa973edd39f2ac863f82db0d602280004.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "type": "text", + "content": "to remain unchanged in most video editing tasks. Relying on these blocks for guidance may, therefore, misdirect the model, diverting its attention from building the foundational layout and movement of the input video, ultimately resulting in reduced performance. These results underscore the effectiveness of our frequency-aware factorization approach, demonstrating that our selective focus on sketching blocks provides an optimal balance between maintaining high-quality edits and achieving computational efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "Impact of Spectrum-Guided Modulation. In Figure 5 and Table 2, we examine our spectrum-guided modulation by removing the low-pass filter (w/o filter) and analyzing its effect on the editing performance. When we use the entire spectrum of attention outputs, high-frequency components are retained in the guidance term, causing the target object to retain too much of the source's original characteristics, which weakens the text-video alignment. Experiments show that using about two-thirds of the frequency components in the guidance term balances reconstruction and editing quality. The quantitative results further underscore the efficacy of spectrum-guided modulation, as shown by superior performance on key metrics like textual alignment and temporal consistency. By selectively modulating the frequency components, our method achieves precise edits that align well with the given textual prompt and maintain consistency across frames, both of which are crucial for high-quality video editing. These findings highlight that spectrum-guided modulation not only supports a stable structural foundation for the edits but also enhances flexibil" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 145, + 552, + 208 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": "Table 2. Ablations studies. We perform ablations to verify the effectiveness of the components in FADE and the impact of the model factorization. We find that FADE achieves similar performance with only a subset of sketching blocks, compared to using symmetric or all blocks. Moreover, spectrum-guided modulation further enhances both preservation and precision in edits." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 145, + 552, + 208 + ], + "lines": [ + { + "bbox": [ + 315, + 145, + 552, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 145, + 552, + 208 + ], + "type": "table", + "html": "
MethodCLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Test-time
symm. blocks0.365920.730.336732.615 min
w/o factor.0.369120.940.332832.0512 min
w/o filter0.361220.890.336432.283 min
FADE (Ours)0.372820.870.335231.773 min
", + "image_path": "6e29b377fd380ca94abf2795bfd230389479d3bcb557db53b11f011a6c22b3ea.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "content": "ity, allowing detailed adjustments for various editing tasks while ensuring input fidelity across the video." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 240, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 456 + ], + "type": "text", + "content": "Effective Exploitation of Video Diffusion Model. Our framework capitalizes on the rich video priors embedded in the pre-trained T2V model, enabling more effective utilization of the model's inherent knowledge of spatial layouts, temporal coherence, and object motions. The video diffusion model captures nuanced understanding across frames, making it an ideal foundation for edits that require both high fidelity and temporal consistency. To validate our superior exploitation of these priors, we compare FADE with CogVideoX-V2V [35], which uses the same underlying T2V model. As shown in Figure 4 and Table 1, FADE outperforms it both qualitatively and quantitatively across various editing tasks, demonstrating our framework's ability to harness the model's rich knowledge more effectively. Limitations. Despite FADE's effectiveness, its performance relies on the underlying video model and struggles with heavy occlusions requiring advanced temporal reasoning. We will address these in future work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we present FADE, a novel, training-free framework tailored for a diverse range of video editing tasks, encompassing both appearance and motion adjustments. Through frequency-aware factorization of the T2V model, FADE efficiently harnesses the model's inherent structural and temporal patterns, allowing each block to excel in its designated role. We further design the spectrum-guided modulation, which adjusts the sampling trajectory using frequency cues to preserve foundational video elements and enhance text-video alignment. Extensive experiments demonstrate our framework's competitive performance and generalization across various video editing challenges. We envision our work will inspire future research on versatile video editing and efficient video diffusion model. Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2024YFB4708100, and in part by the National Natural Science Foundation of China under Grant Grant 62321005, Grant 62306031" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28433" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 294, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 214 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 270 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 294, + 316 + ], + "type": "text", + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 317, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 348 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 294, + 384 + ], + "type": "text", + "content": "[7] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "text", + "content": "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 430, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 294, + 462 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 465, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 294, + 520 + ], + "type": "text", + "content": "[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 521, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 294, + 553 + ], + "type": "text", + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 587 + ], + "type": "text", + "content": "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[15] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In CVPR, pages 8599–8608, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[16] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[17] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 165, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 553, + 207 + ], + "type": "text", + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2, 3, 5, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 211, + 553, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 211, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 553, + 254 + ], + "type": "text", + "content": "[19] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 256, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 256, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 256, + 553, + 289 + ], + "type": "text", + "content": "[20] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, pages 1-11, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 291, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 312 + ], + "type": "text", + "content": "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 315, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 315, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 315, + 553, + 357 + ], + "type": "text", + "content": "[22] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "type": "text", + "content": "[23] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 462, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 462, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 317, + 462, + 553, + 506 + ], + "type": "text", + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 508, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 508, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 317, + 508, + 553, + 541 + ], + "type": "text", + "content": "[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 543, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 543, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 543, + 553, + 586 + ], + "type": "text", + "content": "[27] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. In ACML, pages 1215-1230. PMLR, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 589, + 553, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 553, + 642 + ], + "type": "text", + "content": "[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 646, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 667 + ], + "type": "text", + "content": "[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[30] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-rescale: Self-guidance mechanism for effective tuning-free real image editing. arXiv preprint arXiv:2409.01322, 2024. 2, 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28434" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 464 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[33] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[35] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3, 4, 7, 8, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 429 + ], + "type": "text", + "content": "[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. URL https://github.com/hpcaitech/Open-Sora, 2024.3,4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 464 + ], + "type": "text", + "content": "[38] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28435" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_content_list.json b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e520f1120af1dc60718eed17ee1366416d4813cd --- /dev/null +++ b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_content_list.json @@ -0,0 +1,1507 @@ +[ + { + "type": "text", + "text": "FALCON: Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding", + "text_level": 1, + "bbox": [ + 130, + 125, + 869, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thanh-Dat Truong $^{1}$ , Utsav Prabhu $^{2}$ , Bhiksha Raj $^{3,4}$ , Jackson Cothren $^{5}$ , Khoa Luu $^{1}$ $^{1}$ CVIU Lab, University of Arkansas, USA $^{2}$ Google DeepMind, USA \n $^{3}$ Carnegie Mellon University, USA $^{4}$ Mohammed bin Zayed University of AI, UAE \n $^{5}$ Dep. of Geosciences, University of Arkansas, USA", + "bbox": [ + 155, + 191, + 844, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tt032, jcothre, khoaluu}@uark.edu, bhiksha@cs.cmu.edu, utsavprabhu@google.com http://uark-cviu.github.io/projects/FALCON", + "bbox": [ + 153, + 266, + 844, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 319, + 326, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continual Learning in semantic scene segmentation aims to continually learn new unseen classes in dynamic environments while maintaining previously learned knowledge. Prior studies focused on modeling the catastrophic forgetting and background shift challenges in continual learning. However, fairness, another major challenge that causes unfair predictions leading to low performance among major and minor classes, still needs to be well addressed. In addition, prior methods have yet to model the unknown classes well, thus resulting in producing non-discriminative features among unknown classes. This work presents a novel Fairness Learning via Contrastive Attention Approach to continual learning in semantic scene understanding. In particular, we first introduce a new Fairness Contrastive Clustering loss to address the problems of catastrophic forgetting and fairness. Then, we propose an attention-based visual grammar approach to effectively model the background shift problem and unknown classes, producing better feature representations for different unknown classes. Through our experiments, our proposed approach achieves State-of-the-Art (SoTA) performance on different continual learning benchmarks, i.e., ADE20K, Cityscapes, and Pascal VOC. It promotes the fairness of the continual semantic segmentation model.", + "bbox": [ + 86, + 354, + 485, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The semantic segmentation networks, e.g., Transformers [69] and Convolutional Neural Networks [7], learned from data with a closed-set of known classes, have shown outstanding performance. However, they often suffer performance degradation when encountering novel objects or classes in new dynamic environments [2, 20, 62, 63]. To improve their performance, several transfer learning and domain adaptation methods [1, 28, 29, 40, 57-60, 62, 64] were introduced to adapt trained models into deployed environments. While", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c59bd327f32f50759195ae02cd7f7ed871e4b8b933cae3b50922ea91065c27b3.jpg", + "image_caption": [ + "Figure 1. Our Fairness Learning via Contrastive Attention to Continual Semantic Segmentation. The Fairness Contrastive Clustering Loss promotes the fairness of the model while the Attention-based Visual Grammar models the unknown classes." + ], + "image_footnote": [], + "bbox": [ + 516, + 316, + 906, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the former often aims to fine-tune the model on labeled data collected in the new environments, the latter adapts the model to the new domains in an unsupervised manner [1, 28]. However, these methods cannot handle novel objects well due to their close-set learning. In practice, the semantic segmentation models should be able to adaptively and continually learn the new knowledge of novel classes. It motivates the development of Continual Learning paradigm [2, 20, 46], a.k.a, Continual Semantic Segmentation (CSS), where the segmentation models are learned sequentially to new contents of data.", + "bbox": [ + 511, + 564, + 908, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Far apart from prior segmentation methods [7, 69] that learn one time on static, closed-set data, Continual Learning requires the segmentation models to learn from dynamic, open-set data [20, 62]. In a particular scenario, accessing previous learning data is restricted due to privacy concerns. In CSS, three challenges have been identified, including (1) Catastrophic Forgetting, (2) Background Shift, and (3) Fairness. While the catastrophic forgetting problem [24, 50, 56] depicts the segmentation model tends to forget its knowledge when learning new data, background shift indicates the problem of classes of previous or future data (unknown classes)", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7477b3592d42ef495fc1f16284bddfe5ac5f7b390885caa6d5959143cf357d37.jpg", + "image_caption": [ + "Figure 2. The Data Class Distribution of ADE20K. The major classes occupy more than $75\\%$ of the total pixels of the dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "have collapsed into a background class [2, 20, 75]. Prior methods [4, 20, 46, 62] addressed these two problems by introducing knowledge distillation and pseudo labels. However, these methods can not handle unknown classes since they either consider these unknown classes as a background class or assign unknown pixels by a pseudo label of prior known classes [20, 62]. More importantly, the last problem, fairness, is a significant challenge that limits the performance of CSS models (Fig. 1).", + "bbox": [ + 88, + 279, + 482, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the number of pixels of each class in training data have been imbalanced among classes and significantly decreased after each task. Thus, this bias influences the learning procedure and model predictions that later cause unfair class predictions. However, limited studies are taking the fairness problem into account. [61] presented a similar problem in domain adaptation and extended it to continual learning [62]. These methods rely on the assumption of ideal fair or balanced data distributions. However, it is not applicable in practice since the size, i.e., the number of pixels, of several classes can never be more significant than others. For example, the bottle size should not be more significant than the size of a car. Meanwhile, the current knowledge distillation methods [3, 20, 45] in CSS are unable to handle the fairness problem since they focus on modeling catastrophic forgetting and background shift problems. Therefore, it is essential to develop a new CSS approach to address these limitations.", + "bbox": [ + 91, + 415, + 483, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions of This Work. This work presents a novel Fairness Learning via Contrastive Attention Approach (FALCON) to Continual Semantic Segmentation (as shown in Fig. 1). Our contributions can be summarized as follows. First, we introduce a novel Contrastive Clustering Paradigm approach to Continual Learning that models the catastrophic forgetting problem. Second, by analyzing the limitation of vanilla Contrastive Clustering in biased data, we introduce a novel Fairness Contrastive Clustering loss to model the fairness problem in continual learning efficiently. Third, to effectively model the background shift problem, we introduce a new Attention-based Visual Grammar that model the topological structures of feature distribution to handle the unknown classes effectively. Finally, the ablation studies", + "bbox": [ + 89, + 688, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "illustrate the effectiveness of the proposed approach in different aspects of fairness promotion in CSS models. Compared with prior methods, our approach achieves SoTA performance on different settings of three standard benchmarks of CSS, including ADE20K, Pascal VOC, and Cityscapes.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 180, + 653, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Semantic Segmentation. Several studies were introduced to address catastrophic forgetting and background shift problems [19-21, 32, 37, 39, 43, 44]. The common CSS approach adopts knowledge distillation [20] and pseudo labels [2] to model catastrophic forgetting and background shift, respectively. Later, it was further improved by decoupling knowledge representations [75], modeling the inter- and intra-class knowledge [15, 46], distinguishing the feature representations of the future classes [4, 45, 67, 70], reducing background confusion [73, 74], using visual prompting [31, 35], or modeling distillation loss via the geodesic flow [54]. Another approach [3, 5, 25] adopted the mask-based segmentation networks [12, 13] to improve the performance. Recent studies have introduced CSS under the unsupervised domain adaptation settings [51, 52, 63, 66]. However, prior studies have yet to well model different unknown classes. Particularly, as previous methods [3, 20] use the pseudo labels to model the unknown classes, the future classes will be treated as a single background class. Then, Joseph et al. [30] improves the unknown class modeling using clustering but this method considers all different unknown classes as a single cluster, leading to the non-discriminative features among unknown classes.", + "bbox": [ + 511, + 205, + 906, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Learning (CL). CL is a common learning approach [9, 10, 42] to structure the deep feature representations in the deep latent space. Oorde et al. [42] first introduced the Noise-Contrastive Estimation (InfoNCE) learning framework. Then, Chen et al. [9] presented SimCLR, a self-supervised contrastive learning approach to improve the representation power of Residual Networks. He [27] proposed a Momentum Contrast framework for unsupervised representation learning. Later, it was further improved by using MLP projection head [10] and extended to improve the self-supervised training process of vision transformers [11]. Cui et al. [17] introduced a supervised parametric contrastive learning loss to address the long-tailed recognition. Li et al. [34] adopted contrastive learning to develop the one-stage online contrastive clustering method. Radford et al. [47] presents a contrastive framework to learn the vision-language model. Later, several methods also adopted this framework for vision-language pretraining [33, 47].", + "bbox": [ + 511, + 553, + 908, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imbalanced and Fairness Learning. The early methods utilized the balanced Softmax loss [49] to alleviate the impact of imbalanced data distribution. Later, Wang et al. [68] introduced a Seesaw loss to re-balance the contributions of positive and negative instances via the mitigation", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15066", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and compensation modules. Ziwei et al. [36] introduced a dynamic meta-embedding to model the imbalanced classification problem. Chu et al. [14] reduce the bias in the segmentation model by presenting a new stochastic training scheme. Szabo et al. [55] presented a tilted cross-entropy loss to promote class-relevant fairness. However, there are limited studies that address the fairness problem in CSS. Truong et al. [61] introduced a fairness domain adaptation approach to semantic segmentation and later extended it into continual learning setting [62]. However, these methods [61, 62] rely on the assumption of ideal balanced data, which could not be achieved by nature. To address the limitations in prior work, this paper will introduce a novel approach to effectively model the fairness problem and unknown classes in the continual learning setting.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed FALCON Approach", + "text_level": 1, + "bbox": [ + 89, + 330, + 405, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CSS aims to learn a segmentation network $F$ on sequence data $\\mathcal{D} = \\{\\mathcal{D}^1,\\dots,\\mathcal{D}^T\\}$ where $T$ is the number of learning steps. At learning step $t$ , the model $F$ encounters a dataset $\\mathcal{D}_t = \\{(\\mathbf{x}^t,\\hat{\\mathbf{y}}^t)\\}$ where $\\mathbf{x}^t\\in \\mathbb{R}^{H\\times W\\times 3}$ is the image and $\\mathbf{y}\\in \\mathbb{R}^{H\\times W}$ is a segmentation label of $\\mathbf{x}^t$ . The ground truths at learning step $t$ only consist of current classes $\\mathcal{C}^t$ , while the class labels of the previous $\\mathcal{C}^{1\\dots t - 1}$ and future steps $\\mathcal{C}^{t + 1\\dots T}$ are collapsed into a background class. Formally, learning the CSS model at step $t$ can be formed as Eqn. (1).", + "bbox": [ + 89, + 354, + 483, + 492 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} ^ {*} = \\arg \\min _ {\\theta_ {t}} \\mathbb {E} _ {\\mathbf {x} ^ {t}, \\hat {\\mathbf {y}} ^ {t} \\in \\mathcal {D} ^ {t}} \\left[ \\mathcal {L} _ {C E} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} ^ {t}\\right) + \\lambda_ {C L} \\mathcal {L} _ {C L} \\left(F (\\mathbf {x} ^ {t})\\right) \\right] (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 494, + 483, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where, $\\mathbf{y}^t = F(\\mathbf{x}^t,\\theta_t)$ , $\\theta_{t}$ is the parameter of $F$ at current learning step $t$ , $\\mathcal{L}_{CE}$ is the cross-entropy loss, $\\lambda_{CL}$ is the balanced weight. and $\\mathcal{L}_{CL}$ is the CSS objective. At learning step $t$ , the segmentation model $F$ is required to be able to predict both previously learned classes $\\mathcal{C}^{1\\dots t - 1}$ and current new classes $\\mathcal{C}^t$ . Under this learning scenario, three challenges have been identified, i.e., Catastrophic Forgetting, Background Shift, and Fairness. Several prior methods were presented to model the two first issues in CSS using knowledge distillation [3, 20]. The last issue has not been well addressed yet due to its challenges [62]. Prior methods [2, 3, 20, 53, 75] adopt knowledge distillation to design $\\mathcal{L}_{CL}$ . However, this method prevents the CSS model from diverging knowledge learned previously, therefore resulting in limiting the ability to adopt new knowledge [62]. In addition, these methods have not addressed fairness and background shift problems due to their dedicated design for maintaining knowledge via distillation [3, 20, 45]. Therefore, to address these problems, we introduce a novel Fairness Learning via Contrastive Attention Approach to CSS.", + "bbox": [ + 89, + 521, + 483, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Continual Learning via Contrastive Clustering", + "text_level": 1, + "bbox": [ + 89, + 833, + 483, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Apart from prior methods [3, 20, 45], our CSS (Fig. 3) is defined as Contrastive Clustering Learning. Given a set of centroid vectors $\\{\\mathbf{c}_i\\}_{i=1}^{N_K + N_U}$ where $N_K = |\\mathcal{C}^{1..t}|$ and $N_U$", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is the number of known and unknown classes up to current learning tasks. Prior work [3, 30, 62] often defined the number of unknown classes as 1 where background classes are considered as a single unknown class. Formally, our Contrastive Clustering Learning for CSS can be defined as in Eqn. (2).", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {C L} \\left(F (\\mathbf {x} ^ {t})\\right) = \\sum_ {\\mathbf {c} _ {i}} \\mathcal {L} _ {C o n t} (\\mathbf {F} ^ {t}, \\mathbf {c} _ {i}) \\\\ = \\sum_ {\\mathbf {c} _ {i}} \\sum_ {h, w} - \\phi \\left(\\mathbf {f} _ {h, w} ^ {t}, \\mathbf {c} _ {i}\\right) \\log \\frac {\\exp \\left(\\mathbf {f} _ {h , w} ^ {t} \\times \\mathbf {c} _ {i}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c} _ {i}\\right)} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 188, + 905, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{F}^t\\in \\mathbb{R}^{H\\times W\\times D}$ is the feature maps extracted from the input image $\\mathbf{x}^t$ by the segmentation network $F$ , $\\mathbf{f}_{h,w}^t\\in \\mathbb{R}^D$ is the feature at the pixel location $(h,w)$ of features $\\mathbf{F}^t$ , $\\sum_{\\mathbf{f}'}$ means the summation over all feature representations $\\mathbf{f}'\\in \\mathbb{R}^D$ , and $\\phi :\\mathbb{R}^D\\times \\mathbb{R}^D\\to [0,1]$ is the function that determines either $\\mathbf{f}_{h,w}^t$ belongs to the cluster $\\mathbf{c}_i$ or not.", + "bbox": [ + 511, + 256, + 906, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By defining CSS as contrastive clustering learning, the knowledge of the segmentation model has been well maintained via the cluster vectors $\\mathbf{c}$ . Then, minimizing Eqn. (2) will separate the representations of different classes while gathering the features of the same class into the same cluster. As the cluster vectors $\\mathbf{c}$ of the old classes $\\mathcal{C}^{1..t-1}$ have been well learned to represent their knowledge, these vectors are frozen at learning step $t$ to maintain the knowledge representations of previous classes to address the catastrophic forgetting problem. To effectively learn cluster vectors $\\mathbf{c}$ , the cluster vector $\\mathbf{c}$ will periodically updated after each $K$ steps by the momentum update [27, 30, 62] based on the features $\\mathbf{f}_{h,w}^t$ assigned to cluster $\\mathbf{c}$ . However, there are two major problems in contrastive clustering learning. First, since the training data suffer the bias among classes as shown in Fig. 2, this bias will influence Eqn. (2) and cause the unfair predictions. Second, as the function $\\phi$ requires the labels to determine the features belonging to clusters, it limits the ability to model the unknown classes where the labels are unavailable. Therefore, Secs. 3.2-3.3 will present a novel approach to tackle these problems.", + "bbox": [ + 511, + 351, + 908, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Fairness Contrastive Clustering Learning", + "text_level": 1, + "bbox": [ + 511, + 676, + 867, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While contrastive clustering learning in Eqn. (2) promotes the compact representations of features around their clusters, inspired by [8, 17, 77], we observe that the imbalanced class distribution will influence unfair behaviors among classes. In particular, for simplicity, we consider $\\{\\mathbf{f}_i^t\\}_{i=1}^L$ is the set of features that belong to the cluster $\\mathbf{c}$ at learning step $t$ (i.e., $\\phi(\\mathbf{f}_i^t, \\mathbf{c}) = 1$ ) and $L$ is the number of features (in this case, $L$ is the total number of pixels belong to the class of cluster $\\mathbf{c}$ ). Let us define the enforcement between the feature $\\mathbf{f}_t^t$ and the cluster $\\mathbf{c}$ as $\\ell_i = \\frac{\\exp(\\mathbf{f}_i^t \\times \\mathbf{c})}{\\sum_{\\mathbf{f}' \\in L} \\exp(\\mathbf{f}' \\times \\mathbf{c})}$ . Hence, the lower the value of the enforcement $\\ell_i$ is, the more compact the representation of visual features and clusters is. Then, the contrastive clustering learning loss in Eqn. (2) of entire", + "bbox": [ + 511, + 699, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c942f7a1568715b6935ae538020f155f335c8fb9a5ba1c3e0c5aa88913f00cd3.jpg", + "image_caption": [ + "Figure 3. The Proposed Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding." + ], + "image_footnote": [], + "bbox": [ + 171, + 87, + 823, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cluster $\\mathbf{c}$ can be defined as in Eqn. (3).", + "bbox": [ + 89, + 262, + 346, + 276 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {C o n t}} (;;, \\mathbf {c}) = - \\sum_ {i = 1} ^ {L} \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {t} \\times \\mathbf {c}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} = - \\sum_ {i = 1} ^ {L} \\log \\ell_ {i} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 286, + 483, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1. If the contrastive clustering loss $\\mathcal{L}_{Cont}(:,\\mathbf{c})$ achieves the optimal value, the enforcement $\\ell_{i}$ between the feature and the cluster will converge to $\\ell_{i} = L^{-1}$ .", + "bbox": [ + 88, + 334, + 482, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1 has implied that the class with more samples will result in a lower value of the enforcement and produce a more compact representation while the class having fewer samples will be more scattered in the feature space due to the higher value of the enforcement. In particular, let $L_{major}$ and $L_{minor}$ be the number of samples of the major and minor class where $L_{major} >> L_{minor}$ . Then, based on Proposition 1, the enforcement between features and the cluster of the major class will be significantly lower than the one of the minor class, i.e., $L_{major}^{-1} << L_{minor}^{-1}$ . Therefore, a direct adoption of the contrastive clustering loss in Eqn. (2) will result in an unfair CSS model. In addition, for classes in the minority group, the lower value of $L$ results in the feature presentations of classes being far away from their clusters. Thus, the model will produce non-discriminative features compared to the ones in the majority group. Moreover, if the loss is applied to the cases of unknown labels, these feature representations can be scattered in the latent space and pulled into the incorrect clusters due to weak enforcement between features and clusters (Fig. 4).", + "bbox": [ + 91, + 380, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the unfair problem in contrastive clustering learning, inspired by [8, 17, 77], we introduce a scaling factor $\\alpha$ and a learnable transition vector $\\mathbf{v}$ for each cluster $\\mathbf{c}$ (all clusters have the same value of $\\alpha$ but different vector $\\mathbf{v}$ ). Our Fairness Contrastive Clustering Learning Loss for the entire cluster in Eqn. (3) can be re-formed as follows,", + "bbox": [ + 89, + 681, + 483, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {C o n t}} ^ {\\alpha} (;, \\mathbf {c}) = - \\alpha \\sum_ {i = 1} ^ {L} \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {t} \\times \\mathbf {c}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} - \\log \\frac {\\exp (\\mathbf {v} \\times \\mathbf {c})}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 779, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, the scaling factor $\\alpha$ will help to re-scale the impact of the enforcement in learning, and the transitive vector $\\mathbf{v}$ assists in translating the center cluster into the proper position of the latent space. This action promotes the compactness of clusters in the minority group.", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 2. If the fairness contrastive clustering loss $\\mathcal{L}_{Cont}^{\\alpha}(:,\\mathbf{c})$ achieves the optimal value, the enforcement $\\ell_{i}$ between the feature and the cluster will converge to $\\ell_{i} = (\\alpha^{-1} + L)^{-1}$ . Proofs of Propositions 1-2 are in the appendix.", + "bbox": [ + 511, + 262, + 903, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under the Proposition 2, when the value of $\\alpha$ is small, the divergence of the enforcement between major and minor classes will be smaller, i.e., $||(\\alpha^{-1} + L_{major})^{-1} - (\\alpha^{-1} + L_{minor})^{-1}|| < ||L_{major}^{-1} - L_{minor}^{-1}||$ . Fig. 4 has illustrated the impact of fairness contrastive clustering loss. Therefore, our designed proposed fairness contrastive loss has effectively addressed the fairness issue in Eqn. (2). It should be noted that although the smaller $\\alpha$ results in the fairer enforcement varied from major to minor classes. However, if the value of scaling factor $\\alpha$ is too small, the contrastive clustering loss will rely more on the enforcement of the transitive vector $\\mathbf{v}$ , and the distribution of features $\\mathbf{f}_i^t$ around its cluster $\\mathbf{c}$ will be scattered due the weak enforcement caused by small $\\alpha$ . Therefore, the value of scaling factor $\\alpha$ in practice should be carefully selected.", + "bbox": [ + 511, + 321, + 906, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. An Efficient Unknown Class Modeling", + "text_level": 1, + "bbox": [ + 511, + 556, + 843, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An ideal CSS approach must be able to model the unknown classes without supervision, especially in open-set contexts", + "bbox": [ + 511, + 579, + 903, + 609 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd774369ab042b69c737d32bebdef9bed92f826bef9d7b9b940de6ea8260a0d3.jpg", + "image_caption": [ + "Figure 4. The Enforcement Loss of Contrastive Clustering $\\mathcal{L}_{Cont}$ and Fairness Contrastive Clustering $\\mathcal{L}_{Cont}^{\\alpha}$ on Pascal VOC. Since $\\mathcal{L}_{Cont}$ suffers severe biased, its clusters of minor classes remain scattered. Our $\\mathcal{L}_{Cont}^{\\alpha}$ produces a more uniform loss among classes, promotes fairness and compactness of clusters." + ], + "image_footnote": [], + "bbox": [ + 516, + 616, + 903, + 821 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[30, 62] where there could be multiple unknown classes or objects. Prior studies have adopted the pseudo-label strategies [3, 20] based on the model predictions to assign labels for seen classes while unseen classes have been ignored, thus resulting in non-discriminative features. [30, 62] improved the background modeling by using an additional prototypical representation for unknown classes. However, these approaches consider different unknown classes as one (i.e., $N_U = 1$ ) resulting in non-distinguished representations of different unknown classes. Thus, modeling function $\\phi$ in Eqn. (2) without supervision of different unknown classes (i.e., $N_U > 1$ ) is challenging.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although modeling $\\phi$ to determine the single feature $\\mathbf{f}$ belonging to the cluster $\\mathbf{c}$ is challenging, prior studies in clustering [41, 71, 72] have suggested that determine a set of features $\\{\\mathbf{f}_i^t\\}_{i = 1}^M$ belonging to cluster $\\mathbf{c}$ should be easier. This derives from the fact that even though the feature representations of different classes are different, the distributions of features around its cluster (termed as Visual Grammar) in the feature space should be similar among classes or clusters. As a result, by learning the distribution of features and their clusters, the model $\\phi$ can determine whether a feature belongs to a cluster. Then, by learning the model $\\phi$ on prior known clusters and features, the knowledge of $\\phi$ can be adaptively applied to unknown clusters. Fig. 5 illustrates our visual grammar model of the cluster distributions.", + "bbox": [ + 89, + 276, + 483, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations of Prior Clustering Methods. The traditional methods in clustering, e.g., KNN or density-based clustering [22], remain limited to noisy features leading to producing the incorrect cluster assignment. Meanwhile, the modern clustering methods, e.g., Graph Neural Networks (GNNs) [71, 72], require a large memory to build the affinity graph for clusters. In addition, GNNs often learn the local structures of graphs (or clusters) and accumulate them via the aggregation layers. Hence, the global structures of the clusters, i.e., visual grammar, are not well modeled by GNNs [41]. Therefore, to address these limitations, we introduced a new Attention-based Visual Grammar approach to efficiently model the distribution of features and their clusters via the self-attention mechanism [65].", + "bbox": [ + 89, + 492, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1: Given a center $\\mathbf{c}$ and a set of $M$ features $\\{\\mathbf{f}_i^{\\mathbf{c}}\\}_{i = 1}^{M}$ where $\\mathbf{f}_i^{\\mathbf{c}}$ denotes the feature $\\mathbf{f}_i$ belonging to the cluster $\\mathbf{c}$ , and $\\forall i\\in [1..M - 1]:\\cos (\\mathbf{f}_i^{\\mathbf{c}},\\mathbf{c})\\geq \\cos (\\mathbf{f}_{i + 1}^{\\mathbf{c}},\\mathbf{c})$ the Visual Grammar of the cluster c parameterized by $\\Theta$ can be defined", + "bbox": [ + 89, + 707, + 483, + 768 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6384771357077251daa8b9897a43d1e44dd7c7be51c270b49d96b82029fcafe5.jpg", + "image_caption": [ + "Figure 5. The Proposed Visual Grammar Model." + ], + "image_footnote": [], + "bbox": [ + 91, + 784, + 483, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as in Eqn. (5).", + "bbox": [ + 513, + 90, + 611, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\{\\mathbf {f} _ {i, j} ^ {\\mathbf {c}} \\} ^ {M}} [ - \\log p (\\mathbf {f} _ {1} ^ {\\mathbf {c}}, \\mathbf {f} _ {2} ^ {\\mathbf {c}}, \\dots , \\mathbf {f} _ {M} ^ {\\mathbf {c}}, \\mathbf {c}, \\Theta) ] \\\\ \\Leftrightarrow \\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\left\\{\\mathbf {f} _ {i} ^ {\\mathbf {c}} \\right\\} _ {i = 1} ^ {M}} [ - \\log p \\left(\\boldsymbol {\\Delta} _ {1} ^ {\\mathbf {c}}, \\boldsymbol {\\Delta} _ {2} ^ {\\mathbf {c}}, \\dots , \\boldsymbol {\\Delta} _ {M} ^ {\\mathbf {c}}, \\mathbf {c}, \\Theta\\right) ] \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 112, + 903, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta_i^{\\mathbf{c}} = \\mathbf{f}_i^{\\mathbf{c}} - \\mathbf{c}$ . Eqn. (5) defines the visual grammar of the cluster by modeling the feature distribution of $\\mathbf{f}_i^{\\mathbf{c}}$ and its cluster center $\\mathbf{c}$ . Let $\\phi : \\mathbb{R}^{(M + 1)\\times D} \\to [0,1]^M$ be a function receiving a center $\\mathbf{c}$ and a set of $M$ features $\\{\\mathbf{f}_i\\}_{i = 1}^M$ ( $\\cos(\\mathbf{f}_i,\\mathbf{c}) \\geq \\cos(\\mathbf{f}_{i + 1},\\mathbf{c})$ ) to determine whether $\\mathbf{f}_i$ belonging to $\\mathbf{c}$ , i.e., $\\mathbf{u} = \\phi(\\Delta_1,\\Delta_2,\\dots,\\Delta_M,\\mathbf{c})$ where $\\Delta_i = \\mathbf{f}_i - \\mathbf{c}$ , $\\mathbf{u} = [u_1,u_2,\\dots,u_M]$ and $u_i = 1$ denotes $\\mathbf{f}_i$ belong to cluster $\\mathbf{c}$ and vice versa. Hence, the visual grammar model in Eqn. (5) can be modeled by the network $\\phi$ as follows with parameter $\\Theta$ as in Eqn. (6).", + "bbox": [ + 511, + 157, + 906, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\left\\{\\mathbf {f} _ {i} \\right\\} _ {i = 1} ^ {M}} \\left[ - \\log p (\\mathbf {u} | \\boldsymbol {\\Delta} _ {1}, \\boldsymbol {\\Delta} _ {2}, \\dots , \\boldsymbol {\\Delta} _ {M}, \\mathbf {c}, \\Theta) \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 316, + 906, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eqn. (6) aims to model the distribution of features around its cluster by learning the correlation of relatively topological structures $\\pmb{\\Delta}_i$ of features $\\mathbf{f}_i$ around cluster $\\mathbf{c}$ . Then, based on knowledge of the cluster distribution, the model $\\phi$ is able to determine whether a feature $\\mathbf{f}_i$ belongs to cluster $\\mathbf{c}$ . Hence, it is essential that the model $\\phi$ has the ability to exploit the correlation between features $\\mathbf{f}_i$ and cluster $\\mathbf{c}$ to learn the topological structure of visual grammar. Therefore, we adopt the self-attention mechanism [41, 65] to efficiently model these feature correlations. Particularly, the model $\\phi$ is formed by $L_{\\phi}$ blocks of self-attention as follows,", + "bbox": [ + 511, + 342, + 906, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {z} _ {0} = \\mathrm {L N} ([ \\boldsymbol {\\Delta} _ {1}, \\dots , \\boldsymbol {\\Delta} _ {\\mathrm {M}}, \\mathbf {c} ]) + \\boldsymbol {\\beta}, \\quad \\mathbf {a} _ {l} = \\mathbf {z} _ {l} + \\operatorname {M H S A} (\\mathbf {z}) \\\\ \\mathbf {z} _ {l + 1} = \\mathbf {a} _ {l} + \\operatorname {M L P} \\left(\\ln \\left(\\mathbf {a} _ {l}\\right)\\right), \\quad \\mathbf {u} = \\operatorname {P r o j} \\left(\\mathbf {z} _ {L _ {\\phi}}\\right) \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 515, + 905, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is the positional embedding, LN is Layer Normalization, MHSA is multi-head self-attention, MLP is the multi-layer perception, and Proj is the linear projection. By using Transformers, the correlation of cluster distributions can be well modeled by the self-attention mechanism.", + "bbox": [ + 511, + 551, + 906, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cluster Assignment via Visual Grammar. Instead of assigning the clusters based on the model prediction [3, 20, 45] or nearest cluster [30, 62] that are less effective, the cluster assignment in our approach will be performed by the visual grammar model, i.e., the visual grammar model will consider the $M$ closest features around cluster $\\mathbf{c}$ to assign the cluster for these features. Then, the cluster assignments are used to compute our Fairness Contrastive Clustering loss. In addition, following common practices [3, 20, 45], we improve background shift modeling by using the cluster assignments of features as the pseudo labels of pixels.", + "bbox": [ + 511, + 628, + 908, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unknown Cluster Initialization. Prior work [30, 62] initialized a single unknown cluster $(N_U = 1)$ , thus resulting in producing non-discriminative class-wise features. However, there should be more than a single unknown cluster $(N_U > 1)$ to produce discriminative features for different unknown classes. Therefore, our approach first initializes a list of potential unknown clusters at each learning step via", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DB-SCAN [22] on the features of unknown classes extracted by the current CSS model. For the new known class $\\mathcal{C}^t$ , we initialize these clusters based on the mean of their feature representations. Meanwhile, the clusters of known classes learned in previous steps are maintained. Further details can be found in the appendix.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Continual Learning Procedure", + "text_level": 1, + "bbox": [ + 89, + 190, + 362, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the training procedure of our continual learning approach. At each learning step $t$ , the CSS model $F$ with $\\theta_{t}$ is trained with the Fairness Contrastive Clustering loss defined in Eqn. (4) and the previous visual grammar model $\\phi$ with $\\Theta_{t-1}$ . In addition, we introduce a cluster regularizer $\\mathcal{R}_{C}$ to avoid the clusters of different classes collapsing into a single cluster. Therefore, the entire CSS learning objective in our approach can be formed as in Eqn. (8).", + "bbox": [ + 89, + 212, + 483, + 333 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\boldsymbol {\\theta} _ {t}} \\mathbb {E} _ {\\mathbf {x} ^ {t}, \\hat {\\mathbf {y}} ^ {t}} \\left[ \\mathcal {L} _ {C E} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} ^ {t}\\right) + \\lambda_ {C L} \\sum_ {\\mathbf {c} _ {i}} \\mathcal {L} _ {C o n t} ^ {\\alpha} \\left(\\mathbf {F} ^ {t}, \\mathbf {c} _ {i}\\right) + \\lambda_ {C} \\mathcal {R} _ {C} (\\mathbf {c}) \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 342, + 485, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_C(\\mathbf{c}) = \\sum_{\\mathbf{c}_i,\\mathbf{c}_j}\\{\\max (0,2\\nabla -||\\mathbf{c}_i - \\mathbf{c}_j||)\\} ^2$ is the regularizer to avoid the cluster collapsing, $\\lambda_C$ is the balanced weight, and $\\nabla$ is the margin between clusters.", + "bbox": [ + 89, + 372, + 482, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Procedure of Visual Grammar Model. At CSS learning step $t$ , we adopt the visual grammar model trained on the previous learning step, i.e., $\\phi$ with $\\Theta_{t-1}$ , to perform the cluster assignment for the contrastive clustering loss defined in Eqn. (2). Then, the visual grammar model $\\phi$ with $\\Theta_t$ at learning step $t$ will be learned (initialized from $\\Theta_{t-1}$ ) on the features extracted from the dataset and the set of known clusters $\\mathbf{c}$ up to the current learning step. Following [41], we sample a center $\\mathbf{c}$ from the known clusters and its $M$ closest features to train the visual grammar model.", + "bbox": [ + 89, + 417, + 483, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initial Visual Grammar Model. At the first learning step $t = 1$ , since no clusters have been learned initially, the visual grammar model $\\phi$ with $\\Theta_0$ is unavailable. However, as common practices in CSS [3, 4, 20], the segmentation model is typically trained from a pre-trained backbone on ImageNet [18]. As a result, the features extracted at the first learning step are characterized by the ImageNet features. Therefore, we adopt this philosophy to initialize our visual grammar model $(\\phi$ with $\\Theta_0$ ) by pre-training the visual grammar model on the ImageNet dataset. Then, during CSS training, we will progressively train our visual grammar model at each learning step as aforementioned.", + "bbox": [ + 89, + 569, + 483, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 763, + 223, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementations and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 89, + 787, + 454, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. Following common practices [2, 20, 62], we adopt DeepLab-V3 [6] with ResNet-101 [26] and SegFormer [69] with MiT-B3 [69] in our experiments. For the Visual Grammar model, we adopt the design of [41] with $L_{\\phi} = 12$ blocks of multi-head self-attention layers. The feature vectors from the last layer of the decoder are used", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/80a7d749960cd6df7605a550fa69e8763d55afffefb39c8f247ad1c6e8cb82e1.jpg", + "table_caption": [ + "Table 1. Effectiveness of Fairness Contrastive Learning Loss." + ], + "table_footnote": [], + "table_body": "
(a) ADE20K 100-50
\\( \\mathcal{L}_{Cont} \\)\\( \\mathcal{L}_{Cont}^{\\alpha} \\)0-100100-150allMajor Minor
✓44.615.234.851.5 26.4
✓44.624.537.952.1 30.8
(b) ADE20K 100-10
\\( \\mathcal{L}_{Cont} \\)\\( \\mathcal{L}_{Cont}^{\\alpha} \\)0-100100-150allMajor Minor
✓41.916.033.249.9 24.9
✓44.420.436.451.8 28.7
", + "bbox": [ + 517, + 103, + 901, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for our $\\mathcal{L}_{Cont}^{\\alpha}$ loss. The value $\\alpha$ is set individually for each dataset, i.e., $\\alpha = 5 \\times 10^{-2}$ for ADE20K, $\\alpha = 10^{-2}$ for VOC for Cityscapes. The details of our hyper-parameters are provided in the appendix.", + "bbox": [ + 511, + 223, + 903, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocols. We evaluate models on three standard datasets of CSS, i.e., ADE20K [76], Pascal VOC [23], and Cityscapes [16]. Following common practices [3, 4, 62], our experiments are conducted on the overlapped CSS settings. In particular, on ADE20K, we use three different settings, i.e., ADE20K 100-50 (2 steps), ADE20K 100-10 (6 steps), and ADE20K 100-5 (11 steps). On Pascal VOC, we evaluate FALCON in three benchmarks, i.e., VOC 15-5 (2 steps), VOC 15-1 (6 steps), and VOC 10-1 (11 steps). On Cityscapes, we conduct domain incremental experiments with three settings, i.e., Cityscapes 11-5 (3 steps), Cityscapes 11-1 (11 steps), and Cityscapes 1-1 (21 steps). Following [3, 20], the mean Intersection over Union (mIoU) metric is adopted in our comparison, including mIoU of the last learning step on initial classes, incremental classes, and all classes. In addition, to illustrate the fairness improvement, we report the mIoU of major and minor classes.", + "bbox": [ + 511, + 284, + 906, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 550, + 666, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of Fairness Contrastive Clustering. Table 1 presents our results using DeepLab-V3 [7] with Resnet101 on ADE20K 100-50 and ADE20K 100-10 benchmarks. We evaluate the impact of the fairness contrastive clustering loss $\\mathcal{L}_{Cont}^{\\alpha}$ by comparing it with the vanilla contrastive clustering loss $\\mathcal{L}_{Cont}$ . As shown in our results, the overall performance has been significantly improved to $37.9\\%$ and $36.4\\%$ on ADE20K 100-50 and ADE20K 100-10, respectively. In addition, the fairness has been promoted due to the mIoU improvement of minor groups. We also study the impact of network backbones and cluster margin $\\nabla$ in our appendix.", + "bbox": [ + 511, + 571, + 906, + 738 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6de08428c46c8ade2cd12713a846b9726af096f3c46affbe9b77ce6c41f87f8e.jpg", + "table_caption": [ + "Table 2. Effectiveness of Visual Grammar." + ], + "table_footnote": [], + "table_body": "
(a) ADE20K 100-50
0-100101-150allMajorMinor
Nearest Cluster44.311.533.451.524.3
Fixed φ44.617.635.652.027.4
Adaptive φ44.624.537.952.130.8
(b) ADE20K 100-10
0-100101-150allMajorMinor
Nearest Cluster40.114.331.548.722.9
Fixed φ43.018.534.950.627.0
Adaptive φ44.420.436.451.828.7
", + "bbox": [ + 517, + 753, + 901, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15070", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0d0fdd972919bf209c99284a59c094390b9351735b8b133afa5b3f2fb9c4ed29.jpg", + "table_caption": [ + "Table 3. Effectiveness of Scaling Factor $\\alpha$ .", + "(a) ADE20K 100-50" + ], + "table_footnote": [], + "table_body": "
α0-1516-20allMajorMinor
α = 0.143.119.835.350.627.7
α = 0.0544.624.537.952.130.8
α = 0.0143.621.336.251.028.7
α = 0.00542.418.634.550.126.6
", + "bbox": [ + 96, + 117, + 478, + 185 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fe99027b6c7156492607cdcd54903336cfb9026989a3e6b364f3324571dbc989.jpg", + "table_caption": [ + "(b) Pascal VOC 15-5" + ], + "table_footnote": [], + "table_body": "
α0-1516-20allMajorMinor
α = 0.174.851.669.376.963.5
α = 0.0576.251.370.379.063.8
α = 0.0179.454.873.581.367.7
α = 0.00574.648.968.577.661.6
", + "bbox": [ + 96, + 196, + 478, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Scaling Factor of Cluster. Table 3 illustrates the experimental results of the impact of different scaling factor $\\alpha$ on ADE20K 100-50 and Pascal VOC 15-5 benchmarks. As shown in Table 3, when the value of scaling factor $\\alpha$ gradually decreases, the performance of our proposed approach is improved accordingly since the fairness contrastive loss in Eqn (4) tends to be more uniform across major and minor classes. However, when the scaling factor is too small ( $\\alpha = 0.005$ ), the impact of the loss enforcement becomes weaker leading to the weaker enforcement of the fairness contrastive clustering, resulting in lower overall performance. In addition, we have observed that the higher the number of classes demands the higher the value of $\\alpha$ since it will increase the compactness of more clusters.", + "bbox": [ + 89, + 273, + 483, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Loss Contributions. Table 4 illustrates the contributions of proposed learning objectives. For the model without using visual grammar, we only use a single unknown cluster $(N_U = 1)$ and adopt the nearest cluster strategies to assign clusters of unknown pixels. By using only cross-entropy loss, the mIoU performance remains low due to catastrophic forgetting and background shift problems. Meanwhile, with our fairness clustering loss $\\mathcal{L}_{Cont}^{\\alpha}$ , visual grammar model $\\phi$ , and the cluster regularizer $\\mathcal{R}$ , the mIoU performance has been significantly improved to $37.9\\%$ and $36.4\\%$ on ADE20K 100-50 and ADE20K 100-10, respectively. Moreover, our FALCON has significantly promoted the fairness of segmentation models illustrated by the mIoU improvement of major and minor groups.", + "bbox": [ + 89, + 487, + 483, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Visual Grammar. We evaluate FALCON under three settings, i.e., Nearest Cluster, Fixed $\\phi$ pretrained", + "bbox": [ + 89, + 700, + 483, + 731 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a5141940d57570322073bf2076ff029cfa38919a60e6449d274a974127809ea.jpg", + "table_caption": [ + "Table 4. Effectiveness of Our Proposed Losses." + ], + "table_footnote": [], + "table_body": "
(a) ADE20K 100-50
\\( \\mathcal{L}_{CE} \\)\\( \\mathcal{L}_{Cont}^{\\alpha } \\)\\( \\phi \\)\\( \\mathcal{R}_{C} \\)0-100101-150allMajorMinor
✓0.018.96.30.09.4
✓✓44.07.931.951.622.1
✓✓✓43.821.836.451.129.1
✓✓✓✓44.624.537.952.130.8
", + "bbox": [ + 94, + 751, + 478, + 825 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7490d9b1e1e3de4bd66522df0318143a857798a49a6b59f3befa885dac3659ad.jpg", + "table_caption": [ + "(b) ADE20K 100-10" + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{CE} \\)\\( \\mathcal{L}_{Cont}^{\\alpha } \\)\\( \\phi \\)\\( \\mathcal{R}_C \\)0-100101-150allMajorMinor
✓0.03.51.20.01.8
✓✓39.013.130.447.821.6
✓✓✓43.418.535.151.227.1
✓✓✓✓44.420.436.451.828.7
", + "bbox": [ + 96, + 835, + 478, + 897 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ff1001ed48c41e4d2174c4832b287136ba02bd962dbe75b777e76f28cce5aa4.jpg", + "image_caption": [ + "Figure 6. Cluster Distribution at Learning Step $t = 1$ of ADE20K 100-50 (classes 109-144 are future classes) without (left) and with (right) Fairness Learning via Contrastive Attention." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 218 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ImageNet (without updating on each learning step), and Adaptive $\\phi$ (with updating on each learning step). As in Table 2, the mIoU result using only the nearest cluster remains ineffective. Meanwhile, the adaptive visual grammar model updated at each learning step further boosts the mIoU performance and promotes fairness, i.e., increased by $4.5\\%$ and $4.9\\%$ on ADE20K 100-50 and ADE20K 100-10 compared to the nearest cluster approach. In addition, we study the impact of choosing the number of features $M$ in the visual grammar model in our appendix. Fig. 6 illustrates the feature distributions of unknown classes (future class). As a result, our FALCON approach is able to model features of unknown classes into different clusters and produce better and more compact clusters compared to the one without Fairness Learning via Contrastive Attention.", + "bbox": [ + 511, + 268, + 906, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Forgetting Each Learning Step. Table 5 reports the results of our model evaluated at each learning step. The performance of our model remains stable at each learning step. In particular, the maximum performance drop at learning step 4 of classes 111-120 due forgetting is only $0.8\\%$ . Meanwhile, at the last learning task, the mIoU results of old classes are well maintained, with a slight drop of $0.4\\%$ in classes 131-140. As shown in the result, our model has effectively maintained the performance to avoid catastrophic forgetting while enabling the capability of learning classes.", + "bbox": [ + 511, + 494, + 908, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with Prior SoTA Methods", + "text_level": 1, + "bbox": [ + 511, + 657, + 846, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ADE20K. Table 6 presents our experimental results compared to prior CSS methods. By using DeepLab-V3, our approach has achieved SoTA performance, i.e., the mIoU results of $37.9\\%$ and $+36.4\\%$ on ADE20K 100-50 and ADE20K 100-10 benchmarks, higher than prior FairCL [62]. Meanwhile, FALCON using Transformer has outperformed", + "bbox": [ + 511, + 681, + 908, + 772 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ac24a0023f80c534a767ed4b8885945355b3ed69e5ac178c4a84c66a93a67595.jpg", + "table_caption": [ + "Table 5. Effectiveness of CSS Model At Each Learning Step." + ], + "table_footnote": [], + "table_body": "
ADE20K 100-10
0-100101-110111-120121-130131-140141-150all
Step 144.9-----44.9
Step 244.815.8----42.2
Step 344.715.722.1---40.3
Step 444.615.621.326.3--39.2
Step 544.515.421.226.124.4-37.9
Step 644.415.421.226.124.015.536.4
avg40.1
", + "bbox": [ + 517, + 797, + 901, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/afa9f66a1a896043680f1bae18cc2f78596ed042b5d902e421856661aa4865e9.jpg", + "table_caption": [ + "Table 6. Comparison with Prior Methods on ADE20K Benchmarks (Note: The results of MiB [2], PLOP [20], and FairCL [62] using Transformer on ADE20K 100-5 were not reported in prior studies. The upper bound results are not trained with fairness objective)." + ], + "table_footnote": [], + "table_body": "
NetworkMethodADE20K 100-50ADE20K 100-10ADE20K 100-5
0-100101-150allavg0-100101-150allavg0-100101-150allavg
DeepLab-V3PLOP [20]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [75]42.318.834.5-39.317.632.1-38.511.529.6-
REIMINDER [45]41.619.234.1-39.021.333.1-36.116.429.5-
RCIL+LGKD [73]43.325.137.2-42.220.434.9-----
FairCL [62]43.424.037.040.541.720.434.739.0----
FALCON44.624.537.941.344.420.436.440.138.016.130.737.6
Upper Bound44.328.238.9-44.328.238.9-44.328.238.9-
Mask2FormerMiB [2]37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP [20]44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer [3]44.726.238.441.240.615.632.337.439.513.630.936.5
ECLIPSE [31]45.021.737.1-43.417.434.6-43.316.334.2-
CoMasTRe [25]45.726.039.241.642.318.434.438.440.815.832.638.6
TransformerMiB [2]43.430.639.238.739.120.434.239.5----
PLOP [20]43.826.238.038.143.324.136.240.3----
FairCL [62]43.625.537.640.742.221.935.539.4----
FALCON47.530.641.943.547.326.240.342.840.818.933.538.1
Upper Bound48.739.045.5-48.739.045.5-48.739.045.5-
", + "bbox": [ + 94, + 145, + 565, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the prior CoMFormer [3] model by $+3.5\\%$ , $+8.0\\%$ , and $+2.6\\%$ on ADE20K 100-50, ADE20K 100-10, and ADE20K 100-5. In addition, our mIoU results on the initial classes remain competitive with the upper-bounded results because FALCON can well handle the fairness problem compared to fully supervised learning. Although ECLIPSE [31] exhibits slightly better on ADE20K 100-5 by only $0.7\\%$ (all) due to the better segmentation framework (i.e., Mask2Former), FALCON significantly outperforms ECLIPSE on ADE20K 100-50 and ADE20K 100-10. We also report our results of ADE20K 50-50 in the appendix. As in Fig. 7, FALCON produces better segmentation maps than prior methods.", + "bbox": [ + 88, + 334, + 480, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pascal VOC. Table 7 presents our results on Pascal VOC benchmarks. FALCON has consistently achieved the SoTA performance on three benchmarks. In particular, compared to the prior FairCL [62] approach, our methods using DeepLab-V3 have improved the mIoU performance up to $73.50\\%$ , $69.83\\%$ , and $62.41\\%$ on Pascal VOC 15-5, Pascal VOC 15-1, and Pascal VOC 10-1, respectively. Additionally, by using the better backbone, the performance is further improved, and the gap with the upper-bound result is reduced.", + "bbox": [ + 89, + 517, + 485, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cityscapes. Table 8 reports the performance of our approach using DeepLab-V3 compared to prior methods on three different settings, i.e., Cityscapes 11-5, Cityscapes", + "bbox": [ + 89, + 652, + 483, + 698 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f4de9087e49e5ac462c81197945df54f737630d27614a25cc819ded3bf2bf14.jpg", + "image_caption": [ + "Figure 7. The Comparison of Qualitative Results on ADE20K Between Our Approach and RCIL [75]." + ], + "image_footnote": [], + "bbox": [ + 573, + 98, + 901, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "11-1, and Cityscapes 1-1. As in Table 8, the performance of our methods has consistently outperformed prior FairCL [62] approach by $+3.78\\%$ , $+3.14\\%$ , and $+6.02\\%$ on three benchmarks. Similar to our experiments on ADE20K and VOC, the better network brings higher results.", + "bbox": [ + 511, + 334, + 906, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 511, + 417, + 777, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusions. This paper has presented a novel Fairness Learning via Contrastive Attention approach to CSS. In particular, the fairness contrastive clustering loss has been introduced to model both catastrophic forgetting and fairness problems. Then, the visual grammar model was presented to model the unknown classes. The experimental results on different benchmarks have shown the SoTA performance and fairness promotion of our proposed FALCON approach. Limitations. Our study chose a set of learning hyperparameters to support our theoretical analysis. However, it potentially consists of several limitations related to choosing learning parameters and cluster initialization. Further discussion of limitations is presented in the supplementary.", + "bbox": [ + 511, + 438, + 908, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs.", + "bbox": [ + 511, + 638, + 906, + 698 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/56ac49bb9da1ae40a40fca599a264c4b145c3f9eae9596afe151594ecded11b7.jpg", + "table_caption": [ + "Table 7. Comparisons with Prior Methods on Pascal VOC." + ], + "table_footnote": [], + "table_body": "
MethodPascal VOC 15-5Pascal VOC 15-1Pascal VOC 10-1
0-1516-20all0-1516-20all0-1011-20all
DeepLab-V3MiB [2]76.3749.9770.0838.0013.5032.2020.0020.1020.10
PLOP [20]75.7351.7170.0965.1021.1054.6044.0015.5030.50
RCIL [75]---70.6023.7059.4055.4015.1034.30
FairCL [62]---72.0022.7060.3042.3025.6034.40
SSUL [4]77.8250.1071.2277.3136.5967.6171.3145.9859.25
FALCON79.3554.7773.5078.3442.5769.8373.9449.7362.41
Upper Bound79.7772.3577.4379.7772.3577.4378.4176.3577.43
TransformerPLOP [20]72.5148.3766.7664.5937.2358.0848.5333.7141.47
SSUL [4]79.9156.8374.4179.9140.5670.5474.0651.8563.48
FairCL [62]---73.5022.8061.5057.1014.2036.60
FALCON81.2058.0475.6978.7147.5471.2874.9252.5464.26
Upper Bound80.8474.9779.4480.8474.9779.4480.8474.9779.44
", + "bbox": [ + 96, + 719, + 578, + 897 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b207b321aa66a8938add9cc1bb0eb473b97ce1685e98c69e90c50ac6c1fcdcf1.jpg", + "table_caption": [ + "Table 8. Comparisons on Cityscapes." + ], + "table_footnote": [], + "table_body": "
Method11-511-11-1
DeepLab-V3LWF-MC [48]58.9056.9231.24
ILT [38]59.1457.7530.11
MIB [2]61.5160.0242.15
PLOP [20]63.5162.0545.24
RCIL [75]64.3063.0048.90
FairCL [62]66.9666.6149.22
FALCON70.7469.7555.24
Upper Bound79.3079.3079.30
Trans.FairCL [62]67.8567.0955.68
FALCON71.3370.1458.79
Upper Bound83.8083.8083.80
", + "bbox": [ + 602, + 719, + 897, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15072", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nikita Araslanov, and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[2] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. Conf. Comp. Vision Pattern Rec., 2020.", + "[3] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. IEEE/CVF Computer Vision and Pattern Recognition Conference, 2023.", + "[4] Sungmin Cha, beomyoung kim, YoungJoon Yoo, and Taesup Moon. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In Advances in Neural Information Processing Systems, pages 10919-10930. Curran Associates, Inc., 2021.", + "[5] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. arXiv preprint arXiv:2407.16354, 2024.", + "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017.", + "[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. TPAMI, 2018.", + "[8] Mayee Chen, Daniel Y Fu, Avanika Narayan, Michael Zhang, Zhao Song, Kayvon Fatahalian, and Christopher Ré. Perfectly balanced: Improving transfer and robustness of supervised contrastive learning. In International Conference on Machine Learning, pages 3090-3122. PMLR, 2022.", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020.", + "[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021.", + "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021.", + "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022.", + "[14] Sanghyeok Chu, Dongwan Kim, and Bohyung Han. Learning debiased and disentangled representations for semantic segmentation. Advances in Neural Information Processing Systems, 34:8355-8366, 2021.", + "[15] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for in-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cremental semantic segmentation. In European Conference on Computer Vision, pages 244-261. Springer, 2024.", + "[16] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016.", + "[17] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021.", + "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.", + "[19] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proc. Europ. Conf. Computer Vision, 2020.", + "[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4040-4050, 2021.", + "[21] Beyza Ermis, Giovanni Zappella, Martin Wistuba, Aditya Rawal, and Cedric Archambeau. Continual learning with transformers for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3774-3781, 2022.", + "[22] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996.", + "[23] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015.", + "[24] Robert French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 1999.", + "[25] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3857, 2024.", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.", + "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020.", + "[28] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. DAFormer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022.", + "[29] Ibsa Jalata, Naga Venkata Sai Raviteja Chappa, Thanh-Dat Truong, Pierce Helton, Chase Rainwater, and Khoa Luu. Eqadap: Equipollent domain adaptation approach to image deblurring. IEEE Access, 10:93203-93211, 2022.", + "[30] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detect" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15073", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5830-5840, 2021.", + "[31] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024.", + "[32] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. PNAS, 2017.", + "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021.", + "[34] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 8547-8555, 2021.", + "[35] Chang Liu, Giulia Rizzoli, Pietro Zanuttigh, Fu Li, and Yi Niu. Learning from the web: Language drives weakly-supervised incremental learning for semantic segmentation. arXiv preprint arXiv:2407.13363, 2024.", + "[36] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world, 2019.", + "[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In NeurIPS, 2017.", + "[38] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCVWS, 2019.", + "[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019.", + "[40] Pha Nguyen, Thanh-Dat Truong, Miaoqing Huang, Yi Liang, Ngan Le, and Khoa Luu. Self-supervised domain adaptation in crowd counting. In 2022 IEEE international conference on image processing (ICIP), pages 2786-2790. IEEE, 2022.", + "[41] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10847-10856, 2021.", + "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.", + "[43] First Ozdemir and Orcun Goksel. Extending pretrained segmentation networks with additional anatomical structures. International journal of computer assisted radiology and surgery, 14:1187-1195, 2019.", + "[44] First Ozdemir, Philipp Fuernstahl, and Orcun Goksel. Learn the new, keep the old: Extending pretrained models with new anatomy and images. In Medical Image Computing" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part IV 11, pages 361-369. Springer, 2018.", + "[45] Minh Hieu Phan, The-Anh Ta, Son Lam Phung, Long Tran-Thanh, and Abdesselam Bouzerdoum. Class similarity weighted knowledge distillation for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16866-16875, 2022.", + "[46] Yiqiao Qiu, Yixing Shen, Zhuohao Sun, Yanchong Zheng, Xiaobin Chang, Weishi Zheng, and Ruixuan Wang. Sats: Self-attention transfer for continual semantic segmentation. Pattern Recognition, 138:109383, 2023.", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.", + "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proc. Conf. Comp. Vision Pattern Rec., 2017.", + "[49] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition, 2020.", + "[50] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995.", + "[51] Mohammad Rostami. Lifelong domain adaptation via consolidated internal distribution. In NeurIPS, 2021.", + "[52] Antoine Saporta, Arthur Douillard, Tuan-Hung Vu, Patrick Pérez, and Matthieu Cord. Multi-head distillation for continual unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2022.", + "[53] Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. Incremental learning of object detectors without catastrophic forgetting. In Proceedings of the IEEE international conference on computer vision, pages 3400-3409, 2017.", + "[54] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021.", + "[55] Attila Szabó, Hadi Jamali-Rad, and Siva-Datta Mannava. Tilted cross-entropy (tce): Promoting fairness in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2305–2310, 2021.", + "[56] Sebastian Thrun. Lifelong learning algorithms. In *Springer Learning to Learn*, 1998.", + "[57] Thanh-Dat Truong and Khoa Luu. Cross-view action recognition understanding from exocentric to egocentric perspective. Neurocomputing, 614:128731, 2025.", + "[58] Thanh-Dat Truong, Chi Nhan Duong, Khoa Luu, Minh-Triet Tran, and Ngan Le. Domain generalization via universal non-volume preserving approach. In 2020 17th Conference" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15074", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer and Robot Vision (CRV), pages 93-100. IEEE, 2020.", + "[59] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Son Lam Phung, Chase Rainwater, and Khoa Luu. Bimal: Bijective maximum likelihood approach to domain adaptation in semantic scene segmentation. In International Conference on Computer Vision, 2021.", + "[60] Thanh-Dat Truong, Ravi Teja Nvs Chappa, Xuan-Bac Nguyen, Ngan Le, Ashley PG Dowling, and Khoa Luu. Otadapt: Optimal transport-based approach for unsupervised domain adaptation. In 2022 26th international conference on pattern recognition (ICPR), pages 2850-2856. IEEE, 2022.", + "[61] Thanh-Dat Truong, Ngan Le, Bhiksha Raj, Jackson Cothren, and Khoa Luu. Freedom: Fairness domain adaptation approach to semantic scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023.", + "[62] Thanh-Dat Truong, Hoang-Quan Nguyen, Bhiksha Raj, and Khoa Luu. Fairness continual learning approach to semantic scene understanding in open-world environments. In NeurIPS, 2023.", + "[63] Thanh-Dat Truong, Pierce Helton, Ahmed Moustafa, Jackson David Cothren, and Khoa Luu. Conda: Continual unsupervised domain adaptation learning in visual perception for self-driving cars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 5642-5650, 2024.", + "[64] Thanh-Dat Truong, Utsav Prabhu, Dongyi Wang, Bhiksha Raj, Susan Gauch, Jeyamkondan Subbiah, and Khoa Luu. EAGLE: Efficient adaptive geometry-based learning in cross-view understanding. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.", + "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[66] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proc. Conf. Comp. Vision Pattern Rec., 2021.", + "[67] Huyong Wang, Huisi Wu, and Jing Qin. Incremental nuclei segmentation from histopathological images via future-class awareness and compatibility-inspired distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11408-11417, 2024.", + "[68] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021.", + "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021.", + "[70] Zhengyuan Xie, Haiquan Lu, Jia-wen Xiao, Enguang Wang, Le Zhang, and Xialei Liu. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation. In European Conference on Computer Vision, pages 183-201. Springer, 2025.", + "[71] Lei Yang, Xiaohang Zhan, Dapeng Chen, Junjie Yan, Chen Change Loy, and Dahua Lin. Learning to cluster faces on an affinity graph. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2298-2306, 2019.", + "[72] Lei Yang, Dapeng Chen, Xiaohang Zhan, Rui Zhao, Chen Change Loy, and Dahua Lin. Learning to cluster faces via confidence and connectivity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13369-13378, 2020.", + "[73] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023.", + "[74] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. arXiv preprint arXiv:2407.09838, 2024.", + "[75] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022.", + "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017.", + "[77] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022." + ], + "bbox": [ + 516, + 92, + 906, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_model.json b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01195cf70d64eb1bc172adfb6c51683938276a2e --- /dev/null +++ b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.125, + 0.87, + 0.18 + ], + "angle": 0, + "content": "FALCON: Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.192, + 0.845, + 0.265 + ], + "angle": 0, + "content": "Thanh-Dat Truong\\(^{1}\\), Utsav Prabhu\\(^{2}\\), Bhiksha Raj\\(^{3,4}\\), Jackson Cothren\\(^{5}\\), Khoa Luu\\(^{1}\\) \n\\(^{1}\\)CVIU Lab, University of Arkansas, USA \\(^{2}\\)Google DeepMind, USA \n\\(^{3}\\)Carnegie Mellon University, USA \\(^{4}\\)Mohammed bin Zayed University of AI, UAE \n\\(^{5}\\)Dep. of Geosciences, University of Arkansas, USA" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.267, + 0.846, + 0.299 + ], + "angle": 0, + "content": "{tt032, jcothre, khoaluu}@uark.edu, bhiksha@cs.cmu.edu, utsavprabhu@google.com http://uark-cviu.github.io/projects/FALCON" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.32, + 0.327, + 0.336 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.356, + 0.486, + 0.704 + ], + "angle": 0, + "content": "Continual Learning in semantic scene segmentation aims to continually learn new unseen classes in dynamic environments while maintaining previously learned knowledge. Prior studies focused on modeling the catastrophic forgetting and background shift challenges in continual learning. However, fairness, another major challenge that causes unfair predictions leading to low performance among major and minor classes, still needs to be well addressed. In addition, prior methods have yet to model the unknown classes well, thus resulting in producing non-discriminative features among unknown classes. This work presents a novel Fairness Learning via Contrastive Attention Approach to continual learning in semantic scene understanding. In particular, we first introduce a new Fairness Contrastive Clustering loss to address the problems of catastrophic forgetting and fairness. Then, we propose an attention-based visual grammar approach to effectively model the background shift problem and unknown classes, producing better feature representations for different unknown classes. Through our experiments, our proposed approach achieves State-of-the-Art (SoTA) performance on different continual learning benchmarks, i.e., ADE20K, Cityscapes, and Pascal VOC. It promotes the fairness of the continual semantic segmentation model." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.902 + ], + "angle": 0, + "content": "The semantic segmentation networks, e.g., Transformers [69] and Convolutional Neural Networks [7], learned from data with a closed-set of known classes, have shown outstanding performance. However, they often suffer performance degradation when encountering novel objects or classes in new dynamic environments [2, 20, 62, 63]. To improve their performance, several transfer learning and domain adaptation methods [1, 28, 29, 40, 57-60, 62, 64] were introduced to adapt trained models into deployed environments. While" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.318, + 0.907, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Figure 1. Our Fairness Learning via Contrastive Attention to Continual Semantic Segmentation. The Fairness Contrastive Clustering Loss promotes the fairness of the model while the Attention-based Visual Grammar models the unknown classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.732 + ], + "angle": 0, + "content": "the former often aims to fine-tune the model on labeled data collected in the new environments, the latter adapts the model to the new domains in an unsupervised manner [1, 28]. However, these methods cannot handle novel objects well due to their close-set learning. In practice, the semantic segmentation models should be able to adaptively and continually learn the new knowledge of novel classes. It motivates the development of Continual Learning paradigm [2, 20, 46], a.k.a, Continual Semantic Segmentation (CSS), where the segmentation models are learned sequentially to new contents of data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Far apart from prior segmentation methods [7, 69] that learn one time on static, closed-set data, Continual Learning requires the segmentation models to learn from dynamic, open-set data [20, 62]. In a particular scenario, accessing previous learning data is restricted due to privacy concerns. In CSS, three challenges have been identified, including (1) Catastrophic Forgetting, (2) Background Shift, and (3) Fairness. While the catastrophic forgetting problem [24, 50, 56] depicts the segmentation model tends to forget its knowledge when learning new data, background shift indicates the problem of classes of previous or future data (unknown classes)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15065" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Figure 2. The Data Class Distribution of ADE20K. The major classes occupy more than \\(75\\%\\) of the total pixels of the dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.28, + 0.483, + 0.414 + ], + "angle": 0, + "content": "have collapsed into a background class [2, 20, 75]. Prior methods [4, 20, 46, 62] addressed these two problems by introducing knowledge distillation and pseudo labels. However, these methods can not handle unknown classes since they either consider these unknown classes as a background class or assign unknown pixels by a pseudo label of prior known classes [20, 62]. More importantly, the last problem, fairness, is a significant challenge that limits the performance of CSS models (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.416, + 0.484, + 0.687 + ], + "angle": 0, + "content": "As shown in Fig. 2, the number of pixels of each class in training data have been imbalanced among classes and significantly decreased after each task. Thus, this bias influences the learning procedure and model predictions that later cause unfair class predictions. However, limited studies are taking the fairness problem into account. [61] presented a similar problem in domain adaptation and extended it to continual learning [62]. These methods rely on the assumption of ideal fair or balanced data distributions. However, it is not applicable in practice since the size, i.e., the number of pixels, of several classes can never be more significant than others. For example, the bottle size should not be more significant than the size of a car. Meanwhile, the current knowledge distillation methods [3, 20, 45] in CSS are unable to handle the fairness problem since they focus on modeling catastrophic forgetting and background shift problems. Therefore, it is essential to develop a new CSS approach to address these limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Contributions of This Work. This work presents a novel Fairness Learning via Contrastive Attention Approach (FALCON) to Continual Semantic Segmentation (as shown in Fig. 1). Our contributions can be summarized as follows. First, we introduce a novel Contrastive Clustering Paradigm approach to Continual Learning that models the catastrophic forgetting problem. Second, by analyzing the limitation of vanilla Contrastive Clustering in biased data, we introduce a novel Fairness Contrastive Clustering loss to model the fairness problem in continual learning efficiently. Third, to effectively model the background shift problem, we introduce a new Attention-based Visual Grammar that model the topological structures of feature distribution to handle the unknown classes effectively. Finally, the ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.168 + ], + "angle": 0, + "content": "illustrate the effectiveness of the proposed approach in different aspects of fairness promotion in CSS models. Compared with prior methods, our approach achieves SoTA performance on different settings of three standard benchmarks of CSS, including ADE20K, Pascal VOC, and Cityscapes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.181, + 0.655, + 0.196 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.206, + 0.908, + 0.552 + ], + "angle": 0, + "content": "Continual Semantic Segmentation. Several studies were introduced to address catastrophic forgetting and background shift problems [19-21, 32, 37, 39, 43, 44]. The common CSS approach adopts knowledge distillation [20] and pseudo labels [2] to model catastrophic forgetting and background shift, respectively. Later, it was further improved by decoupling knowledge representations [75], modeling the inter- and intra-class knowledge [15, 46], distinguishing the feature representations of the future classes [4, 45, 67, 70], reducing background confusion [73, 74], using visual prompting [31, 35], or modeling distillation loss via the geodesic flow [54]. Another approach [3, 5, 25] adopted the mask-based segmentation networks [12, 13] to improve the performance. Recent studies have introduced CSS under the unsupervised domain adaptation settings [51, 52, 63, 66]. However, prior studies have yet to well model different unknown classes. Particularly, as previous methods [3, 20] use the pseudo labels to model the unknown classes, the future classes will be treated as a single background class. Then, Joseph et al. [30] improves the unknown class modeling using clustering but this method considers all different unknown classes as a single cluster, leading to the non-discriminative features among unknown classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Contrastive Learning (CL). CL is a common learning approach [9, 10, 42] to structure the deep feature representations in the deep latent space. Oorde et al. [42] first introduced the Noise-Contrastive Estimation (InfoNCE) learning framework. Then, Chen et al. [9] presented SimCLR, a self-supervised contrastive learning approach to improve the representation power of Residual Networks. He [27] proposed a Momentum Contrast framework for unsupervised representation learning. Later, it was further improved by using MLP projection head [10] and extended to improve the self-supervised training process of vision transformers [11]. Cui et al. [17] introduced a supervised parametric contrastive learning loss to address the long-tailed recognition. Li et al. [34] adopted contrastive learning to develop the one-stage online contrastive clustering method. Radford et al. [47] presents a contrastive framework to learn the vision-language model. Later, several methods also adopted this framework for vision-language pretraining [33, 47]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Imbalanced and Fairness Learning. The early methods utilized the balanced Softmax loss [49] to alleviate the impact of imbalanced data distribution. Later, Wang et al. [68] introduced a Seesaw loss to re-balance the contributions of positive and negative instances via the mitigation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15066" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.318 + ], + "angle": 0, + "content": "and compensation modules. Ziwei et al. [36] introduced a dynamic meta-embedding to model the imbalanced classification problem. Chu et al. [14] reduce the bias in the segmentation model by presenting a new stochastic training scheme. Szabo et al. [55] presented a tilted cross-entropy loss to promote class-relevant fairness. However, there are limited studies that address the fairness problem in CSS. Truong et al. [61] introduced a fairness domain adaptation approach to semantic segmentation and later extended it into continual learning setting [62]. However, these methods [61, 62] rely on the assumption of ideal balanced data, which could not be achieved by nature. To address the limitations in prior work, this paper will introduce a novel approach to effectively model the fairness problem and unknown classes in the continual learning setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.331, + 0.406, + 0.349 + ], + "angle": 0, + "content": "3. The Proposed FALCON Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.493 + ], + "angle": 0, + "content": "CSS aims to learn a segmentation network \\(F\\) on sequence data \\(\\mathcal{D} = \\{\\mathcal{D}^1,\\dots,\\mathcal{D}^T\\}\\) where \\(T\\) is the number of learning steps. At learning step \\(t\\), the model \\(F\\) encounters a dataset \\(\\mathcal{D}_t = \\{(\\mathbf{x}^t,\\hat{\\mathbf{y}}^t)\\}\\) where \\(\\mathbf{x}^t\\in \\mathbb{R}^{H\\times W\\times 3}\\) is the image and \\(\\mathbf{y}\\in \\mathbb{R}^{H\\times W}\\) is a segmentation label of \\(\\mathbf{x}^t\\). The ground truths at learning step \\(t\\) only consist of current classes \\(\\mathcal{C}^t\\), while the class labels of the previous \\(\\mathcal{C}^{1\\dots t - 1}\\) and future steps \\(\\mathcal{C}^{t + 1\\dots T}\\) are collapsed into a background class. Formally, learning the CSS model at step \\(t\\) can be formed as Eqn. (1)." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.496, + 0.484, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} ^ {*} = \\arg \\min _ {\\theta_ {t}} \\mathbb {E} _ {\\mathbf {x} ^ {t}, \\hat {\\mathbf {y}} ^ {t} \\in \\mathcal {D} ^ {t}} \\left[ \\mathcal {L} _ {C E} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} ^ {t}\\right) + \\lambda_ {C L} \\mathcal {L} _ {C L} \\left(F (\\mathbf {x} ^ {t})\\right) \\right] (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.826 + ], + "angle": 0, + "content": "where, \\(\\mathbf{y}^t = F(\\mathbf{x}^t,\\theta_t)\\), \\(\\theta_{t}\\) is the parameter of \\(F\\) at current learning step \\(t\\), \\(\\mathcal{L}_{CE}\\) is the cross-entropy loss, \\(\\lambda_{CL}\\) is the balanced weight. and \\(\\mathcal{L}_{CL}\\) is the CSS objective. At learning step \\(t\\), the segmentation model \\(F\\) is required to be able to predict both previously learned classes \\(\\mathcal{C}^{1\\dots t - 1}\\) and current new classes \\(\\mathcal{C}^t\\). Under this learning scenario, three challenges have been identified, i.e., Catastrophic Forgetting, Background Shift, and Fairness. Several prior methods were presented to model the two first issues in CSS using knowledge distillation [3, 20]. The last issue has not been well addressed yet due to its challenges [62]. Prior methods [2, 3, 20, 53, 75] adopt knowledge distillation to design \\(\\mathcal{L}_{CL}\\). However, this method prevents the CSS model from diverging knowledge learned previously, therefore resulting in limiting the ability to adopt new knowledge [62]. In addition, these methods have not addressed fairness and background shift problems due to their dedicated design for maintaining knowledge via distillation [3, 20, 45]. Therefore, to address these problems, we introduce a novel Fairness Learning via Contrastive Attention Approach to CSS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.484, + 0.851 + ], + "angle": 0, + "content": "3.1. Continual Learning via Contrastive Clustering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Apart from prior methods [3, 20, 45], our CSS (Fig. 3) is defined as Contrastive Clustering Learning. Given a set of centroid vectors \\(\\{\\mathbf{c}_i\\}_{i=1}^{N_K + N_U}\\) where \\(N_K = |\\mathcal{C}^{1..t}|\\) and \\(N_U\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.182 + ], + "angle": 0, + "content": "is the number of known and unknown classes up to current learning tasks. Prior work [3, 30, 62] often defined the number of unknown classes as 1 where background classes are considered as a single unknown class. Formally, our Contrastive Clustering Learning for CSS can be defined as in Eqn. (2)." + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.189, + 0.906, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {C L} \\left(F (\\mathbf {x} ^ {t})\\right) = \\sum_ {\\mathbf {c} _ {i}} \\mathcal {L} _ {C o n t} (\\mathbf {F} ^ {t}, \\mathbf {c} _ {i}) \\\\ = \\sum_ {\\mathbf {c} _ {i}} \\sum_ {h, w} - \\phi \\left(\\mathbf {f} _ {h, w} ^ {t}, \\mathbf {c} _ {i}\\right) \\log \\frac {\\exp \\left(\\mathbf {f} _ {h , w} ^ {t} \\times \\mathbf {c} _ {i}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c} _ {i}\\right)} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.257, + 0.907, + 0.352 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}^t\\in \\mathbb{R}^{H\\times W\\times D}\\) is the feature maps extracted from the input image \\(\\mathbf{x}^t\\) by the segmentation network \\(F\\), \\(\\mathbf{f}_{h,w}^t\\in \\mathbb{R}^D\\) is the feature at the pixel location \\((h,w)\\) of features \\(\\mathbf{F}^t\\), \\(\\sum_{\\mathbf{f}'}\\) means the summation over all feature representations \\(\\mathbf{f}'\\in \\mathbb{R}^D\\), and \\(\\phi :\\mathbb{R}^D\\times \\mathbb{R}^D\\to [0,1]\\) is the function that determines either \\(\\mathbf{f}_{h,w}^t\\) belongs to the cluster \\(\\mathbf{c}_i\\) or not." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.909, + 0.669 + ], + "angle": 0, + "content": "By defining CSS as contrastive clustering learning, the knowledge of the segmentation model has been well maintained via the cluster vectors \\(\\mathbf{c}\\). Then, minimizing Eqn. (2) will separate the representations of different classes while gathering the features of the same class into the same cluster. As the cluster vectors \\(\\mathbf{c}\\) of the old classes \\(\\mathcal{C}^{1..t-1}\\) have been well learned to represent their knowledge, these vectors are frozen at learning step \\(t\\) to maintain the knowledge representations of previous classes to address the catastrophic forgetting problem. To effectively learn cluster vectors \\(\\mathbf{c}\\), the cluster vector \\(\\mathbf{c}\\) will periodically updated after each \\(K\\) steps by the momentum update [27, 30, 62] based on the features \\(\\mathbf{f}_{h,w}^t\\) assigned to cluster \\(\\mathbf{c}\\). However, there are two major problems in contrastive clustering learning. First, since the training data suffer the bias among classes as shown in Fig. 2, this bias will influence Eqn. (2) and cause the unfair predictions. Second, as the function \\(\\phi\\) requires the labels to determine the features belonging to clusters, it limits the ability to model the unknown classes where the labels are unavailable. Therefore, Secs. 3.2-3.3 will present a novel approach to tackle these problems." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.869, + 0.694 + ], + "angle": 0, + "content": "3.2. Fairness Contrastive Clustering Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.901 + ], + "angle": 0, + "content": "While contrastive clustering learning in Eqn. (2) promotes the compact representations of features around their clusters, inspired by [8, 17, 77], we observe that the imbalanced class distribution will influence unfair behaviors among classes. In particular, for simplicity, we consider \\(\\{\\mathbf{f}_i^t\\}_{i=1}^L\\) is the set of features that belong to the cluster \\(\\mathbf{c}\\) at learning step \\(t\\) (i.e., \\(\\phi(\\mathbf{f}_i^t, \\mathbf{c}) = 1\\)) and \\(L\\) is the number of features (in this case, \\(L\\) is the total number of pixels belong to the class of cluster \\(\\mathbf{c}\\)). Let us define the enforcement between the feature \\(\\mathbf{f}_t^t\\) and the cluster \\(\\mathbf{c}\\) as \\(\\ell_i = \\frac{\\exp(\\mathbf{f}_i^t \\times \\mathbf{c})}{\\sum_{\\mathbf{f}' \\in L} \\exp(\\mathbf{f}' \\times \\mathbf{c})}\\). Hence, the lower the value of the enforcement \\(\\ell_i\\) is, the more compact the representation of visual features and clusters is. Then, the contrastive clustering learning loss in Eqn. (2) of entire" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.088, + 0.825, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.24, + 0.888, + 0.255 + ], + "angle": 0, + "content": "Figure 3. The Proposed Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.263, + 0.347, + 0.277 + ], + "angle": 0, + "content": "cluster \\(\\mathbf{c}\\) can be defined as in Eqn. (3)." + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.287, + 0.484, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {C o n t}} (;;, \\mathbf {c}) = - \\sum_ {i = 1} ^ {L} \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {t} \\times \\mathbf {c}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} = - \\sum_ {i = 1} ^ {L} \\log \\ell_ {i} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.335, + 0.483, + 0.379 + ], + "angle": 0, + "content": "Proposition 1. If the contrastive clustering loss \\(\\mathcal{L}_{Cont}(:,\\mathbf{c})\\) achieves the optimal value, the enforcement \\(\\ell_{i}\\) between the feature and the cluster will converge to \\(\\ell_{i} = L^{-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.381, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Proposition 1 has implied that the class with more samples will result in a lower value of the enforcement and produce a more compact representation while the class having fewer samples will be more scattered in the feature space due to the higher value of the enforcement. In particular, let \\( L_{major} \\) and \\( L_{minor} \\) be the number of samples of the major and minor class where \\( L_{major} >> L_{minor} \\). Then, based on Proposition 1, the enforcement between features and the cluster of the major class will be significantly lower than the one of the minor class, i.e., \\( L_{major}^{-1} << L_{minor}^{-1} \\). Therefore, a direct adoption of the contrastive clustering loss in Eqn. (2) will result in an unfair CSS model. In addition, for classes in the minority group, the lower value of \\( L \\) results in the feature presentations of classes being far away from their clusters. Thus, the model will produce non-discriminative features compared to the ones in the majority group. Moreover, if the loss is applied to the cases of unknown labels, these feature representations can be scattered in the latent space and pulled into the incorrect clusters due to weak enforcement between features and clusters (Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.484, + 0.773 + ], + "angle": 0, + "content": "To address the unfair problem in contrastive clustering learning, inspired by [8, 17, 77], we introduce a scaling factor \\(\\alpha\\) and a learnable transition vector \\(\\mathbf{v}\\) for each cluster \\(\\mathbf{c}\\) (all clusters have the same value of \\(\\alpha\\) but different vector \\(\\mathbf{v}\\)). Our Fairness Contrastive Clustering Learning Loss for the entire cluster in Eqn. (3) can be re-formed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.78, + 0.484, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {C o n t}} ^ {\\alpha} (;, \\mathbf {c}) = - \\alpha \\sum_ {i = 1} ^ {L} \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {t} \\times \\mathbf {c}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} - \\log \\frac {\\exp (\\mathbf {v} \\times \\mathbf {c})}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Intuitively, the scaling factor \\(\\alpha\\) will help to re-scale the impact of the enforcement in learning, and the transitive vector \\(\\mathbf{v}\\) assists in translating the center cluster into the proper position of the latent space. This action promotes the compactness of clusters in the minority group." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.263, + 0.905, + 0.322 + ], + "angle": 0, + "content": "Proposition 2. If the fairness contrastive clustering loss \\(\\mathcal{L}_{Cont}^{\\alpha}(:,\\mathbf{c})\\) achieves the optimal value, the enforcement \\(\\ell_{i}\\) between the feature and the cluster will converge to \\(\\ell_{i} = (\\alpha^{-1} + L)^{-1}\\). Proofs of Propositions 1-2 are in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.908, + 0.549 + ], + "angle": 0, + "content": "Under the Proposition 2, when the value of \\(\\alpha\\) is small, the divergence of the enforcement between major and minor classes will be smaller, i.e., \\(||(\\alpha^{-1} + L_{major})^{-1} - (\\alpha^{-1} + L_{minor})^{-1}|| < ||L_{major}^{-1} - L_{minor}^{-1}||\\). Fig. 4 has illustrated the impact of fairness contrastive clustering loss. Therefore, our designed proposed fairness contrastive loss has effectively addressed the fairness issue in Eqn. (2). It should be noted that although the smaller \\(\\alpha\\) results in the fairer enforcement varied from major to minor classes. However, if the value of scaling factor \\(\\alpha\\) is too small, the contrastive clustering loss will rely more on the enforcement of the transitive vector \\(\\mathbf{v}\\), and the distribution of features \\(\\mathbf{f}_i^t\\) around its cluster \\(\\mathbf{c}\\) will be scattered due the weak enforcement caused by small \\(\\alpha\\). Therefore, the value of scaling factor \\(\\alpha\\) in practice should be carefully selected." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.844, + 0.574 + ], + "angle": 0, + "content": "3.3. An Efficient Unknown Class Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.905, + 0.611 + ], + "angle": 0, + "content": "An ideal CSS approach must be able to model the unknown classes without supervision, especially in open-set contexts" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.823 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.827, + 0.907, + 0.897 + ], + "angle": 0, + "content": "Figure 4. The Enforcement Loss of Contrastive Clustering \\(\\mathcal{L}_{Cont}\\) and Fairness Contrastive Clustering \\(\\mathcal{L}_{Cont}^{\\alpha}\\) on Pascal VOC. Since \\(\\mathcal{L}_{Cont}\\) suffers severe biased, its clusters of minor classes remain scattered. Our \\(\\mathcal{L}_{Cont}^{\\alpha}\\) produces a more uniform loss among classes, promotes fairness and compactness of clusters." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "[30, 62] where there could be multiple unknown classes or objects. Prior studies have adopted the pseudo-label strategies [3, 20] based on the model predictions to assign labels for seen classes while unseen classes have been ignored, thus resulting in non-discriminative features. [30, 62] improved the background modeling by using an additional prototypical representation for unknown classes. However, these approaches consider different unknown classes as one (i.e., \\( N_U = 1 \\)) resulting in non-distinguished representations of different unknown classes. Thus, modeling function \\( \\phi \\) in Eqn. (2) without supervision of different unknown classes (i.e., \\( N_U > 1 \\)) is challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.485, + 0.489 + ], + "angle": 0, + "content": "Although modeling \\(\\phi\\) to determine the single feature \\(\\mathbf{f}\\) belonging to the cluster \\(\\mathbf{c}\\) is challenging, prior studies in clustering [41, 71, 72] have suggested that determine a set of features \\(\\{\\mathbf{f}_i^t\\}_{i = 1}^M\\) belonging to cluster \\(\\mathbf{c}\\) should be easier. This derives from the fact that even though the feature representations of different classes are different, the distributions of features around its cluster (termed as Visual Grammar) in the feature space should be similar among classes or clusters. As a result, by learning the distribution of features and their clusters, the model \\(\\phi\\) can determine whether a feature belongs to a cluster. Then, by learning the model \\(\\phi\\) on prior known clusters and features, the knowledge of \\(\\phi\\) can be adaptively applied to unknown clusters. Fig. 5 illustrates our visual grammar model of the cluster distributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Limitations of Prior Clustering Methods. The traditional methods in clustering, e.g., KNN or density-based clustering [22], remain limited to noisy features leading to producing the incorrect cluster assignment. Meanwhile, the modern clustering methods, e.g., Graph Neural Networks (GNNs) [71, 72], require a large memory to build the affinity graph for clusters. In addition, GNNs often learn the local structures of graphs (or clusters) and accumulate them via the aggregation layers. Hence, the global structures of the clusters, i.e., visual grammar, are not well modeled by GNNs [41]. Therefore, to address these limitations, we introduced a new Attention-based Visual Grammar approach to efficiently model the distribution of features and their clusters via the self-attention mechanism [65]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.77 + ], + "angle": 0, + "content": "Remark 1: Given a center \\(\\mathbf{c}\\) and a set of \\(M\\) features \\(\\{\\mathbf{f}_i^{\\mathbf{c}}\\}_{i = 1}^{M}\\) where \\(\\mathbf{f}_i^{\\mathbf{c}}\\) denotes the feature \\(\\mathbf{f}_i\\) belonging to the cluster \\(\\mathbf{c}\\), and \\(\\forall i\\in [1..M - 1]:\\cos (\\mathbf{f}_i^{\\mathbf{c}},\\mathbf{c})\\geq \\cos (\\mathbf{f}_{i + 1}^{\\mathbf{c}},\\mathbf{c})\\) the Visual Grammar of the cluster c parameterized by \\(\\Theta\\) can be defined" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.785, + 0.485, + 0.879 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.883, + 0.442, + 0.898 + ], + "angle": 0, + "content": "Figure 5. The Proposed Visual Grammar Model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.612, + 0.107 + ], + "angle": 0, + "content": "as in Eqn. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.113, + 0.905, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\{\\mathbf {f} _ {i, j} ^ {\\mathbf {c}} \\} ^ {M}} [ - \\log p (\\mathbf {f} _ {1} ^ {\\mathbf {c}}, \\mathbf {f} _ {2} ^ {\\mathbf {c}}, \\dots , \\mathbf {f} _ {M} ^ {\\mathbf {c}}, \\mathbf {c}, \\Theta) ] \\\\ \\Leftrightarrow \\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\left\\{\\mathbf {f} _ {i} ^ {\\mathbf {c}} \\right\\} _ {i = 1} ^ {M}} [ - \\log p \\left(\\boldsymbol {\\Delta} _ {1} ^ {\\mathbf {c}}, \\boldsymbol {\\Delta} _ {2} ^ {\\mathbf {c}}, \\dots , \\boldsymbol {\\Delta} _ {M} ^ {\\mathbf {c}}, \\mathbf {c}, \\Theta\\right) ] \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.159, + 0.908, + 0.31 + ], + "angle": 0, + "content": "where \\(\\Delta_i^{\\mathbf{c}} = \\mathbf{f}_i^{\\mathbf{c}} - \\mathbf{c}\\). Eqn. (5) defines the visual grammar of the cluster by modeling the feature distribution of \\(\\mathbf{f}_i^{\\mathbf{c}}\\) and its cluster center \\(\\mathbf{c}\\). Let \\(\\phi : \\mathbb{R}^{(M + 1)\\times D} \\to [0,1]^M\\) be a function receiving a center \\(\\mathbf{c}\\) and a set of \\(M\\) features \\(\\{\\mathbf{f}_i\\}_{i = 1}^M\\) (\\(\\cos(\\mathbf{f}_i,\\mathbf{c}) \\geq \\cos(\\mathbf{f}_{i + 1},\\mathbf{c})\\)) to determine whether \\(\\mathbf{f}_i\\) belonging to \\(\\mathbf{c}\\), i.e., \\(\\mathbf{u} = \\phi(\\Delta_1,\\Delta_2,\\dots,\\Delta_M,\\mathbf{c})\\) where \\(\\Delta_i = \\mathbf{f}_i - \\mathbf{c}\\), \\(\\mathbf{u} = [u_1,u_2,\\dots,u_M]\\) and \\(u_i = 1\\) denotes \\(\\mathbf{f}_i\\) belong to cluster \\(\\mathbf{c}\\) and vice versa. Hence, the visual grammar model in Eqn. (5) can be modeled by the network \\(\\phi\\) as follows with parameter \\(\\Theta\\) as in Eqn. (6)." + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.317, + 0.907, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\left\\{\\mathbf {f} _ {i} \\right\\} _ {i = 1} ^ {M}} \\left[ - \\log p (\\mathbf {u} | \\boldsymbol {\\Delta} _ {1}, \\boldsymbol {\\Delta} _ {2}, \\dots , \\boldsymbol {\\Delta} _ {M}, \\mathbf {c}, \\Theta) \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Eqn. (6) aims to model the distribution of features around its cluster by learning the correlation of relatively topological structures \\(\\pmb{\\Delta}_i\\) of features \\(\\mathbf{f}_i\\) around cluster \\(\\mathbf{c}\\). Then, based on knowledge of the cluster distribution, the model \\(\\phi\\) is able to determine whether a feature \\(\\mathbf{f}_i\\) belongs to cluster \\(\\mathbf{c}\\). Hence, it is essential that the model \\(\\phi\\) has the ability to exploit the correlation between features \\(\\mathbf{f}_i\\) and cluster \\(\\mathbf{c}\\) to learn the topological structure of visual grammar. Therefore, we adopt the self-attention mechanism [41, 65] to efficiently model these feature correlations. Particularly, the model \\(\\phi\\) is formed by \\(L_{\\phi}\\) blocks of self-attention as follows," + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.516, + 0.906, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {z} _ {0} = \\mathrm {L N} ([ \\boldsymbol {\\Delta} _ {1}, \\dots , \\boldsymbol {\\Delta} _ {\\mathrm {M}}, \\mathbf {c} ]) + \\boldsymbol {\\beta}, \\quad \\mathbf {a} _ {l} = \\mathbf {z} _ {l} + \\operatorname {M H S A} (\\mathbf {z}) \\\\ \\mathbf {z} _ {l + 1} = \\mathbf {a} _ {l} + \\operatorname {M L P} \\left(\\ln \\left(\\mathbf {a} _ {l}\\right)\\right), \\quad \\mathbf {u} = \\operatorname {P r o j} \\left(\\mathbf {z} _ {L _ {\\phi}}\\right) \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.908, + 0.629 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the positional embedding, LN is Layer Normalization, MHSA is multi-head self-attention, MLP is the multi-layer perception, and Proj is the linear projection. By using Transformers, the correlation of cluster distributions can be well modeled by the self-attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Cluster Assignment via Visual Grammar. Instead of assigning the clusters based on the model prediction [3, 20, 45] or nearest cluster [30, 62] that are less effective, the cluster assignment in our approach will be performed by the visual grammar model, i.e., the visual grammar model will consider the \\(M\\) closest features around cluster \\(\\mathbf{c}\\) to assign the cluster for these features. Then, the cluster assignments are used to compute our Fairness Contrastive Clustering loss. In addition, following common practices [3, 20, 45], we improve background shift modeling by using the cluster assignments of features as the pseudo labels of pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Unknown Cluster Initialization. Prior work [30, 62] initialized a single unknown cluster \\((N_U = 1)\\), thus resulting in producing non-discriminative class-wise features. However, there should be more than a single unknown cluster \\((N_U > 1)\\) to produce discriminative features for different unknown classes. Therefore, our approach first initializes a list of potential unknown clusters at each learning step via" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15069" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "DB-SCAN [22] on the features of unknown classes extracted by the current CSS model. For the new known class \\(\\mathcal{C}^t\\), we initialize these clusters based on the mean of their feature representations. Meanwhile, the clusters of known classes learned in previous steps are maintained. Further details can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.191, + 0.363, + 0.208 + ], + "angle": 0, + "content": "3.4. Continual Learning Procedure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.485, + 0.334 + ], + "angle": 0, + "content": "Fig. 3 illustrates the training procedure of our continual learning approach. At each learning step \\( t \\), the CSS model \\( F \\) with \\( \\theta_{t} \\) is trained with the Fairness Contrastive Clustering loss defined in Eqn. (4) and the previous visual grammar model \\( \\phi \\) with \\( \\Theta_{t-1} \\). In addition, we introduce a cluster regularizer \\( \\mathcal{R}_{C} \\) to avoid the clusters of different classes collapsing into a single cluster. Therefore, the entire CSS learning objective in our approach can be formed as in Eqn. (8)." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.343, + 0.486, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\boldsymbol {\\theta} _ {t}} \\mathbb {E} _ {\\mathbf {x} ^ {t}, \\hat {\\mathbf {y}} ^ {t}} \\left[ \\mathcal {L} _ {C E} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} ^ {t}\\right) + \\lambda_ {C L} \\sum_ {\\mathbf {c} _ {i}} \\mathcal {L} _ {C o n t} ^ {\\alpha} \\left(\\mathbf {F} ^ {t}, \\mathbf {c} _ {i}\\right) + \\lambda_ {C} \\mathcal {R} _ {C} (\\mathbf {c}) \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.418 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_C(\\mathbf{c}) = \\sum_{\\mathbf{c}_i,\\mathbf{c}_j}\\{\\max (0,2\\nabla -||\\mathbf{c}_i - \\mathbf{c}_j||)\\} ^2\\) is the regularizer to avoid the cluster collapsing, \\(\\lambda_C\\) is the balanced weight, and \\(\\nabla\\) is the margin between clusters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Training Procedure of Visual Grammar Model. At CSS learning step \\( t \\), we adopt the visual grammar model trained on the previous learning step, i.e., \\( \\phi \\) with \\( \\Theta_{t-1} \\), to perform the cluster assignment for the contrastive clustering loss defined in Eqn. (2). Then, the visual grammar model \\( \\phi \\) with \\( \\Theta_t \\) at learning step \\( t \\) will be learned (initialized from \\( \\Theta_{t-1} \\)) on the features extracted from the dataset and the set of known clusters \\( \\mathbf{c} \\) up to the current learning step. Following [41], we sample a center \\( \\mathbf{c} \\) from the known clusters and its \\( M \\) closest features to train the visual grammar model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Initial Visual Grammar Model. At the first learning step \\( t = 1 \\), since no clusters have been learned initially, the visual grammar model \\( \\phi \\) with \\( \\Theta_0 \\) is unavailable. However, as common practices in CSS [3, 4, 20], the segmentation model is typically trained from a pre-trained backbone on ImageNet [18]. As a result, the features extracted at the first learning step are characterized by the ImageNet features. Therefore, we adopt this philosophy to initialize our visual grammar model \\( (\\phi \\) with \\( \\Theta_0 \\)) by pre-training the visual grammar model on the ImageNet dataset. Then, during CSS training, we will progressively train our visual grammar model at each learning step as aforementioned." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.224, + 0.781 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.455, + 0.804 + ], + "angle": 0, + "content": "4.1. Implementations and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Implementation. Following common practices [2, 20, 62], we adopt DeepLab-V3 [6] with ResNet-101 [26] and SegFormer [69] with MiT-B3 [69] in our experiments. For the Visual Grammar model, we adopt the design of [41] with \\( L_{\\phi} = 12 \\) blocks of multi-head self-attention layers. The feature vectors from the last layer of the decoder are used" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.09, + 0.892, + 0.103 + ], + "angle": 0, + "content": "Table 1. Effectiveness of Fairness Contrastive Learning Loss." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.104, + 0.902, + 0.219 + ], + "angle": 0, + "content": "
(a) ADE20K 100-50
\\( \\mathcal{L}_{Cont} \\)\\( \\mathcal{L}_{Cont}^{\\alpha} \\)0-100100-150allMajor Minor
✓44.615.234.851.5 26.4
✓44.624.537.952.1 30.8
(b) ADE20K 100-10
\\( \\mathcal{L}_{Cont} \\)\\( \\mathcal{L}_{Cont}^{\\alpha} \\)0-100100-150allMajor Minor
✓41.916.033.249.9 24.9
✓44.420.436.451.8 28.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.905, + 0.284 + ], + "angle": 0, + "content": "for our \\(\\mathcal{L}_{Cont}^{\\alpha}\\) loss. The value \\(\\alpha\\) is set individually for each dataset, i.e., \\(\\alpha = 5 \\times 10^{-2}\\) for ADE20K, \\(\\alpha = 10^{-2}\\) for VOC for Cityscapes. The details of our hyper-parameters are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.908, + 0.542 + ], + "angle": 0, + "content": "Evaluation Protocols. We evaluate models on three standard datasets of CSS, i.e., ADE20K [76], Pascal VOC [23], and Cityscapes [16]. Following common practices [3, 4, 62], our experiments are conducted on the overlapped CSS settings. In particular, on ADE20K, we use three different settings, i.e., ADE20K 100-50 (2 steps), ADE20K 100-10 (6 steps), and ADE20K 100-5 (11 steps). On Pascal VOC, we evaluate FALCON in three benchmarks, i.e., VOC 15-5 (2 steps), VOC 15-1 (6 steps), and VOC 10-1 (11 steps). On Cityscapes, we conduct domain incremental experiments with three settings, i.e., Cityscapes 11-5 (3 steps), Cityscapes 11-1 (11 steps), and Cityscapes 1-1 (21 steps). Following [3, 20], the mean Intersection over Union (mIoU) metric is adopted in our comparison, including mIoU of the last learning step on initial classes, incremental classes, and all classes. In addition, to illustrate the fairness improvement, we report the mIoU of major and minor classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.551, + 0.667, + 0.567 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.907, + 0.739 + ], + "angle": 0, + "content": "Effectiveness of Fairness Contrastive Clustering. Table 1 presents our results using DeepLab-V3 [7] with Resnet101 on ADE20K 100-50 and ADE20K 100-10 benchmarks. We evaluate the impact of the fairness contrastive clustering loss \\(\\mathcal{L}_{Cont}^{\\alpha}\\) by comparing it with the vanilla contrastive clustering loss \\(\\mathcal{L}_{Cont}\\). As shown in our results, the overall performance has been significantly improved to \\(37.9\\%\\) and \\(36.4\\%\\) on ADE20K 100-50 and ADE20K 100-10, respectively. In addition, the fairness has been promoted due to the mIoU improvement of minor groups. We also study the impact of network backbones and cluster margin \\(\\nabla\\) in our appendix." + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.742, + 0.836, + 0.754 + ], + "angle": 0, + "content": "Table 2. Effectiveness of Visual Grammar." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.755, + 0.902, + 0.9 + ], + "angle": 0, + "content": "
(a) ADE20K 100-50
0-100101-150allMajorMinor
Nearest Cluster44.311.533.451.524.3
Fixed φ44.617.635.652.027.4
Adaptive φ44.624.537.952.130.8
(b) ADE20K 100-10
0-100101-150allMajorMinor
Nearest Cluster40.114.331.548.722.9
Fixed φ43.018.534.950.627.0
Adaptive φ44.420.436.451.828.7
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15070" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.159, + 0.09, + 0.414, + 0.103 + ], + "angle": 0, + "content": "Table 3. Effectiveness of Scaling Factor \\( \\alpha \\) ." + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.106, + 0.347, + 0.117 + ], + "angle": 0, + "content": "(a) ADE20K 100-50" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.118, + 0.48, + 0.186 + ], + "angle": 0, + "content": "
α0-1516-20allMajorMinor
α = 0.143.119.835.350.627.7
α = 0.0544.624.537.952.130.8
α = 0.0143.621.336.251.028.7
α = 0.00542.418.634.550.126.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.186, + 0.347, + 0.196 + ], + "angle": 0, + "content": "(b) Pascal VOC 15-5" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.197, + 0.48, + 0.266 + ], + "angle": 0, + "content": "
α0-1516-20allMajorMinor
α = 0.174.851.669.376.963.5
α = 0.0576.251.370.379.063.8
α = 0.0179.454.873.581.367.7
α = 0.00574.648.968.577.661.6
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.486 + ], + "angle": 0, + "content": "Effectiveness of Scaling Factor of Cluster. Table 3 illustrates the experimental results of the impact of different scaling factor \\(\\alpha\\) on ADE20K 100-50 and Pascal VOC 15-5 benchmarks. As shown in Table 3, when the value of scaling factor \\(\\alpha\\) gradually decreases, the performance of our proposed approach is improved accordingly since the fairness contrastive loss in Eqn (4) tends to be more uniform across major and minor classes. However, when the scaling factor is too small (\\(\\alpha = 0.005\\)), the impact of the loss enforcement becomes weaker leading to the weaker enforcement of the fairness contrastive clustering, resulting in lower overall performance. In addition, we have observed that the higher the number of classes demands the higher the value of \\(\\alpha\\) since it will increase the compactness of more clusters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.699 + ], + "angle": 0, + "content": "Effectiveness of Loss Contributions. Table 4 illustrates the contributions of proposed learning objectives. For the model without using visual grammar, we only use a single unknown cluster \\((N_U = 1)\\) and adopt the nearest cluster strategies to assign clusters of unknown pixels. By using only cross-entropy loss, the mIoU performance remains low due to catastrophic forgetting and background shift problems. Meanwhile, with our fairness clustering loss \\(\\mathcal{L}_{Cont}^{\\alpha}\\), visual grammar model \\(\\phi\\), and the cluster regularizer \\(\\mathcal{R}\\), the mIoU performance has been significantly improved to \\(37.9\\%\\) and \\(36.4\\%\\) on ADE20K 100-50 and ADE20K 100-10, respectively. Moreover, our FALCON has significantly promoted the fairness of segmentation models illustrated by the mIoU improvement of major and minor groups." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.701, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Effectiveness of Visual Grammar. We evaluate FALCON under three settings, i.e., Nearest Cluster, Fixed \\(\\phi\\) pretrained" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.737, + 0.427, + 0.751 + ], + "angle": 0, + "content": "Table 4. Effectiveness of Our Proposed Losses." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.752, + 0.48, + 0.826 + ], + "angle": 0, + "content": "
(a) ADE20K 100-50
\\( \\mathcal{L}_{CE} \\)\\( \\mathcal{L}_{Cont}^{\\alpha } \\)\\( \\phi \\)\\( \\mathcal{R}_{C} \\)0-100101-150allMajorMinor
✓0.018.96.30.09.4
✓✓44.07.931.951.622.1
✓✓✓43.821.836.451.129.1
✓✓✓✓44.624.537.952.130.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.234, + 0.826, + 0.341, + 0.836 + ], + "angle": 0, + "content": "(b) ADE20K 100-10" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.837, + 0.48, + 0.898 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{CE} \\)\\( \\mathcal{L}_{Cont}^{\\alpha } \\)\\( \\phi \\)\\( \\mathcal{R}_C \\)0-100101-150allMajorMinor
✓0.03.51.20.01.8
✓✓39.013.130.447.821.6
✓✓✓43.418.535.151.227.1
✓✓✓✓44.420.436.451.828.7
" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "Figure 6. Cluster Distribution at Learning Step \\( t = 1 \\) of ADE20K 100-50 (classes 109-144 are future classes) without (left) and with (right) Fairness Learning via Contrastive Attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.908, + 0.495 + ], + "angle": 0, + "content": "ImageNet (without updating on each learning step), and Adaptive \\(\\phi\\) (with updating on each learning step). As in Table 2, the mIoU result using only the nearest cluster remains ineffective. Meanwhile, the adaptive visual grammar model updated at each learning step further boosts the mIoU performance and promotes fairness, i.e., increased by \\(4.5\\%\\) and \\(4.9\\%\\) on ADE20K 100-50 and ADE20K 100-10 compared to the nearest cluster approach. In addition, we study the impact of choosing the number of features \\(M\\) in the visual grammar model in our appendix. Fig. 6 illustrates the feature distributions of unknown classes (future class). As a result, our FALCON approach is able to model features of unknown classes into different clusters and produce better and more compact clusters compared to the one without Fairness Learning via Contrastive Attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.909, + 0.648 + ], + "angle": 0, + "content": "Effectiveness of Forgetting Each Learning Step. Table 5 reports the results of our model evaluated at each learning step. The performance of our model remains stable at each learning step. In particular, the maximum performance drop at learning step 4 of classes 111-120 due forgetting is only \\(0.8\\%\\). Meanwhile, at the last learning task, the mIoU results of old classes are well maintained, with a slight drop of \\(0.4\\%\\) in classes 131-140. As shown in the result, our model has effectively maintained the performance to avoid catastrophic forgetting while enabling the capability of learning classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.847, + 0.675 + ], + "angle": 0, + "content": "4.3. Comparison with Prior SoTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.909, + 0.773 + ], + "angle": 0, + "content": "ADE20K. Table 6 presents our experimental results compared to prior CSS methods. By using DeepLab-V3, our approach has achieved SoTA performance, i.e., the mIoU results of \\(37.9\\%\\) and \\(+36.4\\%\\) on ADE20K 100-50 and ADE20K 100-10 benchmarks, higher than prior FairCL [62]. Meanwhile, FALCON using Transformer has outperformed" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.782, + 0.89, + 0.797 + ], + "angle": 0, + "content": "Table 5. Effectiveness of CSS Model At Each Learning Step." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.798, + 0.902, + 0.898 + ], + "angle": 0, + "content": "
ADE20K 100-10
0-100101-110111-120121-130131-140141-150all
Step 144.9-----44.9
Step 244.815.8----42.2
Step 344.715.722.1---40.3
Step 444.615.621.326.3--39.2
Step 544.515.421.226.124.4-37.9
Step 644.415.421.226.124.015.536.4
avg40.1
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15071" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.566, + 0.145 + ], + "angle": 0, + "content": "Table 6. Comparison with Prior Methods on ADE20K Benchmarks (Note: The results of MiB [2], PLOP [20], and FairCL [62] using Transformer on ADE20K 100-5 were not reported in prior studies. The upper bound results are not trained with fairness objective)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.146, + 0.566, + 0.332 + ], + "angle": 0, + "content": "
NetworkMethodADE20K 100-50ADE20K 100-10ADE20K 100-5
0-100101-150allavg0-100101-150allavg0-100101-150allavg
DeepLab-V3PLOP [20]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [75]42.318.834.5-39.317.632.1-38.511.529.6-
REIMINDER [45]41.619.234.1-39.021.333.1-36.116.429.5-
RCIL+LGKD [73]43.325.137.2-42.220.434.9-----
FairCL [62]43.424.037.040.541.720.434.739.0----
FALCON44.624.537.941.344.420.436.440.138.016.130.737.6
Upper Bound44.328.238.9-44.328.238.9-44.328.238.9-
Mask2FormerMiB [2]37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP [20]44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer [3]44.726.238.441.240.615.632.337.439.513.630.936.5
ECLIPSE [31]45.021.737.1-43.417.434.6-43.316.334.2-
CoMasTRe [25]45.726.039.241.642.318.434.438.440.815.832.638.6
TransformerMiB [2]43.430.639.238.739.120.434.239.5----
PLOP [20]43.826.238.038.143.324.136.240.3----
FairCL [62]43.625.537.640.742.221.935.539.4----
FALCON47.530.641.943.547.326.240.342.840.818.933.538.1
Upper Bound48.739.045.5-48.739.045.5-48.739.045.5-
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.335, + 0.482, + 0.517 + ], + "angle": 0, + "content": "the prior CoMFormer [3] model by \\(+3.5\\%\\), \\(+8.0\\%\\), and \\(+2.6\\%\\) on ADE20K 100-50, ADE20K 100-10, and ADE20K 100-5. In addition, our mIoU results on the initial classes remain competitive with the upper-bounded results because FALCON can well handle the fairness problem compared to fully supervised learning. Although ECLIPSE [31] exhibits slightly better on ADE20K 100-5 by only \\(0.7\\%\\) (all) due to the better segmentation framework (i.e., Mask2Former), FALCON significantly outperforms ECLIPSE on ADE20K 100-50 and ADE20K 100-10. We also report our results of ADE20K 50-50 in the appendix. As in Fig. 7, FALCON produces better segmentation maps than prior methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.486, + 0.653 + ], + "angle": 0, + "content": "Pascal VOC. Table 7 presents our results on Pascal VOC benchmarks. FALCON has consistently achieved the SoTA performance on three benchmarks. In particular, compared to the prior FairCL [62] approach, our methods using DeepLab-V3 have improved the mIoU performance up to \\(73.50\\%\\), \\(69.83\\%\\), and \\(62.41\\%\\) on Pascal VOC 15-5, Pascal VOC 15-1, and Pascal VOC 10-1, respectively. Additionally, by using the better backbone, the performance is further improved, and the gap with the upper-bound result is reduced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Cityscapes. Table 8 reports the performance of our approach using DeepLab-V3 compared to prior methods on three different settings, i.e., Cityscapes 11-5, Cityscapes" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.099, + 0.903, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.296, + 0.902, + 0.324 + ], + "angle": 0, + "content": "Figure 7. The Comparison of Qualitative Results on ADE20K Between Our Approach and RCIL [75]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.412 + ], + "angle": 0, + "content": "11-1, and Cityscapes 1-1. As in Table 8, the performance of our methods has consistently outperformed prior FairCL [62] approach by \\(+3.78\\%\\), \\(+3.14\\%\\), and \\(+6.02\\%\\) on three benchmarks. Similar to our experiments on ADE20K and VOC, the better network brings higher results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.419, + 0.779, + 0.435 + ], + "angle": 0, + "content": "5. Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.909, + 0.638 + ], + "angle": 0, + "content": "Conclusions. This paper has presented a novel Fairness Learning via Contrastive Attention approach to CSS. In particular, the fairness contrastive clustering loss has been introduced to model both catastrophic forgetting and fairness problems. Then, the visual grammar model was presented to model the unknown classes. The experimental results on different benchmarks have shown the SoTA performance and fairness promotion of our proposed FALCON approach. Limitations. Our study chose a set of learning hyperparameters to support our theoretical analysis. However, it potentially consists of several limitations related to choosing learning parameters and cluster initialization. Further discussion of limitations is presented in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Acknowledgment This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs." + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.705, + 0.51, + 0.718 + ], + "angle": 0, + "content": "Table 7. Comparisons with Prior Methods on Pascal VOC." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.72, + 0.58, + 0.898 + ], + "angle": 0, + "content": "
MethodPascal VOC 15-5Pascal VOC 15-1Pascal VOC 10-1
0-1516-20all0-1516-20all0-1011-20all
DeepLab-V3MiB [2]76.3749.9770.0838.0013.5032.2020.0020.1020.10
PLOP [20]75.7351.7170.0965.1021.1054.6044.0015.5030.50
RCIL [75]---70.6023.7059.4055.4015.1034.30
FairCL [62]---72.0022.7060.3042.3025.6034.40
SSUL [4]77.8250.1071.2277.3136.5967.6171.3145.9859.25
FALCON79.3554.7773.5078.3442.5769.8373.9449.7362.41
Upper Bound79.7772.3577.4379.7772.3577.4378.4176.3577.43
TransformerPLOP [20]72.5148.3766.7664.5937.2358.0848.5333.7141.47
SSUL [4]79.9156.8374.4179.9140.5670.5474.0651.8563.48
FairCL [62]---73.5022.8061.5057.1014.2036.60
FALCON81.2058.0475.6978.7147.5471.2874.9252.5464.26
Upper Bound80.8474.9779.4480.8474.9779.4480.8474.9779.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.638, + 0.705, + 0.86, + 0.719 + ], + "angle": 0, + "content": "Table 8. Comparisons on Cityscapes." + }, + { + "type": "table", + "bbox": [ + 0.603, + 0.72, + 0.898, + 0.898 + ], + "angle": 0, + "content": "
Method11-511-11-1
DeepLab-V3LWF-MC [48]58.9056.9231.24
ILT [38]59.1457.7530.11
MIB [2]61.5160.0242.15
PLOP [20]63.5162.0545.24
RCIL [75]64.3063.0048.90
FairCL [62]66.9666.6149.22
FALCON70.7469.7555.24
Upper Bound79.3079.3079.30
Trans.FairCL [62]67.8567.0955.68
FALCON71.3370.1458.79
Upper Bound83.8083.8083.80
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15072" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Nikita Araslanov, and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. Conf. Comp. Vision Pattern Rec., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[3] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. IEEE/CVF Computer Vision and Pattern Recognition Conference, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[4] Sungmin Cha, beomyoung kim, YoungJoon Yoo, and Taesup Moon. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In Advances in Neural Information Processing Systems, pages 10919-10930. Curran Associates, Inc., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[5] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. arXiv preprint arXiv:2407.16354, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.492 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. TPAMI, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[8] Mayee Chen, Daniel Y Fu, Avanika Narayan, Michael Zhang, Zhao Song, Kayvon Fatahalian, and Christopher Ré. Perfectly balanced: Improving transfer and robustness of supervised contrastive learning. In International Conference on Machine Learning, pages 3090-3122. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[14] Sanghyeok Chu, Dongwan Kim, and Bohyung Han. Learning debiased and disentangled representations for semantic segmentation. Advances in Neural Information Processing Systems, 34:8355-8366, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for in-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "cremental semantic segmentation. In European Conference on Computer Vision, pages 244-261. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[16] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[17] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[19] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proc. Europ. Conf. Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4040-4050, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[21] Beyza Ermis, Giovanni Zappella, Martin Wistuba, Aditya Rawal, and Cedric Archambeau. Continual learning with transformers for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3774-3781, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.907, + 0.53 + ], + "angle": 0, + "content": "[22] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.907, + 0.573 + ], + "angle": 0, + "content": "[23] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[24] Robert French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.907, + 0.671 + ], + "angle": 0, + "content": "[25] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3857, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.907, + 0.771 + ], + "angle": 0, + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[28] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. DAFormer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[29] Ibsa Jalata, Naga Venkata Sai Raviteja Chappa, Thanh-Dat Truong, Pierce Helton, Chase Rainwater, and Khoa Luu. Eqadap: Equipollent domain adaptation approach to image deblurring. IEEE Access, 10:93203-93211, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[30] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detect" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15073" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.121 + ], + "angle": 0, + "content": "tion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5830-5840, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[31] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[32] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. PNAS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.484, + 0.402 + ], + "angle": 0, + "content": "[34] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 8547-8555, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[35] Chang Liu, Giulia Rizzoli, Pietro Zanuttigh, Fu Li, and Yi Niu. Learning from the web: Language drives weakly-supervised incremental learning for semantic segmentation. arXiv preprint arXiv:2407.13363, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[36] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[38] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCVWS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.484, + 0.617 + ], + "angle": 0, + "content": "[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[40] Pha Nguyen, Thanh-Dat Truong, Miaoqing Huang, Yi Liang, Ngan Le, and Khoa Luu. Self-supervised domain adaptation in crowd counting. In 2022 IEEE international conference on image processing (ICIP), pages 2786-2790. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[41] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10847-10856, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[43] First Ozdemir and Orcun Goksel. Extending pretrained segmentation networks with additional anatomical structures. International journal of computer assisted radiology and surgery, 14:1187-1195, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[44] First Ozdemir, Philipp Fuernstahl, and Orcun Goksel. Learn the new, keep the old: Extending pretrained models with new anatomy and images. In Medical Image Computing" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part IV 11, pages 361-369. Springer, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[45] Minh Hieu Phan, The-Anh Ta, Son Lam Phung, Long Tran-Thanh, and Abdesselam Bouzerdoum. Class similarity weighted knowledge distillation for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16866-16875, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.908, + 0.289 + ], + "angle": 0, + "content": "[46] Yiqiao Qiu, Yixing Shen, Zhuohao Sun, Yanchong Zheng, Xiaobin Chang, Weishi Zheng, and Ruixuan Wang. Sats: Self-attention transfer for continual semantic segmentation. Pattern Recognition, 138:109383, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.908, + 0.373 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.908, + 0.43 + ], + "angle": 0, + "content": "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proc. Conf. Comp. Vision Pattern Rec., 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.908, + 0.473 + ], + "angle": 0, + "content": "[49] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[50] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.908, + 0.531 + ], + "angle": 0, + "content": "[51] Mohammad Rostami. Lifelong domain adaptation via consolidated internal distribution. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.908, + 0.602 + ], + "angle": 0, + "content": "[52] Antoine Saporta, Arthur Douillard, Tuan-Hung Vu, Patrick Pérez, and Matthieu Cord. Multi-head distillation for continual unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.908, + 0.658 + ], + "angle": 0, + "content": "[53] Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. Incremental learning of object detectors without catastrophic forgetting. In Proceedings of the IEEE international conference on computer vision, pages 3400-3409, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.908, + 0.716 + ], + "angle": 0, + "content": "[54] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.908, + 0.785 + ], + "angle": 0, + "content": "[55] Attila Szabó, Hadi Jamali-Rad, and Siva-Datta Mannava. Tilted cross-entropy (tce): Promoting fairness in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2305–2310, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[56] Sebastian Thrun. Lifelong learning algorithms. In *Springer Learning to Learn*, 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.908, + 0.858 + ], + "angle": 0, + "content": "[57] Thanh-Dat Truong and Khoa Luu. Cross-view action recognition understanding from exocentric to egocentric perspective. Neurocomputing, 614:128731, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[58] Thanh-Dat Truong, Chi Nhan Duong, Khoa Luu, Minh-Triet Tran, and Ngan Le. Domain generalization via universal non-volume preserving approach. In 2020 17th Conference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15074" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "on Computer and Robot Vision (CRV), pages 93-100. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[59] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Son Lam Phung, Chase Rainwater, and Khoa Luu. Bimal: Bijective maximum likelihood approach to domain adaptation in semantic scene segmentation. In International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[60] Thanh-Dat Truong, Ravi Teja Nvs Chappa, Xuan-Bac Nguyen, Ngan Le, Ashley PG Dowling, and Khoa Luu. Otadapt: Optimal transport-based approach for unsupervised domain adaptation. In 2022 26th international conference on pattern recognition (ICPR), pages 2850-2856. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[61] Thanh-Dat Truong, Ngan Le, Bhiksha Raj, Jackson Cothren, and Khoa Luu. Freedom: Fairness domain adaptation approach to semantic scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[62] Thanh-Dat Truong, Hoang-Quan Nguyen, Bhiksha Raj, and Khoa Luu. Fairness continual learning approach to semantic scene understanding in open-world environments. In NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[63] Thanh-Dat Truong, Pierce Helton, Ahmed Moustafa, Jackson David Cothren, and Khoa Luu. Conda: Continual unsupervised domain adaptation learning in visual perception for self-driving cars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 5642-5650, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[64] Thanh-Dat Truong, Utsav Prabhu, Dongyi Wang, Bhiksha Raj, Susan Gauch, Jeyamkondan Subbiah, and Khoa Luu. EAGLE: Efficient adaptive geometry-based learning in cross-view understanding. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[66] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proc. Conf. Comp. Vision Pattern Rec., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[67] Huyong Wang, Huisi Wu, and Jing Qin. Incremental nuclei segmentation from histopathological images via future-class awareness and compatibility-inspired distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11408-11417, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[68] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[70] Zhengyuan Xie, Haiquan Lu, Jia-wen Xiao, Enguang Wang, Le Zhang, and Xialei Liu. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tation. In European Conference on Computer Vision, pages 183-201. Springer, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[71] Lei Yang, Xiaohang Zhan, Dapeng Chen, Junjie Yan, Chen Change Loy, and Dahua Lin. Learning to cluster faces on an affinity graph. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2298-2306, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[72] Lei Yang, Dapeng Chen, Xiaohang Zhan, Rui Zhao, Chen Change Loy, and Dahua Lin. Learning to cluster faces via confidence and connectivity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13369-13378, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.907, + 0.345 + ], + "angle": 0, + "content": "[73] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[74] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. arXiv preprint arXiv:2407.09838, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[75] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.907, + 0.502 + ], + "angle": 0, + "content": "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.907, + 0.573 + ], + "angle": 0, + "content": "[77] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15075" + } + ] +] \ No newline at end of file diff --git a/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_origin.pdf b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..731a1444dd13d35f387adc9ba64a462a378cb6e6 --- /dev/null +++ b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e05db4ebb3337cd196eee53893a7819531e0d452cd321f6d564e15f9538dc5e +size 9244745 diff --git a/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/full.md b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..100988cf39c190a44db9235f10ccb16ff13a9398 --- /dev/null +++ b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/full.md @@ -0,0 +1,333 @@ +# FALCON: Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding + +Thanh-Dat Truong $^{1}$ , Utsav Prabhu $^{2}$ , Bhiksha Raj $^{3,4}$ , Jackson Cothren $^{5}$ , Khoa Luu $^{1}$ $^{1}$ CVIU Lab, University of Arkansas, USA $^{2}$ Google DeepMind, USA + $^{3}$ Carnegie Mellon University, USA $^{4}$ Mohammed bin Zayed University of AI, UAE + $^{5}$ Dep. of Geosciences, University of Arkansas, USA + +{tt032, jcothre, khoaluu}@uark.edu, bhiksha@cs.cmu.edu, utsavprabhu@google.com http://uark-cviu.github.io/projects/FALCON + +# Abstract + +Continual Learning in semantic scene segmentation aims to continually learn new unseen classes in dynamic environments while maintaining previously learned knowledge. Prior studies focused on modeling the catastrophic forgetting and background shift challenges in continual learning. However, fairness, another major challenge that causes unfair predictions leading to low performance among major and minor classes, still needs to be well addressed. In addition, prior methods have yet to model the unknown classes well, thus resulting in producing non-discriminative features among unknown classes. This work presents a novel Fairness Learning via Contrastive Attention Approach to continual learning in semantic scene understanding. In particular, we first introduce a new Fairness Contrastive Clustering loss to address the problems of catastrophic forgetting and fairness. Then, we propose an attention-based visual grammar approach to effectively model the background shift problem and unknown classes, producing better feature representations for different unknown classes. Through our experiments, our proposed approach achieves State-of-the-Art (SoTA) performance on different continual learning benchmarks, i.e., ADE20K, Cityscapes, and Pascal VOC. It promotes the fairness of the continual semantic segmentation model. + +# 1. Introduction + +The semantic segmentation networks, e.g., Transformers [69] and Convolutional Neural Networks [7], learned from data with a closed-set of known classes, have shown outstanding performance. However, they often suffer performance degradation when encountering novel objects or classes in new dynamic environments [2, 20, 62, 63]. To improve their performance, several transfer learning and domain adaptation methods [1, 28, 29, 40, 57-60, 62, 64] were introduced to adapt trained models into deployed environments. While + +![](images/c59bd327f32f50759195ae02cd7f7ed871e4b8b933cae3b50922ea91065c27b3.jpg) +Figure 1. Our Fairness Learning via Contrastive Attention to Continual Semantic Segmentation. The Fairness Contrastive Clustering Loss promotes the fairness of the model while the Attention-based Visual Grammar models the unknown classes. + +the former often aims to fine-tune the model on labeled data collected in the new environments, the latter adapts the model to the new domains in an unsupervised manner [1, 28]. However, these methods cannot handle novel objects well due to their close-set learning. In practice, the semantic segmentation models should be able to adaptively and continually learn the new knowledge of novel classes. It motivates the development of Continual Learning paradigm [2, 20, 46], a.k.a, Continual Semantic Segmentation (CSS), where the segmentation models are learned sequentially to new contents of data. + +Far apart from prior segmentation methods [7, 69] that learn one time on static, closed-set data, Continual Learning requires the segmentation models to learn from dynamic, open-set data [20, 62]. In a particular scenario, accessing previous learning data is restricted due to privacy concerns. In CSS, three challenges have been identified, including (1) Catastrophic Forgetting, (2) Background Shift, and (3) Fairness. While the catastrophic forgetting problem [24, 50, 56] depicts the segmentation model tends to forget its knowledge when learning new data, background shift indicates the problem of classes of previous or future data (unknown classes) + +![](images/7477b3592d42ef495fc1f16284bddfe5ac5f7b390885caa6d5959143cf357d37.jpg) +Figure 2. The Data Class Distribution of ADE20K. The major classes occupy more than $75\%$ of the total pixels of the dataset. + +have collapsed into a background class [2, 20, 75]. Prior methods [4, 20, 46, 62] addressed these two problems by introducing knowledge distillation and pseudo labels. However, these methods can not handle unknown classes since they either consider these unknown classes as a background class or assign unknown pixels by a pseudo label of prior known classes [20, 62]. More importantly, the last problem, fairness, is a significant challenge that limits the performance of CSS models (Fig. 1). + +As shown in Fig. 2, the number of pixels of each class in training data have been imbalanced among classes and significantly decreased after each task. Thus, this bias influences the learning procedure and model predictions that later cause unfair class predictions. However, limited studies are taking the fairness problem into account. [61] presented a similar problem in domain adaptation and extended it to continual learning [62]. These methods rely on the assumption of ideal fair or balanced data distributions. However, it is not applicable in practice since the size, i.e., the number of pixels, of several classes can never be more significant than others. For example, the bottle size should not be more significant than the size of a car. Meanwhile, the current knowledge distillation methods [3, 20, 45] in CSS are unable to handle the fairness problem since they focus on modeling catastrophic forgetting and background shift problems. Therefore, it is essential to develop a new CSS approach to address these limitations. + +Contributions of This Work. This work presents a novel Fairness Learning via Contrastive Attention Approach (FALCON) to Continual Semantic Segmentation (as shown in Fig. 1). Our contributions can be summarized as follows. First, we introduce a novel Contrastive Clustering Paradigm approach to Continual Learning that models the catastrophic forgetting problem. Second, by analyzing the limitation of vanilla Contrastive Clustering in biased data, we introduce a novel Fairness Contrastive Clustering loss to model the fairness problem in continual learning efficiently. Third, to effectively model the background shift problem, we introduce a new Attention-based Visual Grammar that model the topological structures of feature distribution to handle the unknown classes effectively. Finally, the ablation studies + +illustrate the effectiveness of the proposed approach in different aspects of fairness promotion in CSS models. Compared with prior methods, our approach achieves SoTA performance on different settings of three standard benchmarks of CSS, including ADE20K, Pascal VOC, and Cityscapes. + +# 2. Related Work + +Continual Semantic Segmentation. Several studies were introduced to address catastrophic forgetting and background shift problems [19-21, 32, 37, 39, 43, 44]. The common CSS approach adopts knowledge distillation [20] and pseudo labels [2] to model catastrophic forgetting and background shift, respectively. Later, it was further improved by decoupling knowledge representations [75], modeling the inter- and intra-class knowledge [15, 46], distinguishing the feature representations of the future classes [4, 45, 67, 70], reducing background confusion [73, 74], using visual prompting [31, 35], or modeling distillation loss via the geodesic flow [54]. Another approach [3, 5, 25] adopted the mask-based segmentation networks [12, 13] to improve the performance. Recent studies have introduced CSS under the unsupervised domain adaptation settings [51, 52, 63, 66]. However, prior studies have yet to well model different unknown classes. Particularly, as previous methods [3, 20] use the pseudo labels to model the unknown classes, the future classes will be treated as a single background class. Then, Joseph et al. [30] improves the unknown class modeling using clustering but this method considers all different unknown classes as a single cluster, leading to the non-discriminative features among unknown classes. + +Contrastive Learning (CL). CL is a common learning approach [9, 10, 42] to structure the deep feature representations in the deep latent space. Oorde et al. [42] first introduced the Noise-Contrastive Estimation (InfoNCE) learning framework. Then, Chen et al. [9] presented SimCLR, a self-supervised contrastive learning approach to improve the representation power of Residual Networks. He [27] proposed a Momentum Contrast framework for unsupervised representation learning. Later, it was further improved by using MLP projection head [10] and extended to improve the self-supervised training process of vision transformers [11]. Cui et al. [17] introduced a supervised parametric contrastive learning loss to address the long-tailed recognition. Li et al. [34] adopted contrastive learning to develop the one-stage online contrastive clustering method. Radford et al. [47] presents a contrastive framework to learn the vision-language model. Later, several methods also adopted this framework for vision-language pretraining [33, 47]. + +Imbalanced and Fairness Learning. The early methods utilized the balanced Softmax loss [49] to alleviate the impact of imbalanced data distribution. Later, Wang et al. [68] introduced a Seesaw loss to re-balance the contributions of positive and negative instances via the mitigation + +and compensation modules. Ziwei et al. [36] introduced a dynamic meta-embedding to model the imbalanced classification problem. Chu et al. [14] reduce the bias in the segmentation model by presenting a new stochastic training scheme. Szabo et al. [55] presented a tilted cross-entropy loss to promote class-relevant fairness. However, there are limited studies that address the fairness problem in CSS. Truong et al. [61] introduced a fairness domain adaptation approach to semantic segmentation and later extended it into continual learning setting [62]. However, these methods [61, 62] rely on the assumption of ideal balanced data, which could not be achieved by nature. To address the limitations in prior work, this paper will introduce a novel approach to effectively model the fairness problem and unknown classes in the continual learning setting. + +# 3. The Proposed FALCON Approach + +CSS aims to learn a segmentation network $F$ on sequence data $\mathcal{D} = \{\mathcal{D}^1,\dots,\mathcal{D}^T\}$ where $T$ is the number of learning steps. At learning step $t$ , the model $F$ encounters a dataset $\mathcal{D}_t = \{(\mathbf{x}^t,\hat{\mathbf{y}}^t)\}$ where $\mathbf{x}^t\in \mathbb{R}^{H\times W\times 3}$ is the image and $\mathbf{y}\in \mathbb{R}^{H\times W}$ is a segmentation label of $\mathbf{x}^t$ . The ground truths at learning step $t$ only consist of current classes $\mathcal{C}^t$ , while the class labels of the previous $\mathcal{C}^{1\dots t - 1}$ and future steps $\mathcal{C}^{t + 1\dots T}$ are collapsed into a background class. Formally, learning the CSS model at step $t$ can be formed as Eqn. (1). + +$$ +\theta_ {t} ^ {*} = \arg \min _ {\theta_ {t}} \mathbb {E} _ {\mathbf {x} ^ {t}, \hat {\mathbf {y}} ^ {t} \in \mathcal {D} ^ {t}} \left[ \mathcal {L} _ {C E} \left(\mathbf {y} ^ {t}, \hat {\mathbf {y}} ^ {t}\right) + \lambda_ {C L} \mathcal {L} _ {C L} \left(F (\mathbf {x} ^ {t})\right) \right] (1) +$$ + +where, $\mathbf{y}^t = F(\mathbf{x}^t,\theta_t)$ , $\theta_{t}$ is the parameter of $F$ at current learning step $t$ , $\mathcal{L}_{CE}$ is the cross-entropy loss, $\lambda_{CL}$ is the balanced weight. and $\mathcal{L}_{CL}$ is the CSS objective. At learning step $t$ , the segmentation model $F$ is required to be able to predict both previously learned classes $\mathcal{C}^{1\dots t - 1}$ and current new classes $\mathcal{C}^t$ . Under this learning scenario, three challenges have been identified, i.e., Catastrophic Forgetting, Background Shift, and Fairness. Several prior methods were presented to model the two first issues in CSS using knowledge distillation [3, 20]. The last issue has not been well addressed yet due to its challenges [62]. Prior methods [2, 3, 20, 53, 75] adopt knowledge distillation to design $\mathcal{L}_{CL}$ . However, this method prevents the CSS model from diverging knowledge learned previously, therefore resulting in limiting the ability to adopt new knowledge [62]. In addition, these methods have not addressed fairness and background shift problems due to their dedicated design for maintaining knowledge via distillation [3, 20, 45]. Therefore, to address these problems, we introduce a novel Fairness Learning via Contrastive Attention Approach to CSS. + +# 3.1. Continual Learning via Contrastive Clustering + +Apart from prior methods [3, 20, 45], our CSS (Fig. 3) is defined as Contrastive Clustering Learning. Given a set of centroid vectors $\{\mathbf{c}_i\}_{i=1}^{N_K + N_U}$ where $N_K = |\mathcal{C}^{1..t}|$ and $N_U$ + +is the number of known and unknown classes up to current learning tasks. Prior work [3, 30, 62] often defined the number of unknown classes as 1 where background classes are considered as a single unknown class. Formally, our Contrastive Clustering Learning for CSS can be defined as in Eqn. (2). + +$$ +\begin{array}{l} \mathcal {L} _ {C L} \left(F (\mathbf {x} ^ {t})\right) = \sum_ {\mathbf {c} _ {i}} \mathcal {L} _ {C o n t} (\mathbf {F} ^ {t}, \mathbf {c} _ {i}) \\ = \sum_ {\mathbf {c} _ {i}} \sum_ {h, w} - \phi \left(\mathbf {f} _ {h, w} ^ {t}, \mathbf {c} _ {i}\right) \log \frac {\exp \left(\mathbf {f} _ {h , w} ^ {t} \times \mathbf {c} _ {i}\right)}{\sum_ {\mathbf {f} ^ {\prime}} \exp \left(\mathbf {f} ^ {\prime} \times \mathbf {c} _ {i}\right)} \tag {2} \\ \end{array} +$$ + +where $\mathbf{F}^t\in \mathbb{R}^{H\times W\times D}$ is the feature maps extracted from the input image $\mathbf{x}^t$ by the segmentation network $F$ , $\mathbf{f}_{h,w}^t\in \mathbb{R}^D$ is the feature at the pixel location $(h,w)$ of features $\mathbf{F}^t$ , $\sum_{\mathbf{f}'}$ means the summation over all feature representations $\mathbf{f}'\in \mathbb{R}^D$ , and $\phi :\mathbb{R}^D\times \mathbb{R}^D\to [0,1]$ is the function that determines either $\mathbf{f}_{h,w}^t$ belongs to the cluster $\mathbf{c}_i$ or not. + +By defining CSS as contrastive clustering learning, the knowledge of the segmentation model has been well maintained via the cluster vectors $\mathbf{c}$ . Then, minimizing Eqn. (2) will separate the representations of different classes while gathering the features of the same class into the same cluster. As the cluster vectors $\mathbf{c}$ of the old classes $\mathcal{C}^{1..t-1}$ have been well learned to represent their knowledge, these vectors are frozen at learning step $t$ to maintain the knowledge representations of previous classes to address the catastrophic forgetting problem. To effectively learn cluster vectors $\mathbf{c}$ , the cluster vector $\mathbf{c}$ will periodically updated after each $K$ steps by the momentum update [27, 30, 62] based on the features $\mathbf{f}_{h,w}^t$ assigned to cluster $\mathbf{c}$ . However, there are two major problems in contrastive clustering learning. First, since the training data suffer the bias among classes as shown in Fig. 2, this bias will influence Eqn. (2) and cause the unfair predictions. Second, as the function $\phi$ requires the labels to determine the features belonging to clusters, it limits the ability to model the unknown classes where the labels are unavailable. Therefore, Secs. 3.2-3.3 will present a novel approach to tackle these problems. + +# 3.2. Fairness Contrastive Clustering Learning + +While contrastive clustering learning in Eqn. (2) promotes the compact representations of features around their clusters, inspired by [8, 17, 77], we observe that the imbalanced class distribution will influence unfair behaviors among classes. In particular, for simplicity, we consider $\{\mathbf{f}_i^t\}_{i=1}^L$ is the set of features that belong to the cluster $\mathbf{c}$ at learning step $t$ (i.e., $\phi(\mathbf{f}_i^t, \mathbf{c}) = 1$ ) and $L$ is the number of features (in this case, $L$ is the total number of pixels belong to the class of cluster $\mathbf{c}$ ). Let us define the enforcement between the feature $\mathbf{f}_t^t$ and the cluster $\mathbf{c}$ as $\ell_i = \frac{\exp(\mathbf{f}_i^t \times \mathbf{c})}{\sum_{\mathbf{f}' \in L} \exp(\mathbf{f}' \times \mathbf{c})}$ . Hence, the lower the value of the enforcement $\ell_i$ is, the more compact the representation of visual features and clusters is. Then, the contrastive clustering learning loss in Eqn. (2) of entire + +![](images/c942f7a1568715b6935ae538020f155f335c8fb9a5ba1c3e0c5aa88913f00cd3.jpg) +Figure 3. The Proposed Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding. + +cluster $\mathbf{c}$ can be defined as in Eqn. (3). + +$$ +\mathcal {L} _ {\text {C o n t}} (;;, \mathbf {c}) = - \sum_ {i = 1} ^ {L} \log \frac {\exp \left(\mathbf {f} _ {i} ^ {t} \times \mathbf {c}\right)}{\sum_ {\mathbf {f} ^ {\prime}} \exp \left(\mathbf {f} ^ {\prime} \times \mathbf {c}\right)} = - \sum_ {i = 1} ^ {L} \log \ell_ {i} \tag {3} +$$ + +Proposition 1. If the contrastive clustering loss $\mathcal{L}_{Cont}(:,\mathbf{c})$ achieves the optimal value, the enforcement $\ell_{i}$ between the feature and the cluster will converge to $\ell_{i} = L^{-1}$ . + +Proposition 1 has implied that the class with more samples will result in a lower value of the enforcement and produce a more compact representation while the class having fewer samples will be more scattered in the feature space due to the higher value of the enforcement. In particular, let $L_{major}$ and $L_{minor}$ be the number of samples of the major and minor class where $L_{major} >> L_{minor}$ . Then, based on Proposition 1, the enforcement between features and the cluster of the major class will be significantly lower than the one of the minor class, i.e., $L_{major}^{-1} << L_{minor}^{-1}$ . Therefore, a direct adoption of the contrastive clustering loss in Eqn. (2) will result in an unfair CSS model. In addition, for classes in the minority group, the lower value of $L$ results in the feature presentations of classes being far away from their clusters. Thus, the model will produce non-discriminative features compared to the ones in the majority group. Moreover, if the loss is applied to the cases of unknown labels, these feature representations can be scattered in the latent space and pulled into the incorrect clusters due to weak enforcement between features and clusters (Fig. 4). + +To address the unfair problem in contrastive clustering learning, inspired by [8, 17, 77], we introduce a scaling factor $\alpha$ and a learnable transition vector $\mathbf{v}$ for each cluster $\mathbf{c}$ (all clusters have the same value of $\alpha$ but different vector $\mathbf{v}$ ). Our Fairness Contrastive Clustering Learning Loss for the entire cluster in Eqn. (3) can be re-formed as follows, + +$$ +\mathcal {L} _ {\text {C o n t}} ^ {\alpha} (;, \mathbf {c}) = - \alpha \sum_ {i = 1} ^ {L} \log \frac {\exp \left(\mathbf {f} _ {i} ^ {t} \times \mathbf {c}\right)}{\sum_ {\mathbf {f} ^ {\prime}} \exp \left(\mathbf {f} ^ {\prime} \times \mathbf {c}\right)} - \log \frac {\exp (\mathbf {v} \times \mathbf {c})}{\sum_ {\mathbf {f} ^ {\prime}} \exp \left(\mathbf {f} ^ {\prime} \times \mathbf {c}\right)} \tag {4} +$$ + +Intuitively, the scaling factor $\alpha$ will help to re-scale the impact of the enforcement in learning, and the transitive vector $\mathbf{v}$ assists in translating the center cluster into the proper position of the latent space. This action promotes the compactness of clusters in the minority group. + +Proposition 2. If the fairness contrastive clustering loss $\mathcal{L}_{Cont}^{\alpha}(:,\mathbf{c})$ achieves the optimal value, the enforcement $\ell_{i}$ between the feature and the cluster will converge to $\ell_{i} = (\alpha^{-1} + L)^{-1}$ . Proofs of Propositions 1-2 are in the appendix. + +Under the Proposition 2, when the value of $\alpha$ is small, the divergence of the enforcement between major and minor classes will be smaller, i.e., $||(\alpha^{-1} + L_{major})^{-1} - (\alpha^{-1} + L_{minor})^{-1}|| < ||L_{major}^{-1} - L_{minor}^{-1}||$ . Fig. 4 has illustrated the impact of fairness contrastive clustering loss. Therefore, our designed proposed fairness contrastive loss has effectively addressed the fairness issue in Eqn. (2). It should be noted that although the smaller $\alpha$ results in the fairer enforcement varied from major to minor classes. However, if the value of scaling factor $\alpha$ is too small, the contrastive clustering loss will rely more on the enforcement of the transitive vector $\mathbf{v}$ , and the distribution of features $\mathbf{f}_i^t$ around its cluster $\mathbf{c}$ will be scattered due the weak enforcement caused by small $\alpha$ . Therefore, the value of scaling factor $\alpha$ in practice should be carefully selected. + +# 3.3. An Efficient Unknown Class Modeling + +An ideal CSS approach must be able to model the unknown classes without supervision, especially in open-set contexts + +![](images/fd774369ab042b69c737d32bebdef9bed92f826bef9d7b9b940de6ea8260a0d3.jpg) +Figure 4. The Enforcement Loss of Contrastive Clustering $\mathcal{L}_{Cont}$ and Fairness Contrastive Clustering $\mathcal{L}_{Cont}^{\alpha}$ on Pascal VOC. Since $\mathcal{L}_{Cont}$ suffers severe biased, its clusters of minor classes remain scattered. Our $\mathcal{L}_{Cont}^{\alpha}$ produces a more uniform loss among classes, promotes fairness and compactness of clusters. + +[30, 62] where there could be multiple unknown classes or objects. Prior studies have adopted the pseudo-label strategies [3, 20] based on the model predictions to assign labels for seen classes while unseen classes have been ignored, thus resulting in non-discriminative features. [30, 62] improved the background modeling by using an additional prototypical representation for unknown classes. However, these approaches consider different unknown classes as one (i.e., $N_U = 1$ ) resulting in non-distinguished representations of different unknown classes. Thus, modeling function $\phi$ in Eqn. (2) without supervision of different unknown classes (i.e., $N_U > 1$ ) is challenging. + +Although modeling $\phi$ to determine the single feature $\mathbf{f}$ belonging to the cluster $\mathbf{c}$ is challenging, prior studies in clustering [41, 71, 72] have suggested that determine a set of features $\{\mathbf{f}_i^t\}_{i = 1}^M$ belonging to cluster $\mathbf{c}$ should be easier. This derives from the fact that even though the feature representations of different classes are different, the distributions of features around its cluster (termed as Visual Grammar) in the feature space should be similar among classes or clusters. As a result, by learning the distribution of features and their clusters, the model $\phi$ can determine whether a feature belongs to a cluster. Then, by learning the model $\phi$ on prior known clusters and features, the knowledge of $\phi$ can be adaptively applied to unknown clusters. Fig. 5 illustrates our visual grammar model of the cluster distributions. + +Limitations of Prior Clustering Methods. The traditional methods in clustering, e.g., KNN or density-based clustering [22], remain limited to noisy features leading to producing the incorrect cluster assignment. Meanwhile, the modern clustering methods, e.g., Graph Neural Networks (GNNs) [71, 72], require a large memory to build the affinity graph for clusters. In addition, GNNs often learn the local structures of graphs (or clusters) and accumulate them via the aggregation layers. Hence, the global structures of the clusters, i.e., visual grammar, are not well modeled by GNNs [41]. Therefore, to address these limitations, we introduced a new Attention-based Visual Grammar approach to efficiently model the distribution of features and their clusters via the self-attention mechanism [65]. + +Remark 1: Given a center $\mathbf{c}$ and a set of $M$ features $\{\mathbf{f}_i^{\mathbf{c}}\}_{i = 1}^{M}$ where $\mathbf{f}_i^{\mathbf{c}}$ denotes the feature $\mathbf{f}_i$ belonging to the cluster $\mathbf{c}$ , and $\forall i\in [1..M - 1]:\cos (\mathbf{f}_i^{\mathbf{c}},\mathbf{c})\geq \cos (\mathbf{f}_{i + 1}^{\mathbf{c}},\mathbf{c})$ the Visual Grammar of the cluster c parameterized by $\Theta$ can be defined + +![](images/6384771357077251daa8b9897a43d1e44dd7c7be51c270b49d96b82029fcafe5.jpg) +Figure 5. The Proposed Visual Grammar Model. + +as in Eqn. (5). + +$$ +\begin{array}{l} \Theta^ {*} = \arg \min _ {\Theta} \mathbb {E} _ {\mathbf {c}, \{\mathbf {f} _ {i, j} ^ {\mathbf {c}} \} ^ {M}} [ - \log p (\mathbf {f} _ {1} ^ {\mathbf {c}}, \mathbf {f} _ {2} ^ {\mathbf {c}}, \dots , \mathbf {f} _ {M} ^ {\mathbf {c}}, \mathbf {c}, \Theta) ] \\ \Leftrightarrow \Theta^ {*} = \arg \min _ {\Theta} \mathbb {E} _ {\mathbf {c}, \left\{\mathbf {f} _ {i} ^ {\mathbf {c}} \right\} _ {i = 1} ^ {M}} [ - \log p \left(\boldsymbol {\Delta} _ {1} ^ {\mathbf {c}}, \boldsymbol {\Delta} _ {2} ^ {\mathbf {c}}, \dots , \boldsymbol {\Delta} _ {M} ^ {\mathbf {c}}, \mathbf {c}, \Theta\right) ] \tag {5} \\ \end{array} +$$ + +where $\Delta_i^{\mathbf{c}} = \mathbf{f}_i^{\mathbf{c}} - \mathbf{c}$ . Eqn. (5) defines the visual grammar of the cluster by modeling the feature distribution of $\mathbf{f}_i^{\mathbf{c}}$ and its cluster center $\mathbf{c}$ . Let $\phi : \mathbb{R}^{(M + 1)\times D} \to [0,1]^M$ be a function receiving a center $\mathbf{c}$ and a set of $M$ features $\{\mathbf{f}_i\}_{i = 1}^M$ ( $\cos(\mathbf{f}_i,\mathbf{c}) \geq \cos(\mathbf{f}_{i + 1},\mathbf{c})$ ) to determine whether $\mathbf{f}_i$ belonging to $\mathbf{c}$ , i.e., $\mathbf{u} = \phi(\Delta_1,\Delta_2,\dots,\Delta_M,\mathbf{c})$ where $\Delta_i = \mathbf{f}_i - \mathbf{c}$ , $\mathbf{u} = [u_1,u_2,\dots,u_M]$ and $u_i = 1$ denotes $\mathbf{f}_i$ belong to cluster $\mathbf{c}$ and vice versa. Hence, the visual grammar model in Eqn. (5) can be modeled by the network $\phi$ as follows with parameter $\Theta$ as in Eqn. (6). + +$$ +\Theta^ {*} = \arg \min _ {\Theta} \mathbb {E} _ {\mathbf {c}, \left\{\mathbf {f} _ {i} \right\} _ {i = 1} ^ {M}} \left[ - \log p (\mathbf {u} | \boldsymbol {\Delta} _ {1}, \boldsymbol {\Delta} _ {2}, \dots , \boldsymbol {\Delta} _ {M}, \mathbf {c}, \Theta) \right] \tag {6} +$$ + +Eqn. (6) aims to model the distribution of features around its cluster by learning the correlation of relatively topological structures $\pmb{\Delta}_i$ of features $\mathbf{f}_i$ around cluster $\mathbf{c}$ . Then, based on knowledge of the cluster distribution, the model $\phi$ is able to determine whether a feature $\mathbf{f}_i$ belongs to cluster $\mathbf{c}$ . Hence, it is essential that the model $\phi$ has the ability to exploit the correlation between features $\mathbf{f}_i$ and cluster $\mathbf{c}$ to learn the topological structure of visual grammar. Therefore, we adopt the self-attention mechanism [41, 65] to efficiently model these feature correlations. Particularly, the model $\phi$ is formed by $L_{\phi}$ blocks of self-attention as follows, + +$$ +\begin{array}{l} \mathbf {z} _ {0} = \mathrm {L N} ([ \boldsymbol {\Delta} _ {1}, \dots , \boldsymbol {\Delta} _ {\mathrm {M}}, \mathbf {c} ]) + \boldsymbol {\beta}, \quad \mathbf {a} _ {l} = \mathbf {z} _ {l} + \operatorname {M H S A} (\mathbf {z}) \\ \mathbf {z} _ {l + 1} = \mathbf {a} _ {l} + \operatorname {M L P} \left(\ln \left(\mathbf {a} _ {l}\right)\right), \quad \mathbf {u} = \operatorname {P r o j} \left(\mathbf {z} _ {L _ {\phi}}\right) \tag {7} \\ \end{array} +$$ + +where $\beta$ is the positional embedding, LN is Layer Normalization, MHSA is multi-head self-attention, MLP is the multi-layer perception, and Proj is the linear projection. By using Transformers, the correlation of cluster distributions can be well modeled by the self-attention mechanism. + +Cluster Assignment via Visual Grammar. Instead of assigning the clusters based on the model prediction [3, 20, 45] or nearest cluster [30, 62] that are less effective, the cluster assignment in our approach will be performed by the visual grammar model, i.e., the visual grammar model will consider the $M$ closest features around cluster $\mathbf{c}$ to assign the cluster for these features. Then, the cluster assignments are used to compute our Fairness Contrastive Clustering loss. In addition, following common practices [3, 20, 45], we improve background shift modeling by using the cluster assignments of features as the pseudo labels of pixels. + +Unknown Cluster Initialization. Prior work [30, 62] initialized a single unknown cluster $(N_U = 1)$ , thus resulting in producing non-discriminative class-wise features. However, there should be more than a single unknown cluster $(N_U > 1)$ to produce discriminative features for different unknown classes. Therefore, our approach first initializes a list of potential unknown clusters at each learning step via + +DB-SCAN [22] on the features of unknown classes extracted by the current CSS model. For the new known class $\mathcal{C}^t$ , we initialize these clusters based on the mean of their feature representations. Meanwhile, the clusters of known classes learned in previous steps are maintained. Further details can be found in the appendix. + +# 3.4. Continual Learning Procedure + +Fig. 3 illustrates the training procedure of our continual learning approach. At each learning step $t$ , the CSS model $F$ with $\theta_{t}$ is trained with the Fairness Contrastive Clustering loss defined in Eqn. (4) and the previous visual grammar model $\phi$ with $\Theta_{t-1}$ . In addition, we introduce a cluster regularizer $\mathcal{R}_{C}$ to avoid the clusters of different classes collapsing into a single cluster. Therefore, the entire CSS learning objective in our approach can be formed as in Eqn. (8). + +$$ +\arg \min _ {\boldsymbol {\theta} _ {t}} \mathbb {E} _ {\mathbf {x} ^ {t}, \hat {\mathbf {y}} ^ {t}} \left[ \mathcal {L} _ {C E} \left(\mathbf {y} ^ {t}, \hat {\mathbf {y}} ^ {t}\right) + \lambda_ {C L} \sum_ {\mathbf {c} _ {i}} \mathcal {L} _ {C o n t} ^ {\alpha} \left(\mathbf {F} ^ {t}, \mathbf {c} _ {i}\right) + \lambda_ {C} \mathcal {R} _ {C} (\mathbf {c}) \right] \tag {8} +$$ + +where $\mathcal{R}_C(\mathbf{c}) = \sum_{\mathbf{c}_i,\mathbf{c}_j}\{\max (0,2\nabla -||\mathbf{c}_i - \mathbf{c}_j||)\} ^2$ is the regularizer to avoid the cluster collapsing, $\lambda_C$ is the balanced weight, and $\nabla$ is the margin between clusters. + +Training Procedure of Visual Grammar Model. At CSS learning step $t$ , we adopt the visual grammar model trained on the previous learning step, i.e., $\phi$ with $\Theta_{t-1}$ , to perform the cluster assignment for the contrastive clustering loss defined in Eqn. (2). Then, the visual grammar model $\phi$ with $\Theta_t$ at learning step $t$ will be learned (initialized from $\Theta_{t-1}$ ) on the features extracted from the dataset and the set of known clusters $\mathbf{c}$ up to the current learning step. Following [41], we sample a center $\mathbf{c}$ from the known clusters and its $M$ closest features to train the visual grammar model. + +Initial Visual Grammar Model. At the first learning step $t = 1$ , since no clusters have been learned initially, the visual grammar model $\phi$ with $\Theta_0$ is unavailable. However, as common practices in CSS [3, 4, 20], the segmentation model is typically trained from a pre-trained backbone on ImageNet [18]. As a result, the features extracted at the first learning step are characterized by the ImageNet features. Therefore, we adopt this philosophy to initialize our visual grammar model $(\phi$ with $\Theta_0$ ) by pre-training the visual grammar model on the ImageNet dataset. Then, during CSS training, we will progressively train our visual grammar model at each learning step as aforementioned. + +# 4. Experiments + +# 4.1. Implementations and Evaluation Protocols + +Implementation. Following common practices [2, 20, 62], we adopt DeepLab-V3 [6] with ResNet-101 [26] and SegFormer [69] with MiT-B3 [69] in our experiments. For the Visual Grammar model, we adopt the design of [41] with $L_{\phi} = 12$ blocks of multi-head self-attention layers. The feature vectors from the last layer of the decoder are used + +Table 1. Effectiveness of Fairness Contrastive Learning Loss. + +
(a) ADE20K 100-50
\( \mathcal{L}_{Cont} \)\( \mathcal{L}_{Cont}^{\alpha} \)0-100100-150allMajor Minor
✓44.615.234.851.5 26.4
✓44.624.537.952.1 30.8
(b) ADE20K 100-10
\( \mathcal{L}_{Cont} \)\( \mathcal{L}_{Cont}^{\alpha} \)0-100100-150allMajor Minor
✓41.916.033.249.9 24.9
✓44.420.436.451.8 28.7
+ +for our $\mathcal{L}_{Cont}^{\alpha}$ loss. The value $\alpha$ is set individually for each dataset, i.e., $\alpha = 5 \times 10^{-2}$ for ADE20K, $\alpha = 10^{-2}$ for VOC for Cityscapes. The details of our hyper-parameters are provided in the appendix. + +Evaluation Protocols. We evaluate models on three standard datasets of CSS, i.e., ADE20K [76], Pascal VOC [23], and Cityscapes [16]. Following common practices [3, 4, 62], our experiments are conducted on the overlapped CSS settings. In particular, on ADE20K, we use three different settings, i.e., ADE20K 100-50 (2 steps), ADE20K 100-10 (6 steps), and ADE20K 100-5 (11 steps). On Pascal VOC, we evaluate FALCON in three benchmarks, i.e., VOC 15-5 (2 steps), VOC 15-1 (6 steps), and VOC 10-1 (11 steps). On Cityscapes, we conduct domain incremental experiments with three settings, i.e., Cityscapes 11-5 (3 steps), Cityscapes 11-1 (11 steps), and Cityscapes 1-1 (21 steps). Following [3, 20], the mean Intersection over Union (mIoU) metric is adopted in our comparison, including mIoU of the last learning step on initial classes, incremental classes, and all classes. In addition, to illustrate the fairness improvement, we report the mIoU of major and minor classes. + +# 4.2. Ablation Study + +Effectiveness of Fairness Contrastive Clustering. Table 1 presents our results using DeepLab-V3 [7] with Resnet101 on ADE20K 100-50 and ADE20K 100-10 benchmarks. We evaluate the impact of the fairness contrastive clustering loss $\mathcal{L}_{Cont}^{\alpha}$ by comparing it with the vanilla contrastive clustering loss $\mathcal{L}_{Cont}$ . As shown in our results, the overall performance has been significantly improved to $37.9\%$ and $36.4\%$ on ADE20K 100-50 and ADE20K 100-10, respectively. In addition, the fairness has been promoted due to the mIoU improvement of minor groups. We also study the impact of network backbones and cluster margin $\nabla$ in our appendix. + +Table 2. Effectiveness of Visual Grammar. + +
(a) ADE20K 100-50
0-100101-150allMajorMinor
Nearest Cluster44.311.533.451.524.3
Fixed φ44.617.635.652.027.4
Adaptive φ44.624.537.952.130.8
(b) ADE20K 100-10
0-100101-150allMajorMinor
Nearest Cluster40.114.331.548.722.9
Fixed φ43.018.534.950.627.0
Adaptive φ44.420.436.451.828.7
+ +Table 3. Effectiveness of Scaling Factor $\alpha$ . +(a) ADE20K 100-50 + +
α0-1516-20allMajorMinor
α = 0.143.119.835.350.627.7
α = 0.0544.624.537.952.130.8
α = 0.0143.621.336.251.028.7
α = 0.00542.418.634.550.126.6
+ +(b) Pascal VOC 15-5 + +
α0-1516-20allMajorMinor
α = 0.174.851.669.376.963.5
α = 0.0576.251.370.379.063.8
α = 0.0179.454.873.581.367.7
α = 0.00574.648.968.577.661.6
+ +Effectiveness of Scaling Factor of Cluster. Table 3 illustrates the experimental results of the impact of different scaling factor $\alpha$ on ADE20K 100-50 and Pascal VOC 15-5 benchmarks. As shown in Table 3, when the value of scaling factor $\alpha$ gradually decreases, the performance of our proposed approach is improved accordingly since the fairness contrastive loss in Eqn (4) tends to be more uniform across major and minor classes. However, when the scaling factor is too small ( $\alpha = 0.005$ ), the impact of the loss enforcement becomes weaker leading to the weaker enforcement of the fairness contrastive clustering, resulting in lower overall performance. In addition, we have observed that the higher the number of classes demands the higher the value of $\alpha$ since it will increase the compactness of more clusters. + +Effectiveness of Loss Contributions. Table 4 illustrates the contributions of proposed learning objectives. For the model without using visual grammar, we only use a single unknown cluster $(N_U = 1)$ and adopt the nearest cluster strategies to assign clusters of unknown pixels. By using only cross-entropy loss, the mIoU performance remains low due to catastrophic forgetting and background shift problems. Meanwhile, with our fairness clustering loss $\mathcal{L}_{Cont}^{\alpha}$ , visual grammar model $\phi$ , and the cluster regularizer $\mathcal{R}$ , the mIoU performance has been significantly improved to $37.9\%$ and $36.4\%$ on ADE20K 100-50 and ADE20K 100-10, respectively. Moreover, our FALCON has significantly promoted the fairness of segmentation models illustrated by the mIoU improvement of major and minor groups. + +Effectiveness of Visual Grammar. We evaluate FALCON under three settings, i.e., Nearest Cluster, Fixed $\phi$ pretrained + +Table 4. Effectiveness of Our Proposed Losses. + +
(a) ADE20K 100-50
\( \mathcal{L}_{CE} \)\( \mathcal{L}_{Cont}^{\alpha } \)\( \phi \)\( \mathcal{R}_{C} \)0-100101-150allMajorMinor
✓0.018.96.30.09.4
✓✓44.07.931.951.622.1
✓✓✓43.821.836.451.129.1
✓✓✓✓44.624.537.952.130.8
+ +(b) ADE20K 100-10 + +
\( \mathcal{L}_{CE} \)\( \mathcal{L}_{Cont}^{\alpha } \)\( \phi \)\( \mathcal{R}_C \)0-100101-150allMajorMinor
✓0.03.51.20.01.8
✓✓39.013.130.447.821.6
✓✓✓43.418.535.151.227.1
✓✓✓✓44.420.436.451.828.7
+ +![](images/0ff1001ed48c41e4d2174c4832b287136ba02bd962dbe75b777e76f28cce5aa4.jpg) +Figure 6. Cluster Distribution at Learning Step $t = 1$ of ADE20K 100-50 (classes 109-144 are future classes) without (left) and with (right) Fairness Learning via Contrastive Attention. + +ImageNet (without updating on each learning step), and Adaptive $\phi$ (with updating on each learning step). As in Table 2, the mIoU result using only the nearest cluster remains ineffective. Meanwhile, the adaptive visual grammar model updated at each learning step further boosts the mIoU performance and promotes fairness, i.e., increased by $4.5\%$ and $4.9\%$ on ADE20K 100-50 and ADE20K 100-10 compared to the nearest cluster approach. In addition, we study the impact of choosing the number of features $M$ in the visual grammar model in our appendix. Fig. 6 illustrates the feature distributions of unknown classes (future class). As a result, our FALCON approach is able to model features of unknown classes into different clusters and produce better and more compact clusters compared to the one without Fairness Learning via Contrastive Attention. + +Effectiveness of Forgetting Each Learning Step. Table 5 reports the results of our model evaluated at each learning step. The performance of our model remains stable at each learning step. In particular, the maximum performance drop at learning step 4 of classes 111-120 due forgetting is only $0.8\%$ . Meanwhile, at the last learning task, the mIoU results of old classes are well maintained, with a slight drop of $0.4\%$ in classes 131-140. As shown in the result, our model has effectively maintained the performance to avoid catastrophic forgetting while enabling the capability of learning classes. + +# 4.3. Comparison with Prior SoTA Methods + +ADE20K. Table 6 presents our experimental results compared to prior CSS methods. By using DeepLab-V3, our approach has achieved SoTA performance, i.e., the mIoU results of $37.9\%$ and $+36.4\%$ on ADE20K 100-50 and ADE20K 100-10 benchmarks, higher than prior FairCL [62]. Meanwhile, FALCON using Transformer has outperformed + +Table 5. Effectiveness of CSS Model At Each Learning Step. + +
ADE20K 100-10
0-100101-110111-120121-130131-140141-150all
Step 144.9-----44.9
Step 244.815.8----42.2
Step 344.715.722.1---40.3
Step 444.615.621.326.3--39.2
Step 544.515.421.226.124.4-37.9
Step 644.415.421.226.124.015.536.4
avg40.1
+ +Table 6. Comparison with Prior Methods on ADE20K Benchmarks (Note: The results of MiB [2], PLOP [20], and FairCL [62] using Transformer on ADE20K 100-5 were not reported in prior studies. The upper bound results are not trained with fairness objective). + +
NetworkMethodADE20K 100-50ADE20K 100-10ADE20K 100-5
0-100101-150allavg0-100101-150allavg0-100101-150allavg
DeepLab-V3PLOP [20]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [75]42.318.834.5-39.317.632.1-38.511.529.6-
REIMINDER [45]41.619.234.1-39.021.333.1-36.116.429.5-
RCIL+LGKD [73]43.325.137.2-42.220.434.9-----
FairCL [62]43.424.037.040.541.720.434.739.0----
FALCON44.624.537.941.344.420.436.440.138.016.130.737.6
Upper Bound44.328.238.9-44.328.238.9-44.328.238.9-
Mask2FormerMiB [2]37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP [20]44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer [3]44.726.238.441.240.615.632.337.439.513.630.936.5
ECLIPSE [31]45.021.737.1-43.417.434.6-43.316.334.2-
CoMasTRe [25]45.726.039.241.642.318.434.438.440.815.832.638.6
TransformerMiB [2]43.430.639.238.739.120.434.239.5----
PLOP [20]43.826.238.038.143.324.136.240.3----
FairCL [62]43.625.537.640.742.221.935.539.4----
FALCON47.530.641.943.547.326.240.342.840.818.933.538.1
Upper Bound48.739.045.5-48.739.045.5-48.739.045.5-
+ +the prior CoMFormer [3] model by $+3.5\%$ , $+8.0\%$ , and $+2.6\%$ on ADE20K 100-50, ADE20K 100-10, and ADE20K 100-5. In addition, our mIoU results on the initial classes remain competitive with the upper-bounded results because FALCON can well handle the fairness problem compared to fully supervised learning. Although ECLIPSE [31] exhibits slightly better on ADE20K 100-5 by only $0.7\%$ (all) due to the better segmentation framework (i.e., Mask2Former), FALCON significantly outperforms ECLIPSE on ADE20K 100-50 and ADE20K 100-10. We also report our results of ADE20K 50-50 in the appendix. As in Fig. 7, FALCON produces better segmentation maps than prior methods. + +Pascal VOC. Table 7 presents our results on Pascal VOC benchmarks. FALCON has consistently achieved the SoTA performance on three benchmarks. In particular, compared to the prior FairCL [62] approach, our methods using DeepLab-V3 have improved the mIoU performance up to $73.50\%$ , $69.83\%$ , and $62.41\%$ on Pascal VOC 15-5, Pascal VOC 15-1, and Pascal VOC 10-1, respectively. Additionally, by using the better backbone, the performance is further improved, and the gap with the upper-bound result is reduced. + +Cityscapes. Table 8 reports the performance of our approach using DeepLab-V3 compared to prior methods on three different settings, i.e., Cityscapes 11-5, Cityscapes + +![](images/0f4de9087e49e5ac462c81197945df54f737630d27614a25cc819ded3bf2bf14.jpg) +Figure 7. The Comparison of Qualitative Results on ADE20K Between Our Approach and RCIL [75]. + +11-1, and Cityscapes 1-1. As in Table 8, the performance of our methods has consistently outperformed prior FairCL [62] approach by $+3.78\%$ , $+3.14\%$ , and $+6.02\%$ on three benchmarks. Similar to our experiments on ADE20K and VOC, the better network brings higher results. + +# 5. Conclusions and Limitations + +Conclusions. This paper has presented a novel Fairness Learning via Contrastive Attention approach to CSS. In particular, the fairness contrastive clustering loss has been introduced to model both catastrophic forgetting and fairness problems. Then, the visual grammar model was presented to model the unknown classes. The experimental results on different benchmarks have shown the SoTA performance and fairness promotion of our proposed FALCON approach. Limitations. Our study chose a set of learning hyperparameters to support our theoretical analysis. However, it potentially consists of several limitations related to choosing learning parameters and cluster initialization. Further discussion of limitations is presented in the supplementary. + +Acknowledgment This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs. + +Table 7. Comparisons with Prior Methods on Pascal VOC. + +
MethodPascal VOC 15-5Pascal VOC 15-1Pascal VOC 10-1
0-1516-20all0-1516-20all0-1011-20all
DeepLab-V3MiB [2]76.3749.9770.0838.0013.5032.2020.0020.1020.10
PLOP [20]75.7351.7170.0965.1021.1054.6044.0015.5030.50
RCIL [75]---70.6023.7059.4055.4015.1034.30
FairCL [62]---72.0022.7060.3042.3025.6034.40
SSUL [4]77.8250.1071.2277.3136.5967.6171.3145.9859.25
FALCON79.3554.7773.5078.3442.5769.8373.9449.7362.41
Upper Bound79.7772.3577.4379.7772.3577.4378.4176.3577.43
TransformerPLOP [20]72.5148.3766.7664.5937.2358.0848.5333.7141.47
SSUL [4]79.9156.8374.4179.9140.5670.5474.0651.8563.48
FairCL [62]---73.5022.8061.5057.1014.2036.60
FALCON81.2058.0475.6978.7147.5471.2874.9252.5464.26
Upper Bound80.8474.9779.4480.8474.9779.4480.8474.9779.44
+ +Table 8. Comparisons on Cityscapes. + +
Method11-511-11-1
DeepLab-V3LWF-MC [48]58.9056.9231.24
ILT [38]59.1457.7530.11
MIB [2]61.5160.0242.15
PLOP [20]63.5162.0545.24
RCIL [75]64.3063.0048.90
FairCL [62]66.9666.6149.22
FALCON70.7469.7555.24
Upper Bound79.3079.3079.30
Trans.FairCL [62]67.8567.0955.68
FALCON71.3370.1458.79
Upper Bound83.8083.8083.80
+ +# References + +[1] Nikita Araslanov, and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[2] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. Conf. Comp. Vision Pattern Rec., 2020. +[3] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. IEEE/CVF Computer Vision and Pattern Recognition Conference, 2023. +[4] Sungmin Cha, beomyoung kim, YoungJoon Yoo, and Taesup Moon. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In Advances in Neural Information Processing Systems, pages 10919-10930. Curran Associates, Inc., 2021. +[5] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. arXiv preprint arXiv:2407.16354, 2024. +[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. +[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. TPAMI, 2018. +[8] Mayee Chen, Daniel Y Fu, Avanika Narayan, Michael Zhang, Zhao Song, Kayvon Fatahalian, and Christopher Ré. Perfectly balanced: Improving transfer and robustness of supervised contrastive learning. In International Conference on Machine Learning, pages 3090-3122. PMLR, 2022. +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. +[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. +[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021. +[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. +[14] Sanghyeok Chu, Dongwan Kim, and Bohyung Han. Learning debiased and disentangled representations for semantic segmentation. Advances in Neural Information Processing Systems, 34:8355-8366, 2021. +[15] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for in- + +cremental semantic segmentation. In European Conference on Computer Vision, pages 244-261. Springer, 2024. +[16] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. +[17] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. +[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. +[19] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proc. Europ. Conf. Computer Vision, 2020. +[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4040-4050, 2021. +[21] Beyza Ermis, Giovanni Zappella, Martin Wistuba, Aditya Rawal, and Cedric Archambeau. Continual learning with transformers for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3774-3781, 2022. +[22] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. +[23] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015. +[24] Robert French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 1999. +[25] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3857, 2024. +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. +[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. +[28] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. DAFormer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022. +[29] Ibsa Jalata, Naga Venkata Sai Raviteja Chappa, Thanh-Dat Truong, Pierce Helton, Chase Rainwater, and Khoa Luu. Eqadap: Equipollent domain adaptation approach to image deblurring. IEEE Access, 10:93203-93211, 2022. +[30] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detect + +tion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5830-5840, 2021. +[31] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. +[32] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. +[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021. +[34] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 8547-8555, 2021. +[35] Chang Liu, Giulia Rizzoli, Pietro Zanuttigh, Fu Li, and Yi Niu. Learning from the web: Language drives weakly-supervised incremental learning for semantic segmentation. arXiv preprint arXiv:2407.13363, 2024. +[36] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world, 2019. +[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In NeurIPS, 2017. +[38] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCVWS, 2019. +[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. +[40] Pha Nguyen, Thanh-Dat Truong, Miaoqing Huang, Yi Liang, Ngan Le, and Khoa Luu. Self-supervised domain adaptation in crowd counting. In 2022 IEEE international conference on image processing (ICIP), pages 2786-2790. IEEE, 2022. +[41] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10847-10856, 2021. +[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. +[43] First Ozdemir and Orcun Goksel. Extending pretrained segmentation networks with additional anatomical structures. International journal of computer assisted radiology and surgery, 14:1187-1195, 2019. +[44] First Ozdemir, Philipp Fuernstahl, and Orcun Goksel. Learn the new, keep the old: Extending pretrained models with new anatomy and images. In Medical Image Computing + +and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part IV 11, pages 361-369. Springer, 2018. +[45] Minh Hieu Phan, The-Anh Ta, Son Lam Phung, Long Tran-Thanh, and Abdesselam Bouzerdoum. Class similarity weighted knowledge distillation for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16866-16875, 2022. +[46] Yiqiao Qiu, Yixing Shen, Zhuohao Sun, Yanchong Zheng, Xiaobin Chang, Weishi Zheng, and Ruixuan Wang. Sats: Self-attention transfer for continual semantic segmentation. Pattern Recognition, 138:109383, 2023. +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proc. Conf. Comp. Vision Pattern Rec., 2017. +[49] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition, 2020. +[50] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995. +[51] Mohammad Rostami. Lifelong domain adaptation via consolidated internal distribution. In NeurIPS, 2021. +[52] Antoine Saporta, Arthur Douillard, Tuan-Hung Vu, Patrick Pérez, and Matthieu Cord. Multi-head distillation for continual unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2022. +[53] Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. Incremental learning of object detectors without catastrophic forgetting. In Proceedings of the IEEE international conference on computer vision, pages 3400-3409, 2017. +[54] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021. +[55] Attila Szabó, Hadi Jamali-Rad, and Siva-Datta Mannava. Tilted cross-entropy (tce): Promoting fairness in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2305–2310, 2021. +[56] Sebastian Thrun. Lifelong learning algorithms. In *Springer Learning to Learn*, 1998. +[57] Thanh-Dat Truong and Khoa Luu. Cross-view action recognition understanding from exocentric to egocentric perspective. Neurocomputing, 614:128731, 2025. +[58] Thanh-Dat Truong, Chi Nhan Duong, Khoa Luu, Minh-Triet Tran, and Ngan Le. Domain generalization via universal non-volume preserving approach. In 2020 17th Conference + +on Computer and Robot Vision (CRV), pages 93-100. IEEE, 2020. +[59] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Son Lam Phung, Chase Rainwater, and Khoa Luu. Bimal: Bijective maximum likelihood approach to domain adaptation in semantic scene segmentation. In International Conference on Computer Vision, 2021. +[60] Thanh-Dat Truong, Ravi Teja Nvs Chappa, Xuan-Bac Nguyen, Ngan Le, Ashley PG Dowling, and Khoa Luu. Otadapt: Optimal transport-based approach for unsupervised domain adaptation. In 2022 26th international conference on pattern recognition (ICPR), pages 2850-2856. IEEE, 2022. +[61] Thanh-Dat Truong, Ngan Le, Bhiksha Raj, Jackson Cothren, and Khoa Luu. Freedom: Fairness domain adaptation approach to semantic scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. +[62] Thanh-Dat Truong, Hoang-Quan Nguyen, Bhiksha Raj, and Khoa Luu. Fairness continual learning approach to semantic scene understanding in open-world environments. In NeurIPS, 2023. +[63] Thanh-Dat Truong, Pierce Helton, Ahmed Moustafa, Jackson David Cothren, and Khoa Luu. Conda: Continual unsupervised domain adaptation learning in visual perception for self-driving cars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 5642-5650, 2024. +[64] Thanh-Dat Truong, Utsav Prabhu, Dongyi Wang, Bhiksha Raj, Susan Gauch, Jeyamkondan Subbiah, and Khoa Luu. EAGLE: Efficient adaptive geometry-based learning in cross-view understanding. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[66] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proc. Conf. Comp. Vision Pattern Rec., 2021. +[67] Huyong Wang, Huisi Wu, and Jing Qin. Incremental nuclei segmentation from histopathological images via future-class awareness and compatibility-inspired distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11408-11417, 2024. +[68] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. +[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. +[70] Zhengyuan Xie, Haiquan Lu, Jia-wen Xiao, Enguang Wang, Le Zhang, and Xialei Liu. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen + +tation. In European Conference on Computer Vision, pages 183-201. Springer, 2025. +[71] Lei Yang, Xiaohang Zhan, Dapeng Chen, Junjie Yan, Chen Change Loy, and Dahua Lin. Learning to cluster faces on an affinity graph. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2298-2306, 2019. +[72] Lei Yang, Dapeng Chen, Xiaohang Zhan, Rui Zhao, Chen Change Loy, and Dahua Lin. Learning to cluster faces via confidence and connectivity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13369-13378, 2020. +[73] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023. +[74] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. arXiv preprint arXiv:2407.09838, 2024. +[75] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. +[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. +[77] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. \ No newline at end of file diff --git a/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/images.zip b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1c06078ba337e66c41743098c8d35fef6cd0507 --- /dev/null +++ b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b61280531db4a87e7b72cae4276a6094be55d202b8eae12771a4fe1ef0bd5b3b +size 732959 diff --git a/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/layout.json b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..deff624f029ce2522e3149278c8f1cda336e4c3c --- /dev/null +++ b/2025/FALCON_ Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding/layout.json @@ -0,0 +1,11902 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 99, + 532, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 99, + 532, + 142 + ], + "spans": [ + { + "bbox": [ + 80, + 99, + 532, + 142 + ], + "type": "text", + "content": "FALCON: Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "spans": [ + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": "Thanh-Dat Truong" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": ", Utsav Prabhu" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": ", Bhiksha Raj" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": ", Jackson Cothren" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": ", Khoa Luu" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": "CVIU Lab, University of Arkansas, USA " + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": "Google DeepMind, USA \n" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": "Carnegie Mellon University, USA " + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": "Mohammed bin Zayed University of AI, UAE \n" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 95, + 152, + 517, + 209 + ], + "type": "text", + "content": "Dep. of Geosciences, University of Arkansas, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 211, + 517, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 211, + 517, + 236 + ], + "spans": [ + { + "bbox": [ + 94, + 211, + 517, + 236 + ], + "type": "text", + "content": "{tt032, jcothre, khoaluu}@uark.edu, bhiksha@cs.cmu.edu, utsavprabhu@google.com http://uark-cviu.github.io/projects/FALCON" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 253, + 200, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 253, + 200, + 266 + ], + "spans": [ + { + "bbox": [ + 151, + 253, + 200, + 266 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 297, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 297, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 297, + 557 + ], + "type": "text", + "content": "Continual Learning in semantic scene segmentation aims to continually learn new unseen classes in dynamic environments while maintaining previously learned knowledge. Prior studies focused on modeling the catastrophic forgetting and background shift challenges in continual learning. However, fairness, another major challenge that causes unfair predictions leading to low performance among major and minor classes, still needs to be well addressed. In addition, prior methods have yet to model the unknown classes well, thus resulting in producing non-discriminative features among unknown classes. This work presents a novel Fairness Learning via Contrastive Attention Approach to continual learning in semantic scene understanding. In particular, we first introduce a new Fairness Contrastive Clustering loss to address the problems of catastrophic forgetting and fairness. Then, we propose an attention-based visual grammar approach to effectively model the background shift problem and unknown classes, producing better feature representations for different unknown classes. Through our experiments, our proposed approach achieves State-of-the-Art (SoTA) performance on different continual learning benchmarks, i.e., ADE20K, Cityscapes, and Pascal VOC. It promotes the fairness of the continual semantic segmentation model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "The semantic segmentation networks, e.g., Transformers [69] and Convolutional Neural Networks [7], learned from data with a closed-set of known classes, have shown outstanding performance. However, they often suffer performance degradation when encountering novel objects or classes in new dynamic environments [2, 20, 62, 63]. To improve their performance, several transfer learning and domain adaptation methods [1, 28, 29, 40, 57-60, 62, 64] were introduced to adapt trained models into deployed environments. While" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 251, + 555, + 384 + ], + "blocks": [ + { + "bbox": [ + 316, + 251, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 384 + ], + "type": "image", + "image_path": "c59bd327f32f50759195ae02cd7f7ed871e4b8b933cae3b50922ea91065c27b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 431 + ], + "type": "text", + "content": "Figure 1. Our Fairness Learning via Contrastive Attention to Continual Semantic Segmentation. The Fairness Contrastive Clustering Loss promotes the fairness of the model while the Attention-based Visual Grammar models the unknown classes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "type": "text", + "content": "the former often aims to fine-tune the model on labeled data collected in the new environments, the latter adapts the model to the new domains in an unsupervised manner [1, 28]. However, these methods cannot handle novel objects well due to their close-set learning. In practice, the semantic segmentation models should be able to adaptively and continually learn the new knowledge of novel classes. It motivates the development of Continual Learning paradigm [2, 20, 46], a.k.a, Continual Semantic Segmentation (CSS), where the segmentation models are learned sequentially to new contents of data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Far apart from prior segmentation methods [7, 69] that learn one time on static, closed-set data, Continual Learning requires the segmentation models to learn from dynamic, open-set data [20, 62]. In a particular scenario, accessing previous learning data is restricted due to privacy concerns. In CSS, three challenges have been identified, including (1) Catastrophic Forgetting, (2) Background Shift, and (3) Fairness. While the catastrophic forgetting problem [24, 50, 56] depicts the segmentation model tends to forget its knowledge when learning new data, background shift indicates the problem of classes of previous or future data (unknown classes)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15065" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "type": "image", + "image_path": "7477b3592d42ef495fc1f16284bddfe5ac5f7b390885caa6d5959143cf357d37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "type": "text", + "content": "Figure 2. The Data Class Distribution of ADE20K. The major classes occupy more than " + }, + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "type": "text", + "content": " of the total pixels of the dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 221, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 221, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 295, + 327 + ], + "type": "text", + "content": "have collapsed into a background class [2, 20, 75]. Prior methods [4, 20, 46, 62] addressed these two problems by introducing knowledge distillation and pseudo labels. However, these methods can not handle unknown classes since they either consider these unknown classes as a background class or assign unknown pixels by a pseudo label of prior known classes [20, 62]. More importantly, the last problem, fairness, is a significant challenge that limits the performance of CSS models (Fig. 1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 329, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 329, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 329, + 296, + 544 + ], + "type": "text", + "content": "As shown in Fig. 2, the number of pixels of each class in training data have been imbalanced among classes and significantly decreased after each task. Thus, this bias influences the learning procedure and model predictions that later cause unfair class predictions. However, limited studies are taking the fairness problem into account. [61] presented a similar problem in domain adaptation and extended it to continual learning [62]. These methods rely on the assumption of ideal fair or balanced data distributions. However, it is not applicable in practice since the size, i.e., the number of pixels, of several classes can never be more significant than others. For example, the bottle size should not be more significant than the size of a car. Meanwhile, the current knowledge distillation methods [3, 20, 45] in CSS are unable to handle the fairness problem since they focus on modeling catastrophic forgetting and background shift problems. Therefore, it is essential to develop a new CSS approach to address these limitations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": "Contributions of This Work. This work presents a novel Fairness Learning via Contrastive Attention Approach (FALCON) to Continual Semantic Segmentation (as shown in Fig. 1). Our contributions can be summarized as follows. First, we introduce a novel Contrastive Clustering Paradigm approach to Continual Learning that models the catastrophic forgetting problem. Second, by analyzing the limitation of vanilla Contrastive Clustering in biased data, we introduce a novel Fairness Contrastive Clustering loss to model the fairness problem in continual learning efficiently. Third, to effectively model the background shift problem, we introduce a new Attention-based Visual Grammar that model the topological structures of feature distribution to handle the unknown classes effectively. Finally, the ablation studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "illustrate the effectiveness of the proposed approach in different aspects of fairness promotion in CSS models. Compared with prior methods, our approach achieves SoTA performance on different settings of three standard benchmarks of CSS, including ADE20K, Pascal VOC, and Cityscapes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 143, + 400, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 143, + 400, + 155 + ], + "spans": [ + { + "bbox": [ + 314, + 143, + 400, + 155 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 163, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 437 + ], + "type": "text", + "content": "Continual Semantic Segmentation. Several studies were introduced to address catastrophic forgetting and background shift problems [19-21, 32, 37, 39, 43, 44]. The common CSS approach adopts knowledge distillation [20] and pseudo labels [2] to model catastrophic forgetting and background shift, respectively. Later, it was further improved by decoupling knowledge representations [75], modeling the inter- and intra-class knowledge [15, 46], distinguishing the feature representations of the future classes [4, 45, 67, 70], reducing background confusion [73, 74], using visual prompting [31, 35], or modeling distillation loss via the geodesic flow [54]. Another approach [3, 5, 25] adopted the mask-based segmentation networks [12, 13] to improve the performance. Recent studies have introduced CSS under the unsupervised domain adaptation settings [51, 52, 63, 66]. However, prior studies have yet to well model different unknown classes. Particularly, as previous methods [3, 20] use the pseudo labels to model the unknown classes, the future classes will be treated as a single background class. Then, Joseph et al. [30] improves the unknown class modeling using clustering but this method considers all different unknown classes as a single cluster, leading to the non-discriminative features among unknown classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 438, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 653 + ], + "type": "text", + "content": "Contrastive Learning (CL). CL is a common learning approach [9, 10, 42] to structure the deep feature representations in the deep latent space. Oorde et al. [42] first introduced the Noise-Contrastive Estimation (InfoNCE) learning framework. Then, Chen et al. [9] presented SimCLR, a self-supervised contrastive learning approach to improve the representation power of Residual Networks. He [27] proposed a Momentum Contrast framework for unsupervised representation learning. Later, it was further improved by using MLP projection head [10] and extended to improve the self-supervised training process of vision transformers [11]. Cui et al. [17] introduced a supervised parametric contrastive learning loss to address the long-tailed recognition. Li et al. [34] adopted contrastive learning to develop the one-stage online contrastive clustering method. Radford et al. [47] presents a contrastive framework to learn the vision-language model. Later, several methods also adopted this framework for vision-language pretraining [33, 47]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Imbalanced and Fairness Learning. The early methods utilized the balanced Softmax loss [49] to alleviate the impact of imbalanced data distribution. Later, Wang et al. [68] introduced a Seesaw loss to re-balance the contributions of positive and negative instances via the mitigation" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15066" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "and compensation modules. Ziwei et al. [36] introduced a dynamic meta-embedding to model the imbalanced classification problem. Chu et al. [14] reduce the bias in the segmentation model by presenting a new stochastic training scheme. Szabo et al. [55] presented a tilted cross-entropy loss to promote class-relevant fairness. However, there are limited studies that address the fairness problem in CSS. Truong et al. [61] introduced a fairness domain adaptation approach to semantic segmentation and later extended it into continual learning setting [62]. However, these methods [61, 62] rely on the assumption of ideal balanced data, which could not be achieved by nature. To address the limitations in prior work, this paper will introduce a novel approach to effectively model the fairness problem and unknown classes in the continual learning setting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 248, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 248, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 248, + 276 + ], + "type": "text", + "content": "3. The Proposed FALCON Approach" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": "CSS aims to learn a segmentation network " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " on sequence data " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathcal{D}^1,\\dots,\\mathcal{D}^T\\}" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " is the number of learning steps. At learning step " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": ", the model " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " encounters a dataset " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{(\\mathbf{x}^t,\\hat{\\mathbf{y}}^t)\\}" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " is the image and " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{y}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " is a segmentation label of " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": ". The ground truths at learning step " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " only consist of current classes " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^t" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": ", while the class labels of the previous " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{1\\dots t - 1}" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " and future steps " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{t + 1\\dots T}" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " are collapsed into a background class. Formally, learning the CSS model at step " + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 281, + 296, + 390 + ], + "type": "text", + "content": " can be formed as Eqn. (1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 392, + 296, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 392, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 392, + 296, + 409 + ], + "type": "interline_equation", + "content": "\\theta_ {t} ^ {*} = \\arg \\min _ {\\theta_ {t}} \\mathbb {E} _ {\\mathbf {x} ^ {t}, \\hat {\\mathbf {y}} ^ {t} \\in \\mathcal {D} ^ {t}} \\left[ \\mathcal {L} _ {C E} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} ^ {t}\\right) + \\lambda_ {C L} \\mathcal {L} _ {C L} \\left(F (\\mathbf {x} ^ {t})\\right) \\right] (1)", + "image_path": "812d3cb44cc87e6eb900816dc8e5192b16fd954d13d3b38c8dc00bf0d60e8fb1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^t = F(\\mathbf{x}^t,\\theta_t)" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": " is the parameter of " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": " at current learning step " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": " is the cross-entropy loss, " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\lambda_{CL}" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": " is the balanced weight. and " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CL}" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": " is the CSS objective. At learning step " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": ", the segmentation model " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": " is required to be able to predict both previously learned classes " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{1\\dots t - 1}" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": " and current new classes " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^t" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": ". Under this learning scenario, three challenges have been identified, i.e., Catastrophic Forgetting, Background Shift, and Fairness. Several prior methods were presented to model the two first issues in CSS using knowledge distillation [3, 20]. The last issue has not been well addressed yet due to its challenges [62]. Prior methods [2, 3, 20, 53, 75] adopt knowledge distillation to design " + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CL}" + }, + { + "bbox": [ + 55, + 413, + 296, + 654 + ], + "type": "text", + "content": ". However, this method prevents the CSS model from diverging knowledge learned previously, therefore resulting in limiting the ability to adopt new knowledge [62]. In addition, these methods have not addressed fairness and background shift problems due to their dedicated design for maintaining knowledge via distillation [3, 20, 45]. Therefore, to address these problems, we introduce a novel Fairness Learning via Contrastive Attention Approach to CSS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 660, + 296, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 296, + 673 + ], + "type": "text", + "content": "3.1. Continual Learning via Contrastive Clustering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Apart from prior methods [3, 20, 45], our CSS (Fig. 3) is defined as Contrastive Clustering Learning. Given a set of centroid vectors " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i\\}_{i=1}^{N_K + N_U}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_K = |\\mathcal{C}^{1..t}|" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_U" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "is the number of known and unknown classes up to current learning tasks. Prior work [3, 30, 62] often defined the number of unknown classes as 1 where background classes are considered as a single unknown class. Formally, our Contrastive Clustering Learning for CSS can be defined as in Eqn. (2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 149, + 554, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 149, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 326, + 149, + 554, + 198 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {C L} \\left(F (\\mathbf {x} ^ {t})\\right) = \\sum_ {\\mathbf {c} _ {i}} \\mathcal {L} _ {C o n t} (\\mathbf {F} ^ {t}, \\mathbf {c} _ {i}) \\\\ = \\sum_ {\\mathbf {c} _ {i}} \\sum_ {h, w} - \\phi \\left(\\mathbf {f} _ {h, w} ^ {t}, \\mathbf {c} _ {i}\\right) \\log \\frac {\\exp \\left(\\mathbf {f} _ {h , w} ^ {t} \\times \\mathbf {c} _ {i}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c} _ {i}\\right)} \\tag {2} \\\\ \\end{array}", + "image_path": "4d85d45cf2b6237466b9dc46963724b8772e5f598810d3eb0b23957c881892b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^t\\in \\mathbb{R}^{H\\times W\\times D}" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " is the feature maps extracted from the input image " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " by the segmentation network " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{h,w}^t\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " is the feature at the pixel location " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " of features " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^t" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\sum_{\\mathbf{f}'}" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " means the summation over all feature representations " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\phi :\\mathbb{R}^D\\times \\mathbb{R}^D\\to [0,1]" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " is the function that determines either " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{h,w}^t" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " belongs to the cluster " + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 203, + 555, + 278 + ], + "type": "text", + "content": " or not." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": "By defining CSS as contrastive clustering learning, the knowledge of the segmentation model has been well maintained via the cluster vectors " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": ". Then, minimizing Eqn. (2) will separate the representations of different classes while gathering the features of the same class into the same cluster. As the cluster vectors " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": " of the old classes " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{1..t-1}" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": " have been well learned to represent their knowledge, these vectors are frozen at learning step " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": " to maintain the knowledge representations of previous classes to address the catastrophic forgetting problem. To effectively learn cluster vectors " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": ", the cluster vector " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": " will periodically updated after each " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": " steps by the momentum update [27, 30, 62] based on the features " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{h,w}^t" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": " assigned to cluster " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": ". However, there are two major problems in contrastive clustering learning. First, since the training data suffer the bias among classes as shown in Fig. 2, this bias will influence Eqn. (2) and cause the unfair predictions. Second, as the function " + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 278, + 556, + 529 + ], + "type": "text", + "content": " requires the labels to determine the features belonging to clusters, it limits the ability to model the unknown classes where the labels are unavailable. Therefore, Secs. 3.2-3.3 will present a novel approach to tackle these problems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 536, + 531, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 531, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 531, + 549 + ], + "type": "text", + "content": "3.2. Fairness Contrastive Clustering Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": "While contrastive clustering learning in Eqn. (2) promotes the compact representations of features around their clusters, inspired by [8, 17, 77], we observe that the imbalanced class distribution will influence unfair behaviors among classes. In particular, for simplicity, we consider " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i^t\\}_{i=1}^L" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " is the set of features that belong to the cluster " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " at learning step " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbf{f}_i^t, \\mathbf{c}) = 1" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " is the number of features (in this case, " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " is the total number of pixels belong to the class of cluster " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": "). Let us define the enforcement between the feature " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^t" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " and the cluster " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\ell_i = \\frac{\\exp(\\mathbf{f}_i^t \\times \\mathbf{c})}{\\sum_{\\mathbf{f}' \\in L} \\exp(\\mathbf{f}' \\times \\mathbf{c})}" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": ". Hence, the lower the value of the enforcement " + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\ell_i" + }, + { + "bbox": [ + 313, + 554, + 556, + 713 + ], + "type": "text", + "content": " is, the more compact the representation of visual features and clusters is. Then, the contrastive clustering learning loss in Eqn. (2) of entire" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15067" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 69, + 504, + 190 + ], + "blocks": [ + { + "bbox": [ + 105, + 69, + 504, + 190 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 504, + 190 + ], + "type": "image", + "image_path": "c942f7a1568715b6935ae538020f155f335c8fb9a5ba1c3e0c5aa88913f00cd3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 190, + 543, + 201 + ], + "lines": [ + { + "bbox": [ + 66, + 190, + 543, + 201 + ], + "spans": [ + { + "bbox": [ + 66, + 190, + 543, + 201 + ], + "type": "text", + "content": "Figure 3. The Proposed Fairness Learning via Contrastive Attention Approach to Continual Semantic Scene Understanding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 208, + 212, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 212, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 212, + 219 + ], + "type": "text", + "content": "cluster " + }, + { + "bbox": [ + 55, + 208, + 212, + 219 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 208, + 212, + 219 + ], + "type": "text", + "content": " can be defined as in Eqn. (3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 227, + 296, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 296, + 257 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {C o n t}} (;;, \\mathbf {c}) = - \\sum_ {i = 1} ^ {L} \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {t} \\times \\mathbf {c}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} = - \\sum_ {i = 1} ^ {L} \\log \\ell_ {i} \\tag {3}", + "image_path": "b7e55a4660dd84c2aa3673ceb6f0a13c85b1d784ec84b056b040c146b3051af7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "text", + "content": "Proposition 1. If the contrastive clustering loss " + }, + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}(:,\\mathbf{c})" + }, + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "text", + "content": " achieves the optimal value, the enforcement " + }, + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\ell_{i}" + }, + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "text", + "content": " between the feature and the cluster will converge to " + }, + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\ell_{i} = L^{-1}" + }, + { + "bbox": [ + 54, + 265, + 295, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "text", + "content": "Proposition 1 has implied that the class with more samples will result in a lower value of the enforcement and produce a more compact representation while the class having fewer samples will be more scattered in the feature space due to the higher value of the enforcement. In particular, let " + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "inline_equation", + "content": "L_{major}" + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "inline_equation", + "content": "L_{minor}" + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "text", + "content": " be the number of samples of the major and minor class where " + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "inline_equation", + "content": "L_{major} >> L_{minor}" + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "text", + "content": ". Then, based on Proposition 1, the enforcement between features and the cluster of the major class will be significantly lower than the one of the minor class, i.e., " + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "inline_equation", + "content": "L_{major}^{-1} << L_{minor}^{-1}" + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "text", + "content": ". Therefore, a direct adoption of the contrastive clustering loss in Eqn. (2) will result in an unfair CSS model. In addition, for classes in the minority group, the lower value of " + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 56, + 301, + 296, + 539 + ], + "type": "text", + "content": " results in the feature presentations of classes being far away from their clusters. Thus, the model will produce non-discriminative features compared to the ones in the majority group. Moreover, if the loss is applied to the cases of unknown labels, these feature representations can be scattered in the latent space and pulled into the incorrect clusters due to weak enforcement between features and clusters (Fig. 4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": "To address the unfair problem in contrastive clustering learning, inspired by [8, 17, 77], we introduce a scaling factor " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": " and a learnable transition vector " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": " for each cluster " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": " (all clusters have the same value of " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": " but different vector " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": "). Our Fairness Contrastive Clustering Learning Loss for the entire cluster in Eqn. (3) can be re-formed as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 617, + 296, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 617, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 617, + 296, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {C o n t}} ^ {\\alpha} (;, \\mathbf {c}) = - \\alpha \\sum_ {i = 1} ^ {L} \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {t} \\times \\mathbf {c}\\right)}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} - \\log \\frac {\\exp (\\mathbf {v} \\times \\mathbf {c})}{\\sum_ {\\mathbf {f} ^ {\\prime}} \\exp \\left(\\mathbf {f} ^ {\\prime} \\times \\mathbf {c}\\right)} \\tag {4}", + "image_path": "104cd665aba37ac39857f2a11e8f5be79eafb43b029b5f00d065cbe802db7b07.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Intuitively, the scaling factor " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": " will help to re-scale the impact of the enforcement in learning, and the transitive vector " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": " assists in translating the center cluster into the proper position of the latent space. This action promotes the compactness of clusters in the minority group." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "text", + "content": "Proposition 2. If the fairness contrastive clustering loss " + }, + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}^{\\alpha}(:,\\mathbf{c})" + }, + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "text", + "content": " achieves the optimal value, the enforcement " + }, + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "inline_equation", + "content": "\\ell_{i}" + }, + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "text", + "content": " between the feature and the cluster will converge to " + }, + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "inline_equation", + "content": "\\ell_{i} = (\\alpha^{-1} + L)^{-1}" + }, + { + "bbox": [ + 313, + 208, + 553, + 255 + ], + "type": "text", + "content": ". Proofs of Propositions 1-2 are in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": "Under the Proposition 2, when the value of " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " is small, the divergence of the enforcement between major and minor classes will be smaller, i.e., " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "||(\\alpha^{-1} + L_{major})^{-1} - (\\alpha^{-1} + L_{minor})^{-1}|| < ||L_{major}^{-1} - L_{minor}^{-1}||" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": ". Fig. 4 has illustrated the impact of fairness contrastive clustering loss. Therefore, our designed proposed fairness contrastive loss has effectively addressed the fairness issue in Eqn. (2). It should be noted that although the smaller " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " results in the fairer enforcement varied from major to minor classes. However, if the value of scaling factor " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " is too small, the contrastive clustering loss will rely more on the enforcement of the transitive vector " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": ", and the distribution of features " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^t" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " around its cluster " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " will be scattered due the weak enforcement caused by small " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": ". Therefore, the value of scaling factor " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " in practice should be carefully selected." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 441, + 516, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 516, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 516, + 454 + ], + "type": "text", + "content": "3.3. An Efficient Unknown Class Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "content": "An ideal CSS approach must be able to model the unknown classes without supervision, especially in open-set contexts" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 488, + 553, + 651 + ], + "blocks": [ + { + "bbox": [ + 316, + 488, + 553, + 651 + ], + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 651 + ], + "type": "image", + "image_path": "fd774369ab042b69c737d32bebdef9bed92f826bef9d7b9b940de6ea8260a0d3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "text", + "content": "Figure 4. The Enforcement Loss of Contrastive Clustering " + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}" + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "text", + "content": " and Fairness Contrastive Clustering " + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}^{\\alpha}" + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "text", + "content": " on Pascal VOC. Since " + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}" + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "text", + "content": " suffers severe biased, its clusters of minor classes remain scattered. Our " + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}^{\\alpha}" + }, + { + "bbox": [ + 313, + 654, + 555, + 710 + ], + "type": "text", + "content": " produces a more uniform loss among classes, promotes fairness and compactness of clusters." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15068" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "[30, 62] where there could be multiple unknown classes or objects. Prior studies have adopted the pseudo-label strategies [3, 20] based on the model predictions to assign labels for seen classes while unseen classes have been ignored, thus resulting in non-discriminative features. [30, 62] improved the background modeling by using an additional prototypical representation for unknown classes. However, these approaches consider different unknown classes as one (i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "N_U = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ") resulting in non-distinguished representations of different unknown classes. Thus, modeling function " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " in Eqn. (2) without supervision of different unknown classes (i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "N_U > 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ") is challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": "Although modeling " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " to determine the single feature " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " belonging to the cluster " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " is challenging, prior studies in clustering [41, 71, 72] have suggested that determine a set of features " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i^t\\}_{i = 1}^M" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " belonging to cluster " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " should be easier. This derives from the fact that even though the feature representations of different classes are different, the distributions of features around its cluster (termed as Visual Grammar) in the feature space should be similar among classes or clusters. As a result, by learning the distribution of features and their clusters, the model " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " can determine whether a feature belongs to a cluster. Then, by learning the model " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " on prior known clusters and features, the knowledge of " + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 219, + 296, + 387 + ], + "type": "text", + "content": " can be adaptively applied to unknown clusters. Fig. 5 illustrates our visual grammar model of the cluster distributions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 390, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 558 + ], + "type": "text", + "content": "Limitations of Prior Clustering Methods. The traditional methods in clustering, e.g., KNN or density-based clustering [22], remain limited to noisy features leading to producing the incorrect cluster assignment. Meanwhile, the modern clustering methods, e.g., Graph Neural Networks (GNNs) [71, 72], require a large memory to build the affinity graph for clusters. In addition, GNNs often learn the local structures of graphs (or clusters) and accumulate them via the aggregation layers. Hence, the global structures of the clusters, i.e., visual grammar, are not well modeled by GNNs [41]. Therefore, to address these limitations, we introduced a new Attention-based Visual Grammar approach to efficiently model the distribution of features and their clusters via the self-attention mechanism [65]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": "Remark 1: Given a center " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": " features " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i^{\\mathbf{c}}\\}_{i = 1}^{M}" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathbf{c}}" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": " denotes the feature " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": " belonging to the cluster " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\forall i\\in [1..M - 1]:\\cos (\\mathbf{f}_i^{\\mathbf{c}},\\mathbf{c})\\geq \\cos (\\mathbf{f}_{i + 1}^{\\mathbf{c}},\\mathbf{c})" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": " the Visual Grammar of the cluster c parameterized by " + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 560, + 296, + 609 + ], + "type": "text", + "content": " can be defined" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 621, + 296, + 696 + ], + "blocks": [ + { + "bbox": [ + 56, + 621, + 296, + 696 + ], + "lines": [ + { + "bbox": [ + 56, + 621, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 296, + 696 + ], + "type": "image", + "image_path": "6384771357077251daa8b9897a43d1e44dd7c7be51c270b49d96b82029fcafe5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 699, + 270, + 711 + ], + "lines": [ + { + "bbox": [ + 81, + 699, + 270, + 711 + ], + "spans": [ + { + "bbox": [ + 81, + 699, + 270, + 711 + ], + "type": "text", + "content": "Figure 5. The Proposed Visual Grammar Model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 374, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 374, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 374, + 84 + ], + "type": "text", + "content": "as in Eqn. (5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 322, + 89, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 89, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 322, + 89, + 553, + 121 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\{\\mathbf {f} _ {i, j} ^ {\\mathbf {c}} \\} ^ {M}} [ - \\log p (\\mathbf {f} _ {1} ^ {\\mathbf {c}}, \\mathbf {f} _ {2} ^ {\\mathbf {c}}, \\dots , \\mathbf {f} _ {M} ^ {\\mathbf {c}}, \\mathbf {c}, \\Theta) ] \\\\ \\Leftrightarrow \\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\left\\{\\mathbf {f} _ {i} ^ {\\mathbf {c}} \\right\\} _ {i = 1} ^ {M}} [ - \\log p \\left(\\boldsymbol {\\Delta} _ {1} ^ {\\mathbf {c}}, \\boldsymbol {\\Delta} _ {2} ^ {\\mathbf {c}}, \\dots , \\boldsymbol {\\Delta} _ {M} ^ {\\mathbf {c}}, \\mathbf {c}, \\Theta\\right) ] \\tag {5} \\\\ \\end{array}", + "image_path": "f51f31c5b97129457d8ec9165192a063b8a7a449975d831e37122240a736ac09.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\Delta_i^{\\mathbf{c}} = \\mathbf{f}_i^{\\mathbf{c}} - \\mathbf{c}" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": ". Eqn. (5) defines the visual grammar of the cluster by modeling the feature distribution of " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathbf{c}}" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " and its cluster center " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\phi : \\mathbb{R}^{(M + 1)\\times D} \\to [0,1]^M" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " be a function receiving a center " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " features " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i\\}_{i = 1}^M" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\cos(\\mathbf{f}_i,\\mathbf{c}) \\geq \\cos(\\mathbf{f}_{i + 1},\\mathbf{c})" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": ") to determine whether " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " belonging to " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\phi(\\Delta_1,\\Delta_2,\\dots,\\Delta_M,\\mathbf{c})" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\Delta_i = \\mathbf{f}_i - \\mathbf{c}" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = [u_1,u_2,\\dots,u_M]" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "u_i = 1" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " belong to cluster " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " and vice versa. Hence, the visual grammar model in Eqn. (5) can be modeled by the network " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " as follows with parameter " + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 125, + 555, + 245 + ], + "type": "text", + "content": " as in Eqn. (6)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 251, + 555, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 251, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 323, + 251, + 555, + 265 + ], + "type": "interline_equation", + "content": "\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathbb {E} _ {\\mathbf {c}, \\left\\{\\mathbf {f} _ {i} \\right\\} _ {i = 1} ^ {M}} \\left[ - \\log p (\\mathbf {u} | \\boldsymbol {\\Delta} _ {1}, \\boldsymbol {\\Delta} _ {2}, \\dots , \\boldsymbol {\\Delta} _ {M}, \\mathbf {c}, \\Theta) \\right] \\tag {6}", + "image_path": "1f1aa3b3a13f2b95cac1d5c9e6e8e0a88122099f4ee65b409ad894fc21767a1c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": "Eqn. (6) aims to model the distribution of features around its cluster by learning the correlation of relatively topological structures " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_i" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " of features " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " around cluster " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": ". Then, based on knowledge of the cluster distribution, the model " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " is able to determine whether a feature " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " belongs to cluster " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": ". Hence, it is essential that the model " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " has the ability to exploit the correlation between features " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " and cluster " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " to learn the topological structure of visual grammar. Therefore, we adopt the self-attention mechanism [41, 65] to efficiently model these feature correlations. Particularly, the model " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " is formed by " + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "inline_equation", + "content": "L_{\\phi}" + }, + { + "bbox": [ + 313, + 271, + 555, + 403 + ], + "type": "text", + "content": " blocks of self-attention as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 408, + 554, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 408, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 328, + 408, + 554, + 434 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {z} _ {0} = \\mathrm {L N} ([ \\boldsymbol {\\Delta} _ {1}, \\dots , \\boldsymbol {\\Delta} _ {\\mathrm {M}}, \\mathbf {c} ]) + \\boldsymbol {\\beta}, \\quad \\mathbf {a} _ {l} = \\mathbf {z} _ {l} + \\operatorname {M H S A} (\\mathbf {z}) \\\\ \\mathbf {z} _ {l + 1} = \\mathbf {a} _ {l} + \\operatorname {M L P} \\left(\\ln \\left(\\mathbf {a} _ {l}\\right)\\right), \\quad \\mathbf {u} = \\operatorname {P r o j} \\left(\\mathbf {z} _ {L _ {\\phi}}\\right) \\tag {7} \\\\ \\end{array}", + "image_path": "fa3aa6622ea2b93b2a2ef98e5849fe84374cf31f5531d515c3b1622232fef9a7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": " is the positional embedding, LN is Layer Normalization, MHSA is multi-head self-attention, MLP is the multi-layer perception, and Proj is the linear projection. By using Transformers, the correlation of cluster distributions can be well modeled by the self-attention mechanism." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "text", + "content": "Cluster Assignment via Visual Grammar. Instead of assigning the clusters based on the model prediction [3, 20, 45] or nearest cluster [30, 62] that are less effective, the cluster assignment in our approach will be performed by the visual grammar model, i.e., the visual grammar model will consider the " + }, + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "text", + "content": " closest features around cluster " + }, + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 498, + 556, + 629 + ], + "type": "text", + "content": " to assign the cluster for these features. Then, the cluster assignments are used to compute our Fairness Contrastive Clustering loss. In addition, following common practices [3, 20, 45], we improve background shift modeling by using the cluster assignments of features as the pseudo labels of pixels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Unknown Cluster Initialization. Prior work [30, 62] initialized a single unknown cluster " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "(N_U = 1)" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": ", thus resulting in producing non-discriminative class-wise features. However, there should be more than a single unknown cluster " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "(N_U > 1)" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " to produce discriminative features for different unknown classes. Therefore, our approach first initializes a list of potential unknown clusters at each learning step via" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15069" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "DB-SCAN [22] on the features of unknown classes extracted by the current CSS model. For the new known class " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ", we initialize these clusters based on the mean of their feature representations. Meanwhile, the clusters of known classes learned in previous steps are maintained. Further details can be found in the appendix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 222, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 222, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 222, + 164 + ], + "type": "text", + "content": "3.4. Continual Learning Procedure" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "content": "Fig. 3 illustrates the training procedure of our continual learning approach. At each learning step " + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "content": ", the CSS model " + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "content": " is trained with the Fairness Contrastive Clustering loss defined in Eqn. (4) and the previous visual grammar model " + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\Theta_{t-1}" + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "content": ". In addition, we introduce a cluster regularizer " + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C}" + }, + { + "bbox": [ + 55, + 168, + 296, + 264 + ], + "type": "text", + "content": " to avoid the clusters of different classes collapsing into a single cluster. Therefore, the entire CSS learning objective in our approach can be formed as in Eqn. (8)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 297, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 297, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 297, + 294 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\boldsymbol {\\theta} _ {t}} \\mathbb {E} _ {\\mathbf {x} ^ {t}, \\hat {\\mathbf {y}} ^ {t}} \\left[ \\mathcal {L} _ {C E} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} ^ {t}\\right) + \\lambda_ {C L} \\sum_ {\\mathbf {c} _ {i}} \\mathcal {L} _ {C o n t} ^ {\\alpha} \\left(\\mathbf {F} ^ {t}, \\mathbf {c} _ {i}\\right) + \\lambda_ {C} \\mathcal {R} _ {C} (\\mathbf {c}) \\right] \\tag {8}", + "image_path": "26e4319dc39581619986ebde6cfe7c036b81fa7ffe8b9c75392d1653ea585daa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_C(\\mathbf{c}) = \\sum_{\\mathbf{c}_i,\\mathbf{c}_j}\\{\\max (0,2\\nabla -||\\mathbf{c}_i - \\mathbf{c}_j||)\\} ^2" + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": " is the regularizer to avoid the cluster collapsing, " + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\lambda_C" + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": " is the balanced weight, and " + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": " is the margin between clusters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": "Training Procedure of Visual Grammar Model. At CSS learning step " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": ", we adopt the visual grammar model trained on the previous learning step, i.e., " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\Theta_{t-1}" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": ", to perform the cluster assignment for the contrastive clustering loss defined in Eqn. (2). Then, the visual grammar model " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\Theta_t" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": " at learning step " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": " will be learned (initialized from " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\Theta_{t-1}" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": ") on the features extracted from the dataset and the set of known clusters " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": " up to the current learning step. Following [41], we sample a center " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": " from the known clusters and its " + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": " closest features to train the visual grammar model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "content": "Initial Visual Grammar Model. At the first learning step " + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "content": ", since no clusters have been learned initially, the visual grammar model " + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "content": " is unavailable. However, as common practices in CSS [3, 4, 20], the segmentation model is typically trained from a pre-trained backbone on ImageNet [18]. As a result, the features extracted at the first learning step are characterized by the ImageNet features. Therefore, we adopt this philosophy to initialize our visual grammar model " + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "inline_equation", + "content": "(\\phi" + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "content": ") by pre-training the visual grammar model on the ImageNet dataset. Then, during CSS training, we will progressively train our visual grammar model at each learning step as aforementioned." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 137, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 137, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 137, + 618 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 278, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 278, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 278, + 636 + ], + "type": "text", + "content": "4.1. Implementations and Evaluation Protocols" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Implementation. Following common practices [2, 20, 62], we adopt DeepLab-V3 [6] with ResNet-101 [26] and SegFormer [69] with MiT-B3 [69] in our experiments. For the Visual Grammar model, we adopt the design of [41] with " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "L_{\\phi} = 12" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " blocks of multi-head self-attention layers. The feature vectors from the last layer of the decoder are used" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 82, + 552, + 173 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 545, + 81 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 545, + 81 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 545, + 81 + ], + "type": "text", + "content": "Table 1. Effectiveness of Fairness Contrastive Learning Loss." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 82, + 552, + 173 + ], + "lines": [ + { + "bbox": [ + 317, + 82, + 552, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 82, + 552, + 173 + ], + "type": "table", + "html": "
(a) ADE20K 100-50
\\( \\mathcal{L}_{Cont} \\)\\( \\mathcal{L}_{Cont}^{\\alpha} \\)0-100100-150allMajor Minor
✓44.615.234.851.5 26.4
✓44.624.537.952.1 30.8
(b) ADE20K 100-10
\\( \\mathcal{L}_{Cont} \\)\\( \\mathcal{L}_{Cont}^{\\alpha} \\)0-100100-150allMajor Minor
✓41.916.033.249.9 24.9
✓44.420.436.451.8 28.7
", + "image_path": "80a7d749960cd6df7605a550fa69e8763d55afffefb39c8f247ad1c6e8cb82e1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "text", + "content": "for our " + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}^{\\alpha}" + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "text", + "content": " loss. The value " + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "text", + "content": " is set individually for each dataset, i.e., " + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "inline_equation", + "content": "\\alpha = 5 \\times 10^{-2}" + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "text", + "content": " for ADE20K, " + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "inline_equation", + "content": "\\alpha = 10^{-2}" + }, + { + "bbox": [ + 313, + 177, + 553, + 224 + ], + "type": "text", + "content": " for VOC for Cityscapes. The details of our hyper-parameters are provided in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 225, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 429 + ], + "type": "text", + "content": "Evaluation Protocols. We evaluate models on three standard datasets of CSS, i.e., ADE20K [76], Pascal VOC [23], and Cityscapes [16]. Following common practices [3, 4, 62], our experiments are conducted on the overlapped CSS settings. In particular, on ADE20K, we use three different settings, i.e., ADE20K 100-50 (2 steps), ADE20K 100-10 (6 steps), and ADE20K 100-5 (11 steps). On Pascal VOC, we evaluate FALCON in three benchmarks, i.e., VOC 15-5 (2 steps), VOC 15-1 (6 steps), and VOC 10-1 (11 steps). On Cityscapes, we conduct domain incremental experiments with three settings, i.e., Cityscapes 11-5 (3 steps), Cityscapes 11-1 (11 steps), and Cityscapes 1-1 (21 steps). Following [3, 20], the mean Intersection over Union (mIoU) metric is adopted in our comparison, including mIoU of the last learning step on initial classes, incremental classes, and all classes. In addition, to illustrate the fairness improvement, we report the mIoU of major and minor classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 436, + 408, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 408, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 408, + 449 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": "Effectiveness of Fairness Contrastive Clustering. Table 1 presents our results using DeepLab-V3 [7] with Resnet101 on ADE20K 100-50 and ADE20K 100-10 benchmarks. We evaluate the impact of the fairness contrastive clustering loss " + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}^{\\alpha}" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": " by comparing it with the vanilla contrastive clustering loss " + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": ". As shown in our results, the overall performance has been significantly improved to " + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "inline_equation", + "content": "37.9\\%" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "inline_equation", + "content": "36.4\\%" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": " on ADE20K 100-50 and ADE20K 100-10, respectively. In addition, the fairness has been promoted due to the mIoU improvement of minor groups. We also study the impact of network backbones and cluster margin " + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": " in our appendix." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 597, + 552, + 712 + ], + "blocks": [ + { + "bbox": [ + 356, + 587, + 511, + 597 + ], + "lines": [ + { + "bbox": [ + 356, + 587, + 511, + 597 + ], + "spans": [ + { + "bbox": [ + 356, + 587, + 511, + 597 + ], + "type": "text", + "content": "Table 2. Effectiveness of Visual Grammar." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 597, + 552, + 712 + ], + "lines": [ + { + "bbox": [ + 317, + 597, + 552, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 597, + 552, + 712 + ], + "type": "table", + "html": "
(a) ADE20K 100-50
0-100101-150allMajorMinor
Nearest Cluster44.311.533.451.524.3
Fixed φ44.617.635.652.027.4
Adaptive φ44.624.537.952.130.8
(b) ADE20K 100-10
0-100101-150allMajorMinor
Nearest Cluster40.114.331.548.722.9
Fixed φ43.018.534.950.627.0
Adaptive φ44.420.436.451.828.7
", + "image_path": "6de08428c46c8ade2cd12713a846b9726af096f3c46affbe9b77ce6c41f87f8e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15070" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 93, + 293, + 147 + ], + "blocks": [ + { + "bbox": [ + 97, + 71, + 253, + 81 + ], + "lines": [ + { + "bbox": [ + 97, + 71, + 253, + 81 + ], + "spans": [ + { + "bbox": [ + 97, + 71, + 253, + 81 + ], + "type": "text", + "content": "Table 3. Effectiveness of Scaling Factor " + }, + { + "bbox": [ + 97, + 71, + 253, + 81 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 97, + 71, + 253, + 81 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 83, + 212, + 92 + ], + "lines": [ + { + "bbox": [ + 139, + 83, + 212, + 92 + ], + "spans": [ + { + "bbox": [ + 139, + 83, + 212, + 92 + ], + "type": "text", + "content": "(a) ADE20K 100-50" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 93, + 293, + 147 + ], + "lines": [ + { + "bbox": [ + 59, + 93, + 293, + 147 + ], + "spans": [ + { + "bbox": [ + 59, + 93, + 293, + 147 + ], + "type": "table", + "html": "
α0-1516-20allMajorMinor
α = 0.143.119.835.350.627.7
α = 0.0544.624.537.952.130.8
α = 0.0143.621.336.251.028.7
α = 0.00542.418.634.550.126.6
", + "image_path": "0d0fdd972919bf209c99284a59c094390b9351735b8b133afa5b3f2fb9c4ed29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 156, + 293, + 210 + ], + "blocks": [ + { + "bbox": [ + 139, + 147, + 212, + 155 + ], + "lines": [ + { + "bbox": [ + 139, + 147, + 212, + 155 + ], + "spans": [ + { + "bbox": [ + 139, + 147, + 212, + 155 + ], + "type": "text", + "content": "(b) Pascal VOC 15-5" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 156, + 293, + 210 + ], + "lines": [ + { + "bbox": [ + 59, + 156, + 293, + 210 + ], + "spans": [ + { + "bbox": [ + 59, + 156, + 293, + 210 + ], + "type": "table", + "html": "
α0-1516-20allMajorMinor
α = 0.174.851.669.376.963.5
α = 0.0576.251.370.379.063.8
α = 0.0179.454.873.581.367.7
α = 0.00574.648.968.577.661.6
", + "image_path": "fe99027b6c7156492607cdcd54903336cfb9026989a3e6b364f3324571dbc989.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "text", + "content": "Effectiveness of Scaling Factor of Cluster. Table 3 illustrates the experimental results of the impact of different scaling factor " + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "text", + "content": " on ADE20K 100-50 and Pascal VOC 15-5 benchmarks. As shown in Table 3, when the value of scaling factor " + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "text", + "content": " gradually decreases, the performance of our proposed approach is improved accordingly since the fairness contrastive loss in Eqn (4) tends to be more uniform across major and minor classes. However, when the scaling factor is too small (" + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\alpha = 0.005" + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "text", + "content": "), the impact of the loss enforcement becomes weaker leading to the weaker enforcement of the fairness contrastive clustering, resulting in lower overall performance. In addition, we have observed that the higher the number of classes demands the higher the value of " + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 217, + 296, + 384 + ], + "type": "text", + "content": " since it will increase the compactness of more clusters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "content": "Effectiveness of Loss Contributions. Table 4 illustrates the contributions of proposed learning objectives. For the model without using visual grammar, we only use a single unknown cluster " + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "inline_equation", + "content": "(N_U = 1)" + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "content": " and adopt the nearest cluster strategies to assign clusters of unknown pixels. By using only cross-entropy loss, the mIoU performance remains low due to catastrophic forgetting and background shift problems. Meanwhile, with our fairness clustering loss " + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Cont}^{\\alpha}" + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "content": ", visual grammar model " + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "content": ", and the cluster regularizer " + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "content": ", the mIoU performance has been significantly improved to " + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "inline_equation", + "content": "37.9\\%" + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "inline_equation", + "content": "36.4\\%" + }, + { + "bbox": [ + 55, + 386, + 296, + 553 + ], + "type": "text", + "content": " on ADE20K 100-50 and ADE20K 100-10, respectively. Moreover, our FALCON has significantly promoted the fairness of segmentation models illustrated by the mIoU improvement of major and minor groups." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 579 + ], + "type": "text", + "content": "Effectiveness of Visual Grammar. We evaluate FALCON under three settings, i.e., Nearest Cluster, Fixed " + }, + { + "bbox": [ + 55, + 555, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 555, + 296, + 579 + ], + "type": "text", + "content": " pretrained" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 595, + 293, + 654 + ], + "blocks": [ + { + "bbox": [ + 89, + 583, + 261, + 594 + ], + "lines": [ + { + "bbox": [ + 89, + 583, + 261, + 594 + ], + "spans": [ + { + "bbox": [ + 89, + 583, + 261, + 594 + ], + "type": "text", + "content": "Table 4. Effectiveness of Our Proposed Losses." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 595, + 293, + 654 + ], + "lines": [ + { + "bbox": [ + 58, + 595, + 293, + 654 + ], + "spans": [ + { + "bbox": [ + 58, + 595, + 293, + 654 + ], + "type": "table", + "html": "
(a) ADE20K 100-50
\\( \\mathcal{L}_{CE} \\)\\( \\mathcal{L}_{Cont}^{\\alpha } \\)\\( \\phi \\)\\( \\mathcal{R}_{C} \\)0-100101-150allMajorMinor
✓0.018.96.30.09.4
✓✓44.07.931.951.622.1
✓✓✓43.821.836.451.129.1
✓✓✓✓44.624.537.952.130.8
", + "image_path": "5a5141940d57570322073bf2076ff029cfa38919a60e6449d274a974127809ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 59, + 662, + 293, + 711 + ], + "blocks": [ + { + "bbox": [ + 143, + 654, + 208, + 662 + ], + "lines": [ + { + "bbox": [ + 143, + 654, + 208, + 662 + ], + "spans": [ + { + "bbox": [ + 143, + 654, + 208, + 662 + ], + "type": "text", + "content": "(b) ADE20K 100-10" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 662, + 293, + 711 + ], + "lines": [ + { + "bbox": [ + 59, + 662, + 293, + 711 + ], + "spans": [ + { + "bbox": [ + 59, + 662, + 293, + 711 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{CE} \\)\\( \\mathcal{L}_{Cont}^{\\alpha } \\)\\( \\phi \\)\\( \\mathcal{R}_C \\)0-100101-150allMajorMinor
✓0.03.51.20.01.8
✓✓39.013.130.447.821.6
✓✓✓43.418.535.151.227.1
✓✓✓✓44.420.436.451.828.7
", + "image_path": "7490d9b1e1e3de4bd66522df0318143a857798a49a6b59f3befa885dac3659ad.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 173 + ], + "type": "image", + "image_path": "0ff1001ed48c41e4d2174c4832b287136ba02bd962dbe75b777e76f28cce5aa4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 207 + ], + "type": "text", + "content": "Figure 6. Cluster Distribution at Learning Step " + }, + { + "bbox": [ + 313, + 175, + 553, + 207 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 175, + 553, + 207 + ], + "type": "text", + "content": " of ADE20K 100-50 (classes 109-144 are future classes) without (left) and with (right) Fairness Learning via Contrastive Attention." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "content": "ImageNet (without updating on each learning step), and Adaptive " + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "content": " (with updating on each learning step). As in Table 2, the mIoU result using only the nearest cluster remains ineffective. Meanwhile, the adaptive visual grammar model updated at each learning step further boosts the mIoU performance and promotes fairness, i.e., increased by " + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "content": " on ADE20K 100-50 and ADE20K 100-10 compared to the nearest cluster approach. In addition, we study the impact of choosing the number of features " + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "content": " in the visual grammar model in our appendix. Fig. 6 illustrates the feature distributions of unknown classes (future class). As a result, our FALCON approach is able to model features of unknown classes into different clusters and produce better and more compact clusters compared to the one without Fairness Learning via Contrastive Attention." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 392, + 556, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 513 + ], + "type": "text", + "content": "Effectiveness of Forgetting Each Learning Step. Table 5 reports the results of our model evaluated at each learning step. The performance of our model remains stable at each learning step. In particular, the maximum performance drop at learning step 4 of classes 111-120 due forgetting is only " + }, + { + "bbox": [ + 313, + 392, + 556, + 513 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 392, + 556, + 513 + ], + "type": "text", + "content": ". Meanwhile, at the last learning task, the mIoU results of old classes are well maintained, with a slight drop of " + }, + { + "bbox": [ + 313, + 392, + 556, + 513 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 392, + 556, + 513 + ], + "type": "text", + "content": " in classes 131-140. As shown in the result, our model has effectively maintained the performance to avoid catastrophic forgetting while enabling the capability of learning classes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 518, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 518, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 518, + 534 + ], + "type": "text", + "content": "4.3. Comparison with Prior SoTA Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 540, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 556, + 612 + ], + "type": "text", + "content": "ADE20K. Table 6 presents our experimental results compared to prior CSS methods. By using DeepLab-V3, our approach has achieved SoTA performance, i.e., the mIoU results of " + }, + { + "bbox": [ + 313, + 540, + 556, + 612 + ], + "type": "inline_equation", + "content": "37.9\\%" + }, + { + "bbox": [ + 313, + 540, + 556, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 540, + 556, + 612 + ], + "type": "inline_equation", + "content": "+36.4\\%" + }, + { + "bbox": [ + 313, + 540, + 556, + 612 + ], + "type": "text", + "content": " on ADE20K 100-50 and ADE20K 100-10 benchmarks, higher than prior FairCL [62]. Meanwhile, FALCON using Transformer has outperformed" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 317, + 632, + 552, + 711 + ], + "blocks": [ + { + "bbox": [ + 322, + 619, + 544, + 631 + ], + "lines": [ + { + "bbox": [ + 322, + 619, + 544, + 631 + ], + "spans": [ + { + "bbox": [ + 322, + 619, + 544, + 631 + ], + "type": "text", + "content": "Table 5. Effectiveness of CSS Model At Each Learning Step." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 632, + 552, + 711 + ], + "lines": [ + { + "bbox": [ + 317, + 632, + 552, + 711 + ], + "spans": [ + { + "bbox": [ + 317, + 632, + 552, + 711 + ], + "type": "table", + "html": "
ADE20K 100-10
0-100101-110111-120121-130131-140141-150all
Step 144.9-----44.9
Step 244.815.8----42.2
Step 344.715.722.1---40.3
Step 444.615.621.326.3--39.2
Step 544.515.421.226.124.4-37.9
Step 644.415.421.226.124.015.536.4
avg40.1
", + "image_path": "ac24a0023f80c534a767ed4b8885945355b3ed69e5ac178c4a84c66a93a67595.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15071" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 115, + 346, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 346, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 346, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 346, + 114 + ], + "type": "text", + "content": "Table 6. Comparison with Prior Methods on ADE20K Benchmarks (Note: The results of MiB [2], PLOP [20], and FairCL [62] using Transformer on ADE20K 100-5 were not reported in prior studies. The upper bound results are not trained with fairness objective)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 115, + 346, + 262 + ], + "lines": [ + { + "bbox": [ + 58, + 115, + 346, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 115, + 346, + 262 + ], + "type": "table", + "html": "
NetworkMethodADE20K 100-50ADE20K 100-10ADE20K 100-5
0-100101-150allavg0-100101-150allavg0-100101-150allavg
DeepLab-V3PLOP [20]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [75]42.318.834.5-39.317.632.1-38.511.529.6-
REIMINDER [45]41.619.234.1-39.021.333.1-36.116.429.5-
RCIL+LGKD [73]43.325.137.2-42.220.434.9-----
FairCL [62]43.424.037.040.541.720.434.739.0----
FALCON44.624.537.941.344.420.436.440.138.016.130.737.6
Upper Bound44.328.238.9-44.328.238.9-44.328.238.9-
Mask2FormerMiB [2]37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP [20]44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer [3]44.726.238.441.240.615.632.337.439.513.630.936.5
ECLIPSE [31]45.021.737.1-43.417.434.6-43.316.334.2-
CoMasTRe [25]45.726.039.241.642.318.434.438.440.815.832.638.6
TransformerMiB [2]43.430.639.238.739.120.434.239.5----
PLOP [20]43.826.238.038.143.324.136.240.3----
FairCL [62]43.625.537.640.742.221.935.539.4----
FALCON47.530.641.943.547.326.240.342.840.818.933.538.1
Upper Bound48.739.045.5-48.739.045.5-48.739.045.5-
", + "image_path": "afa9f66a1a896043680f1bae18cc2f78596ed042b5d902e421856661aa4865e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "text", + "content": "the prior CoMFormer [3] model by " + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "inline_equation", + "content": "+3.5\\%" + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "inline_equation", + "content": "+8.0\\%" + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "text", + "content": " on ADE20K 100-50, ADE20K 100-10, and ADE20K 100-5. In addition, our mIoU results on the initial classes remain competitive with the upper-bounded results because FALCON can well handle the fairness problem compared to fully supervised learning. Although ECLIPSE [31] exhibits slightly better on ADE20K 100-5 by only " + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 54, + 265, + 294, + 409 + ], + "type": "text", + "content": " (all) due to the better segmentation framework (i.e., Mask2Former), FALCON significantly outperforms ECLIPSE on ADE20K 100-50 and ADE20K 100-10. We also report our results of ADE20K 50-50 in the appendix. As in Fig. 7, FALCON produces better segmentation maps than prior methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "text", + "content": "Pascal VOC. Table 7 presents our results on Pascal VOC benchmarks. FALCON has consistently achieved the SoTA performance on three benchmarks. In particular, compared to the prior FairCL [62] approach, our methods using DeepLab-V3 have improved the mIoU performance up to " + }, + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "inline_equation", + "content": "73.50\\%" + }, + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "inline_equation", + "content": "69.83\\%" + }, + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "inline_equation", + "content": "62.41\\%" + }, + { + "bbox": [ + 55, + 410, + 297, + 517 + ], + "type": "text", + "content": " on Pascal VOC 15-5, Pascal VOC 15-1, and Pascal VOC 10-1, respectively. Additionally, by using the better backbone, the performance is further improved, and the gap with the upper-bound result is reduced." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 553 + ], + "type": "text", + "content": "Cityscapes. Table 8 reports the performance of our approach using DeepLab-V3 compared to prior methods on three different settings, i.e., Cityscapes 11-5, Cityscapes" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 351, + 78, + 552, + 233 + ], + "blocks": [ + { + "bbox": [ + 351, + 78, + 552, + 233 + ], + "lines": [ + { + "bbox": [ + 351, + 78, + 552, + 233 + ], + "spans": [ + { + "bbox": [ + 351, + 78, + 552, + 233 + ], + "type": "image", + "image_path": "0f4de9087e49e5ac462c81197945df54f737630d27614a25cc819ded3bf2bf14.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 234, + 552, + 256 + ], + "lines": [ + { + "bbox": [ + 350, + 234, + 552, + 256 + ], + "spans": [ + { + "bbox": [ + 350, + 234, + 552, + 256 + ], + "type": "text", + "content": "Figure 7. The Comparison of Qualitative Results on ADE20K Between Our Approach and RCIL [75]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "text", + "content": "11-1, and Cityscapes 1-1. As in Table 8, the performance of our methods has consistently outperformed prior FairCL [62] approach by " + }, + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "inline_equation", + "content": "+3.78\\%" + }, + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "inline_equation", + "content": "+3.14\\%" + }, + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "inline_equation", + "content": "+6.02\\%" + }, + { + "bbox": [ + 313, + 265, + 555, + 326 + ], + "type": "text", + "content": " on three benchmarks. Similar to our experiments on ADE20K and VOC, the better network brings higher results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 476, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 476, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 476, + 344 + ], + "type": "text", + "content": "5. Conclusions and Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 347, + 556, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 556, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 556, + 505 + ], + "type": "text", + "content": "Conclusions. This paper has presented a novel Fairness Learning via Contrastive Attention approach to CSS. In particular, the fairness contrastive clustering loss has been introduced to model both catastrophic forgetting and fairness problems. Then, the visual grammar model was presented to model the unknown classes. The experimental results on different benchmarks have shown the SoTA performance and fairness promotion of our proposed FALCON approach. Limitations. Our study chose a set of learning hyperparameters to support our theoretical analysis. However, it potentially consists of several limitations related to choosing learning parameters and cluster initialization. Further discussion of limitations is presented in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 506, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 553 + ], + "type": "text", + "content": "Acknowledgment This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 59, + 570, + 354, + 711 + ], + "blocks": [ + { + "bbox": [ + 99, + 558, + 312, + 568 + ], + "lines": [ + { + "bbox": [ + 99, + 558, + 312, + 568 + ], + "spans": [ + { + "bbox": [ + 99, + 558, + 312, + 568 + ], + "type": "text", + "content": "Table 7. Comparisons with Prior Methods on Pascal VOC." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 570, + 354, + 711 + ], + "lines": [ + { + "bbox": [ + 59, + 570, + 354, + 711 + ], + "spans": [ + { + "bbox": [ + 59, + 570, + 354, + 711 + ], + "type": "table", + "html": "
MethodPascal VOC 15-5Pascal VOC 15-1Pascal VOC 10-1
0-1516-20all0-1516-20all0-1011-20all
DeepLab-V3MiB [2]76.3749.9770.0838.0013.5032.2020.0020.1020.10
PLOP [20]75.7351.7170.0965.1021.1054.6044.0015.5030.50
RCIL [75]---70.6023.7059.4055.4015.1034.30
FairCL [62]---72.0022.7060.3042.3025.6034.40
SSUL [4]77.8250.1071.2277.3136.5967.6171.3145.9859.25
FALCON79.3554.7773.5078.3442.5769.8373.9449.7362.41
Upper Bound79.7772.3577.4379.7772.3577.4378.4176.3577.43
TransformerPLOP [20]72.5148.3766.7664.5937.2358.0848.5333.7141.47
SSUL [4]79.9156.8374.4179.9140.5670.5474.0651.8563.48
FairCL [62]---73.5022.8061.5057.1014.2036.60
FALCON81.2058.0475.6978.7147.5471.2874.9252.5464.26
Upper Bound80.8474.9779.4480.8474.9779.4480.8474.9779.44
", + "image_path": "56ac49bb9da1ae40a40fca599a264c4b145c3f9eae9596afe151594ecded11b7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 369, + 570, + 549, + 711 + ], + "blocks": [ + { + "bbox": [ + 390, + 558, + 526, + 569 + ], + "lines": [ + { + "bbox": [ + 390, + 558, + 526, + 569 + ], + "spans": [ + { + "bbox": [ + 390, + 558, + 526, + 569 + ], + "type": "text", + "content": "Table 8. Comparisons on Cityscapes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 369, + 570, + 549, + 711 + ], + "lines": [ + { + "bbox": [ + 369, + 570, + 549, + 711 + ], + "spans": [ + { + "bbox": [ + 369, + 570, + 549, + 711 + ], + "type": "table", + "html": "
Method11-511-11-1
DeepLab-V3LWF-MC [48]58.9056.9231.24
ILT [38]59.1457.7530.11
MIB [2]61.5160.0242.15
PLOP [20]63.5162.0545.24
RCIL [75]64.3063.0048.90
FairCL [62]66.9666.6149.22
FALCON70.7469.7555.24
Upper Bound79.3079.3079.30
Trans.FairCL [62]67.8567.0955.68
FALCON71.3370.1458.79
Upper Bound83.8083.8083.80
", + "image_path": "b207b321aa66a8938add9cc1bb0eb473b97ce1685e98c69e90c50ac6c1fcdcf1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15072" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Nikita Araslanov, and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. Conf. Comp. Vision Pattern Rec., 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "text", + "content": "[3] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. IEEE/CVF Computer Vision and Pattern Recognition Conference, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 296, + 278 + ], + "type": "text", + "content": "[4] Sungmin Cha, beomyoung kim, YoungJoon Yoo, and Taesup Moon. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In Advances in Neural Information Processing Systems, pages 10919-10930. Curran Associates, Inc., 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "type": "text", + "content": "[5] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. arXiv preprint arXiv:2407.16354, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. TPAMI, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 391, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 445 + ], + "type": "text", + "content": "[8] Mayee Chen, Daniel Y Fu, Avanika Narayan, Michael Zhang, Zhao Song, Kayvon Fatahalian, and Christopher Ré. Perfectly balanced: Improving transfer and robustness of supervised contrastive learning. In International Conference on Machine Learning, pages 3090-3122. PMLR, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 491, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 296, + 523 + ], + "type": "text", + "content": "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 525, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 555 + ], + "type": "text", + "content": "[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "type": "text", + "content": "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "text", + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[14] Sanghyeok Chu, Dongwan Kim, and Bohyung Han. Learning debiased and disentangled representations for semantic segmentation. Advances in Neural Information Processing Systems, 34:8355-8366, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[15] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for in-" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "cremental semantic segmentation. In European Conference on Computer Vision, pages 244-261. Springer, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 140 + ], + "type": "text", + "content": "[16] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 555, + 184 + ], + "type": "text", + "content": "[17] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 217 + ], + "type": "text", + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 220, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 555, + 262 + ], + "type": "text", + "content": "[19] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proc. Europ. Conf. Computer Vision, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "text", + "content": "[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4040-4050, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 320, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 555, + 374 + ], + "type": "text", + "content": "[21] Beyza Ermis, Giovanni Zappella, Martin Wistuba, Aditya Rawal, and Cedric Archambeau. Continual learning with transformers for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3774-3781, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 376, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 555, + 419 + ], + "type": "text", + "content": "[22] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 421, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 555, + 453 + ], + "type": "text", + "content": "[23] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 455, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 555, + 476 + ], + "type": "text", + "content": "[24] Robert French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 1999." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 478, + 555, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 555, + 531 + ], + "type": "text", + "content": "[25] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3857, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 533, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 555, + 555 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "type": "text", + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 613, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 644 + ], + "type": "text", + "content": "[28] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. DAFormer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 555, + 689 + ], + "type": "text", + "content": "[29] Ibsa Jalata, Naga Venkata Sai Raviteja Chappa, Thanh-Dat Truong, Pierce Helton, Chase Rainwater, and Khoa Luu. Eqadap: Equipollent domain adaptation approach to image deblurring. IEEE Access, 10:93203-93211, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 691, + 555, + 713 + ], + "type": "text", + "content": "[30] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detect" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15073" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "tion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5830-5840, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "type": "text", + "content": "[31] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "text", + "content": "[32] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. PNAS, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 274 + ], + "type": "text", + "content": "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 296, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 296, + 318 + ], + "type": "text", + "content": "[34] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 8547-8555, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 363 + ], + "type": "text", + "content": "[35] Chang Liu, Giulia Rizzoli, Pietro Zanuttigh, Fu Li, and Yi Niu. Learning from the web: Language drives weakly-supervised incremental learning for semantic segmentation. arXiv preprint arXiv:2407.13363, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "text", + "content": "[36] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 421 + ], + "type": "text", + "content": "[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In NeurIPS, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 443 + ], + "type": "text", + "content": "[38] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCVWS, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 296, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 488 + ], + "type": "text", + "content": "[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 296, + 533 + ], + "type": "text", + "content": "[40] Pha Nguyen, Thanh-Dat Truong, Miaoqing Huang, Yi Liang, Ngan Le, and Khoa Luu. Self-supervised domain adaptation in crowd counting. In 2022 IEEE international conference on image processing (ICIP), pages 2786-2790. IEEE, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 599 + ], + "type": "text", + "content": "[41] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10847-10856, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 634 + ], + "type": "text", + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "text", + "content": "[43] First Ozdemir and Orcun Goksel. Extending pretrained segmentation networks with additional anatomical structures. International journal of computer assisted radiology and surgery, 14:1187-1195, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[44] First Ozdemir, Philipp Fuernstahl, and Orcun Goksel. Learn the new, keep the old: Extending pretrained models with new anatomy and images. In Medical Image Computing" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part IV 11, pages 361-369. Springer, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "text", + "content": "[45] Minh Hieu Phan, The-Anh Ta, Son Lam Phung, Long Tran-Thanh, and Abdesselam Bouzerdoum. Class similarity weighted knowledge distillation for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16866-16875, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 185, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 228 + ], + "type": "text", + "content": "[46] Yiqiao Qiu, Yixing Shen, Zhuohao Sun, Yanchong Zheng, Xiaobin Chang, Weishi Zheng, and Ruixuan Wang. Sats: Self-attention transfer for continual semantic segmentation. Pattern Recognition, 138:109383, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 555, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 340 + ], + "type": "text", + "content": "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proc. Conf. Comp. Vision Pattern Rec., 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 374 + ], + "type": "text", + "content": "[49] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 397 + ], + "type": "text", + "content": "[50] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 420 + ], + "type": "text", + "content": "[51] Mohammad Rostami. Lifelong domain adaptation via consolidated internal distribution. In NeurIPS, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 422, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 476 + ], + "type": "text", + "content": "[52] Antoine Saporta, Arthur Douillard, Tuan-Hung Vu, Patrick Pérez, and Matthieu Cord. Multi-head distillation for continual unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "type": "text", + "content": "[53] Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. Incremental learning of object detectors without catastrophic forgetting. In Proceedings of the IEEE international conference on computer vision, pages 3400-3409, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 522, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 555, + 567 + ], + "type": "text", + "content": "[54] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 568, + 555, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 621 + ], + "type": "text", + "content": "[55] Attila Szabó, Hadi Jamali-Rad, and Siva-Datta Mannava. Tilted cross-entropy (tce): Promoting fairness in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2305–2310, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 645 + ], + "type": "text", + "content": "[56] Sebastian Thrun. Lifelong learning algorithms. In *Springer Learning to Learn*, 1998." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 679 + ], + "type": "text", + "content": "[57] Thanh-Dat Truong and Khoa Luu. Cross-view action recognition understanding from exocentric to egocentric perspective. Neurocomputing, 614:128731, 2025." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[58] Thanh-Dat Truong, Chi Nhan Duong, Khoa Luu, Minh-Triet Tran, and Ngan Le. Domain generalization via universal non-volume preserving approach. In 2020 17th Conference" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15074" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "on Computer and Robot Vision (CRV), pages 93-100. IEEE, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[59] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Son Lam Phung, Chase Rainwater, and Khoa Luu. Bimal: Bijective maximum likelihood approach to domain adaptation in semantic scene segmentation. In International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[60] Thanh-Dat Truong, Ravi Teja Nvs Chappa, Xuan-Bac Nguyen, Ngan Le, Ashley PG Dowling, and Khoa Luu. Otadapt: Optimal transport-based approach for unsupervised domain adaptation. In 2022 26th international conference on pattern recognition (ICPR), pages 2850-2856. IEEE, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "text", + "content": "[61] Thanh-Dat Truong, Ngan Le, Bhiksha Raj, Jackson Cothren, and Khoa Luu. Freedom: Fairness domain adaptation approach to semantic scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "text", + "content": "[62] Thanh-Dat Truong, Hoang-Quan Nguyen, Bhiksha Raj, and Khoa Luu. Fairness continual learning approach to semantic scene understanding in open-world environments. In NeurIPS, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 364 + ], + "type": "text", + "content": "[63] Thanh-Dat Truong, Pierce Helton, Ahmed Moustafa, Jackson David Cothren, and Khoa Luu. Conda: Continual unsupervised domain adaptation learning in visual perception for self-driving cars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 5642-5650, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "text", + "content": "[64] Thanh-Dat Truong, Utsav Prabhu, Dongyi Wang, Bhiksha Raj, Susan Gauch, Jeyamkondan Subbiah, and Khoa Luu. EAGLE: Efficient adaptive geometry-based learning in cross-view understanding. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "text", + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "text", + "content": "[66] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proc. Conf. Comp. Vision Pattern Rec., 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 567 + ], + "type": "text", + "content": "[67] Huyong Wang, Huisi Wu, and Jing Qin. Incremental nuclei segmentation from histopathological images via future-class awareness and compatibility-inspired distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11408-11417, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 633 + ], + "type": "text", + "content": "[68] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "text", + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[70] Zhengyuan Xie, Haiquan Lu, Jia-wen Xiao, Enguang Wang, Le Zhang, and Xialei Liu. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 453 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tation. In European Conference on Computer Vision, pages 183-201. Springer, 2025." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[71] Lei Yang, Xiaohang Zhan, Dapeng Chen, Junjie Yan, Chen Change Loy, and Dahua Lin. Learning to cluster faces on an affinity graph. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2298-2306, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[72] Lei Yang, Dapeng Chen, Xiaohang Zhan, Rui Zhao, Chen Change Loy, and Dahua Lin. Learning to cluster faces via confidence and connectivity estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13369-13378, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 273 + ], + "type": "text", + "content": "[73] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 307 + ], + "type": "text", + "content": "[74] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. arXiv preprint arXiv:2407.09838, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "type": "text", + "content": "[75] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 397 + ], + "type": "text", + "content": "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "text", + "content": "[77] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15075" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file